diff --git a/.dockerignore b/.dockerignore
new file mode 100644
index 000000000..fd870df8f
--- /dev/null
+++ b/.dockerignore
@@ -0,0 +1,43 @@
+# The server images intentionally receive only the Server tree and the selected
+# InsightFace inference package; unrelated repository data never enters the
+# Docker build context.
+**
+!server/
+!server/**
+!python-package/
+!python-package/insightface/
+!python-package/insightface/__init__.py
+!python-package/insightface/app/
+!python-package/insightface/app/**
+!python-package/insightface/model_zoo/
+!python-package/insightface/model_zoo/**
+!python-package/insightface/utils/
+!python-package/insightface/utils/**
+!python-package/insightface/data/
+!python-package/insightface/data/__init__.py
+!python-package/insightface/data/image.py
+!python-package/insightface/data/pickle_object.py
+!python-package/insightface/thirdparty/
+!python-package/insightface/thirdparty/__init__.py
+
+# Re-exclude local state and secrets below the whitelisted source trees.
+.private/
+server/.venv/
+server/.data/
+server/.models/
+server/.mypy_cache/
+server/.pytest_cache/
+server/.ruff_cache/
+server/build/
+server/test-results/
+**/__pycache__/
+**/*.py[cod]
+**/._*
+**/.DS_Store
+server/**/dist/
+**/*.onnx
+**/.env
+**/.env.*
+**/*.key
+**/*.pem
+!server/backend/insightface_server/licensing/trusted_keys/*.pem
diff --git a/.github/workflows/manual-python-package-ci.yml b/.github/workflows/manual-python-package-ci.yml
new file mode 100644
index 000000000..ffb9864c7
--- /dev/null
+++ b/.github/workflows/manual-python-package-ci.yml
@@ -0,0 +1,302 @@
+name: Manual Python Package CI
+
+on:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+env:
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+
+jobs:
+  build-package:
+    name: Build package / macOS / Python 3.11
+    runs-on: macos-latest
+    timeout-minutes: 30
+
+    defaults:
+      run:
+        shell: bash
+        working-directory: python-package
+
+    env:
+      PIP_DISABLE_PIP_VERSION_CHECK: "1"
+      PIP_NO_PYTHON_VERSION_WARNING: "1"
+      PIP_NO_CACHE_DIR: "1"
+      PYTHONUTF8: "1"
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Set up Python 3.11
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Upgrade packaging tools
+        run: |
+          python -m pip install --upgrade pip setuptools wheel build twine
+
+      - name: Build wheel and source distribution
+        run: |
+          rm -rf dist build *.egg-info
+          python -m build
+
+      - name: Check package metadata
+        run: |
+          python -m twine check dist/*
+
+      - name: Inspect built package contents
+        run: |
+          python - <<'PY'
+          from pathlib import Path
+          import tarfile
+          from zipfile import ZipFile
+
+          wheels = sorted(Path("dist").glob("insightface-1.0.1-*.whl"))
+          sdists = sorted(Path("dist").glob("insightface-1.0.1.tar.gz"))
+          assert wheels, "No wheel was built."
+          assert sdists, "No source distribution was built."
+          wheel_path = wheels[-1]
+          sdist_path = sdists[-1]
+
+          with ZipFile(wheel_path) as archive:
+              names = archive.namelist()
+              wheel_metadata = archive.read("insightface-1.0.1.dist-info/WHEEL").decode()
+              package_metadata = archive.read("insightface-1.0.1.dist-info/METADATA").decode()
+          with tarfile.open(sdist_path, "r:gz") as archive:
+              sdist_names = archive.getnames()
+
+          assert "Root-Is-Purelib: true" in wheel_metadata
+          assert "Tag: py3-none-any" in wheel_metadata
+          assert "Requires-Dist: onnxruntime" in package_metadata
+          assert "Requires-Dist: opencv-python" in package_metadata
+          assert "Provides-Extra: gui" in package_metadata
+          assert "Requires-Dist: PySide6-Essentials" in package_metadata
+          assert "Requires-Dist: reportlab" in package_metadata
+          assert "Requires-Dist: scikit-learn" in package_metadata
+          assert "Provides-Extra: face3d" in package_metadata
+          assert "Requires-Dist: matplotlib" in package_metadata
+          assert "Requires-Dist: pandas" not in package_metadata
+          assert "Requires-Dist: easydict" not in package_metadata
+          assert "Requires-Dist: prettytable" not in package_metadata
+          assert "License: MIT" not in package_metadata
+          assert "Requires-Dist: PySide6 " not in package_metadata
+          assert "insightface/data/images/t1.jpg" in names
+          assert "insightface/data/objects/meanshape_68.pkl" in names
+          assert "insightface/gui/assets/app_icon.png" in names
+          assert "insightface/gui/assets/app_icon.ico" in names
+          assert "insightface/gui/assets/app_icon.icns" in names
+          assert any(name.endswith("insightface/gui/assets/app_icon.png") for name in sdist_names)
+          assert any(name.endswith("insightface/gui/assets/app_icon.ico") for name in sdist_names)
+          assert any(name.endswith("insightface/gui/assets/app_icon.icns") for name in sdist_names)
+          assert not any(".data/data/insightface/data/" in name for name in names)
+          assert not any(name.endswith((".so", ".pyd", ".dll", ".dylib")) for name in names)
+
+          print(f"Package inspection passed: {wheel_path}, {sdist_path}")
+          PY
+
+      - name: Upload package artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          name: python-package-dist
+          path: python-package/dist/*
+
+  install-test:
+    name: Install test / ${{ matrix.os }} / Python ${{ matrix.python-version }}
+    needs: build-package
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 45
+
+    strategy:
+      fail-fast: false
+      matrix:
+        os:
+          - ubuntu-latest
+          - macos-latest
+          - windows-latest
+        python-version:
+          - "3.9"
+          - "3.10"
+          - "3.11"
+          - "3.12"
+
+    defaults:
+      run:
+        shell: bash
+        working-directory: python-package
+
+    env:
+      QT_QPA_PLATFORM: offscreen
+      PIP_DISABLE_PIP_VERSION_CHECK: "1"
+      PIP_NO_PYTHON_VERSION_WARNING: "1"
+      PIP_NO_CACHE_DIR: "1"
+      PYTHONUTF8: "1"
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v6
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install Linux GUI runtime libraries
+        if: runner.os == 'Linux'
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libegl1 \
+            libgl1 \
+            libglib2.0-0 \
+            libxkbcommon-x11-0 \
+            libxcb-cursor0 \
+            libxcb-icccm4 \
+            libxcb-image0 \
+            libxcb-keysyms1 \
+            libxcb-render-util0 \
+            libxcb-xinerama0
+        working-directory: .
+
+      - name: Download macOS-built package artifacts
+        uses: actions/download-artifact@v5
+        with:
+          name: python-package-dist
+          path: python-package/dist
+
+      - name: Install test tools
+        run: |
+          python -m pip install --upgrade pip setuptools wheel pytest
+
+      - name: Install built wheel with GUI extra
+        run: |
+          python - <<'PY'
+          import subprocess
+          import sys
+          from pathlib import Path
+
+          wheels = sorted(Path("dist").glob("insightface-1.0.1-*.whl"))
+          assert wheels, "No wheel artifact was downloaded."
+          wheel = wheels[-1]
+          subprocess.check_call([sys.executable, "-m", "pip", "install", f"{wheel}[gui]"])
+          print(f"Installed package artifact: {wheel}")
+          PY
+
+      - name: Check PySide6 Essentials platform plugins
+        run: |
+          python - <<'PY'
+          import os
+          import sys
+          from importlib import metadata
+          from pathlib import Path
+
+          assert metadata.version("PySide6-Essentials")
+          assert metadata.version("shiboken6")
+          for dist_name in ("PySide6", "PySide6-Addons"):
+              try:
+                  version = metadata.version(dist_name)
+              except metadata.PackageNotFoundError:
+                  continue
+              raise AssertionError(f"{dist_name} should not be installed in the GUI extra, found {version}")
+
+          from insightface.gui.app import configure_qt_plugin_paths, qt_plugin_root_candidates
+
+          os.environ.pop("QT_PLUGIN_PATH", None)
+          os.environ.pop("QT_QPA_PLATFORM_PLUGIN_PATH", None)
+          configure_qt_plugin_paths()
+
+          if not os.environ.get("QT_PLUGIN_PATH") or not os.environ.get("QT_QPA_PLATFORM_PLUGIN_PATH"):
+              raise AssertionError(
+                  "Qt plugin paths were not configured. "
+                  f"Candidates: {[str(path) for path in qt_plugin_root_candidates()]}"
+              )
+
+          plugin_root = Path(os.environ["QT_PLUGIN_PATH"])
+          platform_dir = Path(os.environ["QT_QPA_PLATFORM_PLUGIN_PATH"])
+          if sys.platform == "darwin":
+              expected = {"libqcocoa.dylib", "libqoffscreen.dylib"}
+          elif sys.platform.startswith("win"):
+              expected = {"qwindows.dll", "qoffscreen.dll"}
+          elif sys.platform.startswith("linux"):
+              expected = {"libqxcb.so", "libqoffscreen.so"}
+          else:
+              expected = set()
+
+          available = {path.name for path in platform_dir.iterdir()}
+          missing = expected - available
+          assert plugin_root.exists(), plugin_root
+          assert platform_dir.exists(), platform_dir
+          assert not missing, f"Missing Qt platform plugins: {sorted(missing)} in {platform_dir}"
+
+          print(f"PySide6 Essentials Qt plugin root: {plugin_root}")
+          print(f"Available platform plugins: {sorted(available)}")
+          PY
+
+      - name: Check installed dependencies
+        run: |
+          python -m pip check
+
+      - name: Verify installed package and entry points
+        run: |
+          python - <<'PY'
+          import subprocess
+          import sys
+          import tempfile
+
+          checks = [
+              [sys.executable, "-c", "import insightface, insightface.gui; assert insightface.__version__ == '1.0.1'; assert insightface.gui.__version__ == '1.0.1'"],
+              [sys.executable, "-m", "insightface.gui", "--version"],
+              ["insightface-gui", "--version"],
+              ["insightface-eval-studio", "--version"],
+              ["insightface-desktop", "--version"],
+          ]
+          with tempfile.TemporaryDirectory() as tmp:
+              for command in checks:
+                  output = subprocess.check_output(command, text=True, cwd=tmp)
+                  if command[-1] == "--version":
+                      assert "InsightFace Evaluation Studio 1.0.1" in output
+                      assert "insightface 1.0.1" in output
+          print("Installed package and console entry points verified.")
+          PY
+
+      - name: Copy GUI tests outside source tree
+        run: |
+          python - <<'PY'
+          import os
+          import shutil
+          from pathlib import Path
+
+          source = Path("tests/gui")
+          target = Path(os.environ["RUNNER_TEMP"]) / "insightface-gui-tests"
+          if target.exists():
+              shutil.rmtree(target)
+          shutil.copytree(source, target)
+          print(f"Copied GUI tests to {target}")
+          PY
+
+      - name: Run GUI and packaging tests against installed wheel
+        run: |
+          python - <<'PY'
+          import os
+          import subprocess
+          import sys
+          from pathlib import Path
+
+          test_dir = Path(os.environ["RUNNER_TEMP"]) / "insightface-gui-tests"
+          subprocess.check_call([sys.executable, "-m", "pytest", str(test_dir), "-q"], cwd=str(test_dir.parent))
+          PY
+
+      - name: Run multilingual GUI tests against installed wheel
+        run: |
+          python - <<'PY'
+          import os
+          import subprocess
+          import sys
+          from pathlib import Path
+
+          test_file = Path(os.environ["RUNNER_TEMP"]) / "insightface-gui-tests" / "test_i18n.py"
+          subprocess.check_call([sys.executable, "-m", "pytest", str(test_file), "-q"], cwd=str(test_file.parent.parent))
+          PY
diff --git a/.gitignore b/.gitignore
index 7bbc71c09..884cd7f03 100644
--- a/.gitignore
+++ b/.gitignore
@@ -99,3 +99,16 @@ ENV/
 
 # mypy
 .mypy_cache/
+
+# InsightFace Server local state and private biometric/model assets
+.private/
+evaluation_data/
+.pytest_cache/
+.ruff_cache/
+server/.data/
+server/.models/*
+!server/.models/.gitkeep
+server/test-results/
+*.onnx
+
+.DS_Store
diff --git a/.gitmodules b/.gitmodules
deleted file mode 100644
index 6c4c7f980..000000000
--- a/.gitmodules
+++ /dev/null
@@ -1,3 +0,0 @@
-[submodule "alignment/SDUNet"]
-	path = alignment/SDUNet
-	url = https://github.com/deepinsight/SDUNet
diff --git a/3rdparty/operator/amsoftmax-inl.h b/3rdparty/operator/amsoftmax-inl.h
deleted file mode 100644
index d899bbbfa..000000000
--- a/3rdparty/operator/amsoftmax-inl.h
+++ /dev/null
@@ -1,287 +0,0 @@
-/*!
- * Copyright (c) 2018 by Contributors
- * \file amsoftmax-inl.h
- * \brief AmSoftmax from <TODO>
- * \author Jia Guo
- */
-#ifndef MXNET_OPERATOR_AMSOFTMAX_INL_H_
-#define MXNET_OPERATOR_AMSOFTMAX_INL_H_
-
-#include <dmlc/logging.h>
-#include <dmlc/parameter.h>
-#include <mxnet/operator.h>
-#include <cmath>
-#include <map>
-#include <vector>
-#include <string>
-#include "./operator_common.h"
-
-namespace mxnet {
-namespace op {
-
-namespace amsoftmax_enum {
-enum AmSoftmaxOpInputs {kData, kWeight, kLabel};
-enum AmSoftmaxOpOutputs {kOut, kOOut};
-enum AmSoftmaxResource {kTempSpace};
-}
-
-struct AmSoftmaxParam : public dmlc::Parameter<AmSoftmaxParam> {
-  float margin;
-  float s;
-  int num_hidden;
-  int verbose;
-  float eps;
-  DMLC_DECLARE_PARAMETER(AmSoftmaxParam) {
-    DMLC_DECLARE_FIELD(margin).set_default(0.5).set_lower_bound(0.0)
-    .describe("AmSoftmax margin");
-    DMLC_DECLARE_FIELD(s).set_default(64.0).set_lower_bound(1.0)
-    .describe("s to X");
-    DMLC_DECLARE_FIELD(num_hidden).set_lower_bound(1)
-    .describe("Number of hidden nodes of the output");
-    DMLC_DECLARE_FIELD(verbose).set_default(0)
-    .describe("Log for beta change");
-    DMLC_DECLARE_FIELD(eps).set_default(1e-10f)
-    .describe("l2 eps");
-  }
-};
-
-template<typename xpu, typename DType>
-class AmSoftmaxOp : public Operator {
- public:
-  explicit AmSoftmaxOp(AmSoftmaxParam param) {
-    this->param_ = param;
-    count_ = 0;
-  }
-
-  virtual void Forward(const OpContext &ctx,
-                       const std::vector<TBlob> &in_data,
-                       const std::vector<OpReqType> &req,
-                       const std::vector<TBlob> &out_data,
-                       const std::vector<TBlob> &aux_args) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    CHECK_EQ(in_data.size(), 3);
-    CHECK_EQ(out_data.size(), 2);
-    CHECK_EQ(req.size(), 2);
-    CHECK_EQ(req[amsoftmax_enum::kOut], kWriteTo);
-    Stream<xpu> *stream = ctx.get_stream<xpu>();
-    const int n = in_data[amsoftmax_enum::kData].size(0); //batch size
-    const int m = in_data[amsoftmax_enum::kWeight].size(0);//num classes
-    Tensor<xpu, 2, DType> x = in_data[amsoftmax_enum::kData].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> w = in_data[amsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 1, DType> label = in_data[amsoftmax_enum::kLabel].get_with_shape<xpu, 1, DType>(Shape1(n), stream);
-    Tensor<xpu, 2, DType> out = out_data[amsoftmax_enum::kOut].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> oout = out_data[amsoftmax_enum::kOOut].get_with_shape<xpu, 2, DType>(Shape2(n,1), stream);
-    //Tensor<xpu, 2, DType> workspace = ctx.requested[amsoftmax_enum::kTempSpace].get_space_typed<xpu, 2, DType>(Shape2(n, 1), stream);
-#if defined(__CUDACC__)
-    CHECK_EQ(stream->blas_handle_ownership_, Stream<xpu>::OwnHandle)
-        << "Must init CuBLAS handle in stream";
-#endif
-    // original fully connected
-    out = dot(x, w.T());
-    if (ctx.is_train) {
-      const DType margin = static_cast<DType>(param_.margin);
-      const DType s = static_cast<DType>(param_.s);
-      AmSoftmaxForward(x, w, label, out, oout, margin, s);
-    }
-  }
-
-  //virtual void GradNorm(mshadow::Tensor<xpu, 2, DType> grad, mshadow::Stream<xpu>* s) {
-  //  using namespace mshadow;
-  //  using namespace mshadow::expr;
-  //  Tensor<cpu, 2, DType> grad_cpu(grad.shape_);
-  //  AllocSpace(&grad_cpu);
-  //  Copy(grad_cpu, grad, s);
-  //  DType grad_norm = param_.eps;
-  //  for(uint32_t i=0;i<grad_cpu.shape_[0];i++) {
-  //    for(uint32_t j=0;j<grad_cpu.shape_[1];j++) {
-  //      grad_norm += grad_cpu[i][j]*grad_cpu[i][j];
-  //    }
-  //  }
-  //  grad_norm = sqrt(grad_norm);
-  //  grad_cpu /= grad_norm;
-  //  Copy(grad, grad_cpu, s);
-  //  FreeSpace(&grad_cpu);
-  //}
-
-  virtual DType GradNorm(mshadow::Tensor<xpu, 2, DType> grad, mshadow::Stream<xpu>* s) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    Tensor<cpu, 2, DType> grad_cpu(grad.shape_);
-    AllocSpace(&grad_cpu);
-    Copy(grad_cpu, grad, s);
-    DType grad_norm = param_.eps;
-    for(uint32_t i=0;i<grad_cpu.shape_[0];i++) {
-      for(uint32_t j=0;j<grad_cpu.shape_[1];j++) {
-        grad_norm += grad_cpu[i][j]*grad_cpu[i][j];
-      }
-    }
-    grad_norm = sqrt(grad_norm);
-    //grad_cpu /= grad_norm;
-    //Copy(grad, grad_cpu, s);
-    FreeSpace(&grad_cpu);
-    return grad_norm;
-  }
-  virtual void Print(mshadow::Tensor<xpu, 2, DType> tensor, mshadow::Stream<xpu>* s) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    Tensor<cpu, 2, DType> tensor_cpu(tensor.shape_);
-    AllocSpace(&tensor_cpu);
-    Copy(tensor_cpu, tensor, s);
-    for(uint32_t i=0;i<tensor_cpu.shape_[0];i++) {
-      for(uint32_t j=0;j<tensor_cpu.shape_[1];j++) {
-        std::cout<<tensor_cpu[i][j]<<",";
-      }
-      std::cout<<std::endl;
-    }
-    FreeSpace(&tensor_cpu);
-  }
-
-  virtual void Backward(const OpContext &ctx,
-                        const std::vector<TBlob> &out_grad,
-                        const std::vector<TBlob> &in_data,
-                        const std::vector<TBlob> &out_data,
-                        const std::vector<OpReqType> &req,
-                        const std::vector<TBlob> &in_grad,
-                        const std::vector<TBlob> &aux_args) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    CHECK_EQ(out_grad.size(), 1);
-    CHECK_EQ(in_data.size(), 3);
-    CHECK_EQ(out_data.size(), 2);
-    CHECK_GE(in_grad.size(), 2);
-    CHECK_GE(req.size(), 2);
-    CHECK_EQ(req[amsoftmax_enum::kData], kWriteTo);
-    CHECK_EQ(req[amsoftmax_enum::kWeight], kWriteTo);
-    Stream<xpu> *stream = ctx.get_stream<xpu>();
-    const int n = in_data[amsoftmax_enum::kData].size(0);
-    const int m = in_data[amsoftmax_enum::kWeight].size(0);
-    Tensor<xpu, 2, DType> x = in_data[amsoftmax_enum::kData].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> w = in_data[amsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 1, DType> label = in_data[amsoftmax_enum::kLabel].get_with_shape<xpu, 1, DType>(Shape1(n), stream);
-    Tensor<xpu, 2, DType> out = out_data[amsoftmax_enum::kOut].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> oout = out_data[amsoftmax_enum::kOOut].get_with_shape<xpu, 2, DType>(Shape2(n,1), stream);
-    Tensor<xpu, 2, DType> o_grad = out_grad[amsoftmax_enum::kOut].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> x_grad = in_grad[amsoftmax_enum::kData].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> w_grad = in_grad[amsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(stream);
-    Tensor<xpu, 2, DType> workspace = ctx.requested[amsoftmax_enum::kTempSpace].get_space_typed<xpu, 2, DType>(Shape2(n, 1), stream);
-#if defined(__CUDACC__)
-    CHECK_EQ(stream->blas_handle_ownership_, Stream<xpu>::OwnHandle)
-        << "Must init CuBLAS handle in stream";
-#endif
-    // original fully connected
-    x_grad = dot(o_grad, w);
-    w_grad = dot(o_grad.T(), x);
-    // large margin fully connected
-    const DType margin = static_cast<DType>(param_.margin);
-    const DType s = static_cast<DType>(param_.s);
-    AmSoftmaxBackward(x, w, label, out, oout, o_grad, x_grad, w_grad, workspace, margin, s);
-    count_+=1;
-    if (param_.verbose) {
-      if(count_%param_.verbose==0) {
-        DType n = GradNorm(x_grad, stream);
-        LOG(INFO)<<"x_grad norm:"<<n;
-        n = GradNorm(w_grad, stream);
-        LOG(INFO)<<"w_grad norm:"<<n;
-        //Print(oout, stream);
-      }
-    }
-  }
-
-
- private:
-  AmSoftmaxParam param_;
-  uint32_t count_;
-};
-
-template<typename xpu>
-Operator *CreateOp(AmSoftmaxParam param, int dtype);
-
-#if DMLC_USE_CXX11
-class AmSoftmaxProp : public OperatorProperty {
- public:
-  void Init(const std::vector<std::pair<std::string, std::string> > &kwargs) override {
-    param_.Init(kwargs);
-  }
-
-  std::map<std::string, std::string> GetParams() const override {
-    return param_.__DICT__();
-  }
-
-  std::vector<std::string> ListArguments() const override {
-    return {"data", "weight", "label"};
-  }
-
-  std::vector<std::string> ListOutputs() const override {
-    return {"output", "ooutput"};
-  }
-
-  int NumOutputs() const override {
-    return 2;
-  }
-
-  int NumVisibleOutputs() const override {
-    return 1;
-  }
-
-  bool InferShape(std::vector<TShape> *in_shape,
-                  std::vector<TShape> *out_shape,
-                  std::vector<TShape> *aux_shape) const override {
-    using namespace mshadow;
-    CHECK_EQ(in_shape->size(), 3) << "Input:[data, label, weight]";
-    const TShape &dshape = in_shape->at(amsoftmax_enum::kData);
-    const TShape &lshape = in_shape->at(amsoftmax_enum::kLabel);
-    CHECK_EQ(dshape.ndim(), 2) << "data shape should be (batch_size, feature_dim)";
-    CHECK_EQ(lshape.ndim(), 1) << "label shape should be (batch_size,)";
-    const int n = dshape[0];
-    const int feature_dim = dshape[1];
-    const int m = param_.num_hidden;
-    SHAPE_ASSIGN_CHECK(*in_shape, amsoftmax_enum::kWeight, Shape2(m, feature_dim));
-    out_shape->clear();
-    out_shape->push_back(Shape2(n, m));  // output
-    out_shape->push_back(Shape2(n, 1));  // output
-    aux_shape->clear();
-    return true;
-  }
-
-  std::vector<ResourceRequest> BackwardResource(
-      const std::vector<TShape> &in_shape) const override {
-    return {ResourceRequest::kTempSpace};
-  }
-
-  std::vector<int> DeclareBackwardDependency(
-      const std::vector<int> &out_grad,
-      const std::vector<int> &in_data,
-      const std::vector<int> &out_data) const override {
-    return {out_grad[amsoftmax_enum::kOut], 
-            in_data[amsoftmax_enum::kData],
-            in_data[amsoftmax_enum::kWeight], in_data[amsoftmax_enum::kLabel]};
-  }
-
-  std::string TypeString() const override {
-    return "AmSoftmax";
-  }
-
-  OperatorProperty *Copy() const override {
-    auto ptr = new AmSoftmaxProp();
-    ptr->param_ = param_;
-    return ptr;
-  }
-
-  Operator *CreateOperator(Context ctx) const override {
-    LOG(FATAL) << "Not Implemented.";
-    return NULL;
-  }
-
-  Operator *CreateOperatorEx(Context ctx, std::vector<TShape> *in_shape,
-                             std::vector<int> *in_type) const override;
-
- private:
-  AmSoftmaxParam param_;
-};
-#endif  // DMLC_USE_CXX11
-
-}  // namespace op
-}  // namespace mxnet
-
-#endif 
diff --git a/3rdparty/operator/amsoftmax.cc b/3rdparty/operator/amsoftmax.cc
deleted file mode 100644
index 4075fd307..000000000
--- a/3rdparty/operator/amsoftmax.cc
+++ /dev/null
@@ -1,64 +0,0 @@
-#include "./amsoftmax-inl.h"
-
-namespace mshadow {
-
-template <typename DType>
-inline void AmSoftmaxForward(const Tensor<cpu, 2, DType> &x,
-                            const Tensor<cpu, 2, DType> &w,
-                            const Tensor<cpu, 1, DType> &label,
-                            const Tensor<cpu, 2, DType> &out,
-                            const Tensor<cpu, 2, DType> &oout,
-                            const DType margin,
-                            const DType s) {
-  LOG(FATAL) << "Not Implemented.";
-}
-
-template <typename DType>
-inline void AmSoftmaxBackward(const Tensor<cpu, 2, DType> &x,
-                             const Tensor<cpu, 2, DType> &w,
-                             const Tensor<cpu, 1, DType> &label,
-                             const Tensor<cpu, 2, DType> &out,
-                             const Tensor<cpu, 2, DType> &oout,
-                             const Tensor<cpu, 2, DType> &o_grad,
-                             const Tensor<cpu, 2, DType> &x_grad,
-                             const Tensor<cpu, 2, DType> &w_grad,
-                             const Tensor<cpu, 2, DType> &workspace,
-                             const DType margin,
-                             const DType s) {
-  LOG(FATAL) << "Not Implemented.";
-}
-
-}  // namespace mshadow
-
-namespace mxnet {
-namespace op {
-
-template<>
-Operator *CreateOp<cpu>(AmSoftmaxParam param, int dtype) {
-  Operator *op = NULL;
-  MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
-    op = new AmSoftmaxOp<cpu, DType>(param);
-  })
-  return op;
-}
-
-Operator *AmSoftmaxProp::CreateOperatorEx(Context ctx, std::vector<TShape> *in_shape,
-                                         std::vector<int> *in_type) const {
-  std::vector<TShape> out_shape, aux_shape;
-  std::vector<int> out_type, aux_type;
-  CHECK(InferType(in_type, &out_type, &aux_type));
-  CHECK(InferShape(in_shape, &out_shape, &aux_shape));
-  DO_BIND_DISPATCH(CreateOp, param_, in_type->at(0));
-}
-
-DMLC_REGISTER_PARAMETER(AmSoftmaxParam);
-
-MXNET_REGISTER_OP_PROPERTY(AmSoftmax, AmSoftmaxProp)
-.describe("AmSoftmax from <TODO>")
-.add_argument("data", "Symbol", "data")
-.add_argument("weight", "Symbol", "weight")
-.add_argument("label", "Symbol", "label")
-.add_arguments(AmSoftmaxParam::__FIELDS__());
-
-}  // namespace op
-}  // namespace mxnet
diff --git a/3rdparty/operator/amsoftmax.cu b/3rdparty/operator/amsoftmax.cu
deleted file mode 100644
index dfd7f1f69..000000000
--- a/3rdparty/operator/amsoftmax.cu
+++ /dev/null
@@ -1,195 +0,0 @@
-#include "./amsoftmax-inl.h"
-#include <math.h>
-
-namespace mshadow {
-namespace cuda {
-
-#define CUDA_KERNEL_LOOP(i, n) \
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; \
-       i < (n); \
-       i += blockDim.x * gridDim.x)
-
-
-template<typename DType>
-__global__ void AmSoftmaxForwardKernel(const Tensor<gpu, 2, DType> x,
-                                      const Tensor<gpu, 2, DType> w,
-                                      const Tensor<gpu, 1, DType> label,
-                                      Tensor<gpu, 2, DType> out,
-                                      Tensor<gpu, 2, DType> oout,
-                                      const DType margin,
-                                      const DType s) {
-  const int n = x.size(0); //batch size
-  const int feature_dim = x.size(1); //embedding size, 512 for example
-  const int m = w.size(0);//num classes
-  const DType cos_m = cos(margin);
-  const DType sin_m = sin(margin);
-  CUDA_KERNEL_LOOP(i, n) {
-    const int yi = static_cast<int>(label[i]);
-    const DType fo_i_yi = out[i][yi];
-    oout[i][0] = fo_i_yi;
-    if(fo_i_yi>=0.0) {
-      const DType cos_t = fo_i_yi / s;
-      const DType sin_t = sqrt(1.0-cos_t*cos_t);
-      out[i][yi] = fo_i_yi*cos_m - (s*sin_t*sin_m);
-    }
-  }
-}
-
-template<typename DType>
-inline void AmSoftmaxForward(const Tensor<gpu, 2, DType> &x,
-                            const Tensor<gpu, 2, DType> &w,
-                            const Tensor<gpu, 1, DType> &label,
-                            const Tensor<gpu, 2, DType> &out,
-                            const Tensor<gpu, 2, DType> &oout,
-                            const DType margin,
-                            const DType s) {
-  const int n = x.size(0);
-  const int m = w.size(0);
-  dim3 dimBlock(kBaseThreadNum);
-  dim3 dimGrid((n + kBaseThreadNum - 1) / kBaseThreadNum);
-  AmSoftmaxForwardKernel<<<dimGrid, dimBlock>>>(x, w, label, out, oout, margin, s);
-}
-
-
-template<typename DType>
-__global__ void AmSoftmaxBackwardXKernel(const Tensor<gpu, 2, DType> x,
-                                        const Tensor<gpu, 2, DType> w,
-                                        const Tensor<gpu, 1, DType> label,
-                                        const Tensor<gpu, 2, DType> out,
-                                        const Tensor<gpu, 2, DType> oout,
-                                        const Tensor<gpu, 2, DType> o_grad,
-                                        Tensor<gpu, 2, DType> x_grad,
-                                        const Tensor<gpu, 2, DType> workspace,
-                                        const DType margin,
-                                        const DType s) {
-  const int nthreads = x.size(0) * x.size(1);
-  //const int nthreads = x.size(0);
-  const int feature_dim = x.size(1);
-  const DType cos_m = cos(margin);
-  const DType nsin_m = sin(margin)*-1.0;
-  const DType ss = s*s;
-  CUDA_KERNEL_LOOP(idx, nthreads) {
-    const int i = idx / feature_dim;
-    const int l = idx % feature_dim;
-    //const int i = idx;
-    const int yi = static_cast<int>(label[i]);
-    if(oout[i][0]>=0.0) {
-      //x_grad[i][l] -= o_grad[i][yi] * w[yi][l];
-      //c = 1-cost*cost, = sint*sint
-      const DType cost = oout[i][0]/s;
-      const DType c = 1.0-cost*cost;
-      const DType dc_dx = -2.0/ss*oout[i][0]*w[yi][l];
-      const DType d_sint_dc = 1.0/(2*sqrt(c));
-      const DType d_sint_dx = dc_dx*d_sint_dc;
-      const DType df_dx = cos_m*w[yi][l] + s*nsin_m*d_sint_dx;
-      x_grad[i][l] += o_grad[i][yi] * (df_dx - w[yi][l]);
-    }
-  }
-}
-
-template<typename DType>
-__global__ void AmSoftmaxBackwardWKernel(const Tensor<gpu, 2, DType> x,
-                                        const Tensor<gpu, 2, DType> w,
-                                        const Tensor<gpu, 1, DType> label,
-                                        const Tensor<gpu, 2, DType> out,
-                                        const Tensor<gpu, 2, DType> oout,
-                                        const Tensor<gpu, 2, DType> o_grad,
-                                        Tensor<gpu, 2, DType> w_grad,
-                                        const Tensor<gpu, 2, DType> workspace,
-                                        const DType margin,
-                                        const DType s) {
-  const int nthreads = w.size(0) * w.size(1);
-  const int n = x.size(0);
-  const int feature_dim = w.size(1);
-  const DType cos_m = cos(margin);
-  const DType nsin_m = sin(margin)*-1.0;
-  const DType ss = s*s;
-  CUDA_KERNEL_LOOP(idx, nthreads) {
-    const int j = idx / feature_dim;
-    const int l = idx % feature_dim;
-    DType dw = 0;
-    for (int i = 0; i < n; ++i) {
-      const int yi = static_cast<int>(label[i]);
-      if (yi == j&&oout[i][0]>=0.0) {
-        const DType cost = oout[i][0]/s;
-        const DType c = 1.0-cost*cost;
-        const DType dc_dw = -2.0/ss*oout[i][0]*x[i][l];
-        const DType d_sint_dc = 1.0/(2*sqrt(c));
-        const DType d_sint_dw = dc_dw*d_sint_dc;
-        const DType df_dw = cos_m*x[i][l] + s*nsin_m*d_sint_dw;
-        dw += o_grad[i][yi] * (df_dw - x[i][l]);
-      }
-    }
-    w_grad[j][l] += dw;
-  }
-}
-
-template<typename DType>
-inline void AmSoftmaxBackward(const Tensor<gpu, 2, DType> &x,
-                             const Tensor<gpu, 2, DType> &w,
-                             const Tensor<gpu, 1, DType> &label,
-                             const Tensor<gpu, 2, DType> &out,
-                             const Tensor<gpu, 2, DType> &oout,
-                             const Tensor<gpu, 2, DType> &o_grad,
-                             const Tensor<gpu, 2, DType> &x_grad,
-                             const Tensor<gpu, 2, DType> &w_grad,
-                             const Tensor<gpu, 2, DType> &workspace,
-                             const DType margin,
-                             const DType s) {
-  const int n = x.size(0);
-  const int feature_dim = x.size(1);
-  const int m = w.size(0);
-  dim3 dimBlock(kBaseThreadNum);
-  dim3 dimGrid((n + kBaseThreadNum - 1) / kBaseThreadNum);
-  dimGrid.x = ((n * feature_dim + kBaseThreadNum - 1) / kBaseThreadNum);
-  AmSoftmaxBackwardXKernel<<<dimGrid, dimBlock>>>(x, w, label, out, oout, o_grad, x_grad, workspace,
-                                                 margin, s);
-  dimGrid.x = ((m * feature_dim + kBaseThreadNum - 1) / kBaseThreadNum);
-  AmSoftmaxBackwardWKernel<<<dimGrid, dimBlock>>>(x, w, label, out, oout, o_grad, w_grad, workspace,
-                                                 margin, s);
-}
-
-}  // namespace cuda
-
-template<typename DType>
-inline void AmSoftmaxForward(const Tensor<gpu, 2, DType> &x,
-                            const Tensor<gpu, 2, DType> &w,
-                            const Tensor<gpu, 1, DType> &label,
-                            const Tensor<gpu, 2, DType> &out,
-                            const Tensor<gpu, 2, DType> &oout,
-                            const DType margin,
-                            const DType s) {
-  cuda::AmSoftmaxForward(x, w, label, out, oout, margin, s);
-}
-
-template<typename DType>
-inline void AmSoftmaxBackward(const Tensor<gpu, 2, DType> &x,
-                             const Tensor<gpu, 2, DType> &w,
-                             const Tensor<gpu, 1, DType> &label,
-                             const Tensor<gpu, 2, DType> &out,
-                             const Tensor<gpu, 2, DType> &oout,
-                             const Tensor<gpu, 2, DType> &o_grad,
-                             const Tensor<gpu, 2, DType> &x_grad,
-                             const Tensor<gpu, 2, DType> &w_grad,
-                             const Tensor<gpu, 2, DType> &workspace,
-                             const DType margin,
-                             const DType s) {
-  cuda::AmSoftmaxBackward(x, w, label, out, oout, o_grad, x_grad, w_grad, workspace, margin, s);
-}
-
-}  // namespace mshadow
-
-namespace mxnet {
-namespace op {
-
-template<>
-Operator *CreateOp<gpu>(AmSoftmaxParam param, int dtype) {
-  Operator *op = NULL;
-  MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
-    op = new AmSoftmaxOp<gpu, DType>(param);
-  })
-  return op;
-}
-
-}  // namespace op
-}  // namespace mxnet
diff --git a/3rdparty/operator/lsoftmax-inl.h b/3rdparty/operator/lsoftmax-inl.h
deleted file mode 100644
index a457eb4f4..000000000
--- a/3rdparty/operator/lsoftmax-inl.h
+++ /dev/null
@@ -1,377 +0,0 @@
-/*!
- * Copyright (c) 2016 by Contributors
- * \file lsoftmax-inl.h
- * \brief LSoftmax from <Large-Margin Softmax Loss for Convolutional Neural Networks>
- * \author luoyetx
- */
-#ifndef MXNET_OPERATOR_LSOFTMAX_INL_H_
-#define MXNET_OPERATOR_LSOFTMAX_INL_H_
-
-#include <dmlc/logging.h>
-#include <dmlc/parameter.h>
-#include <mxnet/operator.h>
-#include <cmath>
-#include <map>
-#include <vector>
-#include <string>
-#include "./operator_common.h"
-
-namespace mxnet {
-namespace op {
-
-namespace lsoftmax_enum {
-enum LSoftmaxOpInputs {kData, kWeight, kLabel};
-enum LSoftmaxOpOutputs {kOut, kDataNorm, kWeightNorm};
-enum LSoftmaxResource {kTempSpace};
-}
-
-struct LSoftmaxParam : public dmlc::Parameter<LSoftmaxParam> {
-  int margin;
-  float beta;
-  float beta_min;
-  float scale;
-  int num_hidden;
-  bool grad_norm;
-  int verbose;
-  float eps;
-  DMLC_DECLARE_PARAMETER(LSoftmaxParam) {
-    DMLC_DECLARE_FIELD(margin).set_default(2).set_lower_bound(1)
-    .describe("LSoftmax margin");
-    DMLC_DECLARE_FIELD(beta).set_default(1).set_lower_bound(0)
-    .describe("LSoftmax beta, same as lambda to weight original value");
-    DMLC_DECLARE_FIELD(beta_min).set_default(0).set_lower_bound(0)
-    .describe("Minimum beta");
-    DMLC_DECLARE_FIELD(scale).set_default(1).set_range(0, 1)
-    .describe("Scale of beta during training for every iteration");
-    DMLC_DECLARE_FIELD(num_hidden).set_lower_bound(1)
-    .describe("Number of hidden nodes of the output");
-    DMLC_DECLARE_FIELD(grad_norm).set_default(false)
-    .describe("do grad norm");
-    DMLC_DECLARE_FIELD(verbose).set_default(0)
-    .describe("Log for beta change");
-    DMLC_DECLARE_FIELD(eps).set_default(1e-10f)
-    .describe("l2 eps");
-  }
-};
-
-template<typename xpu, typename DType>
-class LSoftmaxOp : public Operator {
- public:
-  explicit LSoftmaxOp(LSoftmaxParam param) {
-    this->param_ = param;
-    // setup global lookup table
-    k_table_.clear();
-    c_table_.clear();
-    k_table_.push_back(1);
-    c_table_.push_back(1);
-    const int margin = param.margin;
-    const double pi = std::atan(1) * 4;
-    double factor = 1;
-    for (int i = 1; i <= margin; ++i) {
-      factor = factor * (margin - i + 1) / i;
-      k_table_.push_back(std::cos(i * pi / margin));
-      c_table_.push_back(factor);
-    }
-    //next_beta_ = param.beta * 0.1f;
-    count_ = 0;
-    if(const char* env_p = std::getenv("BETA")) {
-      float _beta = std::atof(env_p);
-      if (param_.verbose) {
-        LOG(INFO)<<"beta:"<<_beta;
-        LOG(INFO)<<param_.margin<<","<<param_.beta<<","<<param_.beta_min<<","<<param_.scale;
-      }
-      param_.beta = _beta;
-    }
-    else if(const char* env_p = std::getenv("GLOBAL_STEP")) {
-      int nbatch = std::atoi(env_p);
-      if (param_.verbose) {
-        LOG(INFO)<<"nbatch:"<<nbatch;
-      }
-      float _beta = param.beta*std::pow((double)param.scale, (double)nbatch);
-      param_.beta = std::max(_beta, param_.beta_min);
-    }
-  }
-
-  virtual void Forward(const OpContext &ctx,
-                       const std::vector<TBlob> &in_data,
-                       const std::vector<OpReqType> &req,
-                       const std::vector<TBlob> &out_data,
-                       const std::vector<TBlob> &aux_args) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    CHECK_EQ(in_data.size(), 3);
-    CHECK_EQ(out_data.size(), 3);
-    CHECK_EQ(req.size(), 3);
-    CHECK_EQ(req[lsoftmax_enum::kOut], kWriteTo);
-    CHECK(req[lsoftmax_enum::kDataNorm] == kNullOp ||
-          req[lsoftmax_enum::kDataNorm] == kWriteTo);
-    CHECK(req[lsoftmax_enum::kWeightNorm] == kNullOp ||
-          req[lsoftmax_enum::kWeightNorm] == kWriteTo);
-    Stream<xpu> *s = ctx.get_stream<xpu>();
-    const int n = in_data[lsoftmax_enum::kData].size(0);
-    const int m = in_data[lsoftmax_enum::kWeight].size(0);
-    Tensor<xpu, 2, DType> x = in_data[lsoftmax_enum::kData].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 2, DType> w = in_data[lsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 1, DType> label = in_data[lsoftmax_enum::kLabel].get_with_shape<xpu, 1, DType>(Shape1(n), s);
-    Tensor<xpu, 2, DType> out = out_data[lsoftmax_enum::kOut].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 1, DType> x_norm = out_data[lsoftmax_enum::kDataNorm].get_with_shape<xpu, 1, DType>(Shape1(n), s);
-    Tensor<xpu, 1, DType> w_norm = out_data[lsoftmax_enum::kWeightNorm].get_with_shape<xpu, 1, DType>(Shape1(m), s);
-#if defined(__CUDACC__)
-    CHECK_EQ(s->blas_handle_ownership_, Stream<xpu>::OwnHandle)
-        << "Must init CuBLAS handle in stream";
-#endif
-    // original fully connected
-    out = dot(x, w.T());
-    if (ctx.is_train) {
-      // large margin fully connected
-      const int margin = param_.margin;
-      if(const char* env_p = std::getenv("BETA")) {
-        float _beta = std::atof(env_p);
-        param_.beta = _beta;
-      }
-      const DType beta = static_cast<DType>(param_.beta);
-      //LOG(INFO)<<"beta:"<<beta<<std::endl;
-      Tensor<cpu, 1, DType> k_table_cpu(k_table_.data(), Shape1(k_table_.size()));
-      Tensor<cpu, 1, DType> c_table_cpu(c_table_.data(), Shape1(c_table_.size()));
-      Tensor<xpu, 1, DType> k_table_xpu(Shape1(k_table_.size()));
-      Tensor<xpu, 1, DType> c_table_xpu(Shape1(c_table_.size()));
-      k_table_xpu.set_stream(s);
-      c_table_xpu.set_stream(s);
-      AllocSpace(&k_table_xpu);
-      AllocSpace(&c_table_xpu);
-      Copy(k_table_xpu, k_table_cpu, s);
-      Copy(c_table_xpu, c_table_cpu, s);
-      LSoftmaxForward(x, w, label, out, x_norm, w_norm, k_table_xpu, c_table_xpu, margin, beta);
-      FreeSpace(&k_table_xpu);
-      FreeSpace(&c_table_xpu);
-    }
-  }
-
-  //virtual void GradNorm(mshadow::Tensor<xpu, 2, DType> grad, mshadow::Stream<xpu>* s) {
-  //  using namespace mshadow;
-  //  using namespace mshadow::expr;
-  //  Tensor<cpu, 2, DType> grad_cpu(grad.shape_);
-  //  AllocSpace(&grad_cpu);
-  //  Copy(grad_cpu, grad, s);
-  //  DType grad_norm = param_.eps;
-  //  for(uint32_t i=0;i<grad_cpu.shape_[0];i++) {
-  //    for(uint32_t j=0;j<grad_cpu.shape_[1];j++) {
-  //      grad_norm += grad_cpu[i][j]*grad_cpu[i][j];
-  //    }
-  //  }
-  //  grad_norm = sqrt(grad_norm);
-  //  grad_cpu /= grad_norm;
-  //  Copy(grad, grad_cpu, s);
-  //  FreeSpace(&grad_cpu);
-  //}
-
-  virtual DType GradNorm(mshadow::Tensor<xpu, 2, DType> grad, mshadow::Stream<xpu>* s) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    Tensor<cpu, 2, DType> grad_cpu(grad.shape_);
-    AllocSpace(&grad_cpu);
-    Copy(grad_cpu, grad, s);
-    DType grad_norm = param_.eps;
-    for(uint32_t i=0;i<grad_cpu.shape_[0];i++) {
-      for(uint32_t j=0;j<grad_cpu.shape_[1];j++) {
-        grad_norm += grad_cpu[i][j]*grad_cpu[i][j];
-      }
-    }
-    grad_norm = sqrt(grad_norm);
-    //grad_cpu /= grad_norm;
-    //Copy(grad, grad_cpu, s);
-    FreeSpace(&grad_cpu);
-    return grad_norm;
-  }
-
-  virtual void Backward(const OpContext &ctx,
-                        const std::vector<TBlob> &out_grad,
-                        const std::vector<TBlob> &in_data,
-                        const std::vector<TBlob> &out_data,
-                        const std::vector<OpReqType> &req,
-                        const std::vector<TBlob> &in_grad,
-                        const std::vector<TBlob> &aux_args) {
-    using namespace mshadow;
-    using namespace mshadow::expr;
-    CHECK_EQ(out_grad.size(), 1);
-    CHECK_EQ(in_data.size(), 3);
-    CHECK_EQ(out_data.size(), 3);
-    CHECK_GE(in_grad.size(), 2);
-    CHECK_GE(req.size(), 2);
-    CHECK_EQ(req[lsoftmax_enum::kData], kWriteTo);
-    CHECK_EQ(req[lsoftmax_enum::kWeight], kWriteTo);
-    Stream<xpu> *s = ctx.get_stream<xpu>();
-    const int n = in_data[lsoftmax_enum::kData].size(0);
-    const int m = in_data[lsoftmax_enum::kWeight].size(0);
-    Tensor<xpu, 2, DType> x = in_data[lsoftmax_enum::kData].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 2, DType> w = in_data[lsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 1, DType> label = in_data[lsoftmax_enum::kLabel].get_with_shape<xpu, 1, DType>(Shape1(n), s);
-    Tensor<xpu, 1, DType> x_norm = out_data[lsoftmax_enum::kDataNorm].get_with_shape<xpu, 1, DType>(Shape1(n), s);
-    Tensor<xpu, 1, DType> w_norm = out_data[lsoftmax_enum::kWeightNorm].get_with_shape<xpu, 1, DType>(Shape1(m), s);
-    Tensor<xpu, 2, DType> o_grad = out_grad[lsoftmax_enum::kOut].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 2, DType> x_grad = in_grad[lsoftmax_enum::kData].FlatTo2D<xpu, DType>(s);
-    Tensor<xpu, 2, DType> w_grad = in_grad[lsoftmax_enum::kWeight].FlatTo2D<xpu, DType>(s);
-    // workspace is used for cos_t, cos_mt, k, sin2_t, fo and cos_t_m for every data point
-    Tensor<xpu, 2, DType> workspace = ctx.requested[lsoftmax_enum::kTempSpace].get_space_typed<xpu, 2, DType>(Shape2(6, n), s);
-#if defined(__CUDACC__)
-    CHECK_EQ(s->blas_handle_ownership_, Stream<xpu>::OwnHandle)
-        << "Must init CuBLAS handle in stream";
-#endif
-    // original fully connected
-    x_grad = dot(o_grad, w);
-    w_grad = dot(o_grad.T(), x);
-    // large margin fully connected
-    const int margin = param_.margin;
-    const DType beta = static_cast<DType>(param_.beta);
-    count_+=1;
-    if (param_.verbose) {
-      if(count_%param_.verbose==0) {
-        LOG(INFO)<<"["<<count_<<"]current beta:"<<beta;
-        DType n = GradNorm(x_grad, s);
-        LOG(INFO)<<"x_grad norm:"<<n;
-        n = GradNorm(w_grad, s);
-        LOG(INFO)<<"w_grad norm:"<<n;
-      }
-    }
-    Tensor<cpu, 1, DType> k_table_cpu(k_table_.data(), Shape1(k_table_.size()));
-    Tensor<cpu, 1, DType> c_table_cpu(c_table_.data(), Shape1(c_table_.size()));
-    Tensor<xpu, 1, DType> k_table_xpu(Shape1(k_table_.size()));
-    Tensor<xpu, 1, DType> c_table_xpu(Shape1(c_table_.size()));
-    k_table_xpu.set_stream(s);
-    c_table_xpu.set_stream(s);
-    AllocSpace(&k_table_xpu);
-    AllocSpace(&c_table_xpu);
-    Copy(k_table_xpu, k_table_cpu, s);
-    Copy(c_table_xpu, c_table_cpu, s);
-    LSoftmaxBackward(x, w, label, x_norm, w_norm, o_grad, x_grad, w_grad, workspace,
-                     k_table_xpu, c_table_xpu, margin, beta);
-    FreeSpace(&k_table_xpu);
-    FreeSpace(&c_table_xpu);
-    //if(param_.grad_norm) {
-    //  GradNorm(x_grad, s);
-    //  GradNorm(w_grad, s);
-    //}
-    // dirty hack, should also work for multi device
-    if(std::getenv("BETA")==NULL) {
-      param_.beta *= param_.scale;
-      param_.beta = std::max(param_.beta, param_.beta_min);
-    }
-    //LOG(INFO)<<"w_grad:"<<w_grad.shape_[0]<<","<<w_grad.shape_[1];
-    //norm = reduce_with_axis<red::sum, false>(F<mxnet::op::mshadow_op::square>(w_grad), 2);
-    //norm = F<mxnet::op::mshadow_op::square_root>(norm + param_.eps);
-    //out = data / broadcast_with_axis(norm, 1, dshape[2]);
-    //if (param_.beta < next_beta_) {
-    //  next_beta_ *= 0.1f;
-    //  if (param_.verbose) {
-    //    LOG(INFO) << "LSoftmax changes beta to " << param_.beta;
-    //  }
-    //}
-  }
-
-  //Tensor<xpu, 2, DType> grad_norm(const Tensor<xpu, 2, DType> grad) {
-  //}
-
-
-
-
- private:
-  LSoftmaxParam param_;
-  // global lookup table
-  std::vector<DType> k_table_;
-  std::vector<DType> c_table_;
-  //float next_beta_;
-  uint32_t count_;
-};  // class LSoftmaxOp
-
-template<typename xpu>
-Operator *CreateOp(LSoftmaxParam param, int dtype);
-
-#if DMLC_USE_CXX11
-class LSoftmaxProp : public OperatorProperty {
- public:
-  void Init(const std::vector<std::pair<std::string, std::string> > &kwargs) override {
-    param_.Init(kwargs);
-  }
-
-  std::map<std::string, std::string> GetParams() const override {
-    return param_.__DICT__();
-  }
-
-  std::vector<std::string> ListArguments() const override {
-    return {"data", "weight", "label"};
-  }
-
-  std::vector<std::string> ListOutputs() const override {
-    return {"output", "data_norm", "weight_norm"};
-  }
-
-  int NumOutputs() const override {
-    return 3;
-  }
-
-  int NumVisibleOutputs() const override {
-    return 1;
-  }
-
-  bool InferShape(std::vector<TShape> *in_shape,
-                  std::vector<TShape> *out_shape,
-                  std::vector<TShape> *aux_shape) const override {
-    using namespace mshadow;
-    CHECK_EQ(in_shape->size(), 3) << "Input:[data, label, weight]";
-    const TShape &dshape = in_shape->at(lsoftmax_enum::kData);
-    const TShape &lshape = in_shape->at(lsoftmax_enum::kLabel);
-    CHECK_EQ(dshape.ndim(), 2) << "data shape should be (batch_size, feature_dim)";
-    CHECK_EQ(lshape.ndim(), 1) << "label shape should be (batch_size,)";
-    const int n = dshape[0];
-    const int feature_dim = dshape[1];
-    const int m = param_.num_hidden;
-    SHAPE_ASSIGN_CHECK(*in_shape, lsoftmax_enum::kWeight, Shape2(m, feature_dim));
-    out_shape->clear();
-    out_shape->push_back(Shape2(n, m));  // output
-    out_shape->push_back(Shape1(n));  // data norm
-    out_shape->push_back(Shape1(m));  // weight norm
-    aux_shape->clear();
-    return true;
-  }
-
-  std::vector<ResourceRequest> BackwardResource(
-      const std::vector<TShape> &in_shape) const override {
-    return {ResourceRequest::kTempSpace};
-  }
-
-  std::vector<int> DeclareBackwardDependency(
-      const std::vector<int> &out_grad,
-      const std::vector<int> &in_data,
-      const std::vector<int> &out_data) const override {
-    return {out_grad[lsoftmax_enum::kOut], out_data[lsoftmax_enum::kDataNorm],
-            out_data[lsoftmax_enum::kWeightNorm], in_data[lsoftmax_enum::kData],
-            in_data[lsoftmax_enum::kWeight], in_data[lsoftmax_enum::kLabel]};
-  }
-
-  std::string TypeString() const override {
-    return "LSoftmax";
-  }
-
-  OperatorProperty *Copy() const override {
-    auto ptr = new LSoftmaxProp();
-    ptr->param_ = param_;
-    return ptr;
-  }
-
-  Operator *CreateOperator(Context ctx) const override {
-    LOG(FATAL) << "Not Implemented.";
-    return NULL;
-  }
-
-  Operator *CreateOperatorEx(Context ctx, std::vector<TShape> *in_shape,
-                             std::vector<int> *in_type) const override;
-
- private:
-  LSoftmaxParam param_;
-};  // class LSoftmaxProp
-#endif  // DMLC_USE_CXX11
-
-}  // namespace op
-}  // namespace mxnet
-
-#endif  // MXNET_OPERATOR_LSOFTMAX_INL_H_
diff --git a/3rdparty/operator/lsoftmax.cc b/3rdparty/operator/lsoftmax.cc
deleted file mode 100644
index cbf708bed..000000000
--- a/3rdparty/operator/lsoftmax.cc
+++ /dev/null
@@ -1,75 +0,0 @@
-/*!
- * Copyright (c) 2016 by Contributors
- * \file lsoftmax.cc
- * \brief LSoftmax from <Large-Margin Softmax Loss for Convolutional Neural Networks>
- * \author luoyetx
- */
-#include "./lsoftmax-inl.h"
-
-namespace mshadow {
-
-template <typename DType>
-inline void LSoftmaxForward(const Tensor<cpu, 2, DType> &x,
-                            const Tensor<cpu, 2, DType> &w,
-                            const Tensor<cpu, 1, DType> &label,
-                            const Tensor<cpu, 2, DType> &out,
-                            const Tensor<cpu, 1, DType> &x_norm,
-                            const Tensor<cpu, 1, DType> &w_norm,
-                            const Tensor<cpu, 1, DType> &k_table,
-                            const Tensor<cpu, 1, DType> &c_table,
-                            const int margin,
-                            const DType beta) {
-  LOG(FATAL) << "Not Implemented.";
-}
-
-template <typename DType>
-inline void LSoftmaxBackward(const Tensor<cpu, 2, DType> &x,
-                             const Tensor<cpu, 2, DType> &w,
-                             const Tensor<cpu, 1, DType> &label,
-                             const Tensor<cpu, 1, DType> &x_norm,
-                             const Tensor<cpu, 1, DType> &w_norm,
-                             const Tensor<cpu, 2, DType> &o_grad,
-                             const Tensor<cpu, 2, DType> &x_grad,
-                             const Tensor<cpu, 2, DType> &w_grad,
-                             const Tensor<cpu, 2, DType> &workspace,
-                             const Tensor<cpu, 1, DType> &k_table,
-                             const Tensor<cpu, 1, DType> &c_table,
-                             const int margin,
-                             const DType beta) {
-  LOG(FATAL) << "Not Implemented.";
-}
-
-}  // namespace mshadow
-
-namespace mxnet {
-namespace op {
-
-template<>
-Operator *CreateOp<cpu>(LSoftmaxParam param, int dtype) {
-  Operator *op = NULL;
-  MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
-    op = new LSoftmaxOp<cpu, DType>(param);
-  })
-  return op;
-}
-
-Operator *LSoftmaxProp::CreateOperatorEx(Context ctx, std::vector<TShape> *in_shape,
-                                         std::vector<int> *in_type) const {
-  std::vector<TShape> out_shape, aux_shape;
-  std::vector<int> out_type, aux_type;
-  CHECK(InferType(in_type, &out_type, &aux_type));
-  CHECK(InferShape(in_shape, &out_shape, &aux_shape));
-  DO_BIND_DISPATCH(CreateOp, param_, in_type->at(0));
-}
-
-DMLC_REGISTER_PARAMETER(LSoftmaxParam);
-
-MXNET_REGISTER_OP_PROPERTY(LSoftmax, LSoftmaxProp)
-.describe("LSoftmax from <Large-Margin Softmax Loss for Convolutional Neural Networks>")
-.add_argument("data", "Symbol", "data")
-.add_argument("weight", "Symbol", "weight")
-.add_argument("label", "Symbol", "label")
-.add_arguments(LSoftmaxParam::__FIELDS__());
-
-}  // namespace op
-}  // namespace mxnet
diff --git a/3rdparty/operator/lsoftmax.cu b/3rdparty/operator/lsoftmax.cu
deleted file mode 100644
index 6055c1bea..000000000
--- a/3rdparty/operator/lsoftmax.cu
+++ /dev/null
@@ -1,322 +0,0 @@
-/*!
- * Copyright (c) 2016 by Contributors
- * \file lsoftmax.cu
- * \brief LSoftmax from <Large-Margin Softmax Loss for Convolutional Neural Networks>
- * \author luoyetx
- */
-#include "./lsoftmax-inl.h"
-
-namespace mshadow {
-namespace cuda {
-
-namespace {
-// workspace variables
-enum LSoftmaxTempSpaceType {kCost, kCosmt, kK, kSin2t, kFo, kCostM};
-}
-
-#define CUDA_KERNEL_LOOP(i, n) \
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; \
-       i < (n); \
-       i += blockDim.x * gridDim.x)
-
-MSHADOW_XINLINE int LSPowOfMO(const int k) {
-  return 1 - ((k&0x01) << 1);
-}
-
-template<typename DType>
-__global__ void LSCalcNorm(const Tensor<gpu, 2, DType> x,
-                           Tensor<gpu, 1, DType> x_norm) {
-  const int n = x.size(0);
-  const int m = x.size(1);
-  CUDA_KERNEL_LOOP(i, n) {
-    DType norm = 0;
-    for (int j = 0; j < m; ++j) {
-      norm += x[i][j] * x[i][j];
-    }
-    x_norm[i] = sqrt(norm);
-  }
-}
-
-template<typename DType>
-__device__ int LSFindK(const DType *k_table, const int n, const DType cos_t) {
-  const DType eps = 1e-5;
-  for (int i = 0; i < n; ++i) {
-    if (((k_table[i+1] < cos_t) || (abs(k_table[i+1] - cos_t) < eps)) &&
-        ((k_table[i] > cos_t) || (abs(k_table[i] - cos_t) < eps))) {
-      return i;
-    }
-  }
-  return 0;
-}
-
-template<typename DType>
-__device__ DType LSCalcCosmt(const DType *c_table, const int n,
-                             const DType cos_t, const int margin) {
-  const DType sin2_t = 1 - cos_t * cos_t;
-  DType cos_t_p = pow(cos_t, margin);
-  DType sin2_t_p = 1;
-  DType cos_mt = cos_t_p;  // p = 0
-  for (int p = 1; p <= margin / 2; ++p) {
-    cos_t_p /= cos_t * cos_t;  // don't replace `cos_t*cos_t` with `1-sin2_t`, this can cause numeric issue if cos_t --> 0
-    sin2_t_p *= sin2_t;
-    cos_mt += LSPowOfMO(p) * c_table[2*p] * cos_t_p * sin2_t_p;
-  }
-  return cos_mt;
-}
-
-template<typename DType>
-__global__ void LSoftmaxForwardKernel(const Tensor<gpu, 2, DType> x,
-                                      const Tensor<gpu, 2, DType> w,
-                                      const Tensor<gpu, 1, DType> label,
-                                      const Tensor<gpu, 1, DType> x_norm,
-                                      const Tensor<gpu, 1, DType> w_norm,
-                                      Tensor<gpu, 2, DType> out,
-                                      const Tensor<gpu, 1, DType> k_table,
-                                      const Tensor<gpu, 1, DType> c_table,
-                                      const int margin,
-                                      const DType beta) {
-  const int n = x.size(0);
-  const int feature_dim = x.size(1);
-  const int m = w.size(0);
-  CUDA_KERNEL_LOOP(i, n) {
-    const int yi = static_cast<int>(label[i]);
-    const DType fo_i_yi = out[i][yi];
-    const DType cos_t = fo_i_yi / (x_norm[i] * w_norm[yi]);
-    const int k = LSFindK(k_table.dptr_, k_table.size(0), cos_t);
-    const DType cos_mt = LSCalcCosmt(c_table.dptr_, c_table.size(0), cos_t, margin);
-    const DType f_i_yi = (LSPowOfMO(k) * cos_mt - 2*k) * (w_norm[yi] * x_norm[i]);
-    out[i][yi] = (f_i_yi + beta * fo_i_yi) / (1 + beta);
-  }
-}
-
-template<typename DType>
-inline void LSoftmaxForward(const Tensor<gpu, 2, DType> &x,
-                            const Tensor<gpu, 2, DType> &w,
-                            const Tensor<gpu, 1, DType> &label,
-                            const Tensor<gpu, 2, DType> &out,
-                            const Tensor<gpu, 1, DType> &x_norm,
-                            const Tensor<gpu, 1, DType> &w_norm,
-                            const Tensor<gpu, 1, DType> &k_table,
-                            const Tensor<gpu, 1, DType> &c_table,
-                            const int margin,
-                            const DType beta) {
-  const int n = x.size(0);
-  const int m = w.size(0);
-  dim3 dimBlock(kBaseThreadNum);
-  dim3 dimGrid((n + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSCalcNorm<<<dimGrid, dimBlock>>>(x, x_norm);
-  dimGrid.x = ((m + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSCalcNorm<<<dimGrid, dimBlock>>>(w, w_norm);
-  dimGrid.x = ((n + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSoftmaxForwardKernel<<<dimGrid, dimBlock>>>(x, w, label, x_norm, w_norm, out, k_table, c_table, margin, beta);
-}
-
-template<typename DType>
-__global__ void LSoftmaxBackwardRequired(const Tensor<gpu, 2, DType> x,
-                                         const Tensor<gpu, 2, DType> w,
-                                         const Tensor<gpu, 1, DType> label,
-                                         const Tensor<gpu, 1, DType> x_norm,
-                                         const Tensor<gpu, 1, DType> w_norm,
-                                         Tensor<gpu, 2, DType> workspace,
-                                         const Tensor<gpu, 1, DType> k_table,
-                                         const Tensor<gpu, 1, DType> c_table,
-                                         const int margin) {
-  const int n = x.size(0);
-  const int feature_dim = x.size(1);
-  CUDA_KERNEL_LOOP(i, n) {
-    const int yi = static_cast<int>(label[i]);
-    // fo_i_yi = dot(w_yi, x_i)
-    DType fo_i_yi = 0;
-    for (int p = 0; p < feature_dim; ++p) {
-      fo_i_yi += w[yi][p] * x[i][p];
-    }
-    const DType cos_t = fo_i_yi / (x_norm[i] * w_norm[yi]);
-    const int k = LSFindK(k_table.dptr_, k_table.size(0), cos_t);
-    const DType cos_mt = LSCalcCosmt(c_table.dptr_, c_table.size(0), cos_t, margin);
-    const DType sin2_t = 1 - cos_t * cos_t;
-    workspace[kCost][i] = cos_t;
-    workspace[kCosmt][i] = cos_mt;
-    workspace[kK][i] = static_cast<DType>(k);
-    workspace[kSin2t][i] = sin2_t;
-    workspace[kFo][i] = fo_i_yi;
-    workspace[kCostM][i] = pow(cos_t, margin - 1);
-  }
-}
-
-template<typename DType>
-__global__ void LSoftmaxBackwardXKernel(const Tensor<gpu, 2, DType> x,
-                                        const Tensor<gpu, 2, DType> w,
-                                        const Tensor<gpu, 1, DType> label,
-                                        const Tensor<gpu, 1, DType> x_norm,
-                                        const Tensor<gpu, 1, DType> w_norm,
-                                        const Tensor<gpu, 2, DType> o_grad,
-                                        Tensor<gpu, 2, DType> x_grad,
-                                        const Tensor<gpu, 2, DType> workspace,
-                                        const Tensor<gpu, 1, DType> c_table,
-                                        const int margin,
-                                        const DType beta) {
-  const int nthreads = x.size(0) * x.size(1);
-  const int feature_dim = x.size(1);
-  CUDA_KERNEL_LOOP(idx, nthreads) {
-    const int i = idx / feature_dim;
-    const int l = idx % feature_dim;
-    const int yi = static_cast<int>(label[i]);
-    const DType cos_t = workspace[kCost][i];
-    const DType cos_mt = workspace[kCosmt][i];
-    const int k = static_cast<int>(workspace[kK][i]);
-    const DType sin2_t = workspace[kSin2t][i];
-    const DType fo_i_yi = workspace[kFo][i];
-    const DType w_norm_yi = w_norm[yi];
-    const DType x_norm_i = x_norm[i];
-
-    const DType dcos_dx = w[yi][l] / (w_norm_yi * x_norm_i) - \
-                          fo_i_yi * x[i][l] / (w_norm_yi * x_norm_i * x_norm_i * x_norm_i);
-    const DType dsin2_dx = -2 * cos_t * dcos_dx;
-    DType cos_t_p = workspace[kCostM][i];
-    DType sin2_t_p = 1;
-    DType dcosm_dx = margin * cos_t_p * dcos_dx;  // p = 0
-    for (int p = 1; p <= margin / 2; ++p) {
-      cos_t_p /= cos_t * cos_t;
-      dcosm_dx += LSPowOfMO(p) * c_table[2*p] * (p * cos_t * dsin2_dx + \
-                    (margin - 2*p) * sin2_t * dcos_dx) * cos_t_p * sin2_t_p;
-      sin2_t_p *= sin2_t;
-    }
-    const DType df_dx = (LSPowOfMO(k) * cos_mt - 2*k) * w_norm_yi / x_norm_i * x[i][l] + \
-                         LSPowOfMO(k) * w_norm_yi * x_norm_i * dcosm_dx;
-    const DType alpha = 1 / (1 + beta);
-    x_grad[i][l] += alpha * o_grad[i][yi] * (df_dx - w[yi][l]);
-  }
-}
-
-template<typename DType>
-__global__ void LSoftmaxBackwardWKernel(const Tensor<gpu, 2, DType> x,
-                                        const Tensor<gpu, 2, DType> w,
-                                        const Tensor<gpu, 1, DType> label,
-                                        const Tensor<gpu, 1, DType> x_norm,
-                                        const Tensor<gpu, 1, DType> w_norm,
-                                        const Tensor<gpu, 2, DType> o_grad,
-                                        Tensor<gpu, 2, DType> w_grad,
-                                        const Tensor<gpu, 2, DType> workspace,
-                                        const Tensor<gpu, 1, DType> c_table,
-                                        const int margin,
-                                        const DType beta) {
-  const int nthreads = w.size(0) * w.size(1);
-  const int n = x.size(0);
-  const int feature_dim = w.size(1);
-  CUDA_KERNEL_LOOP(idx, nthreads) {
-    const int j = idx / feature_dim;
-    const int l = idx % feature_dim;
-    DType dw = 0;
-    for (int i = 0; i < n; ++i) {
-      const int yi = static_cast<int>(label[i]);
-      if (yi == j) {
-        const DType cos_t = workspace[kCost][i];
-        const DType cos_mt = workspace[kCosmt][i];
-        const int k = static_cast<int>(workspace[kK][i]);
-        const DType sin2_t = workspace[kSin2t][i];
-        const DType fo_i_yi = workspace[kFo][i];
-        const DType x_norm_i = x_norm[i];
-        const DType w_norm_yi = w_norm[yi];
-
-        const DType dcos_dw = x[i][l] / (w_norm_yi * x_norm_i) - \
-                              fo_i_yi * w[yi][l] / (x_norm_i * w_norm_yi * w_norm_yi * w_norm_yi);
-        const DType dsin2_dw = -2 * cos_t * dcos_dw;
-        DType cos_t_p = workspace[kCostM][i];
-        DType sin2_t_p = 1;
-        DType dcosm_dw = margin * cos_t_p * dcos_dw;  // p = 0
-        for (int p = 1; p <= margin / 2; ++p) {
-          cos_t_p /= cos_t * cos_t;
-          dcosm_dw += LSPowOfMO(p) * c_table[2*p] * (p * cos_t * dsin2_dw + \
-                        (margin - 2*p) * sin2_t * dcos_dw) * cos_t_p * sin2_t_p;
-          sin2_t_p *= sin2_t;
-        }
-        const DType df_dw_j = (LSPowOfMO(k) * cos_mt - 2*k) * x_norm_i / w_norm_yi * w[yi][l] + \
-                               LSPowOfMO(k) * w_norm_yi * x_norm_i * dcosm_dw;
-        dw += o_grad[i][yi] * (df_dw_j - x[i][l]);
-      }
-    }
-    const DType alpha = 1 / (1 + beta);
-    w_grad[j][l] += alpha * dw;
-  }
-}
-
-template<typename DType>
-inline void LSoftmaxBackward(const Tensor<gpu, 2, DType> &x,
-                             const Tensor<gpu, 2, DType> &w,
-                             const Tensor<gpu, 1, DType> &label,
-                             const Tensor<gpu, 1, DType> &x_norm,
-                             const Tensor<gpu, 1, DType> &w_norm,
-                             const Tensor<gpu, 2, DType> &o_grad,
-                             const Tensor<gpu, 2, DType> &x_grad,
-                             const Tensor<gpu, 2, DType> &w_grad,
-                             const Tensor<gpu, 2, DType> &workspace,
-                             const Tensor<gpu, 1, DType> &k_table,
-                             const Tensor<gpu, 1, DType> &c_table,
-                             const int margin,
-                             const DType beta) {
-  const int n = x.size(0);
-  const int feature_dim = x.size(1);
-  const int m = w.size(0);
-  dim3 dimBlock(kBaseThreadNum);
-  dim3 dimGrid((n + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSoftmaxBackwardRequired<<<dimGrid, dimBlock>>>(x, w, label, x_norm, w_norm, workspace,
-                                                  k_table, c_table, margin);
-  dimGrid.x = ((n * feature_dim + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSoftmaxBackwardXKernel<<<dimGrid, dimBlock>>>(x, w, label, x_norm, w_norm, o_grad, x_grad, workspace,
-                                                 c_table, margin, beta);
-  dimGrid.x = ((m * feature_dim + kBaseThreadNum - 1) / kBaseThreadNum);
-  LSoftmaxBackwardWKernel<<<dimGrid, dimBlock>>>(x, w, label, x_norm, w_norm, o_grad, w_grad, workspace,
-                                                 c_table, margin, beta);
-}
-
-}  // namespace cuda
-
-template<typename DType>
-inline void LSoftmaxForward(const Tensor<gpu, 2, DType> &x,
-                            const Tensor<gpu, 2, DType> &w,
-                            const Tensor<gpu, 1, DType> &label,
-                            const Tensor<gpu, 2, DType> &out,
-                            const Tensor<gpu, 1, DType> &x_norm,
-                            const Tensor<gpu, 1, DType> &w_norm,
-                            const Tensor<gpu, 1, DType> &k_table,
-                            const Tensor<gpu, 1, DType> &c_table,
-                            const int margin,
-                            const DType beta) {
-  cuda::LSoftmaxForward(x, w, label, out, x_norm, w_norm,
-                        k_table, c_table, margin, beta);
-}
-
-template<typename DType>
-inline void LSoftmaxBackward(const Tensor<gpu, 2, DType> &x,
-                             const Tensor<gpu, 2, DType> &w,
-                             const Tensor<gpu, 1, DType> &label,
-                             const Tensor<gpu, 1, DType> &x_norm,
-                             const Tensor<gpu, 1, DType> &w_norm,
-                             const Tensor<gpu, 2, DType> &o_grad,
-                             const Tensor<gpu, 2, DType> &x_grad,
-                             const Tensor<gpu, 2, DType> &w_grad,
-                             const Tensor<gpu, 2, DType> &workspace,
-                             const Tensor<gpu, 1, DType> &k_table,
-                             const Tensor<gpu, 1, DType> &c_table,
-                             const int margin,
-                             const DType beta) {
-  cuda::LSoftmaxBackward(x, w, label, x_norm, w_norm, o_grad, x_grad, w_grad, workspace,
-                         k_table, c_table, margin, beta);
-}
-
-}  // namespace mshadow
-
-namespace mxnet {
-namespace op {
-
-template<>
-Operator *CreateOp<gpu>(LSoftmaxParam param, int dtype) {
-  Operator *op = NULL;
-  MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
-    op = new LSoftmaxOp<gpu, DType>(param);
-  })
-  return op;
-}
-
-}  // namespace op
-}  // namespace mxnet
diff --git a/CODEOWNERS b/CODEOWNERS
new file mode 100644
index 000000000..4e6275c24
--- /dev/null
+++ b/CODEOWNERS
@@ -0,0 +1 @@
+/server/ @nttstar
diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md
new file mode 100644
index 000000000..a1d657f00
--- /dev/null
+++ b/CODE_OF_CONDUCT.md
@@ -0,0 +1,128 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, religion, or sexual identity
+and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the
+  overall community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official e-mail address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement at
+contact@insightface.ai.
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series
+of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or
+permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior,  harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within
+the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.0, available at
+https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct
+enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.
diff --git a/Evaluation/IJB/IJBB_Evaluation_MS1MV2.ipynb b/Evaluation/IJB/IJBB_Evaluation_MS1MV2.ipynb
deleted file mode 100644
index c18234e1a..000000000
--- a/Evaluation/IJB/IJBB_Evaluation_MS1MV2.ipynb
+++ /dev/null
@@ -1,520 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/jd4615/miniconda3/envs/insightface/lib/python2.7/site-packages/sklearn/utils/fixes.py:313: FutureWarning: numpy not_equal will not check object identity in the future. The comparison did not return the same result as suggested by the identity (`is`)) and will change.\n",
-      "  _nan_object_mask = _nan_object_array != _nan_object_array\n"
-     ]
-    }
-   ],
-   "source": [
-    "import os\n",
-    "import numpy as np\n",
-    "import cPickle\n",
-    "from sklearn.metrics import roc_curve, auc\n",
-    "import matplotlib.pyplot as plt\n",
-    "import timeit\n",
-    "import sklearn\n",
-    "import cv2\n",
-    "import sys\n",
-    "import glob\n",
-    "sys.path.append('./recognition')\n",
-    "from embedding import Embedding\n",
-    "from menpo.visualize import print_progress\n",
-    "from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap\n",
-    "from prettytable import PrettyTable\n",
-    "from pathlib import Path\n",
-    "import warnings \n",
-    "warnings.filterwarnings(\"ignore\")  "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_media_list(path):\n",
-    "    ijb_meta = np.loadtxt(path, dtype=str)\n",
-    "    templates = ijb_meta[:,1].astype(np.int)\n",
-    "    medias = ijb_meta[:,2].astype(np.int)\n",
-    "    return templates, medias"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_pair_list(path):\n",
-    "    pairs = np.loadtxt(path, dtype=str)\n",
-    "    t1 = pairs[:,0].astype(np.int)\n",
-    "    t2 = pairs[:,1].astype(np.int)\n",
-    "    label = pairs[:,2].astype(np.int)\n",
-    "    return t1, t2, label"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_image_feature(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def get_image_feature(img_path, img_list_path, model_path, gpu_id):\n",
-    "    img_list = open(img_list_path)\n",
-    "    embedding = Embedding(model_path, 0, gpu_id)\n",
-    "    files = img_list.readlines()\n",
-    "    img_feats = []\n",
-    "    faceness_scores = []\n",
-    "    for img_index, each_line in enumerate(print_progress(files)):\n",
-    "        name_lmk_score = each_line.strip().split(' ')\n",
-    "        img_name = os.path.join(img_path, name_lmk_score[0])\n",
-    "        img = cv2.imread(img_name)\n",
-    "        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)\n",
-    "        lmk = lmk.reshape( (5,2) )\n",
-    "        img_feats.append(embedding.get(img,lmk))\n",
-    "        faceness_scores.append(name_lmk_score[-1])\n",
-    "    img_feats = np.array(img_feats).astype(np.float32)\n",
-    "    faceness_scores = np.array(faceness_scores).astype(np.float32)\n",
-    "    return img_feats, faceness_scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def image2template_feature(img_feats = None, templates = None, medias = None):\n",
-    "    # ==========================================================\n",
-    "    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]\n",
-    "    # 2. compute media feature.\n",
-    "    # 3. compute template feature.\n",
-    "    # ==========================================================    \n",
-    "    unique_templates = np.unique(templates)\n",
-    "    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))\n",
-    "\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        (ind_t,) = np.where(templates == uqt)\n",
-    "        face_norm_feats = img_feats[ind_t]\n",
-    "        face_medias = medias[ind_t]\n",
-    "        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)\n",
-    "        media_norm_feats = []\n",
-    "        for u,ct in zip(unique_medias, unique_media_counts):\n",
-    "            (ind_m,) = np.where(face_medias == u)\n",
-    "            if ct == 1:\n",
-    "                media_norm_feats += [face_norm_feats[ind_m]]\n",
-    "            else: # image features from the same video will be aggregated into one feature\n",
-    "                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]\n",
-    "        media_norm_feats = np.array(media_norm_feats)\n",
-    "        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))\n",
-    "        template_feats[count_template] = np.sum(media_norm_feats, 0)\n",
-    "        if count_template % 2000 == 0: \n",
-    "            print('Finish Calculating {} template features.'.format(count_template))\n",
-    "    template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))\n",
-    "    return template_norm_feats, unique_templates"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):\n",
-    "    # ==========================================================\n",
-    "    #         Compute set-to-set Similarity Score.\n",
-    "    # ==========================================================\n",
-    "    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        template2id[uqt] = count_template\n",
-    "    \n",
-    "    score = np.zeros((len(p1),))   # save cosine distance between pairs \n",
-    "\n",
-    "    total_pairs = np.array(range(len(p1)))\n",
-    "    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation\n",
-    "    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]\n",
-    "    total_sublists = len(sublists)\n",
-    "    for c, s in enumerate(sublists):\n",
-    "        feat1 = template_norm_feats[template2id[p1[s]]]\n",
-    "        feat2 = template_norm_feats[template2id[p2[s]]]\n",
-    "        similarity_score = np.sum(feat1 * feat2, -1)\n",
-    "        score[s] = similarity_score.flatten()\n",
-    "        if c % 10 == 0:\n",
-    "            print('Finish {}/{} pairs.'.format(c, total_sublists))\n",
-    "    return score"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_score(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step1: Load Meta Data"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 0.83 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image and template relationships for template feature embedding\n",
-    "# tid --> template id,  mid --> media id \n",
-    "# format:\n",
-    "#           image_name tid mid\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "templates, medias = read_template_media_list(os.path.join('IJBB/meta', 'ijbb_face_tid_mid.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 31.88 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load template pairs for template-to-template verification\n",
-    "# tid : template id,  label : 1/0\n",
-    "# format:\n",
-    "#           tid_1 tid_2 label\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "p1, p2, label = read_template_pair_list(os.path.join('IJBB/meta', 'ijbb_template_pair_label.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 2: Get Image Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 12,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "('loading', './pretrained_models/MS1MV2-ResNet100-Arcface/model', 0)\n",
-      "[====================] 100% (227630/227630) - done.                             \n",
-      "Time: 3279.69 s. \n",
-      "Feature Shape: (227630 , 1024) .\n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image features \n",
-    "# format:\n",
-    "#           img_feats: [image_num x feats_dim] (227630, 512)\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "#img_feats = read_image_feature('./MS1MV2/IJBB_MS1MV2_r100_arcface.pkl')\n",
-    "img_path = './IJBB/loose_crop'\n",
-    "img_list_path = './IJBB/meta/ijbb_name_5pts_score.txt'\n",
-    "model_path = './pretrained_models/MS1MV2-ResNet100-Arcface/model'\n",
-    "gpu_id = 1\n",
-    "img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, gpu_id)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))\n",
-    "print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step3: Get Template Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 45,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish Calculating 0 template features.\n",
-      "Finish Calculating 2000 template features.\n",
-      "Finish Calculating 4000 template features.\n",
-      "Finish Calculating 6000 template features.\n",
-      "Finish Calculating 8000 template features.\n",
-      "Finish Calculating 10000 template features.\n",
-      "Finish Calculating 12000 template features.\n",
-      "Time: 3.65 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute template features from image features.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "# ========================================================== \n",
-    "# Norm feature before aggregation into template feature?\n",
-    "# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).\n",
-    "# ========================================================== \n",
-    "# 1. FaceScore （Feature Norm）\n",
-    "# 2. FaceScore （Detector）\n",
-    "\n",
-    "use_norm_score = True # if True, TestMode(N1)  \n",
-    "use_detector_score = True # if True, TestMode(D1)\n",
-    "use_flip_test = True # if True, TestMode(F2)\n",
-    "\n",
-    "if use_flip_test:\n",
-    "    # concat --- F1\n",
-    "    # img_input_feats = img_feats \n",
-    "    # add --- F2\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2] + img_feats[:,img_feats.shape[1]/2:]\n",
-    "else:\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2]\n",
-    "    \n",
-    "if use_norm_score:\n",
-    "    img_input_feats = img_input_feats\n",
-    "else:\n",
-    "    # normalise features to remove norm information\n",
-    "    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    \n",
-    "    \n",
-    "if use_detector_score:\n",
-    "    img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])\n",
-    "else:\n",
-    "    img_input_feats = img_input_feats\n",
-    "\n",
-    "template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 4: Get Template Similarity Scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 46,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish 0/81 pairs.\n",
-      "Finish 10/81 pairs.\n",
-      "Finish 20/81 pairs.\n",
-      "Finish 30/81 pairs.\n",
-      "Finish 40/81 pairs.\n",
-      "Finish 50/81 pairs.\n",
-      "Finish 60/81 pairs.\n",
-      "Finish 70/81 pairs.\n",
-      "Finish 80/81 pairs.\n",
-      "Time: 77.30 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute verification scores between template pairs.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "score = verification(template_norm_feats, unique_templates, p1, p2)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 47,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "score_save_name = './IJBB/result/MS1MV2-ResNet100-ArcFace-TestMode(N1D1F2).npy'\n",
-    "np.save(score_save_name, score)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 5: Get ROC Curves and TPR@FPR Table"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 48,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "<Figure size 432x288 with 1 Axes>"
-      ]
-     },
-     "metadata": {
-      "needs_background": "light"
-     },
-     "output_type": "display_data"
-    }
-   ],
-   "source": [
-    "score_save_path = './IJBB/result'\n",
-    "files = glob.glob(score_save_path + '/MS1MV2*.npy')  \n",
-    "methods = []\n",
-    "scores = []\n",
-    "for file in files:\n",
-    "    methods.append(Path(file).stem)\n",
-    "    scores.append(np.load(file)) \n",
-    "methods = np.array(methods)\n",
-    "scores = dict(zip(methods,scores))\n",
-    "colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))\n",
-    "#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]\n",
-    "x_labels = [10**-6, 10**-5, 10**-4,10**-3, 10**-2, 10**-1]\n",
-    "tpr_fpr_table = PrettyTable(['Methods'] + map(str, x_labels))\n",
-    "fig = plt.figure()\n",
-    "for method in methods:\n",
-    "    fpr, tpr, _ = roc_curve(label, scores[method])\n",
-    "    roc_auc = auc(fpr, tpr)\n",
-    "    fpr = np.flipud(fpr)\n",
-    "    tpr = np.flipud(tpr) # select largest tpr at same fpr\n",
-    "    plt.plot(fpr, tpr, color=colours[method], lw=1, label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc*100)))\n",
-    "    tpr_fpr_row = []\n",
-    "    tpr_fpr_row.append(method)\n",
-    "    for fpr_iter in np.arange(len(x_labels)):\n",
-    "        _, min_index = min(list(zip(abs(fpr-x_labels[fpr_iter]), range(len(fpr)))))\n",
-    "        tpr_fpr_row.append('%.4f' % tpr[min_index])\n",
-    "    tpr_fpr_table.add_row(tpr_fpr_row)\n",
-    "plt.xlim([10**-6, 0.1])\n",
-    "plt.ylim([0.3, 1.0])\n",
-    "plt.grid(linestyle='--', linewidth=1)\n",
-    "plt.xticks(x_labels) \n",
-    "plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True)) \n",
-    "plt.xscale('log')\n",
-    "plt.xlabel('False Positive Rate')\n",
-    "plt.ylabel('True Positive Rate')\n",
-    "plt.title('ROC on IJB-B')\n",
-    "plt.legend(loc=\"lower right\")\n",
-    "plt.show()\n",
-    "#fig.savefig('IJB-B.pdf')"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 49,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "|                  Methods                  | 1e-06  | 1e-05  | 0.0001 | 0.001  |  0.01  |  0.1   |\n",
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F1) | 0.4091 | 0.9081 | 0.9477 | 0.9636 | 0.9755 | 0.9863 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D1F2) | 0.4089 | 0.8995 | 0.9463 | 0.9642 | 0.9761 | 0.9867 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F2) | 0.4281 | 0.9082 | 0.9490 | 0.9647 | 0.9767 | 0.9866 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D0F0) | 0.3900 | 0.9042 | 0.9467 | 0.9620 | 0.9761 | 0.9860 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D0F0) | 0.3828 | 0.8933 | 0.9425 | 0.9615 | 0.9751 | 0.9856 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F0) | 0.3930 | 0.9039 | 0.9476 | 0.9630 | 0.9758 | 0.9861 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D1F0) | 0.3892 | 0.8984 | 0.9456 | 0.9626 | 0.9753 | 0.9861 |\n",
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(tpr_fpr_table)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# setting N1D1F2 is the best"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 2",
-   "language": "python",
-   "name": "python2"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 2
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython2",
-   "version": "2.7.15"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/Evaluation/IJB/IJBB_Evaluation_VGG2.ipynb b/Evaluation/IJB/IJBB_Evaluation_VGG2.ipynb
deleted file mode 100644
index 0ed5fa7f2..000000000
--- a/Evaluation/IJB/IJBB_Evaluation_VGG2.ipynb
+++ /dev/null
@@ -1,535 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/jd4615/miniconda3/envs/insightface/lib/python2.7/site-packages/sklearn/utils/fixes.py:313: FutureWarning: numpy not_equal will not check object identity in the future. The comparison did not return the same result as suggested by the identity (`is`)) and will change.\n",
-      "  _nan_object_mask = _nan_object_array != _nan_object_array\n"
-     ]
-    }
-   ],
-   "source": [
-    "import os\n",
-    "import numpy as np\n",
-    "import cPickle\n",
-    "from sklearn.metrics import roc_curve, auc\n",
-    "import matplotlib.pyplot as plt\n",
-    "import timeit\n",
-    "import sklearn\n",
-    "import cv2\n",
-    "import sys\n",
-    "import glob\n",
-    "sys.path.append('./recognition')\n",
-    "from embedding import Embedding\n",
-    "from menpo.visualize import print_progress\n",
-    "from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap\n",
-    "from prettytable import PrettyTable\n",
-    "from pathlib import Path\n",
-    "import warnings \n",
-    "warnings.filterwarnings(\"ignore\")  "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_media_list(path):\n",
-    "    ijb_meta = np.loadtxt(path, dtype=str)\n",
-    "    templates = ijb_meta[:,1].astype(np.int)\n",
-    "    medias = ijb_meta[:,2].astype(np.int)\n",
-    "    return templates, medias"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_pair_list(path):\n",
-    "    pairs = np.loadtxt(path, dtype=str)\n",
-    "    t1 = pairs[:,0].astype(np.int)\n",
-    "    t2 = pairs[:,1].astype(np.int)\n",
-    "    label = pairs[:,2].astype(np.int)\n",
-    "    return t1, t2, label"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_image_feature(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def get_image_feature(img_path, img_list_path, model_path, gpu_id):\n",
-    "    img_list = open(img_list_path)\n",
-    "    embedding = Embedding(model_path, 0, gpu_id)\n",
-    "    files = img_list.readlines()\n",
-    "    img_feats = []\n",
-    "    faceness_scores = []\n",
-    "    for img_index, each_line in enumerate(print_progress(files)):\n",
-    "        name_lmk_score = each_line.strip().split(' ')\n",
-    "        img_name = os.path.join(img_path, name_lmk_score[0])\n",
-    "        img = cv2.imread(img_name)\n",
-    "        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)\n",
-    "        lmk = lmk.reshape( (5,2) )\n",
-    "        img_feats.append(embedding.get(img,lmk))\n",
-    "        faceness_scores.append(name_lmk_score[-1])\n",
-    "    img_feats = np.array(img_feats).astype(np.float32)\n",
-    "    faceness_scores = np.array(faceness_scores).astype(np.float32)\n",
-    "    return img_feats, faceness_scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def image2template_feature(img_feats = None, templates = None, medias = None):\n",
-    "    # ==========================================================\n",
-    "    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]\n",
-    "    # 2. compute media feature.\n",
-    "    # 3. compute template feature.\n",
-    "    # ==========================================================    \n",
-    "    unique_templates = np.unique(templates)\n",
-    "    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))\n",
-    "\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        (ind_t,) = np.where(templates == uqt)\n",
-    "        face_norm_feats = img_feats[ind_t]\n",
-    "        face_medias = medias[ind_t]\n",
-    "        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)\n",
-    "        media_norm_feats = []\n",
-    "        for u,ct in zip(unique_medias, unique_media_counts):\n",
-    "            (ind_m,) = np.where(face_medias == u)\n",
-    "            if ct == 1:\n",
-    "                media_norm_feats += [face_norm_feats[ind_m]]\n",
-    "            else: # image features from the same video will be aggregated into one feature\n",
-    "                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]\n",
-    "        media_norm_feats = np.array(media_norm_feats)\n",
-    "        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))\n",
-    "        template_feats[count_template] = np.sum(media_norm_feats, 0)\n",
-    "        if count_template % 2000 == 0: \n",
-    "            print('Finish Calculating {} template features.'.format(count_template))\n",
-    "    template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))\n",
-    "    return template_norm_feats, unique_templates"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):\n",
-    "    # ==========================================================\n",
-    "    #         Compute set-to-set Similarity Score.\n",
-    "    # ==========================================================\n",
-    "    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        template2id[uqt] = count_template\n",
-    "    \n",
-    "    score = np.zeros((len(p1),))   # save cosine distance between pairs \n",
-    "\n",
-    "    total_pairs = np.array(range(len(p1)))\n",
-    "    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation\n",
-    "    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]\n",
-    "    total_sublists = len(sublists)\n",
-    "    for c, s in enumerate(sublists):\n",
-    "        feat1 = template_norm_feats[template2id[p1[s]]]\n",
-    "        feat2 = template_norm_feats[template2id[p2[s]]]\n",
-    "        similarity_score = np.sum(feat1 * feat2, -1)\n",
-    "        score[s] = similarity_score.flatten()\n",
-    "        if c % 10 == 0:\n",
-    "            print('Finish {}/{} pairs.'.format(c, total_sublists))\n",
-    "    return score"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_score(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step1: Load Meta Data"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 0.83 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image and template relationships for template feature embedding\n",
-    "# tid --> template id,  mid --> media id \n",
-    "# format:\n",
-    "#           image_name tid mid\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "templates, medias = read_template_media_list(os.path.join('IJBB/meta', 'ijbb_face_tid_mid.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 31.75 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load template pairs for template-to-template verification\n",
-    "# tid : template id,  label : 1/0\n",
-    "# format:\n",
-    "#           tid_1 tid_2 label\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "p1, p2, label = read_template_pair_list(os.path.join('IJBB/meta', 'ijbb_template_pair_label.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 2: Get Image Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 11,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "('loading', './pretrained_models/VGG2-ResNet50-Arcface/model', 0)\n",
-      "[====================] 100% (227630/227630) - done.                             \n",
-      "Time: 2386.28 s. \n",
-      "Feature Shape: (227630 , 1024) .\n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image features \n",
-    "# format:\n",
-    "#           img_feats: [image_num x feats_dim] (227630, 512)\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "#img_feats = read_image_feature('./MS1MV2/IJBB_MS1MV2_r100_arcface.pkl')\n",
-    "img_path = './IJBB/loose_crop'\n",
-    "img_list_path = './IJBB/meta/ijbb_name_5pts_score.txt'\n",
-    "model_path = './pretrained_models/VGG2-ResNet50-Arcface/model'\n",
-    "gpu_id = 0\n",
-    "img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, gpu_id)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))\n",
-    "print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step3: Get Template Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 12,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish Calculating 0 template features.\n",
-      "Finish Calculating 2000 template features.\n",
-      "Finish Calculating 4000 template features.\n",
-      "Finish Calculating 6000 template features.\n",
-      "Finish Calculating 8000 template features.\n",
-      "Finish Calculating 10000 template features.\n",
-      "Finish Calculating 12000 template features.\n",
-      "Time: 3.41 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute template features from image features.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "# ========================================================== \n",
-    "# Norm feature before aggregation into template feature?\n",
-    "# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).\n",
-    "# ========================================================== \n",
-    "# 1. FaceScore （Feature Norm）\n",
-    "# 2. FaceScore （Detector）\n",
-    "\n",
-    "use_norm_score = False # if True, TestMode(N1)  \n",
-    "use_detector_score = True # if True, TestMode(D1)\n",
-    "use_flip_test = True # if True, TestMode(F1)\n",
-    "\n",
-    "if use_flip_test:\n",
-    "    # concat --- F1\n",
-    "    #img_input_feats = img_feats \n",
-    "    # add --- F2\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2] + img_feats[:,img_feats.shape[1]/2:]\n",
-    "else:\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2]\n",
-    "    \n",
-    "if use_norm_score:\n",
-    "    img_input_feats = img_input_feats\n",
-    "else:\n",
-    "    # normalise features to remove norm information\n",
-    "    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    \n",
-    "    \n",
-    "if use_detector_score:\n",
-    "    img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])\n",
-    "else:\n",
-    "    img_input_feats = img_input_feats\n",
-    "\n",
-    "template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 4: Get Template Similarity Scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 13,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish 0/81 pairs.\n",
-      "Finish 10/81 pairs.\n",
-      "Finish 20/81 pairs.\n",
-      "Finish 30/81 pairs.\n",
-      "Finish 40/81 pairs.\n",
-      "Finish 50/81 pairs.\n",
-      "Finish 60/81 pairs.\n",
-      "Finish 70/81 pairs.\n",
-      "Finish 80/81 pairs.\n",
-      "Time: 38.38 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute verification scores between template pairs.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "score = verification(template_norm_feats, unique_templates, p1, p2)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 14,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "score_save_name = './IJBB/result/VGG2-ResNet50-ArcFace-TestMode(N0D1F2).npy'\n",
-    "np.save(score_save_name, score)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 5: Get ROC Curves and TPR@FPR Table"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "<Figure size 432x288 with 1 Axes>"
-      ]
-     },
-     "metadata": {
-      "needs_background": "light"
-     },
-     "output_type": "display_data"
-    }
-   ],
-   "source": [
-    "score_save_path = './IJBB/result'\n",
-    "files = glob.glob(score_save_path + '/VGG2*.npy')  \n",
-    "methods = []\n",
-    "scores = []\n",
-    "for file in files:\n",
-    "    methods.append(Path(file).stem)\n",
-    "    scores.append(np.load(file)) \n",
-    "methods = np.array(methods)\n",
-    "scores = dict(zip(methods,scores))\n",
-    "colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))\n",
-    "#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]\n",
-    "x_labels = [10**-6, 10**-5, 10**-4,10**-3, 10**-2, 10**-1]\n",
-    "tpr_fpr_table = PrettyTable(['Methods'] + map(str, x_labels))\n",
-    "fig = plt.figure()\n",
-    "for method in methods:\n",
-    "    fpr, tpr, _ = roc_curve(label, scores[method])\n",
-    "    roc_auc = auc(fpr, tpr)\n",
-    "    fpr = np.flipud(fpr)\n",
-    "    tpr = np.flipud(tpr) # select largest tpr at same fpr\n",
-    "    plt.plot(fpr, tpr, color=colours[method], lw=1, label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc*100)))\n",
-    "    tpr_fpr_row = []\n",
-    "    tpr_fpr_row.append(method)\n",
-    "    for fpr_iter in np.arange(len(x_labels)):\n",
-    "        _, min_index = min(list(zip(abs(fpr-x_labels[fpr_iter]), range(len(fpr)))))\n",
-    "        tpr_fpr_row.append('%.4f' % tpr[min_index])\n",
-    "    tpr_fpr_table.add_row(tpr_fpr_row)\n",
-    "plt.xlim([10**-6, 0.1])\n",
-    "plt.ylim([0.3, 1.0])\n",
-    "plt.grid(linestyle='--', linewidth=1)\n",
-    "plt.xticks(x_labels) \n",
-    "plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True)) \n",
-    "plt.xscale('log')\n",
-    "plt.xlabel('False Positive Rate')\n",
-    "plt.ylabel('True Positive Rate')\n",
-    "plt.title('ROC on IJB-B')\n",
-    "plt.legend(loc=\"lower right\")\n",
-    "plt.show()\n",
-    "#fig.savefig('IJB-B.pdf')"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "|                Methods                 | 1e-06  | 1e-05  | 0.0001 | 0.001  |  0.01  |  0.1   |\n",
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F2) | 0.4044 | 0.8145 | 0.9056 | 0.9497 | 0.9779 | 0.9922 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D0F0) | 0.4035 | 0.8038 | 0.8976 | 0.9437 | 0.9755 | 0.9914 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F1) | 0.3940 | 0.8124 | 0.9028 | 0.9479 | 0.9770 | 0.9919 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D0F0) | 0.3893 | 0.8050 | 0.8990 | 0.9448 | 0.9759 | 0.9918 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F0) | 0.4098 | 0.8123 | 0.9022 | 0.9463 | 0.9766 | 0.9918 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D1F0) | 0.3949 | 0.8130 | 0.9036 | 0.9471 | 0.9767 | 0.9919 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D1F2) | 0.4011 | 0.8210 | 0.9069 | 0.9500 | 0.9779 | 0.9924 |\n",
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(tpr_fpr_table)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# setting N0D1F2 is the best"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Test Setting Conclusions"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### (1) add is better than concat for the flip test (N1D1F2 v.s. N1D1F1)\n",
-    "#### (2) detection score contains some faceness information to decrease weights of noise samples within the template (N0D1F0 v.s. N0D0F0)"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 2",
-   "language": "python",
-   "name": "python2"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 2
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython2",
-   "version": "2.7.15"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/Evaluation/IJB/IJBC_Evaluation_MS1MV2.ipynb b/Evaluation/IJB/IJBC_Evaluation_MS1MV2.ipynb
deleted file mode 100644
index e364be8cc..000000000
--- a/Evaluation/IJB/IJBC_Evaluation_MS1MV2.ipynb
+++ /dev/null
@@ -1,532 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/jd4615/miniconda3/envs/insightface/lib/python2.7/site-packages/sklearn/utils/fixes.py:313: FutureWarning: numpy not_equal will not check object identity in the future. The comparison did not return the same result as suggested by the identity (`is`)) and will change.\n",
-      "  _nan_object_mask = _nan_object_array != _nan_object_array\n"
-     ]
-    }
-   ],
-   "source": [
-    "import os\n",
-    "import numpy as np\n",
-    "import cPickle\n",
-    "from sklearn.metrics import roc_curve, auc\n",
-    "import matplotlib.pyplot as plt\n",
-    "import timeit\n",
-    "import sklearn\n",
-    "import cv2\n",
-    "import sys\n",
-    "import glob\n",
-    "sys.path.append('./recognition')\n",
-    "from embedding import Embedding\n",
-    "from menpo.visualize import print_progress\n",
-    "from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap\n",
-    "from prettytable import PrettyTable\n",
-    "from pathlib import Path\n",
-    "import warnings \n",
-    "warnings.filterwarnings(\"ignore\")  "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_media_list(path):\n",
-    "    ijb_meta = np.loadtxt(path, dtype=str)\n",
-    "    templates = ijb_meta[:,1].astype(np.int)\n",
-    "    medias = ijb_meta[:,2].astype(np.int)\n",
-    "    return templates, medias"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_pair_list(path):\n",
-    "    pairs = np.loadtxt(path, dtype=str)\n",
-    "    t1 = pairs[:,0].astype(np.int)\n",
-    "    t2 = pairs[:,1].astype(np.int)\n",
-    "    label = pairs[:,2].astype(np.int)\n",
-    "    return t1, t2, label"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_image_feature(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def get_image_feature(img_path, img_list_path, model_path, gpu_id):\n",
-    "    img_list = open(img_list_path)\n",
-    "    embedding = Embedding(model_path, 0, gpu_id)\n",
-    "    files = img_list.readlines()\n",
-    "    img_feats = []\n",
-    "    faceness_scores = []\n",
-    "    for img_index, each_line in enumerate(print_progress(files)):\n",
-    "        name_lmk_score = each_line.strip().split(' ')\n",
-    "        img_name = os.path.join(img_path, name_lmk_score[0])\n",
-    "        img = cv2.imread(img_name)\n",
-    "        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)\n",
-    "        lmk = lmk.reshape( (5,2) )\n",
-    "        img_feats.append(embedding.get(img,lmk))\n",
-    "        faceness_scores.append(name_lmk_score[-1])\n",
-    "    img_feats = np.array(img_feats).astype(np.float32)\n",
-    "    faceness_scores = np.array(faceness_scores).astype(np.float32)\n",
-    "    return img_feats, faceness_scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def image2template_feature(img_feats = None, templates = None, medias = None):\n",
-    "    # ==========================================================\n",
-    "    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]\n",
-    "    # 2. compute media feature.\n",
-    "    # 3. compute template feature.\n",
-    "    # ==========================================================    \n",
-    "    unique_templates = np.unique(templates)\n",
-    "    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))\n",
-    "\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        (ind_t,) = np.where(templates == uqt)\n",
-    "        face_norm_feats = img_feats[ind_t]\n",
-    "        face_medias = medias[ind_t]\n",
-    "        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)\n",
-    "        media_norm_feats = []\n",
-    "        for u,ct in zip(unique_medias, unique_media_counts):\n",
-    "            (ind_m,) = np.where(face_medias == u)\n",
-    "            if ct == 1:\n",
-    "                media_norm_feats += [face_norm_feats[ind_m]]\n",
-    "            else: # image features from the same video will be aggregated into one feature\n",
-    "                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]\n",
-    "        media_norm_feats = np.array(media_norm_feats)\n",
-    "        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))\n",
-    "        template_feats[count_template] = np.sum(media_norm_feats, 0)\n",
-    "        if count_template % 2000 == 0: \n",
-    "            print('Finish Calculating {} template features.'.format(count_template))\n",
-    "    template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))\n",
-    "    return template_norm_feats, unique_templates"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 19,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):\n",
-    "    # ==========================================================\n",
-    "    #         Compute set-to-set Similarity Score.\n",
-    "    # ==========================================================\n",
-    "    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        template2id[uqt] = count_template\n",
-    "    \n",
-    "    score = np.zeros((len(p1),))   # save cosine distance between pairs \n",
-    "\n",
-    "    total_pairs = np.array(range(len(p1)))\n",
-    "    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation\n",
-    "    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]\n",
-    "    total_sublists = len(sublists)\n",
-    "    for c, s in enumerate(sublists):\n",
-    "        feat1 = template_norm_feats[template2id[p1[s]]]\n",
-    "        feat2 = template_norm_feats[template2id[p2[s]]]\n",
-    "        similarity_score = np.sum(feat1 * feat2, -1)\n",
-    "        score[s] = similarity_score.flatten()\n",
-    "        if c % 10 == 0:\n",
-    "            print('Finish {}/{} pairs.'.format(c, total_sublists))\n",
-    "    return score"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_score(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step1: Load Meta Data"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 1.73 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image and template relationships for template feature embedding\n",
-    "# tid --> template id,  mid --> media id \n",
-    "# format:\n",
-    "#           image_name tid mid\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "templates, medias = read_template_media_list(os.path.join('IJBC/meta', 'ijbc_face_tid_mid.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 63.98 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load template pairs for template-to-template verification\n",
-    "# tid : template id,  label : 1/0\n",
-    "# format:\n",
-    "#           tid_1 tid_2 label\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "p1, p2, label = read_template_pair_list(os.path.join('IJBC/meta', 'ijbc_template_pair_label.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 2: Get Image Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 12,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "('loading', './pretrained_models/MS1MV2-ResNet100-Arcface/model', 0)\n",
-      "[====================] 100% (469375/469375) - done.                             \n",
-      "Time: 6806.24 s. \n",
-      "Feature Shape: (469375 , 1024) .\n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image features \n",
-    "# format:\n",
-    "#           img_feats: [image_num x feats_dim] (227630, 512)\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "#img_feats = read_image_feature('./MS1MV2/IJBB_MS1MV2_r100_arcface.pkl')\n",
-    "img_path = './IJBC/loose_crop'\n",
-    "img_list_path = './IJBC/meta/ijbc_name_5pts_score.txt'\n",
-    "model_path = './pretrained_models/MS1MV2-ResNet100-Arcface/model'\n",
-    "gpu_id = 1\n",
-    "img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, gpu_id)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))\n",
-    "print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step3: Get Template Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 34,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish Calculating 0 template features.\n",
-      "Finish Calculating 2000 template features.\n",
-      "Finish Calculating 4000 template features.\n",
-      "Finish Calculating 6000 template features.\n",
-      "Finish Calculating 8000 template features.\n",
-      "Finish Calculating 10000 template features.\n",
-      "Finish Calculating 12000 template features.\n",
-      "Finish Calculating 14000 template features.\n",
-      "Finish Calculating 16000 template features.\n",
-      "Finish Calculating 18000 template features.\n",
-      "Finish Calculating 20000 template features.\n",
-      "Finish Calculating 22000 template features.\n",
-      "Time: 7.85 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute template features from image features.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "# ========================================================== \n",
-    "# Norm feature before aggregation into template feature?\n",
-    "# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).\n",
-    "# ========================================================== \n",
-    "# 1. FaceScore （Feature Norm）\n",
-    "# 2. FaceScore （Detector）\n",
-    "\n",
-    "use_norm_score = False # if Ture, TestMode(N1)  \n",
-    "use_detector_score = False # if Ture, TestMode(D1)\n",
-    "use_flip_test = False # if Ture, TestMode(F1)\n",
-    "\n",
-    "if use_flip_test:\n",
-    "    # concat --- F1\n",
-    "    #img_input_feats = img_feats \n",
-    "    # add --- F2\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2] + img_feats[:,img_feats.shape[1]/2:]\n",
-    "else:\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2]\n",
-    "    \n",
-    "if use_norm_score:\n",
-    "    img_input_feats = img_input_feats\n",
-    "else:\n",
-    "    # normalise features to remove norm information\n",
-    "    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    \n",
-    "    \n",
-    "if use_detector_score:\n",
-    "    img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])\n",
-    "else:\n",
-    "    img_input_feats = img_input_feats\n",
-    "\n",
-    "template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 4: Get Template Similarity Scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 35,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish 0/157 pairs.\n",
-      "Finish 10/157 pairs.\n",
-      "Finish 20/157 pairs.\n",
-      "Finish 30/157 pairs.\n",
-      "Finish 40/157 pairs.\n",
-      "Finish 50/157 pairs.\n",
-      "Finish 60/157 pairs.\n",
-      "Finish 70/157 pairs.\n",
-      "Finish 80/157 pairs.\n",
-      "Finish 90/157 pairs.\n",
-      "Finish 100/157 pairs.\n",
-      "Finish 110/157 pairs.\n",
-      "Finish 120/157 pairs.\n",
-      "Finish 130/157 pairs.\n",
-      "Finish 140/157 pairs.\n",
-      "Finish 150/157 pairs.\n",
-      "Time: 67.17 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute verification scores between template pairs.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "score = verification(template_norm_feats, unique_templates, p1, p2)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 36,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "score_save_name = './IJBC/result/MS1MV2-ResNet100-ArcFace-TestMode(N0D0F0).npy'\n",
-    "np.save(score_save_name, score)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 5: Get ROC Curves and TPR@FPR Table"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 38,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "<Figure size 432x288 with 1 Axes>"
-      ]
-     },
-     "metadata": {
-      "needs_background": "light"
-     },
-     "output_type": "display_data"
-    }
-   ],
-   "source": [
-    "score_save_path = './IJBC/result'\n",
-    "files = glob.glob(score_save_path + '/MS1MV2*.npy')  \n",
-    "methods = []\n",
-    "scores = []\n",
-    "for file in files:\n",
-    "    methods.append(Path(file).stem)\n",
-    "    scores.append(np.load(file)) \n",
-    "methods = np.array(methods)\n",
-    "scores = dict(zip(methods,scores))\n",
-    "colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))\n",
-    "#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]\n",
-    "x_labels = [10**-6, 10**-5, 10**-4,10**-3, 10**-2, 10**-1]\n",
-    "tpr_fpr_table = PrettyTable(['Methods'] + map(str, x_labels))\n",
-    "fig = plt.figure()\n",
-    "for method in methods:\n",
-    "    fpr, tpr, _ = roc_curve(label, scores[method])\n",
-    "    roc_auc = auc(fpr, tpr)\n",
-    "    fpr = np.flipud(fpr)\n",
-    "    tpr = np.flipud(tpr) # select largest tpr at same fpr\n",
-    "    plt.plot(fpr, tpr, color=colours[method], lw=1, label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc*100)))\n",
-    "    tpr_fpr_row = []\n",
-    "    tpr_fpr_row.append(method)\n",
-    "    for fpr_iter in np.arange(len(x_labels)):\n",
-    "        _, min_index = min(list(zip(abs(fpr-x_labels[fpr_iter]), range(len(fpr)))))\n",
-    "        tpr_fpr_row.append('%.4f' % tpr[min_index])\n",
-    "    tpr_fpr_table.add_row(tpr_fpr_row)\n",
-    "plt.xlim([10**-6, 0.1])\n",
-    "plt.ylim([0.3, 1.0])\n",
-    "plt.grid(linestyle='--', linewidth=1)\n",
-    "plt.xticks(x_labels) \n",
-    "plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True)) \n",
-    "plt.xscale('log')\n",
-    "plt.xlabel('False Positive Rate')\n",
-    "plt.ylabel('True Positive Rate')\n",
-    "plt.title('ROC on IJB-C')\n",
-    "plt.legend(loc=\"lower right\")\n",
-    "plt.show()\n",
-    "#fig.savefig('IJB-B.pdf')"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 39,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "|                  Methods                  | 1e-06  | 1e-05  | 0.0001 | 0.001  |  0.01  |  0.1   |\n",
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F1) | 0.8997 | 0.9434 | 0.9618 | 0.9744 | 0.9832 | 0.9907 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D1F2) | 0.8829 | 0.9400 | 0.9607 | 0.9746 | 0.9833 | 0.9910 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F2) | 0.8985 | 0.9447 | 0.9628 | 0.9753 | 0.9836 | 0.9908 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D0F0) | 0.8906 | 0.9394 | 0.9603 | 0.9731 | 0.9829 | 0.9904 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D0F0) | 0.8625 | 0.9315 | 0.9565 | 0.9720 | 0.9818 | 0.9901 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F0) | 0.8943 | 0.9413 | 0.9610 | 0.9735 | 0.9829 | 0.9905 |\n",
-      "| MS1MV2-ResNet100-ArcFace-TestMode(N0D1F0) | 0.8795 | 0.9387 | 0.9591 | 0.9731 | 0.9824 | 0.9904 |\n",
-      "+-------------------------------------------+--------+--------+--------+--------+--------+--------+\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(tpr_fpr_table)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# setting N1D1F2 is the best"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 2",
-   "language": "python",
-   "name": "python2"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 2
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython2",
-   "version": "2.7.15"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/Evaluation/IJB/IJBC_Evaluation_VGG2.ipynb b/Evaluation/IJB/IJBC_Evaluation_VGG2.ipynb
deleted file mode 100644
index 8f32634e3..000000000
--- a/Evaluation/IJB/IJBC_Evaluation_VGG2.ipynb
+++ /dev/null
@@ -1,532 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/jd4615/miniconda3/envs/insightface/lib/python2.7/site-packages/sklearn/utils/fixes.py:313: FutureWarning: numpy not_equal will not check object identity in the future. The comparison did not return the same result as suggested by the identity (`is`)) and will change.\n",
-      "  _nan_object_mask = _nan_object_array != _nan_object_array\n"
-     ]
-    }
-   ],
-   "source": [
-    "import os\n",
-    "import numpy as np\n",
-    "import cPickle\n",
-    "from sklearn.metrics import roc_curve, auc\n",
-    "import matplotlib.pyplot as plt\n",
-    "import timeit\n",
-    "import sklearn\n",
-    "import cv2\n",
-    "import sys\n",
-    "import glob\n",
-    "sys.path.append('./recognition')\n",
-    "from embedding import Embedding\n",
-    "from menpo.visualize import print_progress\n",
-    "from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap\n",
-    "from prettytable import PrettyTable\n",
-    "from pathlib import Path\n",
-    "import warnings \n",
-    "warnings.filterwarnings(\"ignore\")  "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_media_list(path):\n",
-    "    ijb_meta = np.loadtxt(path, dtype=str)\n",
-    "    templates = ijb_meta[:,1].astype(np.int)\n",
-    "    medias = ijb_meta[:,2].astype(np.int)\n",
-    "    return templates, medias"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_template_pair_list(path):\n",
-    "    pairs = np.loadtxt(path, dtype=str)\n",
-    "    t1 = pairs[:,0].astype(np.int)\n",
-    "    t2 = pairs[:,1].astype(np.int)\n",
-    "    label = pairs[:,2].astype(np.int)\n",
-    "    return t1, t2, label"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_image_feature(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def get_image_feature(img_path, img_list_path, model_path, gpu_id):\n",
-    "    img_list = open(img_list_path)\n",
-    "    embedding = Embedding(model_path, 0, gpu_id)\n",
-    "    files = img_list.readlines()\n",
-    "    img_feats = []\n",
-    "    faceness_scores = []\n",
-    "    for img_index, each_line in enumerate(print_progress(files)):\n",
-    "        name_lmk_score = each_line.strip().split(' ')\n",
-    "        img_name = os.path.join(img_path, name_lmk_score[0])\n",
-    "        img = cv2.imread(img_name)\n",
-    "        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)\n",
-    "        lmk = lmk.reshape( (5,2) )\n",
-    "        img_feats.append(embedding.get(img,lmk))\n",
-    "        faceness_scores.append(name_lmk_score[-1])\n",
-    "    img_feats = np.array(img_feats).astype(np.float32)\n",
-    "    faceness_scores = np.array(faceness_scores).astype(np.float32)\n",
-    "    return img_feats, faceness_scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def image2template_feature(img_feats = None, templates = None, medias = None):\n",
-    "    # ==========================================================\n",
-    "    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]\n",
-    "    # 2. compute media feature.\n",
-    "    # 3. compute template feature.\n",
-    "    # ==========================================================    \n",
-    "    unique_templates = np.unique(templates)\n",
-    "    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))\n",
-    "\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        (ind_t,) = np.where(templates == uqt)\n",
-    "        face_norm_feats = img_feats[ind_t]\n",
-    "        face_medias = medias[ind_t]\n",
-    "        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)\n",
-    "        media_norm_feats = []\n",
-    "        for u,ct in zip(unique_medias, unique_media_counts):\n",
-    "            (ind_m,) = np.where(face_medias == u)\n",
-    "            if ct == 1:\n",
-    "                media_norm_feats += [face_norm_feats[ind_m]]\n",
-    "            else: # image features from the same video will be aggregated into one feature\n",
-    "                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]\n",
-    "        media_norm_feats = np.array(media_norm_feats)\n",
-    "        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))\n",
-    "        template_feats[count_template] = np.sum(media_norm_feats, 0)\n",
-    "        if count_template % 2000 == 0: \n",
-    "            print('Finish Calculating {} template features.'.format(count_template))\n",
-    "    template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))\n",
-    "    return template_norm_feats, unique_templates"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):\n",
-    "    # ==========================================================\n",
-    "    #         Compute set-to-set Similarity Score.\n",
-    "    # ==========================================================\n",
-    "    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)\n",
-    "    for count_template, uqt in enumerate(unique_templates):\n",
-    "        template2id[uqt] = count_template\n",
-    "    \n",
-    "    score = np.zeros((len(p1),))   # save cosine distance between pairs \n",
-    "\n",
-    "    total_pairs = np.array(range(len(p1)))\n",
-    "    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation\n",
-    "    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]\n",
-    "    total_sublists = len(sublists)\n",
-    "    for c, s in enumerate(sublists):\n",
-    "        feat1 = template_norm_feats[template2id[p1[s]]]\n",
-    "        feat2 = template_norm_feats[template2id[p2[s]]]\n",
-    "        similarity_score = np.sum(feat1 * feat2, -1)\n",
-    "        score[s] = similarity_score.flatten()\n",
-    "        if c % 10 == 0:\n",
-    "            print('Finish {}/{} pairs.'.format(c, total_sublists))\n",
-    "    return score"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def read_score(path):\n",
-    "    with open(path, 'rb') as fid:\n",
-    "        img_feats = cPickle.load(fid)\n",
-    "    return img_feats"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step1: Load Meta Data"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 1.76 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image and template relationships for template feature embedding\n",
-    "# tid --> template id,  mid --> media id \n",
-    "# format:\n",
-    "#           image_name tid mid\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "templates, medias = read_template_media_list(os.path.join('IJBC/meta', 'ijbc_face_tid_mid.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Time: 63.31 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load template pairs for template-to-template verification\n",
-    "# tid : template id,  label : 1/0\n",
-    "# format:\n",
-    "#           tid_1 tid_2 label\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "p1, p2, label = read_template_pair_list(os.path.join('IJBC/meta', 'ijbc_template_pair_label.txt'))\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 2: Get Image Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 11,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "('loading', './pretrained_models/VGG2-ResNet50-Arcface/model', 0)\n",
-      "[====================] 100% (469375/469375) - done.                             \n",
-      "Time: 5087.25 s. \n",
-      "Feature Shape: (469375 , 1024) .\n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# load image features \n",
-    "# format:\n",
-    "#           img_feats: [image_num x feats_dim] (227630, 512)\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "#img_feats = read_image_feature('./MS1MV2/IJBB_MS1MV2_r100_arcface.pkl')\n",
-    "img_path = './IJBC/loose_crop'\n",
-    "img_list_path = './IJBC/meta/ijbc_name_5pts_score.txt'\n",
-    "model_path = './pretrained_models/VGG2-ResNet50-Arcface/model'\n",
-    "gpu_id = 0\n",
-    "img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, gpu_id)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))\n",
-    "print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step3: Get Template Features"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 34,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish Calculating 0 template features.\n",
-      "Finish Calculating 2000 template features.\n",
-      "Finish Calculating 4000 template features.\n",
-      "Finish Calculating 6000 template features.\n",
-      "Finish Calculating 8000 template features.\n",
-      "Finish Calculating 10000 template features.\n",
-      "Finish Calculating 12000 template features.\n",
-      "Finish Calculating 14000 template features.\n",
-      "Finish Calculating 16000 template features.\n",
-      "Finish Calculating 18000 template features.\n",
-      "Finish Calculating 20000 template features.\n",
-      "Finish Calculating 22000 template features.\n",
-      "Time: 9.98 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute template features from image features.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "# ========================================================== \n",
-    "# Norm feature before aggregation into template feature?\n",
-    "# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).\n",
-    "# ========================================================== \n",
-    "# 1. FaceScore （Feature Norm）\n",
-    "# 2. FaceScore （Detector）\n",
-    "\n",
-    "use_norm_score = True # if True, TestMode(N1)  \n",
-    "use_detector_score = True # if True, TestMode(D1)\n",
-    "use_flip_test = True # if True, TestMode(F1)\n",
-    "\n",
-    "if use_flip_test:\n",
-    "    # concat --- F1\n",
-    "    img_input_feats = img_feats \n",
-    "    # add --- F2\n",
-    "    # img_input_feats = img_feats[:,0:img_feats.shape[1]/2] + img_feats[:,img_feats.shape[1]/2:]\n",
-    "else:\n",
-    "    img_input_feats = img_feats[:,0:img_feats.shape[1]/2]\n",
-    "    \n",
-    "if use_norm_score:\n",
-    "    img_input_feats = img_input_feats\n",
-    "else:\n",
-    "    # normalise features to remove norm information\n",
-    "    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    \n",
-    "    \n",
-    "if use_detector_score:\n",
-    "    img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])\n",
-    "else:\n",
-    "    img_input_feats = img_input_feats\n",
-    "\n",
-    "template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 4: Get Template Similarity Scores"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 35,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Finish 0/157 pairs.\n",
-      "Finish 10/157 pairs.\n",
-      "Finish 20/157 pairs.\n",
-      "Finish 30/157 pairs.\n",
-      "Finish 40/157 pairs.\n",
-      "Finish 50/157 pairs.\n",
-      "Finish 60/157 pairs.\n",
-      "Finish 70/157 pairs.\n",
-      "Finish 80/157 pairs.\n",
-      "Finish 90/157 pairs.\n",
-      "Finish 100/157 pairs.\n",
-      "Finish 110/157 pairs.\n",
-      "Finish 120/157 pairs.\n",
-      "Finish 130/157 pairs.\n",
-      "Finish 140/157 pairs.\n",
-      "Finish 150/157 pairs.\n",
-      "Time: 146.08 s. \n"
-     ]
-    }
-   ],
-   "source": [
-    "# =============================================================\n",
-    "# compute verification scores between template pairs.\n",
-    "# =============================================================\n",
-    "start = timeit.default_timer()\n",
-    "score = verification(template_norm_feats, unique_templates, p1, p2)\n",
-    "stop = timeit.default_timer()\n",
-    "print('Time: %.2f s. ' % (stop - start))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 36,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "score_save_name = './IJBC/result/VGG2-ResNet50-ArcFace-TestMode(N1D1F1).npy'\n",
-    "np.save(score_save_name, score)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Step 5: Get ROC Curves and TPR@FPR Table"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 39,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "<Figure size 432x288 with 1 Axes>"
-      ]
-     },
-     "metadata": {
-      "needs_background": "light"
-     },
-     "output_type": "display_data"
-    }
-   ],
-   "source": [
-    "score_save_path = './IJBC/result'\n",
-    "files = glob.glob(score_save_path + '/VGG2*.npy')  \n",
-    "methods = []\n",
-    "scores = []\n",
-    "for file in files:\n",
-    "    methods.append(Path(file).stem)\n",
-    "    scores.append(np.load(file)) \n",
-    "methods = np.array(methods)\n",
-    "scores = dict(zip(methods,scores))\n",
-    "colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))\n",
-    "#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]\n",
-    "x_labels = [10**-6, 10**-5, 10**-4,10**-3, 10**-2, 10**-1]\n",
-    "tpr_fpr_table = PrettyTable(['Methods'] + map(str, x_labels))\n",
-    "fig = plt.figure()\n",
-    "for method in methods:\n",
-    "    fpr, tpr, _ = roc_curve(label, scores[method])\n",
-    "    roc_auc = auc(fpr, tpr)\n",
-    "    fpr = np.flipud(fpr)\n",
-    "    tpr = np.flipud(tpr) # select largest tpr at same fpr\n",
-    "    plt.plot(fpr, tpr, color=colours[method], lw=1, label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc*100)))\n",
-    "    tpr_fpr_row = []\n",
-    "    tpr_fpr_row.append(method)\n",
-    "    for fpr_iter in np.arange(len(x_labels)):\n",
-    "        _, min_index = min(list(zip(abs(fpr-x_labels[fpr_iter]), range(len(fpr)))))\n",
-    "        tpr_fpr_row.append('%.4f' % tpr[min_index])\n",
-    "    tpr_fpr_table.add_row(tpr_fpr_row)\n",
-    "plt.xlim([10**-6, 0.1])\n",
-    "plt.ylim([0.3, 1.0])\n",
-    "plt.grid(linestyle='--', linewidth=1)\n",
-    "plt.xticks(x_labels) \n",
-    "plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True)) \n",
-    "plt.xscale('log')\n",
-    "plt.xlabel('False Positive Rate')\n",
-    "plt.ylabel('True Positive Rate')\n",
-    "plt.title('ROC on IJB-C')\n",
-    "plt.legend(loc=\"lower right\")\n",
-    "plt.show()\n",
-    "#fig.savefig('IJB-C.pdf')"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 38,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "|                Methods                 | 1e-06  | 1e-05  | 0.0001 | 0.001  |  0.01  |  0.1   |\n",
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F2) | 0.7444 | 0.8751 | 0.9279 | 0.9635 | 0.9841 | 0.9939 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D0F0) | 0.6863 | 0.8554 | 0.9199 | 0.9586 | 0.9820 | 0.9934 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F1) | 0.7586 | 0.8717 | 0.9253 | 0.9620 | 0.9836 | 0.9937 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D0F0) | 0.7081 | 0.8612 | 0.9214 | 0.9595 | 0.9823 | 0.9934 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N1D1F0) | 0.7470 | 0.8675 | 0.9245 | 0.9610 | 0.9830 | 0.9935 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D1F0) | 0.7637 | 0.8733 | 0.9258 | 0.9617 | 0.9831 | 0.9936 |\n",
-      "| VGG2-ResNet50-ArcFace-TestMode(N0D1F2) | 0.7668 | 0.8796 | 0.9289 | 0.9636 | 0.9840 | 0.9941 |\n",
-      "+----------------------------------------+--------+--------+--------+--------+--------+--------+\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(tpr_fpr_table)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# setting N0D1F2 is the best"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 2",
-   "language": "python",
-   "name": "python2"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 2
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython2",
-   "version": "2.7.15"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/Evaluation/IJB/IJB_11.py b/Evaluation/IJB/IJB_11.py
deleted file mode 100644
index ce3f9fe96..000000000
--- a/Evaluation/IJB/IJB_11.py
+++ /dev/null
@@ -1,347 +0,0 @@
-# coding: utf-8
-
-import os
-import numpy as np
-#import cPickle
-import pickle
-import pandas as pd
-import matplotlib
-matplotlib.use('Agg')
-import matplotlib.pyplot as plt
-import timeit
-import sklearn
-import argparse
-from sklearn.metrics import roc_curve, auc
-from sklearn import preprocessing
-import cv2
-import sys
-import glob
-sys.path.append('./recognition')
-from embedding import Embedding
-from menpo.visualize import print_progress
-from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
-from prettytable import PrettyTable
-from pathlib import Path
-import warnings 
-warnings.filterwarnings("ignore")  
-
-parser = argparse.ArgumentParser(description='do ijb test')
-# general
-parser.add_argument('--model-prefix', default='', help='path to load model.')
-parser.add_argument('--model-epoch', default=1, type=int, help='')
-parser.add_argument('--gpu', default=7, type=int, help='gpu id')
-parser.add_argument('--batch-size', default=32, type=int, help='')
-parser.add_argument('--job', default='insightface', type=str, help='job name')
-parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
-args = parser.parse_args()
-
-target = args.target
-model_path = args.model_prefix
-gpu_id = args.gpu
-epoch = args.model_epoch
-use_norm_score = True # if Ture, TestMode(N1)  
-use_detector_score = True # if Ture, TestMode(D1)
-use_flip_test = True # if Ture, TestMode(F1)
-job = args.job
-
-
-def read_template_media_list(path):
-    #ijb_meta = np.loadtxt(path, dtype=str)
-    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
-    templates = ijb_meta[:,1].astype(np.int)
-    medias = ijb_meta[:,2].astype(np.int)
-    return templates, medias
-
-# In[ ]:
-
-
-def read_template_pair_list(path):
-    #pairs = np.loadtxt(path, dtype=str)
-    pairs = pd.read_csv(path, sep=' ', header=None).values
-    #print(pairs.shape)
-    #print(pairs[:, 0].astype(np.int))
-    t1 = pairs[:,0].astype(np.int)
-    t2 = pairs[:,1].astype(np.int)
-    label = pairs[:,2].astype(np.int)
-    return t1, t2, label
-
-# In[ ]:
-
-
-def read_image_feature(path):
-    with open(path, 'rb') as fid:
-        img_feats = pickle.load(fid)
-    return img_feats
-
-# In[ ]:
-
-
-def get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id):
-    img_list = open(img_list_path)
-    embedding = Embedding(model_path, epoch, gpu_id)
-    files = img_list.readlines()
-    print('files:', len(files))
-    faceness_scores = []
-    img_feats = []
-    for img_index, each_line in enumerate(files):
-        if img_index%500==0:
-          print('processing', img_index)
-        name_lmk_score = each_line.strip().split(' ')
-        img_name = os.path.join(img_path, name_lmk_score[0])
-        img = cv2.imread(img_name)
-        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)
-        lmk = lmk.reshape( (5,2) )
-        img_feats.append(embedding.get(img,lmk))
-        faceness_scores.append(name_lmk_score[-1])
-    img_feats = np.array(img_feats).astype(np.float32)
-    faceness_scores = np.array(faceness_scores).astype(np.float32)
-
-    #img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
-    #faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
-    return img_feats, faceness_scores
-
-# In[ ]:
-
-
-def image2template_feature(img_feats = None, templates = None, medias = None):
-    # ==========================================================
-    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
-    # 2. compute media feature.
-    # 3. compute template feature.
-    # ==========================================================    
-    unique_templates = np.unique(templates)
-    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
-
-    for count_template, uqt in enumerate(unique_templates):
-        (ind_t,) = np.where(templates == uqt)
-        face_norm_feats = img_feats[ind_t]
-        face_medias = medias[ind_t]
-        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
-        media_norm_feats = []
-        for u,ct in zip(unique_medias, unique_media_counts):
-            (ind_m,) = np.where(face_medias == u)
-            if ct == 1:
-                media_norm_feats += [face_norm_feats[ind_m]]
-            else: # image features from the same video will be aggregated into one feature
-                media_norm_feats += [np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)]
-        media_norm_feats = np.array(media_norm_feats)
-        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
-        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
-        if count_template % 2000 == 0: 
-            print('Finish Calculating {} template features.'.format(count_template))
-    #template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
-    template_norm_feats = sklearn.preprocessing.normalize(template_feats)
-    #print(template_norm_feats.shape)
-    return template_norm_feats, unique_templates
-
-# In[ ]:
-
-
-def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):
-    # ==========================================================
-    #         Compute set-to-set Similarity Score.
-    # ==========================================================
-    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)
-    for count_template, uqt in enumerate(unique_templates):
-        template2id[uqt] = count_template
-    
-    score = np.zeros((len(p1),))   # save cosine distance between pairs 
-
-    total_pairs = np.array(range(len(p1)))
-    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation
-    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
-    total_sublists = len(sublists)
-    for c, s in enumerate(sublists):
-        feat1 = template_norm_feats[template2id[p1[s]]]
-        feat2 = template_norm_feats[template2id[p2[s]]]
-        similarity_score = np.sum(feat1 * feat2, -1)
-        score[s] = similarity_score.flatten()
-        if c % 10 == 0:
-            print('Finish {}/{} pairs.'.format(c, total_sublists))
-    return score
-
-# In[ ]:
-def verification2(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):
-  template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
-  for count_template, uqt in enumerate(unique_templates):
-      template2id[uqt] = count_template
-  score = np.zeros((len(p1),))  # save cosine distance between pairs
-  total_pairs = np.array(range(len(p1)))
-  batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
-  sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
-  total_sublists = len(sublists)
-  for c, s in enumerate(sublists):
-      feat1 = template_norm_feats[template2id[p1[s]]]
-      feat2 = template_norm_feats[template2id[p2[s]]]
-      similarity_score = np.sum(feat1 * feat2, -1)
-      score[s] = similarity_score.flatten()
-      if c % 10 == 0:
-          print('Finish {}/{} pairs.'.format(c, total_sublists))
-  return score
-
-
-def read_score(path):
-    with open(path, 'rb') as fid:
-        img_feats = pickle.load(fid)
-    return img_feats
-
-# # Step1: Load Meta Data
-
-# In[ ]:
-
-assert target=='IJBC' or target=='IJBB'
-
-# =============================================================
-# load image and template relationships for template feature embedding
-# tid --> template id,  mid --> media id 
-# format:
-#           image_name tid mid
-# =============================================================
-start = timeit.default_timer()
-templates, medias = read_template_media_list(os.path.join('%s/meta'%target, '%s_face_tid_mid.txt'%target.lower()))
-stop = timeit.default_timer()
-print('Time: %.2f s. ' % (stop - start))
-
-# In[ ]:
-
-
-# =============================================================
-# load template pairs for template-to-template verification
-# tid : template id,  label : 1/0
-# format:
-#           tid_1 tid_2 label
-# =============================================================
-start = timeit.default_timer()
-p1, p2, label = read_template_pair_list(os.path.join('%s/meta'%target, '%s_template_pair_label.txt'%target.lower()))
-stop = timeit.default_timer()
-print('Time: %.2f s. ' % (stop - start))
-
-# # Step 2: Get Image Features
-
-# In[ ]:
-
-
-# =============================================================
-# load image features 
-# format:
-#           img_feats: [image_num x feats_dim] (227630, 512)
-# =============================================================
-start = timeit.default_timer()
-img_path = './%s/loose_crop' % target
-img_list_path = './%s/meta/%s_name_5pts_score.txt' % (target, target.lower())
-img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id)
-stop = timeit.default_timer()
-print('Time: %.2f s. ' % (stop - start))
-print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
-
-# # Step3: Get Template Features
-
-# In[ ]:
-
-
-# =============================================================
-# compute template features from image features.
-# =============================================================
-start = timeit.default_timer()
-# ========================================================== 
-# Norm feature before aggregation into template feature?
-# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
-# ========================================================== 
-# 1. FaceScore （Feature Norm）
-# 2. FaceScore （Detector）
-
-
-if use_flip_test:
-    # concat --- F1
-    #img_input_feats = img_feats 
-    # add --- F2
-    img_input_feats = img_feats[:,0:img_feats.shape[1]//2] + img_feats[:,img_feats.shape[1]//2:]
-else:
-    img_input_feats = img_feats[:,0:img_feats.shape[1]//2]
-    
-if use_norm_score:
-    img_input_feats = img_input_feats
-else:
-    # normalise features to remove norm information
-    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    
-    
-if use_detector_score:
-    print(img_input_feats.shape, faceness_scores.shape)
-    #img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])
-    img_input_feats = img_input_feats * faceness_scores[:,np.newaxis]
-else:
-    img_input_feats = img_input_feats
-
-template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)
-stop = timeit.default_timer()
-print('Time: %.2f s. ' % (stop - start))
-
-# # Step 4: Get Template Similarity Scores
-
-# In[ ]:
-
-
-# =============================================================
-# compute verification scores between template pairs.
-# =============================================================
-start = timeit.default_timer()
-score = verification(template_norm_feats, unique_templates, p1, p2)
-stop = timeit.default_timer()
-print('Time: %.2f s. ' % (stop - start))
-
-# In[ ]:
-
-save_path = './%s_result' % target
-
-if not os.path.exists(save_path):
-    os.makedirs(save_path)
-
-score_save_file = os.path.join(save_path, "%s.npy"%job)
-np.save(score_save_file, score)
-
-# # Step 5: Get ROC Curves and TPR@FPR Table
-
-# In[ ]:
-
-
-files = [score_save_file]
-methods = []
-scores = []
-for file in files:
-    methods.append(Path(file).stem)
-    scores.append(np.load(file)) 
-
-methods = np.array(methods)
-scores = dict(zip(methods,scores))
-colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
-#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]
-x_labels = [10**-6, 10**-5, 10**-4,10**-3, 10**-2, 10**-1]
-tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
-fig = plt.figure()
-for method in methods:
-    fpr, tpr, _ = roc_curve(label, scores[method])
-    roc_auc = auc(fpr, tpr)
-    fpr = np.flipud(fpr)
-    tpr = np.flipud(tpr) # select largest tpr at same fpr
-    plt.plot(fpr, tpr, color=colours[method], lw=1, label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc*100)))
-    tpr_fpr_row = []
-    tpr_fpr_row.append("%s-%s"%(method, target))
-    for fpr_iter in np.arange(len(x_labels)):
-        _, min_index = min(list(zip(abs(fpr-x_labels[fpr_iter]), range(len(fpr)))))
-        #tpr_fpr_row.append('%.4f' % tpr[min_index])
-        tpr_fpr_row.append('%.2f' % (tpr[min_index]*100))
-    tpr_fpr_table.add_row(tpr_fpr_row)
-plt.xlim([10**-6, 0.1])
-plt.ylim([0.3, 1.0])
-plt.grid(linestyle='--', linewidth=1)
-plt.xticks(x_labels) 
-plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True)) 
-plt.xscale('log')
-plt.xlabel('False Positive Rate')
-plt.ylabel('True Positive Rate')
-plt.title('ROC on IJB')
-plt.legend(loc="lower right")
-#plt.show()
-fig.savefig(os.path.join(save_path, '%s.pdf'%job))
-print(tpr_fpr_table)
-
diff --git a/Evaluation/IJB/IJB_1N.py b/Evaluation/IJB/IJB_1N.py
deleted file mode 100644
index e42b49ba9..000000000
--- a/Evaluation/IJB/IJB_1N.py
+++ /dev/null
@@ -1,314 +0,0 @@
-#!/usr/bin/env python
-# coding: utf-8
-import os
-import numpy as np
-import timeit
-import sklearn
-import cv2
-import sys
-import argparse
-import glob
-import numpy.matlib
-import heapq
-import math
-from datetime import datetime as dt
-
-from sklearn import preprocessing
-sys.path.append('./recognition')
-from embedding import Embedding
-from menpo.visualize import print_progress
-from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
-
-def read_template_subject_id_list(path):
-    ijb_meta = np.loadtxt(path, dtype=str, skiprows=1, delimiter=',')
-    templates = ijb_meta[:, 0].astype(np.int)
-    subject_ids = ijb_meta[:, 1].astype(np.int)
-    return templates, subject_ids
-
-def read_template_media_list(path):
-    ijb_meta = np.loadtxt(path, dtype=str)
-    templates = ijb_meta[:,1].astype(np.int)
-    medias = ijb_meta[:,2].astype(np.int)
-    return templates, medias
-
-def read_template_pair_list(path):
-    pairs = np.loadtxt(path, dtype=str)
-    t1 = pairs[:,0].astype(np.int)
-    t2 = pairs[:,1].astype(np.int)
-    label = pairs[:,2].astype(np.int)
-    return t1, t2, label
-
-
-#def get_image_feature(feature_path, faceness_path):
-#    img_feats = np.loadtxt(feature_path)
-#    faceness_scores = np.loadtxt(faceness_path)
-#    return img_feats, faceness_scores
-def get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id):
-    img_list = open(img_list_path)
-    embedding = Embedding(model_path, epoch, gpu_id)
-    files = img_list.readlines()
-    print('files:', len(files))
-    faceness_scores = []
-    img_feats = []
-    for img_index, each_line in enumerate(files):
-        if img_index%500==0:
-          print('processing', img_index)
-        name_lmk_score = each_line.strip().split(' ')
-        img_name = os.path.join(img_path, name_lmk_score[0])
-        img = cv2.imread(img_name)
-        lmk = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)
-        lmk = lmk.reshape( (5,2) )
-        img_feats.append(embedding.get(img,lmk))
-        faceness_scores.append(name_lmk_score[-1])
-    img_feats = np.array(img_feats).astype(np.float32)
-    faceness_scores = np.array(faceness_scores).astype(np.float32)
-
-    #img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
-    #faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
-    return img_feats, faceness_scores
-
-def image2template_feature(img_feats = None, templates = None, medias = None, choose_templates = None, choose_ids = None):
-    # ==========================================================
-    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
-    # 2. compute media feature.
-    # 3. compute template feature.
-    # ==========================================================    
-    unique_templates, indices = np.unique(choose_templates, return_index=True)
-    unique_subjectids = choose_ids[indices]
-    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
-
-    for count_template, uqt in enumerate(unique_templates):
-        (ind_t,) = np.where(templates == uqt)
-        face_norm_feats = img_feats[ind_t]
-        face_medias = medias[ind_t]
-        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
-        media_norm_feats = []
-        for u,ct in zip(unique_medias, unique_media_counts):
-            (ind_m,) = np.where(face_medias == u)
-            if ct == 1:
-                media_norm_feats += [face_norm_feats[ind_m]]
-            else: # image features from the same video will be aggregated into one feature
-                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]
-        media_norm_feats = np.array(media_norm_feats)
-        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
-        template_feats[count_template] = np.sum(media_norm_feats, 0)
-        if count_template % 2000 == 0: 
-            print('Finish Calculating {} template features.'.format(count_template))
-    template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
-    return template_norm_feats, unique_templates, unique_subjectids
-
-def verification(template_norm_feats = None, unique_templates = None, p1 = None, p2 = None):
-    # ==========================================================
-    #         Compute set-to-set Similarity Score.
-    # ==========================================================
-    template2id = np.zeros((max(unique_templates)+1,1),dtype=int)
-    for count_template, uqt in enumerate(unique_templates):
-        template2id[uqt] = count_template
-    
-    score = np.zeros((len(p1),))   # save cosine distance between pairs 
-
-    total_pairs = np.array(range(len(p1)))
-    batchsize = 100000 # small batchsize instead of all pairs in one batch due to the memory limiation
-    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
-    total_sublists = len(sublists)
-    for c, s in enumerate(sublists):
-        feat1 = template_norm_feats[template2id[p1[s]]]
-        feat2 = template_norm_feats[template2id[p2[s]]]
-        similarity_score = np.sum(feat1 * feat2, -1)
-        score[s] = similarity_score.flatten()
-        if c % 10 == 0:
-            print('Finish {}/{} pairs.'.format(c, total_sublists))
-    return score
-
-def read_score(path):
-    with open(path, 'rb') as fid:
-        img_feats = cPickle.load(fid)
-    return img_feats
-
-def evaluation(query_feats, gallery_feats, mask):
-    Fars = [0.01, 0.1]
-    print(query_feats.shape)
-    print(gallery_feats.shape)
-
-    query_num = query_feats.shape[0]
-    gallery_num = gallery_feats.shape[0]
-
-    similarity = np.dot(query_feats, gallery_feats.T)
-    print('similarity shape', similarity.shape)
-    top_inds = np.argsort(-similarity)
-    print(top_inds.shape)
-
-    # calculate top1 
-    correct_num = 0
-    for i in range(query_num):
-        j = top_inds[i, 0]
-        if j == mask[i]:
-            correct_num += 1
-    print("top1 = {}".format(correct_num/query_num))
-    # calculate top5
-    correct_num = 0
-    for i in range(query_num):
-        j = top_inds[i, 0:5]
-        if mask[i] in j:
-            correct_num += 1
-    print("top5 = {}".format(correct_num/query_num))
-    # calculate 10
-    correct_num = 0
-    for i in range(query_num):
-        j = top_inds[i, 0:10]
-        if mask[i] in j:
-            correct_num += 1
-    print("top10 = {}".format(correct_num/query_num))
-    
-    neg_pair_num = query_num * gallery_num - query_num
-    print(neg_pair_num)
-    required_topk = [math.ceil(query_num * x) for x in Fars]
-    top_sims = similarity
-    # calculate fars and tprs
-    pos_sims = []
-    for i in range(query_num):
-        gt = mask[i]
-        pos_sims.append(top_sims[i, gt])
-        top_sims[i, gt] = -2.0
-    
-    pos_sims = np.array(pos_sims)
-    print(pos_sims.shape)
-    neg_sims = top_sims[np.where(top_sims > -2.0)]
-    print("neg_sims num = {}".format(len(neg_sims)))
-    neg_sims = heapq.nlargest(max(required_topk), neg_sims)  # heap sort
-    print("after sorting , neg_sims num = {}".format(len(neg_sims)))
-    for far, pos in zip(Fars, required_topk):
-        th = neg_sims[pos-1]
-        recall = np.sum(pos_sims > th) / query_num
-        print("far = {:.10f} pr = {:.10f} th = {:.10f}".format(far, recall, th))
-    
-def gen_mask(query_ids, reg_ids):
-    mask = []
-    for query_id in query_ids:
-        pos = [i for i, x in enumerate(reg_ids) if query_id == x]
-        if len(pos) != 1:
-            raise RuntimeError("RegIdsError with id = {}， duplicate = {} ".format(query_id, len(pos)))
-        mask.append(pos[0])
-    return mask
-
-if __name__ == "__main__":
-   parser = argparse.ArgumentParser(description='do ijb 1n test')
-   # general
-   parser.add_argument('--model-prefix', default='', help='path to load model.')
-   parser.add_argument('--model-epoch', default=1, type=int, help='')
-   parser.add_argument('--gpu', default=7, type=int, help='gpu id')
-   parser.add_argument('--batch-size', default=32, type=int, help='')
-   parser.add_argument('--job', default='insightface', type=str, help='job name')
-   parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
-   args = parser.parse_args()
-   target = args.target
-   model_path = args.model_prefix
-   gpu_id = args.gpu
-   epoch = args.model_epoch
-   meta_dir = "%s/meta"%args.target #meta root dir
-   if target=='IJBC':
-       gallery_s1_record = "%s_1N_gallery_G1.csv" % (args.target.lower())
-       gallery_s2_record = "%s_1N_gallery_G2.csv" % (args.target.lower())
-   else:
-       gallery_s1_record = "%s_1N_gallery_S1.csv" % (args.target.lower())
-       gallery_s2_record = "%s_1N_gallery_S2.csv" % (args.target.lower())
-   gallery_s1_templates, gallery_s1_subject_ids = read_template_subject_id_list(os.path.join(meta_dir, gallery_s1_record))
-   print(gallery_s1_templates.shape, gallery_s1_subject_ids.shape)
-
-   gallery_s2_templates, gallery_s2_subject_ids = read_template_subject_id_list(os.path.join(meta_dir, gallery_s2_record))
-   print(gallery_s2_templates.shape, gallery_s2_templates.shape)
-   
-   gallery_templates = np.concatenate([gallery_s1_templates, gallery_s2_templates])
-   gallery_subject_ids = np.concatenate([gallery_s1_subject_ids, gallery_s2_subject_ids])
-   print(gallery_templates.shape, gallery_subject_ids.shape)
-   
-   media_record = "%s_face_tid_mid.txt" % args.target.lower()
-   total_templates, total_medias = read_template_media_list(os.path.join(meta_dir, media_record))
-   print("total_templates", total_templates.shape, total_medias.shape)
-   #load image features
-   start = timeit.default_timer()
-   feature_path = '' #feature path
-   face_path = '' #face path
-   img_path = './%s/loose_crop' % target
-   img_list_path = './%s/meta/%s_name_5pts_score.txt' % (target, target.lower())
-   #img_feats, faceness_scores = get_image_feature(feature_path, face_path)
-   img_feats, faceness_scores = get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id)
-   print('img_feats', img_feats.shape)
-   print('faceness_scores', faceness_scores.shape)
-   stop = timeit.default_timer()
-   print('Time: %.2f s. ' % (stop - start))
-   print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
-   
-   # compute template features from image features.
-   start = timeit.default_timer()
-   # ========================================================== 
-   # Norm feature before aggregation into template feature?
-   # Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
-   # ========================================================== 
-   use_norm_score = True # if True, TestMode(N1)  
-   use_detector_score = True # if True, TestMode(D1)
-   use_flip_test = True # if True, TestMode(F1)
-   
-   if use_flip_test:
-       # concat --- F1
-       #img_input_feats = img_feats 
-       # add --- F2
-       img_input_feats = img_feats[:,0:int(img_feats.shape[1]/2)] + img_feats[:,int(img_feats.shape[1]/2):]
-   else:
-       img_input_feats = img_feats[:,0:int(img_feats.shape[1]/2)]
-       
-   if use_norm_score:
-       img_input_feats = img_input_feats
-   else:
-       # normalise features to remove norm information
-       img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))    
-       
-   if use_detector_score:
-       img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])
-   else:
-       img_input_feats = img_input_feats
-   print("input features shape", img_input_feats.shape)
-   
-   #load gallery feature 
-   gallery_templates_feature, gallery_unique_templates, gallery_unique_subject_ids = image2template_feature(img_input_feats, total_templates, total_medias, gallery_templates, gallery_subject_ids)
-   stop = timeit.default_timer()
-   print('Time: %.2f s. ' % (stop - start))
-   print("gallery_templates_feature", gallery_templates_feature.shape)
-   print("gallery_unique_subject_ids", gallery_unique_subject_ids.shape)
-   #np.savetxt("gallery_templates_feature.txt", gallery_templates_feature)
-   #np.savetxt("gallery_unique_subject_ids.txt", gallery_unique_subject_ids)
-
-   #load prope feature 
-   probe_mixed_record = "%s_1N_probe_mixed.csv" % target.lower()
-   probe_mixed_templates, probe_mixed_subject_ids = read_template_subject_id_list(os.path.join(meta_dir, probe_mixed_record))
-   print(probe_mixed_templates.shape, probe_mixed_subject_ids.shape)
-   probe_mixed_templates_feature, probe_mixed_unique_templates, probe_mixed_unique_subject_ids = image2template_feature(img_input_feats, total_templates, total_medias, probe_mixed_templates, probe_mixed_subject_ids)
-   print("probe_mixed_templates_feature", probe_mixed_templates_feature.shape)
-   print("probe_mixed_unique_subject_ids", probe_mixed_unique_subject_ids.shape)
-   #np.savetxt("probe_mixed_templates_feature.txt", probe_mixed_templates_feature)
-   #np.savetxt("probe_mixed_unique_subject_ids.txt", probe_mixed_unique_subject_ids)
-
-   #root_dir = "" #feature root dir
-   #gallery_id_path = "" #id filepath
-   #gallery_feats_path = "" #feature filelpath
-   #print("{}: start loading gallery feat {}".format(dt.now(), gallery_id_path))
-   #gallery_ids, gallery_feats = load_feat_file(root_dir, gallery_id_path, gallery_feats_path)
-   #print("{}: end loading gallery feat".format(dt.now()))
-   #
-   #probe_id_path = "probe_mixed_unique_subject_ids.txt" #probe id filepath
-   #probe_feats_path = "probe_mixed_templates_feature.txt" #probe feats filepath
-   #print("{}: start loading probe feat {}".format(dt.now(), probe_id_path))
-   #probe_ids, probe_feats = load_feat_file(root_dir, probe_id_path, probe_feats_path)
-   #print("{}: end loading probe feat".format(dt.now()))
-
-   gallery_ids = gallery_unique_subject_ids
-   gallery_feats = gallery_templates_feature
-   probe_ids = probe_mixed_unique_subject_ids
-   probe_feats = probe_mixed_templates_feature
-
-   mask = gen_mask(probe_ids, gallery_ids)
-   
-   print("{}: start evaluation".format(dt.now()))
-   evaluation(probe_feats, gallery_feats, mask)
-   print("{}: end evaluation".format(dt.now()))
-
diff --git a/Evaluation/IJB/README.md b/Evaluation/IJB/README.md
deleted file mode 100644
index 92ed71e92..000000000
--- a/Evaluation/IJB/README.md
+++ /dev/null
@@ -1,36 +0,0 @@
-To reproduce the figures and tables in the notebook, please download everything (model, code, data and meta info) from here:
-[Dropbox](https://www.dropbox.com/s/33a6haw7v79e5qe/IJB_release.tar?dl=0)
-or
-[Baidu Cloud](https://pan.baidu.com/s/1oer0p4_mcOrs4cfdeWfbFg)
-
-Updated Meta data (1:1 and 1:N):
-[IJB-B Dropbox](https://www.dropbox.com/s/5n2ehrsucmu7vsd/IJBB_meta.tar?dl=0) 
-and
-[IJB-C Dropbox](https://www.dropbox.com/s/pgju50f2gcgqkc2/IJBC_meta.tar?dl=0)
-
-Please apply for the IJB-B and IJB-C by yourself and strictly follow their distribution licenses.
-
-## Aknowledgement
-
-Great thanks for Weidi Xie's instruction [2,3,4,5] to evaluate ArcFace [1] on IJB-B[6] and IJB-C[7] (1:1 protocol).
-
-Great thanks for Yuge Huang's code [8] to evaluate ArcFace [1] on IJB-B[6] and IJB-C[7] (1:N protocol). 
-
-## Reference
-
-[1] Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition[J]. arXiv:1801.07698, 2018.
-
-[2] https://github.com/ox-vgg/vgg_face2.
-
-[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. FG, 2018.
-
-[4] Weidi Xie, Andrew Zisserman. Multicolumn Networks for Face Recognition. BMVC 2018.
-
-[5] Weidi Xie, Li Shen, Andrew Zisserman. Comparator Networks. ECCV, 2018.
-
-[6] Whitelam, Cameron, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn C. Adams, Tim Miller, Nathan D. Kalka et al. IARPA Janus Benchmark-B Face Dataset. CVPR Workshops, 2017.
-
-[7] Maze, Brianna, Jocelyn Adams, James A. Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K. Jain et al. IARPA Janus Benchmark–C: Face Dataset and Protocol. ICB, 2018.
-
-[8] Yuge Huang, Pengcheng Shen, Ying Tai, Shaoxin Li, Xiaoming Liu, Jilin Li, Feiyue Huang, Rongrong Ji. Distribution Distillation Loss: Generic Approach for Improving Face Recognition from Hard Samples. arXiv:2002.03662.
-
diff --git a/Evaluation/IJB/example.sh b/Evaluation/IJB/example.sh
deleted file mode 100755
index 9a386c2bf..000000000
--- a/Evaluation/IJB/example.sh
+++ /dev/null
@@ -1,6 +0,0 @@
-#!/usr/bin/env bash
-
-python -u IJB_11.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 1 --gpu 0 --target IJBC --job arcface > ijbc_11.log 2>&1 &
-
-python -u IJB_1N.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 1 --gpu 0 --target IJBB --job arcface > ijbb_1n.log 2>&1 &
-
diff --git a/Evaluation/IJB/readme.txt b/Evaluation/IJB/readme.txt
deleted file mode 100644
index f6027e12a..000000000
--- a/Evaluation/IJB/readme.txt
+++ /dev/null
@@ -1,19 +0,0 @@
-To reproduce the figures and tables in the notebook, please download everything (model, code, data and meta info) from here:
-[Dropbox] https://www.dropbox.com/s/33a6haw7v79e5qe/IJB_release.tar?dl=0
-or
-[Baidu Cloud] https://pan.baidu.com/s/1oer0p4_mcOrs4cfdeWfbFg
-
-Please apply for the IJB-B and IJB-C by yourself and strictly follow their distribution licenses.
-
-Aknowledgement
-Great thanks for Weidi Xie's instruction [2,3,4,5] to evaluate ArcFace [1] on IJB-B[6] and IJB-C[7].
-
-[1] Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition[J]. arXiv:1801.07698, 2018.
-[2] https://github.com/ox-vgg/vgg_face2.
-[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. FG, 2018.
-[4] Weidi Xie, Andrew Zisserman. Multicolumn Networks for Face Recognition. BMVC 2018.
-[5] Weidi Xie, Li Shen, Andrew Zisserman. Comparator Networks. ECCV, 2018.
-[6] Whitelam, Cameron, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn C. Adams, Tim Miller, Nathan D. Kalka et al. IARPA Janus Benchmark-B Face Dataset. CVPR Workshops, 2017.
-[7] Maze, Brianna, Jocelyn Adams, James A. Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K. Jain et al. IARPA Janus Benchmark–C: Face Dataset and Protocol. ICB, 2018.
- 
-
diff --git a/Evaluation/Megaface/README.md b/Evaluation/Megaface/README.md
deleted file mode 100644
index 9ea432e76..000000000
--- a/Evaluation/Megaface/README.md
+++ /dev/null
@@ -1,4 +0,0 @@
-
-Download megaface testpack data from [baiducloud](https://pan.baidu.com/s/1h4ezfwJiXClbZDdg1RX0MQ) or [dropbox](https://www.dropbox.com/s/5ko2dcd1x7vn37w/megaface_testpack_v1.0.zip?dl=0) and unzip it to ``data/``, then check and call ``run.sh`` to evaluate insightface model performance.
-
-[Devkit](https://www.dropbox.com/s/zv1ctjxmghyy7ow/devkit.tar?dl=0)
diff --git a/Evaluation/Megaface/gen_megaface.py b/Evaluation/Megaface/gen_megaface.py
deleted file mode 100644
index 6ec13337d..000000000
--- a/Evaluation/Megaface/gen_megaface.py
+++ /dev/null
@@ -1,179 +0,0 @@
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-
-from easydict import EasyDict as edict
-import time
-import sys
-import numpy as np
-import argparse
-import struct
-import cv2
-import sklearn
-from sklearn.preprocessing import normalize
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-
-
-def read_img(image_path):
-  img = cv2.imread(image_path, cv2.CV_LOAD_IMAGE_COLOR)
-  return img
-
-def get_feature(imgs, nets):
-  count = len(imgs)
-  data = mx.nd.zeros(shape = (count*2, 3, imgs[0].shape[0], imgs[0].shape[1]))
-  for idx, img in enumerate(imgs):
-    img = img[:,:,::-1] #to rgb
-    img = np.transpose( img, (2,0,1) )
-    for flipid in [0,1]:
-      _img = np.copy(img)
-      if flipid==1:
-        _img = _img[:,:,::-1]
-      _img = nd.array(_img)
-      data[count*flipid+idx] = _img
-
-  F = []
-  for net in nets:
-    db = mx.io.DataBatch(data=(data,))
-    net.model.forward(db, is_train=False)
-    x = net.model.get_outputs()[0].asnumpy()
-    embedding = x[0:count,:] + x[count:,:]
-    embedding = sklearn.preprocessing.normalize(embedding)
-    #print('emb', embedding.shape)
-    F.append(embedding)
-  F = np.concatenate(F, axis=1)
-  F = sklearn.preprocessing.normalize(F)
-  #print('F', F.shape)
-  return F
-
-
-def write_bin(path, feature):
-  feature = list(feature)
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', len(feature),1,4,5))
-    f.write(struct.pack("%df"%len(feature), *feature))
-
-def get_and_write(buffer, nets):
-  imgs = []
-  for k in buffer:
-    imgs.append(k[0])
-  features = get_feature(imgs, nets)
-  #print(np.linalg.norm(feature))
-  assert features.shape[0]==len(buffer)
-  for ik,k in enumerate(buffer):
-    out_path = k[1]
-    feature = features[ik].flatten()
-    write_bin(out_path, feature)
-
-def main(args):
-
-  print(args)
-  gpuid = args.gpu
-  ctx = mx.gpu(gpuid)
-  nets = []
-  image_shape = [int(x) for x in args.image_size.split(',')]
-  for model in args.model.split('|'):
-    vec = model.split(',')
-    assert len(vec)>1
-    prefix = vec[0]
-    epoch = int(vec[1])
-    print('loading',prefix, epoch)
-    net = edict()
-    net.ctx = ctx
-    net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(prefix, epoch)
-    all_layers = net.sym.get_internals()
-    net.sym = all_layers['fc1_output']
-    net.model = mx.mod.Module(symbol=net.sym, context=net.ctx, label_names = None)
-    net.model.bind(data_shapes=[('data', (1, 3, image_shape[1], image_shape[2]))])
-    net.model.set_params(net.arg_params, net.aux_params)
-    nets.append(net)
-
-  facescrub_out = os.path.join(args.output, 'facescrub')
-  megaface_out = os.path.join(args.output, 'megaface')
-
-  i = 0
-  succ = 0
-  buffer = []
-  for line in open(args.facescrub_lst, 'r'):
-    if i%1000==0:
-      print("writing fs",i, succ)
-    i+=1
-    image_path = line.strip()
-    _path = image_path.split('/')
-    a,b = _path[-2], _path[-1]
-    out_dir = os.path.join(facescrub_out, a)
-    if not os.path.exists(out_dir):
-      os.makedirs(out_dir)
-    image_path = os.path.join(args.facescrub_root, image_path)
-    img = read_img(image_path)
-    if img is None:
-      print('read error:', image_path)
-      continue
-    out_path = os.path.join(out_dir, b+"_%s.bin"%(args.algo))
-    item = (img, out_path)
-    buffer.append(item)
-    if len(buffer)==args.batch_size:
-      get_and_write(buffer, nets)
-      buffer = []
-    succ+=1
-  if len(buffer)>0:
-    get_and_write(buffer, nets)
-    buffer = []
-  print('fs stat',i, succ)
-
-  i = 0
-  succ = 0
-  buffer = []
-  for line in open(args.megaface_lst, 'r'):
-    if i%1000==0:
-      print("writing mf",i, succ)
-    i+=1
-    image_path = line.strip()
-    _path = image_path.split('/')
-    a1, a2, b = _path[-3], _path[-2], _path[-1]
-    out_dir = os.path.join(megaface_out, a1, a2)
-    if not os.path.exists(out_dir):
-      os.makedirs(out_dir)
-      #continue
-    #print(landmark)
-    image_path = os.path.join(args.megaface_root, image_path)
-    img = read_img(image_path)
-    if img is None:
-      print('read error:', image_path)
-      continue
-    out_path = os.path.join(out_dir, b+"_%s.bin"%(args.algo))
-    item = (img, out_path)
-    buffer.append(item)
-    if len(buffer)==args.batch_size:
-      get_and_write(buffer, nets)
-      buffer = []
-    succ+=1
-  if len(buffer)>0:
-    get_and_write(buffer, nets)
-    buffer = []
-  print('mf stat',i, succ)
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--batch_size', type=int, help='', default=8)
-  parser.add_argument('--image_size', type=str, help='', default='3,112,112')
-  parser.add_argument('--gpu', type=int, help='', default=0)
-  parser.add_argument('--algo', type=str, help='', default='insightface')
-  parser.add_argument('--facescrub-lst', type=str, help='', default='./data/facescrub_lst')
-  parser.add_argument('--megaface-lst', type=str, help='', default='./data/megaface_lst')
-  parser.add_argument('--facescrub-root', type=str, help='', default='./data/facescrub_images')
-  parser.add_argument('--megaface-root', type=str, help='', default='./data/megaface_images')
-  parser.add_argument('--output', type=str, help='', default='./feature_out')
-  parser.add_argument('--model', type=str, help='', default='')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
diff --git a/Evaluation/Megaface/remove_noises.py b/Evaluation/Megaface/remove_noises.py
deleted file mode 100644
index f80f350f9..000000000
--- a/Evaluation/Megaface/remove_noises.py
+++ /dev/null
@@ -1,156 +0,0 @@
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import datetime
-import time
-import shutil
-import sys
-import numpy as np
-import argparse
-import struct
-import cv2
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-feature_dim = 512
-feature_ext = 1
-
-def load_bin(path, fill = 0.0):
-  with open(path, 'rb') as f:
-    bb = f.read(4*4)
-    #print(len(bb))
-    v = struct.unpack('4i', bb)
-    #print(v[0])
-    bb = f.read(v[0]*4)
-    v = struct.unpack("%df"%(v[0]), bb)
-    feature = np.full( (feature_dim+feature_ext,), fill, dtype=np.float32)
-    feature[0:feature_dim] = v
-    #feature = np.array( v, dtype=np.float32)
-  #print(feature.shape)
-  #print(np.linalg.norm(feature))
-  return feature
-
-def write_bin(path, feature):
-  feature = list(feature)
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', len(feature),1,4,5))
-    f.write(struct.pack("%df"%len(feature), *feature))
-
-def main(args):
-
-
-  fs_noise_map = {}
-  for line in open(args.facescrub_noises, 'r'):
-    if line.startswith('#'):
-      continue
-    line = line.strip()
-    fname = line.split('.')[0]
-    p = fname.rfind('_')
-    fname = fname[0:p]
-    fs_noise_map[line] = fname
-
-  print(len(fs_noise_map))
-
-  i=0
-  fname2center = {}
-  noises = []
-  for line in open(args.facescrub_lst, 'r'):
-    if i%1000==0:
-      print("reading fs",i)
-    i+=1
-    image_path = line.strip()
-    _path = image_path.split('/')
-    a, b = _path[-2], _path[-1]
-    feature_path = os.path.join(args.feature_dir_input, 'facescrub', a, "%s_%s.bin"%(b, args.algo))
-    feature_dir_out = os.path.join(args.feature_dir_out, 'facescrub', a)
-    if not os.path.exists(feature_dir_out):
-      os.makedirs(feature_dir_out)
-    feature_path_out = os.path.join(feature_dir_out, "%s_%s.bin"%(b, args.algo))
-    #print(b)
-    if not b in fs_noise_map:
-      #shutil.copyfile(feature_path, feature_path_out)
-      feature = load_bin(feature_path)
-      write_bin(feature_path_out, feature)
-      if not a in fname2center:
-        fname2center[a] = np.zeros((feature_dim+feature_ext,), dtype=np.float32)
-      fname2center[a] += feature
-    else:
-      #print('n', b)
-      noises.append( (a,b) )
-  print(len(noises))
-
-  for k in noises:
-    a,b = k
-    assert a in fname2center
-    center = fname2center[a]
-    g = np.zeros( (feature_dim+feature_ext,), dtype=np.float32)
-    g2 = np.random.uniform(-0.001, 0.001, (feature_dim,))
-    g[0:feature_dim] = g2
-    f = center+g
-    _norm=np.linalg.norm(f)
-    f /= _norm
-    feature_path_out = os.path.join(args.feature_dir_out, 'facescrub', a, "%s_%s.bin"%(b, args.algo))
-    write_bin(feature_path_out, f)
-
-  mf_noise_map = {}
-  for line in open(args.megaface_noises, 'r'):
-    if line.startswith('#'):
-      continue
-    line = line.strip()
-    _vec = line.split("\t")
-    if len(_vec)>1:
-      line = _vec[1]
-    mf_noise_map[line] = 1
-
-  print(len(mf_noise_map))
-
-  i=0
-  nrof_noises = 0
-  for line in open(args.megaface_lst, 'r'):
-    if i%1000==0:
-      print("reading mf",i)
-    i+=1
-    image_path = line.strip()
-    _path = image_path.split('/')
-    a1, a2, b = _path[-3], _path[-2], _path[-1]
-    feature_path = os.path.join(args.feature_dir_input, 'megaface', a1, a2, "%s_%s.bin"%(b, args.algo))
-    feature_dir_out = os.path.join(args.feature_dir_out, 'megaface', a1, a2)
-    if not os.path.exists(feature_dir_out):
-      os.makedirs(feature_dir_out)
-    feature_path_out = os.path.join(feature_dir_out, "%s_%s.bin"%(b, args.algo))
-    bb = '/'.join([a1, a2, b])
-    #print(b)
-    if not bb in mf_noise_map:
-      feature = load_bin(feature_path)
-      write_bin(feature_path_out, feature)
-      #shutil.copyfile(feature_path, feature_path_out)
-    else:
-      feature = load_bin(feature_path, 100.0)
-      write_bin(feature_path_out, feature)
-      #g = np.random.uniform(-0.001, 0.001, (feature_dim,))
-      #print('n', bb)
-      #write_bin(feature_path_out, g)
-      nrof_noises+=1
-  print(nrof_noises)
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--facescrub-noises', type=str, help='', default='./data/facescrub_noises.txt')
-  parser.add_argument('--megaface-noises', type=str, help='', default='./data/megaface_noises.txt')
-  parser.add_argument('--algo', type=str, help='', default='insightface')
-  parser.add_argument('--facescrub-lst', type=str, help='', default='./data/facescrub_lst')
-  parser.add_argument('--megaface-lst', type=str, help='', default='./data/megaface_lst')
-  parser.add_argument('--feature-dir-input', type=str, help='', default='./feature_out')
-  parser.add_argument('--feature-dir-out', type=str, help='', default='./feature_out_clean')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/IFRT/README.md b/IFRT/README.md
deleted file mode 100644
index 83cb2ffda..000000000
--- a/IFRT/README.md
+++ /dev/null
@@ -1,67 +0,0 @@
-# InsightFace Recognition Test (IFRT)
-**IFRT** is a globalised fair benchmark for face recognition algorithms. IFRT evaluates the algorithm performance on worldwide web pictures which contain various sex, age and race groups, but no identification photos.
-
-**IFRT** testset consists of non-celebrities so we can ensure that it has very few overlap with public available face recognition training set, such as MS1M and CASIA as they mostly collected from online celebrities. As the result, we can evaluate the FAIR performance for different algorithms.
-
-Similar to [FRVT](https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt), we encourage participants to prepare a black-box feature extractor. To simplify this process, users can just replace their trained ArcFace model(with or without encryption) in our simple open-sourced pre-packaged software.
-
-~~Submitting features is also allowed, you can send an e-mail to us to request the test image set, with a promise not to redistribute it.~~
-
-## Dataset Statistics and Visualization
-
-IFRT testset contains 242,143 identities and 1,624,305 images.
-
-| Race-Set     | Identities  | Images        |  Positive Pairs   | Negative Pairs        |
-| -------      | ----------  | -----------   |  -----------      | -----------           |
-| African      | 43,874      | 298,010       |  870,091          | 88,808,791,999        |
-| Caucasian    | 103,293     | 697,245       |  2,024,609        | 486,147,868,171       |
-| Indian       | 35,086      | 237,080       |  688,259          | 56,206,001,061        |
-| Asian        | 59,890      | 391,970       |  1,106,078        | 153,638,982,852       |
-| **ALL**      | **242,143** | **1,624,305** |  **4,689,037**    | **2,638,360,419,683** |
-
-<details>
-  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
-  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrtsample_blur.jpg" alt="ifrtsample" width="640">
-</details>
-
-## Evaluation Metric
-
-TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6).
-
-## Baselines
-
-| Backbone     | Dataset     | DESC.   | African | Caucasian | South Asian | East Asian | ALL   |
-| ------------ | ----------- | ------- | ----- | ----- | ------ | ----- | ----- |
-| R100         | CASIA       | ArcFace | 39.67 | 53.93 | 47.81  | 16.17 | 37.53 |
-| R50          | VGG2        | ArcFace | 49.20 | 65.93 | 56.22  | 27.15 | 47.13 |
-| R50          | MS1M-V2     | ArcFace | 71.97 | 83.24 | 79.66  | 22.94 | 56.20 |
-| R50          | MS1M-V3     | ArcFace | 76.24 | 86.21 | 84.44  | 37.43 | 71.02 |
-| R124         | MS1M-V3     | ArcFace | 81.08 | 89.06 | 87.53  | 38.40 | 74.76 |
-| R124         | MS1M-V3     | +FlipTest | 83.22 | 90.43 | 89.22  | 39.61 | 75.69 |
-| R100	       | Glint360k   | PartialFC(r=0.1)| 90.45 | 94.60	| 93.96	| 63.91	| 88.23 |
-| R180 | *InsightFace-Private*   | ArcFace | 94.45 | 96.98 | 96.02  | 91.67 | 96.26 |
-| R180 | *Private*   | block-box | 97.54 | 98.67 | 98.35  | 83.83 | 96.93 |
-
-(MS1M-V2 means MS1M-ArcFace, MS1M-V3 means MS1M-RetinaFace)
-
-(We only consider *African* to *African* comparisons in *African* subset, so others like *African* to *Caucasian* will be ignored)
-
-## How to Participate
-
-Send an e-mail to **insightface.challenge(AT)gmail.com** after preparing your black-box feature extractor or your academic model file(without commercial risk), with your name, organization and submission comments.
-
-There are some ways to submit:
-
-1. (Recommended) Submit black-box face feature extracting tool.
-    * Use python binding to provide python interface: `feat = get_feature(image, bbox, landmark)`, where shape(image)==(H,W,3), shape(bbox)==(4,), shape(landmark)==(5,2) and shape(feat)==(K,). You can either use the provided landmark or detect them by yourself.
-    * In current stage, it should be better to not encrypt your feature embeddings, for fast GPU N:N matrix calculation.
-    * You can add some restrictions on your tool. Such as number of api calls and time constraints.
-2. (Simplest) Submit your recognition model.
-    * Submit MXNet ArcFace model with the same face alignment. In this case, you can just submit the single model file.
-    * In other case, such as PyTorch/TF models or ArcFace models with different face alignment method, please give us an example on how to generate feature embeddings. (eg. provide a function `get_feature(image, bbox, landmark)`)
-
-## Leaderboard
-
-[Leaderboard](http://insightface.ai/IFRT) on insightface.ai. (TODO)
-
-TODO
diff --git a/LICENSE b/LICENSE
deleted file mode 100644
index 1562c52df..000000000
--- a/LICENSE
+++ /dev/null
@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2018 Jiankang Deng and Jia Guo
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
diff --git a/PRNet.mxnet/README.md b/PRNet.mxnet/README.md
deleted file mode 100644
index a9896bf07..000000000
--- a/PRNet.mxnet/README.md
+++ /dev/null
@@ -1,6 +0,0 @@
-MXNet implementation of [Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network](http://openaccess.thecvf.com/content_ECCV_2018/papers/Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf). 
-
-Original [PyTorch implementation](https://github.com/YadiraF/PRNet)
-
-Pretrained Models and details coming soon.
-
diff --git a/PRNet.mxnet/config.py b/PRNet.mxnet/config.py
deleted file mode 100644
index 7ff5ba9e4..000000000
--- a/PRNet.mxnet/config.py
+++ /dev/null
@@ -1,89 +0,0 @@
-import numpy as np
-from easydict import EasyDict as edict
-
-config = edict()
-
-#default training/dataset config
-config.num_classes = 3
-config.input_img_size = 256
-config.output_label_size = 64
-
-# network settings
-network = edict()
-
-network.hourglass = edict()
-network.hourglass.net_sta = 0
-network.hourglass.net_n = 4
-network.hourglass.net_dcn = 0
-network.hourglass.net_stacks = 1
-network.hourglass.net_block = 'resnet'
-network.hourglass.net_binarize = False
-network.hourglass.losstype = 'heatmap'
-network.hourglass.multiplier = 1.0
-
-network.prnet = edict()
-network.prnet.net_sta = 0
-network.prnet.net_n = 5
-network.prnet.net_dcn = 0
-network.prnet.net_stacks = 1
-network.prnet.net_modules = 2
-network.prnet.net_block = 'hpm'
-network.prnet.net_binarize = False
-network.prnet.losstype = 'heatmap'
-network.prnet.multiplier = 0.25
-
-network.hpm = edict()
-network.hpm.net_sta = 0
-network.hpm.net_n = 4
-network.hpm.net_dcn = 0
-network.hpm.net_stacks = 1
-network.hpm.net_block = 'hpm'
-network.hpm.net_binarize = False
-network.hpm.losstype = 'heatmap'
-network.hpm.multiplier = 1.0
-
-
-# dataset settings
-dataset = edict()
-
-
-dataset.prnet = edict()
-dataset.prnet.dataset = '3D'
-dataset.prnet.landmark_type = 'dense'
-dataset.prnet.dataset_path = './data64'
-dataset.prnet.num_classes = 3
-dataset.prnet.input_img_size = 256
-dataset.prnet.output_label_size = 64
-#dataset.prnet.label_xfirst = False
-dataset.prnet.val_targets = ['']
-
-# default settings
-default = edict()
-
-# default network
-default.network = 'hpm'
-default.pretrained = ''
-default.pretrained_epoch = 0
-# default dataset
-default.dataset = 'prnet'
-default.frequent = 20
-default.verbose = 200
-default.kvstore = 'device'
-
-default.prefix = 'model/A'
-default.end_epoch = 10000
-default.lr = 0.00025
-default.wd = 0.0
-default.per_batch_size = 20
-default.lr_step = '16000,24000,30000'
-
-def generate_config(_network, _dataset):
-    for k, v in network[_network].items():
-      config[k] = v
-      default[k] = v
-    for k, v in dataset[_dataset].items():
-      config[k] = v
-      default[k] = v
-    config.network = _network
-    config.dataset = _dataset
-
diff --git a/PRNet.mxnet/data.py b/PRNet.mxnet/data.py
deleted file mode 100644
index 2ad1bfbfc..000000000
--- a/PRNet.mxnet/data.py
+++ /dev/null
@@ -1,164 +0,0 @@
-# pylint: skip-file
-import mxnet as mx
-import numpy as np
-import sys, os
-import random
-import glob
-import math
-import scipy.misc
-import cv2
-import logging
-import sklearn
-import datetime
-import img_helper
-from mxnet.io import DataIter
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-from PIL import Image
-from config import config
-from skimage import transform as tf
-
-
-class FaceSegIter(DataIter):
-    def __init__(self, path, batch_size, 
-                 per_batch_size = 0,
-                 aug_level = 0,
-                 force_mirror = False,
-                 exf = 1,
-                 args = None):
-      self.aug_level = aug_level
-      self.force_mirror = force_mirror
-      self.exf = exf
-      self.batch_size = batch_size
-      self.per_batch_size = per_batch_size
-      self.image_file_list = []
-      self.uv_file_list = []
-      for _file in glob.glob(os.path.join(path, '*.jpg')):
-        self.image_file_list.append(_file)
-      for img in self.image_file_list:
-        uv_file = img[0:-3]+"npy"
-        self.uv_file_list.append(uv_file)
-      self.seq = range(len(self.image_file_list))
-      print('train size', len(self.seq))
-      self.cur = 0
-      self.reset()
-      self.data_shape = (3, config.input_img_size, config.input_img_size)
-      self.num_classes = config.num_classes
-      self.input_img_size = config.input_img_size
-      #self.label_classes = self.num_classes
-      self.output_label_size = config.output_label_size
-      #if aug_level>0:
-      #  self.output_label_size = config.output_label_size
-      #else:
-      #  self.output_label_size = self.input_img_size
-      self.label_shape = (self.num_classes, self.output_label_size, self.output_label_size)
-      self.provide_data = [('data', (batch_size,) + self.data_shape)]
-      self.provide_label = [('softmax_label', (batch_size,) + self.label_shape),
-                            ('mask_label', (batch_size,)+ self.label_shape)]
-      weight_mask = cv2.imread('./uv-data/uv_weight_mask.png')
-      print('weight_mask', weight_mask.shape)
-      if weight_mask.shape[0]!=self.output_label_size:
-        weight_mask = cv2.resize(weight_mask, (self.output_label_size, self.output_label_size) )
-      #idx = np.where(weight_mask>0)[0]
-      #print('weight idx', idx)
-      weight_mask = weight_mask.astype(np.float32)
-      weight_mask /= 255.0
-
-      vis_mask = cv2.imread('./uv-data/uv_face_mask.png')
-      print('vis_mask', vis_mask.shape)
-      if vis_mask.shape[0]!=self.output_label_size:
-        vis_mask = cv2.resize(vis_mask, (self.output_label_size, self.output_label_size) )
-      vis_mask = vis_mask.astype(np.float32)
-      vis_mask /= 255.0
-      weight_mask *= vis_mask
-      print('weight_mask', weight_mask.shape)
-      weight_mask = weight_mask.transpose( (2,0,1) )
-      #WM = np.zeros( (batch_size,)+self.label_shape, dtype=np.float32 )
-      #for i in range(batch_size):
-      #  WM[i] = weight_mask
-      #weight_mask = WM
-      #weight_mask = weight_mask.reshape( (1, 3, weight_mask.shape[0], weight_mask.shape[1]) )
-      weight_mask = weight_mask[np.newaxis,:,:,:]
-      print('weight_mask', weight_mask.shape)
-      weight_mask = np.tile(weight_mask, (batch_size,1,1,1))
-      print('weight_mask', weight_mask.shape)
-      self.weight_mask = nd.array(weight_mask)
-      self.img_num = 0
-      self.invalid_num = 0
-      self.mode = 1
-      self.vis = 0
-      self.stats = [0,0]
-
-    def get_data_shape(self):
-        return self.data_shape
-
-    #def get_label_shape(self):
-    #    return self.label_shape
-
-    def get_shape_dict(self):
-        D = {}
-        for (k,v) in self.provide_data:
-            D[k] = v
-        for (k,v) in self.provide_label:
-            D[k] = v
-        return D
-
-    def get_label_names(self):
-        D = []
-        for (k,v) in self.provide_label:
-            D.append(k)
-        return D
-
-    def reset(self):
-      #print('reset')
-      self.cur = 0
-      if self.aug_level>0:
-        random.shuffle(self.seq)
-
-    def next_sample(self):
-      """Helper function for reading in next sample."""
-      if self.cur >= len(self.seq):
-        raise StopIteration
-      idx = self.seq[self.cur]
-      self.cur += 1
-      uv_path = self.uv_file_list[idx]
-      image_path = self.image_file_list[idx]
-      uvmap = np.load(uv_path)
-      img = cv2.imread(image_path)[:,:,::-1]#to rgb
-      hlabel = uvmap
-      #print(hlabel.shape)
-      #hlabel = np.array(header.label).reshape( (self.output_label_size, self.output_label_size, self.num_classes) )
-      hlabel /= self.input_img_size
-
-      return img, hlabel
-
-
-    def next(self):
-        """Returns the next batch of data."""
-        #print('next')
-        batch_size = self.batch_size
-        batch_data = nd.empty((batch_size,)+self.data_shape)
-        batch_label = nd.empty((batch_size,)+self.label_shape)
-        i = 0
-        #self.cutoff = random.randint(800,1280)
-        try:
-            while i < batch_size:
-                #print('N', i)
-                data, label = self.next_sample()
-                data = nd.array(data)
-                data = nd.transpose(data, axes=(2, 0, 1))
-                label = nd.array(label)
-                label = nd.transpose(label, axes=(2, 0, 1))
-                batch_data[i][:] = data
-                batch_label[i][:] = label
-                i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        #return {self.data_name  :  batch_data,
-        #        self.label_name :  batch_label}
-        #print(batch_data.shape, batch_label.shape)
-        return mx.io.DataBatch([batch_data], [batch_label, self.weight_mask], batch_size - i)
-
diff --git a/PRNet.mxnet/metric.py b/PRNet.mxnet/metric.py
deleted file mode 100644
index 11f3770c5..000000000
--- a/PRNet.mxnet/metric.py
+++ /dev/null
@@ -1,99 +0,0 @@
-import mxnet as mx
-import numpy as np
-import math
-import cv2
-from config import config
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[0].asnumpy()
-    self.sum_metric += np.mean(loss)
-    self.num_inst += 1.0
-
-class NMEMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(NMEMetric, self).__init__(
-        'NME', axis=self.axis,
-        output_names=None, label_names=None)
-    #self.losses = []
-    self.count = 0
-
-  def cal_nme(self, label, pred_label):
-    nme = []
-    for b in xrange(pred_label.shape[0]):
-      record = [None]*6
-      item = []
-      if label.ndim==4:
-          _heatmap = label[b][36]
-          if np.count_nonzero(_heatmap)==0:
-              continue
-      else:#ndim==3
-          #print(label[b])
-          if np.count_nonzero(label[b])==0:
-              continue
-      for p in xrange(pred_label.shape[1]):
-        if label.ndim==4:
-            heatmap_gt = label[b][p]
-            ind_gt = np.unravel_index(np.argmax(heatmap_gt, axis=None), heatmap_gt.shape)
-            ind_gt = np.array(ind_gt)
-        else:
-            ind_gt = label[b][p]
-            #ind_gt = ind_gt.astype(np.int)
-            #print(ind_gt)
-        heatmap_pred = pred_label[b][p]
-        heatmap_pred = cv2.resize(heatmap_pred, (config.input_img_size, config.input_img_size))
-        ind_pred = np.unravel_index(np.argmax(heatmap_pred, axis=None), heatmap_pred.shape)
-        ind_pred = np.array(ind_pred)
-        #print(ind_gt.shape)
-        #print(ind_pred)
-        if p==36:
-            #print('b', b, p, ind_gt, np.count_nonzero(heatmap_gt))
-            record[0] = ind_gt
-        elif p==39:
-            record[1] = ind_gt
-        elif p==42:
-            record[2] = ind_gt
-        elif p==45:
-            record[3] = ind_gt
-        if record[4] is None or record[5] is None:
-            record[4] = ind_gt
-            record[5] = ind_gt
-        else:
-            record[4] = np.minimum(record[4], ind_gt)
-            record[5] = np.maximum(record[5], ind_gt)
-        #print(ind_gt.shape, ind_pred.shape)
-        value = np.sqrt(np.sum(np.square(ind_gt - ind_pred)))
-        item.append(value)
-      _nme = np.mean(item)
-      if config.landmark_type=='2d':
-          left_eye = (record[0]+record[1])/2
-          right_eye = (record[2]+record[3])/2
-          _dist = np.sqrt(np.sum(np.square(left_eye - right_eye)))
-          #print('eye dist', _dist, left_eye, right_eye)
-          _nme /= _dist
-      else:
-          #_dist = np.sqrt(float(label.shape[2]*label.shape[3]))
-          _dist = np.sqrt(np.sum(np.square(record[5] - record[4])))
-          #print(_dist)
-          _nme /= _dist
-      nme.append(_nme)
-    return np.mean(nme)
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0].asnumpy()
-    pred_label = preds[-1].asnumpy()
-    nme = self.cal_nme(label, pred_label)
-
-    #print('nme', nme)
-    #nme = np.mean(nme)
-    self.sum_metric += np.mean(nme)
-    self.num_inst += 1.0
diff --git a/PRNet.mxnet/symbol/sym_heatmap.py b/PRNet.mxnet/symbol/sym_heatmap.py
deleted file mode 100644
index 444f9fae1..000000000
--- a/PRNet.mxnet/symbol/sym_heatmap.py
+++ /dev/null
@@ -1,435 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import mxnet as mx
-import numpy as np
-from config import config
-
-
-ACT_BIT = 1
-bn_mom = 0.9
-workspace = 256
-memonger = False
-
-
-
-def Conv(**kwargs):
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-def Act(data, act_type, name):
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-#def lin(data, num_filter, workspace, name, binarize, dcn):
-#  bit = 1
-#  if not binarize:
-#    if not dcn:
-#        conv1 = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-#                                      no_bias=True, workspace=workspace, name=name + '_conv')
-#        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#        return act1
-#    else:
-#        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#        conv1_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = act1,
-#                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-#        conv1 = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=act1, offset=conv1_offset,
-#                num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=False)
-#        #conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-#        #                              no_bias=False, workspace=workspace, name=name + '_conv')
-#        return conv1
-#  else:
-#    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#    conv1 = mx.sym.QConvolution_v1(data=act1, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-#                               no_bias=True, workspace=workspace, name=name + '_conv', act_bit=ACT_BIT, weight_bit=bit)
-#    conv1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-#    return conv1
-
-def lin3(data, num_filter, workspace, name, k, g=1, d=1):
-    if k!=3:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(k,k), stride=(1,1), pad=((k-1)//2,(k-1)//2), num_group=g,
-                                      no_bias=True, workspace=workspace, name=name + '_conv')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(k,k), stride=(1,1), pad=(d,d), num_group=g, dilate=(d, d),
-                                      no_bias=True, workspace=workspace, name=name + '_conv')
-    bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-    ret = act1
-    return ret
-
-def ConvFactory(data, num_filter, kernel, stride=(1, 1), pad=(0, 0), act_type="relu", mirror_attr={}, with_act=True, dcn=False, name=''):
-    if not dcn:
-      conv = mx.symbol.Convolution(
-          data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad, no_bias=True, workspace=workspace, name=name+'_conv')
-    else:
-        conv_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = data,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-        conv = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=data, offset=conv_offset,
-                num_filter=num_filter, pad=(1,1), kernel=(3,3), num_deformable_group=1, stride=stride, dilate=(1, 1), no_bias=False)
-    bn = mx.symbol.BatchNorm(data=conv, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name+'_bn')
-    if with_act:
-      act = Act(bn, act_type, name=name+'_relu')
-      #act = mx.symbol.Activation(
-      #    data=bn, act_type=act_type, attr=mirror_attr, name=name+'_relu')
-      return act
-    else:
-      return bn
-
-class CAB:
-    def __init__(self, data, nFilters, nModules, n, workspace, name, dilate, group):
-        self.data = data
-        self.nFilters = nFilters
-        self.nModules = nModules
-        self.n = n
-        self.workspace = workspace
-        self.name = name
-        self.dilate = dilate
-        self.group = group
-        self.sym_map = {}
-
-    def get_output(self, w, h):
-        key = (w, h)
-        if key in self.sym_map:
-            return self.sym_map[key]
-        ret = None
-        if h==self.n:
-            if w==self.n:
-                ret = (self.data, self.nFilters)
-            else:
-                x = self.get_output(w+1, h)
-                f = int(x[1]*0.5)
-                if w!=self.n-1:
-                    body = lin3(x[0], f, self.workspace, "%s_w%d_h%d_1"%(self.name, w, h), 3, self.group, 1)
-                else:
-                    body = lin3(x[0], f, self.workspace, "%s_w%d_h%d_1"%(self.name, w, h), 3, self.group, self.dilate)
-                ret = (body,f)
-        else:
-            x = self.get_output(w+1, h+1)
-            y = self.get_output(w, h+1)
-            if h%2==1 and h!=w:
-                xbody = lin3(x[0], x[1], self.workspace, "%s_w%d_h%d_2"%(self.name, w, h), 3, x[1])
-                #xbody = xbody+x[0]
-            else:
-                xbody = x[0]
-            #xbody = x[0]
-            #xbody = lin3(x[0], x[1], self.workspace, "%s_w%d_h%d_2"%(self.name, w, h), 3, x[1])
-            if w==0:
-                ybody = lin3(y[0], y[1], self.workspace, "%s_w%d_h%d_3"%(self.name, w, h), 3, self.group)
-            else:
-                ybody = y[0]
-            ybody = mx.sym.concat(y[0], ybody, dim=1)
-            body = mx.sym.add_n(xbody,ybody, name="%s_w%d_h%d_add"%(self.name, w, h))
-            body = body/2
-            ret = (body, x[1])
-        self.sym_map[key] = ret
-        return ret
-
-    def get(self):
-        return self.get_output(1, 1)[0]
-
-def conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    bit = 1
-    #print('in unit2')
-    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
-    conv1 = Conv(data=act1, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
-    conv2 = Conv(data=act2, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
-    conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                               workspace=workspace, name=name + '_conv3')
-    #if binarize:
-    #  conv3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-    if dim_match:
-        shortcut = data
-    else:
-        shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return conv3 + shortcut
-
-
-def conv_prnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    #print('in unit2')
-    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
-    conv1 = Conv(data=act1, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
-    conv2 = Conv(data=act2, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
-    conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                               workspace=workspace, name=name + '_conv3')
-    if dim_match:
-        shortcut = data
-    else:
-        shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return conv3 + shortcut
-
-def conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilation, **kwargs):
-    bit = 1
-    #print('in unit2')
-    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
-    conv1 = Conv(data=act1, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
-    conv2 = Conv(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
-    conv3 = Conv(data=act3, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation), 
-            no_bias=True, workspace=workspace, name=name + '_conv3')
-    conv4 = mx.symbol.Concat(*[conv1, conv2, conv3])
-    if dim_match:
-        shortcut = data
-    else:
-        shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return conv4 + shortcut
-
-
-def block17(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}, name=''):
-    tower_conv = ConvFactory(net, 192, (1, 1), name=name+'_conv')
-    tower_conv1_0 = ConvFactory(net, 129, (1, 1), name=name+'_conv1_0')
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 160, (1, 7), pad=(1, 2), name=name+'_conv1_1')
-    tower_conv1_2 = ConvFactory(tower_conv1_1, 192, (7, 1), pad=(2, 1), name=name+'_conv1_2')
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False, name=name+'_conv_out')
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-def block35(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}, name=''):
-    M = 1.0
-    tower_conv = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv')
-    tower_conv1_0 = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv1_0')
-    tower_conv1_1 = ConvFactory(tower_conv1_0, int(input_num_channels*0.25*M), (3, 3), pad=(1, 1), name=name+'_conv1_1')
-    tower_conv2_0 = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv2_0')
-    tower_conv2_1 = ConvFactory(tower_conv2_0, int(input_num_channels*0.375*M), (3, 3), pad=(1, 1), name=name+'_conv2_1')
-    tower_conv2_2 = ConvFactory(tower_conv2_1, int(input_num_channels*0.5*M), (3, 3), pad=(1, 1), name=name+'_conv2_2')
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False, name=name+'_conv_out')
-
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-def conv_inception(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    assert not binarize
-    if stride[0]>1 or not dim_match:
-        return conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs)
-    conv4 = block35(data, num_filter, name=name+'_block35')
-    return conv4
-
-def conv_cab(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    if stride[0]>1 or not dim_match:
-        return conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs)
-    cab = CAB(data, num_filter, 1, 4, workspace, name, dilate, 1)
-    return cab.get()
-
-def conv_block(data, num_filter, stride, dim_match, name, binarize, dcn, dilate):
-  if config.net_block=='resnet':
-    return conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='inception':
-    return conv_inception(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='hpm':
-    return conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='cab':
-    return conv_cab(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='prnet':
-    return conv_prnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-
-def hourglass(data, nFilters, nModules, n, workspace, name, binarize, dcn):
-  s = 2
-  _dcn = False
-  up1 = data
-  for i in xrange(nModules):
-    up1 = conv_block(up1, nFilters, (1,1), True, "%s_up1_%d"%(name,i), binarize, _dcn, 1)
-  low1 = mx.sym.Pooling(data=data, kernel=(s, s), stride=(s,s), pad=(0,0), pool_type='max')
-  #low1 = ConvFactory(data, nFilters, (4,4), stride=(2,2), pad=(1,1), name=name+'_conv')
-  #low1 = ConvFactory(data, nFilters, (3,3), stride=(2,2), pad=(1,1), name=name+'_conv')
-  #low1 = ConvFactory(up1, nFilters, (3,3), stride=(2,2), pad=(1,1), name=name+'_conv')
-  for i in xrange(nModules):
-    low1 = conv_block(low1, nFilters, (1,1), True, "%s_low1_%d"%(name,i), binarize, _dcn, 1)
-  if n>1:
-    low2 = hourglass(low1, nFilters, nModules, n-1, workspace, "%s_%d"%(name, n-1), binarize, dcn)
-  else:
-    low2 = low1
-    for i in xrange(nModules):
-      low2 = conv_block(low2, nFilters, (1,1), True, "%s_low2_%d"%(name,i), binarize, _dcn, 1) #TODO
-  low3 = low2
-  for i in xrange(nModules):
-    low3 = conv_block(low3, nFilters, (1,1), True, "%s_low3_%d"%(name,i), binarize, _dcn, 1)
-  up2 = mx.symbol.UpSampling(low3, scale=s, sample_type='nearest', workspace=512, name='%s_upsampling_%s'%(name,n), num_args=1)
-  #up2 = mx.symbol.UpSampling(low3, scale=s, sample_type='bilinear', num_filter=nFilters, workspace=512, name='%s_upsampling_%s'%(name,n), num_args=1)
-  #up2 = mx.symbol.Deconvolution(data=low3, num_filter=nFilters, kernel=(s*2,s*2), 
-  #  stride=(s, s), pad=(s//2, s//2),
-  #  name='%s_upsampling_%s'%(name,n),
-  #  attr={'lr_mult': '0.1'})
-  #return mx.symbol.add_n(up1, up2)
-  return up2
-
-
-def prnet_loss(pred, gt_label, mask_label):
-  loss = pred - gt_label
-  #loss = mx.symbol.smooth_l1(loss, scalar=3.0)
-  loss = mx.symbol.abs(loss)
-  loss = mx.symbol.broadcast_mul(loss, mask_label)
-  #loss = mx.symbol.mean(loss, axis=0)
-  #loss = loss*loss
-  #loss = mx.symbol.mean(loss)
-  return loss
-
-def ce_loss(x, y):
-  #loss = mx.sym.SoftmaxOutput(data = x, label = y, normalization='valid', multi_output=True)
-  x_max = mx.sym.max(x, axis=[2,3], keepdims=True)
-  x = mx.sym.broadcast_minus(x, x_max)
-  body = mx.sym.exp(x)
-  sums = mx.sym.sum(body, axis=[2,3], keepdims=True)
-  body = mx.sym.broadcast_div(body, sums)
-  loss = mx.sym.log(body)
-  loss = loss*y*-1.0
-  #loss = mx.symbol.mean(loss, axis=[1,2,3])
-  loss = mx.symbol.mean(loss)
-  return loss
-
-def get_symbol(num_classes):
-    m = config.multiplier
-    sFilters = max(int(64*m), 16)
-    mFilters = max(int(128*m), 32)
-    nFilters = int(256*m)
-
-    nModules = config.net_modules
-    nStacks = config.net_stacks
-    binarize = config.net_binarize
-    input_size = config.input_img_size
-    label_size = config.output_label_size
-    use_STA = config.net_sta
-    N = config.net_n
-    DCN = config.net_dcn
-    per_batch_size = config.per_batch_size
-    print('binarize', binarize)
-    print('use_STA', use_STA)
-    print('use_N', N)
-    print('use_DCN', DCN)
-    print('per_batch_size', per_batch_size)
-    #assert(label_size==64 or label_size==32)
-    #assert(input_size==128 or input_size==256)
-    D = input_size // label_size
-    print(input_size, label_size, D)
-    data = mx.sym.Variable(name='data')
-    data = data-127.5
-    data = data*0.0078125
-    gt_label = mx.symbol.Variable(name='softmax_label')
-    mask_label = mx.symbol.Variable(name='mask_label')
-    losses = []
-    closses = []
-    #body = Conv(data=data, num_filter=sFilters, kernel=(3, 3), stride=(1,1), pad=(1, 1),
-    #                        no_bias=True, name="conv0", workspace=workspace)
-    body = Conv(data=data, num_filter=sFilters, kernel=(7,7), stride=(2,2), pad=(3,3),
-                            no_bias=True, name="conv0", workspace=workspace)
-    #body = Conv(data=data, num_filter=sFilters, kernel=(4,4), stride=(2,2), pad=(1,1),
-    #                        no_bias=True, name="conv0", workspace=workspace)
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-    body = Act(data=body, act_type='relu', name='relu0')
-
-    dcn = False
-    body = conv_block(body, mFilters, (1,1), sFilters==mFilters, 'res0', False, dcn, 1)
-    body = mx.sym.Pooling(data=body, kernel=(2, 2), stride=(2,2), pad=(0,0), pool_type='max')
-
-    #body = Conv(data=body, num_filter=mFilters, kernel=(4,4), stride=(2,2), pad=(1,1),
-    #                        no_bias=True, name="conv1", workspace=workspace)
-    #body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    #body = Act(data=body, act_type='relu', name='relu1')
-
-    #body = conv_block(body, mFilters, (1,1), True, 'res1', False, dcn, 1) #TODO
-    body = conv_block(body, nFilters, (1,1), mFilters==nFilters, 'res2', binarize, dcn, 1) #binarize=True?
-
-    heatmap = None
-    outs = []
-
-    body = hourglass(body, nFilters, nModules, config.net_n, workspace, 'stack0_hg', binarize, dcn)
-    for j in xrange(nModules):
-      body = conv_block(body, nFilters, (1,1), True, 'stack0_unit%d'%(j), binarize, dcn, 1)
-    _dcn = False
-    ll = ConvFactory(body, nFilters, (1,1), dcn = _dcn, name='stack0_ll')
-    _name = 'heatmap'
-    pred = Conv(data=ll, num_filter=num_classes, kernel=(1, 1), stride=(1,1), pad=(0,0),
-                              name=_name, workspace=workspace)
-    loss = prnet_loss(pred, gt_label, mask_label)
-    outs.append(mx.sym.MakeLoss(loss))
-
-
-    pred = mx.symbol.BlockGrad(pred)
-    #loss = mx.symbol.add_n(*losses)
-    #loss = mx.symbol.MakeLoss(loss)
-    #syms = [loss]
-    outs.append(pred)
-    sym = mx.symbol.Group( outs )
-    return sym
-
-def init_weights(sym, data_shape_dict):
-    #print('in hg')
-    arg_name = sym.list_arguments()
-    aux_name = sym.list_auxiliary_states()
-    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(arg_name, arg_shape))
-    aux_shape_dict = dict(zip(aux_name, aux_shape))
-    #print(aux_shape)
-    #print(aux_params)
-    #print(arg_shape_dict)
-    arg_params = {}
-    aux_params = {}
-    for k,v in arg_shape_dict.iteritems():
-      #print(k,v)
-      if k.endswith('offset_weight') or k.endswith('offset_bias'):
-        print('initializing',k)
-        arg_params[k] = mx.nd.zeros(shape = v)
-      elif k.startswith('fc6_'):
-        if k.endswith('_weight'):
-          print('initializing',k)
-          arg_params[k] = mx.random.normal(0, 0.01, shape=v)
-        elif k.endswith('_bias'):
-          print('initializing',k)
-          arg_params[k] = mx.nd.zeros(shape=v)
-      elif k.find('upsampling')>=0:
-        print('initializing upsampling_weight', k)
-        arg_params[k] = mx.nd.zeros(shape=arg_shape_dict[k])
-        init = mx.init.Initializer()
-        init._init_bilinear(k, arg_params[k])
-    return arg_params, aux_params
-
diff --git a/PRNet.mxnet/train.py b/PRNet.mxnet/train.py
deleted file mode 100644
index c4bb954c8..000000000
--- a/PRNet.mxnet/train.py
+++ /dev/null
@@ -1,215 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import logging
-import argparse
-from data import FaceSegIter
-import mxnet as mx
-import mxnet.optimizer as optimizer
-import numpy as np
-import os
-import sys
-import math
-import random
-import cv2
-from config import config, default, generate_config
-from optimizer import ONadam
-from metric import LossValueMetric, NMEMetric
-sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
-import sym_heatmap
-#import sym_fc
-#from symbol import fc
-
-
-args = None
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-def main(args):
-  _seed = 727
-  random.seed(_seed)
-  np.random.seed(_seed)
-  mx.random.seed(_seed)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in xrange(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  #ctx = [mx.gpu(0)]
-  args.ctx_num = len(ctx)
-
-  args.batch_size = args.per_batch_size*args.ctx_num
-  config.per_batch_size = args.per_batch_size
-
-
-
-  print('Call with', args, config)
-  train_iter = FaceSegIter(path = config.dataset_path,
-      batch_size = args.batch_size,
-      per_batch_size = args.per_batch_size,
-      aug_level = 1,
-      exf = args.exf,
-      args = args,
-      )
-
-  data_shape = train_iter.get_data_shape()
-  #label_shape = train_iter.get_label_shape()
-  sym = sym_heatmap.get_symbol(num_classes=config.num_classes)
-  if len(args.pretrained)==0:
-      #data_shape_dict = {'data' : (args.per_batch_size,)+data_shape, 'softmax_label' : (args.per_batch_size,)+label_shape}
-      data_shape_dict = train_iter.get_shape_dict()
-      arg_params, aux_params = sym_heatmap.init_weights(sym, data_shape_dict)
-  else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      _, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      #sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-
-  model = mx.mod.Module(
-      context       = ctx,
-      symbol        = sym,
-      label_names   = train_iter.get_label_names(),
-  )
-  #lr = 1.0e-3
-  #lr = 2.5e-4
-  #_rescale_grad = 1.0/args.ctx_num
-  _rescale_grad = 1.0/args.batch_size
-  #lr = args.lr
-  #opt = optimizer.Nadam(learning_rate=args.lr, wd=args.wd, rescale_grad=_rescale_grad, clip_gradient=5.0)
-  if args.optimizer=='onadam':
-    opt = ONadam(learning_rate=args.lr, wd=args.wd, rescale_grad=_rescale_grad, clip_gradient=5.0)
-  elif args.optimizer=='nadam':
-    opt = optimizer.Nadam(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  elif args.optimizer=='rmsprop':
-    opt = optimizer.RMSProp(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  elif args.optimizer=='adam':
-    opt = optimizer.Adam(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  else:
-    opt = optimizer.SGD(learning_rate=args.lr, momentum=0.9, wd=args.wd, rescale_grad=_rescale_grad)
-  initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2)
-  _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
-  _metric = LossValueMetric()
-  #_metric = NMEMetric()
-  #_metric2 = AccMetric()
-  #eval_metrics = [_metric, _metric2]
-  eval_metrics = [_metric]
-  lr_steps = [int(x) for x in args.lr_step.split(',')]
-  print('lr-steps', lr_steps)
-  global_step = [0]
-
-  def val_test():
-    all_layers = sym.get_internals()
-    vsym = all_layers['heatmap_output']
-    vmodel = mx.mod.Module(symbol=vsym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    vmodel.bind(data_shapes=[('data', (args.batch_size,)+data_shape)])
-    arg_params, aux_params = model.get_params()
-    vmodel.set_params(arg_params, aux_params)
-    for target in config.val_targets:
-        _file = os.path.join(config.dataset_path, '%s.rec'%target)
-        if not os.path.exists(_file):
-            continue
-        val_iter = FaceSegIter(path_imgrec = _file,
-          batch_size = args.batch_size,
-          #batch_size = 4,
-          aug_level = 0,
-          args = args,
-          )
-        _metric = LossValueMetric()
-        val_metric = mx.metric.create(_metric)
-        val_metric.reset()
-        val_iter.reset()
-        diffs = []
-        for i, eval_batch in enumerate(val_iter):
-          #print(eval_batch.data[0].shape, eval_batch.label[0].shape)
-          batch_data = mx.io.DataBatch(eval_batch.data)
-          model.forward(batch_data, is_train=False)
-          _label = eval_batch.label[0].asnumpy()
-          _pred = model.get_outputs()[-1].asnumpy()
-          _diff = np.abs(_pred-_label)
-          _diff = np.mean(_diff)*config.input_img_size
-          #print('pred', _pred.shape, _label.shape)
-          #print('diff', _diff)
-          diffs.append(_diff)
-          model.update_metric(val_metric, eval_batch.label)
-        nme_value = val_metric.get_name_value()[0][1]
-        print('[%d][%s]LOSS: %f'%(global_step[0], target, nme_value))
-        print('avg diff', np.mean(diffs))
-  
-  def _batch_callback(param):
-    _cb(param)
-    global_step[0]+=1
-    mbatch = global_step[0]
-    for _lr in lr_steps:
-      if mbatch==_lr:
-        if args.optimizer=='sgd':
-          opt.lr *= 0.1
-        else:
-          opt.lr *= 0.5
-        print('lr change to', opt.lr)
-        break
-    if mbatch%1000==0:
-      print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-    if mbatch>0 and mbatch%args.verbose==0:
-      val_test()
-      if args.ckpt==1:
-        msave = mbatch//args.verbose
-        print('saving', msave)
-        arg, aux = model.get_params()
-        mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg, aux)
-    if mbatch==lr_steps[-1]:
-      if args.ckpt==2:
-        #msave = mbatch//args.verbose
-        msave = 1
-        print('saving', msave)
-        arg, aux = model.get_params()
-        mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg, aux)
-      sys.exit(0)
-
-  train_iter = mx.io.PrefetchingIter(train_iter)
-
-  model.fit(train_iter,
-      begin_epoch        = 0,
-      num_epoch          = 9999,
-      #eval_data          = val_iter,
-      eval_data          = None,
-      eval_metric        = eval_metrics,
-      kvstore            = 'device',
-      optimizer          = opt,
-      initializer        = initializer,
-      arg_params         = arg_params,
-      aux_params         = aux_params,
-      allow_missing      = True,
-      batch_end_callback = _batch_callback,
-      epoch_end_callback = None,
-      )
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='Train face alignment')
-  # general
-  parser.add_argument('--network', help='network name', default=default.network, type=str)
-  parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-  args, rest = parser.parse_known_args()
-  generate_config(args.network, args.dataset)
-  parser.add_argument('--prefix', default=default.prefix, help='directory to save model.')
-  parser.add_argument('--pretrained', default=default.pretrained, help='')
-  parser.add_argument('--optimizer', default='nadam', help='')
-  parser.add_argument('--lr', type=float, default=default.lr, help='')
-  parser.add_argument('--wd', type=float, default=default.wd, help='')
-  parser.add_argument('--per-batch-size', type=int, default=default.per_batch_size, help='')
-  parser.add_argument('--lr-step', help='learning rate steps (in epoch)', default=default.lr_step, type=str)
-  parser.add_argument('--ckpt', type=int, default=1, help='')
-  parser.add_argument('--norm', type=int, default=0, help='')
-  parser.add_argument('--exf', type=int, default=1, help='')
-  parser.add_argument('--frequent', type=int, default=default.frequent, help='')
-  parser.add_argument('--verbose', type=int, default=default.verbose, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/README.md b/README.md
index f8963d625..ebdcc56cc 100644
--- a/README.md
+++ b/README.md
@@ -1,240 +1,169 @@
 
 # InsightFace: 2D and 3D Face Analysis Project
 
-By Jia Guo and [Jiankang Deng](https://jiankangdeng.github.io/)
-
-## License
-
-The code of InsightFace is released under the MIT License. There is no limitation for both acadmic and commercial usage.
-
-The training data containing the annotation (and the models trained with these data) are available for non-commercial research purposes only.
-
-## Introduction
-
-InsightFace is an open source 2D&3D deep face analysis toolbox, mainly based on MXNet. 
-
-The master branch works with **MXNet 1.2 to 1.6**, with **Python 3.x**.
-
-
- ## ArcFace Video Demo
-
-[![ArcFace Demo](https://github.com/deepinsight/insightface/blob/master/resources/facerecognitionfromvideo.PNG)](https://www.youtube.com/watch?v=y-D1tReryGA&t=81s)
-
-Please click the image to watch the Youtube video. For Bilibili users, click [here](https://www.bilibili.com/video/av38041494?from=search&seid=11501833604850032313).
-
-## Recent Update
-
-**`2020-10-13`**: A new training method and one large training set(360K IDs) were released [here](https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc) by DeepGlint.
-
-**`2020-10-09`**: We opened a large scale recognition test benchmark [IFRT](https://github.com/deepinsight/insightface/tree/master/IFRT)
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/insightface_logo.jpg_320x320.webp" width="240"/>
+</div>
 
-**`2020-08-01`**: We released lightweight facial landmark models with fast coordinate regression(106 points). See detail [here](https://github.com/deepinsight/insightface/tree/master/alignment/coordinateReg).
+InsightFace project is mainly maintained by [Jia Guo](mailto:guojia@insightface.ai) and [Jiankang Deng](https://jiankangdeng.github.io/). 
 
-**`2020-04-27`**: InsightFace pretrained models and MS1M-Arcface are now specified as the only external training dataset, for iQIYI iCartoonFace challenge, see detail [here](http://challenge.ai.iqiyi.com/detail?raceId=5def71b4e9fcf68aef76a75e).
+For more information, please visit our website at [https://insightface.ai](https://insightface.ai)
 
-**`2020.02.21`**: Instant discussion group created on QQ with group-id: 711302608. For English developers, see install tutorial [here](https://github.com/deepinsight/insightface/issues/1069).
+## InsightFace 1.0 Update
 
-**`2020.02.16`**: RetinaFace now can detect faces with mask, for anti-CoVID19, see detail [here](https://github.com/deepinsight/insightface/tree/master/RetinaFaceAntiCov)
+InsightFace 1.0 focuses on easier local evaluation and desktop usage:
 
-**`2019.08.10`**: We achieved 2nd place at [WIDER Face Detection Challenge 2019](http://wider-challenge.org/2019.html).
+- The default Python package no longer builds the optional C++/Cython `face3d` extension, reducing local compiler requirements.
+- Added **InsightFace Evaluation Studio**, a cross-platform GUI Demo for Windows, macOS, and Linux.
+- The GUI supports local face recognition, enterprise model evaluation/reporting, and basic face swap trials.
 
-**`2019.05.30`**: [Presentation at cvmart](https://pan.baidu.com/s/1v9fFHBJ8Q9Kl9Z6GwhbY6A)
+## InsightFace Server
 
-**`2019.04.30`**: Our Face detector ([RetinaFace](https://github.com/deepinsight/insightface/tree/master/RetinaFace)) obtains state-of-the-art results on [the WiderFace dataset](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html).
+The new [InsightFace Server](server/) provides a simple self-hosted Web UI,
+snake_case REST API, and Python client for detection, comparison,
+registration, and Person search. A single Linux x86_64 CPU or NVIDIA GPU
+container runs local ONNX Runtime inference and SQLite with operator-supplied,
+manifest-verified models. It is a simple AWS Rekognition alternative with 
+accuracy-preserving INT8 embedding quantization and 50M+ image search on one RTX 5090 GPU.
 
-**`2019.04.14`**: We will launch a [Light-weight Face Recognition challenge/workshop](https://github.com/deepinsight/insightface/tree/master/iccv19-challenge) on ICCV 2019.
+## License
 
-**`2019.04.04`**: Arcface achieved state-of-the-art performance (7/109) on the NIST Face Recognition Vendor Test (FRVT) (1:1 verification)
-[report](https://www.nist.gov/sites/default/files/documents/2019/04/04/frvt_report_2019_04_04.pdf) (name: Imperial-000 and Imperial-001). Our solution is based on [MS1MV2+DeepGlintAsian, ResNet100, ArcFace loss]. 
+The code of InsightFace is released under the MIT License. There is no limitation for both academic and commercial usage.
 
-**`2019.02.08`**: Please check [https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace) for our parallel training code which can easily and efficiently support one million identities on a single machine (8* 1080ti).
+The training data containing the annotation (and the models trained with these data) are available for non-commercial research purposes only.
 
-**`2018.12.13`**: Inference acceleration [TVM-Benchmark](https://github.com/deepinsight/insightface/wiki/TVM-Benchmark).
+Both manual-downloading models from our github repo and auto-downloading models with our [python-library](python-package) follow the above license policy(which is for non-commercial research purposes only).
 
-**`2018.10.28`**: Light-weight attribute model [Gender-Age](https://github.com/deepinsight/insightface/tree/master/gender-age). About 1MB, 10ms on single CPU core. Gender accuracy 96% on validation set and 4.1 age MAE.
+`2025-11-24 Update:`
 
-**`2018.10.16`**: We achieved state-of-the-art performance on [Trillionpairs](http://trillionpairs.deepglint.com/results) (name: nttstar) and [IQIYI_VID](http://challenge.ai.iqiyi.com/detail?raceId=5afc36639689443e8f815f9e) (name: WitcheR). 
+1. For inswapper series face swap models (e.g., inswapper_128.onnx/[inswapper-512-live](https://github.com/deepinsight/inswapper-512-live)), please contact [contact@insightface.ai](mailto:contact@insightface.ai) for licensing and additional support.
+2. For open-sourced face recognition models (e.g., buffalo_l package), please contact [recognition-oss-pack@insightface.ai](mailto:recognition-oss-pack@insightface.ai) for licensing.
+3. For advanced face recognition SDK and models (e.g., InspireFace SDK), please contact [contact@insightface.ai](mailto:contact@insightface.ai) for licensing and additional support.
 
 
-## Contents
-[Deep Face Recognition](#deep-face-recognition)
-- [Introduction](#introduction)
-- [Training Data](#training-data)
-- [Train](#train)
-- [Pretrained Models](#pretrained-models)
-- [Verification Results On Combined Margin](#verification-results-on-combined-margin)
-- [Test on MegaFace](#test-on-megaface)
-- [512-D Feature Embedding](#512-d-feature-embedding)
-- [Third-party Re-implementation](#third-party-re-implementation)
+## Top News
 
-[Face Detection](#face-detection)
-- [RetinaFace](#retinaface)
-- [RetinaFaceAntiCov](#retinafaceanticov)
+**`2026-07-27`** [InsightFace Server](server/) Added a simple AWS Rekognition alternative with accuracy-preserving INT8 embedding quantization and 50M+ image search on one RTX 5090 GPU.
 
-[Face Alignment](#face-alignment)
-- [DenseUNet](#denseunet)
-- [CoordinateReg](#coordinatereg)
+**`2026-05-23`** `InsightFace 1.0` Added a cross-platform desktop GUI Demo for face recognition, enterprise evaluation, reports, and face swap trials, with a lighter default Python install that removes C++ build requirements.
 
+**`2025-11-18`** `[Picsi.ai]` Released Live Face Swap macOS & iOS App and updated [Picsi.ai](https://www.picsi.ai) services with our latest series of swap models (incl. [inswapper-512-live](https://github.com/deepinsight/inswapper-512-live)/Cyn/Dax).
 
-[Citation](#citation)
+**`2024-05-04`** `[Picsi.ai]` Released [InspireFace](cpp-package/inspireface), a cross-platform C/C++ face recognition SDK.
 
-[Contact](#contact)
+**`2022-08-12`**: We achieved Rank-1st of 
+[Perspective Projection Based Monocular 3D Face Reconstruction Challenge](https://tianchi.aliyun.com/competition/entrance/531961/introduction)
+of [ECCV-2022 WCPA Workshop](https://sites.google.com/view/wcpa2022), [paper](https://arxiv.org/abs/2208.07142) and [code](reconstruction/jmlr).
 
-## Deep Face Recognition
+**`2021-10-29`**: We achieved 1st place on the [VISA track](https://pages.nist.gov/frvt/plots/11/visa.html) of [NIST-FRVT 1:1](https://pages.nist.gov/frvt/html/frvt11.html) by using Partial FC (Xiang An, Jiankang Deng, Jia Guo).
 
-### Introduction
+## ChangeLogs
 
-In this module, we provide training data, network settings and loss designs for deep face recognition.
-The training data includes, but not limited to the cleaned MS1M, VGG2 and CASIA-Webface datasets, which were already packed in MXNet binary format.
-The network backbones include ResNet, MobilefaceNet, MobileNet, InceptionResNet_v2, DenseNet, etc..
-The loss functions include Softmax, SphereFace, CosineFace, ArcFace, Sub-Center ArcFace and Triplet (Euclidean/Angular) Loss.
+**`2026-07-27`** [InsightFace Server](server/) Added a simple AWS Rekognition alternative with accuracy-preserving INT8 embedding quantization and 50M+ image search on one RTX 5090 GPU.
 
-You can check the detail page of our work [ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace)(which accepted in CVPR-2019) and [SubCenter-ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/SubCenter-ArcFace)(which accepted in ECCV-2020).
+**`2025-11-18`** `[Picsi.ai]` Released Live Face Swap macOS & iOS App and updated [Picsi.ai](https://www.picsi.ai) services with our latest series of swap models (incl. [inswapper-live](https://github.com/deepinsight/inswapper-512-live)/Cyn/Dax).
 
-![margin penalty for target logit](https://github.com/deepinsight/insightface/raw/master/resources/arcface.png)
+**`2024-05-04`** `[Picsi.ai]` Released [InspireFace](cpp-package/inspireface), a cross-platform C/C++ face recognition SDK.
 
-Our method, ArcFace, was initially described in an [arXiv technical report](https://arxiv.org/abs/1801.07698). By using this module, you can simply achieve LFW 99.83%+ and Megaface 98%+ by a single model. This module can help researcher/engineer to develop deep face recognition algorithms quickly by only two steps: download the binary dataset and run the training script.
+**`2024-04-17`**: [Monocular Identity-Conditioned Facial Reflectance Reconstruction](https://arxiv.org/abs/2404.00301) accepted by [CVPR-2024](https://cvpr.thecvf.com/Conferences/2024).
 
-### Training Data
+**`2023-08-08`**: We released the implementation of [Generalizing Gaze Estimation with Weak-Supervision from Synthetic Views](https://arxiv.org/abs/2212.02997) at [reconstruction/gaze](reconstruction/gaze).
 
-All face images are aligned by ficial five landmarks and cropped to 112x112:
+**`2023-05-03`**: We have launched the ongoing version of wild face anti-spoofing challenge. See details [here](https://github.com/deepinsight/insightface/tree/master/challenges/cvpr23-fas-wild#updates).
 
-Please check [Dataset-Zoo](https://github.com/deepinsight/insightface/wiki/Dataset-Zoo) for detail information and dataset downloading.
+**`2023-02-13`**: We launch a large scale in the wild face anti-spoofing challenge on CVPR23 Workshop, see details at [challenges/cvpr23-fas-wild](challenges/cvpr23-fas-wild).
 
+**`2022-11-28`**: Single line code for facial identity swapping in our python packge ver 0.7, please check the example [here](examples/in_swapper).
 
-* Please check *src/data/face2rec2.py* on how to build a binary face dataset. You can either choose *MTCNN* or *RetinaFace* to align the faces.
+**`2022-10-28`**: [MFR-Ongoing](http://iccv21-mfr.com) website is refactored, please create issues if there's any bug.
 
-### Train
+**`2022-09-22`**: Now we have [web-demos](web-demos): [face-localization](http://demo.insightface.ai:7007/), [face-recognition](http://demo.insightface.ai:7008/), and [face-swapping](http://demo.insightface.ai:7009/).
 
-1. Install `MXNet` with GPU support (Python 3.X).
+**`2022-08-12`**: We achieved Rank-1st of 
+[Perspective Projection Based Monocular 3D Face Reconstruction Challenge](https://tianchi.aliyun.com/competition/entrance/531961/introduction)
+of [ECCV-2022 WCPA Workshop](https://sites.google.com/view/wcpa2022), [paper](https://arxiv.org/abs/2208.07142) and [code](reconstruction/jmlr).
 
-```
-pip install mxnet-cu101 # which should match your installed cuda version
-```
+**`2022-03-30`**: [Partial FC](https://arxiv.org/abs/2203.15565) accepted by CVPR-2022.
 
-2. Clone the InsightFace repository. We call the directory insightface as *`INSIGHTFACE_ROOT`*.
+**`2022-02-23`**: [SCRFD](detection/scrfd) accepted by [ICLR-2022](https://iclr.cc/Conferences/2022).
 
-```
-git clone --recursive https://github.com/deepinsight/insightface.git
-```
+**`2021-11-30`**: [MFR-Ongoing](challenges/mfr) challenge launched(same with IFRT), which is an extended version of [iccv21-mfr](challenges/iccv21-mfr).
 
-3. Download the training set (`MS1M-Arcface`) and place it in *`$INSIGHTFACE_ROOT/recognition/datasets/`*. Each training dataset includes at least following 6 files:
+**`2021-10-29`**: We achieved 1st place on the [VISA track](https://pages.nist.gov/frvt/plots/11/visa.html) of [NIST-FRVT 1:1](https://pages.nist.gov/frvt/html/frvt11.html) by using Partial FC (Xiang An, Jiankang Deng, Jia Guo).
 
-```Shell
-    faces_emore/
-       train.idx
-       train.rec
-       property
-       lfw.bin
-       cfp_fp.bin
-       agedb_30.bin
-```
+**`2021-10-11`**: [Leaderboard](https://insightface.ai/mfr21) of [ICCV21 - Masked Face Recognition Challenge](challenges/iccv21-mfr) released. Video: [Youtube](https://www.youtube.com/watch?v=lL-7l5t6x2w), [Bilibili](https://www.bilibili.com/video/BV15b4y1h79N/).
 
-The first three files are the training dataset while the last three files are verification sets.
+**`2021-06-05`**: We launch a [Masked Face Recognition Challenge & Workshop](challenges/iccv21-mfr) on ICCV 2021.
 
-4. Train deep face recognition models.
-In this part, we assume you are in the directory *`$INSIGHTFACE_ROOT/recognition/ArcFace/`*.
 
-Place and edit config file:
-```Shell
-cp sample_config.py config.py
-vim config.py # edit dataset path etc..
-```
 
-We give some examples below. Our experiments were conducted on the Tesla P40 GPU.
+## Introduction
 
-(1). Train ArcFace with LResNet100E-IR.
+[InsightFace](https://insightface.ai) is an open source 2D&3D deep face analysis toolbox, mainly based on PyTorch and MXNet. 
 
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r100 --loss arcface --dataset emore
-```
-It will output verification results of *LFW*, *CFP-FP* and *AgeDB-30* every 2000 batches. You can check all options in *config.py*.
-This model can achieve *LFW 99.83+* and *MegaFace 98.3%+*.
+Please check our [website](https://insightface.ai) for detail.
 
-(2). Train CosineFace with LResNet50E-IR.
+The master branch works with **PyTorch 1.6+** and/or **MXNet=1.6-1.8**, with **Python 3.x**.
 
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r50 --loss cosface --dataset emore
-```
+InsightFace efficiently implements a rich variety of state of the art algorithms of face recognition, face detection and face alignment, which optimized for both training and deployment.
 
-(3). Train Softmax with LMobileNet-GAP.
+## Quick Start
 
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network m1 --loss softmax --dataset emore
-```
+Please start with our [python-package](python-package/), for testing detection, recognition and alignment models on input images.
 
-(4). Fine-turn the above Softmax model with Triplet loss.
 
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network m1 --loss triplet --lr 0.005 --pretrained ./models/m1-softmax-emore,1
-```
+### ArcFace Video Demo
 
-(5). Training in model parallel acceleration.
 
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_parall.py --network r100 --loss arcface --dataset emore
-```
+[<img src=https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/facerecognitionfromvideo.PNG width="760" />](https://www.youtube.com/watch?v=y-D1tReryGA&t=81s)
 
-5. Verification results.
 
-*LResNet100E-IR* network trained on *MS1M-Arcface* dataset with ArcFace loss:
+Please click the image to watch the Youtube video. For Bilibili users, click [here](https://www.bilibili.com/video/av38041494?from=search&seid=11501833604850032313).
 
-| Method  | LFW(%) | CFP-FP(%) | AgeDB-30(%) |  
-| ------- | ------ | --------- | ----------- |  
-|  Ours   | 99.80+ | 98.0+     | 98.20+      |   
 
 
+## Projects
 
-### Pretrained Models
+The [page](https://insightface.ai/projects) on InsightFace website also describes all supported projects in InsightFace.
 
-You can use `$INSIGHTFACE/src/eval/verification.py` to test all the pre-trained models.
+You may also interested in some [challenges](https://insightface.ai/challenges) hold by InsightFace.
 
-**Please check [Model-Zoo](https://github.com/deepinsight/insightface/wiki/Model-Zoo) for more pretrained models.**
 
 
+## Face Recognition
 
-### Verification Results on Combined Margin
+### Introduction
 
-A combined margin method was proposed as a function of target logits value and original `θ`:
+In this module, we provide training data, network settings and loss designs for deep face recognition.
 
-```
-COM(θ) = cos(m_1*θ+m_2) - m_3
-```
+The supported methods are as follows:
 
-For training with `m1=1.0, m2=0.3, m3=0.2`, run following command:
-```
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r100 --loss combined --dataset emore
-```
+- [x] [ArcFace_mxnet (CVPR'2019)](recognition/arcface_mxnet)
+- [x] [ArcFace_torch (CVPR'2019)](recognition/arcface_torch)
+- [x] [SubCenter ArcFace (ECCV'2020)](recognition/subcenter_arcface)
+- [x] [PartialFC_mxnet (CVPR'2022)](recognition/partial_fc)
+- [x] [PartialFC_torch (CVPR'2022)](recognition/arcface_torch)
+- [x] [VPL (CVPR'2021)](recognition/vpl)
+- [x] [Arcface_oneflow](recognition/arcface_oneflow)
+- [x] [ArcFace_Paddle (CVPR'2019)](recognition/arcface_paddle)
 
-Results by using ``MS1M-IBUG(MS1M-V1)``
+Commonly used network backbones are included in most of the methods, such as IResNet, MobilefaceNet, MobileNet, InceptionResNet_v2, DenseNet, etc..
 
-| Method           | m1   | m2   | m3   | LFW   | CFP-FP | AgeDB-30 |
-| ---------------- | ---- | ---- | ---- | ----- | ------ | -------- |
-| W&F Norm Softmax | 1    | 0    | 0    | 99.28 | 88.50  | 95.13    |
-| SphereFace       | 1.5  | 0    | 0    | 99.76 | 94.17  | 97.30    |
-| CosineFace       | 1    | 0    | 0.35 | 99.80 | 94.4   | 97.91    |
-| ArcFace          | 1    | 0.5  | 0    | 99.83 | 94.04  | 98.08    |
-| Combined Margin  | 1.2  | 0.4  | 0    | 99.80 | 94.08  | 98.05    |
-| Combined Margin  | 1.1  | 0    | 0.35 | 99.81 | 94.50  | 98.08    |
-| Combined Margin  | 1    | 0.3  | 0.2  | 99.83 | 94.51  | 98.13    |
-| Combined Margin  | 0.9  | 0.4  | 0.15 | 99.83 | 94.20  | 98.16    |
 
-### Test on MegaFace
+### Datasets
 
-Please check *`$INSIGHTFACE_ROOT/Evaluation/megaface/`* to evaluate the model accuracy on Megaface. All aligned images were already provided.
+The training data includes, but not limited to the cleaned MS1M, VGG2 and CASIA-Webface datasets, which were already packed in MXNet binary format. Please [dataset](recognition/_datasets_) page for detail.
 
+### Evaluation
 
-### 512-D Feature Embedding
+We provide standard IJB and Megaface evaluation pipelines in [evaluation](recognition/_evaluation_)
 
-In this part, we assume you are in the directory *`$INSIGHTFACE_ROOT/deploy/`*. The input face image should be generally centre cropped. We use *RNet+ONet* of *MTCNN* to further align the image before sending it to the feature embedding network.
 
-1. Prepare a pre-trained model.
-2. Put the model under *`$INSIGHTFACE_ROOT/models/`*. For example, *`$INSIGHTFACE_ROOT/models/model-r100-ii`*.
-3. Run the test script *`$INSIGHTFACE_ROOT/deploy/test.py`*.
+### Pretrained Models
 
-For single cropped face image(112x112), total inference time is only 17ms on our testing server(Intel E5-2660 @ 2.00GHz, Tesla M40, *LResNet34E-IR*).
+**Please check [Model-Zoo](https://github.com/deepinsight/insightface/wiki/Model-Zoo) for more pretrained models.**
 
-### Third-party Re-implementation
+### Third-party Re-implementation of ArcFace
 
 - TensorFlow: [InsightFace_TF](https://github.com/auroua/InsightFace_TF)
 - TensorFlow: [tf-insightface](https://github.com/AIInAi/tf-insightface)
@@ -244,55 +173,108 @@ For single cropped face image(112x112), total inference time is only 17ms on our
 - Caffe: [arcface-caffe](https://github.com/xialuxi/arcface-caffe)
 - Caffe: [CombinedMargin-caffe](https://github.com/gehaocool/CombinedMargin-caffe)
 - Tensorflow: [InsightFace-tensorflow](https://github.com/luckycallor/InsightFace-tensorflow)
-- TensorRT: [wang-xinyu/tensorrtx](https://github.com/wang-xinyu/tensorrtx)
+- TensorRT: [wang-xinyu/tensorrtx](https://github.com/wang-xinyu/tensorrtx)  
+- TensorRT: [InsightFace-REST](https://github.com/SthPhoenix/InsightFace-REST)
+- ONNXRuntime C++: [ArcFace-ONNXRuntime](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/ort/cv/glint_arcface.cpp)
+- ONNXRuntime Go: [arcface-go](https://github.com/jack139/arcface-go)
+- MNN: [ArcFace-MNN](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/mnn/cv/mnn_glint_arcface.cpp)
+- TNN: [ArcFace-TNN](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/tnn/cv/tnn_glint_arcface.cpp)
+- NCNN: [ArcFace-NCNN](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/ncnn/cv/ncnn_glint_arcface.cpp)
 
 ## Face Detection
 
-### RetinaFace
+### Introduction
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/11513D05.jpg" width="640"/>
+</div>
 
-  RetinaFace is a practical single-stage [SOTA](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html) face detector which is initially introduced in [arXiv technical report](https://arxiv.org/abs/1905.00641) and then accepted by [CVPR 2020](https://openaccess.thecvf.com/content_CVPR_2020/html/Deng_RetinaFace_Single-Shot_Multi-Level_Face_Localisation_in_the_Wild_CVPR_2020_paper.html). We provide training code, training dataset, pretrained models and evaluation scripts. 
-  
-  ![demoimg1](https://github.com/deepinsight/insightface/blob/master/resources/11513D05.jpg)
-  
-  Please check [RetinaFace](https://github.com/deepinsight/insightface/tree/master/RetinaFace) for detail.
-  
-### RetinaFaceAntiCov 
- 
-  RetinaFaceAntiCov is an experimental module to identify face boxes with masks. Please check [RetinaFaceAntiCov](https://github.com/deepinsight/insightface/tree/master/RetinaFaceAntiCov) for detail.
-  
-  ![demoimg1](https://github.com/deepinsight/insightface/blob/master/resources/cov_test.jpg)
+In this module, we provide training data with annotation, network settings and loss designs for face detection training, evaluation and inference.
 
+The supported methods are as follows:
 
-## Face Alignment
+- [x] [RetinaFace (CVPR'2020)](detection/retinaface)
+- [x] [SCRFD (Arxiv'2021)](detection/scrfd)
+- [x] [blazeface_paddle](detection/blazeface_paddle)
 
-### DenseUNet
+[RetinaFace](detection/retinaface) is a practical single-stage face detector which is accepted by [CVPR 2020](https://openaccess.thecvf.com/content_CVPR_2020/html/Deng_RetinaFace_Single-Shot_Multi-Level_Face_Localisation_in_the_Wild_CVPR_2020_paper.html). We provide training code, training dataset, pretrained models and evaluation scripts. 
 
-  Please check the [Menpo](https://github.com/jiankangdeng/MenpoBenchmark) Benchmark and our [Dense U-Net](https://github.com/deepinsight/insightface/tree/master/alignment/heatmapReg) for detail. We also provide other network settings such as classic hourglass. You can find all of training code, training dataset and evaluation scripts there.
-  
-### CoordinateReg
+[SCRFD](detection/scrfd) is an efficient high accuracy face detection approach which is initialy described in [Arxiv](https://arxiv.org/abs/2105.04714). We provide an easy-to-use pipeline to train high efficiency face detectors with NAS supporting.
 
-  On the other hand, in contrast to heatmap based approaches, we provide some lightweight facial landmark models with fast coordinate regression. The input of these models is loose cropped face image while the output is the direct landmark coordinates. See detail at [alignment-coordinateReg](https://github.com/deepinsight/insightface/tree/master/alignment/coordinateReg). Now only pretrained models available.
-  
- <div align="center">
-	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/t1_out.jpg" alt="imagevis" width="800">
-</div>
 
+## Face Alignment
 
-<div align="center">
-	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/C_jiaguo.gif" alt="videovis" width="240">
+### Introduction
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/thumb_sdunet.png" width="600"/>
 </div>
 
+In this module, we provide datasets and training/inference pipelines for face alignment.
+
+Supported methods:
+
+- [x] [SDUNets (BMVC'2018)](alignment/heatmap)
+- [x] [SimpleRegression](alignment/coordinate_reg)
+
+
+[SDUNets](alignment/heatmap) is a heatmap based method which accepted on [BMVC](http://bmvc2018.org/contents/papers/0051.pdf).
+
+[SimpleRegression](alignment/coordinate_reg) provides very lightweight facial landmark models with fast coordinate regression. The input of these models is loose cropped face image while the output is the direct landmark coordinates.
+
 
 ## Citation
 
 If you find *InsightFace* useful in your research, please consider to cite the following related papers:
 
 ```
-@inproceedings{deng2019retinaface,
-title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
-author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
-booktitle={arxiv},
-year={2019}
+@inproceedings{ren2023pbidr,
+  title={Facial Geometric Detail Recovery via Implicit Representation},
+  author={Ren, Xingyu and Lattas, Alexandros and Gecer, Baris and Deng, Jiankang and Ma, Chao and Yang, Xiaokang},
+  booktitle={2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG)},  
+  year={2023}
+ }
+
+@article{guo2021sample,
+  title={Sample and Computation Redistribution for Efficient Face Detection},
+  author={Guo, Jia and Deng, Jiankang and Lattas, Alexandros and Zafeiriou, Stefanos},
+  journal={arXiv preprint arXiv:2105.04714},
+  year={2021}
+}
+
+@inproceedings{gecer2021ostec,
+  title={OSTeC: One-Shot Texture Completion},
+  author={Gecer, Baris and Deng, Jiankang and Zafeiriou, Stefanos},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year={2021}
+}
+
+@inproceedings{an_2022_pfc_cvpr,
+  title={Killing Two Birds with One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC},
+  author={An, Xiang and Deng, Jiangkang and Guo, Jia and Feng, Ziyong and Zhu, Xuhan and Jing, Yang and Tongliang, Liu},
+  booktitle={CVPR},
+  year={2022}
+}
+@inproceedings{an_2021_pfc_iccvw,
+  title={Partial FC: Training 10 Million Identities on a Single Machine},
+  author={An, Xiang and Zhu, Xuhan and Gao, Yuan and Xiao, Yang and Zhao, Yongle and Feng, Ziyong and Wu, Lan and Qin, Bin and Zhang, Ming and Zhang, Debing and Fu, Ying},
+  booktitle={ICCVW},
+  year={2021},
+}
+
+
+@inproceedings{deng2020subcenter,
+  title={Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces},
+  author={Deng, Jiankang and Guo, Jia and Liu, Tongliang and Gong, Mingming and Zafeiriou, Stefanos},
+  booktitle={Proceedings of the IEEE Conference on European Conference on Computer Vision},
+  year={2020}
+}
+
+@inproceedings{Deng2020CVPR,
+title = {RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild},
+author = {Deng, Jiankang and Guo, Jia and Ververas, Evangelos and Kotsia, Irene and Zafeiriou, Stefanos},
+booktitle = {CVPR},
+year = {2020}
 }
 
 @inproceedings{guo2018stacked,
@@ -317,9 +299,13 @@ year={2019}
 }
 ```
 
-## Contact
+## Contributing
 
-```
-[Jia Guo](guojia[at]gmail.com)
-[Jiankang Deng](jiankangdeng[at]gmail.com)
-```
+Main contributors:
+
+- [Jia Guo](https://github.com/nttstar), ``guojia[at]gmail.com``
+- [Jiankang Deng](https://github.com/jiankangdeng) ``jiankangdeng[at]gmail.com``
+- [Xiang An](https://github.com/anxiangsir) ``anxiangsir[at]gmail.com``
+- [Jack Yu](https://github.com/szad670401) ``jackyu961127[at]gmail.com``
+- [Baris Gecer](https://barisgecer.github.io/) ``barisgecer[at]msn.com``
+ ``
diff --git a/RetinaFace/README.md b/RetinaFace/README.md
deleted file mode 100644
index 65f62431f..000000000
--- a/RetinaFace/README.md
+++ /dev/null
@@ -1,84 +0,0 @@
-# RetinaFace Face Detector
-
-## Introduction
-
-RetinaFace is a practical single-stage [SOTA](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html) face detector which is initially introduced in [arXiv technical report](https://arxiv.org/abs/1905.00641) and then accepted by [CVPR 2020](https://openaccess.thecvf.com/content_CVPR_2020/html/Deng_RetinaFace_Single-Shot_Multi-Level_Face_Localisation_in_the_Wild_CVPR_2020_paper.html).
-
-![demoimg1](https://github.com/deepinsight/insightface/blob/master/resources/11513D05.jpg)
-
-![demoimg2](https://github.com/deepinsight/insightface/blob/master/resources/widerfacevaltest.png)
-
-## Data
-
-1. Download our annotations (face bounding boxes & five facial landmarks) from [baidu cloud](https://pan.baidu.com/s/1Laby0EctfuJGgGMgRRgykA) or [dropbox](https://www.dropbox.com/s/7j70r3eeepe4r2g/retinaface_gt_v1.1.zip?dl=0)
-
-2. Download the [WIDERFACE](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html) dataset.
-
-3. Organise the dataset directory under ``insightface/RetinaFace/`` as follows:
-
-```Shell
-  data/retinaface/
-    train/
-      images/
-      label.txt
-    val/
-      images/
-      label.txt
-    test/
-      images/
-      label.txt
-```
-
-## Install
-
-1. Install MXNet with GPU support.
-2. Install Deformable Convolution V2 operator from [Deformable-ConvNets](https://github.com/msracver/Deformable-ConvNets) if you use the DCN based backbone.
-3. Type ``make`` to build cxx tools.
-
-## Training
-
-Please check ``train.py`` for training.
-
-1. Copy ``rcnn/sample_config.py`` to ``rcnn/config.py``
-2. Download ImageNet pretrained models and put them into ``model/``(these models are not for detection testing/inferencing but training and parameters initialization). 
-
-    ImageNet ResNet50 ([baidu cloud](https://pan.baidu.com/s/1WAkU9ZA_j-OmzO-sdk9whA) and [dropbox](https://www.dropbox.com/s/48b850vmnaaasfl/imagenet-resnet-50.zip?dl=0)). 
-
-    ImageNet ResNet152 ([baidu cloud](https://pan.baidu.com/s/1nzQ6CzmdKFzg8bM8ChZFQg) and [dropbox](https://www.dropbox.com/s/8ypcra4nqvm32v6/imagenet-resnet-152.zip?dl=0)).
-
-3. Start training with ``CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --prefix ./model/retina --network resnet``.  
-Before training, you can check the ``resnet`` network configuration (e.g. pretrained model path, anchor setting and learning rate policy etc..) in ``rcnn/config.py``.
-4. We have two predefined network settings named ``resnet``(for medium and large models) and ``mnet``(for lightweight models).
-
-## Testing
-
-Please check ``test.py`` for testing.
-
-## RetinaFace Pretrained Models
-
-Pretrained Model: RetinaFace-R50 ([baidu cloud](https://pan.baidu.com/s/1C6nKq122gJxRhb37vK0_LQ) or [dropbox](https://www.dropbox.com/s/53ftnlarhyrpkg2/retinaface-R50.zip?dl=0)) is a medium size model with ResNet50 backbone.
-It can output face bounding boxes and five facial landmarks in a single forward pass.
-
-WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4. 
-
-To avoid the confliction with the WiderFace Challenge (ICCV 2019), we postpone the release time of our best model.
-
-## Third-party Models
-
-[yangfly](https://github.com/yangfly): RetinaFace-MobileNet0.25 ([baidu cloud](https://pan.baidu.com/s/1P1ypO7VYUbNAezdvLm2m9w)).
-WiderFace validation mAP: Hard 82.5. (model size: 1.68Mb) 
-
-[clancylian](https://github.com/clancylian/retinaface): C++ version
-
-## References
-
-```  
-@inproceedings{Deng2020CVPR,
-title = {RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild},
-author = {Deng, Jiankang and Guo, Jia and Ververas, Evangelos and Kotsia, Irene and Zafeiriou, Stefanos},
-booktitle = {CVPR},
-year = {2020}
-}
-```
-
-
diff --git a/RetinaFace/rcnn/PY_OP/cascade_refine.py b/RetinaFace/rcnn/PY_OP/cascade_refine.py
deleted file mode 100644
index fc93ee55f..000000000
--- a/RetinaFace/rcnn/PY_OP/cascade_refine.py
+++ /dev/null
@@ -1,453 +0,0 @@
-
-from __future__ import print_function
-import sys
-import mxnet as mx
-import numpy as np
-import datetime
-from distutils.util import strtobool
-from ..config import config, generate_config
-from ..processing.generate_anchor import generate_anchors, anchors_plane
-from ..processing.bbox_transform import bbox_overlaps, bbox_transform, landmark_transform
-
-STAT = {0:0}
-STEP = 28800
-
-class CascadeRefineOperator(mx.operator.CustomOp):
-    def __init__(self, stride=0, network='', dataset='', prefix=''):
-        super(CascadeRefineOperator, self).__init__()
-        self.stride = int(stride)
-        self.prefix = prefix
-        generate_config(network, dataset)
-        self.mode = config.TRAIN.OHEM_MODE #0 for random 10:245, 1 for 10:246, 2 for 10:30, mode 1 for default
-        stride = self.stride
-        sstride = str(stride)
-        base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
-        allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
-        ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
-        scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
-        base_anchors = generate_anchors(base_size=base_size, ratios=list(ratios), scales=np.array(scales, dtype=np.float32), stride = stride, dense_anchor = config.DENSE_ANCHOR)
-        num_anchors = base_anchors.shape[0]
-        feat_height, feat_width = config.SCALES[0][0]//self.stride, config.SCALES[0][0]//self.stride
-        feat_stride = self.stride
-
-        A = num_anchors
-        K = feat_height * feat_width
-        self.A = A
-
-        all_anchors = anchors_plane(feat_height, feat_width, feat_stride, base_anchors)
-        all_anchors = all_anchors.reshape((K * A, 4))
-        self.ori_anchors = all_anchors
-        self.nbatch = 0
-        global STAT
-        for k in config.RPN_FEAT_STRIDE:
-          STAT[k] = [0,0,0]
-
-    def apply_bbox_pred(self, bbox_pred, ind = None):
-      box_deltas = bbox_pred
-      box_deltas[:, 0::4] = box_deltas[:,0::4] * config.TRAIN.BBOX_STDS[0]
-      box_deltas[:, 1::4] = box_deltas[:,1::4] * config.TRAIN.BBOX_STDS[1]
-      box_deltas[:, 2::4] = box_deltas[:,2::4] * config.TRAIN.BBOX_STDS[2]
-      box_deltas[:, 3::4] = box_deltas[:,3::4] * config.TRAIN.BBOX_STDS[3]
-      if ind is None:
-        boxes = self.ori_anchors
-      else:
-        boxes = self.ori_anchors[ind]
-      #print('in apply',self.stride, box_deltas.shape, boxes.shape)
-
-      widths = boxes[:, 2] - boxes[:, 0] + 1.0
-      heights = boxes[:, 3] - boxes[:, 1] + 1.0
-      ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-      ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-      dx = box_deltas[:, 0:1]
-      dy = box_deltas[:, 1:2]
-      dw = box_deltas[:, 2:3]
-      dh = box_deltas[:, 3:4]
-
-      pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-      pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-      pred_w = np.exp(dw) * widths[:, np.newaxis]
-      pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-      pred_boxes = np.zeros(box_deltas.shape)
-      # x1
-      pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-      # y1
-      pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-      # x2
-      pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-      # y2
-      pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-      return pred_boxes
-
-    def assign_anchor_fpn(self, gt_label, anchors, landmark=False, prefix='face'):
-      IOU = config.TRAIN.CASCADE_OVERLAP
-
-      gt_boxes = gt_label['gt_boxes']
-      #_label = gt_label['gt_label']
-      # clean up boxes
-      #nonneg = np.where(_label[:] != -1)[0]
-      #gt_boxes = gt_boxes[nonneg]
-      if landmark:
-        gt_landmarks = gt_label['gt_landmarks']
-        #gt_landmarks = gt_landmarks[nonneg]
-        assert gt_boxes.shape[0]==gt_landmarks.shape[0]
-      #scales = np.array(scales, dtype=np.float32)
-      feat_strides = config.RPN_FEAT_STRIDE
-      bbox_pred_len = 4
-      landmark_pred_len = 10
-      num_anchors = anchors.shape[0]
-      A = self.A
-      total_anchors = num_anchors
-      feat_height, feat_width = config.SCALES[0][0]//self.stride, config.SCALES[0][0]//self.stride
-
-
-      #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
-
-      # label: 1 is positive, 0 is negative, -1 is dont care
-      labels = np.empty((num_anchors,), dtype=np.float32)
-      labels.fill(-1)
-      #print('BB', anchors.shape, len(inds_inside))
-      #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
-      #tb = datetime.datetime.now()
-      #self._times[0] += (tb-ta).total_seconds()
-      #ta = datetime.datetime.now()
-
-      if gt_boxes.size > 0:
-          # overlap between the anchors and the gt boxes
-          # overlaps (ex, gt)
-          overlaps = bbox_overlaps(anchors.astype(np.float), gt_boxes.astype(np.float))
-          argmax_overlaps = overlaps.argmax(axis=1)
-          #print('AAA', argmax_overlaps.shape)
-          max_overlaps = overlaps[np.arange(num_anchors), argmax_overlaps]
-          gt_argmax_overlaps = overlaps.argmax(axis=0)
-          gt_max_overlaps = overlaps[gt_argmax_overlaps, np.arange(overlaps.shape[1])]
-          gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
-
-          if not config.TRAIN.RPN_CLOBBER_POSITIVES:
-              # assign bg labels first so that positive labels can clobber them
-              labels[max_overlaps < IOU[0]] = 0
-
-          # fg label: for each gt, anchor with highest overlap
-          if config.TRAIN.RPN_FORCE_POSITIVE:
-            labels[gt_argmax_overlaps] = 1
-
-          # fg label: above threshold IoU
-          labels[max_overlaps >= IOU[1]] = 1
-
-          if config.TRAIN.RPN_CLOBBER_POSITIVES:
-              # assign bg labels last so that negative labels can clobber positives
-              labels[max_overlaps < IOU[0]] = 0
-      else:
-          labels[:] = 0
-      fg_inds = np.where(labels == 1)[0]
-      #print('fg count', len(fg_inds))
-
-      # subsample positive labels if we have too many
-      if config.TRAIN.RPN_ENABLE_OHEM==0:
-        fg_inds = np.where(labels == 1)[0]
-        num_fg = int(config.TRAIN.RPN_FG_FRACTION * config.TRAIN.RPN_BATCH_SIZE)
-        if len(fg_inds) > num_fg:
-            disable_inds = npr.choice(fg_inds, size=(len(fg_inds) - num_fg), replace=False)
-            if DEBUG:
-                disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
-            labels[disable_inds] = -1
-
-        # subsample negative labels if we have too many
-        num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
-        bg_inds = np.where(labels == 0)[0]
-        if len(bg_inds) > num_bg:
-            disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-            if DEBUG:
-                disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-            labels[disable_inds] = -1
-
-        #fg_inds = np.where(labels == 1)[0]
-        #num_fg = len(fg_inds)
-        #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
-
-        #bg_inds = np.where(labels == 0)[0]
-        #if len(bg_inds) > num_bg:
-        #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-        #    if DEBUG:
-        #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-        #    labels[disable_inds] = -1
-      else:
-        fg_inds = np.where(labels == 1)[0]
-        num_fg = len(fg_inds)
-        bg_inds = np.where(labels == 0)[0]
-        num_bg = len(bg_inds)
-
-      #print('anchor stat', num_fg, num_bg)
-
-
-      bbox_targets = np.zeros((num_anchors, bbox_pred_len), dtype=np.float32)
-      if gt_boxes.size > 0:
-          #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
-          bbox_targets[:,:] = bbox_transform(anchors, gt_boxes[argmax_overlaps, :])
-          #bbox_targets[:,4] = gt_blur
-      #tb = datetime.datetime.now()
-      #self._times[1] += (tb-ta).total_seconds()
-      #ta = datetime.datetime.now()
-
-      bbox_weights = np.zeros((num_anchors, bbox_pred_len), dtype=np.float32)
-      #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
-      bbox_weights[labels == 1, 0:4] = 1.0
-      if bbox_pred_len>4:
-        bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
-
-      if landmark:
-        landmark_targets = np.zeros((num_anchors, landmark_pred_len), dtype=np.float32)
-        landmark_weights = np.zeros((num_anchors, landmark_pred_len), dtype=np.float32)
-        #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
-        if landmark_pred_len==10:
-          landmark_weights[labels == 1, :] = 1.0
-        elif landmark_pred_len==15:
-          v = [1.0, 1.0, 0.1] * 5
-          assert len(v)==15
-          landmark_weights[labels == 1, :] = np.array(v)
-        else:
-          assert False
-        #TODO here
-        if gt_landmarks.size > 0:
-          #print('AAA',argmax_overlaps)
-          a_landmarks = gt_landmarks[argmax_overlaps,:,:]
-          landmark_targets[:] = landmark_transform(anchors, a_landmarks)
-          invalid = np.where(a_landmarks[:,0,2]<0.0)[0]
-          #assert len(invalid)==0
-          #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
-          landmark_weights[invalid, :] = 0.0
-      #tb = datetime.datetime.now()
-      #self._times[2] += (tb-ta).total_seconds()
-      #ta = datetime.datetime.now()
-      bbox_targets[:, 0::4] = bbox_targets[:,0::4] / config.TRAIN.BBOX_STDS[0]
-      bbox_targets[:, 1::4] = bbox_targets[:,1::4] / config.TRAIN.BBOX_STDS[1]
-      bbox_targets[:, 2::4] = bbox_targets[:,2::4] / config.TRAIN.BBOX_STDS[2]
-      bbox_targets[:, 3::4] = bbox_targets[:,3::4] / config.TRAIN.BBOX_STDS[3]
-
-      #print('CC', anchors.shape, len(inds_inside))
-      label = {}
-      _label = labels.reshape((1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
-      _label = _label.reshape((1, A * feat_height * feat_width))
-      bbox_target = bbox_targets.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose(0, 2, 1)
-      bbox_weight = bbox_weights.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose((0, 2, 1))
-      label['%s_label'%prefix] = _label[0]
-      label['%s_bbox_target'%prefix] = bbox_target[0]
-      label['%s_bbox_weight'%prefix] = bbox_weight[0]
-      if landmark:
-        landmark_target = landmark_target.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose(0, 2, 1)
-        landmark_target /= config.TRAIN.LANDMARK_STD
-        landmark_weight = landmark_weight.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose((0, 2, 1))
-        label['%s_landmark_target'%prefix] = landmark_target[0]
-        label['%s_landmark_weight'%prefix] = landmark_weight[0]
-
-      return label
-
-    def forward(self, is_train, req, in_data, out_data, aux):
-        self.nbatch+=1
-        ta = datetime.datetime.now()
-        global STAT
-        A = config.NUM_ANCHORS
-
-        cls_label_t0 = in_data[0].asnumpy() #BS, AHW
-        cls_score_t0 = in_data[1].asnumpy() #BS, C, AHW
-        cls_score = in_data[2].asnumpy() #BS, C, AHW
-        #labels_raw = in_data[1].asnumpy() #BS, ANCHORS
-        bbox_pred_t0 = in_data[3].asnumpy() #BS, AC, HW
-        bbox_target_t0 = in_data[4].asnumpy() #BS, AC, HW
-        cls_label_raw = in_data[5].asnumpy() #BS, AHW
-        gt_boxes = in_data[6].asnumpy() #BS, N, C=4+1
-        #imgs = in_data[7].asnumpy().astype(np.uint8)
-
-        batch_size = cls_score.shape[0]
-        num_anchors = cls_score.shape[2]
-        #print('in cas', cls_score.shape, bbox_target.shape)
-
-        labels_out = np.zeros(shape=(batch_size, num_anchors), dtype=np.float32)
-        bbox_target_out = np.zeros(shape=bbox_target_t0.shape, dtype=np.float32)
-        anchor_weight = np.zeros( (batch_size, num_anchors,1), dtype=np.float32 )
-        valid_count = np.zeros( (batch_size,1), dtype=np.float32 )
-
-        bbox_pred_t0 = bbox_pred_t0.transpose( (0,2,1) )
-        bbox_pred_t0 = bbox_pred_t0.reshape( (bbox_pred_t0.shape[0], -1, 4) ) #BS, H*W*A, C
-        bbox_target_t0 = bbox_target_t0.transpose( (0,2,1) )
-        bbox_target_t0 = bbox_target_t0.reshape( (bbox_target_t0.shape[0], -1, 4) )
-
-        #print('anchor_weight', anchor_weight.shape)
-
-        #assert labels.shape[0]==1
-        #assert cls_score.shape[0]==1
-        #assert bbox_weight.shape[0]==1
-        #print('shape', cls_score.shape, labels.shape, file=sys.stderr)
-        #print('bbox_weight 0', bbox_weight.shape, file=sys.stderr)
-        #bbox_weight = np.zeros( (labels_raw.shape[0], labels_raw.shape[1], 4), dtype=np.float32)
-        _stat = [0,0,0]
-        SEL_TOPK = config.TRAIN.RPN_BATCH_SIZE
-        FAST = False
-        for ibatch in range(batch_size):
-          #bgr = imgs[ibatch].transpose( (1,2,0) )[:,:,::-1]
-
-          if not FAST:
-            _gt_boxes = gt_boxes[ibatch] #N, 4+1
-            _gtind = len(np.where(_gt_boxes[:,4]>=0)[0])
-            #print('gt num', _gtind)
-            _gt_boxes = _gt_boxes[0:_gtind,:]
-
-            #anchors_t1 = self.ori_anchors.copy()
-            #_cls_label_raw = cls_label_raw[ibatch] #AHW
-            #_cls_label_raw = _cls_label_raw.reshape( (A, -1) ).transpose( (1,0) ).reshape( (-1,) ) #HWA
-            #fg_ind_raw = np.where(_cls_label_raw>0)[0]
-            #_bbox_target_t0 = bbox_target_t0[ibatch][fg_ind_raw]
-            #_bbox_pred_t0 = bbox_pred_t0[ibatch][fg_ind_raw]
-            #anchors_t1_pos = self.apply_bbox_pred(_bbox_pred_t0, ind=fg_ind_raw)
-            #anchors_t1[fg_ind_raw,:] = anchors_t1_pos
-
-            anchors_t1 = self.apply_bbox_pred(bbox_pred_t0[ibatch])
-            assert anchors_t1.shape[0]==self.ori_anchors.shape[0]
-
-            #for i in range(_gt_boxes.shape[0]):
-            #  box = _gt_boxes[i].astype(np.int)
-            #  print('%d: gt%d'%(self.nbatch, i), box)
-            #  #color = (0,0,255)
-            #  #cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
-            #for i in range(anchors_t1.shape[0]):
-            #  box1 = self.ori_anchors[i].astype(np.int)
-            #  box2 = anchors_t1[i].astype(np.int)
-            #  print('%d %d: anchorscompare %d'%(self.nbatch, self.stride, i), box1, box2)
-              #color = (255,255,0)
-              #cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
-            #filename = "./debug/%d_%d_%d.jpg"%(self.nbatch, ibatch, stride)
-            #cv2.imwrite(filename, img)
-            #print(filename)
-            #gt_label = {'gt_boxes': gt_anchors, 'gt_label' : labels_raw[ibatch]}
-            gt_label = {'gt_boxes': _gt_boxes}
-            new_label_dict = self.assign_anchor_fpn(gt_label, anchors_t1, False, prefix=self.prefix)
-            labels = new_label_dict['%s_label'%self.prefix] #AHW
-            new_bbox_target = new_label_dict['%s_bbox_target'%self.prefix] #AC,HW
-            #print('assign ret', labels.shape, new_bbox_target.shape)
-            _anchor_weight = np.zeros( (num_anchors,1), dtype=np.float32)
-            fg_score = cls_score[ibatch,1,:] - cls_score[ibatch,0,:]
-            fg_inds = np.where(labels>0)[0]
-            num_fg = int(config.TRAIN.RPN_FG_FRACTION * config.TRAIN.RPN_BATCH_SIZE)
-            origin_num_fg = len(fg_inds)
-            #continue
-            #print('cas fg', len(fg_inds), num_fg, file=sys.stderr)
-            if len(fg_inds) > num_fg:
-              if self.mode==0:
-                disable_inds = np.random.choice(fg_inds, size=(len(fg_inds) - num_fg), replace=False)
-                labels[disable_inds] = -1
-              else:
-                pos_ohem_scores = fg_score[fg_inds]
-                order_pos_ohem_scores = pos_ohem_scores.ravel().argsort()
-                sampled_inds = fg_inds[order_pos_ohem_scores[:num_fg]]
-                labels[fg_inds] = -1
-                labels[sampled_inds] = 1
-
-            n_fg = np.sum(labels>0)
-            fg_inds = np.where(labels>0)[0]
-            num_bg = config.TRAIN.RPN_BATCH_SIZE - n_fg
-            if self.mode==2:
-              num_bg = max(48, n_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1))
-
-            bg_inds = np.where(labels == 0)[0]
-            origin_num_bg = len(bg_inds)
-            if num_bg==0:
-              labels[bg_inds] = -1
-            elif len(bg_inds) > num_bg:
-              # sort ohem scores
-
-              if self.mode==0:
-                disable_inds = np.random.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-                labels[disable_inds] = -1
-              else:
-                neg_ohem_scores = fg_score[bg_inds]
-                order_neg_ohem_scores = neg_ohem_scores.ravel().argsort()[::-1]
-                sampled_inds = bg_inds[order_neg_ohem_scores[:num_bg]]
-                #print('sampled_inds_bg', sampled_inds, file=sys.stderr)
-                labels[bg_inds] = -1
-                labels[sampled_inds] = 0
-
-            if n_fg>0:
-              order0_labels = labels.reshape( (1, A, -1) ).transpose( (0, 2, 1) ).reshape( (-1,) )
-              bbox_fg_inds = np.where(order0_labels>0)[0]
-              #print('bbox_fg_inds, order0 ', bbox_fg_inds, file=sys.stderr)
-              _anchor_weight[bbox_fg_inds,:] = 1.0
-            anchor_weight[ibatch] = _anchor_weight
-            valid_count[ibatch][0] = n_fg
-            labels_out[ibatch] = labels
-            #print('labels_out', self.stride, ibatch, labels)
-            bbox_target_out[ibatch] = new_bbox_target
-            #print('cascade stat', self.stride, ibatch, len(labels), len(np.where(labels==1)[0]), len(np.where(labels==0)[0]))
-          else: #FAST MODE
-            fg_score_t0 = cls_score_t0[ibatch,1,:] - cls_score_t0[ibatch,0,:]
-            sort_idx_t0 = np.argsort(fg_score_t0.flatten())[::-1][0:SEL_TOPK]
-            _bbox_pred_t0 = bbox_pred_t0[ibatch][sort_idx_t0]
-            _bbox_target_t0 = bbox_target_t0[ibatch][sort_idx_t0]
-            #print('SEL fg score:', fg_score_t0[sort_idx[-1]], fg_score_t0[sort_idx[0]])
-            anchors_t0 = self.apply_bbox_pred(_bbox_pred_t0)
-            gt_anchors = self.apply_bbox_pred(_bbox_target_t0)
-            #gt_label = {'gt_boxes': gt_anchors, 'gt_label' : labels_raw[ibatch]}
-            gt_label = {'gt_boxes': gt_anchors}
-            new_label_dict = self.assign_anchor_fpn(gt_label, anchors_t0, False, prefix=self.prefix)
-            labels = new_label_dict['%s_label'%self.prefix]
-            new_bbox_target = new_label_dict['%s_bbox_target'%self.prefix]
-            #print('assign ret', labels.shape, new_bbox_target.shape)
-            _anchor_weight = np.zeros( (num_anchors,1), dtype=np.float32)
-            fg_score = cls_score[ibatch,1,:] - cls_score[ibatch,0,:]
-            fg_inds = np.where(labels>0)[0]
-            _labels = np.empty(shape=labels.shape, dtype=np.float32)
-            _labels.fill(-1)
-            _labels[sort_idx_idx] = labels
-
-            anchor_weight[ibatch] = _anchor_weight
-            valid_count[ibatch][0] = len(fg_inds)
-            labels_out[ibatch] = _labels
-            #print('labels_out', self.stride, ibatch, labels)
-            bbox_target_out[ibatch] = new_bbox_target
-
-        #print('cascade pos stat', self.stride, batch_size, np.sum(valid_count))
-        for ind, val in enumerate([labels_out, bbox_target_out, anchor_weight, valid_count]):
-            val = mx.nd.array(val)
-            self.assign(out_data[ind], req[ind], val)
-        tb = datetime.datetime.now()
-        #print('cascade forward cost', self.stride, (tb-ta).total_seconds())
-
-    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
-        for i in range(len(in_grad)):
-            self.assign(in_grad[i], req[i], 0)
-
-
-@mx.operator.register('cascade_refine')
-class CascadeRefineProp(mx.operator.CustomOpProp):
-    def __init__(self, stride=0, network='', dataset='', prefix=''):
-        super(CascadeRefineProp, self).__init__(need_top_grad=False)
-        self.stride = stride
-        self.network=network
-        self.dataset=dataset
-        self.prefix = prefix
-
-    def list_arguments(self):
-        #return ['cls_label_t0', 'cls_pred_t0', 'cls_pred', 'bbox_pred_t0', 'bbox_label_t0', 'cls_label_raw', 'cas_gt_boxes', 'cas_img']
-        return ['cls_label_t0', 'cls_pred_t0', 'cls_pred', 'bbox_pred_t0', 'bbox_label_t0', 'cls_label_raw', 'cas_gt_boxes']
-
-    def list_outputs(self):
-        return ['cls_label_out', 'bbox_label_out', 'anchor_weight_out', 'pos_count_out']
-
-    def infer_shape(self, in_shape):
-        cls_pred_shape = in_shape[1]
-        bs = cls_pred_shape[0]
-        num_anchors = cls_pred_shape[2]
-        #print('in_rpn_ohem', in_shape[0], in_shape[1], in_shape[2], file=sys.stderr)
-        #print('in_rpn_ohem', labels_shape, anchor_weight_shape)
-        cls_label_shape = [bs, num_anchors]
-
-        return in_shape, \
-               [cls_label_shape, in_shape[4], [bs,num_anchors,1], [bs,1]]
-
-    def create_operator(self, ctx, shapes, dtypes):
-        return CascadeRefineOperator(self.stride, self.network, self.dataset, self.prefix)
-
-    def declare_backward_dependency(self, out_grad, in_data, out_data):
-        return []
-
-
diff --git a/RetinaFace/rcnn/PY_OP/rpn_fpn_ohem3.py b/RetinaFace/rcnn/PY_OP/rpn_fpn_ohem3.py
deleted file mode 100644
index 330db5772..000000000
--- a/RetinaFace/rcnn/PY_OP/rpn_fpn_ohem3.py
+++ /dev/null
@@ -1,167 +0,0 @@
-
-from __future__ import print_function
-import sys
-import mxnet as mx
-import numpy as np
-from distutils.util import strtobool
-from ..config import config, generate_config
-
-
-STAT = {0:0}
-STEP = 28800
-
-class RPNFPNOHEM3Operator(mx.operator.CustomOp):
-    def __init__(self, stride=0, network='', dataset='', prefix=''):
-        super(RPNFPNOHEM3Operator, self).__init__()
-        self.stride = int(stride)
-        self.prefix = prefix
-        generate_config(network, dataset)
-        self.mode = config.TRAIN.OHEM_MODE #0 for random 10:245, 1 for 10:246, 2 for 10:30, mode 1 for default
-        global STAT
-        for k in config.RPN_FEAT_STRIDE:
-          STAT[k] = [0,0,0]
-
-    def forward(self, is_train, req, in_data, out_data, aux):
-        global STAT
-
-        cls_score    = in_data[0].asnumpy() #BS, 2, ANCHORS
-        labels_raw       = in_data[1].asnumpy() # BS, ANCHORS
-
-        A = config.NUM_ANCHORS
-        anchor_weight = np.zeros( (labels_raw.shape[0], labels_raw.shape[1],1), dtype=np.float32 )
-        valid_count = np.zeros( (labels_raw.shape[0],1), dtype=np.float32 )
-        #print('anchor_weight', anchor_weight.shape)
-
-        #assert labels.shape[0]==1
-        #assert cls_score.shape[0]==1
-        #assert bbox_weight.shape[0]==1
-        #print('shape', cls_score.shape, labels.shape, file=sys.stderr)
-        #print('bbox_weight 0', bbox_weight.shape, file=sys.stderr)
-        #bbox_weight = np.zeros( (labels_raw.shape[0], labels_raw.shape[1], 4), dtype=np.float32)
-        _stat = [0,0,0]
-        for ibatch in range(labels_raw.shape[0]):
-          _anchor_weight = np.zeros( (labels_raw.shape[1],1), dtype=np.float32)
-          labels = labels_raw[ibatch]
-          fg_score = cls_score[ibatch,1,:] - cls_score[ibatch,0,:]
-
-
-
-          fg_inds = np.where(labels>0)[0]
-          num_fg = int(config.TRAIN.RPN_FG_FRACTION * config.TRAIN.RPN_BATCH_SIZE)
-          origin_num_fg = len(fg_inds)
-          #print(len(fg_inds), num_fg, file=sys.stderr)
-          if len(fg_inds) > num_fg:
-            if self.mode==0:
-              disable_inds = np.random.choice(fg_inds, size=(len(fg_inds) - num_fg), replace=False)
-              labels[disable_inds] = -1
-            else:
-              pos_ohem_scores = fg_score[fg_inds]
-              order_pos_ohem_scores = pos_ohem_scores.ravel().argsort()
-              sampled_inds = fg_inds[order_pos_ohem_scores[:num_fg]]
-              labels[fg_inds] = -1
-              labels[sampled_inds] = 1
-
-          n_fg = np.sum(labels>0)
-          fg_inds = np.where(labels>0)[0]
-          num_bg = config.TRAIN.RPN_BATCH_SIZE - n_fg
-          if self.mode==2:
-            num_bg = max(48, n_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1))
-
-          bg_inds = np.where(labels == 0)[0]
-          origin_num_bg = len(bg_inds)
-          if num_bg==0:
-            labels[bg_inds] = -1
-          elif len(bg_inds) > num_bg:
-            # sort ohem scores
-
-            if self.mode==0:
-              disable_inds = np.random.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-              labels[disable_inds] = -1
-            else:
-              neg_ohem_scores = fg_score[bg_inds]
-              order_neg_ohem_scores = neg_ohem_scores.ravel().argsort()[::-1]
-              sampled_inds = bg_inds[order_neg_ohem_scores[:num_bg]]
-              #print('sampled_inds_bg', sampled_inds, file=sys.stderr)
-              labels[bg_inds] = -1
-              labels[sampled_inds] = 0
-
-          if n_fg>0:
-            order0_labels = labels.reshape( (1, A, -1) ).transpose( (0, 2, 1) ).reshape( (-1,) )
-            bbox_fg_inds = np.where(order0_labels>0)[0]
-            #print('bbox_fg_inds, order0 ', bbox_fg_inds, file=sys.stderr)
-            _anchor_weight[bbox_fg_inds,:] = 1.0
-          anchor_weight[ibatch] = _anchor_weight
-          valid_count[ibatch][0] = n_fg
-
-          #if self.prefix=='face':
-          #  #print('fg-bg', self.stride, n_fg, num_bg)
-          #  STAT[0]+=1
-          #  STAT[self.stride][0] += config.TRAIN.RPN_BATCH_SIZE
-          #  STAT[self.stride][1] += n_fg
-          #  STAT[self.stride][2] += np.sum(fg_score[fg_inds]>=0)
-          #  #_stat[0] += config.TRAIN.RPN_BATCH_SIZE
-          #  #_stat[1] += n_fg
-          #  #_stat[2] += np.sum(fg_score[fg_inds]>=0)
-          #  #print('stride num_fg', self.stride, n_fg, file=sys.stderr)
-          #  #ACC[self.stride] += np.sum(fg_score[fg_inds]>=0)
-          #  #x = float(labels_raw.shape[0]*len(config.RPN_FEAT_STRIDE))
-          #  x = 1.0
-          #  if STAT[0]%STEP==0:
-          #    _str = ['STAT']
-          #    STAT[0] = 0
-          #    for k in config.RPN_FEAT_STRIDE:
-          #      acc = float(STAT[k][2])/STAT[k][1]
-          #      acc0 = float(STAT[k][1])/STAT[k][0]
-          #      #_str.append("%d: all-fg(%d, %d, %.4f), fg-fgcorrect(%d, %d, %.4f)"%(k,STAT[k][0], STAT[k][1], acc0, STAT[k][1], STAT[k][2], acc))
-          #      _str.append("%d: (%d, %d, %.4f)"%(k, STAT[k][1], STAT[k][2], acc))
-          #      STAT[k] = [0,0,0]
-          #    _str = ' | '.join(_str)
-          #    print(_str, file=sys.stderr)
-          #if self.stride==4 and num_fg>0:
-          #  print('_stat_', self.stride, num_fg, num_bg, file=sys.stderr)
-
-        #labels_ohem = mx.nd.array(labels_raw)
-        #anchor_weight = mx.nd.array(anchor_weight)
-        #print('valid_count', self.stride, np.sum(valid_count))
-        #print('_stat', _stat, valid_count)
-
-        for ind, val in enumerate([labels_raw, anchor_weight, valid_count]):
-            val = mx.nd.array(val)
-            self.assign(out_data[ind], req[ind], val)
-
-    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
-        for i in range(len(in_grad)):
-            self.assign(in_grad[i], req[i], 0)
-
-
-@mx.operator.register('rpn_fpn_ohem3')
-class RPNFPNOHEM3Prop(mx.operator.CustomOpProp):
-    def __init__(self, stride=0, network='', dataset='', prefix=''):
-        super(RPNFPNOHEM3Prop, self).__init__(need_top_grad=False)
-        self.stride = stride
-        self.network=network
-        self.dataset=dataset
-        self.prefix = prefix
-
-    def list_arguments(self):
-        return ['cls_score', 'labels']
-
-    def list_outputs(self):
-        return ['labels_ohem', 'anchor_weight', 'valid_count']
-
-    def infer_shape(self, in_shape):
-        labels_shape = in_shape[1]
-        #print('in_rpn_ohem', in_shape[0], in_shape[1], in_shape[2], file=sys.stderr)
-        anchor_weight_shape = [labels_shape[0], labels_shape[1], 1]
-        #print('in_rpn_ohem', labels_shape, anchor_weight_shape)
-
-        return in_shape, \
-               [labels_shape, anchor_weight_shape, [labels_shape[0], 1]]
-
-    def create_operator(self, ctx, shapes, dtypes):
-        return RPNFPNOHEM3Operator(self.stride, self.network, self.dataset, self.prefix)
-
-    def declare_backward_dependency(self, out_grad, in_data, out_data):
-        return []
-
-
diff --git a/RetinaFace/rcnn/core/callback.py b/RetinaFace/rcnn/core/callback.py
deleted file mode 100644
index 317d5cda2..000000000
--- a/RetinaFace/rcnn/core/callback.py
+++ /dev/null
@@ -1,13 +0,0 @@
-import mxnet as mx
-
-
-def do_checkpoint(prefix, means, stds):
-    def _callback(iter_no, sym, arg, aux):
-      if 'bbox_pred_weight' in arg:
-        arg['bbox_pred_weight_test'] = (arg['bbox_pred_weight'].T * mx.nd.array(stds)).T
-        arg['bbox_pred_bias_test'] = arg['bbox_pred_bias'] * mx.nd.array(stds) + mx.nd.array(means)
-      mx.model.save_checkpoint(prefix, iter_no + 1, sym, arg, aux)
-      if 'bbox_pred_weight' in arg:
-        arg.pop('bbox_pred_weight_test')
-        arg.pop('bbox_pred_bias_test')
-    return _callback
diff --git a/RetinaFace/rcnn/core/loader.py b/RetinaFace/rcnn/core/loader.py
deleted file mode 100644
index 019ae758e..000000000
--- a/RetinaFace/rcnn/core/loader.py
+++ /dev/null
@@ -1,478 +0,0 @@
-from __future__ import print_function
-import sys
-import mxnet as mx
-import numpy as np
-import random
-import datetime
-import multiprocessing
-import cv2
-from mxnet.executor_manager import _split_input_slice
-
-from rcnn.config import config
-from rcnn.io.image import tensor_vstack
-from rcnn.io.rpn import get_rpn_testbatch, get_rpn_batch, assign_anchor_fpn, get_crop_batch, AA
-
-
-class CropLoader(mx.io.DataIter):
-    def __init__(self, feat_sym, roidb, batch_size=1, shuffle=False, ctx=None, work_load_list=None,
-                 aspect_grouping=False):
-        """
-        This Iter will provide roi data to Fast R-CNN network
-        :param feat_sym: to infer shape of assign_output
-        :param roidb: must be preprocessed
-        :param batch_size: must divide BATCH_SIZE(128)
-        :param shuffle: bool
-        :param ctx: list of contexts
-        :param work_load_list: list of work load
-        :param aspect_grouping: group images with similar aspects
-        :return: AnchorLoader
-        """
-        super(CropLoader, self).__init__()
-
-        # save parameters as properties
-        self.feat_sym = feat_sym
-        self.roidb = roidb
-        self.batch_size = batch_size
-        self.shuffle = shuffle
-        self.ctx = ctx
-        if self.ctx is None:
-            self.ctx = [mx.cpu()]
-        self.work_load_list = work_load_list
-        #self.feat_stride = feat_stride
-        #self.anchor_scales = anchor_scales
-        #self.anchor_ratios = anchor_ratios
-        #self.allowed_border = allowed_border
-        self.aspect_grouping = aspect_grouping
-        self.feat_stride = config.RPN_FEAT_STRIDE
-
-        # infer properties from roidb
-        self.size = len(roidb)
-        self.index = np.arange(self.size)
-
-        # decide data and label names
-        #self.data_name = ['data']
-        #self.label_name = []
-        #self.label_name.append('label')
-        #self.label_name.append('bbox_target')
-        #self.label_name.append('bbox_weight')
-
-        self.data_name = ['data']
-        #self.label_name = ['label', 'bbox_target', 'bbox_weight']
-        self.label_name = []
-        prefixes = ['face']
-        if config.HEAD_BOX:
-          prefixes.append('head')
-        names = []
-        for prefix in prefixes:
-          names += [prefix+'_label', prefix+'_bbox_target', prefix+'_bbox_weight']
-          if prefix=='face' and config.FACE_LANDMARK:
-            names += [prefix+'_landmark_target', prefix+'_landmark_weight']
-        #names = ['label', 'bbox_weight']
-        for stride in self.feat_stride:
-          for n in names:
-            k = "%s_stride%d"%(n,stride)
-            self.label_name.append(k)
-        if config.CASCADE>0:
-          self.label_name.append('gt_boxes')
-
-        # status variable for synchronization between get_data and get_label
-        self.cur = 0
-        self.batch = None
-        self.data = None
-        self.label = None
-        # infer shape
-        feat_shape_list = []
-        _data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-        _data_shape = dict(_data_shape)
-        for i in range(len(self.feat_stride)):
-            _, feat_shape, _ = self.feat_sym[i].infer_shape(**_data_shape)
-            feat_shape = [int(i) for i in feat_shape[0]]
-            feat_shape_list.append(feat_shape)
-        self.aa = AA(feat_shape_list)
-
-        self._debug = False
-        self._debug_id = 0
-        self._times = [0.0, 0.0, 0.0, 0.0]
-
-        # get first batch to fill in provide_data and provide_label
-        self.reset()
-        self.get_batch()
-
-    @property
-    def provide_data(self):
-        return [(k, v.shape) for k, v in zip(self.data_name, self.data)]
-
-    @property
-    def provide_label(self):
-        return [(k, v.shape) for k, v in zip(self.label_name, self.label)]
-
-    def reset(self):
-        self.cur = 0
-        if self.shuffle:
-          np.random.shuffle(self.index)
-
-    def iter_next(self):
-        return self.cur + self.batch_size <= self.size
-
-    def next(self):
-        if self.iter_next():
-            self.get_batch()
-            self.cur += self.batch_size
-            return mx.io.DataBatch(data=self.data, label=self.label,
-                                   pad=self.getpad(), index=self.getindex(),
-                                   provide_data=self.provide_data, provide_label=self.provide_label)
-        else:
-            raise StopIteration
-
-    def getindex(self):
-        return self.cur / self.batch_size
-
-    def getpad(self):
-        if self.cur + self.batch_size > self.size:
-            return self.cur + self.batch_size - self.size
-        else:
-            return 0
-
-    def infer_shape(self, max_data_shape=None, max_label_shape=None):
-        """ Return maximum data and label shape for single gpu """
-        if max_data_shape is None:
-            max_data_shape = []
-        if max_label_shape is None:
-            max_label_shape = []
-        max_shapes = dict(max_data_shape + max_label_shape)
-        input_batch_size = max_shapes['data'][0]
-        dummy_boxes = np.zeros((0, 5))
-        dummy_info = [ [max_shapes['data'][2], max_shapes['data'][3], 1.0] ]
-        dummy_label = {'gt_boxes' : dummy_boxes}
-        dummy_blur = np.zeros((0,))
-        dummy_label['gt_blur'] = dummy_blur
-
-
-        label_dict = {}
-        if config.HEAD_BOX:
-          head_label_dict = self.aa.assign_anchor_fpn(dummy_label, dummy_info, False, prefix='head')
-          label_dict.update(head_label_dict)
-
-        if config.FACE_LANDMARK:
-          dummy_landmarks = np.zeros( (0,5,3) )
-          dummy_label['gt_landmarks'] = dummy_landmarks
-        face_label_dict = self.aa.assign_anchor_fpn(dummy_label, dummy_info, config.FACE_LANDMARK, prefix='face')
-        label_dict.update(face_label_dict)
-        if config.CASCADE>0:
-          label_dict['gt_boxes'] = np.zeros((0, config.TRAIN.MAX_BBOX_PER_IMAGE, 5), dtype=np.float32)
-
-        label_list = []
-        for k in self.label_name:
-          label_list.append(label_dict[k])
-        label_shape = [(k, tuple([input_batch_size] + list(v.shape[1:]))) for k, v in zip(self.label_name, label_list)]
-        return max_data_shape, label_shape
-
-    def get_batch(self):
-        # slice roidb
-        cur_from = self.cur
-        cur_to = min(cur_from + self.batch_size, self.size)
-        assert cur_to==cur_from+self.batch_size
-        roidb = [self.roidb[self.index[i]] for i in range(cur_from, cur_to)]
-
-        # decide multi device slice
-        work_load_list = self.work_load_list
-        ctx = self.ctx
-        if work_load_list is None:
-            work_load_list = [1] * len(ctx)
-        assert isinstance(work_load_list, list) and len(work_load_list) == len(ctx), \
-            "Invalid settings for work load. "
-        slices = _split_input_slice(self.batch_size, work_load_list)
-
-        # get testing data for multigpu
-        data_list = []
-        label_list = []
-        for islice in slices:
-            iroidb = [roidb[i] for i in range(islice.start, islice.stop)]
-            data, label = get_crop_batch(iroidb)
-            data_list += data
-            label_list += label
-            #data_list.append(data)
-            #label_list.append(label)
-
-        # pad data first and then assign anchor (read label)
-        #data_tensor = tensor_vstack([batch['data'] for batch in data_list])
-        #for i_card in range(len(data_list)):
-        #    data_list[i_card]['data'] = data_tensor[
-        #                                i_card * config.TRAIN.BATCH_IMAGES:(1 + i_card) * config.TRAIN.BATCH_IMAGES]
-
-        #iiddxx = 0
-        select_stride = 0
-        if config.RANDOM_FEAT_STRIDE:
-          select_stride = random.choice(config.RPN_FEAT_STRIDE)
-
-        for data, label in zip(data_list, label_list):
-            data_shape = {k: v.shape for k, v in data.items()}
-            del data_shape['im_info']
-            feat_shape_list = []
-            for s in range(len(self.feat_stride)):
-                _, feat_shape, _ = self.feat_sym[s].infer_shape(**data_shape)
-                feat_shape = [int(i) for i in feat_shape[0]]
-                feat_shape_list.append(feat_shape)
-            im_info = data['im_info']
-            gt_boxes = label['gt_boxes']
-            gt_label = {'gt_boxes':gt_boxes}
-            if config.USE_BLUR:
-              gt_blur = label['gt_blur']
-              gt_label['gt_blur'] = gt_blur
-            if self._debug:
-              img = data['data'].copy()[0].transpose( (1,2,0) )[:,:,::-1].copy()
-              print('DEBUG SHAPE', data['data'].shape, label['gt_boxes'].shape)
-
-              box = label['gt_boxes'].copy()[0][0:4].astype(np.int)
-              cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), (0, 255, 0), 2)
-              filename = './debugout/%d.png' % (self._debug_id)
-              print('debug write', filename)
-              cv2.imwrite(filename, img)
-              self._debug_id+=1
-              #print('DEBUG', img.shape, bbox.shape)
-            label_dict = {}
-            if config.HEAD_BOX:
-              head_label_dict = self.aa.assign_anchor_fpn(gt_label, im_info, False, prefix='head', select_stride = select_stride)
-              label_dict.update(head_label_dict)
-            if config.FACE_LANDMARK:
-              gt_landmarks = label['gt_landmarks']
-              gt_label['gt_landmarks'] = gt_landmarks
-            #ta = datetime.datetime.now()
-            #face_label_dict = assign_anchor_fpn(feat_shape_list, gt_label, im_info, config.FACE_LANDMARK, prefix='face', select_stride = select_stride)
-            face_label_dict = self.aa.assign_anchor_fpn(gt_label, im_info, config.FACE_LANDMARK, prefix='face', select_stride = select_stride)
-            #tb = datetime.datetime.now()
-            #self._times[0] += (tb-ta).total_seconds()
-            label_dict.update(face_label_dict)
-            #for k in label_dict:
-            #  print(k, label_dict[k].shape)
-
-            if config.CASCADE>0:
-              pad_gt_boxes = np.empty( (1, config.TRAIN.MAX_BBOX_PER_IMAGE, 5), dtype=np.float32)
-              pad_gt_boxes.fill(-1)
-              pad_gt_boxes[0, 0:gt_boxes.shape[0],:] = gt_boxes
-              label_dict['gt_boxes'] = pad_gt_boxes
-            #print('im_info', im_info.shape)
-            #print(gt_boxes.shape)
-            for k in self.label_name:
-              label[k] = label_dict[k]
-
-        all_data = dict()
-        for key in self.data_name:
-            all_data[key] = tensor_vstack([batch[key] for batch in data_list])
-
-        all_label = dict()
-        for key in self.label_name:
-            pad = 0 if key.startswith('bbox_') else -1
-            #print('label vstack', key, pad, len(label_list), file=sys.stderr)
-            all_label[key] = tensor_vstack([batch[key] for batch in label_list], pad=pad)
-
-        self.data = [mx.nd.array(all_data[key]) for key in self.data_name]
-        self.label = [mx.nd.array(all_label[key]) for key in self.label_name]
-        #for _label in self.label:
-        #  print('LABEL SHAPE', _label.shape)
-        #print(self._times)
-
-class CropLoader2(mx.io.DataIter):
-    def __init__(self, feat_sym, roidb, batch_size=1, shuffle=False, ctx=None, work_load_list=None,
-                 aspect_grouping=False):
-        """
-        This Iter will provide roi data to Fast R-CNN network
-        :param feat_sym: to infer shape of assign_output
-        :param roidb: must be preprocessed
-        :param batch_size: must divide BATCH_SIZE(128)
-        :param shuffle: bool
-        :param ctx: list of contexts
-        :param work_load_list: list of work load
-        :param aspect_grouping: group images with similar aspects
-        :return: AnchorLoader
-        """
-        super(CropLoader2, self).__init__()
-
-        # save parameters as properties
-        self.feat_sym = feat_sym
-        self.roidb = roidb
-        self.batch_size = batch_size
-        self.shuffle = shuffle
-        self.ctx = ctx
-        if self.ctx is None:
-            self.ctx = [mx.cpu()]
-        self.work_load_list = work_load_list
-        #self.feat_stride = feat_stride
-        #self.anchor_scales = anchor_scales
-        #self.anchor_ratios = anchor_ratios
-        #self.allowed_border = allowed_border
-        self.aspect_grouping = aspect_grouping
-        self.feat_stride = config.RPN_FEAT_STRIDE
-
-        # infer properties from roidb
-        self.size = len(roidb)
-
-        # decide data and label names
-        #self.data_name = ['data']
-        #self.label_name = []
-        #self.label_name.append('label')
-        #self.label_name.append('bbox_target')
-        #self.label_name.append('bbox_weight')
-
-        self.data_name = ['data']
-        #self.label_name = ['label', 'bbox_target', 'bbox_weight']
-        self.label_name = []
-        prefixes = ['face']
-        if config.HEAD_BOX:
-          prefixes.append('head')
-        names = []
-        for prefix in prefixes:
-          names += [prefix+'_label', prefix+'_bbox_target', prefix+'_bbox_weight']
-          if prefix=='face' and config.FACE_LANDMARK:
-            names += [prefix+'_landmark_target', prefix+'_landmark_weight']
-        #names = ['label', 'bbox_weight']
-        for stride in self.feat_stride:
-          for n in names:
-            k = "%s_stride%d"%(n,stride)
-            self.label_name.append(k)
-        # status variable for synchronization between get_data and get_label
-        self.cur = 0
-        self.batch = None
-        self.data = None
-        self.label = None
-
-        # get first batch to fill in provide_data and provide_label
-        self.reset()
-        self.q_in = [multiprocessing.Queue(1024) for i in range(config.NUM_CPU)]
-        #self.q_in = multiprocessing.Queue(1024)
-        self.q_out = multiprocessing.Queue(1024)
-        self.start()
-        self.get_batch()
-
-    @property
-    def provide_data(self):
-        return [(k, v.shape) for k, v in zip(self.data_name, self.data)]
-
-    @property
-    def provide_label(self):
-        return [(k, v.shape) for k, v in zip(self.label_name, self.label)]
-
-    def reset(self):
-      pass
-
-    @staticmethod
-    def input_worker(q_in, roidb, batch_size):
-      index = np.arange(len(roidb))
-      np.random.shuffle(index)
-      cur_from = 0
-      while True:
-        cur_to = cur_from + batch_size
-        if cur_to>len(roidb):
-          np.random.shuffle(index)
-          cur_from = 0
-          continue
-        _roidb = [roidb[index[i]] for i in range(cur_from, cur_to)]
-        istart = index[cur_from]
-        q_in[istart%len(q_in)].put(_roidb)
-        cur_from = cur_to
-
-    @staticmethod
-    def gen_worker(q_in, q_out):
-      while True:
-        deq = q_in.get()
-        if deq is None:
-          break
-        _roidb = deq
-        data, label = get_crop_batch(_roidb)
-        print('generated')
-        q_out.put( (data, label) )
-
-    def start(self):
-      input_process = multiprocessing.Process(target=CropLoader2.input_worker, args=(self.q_in, self.roidb, self.batch_size))
-      #gen_process = multiprocessing.Process(target=gen_worker, args=(q_in, q_out))
-      gen_process = [multiprocessing.Process(target=CropLoader2.gen_worker, args=(self.q_in[i], self.q_out)) \
-                for i in range(config.NUM_CPU)]
-      input_process.start()
-      for p in gen_process:
-        p.start()
-
-
-    def next(self):
-        self.get_batch()
-        return mx.io.DataBatch(data=self.data, label=self.label,
-                               provide_data=self.provide_data, provide_label=self.provide_label)
-
-    def infer_shape(self, max_data_shape=None, max_label_shape=None):
-        """ Return maximum data and label shape for single gpu """
-        if max_data_shape is None:
-            max_data_shape = []
-        if max_label_shape is None:
-            max_label_shape = []
-        max_shapes = dict(max_data_shape + max_label_shape)
-        input_batch_size = max_shapes['data'][0]
-        dummy_boxes = np.zeros((0, 5))
-        dummy_info = [ [max_shapes['data'][2], max_shapes['data'][3], 1.0] ]
-        dummy_label = {'gt_boxes' : dummy_boxes}
-
-        # infer shape
-        feat_shape_list = []
-        for i in range(len(self.feat_stride)):
-            _, feat_shape, _ = self.feat_sym[i].infer_shape(**max_shapes)
-            feat_shape = [int(i) for i in feat_shape[0]]
-            feat_shape_list.append(feat_shape)
-
-        label_dict = {}
-        if config.HEAD_BOX:
-          head_label_dict = assign_anchor_fpn(feat_shape_list, dummy_label, dummy_info, False, prefix='head')
-          label_dict.update(head_label_dict)
-
-        if config.FACE_LANDMARK:
-          dummy_landmarks = np.zeros( (0,11) )
-          dummy_label['gt_landmarks'] = dummy_landmarks
-        face_label_dict = assign_anchor_fpn(feat_shape_list, dummy_label, dummy_info, config.FACE_LANDMARK, prefix='face')
-        label_dict.update(face_label_dict)
-
-        label_list = []
-        for k in self.label_name:
-          label_list.append(label_dict[k])
-        label_shape = [(k, tuple([input_batch_size] + list(v.shape[1:]))) for k, v in zip(self.label_name, label_list)]
-        return max_data_shape, label_shape
-
-    def get_batch(self):
-        deq = self.q_out.get()
-        print('q_out got')
-        data_list, label_list = deq
-
-        for data, label in zip(data_list, label_list):
-            data_shape = {k: v.shape for k, v in data.items()}
-            del data_shape['im_info']
-            feat_shape_list = []
-            for s in range(len(self.feat_stride)):
-                _, feat_shape, _ = self.feat_sym[s].infer_shape(**data_shape)
-                feat_shape = [int(i) for i in feat_shape[0]]
-                feat_shape_list.append(feat_shape)
-            #for k in self.label_name:
-            #  label[k] = [0 for i in range(config.TRAIN.BATCH_IMAGES)]
-            im_info = data['im_info']
-            gt_boxes = label['gt_boxes']
-            gt_label = {'gt_boxes':gt_boxes}
-            label_dict = {}
-            head_label_dict = assign_anchor_fpn(feat_shape_list, gt_label, im_info, False, prefix='head')
-            label_dict.update(head_label_dict)
-            if config.FACE_LANDMARK:
-              gt_landmarks = label['gt_landmarks']
-              gt_label['gt_landmarks'] = gt_landmarks
-            face_label_dict = assign_anchor_fpn(feat_shape_list, gt_label, im_info, config.FACE_LANDMARK, prefix='face')
-            label_dict.update(face_label_dict)
-            #print('im_info', im_info.shape)
-            #print(gt_boxes.shape)
-            for k in self.label_name:
-              label[k] = label_dict[k]
-
-        all_data = dict()
-        for key in self.data_name:
-            all_data[key] = tensor_vstack([batch[key] for batch in data_list])
-
-        all_label = dict()
-        for key in self.label_name:
-            pad = 0 if key.startswith('bbox_') else -1
-            #print('label vstack', key, pad, len(label_list), file=sys.stderr)
-            all_label[key] = tensor_vstack([batch[key] for batch in label_list], pad=pad)
-        self.data = [mx.nd.array(all_data[key]) for key in self.data_name]
-        self.label = [mx.nd.array(all_label[key]) for key in self.label_name]
-
diff --git a/RetinaFace/rcnn/core/metric.py b/RetinaFace/rcnn/core/metric.py
deleted file mode 100644
index d600b20e5..000000000
--- a/RetinaFace/rcnn/core/metric.py
+++ /dev/null
@@ -1,168 +0,0 @@
-from __future__ import print_function
-import sys
-import mxnet as mx
-import numpy as np
-
-from rcnn.config import config
-
-
-def get_rpn_names():
-    pred = ['rpn_cls_prob', 'rpn_bbox_loss', 'rpn_label', 'rpn_bbox_weight']
-    label = ['rpn_label', 'rpn_bbox_target', 'rpn_bbox_weight']
-    return pred, label
-
-
-
-class RPNAccMetric(mx.metric.EvalMetric):
-    def __init__(self, pred_idx=-1, label_idx=-1,name='RPNAcc'):
-        super(RPNAccMetric, self).__init__(name)
-        self.pred, self.label = get_rpn_names()
-        #self.name = 'RPNAcc'
-        self.name = [name, name+'_BG', name+'_FG']
-        self.pred_idx = pred_idx
-        self.label_idx = label_idx
-        self.STAT = [0, 0, 0]
-
-    def reset(self):
-        """Clear the internal statistics to initial state."""
-        if isinstance(self.name, str):
-          self.num_inst = 0
-          self.sum_metric = 0.0
-        else:
-          #print('reset to ',len(self.name), self.name, file=sys.stderr)
-          self.num_inst = [0] * len(self.name)
-          self.sum_metric = [0.0] * len(self.name)
-
-
-    def get(self):
-        if isinstance(self.name, str):
-            if self.num_inst == 0:
-                return (self.name, float('nan'))
-            else:
-                return (self.name, self.sum_metric / self.num_inst)
-        else:
-            names = ['%s'%(self.name[i]) for i in range(len(self.name))]
-            values = [x / y if y != 0 else float('nan') \
-                for x, y in zip(self.sum_metric, self.num_inst)]
-            return (names, values)
-
-    def update(self, labels, preds):
-        if self.pred_idx>=0 and self.label_idx>=0:
-          pred = preds[self.pred_idx]
-          label = preds[self.label_idx]
-        else:
-          pred = preds[self.pred.index('rpn_cls_prob')]
-          label = labels[self.label.index('rpn_label')]
-          #label = preds[self.pred.index('rpn_label')]
-
-        num_images = pred.shape[0]
-        #print(pred.shape, label.shape, file=sys.stderr)
-        # pred (b, c, p) or (b, c, h, w)
-        pred_label = mx.ndarray.argmax_channel(pred).asnumpy().astype('int32')
-        #pred_label = pred_label.reshape((pred_label.shape[0], -1))
-        pred_label = pred_label.reshape(-1,)
-        # label (b, p)
-        label = label.asnumpy().astype('int32').reshape(-1,)
-        #print(pred_label.shape, label.shape)
-
-        # filter with keep_inds
-        keep_inds = np.where(label != -1)[0]
-        #print('in_metric acc', pred_label.shape, label.shape, len(keep_inds), file=sys.stderr)
-        #print(keep_inds, file=sys.stderr)
-        _pred_label = pred_label[keep_inds]
-        _label = label[keep_inds]
-        #print('in_metric2', pred_label.shape, label.shape, len(keep_inds), file=sys.stderr)
-        if isinstance(self.name, str):
-          self.sum_metric += np.sum(_pred_label.flat == _label.flat)
-          self.num_inst += len(_pred_label.flat)
-        else:
-          self.sum_metric[0] += np.sum(_pred_label.flat == _label.flat)
-          self.num_inst[0] += len(_pred_label.flat)
-
-          keep_inds = np.where(label == 0)[0]
-          _pred_label = pred_label[keep_inds]
-          _label = label[keep_inds]
-          self.sum_metric[1] += np.sum(_pred_label.flat == _label.flat)
-          self.num_inst[1] += len(_pred_label.flat)
-
-          keep_inds = np.where(label == 1)[0]
-          _pred_label = pred_label[keep_inds]
-          _label = label[keep_inds]
-          a = np.sum(_pred_label.flat == _label.flat)
-          b = len(_pred_label.flat)
-          self.sum_metric[2] += a
-          self.num_inst[2] += b
-
-          #self.STAT[0]+=a
-          #self.STAT[1]+=b
-          #self.STAT[2]+=num_images
-          #if self.STAT[2]%400==0:
-          #  print('FG_ACC', self.pred_idx, self.STAT[2], self.STAT[0], self.STAT[1], float(self.STAT[0])/self.STAT[1], file=sys.stderr)
-          #  self.STAT = [0,0,0]
-
-
-class RPNLogLossMetric(mx.metric.EvalMetric):
-    def __init__(self, pred_idx=-1, label_idx=-1):
-        super(RPNLogLossMetric, self).__init__('RPNLogLoss')
-        self.pred, self.label = get_rpn_names()
-        self.pred_idx = pred_idx
-        self.label_idx = label_idx
-
-    def update(self, labels, preds):
-        if self.pred_idx>=0 and self.label_idx>=0:
-          pred = preds[self.pred_idx]
-          label = preds[self.label_idx]
-        else:
-          pred = preds[self.pred.index('rpn_cls_prob')]
-          label = labels[self.label.index('rpn_label')]
-          #label = preds[self.pred.index('rpn_label')]
-
-        # label (b, p)
-        label = label.asnumpy().astype('int32').reshape((-1))
-        # pred (b, c, p) or (b, c, h, w) --> (b, p, c) --> (b*p, c)
-        pred = pred.asnumpy().reshape((pred.shape[0], pred.shape[1], -1)).transpose((0, 2, 1))
-        pred = pred.reshape((label.shape[0], -1))
-
-        # filter with keep_inds
-        keep_inds = np.where(label != -1)[0]
-        label = label[keep_inds]
-        cls = pred[keep_inds, label]
-        #print('in_metric log', label.shape, cls.shape, file=sys.stderr)
-
-        cls += 1e-14
-        cls_loss = -1 * np.log(cls)
-        cls_loss = np.sum(cls_loss)
-        self.sum_metric += cls_loss
-        self.num_inst += label.shape[0]
-
-
-class RPNL1LossMetric(mx.metric.EvalMetric):
-    def __init__(self, loss_idx=-1, weight_idx=-1, name='RPNL1Loss'):
-        super(RPNL1LossMetric, self).__init__(name)
-        self.pred, self.label = get_rpn_names()
-        self.loss_idx = loss_idx
-        self.weight_idx = weight_idx
-        self.name = name
-
-    def update(self, labels, preds):
-        if self.loss_idx>=0 and self.weight_idx>=0:
-          bbox_loss = preds[self.loss_idx].asnumpy()
-          bbox_weight = preds[self.weight_idx].asnumpy()
-        else:
-          bbox_loss = preds[self.pred.index('rpn_bbox_loss')].asnumpy()
-          bbox_weight = labels[self.label.index('rpn_bbox_weight')].asnumpy()
-          #bbox_weight = preds[self.pred.index('rpn_bbox_weight')].asnumpy()
-
-        #print('in_metric', self.name, bbox_weight.shape, bbox_loss.shape)
-
-        # calculate num_inst (average on those fg anchors)
-        if config.LR_MODE==0:
-          num_inst = np.sum(bbox_weight > 0) / (bbox_weight.shape[1]/config.NUM_ANCHORS)
-        else:
-          num_inst = 1
-        #print('in_metric log', bbox_loss.shape, num_inst, file=sys.stderr)
-
-        self.sum_metric += np.sum(bbox_loss)
-        self.num_inst += num_inst
-
-
diff --git a/RetinaFace/rcnn/core/module.py b/RetinaFace/rcnn/core/module.py
deleted file mode 100644
index bf28f8ee5..000000000
--- a/RetinaFace/rcnn/core/module.py
+++ /dev/null
@@ -1,215 +0,0 @@
-"""A `MutableModule` implement the `BaseModule` API, and allows input shape
-varying with training iterations. If shapes vary, executors will rebind,
-using shared arrays from the initial module binded with maximum shape.
-"""
-
-import logging
-
-from mxnet import context as ctx
-from mxnet.initializer import Uniform
-from mxnet.module.base_module import BaseModule
-from mxnet.module.module import Module
-
-class MutableModule(BaseModule):
-    """A mutable module is a module that supports variable input data.
-
-    Parameters
-    ----------
-    symbol : Symbol
-    data_names : list of str
-    label_names : list of str
-    logger : Logger
-    context : Context or list of Context
-    work_load_list : list of number
-    max_data_shapes : list of (name, shape) tuple, designating inputs whose shape vary
-    max_label_shapes : list of (name, shape) tuple, designating inputs whose shape vary
-    fixed_param_prefix : list of str, indicating fixed parameters
-    """
-    def __init__(self, symbol, data_names, label_names,
-                 logger=logging, context=ctx.cpu(), work_load_list=None,
-                 max_data_shapes=None, max_label_shapes=None, fixed_param_prefix=None):
-        super(MutableModule, self).__init__(logger=logger)
-        self._symbol = symbol
-        self._data_names = data_names
-        self._label_names = label_names
-        self._context = context
-        self._work_load_list = work_load_list
-
-        self._curr_module = None
-        self._max_data_shapes = max_data_shapes
-        self._max_label_shapes = max_label_shapes
-        self._fixed_param_prefix = fixed_param_prefix
-
-        fixed_param_names = list()
-        if fixed_param_prefix is not None:
-            for name in self._symbol.list_arguments():
-                for prefix in self._fixed_param_prefix:
-                    if prefix in name:
-                        fixed_param_names.append(name)
-        self._fixed_param_names = fixed_param_names
-
-    def _reset_bind(self):
-        self.binded = False
-        self._curr_module = None
-
-    @property
-    def data_names(self):
-        return self._data_names
-
-    @property
-    def output_names(self):
-        return self._symbol.list_outputs()
-
-    @property
-    def data_shapes(self):
-        assert self.binded
-        return self._curr_module.data_shapes
-
-    @property
-    def label_shapes(self):
-        assert self.binded
-        return self._curr_module.label_shapes
-
-    @property
-    def output_shapes(self):
-        assert self.binded
-        return self._curr_module.output_shapes
-
-    def get_params(self):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_params()
-
-    def init_params(self, initializer=Uniform(0.01), arg_params=None, aux_params=None,
-                    allow_missing=False, force_init=False, allow_extra=False):
-        if self.params_initialized and not force_init:
-            return
-        assert self.binded, 'call bind before initializing the parameters'
-        self._curr_module.init_params(initializer=initializer, arg_params=arg_params,
-                                      aux_params=aux_params, allow_missing=allow_missing,
-                                      force_init=force_init, allow_extra=allow_extra)
-        self.params_initialized = True
-
-    def bind(self, data_shapes, label_shapes=None, for_training=True,
-             inputs_need_grad=False, force_rebind=False, shared_module=None):
-        # in case we already initialized params, keep it
-        if self.params_initialized:
-            arg_params, aux_params = self.get_params()
-
-        # force rebinding is typically used when one want to switch from
-        # training to prediction phase.
-        if force_rebind:
-            self._reset_bind()
-
-        if self.binded:
-            self.logger.warning('Already binded, ignoring bind()')
-            return
-
-        assert shared_module is None, 'shared_module for MutableModule is not supported'
-
-        self.for_training = for_training
-        self.inputs_need_grad = inputs_need_grad
-        self.binded = True
-
-        max_shapes_dict = dict()
-        if self._max_data_shapes is not None:
-            max_shapes_dict.update(dict(self._max_data_shapes))
-        if self._max_label_shapes is not None:
-            max_shapes_dict.update(dict(self._max_label_shapes))
-
-        max_data_shapes = list()
-        for name, shape in data_shapes:
-            if name in max_shapes_dict:
-                max_data_shapes.append((name, max_shapes_dict[name]))
-            else:
-                max_data_shapes.append((name, shape))
-
-        max_label_shapes = list()
-        if label_shapes is not None:
-            for name, shape in label_shapes:
-                if name in max_shapes_dict:
-                    max_label_shapes.append((name, max_shapes_dict[name]))
-                else:
-                    max_label_shapes.append((name, shape))
-
-        if len(max_label_shapes) == 0:
-            max_label_shapes = None
-
-        module = Module(self._symbol, self._data_names, self._label_names, logger=self.logger,
-                        context=self._context, work_load_list=self._work_load_list,
-                        fixed_param_names=self._fixed_param_names)
-        module.bind(max_data_shapes, max_label_shapes, for_training, inputs_need_grad,
-                    force_rebind=False, shared_module=None)
-        self._curr_module = module
-
-        # copy back saved params, if already initialized
-        if self.params_initialized:
-            self.set_params(arg_params, aux_params)
-
-    def init_optimizer(self, kvstore='local', optimizer='sgd',
-                       optimizer_params=(('learning_rate', 0.01),), force_init=False):
-        assert self.binded and self.params_initialized
-        if self.optimizer_initialized and not force_init:
-            self.logger.warning('optimizer already initialized, ignoring.')
-            return
-
-        self._curr_module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                         force_init=force_init)
-        self.optimizer_initialized = True
-
-    def forward(self, data_batch, is_train=None):
-        assert self.binded and self.params_initialized
-
-        # get current_shapes
-        if self._curr_module.label_shapes is not None:
-            current_shapes = dict(self._curr_module.data_shapes + self._curr_module.label_shapes)
-        else:
-            current_shapes = dict(self._curr_module.data_shapes)
-
-        # get input_shapes
-        if data_batch.provide_label is not None:
-            input_shapes = dict(data_batch.provide_data + data_batch.provide_label)
-        else:
-            input_shapes = dict(data_batch.provide_data)
-
-        # decide if shape changed
-        shape_changed = False
-        for k, v in current_shapes.items():
-            if v != input_shapes[k]:
-                shape_changed = True
-
-        if shape_changed:
-            module = Module(self._symbol, self._data_names, self._label_names,
-                            logger=self.logger, context=self._context,
-                            work_load_list=self._work_load_list,
-                            fixed_param_names=self._fixed_param_names)
-            module.bind(data_batch.provide_data, data_batch.provide_label, self._curr_module.for_training,
-                        self._curr_module.inputs_need_grad, force_rebind=False,
-                        shared_module=self._curr_module)
-            self._curr_module = module
-
-        self._curr_module.forward(data_batch, is_train=is_train)
-
-    def backward(self, out_grads=None):
-        assert self.binded and self.params_initialized
-        self._curr_module.backward(out_grads=out_grads)
-
-    def update(self):
-        assert self.binded and self.params_initialized and self.optimizer_initialized
-        self._curr_module.update()
-
-    def get_outputs(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_outputs(merge_multi_context=merge_multi_context)
-
-    def get_input_grads(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized and self.inputs_need_grad
-        return self._curr_module.get_input_grads(merge_multi_context=merge_multi_context)
-
-    def update_metric(self, eval_metric, labels):
-        assert self.binded and self.params_initialized
-        self._curr_module.update_metric(eval_metric, labels)
-
-    def install_monitor(self, mon):
-        """ Install monitor on all executors """
-        assert self.binded
-        self._curr_module.install_monitor(mon)
diff --git a/RetinaFace/rcnn/core/module_bak.py b/RetinaFace/rcnn/core/module_bak.py
deleted file mode 100644
index 569d50dd9..000000000
--- a/RetinaFace/rcnn/core/module_bak.py
+++ /dev/null
@@ -1,215 +0,0 @@
-"""A `MutableModule` implement the `BaseModule` API, and allows input shape
-varying with training iterations. If shapes vary, executors will rebind,
-using shared arrays from the initial module binded with maximum shape.
-"""
-
-import logging
-
-from mxnet import context as ctx
-from mxnet.initializer import Uniform
-from mxnet.module.base_module import BaseModule
-from mxnet.module.module import Module
-
-class MutableModule(BaseModule):
-    """A mutable module is a module that supports variable input data.
-
-    Parameters
-    ----------
-    symbol : Symbol
-    data_names : list of str
-    label_names : list of str
-    logger : Logger
-    context : Context or list of Context
-    work_load_list : list of number
-    max_data_shapes : list of (name, shape) tuple, designating inputs whose shape vary
-    max_label_shapes : list of (name, shape) tuple, designating inputs whose shape vary
-    fixed_param_prefix : list of str, indicating fixed parameters
-    """
-    def __init__(self, symbol, data_names, label_names,
-                 logger=logging, context=ctx.cpu(), work_load_list=None,
-                 max_data_shapes=None, max_label_shapes=None, fixed_param_prefix=None):
-        super(MutableModule, self).__init__(logger=logger)
-        self._symbol = symbol
-        self._data_names = data_names
-        self._label_names = label_names
-        self._context = context
-        self._work_load_list = work_load_list
-
-        self._curr_module = None
-        self._max_data_shapes = max_data_shapes
-        self._max_label_shapes = max_label_shapes
-        self._fixed_param_prefix = fixed_param_prefix
-
-        fixed_param_names = list()
-        if fixed_param_prefix is not None:
-            for name in self._symbol.list_arguments():
-                for prefix in self._fixed_param_prefix:
-                    if prefix in name:
-                        fixed_param_names.append(name)
-        self._fixed_param_names = fixed_param_names
-
-    def _reset_bind(self):
-        self.binded = False
-        self._curr_module = None
-
-    @property
-    def data_names(self):
-        return self._data_names
-
-    @property
-    def output_names(self):
-        return self._symbol.list_outputs()
-
-    @property
-    def data_shapes(self):
-        assert self.binded
-        return self._curr_module.data_shapes
-
-    @property
-    def label_shapes(self):
-        assert self.binded
-        return self._curr_module.label_shapes
-
-    @property
-    def output_shapes(self):
-        assert self.binded
-        return self._curr_module.output_shapes
-
-    def get_params(self):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_params()
-
-    def init_params(self, initializer=Uniform(0.01), arg_params=None, aux_params=None,
-                    allow_missing=False, force_init=False, allow_extra=False):
-        if self.params_initialized and not force_init:
-            return
-        assert self.binded, 'call bind before initializing the parameters'
-        self._curr_module.init_params(initializer=initializer, arg_params=arg_params,
-                                      aux_params=aux_params, allow_missing=allow_missing,
-                                      force_init=force_init, allow_extra=allow_extra)
-        self.params_initialized = True
-
-    def bind(self, data_shapes, label_shapes=None, for_training=True,
-             inputs_need_grad=False, force_rebind=False, shared_module=None, grad_req='write'):
-        # in case we already initialized params, keep it
-        if self.params_initialized:
-            arg_params, aux_params = self.get_params()
-
-        # force rebinding is typically used when one want to switch from
-        # training to prediction phase.
-        if force_rebind:
-            self._reset_bind()
-
-        if self.binded:
-            self.logger.warning('Already binded, ignoring bind()')
-            return
-
-        assert shared_module is None, 'shared_module for MutableModule is not supported'
-
-        self.for_training = for_training
-        self.inputs_need_grad = inputs_need_grad
-        self.binded = True
-
-        max_shapes_dict = dict()
-        if self._max_data_shapes is not None:
-            max_shapes_dict.update(dict(self._max_data_shapes))
-        if self._max_label_shapes is not None:
-            max_shapes_dict.update(dict(self._max_label_shapes))
-
-        max_data_shapes = list()
-        for name, shape in data_shapes:
-            if name in max_shapes_dict:
-                max_data_shapes.append((name, max_shapes_dict[name]))
-            else:
-                max_data_shapes.append((name, shape))
-
-        max_label_shapes = list()
-        if label_shapes is not None:
-            for name, shape in label_shapes:
-                if name in max_shapes_dict:
-                    max_label_shapes.append((name, max_shapes_dict[name]))
-                else:
-                    max_label_shapes.append((name, shape))
-
-        if len(max_label_shapes) == 0:
-            max_label_shapes = None
-
-        module = Module(self._symbol, self._data_names, self._label_names, logger=self.logger,
-                        context=self._context, work_load_list=self._work_load_list,
-                        fixed_param_names=self._fixed_param_names)
-        module.bind(max_data_shapes, max_label_shapes, for_training, inputs_need_grad,
-                    force_rebind=False, shared_module=None)
-        self._curr_module = module
-
-        # copy back saved params, if already initialized
-        if self.params_initialized:
-            self.set_params(arg_params, aux_params)
-
-    def init_optimizer(self, kvstore='local', optimizer='sgd',
-                       optimizer_params=(('learning_rate', 0.01),), force_init=False):
-        assert self.binded and self.params_initialized
-        if self.optimizer_initialized and not force_init:
-            self.logger.warning('optimizer already initialized, ignoring.')
-            return
-
-        self._curr_module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                         force_init=force_init)
-        self.optimizer_initialized = True
-
-    def forward(self, data_batch, is_train=None):
-        assert self.binded and self.params_initialized
-
-        # get current_shapes
-        if self._curr_module.label_shapes is not None:
-            current_shapes = dict(self._curr_module.data_shapes + self._curr_module.label_shapes)
-        else:
-            current_shapes = dict(self._curr_module.data_shapes)
-
-        # get input_shapes
-        if data_batch.provide_label is not None:
-            input_shapes = dict(data_batch.provide_data + data_batch.provide_label)
-        else:
-            input_shapes = dict(data_batch.provide_data)
-
-        # decide if shape changed
-        shape_changed = False
-        for k, v in current_shapes.items():
-            if v != input_shapes[k]:
-                shape_changed = True
-
-        if shape_changed:
-            module = Module(self._symbol, self._data_names, self._label_names,
-                            logger=self.logger, context=self._context,
-                            work_load_list=self._work_load_list,
-                            fixed_param_names=self._fixed_param_names)
-            module.bind(data_batch.provide_data, data_batch.provide_label, self._curr_module.for_training,
-                        self._curr_module.inputs_need_grad, force_rebind=False,
-                        shared_module=self._curr_module)
-            self._curr_module = module
-
-        self._curr_module.forward(data_batch, is_train=is_train)
-
-    def backward(self, out_grads=None):
-        assert self.binded and self.params_initialized
-        self._curr_module.backward(out_grads=out_grads)
-
-    def update(self):
-        assert self.binded and self.params_initialized and self.optimizer_initialized
-        self._curr_module.update()
-
-    def get_outputs(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_outputs(merge_multi_context=merge_multi_context)
-
-    def get_input_grads(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized and self.inputs_need_grad
-        return self._curr_module.get_input_grads(merge_multi_context=merge_multi_context)
-
-    def update_metric(self, eval_metric, labels):
-        assert self.binded and self.params_initialized
-        self._curr_module.update_metric(eval_metric, labels)
-
-    def install_monitor(self, mon):
-        """ Install monitor on all executors """
-        assert self.binded
-        self._curr_module.install_monitor(mon)
diff --git a/RetinaFace/rcnn/core/tester.py b/RetinaFace/rcnn/core/tester.py
deleted file mode 100644
index 8d981fa1e..000000000
--- a/RetinaFace/rcnn/core/tester.py
+++ /dev/null
@@ -1,485 +0,0 @@
-from __future__ import print_function
-try:
-    import cPickle as pickle
-except ImportError:
-    import pickle
-import os
-import sys
-import time
-import mxnet as mx
-import numpy as np
-from builtins import range
-
-from mxnet.module import Module
-from .module import MutableModule
-from rcnn.logger import logger
-from rcnn.config import config
-from rcnn.io import image
-from rcnn.processing.bbox_transform import bbox_pred, clip_boxes
-from rcnn.processing.nms import py_nms_wrapper, cpu_nms_wrapper, gpu_nms_wrapper
-from rcnn.processing.bbox_transform import bbox_overlaps
-
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-class Predictor(object):
-    def __init__(self, symbol, data_names, label_names,
-                 context=mx.cpu(), max_data_shapes=None,
-                 provide_data=None, provide_label=None,
-                 arg_params=None, aux_params=None):
-        #self._mod = MutableModule(symbol, data_names, label_names,
-        #                          context=context, max_data_shapes=max_data_shapes)
-        self._mod = Module(symbol, data_names, label_names, context=context)
-        self._mod.bind(provide_data, provide_label, for_training=False)
-        self._mod.init_params(arg_params=arg_params, aux_params=aux_params)
-
-    def predict(self, data_batch):
-        self._mod.forward(data_batch)
-        return dict(zip(self._mod.output_names, self._mod.get_outputs())) #TODO
-        #return self._mod.get_outputs()
-
-
-def im_proposal(predictor, data_batch, data_names, scale):
-    data_dict = dict(zip(data_names, data_batch.data))
-    output = predictor.predict(data_batch)
-
-    # drop the batch index
-    boxes = output['rois_output'].asnumpy()[:, 1:]
-    scores = output['rois_score'].asnumpy()
-
-    # transform to original scale
-    boxes = boxes / scale
-
-    return scores, boxes, data_dict
-
-def _im_proposal(predictor, data_batch, data_names, scale):
-    data_dict = dict(zip(data_names, data_batch.data))
-    output = predictor.predict(data_batch)
-    print('output', output)
-
-    # drop the batch index
-    boxes = output['rois_output'].asnumpy()[:, 1:]
-    scores = output['rois_score'].asnumpy()
-
-    # transform to original scale
-    boxes = boxes / scale
-
-    return scores, boxes, data_dict
-
-
-def generate_proposals(predictor, test_data, imdb, vis=False, thresh=0.):
-    """
-    Generate detections results using RPN.
-    :param predictor: Predictor
-    :param test_data: data iterator, must be non-shuffled
-    :param imdb: image database
-    :param vis: controls visualization
-    :param thresh: thresh for valid detections
-    :return: list of detected boxes
-    """
-    assert vis or not test_data.shuffle
-    data_names = [k[0] for k in test_data.provide_data]
-
-    i = 0
-    t = time.time()
-    imdb_boxes = list()
-    original_boxes = list()
-    for im_info, data_batch in test_data:
-        t1 = time.time() - t
-        t = time.time()
-
-        scale = im_info[0, 2]
-        scores, boxes, data_dict = im_proposal(predictor, data_batch, data_names, scale)
-        print(scores.shape, boxes.shape, file=sys.stderr)
-        t2 = time.time() - t
-        t = time.time()
-
-        # assemble proposals
-        dets = np.hstack((boxes, scores))
-        original_boxes.append(dets)
-
-        # filter proposals
-        keep = np.where(dets[:, 4:] > thresh)[0]
-        dets = dets[keep, :]
-        imdb_boxes.append(dets)
-
-        if vis:
-            vis_all_detection(data_dict['data'].asnumpy(), [dets], ['obj'], scale)
-
-        logger.info('generating %d/%d ' % (i + 1, imdb.num_images) +
-                    'proposal %d ' % (dets.shape[0]) +
-                    'data %.4fs net %.4fs' % (t1, t2))
-        i += 1
-
-    assert len(imdb_boxes) == imdb.num_images, 'calculations not complete'
-
-    # save results
-    rpn_folder = os.path.join(imdb.root_path, 'rpn_data')
-    if not os.path.exists(rpn_folder):
-        os.mkdir(rpn_folder)
-
-    rpn_file = os.path.join(rpn_folder, imdb.name + '_rpn.pkl')
-    with open(rpn_file, 'wb') as f:
-        pickle.dump(imdb_boxes, f, pickle.HIGHEST_PROTOCOL)
-
-    if thresh > 0:
-        full_rpn_file = os.path.join(rpn_folder, imdb.name + '_full_rpn.pkl')
-        with open(full_rpn_file, 'wb') as f:
-            pickle.dump(original_boxes, f, pickle.HIGHEST_PROTOCOL)
-
-    logger.info('wrote rpn proposals to %s' % rpn_file)
-    return imdb_boxes
-
-def test_proposals(predictor, test_data, imdb, roidb, vis=False):
-    """
-    Test detections results using RPN.
-    :param predictor: Predictor
-    :param test_data: data iterator, must be non-shuffled
-    :param imdb: image database
-    :param roidb: roidb 
-    :param vis: controls visualization
-    :return: recall, mAP
-    """
-    assert vis or not test_data.shuffle
-    data_names = [k[0] for k in test_data.provide_data]
-
-    #bbox_file = os.path.join(rpn_folder, imdb.name + '_bbox.txt')
-    #bbox_f = open(bbox_file, 'w')
-
-    i = 0
-    t = time.time()
-    output_folder = os.path.join(imdb.root_path, 'output')
-    if not os.path.exists(output_folder):
-        os.mkdir(output_folder)
-    imdb_boxes = list()
-    original_boxes = list()
-    gt_overlaps = np.zeros(0)
-    overall = [0.0, 0.0]
-    gt_max = np.array( (0.0, 0.0) )
-    num_pos = 0
-    #apply scale, for SSH
-    #_, roidb = image.get_image(roidb)
-    for im_info, data_batch in test_data:
-        t1 = time.time() - t
-        t = time.time()
-
-        oscale = im_info[0, 2]
-        #print('scale', scale, file=sys.stderr)
-        scale = 1.0 #fix scale=1.0 for SSH face detector
-        scores, boxes, data_dict = im_proposal(predictor, data_batch, data_names, scale)
-        #print(scores.shape, boxes.shape, file=sys.stderr)
-        t2 = time.time() - t
-        t = time.time()
-
-        # assemble proposals
-        dets = np.hstack((boxes, scores))
-        original_boxes.append(dets)
-
-        # filter proposals
-        keep = np.where(dets[:, 4:] > config.TEST.SCORE_THRESH)[0]
-        dets = dets[keep, :]
-        imdb_boxes.append(dets)
-
-
-        logger.info('generating %d/%d ' % (i + 1, imdb.num_images) +
-                    'proposal %d ' % (dets.shape[0]) +
-                    'data %.4fs net %.4fs' % (t1, t2))
-
-        #if dets.shape[0]==0:
-        #  continue
-        if vis:
-            vis_all_detection(data_dict['data'].asnumpy(), [dets], ['obj'], scale)
-        boxes = dets
-        #max_gt_overlaps = roidb[i]['gt_overlaps'].max(axis=1)
-        #gt_inds = np.where((roidb[i]['gt_classes'] > 0) & (max_gt_overlaps == 1))[0]
-        #gt_boxes = roidb[i]['boxes'][gt_inds, :]
-        gt_boxes = roidb[i]['boxes'].copy() * oscale # as roidb is the original one, need to scale GT for SSH
-        gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] + 1) * (gt_boxes[:, 3] - gt_boxes[:, 1] + 1)
-        num_pos += gt_boxes.shape[0]
-
-        overlaps = bbox_overlaps(boxes.astype(np.float), gt_boxes.astype(np.float))
-        #print(im_info, gt_boxes.shape, boxes.shape, overlaps.shape, file=sys.stderr)
-
-        _gt_overlaps = np.zeros((gt_boxes.shape[0]))
-        # choose whatever is smaller to iterate
-
-        #for j in range(gt_boxes.shape[0]):
-        #  print('gt %d,%d,%d,%d'% (gt_boxes[j][0], gt_boxes[j][1], gt_boxes[j][2]-gt_boxes[j][0], gt_boxes[j][3]-gt_boxes[j][1]), file=sys.stderr)
-        #  gt_max = np.maximum( gt_max, np.array( (gt_boxes[j][2], gt_boxes[j][3]) ) )
-        #print('gt max', gt_max, file=sys.stderr)
-        #for j in range(boxes.shape[0]):
-        #  print('anchor_box %.2f,%.2f,%.2f,%.2f'% (boxes[j][0], boxes[j][1], boxes[j][2]-boxes[j][0], boxes[j][3]-boxes[j][1]), file=sys.stderr)
-
-        #rounds = min(boxes.shape[0], gt_boxes.shape[0])
-        #for j in range(rounds):
-        #    # find which proposal maximally covers each gt box
-        #    argmax_overlaps = overlaps.argmax(axis=0)
-        #    print(j, 'argmax_overlaps', argmax_overlaps, file=sys.stderr)
-        #    # get the IoU amount of coverage for each gt box
-        #    max_overlaps = overlaps.max(axis=0)
-        #    print(j, 'max_overlaps', max_overlaps, file=sys.stderr)
-        #    # find which gt box is covered by most IoU
-        #    gt_ind = max_overlaps.argmax()
-        #    gt_ovr = max_overlaps.max()
-        #    assert (gt_ovr >= 0), '%s\n%s\n%s' % (boxes, gt_boxes, overlaps)
-        #    # find the proposal box that covers the best covered gt box
-        #    box_ind = argmax_overlaps[gt_ind]
-        #    print('max box', gt_ind, box_ind, (boxes[box_ind][0], boxes[box_ind][1], boxes[box_ind][2]-boxes[box_ind][0], boxes[box_ind][3]-boxes[box_ind][1], boxes[box_ind][4]), file=sys.stderr)
-        #    # record the IoU coverage of this gt box
-        #    _gt_overlaps[j] = overlaps[box_ind, gt_ind]
-        #    assert (_gt_overlaps[j] == gt_ovr)
-        #    # mark the proposal box and the gt box as used
-        #    overlaps[box_ind, :] = -1
-        #    overlaps[:, gt_ind] = -1
-
-        if boxes.shape[0]>0:
-          _gt_overlaps = overlaps.max(axis=0)
-          #print('max_overlaps', _gt_overlaps, file=sys.stderr)
-          for j in range(len(_gt_overlaps)):
-            if _gt_overlaps[j]>config.TEST.IOU_THRESH:
-              continue
-            print(j, 'failed', gt_boxes[j],  'max_overlap:', _gt_overlaps[j], file=sys.stderr)
-            #_idx = np.where(overlaps[:,j]>0.4)[0]
-            #print(j, _idx, file=sys.stderr)
-            #print(overlaps[_idx,j], file=sys.stderr)
-            #for __idx in _idx:
-            #  print(gt_boxes[j], boxes[__idx], overlaps[__idx,j], IOU(gt_boxes[j], boxes[__idx,0:4]), file=sys.stderr)
-
-          # append recorded IoU coverage level
-          found = (_gt_overlaps > config.TEST.IOU_THRESH).sum()
-          _recall = found / float(gt_boxes.shape[0])
-          print('recall', _recall, gt_boxes.shape[0], boxes.shape[0], gt_areas, file=sys.stderr)
-          overall[0]+=found
-          overall[1]+=gt_boxes.shape[0]
-          #gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
-          #_recall = (gt_overlaps >= threshold).sum() / float(num_pos)
-          _recall = float(overall[0])/overall[1]
-          print('recall_all', _recall, file=sys.stderr)
-
-
-        boxes[:,0:4] /= oscale
-        _vec = roidb[i]['image'].split('/')
-        out_dir = os.path.join(output_folder, _vec[-2])
-        if not os.path.exists(out_dir):
-            os.mkdir(out_dir)
-        out_file = os.path.join(out_dir, _vec[-1].replace('jpg', 'txt'))
-        with open(out_file, 'w') as f:
-          name = '/'.join(roidb[i]['image'].split('/')[-2:])
-          f.write("%s\n"%(name))
-          f.write("%d\n"%(boxes.shape[0]))
-          for b in range(boxes.shape[0]):
-            box = boxes[b]
-            f.write("%d %d %d %d %g \n"%(box[0], box[1], box[2]-box[0], box[3]-box[1], box[4]))
-        i += 1
-
-    #bbox_f.close()
-    return
-    gt_overlaps = np.sort(gt_overlaps)
-    recalls = np.zeros_like(thresholds)
-
-    # compute recall for each IoU threshold
-    for i, t in enumerate(thresholds):
-        recalls[i] = (gt_overlaps >= t).sum() / float(num_pos)
-    ar = recalls.mean()
-
-    # print results
-    print('average recall for {}: {:.3f}'.format(area_name, ar))
-    for threshold, recall in zip(thresholds, recalls):
-        print('recall @{:.2f}: {:.3f}'.format(threshold, recall))
-
-
-
-
-    assert len(imdb_boxes) == imdb.num_images, 'calculations not complete'
-
-    # save results
-
-    rpn_file = os.path.join(rpn_folder, imdb.name + '_rpn.pkl')
-    with open(rpn_file, 'wb') as f:
-        pickle.dump(imdb_boxes, f, pickle.HIGHEST_PROTOCOL)
-
-    logger.info('wrote rpn proposals to %s' % rpn_file)
-    return imdb_boxes
-
-def im_detect(predictor, data_batch, data_names, scale):
-    output = predictor.predict(data_batch)
-
-    data_dict = dict(zip(data_names, data_batch.data))
-    if config.TEST.HAS_RPN:
-        rois = output['rois_output'].asnumpy()[:, 1:]
-    else:
-        rois = data_dict['rois'].asnumpy().reshape((-1, 5))[:, 1:]
-    im_shape = data_dict['data'].shape
-
-    # save output
-    scores = output['cls_prob_reshape_output'].asnumpy()[0]
-    bbox_deltas = output['bbox_pred_reshape_output'].asnumpy()[0]
-
-    # post processing
-    pred_boxes = bbox_pred(rois, bbox_deltas)
-    pred_boxes = clip_boxes(pred_boxes, im_shape[-2:])
-
-    # we used scaled image & roi to train, so it is necessary to transform them back
-    pred_boxes = pred_boxes / scale
-
-    return scores, pred_boxes, data_dict
-
-
-def pred_eval(predictor, test_data, imdb, vis=False, thresh=1e-3):
-    """
-    wrapper for calculating offline validation for faster data analysis
-    in this example, all threshold are set by hand
-    :param predictor: Predictor
-    :param test_data: data iterator, must be non-shuffle
-    :param imdb: image database
-    :param vis: controls visualization
-    :param thresh: valid detection threshold
-    :return:
-    """
-    assert vis or not test_data.shuffle
-    data_names = [k[0] for k in test_data.provide_data]
-
-    nms = py_nms_wrapper(config.TEST.NMS)
-
-    # limit detections to max_per_image over all classes
-    max_per_image = -1
-
-    num_images = imdb.num_images
-    # all detections are collected into:
-    #    all_boxes[cls][image] = N x 5 array of detections in
-    #    (x1, y1, x2, y2, score)
-    all_boxes = [[[] for _ in range(num_images)]
-                 for _ in range(imdb.num_classes)]
-
-    i = 0
-    t = time.time()
-    for im_info, data_batch in test_data:
-        t1 = time.time() - t
-        t = time.time()
-
-        scale = im_info[0, 2]
-        scores, boxes, data_dict = im_detect(predictor, data_batch, data_names, scale)
-
-        t2 = time.time() - t
-        t = time.time()
-
-        for j in range(1, imdb.num_classes):
-            indexes = np.where(scores[:, j] > thresh)[0]
-            cls_scores = scores[indexes, j, np.newaxis]
-            cls_boxes = boxes[indexes, j * 4:(j + 1) * 4]
-            cls_dets = np.hstack((cls_boxes, cls_scores))
-            keep = nms(cls_dets)
-            all_boxes[j][i] = cls_dets[keep, :]
-
-        if max_per_image > 0:
-            image_scores = np.hstack([all_boxes[j][i][:, -1]
-                                      for j in range(1, imdb.num_classes)])
-            if len(image_scores) > max_per_image:
-                image_thresh = np.sort(image_scores)[-max_per_image]
-                for j in range(1, imdb.num_classes):
-                    keep = np.where(all_boxes[j][i][:, -1] >= image_thresh)[0]
-                    all_boxes[j][i] = all_boxes[j][i][keep, :]
-
-        if vis:
-            boxes_this_image = [[]] + [all_boxes[j][i] for j in range(1, imdb.num_classes)]
-            vis_all_detection(data_dict['data'].asnumpy(), boxes_this_image, imdb.classes, scale)
-
-        t3 = time.time() - t
-        t = time.time()
-        logger.info('testing %d/%d data %.4fs net %.4fs post %.4fs' % (i, imdb.num_images, t1, t2, t3))
-        i += 1
-
-    det_file = os.path.join(imdb.cache_path, imdb.name + '_detections.pkl')
-    with open(det_file, 'wb') as f:
-        pickle.dump(all_boxes, f, protocol=pickle.HIGHEST_PROTOCOL)
-
-    imdb.evaluate_detections(all_boxes)
-
-
-def vis_all_detection(im_array, detections, class_names, scale):
-    """
-    visualize all detections in one image
-    :param im_array: [b=1 c h w] in rgb
-    :param detections: [ numpy.ndarray([[x1 y1 x2 y2 score]]) for j in classes ]
-    :param class_names: list of names in imdb
-    :param scale: visualize the scaled image
-    :return:
-    """
-    import matplotlib.pyplot as plt
-    import random
-    im = image.transform_inverse(im_array, config.PIXEL_MEANS)
-    plt.imshow(im)
-    for j, name in enumerate(class_names):
-        if name == '__background__':
-            continue
-        color = (random.random(), random.random(), random.random())  # generate a random color
-        dets = detections[j]
-        for det in dets:
-            bbox = det[:4] * scale
-            score = det[-1]
-            rect = plt.Rectangle((bbox[0], bbox[1]),
-                                 bbox[2] - bbox[0],
-                                 bbox[3] - bbox[1], fill=False,
-                                 edgecolor=color, linewidth=3.5)
-            plt.gca().add_patch(rect)
-            plt.gca().text(bbox[0], bbox[1] - 2,
-                           '{:s} {:.3f}'.format(name, score),
-                           bbox=dict(facecolor=color, alpha=0.5), fontsize=12, color='white')
-    plt.show()
-
-
-def draw_all_detection(im_array, detections, class_names, scale):
-    """
-    visualize all detections in one image
-    :param im_array: [b=1 c h w] in rgb
-    :param detections: [ numpy.ndarray([[x1 y1 x2 y2 score]]) for j in classes ]
-    :param class_names: list of names in imdb
-    :param scale: visualize the scaled image
-    :return:
-    """
-    import cv2
-    import random
-    color_white = (255, 255, 255)
-    im = image.transform_inverse(im_array, config.PIXEL_MEANS)
-    # change to bgr
-    im = cv2.cvtColor(im, cv2.cv.CV_RGB2BGR)
-    for j, name in enumerate(class_names):
-        if name == '__background__':
-            continue
-        color = (random.randint(0, 256), random.randint(0, 256), random.randint(0, 256))  # generate a random color
-        dets = detections[j]
-        for det in dets:
-            bbox = det[:4] * scale
-            score = det[-1]
-            bbox = map(int, bbox)
-            cv2.rectangle(im, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color=color, thickness=2)
-            cv2.putText(im, '%s %.3f' % (class_names[j], score), (bbox[0], bbox[1] + 10),
-                        color=color_white, fontFace=cv2.FONT_HERSHEY_COMPLEX, fontScale=0.5)
-    return im
diff --git a/RetinaFace/rcnn/cython/setup.py b/RetinaFace/rcnn/cython/setup.py
deleted file mode 100644
index 3e27add87..000000000
--- a/RetinaFace/rcnn/cython/setup.py
+++ /dev/null
@@ -1,169 +0,0 @@
-# --------------------------------------------------------
-# Fast R-CNN
-# Copyright (c) 2015 Microsoft
-# Licensed under The MIT License [see LICENSE for details]
-# Written by Ross Girshick
-# --------------------------------------------------------
-
-import os
-from os.path import join as pjoin
-from setuptools import setup
-from distutils.extension import Extension
-from Cython.Distutils import build_ext
-import numpy as np
-
-
-def find_in_path(name, path):
-    "Find a file in a search path"
-    # Adapted fom
-    # http://code.activestate.com/recipes/52224-find-a-file-given-a-search-path/
-    for dir in path.split(os.pathsep):
-        binpath = pjoin(dir, name)
-        if os.path.exists(binpath):
-            return os.path.abspath(binpath)
-    return None
-
-
-def locate_cuda():
-    """Locate the CUDA environment on the system
-
-    Returns a dict with keys 'home', 'nvcc', 'include', and 'lib64'
-    and values giving the absolute path to each directory.
-
-    Starts by looking for the CUDAHOME env variable. If not found, everything
-    is based on finding 'nvcc' in the PATH.
-    """
-
-    # first check if the CUDAHOME env variable is in use
-    if 'CUDAHOME' in os.environ:
-        home = os.environ['CUDAHOME']
-        nvcc = pjoin(home, 'bin', 'nvcc')
-    else:
-        # otherwise, search the PATH for NVCC
-        default_path = pjoin(os.sep, 'usr', 'local', 'cuda', 'bin')
-        nvcc = find_in_path('nvcc', os.environ['PATH'] + os.pathsep + default_path)
-        if nvcc is None:
-            raise EnvironmentError('The nvcc binary could not be '
-                'located in your $PATH. Either add it to your path, or set $CUDAHOME')
-        home = os.path.dirname(os.path.dirname(nvcc))
-
-    cudaconfig = {'home':home, 'nvcc':nvcc,
-                  'include': pjoin(home, 'include'),
-                  'lib64': pjoin(home, 'lib64')}
-    for k, v in cudaconfig.items():
-        if not os.path.exists(v):
-            raise EnvironmentError('The CUDA %s path could not be located in %s' % (k, v))
-
-    return cudaconfig
-
-
-# Test if cuda could be foun
-try:
-    CUDA = locate_cuda()
-except EnvironmentError:
-    CUDA = None
-
-
-# Obtain the numpy include directory.  This logic works across numpy versions.
-try:
-    numpy_include = np.get_include()
-except AttributeError:
-    numpy_include = np.get_numpy_include()
-
-
-def customize_compiler_for_nvcc(self):
-    """inject deep into distutils to customize how the dispatch
-    to gcc/nvcc works.
-
-    If you subclass UnixCCompiler, it's not trivial to get your subclass
-    injected in, and still have the right customizations (i.e.
-    distutils.sysconfig.customize_compiler) run on it. So instead of going
-    the OO route, I have this. Note, it's kindof like a wierd functional
-    subclassing going on."""
-
-    # tell the compiler it can processes .cu
-    self.src_extensions.append('.cu')
-
-    # save references to the default compiler_so and _comple methods
-    default_compiler_so = self.compiler_so
-    super = self._compile
-
-    # now redefine the _compile method. This gets executed for each
-    # object but distutils doesn't have the ability to change compilers
-    # based on source extension: we add it.
-    def _compile(obj, src, ext, cc_args, extra_postargs, pp_opts):
-        if os.path.splitext(src)[1] == '.cu':
-            # use the cuda for .cu files
-            self.set_executable('compiler_so', CUDA['nvcc'])
-            # use only a subset of the extra_postargs, which are 1-1 translated
-            # from the extra_compile_args in the Extension class
-            postargs = extra_postargs['nvcc']
-        else:
-            postargs = extra_postargs['gcc']
-
-        super(obj, src, ext, cc_args, postargs, pp_opts)
-        # reset the default compiler_so, which we might have changed for cuda
-        self.compiler_so = default_compiler_so
-
-    # inject our redefined _compile method into the class
-    self._compile = _compile
-
-
-# run the customize_compiler
-class custom_build_ext(build_ext):
-    def build_extensions(self):
-        customize_compiler_for_nvcc(self.compiler)
-        build_ext.build_extensions(self)
-
-
-ext_modules = [
-    Extension(
-        "bbox",
-        ["bbox.pyx"],
-        extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
-        include_dirs=[numpy_include]
-    ),
-    Extension(
-        "anchors",
-        ["anchors.pyx"],
-        extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
-        include_dirs=[numpy_include]
-    ),
-    Extension(
-        "cpu_nms",
-        ["cpu_nms.pyx"],
-        extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
-        include_dirs = [numpy_include]
-    ),
-]
-
-if CUDA is not None:
-    ext_modules.append(
-        Extension('gpu_nms',
-            ['nms_kernel.cu', 'gpu_nms.pyx'],
-            library_dirs=[CUDA['lib64']],
-            libraries=['cudart'],
-            language='c++',
-            runtime_library_dirs=[CUDA['lib64']],
-            # this syntax is specific to this build system
-            # we're only going to use certain compiler args with nvcc and not with
-            # gcc the implementation of this trick is in customize_compiler() below
-            extra_compile_args={'gcc': ["-Wno-unused-function"],
-                                'nvcc': ['-arch=sm_35',
-                                         '--ptxas-options=-v',
-                                         '-c',
-                                         '--compiler-options',
-                                         "'-fPIC'"]},
-            include_dirs = [numpy_include, CUDA['include']]
-        )
-    )
-else:
-    print('Skipping GPU_NMS')
-
-
-setup(
-    name='frcnn_cython',
-    ext_modules=ext_modules,
-    # inject our custom trigger
-    cmdclass={'build_ext': custom_build_ext},
-)
diff --git a/RetinaFace/rcnn/dataset/imdb.py b/RetinaFace/rcnn/dataset/imdb.py
deleted file mode 100644
index d03420696..000000000
--- a/RetinaFace/rcnn/dataset/imdb.py
+++ /dev/null
@@ -1,318 +0,0 @@
-"""
-General image database
-An image database creates a list of relative image path called image_set_index and
-transform index to absolute image path. As to training, it is necessary that ground
-truth and proposals are mixed together for training.
-roidb
-basic format [image_index]
-['image', 'height', 'width', 'flipped',
-'boxes', 'gt_classes', 'gt_overlaps', 'max_classes', 'max_overlaps', 'bbox_targets']
-"""
-
-from ..logger import logger
-import os
-try:
-    import cPickle as pickle
-except ImportError:
-    import pickle
-import numpy as np
-from ..processing.bbox_transform import bbox_overlaps
-
-
-class IMDB(object):
-    def __init__(self, name, image_set, root_path, dataset_path):
-        """
-        basic information about an image database
-        :param name: name of image database will be used for any output
-        :param root_path: root path store cache and proposal data
-        :param dataset_path: dataset path store images and image lists
-        """
-        self.name = name + '_' + image_set
-        self.image_set = image_set
-        self.root_path = root_path
-        self.data_path = dataset_path
-
-        # abstract attributes
-        self.classes = []
-        self.num_classes = 0
-        self.image_set_index = []
-        self.num_images = 0
-
-        self.config = {}
-
-    def image_path_from_index(self, index):
-        raise NotImplementedError
-
-    def gt_roidb(self):
-        raise NotImplementedError
-
-    def evaluate_detections(self, detections):
-        raise NotImplementedError
-
-    @property
-    def cache_path(self):
-        """
-        make a directory to store all caches
-        :return: cache path
-        """
-        cache_path = os.path.join(self.root_path, 'cache')
-        if not os.path.exists(cache_path):
-            os.mkdir(cache_path)
-        return cache_path
-
-    def image_path_at(self, index):
-        """
-        access image at index in image database
-        :param index: image index in image database
-        :return: image path
-        """
-        return self.image_path_from_index(self.image_set_index[index])
-
-    def load_rpn_data(self, full=False):
-        if full:
-            rpn_file = os.path.join(self.root_path, 'rpn_data', self.name + '_full_rpn.pkl')
-        else:
-            rpn_file = os.path.join(self.root_path, 'rpn_data', self.name + '_rpn.pkl')
-        assert os.path.exists(rpn_file), '%s rpn data not found at %s' % (self.name, rpn_file)
-        logger.info('%s loading rpn data from %s' % (self.name, rpn_file))
-        with open(rpn_file, 'rb') as f:
-            box_list = pickle.load(f)
-        return box_list
-
-    def load_rpn_roidb(self, gt_roidb):
-        """
-        turn rpn detection boxes into roidb
-        :param gt_roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        """
-        box_list = self.load_rpn_data()
-        return self.create_roidb_from_box_list(box_list, gt_roidb)
-
-    def rpn_roidb(self, gt_roidb, append_gt=False):
-        """
-        get rpn roidb and ground truth roidb
-        :param gt_roidb: ground truth roidb
-        :param append_gt: append ground truth
-        :return: roidb of rpn
-        """
-        if append_gt:
-            logger.info('%s appending ground truth annotations' % self.name)
-            rpn_roidb = self.load_rpn_roidb(gt_roidb)
-            roidb = IMDB.merge_roidbs(gt_roidb, rpn_roidb)
-        else:
-            roidb = self.load_rpn_roidb(gt_roidb)
-        return roidb
-
-    def create_roidb_from_box_list(self, box_list, gt_roidb):
-        """
-        given ground truth, prepare roidb
-        :param box_list: [image_index] ndarray of [box_index][x1, x2, y1, y2]
-        :param gt_roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        """
-        assert len(box_list) == self.num_images, 'number of boxes matrix must match number of images'
-        roidb = []
-        for i in range(self.num_images):
-            roi_rec = dict()
-            roi_rec['image'] = gt_roidb[i]['image']
-            roi_rec['height'] = gt_roidb[i]['height']
-            roi_rec['width'] = gt_roidb[i]['width']
-
-            boxes = box_list[i]
-            if boxes.shape[1] == 5:
-                boxes = boxes[:, :4]
-            num_boxes = boxes.shape[0]
-            overlaps = np.zeros((num_boxes, self.num_classes), dtype=np.float32)
-            if gt_roidb is not None and gt_roidb[i]['boxes'].size > 0:
-                gt_boxes = gt_roidb[i]['boxes']
-                gt_classes = gt_roidb[i]['gt_classes']
-                # n boxes and k gt_boxes => n * k overlap
-                gt_overlaps = bbox_overlaps(boxes.astype(np.float), gt_boxes.astype(np.float))
-                # for each box in n boxes, select only maximum overlap (must be greater than zero)
-                argmaxes = gt_overlaps.argmax(axis=1)
-                maxes = gt_overlaps.max(axis=1)
-                I = np.where(maxes > 0)[0]
-                overlaps[I, gt_classes[argmaxes[I]]] = maxes[I]
-
-            roi_rec.update({'boxes': boxes,
-                            'gt_classes': np.zeros((num_boxes,), dtype=np.int32),
-                            'gt_overlaps': overlaps,
-                            'max_classes': overlaps.argmax(axis=1),
-                            'max_overlaps': overlaps.max(axis=1),
-                            'flipped': False})
-
-            # background roi => background class
-            zero_indexes = np.where(roi_rec['max_overlaps'] == 0)[0]
-            assert all(roi_rec['max_classes'][zero_indexes] == 0)
-            # foreground roi => foreground class
-            nonzero_indexes = np.where(roi_rec['max_overlaps'] > 0)[0]
-            assert all(roi_rec['max_classes'][nonzero_indexes] != 0)
-
-            roidb.append(roi_rec)
-
-        return roidb
-
-    def append_flipped_images(self, roidb):
-        """
-        append flipped images to an roidb
-        flip boxes coordinates, images will be actually flipped when loading into network
-        :param roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
-        """
-        logger.info('%s append flipped images to roidb' % self.name)
-        assert self.num_images == len(roidb)
-        for i in range(self.num_images):
-            roi_rec = roidb[i]
-            entry = {'image': roi_rec['image'],
-                     'stream': roi_rec['stream'],
-                     'height': roi_rec['height'],
-                     'width': roi_rec['width'],
-                     #'boxes': boxes,
-                     'gt_classes': roidb[i]['gt_classes'],
-                     'gt_overlaps': roidb[i]['gt_overlaps'],
-                     'max_classes': roidb[i]['max_classes'],
-                     'max_overlaps': roidb[i]['max_overlaps'],
-                     'flipped': True}
-            for k in roi_rec:
-              if not k.startswith('boxes'):
-                continue
-              boxes = roi_rec[k].copy()
-              oldx1 = boxes[:, 0].copy()
-              oldx2 = boxes[:, 2].copy()
-              boxes[:, 0] = roi_rec['width'] - oldx2 - 1
-              boxes[:, 2] = roi_rec['width'] - oldx1 - 1
-              assert (boxes[:, 2] >= boxes[:, 0]).all()
-              entry[k] = boxes
-            if 'landmarks' in roi_rec:
-              k = 'landmarks'
-              landmarks = roi_rec[k].copy()
-              landmarks[:,:,0] *= -1
-              landmarks[:,:,0] += (roi_rec['width']-1)
-              #for a in range(0,10,2):
-              #  oldx1 = landmarks[:, a].copy()
-              #  landmarks[:,a] = roi_rec['width'] - oldx1 - 1
-              order = [1,0,2,4,3]
-              flandmarks = landmarks.copy()
-              for idx, a in enumerate(order):
-                flandmarks[:, idx,:] = landmarks[:,a,:]
-
-              entry[k] = flandmarks
-            if 'blur' in roi_rec:
-              entry['blur'] = roi_rec['blur']
-            roidb.append(entry)
-
-        self.image_set_index *= 2
-        return roidb
-
-    def evaluate_recall(self, roidb, candidate_boxes=None, thresholds=None):
-        """
-        evaluate detection proposal recall metrics
-        record max overlap value for each gt box; return vector of overlap values
-        :param roidb: used to evaluate
-        :param candidate_boxes: if not given, use roidb's non-gt boxes
-        :param thresholds: array-like recall threshold
-        :return: None
-        ar: average recall, recalls: vector recalls at each IoU overlap threshold
-        thresholds: vector of IoU overlap threshold, gt_overlaps: vector of all ground-truth overlaps
-        """
-        area_names = ['all', '0-25', '25-50', '50-100',
-                      '100-200', '200-300', '300-inf']
-        area_ranges = [[0**2, 1e5**2], [0**2, 25**2], [25**2, 50**2], [50**2, 100**2],
-                       [100**2, 200**2], [200**2, 300**2], [300**2, 1e5**2]]
-        area_counts = []
-        for area_name, area_range in zip(area_names[1:], area_ranges[1:]):
-            area_count = 0
-            for i in range(self.num_images):
-                if candidate_boxes is None:
-                    # default is use the non-gt boxes from roidb
-                    non_gt_inds = np.where(roidb[i]['gt_classes'] == 0)[0]
-                    boxes = roidb[i]['boxes'][non_gt_inds, :]
-                else:
-                    boxes = candidate_boxes[i]
-                boxes_areas = (boxes[:, 2] - boxes[:, 0] + 1) * (boxes[:, 3] - boxes[:, 1] + 1)
-                valid_range_inds = np.where((boxes_areas >= area_range[0]) & (boxes_areas < area_range[1]))[0]
-                area_count += len(valid_range_inds)
-            area_counts.append(area_count)
-        total_counts = float(sum(area_counts))
-        for area_name, area_count in zip(area_names[1:], area_counts):
-            logger.info('percentage of %s is %f' % (area_name, area_count / total_counts))
-        logger.info('average number of proposal is %f' % (total_counts / self.num_images))
-        for area_name, area_range in zip(area_names, area_ranges):
-            gt_overlaps = np.zeros(0)
-            num_pos = 0
-            for i in range(self.num_images):
-                # check for max_overlaps == 1 avoids including crowd annotations
-                max_gt_overlaps = roidb[i]['gt_overlaps'].max(axis=1)
-                gt_inds = np.where((roidb[i]['gt_classes'] > 0) & (max_gt_overlaps == 1))[0]
-                gt_boxes = roidb[i]['boxes'][gt_inds, :]
-                gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] + 1) * (gt_boxes[:, 3] - gt_boxes[:, 1] + 1)
-                valid_gt_inds = np.where((gt_areas >= area_range[0]) & (gt_areas < area_range[1]))[0]
-                gt_boxes = gt_boxes[valid_gt_inds, :]
-                num_pos += len(valid_gt_inds)
-
-                if candidate_boxes is None:
-                    # default is use the non-gt boxes from roidb
-                    non_gt_inds = np.where(roidb[i]['gt_classes'] == 0)[0]
-                    boxes = roidb[i]['boxes'][non_gt_inds, :]
-                else:
-                    boxes = candidate_boxes[i]
-                if boxes.shape[0] == 0:
-                    continue
-
-                overlaps = bbox_overlaps(boxes.astype(np.float), gt_boxes.astype(np.float))
-
-                _gt_overlaps = np.zeros((gt_boxes.shape[0]))
-                # choose whatever is smaller to iterate
-                rounds = min(boxes.shape[0], gt_boxes.shape[0])
-                for j in range(rounds):
-                    # find which proposal maximally covers each gt box
-                    argmax_overlaps = overlaps.argmax(axis=0)
-                    # get the IoU amount of coverage for each gt box
-                    max_overlaps = overlaps.max(axis=0)
-                    # find which gt box is covered by most IoU
-                    gt_ind = max_overlaps.argmax()
-                    gt_ovr = max_overlaps.max()
-                    assert (gt_ovr >= 0), '%s\n%s\n%s' % (boxes, gt_boxes, overlaps)
-                    # find the proposal box that covers the best covered gt box
-                    box_ind = argmax_overlaps[gt_ind]
-                    # record the IoU coverage of this gt box
-                    _gt_overlaps[j] = overlaps[box_ind, gt_ind]
-                    assert (_gt_overlaps[j] == gt_ovr)
-                    # mark the proposal box and the gt box as used
-                    overlaps[box_ind, :] = -1
-                    overlaps[:, gt_ind] = -1
-                # append recorded IoU coverage level
-                gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
-
-            gt_overlaps = np.sort(gt_overlaps)
-            if thresholds is None:
-                step = 0.05
-                thresholds = np.arange(0.5, 0.95 + 1e-5, step)
-            recalls = np.zeros_like(thresholds)
-
-            # compute recall for each IoU threshold
-            for i, t in enumerate(thresholds):
-                recalls[i] = (gt_overlaps >= t).sum() / float(num_pos)
-            ar = recalls.mean()
-
-            # print results
-            print('average recall for {}: {:.3f}, number:{}'.format(area_name, ar, num_pos))
-            for threshold, recall in zip(thresholds, recalls):
-                print('recall @{:.2f}: {:.3f}'.format(threshold, recall))
-
-    @staticmethod
-    def merge_roidbs(a, b):
-        """
-        merge roidbs into one
-        :param a: roidb to be merged into
-        :param b: roidb to be merged
-        :return: merged imdb
-        """
-        assert len(a) == len(b)
-        for i in range(len(a)):
-            a[i]['boxes'] = np.vstack((a[i]['boxes'], b[i]['boxes']))
-            a[i]['gt_classes'] = np.hstack((a[i]['gt_classes'], b[i]['gt_classes']))
-            a[i]['gt_overlaps'] = np.vstack((a[i]['gt_overlaps'], b[i]['gt_overlaps']))
-            a[i]['max_classes'] = np.hstack((a[i]['max_classes'], b[i]['max_classes']))
-            a[i]['max_overlaps'] = np.hstack((a[i]['max_overlaps'], b[i]['max_overlaps']))
-        return a
diff --git a/RetinaFace/rcnn/dataset/retinaface.py b/RetinaFace/rcnn/dataset/retinaface.py
deleted file mode 100644
index 53fe1aba2..000000000
--- a/RetinaFace/rcnn/dataset/retinaface.py
+++ /dev/null
@@ -1,186 +0,0 @@
-from __future__ import print_function
-try:
-    import cPickle as pickle
-except ImportError:
-    import pickle
-import cv2
-import os
-import numpy as np
-import json
-#from PIL import Image
-
-from ..logger import logger
-from .imdb import IMDB
-from .ds_utils import unique_boxes, filter_small_boxes
-from ..config import config
-
-class retinaface(IMDB):
-    def __init__(self, image_set, root_path, data_path):
-        super(retinaface, self).__init__('retinaface', image_set, root_path, data_path)
-        #assert image_set=='train'
-
-        split = image_set
-        self._split = image_set
-        self._image_set = image_set
-
-
-        self.root_path = root_path
-        self.data_path = data_path
-
-
-        self._dataset_path = self.data_path
-        self._imgs_path = os.path.join(self._dataset_path, image_set, 'images')
-        self._fp_bbox_map = {}
-        label_file = os.path.join(self._dataset_path, image_set, 'label.txt')
-        name = None
-        for line in open(label_file, 'r'):
-          line = line.strip()
-          if line.startswith('#'):
-            name = line[1:].strip()
-            self._fp_bbox_map[name] = []
-            continue
-          assert name is not None
-          assert name in self._fp_bbox_map
-          self._fp_bbox_map[name].append(line)
-        print('origin image size', len(self._fp_bbox_map))
-
-        #self.num_images = len(self._image_paths)
-        #self._image_index = range(len(self._image_paths))
-        self.classes = ['bg', 'face']
-        self.num_classes = len(self.classes)
-
-
-    def gt_roidb(self):
-        cache_file = os.path.join(self.cache_path, '{}_{}_gt_roidb.pkl'.format(self.name, self._split))
-        if os.path.exists(cache_file):
-            with open(cache_file, 'rb') as fid:
-                roidb = pickle.load(fid)
-            print('{} gt roidb loaded from {}'.format(self.name, cache_file))
-            self.num_images = len(roidb)
-            return roidb
-
-        roidb = []
-        max_num_boxes = 0
-        nonattr_box_num = 0
-        landmark_num = 0
-
-        pp = 0
-        for fp in self._fp_bbox_map:
-            pp += 1
-            if pp%1000==0:
-              print('loading', pp)
-            if self._split=='test':
-              image_path = os.path.join(self._imgs_path, fp)
-              roi = {'image': image_path}
-              roidb.append(roi)
-              continue
-            boxes = np.zeros([len(self._fp_bbox_map[fp]), 4], np.float)
-            landmarks = np.zeros([len(self._fp_bbox_map[fp]), 5, 3], np.float)
-            blur = np.zeros((len(self._fp_bbox_map[fp]),), np.float)
-            boxes_mask = []
-
-            gt_classes = np.ones([len(self._fp_bbox_map[fp])], np.int32)
-            overlaps = np.zeros([len(self._fp_bbox_map[fp]), 2], np.float)
-
-            imsize = cv2.imread(os.path.join(self._imgs_path, fp)).shape[0:2][::-1]
-            ix = 0
-
-            for aline in self._fp_bbox_map[fp]:
-                #imsize = Image.open(os.path.join(self._imgs_path, fp)).size
-                values = [float(x) for x in aline.strip().split()]
-                bbox = [values[0], values[1], values[0]+values[2], values[1]+values[3]]
-
-                x1 = bbox[0]
-                y1 = bbox[1]
-                x2 = min(imsize[0], bbox[2])
-                y2 = min(imsize[1], bbox[3])
-                if x1>=x2 or y1>=y2:
-                  continue
-
-                if config.BBOX_MASK_THRESH>0:
-                  if (x2 - x1) < config.BBOX_MASK_THRESH or y2 - y1 < config.BBOX_MASK_THRESH:
-                    boxes_mask.append(np.array([x1, y1, x2, y2], np.float))
-                    continue
-                if (x2 - x1) < config.TRAIN.MIN_BOX_SIZE or y2 - y1 < config.TRAIN.MIN_BOX_SIZE:
-                    continue
-
-                boxes[ix, :] = np.array([x1, y1, x2, y2], np.float)
-                if self._split=='train':
-                  landmark = np.array( values[4:19], dtype=np.float32 ).reshape((5,3))
-                  for li in range(5):
-                    #print(landmark)
-                    if landmark[li][0]==-1. and landmark[li][1]==-1.: #missing landmark
-                      assert landmark[li][2]==-1
-                    else:
-                      assert landmark[li][2]>=0
-                      if li==0:
-                        landmark_num+=1
-                      if landmark[li][2]==0.0:#visible
-                        landmark[li][2] = 1.0
-                      else:
-                        landmark[li][2] = 0.0
-
-                  landmarks[ix] = landmark
-
-                  blur[ix] = values[19]
-                  #print(aline, blur[ix])
-                  if blur[ix]<0.0:
-                    blur[ix] = 0.3
-                    nonattr_box_num+=1
-
-                cls = int(1)
-                gt_classes[ix] = cls
-                overlaps[ix, cls] = 1.0
-                ix += 1
-            max_num_boxes = max(max_num_boxes, ix)
-            #overlaps = scipy.sparse.csr_matrix(overlaps)
-            if self._split=='train' and ix==0:
-              continue
-            boxes = boxes[:ix,:]
-            landmarks = landmarks[:ix,:,:]
-            blur = blur[:ix]
-            gt_classes = gt_classes[:ix]
-            overlaps = overlaps[:ix,:]
-            image_path = os.path.join(self._imgs_path, fp)
-            with open(image_path, 'rb') as fin:
-                stream = fin.read()
-            stream = np.fromstring(stream, dtype=np.uint8)
-
-            roi = {
-              'image': image_path,
-              'stream': stream,
-              'height': imsize[1],
-              'width': imsize[0],
-              'boxes': boxes,
-              'landmarks': landmarks,
-              'blur': blur,
-              'gt_classes': gt_classes,
-              'gt_overlaps': overlaps,
-              'max_classes': overlaps.argmax(axis=1),
-              'max_overlaps': overlaps.max(axis=1),
-              'flipped': False,
-            }
-            if len(boxes_mask)>0:
-              boxes_mask = np.array(boxes_mask)
-              roi['boxes_mask'] = boxes_mask
-            roidb.append(roi)
-        for roi in roidb:
-          roi['max_num_boxes'] = max_num_boxes
-        self.num_images = len(roidb)
-        print('roidb size', len(roidb))
-        print('non attr box num', nonattr_box_num)
-        print('landmark num', landmark_num)
-        with open(cache_file, 'wb') as fid:
-            pickle.dump(roidb, fid, pickle.HIGHEST_PROTOCOL)
-        print('wrote gt roidb to {}'.format(cache_file))
-
-        return roidb
-
-    def write_detections(self, all_boxes, output_dir='./output/'):
-      pass
-
-        
-    def evaluate_detections(self, all_boxes, output_dir='./output/',method_name='insightdetection'):
-      pass
-
-
diff --git a/RetinaFace/rcnn/io/image.py b/RetinaFace/rcnn/io/image.py
deleted file mode 100644
index 32158f90e..000000000
--- a/RetinaFace/rcnn/io/image.py
+++ /dev/null
@@ -1,808 +0,0 @@
-from __future__ import print_function
-import numpy as np
-import cv2
-import os
-import math
-import sys
-import random
-from ..config import config
-
-def brightness_aug(src, x):
-  alpha = 1.0 + random.uniform(-x, x)
-  src *= alpha
-  return src
-
-def contrast_aug(src, x):
-  alpha = 1.0 + random.uniform(-x, x)
-  coef = np.array([[[0.299, 0.587, 0.114]]])
-  gray = src * coef
-  gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
-  src *= alpha
-  src += gray
-  return src
-
-def saturation_aug(src, x):
-  alpha = 1.0 + random.uniform(-x, x)
-  coef = np.array([[[0.299, 0.587, 0.114]]])
-  gray = src * coef
-  gray = np.sum(gray, axis=2, keepdims=True)
-  gray *= (1.0 - alpha)
-  src *= alpha
-  src += gray
-  return src
-
-def color_aug(img, x):
-  if config.COLOR_MODE>1:
-    augs = [brightness_aug, contrast_aug, saturation_aug]
-    random.shuffle(augs)
-  else:
-    augs = [brightness_aug]
-  for aug in augs:
-    #print(img.shape)
-    img = aug(img, x)
-    #print(img.shape)
-  return img
-
-def get_image(roidb, scale=False):
-    """
-    preprocess image and return processed roidb
-    :param roidb: a list of roidb
-    :return: list of img as in mxnet format
-    roidb add new item['im_info']
-    0 --- x (width, second dim of im)
-    |
-    y (height, first dim of im)
-    """
-    num_images = len(roidb)
-    processed_ims = []
-    processed_roidb = []
-    for i in range(num_images):
-        roi_rec = roidb[i]
-        if 'stream' in roi_rec:
-          im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
-        else:
-          assert os.path.exists(roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
-          im = cv2.imread(roi_rec['image'])
-        if roidb[i]['flipped']:
-            im = im[:, ::-1, :]
-        new_rec = roi_rec.copy()
-        if scale:
-          scale_range = config.TRAIN.SCALE_RANGE
-          im_scale = np.random.uniform(scale_range[0], scale_range[1])
-          im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-        elif not config.ORIGIN_SCALE:
-          scale_ind = random.randrange(len(config.SCALES))
-          target_size = config.SCALES[scale_ind][0]
-          max_size = config.SCALES[scale_ind][1]
-          im, im_scale = resize(im, target_size, max_size, stride=config.IMAGE_STRIDE)
-        else:
-          im_scale = 1.0
-        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS)
-        if 'boxes_mask' in roi_rec:
-          im = im.astype(np.float32)
-          boxes_mask = roi_rec['boxes_mask'].copy() * im_scale
-          boxes_mask = boxes_mask.astype(np.int)
-          for j in range(boxes_mask.shape[0]):
-            m = boxes_mask[j]
-            im_tensor[:,:,m[1]:m[3],m[0]:m[2]] = 0.0
-            #print('find mask', m, file=sys.stderr)
-        processed_ims.append(im_tensor)
-        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
-        if config.TRAIN.IMAGE_ALIGN>0:
-          if im_tensor.shape[2]%config.TRAIN.IMAGE_ALIGN!=0 or im_tensor.shape[3]%config.TRAIN.IMAGE_ALIGN!=0:
-            new_height = math.ceil(float(im_tensor.shape[2])/config.TRAIN.IMAGE_ALIGN)*config.TRAIN.IMAGE_ALIGN
-            new_width = math.ceil(float(im_tensor.shape[3])/config.TRAIN.IMAGE_ALIGN)*config.TRAIN.IMAGE_ALIGN
-            new_im_tensor = np.zeros((1, 3, int(new_height), int(new_width)))
-            new_im_tensor[:,:,0:im_tensor.shape[2],0:im_tensor.shape[3]] = im_tensor
-            print(im_tensor.shape, new_im_tensor.shape, file=sys.stderr)
-            im_tensor = new_im_tensor
-        #print('boxes', new_rec['boxes'], file=sys.stderr)
-        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
-        new_rec['im_info'] = im_info
-        processed_roidb.append(new_rec)
-    return processed_ims, processed_roidb
-
-TMP_ID = -1
-#bakup method
-def __get_crop_image(roidb):
-    """
-    preprocess image and return processed roidb
-    :param roidb: a list of roidb
-    :return: list of img as in mxnet format
-    roidb add new item['im_info']
-    0 --- x (width, second dim of im)
-    |
-    y (height, first dim of im)
-    """
-    #roidb and each roi_rec can not be changed as it will be reused in next epoch
-    num_images = len(roidb)
-    processed_ims = []
-    processed_roidb = []
-    for i in range(num_images):
-        roi_rec = roidb[i]
-        if 'stream' in roi_rec:
-          im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
-        else:
-          assert os.path.exists(roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
-          im = cv2.imread(roi_rec['image'])
-        if roidb[i]['flipped']:
-            im = im[:, ::-1, :]
-        if 'boxes_mask' in roi_rec:
-          #im = im.astype(np.float32)
-          boxes_mask = roi_rec['boxes_mask'].copy()
-          boxes_mask = boxes_mask.astype(np.int)
-          for j in range(boxes_mask.shape[0]):
-            m = boxes_mask[j]
-            im[m[1]:m[3],m[0]:m[2],:] = 0
-            #print('find mask', m, file=sys.stderr)
-        new_rec = roi_rec.copy()
-
-
-        #choose one gt randomly
-        SIZE = config.SCALES[0][0]
-        TARGET_BOX_SCALES = np.array([16,32,64,128,256,512])
-        assert roi_rec['boxes'].shape[0]>0
-        candidates = []
-        for i in range(roi_rec['boxes'].shape[0]):
-          box = roi_rec['boxes'][i]
-          box_size = max(box[2]-box[0], box[3]-box[1])
-          if box_size<config.TRAIN.MIN_BOX_SIZE:
-            continue
-          #if box[0]<0 or box[1]<0:
-          #  continue
-          #if box[2]>im.shape[1] or box[3]>im.shape[0]:
-          #  continue;
-          candidates.append(i)
-        assert len(candidates)>0
-        box_ind = random.choice(candidates)
-        box = roi_rec['boxes'][box_ind]
-        box_size = max(box[2]-box[0], box[3]-box[1])
-        dist = np.abs(TARGET_BOX_SCALES - box_size)
-        nearest = np.argmin(dist)
-        target_ind = random.randrange(min(len(TARGET_BOX_SCALES), nearest+2))
-        target_box_size = TARGET_BOX_SCALES[target_ind]
-        im_scale = float(target_box_size) / box_size
-        #min_scale = float(SIZE)/np.min(im.shape[0:2])
-        #if im_scale<min_scale:
-        #  im_scale = min_scale
-        im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-        new_rec['boxes'] = roi_rec['boxes'].copy()*im_scale
-        box_scale = new_rec['boxes'][box_ind].copy().astype(np.int)
-        ul_min = box_scale[2:4] - SIZE
-        ul_max = box_scale[0:2]
-        assert ul_min[0]<=ul_max[0]
-        assert ul_min[1]<=ul_max[1]
-        #print('ul', ul_min, ul_max, box)
-        up, left = np.random.randint(ul_min[1], ul_max[1]+1), np.random.randint(ul_min[0], ul_max[0]+1)
-        #print('box', box, up, left)
-        M = [ [1.0, 0.0, -left],
-              [0.0, 1.0, -up], ]
-        M = np.array(M)
-        im = cv2.warpAffine(im, M, (SIZE, SIZE), borderValue = tuple(config.PIXEL_MEANS))
-        #tbox = np.array([left, left+SIZE, up, up+SIZE], dtype=np.int)
-        #im_new = np.zeros( (SIZE, SIZE,3), dtype=im.dtype)
-        #for i in range(3):
-        #  im_new[:,:,i] = config.PIXEL_MEANS[i]
-        new_rec['boxes'][:,0] -= left
-        new_rec['boxes'][:,2] -= left
-        new_rec['boxes'][:,1] -= up
-        new_rec['boxes'][:,3] -= up
-        box_trans = new_rec['boxes'][box_ind].copy().astype(np.int)
-        #print('sel box', im_scale, box, box_scale, box_trans, file=sys.stderr)
-        #print('before', new_rec['boxes'].shape[0])
-        boxes_new = []
-        classes_new = []
-        for i in range(new_rec['boxes'].shape[0]):
-          box = new_rec['boxes'][i]
-          box_size = max(box[2]-box[0], box[3]-box[1])
-          center = np.array(([box[0], box[1]]+[box[2], box[3]]))/2
-          if center[0]<0 or center[1]<0 or center[0]>=im.shape[1] or center[1]>=im.shape[0]:
-            continue
-          if box_size<config.TRAIN.MIN_BOX_SIZE:
-            continue
-          boxes_new.append(box)
-          classes_new.append(new_rec['gt_classes'][i])
-        new_rec['boxes'] = np.array(boxes_new)
-        new_rec['gt_classes'] = np.array(classes_new)
-        #print('after', new_rec['boxes'].shape[0])
-        #assert new_rec['boxes'].shape[0]>0
-        DEBUG = True
-        if DEBUG:
-          global TMP_ID
-          if TMP_ID<10:
-            tim = im.copy()
-            for i in range(new_rec['boxes'].shape[0]):
-              box = new_rec['boxes'][i].copy().astype(np.int)
-              cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]), (255, 0, 0), 1)
-            filename = './trainimages/train%d.png' % TMP_ID
-            TMP_ID+=1
-            cv2.imwrite(filename, tim)
-
-        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS, config.PIXEL_SCALE)
-
-        processed_ims.append(im_tensor)
-        #print('boxes', new_rec['boxes'], file=sys.stderr)
-        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
-        new_rec['im_info'] = im_info
-        processed_roidb.append(new_rec)
-    return processed_ims, processed_roidb
-
-def expand_bboxes(bboxes,
-                  image_width,
-                  image_height,
-                  expand_left=2.,
-                  expand_up=2.,
-                  expand_right=2.,
-                  expand_down=2.):
-    """
-    Expand bboxes, expand 2 times by defalut.
-    """
-    expand_boxes = []
-    for bbox in bboxes:
-        xmin = bbox[0]
-        ymin = bbox[1]
-        xmax = bbox[2]
-        ymax = bbox[3]
-        w = xmax - xmin
-        h = ymax - ymin
-        ex_xmin = max(xmin - w / expand_left, 0.)
-        ex_ymin = max(ymin - h / expand_up, 0.)
-        ex_xmax = min(xmax + w / expand_right, image_width)
-        ex_ymax = min(ymax + h / expand_down, image_height)
-        expand_boxes.append([ex_xmin, ex_ymin, ex_xmax, ex_ymax])
-    return expand_boxes
-
-def get_crop_image1(roidb):
-    """
-    preprocess image and return processed roidb
-    :param roidb: a list of roidb
-    :return: list of img as in mxnet format
-    roidb add new item['im_info']
-    0 --- x (width, second dim of im)
-    |
-    y (height, first dim of im)
-    """
-    #roidb and each roi_rec can not be changed as it will be reused in next epoch
-    num_images = len(roidb)
-    processed_ims = []
-    processed_roidb = []
-    for i in range(num_images):
-        roi_rec = roidb[i]
-        if 'stream' in roi_rec:
-          im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
-        else:
-          assert os.path.exists(roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
-          im = cv2.imread(roi_rec['image'])
-        if roidb[i]['flipped']:
-            im = im[:, ::-1, :]
-        if 'boxes_mask' in roi_rec:
-          #im = im.astype(np.float32)
-          boxes_mask = roi_rec['boxes_mask'].copy()
-          boxes_mask = boxes_mask.astype(np.int)
-          for j in range(boxes_mask.shape[0]):
-            m = boxes_mask[j]
-            im[m[1]:m[3],m[0]:m[2],:] = 127
-            #print('find mask', m, file=sys.stderr)
-        SIZE = config.SCALES[0][0]
-        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
-        #PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0, 0.8, 1.0, 0.8, 1.0]
-        _scale = random.choice(PRE_SCALES)
-        #_scale = np.random.uniform(PRE_SCALES[0], PRE_SCALES[-1])
-        size = int(np.min(im.shape[0:2])*_scale)
-        #size = int(np.round(_scale*np.min(im.shape[0:2])))
-        im_scale = float(SIZE)/size
-        #origin_im_scale = im_scale
-        #size = np.round(np.min(im.shape[0:2])*im_scale)
-        #im_scale *= (float(SIZE)/size)
-        origin_shape = im.shape
-        if _scale>10.0: #avoid im.size<SIZE, never?
-          sizex = int(np.round(im.shape[1]*im_scale))
-          sizey = int(np.round(im.shape[0]*im_scale))
-          if sizex<SIZE:
-            sizex = SIZE
-            print('keepx', sizex)
-          if sizey<SIZE:
-            sizey = SIZE
-            print('keepy', sizex)
-          im = cv2.resize(im, (sizex, sizey), interpolation=cv2.INTER_LINEAR)
-        else:
-          im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-
-        assert im.shape[0]>=SIZE and im.shape[1]>=SIZE
-        #print('image size', origin_shape, _scale, SIZE, size, im_scale)
-
-        new_rec = roi_rec.copy()
-        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
-        if config.FACE_LANDMARK:
-          new_rec['landmarks'] = roi_rec['landmarks'].copy()
-          new_rec['landmarks'][:,:,0:2] *= im_scale
-        retry = 0
-        LIMIT = 25
-        size = SIZE
-        while retry<LIMIT:
-          up, left = (np.random.randint(0, im.shape[0]-size+1), np.random.randint(0, im.shape[1]-size+1))
-          boxes_new = new_rec['boxes'].copy()
-          im_new = im[up:(up+size), left:(left+size), :]
-          #print('crop', up, left, size, im_scale)
-          boxes_new[:,0] -= left
-          boxes_new[:,2] -= left
-          boxes_new[:,1] -= up
-          boxes_new[:,3] -= up
-          if config.FACE_LANDMARK:
-            landmarks_new = new_rec['landmarks'].copy()
-            landmarks_new[:,:,0] -= left
-            landmarks_new[:,:,1] -= up
-            #for i in range(0,10,2):
-            #  landmarks_new[:,i] -= left
-            #for i in range(1,10,2):
-            #  landmarks_new[:,i] -= up
-            valid_landmarks = []
-          #im_new = cv2.resize(im_new, (SIZE, SIZE), interpolation=cv2.INTER_LINEAR)
-          #boxes_new *= im_scale
-          #print(origin_shape, im_new.shape, im_scale)
-          valid = []
-          valid_boxes = []
-          for i in range(boxes_new.shape[0]):
-            box = boxes_new[i]
-            #center = np.array(([box[0], box[1]]+[box[2], box[3]]))/2
-            centerx = (box[0]+box[2])/2
-            centery = (box[1]+box[3])/2
-
-            #box[0] = max(0, box[0])
-            #box[1] = max(0, box[1])
-            #box[2] = min(im_new.shape[1], box[2])
-            #box[3] = min(im_new.shape[0], box[3])
-            box_size = max(box[2]-box[0], box[3]-box[1])
-
-            if centerx<0 or centery<0 or centerx>=im_new.shape[1] or centery>=im_new.shape[0]:
-              continue
-            if box_size<config.TRAIN.MIN_BOX_SIZE:
-              continue
-            #filter by landmarks? TODO
-            valid.append(i)
-            valid_boxes.append(box)
-            if config.FACE_LANDMARK:
-              valid_landmarks.append(landmarks_new[i])
-          if len(valid)>0 or retry==LIMIT-1:
-            im = im_new
-            new_rec['boxes'] = np.array(valid_boxes)
-            new_rec['gt_classes'] = new_rec['gt_classes'][valid]
-            if config.FACE_LANDMARK:
-              new_rec['landmarks'] = np.array(valid_landmarks)
-            if config.HEAD_BOX:
-              face_box = new_rec['boxes']
-              head_box = expand_bboxes(face_box, image_width=im.shape[1], image_height=im.shape[0])
-              new_rec['boxes_head'] = np.array(head_box)
-            break
-
-          retry+=1
-
-        if config.COLOR_MODE>0 and config.COLOR_JITTERING>0.0:
-          im = im.astype(np.float32)
-          im = color_aug(im, config.COLOR_JITTERING)
-
-        #assert np.all(new_rec['landmarks'][:,10]>0.0)
-        global TMP_ID
-        if TMP_ID>=0 and TMP_ID<10:
-          tim = im.copy().astype(np.uint8)
-          for i in range(new_rec['boxes'].shape[0]):
-            box = new_rec['boxes'][i].copy().astype(np.int)
-            cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]), (255, 0, 0), 1)
-            print('draw box:', box)
-          if config.FACE_LANDMARK:
-            for i in range(new_rec['landmarks'].shape[0]):
-              landmark = new_rec['landmarks'][i].copy()
-              if landmark[0][2]<0:
-                print('zero', landmark)
-                continue
-              landmark = landmark.astype(np.int)
-              print('draw landmark', landmark)
-              for k in range(5):
-                color = (0, 0, 255)
-                if k==0 or k==3:
-                  color = (0, 255, 0)
-                pp = (landmark[k][0], landmark[k][1])
-                cv2.circle(tim, (pp[0], pp[1]), 1, color, 2)
-          filename = './trainimages/train%d.png' % TMP_ID
-          print('write', filename)
-          cv2.imwrite(filename, tim)
-          TMP_ID+=1
-
-        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS, config.PIXEL_SCALE)
-
-        processed_ims.append(im_tensor)
-        #print('boxes', new_rec['boxes'], file=sys.stderr)
-        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
-        new_rec['im_info'] = np.array(im_info, dtype=np.float32)
-        processed_roidb.append(new_rec)
-    return processed_ims, processed_roidb
-
-def get_crop_image2(roidb):
-    """
-    preprocess image and return processed roidb
-    :param roidb: a list of roidb
-    :return: list of img as in mxnet format
-    roidb add new item['im_info']
-    0 --- x (width, second dim of im)
-    |
-    y (height, first dim of im)
-    """
-    #roidb and each roi_rec can not be changed as it will be reused in next epoch
-    num_images = len(roidb)
-    processed_ims = []
-    processed_roidb = []
-    for i in range(num_images):
-        roi_rec = roidb[i]
-        if 'stream' in roi_rec:
-          im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
-        else:
-          assert os.path.exists(roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
-          im = cv2.imread(roi_rec['image'])
-        if roidb[i]['flipped']:
-            im = im[:, ::-1, :]
-        if 'boxes_mask' in roi_rec:
-          #im = im.astype(np.float32)
-          boxes_mask = roi_rec['boxes_mask'].copy()
-          boxes_mask = boxes_mask.astype(np.int)
-          for j in range(boxes_mask.shape[0]):
-            m = boxes_mask[j]
-            im[m[1]:m[3],m[0]:m[2],:] = 0
-            #print('find mask', m, file=sys.stderr)
-        SIZE = config.SCALES[0][0]
-        scale_array = np.array([16,32,64,128,256,512], dtype=np.float32)
-        candidates = []
-        for i in range(roi_rec['boxes'].shape[0]):
-          box = roi_rec['boxes'][i]
-          box_size = max(box[2]-box[0], box[3]-box[1])
-          if box_size<config.TRAIN.MIN_BOX_SIZE:
-            continue
-          #if box[0]<0 or box[1]<0:
-          #  continue
-          #if box[2]>im.shape[1] or box[3]>im.shape[0]:
-          #  continue;
-          candidates.append(i)
-        assert len(candidates)>0
-        box_ind = random.choice(candidates)
-        box = roi_rec['boxes'][box_ind]
-        width = box[2]-box[0]
-        height = box[3]-box[1]
-        wid = width
-        hei = height
-        resize_width, resize_height = config.SCALES[0]
-        image_width = im.shape[0]
-        image_height = im.shape[1]
-        area = width*height
-        range_size = 0
-        for scale_ind in range(0, len(scale_array) - 1):
-            if area > scale_array[scale_ind] ** 2 and area < \
-                    scale_array[scale_ind + 1] ** 2:
-                range_size = scale_ind + 1
-                break
-
-        if area > scale_array[len(scale_array) - 2]**2:
-            range_size = len(scale_array) - 2
-        scale_choose = 0.0
-        if range_size == 0:
-            rand_idx_size = 0
-        else:
-            # np.random.randint range: [low, high)
-            rng_rand_size = np.random.randint(0, range_size + 1)
-            rand_idx_size = rng_rand_size % (range_size + 1)
-
-        if rand_idx_size == range_size:
-            min_resize_val = scale_array[rand_idx_size] / 2.0
-            max_resize_val = min(2.0 * scale_array[rand_idx_size],
-                                 2 * math.sqrt(wid * hei))
-            scale_choose = random.uniform(min_resize_val, max_resize_val)
-        else:
-            min_resize_val = scale_array[rand_idx_size] / 2.0
-            max_resize_val = 2.0 * scale_array[rand_idx_size]
-            scale_choose = random.uniform(min_resize_val, max_resize_val)
-
-        sample_bbox_size = wid * resize_width / scale_choose
-
-        w_off_orig = 0.0
-        h_off_orig = 0.0
-        if sample_bbox_size < max(image_height, image_width):
-            if wid <= sample_bbox_size:
-                w_off_orig = np.random.uniform(xmin + wid - sample_bbox_size,
-                                               xmin)
-            else:
-                w_off_orig = np.random.uniform(xmin,
-                                               xmin + wid - sample_bbox_size)
-
-            if hei <= sample_bbox_size:
-                h_off_orig = np.random.uniform(ymin + hei - sample_bbox_size,
-                                               ymin)
-            else:
-                h_off_orig = np.random.uniform(ymin,
-                                               ymin + hei - sample_bbox_size)
-
-        else:
-            w_off_orig = np.random.uniform(image_width - sample_bbox_size, 0.0)
-            h_off_orig = np.random.uniform(image_height - sample_bbox_size, 0.0)
-
-        w_off_orig = math.floor(w_off_orig)
-        h_off_orig = math.floor(h_off_orig)
-
-        # Figure out top left coordinates.
-        w_off = 0.0
-        h_off = 0.0
-        w_off = float(w_off_orig / image_width)
-        h_off = float(h_off_orig / image_height)
-        im_new = im[up:(up+size), left:(left+size), :]
-
-        sampled_bbox = bbox(w_off, h_off,
-                            w_off + float(sample_bbox_size / image_width),
-                            h_off + float(sample_bbox_size / image_height))
-        return sampled_bbox
-
-        box_size = max(box[2]-box[0], box[3]-box[1])
-        dist = np.abs(TARGET_BOX_SCALES - box_size)
-        nearest = np.argmin(dist)
-        target_ind = random.randrange(min(len(TARGET_BOX_SCALES), nearest+2))
-        target_box_size = TARGET_BOX_SCALES[target_ind]
-        im_scale = float(target_box_size) / box_size
-        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
-        _scale = random.choice(PRE_SCALES)
-        #_scale = np.random.uniform(PRE_SCALES[0], PRE_SCALES[-1])
-        size = int(np.round(_scale*np.min(im.shape[0:2])))
-        im_scale = float(SIZE)/size
-        #origin_im_scale = im_scale
-        #size = np.round(np.min(im.shape[0:2])*im_scale)
-        #im_scale *= (float(SIZE)/size)
-        origin_shape = im.shape
-        if _scale>10.0: #avoid im.size<SIZE, never?
-          sizex = int(np.round(im.shape[1]*im_scale))
-          sizey = int(np.round(im.shape[0]*im_scale))
-          if sizex<SIZE:
-            sizex = SIZE
-            print('keepx', sizex)
-          if sizey<SIZE:
-            sizey = SIZE
-            print('keepy', sizex)
-          im = cv2.resize(im, (sizex, sizey), interpolation=cv2.INTER_LINEAR)
-        else:
-          im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-        assert im.shape[0]>=SIZE and im.shape[1]>=SIZE
-
-        new_rec = roi_rec.copy()
-        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
-        if config.FACE_LANDMARK:
-          new_rec['landmarks'] = roi_rec['landmarks'].copy() * im_scale
-        retry = 0
-        LIMIT = 25
-        size = SIZE
-        while retry<LIMIT:
-          up, left = (np.random.randint(0, im.shape[0]-size+1), np.random.randint(0, im.shape[1]-size+1))
-          boxes_new = new_rec['boxes'].copy()
-          im_new = im[up:(up+size), left:(left+size), :]
-          #print('crop', up, left, size, im_scale)
-          boxes_new[:,0] -= left
-          boxes_new[:,2] -= left
-          boxes_new[:,1] -= up
-          boxes_new[:,3] -= up
-          if config.FACE_LANDMARK:
-            landmarks_new = new_rec['landmarks'].copy()
-            for i in range(0,10,2):
-              landmarks_new[:,i] -= left
-            for i in range(1,10,2):
-              landmarks_new[:,i] -= up
-            valid_landmarks = []
-          #im_new = cv2.resize(im_new, (SIZE, SIZE), interpolation=cv2.INTER_LINEAR)
-          #boxes_new *= im_scale
-          #print(origin_shape, im_new.shape, im_scale)
-          valid = []
-          valid_boxes = []
-          for i in range(boxes_new.shape[0]):
-            box = boxes_new[i]
-            #center = np.array(([box[0], box[1]]+[box[2], box[3]]))/2
-            centerx = (box[0]+box[2])/2
-            centery = (box[1]+box[3])/2
-
-            #box[0] = max(0, box[0])
-            #box[1] = max(0, box[1])
-            #box[2] = min(im_new.shape[1], box[2])
-            #box[3] = min(im_new.shape[0], box[3])
-            box_size = max(box[2]-box[0], box[3]-box[1])
-
-            if centerx<0 or centery<0 or centerx>=im_new.shape[1] or centery>=im_new.shape[0]:
-              continue
-            if box_size<config.TRAIN.MIN_BOX_SIZE:
-              continue
-            #filter by landmarks? TODO
-            valid.append(i)
-            valid_boxes.append(box)
-            if config.FACE_LANDMARK:
-              valid_landmarks.append(landmarks_new[i])
-          if len(valid)>0 or retry==LIMIT-1:
-            im = im_new
-            new_rec['boxes'] = np.array(valid_boxes)
-            new_rec['gt_classes'] = new_rec['gt_classes'][valid]
-            if config.FACE_LANDMARK:
-              new_rec['landmarks'] = np.array(valid_landmarks)
-            break
-
-          retry+=1
-
-        if config.COLOR_JITTERING>0.0:
-          im = im.astype(np.float32)
-          im = color_aug(im, config.COLOR_JITTERING)
-
-        #assert np.all(new_rec['landmarks'][:,10]>0.0)
-        global TMP_ID
-        if TMP_ID>=0 and TMP_ID<10:
-          tim = im.copy().astype(np.uint8)
-          for i in range(new_rec['boxes'].shape[0]):
-            box = new_rec['boxes'][i].copy().astype(np.int)
-            cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]), (255, 0, 0), 1)
-            print('draw box:', box)
-          if config.FACE_LANDMARK:
-            for i in range(new_rec['landmarks'].shape[0]):
-              landmark = new_rec['landmarks'][i].copy()
-              if landmark[10]==0.0:
-                print('zero', landmark)
-                continue
-              landmark = landmark.astype(np.int)
-              print('draw landmark', landmark)
-              for k in range(5):
-                color = (0, 0, 255)
-                if k==0 or k==3:
-                  color = (0, 255, 0)
-                pp = (landmark[k*2], landmark[1+k*2])
-                cv2.circle(tim, (pp[0], pp[1]), 1, color, 2)
-          filename = './trainimages/train%d.png' % TMP_ID
-          print('write', filename)
-          cv2.imwrite(filename, tim)
-          TMP_ID+=1
-
-        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS, config.PIXEL_SCALE)
-
-        processed_ims.append(im_tensor)
-        #print('boxes', new_rec['boxes'], file=sys.stderr)
-        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
-        new_rec['im_info'] = np.array(im_info, dtype=np.float32)
-        processed_roidb.append(new_rec)
-    return processed_ims, processed_roidb
-
-def do_mixup(im1, roidb1, im2, roidb2):
-  im = (im1+im2)/2.0
-  roidb = {}
-  #print(roidb1.keys())
-  #for k in roidb1:
-  for k in ['boxes', 'landmarks', 'gt_classes', 'im_info']:
-    v1 = roidb1[k]
-    v2 = roidb2[k]
-    if k!='im_info':
-      #print('try', k, v1.shape, v2.shape)
-      if v1.shape[0]>0 and v2.shape[0]>0:
-        v = np.concatenate( (v1, v2), axis=0 )
-      else:
-        v = v1
-    else:
-      v = v1
-    #print(k, v1.shape, v2.shape, v.shape)
-    roidb[k] = v
-  return im, roidb
-
-def get_crop_image(roidb):
-    ims, roidbs = get_crop_image1(roidb)
-    if config.MIXUP>0.0 and np.random.random()<config.MIXUP:
-      for i in range(len(ims)):
-        im = ims[i]
-        roidb = roidbs[i]
-        j = np.random.randint(0, len(ims)-1)
-        if j>=i:
-          j+=1
-        im, roidb = do_mixup(im, roidb, ims[j], roidbs[j])
-        ims[i] = im
-        roidbs[i] = roidb
-    return ims, roidbs
-
-def resize(im, target_size, max_size, stride=0, min_size=0):
-    """
-    only resize input image to target size and return scale
-    :param im: BGR image input by opencv
-    :param target_size: one dimensional size (the short side)
-    :param max_size: one dimensional max size (the long side)
-    :param stride: if given, pad the image to designated stride
-    :return:
-    """
-    im_shape = im.shape
-    im_size_min = np.min(im_shape[0:2])
-    im_size_max = np.max(im_shape[0:2])
-    im_scale = float(target_size) / float(im_size_min)
-    # prevent bigger axis from being more than max_size:
-    if np.round(im_scale * im_size_max) > max_size:
-        im_scale = float(max_size) / float(im_size_max)
-        if min_size>0 and np.round(im_scale*im_size_min)<min_size:
-          im_scale = float(min_size) / float(im_size_min)
-    im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-
-    if stride == 0:
-        return im, im_scale
-    else:
-        # pad to product of stride
-        im_height = int(np.ceil(im.shape[0] / float(stride)) * stride)
-        im_width = int(np.ceil(im.shape[1] / float(stride)) * stride)
-        im_channel = im.shape[2]
-        padded_im = np.zeros((im_height, im_width, im_channel))
-        padded_im[:im.shape[0], :im.shape[1], :] = im
-        return padded_im, im_scale
-
-
-def transform(im, pixel_means, pixel_stds, pixel_scale):
-    """
-    transform into mxnet tensor,
-    subtract pixel size and transform to correct format
-    :param im: [height, width, channel] in BGR
-    :param pixel_means: [B, G, R pixel means]
-    :return: [batch, channel, height, width]
-    """
-    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-    for i in range(3):
-        im_tensor[0, i, :, :] = (im[:, :, 2 - i]/pixel_scale - pixel_means[2 - i]) / pixel_stds[2 - i]
-    return im_tensor
-
-
-def transform_inverse(im_tensor, pixel_means):
-    """
-    transform from mxnet im_tensor to ordinary RGB image
-    im_tensor is limited to one image
-    :param im_tensor: [batch, channel, height, width]
-    :param pixel_means: [B, G, R pixel means]
-    :return: im [height, width, channel(RGB)]
-    """
-    assert im_tensor.shape[0] == 1
-    im_tensor = im_tensor.copy()
-    # put channel back
-    channel_swap = (0, 2, 3, 1)
-    im_tensor = im_tensor.transpose(channel_swap)
-    im = im_tensor[0]
-    assert im.shape[2] == 3
-    im += pixel_means[[2, 1, 0]]
-    im = im.astype(np.uint8)
-    return im
-
-
-def tensor_vstack(tensor_list, pad=0):
-    """
-    vertically stack tensors
-    :param tensor_list: list of tensor to be stacked vertically
-    :param pad: label to pad with
-    :return: tensor with max shape
-    """
-    ndim = len(tensor_list[0].shape)
-    dtype = tensor_list[0].dtype
-    islice = tensor_list[0].shape[0]
-    dimensions = []
-    first_dim = sum([tensor.shape[0] for tensor in tensor_list])
-    dimensions.append(first_dim)
-    for dim in range(1, ndim):
-        dimensions.append(max([tensor.shape[dim] for tensor in tensor_list]))
-    if pad == 0:
-        all_tensor = np.zeros(tuple(dimensions), dtype=dtype)
-    elif pad == 1:
-        all_tensor = np.ones(tuple(dimensions), dtype=dtype)
-    else:
-        all_tensor = np.full(tuple(dimensions), pad, dtype=dtype)
-    if ndim == 1:
-        for ind, tensor in enumerate(tensor_list):
-            all_tensor[ind*islice:(ind+1)*islice] = tensor
-    elif ndim == 2:
-        for ind, tensor in enumerate(tensor_list):
-            all_tensor[ind*islice:(ind+1)*islice, :tensor.shape[1]] = tensor
-    elif ndim == 3:
-        for ind, tensor in enumerate(tensor_list):
-            all_tensor[ind*islice:(ind+1)*islice, :tensor.shape[1], :tensor.shape[2]] = tensor
-    elif ndim == 4:
-        for ind, tensor in enumerate(tensor_list):
-            all_tensor[ind*islice:(ind+1)*islice, :tensor.shape[1], :tensor.shape[2], :tensor.shape[3]] = tensor
-    elif ndim == 5:
-        for ind, tensor in enumerate(tensor_list):
-            all_tensor[ind*islice:(ind+1)*islice, :tensor.shape[1], :tensor.shape[2], :tensor.shape[3], :tensor.shape[4]] = tensor
-    else:
-        print(tensor_list[0].shape)
-        raise Exception('Sorry, unimplemented.')
-    return all_tensor
-
diff --git a/RetinaFace/rcnn/io/rcnn.py b/RetinaFace/rcnn/io/rcnn.py
deleted file mode 100644
index 298df8e85..000000000
--- a/RetinaFace/rcnn/io/rcnn.py
+++ /dev/null
@@ -1,579 +0,0 @@
-"""
-Fast R-CNN:
-data =
-    {'data': [num_images, c, h, w],
-    'rois': [num_rois, 5]}
-label =
-    {'label': [num_rois],
-    'bbox_target': [num_rois, 4 * num_classes],
-    'bbox_weight': [num_rois, 4 * num_classes]}
-roidb extended format [image_index]
-    ['image', 'height', 'width', 'flipped',
-     'boxes', 'gt_classes', 'gt_overlaps', 'max_classes', 'max_overlaps', 'bbox_targets']
-"""
-
-import numpy as np
-import numpy.random as npr
-
-from ..config import config
-from ..io.image import get_image, tensor_vstack
-from ..processing.bbox_transform import bbox_overlaps, bbox_transform
-from ..processing.bbox_regression import expand_bbox_regression_targets
-
-
-def get_rcnn_testbatch(roidb):
-    """
-    return a dict of testbatch
-    :param roidb: ['image', 'flipped'] + ['boxes']
-    :return: data, label, im_info
-    """
-    assert len(roidb) == 1, 'Single batch only'
-    imgs, roidb = get_image(roidb)
-    im_array = imgs[0]
-    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
-
-    im_rois = roidb[0]['boxes']
-    rois = im_rois
-    batch_index = 0 * np.ones((rois.shape[0], 1))
-    rois_array = np.hstack((batch_index, rois))[np.newaxis, :]
-
-    data = {'data': im_array,
-            'rois': rois_array,
-            'im_info': im_info}
-    label = {}
-
-    return data, label
-
-
-def get_rcnn_batch(roidb):
-    """
-    return a dict of multiple images
-    :param roidb: a list of dict, whose length controls batch size
-    ['images', 'flipped'] + ['gt_boxes', 'boxes', 'gt_overlap'] => ['bbox_targets']
-    :return: data, label
-    """
-    num_images = len(roidb)
-    imgs, roidb = get_image(roidb)
-    im_array = tensor_vstack(imgs)
-
-    assert config.TRAIN.BATCH_ROIS % config.TRAIN.BATCH_IMAGES == 0, \
-        'BATCHIMAGES {} must divide BATCH_ROIS {}'.format(config.TRAIN.BATCH_IMAGES, config.TRAIN.BATCH_ROIS)
-    rois_per_image = int(config.TRAIN.BATCH_ROIS / config.TRAIN.BATCH_IMAGES)
-    fg_rois_per_image = int(round(config.TRAIN.FG_FRACTION * rois_per_image))
-
-    rois_array = list()
-    labels_array = list()
-    bbox_targets_array = list()
-    bbox_weights_array = list()
-
-    for im_i in range(num_images):
-        roi_rec = roidb[im_i]
-
-        # infer num_classes from gt_overlaps
-        num_classes = roi_rec['gt_overlaps'].shape[1]
-
-        # label = class RoI has max overlap with
-        rois = roi_rec['boxes']
-        labels = roi_rec['max_classes']
-        overlaps = roi_rec['max_overlaps']
-        bbox_targets = roi_rec['bbox_targets']
-
-        im_rois, labels, bbox_targets, bbox_weights = \
-            sample_rois(rois, fg_rois_per_image, rois_per_image, num_classes,
-                        labels, overlaps, bbox_targets)
-
-        # project im_rois
-        # do not round roi
-        rois = im_rois
-        batch_index = im_i * np.ones((rois.shape[0], 1))
-        rois_array_this_image = np.hstack((batch_index, rois))
-        rois_array.append(rois_array_this_image)
-
-        # add labels
-        labels_array.append(labels)
-        bbox_targets_array.append(bbox_targets)
-        bbox_weights_array.append(bbox_weights)
-
-    rois_array = np.array(rois_array)
-    labels_array = np.array(labels_array)
-    bbox_targets_array = np.array(bbox_targets_array)
-    bbox_weights_array = np.array(bbox_weights_array)
-
-    data = {'data': im_array,
-            'rois': rois_array}
-    label = {'label': labels_array,
-             'bbox_target': bbox_targets_array,
-             'bbox_weight': bbox_weights_array}
-
-    return data, label
-
-
-def sample_rois(rois, fg_rois_per_image, rois_per_image, num_classes,
-                labels=None, overlaps=None, bbox_targets=None, gt_boxes=None):
-    """
-    generate random sample of ROIs comprising foreground and background examples
-    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
-    :param fg_rois_per_image: foreground roi number
-    :param rois_per_image: total roi number
-    :param num_classes: number of classes
-    :param labels: maybe precomputed
-    :param overlaps: maybe precomputed (max_overlaps)
-    :param bbox_targets: maybe precomputed
-    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
-    :return: (labels, rois, bbox_targets, bbox_weights)
-    """
-    if labels is None:
-        overlaps = bbox_overlaps(rois[:, 1:].astype(np.float), gt_boxes[:, :4].astype(np.float))
-        gt_assignment = overlaps.argmax(axis=1)
-        overlaps = overlaps.max(axis=1)
-        labels = gt_boxes[gt_assignment, 4]
-
-    # foreground RoI with FG_THRESH overlap
-    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
-    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
-    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
-    # Sample foreground regions without replacement
-    if len(fg_indexes) > fg_rois_per_this_image:
-        fg_indexes = npr.choice(fg_indexes, size=fg_rois_per_this_image, replace=False)
-
-    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
-    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI) & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
-    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
-    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
-    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image, bg_indexes.size)
-    # Sample foreground regions without replacement
-    if len(bg_indexes) > bg_rois_per_this_image:
-        bg_indexes = npr.choice(bg_indexes, size=bg_rois_per_this_image, replace=False)
-
-    # indexes selected
-    keep_indexes = np.append(fg_indexes, bg_indexes)
-    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
-    neg_rois = rois[neg_idx]
-    # pad more to ensure a fixed minibatch size
-    while keep_indexes.shape[0] < rois_per_image:
-        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
-        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
-        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
-
-    # select labels
-    labels = labels[keep_indexes]
-    # set labels of bg_rois to be 0
-    labels[fg_rois_per_this_image:] = 0
-    rois = rois[keep_indexes]
-
-    # load or compute bbox_target
-    if bbox_targets is not None:
-        bbox_target_data = bbox_targets[keep_indexes, :]
-    else:
-        targets = bbox_transform(rois[:, 1:], gt_boxes[gt_assignment[keep_indexes], :4])
-        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS))
-                       / np.array(config.TRAIN.BBOX_STDS))
-        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
-
-    bbox_targets, bbox_weights = \
-        expand_bbox_regression_targets(bbox_target_data, num_classes)
-
-    return rois, labels, bbox_targets, bbox_weights
-
-def get_fpn_rcnn_testbatch(roidb):
-    """
-    return a dict of testbatch
-    :param roidb: ['image', 'flipped'] + ['boxes']
-    :return: data, label, im_info
-    """
-    assert len(roidb) == 1, 'Single batch only'
-    imgs, roidb = get_image(roidb)
-    im_array = imgs[0]
-    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
-
-    im_rois = roidb[0]['boxes']
-    rois = im_rois
-
-    # assign rois
-    rois_area = np.sqrt((rois[:, 2] - rois[:, 0]) * (rois[:, 3] - rois[:, 1]))
-    area_threshold = {'P5': 448, 'P4': 224, 'P3': 112}
-    rois_p5 = rois[area_threshold['P5'] <= rois_area]
-    rois_p4 = rois[np.logical_and(area_threshold['P4'] <= rois_area, rois_area < area_threshold['P5'])]
-    rois_p3 = rois[np.logical_and(area_threshold['P3'] <= rois_area, rois_area < area_threshold['P4'])]
-    rois_p2 = rois[np.logical_and(0 < rois_area, rois_area < area_threshold['P3'])]
-
-    # pad a virtual rois if on rois assigned
-    if rois_p5.size == 0:
-        rois_p5 = np.array([[12,34,56,78]])
-    if rois_p4.size == 0:
-        rois_p4 = np.array([[12,34,56,78]])
-    if rois_p3.size == 0:
-        rois_p3 = np.array([[12,34,56,78]])
-    if rois_p2.size == 0:
-        rois_p2 = np.array([[12,34,56,78]])
-
-    p5_batch_index = 0 * np.ones((rois_p5.shape[0], 1))
-    rois_p5_array = np.hstack((p5_batch_index, rois_p5))[np.newaxis, :]
-
-    p4_batch_index = 0 * np.ones((rois_p4.shape[0], 1))
-    rois_p4_array = np.hstack((p4_batch_index, rois_p4))[np.newaxis, :]
-
-    p3_batch_index = 0 * np.ones((rois_p3.shape[0], 1))
-    rois_p3_array = np.hstack((p3_batch_index, rois_p3))[np.newaxis, :]
-
-    p2_batch_index = 0 * np.ones((rois_p2.shape[0], 1))
-    rois_p2_array = np.hstack((p2_batch_index, rois_p2))[np.newaxis, :]
-
-    data = {'data': im_array,
-            'rois_stride32': rois_p5_array,
-            'rois_stride16': rois_p4_array,
-            'rois_stride8': rois_p3_array,
-            'rois_stride4': rois_p2_array}
-    label = {}
-
-    return data, label, im_info
-
-def get_fpn_maskrcnn_batch(roidb):
-    """
-    return a dictionary that contains raw data.
-    """
-    num_images = len(roidb)
-    imgs, roidb = get_image(roidb, scale=config.TRAIN.SCALE) #TODO
-    #imgs, roidb = get_image(roidb)
-    im_array = tensor_vstack(imgs)
-
-    assert config.TRAIN.BATCH_ROIS % config.TRAIN.BATCH_IMAGES == 0, \
-        'BATCHIMAGES {} must divide BATCH_ROIS {}'.format(config.TRAIN.BATCH_IMAGES, config.TRAIN.BATCH_ROIS)
-    rois_per_image = config.TRAIN.BATCH_ROIS / config.TRAIN.BATCH_IMAGES
-    fg_rois_per_image = np.round(config.TRAIN.FG_FRACTION * rois_per_image).astype(int)
-
-    rois_on_imgs = dict()
-    labels_on_imgs = dict()
-    bbox_targets_on_imgs = dict()
-    bbox_weights_on_imgs = dict()
-    mask_targets_on_imgs = dict()
-    mask_weights_on_imgs = dict()
-    for s in config.RCNN_FEAT_STRIDE:
-        rois_on_imgs.update({'stride%s' % s : list()})
-        labels_on_imgs.update({'stride%s' % s : list()})
-        bbox_targets_on_imgs.update({'stride%s' % s : list()})
-        bbox_weights_on_imgs.update({'stride%s' % s : list()})
-        mask_targets_on_imgs.update({'stride%s' % s : list()})
-        mask_weights_on_imgs.update({'stride%s' % s : list()})
-
-    # Sample rois
-    level_related_data_on_imgs = {}
-    for im_i in range(num_images):
-        roi_rec = roidb[im_i]
-        # infer num_classes from gt_overlaps
-        num_classes = roi_rec['gt_overlaps'].shape[1]
-        # label = class RoI has max overlap with
-        rois = roi_rec['boxes']
-        labels = roi_rec['max_classes']
-        overlaps = roi_rec['max_overlaps']
-        bbox_targets = roi_rec['bbox_targets']
-        im_info = roi_rec['im_info']
-
-        mask_targets = roi_rec['mask_targets']
-        mask_labels = roi_rec['mask_labels']
-        mask_inds = roi_rec['mask_inds']
-
-        assign_levels = roi_rec['assign_levels']
-
-        im_rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels, mask_targets_on_levels, mask_weights_on_levels = \
-            sample_rois_fpn(rois, assign_levels, fg_rois_per_image, rois_per_image, num_classes,
-                            labels, overlaps, bbox_targets, mask_targets=mask_targets, mask_labels=mask_labels, mask_inds=mask_inds, im_info=im_info)
-
-        level_related_data_on_imgs.update({'img_%s' % im_i: {'rois_on_levels': im_rois_on_levels,
-                                                             'labels_on_levels': labels_on_levels,
-                                                             'bbox_targets_on_levels': bbox_targets_on_levels,
-                                                             'bbox_weights_on_levels': bbox_weights_on_levels,
-                                                             'mask_targets_on_levels': mask_targets_on_levels,
-                                                             'mask_weights_on_levels': mask_weights_on_levels, }})
-
-    return im_array, level_related_data_on_imgs
-
-
-def sample_rois(rois, fg_rois_per_image, rois_per_image, num_classes,
-                labels=None, overlaps=None, bbox_targets=None, gt_boxes=None, mask_targets=None,
-                mask_labels=None, mask_inds=None):
-    """
-    generate random sample of ROIs comprising foreground and background examples
-    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
-    :param fg_rois_per_image: foreground roi number
-    :param rois_per_image: total roi number
-    :param num_classes: number of classes
-    :param labels: maybe precomputed
-    :param overlaps: maybe precomputed (max_overlaps)
-    :param bbox_targets: maybe precomputed
-    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
-    :return: (rois, labels, bbox_targets, bbox_weights)
-    """
-    if labels is None:
-        if len(gt_boxes) == 0:
-            gt_boxes = np.zeros((1, 5))
-            gt_assignment = np.zeros((len(rois), ), dtype=np.int32)
-            overlaps = np.zeros((len(rois), ))
-            labels = np.zeros((len(rois), ))
-        else:
-            overlaps = bbox_overlaps(rois[:, 1:].astype(np.float), gt_boxes[:, :4].astype(np.float))
-            gt_assignment = overlaps.argmax(axis=1)
-            overlaps = overlaps.max(axis=1)
-            labels = gt_boxes[gt_assignment, 4]
-
-    num_rois = rois.shape[0]
-    # foreground RoI with FG_THRESH overlap
-    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
-    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
-    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
-    # Sample foreground regions without replacement
-    if len(fg_indexes) > fg_rois_per_this_image:
-        fg_indexes = npr.choice(fg_indexes, size=fg_rois_per_this_image, replace=False)
-
-    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
-    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI) & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
-    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
-    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
-    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image, bg_indexes.size)
-    # Sample foreground regions without replacement
-    if len(bg_indexes) > bg_rois_per_this_image:
-        bg_indexes = npr.choice(bg_indexes, size=bg_rois_per_this_image, replace=False)
-
-    # indexes selected
-    keep_indexes = np.append(fg_indexes, bg_indexes)
-
-    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
-    neg_rois = rois[neg_idx]
-
-    # pad more to ensure a fixed minibatch size
-    while keep_indexes.shape[0] < rois_per_image:
-        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
-        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
-        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
-
-    # select labels
-    labels = labels[keep_indexes]
-    # set labels of bg_rois to be 0
-    labels[fg_rois_per_this_image:] = 0
-    rois = rois[keep_indexes]
-    if mask_targets is not None:
-        assert mask_labels is not None
-        assert mask_inds is not None
-        def _mask_umap(mask_targets, mask_labels, mask_inds):
-            _mask_targets = np.zeros((num_rois, num_classes, 28, 28), dtype=np.int8)
-            _mask_weights = np.zeros((num_rois, num_classes, 28, 28), dtype=np.int8)
-            _mask_targets[mask_inds, mask_labels] = mask_targets
-            _mask_weights[mask_inds, mask_labels] = 1
-            _mask_weights[:, 0] = 0 # set background mask weight to zeros
-            return _mask_targets, _mask_weights # [num_rois, num_classes, 28, 28]
-        mask_targets, mask_weights = _mask_umap(mask_targets, mask_labels, mask_inds)
-        mask_targets = mask_targets[keep_indexes]
-        mask_weights = mask_weights[keep_indexes]
-
-    # load or compute bbox_target
-    if bbox_targets is not None:
-        bbox_target_data = bbox_targets[keep_indexes, :]
-    else:
-        targets = bbox_transform(rois[:, 1:], gt_boxes[gt_assignment[keep_indexes], :4])
-        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS))
-                       / np.array(config.TRAIN.BBOX_STDS))
-        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
-
-    bbox_targets, bbox_weights = \
-        expand_bbox_regression_targets(bbox_target_data, num_classes)
-    
-    if mask_targets is not None:
-        return rois, labels, bbox_targets, bbox_weights, mask_targets, mask_weights
-    else:
-        return rois, labels, bbox_targets, bbox_weights
-
-def sample_rois_fpn(rois, assign_levels, fg_rois_per_image, rois_per_image, num_classes,
-                    labels=None, overlaps=None, bbox_targets=None, mask_targets=None, mask_labels=None, mask_inds=None, gt_boxes=None, im_info=None):
-    """
-    generate random sample of ROIs comprising foreground and background examples
-    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
-    :param assign_levels: [n]
-    :param fg_rois_per_image: foreground roi number
-    :param rois_per_image: total roi number
-    :param num_classes: number of classes
-    :param labels: maybe precomputed
-    :param overlaps: maybe precomputed (max_overlaps)
-    :param bbox_targets: maybe precomputed
-    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
-    :return: (rois, labels, bbox_targets, bbox_weights)
-    """
-    DEBUG = False
-    if labels is None:
-        if len(gt_boxes) == 0:
-            gt_boxes = np.zeros((1, 5))
-            gt_assignment = np.zeros((len(rois), ), dtype=np.int32)
-            overlaps = np.zeros((len(rois), ))
-            labels = np.zeros((len(rois), ))
-        else:
-            overlaps = bbox_overlaps(rois[:, 1:].astype(np.float), gt_boxes[:, :4].astype(np.float))
-            gt_assignment = overlaps.argmax(axis=1)
-            overlaps = overlaps.max(axis=1)
-            labels = gt_boxes[gt_assignment, 4]
-
-    num_rois = rois.shape[0]
-    # foreground RoI with FG_THRESH overlap
-    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
-    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
-    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
-
-    if DEBUG:
-        print 'fg total num:', len(fg_indexes)
-
-    # Sample foreground regions without replacement
-    if len(fg_indexes) > fg_rois_per_this_image:
-        fg_indexes = npr.choice(fg_indexes, size=fg_rois_per_this_image, replace=False)
-
-    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
-    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI) & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
-    if DEBUG:
-        print 'bg total num:', len(bg_indexes)
-    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
-    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
-    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image, bg_indexes.size)
-    # Sample foreground regions without replacement
-    if len(bg_indexes) > bg_rois_per_this_image:
-        bg_indexes = npr.choice(bg_indexes, size=bg_rois_per_this_image, replace=False)
-    if DEBUG:
-        print 'fg num:', len(fg_indexes)
-        print 'bg num:', len(bg_indexes)
-
-    # bg rois statistics
-    if DEBUG:
-        bg_assign = assign_levels[bg_indexes]
-        bg_rois_on_levels = dict()
-        for i, s in enumerate(config.RCNN_FEAT_STRIDE):
-            bg_rois_on_levels.update({'stride%s'%s:len(np.where(bg_assign == s)[0])})
-        print bg_rois_on_levels
-
-    # indexes selected
-    keep_indexes = np.append(fg_indexes, bg_indexes)
-
-    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
-    neg_rois = rois[neg_idx]
-
-    # pad more to ensure a fixed minibatch size
-    while keep_indexes.shape[0] < rois_per_image:
-        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
-        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
-        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
-
-    # select labels
-    labels = labels[keep_indexes]
-    # set labels of bg_rois to be 0
-    labels[fg_rois_per_this_image:] = 0
-    rois = rois[keep_indexes]
-    assign_levels = assign_levels[keep_indexes]
-
-    if mask_targets is not None:
-        assert mask_labels is not None
-        assert mask_inds is not None
-        def _mask_umap(mask_targets, mask_labels, mask_inds):
-            _mask_targets = np.zeros((num_rois, num_classes, 28, 28), dtype=np.int8)
-            _mask_weights = np.zeros((num_rois, num_classes, 1, 1), dtype=np.int8)
-            _mask_targets[mask_inds, mask_labels] = mask_targets
-            _mask_weights[mask_inds, mask_labels] = 1
-            return _mask_targets, _mask_weights # [num_rois, num_classes, 28, 28]
-        mask_targets, mask_weights = _mask_umap(mask_targets, mask_labels, mask_inds)
-        mask_targets = mask_targets[keep_indexes]
-        mask_weights = mask_weights[keep_indexes]
-
-    # load or compute bbox_target
-    if bbox_targets is not None:
-        bbox_target_data = bbox_targets[keep_indexes, :]
-    else:
-        targets = bbox_transform(rois[:, 1:], gt_boxes[gt_assignment[keep_indexes], :4])
-        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS))
-                       / np.array(config.TRAIN.BBOX_STDS))
-        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
-
-    bbox_targets, bbox_weights = \
-        expand_bbox_regression_targets(bbox_target_data, num_classes)
-
-    # Assign to levels
-    rois_on_levels = dict()
-    labels_on_levels = dict()
-    bbox_targets_on_levels = dict()
-    bbox_weights_on_levels = dict()
-    if mask_targets is not None:
-        mask_targets_on_levels = dict()
-        mask_weights_on_levels = dict()
-    for i, s in enumerate(config.RCNN_FEAT_STRIDE):
-        index = np.where(assign_levels == s)
-        _rois         = rois[index]
-        _labels       = labels[index]
-        _bbox_targets = bbox_targets[index]
-        _bbox_weights = bbox_weights[index]
-        if mask_targets is not None:
-            _mask_targets = mask_targets[index]
-            _mask_weights = mask_weights[index]
-
-        rois_on_levels.update({'stride%s' % s: _rois})
-        labels_on_levels.update({'stride%s' % s: _labels})
-        bbox_targets_on_levels.update({'stride%s' % s: _bbox_targets})
-        bbox_weights_on_levels.update({'stride%s' % s: _bbox_weights})
-        if mask_targets is not None:
-            mask_targets_on_levels.update({'stride%s' % s: _mask_targets})
-            mask_weights_on_levels.update({'stride%s' % s: _mask_weights})
-
-    if mask_targets is not None:
-        return rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels,mask_targets_on_levels,mask_weights_on_levels
-    else:
-        return rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels
-
-def get_rois(rois, rois_per_image, num_classes,
-             labels=None, overlaps=None, bbox_targets=None, gt_boxes=None):
-    """
-    get top N ROIs, used in online hard example mining
-    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
-    :param rois_per_image: total roi number
-    :param num_classes: number of classes
-    :param labels: maybe precomputed
-    :param overlaps: maybe precomputed (max_overlaps)
-    :param bbox_targets: maybe precomputed
-    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
-    :return: (rois, labels, bbox_targets, bbox_weights)
-    """
-    if labels is None:
-        if len(gt_boxes) == 0:
-            gt_boxes = np.array([[1, 1, 1, 1, 0]])
-        overlaps = bbox_overlaps(rois[:, 1:].astype(np.float), gt_boxes[:, :4].astype(np.float))
-        gt_assignment = overlaps.argmax(axis=1)
-        overlaps = overlaps.max(axis=1)
-        labels = gt_boxes[gt_assignment, 4]
-
-    # select indices
-    keep_indexes = np.arange(rois.shape[0])
-    if keep_indexes.shape[0] > rois_per_image:
-        keep_indexes = npr.choice(keep_indexes, size=rois_per_image, replace=False)
-
-    # if not enough, pad until rois_per_image is satisfied
-    while keep_indexes.shape[0] < rois_per_image:
-        gap = np.minimum(rois_per_image - keep_indexes.shape[0], len(rois))
-        gap_indexes = npr.choice(range(len(rois)), size=gap, replace=False)
-        keep_indexes = np.append(keep_indexes, gap_indexes)
-
-    # suppress any bg defined by overlap
-    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI) & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
-    labels[bg_indexes] = 0
-
-    labels = labels[keep_indexes]
-    rois = rois[keep_indexes]
-
-    # load or compute bbox_target
-    if bbox_targets is not None:
-        bbox_target_data = bbox_targets[keep_indexes, :]
-    else:
-        targets = bbox_transform(rois[:, 1:], gt_boxes[gt_assignment[keep_indexes], :4])
-        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS))
-                       / np.array(config.TRAIN.BBOX_STDS))
-        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
-
-    bbox_targets, bbox_weights = \
-        expand_bbox_regression_targets(bbox_target_data, num_classes)
-
-    return rois, labels, bbox_targets, bbox_weights
-
diff --git a/RetinaFace/rcnn/io/rpn.py b/RetinaFace/rcnn/io/rpn.py
deleted file mode 100644
index cd56941d3..000000000
--- a/RetinaFace/rcnn/io/rpn.py
+++ /dev/null
@@ -1,765 +0,0 @@
-"""
-RPN:
-data =
-    {'data': [num_images, c, h, w],
-     'im_info': [num_images, 4] (optional)}
-label =
-    {'gt_boxes': [num_boxes, 5] (optional),
-     'label': [batch_size, 1] <- [batch_size, num_anchors, feat_height, feat_width],
-     'bbox_target': [batch_size, num_anchors, feat_height, feat_width],
-     'bbox_weight': [batch_size, num_anchors, feat_height, feat_width]}
-"""
-
-from __future__ import print_function
-import sys
-import logging
-import datetime
-import numpy as np
-import numpy.random as npr
-
-from ..logger import logger
-from ..config import config
-from .image import get_image, tensor_vstack, get_crop_image
-from ..processing.generate_anchor import generate_anchors, anchors_plane
-from ..processing.bbox_transform import bbox_overlaps, bbox_transform, landmark_transform
-
-STAT = {0:0, 8:0, 16:0, 32:0}
-
-def get_rpn_testbatch(roidb):
-    """
-    return a dict of testbatch
-    :param roidb: ['image', 'flipped']
-    :return: data, label, im_info
-    """
-    assert len(roidb) == 1, 'Single batch only'
-    imgs, roidb = get_image(roidb)
-    im_array = imgs[0]
-    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
-
-    data = {'data': im_array,
-            'im_info': im_info}
-    label = {}
-
-    return data, label, im_info
-
-def get_rpn_batch(roidb):
-    """
-    prototype for rpn batch: data, im_info, gt_boxes
-    :param roidb: ['image', 'flipped'] + ['gt_boxes', 'boxes', 'gt_classes']
-    :return: data, label
-    """
-    assert len(roidb) == 1, 'Single batch only'
-    imgs, roidb = get_image(roidb)
-    im_array = imgs[0]
-    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
-
-    # gt boxes: (x1, y1, x2, y2, cls)
-    if roidb[0]['gt_classes'].size > 0:
-        gt_inds = np.where(roidb[0]['gt_classes'] != 0)[0]
-        gt_boxes = np.empty((roidb[0]['boxes'].shape[0], 5), dtype=np.float32)
-        gt_boxes[:, 0:4] = roidb[0]['boxes'][gt_inds, :]
-        gt_boxes[:, 4] = roidb[0]['gt_classes'][gt_inds]
-    else:
-        gt_boxes = np.empty((0, 5), dtype=np.float32)
-
-    data = {'data': im_array,
-            'im_info': im_info}
-    label = {'gt_boxes': gt_boxes}
-
-    return data, label
-
-def get_crop_batch(roidb):
-    """
-    prototype for rpn batch: data, im_info, gt_boxes
-    :param roidb: ['image', 'flipped'] + ['gt_boxes', 'boxes', 'gt_classes']
-    :return: data, label
-    """
-    #assert len(roidb) == 1, 'Single batch only'
-    data_list = []
-    label_list = []
-    imgs, roidb = get_crop_image(roidb)
-    assert len(imgs)==len(roidb)
-    for i in range(len(imgs)):
-      im_array = imgs[i]
-      im_info = np.array([roidb[i]['im_info']], dtype=np.float32)
-
-      # gt boxes: (x1, y1, x2, y2, cls)
-      if roidb[i]['gt_classes'].size > 0:
-          gt_inds = np.where(roidb[i]['gt_classes'] != 0)[0]
-          gt_boxes = np.empty((roidb[i]['boxes'].shape[0], 5), dtype=np.float32)
-          gt_boxes[:, 0:4] = roidb[i]['boxes'][gt_inds, :]
-          gt_boxes[:, 4] = roidb[i]['gt_classes'][gt_inds]
-          if config.USE_BLUR:
-            gt_blur = roidb[i]['blur']
-          if config.FACE_LANDMARK:
-            #gt_landmarks = np.empty((roidb[i]['landmarks'].shape[0], 11), dtype=np.float32)
-            gt_landmarks = roidb[i]['landmarks'][gt_inds,:,:]
-          if config.HEAD_BOX:
-            gt_boxes_head = np.empty((roidb[i]['boxes_head'].shape[0], 5), dtype=np.float32)
-            gt_boxes_head[:, 0:4] = roidb[i]['boxes_head'][gt_inds, :]
-            gt_boxes_head[:, 4] = roidb[i]['gt_classes'][gt_inds]
-      else:
-          gt_boxes = np.empty((0, 5), dtype=np.float32)
-          if config.USE_BLUR:
-            gt_blur = np.empty((0,), dtype=np.float32)
-          if config.FACE_LANDMARK:
-            gt_landmarks = np.empty((0, 5, 3), dtype=np.float32)
-          if config.HEAD_BOX:
-            gt_boxes_head = np.empty((0, 5), dtype=np.float32)
-
-      data = {'data': im_array,
-              'im_info': im_info}
-      label = {'gt_boxes': gt_boxes}
-      if config.USE_BLUR:
-        label['gt_blur'] = gt_blur
-      if config.FACE_LANDMARK:
-        label['gt_landmarks'] = gt_landmarks
-      if config.HEAD_BOX:
-        label['gt_boxes_head'] = gt_boxes_head
-      data_list.append(data)
-      label_list.append(label)
-
-    return data_list, label_list
-
-def assign_anchor_fpn(feat_shape, gt_label, im_info, landmark=False, prefix='face', select_stride=0):
-    """
-    assign ground truth boxes to anchor positions
-    :param feat_shape: infer output shape
-    :param gt_boxes: assign ground truth
-    :param im_info: filter out anchors overlapped with edges
-    :return: tuple
-    labels: of shape (batch_size, 1) <- (batch_size, num_anchors, feat_height, feat_width)
-    bbox_targets: of shape (batch_size, num_anchors * 4, feat_height, feat_width)
-    bbox_weights: mark the assigned anchors
-    """
-    def _unmap(data, count, inds, fill=0):
-        """" unmap a subset inds of data into original data of size count """
-        if len(data.shape) == 1:
-            ret = np.empty((count,), dtype=np.float32)
-            ret.fill(fill)
-            ret[inds] = data
-        else:
-            ret = np.empty((count,) + data.shape[1:], dtype=np.float32)
-            ret.fill(fill)
-            ret[inds, :] = data
-        return ret
-
-    global STAT
-    DEBUG = False
-
-    im_info = im_info[0]
-    gt_boxes = gt_label['gt_boxes']
-    # clean up boxes
-    nonneg = np.where(gt_boxes[:, 4] != -1)[0]
-    gt_boxes = gt_boxes[nonneg]
-    if config.USE_BLUR:
-      gt_blur = gt_label['gt_blur']
-      gt_blur = gt_blur[nonneg]
-    if landmark:
-      gt_landmarks = gt_label['gt_landmarks']
-      gt_landmarks = gt_landmarks[nonneg]
-      assert gt_boxes.shape[0]==gt_landmarks.shape[0]
-    #scales = np.array(scales, dtype=np.float32)
-    feat_strides = config.RPN_FEAT_STRIDE
-    bbox_pred_len = 4
-    landmark_pred_len = 10
-    if config.USE_BLUR:
-      gt_boxes[:,4] = gt_blur
-      bbox_pred_len = 5
-    if config.USE_OCCLUSION:
-      landmark_pred_len = 15
-
-    anchors_list = []
-    anchors_num_list = []
-    inds_inside_list = []
-    feat_infos = []
-    A_list = []
-    for i in range(len(feat_strides)):
-        stride = feat_strides[i]
-        sstride = str(stride)
-        base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
-        allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
-        ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
-        scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
-        base_anchors = generate_anchors(base_size=base_size, ratios=list(ratios), scales=np.array(scales, dtype=np.float32), stride = stride, dense_anchor = config.DENSE_ANCHOR)
-        num_anchors = base_anchors.shape[0]
-        feat_height, feat_width = feat_shape[i][-2:]
-        feat_stride = feat_strides[i]
-        feat_infos.append([feat_height, feat_width])
-
-        A = num_anchors
-        A_list.append(A)
-        K = feat_height * feat_width
-
-        all_anchors = anchors_plane(feat_height, feat_width, feat_stride, base_anchors)
-        all_anchors = all_anchors.reshape((K * A, 4))
-        #print('anchor0', stride, all_anchors[0])
-
-        total_anchors = int(K * A)
-        anchors_num_list.append(total_anchors)
-        # only keep anchors inside the image
-        inds_inside = np.where((all_anchors[:, 0] >= -allowed_border) &
-                               (all_anchors[:, 1] >= -allowed_border) &
-                               (all_anchors[:, 2] < im_info[1] + allowed_border) &
-                               (all_anchors[:, 3] < im_info[0] + allowed_border))[0]
-        if DEBUG:
-            print('total_anchors', total_anchors)
-            print('inds_inside', len(inds_inside))
-
-        # keep only inside anchors
-        anchors = all_anchors[inds_inside, :]
-        #print('AA', anchors.shape, len(inds_inside))
-
-        anchors_list.append(anchors)
-        inds_inside_list.append(inds_inside)
-
-    # Concat anchors from each level
-    anchors = np.concatenate(anchors_list)
-    for i in range(1, len(inds_inside_list)):
-        inds_inside_list[i] = inds_inside_list[i] + sum(anchors_num_list[:i])
-    inds_inside = np.concatenate(inds_inside_list)
-    total_anchors = sum(anchors_num_list)
-    #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
-
-    # label: 1 is positive, 0 is negative, -1 is dont care
-    labels = np.empty((len(inds_inside),), dtype=np.float32)
-    labels.fill(-1)
-    #print('BB', anchors.shape, len(inds_inside))
-    #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
-
-    if gt_boxes.size > 0:
-        # overlap between the anchors and the gt boxes
-        # overlaps (ex, gt)
-        overlaps = bbox_overlaps(anchors.astype(np.float), gt_boxes.astype(np.float))
-        argmax_overlaps = overlaps.argmax(axis=1)
-        #print('AAA', argmax_overlaps.shape)
-        max_overlaps = overlaps[np.arange(len(inds_inside)), argmax_overlaps]
-        gt_argmax_overlaps = overlaps.argmax(axis=0)
-        gt_max_overlaps = overlaps[gt_argmax_overlaps, np.arange(overlaps.shape[1])]
-        gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
-
-        if not config.TRAIN.RPN_CLOBBER_POSITIVES:
-            # assign bg labels first so that positive labels can clobber them
-            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
-
-        # fg label: for each gt, anchor with highest overlap
-        if config.TRAIN.RPN_FORCE_POSITIVE:
-          labels[gt_argmax_overlaps] = 1
-
-        # fg label: above threshold IoU
-        labels[max_overlaps >= config.TRAIN.RPN_POSITIVE_OVERLAP] = 1
-
-        if config.TRAIN.RPN_CLOBBER_POSITIVES:
-            # assign bg labels last so that negative labels can clobber positives
-            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
-    else:
-        labels[:] = 0
-    fg_inds = np.where(labels == 1)[0]
-    #print('fg count', len(fg_inds))
-
-    # subsample positive labels if we have too many
-    if config.TRAIN.RPN_ENABLE_OHEM==0:
-      fg_inds = np.where(labels == 1)[0]
-      num_fg = int(config.TRAIN.RPN_FG_FRACTION * config.TRAIN.RPN_BATCH_SIZE)
-      if len(fg_inds) > num_fg:
-          disable_inds = npr.choice(fg_inds, size=(len(fg_inds) - num_fg), replace=False)
-          if DEBUG:
-              disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
-          labels[disable_inds] = -1
-
-      # subsample negative labels if we have too many
-      num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
-      bg_inds = np.where(labels == 0)[0]
-      if len(bg_inds) > num_bg:
-          disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-          if DEBUG:
-              disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-          labels[disable_inds] = -1
-
-      #fg_inds = np.where(labels == 1)[0]
-      #num_fg = len(fg_inds)
-      #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
-
-      #bg_inds = np.where(labels == 0)[0]
-      #if len(bg_inds) > num_bg:
-      #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-      #    if DEBUG:
-      #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-      #    labels[disable_inds] = -1
-    else:
-      fg_inds = np.where(labels == 1)[0]
-      num_fg = len(fg_inds)
-      bg_inds = np.where(labels == 0)[0]
-      num_bg = len(bg_inds)
-
-    #print('anchor stat', num_fg, num_bg)
-
-
-    bbox_targets = np.zeros((len(inds_inside), bbox_pred_len), dtype=np.float32)
-    if gt_boxes.size > 0:
-        #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
-        bbox_targets[:,:] = bbox_transform(anchors, gt_boxes[argmax_overlaps, :])
-        #bbox_targets[:,4] = gt_blur
-
-    bbox_weights = np.zeros((len(inds_inside), bbox_pred_len), dtype=np.float32)
-    #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
-    bbox_weights[labels == 1, 0:4] = 1.0
-    if bbox_pred_len>4:
-      bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
-
-    if landmark:
-      landmark_targets = np.zeros((len(inds_inside), landmark_pred_len), dtype=np.float32)
-      #landmark_weights = np.zeros((len(inds_inside), 10), dtype=np.float32)
-      landmark_weights = np.zeros((len(inds_inside), landmark_pred_len), dtype=np.float32)
-      #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
-      if landmark_pred_len==10:
-        landmark_weights[labels == 1, :] = 1.0
-      elif landmark_pred_len==15:
-        v = [1.0, 1.0, 0.1] * 5
-        assert len(v)==15
-        landmark_weights[labels == 1, :] = np.array(v)
-      else:
-        assert False
-      #TODO here
-      if gt_landmarks.size > 0:
-        #print('AAA',argmax_overlaps)
-        a_landmarks = gt_landmarks[argmax_overlaps,:,:]
-        landmark_targets[:] = landmark_transform(anchors, a_landmarks)
-        invalid = np.where(a_landmarks[:,0,2]<0.0)[0]
-        #assert len(invalid)==0
-        #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
-        landmark_weights[invalid, :] = 0.0
-
-    #if DEBUG:
-    #    _sums = bbox_targets[labels == 1, :].sum(axis=0)
-    #    _squared_sums = (bbox_targets[labels == 1, :] ** 2).sum(axis=0)
-    #    _counts = np.sum(labels == 1)
-    #    means = _sums / (_counts + 1e-14)
-    #    stds = np.sqrt(_squared_sums / _counts - means ** 2)
-    #    print 'means', means
-    #    print 'stdevs', stds
-    # map up to original set of anchors
-    #print(labels.shape, total_anchors, inds_inside.shape, inds_inside[0], inds_inside[-1])
-    labels = _unmap(labels, total_anchors, inds_inside, fill=-1)
-    bbox_targets = _unmap(bbox_targets, total_anchors, inds_inside, fill=0)
-    bbox_weights = _unmap(bbox_weights, total_anchors, inds_inside, fill=0)
-    if landmark:
-      landmark_targets = _unmap(landmark_targets, total_anchors, inds_inside, fill=0)
-      landmark_weights = _unmap(landmark_weights, total_anchors, inds_inside, fill=0)
-    #print('CC', anchors.shape, len(inds_inside))
-
-    #if DEBUG:
-    #    if gt_boxes.size > 0:
-    #        print 'rpn: max max_overlaps', np.max(max_overlaps)
-    #    print 'rpn: num_positives', np.sum(labels == 1)
-    #    print 'rpn: num_negatives', np.sum(labels == 0)
-    #    _fg_sum = np.sum(labels == 1)
-    #    _bg_sum = np.sum(labels == 0)
-    #    _count = 1
-    #    print 'rpn: num_positive avg', _fg_sum / _count
-    #    print 'rpn: num_negative avg', _bg_sum / _count
-
-    # resahpe
-    label_list = list()
-    bbox_target_list = list()
-    bbox_weight_list = list()
-    if landmark:
-      landmark_target_list = list()
-      landmark_weight_list = list()
-    anchors_num_range = [0] + anchors_num_list
-    label = {}
-    for i in range(len(feat_strides)):
-        stride = feat_strides[i]
-        feat_height, feat_width = feat_infos[i]
-        A = A_list[i]
-        _label = labels[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        if select_stride>0 and stride!=select_stride:
-          #print('set', stride, select_stride)
-          _label[:] = -1
-        #print('_label', _label.shape, select_stride)
-        #_fg_inds = np.where(_label == 1)[0]
-        #n_fg = len(_fg_inds)
-        #STAT[0]+=1
-        #STAT[stride]+=n_fg
-        #if STAT[0]%100==0:
-        #  print('rpn_stat', STAT, file=sys.stderr)
-        bbox_target = bbox_targets[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        bbox_weight = bbox_weights[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        if landmark:
-          landmark_target = landmark_targets[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-          landmark_weight = landmark_weights[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-
-        _label = _label.reshape((1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
-        _label = _label.reshape((1, A * feat_height * feat_width))
-        bbox_target = bbox_target.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose(0, 2, 1)
-        bbox_weight = bbox_weight.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose((0, 2, 1))
-        label['%s_label_stride%d'%(prefix, stride)] = _label
-        label['%s_bbox_target_stride%d'%(prefix,stride)] = bbox_target
-        label['%s_bbox_weight_stride%d'%(prefix,stride)] = bbox_weight
-        if landmark:
-          landmark_target = landmark_target.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose(0, 2, 1)
-          landmark_weight = landmark_weight.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose((0, 2, 1))
-          label['%s_landmark_target_stride%d'%(prefix,stride)] = landmark_target
-          label['%s_landmark_weight_stride%d'%(prefix,stride)] = landmark_weight
-        #print('in_rpn', stride,_label.shape, bbox_target.shape, bbox_weight.shape, file=sys.stderr)
-        label_list.append(_label)
-        #print('DD', _label.shape)
-        bbox_target_list.append(bbox_target)
-        bbox_weight_list.append(bbox_weight)
-        if landmark:
-          landmark_target_list.append(landmark_target)
-          landmark_weight_list.append(landmark_weight)
-
-    label_concat = np.concatenate(label_list, axis=1)
-    bbox_target_concat = np.concatenate(bbox_target_list, axis=2)
-    bbox_weight_concat = np.concatenate(bbox_weight_list, axis=2)
-    #fg_inds = np.where(label_concat[0] == 1)[0]
-    #print('fg_inds_in_rpn2', fg_inds, file=sys.stderr)
-
-    label.update({'%s_label'%prefix: label_concat,
-            '%s_bbox_target'%prefix: bbox_target_concat,
-            '%s_bbox_weight'%prefix: bbox_weight_concat}
-            )
-    if landmark:
-      landmark_target_concat = np.concatenate(landmark_target_list, axis=2)
-      landmark_weight_concat = np.concatenate(landmark_weight_list, axis=2)
-      label['%s_landmark_target'%prefix] = landmark_target_concat
-      label['%s_landmark_weight'%prefix] = landmark_weight_concat
-    return label
-
-
-class AA:
-  def __init__(self, feat_shape):
-    self.feat_shape = feat_shape
-    feat_strides = config.RPN_FEAT_STRIDE
-    anchors_list = []
-    anchors_num_list = []
-    inds_inside_list = []
-    feat_infos = []
-    A_list = []
-    DEBUG = False
-    for i in range(len(feat_strides)):
-        stride = feat_strides[i]
-        sstride = str(stride)
-        base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
-        allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
-        ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
-        scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
-        base_anchors = generate_anchors(base_size=base_size, ratios=list(ratios), scales=np.array(scales, dtype=np.float32), stride = stride, dense_anchor = config.DENSE_ANCHOR)
-        num_anchors = base_anchors.shape[0]
-        feat_height, feat_width = feat_shape[i][-2:]
-        feat_stride = feat_strides[i]
-        feat_infos.append([feat_height, feat_width])
-
-        A = num_anchors
-        A_list.append(A)
-        K = feat_height * feat_width
-
-        all_anchors = anchors_plane(feat_height, feat_width, feat_stride, base_anchors)
-        all_anchors = all_anchors.reshape((K * A, 4))
-        #print('anchor0', stride, all_anchors[0])
-
-        total_anchors = int(K * A)
-        anchors_num_list.append(total_anchors)
-        # only keep anchors inside the image
-        inds_inside = np.where((all_anchors[:, 0] >= -allowed_border) &
-                               (all_anchors[:, 1] >= -allowed_border) &
-                               (all_anchors[:, 2] < config.SCALES[0][1] + allowed_border) &
-                               (all_anchors[:, 3] < config.SCALES[0][1] + allowed_border))[0]
-        if DEBUG:
-            print('total_anchors', total_anchors)
-            print('inds_inside', len(inds_inside))
-
-        # keep only inside anchors
-        anchors = all_anchors[inds_inside, :]
-        #print('AA', anchors.shape, len(inds_inside))
-
-        anchors_list.append(anchors)
-        inds_inside_list.append(inds_inside)
-    anchors = np.concatenate(anchors_list)
-    for i in range(1, len(inds_inside_list)):
-        inds_inside_list[i] = inds_inside_list[i] + sum(anchors_num_list[:i])
-    inds_inside = np.concatenate(inds_inside_list)
-    #self.anchors_list = anchors_list
-    #self.inds_inside_list = inds_inside_list
-    self.anchors = anchors
-    self.inds_inside = inds_inside
-    self.anchors_num_list = anchors_num_list
-    self.feat_infos = feat_infos
-    self.A_list = A_list
-    self._times = [0.0, 0.0, 0.0, 0.0]
-
-  @staticmethod
-  def _unmap(data, count, inds, fill=0):
-      """" unmap a subset inds of data into original data of size count """
-      if len(data.shape) == 1:
-          ret = np.empty((count,), dtype=np.float32)
-          ret.fill(fill)
-          ret[inds] = data
-      else:
-          ret = np.empty((count,) + data.shape[1:], dtype=np.float32)
-          ret.fill(fill)
-          ret[inds, :] = data
-      return ret
-
-  def assign_anchor_fpn(self, gt_label, im_info, landmark=False, prefix='face', select_stride=0):
-
-    #ta = datetime.datetime.now()
-    im_info = im_info[0]
-    gt_boxes = gt_label['gt_boxes']
-    # clean up boxes
-    nonneg = np.where(gt_boxes[:, 4] != -1)[0]
-    gt_boxes = gt_boxes[nonneg]
-    if config.USE_BLUR:
-      gt_blur = gt_label['gt_blur']
-      gt_blur = gt_blur[nonneg]
-    if landmark:
-      gt_landmarks = gt_label['gt_landmarks']
-      gt_landmarks = gt_landmarks[nonneg]
-      assert gt_boxes.shape[0]==gt_landmarks.shape[0]
-    #scales = np.array(scales, dtype=np.float32)
-    feat_strides = config.RPN_FEAT_STRIDE
-    bbox_pred_len = 4
-    landmark_pred_len = 10
-    if config.USE_BLUR:
-      gt_boxes[:,4] = gt_blur
-      bbox_pred_len = 5
-    if config.USE_OCCLUSION:
-      landmark_pred_len = 15
-
-    #anchors_list = self.anchors_list
-    #inds_inside_list = self.inds_inside_list
-    anchors = self.anchors
-    inds_inside = self.inds_inside
-    anchors_num_list = self.anchors_num_list
-    feat_infos = self.feat_infos
-    A_list = self.A_list
-
-    total_anchors = sum(anchors_num_list)
-    #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
-
-    # label: 1 is positive, 0 is negative, -1 is dont care
-    labels = np.empty((len(inds_inside),), dtype=np.float32)
-    labels.fill(-1)
-    #print('BB', anchors.shape, len(inds_inside))
-    #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
-    #tb = datetime.datetime.now()
-    #self._times[0] += (tb-ta).total_seconds()
-    #ta = datetime.datetime.now()
-
-    if gt_boxes.size > 0:
-        # overlap between the anchors and the gt boxes
-        # overlaps (ex, gt)
-        overlaps = bbox_overlaps(anchors.astype(np.float), gt_boxes.astype(np.float))
-        argmax_overlaps = overlaps.argmax(axis=1)
-        #print('AAA', argmax_overlaps.shape)
-        max_overlaps = overlaps[np.arange(len(inds_inside)), argmax_overlaps]
-        gt_argmax_overlaps = overlaps.argmax(axis=0)
-        gt_max_overlaps = overlaps[gt_argmax_overlaps, np.arange(overlaps.shape[1])]
-        gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
-
-        if not config.TRAIN.RPN_CLOBBER_POSITIVES:
-            # assign bg labels first so that positive labels can clobber them
-            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
-
-        # fg label: for each gt, anchor with highest overlap
-        if config.TRAIN.RPN_FORCE_POSITIVE:
-          labels[gt_argmax_overlaps] = 1
-
-        # fg label: above threshold IoU
-        labels[max_overlaps >= config.TRAIN.RPN_POSITIVE_OVERLAP] = 1
-
-        if config.TRAIN.RPN_CLOBBER_POSITIVES:
-            # assign bg labels last so that negative labels can clobber positives
-            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
-    else:
-        labels[:] = 0
-    fg_inds = np.where(labels == 1)[0]
-    #print('fg count', len(fg_inds))
-
-    # subsample positive labels if we have too many
-    if config.TRAIN.RPN_ENABLE_OHEM==0:
-      fg_inds = np.where(labels == 1)[0]
-      num_fg = int(config.TRAIN.RPN_FG_FRACTION * config.TRAIN.RPN_BATCH_SIZE)
-      if len(fg_inds) > num_fg:
-          disable_inds = npr.choice(fg_inds, size=(len(fg_inds) - num_fg), replace=False)
-          if DEBUG:
-              disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
-          labels[disable_inds] = -1
-
-      # subsample negative labels if we have too many
-      num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
-      bg_inds = np.where(labels == 0)[0]
-      if len(bg_inds) > num_bg:
-          disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-          if DEBUG:
-              disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-          labels[disable_inds] = -1
-
-      #fg_inds = np.where(labels == 1)[0]
-      #num_fg = len(fg_inds)
-      #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
-
-      #bg_inds = np.where(labels == 0)[0]
-      #if len(bg_inds) > num_bg:
-      #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
-      #    if DEBUG:
-      #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
-      #    labels[disable_inds] = -1
-    else:
-      fg_inds = np.where(labels == 1)[0]
-      num_fg = len(fg_inds)
-      bg_inds = np.where(labels == 0)[0]
-      num_bg = len(bg_inds)
-
-    #print('anchor stat', num_fg, num_bg)
-
-
-    bbox_targets = np.zeros((len(inds_inside), bbox_pred_len), dtype=np.float32)
-    if gt_boxes.size > 0:
-        #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
-        bbox_targets[:,:] = bbox_transform(anchors, gt_boxes[argmax_overlaps, :])
-        #bbox_targets[:,4] = gt_blur
-    #tb = datetime.datetime.now()
-    #self._times[1] += (tb-ta).total_seconds()
-    #ta = datetime.datetime.now()
-
-    bbox_weights = np.zeros((len(inds_inside), bbox_pred_len), dtype=np.float32)
-    #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
-    bbox_weights[labels == 1, 0:4] = 1.0
-    if bbox_pred_len>4:
-      bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
-
-    if landmark:
-      landmark_targets = np.zeros((len(inds_inside), landmark_pred_len), dtype=np.float32)
-      #landmark_weights = np.zeros((len(inds_inside), 10), dtype=np.float32)
-      landmark_weights = np.zeros((len(inds_inside), landmark_pred_len), dtype=np.float32)
-      #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
-      if landmark_pred_len==10:
-        landmark_weights[labels == 1, :] = 1.0
-      elif landmark_pred_len==15:
-        v = [1.0, 1.0, 0.1] * 5
-        assert len(v)==15
-        landmark_weights[labels == 1, :] = np.array(v)
-      else:
-        assert False
-      #TODO here
-      if gt_landmarks.size > 0:
-        #print('AAA',argmax_overlaps)
-        a_landmarks = gt_landmarks[argmax_overlaps,:,:]
-        landmark_targets[:] = landmark_transform(anchors, a_landmarks)
-        invalid = np.where(a_landmarks[:,0,2]<0.0)[0]
-        #assert len(invalid)==0
-        #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
-        landmark_weights[invalid, :] = 0.0
-    #tb = datetime.datetime.now()
-    #self._times[2] += (tb-ta).total_seconds()
-    #ta = datetime.datetime.now()
-
-    #if DEBUG:
-    #    _sums = bbox_targets[labels == 1, :].sum(axis=0)
-    #    _squared_sums = (bbox_targets[labels == 1, :] ** 2).sum(axis=0)
-    #    _counts = np.sum(labels == 1)
-    #    means = _sums / (_counts + 1e-14)
-    #    stds = np.sqrt(_squared_sums / _counts - means ** 2)
-    #    print 'means', means
-    #    print 'stdevs', stds
-    # map up to original set of anchors
-    #print(labels.shape, total_anchors, inds_inside.shape, inds_inside[0], inds_inside[-1])
-    labels = AA._unmap(labels, total_anchors, inds_inside, fill=-1)
-    bbox_targets = AA._unmap(bbox_targets, total_anchors, inds_inside, fill=0)
-    bbox_weights = AA._unmap(bbox_weights, total_anchors, inds_inside, fill=0)
-    if landmark:
-      landmark_targets = AA._unmap(landmark_targets, total_anchors, inds_inside, fill=0)
-      landmark_weights = AA._unmap(landmark_weights, total_anchors, inds_inside, fill=0)
-    #print('CC', anchors.shape, len(inds_inside))
-
-    bbox_targets[:, 0::4] = bbox_targets[:,0::4] / config.TRAIN.BBOX_STDS[0]
-    bbox_targets[:, 1::4] = bbox_targets[:,1::4] / config.TRAIN.BBOX_STDS[1]
-    bbox_targets[:, 2::4] = bbox_targets[:,2::4] / config.TRAIN.BBOX_STDS[2]
-    bbox_targets[:, 3::4] = bbox_targets[:,3::4] / config.TRAIN.BBOX_STDS[3]
-    landmark_targets /= config.TRAIN.LANDMARK_STD
-    #print('applied STD')
-
-    #if DEBUG:
-    #    if gt_boxes.size > 0:
-    #        print 'rpn: max max_overlaps', np.max(max_overlaps)
-    #    print 'rpn: num_positives', np.sum(labels == 1)
-    #    print 'rpn: num_negatives', np.sum(labels == 0)
-    #    _fg_sum = np.sum(labels == 1)
-    #    _bg_sum = np.sum(labels == 0)
-    #    _count = 1
-    #    print 'rpn: num_positive avg', _fg_sum / _count
-    #    print 'rpn: num_negative avg', _bg_sum / _count
-
-    # resahpe
-    label_list = list()
-    bbox_target_list = list()
-    bbox_weight_list = list()
-    if landmark:
-      landmark_target_list = list()
-      landmark_weight_list = list()
-    anchors_num_range = [0] + anchors_num_list
-    label = {}
-    for i in range(len(feat_strides)):
-        stride = feat_strides[i]
-        feat_height, feat_width = feat_infos[i]
-        A = A_list[i]
-        _label = labels[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        if select_stride>0 and stride!=select_stride:
-          #print('set', stride, select_stride)
-          _label[:] = -1
-        #print('_label', _label.shape, select_stride)
-        #_fg_inds = np.where(_label == 1)[0]
-        #n_fg = len(_fg_inds)
-        #STAT[0]+=1
-        #STAT[stride]+=n_fg
-        #if STAT[0]%100==0:
-        #  print('rpn_stat', STAT, file=sys.stderr)
-        bbox_target = bbox_targets[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        bbox_weight = bbox_weights[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-        if landmark:
-          landmark_target = landmark_targets[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-          landmark_weight = landmark_weights[sum(anchors_num_range[:i+1]):sum(anchors_num_range[:i+1])+anchors_num_range[i+1]]
-
-        _label = _label.reshape((1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
-        _label = _label.reshape((1, A * feat_height * feat_width))
-        bbox_target = bbox_target.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose(0, 2, 1)
-        bbox_weight = bbox_weight.reshape((1, feat_height*feat_width, A * bbox_pred_len)).transpose((0, 2, 1))
-        label['%s_label_stride%d'%(prefix, stride)] = _label
-        label['%s_bbox_target_stride%d'%(prefix,stride)] = bbox_target
-        label['%s_bbox_weight_stride%d'%(prefix,stride)] = bbox_weight
-        if landmark:
-          landmark_target = landmark_target.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose(0, 2, 1)
-          landmark_weight = landmark_weight.reshape((1, feat_height*feat_width, A * landmark_pred_len)).transpose((0, 2, 1))
-          label['%s_landmark_target_stride%d'%(prefix,stride)] = landmark_target
-          label['%s_landmark_weight_stride%d'%(prefix,stride)] = landmark_weight
-        #print('in_rpn', stride,_label.shape, bbox_target.shape, bbox_weight.shape, file=sys.stderr)
-        label_list.append(_label)
-        #print('DD', _label.shape)
-        bbox_target_list.append(bbox_target)
-        bbox_weight_list.append(bbox_weight)
-        if landmark:
-          landmark_target_list.append(landmark_target)
-          landmark_weight_list.append(landmark_weight)
-
-    label_concat = np.concatenate(label_list, axis=1)
-    bbox_target_concat = np.concatenate(bbox_target_list, axis=2)
-    bbox_weight_concat = np.concatenate(bbox_weight_list, axis=2)
-    #fg_inds = np.where(label_concat[0] == 1)[0]
-    #print('fg_inds_in_rpn2', fg_inds, file=sys.stderr)
-
-    label.update({'%s_label'%prefix: label_concat,
-            '%s_bbox_target'%prefix: bbox_target_concat,
-            '%s_bbox_weight'%prefix: bbox_weight_concat}
-            )
-    if landmark:
-      landmark_target_concat = np.concatenate(landmark_target_list, axis=2)
-      landmark_weight_concat = np.concatenate(landmark_weight_list, axis=2)
-      label['%s_landmark_target'%prefix] = landmark_target_concat
-      label['%s_landmark_weight'%prefix] = landmark_weight_concat
-    #tb = datetime.datetime.now()
-    #self._times[3] += (tb-ta).total_seconds()
-    #ta = datetime.datetime.now()
-    #print(self._times)
-    return label
diff --git a/RetinaFace/rcnn/processing/assign_levels.py b/RetinaFace/rcnn/processing/assign_levels.py
deleted file mode 100755
index b237439fc..000000000
--- a/RetinaFace/rcnn/processing/assign_levels.py
+++ /dev/null
@@ -1,37 +0,0 @@
-from rcnn.config import config
-import numpy as np
-
-
-def compute_assign_targets(rois, threshold):
-    rois_area = np.sqrt((rois[:, 2] - rois[:, 0] + 1) * (rois[:, 3] - rois[:, 1] + 1))
-    num_rois = np.shape(rois)[0]
-    assign_levels = np.zeros(num_rois, dtype=np.uint8)
-    for i, stride in enumerate(config.RCNN_FEAT_STRIDE):
-        thd = threshold[i]
-        idx = np.logical_and(thd[1] <= rois_area, rois_area < thd[0])
-        assign_levels[idx] = stride
-
-    assert 0 not in assign_levels, "All rois should assign to specify levels."
-    return assign_levels
-
-
-def add_assign_targets(roidb):
-    """
-    given roidb, add ['assign_level']
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    """
-    print 'add assign targets'
-    assert len(roidb) > 0
-    assert 'boxes' in roidb[0]
-
-    area_threshold = [[np.inf, 448],
-                      [448,    224],
-                      [224,    112],
-                      [112,     0]]
-
-    assert len(config.RCNN_FEAT_STRIDE) == len(area_threshold)
-
-    num_images = len(roidb)
-    for im_i in range(num_images):
-        rois = roidb[im_i]['boxes']
-        roidb[im_i]['assign_levels'] = compute_assign_targets(rois, area_threshold)
diff --git a/RetinaFace/rcnn/processing/bbox_regression.py b/RetinaFace/rcnn/processing/bbox_regression.py
deleted file mode 100644
index 11d1a0236..000000000
--- a/RetinaFace/rcnn/processing/bbox_regression.py
+++ /dev/null
@@ -1,255 +0,0 @@
-"""
-This file has functions about generating bounding box regression targets
-"""
-
-from ..pycocotools.mask import encode
-import numpy as np
-
-from ..logger import logger
-from .bbox_transform import bbox_overlaps, bbox_transform
-from rcnn.config import config
-import math
-import cv2
-import PIL.Image as Image
-import threading
-import Queue
-
-
-def compute_bbox_regression_targets(rois, overlaps, labels):
-    """
-    given rois, overlaps, gt labels, compute bounding box regression targets
-    :param rois: roidb[i]['boxes'] k * 4
-    :param overlaps: roidb[i]['max_overlaps'] k * 1
-    :param labels: roidb[i]['max_classes'] k * 1
-    :return: targets[i][class, dx, dy, dw, dh] k * 5
-    """
-    # Ensure ROIs are floats
-    rois = rois.astype(np.float, copy=False)
-
-    # Sanity check
-    if len(rois) != len(overlaps):
-        logger.warning('bbox regression: len(rois) != len(overlaps)')
-
-    # Indices of ground-truth ROIs
-    gt_inds = np.where(overlaps == 1)[0]
-    if len(gt_inds) == 0:
-        logger.warning('bbox regression: len(gt_inds) == 0')
-
-    # Indices of examples for which we try to make predictions
-    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
-
-    # Get IoU overlap between each ex ROI and gt ROI
-    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
-
-    # Find which gt ROI each ex ROI has max overlap with:
-    # this will be the ex ROI's gt target
-    gt_assignment = ex_gt_overlaps.argmax(axis=1)
-    gt_rois = rois[gt_inds[gt_assignment], :]
-    ex_rois = rois[ex_inds, :]
-
-    targets = np.zeros((rois.shape[0], 5), dtype=np.float32)
-    targets[ex_inds, 0] = labels[ex_inds]
-    targets[ex_inds, 1:] = bbox_transform(ex_rois, gt_rois)
-    return targets
-
-
-def add_bbox_regression_targets(roidb):
-    """
-    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    :return: means, std variances of targets
-    """
-    logger.info('bbox regression: add bounding box regression targets')
-    assert len(roidb) > 0
-    assert 'max_classes' in roidb[0]
-
-    num_images = len(roidb)
-    num_classes = roidb[0]['gt_overlaps'].shape[1]
-    for im_i in range(num_images):
-        rois = roidb[im_i]['boxes']
-        max_overlaps = roidb[im_i]['max_overlaps']
-        max_classes = roidb[im_i]['max_classes']
-        roidb[im_i]['bbox_targets'] = compute_bbox_regression_targets(rois, max_overlaps, max_classes)
-
-    if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-        # use fixed / precomputed means and stds instead of empirical values
-        means = np.tile(np.array(config.TRAIN.BBOX_MEANS), (num_classes, 1))
-        stds = np.tile(np.array(config.TRAIN.BBOX_STDS), (num_classes, 1))
-    else:
-        # compute mean, std values
-        class_counts = np.zeros((num_classes, 1)) + 1e-14
-        sums = np.zeros((num_classes, 4))
-        squared_sums = np.zeros((num_classes, 4))
-        for im_i in range(num_images):
-            targets = roidb[im_i]['bbox_targets']
-            for cls in range(1, num_classes):
-                cls_indexes = np.where(targets[:, 0] == cls)[0]
-                if cls_indexes.size > 0:
-                    class_counts[cls] += cls_indexes.size
-                    sums[cls, :] += targets[cls_indexes, 1:].sum(axis=0)
-                    squared_sums[cls, :] += (targets[cls_indexes, 1:] ** 2).sum(axis=0)
-
-        means = sums / class_counts
-        # var(x) = E(x^2) - E(x)^2
-        stds = np.sqrt(squared_sums / class_counts - means ** 2)
-
-    # normalized targets
-    for im_i in range(num_images):
-        targets = roidb[im_i]['bbox_targets']
-        for cls in range(1, num_classes):
-            cls_indexes = np.where(targets[:, 0] == cls)[0]
-            roidb[im_i]['bbox_targets'][cls_indexes, 1:] -= means[cls, :]
-            roidb[im_i]['bbox_targets'][cls_indexes, 1:] /= stds[cls, :]
-
-    return means.ravel(), stds.ravel()
-
-
-def expand_bbox_regression_targets(bbox_targets_data, num_classes):
-    """
-    expand from 5 to 4 * num_classes; only the right class has non-zero bbox regression targets
-    :param bbox_targets_data: [k * 5]
-    :param num_classes: number of classes
-    :return: bbox target processed [k * 4 num_classes]
-    bbox_weights ! only foreground boxes have bbox regression computation!
-    """
-    classes = bbox_targets_data[:, 0]
-    bbox_targets = np.zeros((classes.size, 4 * num_classes), dtype=np.float32)
-    bbox_weights = np.zeros(bbox_targets.shape, dtype=np.float32)
-    indexes = np.where(classes > 0)[0]
-    for index in indexes:
-        cls = classes[index]
-        start = int(4 * cls)
-        end = start + 4
-        bbox_targets[index, start:end] = bbox_targets_data[index, 1:]
-        bbox_weights[index, start:end] = config.TRAIN.BBOX_WEIGHTS
-    return bbox_targets, bbox_weights
-
-
-def compute_mask_and_label(ex_rois, ex_labels, seg, flipped):
-    # assert os.path.exists(seg_gt), 'Path does not exist: {}'.format(seg_gt)
-    # im = Image.open(seg_gt)
-    # pixel = list(im.getdata())
-    # pixel = np.array(pixel).reshape([im.size[1], im.size[0]])
-    im = Image.open(seg)
-    pixel = list(im.getdata())
-    ins_seg = np.array(pixel).reshape([im.size[1], im.size[0]])
-    if flipped:
-        ins_seg = ins_seg[:, ::-1]
-    rois = ex_rois
-    n_rois = ex_rois.shape[0]
-    label = ex_labels
-    class_id = config.CLASS_ID
-    mask_target = np.zeros((n_rois, 28, 28), dtype=np.int8)
-    mask_label = np.zeros((n_rois), dtype=np.int8)
-    for n in range(n_rois):
-        target = ins_seg[int(rois[n, 1]): int(rois[n, 3]), int(rois[n, 0]): int(rois[n, 2])]
-        ids = np.unique(target)
-        ins_id = 0
-        max_count = 0
-        for id in ids:
-            if math.floor(id / 1000) == class_id[int(label[int(n)])]:
-                px = np.where(ins_seg == int(id))
-                x_min = np.min(px[1])
-                y_min = np.min(px[0])
-                x_max = np.max(px[1])
-                y_max = np.max(px[0])
-                x1 = max(rois[n, 0], x_min)
-                y1 = max(rois[n, 1], y_min)
-                x2 = min(rois[n, 2], x_max)
-                y2 = min(rois[n, 3], y_max)
-                iou = (x2 - x1) * (y2 - y1)
-                iou = iou / ((rois[n, 2] - rois[n, 0]) * (rois[n, 3] - rois[n, 1])
-                             + (x_max - x_min) * (y_max - y_min) - iou)
-                if iou > max_count:
-                    ins_id = id
-                    max_count = iou
-
-        if max_count == 0:
-            continue
-        # print max_count
-        mask = np.zeros(target.shape)
-        idx = np.where(target == ins_id)
-        mask[idx] = 1
-        mask = cv2.resize(mask, (28, 28), interpolation=cv2.INTER_NEAREST)
-
-        mask_target[n] = mask
-        mask_label[n] = label[int(n)]
-    return mask_target, mask_label
-
-
-def compute_bbox_mask_targets_and_label(rois, overlaps, labels, seg, flipped):
-    """
-    given rois, overlaps, gt labels, seg, compute bounding box mask targets
-    :param rois: roidb[i]['boxes'] k * 4
-    :param overlaps: roidb[i]['max_overlaps'] k * 1
-    :param labels: roidb[i]['max_classes'] k * 1
-    :return: targets[i][class, dx, dy, dw, dh] k * 5
-    """
-    # Ensure ROIs are floats
-    rois = rois.astype(np.float, copy=False)
-
-    # Sanity check
-    if len(rois) != len(overlaps):
-        print 'bbox regression: this should not happen'
-
-    # Indices of ground-truth ROIs
-    gt_inds = np.where(overlaps == 1)[0]
-    if len(gt_inds) == 0:
-        print 'something wrong : zero ground truth rois'
-    # Indices of examples for which we try to make predictions
-    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
-
-    # Get IoU overlap between each ex ROI and gt ROI
-    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
-
-
-    # Find which gt ROI each ex ROI has max overlap with:
-    # this will be the ex ROI's gt target
-    gt_assignment = ex_gt_overlaps.argmax(axis=1)
-    gt_rois = rois[gt_inds[gt_assignment], :]
-    ex_rois = rois[ex_inds, :]
-
-    mask_targets, mask_label = compute_mask_and_label(ex_rois, labels[ex_inds], seg, flipped)
-    return mask_targets, mask_label, ex_inds
-
-def add_mask_targets(roidb):
-    """
-    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    :return: means, std variances of targets
-    """
-    print 'add bounding box mask targets'
-    assert len(roidb) > 0
-    assert 'max_classes' in roidb[0]
-
-    num_images = len(roidb)
-
-    # Multi threads processing
-    im_quene = Queue.Queue(maxsize=0)
-    for im_i in range(num_images):
-        im_quene.put(im_i)
-
-    def process():
-        while not im_quene.empty():
-            im_i = im_quene.get()
-            print "-----process img {}".format(im_i)
-            rois = roidb[im_i]['boxes']
-            max_overlaps = roidb[im_i]['max_overlaps']
-            max_classes = roidb[im_i]['max_classes']
-            ins_seg = roidb[im_i]['ins_seg']
-            flipped = roidb[im_i]['flipped']
-            roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
-                compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg, flipped)
-    threads = [threading.Thread(target=process, args=()) for i in range(10)]
-    for t in threads: t.start()
-    for t in threads: t.join()
-    # Single thread
-    # for im_i in range(num_images):
-    #     print "-----processing img {}".format(im_i)
-    #     rois = roidb[im_i]['boxes']
-    #     max_overlaps = roidb[im_i]['max_overlaps']
-    #     max_classes = roidb[im_i]['max_classes']
-    #     ins_seg = roidb[im_i]['ins_seg']
-    #     # roidb[im_i]['mask_targets'] = compute_bbox_mask_targets(rois, max_overlaps, max_classes, ins_seg)
-    #     roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
-    #         compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg)
diff --git a/RetinaFace/rcnn/processing/bbox_transform.py b/RetinaFace/rcnn/processing/bbox_transform.py
deleted file mode 100644
index 0c4fc7c8b..000000000
--- a/RetinaFace/rcnn/processing/bbox_transform.py
+++ /dev/null
@@ -1,218 +0,0 @@
-import numpy as np
-from ..cython.bbox import bbox_overlaps_cython
-#from rcnn.config import config
-
-
-def bbox_overlaps(boxes, query_boxes):
-    return bbox_overlaps_cython(boxes, query_boxes)
-
-
-def bbox_overlaps_py(boxes, query_boxes):
-    """
-    determine overlaps between boxes and query_boxes
-    :param boxes: n * 4 bounding boxes
-    :param query_boxes: k * 4 bounding boxes
-    :return: overlaps: n * k overlaps
-    """
-    n_ = boxes.shape[0]
-    k_ = query_boxes.shape[0]
-    overlaps = np.zeros((n_, k_), dtype=np.float)
-    for k in range(k_):
-        query_box_area = (query_boxes[k, 2] - query_boxes[k, 0] + 1) * (query_boxes[k, 3] - query_boxes[k, 1] + 1)
-        for n in range(n_):
-            iw = min(boxes[n, 2], query_boxes[k, 2]) - max(boxes[n, 0], query_boxes[k, 0]) + 1
-            if iw > 0:
-                ih = min(boxes[n, 3], query_boxes[k, 3]) - max(boxes[n, 1], query_boxes[k, 1]) + 1
-                if ih > 0:
-                    box_area = (boxes[n, 2] - boxes[n, 0] + 1) * (boxes[n, 3] - boxes[n, 1] + 1)
-                    all_area = float(box_area + query_box_area - iw * ih)
-                    overlaps[n, k] = iw * ih / all_area
-    return overlaps
-
-
-def clip_boxes(boxes, im_shape):
-    """
-    Clip boxes to image boundaries.
-    :param boxes: [N, 4* num_classes]
-    :param im_shape: tuple of 2
-    :return: [N, 4* num_classes]
-    """
-    # x1 >= 0
-    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
-    # y1 >= 0
-    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
-    # x2 < im_shape[1]
-    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
-    # y2 < im_shape[0]
-    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
-    return boxes
-
-
-def nonlinear_transform(ex_rois, gt_rois):
-    """
-    compute bounding box regression targets from ex_rois to gt_rois
-    :param ex_rois: [N, 4]
-    :param gt_rois: [N, 4]
-    :return: [N, 4]
-    """
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-
-    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
-    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
-    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
-    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
-
-    gt_widths = gt_rois[:, 2] - gt_rois[:, 0] + 1.0
-    gt_heights = gt_rois[:, 3] - gt_rois[:, 1] + 1.0
-    gt_ctr_x = gt_rois[:, 0] + 0.5 * (gt_widths - 1.0)
-    gt_ctr_y = gt_rois[:, 1] + 0.5 * (gt_heights - 1.0)
-
-    targets_dx = (gt_ctr_x - ex_ctr_x) / (ex_widths + 1e-14)
-    targets_dy = (gt_ctr_y - ex_ctr_y) / (ex_heights + 1e-14)
-    targets_dw = np.log(gt_widths / ex_widths)
-    targets_dh = np.log(gt_heights / ex_heights)
-
-    if gt_rois.shape[1]<=4:
-      targets = np.vstack(
-          (targets_dx, targets_dy, targets_dw, targets_dh)).transpose()
-      return targets
-    else:
-      targets = [targets_dx, targets_dy, targets_dw, targets_dh]
-      #if config.USE_BLUR:
-      #  for i in range(4, gt_rois.shape[1]):
-      #    t = gt_rois[:,i]
-      #    targets.append(t)
-      targets = np.vstack(targets).transpose()
-      return targets
-
-def landmark_transform(ex_rois, gt_rois):
-
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-
-    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
-    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
-    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
-    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
-
-    
-    targets = []
-    for i in range(gt_rois.shape[1]):
-      for j in range(gt_rois.shape[2]):
-        #if not config.USE_OCCLUSION and j==2:
-        #  continue
-        if j==2:
-          continue
-        if j==0: #w
-          target = (gt_rois[:,i,j] - ex_ctr_x) / (ex_widths + 1e-14)
-        elif j==1: #h
-          target = (gt_rois[:,i,j] - ex_ctr_y) / (ex_heights + 1e-14)
-        else: #visibile
-          target = gt_rois[:,i,j]
-        targets.append(target)
-
-
-    targets = np.vstack(targets).transpose()
-    return targets
-
-
-def nonlinear_pred(boxes, box_deltas):
-    """
-    Transform the set of class-agnostic boxes into class-specific boxes
-    by applying the predicted offsets (box_deltas)
-    :param boxes: !important [N 4]
-    :param box_deltas: [N, 4 * num_classes]
-    :return: [N 4 * num_classes]
-    """
-    if boxes.shape[0] == 0:
-        return np.zeros((0, box_deltas.shape[1]))
-
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-    dx = box_deltas[:, 0::4]
-    dy = box_deltas[:, 1::4]
-    dw = box_deltas[:, 2::4]
-    dh = box_deltas[:, 3::4]
-
-    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-    pred_w = np.exp(dw) * widths[:, np.newaxis]
-    pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-    pred_boxes = np.zeros(box_deltas.shape)
-    # x1
-    pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-    # y1
-    pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-    # x2
-    pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-    # y2
-    pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-
-    return pred_boxes
-
-def landmark_pred(boxes, landmark_deltas):
-    if boxes.shape[0] == 0:
-        return np.zeros((0, landmark_deltas.shape[1]))
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-    preds = []
-    for i in range(landmark_deltas.shape[1]):
-      if i%2==0:
-        pred = (landmark_deltas[:,i]*widths + ctr_x)
-      else:
-        pred = (landmark_deltas[:,i]*heights + ctr_y)
-      preds.append(pred)
-    preds = np.vstack(preds).transpose()
-    return preds
-
-def iou_transform(ex_rois, gt_rois):
-    """ return bbox targets, IoU loss uses gt_rois as gt """
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-    return gt_rois
-
-
-def iou_pred(boxes, box_deltas):
-    """
-    Transform the set of class-agnostic boxes into class-specific boxes
-    by applying the predicted offsets (box_deltas)
-    :param boxes: !important [N 4]
-    :param box_deltas: [N, 4 * num_classes]
-    :return: [N 4 * num_classes]
-    """
-    if boxes.shape[0] == 0:
-        return np.zeros((0, box_deltas.shape[1]))
-
-    boxes = boxes.astype(np.float, copy=False)
-    x1 = boxes[:, 0]
-    y1 = boxes[:, 1]
-    x2 = boxes[:, 2]
-    y2 = boxes[:, 3]
-
-    dx1 = box_deltas[:, 0::4]
-    dy1 = box_deltas[:, 1::4]
-    dx2 = box_deltas[:, 2::4]
-    dy2 = box_deltas[:, 3::4]
-
-    pred_boxes = np.zeros(box_deltas.shape)
-    # x1
-    pred_boxes[:, 0::4] = dx1 + x1[:, np.newaxis]
-    # y1
-    pred_boxes[:, 1::4] = dy1 + y1[:, np.newaxis]
-    # x2
-    pred_boxes[:, 2::4] = dx2 + x2[:, np.newaxis]
-    # y2
-    pred_boxes[:, 3::4] = dy2 + y2[:, np.newaxis]
-
-    return pred_boxes
-
-
-# define bbox_transform and bbox_pred
-bbox_transform = nonlinear_transform
-bbox_pred = nonlinear_pred
diff --git a/RetinaFace/rcnn/processing/generate_anchor.py b/RetinaFace/rcnn/processing/generate_anchor.py
deleted file mode 100644
index eb47d8ea5..000000000
--- a/RetinaFace/rcnn/processing/generate_anchor.py
+++ /dev/null
@@ -1,128 +0,0 @@
-"""
-Generate base anchors on index 0
-"""
-from __future__ import print_function
-import sys
-from builtins import range
-import numpy as np
-from ..cython.anchors import anchors_cython
-#from ..config import config
-
-
-def anchors_plane(feat_h, feat_w, stride, base_anchor):
-    return anchors_cython(feat_h, feat_w, stride, base_anchor)
-
-def generate_anchors(base_size=16, ratios=[0.5, 1, 2],
-                     scales=2 ** np.arange(3, 6), stride=16, dense_anchor=False):
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-
-    base_anchor = np.array([1, 1, base_size, base_size]) - 1
-    ratio_anchors = _ratio_enum(base_anchor, ratios)
-    anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales)
-                         for i in range(ratio_anchors.shape[0])])
-    if dense_anchor:
-      assert stride%2==0
-      anchors2 = anchors.copy()
-      anchors2[:,:] += int(stride/2)
-      anchors = np.vstack( (anchors, anchors2) )
-    #print('GA',base_anchor.shape, ratio_anchors.shape, anchors.shape)
-    return anchors
-
-#def generate_anchors_fpn(base_size=[64,32,16,8,4], ratios=[0.5, 1, 2], scales=8):
-#    """
-#    Generate anchor (reference) windows by enumerating aspect ratios X
-#    scales wrt a reference (0, 0, 15, 15) window.
-#    """
-#    anchors = []
-#    _ratios = ratios.reshape( (len(base_size), -1) )
-#    _scales = scales.reshape( (len(base_size), -1) )
-#    for i,bs in enumerate(base_size):
-#      __ratios = _ratios[i]
-#      __scales = _scales[i]
-#      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
-#      r = generate_anchors(bs, __ratios, __scales)
-#      #print('anchors_fpn', r.shape, file=sys.stderr)
-#      anchors.append(r)
-#    return anchors
-
-def generate_anchors_fpn(dense_anchor=False, cfg = None):
-    #assert(False)
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-    if cfg is None:
-      from ..config import config
-      cfg = config.RPN_ANCHOR_CFG
-    RPN_FEAT_STRIDE = []
-    for k in cfg:
-      RPN_FEAT_STRIDE.append( int(k) )
-    RPN_FEAT_STRIDE = sorted(RPN_FEAT_STRIDE, reverse=True)
-    anchors = []
-    for k in RPN_FEAT_STRIDE:
-      v = cfg[str(k)]
-      bs = v['BASE_SIZE']
-      __ratios = np.array(v['RATIOS'])
-      __scales = np.array(v['SCALES'])
-      stride = int(k)
-      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
-      r = generate_anchors(bs, __ratios, __scales, stride, dense_anchor)
-      #print('anchors_fpn', r.shape, file=sys.stderr)
-      anchors.append(r)
-
-    return anchors
-
-def _whctrs(anchor):
-    """
-    Return width, height, x center, and y center for an anchor (window).
-    """
-
-    w = anchor[2] - anchor[0] + 1
-    h = anchor[3] - anchor[1] + 1
-    x_ctr = anchor[0] + 0.5 * (w - 1)
-    y_ctr = anchor[1] + 0.5 * (h - 1)
-    return w, h, x_ctr, y_ctr
-
-
-def _mkanchors(ws, hs, x_ctr, y_ctr):
-    """
-    Given a vector of widths (ws) and heights (hs) around a center
-    (x_ctr, y_ctr), output a set of anchors (windows).
-    """
-
-    ws = ws[:, np.newaxis]
-    hs = hs[:, np.newaxis]
-    anchors = np.hstack((x_ctr - 0.5 * (ws - 1),
-                         y_ctr - 0.5 * (hs - 1),
-                         x_ctr + 0.5 * (ws - 1),
-                         y_ctr + 0.5 * (hs - 1)))
-    return anchors
-
-
-def _ratio_enum(anchor, ratios):
-    """
-    Enumerate a set of anchors for each aspect ratio wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    size = w * h
-    size_ratios = size / ratios
-    ws = np.round(np.sqrt(size_ratios))
-    hs = np.round(ws * ratios)
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
-
-
-def _scale_enum(anchor, scales):
-    """
-    Enumerate a set of anchors for each scale wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    ws = w * scales
-    hs = h * scales
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
diff --git a/RetinaFace/rcnn/processing/nms.py b/RetinaFace/rcnn/processing/nms.py
deleted file mode 100644
index 230139c41..000000000
--- a/RetinaFace/rcnn/processing/nms.py
+++ /dev/null
@@ -1,64 +0,0 @@
-import numpy as np
-from ..cython.cpu_nms import cpu_nms
-try:
-    from ..cython.gpu_nms import gpu_nms
-except ImportError:
-    gpu_nms = None
-
-
-def py_nms_wrapper(thresh):
-    def _nms(dets):
-        return nms(dets, thresh)
-    return _nms
-
-
-def cpu_nms_wrapper(thresh):
-    def _nms(dets):
-        return cpu_nms(dets, thresh)
-    return _nms
-
-
-def gpu_nms_wrapper(thresh, device_id):
-    def _nms(dets):
-        return gpu_nms(dets, thresh, device_id)
-    if gpu_nms is not None:
-        return _nms
-    else:
-        return cpu_nms_wrapper(thresh)
-
-
-def nms(dets, thresh):
-    """
-    greedily select boxes with high confidence and overlap with current maximum <= thresh
-    rule out overlap >= thresh
-    :param dets: [[x1, y1, x2, y2 score]]
-    :param thresh: retain overlap < thresh
-    :return: indexes to keep
-    """
-    x1 = dets[:, 0]
-    y1 = dets[:, 1]
-    x2 = dets[:, 2]
-    y2 = dets[:, 3]
-    scores = dets[:, 4]
-
-    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
-    order = scores.argsort()[::-1]
-
-    keep = []
-    while order.size > 0:
-        i = order[0]
-        keep.append(i)
-        xx1 = np.maximum(x1[i], x1[order[1:]])
-        yy1 = np.maximum(y1[i], y1[order[1:]])
-        xx2 = np.minimum(x2[i], x2[order[1:]])
-        yy2 = np.minimum(y2[i], y2[order[1:]])
-
-        w = np.maximum(0.0, xx2 - xx1 + 1)
-        h = np.maximum(0.0, yy2 - yy1 + 1)
-        inter = w * h
-        ovr = inter / (areas[i] + areas[order[1:]] - inter)
-
-        inds = np.where(ovr <= thresh)[0]
-        order = order[inds + 1]
-
-    return keep
diff --git a/RetinaFace/rcnn/pycocotools/coco.py b/RetinaFace/rcnn/pycocotools/coco.py
deleted file mode 100644
index 2b63ea1ea..000000000
--- a/RetinaFace/rcnn/pycocotools/coco.py
+++ /dev/null
@@ -1,428 +0,0 @@
-__author__ = 'tylin'
-__version__ = '2.0'
-# Interface for accessing the Microsoft COCO dataset.
-
-# Microsoft COCO is a large image dataset designed for object detection,
-# segmentation, and caption generation. pycocotools is a Python API that
-# assists in loading, parsing and visualizing the annotations in COCO.
-# Please visit http://mscoco.org/ for more information on COCO, including
-# for the data, paper, and tutorials. The exact format of the annotations
-# is also described on the COCO website. For example usage of the pycocotools
-# please see pycocotools_demo.ipynb. In addition to this API, please download both
-# the COCO images and annotations in order to run the demo.
-
-# An alternative to using the API is to load the annotations directly
-# into Python dictionary
-# Using the API provides additional utility functions. Note that this API
-# supports both *instance* and *caption* annotations. In the case of
-# captions not all functions are defined (e.g. categories are undefined).
-
-# The following API functions are defined:
-#  COCO       - COCO api class that loads COCO annotation file and prepare data structures.
-#  decodeMask - Decode binary mask M encoded via run-length encoding.
-#  encodeMask - Encode binary mask M using run-length encoding.
-#  getAnnIds  - Get ann ids that satisfy given filter conditions.
-#  getCatIds  - Get cat ids that satisfy given filter conditions.
-#  getImgIds  - Get img ids that satisfy given filter conditions.
-#  loadAnns   - Load anns with the specified ids.
-#  loadCats   - Load cats with the specified ids.
-#  loadImgs   - Load imgs with the specified ids.
-#  annToMask  - Convert segmentation in an annotation to binary mask.
-#  showAnns   - Display the specified annotations.
-#  loadRes    - Load algorithm results and create API for accessing them.
-#  download   - Download COCO images from mscoco.org server.
-# Throughout the API "ann"=annotation, "cat"=category, and "img"=image.
-# Help on each functions can be accessed by: "help COCO>function".
-
-# See also COCO>decodeMask,
-# COCO>encodeMask, COCO>getAnnIds, COCO>getCatIds,
-# COCO>getImgIds, COCO>loadAnns, COCO>loadCats,
-# COCO>loadImgs, COCO>annToMask, COCO>showAnns
-
-# Microsoft COCO Toolbox.      version 2.0
-# Data, paper, and tutorials available at:  http://mscoco.org/
-# Code written by Piotr Dollar and Tsung-Yi Lin, 2014.
-# Licensed under the Simplified BSD License [see bsd.txt]
-
-import json
-import time
-import matplotlib.pyplot as plt
-from matplotlib.collections import PatchCollection
-from matplotlib.patches import Polygon
-import numpy as np
-import copy
-import itertools
-from . import mask as maskUtils
-import os
-from collections import defaultdict
-import sys
-PYTHON_VERSION = sys.version_info[0]
-if PYTHON_VERSION == 2:
-    from urllib import urlretrieve
-elif PYTHON_VERSION == 3:
-    from urllib.request import urlretrieve
-
-class COCO:
-    def __init__(self, annotation_file=None):
-        """
-        Constructor of Microsoft COCO helper class for reading and visualizing annotations.
-        :param annotation_file (str): location of annotation file
-        :param image_folder (str): location to the folder that hosts images.
-        :return:
-        """
-        # load dataset
-        self.dataset,self.anns,self.cats,self.imgs = dict(),dict(),dict(),dict()
-        self.imgToAnns, self.catToImgs = defaultdict(list), defaultdict(list)
-        if not annotation_file == None:
-            print('loading annotations into memory...')
-            tic = time.time()
-            dataset = json.load(open(annotation_file, 'r'))
-            assert type(dataset)==dict, 'annotation file format {} not supported'.format(type(dataset))
-            print('Done (t={:0.2f}s)'.format(time.time()- tic))
-            self.dataset = dataset
-            self.createIndex()
-
-    def createIndex(self):
-        # create index
-        print('creating index...')
-        anns, cats, imgs = {}, {}, {}
-        imgToAnns,catToImgs = defaultdict(list),defaultdict(list)
-        if 'annotations' in self.dataset:
-            for ann in self.dataset['annotations']:
-                imgToAnns[ann['image_id']].append(ann)
-                anns[ann['id']] = ann
-
-        if 'images' in self.dataset:
-            for img in self.dataset['images']:
-                imgs[img['id']] = img
-
-        if 'categories' in self.dataset:
-            for cat in self.dataset['categories']:
-                cats[cat['id']] = cat
-
-        if 'annotations' in self.dataset and 'categories' in self.dataset:
-            for ann in self.dataset['annotations']:
-                catToImgs[ann['category_id']].append(ann['image_id'])
-
-        print('index created!')
-
-        # create class members
-        self.anns = anns
-        self.imgToAnns = imgToAnns
-        self.catToImgs = catToImgs
-        self.imgs = imgs
-        self.cats = cats
-
-    def info(self):
-        """
-        Print information about the annotation file.
-        :return:
-        """
-        for key, value in self.dataset['info'].items():
-            print('{}: {}'.format(key, value))
-
-    def getAnnIds(self, imgIds=[], catIds=[], areaRng=[], iscrowd=None):
-        """
-        Get ann ids that satisfy given filter conditions. default skips that filter
-        :param imgIds  (int array)     : get anns for given imgs
-               catIds  (int array)     : get anns for given cats
-               areaRng (float array)   : get anns for given area range (e.g. [0 inf])
-               iscrowd (boolean)       : get anns for given crowd label (False or True)
-        :return: ids (int array)       : integer array of ann ids
-        """
-        imgIds = imgIds if type(imgIds) == list else [imgIds]
-        catIds = catIds if type(catIds) == list else [catIds]
-
-        if len(imgIds) == len(catIds) == len(areaRng) == 0:
-            anns = self.dataset['annotations']
-        else:
-            if not len(imgIds) == 0:
-                lists = [self.imgToAnns[imgId] for imgId in imgIds if imgId in self.imgToAnns]
-                anns = list(itertools.chain.from_iterable(lists))
-            else:
-                anns = self.dataset['annotations']
-            anns = anns if len(catIds)  == 0 else [ann for ann in anns if ann['category_id'] in catIds]
-            anns = anns if len(areaRng) == 0 else [ann for ann in anns if ann['area'] > areaRng[0] and ann['area'] < areaRng[1]]
-        if not iscrowd == None:
-            ids = [ann['id'] for ann in anns if ann['iscrowd'] == iscrowd]
-        else:
-            ids = [ann['id'] for ann in anns]
-        return ids
-
-    def getCatIds(self, catNms=[], supNms=[], catIds=[]):
-        """
-        filtering parameters. default skips that filter.
-        :param catNms (str array)  : get cats for given cat names
-        :param supNms (str array)  : get cats for given supercategory names
-        :param catIds (int array)  : get cats for given cat ids
-        :return: ids (int array)   : integer array of cat ids
-        """
-        catNms = catNms if type(catNms) == list else [catNms]
-        supNms = supNms if type(supNms) == list else [supNms]
-        catIds = catIds if type(catIds) == list else [catIds]
-
-        if len(catNms) == len(supNms) == len(catIds) == 0:
-            cats = self.dataset['categories']
-        else:
-            cats = self.dataset['categories']
-            cats = cats if len(catNms) == 0 else [cat for cat in cats if cat['name']          in catNms]
-            cats = cats if len(supNms) == 0 else [cat for cat in cats if cat['supercategory'] in supNms]
-            cats = cats if len(catIds) == 0 else [cat for cat in cats if cat['id']            in catIds]
-        ids = [cat['id'] for cat in cats]
-        return ids
-
-    def getImgIds(self, imgIds=[], catIds=[]):
-        '''
-        Get img ids that satisfy given filter conditions.
-        :param imgIds (int array) : get imgs for given ids
-        :param catIds (int array) : get imgs with all given cats
-        :return: ids (int array)  : integer array of img ids
-        '''
-        imgIds = imgIds if type(imgIds) == list else [imgIds]
-        catIds = catIds if type(catIds) == list else [catIds]
-
-        if len(imgIds) == len(catIds) == 0:
-            ids = self.imgs.keys()
-        else:
-            ids = set(imgIds)
-            for i, catId in enumerate(catIds):
-                if i == 0 and len(ids) == 0:
-                    ids = set(self.catToImgs[catId])
-                else:
-                    ids &= set(self.catToImgs[catId])
-        return list(ids)
-
-    def loadAnns(self, ids=[]):
-        """
-        Load anns with the specified ids.
-        :param ids (int array)       : integer ids specifying anns
-        :return: anns (object array) : loaded ann objects
-        """
-        if type(ids) == list:
-            return [self.anns[id] for id in ids]
-        elif type(ids) == int:
-            return [self.anns[ids]]
-
-    def loadCats(self, ids=[]):
-        """
-        Load cats with the specified ids.
-        :param ids (int array)       : integer ids specifying cats
-        :return: cats (object array) : loaded cat objects
-        """
-        if type(ids) == list:
-            return [self.cats[id] for id in ids]
-        elif type(ids) == int:
-            return [self.cats[ids]]
-
-    def loadImgs(self, ids=[]):
-        """
-        Load anns with the specified ids.
-        :param ids (int array)       : integer ids specifying img
-        :return: imgs (object array) : loaded img objects
-        """
-        if type(ids) == list:
-            return [self.imgs[id] for id in ids]
-        elif type(ids) == int:
-            return [self.imgs[ids]]
-
-    def showAnns(self, anns):
-        """
-        Display the specified annotations.
-        :param anns (array of object): annotations to display
-        :return: None
-        """
-        if len(anns) == 0:
-            return 0
-        if 'segmentation' in anns[0] or 'keypoints' in anns[0]:
-            datasetType = 'instances'
-        elif 'caption' in anns[0]:
-            datasetType = 'captions'
-        else:
-            raise Exception('datasetType not supported')
-        if datasetType == 'instances':
-            ax = plt.gca()
-            ax.set_autoscale_on(False)
-            polygons = []
-            color = []
-            for ann in anns:
-                c = (np.random.random((1, 3))*0.6+0.4).tolist()[0]
-                if 'segmentation' in ann:
-                    if type(ann['segmentation']) == list:
-                        # polygon
-                        for seg in ann['segmentation']:
-                            poly = np.array(seg).reshape((int(len(seg)/2), 2))
-                            polygons.append(Polygon(poly))
-                            color.append(c)
-                    else:
-                        # mask
-                        t = self.imgs[ann['image_id']]
-                        if type(ann['segmentation']['counts']) == list:
-                            rle = maskUtils.frPyObjects([ann['segmentation']], t['height'], t['width'])
-                        else:
-                            rle = [ann['segmentation']]
-                        m = maskUtils.decode(rle)
-                        img = np.ones( (m.shape[0], m.shape[1], 3) )
-                        if ann['iscrowd'] == 1:
-                            color_mask = np.array([2.0,166.0,101.0])/255
-                        if ann['iscrowd'] == 0:
-                            color_mask = np.random.random((1, 3)).tolist()[0]
-                        for i in range(3):
-                            img[:,:,i] = color_mask[i]
-                        ax.imshow(np.dstack( (img, m*0.5) ))
-                if 'keypoints' in ann and type(ann['keypoints']) == list:
-                    # turn skeleton into zero-based index
-                    sks = np.array(self.loadCats(ann['category_id'])[0]['skeleton'])-1
-                    kp = np.array(ann['keypoints'])
-                    x = kp[0::3]
-                    y = kp[1::3]
-                    v = kp[2::3]
-                    for sk in sks:
-                        if np.all(v[sk]>0):
-                            plt.plot(x[sk],y[sk], linewidth=3, color=c)
-                    plt.plot(x[v>0], y[v>0],'o',markersize=8, markerfacecolor=c, markeredgecolor='k',markeredgewidth=2)
-                    plt.plot(x[v>1], y[v>1],'o',markersize=8, markerfacecolor=c, markeredgecolor=c, markeredgewidth=2)
-            p = PatchCollection(polygons, facecolor=color, linewidths=0, alpha=0.4)
-            ax.add_collection(p)
-            p = PatchCollection(polygons, facecolor='none', edgecolors=color, linewidths=2)
-            ax.add_collection(p)
-        elif datasetType == 'captions':
-            for ann in anns:
-                print(ann['caption'])
-
-    def loadRes(self, resFile):
-        """
-        Load result file and return a result api object.
-        :param   resFile (str)     : file name of result file
-        :return: res (obj)         : result api object
-        """
-        res = COCO()
-        res.dataset['images'] = [img for img in self.dataset['images']]
-
-        print('Loading and preparing results...')
-        tic = time.time()
-        if type(resFile) == str or type(resFile) == unicode:
-            anns = json.load(open(resFile))
-        elif type(resFile) == np.ndarray:
-            anns = self.loadNumpyAnnotations(resFile)
-        else:
-            anns = resFile
-        assert type(anns) == list, 'results in not an array of objects'
-        annsImgIds = [ann['image_id'] for ann in anns]
-        assert set(annsImgIds) == (set(annsImgIds) & set(self.getImgIds())), \
-               'Results do not correspond to current coco set'
-        if 'caption' in anns[0]:
-            imgIds = set([img['id'] for img in res.dataset['images']]) & set([ann['image_id'] for ann in anns])
-            res.dataset['images'] = [img for img in res.dataset['images'] if img['id'] in imgIds]
-            for id, ann in enumerate(anns):
-                ann['id'] = id+1
-        elif 'bbox' in anns[0] and not anns[0]['bbox'] == []:
-            res.dataset['categories'] = copy.deepcopy(self.dataset['categories'])
-            for id, ann in enumerate(anns):
-                bb = ann['bbox']
-                x1, x2, y1, y2 = [bb[0], bb[0]+bb[2], bb[1], bb[1]+bb[3]]
-                if not 'segmentation' in ann:
-                    ann['segmentation'] = [[x1, y1, x1, y2, x2, y2, x2, y1]]
-                ann['area'] = bb[2]*bb[3]
-                ann['id'] = id+1
-                ann['iscrowd'] = 0
-        elif 'segmentation' in anns[0]:
-            res.dataset['categories'] = copy.deepcopy(self.dataset['categories'])
-            for id, ann in enumerate(anns):
-                # now only support compressed RLE format as segmentation results
-                ann['area'] = maskUtils.area(ann['segmentation'])
-                if not 'bbox' in ann:
-                    ann['bbox'] = maskUtils.toBbox(ann['segmentation'])
-                ann['id'] = id+1
-                ann['iscrowd'] = 0
-        elif 'keypoints' in anns[0]:
-            res.dataset['categories'] = copy.deepcopy(self.dataset['categories'])
-            for id, ann in enumerate(anns):
-                s = ann['keypoints']
-                x = s[0::3]
-                y = s[1::3]
-                x0,x1,y0,y1 = np.min(x), np.max(x), np.min(y), np.max(y)
-                ann['area'] = (x1-x0)*(y1-y0)
-                ann['id'] = id + 1
-                ann['bbox'] = [x0,y0,x1-x0,y1-y0]
-        print('DONE (t={:0.2f}s)'.format(time.time()- tic))
-
-        res.dataset['annotations'] = anns
-        res.createIndex()
-        return res
-
-    def download(self, tarDir = None, imgIds = [] ):
-        '''
-        Download COCO images from mscoco.org server.
-        :param tarDir (str): COCO results directory name
-               imgIds (list): images to be downloaded
-        :return:
-        '''
-        if tarDir is None:
-            print('Please specify target directory')
-            return -1
-        if len(imgIds) == 0:
-            imgs = self.imgs.values()
-        else:
-            imgs = self.loadImgs(imgIds)
-        N = len(imgs)
-        if not os.path.exists(tarDir):
-            os.makedirs(tarDir)
-        for i, img in enumerate(imgs):
-            tic = time.time()
-            fname = os.path.join(tarDir, img['file_name'])
-            if not os.path.exists(fname):
-                urlretrieve(img['coco_url'], fname)
-            print('downloaded {}/{} images (t={:0.1f}s)'.format(i, N, time.time()- tic))
-
-    def loadNumpyAnnotations(self, data):
-        """
-        Convert result data from a numpy array [Nx7] where each row contains {imageID,x1,y1,w,h,score,class}
-        :param  data (numpy.ndarray)
-        :return: annotations (python nested list)
-        """
-        print('Converting ndarray to lists...')
-        assert(type(data) == np.ndarray)
-        print(data.shape)
-        assert(data.shape[1] == 7)
-        N = data.shape[0]
-        ann = []
-        for i in range(N):
-            if i % 1000000 == 0:
-                print('{}/{}'.format(i,N))
-            ann += [{
-                'image_id'  : int(data[i, 0]),
-                'bbox'  : [ data[i, 1], data[i, 2], data[i, 3], data[i, 4] ],
-                'score' : data[i, 5],
-                'category_id': int(data[i, 6]),
-                }]
-        return ann
-
-    def annToRLE(self, ann):
-        """
-        Convert annotation which can be polygons, uncompressed RLE to RLE.
-        :return: binary mask (numpy 2D array)
-        """
-        t = self.imgs[ann['image_id']]
-        h, w = t['height'], t['width']
-        segm = ann['segmentation']
-        if type(segm) == list:
-            # polygon -- a single object might consist of multiple parts
-            # we merge all parts into one mask rle code
-            rles = maskUtils.frPyObjects(segm, h, w)
-            rle = maskUtils.merge(rles)
-        elif type(segm['counts']) == list:
-            # uncompressed RLE
-            rle = maskUtils.frPyObjects(segm, h, w)
-        else:
-            # rle
-            rle = ann['segmentation']
-        return rle
-
-    def annToMask(self, ann):
-        """
-        Convert annotation which can be polygons, uncompressed RLE, or RLE to binary mask.
-        :return: binary mask (numpy 2D array)
-        """
-        rle = self.annToRLE(ann)
-        m = maskUtils.decode(rle)
-        return m
diff --git a/RetinaFace/rcnn/pycocotools/cocoeval.py b/RetinaFace/rcnn/pycocotools/cocoeval.py
deleted file mode 100644
index 58975db15..000000000
--- a/RetinaFace/rcnn/pycocotools/cocoeval.py
+++ /dev/null
@@ -1,528 +0,0 @@
-__author__ = 'tsungyi'
-
-import numpy as np
-import datetime
-import time
-from collections import defaultdict
-from .mask import *
-import copy
-
-class COCOeval:
-    # Interface for evaluating detection on the Microsoft COCO dataset.
-    #
-    # The usage for CocoEval is as follows:
-    #  cocoGt=..., cocoDt=...       # load dataset and results
-    #  E = CocoEval(cocoGt,cocoDt); # initialize CocoEval object
-    #  E.params.recThrs = ...;      # set parameters as desired
-    #  E.evaluate();                # run per image evaluation
-    #  E.accumulate();              # accumulate per image results
-    #  E.summarize();               # display summary metrics of results
-    # For example usage see evalDemo.m and http://mscoco.org/.
-    #
-    # The evaluation parameters are as follows (defaults in brackets):
-    #  imgIds     - [all] N img ids to use for evaluation
-    #  catIds     - [all] K cat ids to use for evaluation
-    #  iouThrs    - [.5:.05:.95] T=10 IoU thresholds for evaluation
-    #  recThrs    - [0:.01:1] R=101 recall thresholds for evaluation
-    #  areaRng    - [...] A=4 object area ranges for evaluation
-    #  maxDets    - [1 10 100] M=3 thresholds on max detections per image
-    #  iouType    - ['segm'] set iouType to 'segm', 'bbox' or 'keypoints'
-    #  iouType replaced the now DEPRECATED useSegm parameter.
-    #  useCats    - [1] if true use category labels for evaluation
-    # Note: if useCats=0 category labels are ignored as in proposal scoring.
-    # Note: multiple areaRngs [Ax2] and maxDets [Mx1] can be specified.
-    #
-    # evaluate(): evaluates detections on every image and every category and
-    # concats the results into the "evalImgs" with fields:
-    #  dtIds      - [1xD] id for each of the D detections (dt)
-    #  gtIds      - [1xG] id for each of the G ground truths (gt)
-    #  dtMatches  - [TxD] matching gt id at each IoU or 0
-    #  gtMatches  - [TxG] matching dt id at each IoU or 0
-    #  dtScores   - [1xD] confidence of each dt
-    #  gtIgnore   - [1xG] ignore flag for each gt
-    #  dtIgnore   - [TxD] ignore flag for each dt at each IoU
-    #
-    # accumulate(): accumulates the per-image, per-category evaluation
-    # results in "evalImgs" into the dictionary "eval" with fields:
-    #  params     - parameters used for evaluation
-    #  date       - date evaluation was performed
-    #  counts     - [T,R,K,A,M] parameter dimensions (see above)
-    #  precision  - [TxRxKxAxM] precision for every evaluation setting
-    #  recall     - [TxKxAxM] max recall for every evaluation setting
-    # Note: precision and recall==-1 for settings with no gt objects.
-    #
-    # See also coco, mask, pycocoDemo, pycocoEvalDemo
-    #
-    # Microsoft COCO Toolbox.      version 2.0
-    # Data, paper, and tutorials available at:  http://mscoco.org/
-    # Code written by Piotr Dollar and Tsung-Yi Lin, 2015.
-    # Licensed under the Simplified BSD License [see coco/license.txt]
-    def __init__(self, cocoGt=None, cocoDt=None, iouType='segm'):
-        '''
-        Initialize CocoEval using coco APIs for gt and dt
-        :param cocoGt: coco object with ground truth annotations
-        :param cocoDt: coco object with detection results
-        :return: None
-        '''
-        if not iouType:
-            print('iouType not specified. use default iouType segm')
-        self.cocoGt   = cocoGt              # ground truth COCO API
-        self.cocoDt   = cocoDt              # detections COCO API
-        self.params   = {}                  # evaluation parameters
-        self.evalImgs = defaultdict(list)   # per-image per-category evaluation results [KxAxI] elements
-        self.eval     = {}                  # accumulated evaluation results
-        self._gts = defaultdict(list)       # gt for evaluation
-        self._dts = defaultdict(list)       # dt for evaluation
-        self.params = Params(iouType=iouType) # parameters
-        self._paramsEval = {}               # parameters for evaluation
-        self.stats = []                     # result summarization
-        self.ious = {}                      # ious between all gts and dts
-        if not cocoGt is None:
-            self.params.imgIds = sorted(cocoGt.getImgIds())
-            self.params.catIds = sorted(cocoGt.getCatIds())
-
-
-    def _prepare(self):
-        '''
-        Prepare ._gts and ._dts for evaluation based on params
-        :return: None
-        '''
-        def _toMask(anns, coco):
-            # modify ann['segmentation'] by reference
-            for ann in anns:
-                rle = coco.annToRLE(ann)
-                ann['segmentation'] = rle
-        p = self.params
-        if p.useCats:
-            gts=self.cocoGt.loadAnns(self.cocoGt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
-            dts=self.cocoDt.loadAnns(self.cocoDt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
-        else:
-            gts=self.cocoGt.loadAnns(self.cocoGt.getAnnIds(imgIds=p.imgIds))
-            dts=self.cocoDt.loadAnns(self.cocoDt.getAnnIds(imgIds=p.imgIds))
-
-        # convert ground truth to mask if iouType == 'segm'
-        if p.iouType == 'segm':
-            _toMask(gts, self.cocoGt)
-            _toMask(dts, self.cocoDt)
-        # set ignore flag
-        for gt in gts:
-            gt['ignore'] = gt['ignore'] if 'ignore' in gt else 0
-            gt['ignore'] = 'iscrowd' in gt and gt['iscrowd']
-            if p.iouType == 'keypoints':
-                gt['ignore'] = (gt['num_keypoints'] == 0) or gt['ignore']
-        self._gts = defaultdict(list)       # gt for evaluation
-        self._dts = defaultdict(list)       # dt for evaluation
-        for gt in gts:
-            self._gts[gt['image_id'], gt['category_id']].append(gt)
-        for dt in dts:
-            self._dts[dt['image_id'], dt['category_id']].append(dt)
-        self.evalImgs = defaultdict(list)   # per-image per-category evaluation results
-        self.eval     = {}                  # accumulated evaluation results
-
-    def evaluate(self):
-        '''
-        Run per image evaluation on given images and store results (a list of dict) in self.evalImgs
-        :return: None
-        '''
-        tic = time.time()
-        print('Running per image evaluation...')
-        p = self.params
-        # add backward compatibility if useSegm is specified in params
-        if not p.useSegm is None:
-            p.iouType = 'segm' if p.useSegm == 1 else 'bbox'
-            print('useSegm (deprecated) is not None. Running {} evaluation'.format(p.iouType))
-        print('Evaluate annotation type *{}*'.format(p.iouType))
-        p.imgIds = list(np.unique(p.imgIds))
-        if p.useCats:
-            p.catIds = list(np.unique(p.catIds))
-        p.maxDets = sorted(p.maxDets)
-        self.params=p
-
-        self._prepare()
-        # loop through images, area range, max detection number
-        catIds = p.catIds if p.useCats else [-1]
-
-        if p.iouType == 'segm' or p.iouType == 'bbox':
-            computeIoU = self.computeIoU
-        elif p.iouType == 'keypoints':
-            computeIoU = self.computeOks
-        self.ious = {(imgId, catId): computeIoU(imgId, catId) \
-                        for imgId in p.imgIds
-                        for catId in catIds}
-
-        evaluateImg = self.evaluateImg
-        maxDet = p.maxDets[-1]
-        self.evalImgs = [evaluateImg(imgId, catId, areaRng, maxDet)
-                 for catId in catIds
-                 for areaRng in p.areaRng
-                 for imgId in p.imgIds
-             ]
-        self._paramsEval = copy.deepcopy(self.params)
-        toc = time.time()
-        print('DONE (t={:0.2f}s).'.format(toc-tic))
-
-    def computeIoU(self, imgId, catId):
-        p = self.params
-        if p.useCats:
-            gt = self._gts[imgId,catId]
-            dt = self._dts[imgId,catId]
-        else:
-            gt = [_ for cId in p.catIds for _ in self._gts[imgId,cId]]
-            dt = [_ for cId in p.catIds for _ in self._dts[imgId,cId]]
-        if len(gt) == 0 and len(dt) ==0:
-            return []
-        inds = np.argsort([-d['score'] for d in dt], kind='mergesort')
-        dt = [dt[i] for i in inds]
-        if len(dt) > p.maxDets[-1]:
-            dt=dt[0:p.maxDets[-1]]
-
-        if p.iouType == 'segm':
-            g = [g['segmentation'] for g in gt]
-            d = [d['segmentation'] for d in dt]
-        elif p.iouType == 'bbox':
-            g = [g['bbox'] for g in gt]
-            d = [d['bbox'] for d in dt]
-        else:
-            raise Exception('unknown iouType for iou computation')
-
-        # compute iou between each dt and gt region
-        iscrowd = [int(o['iscrowd']) for o in gt]
-        ious = iou(d,g,iscrowd)
-        return ious
-
-    def computeOks(self, imgId, catId):
-        p = self.params
-        # dimention here should be Nxm
-        gts = self._gts[imgId, catId]
-        dts = self._dts[imgId, catId]
-        inds = np.argsort([-d['score'] for d in dts], kind='mergesort')
-        dts = [dts[i] for i in inds]
-        if len(dts) > p.maxDets[-1]:
-            dts = dts[0:p.maxDets[-1]]
-        # if len(gts) == 0 and len(dts) == 0:
-        if len(gts) == 0 or len(dts) == 0:
-            return []
-        ious = np.zeros((len(dts), len(gts)))
-        sigmas = np.array([.26, .25, .25, .35, .35, .79, .79, .72, .72, .62,.62, 1.07, 1.07, .87, .87, .89, .89])/10.0
-        vars = (sigmas * 2)**2
-        k = len(sigmas)
-        # compute oks between each detection and ground truth object
-        for j, gt in enumerate(gts):
-            # create bounds for ignore regions(double the gt bbox)
-            g = np.array(gt['keypoints'])
-            xg = g[0::3]; yg = g[1::3]; vg = g[2::3]
-            k1 = np.count_nonzero(vg > 0)
-            bb = gt['bbox']
-            x0 = bb[0] - bb[2]; x1 = bb[0] + bb[2] * 2
-            y0 = bb[1] - bb[3]; y1 = bb[1] + bb[3] * 2
-            for i, dt in enumerate(dts):
-                d = np.array(dt['keypoints'])
-                xd = d[0::3]; yd = d[1::3]
-                if k1>0:
-                    # measure the per-keypoint distance if keypoints visible
-                    dx = xd - xg
-                    dy = yd - yg
-                else:
-                    # measure minimum distance to keypoints in (x0,y0) & (x1,y1)
-                    z = np.zeros((k))
-                    dx = np.max((z, x0-xd),axis=0)+np.max((z, xd-x1),axis=0)
-                    dy = np.max((z, y0-yd),axis=0)+np.max((z, yd-y1),axis=0)
-                e = (dx**2 + dy**2) / vars / (gt['area']+np.spacing(1)) / 2
-                if k1 > 0:
-                    e=e[vg > 0]
-                ious[i, j] = np.sum(np.exp(-e)) / e.shape[0]
-        return ious
-
-    def evaluateImg(self, imgId, catId, aRng, maxDet):
-        '''
-        perform evaluation for single category and image
-        :return: dict (single image results)
-        '''
-        p = self.params
-        if p.useCats:
-            gt = self._gts[imgId,catId]
-            dt = self._dts[imgId,catId]
-        else:
-            gt = [_ for cId in p.catIds for _ in self._gts[imgId,cId]]
-            dt = [_ for cId in p.catIds for _ in self._dts[imgId,cId]]
-        if len(gt) == 0 and len(dt) ==0:
-            return None
-
-        for g in gt:
-            if g['ignore'] or (g['area']<aRng[0] or g['area']>aRng[1]):
-                g['_ignore'] = 1
-            else:
-                g['_ignore'] = 0
-
-        # sort dt highest score first, sort gt ignore last
-        gtind = np.argsort([g['_ignore'] for g in gt], kind='mergesort')
-        gt = [gt[i] for i in gtind]
-        dtind = np.argsort([-d['score'] for d in dt], kind='mergesort')
-        dt = [dt[i] for i in dtind[0:maxDet]]
-        iscrowd = [int(o['iscrowd']) for o in gt]
-        # load computed ious
-        ious = self.ious[imgId, catId][:, gtind] if len(self.ious[imgId, catId]) > 0 else self.ious[imgId, catId]
-
-        T = len(p.iouThrs)
-        G = len(gt)
-        D = len(dt)
-        gtm  = np.zeros((T,G))
-        dtm  = np.zeros((T,D))
-        gtIg = np.array([g['_ignore'] for g in gt])
-        dtIg = np.zeros((T,D))
-        if not len(ious)==0:
-            for tind, t in enumerate(p.iouThrs):
-                for dind, d in enumerate(dt):
-                    # information about best match so far (m=-1 -> unmatched)
-                    iou = min([t,1-1e-10])
-                    m   = -1
-                    for gind, g in enumerate(gt):
-                        # if this gt already matched, and not a crowd, continue
-                        if gtm[tind,gind]>0 and not iscrowd[gind]:
-                            continue
-                        # if dt matched to reg gt, and on ignore gt, stop
-                        if m>-1 and gtIg[m]==0 and gtIg[gind]==1:
-                            break
-                        # continue to next gt unless better match made
-                        if ious[dind,gind] < iou:
-                            continue
-                        # if match successful and best so far, store appropriately
-                        iou=ious[dind,gind]
-                        m=gind
-                    # if match made store id of match for both dt and gt
-                    if m ==-1:
-                        continue
-                    dtIg[tind,dind] = gtIg[m]
-                    dtm[tind,dind]  = gt[m]['id']
-                    gtm[tind,m]     = d['id']
-        # set unmatched detections outside of area range to ignore
-        a = np.array([d['area']<aRng[0] or d['area']>aRng[1] for d in dt]).reshape((1, len(dt)))
-        dtIg = np.logical_or(dtIg, np.logical_and(dtm==0, np.repeat(a,T,0)))
-        # store results for given image and category
-        return {
-                'image_id':     imgId,
-                'category_id':  catId,
-                'aRng':         aRng,
-                'maxDet':       maxDet,
-                'dtIds':        [d['id'] for d in dt],
-                'gtIds':        [g['id'] for g in gt],
-                'dtMatches':    dtm,
-                'gtMatches':    gtm,
-                'dtScores':     [d['score'] for d in dt],
-                'gtIgnore':     gtIg,
-                'dtIgnore':     dtIg,
-            }
-
-    def accumulate(self, p = None):
-        '''
-        Accumulate per image evaluation results and store the result in self.eval
-        :param p: input params for evaluation
-        :return: None
-        '''
-        print('Accumulating evaluation results...')
-        tic = time.time()
-        if not self.evalImgs:
-            print('Please run evaluate() first')
-        # allows input customized parameters
-        if p is None:
-            p = self.params
-        p.catIds = p.catIds if p.useCats == 1 else [-1]
-        T           = len(p.iouThrs)
-        R           = len(p.recThrs)
-        K           = len(p.catIds) if p.useCats else 1
-        A           = len(p.areaRng)
-        M           = len(p.maxDets)
-        precision   = -np.ones((T,R,K,A,M)) # -1 for the precision of absent categories
-        recall      = -np.ones((T,K,A,M))
-
-        # create dictionary for future indexing
-        _pe = self._paramsEval
-        catIds = _pe.catIds if _pe.useCats else [-1]
-        setK = set(catIds)
-        setA = set(map(tuple, _pe.areaRng))
-        setM = set(_pe.maxDets)
-        setI = set(_pe.imgIds)
-        # get inds to evaluate
-        k_list = [n for n, k in enumerate(p.catIds)  if k in setK]
-        m_list = [m for n, m in enumerate(p.maxDets) if m in setM]
-        a_list = [n for n, a in enumerate(map(lambda x: tuple(x), p.areaRng)) if a in setA]
-        i_list = [n for n, i in enumerate(p.imgIds)  if i in setI]
-        I0 = len(_pe.imgIds)
-        A0 = len(_pe.areaRng)
-        # retrieve E at each category, area range, and max number of detections
-        for k, k0 in enumerate(k_list):
-            Nk = k0*A0*I0
-            for a, a0 in enumerate(a_list):
-                Na = a0*I0
-                for m, maxDet in enumerate(m_list):
-                    E = [self.evalImgs[Nk + Na + i] for i in i_list]
-                    E = [e for e in E if not e is None]
-                    if len(E) == 0:
-                        continue
-                    dtScores = np.concatenate([e['dtScores'][0:maxDet] for e in E])
-
-                    # different sorting method generates slightly different results.
-                    # mergesort is used to be consistent as Matlab implementation.
-                    inds = np.argsort(-dtScores, kind='mergesort')
-
-                    dtm  = np.concatenate([e['dtMatches'][:,0:maxDet] for e in E], axis=1)[:,inds]
-                    dtIg = np.concatenate([e['dtIgnore'][:,0:maxDet]  for e in E], axis=1)[:,inds]
-                    gtIg = np.concatenate([e['gtIgnore'] for e in E])
-                    npig = np.count_nonzero(gtIg==0 )
-                    if npig == 0:
-                        continue
-                    tps = np.logical_and(               dtm,  np.logical_not(dtIg) )
-                    fps = np.logical_and(np.logical_not(dtm), np.logical_not(dtIg) )
-
-                    tp_sum = np.cumsum(tps, axis=1).astype(dtype=np.float)
-                    fp_sum = np.cumsum(fps, axis=1).astype(dtype=np.float)
-                    for t, (tp, fp) in enumerate(zip(tp_sum, fp_sum)):
-                        tp = np.array(tp)
-                        fp = np.array(fp)
-                        nd = len(tp)
-                        rc = tp / npig
-                        pr = tp / (fp+tp+np.spacing(1))
-                        q  = np.zeros((R,))
-
-                        if nd:
-                            recall[t,k,a,m] = rc[-1]
-                        else:
-                            recall[t,k,a,m] = 0
-
-                        # numpy is slow without cython optimization for accessing elements
-                        # use python array gets significant speed improvement
-                        pr = pr.tolist(); q = q.tolist()
-
-                        for i in range(nd-1, 0, -1):
-                            if pr[i] > pr[i-1]:
-                                pr[i-1] = pr[i]
-
-                        inds = np.searchsorted(rc, p.recThrs, side='left')
-                        try:
-                            for ri, pi in enumerate(inds):
-                                q[ri] = pr[pi]
-                        except:
-                            pass
-                        precision[t,:,k,a,m] = np.array(q)
-        self.eval = {
-            'params': p,
-            'counts': [T, R, K, A, M],
-            'date': datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
-            'precision': precision,
-            'recall':   recall,
-        }
-        toc = time.time()
-        print('DONE (t={:0.2f}s).'.format( toc-tic))
-
-    def summarize(self):
-        '''
-        Compute and display summary metrics for evaluation results.
-        Note this functin can *only* be applied on the default parameter setting
-        '''
-        def _summarize( ap=1, iouThr=None, areaRng='all', maxDets=100 ):
-            p = self.params
-            iStr = ' {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}'
-            titleStr = 'Average Precision' if ap == 1 else 'Average Recall'
-            typeStr = '(AP)' if ap==1 else '(AR)'
-            iouStr = '{:0.2f}:{:0.2f}'.format(p.iouThrs[0], p.iouThrs[-1]) \
-                if iouThr is None else '{:0.2f}'.format(iouThr)
-
-            aind = [i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng]
-            mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets]
-            if ap == 1:
-                # dimension of precision: [TxRxKxAxM]
-                s = self.eval['precision']
-                # IoU
-                if iouThr is not None:
-                    t = np.where(iouThr == p.iouThrs)[0]
-                    s = s[t]
-                s = s[:,:,:,aind,mind]
-            else:
-                # dimension of recall: [TxKxAxM]
-                s = self.eval['recall']
-                if iouThr is not None:
-                    t = np.where(iouThr == p.iouThrs)[0]
-                    s = s[t]
-                s = s[:,:,aind,mind]
-            if len(s[s>-1])==0:
-                mean_s = -1
-            else:
-                mean_s = np.mean(s[s>-1])
-            print(iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets, mean_s))
-            return mean_s
-        def _summarizeDets():
-            stats = np.zeros((12,))
-            stats[0] = _summarize(1)
-            stats[1] = _summarize(1, iouThr=.5, maxDets=self.params.maxDets[2])
-            stats[2] = _summarize(1, iouThr=.75, maxDets=self.params.maxDets[2])
-            stats[3] = _summarize(1, areaRng='small', maxDets=self.params.maxDets[2])
-            stats[4] = _summarize(1, areaRng='medium', maxDets=self.params.maxDets[2])
-            stats[5] = _summarize(1, areaRng='large', maxDets=self.params.maxDets[2])
-            stats[6] = _summarize(0, maxDets=self.params.maxDets[0])
-            stats[7] = _summarize(0, maxDets=self.params.maxDets[1])
-            stats[8] = _summarize(0, maxDets=self.params.maxDets[2])
-            stats[9] = _summarize(0, areaRng='small', maxDets=self.params.maxDets[2])
-            stats[10] = _summarize(0, areaRng='medium', maxDets=self.params.maxDets[2])
-            stats[11] = _summarize(0, areaRng='large', maxDets=self.params.maxDets[2])
-            return stats
-        def _summarizeKps():
-            stats = np.zeros((10,))
-            stats[0] = _summarize(1, maxDets=20)
-            stats[1] = _summarize(1, maxDets=20, iouThr=.5)
-            stats[2] = _summarize(1, maxDets=20, iouThr=.75)
-            stats[3] = _summarize(1, maxDets=20, areaRng='medium')
-            stats[4] = _summarize(1, maxDets=20, areaRng='large')
-            stats[5] = _summarize(0, maxDets=20)
-            stats[6] = _summarize(0, maxDets=20, iouThr=.5)
-            stats[7] = _summarize(0, maxDets=20, iouThr=.75)
-            stats[8] = _summarize(0, maxDets=20, areaRng='medium')
-            stats[9] = _summarize(0, maxDets=20, areaRng='large')
-            return stats
-        if not self.eval:
-            raise Exception('Please run accumulate() first')
-        iouType = self.params.iouType
-        if iouType == 'segm' or iouType == 'bbox':
-            summarize = _summarizeDets
-        elif iouType == 'keypoints':
-            summarize = _summarizeKps
-        self.stats = summarize()
-
-    def __str__(self):
-        self.summarize()
-
-class Params:
-    '''
-    Params for coco evaluation api
-    '''
-    def setDetParams(self):
-        self.imgIds = []
-        self.catIds = []
-        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
-        self.iouThrs = np.linspace(.5, 0.95, np.round((0.95 - .5) / .05) + 1, endpoint=True)
-        self.recThrs = np.linspace(.0, 1.00, np.round((1.00 - .0) / .01) + 1, endpoint=True)
-        self.maxDets = [1, 10, 100]
-        self.areaRng = [[0 ** 2, 1e5 ** 2], [0 ** 2, 32 ** 2], [32 ** 2, 96 ** 2], [96 ** 2, 1e5 ** 2]]
-        self.areaRngLbl = ['all', 'small', 'medium', 'large']
-        self.useCats = 1
-
-    def setKpParams(self):
-        self.imgIds = []
-        self.catIds = []
-        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
-        self.iouThrs = np.linspace(.5, 0.95, np.round((0.95 - .5) / .05) + 1, endpoint=True)
-        self.recThrs = np.linspace(.0, 1.00, np.round((1.00 - .0) / .01) + 1, endpoint=True)
-        self.maxDets = [20]
-        self.areaRng = [[0 ** 2, 1e5 ** 2], [32 ** 2, 96 ** 2], [96 ** 2, 1e5 ** 2]]
-        self.areaRngLbl = ['all', 'medium', 'large']
-        self.useCats = 1
-
-    def __init__(self, iouType='segm'):
-        if iouType == 'segm' or iouType == 'bbox':
-            self.setDetParams()
-        elif iouType == 'keypoints':
-            self.setKpParams()
-        else:
-            raise Exception('iouType not supported')
-        self.iouType = iouType
-        # useSegm is deprecated
-        self.useSegm = None
diff --git a/RetinaFace/rcnn/pycocotools/setup.py b/RetinaFace/rcnn/pycocotools/setup.py
deleted file mode 100644
index 5e836f1b4..000000000
--- a/RetinaFace/rcnn/pycocotools/setup.py
+++ /dev/null
@@ -1,20 +0,0 @@
-from distutils.core import setup
-from Cython.Build import cythonize
-from distutils.extension import Extension
-import numpy as np
-
-# To compile and install locally run "python setup.py build_ext --inplace"
-# To install library to Python site-packages run "python setup.py build_ext install"
-
-ext_modules = [
-    Extension(
-        '_mask',
-        sources=['maskApi.c', '_mask.pyx'],
-        include_dirs=[np.get_include()],
-        extra_compile_args=['-Wno-cpp', '-Wno-unused-function', '-std=c99'],
-    )
-]
-
-setup(name='pycocotools',
-      ext_modules=cythonize(ext_modules)
-)
diff --git a/RetinaFace/rcnn/sample_config.py b/RetinaFace/rcnn/sample_config.py
deleted file mode 100644
index e07f77c10..000000000
--- a/RetinaFace/rcnn/sample_config.py
+++ /dev/null
@@ -1,304 +0,0 @@
-import numpy as np
-from easydict import EasyDict as edict
-
-config = edict()
-
-# network related params
-config.PIXEL_MEANS = np.array([103.939, 116.779, 123.68])
-config.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
-config.PIXEL_SCALE = 1.0
-config.IMAGE_STRIDE = 0
-
-# dataset related params
-config.NUM_CLASSES = 2
-config.PRE_SCALES = [(1200, 1600)]  # first is scale (the shorter side); second is max size
-config.SCALES = [(640, 640)]  # first is scale (the shorter side); second is max size
-#config.SCALES = [(800, 800)]  # first is scale (the shorter side); second is max size
-config.ORIGIN_SCALE = False
-
-_ratio = (1.,)
-
-RAC_SSH = {
-    '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-}
-
-_ratio = (1.,1.5)
-RAC_SSH2 = {
-    '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-}
-
-_ratio = (1.,1.5)
-RAC_SSH3 = {
-    '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-    '4': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-}
-
-RAC_RETINA = {}
-_ratios = (1.0,)
-_ass = 2.0**(1.0/3)
-_basescale = 1.0
-for _stride in [4, 8, 16, 32, 64]:
-  key = str(_stride)
-  value = {'BASE_SIZE': 16, 'RATIOS': _ratios, 'ALLOWED_BORDER': 9999}
-  scales = []
-  for _ in range(3):
-    scales.append(_basescale)
-    _basescale *= _ass
-  value['SCALES'] = tuple(scales)
-  RAC_RETINA[key] = value
-
-
-config.RPN_ANCHOR_CFG = RAC_SSH #default
-
-config.NET_MODE = 2
-config.HEAD_MODULE = 'SSH'
-#config.HEAD_MODULE = 'RF'
-config.LR_MODE = 0
-config.LANDMARK_LR_MULT = 2.0
-config.HEAD_FILTER_NUM = 256
-config.CONTEXT_FILTER_RATIO = 1
-config.max_feat_channel = 9999
-
-config.USE_CROP = True
-config.USE_FPN = True
-config.USE_DCN = 0
-config.FACE_LANDMARK = True
-config.USE_OCCLUSION = False
-config.USE_BLUR = False
-config.MORE_SMALL_BOX = True
-
-config.LAYER_FIX = False
-
-config.CASCADE = 0
-config.CASCADE_MODE = 1
-#config.CASCADE_CLS_STRIDES = [16,8,4]
-#config.CASCADE_BBOX_STRIDES = [64,32]
-config.CASCADE_CLS_STRIDES = [64,32,16,8,4]
-config.CASCADE_BBOX_STRIDES = [64,32,16,8,4]
-#config.CASCADE_BBOX_STRIDES = [64,32,16,8]
-
-config.HEAD_BOX = False
-config.DENSE_ANCHOR = False
-config.USE_MAXOUT = 0
-config.SHARE_WEIGHT_BBOX = False
-config.SHARE_WEIGHT_LANDMARK = False
-
-config.RANDOM_FEAT_STRIDE = False
-config.NUM_CPU = 4
-config.MIXUP = 0.0
-config.USE_3D = False
-
-#config.BBOX_MASK_THRESH = 0
-config.COLOR_MODE = 2
-config.COLOR_JITTERING = 0.125
-#config.COLOR_JITTERING = 0
-#config.COLOR_JITTERING = 0.2
-
-
-config.TRAIN = edict()
-
-config.TRAIN.IMAGE_ALIGN = 0
-config.TRAIN.MIN_BOX_SIZE = 0
-config.BBOX_MASK_THRESH = config.TRAIN.MIN_BOX_SIZE
-# R-CNN and RPN
-# size of images for each device, 2 for rcnn, 1 for rpn and e2e
-config.TRAIN.BATCH_IMAGES = 8
-# e2e changes behavior of anchor loader and metric
-config.TRAIN.END2END = True
-# group images with similar aspect ratio
-config.TRAIN.ASPECT_GROUPING = False
-
-# RPN anchor loader
-# rpn anchors batch size
-config.TRAIN.RPN_ENABLE_OHEM = 2
-config.TRAIN.OHEM_MODE = 1
-config.TRAIN.RPN_BATCH_SIZE = 256
-# rpn anchors sampling params
-config.TRAIN.RPN_FG_FRACTION = 0.25
-config.TRAIN.RPN_POSITIVE_OVERLAP = 0.5
-config.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3
-if config.CASCADE>0:
-    config.TRAIN.RPN_POSITIVE_OVERLAP = 0.7
-config.TRAIN.CASCADE_OVERLAP = [0.4, 0.5]
-config.TRAIN.RPN_CLOBBER_POSITIVES = False
-config.TRAIN.RPN_FORCE_POSITIVE = False
-# rpn bounding box regression params
-config.TRAIN.BBOX_STDS = (1.0, 1.0, 1.0, 1.0)
-config.TRAIN.LANDMARK_STD = 1.0
-
-
-config.TEST = edict()
-
-# R-CNN testing
-# use rpn to generate proposal
-config.TEST.HAS_RPN = False
-# size of images for each device
-config.TEST.BATCH_IMAGES = 1
-
-# RPN proposal
-config.TEST.CXX_PROPOSAL = True
-config.TEST.RPN_NMS_THRESH = 0.3
-config.TEST.RPN_PRE_NMS_TOP_N = 1000
-config.TEST.RPN_POST_NMS_TOP_N = 3000
-#config.TEST.RPN_MIN_SIZE = config.RPN_FEAT_STRIDE
-#config.TEST.RPN_MIN_SIZE = [0,0,0]
-
-# RCNN nms
-config.TEST.NMS = 0.3
-
-config.TEST.SCORE_THRESH = 0.05
-config.TEST.IOU_THRESH = 0.5
-
-
-# network settings
-network = edict()
-
-network.ssh = edict()
-
-network.mnet = edict()
-#network.mnet.pretrained = 'model/mnasnet'
-#network.mnet.pretrained = 'model/mobilenetv2_0_5'
-#network.mnet.pretrained = 'model/mobilenet_0_5'
-#network.mnet.MULTIPLIER = 0.5
-#network.mnet.pretrained = 'model/mobilenet_0_25'
-#network.mnet.pretrained_epoch = 0
-#network.mnet.PIXEL_MEANS = np.array([0.406, 0.456, 0.485])
-#network.mnet.PIXEL_STDS = np.array([0.225, 0.224, 0.229])
-#network.mnet.PIXEL_SCALE = 255.0
-network.mnet.FIXED_PARAMS = ['^stage1', '^.*upsampling']
-network.mnet.BATCH_IMAGES = 16
-network.mnet.HEAD_FILTER_NUM = 64
-network.mnet.CONTEXT_FILTER_RATIO = 1
-
-network.mnet.PIXEL_MEANS = np.array([0.0, 0.0, 0.0])
-network.mnet.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
-network.mnet.PIXEL_SCALE = 1.0
-#network.mnet.pretrained = 'model/mobilenetfd_0_25' #78
-#network.mnet.pretrained = 'model/mobilenetfd2' #75
-network.mnet.pretrained = 'model/mobilenet025fd0' #78
-#network.mnet.pretrained = 'model/mobilenet025fd1' #75
-#network.mnet.pretrained = 'model/mobilenet025fd2' #
-network.mnet.pretrained_epoch = 0
-network.mnet.max_feat_channel = 8888
-network.mnet.COLOR_MODE = 1
-network.mnet.USE_CROP = True
-network.mnet.RPN_ANCHOR_CFG = RAC_SSH
-network.mnet.LAYER_FIX = True
-network.mnet.LANDMARK_LR_MULT = 2.5
-
-
-network.resnet = edict()
-#network.resnet.pretrained = 'model/ResNet50_v1d'
-#network.resnet.pretrained = 'model/resnet-50'
-network.resnet.pretrained = 'model/resnet-152'
-#network.resnet.pretrained = 'model/senet154'
-#network.resnet.pretrained = 'model/densenet161'
-network.resnet.pretrained_epoch = 0
-#network.mnet.PIXEL_MEANS = np.array([103.939, 116.779, 123.68])
-#network.mnet.PIXEL_STDS = np.array([57.375, 57.12, 58.393])
-#network.resnet.PIXEL_MEANS = np.array([0.406, 0.456, 0.485])
-#network.resnet.PIXEL_STDS = np.array([0.225, 0.224, 0.229])
-#network.resnet.PIXEL_SCALE = 255.0
-network.resnet.lr_step = '1,2,3,4,5,55,68,80'
-network.resnet.lr = 0.001
-network.resnet.PIXEL_MEANS = np.array([0.0, 0.0, 0.0])
-network.resnet.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
-network.resnet.PIXEL_SCALE = 1.0
-network.resnet.FIXED_PARAMS = ['^stage1', '^.*upsampling']
-network.resnet.BATCH_IMAGES = 8
-network.resnet.HEAD_FILTER_NUM = 256
-network.resnet.CONTEXT_FILTER_RATIO = 1
-network.resnet.USE_DCN = 2
-network.resnet.RPN_BATCH_SIZE = 256
-network.resnet.RPN_ANCHOR_CFG = RAC_RETINA
-
-network.resnet.USE_DCN = 0
-network.resnet.pretrained = 'model/resnet-50'
-network.resnet.RPN_ANCHOR_CFG = RAC_SSH
-
-
-# dataset settings
-dataset = edict()
-
-dataset.widerface = edict()
-dataset.widerface.dataset = 'widerface'
-dataset.widerface.image_set = 'train'
-dataset.widerface.test_image_set = 'val'
-dataset.widerface.root_path = 'data'
-dataset.widerface.dataset_path = 'data/widerface'
-dataset.widerface.NUM_CLASSES = 2
-
-dataset.retinaface = edict()
-dataset.retinaface.dataset = 'retinaface'
-dataset.retinaface.image_set = 'train'
-dataset.retinaface.test_image_set = 'val'
-dataset.retinaface.root_path = 'data'
-dataset.retinaface.dataset_path = 'data/retinaface'
-dataset.retinaface.NUM_CLASSES = 2
-
-# default settings
-default = edict()
-
-config.FIXED_PARAMS = ['^conv1', '^conv2', '^conv3', '^.*upsampling']
-#config.FIXED_PARAMS = ['^.*upsampling']
-#config.FIXED_PARAMS = ['^conv1', '^conv2', '^conv3']
-#config.FIXED_PARAMS = ['^conv0', '^stage1', 'gamma', 'beta']  #for resnet
-
-# default network
-default.network = 'resnet'
-default.pretrained = 'model/resnet-152'
-#default.network = 'resnetssh'
-default.pretrained_epoch = 0
-# default dataset
-default.dataset = 'retinaface'
-default.image_set = 'train'
-default.test_image_set = 'val'
-default.root_path = 'data'
-default.dataset_path = 'data/retinaface'
-# default training
-default.frequent = 20
-default.kvstore = 'device'
-# default e2e
-default.prefix = 'model/retinaface'
-default.end_epoch = 10000
-default.lr_step = '55,68,80'
-default.lr = 0.01
-default.wd = 0.0005
-
-def generate_config(_network, _dataset):
-    for k, v in network[_network].items():
-        if k in config:
-            config[k] = v
-        elif k in default:
-            default[k] = v
-        if k in config.TRAIN:
-          config.TRAIN[k] = v
-    for k, v in dataset[_dataset].items():
-        if k in config:
-            config[k] = v
-        elif k in default:
-            default[k] = v
-        if k in config.TRAIN:
-          config.TRAIN[k] = v
-    config.network = _network
-    config.dataset = _dataset
-    config.RPN_FEAT_STRIDE = []
-    num_anchors = []
-    for k in config.RPN_ANCHOR_CFG:
-      config.RPN_FEAT_STRIDE.append( int(k) )
-      _num_anchors = len(config.RPN_ANCHOR_CFG[k]['SCALES'])*len(config.RPN_ANCHOR_CFG[k]['RATIOS'])
-      if config.DENSE_ANCHOR:
-        _num_anchors *= 2
-      config.RPN_ANCHOR_CFG[k]['NUM_ANCHORS'] = _num_anchors
-      num_anchors.append(_num_anchors)
-    config.RPN_FEAT_STRIDE = sorted(config.RPN_FEAT_STRIDE, reverse=True)
-    for j in range(1,len(num_anchors)):
-      assert num_anchors[0]==num_anchors[j]
-    config.NUM_ANCHORS = num_anchors[0]
-
diff --git a/RetinaFace/rcnn/symbol/pyramidbox.py b/RetinaFace/rcnn/symbol/pyramidbox.py
deleted file mode 100644
index dae2537ea..000000000
--- a/RetinaFace/rcnn/symbol/pyramidbox.py
+++ /dev/null
@@ -1,427 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import numpy as np
-import six
-import paddle.fluid as fluid
-from paddle.fluid.param_attr import ParamAttr
-from paddle.fluid.initializer import Xavier
-from paddle.fluid.initializer import Constant
-from paddle.fluid.initializer import Bilinear
-from paddle.fluid.regularizer import L2Decay
-
-
-def conv_bn(input, filter, ksize, stride, padding, act='relu', bias_attr=False):
-    conv = fluid.layers.conv2d(
-        input=input,
-        filter_size=ksize,
-        num_filters=filter,
-        stride=stride,
-        padding=padding,
-        act=None,
-        bias_attr=bias_attr)
-    return fluid.layers.batch_norm(input=conv, act=act)
-
-
-def conv_block(input, groups, filters, ksizes, strides=None, with_pool=True):
-    assert len(filters) == groups
-    assert len(ksizes) == groups
-    strides = [1] * groups if strides is None else strides
-    w_attr = ParamAttr(learning_rate=1., initializer=Xavier())
-    b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
-    conv = input
-    for i in six.moves.xrange(groups):
-        conv = fluid.layers.conv2d(
-            input=conv,
-            num_filters=filters[i],
-            filter_size=ksizes[i],
-            stride=strides[i],
-            padding=(ksizes[i] - 1) // 2,
-            param_attr=w_attr,
-            bias_attr=b_attr,
-            act='relu')
-    if with_pool:
-        pool = fluid.layers.pool2d(
-            input=conv,
-            pool_size=2,
-            pool_type='max',
-            pool_stride=2,
-            ceil_mode=True)
-        return conv, pool
-    else:
-        return conv
-
-
-class PyramidBox(object):
-    def __init__(self,
-                 data_shape,
-                 num_classes=None,
-                 use_transposed_conv2d=True,
-                 is_infer=False,
-                 sub_network=False):
-        """
-        TODO(qingqing): add comments.
-        """
-        self.data_shape = data_shape
-        self.min_sizes = [16., 32., 64., 128., 256., 512.]
-        self.steps = [4., 8., 16., 32., 64., 128.]
-        self.num_classes = num_classes
-        self.use_transposed_conv2d = use_transposed_conv2d
-        self.is_infer = is_infer
-        self.sub_network = sub_network
-
-        # the base network is VGG with atrous layers
-        self._input()
-        self._vgg()
-        if sub_network:
-            self._low_level_fpn()
-            self._cpm_module()
-            self._pyramidbox()
-        else:
-            self._vgg_ssd()
-
-    def feeds(self):
-        if self.is_infer:
-            return [self.image]
-        else:
-            return [self.image, self.face_box, self.head_box, self.gt_label]
-
-    def _input(self):
-        self.image = fluid.layers.data(
-            name='image', shape=self.data_shape, dtype='float32')
-        if not self.is_infer:
-            self.face_box = fluid.layers.data(
-                name='face_box', shape=[4], dtype='float32', lod_level=1)
-            self.head_box = fluid.layers.data(
-                name='head_box', shape=[4], dtype='float32', lod_level=1)
-            self.gt_label = fluid.layers.data(
-                name='gt_label', shape=[1], dtype='int32', lod_level=1)
-
-    def _vgg(self):
-        self.conv1, self.pool1 = conv_block(self.image, 2, [64] * 2, [3] * 2)
-        self.conv2, self.pool2 = conv_block(self.pool1, 2, [128] * 2, [3] * 2)
-
-        #priorbox min_size is 16
-        self.conv3, self.pool3 = conv_block(self.pool2, 3, [256] * 3, [3] * 3)
-        #priorbox min_size is 32
-        self.conv4, self.pool4 = conv_block(self.pool3, 3, [512] * 3, [3] * 3)
-        #priorbox min_size is 64
-        self.conv5, self.pool5 = conv_block(self.pool4, 3, [512] * 3, [3] * 3)
-
-        # fc6 and fc7 in paper, priorbox min_size is 128
-        self.conv6 = conv_block(
-            self.pool5, 2, [1024, 1024], [3, 1], with_pool=False)
-        # conv6_1 and conv6_2 in paper, priorbox min_size is 256
-        self.conv7 = conv_block(
-            self.conv6, 2, [256, 512], [1, 3], [1, 2], with_pool=False)
-        # conv7_1 and conv7_2 in paper, priorbox mini_size is 512
-        self.conv8 = conv_block(
-            self.conv7, 2, [128, 256], [1, 3], [1, 2], with_pool=False)
-
-    def _low_level_fpn(self):
-        """
-        Low-level feature pyramid network.
-        """
-
-        def fpn(up_from, up_to):
-            ch = up_to.shape[1]
-            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
-            conv1 = fluid.layers.conv2d(
-                up_from, ch, 1, act='relu', bias_attr=b_attr)
-            if self.use_transposed_conv2d:
-                w_attr = ParamAttr(
-                    learning_rate=0.,
-                    regularizer=L2Decay(0.),
-                    initializer=Bilinear())
-                upsampling = fluid.layers.conv2d_transpose(
-                    conv1,
-                    ch,
-                    output_size=None,
-                    filter_size=4,
-                    padding=1,
-                    stride=2,
-                    groups=ch,
-                    param_attr=w_attr,
-                    bias_attr=False,
-                    use_cudnn=True)
-            else:
-                upsampling = fluid.layers.resize_bilinear(
-                    conv1, out_shape=up_to.shape[2:])
-
-            conv2 = fluid.layers.conv2d(
-                up_to, ch, 1, act='relu', bias_attr=b_attr)
-            if self.is_infer:
-                upsampling = fluid.layers.crop(upsampling, shape=conv2)
-            # eltwise mul
-            conv_fuse = upsampling * conv2
-            return conv_fuse
-
-        self.lfpn2_on_conv5 = fpn(self.conv6, self.conv5)
-        self.lfpn1_on_conv4 = fpn(self.lfpn2_on_conv5, self.conv4)
-        self.lfpn0_on_conv3 = fpn(self.lfpn1_on_conv4, self.conv3)
-
-    def _cpm_module(self):
-        """
-        Context-sensitive Prediction Module 
-        """
-
-        def cpm(input):
-            # residual
-            branch1 = conv_bn(input, 1024, 1, 1, 0, None)
-            branch2a = conv_bn(input, 256, 1, 1, 0, act='relu')
-            branch2b = conv_bn(branch2a, 256, 3, 1, 1, act='relu')
-            branch2c = conv_bn(branch2b, 1024, 1, 1, 0, None)
-            sum = branch1 + branch2c
-            rescomb = fluid.layers.relu(x=sum)
-
-            # ssh
-            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
-            ssh_1 = fluid.layers.conv2d(rescomb, 256, 3, 1, 1, bias_attr=b_attr)
-            ssh_dimred = fluid.layers.conv2d(
-                rescomb, 128, 3, 1, 1, act='relu', bias_attr=b_attr)
-            ssh_2 = fluid.layers.conv2d(
-                ssh_dimred, 128, 3, 1, 1, bias_attr=b_attr)
-            ssh_3a = fluid.layers.conv2d(
-                ssh_dimred, 128, 3, 1, 1, act='relu', bias_attr=b_attr)
-            ssh_3b = fluid.layers.conv2d(ssh_3a, 128, 3, 1, 1, bias_attr=b_attr)
-
-            ssh_concat = fluid.layers.concat([ssh_1, ssh_2, ssh_3b], axis=1)
-            ssh_out = fluid.layers.relu(x=ssh_concat)
-            return ssh_out
-
-        self.ssh_conv3 = cpm(self.lfpn0_on_conv3)
-        self.ssh_conv4 = cpm(self.lfpn1_on_conv4)
-        self.ssh_conv5 = cpm(self.lfpn2_on_conv5)
-        self.ssh_conv6 = cpm(self.conv6)
-        self.ssh_conv7 = cpm(self.conv7)
-        self.ssh_conv8 = cpm(self.conv8)
-
-    def _l2_norm_scale(self, input, init_scale=1.0, channel_shared=False):
-        from paddle.fluid.layer_helper import LayerHelper
-        helper = LayerHelper("Scale")
-        l2_norm = fluid.layers.l2_normalize(
-            input, axis=1)  # l2 norm along channel
-        shape = [1] if channel_shared else [input.shape[1]]
-        scale = helper.create_parameter(
-            attr=helper.param_attr,
-            shape=shape,
-            dtype=input.dtype,
-            default_initializer=Constant(init_scale))
-        out = fluid.layers.elementwise_mul(
-            x=l2_norm, y=scale, axis=-1 if channel_shared else 1)
-        return out
-
-    def _pyramidbox(self):
-        """
-        Get prior-boxes and pyramid-box
-        """
-        self.ssh_conv3_norm = self._l2_norm_scale(
-            self.ssh_conv3, init_scale=10.)
-        self.ssh_conv4_norm = self._l2_norm_scale(self.ssh_conv4, init_scale=8.)
-        self.ssh_conv5_norm = self._l2_norm_scale(self.ssh_conv5, init_scale=5.)
-
-        def permute_and_reshape(input, last_dim):
-            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
-            compile_shape = [
-                trans.shape[0], np.prod(trans.shape[1:]) // last_dim, last_dim
-            ]
-            run_shape = fluid.layers.assign(
-                np.array([0, -1, last_dim]).astype("int32"))
-            return fluid.layers.reshape(
-                trans, shape=compile_shape, actual_shape=run_shape)
-
-        face_locs, face_confs = [], []
-        head_locs, head_confs = [], []
-        boxes, vars = [], []
-        inputs = [
-            self.ssh_conv3_norm, self.ssh_conv4_norm, self.ssh_conv5_norm,
-            self.ssh_conv6, self.ssh_conv7, self.ssh_conv8
-        ]
-        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
-        for i, input in enumerate(inputs):
-            mbox_loc = fluid.layers.conv2d(input, 8, 3, 1, 1, bias_attr=b_attr)
-            face_loc, head_loc = fluid.layers.split(
-                mbox_loc, num_or_sections=2, dim=1)
-            face_loc = permute_and_reshape(face_loc, 4)
-            head_loc = permute_and_reshape(head_loc, 4)
-
-            mbox_conf = fluid.layers.conv2d(input, 6, 3, 1, 1, bias_attr=b_attr)
-            face_conf1, face_conf3, head_conf = fluid.layers.split(
-                mbox_conf, num_or_sections=[1, 3, 2], dim=1)
-            face_conf3_maxin = fluid.layers.reduce_max(
-                face_conf3, dim=1, keep_dim=True)
-            face_conf = fluid.layers.concat(
-                [face_conf1, face_conf3_maxin], axis=1)
-
-            face_conf = permute_and_reshape(face_conf, 2)
-            head_conf = permute_and_reshape(head_conf, 2)
-
-            face_locs.append(face_loc)
-            face_confs.append(face_conf)
-
-            head_locs.append(head_loc)
-            head_confs.append(head_conf)
-
-            box, var = fluid.layers.prior_box(
-                input,
-                self.image,
-                min_sizes=[self.min_sizes[i]],
-                steps=[self.steps[i]] * 2,
-                aspect_ratios=[1.],
-                clip=False,
-                flip=True,
-                offset=0.5)
-            box = fluid.layers.reshape(box, shape=[-1, 4])
-            var = fluid.layers.reshape(var, shape=[-1, 4])
-
-            boxes.append(box)
-            vars.append(var)
-
-        self.face_mbox_loc = fluid.layers.concat(face_locs, axis=1)
-        self.face_mbox_conf = fluid.layers.concat(face_confs, axis=1)
-
-        self.head_mbox_loc = fluid.layers.concat(head_locs, axis=1)
-        self.head_mbox_conf = fluid.layers.concat(head_confs, axis=1)
-
-        self.prior_boxes = fluid.layers.concat(boxes)
-        self.box_vars = fluid.layers.concat(vars)
-
-    def _vgg_ssd(self):
-        self.conv3_norm = self._l2_norm_scale(self.conv3, init_scale=10.)
-        self.conv4_norm = self._l2_norm_scale(self.conv4, init_scale=8.)
-        self.conv5_norm = self._l2_norm_scale(self.conv5, init_scale=5.)
-
-        def permute_and_reshape(input, last_dim):
-            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
-            compile_shape = [
-                trans.shape[0], np.prod(trans.shape[1:]) // last_dim, last_dim
-            ]
-            run_shape = fluid.layers.assign(
-                np.array([0, -1, last_dim]).astype("int32"))
-            return fluid.layers.reshape(
-                trans, shape=compile_shape, actual_shape=run_shape)
-
-        locs, confs = [], []
-        boxes, vars = [], []
-        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
-
-        # conv3
-        mbox_loc = fluid.layers.conv2d(
-            self.conv3_norm, 4, 3, 1, 1, bias_attr=b_attr)
-        loc = permute_and_reshape(mbox_loc, 4)
-        mbox_conf = fluid.layers.conv2d(
-            self.conv3_norm, 4, 3, 1, 1, bias_attr=b_attr)
-        conf1, conf3 = fluid.layers.split(
-            mbox_conf, num_or_sections=[1, 3], dim=1)
-        conf3_maxin = fluid.layers.reduce_max(conf3, dim=1, keep_dim=True)
-        conf = fluid.layers.concat([conf1, conf3_maxin], axis=1)
-        conf = permute_and_reshape(conf, 2)
-        box, var = fluid.layers.prior_box(
-            self.conv3_norm,
-            self.image,
-            min_sizes=[16.],
-            steps=[4, 4],
-            aspect_ratios=[1.],
-            clip=False,
-            flip=True,
-            offset=0.5)
-        box = fluid.layers.reshape(box, shape=[-1, 4])
-        var = fluid.layers.reshape(var, shape=[-1, 4])
-
-        locs.append(loc)
-        confs.append(conf)
-        boxes.append(box)
-        vars.append(var)
-
-        min_sizes = [32., 64., 128., 256., 512.]
-        steps = [8., 16., 32., 64., 128.]
-        inputs = [
-            self.conv4_norm, self.conv5_norm, self.conv6, self.conv7, self.conv8
-        ]
-        for i, input in enumerate(inputs):
-            mbox_loc = fluid.layers.conv2d(input, 4, 3, 1, 1, bias_attr=b_attr)
-            loc = permute_and_reshape(mbox_loc, 4)
-
-            mbox_conf = fluid.layers.conv2d(input, 2, 3, 1, 1, bias_attr=b_attr)
-            conf = permute_and_reshape(mbox_conf, 2)
-            box, var = fluid.layers.prior_box(
-                input,
-                self.image,
-                min_sizes=[min_sizes[i]],
-                steps=[steps[i]] * 2,
-                aspect_ratios=[1.],
-                clip=False,
-                flip=True,
-                offset=0.5)
-            box = fluid.layers.reshape(box, shape=[-1, 4])
-            var = fluid.layers.reshape(var, shape=[-1, 4])
-
-            locs.append(loc)
-            confs.append(conf)
-            boxes.append(box)
-            vars.append(var)
-
-        self.face_mbox_loc = fluid.layers.concat(locs, axis=1)
-        self.face_mbox_conf = fluid.layers.concat(confs, axis=1)
-        self.prior_boxes = fluid.layers.concat(boxes)
-        self.box_vars = fluid.layers.concat(vars)
-
-    def vgg_ssd_loss(self):
-        loss = fluid.layers.ssd_loss(
-            self.face_mbox_loc,
-            self.face_mbox_conf,
-            self.face_box,
-            self.gt_label,
-            self.prior_boxes,
-            self.box_vars,
-            overlap_threshold=0.35,
-            neg_overlap=0.35)
-        loss = fluid.layers.reduce_sum(loss)
-        return loss
-
-    def train(self):
-        face_loss = fluid.layers.ssd_loss(
-            self.face_mbox_loc,
-            self.face_mbox_conf,
-            self.face_box,
-            self.gt_label,
-            self.prior_boxes,
-            self.box_vars,
-            overlap_threshold=0.35,
-            neg_overlap=0.35)
-        face_loss.persistable = True
-        head_loss = fluid.layers.ssd_loss(
-            self.head_mbox_loc,
-            self.head_mbox_conf,
-            self.head_box,
-            self.gt_label,
-            self.prior_boxes,
-            self.box_vars,
-            overlap_threshold=0.35,
-            neg_overlap=0.35)
-        head_loss.persistable = True
-        face_loss = fluid.layers.reduce_sum(face_loss)
-        face_loss.persistable = True
-        head_loss = fluid.layers.reduce_sum(head_loss)
-        head_loss.persistable = True
-        total_loss = face_loss + head_loss
-        total_loss.persistable = True
-        return face_loss, head_loss, total_loss
-
-    def infer(self, main_program=None):
-        if main_program is None:
-            test_program = fluid.default_main_program().clone(for_test=True)
-        else:
-            test_program = main_program.clone(for_test=True)
-        with fluid.program_guard(test_program):
-            face_nmsed_out = fluid.layers.detection_output(
-                self.face_mbox_loc,
-                self.face_mbox_conf,
-                self.prior_boxes,
-                self.box_vars,
-                nms_threshold=0.3,
-                nms_top_k=5000,
-                keep_top_k=750,
-                score_threshold=0.01)
-        return test_program, face_nmsed_out
diff --git a/RetinaFace/rcnn/symbol/symbol_common.py b/RetinaFace/rcnn/symbol/symbol_common.py
deleted file mode 100644
index 4db261836..000000000
--- a/RetinaFace/rcnn/symbol/symbol_common.py
+++ /dev/null
@@ -1,472 +0,0 @@
-import mxnet as mx
-import mxnet.ndarray as nd
-import numpy as np
-from rcnn.config import config
-from rcnn.PY_OP import rpn_fpn_ohem3, cascade_refine
-
-PREFIX = 'RF'
-F1 = 0
-F2 = 0
-_bwm = 1.0
-
-def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), bias_wd_mult=0.0, shared_weight=None, shared_bias = None):
-  if shared_weight is None:
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-  else:
-    weight = shared_weight
-    bias = shared_bias
-    print('reuse shared var in', name)
-  conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-      stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
-  return conv
-
-def conv_deformable(net, num_filter, num_group=1, act_type='relu',name=''):
-  if config.USE_DCN==1:
-    f = num_group*18
-    conv_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = net,
-                        num_filter=f, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-    net = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=net, offset=conv_offset,
-                        num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=num_group, stride=(1, 1), no_bias=False)
-  else:
-    print('use dcnv2 at', name)
-    lr_mult = 0.1
-    weight_var = mx.sym.Variable(name=name+'_conv2_offset_weight', init=mx.init.Zero(), lr_mult=lr_mult)
-    bias_var = mx.sym.Variable(name=name+'_conv2_offset_bias', init=mx.init.Zero(), lr_mult=lr_mult)
-    conv2_offset = mx.symbol.Convolution(name=name + '_conv2_offset', data=net, num_filter=27,
-      pad=(1, 1), kernel=(3, 3), stride=(1,1), weight=weight_var, bias=bias_var, lr_mult=lr_mult)
-    conv2_offset_t = mx.sym.slice_axis(conv2_offset, axis=1, begin=0, end=18)
-    conv2_mask =  mx.sym.slice_axis(conv2_offset, axis=1, begin=18, end=None)
-    conv2_mask = 2 * mx.sym.Activation(conv2_mask, act_type='sigmoid')
-
-    conv2 = mx.contrib.symbol.ModulatedDeformableConvolution(name=name + '_conv2', data=net, offset=conv2_offset_t, mask=conv2_mask,
-        num_filter=num_filter, pad=(1, 1), kernel=(3, 3), stride=(1,1), 
-        num_deformable_group=num_group, no_bias=True)
-    net = conv2
-  net = mx.sym.BatchNorm(data=net, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-  if len(act_type)>0:
-    net = mx.symbol.Activation(data=net, act_type=act_type, name=name+'_act')
-  return net
-
-def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
-    assert kernel[0]==3
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
-    conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
-
-    if config.USE_DCN>1 and kernel==(3,3) and pad==(1,1) and stride==(1,1) and not separable:
-      return conv_deformable(from_layer, num_filter, num_group=1, act_type = act_type, name=name)
-
-    if separable:
-      assert kernel[0]>1
-      assert filter_in>0
-    if not separable:
-      weight = mx.symbol.Variable(name="{}_weight".format(name),   
-          init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-      bias = mx.symbol.Variable(name="{}_bias".format(name),   
-          init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    else:
-      if filter_in<0:
-        filter_in = num_filter
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_sep_bn')
-      conv = mx.symbol.Activation(data=conv, act_type='relu', \
-          name="{}_sep_bn_relu".format(name))
-      conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
-          stride=(1,1), num_filter=num_filter, name="{}".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def ssh_context_module(body, num_filter, filter_in, name):
-  conv_dimred = conv_act_layer(body, name+'_conv1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=False, filter_in = filter_in)
-  conv5x5 = conv_act_layer(conv_dimred, name+'_conv2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=False)
-  conv7x7_1 = conv_act_layer(conv_dimred, name+'_conv3_1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=False)
-  conv7x7 = conv_act_layer(conv7x7_1, name+'_conv3_2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=False)
-  return (conv5x5, conv7x7)
-
-
-def ssh_detection_module(body, num_filter, filter_in, name):
-  assert num_filter%4==0
-  conv3x3 = conv_act_layer(body, name+'_conv1',
-      num_filter//2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=False, filter_in=filter_in)
-  #_filter = max(num_filter//4, 16)
-  _filter = num_filter//4
-  conv5x5, conv7x7 = ssh_context_module(body, _filter, filter_in, name+'_context')
-  ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7], dim=1, name = name+'_concat')
-  ret = mx.symbol.Activation(data=ret, act_type='relu', name=name+'_concat_relu')
-  out_filter = num_filter//2+_filter*2
-  if config.USE_DCN>0:
-    ret = conv_deformable(ret, num_filter = out_filter, name = name+'_concat_dcn')
-  return ret
-
-#def retina_context_module(body, kernel, num_filter, filter_in, name):
-#  conv_dimred = conv_act_layer(body, name+'_conv0',
-#      num_filter, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False, filter_in = filter_in)
-#  conv1 = conv_act_layer(conv_dimred, name+'_conv1',
-#      num_filter*6, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False, filter_in = filter_in)
-#  conv2 = conv_act_layer(conv1, name+'_conv2',
-#      num_filter*6, kernel=kernel, pad=((kernel[0]-1)//2, (kernel[1]-1)//2), stride=(1, 1), act_type='relu', separable=True, filter_in = num_filter*6)
-#  conv3 = conv_act_layer(conv2, name+'_conv3',
-#      num_filter, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False)
-#  conv3 = conv3 + conv_dimred
-#  return conv3
-
-def retina_detection_module(body, num_filter, filter_in, name):
-  assert num_filter%4==0
-  conv1 = conv_act_layer(body, name+'_conv1',
-      num_filter//2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=False, filter_in=filter_in)
-  conv2 = conv_act_layer(conv1, name+'_conv2',
-      num_filter//2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=False, filter_in=num_filter//2)
-  conv3 = conv_act_layer(conv2, name+'_conv3',
-      num_filter//2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=False, filter_in=num_filter//2)
-  conv4 = conv2 + conv3
-  body = mx.sym.concat(*[conv1, conv4], dim=1, name = name+'_concat')
-  if config.USE_DCN>0:
-    body = conv_deformable(body, num_filter = num_filter, name = name+'_concat_dcn')
-  return body
-
-
-def head_module(body, num_filter, filter_in, name):
-  if config.HEAD_MODULE=='SSH':
-    return ssh_detection_module(body, num_filter, filter_in, name)
-  else:
-    return retina_detection_module(body, num_filter, filter_in, name)
-
-
-def upsampling(data, num_filter, name):
-    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-    #    name=name)
-    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(2,2),  stride=(2, 2), pad=(0,0),
-    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-    #    name=name)
-    ret = mx.symbol.UpSampling(data, scale=2, sample_type='nearest', workspace=512, name=name, num_args=1)
-    return ret
-
-def get_sym_by_name(name, sym_buffer):
-  if name in sym_buffer:
-    return sym_buffer[name]
-  ret = None
-  name_key = name[0:1]
-  name_num = int(name[1:])
-  #print('getting', name, name_key, name_num)
-  if name_key=='C':
-    assert name_num%2==0
-    bottom = get_sym_by_name('C%d'%(name_num//2), sym_buffer)
-    ret = conv_act_layer(bottom, '%s_C%d'(PREFIX, name_num), 
-        F1, kernel=(3, 3), pad=(1, 1), stride=(2, 2), act_type='relu', bias_wd_mult=_bwm)
-  elif name_key=='P':
-    assert name_num%2==0
-    assert name_num<=max(config.RPN_FEAT_STRIDE)
-    lateral = get_sym_by_name('L%d'%(name_num), sym_buffer)
-    if name_num==max(config.RPN_FEAT_STRIDE) or name_num>32:
-      ret = mx.sym.identity(lateral, name='%s_P%d'%(PREFIX, name_num))
-    else:
-      bottom = get_sym_by_name('L%d'%(name_num*2), sym_buffer)
-      bottom_up = upsampling(bottom, F1, '%s_U%d'%(PREFIX, name_num))
-      if config.USE_CROP:
-        bottom_up = mx.symbol.Crop(*[bottom_up, lateral])
-      aggr = lateral + bottom_up
-      aggr = conv_act_layer(aggr, '%s_A%d'%(PREFIX, name_num),
-          F1, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      ret = mx.sym.identity(aggr, name='%s_P%d'%(PREFIX, name_num))
-  elif name_key=='L':
-    c = get_sym_by_name('C%d'%(name_num), sym_buffer)
-    #print('L', name, F1)
-    ret = conv_act_layer(c, '%s_L%d'%(PREFIX, name_num),
-        F1, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-  else:
-    raise RuntimeError('%s is not a valid sym key name'%name)
-  sym_buffer[name] = ret
-  return ret
-
-
-def get_sym_conv(data, sym):
-    all_layers = sym.get_internals()
-
-    isize = 640
-    _, out_shape, _ = all_layers.infer_shape(data = (1,3,isize,isize))
-    last_entry = None
-    c1 = None
-    c2 = None
-    c3 = None
-    c1_name = None
-    c2_name = None
-    c3_name = None
-    c1_filter = -1
-    c2_filter = -1
-    c3_filter = -1
-    #print(len(all_layers), len(out_shape))
-    #print(all_layers.__class__)
-    outputs = all_layers.list_outputs()
-    #print(outputs.__class__, len(outputs))
-    count = len(outputs)
-    stride2name = {}
-    stride2layer = {}
-    stride2shape = {}
-    for i in range(count):
-      name = outputs[i]
-      shape = out_shape[i]
-      print(i, name, count, shape)
-      if not name.endswith('_output'):
-        continue
-      if len(shape)!=4:
-        continue
-      assert isize%shape[2]==0
-      if shape[1]>config.max_feat_channel:
-        break
-      stride = isize//shape[2]
-      stride2name[stride] = name
-      stride2layer[stride] = all_layers[name]
-      stride2shape[stride] = shape
-
-    strides = sorted(stride2name.keys())
-    for stride in strides:
-      print('stride', stride, stride2name[stride], stride2shape[stride])
-    print('F1_F2', F1, F2)
-    #print('cnames', c1_name, c2_name, c3_name, F1, F2)
-    _bwm = 1.0
-    ret = {}
-    sym_buffer = {}
-    for stride in [4,8,16,32]:
-      sym_buffer['C%d'%stride] = stride2layer[stride]
-    if not config.USE_FPN:
-      for stride in config.RPN_FEAT_STRIDE:
-        name = 'L%d'%stride
-        ret[stride] = get_sym_by_name(name, sym_buffer)
-    else:
-      for stride in config.RPN_FEAT_STRIDE:
-        name = 'P%d'%stride
-        ret[stride] = get_sym_by_name(name, sym_buffer)
-
-    return ret
-
-def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0, gt_boxes=None):
-    A = config.NUM_ANCHORS
-    bbox_pred_len = 4
-    landmark_pred_len = 10
-    if config.USE_BLUR:
-      bbox_pred_len = 5
-    if config.USE_OCCLUSION:
-      landmark_pred_len = 15
-    ret_group = []
-    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
-    cls_label = mx.symbol.Variable(name='%s_label_stride%d'%(prefix,stride))
-    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d'%(prefix,stride))
-    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d'%(prefix,stride))
-    if landmark:
-      landmark_target = mx.symbol.Variable(name='%s_landmark_target_stride%d'%(prefix,stride))
-      landmark_weight = mx.symbol.Variable(name='%s_landmark_weight_stride%d'%(prefix,stride))
-    conv_feat = conv_fpn_feat[stride]
-    rpn_relu = head_module(conv_feat, F2*config.CONTEXT_FILTER_RATIO, F1, 'rf_head_stride%d'%stride)
-
-    rpn_cls_score = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d'%(prefix, stride), 2*num_anchors,
-        kernel=(1,1), pad=(0,0), stride=(1, 1))
-
-    rpn_bbox_pred = conv_only(rpn_relu, '%s_rpn_bbox_pred_stride%d'%(prefix,stride), bbox_pred_len*num_anchors,
-        kernel=(1,1), pad=(0,0), stride=(1, 1))
-
-    # prepare rpn data
-    rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                              shape=(0, 2, -1),
-                                              name="%s_rpn_cls_score_reshape_stride%s" % (prefix,stride))
-
-    rpn_bbox_pred_reshape = mx.symbol.Reshape(data=rpn_bbox_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix,stride))
-    if landmark:
-      rpn_landmark_pred = conv_only(rpn_relu, '%s_rpn_landmark_pred_stride%d'%(prefix,stride), landmark_pred_len*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1))
-      rpn_landmark_pred_reshape = mx.symbol.Reshape(data=rpn_landmark_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix,stride))
-
-    if config.TRAIN.RPN_ENABLE_OHEM>=2:
-      label, anchor_weight, pos_count = mx.sym.Custom(op_type='rpn_fpn_ohem3', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_score=rpn_cls_score_reshape, labels = cls_label)
-
-      _bbox_weight = mx.sym.tile(anchor_weight, (1,1,bbox_pred_len))
-      _bbox_weight = _bbox_weight.reshape((0, -1, A * bbox_pred_len)).transpose((0,2,1))
-      bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
-
-      if landmark:
-        _landmark_weight = mx.sym.tile(anchor_weight, (1,1,landmark_pred_len))
-        _landmark_weight = _landmark_weight.reshape((0, -1, A * landmark_pred_len)).transpose((0,2,1))
-        landmark_weight = mx.sym.elemwise_mul(landmark_weight, _landmark_weight, name='%s_landmark_weight_mul_stride%s'%(prefix,stride))
-    else:
-      label = cls_label
-      #if not config.FACE_LANDMARK:
-      #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
-      #else:
-      #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
-    #cls loss
-    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
-                                           label=label,
-                                           multi_output=True,
-                                           normalization='valid', use_ignore=True, ignore_label=-1,
-                                           grad_scale = lr_mult,
-                                           name='%s_rpn_cls_prob_stride%d'%(prefix,stride))
-    ret_group.append(rpn_cls_prob)
-    ret_group.append(mx.sym.BlockGrad(label))
-
-    pos_count = mx.symbol.sum(pos_count)
-    pos_count = pos_count + 0.001 #avoid zero
-
-    #bbox loss
-    bbox_diff = rpn_bbox_pred_reshape-bbox_target
-    bbox_diff = bbox_diff * bbox_weight
-    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_'%(prefix,stride), scalar=3.0, data=bbox_diff)
-    bbox_lr_mode0 = 0.25*lr_mult*config.TRAIN.BATCH_IMAGES / config.TRAIN.RPN_BATCH_SIZE
-    landmark_lr_mode0 = 0.4*config.LANDMARK_LR_MULT*bbox_lr_mode0
-    if config.LR_MODE==0:
-      rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d'%(prefix,stride), data=rpn_bbox_loss_, grad_scale=bbox_lr_mode0)
-    else:
-      rpn_bbox_loss_ = mx.symbol.broadcast_div(rpn_bbox_loss_, pos_count)
-      rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d'%(prefix,stride), data=rpn_bbox_loss_, grad_scale=0.5*lr_mult)
-    ret_group.append(rpn_bbox_loss)
-    ret_group.append(mx.sym.BlockGrad(bbox_weight))
-
-    #landmark loss
-    if landmark:
-      landmark_diff = rpn_landmark_pred_reshape-landmark_target
-      landmark_diff = landmark_diff * landmark_weight
-      rpn_landmark_loss_ = mx.symbol.smooth_l1(name='%s_rpn_landmark_loss_stride%d_'%(prefix,stride), scalar=3.0, data=landmark_diff)
-      if config.LR_MODE==0:
-        rpn_landmark_loss = mx.sym.MakeLoss(name='%s_rpn_landmark_loss_stride%d'%(prefix,stride), data=rpn_landmark_loss_, grad_scale=landmark_lr_mode0)
-      else:
-        rpn_landmark_loss_ = mx.symbol.broadcast_div(rpn_landmark_loss_, pos_count)
-        rpn_landmark_loss = mx.sym.MakeLoss(name='%s_rpn_landmark_loss_stride%d'%(prefix,stride), data=rpn_landmark_loss_, grad_scale=0.2*config.LANDMARK_LR_MULT*lr_mult)
-      ret_group.append(rpn_landmark_loss)
-      ret_group.append(mx.sym.BlockGrad(landmark_weight))
-    if config.USE_3D:
-      from rcnn.PY_OP import rpn_3d_mesh
-      pass
-    if config.CASCADE>0:
-      if config.CASCADE_MODE==0:
-        body = rpn_relu
-      elif config.CASCADE_MODE==1:
-        body = head_module(conv_feat, F2*config.CONTEXT_FILTER_RATIO, F1, '%s_head_stride%d_cas'%(PREFIX, stride))
-      elif config.CASCADE_MODE==2:
-        body = conv_feat + rpn_relu
-        body = head_module(body, F2*config.CONTEXT_FILTER_RATIO, F1, '%s_head_stride%d_cas'%(PREFIX,stride))
-      else:
-        body = head_module(conv_feat, F2*config.CONTEXT_FILTER_RATIO, F1, '%s_head_stride%d_cas'%(PREFIX, stride))
-        body = mx.sym.concat(body, rpn_cls_score, rpn_bbox_pred, rpn_landmark_pred, dim=1)
-
-      #cls_pred = rpn_cls_prob
-      cls_pred_t0 = rpn_cls_score_reshape
-      cls_label_raw = cls_label
-      cls_label_t0 = label
-      bbox_pred_t0 = rpn_bbox_pred_reshape
-      #bbox_pred = rpn_bbox_pred
-      #bbox_pred = mx.sym.transpose(bbox_pred, (0, 2, 3, 1))
-      #bbox_pred_len = 4
-      #bbox_pred = mx.sym.reshape(bbox_pred, (0, -1, bbox_pred_len))
-      bbox_label_t0 = bbox_target
-      #prefix = prefix+'2'
-      for casid in range(config.CASCADE):
-        #pseudo-code
-        #anchor_label = GENANCHOR(bbox_label, bbox_pred, stride)
-        #bbox_label = F(anchor_label, bbox_pred)
-        #bbox_label = bbox_label - bbox_pred
-        cls_pred = conv_only(body, '%s_rpn_cls_score_stride%d_cas%d'%(prefix, stride, casid), 2*num_anchors,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        rpn_cls_score_reshape = mx.symbol.Reshape(data=cls_pred,
-                                                  shape=(0, 2, -1),
-                                                  name="%s_rpn_cls_score_reshape_stride%s_cas%d" % (prefix,stride, casid))
-
-        #bbox_label equals to bbox_target
-        #cls_pred, cls_label, bbox_pred, bbox_label, bbox_weight, pos_count = mx.sym.Custom(op_type='cascade_refine', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_pred=cls_pred, cls_label = cls_label, bbox_pred = bbox_pred, bbox_label = bbox_label)
-        #cls_label, bbox_label, anchor_weight, pos_count = mx.sym.Custom(op_type='cascade_refine', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_pred_t0=cls_pred_t0, cls_label_t0 = cls_label_t0, cls_pred = rpn_cls_score_reshape, bbox_pred_t0 = bbox_pred_t0, bbox_label_t0 = bbox_label_t0)
-        cls_label, bbox_label, anchor_weight, pos_count = mx.sym.Custom(op_type='cascade_refine', stride=int(stride), network=config.network, 
-            dataset=config.dataset, prefix=prefix, 
-            cls_label_t0 = cls_label_t0, cls_pred_t0=cls_pred_t0, cls_pred = rpn_cls_score_reshape, 
-            bbox_pred_t0 = bbox_pred_t0, bbox_label_t0 = bbox_label_t0, 
-            cls_label_raw = cls_label_raw, cas_gt_boxes = gt_boxes)
-        if stride in config.CASCADE_CLS_STRIDES:
-          rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
-                                                 label=cls_label,
-                                                 multi_output=True,
-                                                 normalization='valid', use_ignore=True, ignore_label=-1,
-                                                 grad_scale = lr_mult,
-                                                 name='%s_rpn_cls_prob_stride%d_cas%d'%(prefix,stride,casid))
-          ret_group.append(rpn_cls_prob)
-          ret_group.append(mx.sym.BlockGrad(cls_label))
-        if stride in config.CASCADE_BBOX_STRIDES:
-          bbox_pred = conv_only(body, '%s_rpn_bbox_pred_stride%d_cas%d'%(prefix,stride,casid), bbox_pred_len*num_anchors,
-              kernel=(1,1), pad=(0,0), stride=(1, 1))
-
-          rpn_bbox_pred_reshape = mx.symbol.Reshape(data=bbox_pred,
-                                                    shape=(0, 0, -1),
-                                                    name="%s_rpn_bbox_pred_reshape_stride%s_cas%d" % (prefix,stride,casid))
-          _bbox_weight = mx.sym.tile(anchor_weight, (1,1,bbox_pred_len))
-          _bbox_weight = _bbox_weight.reshape((0, -1, A * bbox_pred_len)).transpose((0,2,1))
-          bbox_weight = _bbox_weight
-          pos_count = mx.symbol.sum(pos_count)
-          pos_count = pos_count + 0.01 #avoid zero
-          #bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
-          #bbox loss
-          bbox_diff = rpn_bbox_pred_reshape-bbox_label
-          bbox_diff = bbox_diff * bbox_weight
-          rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_cas%d'%(prefix,stride,casid), scalar=3.0, data=bbox_diff)
-          if config.LR_MODE==0:
-            rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d_cas%d'%(prefix,stride,casid), data=rpn_bbox_loss_, grad_scale=bbox_lr_mode0)
-          else:
-            rpn_bbox_loss_ = mx.symbol.broadcast_div(rpn_bbox_loss_, pos_count)
-            rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d_cas%d'%(prefix,stride,casid), data=rpn_bbox_loss_, grad_scale=0.5*lr_mult)
-          ret_group.append(rpn_bbox_loss)
-          ret_group.append(mx.sym.BlockGrad(bbox_weight))
-          #bbox_pred = rpn_bbox_pred_reshape
-
-    return ret_group
-
-def get_sym_train(sym):
-    data = mx.symbol.Variable(name="data")
-    global F1, F2
-    F1 = config.HEAD_FILTER_NUM
-    F2 = F1
-
-    # shared convolutional layers
-    conv_fpn_feat = get_sym_conv(data, sym)
-    ret_group = []
-    gt_boxes = None
-    if config.CASCADE>0:
-      gt_boxes = mx.sym.Variable('gt_boxes')
-
-
-    for stride in config.RPN_FEAT_STRIDE:
-      ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0, gt_boxes = gt_boxes)
-      ret_group += ret
-
-    return mx.sym.Group(ret_group)
-
-
diff --git a/RetinaFace/rcnn/symbol/symbol_mnet.py b/RetinaFace/rcnn/symbol/symbol_mnet.py
deleted file mode 100644
index 4a8e2b5e4..000000000
--- a/RetinaFace/rcnn/symbol/symbol_mnet.py
+++ /dev/null
@@ -1,492 +0,0 @@
-import mxnet as mx
-import mxnet.ndarray as nd
-import mxnet.gluon as gluon
-import mxnet.gluon.nn as nn
-import mxnet.autograd as ag
-import numpy as np
-from rcnn.config import config
-from rcnn.PY_OP import rpn_fpn_ohem3
-from rcnn.symbol.symbol_common import get_sym_train
-
-
-def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), bias_wd_mult=0.0, shared_weight=None, shared_bias = None):
-  if shared_weight is None:
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-  else:
-    weight = shared_weight
-    bias = shared_bias
-    print('reuse shared var in', name)
-  conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-      stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
-  return conv
-
-def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
-    assert kernel[0]==3
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
-    conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
-
-    separable = False
-    if separable:
-      assert kernel[0]==3
-    if not separable:
-      weight = mx.symbol.Variable(name="{}_weight".format(name),   
-          init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-      bias = mx.symbol.Variable(name="{}_bias".format(name),   
-          init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    else:
-      if filter_in<0:
-        filter_in = num_filter
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_sep_bn')
-      conv = mx.symbol.Activation(data=conv, act_type='relu', \
-          name="{}_sep_bn_relu".format(name))
-      conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
-          stride=(1,1), num_filter=num_filter, name="{}".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def ssh_context_module(body, num_filter, filter_in, name):
-  conv_dimred = conv_act_layer(body, name+'_conv1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=True, filter_in = filter_in)
-  conv5x5 = conv_act_layer(conv_dimred, name+'_conv2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True)
-  conv7x7_1 = conv_act_layer(conv_dimred, name+'_conv3_1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=True)
-  conv7x7 = conv_act_layer(conv7x7_1, name+'_conv3_2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True)
-  return (conv5x5, conv7x7)
-
-def ssh_detection_module(body, num_filter, filter_in, name):
-  conv3x3 = conv_act_layer(body, name+'_conv1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True, filter_in=filter_in)
-  conv5x5, conv7x7 = ssh_context_module(body, num_filter//2, filter_in, name+'_context')
-  ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7], dim=1, name = name+'_concat')
-  ret = mx.symbol.Activation(data=ret, act_type='relu', name=name+'_concat_relu')
-  return ret
-
-
-def upsampling(data, num_filter, name):
-    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-    #    name=name)
-    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(2,2),  stride=(2, 2), pad=(0,0),
-    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-    #    name=name)
-    ret = mx.symbol.UpSampling(data, scale=2, sample_type='nearest', workspace=512, name=name, num_args=1)
-    return ret
-
-def get_mnet_conv(data, sym):
-    mm = config.MULTIPLIER
-    all_layers = sym.get_internals()
-    #print(all_layers)
-    ##c1 = all_layers['mobilenetv20_features_linearbottleneck6_relu60_relu6_output'] #96
-    #c1 = all_layers['mobilenetv20_features_linearbottleneck5_elemwise_add0_output'] # 16
-    ##c2 = all_layers['mobilenetv20_features_linearbottleneck13_relu60_relu6_output']
-    #c2 = all_layers['mobilenetv20_features_linearbottleneck12_elemwise_add0_output'] # 48
-    ##c3 = all_layers['mobilenetv20_features_linearbottleneck16_batchnorm2_fwd_output'] # 160
-    #c3 = all_layers['mobilenetv20_features_linearbottleneck13_batchnorm2_fwd_output'] # 80
-    #c1_filter = int(32*mm)
-    #c2_filter = int(96*mm)
-    #c3_filter = int(160*mm)
-
-    #c1 = all_layers['mobilenet0_relu10_fwd_output']
-    #c2 = all_layers['mobilenet0_relu22_fwd_output']
-    #c3 = all_layers['mobilenet0_relu26_fwd_output']
-
-    #c1 = all_layers['conv_6_relu_output']
-    #c2 = all_layers['conv_12_relu_output']
-    #c3 = all_layers['conv_14_relu_output']
-    #c1_filter = int(256*mm)
-    #c2_filter = int(512*mm)
-    #c3_filter = int(1024*mm)
-
-    isize = 640
-    _, out_shape, _ = all_layers.infer_shape(data = (1,3,isize,isize))
-    last_entry = None
-    c1 = None
-    c2 = None
-    c3 = None
-    c1_name = None
-    c2_name = None
-    c3_name = None
-    c1_filter = -1
-    c2_filter = -1
-    c3_filter = -1
-    #print(len(all_layers), len(out_shape))
-    #print(all_layers.__class__)
-    outputs = all_layers.list_outputs()
-    #print(outputs.__class__, len(outputs))
-    count = len(outputs)
-    for i in range(count):
-      name = outputs[i]
-      shape = out_shape[i]
-      if not name.endswith('_output'):
-        continue
-      if len(shape)!=4:
-        continue
-      #print(name, shape)
-      if c1 is None and shape[2]==isize//16:
-        cname = last_entry[0]
-        #print('c1', last_entry)
-        c1 = all_layers[cname]
-        c1_name = cname
-      if c2 is None and shape[2]==isize//32:
-        cname = last_entry[0]
-        #print('c2', last_entry)
-        c2 = all_layers[cname]
-        c2_name = cname
-      if shape[2]==isize//32:
-        c3 = all_layers[name]
-        #print('c3', name, shape)
-        c3_name = name
-
-      last_entry = (name, shape)
-    print('cnames', c1_name, c2_name, c3_name)
-
-    F1 = int(256*mm)
-    F2 = int(128*mm)
-    if config.SHARE_WEIGHT_BBOX or config.SHARE_WEIGHT_LANDMARK:
-      F2 = F1
-    _bwm = 1.0
-    if config.NET_MODE==0:
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = conv_act_layer(c2, 'ssh_m2_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-      #    num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-      #    name='ssh_m2_red_upsampling')
-      #c2_up = mx.symbol.UpSampling(c2_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2_lateral, F2, 'ssh_m2_red_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-
-      c1 = c1_lateral+c2_up
-
-      c1 = conv_act_layer(c1, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==1:
-      c3_lateral = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c3_up = mx.symbol.UpSampling(c3_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c3_up = upsampling(c3_lateral, F2, 'ssh_c3_upsampling')
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-
-      c1 = c1_lateral+c2_up
-
-      c1 = conv_act_layer(c1, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==2:
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==3:
-      #c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3 = ssh_detection_module(c3, F2//2, c3_filter, 'ssh_c3_lateral')
-      #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
-      #c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = ssh_detection_module(c2, F2//2, c2_filter, 'ssh_c2_lateral')
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = ssh_detection_module(c1, F2//2, c1_filter, 'ssh_c1_lateral')
-      #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==4:
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2//2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1//2, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1//2, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==5:
-      c3 = conv_act_layer_dw(c3, 'ssh_c3_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
-      c2 = conv_act_layer_dw(c2, 'ssh_c2_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1 = conv_act_layer_dw(c1, 'ssh_c1_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-
-    return {8: m1, 16:m2, 32: m3}
-
-def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0, shared_vars = None):
-    A = config.NUM_ANCHORS
-    bbox_pred_len = 4
-    landmark_pred_len = 10
-    if config.USE_BLUR:
-      bbox_pred_len = 5
-    if config.USE_OCCLUSION:
-      landmark_pred_len = 15
-    ret_group = []
-    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
-    label = mx.symbol.Variable(name='%s_label_stride%d'%(prefix,stride))
-    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d'%(prefix,stride))
-    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d'%(prefix,stride))
-    if landmark:
-      landmark_target = mx.symbol.Variable(name='%s_landmark_target_stride%d'%(prefix,stride))
-      landmark_weight = mx.symbol.Variable(name='%s_landmark_weight_stride%d'%(prefix,stride))
-    rpn_relu = conv_fpn_feat[stride]
-    maxout_stat = 0
-    if config.USE_MAXOUT>=1 and stride==config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 1
-    if config.USE_MAXOUT>=2 and stride!=config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 2
-
-
-    if maxout_stat==0:
-      rpn_cls_score = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d'%(prefix, stride), 2*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1), shared_weight = shared_vars[0][0], shared_bias = shared_vars[0][1])
-    elif maxout_stat==1:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-    else:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-
-    rpn_bbox_pred = conv_only(rpn_relu, '%s_rpn_bbox_pred_stride%d'%(prefix,stride), bbox_pred_len*num_anchors,
-        kernel=(1,1), pad=(0,0), stride=(1, 1), shared_weight = shared_vars[1][0], shared_bias = shared_vars[1][1])
-
-    # prepare rpn data
-    if not config.FBN:
-      rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                                shape=(0, 2, -1),
-                                                name="%s_rpn_cls_score_reshape_stride%s" % (prefix,stride))
-    else:
-      rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                                shape=(0, 2, -1),
-                                                name="%s_rpn_cls_score_reshape_stride%s_pre" % (prefix,stride))
-      rpn_cls_score_reshape = mx.symbol.BatchNorm(rpn_cls_score_reshape, fix_gamma=True, eps=2e-5, name="%s_rpn_cls_score_reshape_stride%s"%(prefix, stride))
-
-    rpn_bbox_pred_reshape = mx.symbol.Reshape(data=rpn_bbox_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix,stride))
-    if landmark:
-      rpn_landmark_pred = conv_only(rpn_relu, '%s_rpn_landmark_pred_stride%d'%(prefix,stride), landmark_pred_len*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1), shared_weight = shared_vars[2][0], shared_bias = shared_vars[2][1])
-      rpn_landmark_pred_reshape = mx.symbol.Reshape(data=rpn_landmark_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix,stride))
-
-    if config.TRAIN.RPN_ENABLE_OHEM>=2:
-      label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_score=rpn_cls_score_reshape, labels = label)
-
-      _bbox_weight = mx.sym.tile(anchor_weight, (1,1,bbox_pred_len))
-      _bbox_weight = _bbox_weight.reshape((0, -1, A * bbox_pred_len)).transpose((0,2,1))
-      bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
-
-      if landmark:
-        _landmark_weight = mx.sym.tile(anchor_weight, (1,1,landmark_pred_len))
-        _landmark_weight = _landmark_weight.reshape((0, -1, A * landmark_pred_len)).transpose((0,2,1))
-        landmark_weight = mx.sym.elemwise_mul(landmark_weight, _landmark_weight, name='%s_landmark_weight_mul_stride%s'%(prefix,stride))
-      #if not config.FACE_LANDMARK:
-      #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
-      #else:
-      #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
-    #cls loss
-    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
-                                           label=label,
-                                           multi_output=True,
-                                           normalization='valid', use_ignore=True, ignore_label=-1,
-                                           grad_scale = lr_mult,
-                                           name='%s_rpn_cls_prob_stride%d'%(prefix,stride))
-    ret_group.append(rpn_cls_prob)
-    ret_group.append(mx.sym.BlockGrad(label))
-
-    #bbox loss
-    bbox_diff = rpn_bbox_pred_reshape-bbox_target
-    bbox_diff = bbox_diff * bbox_weight
-    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_'%(prefix,stride), scalar=3.0, data=bbox_diff)
-    rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d'%(prefix,stride), data=rpn_bbox_loss_, grad_scale=1.0*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-    ret_group.append(rpn_bbox_loss)
-    ret_group.append(mx.sym.BlockGrad(bbox_weight))
-
-    #landmark loss
-    if landmark:
-      landmark_diff = rpn_landmark_pred_reshape-landmark_target
-      landmark_diff = landmark_diff * landmark_weight
-      rpn_landmark_loss_ = mx.symbol.smooth_l1(name='%s_rpn_landmark_loss_stride%d_'%(prefix,stride), scalar=3.0, data=landmark_diff)
-      rpn_landmark_loss = mx.sym.MakeLoss(name='%s_rpn_landmark_loss_stride%d'%(prefix,stride), data=rpn_landmark_loss_, grad_scale=0.5*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-      ret_group.append(rpn_landmark_loss)
-      ret_group.append(mx.sym.BlockGrad(landmark_weight))
-    return ret_group
-
-def get_mnet_train(sym):
-    return get_sym_train(sym)
-    #data = mx.symbol.Variable(name="data")
-    ## shared convolutional layers
-    #conv_fpn_feat = get_mnet_conv(data, sym)
-    #ret_group = []
-    #shared_vars = []
-    #if config.SHARE_WEIGHT_BBOX:
-    #  assert config.USE_MAXOUT==0
-    #  _name = 'face_rpn_cls_score_share'
-    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),   
-    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),   
-    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
-    #  shared_vars.append( [shared_weight, shared_bias] )
-    #  _name = 'face_rpn_bbox_pred_share'
-    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),   
-    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),   
-    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
-    #  shared_vars.append( [shared_weight, shared_bias] )
-    #else:
-    #  shared_vars.append( [None, None] )
-    #  shared_vars.append( [None, None] )
-    #if config.SHARE_WEIGHT_LANDMARK:
-    #  _name = 'face_rpn_landmark_pred_share'
-    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),   
-    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),   
-    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
-    #  shared_vars.append( [shared_weight, shared_bias] )
-    #else:
-    #  shared_vars.append( [None, None] )
-
-    #for stride in config.RPN_FEAT_STRIDE:
-    #  ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0, shared_vars = shared_vars)
-    #  ret_group += ret
-    #  if config.HEAD_BOX:
-    #    ret = get_out(conv_fpn_feat, 'head', stride, False, lr_mult=0.5)
-    #    ret_group += ret
-
-    #return mx.sym.Group(ret_group)
-
-
diff --git a/RetinaFace/rcnn/symbol/symbol_resnet.py b/RetinaFace/rcnn/symbol/symbol_resnet.py
deleted file mode 100644
index 8df5f1adc..000000000
--- a/RetinaFace/rcnn/symbol/symbol_resnet.py
+++ /dev/null
@@ -1,423 +0,0 @@
-import mxnet as mx
-import mxnet.ndarray as nd
-import mxnet.gluon as gluon
-import mxnet.gluon.nn as nn
-import mxnet.autograd as ag
-import numpy as np
-from rcnn.config import config
-from rcnn.PY_OP import rpn_fpn_ohem3
-from rcnn.symbol.symbol_common import get_sym_train
-
-def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), bias_wd_mult=0.0):
-  weight = mx.symbol.Variable(name="{}_weight".format(name),   
-      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-  bias = mx.symbol.Variable(name="{}_bias".format(name),   
-      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-  conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-      stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
-  return conv
-
-def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
-    assert kernel[0]==3
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
-    conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
-
-    separable = False
-    if separable:
-      assert kernel[0]==3
-    if not separable:
-      weight = mx.symbol.Variable(name="{}_weight".format(name),   
-          init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-      bias = mx.symbol.Variable(name="{}_bias".format(name),   
-          init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    else:
-      if filter_in<0:
-        filter_in = num_filter
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_sep_bn')
-      conv = mx.symbol.Activation(data=conv, act_type='relu', \
-          name="{}_sep_bn_relu".format(name))
-      conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
-          stride=(1,1), num_filter=num_filter, name="{}".format(name))
-      conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def ssh_context_module(body, num_filter, filter_in, name):
-  conv_dimred = conv_act_layer(body, name+'_conv1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=True, filter_in = filter_in)
-  conv5x5 = conv_act_layer(conv_dimred, name+'_conv2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True)
-  conv7x7_1 = conv_act_layer(conv_dimred, name+'_conv3_1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', separable=True)
-  conv7x7 = conv_act_layer(conv7x7_1, name+'_conv3_2',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True)
-  return (conv5x5, conv7x7)
-
-def conv_deformable(net, num_filter, num_group=1, act_type='relu',name=''):
-  f = num_group*18
-  conv_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = net,
-                      num_filter=f, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-  net = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=net, offset=conv_offset,
-                      num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=num_group, stride=(1, 1), no_bias=False)
-  net = mx.sym.BatchNorm(data=net, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-  if len(act_type)>0:
-    net = mx.symbol.Activation(data=net, act_type=act_type, name=name+'_act')
-  return net
-
-def ssh_detection_module(body, num_filter, filter_in, name):
-  conv3x3 = conv_act_layer(body, name+'_conv1',
-      num_filter, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', separable=True, filter_in=filter_in)
-  conv5x5, conv7x7 = ssh_context_module(body, num_filter//2, filter_in, name+'_context')
-  ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7], dim=1, name = name+'_concat')
-  ret = mx.symbol.Activation(data=ret, act_type='relu', name=name+'_concat_relu')
-  if config.USE_DCN==1:
-    ret = conv_deformable(ret, num_filter = num_filter*2, name = name+'_concat_dcn')
-  elif config.USE_DCN==2:
-    ret = conv_deformable2(ret, num_filter = num_filter*2, name = name+'_concat_dcn')
-  return ret
-
-
-def get_resnet_conv(data, sym):
-    all_layers = sym.get_internals()
-    isize = 640
-    _, out_shape, _ = all_layers.infer_shape(data = (1,3,isize,isize))
-    last_entry = None
-    c1 = None
-    c2 = None
-    c3 = None
-    #print(len(all_layers), len(out_shape))
-    #print(all_layers.__class__)
-    outputs = all_layers.list_outputs()
-    #print(outputs.__class__, len(outputs))
-    count = len(outputs)
-    for i in range(count):
-      name = outputs[i]
-      shape = out_shape[i]
-      if not name.endswith('_output'):
-        continue
-      if len(shape)!=4:
-        continue
-      print(name, shape)
-      if c1 is None and shape[2]==isize//16:
-        cname = last_entry[0]
-        print('c1', last_entry)
-        c1 = all_layers[cname]
-      if c2 is None and shape[2]==isize//32:
-        cname = last_entry[0]
-        print('c2', last_entry)
-        c2 = all_layers[cname]
-      if shape[2]==isize//32:
-        c3 = all_layers[name]
-        print('c3', name, shape)
-
-      last_entry = (name, shape)
-
-    c1_filter = -1
-    c2_filter = -1
-    c3_filter = -1
-
-    F1 = 256
-    F2 = 256
-    _bwm = 1.0
-    if config.NET_MODE==0:
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = conv_act_layer(c2, 'ssh_m2_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-      #    num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-      #    name='ssh_m2_red_upsampling')
-      c2_up = mx.symbol.UpSampling(c2_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-
-      c1 = c1_lateral+c2_up
-
-      c1 = conv_act_layer(c1, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==1:
-      c3_lateral = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.UpSampling(c3_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-
-      c1 = c1_lateral+c2_up
-
-      c1 = conv_act_layer(c1, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==2:
-      n1 = ssh_detection_module(c1, F2, F2, 'ssh_n1_det')
-      n2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_n2_det')
-      n3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_n3_det')
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==3:
-      #c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3 = ssh_detection_module(c3, F2//2, c3_filter, 'ssh_c3_lateral')
-      c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      #c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = ssh_detection_module(c2, F2//2, c2_filter, 'ssh_c2_lateral')
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      #c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-      #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = ssh_detection_module(c1, F2//2, c1_filter, 'ssh_c1_lateral')
-      c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==4:
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2//2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1//2, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1//2, c3_filter, 'ssh_m3_det')
-    elif config.NET_MODE==5:
-      c3 = conv_act_layer_dw(c3, 'ssh_c3_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3 = conv_act_layer(c3, 'ssh_c3_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
-      c2 = conv_act_layer_dw(c2, 'ssh_c2_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
-      c2 = c2_lateral+c3_up
-      c2 = conv_act_layer(c2, 'ssh_c2_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1 = conv_act_layer_dw(c1, 'ssh_c1_lateral_m',
-          F2, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
-      c1 = c1_lateral+c2_up
-      c1 = conv_act_layer(c1, 'ssh_c1_aggr',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-
-      m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
-      m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
-
-    return {8: m1, 16:m2, 32: m3}, {8: n1, 16:n2, 32: n3}
-
-def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0):
-    A = config.NUM_ANCHORS
-    bbox_pred_len = 4
-    landmark_pred_len = 10
-    if config.USE_BLUR:
-      bbox_pred_len = 5
-    if config.USE_OCCLUSION:
-      landmark_pred_len = 15
-    ret_group = []
-    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
-    label = mx.symbol.Variable(name='%s_label_stride%d'%(prefix,stride))
-    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d'%(prefix,stride))
-    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d'%(prefix,stride))
-    if landmark:
-      landmark_target = mx.symbol.Variable(name='%s_landmark_target_stride%d'%(prefix,stride))
-      landmark_weight = mx.symbol.Variable(name='%s_landmark_weight_stride%d'%(prefix,stride))
-    rpn_relu = conv_fpn_feat[stride]
-    maxout_stat = 0
-    if config.USE_MAXOUT>=1 and stride==config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 1
-    if config.USE_MAXOUT>=2 and stride!=config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 2
-
-    if maxout_stat==0:
-      rpn_cls_score = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d'%(prefix, stride), 2*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1))
-    elif maxout_stat==1:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-    else:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_only(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1))
-        rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-
-    rpn_bbox_pred = conv_only(rpn_relu, '%s_rpn_bbox_pred_stride%d'%(prefix,stride), bbox_pred_len*num_anchors,
-        kernel=(1,1), pad=(0,0), stride=(1, 1))
-
-    # prepare rpn data
-    if not config.FBN:
-      rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                                shape=(0, 2, -1),
-                                                name="%s_rpn_cls_score_reshape_stride%s" % (prefix,stride))
-    else:
-      rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                                shape=(0, 2, -1),
-                                                name="%s_rpn_cls_score_reshape_stride%s_pre" % (prefix,stride))
-      rpn_cls_score_reshape = mx.symbol.BatchNorm(rpn_cls_score_reshape, fix_gamma=True, eps=2e-5, name="%s_rpn_cls_score_reshape_stride%s"%(prefix, stride))
-
-    rpn_bbox_pred_reshape = mx.symbol.Reshape(data=rpn_bbox_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix,stride))
-    if landmark:
-      rpn_landmark_pred = conv_only(rpn_relu, '%s_rpn_landmark_pred_stride%d'%(prefix,stride), landmark_pred_len*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1))
-      rpn_landmark_pred_reshape = mx.symbol.Reshape(data=rpn_landmark_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix,stride))
-
-    if config.TRAIN.RPN_ENABLE_OHEM>=2:
-      label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_score=rpn_cls_score_reshape, labels = label)
-
-      _bbox_weight = mx.sym.tile(anchor_weight, (1,1,bbox_pred_len))
-      _bbox_weight = _bbox_weight.reshape((0, -1, A * bbox_pred_len)).transpose((0,2,1))
-      bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
-
-      if landmark:
-        _landmark_weight = mx.sym.tile(anchor_weight, (1,1,landmark_pred_len))
-        _landmark_weight = _landmark_weight.reshape((0, -1, A * landmark_pred_len)).transpose((0,2,1))
-        landmark_weight = mx.sym.elemwise_mul(landmark_weight, _landmark_weight, name='%s_landmark_weight_mul_stride%s'%(prefix,stride))
-      #if not config.FACE_LANDMARK:
-      #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
-      #else:
-      #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
-    #cls loss
-    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
-                                           label=label,
-                                           multi_output=True,
-                                           normalization='valid', use_ignore=True, ignore_label=-1,
-                                           grad_scale = lr_mult,
-                                           name='%s_rpn_cls_prob_stride%d'%(prefix,stride))
-    ret_group.append(rpn_cls_prob)
-    ret_group.append(mx.sym.BlockGrad(label))
-
-    #bbox loss
-    bbox_diff = rpn_bbox_pred_reshape-bbox_target
-    bbox_diff = bbox_diff * bbox_weight
-    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_'%(prefix,stride), scalar=3.0, data=bbox_diff)
-    rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d'%(prefix,stride), data=rpn_bbox_loss_, grad_scale=1.0*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-    ret_group.append(rpn_bbox_loss)
-    ret_group.append(mx.sym.BlockGrad(bbox_weight))
-
-    #landmark loss
-    if landmark:
-      landmark_diff = rpn_landmark_pred_reshape-landmark_target
-      landmark_diff = landmark_diff * landmark_weight
-      rpn_landmark_loss_ = mx.symbol.smooth_l1(name='%s_rpn_landmark_loss_stride%d_'%(prefix,stride), scalar=3.0, data=landmark_diff)
-      rpn_landmark_loss = mx.sym.MakeLoss(name='%s_rpn_landmark_loss_stride%d'%(prefix,stride), data=rpn_landmark_loss_, grad_scale=0.5*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-      ret_group.append(rpn_landmark_loss)
-      ret_group.append(mx.sym.BlockGrad(landmark_weight))
-    return ret_group
-
-def get_resnet_train(sym):
-    return get_sym_train(sym)
-    #data = mx.symbol.Variable(name="data")
-    ## shared convolutional layers
-    #conv_fpn_feat, conv_fpn_feat2 = get_resnet_conv(data, sym)
-    #ret_group = []
-    #for stride in config.RPN_FEAT_STRIDE:
-    #  ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0)
-    #  ret_group += ret
-    #  if config.HEAD_BOX:
-    #    ret = get_out(conv_fpn_feat2, 'head', stride, False, lr_mult=1.0)
-    #    ret_group += ret
-
-    #return mx.sym.Group(ret_group)
-
-
diff --git a/RetinaFace/rcnn/symbol/symbol_ssh.py b/RetinaFace/rcnn/symbol/symbol_ssh.py
deleted file mode 100644
index 31541ee23..000000000
--- a/RetinaFace/rcnn/symbol/symbol_ssh.py
+++ /dev/null
@@ -1,365 +0,0 @@
-import mxnet as mx
-import numpy as np
-from rcnn.config import config
-from rcnn.PY_OP import rpn_fpn_ohem3
-FPN = False
-USE_DCN=False
-
-def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
-    stride=(1,1), act_type="relu", bias_wd_mult=0.0, dcn=False):
-
-    weight = mx.symbol.Variable(name="{}_weight".format(name),   
-        init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
-    bias = mx.symbol.Variable(name="{}_bias".format(name),   
-        init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-    if not dcn:
-      conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
-          stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
-    else:
-      assert kernel[0]==3 and kernel[1]==3
-      num_group = 1
-      f = num_group*18
-      offset_weight = mx.symbol.Variable(name="{}_offset_weight".format(name),   
-          init=mx.init.Constant(0.0), attr={'__lr_mult__': '1.0'})
-      offset_bias = mx.symbol.Variable(name="{}_offset_bias".format(name),   
-          init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(bias_wd_mult)})
-      conv_offset = mx.symbol.Convolution(name=name+'_offset', data = from_layer, weight=offset_weight, bias=offset_bias,
-                          num_filter=f, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-      conv = mx.contrib.symbol.DeformableConvolution(name=name, data=from_layer, offset=conv_offset, weight=weight, bias=bias,
-                          num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=num_group, stride=(1, 1), no_bias=False)
-    if len(act_type)>0:
-      relu = mx.symbol.Activation(data=conv, act_type=act_type, \
-          name="{}_{}".format(name, act_type))
-    else:
-      relu = conv
-    return relu
-
-def ssh_context_module(body, num_filters, name):
-  conv_dimred = conv_act_layer(body, name+'_conv1',
-      num_filters, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', dcn=False)
-  conv5x5 = conv_act_layer(conv_dimred, name+'_conv2',
-      num_filters, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', dcn=USE_DCN)
-  conv7x7_1 = conv_act_layer(conv_dimred, name+'_conv3_1',
-      num_filters, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', dcn=False)
-  conv7x7 = conv_act_layer(conv7x7_1, name+'_conv3_2',
-      num_filters, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', dcn=USE_DCN)
-  return (conv5x5, conv7x7)
-
-def ssh_detection_module(body, num_filters, name):
-  conv3x3 = conv_act_layer(body, name+'_conv1',
-      num_filters, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='', dcn=USE_DCN)
-  conv5x5, conv7x7 = ssh_context_module(body, num_filters//2, name+'_context')
-  ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7], dim=1, name = name+'_concat')
-  ret = mx.symbol.Activation(data=ret, act_type='relu', name=name+'_concat_relu')
-  return ret
-
-def conv_bn(input, filter, ksize, stride, padding, act_type='relu', name=''):
-  conv = mx.symbol.Convolution(data=input, kernel=(ksize,ksize), pad=(padding,padding), \
-      stride=(stride,stride), num_filter=filter, name=name+"_conv")
-  ret = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=0.9, name=name + '_bn')
-  if act_type is not None:
-    ret = mx.symbol.Activation(data=ret, act_type=act_type, \
-        name="{}_{}".format(name, act_type))
-  return ret
-
-def cpm(input, name):
-    # residual
-    branch1 = conv_bn(input, 1024, 1, 1, 0, act_type=None, name=name+"_branch1")
-    branch2a = conv_bn(input, 256, 1, 1, 0, act_type='relu', name=name+"_branch2a")
-    branch2b = conv_bn(branch2a, 256, 3, 1, 1, act_type='relu', name=name+"_branch2b")
-    branch2c = conv_bn(branch2b, 1024, 1, 1, 0, act_type=None, name=name+"_branch2c")
-    sum = branch1 + branch2c
-    rescomb = mx.symbol.Activation(data=sum, act_type='relu', name="%s_relu2"%(name))
-
-    ssh_out = ssh_detection_module(rescomb, 256, name=name+"_ssh")
-    return ssh_out
-
-def get_feat_down(conv_feat):
-    #P5 = mx.symbol.Convolution(data=conv_feat[0], kernel=(1, 1), num_filter=256, name="P5_lateral")
-    P5 = conv_act_layer(conv_feat[0], 'P5_lateral',
-        256, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu')
-
-    # P5 2x upsampling + C4 = P4
-    P5_up   = mx.symbol.UpSampling(P5, scale=2, sample_type='nearest', workspace=512, name='P5_upsampling', num_args=1)
-    #P4_la   = mx.symbol.Convolution(data=conv_feat[1], kernel=(1, 1), num_filter=256, name="P4_lateral")
-    P4_la = conv_act_layer(conv_feat[1], 'P4_lateral',
-        256, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu')
-    P5_clip = mx.symbol.Crop(*[P5_up, P4_la], name="P4_clip")
-    P4      = mx.sym.ElementWiseSum(*[P5_clip, P4_la], name="P4_sum")
-    #P4      = mx.symbol.Convolution(data=P4, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P4_aggregate")
-    P4 = conv_act_layer(P4, 'P4_aggregate',
-        256, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-
-    # P4 2x upsampling + C3 = P3
-    P4_up   = mx.symbol.UpSampling(P4, scale=2, sample_type='nearest', workspace=512, name='P4_upsampling', num_args=1)
-    #P3_la   = mx.symbol.Convolution(data=conv_feat[2], kernel=(1, 1), num_filter=256, name="P3_lateral")
-    P3_la = conv_act_layer(conv_feat[2], 'P3_lateral',
-        256, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu')
-    P4_clip = mx.symbol.Crop(*[P4_up, P3_la], name="P3_clip")
-    P3      = mx.sym.ElementWiseSum(*[P4_clip, P3_la], name="P3_sum")
-    #P3      = mx.symbol.Convolution(data=P3, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P3_aggregate")
-    P3 = conv_act_layer(P3, 'P3_aggregate',
-        256, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-
-    return P3, P4, P5
-
-def get_ssh_conv(data):
-    """
-    shared convolutional layers
-    :param data: Symbol
-    :return: Symbol
-    """
-    # group 1
-    #conv1_1 = mx.symbol.Convolution(
-    #    data=data, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_1")
-    #relu1_1 = mx.symbol.Activation(data=conv1_1, act_type="relu", name="relu1_1")
-    relu1_1 = conv_act_layer(data, 'conv1_1',
-        64, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv1_2 = mx.symbol.Convolution(
-    #    data=relu1_1, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_2")
-    #relu1_2 = mx.symbol.Activation(data=conv1_2, act_type="relu", name="relu1_2")
-    relu1_2 = conv_act_layer(relu1_1, 'conv1_2',
-        64, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    pool1 = mx.symbol.Pooling(
-        data=relu1_2, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool1")
-    # group 2
-    #conv2_1 = mx.symbol.Convolution(
-    #    data=pool1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_1")
-    #relu2_1 = mx.symbol.Activation(data=conv2_1, act_type="relu", name="relu2_1")
-    relu2_1 = conv_act_layer(pool1, 'conv2_1',
-        128, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv2_2 = mx.symbol.Convolution(
-    #    data=relu2_1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_2")
-    #relu2_2 = mx.symbol.Activation(data=conv2_2, act_type="relu", name="relu2_2")
-    relu2_2 = conv_act_layer(relu2_1, 'conv2_2',
-        128, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    pool2 = mx.symbol.Pooling(
-        data=relu2_2, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool2")
-    # group 3
-    #conv3_1 = mx.symbol.Convolution(
-    #    data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_1")
-    #relu3_1 = mx.symbol.Activation(data=conv3_1, act_type="relu", name="relu3_1")
-    relu3_1 = conv_act_layer(pool2, 'conv3_1',
-        256, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv3_2 = mx.symbol.Convolution(
-    #    data=relu3_1, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_2")
-    #relu3_2 = mx.symbol.Activation(data=conv3_2, act_type="relu", name="relu3_2")
-    relu3_2 = conv_act_layer(relu3_1, 'conv3_2',
-        256, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv3_3 = mx.symbol.Convolution(
-    #    data=relu3_2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_3")
-    #relu3_3 = mx.symbol.Activation(data=conv3_3, act_type="relu", name="relu3_3")
-    relu3_3 = conv_act_layer(relu3_2, 'conv3_3',
-        256, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    pool3 = mx.symbol.Pooling(
-        data=relu3_3, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool3")
-    # group 4
-    #conv4_1 = mx.symbol.Convolution(
-    #    data=pool3, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_1")
-    #relu4_1 = mx.symbol.Activation(data=conv4_1, act_type="relu", name="relu4_1")
-    relu4_1 = conv_act_layer(pool3, 'conv4_1',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv4_2 = mx.symbol.Convolution(
-    #    data=relu4_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_2")
-    #relu4_2 = mx.symbol.Activation(data=conv4_2, act_type="relu", name="relu4_2")
-    relu4_2 = conv_act_layer(relu4_1, 'conv4_2',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv4_3 = mx.symbol.Convolution(
-    #    data=relu4_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_3")
-    #relu4_3 = mx.symbol.Activation(data=conv4_3, act_type="relu", name="relu4_3")
-    relu4_3 = conv_act_layer(relu4_2, 'conv4_3',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    pool4 = mx.symbol.Pooling(
-        data=relu4_3, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool4")
-    # group 5
-    #conv5_1 = mx.symbol.Convolution(
-    #    data=pool4, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_1")
-    #relu5_1 = mx.symbol.Activation(data=conv5_1, act_type="relu", name="relu5_1")
-    relu5_1 = conv_act_layer(pool4, 'conv5_1',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv5_2 = mx.symbol.Convolution(
-    #    data=relu5_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_2")
-    #relu5_2 = mx.symbol.Activation(data=conv5_2, act_type="relu", name="relu5_2")
-    relu5_2 = conv_act_layer(relu5_1, 'conv5_2',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    #conv5_3 = mx.symbol.Convolution(
-    #    data=relu5_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_3")
-    #relu5_3 = mx.symbol.Activation(data=conv5_3, act_type="relu", name="relu5_3")
-    relu5_3 = conv_act_layer(relu5_2, 'conv5_3',
-        512, kernel=(3,3), pad=(1,1), stride=(1, 1), act_type='relu')
-    m3_pool = mx.sym.Pooling(data=relu5_3, kernel=(2, 2), stride=(2,2), pad=(0,0), pool_type='max')
-    if config.SSH_MODE<=5:
-      #if FPN:
-      #  relu4_3, relu5_3, m3_pool = get_feat_down([m3_pool, relu5_3, relu4_3])
-
-      F1 = 256
-      F2 = 128
-      if config.SSH_MODE==1:
-        F2 = 256
-      _bwm = 1.0
-      conv4_128 = conv_act_layer(relu4_3, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      conv5_128 = conv_act_layer(relu5_3, 'ssh_m2_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-          num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-          name='ssh_m2_red_upsampling')
-      #conv5_128_up = mx.symbol.UpSampling(conv5_128, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      #conv5_128_up = mx.symbol.Crop(*[conv5_128_up, conv4_128])
-
-      conv_sum = conv4_128+conv5_128_up
-      #conv_sum = conv_1x1
-
-      m1_conv = conv_act_layer(conv_sum, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = ssh_detection_module(m1_conv, F2, 'ssh_m1_det')
-      m2 = ssh_detection_module(relu5_3, F1, 'ssh_m2_det')
-      m3 = ssh_detection_module(m3_pool, F1, 'ssh_m3_det')
-      return {8: m1, 16:m2, 32: m3}
-    else:
-      F1 = 256
-      F2 = 256
-      _bwm = 1.0
-      conv4_128 = conv_act_layer(relu4_3, 'ssh_m1_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      conv5_128 = conv_act_layer(relu5_3, 'ssh_m2_red_conv',
-          F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
-          num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
-          name='ssh_m2_red_upsampling')
-      #conv5_128_up = mx.symbol.UpSampling(conv5_128, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
-      conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
-      #conv5_128_up = mx.symbol.Crop(*[conv5_128_up, conv4_128])
-
-      conv_sum = conv4_128+conv5_128_up
-      m1_conv = conv_act_layer(conv_sum, 'ssh_m1_conv',
-          F2, kernel=(3, 3), pad=(1, 1), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
-      m1 = cpm(m1_conv, 'ssh_m1_det')
-      m2 = cpm(relu5_3, 'ssh_m2_det')
-      m3 = cpm(m3_pool, 'ssh_m3_det')
-      return {8: m1, 16:m2, 32: m3}
-
-def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0):
-    A = config.NUM_ANCHORS
-    ret_group = []
-    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
-    label = mx.symbol.Variable(name='%s_label_stride%d'%(prefix,stride))
-    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d'%(prefix,stride))
-    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d'%(prefix,stride))
-    if landmark:
-      landmark_target = mx.symbol.Variable(name='%s_landmark_target_stride%d'%(prefix,stride))
-      landmark_weight = mx.symbol.Variable(name='%s_landmark_weight_stride%d'%(prefix,stride))
-    rpn_relu = conv_fpn_feat[stride]
-    maxout_stat = 0
-    if config.USE_MAXOUT>=1 and stride==config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 1
-    if config.USE_MAXOUT>=2 and stride!=config.RPN_FEAT_STRIDE[-1]:
-      maxout_stat = 2
-
-    if maxout_stat==0:
-      rpn_cls_score = conv_act_layer(rpn_relu, '%s_rpn_cls_score_stride%d'%(prefix, stride), 2*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-    elif maxout_stat==1:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_act_layer(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-        rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_act_layer(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-    else:
-      cls_list = []
-      for a in range(num_anchors):
-        rpn_cls_score_bg = conv_act_layer(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_bg'%(prefix,stride,a), 1,
-            kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-        cls_list.append(rpn_cls_score_bg)
-        rpn_cls_score_fg = conv_act_layer(rpn_relu, '%s_rpn_cls_score_stride%d_anchor%d_fg'%(prefix,stride,a), 3,
-            kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-        rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg, axis=1, keepdims=True)
-        cls_list.append(rpn_cls_score_fg)
-      rpn_cls_score = mx.sym.concat(*cls_list, dim=1, name='%s_rpn_cls_score_stride%d'%(prefix,stride))
-
-    rpn_bbox_pred = conv_act_layer(rpn_relu, '%s_rpn_bbox_pred_stride%d'%(prefix,stride), 4*num_anchors,
-        kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-
-    # prepare rpn data
-    rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                              shape=(0, 2, -1),
-                                              name="%s_rpn_cls_score_reshape_stride%s" % (prefix,stride))
-    rpn_bbox_pred_reshape = mx.symbol.Reshape(data=rpn_bbox_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix,stride))
-    if landmark:
-      rpn_landmark_pred = conv_act_layer(rpn_relu, '%s_rpn_landmark_pred_stride%d'%(prefix,stride), 10*num_anchors,
-          kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='')
-      rpn_landmark_pred_reshape = mx.symbol.Reshape(data=rpn_landmark_pred,
-                                              shape=(0, 0, -1),
-                                              name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix,stride))
-
-    if config.TRAIN.RPN_ENABLE_OHEM>=2:
-      label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_score=rpn_cls_score_reshape, labels = label)
-
-      _bbox_weight = mx.sym.tile(anchor_weight, (1,1,4))
-      _bbox_weight = _bbox_weight.reshape((0, -1, A * 4)).transpose((0,2,1))
-      bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
-
-      if landmark:
-        _landmark_weight = mx.sym.tile(anchor_weight, (1,1,10))
-        _landmark_weight = _landmark_weight.reshape((0, -1, A * 10)).transpose((0,2,1))
-        landmark_weight = mx.sym.elemwise_mul(landmark_weight, _landmark_weight, name='%s_landmark_weight_mul_stride%s'%(prefix,stride))
-      #if not config.FACE_LANDMARK:
-      #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
-      #else:
-      #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
-    #cls loss
-    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
-                                           label=label,
-                                           multi_output=True,
-                                           normalization='valid', use_ignore=True, ignore_label=-1,
-                                           grad_scale = lr_mult,
-                                           name='%s_rpn_cls_prob_stride%d'%(prefix,stride))
-    ret_group.append(rpn_cls_prob)
-    ret_group.append(mx.sym.BlockGrad(label))
-
-    #bbox loss
-    bbox_diff = rpn_bbox_pred_reshape-bbox_target
-    bbox_diff = bbox_diff * bbox_weight
-    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_'%(prefix,stride), scalar=3.0, data=bbox_diff)
-    rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d'%(prefix,stride), data=rpn_bbox_loss_, grad_scale=1.0*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-    ret_group.append(rpn_bbox_loss)
-    ret_group.append(mx.sym.BlockGrad(bbox_weight))
-
-    #landmark loss
-    if landmark:
-      landmark_diff = rpn_landmark_pred_reshape-landmark_target
-      landmark_diff = landmark_diff * landmark_weight
-      rpn_landmark_loss_ = mx.symbol.smooth_l1(name='%s_rpn_landmark_loss_stride%d_'%(prefix,stride), scalar=3.0, data=landmark_diff)
-      rpn_landmark_loss = mx.sym.MakeLoss(name='%s_rpn_landmark_loss_stride%d'%(prefix,stride), data=rpn_landmark_loss_, grad_scale=0.5*lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
-      ret_group.append(rpn_landmark_loss)
-      ret_group.append(mx.sym.BlockGrad(landmark_weight))
-    return ret_group
-
-def get_ssh_train():
-    """
-    Region Proposal Network with VGG
-    :return: Symbol
-    """
-    data = mx.symbol.Variable(name="data")
-
-    # shared convolutional layers
-    conv_fpn_feat = get_ssh_conv(data)
-    ret_group = []
-    for stride in config.RPN_FEAT_STRIDE:
-      ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0)
-      ret_group += ret
-      if config.HEAD_BOX:
-        ret = get_out(conv_fpn_feat, 'head', stride, False, lr_mult=1.0)
-        ret_group += ret
-
-    return mx.sym.Group(ret_group)
-
-
diff --git a/RetinaFace/rcnn/tools/reeval.py b/RetinaFace/rcnn/tools/reeval.py
deleted file mode 100644
index 1bc9c5b8a..000000000
--- a/RetinaFace/rcnn/tools/reeval.py
+++ /dev/null
@@ -1,50 +0,0 @@
-import argparse
-try:
-    import cPickle as pickle
-except ImportError:
-    import pickle
-import os
-import mxnet as mx
-
-from ..logger import logger
-from ..config import config, default, generate_config
-from ..dataset import *
-
-
-def reeval(args):
-    # load imdb
-    imdb = eval(args.dataset)(args.image_set, args.root_path, args.dataset_path)
-
-    # load detection results
-    cache_file = os.path.join(imdb.cache_path, imdb.name, 'detections.pkl')
-    with open(cache_file) as f:
-        detections = pickle.load(f)
-
-    # eval
-    imdb.evaluate_detections(detections)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='imdb test')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # other
-    parser.add_argument('--no_shuffle', help='disable random shuffle', action='store_true')
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    logger.info('Called with argument: %s' % args)
-    reeval(args)
-
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFace/rcnn/tools/test_rcnn.py b/RetinaFace/rcnn/tools/test_rcnn.py
deleted file mode 100644
index 5e5310632..000000000
--- a/RetinaFace/rcnn/tools/test_rcnn.py
+++ /dev/null
@@ -1,109 +0,0 @@
-import argparse
-import pprint
-import mxnet as mx
-
-from ..logger import logger
-from ..config import config, default, generate_config
-from ..symbol import *
-from ..dataset import *
-from ..core.loader import TestLoader
-from ..core.tester import Predictor, pred_eval
-from ..utils.load_model import load_param
-
-
-def test_rcnn(network, dataset, image_set, root_path, dataset_path,
-              ctx, prefix, epoch,
-              vis, shuffle, has_rpn, proposal, thresh):
-    # set config
-    if has_rpn:
-        config.TEST.HAS_RPN = True
-
-    # print config
-    logger.info(pprint.pformat(config))
-
-    # load symbol and testing data
-    if has_rpn:
-        sym = eval('get_' + network + '_test')(num_classes=config.NUM_CLASSES, num_anchors=config.NUM_ANCHORS)
-        imdb = eval(dataset)(image_set, root_path, dataset_path)
-        roidb = imdb.gt_roidb()
-    else:
-        sym = eval('get_' + network + '_rcnn_test')(num_classes=config.NUM_CLASSES)
-        imdb = eval(dataset)(image_set, root_path, dataset_path)
-        gt_roidb = imdb.gt_roidb()
-        roidb = eval('imdb.' + proposal + '_roidb')(gt_roidb)
-
-    # get test data iter
-    test_data = TestLoader(roidb, batch_size=1, shuffle=shuffle, has_rpn=has_rpn)
-
-    # load model
-    arg_params, aux_params = load_param(prefix, epoch, convert=True, ctx=ctx, process=True)
-
-    # infer shape
-    data_shape_dict = dict(test_data.provide_data)
-    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
-    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
-
-    # check parameters
-    for k in sym.list_arguments():
-        if k in data_shape_dict or 'label' in k:
-            continue
-        assert k in arg_params, k + ' not initialized'
-        assert arg_params[k].shape == arg_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
-    for k in sym.list_auxiliary_states():
-        assert k in aux_params, k + ' not initialized'
-        assert aux_params[k].shape == aux_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
-
-    # decide maximum shape
-    data_names = [k[0] for k in test_data.provide_data]
-    label_names = None
-    max_data_shape = [('data', (1, 3, max([v[0] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-    if not has_rpn:
-        max_data_shape.append(('rois', (1, config.TEST.PROPOSAL_POST_NMS_TOP_N + 30, 5)))
-
-    # create predictor
-    predictor = Predictor(sym, data_names, label_names,
-                          context=ctx, max_data_shapes=max_data_shape,
-                          provide_data=test_data.provide_data, provide_label=test_data.provide_label,
-                          arg_params=arg_params, aux_params=aux_params)
-
-    # start detection
-    pred_eval(predictor, test_data, imdb, vis=vis, thresh=thresh)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Test a Fast R-CNN network')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.test_image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # testing
-    parser.add_argument('--prefix', help='model to test with', default=default.rcnn_prefix, type=str)
-    parser.add_argument('--epoch', help='model to test with', default=default.rcnn_epoch, type=int)
-    parser.add_argument('--gpu', help='GPU device to test with', default=0, type=int)
-    # rcnn
-    parser.add_argument('--vis', help='turn on visualization', action='store_true')
-    parser.add_argument('--thresh', help='valid detection threshold', default=1e-3, type=float)
-    parser.add_argument('--shuffle', help='shuffle data on visualization', action='store_true')
-    parser.add_argument('--has_rpn', help='generate proposals on the fly', action='store_true')
-    parser.add_argument('--proposal', help='can be ss for selective search or rpn', default='rpn', type=str)
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    logger.info('Called with argument: %s' % args)
-    ctx = mx.gpu(args.gpu)
-    test_rcnn(args.network, args.dataset, args.image_set, args.root_path, args.dataset_path,
-              ctx, args.prefix, args.epoch,
-              args.vis, args.shuffle, args.has_rpn, args.proposal, args.thresh)
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFace/rcnn/tools/test_rpn.py b/RetinaFace/rcnn/tools/test_rpn.py
deleted file mode 100644
index 5622227dc..000000000
--- a/RetinaFace/rcnn/tools/test_rpn.py
+++ /dev/null
@@ -1,102 +0,0 @@
-import argparse
-import pprint
-import mxnet as mx
-
-from ..logger import logger
-from ..config import config, default, generate_config
-from ..symbol import *
-from ..dataset import *
-from ..core.loader import TestLoader
-from ..core.tester import Predictor, generate_proposals, test_proposals
-from ..utils.load_model import load_param
-
-
-def test_rpn(network, dataset, image_set, root_path, dataset_path,
-             ctx, prefix, epoch,
-             vis, shuffle, thresh, test_output=False):
-    # rpn generate proposal config
-    config.TEST.HAS_RPN = True
-
-    # print config
-    logger.info(pprint.pformat(config))
-
-    # load symbol
-    sym = eval('get_' + network + '_rpn_test')()
-
-    # load dataset and prepare imdb for training
-    imdb = eval(dataset)(image_set, root_path, dataset_path)
-    roidb = imdb.gt_roidb()
-    test_data = TestLoader(roidb, batch_size=1, shuffle=shuffle, has_rpn=True, withlabel=True)
-
-    # load model
-    arg_params, aux_params = load_param(prefix, epoch, convert=True, ctx=ctx)
-
-    # infer shape
-    data_shape_dict = dict(test_data.provide_data)
-    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
-    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
-
-    # check parameters
-    for k in sym.list_arguments():
-        if k in data_shape_dict or 'label' in k:
-            continue
-        assert k in arg_params, k + ' not initialized'
-        assert arg_params[k].shape == arg_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
-    for k in sym.list_auxiliary_states():
-        assert k in aux_params, k + ' not initialized'
-        assert aux_params[k].shape == aux_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
-
-    # decide maximum shape
-    data_names = [k[0] for k in test_data.provide_data]
-    label_names = None if test_data.provide_label is None else [k[0] for k in test_data.provide_label]
-    max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-
-    # create predictor
-    predictor = Predictor(sym, data_names, label_names,
-                          context=ctx, max_data_shapes=max_data_shape,
-                          provide_data=test_data.provide_data, provide_label=test_data.provide_label,
-                          arg_params=arg_params, aux_params=aux_params)
-
-    # start testing
-    if not test_output:
-      imdb_boxes = generate_proposals(predictor, test_data, imdb, vis=vis, thresh=thresh)
-      imdb.evaluate_recall(roidb, candidate_boxes=imdb_boxes)
-    else:
-      test_proposals(predictor, test_data, imdb, roidb, vis=vis)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Test a Region Proposal Network')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.test_image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # testing
-    parser.add_argument('--prefix', help='model to test with', default=default.rpn_prefix, type=str)
-    parser.add_argument('--epoch', help='model to test with', default=default.rpn_epoch, type=int)
-    # rpn
-    parser.add_argument('--gpu', help='GPU device to test with', default=0, type=int)
-    parser.add_argument('--vis', help='turn on visualization', action='store_true')
-    parser.add_argument('--thresh', help='rpn proposal threshold', default=0, type=float)
-    parser.add_argument('--shuffle', help='shuffle data on visualization', action='store_true')
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    logger.info('Called with argument: %s' % args)
-    ctx = mx.gpu(args.gpu)
-    test_rpn(args.network, args.dataset, args.image_set, args.root_path, args.dataset_path,
-             ctx, args.prefix, args.epoch,
-             args.vis, args.shuffle, args.thresh)
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFace/rcnn/tools/train_rcnn.py b/RetinaFace/rcnn/tools/train_rcnn.py
deleted file mode 100644
index 4d2f79784..000000000
--- a/RetinaFace/rcnn/tools/train_rcnn.py
+++ /dev/null
@@ -1,172 +0,0 @@
-import argparse
-import pprint
-import mxnet as mx
-
-from ..logger import logger
-from ..config import config, default, generate_config
-from ..symbol import *
-from ..core import callback, metric
-from ..core.loader import ROIIter
-from ..core.module import MutableModule
-from ..processing.bbox_regression import add_bbox_regression_targets
-from ..utils.load_data import load_proposal_roidb, merge_roidb, filter_roidb
-from ..utils.load_model import load_param
-
-
-def train_rcnn(network, dataset, image_set, root_path, dataset_path,
-               frequent, kvstore, work_load_list, no_flip, no_shuffle, resume,
-               ctx, pretrained, epoch, prefix, begin_epoch, end_epoch,
-               train_shared, lr, lr_step, proposal):
-    # set up config
-    config.TRAIN.BATCH_IMAGES = 2
-    config.TRAIN.BATCH_ROIS = 128
-    if proposal == 'ss':
-        config.TRAIN.BG_THRESH_LO = 0.1  # reproduce Fast R-CNN
-
-    # load symbol
-    sym = eval('get_' + network + '_rcnn')(num_classes=config.NUM_CLASSES)
-
-    # setup multi-gpu
-    batch_size = len(ctx)
-    input_batch_size = config.TRAIN.BATCH_IMAGES * batch_size
-
-    # print config
-    logger.info(pprint.pformat(config))
-
-    # load dataset and prepare imdb for training
-    image_sets = [iset for iset in image_set.split('+')]
-    roidbs = [load_proposal_roidb(dataset, image_set, root_path, dataset_path,
-                                  proposal=proposal, append_gt=True, flip=not no_flip)
-              for image_set in image_sets]
-    roidb = merge_roidb(roidbs)
-    roidb = filter_roidb(roidb)
-    means, stds = add_bbox_regression_targets(roidb)
-
-    # load training data
-    train_data = ROIIter(roidb, batch_size=input_batch_size, shuffle=not no_shuffle,
-                         ctx=ctx, work_load_list=work_load_list, aspect_grouping=config.TRAIN.ASPECT_GROUPING)
-
-    # infer max shape
-    max_data_shape = [('data', (input_batch_size, 3, max([v[0] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-    logger.info('providing maximum shape %s' % max_data_shape)
-
-    # infer shape
-    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
-    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
-    out_shape_dict = dict(zip(sym.list_outputs(), out_shape))
-    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
-    logger.info('output shape %s' % pprint.pformat(out_shape_dict))
-
-    # load and initialize params
-    if resume:
-        arg_params, aux_params = load_param(prefix, begin_epoch, convert=True)
-    else:
-        arg_params, aux_params = load_param(pretrained, epoch, convert=True)
-        arg_params['cls_score_weight'] = mx.random.normal(0, 0.01, shape=arg_shape_dict['cls_score_weight'])
-        arg_params['cls_score_bias'] = mx.nd.zeros(shape=arg_shape_dict['cls_score_bias'])
-        arg_params['bbox_pred_weight'] = mx.random.normal(0, 0.001, shape=arg_shape_dict['bbox_pred_weight'])
-        arg_params['bbox_pred_bias'] = mx.nd.zeros(shape=arg_shape_dict['bbox_pred_bias'])
-
-    # check parameter shapes
-    for k in sym.list_arguments():
-        if k in data_shape_dict:
-            continue
-        assert k in arg_params, k + ' not initialized'
-        assert arg_params[k].shape == arg_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
-    for k in sym.list_auxiliary_states():
-        assert k in aux_params, k + ' not initialized'
-        assert aux_params[k].shape == aux_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
-
-    # prepare training
-    # create solver
-    data_names = [k[0] for k in train_data.provide_data]
-    label_names = [k[0] for k in train_data.provide_label]
-    if train_shared:
-        fixed_param_prefix = config.FIXED_PARAMS_SHARED
-    else:
-        fixed_param_prefix = config.FIXED_PARAMS
-    mod = MutableModule(sym, data_names=data_names, label_names=label_names,
-                        logger=logger, context=ctx, work_load_list=work_load_list,
-                        max_data_shapes=max_data_shape, fixed_param_prefix=fixed_param_prefix)
-
-    # decide training params
-    # metric
-    eval_metric = metric.RCNNAccMetric()
-    cls_metric = metric.RCNNLogLossMetric()
-    bbox_metric = metric.RCNNL1LossMetric()
-    eval_metrics = mx.metric.CompositeEvalMetric()
-    for child_metric in [eval_metric, cls_metric, bbox_metric]:
-        eval_metrics.add(child_metric)
-    # callback
-    batch_end_callback = mx.callback.Speedometer(train_data.batch_size, frequent=frequent, auto_reset=False)
-    epoch_end_callback = callback.do_checkpoint(prefix, means, stds)
-    # decide learning rate
-    base_lr = lr
-    lr_factor = 0.1
-    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
-    lr_epoch_diff = [epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch]
-    lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
-    lr_iters = [int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff]
-    logger.info('lr %f lr_epoch_diff %s lr_iters %s' % (lr, lr_epoch_diff, lr_iters))
-    lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
-    # optimizer
-    optimizer_params = {'momentum': 0.9,
-                        'wd': 0.0005,
-                        'learning_rate': lr,
-                        'lr_scheduler': lr_scheduler,
-                        'rescale_grad': (1.0 / batch_size),
-                        'clip_gradient': 5}
-
-    # train
-    mod.fit(train_data, eval_metric=eval_metrics, epoch_end_callback=epoch_end_callback,
-            batch_end_callback=batch_end_callback, kvstore=kvstore,
-            optimizer='sgd', optimizer_params=optimizer_params,
-            arg_params=arg_params, aux_params=aux_params, begin_epoch=begin_epoch, num_epoch=end_epoch)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Train a Fast R-CNN Network')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # training
-    parser.add_argument('--frequent', help='frequency of logging', default=default.frequent, type=int)
-    parser.add_argument('--kvstore', help='the kv-store type', default=default.kvstore, type=str)
-    parser.add_argument('--work_load_list', help='work load for different devices', default=None, type=list)
-    parser.add_argument('--no_flip', help='disable flip images', action='store_true')
-    parser.add_argument('--no_shuffle', help='disable random shuffle', action='store_true')
-    parser.add_argument('--resume', help='continue training', action='store_true')
-    # rcnn
-    parser.add_argument('--gpus', help='GPU device to train with', default='0', type=str)
-    parser.add_argument('--pretrained', help='pretrained model prefix', default=default.pretrained, type=str)
-    parser.add_argument('--pretrained_epoch', help='pretrained model epoch', default=default.pretrained_epoch, type=int)
-    parser.add_argument('--prefix', help='new model prefix', default=default.rcnn_prefix, type=str)
-    parser.add_argument('--begin_epoch', help='begin epoch of training', default=0, type=int)
-    parser.add_argument('--end_epoch', help='end epoch of training', default=default.rcnn_epoch, type=int)
-    parser.add_argument('--lr', help='base learning rate', default=default.rcnn_lr, type=float)
-    parser.add_argument('--lr_step', help='learning rate steps (in epoch)', default=default.rcnn_lr_step, type=str)
-    parser.add_argument('--train_shared', help='second round train shared params', action='store_true')
-    parser.add_argument('--proposal', help='can be ss for selective search or rpn', default='rpn', type=str)
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    logger.info('Called with argument: %s' % args)
-    ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
-    train_rcnn(args.network, args.dataset, args.image_set, args.root_path, args.dataset_path,
-               args.frequent, args.kvstore, args.work_load_list, args.no_flip, args.no_shuffle, args.resume,
-               ctx, args.pretrained, args.pretrained_epoch, args.prefix, args.begin_epoch, args.end_epoch,
-               train_shared=args.train_shared, lr=args.lr, lr_step=args.lr_step, proposal=args.proposal)
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFace/rcnn/tools/train_rpn.py b/RetinaFace/rcnn/tools/train_rpn.py
deleted file mode 100644
index 5a9e296b5..000000000
--- a/RetinaFace/rcnn/tools/train_rpn.py
+++ /dev/null
@@ -1,195 +0,0 @@
-import argparse
-import logging
-import pprint
-import mxnet as mx
-
-from ..config import config, default, generate_config
-from ..symbol import *
-from ..core import callback, metric
-from ..core.loader import AnchorLoaderFPN
-from ..core.module import MutableModule
-from ..utils.load_data import load_gt_roidb, merge_roidb, filter_roidb
-from ..utils.load_model import load_param
-
-
-def train_rpn(network, dataset, image_set, root_path, dataset_path,
-              frequent, kvstore, work_load_list, no_flip, no_shuffle, resume,
-              ctx, pretrained, epoch, prefix, begin_epoch, end_epoch,
-              train_shared, lr, lr_step):
-    # set up logger
-    logging.basicConfig()
-    logger = logging.getLogger()
-    logger.setLevel(logging.INFO)
-
-    # setup config
-    assert config.TRAIN.BATCH_IMAGES==1
-
-    # load symbol
-    sym = eval('get_' + network + '_rpn')()
-    feat_sym = []
-    for stride in config.RPN_FEAT_STRIDE:
-        feat_sym.append(sym.get_internals()['rpn_cls_score_stride%s_output' % stride])
-
-
-    # setup multi-gpu
-    batch_size = len(ctx)
-    input_batch_size = config.TRAIN.BATCH_IMAGES * batch_size
-
-    # print config
-    pprint.pprint(config)
-
-    # load dataset and prepare imdb for training
-    image_sets = [iset for iset in image_set.split('+')]
-    roidbs = [load_gt_roidb(dataset, image_set, root_path, dataset_path,
-                            flip=not no_flip)
-              for image_set in image_sets]
-    roidb = merge_roidb(roidbs)
-    roidb = filter_roidb(roidb)
-
-    # load training data
-    #train_data = AnchorLoaderFPN(feat_sym, roidb, batch_size=input_batch_size, shuffle=not no_shuffle,
-    #                              ctx=ctx, work_load_list=work_load_list,
-    #                              feat_stride=config.RPN_FEAT_STRIDE, anchor_scales=config.ANCHOR_SCALES,
-    #                              anchor_ratios=config.ANCHOR_RATIOS, aspect_grouping=config.TRAIN.ASPECT_GROUPING,
-    #                              allowed_border=9999)
-    train_data = AnchorLoaderFPN(feat_sym, roidb, batch_size=input_batch_size, shuffle=not no_shuffle,
-                                  ctx=ctx, work_load_list=work_load_list)
-
-    # infer max shape
-    max_data_shape = [('data', (input_batch_size, 3, max([v[0] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-    max_data_shape, max_label_shape = train_data.infer_shape(max_data_shape)
-    print 'providing maximum shape', max_data_shape, max_label_shape
-
-    # infer shape
-    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
-    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
-    out_shape_dict = zip(sym.list_outputs(), out_shape)
-    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
-    print 'output shape'
-    pprint.pprint(out_shape_dict)
-
-    # load and initialize params
-    if resume:
-        arg_params, aux_params = load_param(prefix, begin_epoch, convert=True)
-    else:
-        arg_params, aux_params = load_param(pretrained, epoch, convert=True)
-        init = mx.init.Xavier(factor_type="in", rnd_type='gaussian', magnitude=2)
-        init_internal = mx.init.Normal(sigma=0.01)
-        for k in sym.list_arguments():
-            if k in data_shape_dict:
-                continue
-            if k not in arg_params:
-                print 'init', k
-                arg_params[k] = mx.nd.zeros(shape=arg_shape_dict[k])
-                if not k.endswith('bias'):
-                    init_internal(k, arg_params[k])
-
-        for k in sym.list_auxiliary_states():
-            if k not in aux_params:
-                print 'init', k
-                aux_params[k] = mx.nd.zeros(shape=aux_shape_dict[k])
-                init(k, aux_params[k])
-
-    # check parameter shapes
-    for k in sym.list_arguments():
-        if k in data_shape_dict:
-            continue
-        assert k in arg_params, k + ' not initialized'
-        assert arg_params[k].shape == arg_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
-    for k in sym.list_auxiliary_states():
-        assert k in aux_params, k + ' not initialized'
-        assert aux_params[k].shape == aux_shape_dict[k], \
-            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
-
-    # create solver
-    data_names = [k[0] for k in train_data.provide_data]
-    label_names = [k[0] for k in train_data.provide_label]
-    if train_shared:
-        fixed_param_prefix = config.FIXED_PARAMS_SHARED
-    else:
-        fixed_param_prefix = config.FIXED_PARAMS
-    mod = MutableModule(sym, data_names=data_names, label_names=label_names,
-                        logger=logger, context=ctx, work_load_list=work_load_list,
-                        max_data_shapes=max_data_shape, max_label_shapes=max_label_shape,
-                        fixed_param_prefix=fixed_param_prefix)
-
-    # decide training params
-    # metric
-    eval_metric = metric.RPNAccMetric()
-    cls_metric = metric.RPNLogLossMetric()
-    bbox_metric = metric.RPNL1LossMetric()
-    eval_metrics = mx.metric.CompositeEvalMetric()
-    for child_metric in [eval_metric,cls_metric,bbox_metric]:
-        eval_metrics.add(child_metric)
-    # callback
-    batch_end_callback = []
-    batch_end_callback.append(mx.callback.Speedometer(train_data.batch_size, frequent=frequent))
-    epoch_end_callback = mx.callback.do_checkpoint(prefix)
-    # decide learning rate
-    base_lr = lr
-    lr_factor = 0.1
-    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
-    lr_epoch_diff = [epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch]
-    lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
-    lr_iters = [int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff]
-    print 'lr', lr, 'lr_epoch_diff', lr_epoch_diff, 'lr_iters', lr_iters
-    lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
-    # optimizer
-    optimizer_params = {'momentum': 0.9,
-                        'wd': 0.0001,
-                        'learning_rate': lr,
-                        'lr_scheduler': lr_scheduler,
-                        'rescale_grad': (1.0 / batch_size),
-                        'clip_gradient': 5}
-
-    # train
-    mod.fit(train_data, eval_metric=eval_metrics, epoch_end_callback=epoch_end_callback,
-            batch_end_callback=batch_end_callback, kvstore=kvstore,
-            optimizer='sgd', optimizer_params=optimizer_params,
-            arg_params=arg_params, aux_params=aux_params, begin_epoch=begin_epoch, num_epoch=end_epoch)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Train a Region Proposal Network')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # training
-    parser.add_argument('--frequent', help='frequency of logging', default=default.frequent, type=int)
-    parser.add_argument('--kvstore', help='the kv-store type', default=default.kvstore, type=str)
-    parser.add_argument('--work_load_list', help='work load for different devices', default=None, type=list)
-    parser.add_argument('--no_flip', help='disable flip images', action='store_true')
-    parser.add_argument('--no_shuffle', help='disable random shuffle', action='store_true')
-    parser.add_argument('--resume', help='continue training', action='store_true')
-    # rpn
-    parser.add_argument('--gpus', help='GPU device to train with', default='0', type=str)
-    parser.add_argument('--pretrained', help='pretrained model prefix', default=default.pretrained, type=str)
-    parser.add_argument('--pretrained_epoch', help='pretrained model epoch', default=default.pretrained_epoch, type=int)
-    parser.add_argument('--prefix', help='new model prefix', default=default.rpn_prefix, type=str)
-    parser.add_argument('--begin_epoch', help='begin epoch of training', default=0, type=int)
-    parser.add_argument('--end_epoch', help='end epoch of training', default=default.rpn_epoch, type=int)
-    parser.add_argument('--lr', help='base learning rate', default=default.rpn_lr, type=float)
-    parser.add_argument('--lr_step', help='learning rate steps (in epoch)', default=default.rpn_lr_step, type=str)
-    parser.add_argument('--train_shared', help='second round train shared params', action='store_true')
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    print 'Called with argument:', args
-    ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
-    train_rpn(args.network, args.dataset, args.image_set, args.root_path, args.dataset_path,
-              args.frequent, args.kvstore, args.work_load_list, args.no_flip, args.no_shuffle, args.resume,
-              ctx, args.pretrained, args.pretrained_epoch, args.prefix, args.begin_epoch, args.end_epoch,
-              train_shared=args.train_shared, lr=args.lr, lr_step=args.lr_step)
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFace/rcnn/utils/load_data.py b/RetinaFace/rcnn/utils/load_data.py
deleted file mode 100644
index 180814c5b..000000000
--- a/RetinaFace/rcnn/utils/load_data.py
+++ /dev/null
@@ -1,55 +0,0 @@
-import numpy as np
-from ..logger import logger
-from ..config import config
-from ..dataset import *
-
-
-def load_gt_roidb(dataset_name, image_set_name, root_path, dataset_path,
-                  flip=False):
-    """ load ground truth roidb """
-    imdb = eval(dataset_name)(image_set_name, root_path, dataset_path)
-    roidb = imdb.gt_roidb()
-    print('roidb size', len(roidb))
-    if flip:
-        roidb = imdb.append_flipped_images(roidb)
-    print('flipped roidb size', len(roidb))
-    return roidb
-
-
-def load_proposal_roidb(dataset_name, image_set_name, root_path, dataset_path,
-                        proposal='rpn', append_gt=True, flip=False):
-    """ load proposal roidb (append_gt when training) """
-    imdb = eval(dataset_name)(image_set_name, root_path, dataset_path)
-    gt_roidb = imdb.gt_roidb()
-    roidb = eval('imdb.' + proposal + '_roidb')(gt_roidb, append_gt)
-    if flip:
-        roidb = imdb.append_flipped_images(roidb)
-    return roidb
-
-
-def merge_roidb(roidbs):
-    """ roidb are list, concat them together """
-    roidb = roidbs[0]
-    for r in roidbs[1:]:
-        roidb.extend(r)
-    return roidb
-
-
-def filter_roidb(roidb):
-    """ remove roidb entries without usable rois """
-
-    def is_valid(entry):
-        """ valid images have at least 1 fg or bg roi """
-        overlaps = entry['max_overlaps']
-        fg_inds = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
-        bg_inds = np.where((overlaps < config.TRAIN.BG_THRESH_HI) & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
-        valid = len(fg_inds) > 0 or len(bg_inds) > 0
-        #valid = len(fg_inds) > 0
-        return valid
-
-    num = len(roidb)
-    filtered_roidb = [entry for entry in roidb if is_valid(entry)]
-    num_after = len(filtered_roidb)
-    logger.info('load data: filtered %d roidb entries: %d -> %d' % (num - num_after, num, num_after))
-
-    return filtered_roidb
diff --git a/RetinaFace/retinaface.py b/RetinaFace/retinaface.py
deleted file mode 100644
index 624452b9d..000000000
--- a/RetinaFace/retinaface.py
+++ /dev/null
@@ -1,698 +0,0 @@
-from __future__ import print_function
-import sys
-import os
-import datetime
-import time
-import numpy as np
-import mxnet as mx
-from mxnet import ndarray as nd
-import cv2
-#from rcnn import config
-from rcnn.logger import logger
-#from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
-from rcnn.processing.bbox_transform import clip_boxes
-from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
-from rcnn.processing.nms import gpu_nms_wrapper, cpu_nms_wrapper
-from rcnn.processing.bbox_transform import bbox_overlaps
-
-class RetinaFace:
-  def __init__(self, prefix, epoch, ctx_id=0, network='net3', nms=0.4, nocrop=False, decay4 = 0.5, vote=False):
-    self.ctx_id = ctx_id
-    self.network = network
-    self.decay4 = decay4
-    self.nms_threshold = nms
-    self.vote = vote
-    self.nocrop = nocrop
-    self.debug = False
-    self.fpn_keys = []
-    self.anchor_cfg = None
-    pixel_means=[0.0, 0.0, 0.0]
-    pixel_stds=[1.0, 1.0, 1.0]
-    pixel_scale = 1.0
-    self.preprocess = False
-    _ratio = (1.,)
-    fmc = 3
-    if network=='ssh' or network=='vgg':
-      pixel_means=[103.939, 116.779, 123.68]
-      self.preprocess = True
-    elif network=='net3':
-      _ratio = (1.,)
-    elif network=='net3a':
-      _ratio = (1.,1.5)
-    elif network=='net6': #like pyramidbox or s3fd
-      fmc = 6
-    elif network=='net5': #retinaface
-      fmc = 5
-    elif network=='net5a':
-      fmc = 5
-      _ratio = (1.,1.5)
-    elif network=='net4':
-      fmc = 4
-    elif network=='net4a':
-      fmc = 4
-      _ratio = (1.,1.5)
-    elif network=='x5':
-      fmc = 5
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    elif network=='x3':
-      fmc = 3
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    elif network=='x3a':
-      fmc = 3
-      _ratio = (1.,1.5)
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    else:
-      assert False, 'network setting error %s'%network
-
-    if fmc==3:
-      self._feat_stride_fpn = [32, 16, 8]
-      self.anchor_cfg = {
-          '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==4:
-      self._feat_stride_fpn = [32, 16, 8, 4]
-      self.anchor_cfg = {
-          '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '4': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==6:
-      self._feat_stride_fpn = [128, 64, 32, 16, 8, 4]
-      self.anchor_cfg = {
-          '128': {'SCALES': (32,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '64': {'SCALES': (16,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '32': {'SCALES': (8,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (4,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '4': {'SCALES': (1,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==5:
-      self._feat_stride_fpn = [64, 32, 16, 8, 4]
-      self.anchor_cfg = {}
-      _ass = 2.0**(1.0/3)
-      _basescale = 1.0
-      for _stride in [4, 8, 16, 32, 64]:
-        key = str(_stride)
-        value = {'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999}
-        scales = []
-        for _ in range(3):
-          scales.append(_basescale)
-          _basescale *= _ass
-        value['SCALES'] = tuple(scales)
-        self.anchor_cfg[key] = value
-
-    print(self._feat_stride_fpn, self.anchor_cfg)
-
-    for s in self._feat_stride_fpn:
-        self.fpn_keys.append('stride%s'%s)
-
-
-    dense_anchor = False
-    #self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(base_size=fpn_base_size, scales=self._scales, ratios=self._ratios)))
-    self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(dense_anchor=dense_anchor, cfg=self.anchor_cfg)))
-    for k in self._anchors_fpn:
-      v = self._anchors_fpn[k].astype(np.float32)
-      self._anchors_fpn[k] = v
-
-    self._num_anchors = dict(zip(self.fpn_keys, [anchors.shape[0] for anchors in self._anchors_fpn.values()]))
-    #self._bbox_pred = nonlinear_pred
-    #self._landmark_pred = landmark_pred
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    if self.ctx_id>=0:
-      self.ctx = mx.gpu(self.ctx_id)
-      self.nms = gpu_nms_wrapper(self.nms_threshold, self.ctx_id)
-    else:
-      self.ctx = mx.cpu()
-      self.nms = cpu_nms_wrapper(self.nms_threshold)
-    self.pixel_means = np.array(pixel_means, dtype=np.float32)
-    self.pixel_stds = np.array(pixel_stds, dtype=np.float32)
-    self.pixel_scale = float(pixel_scale)
-    print('means', self.pixel_means)
-    self.use_landmarks = False
-    if len(sym)//len(self._feat_stride_fpn)>=3:
-      self.use_landmarks = True
-    print('use_landmarks', self.use_landmarks)
-    self.cascade = 0
-    if float(len(sym))//len(self._feat_stride_fpn)>3.0:
-      self.cascade = 1
-    print('cascade', self.cascade)
-    #self.bbox_stds = [0.1, 0.1, 0.2, 0.2]
-    #self.landmark_std = 0.1
-    self.bbox_stds = [1.0, 1.0, 1.0, 1.0]
-    self.landmark_std = 1.0
-
-    if self.debug:
-      c = len(sym)//len(self._feat_stride_fpn)
-      sym = sym[(c*0):]
-      self._feat_stride_fpn = [32,16,8]
-    print('sym size:', len(sym))
-
-    image_size = (640, 640)
-    self.model = mx.mod.Module(symbol=sym, context=self.ctx, label_names = None)
-    self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
-    self.model.set_params(arg_params, aux_params)
-
-  def get_input(self, img):
-    im = img.astype(np.float32)
-    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-    for i in range(3):
-        im_tensor[0, i, :, :] = (im[:, :, 2 - i]/self.pixel_scale - self.pixel_means[2 - i])/self.pixel_stds[2-i]
-    #if self.debug:
-    #  timeb = datetime.datetime.now()
-    #  diff = timeb - timea
-    #  print('X2 uses', diff.total_seconds(), 'seconds')
-    data = nd.array(im_tensor)
-    return data
-
-  def detect(self, img, threshold=0.5, scales=[1.0], do_flip=False):
-    #print('in_detect', threshold, scales, do_flip, do_nms)
-    proposals_list = []
-    scores_list = []
-    landmarks_list = []
-    strides_list = []
-    timea = datetime.datetime.now()
-    flips = [0]
-    if do_flip:
-      flips = [0, 1]
-
-    imgs = [img]
-    if isinstance(img, list):
-      imgs = img
-    for img in imgs:
-      for im_scale in scales:
-        for flip in flips:
-          if im_scale!=1.0:
-            im = cv2.resize(img, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-          else:
-            im = img.copy()
-          if flip:
-            im = im[:,::-1,:]
-          if self.nocrop:
-            if im.shape[0]%32==0:
-              h = im.shape[0]
-            else:
-              h = (im.shape[0]//32+1)*32
-            if im.shape[1]%32==0:
-              w = im.shape[1]
-            else:
-              w = (im.shape[1]//32+1)*32
-            _im = np.zeros( (h, w, 3), dtype=np.float32 )
-            _im[0:im.shape[0], 0:im.shape[1], :] = im
-            im = _im
-          else:
-            im = im.astype(np.float32)
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X1 uses', diff.total_seconds(), 'seconds')
-          #self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
-          #im_info = [im.shape[0], im.shape[1], im_scale]
-          im_info = [im.shape[0], im.shape[1]]
-          im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-          for i in range(3):
-              im_tensor[0, i, :, :] = (im[:, :, 2 - i]/self.pixel_scale - self.pixel_means[2 - i])/self.pixel_stds[2-i]
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X2 uses', diff.total_seconds(), 'seconds')
-          data = nd.array(im_tensor)
-          db = mx.io.DataBatch(data=(data,), provide_data=[('data', data.shape)])
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X3 uses', diff.total_seconds(), 'seconds')
-          self.model.forward(db, is_train=False)
-          net_out = self.model.get_outputs()
-          #post_nms_topN = self._rpn_post_nms_top_n
-          #min_size_dict = self._rpn_min_size_fpn
-
-          sym_idx = 0
-
-          for _idx,s in enumerate(self._feat_stride_fpn):
-              #if len(scales)>1 and s==32 and im_scale==scales[-1]:
-              #  continue
-              _key = 'stride%s'%s
-              stride = int(s)
-              is_cascade = False
-              if self.cascade:
-                is_cascade = True
-              #if self.vote and stride==4 and len(scales)>2 and (im_scale==scales[0]):
-              #  continue
-              #print('getting', im_scale, stride, idx, len(net_out), data.shape, file=sys.stderr)
-              scores = net_out[sym_idx].asnumpy()
-              if self.debug:
-                timeb = datetime.datetime.now()
-                diff = timeb - timea
-                print('A uses', diff.total_seconds(), 'seconds')
-              #print(scores.shape)
-              #print('scores',stride, scores.shape, file=sys.stderr)
-              scores = scores[:, self._num_anchors['stride%s'%s]:, :, :]
-
-              bbox_deltas = net_out[sym_idx+1].asnumpy()
-
-              #if DEBUG:
-              #    print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
-              #    print 'scale: {}'.format(im_info[2])
-
-              #_height, _width = int(im_info[0] / stride), int(im_info[1] / stride)
-              height, width = bbox_deltas.shape[2], bbox_deltas.shape[3]
-
-              A = self._num_anchors['stride%s'%s]
-              K = height * width
-              anchors_fpn = self._anchors_fpn['stride%s'%s]
-              anchors = anchors_plane(height, width, stride, anchors_fpn)
-              #print((height, width), (_height, _width), anchors.shape, bbox_deltas.shape, scores.shape, file=sys.stderr)
-              anchors = anchors.reshape((K * A, 4))
-              #print('num_anchors', self._num_anchors['stride%s'%s], file=sys.stderr)
-              #print('HW', (height, width), file=sys.stderr)
-              #print('anchors_fpn', anchors_fpn.shape, file=sys.stderr)
-              #print('anchors', anchors.shape, file=sys.stderr)
-              #print('bbox_deltas', bbox_deltas.shape, file=sys.stderr)
-              #print('scores', scores.shape, file=sys.stderr)
-
-
-              #scores = self._clip_pad(scores, (height, width))
-              scores = scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
-
-              #print('pre', bbox_deltas.shape, height, width)
-              #bbox_deltas = self._clip_pad(bbox_deltas, (height, width))
-              #print('after', bbox_deltas.shape, height, width)
-              bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
-              bbox_pred_len = bbox_deltas.shape[3]//A
-              #print(bbox_deltas.shape)
-              bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
-              bbox_deltas[:, 0::4] = bbox_deltas[:,0::4] * self.bbox_stds[0]
-              bbox_deltas[:, 1::4] = bbox_deltas[:,1::4] * self.bbox_stds[1]
-              bbox_deltas[:, 2::4] = bbox_deltas[:,2::4] * self.bbox_stds[2]
-              bbox_deltas[:, 3::4] = bbox_deltas[:,3::4] * self.bbox_stds[3]
-              proposals = self.bbox_pred(anchors, bbox_deltas)
-
-
-              #print(anchors.shape, bbox_deltas.shape, A, K, file=sys.stderr)
-              if is_cascade:
-                cascade_sym_num = 0
-                cls_cascade = False
-                bbox_cascade = False
-                __idx = [3,4]
-                if not self.use_landmarks:
-                  __idx = [2,3]
-                for diff_idx in __idx:
-                  if sym_idx+diff_idx>=len(net_out):
-                    break
-                  body = net_out[sym_idx+diff_idx].asnumpy()
-                  if body.shape[1]//A==2: #cls branch
-                    if cls_cascade or bbox_cascade:
-                      break
-                    else:
-                      cascade_scores = body[:, self._num_anchors['stride%s'%s]:, :, :]
-                      cascade_scores = cascade_scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
-                      #scores = (scores+cascade_scores)/2.0
-                      scores = cascade_scores #TODO?
-                      cascade_sym_num+=1
-                      cls_cascade = True
-                      #print('find cascade cls at stride', stride)
-                  elif body.shape[1]//A==4: #bbox branch
-                    cascade_deltas = body.transpose((0,2,3,1)).reshape( (-1, bbox_pred_len) )
-                    cascade_deltas[:, 0::4] = cascade_deltas[:,0::4] * self.bbox_stds[0]
-                    cascade_deltas[:, 1::4] = cascade_deltas[:,1::4] * self.bbox_stds[1]
-                    cascade_deltas[:, 2::4] = cascade_deltas[:,2::4] * self.bbox_stds[2]
-                    cascade_deltas[:, 3::4] = cascade_deltas[:,3::4] * self.bbox_stds[3]
-                    proposals = self.bbox_pred(proposals, cascade_deltas)
-                    cascade_sym_num+=1
-                    bbox_cascade = True
-                    #print('find cascade bbox at stride', stride)
-
-
-              proposals = clip_boxes(proposals, im_info[:2])
-
-              #if self.vote:
-              #  if im_scale>1.0:
-              #    keep = self._filter_boxes2(proposals, 160*im_scale, -1)
-              #  else:
-              #    keep = self._filter_boxes2(proposals, -1, 100*im_scale)
-              #  if stride==4:
-              #    keep = self._filter_boxes2(proposals, 12*im_scale, -1)
-              #    proposals = proposals[keep, :]
-              #    scores = scores[keep]
-
-              #keep = self._filter_boxes(proposals, min_size_dict['stride%s'%s] * im_info[2])
-              #proposals = proposals[keep, :]
-              #scores = scores[keep]
-              #print('333', proposals.shape)
-              if stride==4 and self.decay4<1.0:
-                scores *= self.decay4
-
-              scores_ravel = scores.ravel()
-              #print('__shapes', proposals.shape, scores_ravel.shape)
-              #print('max score', np.max(scores_ravel))
-              order = np.where(scores_ravel>=threshold)[0]
-                #_scores = scores_ravel[order]
-                #_order = _scores.argsort()[::-1]
-                #order = order[_order]
-              proposals = proposals[order, :]
-              scores = scores[order]
-              if flip:
-                oldx1 = proposals[:, 0].copy()
-                oldx2 = proposals[:, 2].copy()
-                proposals[:, 0] = im.shape[1] - oldx2 - 1
-                proposals[:, 2] = im.shape[1] - oldx1 - 1
-
-              proposals[:,0:4] /= im_scale
-
-              proposals_list.append(proposals)
-              scores_list.append(scores)
-              if self.nms_threshold<0.0:
-                _strides = np.empty(shape=(scores.shape), dtype=np.float32)
-                _strides.fill(stride)
-                strides_list.append(_strides)
-
-              if not self.vote and self.use_landmarks:
-                landmark_deltas = net_out[sym_idx+2].asnumpy()
-                #landmark_deltas = self._clip_pad(landmark_deltas, (height, width))
-                landmark_pred_len = landmark_deltas.shape[1]//A
-                landmark_deltas = landmark_deltas.transpose((0, 2, 3, 1)).reshape((-1, 5, landmark_pred_len//5))
-                landmark_deltas *= self.landmark_std
-                #print(landmark_deltas.shape, landmark_deltas)
-                landmarks = self.landmark_pred(anchors, landmark_deltas)
-                landmarks = landmarks[order, :]
-
-                if flip:
-                  landmarks[:,:,0] = im.shape[1] - landmarks[:,:,0] - 1
-                  #for a in range(5):
-                  #  oldx1 = landmarks[:, a].copy()
-                  #  landmarks[:,a] = im.shape[1] - oldx1 - 1
-                  order = [1,0,2,4,3]
-                  flandmarks = landmarks.copy()
-                  for idx, a in enumerate(order):
-                    flandmarks[:,idx,:] = landmarks[:,a,:]
-                    #flandmarks[:, idx*2] = landmarks[:,a*2]
-                    #flandmarks[:, idx*2+1] = landmarks[:,a*2+1]
-                  landmarks = flandmarks
-                landmarks[:,:,0:2] /= im_scale
-                #landmarks /= im_scale
-                #landmarks = landmarks.reshape( (-1, landmark_pred_len) )
-                landmarks_list.append(landmarks)
-                #proposals = np.hstack((proposals, landmarks))
-              if self.use_landmarks:
-                sym_idx += 3
-              else:
-                sym_idx += 2
-              if is_cascade:
-                sym_idx += cascade_sym_num
-
-    if self.debug:
-      timeb = datetime.datetime.now()
-      diff = timeb - timea
-      print('B uses', diff.total_seconds(), 'seconds')
-    proposals = np.vstack(proposals_list)
-    landmarks = None
-    if proposals.shape[0]==0:
-      if self.use_landmarks:
-        landmarks = np.zeros( (0,5,2) )
-      if self.nms_threshold<0.0:
-        return np.zeros( (0,6) ), landmarks
-      else:
-        return np.zeros( (0,5) ), landmarks
-    scores = np.vstack(scores_list)
-    #print('shapes', proposals.shape, scores.shape)
-    scores_ravel = scores.ravel()
-    order = scores_ravel.argsort()[::-1]
-    #if config.TEST.SCORE_THRESH>0.0:
-    #  _count = np.sum(scores_ravel>config.TEST.SCORE_THRESH)
-    #  order = order[:_count]
-    proposals = proposals[order, :]
-    scores = scores[order]
-    if self.nms_threshold<0.0:
-      strides = np.vstack(strides_list)
-      strides = strides[order]
-    if not self.vote and self.use_landmarks:
-      landmarks = np.vstack(landmarks_list)
-      landmarks = landmarks[order].astype(np.float32, copy=False)
-
-    if self.nms_threshold>0.0:
-      pre_det = np.hstack((proposals[:,0:4], scores)).astype(np.float32, copy=False)
-      if not self.vote:
-        keep = self.nms(pre_det)
-        det = np.hstack( (pre_det, proposals[:,4:]) )
-        det = det[keep, :]
-        if self.use_landmarks:
-          landmarks = landmarks[keep]
-      else:
-        det = np.hstack( (pre_det, proposals[:,4:]) )
-        det = self.bbox_vote(det)
-    elif self.nms_threshold<0.0:
-      det = np.hstack((proposals[:,0:4], scores, strides)).astype(np.float32, copy=False)
-    else:
-      det = np.hstack((proposals[:,0:4], scores)).astype(np.float32, copy=False)
-
-
-    if self.debug:
-      timeb = datetime.datetime.now()
-      diff = timeb - timea
-      print('C uses', diff.total_seconds(), 'seconds')
-    return det, landmarks
-
-  def detect_center(self, img, threshold=0.5, scales=[1.0], do_flip=False):
-    det, landmarks = self.detect(img, threshold, scales, do_flip)
-    if det.shape[0]==0:
-      return None, None
-    bindex = 0
-    if det.shape[0]>1:
-      img_size = np.asarray(img.shape)[0:2]
-      bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-      img_center = img_size / 2
-      offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-      offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-      bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-    bbox = det[bindex,:]
-    landmark = landmarks[bindex, :, :]
-    return bbox, landmark
-
-  @staticmethod
-  def check_large_pose(landmark, bbox):
-    assert landmark.shape==(5,2)
-    assert len(bbox)==4
-    def get_theta(base, x, y):
-      vx = x-base
-      vy = y-base
-      vx[1] *= -1
-      vy[1] *= -1
-      tx = np.arctan2(vx[1], vx[0])
-      ty = np.arctan2(vy[1], vy[0])
-      d = ty-tx
-      d = np.degrees(d)
-      #print(vx, tx, vy, ty, d)
-      #if d<-1.*math.pi:
-      #  d+=2*math.pi
-      #elif d>math.pi:
-      #  d-=2*math.pi
-      if d<-180.0:
-        d+=360.
-      elif d>180.0:
-        d-=360.0
-      return d
-    landmark = landmark.astype(np.float32)
-
-    theta1 = get_theta(landmark[0], landmark[3], landmark[2])
-    theta2 = get_theta(landmark[1], landmark[2], landmark[4])
-    #print(va, vb, theta2)
-    theta3 = get_theta(landmark[0], landmark[2], landmark[1])
-    theta4 = get_theta(landmark[1], landmark[0], landmark[2])
-    theta5 = get_theta(landmark[3], landmark[4], landmark[2])
-    theta6 = get_theta(landmark[4], landmark[2], landmark[3])
-    theta7 = get_theta(landmark[3], landmark[2], landmark[0])
-    theta8 = get_theta(landmark[4], landmark[1], landmark[2])
-    #print(theta1, theta2, theta3, theta4, theta5, theta6, theta7, theta8)
-    left_score = 0.0
-    right_score = 0.0
-    up_score = 0.0
-    down_score = 0.0
-    if theta1<=0.0:
-      left_score = 10.0
-    elif theta2<=0.0:
-      right_score = 10.0
-    else:
-      left_score = theta2/theta1
-      right_score = theta1/theta2
-    if theta3<=10.0 or theta4<=10.0:
-      up_score = 10.0
-    else:
-      up_score = max(theta1/theta3, theta2/theta4)
-    if theta5<=10.0 or theta6<=10.0:
-      down_score = 10.0
-    else:
-      down_score = max(theta7/theta5, theta8/theta6)
-    mleft = (landmark[0][0]+landmark[3][0])/2
-    mright = (landmark[1][0]+landmark[4][0])/2
-    box_center = ( (bbox[0]+bbox[2])/2,  (bbox[1]+bbox[3])/2 )
-    ret = 0
-    if left_score>=3.0:
-      ret = 1
-    if ret==0 and left_score>=2.0:
-      if mright<=box_center[0]:
-        ret = 1
-    if ret==0 and right_score>=3.0:
-      ret = 2
-    if ret==0 and right_score>=2.0:
-      if mleft>=box_center[0]:
-        ret = 2
-    if ret==0 and up_score>=2.0:
-      ret = 3
-    if ret==0 and down_score>=5.0:
-      ret = 4
-    return ret, left_score, right_score, up_score, down_score
-
-  @staticmethod
-  def _filter_boxes(boxes, min_size):
-      """ Remove all boxes with any side smaller than min_size """
-      ws = boxes[:, 2] - boxes[:, 0] + 1
-      hs = boxes[:, 3] - boxes[:, 1] + 1
-      keep = np.where((ws >= min_size) & (hs >= min_size))[0]
-      return keep
-
-  @staticmethod
-  def _filter_boxes2(boxes, max_size, min_size):
-      """ Remove all boxes with any side smaller than min_size """
-      ws = boxes[:, 2] - boxes[:, 0] + 1
-      hs = boxes[:, 3] - boxes[:, 1] + 1
-      if max_size>0:
-        keep = np.where( np.minimum(ws, hs)<max_size )[0]
-      elif min_size>0:
-        keep = np.where( np.maximum(ws, hs)>min_size )[0]
-      return keep
-
-  @staticmethod
-  def _clip_pad(tensor, pad_shape):
-      """
-      Clip boxes of the pad area.
-      :param tensor: [n, c, H, W]
-      :param pad_shape: [h, w]
-      :return: [n, c, h, w]
-      """
-      H, W = tensor.shape[2:]
-      h, w = pad_shape
-
-      if h < H or w < W:
-        tensor = tensor[:, :, :h, :w].copy()
-
-      return tensor
-
-  @staticmethod
-  def bbox_pred(boxes, box_deltas):
-      """
-      Transform the set of class-agnostic boxes into class-specific boxes
-      by applying the predicted offsets (box_deltas)
-      :param boxes: !important [N 4]
-      :param box_deltas: [N, 4 * num_classes]
-      :return: [N 4 * num_classes]
-      """
-      if boxes.shape[0] == 0:
-          return np.zeros((0, box_deltas.shape[1]))
-
-      boxes = boxes.astype(np.float, copy=False)
-      widths = boxes[:, 2] - boxes[:, 0] + 1.0
-      heights = boxes[:, 3] - boxes[:, 1] + 1.0
-      ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-      ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-      dx = box_deltas[:, 0:1]
-      dy = box_deltas[:, 1:2]
-      dw = box_deltas[:, 2:3]
-      dh = box_deltas[:, 3:4]
-
-      pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-      pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-      pred_w = np.exp(dw) * widths[:, np.newaxis]
-      pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-      pred_boxes = np.zeros(box_deltas.shape)
-      # x1
-      pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-      # y1
-      pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-      # x2
-      pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-      # y2
-      pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-
-      if box_deltas.shape[1]>4:
-        pred_boxes[:,4:] = box_deltas[:,4:]
-
-      return pred_boxes
-
-  @staticmethod
-  def landmark_pred(boxes, landmark_deltas):
-      if boxes.shape[0] == 0:
-          return np.zeros((0, landmark_deltas.shape[1]))
-      boxes = boxes.astype(np.float, copy=False)
-      widths = boxes[:, 2] - boxes[:, 0] + 1.0
-      heights = boxes[:, 3] - boxes[:, 1] + 1.0
-      ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-      ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-      pred = landmark_deltas.copy()
-      for i in range(5):
-        pred[:,i,0] = landmark_deltas[:,i,0]*widths + ctr_x
-        pred[:,i,1] = landmark_deltas[:,i,1]*heights + ctr_y
-      return pred
-      #preds = []
-      #for i in range(landmark_deltas.shape[1]):
-      #  if i%2==0:
-      #    pred = (landmark_deltas[:,i]*widths + ctr_x)
-      #  else:
-      #    pred = (landmark_deltas[:,i]*heights + ctr_y)
-      #  preds.append(pred)
-      #preds = np.vstack(preds).transpose()
-      #return preds
-
-  def bbox_vote(self, det):
-      #order = det[:, 4].ravel().argsort()[::-1]
-      #det = det[order, :]
-      if det.shape[0] == 0:
-        return np.zeros( (0, 5) )
-          #dets = np.array([[10, 10, 20, 20, 0.002]])
-          #det = np.empty(shape=[0, 5])
-      dets = None
-      while det.shape[0] > 0:
-          if dets is not None and dets.shape[0]>=750:
-              break
-          # IOU
-          area = (det[:, 2] - det[:, 0] + 1) * (det[:, 3] - det[:, 1] + 1)
-          xx1 = np.maximum(det[0, 0], det[:, 0])
-          yy1 = np.maximum(det[0, 1], det[:, 1])
-          xx2 = np.minimum(det[0, 2], det[:, 2])
-          yy2 = np.minimum(det[0, 3], det[:, 3])
-          w = np.maximum(0.0, xx2 - xx1 + 1)
-          h = np.maximum(0.0, yy2 - yy1 + 1)
-          inter = w * h
-          o = inter / (area[0] + area[:] - inter)
-
-          # nms
-          merge_index = np.where(o >= self.nms_threshold)[0]
-          det_accu = det[merge_index, :]
-          det = np.delete(det, merge_index, 0)
-          if merge_index.shape[0] <= 1:
-              if det.shape[0] == 0:
-                  try:
-                      dets = np.row_stack((dets, det_accu))
-                  except:
-                      dets = det_accu
-              continue
-          det_accu[:, 0:4] = det_accu[:, 0:4] * np.tile(det_accu[:, -1:], (1, 4))
-          max_score = np.max(det_accu[:, 4])
-          det_accu_sum = np.zeros((1, 5))
-          det_accu_sum[:, 0:4] = np.sum(det_accu[:, 0:4],
-                                        axis=0) / np.sum(det_accu[:, -1:])
-          det_accu_sum[:, 4] = max_score
-          if dets is None:
-              dets = det_accu_sum
-          else:
-              dets = np.row_stack((dets, det_accu_sum))
-      dets = dets[0:750, :]
-      return dets
-
diff --git a/RetinaFace/test.py b/RetinaFace/test.py
deleted file mode 100644
index 748fd013a..000000000
--- a/RetinaFace/test.py
+++ /dev/null
@@ -1,60 +0,0 @@
-import cv2
-import sys
-import numpy as np
-import datetime
-import os
-import glob
-from retinaface import RetinaFace
-
-thresh = 0.8
-scales = [1024, 1980]
-
-count = 1
-
-gpuid = 0
-detector = RetinaFace('./model/R50', 0, gpuid, 'net3')
-
-img = cv2.imread('t1.jpg')
-print(img.shape)
-im_shape = img.shape
-target_size = scales[0]
-max_size = scales[1]
-im_size_min = np.min(im_shape[0:2])
-im_size_max = np.max(im_shape[0:2])
-#im_scale = 1.0
-#if im_size_min>target_size or im_size_max>max_size:
-im_scale = float(target_size) / float(im_size_min)
-# prevent bigger axis from being more than max_size:
-if np.round(im_scale * im_size_max) > max_size:
-    im_scale = float(max_size) / float(im_size_max)
-
-print('im_scale', im_scale)
-
-scales = [im_scale]
-flip = False
-
-for c in range(count):
-  faces, landmarks = detector.detect(img, thresh, scales=scales, do_flip=flip)
-  print(c, faces.shape, landmarks.shape)
-
-if faces is not None:
-  print('find', faces.shape[0], 'faces')
-  for i in range(faces.shape[0]):
-    #print('score', faces[i][4])
-    box = faces[i].astype(np.int)
-    #color = (255,0,0)
-    color = (0,0,255)
-    cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
-    if landmarks is not None:
-      landmark5 = landmarks[i].astype(np.int)
-      #print(landmark.shape)
-      for l in range(landmark5.shape[0]):
-        color = (0,0,255)
-        if l==0 or l==3:
-          color = (0,255,0)
-        cv2.circle(img, (landmark5[l][0], landmark5[l][1]), 1, color, 2)
-
-  filename = './detector_test.jpg'
-  print('writing', filename)
-  cv2.imwrite(filename, img)
-
diff --git a/RetinaFace/test_widerface.py b/RetinaFace/test_widerface.py
deleted file mode 100644
index d0a3eeba8..000000000
--- a/RetinaFace/test_widerface.py
+++ /dev/null
@@ -1,201 +0,0 @@
-from __future__ import print_function
-
-import argparse
-import sys
-import os
-import time
-import numpy as np
-import mxnet as mx
-from mxnet import ndarray as nd
-import cv2
-from rcnn.logger import logger
-#from rcnn.config import config, default, generate_config
-#from rcnn.tools.test_rcnn import test_rcnn
-#from rcnn.tools.test_rpn import test_rpn
-from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
-from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
-from rcnn.processing.nms import gpu_nms_wrapper
-from rcnn.processing.bbox_transform import bbox_overlaps
-from rcnn.dataset import retinaface
-from retinaface import RetinaFace
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Test widerface by retinaface detector')
-    # general
-    parser.add_argument('--network', help='network name', default='net3', type=str)
-    parser.add_argument('--dataset', help='dataset name', default='retinaface', type=str)
-    parser.add_argument('--image-set', help='image_set name', default='val', type=str)
-    parser.add_argument('--root-path', help='output data folder', default='./data', type=str)
-    parser.add_argument('--dataset-path', help='dataset path', default='./data/retinaface', type=str)
-    parser.add_argument('--gpu', help='GPU device to test with', default=0, type=int)
-    # testing
-    parser.add_argument('--prefix', help='model to test with', default='', type=str)
-    parser.add_argument('--epoch', help='model to test with', default=0, type=int)
-    parser.add_argument('--output', help='output folder', default='./wout', type=str)
-    parser.add_argument('--nocrop', help='', action='store_true')
-    parser.add_argument('--thresh', help='valid detection threshold', default=0.02, type=float)
-    parser.add_argument('--mode', help='test mode, 0 for fast, 1 for accurate', default=1, type=int)
-    #parser.add_argument('--pyramid', help='enable pyramid test', action='store_true')
-    #parser.add_argument('--bbox-vote', help='', action='store_true')
-    parser.add_argument('--part', help='', default=0, type=int)
-    parser.add_argument('--parts', help='', default=1, type=int)
-    args = parser.parse_args()
-    return args
-
-detector = None
-args = None
-imgid = -1
-
-def get_boxes(roi, pyramid):
-  global imgid
-  im = cv2.imread(roi['image'])
-  do_flip = False
-  if not pyramid:
-    target_size = 1200
-    max_size = 1600
-    #do_flip = True
-    target_size = 1504
-    max_size = 2000
-    target_size = 1600
-    max_size = 2150
-    im_shape = im.shape
-    im_size_min = np.min(im_shape[0:2])
-    im_size_max = np.max(im_shape[0:2])
-    im_scale = float(target_size) / float(im_size_min)
-    # prevent bigger axis from being more than max_size:
-    if np.round(im_scale * im_size_max) > max_size:
-        im_scale = float(max_size) / float(im_size_max)
-    scales = [im_scale]
-  else:
-    do_flip = True
-    #TEST_SCALES = [500, 800, 1200, 1600]
-    TEST_SCALES = [500, 800, 1100, 1400, 1700]
-    target_size = 800
-    max_size = 1200
-    im_shape = im.shape
-    im_size_min = np.min(im_shape[0:2])
-    im_size_max = np.max(im_shape[0:2])
-    im_scale = float(target_size) / float(im_size_min)
-    # prevent bigger axis from being more than max_size:
-    if np.round(im_scale * im_size_max) > max_size:
-        im_scale = float(max_size) / float(im_size_max)
-    scales = [float(scale)/target_size*im_scale for scale in TEST_SCALES]
-  boxes, landmarks = detector.detect(im, threshold=args.thresh, scales = scales, do_flip=do_flip)
-  #print(boxes.shape, landmarks.shape)
-  if imgid>=0 and imgid<100:
-    font = cv2.FONT_HERSHEY_SIMPLEX
-    for i in range(boxes.shape[0]):
-      box = boxes[i]
-      ibox = box[0:4].copy().astype(np.int)
-      cv2.rectangle(im, (ibox[0], ibox[1]), (ibox[2], ibox[3]), (255, 0, 0), 2)
-      #print('box', ibox)
-      #if len(ibox)>5:
-      #  for l in range(5):
-      #    pp = (ibox[5+l*2], ibox[6+l*2])
-      #    cv2.circle(im, (pp[0], pp[1]), 1, (0, 0, 255), 1)
-      blur = box[5]
-      k = "%.3f"%blur
-      cv2.putText(im,k,(ibox[0]+2,ibox[1]+14), font, 0.6, (0,255,0), 2)
-      #landmarks = box[6:21].reshape( (5,3) )
-      if landmarks is not None:
-        for l in range(5):
-          color = (0,255,0)
-          landmark = landmarks[i][l]
-          pp = (int(landmark[0]), int(landmark[1]))
-          if landmark[2]-0.5<0.0:
-            color = (0,0,255)
-          cv2.circle(im, (pp[0], pp[1]), 1, color, 2)
-    filename = './testimages/%d.jpg'%imgid
-    cv2.imwrite(filename, im)
-    print(filename, 'wrote')
-    imgid+=1
-
-  return boxes
-
-
-def test(args):
-  print('test with', args)
-  global detector
-  output_folder = args.output
-  if not os.path.exists(output_folder):
-    os.mkdir(output_folder)
-  detector = RetinaFace(args.prefix, args.epoch, args.gpu, network=args.network, nocrop=args.nocrop, vote=args.bbox_vote)
-  imdb = eval(args.dataset)(args.image_set, args.root_path, args.dataset_path)
-  roidb = imdb.gt_roidb()
-  gt_overlaps = np.zeros(0)
-  overall = [0.0, 0.0]
-  gt_max = np.array( (0.0, 0.0) )
-  num_pos = 0
-  print('roidb size', len(roidb))
-
-  for i in range(len(roidb)):
-    if i%args.parts!=args.part:
-      continue
-    #if i%10==0:
-    #  print('processing', i, file=sys.stderr)
-    roi = roidb[i]
-    boxes = get_boxes(roi, args.pyramid)
-    if 'boxes' in roi:
-      gt_boxes = roi['boxes'].copy()
-      gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] + 1) * (gt_boxes[:, 3] - gt_boxes[:, 1] + 1)
-      num_pos += gt_boxes.shape[0]
-
-      overlaps = bbox_overlaps(boxes.astype(np.float), gt_boxes.astype(np.float))
-      #print(im_info, gt_boxes.shape, boxes.shape, overlaps.shape, file=sys.stderr)
-
-      _gt_overlaps = np.zeros((gt_boxes.shape[0]))
-
-      if boxes.shape[0]>0:
-        _gt_overlaps = overlaps.max(axis=0)
-        #print('max_overlaps', _gt_overlaps, file=sys.stderr)
-        for j in range(len(_gt_overlaps)):
-          if _gt_overlaps[j]>0.5:
-            continue
-          #print(j, 'failed', gt_boxes[j],  'max_overlap:', _gt_overlaps[j], file=sys.stderr)
-
-        # append recorded IoU coverage level
-        found = (_gt_overlaps > 0.5).sum()
-        recall = found / float(gt_boxes.shape[0])
-        #print('recall', _recall, gt_boxes.shape[0], boxes.shape[0], gt_areas, 'num:', i, file=sys.stderr)
-        overall[0]+=found
-        overall[1]+=gt_boxes.shape[0]
-        #gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
-        #_recall = (gt_overlaps >= threshold).sum() / float(num_pos)
-        recall_all = float(overall[0])/overall[1]
-        #print('recall_all', _recall, file=sys.stderr)
-        print('[%d]'%i, 'recall', recall, (gt_boxes.shape[0], boxes.shape[0]), 'all:', recall_all, file=sys.stderr)
-    else:
-      print('[%d]'%i, 'detect %d faces'%boxes.shape[0])
-
-
-    _vec = roidb[i]['image'].split('/')
-    out_dir = os.path.join(output_folder, _vec[-2])
-    if not os.path.exists(out_dir):
-        os.mkdir(out_dir)
-    out_file = os.path.join(out_dir, _vec[-1].replace('jpg', 'txt'))
-    with open(out_file, 'w') as f:
-      name = '/'.join(roidb[i]['image'].split('/')[-2:])
-      f.write("%s\n"%(name))
-      f.write("%d\n"%(boxes.shape[0]))
-      for b in range(boxes.shape[0]):
-        box = boxes[b]
-        f.write("%d %d %d %d %g \n"%(box[0], box[1], box[2]-box[0], box[3]-box[1], box[4]))
-
-def main():
-    global args
-    args = parse_args()
-    args.pyramid = False
-    args.bbox_vote = False
-    if args.mode==1:
-      args.pyramid = True
-      args.bbox_vote = True
-    elif args.mode==2:
-      args.pyramid = True
-      args.bbox_vote = False
-    logger.info('Called with argument: %s' % args)
-    test(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/RetinaFace/train.py b/RetinaFace/train.py
deleted file mode 100644
index 3d17ffc75..000000000
--- a/RetinaFace/train.py
+++ /dev/null
@@ -1,372 +0,0 @@
-from __future__ import print_function
-import sys
-import argparse
-import os
-import pprint
-import re
-import mxnet as mx
-import numpy as np
-from mxnet.module import Module
-import mxnet.optimizer as optimizer
-
-from rcnn.logger import logger
-from rcnn.config import config, default, generate_config
-from rcnn.symbol import *
-from rcnn.core import callback, metric
-from rcnn.core.loader import CropLoader, CropLoader2
-from rcnn.core.module import MutableModule
-from rcnn.utils.load_data import load_gt_roidb, merge_roidb, filter_roidb
-from rcnn.utils.load_model import load_param
-
-
-def get_fixed_params(symbol, fixed_param):
-    if not config.LAYER_FIX:
-      return []
-    fixed_param_names = []
-    #for name in symbol.list_arguments():
-    #  for f in fixed_param:
-    #    if re.match(f, name):
-    #      fixed_param_names.append(name)
-    #pre = 'mobilenetv20_features_linearbottleneck'
-    idx = 0
-    for name in symbol.list_arguments():
-      #print(idx, name)
-      if idx<7 and name!='data':
-        fixed_param_names.append(name)
-      #elif name.startswith('stage1_'):
-      #  fixed_param_names.append(name)
-      if name.find('upsampling')>=0:
-        fixed_param_names.append(name)
-
-      idx+=1
-    return fixed_param_names
-
-def train_net(args, ctx, pretrained, epoch, prefix, begin_epoch, end_epoch,
-              lr=0.001, lr_step='5'):
-    # setup config
-    #init_config()
-    #print(config)
-    # setup multi-gpu
-
-    input_batch_size = config.TRAIN.BATCH_IMAGES * len(ctx)
-
-    # print config
-    logger.info(pprint.pformat(config))
-
-    # load dataset and prepare imdb for training
-    image_sets = [iset for iset in args.image_set.split('+')]
-    roidbs = [load_gt_roidb(args.dataset, image_set, args.root_path, args.dataset_path,
-                            flip=not args.no_flip)
-              for image_set in image_sets]
-    #roidb = merge_roidb(roidbs)
-    #roidb = filter_roidb(roidb)
-    roidb = roidbs[0]
-
-    # load symbol
-    #sym = eval('get_' + args.network + '_train')(num_classes=config.NUM_CLASSES, num_anchors=config.NUM_ANCHORS)
-    #feat_sym = sym.get_internals()['rpn_cls_score_output']
-    #train_data = AnchorLoader(feat_sym, roidb, batch_size=input_batch_size, shuffle=not args.no_shuffle,
-    #                          ctx=ctx, work_load_list=args.work_load_list,
-    #                          feat_stride=config.RPN_FEAT_STRIDE, anchor_scales=config.ANCHOR_SCALES,
-    #                          anchor_ratios=config.ANCHOR_RATIOS, aspect_grouping=config.TRAIN.ASPECT_GROUPING)
-
-    # load and initialize params
-    sym = None
-    if len(pretrained)==0:
-        arg_params = {}
-        aux_params = {}
-    else:
-        logger.info('loading %s,%d'%(pretrained, epoch))
-        sym, arg_params, aux_params = mx.model.load_checkpoint(pretrained, epoch)
-        #arg_params, aux_params = load_param(pretrained, epoch, convert=True)
-        #for k in ['rpn_conv_3x3', 'rpn_cls_score', 'rpn_bbox_pred', 'cls_score', 'bbox_pred']:
-        #  _k = k+"_weight"
-        #  if _k in arg_shape_dict:
-        #    v = 0.001 if _k.startswith('bbox_') else 0.01
-        #    arg_params[_k] = mx.random.normal(0, v, shape=arg_shape_dict[_k])
-        #    print('init %s with normal %.5f'%(_k,v))
-        #  _k = k+"_bias"
-        #  if _k in arg_shape_dict:
-        #    arg_params[_k] = mx.nd.zeros(shape=arg_shape_dict[_k])
-        #    print('init %s with zero'%(_k))
-
-    sym = eval('get_' + args.network + '_train')(sym)
-    #print(sym.get_internals())
-    feat_sym = []
-    for stride in config.RPN_FEAT_STRIDE:
-        feat_sym.append(sym.get_internals()['face_rpn_cls_score_stride%s_output' % stride])
-
-
-
-    train_data = CropLoader(feat_sym, roidb, batch_size=input_batch_size, shuffle=not args.no_shuffle,
-                                  ctx=ctx, work_load_list=args.work_load_list)
-
-
-    # infer max shape
-    max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-    #max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
-    max_data_shape, max_label_shape = train_data.infer_shape(max_data_shape)
-    max_data_shape.append(('gt_boxes', (1, roidb[0]['max_num_boxes'], 5)))
-    logger.info('providing maximum shape %s %s' % (max_data_shape, max_label_shape))
-
-    # infer shape
-    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
-    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
-    out_shape_dict = dict(zip(sym.list_outputs(), out_shape))
-    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
-    logger.info('output shape %s' % pprint.pformat(out_shape_dict))
-
-
-    for k in arg_shape_dict:
-      v = arg_shape_dict[k]
-      if k.find('upsampling')>=0:
-        print('initializing upsampling_weight', k)
-        arg_params[k] = mx.nd.zeros(shape=v)
-        init = mx.init.Initializer()
-        init._init_bilinear(k, arg_params[k])
-            #print(args[k])
-
-    # check parameter shapes
-    #for k in sym.list_arguments():
-    #    if k in data_shape_dict:
-    #        continue
-    #    assert k in arg_params, k + ' not initialized'
-    #    assert arg_params[k].shape == arg_shape_dict[k], \
-    #        'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
-    #for k in sym.list_auxiliary_states():
-    #    assert k in aux_params, k + ' not initialized'
-    #    assert aux_params[k].shape == aux_shape_dict[k], \
-    #        'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
-
-    fixed_param_prefix = config.FIXED_PARAMS
-    # create solver
-    data_names = [k[0] for k in train_data.provide_data]
-    label_names = [k[0] for k in train_data.provide_label]
-    fixed_param_names = get_fixed_params(sym, fixed_param_prefix)
-    print('fixed', fixed_param_names, file=sys.stderr)
-    mod = Module(sym, data_names=data_names, label_names=label_names,
-                        logger=logger, context=ctx, work_load_list=args.work_load_list,
-                        fixed_param_names=fixed_param_names)
-
-    # metric
-    eval_metrics = mx.metric.CompositeEvalMetric()
-    mid=0
-    for m in range(len(config.RPN_FEAT_STRIDE)):
-      stride = config.RPN_FEAT_STRIDE[m]
-      #mid = m*MSTEP
-      _metric = metric.RPNAccMetric(pred_idx=mid, label_idx=mid+1, name='RPNAcc_s%s'%stride)
-      eval_metrics.add(_metric)
-      mid+=2
-      #_metric = metric.RPNLogLossMetric(pred_idx=mid, label_idx=mid+1)
-      #eval_metrics.add(_metric)
-
-      _metric = metric.RPNL1LossMetric(loss_idx=mid, weight_idx=mid+1, name='RPNL1Loss_s%s'%stride)
-      eval_metrics.add(_metric)
-      mid+=2
-      if config.FACE_LANDMARK:
-        _metric = metric.RPNL1LossMetric(loss_idx=mid, weight_idx=mid+1, name='RPNLandMarkL1Loss_s%s'%stride)
-        eval_metrics.add(_metric)
-        mid+=2
-      if config.HEAD_BOX:
-        _metric = metric.RPNAccMetric(pred_idx=mid, label_idx=mid+1, name='RPNAcc_head_s%s'%stride)
-        eval_metrics.add(_metric)
-        mid+=2
-        #_metric = metric.RPNLogLossMetric(pred_idx=mid, label_idx=mid+1)
-        #eval_metrics.add(_metric)
-
-        _metric = metric.RPNL1LossMetric(loss_idx=mid, weight_idx=mid+1, name='RPNL1Loss_head_s%s'%stride)
-        eval_metrics.add(_metric)
-        mid+=2
-      if config.CASCADE>0:
-        for _idx in range(config.CASCADE):
-          if stride in config.CASCADE_CLS_STRIDES:
-            _metric = metric.RPNAccMetric(pred_idx=mid, label_idx=mid+1, name='RPNAccCAS%d_s%s'%(_idx,stride))
-            eval_metrics.add(_metric)
-            mid+=2
-          if stride in config.CASCADE_BBOX_STRIDES:
-            _metric = metric.RPNL1LossMetric(loss_idx=mid, weight_idx=mid+1, name='RPNL1LossCAS%d_s%s'%(_idx,stride))
-            eval_metrics.add(_metric)
-            mid+=2
-
-    # callback
-    #means = np.tile(np.array(config.TRAIN.BBOX_MEANS), config.NUM_CLASSES)
-    #stds = np.tile(np.array(config.TRAIN.BBOX_STDS), config.NUM_CLASSES)
-    #epoch_end_callback = callback.do_checkpoint(prefix, means, stds)
-    epoch_end_callback = None
-    # decide learning rate
-    #base_lr = lr
-    #lr_factor = 0.1
-    #lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
-
-    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
-    lr_epoch_diff = [epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch]
-    lr_iters = [int(epoch * len(roidb) / input_batch_size) for epoch in lr_epoch_diff]
-    iter_per_epoch = int(len(roidb)/input_batch_size)
-
-    lr_steps = []
-    if len(lr_iters)==5:
-      factors = [0.5, 0.5, 0.4, 0.1, 0.1]
-      for i in range(5):
-        lr_steps.append( (lr_iters[i], factors[i]) )
-    elif len(lr_iters)==8: #warmup
-      for li in lr_iters[0:5]:
-        lr_steps.append( (li, 1.5849) )
-      for li in lr_iters[5:]:
-        lr_steps.append( (li, 0.1) )
-    else:
-      for li in lr_iters:
-        lr_steps.append( (li, 0.1) )
-    #lr_steps = [ (10,0.1), (20, 0.1) ] #XXX
-
-    end_epoch = 10000
-    logger.info('lr %f lr_epoch_diff %s lr_steps %s' % (lr, lr_epoch_diff, lr_steps))
-    # optimizer
-    opt = optimizer.SGD(learning_rate=lr, momentum=0.9, wd=args.wd, rescale_grad=1.0/len(ctx), clip_gradient=None)
-    initializer=mx.init.Xavier()
-    #initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-
-    train_data = mx.io.PrefetchingIter(train_data)
-
-    _cb = mx.callback.Speedometer(train_data.batch_size, frequent=args.frequent, auto_reset=False)
-    global_step = [0]
-
-    def save_model(epoch):
-      arg, aux = mod.get_params()
-      all_layers = mod.symbol.get_internals()
-      outs = []
-      for stride in config.RPN_FEAT_STRIDE:
-        num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
-        if config.CASCADE>0:
-          _name = 'face_rpn_cls_score_stride%d_output' % (stride)
-          cls_pred = all_layers[_name]
-          cls_pred = mx.symbol.Reshape(data=cls_pred, shape=(0, 2, -1, 0))
-
-          cls_pred = mx.symbol.SoftmaxActivation(data=cls_pred, mode="channel")
-          cls_pred = mx.symbol.Reshape(data=cls_pred, shape=(0, 2 * num_anchors, -1, 0))
-          outs.append(cls_pred)
-          _name = 'face_rpn_bbox_pred_stride%d_output' % stride
-          rpn_bbox_pred = all_layers[_name]
-          outs.append(rpn_bbox_pred)
-          if config.FACE_LANDMARK:
-            _name = 'face_rpn_landmark_pred_stride%d_output' % stride
-            rpn_landmark_pred = all_layers[_name]
-            outs.append(rpn_landmark_pred)
-          for casid in range(config.CASCADE):
-            if stride in config.CASCADE_CLS_STRIDES:
-              _name = 'face_rpn_cls_score_stride%d_cas%d_output' % (stride, casid)
-              cls_pred = all_layers[_name]
-              cls_pred = mx.symbol.Reshape(data=cls_pred, shape=(0, 2, -1, 0))
-              cls_pred = mx.symbol.SoftmaxActivation(data=cls_pred, mode="channel")
-              cls_pred = mx.symbol.Reshape(data=cls_pred, shape=(0, 2 * num_anchors, -1, 0))
-              outs.append(cls_pred)
-            if stride in config.CASCADE_BBOX_STRIDES:
-              _name = 'face_rpn_bbox_pred_stride%d_cas%d_output' % (stride, casid)
-              bbox_pred = all_layers[_name]
-              outs.append(bbox_pred)
-        else:
-          _name = 'face_rpn_cls_score_stride%d_output' % stride
-          rpn_cls_score = all_layers[_name]
-
-
-          # prepare rpn data
-          rpn_cls_score_reshape = mx.symbol.Reshape(data=rpn_cls_score,
-                                                    shape=(0, 2, -1, 0),
-                                                    name="face_rpn_cls_score_reshape_stride%d" % stride)
-
-          rpn_cls_prob = mx.symbol.SoftmaxActivation(data=rpn_cls_score_reshape,
-                                                     mode="channel",
-                                                     name="face_rpn_cls_prob_stride%d" % stride)
-          rpn_cls_prob_reshape = mx.symbol.Reshape(data=rpn_cls_prob,
-                                                   shape=(0, 2 * num_anchors, -1, 0),
-                                                   name='face_rpn_cls_prob_reshape_stride%d' % stride)
-          _name = 'face_rpn_bbox_pred_stride%d_output' % stride
-          rpn_bbox_pred = all_layers[_name]
-          outs.append(rpn_cls_prob_reshape)
-          outs.append(rpn_bbox_pred)
-          if config.FACE_LANDMARK:
-            _name = 'face_rpn_landmark_pred_stride%d_output' % stride
-            rpn_landmark_pred = all_layers[_name]
-            outs.append(rpn_landmark_pred)
-      _sym = mx.sym.Group(outs)
-      mx.model.save_checkpoint(prefix, epoch, _sym, arg, aux)
-
-    def _batch_callback(param):
-      #global global_step
-      _cb(param)
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for step in lr_steps:
-        if mbatch==step[0]:
-          opt.lr *= step[1]
-          print('lr change to', opt.lr,' in batch', mbatch, file=sys.stderr)
-          break
-
-      if mbatch%iter_per_epoch==0:
-        print('saving checkpoint', mbatch, file=sys.stderr)
-        save_model(0)
-      if mbatch==lr_steps[-1][0]:
-        print('saving final checkpoint', mbatch, file=sys.stderr)
-        save_model(0)
-        #arg, aux = mod.get_params()
-        #mx.model.save_checkpoint(prefix, 99, mod.symbol, arg, aux)
-        sys.exit(0)
-
-    # train
-    mod.fit(train_data, eval_metric=eval_metrics, epoch_end_callback=epoch_end_callback,
-            batch_end_callback=_batch_callback, kvstore=args.kvstore,
-            optimizer=opt,
-            initializer = initializer,
-            allow_missing=True,
-            arg_params=arg_params, aux_params=aux_params, begin_epoch=begin_epoch, num_epoch=end_epoch)
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Train RetinaFace')
-    # general
-    parser.add_argument('--network', help='network name', default=default.network, type=str)
-    parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-    args, rest = parser.parse_known_args()
-    generate_config(args.network, args.dataset)
-    parser.add_argument('--image_set', help='image_set name', default=default.image_set, type=str)
-    parser.add_argument('--root_path', help='output data folder', default=default.root_path, type=str)
-    parser.add_argument('--dataset_path', help='dataset path', default=default.dataset_path, type=str)
-    # training
-    parser.add_argument('--frequent', help='frequency of logging', default=default.frequent, type=int)
-    parser.add_argument('--kvstore', help='the kv-store type', default=default.kvstore, type=str)
-    parser.add_argument('--work_load_list', help='work load for different devices', default=None, type=list)
-    parser.add_argument('--no_flip', help='disable flip images', action='store_true')
-    parser.add_argument('--no_shuffle', help='disable random shuffle', action='store_true')
-    # e2e
-    #parser.add_argument('--gpus', help='GPU device to train with', default='0,1,2,3', type=str)
-    parser.add_argument('--pretrained', help='pretrained model prefix', default=default.pretrained, type=str)
-    parser.add_argument('--pretrained_epoch', help='pretrained model epoch', default=default.pretrained_epoch, type=int)
-    parser.add_argument('--prefix', help='new model prefix', default=default.prefix, type=str)
-    parser.add_argument('--begin_epoch', help='begin epoch of training, use with resume', default=0, type=int)
-    parser.add_argument('--end_epoch', help='end epoch of training', default=default.end_epoch, type=int)
-    parser.add_argument('--lr', help='base learning rate', default=default.lr, type=float)
-    parser.add_argument('--lr_step', help='learning rate steps (in epoch)', default=default.lr_step, type=str)
-    parser.add_argument('--wd', help='weight decay', default=default.wd, type=float)
-    args = parser.parse_args()
-    return args
-
-
-def main():
-    args = parse_args()
-    logger.info('Called with argument: %s' % args)
-    #ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in range(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    train_net(args, ctx, args.pretrained, args.pretrained_epoch, args.prefix, args.begin_epoch, args.end_epoch,
-              lr=args.lr, lr_step=args.lr_step)
-
-if __name__ == '__main__':
-    main()
diff --git a/RetinaFaceAntiCov/README.md b/RetinaFaceAntiCov/README.md
deleted file mode 100644
index ae6c222ce..000000000
--- a/RetinaFaceAntiCov/README.md
+++ /dev/null
@@ -1,36 +0,0 @@
-# RetinaFace Anti Cov Face Detector
-
-## Introduction
-
-RetinaFace-Anti-Cov is a customized one stage face detector to help people protect themselves from CovID-19.
-
-![demoimg1](https://github.com/deepinsight/insightface/blob/master/resources/cov_test.jpg)
-
-
-## Testing
-
-Please check ``test.py`` for testing.
-
-Make sure that you set ``network='net3l'`` instead of ``'net3'`` for 'mnet_cov2' model, otherwise you will get incorrect landmarks.
-
-## Pretrained Models
-
-~~MobileNet0.25([baidu cloud](https://pan.baidu.com/s/1p8n4R2W-9WmmBWxYQEFcWg),code: fmfm)~~
-
-Better: MobileNet0.25 ([baidu cloud](https://pan.baidu.com/s/16ihzPxjTObdbv0D6P6LmEQ), code: j3b6, [dropbox](https://www.dropbox.com/s/6rhhxsbh2qik65k/cov2.zip?dl=0))
-
-
-
-## References
-
-```
-  
-@inproceedings{deng2019retinaface,
-title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
-author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
-booktitle={arxiv},
-year={2019}
-}
-```
-
-
diff --git a/RetinaFaceAntiCov/rcnn/processing/assign_levels.py b/RetinaFaceAntiCov/rcnn/processing/assign_levels.py
deleted file mode 100755
index b237439fc..000000000
--- a/RetinaFaceAntiCov/rcnn/processing/assign_levels.py
+++ /dev/null
@@ -1,37 +0,0 @@
-from rcnn.config import config
-import numpy as np
-
-
-def compute_assign_targets(rois, threshold):
-    rois_area = np.sqrt((rois[:, 2] - rois[:, 0] + 1) * (rois[:, 3] - rois[:, 1] + 1))
-    num_rois = np.shape(rois)[0]
-    assign_levels = np.zeros(num_rois, dtype=np.uint8)
-    for i, stride in enumerate(config.RCNN_FEAT_STRIDE):
-        thd = threshold[i]
-        idx = np.logical_and(thd[1] <= rois_area, rois_area < thd[0])
-        assign_levels[idx] = stride
-
-    assert 0 not in assign_levels, "All rois should assign to specify levels."
-    return assign_levels
-
-
-def add_assign_targets(roidb):
-    """
-    given roidb, add ['assign_level']
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    """
-    print 'add assign targets'
-    assert len(roidb) > 0
-    assert 'boxes' in roidb[0]
-
-    area_threshold = [[np.inf, 448],
-                      [448,    224],
-                      [224,    112],
-                      [112,     0]]
-
-    assert len(config.RCNN_FEAT_STRIDE) == len(area_threshold)
-
-    num_images = len(roidb)
-    for im_i in range(num_images):
-        rois = roidb[im_i]['boxes']
-        roidb[im_i]['assign_levels'] = compute_assign_targets(rois, area_threshold)
diff --git a/RetinaFaceAntiCov/rcnn/processing/bbox_regression.py b/RetinaFaceAntiCov/rcnn/processing/bbox_regression.py
deleted file mode 100644
index 11d1a0236..000000000
--- a/RetinaFaceAntiCov/rcnn/processing/bbox_regression.py
+++ /dev/null
@@ -1,255 +0,0 @@
-"""
-This file has functions about generating bounding box regression targets
-"""
-
-from ..pycocotools.mask import encode
-import numpy as np
-
-from ..logger import logger
-from .bbox_transform import bbox_overlaps, bbox_transform
-from rcnn.config import config
-import math
-import cv2
-import PIL.Image as Image
-import threading
-import Queue
-
-
-def compute_bbox_regression_targets(rois, overlaps, labels):
-    """
-    given rois, overlaps, gt labels, compute bounding box regression targets
-    :param rois: roidb[i]['boxes'] k * 4
-    :param overlaps: roidb[i]['max_overlaps'] k * 1
-    :param labels: roidb[i]['max_classes'] k * 1
-    :return: targets[i][class, dx, dy, dw, dh] k * 5
-    """
-    # Ensure ROIs are floats
-    rois = rois.astype(np.float, copy=False)
-
-    # Sanity check
-    if len(rois) != len(overlaps):
-        logger.warning('bbox regression: len(rois) != len(overlaps)')
-
-    # Indices of ground-truth ROIs
-    gt_inds = np.where(overlaps == 1)[0]
-    if len(gt_inds) == 0:
-        logger.warning('bbox regression: len(gt_inds) == 0')
-
-    # Indices of examples for which we try to make predictions
-    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
-
-    # Get IoU overlap between each ex ROI and gt ROI
-    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
-
-    # Find which gt ROI each ex ROI has max overlap with:
-    # this will be the ex ROI's gt target
-    gt_assignment = ex_gt_overlaps.argmax(axis=1)
-    gt_rois = rois[gt_inds[gt_assignment], :]
-    ex_rois = rois[ex_inds, :]
-
-    targets = np.zeros((rois.shape[0], 5), dtype=np.float32)
-    targets[ex_inds, 0] = labels[ex_inds]
-    targets[ex_inds, 1:] = bbox_transform(ex_rois, gt_rois)
-    return targets
-
-
-def add_bbox_regression_targets(roidb):
-    """
-    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    :return: means, std variances of targets
-    """
-    logger.info('bbox regression: add bounding box regression targets')
-    assert len(roidb) > 0
-    assert 'max_classes' in roidb[0]
-
-    num_images = len(roidb)
-    num_classes = roidb[0]['gt_overlaps'].shape[1]
-    for im_i in range(num_images):
-        rois = roidb[im_i]['boxes']
-        max_overlaps = roidb[im_i]['max_overlaps']
-        max_classes = roidb[im_i]['max_classes']
-        roidb[im_i]['bbox_targets'] = compute_bbox_regression_targets(rois, max_overlaps, max_classes)
-
-    if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
-        # use fixed / precomputed means and stds instead of empirical values
-        means = np.tile(np.array(config.TRAIN.BBOX_MEANS), (num_classes, 1))
-        stds = np.tile(np.array(config.TRAIN.BBOX_STDS), (num_classes, 1))
-    else:
-        # compute mean, std values
-        class_counts = np.zeros((num_classes, 1)) + 1e-14
-        sums = np.zeros((num_classes, 4))
-        squared_sums = np.zeros((num_classes, 4))
-        for im_i in range(num_images):
-            targets = roidb[im_i]['bbox_targets']
-            for cls in range(1, num_classes):
-                cls_indexes = np.where(targets[:, 0] == cls)[0]
-                if cls_indexes.size > 0:
-                    class_counts[cls] += cls_indexes.size
-                    sums[cls, :] += targets[cls_indexes, 1:].sum(axis=0)
-                    squared_sums[cls, :] += (targets[cls_indexes, 1:] ** 2).sum(axis=0)
-
-        means = sums / class_counts
-        # var(x) = E(x^2) - E(x)^2
-        stds = np.sqrt(squared_sums / class_counts - means ** 2)
-
-    # normalized targets
-    for im_i in range(num_images):
-        targets = roidb[im_i]['bbox_targets']
-        for cls in range(1, num_classes):
-            cls_indexes = np.where(targets[:, 0] == cls)[0]
-            roidb[im_i]['bbox_targets'][cls_indexes, 1:] -= means[cls, :]
-            roidb[im_i]['bbox_targets'][cls_indexes, 1:] /= stds[cls, :]
-
-    return means.ravel(), stds.ravel()
-
-
-def expand_bbox_regression_targets(bbox_targets_data, num_classes):
-    """
-    expand from 5 to 4 * num_classes; only the right class has non-zero bbox regression targets
-    :param bbox_targets_data: [k * 5]
-    :param num_classes: number of classes
-    :return: bbox target processed [k * 4 num_classes]
-    bbox_weights ! only foreground boxes have bbox regression computation!
-    """
-    classes = bbox_targets_data[:, 0]
-    bbox_targets = np.zeros((classes.size, 4 * num_classes), dtype=np.float32)
-    bbox_weights = np.zeros(bbox_targets.shape, dtype=np.float32)
-    indexes = np.where(classes > 0)[0]
-    for index in indexes:
-        cls = classes[index]
-        start = int(4 * cls)
-        end = start + 4
-        bbox_targets[index, start:end] = bbox_targets_data[index, 1:]
-        bbox_weights[index, start:end] = config.TRAIN.BBOX_WEIGHTS
-    return bbox_targets, bbox_weights
-
-
-def compute_mask_and_label(ex_rois, ex_labels, seg, flipped):
-    # assert os.path.exists(seg_gt), 'Path does not exist: {}'.format(seg_gt)
-    # im = Image.open(seg_gt)
-    # pixel = list(im.getdata())
-    # pixel = np.array(pixel).reshape([im.size[1], im.size[0]])
-    im = Image.open(seg)
-    pixel = list(im.getdata())
-    ins_seg = np.array(pixel).reshape([im.size[1], im.size[0]])
-    if flipped:
-        ins_seg = ins_seg[:, ::-1]
-    rois = ex_rois
-    n_rois = ex_rois.shape[0]
-    label = ex_labels
-    class_id = config.CLASS_ID
-    mask_target = np.zeros((n_rois, 28, 28), dtype=np.int8)
-    mask_label = np.zeros((n_rois), dtype=np.int8)
-    for n in range(n_rois):
-        target = ins_seg[int(rois[n, 1]): int(rois[n, 3]), int(rois[n, 0]): int(rois[n, 2])]
-        ids = np.unique(target)
-        ins_id = 0
-        max_count = 0
-        for id in ids:
-            if math.floor(id / 1000) == class_id[int(label[int(n)])]:
-                px = np.where(ins_seg == int(id))
-                x_min = np.min(px[1])
-                y_min = np.min(px[0])
-                x_max = np.max(px[1])
-                y_max = np.max(px[0])
-                x1 = max(rois[n, 0], x_min)
-                y1 = max(rois[n, 1], y_min)
-                x2 = min(rois[n, 2], x_max)
-                y2 = min(rois[n, 3], y_max)
-                iou = (x2 - x1) * (y2 - y1)
-                iou = iou / ((rois[n, 2] - rois[n, 0]) * (rois[n, 3] - rois[n, 1])
-                             + (x_max - x_min) * (y_max - y_min) - iou)
-                if iou > max_count:
-                    ins_id = id
-                    max_count = iou
-
-        if max_count == 0:
-            continue
-        # print max_count
-        mask = np.zeros(target.shape)
-        idx = np.where(target == ins_id)
-        mask[idx] = 1
-        mask = cv2.resize(mask, (28, 28), interpolation=cv2.INTER_NEAREST)
-
-        mask_target[n] = mask
-        mask_label[n] = label[int(n)]
-    return mask_target, mask_label
-
-
-def compute_bbox_mask_targets_and_label(rois, overlaps, labels, seg, flipped):
-    """
-    given rois, overlaps, gt labels, seg, compute bounding box mask targets
-    :param rois: roidb[i]['boxes'] k * 4
-    :param overlaps: roidb[i]['max_overlaps'] k * 1
-    :param labels: roidb[i]['max_classes'] k * 1
-    :return: targets[i][class, dx, dy, dw, dh] k * 5
-    """
-    # Ensure ROIs are floats
-    rois = rois.astype(np.float, copy=False)
-
-    # Sanity check
-    if len(rois) != len(overlaps):
-        print 'bbox regression: this should not happen'
-
-    # Indices of ground-truth ROIs
-    gt_inds = np.where(overlaps == 1)[0]
-    if len(gt_inds) == 0:
-        print 'something wrong : zero ground truth rois'
-    # Indices of examples for which we try to make predictions
-    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
-
-    # Get IoU overlap between each ex ROI and gt ROI
-    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
-
-
-    # Find which gt ROI each ex ROI has max overlap with:
-    # this will be the ex ROI's gt target
-    gt_assignment = ex_gt_overlaps.argmax(axis=1)
-    gt_rois = rois[gt_inds[gt_assignment], :]
-    ex_rois = rois[ex_inds, :]
-
-    mask_targets, mask_label = compute_mask_and_label(ex_rois, labels[ex_inds], seg, flipped)
-    return mask_targets, mask_label, ex_inds
-
-def add_mask_targets(roidb):
-    """
-    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
-    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
-    :return: means, std variances of targets
-    """
-    print 'add bounding box mask targets'
-    assert len(roidb) > 0
-    assert 'max_classes' in roidb[0]
-
-    num_images = len(roidb)
-
-    # Multi threads processing
-    im_quene = Queue.Queue(maxsize=0)
-    for im_i in range(num_images):
-        im_quene.put(im_i)
-
-    def process():
-        while not im_quene.empty():
-            im_i = im_quene.get()
-            print "-----process img {}".format(im_i)
-            rois = roidb[im_i]['boxes']
-            max_overlaps = roidb[im_i]['max_overlaps']
-            max_classes = roidb[im_i]['max_classes']
-            ins_seg = roidb[im_i]['ins_seg']
-            flipped = roidb[im_i]['flipped']
-            roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
-                compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg, flipped)
-    threads = [threading.Thread(target=process, args=()) for i in range(10)]
-    for t in threads: t.start()
-    for t in threads: t.join()
-    # Single thread
-    # for im_i in range(num_images):
-    #     print "-----processing img {}".format(im_i)
-    #     rois = roidb[im_i]['boxes']
-    #     max_overlaps = roidb[im_i]['max_overlaps']
-    #     max_classes = roidb[im_i]['max_classes']
-    #     ins_seg = roidb[im_i]['ins_seg']
-    #     # roidb[im_i]['mask_targets'] = compute_bbox_mask_targets(rois, max_overlaps, max_classes, ins_seg)
-    #     roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
-    #         compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg)
diff --git a/RetinaFaceAntiCov/rcnn/processing/bbox_transform.py b/RetinaFaceAntiCov/rcnn/processing/bbox_transform.py
deleted file mode 100644
index 0c4fc7c8b..000000000
--- a/RetinaFaceAntiCov/rcnn/processing/bbox_transform.py
+++ /dev/null
@@ -1,218 +0,0 @@
-import numpy as np
-from ..cython.bbox import bbox_overlaps_cython
-#from rcnn.config import config
-
-
-def bbox_overlaps(boxes, query_boxes):
-    return bbox_overlaps_cython(boxes, query_boxes)
-
-
-def bbox_overlaps_py(boxes, query_boxes):
-    """
-    determine overlaps between boxes and query_boxes
-    :param boxes: n * 4 bounding boxes
-    :param query_boxes: k * 4 bounding boxes
-    :return: overlaps: n * k overlaps
-    """
-    n_ = boxes.shape[0]
-    k_ = query_boxes.shape[0]
-    overlaps = np.zeros((n_, k_), dtype=np.float)
-    for k in range(k_):
-        query_box_area = (query_boxes[k, 2] - query_boxes[k, 0] + 1) * (query_boxes[k, 3] - query_boxes[k, 1] + 1)
-        for n in range(n_):
-            iw = min(boxes[n, 2], query_boxes[k, 2]) - max(boxes[n, 0], query_boxes[k, 0]) + 1
-            if iw > 0:
-                ih = min(boxes[n, 3], query_boxes[k, 3]) - max(boxes[n, 1], query_boxes[k, 1]) + 1
-                if ih > 0:
-                    box_area = (boxes[n, 2] - boxes[n, 0] + 1) * (boxes[n, 3] - boxes[n, 1] + 1)
-                    all_area = float(box_area + query_box_area - iw * ih)
-                    overlaps[n, k] = iw * ih / all_area
-    return overlaps
-
-
-def clip_boxes(boxes, im_shape):
-    """
-    Clip boxes to image boundaries.
-    :param boxes: [N, 4* num_classes]
-    :param im_shape: tuple of 2
-    :return: [N, 4* num_classes]
-    """
-    # x1 >= 0
-    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
-    # y1 >= 0
-    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
-    # x2 < im_shape[1]
-    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
-    # y2 < im_shape[0]
-    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
-    return boxes
-
-
-def nonlinear_transform(ex_rois, gt_rois):
-    """
-    compute bounding box regression targets from ex_rois to gt_rois
-    :param ex_rois: [N, 4]
-    :param gt_rois: [N, 4]
-    :return: [N, 4]
-    """
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-
-    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
-    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
-    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
-    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
-
-    gt_widths = gt_rois[:, 2] - gt_rois[:, 0] + 1.0
-    gt_heights = gt_rois[:, 3] - gt_rois[:, 1] + 1.0
-    gt_ctr_x = gt_rois[:, 0] + 0.5 * (gt_widths - 1.0)
-    gt_ctr_y = gt_rois[:, 1] + 0.5 * (gt_heights - 1.0)
-
-    targets_dx = (gt_ctr_x - ex_ctr_x) / (ex_widths + 1e-14)
-    targets_dy = (gt_ctr_y - ex_ctr_y) / (ex_heights + 1e-14)
-    targets_dw = np.log(gt_widths / ex_widths)
-    targets_dh = np.log(gt_heights / ex_heights)
-
-    if gt_rois.shape[1]<=4:
-      targets = np.vstack(
-          (targets_dx, targets_dy, targets_dw, targets_dh)).transpose()
-      return targets
-    else:
-      targets = [targets_dx, targets_dy, targets_dw, targets_dh]
-      #if config.USE_BLUR:
-      #  for i in range(4, gt_rois.shape[1]):
-      #    t = gt_rois[:,i]
-      #    targets.append(t)
-      targets = np.vstack(targets).transpose()
-      return targets
-
-def landmark_transform(ex_rois, gt_rois):
-
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-
-    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
-    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
-    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
-    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
-
-    
-    targets = []
-    for i in range(gt_rois.shape[1]):
-      for j in range(gt_rois.shape[2]):
-        #if not config.USE_OCCLUSION and j==2:
-        #  continue
-        if j==2:
-          continue
-        if j==0: #w
-          target = (gt_rois[:,i,j] - ex_ctr_x) / (ex_widths + 1e-14)
-        elif j==1: #h
-          target = (gt_rois[:,i,j] - ex_ctr_y) / (ex_heights + 1e-14)
-        else: #visibile
-          target = gt_rois[:,i,j]
-        targets.append(target)
-
-
-    targets = np.vstack(targets).transpose()
-    return targets
-
-
-def nonlinear_pred(boxes, box_deltas):
-    """
-    Transform the set of class-agnostic boxes into class-specific boxes
-    by applying the predicted offsets (box_deltas)
-    :param boxes: !important [N 4]
-    :param box_deltas: [N, 4 * num_classes]
-    :return: [N 4 * num_classes]
-    """
-    if boxes.shape[0] == 0:
-        return np.zeros((0, box_deltas.shape[1]))
-
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-    dx = box_deltas[:, 0::4]
-    dy = box_deltas[:, 1::4]
-    dw = box_deltas[:, 2::4]
-    dh = box_deltas[:, 3::4]
-
-    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-    pred_w = np.exp(dw) * widths[:, np.newaxis]
-    pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-    pred_boxes = np.zeros(box_deltas.shape)
-    # x1
-    pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-    # y1
-    pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-    # x2
-    pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-    # y2
-    pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-
-    return pred_boxes
-
-def landmark_pred(boxes, landmark_deltas):
-    if boxes.shape[0] == 0:
-        return np.zeros((0, landmark_deltas.shape[1]))
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-    preds = []
-    for i in range(landmark_deltas.shape[1]):
-      if i%2==0:
-        pred = (landmark_deltas[:,i]*widths + ctr_x)
-      else:
-        pred = (landmark_deltas[:,i]*heights + ctr_y)
-      preds.append(pred)
-    preds = np.vstack(preds).transpose()
-    return preds
-
-def iou_transform(ex_rois, gt_rois):
-    """ return bbox targets, IoU loss uses gt_rois as gt """
-    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
-    return gt_rois
-
-
-def iou_pred(boxes, box_deltas):
-    """
-    Transform the set of class-agnostic boxes into class-specific boxes
-    by applying the predicted offsets (box_deltas)
-    :param boxes: !important [N 4]
-    :param box_deltas: [N, 4 * num_classes]
-    :return: [N 4 * num_classes]
-    """
-    if boxes.shape[0] == 0:
-        return np.zeros((0, box_deltas.shape[1]))
-
-    boxes = boxes.astype(np.float, copy=False)
-    x1 = boxes[:, 0]
-    y1 = boxes[:, 1]
-    x2 = boxes[:, 2]
-    y2 = boxes[:, 3]
-
-    dx1 = box_deltas[:, 0::4]
-    dy1 = box_deltas[:, 1::4]
-    dx2 = box_deltas[:, 2::4]
-    dy2 = box_deltas[:, 3::4]
-
-    pred_boxes = np.zeros(box_deltas.shape)
-    # x1
-    pred_boxes[:, 0::4] = dx1 + x1[:, np.newaxis]
-    # y1
-    pred_boxes[:, 1::4] = dy1 + y1[:, np.newaxis]
-    # x2
-    pred_boxes[:, 2::4] = dx2 + x2[:, np.newaxis]
-    # y2
-    pred_boxes[:, 3::4] = dy2 + y2[:, np.newaxis]
-
-    return pred_boxes
-
-
-# define bbox_transform and bbox_pred
-bbox_transform = nonlinear_transform
-bbox_pred = nonlinear_pred
diff --git a/RetinaFaceAntiCov/rcnn/processing/generate_anchor.py b/RetinaFaceAntiCov/rcnn/processing/generate_anchor.py
deleted file mode 100644
index eb47d8ea5..000000000
--- a/RetinaFaceAntiCov/rcnn/processing/generate_anchor.py
+++ /dev/null
@@ -1,128 +0,0 @@
-"""
-Generate base anchors on index 0
-"""
-from __future__ import print_function
-import sys
-from builtins import range
-import numpy as np
-from ..cython.anchors import anchors_cython
-#from ..config import config
-
-
-def anchors_plane(feat_h, feat_w, stride, base_anchor):
-    return anchors_cython(feat_h, feat_w, stride, base_anchor)
-
-def generate_anchors(base_size=16, ratios=[0.5, 1, 2],
-                     scales=2 ** np.arange(3, 6), stride=16, dense_anchor=False):
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-
-    base_anchor = np.array([1, 1, base_size, base_size]) - 1
-    ratio_anchors = _ratio_enum(base_anchor, ratios)
-    anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales)
-                         for i in range(ratio_anchors.shape[0])])
-    if dense_anchor:
-      assert stride%2==0
-      anchors2 = anchors.copy()
-      anchors2[:,:] += int(stride/2)
-      anchors = np.vstack( (anchors, anchors2) )
-    #print('GA',base_anchor.shape, ratio_anchors.shape, anchors.shape)
-    return anchors
-
-#def generate_anchors_fpn(base_size=[64,32,16,8,4], ratios=[0.5, 1, 2], scales=8):
-#    """
-#    Generate anchor (reference) windows by enumerating aspect ratios X
-#    scales wrt a reference (0, 0, 15, 15) window.
-#    """
-#    anchors = []
-#    _ratios = ratios.reshape( (len(base_size), -1) )
-#    _scales = scales.reshape( (len(base_size), -1) )
-#    for i,bs in enumerate(base_size):
-#      __ratios = _ratios[i]
-#      __scales = _scales[i]
-#      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
-#      r = generate_anchors(bs, __ratios, __scales)
-#      #print('anchors_fpn', r.shape, file=sys.stderr)
-#      anchors.append(r)
-#    return anchors
-
-def generate_anchors_fpn(dense_anchor=False, cfg = None):
-    #assert(False)
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-    if cfg is None:
-      from ..config import config
-      cfg = config.RPN_ANCHOR_CFG
-    RPN_FEAT_STRIDE = []
-    for k in cfg:
-      RPN_FEAT_STRIDE.append( int(k) )
-    RPN_FEAT_STRIDE = sorted(RPN_FEAT_STRIDE, reverse=True)
-    anchors = []
-    for k in RPN_FEAT_STRIDE:
-      v = cfg[str(k)]
-      bs = v['BASE_SIZE']
-      __ratios = np.array(v['RATIOS'])
-      __scales = np.array(v['SCALES'])
-      stride = int(k)
-      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
-      r = generate_anchors(bs, __ratios, __scales, stride, dense_anchor)
-      #print('anchors_fpn', r.shape, file=sys.stderr)
-      anchors.append(r)
-
-    return anchors
-
-def _whctrs(anchor):
-    """
-    Return width, height, x center, and y center for an anchor (window).
-    """
-
-    w = anchor[2] - anchor[0] + 1
-    h = anchor[3] - anchor[1] + 1
-    x_ctr = anchor[0] + 0.5 * (w - 1)
-    y_ctr = anchor[1] + 0.5 * (h - 1)
-    return w, h, x_ctr, y_ctr
-
-
-def _mkanchors(ws, hs, x_ctr, y_ctr):
-    """
-    Given a vector of widths (ws) and heights (hs) around a center
-    (x_ctr, y_ctr), output a set of anchors (windows).
-    """
-
-    ws = ws[:, np.newaxis]
-    hs = hs[:, np.newaxis]
-    anchors = np.hstack((x_ctr - 0.5 * (ws - 1),
-                         y_ctr - 0.5 * (hs - 1),
-                         x_ctr + 0.5 * (ws - 1),
-                         y_ctr + 0.5 * (hs - 1)))
-    return anchors
-
-
-def _ratio_enum(anchor, ratios):
-    """
-    Enumerate a set of anchors for each aspect ratio wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    size = w * h
-    size_ratios = size / ratios
-    ws = np.round(np.sqrt(size_ratios))
-    hs = np.round(ws * ratios)
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
-
-
-def _scale_enum(anchor, scales):
-    """
-    Enumerate a set of anchors for each scale wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    ws = w * scales
-    hs = h * scales
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
diff --git a/RetinaFaceAntiCov/rcnn/processing/nms.py b/RetinaFaceAntiCov/rcnn/processing/nms.py
deleted file mode 100644
index 230139c41..000000000
--- a/RetinaFaceAntiCov/rcnn/processing/nms.py
+++ /dev/null
@@ -1,64 +0,0 @@
-import numpy as np
-from ..cython.cpu_nms import cpu_nms
-try:
-    from ..cython.gpu_nms import gpu_nms
-except ImportError:
-    gpu_nms = None
-
-
-def py_nms_wrapper(thresh):
-    def _nms(dets):
-        return nms(dets, thresh)
-    return _nms
-
-
-def cpu_nms_wrapper(thresh):
-    def _nms(dets):
-        return cpu_nms(dets, thresh)
-    return _nms
-
-
-def gpu_nms_wrapper(thresh, device_id):
-    def _nms(dets):
-        return gpu_nms(dets, thresh, device_id)
-    if gpu_nms is not None:
-        return _nms
-    else:
-        return cpu_nms_wrapper(thresh)
-
-
-def nms(dets, thresh):
-    """
-    greedily select boxes with high confidence and overlap with current maximum <= thresh
-    rule out overlap >= thresh
-    :param dets: [[x1, y1, x2, y2 score]]
-    :param thresh: retain overlap < thresh
-    :return: indexes to keep
-    """
-    x1 = dets[:, 0]
-    y1 = dets[:, 1]
-    x2 = dets[:, 2]
-    y2 = dets[:, 3]
-    scores = dets[:, 4]
-
-    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
-    order = scores.argsort()[::-1]
-
-    keep = []
-    while order.size > 0:
-        i = order[0]
-        keep.append(i)
-        xx1 = np.maximum(x1[i], x1[order[1:]])
-        yy1 = np.maximum(y1[i], y1[order[1:]])
-        xx2 = np.minimum(x2[i], x2[order[1:]])
-        yy2 = np.minimum(y2[i], y2[order[1:]])
-
-        w = np.maximum(0.0, xx2 - xx1 + 1)
-        h = np.maximum(0.0, yy2 - yy1 + 1)
-        inter = w * h
-        ovr = inter / (areas[i] + areas[order[1:]] - inter)
-
-        inds = np.where(ovr <= thresh)[0]
-        order = order[inds + 1]
-
-    return keep
diff --git a/RetinaFaceAntiCov/retinaface_cov.py b/RetinaFaceAntiCov/retinaface_cov.py
deleted file mode 100644
index 57d1ef23e..000000000
--- a/RetinaFaceAntiCov/retinaface_cov.py
+++ /dev/null
@@ -1,637 +0,0 @@
-from __future__ import print_function
-import sys
-import os
-import datetime
-import time
-import numpy as np
-import mxnet as mx
-from mxnet import ndarray as nd
-import cv2
-#from rcnn import config
-#from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
-from rcnn.processing.bbox_transform import clip_boxes
-from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
-from rcnn.processing.nms import gpu_nms_wrapper, cpu_nms_wrapper
-from rcnn.processing.bbox_transform import bbox_overlaps
-
-class RetinaFaceCoV:
-  def __init__(self, prefix, epoch, ctx_id=0, network='net3', nms=0.4, nocrop=False):
-    self.ctx_id = ctx_id
-    self.network = network
-    self.nms_threshold = nms
-    self.nocrop = nocrop
-    self.debug = False
-    self.fpn_keys = []
-    self.anchor_cfg = None
-    pixel_means=[0.0, 0.0, 0.0]
-    pixel_stds=[1.0, 1.0, 1.0]
-    pixel_scale = 1.0
-    self.bbox_stds = [1.0, 1.0, 1.0, 1.0]
-    self.landmark_std = 1.0
-    self.preprocess = False
-    _ratio = (1.,)
-    fmc = 3
-    if network=='ssh' or network=='vgg':
-      pixel_means=[103.939, 116.779, 123.68]
-      self.preprocess = True
-    elif network=='net3':
-      _ratio = (1.,)
-    elif network=='net3l':
-      _ratio = (1.,)
-      self.landmark_std = 0.2
-    elif network=='net3a':
-      _ratio = (1.,1.5)
-    elif network=='net6': #like pyramidbox or s3fd
-      fmc = 6
-    elif network=='net5': #retinaface
-      fmc = 5
-    elif network=='net5a':
-      fmc = 5
-      _ratio = (1.,1.5)
-    elif network=='net4':
-      fmc = 4
-    elif network=='net4a':
-      fmc = 4
-      _ratio = (1.,1.5)
-    elif network=='x5':
-      fmc = 5
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    elif network=='x3':
-      fmc = 3
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    elif network=='x3a':
-      fmc = 3
-      _ratio = (1.,1.5)
-      pixel_means=[103.52, 116.28, 123.675]
-      pixel_stds=[57.375, 57.12, 58.395]
-    else:
-      assert False, 'network setting error %s'%network
-
-    if fmc==3:
-      self._feat_stride_fpn = [32, 16, 8]
-      self.anchor_cfg = {
-          '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==4:
-      self._feat_stride_fpn = [32, 16, 8, 4]
-      self.anchor_cfg = {
-          '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '4': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==6:
-      self._feat_stride_fpn = [128, 64, 32, 16, 8, 4]
-      self.anchor_cfg = {
-          '128': {'SCALES': (32,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '64': {'SCALES': (16,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '32': {'SCALES': (8,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '16': {'SCALES': (4,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '8': {'SCALES': (2,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-          '4': {'SCALES': (1,), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-      }
-    elif fmc==5:
-      self._feat_stride_fpn = [64, 32, 16, 8, 4]
-      self.anchor_cfg = {}
-      _ass = 2.0**(1.0/3)
-      _basescale = 1.0
-      for _stride in [4, 8, 16, 32, 64]:
-        key = str(_stride)
-        value = {'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999}
-        scales = []
-        for _ in range(3):
-          scales.append(_basescale)
-          _basescale *= _ass
-        value['SCALES'] = tuple(scales)
-        self.anchor_cfg[key] = value
-
-    #print(self._feat_stride_fpn, self.anchor_cfg)
-
-    for s in self._feat_stride_fpn:
-        self.fpn_keys.append('stride%s'%s)
-
-
-    dense_anchor = False
-    #self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(base_size=fpn_base_size, scales=self._scales, ratios=self._ratios)))
-    self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(dense_anchor=dense_anchor, cfg=self.anchor_cfg)))
-    for k in self._anchors_fpn:
-      v = self._anchors_fpn[k].astype(np.float32)
-      self._anchors_fpn[k] = v
-
-    self._num_anchors = dict(zip(self.fpn_keys, [anchors.shape[0] for anchors in self._anchors_fpn.values()]))
-    #self._bbox_pred = nonlinear_pred
-    #self._landmark_pred = landmark_pred
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    if self.ctx_id>=0:
-      self.ctx = mx.gpu(self.ctx_id)
-      self.nms = gpu_nms_wrapper(self.nms_threshold, self.ctx_id)
-    else:
-      self.ctx = mx.cpu()
-      self.nms = cpu_nms_wrapper(self.nms_threshold)
-    self.pixel_means = np.array(pixel_means, dtype=np.float32)
-    self.pixel_stds = np.array(pixel_stds, dtype=np.float32)
-    self.pixel_scale = float(pixel_scale)
-    #print('means', self.pixel_means)
-    self.use_landmarks = True
-    #print('use_landmarks', self.use_landmarks)
-    self.cascade = 0
-
-    if self.debug:
-      c = len(sym)//len(self._feat_stride_fpn)
-      sym = sym[(c*0):]
-      self._feat_stride_fpn = [32,16,8]
-    #print('sym size:', len(sym))
-
-    image_size = (640, 640)
-    self.model = mx.mod.Module(symbol=sym, context=self.ctx, label_names = None)
-    self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
-    self.model.set_params(arg_params, aux_params)
-
-  def get_input(self, img):
-    im = img.astype(np.float32)
-    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-    for i in range(3):
-        im_tensor[0, i, :, :] = (im[:, :, 2 - i]/self.pixel_scale - self.pixel_means[2 - i])/self.pixel_stds[2-i]
-    #if self.debug:
-    #  timeb = datetime.datetime.now()
-    #  diff = timeb - timea
-    #  print('X2 uses', diff.total_seconds(), 'seconds')
-    data = nd.array(im_tensor)
-    return data
-
-  def detect(self, img, threshold=0.5, scales=[1.0], do_flip=False):
-    #print('in_detect', threshold, scales, do_flip, do_nms)
-    proposals_list = []
-    scores_list = []
-    mask_scores_list = []
-    landmarks_list = []
-    strides_list = []
-    timea = datetime.datetime.now()
-    flips = [0]
-    if do_flip:
-      flips = [0, 1]
-
-    imgs = [img]
-    if isinstance(img, list):
-      imgs = img
-    for img in imgs:
-      for im_scale in scales:
-        for flip in flips:
-          if im_scale!=1.0:
-            im = cv2.resize(img, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
-          else:
-            im = img.copy()
-          if flip:
-            im = im[:,::-1,:]
-          if self.nocrop:
-            if im.shape[0]%32==0:
-              h = im.shape[0]
-            else:
-              h = (im.shape[0]//32+1)*32
-            if im.shape[1]%32==0:
-              w = im.shape[1]
-            else:
-              w = (im.shape[1]//32+1)*32
-            _im = np.zeros( (h, w, 3), dtype=np.float32 )
-            _im[0:im.shape[0], 0:im.shape[1], :] = im
-            im = _im
-          else:
-            im = im.astype(np.float32)
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X1 uses', diff.total_seconds(), 'seconds')
-          #self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
-          #im_info = [im.shape[0], im.shape[1], im_scale]
-          im_info = [im.shape[0], im.shape[1]]
-          im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-          for i in range(3):
-              im_tensor[0, i, :, :] = (im[:, :, 2 - i]/self.pixel_scale - self.pixel_means[2 - i])/self.pixel_stds[2-i]
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X2 uses', diff.total_seconds(), 'seconds')
-          data = nd.array(im_tensor)
-          db = mx.io.DataBatch(data=(data,), provide_data=[('data', data.shape)])
-          if self.debug:
-            timeb = datetime.datetime.now()
-            diff = timeb - timea
-            print('X3 uses', diff.total_seconds(), 'seconds')
-          self.model.forward(db, is_train=False)
-          net_out = self.model.get_outputs()
-          #post_nms_topN = self._rpn_post_nms_top_n
-          #min_size_dict = self._rpn_min_size_fpn
-
-          sym_idx = 0
-
-          for _idx,s in enumerate(self._feat_stride_fpn):
-              #if len(scales)>1 and s==32 and im_scale==scales[-1]:
-              #  continue
-              _key = 'stride%s'%s
-              stride = int(s)
-              is_cascade = False
-              #if self.vote and stride==4 and len(scales)>2 and (im_scale==scales[0]):
-              #  continue
-              #print('getting', im_scale, stride, idx, len(net_out), data.shape, file=sys.stderr)
-              scores = net_out[sym_idx].asnumpy()
-              type_scores = net_out[sym_idx+3].asnumpy()
-              print(scores.shape, type_scores.shape)
-              if self.debug:
-                timeb = datetime.datetime.now()
-                diff = timeb - timea
-                print('A uses', diff.total_seconds(), 'seconds')
-              A = self._num_anchors['stride%s'%s]
-              #print(scores.shape)
-              #print('scores',stride, scores.shape, file=sys.stderr)
-              scores = scores[:, A:, :, :]
-              mask_scores = type_scores[:, A*2:, :, :] #x, A, x, x
-
-              bbox_deltas = net_out[sym_idx+1].asnumpy()
-
-              #if DEBUG:
-              #    print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
-              #    print 'scale: {}'.format(im_info[2])
-
-              #_height, _width = int(im_info[0] / stride), int(im_info[1] / stride)
-              height, width = bbox_deltas.shape[2], bbox_deltas.shape[3]
-
-              K = height * width
-              anchors_fpn = self._anchors_fpn['stride%s'%s]
-              anchors = anchors_plane(height, width, stride, anchors_fpn)
-              #print((height, width), (_height, _width), anchors.shape, bbox_deltas.shape, scores.shape, file=sys.stderr)
-              anchors = anchors.reshape((K * A, 4))
-              #print('num_anchors', self._num_anchors['stride%s'%s], file=sys.stderr)
-              #print('HW', (height, width), file=sys.stderr)
-              #print('anchors_fpn', anchors_fpn.shape, file=sys.stderr)
-              #print('anchors', anchors.shape, file=sys.stderr)
-              #print('bbox_deltas', bbox_deltas.shape, file=sys.stderr)
-              #print('scores', scores.shape, file=sys.stderr)
-
-
-              #scores = self._clip_pad(scores, (height, width))
-              scores = scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
-              mask_scores = mask_scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
-
-              #print('pre', bbox_deltas.shape, height, width)
-              #bbox_deltas = self._clip_pad(bbox_deltas, (height, width))
-              #print('after', bbox_deltas.shape, height, width)
-              bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
-              bbox_pred_len = bbox_deltas.shape[3]//A
-              #print(bbox_deltas.shape)
-              bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
-              bbox_deltas[:, 0::4] = bbox_deltas[:,0::4] * self.bbox_stds[0]
-              bbox_deltas[:, 1::4] = bbox_deltas[:,1::4] * self.bbox_stds[1]
-              bbox_deltas[:, 2::4] = bbox_deltas[:,2::4] * self.bbox_stds[2]
-              bbox_deltas[:, 3::4] = bbox_deltas[:,3::4] * self.bbox_stds[3]
-              proposals = self.bbox_pred(anchors, bbox_deltas)
-
-
-
-              proposals = clip_boxes(proposals, im_info[:2])
-
-              #if self.vote:
-              #  if im_scale>1.0:
-              #    keep = self._filter_boxes2(proposals, 160*im_scale, -1)
-              #  else:
-              #    keep = self._filter_boxes2(proposals, -1, 100*im_scale)
-              #  if stride==4:
-              #    keep = self._filter_boxes2(proposals, 12*im_scale, -1)
-              #    proposals = proposals[keep, :]
-              #    scores = scores[keep]
-
-              #keep = self._filter_boxes(proposals, min_size_dict['stride%s'%s] * im_info[2])
-              #proposals = proposals[keep, :]
-              #scores = scores[keep]
-              #print('333', proposals.shape)
-              if stride==4 and self.decay4<1.0:
-                scores *= self.decay4
-
-              scores_ravel = scores.ravel()
-              #mask_scores_ravel = mask_scores.ravel()
-              #print('__shapes', proposals.shape, scores_ravel.shape)
-              #print('max score', np.max(scores_ravel))
-              order = np.where(scores_ravel>=threshold)[0]
-                #_scores = scores_ravel[order]
-                #_order = _scores.argsort()[::-1]
-                #order = order[_order]
-              proposals = proposals[order, :]
-              scores = scores[order]
-              mask_scores = mask_scores[order]
-              if flip:
-                oldx1 = proposals[:, 0].copy()
-                oldx2 = proposals[:, 2].copy()
-                proposals[:, 0] = im.shape[1] - oldx2 - 1
-                proposals[:, 2] = im.shape[1] - oldx1 - 1
-
-              proposals[:,0:4] /= im_scale
-
-              proposals_list.append(proposals)
-              scores_list.append(scores)
-              mask_scores_list.append(mask_scores)
-
-              landmark_deltas = net_out[sym_idx+2].asnumpy()
-              #landmark_deltas = self._clip_pad(landmark_deltas, (height, width))
-              landmark_pred_len = landmark_deltas.shape[1]//A
-              landmark_deltas = landmark_deltas.transpose((0, 2, 3, 1)).reshape((-1, 5, landmark_pred_len//5))
-              landmark_deltas *= self.landmark_std
-              #print(landmark_deltas.shape, landmark_deltas)
-              landmarks = self.landmark_pred(anchors, landmark_deltas)
-              landmarks = landmarks[order, :]
-
-              if flip:
-                landmarks[:,:,0] = im.shape[1] - landmarks[:,:,0] - 1
-                #for a in range(5):
-                #  oldx1 = landmarks[:, a].copy()
-                #  landmarks[:,a] = im.shape[1] - oldx1 - 1
-                order = [1,0,2,4,3]
-                flandmarks = landmarks.copy()
-                for idx, a in enumerate(order):
-                  flandmarks[:,idx,:] = landmarks[:,a,:]
-                  #flandmarks[:, idx*2] = landmarks[:,a*2]
-                  #flandmarks[:, idx*2+1] = landmarks[:,a*2+1]
-                landmarks = flandmarks
-              landmarks[:,:,0:2] /= im_scale
-              #landmarks /= im_scale
-              #landmarks = landmarks.reshape( (-1, landmark_pred_len) )
-              landmarks_list.append(landmarks)
-              #proposals = np.hstack((proposals, landmarks))
-              sym_idx += 4
-
-    if self.debug:
-      timeb = datetime.datetime.now()
-      diff = timeb - timea
-      print('B uses', diff.total_seconds(), 'seconds')
-    proposals = np.vstack(proposals_list)
-    landmarks = None
-    if proposals.shape[0]==0:
-      landmarks = np.zeros( (0,5,2) )
-      return np.zeros( (0,6) ), landmarks
-    scores = np.vstack(scores_list)
-    mask_scores = np.vstack(mask_scores_list)
-    #print('shapes', proposals.shape, scores.shape)
-    scores_ravel = scores.ravel()
-    order = scores_ravel.argsort()[::-1]
-    #if config.TEST.SCORE_THRESH>0.0:
-    #  _count = np.sum(scores_ravel>config.TEST.SCORE_THRESH)
-    #  order = order[:_count]
-    proposals = proposals[order, :]
-    scores = scores[order]
-    mask_scores = mask_scores[order]
-    landmarks = np.vstack(landmarks_list)
-    landmarks = landmarks[order].astype(np.float32, copy=False)
-
-    pre_det = np.hstack((proposals[:,0:4], scores)).astype(np.float32, copy=False)
-    keep = self.nms(pre_det)
-    det = np.hstack( (pre_det, mask_scores) )
-    det = det[keep, :]
-    landmarks = landmarks[keep]
-
-
-    if self.debug:
-      timeb = datetime.datetime.now()
-      diff = timeb - timea
-      print('C uses', diff.total_seconds(), 'seconds')
-    return det, landmarks
-
-  def detect_center(self, img, threshold=0.5, scales=[1.0], do_flip=False):
-    det, landmarks = self.detect(img, threshold, scales, do_flip)
-    if det.shape[0]==0:
-      return None, None
-    bindex = 0
-    if det.shape[0]>1:
-      img_size = np.asarray(img.shape)[0:2]
-      bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-      img_center = img_size / 2
-      offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-      offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-      bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-    bbox = det[bindex,:]
-    landmark = landmarks[bindex, :, :]
-    return bbox, landmark
-
-  @staticmethod
-  def check_large_pose(landmark, bbox):
-    assert landmark.shape==(5,2)
-    assert len(bbox)==4
-    def get_theta(base, x, y):
-      vx = x-base
-      vy = y-base
-      vx[1] *= -1
-      vy[1] *= -1
-      tx = np.arctan2(vx[1], vx[0])
-      ty = np.arctan2(vy[1], vy[0])
-      d = ty-tx
-      d = np.degrees(d)
-      #print(vx, tx, vy, ty, d)
-      #if d<-1.*math.pi:
-      #  d+=2*math.pi
-      #elif d>math.pi:
-      #  d-=2*math.pi
-      if d<-180.0:
-        d+=360.
-      elif d>180.0:
-        d-=360.0
-      return d
-    landmark = landmark.astype(np.float32)
-
-    theta1 = get_theta(landmark[0], landmark[3], landmark[2])
-    theta2 = get_theta(landmark[1], landmark[2], landmark[4])
-    #print(va, vb, theta2)
-    theta3 = get_theta(landmark[0], landmark[2], landmark[1])
-    theta4 = get_theta(landmark[1], landmark[0], landmark[2])
-    theta5 = get_theta(landmark[3], landmark[4], landmark[2])
-    theta6 = get_theta(landmark[4], landmark[2], landmark[3])
-    theta7 = get_theta(landmark[3], landmark[2], landmark[0])
-    theta8 = get_theta(landmark[4], landmark[1], landmark[2])
-    #print(theta1, theta2, theta3, theta4, theta5, theta6, theta7, theta8)
-    left_score = 0.0
-    right_score = 0.0
-    up_score = 0.0
-    down_score = 0.0
-    if theta1<=0.0:
-      left_score = 10.0
-    elif theta2<=0.0:
-      right_score = 10.0
-    else:
-      left_score = theta2/theta1
-      right_score = theta1/theta2
-    if theta3<=10.0 or theta4<=10.0:
-      up_score = 10.0
-    else:
-      up_score = max(theta1/theta3, theta2/theta4)
-    if theta5<=10.0 or theta6<=10.0:
-      down_score = 10.0
-    else:
-      down_score = max(theta7/theta5, theta8/theta6)
-    mleft = (landmark[0][0]+landmark[3][0])/2
-    mright = (landmark[1][0]+landmark[4][0])/2
-    box_center = ( (bbox[0]+bbox[2])/2,  (bbox[1]+bbox[3])/2 )
-    ret = 0
-    if left_score>=3.0:
-      ret = 1
-    if ret==0 and left_score>=2.0:
-      if mright<=box_center[0]:
-        ret = 1
-    if ret==0 and right_score>=3.0:
-      ret = 2
-    if ret==0 and right_score>=2.0:
-      if mleft>=box_center[0]:
-        ret = 2
-    if ret==0 and up_score>=2.0:
-      ret = 3
-    if ret==0 and down_score>=5.0:
-      ret = 4
-    return ret, left_score, right_score, up_score, down_score
-
-  @staticmethod
-  def _filter_boxes(boxes, min_size):
-      """ Remove all boxes with any side smaller than min_size """
-      ws = boxes[:, 2] - boxes[:, 0] + 1
-      hs = boxes[:, 3] - boxes[:, 1] + 1
-      keep = np.where((ws >= min_size) & (hs >= min_size))[0]
-      return keep
-
-  @staticmethod
-  def _filter_boxes2(boxes, max_size, min_size):
-      """ Remove all boxes with any side smaller than min_size """
-      ws = boxes[:, 2] - boxes[:, 0] + 1
-      hs = boxes[:, 3] - boxes[:, 1] + 1
-      if max_size>0:
-        keep = np.where( np.minimum(ws, hs)<max_size )[0]
-      elif min_size>0:
-        keep = np.where( np.maximum(ws, hs)>min_size )[0]
-      return keep
-
-  @staticmethod
-  def _clip_pad(tensor, pad_shape):
-      """
-      Clip boxes of the pad area.
-      :param tensor: [n, c, H, W]
-      :param pad_shape: [h, w]
-      :return: [n, c, h, w]
-      """
-      H, W = tensor.shape[2:]
-      h, w = pad_shape
-
-      if h < H or w < W:
-        tensor = tensor[:, :, :h, :w].copy()
-
-      return tensor
-
-  @staticmethod
-  def bbox_pred(boxes, box_deltas):
-      """
-      Transform the set of class-agnostic boxes into class-specific boxes
-      by applying the predicted offsets (box_deltas)
-      :param boxes: !important [N 4]
-      :param box_deltas: [N, 4 * num_classes]
-      :return: [N 4 * num_classes]
-      """
-      if boxes.shape[0] == 0:
-          return np.zeros((0, box_deltas.shape[1]))
-
-      boxes = boxes.astype(np.float, copy=False)
-      widths = boxes[:, 2] - boxes[:, 0] + 1.0
-      heights = boxes[:, 3] - boxes[:, 1] + 1.0
-      ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-      ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-      dx = box_deltas[:, 0:1]
-      dy = box_deltas[:, 1:2]
-      dw = box_deltas[:, 2:3]
-      dh = box_deltas[:, 3:4]
-
-      pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-      pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-      pred_w = np.exp(dw) * widths[:, np.newaxis]
-      pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-      pred_boxes = np.zeros(box_deltas.shape)
-      # x1
-      pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-      # y1
-      pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-      # x2
-      pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-      # y2
-      pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-
-      if box_deltas.shape[1]>4:
-        pred_boxes[:,4:] = box_deltas[:,4:]
-
-      return pred_boxes
-
-  @staticmethod
-  def landmark_pred(boxes, landmark_deltas):
-      if boxes.shape[0] == 0:
-          return np.zeros((0, landmark_deltas.shape[1]))
-      boxes = boxes.astype(np.float, copy=False)
-      widths = boxes[:, 2] - boxes[:, 0] + 1.0
-      heights = boxes[:, 3] - boxes[:, 1] + 1.0
-      ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-      ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-      pred = landmark_deltas.copy()
-      for i in range(5):
-        pred[:,i,0] = landmark_deltas[:,i,0]*widths + ctr_x
-        pred[:,i,1] = landmark_deltas[:,i,1]*heights + ctr_y
-      return pred
-      #preds = []
-      #for i in range(landmark_deltas.shape[1]):
-      #  if i%2==0:
-      #    pred = (landmark_deltas[:,i]*widths + ctr_x)
-      #  else:
-      #    pred = (landmark_deltas[:,i]*heights + ctr_y)
-      #  preds.append(pred)
-      #preds = np.vstack(preds).transpose()
-      #return preds
-
-  def vote(self, det):
-      #order = det[:, 4].ravel().argsort()[::-1]
-      #det = det[order, :]
-      if det.shape[0] == 0:
-        return np.zeros( (0, 5) )
-          #dets = np.array([[10, 10, 20, 20, 0.002]])
-          #det = np.empty(shape=[0, 5])
-      dets = None
-      while det.shape[0] > 0:
-          if dets is not None and dets.shape[0]>=750:
-              break
-          # IOU
-          area = (det[:, 2] - det[:, 0] + 1) * (det[:, 3] - det[:, 1] + 1)
-          xx1 = np.maximum(det[0, 0], det[:, 0])
-          yy1 = np.maximum(det[0, 1], det[:, 1])
-          xx2 = np.minimum(det[0, 2], det[:, 2])
-          yy2 = np.minimum(det[0, 3], det[:, 3])
-          w = np.maximum(0.0, xx2 - xx1 + 1)
-          h = np.maximum(0.0, yy2 - yy1 + 1)
-          inter = w * h
-          o = inter / (area[0] + area[:] - inter)
-
-          # nms
-          merge_index = np.where(o >= self.nms_threshold)[0]
-          det_accu = det[merge_index, :]
-          det = np.delete(det, merge_index, 0)
-          if merge_index.shape[0] <= 1:
-              if det.shape[0] == 0:
-                  try:
-                      dets = np.row_stack((dets, det_accu))
-                  except:
-                      dets = det_accu
-              continue
-          det_accu[:, 0:4] = det_accu[:, 0:4] * np.tile(det_accu[:, -1:], (1, 4))
-          max_score = np.max(det_accu[:, 4])
-          det_accu_sum = np.zeros((1, 5))
-          det_accu_sum[:, 0:4] = np.sum(det_accu[:, 0:4],
-                                        axis=0) / np.sum(det_accu[:, -1:])
-          det_accu_sum[:, 4] = max_score
-          if dets is None:
-              dets = det_accu_sum
-          else:
-              dets = np.row_stack((dets, det_accu_sum))
-      dets = dets[0:750, :]
-      return dets
-
diff --git a/RetinaFaceAntiCov/test.py b/RetinaFaceAntiCov/test.py
deleted file mode 100644
index be177dcb0..000000000
--- a/RetinaFaceAntiCov/test.py
+++ /dev/null
@@ -1,65 +0,0 @@
-import cv2
-import sys
-import numpy as np
-import datetime
-import os
-import glob
-from retinaface_cov import RetinaFaceCoV
-
-thresh = 0.8
-mask_thresh = 0.2
-scales = [640, 1080]
-
-count = 1
-
-gpuid = 0
-#detector = RetinaFaceCoV('./model/mnet_cov1', 0, gpuid, 'net3')
-detector = RetinaFaceCoV('./model/mnet_cov2', 0, gpuid, 'net3l')
-
-img = cv2.imread('n1.jpg')
-print(img.shape)
-im_shape = img.shape
-target_size = scales[0]
-max_size = scales[1]
-im_size_min = np.min(im_shape[0:2])
-im_size_max = np.max(im_shape[0:2])
-#im_scale = 1.0
-#if im_size_min>target_size or im_size_max>max_size:
-im_scale = float(target_size) / float(im_size_min)
-# prevent bigger axis from being more than max_size:
-if np.round(im_scale * im_size_max) > max_size:
-    im_scale = float(max_size) / float(im_size_max)
-
-print('im_scale', im_scale)
-
-scales = [im_scale]
-flip = False
-
-for c in range(count):
-  faces, landmarks = detector.detect(img, thresh, scales=scales, do_flip=flip)
-
-
-if faces is not None:
-  print('find', faces.shape[0], 'faces')
-  for i in range(faces.shape[0]):
-    #print('score', faces[i][4])
-    face = faces[i]
-    box = face[0:4].astype(np.int)
-    mask = face[5]
-    print(i,box,mask)
-    #color = (255,0,0)
-    if mask>=mask_thresh:
-      color = (0,0,255)
-    else:
-      color = (0,255,0)
-    cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
-    landmark5 = landmarks[i].astype(np.int)
-    #print(landmark.shape)
-    for l in range(landmark5.shape[0]):
-      color = (255,0,0)
-      cv2.circle(img, (landmark5[l][0], landmark5[l][1]), 1, color, 2)
-
-  filename = './cov_test.jpg'
-  print('writing', filename)
-  cv2.imwrite(filename, img)
-
diff --git a/alignment/README.md b/alignment/README.md
index 9656d13f6..bacfbd80a 100644
--- a/alignment/README.md
+++ b/alignment/README.md
@@ -1,4 +1,42 @@
-You can now find heatmap based approaches under ``heatmapReg`` directory.
+## Face Alignment
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/logo3.jpg" width="240"/>
+</div>
+
+
+## Introduction
+
+These are the face alignment methods of [InsightFace](https://insightface.ai)
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/thumb_sdunet.png" width="600"/>
+</div>
+
+
+### Datasets
+
+  Please refer to [datasets](_datasets_) page for the details of face alignment datasets used for training and evaluation.
+
+### Evaluation
+
+  Please refer to [evaluation](_evaluation_) page for the details of face alignment evaluation.
+
+
+## Methods
+
+
+Supported methods:
+
+- [x] [SDUNets (BMVC'2018)](heatmap)
+- [x] [SimpleRegression](coordinate_reg)
+- [x] [Alignment By Face Synthetics](synthetics)
+
+
+## Contributing
+
+We appreciate all contributions to improve the face alignment model zoo of InsightFace. 
 
-You can now find coordinate regression approaches under ``coordinateReg`` directory.
 
diff --git a/alignment/_datasets_/README.md b/alignment/_datasets_/README.md
new file mode 100644
index 000000000..dcde280b4
--- /dev/null
+++ b/alignment/_datasets_/README.md
@@ -0,0 +1,57 @@
+# Face Alignment Datasets
+
+(Updating)
+
+## Training Datasets
+
+### Menpo2D-Train
+
+https://ibug.doc.ic.ac.uk/resources/2nd-facial-landmark-tracking-competition-menpo-ben/
+
+### 300W-Train
+
+https://ibug.doc.ic.ac.uk/resources/300-W/
+
+
+### LFPW
+
+https://neerajkumar.org/databases/lfpw/
+
+### Helen
+
+http://www.ifp.illinois.edu/~vuongle2/helen/
+
+### AFW
+
+### AFLW
+
+https://www.tugraz.at/institute/icg/research/team-bischof/learning-recognition-surveillance/downloads/aflw
+
+### FDDB
+
+
+### Face Synthetics
+
+https://github.com/microsoft/FaceSynthetics
+
+### 300W-LP (3D annotation)
+
+http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm
+
+## Test Datasets
+
+### 300W-Test
+
+https://ibug.doc.ic.ac.uk/resources/300-W/
+
+### COFW
+
+http://www.vision.caltech.edu/xpburgos/ICCV13/#dataset
+
+### Menpo2D-Test
+
+https://ibug.doc.ic.ac.uk/resources/2nd-facial-landmark-tracking-competition-menpo-ben/
+
+### AFLW2000-3D (3D annotation)
+
+http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm
diff --git a/alignment/coordinateReg/README.md b/alignment/coordinateReg/README.md
deleted file mode 100644
index 39f1fbb21..000000000
--- a/alignment/coordinateReg/README.md
+++ /dev/null
@@ -1,54 +0,0 @@
-### Introduction
-
- Here we provide some lightweight facial landmark models with fast coordinate regression.
- The input of these models is loose cropped face image while the output is the direct landmark coordinates. 
-
-
-### Pretrained Models
-
-- **Model ``2d106det``**
-
-  Given face detection bounding box, predict 2d-106 landmarks. Mainly used for static image inference. Please check ``image_infer.py`` for detail.
-  
-  Backbone: MobileNet-0.5, size 5MB.
-  
-  Input: size 192x192, loose cropped detection bounding-box.
-  
-  Download link:
-  
-  [baidu cloud](https://pan.baidu.com/s/10m5GmtNV5snynDrq3KqIdg) (code: ``lqvv``)
-  
-  [google drive](https://drive.google.com/file/d/1MBWbTEYRhZFzj_O2f2Dc6fWGXFWtbMFw/view?usp=sharing)
-
-- **Model ``2d106track``** 
-
-  Given landmarks bounding box, predict 2d-106 landmarks. Used for video landmarks tracking.
-  
-  Download link: coming soon
-
-### Visualization
-
-
-<p align="center">Points mark-up(ordered by point names):</p>
-
-<div align="center">
-	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/2d106markup.jpg" alt="markup" width="320">
-</div>
-
-
-<p align="center">Image result:</p>
-
-<div align="center">
-	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/t1_out.jpg" alt="imagevis" width="800">
-</div>
-
-
-<p align="center">Video result:</p>
-
-<div align="center">
-	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/C_jiaguo.gif" alt="videovis" width="240">
-</div>
-
-
-### FAQ
-
diff --git a/alignment/coordinateReg/image_infer.py b/alignment/coordinateReg/image_infer.py
deleted file mode 100644
index 3b2fa6c8a..000000000
--- a/alignment/coordinateReg/image_infer.py
+++ /dev/null
@@ -1,154 +0,0 @@
-import argparse
-import cv2
-import sys
-import numpy as np
-import os
-import mxnet as mx
-import datetime
-from skimage import transform as trans
-import insightface
-
-def square_crop(im, S):
-  if im.shape[0]>im.shape[1]:
-    height = S
-    width = int( float(im.shape[1]) / im.shape[0] * S )
-    scale = float(S) / im.shape[0]
-  else:
-    width = S
-    height = int( float(im.shape[0]) / im.shape[1] * S )
-    scale = float(S) / im.shape[1]
-  resized_im = cv2.resize(im, (width, height))
-  det_im = np.zeros( (S, S, 3), dtype=np.uint8 )
-  det_im[:resized_im.shape[0], :resized_im.shape[1], :] = resized_im
-  return det_im, scale
-
-def transform(data, center, output_size, scale, rotation):
-    scale_ratio = scale
-    rot = float(rotation)*np.pi/180.0
-    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
-    t1 = trans.SimilarityTransform(scale=scale_ratio)
-    cx = center[0]*scale_ratio
-    cy = center[1]*scale_ratio
-    t2 = trans.SimilarityTransform(translation=(-1*cx, -1*cy))
-    t3 = trans.SimilarityTransform(rotation=rot)
-    t4 = trans.SimilarityTransform(translation=(output_size/2, output_size/2))
-    t = t1+t2+t3+t4
-    M = t.params[0:2]
-    cropped = cv2.warpAffine(data,M,(output_size, output_size), borderValue = 0.0)
-    return cropped, M
-
-def trans_points2d(pts, M):
-    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
-    for i in range(pts.shape[0]):
-        pt = pts[i]
-        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
-        new_pt = np.dot(M, new_pt)
-        #print('new_pt', new_pt.shape, new_pt)
-        new_pts[i] = new_pt[0:2]
-
-    return new_pts
-
-def trans_points3d(pts, M):
-    scale = np.sqrt(M[0][0]*M[0][0] + M[0][1]*M[0][1])
-    #print(scale)
-    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
-    for i in range(pts.shape[0]):
-        pt = pts[i]
-        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
-        new_pt = np.dot(M, new_pt)
-        #print('new_pt', new_pt.shape, new_pt)
-        new_pts[i][0:2] = new_pt[0:2]
-        new_pts[i][2] = pts[i][2]*scale
-
-    return new_pts
-
-def trans_points(pts, M):
-  if pts.shape[1]==2:
-    return trans_points2d(pts, M)
-  else:
-    return trans_points3d(pts, M)
-
-
-class Handler:
-  def __init__(self, prefix, epoch, im_size=192, det_size=224, ctx_id=0):
-    print('loading',prefix, epoch)
-    if ctx_id>=0:
-      ctx = mx.gpu(ctx_id)
-    else:
-      ctx = mx.cpu()
-    image_size = (im_size, im_size)
-    self.detector = insightface.model_zoo.get_model('retinaface_mnet025_v2') #can replace with your own face detector
-    #self.detector = insightface.model_zoo.get_model('retinaface_r50_v1')
-    self.detector.prepare(ctx_id=ctx_id)
-    self.det_size = det_size
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    self.image_size = image_size
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    model.bind(for_training=False, data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    self.model = model
-    self.image_size = image_size
-
-
-  
-  def get(self, img, get_all=False):
-    out = []
-    det_im, det_scale = square_crop(img, self.det_size)
-    bboxes, _ = self.detector.detect(det_im)
-    if bboxes.shape[0]==0:
-        return out
-    bboxes /= det_scale
-    if not get_all:
-      areas = []
-      for i in range(bboxes.shape[0]):
-        x = bboxes[i]
-        area = (x[2]-x[0])*(x[3]-x[1])
-        areas.append(area)
-      m = np.argsort(areas)[-1]
-      bboxes = bboxes[m:m+1]
-    for i in range(bboxes.shape[0]):
-      bbox = bboxes[i]
-      input_blob = np.zeros( (1, 3)+self.image_size,dtype=np.float32)
-      w, h = (bbox[2]-bbox[0]), (bbox[3]-bbox[1])
-      center = (bbox[2]+bbox[0])/2, (bbox[3]+bbox[1])/2
-      rotate = 0
-      _scale = self.image_size[0]*2/3.0/max(w,h)
-      rimg, M = transform(img, center, self.image_size[0], _scale, rotate)
-      rimg = cv2.cvtColor(rimg, cv2.COLOR_BGR2RGB)
-      rimg = np.transpose(rimg, (2,0,1)) #3*112*112, RGB
-      input_blob[0] = rimg
-      data = mx.nd.array(input_blob)
-      db = mx.io.DataBatch(data=(data,))
-      self.model.forward(db, is_train=False)
-      pred = self.model.get_outputs()[-1].asnumpy()[0]
-      if pred.shape[0]>=3000:
-        pred = pred.reshape( (-1, 3) )
-      else:
-        pred = pred.reshape( (-1, 2) )
-      pred[:,0:2] += 1
-      pred[:,0:2] *= (self.image_size[0]//2)
-      if pred.shape[1]==3:
-        pred[:,2] *= (self.image_size[0]//2)
-
-      IM = cv2.invertAffineTransform(M)
-      pred = trans_points(pred, IM)
-      out.append(pred)
-    return out
-
-
-if __name__ == '__main__':
-  handler = Handler('./model/2d106_det', 0, ctx_id=7, det_size=640)
-  im = cv2.imread('../../sample-images/t1.jpg')
-  tim = im.copy()
-  preds = handler.get(im, get_all=True)
-  color = (200, 160, 75)
-  for pred in preds:
-    pred = np.round(pred).astype(np.int)
-    for i in range(pred.shape[0]):
-      p = tuple(pred[i])
-      cv2.circle(tim, p, 1, color, 1,cv2.LINE_AA)
-  cv2.imwrite('./test_out.jpg', tim)
-
-
diff --git a/alignment/coordinate_reg/README.md b/alignment/coordinate_reg/README.md
new file mode 100644
index 000000000..c448cdb4b
--- /dev/null
+++ b/alignment/coordinate_reg/README.md
@@ -0,0 +1,58 @@
+### Introduction
+
+ Here we provide some lightweight facial landmark models with fast coordinate regression.
+ The input of these models is loose cropped face image while the output is the direct landmark coordinates. 
+
+
+### Pretrained Models
+
+- **Model ``2d106det``**
+
+  **2021.07: We now support model inference by our `insightface` python package, please check [image_infer.py](image_infer.py) for detail.**
+
+  Given face detection bounding box, predict 2d-106 landmarks. Mainly used for static image inference. 
+  
+  Backbone: MobileNet-0.5, size 5MB.
+  
+  Input: size 192x192, loose cropped detection bounding-box.
+  
+  Download link:
+  
+  [baidu cloud](https://pan.baidu.com/s/10m5GmtNV5snynDrq3KqIdg) (code: ``lqvv``)
+  
+  [google drive](https://drive.google.com/file/d/13Pz8mH-a1s7RXpq_jFUXxaqCpDUE0oSr/view?usp=sharing)
+  
+  
+
+- **Model ``2d106track``** 
+
+  Given landmarks bounding box, predict 2d-106 landmarks. Used for video landmarks tracking.
+  
+  Download link: coming soon
+
+### Visualization
+
+
+<p align="center">Points mark-up(ordered by point names):</p>
+
+<div align="center">
+	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/2d106markup.jpg" alt="markup" width="320">
+</div>
+
+
+<p align="center">Image result:</p>
+
+<div align="center">
+	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/t1_out.jpg" alt="imagevis" width="800">
+</div>
+
+
+<p align="center">Video result:</p>
+
+<div align="center">
+	<img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/C_jiaguo.gif" alt="videovis" width="240">
+</div>
+
+
+### FAQ
+
diff --git a/alignment/coordinate_reg/image_infer.py b/alignment/coordinate_reg/image_infer.py
new file mode 100644
index 000000000..a336221e8
--- /dev/null
+++ b/alignment/coordinate_reg/image_infer.py
@@ -0,0 +1,23 @@
+import cv2
+import numpy as np
+import os
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+if __name__ == '__main__':
+    app = FaceAnalysis(allowed_modules=['detection', 'landmark_2d_106'])
+    app.prepare(ctx_id=0, det_size=(640, 640))
+    img = ins_get_image('t1')
+    faces = app.get(img)
+    #assert len(faces)==6
+    tim = img.copy()
+    color = (200, 160, 75)
+    for face in faces:
+        lmk = face.landmark_2d_106
+        lmk = np.round(lmk).astype(np.int)
+        for i in range(lmk.shape[0]):
+            p = tuple(lmk[i])
+            cv2.circle(tim, p, 1, color, 1, cv2.LINE_AA)
+    cv2.imwrite('./test_out.jpg', tim)
+
diff --git a/alignment/heatmapReg/README.md b/alignment/heatmap/README.md
similarity index 100%
rename from alignment/heatmapReg/README.md
rename to alignment/heatmap/README.md
diff --git a/alignment/heatmap/data.py b/alignment/heatmap/data.py
new file mode 100644
index 000000000..eeaca9c51
--- /dev/null
+++ b/alignment/heatmap/data.py
@@ -0,0 +1,354 @@
+# pylint: skip-file
+import mxnet as mx
+import numpy as np
+import sys, os
+import random
+import math
+import scipy.misc
+import cv2
+import logging
+import sklearn
+import datetime
+import img_helper
+from mxnet.io import DataIter
+from mxnet import ndarray as nd
+from mxnet import io
+from mxnet import recordio
+from PIL import Image
+from config import config
+from skimage import transform as tf
+
+
+class FaceSegIter(DataIter):
+    def __init__(self,
+                 batch_size,
+                 per_batch_size=0,
+                 path_imgrec=None,
+                 aug_level=0,
+                 force_mirror=False,
+                 exf=1,
+                 use_coherent=0,
+                 args=None,
+                 data_name="data",
+                 label_name="softmax_label"):
+        self.aug_level = aug_level
+        self.force_mirror = force_mirror
+        self.use_coherent = use_coherent
+        self.exf = exf
+        self.batch_size = batch_size
+        self.per_batch_size = per_batch_size
+        self.data_name = data_name
+        self.label_name = label_name
+        assert path_imgrec
+        logging.info('loading recordio %s...', path_imgrec)
+        path_imgidx = path_imgrec[0:-4] + ".idx"
+        self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec,
+                                                    'r')  # pylint: disable=redefined-variable-type
+        self.oseq = list(self.imgrec.keys)
+        print('train size', len(self.oseq))
+        self.cur = 0
+        self.reset()
+        self.data_shape = (3, config.input_img_size, config.input_img_size)
+        self.num_classes = config.num_classes
+        self.input_img_size = config.input_img_size
+        #self.label_classes = self.num_classes
+        if config.losstype == 'heatmap':
+            if aug_level > 0:
+                self.output_label_size = config.output_label_size
+                self.label_shape = (self.num_classes, self.output_label_size,
+                                    self.output_label_size)
+            else:
+                self.output_label_size = self.input_img_size
+                #self.label_shape = (self.num_classes, 2)
+                self.label_shape = (self.num_classes, self.output_label_size,
+                                    self.output_label_size)
+        else:
+            if aug_level > 0:
+                self.output_label_size = config.output_label_size
+                self.label_shape = (self.num_classes, 2)
+            else:
+                self.output_label_size = self.input_img_size
+                #self.label_shape = (self.num_classes, 2)
+                self.label_shape = (self.num_classes, 2)
+        self.provide_data = [(data_name, (batch_size, ) + self.data_shape)]
+        self.provide_label = [(label_name, (batch_size, ) + self.label_shape)]
+        self.img_num = 0
+        self.invalid_num = 0
+        self.mode = 1
+        self.vis = 0
+        self.stats = [0, 0]
+        self.flip_order = [
+            16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 26, 25,
+            24, 23, 22, 21, 20, 19, 18, 17, 27, 28, 29, 30, 35, 34, 33, 32, 31,
+            45, 44, 43, 42, 47, 46, 39, 38, 37, 36, 41, 40, 54, 53, 52, 51, 50,
+            49, 48, 59, 58, 57, 56, 55, 64, 63, 62, 61, 60, 67, 66, 65
+        ]
+        #self.mirror_set = [
+        #        (22,23),
+        #        (21,24),
+        #        (20,25),
+        #        (19,26),
+        #        (18,27),
+        #        (40,43),
+        #        (39,44),
+        #        (38,45),
+        #        (37,46),
+        #        (42,47),
+        #        (41,48),
+        #        (33,35),
+        #        (32,36),
+        #        (51,53),
+        #        (50,54),
+        #        (62,64),
+        #        (61,65),
+        #        (49,55),
+        #        (49,55),
+        #        (68,66),
+        #        (60,56),
+        #        (59,57),
+        #        (1,17),
+        #        (2,16),
+        #        (3,15),
+        #        (4,14),
+        #        (5,13),
+        #        (6,12),
+        #        (7,11),
+        #        (8,10),
+        #    ]
+
+    def get_data_shape(self):
+        return self.data_shape
+
+    #def get_label_shape(self):
+    #    return self.label_shape
+
+    def get_shape_dict(self):
+        D = {}
+        for (k, v) in self.provide_data:
+            D[k] = v
+        for (k, v) in self.provide_label:
+            D[k] = v
+        return D
+
+    def get_label_names(self):
+        D = []
+        for (k, v) in self.provide_label:
+            D.append(k)
+        return D
+
+    def reset(self):
+        #print('reset')
+        if self.aug_level == 0:
+            self.seq = self.oseq
+        else:
+            self.seq = []
+            for _ in range(self.exf):
+                _seq = self.oseq[:]
+                random.shuffle(_seq)
+                self.seq += _seq
+            print('train size after reset', len(self.seq))
+        self.cur = 0
+
+    def next_sample(self):
+        """Helper function for reading in next sample."""
+        if self.cur >= len(self.seq):
+            raise StopIteration
+        idx = self.seq[self.cur]
+        self.cur += 1
+        s = self.imgrec.read_idx(idx)
+        header, img = recordio.unpack(s)
+        img = mx.image.imdecode(img).asnumpy()
+        hlabel = np.array(header.label).reshape((self.num_classes, 2))
+        if not config.label_xfirst:
+            hlabel = hlabel[:, ::-1]  #convert to X/W first
+        annot = {'scale': config.base_scale}
+
+        #ul = np.array( (50000,50000), dtype=np.int32)
+        #br = np.array( (0,0), dtype=np.int32)
+        #for i in range(hlabel.shape[0]):
+        #  h = int(hlabel[i][0])
+        #  w = int(hlabel[i][1])
+        #  key = np.array((h,w))
+        #  ul = np.minimum(key, ul)
+        #  br = np.maximum(key, br)
+
+        return img, hlabel, annot
+
+    def get_flip(self, data, label):
+        data_flip = np.zeros_like(data)
+        label_flip = np.zeros_like(label)
+        for k in range(data_flip.shape[2]):
+            data_flip[:, :, k] = np.fliplr(data[:, :, k])
+        for k in range(label_flip.shape[0]):
+            label_flip[k, :] = np.fliplr(label[k, :])
+        #print(label[0,:].shape)
+        label_flip = label_flip[self.flip_order, :]
+        return data_flip, label_flip
+
+    def get_data(self, data, label, annot):
+        if self.vis:
+            self.img_num += 1
+            #if self.img_num<=self.vis:
+            #  filename = './vis/raw_%d.jpg' % (self.img_num)
+            #  print('save', filename)
+            #  draw = data.copy()
+            #  for i in range(label.shape[0]):
+            #    cv2.circle(draw, (label[i][1], label[i][0]), 1, (0, 0, 255), 2)
+            #  scipy.misc.imsave(filename, draw)
+
+        rotate = 0
+        #scale = 1.0
+        if 'scale' in annot:
+            scale = annot['scale']
+        else:
+            scale = max(data.shape[0], data.shape[1])
+        if 'center' in annot:
+            center = annot['center']
+        else:
+            center = np.array((data.shape[1] / 2, data.shape[0] / 2))
+        max_retry = 3
+        if self.aug_level == 0:  #validation mode
+            max_retry = 6
+        retry = 0
+        found = False
+        base_scale = scale
+        while retry < max_retry:
+            retry += 1
+            succ = True
+            _scale = base_scale
+            if self.aug_level > 0:
+                rotate = np.random.randint(-40, 40)
+                scale_config = 0.2
+                #rotate = 0
+                #scale_config = 0.0
+                scale_ratio = min(
+                    1 + scale_config,
+                    max(1 - scale_config,
+                        (np.random.randn() * scale_config) + 1))
+                _scale = int(base_scale * scale_ratio)
+                #translate = np.random.randint(-5, 5, size=(2,))
+                #center += translate
+            data_out, trans = img_helper.transform(data, center,
+                                                   self.input_img_size, _scale,
+                                                   rotate)
+            #data_out = img_helper.crop2(data, center, _scale, (self.input_img_size, self.input_img_size), rot=rotate)
+            label_out = np.zeros(self.label_shape, dtype=np.float32)
+            #print('out shapes', data_out.shape, label_out.shape)
+            for i in range(label.shape[0]):
+                pt = label[i].copy()
+                #pt = pt[::-1]
+                npt = img_helper.transform_pt(pt, trans)
+                if npt[0] >= data_out.shape[1] or npt[1] >= data_out.shape[
+                        0] or npt[0] < 0 or npt[1] < 0:
+                    succ = False
+                    #print('err npt', npt)
+                    break
+                if config.losstype == 'heatmap':
+                    pt_scale = float(
+                        self.output_label_size) / self.input_img_size
+                    npt *= pt_scale
+                    npt = npt.astype(np.int32)
+                    img_helper.gaussian(label_out[i], npt, config.gaussian)
+                else:
+                    label_out[i] = (npt / self.input_img_size)
+                #print('before gaussian', label_out[i].shape, pt.shape)
+                #trans = img_helper.transform(pt, center, _scale, (self.output_label_size, self.output_label_size), rot=rotate)
+                #print(trans.shape)
+                #if not img_helper.gaussian(label_out[i], trans, _g):
+                #    succ = False
+                #    break
+            if not succ:
+                if self.aug_level == 0:
+                    base_scale += 20
+                continue
+
+            flip_data_out = None
+            flip_label_out = None
+            if config.net_coherent:
+                flip_data_out, flip_label_out = self.get_flip(
+                    data_out, label_out)
+            elif ((self.aug_level > 0 and np.random.rand() < 0.5)
+                  or self.force_mirror):  #flip aug
+                flip_data_out, flip_label_out = self.get_flip(
+                    data_out, label_out)
+                data_out, label_out = flip_data_out, flip_label_out
+
+            found = True
+            break
+
+        #self.stats[0]+=1
+        if not found:
+            #self.stats[1]+=1
+            #print('find aug error', retry)
+            #print(self.stats)
+            #print('!!!ERR')
+            return None
+        #print('found with scale', _scale, rotate)
+
+        if self.vis > 0 and self.img_num <= self.vis:
+            print('crop', data.shape, center, _scale, rotate, data_out.shape)
+            filename = './vis/cropped_%d.jpg' % (self.img_num)
+            print('save', filename)
+            draw = data_out.copy()
+            alabel = label_out.copy()
+            for i in range(label.shape[0]):
+                a = cv2.resize(alabel[i],
+                               (self.input_img_size, self.input_img_size))
+                ind = np.unravel_index(np.argmax(a, axis=None), a.shape)
+                cv2.circle(draw, (ind[1], ind[0]), 1, (0, 0, 255), 2)
+            scipy.misc.imsave(filename, draw)
+            filename = './vis/raw_%d.jpg' % (self.img_num)
+            scipy.misc.imsave(filename, data)
+
+        return data_out, label_out, flip_data_out, flip_label_out
+
+    def next(self):
+        """Returns the next batch of data."""
+        #print('next')
+        batch_size = self.batch_size
+        batch_data = nd.empty((batch_size, ) + self.data_shape)
+        batch_label = nd.empty((batch_size, ) + self.label_shape)
+        i = 0
+        #self.cutoff = random.randint(800,1280)
+        try:
+            while i < batch_size:
+                #print('N', i)
+                data, label, annot = self.next_sample()
+                R = self.get_data(data, label, annot)
+                if R is None:
+                    continue
+                data_out, label_out, flip_data_out, flip_label_out = R
+                if not self.use_coherent:
+                    data = nd.array(data_out)
+                    data = nd.transpose(data, axes=(2, 0, 1))
+                    label = nd.array(label_out)
+                    #print(data.shape, label.shape)
+                    batch_data[i][:] = data
+                    batch_label[i][:] = label
+                    i += 1
+                else:
+                    data = nd.array(data_out)
+                    data = nd.transpose(data, axes=(2, 0, 1))
+                    label = nd.array(label_out)
+                    data2 = nd.array(flip_data_out)
+                    data2 = nd.transpose(data2, axes=(2, 0, 1))
+                    label2 = nd.array(flip_label_out)
+                    #M = nd.array(M)
+                    #print(data.shape, label.shape)
+                    batch_data[i][:] = data
+                    batch_label[i][:] = label
+                    #i+=1
+                    j = i + self.per_batch_size // 2
+                    batch_data[j][:] = data2
+                    batch_label[j][:] = label2
+                    i += 1
+                    if j % self.per_batch_size == self.per_batch_size - 1:
+                        i = j + 1
+        except StopIteration:
+            if i < batch_size:
+                raise StopIteration
+
+        #return {self.data_name  :  batch_data,
+        #        self.label_name :  batch_label}
+        #print(batch_data.shape, batch_label.shape)
+        return mx.io.DataBatch([batch_data], [batch_label], batch_size - i)
diff --git a/alignment/heatmap/img_helper.py b/alignment/heatmap/img_helper.py
new file mode 100644
index 000000000..b37123ff9
--- /dev/null
+++ b/alignment/heatmap/img_helper.py
@@ -0,0 +1,86 @@
+import numpy as np
+import math
+import cv2
+from skimage import transform as stf
+
+
+def transform(data, center, output_size, scale, rotation):
+    scale_ratio = float(output_size) / scale
+    rot = float(rotation) * np.pi / 180.0
+    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
+    t1 = stf.SimilarityTransform(scale=scale_ratio)
+    cx = center[0] * scale_ratio
+    cy = center[1] * scale_ratio
+    t2 = stf.SimilarityTransform(translation=(-1 * cx, -1 * cy))
+    t3 = stf.SimilarityTransform(rotation=rot)
+    t4 = stf.SimilarityTransform(translation=(output_size / 2,
+                                              output_size / 2))
+    t = t1 + t2 + t3 + t4
+    trans = t.params[0:2]
+    #print('M', scale, rotation, trans)
+    cropped = cv2.warpAffine(data,
+                             trans, (output_size, output_size),
+                             borderValue=0.0)
+    return cropped, trans
+
+
+def transform_pt(pt, trans):
+    new_pt = np.array([pt[0], pt[1], 1.]).T
+    new_pt = np.dot(trans, new_pt)
+    #print('new_pt', new_pt.shape, new_pt)
+    return new_pt[:2]
+
+
+def gaussian(img, pt, sigma):
+    # Draw a 2D gaussian
+    assert (sigma >= 0)
+    if sigma == 0:
+        img[pt[1], pt[0]] = 1.0
+        return True
+    #assert pt[0]<=img.shape[1]
+    #assert pt[1]<=img.shape[0]
+
+    # Check that any part of the gaussian is in-bounds
+    ul = [int(pt[0] - 3 * sigma), int(pt[1] - 3 * sigma)]
+    br = [int(pt[0] + 3 * sigma + 1), int(pt[1] + 3 * sigma + 1)]
+    if (ul[0] > img.shape[1] or ul[1] >= img.shape[0] or br[0] < 0
+            or br[1] < 0):
+        # If not, just return the image as is
+        #print('gaussian error')
+        return False
+        #return img
+
+    # Generate gaussian
+    size = 6 * sigma + 1
+    x = np.arange(0, size, 1, float)
+    y = x[:, np.newaxis]
+    x0 = y0 = size // 2
+    # The gaussian is not normalized, we want the center value to equal 1
+    g = np.exp(-((x - x0)**2 + (y - y0)**2) / (2 * sigma**2))
+
+    # Usable gaussian range
+    g_x = max(0, -ul[0]), min(br[0], img.shape[1]) - ul[0]
+    g_y = max(0, -ul[1]), min(br[1], img.shape[0]) - ul[1]
+    # Image range
+    img_x = max(0, ul[0]), min(br[0], img.shape[1])
+    img_y = max(0, ul[1]), min(br[1], img.shape[0])
+
+    img[img_y[0]:img_y[1], img_x[0]:img_x[1]] = g[g_y[0]:g_y[1], g_x[0]:g_x[1]]
+    return True
+    #return img
+
+
+def estimate_trans_bbox(face, input_size, s=2.0):
+    w = face[2] - face[0]
+    h = face[3] - face[1]
+    wc = int((face[2] + face[0]) / 2)
+    hc = int((face[3] + face[1]) / 2)
+    im_size = max(w, h)
+    #size = int(im_size*1.2)
+    scale = input_size / (max(w, h) * s)
+    M = [
+        [scale, 0, input_size / 2 - wc * scale],
+        [0, scale, input_size / 2 - hc * scale],
+    ]
+    M = np.array(M)
+    return M
diff --git a/alignment/heatmap/metric.py b/alignment/heatmap/metric.py
new file mode 100644
index 000000000..2ddc96c9f
--- /dev/null
+++ b/alignment/heatmap/metric.py
@@ -0,0 +1,107 @@
+import mxnet as mx
+import numpy as np
+import math
+import cv2
+from config import config
+
+
+class LossValueMetric(mx.metric.EvalMetric):
+    def __init__(self):
+        self.axis = 1
+        super(LossValueMetric, self).__init__('lossvalue',
+                                              axis=self.axis,
+                                              output_names=None,
+                                              label_names=None)
+        self.losses = []
+
+    def update(self, labels, preds):
+        loss = preds[0].asnumpy()[0]
+        self.sum_metric += loss
+        self.num_inst += 1.0
+
+
+class NMEMetric(mx.metric.EvalMetric):
+    def __init__(self):
+        self.axis = 1
+        super(NMEMetric, self).__init__('NME',
+                                        axis=self.axis,
+                                        output_names=None,
+                                        label_names=None)
+        #self.losses = []
+        self.count = 0
+
+    def cal_nme(self, label, pred_label):
+        nme = []
+        for b in range(pred_label.shape[0]):
+            record = [None] * 6
+            item = []
+            if label.ndim == 4:
+                _heatmap = label[b][36]
+                if np.count_nonzero(_heatmap) == 0:
+                    continue
+            else:  #ndim==3
+                #print(label[b])
+                if np.count_nonzero(label[b]) == 0:
+                    continue
+            for p in range(pred_label.shape[1]):
+                if label.ndim == 4:
+                    heatmap_gt = label[b][p]
+                    ind_gt = np.unravel_index(np.argmax(heatmap_gt, axis=None),
+                                              heatmap_gt.shape)
+                    ind_gt = np.array(ind_gt)
+                else:
+                    ind_gt = label[b][p]
+                    #ind_gt = ind_gt.astype(np.int)
+                    #print(ind_gt)
+                heatmap_pred = pred_label[b][p]
+                heatmap_pred = cv2.resize(
+                    heatmap_pred,
+                    (config.input_img_size, config.input_img_size))
+                ind_pred = np.unravel_index(np.argmax(heatmap_pred, axis=None),
+                                            heatmap_pred.shape)
+                ind_pred = np.array(ind_pred)
+                #print(ind_gt.shape)
+                #print(ind_pred)
+                if p == 36:
+                    #print('b', b, p, ind_gt, np.count_nonzero(heatmap_gt))
+                    record[0] = ind_gt
+                elif p == 39:
+                    record[1] = ind_gt
+                elif p == 42:
+                    record[2] = ind_gt
+                elif p == 45:
+                    record[3] = ind_gt
+                if record[4] is None or record[5] is None:
+                    record[4] = ind_gt
+                    record[5] = ind_gt
+                else:
+                    record[4] = np.minimum(record[4], ind_gt)
+                    record[5] = np.maximum(record[5], ind_gt)
+                #print(ind_gt.shape, ind_pred.shape)
+                value = np.sqrt(np.sum(np.square(ind_gt - ind_pred)))
+                item.append(value)
+            _nme = np.mean(item)
+            if config.landmark_type == '2d':
+                left_eye = (record[0] + record[1]) / 2
+                right_eye = (record[2] + record[3]) / 2
+                _dist = np.sqrt(np.sum(np.square(left_eye - right_eye)))
+                #print('eye dist', _dist, left_eye, right_eye)
+                _nme /= _dist
+            else:
+                #_dist = np.sqrt(float(label.shape[2]*label.shape[3]))
+                _dist = np.sqrt(np.sum(np.square(record[5] - record[4])))
+                #print(_dist)
+                _nme /= _dist
+            nme.append(_nme)
+        return np.mean(nme)
+
+    def update(self, labels, preds):
+        self.count += 1
+        label = labels[0].asnumpy()
+        pred_label = preds[-1].asnumpy()
+        nme = self.cal_nme(label, pred_label)
+
+        #print('nme', nme)
+        #nme = np.mean(nme)
+        self.sum_metric += np.mean(nme)
+        self.num_inst += 1.0
diff --git a/alignment/heatmap/optimizer.py b/alignment/heatmap/optimizer.py
new file mode 100644
index 000000000..36290de84
--- /dev/null
+++ b/alignment/heatmap/optimizer.py
@@ -0,0 +1,65 @@
+import mxnet as mx
+import mxnet.optimizer as optimizer
+from mxnet.ndarray import (NDArray, zeros, clip, sqrt, cast, maximum, abs as
+                           NDabs)
+#from mxnet.ndarray import (sgd_update, sgd_mom_update, adam_update, rmsprop_update, rmspropalex_update,
+#                      mp_sgd_update, mp_sgd_mom_update, square, ftrl_update)
+
+
+class ONadam(optimizer.Optimizer):
+    def __init__(self,
+                 learning_rate=0.001,
+                 beta1=0.9,
+                 beta2=0.999,
+                 epsilon=1e-8,
+                 schedule_decay=0.004,
+                 **kwargs):
+        super(ONadam, self).__init__(learning_rate=learning_rate, **kwargs)
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+        self.schedule_decay = schedule_decay
+        self.m_schedule = 1.
+
+    def create_state(self, index, weight):
+        return (
+            zeros(weight.shape, weight.context, dtype=weight.dtype),  # mean
+            zeros(weight.shape, weight.context,
+                  dtype=weight.dtype))  # variance
+
+    def update(self, index, weight, grad, state):
+        assert (isinstance(weight, NDArray))
+        assert (isinstance(grad, NDArray))
+        self._update_count(index)
+        lr = self._get_lr(index)
+        wd = self._get_wd(index)
+
+        t = self._index_update_count[index]
+
+        # preprocess grad
+        #grad = grad * self.rescale_grad + wd * weight
+        grad *= self.rescale_grad + wd * weight
+        if self.clip_gradient is not None:
+            grad = clip(grad, -self.clip_gradient, self.clip_gradient)
+
+        # warming momentum schedule
+        momentum_t = self.beta1 * (1. - 0.5 *
+                                   (pow(0.96, t * self.schedule_decay)))
+        momentum_t_1 = self.beta1 * (1. - 0.5 *
+                                     (pow(0.96,
+                                          (t + 1) * self.schedule_decay)))
+        self.m_schedule = self.m_schedule * momentum_t
+        m_schedule_next = self.m_schedule * momentum_t_1
+
+        # update m_t and v_t
+        m_t, v_t = state
+        m_t[:] = self.beta1 * m_t + (1. - self.beta1) * grad
+        v_t[:] = self.beta2 * v_t + (1. - self.beta2) * grad * grad
+
+        grad_prime = grad / (1. - self.m_schedule)
+        m_t_prime = m_t / (1. - m_schedule_next)
+        v_t_prime = v_t / (1. - pow(self.beta2, t))
+        m_t_bar = (1. - momentum_t) * grad_prime + momentum_t_1 * m_t_prime
+
+        # update weight
+        weight[:] -= lr * m_t_bar / (sqrt(v_t_prime) + self.epsilon)
diff --git a/alignment/heatmap/sample_config.py b/alignment/heatmap/sample_config.py
new file mode 100644
index 000000000..bb9ada185
--- /dev/null
+++ b/alignment/heatmap/sample_config.py
@@ -0,0 +1,98 @@
+import numpy as np
+from easydict import EasyDict as edict
+
+config = edict()
+
+#default training/dataset config
+config.num_classes = 68
+config.record_img_size = 384
+config.base_scale = 256
+config.input_img_size = 128
+config.output_label_size = 64
+config.label_xfirst = False
+config.losstype = 'heatmap'
+config.net_coherent = False
+config.multiplier = 1.0
+
+config.gaussian = 0
+
+# network settings
+network = edict()
+
+network.hourglass = edict()
+network.hourglass.net_coherent = False
+network.hourglass.net_sta = 0
+network.hourglass.net_n = 3
+network.hourglass.net_dcn = 0
+network.hourglass.net_stacks = 2
+network.hourglass.net_block = 'resnet'
+network.hourglass.net_binarize = False
+network.hourglass.losstype = 'heatmap'
+
+network.sdu = edict()
+network.sdu.net_coherent = False
+network.sdu.net_sta = 1
+network.sdu.net_n = 3
+network.sdu.net_dcn = 3
+network.sdu.net_stacks = 2
+network.sdu.net_block = 'cab'
+network.sdu.net_binarize = False
+network.sdu.losstype = 'heatmap'
+
+# dataset settings
+dataset = edict()
+
+dataset.i2d = edict()
+dataset.i2d.dataset = '2D'
+dataset.i2d.landmark_type = '2d'
+dataset.i2d.dataset_path = './data_2d'
+dataset.i2d.num_classes = 68
+dataset.i2d.record_img_size = 384
+dataset.i2d.base_scale = 256
+dataset.i2d.input_img_size = 128
+dataset.i2d.output_label_size = 64
+dataset.i2d.label_xfirst = False
+dataset.i2d.val_targets = ['ibug', 'cofw_testset', '300W']
+
+dataset.i3d = edict()
+dataset.i3d.dataset = '3D'
+dataset.i3d.landmark_type = '3d'
+dataset.i3d.dataset_path = './data_3d'
+dataset.i3d.num_classes = 68
+dataset.i3d.record_img_size = 384
+dataset.i3d.base_scale = 256
+dataset.i3d.input_img_size = 128
+dataset.i3d.output_label_size = 64
+dataset.i3d.label_xfirst = False
+dataset.i3d.val_targets = ['AFLW2000-3D']
+
+# default settings
+default = edict()
+
+# default network
+default.network = 'hourglass'
+default.pretrained = ''
+default.pretrained_epoch = 0
+# default dataset
+default.dataset = 'i2d'
+default.frequent = 20
+default.verbose = 200
+default.kvstore = 'device'
+
+default.prefix = 'model/A'
+default.end_epoch = 10000
+default.lr = 0.00025
+default.wd = 0.0
+default.per_batch_size = 20
+default.lr_step = '16000,24000,30000'
+
+
+def generate_config(_network, _dataset):
+    for k, v in network[_network].items():
+        config[k] = v
+        default[k] = v
+    for k, v in dataset[_dataset].items():
+        config[k] = v
+        default[k] = v
+    config.network = _network
+    config.dataset = _dataset
diff --git a/alignment/heatmap/symbol/sym_heatmap.py b/alignment/heatmap/symbol/sym_heatmap.py
new file mode 100644
index 000000000..338411292
--- /dev/null
+++ b/alignment/heatmap/symbol/sym_heatmap.py
@@ -0,0 +1,1085 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import mxnet as mx
+import numpy as np
+from config import config
+
+ACT_BIT = 1
+bn_mom = 0.9
+workspace = 256
+memonger = False
+
+
+def Conv(**kwargs):
+    body = mx.sym.Convolution(**kwargs)
+    return body
+
+
+def Act(data, act_type, name):
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+#def lin(data, num_filter, workspace, name, binarize, dcn):
+#  bit = 1
+#  if not binarize:
+#    if not dcn:
+#        conv1 = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
+#                                      no_bias=True, workspace=workspace, name=name + '_conv')
+#        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
+#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
+#        return act1
+#    else:
+#        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
+#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
+#        conv1_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = act1,
+#                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
+#        conv1 = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=act1, offset=conv1_offset,
+#                num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=False)
+#        #conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
+#        #                              no_bias=False, workspace=workspace, name=name + '_conv')
+#        return conv1
+#  else:
+#    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
+#    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
+#    conv1 = mx.sym.QConvolution_v1(data=act1, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
+#                               no_bias=True, workspace=workspace, name=name + '_conv', act_bit=ACT_BIT, weight_bit=bit)
+#    conv1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
+#    return conv1
+
+
+def lin3(data, num_filter, workspace, name, k, g=1, d=1):
+    if k != 3:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(k, k),
+                     stride=(1, 1),
+                     pad=((k - 1) // 2, (k - 1) // 2),
+                     num_group=g,
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(k, k),
+                     stride=(1, 1),
+                     pad=(d, d),
+                     num_group=g,
+                     dilate=(d, d),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv')
+    bn1 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           momentum=bn_mom,
+                           eps=2e-5,
+                           name=name + '_bn')
+    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
+    ret = act1
+    return ret
+
+
+def ConvFactory(data,
+                num_filter,
+                kernel,
+                stride=(1, 1),
+                pad=(0, 0),
+                act_type="relu",
+                mirror_attr={},
+                with_act=True,
+                dcn=False,
+                name=''):
+    if not dcn:
+        conv = mx.symbol.Convolution(data=data,
+                                     num_filter=num_filter,
+                                     kernel=kernel,
+                                     stride=stride,
+                                     pad=pad,
+                                     no_bias=True,
+                                     workspace=workspace,
+                                     name=name + '_conv')
+    else:
+        conv_offset = mx.symbol.Convolution(name=name + '_conv_offset',
+                                            data=data,
+                                            num_filter=18,
+                                            pad=(1, 1),
+                                            kernel=(3, 3),
+                                            stride=(1, 1))
+        conv = mx.contrib.symbol.DeformableConvolution(name=name + "_conv",
+                                                       data=data,
+                                                       offset=conv_offset,
+                                                       num_filter=num_filter,
+                                                       pad=(1, 1),
+                                                       kernel=(3, 3),
+                                                       num_deformable_group=1,
+                                                       stride=stride,
+                                                       dilate=(1, 1),
+                                                       no_bias=False)
+    bn = mx.symbol.BatchNorm(data=conv,
+                             fix_gamma=False,
+                             momentum=bn_mom,
+                             eps=2e-5,
+                             name=name + '_bn')
+    if with_act:
+        act = Act(bn, act_type, name=name + '_relu')
+        #act = mx.symbol.Activation(
+        #    data=bn, act_type=act_type, attr=mirror_attr, name=name+'_relu')
+        return act
+    else:
+        return bn
+
+
+class CAB:
+    def __init__(self, data, nFilters, nModules, n, workspace, name, dilate,
+                 group):
+        self.data = data
+        self.nFilters = nFilters
+        self.nModules = nModules
+        self.n = n
+        self.workspace = workspace
+        self.name = name
+        self.dilate = dilate
+        self.group = group
+        self.sym_map = {}
+
+    def get_output(self, w, h):
+        key = (w, h)
+        if key in self.sym_map:
+            return self.sym_map[key]
+        ret = None
+        if h == self.n:
+            if w == self.n:
+                ret = (self.data, self.nFilters)
+            else:
+                x = self.get_output(w + 1, h)
+                f = int(x[1] * 0.5)
+                if w != self.n - 1:
+                    body = lin3(x[0], f, self.workspace,
+                                "%s_w%d_h%d_1" % (self.name, w, h), 3,
+                                self.group, 1)
+                else:
+                    body = lin3(x[0], f, self.workspace,
+                                "%s_w%d_h%d_1" % (self.name, w, h), 3,
+                                self.group, self.dilate)
+                ret = (body, f)
+        else:
+            x = self.get_output(w + 1, h + 1)
+            y = self.get_output(w, h + 1)
+            if h % 2 == 1 and h != w:
+                xbody = lin3(x[0], x[1], self.workspace,
+                             "%s_w%d_h%d_2" % (self.name, w, h), 3, x[1])
+                #xbody = xbody+x[0]
+            else:
+                xbody = x[0]
+            #xbody = x[0]
+            #xbody = lin3(x[0], x[1], self.workspace, "%s_w%d_h%d_2"%(self.name, w, h), 3, x[1])
+            if w == 0:
+                ybody = lin3(y[0], y[1], self.workspace,
+                             "%s_w%d_h%d_3" % (self.name, w, h), 3, self.group)
+            else:
+                ybody = y[0]
+            ybody = mx.sym.concat(y[0], ybody, dim=1)
+            body = mx.sym.add_n(xbody,
+                                ybody,
+                                name="%s_w%d_h%d_add" % (self.name, w, h))
+            body = body / 2
+            ret = (body, x[1])
+        self.sym_map[key] = ret
+        return ret
+
+    def get(self):
+        return self.get_output(1, 1)[0]
+
+
+def conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn,
+                dilate, **kwargs):
+    bit = 1
+    #print('in unit2')
+    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    if not binarize:
+        act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
+        conv1 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.5),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+    else:
+        act1 = mx.sym.QActivation(data=bn1,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu1',
+                                  backward_only=True)
+        conv1 = mx.sym.QConvolution(data=act1,
+                                    num_filter=int(num_filter * 0.5),
+                                    kernel=(1, 1),
+                                    stride=(1, 1),
+                                    pad=(0, 0),
+                                    no_bias=True,
+                                    workspace=workspace,
+                                    name=name + '_conv1',
+                                    act_bit=ACT_BIT,
+                                    weight_bit=bit)
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    if not binarize:
+        act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
+        conv2 = Conv(data=act2,
+                     num_filter=int(num_filter * 0.5),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+    else:
+        act2 = mx.sym.QActivation(data=bn2,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu2',
+                                  backward_only=True)
+        conv2 = mx.sym.QConvolution(data=act2,
+                                    num_filter=int(num_filter * 0.5),
+                                    kernel=(3, 3),
+                                    stride=(1, 1),
+                                    pad=(1, 1),
+                                    no_bias=True,
+                                    workspace=workspace,
+                                    name=name + '_conv2',
+                                    act_bit=ACT_BIT,
+                                    weight_bit=bit)
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+    if not binarize:
+        act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
+        conv3 = Conv(data=act3,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+    else:
+        act3 = mx.sym.QActivation(data=bn3,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu3',
+                                  backward_only=True)
+        conv3 = mx.sym.QConvolution(data=act3,
+                                    num_filter=num_filter,
+                                    kernel=(1, 1),
+                                    stride=(1, 1),
+                                    pad=(0, 0),
+                                    no_bias=True,
+                                    workspace=workspace,
+                                    name=name + '_conv3',
+                                    act_bit=ACT_BIT,
+                                    weight_bit=bit)
+    #if binarize:
+    #  conv3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
+    if dim_match:
+        shortcut = data
+    else:
+        if not binarize:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        else:
+            shortcut = mx.sym.QConvolution(data=act1,
+                                           num_filter=num_filter,
+                                           kernel=(1, 1),
+                                           stride=stride,
+                                           pad=(0, 0),
+                                           no_bias=True,
+                                           workspace=workspace,
+                                           name=name + '_sc',
+                                           act_bit=ACT_BIT,
+                                           weight_bit=bit)
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return conv3 + shortcut
+
+
+def conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn,
+             dilation, **kwargs):
+    bit = 1
+    #print('in unit2')
+    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    if not binarize:
+        act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
+        if not dcn:
+            conv1 = Conv(data=act1,
+                         num_filter=int(num_filter * 0.5),
+                         kernel=(3, 3),
+                         stride=(1, 1),
+                         pad=(dilation, dilation),
+                         dilate=(dilation, dilation),
+                         no_bias=True,
+                         workspace=workspace,
+                         name=name + '_conv1')
+        else:
+            conv1_offset = mx.symbol.Convolution(name=name + '_conv1_offset',
+                                                 data=act1,
+                                                 num_filter=18,
+                                                 pad=(1, 1),
+                                                 kernel=(3, 3),
+                                                 stride=(1, 1))
+            conv1 = mx.contrib.symbol.DeformableConvolution(
+                name=name + '_conv1',
+                data=act1,
+                offset=conv1_offset,
+                num_filter=int(num_filter * 0.5),
+                pad=(1, 1),
+                kernel=(3, 3),
+                num_deformable_group=1,
+                stride=(1, 1),
+                dilate=(1, 1),
+                no_bias=True)
+    else:
+        act1 = mx.sym.QActivation(data=bn1,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu1',
+                                  backward_only=True)
+        conv1 = mx.sym.QConvolution_v1(data=act1,
+                                       num_filter=int(num_filter * 0.5),
+                                       kernel=(3, 3),
+                                       stride=(1, 1),
+                                       pad=(1, 1),
+                                       no_bias=True,
+                                       workspace=workspace,
+                                       name=name + '_conv1',
+                                       act_bit=ACT_BIT,
+                                       weight_bit=bit)
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    if not binarize:
+        act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
+        if not dcn:
+            conv2 = Conv(data=act2,
+                         num_filter=int(num_filter * 0.25),
+                         kernel=(3, 3),
+                         stride=(1, 1),
+                         pad=(dilation, dilation),
+                         dilate=(dilation, dilation),
+                         no_bias=True,
+                         workspace=workspace,
+                         name=name + '_conv2')
+        else:
+            conv2_offset = mx.symbol.Convolution(name=name + '_conv2_offset',
+                                                 data=act2,
+                                                 num_filter=18,
+                                                 pad=(1, 1),
+                                                 kernel=(3, 3),
+                                                 stride=(1, 1))
+            conv2 = mx.contrib.symbol.DeformableConvolution(
+                name=name + '_conv2',
+                data=act2,
+                offset=conv2_offset,
+                num_filter=int(num_filter * 0.25),
+                pad=(1, 1),
+                kernel=(3, 3),
+                num_deformable_group=1,
+                stride=(1, 1),
+                dilate=(1, 1),
+                no_bias=True)
+    else:
+        act2 = mx.sym.QActivation(data=bn2,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu2',
+                                  backward_only=True)
+        conv2 = mx.sym.QConvolution_v1(data=act2,
+                                       num_filter=int(num_filter * 0.25),
+                                       kernel=(3, 3),
+                                       stride=(1, 1),
+                                       pad=(1, 1),
+                                       no_bias=True,
+                                       workspace=workspace,
+                                       name=name + '_conv2',
+                                       act_bit=ACT_BIT,
+                                       weight_bit=bit)
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+    if not binarize:
+        act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
+        if not dcn:
+            conv3 = Conv(data=act3,
+                         num_filter=int(num_filter * 0.25),
+                         kernel=(3, 3),
+                         stride=(1, 1),
+                         pad=(dilation, dilation),
+                         dilate=(dilation, dilation),
+                         no_bias=True,
+                         workspace=workspace,
+                         name=name + '_conv3')
+        else:
+            conv3_offset = mx.symbol.Convolution(name=name + '_conv3_offset',
+                                                 data=act3,
+                                                 num_filter=18,
+                                                 pad=(1, 1),
+                                                 kernel=(3, 3),
+                                                 stride=(1, 1))
+            conv3 = mx.contrib.symbol.DeformableConvolution(
+                name=name + '_conv3',
+                data=act3,
+                offset=conv3_offset,
+                num_filter=int(num_filter * 0.25),
+                pad=(1, 1),
+                kernel=(3, 3),
+                num_deformable_group=1,
+                stride=(1, 1),
+                dilate=(1, 1),
+                no_bias=True)
+    else:
+        act3 = mx.sym.QActivation(data=bn3,
+                                  act_bit=ACT_BIT,
+                                  name=name + '_relu3',
+                                  backward_only=True)
+        conv3 = mx.sym.QConvolution_v1(data=act3,
+                                       num_filter=int(num_filter * 0.25),
+                                       kernel=(3, 3),
+                                       stride=(1, 1),
+                                       pad=(1, 1),
+                                       no_bias=True,
+                                       workspace=workspace,
+                                       name=name + '_conv3',
+                                       act_bit=ACT_BIT,
+                                       weight_bit=bit)
+    conv4 = mx.symbol.Concat(*[conv1, conv2, conv3])
+    if binarize:
+        conv4 = mx.sym.BatchNorm(data=conv4,
+                                 fix_gamma=False,
+                                 eps=2e-5,
+                                 momentum=bn_mom,
+                                 name=name + '_bn4')
+    if dim_match:
+        shortcut = data
+    else:
+        if not binarize:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        else:
+            #assert(False)
+            shortcut = mx.sym.QConvolution_v1(data=act1,
+                                              num_filter=num_filter,
+                                              kernel=(1, 1),
+                                              stride=stride,
+                                              pad=(0, 0),
+                                              no_bias=True,
+                                              workspace=workspace,
+                                              name=name + '_sc',
+                                              act_bit=ACT_BIT,
+                                              weight_bit=bit)
+            shortcut = mx.sym.BatchNorm(data=shortcut,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc_bn')
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return conv4 + shortcut
+    #return bn4 + shortcut
+    #return act4 + shortcut
+
+
+def block17(net,
+            input_num_channels,
+            scale=1.0,
+            with_act=True,
+            act_type='relu',
+            mirror_attr={},
+            name=''):
+    tower_conv = ConvFactory(net, 192, (1, 1), name=name + '_conv')
+    tower_conv1_0 = ConvFactory(net, 129, (1, 1), name=name + '_conv1_0')
+    tower_conv1_1 = ConvFactory(tower_conv1_0,
+                                160, (1, 7),
+                                pad=(1, 2),
+                                name=name + '_conv1_1')
+    tower_conv1_2 = ConvFactory(tower_conv1_1,
+                                192, (7, 1),
+                                pad=(2, 1),
+                                name=name + '_conv1_2')
+    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_2])
+    tower_out = ConvFactory(tower_mixed,
+                            input_num_channels, (1, 1),
+                            with_act=False,
+                            name=name + '_conv_out')
+    net = net + scale * tower_out
+    if with_act:
+        act = mx.symbol.Activation(data=net,
+                                   act_type=act_type,
+                                   attr=mirror_attr)
+        return act
+    else:
+        return net
+
+
+def block35(net,
+            input_num_channels,
+            scale=1.0,
+            with_act=True,
+            act_type='relu',
+            mirror_attr={},
+            name=''):
+    M = 1.0
+    tower_conv = ConvFactory(net,
+                             int(input_num_channels * 0.25 * M), (1, 1),
+                             name=name + '_conv')
+    tower_conv1_0 = ConvFactory(net,
+                                int(input_num_channels * 0.25 * M), (1, 1),
+                                name=name + '_conv1_0')
+    tower_conv1_1 = ConvFactory(tower_conv1_0,
+                                int(input_num_channels * 0.25 * M), (3, 3),
+                                pad=(1, 1),
+                                name=name + '_conv1_1')
+    tower_conv2_0 = ConvFactory(net,
+                                int(input_num_channels * 0.25 * M), (1, 1),
+                                name=name + '_conv2_0')
+    tower_conv2_1 = ConvFactory(tower_conv2_0,
+                                int(input_num_channels * 0.375 * M), (3, 3),
+                                pad=(1, 1),
+                                name=name + '_conv2_1')
+    tower_conv2_2 = ConvFactory(tower_conv2_1,
+                                int(input_num_channels * 0.5 * M), (3, 3),
+                                pad=(1, 1),
+                                name=name + '_conv2_2')
+    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])
+    tower_out = ConvFactory(tower_mixed,
+                            input_num_channels, (1, 1),
+                            with_act=False,
+                            name=name + '_conv_out')
+
+    net = net + scale * tower_out
+    if with_act:
+        act = mx.symbol.Activation(data=net,
+                                   act_type=act_type,
+                                   attr=mirror_attr)
+        return act
+    else:
+        return net
+
+
+def conv_inception(data, num_filter, stride, dim_match, name, binarize, dcn,
+                   dilate, **kwargs):
+    assert not binarize
+    if stride[0] > 1 or not dim_match:
+        return conv_resnet(data, num_filter, stride, dim_match, name, binarize,
+                           dcn, dilate, **kwargs)
+    conv4 = block35(data, num_filter, name=name + '_block35')
+    return conv4
+
+
+def conv_cab(data, num_filter, stride, dim_match, name, binarize, dcn, dilate,
+             **kwargs):
+    if stride[0] > 1 or not dim_match:
+        return conv_hpm(data, num_filter, stride, dim_match, name, binarize,
+                        dcn, dilate, **kwargs)
+    cab = CAB(data, num_filter, 1, 4, workspace, name, dilate, 1)
+    return cab.get()
+
+
+def conv_block(data, num_filter, stride, dim_match, name, binarize, dcn,
+               dilate):
+    if config.net_block == 'resnet':
+        return conv_resnet(data, num_filter, stride, dim_match, name, binarize,
+                           dcn, dilate)
+    elif config.net_block == 'inception':
+        return conv_inception(data, num_filter, stride, dim_match, name,
+                              binarize, dcn, dilate)
+    elif config.net_block == 'hpm':
+        return conv_hpm(data, num_filter, stride, dim_match, name, binarize,
+                        dcn, dilate)
+    elif config.net_block == 'cab':
+        return conv_cab(data, num_filter, stride, dim_match, name, binarize,
+                        dcn, dilate)
+
+
+def hourglass(data, nFilters, nModules, n, workspace, name, binarize, dcn):
+    s = 2
+    _dcn = False
+    up1 = data
+    for i in range(nModules):
+        up1 = conv_block(up1, nFilters, (1, 1), True, "%s_up1_%d" % (name, i),
+                         binarize, _dcn, 1)
+    low1 = mx.sym.Pooling(data=data,
+                          kernel=(s, s),
+                          stride=(s, s),
+                          pad=(0, 0),
+                          pool_type='max')
+    for i in range(nModules):
+        low1 = conv_block(low1, nFilters, (1, 1), True,
+                          "%s_low1_%d" % (name, i), binarize, _dcn, 1)
+    if n > 1:
+        low2 = hourglass(low1, nFilters, nModules, n - 1, workspace,
+                         "%s_%d" % (name, n - 1), binarize, dcn)
+    else:
+        low2 = low1
+        for i in range(nModules):
+            low2 = conv_block(low2, nFilters, (1, 1), True,
+                              "%s_low2_%d" % (name, i), binarize, _dcn,
+                              1)  #TODO
+    low3 = low2
+    for i in range(nModules):
+        low3 = conv_block(low3, nFilters, (1, 1), True,
+                          "%s_low3_%d" % (name, i), binarize, _dcn, 1)
+    up2 = mx.symbol.UpSampling(low3,
+                               scale=s,
+                               sample_type='nearest',
+                               workspace=512,
+                               name='%s_upsampling_%s' % (name, n),
+                               num_args=1)
+    return mx.symbol.add_n(up1, up2)
+
+
+class STA:
+    def __init__(self, data, nFilters, nModules, n, workspace, name):
+        self.data = data
+        self.nFilters = nFilters
+        self.nModules = nModules
+        self.n = n
+        self.workspace = workspace
+        self.name = name
+        self.sym_map = {}
+
+    def get_conv(self, data, name, dilate=1, group=1):
+        cab = CAB(data, self.nFilters, self.nModules, 4, self.workspace, name,
+                  dilate, group)
+        return cab.get()
+
+    def get_output(self, w, h):
+        #print(w,h)
+        assert w >= 1 and w <= config.net_n + 1
+        assert h >= 1 and h <= config.net_n + 1
+        s = 2
+        bn_mom = 0.9
+        key = (w, h)
+        if key in self.sym_map:
+            return self.sym_map[key]
+        ret = None
+        if h == self.n:
+            if w == self.n:
+                ret = self.data, 64
+            else:
+                x = self.get_output(w + 1, h)
+                body = self.get_conv(x[0], "%s_w%d_h%d_1" % (self.name, w, h))
+                body = mx.sym.Pooling(data=body,
+                                      kernel=(s, s),
+                                      stride=(s, s),
+                                      pad=(0, 0),
+                                      pool_type='max')
+                body = self.get_conv(body, "%s_w%d_h%d_2" % (self.name, w, h))
+                ret = body, x[1] // 2
+        else:
+            x = self.get_output(w + 1, h + 1)
+            y = self.get_output(w, h + 1)
+
+            HC = False
+
+            if h % 2 == 1 and h != w:
+                xbody = lin3(x[0], self.nFilters, self.workspace,
+                             "%s_w%d_h%d_x" % (self.name, w, h), 3,
+                             self.nFilters, 1)
+                HC = True
+                #xbody = x[0]
+            else:
+                xbody = x[0]
+            if x[1] // y[1] == 2:
+                if w > 1:
+                    ybody = mx.symbol.Deconvolution(
+                        data=y[0],
+                        num_filter=self.nFilters,
+                        kernel=(s, s),
+                        stride=(s, s),
+                        name='%s_upsampling_w%d_h%d' % (self.name, w, h),
+                        attr={'lr_mult': '1.0'},
+                        workspace=self.workspace)
+                    ybody = mx.sym.BatchNorm(data=ybody,
+                                             fix_gamma=False,
+                                             momentum=bn_mom,
+                                             eps=2e-5,
+                                             name="%s_w%d_h%d_y_bn" %
+                                             (self.name, w, h))
+                    ybody = Act(data=ybody,
+                                act_type='relu',
+                                name="%s_w%d_h%d_y_act" % (self.name, w, h))
+                else:
+                    if h >= 1:
+                        ybody = mx.symbol.UpSampling(
+                            y[0],
+                            scale=s,
+                            sample_type='nearest',
+                            workspace=512,
+                            name='%s_upsampling_w%d_h%d' % (self.name, w, h),
+                            num_args=1)
+                        ybody = self.get_conv(
+                            ybody, "%s_w%d_h%d_4" % (self.name, w, h))
+                    else:
+                        ybody = mx.symbol.Deconvolution(
+                            data=y[0],
+                            num_filter=self.nFilters,
+                            kernel=(s, s),
+                            stride=(s, s),
+                            name='%s_upsampling_w%d_h%d' % (self.name, w, h),
+                            attr={'lr_mult': '1.0'},
+                            workspace=self.workspace)
+                        ybody = mx.sym.BatchNorm(data=ybody,
+                                                 fix_gamma=False,
+                                                 momentum=bn_mom,
+                                                 eps=2e-5,
+                                                 name="%s_w%d_h%d_y_bn" %
+                                                 (self.name, w, h))
+                        ybody = Act(data=ybody,
+                                    act_type='relu',
+                                    name="%s_w%d_h%d_y_act" %
+                                    (self.name, w, h))
+                        ybody = Conv(data=ybody,
+                                     num_filter=self.nFilters,
+                                     kernel=(3, 3),
+                                     stride=(1, 1),
+                                     pad=(1, 1),
+                                     no_bias=True,
+                                     name="%s_w%d_h%d_y_conv2" %
+                                     (self.name, w, h),
+                                     workspace=self.workspace)
+                        ybody = mx.sym.BatchNorm(data=ybody,
+                                                 fix_gamma=False,
+                                                 momentum=bn_mom,
+                                                 eps=2e-5,
+                                                 name="%s_w%d_h%d_y_bn2" %
+                                                 (self.name, w, h))
+                        ybody = Act(data=ybody,
+                                    act_type='relu',
+                                    name="%s_w%d_h%d_y_act2" %
+                                    (self.name, w, h))
+            else:
+                ybody = self.get_conv(y[0], "%s_w%d_h%d_5" % (self.name, w, h))
+            #if not HC:
+            if config.net_sta == 2 and h == 3 and w == 2:
+                z = self.get_output(w + 1, h)
+                zbody = z[0]
+                zbody = mx.sym.Pooling(data=zbody,
+                                       kernel=(z[1], z[1]),
+                                       stride=(z[1], z[1]),
+                                       pad=(0, 0),
+                                       pool_type='avg')
+                body = xbody + ybody
+                body = body / 2
+                body = mx.sym.broadcast_mul(body, zbody)
+            else:  #sta==1
+                body = xbody + ybody
+                body = body / 2
+            ret = body, x[1]
+
+        assert ret is not None
+        self.sym_map[key] = ret
+        return ret
+
+    def get(self):
+        return self.get_output(1, 1)[0]
+
+
+class SymCoherent:
+    def __init__(self, per_batch_size):
+        self.per_batch_size = per_batch_size
+        self.flip_order = [
+            16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 26, 25,
+            24, 23, 22, 21, 20, 19, 18, 17, 27, 28, 29, 30, 35, 34, 33, 32, 31,
+            45, 44, 43, 42, 47, 46, 39, 38, 37, 36, 41, 40, 54, 53, 52, 51, 50,
+            49, 48, 59, 58, 57, 56, 55, 64, 63, 62, 61, 60, 67, 66, 65
+        ]
+
+    def get(self, data):
+        #data.shape[0]==per_batch_size
+        b = self.per_batch_size // 2
+        ux = mx.sym.slice_axis(data, axis=0, begin=0, end=b)
+        dx = mx.sym.slice_axis(data, axis=0, begin=b, end=b * 2)
+        ux = mx.sym.flip(ux, axis=3)
+        #ux = mx.sym.take(ux, indices = self.flip_order, axis=0)
+        ux_list = []
+        for o in self.flip_order:
+            _ux = mx.sym.slice_axis(ux, axis=1, begin=o, end=o + 1)
+            ux_list.append(_ux)
+        ux = mx.sym.concat(*ux_list, dim=1)
+        return ux, dx
+
+
+def l2_loss(x, y):
+    loss = x - y
+    loss = mx.symbol.smooth_l1(loss, scalar=1.0)
+    #loss = loss*loss
+    loss = mx.symbol.mean(loss)
+    return loss
+
+
+def ce_loss(x, y):
+    #loss = mx.sym.SoftmaxOutput(data = x, label = y, normalization='valid', multi_output=True)
+    x_max = mx.sym.max(x, axis=[2, 3], keepdims=True)
+    x = mx.sym.broadcast_minus(x, x_max)
+    body = mx.sym.exp(x)
+    sums = mx.sym.sum(body, axis=[2, 3], keepdims=True)
+    body = mx.sym.broadcast_div(body, sums)
+    loss = mx.sym.log(body)
+    loss = loss * y * -1.0
+    loss = mx.symbol.mean(loss, axis=[1, 2, 3])
+    #loss = mx.symbol.mean(loss)
+    return loss
+
+
+def get_symbol(num_classes):
+    m = config.multiplier
+    sFilters = max(int(64 * m), 32)
+    mFilters = max(int(128 * m), 32)
+    nFilters = int(256 * m)
+
+    nModules = 1
+    nStacks = config.net_stacks
+    binarize = config.net_binarize
+    input_size = config.input_img_size
+    label_size = config.output_label_size
+    use_coherent = config.net_coherent
+    use_STA = config.net_sta
+    N = config.net_n
+    DCN = config.net_dcn
+    per_batch_size = config.per_batch_size
+    print('binarize', binarize)
+    print('use_coherent', use_coherent)
+    print('use_STA', use_STA)
+    print('use_N', N)
+    print('use_DCN', DCN)
+    print('per_batch_size', per_batch_size)
+    #assert(label_size==64 or label_size==32)
+    #assert(input_size==128 or input_size==256)
+    coherentor = SymCoherent(per_batch_size)
+    D = input_size // label_size
+    print(input_size, label_size, D)
+    data = mx.sym.Variable(name='data')
+    data = data - 127.5
+    data = data * 0.0078125
+    gt_label = mx.symbol.Variable(name='softmax_label')
+    losses = []
+    closses = []
+    ref_label = gt_label
+    if D == 4:
+        body = Conv(data=data,
+                    num_filter=sFilters,
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+    else:
+        body = Conv(data=data,
+                    num_filter=sFilters,
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+    body = mx.sym.BatchNorm(data=body,
+                            fix_gamma=False,
+                            eps=2e-5,
+                            momentum=bn_mom,
+                            name='bn0')
+    body = Act(data=body, act_type='relu', name='relu0')
+
+    dcn = False
+    body = conv_block(body, mFilters, (1, 1), sFilters == mFilters, 'res0',
+                      False, dcn, 1)
+
+    body = mx.sym.Pooling(data=body,
+                          kernel=(2, 2),
+                          stride=(2, 2),
+                          pad=(0, 0),
+                          pool_type='max')
+
+    body = conv_block(body, mFilters, (1, 1), True, 'res1', False, dcn,
+                      1)  #TODO
+    body = conv_block(body, nFilters, (1, 1), mFilters == nFilters, 'res2',
+                      binarize, dcn, 1)  #binarize=True?
+
+    heatmap = None
+
+    for i in range(nStacks):
+        shortcut = body
+        if config.net_sta > 0:
+            sta = STA(body, nFilters, nModules, config.net_n + 1, workspace,
+                      'sta%d' % (i))
+            body = sta.get()
+        else:
+            body = hourglass(body, nFilters, nModules, config.net_n, workspace,
+                             'stack%d_hg' % (i), binarize, dcn)
+        for j in range(nModules):
+            body = conv_block(body, nFilters, (1, 1), True,
+                              'stack%d_unit%d' % (i, j), binarize, dcn, 1)
+        _dcn = True if config.net_dcn >= 2 else False
+        ll = ConvFactory(body,
+                         nFilters, (1, 1),
+                         dcn=_dcn,
+                         name='stack%d_ll' % (i))
+        _name = "heatmap%d" % (i) if i < nStacks - 1 else "heatmap"
+        _dcn = True if config.net_dcn >= 2 else False
+        if not _dcn:
+            out = Conv(data=ll,
+                       num_filter=num_classes,
+                       kernel=(1, 1),
+                       stride=(1, 1),
+                       pad=(0, 0),
+                       name=_name,
+                       workspace=workspace)
+        else:
+            out_offset = mx.symbol.Convolution(name=_name + '_offset',
+                                               data=ll,
+                                               num_filter=18,
+                                               pad=(1, 1),
+                                               kernel=(3, 3),
+                                               stride=(1, 1))
+            out = mx.contrib.symbol.DeformableConvolution(
+                name=_name,
+                data=ll,
+                offset=out_offset,
+                num_filter=num_classes,
+                pad=(1, 1),
+                kernel=(3, 3),
+                num_deformable_group=1,
+                stride=(1, 1),
+                dilate=(1, 1),
+                no_bias=False)
+            #out = Conv(data=ll, num_filter=num_classes, kernel=(3,3), stride=(1,1), pad=(1,1),
+            #                          name=_name, workspace=workspace)
+        if i == nStacks - 1:
+            heatmap = out
+        loss = ce_loss(out, ref_label)
+        #loss = loss/nStacks
+        #loss = l2_loss(out, ref_label)
+        losses.append(loss)
+        if config.net_coherent > 0:
+            ux, dx = coherentor.get(out)
+            closs = l2_loss(ux, dx)
+            closs = closs / nStacks
+            closses.append(closs)
+
+        if i < nStacks - 1:
+            ll2 = Conv(data=ll,
+                       num_filter=nFilters,
+                       kernel=(1, 1),
+                       stride=(1, 1),
+                       pad=(0, 0),
+                       name="stack%d_ll2" % (i),
+                       workspace=workspace)
+            out2 = Conv(data=out,
+                        num_filter=nFilters,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name="stack%d_out2" % (i),
+                        workspace=workspace)
+            body = mx.symbol.add_n(shortcut, ll2, out2)
+            _dcn = True if (config.net_dcn == 1
+                            or config.net_dcn == 3) else False
+            if _dcn:
+                _name = "stack%d_out3" % (i)
+                out3_offset = mx.symbol.Convolution(name=_name + '_offset',
+                                                    data=body,
+                                                    num_filter=18,
+                                                    pad=(1, 1),
+                                                    kernel=(3, 3),
+                                                    stride=(1, 1))
+                out3 = mx.contrib.symbol.DeformableConvolution(
+                    name=_name,
+                    data=body,
+                    offset=out3_offset,
+                    num_filter=nFilters,
+                    pad=(1, 1),
+                    kernel=(3, 3),
+                    num_deformable_group=1,
+                    stride=(1, 1),
+                    dilate=(1, 1),
+                    no_bias=False)
+                body = out3
+
+    pred = mx.symbol.BlockGrad(heatmap)
+    #loss = mx.symbol.add_n(*losses)
+    #loss = mx.symbol.MakeLoss(loss)
+    #syms = [loss]
+    syms = []
+    for loss in losses:
+        loss = mx.symbol.MakeLoss(loss)
+        syms.append(loss)
+    if len(closses) > 0:
+        coherent_weight = 0.0001
+        closs = mx.symbol.add_n(*closses)
+        closs = mx.symbol.MakeLoss(closs, grad_scale=coherent_weight)
+        syms.append(closs)
+    syms.append(pred)
+    sym = mx.symbol.Group(syms)
+    return sym
+
+
+def init_weights(sym, data_shape_dict):
+    #print('in hg')
+    arg_name = sym.list_arguments()
+    aux_name = sym.list_auxiliary_states()
+    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(arg_name, arg_shape))
+    aux_shape_dict = dict(zip(aux_name, aux_shape))
+    #print(aux_shape)
+    #print(aux_params)
+    #print(arg_shape_dict)
+    arg_params = {}
+    aux_params = {}
+    for k in arg_shape_dict:
+        v = arg_shape_dict[k]
+        #print(k,v)
+        if k.endswith('offset_weight') or k.endswith('offset_bias'):
+            print('initializing', k)
+            arg_params[k] = mx.nd.zeros(shape=v)
+        elif k.startswith('fc6_'):
+            if k.endswith('_weight'):
+                print('initializing', k)
+                arg_params[k] = mx.random.normal(0, 0.01, shape=v)
+            elif k.endswith('_bias'):
+                print('initializing', k)
+                arg_params[k] = mx.nd.zeros(shape=v)
+        elif k.find('upsampling') >= 0:
+            print('initializing upsampling_weight', k)
+            arg_params[k] = mx.nd.zeros(shape=arg_shape_dict[k])
+            init = mx.init.Initializer()
+            init._init_bilinear(k, arg_params[k])
+    return arg_params, aux_params
diff --git a/alignment/heatmap/test.py b/alignment/heatmap/test.py
new file mode 100644
index 000000000..160e5acaf
--- /dev/null
+++ b/alignment/heatmap/test.py
@@ -0,0 +1,100 @@
+import argparse
+import cv2
+import sys
+import numpy as np
+import os
+import mxnet as mx
+import datetime
+import img_helper
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'deploy'))
+from mtcnn_detector import MtcnnDetector
+
+
+class Handler:
+    def __init__(self, prefix, epoch, ctx_id=0):
+        print('loading', prefix, epoch)
+        if ctx_id >= 0:
+            ctx = mx.gpu(ctx_id)
+        else:
+            ctx = mx.cpu()
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        all_layers = sym.get_internals()
+        sym = all_layers['heatmap_output']
+        image_size = (128, 128)
+        self.image_size = image_size
+        model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+        #model = mx.mod.Module(symbol=sym, context=ctx)
+        model.bind(for_training=False,
+                   data_shapes=[('data', (1, 3, image_size[0], image_size[1]))
+                                ])
+        model.set_params(arg_params, aux_params)
+        self.model = model
+        mtcnn_path = os.path.join(os.path.dirname(__file__), '..', 'deploy',
+                                  'mtcnn-model')
+        self.det_threshold = [0.6, 0.7, 0.8]
+        self.detector = MtcnnDetector(model_folder=mtcnn_path,
+                                      ctx=ctx,
+                                      num_worker=1,
+                                      accurate_landmark=True,
+                                      threshold=self.det_threshold)
+
+    def get(self, img):
+        ret = self.detector.detect_face(img, det_type=0)
+        if ret is None:
+            return None
+        bbox, points = ret
+        if bbox.shape[0] == 0:
+            return None
+        bbox = bbox[0, 0:4]
+        points = points[0, :].reshape((2, 5)).T
+        M = img_helper.estimate_trans_bbox(bbox, self.image_size[0], s=2.0)
+        rimg = cv2.warpAffine(img, M, self.image_size, borderValue=0.0)
+        img = cv2.cvtColor(rimg, cv2.COLOR_BGR2RGB)
+        img = np.transpose(img, (2, 0, 1))  #3*112*112, RGB
+        input_blob = np.zeros((1, 3, self.image_size[1], self.image_size[0]),
+                              dtype=np.uint8)
+        input_blob[0] = img
+        ta = datetime.datetime.now()
+        data = mx.nd.array(input_blob)
+        db = mx.io.DataBatch(data=(data, ))
+        self.model.forward(db, is_train=False)
+        alabel = self.model.get_outputs()[-1].asnumpy()[0]
+        tb = datetime.datetime.now()
+        print('module time cost', (tb - ta).total_seconds())
+        ret = np.zeros((alabel.shape[0], 2), dtype=np.float32)
+        for i in range(alabel.shape[0]):
+            a = cv2.resize(alabel[i], (self.image_size[1], self.image_size[0]))
+            ind = np.unravel_index(np.argmax(a, axis=None), a.shape)
+            #ret[i] = (ind[0], ind[1]) #h, w
+            ret[i] = (ind[1], ind[0])  #w, h
+        return ret, M
+
+
+ctx_id = 4
+img_path = '../deploy/Tom_Hanks_54745.png'
+img = cv2.imread(img_path)
+#img = np.zeros( (128,128,3), dtype=np.uint8 )
+
+handler = Handler('./model/HG', 1, ctx_id)
+for _ in range(10):
+    ta = datetime.datetime.now()
+    landmark, M = handler.get(img)
+    tb = datetime.datetime.now()
+    print('get time cost', (tb - ta).total_seconds())
+#visualize landmark
+IM = cv2.invertAffineTransform(M)
+for i in range(landmark.shape[0]):
+    p = landmark[i]
+    point = np.ones((3, ), dtype=np.float32)
+    point[0:2] = p
+    point = np.dot(IM, point)
+    landmark[i] = point[0:2]
+
+for i in range(landmark.shape[0]):
+    p = landmark[i]
+    point = (int(p[0]), int(p[1]))
+    cv2.circle(img, point, 1, (0, 255, 0), 2)
+
+filename = './landmark_test.png'
+print('writing', filename)
+cv2.imwrite(filename, img)
diff --git a/alignment/heatmap/test_rec_nme.py b/alignment/heatmap/test_rec_nme.py
new file mode 100644
index 000000000..baad7c7ba
--- /dev/null
+++ b/alignment/heatmap/test_rec_nme.py
@@ -0,0 +1,71 @@
+import argparse
+import cv2
+import sys
+import numpy as np
+import os
+import mxnet as mx
+import datetime
+import img_helper
+from config import config
+from data import FaceSegIter
+from metric import LossValueMetric, NMEMetric
+
+parser = argparse.ArgumentParser(description='test nme on rec data')
+# general
+parser.add_argument('--rec',
+                    default='./data_2d/ibug.rec',
+                    help='rec data path')
+parser.add_argument('--prefix', default='', help='model prefix')
+parser.add_argument('--epoch', type=int, default=1, help='model epoch')
+parser.add_argument('--gpu', type=int, default=0, help='')
+parser.add_argument('--landmark-type', default='2d', help='')
+parser.add_argument('--image-size', type=int, default=128, help='')
+args = parser.parse_args()
+
+rec_path = args.rec
+ctx_id = args.gpu
+prefix = args.prefix
+epoch = args.epoch
+image_size = (args.image_size, args.image_size)
+config.landmark_type = args.landmark_type
+config.input_img_size = image_size[0]
+
+if ctx_id >= 0:
+    ctx = mx.gpu(ctx_id)
+else:
+    ctx = mx.cpu()
+sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+all_layers = sym.get_internals()
+sym = all_layers['heatmap_output']
+#model = mx.mod.Module(symbol=sym, context=ctx, data_names=['data'], label_names=['softmax_label'])
+model = mx.mod.Module(symbol=sym,
+                      context=ctx,
+                      data_names=['data'],
+                      label_names=None)
+#model = mx.mod.Module(symbol=sym, context=ctx)
+model.bind(for_training=False,
+           data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
+model.set_params(arg_params, aux_params)
+
+val_iter = FaceSegIter(
+    path_imgrec=rec_path,
+    batch_size=1,
+    aug_level=0,
+)
+_metric = NMEMetric()
+#val_metric = mx.metric.create(_metric)
+#val_metric.reset()
+#val_iter.reset()
+nme = []
+for i, eval_batch in enumerate(val_iter):
+    if i % 10 == 0:
+        print('processing', i)
+    #print(eval_batch.data[0].shape, eval_batch.label[0].shape)
+    batch_data = mx.io.DataBatch(eval_batch.data)
+    model.forward(batch_data, is_train=False)
+    #model.update_metric(val_metric, eval_batch.label, True)
+    pred_label = model.get_outputs()[-1].asnumpy()
+    label = eval_batch.label[0].asnumpy()
+    _nme = _metric.cal_nme(label, pred_label)
+    nme.append(_nme)
+print(np.mean(nme))
diff --git a/alignment/heatmap/train.py b/alignment/heatmap/train.py
new file mode 100644
index 000000000..81412e47f
--- /dev/null
+++ b/alignment/heatmap/train.py
@@ -0,0 +1,236 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import logging
+import argparse
+from data import FaceSegIter
+import mxnet as mx
+import mxnet.optimizer as optimizer
+import numpy as np
+import os
+import sys
+import math
+import random
+import cv2
+from config import config, default, generate_config
+from optimizer import ONadam
+from metric import LossValueMetric, NMEMetric
+sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
+import sym_heatmap
+#import sym_fc
+#from symbol import fc
+
+args = None
+logger = logging.getLogger()
+logger.setLevel(logging.INFO)
+
+
+def main(args):
+    _seed = 727
+    random.seed(_seed)
+    np.random.seed(_seed)
+    mx.random.seed(_seed)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    #ctx = [mx.gpu(0)]
+    args.ctx_num = len(ctx)
+
+    args.batch_size = args.per_batch_size * args.ctx_num
+    config.per_batch_size = args.per_batch_size
+
+    print('Call with', args, config)
+    train_iter = FaceSegIter(
+        path_imgrec=os.path.join(config.dataset_path, 'train.rec'),
+        batch_size=args.batch_size,
+        per_batch_size=args.per_batch_size,
+        aug_level=1,
+        exf=args.exf,
+        args=args,
+    )
+
+    data_shape = train_iter.get_data_shape()
+    #label_shape = train_iter.get_label_shape()
+    sym = sym_heatmap.get_symbol(num_classes=config.num_classes)
+    if len(args.pretrained) == 0:
+        #data_shape_dict = {'data' : (args.per_batch_size,)+data_shape, 'softmax_label' : (args.per_batch_size,)+label_shape}
+        data_shape_dict = train_iter.get_shape_dict()
+        arg_params, aux_params = sym_heatmap.init_weights(sym, data_shape_dict)
+    else:
+        vec = args.pretrained.split(',')
+        print('loading', vec)
+        _, arg_params, aux_params = mx.model.load_checkpoint(
+            vec[0], int(vec[1]))
+        #sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
+
+    model = mx.mod.Module(
+        context=ctx,
+        symbol=sym,
+        label_names=train_iter.get_label_names(),
+    )
+    #lr = 1.0e-3
+    #lr = 2.5e-4
+    _rescale_grad = 1.0 / args.ctx_num
+    #_rescale_grad = 1.0/args.batch_size
+    #lr = args.lr
+    #opt = optimizer.Nadam(learning_rate=args.lr, wd=args.wd, rescale_grad=_rescale_grad, clip_gradient=5.0)
+    if args.optimizer == 'onadam':
+        opt = ONadam(learning_rate=args.lr,
+                     wd=args.wd,
+                     rescale_grad=_rescale_grad,
+                     clip_gradient=5.0)
+    elif args.optimizer == 'nadam':
+        opt = optimizer.Nadam(learning_rate=args.lr,
+                              rescale_grad=_rescale_grad)
+    elif args.optimizer == 'rmsprop':
+        opt = optimizer.RMSProp(learning_rate=args.lr,
+                                rescale_grad=_rescale_grad)
+    elif args.optimizer == 'adam':
+        opt = optimizer.Adam(learning_rate=args.lr, rescale_grad=_rescale_grad)
+    else:
+        opt = optimizer.SGD(learning_rate=args.lr,
+                            momentum=0.9,
+                            wd=args.wd,
+                            rescale_grad=_rescale_grad)
+    initializer = mx.init.Xavier(rnd_type='gaussian',
+                                 factor_type="in",
+                                 magnitude=2)
+    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
+    _metric = LossValueMetric()
+    #_metric = NMEMetric()
+    #_metric2 = AccMetric()
+    #eval_metrics = [_metric, _metric2]
+    eval_metrics = [_metric]
+    lr_steps = [int(x) for x in args.lr_step.split(',')]
+    print('lr-steps', lr_steps)
+    global_step = [0]
+
+    def val_test():
+        all_layers = sym.get_internals()
+        vsym = all_layers['heatmap_output']
+        vmodel = mx.mod.Module(symbol=vsym, context=ctx, label_names=None)
+        #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+        vmodel.bind(data_shapes=[('data', (args.batch_size, ) + data_shape)])
+        arg_params, aux_params = model.get_params()
+        vmodel.set_params(arg_params, aux_params)
+        for target in config.val_targets:
+            _file = os.path.join(config.dataset_path, '%s.rec' % target)
+            if not os.path.exists(_file):
+                continue
+            val_iter = FaceSegIter(
+                path_imgrec=_file,
+                batch_size=args.batch_size,
+                #batch_size = 4,
+                aug_level=0,
+                args=args,
+            )
+            _metric = NMEMetric()
+            val_metric = mx.metric.create(_metric)
+            val_metric.reset()
+            val_iter.reset()
+            for i, eval_batch in enumerate(val_iter):
+                #print(eval_batch.data[0].shape, eval_batch.label[0].shape)
+                batch_data = mx.io.DataBatch(eval_batch.data)
+                model.forward(batch_data, is_train=False)
+                model.update_metric(val_metric, eval_batch.label)
+            nme_value = val_metric.get_name_value()[0][1]
+            print('[%d][%s]NME: %f' % (global_step[0], target, nme_value))
+
+    def _batch_callback(param):
+        _cb(param)
+        global_step[0] += 1
+        mbatch = global_step[0]
+        for _lr in lr_steps:
+            if mbatch == _lr:
+                opt.lr *= 0.2
+                print('lr change to', opt.lr)
+                break
+        if mbatch % 1000 == 0:
+            print('lr-batch-epoch:', opt.lr, param.nbatch, param.epoch)
+        if mbatch > 0 and mbatch % args.verbose == 0:
+            val_test()
+            if args.ckpt == 1:
+                msave = mbatch // args.verbose
+                print('saving', msave)
+                arg, aux = model.get_params()
+                mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg,
+                                         aux)
+        if mbatch == lr_steps[-1]:
+            if args.ckpt == 2:
+                #msave = mbatch//args.verbose
+                msave = 1
+                print('saving', msave)
+                arg, aux = model.get_params()
+                mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg,
+                                         aux)
+            sys.exit(0)
+
+    train_iter = mx.io.PrefetchingIter(train_iter)
+
+    model.fit(
+        train_iter,
+        begin_epoch=0,
+        num_epoch=9999,
+        #eval_data          = val_iter,
+        eval_data=None,
+        eval_metric=eval_metrics,
+        kvstore='device',
+        optimizer=opt,
+        initializer=initializer,
+        arg_params=arg_params,
+        aux_params=aux_params,
+        allow_missing=True,
+        batch_end_callback=_batch_callback,
+        epoch_end_callback=None,
+    )
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Train face alignment')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--prefix',
+                        default=default.prefix,
+                        help='directory to save model.')
+    parser.add_argument('--pretrained', default=default.pretrained, help='')
+    parser.add_argument('--optimizer', default='nadam', help='')
+    parser.add_argument('--lr', type=float, default=default.lr, help='')
+    parser.add_argument('--wd', type=float, default=default.wd, help='')
+    parser.add_argument('--per-batch-size',
+                        type=int,
+                        default=default.per_batch_size,
+                        help='')
+    parser.add_argument('--lr-step',
+                        help='learning rate steps (in epoch)',
+                        default=default.lr_step,
+                        type=str)
+    parser.add_argument('--ckpt', type=int, default=1, help='')
+    parser.add_argument('--norm', type=int, default=0, help='')
+    parser.add_argument('--exf', type=int, default=1, help='')
+    parser.add_argument('--frequent',
+                        type=int,
+                        default=default.frequent,
+                        help='')
+    parser.add_argument('--verbose',
+                        type=int,
+                        default=default.verbose,
+                        help='')
+    args = parser.parse_args()
+    main(args)
diff --git a/alignment/heatmapReg/data.py b/alignment/heatmapReg/data.py
deleted file mode 100644
index de6bcb0ca..000000000
--- a/alignment/heatmapReg/data.py
+++ /dev/null
@@ -1,340 +0,0 @@
-# pylint: skip-file
-import mxnet as mx
-import numpy as np
-import sys, os
-import random
-import math
-import scipy.misc
-import cv2
-import logging
-import sklearn
-import datetime
-import img_helper
-from mxnet.io import DataIter
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-from PIL import Image
-from config import config
-from skimage import transform as tf
-
-class FaceSegIter(DataIter):
-    def __init__(self, batch_size, 
-                 per_batch_size = 0,
-                 path_imgrec = None,
-                 aug_level = 0,
-                 force_mirror = False,
-                 exf = 1,
-                 use_coherent = 0,
-                 args = None,
-                 data_name = "data",
-                 label_name = "softmax_label"):
-      self.aug_level = aug_level
-      self.force_mirror = force_mirror
-      self.use_coherent = use_coherent
-      self.exf = exf
-      self.batch_size = batch_size
-      self.per_batch_size = per_batch_size
-      self.data_name = data_name
-      self.label_name = label_name
-      assert path_imgrec
-      logging.info('loading recordio %s...',
-                   path_imgrec)
-      path_imgidx = path_imgrec[0:-4]+".idx"
-      self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-      self.oseq = list(self.imgrec.keys)
-      print('train size', len(self.oseq))
-      self.cur = 0
-      self.reset()
-      self.data_shape = (3, config.input_img_size, config.input_img_size)
-      self.num_classes = config.num_classes
-      self.input_img_size = config.input_img_size
-      #self.label_classes = self.num_classes
-      if config.losstype=='heatmap':
-        if aug_level>0:
-          self.output_label_size = config.output_label_size
-          self.label_shape = (self.num_classes, self.output_label_size, self.output_label_size)
-        else:
-          self.output_label_size = self.input_img_size
-          #self.label_shape = (self.num_classes, 2)
-          self.label_shape = (self.num_classes, self.output_label_size, self.output_label_size)
-      else:
-        if aug_level>0:
-          self.output_label_size = config.output_label_size
-          self.label_shape = (self.num_classes, 2)
-        else:
-          self.output_label_size = self.input_img_size
-          #self.label_shape = (self.num_classes, 2)
-          self.label_shape = (self.num_classes, 2)
-      self.provide_data = [(data_name, (batch_size,) + self.data_shape)]
-      self.provide_label = [(label_name, (batch_size,) + self.label_shape)]
-      self.img_num = 0
-      self.invalid_num = 0
-      self.mode = 1
-      self.vis = 0
-      self.stats = [0,0]
-      self.flip_order = [16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 
-          26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 27, 28, 29, 30, 35, 34, 33, 32, 31, 
-          45, 44, 43, 42, 47, 46, 39, 38, 37, 36, 41, 40, 54, 53, 52, 51, 50, 49, 48, 
-          59, 58, 57, 56, 55, 64, 63, 62, 61, 60, 67, 66, 65]
-      #self.mirror_set = [
-      #        (22,23),
-      #        (21,24),
-      #        (20,25),
-      #        (19,26),
-      #        (18,27),
-      #        (40,43),
-      #        (39,44),
-      #        (38,45),
-      #        (37,46),
-      #        (42,47),
-      #        (41,48),
-      #        (33,35),
-      #        (32,36),
-      #        (51,53),
-      #        (50,54),
-      #        (62,64),
-      #        (61,65),
-      #        (49,55),
-      #        (49,55),
-      #        (68,66),
-      #        (60,56),
-      #        (59,57),
-      #        (1,17),
-      #        (2,16),
-      #        (3,15),
-      #        (4,14),
-      #        (5,13),
-      #        (6,12),
-      #        (7,11),
-      #        (8,10),
-      #    ]
-
-    def get_data_shape(self):
-        return self.data_shape
-
-    #def get_label_shape(self):
-    #    return self.label_shape
-
-    def get_shape_dict(self):
-        D = {}
-        for (k,v) in self.provide_data:
-            D[k] = v
-        for (k,v) in self.provide_label:
-            D[k] = v
-        return D
-
-    def get_label_names(self):
-        D = []
-        for (k,v) in self.provide_label:
-            D.append(k)
-        return D
-
-    def reset(self):
-      #print('reset')
-      if self.aug_level==0:
-        self.seq = self.oseq
-      else:
-        self.seq = []
-        for _ in range(self.exf):
-          _seq = self.oseq[:]
-          random.shuffle(_seq)
-          self.seq += _seq
-        print('train size after reset', len(self.seq))
-      self.cur = 0
-
-    def next_sample(self):
-      """Helper function for reading in next sample."""
-      if self.cur >= len(self.seq):
-        raise StopIteration
-      idx = self.seq[self.cur]
-      self.cur += 1
-      s = self.imgrec.read_idx(idx)
-      header, img = recordio.unpack(s)
-      img = mx.image.imdecode(img).asnumpy()
-      hlabel = np.array(header.label).reshape( (self.num_classes,2) )
-      if not config.label_xfirst:
-        hlabel = hlabel[:,::-1] #convert to X/W first
-      annot = {'scale': config.base_scale}
-
-      #ul = np.array( (50000,50000), dtype=np.int32)
-      #br = np.array( (0,0), dtype=np.int32)
-      #for i in range(hlabel.shape[0]):
-      #  h = int(hlabel[i][0])
-      #  w = int(hlabel[i][1])
-      #  key = np.array((h,w))
-      #  ul = np.minimum(key, ul)
-      #  br = np.maximum(key, br)
-
-      return img, hlabel, annot
-
-    def get_flip(self, data, label):
-      data_flip = np.zeros_like(data)
-      label_flip = np.zeros_like(label)
-      for k in range(data_flip.shape[2]):
-          data_flip[:,:,k] = np.fliplr(data[:,:,k])
-      for k in range(label_flip.shape[0]):
-          label_flip[k,:] = np.fliplr(label[k,:])
-      #print(label[0,:].shape)
-      label_flip = label_flip[self.flip_order,:]
-      return data_flip, label_flip
-
-    def get_data(self, data, label, annot):
-      if self.vis:
-        self.img_num+=1
-        #if self.img_num<=self.vis:
-        #  filename = './vis/raw_%d.jpg' % (self.img_num)
-        #  print('save', filename)
-        #  draw = data.copy()
-        #  for i in range(label.shape[0]):
-        #    cv2.circle(draw, (label[i][1], label[i][0]), 1, (0, 0, 255), 2)
-        #  scipy.misc.imsave(filename, draw)
-
-      rotate = 0
-      #scale = 1.0
-      if 'scale' in annot:
-          scale = annot['scale']
-      else:
-          scale = max(data.shape[0], data.shape[1])
-      if 'center' in annot:
-        center = annot['center']
-      else:
-        center = np.array( (data.shape[1]/2, data.shape[0]/2) )
-      max_retry = 3
-      if self.aug_level==0: #validation mode
-          max_retry = 6
-      retry = 0
-      found = False
-      base_scale = scale
-      while retry<max_retry:
-          retry+=1
-          succ = True
-          _scale = base_scale
-          if self.aug_level>0:
-            rotate = np.random.randint(-40, 40)
-            scale_config = 0.2
-            #rotate = 0
-            #scale_config = 0.0
-            scale_ratio = min(1+scale_config, max(1-scale_config, (np.random.randn() * scale_config) + 1))
-            _scale = int(base_scale * scale_ratio)
-            #translate = np.random.randint(-5, 5, size=(2,))
-            #center += translate
-          data_out, trans = img_helper.transform(data, center, self.input_img_size, _scale, rotate)
-          #data_out = img_helper.crop2(data, center, _scale, (self.input_img_size, self.input_img_size), rot=rotate)
-          label_out = np.zeros(self.label_shape, dtype=np.float32)
-          #print('out shapes', data_out.shape, label_out.shape)
-          for i in range(label.shape[0]):
-            pt = label[i].copy()
-            #pt = pt[::-1]
-            npt = img_helper.transform_pt(pt, trans)
-            if npt[0]>=data_out.shape[1] or npt[1]>=data_out.shape[0] or npt[0]<0 or npt[1]<0:
-              succ = False
-              #print('err npt', npt)
-              break
-            if config.losstype=='heatmap':
-              pt_scale = float(self.output_label_size)/self.input_img_size
-              npt *= pt_scale
-              npt = npt.astype(np.int32)
-              img_helper.gaussian(label_out[i], npt, config.gaussian)
-            else:
-              label_out[i] = (npt/self.input_img_size)
-            #print('before gaussian', label_out[i].shape, pt.shape)
-            #trans = img_helper.transform(pt, center, _scale, (self.output_label_size, self.output_label_size), rot=rotate)
-            #print(trans.shape)
-            #if not img_helper.gaussian(label_out[i], trans, _g):
-            #    succ = False
-            #    break
-          if not succ:
-              if self.aug_level==0:
-                  base_scale+=20
-              continue
-          
-          flip_data_out = None
-          flip_label_out = None
-          if config.net_coherent:
-            flip_data_out, flip_label_out = self.get_flip(data_out, label_out)
-          elif ((self.aug_level>0 and np.random.rand() < 0.5) or self.force_mirror): #flip aug
-            flip_data_out, flip_label_out = self.get_flip(data_out, label_out)
-            data_out, label_out = flip_data_out, flip_label_out
-
-          found = True
-          break
-
-
-      #self.stats[0]+=1
-      if not found:
-          #self.stats[1]+=1
-          #print('find aug error', retry)
-          #print(self.stats)
-          #print('!!!ERR')
-          return None
-      #print('found with scale', _scale, rotate)
-
-
-      if self.vis>0 and self.img_num<=self.vis:
-        print('crop', data.shape, center, _scale, rotate, data_out.shape)
-        filename = './vis/cropped_%d.jpg' % (self.img_num)
-        print('save', filename)
-        draw = data_out.copy()
-        alabel = label_out.copy()
-        for i in range(label.shape[0]):
-          a = cv2.resize(alabel[i], (self.input_img_size, self.input_img_size))
-          ind = np.unravel_index(np.argmax(a, axis=None), a.shape)
-          cv2.circle(draw, (ind[1], ind[0]), 1, (0, 0, 255), 2)
-        scipy.misc.imsave(filename, draw)
-        filename = './vis/raw_%d.jpg' % (self.img_num)
-        scipy.misc.imsave(filename, data)
-
-      return data_out, label_out, flip_data_out,flip_label_out 
-
-    def next(self):
-        """Returns the next batch of data."""
-        #print('next')
-        batch_size = self.batch_size
-        batch_data = nd.empty((batch_size,)+self.data_shape)
-        batch_label = nd.empty((batch_size,)+self.label_shape)
-        i = 0
-        #self.cutoff = random.randint(800,1280)
-        try:
-            while i < batch_size:
-                #print('N', i)
-                data, label, annot = self.next_sample()
-                R = self.get_data(data, label, annot)
-                if R is None:
-                  continue
-                data_out, label_out, flip_data_out, flip_label_out = R
-                if not self.use_coherent:
-                    data = nd.array(data_out)
-                    data = nd.transpose(data, axes=(2, 0, 1))
-                    label = nd.array(label_out)
-                    #print(data.shape, label.shape)
-                    batch_data[i][:] = data
-                    batch_label[i][:] = label
-                    i += 1
-                else:
-                    data = nd.array(data_out)
-                    data = nd.transpose(data, axes=(2, 0, 1))
-                    label = nd.array(label_out)
-                    data2 = nd.array(flip_data_out)
-                    data2 = nd.transpose(data2, axes=(2, 0, 1))
-                    label2 = nd.array(flip_label_out)
-                    #M = nd.array(M)
-                    #print(data.shape, label.shape)
-                    batch_data[i][:] = data
-                    batch_label[i][:] = label
-                    #i+=1
-                    j = i+self.per_batch_size//2
-                    batch_data[j][:] = data2
-                    batch_label[j][:] = label2
-                    i += 1
-                    if j%self.per_batch_size==self.per_batch_size-1:
-                        i = j+1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        #return {self.data_name  :  batch_data,
-        #        self.label_name :  batch_label}
-        #print(batch_data.shape, batch_label.shape)
-        return mx.io.DataBatch([batch_data], [batch_label], batch_size - i)
-
diff --git a/alignment/heatmapReg/img_helper.py b/alignment/heatmapReg/img_helper.py
deleted file mode 100644
index 86a1d5777..000000000
--- a/alignment/heatmapReg/img_helper.py
+++ /dev/null
@@ -1,80 +0,0 @@
-import numpy as np
-import math
-import cv2
-from skimage import transform as stf
-
-def transform(data, center, output_size, scale, rotation):
-    scale_ratio = float(output_size)/scale
-    rot = float(rotation)*np.pi/180.0
-    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
-    t1 = stf.SimilarityTransform(scale=scale_ratio)
-    cx = center[0]*scale_ratio
-    cy = center[1]*scale_ratio
-    t2 = stf.SimilarityTransform(translation=(-1*cx, -1*cy))
-    t3 = stf.SimilarityTransform(rotation=rot)
-    t4 = stf.SimilarityTransform(translation=(output_size/2, output_size/2))
-    t = t1+t2+t3+t4
-    trans = t.params[0:2]
-    #print('M', scale, rotation, trans)
-    cropped = cv2.warpAffine(data,trans,(output_size, output_size), borderValue = 0.0)
-    return cropped, trans
-
-def transform_pt(pt, trans):
-    new_pt = np.array([pt[0], pt[1], 1.]).T
-    new_pt = np.dot(trans, new_pt)
-    #print('new_pt', new_pt.shape, new_pt)
-    return new_pt[:2]
-
-def gaussian(img, pt, sigma):
-    # Draw a 2D gaussian
-    assert(sigma>=0)
-    if sigma==0:
-      img[pt[1], pt[0]] = 1.0
-      return True
-    #assert pt[0]<=img.shape[1]
-    #assert pt[1]<=img.shape[0]
-
-    # Check that any part of the gaussian is in-bounds
-    ul = [int(pt[0] - 3 * sigma), int(pt[1] - 3 * sigma)]
-    br = [int(pt[0] + 3 * sigma + 1), int(pt[1] + 3 * sigma + 1)]
-    if (ul[0] > img.shape[1] or ul[1] >= img.shape[0] or
-            br[0] < 0 or br[1] < 0):
-        # If not, just return the image as is
-        #print('gaussian error')
-        return False
-        #return img
-
-    # Generate gaussian
-    size = 6 * sigma + 1
-    x = np.arange(0, size, 1, float)
-    y = x[:, np.newaxis]
-    x0 = y0 = size // 2
-    # The gaussian is not normalized, we want the center value to equal 1
-    g = np.exp(- ((x - x0) ** 2 + (y - y0) ** 2) / (2 * sigma ** 2))
-
-    # Usable gaussian range
-    g_x = max(0, -ul[0]), min(br[0], img.shape[1]) - ul[0]
-    g_y = max(0, -ul[1]), min(br[1], img.shape[0]) - ul[1]
-    # Image range
-    img_x = max(0, ul[0]), min(br[0], img.shape[1])
-    img_y = max(0, ul[1]), min(br[1], img.shape[0])
-
-    img[img_y[0]:img_y[1], img_x[0]:img_x[1]] = g[g_y[0]:g_y[1], g_x[0]:g_x[1]]
-    return True
-    #return img
-
-def estimate_trans_bbox(face, input_size, s = 2.0):
-  w = face[2] - face[0]
-  h = face[3] - face[1]
-  wc = int( (face[2]+face[0])/2 )
-  hc = int( (face[3]+face[1])/2 )
-  im_size = max(w, h)
-  #size = int(im_size*1.2)
-  scale = input_size/(max(w,h)*s)
-  M = [ 
-        [scale, 0, input_size/2-wc*scale],
-        [0, scale, input_size/2-hc*scale],
-      ]
-  M = np.array(M)
-  return M
-
diff --git a/alignment/heatmapReg/metric.py b/alignment/heatmapReg/metric.py
deleted file mode 100644
index 16ffe1c95..000000000
--- a/alignment/heatmapReg/metric.py
+++ /dev/null
@@ -1,99 +0,0 @@
-import mxnet as mx
-import numpy as np
-import math
-import cv2
-from config import config
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[0].asnumpy()[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-
-class NMEMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(NMEMetric, self).__init__(
-        'NME', axis=self.axis,
-        output_names=None, label_names=None)
-    #self.losses = []
-    self.count = 0
-
-  def cal_nme(self, label, pred_label):
-    nme = []
-    for b in range(pred_label.shape[0]):
-      record = [None]*6
-      item = []
-      if label.ndim==4:
-          _heatmap = label[b][36]
-          if np.count_nonzero(_heatmap)==0:
-              continue
-      else:#ndim==3
-          #print(label[b])
-          if np.count_nonzero(label[b])==0:
-              continue
-      for p in range(pred_label.shape[1]):
-        if label.ndim==4:
-            heatmap_gt = label[b][p]
-            ind_gt = np.unravel_index(np.argmax(heatmap_gt, axis=None), heatmap_gt.shape)
-            ind_gt = np.array(ind_gt)
-        else:
-            ind_gt = label[b][p]
-            #ind_gt = ind_gt.astype(np.int)
-            #print(ind_gt)
-        heatmap_pred = pred_label[b][p]
-        heatmap_pred = cv2.resize(heatmap_pred, (config.input_img_size, config.input_img_size))
-        ind_pred = np.unravel_index(np.argmax(heatmap_pred, axis=None), heatmap_pred.shape)
-        ind_pred = np.array(ind_pred)
-        #print(ind_gt.shape)
-        #print(ind_pred)
-        if p==36:
-            #print('b', b, p, ind_gt, np.count_nonzero(heatmap_gt))
-            record[0] = ind_gt
-        elif p==39:
-            record[1] = ind_gt
-        elif p==42:
-            record[2] = ind_gt
-        elif p==45:
-            record[3] = ind_gt
-        if record[4] is None or record[5] is None:
-            record[4] = ind_gt
-            record[5] = ind_gt
-        else:
-            record[4] = np.minimum(record[4], ind_gt)
-            record[5] = np.maximum(record[5], ind_gt)
-        #print(ind_gt.shape, ind_pred.shape)
-        value = np.sqrt(np.sum(np.square(ind_gt - ind_pred)))
-        item.append(value)
-      _nme = np.mean(item)
-      if config.landmark_type=='2d':
-          left_eye = (record[0]+record[1])/2
-          right_eye = (record[2]+record[3])/2
-          _dist = np.sqrt(np.sum(np.square(left_eye - right_eye)))
-          #print('eye dist', _dist, left_eye, right_eye)
-          _nme /= _dist
-      else:
-          #_dist = np.sqrt(float(label.shape[2]*label.shape[3]))
-          _dist = np.sqrt(np.sum(np.square(record[5] - record[4])))
-          #print(_dist)
-          _nme /= _dist
-      nme.append(_nme)
-    return np.mean(nme)
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0].asnumpy()
-    pred_label = preds[-1].asnumpy()
-    nme = self.cal_nme(label, pred_label)
-
-    #print('nme', nme)
-    #nme = np.mean(nme)
-    self.sum_metric += np.mean(nme)
-    self.num_inst += 1.0
diff --git a/alignment/heatmapReg/optimizer.py b/alignment/heatmapReg/optimizer.py
deleted file mode 100644
index d86970c95..000000000
--- a/alignment/heatmapReg/optimizer.py
+++ /dev/null
@@ -1,54 +0,0 @@
-import mxnet as mx
-import mxnet.optimizer as optimizer
-from mxnet.ndarray import (NDArray, zeros, clip, sqrt, cast, maximum, abs as NDabs)
-#from mxnet.ndarray import (sgd_update, sgd_mom_update, adam_update, rmsprop_update, rmspropalex_update,
-#                      mp_sgd_update, mp_sgd_mom_update, square, ftrl_update)
-
-class ONadam(optimizer.Optimizer):
-    def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8,
-                 schedule_decay=0.004, **kwargs):
-        super(ONadam, self).__init__(learning_rate=learning_rate, **kwargs)
-        self.beta1 = beta1
-        self.beta2 = beta2
-        self.epsilon = epsilon
-        self.schedule_decay = schedule_decay
-        self.m_schedule = 1.
-
-    def create_state(self, index, weight):
-        return (zeros(weight.shape, weight.context, dtype=weight.dtype),  # mean
-                zeros(weight.shape, weight.context, dtype=weight.dtype))  # variance
-
-    def update(self, index, weight, grad, state):
-        assert(isinstance(weight, NDArray))
-        assert(isinstance(grad, NDArray))
-        self._update_count(index)
-        lr = self._get_lr(index)
-        wd = self._get_wd(index)
-
-        t = self._index_update_count[index]
-
-        # preprocess grad
-        #grad = grad * self.rescale_grad + wd * weight
-        grad *= self.rescale_grad + wd * weight
-        if self.clip_gradient is not None:
-            grad = clip(grad, -self.clip_gradient, self.clip_gradient)
-
-        # warming momentum schedule
-        momentum_t = self.beta1 * (1. - 0.5 * (pow(0.96, t * self.schedule_decay)))
-        momentum_t_1 = self.beta1 * (1. - 0.5 * (pow(0.96, (t + 1) * self.schedule_decay)))
-        self.m_schedule = self.m_schedule * momentum_t
-        m_schedule_next = self.m_schedule * momentum_t_1
-
-        # update m_t and v_t
-        m_t, v_t = state
-        m_t[:] = self.beta1 * m_t + (1. - self.beta1) * grad
-        v_t[:] = self.beta2 * v_t + (1. - self.beta2) * grad * grad
-
-        grad_prime = grad / (1. - self.m_schedule)
-        m_t_prime = m_t / (1. - m_schedule_next)
-        v_t_prime = v_t / (1. - pow(self.beta2, t))
-        m_t_bar = (1. - momentum_t) * grad_prime + momentum_t_1 * m_t_prime
-
-        # update weight
-        weight[:] -= lr * m_t_bar / (sqrt(v_t_prime) + self.epsilon)
-
diff --git a/alignment/heatmapReg/sample_config.py b/alignment/heatmapReg/sample_config.py
deleted file mode 100644
index 087063894..000000000
--- a/alignment/heatmapReg/sample_config.py
+++ /dev/null
@@ -1,100 +0,0 @@
-import numpy as np
-from easydict import EasyDict as edict
-
-config = edict()
-
-#default training/dataset config
-config.num_classes = 68
-config.record_img_size = 384
-config.base_scale = 256
-config.input_img_size = 128
-config.output_label_size = 64
-config.label_xfirst = False
-config.losstype = 'heatmap'
-config.net_coherent = False
-config.multiplier = 1.0
-
-config.gaussian = 0
-
-# network settings
-network = edict()
-
-network.hourglass = edict()
-network.hourglass.net_coherent = False
-network.hourglass.net_sta = 0
-network.hourglass.net_n = 3
-network.hourglass.net_dcn = 0
-network.hourglass.net_stacks = 2
-network.hourglass.net_block = 'resnet'
-network.hourglass.net_binarize = False
-network.hourglass.losstype = 'heatmap'
-
-network.sdu = edict()
-network.sdu.net_coherent = False
-network.sdu.net_sta = 1
-network.sdu.net_n = 3
-network.sdu.net_dcn = 3
-network.sdu.net_stacks = 2
-network.sdu.net_block = 'cab'
-network.sdu.net_binarize = False
-network.sdu.losstype = 'heatmap'
-
-
-# dataset settings
-dataset = edict()
-
-dataset.i2d = edict()
-dataset.i2d.dataset = '2D'
-dataset.i2d.landmark_type = '2d'
-dataset.i2d.dataset_path = './data_2d'
-dataset.i2d.num_classes = 68
-dataset.i2d.record_img_size = 384
-dataset.i2d.base_scale = 256
-dataset.i2d.input_img_size = 128
-dataset.i2d.output_label_size = 64
-dataset.i2d.label_xfirst = False
-dataset.i2d.val_targets = ['ibug', 'cofw_testset', '300W']
-
-dataset.i3d = edict()
-dataset.i3d.dataset = '3D'
-dataset.i3d.landmark_type = '3d'
-dataset.i3d.dataset_path = './data_3d'
-dataset.i3d.num_classes = 68
-dataset.i3d.record_img_size = 384
-dataset.i3d.base_scale = 256
-dataset.i3d.input_img_size = 128
-dataset.i3d.output_label_size = 64
-dataset.i3d.label_xfirst = False
-dataset.i3d.val_targets = ['AFLW2000-3D']
-
-
-# default settings
-default = edict()
-
-# default network
-default.network = 'hourglass'
-default.pretrained = ''
-default.pretrained_epoch = 0
-# default dataset
-default.dataset = 'i2d'
-default.frequent = 20
-default.verbose = 200
-default.kvstore = 'device'
-
-default.prefix = 'model/A'
-default.end_epoch = 10000
-default.lr = 0.00025
-default.wd = 0.0
-default.per_batch_size = 20
-default.lr_step = '16000,24000,30000'
-
-def generate_config(_network, _dataset):
-    for k, v in network[_network].items():
-      config[k] = v
-      default[k] = v
-    for k, v in dataset[_dataset].items():
-      config[k] = v
-      default[k] = v
-    config.network = _network
-    config.dataset = _dataset
-
diff --git a/alignment/heatmapReg/symbol/sym_heatmap.py b/alignment/heatmapReg/symbol/sym_heatmap.py
deleted file mode 100644
index 12a7e94bf..000000000
--- a/alignment/heatmapReg/symbol/sym_heatmap.py
+++ /dev/null
@@ -1,618 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import mxnet as mx
-import numpy as np
-from config import config
-
-
-ACT_BIT = 1
-bn_mom = 0.9
-workspace = 256
-memonger = False
-
-
-
-def Conv(**kwargs):
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-def Act(data, act_type, name):
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-#def lin(data, num_filter, workspace, name, binarize, dcn):
-#  bit = 1
-#  if not binarize:
-#    if not dcn:
-#        conv1 = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-#                                      no_bias=True, workspace=workspace, name=name + '_conv')
-#        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#        return act1
-#    else:
-#        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#        act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#        conv1_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = act1,
-#                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-#        conv1 = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=act1, offset=conv1_offset,
-#                num_filter=num_filter, pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=False)
-#        #conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-#        #                              no_bias=False, workspace=workspace, name=name + '_conv')
-#        return conv1
-#  else:
-#    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-#    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-#    conv1 = mx.sym.QConvolution_v1(data=act1, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-#                               no_bias=True, workspace=workspace, name=name + '_conv', act_bit=ACT_BIT, weight_bit=bit)
-#    conv1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-#    return conv1
-
-def lin3(data, num_filter, workspace, name, k, g=1, d=1):
-    if k!=3:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(k,k), stride=(1,1), pad=((k-1)//2,(k-1)//2), num_group=g,
-                                      no_bias=True, workspace=workspace, name=name + '_conv')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(k,k), stride=(1,1), pad=(d,d), num_group=g, dilate=(d, d),
-                                      no_bias=True, workspace=workspace, name=name + '_conv')
-    bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn')
-    act1 = Act(data=bn1, act_type='relu', name=name + '_relu')
-    ret = act1
-    return ret
-
-def ConvFactory(data, num_filter, kernel, stride=(1, 1), pad=(0, 0), act_type="relu", mirror_attr={}, with_act=True, dcn=False, name=''):
-    if not dcn:
-      conv = mx.symbol.Convolution(
-          data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad, no_bias=True, workspace=workspace, name=name+'_conv')
-    else:
-        conv_offset = mx.symbol.Convolution(name=name+'_conv_offset', data = data,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-        conv = mx.contrib.symbol.DeformableConvolution(name=name+"_conv", data=data, offset=conv_offset,
-                num_filter=num_filter, pad=(1,1), kernel=(3,3), num_deformable_group=1, stride=stride, dilate=(1, 1), no_bias=False)
-    bn = mx.symbol.BatchNorm(data=conv, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name+'_bn')
-    if with_act:
-      act = Act(bn, act_type, name=name+'_relu')
-      #act = mx.symbol.Activation(
-      #    data=bn, act_type=act_type, attr=mirror_attr, name=name+'_relu')
-      return act
-    else:
-      return bn
-
-class CAB:
-    def __init__(self, data, nFilters, nModules, n, workspace, name, dilate, group):
-        self.data = data
-        self.nFilters = nFilters
-        self.nModules = nModules
-        self.n = n
-        self.workspace = workspace
-        self.name = name
-        self.dilate = dilate
-        self.group = group
-        self.sym_map = {}
-
-    def get_output(self, w, h):
-        key = (w, h)
-        if key in self.sym_map:
-            return self.sym_map[key]
-        ret = None
-        if h==self.n:
-            if w==self.n:
-                ret = (self.data, self.nFilters)
-            else:
-                x = self.get_output(w+1, h)
-                f = int(x[1]*0.5)
-                if w!=self.n-1:
-                    body = lin3(x[0], f, self.workspace, "%s_w%d_h%d_1"%(self.name, w, h), 3, self.group, 1)
-                else:
-                    body = lin3(x[0], f, self.workspace, "%s_w%d_h%d_1"%(self.name, w, h), 3, self.group, self.dilate)
-                ret = (body,f)
-        else:
-            x = self.get_output(w+1, h+1)
-            y = self.get_output(w, h+1)
-            if h%2==1 and h!=w:
-                xbody = lin3(x[0], x[1], self.workspace, "%s_w%d_h%d_2"%(self.name, w, h), 3, x[1])
-                #xbody = xbody+x[0]
-            else:
-                xbody = x[0]
-            #xbody = x[0]
-            #xbody = lin3(x[0], x[1], self.workspace, "%s_w%d_h%d_2"%(self.name, w, h), 3, x[1])
-            if w==0:
-                ybody = lin3(y[0], y[1], self.workspace, "%s_w%d_h%d_3"%(self.name, w, h), 3, self.group)
-            else:
-                ybody = y[0]
-            ybody = mx.sym.concat(y[0], ybody, dim=1)
-            body = mx.sym.add_n(xbody,ybody, name="%s_w%d_h%d_add"%(self.name, w, h))
-            body = body/2
-            ret = (body, x[1])
-        self.sym_map[key] = ret
-        return ret
-
-    def get(self):
-        return self.get_output(1, 1)[0]
-
-def conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    bit = 1
-    #print('in unit2')
-    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    if not binarize:
-      act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
-      conv1 = Conv(data=act1, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                 no_bias=True, workspace=workspace, name=name + '_conv1')
-    else:
-      act1 = mx.sym.QActivation(data=bn1, act_bit=ACT_BIT, name=name + '_relu1', backward_only=True)
-      conv1 = mx.sym.QConvolution(data=act1, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                 no_bias=True, workspace=workspace, name=name + '_conv1', act_bit=ACT_BIT, weight_bit=bit)
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    if not binarize:
-      act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
-      conv2 = Conv(data=act2, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                 no_bias=True, workspace=workspace, name=name + '_conv2')
-    else:
-      act2 = mx.sym.QActivation(data=bn2, act_bit=ACT_BIT, name=name + '_relu2', backward_only=True)
-      conv2 = mx.sym.QConvolution(data=act2, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                 no_bias=True, workspace=workspace, name=name + '_conv2', act_bit=ACT_BIT, weight_bit=bit)
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    if not binarize:
-      act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
-      conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                 workspace=workspace, name=name + '_conv3')
-    else:
-      act3 = mx.sym.QActivation(data=bn3, act_bit=ACT_BIT, name=name + '_relu3', backward_only=True)
-      conv3 = mx.sym.QConvolution(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                 no_bias=True, workspace=workspace, name=name + '_conv3', act_bit=ACT_BIT, weight_bit=bit)
-    #if binarize:
-    #  conv3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-    if dim_match:
-        shortcut = data
-    else:
-        if not binarize:
-          shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                          workspace=workspace, name=name+'_sc')
-        else:
-          shortcut = mx.sym.QConvolution(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_sc', act_bit=ACT_BIT, weight_bit=bit)
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return conv3 + shortcut
-
-
-def conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilation, **kwargs):
-    bit = 1
-    #print('in unit2')
-    # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    if not binarize:
-      act1 = Act(data=bn1, act_type='relu', name=name + '_relu1')
-      if not dcn:
-          conv1 = Conv(data=act1, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation),
-                                     no_bias=True, workspace=workspace, name=name + '_conv1')
-      else:
-          conv1_offset = mx.symbol.Convolution(name=name+'_conv1_offset', data = act1,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-          conv1 = mx.contrib.symbol.DeformableConvolution(name=name+'_conv1', data=act1, offset=conv1_offset,
-                num_filter=int(num_filter*0.5), pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=True)
-    else:
-      act1 = mx.sym.QActivation(data=bn1, act_bit=ACT_BIT, name=name + '_relu1', backward_only=True)
-      conv1 = mx.sym.QConvolution_v1(data=act1, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                 no_bias=True, workspace=workspace, name=name + '_conv1', act_bit=ACT_BIT, weight_bit=bit)
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    if not binarize:
-      act2 = Act(data=bn2, act_type='relu', name=name + '_relu2')
-      if not dcn:
-          conv2 = Conv(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation),
-                                     no_bias=True, workspace=workspace, name=name + '_conv2')
-      else:
-          conv2_offset = mx.symbol.Convolution(name=name+'_conv2_offset', data = act2,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-          conv2 = mx.contrib.symbol.DeformableConvolution(name=name+'_conv2', data=act2, offset=conv2_offset,
-                num_filter=int(num_filter*0.25), pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=True)
-    else:
-      act2 = mx.sym.QActivation(data=bn2, act_bit=ACT_BIT, name=name + '_relu2', backward_only=True)
-      conv2 = mx.sym.QConvolution_v1(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                 no_bias=True, workspace=workspace, name=name + '_conv2', act_bit=ACT_BIT, weight_bit=bit)
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    if not binarize:
-      act3 = Act(data=bn3, act_type='relu', name=name + '_relu3')
-      if not dcn:
-          conv3 = Conv(data=act3, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(dilation,dilation), dilate=(dilation,dilation), 
-                  no_bias=True, workspace=workspace, name=name + '_conv3')
-      else:
-          conv3_offset = mx.symbol.Convolution(name=name+'_conv3_offset', data = act3,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-          conv3 = mx.contrib.symbol.DeformableConvolution(name=name+'_conv3', data=act3, offset=conv3_offset,
-                num_filter=int(num_filter*0.25), pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=True)
-    else:
-      act3 = mx.sym.QActivation(data=bn3, act_bit=ACT_BIT, name=name + '_relu3', backward_only=True)
-      conv3 = mx.sym.QConvolution_v1(data=act3, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                 no_bias=True, workspace=workspace, name=name + '_conv3', act_bit=ACT_BIT, weight_bit=bit)
-    conv4 = mx.symbol.Concat(*[conv1, conv2, conv3])
-    if binarize:
-      conv4 = mx.sym.BatchNorm(data=conv4, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-    if dim_match:
-        shortcut = data
-    else:
-        if not binarize:
-          shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                          workspace=workspace, name=name+'_sc')
-        else:
-          #assert(False)
-          shortcut = mx.sym.QConvolution_v1(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_sc', act_bit=ACT_BIT, weight_bit=bit)
-          shortcut = mx.sym.BatchNorm(data=shortcut, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc_bn')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return conv4 + shortcut
-    #return bn4 + shortcut
-    #return act4 + shortcut
-
-
-def block17(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}, name=''):
-    tower_conv = ConvFactory(net, 192, (1, 1), name=name+'_conv')
-    tower_conv1_0 = ConvFactory(net, 129, (1, 1), name=name+'_conv1_0')
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 160, (1, 7), pad=(1, 2), name=name+'_conv1_1')
-    tower_conv1_2 = ConvFactory(tower_conv1_1, 192, (7, 1), pad=(2, 1), name=name+'_conv1_2')
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False, name=name+'_conv_out')
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-def block35(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}, name=''):
-    M = 1.0
-    tower_conv = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv')
-    tower_conv1_0 = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv1_0')
-    tower_conv1_1 = ConvFactory(tower_conv1_0, int(input_num_channels*0.25*M), (3, 3), pad=(1, 1), name=name+'_conv1_1')
-    tower_conv2_0 = ConvFactory(net, int(input_num_channels*0.25*M), (1, 1), name=name+'_conv2_0')
-    tower_conv2_1 = ConvFactory(tower_conv2_0, int(input_num_channels*0.375*M), (3, 3), pad=(1, 1), name=name+'_conv2_1')
-    tower_conv2_2 = ConvFactory(tower_conv2_1, int(input_num_channels*0.5*M), (3, 3), pad=(1, 1), name=name+'_conv2_2')
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False, name=name+'_conv_out')
-
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-def conv_inception(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    assert not binarize
-    if stride[0]>1 or not dim_match:
-        return conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs)
-    conv4 = block35(data, num_filter, name=name+'_block35')
-    return conv4
-
-def conv_cab(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs):
-    if stride[0]>1 or not dim_match:
-        return conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilate, **kwargs)
-    cab = CAB(data, num_filter, 1, 4, workspace, name, dilate, 1)
-    return cab.get()
-
-def conv_block(data, num_filter, stride, dim_match, name, binarize, dcn, dilate):
-  if config.net_block=='resnet':
-    return conv_resnet(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='inception':
-    return conv_inception(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='hpm':
-    return conv_hpm(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-  elif config.net_block=='cab':
-    return conv_cab(data, num_filter, stride, dim_match, name, binarize, dcn, dilate)
-
-def hourglass(data, nFilters, nModules, n, workspace, name, binarize, dcn):
-  s = 2
-  _dcn = False
-  up1 = data
-  for i in range(nModules):
-    up1 = conv_block(up1, nFilters, (1,1), True, "%s_up1_%d"%(name,i), binarize, _dcn, 1)
-  low1 = mx.sym.Pooling(data=data, kernel=(s, s), stride=(s,s), pad=(0,0), pool_type='max')
-  for i in range(nModules):
-    low1 = conv_block(low1, nFilters, (1,1), True, "%s_low1_%d"%(name,i), binarize, _dcn, 1)
-  if n>1:
-    low2 = hourglass(low1, nFilters, nModules, n-1, workspace, "%s_%d"%(name, n-1), binarize, dcn)
-  else:
-    low2 = low1
-    for i in range(nModules):
-      low2 = conv_block(low2, nFilters, (1,1), True, "%s_low2_%d"%(name,i), binarize, _dcn, 1) #TODO
-  low3 = low2
-  for i in range(nModules):
-    low3 = conv_block(low3, nFilters, (1,1), True, "%s_low3_%d"%(name,i), binarize, _dcn, 1)
-  up2 = mx.symbol.UpSampling(low3, scale=s, sample_type='nearest', workspace=512, name='%s_upsampling_%s'%(name,n), num_args=1)
-  return mx.symbol.add_n(up1, up2)
-
-
-class STA:
-    def __init__(self, data, nFilters, nModules, n, workspace, name):
-        self.data = data
-        self.nFilters = nFilters
-        self.nModules = nModules
-        self.n = n
-        self.workspace = workspace
-        self.name = name
-        self.sym_map = {}
-
-
-    def get_conv(self, data, name, dilate=1, group=1):
-        cab = CAB(data, self.nFilters, self.nModules, 4, self.workspace, name, dilate, group)
-        return cab.get()
-
-    def get_output(self, w, h):
-        #print(w,h)
-        assert w>=1 and w<=config.net_n+1
-        assert h>=1 and h<=config.net_n+1
-        s = 2
-        bn_mom = 0.9
-        key = (w,h)
-        if key in self.sym_map:
-            return self.sym_map[key]
-        ret = None
-        if h==self.n:
-            if w==self.n:
-                ret = self.data,64
-            else:
-                x = self.get_output(w+1, h)
-                body = self.get_conv(x[0], "%s_w%d_h%d_1"%(self.name, w, h))
-                body = mx.sym.Pooling(data=body, kernel=(s, s), stride=(s,s), pad=(0,0), pool_type='max')
-                body = self.get_conv(body, "%s_w%d_h%d_2"%(self.name, w, h))
-                ret = body, x[1]//2
-        else:
-            x = self.get_output(w+1, h+1)
-            y = self.get_output(w, h+1)
-
-            HC = False
-
-            if h%2==1 and h!=w:
-                xbody = lin3(x[0], self.nFilters, self.workspace, "%s_w%d_h%d_x"%(self.name, w, h), 3, self.nFilters, 1)
-                HC = True
-                #xbody = x[0]
-            else:
-                xbody = x[0]
-            if x[1]//y[1]==2:
-                if w>1:
-                    ybody = mx.symbol.Deconvolution(data=y[0], num_filter=self.nFilters, kernel=(s,s), 
-                      stride=(s, s),
-                      name='%s_upsampling_w%d_h%d'%(self.name,w, h),
-                      attr={'lr_mult': '1.0'}, workspace=self.workspace)
-                    ybody = mx.sym.BatchNorm(data=ybody, fix_gamma=False, momentum=bn_mom, eps=2e-5, name="%s_w%d_h%d_y_bn"%(self.name, w, h))
-                    ybody = Act(data=ybody, act_type='relu', name="%s_w%d_h%d_y_act"%(self.name, w, h))
-                else:
-                    if h>=1:
-                        ybody = mx.symbol.UpSampling(y[0], scale=s, sample_type='nearest', workspace=512, name='%s_upsampling_w%d_h%d'%(self.name,w, h), num_args=1)
-                        ybody = self.get_conv(ybody, "%s_w%d_h%d_4"%(self.name, w, h))
-                    else:
-                        ybody = mx.symbol.Deconvolution(data=y[0], num_filter=self.nFilters, kernel=(s,s), 
-                          stride=(s, s),
-                          name='%s_upsampling_w%d_h%d'%(self.name,w, h),
-                          attr={'lr_mult': '1.0'}, workspace=self.workspace)
-                        ybody = mx.sym.BatchNorm(data=ybody, fix_gamma=False, momentum=bn_mom, eps=2e-5, name="%s_w%d_h%d_y_bn"%(self.name, w, h))
-                        ybody = Act(data=ybody, act_type='relu', name="%s_w%d_h%d_y_act"%(self.name, w, h))
-                        ybody = Conv(data=ybody, num_filter=self.nFilters, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                              no_bias=True, name="%s_w%d_h%d_y_conv2"%(self.name, w, h), workspace=self.workspace)
-                        ybody = mx.sym.BatchNorm(data=ybody, fix_gamma=False, momentum=bn_mom, eps=2e-5, name="%s_w%d_h%d_y_bn2"%(self.name, w, h))
-                        ybody = Act(data=ybody, act_type='relu', name="%s_w%d_h%d_y_act2"%(self.name, w, h))
-            else:
-                ybody = self.get_conv(y[0], "%s_w%d_h%d_5"%(self.name, w, h))
-            #if not HC:
-            if config.net_sta==2 and h==3 and w==2:
-              z = self.get_output(w+1, h)
-              zbody = z[0]
-              zbody = mx.sym.Pooling(data=zbody, kernel=(z[1], z[1]), stride=(z[1],z[1]), pad=(0,0), pool_type='avg')
-              body = xbody+ybody
-              body = body/2
-              body = mx.sym.broadcast_mul(body, zbody)
-            else: #sta==1
-              body = xbody+ybody
-              body = body/2
-            ret = body, x[1]
-
-        assert ret is not None
-        self.sym_map[key] = ret
-        return ret
-
-    def get(self):
-        return self.get_output(1, 1)[0]
-
-class SymCoherent:
-  def __init__(self, per_batch_size):
-    self.per_batch_size = per_batch_size
-    self.flip_order = [16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 
-        26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 27, 28, 29, 30, 35, 34, 33, 32, 31, 
-        45, 44, 43, 42, 47, 46, 39, 38, 37, 36, 41, 40, 54, 53, 52, 51, 50, 49, 48, 
-        59, 58, 57, 56, 55, 64, 63, 62, 61, 60, 67, 66, 65]
-
-  def get(self, data):
-    #data.shape[0]==per_batch_size
-    b = self.per_batch_size//2
-    ux = mx.sym.slice_axis(data, axis=0, begin=0, end=b)
-    dx = mx.sym.slice_axis(data, axis=0, begin=b, end=b*2)
-    ux = mx.sym.flip(ux, axis=3)
-    #ux = mx.sym.take(ux, indices = self.flip_order, axis=0)
-    ux_list = []
-    for o in self.flip_order:
-      _ux = mx.sym.slice_axis(ux, axis=1, begin=o, end=o+1)
-      ux_list.append(_ux)
-    ux = mx.sym.concat(*ux_list, dim=1)
-    return ux, dx
-
-def l2_loss(x, y):
-  loss = x-y
-  loss = mx.symbol.smooth_l1(loss, scalar=1.0)
-  #loss = loss*loss
-  loss = mx.symbol.mean(loss)
-  return loss
-
-def ce_loss(x, y):
-  #loss = mx.sym.SoftmaxOutput(data = x, label = y, normalization='valid', multi_output=True)
-  x_max = mx.sym.max(x, axis=[2,3], keepdims=True)
-  x = mx.sym.broadcast_minus(x, x_max)
-  body = mx.sym.exp(x)
-  sums = mx.sym.sum(body, axis=[2,3], keepdims=True)
-  body = mx.sym.broadcast_div(body, sums)
-  loss = mx.sym.log(body)
-  loss = loss*y*-1.0
-  loss = mx.symbol.mean(loss, axis=[1,2,3])
-  #loss = mx.symbol.mean(loss)
-  return loss
-
-def get_symbol(num_classes):
-    m = config.multiplier
-    sFilters = max(int(64*m), 32)
-    mFilters = max(int(128*m), 32)
-    nFilters = int(256*m)
-
-    nModules = 1
-    nStacks = config.net_stacks
-    binarize = config.net_binarize
-    input_size = config.input_img_size
-    label_size = config.output_label_size
-    use_coherent = config.net_coherent
-    use_STA = config.net_sta
-    N = config.net_n
-    DCN = config.net_dcn
-    per_batch_size = config.per_batch_size
-    print('binarize', binarize)
-    print('use_coherent', use_coherent)
-    print('use_STA', use_STA)
-    print('use_N', N)
-    print('use_DCN', DCN)
-    print('per_batch_size', per_batch_size)
-    #assert(label_size==64 or label_size==32)
-    #assert(input_size==128 or input_size==256)
-    coherentor = SymCoherent(per_batch_size)
-    D = input_size // label_size
-    print(input_size, label_size, D)
-    data = mx.sym.Variable(name='data')
-    data = data-127.5
-    data = data*0.0078125
-    gt_label = mx.symbol.Variable(name='softmax_label')
-    losses = []
-    closses = []
-    ref_label = gt_label
-    if D==4:
-      body = Conv(data=data, num_filter=sFilters, kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                              no_bias=True, name="conv0", workspace=workspace)
-    else:
-      body = Conv(data=data, num_filter=sFilters, kernel=(3, 3), stride=(1,1), pad=(1, 1),
-                              no_bias=True, name="conv0", workspace=workspace)
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-    body = Act(data=body, act_type='relu', name='relu0')
-
-    dcn = False
-    body = conv_block(body, mFilters, (1,1), sFilters==mFilters, 'res0', False, dcn, 1)
-
-    body = mx.sym.Pooling(data=body, kernel=(2, 2), stride=(2,2), pad=(0,0), pool_type='max')
-
-    body = conv_block(body, mFilters, (1,1), True, 'res1', False, dcn, 1) #TODO
-    body = conv_block(body, nFilters, (1,1), mFilters==nFilters, 'res2', binarize, dcn, 1) #binarize=True?
-
-    heatmap = None
-
-    for i in range(nStacks):
-      shortcut = body
-      if config.net_sta>0:
-        sta = STA(body, nFilters, nModules, config.net_n+1, workspace, 'sta%d'%(i))
-        body = sta.get()
-      else:
-        body = hourglass(body, nFilters, nModules, config.net_n, workspace, 'stack%d_hg'%(i), binarize, dcn)
-      for j in range(nModules):
-        body = conv_block(body, nFilters, (1,1), True, 'stack%d_unit%d'%(i,j), binarize, dcn, 1)
-      _dcn = True if config.net_dcn>=2 else False
-      ll = ConvFactory(body, nFilters, (1,1), dcn = _dcn, name='stack%d_ll'%(i))
-      _name = "heatmap%d"%(i) if i<nStacks-1 else "heatmap"
-      _dcn = True if config.net_dcn>=2 else False
-      if not _dcn:
-          out = Conv(data=ll, num_filter=num_classes, kernel=(1, 1), stride=(1,1), pad=(0,0),
-                                    name=_name, workspace=workspace)
-      else:
-          out_offset = mx.symbol.Convolution(name=_name+'_offset', data = ll,
-                num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-          out = mx.contrib.symbol.DeformableConvolution(name=_name, data=ll, offset=out_offset,
-                num_filter=num_classes, pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=False)
-          #out = Conv(data=ll, num_filter=num_classes, kernel=(3,3), stride=(1,1), pad=(1,1),
-          #                          name=_name, workspace=workspace)
-      if i==nStacks-1:
-          heatmap = out
-      loss = ce_loss(out, ref_label)
-      #loss = loss/nStacks
-      #loss = l2_loss(out, ref_label)
-      losses.append(loss)
-      if config.net_coherent>0:
-          ux, dx = coherentor.get(out)
-          closs = l2_loss(ux, dx)
-          closs = closs/nStacks
-          closses.append(closs)
-
-      if i<nStacks-1:
-        ll2 = Conv(data=ll, num_filter=nFilters, kernel=(1, 1), stride=(1,1), pad=(0,0),
-                                  name="stack%d_ll2"%(i), workspace=workspace)
-        out2 = Conv(data=out, num_filter=nFilters, kernel=(1, 1), stride=(1,1), pad=(0,0),
-                                  name="stack%d_out2"%(i), workspace=workspace)
-        body = mx.symbol.add_n(shortcut, ll2, out2)
-        _dcn = True if (config.net_dcn==1 or config.net_dcn==3) else False
-        if _dcn:
-            _name = "stack%d_out3" % (i)
-            out3_offset = mx.symbol.Convolution(name=_name+'_offset', data = body,
-                  num_filter=18, pad=(1, 1), kernel=(3, 3), stride=(1, 1))
-            out3 = mx.contrib.symbol.DeformableConvolution(name=_name, data=body, offset=out3_offset,
-                  num_filter=nFilters, pad=(1,1), kernel=(3, 3), num_deformable_group=1, stride=(1, 1), dilate=(1, 1), no_bias=False)
-            body = out3
-
-    pred = mx.symbol.BlockGrad(heatmap)
-    #loss = mx.symbol.add_n(*losses)
-    #loss = mx.symbol.MakeLoss(loss)
-    #syms = [loss]
-    syms = []
-    for loss in losses:
-      loss = mx.symbol.MakeLoss(loss)
-      syms.append(loss)
-    if len(closses)>0:
-        coherent_weight = 0.0001
-        closs = mx.symbol.add_n(*closses)
-        closs = mx.symbol.MakeLoss(closs, grad_scale = coherent_weight)
-        syms.append(closs)
-    syms.append(pred)
-    sym = mx.symbol.Group( syms )
-    return sym
-
-def init_weights(sym, data_shape_dict):
-    #print('in hg')
-    arg_name = sym.list_arguments()
-    aux_name = sym.list_auxiliary_states()
-    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
-    arg_shape_dict = dict(zip(arg_name, arg_shape))
-    aux_shape_dict = dict(zip(aux_name, aux_shape))
-    #print(aux_shape)
-    #print(aux_params)
-    #print(arg_shape_dict)
-    arg_params = {}
-    aux_params = {}
-    for k in arg_shape_dict:
-      v = arg_shape_dict[k]
-      #print(k,v)
-      if k.endswith('offset_weight') or k.endswith('offset_bias'):
-        print('initializing',k)
-        arg_params[k] = mx.nd.zeros(shape = v)
-      elif k.startswith('fc6_'):
-        if k.endswith('_weight'):
-          print('initializing',k)
-          arg_params[k] = mx.random.normal(0, 0.01, shape=v)
-        elif k.endswith('_bias'):
-          print('initializing',k)
-          arg_params[k] = mx.nd.zeros(shape=v)
-      elif k.find('upsampling')>=0:
-        print('initializing upsampling_weight', k)
-        arg_params[k] = mx.nd.zeros(shape=arg_shape_dict[k])
-        init = mx.init.Initializer()
-        init._init_bilinear(k, arg_params[k])
-    return arg_params, aux_params
-
diff --git a/alignment/heatmapReg/test.py b/alignment/heatmapReg/test.py
deleted file mode 100644
index 8c753c3f6..000000000
--- a/alignment/heatmapReg/test.py
+++ /dev/null
@@ -1,93 +0,0 @@
-import argparse
-import cv2
-import sys
-import numpy as np
-import os
-import mxnet as mx
-import datetime
-import img_helper
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'deploy'))
-from mtcnn_detector import MtcnnDetector
-
-
-class Handler:
-  def __init__(self, prefix, epoch, ctx_id=0):
-    print('loading',prefix, epoch)
-    if ctx_id>=0:
-      ctx = mx.gpu(ctx_id)
-    else:
-      ctx = mx.cpu()
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    all_layers = sym.get_internals()
-    sym = all_layers['heatmap_output']
-    image_size = (128, 128)
-    self.image_size = image_size
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    #model = mx.mod.Module(symbol=sym, context=ctx)
-    model.bind(for_training=False, data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    self.model = model
-    mtcnn_path = os.path.join(os.path.dirname(__file__), '..', 'deploy', 'mtcnn-model')
-    self.det_threshold = [0.6,0.7,0.8]
-    self.detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=self.det_threshold)
-  
-  def get(self, img):
-    ret = self.detector.detect_face(img, det_type = 0)
-    if ret is None:
-      return None
-    bbox, points = ret
-    if bbox.shape[0]==0:
-      return None
-    bbox = bbox[0,0:4]
-    points = points[0,:].reshape((2,5)).T
-    M = img_helper.estimate_trans_bbox(bbox, self.image_size[0], s = 2.0)
-    rimg = cv2.warpAffine(img, M, self.image_size, borderValue = 0.0)
-    img = cv2.cvtColor(rimg, cv2.COLOR_BGR2RGB)
-    img = np.transpose(img, (2,0,1)) #3*112*112, RGB
-    input_blob = np.zeros( (1, 3, self.image_size[1], self.image_size[0]),dtype=np.uint8 )
-    input_blob[0] = img
-    ta = datetime.datetime.now()
-    data = mx.nd.array(input_blob)
-    db = mx.io.DataBatch(data=(data,))
-    self.model.forward(db, is_train=False)
-    alabel = self.model.get_outputs()[-1].asnumpy()[0]
-    tb = datetime.datetime.now()
-    print('module time cost', (tb-ta).total_seconds())
-    ret = np.zeros( (alabel.shape[0], 2), dtype=np.float32)
-    for i in range(alabel.shape[0]):
-      a = cv2.resize(alabel[i], (self.image_size[1], self.image_size[0]))
-      ind = np.unravel_index(np.argmax(a, axis=None), a.shape)
-      #ret[i] = (ind[0], ind[1]) #h, w
-      ret[i] = (ind[1], ind[0]) #w, h
-    return ret, M
-
-ctx_id = 4
-img_path = '../deploy/Tom_Hanks_54745.png'
-img = cv2.imread(img_path)
-#img = np.zeros( (128,128,3), dtype=np.uint8 )
-
-handler = Handler('./model/HG', 1, ctx_id)
-for _ in range(10):
-  ta = datetime.datetime.now()
-  landmark, M = handler.get(img)
-  tb = datetime.datetime.now()
-  print('get time cost', (tb-ta).total_seconds())
-#visualize landmark
-IM = cv2.invertAffineTransform(M)
-for i in range(landmark.shape[0]):
-  p = landmark[i]
-  point = np.ones( (3,), dtype=np.float32)
-  point[0:2] = p
-  point = np.dot(IM, point)
-  landmark[i] = point[0:2]
-
-for i in range(landmark.shape[0]):
-  p = landmark[i]
-  point = (int(p[0]), int(p[1]))
-  cv2.circle(img, point, 1, (0, 255, 0), 2)
-
-filename = './landmark_test.png'
-print('writing', filename)
-cv2.imwrite(filename, img)
-
-
diff --git a/alignment/heatmapReg/test_rec_nme.py b/alignment/heatmapReg/test_rec_nme.py
deleted file mode 100644
index af1197178..000000000
--- a/alignment/heatmapReg/test_rec_nme.py
+++ /dev/null
@@ -1,65 +0,0 @@
-import argparse
-import cv2
-import sys
-import numpy as np
-import os
-import mxnet as mx
-import datetime
-import img_helper
-from config import config
-from data import FaceSegIter
-from metric import LossValueMetric, NMEMetric
-
-parser = argparse.ArgumentParser(description='test nme on rec data')
-# general
-parser.add_argument('--rec', default='./data_2d/ibug.rec', help='rec data path')
-parser.add_argument('--prefix', default='', help='model prefix')
-parser.add_argument('--epoch', type=int, default=1, help='model epoch')
-parser.add_argument('--gpu', type=int, default=0, help='')
-parser.add_argument('--landmark-type', default='2d', help='')
-parser.add_argument('--image-size', type=int, default=128, help='')
-args = parser.parse_args()
-
-rec_path = args.rec
-ctx_id = args.gpu
-prefix = args.prefix
-epoch = args.epoch
-image_size = (args.image_size, args.image_size)
-config.landmark_type = args.landmark_type
-config.input_img_size = image_size[0]
-
-if ctx_id>=0:
-  ctx = mx.gpu(ctx_id)
-else:
-  ctx = mx.cpu()
-sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-all_layers = sym.get_internals()
-sym = all_layers['heatmap_output']
-#model = mx.mod.Module(symbol=sym, context=ctx, data_names=['data'], label_names=['softmax_label'])
-model = mx.mod.Module(symbol=sym, context=ctx, data_names=['data'], label_names=None)
-#model = mx.mod.Module(symbol=sym, context=ctx)
-model.bind(for_training=False, data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-model.set_params(arg_params, aux_params)
-
-val_iter = FaceSegIter(path_imgrec = rec_path,
-  batch_size = 1,
-  aug_level = 0,
-  )
-_metric = NMEMetric()
-#val_metric = mx.metric.create(_metric)
-#val_metric.reset()
-#val_iter.reset()
-nme = []
-for i, eval_batch in enumerate(val_iter):
-  if i%10==0:
-    print('processing', i)
-  #print(eval_batch.data[0].shape, eval_batch.label[0].shape)
-  batch_data = mx.io.DataBatch(eval_batch.data)
-  model.forward(batch_data, is_train=False)
-  #model.update_metric(val_metric, eval_batch.label, True)
-  pred_label = model.get_outputs()[-1].asnumpy()
-  label = eval_batch.label[0].asnumpy()
-  _nme = _metric.cal_nme(label, pred_label)
-  nme.append(_nme)
-print(np.mean(nme))
-
diff --git a/alignment/heatmapReg/train.py b/alignment/heatmapReg/train.py
deleted file mode 100644
index c92dcf3a7..000000000
--- a/alignment/heatmapReg/train.py
+++ /dev/null
@@ -1,203 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import logging
-import argparse
-from data import FaceSegIter
-import mxnet as mx
-import mxnet.optimizer as optimizer
-import numpy as np
-import os
-import sys
-import math
-import random
-import cv2
-from config import config, default, generate_config
-from optimizer import ONadam
-from metric import LossValueMetric, NMEMetric
-sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
-import sym_heatmap
-#import sym_fc
-#from symbol import fc
-
-
-args = None
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-def main(args):
-  _seed = 727
-  random.seed(_seed)
-  np.random.seed(_seed)
-  mx.random.seed(_seed)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in range(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  #ctx = [mx.gpu(0)]
-  args.ctx_num = len(ctx)
-
-  args.batch_size = args.per_batch_size*args.ctx_num
-  config.per_batch_size = args.per_batch_size
-
-
-
-  print('Call with', args, config)
-  train_iter = FaceSegIter(path_imgrec = os.path.join(config.dataset_path, 'train.rec'),
-      batch_size = args.batch_size,
-      per_batch_size = args.per_batch_size,
-      aug_level = 1,
-      exf = args.exf,
-      args = args,
-      )
-
-  data_shape = train_iter.get_data_shape()
-  #label_shape = train_iter.get_label_shape()
-  sym = sym_heatmap.get_symbol(num_classes=config.num_classes)
-  if len(args.pretrained)==0:
-      #data_shape_dict = {'data' : (args.per_batch_size,)+data_shape, 'softmax_label' : (args.per_batch_size,)+label_shape}
-      data_shape_dict = train_iter.get_shape_dict()
-      arg_params, aux_params = sym_heatmap.init_weights(sym, data_shape_dict)
-  else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      _, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      #sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-
-  model = mx.mod.Module(
-      context       = ctx,
-      symbol        = sym,
-      label_names   = train_iter.get_label_names(),
-  )
-  #lr = 1.0e-3
-  #lr = 2.5e-4
-  _rescale_grad = 1.0/args.ctx_num
-  #_rescale_grad = 1.0/args.batch_size
-  #lr = args.lr
-  #opt = optimizer.Nadam(learning_rate=args.lr, wd=args.wd, rescale_grad=_rescale_grad, clip_gradient=5.0)
-  if args.optimizer=='onadam':
-    opt = ONadam(learning_rate=args.lr, wd=args.wd, rescale_grad=_rescale_grad, clip_gradient=5.0)
-  elif args.optimizer=='nadam':
-    opt = optimizer.Nadam(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  elif args.optimizer=='rmsprop':
-    opt = optimizer.RMSProp(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  elif args.optimizer=='adam':
-    opt = optimizer.Adam(learning_rate=args.lr, rescale_grad=_rescale_grad)
-  else:
-    opt = optimizer.SGD(learning_rate=args.lr, momentum=0.9, wd=args.wd, rescale_grad=_rescale_grad)
-  initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2)
-  _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
-  _metric = LossValueMetric()
-  #_metric = NMEMetric()
-  #_metric2 = AccMetric()
-  #eval_metrics = [_metric, _metric2]
-  eval_metrics = [_metric]
-  lr_steps = [int(x) for x in args.lr_step.split(',')]
-  print('lr-steps', lr_steps)
-  global_step = [0]
-
-  def val_test():
-    all_layers = sym.get_internals()
-    vsym = all_layers['heatmap_output']
-    vmodel = mx.mod.Module(symbol=vsym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    vmodel.bind(data_shapes=[('data', (args.batch_size,)+data_shape)])
-    arg_params, aux_params = model.get_params()
-    vmodel.set_params(arg_params, aux_params)
-    for target in config.val_targets:
-        _file = os.path.join(config.dataset_path, '%s.rec'%target)
-        if not os.path.exists(_file):
-            continue
-        val_iter = FaceSegIter(path_imgrec = _file,
-          batch_size = args.batch_size,
-          #batch_size = 4,
-          aug_level = 0,
-          args = args,
-          )
-        _metric = NMEMetric()
-        val_metric = mx.metric.create(_metric)
-        val_metric.reset()
-        val_iter.reset()
-        for i, eval_batch in enumerate(val_iter):
-          #print(eval_batch.data[0].shape, eval_batch.label[0].shape)
-          batch_data = mx.io.DataBatch(eval_batch.data)
-          model.forward(batch_data, is_train=False)
-          model.update_metric(val_metric, eval_batch.label)
-        nme_value = val_metric.get_name_value()[0][1]
-        print('[%d][%s]NME: %f'%(global_step[0], target, nme_value))
-  
-  def _batch_callback(param):
-    _cb(param)
-    global_step[0]+=1
-    mbatch = global_step[0]
-    for _lr in lr_steps:
-      if mbatch==_lr:
-        opt.lr *= 0.2
-        print('lr change to', opt.lr)
-        break
-    if mbatch%1000==0:
-      print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-    if mbatch>0 and mbatch%args.verbose==0:
-      val_test()
-      if args.ckpt==1:
-        msave = mbatch//args.verbose
-        print('saving', msave)
-        arg, aux = model.get_params()
-        mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg, aux)
-    if mbatch==lr_steps[-1]:
-      if args.ckpt==2:
-        #msave = mbatch//args.verbose
-        msave = 1
-        print('saving', msave)
-        arg, aux = model.get_params()
-        mx.model.save_checkpoint(args.prefix, msave, model.symbol, arg, aux)
-      sys.exit(0)
-
-  train_iter = mx.io.PrefetchingIter(train_iter)
-
-  model.fit(train_iter,
-      begin_epoch        = 0,
-      num_epoch          = 9999,
-      #eval_data          = val_iter,
-      eval_data          = None,
-      eval_metric        = eval_metrics,
-      kvstore            = 'device',
-      optimizer          = opt,
-      initializer        = initializer,
-      arg_params         = arg_params,
-      aux_params         = aux_params,
-      allow_missing      = True,
-      batch_end_callback = _batch_callback,
-      epoch_end_callback = None,
-      )
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='Train face alignment')
-  # general
-  parser.add_argument('--network', help='network name', default=default.network, type=str)
-  parser.add_argument('--dataset', help='dataset name', default=default.dataset, type=str)
-  args, rest = parser.parse_known_args()
-  generate_config(args.network, args.dataset)
-  parser.add_argument('--prefix', default=default.prefix, help='directory to save model.')
-  parser.add_argument('--pretrained', default=default.pretrained, help='')
-  parser.add_argument('--optimizer', default='nadam', help='')
-  parser.add_argument('--lr', type=float, default=default.lr, help='')
-  parser.add_argument('--wd', type=float, default=default.wd, help='')
-  parser.add_argument('--per-batch-size', type=int, default=default.per_batch_size, help='')
-  parser.add_argument('--lr-step', help='learning rate steps (in epoch)', default=default.lr_step, type=str)
-  parser.add_argument('--ckpt', type=int, default=1, help='')
-  parser.add_argument('--norm', type=int, default=0, help='')
-  parser.add_argument('--exf', type=int, default=1, help='')
-  parser.add_argument('--frequent', type=int, default=default.frequent, help='')
-  parser.add_argument('--verbose', type=int, default=default.verbose, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/alignment/synthetics/README.md b/alignment/synthetics/README.md
new file mode 100644
index 000000000..698cc3ee7
--- /dev/null
+++ b/alignment/synthetics/README.md
@@ -0,0 +1,63 @@
+# Introduction
+
+We provide training and testing tools on synthetics data.
+
+
+## Dataset
+
+### Training dataset
+
+Download `Face Synthetics dataset` from [https://github.com/microsoft/FaceSynthetics](https://github.com/microsoft/FaceSynthetics) and put it somewhere.
+
+<div align="left">
+  <img src="https://github.com/microsoft/FaceSynthetics/raw/main/docs/img/dataset_samples_2.jpg" width="640"/>
+</div>
+<br/>
+
+Then use [tools/prepare_synthetics.py](tools/prepare_synthetics.py) for training data preparation.
+
+
+### Testing dataset
+
+[300-W](https://ibug.doc.ic.ac.uk/resources/300-W/)
+
+
+## Pretrained Model
+
+[ResNet50d](https://drive.google.com/file/d/1kNP7qEl3AYNbaHFUg_ZiyRB1CtfDWXR4/view?usp=sharing)
+
+
+## Train and Test
+
+### Prerequisites
+
+- pytorch_lightning
+- timm
+- albumentations
+
+### Training
+
+`` python -u trainer_synthetics.py ``
+
+which uses `resnet50d` as backbone by default, please check the [code](trainer_synthetics.py) for detail.
+
+### Testing
+
+Please check [test_synthetics.py](test_synthetics.py) for detail.
+
+
+## Result Visualization(3D 68 Keypoints)
+
+<div align="left">
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/image_008_1.jpg?raw=true" width="320"/>
+</div>
+
+<div align="left">
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/image_017_1.jpg?raw=true" width="320"/>
+</div>
+
+<div align="left">
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/alignment/images/image_039.jpg?raw=true" width="320"/>
+</div>
+
+
diff --git a/alignment/synthetics/datasets/augs.py b/alignment/synthetics/datasets/augs.py
new file mode 100644
index 000000000..08f47fccd
--- /dev/null
+++ b/alignment/synthetics/datasets/augs.py
@@ -0,0 +1,40 @@
+import numpy as np
+import albumentations as A
+from albumentations.core.transforms_interface import ImageOnlyTransform
+
+class RectangleBorderAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            fill_value = 0,
+            limit = 0.3,
+            always_apply=False,
+            p=1.0,
+            ):
+        super(RectangleBorderAugmentation, self).__init__(always_apply, p)
+        assert limit>0.0 and limit<1.0
+        self.fill_value = 0
+        self.limit = limit
+
+
+    def apply(self, image, border_size_limit, **params):
+        assert len(border_size_limit)==4
+        border_size = border_size_limit.copy()
+        border_size[0] *= image.shape[1]
+        border_size[2] *= image.shape[1]
+        border_size[1] *= image.shape[0]
+        border_size[3] *= image.shape[0]
+        border_size = border_size.astype(np.int)
+        image[:,:border_size[0],:] = self.fill_value
+        image[:border_size[1],:,:] = self.fill_value
+        image[:,-border_size[2]:,:] = self.fill_value
+        image[-border_size[3]:,:,:] = self.fill_value
+        return image
+
+    def get_params(self):
+        border_size_limit = np.random.uniform(0.0, self.limit, size=4)
+        return {'border_size_limit': border_size_limit}
+
+    def get_transform_init_args_names(self):
+        return ('fill_value', 'limit')
+
diff --git a/alignment/synthetics/datasets/dataset_synthetics.py b/alignment/synthetics/datasets/dataset_synthetics.py
new file mode 100644
index 000000000..eed440c0c
--- /dev/null
+++ b/alignment/synthetics/datasets/dataset_synthetics.py
@@ -0,0 +1,163 @@
+import os
+import os.path as osp
+import queue as Queue
+import pickle
+import threading
+import logging
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+import cv2
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+from .augs import RectangleBorderAugmentation
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank,
+                                                 non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+
+class FaceDataset(Dataset):
+    def __init__(self, root_dir, is_train):
+        super(FaceDataset, self).__init__()
+
+        #self.local_rank = local_rank
+        self.is_train = is_train
+        self.input_size = 256
+        self.num_kps = 68
+        transform_list = []
+        if is_train:
+            transform_list += \
+                [
+                    A.ColorJitter(brightness=0.8, contrast=0.5, p=0.5),
+                    A.ToGray(p=0.1),
+                    A.ISONoise(p=0.1),
+                    A.MedianBlur(blur_limit=(1,7), p=0.1),
+                    A.GaussianBlur(blur_limit=(1,7), p=0.1),
+                    A.MotionBlur(blur_limit=(5,12), p=0.1),
+                    A.ImageCompression(quality_lower=50, quality_upper=90, p=0.05),
+                    A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.2, rotate_limit=40, interpolation=cv2.INTER_LINEAR, 
+                        border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.8),
+                    A.HorizontalFlip(p=0.5),
+                    RectangleBorderAugmentation(limit=0.33, fill_value=0, p=0.2),
+                ]
+        transform_list += \
+            [
+                A.geometric.resize.Resize(self.input_size, self.input_size, interpolation=cv2.INTER_LINEAR, always_apply=True),
+                A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+                ToTensorV2(),
+            ]
+        self.transform = A.ReplayCompose(
+            transform_list,
+            keypoint_params=A.KeypointParams(format='xy', remove_invisible=False)
+        )
+        self.root_dir = root_dir
+        with open(osp.join(root_dir, 'annot.pkl'), 'rb') as f:
+            annot = pickle.load(f)
+            self.X, self.Y = annot
+        train_size = int(len(self.X)*0.99)
+        
+        if is_train:
+            self.X = self.X[:train_size]
+            self.Y = self.Y[:train_size]
+        else:
+            self.X = self.X[train_size:]
+            self.Y = self.Y[train_size:]
+        #if local_rank==0:
+        #    logging.info('data_transform_list:%s'%transform_list)
+        flip_parts = ([1, 17], [2, 16], [3, 15], [4, 14], [5, 13], [6, 12], [7, 11], [8, 10],
+            [18, 27], [19, 26], [20, 25], [21, 24], [22, 23],
+            [32, 36], [33, 35],
+            [37, 46], [38, 45], [39, 44], [40, 43], [41, 48], [42, 47],
+            [49, 55], [50, 54], [51, 53], [62, 64], [61, 65], [68, 66], [59, 57], [60, 56])
+        self.flip_order = np.arange(self.num_kps)
+        for pair in flip_parts:
+            self.flip_order[pair[1]-1] = pair[0]-1
+            self.flip_order[pair[0]-1] = pair[1]-1
+        logging.info('len:%d'%len(self.X))
+        print('!!!len:%d'%len(self.X))
+
+    def __getitem__(self, index):
+        x = self.X[index]
+        y = self.Y[index]
+        image_path = os.path.join(self.root_dir, x)
+        img = cv2.imread(image_path)[:,:,::-1]
+        label = y
+        if self.transform is not None:
+            t = self.transform(image=img, keypoints=label)
+            flipped = False
+            for trans in t["replay"]["transforms"]:
+                if trans["__class_fullname__"].endswith('HorizontalFlip'):
+                    if trans["applied"]:
+                        flipped = True
+            img = t['image']
+            label = t['keypoints']
+            label = np.array(label, dtype=np.float32)
+            #print(img.shape)
+            if flipped:
+                #label[:, 0] = self.input_size - 1 - label[:, 0]  #already applied in horizantal flip aug
+                label = label[self.flip_order,:]
+            label /= (self.input_size/2)
+            label -= 1.0
+            label = label.flatten()
+            label = torch.tensor(label, dtype=torch.float32)
+        return img, label
+
+    def __len__(self):
+        return len(self.X)
+
diff --git a/alignment/synthetics/test_synthetics.py b/alignment/synthetics/test_synthetics.py
new file mode 100644
index 000000000..e1a46f810
--- /dev/null
+++ b/alignment/synthetics/test_synthetics.py
@@ -0,0 +1,104 @@
+
+from trainer_synthetics import FaceSynthetics
+import sys
+import glob
+import torch
+import os
+import numpy as np
+import cv2
+import os.path as osp
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.utils import face_align
+
+flip_parts = ([1, 17], [2, 16], [3, 15], [4, 14], [5, 13], [6, 12], [7, 11], [8, 10],
+    [18, 27], [19, 26], [20, 25], [21, 24], [22, 23],
+    [32, 36], [33, 35],
+    [37, 46], [38, 45], [39, 44], [40, 43], [41, 48], [42, 47],
+    [49, 55], [50, 54], [51, 53], [62, 64], [61, 65], [68, 66], [59, 57], [60, 56])
+
+app = FaceAnalysis()
+app.prepare(ctx_id=0, det_size=(224, 224))
+input_size = 256
+USE_FLIP = False
+
+root = 'data/300W/Validation'
+output_dir = 'outputs/'
+
+if not osp.exists(output_dir):
+    os.makedirs(output_dir)
+
+outf = open(osp.join(output_dir, 'pred.txt'), 'w')
+
+model = FaceSynthetics.load_from_checkpoint(sys.argv[1]).cuda()
+model.eval()
+for line in open(osp.join(root, '300W_validation.txt'), 'r'):
+    line = line.strip().split()
+    img_path = osp.join(root, line[0])
+    gt = line[1:]
+    #print(len(gt))
+    name = img_path.split('/')[-1]
+    img = cv2.imread(img_path)
+    dimg = img.copy()
+    faces = app.get(img, max_num=1)
+    if len(faces)!=1:
+        continue
+    bbox = faces[0].bbox
+    w, h = (bbox[2] - bbox[0]), (bbox[3] - bbox[1])
+    center = (bbox[2] + bbox[0]) / 2, (bbox[3] + bbox[1]) / 2
+    rotate = 0
+    _scale = input_size  / (max(w, h)*1.5)
+    aimg, M = face_align.transform(img, center, input_size, _scale, rotate)
+    #cv2.imwrite("outputs/a_%s"%name, aimg)
+    aimg = cv2.cvtColor(aimg, cv2.COLOR_BGR2RGB)
+    kps = None
+    flips = [0, 1] if USE_FLIP else [0]
+    for flip in flips:
+        input = aimg.copy()
+        if flip:
+            input = input[:,::-1,:].copy()
+        input = np.transpose(input, (2, 0, 1))
+        input = np.expand_dims(input, 0)
+        imgs = torch.Tensor(input).cuda()
+        imgs.div_(255).sub_(0.5).div_(0.5)
+        pred = model(imgs).detach().cpu().numpy().flatten().reshape( (-1, 2) )
+        pred[:, 0:2] += 1
+        pred[:, 0:2] *= (input_size // 2)
+        if flip:
+            pred_flip = pred.copy()
+            pred_flip[:, 0] = input_size - 1 - pred_flip[:, 0] 
+            for pair in flip_parts:
+                tmp = pred_flip[pair[0] - 1, :].copy()
+                pred_flip[pair[0] - 1, :] = pred_flip[pair[1] - 1, :]
+                pred_flip[pair[1] - 1, :] = tmp
+            pred = pred_flip
+        if kps is None:
+            kps = pred
+        else:
+            kps += pred
+            kps /= 2.0
+    #print(pred.shape)
+
+    IM = cv2.invertAffineTransform(M)
+    kps = face_align.trans_points(kps, IM)
+    outf.write(line[0])
+    outf.write(' ')
+    outf.write(' '.join(["%.5f"%x for x in kps.flatten()]))
+    outf.write("\n")
+    box = bbox.astype(np.int)
+    color = (0, 0, 255)
+    cv2.rectangle(dimg, (box[0], box[1]), (box[2], box[3]), color, 2)
+    kps = kps.astype(np.int)
+    #print(landmark.shape)
+    for l in range(kps.shape[0]):
+        color = (0, 0, 255)
+        cv2.circle(dimg, (kps[l][0], kps[l][1]), 1, color, 2)
+
+    cv2.imwrite("outputs/%s"%name, dimg)
+
+    #ret = np.argmax(feat)
+    #print(feat)
+    #outf.write("%s %.4f %.4f %.4f\n"%(line[0], feat[0], feat[1], feat[2]))
+
+outf.close()
+
diff --git a/alignment/synthetics/tools/prepare_synthetics.py b/alignment/synthetics/tools/prepare_synthetics.py
new file mode 100644
index 000000000..5e977d611
--- /dev/null
+++ b/alignment/synthetics/tools/prepare_synthetics.py
@@ -0,0 +1,70 @@
+
+import sys
+import glob
+import torch
+import pickle
+import os
+import numpy as np
+import cv2
+import os.path as osp
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.utils import face_align
+
+app = FaceAnalysis()
+app.prepare(ctx_id=0, det_size=(224, 224))
+output_size = 384
+
+input_dir = '/root/codebase/FaceSynthetics'
+output_dir = 'data/synthetics'
+
+if not osp.exists(output_dir):
+    os.makedirs(output_dir)
+
+X = []
+Y = []
+
+for i in range(0, 100000):
+    if i%1000==0:
+        print('loading', i)
+    x = "%06d.png"%i
+    img_path = osp.join(input_dir, x)
+    img = cv2.imread(img_path)
+    dimg = img.copy()
+    ylines = open(osp.join(input_dir, "%06d_ldmks.txt"%i)).readlines()
+    ylines = ylines[:68]
+    y = []
+    for yline in ylines:
+        lmk = [float(x) for x in yline.strip().split()]
+        y.append( tuple(lmk) )
+    pred = np.array(y)
+    faces = app.get(img, max_num=1)
+    if len(faces)!=1:
+        continue
+    bbox = faces[0].bbox
+    w, h = (bbox[2] - bbox[0]), (bbox[3] - bbox[1])
+    center = (bbox[2] + bbox[0]) / 2, (bbox[3] + bbox[1]) / 2
+    rotate = 0
+    _scale = output_size  / (max(w, h)*1.5)
+    aimg, M = face_align.transform(dimg, center, output_size, _scale, rotate)
+    pred = face_align.trans_points(pred, M)
+    #box = bbox.astype(np.int)
+    #color = (0, 0, 255)
+    #cv2.rectangle(dimg, (box[0], box[1]), (box[2], box[3]), color, 2)
+
+    #kps = pred.astype(np.int)
+    #for l in range(kps.shape[0]):
+    #    color = (0, 0, 255)
+    #    cv2.circle(aimg, (kps[l][0], kps[l][1]), 1, color, 2)
+    x = x.replace('png', 'jpg')
+    X.append(x)
+    y = []
+    for k in range(pred.shape[0]):
+        y.append( (pred[k][0], pred[k][1]) )
+    Y.append(y)
+    cv2.imwrite("%s/%s"%(output_dir, x), aimg)
+
+
+with open(osp.join(output_dir, 'annot.pkl'), 'wb') as pfile:
+    pickle.dump((X, Y), pfile, protocol=pickle.HIGHEST_PROTOCOL)
+
diff --git a/alignment/synthetics/trainer_synthetics.py b/alignment/synthetics/trainer_synthetics.py
new file mode 100644
index 000000000..5de7743db
--- /dev/null
+++ b/alignment/synthetics/trainer_synthetics.py
@@ -0,0 +1,140 @@
+from argparse import ArgumentParser
+
+import os
+import os.path as osp
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+import pytorch_lightning as pl
+from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import LearningRateMonitor
+from pytorch_lightning.loggers import TensorBoardLogger
+import timm
+from datasets.dataset_synthetics import FaceDataset, DataLoaderX
+
+
+class FaceSynthetics(pl.LightningModule):
+    def __init__(self, backbone):
+        super().__init__()
+        self.save_hyperparameters()
+        backbone = timm.create_model(backbone, num_classes=68*2)
+        self.backbone = backbone
+        self.loss = nn.L1Loss(reduction='mean')
+        self.hard_mining = False
+
+    def forward(self, x):
+        # use forward for inference/predictions
+        y = self.backbone(x)
+        return y
+
+    def training_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        if self.hard_mining:
+            loss = torch.abs(y_hat - y) #(B,K)
+            loss = torch.mean(loss, dim=1) #(B,)
+            B = len(loss)
+            S = int(B*0.5)
+            loss, _ = torch.sort(loss, descending=True)
+            loss = loss[:S]
+            loss = torch.mean(loss) * 5.0
+        else:
+            loss = self.loss(y_hat, y) * 5.0
+        self.log('train_loss', loss, on_epoch=True)
+        return loss
+
+    def validation_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        loss = self.loss(y_hat, y)
+        self.log('val_loss', loss, on_step=True)
+
+    def test_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        loss = self.loss(y_hat, y)
+        self.log('test_loss', loss)
+
+    def configure_optimizers(self):
+        #return torch.optim.Adam(self.parameters(), lr=0.0002)
+        opt = torch.optim.SGD(self.parameters(), lr = 0.1, momentum=0.9, weight_decay = 0.0005)
+        def lr_step_func(epoch):
+            return 0.1 ** len([m for m in [15, 25, 28] if m <= epoch])
+        scheduler = torch.optim.lr_scheduler.LambdaLR(
+                optimizer=opt, lr_lambda=lr_step_func)
+        lr_scheduler = {
+                'scheduler': scheduler,
+                'name': 'learning_rate',
+                'interval':'epoch',
+                'frequency': 1}
+        return [opt], [lr_scheduler]
+
+
+
+def cli_main():
+    pl.seed_everything(727)
+
+    # ------------
+    # args
+    # ------------
+    parser = ArgumentParser()
+    parser.add_argument('--backbone', default='resnet50d', type=str)
+    parser.add_argument('--batch_size', default=64, type=int)
+    parser.add_argument('--root', default='data/synthetics', type=str)
+    parser.add_argument('--num-gpus', default=2, type=int)
+    parser.add_argument('--tf32', action='store_true')
+    parser = pl.Trainer.add_argparse_args(parser)
+    args = parser.parse_args()
+
+    if not args.tf32:
+        torch.backends.cuda.matmul.allow_tf32 = False
+        torch.backends.cudnn.allow_tf32 = False
+    else:
+        torch.backends.cuda.matmul.allow_tf32 = True
+        torch.backends.cudnn.allow_tf32 = True
+    torch.backends.cudnn.benchmark = True
+
+    # ------------
+    # data
+    # ------------
+    train_set = FaceDataset(root_dir=args.root, is_train=True)
+    val_set = FaceDataset(root_dir=args.root, is_train=False)
+
+    train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=3, pin_memory=True)
+    val_loader = DataLoader(val_set, batch_size=args.batch_size, shuffle=False)
+
+    # ------------
+    # model
+    # ------------
+    model = FaceSynthetics(backbone=args.backbone)
+    ckpt_path = 'work_dirs/synthetics'
+    if not os.path.exists(ckpt_path):
+        os.makedirs(ckpt_path)
+
+    # ------------
+    # training
+    # ------------
+    checkpoint_callback = ModelCheckpoint(
+            monitor='val_loss',
+            dirpath=ckpt_path,
+            filename='{epoch:02d}-{val_loss:.6f}',
+            save_top_k=10,
+            mode='min',
+            )
+    lr_monitor = LearningRateMonitor(logging_interval='step')
+    trainer = pl.Trainer(
+        gpus = args.num_gpus,
+        accelerator="ddp",
+        benchmark=True,
+		logger=TensorBoardLogger(osp.join(ckpt_path, 'logs')),
+        callbacks=[checkpoint_callback, lr_monitor],
+        check_val_every_n_epoch=1,
+        progress_bar_refresh_rate=1,
+        max_epochs=30,
+    )
+    trainer.fit(model, train_loader, val_loader)
+
+if __name__ == '__main__':
+    cli_main()
+
diff --git a/attribute/README.md b/attribute/README.md
new file mode 100644
index 000000000..a926b1dd2
--- /dev/null
+++ b/attribute/README.md
@@ -0,0 +1,33 @@
+## Face Attribute
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/logo3.jpg" width="320"/>
+</div>
+
+
+## Introduction
+
+These are the face attribute methods of [InsightFace](https://insightface.ai)
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/t1_genderage.jpg" width="600"/>
+</div>
+
+
+
+## Methods
+
+
+Supported methods:
+
+- [x] [Gender_Age](gender_age)
+
+
+
+## Contributing
+
+We appreciate all contributions to improve the face attribute module of InsightFace. 
+
+
diff --git a/attribute/_datasets_/README.md b/attribute/_datasets_/README.md
new file mode 100644
index 000000000..7b0096cf8
--- /dev/null
+++ b/attribute/_datasets_/README.md
@@ -0,0 +1,15 @@
+# Face Attribute Datasets
+
+(Updating)
+
+## Training Datasets
+
+### CelebA
+
+https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
+
+
+
+## Test Datasets
+
+
diff --git a/attribute/gender_age/test.py b/attribute/gender_age/test.py
new file mode 100644
index 000000000..a92b21645
--- /dev/null
+++ b/attribute/gender_age/test.py
@@ -0,0 +1,24 @@
+import argparse
+import cv2
+import sys
+import numpy as np
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+
+parser = argparse.ArgumentParser(description='insightface gender-age test')
+# general
+parser.add_argument('--ctx', default=0, type=int, help='ctx id, <0 means using cpu')
+args = parser.parse_args()
+
+app = FaceAnalysis(allowed_modules=['detection', 'genderage'])
+app.prepare(ctx_id=args.ctx, det_size=(640,640))
+
+img = ins_get_image('t1')
+faces = app.get(img)
+assert len(faces)==6
+for face in faces:
+    print(face.bbox)
+    print(face.sex, face.age)
+
diff --git a/benchmarks/train/nvidia_a10.md b/benchmarks/train/nvidia_a10.md
new file mode 100644
index 000000000..f18620bdd
--- /dev/null
+++ b/benchmarks/train/nvidia_a10.md
@@ -0,0 +1,48 @@
+# Training performance report on NVIDIA A10
+
+[NVIDIA A10 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/products/a10-gpu/) 
+
+We can use A10 to train deep learning models by its FP16 and TF32 supports.
+
+
+
+## Test Server Spec
+
+| Key          | Value                                            |
+| ------------ | ------------------------------------------------ |
+| System       | ServMax G408-X2 Rackmountable Server             |
+| CPU          | 2 x Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz    |
+| Memory       | 384GB, 12 x Samsung 32GB DDR4-2933               |
+| GPU          | 8 x NVIDIA A10 22GB                              |
+| Cooling      | 2x Customized GPU Kit for GPU support FAN-1909L2 |
+| Hard Drive   | Intel SSD S4500 1.9TB/SATA/TLC/2.5"              |
+| OS           | Ubuntu 16.04.7 LTS                               |
+| Installation | CUDA 11.1, cuDNN 8.0.5                           |
+| Installation | Python 3.7.10                                    |
+| Installation | PyTorch 1.9.0 (conda)                            |
+
+This server is donated by [AMAX](https://www.amaxchina.com/), many thanks!
+
+
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is in mxnet record format and located on SSD hard drive.
+
+2. Embedding-size are all set to 512.
+
+3. We use a large dataset which contains about 618K identities to simulate real cases.
+
+| Dataset     | Classes | Backbone    | Batch-size | FP16 | TF32 | Samples/sec |
+| ----------- | ------- | ----------- | ---------- | ---- | ---- | ----------- |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | ×    | ~2040       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | √    | ~2255       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | √    | ×    | ~3300       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | √    | √    | ~3360       |
+| WebFace600K | 618K    | IResNet-50  | 2048       | √    | √    | ~3940       |
+| WebFace600K | 618K    | IResNet-100 | 1024       | √    | √    | ~2210       |
+| WebFace600K | 618K    | IResNet-180 | 1024       | √    | √    | ~1410       |
+
+
diff --git a/benchmarks/train/nvidia_a100.md b/benchmarks/train/nvidia_a100.md
new file mode 100644
index 000000000..9af56fcfd
--- /dev/null
+++ b/benchmarks/train/nvidia_a100.md
@@ -0,0 +1,53 @@
+# Training performance report on NVIDIA A100
+
+[NVIDIA A100 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/a100/) 
+
+
+
+## Test Server Spec
+
+| Key          | Value                                            |
+| ------------ | ------------------------------------------------ |
+| System       | ServMax G408-X2 Rackmountable Server             |
+| CPU          | 2 x Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz    |
+| Memory       | 384GB, 12 x Samsung 32GB DDR4-2933               |
+| GPU          | 8 x NVIDIA A100 80GB                             |
+| Cooling      | 2x Customized GPU Kit for GPU support FAN-1909L2 |
+| Hard Drive   | Intel SSD S4500 1.9TB/SATA/TLC/2.5"              |
+| OS           | Ubuntu 16.04.7 LTS                               |
+| Installation | CUDA 11.1, cuDNN 8.0.5                           |
+| Installation | Python 3.7.10                                    |
+| Installation | PyTorch 1.9.0 (conda)                            |
+
+This server is donated by [AMAX](https://www.amaxchina.com/), many thanks!
+
+
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is in mxnet record format and located on SSD hard drive.
+2. Embedding-size are all set to 512.
+3. We use large datasets with about 618K/2M identities to simulate real cases.
+4. We test the 10K batch-size on real dataset to take the full advantage of 80GB memory.
+5. We also test on huge synthetic datasets which include 50M~80M classes.
+
+| Dataset     | Classes | Backbone    | Batch-size | PFC  | FP16 | TF32 | Samples/sec | GPU Mem(GB) |
+| ----------- | ------- | ----------- | ---------- | ---- | ---- | ---- | ----------- | ----------- |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | ×    | ×    | ~3670       | ~18.2       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | ×    | √    | ~4760       | ~15.0       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | √    | ×    | ~5170       | ~10.1       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | √    | √    | ~5400       | ~10.1       |
+| WebFace600K | 618K    | IResNet-50  | 2048       | ×    | √    | √    | ~7780       | ~16.4       |
+| WebFace600K | 618K    | IResNet-50  | 10240      | ×    | √    | √    | ~9400       | ~66.7       |
+| WebFace600K | 618K    | IResNet-100 | 1024       | ×    | √    | √    | ~3700       | ~13.1       |
+| WebFace600K | 618K    | IResNet-180 | 1024       | ×    | √    | √    | ~2380       | ~17.5       |
+| WebFace2M   | 2M      | IResNet-100 | 1024       | ×    | √    | √    | ~3480       | ~20.5       |
+| WebFace2M   | 2M      | IResNet-180 | 1024       | ×    | √    | √    | ~2350       | ~25.0       |
+| WebFace2M   | 2M      | IResNet-300 | 1024       | ×    | √    | √    | ~1541       | ~32.6       |
+| Virtual     | 50M     | IResNet-50  | 1024       | 0.1  | √    | √    | ~2700       | ~54.1       |
+| Virtual     | 70M     | IResNet-50  | 1024       | 0.1  | √    | √    | ~2170       | ~73.7       |
+| Virtual     | 80M     | IResNet-50  | 1024       | 0.1  | √    | √    | ~1080       | ~79.6       |
+
+
diff --git a/benchmarks/train/nvidia_a30.md b/benchmarks/train/nvidia_a30.md
new file mode 100644
index 000000000..2053300a1
--- /dev/null
+++ b/benchmarks/train/nvidia_a30.md
@@ -0,0 +1,52 @@
+# Training performance report on NVIDIA A30
+
+[NVIDIA A30 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/products/a30-gpu/) is the most versatile mainstream
+compute GPU for AI inference and mainstream enterprise
+workloads. 
+
+Besides, we can also use A30 to train deep learning models by its FP16 and TF32 supports.
+
+
+
+## Test Server Spec
+
+| Key          | Value                                            |
+| ------------ | ------------------------------------------------ |
+| System       | ServMax G408-X2 Rackmountable Server             |
+| CPU          | 2 x Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz    |
+| Memory       | 384GB, 12 x Samsung 32GB DDR4-2933               |
+| GPU          | 8 x NVIDIA A30 24GB                              |
+| Cooling      | 2x Customized GPU Kit for GPU support FAN-1909L2 |
+| Hard Drive   | Intel SSD S4500 1.9TB/SATA/TLC/2.5"              |
+| OS           | Ubuntu 16.04.7 LTS                               |
+| Installation | CUDA 11.1, cuDNN 8.0.5                           |
+| Installation | Python 3.7.10                                    |
+| Installation | PyTorch 1.9 (conda)                              |
+
+This server is donated by [AMAX](https://www.amaxchina.com/), many thanks!
+
+
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is in mxnet record format and located on SSD hard drive.
+
+2. Embedding-size are all set to 512.
+
+3. We use a large dataset which contains about 618K identities to simulate real cases.
+
+| Dataset     | Classes | Backbone    | Batch-size | FP16 | TF32 | Samples/sec |
+| ----------- | ------- | ----------- | ---------- | ---- | ---- | ----------- |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | ×    | ~2230       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | ×    | √    | ~3200       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | √    | ×    | ~3940       |
+| WebFace600K | 618K    | IResNet-50  | 1024       | √    | √    | ~4350       |
+| WebFace600K | 618K    | IResNet-50  | 2048       | √    | √    | ~5100       |
+| WebFace600K | 618K    | IResNet-100 | 1024       | √    | √    | ~2810       |
+| WebFace600K | 618K    | IResNet-180 | 1024       | √    | √    | ~1800       |
+
+
+
+
diff --git a/benchmarks/train/nvidia_rtx3080.md b/benchmarks/train/nvidia_rtx3080.md
new file mode 100644
index 000000000..80af02d58
--- /dev/null
+++ b/benchmarks/train/nvidia_rtx3080.md
@@ -0,0 +1,58 @@
+# Training performance report on NVIDIA RTX3080
+
+[GeForce RTX 3080](https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3080-3080ti/) 
+The GeForce RTX™ 3080 Ti and RTX 3080 graphics cards deliver the ultra performance that gamers crave, powered by Ampere—NVIDIA’s 2nd gen RTX architecture. They are built with enhanced RT Cores and Tensor Cores, new streaming multiprocessors, and superfast G6X memory for an amazing gaming experience.
+
+Besides, we can also use GeForce RTX™ 3080 to train deep learning models by its FP16 and TF32 supports.
+
+
+
+## Test Server Spec
+
+| Key          | Value                                             |
+|--------------|---------------------------------------------------|
+| CPU          | 2 x Intel(R) Xeon(R) Platinum 8255C CPU @ 2.50GHz |
+| Memory       | 384GB                                             |
+| GPU          | 8 x GeForce RTX™ 3080                             |
+| OS           | Ubuntu 18.04.4 LTS                                |
+| Installation | CUDA 11.1,                                        |
+| Installation | Python 3.7.3                                      |
+| Installation | PyTorch 1.9.0 (pip)                               |
+
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is SyntheticDataset.
+
+2. Embedding-size are all set to 512.
+
+
+### 1. 2 Million Identities
+
+We use a large dataset which contains about 2 millions identities to simulate real cases.
+
+
+| Dataset    | Classes    | Backbone   | Batch-size | FP16 | Partial FC | Samples/sec |
+|------------|------------|------------|------------|------|------------|-------------|
+| WebFace40M | 2 Millions | IResNet-50 | 512        | ×    | ×          | Fail        |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | x    | √          | ~2190       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | ×          | Fail        |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | √          | ~2620       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | ×    | ×          | Fail        |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | x    | √          | Fail        |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | ×          | Fail        |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | √          | ~3800       |
+
+### 2. 600K Identities
+
+We use a large dataset which contains about 600k identities to simulate real cases.
+
+| Dataset     | Classes | Backbone   | Batch-size | Partial FC | FP16 | Samples/sec |
+|-------------|---------|------------|------------|------------|------|-------------|
+| WebFace600K | 618K    | IResNet-50 | 512        | ×          | ×    | ~2023       |
+| WebFace600K | 618K    | IResNet-50 | 512        | ×          | √    | ~2392       |
+| WebFace600K | 618K    | IResNet-50 | 1024       | ×          | ×    | Fail        |
+| WebFace600K | 618K    | IResNet-50 | 1024       | ×          | √    | Fail        |
+| WebFace600K | 618K    | IResNet-50 | 1024       | √          | √    | ~4010       |
diff --git a/benchmarks/train/nvidia_rtx3090.md b/benchmarks/train/nvidia_rtx3090.md
new file mode 100644
index 000000000..cbc15c84f
--- /dev/null
+++ b/benchmarks/train/nvidia_rtx3090.md
@@ -0,0 +1,57 @@
+# Training performance report on NVIDIA RTX3090
+
+[GEFORCE RTX 3090](https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090/) 
+The GeForce RTX™ 3090 is a big ferocious GPU (BFGPU) with TITAN class performance.
+
+Besides, we can also use GeForce RTX™ 3090 to train deep learning models by its FP16 and TF32 supports.
+
+
+
+## Test Server Spec
+
+| Key          | Value                                             |
+|--------------|---------------------------------------------------|
+| CPU          | 2 x Intel(R) Xeon(R) Platinum 8255C CPU @ 2.50GHz |
+| Memory       | 384GB                                             |
+| GPU          | 8 x GeForce RTX™ 3090                             |
+| OS           | Ubuntu 18.04.4 LTS                                |
+| Installation | CUDA 11.1,                                        |
+| Installation | Python 3.7.3                                      |
+| Installation | PyTorch 1.9.0 (pip)                               |
+
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is SyntheticDataset.
+
+2. Embedding-size are all set to 512.
+
+
+### 1. 2 Million Identities
+
+We use a large dataset which contains about 2 millions identities to simulate real cases.
+
+| Dataset    | Classes    | Backbone   | Batch-size | FP16 | TF32 | Partial FC | Samples/sec |
+|------------|------------|------------|------------|------|------|------------|-------------|
+| WebFace40M | 2 Millions | IResNet-50 | 512        | ×    | ×    | ×          | ~1750       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | ×    | √    | ×          | ~1810       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | √    | ×          | ~2056       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | √    | √          | ~2850       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | √    | ×          | ~2810       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | √    | √          | ~4220       |
+| WebFace40M | 2 Millions | IResNet-50 | 2048       | √    | √    | √          | ~5330       |
+
+
+### 2. 600K Identities
+
+We use a large dataset which contains about 600k identities to simulate real cases.
+
+| Dataset     | Classes | Backbone   | Batch-size | FP16 | Samples/sec |
+|-------------|---------|------------|------------|------|-------------|
+| WebFace600K | 618K    | IResNet-50 | 512        | ×    | ~2220       |
+| WebFace600K | 618K    | IResNet-50 | 512        | √    | ~2610       |
+| WebFace600K | 618K    | IResNet-50 | 1024       | ×    | ~2940       |
+| WebFace600K | 618K    | IResNet-50 | 1024       | √    | ~3790       |
+| WebFace600K | 618K    | IResNet-50 | 2048       | √    | ~4680       |
\ No newline at end of file
diff --git a/benchmarks/train/nvidia_v100.md b/benchmarks/train/nvidia_v100.md
new file mode 100644
index 000000000..6f3aee107
--- /dev/null
+++ b/benchmarks/train/nvidia_v100.md
@@ -0,0 +1,54 @@
+# Training performance report on NVIDIA® V100
+
+[NVIDIA® V100](https://www.nvidia.com/en-us/data-center/v100/) 
+NVIDIA® V100 Tensor Core is the most advanced data center GPU ever built to accelerate AI, high performance computing (HPC), data science and graphics. It’s powered by NVIDIA Volta architecture, comes in 16 and 32GB configurations, and offers the performance of up to 32 CPUs in a single GPU.
+
+Besides, we can also use NVIDIA® V100 to train deep learning models by its FP16 and FP32 supports.
+
+## Test Server Spec
+
+| Key          | Value                                        |
+|--------------|----------------------------------------------|
+| CPU          | 2 x Intel(R) Xeon(R) Gold 6133 CPU @ 2.50GHz |
+| Memory       | 384GB                                        |
+| GPU          | 8 x Tesla V100-SXM2-32GB                     |
+| OS           | Ubuntu 16.04 LTS                             |
+| Installation | CUDA 10.2                                    |
+| Installation | Python 3.7.3                                 |
+| Installation | PyTorch 1.9.0 (pip)                          |
+
+## Experiments on arcface_torch
+
+We report training speed in following table, please also note that:
+
+1. The training dataset is SyntheticDataset.
+
+2. Embedding-size are all set to 512.
+
+### 1. 2 Million Identities
+
+We use a large dataset which contains about 2 millions identities to simulate real cases.
+
+| Dataset    | Classes    | Backbone   | Batch-size | FP16 | Partial FC | Samples/sec |
+|------------|------------|------------|------------|------|------------|-------------|
+| WebFace40M | 2 Millions | IResNet-50 | 512        | ×    | ×          | ~1868       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | x    | √          | ~2712       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | ×          | ~2576       |
+| WebFace40M | 2 Millions | IResNet-50 | 512        | √    | √          | ~4501       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | ×    | ×          | ~1960       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | x    | √          | ~2922       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | ×          | ~2810       |
+| WebFace40M | 2 Millions | IResNet-50 | 1024       | √    | √          | ~5430       |
+| WebFace40M | 2 Millions | IResNet-50 | 2048       | √    | √          | ~6095       |
+
+### 2. 600K Identities
+
+We use a large dataset which contains about 600k identities to simulate real cases.
+
+| Dataset     | Classes | Backbone   | Batch-size | FP16 | Samples/sec |
+|-------------|---------|------------|------------|------|-------------|
+| WebFace600K | 618K    | IResNet-50 | 512        | ×    | ~2430       |
+| WebFace600K | 618K    | IResNet-50 | 512        | √    | ~3889       |
+| WebFace600K | 618K    | IResNet-50 | 1024       | ×    | ~2607       |
+| WebFace600K | 618K    | IResNet-50 | 1024       | √    | ~4322       |
+| WebFace600K | 618K    | IResNet-50 | 2048       | √    | ~4921       |
diff --git a/challenges/README.md b/challenges/README.md
new file mode 100644
index 000000000..80f44e96e
--- /dev/null
+++ b/challenges/README.md
@@ -0,0 +1,29 @@
+## Challenges
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/logo3.jpg" width="240"/>
+</div>
+
+
+## Introduction
+
+These benchmarks are maintained by [InsightFace](https://insightface.ai)
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/thumb_ifrt.png" width="480"/>
+</div>
+
+
+
+## Supported Benchmarks
+- [MFR-Ongoing](mfr) (Ongoing version of iccv21-mfr)
+- [MFR21 (ICCVW'2021)](iccv21-mfr)
+- [LFR19 (ICCVW'2019)](iccv19-lfr)
+
+
+
+
+
+
diff --git a/challenges/cvpr23-fas-wild/README.md b/challenges/cvpr23-fas-wild/README.md
new file mode 100644
index 000000000..49d30ff7f
--- /dev/null
+++ b/challenges/cvpr23-fas-wild/README.md
@@ -0,0 +1,172 @@
+# 4th Face Anti-spoofing Workshop and Challenge@CVPR2023, Wild Track
+
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/faswild_large.png" width="640"/>
+</div>
+
+## Updates
+
+**``2023-05-03``**: We have launched the ongoing version of this challenge. You can start submitting your test results at this [new link](https://codalab.lisn.upsaclay.fr/competitions/12933). If you have not applied for the dataset yet, you need to send an application email to both of ``insightface.challenge@gmail.com`` and ``wangdong@moredian.com``. 
+
+
+
+**``2023-02-15``**: The annotation format in readme is fixed: 
+  ```
+    - e.g: Train/spoof/2D-Display-Phone/000001/000001.txt
+    192 148 (bbox left top)
+    234 203 (bbox right bottom)
+    216 171 (landmark left eye)
+    230 168 (landmark right eye)
+    231 180 (landmark nose)
+    218 190 (landmark left mouth )
+    229 188 (landmark right mouth )
+  ```
+
+## Challenge
+
+We host the WILD track of Face Anti-spoofing Workshop and Challenge@CVPR2023 here. The challenge will officially start together with [4th Face Anti-spoofing Workshop](https://sites.google.com/view/face-anti-spoofing-challenge/welcome/challengecvpr2023). 
+
+[Registration](https://codalab.lisn.upsaclay.fr/competitions/10670) is now open on codalab.
+
+Our competition encompasses over 800K spoof photos and over 500K live photos. In the spoof photos, there are three major categories and 17 subcategories.
+
+### Rules and Regulations
+
+1) Any extra data or pretrained model trained from extra data cannot be used in this challenge.
+
+2) Only one DL model can be used, we can not accept the fusion results from many DL models. The computational cost of a single DL model should be **less than 5G FLOPs**. (FLOPs can be calculated using `ptflops` python library)
+
+3) The top-3 winners are required to submit the code for the entire method, ensuring reproducibility of the results and compliance with all contest rules, otherwise the score will be disqualified.
+
+### Timeline
+
+| Phase | Start Date | End Date            | Intro                                                               |
+|-------|------------|---------------------|---------------------------------------------------------------------|
+| Dev   | 2023-02-13 | 2023-03-15          | evaluate the accuracy on dev set                                    |
+| Test  | 2023-03-15 | 2023-03-28 23:59:59 | evaluate the accuracy on test set, using the threshold from dev set |
+
+
+### Rewards
+
+
+| Rank      | Prize   |
+|-----------|---------|
+| 1st place | $ 3,000 |
+| 2nd place | $ 1,500 |
+| 3rd place | $ 500   |
+
+Sponsors: ``Moredian Technology``
+
+## Evaluation
+
+### Evaluation Criteria
+
+For the performance evaluation, we selected the recently standardized ISO/IEC 30107-3 metrics: Attack Presentation Classification Error Rate (APCER), Normal/Bona Fide Presentation Classification Error Rate (NPCER/BPCER) and Average Classification Error Rate (ACER) as the evaluation metric, in which APCER and BPCER/NPCER are used to measure the error rate of fake or live samples, respectively. The ACER value is used as the final evaluation criterion.
+
+
+### Submission Format
+
+**Phase1**: training dataset is used to train the model (Label: live=1, fake=0). Then the trained model is used to predict the sample scores in dev.txt. Participants can directly submit the predicted score file in codalab system. Note that the order of the samples in dev.txt cannot be changed. The final submitted file contains a total of 140,058 lines. Each line in the file contains two parts separated by a space. The first part is the path of each image in dev.txt and must contain the set name(``dev/``), and the second part is the prediction score given by the model (representing the probability that the sample belongs to the live face, which must be in the range of [0.0, 1.0]). Such as:
+```
+dev/000001.jpg 0.15361                   #Note:  line 1- the first row of dev.txt
+
+......
+
+dev/140058.jpg 0.23394                   #Note:  line 140,058 the last row of dev.txt
+```
+The predicted file should be a ``.txt`` file and compressed into a ZIP file (do not add any folder in the ZIP).
+
+
+**Phase2**: In order to submit results at one time, participants need to combine the dev and test predictions into one file before result submission via codalab system. Note that the order of the samples cannot be changed and the dev sample list needs to be written before the test samples.
+
+The final submission file contains a total of 895,237 lines. Each line in the file contains two parts separated by a space. Such as: 
+```
+dev/000001.jpg 0.15361                   #Note:  line 1- the first row of dev.txt
+
+......
+
+dev/140058.jpg 0.23394                   #Note:  line 140,058 the last row of dev.txt
+test/000001.jpg 0.15361                   #Note:  line 140,059 the first row of test.txt  
+
+......   
+
+test/755179.jpg 0.23394                   #Note:  line 895,237 the last row of test.txt
+```
+The predicted file should be a ``.txt`` file and compressed into a ZIP file (do not add any folder in the ZIP).
+
+## Dataset
+
+### Rules
+
+1. The dataset and its subsets can only be used for academic research purposes.
+2. The user is not allowed to use the dataset or its subsets for any type of commercial purpose.
+3. Any form of usage of the dataset in defamatory, pornographic, or any other unlawful manner, or violation of any applicable regulations or laws is forbidden. We are not responsible for any consequences caused by the above behaviors.
+4. The User is not allowed to distribute, broadcast, or reproduce the dataset or its subsets in any way without official permission.
+5. The user is not allowed to share, transfer, sell or resell the dataset or its subsets to any third party for any purpose. HOWEVER, providing the dataset access to user’s research associates, colleagues or team member is allowed if user’s research associates, colleagues or team member agree to be bound by these usage rules.
+6. All images in this dataset can be used for academic research purposes, BUT, only the approved images of the dataset can be exhibited on user’s publications(including but not limited to research paper, presentations for conferences or educational purposes). The approved images have special marks and are listed in a appendix.
+7. We reserve the right to interpret and amend these rules.
+8. please cite us if the InsightFace Wild Anti-Spoofing Dataset or its subset is used in your research:
+```
+@misc{wang2023wild,
+      title={Wild Face Anti-Spoofing Challenge 2023: Benchmark and Results}, 
+      author={Dong Wang and Jia Guo and Qiqi Shao and Haochi He and Zhian Chen and Chuanbao Xiao and Ajian Liu and Sergio Escalera and Hugo Jair Escalante and Lei Zhen and Jun Wan and Jiankang Deng},
+      year={2023},
+      eprint={2304.05753},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```
+
+
+
+### Download
+
+All users can obtain and use this dataset and its subsets only after signing the [Agreement](https://github.com/nttstar/insightface-resources/raw/master/files/License%20Agreement%20for%20InsightFace%20Wild%20Anti-Spoofing%20Dataset.pdf) and sending it to the official e-mail ``insightface.challenge#gmail.com``.
+
+
+### Dataset Annotations
+
+Please refer to the following table for detailed information on the number of labeled data and examples in the dataset:
+
+#### Spoof Images, Training Set
+
+Training Subset, live/spoof labels and categorization information are given:
+
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/faswild_train_dataset.png" width="1024"/>
+</div>
+
+#### Spoof Images, Dev and Test Sets
+
+Dev and Test Subsets, where dev set is used to select the threshold.
+
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/faswild_devtest_dataset.png" width="1024"/>
+</div>
+
+#### Live Images
+
+There're 205,146 live images in training set, and 51,299/273,126 images in dev and test sets respectively.
+
+
+## Baselines
+
+### Dev
+
+| Backbone | Input Crop | FLOPs | APCER  | BPCER  | ACER   |
+|----------|------------|-------|--------|--------|--------|
+| ResNet18 | 224x224    | 1.8G    | 4.244% | 4.245% | 4.245% |
+
+
+### Test
+
+| Backbone | Input Crop | FLOPs | APCER  | BPCER  | ACER   |
+|----------|------------|-------|--------|--------|--------|
+| ResNet18 | 224x224    | 1.8G    | 6.145% | 8.874% | 7.509% |
+
+
+## Feedback
+
+1) If you have any questions regarding the challenge, kindly open an issue on insightface github. (recommended)
+2) Or you can send an e-mail to ``insightface.challenge#gmail.com``
+
diff --git a/challenges/frvt-impl/README.md b/challenges/frvt-impl/README.md
new file mode 100644
index 000000000..a6c3ea5c6
--- /dev/null
+++ b/challenges/frvt-impl/README.md
@@ -0,0 +1,21 @@
+## FRVT-Implementation
+
+
+<div align="left">
+  <img src="https://insightface.ai/assets/img/custom/logo3.jpg" width="240"/>
+</div>
+
+
+## Introduction
+
+We provide a [FRVT-1:1](https://pages.nist.gov/frvt/html/frvt11.html) implementation example here. One can easily build FRVT-1:1 submission by simply putting insightface trained ONNX models into the codebase.
+
+
+
+Coming soon. 
+
+
+
+
+
+
diff --git a/challenges/iccv19-lfr/README.md b/challenges/iccv19-lfr/README.md
new file mode 100644
index 000000000..12aa53549
--- /dev/null
+++ b/challenges/iccv19-lfr/README.md
@@ -0,0 +1,111 @@
+[The Lightweight Face Recognition Challenge & Workshop](https://ibug.doc.ic.ac.uk/resources/lightweight-face-recognition-challenge-workshop/) will be held in conjunction with the International Conference on Computer Vision (ICCV) 2019, Seoul Korea. 
+
+Please strictly follow the rules. For example, please use the same [method](https://github.com/deepinsight/insightface/blob/master/common/flops_counter.py) for the FLOPs calculation regardless of your training framework is insightface or not.
+
+[Test Server](http://www.insightface-challenge.com/overview) 
+
+**Sponsors:**
+
+The Lightweight Face Recognition Challenge has been supported by 
+
+EPSRC project FACER2VM (EP/N007743/1)
+
+Huawei (5000$)
+
+DeepGlint (3000$)
+
+iQIYI (3000$)
+
+Kingsoft Cloud (3000$)
+
+Pensees (3000$)
+
+Dynamic funding pool: (17000$)
+
+Cash sponsors and gift donations are welcome.
+
+Contact:
+insightface.challenge@gmail.com
+
+**Discussion Group**
+
+*For Chinese:*
+
+![wechat](https://insightface.ai/assets/img/github/lfr19_wechat1.jpg)
+
+*For English:*
+
+(in #lfr2019 channel)
+https://join.slack.com/t/insightface/shared_invite/enQtNjU0NDk2MjYyMTMzLTIzNDEwNmIxMjU5OGYzYzFhMjlkNjlhMTBkNWFiNjU4MTVhNTgzYjQ5ZTZiMGM3MzUyNzQ3OTBhZTg3MzM5M2I
+
+
+**NEWS**
+
+``2019.06.21`` We updated the groundtruth of Glint test dataset.
+
+``2019.06.04`` We will clean the groundtruth on deepglint testset.
+
+``2019.05.21`` Baseline models and training logs available.
+
+``2019.05.16`` The four tracks (deepglint-light, deepglint-large, iQIYI-light, iQIYI-large) will equally share the dynamic funding pool (14000$). From each track, the top 3 players will share the funding pool for 50%, 30% and 20% respectively.
+
+==================
+
+**How To Start:**
+
+**Training:**
+
+1. Download ms1m-retinaface from [baiducloud](https://pan.baidu.com/s/14z7qbi0K8aAYDcgT4ArnWg)(code:4ouw) or [onedrive](https://1drv.ms/u/s!AswpsDO2toNKrjhJhMRoxr-HlECx?e=VSXTmv) and unzip it to `$INSIGHTFACE_ROOT/datasets/`
+2. Go into `$INSIGHTFACE_ROOT/recognition/`
+3. Refer to the `retina` dataset configuration section in `sample_config.py` and copy it as your own configuration file `config.py`.
+4. Start training with `CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --dataset retina --network [your-network] --loss arcface`. It will output the accuracy of lfw, cfp_fp and agedb_30 every 2000 batches by default.
+5. Putting the training dataset on SSD hard disk will achieve better training efficiency.
+
+------------------
+
+**Testing:**
+
+1. Download testdata-image from [baiducloud](https://pan.baidu.com/s/1UKUYsRfVTSzj1tfU3BVFrw) or [dropbox](https://www.dropbox.com/s/r5y6xt754m36rh8/iccv19-challenge-data-v1.zip?dl=0). These face images are all pre-processed and aligned.
+2. To download testdata-video from iQIYI, please visit <http://challenge.ai.iqiyi.com/data-cluster>. You need to download iQIYI-VID-FACE.z01, iQIYI-VID-FACE.z02 and iQIYI-VID-FACE.zip after registration. These face frames are also pre-processed and aligned.
+   1. Unzip: ``zip iQIYI_VID_FACE.zip -s=0 --out iQIYI_VID_FACE_ALL.zip; unzip iQIYI_VID_FACE_ALL.zip``
+   2. We can get a directory named ``iQIYI_VID_FACE`` after decompression. Then, we have to move ``video_filelist.txt`` in testdata-image package to ``iQIYI_VID_FACE/filelist.txt``, to indicate the order of videos in our submission feature file.
+3. To generate image feature submission file: check ``gen_image_feature.py``
+4. To generate video feature submission file: check ``gen_video_feature.py``
+5. Submit binary feature to the right track of the test server.
+
+You can also check the verification performance during training time on LFW,CFP_FP,AgeDB_30 datasets.
+
+------------------
+
+**Evaluation:**
+
+Final ranking is determined by the TAR under 1:1 protocal only, for all valid submissions. 
+
+For image testset, we evaluate the TAR under FAR@e-8 while we choose the TAR under FAR@e-4 for video testset.
+
+------------------
+
+**Baseline:**
+
+1. Network y2(a deeper mobilefacenet): 933M FLOPs. TAR_image: 0.64691, TAR_video: 0.47191
+2. Network r100fc(ResNet100FC-IR): 24G FLOPs. TAR_image: 0.80312, TAR_video: 0.64894
+
+Baseline models download link: [baidu cloud](https://pan.baidu.com/s/1Em0ZFnefSoTsZoTd-9m8Nw)    [dropbox](https://www.dropbox.com/s/yqaziktiv38ehrv/iccv19-baseline-models.zip?dl=0)
+
+Training logs: [baidu cloud](https://pan.baidu.com/s/12rsp-oMzsjTeU6nugEvA9g)   [dropbox](https://www.dropbox.com/s/4ufb9g7n76rfav5/iccv-baseline-log.zip?dl=0)
+
+------------------
+
+**Discussion:**
+
+[https://github.com/deepinsight/insightface/issues/632](https://github.com/deepinsight/insightface/issues/632)
+
+------------------
+
+**Candidate solutions:**
+
+1. Manually design or automatically search different networks/losses.
+2. Use slightly deeper or wider mobile-level networks.
+3. [OctConv](https://arxiv.org/abs/1904.05049), to reduce FLOPs.
+4. [HRNet](https://arxiv.org/abs/1904.04514), for large FLOPs track.
+and so on
diff --git a/challenges/iccv19-lfr/gen_image_feature.py b/challenges/iccv19-lfr/gen_image_feature.py
new file mode 100644
index 000000000..c555227b9
--- /dev/null
+++ b/challenges/iccv19-lfr/gen_image_feature.py
@@ -0,0 +1,157 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+from datetime import datetime
+import os.path
+from easydict import EasyDict as edict
+import time
+import json
+import sys
+import numpy as np
+import importlib
+import itertools
+import argparse
+import struct
+import cv2
+import sklearn
+from sklearn.preprocessing import normalize
+import mxnet as mx
+from mxnet import ndarray as nd
+
+image_shape = None
+net = None
+data_size = 1862120
+emb_size = 0
+use_flip = True
+
+
+def do_flip(data):
+    for idx in range(data.shape[0]):
+        data[idx, :, :] = np.fliplr(data[idx, :, :])
+
+
+def get_feature(buffer):
+    global emb_size
+    if use_flip:
+        input_blob = np.zeros(
+            (len(buffer) * 2, 3, image_shape[1], image_shape[2]))
+    else:
+        input_blob = np.zeros((len(buffer), 3, image_shape[1], image_shape[2]))
+    idx = 0
+    for item in buffer:
+        img = cv2.imread(item)[:, :, ::-1]  #to rgb
+        img = np.transpose(img, (2, 0, 1))
+        attempts = [0, 1] if use_flip else [0]
+        for flipid in attempts:
+            _img = np.copy(img)
+            if flipid == 1:
+                do_flip(_img)
+            input_blob[idx] = _img
+            idx += 1
+    data = mx.nd.array(input_blob)
+    db = mx.io.DataBatch(data=(data, ))
+    net.model.forward(db, is_train=False)
+    _embedding = net.model.get_outputs()[0].asnumpy()
+    if emb_size == 0:
+        emb_size = _embedding.shape[1]
+        print('set emb_size to ', emb_size)
+    embedding = np.zeros((len(buffer), emb_size), dtype=np.float32)
+    if use_flip:
+        embedding1 = _embedding[0::2]
+        embedding2 = _embedding[1::2]
+        embedding = embedding1 + embedding2
+    else:
+        embedding = _embedding
+    embedding = sklearn.preprocessing.normalize(embedding)
+    return embedding
+
+
+def write_bin(path, m):
+    rows, cols = m.shape
+    with open(path, 'wb') as f:
+        f.write(struct.pack('4i', rows, cols, cols * 4, 5))
+        f.write(m.data)
+
+
+def main(args):
+    global image_shape
+    global net
+
+    print(args)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    image_shape = [int(x) for x in args.image_size.split(',')]
+    vec = args.model.split(',')
+    assert len(vec) > 1
+    prefix = vec[0]
+    epoch = int(vec[1])
+    print('loading', prefix, epoch)
+    net = edict()
+    net.ctx = ctx
+    net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(
+        prefix, epoch)
+    #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
+    all_layers = net.sym.get_internals()
+    net.sym = all_layers['fc1_output']
+    net.model = mx.mod.Module(symbol=net.sym,
+                              context=net.ctx,
+                              label_names=None)
+    net.model.bind(data_shapes=[('data', (args.batch_size, 3, image_shape[1],
+                                          image_shape[2]))])
+    net.model.set_params(net.arg_params, net.aux_params)
+
+    features_all = None
+
+    i = 0
+    fstart = 0
+    buffer = []
+    for line in open(os.path.join(args.input, 'filelist.txt'), 'r'):
+        if i % 1000 == 0:
+            print("processing ", i)
+        i += 1
+        line = line.strip()
+        image_path = os.path.join(args.input, line)
+        buffer.append(image_path)
+        if len(buffer) == args.batch_size:
+            embedding = get_feature(buffer)
+            buffer = []
+            fend = fstart + embedding.shape[0]
+            if features_all is None:
+                features_all = np.zeros((data_size, emb_size),
+                                        dtype=np.float32)
+            #print('writing', fstart, fend)
+            features_all[fstart:fend, :] = embedding
+            fstart = fend
+    if len(buffer) > 0:
+        embedding = get_feature(buffer)
+        fend = fstart + embedding.shape[0]
+        print('writing', fstart, fend)
+        features_all[fstart:fend, :] = embedding
+    write_bin(args.output, features_all)
+    #os.system("bypy upload %s"%args.output)
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--batch_size', type=int, help='', default=32)
+    parser.add_argument('--image_size', type=str, help='', default='3,112,112')
+    parser.add_argument('--input', type=str, help='', default='')
+    parser.add_argument('--output', type=str, help='', default='')
+    parser.add_argument('--model', type=str, help='', default='')
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    main(parse_arguments(sys.argv[1:]))
diff --git a/challenges/iccv19-lfr/gen_video_feature.py b/challenges/iccv19-lfr/gen_video_feature.py
new file mode 100644
index 000000000..8b3d7fc29
--- /dev/null
+++ b/challenges/iccv19-lfr/gen_video_feature.py
@@ -0,0 +1,215 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+from datetime import datetime
+import os.path
+from easydict import EasyDict as edict
+import time
+import json
+import glob
+import sys
+import numpy as np
+import importlib
+import itertools
+import argparse
+import struct
+import cv2
+import sklearn
+from sklearn.preprocessing import normalize
+import mxnet as mx
+from mxnet import ndarray as nd
+
+image_shape = None
+net = None
+data_size = 203848
+emb_size = 0
+use_flip = False
+ctx_num = 0
+
+
+def do_flip(data):
+    for idx in range(data.shape[0]):
+        data[idx, :, :] = np.fliplr(data[idx, :, :])
+
+
+def get_feature(buffer):
+    global emb_size
+    input_count = len(buffer)
+    if use_flip:
+        input_count *= 2
+    network_count = input_count
+    if input_count % ctx_num != 0:
+        network_count = (input_count // ctx_num + 1) * ctx_num
+
+    input_blob = np.zeros((network_count, 3, image_shape[1], image_shape[2]),
+                          dtype=np.float32)
+    idx = 0
+    for item in buffer:
+        img = cv2.imread(item)[:, :, ::-1]  #to rgb
+        img = np.transpose(img, (2, 0, 1))
+        attempts = [0, 1] if use_flip else [0]
+        for flipid in attempts:
+            _img = np.copy(img)
+            if flipid == 1:
+                do_flip(_img)
+            input_blob[idx] = _img
+            idx += 1
+    data = mx.nd.array(input_blob)
+    db = mx.io.DataBatch(data=(data, ))
+    net.model.forward(db, is_train=False)
+    _embedding = net.model.get_outputs()[0].asnumpy()
+    _embedding = _embedding[0:input_count]
+    if emb_size == 0:
+        emb_size = _embedding.shape[1]
+        print('set emb_size to ', emb_size)
+    embedding = np.zeros((len(buffer), emb_size), dtype=np.float32)
+    if use_flip:
+        embedding1 = _embedding[0::2]
+        embedding2 = _embedding[1::2]
+        embedding = embedding1 + embedding2
+    else:
+        embedding = _embedding
+    embedding = sklearn.preprocessing.normalize(embedding)
+    return embedding
+
+
+def write_bin(path, m):
+    rows, cols = m.shape
+    with open(path, 'wb') as f:
+        f.write(struct.pack('4i', rows, cols, cols * 4, 5))
+        f.write(m.data)
+
+
+def main(args):
+    global image_shape
+    global net
+    global ctx_num
+
+    print(args)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    ctx_num = len(ctx)
+    image_shape = [int(x) for x in args.image_size.split(',')]
+    vec = args.model.split(',')
+    assert len(vec) > 1
+    prefix = vec[0]
+    epoch = int(vec[1])
+    print('loading', prefix, epoch)
+    net = edict()
+    net.ctx = ctx
+    net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(
+        prefix, epoch)
+    #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
+    all_layers = net.sym.get_internals()
+    net.sym = all_layers['fc1_output']
+    net.model = mx.mod.Module(symbol=net.sym,
+                              context=net.ctx,
+                              label_names=None)
+    net.model.bind(data_shapes=[('data', (args.batch_size, 3, image_shape[1],
+                                          image_shape[2]))])
+    net.model.set_params(net.arg_params, net.aux_params)
+
+    features_all = None
+
+    i = 0
+    filelist = os.path.join(args.input, 'filelist.txt')
+    #print(filelist)
+    buffer_images = []
+    buffer_embedding = np.zeros((0, 0), dtype=np.float32)
+    aggr_nums = []
+    row_idx = 0
+    for line in open(filelist, 'r'):
+        if i % 1000 == 0:
+            print("processing ", i)
+        i += 1
+        #print('stat', i, len(buffer_images), buffer_embedding.shape, aggr_nums, row_idx)
+        videoname = line.strip().split()[0]
+        images = glob.glob("%s/%s/*.jpg" % (args.input, videoname))
+        assert len(images) > 0
+        image_features = []
+        for image_path in images:
+            buffer_images.append(image_path)
+        aggr_nums.append(len(images))
+        while len(buffer_images) >= args.batch_size:
+            embedding = get_feature(buffer_images[0:args.batch_size])
+            buffer_images = buffer_images[args.batch_size:]
+            if buffer_embedding.shape[1] == 0:
+                buffer_embedding = embedding.copy()
+            else:
+                buffer_embedding = np.concatenate(
+                    (buffer_embedding, embedding), axis=0)
+        buffer_idx = 0
+        acount = 0
+        for anum in aggr_nums:
+            if buffer_embedding.shape[0] >= anum + buffer_idx:
+                image_features = buffer_embedding[buffer_idx:buffer_idx + anum]
+                video_feature = np.sum(image_features, axis=0, keepdims=True)
+                video_feature = sklearn.preprocessing.normalize(video_feature)
+                if features_all is None:
+                    features_all = np.zeros(
+                        (data_size, video_feature.shape[1]), dtype=np.float32)
+                #print('write to', row_idx, anum, buffer_embedding.shape)
+                features_all[row_idx] = video_feature.flatten()
+                row_idx += 1
+                buffer_idx += anum
+                acount += 1
+            else:
+                break
+        aggr_nums = aggr_nums[acount:]
+        buffer_embedding = buffer_embedding[buffer_idx:]
+
+    if len(buffer_images) > 0:
+        embedding = get_feature(buffer_images)
+        buffer_images = buffer_images[args.batch_size:]
+        buffer_embedding = np.concatenate((buffer_embedding, embedding),
+                                          axis=0)
+    buffer_idx = 0
+    acount = 0
+    for anum in aggr_nums:
+        assert buffer_embedding.shape[0] >= anum + buffer_idx
+        image_features = buffer_embedding[buffer_idx:buffer_idx + anum]
+        video_feature = np.sum(image_features, axis=0, keepdims=True)
+        video_feature = sklearn.preprocessing.normalize(video_feature)
+        #print('last write to', row_idx, anum, buffer_embedding.shape)
+        features_all[row_idx] = video_feature.flatten()
+        row_idx += 1
+        buffer_idx += anum
+        acount += 1
+
+    aggr_nums = aggr_nums[acount:]
+    buffer_embedding = buffer_embedding[buffer_idx:]
+    assert len(aggr_nums) == 0
+    assert buffer_embedding.shape[0] == 0
+
+    write_bin(args.output, features_all)
+    print(row_idx, features_all.shape)
+    #os.system("bypy upload %s"%args.output)
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--batch_size', type=int, help='', default=32)
+    parser.add_argument('--image_size', type=str, help='', default='3,112,112')
+    parser.add_argument('--input',
+                        type=str,
+                        help='',
+                        default='./testdata-video')
+    parser.add_argument('--output', type=str, help='', default='')
+    parser.add_argument('--model', type=str, help='', default='')
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    main(parse_arguments(sys.argv[1:]))
diff --git a/challenges/iccv21-mfr/README.md b/challenges/iccv21-mfr/README.md
new file mode 100644
index 000000000..286511b42
--- /dev/null
+++ b/challenges/iccv21-mfr/README.md
@@ -0,0 +1,187 @@
+# InsightFace Track of ICCV21-MFR
+
+## NEWS
+
+**``2021-11-30``** [MFR-Ongoing](../mfr) is now available.
+
+**``2021-10-26``** Please send the onnx models to us(insightface.challenge[at]gmail.com) if you want to test the MFR accuracy before our system rebooting(may be in Nov.).
+
+**``2021-10-11``** [Final Leaderboard](https://insightface.ai/mfr21)
+
+**``2021-10-04``** Please fix the public leaderboard scores before ``2021-10-05 20:00(UTC+8 Time)``
+
+**``2021-07-16``**  Implicit batch inference is prohibited. For example, inserting some data-related OPs to onnx graph to enable automatic flip-test is not allowed(or similar ideas). We will check it after submission closed, to ensure fairness.
+
+**``2021-06-17``**  Participants are now ordered in terms of highest scores across two datasets: **TAR@Mask** and **TAR@MR-All**, by the formula of ``0.25 * TAR@Mask + 0.75 * TAR@MR-All``.
+
+
+## Introduction
+
+The Masked Face Recognition Challenge & Workshop(MFR) will be held in conjunction with the International Conference on Computer Vision (ICCV) 2021. 
+
+[Workshop-Homepage](https://ibug.doc.ic.ac.uk/resources/masked-face-recognition-challenge-workshop-iccv-21/). 
+
+There're InsightFace track here and [Webface260M](https://www.face-benchmark.org/challenge.html) track(with larger training set) in this workshop.
+
+[**Challenge Leaderboard**](https://insightface.ai/mfr21)
+
+Submission server link: [http://iccv21-mfr.com/](http://iccv21-mfr.com/)
+
+An alternative submission server for Non-Chinese users: [http://124.156.136.55/](http://124.156.136.55/)
+
+
+### Discussion group:
+
+WeChat:
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/mfr_wechat_group.png" alt="mfr_group" width="360">
+
+QQ Group: 711302608, *answer: github*
+
+Online issue discussion: [https://github.com/deepinsight/insightface/issues/1564](https://github.com/deepinsight/insightface/issues/1564)
+
+## Testsets for insightface track
+
+In this challenge, we will evaluate the accuracy of following testsets:
+
+  * Accuracy between masked and non-masked faces.
+  * Accuracy among children(2~16 years old).
+  * Accuracy of globalised multi-racial benchmarks.
+
+We ensure that there's no overlap between these testsets and public available training datasets, as they are not collected from online celebrities.
+
+Our test datasets mainly comes from [IFRT](../ifrt).
+
+### ``Mask test-set:``
+
+Mask testset contains 6,964 identities, 6,964 masked images and 13,928 non-masked images. There are totally 13,928 positive pairs and 96,983,824 negative pairs.
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrt_mask_sample.jpg" alt="ifrtsample" width="360">
+</details>
+
+### ``Children test-set:``
+
+Children testset contains 14,344 identities and 157,280 images. There are totally 1,773,428 positive pairs and 24,735,067,692 negative pairs.
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrt_children_sample.jpg" alt="ifrtsample" width="360">
+</details>
+
+### ``Multi-racial test-set (MR in short):``
+
+The globalised multi-racial testset contains 242,143 identities and 1,624,305 images.
+
+| Race-Set     | Identities  | Images        |  Positive Pairs   | Negative Pairs        |
+| -------      | ----------  | -----------   |  -----------      | -----------           |
+| African      | 43,874      | 298,010       |  870,091          | 88,808,791,999        |
+| Caucasian    | 103,293     | 697,245       |  2,024,609        | 486,147,868,171       |
+| Indian       | 35,086      | 237,080       |  688,259          | 56,206,001,061        |
+| Asian        | 59,890      | 391,970       |  1,106,078        | 153,638,982,852       |
+| **ALL**      | **242,143** | **1,624,305** |  **4,689,037**    | **2,638,360,419,683** |
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrtsample_blur.jpg" alt="ifrtsample" width="640">
+</details>
+
+## Evaluation Metric
+
+For ``Mask`` set, TAR is measured on mask-to-nonmask 1:1 protocal, with FAR less than 0.0001(e-4).
+
+For ``Children`` set, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.0001(e-4).
+
+For other sets, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6).
+
+Participants are ordered in terms of highest scores across two datasets: **TAR@Mask** and **TAR@MR-All**, by the formula of ``0.25 * TAR@Mask + 0.75 * TAR@MR-All``.
+
+
+## Baselines
+
+
+| Backbone   | Dataset    | Method     | Mask   | Children | African | Caucasian | South Asian | East Asian | MR-All    | size(mb) | infer(ms) | link |
+|------------|------------|------------|--------|----------|---------|-----------|-------------|------------|--------|----------|-----------|-----------|
+| R100  | Casia  | ArcFace  | 26.623 | 30.359   | 39.666  | 53.933    | 47.807      | 21.572     | 42.735 | 248.904  | 7.073     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUJpk8zC61HVN7Kg?e=zE9JDd) |
+| R100  | MS1MV2  | ArcFace  | 65.767 | 60.496   | 79.117  | 87.176    | 85.501      | 55.807     | 80.725 | 248.904  | 7.028     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUTlYEHJCHg3UYM-?e=ihxMpS) |
+| R18  | MS1MV3  | ArcFace | 47.853 | 41.047   | 62.613  | 75.125    | 70.213      | 43.859     | 68.326 | 91.658   | 1.856     | [download](https://1drv.ms/u/s!AswpsDO2toNKrTxlT6w1Jo02yzSh?e=KDhFAA) |
+| R34  | MS1MV3  | ArcFace | 58.723 | 55.834   | 71.644  | 83.291    | 80.084      | 53.712     | 77.365 | 130.245  | 3.054     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT2O5pgyVtwnjeMq?e=16S8LI) |
+| R50  | MS1MV3  | ArcFace | 63.850 | 60.457   | 75.488  | 86.115    | 84.305      | 57.352     | 80.533 | 166.305  | 4.262     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUUWd5i3a5OlFpM_?e=ExBDBN) |
+| R100 | MS1MV3 | ArcFace | 69.091 | 66.864   | 81.083  | 89.040    | 88.082      | 62.193     | 84.312 | 248.590  | 7.031     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUPwyqWvNXUlNd3P?e=pTLw9A) |
+| R18   | Glint360K   | ArcFace | 53.317 | 48.113   | 68.230  | 80.575    | 75.852      | 47.831     | 72.074 | 91.658   | 2.013     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT5ey4lCqFzlpzDd?e=VWP28J) |
+| R34   | Glint360K   | ArcFace | 65.106 | 65.454   | 79.907  | 88.620    | 86.815      | 60.604     | 83.015 | 130.245  | 3.044     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUBcgGkiuUS11Hsd?e=ISGDnP) |
+| R50   | Glint360K   | ArcFace | 70.233 | 69.952   | 85.272  | 91.617    | 90.541      | 66.813     | 87.077 | 166.305  | 4.340     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT8jbvHxjqCY0d08?e=igfdrd) |
+| R100  | Glint360K  | ArcFace | 75.567 | 75.202   | 89.488  | 94.285    | 93.434      | 72.528     | 90.659 | 248.590  | 7.038     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUFgLEIj-mnkb51b?e=vWqy2q) |
+| -       | *Private*     | <div style="width: 50pt">insightface-000 of frvt  | 97.760 | 93.358   | 98.850  | 99.372    | 99.058      | 87.694     | 97.481 | -  | -    |   -  |
+
+
+(MS1M-V2 means MS1M-ArcFace, MS1M-V3 means MS1M-RetinaFace).
+
+Inference time was evaluated on Tesla V100 GPU, using onnxruntime-gpu==1.6.
+
+
+## Rules
+
+1. We have two sub-tracks, determined by the size of training dataset and inference time limitation.
+  * Sub-Track A: Use MS1M-V3 as training set, download: [ref-link](https://github.com/deepinsight/insightface/blob/master/recognition/_datasets_/README.md#ms1m-retinaface), feature length must <= 512, and the inference time must <= 10ms on Tesla V100 GPU.
+  * Sub-Track B: Use Glint360K as training set, download: [ref-link](https://github.com/deepinsight/insightface/blob/master/recognition/_datasets_/README.md#deepglint-181k-ids675m-images-8), feature length must <= 1024, and the inference time must <= 20ms on Tesla V100 GPU.
+2. Training set and testing set are both aligned to 112x112, re-alignment is prohibited.
+3. Mask data-augmentation is allowed, such as [this](../../recognition/_tools_). The applied mask augmentation tool should be reproducible. 
+4. External dataset and pretrained models are both prohibited.
+5. Participants submit onnx model, then get scores by our online evaluation. Test images are invisible.
+6. Matching score is measured by cosine similarity.
+7. Model size must <= 1GB.
+8. The input shape of submission model should equal to 3x112x112 (RGB order).
+9. Online evaluation server uses onnxruntime-gpu==1.6, cuda==10.2, cudnn==8.0.5.
+10. Any float-16 model weights is prohibited, as it will lead to incorrect model size estimiation.
+11. Please use ``onnx_helper.py`` to check whether the model is valid.
+12. Participants are now ordered in terms of highest scores across two datasets: **TAR@Mask** and **TAR@MR-All**, by the formula of ``0.25 * TAR@Mask + 0.75 * TAR@MR-All``.
+13. Top-ranked participants should provide their solutions and codes to ensure their validity after submission closed.
+
+
+## Tutorial 
+
+1. ArcFace-PyTorch (with Partial-FC), [code](../../recognition/arcface_torch), [tutorial-cn](tutorial_pytorch_cn.md)
+2. OneFlow, [code](../../recognition/oneflow_face)
+3. MXNet, [code](../../recognition/arcface_mxnet)
+
+## Submission Guide
+
+1. Participants must package the onnx model for submission using ``zip xxx.zip model.onnx``.
+2. Each participant can submit three times a day at most.
+3. Please sign-up with the real organization name. You can hide the organization name in our system if you like.
+4. You can decide which submission to be displayed on the leaderboard by clicking 'Set Public' button.
+5. Please click 'sign-in' on submission server if find you're not logged in.
+
+Server link: [http://iccv21-mfr.com/](http://iccv21-mfr.com/)
+
+## Timelines
+
+* 1 June - Release of the training data, baseline solutions and testing leader-board
+* 1 October - Stop leader-board submission (11:59 PM Pacific Time)
+* 7 October - Winners notification
+
+
+## Sponsors
+
+(in alphabetical order)
+ 
+
+**[DeepGlint](http://www.deepglint.com/)**
+
+**[Kiwi Tech](http://www.kiwiar.com)**
+
+**[OneFlow](https://www.oneflow.org)**
+
+ **[Zoloz](https://www.zoloz.com)**
+
+
+## Bonus Share
+
+|           | Sub-Track A | Sub-Track B |
+| --------- | ------- | ------- |
+| 1st place | 30%     | 30%     |
+| 2nd place | 15%     | 15%     |
+| 3rd place | 5%      | 5%      |
+
diff --git a/challenges/iccv21-mfr/dataset_mask.py b/challenges/iccv21-mfr/dataset_mask.py
new file mode 100644
index 000000000..b221c95c5
--- /dev/null
+++ b/challenges/iccv21-mfr/dataset_mask.py
@@ -0,0 +1,208 @@
+import numbers
+import os
+import queue as Queue
+import threading
+
+import mxnet as mx
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+import cv2
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+from insightface.app import MaskAugmentation
+
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank,
+                                                 non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+class MXFaceDataset(Dataset):
+    def __init__(self, root_dir, local_rank, aug_modes="brightness=0.1+mask=0.1"):
+        super(MXFaceDataset, self).__init__()
+        default_aug_probs = {
+                'brightness' : 0.2,
+                'blur': 0.1,
+                'mask': 0.1,
+                }
+
+        aug_mode_list = aug_modes.lower().split('+')
+        aug_mode_map = {}
+        for aug_mode_str in aug_mode_list:
+            _aug = aug_mode_str.split('=')
+            aug_key = _aug[0]
+            if len(_aug)>1:
+                aug_prob = float(_aug[1])
+            else:
+                aug_prob = default_aug_probs[aug_key]
+            aug_mode_map[aug_key] = aug_prob
+
+        transform_list = []
+        self.mask_aug = False
+        self.mask_prob = 0.0
+        key = 'mask'
+        if key in aug_mode_map:
+            self.mask_aug = True
+            self.mask_prob = aug_mode_map[key]
+            transform_list.append(
+                MaskAugmentation(mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'], mask_probs=[0.4, 0.4, 0.1, 0.1], h_low=0.33, h_high=0.4, p=self.mask_prob)
+                )
+        if local_rank==0:
+            print('data_transform_list:', transform_list)
+            print('mask:', self.mask_aug, self.mask_prob)
+        key = 'brightness'
+        if key in aug_mode_map:
+            prob = aug_mode_map[key]
+            transform_list.append(
+                A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=prob)
+                )
+        key = 'blur'
+        if key in aug_mode_map:
+            prob = aug_mode_map[key]
+            transform_list.append(
+                A.ImageCompression(quality_lower=30, quality_upper=80, p=prob)
+                )
+            transform_list.append(
+                A.MedianBlur(blur_limit=(1,7), p=prob)
+                )
+            transform_list.append(
+                A.MotionBlur(blur_limit=(5,12), p=prob)
+                )
+        transform_list += \
+            [
+                A.HorizontalFlip(p=0.5),
+                A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+                ToTensorV2(),
+            ]
+        #here, the input for A transform is rgb cv2 img
+        self.transform = A.Compose(
+            transform_list 
+        )
+        self.root_dir = root_dir
+        self.local_rank = local_rank
+        path_imgrec = os.path.join(root_dir, 'train.rec')
+        path_imgidx = os.path.join(root_dir, 'train.idx')
+        self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+        s = self.imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        #print(header)
+        #print(len(self.imgrec.keys))
+        if header.flag > 0:
+            if len(header.label)==2:
+                self.imgidx = np.array(range(1, int(header.label[0])))
+            else:
+                self.imgidx = np.array(list(self.imgrec.keys))
+        else:
+            self.imgidx = np.array(list(self.imgrec.keys))
+        #print('imgidx len:', len(self.imgidx))
+
+    def __getitem__(self, index):
+        idx = self.imgidx[index]
+        s = self.imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        hlabel = header.label
+        #print('hlabel:', hlabel.__class__)
+        sample = mx.image.imdecode(img).asnumpy()
+        if not isinstance(hlabel, numbers.Number):
+            idlabel = hlabel[0]
+        else:
+            idlabel = hlabel
+        label = torch.tensor(idlabel, dtype=torch.long)
+        if self.transform is not None:
+            sample = self.transform(image=sample, hlabel=hlabel)['image']
+        return sample, label
+
+    def __len__(self):
+        return len(self.imgidx)
+
+if __name__ == "__main__":
+    import argparse, cv2, copy
+    parser = argparse.ArgumentParser(description='dataset test')
+    parser.add_argument('--dataset', type=str,  help='dataset path')
+    parser.add_argument('--samples', type=int, default=256, help='')
+    parser.add_argument('--cols', type=int, default=16, help='')
+    args = parser.parse_args()
+    assert args.samples%args.cols==0
+    assert args.cols%2==0
+    samples = args.samples
+    cols = args.cols
+    rows = args.samples // args.cols
+    dataset = MXFaceDataset(root_dir=args.dataset, local_rank=0, aug_modes='mask=1.0')
+    dataset.transform = A.Compose([t for t in dataset.transform if not isinstance(t, (A.Normalize, ToTensorV2))])
+    dataset_0 = copy.deepcopy(dataset)
+    #dataset_0.transform = None
+    dataset_1 = copy.deepcopy(dataset)
+    #dataset_1.transform = A.Compose(
+    #    [
+    #        A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=1.0),
+    #        A.ImageCompression(quality_lower=30, quality_upper=80, p=1.0),
+    #        A.MedianBlur(blur_limit=(1,7), p=1.0),
+    #        A.MotionBlur(blur_limit=(5,12), p=1.0),
+    #        A.Affine(scale=(0.92, 1.08),  translate_percent=(-0.06, 0.06), rotate=(-6, 6), shear=None, interpolation=cv2.INTER_LINEAR, p=1.0),
+    #    ]
+    #)
+    fig = np.zeros( (112*rows, 112*cols, 3), dtype=np.uint8 )
+    for idx in range(samples):
+        if idx%2==0:
+            image, _ = dataset_0[idx//2]
+        else:
+            image, _ = dataset_1[idx//2]
+        row_idx = idx // cols
+        col_idx = idx % cols
+        fig[row_idx*112:(row_idx+1)*112, col_idx*112:(col_idx+1)*112,:] = image[:,:,::-1] # to bgr
+    cv2.imwrite("./datasets.png", fig)
diff --git a/challenges/iccv21-mfr/mxnet_to_ort.py b/challenges/iccv21-mfr/mxnet_to_ort.py
new file mode 100644
index 000000000..16f9f2f6a
--- /dev/null
+++ b/challenges/iccv21-mfr/mxnet_to_ort.py
@@ -0,0 +1,115 @@
+import sys
+import os
+import argparse
+import onnx
+import mxnet as mx
+from onnx import helper
+from onnx import TensorProto
+from onnx import numpy_helper
+
+print('mxnet version:', mx.__version__)
+print('onnx version:', onnx.__version__)
+
+assert mx.__version__ >= '1.8', 'mxnet version should >= 1.8'
+assert onnx.__version__ >= '1.2.1', 'onnx version should >= 1.2.1'
+
+import numpy as np
+from mxnet.contrib import onnx as onnx_mxnet
+
+def create_map(graph_member_list):
+    member_map={}
+    for n in graph_member_list:
+        member_map[n.name]=n
+    return member_map
+
+
+parser = argparse.ArgumentParser(description='convert arcface models to onnx')
+# general
+parser.add_argument('params', default='./r100a/model-0000.params', help='mxnet params to load.')
+parser.add_argument('output', default='./r100a.onnx', help='path to write onnx model.')
+parser.add_argument('--eps', default=1.0e-8, type=float, help='eps for weights.')
+parser.add_argument('--input-shape', default='3,112,112', help='input shape.')
+args = parser.parse_args()
+input_shape = (1,) + tuple( [int(x) for x in args.input_shape.split(',')] )
+
+params_file = args.params
+pos = params_file.rfind('-')
+prefix = params_file[:pos]
+epoch = int(params_file[pos+1:pos+5])
+sym_file = prefix + "-symbol.json"
+assert os.path.exists(sym_file)
+assert os.path.exists(params_file)
+
+sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+eps = args.eps
+
+arg = {}
+aux = {}
+invalid = 0
+ac = 0
+for k in arg_params:
+    v = arg_params[k]
+    nv = v.asnumpy()
+    #print(k, nv.dtype)
+    nv = nv.astype(np.float32)
+    ac += nv.size
+    invalid += np.count_nonzero(np.abs(nv)<eps)
+    nv[np.abs(nv) < eps] = 0.0
+    arg[k] = mx.nd.array(nv, dtype='float32')
+print(invalid, ac)
+arg_params = arg
+invalid = 0
+ac = 0
+for k in aux_params:
+    v = aux_params[k]
+    nv = v.asnumpy().astype(np.float32)
+    ac += nv.size
+    invalid += np.count_nonzero(np.abs(nv)<eps)
+    nv[np.abs(nv) < eps] = 0.0
+    aux[k] = mx.nd.array(nv, dtype='float32')
+print(invalid, ac)
+aux_params = aux
+
+all_args = {}
+all_args.update(arg_params)
+all_args.update(aux_params)
+converted_model_path = onnx_mxnet.export_model(sym, all_args, [input_shape], np.float32, args.output, opset_version=11)
+
+model = onnx.load(args.output)
+graph = model.graph
+input_map = create_map(graph.input)
+node_map = create_map(graph.node)
+init_map = create_map(graph.initializer)
+
+#fix PRelu issue
+for input_name in input_map.keys():
+    if input_name.endswith('_gamma'):
+        node_name = input_name[:-6]
+        if not node_name in node_map:
+            continue
+        node = node_map[node_name]
+        if node.op_type!='PRelu':
+            continue
+        input_shape = input_map[input_name].type.tensor_type.shape.dim
+        input_dim_val=input_shape[0].dim_value
+        
+        graph.initializer.remove(init_map[input_name])
+        weight_array = numpy_helper.to_array(init_map[input_name])
+        
+        b=[]
+        for w in weight_array:
+            b.append(w)
+        new_nv = helper.make_tensor(input_name, TensorProto.FLOAT, [input_dim_val,1,1], b)
+        graph.initializer.extend([new_nv])
+
+for init_name in init_map.keys():
+    weight_array = numpy_helper.to_array(init_map[init_name])
+    assert weight_array.dtype==np.float32
+    if init_name in input_map:
+        graph.input.remove(input_map[init_name])
+
+#support batch-inference
+graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+
+onnx.save(model, args.output)
+
diff --git a/challenges/iccv21-mfr/onnx_helper.py b/challenges/iccv21-mfr/onnx_helper.py
new file mode 100644
index 000000000..1d44aceda
--- /dev/null
+++ b/challenges/iccv21-mfr/onnx_helper.py
@@ -0,0 +1,253 @@
+from __future__ import division
+import datetime
+import os
+import os.path as osp
+import glob
+import numpy as np
+import cv2
+import sys
+import onnxruntime
+import onnx
+import argparse
+from onnx import numpy_helper
+from insightface.data import get_image
+
+class ArcFaceORT:
+    def __init__(self, model_path, cpu=False):
+        self.model_path = model_path
+        self.cpu = cpu
+
+    #input_size is (w,h), return error message, return None if success
+    def check(self, track='cfat', test_img = None):
+        #default is cfat
+        max_model_size_mb=1024
+        max_feat_dim=512
+        max_time_cost=15
+        if track.startswith('ms1m'):
+            max_model_size_mb=1024
+            max_feat_dim=512
+            max_time_cost=10
+        elif track.startswith('glint'):
+            max_model_size_mb=1024
+            max_feat_dim=1024
+            max_time_cost=20
+        elif track.startswith('cfat'):
+            max_model_size_mb = 1024
+            max_feat_dim = 512
+            max_time_cost = 15
+        elif track.startswith('unconstrained'):
+            max_model_size_mb=1024
+            max_feat_dim=1024
+            max_time_cost=30
+        else:
+            return "track not found"
+
+        if not os.path.exists(self.model_path):
+            return "model_path not exists"
+        if not os.path.isdir(self.model_path):
+            return "model_path should be directory"
+        onnx_files = []
+        for _file in os.listdir(self.model_path):
+            if _file.endswith('.onnx'):
+                onnx_files.append(osp.join(self.model_path, _file))
+        if len(onnx_files)==0:
+            return "do not have onnx files"
+        self.model_file = sorted(onnx_files)[-1]
+        print('use onnx-model:', self.model_file)
+        try:
+            session = onnxruntime.InferenceSession(self.model_file, None)
+        except:
+            return "load onnx failed"
+        if self.cpu:
+            session.set_providers(['CPUExecutionProvider'])
+        input_cfg = session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        print('input-shape:', input_shape)
+        if len(input_shape)!=4:
+            return "length of input_shape should be 4"
+        if not isinstance(input_shape[0], str):
+            #return "input_shape[0] should be str to support batch-inference"
+            print('reset input-shape[0] to None')
+            model = onnx.load(self.model_file)
+            model.graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+            new_model_file = osp.join(self.model_path, 'zzzzrefined.onnx')
+            onnx.save(model, new_model_file)
+            self.model_file = new_model_file
+            print('use new onnx-model:', self.model_file)
+            try:
+                session = onnxruntime.InferenceSession(self.model_file, None)
+            except:
+                return "load onnx failed"
+            if self.cpu:
+                session.set_providers(['CPUExecutionProvider'])
+            input_cfg = session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            print('new-input-shape:', input_shape)
+
+        self.image_size = tuple(input_shape[2:4][::-1])
+        #print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        outputs = session.get_outputs()
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+            #print(o.name, o.shape)
+        if len(output_names)!=1:
+            return "number of output nodes should be 1"
+        self.session = session
+        self.input_name = input_name
+        self.output_names = output_names
+        #print(self.output_names)
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        if len(graph.node)<8:
+            return "too small onnx graph"
+
+        input_size = (112,112)
+        self.crop = None
+        if track=='cfat':
+            crop_file = osp.join(self.model_path, 'crop.txt')
+            if osp.exists(crop_file):
+                lines = open(crop_file,'r').readlines()
+                if len(lines)!=6:
+                    return "crop.txt should contain 6 lines"
+                lines = [int(x) for x in lines]
+                self.crop = lines[:4]
+                input_size = tuple(lines[4:6])
+        if input_size!=self.image_size:
+            return "input-size is inconsistant with onnx model input, %s vs %s"%(input_size, self.image_size)
+
+        self.model_size_mb = os.path.getsize(self.model_file) / float(1024*1024)
+        if self.model_size_mb > max_model_size_mb:
+            return "max model size exceed, given %.3f-MB"%self.model_size_mb
+
+        input_mean = None
+        input_std = None
+        if track=='cfat':
+            pn_file = osp.join(self.model_path, 'pixel_norm.txt')
+            if osp.exists(pn_file):
+                lines = open(pn_file,'r').readlines()
+                if len(lines)!=2:
+                    return "pixel_norm.txt should contain 2 lines"
+                input_mean = float(lines[0])
+                input_std = float(lines[1])
+        if input_mean is not None or input_std is not None:
+            if input_mean is None or input_std is None:
+                return "please set input_mean and input_std simultaneously"
+        else:
+            find_sub = False
+            find_mul = False
+            for nid, node in enumerate(graph.node[:8]):
+                print(nid, node.name)
+                if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                    find_sub = True
+                if node.name.startswith('Mul') or node.name.startswith('_mul') or node.name.startswith('Div'):
+                    find_mul = True
+            if find_sub and find_mul:
+                print("find sub and mul")
+                #mxnet arcface model
+                input_mean = 0.0
+                input_std = 1.0
+            else:
+                input_mean = 127.5
+                input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        for initn in graph.initializer:
+            weight_array = numpy_helper.to_array(initn)
+            dt = weight_array.dtype
+            if dt.itemsize<4:
+                return 'invalid weight type - (%s:%s)' % (initn.name, dt.name)
+        if test_img is None:
+            test_img = get_image('Tom_Hanks_54745')
+            test_img = cv2.resize(test_img, self.image_size)
+        else:
+            test_img = cv2.resize(test_img, self.image_size)
+        feat, cost = self.benchmark(test_img)
+        batch_result = self.check_batch(test_img)
+        batch_result_sum = float(np.sum(batch_result))
+        if batch_result_sum in [float('inf'), -float('inf')] or batch_result_sum != batch_result_sum:
+            print(batch_result)
+            print(batch_result_sum)
+            return "batch result output contains NaN!"
+
+        if len(feat.shape) < 2:
+           return "the shape of the feature must be two, but get {}".format(str(feat.shape))
+
+        if feat.shape[1] > max_feat_dim:
+            return "max feat dim exceed, given %d"%feat.shape[1]
+        self.feat_dim = feat.shape[1]
+        cost_ms = cost*1000
+        if cost_ms>max_time_cost:
+            return "max time cost exceed, given %.4f"%cost_ms
+        self.cost_ms = cost_ms
+        print('check stat:, model-size-mb: %.4f, feat-dim: %d, time-cost-ms: %.4f, input-mean: %.3f, input-std: %.3f'%(self.model_size_mb, self.feat_dim, self.cost_ms, self.input_mean, self.input_std))
+        return None
+
+    def check_batch(self, img):
+        if not isinstance(img, list):
+            imgs = [img, ] * 32
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1]:self.crop[3], self.crop[0]:self.crop[2], :]
+                if nimg.shape[0] != self.image_size[1] or nimg.shape[1] != self.image_size[0]:
+                    nimg = cv2.resize(nimg, self.image_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(
+            images=imgs, scalefactor=1.0 / self.input_std, size=self.image_size,
+            mean=(self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+
+    def meta_info(self):
+        return {'model-size-mb':self.model_size_mb, 'feature-dim':self.feat_dim, 'infer': self.cost_ms}
+
+
+    def forward(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.image_size
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1]:self.crop[3],self.crop[0]:self.crop[2],:]
+                if nimg.shape[0]!=input_size[1] or nimg.shape[1]!=input_size[0]:
+                    nimg = cv2.resize(nimg, input_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(imgs, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name : blob})[0]
+        return net_out
+
+    def benchmark(self, img):
+        input_size = self.image_size
+        if self.crop is not None:
+            nimg = img[self.crop[1]:self.crop[3],self.crop[0]:self.crop[2],:]
+            if nimg.shape[0]!=input_size[1] or nimg.shape[1]!=input_size[0]:
+                nimg = cv2.resize(nimg, input_size)
+            img = nimg
+        blob = cv2.dnn.blobFromImage(img, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        costs = []
+        for _ in range(50):
+            ta = datetime.datetime.now()
+            net_out = self.session.run(self.output_names, {self.input_name : blob})[0]
+            tb = datetime.datetime.now()
+            cost = (tb-ta).total_seconds()
+            costs.append(cost)
+        costs = sorted(costs)
+        cost = costs[5]
+        return net_out, cost
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='')
+    # general
+    parser.add_argument('workdir', help='submitted work dir', type=str)
+    parser.add_argument('--track', help='track name, for different challenge', type=str, default='cfat')
+    args = parser.parse_args()
+    handler = ArcFaceORT(args.workdir)
+    err = handler.check(args.track)
+    print('err:', err)
diff --git a/challenges/iccv21-mfr/tutorial_pytorch_cn.md b/challenges/iccv21-mfr/tutorial_pytorch_cn.md
new file mode 100644
index 000000000..39f5b9acb
--- /dev/null
+++ b/challenges/iccv21-mfr/tutorial_pytorch_cn.md
@@ -0,0 +1,167 @@
+## pytorch 训练样例
+
+[训练样例地址]()
+
+### 下载数据集
+
+* 下载 MS1MV3 [Link](https://github.com/deepinsight/insightface/tree/master/challenges/iccv19-lfr)
+* 下载 Glint360K [Link](https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc#4-download)
+
+### 服务器提交地址
+
+http://iccv21-mfr.com/
+
+### 安装依赖
+
+1. 安装 pytorch 1.7.1
+
+假设你已经安装好了GPU驱动和CUDA，根据你的CUDA版本，来选择你要安装的pytorch命令。  
+查看CUDA版本的命令为: `/usr/local/cuda/bin/nvcc -V`。
+
+Linux and Windows  
+```shell
+# CUDA 11.0
+pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
+
+# CUDA 10.2
+pip install torch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2
+
+# CUDA 10.1
+pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
+
+# CUDA 9.2
+pip install torch==1.7.1+cu92 torchvision==0.8.2+cu92 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
+
+```
+
+你也可以安装pytorch的其他版本，例如1.6.0或者更高的版本。
+
+2. 安装其他依赖
+
+```shell
+pip install -r requirement.txt
+```
+
+### 运行
+根据你的服务器，选择你要运行的命令。
+
+* 一台服务器，四张GPU运行
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py
+```
+
+* 一台服务器，八张GPU运行
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py
+```
+
+* 多台服务器，每台服务器8张GPU
+
+1. 节点0
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr="ip1" --master_port=1234 train.py
+```
+
+2. 节点1
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="ip1" --master_port=1234 train.py
+```
+
+
+### 提交
+
+1. 提交onnx模型
+
+竞赛要求模型转换为`onnx`模型提交，arcface_torch工程在保存模型时，会自动转换成为onnx，其地址为`${cfg.output}/backbone.onnx`。
+
+模型checkpoint介绍：
+```shell
+├── backbone.onnx                        # 需要提交的模型
+├── backbone.pth                         # pytorch 保存的模型
+├── rank_0_softmax_weight_mom.pt         # 模型并行原因，每张卡保存softmax独有的参数
+├── rank_0_softmax_weight.pt
+├── rank_1_softmax_weight_mom.pt
+├── rank_1_softmax_weight.pt
+├── ... ...
+└── training.log                          # 训练日志
+```
+
+2. 检查onnx模型是否规范
+
+提交模型前检查一下提交的模型是否规范，并测试模型的推理时间  
+
+
+测试命令：
+```shell
+python onnx_helper_sample.py --model_root ms1mv3_arcface_r50/
+```
+
+也可以先测试一下onnx模型在公开测试集IJBC上的性能：
+https://github.com/deepinsight/insightface/blob/master/recognition/arcface_torch/onnx_ijbc.py
+
+测试命令：
+```shell
+CUDA_VISIBLE_DEVICES=0 python onnx_ijbc.py --model-root ms1mv3_arcface_r50 --image-path IJB_release/IJBC --result-dir ms1mv3_arcface_r50
+```
+
+3. 模型大小参考
+
+推理时间是在`Tesla V100 GPU`中测试, 其中 onnxruntime-gpu==1.6。
+
+| 模型名称      | 大小/MB     | 推理时间/ms      |  
+| -------      | ----------  | -----------   |
+| R50          | 166         | 4.262         | 
+| R100         | 248         | 7.031         |  
+| R200         | 476         | 13.48         | 
+
+### 提示与技巧
+
+1. 训练加速-混合精度训练
+
+当时使用图灵架构的GPU时候，强烈建议开启混合精度训练模型，在`config.py`中，将`config.fp16`设置为True，可以节省大量显存和提升训练速度，例如：
+
+训练设置：
+MS1MV3(SSD) + 4*V100 + R100 + BatchSize 4*128
+
+- 开启混合精度训练前
+```python3
+# training log
+Training: 2021-05-12 00:00:42,110-Speed 884.42 samples/sec   Loss 47.2532   Epoch: 0   Global Step: 100
+Training: 2021-05-12 00:01:10,979-Speed 886.77 samples/sec   Loss 47.3550   Epoch: 0   Global Step: 150
+Training: 2021-05-12 00:01:43,936-Speed 776.80 samples/sec   Loss 47.0214   Epoch: 0   Global Step: 200
+Training: 2021-05-12 00:02:16,064-Speed 796.83 samples/sec   Loss 46.7781   Epoch: 0   Global Step: 250
+Training: 2021-05-12 00:02:45,018-Speed 884.18 samples/sec   Loss 46.3187   Epoch: 0   Global Step: 300
+# gpustat -i
+[0] Tesla V100-SXM2-32GB | 67 C,  99 % | 17844 / 32510 MB 
+[1] Tesla V100-SXM2-32GB | 64 C,  98 % | 17844 / 32510 MB 
+[2] Tesla V100-SXM2-32GB | 65 C,  93 % | 17916 / 32510 MB 
+[3] Tesla V100-SXM2-32GB | 72 C,  82 % | 17910 / 32510 MB 
+```
+
+- 开启混合精度训练后
+
+```python3
+# training log
+Training: 2021-05-12 00:04:27,869-Speed 1604.59 samples/sec   Loss 47.6050   Epoch: 0   Global Step: 100
+Training: 2021-05-12 00:04:43,681-Speed 1619.08 samples/sec   Loss 47.5865   Epoch: 0   Global Step: 150
+Training: 2021-05-12 00:04:59,460-Speed 1622.39 samples/sec   Loss 47.2380   Epoch: 0   Global Step: 200
+Training: 2021-05-12 00:05:15,271-Speed 1619.25 samples/sec   Loss 46.9030   Epoch: 0   Global Step: 250
+Training: 2021-05-12 00:05:31,065-Speed 1620.86 samples/sec   Loss 46.4425   Epoch: 0   Global Step: 300
+# gpustat -i
+[0] Tesla V100-SXM2-32GB | 64 C,  96 % | 10664 / 32510 M  
+[1] Tesla V100-SXM2-32GB | 63 C,  96 % | 10630 / 32510 MB 
+[2] Tesla V100-SXM2-32GB | 63 C,  79 % | 10736 / 32510 MB 
+[3] Tesla V100-SXM2-32GB | 70 C,  86 % | 10736 / 32510 MB
+```
+
+2. 训练加速-将数据挂载到内存盘来提升训练速度  
+使用如下的命令：
+```shell
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=40G  tmpfs /train_tmp
+```
+
+让后将训练集拷贝到目录`/train_tmp`下，然后开始训练。
diff --git a/challenges/iccv21-mfr/tutorial_pytorch_mask_aug.md b/challenges/iccv21-mfr/tutorial_pytorch_mask_aug.md
new file mode 100644
index 000000000..9bf2ba876
--- /dev/null
+++ b/challenges/iccv21-mfr/tutorial_pytorch_mask_aug.md
@@ -0,0 +1,34 @@
+# A tutorial on how to enable mask augmentation on arcface_torch training.
+
+The python package insightface==0.3.2 provides utilities to enable mask augmentation within one line:
+
+```
+transform_list.append(
+   MaskAugmentation(
+      mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'], 
+      mask_probs=[0.4, 0.4, 0.1, 0.1], h_low=0.33, h_high=0.4, p=self.mask_prob)
+   )
+```
+
+### Prepare
+
+1. Download antelope model pack by `bash> insightface-cli model.download antelope` which will be located at `~/.insightface/models/antelope`
+2. Generate BFM.mat and BFM_UV.mat following [here](https://github.com/deepinsight/insightface/tree/master/recognition/tools#data-prepare), for license concern.
+3. Generate new mask-rec dataset by `bash> insightface-cli rec.addmaskparam /data/ms1m-retinaface-t1 /data/ms1m-retinaface-t1mask` which generates and writes the mask params of each image into the record.
+
+
+### Add Mask Renderer Augmentation
+just by following code:
+```
+from insightface.app import MaskAugmentation
+self.transform_list.append(
+    MaskAugmentation(
+    mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'], 
+    mask_probs=[0.4, 0.4, 0.1, 0.1], 
+    h_low=0.33, h_high=0.4, p=0.1)
+)
+```
+
+Please check [dataset_mask.py](https://github.com/deepinsight/insightface/blob/master/challenges/iccv21-mfr/dataset_mask.py) for detail. 
+
+You can override the original dataset.py with this file to simply enable mask augmentation.
diff --git a/challenges/mfr/README.md b/challenges/mfr/README.md
new file mode 100644
index 000000000..8d374fbee
--- /dev/null
+++ b/challenges/mfr/README.md
@@ -0,0 +1,139 @@
+# MFR Ongoing
+
+
+💡 **Due to GPU resource constraints, we are temporarily taking this service offline. If any companies or individuals are interested in hosting this service, please contact me via email: ``guojia AT insightface.ai``**
+
+
+This is the ongoing version of [ICCV-2021 Masked Face Recognition Challenge & Workshop(MFR)](https://ibug.doc.ic.ac.uk/resources/masked-face-recognition-challenge-workshop-iccv-21/). We also extend it to involve some public available and popular benchmarks such as IJBC, LFW, CFPFP and AgeDB.
+
+
+(:bulb: :bulb: Once you find the name **IFRT** which is *InsightFace Recognition Test* in short anywhere, it is the same as MFR-Ongoing.)
+
+
+For detail, please check our ICCV 2021 workshop [paper](https://openaccess.thecvf.com/content/ICCV2021W/MFR/papers/Deng_Masked_Face_Recognition_Challenge_The_InsightFace_Track_Report_ICCVW_2021_paper.pdf).
+
+More information about the workshop challenge can be found [here](../iccv21-mfr), for reference.
+
+**MFR** testset consists of non-celebrities so we can ensure that it has very few overlap with public available face recognition training set, such as MS1M and CASIA as they mostly collected from online celebrities. As the result, we can evaluate the FAIR performance for different algorithms.
+
+In recent changes, we also add public available popular benchmarks such as IJBC, LFW, CFPFP, AgeDB into **MFR-Ongoing**.
+
+
+Current submission server link: [http://iccv21-mfr.com/](http://iccv21-mfr.com/)
+
+For any question, please send email to `insightface.challenge AT gmail.com`
+
+## Testsets
+
+In MFR-Ongoing, we will evaluate the accuracy of following testsets:
+
+  * **Accuracy between masked and non-masked faces.**
+  * **Accuracy among children(2~16 years old).**
+  * **Accuracy of globalised multi-racial benchmarks.**
+
+
+We ensure that there's no overlap between the above testsets and public available training datasets, as they are not collected from online celebrities.
+
+We also evaluate below public available popular benchmarks:
+
+  * **IJBC under FAR<=e-5 and FAR<=e-4.**
+  * **Some 1:1 verification testsets, such as LFW, CFPFP, AgeDB-30.**
+
+
+### ``Mask test-set:``
+
+Mask testset contains 6,964 identities, 6,964 masked images and 13,928 non-masked images. There are totally 13,928 positive pairs and 96,983,824 negative pairs.
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrt_mask_sample.jpg" alt="ifrtsample" width="360">
+</details>
+
+### ``Children test-set:``
+
+Children testset contains 14,344 identities and 157,280 images. There are totally 1,773,428 positive pairs and 24,735,067,692 negative pairs.
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrt_children_sample.jpg" alt="ifrtsample" width="360">
+</details>
+
+### ``Multi-racial test-set (MR in short):``
+
+The globalised multi-racial testset contains 242,143 identities and 1,624,305 images.
+
+| Race-Set     | Identities  | Images        |  Positive Pairs   | Negative Pairs        |
+| -------      | ----------  | -----------   |  -----------      | -----------           |
+| African      | 43,874      | 298,010       |  870,091          | 88,808,791,999        |
+| Caucasian    | 103,293     | 697,245       |  2,024,609        | 486,147,868,171       |
+| Indian       | 35,086      | 237,080       |  688,259          | 56,206,001,061        |
+| Asian        | 59,890      | 391,970       |  1,106,078        | 153,638,982,852       |
+| **ALL**      | **242,143** | **1,624,305** |  **4,689,037**    | **2,638,360,419,683** |
+
+<details>
+  <summary>Click to check the sample images(here we manually blur it to protect privacy) </summary>
+  <img src="https://github.com/nttstar/insightface-resources/blob/master/images/ifrtsample_blur.jpg" alt="ifrtsample" width="640">
+</details>
+
+## Evaluation Metric
+
+For ``Mask`` set, TAR is measured on mask-to-nonmask 1:1 protocal, with FAR less than 0.0001(e-4).
+
+For ``Children`` set, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.0001(e-4).
+
+For multi-racial sets, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6).
+
+For IJBC and verification test-set, we use the most common test protocal.
+
+Participants are ordered in terms of highest scores across two datasets: **TAR@Mask** and **TAR@MR-All**, by the formula of ``0.25 * TAR@Mask + 0.75 * TAR@MR-All``.
+
+
+
+
+## Baselines
+
+**``2021.04.25``** We made a clean on East Asian subset, by removing children images.
+
+**``2021.04.27``** Add onnx download links.
+
+| Backbone   | Dataset    | Method     | Mask   | Children | African | Caucasian | South Asian | East Asian | All    | size(mb) | infer(ms) | link |
+|------------|------------|------------|--------|----------|---------|-----------|-------------|------------|--------|----------|-----------|-----------|
+| R100  | Casia  | ArcFace  | 26.623 | 30.359   | 39.666  | 53.933    | 47.807      | 21.572     | 42.735 | 248.904  | 7.073     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUJpk8zC61HVN7Kg?e=zE9JDd) |
+| R100  | MS1MV2  | ArcFace  | 65.767 | 60.496   | 79.117  | 87.176    | 85.501      | 55.807     | 80.725 | 248.904  | 7.028     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUTlYEHJCHg3UYM-?e=ihxMpS) |
+| R18  | MS1MV3  | ArcFace | 47.853 | 41.047   | 62.613  | 75.125    | 70.213      | 43.859     | 68.326 | 91.658   | 1.856     | [download](https://1drv.ms/u/s!AswpsDO2toNKrTxlT6w1Jo02yzSh?e=KDhFAA) |
+| R34  | MS1MV3  | ArcFace | 58.723 | 55.834   | 71.644  | 83.291    | 80.084      | 53.712     | 77.365 | 130.245  | 3.054     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT2O5pgyVtwnjeMq?e=16S8LI) |
+| R50  | MS1MV3  | ArcFace | 63.850 | 60.457   | 75.488  | 86.115    | 84.305      | 57.352     | 80.533 | 166.305  | 4.262     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUUWd5i3a5OlFpM_?e=ExBDBN) |
+| R100 | MS1MV3 | ArcFace | 69.091 | 66.864   | 81.083  | 89.040    | 88.082      | 62.193     | 84.312 | 248.590  | 7.031     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUPwyqWvNXUlNd3P?e=pTLw9A) |
+| R18   | Glint360K   | ArcFace | 53.317 | 48.113   | 68.230  | 80.575    | 75.852      | 47.831     | 72.074 | 91.658   | 2.013     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT5ey4lCqFzlpzDd?e=VWP28J) |
+| R34   | Glint360K   | ArcFace | 65.106 | 65.454   | 79.907  | 88.620    | 86.815      | 60.604     | 83.015 | 130.245  | 3.044     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUBcgGkiuUS11Hsd?e=ISGDnP) |
+| R50   | Glint360K   | ArcFace | 70.233 | 69.952   | 85.272  | 91.617    | 90.541      | 66.813     | 87.077 | 166.305  | 4.340     | [download](https://1drv.ms/u/s!AswpsDO2toNKrT8jbvHxjqCY0d08?e=igfdrd) |
+| R100  | Glint360K  | ArcFace | 75.567 | 75.202   | 89.488  | 94.285    | 93.434      | 72.528     | 90.659 | 248.590  | 7.038     | [download](https://1drv.ms/u/s!AswpsDO2toNKrUFgLEIj-mnkb51b?e=vWqy2q) |
+| -       | *Private*     | <div style="width: 50pt">insightface-000 of frvt  | 97.760 | 93.358   | 98.850  | 99.372    | 99.058      | 87.694     | 97.481 | -  | -    |   -  |
+
+
+(MS1M-V2 means MS1M-ArcFace, MS1M-V3 means MS1M-RetinaFace).
+
+Inference time in above table was evaluated on Tesla V100 GPU, using onnxruntime-gpu==1.6.
+
+## Rules
+
+1. We have two tracks, academic and unconstrained.
+2. Please **DO NOT** register the account with messy or random characters(for both username and organization).
+3. **For academic submissions, we recommend to set the username as the name of your proposed paper or method. Orgnization hiding is not allowed(or the score will be banned) for this track but you can set the submission as private. You can also create multiple accounts, one account for one method.**
+4. Right now we only support 112x112 input, so make sure that the submission model accepts the correct input shape(['*',3,112,112]), in RGB order. Add an interpolate operator into the first layer of the submission model if you need a different input resolution.
+5. Participants submit onnx model, then get scores by our online evaluation. 
+6. Matching score is measured by cosine similarity.
+7. **Online evaluation server uses onnxruntime-gpu==1.8, cuda==11.1, cudnn==8.0.5, GPU is RTX3090.**
+8. Any float-16 model weights is prohibited, as it will lead to incorrect model size estimiation.
+9. Please use ``onnx_helper.py`` to check whether the model is valid.
+10. Leaderboard is now ordered in terms of highest scores across two datasets: **TAR@Mask** and **TAR@MR-All**, by the formula of ``0.25 * TAR@Mask + 0.75 * TAR@MR-All``.
+
+
+
+## Submission Guide
+
+1. Participants must package the onnx model for submission using ``zip xxx.zip model.onnx``.
+2. Each participant can submit three times a day at most.
+3. Please sign-up with the real organization name. You can hide the organization name in our system if you like(not allowed for academic track).
+4. You can decide which submission to be displayed on the leaderboard by clicking 'Set Public' button.
+5. Please click 'sign-in' on submission server if find you're not logged in.
diff --git a/common/face_align.py b/common/face_align.py
deleted file mode 100644
index 6cf5a2e64..000000000
--- a/common/face_align.py
+++ /dev/null
@@ -1,88 +0,0 @@
-
-import cv2
-import numpy as np
-from skimage import transform as trans
-
-src1 = np.array([
-     [51.642,50.115],
-     [57.617,49.990],
-     [35.740,69.007],
-     [51.157,89.050],
-     [57.025,89.702]], dtype=np.float32)
-#<--left 
-src2 = np.array([
-    [45.031,50.118],
-    [65.568,50.872],
-    [39.677,68.111],
-    [45.177,86.190],
-    [64.246,86.758]], dtype=np.float32)
-
-#---frontal
-src3 = np.array([
-    [39.730,51.138],
-    [72.270,51.138],
-    [56.000,68.493],
-    [42.463,87.010],
-    [69.537,87.010]], dtype=np.float32)
-
-#-->right
-src4 = np.array([
-    [46.845,50.872],
-    [67.382,50.118],
-    [72.737,68.111],
-    [48.167,86.758],
-    [67.236,86.190]], dtype=np.float32)
-
-#-->right profile
-src5 = np.array([
-    [54.796,49.990],
-    [60.771,50.115],
-    [76.673,69.007],
-    [55.388,89.702],
-    [61.257,89.050]], dtype=np.float32)
-
-src = np.array([src1,src2,src3,src4,src5])
-src_map = {112 : src, 224 : src*2}
-
-arcface_src = np.array([
-  [38.2946, 51.6963],
-  [73.5318, 51.5014],
-  [56.0252, 71.7366],
-  [41.5493, 92.3655],
-  [70.7299, 92.2041] ], dtype=np.float32 )
-
-arcface_src = np.expand_dims(arcface_src, axis=0)
-
-# In[66]:
-
-# lmk is prediction; src is template
-def estimate_norm(lmk, image_size = 112, mode='arcface'):
-  assert lmk.shape==(5,2)
-  tform = trans.SimilarityTransform()
-  lmk_tran = np.insert(lmk, 2, values=np.ones(5), axis=1)
-  min_M = []
-  min_index = []
-  min_error = float('inf') 
-  if mode=='arcface':
-    assert image_size==112
-    src = arcface_src
-  else:
-    src = src_map[image_size]
-  for i in np.arange(src.shape[0]):
-    tform.estimate(lmk, src[i])
-    M = tform.params[0:2,:]
-    results = np.dot(M, lmk_tran.T)
-    results = results.T
-    error = np.sum(np.sqrt(np.sum((results - src[i]) ** 2,axis=1)))
-#         print(error)
-    if error< min_error:
-        min_error = error
-        min_M = M
-        min_index = i
-  return min_M, min_index
-
-def norm_crop(img, landmark, image_size=112, mode='arcface'):
-  M, pose_index = estimate_norm(landmark, image_size, mode)
-  warped = cv2.warpAffine(img,M, (image_size, image_size), borderValue = 0.0)
-  return warped
-
diff --git a/common/flops_counter.py b/common/flops_counter.py
deleted file mode 100644
index 64469ec85..000000000
--- a/common/flops_counter.py
+++ /dev/null
@@ -1,114 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import sys
-import os
-import json
-import argparse
-import numpy as np
-import mxnet as mx
-
-
-def is_no_bias(attr):
-  ret = False
-  if 'no_bias' in attr and (attr['no_bias']==True or attr['no_bias']=='True'):
-    ret = True
-  return ret
-
-def count_fc_flops(input_filter, output_filter, attr):
-  #print(input_filter, output_filter ,attr)
-  ret = 2*input_filter*output_filter
-  if is_no_bias(attr):
-    ret -= output_filter
-  return int(ret)
-
-
-def count_conv_flops(input_shape, output_shape, attr):
-  kernel = attr['kernel'][1:-1].split(',')
-  kernel = [int(x) for x in kernel]
-
-  #print('kernel', kernel)
-  if is_no_bias(attr):
-    ret = (2*input_shape[1]*kernel[0]*kernel[1]-1)*output_shape[2]*output_shape[3]*output_shape[1]
-  else:
-    ret = 2*input_shape[1]*kernel[0]*kernel[1]*output_shape[2]*output_shape[3]*output_shape[1]
-  num_group = 1
-  if 'num_group' in attr:
-    num_group = int(attr['num_group'])
-  ret /= num_group
-  return int(ret)
-
-
-def count_flops(sym, **data_shapes):
-  all_layers = sym.get_internals()
-  #print(all_layers)
-  arg_shapes, out_shapes, aux_shapes = all_layers.infer_shape(**data_shapes)
-  out_shape_dict = dict(zip(all_layers.list_outputs(), out_shapes))
-
-  nodes = json.loads(sym.tojson())['nodes']
-  nodeid_shape = {}
-  for nodeid, node in enumerate(nodes):
-    name = node['name']
-    layer_name = name+"_output"
-    if layer_name in out_shape_dict:
-      nodeid_shape[nodeid] = out_shape_dict[layer_name]
-  #print(nodeid_shape)
-  FLOPs = 0
-  for nodeid, node in enumerate(nodes):
-    flops = 0
-    if node['op']=='Convolution':
-      output_shape = nodeid_shape[nodeid]
-      name = node['name']
-      attr = node['attrs']
-      input_nodeid = node['inputs'][0][0]
-      input_shape = nodeid_shape[input_nodeid]
-      flops = count_conv_flops(input_shape, output_shape, attr)
-    elif node['op']=='FullyConnected':
-      attr = node['attrs']
-      output_shape = nodeid_shape[nodeid]
-      input_nodeid = node['inputs'][0][0]
-      input_shape = nodeid_shape[input_nodeid]
-      output_filter = output_shape[1]
-      input_filter = input_shape[1]*input_shape[2]*input_shape[3]
-      #assert len(input_shape)==4 and input_shape[2]==1 and input_shape[3]==1
-      flops = count_fc_flops(input_filter, output_filter, attr)
-    #print(node, flops)
-    FLOPs += flops
-
-  return FLOPs
-
-def flops_str(FLOPs):
-  preset = [ (1e12, 'T'), (1e9, 'G'), (1e6, 'M'), (1e3, 'K') ]
-
-  for p in preset:
-    if FLOPs//p[0]>0:
-      N = FLOPs/p[0]
-      ret = "%.1f%s"%(N, p[1])
-      return ret
-  ret = "%.1f"%(FLOPs)
-  return ret
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='flops counter')
-  # general
-  #parser.add_argument('--model', default='../models2/y2-arcface-retinat1/model,1', help='path to load model.')
-  #parser.add_argument('--model', default='../models2/r100fc-arcface-retinaa/model,1', help='path to load model.')
-  parser.add_argument('--model', default='../models2/r50fc-arcface-emore/model,1', help='path to load model.')
-  args = parser.parse_args()
-  _vec = args.model.split(',')
-  assert len(_vec)==2
-  prefix = _vec[0]
-  epoch = int(_vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  FLOPs = count_flops(sym, data=(1,3,112,112))
-  print('FLOPs:', FLOPs)
-
diff --git a/cpp-align/FacePreprocess.h b/cpp-align/FacePreprocess.h
deleted file mode 100644
index aee9ec942..000000000
--- a/cpp-align/FacePreprocess.h
+++ /dev/null
@@ -1,150 +0,0 @@
-//
-// Created by Jack Yu on 23/03/2018.
-//
-
-#ifndef FACE_DEMO_FACEPREPROCESS_H
-#define FACE_DEMO_FACEPREPROCESS_H
-
-#include<opencv2/opencv.hpp>
-
-
-namespace FacePreprocess {
-
-    cv::Mat meanAxis0(const cv::Mat &src)
-    {
-        int num = src.rows;
-        int dim = src.cols;
-
-        // x1 y1
-        // x2 y2
-
-        cv::Mat output(1,dim,CV_32F);
-        for(int i = 0 ; i <  dim; i ++)
-        {
-            float sum = 0 ;
-            for(int j = 0 ; j < num ; j++)
-            {
-                sum+=src.at<float>(j,i);
-            }
-            output.at<float>(0,i) = sum/num;
-        }
-
-        return output;
-    }
-
-    cv::Mat elementwiseMinus(const cv::Mat &A,const cv::Mat &B)
-    {
-        cv::Mat output(A.rows,A.cols,A.type());
-
-        assert(B.cols == A.cols);
-        if(B.cols == A.cols)
-        {
-            for(int i = 0 ; i <  A.rows; i ++)
-            {
-                for(int j = 0 ; j < B.cols; j++)
-                {
-                    output.at<float>(i,j) = A.at<float>(i,j) - B.at<float>(0,j);
-                }
-            }
-        }
-        return output;
-    }
-
-
-    cv::Mat varAxis0(const cv::Mat &src)
-    {
-        cv:Mat temp_ = elementwiseMinus(src,meanAxis0(src));
-        cv::multiply(temp_ ,temp_ ,temp_ );
-        return meanAxis0(temp_);
-
-    }
-
-
-
-    int MatrixRank(cv::Mat M)
-    {
-        Mat w, u, vt;
-        SVD::compute(M, w, u, vt);
-        Mat1b nonZeroSingularValues = w > 0.0001;
-        int rank = countNonZero(nonZeroSingularValues);
-        return rank;
-
-    }
-
-//    References
-//    ----------
-//    .. [1] "Least-squares estimation of transformation parameters between two
-//    point patterns", Shinji Umeyama, PAMI 1991, DOI: 10.1109/34.88573
-//
-//    """
-//
-//    Anthor:Jack Yu
-    cv::Mat similarTransform(cv::Mat src,cv::Mat dst) {
-        int num = src.rows;
-        int dim = src.cols;
-        cv::Mat src_mean = meanAxis0(src);
-        cv::Mat dst_mean = meanAxis0(dst);
-        cv::Mat src_demean = elementwiseMinus(src, src_mean);
-        cv::Mat dst_demean = elementwiseMinus(dst, dst_mean);
-        cv::Mat A = (dst_demean.t() * src_demean) / static_cast<float>(num);
-        cv::Mat d(dim, 1, CV_32F);
-        d.setTo(1.0f);
-        if (cv::determinant(A) < 0) {
-            d.at<float>(dim - 1, 0) = -1;
-
-        }
-        Mat T = cv::Mat::eye(dim + 1, dim + 1, CV_32F);
-        cv::Mat U, S, V;
-        SVD::compute(A, S,U, V);
-
-        // the SVD function in opencv differ from scipy .
-
-
-        int rank = MatrixRank(A);
-        if (rank == 0) {
-            assert(rank == 0);
-
-        } else if (rank == dim - 1) {
-            if (cv::determinant(U) * cv::determinant(V) > 0) {
-                T.rowRange(0, dim).colRange(0, dim) = U * V;
-            } else {
-//            s = d[dim - 1]
-//            d[dim - 1] = -1
-//            T[:dim, :dim] = np.dot(U, np.dot(np.diag(d), V))
-//            d[dim - 1] = s
-                int s = d.at<float>(dim - 1, 0) = -1;
-                d.at<float>(dim - 1, 0) = -1;
-
-                T.rowRange(0, dim).colRange(0, dim) = U * V;
-                cv::Mat diag_ = cv::Mat::diag(d);
-                cv::Mat twp = diag_*V; //np.dot(np.diag(d), V.T)
-                Mat B = Mat::zeros(3, 3, CV_8UC1);
-                Mat C = B.diag(0);
-                T.rowRange(0, dim).colRange(0, dim) = U* twp;
-                d.at<float>(dim - 1, 0) = s;
-            }
-        }
-        else{
-            cv::Mat diag_ = cv::Mat::diag(d);
-            cv::Mat twp = diag_*V.t(); //np.dot(np.diag(d), V.T)
-            cv::Mat res = U* twp; // U
-            T.rowRange(0, dim).colRange(0, dim) = -U.t()* twp;
-        }
-        cv::Mat var_ = varAxis0(src_demean);
-        float val = cv::sum(var_).val[0];
-        cv::Mat res;
-        cv::multiply(d,S,res);
-        float scale =  1.0/val*cv::sum(res).val[0];
-        T.rowRange(0, dim).colRange(0, dim) = - T.rowRange(0, dim).colRange(0, dim).t();
-        cv::Mat  temp1 = T.rowRange(0, dim).colRange(0, dim); // T[:dim, :dim]
-        cv::Mat  temp2 = src_mean.t(); //src_mean.T
-        cv::Mat  temp3 = temp1*temp2; // np.dot(T[:dim, :dim], src_mean.T)
-        cv::Mat temp4 = scale*temp3;
-        T.rowRange(0, dim).colRange(dim, dim+1)=  -(temp4 - dst_mean.t()) ;
-        T.rowRange(0, dim).colRange(0, dim) *= scale;
-        return T;
-    }
-
-
-}
-#endif //FACE_DEMO_FACEPREPROCESS_H
diff --git a/cpp-package/inspireface/.clang-format b/cpp-package/inspireface/.clang-format
new file mode 100644
index 000000000..0d41632c5
--- /dev/null
+++ b/cpp-package/inspireface/.clang-format
@@ -0,0 +1,112 @@
+---
+Language:        Cpp
+# BasedOnStyle:  Google
+AccessModifierOffset: -4
+AlignAfterOpenBracket: Align
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlines: Left
+AlignOperands:   true
+AlignTrailingComments: true
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: false
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: Empty
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: true
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: true
+BinPackParameters: true
+BraceWrapping:   
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  AfterExternBlock: false
+  BeforeCatch:     false
+  BeforeElse:      false
+  IndentBraces:    false
+  SplitEmptyFunction: true
+  SplitEmptyRecord: true
+  SplitEmptyNamespace: true
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Attach
+BreakBeforeInheritanceComma: false
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+BreakConstructorInitializers: BeforeColon
+BreakAfterJavaFieldAnnotations: false
+BreakStringLiterals: true
+ColumnLimit:     150
+CommentPragmas:  '^ IWYU pragma:'
+CompactNamespaces: false
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 0
+ContinuationIndentWidth: 2
+Cpp11BracedListStyle: true
+DerivePointerAlignment: true
+DisableFormat:   false
+ExperimentalAutoDetectBinPacking: false
+FixNamespaceComments: true
+ForEachMacros:   
+  - foreach
+  - Q_FOREACH
+  - BOOST_FOREACH
+IncludeBlocks:   Preserve
+IncludeCategories: 
+  - Regex:           '^<ext/.*\.h>'
+    Priority:        2
+  - Regex:           '^<.*\.h>'
+    Priority:        1
+  - Regex:           '^<.*'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        3
+IncludeIsMainRegex: '([-_](test|unittest))?$'
+IndentCaseLabels: true
+IndentPPDirectives: None
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+JavaScriptQuotes: Leave
+JavaScriptWrapImports: true
+KeepEmptyLinesAtTheStartOfBlocks: false
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: false
+PenaltyBreakAssignment: 2
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Left
+ReflowComments:  true
+SortIncludes:    false
+SortUsingDeclarations: true
+SpaceAfterCStyleCast: false
+SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: ControlStatements
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 2
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Auto
+TabWidth:        4
+UseTab:          Never
+...
\ No newline at end of file
diff --git a/cpp-package/inspireface/.gitignore b/cpp-package/inspireface/.gitignore
new file mode 100644
index 000000000..604de0eb2
--- /dev/null
+++ b/cpp-package/inspireface/.gitignore
@@ -0,0 +1,31 @@
+.idea/*
+cmake-build-debug/*
+build/*
+test_*.rknn
+test_zip_*
+test_res/pack/*
+test_res/save/*
+!test_res/save/.gitkeep
+# !test_res/save/video_frames/
+test_res/save/video_frames/*
+!test_res/save/video_frames/.gitkeep
+resource/*
+pack/*
+*.zip
+3rdparty/
+.macos_cache/
+*.framework
+.cache/*
+.vscode/*
+build_local/*
+local_build/*
+cpp/inspireface/meta.cpp
+cpp/inspireface/version.txt
+.DS_Store
+._.DS_Store
+**/.DS_Store
+**/._.DS_Store
+.rknpu2_cache/
+test_res/data/video_frames/
+*.mp4
+build_*/
\ No newline at end of file
diff --git a/cpp-package/inspireface/CMakeLists.txt b/cpp-package/inspireface/CMakeLists.txt
new file mode 100644
index 000000000..3ff2f74e9
--- /dev/null
+++ b/cpp-package/inspireface/CMakeLists.txt
@@ -0,0 +1,379 @@
+cmake_minimum_required(VERSION 3.20)
+project(InspireFace)
+
+set(CMAKE_CXX_STANDARD 14)
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++14")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+
+set(INSPIRE_FACE_SERIES_NAME "Community Edition" CACHE STRING "Series name")
+
+# Hide symbols
+option(ISF_ENABLE_SYMBOL_HIDING "Enable symbol hiding." ON)
+if(ISF_ENABLE_SYMBOL_HIDING)
+    if(NOT WIN32)
+        set(CMAKE_C_VISIBILITY_PRESET hidden)
+        set(CMAKE_CXX_VISIBILITY_PRESET hidden)
+        set(CMAKE_VISIBILITY_INLINES_HIDDEN YES)
+    endif()
+else()
+    set(CMAKE_C_VISIBILITY_PRESET default)
+    set(CMAKE_CXX_VISIBILITY_PRESET default)
+    set(CMAKE_VISIBILITY_INLINES_HIDDEN NO)
+endif()
+
+# Current version
+set(INSPIRE_FACE_VERSION_MAJOR 1)
+set(INSPIRE_FACE_VERSION_MINOR 2)
+set(INSPIRE_FACE_VERSION_PATCH 3)
+
+# Converts the version number to a string
+string(CONCAT INSPIRE_FACE_VERSION_MAJOR_STR ${INSPIRE_FACE_VERSION_MAJOR})
+string(CONCAT INSPIRE_FACE_VERSION_MINOR_STR ${INSPIRE_FACE_VERSION_MINOR})
+string(CONCAT INSPIRE_FACE_VERSION_PATCH_STR ${INSPIRE_FACE_VERSION_PATCH})
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/cpp/inspireface/version.txt.in ${CMAKE_CURRENT_SOURCE_DIR}/cpp/inspireface/version.txt)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/python/version.txt.in ${CMAKE_CURRENT_SOURCE_DIR}/python/version.txt)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/cpp/inspireface/meta.cpp.in ${CMAKE_CURRENT_SOURCE_DIR}/cpp/inspireface/meta.cpp)
+
+# Check that the 3rdparty folder exists
+if(NOT EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/3rdparty")
+    message(STATUS "3rdparty directory not found, cloning from github...")
+    
+    # Execute git clone and directly specify it in the current directory
+    execute_process(
+        COMMAND git clone --recurse-submodules https://github.com/tunmx/inspireface-3rdparty.git "${CMAKE_CURRENT_SOURCE_DIR}/3rdparty"
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        RESULT_VARIABLE GIT_RESULT
+    )
+    
+    if(NOT GIT_RESULT EQUAL "0")
+        message(FATAL_ERROR "Failed to clone 3rdparty repository")
+    else()
+        message(STATUS "Successfully cloned 3rdparty repository")
+    endif()
+else()
+    message(STATUS "3rdparty directory already exists")
+endif()
+
+# Install cpp api header file
+option(ISF_INSTALL_CPP_HEADER "Install cpp api header file." ON)
+
+# Set the ISF_THIRD_PARTY_DIR variable to allow it to be set externally from the command line, or use the default path if it is not set
+set(ISF_THIRD_PARTY_DIR "${CMAKE_CURRENT_SOURCE_DIR}/3rdparty" CACHE PATH "Path to the third-party libraries directory")
+
+# Check that the SANITIZE compile option is enabled
+set(ISF_SANITIZE_ADDRESS OFF CACHE BOOL "Enable AddressSanitizer")
+set(ISF_SANITIZE_LEAK OFF CACHE BOOL "Enable LeakSanitizer")
+
+if (ISF_SANITIZE_ADDRESS AND ISF_SANITIZE_LEAK)
+    message(FATAL_ERROR "Cannot enable both AddressSanitizer and LeakSanitizer.")
+endif()
+
+# Enable AddressSanitizer
+if(ISF_SANITIZE_ADDRESS)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=address")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=address")
+endif()
+
+# Enable LeakSanitizer
+if(ISF_SANITIZE_LEAK)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=leak")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=leak")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=leak")
+endif()
+
+set(TARGET_PLATFORM "drawin-x86" CACHE STRING "Target platform")
+
+# TensorRT dependency configuration
+set(ISF_ENABLE_TENSORRT OFF CACHE BOOL "Enable TensorRT Inference")
+set(TENSORRT_ROOT "/usr/local/TensorRT" CACHE PATH "Path to TensorRT installation")
+if(ISF_ENABLE_TENSORRT)
+    list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/toolchain)
+    include(FindTensorRT)
+    add_definitions("-DISF_ENABLE_TENSORRT")
+    add_definitions("-DINFERENCE_WRAPPER_ENABLE_TENSORRT")
+endif()
+
+# RKNN dependency configuration
+option(ISF_ENABLE_RKNN "Use RKNPU." OFF)
+set(ISF_RK_DEVICE_TYPE "RV1109RV1126" CACHE STRING "Type of the device")
+set(ISF_RK_COMPILER_TYPE "armhf" CACHE STRING "Type of the compiler")
+if (ISF_ENABLE_RKNN)
+    add_definitions("-DISF_ENABLE_RKNN")
+    # Device list
+    set(RKNPU1_DEVICES "RV1109RV1126")
+    set(RKNPU2_DEVICES "RK356X" "RK3588" "RV1106")
+    set(ISF_RKNPU_MAJOR "")
+    list(FIND RKNPU1_DEVICES "${ISF_RK_DEVICE_TYPE}" INDEX_RKNPU1)
+    list(FIND RKNPU2_DEVICES "${ISF_RK_DEVICE_TYPE}" INDEX_RKNPU2)
+    if(INDEX_RKNPU1 GREATER_EQUAL 0)
+        # match rknn1
+        set(ISF_RKNPU_MAJOR "rknpu1")
+    elseif(INDEX_RKNPU2 GREATER_EQUAL 0)
+        # match rknn2
+        set(ISF_RKNPU_MAJOR "rknpu2")
+    endif()
+    # Result
+    message(STATUS "Use ${ISF_RKNPU_MAJOR}")
+
+    if(ISF_RK_DEVICE_TYPE STREQUAL "RV1106")
+        set(ISF_RKNPU_RV1106 ON)
+        add_definitions("-DISF_RKNPU_RV1106")
+    endif()
+    
+endif ()
+
+# Enable Rockchip RGA
+option(ISF_ENABLE_RGA "Use Rockchip RGA." OFF)
+if(ISF_ENABLE_RGA)
+    if(NOT ISF_ENABLE_RKNN)
+        message(FATAL_ERROR "RGA requires RKNN to be enabled. Please enable ISF_ENABLE_RKNN first.")
+    endif()
+    if(NOT ISF_RKNPU_MAJOR STREQUAL "rknpu2")
+        message(FATAL_ERROR "RGA requires RKNPU2. Please use a device that supports RKNPU2.")
+    endif()
+
+    add_definitions("-DISF_ENABLE_RGA")
+    if(ANDROID)
+        set(PLAT Android)
+    else()
+        set(PLAT Linux)
+    endif()
+    set(RGA_LIBS ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/librga/${PLAT}/${ISF_RK_COMPILER_TYPE}/librga.a)
+    set(RGA_INCLUDE_DIRS ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/librga/include)
+endif()
+
+option(ISF_ENABLE_COST_TIME "Enable cost time." OFF)
+if(ISF_ENABLE_COST_TIME)
+    add_definitions("-DISF_ENABLE_COST_TIME")
+endif()
+
+# Apple device dedicated expansion services, support for some models coreml reasoning and so on
+set(ISF_ENABLE_APPLE_EXTENSION OFF CACHE BOOL "Enable Apple device extension services, such as coreml, etc.")
+if(ISF_ENABLE_APPLE_EXTENSION)  
+    add_definitions("-DISF_ENABLE_APPLE_EXTENSION")
+    add_definitions("-DINFERENCE_WRAPPER_ENABLE_COREML")
+endif()
+
+# Platform configuration
+option(ISF_BUILD_LINUX_ARM7 "Platform Armv7." OFF)
+option(ISF_BUILD_LINUX_AARCH64 "Platform Armv8." OFF)
+option(ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA "The global inference backend uses MNN CUDA." OFF)
+
+
+if (ISF_BUILD_LINUX_ARM7)
+    set(CPU_ARCH "armhf")
+endif()
+if (ISF_BUILD_LINUX_AARCH64)
+    set(CPU_ARCH "aarch64")
+endif()
+
+# If you want to build the unit-test, you need to set this to ON
+option(ISF_BUILD_WITH_TEST "Open Build Unit-Test." ON)
+
+# If you want to build the sample, you need to set this to ON
+option(ISF_BUILD_WITH_SAMPLE "Open Build Sample Exec." ON)
+
+set(SRC_DIR ${CMAKE_CURRENT_SOURCE_DIR}/cpp/)
+
+option(ISF_NEVER_USE_OPENCV "Never use opencv." ON)
+if(ISF_NEVER_USE_OPENCV)
+    set(INSPIRECV_BACKEND_OPENCV OFF)
+    set(INSPIRECV_BACKEND_OKCV_USE_OPENCV OFF)
+    set(INSPIRECV_BACKEND_OKCV_USE_OPENCV_IO OFF)
+    set(INSPIRECV_BACKEND_OKCV_USE_OPENCV_GUI OFF)
+endif()
+
+
+option(ISF_ENABLE_OPENCV "Depends on opencv." OFF)
+if(INSPIRECV_BACKEND_OPENCV OR INSPIRECV_BACKEND_OKCV_USE_OPENCV OR INSPIRECV_BACKEND_OKCV_USE_OPENCV_IO OR INSPIRECV_BACKEND_OKCV_USE_OPENCV_GUI)
+    set(ISF_ENABLE_OPENCV ON)
+    add_definitions("-DISF_ENABLE_OPENCV")
+endif()
+
+
+include_directories(${ISF_THIRD_PARTY_DIR}/InspireCV/3rdparty/Eigen-3.4.0-Headers)
+
+if(APPLE)
+    if(IOS)
+        add_definitions(-DTARGET_OS_IOS)
+        message(IOS_3RDPARTY=${IOS_3RDPARTY})
+        set(CMAKE_XCODE_ATTRIBUTE_ONLY_ACTIVE_ARCH NO)
+        set(CMAKE_XCODE_ATTRIBUTE_ENABLE_BITCODE NO)
+            add_definitions(-DTARGET_OS_IOS)
+        message(IOS_3RDPARTY=${IOS_3RDPARTY})
+        set(CMAKE_XCODE_ATTRIBUTE_ONLY_ACTIVE_ARCH NO)
+        set(CMAKE_XCODE_ATTRIBUTE_ENABLE_BITCODE NO)
+
+        set(MNN_FRAMEWORK_PATH "${IOS_3RDPARTY}/MNN.framework")
+        include_directories("${MNN_FRAMEWORK_PATH}/")
+        link_directories(${IOS_3RDPARTY})
+    else()
+        message("Use apple device")
+        set(PLAT darwin)
+    endif()
+else()
+    message("Use linux device")
+    set(PLAT linux)
+endif()
+
+# OpenCV dependency configuration
+if(ISF_ENABLE_OPENCV)
+    if (APPLE)
+        if (IOS)
+            link_directories(${IOS_3RDPARTY})
+            include_directories(${IOS_3RDPARTY}/opencv2.framework)
+            set(OpenCV_DIR "${IOS_3RDPARTY}/opencv2.framework")
+        else()
+            message("Use apple device")
+            set(PLAT darwin)
+            find_package(OpenCV REQUIRED)
+        endif ()
+    else()
+        if (ISF_BUILD_LINUX_ARM7 OR ISF_BUILD_LINUX_AARCH64)
+            set(DISABLE_GUI ON)
+            add_definitions("-DDISABLE_GUI")
+            if (ISF_RK_DEVICE_TYPE STREQUAL "RV1109RV1126" AND ISF_ENABLE_RKNN)
+                # In special cases, specialize for that version
+                message("The OpenCV that builds the RV1109RV1126 version depends on is specialized!")
+                set(OpenCV_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-armhf/share/OpenCV)
+                set(OpenCV_STATIC_INCLUDE_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-armhf/include/)
+                set(PLAT linux-arm7)
+            else()
+                if (VERSION_MAJOR STREQUAL "3")
+                    set(CV_CMAKE_FOLDER share/OpenCV)
+                elseif(VERSION_MAJOR STREQUAL "4")
+                    set(CV_CMAKE_FOLDER lib/cmake/opencv4)
+                endif ()
+                if(ISF_BUILD_LINUX_ARM7)
+                    set(PLAT linux-arm7)
+                    message("The OpenCV that builds the gnueabihf version depends on is specialized!")
+                    set(OpenCV_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-armhf/share/OpenCV)
+                    set(OpenCV_STATIC_INCLUDE_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-armhf/include/)
+                elseif(ISF_BUILD_LINUX_AARCH64)
+                    set(PLAT linux-aarch64)
+                    message("The OpenCV that builds the aarch64 version depends on is specialized!")
+                    set(OpenCV_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-aarch64/share/OpenCV)
+                    set(OpenCV_STATIC_INCLUDE_DIR ${OPENCV_PRECOMPILED_DIR}/opencv/3.4.5/opencv-linux-aarch64/include/)
+                endif()
+            endif()
+        else ()
+            set(PLAT linux)
+        endif ()
+        find_package(OpenCV REQUIRED)
+    endif ()
+endif ()
+
+# InspireCV dependency configuration
+set(INSPIRECV_BUILD_OBJECT_LIBS ON CACHE BOOL "Build InspireCV as object library")
+add_subdirectory(${ISF_THIRD_PARTY_DIR}/InspireCV EXCLUDE_FROM_ALL)
+set(INSPIRECV_INCLUDE_PATH "${ISF_THIRD_PARTY_DIR}/InspireCV/include")
+set(INSPIRECV_LIBS InspireCV)
+
+# If you need using CUDA-enabled MNN, you need to manually configure the pre-compiled CUDA-enabled MNN library path
+set(ISF_LINUX_MNN_CUDA "" CACHE STRING "Path to CUDA directory")
+
+if (ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA)
+    message("Global MNN CUDA device inference")
+    add_definitions("-DISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA")
+    # Use MNN Cuda
+    if (NOT DEFINED ISF_LINUX_MNN_CUDA)
+        # MNN Options
+        set(MNN_CUDA ON)
+        add_subdirectory(${ISF_THIRD_PARTY_DIR}/MNN EXCLUDE_FROM_ALL)
+        set(MNN_INCLUDE_DIRS "${ISF_THIRD_PARTY_DIR}/MNN/include")
+        set(MNN_LIBS MNN)
+    else()
+        message(ISF_LINUX_MNN_CUDA=${ISF_LINUX_MNN_CUDA})
+        set(MNN_INCLUDE_DIRS ${ISF_LINUX_MNN_CUDA}/include)
+        link_directories(${ISF_LINUX_MNN_CUDA}/)
+        set(MNN_LIBS MNN)
+    endif()
+
+elseif(DEFINED MNN_STATIC_PATH)
+    message("Using static MNN from specified path: ${MNN_STATIC_PATH}")
+    set(MNN_INCLUDE_DIRS "${MNN_STATIC_PATH}/include")
+    set(MNN_LIBS "${MNN_STATIC_PATH}/lib/libMNN.a")
+elseif(IOS)
+    message(Build iOS)
+
+elseif(DEFINED ISF_MNN_CUSTOM_SOURCE)
+    message("Using custom external MNN source path: ${ISF_MNN_CUSTOM_SOURCE}")
+    # In particular, rknpu2 uses a lower version of mnn
+    set(MNN_BUILD_SHARED_LIBS OFF CACHE BOOL "Build MNN as a shared library")
+    add_subdirectory(${ISF_MNN_CUSTOM_SOURCE} ${CMAKE_BINARY_DIR}/MNN EXCLUDE_FROM_ALL)
+    set(MNN_INCLUDE_DIRS "${ISF_MNN_CUSTOM_SOURCE}/include")
+    set(MNN_LIBS MNN)
+    
+else ()
+    # Default or fallback case for MNN setup
+    message("Default or fallback case for MNN setup")
+    # MNN Options
+    set(MNN_BUILD_SHARED_LIBS OFF CACHE BOOL "Build MNN as a shared library")
+
+    # Set MNN build output directory to CMAKE_BINARY_DIR/MNN
+    add_subdirectory(${ISF_THIRD_PARTY_DIR}/MNN ${CMAKE_BINARY_DIR}/MNN EXCLUDE_FROM_ALL)
+    set(MNN_INCLUDE_DIRS "${ISF_THIRD_PARTY_DIR}/MNN/include")
+    set(MNN_LIBS MNN)
+
+endif ()
+
+
+# Set the installation directory to the build directory
+set(CMAKE_INSTALL_PREFIX "${CMAKE_BINARY_DIR}/install")
+
+# Add a child project: InspireFace Source
+add_subdirectory(cpp/inspireface)
+get_property(InspireFace TARGET InspireFace PROPERTY InspireFace)
+
+# Add a child project: Samples
+if (ISF_BUILD_WITH_SAMPLE)
+    add_subdirectory(cpp/sample)  # Add a child project: Samples
+endif ()
+
+# Add a child project: Unit-Test
+if (ISF_BUILD_WITH_TEST)
+    add_subdirectory(cpp/test)  # Add a child project: Unit-Test
+endif ()
+
+# Print Message
+message(STATUS ">>>>>>>>>>>>>")
+message(STATUS "InspireFace Project Global:")
+message(STATUS "\t CMAKE_SYSTEM_NAME: ${CMAKE_SYSTEM_NAME}")
+message(STATUS "\t CMAKE_BUILD_TYPE: ${CMAKE_BUILD_TYPE}")
+message(STATUS "\t ISF_THIRD_PARTY_DIR: ${ISF_THIRD_PARTY_DIR}")
+message(STATUS "\t ISF_SANITIZE_ADDRESS: ${ISF_SANITIZE_ADDRESS}")
+message(STATUS "\t ISF_SANITIZE_LEAK: ${ISF_SANITIZE_LEAK}")
+message(STATUS "\t ISF_ENABLE_RKNN: ${ISF_ENABLE_RKNN}")
+if (ISF_ENABLE_RKNN)
+    message(STATUS "\t ISF_RKNPU_MAJOR: ${ISF_RKNPU_MAJOR}")
+    message(STATUS "\t ISF_RK_DEVICE_TYPE: ${ISF_RK_DEVICE_TYPE}")
+    message(STATUS "\t ISF_RK_COMPILER_TYPE: ${ISF_RK_COMPILER_TYPE}")
+    if(ISF_RKNPU_RV1106)
+        message(STATUS "\t ISF_RKNPU_RV1106: ${ISF_RKNPU_RV1106}")
+    endif()
+endif ()
+message(STATUS "\t ISF_BUILD_LINUX_ARM7: ${ISF_BUILD_LINUX_ARM7}")
+message(STATUS "\t ISF_BUILD_LINUX_AARCH64: ${ISF_BUILD_LINUX_AARCH64}")
+message(STATUS "\t ISF_BUILD_WITH_TEST: ${ISF_BUILD_WITH_TEST}")
+message(STATUS "\t ISF_BUILD_WITH_SAMPLE: ${ISF_BUILD_WITH_SAMPLE}")
+message(STATUS "\t OpenCV_DIR: ${OpenCV_DIR}")
+if(ISF_ENABLE_OPENCV)
+    message(STATUS "\t OpenCV: Enabled")
+else()
+    message(STATUS "\t OpenCV: Disabled")
+endif()
+
+if(ISF_ENABLE_APPLE_EXTENSION)
+    message(STATUS "\t ISF_ENABLE_APPLE_EXTENSION: ${ISF_ENABLE_APPLE_EXTENSION}")
+endif()
+
+# Install cpp api header file
+if(ISF_INSTALL_CPP_HEADER)
+    message(STATUS "\t ISF_INSTALL_CPP_HEADER: ${ISF_INSTALL_CPP_HEADER}")
+endif()
+
+
+message(STATUS "\t CMAKE_INSTALL_PREFIX: ${CMAKE_INSTALL_PREFIX}")
diff --git a/cpp-package/inspireface/README.md b/cpp-package/inspireface/README.md
new file mode 100644
index 000000000..f28493999
--- /dev/null
+++ b/cpp-package/inspireface/README.md
@@ -0,0 +1,818 @@
+# InspireFace
+[![GitHub release](https://img.shields.io/github/v/release/HyperInspire/InspireFace.svg?style=for-the-badge&color=blue&label=Github+release&logo=github)](https://github.com/HyperInspire/InspireFace/releases/latest)
+[![Model](https://img.shields.io/github/v/release/HyperInspire/InspireFace.svg?style=for-the-badge&color=blue&label=Model+Zoo&logo=github)](https://github.com/HyperInspire/InspireFace/releases/tag/v1.x)
+[![pypi](https://img.shields.io/pypi/v/inspireface.svg?style=for-the-badge&color=orange&label=PYPI+release&logo=python)](https://pypi.org/project/inspireface/)
+[![JitPack](https://img.shields.io/jitpack/v/github/HyperInspire/inspireface-android-sdk?style=for-the-badge&color=green&label=JitPack&logo=android)](https://jitpack.io/#HyperInspire/inspireface-android-sdk)
+[![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?&style=for-the-badge&label=building&logo=cmake)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml)
+[![test](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?&style=for-the-badge&label=testing&logo=c)](https://github.com/HyperInspire/InspireFace/actions/workflows/test_ubuntu_x86_Pikachu.yaml)
+[![Document](https://img.shields.io/badge/Document-Building-blue?style=for-the-badge&logo=readthedocs)](https://doc.inspireface.online/)
+
+
+
+InspireFace is a cross-platform face recognition SDK developed in C/C++, supporting multiple operating systems and various backend types for inference, such as CPU, GPU, and NPU.
+
+If you require further information on tracking development branches, CI/CD processes, or downloading pre-compiled libraries, please visit our [development repository](https://github.com/HyperInspire/InspireFace).
+
+Please contact [contact@insightface.ai](mailto:contact@insightface.ai?subject=InspireFace) using your company e-mail for commercial support, including obtaining and integrating higher accuracy models, as well as custom development.
+
+<img src="images/banner.jpg" alt="banner" style="zoom:80%;" />
+
+---
+
+📘 [Documentation](https://doc.inspireface.online/) is a **work in progress**.  
+We welcome your questions💬, they help guide and accelerate its development.
+
+## Change Logs
+
+**`2026-07-22`** Add an Android example app.
+
+**`2025-08-03`** Add a multi-link model download channel for the Python-SDK.
+
+**`2025-06-15`** The [ErrorCode-Table](/doc/Error-Feedback-Codes.md) has been reorganized and streamlined.
+
+**`2025-06-08`** Add facial expression recognition.
+
+**`2025-04-27`** Optimize some issues and provide a stable version.
+
+**`2025-03-16`** Acceleration using NVIDIA-GPU (**CUDA**) devices is already supported.
+
+**`2025-03-09`** Release of android sdk in JitPack.
+
+**`2025-02-20`** Upgrade the face landmark model.
+
+**`2025-01-21`** Update all models to t3 and add tool to convert cosine similarity to percentage.
+
+**`2025-01-08`** Support inference on Rockchip devices **RK3566/RK3568** NPU.
+
+**`2024-12-25`** Add support for optional **RKRGA** image acceleration processing on Rockchip devices.
+
+**`2024-12-22`** Started adapting for multiple Rockchip devices with NPU support, beginning with **RV1103/RV1106** support.
+
+**`2024-12-10`** Added support for quick installation via Python package manager.
+
+**`2024-10-09`** Added system resource monitoring and session statistics.
+
+**`2024-09-30`** Fixed some bugs in the feature hub.
+
+**`2024-08-18`** Updating [Benchmark](doc/Benchmark-Remark(Updating).md): Using CoreML with Apple's Neural Engine (ANE) on the iPhone 13, the combined processes of **Face Detection** + **Alignment** + **Feature Extraction** take less than **2ms**.
+
+**`2024-07-17`** Add global resource statistics monitoring to prevent memory leaks.
+
+**`2024-07-07`** Add some face action detection to the face interaction module.
+
+**`2024-07-05`** Fixed some bugs in the python ctypes interface.
+
+**`2024-07-03`** Add the blink detection algorithm of face interaction module.
+
+**`2024-07-02`** Fixed several bugs in the face detector with multi-level input.
+
+## License
+
+The licensing of the open-source models employed by InspireFace adheres to the same requirements as InsightFace, specifying their use solely for academic purposes and explicitly prohibiting commercial applications.
+
+## Quick Start
+
+### Try the Android Example App
+
+<p>
+  <a href="http://fir.tunm.top/pro/pz7b3dgv">
+    <img src="images/inspireface-android-example-app-download.png" alt="Scan the QR code to download the InspireFace Android example app" width="220" />
+  </a>
+</p>
+
+<p>
+  <strong>Scan the QR code to download the Android example app</strong><br />
+  Or <a href="http://fir.tunm.top/pro/pz7b3dgv">download the app directly</a>.
+</p>
+
+### Python Quick Start
+
+For Python users on **Linux and MacOS**, InspireFace can be quickly installed via pip:
+
+```bash
+pip install -U inspireface
+```
+
+After installation, you can use inspireface like this:
+
+```Python
+import cv2
+import inspireface as isf
+
+# Create a session with optional features
+opt = isf.HF_ENABLE_NONE
+session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+
+# Load the image using OpenCV.
+image = cv2.imread(image_path)
+
+# Perform face detection on the image.
+faces = session.face_detection(image)
+
+for face in faces:
+    x1, y1, x2, y2 = face.location
+    rect = ((x1, y1), (x2, y2), face.roll)
+     # Calculate center, size, and angle
+    center = ((x1 + x2) / 2, (y1 + y2) / 2)
+    size = (x2 - x1, y2 - y1)
+    angle = face.roll
+
+    # Apply rotation to the bounding box corners
+    rect = ((center[0], center[1]), (size[0], size[1]), angle)
+    box = cv2.boxPoints(rect)
+    box = box.astype(int)
+
+    # Draw the rotated bounding box
+    cv2.drawContours(image, [box], 0, (100, 180, 29), 2)
+
+cv2.imshow("face detection", image)
+cv2.waitKey(0)
+cv2.destroyAllWindows()
+```
+
+⚠️The project is currently in a rapid iteration phase, **before each update**, please pull the latest model from the remote side!
+
+```python
+import inspireface
+
+for model in ["Pikachu", "Megatron"]:
+    inspireface.pull_latest_model(model)
+```
+
+More examples can be found in the [python](python/) directory.
+
+## Preparation
+### Clone 3rdparty
+
+Clone the `3rdparty` repository from the remote repository into the root directory of the project. Note that this repository contains some submodules. When cloning, you should use the `--recurse-submodules` parameter, or after entering the directory, use `git submodule update --init --recursive` to fetch and synchronize the latest submodules:
+
+```Bash
+# Must enter this directory
+cd InspireFace
+# Clone the repository and pull submodules
+git clone --recurse-submodules https://github.com/tunmx/inspireface-3rdparty.git 3rdparty
+```
+
+If you need to update the `3rdparty` repository to ensure it is current, or if you didn't use the `--recursive` parameter during the initial pull, you can run `git submodule update --init --recursive`:
+
+```bash
+# Must enter this directory
+cd InspireFace
+# If you're not using recursive pull
+git clone https://github.com/tunmx/inspireface-3rdparty.git 3rdparty
+
+cd 3rdparty
+git pull
+# Update submodules
+git submodule update --init --recursive
+```
+
+### Downloading Model Package Files
+
+You can download the model package files containing models and configurations needed for compilation from [Release Page](https://github.com/HyperInspire/InspireFace/releases/tag/v1.x) and extract them to any location. 
+
+You can use the **command/download_models_general.sh** command to download resource files, which will be downloaded to the **test_res/pack** directory. This way, when running the Test program, it can access and read the resource files from this path by default.
+
+⚠️The project is currently in a rapid iteration phase, **before each update**, please pull the latest model from the remote side!
+
+```bash
+# Download lightweight resource files for mobile device
+bash command/download_models_general.sh Pikachu
+# Download resource files for mobile device or PC/server
+bash command/download_models_general.sh Megatron
+# Download resource files for RV1109
+bash command/download_models_general.sh Gundam_RV1109
+# Download resource files for RV1106
+bash command/download_models_general.sh Gundam_RV1106
+# Download resource files for RK356X
+bash command/download_models_general.sh Gundam_RK356X
+# Download resource files for RK3588
+bash command/download_models_general.sh Gundam_RK3588
+# Download resource files for NVIDIA-GPU Device(TensorRT)
+bash command/download_models_general.sh Megatron_TRT
+
+# Download all model files
+bash command/download_models_general.sh
+```
+
+### Installing MNN
+The '**3rdparty**' directory already includes the MNN library and specifies a particular version as the stable version. If you need to enable or disable additional configuration options during compilation, you can refer to the CMake Options provided by MNN. If you need to use your own precompiled version, feel free to replace it.
+
+### Requirements
+
+- CMake (version 3.20 or higher)
+- NDK (version 16 or higher, only required for Android) [**Optional**]
+- MNN (version 3.x or higher)
+- C++ Compiler
+    - Either GCC or Clang can be used (macOS does not require additional installation as Xcode is included)
+        - Recommended GCC version is 4.9 or higher
+            - Note that in some distributions, GCC (GNU C Compiler) and G++ (GNU C++ Compiler) are installed separately.
+            - For instance, on Ubuntu, you need to install both gcc and g++
+        - Recommended Clang version is 3.9 or higher
+    - arm-linux-gnueabihf (for RV1109/RV1126) [**Optional**]
+        - Prepare the cross-compilation toolchain in advance, such as gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf
+- CUDA (version 11.x or higher) [**Optional**]
+    - GPU-based inference requires installing NVIDIA's CUDA dependencies on the device.
+- TensorRT (version 10 or higher) [**Optional**]
+- Eigen3
+
+- RKNN [**Optional**]
+    - Adjust and select versions currently supported for specific requirements.
+
+## Compilation
+CMake option are used to control the various details of the compilation phase. Please select according to your actual requirements. [CMake Option](doc/CMake-Option.md).
+
+### Local Compilation
+If you are using macOS or Linux, you can quickly compile using the shell scripts provided in the `command` folder at the project root:
+```bash
+cd InspireFace/
+# Execute the local compilation script
+bash command/build.sh
+```
+After compilation, you can find the local file in the build directory, which contains the compilation results. The install directory structure is as follows:
+```bash
+inspireface-linux
+   ├── include
+   │   ├── herror.h
+   │   ├── intypedef.h
+   │   ├── inspireface.h
+   │   ├── inspirecv/
+   │   └── inspireface/
+   └── lib
+       └── libInspireFace.so
+```
+
+- **libInspireFace.so**：Compiled dynamic linking library.
+- **inspireface.h**：Header file definition.
+- **herror.h**：Reference error number definition.
+- **intypedef.h**: Type definition file.
+- **inspirecv**: Simple cv library CPP header file folder.
+- **inspireface**: inspireface cpp header folder.
+### Cross Compilation
+Cross compilation requires you to prepare the target platform's cross-compilation toolchain on the host machine in advance. Here, compiling for Rockchip's embedded devices RV1106 is used as an example:
+```bash
+# Set the path for the cross-compilation toolchain
+export ARM_CROSS_COMPILE_TOOLCHAIN=YOUR_DIR/arm-rockchip830-linux-uclibcgnueabihf
+# Execute the cross-compilation script for RV1106
+bash command/build_cross_rv1106_armhf_uclibc.sh
+```
+After the compilation is complete, you can find the compiled results in the `build/inspireface-linux-armv7-rv1106-armhf-uclibc` directory.
+
+### iOS Compilation
+
+To compile for iOS, ensure you are using a Mac device. The script will automatically download third-party dependencies into the `.macos_cache` directory.
+
+```
+bash command/build_ios.sh
+```
+
+After the compilation is complete, `inspireface.framework` will be placed in the `build/inspireface-ios` directory.
+
+### Android Compilation
+
+You can compile for Android using the following command, but first you need to set your Android NDK path:
+
+```
+export ANDROID_NDK=YOUR_ANDROID_NDK_PATH
+bash command/build_android.sh
+```
+
+After the compilation is complete, arm64-v8a and armeabi-v7a libraries will be placed in the `build/inspireface-android` directory.
+
+### Linux-based NVIDIA GPU Acceleration with TensorRT Compilation
+
+If you want to use NVIDIA GPU devices for accelerated inference on Linux, you need to install **CUDA**, **cuDNN**, and **TensorRT-10** on your device, and configure the relevant environment variables.
+
+```bash
+# Example, Change to your TensorRT-10 path
+export TENSORRT_ROOT=/user/tunm/software/TensorRT-10
+```
+
+Before compiling, please ensure that your related environments such as CUDA and TensorRT-10 are available. If you encounter issues with finding CUDA libraries during the compilation process, you may need to check whether the relevant environment variables have been configured: `CUDA_TOOLKIT_ROOT_DIR`, `CUDA_CUDART_LIBRARY`.
+
+```bash
+bash command/build_linux_tensorrt.sh
+```
+
+Additionally, you can use **NVIDIA's Docker images** for compilation. For example, to compile using a **CUDA 12** and **TensorRT-10** image on Ubuntu 22.04, you can execute the following commands:
+
+```bash
+docker-compose up build-tensorrt-cuda12-ubuntu22
+```
+
+If you want to use pre-compiled libraries, you can use **[FindTensorRT.cmake](toolchain/FindTensorRT.cmake)** to create links to CUDA and TensorRT.
+
+### React Native
+
+For Android and iOS, in addition to the native interface, you can use the React Native library powered by Nitro Modules and JSI—providing ultra-fast, seamless bindings to the InspireFace SDK. For more details, check out the [react-native-nitro-inspire-face](https://github.com/ronickg/react-native-nitro-inspire-face) repository or the [documentation](https://ronickg.github.io/react-native-nitro-inspire-face). Author: ronickg.
+
+### Supported Platforms and Architectures
+
+We have completed the adaptation and testing of the software across various operating systems and CPU architectures. This includes compatibility verification for platforms such as Linux, macOS, iOS, and Android, as well as testing for specific hardware support to ensure stable operation in diverse environments.
+
+| No. | Platform | Architecture<sup><br/>(CPU) | Device<sup><br/>(Special) | **Supported** | Passed Tests | Release<sup><br/>(Online) |
+| ------- | -------------------- | --------------------- | -------------------------- | :-----------: | :----------------: | :----------------: |
+| 1       | **Linux**<sup><br/>(CPU)      | ARMv7                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 2       |                      | ARMv8                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 3       |                      | x86/x86_64            | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 4       | **Linux**<sup><br/>(Rockchip) | ARMv7                 | RV1109/RV1126              | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 5 | | ARMv7 | RV1103/RV1106 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 6 | | ARMv8 | RK3566/RK3568 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 7 | | ARMv8 | RK3588 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 8      | **Linux**<sup><br/>(MNN_CUDA) | x86/x86_64            | NVIDIA-GPU          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - |
+| 9      | **Linux**<sup><br/>(CUDA) | x86/x86_64            | NVIDIA-GPU          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 10      | **MacOS**           | Intel       | CPU/Metal/**ANE** | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 11   |                      | Apple Silicon         | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 12     | **iOS**              | ARM                   | CPU/Metal/**ANE**         | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 13     | **Android**          | ARMv7                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 14     |                      | ARMv8                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 15 | | x86_64 | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 16 | **Android**<sup><br/>(Rockchip) | ARMv8 | RK3566/RK3568 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 17 |  | ARMv8 | RK3588 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![build](https://img.shields.io/github/actions/workflow/status/HyperInspire/InspireFace/release-sdks.yaml?label=✓&labelColor=success&color=success&failedLabel=✗&failedColor=critical&logo=github&logoColor=white)](https://github.com/HyperInspire/InspireFace/actions/workflows/release-sdks.yaml) |
+| 18 | **HarmonyOS** | ARMv8 | - | - | - | - |
+| 19 | **Linux**<sup><br/>(Jetson series) | ARMv8 | Jetson series | - | - | - |
+
+- **Device**: Some special device support, primarily focused on computing power devices.
+- **Supported**: The solution has been fully developed and successfully verified on offline devices.
+- **Passed Tests**: The feature has at least **passed unit tests** on offline devices.
+- **Release**: The solution is already supported and has been successfully compiled and released through **[GitHub Actions](https://github.com/HyperInspire/InspireFace/actions/workflows/built_release_from_docker.yaml)**.
+
+### Multi-platform compilation using Docker
+
+We offer a method for rapid multi-platform compilation using Docker, provided that Docker is installed beforehand, and the appropriate commands are executed:
+```Bash
+# Build x86 Ubuntu18.04
+docker-compose up build-ubuntu18
+
+# Build armv7 cross-compile
+docker-compose up build-cross-armv7-armhf
+
+# Build armv7 with support RV1109RV1126 device NPU cross-complie
+docker-compose up build-cross-rv1109rv1126-armhf
+
+# Build armv7 with support RV1106 device NPU cross-complie
+docker-compose up build-cross-rv1106-armhf-uclibc
+
+# Build armv8 with support RK356x device NPU cross-complie
+docker-compose up build-cross-rk356x-aarch64
+
+# Build Android with support arm64-v8a and armeabi-v7a
+docker-compose up build-cross-android
+
+# Compile the tensorRT back-end based on CUDA12 and then Ubuntu22.04
+docker-compose up build-tensorrt-cuda12-ubuntu22
+
+# Build all
+docker-compose up
+```
+
+## Example
+### C/C++ Sample: Use the recommended CAPI Interface
+To integrate InspireFace into a C/C++ project, you simply need to link the InspireFace library and include the appropriate header files(We recommend using the more compatible **CAPI** headers). Below is a basic example demonstrating face detection:
+
+```c
+#include <inspireface.h>
+#include <herror.h>
+
+...
+  
+HResult ret;
+// The resource file must be loaded before it can be used
+ret = HFLaunchInspireFace(packPath);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+    return ret;
+}
+
+// Enable the functions in the pipeline: mask detection, live detection, and face quality
+// detection
+HOption option = HF_ENABLE_QUALITY | HF_ENABLE_MASK_DETECT | HF_ENABLE_LIVENESS;
+// Non-video or frame sequence mode uses IMAGE-MODE, which is always face detection without
+// tracking
+HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+// Maximum number of faces detected
+HInt32 maxDetectNum = 20;
+// Face detection image input level
+HInt32 detectPixelLevel = 160;
+// Handle of the current face SDK algorithm context
+HFSession session = {0};
+ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Create FaceContext error: %d", ret);
+    return ret;
+}
+
+// Configure some detection parameters
+HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+HFSessionSetFilterMinimumFacePixelSize(session, 4);
+
+// Load a image
+HFImageBitmap image;
+ret = HFCreateImageBitmapFromFilePath(sourcePath, 3, &image);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "The source entered is not a picture or read error.");
+    return ret;
+}
+// Prepare an image parameter structure for configuration
+HFImageStream imageHandle = {0};
+ret = HFCreateImageStreamFromImageBitmap(image, rotation_enum, &imageHandle);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Create ImageStream error: %d", ret);
+    return ret;
+}
+
+// Execute HF_FaceContextRunFaceTrack captures face information in an image
+HFMultipleFaceData multipleFaceData = {0};
+ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Execute HFExecuteFaceTrack error: %d", ret);
+    return ret;
+}
+
+// Print the number of faces detected
+auto faceNum = multipleFaceData.detectedNum;
+HFLogPrint(HF_LOG_INFO, "Num of face: %d", faceNum);
+
+// The memory must be freed at the end of the program
+ret = HFReleaseImageBitmap(image);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Release image bitmap error: %d", ret);
+    return ret;
+}
+
+ret = HFReleaseImageStream(imageHandle);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Release image stream error: %d", ret);
+}
+
+ret = HFReleaseInspireFaceSession(session);
+if (ret != HSUCCEED) {
+    HFLogPrint(HF_LOG_ERROR, "Release session error: %d", ret);
+    return ret;
+}
+
+...
+```
+For more examples, you can refer to the `cpp/sample` sub-project located in the root directory. You can compile these sample executables by enabling the `ISF_BUILD_WITH_SAMPLE` option during the compilation process.
+
+**Note**: For each error code feedback, you can click on this [link](doc/Error-Feedback-Codes.md) to view detailed explanations.
+
+### C++ Sample: Use the C++ version of the header files
+
+If you want to use C++ header files, then you need to enable **ISF_INSTALL_CPP_HEADER** during compilation. Executing the install command will add the C++ header files.
+
+```c++
+#include <iostream>
+#include <memory>
+#include <inspireface/inspireface.hpp>
+
+...
+
+// Set log level to info
+INSPIRE_SET_LOG_LEVEL(inspire::LogLevel::ISF_LOG_INFO);
+
+int32_t ret = 0;
+// Global init(you only need to call once)
+ret = INSPIREFACE_CONTEXT->Load("Pikachu");
+INSPIREFACE_CHECK_MSG(ret == HSUCCEED, "Load model failed");
+
+// Create face algorithm session
+inspire::ContextCustomParameter custom_param;
+custom_param.enable_recognition = true;
+auto max_detect_face = 5;
+auto detect_level_px = 320; // 160, 320, 640
+
+// Create a face algorithm session
+std::shared_ptr<inspire::Session> session(
+    inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, max_detect_face, custom_param, detect_level_px));
+
+// Load image(default format is BGR)
+inspirecv::Image image = inspirecv::Image::Create("face.jpg");
+
+// Create frame process
+inspirecv::FrameProcess process =
+    inspirecv::FrameProcess::Create(image, inspirecv::BGR, inspirecv::ROTATION_0);
+
+// Detect face
+std::vector<inspire::FaceTrackWrap> detect_results;
+ret = session->FaceDetectAndTrack(process, detect_results);
+INSPIRE_LOGI("Number of faces detected: %d", detect_results.size());
+if (detect_results.size() == 0)
+{
+    INSPIRE_LOGW("No face detected");
+    return -1;
+}
+
+// Copy image
+inspirecv::Image image_copy = image.Clone();
+// Draw face
+auto thickness = 2;
+for (auto &face : detect_results)
+{
+    auto rect = session->GetFaceBoundingBox(face);
+    auto lmk = session->GetNumOfFaceDenseLandmark(face);
+    image_copy.DrawRect(rect, inspirecv::Color::Red, thickness);
+    for (auto &point : lmk)
+    {
+        image_copy.DrawCircle(point.As<int>(), 0, inspirecv::Color::Orange, thickness);
+    }
+}
+// Save draw image
+image_copy.Write("result.jpg");
+
+// Face Embedding extract
+inspire::FaceEmbedding face_embedding;
+// Extract the first face feature
+ret = session->FaceFeatureExtract(process, detect_results[0], face_embedding);
+INSPIRE_LOGI("Length of face embedding: %d", face_embedding.embedding.size());
+
+...
+```
+
+Please note that the C++ interface has not been fully tested. It is recommended to use the **CAPI** interface as the primary option.
+
+**More detailed cases**:
+
+- [C Sample](cpp/sample/api/)
+- [C++ Sample](cpp/sample/cpp_api/)
+
+### Python Native Sample
+
+The Python implementation is compiled based on InspireFace source code, and is integrated using a native interface approach.
+
+#### Use pip to install InspireFace
+
+You can use pip to install the InspireFace Python package:
+
+```bash
+pip install inspireface
+```
+
+#### Python Native Sample
+
+We provide a Python API that allows for more efficient use of the InspireFace library. After compiling the dynamic link library, you need to either symlink or copy it to the `python/inspireface/modules/core` directory within the root directory. You can then start testing by navigating to the **[python](python/)** directory. Your Python environment will need to have some dependencies installed:
+
+- python >= 3.7
+- opencv-python
+- loguru
+- tqdm
+- numpy
+- ctypes
+```bash
+# Use a symbolic link
+ln -s YOUR_BUILD_DIR/install/InspireFace/lib/libInspireFace.so python/inspireface/modules/core/PLATFORM/ARCH/
+# Navigate to the sub-project directory
+cd python
+```
+
+Import inspireface for a quick facial detection example:
+```python
+import cv2
+import inspireface as isf
+
+# Step 1: Initialize the SDK globally (only needs to be called once per application)
+ret = isf.reload()
+assert ret, "Launch failure. Please ensure the resource path is correct."
+
+# Optional features, loaded during session creation based on the modules specified.
+opt = isf.HF_ENABLE_FACE_POSE
+session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+
+# Load the image using OpenCV.
+image = cv2.imread(image_path)
+assert image is not None, "Please check that the image path is correct."
+
+# Perform face detection on the image.
+faces = session.face_detection(image)
+print(f"face detection: {len(faces)} found")
+
+# Copy the image for drawing the bounding boxes.
+draw = image.copy()
+for idx, face in enumerate(faces):
+    print(f"{'==' * 20}")
+    print(f"idx: {idx}")
+    # Print Euler angles of the face.
+    print(f"roll: {face.roll}, yaw: {face.yaw}, pitch: {face.pitch}")
+    # Draw bounding box around the detected face.
+    x1, y1, x2, y2 = face.location
+    cv2.rectangle(draw, (x1, y1), (x2, y2), (0, 0, 255), 2)
+```
+In the project, more usage examples are provided:
+
+- `sample_face_detection.py`: Facial detection example
+- `sample_face_recognition.py`: Facial recognition example
+- `sample_face_track_from_video.py`: Facial tracking from video stream example
+
+### Java and Android platform API
+
+We have an [Android SDK project](https://github.com/HyperInspire/inspireface-android-sdk) that integrates pre-compiled dynamic libraries, and you can use it directly.
+
+Precompiled library support: 
+- arm64-v8a
+- armeabi-v7a
+- x86_64
+
+#### a. Quick to use in Android
+
+We released InspireFace's Android SDK on JitPack, which you can incorporate into your android projects in the following ways.
+
+- Step 1. Add the JitPack repository to your build file add it in your root **build.gradle** at the end of repositories:
+
+  ```groovy
+  allprojects {
+      repositories {
+         ...
+         maven { url 'https://jitpack.io' }
+      }
+  }
+  ```
+
+- Step 2. Add the dependency
+
+  ```groovy
+  dependencies {
+      implementation 'com.github.HyperInspire:inspireface-android-sdk:1.2.3.post4'
+  }
+  ```
+
+#### b. Use the Android example project
+
+We have prepared an  [Android SDK project](https://github.com/HyperInspire/inspireface-android-sdk). You can download library from the [Release Page](https://github.com/HyperInspire/InspireFace/releases) or compile the Android library yourself and place it in the `inspireface/libs` directory of the Android sample project. You can compile and run this project using Android Studio.
+
+```bash
+inspireface-android-sdk/inspireface/libs
+├── arm64-v8a
+│   └── libInspireFace.so
+└── armeabi-v7a
+    └── libInspireFace.so
+```
+
+You need to get the resource file from the release  [Release Page](https://github.com/HyperInspire/InspireFace/releases) and place it in the `asset/inspireface` in your android project:
+
+```
+asset/
+└── inspireface/
+    │── Pikachu
+    │── Megatron
+    │── Gundam_RK356X
+    └── Gundam_RK3588
+```
+
+#### How to use the Android/Java API
+
+We provide a Java API for Android devices, which is implemented using Java Native Interface(JNI). 
+
+```java
+// Launch InspireFace, only need to call once
+boolean launchStatus = InspireFace.GlobalLaunch(this, InspireFace.PIKACHU);
+if (!launchStatus) {
+    Log.e(TAG, "Failed to launch InspireFace");
+}
+
+// Create a ImageStream
+ImageStream stream = InspireFace.CreateImageStreamFromBitmap(img, InspireFace.CAMERA_ROTATION_0);
+
+// Create a session
+CustomParameter parameter = InspireFace.CreateCustomParameter()
+                .enableRecognition(true)
+                .enableFaceQuality(true)
+                .enableFaceAttribute(true)
+                .enableInteractionLiveness(true)
+                .enableLiveness(true)
+                .enableMaskDetect(true);
+Session session = InspireFace.CreateSession(parameter, InspireFace.DETECT_MODE_ALWAYS_DETECT, 10, -1, -1);
+
+// Execute face detection
+MultipleFaceData multipleFaceData = InspireFace.ExecuteFaceTrack(session, stream);
+if (multipleFaceData.detectedNum > 0) {
+    // Get face feature
+    FaceFeature feature = InspireFace.ExtractFaceFeature(session, stream, multipleFaceData.tokens[0]);
+    // ....
+}
+
+// .... 
+
+// Release resource
+InspireFace.ReleaseSession(session);
+InspireFace.ReleaseImageStream(stream);
+
+// Global release
+InspireFace.GlobalRelease();
+```
+
+## Test
+
+In the project, there is a subproject called `cpp/test`. To compile it, you need to enable the `ISF_BUILD_WITH_TEST` switch, which will allow you to compile executable programs for testing.
+
+```bash
+cmake -DISF_BUILD_WITH_TEST=ON ..
+```
+To run the test modules in the project, first check if the resource files exist in the test_res/pack directory. If they don't exist, you can either execute **command/download_models_general.sh** to download the required files, or download the files from the [Release Page](https://github.com/HyperInspire/InspireFace/releases/tag/v1.x) and manually place them in this directory.
+
+```bash
+
+test_res
+├── data
+├── images
+├── pack	<-- The model package files are here
+├── save
+├── valid_lfw_funneled.txt
+├── video
+└── video_frames
+
+```
+After compilation, you can find the executable program "**Test**" in `YOUR_BUILD_FOLDER/test`. The program accepts two optional parameters:
+
+- **test_dir**：Path to the test resource files
+- **pack**：Name of the model to be tested
+```bash
+./Test --test_dir PATH/test_res --pack Pikachu
+```
+During the process of building the test program using CMake, it will involve selecting CMake parameters. For specific details, you can refer to the parameter configuration table.
+
+**Note**: If you want to view the benchmark test report, you can click on the [link](doc/Benchmark-Remark(Updating).md).
+
+### Quick Test
+
+If you need to perform a quick test, you can use the script we provide. This script will automatically download the test file `test_res` and build the test program to run the test. 
+
+*Note: If you need to enable more comprehensive tests, you can adjust the options in the script as needed.*
+
+```bash
+# If you are using Ubuntu, you can execute this.
+bash ci/quick_test_linux_x86_usual.sh
+
+# If you are using another system (including Ubuntu), you can execute this.
+bash ci/quick_test_local.sh
+```
+
+Every time code is committed, tests are run on GitHub Actions.
+
+## Features
+The following Features and technologies are currently supported.
+
+| Feature | CPU | RKNPU<sup><br/>(RV1109/1126) | RKNPU<sup><br/>(RV1103/1106) | RKNPU<sup><br/>(RK3566/3568/3588) | ANE<sup><br/>(MacOS/iOS) | GPU<sup><br/>(TensorRT) |
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| Face Detection | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Landmark | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Embeddings | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Comparison | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+| Face Recognition | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Alignment | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+| Tracking | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Mask Detection | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - |
+| Silent Liveness | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - |
+| Face Quality | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Pose Estimation | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Attribute | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Cooperative Liveness | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Emotion<sup>New</sup> | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Embedding Management | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+
+- Some models and features that do **not support** NPU or GPU will **automatically use CPU** for computation when running the program.
+
+## Resource Package List
+
+For different scenarios, we currently provide several Packs, each containing multiple models and configurations.
+
+| Name | Supported Devices | Note | Last Update | Link |
+| --- | --- | --- | --- | --- |
+| Pikachu | CPU | Lightweight edge-side models | Jun 22, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Pikachu) |
+| Megatron | CPU, GPU | Mobile and server models | Jun 15, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Megatron) |
+| Megatron_TRT | GPU | CUDA-based server models | Jun 15, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Megatron_TRT) |
+| Gundam-RV1109 | RKNPU | Supports RK1109 and RK1126 | Jun 15, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RV1109) |
+| Gundam-RV1106 | RKNPU | Supports RV1103 and RV1106 | Jul 6, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RV1106) |
+| Gundam-RK356X | RKNPU | Supports RK3566 and RK3568 | Jun 15, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RK356X) |
+| Gundam-RK3588 | RKNPU | Supports RK3588 | Jun 15, 2025 | [Download](https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RK3588) |
+
+## Short-Term Plan
+
+- [x] Add TensorRT backend support.
+- [x] Add Add C++ style header files.
+- [x] Add the RKNPU backend support for Android .
+- [ ] Python packages that support more platforms.
+- [ ] Example app project for Android and iOS samples.
+- [ ] Add the batch forward feature.
+- [ ] Design a scheme that can be adapted to multiple CUDA devices.
+
+- Continue to provide more support for Rockchip NPU2 devices:
+- [ ] RK3576 Series
+- [ ] RK3562 Series
+- [ ] RV1103B/RV1106B
+- [ ] RV1126B
+- [ ] RK2118
+
+## Acknowledgement
+
+InspireFace is built on the following libraries:
+
+- [MNN](https://github.com/alibaba/MNN)
+- [RKNN](https://github.com/rockchip-linux/rknn-toolkit)
+- [RKNN2](https://github.com/airockchip/rknn-toolkit2.git)
+- [librga](https://github.com/airockchip/librga.git)
+- [Eigen](https://eigen.tuxfamily.org/index.php?title=Main_Page)
+- [sqlite](https://www.sqlite.org/index.html)
+- [sqlite-vec](https://github.com/asg017/sqlite-vec)
+- [Catch2](https://github.com/catchorg/Catch2)
+- [yaml-cpp](https://github.com/jbeder/yaml-cpp)
+- [TensorRT](https://github.com/NVIDIA/TensorRT)
diff --git a/cpp-package/inspireface/android/InspireFaceExample/.gitignore b/cpp-package/inspireface/android/InspireFaceExample/.gitignore
new file mode 100644
index 000000000..0ddb38199
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/.gitignore
@@ -0,0 +1,26 @@
+# --- Gradle ---
+.gradle/
+build/
+/captures
+
+# --- Local machine config (regenerated by Android Studio) ---
+local.properties
+
+# --- IDE ---
+.idea/
+*.iml
+*.hprof
+
+# --- Native build intermediates ---
+.externalNativeBuild/
+.cxx/
+
+# --- OS noise ---
+.DS_Store
+Thumbs.db
+
+# --- Signing & release artifacts (keep keys out of the repo) ---
+*.jks
+*.keystore
+/keystore.properties
+/app/release/
diff --git a/cpp-package/inspireface/android/InspireFaceExample/README.md b/cpp-package/inspireface/android/InspireFaceExample/README.md
new file mode 100644
index 000000000..0f5dd2b3f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/README.md
@@ -0,0 +1,179 @@
+# InspireFace Android Example
+
+A CameraX-based InspireFace Android SDK (1.2.0) example. The launcher is a square-grid
+feature menu with a global model selector (`Pikachu` / `Megatron`). The selected model is
+loaded when a feature page opens and is shown in a small label on every feature page.
+
+## Try the Android app
+
+<p>
+  <a href="http://fir.tunm.top/pro/pz7b3dgv">
+    <img src="docs/images/inspireface-android-example-app-download.png" width="220" alt="Download the InspireFace Android example">
+  </a>
+</p>
+
+<p>
+  <strong>Download the Android example and try it now.</strong><br>
+  Scan the QR code or <strong><a href="http://fir.tunm.top/pro/pz7b3dgv">Download App</a></strong>.
+</p>
+
+The current menu contains eight dedicated pages:
+
+- **Silent liveness (RGB anti-spoofing)**: streams a per-frame liveness score for the current face, averages it over a sliding window, and labels the face real/spoof against a threshold.
+- **Action liveness (cooperative)**: generates a random challenge sequence (blink / head shake / mouth open / head raise), prompts each action in turn, and detects completion with per-action timeouts and face-loss failure handling.
+- **Pose recognition**: displays whatever action you perform — the latest action in large text, with a fading row of smaller history entries below (at most 6 shown; rising-edge debounced, so a held pose is recorded once).
+- **Face 1:1**: selects two local images, detects and numbers every face, selects face 1 by default, and lets you tap any A/B face box to immediately rerun the comparison. The circular gauge shows the converted similarity percentage and the SDK-recommended threshold verdict.
+- **Face management**: searches, adds, renames, replaces, and deletes identities in the currently selected model library. Enrollment supports either gallery multi-face selection or an automatic camera flow that tracks face 1, waits for 1 stable second, then fills a 2-second red/yellow/green ring. Motion immediately resets and hides the ring.
+- **Face recognition**: switches between Photo input and Video stream tabs. Photo mode detects and numbers faces, searches a single face immediately, and reruns the search when a numbered face is tapped. Its collapsed-by-default settings panel sits below the photo picker; detection input px, maximum face count, and minimum face px rebuild the Session and persist locally. Video mode reuses the CameraX tracking pipeline, tracks only face 0, searches after roughly one stable second, renders the result at the bottom, and supports front/rear cameras.
+- **Face tracking**: detects and numbers every image face, then displays the selected face's SDK-native 106 dense landmarks. Medium and large faces are annotated directly; genuinely small faces use a bottom-right 148dp magnifier cropped to 2.4× the detected box. The Video tracking tab renders track-ID colors, four-corner boxes and 106 points with OpenGL.
+- **Face attributes**: analyzes a selected image face for mask state, age bracket, image quality, expression state, ethnicity, gender and left/right eye state. Tapping another numbered face updates the result immediately, and small faces reuse the expanded bottom-right crop magnifier.
+
+Debug aids on each camera page:
+
+- **Euler angles** switch: live Yaw / Pitch / Roll readout for the tracked face (~10 Hz; note that in the 1.2.0 JNI only `angles[0]` is trustworthy, so with multiple faces the first face is shown).
+- **Landmarks** rendering is currently hidden; `LandmarkGlView` remains available as an OpenGL ES 2.0 overlay for a future menu entry or debug switch.
+
+A **Flip camera** chip switches between the front and back lens at runtime. No SDK-side
+changes are needed for that: every frame is pre-rotated by its own `rotationDegrees`
+before being handed to InspireFace as an upright `CAMERA_ROTATION_0` buffer, so the new
+lens's sensor orientation is absorbed per frame — only the display mirroring flips, and
+the mode state machine restarts.
+
+Language: the first launch defaults to **English** regardless of the system language. The
+language chip switches between English and Chinese, and Android 13+ also exposes both in
+the system per-app language settings. The selected language persists across launches.
+
+## Architecture
+
+```
+CameraX ImageAnalysis (YUV_420_888, 640x480, KEEP_ONLY_LATEST)
+   └─ UprightFaceCameraAnalyzer (single-threaded analysis executor)
+        ├─ Nv21Converter.convert()        YUV_420_888 → tight NV21 (VU interleave probed once, then bulk-copied)
+        ├─ Nv21Converter.rotateUpright()  Java-side rotation to upright
+        ├─ CreateImageStreamFromByteBuffer(nv21, CAMERA_ROTATION_0)
+        ├─ ExecuteFaceTrack               LIGHT_TRACK mode
+        ├─ FaceAnalyzer / EnrollmentFaceAnalyzer
+        │    └─ mode pipeline or first-face stability state machine
+        └─ ReleaseImageStream             released within the same frame
+```
+
+- `HomeActivity` — square-grid feature menu and global model selection
+- `FaceCompareActivity` — local image decoding, face feature extraction and 1:1 comparison
+- `FaceManagementActivity` — CRUD UI for model-isolated identities, crops and FeatureHub data
+- `FaceRecognitionActivity` / `StillImageSessionSettings` — photo multi-face selection, model-scoped 1:N search and persisted Session parameters
+- `view/RecognitionFaceAnalyzer` — video face-0 stability gate, feature extraction and model-library search
+- `FaceDetectionActivity` / `widget/FaceLandmarkOverlayView` — image multi-face detection, 106-point overlay and small-face magnifier
+- `FaceAttributeActivity` / `face/FaceAttributeProcessor` — selectable still-image mask, quality, demographic and interaction attributes
+- `view/FaceCaptureActivity` / `EnrollmentFaceAnalyzer` — first-face stable camera enrollment and automatic capture
+- `view/CameraPreviewController` — reusable CameraX preview, 4:3 analysis, lens fallback and front/rear switching
+- `view/UprightFaceCameraAnalyzer` — shared YUV→upright NV21, face tracking and native stream/session lifecycle
+- `face/FaceImageProcessor` / `face/FaceCropUtils` / `widget/FaceImageOverlayView` — shared multi-face extraction, expanded crop and tappable numbered boxes
+- `face/FaceRepository` — model-scoped persistent FeatureHub, crop files and metadata
+- `view/LivenessActivity` — shared CameraX screen and the silent-liveness entry
+- `view/ActionLivenessActivity` / `PoseActivity` — dedicated routes that select their fixed controller mode
+- `view/FaceAnalyzer` — liveness-mode pipeline, performance stats and debug readouts
+- `view/LivenessController` — the state machines for all three modes (tunables live at the top of this class)
+- `view/Nv21Converter` — fast YUV→NV21 conversion + NV21 rotation
+- `view/FaceOverlayView` — face bracket overlay (center-crop mapping + front mirror)
+- `view/LandmarkGlView` — OpenGL landmark overlay
+- `view/FaceEngine` — model-aware GlobalLaunch/GlobalTerminate and session creation
+- `FaceModelPrefs` / `LocalePrefs` / `App` — persisted global model and per-app language
+
+## Model-isolated face storage
+
+Pikachu and Megatron never share face features, crop images, metadata, or ID sequences.
+The app stores them under separate app-private paths:
+
+```text
+files/face_hub/Pikachu/features.db
+files/face_hub/Pikachu/crops/
+shared_prefs/face_records_Pikachu.xml
+
+files/face_hub/Megatron/features.db
+files/face_hub/Megatron/crops/
+shared_prefs/face_records_Megatron.xml
+```
+
+FeatureHub uses manual primary keys and persistent storage. Switching the global model
+therefore opens a different native database as well as a different crop/metadata set.
+
+## Key design decisions (verified against SDK source)
+
+1. **Pre-rotate NV21 on the Java side and always pass `CAMERA_ROTATION_0`.**
+   The SDK (≤1.2.3) crops RGB-liveness input using "the rotated upright full frame + an
+   un-rotated face rect", so passing 90/270 rotation constants misplaces the crop and
+   corrupts silent-liveness scores. Pre-rotating (~1–2 ms at 640×480) sidesteps that
+   entirely, and every SDK output coordinate lands directly in display orientation, so
+   the overlays only need the front-camera mirror. Also note the SDK's rotation
+   constants are the *opposite* of Android's `rotationDegrees` (Android 90 → SDK
+   ROTATION_270); pre-rotation avoids that trap too.
+
+2. **Action liveness has three hard prerequisites** (miss one and actions never fire):
+   - `DETECT_MODE_LIGHT_TRACK` (other modes rebuild tracked faces every frame, so the
+     temporal action window never accumulates);
+   - `enableInteractionLiveness` at session creation;
+   - `enableFaceQuality` at session creation (loads the pose model — without it yaw/pitch
+     stay 0 and shake/head-raise can never trigger).
+
+3. **Action flag semantics** (SDK-internal 10-frame sliding window + rules):
+   blink is a one-call pulse (the window resets after it); shake latches while both yaw
+   extremes sit in the rolling window (~10 calls); mouth-open/head-raise are
+   level-triggered. The controller therefore uses **edge gates**: each challenge step
+   must observe the flag at 0 before a 1 counts, and the SDK's `normal` flag (warm-up
+   indicator, also raised for ~9 calls after every blink-induced reset) quarantines the
+   placeholder zeros so a pose held through a natural blink is neither double-counted
+   (pose mode) nor accepted as fresh (action mode).
+
+4. **`CreateImageStreamFromByteBuffer` does not copy** — the native stream aliases the
+   byte[] until `ReleaseImageStream`. Safe pattern: create → track → pipeline → release
+   within one frame, never overwriting the byte[] before release (this implementation
+   reuses two persistent buffers on a single thread, which satisfies that naturally).
+
+5. **Silent liveness**: single-frame score with the author-encoded 0.88 decision
+   boundary; every pipeline call converts the full frame internally (a known SDK hot
+   spot), so the pipeline runs every 2nd frame with an 8-sample sliding average — same
+   accuracy, half the cost.
+
+6. **`MultipleFaceData.angles[i]` is only valid for `i == 0`** (a 1.2.0 JNI bug writes
+   face[0]'s angles into every slot); this app only reads it in single-face flows.
+
+## Tunables
+
+At the top of `LivenessController`:
+
+| Parameter | Default | Meaning |
+|---|---|---|
+| `RGB_LIVENESS_THRESHOLD` | 0.88 | silent-liveness real/spoof boundary |
+| `SCORE_WINDOW` | 8 | sliding average window for scores |
+| `SILENT_PIPELINE_INTERVAL` | 2 | run the anti-spoofing pipeline every N frames |
+| `ACTIONS_PER_RUN` | 3 | challenge actions per round |
+| `ACTION_TIMEOUT_MS` | 8000 | per-action timeout |
+| `MIN_FACE_WIDTH_RATIO` | 0.18 | minimum face width as a fraction of frame width |
+| `POSE_HISTORY_MAX` | 6 | pose-mode entries shown (1 large + 5 history) |
+
+## Requirements
+
+- JDK 17 (required by AGP 8.6.1; Android Studio's embedded JDK works)
+- Android Studio Ladybug+ — the Gradle 8.7 wrapper is committed, no local Gradle needed
+- Network access to `google()`, `mavenCentral()` and `jitpack.io` on first sync
+  (the InspireFace SDK and its bundled model packs resolve from JitPack)
+- An ARM Android device running Android 7.0 / API 24 or newer. The app compiles and targets
+  Android 15 / API 35; Android has no declared upper install limit.
+- The SDK ships arm64-v8a / armeabi-v7a only, so x86/x86_64 emulators and Intel-only
+  ChromeOS devices cannot run the native face engine. The arm64 native libraries and the
+  compatibility bridge are built/aligned for Android 15's 16 KB page-size devices.
+
+`local.properties` is intentionally not committed; Android Studio regenerates it, or set
+`ANDROID_HOME` for command-line builds.
+
+## Running
+
+The first feature launch is slower while the bundled model packs are unpacked from assets.
+
+```bash
+./gradlew :app:installDebug
+```
+
+This project now covers liveness, pose, 1:1 comparison, face management, and FeatureHub
+1:N photo search. For other SDK capabilities, continue with the upstream
+[InspireFace](https://github.com/HyperInspire/InspireFace) Android example.
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/.gitignore b/cpp-package/inspireface/android/InspireFaceExample/app/.gitignore
new file mode 100644
index 000000000..42afabfd2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/.gitignore
@@ -0,0 +1 @@
+/build
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/build.gradle b/cpp-package/inspireface/android/InspireFaceExample/app/build.gradle
new file mode 100644
index 000000000..66118f1f4
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/build.gradle
@@ -0,0 +1,84 @@
+plugins {
+    alias(libs.plugins.android.application)
+}
+
+def releaseKeystoreProperties = new Properties()
+def releaseKeystorePropertiesFile = rootProject.file("keystore.properties")
+def hasReleaseKeystore = releaseKeystorePropertiesFile.exists()
+if (hasReleaseKeystore) {
+    releaseKeystorePropertiesFile.withInputStream {
+        releaseKeystoreProperties.load(it)
+    }
+}
+
+android {
+    namespace 'com.example.inspireface_example'
+    compileSdk 35
+    ndkVersion '28.1.13356709'
+
+    defaultConfig {
+        applicationId "com.example.inspireface_example"
+        minSdk 24
+        targetSdk 35
+        versionCode 1
+        versionName "1.0.0"
+
+        ndk {
+            // InspireFace 1.2.0 ships these two ABIs only.
+            abiFilters 'arm64-v8a', 'armeabi-v7a'
+        }
+
+        testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
+    }
+
+    signingConfigs {
+        release {
+            if (hasReleaseKeystore) {
+                storeFile rootProject.file(releaseKeystoreProperties['storeFile'])
+                storePassword releaseKeystoreProperties['storePassword']
+                storeType releaseKeystoreProperties.getProperty('storeType', 'JKS')
+                keyAlias releaseKeystoreProperties['keyAlias']
+                keyPassword releaseKeystoreProperties['keyPassword']
+                enableV1Signing true
+                enableV2Signing true
+                enableV3Signing true
+                enableV4Signing true
+            }
+        }
+    }
+    buildTypes {
+        release {
+            if (hasReleaseKeystore) {
+                signingConfig signingConfigs.release
+            }
+            minifyEnabled false
+            proguardFiles getDefaultProguardFile('proguard-android-optimize.txt'), 'proguard-rules.pro'
+        }
+    }
+    compileOptions {
+        sourceCompatibility JavaVersion.VERSION_1_8
+        targetCompatibility JavaVersion.VERSION_1_8
+    }
+    externalNativeBuild {
+        ndkBuild {
+            path file('src/main/cpp/Android.mk')
+        }
+    }
+}
+
+dependencies {
+
+    implementation libs.appcompat
+    implementation libs.material
+    implementation libs.activity
+    implementation libs.constraintlayout
+    implementation libs.camera.core
+    implementation libs.camera.camera2
+    implementation libs.camera.lifecycle
+    implementation libs.camera.view
+    implementation libs.exifinterface
+    testImplementation libs.junit
+    androidTestImplementation libs.ext.junit
+    androidTestImplementation libs.espresso.core
+    implementation libs.inspireface.android.sdk
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/proguard-rules.pro b/cpp-package/inspireface/android/InspireFaceExample/app/proguard-rules.pro
new file mode 100644
index 000000000..481bb4348
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/proguard-rules.pro
@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java
new file mode 100644
index 000000000..bfd1c0ded
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java
@@ -0,0 +1,26 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+
+import androidx.test.platform.app.InstrumentationRegistry;
+import androidx.test.ext.junit.runners.AndroidJUnit4;
+
+import org.junit.Test;
+import org.junit.runner.RunWith;
+
+import static org.junit.Assert.*;
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * @see <a href="http://d.android.com/tools/testing">Testing documentation</a>
+ */
+@RunWith(AndroidJUnit4.class)
+public class ExampleInstrumentedTest {
+    @Test
+    public void useAppContext() {
+        // Context of the app under test.
+        Context appContext = InstrumentationRegistry.getInstrumentation().getTargetContext();
+        assertEquals("com.example.inspireface_example", appContext.getPackageName());
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/AndroidManifest.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/AndroidManifest.xml
new file mode 100644
index 000000000..41a999a5d
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/AndroidManifest.xml
@@ -0,0 +1,76 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:tools="http://schemas.android.com/tools">
+
+    <uses-permission android:name="android.permission.CAMERA" />
+    <uses-feature
+        android:name="android.hardware.camera"
+        android:required="false" />
+    <uses-feature
+        android:name="android.hardware.camera.autofocus"
+        android:required="false" />
+    <uses-feature
+        android:name="android.hardware.camera.front"
+        android:required="false" />
+
+    <application
+        android:name=".App"
+        android:allowBackup="true"
+        android:dataExtractionRules="@xml/data_extraction_rules"
+        android:fullBackupContent="@xml/backup_rules"
+        android:icon="@mipmap/ic_launcher"
+        android:label="@string/app_name"
+        android:localeConfig="@xml/locales_config"
+        android:roundIcon="@mipmap/ic_launcher_round"
+        android:supportsRtl="true"
+        android:theme="@style/Theme.InspireFaceExample"
+        tools:targetApi="33">
+        <activity
+            android:name=".view.FaceCaptureActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Liveness" />
+        <activity
+            android:name=".FaceManagementActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Home" />
+        <activity
+            android:name=".FaceRecognitionActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Home" />
+        <activity
+            android:name=".FaceDetectionActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Home" />
+        <activity
+            android:name=".FaceAttributeActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Home" />
+        <activity
+            android:name=".FaceCompareActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Home" />
+        <activity
+            android:name=".view.PoseActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Liveness" />
+        <activity
+            android:name=".view.ActionLivenessActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Liveness" />
+        <activity
+            android:name=".view.LivenessActivity"
+            android:exported="false"
+            android:theme="@style/Theme.InspireFaceExample.Liveness" />
+        <activity
+            android:name=".HomeActivity"
+            android:exported="true"
+            android:theme="@style/Theme.InspireFaceExample.Home">
+            <intent-filter>
+                <action android:name="android.intent.action.MAIN" />
+
+                <category android:name="android.intent.category.LAUNCHER" />
+            </intent-filter>
+        </activity>
+    </application>
+
+</manifest>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Android.mk b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Android.mk
new file mode 100644
index 000000000..890ef6876
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Android.mk
@@ -0,0 +1,8 @@
+LOCAL_PATH := $(call my-dir)
+
+include $(CLEAR_VARS)
+LOCAL_MODULE := inspireface_session_bridge
+LOCAL_SRC_FILES := inspireface_session_bridge.cpp
+LOCAL_CPPFLAGS := -std=c++17 -Wall -Wextra
+LOCAL_LDLIBS := -ldl -llog
+include $(BUILD_SHARED_LIBRARY)
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Application.mk b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Application.mk
new file mode 100644
index 000000000..e9e67aec5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/Application.mk
@@ -0,0 +1,4 @@
+APP_ABI := arm64-v8a armeabi-v7a
+APP_PLATFORM := android-24
+APP_STL := c++_static
+APP_SUPPORT_FLEXIBLE_PAGE_SIZES := true
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/inspireface_session_bridge.cpp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/inspireface_session_bridge.cpp
new file mode 100644
index 000000000..c67521ff0
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/cpp/inspireface_session_bridge.cpp
@@ -0,0 +1,50 @@
+#include <android/log.h>
+#include <dlfcn.h>
+#include <jni.h>
+
+namespace {
+
+constexpr const char* kLogTag = "InspireFaceBridge";
+using CreateSessionOptional = long (*)(int custom_option,
+                                       int detect_mode,
+                                       int max_detect_face_num,
+                                       int detect_pixel_level,
+                                       int track_by_detect_fps,
+                                       void** session);
+
+}  // namespace
+
+extern "C" JNIEXPORT jlong JNICALL
+Java_com_example_inspireface_1example_view_NativeSessionBridge_nativeCreateLandmarkSession(
+        JNIEnv*, jclass, jint custom_options, jint detect_mode, jint max_faces,
+        jint detect_pixel_level, jint track_by_detect_fps) {
+    void* library = dlopen("libInspireFace.so", RTLD_NOW | RTLD_LOCAL);
+    if (library == nullptr) {
+        __android_log_print(ANDROID_LOG_ERROR, kLogTag,
+                            "Could not load libInspireFace.so: %s", dlerror());
+        return 0;
+    }
+
+    dlerror();
+    auto create_session = reinterpret_cast<CreateSessionOptional>(
+            dlsym(library, "HFCreateInspireFaceSessionOptional"));
+    const char* symbol_error = dlerror();
+    if (symbol_error != nullptr || create_session == nullptr) {
+        __android_log_print(ANDROID_LOG_ERROR, kLogTag,
+                            "Optional Session API unavailable: %s",
+                            symbol_error == nullptr ? "unknown error" : symbol_error);
+        dlclose(library);
+        return 0;
+    }
+
+    void* session = nullptr;
+    long result = create_session(custom_options, detect_mode, max_faces,
+                                 detect_pixel_level, track_by_detect_fps, &session);
+    dlclose(library);
+    if (result != 0 || session == nullptr) {
+        __android_log_print(ANDROID_LOG_ERROR, kLogTag,
+                            "Could not create landmark Session, error=%ld", result);
+        return 0;
+    }
+    return reinterpret_cast<jlong>(session);
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/App.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/App.java
new file mode 100644
index 000000000..5d2a0e4aa
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/App.java
@@ -0,0 +1,12 @@
+package com.example.inspireface_example;
+
+import android.app.Application;
+
+public class App extends Application {
+
+    @Override
+    public void onCreate() {
+        super.onCreate();
+        LocalePrefs.applyStored(this);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/DetectorDefaults.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/DetectorDefaults.java
new file mode 100644
index 000000000..744059842
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/DetectorDefaults.java
@@ -0,0 +1,10 @@
+package com.example.inspireface_example;
+
+/** Shared defaults used by every still-image and camera face detector. */
+public final class DetectorDefaults {
+
+    public static final int INPUT_PX = 320;
+
+    private DetectorDefaults() {
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceAttributeActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceAttributeActivity.java
new file mode 100644
index 000000000..25d555e53
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceAttributeActivity.java
@@ -0,0 +1,392 @@
+package com.example.inspireface_example;
+
+import android.graphics.Bitmap;
+import android.net.Uri;
+import android.os.Bundle;
+import android.view.View;
+import android.widget.ImageView;
+import android.widget.TextView;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.annotation.ColorRes;
+import androidx.annotation.Nullable;
+import androidx.annotation.StringRes;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.content.ContextCompat;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.face.FaceAttributeProcessor;
+import com.example.inspireface_example.face.FaceCropUtils;
+import com.example.inspireface_example.face.FaceImageProcessor;
+import com.example.inspireface_example.face.ImageBitmapLoader;
+import com.example.inspireface_example.view.FaceEngine;
+import com.example.inspireface_example.widget.FaceImageOverlayView;
+import com.google.android.material.button.MaterialButton;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** Selectable still-image face attribute analysis. */
+public final class FaceAttributeActivity extends AppCompatActivity {
+
+    private static final int MAX_IMAGE_DIMENSION = 2048;
+    private static final float DIRECT_FACE_AREA_RATIO = 0.05f;
+    private static final float MAGNIFIER_CROP_SCALE = 2.4f;
+    private static final float BINARY_CONFIDENCE_THRESHOLD = 0.5f;
+
+    private final ExecutorService sdkExecutor = Executors.newSingleThreadExecutor();
+
+    private Session session;
+    private volatile boolean destroyed;
+    private volatile int imageVersion;
+    private boolean busy = true;
+
+    private MaterialButton choosePhotoButton;
+    private ImageView imageView;
+    private View imagePlaceholder;
+    private TextView imageStatus;
+    private FaceImageOverlayView faceOverlay;
+    private View magnifierCard;
+    private ImageView magnifierImage;
+    private View resultValues;
+    private TextView resultPlaceholder;
+    private TextView maskValue;
+    private TextView ageValue;
+    private TextView qualityValue;
+    private TextView expressionValue;
+    private TextView raceValue;
+    private TextView genderValue;
+    private TextView leftEyeValue;
+    private TextView rightEyeValue;
+
+    private Bitmap imageBitmap;
+    private Bitmap magnifierBitmap;
+    private FaceAttributeProcessor.Result analysis;
+    private int selectedIndex = -1;
+
+    private final ActivityResultLauncher<String> photoPicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null && !busy && session != null) {
+                    loadPhoto(uri);
+                }
+            });
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_face_attribute);
+        applyWindowInsets();
+        bindViews();
+
+        ((TextView) findViewById(R.id.currentModel)).setText(getString(
+                R.string.current_model, FaceModelPrefs.get(this).sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        choosePhotoButton.setOnClickListener(v -> photoPicker.launch("image/*"));
+        findViewById(R.id.attributeImageCard).setOnClickListener(v -> {
+            if (!busy && session != null) {
+                photoPicker.launch("image/*");
+            }
+        });
+        faceOverlay.setOnFaceSelectedListener(this::selectFace);
+        setControlsEnabled(false);
+        showResultPlaceholder(R.string.attribute_result_empty);
+        sdkExecutor.execute(this::initializeEngine);
+    }
+
+    private void bindViews() {
+        choosePhotoButton = findViewById(R.id.btnChooseAttributePhoto);
+        imageView = findViewById(R.id.attributeImage);
+        imagePlaceholder = findViewById(R.id.attributeImagePlaceholder);
+        imageStatus = findViewById(R.id.attributeImageStatus);
+        faceOverlay = findViewById(R.id.attributeFaceOverlay);
+        magnifierCard = findViewById(R.id.attributeMagnifierCard);
+        magnifierImage = findViewById(R.id.attributeMagnifierImage);
+        resultValues = findViewById(R.id.attributeResultValues);
+        resultPlaceholder = findViewById(R.id.attributeResultPlaceholder);
+        maskValue = findViewById(R.id.attributeMaskValue);
+        ageValue = findViewById(R.id.attributeAgeValue);
+        qualityValue = findViewById(R.id.attributeQualityValue);
+        expressionValue = findViewById(R.id.attributeExpressionValue);
+        raceValue = findViewById(R.id.attributeRaceValue);
+        genderValue = findViewById(R.id.attributeGenderValue);
+        leftEyeValue = findViewById(R.id.attributeLeftEyeValue);
+        rightEyeValue = findViewById(R.id.attributeRightEyeValue);
+    }
+
+    private void initializeEngine() {
+        boolean ready = FaceEngine.ensureLaunched(this);
+        if (ready) {
+            session = FaceEngine.createAttributeSession();
+            ready = session != null;
+        }
+        boolean finalReady = ready;
+        runOnUiThread(() -> {
+            if (destroyed) {
+                return;
+            }
+            busy = false;
+            setControlsEnabled(finalReady);
+            if (finalReady) {
+                setImageStatus(R.string.no_image_selected, R.color.home_text_secondary);
+            } else {
+                setImageStatus(R.string.compare_engine_failed, R.color.liveness_fail);
+                showResultPlaceholder(R.string.attribute_analysis_failed);
+            }
+        });
+    }
+
+    private void loadPhoto(Uri uri) {
+        int requestVersion = ++imageVersion;
+        busy = true;
+        analysis = null;
+        selectedIndex = -1;
+        faceOverlay.clearFace();
+        clearMagnifier();
+        showResultPlaceholder(R.string.attribute_analyzing);
+        setImageStatus(R.string.attribute_analyzing, R.color.home_text_secondary);
+        setControlsEnabled(false);
+
+        sdkExecutor.execute(() -> {
+            Bitmap bitmap = null;
+            FaceAttributeProcessor.Result result = null;
+            try {
+                bitmap = ImageBitmapLoader.decode(this, uri, MAX_IMAGE_DIMENSION);
+                if (destroyed || imageVersion != requestVersion) {
+                    recycle(bitmap);
+                    return;
+                }
+                if (session != null) {
+                    result = FaceAttributeProcessor.analyze(session, bitmap);
+                }
+            } catch (Exception ignored) {
+                // Null result is presented as a load or analysis failure.
+            }
+            Bitmap deliveredBitmap = bitmap;
+            FaceAttributeProcessor.Result deliveredResult = result;
+            runOnUiThread(() -> applyResult(
+                    requestVersion, deliveredBitmap, deliveredResult));
+        });
+    }
+
+    private void applyResult(int requestVersion, @Nullable Bitmap bitmap,
+                             @Nullable FaceAttributeProcessor.Result result) {
+        if (destroyed || imageVersion != requestVersion) {
+            recycle(bitmap);
+            return;
+        }
+        busy = false;
+        setControlsEnabled(session != null);
+        Bitmap previous = imageBitmap;
+        imageBitmap = bitmap;
+        analysis = result;
+        selectedIndex = result != null && result.candidates.length > 0 ? 0 : -1;
+
+        if (bitmap != null) {
+            imageView.setImageBitmap(bitmap);
+            imagePlaceholder.setVisibility(View.GONE);
+            faceOverlay.showFaces(bitmap.getWidth(), bitmap.getHeight(),
+                    result == null ? null : result.faceRects(), selectedIndex);
+        } else {
+            imageView.setImageDrawable(null);
+            imagePlaceholder.setVisibility(View.VISIBLE);
+            faceOverlay.clearFace();
+        }
+        if (previous != bitmap) {
+            recycle(previous);
+        }
+
+        if (result == null) {
+            setImageStatus(bitmap == null ? R.string.image_load_failed
+                    : R.string.attribute_analysis_failed, R.color.liveness_fail);
+            showResultPlaceholder(R.string.attribute_analysis_failed);
+        } else if (result.status == FaceAttributeProcessor.Status.NO_FACE) {
+            setImageStatus(R.string.image_no_face, R.color.liveness_fail);
+            showResultPlaceholder(R.string.attribute_no_face_result);
+        } else if (result.status != FaceAttributeProcessor.Status.READY
+                || result.attributes.length != result.candidates.length) {
+            setImageStatus(R.string.attribute_analysis_failed, R.color.liveness_fail);
+            showResultPlaceholder(R.string.attribute_analysis_failed);
+        } else {
+            renderSelectedFace();
+        }
+    }
+
+    private void selectFace(int index) {
+        if (busy || analysis == null || index < 0
+                || index >= analysis.candidates.length) {
+            return;
+        }
+        selectedIndex = index;
+        faceOverlay.setSelectedIndex(index);
+        renderSelectedFace();
+    }
+
+    private void renderSelectedFace() {
+        if (analysis == null || selectedIndex < 0
+                || selectedIndex >= analysis.attributes.length) {
+            showResultPlaceholder(R.string.attribute_analysis_failed);
+            return;
+        }
+        setImageStatus(getString(R.string.attribute_face_selected,
+                selectedIndex + 1, analysis.candidates.length), R.color.liveness_accent);
+        updateMagnifier(analysis.candidates[selectedIndex]);
+        FaceAttributeProcessor.Attribute attribute = analysis.attributes[selectedIndex];
+        maskValue.setText(binaryValue(attribute.maskConfidence,
+                R.string.attribute_mask_yes, R.string.attribute_mask_no));
+        ageValue.setText(age(attribute.ageBracket));
+        qualityValue.setText(validConfidence(attribute.qualityScore)
+                ? getString(R.string.attribute_score_format, attribute.qualityScore)
+                : getString(R.string.attribute_unknown));
+        expressionValue.setText(attribute.jawOpen > 0
+                ? R.string.attribute_expression_mouth_open
+                : attribute.jawOpen == 0
+                ? R.string.attribute_expression_neutral
+                : R.string.attribute_unknown);
+        raceValue.setText(race(attribute.race));
+        genderValue.setText(gender(attribute.gender));
+        leftEyeValue.setText(binaryValue(attribute.leftEyeConfidence,
+                R.string.attribute_eye_open, R.string.attribute_eye_closed));
+        rightEyeValue.setText(binaryValue(attribute.rightEyeConfidence,
+                R.string.attribute_eye_open, R.string.attribute_eye_closed));
+        resultPlaceholder.setVisibility(View.GONE);
+        resultValues.setVisibility(View.VISIBLE);
+    }
+
+    private void updateMagnifier(FaceImageProcessor.Candidate selected) {
+        clearMagnifier();
+        Bitmap source = imageBitmap;
+        if (source == null || source.isRecycled()) {
+            return;
+        }
+        float imageArea = (float) source.getWidth() * source.getHeight();
+        float faceArea = Math.max(0f, selected.rect.width())
+                * Math.max(0f, selected.rect.height());
+        if (imageArea > 0f && faceArea / imageArea >= DIRECT_FACE_AREA_RATIO) {
+            return;
+        }
+        FaceCropUtils.SquareCrop crop = FaceCropUtils.createSquare(
+                source, selected.rect, MAGNIFIER_CROP_SCALE);
+        if (crop != null) {
+            magnifierBitmap = crop.bitmap;
+            magnifierImage.setImageBitmap(crop.bitmap);
+            magnifierCard.setVisibility(View.VISIBLE);
+        }
+    }
+
+    private CharSequence binaryValue(float confidence,
+                                     @StringRes int high, @StringRes int low) {
+        return validConfidence(confidence)
+                ? getString(confidence >= BINARY_CONFIDENCE_THRESHOLD ? high : low)
+                : getString(R.string.attribute_unknown);
+    }
+
+    private CharSequence age(int bracket) {
+        switch (bracket) {
+            case 0: return getString(R.string.attribute_age_0_2);
+            case 1: return getString(R.string.attribute_age_3_9);
+            case 2: return getString(R.string.attribute_age_10_19);
+            case 3: return getString(R.string.attribute_age_20_29);
+            case 4: return getString(R.string.attribute_age_30_39);
+            case 5: return getString(R.string.attribute_age_40_49);
+            case 6: return getString(R.string.attribute_age_50_59);
+            case 7: return getString(R.string.attribute_age_60_69);
+            case 8: return getString(R.string.attribute_age_70_plus);
+            default: return getString(R.string.attribute_unknown);
+        }
+    }
+
+    private CharSequence race(int value) {
+        switch (value) {
+            case 0: return getString(R.string.attribute_race_black);
+            case 1: return getString(R.string.attribute_race_asian);
+            case 2: return getString(R.string.attribute_race_latino);
+            case 3: return getString(R.string.attribute_race_middle_eastern);
+            case 4: return getString(R.string.attribute_race_white);
+            default: return getString(R.string.attribute_unknown);
+        }
+    }
+
+    private CharSequence gender(int value) {
+        if (value == 0) {
+            return getString(R.string.attribute_gender_female);
+        }
+        if (value == 1) {
+            return getString(R.string.attribute_gender_male);
+        }
+        return getString(R.string.attribute_unknown);
+    }
+
+    private void showResultPlaceholder(@StringRes int message) {
+        resultValues.setVisibility(View.GONE);
+        resultPlaceholder.setText(message);
+        resultPlaceholder.setVisibility(View.VISIBLE);
+    }
+
+    private void clearMagnifier() {
+        magnifierImage.setImageDrawable(null);
+        magnifierCard.setVisibility(View.GONE);
+        recycle(magnifierBitmap);
+        magnifierBitmap = null;
+    }
+
+    private void setControlsEnabled(boolean enabled) {
+        choosePhotoButton.setEnabled(enabled && !busy);
+    }
+
+    private void setImageStatus(@StringRes int text, @ColorRes int colorRes) {
+        setImageStatus(getString(text), colorRes);
+    }
+
+    private void setImageStatus(CharSequence text, @ColorRes int colorRes) {
+        imageStatus.setText(text);
+        imageStatus.setTextColor(ContextCompat.getColor(this, colorRes));
+    }
+
+    private static boolean validConfidence(float value) {
+        return !Float.isNaN(value) && !Float.isInfinite(value)
+                && value >= 0f && value <= 1f;
+    }
+
+    private static void recycle(@Nullable Bitmap bitmap) {
+        if (bitmap != null && !bitmap.isRecycled()) {
+            bitmap.recycle();
+        }
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.attributeRoot);
+        int horizontal = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(horizontal, top + bars.top, horizontal, bottom + bars.bottom);
+            return insets;
+        });
+    }
+
+    @Override
+    protected void onDestroy() {
+        destroyed = true;
+        imageVersion++;
+        clearMagnifier();
+        recycle(imageBitmap);
+        imageBitmap = null;
+        sdkExecutor.execute(() -> {
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        });
+        sdkExecutor.shutdown();
+        super.onDestroy();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceCompareActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceCompareActivity.java
new file mode 100644
index 000000000..073a1ab42
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceCompareActivity.java
@@ -0,0 +1,396 @@
+package com.example.inspireface_example;
+
+import android.graphics.Bitmap;
+import android.net.Uri;
+import android.os.Bundle;
+import android.view.View;
+import android.widget.ImageView;
+import android.widget.TextView;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.annotation.ColorRes;
+import androidx.annotation.Nullable;
+import androidx.annotation.StringRes;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.content.ContextCompat;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.face.FaceImageProcessor;
+import com.example.inspireface_example.face.ImageBitmapLoader;
+import com.example.inspireface_example.view.FaceEngine;
+import com.example.inspireface_example.widget.FaceImageOverlayView;
+import com.example.inspireface_example.widget.SimilarityGaugeView;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.Session;
+import com.insightface.sdk.inspireface.base.SimilarityConverterConfig;
+
+import java.util.Locale;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** Still-image 1:1 face comparison using the globally selected InspireFace model. */
+public class FaceCompareActivity extends AppCompatActivity {
+
+    private static final int MAX_IMAGE_DIMENSION = 2048;
+
+    private final ExecutorService sdkExecutor = Executors.newSingleThreadExecutor();
+    private final ImageSlot slotA = new ImageSlot();
+    private final ImageSlot slotB = new ImageSlot();
+
+    private SimilarityGaugeView similarityGauge;
+    private TextView comparisonDetails;
+    private Session session;
+    private volatile boolean destroyed;
+
+    private final ActivityResultLauncher<String> imageAPicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null) {
+                    selectImage(slotA, uri);
+                }
+            });
+
+    private final ActivityResultLauncher<String> imageBPicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null) {
+                    selectImage(slotB, uri);
+                }
+            });
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_face_compare);
+        applyWindowInsets();
+
+        bindSlot(slotA, R.id.imageA, R.id.faceOverlayA, R.id.placeholderA, R.id.statusA,
+                R.id.compareCropCardA, R.id.compareCropA);
+        bindSlot(slotB, R.id.imageB, R.id.faceOverlayB, R.id.placeholderB, R.id.statusB,
+                R.id.compareCropCardB, R.id.compareCropB);
+        similarityGauge = findViewById(R.id.similarityGauge);
+        comparisonDetails = findViewById(R.id.comparisonDetails);
+
+        ((TextView) findViewById(R.id.currentModel)).setText(
+                getString(R.string.current_model, FaceModelPrefs.get(this).sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        findViewById(R.id.cardImageA).setOnClickListener(v -> imageAPicker.launch("image/*"));
+        findViewById(R.id.cardImageB).setOnClickListener(v -> imageBPicker.launch("image/*"));
+
+        similarityGauge.showMessage(getString(R.string.compare_select_two));
+        sdkExecutor.execute(this::initializeEngine);
+    }
+
+    private void bindSlot(ImageSlot slot, int imageId, int overlayId,
+                          int placeholderId, int statusId, int cropCardId, int cropId) {
+        slot.imageView = findViewById(imageId);
+        slot.faceOverlay = findViewById(overlayId);
+        slot.faceOverlay.setOnFaceSelectedListener(index -> selectFace(slot, index));
+        slot.placeholder = findViewById(placeholderId);
+        slot.statusView = findViewById(statusId);
+        slot.cropCard = findViewById(cropCardId);
+        slot.cropView = findViewById(cropId);
+    }
+
+    private void initializeEngine() {
+        boolean launched = FaceEngine.ensureLaunched(this);
+        if (launched) {
+            session = FaceEngine.createRecognitionSession();
+        }
+        if (!launched || session == null) {
+            postUi(() -> {
+                similarityGauge.showMessage(getString(R.string.compare_engine_failed));
+                setSlotStatus(slotA, R.string.compare_engine_failed, R.color.liveness_fail);
+                setSlotStatus(slotB, R.string.compare_engine_failed, R.color.liveness_fail);
+            });
+        }
+    }
+
+    private void selectImage(ImageSlot slot, Uri uri) {
+        int requestVersion = ++slot.version;
+        FaceImageProcessor.Candidate[] previousCandidates = slot.candidates;
+        clearSelectedCrop(slot);
+        slot.feature = null;
+        slot.candidates = null;
+        slot.selectedIndex = -1;
+        slot.pending = true;
+        FaceImageProcessor.recycleCrops(previousCandidates, -1);
+        slot.faceOverlay.clearFace();
+        setSlotStatus(slot, R.string.image_analyzing, R.color.home_text_secondary);
+        comparisonDetails.setText(null);
+        updateGaugeForSlots();
+        sdkExecutor.execute(() -> processImage(slot, uri, requestVersion));
+    }
+
+    private void processImage(ImageSlot slot, Uri uri, int requestVersion) {
+        Bitmap bitmap = null;
+        FaceImageProcessor.Candidate[] candidates = null;
+        int statusRes;
+        try {
+            bitmap = ImageBitmapLoader.decode(this, uri, MAX_IMAGE_DIMENSION);
+            if (slot.version != requestVersion || destroyed) {
+                bitmap.recycle();
+                return;
+            }
+            if (session == null) {
+                statusRes = R.string.compare_engine_failed;
+            } else {
+                FaceImageProcessor.Result result =
+                        FaceImageProcessor.detect(session, bitmap, true);
+                candidates = result.candidates;
+                if (result.status == FaceImageProcessor.Status.NO_FACE) {
+                    statusRes = R.string.image_no_face;
+                } else if (result.status != FaceImageProcessor.Status.READY) {
+                    statusRes = R.string.face_extract_failed;
+                } else if (candidates[0].feature == null) {
+                    statusRes = R.string.face_extract_failed;
+                } else {
+                    statusRes = R.string.image_face_ready;
+                }
+            }
+        } catch (Exception e) {
+            statusRes = R.string.image_load_failed;
+        }
+
+        Bitmap deliveredBitmap = bitmap;
+        FaceImageProcessor.Candidate[] deliveredCandidates = candidates;
+        int deliveredStatus = statusRes;
+        // applyImageResult owns cleanup for obsolete results, including Activity teardown.
+        runOnUiThread(() -> applyImageResult(slot, requestVersion, deliveredBitmap,
+                deliveredCandidates, deliveredStatus));
+    }
+
+    private void applyImageResult(ImageSlot slot, int requestVersion, Bitmap bitmap,
+                                  FaceImageProcessor.Candidate[] candidates,
+                                  @StringRes int statusRes) {
+        if (destroyed || slot.version != requestVersion) {
+            if (bitmap != null && !bitmap.isRecycled()) {
+                bitmap.recycle();
+            }
+            FaceImageProcessor.recycleCrops(candidates, -1);
+            return;
+        }
+        slot.pending = false;
+        FaceImageProcessor.Candidate[] previousCandidates = slot.candidates;
+        if (previousCandidates != candidates) {
+            clearSelectedCrop(slot);
+            FaceImageProcessor.recycleCrops(previousCandidates, -1);
+        }
+        slot.candidates = candidates;
+        slot.selectedIndex = candidates != null && candidates.length > 0 ? 0 : -1;
+        slot.feature = slot.selectedIndex >= 0 ? candidates[0].feature : null;
+        if (bitmap != null) {
+            Bitmap previous = slot.bitmap;
+            slot.bitmap = bitmap;
+            slot.imageView.setImageBitmap(bitmap);
+            slot.faceOverlay.showFaces(bitmap.getWidth(), bitmap.getHeight(),
+                    candidates == null ? null : faceRects(candidates), slot.selectedIndex);
+            slot.placeholder.setVisibility(View.GONE);
+            if (previous != null && previous != bitmap && !previous.isRecycled()) {
+                previous.recycle();
+            }
+        }
+        updateSelectedCrop(slot);
+        if (slot.feature != null && candidates.length > 1) {
+            setSlotStatus(slot, getString(R.string.image_face_selected,
+                    slot.selectedIndex + 1, candidates.length), R.color.liveness_accent);
+        } else {
+            setSlotStatus(slot, getString(statusRes),
+                    slot.feature != null ? R.color.liveness_accent : R.color.liveness_fail);
+        }
+        updateGaugeForSlots();
+    }
+
+    private void selectFace(ImageSlot slot, int index) {
+        if (slot.pending || slot.candidates == null
+                || index < 0 || index >= slot.candidates.length) {
+            return;
+        }
+        slot.selectedIndex = index;
+        slot.feature = slot.candidates[index].feature;
+        slot.faceOverlay.setSelectedIndex(index);
+        updateSelectedCrop(slot);
+        comparisonDetails.setText(null);
+        if (slot.feature == null) {
+            setSlotStatus(slot, getString(R.string.face_extract_failed), R.color.liveness_fail);
+        } else if (slot.candidates.length > 1) {
+            setSlotStatus(slot, getString(R.string.image_face_selected,
+                    index + 1, slot.candidates.length), R.color.liveness_accent);
+        } else {
+            setSlotStatus(slot, getString(R.string.image_face_ready), R.color.liveness_accent);
+        }
+        updateGaugeForSlots();
+    }
+
+    private void updateSelectedCrop(ImageSlot slot) {
+        if (slot.candidates == null || slot.selectedIndex < 0
+                || slot.selectedIndex >= slot.candidates.length) {
+            clearSelectedCrop(slot);
+            return;
+        }
+        Bitmap crop = slot.candidates[slot.selectedIndex].crop;
+        if (crop == null || crop.isRecycled()) {
+            clearSelectedCrop(slot);
+            return;
+        }
+        slot.cropView.setImageBitmap(crop);
+        slot.cropCard.setVisibility(View.VISIBLE);
+    }
+
+    private static void clearSelectedCrop(ImageSlot slot) {
+        if (slot.cropView != null) {
+            slot.cropView.setImageDrawable(null);
+        }
+        if (slot.cropCard != null) {
+            slot.cropCard.setVisibility(View.INVISIBLE);
+        }
+    }
+
+    private static android.graphics.RectF[] faceRects(
+            FaceImageProcessor.Candidate[] candidates) {
+        android.graphics.RectF[] rects = new android.graphics.RectF[candidates.length];
+        for (int i = 0; i < candidates.length; i++) {
+            rects[i] = candidates[i].rect;
+        }
+        return rects;
+    }
+
+    private void updateGaugeForSlots() {
+        if (slotA.pending || slotB.pending) {
+            similarityGauge.showMessage(getString(R.string.compare_analyzing));
+            return;
+        }
+        if (slotA.feature == null || slotB.feature == null) {
+            similarityGauge.showMessage(getString(
+                    slotA.bitmap == null && slotB.bitmap == null
+                            ? R.string.compare_select_two : R.string.compare_waiting_face));
+            return;
+        }
+
+        int versionA = slotA.version;
+        int versionB = slotB.version;
+        FaceFeature featureA = slotA.feature;
+        FaceFeature featureB = slotB.feature;
+        similarityGauge.showMessage(getString(R.string.compare_comparing));
+        sdkExecutor.execute(() -> compareFeatures(
+                versionA, versionB, featureA, featureB));
+    }
+
+    private void compareFeatures(int versionA, int versionB,
+                                 FaceFeature featureA, FaceFeature featureB) {
+        float cosine = InspireFace.FaceComparison(featureA, featureB);
+        float threshold = InspireFace.GetRecommendedCosineThreshold();
+        SimilarityConverterConfig converter = InspireFace.GetCosineSimilarityConverter();
+        float converted = InspireFace.CosineSimilarityConvertToPercentage(cosine);
+        boolean converterUsesUnitRange = converter == null
+                ? converted >= 0f && converted <= 1f
+                : converter.outputMax <= 1.0001f;
+        float percent = converterUsesUnitRange ? converted * 100f : converted;
+        if (Float.isNaN(percent) || Float.isInfinite(percent)) {
+            postUi(() -> similarityGauge.showMessage(getString(R.string.compare_failed)));
+            return;
+        }
+        percent = Math.max(0f, Math.min(100f, percent));
+        boolean matched = cosine >= threshold;
+        float finalPercent = percent;
+        postUi(() -> {
+            if (slotA.version != versionA || slotB.version != versionB
+                    || slotA.feature != featureA || slotB.feature != featureB) {
+                return;
+            }
+            similarityGauge.showResult(finalPercent, matched,
+                    getString(matched ? R.string.compare_same_person
+                            : R.string.compare_different_person));
+            comparisonDetails.setText(String.format(Locale.US,
+                    getString(R.string.compare_details), cosine, threshold));
+        });
+    }
+
+    private void setSlotStatus(ImageSlot slot, @StringRes int statusRes,
+                               @ColorRes int colorRes) {
+        setSlotStatus(slot, getString(statusRes), colorRes);
+    }
+
+    private void setSlotStatus(ImageSlot slot, CharSequence status,
+                               @ColorRes int colorRes) {
+        slot.statusView.setText(status);
+        slot.statusView.setTextColor(ContextCompat.getColor(this, colorRes));
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.compareRoot);
+        int left = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int right = root.getPaddingRight();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(left + bars.left, top + bars.top,
+                    right + bars.right, bottom + bars.bottom);
+            return insets;
+        });
+    }
+
+    private void postUi(Runnable action) {
+        runOnUiThread(() -> {
+            if (!destroyed && !isFinishing()) {
+                action.run();
+            }
+        });
+    }
+
+    @Override
+    protected void onDestroy() {
+        destroyed = true;
+        slotA.version++;
+        slotB.version++;
+        Bitmap bitmapA = slotA.bitmap;
+        Bitmap bitmapB = slotB.bitmap;
+        FaceImageProcessor.Candidate[] candidatesA = slotA.candidates;
+        FaceImageProcessor.Candidate[] candidatesB = slotB.candidates;
+        clearSelectedCrop(slotA);
+        clearSelectedCrop(slotB);
+        slotA.bitmap = null;
+        slotB.bitmap = null;
+        slotA.candidates = null;
+        slotB.candidates = null;
+        sdkExecutor.execute(() -> {
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+            FaceImageProcessor.recycleCrops(candidatesA, -1);
+            FaceImageProcessor.recycleCrops(candidatesB, -1);
+            if (bitmapA != null && !bitmapA.isRecycled()) {
+                bitmapA.recycle();
+            }
+            if (bitmapB != null && bitmapB != bitmapA && !bitmapB.isRecycled()) {
+                bitmapB.recycle();
+            }
+        });
+        sdkExecutor.shutdown();
+        super.onDestroy();
+    }
+
+    private static final class ImageSlot {
+        volatile int version;
+        boolean pending;
+        Bitmap bitmap;
+        FaceFeature feature;
+        FaceImageProcessor.Candidate[] candidates;
+        int selectedIndex = -1;
+        ImageView imageView;
+        FaceImageOverlayView faceOverlay;
+        View placeholder;
+        TextView statusView;
+        View cropCard;
+        ImageView cropView;
+    }
+
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceDetectionActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceDetectionActivity.java
new file mode 100644
index 000000000..6c7a776ac
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceDetectionActivity.java
@@ -0,0 +1,914 @@
+package com.example.inspireface_example;
+
+import android.graphics.Bitmap;
+import android.graphics.RectF;
+import android.net.Uri;
+import android.os.Bundle;
+import android.view.View;
+import android.widget.ImageView;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.annotation.ColorRes;
+import androidx.annotation.Nullable;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.camera.view.PreviewView;
+import androidx.core.content.ContextCompat;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.face.FaceImageProcessor;
+import com.example.inspireface_example.face.FaceCropUtils;
+import com.example.inspireface_example.face.ImageBitmapLoader;
+import com.example.inspireface_example.permission.CameraPermissionCoordinator;
+import com.example.inspireface_example.view.CameraPreviewController;
+import com.example.inspireface_example.view.FaceEngine;
+import com.example.inspireface_example.view.FaceTrackingAnalyzer;
+import com.example.inspireface_example.view.FaceTrackingGlView;
+import com.example.inspireface_example.widget.FaceImageOverlayView;
+import com.example.inspireface_example.widget.FaceLandmarkOverlayView;
+import com.google.android.material.button.MaterialButton;
+import com.google.android.material.chip.ChipGroup;
+import com.google.android.material.switchmaterial.SwitchMaterial;
+import com.google.android.material.tabs.TabLayout;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** Still-image first step of the face detection and tracking demo. */
+public final class FaceDetectionActivity extends AppCompatActivity {
+
+    private static final int MAX_IMAGE_DIMENSION = 2048;
+    /** Only genuinely small faces need the landmark magnifier. */
+    private static final float DIRECT_LANDMARK_FACE_AREA_RATIO = 0.05f;
+    private static final float LANDMARK_MAGNIFIER_CROP_SCALE = 2.4f;
+
+    private final ExecutorService sdkExecutor = Executors.newSingleThreadExecutor();
+    private final ExecutorService trackingExecutor = Executors.newSingleThreadExecutor();
+
+    private Session session;
+    private StillImageSessionSettings.Values savedSettings;
+    private FaceTrackingSessionSettings.Values trackingSettings;
+    private volatile boolean destroyed;
+    private volatile int imageVersion;
+    private boolean sessionBusy = true;
+    private boolean engineReady;
+
+    private ChipGroup inputPxGroup;
+    private ChipGroup maxFacesGroup;
+    private ChipGroup minFaceGroup;
+    private MaterialButton applySessionButton;
+    private MaterialButton resetSessionButton;
+    private MaterialButton choosePhotoButton;
+    private TextView sessionStatus;
+    private View sessionSettingsContent;
+    private TextView sessionSettingsToggleText;
+    private SwitchMaterial landmarkSwitch;
+    private ImageView imageView;
+    private View imagePlaceholder;
+    private TextView imageStatus;
+    private FaceImageOverlayView faceOverlay;
+    private FaceLandmarkOverlayView sourceLandmarkOverlay;
+    private View magnifierCard;
+    private ImageView magnifierImage;
+    private FaceLandmarkOverlayView magnifierLandmarkOverlay;
+    private View magnifierLandmarkBadge;
+    private TextView landmarkDisplayHint;
+
+    private PreviewView trackingPreview;
+    private FaceTrackingGlView trackingGlOverlay;
+    private TextView trackingStatus;
+    private TextView trackingSessionStatus;
+    private TextView trackingSettingsToggleText;
+    private View trackingSettingsContent;
+    private ChipGroup trackingModeGroup;
+    private ChipGroup trackingInputPxGroup;
+    private ChipGroup trackingMaxFacesGroup;
+    private ChipGroup trackingMinFaceGroup;
+    private MaterialButton resetTrackingButton;
+    private View trackingLoadingIndicator;
+    private View flipTrackingButton;
+    private CameraPreviewController trackingCameraController;
+    private FaceTrackingAnalyzer trackingAnalyzer;
+    private boolean videoSelected;
+    private boolean trackingStarting;
+    private CameraPermissionCoordinator cameraPermission;
+    private boolean trackingFrontCamera = true;
+    private boolean suppressTrackingSettingChanges;
+    private int trackingGeneration;
+
+    private Bitmap imageBitmap;
+    private Bitmap magnifierBitmap;
+    private FaceImageProcessor.Candidate[] candidates;
+    private int selectedIndex = -1;
+
+    private final ActivityResultLauncher<String> photoPicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null && !sessionBusy && session != null) {
+                    loadPhoto(uri);
+                }
+            });
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        cameraPermission = new CameraPermissionCoordinator(this,
+                new CameraPermissionCoordinator.Listener() {
+                    @Override
+                    public void onCameraPermissionGranted() {
+                        startVideoTracking();
+                    }
+
+                    @Override
+                    public void onCameraPermissionBlocked(boolean requiresSettings) {
+                        if (videoSelected) {
+                            setTrackingLoading(false);
+                            showTrackingMessage(requiresSettings
+                                            ? R.string.camera_permission_settings_hint
+                                            : R.string.camera_permission_retry_hint,
+                                    R.color.liveness_fail);
+                        }
+                    }
+                });
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_face_detection);
+        applyWindowInsets();
+
+        savedSettings = StillImageSessionSettings.load(this);
+        trackingSettings = FaceTrackingSessionSettings.load(this);
+        bindViews();
+        cameraPermission.bindRecoveryButton(
+                findViewById(R.id.btnCameraPermissionAction));
+        ((TextView) findViewById(R.id.currentModel)).setText(getString(
+                R.string.current_model, FaceModelPrefs.get(this).sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        choosePhotoButton.setOnClickListener(v -> photoPicker.launch("image/*"));
+        findViewById(R.id.detectionImageCard).setOnClickListener(v -> {
+            if (!sessionBusy && session != null) {
+                photoPicker.launch("image/*");
+            }
+        });
+        faceOverlay.setOnFaceSelectedListener(this::selectFace);
+        landmarkSwitch.setOnCheckedChangeListener((button, checked) ->
+                renderSelectedLandmarks());
+        findViewById(R.id.sessionSettingsHeader).setOnClickListener(v ->
+                setSettingsExpanded(sessionSettingsContent.getVisibility() != View.VISIBLE));
+        applySessionButton.setOnClickListener(v -> rebuildSession(selectedSettings()));
+        resetSessionButton.setOnClickListener(v -> {
+            StillImageSessionSettings.Values defaults = StillImageSessionSettings.defaults();
+            selectParameters(defaults);
+            rebuildSession(defaults);
+        });
+        flipTrackingButton.setOnClickListener(v -> flipTrackingCamera());
+        findViewById(R.id.trackingSettingsHeader).setOnClickListener(v ->
+                setTrackingSettingsExpanded(
+                        trackingSettingsContent.getVisibility() != View.VISIBLE));
+        resetTrackingButton.setOnClickListener(v -> {
+            FaceTrackingSessionSettings.Values defaults =
+                    FaceTrackingSessionSettings.defaults();
+            selectTrackingParameters(defaults);
+            applyTrackingSettings(defaults);
+        });
+        configureTabs();
+        selectParameters(savedSettings);
+        selectTrackingParameters(trackingSettings);
+        configureTrackingAutoApply();
+        setSettingsExpanded(false);
+        setTrackingSettingsExpanded(false);
+        setControlsEnabled(false);
+        setTrackingControlsEnabled(false);
+        sessionStatus.setText(R.string.recognition_session_rebuilding);
+        showTrackingSummary(trackingSettings);
+        sdkExecutor.execute(this::initializeDetection);
+    }
+
+    private void bindViews() {
+        inputPxGroup = findViewById(R.id.inputPxGroup);
+        maxFacesGroup = findViewById(R.id.maxFacesGroup);
+        minFaceGroup = findViewById(R.id.minFaceGroup);
+        applySessionButton = findViewById(R.id.btnApplySession);
+        resetSessionButton = findViewById(R.id.btnResetSession);
+        choosePhotoButton = findViewById(R.id.btnChooseDetectionPhoto);
+        sessionStatus = findViewById(R.id.sessionStatus);
+        sessionSettingsContent = findViewById(R.id.sessionSettingsContent);
+        sessionSettingsToggleText = findViewById(R.id.sessionSettingsToggleText);
+        landmarkSwitch = findViewById(R.id.switchDenseLandmarks);
+        imageView = findViewById(R.id.detectionImage);
+        imagePlaceholder = findViewById(R.id.detectionImagePlaceholder);
+        imageStatus = findViewById(R.id.detectionImageStatus);
+        faceOverlay = findViewById(R.id.detectionFaceOverlay);
+        sourceLandmarkOverlay = findViewById(R.id.detectionLandmarkOverlay);
+        magnifierCard = findViewById(R.id.landmarkMagnifierCard);
+        magnifierImage = findViewById(R.id.landmarkMagnifierImage);
+        magnifierLandmarkOverlay = findViewById(R.id.landmarkMagnifierOverlay);
+        magnifierLandmarkBadge = findViewById(R.id.landmarkMagnifierBadge);
+        landmarkDisplayHint = findViewById(R.id.landmarkDisplayHint);
+        trackingPreview = findViewById(R.id.detectionTrackingPreview);
+        trackingGlOverlay = findViewById(R.id.detectionTrackingGlOverlay);
+        trackingStatus = findViewById(R.id.detectionTrackingStatus);
+        trackingSessionStatus = findViewById(R.id.trackingSessionStatus);
+        trackingSettingsToggleText = findViewById(R.id.trackingSettingsToggleText);
+        trackingSettingsContent = findViewById(R.id.trackingSettingsContent);
+        trackingModeGroup = findViewById(R.id.trackingModeGroup);
+        trackingInputPxGroup = findViewById(R.id.trackingInputPxGroup);
+        trackingMaxFacesGroup = findViewById(R.id.trackingMaxFacesGroup);
+        trackingMinFaceGroup = findViewById(R.id.trackingMinFaceGroup);
+        resetTrackingButton = findViewById(R.id.btnResetTrackingSession);
+        trackingLoadingIndicator = findViewById(R.id.trackingLoadingIndicator);
+        flipTrackingButton = findViewById(R.id.btnFlipTrackingCamera);
+    }
+
+    private void configureTabs() {
+        View imagePanel = findViewById(R.id.imageDetectionPanel);
+        View trackingPanel = findViewById(R.id.videoTrackingPanel);
+        ((TabLayout) findViewById(R.id.detectionTabs))
+                .addOnTabSelectedListener(new TabLayout.OnTabSelectedListener() {
+                    @Override
+                    public void onTabSelected(TabLayout.Tab tab) {
+                        boolean image = tab.getPosition() == 0;
+                        videoSelected = !image;
+                        imagePanel.setVisibility(image ? View.VISIBLE : View.GONE);
+                        trackingPanel.setVisibility(image ? View.GONE : View.VISIBLE);
+                        if (image) {
+                            stopVideoTracking();
+                        } else {
+                            startVideoTracking();
+                        }
+                    }
+
+                    @Override
+                    public void onTabUnselected(TabLayout.Tab tab) {
+                    }
+
+                    @Override
+                    public void onTabReselected(TabLayout.Tab tab) {
+                    }
+                });
+    }
+
+    private void initializeDetection() {
+        StillImageSessionSettings.Values initial = savedSettings;
+        boolean ready = FaceEngine.ensureLaunched(this);
+        if (ready) {
+            session = FaceEngine.createDetectionSession(
+                    initial.inputPx, initial.maxFaces, initial.minFacePx);
+            ready = session != null;
+        }
+        boolean finalReady = ready;
+        postUi(() -> {
+            sessionBusy = false;
+            engineReady = finalReady;
+            setControlsEnabled(finalReady);
+            setTrackingControlsEnabled(finalReady);
+            if (finalReady) {
+                showSessionSummary(initial);
+                if (videoSelected) {
+                    startVideoTracking();
+                }
+            } else {
+                setTrackingLoading(false);
+                sessionStatus.setText(R.string.recognition_session_failed);
+                sessionStatus.setTextColor(color(R.color.liveness_fail));
+                setImageStatus(R.string.compare_engine_failed, R.color.liveness_fail);
+            }
+        });
+    }
+
+    private void loadPhoto(Uri uri) {
+        int requestVersion = ++imageVersion;
+        StillImageSessionSettings.Values settings = savedSettings;
+        sessionBusy = true;
+        candidates = null;
+        selectedIndex = -1;
+        faceOverlay.clearFace();
+        clearLandmarkPresentation();
+        setControlsEnabled(false);
+        setImageStatus(R.string.image_analyzing, R.color.home_text_secondary);
+        sdkExecutor.execute(() -> {
+            Bitmap bitmap = null;
+            FaceImageProcessor.Result result = null;
+            try {
+                bitmap = ImageBitmapLoader.decode(this, uri, MAX_IMAGE_DIMENSION);
+                if (destroyed || imageVersion != requestVersion) {
+                    recycle(bitmap);
+                    return;
+                }
+                // The 1.2.0 landmark workaround uses LIGHT_TRACK. Start from a clean tracker
+                // for every unrelated still image so cached boxes from the previous photo
+                // cannot affect this detection.
+                replaceDetectionSession(settings);
+                if (session != null) {
+                    result = FaceImageProcessor.detectWithLandmarks(session, bitmap);
+                }
+            } catch (Exception ignored) {
+                // Null result is rendered as a load/detection failure.
+            }
+            Bitmap deliveredBitmap = bitmap;
+            FaceImageProcessor.Result deliveredResult = result;
+            runOnUiThread(() -> applyPhotoResult(
+                    requestVersion, deliveredBitmap, deliveredResult));
+        });
+    }
+
+    /** Called only from {@link #sdkExecutor}. */
+    private void replaceDetectionSession(StillImageSessionSettings.Values settings) {
+        if (session != null) {
+            FaceEngine.releaseSession(session);
+            session = null;
+        }
+        session = FaceEngine.createDetectionSession(
+                settings.inputPx, settings.maxFaces, settings.minFacePx);
+    }
+
+    private void applyPhotoResult(int requestVersion, @Nullable Bitmap bitmap,
+                                  @Nullable FaceImageProcessor.Result result) {
+        if (destroyed || imageVersion != requestVersion) {
+            recycle(bitmap);
+            return;
+        }
+        Bitmap previous = imageBitmap;
+        imageBitmap = bitmap;
+        candidates = result == null ? null : result.candidates;
+        selectedIndex = candidates != null && candidates.length > 0 ? 0 : -1;
+        if (bitmap != null) {
+            imageView.setImageBitmap(bitmap);
+            imagePlaceholder.setVisibility(View.GONE);
+            faceOverlay.showFaces(bitmap.getWidth(), bitmap.getHeight(),
+                    result == null ? null : result.faceRects(), selectedIndex);
+        } else {
+            imageView.setImageDrawable(null);
+            imagePlaceholder.setVisibility(View.VISIBLE);
+            faceOverlay.clearFace();
+        }
+        if (previous != bitmap) {
+            recycle(previous);
+        }
+        finishDetectionUi(result);
+    }
+
+    private void finishDetectionUi(@Nullable FaceImageProcessor.Result result) {
+        sessionBusy = false;
+        setControlsEnabled(session != null);
+        if (result == null) {
+            clearLandmarkPresentation();
+            setImageStatus(R.string.image_load_failed, R.color.liveness_fail);
+        } else if (result.status == FaceImageProcessor.Status.NO_FACE) {
+            clearLandmarkPresentation();
+            setImageStatus(R.string.image_no_face, R.color.liveness_fail);
+        } else if (result.status != FaceImageProcessor.Status.READY
+                || candidates == null || candidates.length == 0) {
+            clearLandmarkPresentation();
+            setImageStatus(R.string.face_detection_failed, R.color.liveness_fail);
+        } else {
+            updateSelectionStatus();
+            renderSelectedLandmarks();
+        }
+    }
+
+    private void selectFace(int index) {
+        if (sessionBusy || candidates == null || index < 0 || index >= candidates.length) {
+            return;
+        }
+        selectedIndex = index;
+        faceOverlay.setSelectedIndex(index);
+        updateSelectionStatus();
+        renderSelectedLandmarks();
+    }
+
+    private void updateSelectionStatus() {
+        if (candidates == null || candidates.length == 0) {
+            return;
+        }
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        int landmarkCount = selected == null || selected.denseLandmarks == null
+                ? 0 : selected.denseLandmarks.length;
+        imageStatus.setText(getString(R.string.detection_face_selected,
+                selectedIndex + 1, candidates.length, landmarkCount));
+        imageStatus.setTextColor(color(R.color.liveness_accent));
+    }
+
+    @Nullable
+    private FaceImageProcessor.Candidate selectedCandidate() {
+        return candidates != null && selectedIndex >= 0 && selectedIndex < candidates.length
+                ? candidates[selectedIndex] : null;
+    }
+
+    private void renderSelectedLandmarks() {
+        clearLandmarkPresentation();
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        Bitmap source = imageBitmap;
+        if (selected == null || source == null || source.isRecycled()) {
+            landmarkDisplayHint.setText(R.string.detection_landmarks_failed);
+            return;
+        }
+
+        boolean showLandmarks = landmarkSwitch.isChecked();
+        boolean landmarksAvailable = selected.denseLandmarks != null
+                && selected.denseLandmarks.length > 0;
+        float imageArea = (float) source.getWidth() * source.getHeight();
+        float faceArea = Math.max(0f, selected.rect.width())
+                * Math.max(0f, selected.rect.height());
+        float ratio = imageArea <= 0f ? 0f : faceArea / imageArea;
+        int percent = Math.round(ratio * 100f);
+        boolean magnified = ratio < DIRECT_LANDMARK_FACE_AREA_RATIO
+                && showLandmarkMagnifier(
+                        selected, showLandmarks && landmarksAvailable);
+        if (!showLandmarks) {
+            landmarkDisplayHint.setText(R.string.detection_landmarks_hidden);
+            return;
+        }
+        if (!landmarksAvailable) {
+            landmarkDisplayHint.setText(R.string.detection_landmarks_failed);
+            return;
+        }
+        if (magnified) {
+            landmarkDisplayHint.setText(getString(
+                    R.string.detection_landmarks_in_magnifier, percent));
+        } else {
+            sourceLandmarkOverlay.showPoints(source.getWidth(), source.getHeight(),
+                    selected.denseLandmarks);
+            landmarkDisplayHint.setText(getString(
+                    R.string.detection_landmarks_on_source, percent));
+        }
+    }
+
+    private boolean showLandmarkMagnifier(FaceImageProcessor.Candidate selected,
+                                          boolean showLandmarks) {
+        Bitmap source = imageBitmap;
+        if (source == null) {
+            return false;
+        }
+        FaceCropUtils.SquareCrop crop = FaceCropUtils.createSquare(
+                source, selected.rect, LANDMARK_MAGNIFIER_CROP_SCALE);
+        if (crop == null) {
+            return false;
+        }
+        magnifierBitmap = crop.bitmap;
+        magnifierImage.setImageBitmap(crop.bitmap);
+        if (showLandmarks) {
+            magnifierLandmarkOverlay.showPoints(crop.size, crop.size,
+                    selected.denseLandmarks, crop.left, crop.top);
+            magnifierLandmarkBadge.setVisibility(View.VISIBLE);
+        }
+        magnifierCard.setVisibility(View.VISIBLE);
+        return true;
+    }
+
+    private void clearLandmarkPresentation() {
+        sourceLandmarkOverlay.clearPoints();
+        magnifierLandmarkOverlay.clearPoints();
+        magnifierImage.setImageDrawable(null);
+        magnifierLandmarkBadge.setVisibility(View.GONE);
+        magnifierCard.setVisibility(View.GONE);
+        recycle(magnifierBitmap);
+        magnifierBitmap = null;
+    }
+
+    private void rebuildSession(StillImageSessionSettings.Values requested) {
+        if (sessionBusy) {
+            return;
+        }
+        int version = ++imageVersion;
+        Bitmap currentImage = imageBitmap;
+        sessionBusy = true;
+        candidates = null;
+        selectedIndex = -1;
+        faceOverlay.clearFace();
+        clearLandmarkPresentation();
+        setControlsEnabled(false);
+        sessionStatus.setText(R.string.recognition_session_rebuilding);
+        sessionStatus.setTextColor(color(R.color.home_text_secondary));
+        if (currentImage != null) {
+            setImageStatus(R.string.image_analyzing, R.color.home_text_secondary);
+        }
+        sdkExecutor.execute(() -> {
+            replaceDetectionSession(requested);
+            FaceImageProcessor.Result result = session != null && currentImage != null
+                    ? FaceImageProcessor.detectWithLandmarks(session, currentImage) : null;
+            boolean ready = session != null;
+            postUi(() -> {
+                if (imageVersion != version) {
+                    return;
+                }
+                sessionBusy = false;
+                setControlsEnabled(ready);
+                if (!ready) {
+                    sessionStatus.setText(R.string.recognition_session_failed);
+                    sessionStatus.setTextColor(color(R.color.liveness_fail));
+                    if (currentImage != null) {
+                        setImageStatus(R.string.compare_engine_failed, R.color.liveness_fail);
+                    }
+                    return;
+                }
+                savedSettings = requested;
+                StillImageSessionSettings.save(this, requested);
+                showSessionSummary(requested);
+                if (currentImage != null) {
+                    candidates = result == null ? null : result.candidates;
+                    selectedIndex = candidates != null && candidates.length > 0 ? 0 : -1;
+                    faceOverlay.showFaces(currentImage.getWidth(), currentImage.getHeight(),
+                            result == null ? null : result.faceRects(), selectedIndex);
+                    finishDetectionUi(result);
+                }
+            });
+        });
+    }
+
+    private void showSessionSummary(StillImageSessionSettings.Values settings) {
+        sessionStatus.setText(getString(R.string.recognition_session_summary,
+                settings.inputPx, settings.maxFaces, settings.minFacePx));
+        sessionStatus.setTextColor(color(R.color.liveness_accent));
+    }
+
+    private void selectParameters(StillImageSessionSettings.Values settings) {
+        inputPxGroup.check(settings.inputPx == 320 ? R.id.inputPx320
+                : settings.inputPx == 1280 ? R.id.inputPx1280 : R.id.inputPx640);
+        maxFacesGroup.check(settings.maxFaces == 1 ? R.id.maxFaces1
+                : settings.maxFaces == 3 ? R.id.maxFaces3
+                : settings.maxFaces == 5 ? R.id.maxFaces5 : R.id.maxFaces10);
+        minFaceGroup.check(settings.minFacePx == 48 ? R.id.minFace48
+                : settings.minFacePx == 64 ? R.id.minFace64
+                : settings.minFacePx == 128 ? R.id.minFace128 : R.id.minFace24);
+    }
+
+    private StillImageSessionSettings.Values selectedSettings() {
+        int inputId = inputPxGroup.getCheckedChipId();
+        int inputPx = inputId == R.id.inputPx320 ? 320
+                : inputId == R.id.inputPx1280 ? 1280 : 640;
+        int maxId = maxFacesGroup.getCheckedChipId();
+        int maxFaces = maxId == R.id.maxFaces1 ? 1 : maxId == R.id.maxFaces3 ? 3
+                : maxId == R.id.maxFaces5 ? 5 : 10;
+        int minId = minFaceGroup.getCheckedChipId();
+        int minFacePx = minId == R.id.minFace48 ? 48 : minId == R.id.minFace64 ? 64
+                : minId == R.id.minFace128 ? 128 : 24;
+        return new StillImageSessionSettings.Values(inputPx, maxFaces, minFacePx);
+    }
+
+    private void setSettingsExpanded(boolean expanded) {
+        sessionSettingsContent.setVisibility(expanded ? View.VISIBLE : View.GONE);
+        sessionSettingsToggleText.setText(expanded
+                ? R.string.recognition_settings_collapse
+                : R.string.recognition_settings_expand);
+    }
+
+    private void setTrackingSettingsExpanded(boolean expanded) {
+        trackingSettingsContent.setVisibility(expanded ? View.VISIBLE : View.GONE);
+        trackingSettingsToggleText.setText(expanded
+                ? R.string.recognition_settings_collapse
+                : R.string.recognition_settings_expand);
+    }
+
+    private void startVideoTracking() {
+        if (!videoSelected || destroyed || trackingCameraController != null
+                || trackingStarting) {
+            return;
+        }
+        setTrackingLoading(true);
+        if (!engineReady) {
+            showTrackingMessage(
+                    R.string.detection_tracking_initializing, R.color.white);
+            return;
+        }
+        if (!cameraPermission.hasPermission()) {
+            showTrackingMessage(
+                    R.string.msg_permission_required, R.color.liveness_fail);
+            cameraPermission.requestAccess();
+            return;
+        }
+
+        trackingStarting = true;
+        setTrackingControlsEnabled(false);
+        showTrackingMessage(R.string.detection_tracking_initializing, R.color.white);
+        int generation = ++trackingGeneration;
+        FaceTrackingSessionSettings.Values settings = trackingSettings;
+        FaceTrackingAnalyzer analyzer = new FaceTrackingAnalyzer(
+                trackingGlOverlay, settings.mode, settings.inputPx,
+                settings.maxFaces, settings.minFacePx,
+                new FaceTrackingAnalyzer.Listener() {
+                    @Override
+                    public void onSessionReady() {
+                        postTrackingUi(generation, () -> {
+                            trackingStarting = false;
+                            setTrackingLoading(false);
+                            setTrackingControlsEnabled(engineReady);
+                        });
+                    }
+
+                    @Override
+                    public void onStats(double fps, long latencyMs) {
+                        postTrackingUi(generation,
+                                () -> renderTrackingStats(fps, latencyMs));
+                    }
+
+                    @Override
+                    public void onSessionError() {
+                        postTrackingUi(generation, () -> {
+                            trackingStarting = false;
+                            setTrackingLoading(false);
+                            setTrackingControlsEnabled(engineReady);
+                            showTrackingMessage(
+                                    R.string.detection_tracking_session_failed,
+                                    R.color.liveness_fail);
+                        });
+                    }
+                });
+        trackingAnalyzer = analyzer;
+        trackingCameraController = new CameraPreviewController(
+                this, this, trackingPreview, trackingExecutor, analyzer,
+                new CameraPreviewController.Listener() {
+                    @Override
+                    public void onCameraReady(boolean frontCamera) {
+                        if (!isCurrentTracking(generation)) {
+                            return;
+                        }
+                        trackingFrontCamera = frontCamera;
+                        analyzer.setMirrored(frontCamera);
+                        showTrackingMessage(
+                                R.string.detection_tracking_initializing, R.color.white);
+                    }
+
+                    @Override
+                    public void onLensChanged(boolean frontCamera) {
+                        if (!isCurrentTracking(generation)) {
+                            return;
+                        }
+                        trackingFrontCamera = frontCamera;
+                        analyzer.setMirrored(frontCamera);
+                        showTrackingMessage(
+                                R.string.detection_tracking_initializing, R.color.white);
+                    }
+
+                    @Override
+                    public void onCameraError(int messageRes) {
+                        if (!isCurrentTracking(generation)) {
+                            return;
+                        }
+                        trackingStarting = false;
+                        setTrackingLoading(false);
+                        setTrackingControlsEnabled(engineReady);
+                        showTrackingMessage(messageRes, R.color.liveness_fail);
+                    }
+                });
+        trackingCameraController.start(trackingFrontCamera);
+    }
+
+    private void stopVideoTracking() {
+        trackingGeneration++;
+        trackingStarting = false;
+        setTrackingLoading(false);
+        if (trackingCameraController != null) {
+            trackingCameraController.stop();
+            trackingCameraController = null;
+        }
+        if (trackingAnalyzer != null) {
+            FaceTrackingAnalyzer toRelease = trackingAnalyzer;
+            trackingAnalyzer = null;
+            trackingExecutor.execute(toRelease::release);
+        }
+        if (trackingGlOverlay != null) {
+            trackingGlOverlay.clearTracking();
+        }
+        if (!destroyed) {
+            setTrackingControlsEnabled(engineReady);
+        }
+    }
+
+    private void flipTrackingCamera() {
+        if (trackingCameraController == null || !trackingCameraController.flipCamera()) {
+            Toast.makeText(this, R.string.msg_camera_unavailable, Toast.LENGTH_SHORT).show();
+        }
+    }
+
+    private void applyTrackingSettings(FaceTrackingSessionSettings.Values requested) {
+        if (!engineReady || trackingStarting) {
+            return;
+        }
+        if (sameTrackingSettings(requested, trackingSettings)) {
+            return;
+        }
+        setTrackingLoading(true);
+        stopVideoTracking();
+        trackingSettings = requested;
+        FaceTrackingSessionSettings.save(this, requested);
+        showTrackingSummary(requested);
+        showTrackingMessage(
+                R.string.detection_tracking_restarting, R.color.liveness_warn);
+        if (videoSelected) {
+            startVideoTracking();
+        }
+    }
+
+    private boolean isCurrentTracking(int generation) {
+        return !destroyed && videoSelected && generation == trackingGeneration
+                && trackingAnalyzer != null;
+    }
+
+    private void postTrackingUi(int generation, Runnable action) {
+        runOnUiThread(() -> {
+            if (isCurrentTracking(generation)) {
+                action.run();
+            }
+        });
+    }
+
+    private void renderTrackingStats(double fps, long latencyMs) {
+        trackingStatus.setText(getString(
+                R.string.detection_tracking_stats, fps, latencyMs));
+        trackingStatus.setTextColor(color(R.color.liveness_accent));
+    }
+
+    private void showTrackingMessage(int statusRes, @ColorRes int colorRes) {
+        trackingStatus.setText(statusRes);
+        trackingStatus.setTextColor(color(colorRes));
+    }
+
+    private void showTrackingSummary(FaceTrackingSessionSettings.Values settings) {
+        trackingSessionStatus.setText(getString(R.string.detection_tracking_summary,
+                trackingModeLabel(settings.mode), settings.inputPx,
+                settings.maxFaces, settings.minFacePx));
+        trackingSessionStatus.setTextColor(color(R.color.liveness_accent));
+    }
+
+    private String trackingModeLabel(int mode) {
+        return getString(mode == InspireFace.DETECT_MODE_TRACK_BY_DETECTION
+                ? R.string.detection_tracking_mode_tbd
+                : R.string.detection_tracking_mode_light);
+    }
+
+    private void selectTrackingParameters(FaceTrackingSessionSettings.Values settings) {
+        suppressTrackingSettingChanges = true;
+        trackingModeGroup.check(settings.mode == InspireFace.DETECT_MODE_TRACK_BY_DETECTION
+                ? R.id.trackingModeTbd : R.id.trackingModeLight);
+        trackingInputPxGroup.check(settings.inputPx == 320 ? R.id.trackingInputPx320
+                : settings.inputPx == 1280
+                ? R.id.trackingInputPx1280 : R.id.trackingInputPx640);
+        trackingMaxFacesGroup.check(settings.maxFaces == 1 ? R.id.trackingMaxFaces1
+                : settings.maxFaces == 3 ? R.id.trackingMaxFaces3
+                : settings.maxFaces == 5
+                ? R.id.trackingMaxFaces5 : R.id.trackingMaxFaces10);
+        trackingMinFaceGroup.check(settings.minFacePx == 48 ? R.id.trackingMinFace48
+                : settings.minFacePx == 64 ? R.id.trackingMinFace64
+                : settings.minFacePx == 128
+                ? R.id.trackingMinFace128 : R.id.trackingMinFace24);
+        suppressTrackingSettingChanges = false;
+    }
+
+    private void configureTrackingAutoApply() {
+        trackingModeGroup.setOnCheckedStateChangeListener(
+                (group, checkedIds) -> onTrackingSettingChanged());
+        trackingInputPxGroup.setOnCheckedStateChangeListener(
+                (group, checkedIds) -> onTrackingSettingChanged());
+        trackingMaxFacesGroup.setOnCheckedStateChangeListener(
+                (group, checkedIds) -> onTrackingSettingChanged());
+        trackingMinFaceGroup.setOnCheckedStateChangeListener(
+                (group, checkedIds) -> onTrackingSettingChanged());
+    }
+
+    private void onTrackingSettingChanged() {
+        if (!suppressTrackingSettingChanges) {
+            applyTrackingSettings(selectedTrackingSettings());
+        }
+    }
+
+    private static boolean sameTrackingSettings(
+            FaceTrackingSessionSettings.Values left,
+            FaceTrackingSessionSettings.Values right) {
+        return left.mode == right.mode && left.inputPx == right.inputPx
+                && left.maxFaces == right.maxFaces
+                && left.minFacePx == right.minFacePx;
+    }
+
+    private FaceTrackingSessionSettings.Values selectedTrackingSettings() {
+        int mode = trackingModeGroup.getCheckedChipId() == R.id.trackingModeTbd
+                ? InspireFace.DETECT_MODE_TRACK_BY_DETECTION
+                : InspireFace.DETECT_MODE_LIGHT_TRACK;
+        int inputId = trackingInputPxGroup.getCheckedChipId();
+        int inputPx = inputId == R.id.trackingInputPx320 ? 320
+                : inputId == R.id.trackingInputPx1280 ? 1280 : 640;
+        int maxId = trackingMaxFacesGroup.getCheckedChipId();
+        int maxFaces = maxId == R.id.trackingMaxFaces1 ? 1
+                : maxId == R.id.trackingMaxFaces3 ? 3
+                : maxId == R.id.trackingMaxFaces5 ? 5 : 10;
+        int minId = trackingMinFaceGroup.getCheckedChipId();
+        int minFacePx = minId == R.id.trackingMinFace48 ? 48
+                : minId == R.id.trackingMinFace64 ? 64
+                : minId == R.id.trackingMinFace128 ? 128 : 24;
+        return new FaceTrackingSessionSettings.Values(
+                mode, inputPx, maxFaces, minFacePx);
+    }
+
+    private void setControlsEnabled(boolean enabled) {
+        choosePhotoButton.setEnabled(enabled);
+        applySessionButton.setEnabled(enabled);
+        resetSessionButton.setEnabled(enabled);
+        landmarkSwitch.setEnabled(enabled);
+        setChipGroupEnabled(inputPxGroup, enabled);
+        setChipGroupEnabled(maxFacesGroup, enabled);
+        setChipGroupEnabled(minFaceGroup, enabled);
+    }
+
+    private void setTrackingControlsEnabled(boolean enabled) {
+        resetTrackingButton.setEnabled(enabled);
+        setChipGroupEnabled(trackingModeGroup, enabled);
+        setChipGroupEnabled(trackingInputPxGroup, enabled);
+        setChipGroupEnabled(trackingMaxFacesGroup, enabled);
+        setChipGroupEnabled(trackingMinFaceGroup, enabled);
+    }
+
+    private void setTrackingLoading(boolean loading) {
+        trackingLoadingIndicator.setVisibility(loading ? View.VISIBLE : View.GONE);
+    }
+
+    private static void setChipGroupEnabled(ChipGroup group, boolean enabled) {
+        for (int i = 0; i < group.getChildCount(); i++) {
+            group.getChildAt(i).setEnabled(enabled);
+        }
+    }
+
+    private void setImageStatus(int textRes, @ColorRes int colorRes) {
+        imageStatus.setText(textRes);
+        imageStatus.setTextColor(color(colorRes));
+    }
+
+    private int color(@ColorRes int colorRes) {
+        return ContextCompat.getColor(this, colorRes);
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.detectionRoot);
+        int left = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int right = root.getPaddingRight();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(left + bars.left, top + bars.top,
+                    right + bars.right, bottom + bars.bottom);
+            return insets;
+        });
+    }
+
+    private void postUi(Runnable action) {
+        runOnUiThread(() -> {
+            if (!destroyed && !isFinishing()) {
+                action.run();
+            }
+        });
+    }
+
+    private static void recycle(@Nullable Bitmap bitmap) {
+        if (bitmap != null && !bitmap.isRecycled()) {
+            bitmap.recycle();
+        }
+    }
+
+    @Override
+    protected void onResume() {
+        super.onResume();
+        cameraPermission.onResume();
+    }
+
+    @Override
+    protected void onStart() {
+        super.onStart();
+        trackingGlOverlay.onResume();
+    }
+
+    @Override
+    protected void onStop() {
+        trackingGlOverlay.onPause();
+        super.onStop();
+    }
+
+    @Override
+    protected void onDestroy() {
+        cameraPermission.close();
+        destroyed = true;
+        engineReady = false;
+        imageVersion++;
+        stopVideoTracking();
+        Bitmap imageToRecycle = imageBitmap;
+        Bitmap magnifierToRecycle = magnifierBitmap;
+        imageBitmap = null;
+        magnifierBitmap = null;
+        sdkExecutor.execute(() -> {
+            recycle(imageToRecycle);
+            recycle(magnifierToRecycle);
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        });
+        sdkExecutor.shutdown();
+        trackingExecutor.shutdown();
+        super.onDestroy();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceManagementActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceManagementActivity.java
new file mode 100644
index 000000000..28d93017b
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceManagementActivity.java
@@ -0,0 +1,603 @@
+package com.example.inspireface_example;
+
+import android.content.Intent;
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.net.Uri;
+import android.os.Bundle;
+import android.text.Editable;
+import android.text.TextWatcher;
+import android.util.LruCache;
+import android.view.LayoutInflater;
+import android.view.View;
+import android.view.ViewGroup;
+import android.widget.BaseAdapter;
+import android.widget.Button;
+import android.widget.EditText;
+import android.widget.ImageView;
+import android.widget.ListView;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.annotation.Nullable;
+import androidx.appcompat.app.AlertDialog;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.content.ContextCompat;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.face.FaceImageProcessor;
+import com.example.inspireface_example.face.FaceRecord;
+import com.example.inspireface_example.face.FaceRepository;
+import com.example.inspireface_example.face.ImageBitmapLoader;
+import com.example.inspireface_example.view.FaceCaptureActivity;
+import com.example.inspireface_example.view.FaceEngine;
+import com.example.inspireface_example.widget.FaceImageOverlayView;
+import com.google.android.material.button.MaterialButton;
+import com.google.android.material.dialog.MaterialAlertDialogBuilder;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.io.File;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** CRUD demo for a model-isolated persistent FeatureHub and its face crops. */
+public class FaceManagementActivity extends AppCompatActivity {
+
+    private static final int MAX_IMAGE_DIMENSION = 2048;
+
+    private final ExecutorService sdkExecutor = Executors.newSingleThreadExecutor();
+    private final FaceRecordAdapter adapter = new FaceRecordAdapter();
+
+    private FaceModelPrefs.Model model;
+    private FaceRepository repository;
+    private Session session;
+    private MaterialButton addButton;
+    private EditText searchInput;
+    private TextView faceCount;
+    private View loadingIndicator;
+    private View emptyState;
+    private volatile boolean destroyed;
+
+    // Editor dialog state
+    private AlertDialog editorDialog;
+    private FaceRecord editingRecord;
+    private ImageView editorImage;
+    private FaceImageOverlayView editorOverlay;
+    private View editorPlaceholder;
+    private TextView editorStatus;
+    private EditText editorName;
+    private MaterialButton captureFaceButton;
+    private MaterialButton chooseImageButton;
+    private Button saveButton;
+    private Bitmap editorBitmap;
+    private FaceImageProcessor.Candidate[] editorCandidates;
+    private int editorSelectedIndex = -1;
+    private volatile int editorVersion;
+    private boolean editorSaving;
+
+    private final ActivityResultLauncher<String> editorImagePicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null && editorDialog != null) {
+                    loadEditorImage(uri);
+                }
+            });
+
+    private final ActivityResultLauncher<Intent> editorCamera =
+            registerForActivityResult(new ActivityResultContracts.StartActivityForResult(), result -> {
+                Intent data = result.getData();
+                String path = data == null ? null
+                        : data.getStringExtra(FaceCaptureActivity.EXTRA_CAPTURE_PATH);
+                if (result.getResultCode() == RESULT_OK && path != null) {
+                    if (editorDialog != null) {
+                        loadEditorCapture(path);
+                    } else {
+                        new File(path).delete();
+                    }
+                }
+            });
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_face_management);
+        applyWindowInsets();
+
+        model = FaceModelPrefs.get(this);
+        repository = new FaceRepository(this, model);
+        addButton = findViewById(R.id.btnAddFace);
+        searchInput = findViewById(R.id.searchInput);
+        faceCount = findViewById(R.id.faceCount);
+        loadingIndicator = findViewById(R.id.loadingIndicator);
+        emptyState = findViewById(R.id.emptyState);
+        ListView faceList = findViewById(R.id.faceList);
+        faceList.setAdapter(adapter);
+        faceList.setEmptyView(emptyState);
+        emptyState.setVisibility(View.GONE);
+
+        ((TextView) findViewById(R.id.currentModel)).setText(
+                getString(R.string.current_model, model.sdkName()));
+        ((TextView) findViewById(R.id.storageScope)).setText(
+                getString(R.string.face_storage_scope, model.sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        addButton.setOnClickListener(v -> openEditor(null));
+        searchInput.addTextChangedListener(new SimpleTextWatcher() {
+            @Override
+            public void afterTextChanged(Editable editable) {
+                refreshRecords();
+            }
+        });
+
+        sdkExecutor.execute(this::initializeLibrary);
+    }
+
+    private void initializeLibrary() {
+        boolean ready = FaceEngine.ensureLaunched(this);
+        if (ready) {
+            session = FaceEngine.createRecognitionSession();
+            ready = session != null;
+        }
+        if (ready) {
+            // Keep an active session for the whole lifetime of FeatureHub. FaceEngine then
+            // cannot terminate or switch the loaded model while this repository is open.
+            ready = repository.open();
+        }
+        if (!ready) {
+            repository.close();
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        }
+        boolean finalReady = ready;
+        postUi(() -> {
+            loadingIndicator.setVisibility(View.GONE);
+            addButton.setEnabled(finalReady);
+            if (finalReady) {
+                refreshRecords();
+            } else {
+                emptyState.setVisibility(View.VISIBLE);
+                Toast.makeText(this, R.string.face_library_failed, Toast.LENGTH_LONG).show();
+            }
+        });
+    }
+
+    private void refreshRecords() {
+        if (repository == null) {
+            return;
+        }
+        String keyword = searchInput == null ? null : searchInput.getText().toString();
+        List<FaceRecord> records = repository.query(keyword);
+        adapter.setRecords(records);
+        faceCount.setText(getString(R.string.face_list_count,
+                keyword == null || keyword.trim().isEmpty()
+                        ? records.size() : repository.query(null).size()));
+    }
+
+    private void openEditor(@Nullable FaceRecord record) {
+        if (session == null || editorDialog != null) {
+            return;
+        }
+        editingRecord = record;
+        editorCandidates = null;
+        editorSelectedIndex = -1;
+        editorSaving = false;
+        editorVersion++;
+
+        View content = getLayoutInflater().inflate(R.layout.dialog_face_editor, null, false);
+        editorImage = content.findViewById(R.id.editorImage);
+        editorOverlay = content.findViewById(R.id.editorFaceOverlay);
+        editorPlaceholder = content.findViewById(R.id.editorPlaceholder);
+        editorStatus = content.findViewById(R.id.editorImageStatus);
+        editorName = content.findViewById(R.id.editorName);
+        captureFaceButton = content.findViewById(R.id.btnCaptureEditorFace);
+        chooseImageButton = content.findViewById(R.id.btnChooseEditorImage);
+        editorOverlay.setOnFaceSelectedListener(this::selectEditorFace);
+
+        if (record != null) {
+            editorName.setText(record.name);
+            editorBitmap = BitmapFactory.decodeFile(record.cropPath);
+            if (editorBitmap != null) {
+                editorImage.setImageBitmap(editorBitmap);
+                editorPlaceholder.setVisibility(View.GONE);
+                setEditorStatus(getString(R.string.stored_face_crop), true);
+            }
+        } else {
+            editorBitmap = null;
+        }
+
+        editorDialog = new MaterialAlertDialogBuilder(this)
+                .setTitle(record == null ? R.string.add_face_title : R.string.edit_face_title)
+                .setView(content)
+                .setNegativeButton(R.string.cancel, null)
+                .setPositiveButton(R.string.save, null)
+                .create();
+        editorDialog.setOnShowListener(ignored -> {
+            saveButton = editorDialog.getButton(AlertDialog.BUTTON_POSITIVE);
+            saveButton.setEnabled(record != null);
+            saveButton.setOnClickListener(v -> saveEditor());
+            captureFaceButton.setOnClickListener(v -> editorCamera.launch(
+                    new Intent(this, FaceCaptureActivity.class)));
+            chooseImageButton.setOnClickListener(v -> editorImagePicker.launch("image/*"));
+        });
+        editorDialog.setOnDismissListener(ignored -> clearEditorState());
+        editorDialog.show();
+    }
+
+    private void loadEditorImage(Uri uri) {
+        loadEditorBitmap(() -> ImageBitmapLoader.decode(
+                this, uri, MAX_IMAGE_DIMENSION), null);
+    }
+
+    private void loadEditorCapture(String path) {
+        File temporaryCapture = new File(path);
+        loadEditorBitmap(() -> {
+            Bitmap bitmap = BitmapFactory.decodeFile(temporaryCapture.getAbsolutePath());
+            if (bitmap == null) {
+                throw new IOException("Unable to decode camera capture");
+            }
+            return bitmap;
+        }, temporaryCapture);
+    }
+
+    private void loadEditorBitmap(EditorBitmapSource source,
+                                  @Nullable File deleteAfterDecode) {
+        int requestVersion = ++editorVersion;
+        editorSelectedIndex = -1;
+        editorOverlay.clearFace();
+        saveButton.setEnabled(false);
+        setEditorStatus(getString(R.string.image_analyzing), false);
+        sdkExecutor.execute(() -> {
+            Bitmap bitmap = null;
+            FaceImageProcessor.Result result = null;
+            try {
+                bitmap = source.decode();
+                if (editorVersion != requestVersion || destroyed) {
+                    bitmap.recycle();
+                    return;
+                }
+                result = FaceImageProcessor.detect(session, bitmap, true);
+            } catch (Exception ignored) {
+                // The UI maps a null result to a load failure below.
+            } finally {
+                if (deleteAfterDecode != null) {
+                    deleteAfterDecode.delete();
+                }
+            }
+            Bitmap deliveredBitmap = bitmap;
+            FaceImageProcessor.Result deliveredResult = result;
+            // Always deliver detector ownership to the UI callback. applyEditorImage also
+            // handles a dismissed/destroyed editor and recycles an obsolete result.
+            runOnUiThread(() -> applyEditorImage(
+                    requestVersion, deliveredBitmap, deliveredResult));
+        });
+    }
+
+    private void applyEditorImage(int requestVersion, @Nullable Bitmap bitmap,
+                                  @Nullable FaceImageProcessor.Result result) {
+        if (editorDialog == null || editorVersion != requestVersion) {
+            recycleEditorResult(bitmap, result);
+            return;
+        }
+        Bitmap previousBitmap = editorBitmap;
+        FaceImageProcessor.Candidate[] previousCandidates = editorCandidates;
+        editorBitmap = bitmap;
+        editorCandidates = result == null ? null : result.candidates;
+        editorSelectedIndex = editorCandidates != null && editorCandidates.length > 0 ? 0 : -1;
+        if (bitmap != null) {
+            editorImage.setImageBitmap(bitmap);
+            editorPlaceholder.setVisibility(View.GONE);
+            editorOverlay.showFaces(bitmap.getWidth(), bitmap.getHeight(),
+                    result == null ? null : result.faceRects(), editorSelectedIndex);
+        } else {
+            editorImage.setImageDrawable(null);
+            editorPlaceholder.setVisibility(View.VISIBLE);
+            editorOverlay.clearFace();
+        }
+        recycleEditorAssets(previousBitmap, previousCandidates);
+        if (result == null) {
+            setEditorStatus(getString(R.string.image_load_failed), false);
+        } else if (result.status == FaceImageProcessor.Status.NO_FACE) {
+            setEditorStatus(getString(R.string.image_no_face), false);
+        } else if (result.status != FaceImageProcessor.Status.READY) {
+            setEditorStatus(getString(R.string.face_extract_failed), false);
+        } else {
+            updateEditorSelectionState();
+        }
+    }
+
+    private void selectEditorFace(int index) {
+        if (editorCandidates == null || index < 0 || index >= editorCandidates.length) {
+            return;
+        }
+        editorSelectedIndex = index;
+        editorOverlay.setSelectedIndex(index);
+        updateEditorSelectionState();
+    }
+
+    private void updateEditorSelectionState() {
+        FaceImageProcessor.Candidate candidate = selectedEditorCandidate();
+        boolean valid = candidate != null && candidate.feature != null && candidate.crop != null;
+        if (editorCandidates != null && editorCandidates.length > 1) {
+            setEditorStatus(getString(R.string.image_face_selected,
+                    editorSelectedIndex + 1, editorCandidates.length), valid);
+        } else {
+            setEditorStatus(getString(valid ? R.string.image_face_ready
+                    : R.string.face_extract_failed), valid);
+        }
+        saveButton.setEnabled(valid);
+    }
+
+    @Nullable
+    private FaceImageProcessor.Candidate selectedEditorCandidate() {
+        return editorCandidates != null && editorSelectedIndex >= 0
+                && editorSelectedIndex < editorCandidates.length
+                ? editorCandidates[editorSelectedIndex] : null;
+    }
+
+    private void saveEditor() {
+        FaceImageProcessor.Candidate selected = selectedEditorCandidate();
+        if (editingRecord == null && selected == null) {
+            return;
+        }
+        String name = editorName.getText() == null ? "" : editorName.getText().toString();
+        FaceFeature feature = selected == null ? null : selected.feature;
+        Bitmap crop = selected == null ? null : selected.crop;
+        long editingId = editingRecord == null ? -1L : editingRecord.id;
+        editorSaving = true;
+        editorDialog.setCancelable(false);
+        saveButton.setEnabled(false);
+        editorDialog.getButton(AlertDialog.BUTTON_NEGATIVE).setEnabled(false);
+        captureFaceButton.setEnabled(false);
+        chooseImageButton.setEnabled(false);
+        setEditorStatus(getString(R.string.saving_face), true);
+
+        sdkExecutor.execute(() -> {
+            boolean success;
+            if (editingId < 0) {
+                success = repository.insert(name, feature, crop).success;
+            } else {
+                success = repository.update(editingId, name, feature, crop);
+            }
+            postUi(() -> {
+                editorSaving = false;
+                if (success) {
+                    Toast.makeText(this, R.string.face_saved, Toast.LENGTH_SHORT).show();
+                    adapter.clearCropCache();
+                    editorDialog.dismiss();
+                    refreshRecords();
+                } else {
+                    editorDialog.setCancelable(true);
+                    editorDialog.getButton(AlertDialog.BUTTON_NEGATIVE).setEnabled(true);
+                    captureFaceButton.setEnabled(true);
+                    chooseImageButton.setEnabled(true);
+                    saveButton.setEnabled(editingRecord != null || selectedEditorCandidate() != null);
+                    setEditorStatus(getString(R.string.face_save_failed), false);
+                }
+            });
+        });
+    }
+
+    private void confirmDelete(FaceRecord record) {
+        new MaterialAlertDialogBuilder(this)
+                .setTitle(R.string.delete_face_title)
+                .setMessage(getString(R.string.delete_face_message, record.name, model.sdkName()))
+                .setNegativeButton(R.string.cancel, null)
+                .setPositiveButton(R.string.delete, (dialog, which) -> deleteRecord(record))
+                .show();
+    }
+
+    private void deleteRecord(FaceRecord record) {
+        addButton.setEnabled(false);
+        sdkExecutor.execute(() -> {
+            boolean deleted = repository.delete(record.id);
+            postUi(() -> {
+                addButton.setEnabled(true);
+                Toast.makeText(this, deleted ? R.string.face_deleted
+                        : R.string.face_delete_failed, Toast.LENGTH_SHORT).show();
+                if (deleted) {
+                    adapter.removeCrop(record.cropPath);
+                    refreshRecords();
+                }
+            });
+        });
+    }
+
+    private void setEditorStatus(CharSequence message, boolean positive) {
+        editorStatus.setText(message);
+        editorStatus.setTextColor(ContextCompat.getColor(this,
+                positive ? R.color.liveness_accent : R.color.home_text_secondary));
+    }
+
+    private void clearEditorState() {
+        editorVersion++;
+        if (!editorSaving) {
+            recycleEditorAssets(editorBitmap, editorCandidates);
+        }
+        editorDialog = null;
+        editingRecord = null;
+        editorImage = null;
+        editorOverlay = null;
+        editorPlaceholder = null;
+        editorStatus = null;
+        editorName = null;
+        captureFaceButton = null;
+        chooseImageButton = null;
+        saveButton = null;
+        editorBitmap = null;
+        editorCandidates = null;
+        editorSelectedIndex = -1;
+    }
+
+    private static void recycleEditorResult(@Nullable Bitmap bitmap,
+                                            @Nullable FaceImageProcessor.Result result) {
+        recycleEditorAssets(bitmap, result == null ? null : result.candidates);
+    }
+
+    private static void recycleEditorAssets(@Nullable Bitmap bitmap,
+                                            @Nullable FaceImageProcessor.Candidate[] candidates) {
+        FaceImageProcessor.recycleCrops(candidates, -1);
+        if (bitmap != null && !bitmap.isRecycled()) {
+            bitmap.recycle();
+        }
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.managementRoot);
+        int left = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int right = root.getPaddingRight();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(left + bars.left, top + bars.top,
+                    right + bars.right, bottom + bars.bottom);
+            return insets;
+        });
+    }
+
+    private void postUi(Runnable action) {
+        runOnUiThread(() -> {
+            if (!destroyed && !isFinishing()) {
+                action.run();
+            }
+        });
+    }
+
+    @Override
+    protected void onDestroy() {
+        destroyed = true;
+        editorVersion++;
+        Bitmap deferredEditorBitmap = editorBitmap;
+        FaceImageProcessor.Candidate[] deferredEditorCandidates = editorCandidates;
+        if (editorDialog != null && !editorSaving) {
+            editorDialog.dismiss();
+        }
+        // The save task may still own the selected crop for JPEG compression. Queue a
+        // second, idempotent cleanup behind all SDK tasks; this also covers dialog teardown.
+        Bitmap bitmapToRecycle = deferredEditorBitmap;
+        FaceImageProcessor.Candidate[] candidatesToRecycle = deferredEditorCandidates;
+        sdkExecutor.execute(() -> {
+            recycleEditorAssets(bitmapToRecycle, candidatesToRecycle);
+            repository.close();
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        });
+        sdkExecutor.shutdown();
+        super.onDestroy();
+    }
+
+    private final class FaceRecordAdapter extends BaseAdapter {
+        private final List<FaceRecord> records = new ArrayList<>();
+        private final LruCache<String, Bitmap> cropCache = new LruCache<String, Bitmap>(8) {
+            @Override
+            protected int sizeOf(String key, Bitmap value) {
+                return 1;
+            }
+        };
+
+        void setRecords(List<FaceRecord> newRecords) {
+            records.clear();
+            records.addAll(newRecords);
+            notifyDataSetChanged();
+        }
+
+        void removeCrop(String path) {
+            cropCache.remove(path);
+        }
+
+        void clearCropCache() {
+            cropCache.evictAll();
+        }
+
+        @Override
+        public int getCount() {
+            return records.size();
+        }
+
+        @Override
+        public FaceRecord getItem(int position) {
+            return records.get(position);
+        }
+
+        @Override
+        public long getItemId(int position) {
+            return getItem(position).id;
+        }
+
+        @Override
+        public boolean hasStableIds() {
+            return true;
+        }
+
+        @Override
+        public View getView(int position, View convertView, ViewGroup parent) {
+            ViewHolder holder;
+            if (convertView == null) {
+                convertView = LayoutInflater.from(parent.getContext())
+                        .inflate(R.layout.item_face_record, parent, false);
+                holder = new ViewHolder(convertView);
+                convertView.setTag(holder);
+            } else {
+                holder = (ViewHolder) convertView.getTag();
+            }
+            FaceRecord record = getItem(position);
+            holder.name.setText(record.name);
+            holder.id.setText(getString(R.string.face_id_format, record.id));
+            Bitmap crop = cropCache.get(record.cropPath);
+            if (crop == null || crop.isRecycled()) {
+                crop = BitmapFactory.decodeFile(record.cropPath);
+                if (crop != null) {
+                    cropCache.put(record.cropPath, crop);
+                }
+            }
+            holder.crop.setImageBitmap(crop);
+            holder.edit.setOnClickListener(v -> openEditor(record));
+            holder.delete.setOnClickListener(v -> confirmDelete(record));
+            return convertView;
+        }
+    }
+
+    private static final class ViewHolder {
+        final ImageView crop;
+        final TextView name;
+        final TextView id;
+        final View edit;
+        final View delete;
+
+        ViewHolder(View root) {
+            crop = root.findViewById(R.id.faceCrop);
+            name = root.findViewById(R.id.faceName);
+            id = root.findViewById(R.id.faceId);
+            edit = root.findViewById(R.id.btnEditFace);
+            delete = root.findViewById(R.id.btnDeleteFace);
+        }
+    }
+
+    private abstract static class SimpleTextWatcher implements TextWatcher {
+        @Override
+        public void beforeTextChanged(CharSequence s, int start, int count, int after) {
+        }
+
+        @Override
+        public void onTextChanged(CharSequence s, int start, int before, int count) {
+        }
+    }
+
+    private interface EditorBitmapSource {
+        Bitmap decode() throws Exception;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceModelPrefs.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceModelPrefs.java
new file mode 100644
index 000000000..533a4e4a6
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceModelPrefs.java
@@ -0,0 +1,49 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+
+/** Process-wide model selection shared by the home screen and all feature screens. */
+public final class FaceModelPrefs {
+
+    public enum Model {
+        PIKACHU("Pikachu"),
+        MEGATRON("Megatron");
+
+        private final String sdkName;
+
+        Model(String sdkName) {
+            this.sdkName = sdkName;
+        }
+
+        public String sdkName() {
+            return sdkName;
+        }
+
+        static Model fromStored(String value) {
+            for (Model model : values()) {
+                if (model.sdkName.equals(value)) {
+                    return model;
+                }
+            }
+            return MEGATRON;
+        }
+    }
+
+    private static final String PREFS = "settings";
+    private static final String KEY_MODEL = "face_model";
+
+    private FaceModelPrefs() {
+    }
+
+    /** Megatron remains the default to preserve the project's previous behavior. */
+    public static Model get(Context context) {
+        String value = context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .getString(KEY_MODEL, Model.MEGATRON.sdkName());
+        return Model.fromStored(value);
+    }
+
+    public static void set(Context context, Model model) {
+        context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .edit().putString(KEY_MODEL, model.sdkName()).apply();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceRecognitionActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceRecognitionActivity.java
new file mode 100644
index 000000000..b5daab87c
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceRecognitionActivity.java
@@ -0,0 +1,918 @@
+package com.example.inspireface_example;
+
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.graphics.RectF;
+import android.net.Uri;
+import android.os.Bundle;
+import android.view.View;
+import android.widget.ImageView;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.annotation.ColorRes;
+import androidx.annotation.Nullable;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.camera.view.PreviewView;
+import androidx.core.content.ContextCompat;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.face.FaceImageProcessor;
+import com.example.inspireface_example.face.FaceRecord;
+import com.example.inspireface_example.face.FaceRepository;
+import com.example.inspireface_example.face.ImageBitmapLoader;
+import com.example.inspireface_example.permission.CameraPermissionCoordinator;
+import com.example.inspireface_example.view.CameraPreviewController;
+import com.example.inspireface_example.view.FaceEngine;
+import com.example.inspireface_example.view.FaceOverlayView;
+import com.example.inspireface_example.view.RecognitionFaceAnalyzer;
+import com.example.inspireface_example.widget.FaceImageOverlayView;
+import com.google.android.material.button.MaterialButton;
+import com.google.android.material.chip.ChipGroup;
+import com.google.android.material.tabs.TabLayout;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** Model-scoped 1:N recognition from selectable still-image faces or camera face 0. */
+public final class FaceRecognitionActivity extends AppCompatActivity {
+
+    private static final int MAX_IMAGE_DIMENSION = 2048;
+    private static final float SELECTED_CROP_HIDE_AREA_RATIO = 0.15f;
+    private static final float SELECTED_CROP_SCALE = 1.5f;
+
+    private final ExecutorService sdkExecutor = Executors.newSingleThreadExecutor();
+
+    private FaceModelPrefs.Model model;
+    private FaceRepository repository;
+    private Session session;
+    private volatile boolean destroyed;
+    private volatile int imageVersion;
+    private boolean libraryEmpty;
+    private boolean sessionBusy = true;
+    private boolean recognitionReady;
+    private StillImageSessionSettings.Values savedSettings;
+
+    private ChipGroup inputPxGroup;
+    private ChipGroup maxFacesGroup;
+    private ChipGroup minFaceGroup;
+    private MaterialButton applySessionButton;
+    private MaterialButton resetSessionButton;
+    private MaterialButton choosePhotoButton;
+    private TextView sessionStatus;
+    private View emptyLibraryTip;
+    private TextView emptyLibraryTipText;
+    private ImageView imageView;
+    private FaceImageOverlayView faceOverlay;
+    private View selectedFaceCropCard;
+    private ImageView selectedFaceCropView;
+    private View imagePlaceholder;
+    private TextView imageStatus;
+    private View resultCard;
+    private ImageView resultCropView;
+    private TextView resultStatus;
+    private TextView resultName;
+    private TextView resultDetails;
+    private View sessionSettingsContent;
+    private TextView sessionSettingsToggleText;
+
+    private PreviewView videoPreview;
+    private FaceOverlayView videoOverlay;
+    private TextView videoStatus;
+    private TextView videoName;
+    private TextView videoDetails;
+    private View flipVideoButton;
+    private CameraPreviewController videoCameraController;
+    private RecognitionFaceAnalyzer videoAnalyzer;
+    private boolean videoSelected;
+    private boolean videoStarting;
+    private CameraPermissionCoordinator cameraPermission;
+    private int videoGeneration;
+
+    private Bitmap imageBitmap;
+    private Bitmap resultCropBitmap;
+    private Bitmap selectedFaceCropBitmap;
+    private FaceImageProcessor.Candidate[] candidates;
+    private int selectedIndex = -1;
+
+    private final ActivityResultLauncher<String> photoPicker =
+            registerForActivityResult(new ActivityResultContracts.GetContent(), uri -> {
+                if (uri != null && !sessionBusy && session != null) {
+                    loadPhoto(uri);
+                }
+            });
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        cameraPermission = new CameraPermissionCoordinator(this,
+                new CameraPermissionCoordinator.Listener() {
+                    @Override
+                    public void onCameraPermissionGranted() {
+                        startVideoRecognition();
+                    }
+
+                    @Override
+                    public void onCameraPermissionBlocked(boolean requiresSettings) {
+                        if (videoSelected) {
+                            showVideoError(R.string.msg_permission_required,
+                                    requiresSettings
+                                            ? R.string.camera_permission_settings_hint
+                                            : R.string.camera_permission_retry_hint);
+                        }
+                    }
+                });
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_face_recognition);
+        applyWindowInsets();
+
+        model = FaceModelPrefs.get(this);
+        repository = new FaceRepository(this, model);
+        savedSettings = StillImageSessionSettings.load(this);
+        bindViews();
+        cameraPermission.bindRecoveryButton(
+                findViewById(R.id.btnCameraPermissionAction));
+        ((TextView) findViewById(R.id.currentModel)).setText(
+                getString(R.string.current_model, model.sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        choosePhotoButton.setOnClickListener(v -> photoPicker.launch("image/*"));
+        findViewById(R.id.recognitionImageCard)
+                .setOnClickListener(v -> {
+                    if (!sessionBusy && session != null) {
+                        photoPicker.launch("image/*");
+                    }
+                });
+        faceOverlay.setOnFaceSelectedListener(this::selectFace);
+        findViewById(R.id.sessionSettingsHeader).setOnClickListener(
+                v -> setSettingsExpanded(
+                        sessionSettingsContent.getVisibility() != View.VISIBLE));
+        applySessionButton.setOnClickListener(v -> rebuildSession(
+                selectedSettings(), true));
+        resetSessionButton.setOnClickListener(v -> {
+            StillImageSessionSettings.Values defaults = StillImageSessionSettings.defaults();
+            selectParameters(defaults);
+            rebuildSession(defaults, true);
+        });
+        flipVideoButton.setOnClickListener(v -> flipVideoCamera());
+        configureTabs();
+        selectParameters(savedSettings);
+        setSettingsExpanded(false);
+        setControlsEnabled(false);
+        sessionStatus.setText(R.string.recognition_session_rebuilding);
+        sdkExecutor.execute(this::initializeRecognition);
+    }
+
+    private void bindViews() {
+        inputPxGroup = findViewById(R.id.inputPxGroup);
+        maxFacesGroup = findViewById(R.id.maxFacesGroup);
+        minFaceGroup = findViewById(R.id.minFaceGroup);
+        applySessionButton = findViewById(R.id.btnApplySession);
+        resetSessionButton = findViewById(R.id.btnResetSession);
+        choosePhotoButton = findViewById(R.id.btnChooseRecognitionPhoto);
+        sessionStatus = findViewById(R.id.sessionStatus);
+        emptyLibraryTip = findViewById(R.id.emptyLibraryTip);
+        emptyLibraryTipText = findViewById(R.id.emptyLibraryTipText);
+        imageView = findViewById(R.id.recognitionImage);
+        faceOverlay = findViewById(R.id.recognitionFaceOverlay);
+        selectedFaceCropCard = findViewById(R.id.recognitionSelectedFaceCropCard);
+        selectedFaceCropView = findViewById(R.id.recognitionSelectedFaceCrop);
+        imagePlaceholder = findViewById(R.id.recognitionImagePlaceholder);
+        imageStatus = findViewById(R.id.recognitionImageStatus);
+        resultCard = findViewById(R.id.recognitionResultCard);
+        resultCropView = findViewById(R.id.recognitionResultCrop);
+        resultStatus = findViewById(R.id.recognitionResultStatus);
+        resultName = findViewById(R.id.recognitionResultName);
+        resultDetails = findViewById(R.id.recognitionResultDetails);
+        sessionSettingsContent = findViewById(R.id.sessionSettingsContent);
+        sessionSettingsToggleText = findViewById(R.id.sessionSettingsToggleText);
+        videoPreview = findViewById(R.id.recognitionVideoPreview);
+        videoOverlay = findViewById(R.id.recognitionVideoOverlay);
+        videoStatus = findViewById(R.id.videoRecognitionStatus);
+        videoName = findViewById(R.id.videoRecognitionName);
+        videoDetails = findViewById(R.id.videoRecognitionDetails);
+        flipVideoButton = findViewById(R.id.btnFlipRecognitionCamera);
+    }
+
+    private void configureTabs() {
+        View photoPanel = findViewById(R.id.photoRecognitionPanel);
+        View videoPanel = findViewById(R.id.videoRecognitionPanel);
+        ((TabLayout) findViewById(R.id.recognitionTabs))
+                .addOnTabSelectedListener(new TabLayout.OnTabSelectedListener() {
+                    @Override
+                    public void onTabSelected(TabLayout.Tab tab) {
+                        boolean photo = tab.getPosition() == 0;
+                        videoSelected = !photo;
+                        photoPanel.setVisibility(photo ? View.VISIBLE : View.GONE);
+                        videoPanel.setVisibility(photo ? View.GONE : View.VISIBLE);
+                        if (photo) {
+                            stopVideoRecognition();
+                        } else {
+                            startVideoRecognition();
+                        }
+                    }
+
+                    @Override
+                    public void onTabUnselected(TabLayout.Tab tab) {
+                    }
+
+                    @Override
+                    public void onTabReselected(TabLayout.Tab tab) {
+                    }
+                });
+    }
+
+    private void initializeRecognition() {
+        StillImageSessionSettings.Values initialSettings = savedSettings;
+        boolean ready = FaceEngine.ensureLaunched(this);
+        if (ready) {
+            session = FaceEngine.createRecognitionSession(
+                    initialSettings.inputPx,
+                    initialSettings.maxFaces,
+                    initialSettings.minFacePx);
+            ready = session != null;
+        }
+        if (ready) {
+            ready = repository.open();
+        }
+        if (!ready) {
+            repository.close();
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        }
+        boolean finalReady = ready;
+        boolean empty = ready && repository.query(null).isEmpty();
+        postUi(() -> {
+            sessionBusy = false;
+            recognitionReady = finalReady;
+            libraryEmpty = empty;
+            updateEmptyLibraryTip();
+            setControlsEnabled(finalReady);
+            if (finalReady) {
+                showSessionSummary(initialSettings);
+                if (videoSelected) {
+                    startVideoRecognition();
+                }
+            } else {
+                sessionStatus.setText(R.string.recognition_session_failed);
+                sessionStatus.setTextColor(color(R.color.liveness_fail));
+                setImageStatus(R.string.compare_engine_failed, R.color.liveness_fail);
+            }
+        });
+    }
+
+    private void updateEmptyLibraryTip() {
+        emptyLibraryTip.setVisibility(libraryEmpty ? View.VISIBLE : View.GONE);
+        if (libraryEmpty) {
+            emptyLibraryTipText.setText(
+                    getString(R.string.recognition_empty_library, model.sdkName()));
+        }
+    }
+
+    private void loadPhoto(Uri uri) {
+        int requestVersion = ++imageVersion;
+        sessionBusy = true;
+        candidates = null;
+        selectedIndex = -1;
+        faceOverlay.clearFace();
+        clearSelectedFaceCrop();
+        hideResult();
+        setControlsEnabled(false);
+        setImageStatus(R.string.image_analyzing, R.color.home_text_secondary);
+        sdkExecutor.execute(() -> {
+            Bitmap bitmap = null;
+            FaceImageProcessor.Result result = null;
+            try {
+                bitmap = ImageBitmapLoader.decode(this, uri, MAX_IMAGE_DIMENSION);
+                if (destroyed || imageVersion != requestVersion) {
+                    bitmap.recycle();
+                    return;
+                }
+                result = FaceImageProcessor.detect(session, bitmap, false);
+            } catch (Exception ignored) {
+                // A null result is rendered as a load failure.
+            }
+            Bitmap deliveredBitmap = bitmap;
+            FaceImageProcessor.Result deliveredResult = result;
+            runOnUiThread(() -> applyPhotoResult(
+                    requestVersion, deliveredBitmap, deliveredResult));
+        });
+    }
+
+    private void applyPhotoResult(int requestVersion, @Nullable Bitmap bitmap,
+                                  @Nullable FaceImageProcessor.Result result) {
+        if (destroyed || imageVersion != requestVersion) {
+            recycle(bitmap);
+            return;
+        }
+        Bitmap previous = imageBitmap;
+        imageBitmap = bitmap;
+        candidates = result == null ? null : result.candidates;
+        selectedIndex = candidates != null && candidates.length > 0 ? 0 : -1;
+        if (bitmap != null) {
+            imageView.setImageBitmap(bitmap);
+            imagePlaceholder.setVisibility(View.GONE);
+            faceOverlay.showFaces(bitmap.getWidth(), bitmap.getHeight(),
+                    result == null ? null : result.faceRects(), selectedIndex);
+            updateSelectedFaceCrop();
+        } else {
+            imageView.setImageDrawable(null);
+            imagePlaceholder.setVisibility(View.VISIBLE);
+            faceOverlay.clearFace();
+            clearSelectedFaceCrop();
+        }
+        if (previous != bitmap) {
+            recycle(previous);
+        }
+        finishDetectionUi(result);
+    }
+
+    private void finishDetectionUi(@Nullable FaceImageProcessor.Result result) {
+        sessionBusy = false;
+        setControlsEnabled(session != null);
+        if (result == null) {
+            setImageStatus(R.string.image_load_failed, R.color.liveness_fail);
+        } else if (result.status == FaceImageProcessor.Status.NO_FACE) {
+            setImageStatus(R.string.image_no_face, R.color.liveness_fail);
+        } else if (result.status != FaceImageProcessor.Status.READY
+                || candidates == null || candidates.length == 0) {
+            setImageStatus(R.string.face_extract_failed, R.color.liveness_fail);
+        } else {
+            updateSelectedFaceStatus();
+            recognizeSelectedFace();
+        }
+    }
+
+    private void selectFace(int index) {
+        if (sessionBusy || candidates == null || index < 0 || index >= candidates.length) {
+            return;
+        }
+        selectedIndex = index;
+        faceOverlay.setSelectedIndex(index);
+        updateSelectedFaceCrop();
+        updateSelectedFaceStatus();
+        recognizeSelectedFace();
+    }
+
+    private void updateSelectedFaceStatus() {
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        if (selected == null || selected.feature == null) {
+            setImageStatus(R.string.face_extract_failed, R.color.liveness_fail);
+        } else if (candidates.length > 1) {
+            imageStatus.setText(getString(R.string.image_face_selected,
+                    selectedIndex + 1, candidates.length));
+            imageStatus.setTextColor(color(R.color.liveness_accent));
+        } else {
+            setImageStatus(R.string.image_face_ready, R.color.liveness_accent);
+        }
+    }
+
+    @Nullable
+    private FaceImageProcessor.Candidate selectedCandidate() {
+        return candidates != null && selectedIndex >= 0 && selectedIndex < candidates.length
+                ? candidates[selectedIndex] : null;
+    }
+
+    private void updateSelectedFaceCrop() {
+        clearSelectedFaceCrop();
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        Bitmap source = imageBitmap;
+        if (selected == null || source == null || source.isRecycled()) {
+            return;
+        }
+        RectF face = selected.rect;
+        float imageArea = (float) source.getWidth() * source.getHeight();
+        float faceArea = Math.max(0f, face.width()) * Math.max(0f, face.height());
+        if (imageArea <= 0f || faceArea / imageArea >= SELECTED_CROP_HIDE_AREA_RATIO) {
+            return;
+        }
+
+        int cropSize = Math.max(2, Math.round(
+                Math.max(face.width(), face.height()) * SELECTED_CROP_SCALE));
+        cropSize = Math.min(cropSize, Math.min(source.getWidth(), source.getHeight()));
+        int left = Math.round(face.centerX() - cropSize / 2f);
+        int top = Math.round(face.centerY() - cropSize / 2f);
+        left = Math.max(0, Math.min(left, source.getWidth() - cropSize));
+        top = Math.max(0, Math.min(top, source.getHeight() - cropSize));
+        try {
+            Bitmap crop = Bitmap.createBitmap(source, left, top, cropSize, cropSize);
+            if (crop == source) {
+                return;
+            }
+            selectedFaceCropBitmap = crop;
+            selectedFaceCropView.setImageBitmap(crop);
+            selectedFaceCropCard.setVisibility(View.VISIBLE);
+        } catch (IllegalArgumentException ignored) {
+            // A malformed SDK rectangle simply leaves the optional preview hidden.
+        }
+    }
+
+    private void clearSelectedFaceCrop() {
+        if (selectedFaceCropView != null) {
+            selectedFaceCropView.setImageDrawable(null);
+        }
+        if (selectedFaceCropCard != null) {
+            selectedFaceCropCard.setVisibility(View.GONE);
+        }
+        recycle(selectedFaceCropBitmap);
+        selectedFaceCropBitmap = null;
+    }
+
+    private void recognizeSelectedFace() {
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        if (selected == null || selected.feature == null) {
+            hideResult();
+            return;
+        }
+        if (libraryEmpty) {
+            showEmptyLibraryResult();
+            return;
+        }
+        int version = imageVersion;
+        int faceIndex = selectedIndex;
+        FaceFeature feature = selected.feature;
+        showSearching();
+        sdkExecutor.execute(() -> {
+            FaceRepository.SearchResult search = repository.search(feature);
+            Bitmap crop = search.matched && search.record != null
+                    ? BitmapFactory.decodeFile(search.record.cropPath) : null;
+            runOnUiThread(() -> applySearchResult(
+                    version, faceIndex, feature, search, crop));
+        });
+    }
+
+    private void applySearchResult(int version, int faceIndex, FaceFeature feature,
+                                   FaceRepository.SearchResult search,
+                                   @Nullable Bitmap crop) {
+        FaceImageProcessor.Candidate selected = selectedCandidate();
+        if (destroyed || imageVersion != version || selectedIndex != faceIndex
+                || selected == null || selected.feature != feature) {
+            recycle(crop);
+            return;
+        }
+        if (search.matched && search.record != null) {
+            showMatch(search.record, search.confidence, search.threshold, crop);
+        } else {
+            recycle(crop);
+            showNoMatch(search);
+        }
+    }
+
+    private void showSearching() {
+        clearResultCrop();
+        resultCard.setVisibility(View.VISIBLE);
+        resultStatus.setText(R.string.recognition_searching);
+        resultStatus.setTextColor(color(R.color.home_text_secondary));
+        resultName.setText(R.string.recognition_result_name_unknown);
+        resultDetails.setText(null);
+    }
+
+    private void showMatch(FaceRecord record, float confidence,
+                           float threshold, @Nullable Bitmap crop) {
+        clearResultCrop();
+        resultCropBitmap = crop;
+        resultCropView.setImageBitmap(crop);
+        resultCropView.setVisibility(crop == null ? View.GONE : View.VISIBLE);
+        resultCard.setVisibility(View.VISIBLE);
+        resultStatus.setText(R.string.recognition_match_found);
+        resultStatus.setTextColor(color(R.color.liveness_accent));
+        resultName.setText(record.name);
+        resultDetails.setText(getString(R.string.recognition_result_details,
+                record.id, confidence, threshold));
+    }
+
+    private void showNoMatch(FaceRepository.SearchResult search) {
+        clearResultCrop();
+        resultCard.setVisibility(View.VISIBLE);
+        resultStatus.setText(R.string.recognition_no_match);
+        resultStatus.setTextColor(color(R.color.liveness_fail));
+        resultName.setText(R.string.recognition_result_name_unknown);
+        if (!Float.isNaN(search.confidence)) {
+            resultDetails.setText(getString(R.string.recognition_no_match_details,
+                    search.confidence, search.threshold));
+        } else {
+            resultDetails.setText(R.string.recognition_no_confidence);
+        }
+    }
+
+    private void showEmptyLibraryResult() {
+        clearResultCrop();
+        resultCard.setVisibility(View.VISIBLE);
+        resultStatus.setText(R.string.recognition_library_empty_result);
+        resultStatus.setTextColor(color(R.color.liveness_warn));
+        resultName.setText(model.sdkName());
+        resultDetails.setText(getString(
+                R.string.recognition_empty_library, model.sdkName()));
+    }
+
+    private void hideResult() {
+        clearResultCrop();
+        resultCard.setVisibility(View.GONE);
+    }
+
+    private void clearResultCrop() {
+        resultCropView.setImageDrawable(null);
+        resultCropView.setVisibility(View.GONE);
+        recycle(resultCropBitmap);
+        resultCropBitmap = null;
+    }
+
+    private void rebuildSession(StillImageSessionSettings.Values requested,
+                                boolean persistOnSuccess) {
+        if (sessionBusy) {
+            return;
+        }
+        int version = ++imageVersion;
+        Bitmap currentImage = imageBitmap;
+        sessionBusy = true;
+        candidates = null;
+        selectedIndex = -1;
+        faceOverlay.clearFace();
+        clearSelectedFaceCrop();
+        hideResult();
+        setControlsEnabled(false);
+        sessionStatus.setText(R.string.recognition_session_rebuilding);
+        sessionStatus.setTextColor(color(R.color.home_text_secondary));
+        if (currentImage != null) {
+            setImageStatus(R.string.image_analyzing, R.color.home_text_secondary);
+        }
+        sdkExecutor.execute(() -> {
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+            session = FaceEngine.createRecognitionSession(
+                    requested.inputPx, requested.maxFaces, requested.minFacePx);
+            FaceImageProcessor.Result result = session != null && currentImage != null
+                    ? FaceImageProcessor.detect(session, currentImage, false) : null;
+            boolean ready = session != null;
+            postUi(() -> {
+                if (imageVersion != version) {
+                    return;
+                }
+                sessionBusy = false;
+                setControlsEnabled(ready);
+                if (!ready) {
+                    sessionStatus.setText(R.string.recognition_session_failed);
+                    sessionStatus.setTextColor(color(R.color.liveness_fail));
+                    if (currentImage != null) {
+                        setImageStatus(R.string.compare_engine_failed, R.color.liveness_fail);
+                    }
+                    return;
+                }
+                savedSettings = requested;
+                if (persistOnSuccess) {
+                    StillImageSessionSettings.save(this, requested);
+                }
+                showSessionSummary(requested);
+                if (currentImage != null) {
+                    candidates = result == null ? null : result.candidates;
+                    selectedIndex = candidates != null && candidates.length > 0 ? 0 : -1;
+                    faceOverlay.showFaces(currentImage.getWidth(), currentImage.getHeight(),
+                            result == null ? null : result.faceRects(), selectedIndex);
+                    updateSelectedFaceCrop();
+                    finishDetectionUi(result);
+                }
+            });
+        });
+    }
+
+    private void showSessionSummary(StillImageSessionSettings.Values settings) {
+        sessionStatus.setText(getString(R.string.recognition_session_summary,
+                settings.inputPx, settings.maxFaces, settings.minFacePx));
+        sessionStatus.setTextColor(color(R.color.liveness_accent));
+    }
+
+    private void selectParameters(StillImageSessionSettings.Values settings) {
+        inputPxGroup.check(settings.inputPx == 320 ? R.id.inputPx320
+                : settings.inputPx == 1280 ? R.id.inputPx1280 : R.id.inputPx640);
+        if (settings.maxFaces == 1) {
+            maxFacesGroup.check(R.id.maxFaces1);
+        } else if (settings.maxFaces == 3) {
+            maxFacesGroup.check(R.id.maxFaces3);
+        } else if (settings.maxFaces == 5) {
+            maxFacesGroup.check(R.id.maxFaces5);
+        } else {
+            maxFacesGroup.check(R.id.maxFaces10);
+        }
+        if (settings.minFacePx == 48) {
+            minFaceGroup.check(R.id.minFace48);
+        } else if (settings.minFacePx == 64) {
+            minFaceGroup.check(R.id.minFace64);
+        } else if (settings.minFacePx == 128) {
+            minFaceGroup.check(R.id.minFace128);
+        } else {
+            minFaceGroup.check(R.id.minFace24);
+        }
+    }
+
+    private StillImageSessionSettings.Values selectedSettings() {
+        return new StillImageSessionSettings.Values(
+                selectedInputPx(), selectedMaxFaces(), selectedMinFacePx());
+    }
+
+    private void setSettingsExpanded(boolean expanded) {
+        sessionSettingsContent.setVisibility(expanded ? View.VISIBLE : View.GONE);
+        sessionSettingsToggleText.setText(expanded
+                ? R.string.recognition_settings_collapse
+                : R.string.recognition_settings_expand);
+    }
+
+    private void startVideoRecognition() {
+        if (!videoSelected || destroyed || videoCameraController != null || videoStarting) {
+            return;
+        }
+        if (!recognitionReady) {
+            showVideoError(R.string.recognition_video_initializing,
+                    R.string.recognition_video_first_face_hint);
+            return;
+        }
+        if (!cameraPermission.hasPermission()) {
+            showVideoError(R.string.msg_permission_required,
+                    R.string.capture_permission_hint);
+            cameraPermission.requestAccess();
+            return;
+        }
+
+        videoStarting = true;
+        showVideoError(R.string.recognition_video_initializing,
+                R.string.recognition_video_first_face_hint);
+        int generation = ++videoGeneration;
+        RecognitionFaceAnalyzer analyzer = new RecognitionFaceAnalyzer(
+                this, videoOverlay, repository, libraryEmpty,
+                new RecognitionFaceAnalyzer.Listener() {
+                    @Override
+                    public void onState(RecognitionFaceAnalyzer.State state,
+                                        @Nullable FaceRecord record,
+                                        float confidence, float threshold) {
+                        postVideoUi(generation,
+                                () -> renderVideoState(state, record, confidence, threshold));
+                    }
+
+                    @Override
+                    public void onSessionError() {
+                        postVideoUi(generation, () -> {
+                            videoStarting = false;
+                            showVideoError(R.string.compare_engine_failed,
+                                    R.string.recognition_video_first_face_hint);
+                        });
+                    }
+                });
+        videoAnalyzer = analyzer;
+        videoCameraController = new CameraPreviewController(
+                this, this, videoPreview, sdkExecutor, analyzer,
+                new CameraPreviewController.Listener() {
+                    @Override
+                    public void onCameraReady(boolean frontCamera) {
+                        if (!isCurrentVideo(generation)) {
+                            return;
+                        }
+                        videoStarting = false;
+                        analyzer.setMirrored(frontCamera);
+                        renderVideoState(RecognitionFaceAnalyzer.State.NO_FACE,
+                                null, Float.NaN, Float.NaN);
+                    }
+
+                    @Override
+                    public void onLensChanged(boolean frontCamera) {
+                        if (!isCurrentVideo(generation)) {
+                            return;
+                        }
+                        analyzer.setMirrored(frontCamera);
+                        analyzer.resetTracking();
+                        renderVideoState(RecognitionFaceAnalyzer.State.NO_FACE,
+                                null, Float.NaN, Float.NaN);
+                    }
+
+                    @Override
+                    public void onCameraError(int messageRes) {
+                        if (!isCurrentVideo(generation)) {
+                            return;
+                        }
+                        videoStarting = false;
+                        showVideoError(messageRes,
+                                R.string.recognition_video_first_face_hint);
+                    }
+                });
+        videoCameraController.start();
+    }
+
+    private void stopVideoRecognition() {
+        videoGeneration++;
+        videoStarting = false;
+        if (videoCameraController != null) {
+            videoCameraController.stop();
+            videoCameraController = null;
+        }
+        if (videoAnalyzer != null) {
+            RecognitionFaceAnalyzer toRelease = videoAnalyzer;
+            videoAnalyzer = null;
+            sdkExecutor.execute(toRelease::release);
+        }
+        if (videoOverlay != null) {
+            videoOverlay.submit(null);
+        }
+    }
+
+    private void flipVideoCamera() {
+        if (videoCameraController == null || !videoCameraController.flipCamera()) {
+            Toast.makeText(this, R.string.msg_camera_unavailable, Toast.LENGTH_SHORT).show();
+        }
+    }
+
+    private boolean isCurrentVideo(int generation) {
+        return !destroyed && videoSelected && generation == videoGeneration
+                && videoAnalyzer != null;
+    }
+
+    private void postVideoUi(int generation, Runnable action) {
+        runOnUiThread(() -> {
+            if (isCurrentVideo(generation)) {
+                action.run();
+            }
+        });
+    }
+
+    private void renderVideoState(RecognitionFaceAnalyzer.State state,
+                                  @Nullable FaceRecord record,
+                                  float confidence, float threshold) {
+        videoName.setVisibility(View.GONE);
+        switch (state) {
+            case MOVE_CLOSER:
+                showVideoError(R.string.capture_move_closer,
+                        R.string.recognition_video_first_face_hint);
+                videoStatus.setTextColor(color(R.color.liveness_warn));
+                break;
+            case HOLD_STILL:
+                showVideoError(R.string.recognition_video_hold_still,
+                        R.string.recognition_video_first_face_hint);
+                videoStatus.setTextColor(color(R.color.white));
+                break;
+            case SEARCHING:
+                showVideoError(R.string.recognition_searching,
+                        R.string.recognition_video_first_face_hint);
+                videoStatus.setTextColor(color(R.color.liveness_warn));
+                break;
+            case MATCHED:
+                videoStatus.setText(R.string.recognition_match_found);
+                videoStatus.setTextColor(color(R.color.liveness_accent));
+                videoName.setVisibility(View.VISIBLE);
+                videoName.setText(record == null
+                        ? getString(R.string.recognition_result_name_unknown) : record.name);
+                if (record != null) {
+                    videoDetails.setText(getString(R.string.recognition_result_details,
+                            record.id, confidence, threshold));
+                } else {
+                    videoDetails.setText(R.string.recognition_no_confidence);
+                }
+                break;
+            case NO_MATCH:
+                videoStatus.setText(R.string.recognition_no_match);
+                videoStatus.setTextColor(color(R.color.liveness_fail));
+                videoName.setVisibility(View.VISIBLE);
+                videoName.setText(R.string.recognition_result_name_unknown);
+                videoDetails.setText(Float.isNaN(confidence)
+                        ? getString(R.string.recognition_no_confidence)
+                        : getString(R.string.recognition_no_match_details,
+                        confidence, threshold));
+                break;
+            case EMPTY_LIBRARY:
+                videoStatus.setText(R.string.recognition_library_empty_result);
+                videoStatus.setTextColor(color(R.color.liveness_warn));
+                videoName.setVisibility(View.VISIBLE);
+                videoName.setText(model.sdkName());
+                videoDetails.setText(getString(
+                        R.string.recognition_empty_library, model.sdkName()));
+                break;
+            case NO_FACE:
+            default:
+                showVideoError(R.string.recognition_video_no_face,
+                        R.string.recognition_video_first_face_hint);
+                videoStatus.setTextColor(color(R.color.white));
+                break;
+        }
+    }
+
+    private void showVideoError(int titleRes, int detailsRes) {
+        videoStatus.setText(titleRes);
+        videoStatus.setTextColor(color(R.color.white));
+        videoName.setVisibility(View.GONE);
+        videoDetails.setText(detailsRes);
+    }
+
+    private int selectedInputPx() {
+        int id = inputPxGroup.getCheckedChipId();
+        return id == R.id.inputPx320 ? 320 : id == R.id.inputPx1280 ? 1280 : 640;
+    }
+
+    private int selectedMaxFaces() {
+        int id = maxFacesGroup.getCheckedChipId();
+        if (id == R.id.maxFaces1) return 1;
+        if (id == R.id.maxFaces3) return 3;
+        if (id == R.id.maxFaces5) return 5;
+        return 10;
+    }
+
+    private int selectedMinFacePx() {
+        int id = minFaceGroup.getCheckedChipId();
+        if (id == R.id.minFace48) return 48;
+        if (id == R.id.minFace64) return 64;
+        if (id == R.id.minFace128) return 128;
+        return 24;
+    }
+
+    private void setControlsEnabled(boolean enabled) {
+        choosePhotoButton.setEnabled(enabled);
+        applySessionButton.setEnabled(enabled);
+        resetSessionButton.setEnabled(enabled);
+        setChipGroupEnabled(inputPxGroup, enabled);
+        setChipGroupEnabled(maxFacesGroup, enabled);
+        setChipGroupEnabled(minFaceGroup, enabled);
+    }
+
+    private static void setChipGroupEnabled(ChipGroup group, boolean enabled) {
+        for (int i = 0; i < group.getChildCount(); i++) {
+            group.getChildAt(i).setEnabled(enabled);
+        }
+    }
+
+    private void setImageStatus(int textRes, @ColorRes int colorRes) {
+        imageStatus.setText(textRes);
+        imageStatus.setTextColor(color(colorRes));
+    }
+
+    private int color(@ColorRes int colorRes) {
+        return ContextCompat.getColor(this, colorRes);
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.recognitionRoot);
+        int left = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int right = root.getPaddingRight();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(left + bars.left, top + bars.top,
+                    right + bars.right, bottom + bars.bottom);
+            return insets;
+        });
+    }
+
+    private void postUi(Runnable action) {
+        runOnUiThread(() -> {
+            if (!destroyed && !isFinishing()) {
+                action.run();
+            }
+        });
+    }
+
+    private static void recycle(@Nullable Bitmap bitmap) {
+        if (bitmap != null && !bitmap.isRecycled()) {
+            bitmap.recycle();
+        }
+    }
+
+    @Override
+    protected void onResume() {
+        super.onResume();
+        cameraPermission.onResume();
+    }
+
+    @Override
+    protected void onDestroy() {
+        cameraPermission.close();
+        destroyed = true;
+        recognitionReady = false;
+        imageVersion++;
+        stopVideoRecognition();
+        Bitmap imageToRecycle = imageBitmap;
+        Bitmap cropToRecycle = resultCropBitmap;
+        Bitmap selectedCropToRecycle = selectedFaceCropBitmap;
+        imageBitmap = null;
+        resultCropBitmap = null;
+        selectedFaceCropBitmap = null;
+        sdkExecutor.execute(() -> {
+            recycle(imageToRecycle);
+            recycle(cropToRecycle);
+            recycle(selectedCropToRecycle);
+            repository.close();
+            if (session != null) {
+                FaceEngine.releaseSession(session);
+                session = null;
+            }
+        });
+        sdkExecutor.shutdown();
+        super.onDestroy();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceTrackingSessionSettings.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceTrackingSessionSettings.java
new file mode 100644
index 000000000..adebf6a6b
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/FaceTrackingSessionSettings.java
@@ -0,0 +1,80 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+import android.content.SharedPreferences;
+
+import com.insightface.sdk.inspireface.InspireFace;
+
+/** Locally cached configuration for the continuous face-tracking camera Session. */
+final class FaceTrackingSessionSettings {
+
+    static final int DEFAULT_MODE = InspireFace.DETECT_MODE_LIGHT_TRACK;
+    static final int DEFAULT_INPUT_PX = DetectorDefaults.INPUT_PX;
+    static final int DEFAULT_MAX_FACES = 10;
+    static final int DEFAULT_MIN_FACE_PX = 24;
+
+    static final class Values {
+        final int mode;
+        final int inputPx;
+        final int maxFaces;
+        final int minFacePx;
+
+        Values(int mode, int inputPx, int maxFaces, int minFacePx) {
+            this.mode = validMode(mode);
+            this.inputPx = validInputPx(inputPx);
+            this.maxFaces = validMaxFaces(maxFaces);
+            this.minFacePx = validMinFacePx(minFacePx);
+        }
+    }
+
+    private static final String PREFS = "face_tracking_session";
+    private static final String KEY_MODE = "detect_mode";
+    private static final String KEY_INPUT_PX = "input_px";
+    private static final String KEY_MAX_FACES = "max_faces";
+    private static final String KEY_MIN_FACE_PX = "min_face_px";
+
+    private FaceTrackingSessionSettings() {
+    }
+
+    static Values defaults() {
+        return new Values(DEFAULT_MODE, DEFAULT_INPUT_PX,
+                DEFAULT_MAX_FACES, DEFAULT_MIN_FACE_PX);
+    }
+
+    static Values load(Context context) {
+        SharedPreferences preferences = context.getSharedPreferences(PREFS, Context.MODE_PRIVATE);
+        return new Values(
+                preferences.getInt(KEY_MODE, DEFAULT_MODE),
+                preferences.getInt(KEY_INPUT_PX, DEFAULT_INPUT_PX),
+                preferences.getInt(KEY_MAX_FACES, DEFAULT_MAX_FACES),
+                preferences.getInt(KEY_MIN_FACE_PX, DEFAULT_MIN_FACE_PX));
+    }
+
+    static void save(Context context, Values values) {
+        context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .edit()
+                .putInt(KEY_MODE, values.mode)
+                .putInt(KEY_INPUT_PX, values.inputPx)
+                .putInt(KEY_MAX_FACES, values.maxFaces)
+                .putInt(KEY_MIN_FACE_PX, values.minFacePx)
+                .apply();
+    }
+
+    private static int validMode(int value) {
+        return value == InspireFace.DETECT_MODE_TRACK_BY_DETECTION
+                ? value : DEFAULT_MODE;
+    }
+
+    private static int validInputPx(int value) {
+        return value == 320 || value == 1280 ? value : DEFAULT_INPUT_PX;
+    }
+
+    private static int validMaxFaces(int value) {
+        return value == 1 || value == 3 || value == 5 ? value : DEFAULT_MAX_FACES;
+    }
+
+    private static int validMinFacePx(int value) {
+        return value == 48 || value == 64 || value == 128
+                ? value : DEFAULT_MIN_FACE_PX;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/HomeActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/HomeActivity.java
new file mode 100644
index 000000000..b33b7d1c7
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/HomeActivity.java
@@ -0,0 +1,76 @@
+package com.example.inspireface_example;
+
+import android.content.Intent;
+import android.os.Bundle;
+import android.view.View;
+
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.view.ActionLivenessActivity;
+import com.example.inspireface_example.view.LivenessActivity;
+import com.example.inspireface_example.view.PoseActivity;
+import com.google.android.material.button.MaterialButtonToggleGroup;
+
+/** Launcher page: model selection plus a square-grid menu for the available demos. */
+public class HomeActivity extends AppCompatActivity {
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        WindowCompat.getInsetsController(getWindow(), getWindow().getDecorView())
+                .setAppearanceLightNavigationBars(false);
+        setContentView(R.layout.activity_home);
+        applyWindowInsets();
+
+        MaterialButtonToggleGroup modelToggle = findViewById(R.id.modelToggle);
+        FaceModelPrefs.Model selected = FaceModelPrefs.get(this);
+        modelToggle.check(selected == FaceModelPrefs.Model.PIKACHU
+                ? R.id.btnModelPikachu : R.id.btnModelMegatron);
+        modelToggle.addOnButtonCheckedListener((group, checkedId, isChecked) -> {
+            if (!isChecked) {
+                return;
+            }
+            FaceModelPrefs.set(this, checkedId == R.id.btnModelPikachu
+                    ? FaceModelPrefs.Model.PIKACHU : FaceModelPrefs.Model.MEGATRON);
+        });
+
+        findViewById(R.id.cardSilent).setOnClickListener(
+                v -> openFeature(LivenessActivity.class));
+        findViewById(R.id.cardAction).setOnClickListener(
+                v -> openFeature(ActionLivenessActivity.class));
+        findViewById(R.id.cardPose).setOnClickListener(
+                v -> openFeature(PoseActivity.class));
+        findViewById(R.id.cardCompare).setOnClickListener(
+                v -> openFeature(FaceCompareActivity.class));
+        findViewById(R.id.cardManagement).setOnClickListener(
+                v -> openFeature(FaceManagementActivity.class));
+        findViewById(R.id.cardRecognition).setOnClickListener(
+                v -> openFeature(FaceRecognitionActivity.class));
+        findViewById(R.id.cardDetection).setOnClickListener(
+                v -> openFeature(FaceDetectionActivity.class));
+        findViewById(R.id.cardAttribute).setOnClickListener(
+                v -> openFeature(FaceAttributeActivity.class));
+        findViewById(R.id.langSwitch).setOnClickListener(v -> LocalePrefs.toggle(this));
+    }
+
+    private void openFeature(Class<?> activityClass) {
+        startActivity(new Intent(this, activityClass));
+    }
+
+    private void applyWindowInsets() {
+        View root = findViewById(R.id.homeRoot);
+        int horizontal = root.getPaddingLeft();
+        int top = root.getPaddingTop();
+        int bottom = root.getPaddingBottom();
+        ViewCompat.setOnApplyWindowInsetsListener(root, (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(horizontal, top + bars.top, horizontal, bottom + bars.bottom);
+            return insets;
+        });
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/LocalePrefs.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/LocalePrefs.java
new file mode 100644
index 000000000..8036be14f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/LocalePrefs.java
@@ -0,0 +1,72 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+import android.os.Build;
+
+import androidx.appcompat.app.AppCompatDelegate;
+import androidx.core.os.LocaleListCompat;
+
+import java.util.Locale;
+
+/**
+ * Per-app language selection. The demo defaults to English regardless of the system
+ * language; the in-app language controls switch to Chinese and back.
+ * The choice is stored in SharedPreferences and re-applied on every app start.
+ */
+public final class LocalePrefs {
+
+    private static final String PREFS = "settings";
+    private static final String KEY_LOCALE = "app_locale";
+    private static final String ENGLISH = "en";
+    private static final String CHINESE = "zh";
+
+    private LocalePrefs() {
+    }
+
+    /** Applies the stored language (English by default). Call from Application.onCreate. */
+    public static void applyStored(Context context) {
+        String selected = stored(context);
+        if (Build.VERSION.SDK_INT >= 33) {
+            String platformSelection = supportedLanguage(
+                    AppCompatDelegate.getApplicationLocales());
+            if (platformSelection != null) {
+                // Respect a language selected from Android 13+'s per-app language screen.
+                selected = platformSelection;
+            }
+        }
+        context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .edit().putString(KEY_LOCALE, selected).apply();
+        AppCompatDelegate.setApplicationLocales(
+                LocaleListCompat.forLanguageTags(selected));
+    }
+
+    /** Switches between English and Chinese; running activities recreate automatically. */
+    public static void toggle(Context context) {
+        String active = supportedLanguage(AppCompatDelegate.getApplicationLocales());
+        String next = CHINESE.equals(active != null ? active : stored(context))
+                ? ENGLISH : CHINESE;
+        context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .edit().putString(KEY_LOCALE, next).apply();
+        AppCompatDelegate.setApplicationLocales(LocaleListCompat.forLanguageTags(next));
+    }
+
+    private static String stored(Context context) {
+        String value = context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .getString(KEY_LOCALE, ENGLISH);
+        return CHINESE.equals(value) ? CHINESE : ENGLISH;
+    }
+
+    private static String supportedLanguage(LocaleListCompat locales) {
+        if (locales == null || locales.isEmpty()) {
+            return null;
+        }
+        Locale locale = locales.get(0);
+        if (locale == null) {
+            return null;
+        }
+        if (CHINESE.equals(locale.getLanguage())) {
+            return CHINESE;
+        }
+        return ENGLISH.equals(locale.getLanguage()) ? ENGLISH : null;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/StillImageSessionSettings.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/StillImageSessionSettings.java
new file mode 100644
index 000000000..016918ce8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/StillImageSessionSettings.java
@@ -0,0 +1,67 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+import android.content.SharedPreferences;
+
+/** Cached still-image detection settings shared by photo-based SDK demos. */
+final class StillImageSessionSettings {
+
+    static final int DEFAULT_INPUT_PX = DetectorDefaults.INPUT_PX;
+    static final int DEFAULT_MAX_FACES = 10;
+    static final int DEFAULT_MIN_FACE_PX = 24;
+
+    static final class Values {
+        final int inputPx;
+        final int maxFaces;
+        final int minFacePx;
+
+        Values(int inputPx, int maxFaces, int minFacePx) {
+            this.inputPx = validInputPx(inputPx);
+            this.maxFaces = validMaxFaces(maxFaces);
+            this.minFacePx = validMinFacePx(minFacePx);
+        }
+    }
+
+    // Keep the original preference name so existing recognition settings migrate intact.
+    private static final String PREFS = "face_recognition_session";
+    private static final String KEY_INPUT_PX = "input_px";
+    private static final String KEY_MAX_FACES = "max_faces";
+    private static final String KEY_MIN_FACE_PX = "min_face_px";
+
+    private StillImageSessionSettings() {
+    }
+
+    static Values defaults() {
+        return new Values(DEFAULT_INPUT_PX, DEFAULT_MAX_FACES, DEFAULT_MIN_FACE_PX);
+    }
+
+    static Values load(Context context) {
+        SharedPreferences preferences = context.getSharedPreferences(PREFS, Context.MODE_PRIVATE);
+        return new Values(
+                preferences.getInt(KEY_INPUT_PX, DEFAULT_INPUT_PX),
+                preferences.getInt(KEY_MAX_FACES, DEFAULT_MAX_FACES),
+                preferences.getInt(KEY_MIN_FACE_PX, DEFAULT_MIN_FACE_PX));
+    }
+
+    static void save(Context context, Values values) {
+        context.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+                .edit()
+                .putInt(KEY_INPUT_PX, values.inputPx)
+                .putInt(KEY_MAX_FACES, values.maxFaces)
+                .putInt(KEY_MIN_FACE_PX, values.minFacePx)
+                .apply();
+    }
+
+    private static int validInputPx(int value) {
+        return value == 320 || value == 1280 ? value : DEFAULT_INPUT_PX;
+    }
+
+    private static int validMaxFaces(int value) {
+        return value == 1 || value == 3 || value == 5 ? value : DEFAULT_MAX_FACES;
+    }
+
+    private static int validMinFacePx(int value) {
+        return value == 48 || value == 64 || value == 128
+                ? value : DEFAULT_MIN_FACE_PX;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceAttributeProcessor.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceAttributeProcessor.java
new file mode 100644
index 000000000..8dd4a6464
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceAttributeProcessor.java
@@ -0,0 +1,162 @@
+package com.example.inspireface_example.face;
+
+import android.graphics.Bitmap;
+import android.graphics.RectF;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.CustomParameter;
+import com.insightface.sdk.inspireface.base.FaceAttributeResult;
+import com.insightface.sdk.inspireface.base.FaceInteractionState;
+import com.insightface.sdk.inspireface.base.FaceInteractionsActions;
+import com.insightface.sdk.inspireface.base.FaceMaskConfidence;
+import com.insightface.sdk.inspireface.base.FaceQualityConfidence;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Session;
+
+/** Runs still-image face detection and the SDK attribute pipeline in one stream lifetime. */
+public final class FaceAttributeProcessor {
+
+    private static final int INTERACTION_PIPELINE_CALLS = 10;
+
+    public enum Status { READY, NO_FACE, PROCESS_FAILED }
+
+    public static final class Attribute {
+        public final float maskConfidence;
+        public final int ageBracket;
+        public final float qualityScore;
+        public final int jawOpen;
+        public final int race;
+        public final int gender;
+        public final float leftEyeConfidence;
+        public final float rightEyeConfidence;
+
+        Attribute(float maskConfidence, int ageBracket, float qualityScore,
+                  int jawOpen, int race, int gender,
+                  float leftEyeConfidence, float rightEyeConfidence) {
+            this.maskConfidence = maskConfidence;
+            this.ageBracket = ageBracket;
+            this.qualityScore = qualityScore;
+            this.jawOpen = jawOpen;
+            this.race = race;
+            this.gender = gender;
+            this.leftEyeConfidence = leftEyeConfidence;
+            this.rightEyeConfidence = rightEyeConfidence;
+        }
+    }
+
+    public static final class Result {
+        public final Status status;
+        public final FaceImageProcessor.Candidate[] candidates;
+        public final Attribute[] attributes;
+
+        Result(Status status, FaceImageProcessor.Candidate[] candidates,
+               Attribute[] attributes) {
+            this.status = status;
+            this.candidates = candidates;
+            this.attributes = attributes;
+        }
+
+        public RectF[] faceRects() {
+            RectF[] rects = new RectF[candidates.length];
+            for (int i = 0; i < candidates.length; i++) {
+                rects[i] = new RectF(candidates[i].rect);
+            }
+            return rects;
+        }
+    }
+
+    private FaceAttributeProcessor() {
+    }
+
+    public static Result analyze(Session session, Bitmap bitmap) {
+        ImageStream stream = InspireFace.CreateImageStreamFromBitmap(
+                bitmap, InspireFace.CAMERA_ROTATION_0);
+        if (stream == null) {
+            return empty(Status.PROCESS_FAILED);
+        }
+        try {
+            MultipleFaceData faces = InspireFace.ExecuteFaceTrack(session, stream);
+            if (faces == null) {
+                return empty(Status.PROCESS_FAILED);
+            }
+            if (faces.detectedNum <= 0) {
+                return empty(Status.NO_FACE);
+            }
+
+            FaceImageProcessor.Candidate[] candidates = candidates(faces);
+            CustomParameter pipeline = InspireFace.CreateCustomParameter()
+                    .enableMaskDetect(true)
+                    .enableFaceQuality(true)
+                    .enableFaceAttribute(true)
+                    .enableInteractionLiveness(true);
+            if (!InspireFace.MultipleFacePipelineProcess(
+                    session, stream, faces, pipeline)) {
+                return new Result(Status.PROCESS_FAILED, candidates, new Attribute[0]);
+            }
+
+            FaceMaskConfidence masks = InspireFace.GetFaceMaskConfidence(session);
+            FaceQualityConfidence qualities = InspireFace.GetFaceQualityConfidence(session);
+            FaceAttributeResult faceAttributes = InspireFace.GetFaceAttributeResult(session);
+
+            // The interaction module has a short fresh-track warm-up. Reusing the same
+            // still frame lets it produce deterministic eye and jaw state without changing
+            // the already cached mask, quality and demographic results above.
+            CustomParameter interaction = InspireFace.CreateCustomParameter()
+                    .enableInteractionLiveness(true);
+            for (int i = 1; i < INTERACTION_PIPELINE_CALLS; i++) {
+                InspireFace.MultipleFacePipelineProcess(
+                        session, stream, faces, interaction);
+            }
+            FaceInteractionState eyeStates =
+                    InspireFace.GetFaceInteractionStateResult(session);
+            FaceInteractionsActions actions =
+                    InspireFace.GetFaceInteractionActionsResult(session);
+
+            Attribute[] attributes = new Attribute[candidates.length];
+            for (int i = 0; i < attributes.length; i++) {
+                attributes[i] = new Attribute(
+                        floatAt(masks == null ? null : masks.confidence, i),
+                        intAt(faceAttributes == null ? null : faceAttributes.ageBracket, i),
+                        floatAt(qualities == null ? null : qualities.confidence, i),
+                        intAt(actions == null ? null : actions.jawOpen, i),
+                        intAt(faceAttributes == null ? null : faceAttributes.race, i),
+                        intAt(faceAttributes == null ? null : faceAttributes.gender, i),
+                        floatAt(eyeStates == null ? null
+                                : eyeStates.leftEyeStatusConfidence, i),
+                        floatAt(eyeStates == null ? null
+                                : eyeStates.rightEyeStatusConfidence, i));
+            }
+            return new Result(Status.READY, candidates, attributes);
+        } finally {
+            InspireFace.ReleaseImageStream(stream);
+        }
+    }
+
+    private static FaceImageProcessor.Candidate[] candidates(MultipleFaceData faces) {
+        FaceImageProcessor.Candidate[] candidates =
+                new FaceImageProcessor.Candidate[faces.detectedNum];
+        for (int i = 0; i < faces.detectedNum; i++) {
+            FaceRect face = faces.rects[i];
+            RectF rect = new RectF(face.x, face.y,
+                    face.x + face.width, face.y + face.height);
+            candidates[i] = new FaceImageProcessor.Candidate(
+                    rect, null, null, null);
+        }
+        return candidates;
+    }
+
+    private static Result empty(Status status) {
+        return new Result(status, new FaceImageProcessor.Candidate[0], new Attribute[0]);
+    }
+
+    private static float floatAt(float[] values, int index) {
+        return values != null && index >= 0 && index < values.length
+                ? values[index] : Float.NaN;
+    }
+
+    private static int intAt(int[] values, int index) {
+        return values != null && index >= 0 && index < values.length ? values[index] : -1;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceCropUtils.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceCropUtils.java
new file mode 100644
index 000000000..c5c4aaf20
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceCropUtils.java
@@ -0,0 +1,48 @@
+package com.example.inspireface_example.face;
+
+import android.graphics.Bitmap;
+import android.graphics.RectF;
+
+import androidx.annotation.Nullable;
+
+/** Shared expanded square crop used by selectable-face magnifiers. */
+public final class FaceCropUtils {
+
+    public static final class SquareCrop {
+        public final Bitmap bitmap;
+        public final int left;
+        public final int top;
+        public final int size;
+
+        SquareCrop(Bitmap bitmap, int left, int top, int size) {
+            this.bitmap = bitmap;
+            this.left = left;
+            this.top = top;
+            this.size = size;
+        }
+    }
+
+    private FaceCropUtils() {
+    }
+
+    @Nullable
+    public static SquareCrop createSquare(Bitmap source, RectF face, float scale) {
+        if (source == null || source.isRecycled() || face == null
+                || source.getWidth() < 2 || source.getHeight() < 2) {
+            return null;
+        }
+        int cropSize = Math.max(2, Math.round(
+                Math.max(face.width(), face.height()) * scale));
+        cropSize = Math.min(cropSize, Math.min(source.getWidth(), source.getHeight()));
+        int left = Math.round(face.centerX() - cropSize / 2f);
+        int top = Math.round(face.centerY() - cropSize / 2f);
+        left = Math.max(0, Math.min(left, source.getWidth() - cropSize));
+        top = Math.max(0, Math.min(top, source.getHeight() - cropSize));
+        try {
+            Bitmap crop = Bitmap.createBitmap(source, left, top, cropSize, cropSize);
+            return crop == source ? null : new SquareCrop(crop, left, top, cropSize);
+        } catch (IllegalArgumentException ignored) {
+            return null;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceImageProcessor.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceImageProcessor.java
new file mode 100644
index 000000000..4fd91c3f0
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceImageProcessor.java
@@ -0,0 +1,142 @@
+package com.example.inspireface_example.face;
+
+import android.graphics.Bitmap;
+import android.graphics.RectF;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Point2f;
+import com.insightface.sdk.inspireface.base.Session;
+
+/** Detects all faces once and materializes optional features, crops, or dense landmarks. */
+public final class FaceImageProcessor {
+
+    public enum Status { READY, NO_FACE, PROCESS_FAILED }
+
+    public static final class Candidate {
+        public final RectF rect;
+        public final FaceFeature feature;
+        public final Bitmap crop;
+        public final Point2f[] denseLandmarks;
+
+        Candidate(RectF rect, FaceFeature feature, Bitmap crop,
+                  Point2f[] denseLandmarks) {
+            this.rect = rect;
+            this.feature = feature;
+            this.crop = crop;
+            this.denseLandmarks = denseLandmarks;
+        }
+    }
+
+    public static final class Result {
+        public final Status status;
+        public final Candidate[] candidates;
+
+        Result(Status status, Candidate[] candidates) {
+            this.status = status;
+            this.candidates = candidates;
+        }
+
+        public RectF[] faceRects() {
+            RectF[] rects = new RectF[candidates.length];
+            for (int i = 0; i < candidates.length; i++) {
+                rects[i] = new RectF(candidates[i].rect);
+            }
+            return rects;
+        }
+
+    }
+
+    private FaceImageProcessor() {
+    }
+
+    /** All calls must remain on the thread that owns {@code session}. */
+    public static Result detect(Session session, Bitmap bitmap, boolean createCrops) {
+        return detect(session, bitmap, createCrops, true, false);
+    }
+
+    /** Detection-only variant that materializes the SDK's native 106-point landmarks. */
+    public static Result detectWithLandmarks(Session session, Bitmap bitmap) {
+        return detect(session, bitmap, false, false, true);
+    }
+
+    private static Result detect(Session session, Bitmap bitmap, boolean createCrops,
+                                 boolean extractFeatures, boolean extractLandmarks) {
+        ImageStream stream = InspireFace.CreateImageStreamFromBitmap(
+                bitmap, InspireFace.CAMERA_ROTATION_0);
+        if (stream == null) {
+            return new Result(Status.PROCESS_FAILED, new Candidate[0]);
+        }
+        try {
+            MultipleFaceData faces = InspireFace.ExecuteFaceTrack(session, stream);
+            if (faces == null) {
+                return new Result(Status.PROCESS_FAILED, new Candidate[0]);
+            }
+            if (faces.detectedNum == 0) {
+                return new Result(Status.NO_FACE, new Candidate[0]);
+            }
+            Candidate[] candidates = new Candidate[faces.detectedNum];
+            for (int i = 0; i < faces.detectedNum; i++) {
+                FaceRect face = faces.rects[i];
+                RectF rect = new RectF(face.x, face.y,
+                        face.x + face.width, face.y + face.height);
+                FaceFeature feature = extractFeatures
+                        ? InspireFace.ExtractFaceFeature(session, stream, faces.tokens[i])
+                        : null;
+                Point2f[] denseLandmarks = extractLandmarks
+                        ? InspireFace.GetFaceDenseLandmarkFromFaceToken(faces.tokens[i])
+                        : null;
+                Bitmap crop = null;
+                if (createCrops) {
+                    crop = InspireFace.GetFaceAlignmentImage(session, stream, faces.tokens[i]);
+                    if (crop == null) {
+                        crop = cropFace(bitmap, rect);
+                    } else if (crop == bitmap) {
+                        // Keep crop ownership independent from the source image. The editor
+                        // recycles source and crop bitmaps on separate lifecycle paths.
+                        crop = independentCopy(bitmap);
+                    }
+                }
+                candidates[i] = new Candidate(rect, feature, crop, denseLandmarks);
+            }
+            return new Result(Status.READY, candidates);
+        } finally {
+            InspireFace.ReleaseImageStream(stream);
+        }
+    }
+
+    public static void recycleCrops(Candidate[] candidates, int exceptIndex) {
+        if (candidates == null) {
+            return;
+        }
+        Bitmap keptCrop = exceptIndex >= 0 && exceptIndex < candidates.length
+                ? candidates[exceptIndex].crop : null;
+        for (int i = 0; i < candidates.length; i++) {
+            Bitmap crop = candidates[i].crop;
+            if (i != exceptIndex && crop != null && crop != keptCrop && !crop.isRecycled()) {
+                crop.recycle();
+            }
+        }
+    }
+
+    private static Bitmap cropFace(Bitmap bitmap, RectF face) {
+        float paddingX = face.width() * 0.18f;
+        float paddingY = face.height() * 0.18f;
+        int left = Math.max(0, Math.round(face.left - paddingX));
+        int top = Math.max(0, Math.round(face.top - paddingY));
+        int right = Math.min(bitmap.getWidth(), Math.round(face.right + paddingX));
+        int bottom = Math.min(bitmap.getHeight(), Math.round(face.bottom + paddingY));
+        if (right <= left || bottom <= top) {
+            return null;
+        }
+        Bitmap crop = Bitmap.createBitmap(bitmap, left, top, right - left, bottom - top);
+        return crop == bitmap ? independentCopy(bitmap) : crop;
+    }
+
+    private static Bitmap independentCopy(Bitmap source) {
+        return source.copy(Bitmap.Config.ARGB_8888, false);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRecord.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRecord.java
new file mode 100644
index 000000000..e1653c4f5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRecord.java
@@ -0,0 +1,16 @@
+package com.example.inspireface_example.face;
+
+/** Local metadata paired with one FeatureHub identity. */
+public final class FaceRecord {
+    public final long id;
+    public final String name;
+    public final String cropPath;
+    public final long updatedAt;
+
+    public FaceRecord(long id, String name, String cropPath, long updatedAt) {
+        this.id = id;
+        this.name = name;
+        this.cropPath = cropPath;
+        this.updatedAt = updatedAt;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRepository.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRepository.java
new file mode 100644
index 000000000..94eef7af2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/FaceRepository.java
@@ -0,0 +1,435 @@
+package com.example.inspireface_example.face;
+
+import android.content.Context;
+import android.content.SharedPreferences;
+import android.graphics.Bitmap;
+
+import androidx.annotation.Nullable;
+
+import com.example.inspireface_example.FaceModelPrefs;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.FaceFeatureIdentity;
+import com.insightface.sdk.inspireface.base.FeatureHubConfiguration;
+import com.insightface.sdk.inspireface.base.SearchTopKResults;
+
+import org.json.JSONException;
+import org.json.JSONObject;
+
+import java.io.File;
+import java.io.FileOutputStream;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.Locale;
+import java.util.Map;
+
+/**
+ * Model-scoped FeatureHub storage. Each model gets its own native DB, crop directory,
+ * metadata preferences and ID sequence; no feature or image is shared across models.
+ */
+public final class FaceRepository {
+
+    public static final class SearchResult {
+        public final boolean matched;
+        public final FaceRecord record;
+        public final float confidence;
+        public final float threshold;
+
+        SearchResult(boolean matched, @Nullable FaceRecord record,
+                     float confidence, float threshold) {
+            this.matched = matched;
+            this.record = record;
+            this.confidence = confidence;
+            this.threshold = threshold;
+        }
+    }
+
+    public static final class InsertResult {
+        public final boolean success;
+        public final FaceRecord record;
+
+        InsertResult(boolean success, @Nullable FaceRecord record) {
+            this.success = success;
+            this.record = record;
+        }
+    }
+
+    private static final String KEY_PREFIX = "record.";
+    private static final String KEY_NEXT_ID = "next_id";
+    private static final Object HUB_LOCK = new Object();
+    private static String activeDatabasePath;
+    private static int hubReferences;
+
+    private final SharedPreferences metadata;
+    private final File modelDirectory;
+    private final File cropDirectory;
+    private final File databaseFile;
+    private boolean hubAcquired;
+
+    public FaceRepository(Context context, FaceModelPrefs.Model model) {
+        Context app = context.getApplicationContext();
+        modelDirectory = new File(new File(app.getFilesDir(), "face_hub"), model.sdkName());
+        cropDirectory = new File(modelDirectory, "crops");
+        databaseFile = new File(modelDirectory, "features.db");
+        metadata = app.getSharedPreferences(
+                "face_records_" + model.sdkName(), Context.MODE_PRIVATE);
+    }
+
+    /** Must be called after GlobalLaunch and from the repository's SDK executor. */
+    public boolean open() {
+        synchronized (HUB_LOCK) {
+            if (hubAcquired) {
+                return true;
+            }
+            String requestedPath = databaseFile.getAbsolutePath();
+            if (hubReferences > 0) {
+                if (!requestedPath.equals(activeDatabasePath)) {
+                    return false;
+                }
+                hubReferences++;
+                hubAcquired = true;
+                return true;
+            }
+            if (!cropDirectory.exists() && !cropDirectory.mkdirs()) {
+                return false;
+            }
+            FeatureHubConfiguration configuration = InspireFace.CreateFeatureHubConfiguration()
+                    .setPrimaryKeyMode(InspireFace.PK_MANUAL_INPUT)
+                    .setEnablePersistence(true)
+                    .setPersistenceDbPath(requestedPath)
+                    .setSearchThreshold(InspireFace.GetRecommendedCosineThreshold())
+                    .setSearchMode(InspireFace.SEARCH_MODE_EXHAUSTIVE);
+            if (!InspireFace.FeatureHubDataEnable(configuration)) {
+                return false;
+            }
+            activeDatabasePath = requestedPath;
+            hubReferences = 1;
+            hubAcquired = true;
+            return true;
+        }
+    }
+
+    public void close() {
+        synchronized (HUB_LOCK) {
+            if (!hubAcquired) {
+                return;
+            }
+            hubAcquired = false;
+            hubReferences = Math.max(0, hubReferences - 1);
+            if (hubReferences == 0) {
+                InspireFace.FeatureHubDataDisable();
+                activeDatabasePath = null;
+            }
+        }
+    }
+
+    public List<FaceRecord> query(@Nullable String keyword) {
+        String normalized = keyword == null ? ""
+                : keyword.trim().toLowerCase(Locale.ROOT);
+        List<FaceRecord> records = new ArrayList<>();
+        for (Map.Entry<String, ?> entry : metadata.getAll().entrySet()) {
+            if (!entry.getKey().startsWith(KEY_PREFIX) || !(entry.getValue() instanceof String)) {
+                continue;
+            }
+            FaceRecord record = decode((String) entry.getValue());
+            if (record != null && (normalized.isEmpty()
+                    || record.name.toLowerCase(Locale.ROOT).contains(normalized)
+                    || String.valueOf(record.id).contains(normalized))) {
+                records.add(record);
+            }
+        }
+        Collections.sort(records,
+                (left, right) -> Long.compare(right.updatedAt, left.updatedAt));
+        return records;
+    }
+
+    /** Searches the active model's native FeatureHub and joins the best ID to metadata. */
+    public SearchResult search(FaceFeature feature) {
+        float threshold = InspireFace.GetRecommendedCosineThreshold();
+        if (!hubAcquired || feature == null) {
+            return new SearchResult(false, null, Float.NaN, threshold);
+        }
+        SearchTopKResults results = InspireFace.FeatureHubFaceSearchTopK(feature, 1);
+        if (results == null || results.num <= 0 || results.ids == null
+                || results.confidence == null || results.ids.length == 0
+                || results.confidence.length == 0) {
+            return new SearchResult(false, null, Float.NaN, threshold);
+        }
+        float confidence = results.confidence[0];
+        FaceRecord record = get(results.ids[0]);
+        return new SearchResult(record != null && confidence >= threshold,
+                record, confidence, threshold);
+    }
+
+    @Nullable
+    public FaceRecord get(long id) {
+        return decode(metadata.getString(key(id), null));
+    }
+
+    public InsertResult insert(String name, FaceFeature feature, Bitmap crop) {
+        if (!hubAcquired || feature == null || crop == null) {
+            return new InsertResult(false, null);
+        }
+        long id = Math.max(1L, metadata.getLong(KEY_NEXT_ID, 1L));
+        while (metadata.contains(key(id))) {
+            id++;
+        }
+        File cropFile = cropFile(id);
+        File stagedCrop = stageCrop(cropFile, crop);
+        if (stagedCrop == null) {
+            return new InsertResult(false, null);
+        }
+        File cropBackup = backupFile(cropFile);
+        boolean hadPreviousCrop = cropFile.exists();
+        FaceFeatureIdentity identity = FaceFeatureIdentity.create(id, feature);
+        if (!InspireFace.FeatureHubInsertFeature(identity)) {
+            stagedCrop.delete();
+            return new InsertResult(false, null);
+        }
+        if (!commitStagedCrop(cropFile, stagedCrop, cropBackup)) {
+            InspireFace.FeatureHubFaceRemove(id);
+            return new InsertResult(false, null);
+        }
+        FaceRecord record = new FaceRecord(
+                id, normalizedName(name, id), cropFile.getAbsolutePath(),
+                System.currentTimeMillis());
+        boolean saved = metadata.edit()
+                .putString(key(id), encode(record))
+                .putLong(KEY_NEXT_ID, id + 1)
+                .commit();
+        if (!saved) {
+            InspireFace.FeatureHubFaceRemove(id);
+            restoreCrop(cropFile, cropBackup, hadPreviousCrop);
+            return new InsertResult(false, null);
+        }
+        cropBackup.delete();
+        return new InsertResult(true, record);
+    }
+
+    /** Passing null feature/crop performs a metadata-only rename. */
+    public boolean update(long id, String name,
+                          @Nullable FaceFeature feature, @Nullable Bitmap crop) {
+        FaceRecord old = get(id);
+        if (!hubAcquired || old == null) {
+            return false;
+        }
+        File cropFile = new File(old.cropPath);
+        File stagedCrop = crop == null ? null : stageCrop(cropFile, crop);
+        if (crop != null && stagedCrop == null) {
+            return false;
+        }
+        File cropBackup = backupFile(cropFile);
+        boolean hadPreviousCrop = cropFile.exists();
+
+        FaceFeatureIdentity previousIdentity = null;
+        if (feature != null) {
+            previousIdentity = InspireFace.FeatureHubGetFaceIdentity(id);
+            if (previousIdentity == null || !InspireFace.FeatureHubFaceUpdate(
+                    FaceFeatureIdentity.create(id, feature))) {
+                if (stagedCrop != null) {
+                    stagedCrop.delete();
+                }
+                return false;
+            }
+        }
+        if (stagedCrop != null
+                && !commitStagedCrop(cropFile, stagedCrop, cropBackup)) {
+            rollbackFeature(previousIdentity);
+            return false;
+        }
+        FaceRecord updated = new FaceRecord(id, normalizedName(name, id),
+                cropFile.getAbsolutePath(), System.currentTimeMillis());
+        boolean saved = metadata.edit().putString(key(id), encode(updated)).commit();
+        if (!saved) {
+            if (stagedCrop != null) {
+                restoreCrop(cropFile, cropBackup, hadPreviousCrop);
+            }
+            rollbackFeature(previousIdentity);
+            return false;
+        }
+        cropBackup.delete();
+        return true;
+    }
+
+    public boolean delete(long id) {
+        FaceRecord record = get(id);
+        if (!hubAcquired || record == null) {
+            return false;
+        }
+        File crop = new File(record.cropPath);
+        File pendingDelete = new File(crop.getParentFile(), crop.getName() + ".delete");
+        if (!recoverPendingDelete(crop, pendingDelete)) {
+            return false;
+        }
+        boolean hadCrop = crop.exists();
+        if (hadCrop && !crop.renameTo(pendingDelete)) {
+            return false;
+        }
+
+        FaceFeatureIdentity previousIdentity = InspireFace.FeatureHubGetFaceIdentity(id);
+        if (previousIdentity != null && !InspireFace.FeatureHubFaceRemove(id)
+                && InspireFace.FeatureHubGetFaceIdentity(id) != null) {
+            restorePendingDelete(crop, pendingDelete, hadCrop);
+            return false;
+        }
+        boolean metadataRemoved = metadata.edit().remove(key(id)).commit();
+        if (!metadataRemoved) {
+            if (previousIdentity != null
+                    && InspireFace.FeatureHubGetFaceIdentity(id) == null) {
+                InspireFace.FeatureHubInsertFeature(previousIdentity);
+            }
+            restorePendingDelete(crop, pendingDelete, hadCrop);
+            return false;
+        }
+        // The identity is already logically deleted. A leftover tombstone is harmless and
+        // will be cleaned the next time this path is touched.
+        pendingDelete.delete();
+        return true;
+    }
+
+    public File databaseFile() {
+        return databaseFile;
+    }
+
+    public File cropDirectory() {
+        return cropDirectory;
+    }
+
+    private File cropFile(long id) {
+        return new File(cropDirectory, id + ".jpg");
+    }
+
+    @Nullable
+    private static File stageCrop(File destination, Bitmap crop) {
+        File parent = destination.getParentFile();
+        if (parent == null || (!parent.exists() && !parent.mkdirs())) {
+            return null;
+        }
+        File backup = backupFile(destination);
+        if (!recoverCropReplacement(destination, backup)) {
+            return null;
+        }
+        File temp = new File(parent, destination.getName() + ".tmp");
+        if (temp.exists() && !temp.delete()) {
+            return null;
+        }
+        try (FileOutputStream output = new FileOutputStream(temp)) {
+            if (!crop.compress(Bitmap.CompressFormat.JPEG, 92, output)) {
+                temp.delete();
+                return null;
+            }
+            output.flush();
+            output.getFD().sync();
+        } catch (IOException e) {
+            temp.delete();
+            return null;
+        }
+        return temp;
+    }
+
+    private static boolean commitStagedCrop(File destination, File staged, File backup) {
+        if (backup.exists() && !backup.delete()) {
+            return false;
+        }
+        boolean hadDestination = destination.exists();
+        if (hadDestination && !destination.renameTo(backup)) {
+            staged.delete();
+            return false;
+        }
+        if (!staged.renameTo(destination)) {
+            if (hadDestination) {
+                backup.renameTo(destination);
+            }
+            staged.delete();
+            return false;
+        }
+        return true;
+    }
+
+    private static void restoreCrop(File destination, File backup, boolean hadPreviousCrop) {
+        if (destination.exists()) {
+            destination.delete();
+        }
+        if (hadPreviousCrop && backup.exists()) {
+            backup.renameTo(destination);
+        } else {
+            backup.delete();
+        }
+    }
+
+    private static boolean recoverCropReplacement(File destination, File backup) {
+        if (!backup.exists()) {
+            return true;
+        }
+        if (destination.exists()) {
+            return backup.delete();
+        }
+        return backup.renameTo(destination);
+    }
+
+    private static boolean recoverPendingDelete(File crop, File pendingDelete) {
+        if (!pendingDelete.exists()) {
+            return true;
+        }
+        if (crop.exists()) {
+            return pendingDelete.delete();
+        }
+        return pendingDelete.renameTo(crop);
+    }
+
+    private static void restorePendingDelete(File crop, File pendingDelete, boolean hadCrop) {
+        if (hadCrop && pendingDelete.exists()) {
+            pendingDelete.renameTo(crop);
+        } else if (!hadCrop) {
+            pendingDelete.delete();
+        }
+    }
+
+    private static void rollbackFeature(@Nullable FaceFeatureIdentity previous) {
+        if (previous != null) {
+            InspireFace.FeatureHubFaceUpdate(previous);
+        }
+    }
+
+    private static File backupFile(File destination) {
+        return new File(destination.getParentFile(), destination.getName() + ".bak");
+    }
+
+    private static String normalizedName(String name, long id) {
+        String trimmed = name == null ? "" : name.trim();
+        return trimmed.isEmpty() ? "Face " + id : trimmed;
+    }
+
+    private static String key(long id) {
+        return KEY_PREFIX + id;
+    }
+
+    private static String encode(FaceRecord record) {
+        try {
+            return new JSONObject()
+                    .put("id", record.id)
+                    .put("name", record.name)
+                    .put("crop", record.cropPath)
+                    .put("updated", record.updatedAt)
+                    .toString();
+        } catch (JSONException impossible) {
+            throw new IllegalStateException(impossible);
+        }
+    }
+
+    @Nullable
+    private static FaceRecord decode(@Nullable String value) {
+        if (value == null) {
+            return null;
+        }
+        try {
+            JSONObject json = new JSONObject(value);
+            return new FaceRecord(json.getLong("id"), json.getString("name"),
+                    json.getString("crop"), json.getLong("updated"));
+        } catch (JSONException e) {
+            return null;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/ImageBitmapLoader.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/ImageBitmapLoader.java
new file mode 100644
index 000000000..0e5421732
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/face/ImageBitmapLoader.java
@@ -0,0 +1,106 @@
+package com.example.inspireface_example.face;
+
+import android.content.Context;
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.graphics.Matrix;
+import android.net.Uri;
+
+import androidx.exifinterface.media.ExifInterface;
+
+import java.io.IOException;
+import java.io.InputStream;
+
+/** Decodes a content Uri with bounded memory and applies its EXIF orientation. */
+public final class ImageBitmapLoader {
+
+    private ImageBitmapLoader() {
+    }
+
+    public static Bitmap decode(Context context, Uri uri, int maxDimension) throws IOException {
+        int orientation = readOrientation(context, uri);
+        BitmapFactory.Options bounds = new BitmapFactory.Options();
+        bounds.inJustDecodeBounds = true;
+        try (InputStream input = openInput(context, uri)) {
+            BitmapFactory.decodeStream(input, null, bounds);
+        }
+
+        BitmapFactory.Options options = new BitmapFactory.Options();
+        options.inPreferredConfig = Bitmap.Config.ARGB_8888;
+        options.inSampleSize = calculateSampleSize(
+                bounds.outWidth, bounds.outHeight, maxDimension);
+        Bitmap decoded;
+        try (InputStream input = openInput(context, uri)) {
+            decoded = BitmapFactory.decodeStream(input, null, options);
+        }
+        if (decoded == null) {
+            throw new IOException("Unable to decode selected image");
+        }
+        return applyExifOrientation(decoded, orientation);
+    }
+
+    private static int readOrientation(Context context, Uri uri) {
+        try (InputStream input = openInput(context, uri)) {
+            return new ExifInterface(input).getAttributeInt(
+                    ExifInterface.TAG_ORIENTATION, ExifInterface.ORIENTATION_NORMAL);
+        } catch (IOException | RuntimeException ignored) {
+            return ExifInterface.ORIENTATION_NORMAL;
+        }
+    }
+
+    private static InputStream openInput(Context context, Uri uri) throws IOException {
+        InputStream input = context.getContentResolver().openInputStream(uri);
+        if (input == null) {
+            throw new IOException("Unable to open selected image");
+        }
+        return input;
+    }
+
+    private static int calculateSampleSize(int width, int height, int maxDimension) {
+        int longest = Math.max(width, height);
+        int sample = 1;
+        while (longest > 0 && longest / sample > maxDimension) {
+            sample *= 2;
+        }
+        return sample;
+    }
+
+    private static Bitmap applyExifOrientation(Bitmap bitmap, int orientation) {
+        Matrix matrix = new Matrix();
+        switch (orientation) {
+            case ExifInterface.ORIENTATION_FLIP_HORIZONTAL:
+                matrix.setScale(-1f, 1f);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_180:
+                matrix.setRotate(180f);
+                break;
+            case ExifInterface.ORIENTATION_FLIP_VERTICAL:
+                matrix.setRotate(180f);
+                matrix.postScale(-1f, 1f);
+                break;
+            case ExifInterface.ORIENTATION_TRANSPOSE:
+                matrix.setRotate(90f);
+                matrix.postScale(-1f, 1f);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_90:
+                matrix.setRotate(90f);
+                break;
+            case ExifInterface.ORIENTATION_TRANSVERSE:
+                matrix.setRotate(-90f);
+                matrix.postScale(-1f, 1f);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_270:
+                matrix.setRotate(-90f);
+                break;
+            case ExifInterface.ORIENTATION_NORMAL:
+            default:
+                return bitmap;
+        }
+        Bitmap oriented = Bitmap.createBitmap(
+                bitmap, 0, 0, bitmap.getWidth(), bitmap.getHeight(), matrix, true);
+        if (oriented != bitmap) {
+            bitmap.recycle();
+        }
+        return oriented;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionCoordinator.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionCoordinator.java
new file mode 100644
index 000000000..7ffb41fe5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionCoordinator.java
@@ -0,0 +1,236 @@
+package com.example.inspireface_example.permission;
+
+import android.Manifest;
+import android.content.Context;
+import android.content.Intent;
+import android.content.SharedPreferences;
+import android.content.pm.PackageManager;
+import android.net.Uri;
+import android.provider.Settings;
+import android.view.View;
+
+import androidx.activity.result.ActivityResultLauncher;
+import androidx.activity.result.contract.ActivityResultContracts;
+import androidx.appcompat.app.AlertDialog;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.app.ActivityCompat;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+import com.google.android.material.button.MaterialButton;
+import com.google.android.material.dialog.MaterialAlertDialogBuilder;
+
+/**
+ * Shared camera-permission UX for every live feature.
+ *
+ * <p>It owns the one-time on-device privacy notice, Android permission request,
+ * rationale, permanent-denial settings recovery, and the return-from-settings check.</p>
+ */
+public final class CameraPermissionCoordinator {
+
+    public interface Listener {
+        void onCameraPermissionGranted();
+
+        void onCameraPermissionBlocked(boolean requiresSettings);
+    }
+
+    public static final String PREFERENCES_NAME = "camera_permission";
+    private static final String KEY_NOTICE_ACCEPTED = "notice_accepted";
+    private static final String KEY_REQUESTED_BEFORE = "requested_before";
+
+    private final AppCompatActivity activity;
+    private final Listener listener;
+    private final SharedPreferences preferences;
+    private final ActivityResultLauncher<String> permissionLauncher;
+
+    private MaterialButton recoveryButton;
+    private AlertDialog activeDialog;
+    private boolean permissionRequestInFlight;
+    private boolean awaitingSettings;
+    private boolean closed;
+
+    public CameraPermissionCoordinator(AppCompatActivity activity, Listener listener) {
+        this.activity = activity;
+        this.listener = listener;
+        preferences = activity.getSharedPreferences(
+                PREFERENCES_NAME, Context.MODE_PRIVATE);
+        permissionLauncher = activity.registerForActivityResult(
+                new ActivityResultContracts.RequestPermission(), this::onPermissionResult);
+    }
+
+    /** Binds the persistent Try again / Open settings action shown by each camera page. */
+    public void bindRecoveryButton(MaterialButton button) {
+        recoveryButton = button;
+        recoveryButton.setOnClickListener(v -> performRecoveryAction());
+        recoveryButton.setVisibility(View.GONE);
+    }
+
+    /** Starts or resumes the permission flow without issuing duplicate requests/dialogs. */
+    public void requestAccess() {
+        if (closed || activity.isFinishing() || activity.isDestroyed()
+                || permissionRequestInFlight || isDialogShowing()) {
+            return;
+        }
+        CameraPermissionPolicy.Action action = CameraPermissionPolicy.nextAction(
+                preferences.getBoolean(KEY_NOTICE_ACCEPTED, false),
+                hasPermission(),
+                preferences.getBoolean(KEY_REQUESTED_BEFORE, false),
+                shouldShowRationale());
+        switch (action) {
+            case SHOW_NOTICE:
+                showPrivacyNotice();
+                break;
+            case GRANTED:
+                notifyGranted();
+                break;
+            case REQUEST_SYSTEM_PERMISSION:
+                launchSystemPermission();
+                break;
+            case SHOW_RATIONALE:
+                showRationale();
+                break;
+            case OPEN_SETTINGS:
+                notifyBlocked(true);
+                break;
+        }
+    }
+
+    /** Call from Activity.onResume so a Settings grant takes effect without reopening the page. */
+    public void onResume() {
+        if (closed || !awaitingSettings) {
+            return;
+        }
+        awaitingSettings = false;
+        if (hasPermission()) {
+            notifyGranted();
+        } else {
+            notifyBlocked(true);
+        }
+    }
+
+    public boolean hasPermission() {
+        return ContextCompat.checkSelfPermission(activity, Manifest.permission.CAMERA)
+                == PackageManager.PERMISSION_GRANTED;
+    }
+
+    public void close() {
+        closed = true;
+        if (activeDialog != null) {
+            activeDialog.dismiss();
+            activeDialog = null;
+        }
+        recoveryButton = null;
+    }
+
+    private void showPrivacyNotice() {
+        activeDialog = new MaterialAlertDialogBuilder(activity)
+                .setTitle(R.string.camera_privacy_notice_title)
+                .setMessage(R.string.camera_privacy_notice_message)
+                .setNegativeButton(R.string.camera_permission_not_now, (dialog, which) -> {
+                    activeDialog = null;
+                    notifyBlocked(false);
+                })
+                .setPositiveButton(R.string.camera_permission_continue, (dialog, which) -> {
+                    activeDialog = null;
+                    preferences.edit().putBoolean(KEY_NOTICE_ACCEPTED, true).apply();
+                    requestAccess();
+                })
+                .setOnCancelListener(dialog -> {
+                    activeDialog = null;
+                    notifyBlocked(false);
+                })
+                .show();
+    }
+
+    private void showRationale() {
+        activeDialog = new MaterialAlertDialogBuilder(activity)
+                .setTitle(R.string.camera_permission_rationale_title)
+                .setMessage(R.string.camera_permission_rationale_message)
+                .setNegativeButton(android.R.string.cancel, (dialog, which) -> {
+                    activeDialog = null;
+                    notifyBlocked(false);
+                })
+                .setPositiveButton(R.string.camera_permission_try_again, (dialog, which) -> {
+                    activeDialog = null;
+                    launchSystemPermission();
+                })
+                .setOnCancelListener(dialog -> {
+                    activeDialog = null;
+                    notifyBlocked(false);
+                })
+                .show();
+    }
+
+    private void launchSystemPermission() {
+        preferences.edit().putBoolean(KEY_REQUESTED_BEFORE, true).apply();
+        permissionRequestInFlight = true;
+        permissionLauncher.launch(Manifest.permission.CAMERA);
+    }
+
+    private void onPermissionResult(boolean granted) {
+        permissionRequestInFlight = false;
+        if (closed) {
+            return;
+        }
+        if (granted) {
+            notifyGranted();
+        } else {
+            notifyBlocked(!shouldShowRationale());
+        }
+    }
+
+    private void performRecoveryAction() {
+        if (closed) {
+            return;
+        }
+        if (hasPermission()) {
+            notifyGranted();
+            return;
+        }
+        boolean requestedBefore = preferences.getBoolean(KEY_REQUESTED_BEFORE, false);
+        if (requestedBefore && !shouldShowRationale()) {
+            openApplicationSettings();
+        } else {
+            requestAccess();
+        }
+    }
+
+    private void openApplicationSettings() {
+        awaitingSettings = true;
+        Intent intent = new Intent(Settings.ACTION_APPLICATION_DETAILS_SETTINGS,
+                Uri.fromParts("package", activity.getPackageName(), null));
+        activity.startActivity(intent);
+    }
+
+    private void notifyGranted() {
+        if (closed) {
+            return;
+        }
+        if (recoveryButton != null) {
+            recoveryButton.setVisibility(View.GONE);
+        }
+        listener.onCameraPermissionGranted();
+    }
+
+    private void notifyBlocked(boolean requiresSettings) {
+        if (closed) {
+            return;
+        }
+        if (recoveryButton != null) {
+            recoveryButton.setText(requiresSettings
+                    ? R.string.camera_permission_open_settings
+                    : R.string.camera_permission_try_again);
+            recoveryButton.setVisibility(View.VISIBLE);
+        }
+        listener.onCameraPermissionBlocked(requiresSettings);
+    }
+
+    private boolean shouldShowRationale() {
+        return ActivityCompat.shouldShowRequestPermissionRationale(
+                activity, Manifest.permission.CAMERA);
+    }
+
+    private boolean isDialogShowing() {
+        return activeDialog != null && activeDialog.isShowing();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionPolicy.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionPolicy.java
new file mode 100644
index 000000000..087a6d3a0
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/permission/CameraPermissionPolicy.java
@@ -0,0 +1,32 @@
+package com.example.inspireface_example.permission;
+
+/** Pure decision table for the camera permission flow; kept Android-free for unit tests. */
+public final class CameraPermissionPolicy {
+
+    public enum Action {
+        SHOW_NOTICE,
+        GRANTED,
+        REQUEST_SYSTEM_PERMISSION,
+        SHOW_RATIONALE,
+        OPEN_SETTINGS
+    }
+
+    private CameraPermissionPolicy() {
+    }
+
+    public static Action nextAction(boolean noticeAccepted,
+                                    boolean permissionGranted,
+                                    boolean requestedBefore,
+                                    boolean shouldShowRationale) {
+        if (!noticeAccepted) {
+            return Action.SHOW_NOTICE;
+        }
+        if (permissionGranted) {
+            return Action.GRANTED;
+        }
+        if (!requestedBefore) {
+            return Action.REQUEST_SYSTEM_PERMISSION;
+        }
+        return shouldShowRationale ? Action.SHOW_RATIONALE : Action.OPEN_SETTINGS;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/ActionLivenessActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/ActionLivenessActivity.java
new file mode 100644
index 000000000..786bfcbc7
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/ActionLivenessActivity.java
@@ -0,0 +1,17 @@
+package com.example.inspireface_example.view;
+
+import com.example.inspireface_example.R;
+
+/** Dedicated cooperative action-liveness screen. */
+public final class ActionLivenessActivity extends LivenessActivity {
+
+    @Override
+    protected LivenessController.Mode initialMode() {
+        return LivenessController.Mode.ACTION;
+    }
+
+    @Override
+    protected int pageTitleRes() {
+        return R.string.mode_action;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/CameraPreviewController.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/CameraPreviewController.java
new file mode 100644
index 000000000..af3b37164
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/CameraPreviewController.java
@@ -0,0 +1,187 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.util.Size;
+
+import androidx.annotation.StringRes;
+import androidx.camera.core.CameraInfoUnavailableException;
+import androidx.camera.core.CameraSelector;
+import androidx.camera.core.ImageAnalysis;
+import androidx.camera.core.Preview;
+import androidx.camera.core.resolutionselector.AspectRatioStrategy;
+import androidx.camera.core.resolutionselector.ResolutionSelector;
+import androidx.camera.core.resolutionselector.ResolutionStrategy;
+import androidx.camera.lifecycle.ProcessCameraProvider;
+import androidx.camera.view.PreviewView;
+import androidx.core.content.ContextCompat;
+import androidx.lifecycle.LifecycleOwner;
+
+import com.example.inspireface_example.R;
+import com.google.common.util.concurrent.ListenableFuture;
+
+import java.util.concurrent.ExecutionException;
+import java.util.concurrent.Executor;
+
+/**
+ * Reusable CameraX preview + analysis component shared by live demos and enrollment.
+ * It owns lens fallback/switching and keeps both use cases on the same 4:3 crop.
+ */
+public final class CameraPreviewController {
+
+    public interface Listener {
+        /** Initial lens is ready; false means the device fell back to the rear camera. */
+        void onCameraReady(boolean frontCamera);
+
+        /** Called only after a user-requested lens switch succeeds. */
+        void onLensChanged(boolean frontCamera);
+
+        void onCameraError(@StringRes int messageRes);
+    }
+
+    private final Context context;
+    private final LifecycleOwner lifecycleOwner;
+    private final PreviewView previewView;
+    private final Executor analysisExecutor;
+    private final ImageAnalysis.Analyzer analyzer;
+    private final Listener listener;
+
+    private ProcessCameraProvider cameraProvider;
+    private Preview preview;
+    private ImageAnalysis analysis;
+    private boolean useFrontCamera = true;
+    private boolean stopped;
+
+    public CameraPreviewController(Context context, LifecycleOwner lifecycleOwner,
+                                   PreviewView previewView, Executor analysisExecutor,
+                                   ImageAnalysis.Analyzer analyzer, Listener listener) {
+        this.context = context.getApplicationContext();
+        this.lifecycleOwner = lifecycleOwner;
+        this.previewView = previewView;
+        this.analysisExecutor = analysisExecutor;
+        this.analyzer = analyzer;
+        this.listener = listener;
+    }
+
+    /** Asynchronously acquires CameraX and binds a front camera, with rear fallback. */
+    public void start() {
+        start(true);
+    }
+
+    /** Asynchronously acquires CameraX and binds the requested lens, with opposite fallback. */
+    public void start(boolean preferFrontCamera) {
+        stopped = false;
+        useFrontCamera = preferFrontCamera;
+        ListenableFuture<ProcessCameraProvider> future =
+                ProcessCameraProvider.getInstance(context);
+        future.addListener(() -> {
+            if (stopped) {
+                return;
+            }
+            try {
+                cameraProvider = future.get();
+            } catch (ExecutionException e) {
+                listener.onCameraError(R.string.msg_engine_failed);
+                return;
+            } catch (InterruptedException e) {
+                Thread.currentThread().interrupt();
+                listener.onCameraError(R.string.msg_engine_failed);
+                return;
+            }
+            CameraSelector requested = useFrontCamera
+                    ? CameraSelector.DEFAULT_FRONT_CAMERA
+                    : CameraSelector.DEFAULT_BACK_CAMERA;
+            if (!hasCamera(requested)) {
+                CameraSelector fallback = useFrontCamera
+                        ? CameraSelector.DEFAULT_BACK_CAMERA
+                        : CameraSelector.DEFAULT_FRONT_CAMERA;
+                if (hasCamera(fallback)) {
+                    useFrontCamera = !useFrontCamera;
+                } else {
+                    listener.onCameraError(R.string.msg_no_front_camera);
+                    return;
+                }
+            }
+            createUseCases();
+            if (bindCurrentLens()) {
+                listener.onCameraReady(useFrontCamera);
+            }
+        }, ContextCompat.getMainExecutor(context));
+    }
+
+    /** Returns false if CameraX is not ready or the opposite lens does not exist. */
+    public boolean flipCamera() {
+        if (cameraProvider == null || stopped) {
+            return false;
+        }
+        boolean targetFront = !useFrontCamera;
+        if (!hasCamera(targetFront ? CameraSelector.DEFAULT_FRONT_CAMERA
+                : CameraSelector.DEFAULT_BACK_CAMERA)) {
+            return false;
+        }
+        useFrontCamera = targetFront;
+        if (!bindCurrentLens()) {
+            useFrontCamera = !targetFront;
+            bindCurrentLens();
+            return false;
+        }
+        listener.onLensChanged(useFrontCamera);
+        return true;
+    }
+
+    /** Stops only this component's use cases; it does not disturb another Activity. */
+    public void stop() {
+        stopped = true;
+        if (analysis != null) {
+            analysis.clearAnalyzer();
+        }
+        if (cameraProvider != null && preview != null && analysis != null) {
+            cameraProvider.unbind(preview, analysis);
+        }
+    }
+
+    private void createUseCases() {
+        ResolutionSelector analysisResolution = new ResolutionSelector.Builder()
+                .setAspectRatioStrategy(AspectRatioStrategy.RATIO_4_3_FALLBACK_AUTO_STRATEGY)
+                .setResolutionStrategy(new ResolutionStrategy(new Size(640, 480),
+                        ResolutionStrategy.FALLBACK_RULE_CLOSEST_HIGHER_THEN_LOWER))
+                .build();
+        analysis = new ImageAnalysis.Builder()
+                .setResolutionSelector(analysisResolution)
+                .setBackpressureStrategy(ImageAnalysis.STRATEGY_KEEP_ONLY_LATEST)
+                .build();
+        analysis.setAnalyzer(analysisExecutor, analyzer);
+
+        preview = new Preview.Builder()
+                .setResolutionSelector(new ResolutionSelector.Builder()
+                        .setAspectRatioStrategy(
+                                AspectRatioStrategy.RATIO_4_3_FALLBACK_AUTO_STRATEGY)
+                        .build())
+                .build();
+        preview.setSurfaceProvider(previewView.getSurfaceProvider());
+    }
+
+    private boolean bindCurrentLens() {
+        if (cameraProvider == null || preview == null || analysis == null || stopped) {
+            return false;
+        }
+        try {
+            cameraProvider.unbind(preview, analysis);
+            cameraProvider.bindToLifecycle(lifecycleOwner,
+                    useFrontCamera ? CameraSelector.DEFAULT_FRONT_CAMERA
+                            : CameraSelector.DEFAULT_BACK_CAMERA,
+                    preview, analysis);
+            return true;
+        } catch (RuntimeException e) {
+            listener.onCameraError(R.string.msg_camera_unavailable);
+            return false;
+        }
+    }
+
+    private boolean hasCamera(CameraSelector selector) {
+        try {
+            return cameraProvider != null && cameraProvider.hasCamera(selector);
+        } catch (CameraInfoUnavailableException e) {
+            return false;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/EnrollmentFaceAnalyzer.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/EnrollmentFaceAnalyzer.java
new file mode 100644
index 000000000..b2575cea9
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/EnrollmentFaceAnalyzer.java
@@ -0,0 +1,214 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.ImageFormat;
+import android.graphics.Rect;
+import android.graphics.RectF;
+import android.graphics.YuvImage;
+import android.os.SystemClock;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.io.File;
+import java.io.FileOutputStream;
+import java.io.IOException;
+
+/** Tracks SDK face 0 and captures it after one warm-up second plus two stable seconds. */
+final class EnrollmentFaceAnalyzer extends UprightFaceCameraAnalyzer {
+
+    enum Stage { NO_FACE, MOVE_CLOSER, HOLD_STILL, CAPTURING, COMPLETE }
+
+    interface Listener {
+        /** Called on the analysis thread. Progress is meaningful only for CAPTURING. */
+        void onState(Stage stage, float progress);
+
+        /** Called once with an app-private temporary JPEG. */
+        void onCaptured(String path);
+
+        void onCaptureError();
+
+        void onSessionError();
+    }
+
+    private static final long STATE_REPORT_INTERVAL_MS = 80L;
+    private static final float MIN_FACE_WIDTH_RATIO = 0.18f;
+
+    private final FaceOverlayView overlay;
+    private final Listener listener;
+    private final File captureDirectory;
+    private final FaceStabilityGate stabilityGate = new FaceStabilityGate();
+    private final int accentColor;
+    private final int redColor;
+    private final int yellowColor;
+    private final int greenColor;
+    private long lastStateReport;
+    private Stage lastStage;
+    private int lastProgressBucket = -1;
+    private volatile boolean mirrored = true;
+    private volatile boolean resetRequested;
+    private boolean captured;
+
+    EnrollmentFaceAnalyzer(Context context, FaceOverlayView overlay,
+                           File captureDirectory, Listener listener) {
+        this.overlay = overlay;
+        this.captureDirectory = captureDirectory;
+        this.listener = listener;
+        accentColor = ContextCompat.getColor(context, R.color.liveness_accent);
+        redColor = ContextCompat.getColor(context, R.color.liveness_fail);
+        yellowColor = ContextCompat.getColor(context, R.color.liveness_warn);
+        greenColor = accentColor;
+    }
+
+    @Override
+    protected boolean shouldSkipFrame() {
+        return captured;
+    }
+
+    @Override
+    protected void beforeFrame() {
+        if (resetRequested) {
+            resetRequested = false;
+            clearStability();
+        }
+    }
+
+    @Override
+    protected Session createSession() {
+        return FaceEngine.createTrackingSession();
+    }
+
+    @Override
+    protected void onFaces(Session session, ImageStream stream,
+                           @Nullable MultipleFaceData faces, byte[] upright,
+                           int uprightWidth, int uprightHeight, long frameStart) {
+        if (faces == null || faces.detectedNum == 0) {
+            clearStability();
+            overlay.submit(null);
+            reportState(Stage.NO_FACE, 0f);
+            return;
+        }
+        FaceRect first = faces.rects[0];
+        RectF face = new RectF(first.x, first.y,
+                first.x + first.width, first.y + first.height);
+        int trackId = faces.trackIds != null && faces.trackIds.length > 0
+                ? faces.trackIds[0] : 0;
+        if (face.width() < uprightWidth * MIN_FACE_WIDTH_RATIO) {
+            clearStability();
+            submitFrame(face, uprightWidth, uprightHeight,
+                    -1f, redColor, redColor);
+            reportState(Stage.MOVE_CLOSER, 0f);
+            return;
+        }
+
+        float progress = stabilityGate.update(trackId,
+                face.left, face.top, face.right, face.bottom,
+                SystemClock.elapsedRealtime());
+        if (progress < 0f) {
+            submitFrame(face, uprightWidth, uprightHeight,
+                    -1f, accentColor, accentColor);
+            reportState(Stage.HOLD_STILL, 0f);
+            return;
+        }
+
+        int progressColor = progress >= 1f ? greenColor
+                : progress >= 0.5f ? yellowColor : redColor;
+        submitFrame(face, uprightWidth, uprightHeight,
+                progress, progressColor, progressColor);
+        reportState(progress >= 1f ? Stage.COMPLETE : Stage.CAPTURING, progress);
+        if (progress >= 1f) {
+            File capture = writeCapture(upright, uprightWidth, uprightHeight, face);
+            if (capture == null) {
+                clearStability();
+                listener.onCaptureError();
+            } else {
+                captured = true;
+                listener.onCaptured(capture.getAbsolutePath());
+            }
+        }
+    }
+
+    @Override
+    protected void onSessionError() {
+        listener.onSessionError();
+    }
+
+    void setMirrored(boolean mirrored) {
+        this.mirrored = mirrored;
+    }
+
+    void resetTracking() {
+        resetRequested = true;
+        overlay.submit(null);
+    }
+
+    private void clearStability() {
+        stabilityGate.reset();
+    }
+
+    private void submitFrame(RectF face, int imageWidth, int imageHeight,
+                             float progress, int boxColor, int progressColor) {
+        overlay.submit(new FaceOverlayView.Frame(imageWidth, imageHeight, mirrored,
+                new RectF[]{new RectF(face)}, boxColor, progress, progressColor));
+    }
+
+    private void reportState(Stage stage, float progress) {
+        long now = SystemClock.elapsedRealtime();
+        int bucket = Math.round(progress * 100f);
+        if (stage != lastStage || bucket != lastProgressBucket
+                || now - lastStateReport >= STATE_REPORT_INTERVAL_MS) {
+            lastStage = stage;
+            lastProgressBucket = bucket;
+            lastStateReport = now;
+            listener.onState(stage, progress);
+        }
+    }
+
+    private File writeCapture(byte[] nv21, int width, int height, RectF face) {
+        if ((!captureDirectory.exists() && !captureDirectory.mkdirs())
+                || !captureDirectory.isDirectory()) {
+            return null;
+        }
+        File destination;
+        try {
+            destination = File.createTempFile("face_", ".jpg", captureDirectory);
+        } catch (IOException e) {
+            return null;
+        }
+        Rect crop = paddedEvenCrop(face, width, height);
+        try (FileOutputStream output = new FileOutputStream(destination)) {
+            boolean compressed = new YuvImage(
+                    nv21, ImageFormat.NV21, width, height, null)
+                    .compressToJpeg(crop, 94, output);
+            if (!compressed) {
+                destination.delete();
+                return null;
+            }
+            output.flush();
+            output.getFD().sync();
+            return destination;
+        } catch (IOException | RuntimeException e) {
+            destination.delete();
+            return null;
+        }
+    }
+
+    private static Rect paddedEvenCrop(RectF face, int width, int height) {
+        float paddingX = face.width() * 0.45f;
+        float paddingY = face.height() * 0.55f;
+        int left = Math.max(0, ((int) Math.floor(face.left - paddingX)) & ~1);
+        int top = Math.max(0, ((int) Math.floor(face.top - paddingY)) & ~1);
+        int right = Math.min(width, ((int) Math.ceil(face.right + paddingX) + 1) & ~1);
+        int bottom = Math.min(height, ((int) Math.ceil(face.bottom + paddingY) + 1) & ~1);
+        if (right <= left + 2 || bottom <= top + 2) {
+            return new Rect(0, 0, width & ~1, height & ~1);
+        }
+        return new Rect(left, top, right, bottom);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceAnalyzer.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceAnalyzer.java
new file mode 100644
index 000000000..b657af6b8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceAnalyzer.java
@@ -0,0 +1,195 @@
+package com.example.inspireface_example.view;
+
+import android.graphics.RectF;
+import android.os.SystemClock;
+
+import androidx.annotation.Nullable;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceEulerAngle;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Point2f;
+import com.insightface.sdk.inspireface.base.Session;
+
+/**
+ * Per-frame pipeline: YUV_420_888 → NV21 → rotate upright → InspireFace track → mode
+ * logic → UI.
+ *
+ * Frames are handed to the SDK already upright with CAMERA_ROTATION_0 (see
+ * {@link Nv21Converter} for why), so every SDK coordinate is directly in preview
+ * orientation. Everything runs on the single-threaded analysis executor, which also
+ * keeps all session calls serialized. The ImageProxy is closed as soon as the NV21 copy
+ * exists so CameraX can refill the buffer while inference runs, and the stream is
+ * created, used and released within this frame because the SDK aliases the byte[]
+ * rather than copying it.
+ */
+final class FaceAnalyzer extends UprightFaceCameraAnalyzer {
+
+    interface Listener {
+        /** Called on the analysis thread. */
+        void onUiState(LivenessController.UiState state);
+
+        /** Called on the analysis thread, throttled. */
+        void onPerf(double fps, long latencyMs);
+
+        /**
+         * Called on the analysis thread while the euler readout is enabled, throttled.
+         * {@code angle} is null when no face is tracked.
+         */
+        void onEulerAngles(FaceEulerAngle angle);
+
+        /** Called once on the analysis thread if the session cannot be created. */
+        void onSessionError();
+    }
+
+    private static final long PERF_REPORT_INTERVAL_MS = 500;
+    private static final long EULER_REPORT_INTERVAL_MS = 100;
+
+    private static final int LANDMARK_FLOATS = 106 * 2;
+
+    private final LivenessController controller;
+    private final FaceOverlayView overlay;
+    private final LandmarkGlView landmarkView;
+    private final Listener listener;
+    /** Front camera previews are displayed mirrored, back camera ones are not. */
+    private volatile boolean mirrored;
+    private float[] landmarkScratch = new float[LANDMARK_FLOATS];
+
+    private volatile boolean eulerEnabled;
+    private volatile boolean landmarksEnabled;
+    private long lastEulerReport;
+
+    // Perf tracking
+    private long windowStart;
+    private int windowFrames;
+    private double fps;
+    private double emaLatencyMs;
+    private long lastPerfReport;
+
+    FaceAnalyzer(LivenessController controller, FaceOverlayView overlay,
+                 LandmarkGlView landmarkView, Listener listener, boolean mirrored) {
+        this.controller = controller;
+        this.overlay = overlay;
+        this.landmarkView = landmarkView;
+        this.listener = listener;
+        this.mirrored = mirrored;
+    }
+
+    @Override
+    protected Session createSession() {
+        return FaceEngine.createPreviewSession();
+    }
+
+    @Override
+    protected void onFaces(Session session, ImageStream stream,
+                           @Nullable MultipleFaceData faces, byte[] uprightNv21,
+                           int uprightWidth, int uprightHeight, long frameStart) {
+        if (faces == null) {
+            return;
+        }
+        LivenessController.UiState state =
+                controller.onFrame(session, stream, faces, uprightWidth, uprightHeight);
+        overlay.submit(buildOverlayFrame(faces, uprightWidth, uprightHeight, state.boxColor));
+        listener.onUiState(state);
+        reportEulerAngles(faces);
+        renderLandmarks(faces, uprightWidth, uprightHeight);
+        trackPerf(frameStart);
+    }
+
+    @Override
+    protected void onSessionError() {
+        listener.onSessionError();
+    }
+
+    /** Safe to call from any thread. */
+    void setEulerEnabled(boolean enabled) {
+        eulerEnabled = enabled;
+    }
+
+    /**
+     * Safe to call from any thread; flip together with the camera lens. Rotation needs
+     * no per-lens handling — every frame is pre-rotated by its own rotationDegrees — so
+     * mirroring is the only display difference between the lenses.
+     */
+    void setMirrored(boolean mirrored) {
+        this.mirrored = mirrored;
+    }
+
+    /** Safe to call from any thread. */
+    void setLandmarksEnabled(boolean enabled) {
+        landmarksEnabled = enabled;
+    }
+
+    /** Dense 106-point landmarks for every tracked face, batched into one GL submission. */
+    private void renderLandmarks(MultipleFaceData faces, int uprightWidth, int uprightHeight) {
+        if (!landmarksEnabled) {
+            return;
+        }
+        if (faces.detectedNum == 0) {
+            landmarkView.clearPoints();
+            return;
+        }
+        int needed = faces.detectedNum * LANDMARK_FLOATS;
+        if (landmarkScratch.length < needed) {
+            landmarkScratch = new float[needed];
+        }
+        int n = 0;
+        for (int i = 0; i < faces.detectedNum; i++) {
+            Point2f[] landmarks = InspireFace.GetFaceDenseLandmarkFromFaceToken(faces.tokens[i]);
+            if (landmarks == null) {
+                continue;
+            }
+            for (Point2f p : landmarks) {
+                landmarkScratch[n++] = p.x;
+                landmarkScratch[n++] = p.y;
+            }
+        }
+        landmarkView.submitPoints(landmarkScratch, n / 2, uprightWidth, uprightHeight, mirrored);
+    }
+
+    private void reportEulerAngles(MultipleFaceData faces) {
+        if (!eulerEnabled) {
+            return;
+        }
+        long now = SystemClock.elapsedRealtime();
+        if (now - lastEulerReport < EULER_REPORT_INTERVAL_MS) {
+            return;
+        }
+        lastEulerReport = now;
+        // Only angles[0] is trustworthy — the 1.2.0 JNI writes face[0]'s angles into
+        // every slot, so per-face readout beyond the first face is impossible anyway.
+        listener.onEulerAngles(faces.detectedNum > 0 ? faces.angles[0] : null);
+    }
+
+    /** The stream is already upright, so SDK rects map to the preview directly. */
+    private FaceOverlayView.Frame buildOverlayFrame(MultipleFaceData faces,
+                                                    int uprightWidth, int uprightHeight, int color) {
+        RectF[] rects = new RectF[faces.detectedNum];
+        for (int i = 0; i < faces.detectedNum; i++) {
+            rects[i] = new RectF(faces.rects[i].x, faces.rects[i].y,
+                    faces.rects[i].x + faces.rects[i].width,
+                    faces.rects[i].y + faces.rects[i].height);
+        }
+        return new FaceOverlayView.Frame(
+                uprightWidth, uprightHeight, mirrored, rects, color);
+    }
+
+    private void trackPerf(long start) {
+        long now = SystemClock.elapsedRealtime();
+        long cost = now - start;
+        emaLatencyMs = emaLatencyMs == 0 ? cost : emaLatencyMs * 0.85 + cost * 0.15;
+        windowFrames++;
+        if (windowStart == 0) {
+            windowStart = now;
+        } else if (now - windowStart >= 1000) {
+            fps = windowFrames * 1000.0 / (now - windowStart);
+            windowFrames = 0;
+            windowStart = now;
+        }
+        if (now - lastPerfReport >= PERF_REPORT_INTERVAL_MS && fps > 0) {
+            lastPerfReport = now;
+            listener.onPerf(fps, Math.round(emaLatencyMs));
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceCaptureActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceCaptureActivity.java
new file mode 100644
index 000000000..3029554e8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceCaptureActivity.java
@@ -0,0 +1,263 @@
+package com.example.inspireface_example.view;
+
+import android.content.Intent;
+import android.os.Bundle;
+import android.os.Handler;
+import android.os.Looper;
+import android.view.View;
+import android.view.ViewGroup;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.camera.view.PreviewView;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.FaceModelPrefs;
+import com.example.inspireface_example.R;
+import com.example.inspireface_example.permission.CameraPermissionCoordinator;
+import com.google.android.material.card.MaterialCardView;
+
+import java.io.File;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/** Camera enrollment flow with first-face tracking and a reset-on-motion stability ring. */
+public final class FaceCaptureActivity extends AppCompatActivity
+        implements EnrollmentFaceAnalyzer.Listener {
+
+    public static final String EXTRA_CAPTURE_PATH = "capture_path";
+    private static final long COMPLETE_DISPLAY_MS = 350L;
+
+    private final ExecutorService analysisExecutor = Executors.newSingleThreadExecutor();
+    private final Handler mainHandler = new Handler(Looper.getMainLooper());
+
+    private PreviewView previewView;
+    private FaceOverlayView overlayView;
+    private TextView promptTitle;
+    private TextView promptSubtitle;
+    private View flipButton;
+    private CameraPreviewController cameraController;
+    private EnrollmentFaceAnalyzer analyzer;
+    private boolean resultScheduled;
+    private boolean resultReturned;
+    private String pendingCapturePath;
+    private CameraPermissionCoordinator cameraPermission;
+    private boolean engineStartingOrReady;
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        cameraPermission = new CameraPermissionCoordinator(this,
+                new CameraPermissionCoordinator.Listener() {
+                    @Override
+                    public void onCameraPermissionGranted() {
+                        startEngine();
+                    }
+
+                    @Override
+                    public void onCameraPermissionBlocked(boolean requiresSettings) {
+                        showCameraPermissionBlocked(requiresSettings);
+                    }
+                });
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        setContentView(R.layout.activity_face_capture);
+
+        previewView = findViewById(R.id.capturePreview);
+        overlayView = findViewById(R.id.captureFaceOverlay);
+        promptTitle = findViewById(R.id.capturePromptTitle);
+        promptSubtitle = findViewById(R.id.capturePromptSubtitle);
+        flipButton = findViewById(R.id.btnFlipCaptureCamera);
+        cameraPermission.bindRecoveryButton(
+                findViewById(R.id.btnCameraPermissionAction));
+        ((TextView) findViewById(R.id.currentModel)).setText(
+                getString(R.string.current_model, FaceModelPrefs.get(this).sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+        flipButton.setOnClickListener(v -> flipCamera());
+        applyWindowInsets();
+
+        cameraPermission.requestAccess();
+    }
+
+    private void startEngine() {
+        if (engineStartingOrReady) {
+            return;
+        }
+        engineStartingOrReady = true;
+        promptTitle.setText(R.string.msg_initializing);
+        promptSubtitle.setText(R.string.capture_first_face_hint);
+        analysisExecutor.execute(() -> {
+            boolean ready = FaceEngine.ensureLaunched(this);
+            runOnUiThread(() -> {
+                if (isFinishing() || isDestroyed()) {
+                    return;
+                }
+                if (ready) {
+                    bindCamera();
+                } else {
+                    engineStartingOrReady = false;
+                    promptTitle.setText(R.string.msg_engine_failed);
+                }
+            });
+        });
+    }
+
+    private void showCameraPermissionBlocked(boolean requiresSettings) {
+        promptTitle.setText(R.string.msg_permission_required);
+        promptSubtitle.setText(requiresSettings
+                ? R.string.camera_permission_settings_hint
+                : R.string.camera_permission_retry_hint);
+    }
+
+    private void bindCamera() {
+        File captureDirectory = new File(getCacheDir(), "face_capture");
+        analyzer = new EnrollmentFaceAnalyzer(
+                this, overlayView, captureDirectory, this);
+        cameraController = new CameraPreviewController(
+                this, this, previewView, analysisExecutor, analyzer,
+                new CameraPreviewController.Listener() {
+                    @Override
+                    public void onCameraReady(boolean frontCamera) {
+                        analyzer.setMirrored(frontCamera);
+                        showState(EnrollmentFaceAnalyzer.Stage.NO_FACE, 0f);
+                    }
+
+                    @Override
+                    public void onLensChanged(boolean frontCamera) {
+                        analyzer.setMirrored(frontCamera);
+                        analyzer.resetTracking();
+                        showState(EnrollmentFaceAnalyzer.Stage.NO_FACE, 0f);
+                    }
+
+                    @Override
+                    public void onCameraError(int messageRes) {
+                        promptTitle.setText(messageRes);
+                    }
+                });
+        cameraController.start();
+    }
+
+    private void flipCamera() {
+        if (resultScheduled) {
+            return;
+        }
+        if (cameraController == null || !cameraController.flipCamera()) {
+            Toast.makeText(this, R.string.msg_camera_unavailable, Toast.LENGTH_SHORT).show();
+        }
+    }
+
+    @Override
+    public void onState(EnrollmentFaceAnalyzer.Stage stage, float progress) {
+        runOnUiThread(() -> showState(stage, progress));
+    }
+
+    private void showState(EnrollmentFaceAnalyzer.Stage stage, float progress) {
+        if (isFinishing() || isDestroyed()) {
+            return;
+        }
+        switch (stage) {
+            case MOVE_CLOSER:
+                promptTitle.setText(R.string.capture_move_closer);
+                promptSubtitle.setText(R.string.capture_first_face_hint);
+                break;
+            case HOLD_STILL:
+                promptTitle.setText(R.string.capture_hold_still);
+                promptSubtitle.setText(R.string.capture_warmup_hint);
+                break;
+            case CAPTURING:
+                promptTitle.setText(R.string.capture_keep_still);
+                promptSubtitle.setText(getString(
+                        R.string.capture_progress, Math.round(progress * 100f)));
+                break;
+            case COMPLETE:
+                promptTitle.setText(R.string.capture_complete);
+                promptSubtitle.setText(R.string.capture_complete_hint);
+                break;
+            case NO_FACE:
+            default:
+                promptTitle.setText(R.string.capture_no_face);
+                promptSubtitle.setText(R.string.capture_first_face_hint);
+                break;
+        }
+    }
+
+    @Override
+    public void onCaptured(String path) {
+        runOnUiThread(() -> {
+            if (isFinishing() || isDestroyed()) {
+                new File(path).delete();
+                return;
+            }
+            resultScheduled = true;
+            pendingCapturePath = path;
+            flipButton.setEnabled(false);
+            showState(EnrollmentFaceAnalyzer.Stage.COMPLETE, 1f);
+            mainHandler.postDelayed(() -> {
+                if (isFinishing() || isDestroyed()) {
+                    new File(path).delete();
+                    return;
+                }
+                resultReturned = true;
+                setResult(RESULT_OK,
+                        new Intent().putExtra(EXTRA_CAPTURE_PATH, path));
+                finish();
+            }, COMPLETE_DISPLAY_MS);
+        });
+    }
+
+    @Override
+    public void onCaptureError() {
+        runOnUiThread(() -> {
+            promptTitle.setText(R.string.capture_failed);
+            promptSubtitle.setText(R.string.capture_retry_hint);
+        });
+    }
+
+    @Override
+    public void onSessionError() {
+        runOnUiThread(() -> promptTitle.setText(R.string.msg_engine_failed));
+    }
+
+    private void applyWindowInsets() {
+        View topBar = findViewById(R.id.captureTopBar);
+        MaterialCardView promptCard = findViewById(R.id.capturePromptCard);
+        int baseBottom = (int) (24 * getResources().getDisplayMetrics().density);
+        ViewCompat.setOnApplyWindowInsetsListener(findViewById(R.id.captureRoot), (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            topBar.setPadding(topBar.getPaddingLeft(), bars.top,
+                    topBar.getPaddingRight(), topBar.getPaddingBottom());
+            ViewGroup.MarginLayoutParams params =
+                    (ViewGroup.MarginLayoutParams) promptCard.getLayoutParams();
+            params.bottomMargin = baseBottom + bars.bottom;
+            promptCard.setLayoutParams(params);
+            return insets;
+        });
+    }
+
+    @Override
+    protected void onResume() {
+        super.onResume();
+        cameraPermission.onResume();
+    }
+
+    @Override
+    protected void onDestroy() {
+        cameraPermission.close();
+        mainHandler.removeCallbacksAndMessages(null);
+        if (!resultReturned && pendingCapturePath != null) {
+            new File(pendingCapturePath).delete();
+        }
+        if (cameraController != null) {
+            cameraController.stop();
+        }
+        if (analyzer != null) {
+            EnrollmentFaceAnalyzer toRelease = analyzer;
+            analysisExecutor.execute(toRelease::release);
+        }
+        analysisExecutor.shutdown();
+        super.onDestroy();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceEngine.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceEngine.java
new file mode 100644
index 000000000..0636095dd
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceEngine.java
@@ -0,0 +1,247 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.util.Log;
+
+import com.example.inspireface_example.DetectorDefaults;
+import com.example.inspireface_example.FaceModelPrefs;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.CustomParameter;
+import com.insightface.sdk.inspireface.base.Session;
+
+/**
+ * Process-wide InspireFace lifecycle: one GlobalLaunch, session factory tuned for
+ * low-latency front-camera preview.
+ */
+public final class FaceEngine {
+
+    private static final String TAG = "FaceEngine";
+    private static final long SESSION_RELEASE_WAIT_MS = 2_000;
+
+    /** Max faces tracked per frame — enough to notice "more than one face" and stay cheap. */
+    private static final int PREVIEW_MAX_FACES = 3;
+    private static final int IMAGE_MAX_FACES = 10;
+
+    /**
+     * Long-edge size the tracker scales frames to before detection. Matches the default
+     * 320 detect level; in LIGHT_TRACK mode detection is amortized (~1 in 20 frames) so
+     * this costs little.
+     */
+    private static final int TRACK_PREVIEW_SIZE = 320;
+
+    private static volatile boolean launched;
+    private static String launchedModel;
+    private static int activeSessions;
+
+    private FaceEngine() {
+    }
+
+    /**
+     * Loads the globally selected model. Changing the home-screen selection terminates the
+     * previous model before the next feature screen launches the new one.
+     */
+    public static synchronized boolean ensureLaunched(Context context) {
+        String requestedModel = FaceModelPrefs.get(context).sdkName();
+        if (launched && requestedModel.equals(launchedModel)) {
+            return true;
+        }
+        if (launched) {
+            // Back navigation can reveal the home page a fraction before the previous
+            // analysis executor releases its session. Briefly wait for that hand-off;
+            // wait() releases this monitor so releaseSession can make progress.
+            long deadline = android.os.SystemClock.elapsedRealtime() + SESSION_RELEASE_WAIT_MS;
+            while (activeSessions > 0) {
+                long remaining = deadline - android.os.SystemClock.elapsedRealtime();
+                if (remaining <= 0) {
+                    Log.e(TAG, "Cannot switch model while " + activeSessions
+                            + " session(s) are active");
+                    return false;
+                }
+                try {
+                    FaceEngine.class.wait(remaining);
+                } catch (InterruptedException e) {
+                    Thread.currentThread().interrupt();
+                    return false;
+                }
+            }
+            if (!InspireFace.GlobalTerminate()) {
+                Log.e(TAG, "GlobalTerminate failed for " + launchedModel);
+                return false;
+            }
+            launched = false;
+            launchedModel = null;
+        }
+        if (!launched) {
+            launched = Boolean.TRUE.equals(
+                    InspireFace.GlobalLaunch(context.getApplicationContext(), requestedModel));
+            launchedModel = launched ? requestedModel : null;
+            Log.i(TAG, "GlobalLaunch(" + requestedModel + ") -> " + launched);
+        }
+        return launched;
+    }
+
+    /**
+     * Session for continuous video tracking with both liveness models loaded. All calls on
+     * the returned session must stay on a single thread.
+     */
+    static synchronized Session createPreviewSession() {
+        // Face quality also loads the pose model — without it yaw/pitch stay 0 and the
+        // shake/head-raise actions can never fire.
+        CustomParameter parameter = InspireFace.CreateCustomParameter()
+                .enableLiveness(true)
+                .enableInteractionLiveness(true)
+                .enableFaceQuality(true);
+        Session session = InspireFace.CreateSession(
+                parameter, InspireFace.DETECT_MODE_LIGHT_TRACK, PREVIEW_MAX_FACES,
+                DetectorDefaults.INPUT_PX, -1);
+        if (session == null) {
+            return null;
+        }
+        activeSessions++;
+        InspireFace.SetTrackPreviewSize(session, TRACK_PREVIEW_SIZE);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, 0);
+        return session;
+    }
+
+    /** Lightweight session for camera enrollment: tracking only, no liveness models. */
+    static synchronized Session createTrackingSession() {
+        return createTrackingSession(
+                InspireFace.CreateCustomParameter(), PREVIEW_MAX_FACES, 0);
+    }
+
+    /** First-face camera tracking with feature extraction enabled for live 1:N search. */
+    static synchronized Session createVideoRecognitionSession() {
+        return createTrackingSession(
+                InspireFace.CreateCustomParameter().enableRecognition(true), 1, 24);
+    }
+
+    private static Session createTrackingSession(
+            CustomParameter parameter, int maxFaces, int minimumFacePixelSize) {
+        Session session = InspireFace.CreateSession(
+                parameter, InspireFace.DETECT_MODE_LIGHT_TRACK,
+                maxFaces, DetectorDefaults.INPUT_PX, -1);
+        if (session == null) {
+            return null;
+        }
+        activeSessions++;
+        InspireFace.SetTrackPreviewSize(session, TRACK_PREVIEW_SIZE);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, minimumFacePixelSize);
+        return session;
+    }
+
+    /** Session for still-image detection and feature extraction. */
+    public static synchronized Session createRecognitionSession() {
+        return createRecognitionSession(DetectorDefaults.INPUT_PX, IMAGE_MAX_FACES, 0);
+    }
+
+    /** Still-image Session with every attribute model used by the attribute demo. */
+    public static synchronized Session createAttributeSession() {
+        CustomParameter parameter = InspireFace.CreateCustomParameter()
+                .enableMaskDetect(true)
+                .enableFaceQuality(true)
+                .enableFaceAttribute(true)
+                .enableInteractionLiveness(true);
+        return createStillImageSession(
+                parameter, DetectorDefaults.INPUT_PX, IMAGE_MAX_FACES, 0);
+    }
+
+    /** Configurable still-image session used by the recognition parameter demo. */
+    public static synchronized Session createRecognitionSession(
+            int detectPixelLevel, int maxDetectFaceNum, int minimumFacePixelSize) {
+        int safeMaxFaces = Math.max(1, maxDetectFaceNum);
+        int safeMinimumFacePixelSize = Math.max(0, minimumFacePixelSize);
+        return createStillImageSession(
+                InspireFace.CreateCustomParameter().enableRecognition(true),
+                detectPixelLevel, safeMaxFaces, safeMinimumFacePixelSize);
+    }
+
+    /** Detection-only still-image Session used by the landmark demo. */
+    public static synchronized Session createDetectionSession(
+            int detectPixelLevel, int maxDetectFaceNum, int minimumFacePixelSize) {
+        int safeMaxFaces = Math.max(1, maxDetectFaceNum);
+        int safeMinimumFacePixelSize = Math.max(0, minimumFacePixelSize);
+        int previewSize = detectPixelLevel > 0
+                ? detectPixelLevel : DetectorDefaults.INPUT_PX;
+
+        /*
+         * Android SDK 1.2.0 exposes enableDetectModeLandmark, but its CreateSession JNI
+         * implementation never copies that Java field into HFSessionCustomParameter. As a
+         * result ALWAYS_DETECT returns valid face tokens without dense landmarks. LIGHT_TRACK
+         * forces landmarks on inside the native tracker (the same path used by the working
+         * camera demo), so use it for this compatibility session. The still-image screen
+         * recreates the session for each new image to prevent tracking state carrying over.
+         */
+        Session session = InspireFace.CreateSession(
+                InspireFace.CreateCustomParameter(), InspireFace.DETECT_MODE_LIGHT_TRACK,
+                safeMaxFaces, detectPixelLevel, -1);
+        if (session == null) {
+            return null;
+        }
+        activeSessions++;
+        InspireFace.SetTrackPreviewSize(session, previewSize);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, safeMinimumFacePixelSize);
+        return session;
+    }
+
+    /**
+     * Configurable continuous tracker with deterministic 106-point tokens in both tracking
+     * modes. The native compatibility bridge is required because the 1.2.0 Java CreateSession
+     * wrapper omits the detect-mode-landmark field.
+     */
+    public static synchronized Session createFaceTrackingSession(
+            int detectMode, int detectPixelLevel,
+            int maxDetectFaceNum, int minimumFacePixelSize) {
+        int safeMode = detectMode == InspireFace.DETECT_MODE_TRACK_BY_DETECTION
+                ? detectMode : InspireFace.DETECT_MODE_LIGHT_TRACK;
+        int safeMaxFaces = Math.max(1, Math.min(maxDetectFaceNum, IMAGE_MAX_FACES));
+        int safeMinimumFacePixelSize = Math.max(0, minimumFacePixelSize);
+        int previewSize = detectPixelLevel > 0 ? detectPixelLevel : TRACK_PREVIEW_SIZE;
+        Session session = NativeSessionBridge.createLandmarkSession(
+                safeMode, safeMaxFaces, detectPixelLevel, 30);
+        if (session == null && safeMode == InspireFace.DETECT_MODE_LIGHT_TRACK) {
+            // LIGHT_TRACK forces landmarks internally and remains a safe fallback if the
+            // bridge cannot be loaded on an unusual device.
+            session = InspireFace.CreateSession(InspireFace.CreateCustomParameter(),
+                    safeMode, safeMaxFaces, detectPixelLevel, -1);
+        }
+        if (session == null) {
+            return null;
+        }
+        activeSessions++;
+        InspireFace.SetTrackPreviewSize(session, previewSize);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, safeMinimumFacePixelSize);
+        return session;
+    }
+
+    private static Session createStillImageSession(
+            CustomParameter parameter, int detectPixelLevel,
+            int maxDetectFaceNum, int minimumFacePixelSize) {
+        int safeMaxFaces = Math.max(1, maxDetectFaceNum);
+        int safeMinimumFacePixelSize = Math.max(0, minimumFacePixelSize);
+        int previewSize = detectPixelLevel > 0
+                ? detectPixelLevel : DetectorDefaults.INPUT_PX;
+        Session session = InspireFace.CreateSession(
+                parameter, InspireFace.DETECT_MODE_ALWAYS_DETECT,
+                safeMaxFaces, detectPixelLevel, -1);
+        if (session == null) {
+            return null;
+        }
+        activeSessions++;
+        InspireFace.SetTrackPreviewSize(session, previewSize);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, safeMinimumFacePixelSize);
+        return session;
+    }
+
+    public static synchronized void releaseSession(Session session) {
+        InspireFace.ReleaseSession(session);
+        if (activeSessions > 0) {
+            activeSessions--;
+        }
+        FaceEngine.class.notifyAll();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceOverlayView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceOverlayView.java
new file mode 100644
index 000000000..3683c5e3d
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceOverlayView.java
@@ -0,0 +1,154 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.Canvas;
+import android.graphics.Paint;
+import android.graphics.RectF;
+import android.util.AttributeSet;
+import android.view.View;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+
+/**
+ * Draws face bounding brackets on top of the camera preview.
+ *
+ * Face rectangles are supplied in the upright image coordinate space (the space the
+ * detector worked in, after rotation). The view maps them to screen space assuming the
+ * preview uses FILL_CENTER (center-crop), and mirrors horizontally for the front camera.
+ */
+public class FaceOverlayView extends View {
+
+    /** Immutable per-frame snapshot handed over from the analysis thread. */
+    public static final class Frame {
+        final int imageWidth;
+        final int imageHeight;
+        final boolean mirrored;
+        final RectF[] rects;
+        final int color;
+        final float progress;
+        final int progressColor;
+
+        public Frame(int imageWidth, int imageHeight, boolean mirrored,
+                     RectF[] rects, int color) {
+            this(imageWidth, imageHeight, mirrored, rects, color, -1f, color);
+        }
+
+        public Frame(int imageWidth, int imageHeight, boolean mirrored,
+                     RectF[] rects, int color, float progress, int progressColor) {
+            this.imageWidth = imageWidth;
+            this.imageHeight = imageHeight;
+            this.mirrored = mirrored;
+            this.rects = rects;
+            this.color = color;
+            this.progress = progress;
+            this.progressColor = progressColor;
+        }
+    }
+
+    private final Paint boxPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint progressTrackPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint progressPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final RectF mapped = new RectF();
+    private final RectF progressBounds = new RectF();
+    private volatile Frame frame;
+
+    public FaceOverlayView(Context context) {
+        this(context, null);
+    }
+
+    public FaceOverlayView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        boxPaint.setStyle(Paint.Style.STROKE);
+        boxPaint.setStrokeWidth(dp(3));
+        boxPaint.setStrokeCap(Paint.Cap.ROUND);
+        boxPaint.setColor(ContextCompat.getColor(context, R.color.liveness_accent));
+        progressTrackPaint.setStyle(Paint.Style.STROKE);
+        progressTrackPaint.setStrokeWidth(dp(7));
+        progressTrackPaint.setColor(0x66000000);
+        progressPaint.setStyle(Paint.Style.STROKE);
+        progressPaint.setStrokeWidth(dp(7));
+        progressPaint.setStrokeCap(Paint.Cap.ROUND);
+    }
+
+    /** Safe to call from any thread. Pass {@code null} rects to clear. */
+    public void submit(@Nullable Frame f) {
+        frame = f;
+        postInvalidateOnAnimation();
+    }
+
+    @Override
+    protected void onDraw(Canvas canvas) {
+        super.onDraw(canvas);
+        Frame f = frame;
+        if (f == null || f.rects == null || f.rects.length == 0
+                || f.imageWidth <= 0 || f.imageHeight <= 0) {
+            return;
+        }
+        float vw = getWidth();
+        float vh = getHeight();
+        // FILL_CENTER: uniform scale that covers the view, centered.
+        float scale = Math.max(vw / f.imageWidth, vh / f.imageHeight);
+        float dx = (vw - f.imageWidth * scale) / 2f;
+        float dy = (vh - f.imageHeight * scale) / 2f;
+
+        boxPaint.setColor(f.color);
+        for (int i = 0; i < f.rects.length; i++) {
+            RectF r = f.rects[i];
+            float left = r.left;
+            float right = r.right;
+            if (f.mirrored) {
+                float l = f.imageWidth - right;
+                right = f.imageWidth - left;
+                left = l;
+            }
+            mapped.set(left * scale + dx, r.top * scale + dy,
+                    right * scale + dx, r.bottom * scale + dy);
+            drawBrackets(canvas, mapped);
+            if (i == 0 && f.progress >= 0f) {
+                drawProgressRing(canvas, mapped, f.progress, f.progressColor);
+            }
+        }
+    }
+
+    private void drawProgressRing(Canvas canvas, RectF face, float progress, int color) {
+        float size = Math.max(face.width(), face.height()) + dp(28);
+        float cx = face.centerX();
+        float cy = face.centerY();
+        progressBounds.set(cx - size / 2f, cy - size / 2f,
+                cx + size / 2f, cy + size / 2f);
+        canvas.drawOval(progressBounds, progressTrackPaint);
+        progressPaint.setColor(color);
+        float clamped = Math.max(0f, Math.min(1f, progress));
+        canvas.drawArc(progressBounds, -90f, Math.max(1f, clamped * 360f),
+                false, progressPaint);
+    }
+
+    /** Corner brackets read better over video than a full box. */
+    private void drawBrackets(Canvas canvas, RectF r) {
+        float len = Math.min(r.width(), r.height()) * 0.22f;
+        float radius = dp(6);
+        // Top-left
+        canvas.drawLine(r.left, r.top + len, r.left, r.top + radius, boxPaint);
+        canvas.drawLine(r.left + radius, r.top, r.left + len, r.top, boxPaint);
+        canvas.drawArc(r.left, r.top, r.left + 2 * radius, r.top + 2 * radius, 180, 90, false, boxPaint);
+        // Top-right
+        canvas.drawLine(r.right - len, r.top, r.right - radius, r.top, boxPaint);
+        canvas.drawLine(r.right, r.top + radius, r.right, r.top + len, boxPaint);
+        canvas.drawArc(r.right - 2 * radius, r.top, r.right, r.top + 2 * radius, 270, 90, false, boxPaint);
+        // Bottom-left
+        canvas.drawLine(r.left, r.bottom - len, r.left, r.bottom - radius, boxPaint);
+        canvas.drawLine(r.left + radius, r.bottom, r.left + len, r.bottom, boxPaint);
+        canvas.drawArc(r.left, r.bottom - 2 * radius, r.left + 2 * radius, r.bottom, 90, 90, false, boxPaint);
+        // Bottom-right
+        canvas.drawLine(r.right - len, r.bottom, r.right - radius, r.bottom, boxPaint);
+        canvas.drawLine(r.right, r.bottom - radius, r.right, r.bottom - len, boxPaint);
+        canvas.drawArc(r.right - 2 * radius, r.bottom - 2 * radius, r.right, r.bottom, 0, 90, false, boxPaint);
+    }
+
+    private float dp(float v) {
+        return v * getResources().getDisplayMetrics().density;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceStabilityGate.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceStabilityGate.java
new file mode 100644
index 000000000..106fb7c2f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceStabilityGate.java
@@ -0,0 +1,99 @@
+package com.example.inspireface_example.view;
+
+/** Pure-Java first-face stability timer, separated so reset/timing behavior is testable. */
+final class FaceStabilityGate {
+
+    static final long WARMUP_MS = 1_000L;
+    static final long PROGRESS_MS = 2_000L;
+    private static final float FRAME_CENTER_TOLERANCE = 0.045f;
+    private static final float FRAME_SIZE_TOLERANCE = 0.07f;
+    private static final float ANCHOR_CENTER_TOLERANCE = 0.11f;
+    private static final float ANCHOR_SIZE_TOLERANCE = 0.13f;
+
+    private final long warmupMs;
+    private final long progressMs;
+
+    private boolean tracking;
+    private int trackId;
+    private long stableSince;
+    private float anchorX;
+    private float anchorY;
+    private float anchorWidth;
+    private float anchorHeight;
+    private float previousX;
+    private float previousY;
+    private float previousWidth;
+    private float previousHeight;
+
+    FaceStabilityGate() {
+        this(WARMUP_MS, PROGRESS_MS);
+    }
+
+    FaceStabilityGate(long warmupMs, long progressMs) {
+        this.warmupMs = Math.max(0L, warmupMs);
+        this.progressMs = Math.max(1L, progressMs);
+    }
+
+    /** Returns -1 during the configured warm-up, otherwise progress in [0, 1]. */
+    float update(int id, float left, float top, float right, float bottom, long now) {
+        float width = Math.max(1f, right - left);
+        float height = Math.max(1f, bottom - top);
+        float centerX = (left + right) * 0.5f;
+        float centerY = (top + bottom) * 0.5f;
+        if (!tracking || trackId != id || now < stableSince
+                || !isStable(centerX, centerY, width, height)) {
+            begin(id, centerX, centerY, width, height, now);
+            return -1f;
+        }
+        previousX = centerX;
+        previousY = centerY;
+        previousWidth = width;
+        previousHeight = height;
+        long elapsed = now - stableSince;
+        if (elapsed < warmupMs) {
+            return -1f;
+        }
+        return Math.min(1f, (elapsed - warmupMs) / (float) progressMs);
+    }
+
+    void reset() {
+        tracking = false;
+        stableSince = 0L;
+    }
+
+    private void begin(int id, float centerX, float centerY,
+                       float width, float height, long now) {
+        tracking = true;
+        trackId = id;
+        stableSince = now;
+        anchorX = previousX = centerX;
+        anchorY = previousY = centerY;
+        anchorWidth = previousWidth = width;
+        anchorHeight = previousHeight = height;
+    }
+
+    private boolean isStable(float centerX, float centerY, float width, float height) {
+        return centerDelta(centerX, centerY,
+                previousX, previousY, previousWidth, previousHeight)
+                <= FRAME_CENTER_TOLERANCE
+                && sizeDelta(width, height, previousWidth, previousHeight)
+                <= FRAME_SIZE_TOLERANCE
+                && centerDelta(centerX, centerY,
+                anchorX, anchorY, anchorWidth, anchorHeight)
+                <= ANCHOR_CENTER_TOLERANCE
+                && sizeDelta(width, height, anchorWidth, anchorHeight)
+                <= ANCHOR_SIZE_TOLERANCE;
+    }
+
+    private static float centerDelta(float x, float y, float referenceX, float referenceY,
+                                     float referenceWidth, float referenceHeight) {
+        float scale = Math.max(1f, Math.max(referenceWidth, referenceHeight));
+        return (float) Math.hypot(x - referenceX, y - referenceY) / scale;
+    }
+
+    private static float sizeDelta(float width, float height,
+                                   float referenceWidth, float referenceHeight) {
+        return Math.max(Math.abs(width - referenceWidth) / Math.max(1f, referenceWidth),
+                Math.abs(height - referenceHeight) / Math.max(1f, referenceHeight));
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingAnalyzer.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingAnalyzer.java
new file mode 100644
index 000000000..bfc939821
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingAnalyzer.java
@@ -0,0 +1,247 @@
+package com.example.inspireface_example.view;
+
+import android.graphics.Color;
+import android.os.SystemClock;
+import android.util.SparseIntArray;
+
+import androidx.annotation.Nullable;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Point2f;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.util.Arrays;
+
+/** Runs configurable multi-face tracking and submits one batched GL frame per camera frame. */
+public final class FaceTrackingAnalyzer extends UprightFaceCameraAnalyzer {
+
+    public interface Listener {
+        void onSessionReady();
+
+        /** Throttled callback on the camera analysis executor. */
+        void onStats(double fps, long latencyMs);
+
+        void onSessionError();
+    }
+
+    private static final long REPORT_INTERVAL_MS = 300L;
+    private static final int[] TRACK_COLORS = {
+            0xFF00E5A0, 0xFFFFC400, 0xFF40C4FF, 0xFFFF6E9C, 0xFFC6A5FF,
+            0xFFFF8A65, 0xFF76FF03, 0xFF18FFFF, 0xFFFFD180, 0xFFEA80FC
+    };
+
+    private final FaceTrackingGlView overlay;
+    private final int detectMode;
+    private final int detectPixelLevel;
+    private final int maxFaces;
+    private final int minimumFacePixelSize;
+    private final Listener listener;
+    private final float[] pointVertices = new float[
+            FaceTrackingGlView.MAX_FACES * FaceTrackingGlView.LANDMARKS_PER_FACE
+                    * FaceTrackingGlView.FLOATS_PER_VERTEX];
+    private final float[] boxVertices = new float[
+            FaceTrackingGlView.MAX_FACES * FaceTrackingGlView.BOX_VERTICES_PER_FACE
+                    * FaceTrackingGlView.FLOATS_PER_VERTEX];
+    private final SparseIntArray colorSlotByTrackId = new SparseIntArray();
+    private final int[] currentTrackIds = new int[FaceTrackingGlView.MAX_FACES];
+    private final boolean[] usedColorSlots = new boolean[TRACK_COLORS.length];
+
+    private volatile boolean mirrored = true;
+    private long fpsWindowStart;
+    private int fpsWindowFrames;
+    private double fps;
+    private double latencyEmaMs;
+    private long lastReport;
+
+    public FaceTrackingAnalyzer(FaceTrackingGlView overlay,
+                                int detectMode, int detectPixelLevel,
+                                int maxFaces, int minimumFacePixelSize,
+                                Listener listener) {
+        this.overlay = overlay;
+        this.detectMode = detectMode;
+        this.detectPixelLevel = detectPixelLevel;
+        this.maxFaces = maxFaces;
+        this.minimumFacePixelSize = minimumFacePixelSize;
+        this.listener = listener;
+    }
+
+    @Override
+    protected Session createSession() {
+        return FaceEngine.createFaceTrackingSession(
+                detectMode, detectPixelLevel, maxFaces, minimumFacePixelSize);
+    }
+
+    @Override
+    protected void onSessionReady() {
+        listener.onSessionReady();
+    }
+
+    @Override
+    protected void onFaces(Session session, ImageStream stream,
+                           @Nullable MultipleFaceData faces, byte[] uprightNv21,
+                           int uprightWidth, int uprightHeight, long frameStart) {
+        int detected = faces == null ? 0
+                : Math.min(faces.detectedNum, FaceTrackingGlView.MAX_FACES);
+        if (detected <= 0) {
+            overlay.clearTracking();
+            reportStats(frameStart);
+            return;
+        }
+
+        int pointCount = 0;
+        int boxCount = 0;
+        updateTrackColors(faces, detected);
+        for (int i = 0; i < detected; i++) {
+            int color = TRACK_COLORS[colorSlotByTrackId.get(currentTrackIds[i])];
+            Point2f[] landmarks = faces.tokens != null && i < faces.tokens.length
+                    && faces.tokens[i] != null
+                    ? InspireFace.GetFaceDenseLandmarkFromFaceToken(faces.tokens[i]) : null;
+            if (faces.rects != null && i < faces.rects.length && faces.rects[i] != null) {
+                boxCount = appendBrackets(
+                        boxVertices, boxCount, faces.rects[i], color);
+            }
+            if (landmarks == null || landmarks.length == 0) {
+                continue;
+            }
+            for (Point2f point : landmarks) {
+                if (point == null || pointCount >= FaceTrackingGlView.MAX_FACES
+                        * FaceTrackingGlView.LANDMARKS_PER_FACE) {
+                    continue;
+                }
+                appendVertex(pointVertices, pointCount++, point.x, point.y, color);
+            }
+        }
+
+        overlay.submit(pointVertices, pointCount, boxVertices, boxCount,
+                uprightWidth, uprightHeight, mirrored);
+        reportStats(frameStart);
+    }
+
+    @Override
+    protected void onSessionError() {
+        overlay.clearTracking();
+        listener.onSessionError();
+    }
+
+    public void setMirrored(boolean mirrored) {
+        this.mirrored = mirrored;
+        overlay.clearTracking();
+    }
+
+    public void clearTracking() {
+        overlay.clearTracking();
+    }
+
+    private void reportStats(long frameStart) {
+        long now = SystemClock.elapsedRealtime();
+        long latency = Math.max(0L, now - frameStart);
+        latencyEmaMs = latencyEmaMs == 0.0
+                ? latency : latencyEmaMs * 0.85 + latency * 0.15;
+        fpsWindowFrames++;
+        if (fpsWindowStart == 0L) {
+            fpsWindowStart = now;
+        } else if (now - fpsWindowStart >= 1_000L) {
+            fps = fpsWindowFrames * 1_000.0 / (now - fpsWindowStart);
+            fpsWindowFrames = 0;
+            fpsWindowStart = now;
+        }
+        if (now - lastReport >= REPORT_INTERVAL_MS) {
+            lastReport = now;
+            listener.onStats(fps, Math.round(latencyEmaMs));
+        }
+    }
+
+    /** Keeps active IDs on distinct palette slots while preserving each surviving ID's color. */
+    private void updateTrackColors(MultipleFaceData faces, int detected) {
+        for (int i = 0; i < detected; i++) {
+            int sdkTrackId = faces.trackIds != null && i < faces.trackIds.length
+                    ? faces.trackIds[i] : -1;
+            // A missing SDK ID still needs a unique key for this frame.
+            currentTrackIds[i] = sdkTrackId >= 0 ? sdkTrackId : Integer.MIN_VALUE + i;
+        }
+        for (int i = colorSlotByTrackId.size() - 1; i >= 0; i--) {
+            if (!containsTrackId(currentTrackIds, detected, colorSlotByTrackId.keyAt(i))) {
+                colorSlotByTrackId.removeAt(i);
+            }
+        }
+        Arrays.fill(usedColorSlots, false);
+        for (int i = 0; i < colorSlotByTrackId.size(); i++) {
+            usedColorSlots[colorSlotByTrackId.valueAt(i)] = true;
+        }
+        for (int i = 0; i < detected; i++) {
+            int trackId = currentTrackIds[i];
+            if (colorSlotByTrackId.indexOfKey(trackId) >= 0) {
+                continue;
+            }
+            int slot = firstFreeColorSlot();
+            colorSlotByTrackId.put(trackId, slot);
+            usedColorSlots[slot] = true;
+        }
+    }
+
+    private int firstFreeColorSlot() {
+        for (int i = 0; i < usedColorSlots.length; i++) {
+            if (!usedColorSlots[i]) {
+                return i;
+            }
+        }
+        return 0;
+    }
+
+    private static boolean containsTrackId(int[] ids, int count, int wanted) {
+        for (int i = 0; i < count; i++) {
+            if (ids[i] == wanted) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    private static int appendBrackets(float[] target, int vertexCount,
+                                      FaceRect rect, int color) {
+        float left = rect.x;
+        float top = rect.y;
+        float right = rect.x + rect.width;
+        float bottom = rect.y + rect.height;
+        float length = Math.min(rect.width, rect.height) * 0.22f;
+
+        vertexCount = appendLine(target, vertexCount,
+                left, top + length, left, top, color);
+        vertexCount = appendLine(target, vertexCount,
+                left, top, left + length, top, color);
+        vertexCount = appendLine(target, vertexCount,
+                right - length, top, right, top, color);
+        vertexCount = appendLine(target, vertexCount,
+                right, top, right, top + length, color);
+        vertexCount = appendLine(target, vertexCount,
+                left, bottom - length, left, bottom, color);
+        vertexCount = appendLine(target, vertexCount,
+                left, bottom, left + length, bottom, color);
+        vertexCount = appendLine(target, vertexCount,
+                right - length, bottom, right, bottom, color);
+        vertexCount = appendLine(target, vertexCount,
+                right, bottom, right, bottom - length, color);
+        return vertexCount;
+    }
+
+    private static int appendLine(float[] target, int vertexCount,
+                                  float x0, float y0, float x1, float y1, int color) {
+        appendVertex(target, vertexCount++, x0, y0, color);
+        appendVertex(target, vertexCount++, x1, y1, color);
+        return vertexCount;
+    }
+
+    private static void appendVertex(float[] target, int vertexIndex,
+                                     float x, float y, int color) {
+        int offset = vertexIndex * FaceTrackingGlView.FLOATS_PER_VERTEX;
+        target[offset] = x;
+        target[offset + 1] = y;
+        target[offset + 2] = Color.red(color) / 255f;
+        target[offset + 3] = Color.green(color) / 255f;
+        target[offset + 4] = Color.blue(color) / 255f;
+        target[offset + 5] = Color.alpha(color) / 255f;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingGlView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingGlView.java
new file mode 100644
index 000000000..4a76a1794
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/FaceTrackingGlView.java
@@ -0,0 +1,264 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.PixelFormat;
+import android.opengl.GLES20;
+import android.opengl.GLSurfaceView;
+import android.util.AttributeSet;
+
+import androidx.annotation.Nullable;
+
+import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
+import java.nio.FloatBuffer;
+
+import javax.microedition.khronos.egl.EGLConfig;
+import javax.microedition.khronos.opengles.GL10;
+
+/**
+ * Transparent OpenGL overlay for multi-face tracking. Colored boxes and dense landmarks are
+ * uploaded in two batches, then rendered on GLSurfaceView's own thread only when a new tracking
+ * result arrives. Every vertex carries its track-ID color, keeping UI work off the main thread.
+ */
+public final class FaceTrackingGlView extends GLSurfaceView {
+
+    /** x, y, red, green, blue, alpha. */
+    static final int FLOATS_PER_VERTEX = 6;
+    static final int MAX_FACES = 10;
+    static final int LANDMARKS_PER_FACE = 106;
+    static final int BOX_VERTICES_PER_FACE = 16;
+
+    private final TrackingRenderer renderer;
+
+    public FaceTrackingGlView(Context context) {
+        this(context, null);
+    }
+
+    public FaceTrackingGlView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        float density = getResources().getDisplayMetrics().density;
+        renderer = new TrackingRenderer(4.5f * density, 2.2f * density);
+        setEGLContextClientVersion(2);
+        setEGLConfigChooser(8, 8, 8, 8, 0, 0);
+        getHolder().setFormat(PixelFormat.TRANSLUCENT);
+        setZOrderMediaOverlay(true);
+        setRenderer(renderer);
+        setRenderMode(RENDERMODE_WHEN_DIRTY);
+        setImportantForAccessibility(IMPORTANT_FOR_ACCESSIBILITY_NO);
+    }
+
+    /** Safe from the CameraX analysis thread; data is copied before returning. */
+    void submit(float[] pointVertices, int pointCount,
+                float[] boxVertices, int boxVertexCount,
+                int imageWidth, int imageHeight, boolean mirrored) {
+        renderer.setFrame(pointVertices, pointCount, boxVertices, boxVertexCount,
+                imageWidth, imageHeight, mirrored);
+        requestRender();
+    }
+
+    /** Safe from any thread. */
+    public void clearTracking() {
+        renderer.setFrame(null, 0, null, 0, 0, 0, false);
+        requestRender();
+    }
+
+    private static final class TrackingRenderer implements Renderer {
+
+        private static final int MAX_POINT_VERTICES = MAX_FACES * LANDMARKS_PER_FACE;
+        private static final int MAX_BOX_VERTICES = MAX_FACES * BOX_VERTICES_PER_FACE;
+        private static final int STRIDE_BYTES = FLOATS_PER_VERTEX * 4;
+
+        private static final String VERTEX_SHADER = ""
+                + "attribute vec2 aPos;\n"
+                + "attribute vec4 aColor;\n"
+                + "uniform vec4 uXform;\n"
+                + "uniform float uPointSize;\n"
+                + "varying vec4 vColor;\n"
+                + "void main() {\n"
+                + "  gl_Position = vec4(aPos.x * uXform.x + uXform.y,\n"
+                + "                     aPos.y * uXform.z + uXform.w, 0.0, 1.0);\n"
+                + "  gl_PointSize = uPointSize;\n"
+                + "  vColor = aColor;\n"
+                + "}\n";
+
+        private static final String FRAGMENT_SHADER = ""
+                + "precision mediump float;\n"
+                + "varying vec4 vColor;\n"
+                + "uniform float uPointPass;\n"
+                + "void main() {\n"
+                + "  float a = vColor.a;\n"
+                + "  if (uPointPass > 0.5) {\n"
+                + "    float dist = length(gl_PointCoord - vec2(0.5));\n"
+                + "    a *= 1.0 - smoothstep(0.35, 0.5, dist);\n"
+                + "    if (a <= 0.01) discard;\n"
+                + "  }\n"
+                + "  gl_FragColor = vec4(vColor.rgb * a, a);\n"
+                + "}\n";
+
+        private final Object lock = new Object();
+        private final float[] stagedPoints =
+                new float[MAX_POINT_VERTICES * FLOATS_PER_VERTEX];
+        private final float[] stagedBoxes =
+                new float[MAX_BOX_VERTICES * FLOATS_PER_VERTEX];
+        private int stagedPointCount;
+        private int stagedBoxCount;
+        private int imageWidth;
+        private int imageHeight;
+        private boolean mirrored;
+
+        private final FloatBuffer pointBuffer = allocate(MAX_POINT_VERTICES);
+        private final FloatBuffer boxBuffer = allocate(MAX_BOX_VERTICES);
+        private final float pointSizePx;
+        private final float boxWidthPx;
+
+        private int program;
+        private int aPosLoc;
+        private int aColorLoc;
+        private int uXformLoc;
+        private int uPointSizeLoc;
+        private int uPointPassLoc;
+        private int viewWidth;
+        private int viewHeight;
+
+        TrackingRenderer(float pointSizePx, float boxWidthPx) {
+            this.pointSizePx = pointSizePx;
+            this.boxWidthPx = boxWidthPx;
+        }
+
+        void setFrame(float[] points, int pointCount,
+                      float[] boxes, int boxCount,
+                      int imageWidth, int imageHeight, boolean mirrored) {
+            synchronized (lock) {
+                stagedPointCount = points == null ? 0
+                        : Math.min(pointCount, MAX_POINT_VERTICES);
+                stagedBoxCount = boxes == null ? 0
+                        : Math.min(boxCount, MAX_BOX_VERTICES);
+                if (stagedPointCount > 0) {
+                    System.arraycopy(points, 0, stagedPoints, 0,
+                            stagedPointCount * FLOATS_PER_VERTEX);
+                }
+                if (stagedBoxCount > 0) {
+                    System.arraycopy(boxes, 0, stagedBoxes, 0,
+                            stagedBoxCount * FLOATS_PER_VERTEX);
+                }
+                this.imageWidth = imageWidth;
+                this.imageHeight = imageHeight;
+                this.mirrored = mirrored;
+            }
+        }
+
+        @Override
+        public void onSurfaceCreated(GL10 gl, EGLConfig config) {
+            synchronized (lock) {
+                stagedPointCount = 0;
+                stagedBoxCount = 0;
+            }
+            program = GLES20.glCreateProgram();
+            GLES20.glAttachShader(program,
+                    compileShader(GLES20.GL_VERTEX_SHADER, VERTEX_SHADER));
+            GLES20.glAttachShader(program,
+                    compileShader(GLES20.GL_FRAGMENT_SHADER, FRAGMENT_SHADER));
+            GLES20.glLinkProgram(program);
+            aPosLoc = GLES20.glGetAttribLocation(program, "aPos");
+            aColorLoc = GLES20.glGetAttribLocation(program, "aColor");
+            uXformLoc = GLES20.glGetUniformLocation(program, "uXform");
+            uPointSizeLoc = GLES20.glGetUniformLocation(program, "uPointSize");
+            uPointPassLoc = GLES20.glGetUniformLocation(program, "uPointPass");
+            GLES20.glClearColor(0f, 0f, 0f, 0f);
+            GLES20.glEnable(GLES20.GL_BLEND);
+            GLES20.glBlendFunc(GLES20.GL_ONE, GLES20.GL_ONE_MINUS_SRC_ALPHA);
+        }
+
+        @Override
+        public void onSurfaceChanged(GL10 gl, int width, int height) {
+            GLES20.glViewport(0, 0, width, height);
+            viewWidth = width;
+            viewHeight = height;
+        }
+
+        @Override
+        public void onDrawFrame(GL10 gl) {
+            GLES20.glClear(GLES20.GL_COLOR_BUFFER_BIT);
+
+            int pointCount;
+            int boxCount;
+            float ax;
+            float bx;
+            float ay;
+            float by;
+            synchronized (lock) {
+                pointCount = stagedPointCount;
+                boxCount = stagedBoxCount;
+                if ((pointCount == 0 && boxCount == 0)
+                        || imageWidth <= 0 || imageHeight <= 0
+                        || viewWidth <= 0 || viewHeight <= 0) {
+                    return;
+                }
+                put(pointBuffer, stagedPoints, pointCount);
+                put(boxBuffer, stagedBoxes, boxCount);
+
+                // PreviewView defaults to FILL_CENTER; mirror only the front-camera x axis.
+                float scale = Math.max((float) viewWidth / imageWidth,
+                        (float) viewHeight / imageHeight);
+                float dx = (viewWidth - imageWidth * scale) / 2f;
+                float dy = (viewHeight - imageHeight * scale) / 2f;
+                float mx = mirrored ? -scale : scale;
+                float cx = mirrored ? imageWidth * scale + dx : dx;
+                ax = 2f * mx / viewWidth;
+                bx = 2f * cx / viewWidth - 1f;
+                ay = -2f * scale / viewHeight;
+                by = 1f - 2f * dy / viewHeight;
+            }
+
+            GLES20.glUseProgram(program);
+            GLES20.glUniform4f(uXformLoc, ax, bx, ay, by);
+            GLES20.glUniform1f(uPointSizeLoc, pointSizePx);
+            GLES20.glEnableVertexAttribArray(aPosLoc);
+            GLES20.glEnableVertexAttribArray(aColorLoc);
+
+            if (boxCount > 0) {
+                bind(boxBuffer);
+                GLES20.glUniform1f(uPointPassLoc, 0f);
+                GLES20.glLineWidth(boxWidthPx);
+                GLES20.glDrawArrays(GLES20.GL_LINES, 0, boxCount);
+            }
+            if (pointCount > 0) {
+                bind(pointBuffer);
+                GLES20.glUniform1f(uPointPassLoc, 1f);
+                GLES20.glDrawArrays(GLES20.GL_POINTS, 0, pointCount);
+            }
+
+            GLES20.glDisableVertexAttribArray(aColorLoc);
+            GLES20.glDisableVertexAttribArray(aPosLoc);
+        }
+
+        private void bind(FloatBuffer buffer) {
+            buffer.position(0);
+            GLES20.glVertexAttribPointer(aPosLoc, 2, GLES20.GL_FLOAT,
+                    false, STRIDE_BYTES, buffer);
+            buffer.position(2);
+            GLES20.glVertexAttribPointer(aColorLoc, 4, GLES20.GL_FLOAT,
+                    false, STRIDE_BYTES, buffer);
+        }
+
+        private static FloatBuffer allocate(int vertexCount) {
+            return ByteBuffer.allocateDirect(
+                            vertexCount * FLOATS_PER_VERTEX * 4)
+                    .order(ByteOrder.nativeOrder())
+                    .asFloatBuffer();
+        }
+
+        private static void put(FloatBuffer buffer, float[] source, int vertexCount) {
+            buffer.position(0);
+            buffer.put(source, 0, vertexCount * FLOATS_PER_VERTEX);
+            buffer.position(0);
+        }
+
+        private static int compileShader(int type, String source) {
+            int shader = GLES20.glCreateShader(type);
+            GLES20.glShaderSource(shader, source);
+            GLES20.glCompileShader(shader);
+            return shader;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LandmarkGlView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LandmarkGlView.java
new file mode 100644
index 000000000..ce0e1c59d
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LandmarkGlView.java
@@ -0,0 +1,213 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.Color;
+import android.graphics.PixelFormat;
+import android.opengl.GLES20;
+import android.opengl.GLSurfaceView;
+import android.util.AttributeSet;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+
+import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
+import java.nio.FloatBuffer;
+
+import javax.microedition.khronos.egl.EGLConfig;
+import javax.microedition.khronos.opengles.GL10;
+
+/**
+ * Transparent GL overlay that renders the dense facial landmarks as one GL_POINTS draw
+ * call: a single ~1 KB vertex upload per frame, the image→screen mapping folded into a
+ * vertex-shader uniform, and round anti-aliased sprites from gl_PointCoord. Renders only
+ * when a new frame of points arrives (RENDERMODE_WHEN_DIRTY); hidden, the surface is
+ * destroyed and costs nothing.
+ *
+ * Points are supplied in the upright image space the SDK reports (same space as the face
+ * rects) and mapped with the same FILL_CENTER + front-mirror convention as
+ * {@link FaceOverlayView}.
+ */
+public class LandmarkGlView extends GLSurfaceView {
+
+    private final PointsRenderer renderer;
+
+    public LandmarkGlView(Context context) {
+        this(context, null);
+    }
+
+    public LandmarkGlView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        renderer = new PointsRenderer(
+                4f * getResources().getDisplayMetrics().density,
+                ContextCompat.getColor(context, R.color.liveness_accent));
+        setEGLContextClientVersion(2);
+        setEGLConfigChooser(8, 8, 8, 8, 0, 0);
+        getHolder().setFormat(PixelFormat.TRANSLUCENT);
+        setZOrderMediaOverlay(true); // above the camera surface, below regular views
+        setRenderer(renderer);
+        setRenderMode(RENDERMODE_WHEN_DIRTY);
+    }
+
+    /**
+     * Safe to call from any thread. {@code xy} holds {@code count} interleaved (x, y)
+     * pairs in upright image space; the array is copied before returning.
+     */
+    void submitPoints(float[] xy, int count, int imageWidth, int imageHeight, boolean mirrored) {
+        renderer.setPoints(xy, count, imageWidth, imageHeight, mirrored);
+        requestRender();
+    }
+
+    /** Safe to call from any thread. */
+    void clearPoints() {
+        renderer.setPoints(null, 0, 0, 0, false);
+        requestRender();
+    }
+
+    private static final class PointsRenderer implements Renderer {
+
+        private static final int MAX_POINTS = 512;
+
+        private static final String VERTEX_SHADER = ""
+                + "attribute vec2 aPos;\n"
+                + "uniform vec4 uXform;\n" // (scaleX, offsetX, scaleY, offsetY) image → NDC
+                + "uniform float uPointSize;\n"
+                + "void main() {\n"
+                + "  gl_Position = vec4(aPos.x * uXform.x + uXform.y,\n"
+                + "                     aPos.y * uXform.z + uXform.w, 0.0, 1.0);\n"
+                + "  gl_PointSize = uPointSize;\n"
+                + "}\n";
+
+        // Premultiplied output: the compositor treats a TRANSLUCENT surface as
+        // premultiplied, so straight alpha would leave dst alpha = a² and draw an
+        // over-bright halo on the anti-aliased fringe.
+        private static final String FRAGMENT_SHADER = ""
+                + "precision mediump float;\n"
+                + "uniform vec4 uColor;\n"
+                + "void main() {\n"
+                + "  float dist = length(gl_PointCoord - vec2(0.5));\n"
+                + "  float a = uColor.a * (1.0 - smoothstep(0.35, 0.5, dist));\n"
+                + "  if (a <= 0.01) discard;\n"
+                + "  gl_FragColor = vec4(uColor.rgb * a, a);\n"
+                + "}\n";
+
+        private final Object lock = new Object();
+        private final float[] staging = new float[MAX_POINTS * 2];
+        private int stagingCount;
+        private int imageWidth;
+        private int imageHeight;
+        private boolean mirrored;
+
+        private final FloatBuffer vertexBuffer;
+        private final float pointSizePx;
+        private final float[] color = new float[4];
+
+        private int program;
+        private int aPosLoc;
+        private int uXformLoc;
+        private int uPointSizeLoc;
+        private int uColorLoc;
+        private int viewWidth;
+        private int viewHeight;
+
+        PointsRenderer(float pointSizePx, int argbColor) {
+            this.pointSizePx = pointSizePx;
+            color[0] = Color.red(argbColor) / 255f;
+            color[1] = Color.green(argbColor) / 255f;
+            color[2] = Color.blue(argbColor) / 255f;
+            color[3] = Color.alpha(argbColor) / 255f;
+            vertexBuffer = ByteBuffer.allocateDirect(MAX_POINTS * 2 * 4)
+                    .order(ByteOrder.nativeOrder())
+                    .asFloatBuffer();
+        }
+
+        void setPoints(float[] xy, int count, int imageWidth, int imageHeight, boolean mirrored) {
+            synchronized (lock) {
+                stagingCount = xy == null ? 0 : Math.min(count, MAX_POINTS);
+                if (stagingCount > 0) {
+                    System.arraycopy(xy, 0, staging, 0, stagingCount * 2);
+                }
+                this.imageWidth = imageWidth;
+                this.imageHeight = imageHeight;
+                this.mirrored = mirrored;
+            }
+        }
+
+        @Override
+        public void onSurfaceCreated(GL10 gl, EGLConfig config) {
+            // A recreated surface must start empty: GLSurfaceView always draws one frame
+            // after creation, which would otherwise flash the pre-destruction points.
+            synchronized (lock) {
+                stagingCount = 0;
+            }
+            // The context is recreated whenever the view is shown again — rebuild everything.
+            program = GLES20.glCreateProgram();
+            GLES20.glAttachShader(program, compileShader(GLES20.GL_VERTEX_SHADER, VERTEX_SHADER));
+            GLES20.glAttachShader(program, compileShader(GLES20.GL_FRAGMENT_SHADER, FRAGMENT_SHADER));
+            GLES20.glLinkProgram(program);
+            aPosLoc = GLES20.glGetAttribLocation(program, "aPos");
+            uXformLoc = GLES20.glGetUniformLocation(program, "uXform");
+            uPointSizeLoc = GLES20.glGetUniformLocation(program, "uPointSize");
+            uColorLoc = GLES20.glGetUniformLocation(program, "uColor");
+            GLES20.glClearColor(0f, 0f, 0f, 0f);
+            GLES20.glEnable(GLES20.GL_BLEND);
+            GLES20.glBlendFunc(GLES20.GL_ONE, GLES20.GL_ONE_MINUS_SRC_ALPHA);
+        }
+
+        @Override
+        public void onSurfaceChanged(GL10 gl, int width, int height) {
+            GLES20.glViewport(0, 0, width, height);
+            viewWidth = width;
+            viewHeight = height;
+        }
+
+        @Override
+        public void onDrawFrame(GL10 gl) {
+            GLES20.glClear(GLES20.GL_COLOR_BUFFER_BIT);
+
+            int count;
+            float ax;
+            float bx;
+            float ay;
+            float by;
+            synchronized (lock) {
+                count = stagingCount;
+                if (count == 0 || imageWidth <= 0 || imageHeight <= 0
+                        || viewWidth <= 0 || viewHeight <= 0) {
+                    return;
+                }
+                vertexBuffer.position(0);
+                vertexBuffer.put(staging, 0, count * 2);
+                vertexBuffer.position(0);
+                // Same FILL_CENTER mapping as FaceOverlayView, folded into x' = ax*x + bx.
+                float scale = Math.max((float) viewWidth / imageWidth, (float) viewHeight / imageHeight);
+                float dx = (viewWidth - imageWidth * scale) / 2f;
+                float dy = (viewHeight - imageHeight * scale) / 2f;
+                float mx = mirrored ? -scale : scale;
+                float cx = mirrored ? imageWidth * scale + dx : dx;
+                ax = 2f * mx / viewWidth;
+                bx = 2f * cx / viewWidth - 1f;
+                ay = -2f * scale / viewHeight;
+                by = 1f - 2f * dy / viewHeight;
+            }
+
+            GLES20.glUseProgram(program);
+            GLES20.glEnableVertexAttribArray(aPosLoc);
+            GLES20.glVertexAttribPointer(aPosLoc, 2, GLES20.GL_FLOAT, false, 0, vertexBuffer);
+            GLES20.glUniform4f(uXformLoc, ax, bx, ay, by);
+            GLES20.glUniform1f(uPointSizeLoc, pointSizePx);
+            GLES20.glUniform4f(uColorLoc, color[0], color[1], color[2], color[3]);
+            GLES20.glDrawArrays(GLES20.GL_POINTS, 0, count);
+            GLES20.glDisableVertexAttribArray(aPosLoc);
+        }
+
+        private static int compileShader(int type, String source) {
+            int shader = GLES20.glCreateShader(type);
+            GLES20.glShaderSource(shader, source);
+            GLES20.glCompileShader(shader);
+            return shader;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessActivity.java
new file mode 100644
index 000000000..d5b87191d
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessActivity.java
@@ -0,0 +1,298 @@
+package com.example.inspireface_example.view;
+
+import android.os.Bundle;
+import android.view.View;
+import android.view.ViewGroup;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.camera.view.PreviewView;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.example.inspireface_example.FaceModelPrefs;
+import com.example.inspireface_example.LocalePrefs;
+import com.example.inspireface_example.R;
+import com.example.inspireface_example.permission.CameraPermissionCoordinator;
+import com.google.android.material.card.MaterialCardView;
+import com.google.android.material.progressindicator.LinearProgressIndicator;
+import com.google.android.material.switchmaterial.SwitchMaterial;
+import com.insightface.sdk.inspireface.base.FaceEulerAngle;
+
+import java.util.Locale;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/**
+ * Shared camera screen used by the three dedicated feature activities. Subclasses choose
+ * one fixed controller mode while this class owns the common CameraX and overlay lifecycle.
+ */
+public class LivenessActivity extends AppCompatActivity implements FaceAnalyzer.Listener {
+
+    private PreviewView previewView;
+    private FaceOverlayView overlayView;
+    private LandmarkGlView landmarkView;
+    private TextView promptTitle;
+    private TextView promptSub;
+    private TextView perfText;
+    private TextView eulerText;
+    private SwitchMaterial switchEuler;
+    private LinearProgressIndicator promptProgress;
+    private View btnRestart;
+
+    private final ExecutorService analysisExecutor = Executors.newSingleThreadExecutor();
+    private LivenessController controller;
+    private FaceAnalyzer analyzer;
+    private LivenessController.UiState lastState;
+    private CameraPreviewController cameraController;
+    private CameraPermissionCoordinator cameraPermission;
+    private boolean engineStartingOrReady;
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        cameraPermission = new CameraPermissionCoordinator(this,
+                new CameraPermissionCoordinator.Listener() {
+                    @Override
+                    public void onCameraPermissionGranted() {
+                        startEngine();
+                    }
+
+                    @Override
+                    public void onCameraPermissionBlocked(boolean requiresSettings) {
+                        showCameraPermissionBlocked(requiresSettings);
+                    }
+                });
+        WindowCompat.setDecorFitsSystemWindows(getWindow(), false);
+        setContentView(R.layout.activity_liveness);
+
+        previewView = findViewById(R.id.previewView);
+        overlayView = findViewById(R.id.faceOverlay);
+        landmarkView = findViewById(R.id.landmarkGlView);
+        promptTitle = findViewById(R.id.promptTitle);
+        promptSub = findViewById(R.id.promptSub);
+        perfText = findViewById(R.id.perfText);
+        eulerText = findViewById(R.id.eulerText);
+        switchEuler = findViewById(R.id.switchEuler);
+        promptProgress = findViewById(R.id.promptProgress);
+        btnRestart = findViewById(R.id.btnRestart);
+        cameraPermission.bindRecoveryButton(
+                findViewById(R.id.btnCameraPermissionAction));
+
+        ((TextView) findViewById(R.id.pageTitle)).setText(pageTitleRes());
+        ((TextView) findViewById(R.id.currentModel)).setText(
+                getString(R.string.current_model, FaceModelPrefs.get(this).sdkName()));
+        findViewById(R.id.btnBack).setOnClickListener(v -> finish());
+
+        switchEuler.setOnCheckedChangeListener((button, checked) -> {
+            if (analyzer != null) {
+                analyzer.setEulerEnabled(checked);
+            }
+            eulerText.setText(R.string.euler_no_face);
+            eulerText.setVisibility(checked ? View.VISIBLE : View.GONE);
+        });
+        applyWindowInsets();
+
+        controller = new LivenessController(this);
+        controller.setMode(initialMode());
+        btnRestart.setOnClickListener(v -> controller.restart());
+        // In-app language toggle (bottom-right): English by default, Chinese on demand.
+        findViewById(R.id.langSwitch).setOnClickListener(v -> LocalePrefs.toggle(this));
+        findViewById(R.id.btnFlipCamera).setOnClickListener(v -> flipCamera());
+
+        cameraPermission.requestAccess();
+    }
+
+    private void applyWindowInsets() {
+        View topBar = findViewById(R.id.topBar);
+        MaterialCardView promptCard = findViewById(R.id.promptCard);
+        int cardBaseMargin = (int) (24 * getResources().getDisplayMetrics().density);
+        ViewCompat.setOnApplyWindowInsetsListener(findViewById(R.id.livenessRoot), (v, insets) -> {
+            Insets bars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            topBar.setPadding(topBar.getPaddingLeft(), bars.top,
+                    topBar.getPaddingRight(), topBar.getPaddingBottom());
+            ViewGroup.MarginLayoutParams lp =
+                    (ViewGroup.MarginLayoutParams) promptCard.getLayoutParams();
+            lp.bottomMargin = cardBaseMargin + bars.bottom;
+            promptCard.setLayoutParams(lp);
+            return insets;
+        });
+    }
+
+    /** GlobalLaunch copies model assets on first run — keep it off the main thread. */
+    private void startEngine() {
+        if (engineStartingOrReady) {
+            return;
+        }
+        engineStartingOrReady = true;
+        promptTitle.setText(R.string.msg_initializing);
+        promptSub.setVisibility(View.GONE);
+        analysisExecutor.execute(() -> {
+            boolean ok = FaceEngine.ensureLaunched(this);
+            runOnUiThread(() -> {
+                if (isDestroyed() || isFinishing()) {
+                    return; // model copy can outlive the activity — don't bind a dead lifecycle
+                }
+                if (ok) {
+                    bindCamera();
+                } else {
+                    engineStartingOrReady = false;
+                    promptTitle.setText(R.string.msg_engine_failed);
+                }
+            });
+        });
+    }
+
+    private void showCameraPermissionBlocked(boolean requiresSettings) {
+        promptTitle.setText(R.string.msg_permission_required);
+        promptSub.setText(requiresSettings
+                ? R.string.camera_permission_settings_hint
+                : R.string.camera_permission_retry_hint);
+        promptSub.setVisibility(View.VISIBLE);
+    }
+
+    private void bindCamera() {
+        analyzer = new FaceAnalyzer(controller, overlayView, landmarkView, this, true);
+        analyzer.setEulerEnabled(switchEuler.isChecked());
+        analyzer.setLandmarksEnabled(false);
+        cameraController = new CameraPreviewController(
+                this, this, previewView, analysisExecutor, analyzer,
+                new CameraPreviewController.Listener() {
+                    @Override
+                    public void onCameraReady(boolean frontCamera) {
+                        analyzer.setMirrored(frontCamera);
+                    }
+
+                    @Override
+                    public void onLensChanged(boolean frontCamera) {
+                        analyzer.setMirrored(frontCamera);
+                        controller.restart();
+                        overlayView.submit(null);
+                        landmarkView.clearPoints();
+                    }
+
+                    @Override
+                    public void onCameraError(int messageRes) {
+                        promptTitle.setText(messageRes);
+                    }
+                });
+        cameraController.start();
+    }
+
+    /**
+     * Switches between the front and back lens. The SDK needs no per-lens handling:
+     * every frame is pre-rotated by its own rotationDegrees before
+     * CreateImageStreamFromByteBuffer (always CAMERA_ROTATION_0), so the new lens's
+     * different sensor orientation is absorbed per frame. Only the display mirroring
+     * flips, and the mode state restarts so stale tracking can't leak across lenses.
+     */
+    private void flipCamera() {
+        if (cameraController == null || !cameraController.flipCamera()) {
+            Toast.makeText(this, R.string.msg_camera_unavailable, Toast.LENGTH_SHORT).show();
+        }
+    }
+
+    /** Silent liveness is the behavior of the original activity and the first home tile. */
+    protected LivenessController.Mode initialMode() {
+        return LivenessController.Mode.SILENT;
+    }
+
+    protected int pageTitleRes() {
+        return R.string.mode_silent;
+    }
+
+    // ------------------------------------------------------------------
+    // FaceAnalyzer.Listener (analysis thread)
+    // ------------------------------------------------------------------
+
+    @Override
+    public void onUiState(LivenessController.UiState state) {
+        runOnUiThread(() -> {
+            if (state.sameContent(lastState)) {
+                return;
+            }
+            lastState = state;
+            promptTitle.setText(state.title);
+            if (state.subtitle != null) {
+                promptSub.setText(state.subtitle);
+                promptSub.setVisibility(View.VISIBLE);
+            } else {
+                promptSub.setVisibility(View.GONE);
+            }
+            if (state.progress >= 0) {
+                promptProgress.setProgress(state.progress);
+                promptProgress.setVisibility(View.VISIBLE);
+            } else {
+                promptProgress.setVisibility(View.GONE);
+            }
+            btnRestart.setVisibility(state.showRestart ? View.VISIBLE : View.GONE);
+        });
+    }
+
+    @Override
+    public void onPerf(double fps, long latencyMs) {
+        runOnUiThread(() -> {
+            perfText.setVisibility(View.VISIBLE);
+            perfText.setText(String.format(Locale.US,
+                    getString(R.string.perf_format), fps, latencyMs));
+        });
+    }
+
+    @Override
+    public void onEulerAngles(FaceEulerAngle angle) {
+        runOnUiThread(() -> {
+            if (!switchEuler.isChecked()) {
+                return;
+            }
+            if (angle == null) {
+                eulerText.setText(R.string.euler_no_face);
+            } else {
+                eulerText.setText(String.format(Locale.US,
+                        getString(R.string.euler_format), angle.yaw, angle.pitch, angle.roll));
+            }
+        });
+    }
+
+    @Override
+    public void onSessionError() {
+        runOnUiThread(() -> promptTitle.setText(R.string.msg_engine_failed));
+    }
+
+    // GL lifecycle follows visibility (start/stop), not focus (resume/pause): in
+    // multi-window the activity can be paused but visible with the camera and analysis
+    // still running — the landmark overlay must keep rendering there.
+    @Override
+    protected void onResume() {
+        super.onResume();
+        cameraPermission.onResume();
+    }
+
+    @Override
+    protected void onStart() {
+        super.onStart();
+        landmarkView.onResume();
+    }
+
+    @Override
+    protected void onStop() {
+        landmarkView.onPause();
+        super.onStop();
+    }
+
+    @Override
+    protected void onDestroy() {
+        cameraPermission.close();
+        if (cameraController != null) {
+            cameraController.stop();
+        }
+        if (analyzer != null) {
+            FaceAnalyzer toRelease = analyzer;
+            analysisExecutor.execute(toRelease::release);
+        }
+        analysisExecutor.shutdown();
+        super.onDestroy();
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessController.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessController.java
new file mode 100644
index 000000000..635c1fd8b
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/LivenessController.java
@@ -0,0 +1,559 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.Color;
+import android.os.SystemClock;
+import android.text.SpannableStringBuilder;
+import android.text.Spanned;
+import android.text.style.ForegroundColorSpan;
+
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.CustomParameter;
+import com.insightface.sdk.inspireface.base.FaceInteractionsActions;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.RGBLivenessConfidence;
+import com.insightface.sdk.inspireface.base.Session;
+
+import java.util.ArrayDeque;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.Collections;
+import java.util.List;
+import java.util.Random;
+
+/**
+ * Per-frame liveness logic for both modes. All {@link #onFrame} calls must come from the
+ * single analysis thread; mode switches from the UI thread are serialized with
+ * {@code synchronized}.
+ */
+final class LivenessController {
+
+    enum Mode { SILENT, ACTION, POSE }
+
+    /** Cooperative challenges the SDK can recognize. */
+    enum ActionType {
+        BLINK(R.string.action_blink, R.string.name_blink),
+        SHAKE(R.string.action_shake, R.string.name_shake),
+        JAW_OPEN(R.string.action_jaw_open, R.string.name_jaw_open),
+        HEAD_RAISE(R.string.action_head_raise, R.string.name_head_raise);
+
+        final int promptRes;
+        final int nameRes;
+
+        ActionType(int promptRes, int nameRes) {
+            this.promptRes = promptRes;
+            this.nameRes = nameRes;
+        }
+    }
+
+    /** Snapshot the UI renders; immutable, built on the analysis thread. */
+    static final class UiState {
+        final String title;
+        final CharSequence subtitle;   // null hides the row
+        final int progress;            // 0..100, -1 hides the bar
+        final boolean showRestart;
+        final int boxColor;
+
+        UiState(String title, CharSequence subtitle, int progress, boolean showRestart, int boxColor) {
+            this.title = title;
+            this.subtitle = subtitle;
+            this.progress = progress;
+            this.showRestart = showRestart;
+            this.boxColor = boxColor;
+        }
+
+        boolean sameContent(UiState o) {
+            return o != null
+                    && progress == o.progress
+                    && showRestart == o.showRestart
+                    && boxColor == o.boxColor
+                    && title.equals(o.title)
+                    && (subtitle == null ? o.subtitle == null
+                        : o.subtitle != null && subtitle.toString().equals(o.subtitle.toString()));
+        }
+    }
+
+    // ---- Silent-mode tunables ----
+    /** RGB anti-spoofing score above which a face counts as live (SDK author default). */
+    private static final float RGB_LIVENESS_THRESHOLD = 0.88f;
+    /** Frames averaged before giving a verdict; smooths single-frame noise. */
+    private static final int SCORE_WINDOW = 8;
+    private static final int MIN_VERDICT_SAMPLES = 3;
+    /**
+     * Run the anti-spoofing pipeline on every Nth tracked frame. Each call converts the
+     * full frame internally (the SDK's own acknowledged hot spot), so skipping frames
+     * buys preview smoothness at no visible cost given the score averaging.
+     */
+    private static final int SILENT_PIPELINE_INTERVAL = 2;
+
+    // ---- Action-mode tunables ----
+    private static final int ACTIONS_PER_RUN = 3;
+    private static final long ACTION_TIMEOUT_MS = 8_000;
+    /**
+     * Consecutive good frames required before the challenge sequence starts. Must be
+     * >= the SDK's 9-call action warm-up so flags are live for the first challenge.
+     */
+    private static final int STABLE_FRAMES_TO_START = 10;
+    /**
+     * Shake latches while both yaw extremes sit in the SDK's rolling 10-slot window, so
+     * a single 0 read can still hide a half-shake residue from the previous step. Only
+     * arm SHAKE after the window has provably flushed.
+     */
+    private static final int SHAKE_ARM_ZERO_FRAMES = 10;
+    private static final long FACE_LOST_GRACE_MS = 800;
+    private static final float MAX_START_YAW_DEG = 20f;
+    private static final float MAX_START_PITCH_DEG = 15f;
+
+    // ---- Pose-mode tunables ----
+    /** Latest action shown large plus up to 5 smaller history entries. */
+    private static final int POSE_HISTORY_MAX = 6;
+
+    // ---- Shared gating ----
+    /** Faces narrower than this fraction of the frame width are too far away to judge. */
+    private static final float MIN_FACE_WIDTH_RATIO = 0.18f;
+
+    private enum Phase { WAIT_FACE, CHALLENGE, PASSED, FAILED }
+
+    private final Context context;
+    private final Random random = new Random();
+
+    private final CustomParameter silentParam = InspireFace.CreateCustomParameter().enableLiveness(true);
+    private final CustomParameter actionParam = InspireFace.CreateCustomParameter().enableInteractionLiveness(true);
+
+    private final int colorNeutral;
+    private final int colorWarn;
+    private final int colorFail;
+
+    private Mode mode = Mode.SILENT;
+
+    // Silent state
+    private final ArrayDeque<Float> scoreWindow = new ArrayDeque<>();
+    private int silentFrameCounter;
+    private UiState lastSilentState;
+    private int silentTrackId = -1;
+
+    // Pose state
+    private final ArrayDeque<ActionType> poseHistory = new ArrayDeque<>();
+    private final int[] posePrevFlags = new int[ActionType.values().length];
+    private int poseTrackId = -1;
+
+    // Action state
+    private Phase phase = Phase.WAIT_FACE;
+    private final List<ActionType> sequence = new ArrayList<>();
+    private int actionIndex;
+    private long actionDeadline;
+    /**
+     * Edge trigger: jawOpen/headRaise stay 1 while the pose is held and shake latches for
+     * ~10 pipeline calls, so a leftover from the previous step could complete the next one
+     * instantly. Each step first has to observe the flag at 0 before a 1 counts —
+     * SHAKE needs a full window of zeros (see SHAKE_ARM_ZERO_FRAMES).
+     */
+    private boolean actionArmed;
+    private int armZeroStreak;
+    private int stableFrames;
+    private int lockedTrackId = -1;
+    private long faceLostSince;
+    private int failReasonRes;
+
+    LivenessController(Context context) {
+        // Deliberately the activity context, not the application one: per-app locales
+        // (AppCompatDelegate) only wrap activity contexts below API 33, and every prompt
+        // string is resolved through this reference. Lifetime matches the activity.
+        this.context = context;
+        colorNeutral = ContextCompat.getColor(context, R.color.liveness_accent);
+        colorWarn = ContextCompat.getColor(context, R.color.liveness_warn);
+        colorFail = ContextCompat.getColor(context, R.color.liveness_fail);
+    }
+
+    synchronized Mode getMode() {
+        return mode;
+    }
+
+    synchronized void setMode(Mode newMode) {
+        if (mode != newMode) {
+            mode = newMode;
+            reset();
+        }
+    }
+
+    synchronized void restart() {
+        reset();
+    }
+
+    private void reset() {
+        clearSilentState();
+        phase = Phase.WAIT_FACE;
+        sequence.clear();
+        actionIndex = 0;
+        stableFrames = 0;
+        lockedTrackId = -1;
+        faceLostSince = 0;
+        poseHistory.clear();
+        resetPoseEdges();
+    }
+
+    /**
+     * Consumes one tracked frame. Runs the pipeline appropriate for the current mode while
+     * {@code stream} is still alive and returns what the UI should display.
+     */
+    synchronized UiState onFrame(Session session, ImageStream stream,
+                                 MultipleFaceData faces, int uprightWidth, int uprightHeight) {
+        switch (mode) {
+            case SILENT:
+                return onSilentFrame(session, stream, faces, uprightWidth);
+            case POSE:
+                return onPoseFrame(session, stream, faces);
+            case ACTION:
+            default:
+                return onActionFrame(session, stream, faces, uprightWidth);
+        }
+    }
+
+    // ------------------------------------------------------------------
+    // Silent (RGB anti-spoofing) mode
+    // ------------------------------------------------------------------
+
+    private UiState onSilentFrame(Session session, ImageStream stream,
+                                  MultipleFaceData faces, int uprightWidth) {
+        if (faces.detectedNum == 0) {
+            clearSilentState();
+            return new UiState(str(R.string.msg_no_face), null, -1, false, colorNeutral);
+        }
+        if (faces.detectedNum > 1) {
+            clearSilentState();
+            return new UiState(str(R.string.msg_multiple_faces), null, -1, false, colorWarn);
+        }
+        int idx = largestFaceIndex(faces);
+        if (!isFaceBigEnough(faces.rects[idx], uprightWidth)) {
+            clearSilentState();
+            return new UiState(str(R.string.msg_face_too_small), null, -1, false, colorWarn);
+        }
+        // A different track id means a different subject — never blend its scores or show
+        // the previous subject's cached verdict.
+        if (faces.trackIds[idx] != silentTrackId) {
+            clearSilentState();
+            silentTrackId = faces.trackIds[idx];
+        }
+
+        if (silentFrameCounter++ % SILENT_PIPELINE_INTERVAL != 0 && lastSilentState != null) {
+            return lastSilentState;
+        }
+        if (!InspireFace.MultipleFacePipelineProcess(session, stream, faces, silentParam)) {
+            return new UiState(str(R.string.silent_analyzing), null, -1, false, colorNeutral);
+        }
+        RGBLivenessConfidence liveness = InspireFace.GetRGBLivenessConfidence(session);
+        if (liveness == null || liveness.num <= idx) {
+            return new UiState(str(R.string.silent_analyzing), null, -1, false, colorNeutral);
+        }
+
+        scoreWindow.addLast(liveness.confidence[idx]);
+        while (scoreWindow.size() > SCORE_WINDOW) {
+            scoreWindow.removeFirst();
+        }
+        float sum = 0f;
+        for (float v : scoreWindow) {
+            sum += v;
+        }
+        float avg = sum / scoreWindow.size();
+        String scoreText = str(R.string.silent_score, avg);
+        int progress = Math.round(avg * 100);
+
+        UiState state;
+        if (scoreWindow.size() < MIN_VERDICT_SAMPLES) {
+            state = new UiState(str(R.string.silent_analyzing), scoreText, progress, false, colorNeutral);
+        } else {
+            boolean live = avg >= RGB_LIVENESS_THRESHOLD;
+            state = new UiState(
+                    str(live ? R.string.silent_real : R.string.silent_fake),
+                    scoreText, progress, false, live ? colorNeutral : colorFail);
+        }
+        lastSilentState = state;
+        return state;
+    }
+
+    private void clearSilentState() {
+        scoreWindow.clear();
+        silentFrameCounter = 0;
+        lastSilentState = null;
+        silentTrackId = -1;
+    }
+
+    // ------------------------------------------------------------------
+    // Cooperative action mode
+    // ------------------------------------------------------------------
+
+    private UiState onActionFrame(Session session, ImageStream stream,
+                                  MultipleFaceData faces, int uprightWidth) {
+        long now = SystemClock.elapsedRealtime();
+        switch (phase) {
+            case PASSED:
+                return new UiState(str(R.string.action_passed), null, 100, true, colorNeutral);
+            case FAILED:
+                return new UiState(str(failReasonRes), null, -1, true, colorFail);
+            case WAIT_FACE:
+                return onWaitFace(session, stream, faces, uprightWidth);
+            case CHALLENGE:
+            default:
+                return onChallenge(session, stream, faces, now);
+        }
+    }
+
+    private UiState onWaitFace(Session session, ImageStream stream,
+                               MultipleFaceData faces, int uprightWidth) {
+        if (faces.detectedNum == 0) {
+            stableFrames = 0;
+            return new UiState(str(R.string.msg_no_face), null, -1, false, colorNeutral);
+        }
+        if (faces.detectedNum > 1) {
+            stableFrames = 0;
+            return new UiState(str(R.string.msg_multiple_faces), null, -1, false, colorWarn);
+        }
+        int idx = largestFaceIndex(faces);
+        if (!isFaceBigEnough(faces.rects[idx], uprightWidth)) {
+            stableFrames = 0;
+            return new UiState(str(R.string.msg_face_too_small), null, -1, false, colorWarn);
+        }
+        // Pre-warm the SDK's 10-call action window during get-ready so the first
+        // challenge is detectable the moment it is shown.
+        InspireFace.MultipleFacePipelineProcess(session, stream, faces, actionParam);
+        // angles[] is only trustworthy at index 0 — the 1.2.0 JNI writes face[0]'s angles
+        // into every slot. We require exactly one face here, so idx is always 0.
+        boolean frontal = Math.abs(faces.angles[idx].yaw) <= MAX_START_YAW_DEG
+                && Math.abs(faces.angles[idx].pitch) <= MAX_START_PITCH_DEG;
+        if (!frontal) {
+            stableFrames = 0;
+            return new UiState(str(R.string.action_get_ready), null, -1, false, colorWarn);
+        }
+        stableFrames++;
+        if (stableFrames < STABLE_FRAMES_TO_START) {
+            int progress = stableFrames * 100 / STABLE_FRAMES_TO_START;
+            return new UiState(str(R.string.action_get_ready), null, progress, false, colorNeutral);
+        }
+        startChallenge(faces.trackIds[idx]);
+        return challengeState(SystemClock.elapsedRealtime(), null);
+    }
+
+    private void startChallenge(int trackId) {
+        List<ActionType> pool = new ArrayList<>(Arrays.asList(ActionType.values()));
+        Collections.shuffle(pool, random);
+        sequence.clear();
+        sequence.addAll(pool.subList(0, Math.min(ACTIONS_PER_RUN, pool.size())));
+        actionIndex = 0;
+        lockedTrackId = trackId;
+        faceLostSince = 0;
+        armCurrentAction();
+        phase = Phase.CHALLENGE;
+    }
+
+    private void armCurrentAction() {
+        actionDeadline = SystemClock.elapsedRealtime() + ACTION_TIMEOUT_MS;
+        actionArmed = false;
+        armZeroStreak = 0;
+    }
+
+    private UiState onChallenge(Session session, ImageStream stream,
+                                MultipleFaceData faces, long now) {
+        int idx = indexOfTrackId(faces, lockedTrackId);
+        if (idx < 0) {
+            if (faceLostSince == 0) {
+                faceLostSince = now;
+            }
+            if (now - faceLostSince > FACE_LOST_GRACE_MS) {
+                return fail(R.string.action_face_lost);
+            }
+            return new UiState(str(R.string.msg_no_face), stepText(), -1, false, colorWarn);
+        }
+        faceLostSince = 0;
+
+        if (now > actionDeadline) {
+            return fail(R.string.action_failed_timeout);
+        }
+
+        if (InspireFace.MultipleFacePipelineProcess(session, stream, faces, actionParam)) {
+            FaceInteractionsActions actions = InspireFace.GetFaceInteractionActionsResult(session);
+            if (actions != null && actions.num > idx) {
+                ActionType current = sequence.get(actionIndex);
+                int flag = actionFlag(actions, idx, current);
+                // normal==1 marks the SDK's warm-up (fresh track or post-blink window
+                // reset) during which every flag reads a placeholder 0. Those zeros must
+                // not open the gate, or a pose held through a blink would count as fresh.
+                boolean warmingUp = actions.normal[idx] == 1;
+                if (!actionArmed) {
+                    // -1 means "not evaluated" — only a real 0 opens the gate.
+                    armZeroStreak = !warmingUp && flag == 0 ? armZeroStreak + 1 : 0;
+                    actionArmed = armZeroStreak
+                            >= (current == ActionType.SHAKE ? SHAKE_ARM_ZERO_FRAMES : 1);
+                } else if (flag > 0 && !warmingUp) {
+                    actionIndex++;
+                    if (actionIndex >= sequence.size()) {
+                        phase = Phase.PASSED;
+                        lockedTrackId = -1;
+                        return new UiState(str(R.string.action_passed), null, 100, true, colorNeutral);
+                    }
+                    armCurrentAction();
+                }
+            }
+        }
+        String warn = faces.detectedNum > 1 ? str(R.string.msg_multiple_faces) : null;
+        return challengeState(now, warn);
+    }
+
+    private UiState challengeState(long now, String warnOverride) {
+        long msLeft = Math.max(0, actionDeadline - now);
+        String subtitle = warnOverride != null
+                ? warnOverride
+                : stepText() + " · " + str(R.string.action_time_left, (int) ((msLeft + 999) / 1000));
+        int progress = (int) (msLeft * 100 / ACTION_TIMEOUT_MS);
+        return new UiState(str(sequence.get(actionIndex).promptRes), subtitle, progress, false, colorNeutral);
+    }
+
+    private String stepText() {
+        return str(R.string.action_step, actionIndex + 1, sequence.size());
+    }
+
+    private UiState fail(int reasonRes) {
+        phase = Phase.FAILED;
+        failReasonRes = reasonRes;
+        lockedTrackId = -1;
+        return new UiState(str(reasonRes), null, -1, true, colorFail);
+    }
+
+    private static int actionFlag(FaceInteractionsActions actions, int idx, ActionType type) {
+        switch (type) {
+            case BLINK:
+                return actions.blink[idx];
+            case SHAKE:
+                return actions.shake[idx];
+            case JAW_OPEN:
+                return actions.jawOpen[idx];
+            case HEAD_RAISE:
+            default:
+                return actions.headRaise[idx];
+        }
+    }
+
+    // ------------------------------------------------------------------
+    // Pose recognition mode — display whatever action the user performs
+    // ------------------------------------------------------------------
+
+    private UiState onPoseFrame(Session session, ImageStream stream, MultipleFaceData faces) {
+        if (faces.detectedNum == 0) {
+            resetPoseEdges();
+            poseTrackId = -1;
+            // Without a face there is no "current" action — show the full history small.
+            return new UiState(str(R.string.msg_no_face), poseHistoryLine(true), -1, false, colorNeutral);
+        }
+        if (faces.detectedNum > 1) {
+            // Edge state is keyed to one face; following the "largest" of two similar
+            // faces would flap and fabricate rising edges.
+            return new UiState(str(R.string.msg_multiple_faces), poseHistoryLine(true), -1, false, colorWarn);
+        }
+        int idx = largestFaceIndex(faces);
+        if (faces.trackIds[idx] != poseTrackId) {
+            resetPoseEdges();
+            poseTrackId = faces.trackIds[idx];
+        }
+        if (InspireFace.MultipleFacePipelineProcess(session, stream, faces, actionParam)) {
+            FaceInteractionsActions actions = InspireFace.GetFaceInteractionActionsResult(session);
+            // Skip the SDK warm-up (normal==1: fresh track or post-blink window reset):
+            // its placeholder zeros would read as "pose released" and make a pose held
+            // through a natural blink re-register as a duplicate entry.
+            if (actions != null && actions.num > idx && actions.normal[idx] != 1) {
+                for (ActionType type : ActionType.values()) {
+                    int flag = actionFlag(actions, idx, type);
+                    if (flag < 0) {
+                        continue; // not evaluated this call — keep the previous edge state
+                    }
+                    // Rising edge only: level-triggered flags (jawOpen/headRaise) and the
+                    // window-latched shake must return to 0 before they register again.
+                    if (flag == 1 && posePrevFlags[type.ordinal()] == 0) {
+                        poseHistory.addFirst(type);
+                        while (poseHistory.size() > POSE_HISTORY_MAX) {
+                            poseHistory.removeLast();
+                        }
+                    }
+                    posePrevFlags[type.ordinal()] = flag;
+                }
+            }
+        }
+        String title = poseHistory.isEmpty()
+                ? str(R.string.pose_hint)
+                : str(poseHistory.peekFirst().nameRes);
+        return new UiState(title, poseHistoryLine(false), -1, false, colorNeutral);
+    }
+
+    /** Edge-detector state only; the visible history is kept until reset/mode switch. */
+    private void resetPoseEdges() {
+        Arrays.fill(posePrevFlags, 0);
+    }
+
+    /**
+     * History entries joined oldest-last with decreasing opacity, so past actions visually
+     * fade out as new ones push them toward the tail.
+     */
+    private CharSequence poseHistoryLine(boolean includeHead) {
+        int skip = includeHead ? 0 : 1;
+        if (poseHistory.size() <= skip) {
+            return null;
+        }
+        SpannableStringBuilder line = new SpannableStringBuilder();
+        int i = 0;
+        int shown = 0;
+        int total = poseHistory.size() - skip;
+        for (ActionType type : poseHistory) {
+            if (i++ < skip) {
+                continue;
+            }
+            if (line.length() > 0) {
+                line.append("  ·  ");
+            }
+            int start = line.length();
+            line.append(str(type.nameRes));
+            int alpha = 230 - (total <= 1 ? 0 : 160 * shown / (total - 1));
+            line.setSpan(new ForegroundColorSpan(Color.argb(alpha, 255, 255, 255)),
+                    start, line.length(), Spanned.SPAN_EXCLUSIVE_EXCLUSIVE);
+            shown++;
+        }
+        return line;
+    }
+
+    // ------------------------------------------------------------------
+    // Helpers
+    // ------------------------------------------------------------------
+
+    private static int largestFaceIndex(MultipleFaceData faces) {
+        int best = 0;
+        long bestArea = -1;
+        for (int i = 0; i < faces.detectedNum; i++) {
+            FaceRect r = faces.rects[i];
+            long area = (long) r.width * r.height;
+            if (area > bestArea) {
+                bestArea = area;
+                best = i;
+            }
+        }
+        return best;
+    }
+
+    private static int indexOfTrackId(MultipleFaceData faces, int trackId) {
+        for (int i = 0; i < faces.detectedNum; i++) {
+            if (faces.trackIds[i] == trackId) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    private static boolean isFaceBigEnough(FaceRect rect, int uprightWidth) {
+        return rect.width >= uprightWidth * MIN_FACE_WIDTH_RATIO;
+    }
+
+    private String str(int res, Object... args) {
+        return args.length == 0 ? context.getString(res) : context.getString(res, args);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/NativeSessionBridge.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/NativeSessionBridge.java
new file mode 100644
index 000000000..e11eaa53c
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/NativeSessionBridge.java
@@ -0,0 +1,54 @@
+package com.example.inspireface_example.view;
+
+import android.util.Log;
+
+import androidx.annotation.Nullable;
+
+import com.insightface.sdk.inspireface.base.Session;
+
+/**
+ * Small compatibility bridge for the landmark flag omitted by InspireFace Android 1.2.0's
+ * CreateSession JNI wrapper. Session release and all processing still use the official API.
+ */
+final class NativeSessionBridge {
+
+    private static final String TAG = "NativeSessionBridge";
+    private static final int ENABLE_DETECT_MODE_LANDMARK = 0x00000200;
+    private static final boolean AVAILABLE;
+
+    static {
+        boolean available;
+        try {
+            System.loadLibrary("inspireface_session_bridge");
+            available = true;
+        } catch (UnsatisfiedLinkError error) {
+            Log.e(TAG, "Could not load Session compatibility bridge", error);
+            available = false;
+        }
+        AVAILABLE = available;
+    }
+
+    private NativeSessionBridge() {
+    }
+
+    @Nullable
+    static Session createLandmarkSession(int detectMode, int maxFaces,
+                                         int detectPixelLevel, int trackByDetectFps) {
+        if (!AVAILABLE) {
+            return null;
+        }
+        long handle = nativeCreateLandmarkSession(
+                ENABLE_DETECT_MODE_LANDMARK, detectMode,
+                maxFaces, detectPixelLevel, trackByDetectFps);
+        if (handle == 0L) {
+            return null;
+        }
+        Session session = new Session();
+        session.handle = handle;
+        return session;
+    }
+
+    private static native long nativeCreateLandmarkSession(
+            int customOptions, int detectMode, int maxFaces,
+            int detectPixelLevel, int trackByDetectFps);
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/Nv21Converter.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/Nv21Converter.java
new file mode 100644
index 000000000..675fd9abe
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/Nv21Converter.java
@@ -0,0 +1,191 @@
+package com.example.inspireface_example.view;
+
+import androidx.camera.core.ImageProxy;
+
+import java.nio.ByteBuffer;
+
+/**
+ * Converts CameraX YUV_420_888 frames to tightly packed NV21, plus an NV21 rotation step
+ * so frames can be handed to InspireFace already upright with CAMERA_ROTATION_0.
+ *
+ * Rotating on the Java side is deliberate: the 1.2.0 SDK's RGB anti-spoofing crop mixes
+ * rotated and unrotated coordinate spaces when a non-zero rotation constant is used, which
+ * silently degrades silent-liveness scores. An upright buffer avoids that entirely and
+ * makes every SDK output coordinate match the preview orientation.
+ *
+ * On virtually all camera HALs the U/V planes of YUV_420_888 alias one interleaved
+ * VU buffer (i.e. the memory already is NV21). That is detected once on the first
+ * frame; afterwards the chroma plane is moved with a single bulk copy. All buffers are
+ * reused across frames, so the converter allocates only on size changes.
+ *
+ * Not thread-safe: use one instance per analysis thread.
+ */
+final class Nv21Converter {
+
+    private static final int UNKNOWN = -1;
+    private static final int INTERLEAVED = 1;
+    private static final int PLANAR = 0;
+
+    private byte[] out;
+    private byte[] rotated;
+    private int uvLayout = UNKNOWN;
+
+    /**
+     * Returns the frame as NV21. The returned array is owned by the converter and
+     * overwritten by the next call.
+     */
+    byte[] convert(ImageProxy image) {
+        int width = image.getWidth();
+        int height = image.getHeight();
+        int ySize = width * height;
+        int total = ySize + ySize / 2;
+        if (out == null || out.length != total) {
+            out = new byte[total];
+            uvLayout = UNKNOWN;
+        }
+
+        ImageProxy.PlaneProxy[] planes = image.getPlanes();
+        copyLuma(planes[0], width, height, out);
+
+        ImageProxy.PlaneProxy uPlane = planes[1];
+        ImageProxy.PlaneProxy vPlane = planes[2];
+        if (uvLayout == UNKNOWN) {
+            uvLayout = isVuInterleaved(uPlane, vPlane, width, height) ? INTERLEAVED : PLANAR;
+        }
+        if (uvLayout == INTERLEAVED) {
+            copyChromaInterleaved(uPlane, vPlane, ySize, out);
+        } else {
+            copyChromaPlanar(uPlane, vPlane, width, height, ySize, out);
+        }
+        return out;
+    }
+
+    /**
+     * Rotates a tight NV21 buffer clockwise by the given degrees so it appears upright.
+     * Returns {@code src} itself for 0°; otherwise a reused internal buffer.
+     */
+    byte[] rotateUpright(byte[] src, int width, int height, int rotationDegrees) {
+        if (rotationDegrees == 0) {
+            return src;
+        }
+        if (rotated == null || rotated.length != src.length) {
+            rotated = new byte[src.length];
+        }
+        int ySize = width * height;
+        int i = 0;
+        switch (rotationDegrees) {
+            case 90:
+                for (int x = 0; x < width; x++) {
+                    for (int y = height - 1; y >= 0; y--) {
+                        rotated[i++] = src[y * width + x];
+                    }
+                }
+                for (int x = 0; x < width; x += 2) {
+                    for (int y = height / 2 - 1; y >= 0; y--) {
+                        int s = ySize + y * width + x;
+                        rotated[i++] = src[s];
+                        rotated[i++] = src[s + 1];
+                    }
+                }
+                break;
+            case 180:
+                for (int p = ySize - 1; p >= 0; p--) {
+                    rotated[i++] = src[p];
+                }
+                for (int p = src.length - 2; p >= ySize; p -= 2) {
+                    rotated[i++] = src[p];
+                    rotated[i++] = src[p + 1];
+                }
+                break;
+            case 270:
+            default:
+                for (int x = width - 1; x >= 0; x--) {
+                    for (int y = 0; y < height; y++) {
+                        rotated[i++] = src[y * width + x];
+                    }
+                }
+                for (int x = width - 2; x >= 0; x -= 2) {
+                    for (int y = 0; y < height / 2; y++) {
+                        int s = ySize + y * width + x;
+                        rotated[i++] = src[s];
+                        rotated[i++] = src[s + 1];
+                    }
+                }
+                break;
+        }
+        return rotated;
+    }
+
+    private static void copyLuma(ImageProxy.PlaneProxy yPlane, int width, int height, byte[] out) {
+        ByteBuffer buf = yPlane.getBuffer();
+        int base = buf.position();
+        int rowStride = yPlane.getRowStride();
+        if (rowStride == width) {
+            buf.get(out, 0, width * height);
+        } else {
+            for (int row = 0; row < height; row++) {
+                buf.position(base + row * rowStride);
+                buf.get(out, row * width, width);
+            }
+        }
+        buf.position(base);
+    }
+
+    /** The V buffer already views VUVU…: one bulk copy plus the trailing U byte it cannot see. */
+    private static void copyChromaInterleaved(ImageProxy.PlaneProxy uPlane, ImageProxy.PlaneProxy vPlane,
+                                              int ySize, byte[] out) {
+        ByteBuffer vBuf = vPlane.getBuffer();
+        int vPos = vBuf.position();
+        int vuSize = ySize / 2;
+        int n = Math.min(vBuf.remaining(), vuSize);
+        vBuf.get(out, ySize, n);
+        vBuf.position(vPos);
+        if (n < vuSize) {
+            ByteBuffer uBuf = uPlane.getBuffer();
+            out[ySize + vuSize - 1] = uBuf.get(uBuf.limit() - 1);
+        }
+    }
+
+    private static void copyChromaPlanar(ImageProxy.PlaneProxy uPlane, ImageProxy.PlaneProxy vPlane,
+                                         int width, int height, int ySize, byte[] out) {
+        ByteBuffer uBuf = uPlane.getBuffer();
+        ByteBuffer vBuf = vPlane.getBuffer();
+        int uBase = uBuf.position();
+        int vBase = vBuf.position();
+        int uRowStride = uPlane.getRowStride();
+        int vRowStride = vPlane.getRowStride();
+        int uPixStride = uPlane.getPixelStride();
+        int vPixStride = vPlane.getPixelStride();
+        int pos = ySize;
+        for (int row = 0; row < height / 2; row++) {
+            int uRow = uBase + row * uRowStride;
+            int vRow = vBase + row * vRowStride;
+            for (int col = 0; col < width / 2; col++) {
+                out[pos++] = vBuf.get(vRow + col * vPixStride);
+                out[pos++] = uBuf.get(uRow + col * uPixStride);
+            }
+        }
+    }
+
+    /**
+     * True when the U/V planes alias a single interleaved VU buffer with no row
+     * padding, i.e. shifting V by one byte yields exactly the U plane.
+     */
+    private static boolean isVuInterleaved(ImageProxy.PlaneProxy uPlane, ImageProxy.PlaneProxy vPlane,
+                                           int width, int height) {
+        if (uPlane.getPixelStride() != 2 || vPlane.getPixelStride() != 2) {
+            return false;
+        }
+        ByteBuffer uBuf = uPlane.getBuffer();
+        ByteBuffer vBuf = vPlane.getBuffer();
+        int vPos = vBuf.position();
+        int uLimit = uBuf.limit();
+        vBuf.position(vPos + 1);
+        uBuf.limit(uLimit - 1);
+        boolean interleaved = vBuf.remaining() == (width * height / 2 - 2)
+                && vBuf.compareTo(uBuf) == 0;
+        vBuf.position(vPos);
+        uBuf.limit(uLimit);
+        return interleaved;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/PoseActivity.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/PoseActivity.java
new file mode 100644
index 000000000..4c850b336
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/PoseActivity.java
@@ -0,0 +1,17 @@
+package com.example.inspireface_example.view;
+
+import com.example.inspireface_example.R;
+
+/** Dedicated pose/action-recognition screen. */
+public final class PoseActivity extends LivenessActivity {
+
+    @Override
+    protected LivenessController.Mode initialMode() {
+        return LivenessController.Mode.POSE;
+    }
+
+    @Override
+    protected int pageTitleRes() {
+        return R.string.mode_pose;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/RecognitionFaceAnalyzer.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/RecognitionFaceAnalyzer.java
new file mode 100644
index 000000000..8d720e83e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/RecognitionFaceAnalyzer.java
@@ -0,0 +1,183 @@
+package com.example.inspireface_example.view;
+
+import android.content.Context;
+import android.graphics.RectF;
+import android.os.SystemClock;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+import com.example.inspireface_example.face.FaceRecord;
+import com.example.inspireface_example.face.FaceRepository;
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.FaceFeature;
+import com.insightface.sdk.inspireface.base.FaceRect;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Session;
+
+/** Tracks only SDK face 0 and searches it after roughly one stable second. */
+public final class RecognitionFaceAnalyzer extends UprightFaceCameraAnalyzer {
+
+    public enum State {
+        NO_FACE, MOVE_CLOSER, HOLD_STILL, SEARCHING, MATCHED, NO_MATCH, EMPTY_LIBRARY
+    }
+
+    public interface Listener {
+        /** Called on the camera analysis executor. */
+        void onState(State state, @Nullable FaceRecord record,
+                     float confidence, float threshold);
+
+        void onSessionError();
+    }
+
+    private static final long STABLE_RECOGNITION_MS = 1_000L;
+    private static final float MIN_FACE_WIDTH_RATIO = 0.12f;
+
+    private final FaceOverlayView overlay;
+    private final FaceRepository repository;
+    private final boolean libraryEmpty;
+    private final Listener listener;
+    private final FaceStabilityGate stabilityGate =
+            new FaceStabilityGate(STABLE_RECOGNITION_MS, 1L);
+    private final int waitingColor;
+    private final int matchColor;
+    private final int noMatchColor;
+    private final int warningColor;
+
+    private volatile boolean mirrored = true;
+    private volatile boolean resetRequested;
+    private boolean recognizedStableRun;
+    private State lastState;
+    private int currentBoxColor;
+
+    public RecognitionFaceAnalyzer(Context context, FaceOverlayView overlay,
+                                   FaceRepository repository, boolean libraryEmpty,
+                                   Listener listener) {
+        this.overlay = overlay;
+        this.repository = repository;
+        this.libraryEmpty = libraryEmpty;
+        this.listener = listener;
+        waitingColor = ContextCompat.getColor(context, R.color.liveness_accent);
+        matchColor = waitingColor;
+        noMatchColor = ContextCompat.getColor(context, R.color.liveness_fail);
+        warningColor = ContextCompat.getColor(context, R.color.liveness_warn);
+        currentBoxColor = waitingColor;
+    }
+
+    @Override
+    protected void beforeFrame() {
+        if (resetRequested) {
+            resetRequested = false;
+            resetRecognition();
+        }
+    }
+
+    @Override
+    protected Session createSession() {
+        return FaceEngine.createVideoRecognitionSession();
+    }
+
+    @Override
+    protected void onFaces(Session session, ImageStream stream,
+                           @Nullable MultipleFaceData faces, byte[] uprightNv21,
+                           int uprightWidth, int uprightHeight, long frameStart) {
+        if (faces == null || faces.detectedNum == 0) {
+            resetRecognition();
+            overlay.submit(null);
+            report(State.NO_FACE, null, Float.NaN, Float.NaN);
+            return;
+        }
+
+        FaceRect first = faces.rects[0];
+        RectF face = new RectF(first.x, first.y,
+                first.x + first.width, first.y + first.height);
+        if (face.width() < uprightWidth * MIN_FACE_WIDTH_RATIO) {
+            resetRecognition();
+            currentBoxColor = warningColor;
+            submitFace(face, uprightWidth, uprightHeight);
+            report(State.MOVE_CLOSER, null, Float.NaN, Float.NaN);
+            return;
+        }
+
+        int trackId = faces.trackIds != null && faces.trackIds.length > 0
+                ? faces.trackIds[0] : 0;
+        float stable = stabilityGate.update(trackId,
+                face.left, face.top, face.right, face.bottom,
+                SystemClock.elapsedRealtime());
+        if (stable < 0f) {
+            recognizedStableRun = false;
+            currentBoxColor = waitingColor;
+            submitFace(face, uprightWidth, uprightHeight);
+            report(State.HOLD_STILL, null, Float.NaN, Float.NaN);
+            return;
+        }
+
+        if (!recognizedStableRun) {
+            recognizedStableRun = true;
+            if (libraryEmpty) {
+                currentBoxColor = warningColor;
+                submitFace(face, uprightWidth, uprightHeight);
+                report(State.EMPTY_LIBRARY, null, Float.NaN, Float.NaN);
+            } else {
+                recognize(session, stream, faces, face, uprightWidth, uprightHeight);
+            }
+        } else {
+            submitFace(face, uprightWidth, uprightHeight);
+        }
+    }
+
+    private void recognize(Session session, ImageStream stream, MultipleFaceData faces,
+                           RectF face, int imageWidth, int imageHeight) {
+        currentBoxColor = warningColor;
+        submitFace(face, imageWidth, imageHeight);
+        report(State.SEARCHING, null, Float.NaN, Float.NaN);
+        FaceFeature feature = InspireFace.ExtractFaceFeature(
+                session, stream, faces.tokens[0]);
+        FaceRepository.SearchResult result = repository.search(feature);
+        if (result.matched && result.record != null) {
+            currentBoxColor = matchColor;
+            report(State.MATCHED, result.record, result.confidence, result.threshold);
+        } else {
+            currentBoxColor = noMatchColor;
+            report(State.NO_MATCH, null, result.confidence, result.threshold);
+        }
+        submitFace(face, imageWidth, imageHeight);
+    }
+
+    @Override
+    protected void onSessionError() {
+        listener.onSessionError();
+    }
+
+    public void setMirrored(boolean mirrored) {
+        this.mirrored = mirrored;
+    }
+
+    public void resetTracking() {
+        resetRequested = true;
+        overlay.submit(null);
+    }
+
+    private void resetRecognition() {
+        stabilityGate.reset();
+        recognizedStableRun = false;
+        currentBoxColor = waitingColor;
+    }
+
+    private void submitFace(RectF face, int imageWidth, int imageHeight) {
+        overlay.submit(new FaceOverlayView.Frame(
+                imageWidth, imageHeight, mirrored,
+                new RectF[]{new RectF(face)}, currentBoxColor));
+    }
+
+    private void report(State state, @Nullable FaceRecord record,
+                        float confidence, float threshold) {
+        if (state == lastState && state != State.MATCHED && state != State.NO_MATCH) {
+            return;
+        }
+        lastState = state;
+        listener.onState(state, record, confidence, threshold);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/UprightFaceCameraAnalyzer.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/UprightFaceCameraAnalyzer.java
new file mode 100644
index 000000000..ff41babe4
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/view/UprightFaceCameraAnalyzer.java
@@ -0,0 +1,103 @@
+package com.example.inspireface_example.view;
+
+import android.os.SystemClock;
+
+import androidx.annotation.NonNull;
+import androidx.annotation.Nullable;
+import androidx.camera.core.ImageAnalysis;
+import androidx.camera.core.ImageProxy;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.ImageStream;
+import com.insightface.sdk.inspireface.base.MultipleFaceData;
+import com.insightface.sdk.inspireface.base.Session;
+
+/**
+ * Shared camera-to-InspireFace pipeline. Subclasses receive an upright NV21 frame,
+ * tracked faces and a live stream while this class owns conversion and native lifetimes.
+ */
+public abstract class UprightFaceCameraAnalyzer implements ImageAnalysis.Analyzer {
+
+    private final Nv21Converter converter = new Nv21Converter();
+    private Session session;
+    private boolean sessionFailed;
+    private volatile boolean released;
+
+    @Override
+    public final void analyze(@NonNull ImageProxy image) {
+        if (released || sessionFailed || shouldSkipFrame()) {
+            image.close();
+            return;
+        }
+        beforeFrame();
+        if (session == null) {
+            session = createSession();
+            if (session == null) {
+                sessionFailed = true;
+                image.close();
+                onSessionError();
+                return;
+            }
+            onSessionReady();
+        }
+
+        long frameStart = SystemClock.elapsedRealtime();
+        int width = image.getWidth();
+        int height = image.getHeight();
+        int rotationDegrees = image.getImageInfo().getRotationDegrees();
+        byte[] nv21;
+        try {
+            nv21 = converter.convert(image);
+        } finally {
+            image.close();
+        }
+        byte[] upright = converter.rotateUpright(nv21, width, height, rotationDegrees);
+        boolean swapped = rotationDegrees == 90 || rotationDegrees == 270;
+        int uprightWidth = swapped ? height : width;
+        int uprightHeight = swapped ? width : height;
+
+        ImageStream stream = InspireFace.CreateImageStreamFromByteBuffer(
+                upright, uprightWidth, uprightHeight,
+                InspireFace.STREAM_YUV_NV21, InspireFace.CAMERA_ROTATION_0);
+        if (stream == null) {
+            return;
+        }
+        try {
+            MultipleFaceData faces = InspireFace.ExecuteFaceTrack(session, stream);
+            onFaces(session, stream, faces, upright,
+                    uprightWidth, uprightHeight, frameStart);
+        } finally {
+            InspireFace.ReleaseImageStream(stream);
+        }
+    }
+
+    /** Called before lazily creating the SDK session, on the analysis executor. */
+    protected void beforeFrame() {
+    }
+
+    /** Lets a completed state cheaply close subsequent camera frames. */
+    protected boolean shouldSkipFrame() {
+        return false;
+    }
+
+    protected abstract Session createSession();
+
+    /** Called once after lazy Session creation succeeds, on the analysis executor. */
+    protected void onSessionReady() {
+    }
+
+    protected abstract void onFaces(Session session, ImageStream stream,
+                                    @Nullable MultipleFaceData faces, byte[] uprightNv21,
+                                    int uprightWidth, int uprightHeight, long frameStart);
+
+    protected abstract void onSessionError();
+
+    /** Must be queued on the same executor used by CameraX analysis. */
+    public final void release() {
+        released = true;
+        if (session != null) {
+            FaceEngine.releaseSession(session);
+            session = null;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceImageOverlayView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceImageOverlayView.java
new file mode 100644
index 000000000..7c66d8341
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceImageOverlayView.java
@@ -0,0 +1,227 @@
+package com.example.inspireface_example.widget;
+
+import android.content.Context;
+import android.graphics.Canvas;
+import android.graphics.Paint;
+import android.graphics.RectF;
+import android.util.AttributeSet;
+import android.view.MotionEvent;
+import android.view.View;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+
+/**
+ * Maps SDK face rectangles over a centerCrop ImageView and lets the user select one.
+ * The selected face is green; other detected faces remain visible in white.
+ */
+public final class FaceImageOverlayView extends View {
+
+    public interface OnFaceSelectedListener {
+        void onFaceSelected(int index);
+    }
+
+    private final Paint outlinePaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint facePaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint badgePaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint badgeTextPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final RectF mappedRect = new RectF();
+    private final int accentColor;
+    private int imageWidth;
+    private int imageHeight;
+    private RectF[] faceRects = new RectF[0];
+    private int selectedIndex = -1;
+    private int pressedIndex = -1;
+    private OnFaceSelectedListener listener;
+
+    public FaceImageOverlayView(Context context) {
+        this(context, null);
+    }
+
+    public FaceImageOverlayView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        outlinePaint.setStyle(Paint.Style.STROKE);
+        outlinePaint.setStrokeWidth(dp(5f));
+        outlinePaint.setColor(0xB3000000);
+
+        accentColor = ContextCompat.getColor(context, R.color.liveness_accent);
+        facePaint.setStyle(Paint.Style.STROKE);
+        facePaint.setStrokeCap(Paint.Cap.ROUND);
+
+        badgePaint.setStyle(Paint.Style.FILL);
+        badgeTextPaint.setTextAlign(Paint.Align.CENTER);
+        badgeTextPaint.setFakeBoldText(true);
+        badgeTextPaint.setTextSize(dp(11f));
+        setImportantForAccessibility(IMPORTANT_FOR_ACCESSIBILITY_NO);
+    }
+
+    public void showFace(int imageWidth, int imageHeight, @Nullable RectF faceRect) {
+        showFaces(imageWidth, imageHeight,
+                faceRect == null ? null : new RectF[]{faceRect}, faceRect == null ? -1 : 0);
+    }
+
+    public void showFaces(int imageWidth, int imageHeight,
+                          @Nullable RectF[] faces, int selectedIndex) {
+        this.imageWidth = imageWidth;
+        this.imageHeight = imageHeight;
+        if (faces == null || faces.length == 0) {
+            faceRects = new RectF[0];
+            this.selectedIndex = -1;
+        } else {
+            faceRects = new RectF[faces.length];
+            for (int i = 0; i < faces.length; i++) {
+                faceRects[i] = new RectF(faces[i]);
+            }
+            this.selectedIndex = selectedIndex >= 0 && selectedIndex < faces.length
+                    ? selectedIndex : 0;
+        }
+        invalidate();
+    }
+
+    public void setSelectedIndex(int index) {
+        if (index >= 0 && index < faceRects.length && selectedIndex != index) {
+            selectedIndex = index;
+            invalidate();
+        }
+    }
+
+    public void setOnFaceSelectedListener(@Nullable OnFaceSelectedListener listener) {
+        this.listener = listener;
+    }
+
+    public void clearFace() {
+        faceRects = new RectF[0];
+        selectedIndex = -1;
+        pressedIndex = -1;
+        invalidate();
+    }
+
+    @Override
+    protected void onDraw(Canvas canvas) {
+        super.onDraw(canvas);
+        if (faceRects.length == 0 || imageWidth <= 0 || imageHeight <= 0
+                || getWidth() <= 0 || getHeight() <= 0) {
+            return;
+        }
+        for (int i = 0; i < faceRects.length; i++) {
+            mapRect(faceRects[i], mappedRect);
+            boolean selected = i == selectedIndex;
+            drawBrackets(canvas, mappedRect, outlinePaint);
+            facePaint.setStrokeWidth(dp(selected ? 3f : 2f));
+            facePaint.setColor(selected ? accentColor : 0xE6FFFFFF);
+            drawBrackets(canvas, mappedRect, facePaint);
+            drawBadge(canvas, mappedRect, i, selected);
+        }
+    }
+
+    @Override
+    public boolean onTouchEvent(MotionEvent event) {
+        switch (event.getActionMasked()) {
+            case MotionEvent.ACTION_DOWN:
+                pressedIndex = hitTest(event.getX(), event.getY());
+                return pressedIndex >= 0;
+            case MotionEvent.ACTION_UP:
+                int releasedIndex = hitTest(event.getX(), event.getY());
+                if (pressedIndex >= 0 && releasedIndex == pressedIndex) {
+                    selectedIndex = releasedIndex;
+                    invalidate();
+                    performClick();
+                    if (listener != null) {
+                        listener.onFaceSelected(releasedIndex);
+                    }
+                }
+                pressedIndex = -1;
+                return true;
+            case MotionEvent.ACTION_CANCEL:
+                pressedIndex = -1;
+                return true;
+            default:
+                return pressedIndex >= 0;
+        }
+    }
+
+    @Override
+    public boolean performClick() {
+        super.performClick();
+        return true;
+    }
+
+    private int hitTest(float x, float y) {
+        int hit = -1;
+        float smallestArea = Float.MAX_VALUE;
+        for (int i = 0; i < faceRects.length; i++) {
+            mapRect(faceRects[i], mappedRect);
+            RectF touchTarget = new RectF(mappedRect);
+            touchTarget.inset(-dp(8f), -dp(8f));
+            float area = mappedRect.width() * mappedRect.height();
+            if (touchTarget.contains(x, y) && area < smallestArea) {
+                smallestArea = area;
+                hit = i;
+            }
+        }
+        return hit;
+    }
+
+    private void mapRect(RectF source, RectF out) {
+        // Must match android:scaleType="centerCrop" on the image directly underneath.
+        float scale = Math.max((float) getWidth() / imageWidth,
+                (float) getHeight() / imageHeight);
+        float dx = (getWidth() - imageWidth * scale) / 2f;
+        float dy = (getHeight() - imageHeight * scale) / 2f;
+        out.set(source.left * scale + dx, source.top * scale + dy,
+                source.right * scale + dx, source.bottom * scale + dy);
+    }
+
+    private void drawBrackets(Canvas canvas, RectF face, Paint paint) {
+        float len = Math.min(face.width(), face.height()) * 0.22f;
+        float radius = Math.min(dp(6f), len * 0.45f);
+        canvas.drawLine(face.left, face.top + len,
+                face.left, face.top + radius, paint);
+        canvas.drawLine(face.left + radius, face.top,
+                face.left + len, face.top, paint);
+        canvas.drawArc(face.left, face.top, face.left + 2f * radius,
+                face.top + 2f * radius, 180f, 90f, false, paint);
+
+        canvas.drawLine(face.right - len, face.top,
+                face.right - radius, face.top, paint);
+        canvas.drawLine(face.right, face.top + radius,
+                face.right, face.top + len, paint);
+        canvas.drawArc(face.right - 2f * radius, face.top, face.right,
+                face.top + 2f * radius, 270f, 90f, false, paint);
+
+        canvas.drawLine(face.left, face.bottom - len,
+                face.left, face.bottom - radius, paint);
+        canvas.drawLine(face.left + radius, face.bottom,
+                face.left + len, face.bottom, paint);
+        canvas.drawArc(face.left, face.bottom - 2f * radius,
+                face.left + 2f * radius, face.bottom,
+                90f, 90f, false, paint);
+
+        canvas.drawLine(face.right - len, face.bottom,
+                face.right - radius, face.bottom, paint);
+        canvas.drawLine(face.right, face.bottom - radius,
+                face.right, face.bottom - len, paint);
+        canvas.drawArc(face.right - 2f * radius, face.bottom - 2f * radius,
+                face.right, face.bottom, 0f, 90f, false, paint);
+    }
+
+    private void drawBadge(Canvas canvas, RectF face, int index, boolean selected) {
+        float badgeRadius = dp(10f);
+        float cx = face.left + badgeRadius;
+        float cy = face.top + badgeRadius;
+        cx = Math.max(badgeRadius + dp(2f), cx);
+        cy = Math.max(badgeRadius + dp(2f), cy);
+        badgePaint.setColor(selected ? accentColor : 0xE6FFFFFF);
+        badgeTextPaint.setColor(0xFF07110F);
+        canvas.drawCircle(cx, cy, badgeRadius, badgePaint);
+        Paint.FontMetrics fm = badgeTextPaint.getFontMetrics();
+        float baseline = cy - (fm.ascent + fm.descent) / 2f;
+        canvas.drawText(String.valueOf(index + 1), cx, baseline, badgeTextPaint);
+    }
+
+    private float dp(float value) {
+        return value * getResources().getDisplayMetrics().density;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceLandmarkOverlayView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceLandmarkOverlayView.java
new file mode 100644
index 000000000..e5e398992
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/FaceLandmarkOverlayView.java
@@ -0,0 +1,95 @@
+package com.example.inspireface_example.widget;
+
+import android.content.Context;
+import android.graphics.Canvas;
+import android.graphics.Paint;
+import android.util.AttributeSet;
+import android.view.View;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+import com.insightface.sdk.inspireface.base.Point2f;
+
+/** Draws one selected face's dense landmarks over a centerCrop ImageView. */
+public final class FaceLandmarkOverlayView extends View {
+
+    private final Paint outlinePaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint pointPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private float[] points = new float[0];
+    private int pointCount;
+    private int imageWidth;
+    private int imageHeight;
+
+    public FaceLandmarkOverlayView(Context context) {
+        this(context, null);
+    }
+
+    public FaceLandmarkOverlayView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        outlinePaint.setStyle(Paint.Style.FILL);
+        outlinePaint.setColor(0xCC000000);
+        pointPaint.setStyle(Paint.Style.FILL);
+        pointPaint.setColor(ContextCompat.getColor(context, R.color.liveness_accent));
+        setImportantForAccessibility(IMPORTANT_FOR_ACCESSIBILITY_NO);
+    }
+
+    public void showPoints(int imageWidth, int imageHeight, @Nullable Point2f[] landmarks) {
+        showPoints(imageWidth, imageHeight, landmarks, 0f, 0f);
+    }
+
+    public void showPoints(int imageWidth, int imageHeight,
+                           @Nullable Point2f[] landmarks,
+                           float offsetX, float offsetY) {
+        this.imageWidth = imageWidth;
+        this.imageHeight = imageHeight;
+        if (landmarks == null || landmarks.length == 0) {
+            points = new float[0];
+            pointCount = 0;
+        } else {
+            points = new float[landmarks.length * 2];
+            pointCount = 0;
+            for (Point2f landmark : landmarks) {
+                if (landmark == null) {
+                    continue;
+                }
+                points[pointCount * 2] = landmark.x - offsetX;
+                points[pointCount * 2 + 1] = landmark.y - offsetY;
+                pointCount++;
+            }
+        }
+        invalidate();
+    }
+
+    public void clearPoints() {
+        points = new float[0];
+        pointCount = 0;
+        invalidate();
+    }
+
+    @Override
+    protected void onDraw(Canvas canvas) {
+        super.onDraw(canvas);
+        if (pointCount == 0 || imageWidth <= 0 || imageHeight <= 0
+                || getWidth() <= 0 || getHeight() <= 0) {
+            return;
+        }
+        float scale = Math.max((float) getWidth() / imageWidth,
+                (float) getHeight() / imageHeight);
+        float dx = (getWidth() - imageWidth * scale) / 2f;
+        float dy = (getHeight() - imageHeight * scale) / 2f;
+        float outlineRadius = dp(2.8f);
+        float pointRadius = dp(1.6f);
+        for (int i = 0; i < pointCount; i++) {
+            float x = points[i * 2] * scale + dx;
+            float y = points[i * 2 + 1] * scale + dy;
+            canvas.drawCircle(x, y, outlineRadius, outlinePaint);
+            canvas.drawCircle(x, y, pointRadius, pointPaint);
+        }
+    }
+
+    private float dp(float value) {
+        return value * getResources().getDisplayMetrics().density;
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/SimilarityGaugeView.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/SimilarityGaugeView.java
new file mode 100644
index 000000000..2944771a4
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/java/com/example/inspireface_example/widget/SimilarityGaugeView.java
@@ -0,0 +1,131 @@
+package com.example.inspireface_example.widget;
+
+import android.animation.ValueAnimator;
+import android.content.Context;
+import android.graphics.Canvas;
+import android.graphics.Paint;
+import android.graphics.RectF;
+import android.graphics.Typeface;
+import android.util.AttributeSet;
+import android.util.TypedValue;
+import android.view.View;
+
+import androidx.annotation.Nullable;
+import androidx.core.content.ContextCompat;
+
+import com.example.inspireface_example.R;
+
+import java.util.Locale;
+
+/** Circular similarity result with an animated progress ring and centered verdict. */
+public final class SimilarityGaugeView extends View {
+
+    private final Paint trackPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint progressPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint valuePaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final Paint labelPaint = new Paint(Paint.ANTI_ALIAS_FLAG);
+    private final RectF arcBounds = new RectF();
+
+    private final int accentColor;
+    private final int failColor;
+    private float displayedPercent;
+    private boolean hasResult;
+    private boolean matched;
+    private String label = "";
+    private ValueAnimator animator;
+
+    public SimilarityGaugeView(Context context) {
+        this(context, null);
+    }
+
+    public SimilarityGaugeView(Context context, @Nullable AttributeSet attrs) {
+        super(context, attrs);
+        accentColor = ContextCompat.getColor(context, R.color.liveness_accent);
+        failColor = ContextCompat.getColor(context, R.color.liveness_fail);
+
+        trackPaint.setStyle(Paint.Style.STROKE);
+        trackPaint.setStrokeCap(Paint.Cap.ROUND);
+        trackPaint.setColor(ContextCompat.getColor(context, R.color.compare_ring_track));
+
+        progressPaint.setStyle(Paint.Style.STROKE);
+        progressPaint.setStrokeCap(Paint.Cap.ROUND);
+
+        valuePaint.setTextAlign(Paint.Align.CENTER);
+        valuePaint.setTypeface(Typeface.create(Typeface.DEFAULT, Typeface.BOLD));
+        valuePaint.setColor(ContextCompat.getColor(context, R.color.white));
+
+        labelPaint.setTextAlign(Paint.Align.CENTER);
+        labelPaint.setColor(ContextCompat.getColor(context, R.color.home_text_secondary));
+    }
+
+    public void showMessage(String message) {
+        cancelAnimator();
+        hasResult = false;
+        displayedPercent = 0f;
+        label = message;
+        setContentDescription(message);
+        invalidate();
+    }
+
+    public void showResult(float percent, boolean isMatched, String verdict) {
+        cancelAnimator();
+        hasResult = true;
+        matched = isMatched;
+        label = verdict;
+        float target = Math.max(0f, Math.min(100f, percent));
+        animator = ValueAnimator.ofFloat(0f, target);
+        animator.setDuration(550L);
+        animator.addUpdateListener(animation -> {
+            displayedPercent = (float) animation.getAnimatedValue();
+            invalidate();
+        });
+        animator.start();
+        setContentDescription(String.format(Locale.US, "%.1f%%, %s", target, verdict));
+    }
+
+    @Override
+    protected void onDraw(Canvas canvas) {
+        super.onDraw(canvas);
+        float stroke = dp(11f);
+        trackPaint.setStrokeWidth(stroke);
+        progressPaint.setStrokeWidth(stroke);
+        progressPaint.setColor(matched ? accentColor : failColor);
+
+        float cx = getWidth() / 2f;
+        float cy = getHeight() / 2f;
+        float radius = Math.min(getWidth(), getHeight()) / 2f - stroke - dp(4f);
+        arcBounds.set(cx - radius, cy - radius, cx + radius, cy + radius);
+        canvas.drawArc(arcBounds, -90f, 360f, false, trackPaint);
+        if (hasResult && displayedPercent > 0f) {
+            canvas.drawArc(arcBounds, -90f, displayedPercent * 3.6f, false, progressPaint);
+        }
+
+        valuePaint.setTextSize(sp(31f));
+        labelPaint.setTextSize(sp(13f));
+        String value = hasResult ? String.format(Locale.US, "%.1f%%", displayedPercent) : "—";
+        canvas.drawText(value, cx, cy + dp(3f), valuePaint);
+        canvas.drawText(label, cx, cy + dp(30f), labelPaint);
+    }
+
+    @Override
+    protected void onDetachedFromWindow() {
+        cancelAnimator();
+        super.onDetachedFromWindow();
+    }
+
+    private void cancelAnimator() {
+        if (animator != null) {
+            animator.cancel();
+            animator = null;
+        }
+    }
+
+    private float dp(float value) {
+        return value * getResources().getDisplayMetrics().density;
+    }
+
+    private float sp(float value) {
+        return TypedValue.applyDimension(TypedValue.COMPLEX_UNIT_SP, value,
+                getResources().getDisplayMetrics());
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable-nodpi/ic_launcher_inspireface_foreground.png b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable-nodpi/ic_launcher_inspireface_foreground.png
new file mode 100644
index 000000000..7012ca420
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable-nodpi/ic_launcher_inspireface_foreground.png differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/bg_chip.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/bg_chip.xml
new file mode 100644
index 000000000..4685c95a5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/bg_chip.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<shape xmlns:android="http://schemas.android.com/apk/res/android"
+    android:shape="rectangle">
+    <solid android:color="@color/liveness_chip_bg" />
+    <corners android:radius="16dp" />
+</shape>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_action_liveness.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_action_liveness.xml
new file mode 100644
index 000000000..8ad65e4e5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_action_liveness.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M520,-40v-240l-84,-80 -40,176 -276,-56 16,-80 192,40 64,-324 -72,28v136h-80v-188l158,-68q35,-15 51.5,-19.5T480,-720q21,0 39,11t29,29l40,64q26,42 70.5,69T760,-520v80q-66,0 -123.5,-27.5T540,-540l-24,120 84,80v300h-80ZM483.5,-763.5Q460,-787 460,-820t23.5,-56.5Q507,-900 540,-900t56.5,23.5Q620,-853 620,-820t-23.5,56.5Q573,-740 540,-740t-56.5,-23.5Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_attributes.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_attributes.xml
new file mode 100644
index 000000000..216aecd80
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_attributes.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M480,-880q39,0 80,8.5t80,25.5l-110,51q-12,-2 -24.5,-3.5T480,-800q-14,0 -27,1t-27,4q42,70 114,112.5T700,-640h14l35,77q-90,11 -188,-21T390,-708q-35,85 -95.5,140T160,-486q0,139 93.5,232.5T480,-160q136,0 229,-96t91,-224q0,-14 -1,-24.5t-3,-25.5l50,-111q18,42 26,81t8,80q0,80 -30.5,152.5t-84,127.5Q712,-145 639,-112.5T480,-80q-82,0 -155,-31.5t-127.5,-86Q143,-252 111.5,-325T80,-480q0,-86 33,-159.5t88.5,-127Q257,-820 329,-850t151,-30ZM395.5,-475.5Q410,-461 410,-440t-14.5,35.5Q381,-390 360,-390t-35.5,-14.5Q310,-419 310,-440t14.5,-35.5Q339,-490 360,-490t35.5,14.5ZM573,-720ZM635.5,-475.5Q650,-461 650,-440t-14.5,35.5Q621,-390 600,-390t-35.5,-14.5Q550,-419 550,-440t14.5,-35.5Q579,-490 600,-490t35.5,14.5ZM780,-920l44,96 96,44 -96,44 -44,96 -44,-96 -96,-44 96,-44 44,-96ZM177,-581q51,-29 89,-75t57,-103q-51,29 -89,75t-57,103Zm146,-178Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_compare.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_compare.xml
new file mode 100644
index 000000000..d2985ff27
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_compare.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M320,-160l-56,-57 103,-103H80v-80h287L264,-503l56,-57 200,200 -200,200ZM640,-400L440,-600l200,-200 56,57 -103,103h287v80H593l103,103 -56,57Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_management.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_management.xml
new file mode 100644
index 000000000..1825e4345
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_management.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M287,-527q-47,-47 -47,-113t47,-113q47,-47 113,-47t113,47q47,47 47,113t-47,113q-47,47 -113,47t-113,-47ZM80,-160v-112q0,-33 17,-62t47,-44q51,-26 115,-44t141,-18h14q6,0 12,2 -8,18 -13.5,37.5T404,-360h-4q-71,0 -127.5,18T180,-306q-9,5 -14.5,14t-5.5,20v32h252q6,21 16,41.5t22,38.5H80ZM640,-120l-12,-60q-12,-5 -22.5,-10.5T584,-204l-58,18 -40,-68 46,-40q-2,-14 -2,-26t2,-26l-46,-40 40,-68 58,18q11,-8 21.5,-13.5T628,-460l12,-60h80l12,60q12,5 22.5,11t21.5,15l58,-20 40,70 -46,40q2,12 2,25t-2,25l46,40 -40,68 -58,-18q-11,8 -21.5,13.5T732,-180l-12,60h-80ZM736.5,-263.5Q760,-287 760,-320t-23.5,-56.5Q713,-400 680,-400t-56.5,23.5Q600,-353 600,-320t23.5,56.5Q647,-240 680,-240t56.5,-23.5ZM456.5,-583.5Q480,-607 480,-640t-23.5,-56.5Q433,-720 400,-720t-56.5,23.5Q320,-673 320,-640t23.5,56.5Q367,-560 400,-560t56.5,-23.5ZM400,-640ZM412,-240Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_pose.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_pose.xml
new file mode 100644
index 000000000..c692db655
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_pose.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M423.5,-743.5Q400,-767 400,-800t23.5,-56.5Q447,-880 480,-880t56.5,23.5Q560,-833 560,-800t-23.5,56.5Q513,-720 480,-720t-56.5,-23.5ZM360,-80v-520q-60,-5 -122,-15t-118,-25l20,-80q78,21 166,30.5t174,9.5q86,0 174,-9.5T820,-720l20,80q-56,15 -118,25t-122,15v520h-80v-240h-80v240h-80Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_recognition.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_recognition.xml
new file mode 100644
index 000000000..42ff9c7b6
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_recognition.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M440,-480q-66,0 -113,-47t-47,-113q0,-66 47,-113t113,-47q66,0 113,47t47,113q0,66 -47,113t-113,47ZM440,-560q33,0 56.5,-23.5T520,-640q0,-33 -23.5,-56.5T440,-720q-33,0 -56.5,23.5T360,-640q0,33 23.5,56.5T440,-560ZM884,-20L756,-148q-21,12 -45,20t-51,8q-75,0 -127.5,-52.5T480,-300q0,-75 52.5,-127.5T660,-480q75,0 127.5,52.5T840,-300q0,27 -8,51t-20,45L940,-76l-56,56ZM731,-229q29,-29 29,-71t-29,-71q-29,-29 -71,-29t-71,29q-29,29 -29,71t29,71q29,29 71,29t71,-29ZM120,-160v-111q0,-34 17,-63t47,-44q51,-26 115,-44t142,-18q-12,18 -20.5,38.5T407,-359q-60,5 -107,20.5T221,-306q-10,5 -15.5,14.5T200,-271v31h207q5,22 13.5,42t20.5,38H120ZM440,-640ZM407,-240Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_silent_liveness.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_silent_liveness.xml
new file mode 100644
index 000000000..953e7b530
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_silent_liveness.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M380.5,-480.5Q340,-521 340,-580t40.5,-99.5Q421,-720 480,-720t99.5,40.5Q620,-639 620,-580t-40.5,99.5Q539,-440 480,-440t-99.5,-40.5ZM523,-537q17,-17 17,-43t-17,-43q-17,-17 -43,-17t-43,17q-17,17 -17,43t17,43q17,17 43,17t43,-17ZM480,-80q-139,-35 -229.5,-159.5T160,-516v-244l320,-120 320,120v244q0,152 -90.5,276.5T480,-80ZM480,-480ZM480,-795l-240,90v189q0,54 15,105t41,96q42,-21 88,-33t96,-12q50,0 96,12t88,33q26,-45 41,-96t15,-105v-189l-240,-90ZM410,-272q-34,8 -65,22 29,30 63,52t72,34q38,-12 72,-34t63,-52q-31,-14 -65,-22t-70,-8q-36,0 -70,8Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_tracking.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_tracking.xml
new file mode 100644
index 000000000..0793f23dd
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_home_tracking.xml
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="32dp"
+    android:height="32dp"
+    android:viewportWidth="960"
+    android:viewportHeight="960">
+    <group android:translateY="960">
+        <path
+            android:fillColor="@color/home_tile_text_dark"
+            android:pathData="M367,-367q-47,-47 -47,-113t47,-113q47,-47 113,-47t113,47q47,47 47,113t-47,113q-47,47 -113,47t-113,-47ZM536.5,-423.5Q560,-447 560,-480t-23.5,-56.5Q513,-560 480,-560t-56.5,23.5Q400,-513 400,-480t23.5,56.5Q447,-400 480,-400t56.5,-23.5ZM480,-480ZM200,-120q-33,0 -56.5,-23.5T120,-200v-160h80v160h160v80H200ZM600,-120v-80h160v-160h80v160q0,33 -23.5,56.5T760,-120H600ZM120,-600v-160q0,-33 23.5,-56.5T200,-840h160v80H200v160h-80ZM760,-600v-160H600v-80h160q33,0 56.5,23.5T840,-760v160h-80Z" />
+    </group>
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_background.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_background.xml
new file mode 100644
index 000000000..6a01db9fc
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_background.xml
@@ -0,0 +1,5 @@
+<?xml version="1.0" encoding="utf-8"?>
+<shape xmlns:android="http://schemas.android.com/apk/res/android"
+    android:shape="rectangle">
+    <solid android:color="#0D53B4" />
+</shape>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_foreground.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_foreground.xml
new file mode 100644
index 000000000..c923d2d97
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/drawable/ic_launcher_foreground.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<bitmap xmlns:android="http://schemas.android.com/apk/res/android"
+    android:antialias="false"
+    android:filter="false"
+    android:gravity="fill"
+    android:src="@drawable/ic_launcher_inspireface_foreground" />
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_attribute.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_attribute.xml
new file mode 100644
index 000000000..472efcf36
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_attribute.xml
@@ -0,0 +1,266 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/attributeRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:clipToPadding="false"
+    android:fillViewport="true"
+    android:paddingStart="20dp"
+    android:paddingTop="12dp"
+    android:paddingEnd="20dp"
+    android:paddingBottom="28dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="40dp"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:text="@string/face_attribute_title"
+                android:textColor="@color/white"
+                android:textSize="17sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="24dp"
+            android:text="@string/face_attribute_heading"
+            android:textColor="@color/white"
+            android:textSize="28sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="6dp"
+            android:text="@string/face_attribute_hint"
+            android:textColor="@color/home_text_secondary"
+            android:textSize="14sp" />
+
+        <com.google.android.material.card.MaterialCardView
+            android:id="@+id/attributeImageCard"
+            android:layout_width="match_parent"
+            android:layout_height="340dp"
+            android:layout_marginTop="18dp"
+            android:clickable="true"
+            android:focusable="true"
+            android:foreground="?attr/selectableItemBackground"
+            app:cardBackgroundColor="@color/home_surface"
+            app:cardCornerRadius="20dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/home_stroke"
+            app:strokeWidth="1dp">
+
+            <FrameLayout
+                android:layout_width="match_parent"
+                android:layout_height="match_parent">
+
+                <ImageView
+                    android:id="@+id/attributeImage"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent"
+                    android:contentDescription="@string/attribute_photo"
+                    android:scaleType="centerCrop" />
+
+                <com.example.inspireface_example.widget.FaceImageOverlayView
+                    android:id="@+id/attributeFaceOverlay"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent" />
+
+                <TextView
+                    android:id="@+id/attributeImagePlaceholder"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent"
+                    android:gravity="center"
+                    android:padding="24dp"
+                    android:text="@string/attribute_choose_photo_hint"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="14sp" />
+
+                <com.google.android.material.card.MaterialCardView
+                    android:id="@+id/attributeMagnifierCard"
+                    android:layout_width="132dp"
+                    android:layout_height="132dp"
+                    android:layout_gravity="bottom|end"
+                    android:layout_marginEnd="12dp"
+                    android:layout_marginBottom="46dp"
+                    android:visibility="gone"
+                    app:cardBackgroundColor="@color/home_surface"
+                    app:cardCornerRadius="14dp"
+                    app:cardElevation="5dp"
+                    app:strokeColor="@color/liveness_accent"
+                    app:strokeWidth="2dp">
+
+                    <ImageView
+                        android:id="@+id/attributeMagnifierImage"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/attribute_selected_face_crop"
+                        android:scaleType="centerCrop" />
+                </com.google.android.material.card.MaterialCardView>
+
+                <TextView
+                    android:id="@+id/attributeImageStatus"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_gravity="bottom"
+                    android:background="@color/compare_status_scrim"
+                    android:gravity="center"
+                    android:padding="8dp"
+                    android:text="@string/no_image_selected"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="12sp" />
+            </FrameLayout>
+        </com.google.android.material.card.MaterialCardView>
+
+        <com.google.android.material.button.MaterialButton
+            android:id="@+id/btnChooseAttributePhoto"
+            android:layout_width="match_parent"
+            android:layout_height="48dp"
+            android:layout_marginTop="10dp"
+            android:enabled="false"
+            android:text="@string/attribute_choose_photo"
+            android:textColor="@color/home_tile_text_dark"
+            app:backgroundTint="@color/liveness_accent" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:gravity="center"
+            android:text="@string/attribute_multi_face_hint"
+            android:textColor="@color/home_text_muted"
+            android:textSize="12sp" />
+
+        <com.google.android.material.card.MaterialCardView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="18dp"
+            app:cardBackgroundColor="@color/home_surface"
+            app:cardCornerRadius="20dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/home_stroke"
+            app:strokeWidth="1dp">
+
+            <LinearLayout
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:orientation="vertical"
+                android:padding="18dp">
+
+                <TextView
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:text="@string/attribute_result_title"
+                    android:textColor="@color/white"
+                    android:textSize="18sp"
+                    android:textStyle="bold" />
+
+                <TextView
+                    android:id="@+id/attributeResultPlaceholder"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="16dp"
+                    android:gravity="center"
+                    android:paddingTop="20dp"
+                    android:paddingBottom="20dp"
+                    android:text="@string/attribute_result_empty"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="14sp" />
+
+                <LinearLayout
+                    android:id="@+id/attributeResultValues"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="10dp"
+                    android:orientation="vertical"
+                    android:visibility="gone">
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_mask" />
+                        <TextView android:id="@+id/attributeMaskValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_age" />
+                        <TextView android:id="@+id/attributeAgeValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_quality" />
+                        <TextView android:id="@+id/attributeQualityValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_expression" />
+                        <TextView android:id="@+id/attributeExpressionValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_race" />
+                        <TextView android:id="@+id/attributeRaceValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_gender" />
+                        <TextView android:id="@+id/attributeGenderValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_left_eye" />
+                        <TextView android:id="@+id/attributeLeftEyeValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+
+                    <LinearLayout style="@style/AttributeResultRow">
+                        <TextView style="@style/AttributeResultLabel" android:text="@string/attribute_right_eye" />
+                        <TextView android:id="@+id/attributeRightEyeValue" style="@style/AttributeResultValue" />
+                    </LinearLayout>
+                </LinearLayout>
+            </LinearLayout>
+        </com.google.android.material.card.MaterialCardView>
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_capture.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_capture.xml
new file mode 100644
index 000000000..11cd2bd21
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_capture.xml
@@ -0,0 +1,146 @@
+<?xml version="1.0" encoding="utf-8"?>
+<FrameLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/captureRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:keepScreenOn="true">
+
+    <androidx.camera.view.PreviewView
+        android:id="@+id/capturePreview"
+        android:layout_width="match_parent"
+        android:layout_height="match_parent" />
+
+    <com.example.inspireface_example.view.FaceOverlayView
+        android:id="@+id/captureFaceOverlay"
+        android:layout_width="match_parent"
+        android:layout_height="match_parent" />
+
+    <LinearLayout
+        android:id="@+id/captureTopBar"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_gravity="top"
+        android:gravity="center_horizontal"
+        android:orientation="vertical"
+        android:paddingStart="16dp"
+        android:paddingEnd="16dp">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="40dp"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:text="@string/face_capture_title"
+                android:textColor="@color/white"
+                android:textSize="16sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:id="@+id/btnFlipCaptureCamera"
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="10dp"
+            android:background="@drawable/bg_chip"
+            android:clickable="true"
+            android:focusable="true"
+            android:paddingStart="14dp"
+            android:paddingTop="7dp"
+            android:paddingEnd="14dp"
+            android:paddingBottom="7dp"
+            android:text="@string/flip_camera"
+            android:textColor="@color/white"
+            android:textSize="13sp" />
+    </LinearLayout>
+
+    <com.google.android.material.card.MaterialCardView
+        android:id="@+id/capturePromptCard"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_gravity="bottom"
+        android:layout_marginStart="16dp"
+        android:layout_marginEnd="16dp"
+        android:layout_marginBottom="24dp"
+        app:cardBackgroundColor="@color/liveness_scrim"
+        app:cardCornerRadius="20dp"
+        app:cardElevation="0dp"
+        app:strokeWidth="0dp">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center"
+            android:orientation="vertical"
+            android:padding="20dp">
+
+            <TextView
+                android:id="@+id/capturePromptTitle"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:gravity="center"
+                android:text="@string/msg_initializing"
+                android:textColor="@color/white"
+                android:textSize="22sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/capturePromptSubtitle"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="7dp"
+                android:gravity="center"
+                android:text="@string/capture_first_face_hint"
+                android:textColor="#B3FFFFFF"
+                android:textSize="14sp" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnCameraPermissionAction"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:text="@string/camera_permission_try_again"
+                android:textColor="@color/black"
+                android:visibility="gone"
+                app:backgroundTint="@color/liveness_accent" />
+        </LinearLayout>
+    </com.google.android.material.card.MaterialCardView>
+</FrameLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_compare.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_compare.xml
new file mode 100644
index 000000000..952d82fff
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_compare.xml
@@ -0,0 +1,372 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/compareRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:clipToPadding="false"
+    android:fillViewport="true"
+    android:paddingStart="20dp"
+    android:paddingTop="12dp"
+    android:paddingEnd="20dp"
+    android:paddingBottom="28dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="40dp"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:text="@string/face_compare_title"
+                android:textColor="@color/white"
+                android:textSize="17sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="26dp"
+            android:text="@string/face_compare_heading"
+            android:textColor="@color/white"
+            android:textSize="27sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="7dp"
+            android:lineSpacingExtra="2dp"
+            android:text="@string/face_compare_hint"
+            android:textColor="@color/home_text_secondary"
+            android:textSize="14sp" />
+
+        <androidx.constraintlayout.widget.ConstraintLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="24dp">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardImageA"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="20dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toStartOf="@id/cardImageB"
+                app:layout_constraintHorizontal_chainStyle="spread"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toTopOf="parent"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/imageA"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/image_a"
+                        android:scaleType="centerCrop" />
+
+                    <com.example.inspireface_example.widget.FaceImageOverlayView
+                        android:id="@+id/faceOverlayA"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <LinearLayout
+                        android:id="@+id/placeholderA"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:gravity="center"
+                        android:orientation="vertical"
+                        android:padding="12dp">
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:text="@string/add_symbol"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="34sp" />
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="2dp"
+                            android:text="@string/upload_image_a"
+                            android:textColor="@color/white"
+                            android:textSize="14sp"
+                            android:textStyle="bold" />
+                    </LinearLayout>
+
+                    <TextView
+                        android:id="@+id/statusA"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="bottom"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:maxLines="1"
+                        android:paddingStart="8dp"
+                        android:paddingTop="7dp"
+                        android:paddingEnd="8dp"
+                        android:paddingBottom="7dp"
+                        android:text="@string/image_empty"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="11sp" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardImageB"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginStart="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="20dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="parent"
+                app:layout_constraintStart_toEndOf="@id/cardImageA"
+                app:layout_constraintTop_toTopOf="parent"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/imageB"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/image_b"
+                        android:scaleType="centerCrop" />
+
+                    <com.example.inspireface_example.widget.FaceImageOverlayView
+                        android:id="@+id/faceOverlayB"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <LinearLayout
+                        android:id="@+id/placeholderB"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:gravity="center"
+                        android:orientation="vertical"
+                        android:padding="12dp">
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:text="@string/add_symbol"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="34sp" />
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="2dp"
+                            android:text="@string/upload_image_b"
+                            android:textColor="@color/white"
+                            android:textSize="14sp"
+                            android:textStyle="bold" />
+                    </LinearLayout>
+
+                    <TextView
+                        android:id="@+id/statusB"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="bottom"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:maxLines="1"
+                        android:paddingStart="8dp"
+                        android:paddingTop="7dp"
+                        android:paddingEnd="8dp"
+                        android:paddingBottom="7dp"
+                        android:text="@string/image_empty"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="11sp" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+        </androidx.constraintlayout.widget.ConstraintLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="28dp"
+            android:gravity="center"
+            android:text="@string/similarity_title"
+            android:textColor="@color/white"
+            android:textSize="17sp"
+            android:textStyle="bold" />
+
+        <androidx.constraintlayout.widget.ConstraintLayout
+            android:layout_width="match_parent"
+            android:layout_height="200dp"
+            android:layout_marginTop="10dp">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/compareCropCardA"
+                android:layout_width="58dp"
+                android:layout_height="58dp"
+                android:layout_marginEnd="6dp"
+                android:visibility="invisible"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="14dp"
+                app:cardElevation="3dp"
+                app:layout_constraintBottom_toBottomOf="@id/similarityGauge"
+                app:layout_constraintEnd_toStartOf="@id/similarityGauge"
+                app:layout_constraintHorizontal_chainStyle="packed"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toTopOf="@id/similarityGauge"
+                app:strokeColor="@color/liveness_accent"
+                app:strokeWidth="2dp"
+                tools:visibility="visible">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/compareCropA"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/compare_selected_face_a"
+                        android:scaleType="centerCrop" />
+
+                    <TextView
+                        android:layout_width="20dp"
+                        android:layout_height="20dp"
+                        android:layout_gravity="bottom|end"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:text="@string/compare_face_a_badge"
+                        android:textColor="@color/white"
+                        android:textSize="11sp"
+                        android:textStyle="bold" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.example.inspireface_example.widget.SimilarityGaugeView
+                android:id="@+id/similarityGauge"
+                android:layout_width="0dp"
+                android:layout_height="200dp"
+                app:layout_constraintEnd_toStartOf="@id/compareCropCardB"
+                app:layout_constraintStart_toEndOf="@id/compareCropCardA"
+                app:layout_constraintTop_toTopOf="parent"
+                app:layout_constraintWidth_max="200dp" />
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/compareCropCardB"
+                android:layout_width="58dp"
+                android:layout_height="58dp"
+                android:layout_marginStart="6dp"
+                android:visibility="invisible"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="14dp"
+                app:cardElevation="3dp"
+                app:layout_constraintBottom_toBottomOf="@id/similarityGauge"
+                app:layout_constraintEnd_toEndOf="parent"
+                app:layout_constraintStart_toEndOf="@id/similarityGauge"
+                app:layout_constraintTop_toTopOf="@id/similarityGauge"
+                app:strokeColor="@color/liveness_accent"
+                app:strokeWidth="2dp"
+                tools:visibility="visible">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/compareCropB"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/compare_selected_face_b"
+                        android:scaleType="centerCrop" />
+
+                    <TextView
+                        android:layout_width="20dp"
+                        android:layout_height="20dp"
+                        android:layout_gravity="bottom|end"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:text="@string/compare_face_b_badge"
+                        android:textColor="@color/white"
+                        android:textSize="11sp"
+                        android:textStyle="bold" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+        </androidx.constraintlayout.widget.ConstraintLayout>
+
+        <TextView
+            android:id="@+id/comparisonDetails"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center"
+            android:minHeight="20dp"
+            android:textColor="@color/home_text_muted"
+            android:textSize="11sp"
+            tools:text="cosine confidence 0.82 · threshold 0.45" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="14dp"
+            android:gravity="center"
+            android:text="@string/on_device_privacy"
+            android:textColor="@color/home_text_muted"
+            android:textSize="12sp" />
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_detection.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_detection.xml
new file mode 100644
index 000000000..d6beefbf9
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_detection.xml
@@ -0,0 +1,628 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/detectionRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:clipToPadding="false"
+    android:fillViewport="true"
+    android:paddingStart="20dp"
+    android:paddingTop="12dp"
+    android:paddingEnd="20dp"
+    android:paddingBottom="28dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="40dp"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:text="@string/face_detection_title"
+                android:textColor="@color/white"
+                android:textSize="17sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="24dp"
+            android:text="@string/face_detection_heading"
+            android:textColor="@color/white"
+            android:textSize="28sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="6dp"
+            android:text="@string/face_detection_hint"
+            android:textColor="@color/home_text_secondary"
+            android:textSize="14sp" />
+
+        <com.google.android.material.tabs.TabLayout
+            android:id="@+id/detectionTabs"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="20dp"
+            app:tabIndicatorColor="@color/liveness_accent"
+            app:tabSelectedTextColor="@color/liveness_accent"
+            app:tabTextColor="@color/home_text_secondary">
+
+            <com.google.android.material.tabs.TabItem
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/detection_tab_image" />
+
+            <com.google.android.material.tabs.TabItem
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/detection_tab_tracking" />
+        </com.google.android.material.tabs.TabLayout>
+
+        <LinearLayout
+            android:id="@+id/imageDetectionPanel"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:orientation="vertical">
+
+            <com.google.android.material.card.MaterialCardView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="16dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <com.google.android.material.switchmaterial.SwitchMaterial
+                    android:id="@+id/switchDenseLandmarks"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:checked="true"
+                    android:paddingStart="16dp"
+                    android:paddingTop="10dp"
+                    android:paddingEnd="16dp"
+                    android:paddingBottom="10dp"
+                    android:text="@string/detection_show_dense_landmarks"
+                    android:textColor="@color/white"
+                    android:textSize="14sp" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/detectionImageCard"
+                android:layout_width="match_parent"
+                android:layout_height="360dp"
+                android:layout_marginTop="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="20dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/detectionImage"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/detection_photo"
+                        android:scaleType="centerCrop" />
+
+                    <com.example.inspireface_example.widget.FaceImageOverlayView
+                        android:id="@+id/detectionFaceOverlay"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <com.example.inspireface_example.widget.FaceLandmarkOverlayView
+                        android:id="@+id/detectionLandmarkOverlay"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <TextView
+                        android:id="@+id/detectionImagePlaceholder"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:gravity="center"
+                        android:padding="24dp"
+                        android:text="@string/detection_choose_photo_hint"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="14sp" />
+
+                    <com.google.android.material.card.MaterialCardView
+                        android:id="@+id/landmarkMagnifierCard"
+                        android:layout_width="148dp"
+                        android:layout_height="148dp"
+                        android:layout_gravity="bottom|end"
+                        android:layout_marginEnd="12dp"
+                        android:layout_marginBottom="48dp"
+                        android:visibility="gone"
+                        app:cardBackgroundColor="@color/home_surface"
+                        app:cardCornerRadius="14dp"
+                        app:cardElevation="5dp"
+                        app:strokeColor="@color/liveness_accent"
+                        app:strokeWidth="2dp">
+
+                        <FrameLayout
+                            android:layout_width="match_parent"
+                            android:layout_height="match_parent">
+
+                            <ImageView
+                                android:id="@+id/landmarkMagnifierImage"
+                                android:layout_width="match_parent"
+                                android:layout_height="match_parent"
+                                android:contentDescription="@string/detection_landmark_magnifier"
+                                android:scaleType="centerCrop" />
+
+                            <com.example.inspireface_example.widget.FaceLandmarkOverlayView
+                                android:id="@+id/landmarkMagnifierOverlay"
+                                android:layout_width="match_parent"
+                                android:layout_height="match_parent" />
+
+                            <TextView
+                                android:id="@+id/landmarkMagnifierBadge"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:layout_gravity="top|start"
+                                android:layout_margin="7dp"
+                                android:background="@drawable/bg_chip"
+                                android:paddingStart="7dp"
+                                android:paddingTop="3dp"
+                                android:paddingEnd="7dp"
+                                android:paddingBottom="3dp"
+                                android:text="@string/detection_106_badge"
+                                android:textColor="@color/liveness_accent"
+                                android:textSize="11sp"
+                                android:textStyle="bold" />
+                        </FrameLayout>
+                    </com.google.android.material.card.MaterialCardView>
+
+                    <TextView
+                        android:id="@+id/detectionImageStatus"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="bottom"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:padding="8dp"
+                        android:text="@string/no_image_selected"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="12sp" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnChooseDetectionPhoto"
+                android:layout_width="match_parent"
+                android:layout_height="48dp"
+                android:layout_marginTop="10dp"
+                android:enabled="false"
+                android:text="@string/detection_choose_photo"
+                android:textColor="@color/home_tile_text_dark"
+                app:backgroundTint="@color/liveness_accent" />
+
+            <TextView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="7dp"
+                android:gravity="center"
+                android:text="@string/detection_multi_face_hint"
+                android:textColor="@color/home_text_muted"
+                android:textSize="12sp" />
+
+            <TextView
+                android:id="@+id/landmarkDisplayHint"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="5dp"
+                android:gravity="center"
+                android:text="@string/detection_landmark_behavior_hint"
+                android:textColor="@color/home_text_secondary"
+                android:textSize="12sp" />
+
+            <com.google.android.material.card.MaterialCardView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="18dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <LinearLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:orientation="vertical">
+
+                    <LinearLayout
+                        android:id="@+id/sessionSettingsHeader"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:background="?attr/selectableItemBackground"
+                        android:clickable="true"
+                        android:focusable="true"
+                        android:gravity="center_vertical"
+                        android:minHeight="64dp"
+                        android:orientation="horizontal"
+                        android:padding="16dp">
+
+                        <LinearLayout
+                            android:layout_width="0dp"
+                            android:layout_height="wrap_content"
+                            android:layout_weight="1"
+                            android:orientation="vertical">
+
+                            <TextView
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:text="@string/recognition_session_parameters"
+                                android:textColor="@color/white"
+                                android:textSize="16sp"
+                                android:textStyle="bold" />
+
+                            <TextView
+                                android:id="@+id/sessionStatus"
+                                android:layout_width="match_parent"
+                                android:layout_height="wrap_content"
+                                android:layout_marginTop="3dp"
+                                android:textColor="@color/home_text_secondary"
+                                android:textSize="11sp"
+                                tools:text="Session · 640 px · max 10 · min 24 px" />
+                        </LinearLayout>
+
+                        <TextView
+                            android:id="@+id/sessionSettingsToggleText"
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginStart="12dp"
+                            android:text="@string/recognition_settings_expand"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="13sp" />
+                    </LinearLayout>
+
+                    <LinearLayout
+                        android:id="@+id/sessionSettingsContent"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:orientation="vertical"
+                        android:paddingStart="16dp"
+                        android:paddingEnd="16dp"
+                        android:paddingBottom="16dp"
+                        android:visibility="gone">
+
+                        <TextView
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:text="@string/recognition_session_parameters_hint"
+                            android:textColor="@color/home_text_secondary"
+                            android:textSize="12sp" />
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="14dp"
+                            android:text="@string/recognition_input_px"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/inputPxGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/inputPx640"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx320"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_320" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx640"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_640" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx1280"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_1280" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="10dp"
+                            android:text="@string/recognition_max_faces"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/maxFacesGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/maxFaces10"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces1"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_1" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces3"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_3" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces5"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_5" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces10"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_10" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="10dp"
+                            android:text="@string/recognition_min_face_px"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/minFaceGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/minFace24"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace24"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_24" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace48"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_48" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace64"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_64" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace128"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_128" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <LinearLayout
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="12dp"
+                            android:orientation="horizontal">
+
+                            <com.google.android.material.button.MaterialButton
+                                android:id="@+id/btnApplySession"
+                                android:layout_width="0dp"
+                                android:layout_height="48dp"
+                                android:layout_weight="1"
+                                android:text="@string/recognition_apply_session"
+                                android:textColor="@color/home_tile_text_dark"
+                                app:backgroundTint="@color/liveness_accent" />
+
+                            <com.google.android.material.button.MaterialButton
+                                android:id="@+id/btnResetSession"
+                                style="?attr/materialButtonOutlinedStyle"
+                                android:layout_width="0dp"
+                                android:layout_height="48dp"
+                                android:layout_marginStart="8dp"
+                                android:layout_weight="1"
+                                android:text="@string/recognition_reset_defaults"
+                                android:textColor="@color/white"
+                                app:strokeColor="@color/liveness_accent" />
+                        </LinearLayout>
+                    </LinearLayout>
+                </LinearLayout>
+            </com.google.android.material.card.MaterialCardView>
+        </LinearLayout>
+
+        <LinearLayout
+            android:id="@+id/videoTrackingPanel"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="16dp"
+            android:keepScreenOn="true"
+            android:orientation="vertical"
+            android:visibility="gone">
+
+            <com.google.android.material.card.MaterialCardView
+                android:layout_width="match_parent"
+                android:layout_height="500dp"
+                app:cardBackgroundColor="@color/black"
+                app:cardCornerRadius="20dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <androidx.camera.view.PreviewView
+                        android:id="@+id/detectionTrackingPreview"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <com.example.inspireface_example.view.FaceTrackingGlView
+                        android:id="@+id/detectionTrackingGlOverlay"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <TextView
+                        android:id="@+id/btnFlipTrackingCamera"
+                        android:layout_width="wrap_content"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="top|end"
+                        android:layout_margin="14dp"
+                        android:background="@drawable/bg_chip"
+                        android:clickable="true"
+                        android:focusable="true"
+                        android:paddingStart="14dp"
+                        android:paddingTop="7dp"
+                        android:paddingEnd="14dp"
+                        android:paddingBottom="7dp"
+                        android:text="@string/flip_camera"
+                        android:textColor="@color/white"
+                        android:textSize="13sp" />
+
+                    <com.google.android.material.card.MaterialCardView
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="bottom"
+                        android:layout_margin="14dp"
+                        app:cardBackgroundColor="@color/liveness_scrim"
+                        app:cardCornerRadius="16dp"
+                        app:cardElevation="0dp"
+                        app:strokeWidth="0dp">
+
+                        <LinearLayout
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:gravity="center"
+                            android:orientation="vertical"
+                            android:padding="14dp">
+
+                            <com.google.android.material.progressindicator.CircularProgressIndicator
+                                android:id="@+id/trackingLoadingIndicator"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:layout_marginBottom="8dp"
+                                android:indeterminate="true"
+                                app:indicatorColor="@color/liveness_accent"
+                                app:indicatorSize="24dp"
+                                app:trackThickness="3dp" />
+
+                            <TextView
+                                android:id="@+id/detectionTrackingStatus"
+                                android:layout_width="match_parent"
+                                android:layout_height="wrap_content"
+                                android:gravity="center"
+                                android:text="@string/detection_tracking_initializing"
+                                android:textColor="@color/white"
+                                android:textSize="15sp"
+                                android:textStyle="bold" />
+
+                            <com.google.android.material.button.MaterialButton
+                                android:id="@+id/btnCameraPermissionAction"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:layout_marginTop="12dp"
+                                android:text="@string/camera_permission_try_again"
+                                android:textColor="@color/black"
+                                android:visibility="gone"
+                                app:backgroundTint="@color/liveness_accent" />
+                        </LinearLayout>
+                    </com.google.android.material.card.MaterialCardView>
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <include
+                layout="@layout/view_face_tracking_settings"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp" />
+        </LinearLayout>
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_management.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_management.xml
new file mode 100644
index 000000000..ea9570dae
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_management.xml
@@ -0,0 +1,171 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/managementRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:orientation="vertical"
+    android:paddingStart="20dp"
+    android:paddingTop="12dp"
+    android:paddingEnd="20dp"
+    android:paddingBottom="16dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <TextView
+            android:id="@+id/btnBack"
+            android:layout_width="40dp"
+            android:layout_height="40dp"
+            android:background="@drawable/bg_chip"
+            android:clickable="true"
+            android:focusable="true"
+            android:gravity="center"
+            android:text="@string/back"
+            android:textColor="@color/white"
+            android:textSize="18sp" />
+
+        <TextView
+            android:layout_width="0dp"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:layout_marginEnd="8dp"
+            android:layout_weight="1"
+            android:gravity="center"
+            android:maxLines="1"
+            android:text="@string/face_management_title"
+            android:textColor="@color/white"
+            android:textSize="17sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:id="@+id/currentModel"
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:background="@drawable/bg_chip"
+            android:gravity="center"
+            android:maxLines="1"
+            android:paddingStart="10dp"
+            android:paddingTop="5dp"
+            android:paddingEnd="10dp"
+            android:paddingBottom="5dp"
+            android:textColor="@color/liveness_accent"
+            android:textSize="11sp"
+            tools:text="Model · Megatron" />
+    </LinearLayout>
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="24dp"
+        android:text="@string/face_library_heading"
+        android:textColor="@color/white"
+        android:textSize="28sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:id="@+id/storageScope"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="6dp"
+        android:textColor="@color/home_text_secondary"
+        android:textSize="13sp"
+        tools:text="Megatron uses an isolated feature DB and crop directory." />
+
+    <TextView
+        android:id="@+id/faceCount"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:textColor="@color/liveness_accent"
+        android:textSize="12sp"
+        tools:text="12 stored identities" />
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="20dp"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <com.google.android.material.textfield.TextInputLayout
+            style="@style/Widget.Material3.TextInputLayout.OutlinedBox"
+            android:layout_width="0dp"
+            android:layout_height="wrap_content"
+            android:layout_weight="1"
+            android:hint="@string/search_faces">
+
+            <com.google.android.material.textfield.TextInputEditText
+                android:id="@+id/searchInput"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:imeOptions="actionDone"
+                android:inputType="text"
+                android:maxLines="1"
+                android:textColor="@color/white" />
+        </com.google.android.material.textfield.TextInputLayout>
+
+        <com.google.android.material.button.MaterialButton
+            android:id="@+id/btnAddFace"
+            android:layout_width="wrap_content"
+            android:layout_height="56dp"
+            android:layout_marginStart="10dp"
+            android:enabled="false"
+            android:text="@string/add_face"
+            android:textColor="@color/home_tile_text_dark"
+            app:backgroundTint="@color/liveness_accent" />
+    </LinearLayout>
+
+    <FrameLayout
+        android:layout_width="match_parent"
+        android:layout_height="0dp"
+        android:layout_marginTop="12dp"
+        android:layout_weight="1">
+
+        <ListView
+            android:id="@+id/faceList"
+            android:layout_width="match_parent"
+            android:layout_height="match_parent"
+            android:clipToPadding="false"
+            android:divider="@android:color/transparent"
+            android:dividerHeight="0dp"
+            android:paddingBottom="8dp" />
+
+        <LinearLayout
+            android:id="@+id/emptyState"
+            android:layout_width="match_parent"
+            android:layout_height="match_parent"
+            android:gravity="center"
+            android:orientation="vertical"
+            android:padding="28dp">
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/no_faces_title"
+                android:textColor="@color/white"
+                android:textSize="18sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="6dp"
+                android:gravity="center"
+                android:text="@string/no_faces_hint"
+                android:textColor="@color/home_text_secondary"
+                android:textSize="13sp" />
+        </LinearLayout>
+
+        <com.google.android.material.progressindicator.CircularProgressIndicator
+            android:id="@+id/loadingIndicator"
+            android:layout_width="48dp"
+            android:layout_height="48dp"
+            android:layout_gravity="center"
+            app:indicatorColor="@color/liveness_accent" />
+    </FrameLayout>
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_recognition.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_recognition.xml
new file mode 100644
index 000000000..5b9bf168b
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_face_recognition.xml
@@ -0,0 +1,650 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/recognitionRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:clipToPadding="false"
+    android:fillViewport="true"
+    android:keepScreenOn="true"
+    android:paddingStart="20dp"
+    android:paddingTop="12dp"
+    android:paddingEnd="20dp"
+    android:paddingBottom="28dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="40dp"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:text="@string/face_recognition_title"
+                android:textColor="@color/white"
+                android:textSize="17sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="24dp"
+            android:text="@string/face_recognition_heading"
+            android:textColor="@color/white"
+            android:textSize="28sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="6dp"
+            android:text="@string/face_recognition_hint"
+            android:textColor="@color/home_text_secondary"
+            android:textSize="14sp" />
+
+        <com.google.android.material.tabs.TabLayout
+            android:id="@+id/recognitionTabs"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="20dp"
+            app:tabIndicatorColor="@color/liveness_accent"
+            app:tabSelectedTextColor="@color/liveness_accent"
+            app:tabTextColor="@color/home_text_secondary">
+
+            <com.google.android.material.tabs.TabItem
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/recognition_tab_photo" />
+
+            <com.google.android.material.tabs.TabItem
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/recognition_tab_video" />
+        </com.google.android.material.tabs.TabLayout>
+
+        <com.google.android.material.card.MaterialCardView
+            android:id="@+id/emptyLibraryTip"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="14dp"
+            android:visibility="gone"
+            app:cardBackgroundColor="@color/home_surface"
+            app:cardCornerRadius="14dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/liveness_warn"
+            app:strokeWidth="1dp">
+
+            <TextView
+                android:id="@+id/emptyLibraryTipText"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:padding="14dp"
+                android:textColor="@color/liveness_warn"
+                android:textSize="13sp"
+                tools:text="No identities are enrolled for Megatron." />
+        </com.google.android.material.card.MaterialCardView>
+
+        <LinearLayout
+            android:id="@+id/photoRecognitionPanel"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:orientation="vertical">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/recognitionImageCard"
+                android:layout_width="match_parent"
+                android:layout_height="300dp"
+                android:layout_marginTop="14dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="20dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <FrameLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent">
+
+                    <ImageView
+                        android:id="@+id/recognitionImage"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:contentDescription="@string/recognition_photo"
+                        android:scaleType="centerCrop" />
+
+                    <com.example.inspireface_example.widget.FaceImageOverlayView
+                        android:id="@+id/recognitionFaceOverlay"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent" />
+
+                    <TextView
+                        android:id="@+id/recognitionImagePlaceholder"
+                        android:layout_width="match_parent"
+                        android:layout_height="match_parent"
+                        android:gravity="center"
+                        android:padding="24dp"
+                        android:text="@string/recognition_choose_photo_hint"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="14sp" />
+
+                    <com.google.android.material.card.MaterialCardView
+                        android:id="@+id/recognitionSelectedFaceCropCard"
+                        android:layout_width="82dp"
+                        android:layout_height="82dp"
+                        android:layout_gravity="bottom|end"
+                        android:layout_marginEnd="12dp"
+                        android:layout_marginBottom="46dp"
+                        android:visibility="gone"
+                        app:cardBackgroundColor="@color/home_surface"
+                        app:cardCornerRadius="12dp"
+                        app:cardElevation="4dp"
+                        app:strokeColor="@color/liveness_accent"
+                        app:strokeWidth="2dp">
+
+                        <ImageView
+                            android:id="@+id/recognitionSelectedFaceCrop"
+                            android:layout_width="match_parent"
+                            android:layout_height="match_parent"
+                            android:contentDescription="@string/recognition_selected_face_crop"
+                            android:scaleType="centerCrop" />
+                    </com.google.android.material.card.MaterialCardView>
+
+                    <TextView
+                        android:id="@+id/recognitionImageStatus"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:layout_gravity="bottom"
+                        android:background="@color/compare_status_scrim"
+                        android:gravity="center"
+                        android:padding="8dp"
+                        android:text="@string/no_image_selected"
+                        android:textColor="@color/home_text_secondary"
+                        android:textSize="12sp" />
+                </FrameLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnChooseRecognitionPhoto"
+                android:layout_width="match_parent"
+                android:layout_height="48dp"
+                android:layout_marginTop="10dp"
+                android:enabled="false"
+                android:text="@string/recognition_choose_photo"
+                android:textColor="@color/home_tile_text_dark"
+                app:backgroundTint="@color/liveness_accent" />
+
+            <TextView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="7dp"
+                android:gravity="center"
+                android:text="@string/recognition_multi_face_hint"
+                android:textColor="@color/home_text_muted"
+                android:textSize="12sp" />
+
+            <com.google.android.material.card.MaterialCardView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="18dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <LinearLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:orientation="vertical">
+
+                    <LinearLayout
+                        android:id="@+id/sessionSettingsHeader"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:background="?attr/selectableItemBackground"
+                        android:clickable="true"
+                        android:focusable="true"
+                        android:gravity="center_vertical"
+                        android:minHeight="64dp"
+                        android:orientation="horizontal"
+                        android:padding="16dp">
+
+                        <LinearLayout
+                            android:layout_width="0dp"
+                            android:layout_height="wrap_content"
+                            android:layout_weight="1"
+                            android:orientation="vertical">
+
+                            <TextView
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:text="@string/recognition_session_parameters"
+                                android:textColor="@color/white"
+                                android:textSize="16sp"
+                                android:textStyle="bold" />
+
+                            <TextView
+                                android:id="@+id/sessionStatus"
+                                android:layout_width="match_parent"
+                                android:layout_height="wrap_content"
+                                android:layout_marginTop="3dp"
+                                android:textColor="@color/home_text_secondary"
+                                android:textSize="11sp"
+                                tools:text="Session · 640 px · max 10 · min 24 px" />
+                        </LinearLayout>
+
+                        <TextView
+                            android:id="@+id/sessionSettingsToggleText"
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginStart="12dp"
+                            android:text="@string/recognition_settings_expand"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="13sp" />
+                    </LinearLayout>
+
+                    <LinearLayout
+                        android:id="@+id/sessionSettingsContent"
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:orientation="vertical"
+                        android:paddingStart="16dp"
+                        android:paddingEnd="16dp"
+                        android:paddingBottom="16dp"
+                        android:visibility="gone">
+
+                        <TextView
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:text="@string/recognition_session_parameters_hint"
+                            android:textColor="@color/home_text_secondary"
+                            android:textSize="12sp" />
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="14dp"
+                            android:text="@string/recognition_input_px"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/inputPxGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/inputPx640"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx320"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_320" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx640"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_640" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/inputPx1280"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_1280" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="10dp"
+                            android:text="@string/recognition_max_faces"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/maxFacesGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/maxFaces10"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces1"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_1" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces3"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_3" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces5"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_5" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/maxFaces10"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_10" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <TextView
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="10dp"
+                            android:text="@string/recognition_min_face_px"
+                            android:textColor="@color/white"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.chip.ChipGroup
+                            android:id="@+id/minFaceGroup"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            app:checkedChip="@id/minFace24"
+                            app:selectionRequired="true"
+                            app:singleSelection="true">
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace24"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_24" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace48"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_48" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace64"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_64" />
+
+                            <com.google.android.material.chip.Chip
+                                android:id="@+id/minFace128"
+                                style="@style/Widget.Material3.Chip.Filter"
+                                android:layout_width="wrap_content"
+                                android:layout_height="wrap_content"
+                                android:checkable="true"
+                                android:text="@string/value_128" />
+                        </com.google.android.material.chip.ChipGroup>
+
+                        <LinearLayout
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="12dp"
+                            android:orientation="horizontal">
+
+                            <com.google.android.material.button.MaterialButton
+                                android:id="@+id/btnApplySession"
+                                android:layout_width="0dp"
+                                android:layout_height="48dp"
+                                android:layout_weight="1"
+                                android:text="@string/recognition_apply_session"
+                                android:textColor="@color/home_tile_text_dark"
+                                app:backgroundTint="@color/liveness_accent" />
+
+                            <com.google.android.material.button.MaterialButton
+                                android:id="@+id/btnResetSession"
+                                style="?attr/materialButtonOutlinedStyle"
+                                android:layout_width="0dp"
+                                android:layout_height="48dp"
+                                android:layout_marginStart="8dp"
+                                android:layout_weight="1"
+                                android:text="@string/recognition_reset_defaults"
+                                android:textColor="@color/white"
+                                app:strokeColor="@color/liveness_accent" />
+                        </LinearLayout>
+                    </LinearLayout>
+                </LinearLayout>
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/recognitionResultCard"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:visibility="gone"
+                app:cardBackgroundColor="@color/home_surface"
+                app:cardCornerRadius="18dp"
+                app:cardElevation="0dp"
+                app:strokeColor="@color/home_stroke"
+                app:strokeWidth="1dp">
+
+                <LinearLayout
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:gravity="center_vertical"
+                    android:orientation="horizontal"
+                    android:padding="16dp">
+
+                    <ImageView
+                        android:id="@+id/recognitionResultCrop"
+                        android:layout_width="76dp"
+                        android:layout_height="76dp"
+                        android:contentDescription="@string/stored_face_crop"
+                        android:scaleType="centerCrop"
+                        android:visibility="gone" />
+
+                    <LinearLayout
+                        android:layout_width="0dp"
+                        android:layout_height="wrap_content"
+                        android:layout_marginStart="14dp"
+                        android:layout_weight="1"
+                        android:orientation="vertical">
+
+                        <TextView
+                            android:id="@+id/recognitionResultStatus"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="13sp"
+                            android:textStyle="bold"
+                            tools:text="Match found" />
+
+                        <TextView
+                            android:id="@+id/recognitionResultName"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="4dp"
+                            android:textColor="@color/white"
+                            android:textSize="20sp"
+                            android:textStyle="bold"
+                            tools:text="Ada" />
+
+                        <TextView
+                            android:id="@+id/recognitionResultDetails"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            android:textColor="@color/home_text_secondary"
+                            android:textSize="12sp"
+                            tools:text="ID · 7 · cosine confidence 0.823 · threshold 0.450" />
+                    </LinearLayout>
+                </LinearLayout>
+            </com.google.android.material.card.MaterialCardView>
+        </LinearLayout>
+
+        <com.google.android.material.card.MaterialCardView
+            android:id="@+id/videoRecognitionPanel"
+            android:layout_width="match_parent"
+            android:layout_height="520dp"
+            android:layout_marginTop="16dp"
+            android:visibility="gone"
+            app:cardBackgroundColor="@color/black"
+            app:cardCornerRadius="20dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/home_stroke"
+            app:strokeWidth="1dp">
+
+            <FrameLayout
+                android:layout_width="match_parent"
+                android:layout_height="match_parent">
+
+                <androidx.camera.view.PreviewView
+                    android:id="@+id/recognitionVideoPreview"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent" />
+
+                <com.example.inspireface_example.view.FaceOverlayView
+                    android:id="@+id/recognitionVideoOverlay"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent" />
+
+                <TextView
+                    android:id="@+id/btnFlipRecognitionCamera"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:layout_gravity="top|end"
+                    android:layout_margin="14dp"
+                    android:background="@drawable/bg_chip"
+                    android:clickable="true"
+                    android:focusable="true"
+                    android:paddingStart="14dp"
+                    android:paddingTop="7dp"
+                    android:paddingEnd="14dp"
+                    android:paddingBottom="7dp"
+                    android:text="@string/flip_camera"
+                    android:textColor="@color/white"
+                    android:textSize="13sp" />
+
+                <com.google.android.material.card.MaterialCardView
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_gravity="bottom"
+                    android:layout_margin="14dp"
+                    app:cardBackgroundColor="@color/liveness_scrim"
+                    app:cardCornerRadius="18dp"
+                    app:cardElevation="0dp"
+                    app:strokeWidth="0dp">
+
+                    <LinearLayout
+                        android:layout_width="match_parent"
+                        android:layout_height="wrap_content"
+                        android:gravity="center"
+                        android:orientation="vertical"
+                        android:padding="18dp">
+
+                        <TextView
+                            android:id="@+id/videoRecognitionStatus"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:gravity="center"
+                            android:text="@string/recognition_video_initializing"
+                            android:textColor="@color/white"
+                            android:textSize="19sp"
+                            android:textStyle="bold" />
+
+                        <TextView
+                            android:id="@+id/videoRecognitionName"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="4dp"
+                            android:gravity="center"
+                            android:textColor="@color/liveness_accent"
+                            android:textSize="16sp"
+                            android:textStyle="bold"
+                            android:visibility="gone"
+                            tools:text="Ada" />
+
+                        <TextView
+                            android:id="@+id/videoRecognitionDetails"
+                            android:layout_width="match_parent"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="5dp"
+                            android:gravity="center"
+                            android:text="@string/recognition_video_first_face_hint"
+                            android:textColor="#B3FFFFFF"
+                            android:textSize="13sp" />
+
+                        <com.google.android.material.button.MaterialButton
+                            android:id="@+id/btnCameraPermissionAction"
+                            android:layout_width="wrap_content"
+                            android:layout_height="wrap_content"
+                            android:layout_marginTop="12dp"
+                            android:text="@string/camera_permission_try_again"
+                            android:textColor="@color/black"
+                            android:visibility="gone"
+                            app:backgroundTint="@color/liveness_accent" />
+                    </LinearLayout>
+                </com.google.android.material.card.MaterialCardView>
+            </FrameLayout>
+        </com.google.android.material.card.MaterialCardView>
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_home.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_home.xml
new file mode 100644
index 000000000..d862d9e0f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_home.xml
@@ -0,0 +1,332 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/homeRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:clipToPadding="false"
+    android:fillViewport="true"
+    android:paddingStart="24dp"
+    android:paddingTop="16dp"
+    android:paddingEnd="24dp"
+    android:paddingBottom="24dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_weight="1"
+                android:text="@string/app_name"
+                android:textColor="@color/home_text_secondary"
+                android:textSize="14sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/langSwitch"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:minHeight="40dp"
+                android:gravity="center"
+                android:paddingStart="16dp"
+                android:paddingEnd="16dp"
+                android:text="@string/lang_switch"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+        </LinearLayout>
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="28dp"
+            android:text="@string/home_title"
+            android:textColor="@color/white"
+            android:textSize="34sp"
+            android:textStyle="bold" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:lineSpacingExtra="2dp"
+            android:text="@string/home_subtitle"
+            android:textColor="@color/home_text_secondary"
+            android:textSize="15sp" />
+
+        <com.google.android.material.card.MaterialCardView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="28dp"
+            app:cardBackgroundColor="@color/home_surface"
+            app:cardCornerRadius="20dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/home_stroke"
+            app:strokeWidth="1dp">
+
+            <LinearLayout
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:orientation="vertical"
+                android:padding="18dp">
+
+                <TextView
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:text="@string/model_section_title"
+                    android:textColor="@color/white"
+                    android:textSize="17sp"
+                    android:textStyle="bold" />
+
+                <TextView
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="4dp"
+                    android:text="@string/model_section_hint"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="12sp" />
+
+                <com.google.android.material.button.MaterialButtonToggleGroup
+                    android:id="@+id/modelToggle"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="16dp"
+                    app:checkedButton="@id/btnModelMegatron"
+                    app:selectionRequired="true"
+                    app:singleSelection="true">
+
+                    <com.google.android.material.button.MaterialButton
+                        android:id="@+id/btnModelPikachu"
+                        style="?attr/materialButtonOutlinedStyle"
+                        android:layout_width="0dp"
+                        android:layout_height="48dp"
+                        android:layout_weight="1"
+                        android:text="@string/model_pikachu"
+                        android:textAllCaps="false"
+                        android:textColor="@color/white"
+                        app:strokeColor="@color/liveness_accent" />
+
+                    <com.google.android.material.button.MaterialButton
+                        android:id="@+id/btnModelMegatron"
+                        style="?attr/materialButtonOutlinedStyle"
+                        android:layout_width="0dp"
+                        android:layout_height="48dp"
+                        android:layout_weight="1"
+                        android:text="@string/model_megatron"
+                        android:textAllCaps="false"
+                        android:textColor="@color/white"
+                        app:strokeColor="@color/liveness_accent" />
+                </com.google.android.material.button.MaterialButtonToggleGroup>
+            </LinearLayout>
+        </com.google.android.material.card.MaterialCardView>
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="30dp"
+            android:layout_marginBottom="12dp"
+            android:text="@string/home_section_analysis"
+            android:textColor="@color/white"
+            android:textSize="20sp"
+            android:textStyle="bold" />
+
+        <androidx.constraintlayout.widget.ConstraintLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardDetection"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_yellow"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toStartOf="@id/cardAttribute"
+                app:layout_constraintHorizontal_chainStyle="spread"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_detection" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardAttribute"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginStart="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_lime"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="parent"
+                app:layout_constraintStart_toEndOf="@id/cardDetection"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_attribute" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardPose"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginTop="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_purple"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="@id/cardDetection"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toBottomOf="@id/cardDetection">
+
+                <include layout="@layout/item_home_pose" />
+            </com.google.android.material.card.MaterialCardView>
+        </androidx.constraintlayout.widget.ConstraintLayout>
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="28dp"
+            android:layout_marginBottom="12dp"
+            android:text="@string/home_section_recognition"
+            android:textColor="@color/white"
+            android:textSize="20sp"
+            android:textStyle="bold" />
+
+        <androidx.constraintlayout.widget.ConstraintLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardCompare"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_orange"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toStartOf="@id/cardRecognition"
+                app:layout_constraintHorizontal_chainStyle="spread"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_compare" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardRecognition"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginStart="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_cyan"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="parent"
+                app:layout_constraintStart_toEndOf="@id/cardCompare"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_recognition" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardManagement"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginTop="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_pink"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="@id/cardCompare"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toBottomOf="@id/cardCompare">
+
+                <include layout="@layout/item_home_management" />
+            </com.google.android.material.card.MaterialCardView>
+        </androidx.constraintlayout.widget.ConstraintLayout>
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="28dp"
+            android:layout_marginBottom="12dp"
+            android:text="@string/home_section_anti_fraud"
+            android:textColor="@color/white"
+            android:textSize="20sp"
+            android:textStyle="bold" />
+
+        <androidx.constraintlayout.widget.ConstraintLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content">
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardSilent"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_green"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toStartOf="@id/cardAction"
+                app:layout_constraintHorizontal_chainStyle="spread"
+                app:layout_constraintStart_toStartOf="parent"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_silent" />
+            </com.google.android.material.card.MaterialCardView>
+
+            <com.google.android.material.card.MaterialCardView
+                android:id="@+id/cardAction"
+                android:layout_width="0dp"
+                android:layout_height="0dp"
+                android:layout_marginStart="12dp"
+                android:clickable="true"
+                android:focusable="true"
+                android:foreground="?attr/selectableItemBackground"
+                app:cardBackgroundColor="@color/home_tile_blue"
+                app:cardCornerRadius="22dp"
+                app:cardElevation="0dp"
+                app:layout_constraintDimensionRatio="1:1"
+                app:layout_constraintEnd_toEndOf="parent"
+                app:layout_constraintStart_toEndOf="@id/cardSilent"
+                app:layout_constraintTop_toTopOf="parent">
+
+                <include layout="@layout/item_home_action" />
+            </com.google.android.material.card.MaterialCardView>
+        </androidx.constraintlayout.widget.ConstraintLayout>
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_liveness.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_liveness.xml
new file mode 100644
index 000000000..274ddb83a
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/activity_liveness.xml
@@ -0,0 +1,241 @@
+<?xml version="1.0" encoding="utf-8"?>
+<FrameLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/livenessRoot"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:background="@color/black"
+    android:keepScreenOn="true">
+
+    <androidx.camera.view.PreviewView
+        android:id="@+id/previewView"
+        android:layout_width="match_parent"
+        android:layout_height="match_parent" />
+
+    <com.example.inspireface_example.view.LandmarkGlView
+        android:id="@+id/landmarkGlView"
+        android:layout_width="match_parent"
+        android:layout_height="match_parent"
+        android:visibility="gone" />
+
+    <com.example.inspireface_example.view.FaceOverlayView
+        android:id="@+id/faceOverlay"
+        android:layout_width="match_parent"
+        android:layout_height="match_parent" />
+
+    <LinearLayout
+        android:id="@+id/topBar"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_gravity="top"
+        android:gravity="center_horizontal"
+        android:orientation="vertical"
+        android:paddingStart="16dp"
+        android:paddingEnd="16dp">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <TextView
+                android:id="@+id/btnBack"
+                android:layout_width="wrap_content"
+                android:layout_height="40dp"
+                android:background="@drawable/bg_chip"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:minWidth="40dp"
+                android:text="@string/back"
+                android:textColor="@color/white"
+                android:textSize="18sp" />
+
+            <TextView
+                android:id="@+id/pageTitle"
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="8dp"
+                android:layout_marginEnd="8dp"
+                android:layout_weight="1"
+                android:gravity="center"
+                android:maxLines="1"
+                android:textColor="@color/white"
+                android:textSize="16sp"
+                android:textStyle="bold"
+                tools:text="Silent liveness" />
+
+            <TextView
+                android:id="@+id/currentModel"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:background="@drawable/bg_chip"
+                android:gravity="center"
+                android:maxLines="1"
+                android:paddingStart="10dp"
+                android:paddingTop="5dp"
+                android:paddingEnd="10dp"
+                android:paddingBottom="5dp"
+                android:textColor="@color/liveness_accent"
+                android:textSize="11sp"
+                tools:text="Model · Megatron" />
+        </LinearLayout>
+
+        <TextView
+            android:id="@+id/perfText"
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:background="@drawable/bg_chip"
+            android:paddingStart="12dp"
+            android:paddingTop="4dp"
+            android:paddingEnd="12dp"
+            android:paddingBottom="4dp"
+            android:textColor="@color/white"
+            android:textSize="12sp"
+            android:visibility="gone"
+            tools:text="30 FPS · 12 ms" />
+
+        <LinearLayout
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:orientation="horizontal">
+
+            <com.google.android.material.switchmaterial.SwitchMaterial
+                android:id="@+id/switchEuler"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:text="@string/switch_euler"
+                android:textColor="@color/white"
+                android:textSize="12sp" />
+
+            <TextView
+                android:id="@+id/btnFlipCamera"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="16dp"
+                android:background="@drawable/bg_chip"
+                android:paddingStart="14dp"
+                android:paddingTop="6dp"
+                android:paddingEnd="14dp"
+                android:paddingBottom="6dp"
+                android:text="@string/flip_camera"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+        </LinearLayout>
+
+        <TextView
+            android:id="@+id/eulerText"
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:background="@drawable/bg_chip"
+            android:paddingStart="12dp"
+            android:paddingTop="4dp"
+            android:paddingEnd="12dp"
+            android:paddingBottom="4dp"
+            android:textColor="@color/liveness_accent"
+            android:textSize="12sp"
+            android:visibility="gone"
+            tools:text="Yaw -3.2° · Pitch 5.1° · Roll 0.8°" />
+    </LinearLayout>
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_gravity="bottom"
+        android:orientation="vertical">
+
+    <TextView
+        android:id="@+id/langSwitch"
+        android:layout_width="wrap_content"
+        android:layout_height="wrap_content"
+        android:layout_gravity="end"
+        android:layout_marginEnd="16dp"
+        android:layout_marginBottom="8dp"
+        android:background="@drawable/bg_chip"
+        android:paddingStart="14dp"
+        android:paddingTop="6dp"
+        android:paddingEnd="14dp"
+        android:paddingBottom="6dp"
+        android:text="@string/lang_switch"
+        android:textColor="@color/white"
+        android:textSize="13sp" />
+
+    <com.google.android.material.card.MaterialCardView
+        android:id="@+id/promptCard"
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginStart="16dp"
+        android:layout_marginEnd="16dp"
+        android:layout_marginBottom="24dp"
+        app:cardBackgroundColor="@color/liveness_scrim"
+        app:cardCornerRadius="20dp"
+        app:cardElevation="0dp"
+        app:strokeWidth="0dp">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_horizontal"
+            android:orientation="vertical"
+            android:padding="20dp">
+
+            <TextView
+                android:id="@+id/promptTitle"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:gravity="center"
+                android:text="@string/msg_initializing"
+                android:textColor="@color/white"
+                android:textSize="22sp"
+                android:textStyle="bold" />
+
+            <TextView
+                android:id="@+id/promptSub"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="6dp"
+                android:gravity="center"
+                android:textColor="#B3FFFFFF"
+                android:textSize="14sp"
+                android:visibility="gone" />
+
+            <com.google.android.material.progressindicator.LinearProgressIndicator
+                android:id="@+id/promptProgress"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:max="100"
+                android:visibility="gone"
+                app:indicatorColor="@color/liveness_accent"
+                app:trackColor="#33FFFFFF"
+                app:trackCornerRadius="4dp"
+                app:trackThickness="8dp" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnRestart"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:backgroundTint="@color/liveness_accent"
+                android:text="@string/btn_restart"
+                android:textColor="@color/black"
+                android:visibility="gone" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnCameraPermissionAction"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:text="@string/camera_permission_try_again"
+                android:textColor="@color/black"
+                android:visibility="gone"
+                app:backgroundTint="@color/liveness_accent" />
+        </LinearLayout>
+    </com.google.android.material.card.MaterialCardView>
+    </LinearLayout>
+</FrameLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/dialog_face_editor.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/dialog_face_editor.xml
new file mode 100644
index 000000000..051a4c365
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/dialog_face_editor.xml
@@ -0,0 +1,107 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.core.widget.NestedScrollView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    android:layout_width="match_parent"
+    android:layout_height="wrap_content"
+    android:fillViewport="true">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical"
+        android:paddingStart="20dp"
+        android:paddingEnd="20dp"
+        android:paddingBottom="8dp">
+
+        <com.google.android.material.card.MaterialCardView
+            android:layout_width="match_parent"
+            android:layout_height="240dp"
+            app:cardBackgroundColor="@color/home_background"
+            app:cardCornerRadius="18dp"
+            app:cardElevation="0dp"
+            app:strokeColor="@color/home_stroke"
+            app:strokeWidth="1dp">
+
+            <FrameLayout
+                android:layout_width="match_parent"
+                android:layout_height="match_parent">
+
+                <ImageView
+                    android:id="@+id/editorImage"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent"
+                    android:contentDescription="@string/enrollment_image"
+                    android:scaleType="centerCrop" />
+
+                <com.example.inspireface_example.widget.FaceImageOverlayView
+                    android:id="@+id/editorFaceOverlay"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent" />
+
+                <TextView
+                    android:id="@+id/editorPlaceholder"
+                    android:layout_width="match_parent"
+                    android:layout_height="match_parent"
+                    android:gravity="center"
+                    android:padding="20dp"
+                    android:text="@string/choose_enrollment_image"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="14sp" />
+
+                <TextView
+                    android:id="@+id/editorImageStatus"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_gravity="bottom"
+                    android:background="@color/compare_status_scrim"
+                    android:gravity="center"
+                    android:padding="8dp"
+                    android:text="@string/no_image_selected"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="12sp" />
+            </FrameLayout>
+        </com.google.android.material.card.MaterialCardView>
+
+        <com.google.android.material.button.MaterialButton
+            android:id="@+id/btnCaptureEditorFace"
+            android:layout_width="match_parent"
+            android:layout_height="48dp"
+            android:layout_marginTop="10dp"
+            android:text="@string/capture_or_replace_face"
+            android:textColor="@color/home_tile_text_dark"
+            app:backgroundTint="@color/liveness_accent" />
+
+        <com.google.android.material.button.MaterialButton
+            android:id="@+id/btnChooseEditorImage"
+            style="?attr/materialButtonOutlinedStyle"
+            android:layout_width="match_parent"
+            android:layout_height="48dp"
+            android:layout_marginTop="8dp"
+            android:text="@string/choose_or_replace_image"
+            android:textColor="@color/white"
+            app:strokeColor="@color/liveness_accent" />
+
+        <TextView
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="8dp"
+            android:text="@string/tap_face_to_select"
+            android:textColor="@color/home_text_muted"
+            android:textSize="12sp" />
+
+        <com.google.android.material.textfield.TextInputLayout
+            style="@style/Widget.Material3.TextInputLayout.OutlinedBox"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="14dp"
+            android:hint="@string/face_name">
+
+            <com.google.android.material.textfield.TextInputEditText
+                android:id="@+id/editorName"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:inputType="textPersonName"
+                android:maxLines="1" />
+        </com.google.android.material.textfield.TextInputLayout>
+    </LinearLayout>
+</androidx.core.widget.NestedScrollView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_face_record.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_face_record.xml
new file mode 100644
index 000000000..daedbf8b5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_face_record.xml
@@ -0,0 +1,83 @@
+<?xml version="1.0" encoding="utf-8"?>
+<FrameLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    android:layout_width="match_parent"
+    android:layout_height="wrap_content"
+    android:paddingBottom="10dp">
+
+    <com.google.android.material.card.MaterialCardView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        app:cardBackgroundColor="@color/home_surface"
+        app:cardCornerRadius="18dp"
+        app:cardElevation="0dp"
+        app:strokeColor="@color/home_stroke"
+        app:strokeWidth="1dp">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:minHeight="94dp"
+            android:orientation="horizontal"
+            android:padding="11dp">
+
+            <ImageView
+                android:id="@+id/faceCrop"
+                android:layout_width="72dp"
+                android:layout_height="72dp"
+                android:background="@color/compare_ring_track"
+                android:contentDescription="@string/stored_face_crop"
+                android:scaleType="centerCrop" />
+
+            <LinearLayout
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="13dp"
+                android:layout_weight="1"
+                android:orientation="vertical">
+
+                <TextView
+                    android:id="@+id/faceName"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:maxLines="1"
+                    android:textColor="@color/white"
+                    android:textSize="16sp"
+                    android:textStyle="bold" />
+
+                <TextView
+                    android:id="@+id/faceId"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="5dp"
+                    android:textColor="@color/home_text_muted"
+                    android:textSize="12sp" />
+            </LinearLayout>
+
+            <TextView
+                android:id="@+id/btnEditFace"
+                android:layout_width="44dp"
+                android:layout_height="44dp"
+                android:background="?attr/selectableItemBackgroundBorderless"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/edit_short"
+                android:textColor="@color/liveness_accent"
+                android:textSize="13sp" />
+
+            <TextView
+                android:id="@+id/btnDeleteFace"
+                android:layout_width="44dp"
+                android:layout_height="44dp"
+                android:background="?attr/selectableItemBackgroundBorderless"
+                android:clickable="true"
+                android:focusable="true"
+                android:gravity="center"
+                android:text="@string/delete_short"
+                android:textColor="@color/liveness_fail"
+                android:textSize="13sp" />
+        </LinearLayout>
+    </com.google.android.material.card.MaterialCardView>
+</FrameLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_action.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_action.xml
new file mode 100644
index 000000000..662157921
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_action.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_action_liveness" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_action"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/mode_action"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_action_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_attribute.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_attribute.xml
new file mode 100644
index 000000000..332744339
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_attribute.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_attributes" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_attribute"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/face_attribute_title"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_attribute_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_compare.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_compare.xml
new file mode 100644
index 000000000..1975d55c4
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_compare.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_compare" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_compare"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/face_compare_title"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_compare_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_detection.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_detection.xml
new file mode 100644
index 000000000..7454c26b6
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_detection.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_tracking" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_detection"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/face_detection_title"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_detection_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_management.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_management.xml
new file mode 100644
index 000000000..33d235f10
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_management.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_management" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_management"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/face_management_title"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_management_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_pose.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_pose.xml
new file mode 100644
index 000000000..deac55729
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_pose.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_pose" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_pose"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/mode_pose"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_pose_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_recognition.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_recognition.xml
new file mode 100644
index 000000000..905d9180f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_recognition.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_recognition" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_recognition"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/face_recognition_title"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_recognition_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_silent.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_silent.xml
new file mode 100644
index 000000000..46db968d5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/item_home_silent.xml
@@ -0,0 +1,53 @@
+<?xml version="1.0" encoding="utf-8"?>
+<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    android:gravity="bottom"
+    android:orientation="vertical"
+    android:padding="18dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:gravity="center_vertical"
+        android:orientation="horizontal">
+
+        <ImageView
+            android:layout_width="32dp"
+            android:layout_height="32dp"
+            android:contentDescription="@null"
+            android:importantForAccessibility="no"
+            android:src="@drawable/ic_home_silent_liveness" />
+
+        <TextView
+            android:layout_width="wrap_content"
+            android:layout_height="wrap_content"
+            android:layout_marginStart="8dp"
+            android:text="@string/menu_badge_rgb"
+            android:textColor="@color/home_tile_text_dark"
+            android:textSize="11sp"
+            android:textStyle="bold" />
+    </LinearLayout>
+
+    <Space
+        android:layout_width="1dp"
+        android:layout_height="0dp"
+        android:layout_weight="1" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:text="@string/mode_silent"
+        android:textColor="@color/home_tile_text_dark"
+        android:textSize="19sp"
+        android:textStyle="bold" />
+
+    <TextView
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="5dp"
+        android:maxLines="2"
+        android:text="@string/home_silent_desc"
+        android:textColor="@color/home_tile_text_dark_secondary"
+        android:textSize="12sp" />
+</LinearLayout>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/view_face_tracking_settings.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/view_face_tracking_settings.xml
new file mode 100644
index 000000000..28949662f
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/layout/view_face_tracking_settings.xml
@@ -0,0 +1,268 @@
+<?xml version="1.0" encoding="utf-8"?>
+<com.google.android.material.card.MaterialCardView xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:layout_width="match_parent"
+    android:layout_height="wrap_content"
+    app:cardBackgroundColor="@color/home_surface"
+    app:cardCornerRadius="18dp"
+    app:cardElevation="0dp"
+    app:strokeColor="@color/home_stroke"
+    app:strokeWidth="1dp">
+
+    <LinearLayout
+        android:layout_width="match_parent"
+        android:layout_height="wrap_content"
+        android:orientation="vertical">
+
+        <LinearLayout
+            android:id="@+id/trackingSettingsHeader"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:background="?attr/selectableItemBackground"
+            android:clickable="true"
+            android:focusable="true"
+            android:gravity="center_vertical"
+            android:minHeight="64dp"
+            android:orientation="horizontal"
+            android:padding="16dp">
+
+            <LinearLayout
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_weight="1"
+                android:orientation="vertical">
+
+                <TextView
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:text="@string/detection_tracking_settings"
+                    android:textColor="@color/white"
+                    android:textSize="16sp"
+                    android:textStyle="bold" />
+
+                <TextView
+                    android:id="@+id/trackingSessionStatus"
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="3dp"
+                    android:textColor="@color/home_text_secondary"
+                    android:textSize="11sp"
+                    tools:text="Light · 640 px · max 10 · min 24 px" />
+            </LinearLayout>
+
+            <TextView
+                android:id="@+id/trackingSettingsToggleText"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="12dp"
+                android:text="@string/recognition_settings_expand"
+                android:textColor="@color/liveness_accent"
+                android:textSize="13sp" />
+        </LinearLayout>
+
+        <LinearLayout
+            android:id="@+id/trackingSettingsContent"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:orientation="vertical"
+            android:paddingStart="16dp"
+            android:paddingEnd="16dp"
+            android:paddingBottom="16dp"
+            android:visibility="gone">
+
+            <TextView
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:text="@string/detection_tracking_settings_hint"
+                android:textColor="@color/home_text_secondary"
+                android:textSize="12sp" />
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="14dp"
+                android:text="@string/detection_tracking_mode"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+
+            <com.google.android.material.chip.ChipGroup
+                android:id="@+id/trackingModeGroup"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="5dp"
+                app:checkedChip="@id/trackingModeLight"
+                app:selectionRequired="true"
+                app:singleSelection="true">
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingModeLight"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/detection_tracking_mode_light" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingModeTbd"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/detection_tracking_mode_tbd" />
+            </com.google.android.material.chip.ChipGroup>
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="10dp"
+                android:text="@string/recognition_input_px"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+
+            <com.google.android.material.chip.ChipGroup
+                android:id="@+id/trackingInputPxGroup"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="5dp"
+                app:checkedChip="@id/trackingInputPx640"
+                app:selectionRequired="true"
+                app:singleSelection="true">
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingInputPx320"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_320" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingInputPx640"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_640" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingInputPx1280"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_1280" />
+            </com.google.android.material.chip.ChipGroup>
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="10dp"
+                android:text="@string/recognition_max_faces"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+
+            <com.google.android.material.chip.ChipGroup
+                android:id="@+id/trackingMaxFacesGroup"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="5dp"
+                app:checkedChip="@id/trackingMaxFaces10"
+                app:selectionRequired="true"
+                app:singleSelection="true">
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMaxFaces1"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_1" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMaxFaces3"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_3" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMaxFaces5"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_5" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMaxFaces10"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_10" />
+            </com.google.android.material.chip.ChipGroup>
+
+            <TextView
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="10dp"
+                android:text="@string/recognition_min_face_px"
+                android:textColor="@color/white"
+                android:textSize="13sp" />
+
+            <com.google.android.material.chip.ChipGroup
+                android:id="@+id/trackingMinFaceGroup"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="5dp"
+                app:checkedChip="@id/trackingMinFace24"
+                app:selectionRequired="true"
+                app:singleSelection="true">
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMinFace24"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_24" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMinFace48"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_48" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMinFace64"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_64" />
+
+                <com.google.android.material.chip.Chip
+                    android:id="@+id/trackingMinFace128"
+                    style="@style/Widget.Material3.Chip.Filter"
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:checkable="true"
+                    android:text="@string/value_128" />
+            </com.google.android.material.chip.ChipGroup>
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnResetTrackingSession"
+                style="?attr/materialButtonOutlinedStyle"
+                android:layout_width="match_parent"
+                android:layout_height="48dp"
+                android:layout_marginTop="12dp"
+                android:text="@string/recognition_reset_defaults"
+                android:textColor="@color/white"
+                app:strokeColor="@color/liveness_accent" />
+        </LinearLayout>
+    </LinearLayout>
+</com.google.android.material.card.MaterialCardView>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml
new file mode 100644
index 000000000..6f3b755bf
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+    <monochrome android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml
new file mode 100644
index 000000000..6f3b755bf
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+    <monochrome android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher.webp
new file mode 100644
index 000000000..e3c33935d
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..128155782
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher.webp
new file mode 100644
index 000000000..73d3d88ae
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..92231b352
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher.webp
new file mode 100644
index 000000000..e290e73c5
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..16042379e
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp
new file mode 100644
index 000000000..11fa12e3a
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..77cce0b2f
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp
new file mode 100644
index 000000000..b0929a936
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..86dd51013
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-night/themes.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-night/themes.xml
new file mode 100644
index 000000000..9274570a8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-night/themes.xml
@@ -0,0 +1,7 @@
+<resources xmlns:tools="http://schemas.android.com/tools">
+    <!-- Base application theme. -->
+    <style name="Base.Theme.InspireFaceExample" parent="Theme.Material3.DayNight.NoActionBar">
+        <!-- Customize your dark theme here. -->
+        <!-- <item name="colorPrimary">@color/my_dark_primary</item> -->
+    </style>
+</resources>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-zh/strings.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-zh/strings.xml
new file mode 100644
index 000000000..c2fd38bd5
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values-zh/strings.xml
@@ -0,0 +1,268 @@
+<resources>
+    <string name="app_name">InspireFace 示例</string>
+
+    <!-- 活体检测页面 -->
+    <string name="mode_silent">静默活体</string>
+    <string name="mode_action">动作活体</string>
+    <string name="mode_pose">姿态识别</string>
+    <string name="msg_initializing">引擎初始化中…</string>
+    <string name="msg_no_face">请将脸部对准取景框</string>
+    <string name="msg_multiple_faces">请保持画面中只有一张人脸</string>
+    <string name="msg_face_too_small">请靠近一点</string>
+    <string name="msg_permission_required">需要相机权限才能使用</string>
+    <string name="msg_engine_failed">人脸引擎初始化失败</string>
+    <string name="msg_no_front_camera">此设备没有可用相机</string>
+    <string name="msg_camera_unavailable">该摄像头不可用</string>
+    <string name="flip_camera">切换镜头</string>
+    <string name="back">←</string>
+    <string name="current_model">模型 · %1$s</string>
+    <string name="camera_privacy_notice_title">相机与人脸数据说明</string>
+    <string name="camera_privacy_notice_message">相机仅用于视频功能和相机录入，人脸画面在当前设备处理。录入的人脸裁剪图与特征保存在 App 内部，按模型隔离，不进入设备备份，也不会上传。拒绝相机权限后仍可继续使用照片输入功能。</string>
+    <string name="camera_permission_continue">继续</string>
+    <string name="camera_permission_not_now">暂不使用</string>
+    <string name="camera_permission_rationale_title">允许使用相机</string>
+    <string name="camera_permission_rationale_message">仅当前视频功能需要相机权限；拒绝后仍可继续使用照片输入功能。</string>
+    <string name="camera_permission_try_again">重新授权</string>
+    <string name="camera_permission_open_settings">前往系统设置</string>
+    <string name="camera_permission_retry_hint">相机权限尚未开启，点击“重新授权”后继续。</string>
+    <string name="camera_permission_settings_hint">相机权限已被阻止，请前往系统设置允许“相机”，然后返回此页面。</string>
+
+    <string name="silent_real">真人</string>
+    <string name="silent_fake">疑似攻击</string>
+    <string name="silent_analyzing">检测中…</string>
+    <string name="silent_score">活体分数 %1$.2f</string>
+
+    <string name="action_get_ready">请正对相机并保持稳定</string>
+    <string name="action_blink">请眨眼</string>
+    <string name="action_shake">请左右摇头</string>
+    <string name="action_jaw_open">请张张嘴</string>
+    <string name="action_head_raise">请抬头</string>
+    <string name="action_step">第 %1$d / %2$d 步</string>
+    <string name="action_time_left">剩余 %1$d 秒</string>
+    <string name="action_passed">活体检测通过</string>
+    <string name="action_failed_timeout">超时,检测失败</string>
+    <string name="action_face_lost">人脸丢失,请重新开始</string>
+    <string name="btn_restart">重新开始</string>
+
+    <string name="pose_hint">做个动作试试:眨眼、摇头、张嘴、抬头</string>
+    <string name="name_blink">眨眼</string>
+    <string name="name_shake">摇头</string>
+    <string name="name_jaw_open">张嘴</string>
+    <string name="name_head_raise">抬头</string>
+
+    <string name="perf_format">%1$.1f FPS · %2$d ms</string>
+    <string name="switch_euler">欧拉角</string>
+    <!-- Label of the language toggle: names the language it switches TO. -->
+    <string name="lang_switch">English</string>
+    <string name="euler_format">偏航 %1$.1f° · 俯仰 %2$.1f° · 翻滚 %3$.1f°</string>
+    <string name="euler_no_face">—</string>
+
+    <!-- 主页 -->
+    <string name="home_title">探索人脸智能</string>
+    <string name="home_subtitle">选择全局模型，然后进入一个 InspireFace 功能演示。</string>
+    <string name="model_section_title">全局模型</string>
+    <string name="model_section_hint">所选模型将在打开功能页面时加载。</string>
+    <string name="home_section_analysis">人脸分析</string>
+    <string name="home_section_recognition">人脸识别</string>
+    <string name="home_section_anti_fraud">反欺诈</string>
+    <string name="home_silent_desc">RGB 防伪检测，实时展示活体置信度</string>
+    <string name="home_action_desc">按提示完成随机人脸动作序列</string>
+    <string name="home_pose_desc">识别眨眼、摇头、张嘴与抬头动作</string>
+    <string name="home_compare_desc">上传两张人像照片并比较人脸相似度</string>
+    <string name="home_management_desc">管理当前模型独立的身份、特征和裁剪图</string>
+    <string name="home_recognition_desc">在当前模型的人脸库中识别人脸身份</string>
+    <string name="home_detection_desc">检测人脸并查看原生 106 点稠密关键点</string>
+    <string name="home_attribute_desc">查看口罩、年龄、质量与人脸属性</string>
+
+    <!-- 人脸 1:1 -->
+    <string name="face_compare_title">人脸 1:1</string>
+    <string name="face_compare_heading">是同一个人吗？</string>
+    <string name="face_compare_hint">请在左右两侧分别添加一张清晰图片。默认选择第 1 张人脸；点击任意编号人脸框即可切换，并立即重新比对。</string>
+    <string name="image_a">图像 A</string>
+    <string name="image_b">图像 B</string>
+    <string name="upload_image_a">添加图像 A</string>
+    <string name="upload_image_b">添加图像 B</string>
+    <string name="image_empty">点击选择</string>
+    <string name="image_analyzing">正在检测人脸…</string>
+    <string name="image_face_ready">人脸已就绪</string>
+    <string name="image_no_face">未检测到人脸</string>
+    <string name="image_face_selected">已选择第 %1$d / %2$d 张人脸</string>
+    <string name="image_load_failed">无法读取图片</string>
+    <string name="face_extract_failed">人脸特征提取失败</string>
+    <string name="similarity_title">相似度</string>
+    <string name="compare_selected_face_a">当前参与比对的人脸 A</string>
+    <string name="compare_selected_face_b">当前参与比对的人脸 B</string>
+    <string name="compare_select_two">请选择两张图片</string>
+    <string name="compare_analyzing">正在分析图片…</string>
+    <string name="compare_waiting_face">请添加两张有效人脸图片</string>
+    <string name="compare_comparing">正在比对人脸…</string>
+    <string name="compare_same_person">可能是同一个人</string>
+    <string name="compare_different_person">可能不是同一个人</string>
+    <string name="compare_failed">人脸比对失败</string>
+    <string name="compare_engine_failed">人脸引擎不可用</string>
+    <string name="compare_details">模型置信度 %1$.3f · 阈值 %2$.3f</string>
+    <string name="on_device_privacy">图片仅在当前设备上处理。</string>
+
+    <!-- 人脸管理 -->
+    <string name="face_management_title">人脸管理</string>
+    <string name="face_library_heading">人脸库</string>
+    <string name="face_storage_scope">%1$s 使用独立的特征数据库和人脸裁剪图目录。</string>
+    <string name="search_faces">按姓名或 ID 搜索</string>
+    <string name="add_face">新增</string>
+    <string name="no_faces_title">还没有录入人脸</string>
+    <string name="no_faces_hint">添加一张人像，在当前模型库中创建第一个身份。</string>
+    <string name="stored_face_crop">已存储的人脸裁剪图</string>
+    <string name="face_id_format">ID · %1$d</string>
+    <string name="edit_short">编辑</string>
+    <string name="delete_short">删除</string>
+    <string name="add_face_title">新增人脸</string>
+    <string name="edit_face_title">编辑人脸</string>
+    <string name="choose_enrollment_image">使用相机录入或从相册选择</string>
+    <string name="enrollment_image">人脸录入图片</string>
+    <string name="no_image_selected">尚未选择图片</string>
+    <string name="capture_or_replace_face">使用相机录入</string>
+    <string name="choose_or_replace_image">从相册选择</string>
+    <string name="tap_face_to_select">检测到多张人脸时，点击人脸框选择要录入的身份。</string>
+    <string name="face_name">姓名</string>
+    <string name="save">保存</string>
+    <string name="cancel">取消</string>
+    <string name="saving_face">正在保存…</string>
+    <string name="face_saved">人脸已保存</string>
+    <string name="face_save_failed">无法保存该人脸</string>
+    <string name="face_library_failed">无法打开当前模型的独立人脸库</string>
+    <string name="delete_face_title">删除人脸？</string>
+    <string name="delete_face_message">确定从当前 %2$s 模型库删除 %1$s？对应特征和裁剪图都会被移除。</string>
+    <string name="delete">删除</string>
+    <string name="face_deleted">人脸已删除</string>
+    <string name="face_delete_failed">无法删除该人脸</string>
+    <string name="face_list_count">已存储 %1$d 个身份</string>
+
+    <!-- 人脸识别 -->
+    <string name="face_recognition_title">人脸识别</string>
+    <string name="face_recognition_heading">这是谁？</string>
+    <string name="face_recognition_hint">在当前模型已录入的身份库中搜索人脸。</string>
+    <string name="recognition_tab_photo">照片输入</string>
+    <string name="recognition_tab_video">视频流</string>
+    <string name="recognition_session_parameters">Session 参数</string>
+    <string name="recognition_session_parameters_hint">应用后会重建照片 Session，并将参数保存在当前设备。</string>
+    <string name="recognition_settings_expand">展开</string>
+    <string name="recognition_settings_collapse">收起</string>
+    <string name="recognition_input_px">检测输入 px</string>
+    <string name="recognition_max_faces">最大检测人脸数</string>
+    <string name="recognition_min_face_px">最小人脸 px</string>
+    <string name="recognition_apply_session">应用并重建</string>
+    <string name="recognition_reset_defaults">恢复默认</string>
+    <string name="recognition_session_summary">Session · %1$d px · 最多 %2$d 张 · 最小 %3$d px</string>
+    <string name="recognition_session_rebuilding">正在重建 Session…</string>
+    <string name="recognition_session_failed">无法创建该 Session</string>
+    <string name="recognition_empty_library">%1$s 模型当前还没有录入人脸，请先前往人脸管理添加身份。</string>
+    <string name="recognition_photo">识别照片</string>
+    <string name="recognition_selected_face_crop">当前选中人脸裁剪图</string>
+    <string name="recognition_choose_photo_hint">选择照片后检测并识别人脸</string>
+    <string name="recognition_choose_photo">选择照片</string>
+    <string name="recognition_multi_face_hint">单张人脸会自动搜索；检测到多张人脸时，点击编号框即可切换。</string>
+    <string name="recognition_searching">正在当前模型的人脸库中搜索…</string>
+    <string name="recognition_match_found">识别成功</string>
+    <string name="recognition_no_match">未找到匹配身份</string>
+    <string name="recognition_library_empty_result">当前模型的人脸库为空</string>
+    <string name="recognition_result_name_unknown">未知身份</string>
+    <string name="recognition_result_details">ID · %1$d · 模型置信度 %2$.3f · 阈值 %3$.3f</string>
+    <string name="recognition_no_match_details">模型置信度 %1$.3f · 阈值 %2$.3f</string>
+    <string name="recognition_no_confidence">没有候选身份超过搜索阈值</string>
+    <string name="recognition_video_initializing">正在启动相机…</string>
+    <string name="recognition_video_no_face">请将人脸放入画面</string>
+    <string name="recognition_video_hold_still">请保持稳定，正在准备识别</string>
+    <string name="recognition_video_first_face_hint">只跟踪第一张人脸，稳定约 1 秒后自动识别。</string>
+
+    <!-- 人脸检测和跟踪 -->
+    <string name="face_detection_title">人脸跟踪</string>
+    <string name="face_detection_heading">找到每一张人脸</string>
+    <string name="face_detection_hint">检测并选择照片中的人脸，查看 SDK 原生 106 点稠密关键点。</string>
+    <string name="detection_tab_image">图像检测</string>
+    <string name="detection_tab_tracking">视频跟踪</string>
+    <string name="detection_show_dense_landmarks">显示 106 点稠密关键点</string>
+    <string name="detection_photo">人脸检测照片</string>
+    <string name="detection_choose_photo_hint">选择照片后检测人脸和关键点</string>
+    <string name="detection_choose_photo">选择检测照片</string>
+    <string name="detection_multi_face_hint">默认选择第 1 张人脸，点击任意编号框即可查看对应人脸。</string>
+    <string name="detection_landmark_behavior_hint">大人脸直接在原图显示关键点，小人脸使用右下角放大器。</string>
+    <string name="detection_face_selected">已选择第 %1$d / %2$d 张人脸 · SDK 返回 %3$d 个关键点</string>
+    <string name="face_detection_failed">人脸检测失败</string>
+    <string name="detection_landmarks_hidden">已隐藏 106 点关键点</string>
+    <string name="detection_landmarks_failed">无法读取该人脸的 106 点关键点</string>
+    <string name="detection_landmarks_on_source">大人脸 · 占原图 %1$d%% · 关键点显示在原图</string>
+    <string name="detection_landmarks_in_magnifier">小人脸 · 占原图 %1$d%% · 关键点已在右下角放大</string>
+    <string name="detection_landmark_magnifier">选中人脸放大图</string>
+    <string name="detection_tracking_initializing">正在启动相机跟踪…</string>
+    <string name="detection_tracking_restarting">正在应用配置并重启相机…</string>
+    <string name="detection_tracking_stats">%1$.1f FPS · %2$d ms</string>
+    <string name="detection_tracking_settings">跟踪 Session 参数</string>
+    <string name="detection_tracking_settings_hint">修改后会自动保存配置并重启相机流。</string>
+    <string name="detection_tracking_mode">跟踪模式</string>
+    <string name="detection_tracking_mode_light">Light 轻量跟踪</string>
+    <string name="detection_tracking_mode_tbd">Track-by-detection</string>
+    <string name="detection_tracking_summary">%1$s · %2$d px · 最多 %3$d 张 · 最小 %4$d px</string>
+    <string name="detection_tracking_session_failed">无法创建该跟踪 Session</string>
+
+    <!-- 人脸属性 -->
+    <string name="face_attribute_title">人脸属性分析</string>
+    <string name="face_attribute_heading">分析人脸属性</string>
+    <string name="face_attribute_hint">选择一张照片，点击任意编号人脸即可动态查看对应属性。</string>
+    <string name="attribute_photo">人脸属性照片</string>
+    <string name="attribute_choose_photo_hint">选择照片后分析人脸属性</string>
+    <string name="attribute_choose_photo">选择照片</string>
+    <string name="attribute_multi_face_hint">默认选择第 1 张人脸；选中的小人脸会显示在右下角放大镜中。</string>
+    <string name="attribute_selected_face_crop">选中人脸放大图</string>
+    <string name="attribute_face_selected">已选择第 %1$d / %2$d 张人脸</string>
+    <string name="attribute_result_title">选中人脸属性</string>
+    <string name="attribute_result_empty">选择照片后查看属性</string>
+    <string name="attribute_analyzing">正在分析人脸属性…</string>
+    <string name="attribute_analysis_failed">无法分析人脸属性</string>
+    <string name="attribute_no_face_result">暂无可显示的人脸属性</string>
+    <string name="attribute_mask">佩戴口罩</string>
+    <string name="attribute_age">年龄段</string>
+    <string name="attribute_quality">人脸画面质量得分</string>
+    <string name="attribute_expression">表情</string>
+    <string name="attribute_race">民族</string>
+    <string name="attribute_gender">性别</string>
+    <string name="attribute_left_eye">左眼状态</string>
+    <string name="attribute_right_eye">右眼状态</string>
+    <string name="attribute_mask_yes">是</string>
+    <string name="attribute_mask_no">否</string>
+    <string name="attribute_eye_open">睁开</string>
+    <string name="attribute_eye_closed">闭合</string>
+    <string name="attribute_expression_neutral">自然</string>
+    <string name="attribute_expression_mouth_open">张嘴</string>
+    <string name="attribute_gender_female">女性</string>
+    <string name="attribute_gender_male">男性</string>
+    <string name="attribute_race_black">黑人</string>
+    <string name="attribute_race_asian">亚洲人</string>
+    <string name="attribute_race_latino">拉丁裔 / 西班牙裔</string>
+    <string name="attribute_race_middle_eastern">中东人</string>
+    <string name="attribute_race_white">白人</string>
+    <string name="attribute_age_0_2">0–2 岁</string>
+    <string name="attribute_age_3_9">3–9 岁</string>
+    <string name="attribute_age_10_19">10–19 岁</string>
+    <string name="attribute_age_20_29">20–29 岁</string>
+    <string name="attribute_age_30_39">30–39 岁</string>
+    <string name="attribute_age_40_49">40–49 岁</string>
+    <string name="attribute_age_50_59">50–59 岁</string>
+    <string name="attribute_age_60_69">60–69 岁</string>
+    <string name="attribute_age_70_plus">70 岁以上</string>
+    <string name="attribute_unknown">—</string>
+
+    <!-- 相机人脸录入 -->
+    <string name="face_capture_title">相机录入</string>
+    <string name="capture_permission_hint">允许相机权限后即可继续。</string>
+    <string name="capture_no_face">请将人脸放入画面</string>
+    <string name="capture_first_face_hint">正在跟踪检测到的第一张人脸，请保持清晰并居中。</string>
+    <string name="capture_move_closer">请靠近一点</string>
+    <string name="capture_hold_still">请保持不动</string>
+    <string name="capture_warmup_hint">稳定保持 1 秒后将出现进度圈。</string>
+    <string name="capture_keep_still">请继续保持稳定</string>
+    <string name="capture_progress">稳定进度 %1$d%%</string>
+    <string name="capture_complete">录入采集完成</string>
+    <string name="capture_complete_hint">进度圈已满，正在准备人脸…</string>
+    <string name="capture_failed">无法采集当前画面</string>
+    <string name="capture_retry_hint">请保持稳定，应用会自动重试。</string>
+</resources>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/colors.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/colors.xml
new file mode 100644
index 000000000..ac9be80c7
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/colors.xml
@@ -0,0 +1,31 @@
+<?xml version="1.0" encoding="utf-8"?>
+<resources>
+    <color name="black">#FF000000</color>
+    <color name="white">#FFFFFFFF</color>
+
+    <!-- Liveness screen -->
+    <color name="liveness_accent">#FF00E5A0</color>
+    <color name="liveness_warn">#FFFFC400</color>
+    <color name="liveness_fail">#FFFF5252</color>
+    <color name="liveness_scrim">#99000000</color>
+    <color name="liveness_chip_bg">#B3000000</color>
+
+    <!-- Home screen -->
+    <color name="home_background">#FF07110F</color>
+    <color name="home_surface">#FF101C19</color>
+    <color name="home_stroke">#FF263A35</color>
+    <color name="home_text_secondary">#FF9FB2AD</color>
+    <color name="home_text_muted">#FF71847F</color>
+    <color name="home_tile_green">#FF72E2B5</color>
+    <color name="home_tile_blue">#FF84C8FF</color>
+    <color name="home_tile_purple">#FFC6B5FF</color>
+    <color name="home_tile_orange">#FFFFC988</color>
+    <color name="home_tile_pink">#FFFFB8D1</color>
+    <color name="home_tile_cyan">#FF79DDD9</color>
+    <color name="home_tile_yellow">#FFFFE082</color>
+    <color name="home_tile_lime">#FFB8E986</color>
+    <color name="home_tile_text_dark">#FF07110F</color>
+    <color name="home_tile_text_dark_secondary">#B307110F</color>
+    <color name="compare_ring_track">#FF23332F</color>
+    <color name="compare_status_scrim">#D907110F</color>
+</resources>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/strings.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/strings.xml
new file mode 100644
index 000000000..65407e3cc
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/strings.xml
@@ -0,0 +1,294 @@
+<resources>
+    <string name="app_name">InspireFace-Example</string>
+
+    <!-- Liveness screen -->
+    <string name="mode_silent">Silent liveness</string>
+    <string name="mode_action">Action liveness</string>
+    <string name="mode_pose">Pose recognition</string>
+    <string name="msg_initializing">Initializing engine…</string>
+    <string name="msg_no_face">Align your face within the frame</string>
+    <string name="msg_multiple_faces">Keep only one face in view</string>
+    <string name="msg_face_too_small">Move closer to the camera</string>
+    <string name="msg_permission_required">Camera permission is required</string>
+    <string name="msg_engine_failed">Failed to initialize the face engine</string>
+    <string name="msg_no_front_camera">No camera available on this device</string>
+    <string name="msg_camera_unavailable">That camera is not available</string>
+    <string name="flip_camera">Flip camera</string>
+    <string name="back">←</string>
+    <string name="current_model">Model · %1$s</string>
+    <string name="camera_privacy_notice_title">Camera and face data</string>
+    <string name="camera_privacy_notice_message">Camera access is used only for live demos and camera enrollment. Frames are processed on this device. Enrolled face crops and features stay in this app’s internal storage, are separated by model, excluded from device backup, and are not uploaded. Photo-only demos remain available without camera access.</string>
+    <string name="camera_permission_continue">Continue</string>
+    <string name="camera_permission_not_now">Not now</string>
+    <string name="camera_permission_rationale_title">Allow camera access</string>
+    <string name="camera_permission_rationale_message">Camera access is required only for this live feature. You can continue using photo-only features without it.</string>
+    <string name="camera_permission_try_again">Try again</string>
+    <string name="camera_permission_open_settings">Open settings</string>
+    <string name="camera_permission_retry_hint">Camera access is off. Tap Try again to continue.</string>
+    <string name="camera_permission_settings_hint">Camera access is blocked. Open system settings, allow Camera, then return here.</string>
+
+    <string name="silent_real">Live face</string>
+    <string name="silent_fake">Spoof suspected</string>
+    <string name="silent_analyzing">Analyzing…</string>
+    <string name="silent_score">Liveness score %1$.2f</string>
+
+    <string name="action_get_ready">Face the camera and hold still</string>
+    <string name="action_blink">Please blink</string>
+    <string name="action_shake">Please shake your head</string>
+    <string name="action_jaw_open">Please open your mouth</string>
+    <string name="action_head_raise">Please raise your head</string>
+    <string name="action_step">Step %1$d / %2$d</string>
+    <string name="action_time_left">%1$ds left</string>
+    <string name="action_passed">Liveness check passed</string>
+    <string name="action_failed_timeout">Timed out, check failed</string>
+    <string name="action_face_lost">Face lost, please try again</string>
+    <string name="btn_restart">Restart</string>
+
+    <string name="pose_hint">Try an action: blink, shake, open mouth, raise head</string>
+    <string name="name_blink">Blink</string>
+    <string name="name_shake">Head shake</string>
+    <string name="name_jaw_open">Mouth open</string>
+    <string name="name_head_raise">Head raise</string>
+
+    <string name="perf_format">%1$.1f FPS · %2$d ms</string>
+    <string name="switch_euler">Euler angles</string>
+    <!-- Label of the language toggle: names the language it switches TO. -->
+    <string name="lang_switch">中文</string>
+    <string name="euler_format">Yaw %1$.1f° · Pitch %2$.1f° · Roll %3$.1f°</string>
+    <string name="euler_no_face">—</string>
+
+    <!-- Home screen -->
+    <string name="home_title">Explore face intelligence</string>
+    <string name="home_subtitle">Choose a model, then open an InspireFace capability demo.</string>
+    <string name="model_section_title">Global model</string>
+    <string name="model_section_hint">The selected model is loaded when a feature opens.</string>
+    <string name="home_section_analysis">Face analysis</string>
+    <string name="home_section_recognition">Face recognition</string>
+    <string name="home_section_anti_fraud">Anti-fraud</string>
+    <string name="home_silent_desc">RGB anti-spoofing with a live confidence score</string>
+    <string name="home_action_desc">Complete a randomized sequence of face actions</string>
+    <string name="home_pose_desc">Recognize blink, shake, mouth and head poses</string>
+    <string name="home_compare_desc">Upload two portraits and compare face similarity</string>
+    <string name="home_management_desc">Manage model-specific identities, features and crops</string>
+    <string name="home_recognition_desc">Identify a face against the current model library</string>
+    <string name="home_detection_desc">Detect faces and inspect native 106-point landmarks</string>
+    <string name="home_attribute_desc">Inspect mask, age, quality and facial attributes</string>
+    <string name="model_pikachu" translatable="false">Pikachu</string>
+    <string name="model_megatron" translatable="false">Megatron</string>
+    <string name="menu_badge_rgb" translatable="false">RGB</string>
+    <string name="menu_badge_action" translatable="false">ACT</string>
+    <string name="menu_badge_pose" translatable="false">POSE</string>
+    <string name="menu_badge_compare" translatable="false">1:1</string>
+    <string name="menu_badge_management" translatable="false">DB</string>
+    <string name="menu_badge_recognition" translatable="false">1:N</string>
+    <string name="menu_badge_detection" translatable="false">TRACK</string>
+    <string name="menu_badge_attribute" translatable="false">ATTR</string>
+    <string name="add_symbol" translatable="false">＋</string>
+
+    <!-- 1:1 face comparison -->
+    <string name="face_compare_title">Face 1:1</string>
+    <string name="face_compare_heading">Are they the same person?</string>
+    <string name="face_compare_hint">Add one clear image to each side. Face 1 is selected by default; tap any numbered face box to switch, and the comparison updates immediately.</string>
+    <string name="image_a">Image A</string>
+    <string name="image_b">Image B</string>
+    <string name="upload_image_a">Add image A</string>
+    <string name="upload_image_b">Add image B</string>
+    <string name="image_empty">Tap to choose</string>
+    <string name="image_analyzing">Detecting face…</string>
+    <string name="image_face_ready">Face ready</string>
+    <string name="image_no_face">No face found</string>
+    <string name="image_face_selected">Face %1$d of %2$d selected</string>
+    <string name="image_load_failed">Could not load image</string>
+    <string name="face_extract_failed">Feature extraction failed</string>
+    <string name="similarity_title">Similarity</string>
+    <string name="compare_selected_face_a">Selected face crop A</string>
+    <string name="compare_selected_face_b">Selected face crop B</string>
+    <string name="compare_face_a_badge" translatable="false">A</string>
+    <string name="compare_face_b_badge" translatable="false">B</string>
+    <string name="compare_select_two">Select two images</string>
+    <string name="compare_analyzing">Analyzing image…</string>
+    <string name="compare_waiting_face">Add two valid face images</string>
+    <string name="compare_comparing">Comparing faces…</string>
+    <string name="compare_same_person">Likely the same person</string>
+    <string name="compare_different_person">Likely different people</string>
+    <string name="compare_failed">Comparison failed</string>
+    <string name="compare_engine_failed">Face engine unavailable</string>
+    <string name="compare_details">cosine confidence %1$.3f · threshold %2$.3f</string>
+    <string name="on_device_privacy">Images are processed only on this device.</string>
+
+    <!-- Face management -->
+    <string name="face_management_title">Face management</string>
+    <string name="face_library_heading">Face library</string>
+    <string name="face_storage_scope">%1$s uses an isolated feature DB and crop directory.</string>
+    <string name="search_faces">Search name or ID</string>
+    <string name="add_face">Add</string>
+    <string name="no_faces_title">No faces stored</string>
+    <string name="no_faces_hint">Add a portrait to create the first identity in this model library.</string>
+    <string name="stored_face_crop">Stored face crop</string>
+    <string name="face_id_format">ID · %1$d</string>
+    <string name="edit_short">Edit</string>
+    <string name="delete_short">Del</string>
+    <string name="add_face_title">Add face</string>
+    <string name="edit_face_title">Edit face</string>
+    <string name="choose_enrollment_image">Use the camera or choose an image</string>
+    <string name="enrollment_image">Face enrollment image</string>
+    <string name="no_image_selected">No image selected</string>
+    <string name="capture_or_replace_face">Use camera</string>
+    <string name="choose_or_replace_image">Choose from gallery</string>
+    <string name="tap_face_to_select">If multiple faces are detected, tap a box to choose the identity to store.</string>
+    <string name="face_name">Name</string>
+    <string name="save">Save</string>
+    <string name="cancel">Cancel</string>
+    <string name="saving_face">Saving…</string>
+    <string name="face_saved">Face saved</string>
+    <string name="face_save_failed">Could not save this face</string>
+    <string name="face_library_failed">Could not open the model-specific face library</string>
+    <string name="delete_face_title">Delete face?</string>
+    <string name="delete_face_message">Delete %1$s from the current %2$s library? This removes its feature and crop.</string>
+    <string name="delete">Delete</string>
+    <string name="face_deleted">Face deleted</string>
+    <string name="face_delete_failed">Could not delete this face</string>
+    <string name="face_list_count">%1$d stored identities</string>
+
+    <!-- Face recognition -->
+    <string name="face_recognition_title">Face recognition</string>
+    <string name="face_recognition_heading">Who is this?</string>
+    <string name="face_recognition_hint">Search a face against identities enrolled for the current model.</string>
+    <string name="recognition_tab_photo">Photo input</string>
+    <string name="recognition_tab_video">Video stream</string>
+    <string name="recognition_session_parameters">Session parameters</string>
+    <string name="recognition_session_parameters_hint">Applying changes rebuilds the photo Session and saves these settings on this device.</string>
+    <string name="recognition_settings_expand">Expand</string>
+    <string name="recognition_settings_collapse">Collapse</string>
+    <string name="recognition_input_px">Detection input px</string>
+    <string name="recognition_max_faces">Maximum detected faces</string>
+    <string name="recognition_min_face_px">Minimum face px</string>
+    <string name="recognition_apply_session">Apply Session</string>
+    <string name="recognition_reset_defaults">Reset defaults</string>
+    <string name="recognition_session_summary">Session · %1$d px · max %2$d · min %3$d px</string>
+    <string name="recognition_session_rebuilding">Rebuilding Session…</string>
+    <string name="recognition_session_failed">Could not create this Session</string>
+    <string name="recognition_empty_library">No faces are enrolled for the %1$s model. Add identities in Face management first.</string>
+    <string name="recognition_photo">Recognition photo</string>
+    <string name="recognition_selected_face_crop">Selected face crop</string>
+    <string name="recognition_choose_photo_hint">Choose a photo to detect and identify faces</string>
+    <string name="recognition_choose_photo">Choose photo</string>
+    <string name="recognition_multi_face_hint">A single face is searched automatically. For multiple faces, tap a numbered box to switch.</string>
+    <string name="recognition_searching">Searching the current model library…</string>
+    <string name="recognition_match_found">Match found</string>
+    <string name="recognition_no_match">No matching identity</string>
+    <string name="recognition_library_empty_result">The current model library is empty</string>
+    <string name="recognition_result_name_unknown">Unknown</string>
+    <string name="recognition_result_details">ID · %1$d · cosine confidence %2$.3f · threshold %3$.3f</string>
+    <string name="recognition_no_match_details">cosine confidence %1$.3f · threshold %2$.3f</string>
+    <string name="recognition_no_confidence">No candidate exceeded the search threshold</string>
+    <string name="recognition_video_initializing">Starting camera…</string>
+    <string name="recognition_video_no_face">Place a face in the frame</string>
+    <string name="recognition_video_hold_still">Hold still for recognition</string>
+    <string name="recognition_video_first_face_hint">Only the first tracked face is identified after it stays stable for about 1 second.</string>
+    <string name="value_1" translatable="false">1</string>
+    <string name="value_3" translatable="false">3</string>
+    <string name="value_5" translatable="false">5</string>
+    <string name="value_10" translatable="false">10</string>
+    <string name="value_24" translatable="false">24</string>
+    <string name="value_48" translatable="false">48</string>
+    <string name="value_64" translatable="false">64</string>
+    <string name="value_128" translatable="false">128</string>
+    <string name="value_320" translatable="false">320</string>
+    <string name="value_640" translatable="false">640</string>
+    <string name="value_1280" translatable="false">1280</string>
+
+    <!-- Face detection and tracking -->
+    <string name="face_detection_title">Face tracking</string>
+    <string name="face_detection_heading">Find every face</string>
+    <string name="face_detection_hint">Detect selectable faces and inspect the SDK native 106-point dense landmarks.</string>
+    <string name="detection_tab_image">Image detection</string>
+    <string name="detection_tab_tracking">Video tracking</string>
+    <string name="detection_show_dense_landmarks">Show 106 dense landmarks</string>
+    <string name="detection_photo">Face detection photo</string>
+    <string name="detection_choose_photo_hint">Choose a photo to detect faces and landmarks</string>
+    <string name="detection_choose_photo">Choose detection photo</string>
+    <string name="detection_multi_face_hint">Face 1 is selected by default. Tap any numbered box to inspect that face.</string>
+    <string name="detection_landmark_behavior_hint">Large faces render landmarks on the source; small faces use the bottom-right magnifier.</string>
+    <string name="detection_face_selected">Face %1$d of %2$d selected · %3$d landmarks returned</string>
+    <string name="face_detection_failed">Face detection failed</string>
+    <string name="detection_landmarks_hidden">106 landmarks are hidden</string>
+    <string name="detection_landmarks_failed">Could not read 106 landmarks for this face</string>
+    <string name="detection_landmarks_on_source">Large face · %1$d%% of image · landmarks shown on source</string>
+    <string name="detection_landmarks_in_magnifier">Small face · %1$d%% of image · landmarks enlarged at bottom right</string>
+    <string name="detection_landmark_magnifier">Selected face magnifier</string>
+    <string name="detection_106_badge" translatable="false">106 PTS</string>
+    <string name="detection_tracking_initializing">Starting camera tracking…</string>
+    <string name="detection_tracking_restarting">Applying settings and restarting camera…</string>
+    <string name="detection_tracking_stats">%1$.1f FPS · %2$d ms</string>
+    <string name="detection_tracking_settings">Tracking Session parameters</string>
+    <string name="detection_tracking_settings_hint">Changes are saved automatically and restart the camera stream.</string>
+    <string name="detection_tracking_mode">Tracking mode</string>
+    <string name="detection_tracking_mode_light">Light</string>
+    <string name="detection_tracking_mode_tbd">Track-by-detection</string>
+    <string name="detection_tracking_summary">%1$s · %2$d px · max %3$d · min %4$d px</string>
+    <string name="detection_tracking_session_failed">Could not create this tracking Session</string>
+
+    <!-- Face attributes -->
+    <string name="face_attribute_title">Face attributes</string>
+    <string name="face_attribute_heading">Read facial attributes</string>
+    <string name="face_attribute_hint">Choose a photo, then tap any numbered face to update its attributes.</string>
+    <string name="attribute_photo">Face attribute photo</string>
+    <string name="attribute_choose_photo_hint">Choose a photo to analyze face attributes</string>
+    <string name="attribute_choose_photo">Choose photo</string>
+    <string name="attribute_multi_face_hint">Face 1 is selected by default. Small selected faces appear in the bottom-right magnifier.</string>
+    <string name="attribute_selected_face_crop">Selected face magnifier</string>
+    <string name="attribute_face_selected">Face %1$d of %2$d selected</string>
+    <string name="attribute_result_title">Selected face</string>
+    <string name="attribute_result_empty">Choose a photo to view attributes</string>
+    <string name="attribute_analyzing">Analyzing attributes…</string>
+    <string name="attribute_analysis_failed">Could not analyze face attributes</string>
+    <string name="attribute_no_face_result">No face attributes available</string>
+    <string name="attribute_mask">Wearing mask</string>
+    <string name="attribute_age">Age bracket</string>
+    <string name="attribute_quality">Image quality score</string>
+    <string name="attribute_expression">Expression</string>
+    <string name="attribute_race">Ethnicity</string>
+    <string name="attribute_gender">Gender</string>
+    <string name="attribute_left_eye">Left eye</string>
+    <string name="attribute_right_eye">Right eye</string>
+    <string name="attribute_mask_yes">Yes</string>
+    <string name="attribute_mask_no">No</string>
+    <string name="attribute_eye_open">Open</string>
+    <string name="attribute_eye_closed">Closed</string>
+    <string name="attribute_expression_neutral">Neutral</string>
+    <string name="attribute_expression_mouth_open">Mouth open</string>
+    <string name="attribute_gender_female">Female</string>
+    <string name="attribute_gender_male">Male</string>
+    <string name="attribute_race_black">Black</string>
+    <string name="attribute_race_asian">Asian</string>
+    <string name="attribute_race_latino">Latino / Hispanic</string>
+    <string name="attribute_race_middle_eastern">Middle Eastern</string>
+    <string name="attribute_race_white">White</string>
+    <string name="attribute_age_0_2">0–2</string>
+    <string name="attribute_age_3_9">3–9</string>
+    <string name="attribute_age_10_19">10–19</string>
+    <string name="attribute_age_20_29">20–29</string>
+    <string name="attribute_age_30_39">30–39</string>
+    <string name="attribute_age_40_49">40–49</string>
+    <string name="attribute_age_50_59">50–59</string>
+    <string name="attribute_age_60_69">60–69</string>
+    <string name="attribute_age_70_plus">70+</string>
+    <string name="attribute_score_format" translatable="false">%1$.3f</string>
+    <string name="attribute_unknown">—</string>
+
+    <!-- Camera face enrollment -->
+    <string name="face_capture_title">Camera enrollment</string>
+    <string name="capture_permission_hint">Allow camera access, then try again.</string>
+    <string name="capture_no_face">Place a face in the frame</string>
+    <string name="capture_first_face_hint">The first detected face is tracked. Keep it clear and centered.</string>
+    <string name="capture_move_closer">Move closer</string>
+    <string name="capture_hold_still">Hold still</string>
+    <string name="capture_warmup_hint">Stay steady for 1 second to start the ring.</string>
+    <string name="capture_keep_still">Keep holding still</string>
+    <string name="capture_progress">Stability %1$d%%</string>
+    <string name="capture_complete">Capture complete</string>
+    <string name="capture_complete_hint">The ring is full. Preparing this face…</string>
+    <string name="capture_failed">Could not capture this frame</string>
+    <string name="capture_retry_hint">Hold still and the app will try again.</string>
+</resources>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/themes.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/themes.xml
new file mode 100644
index 000000000..931e28ebb
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/values/themes.xml
@@ -0,0 +1,49 @@
+<resources xmlns:tools="http://schemas.android.com/tools">
+    <!-- Base application theme. -->
+    <style name="Base.Theme.InspireFaceExample" parent="Theme.Material3.DayNight.NoActionBar">
+        <!-- Customize your light theme here. -->
+        <!-- <item name="colorPrimary">@color/my_light_primary</item> -->
+    </style>
+
+    <style name="Theme.InspireFaceExample" parent="Base.Theme.InspireFaceExample" />
+
+    <style name="Theme.InspireFaceExample.Home">
+        <item name="android:windowBackground">@color/home_background</item>
+        <item name="android:statusBarColor">@android:color/transparent</item>
+        <item name="android:navigationBarColor">@color/home_background</item>
+        <item name="android:windowLightStatusBar">false</item>
+    </style>
+
+    <!-- Full-bleed dark theme for the camera liveness screen -->
+    <style name="Theme.InspireFaceExample.Liveness">
+        <item name="android:windowBackground">@android:color/black</item>
+        <item name="android:statusBarColor">@android:color/transparent</item>
+        <item name="android:navigationBarColor">@android:color/transparent</item>
+        <item name="android:windowLightStatusBar">false</item>
+    </style>
+
+    <style name="AttributeResultRow">
+        <item name="android:layout_width">match_parent</item>
+        <item name="android:layout_height">48dp</item>
+        <item name="android:gravity">center_vertical</item>
+        <item name="android:orientation">horizontal</item>
+    </style>
+
+    <style name="AttributeResultLabel">
+        <item name="android:layout_width">0dp</item>
+        <item name="android:layout_height">wrap_content</item>
+        <item name="android:layout_weight">1</item>
+        <item name="android:textColor">@color/home_text_secondary</item>
+        <item name="android:textSize">14sp</item>
+    </style>
+
+    <style name="AttributeResultValue">
+        <item name="android:layout_width">0dp</item>
+        <item name="android:layout_height">wrap_content</item>
+        <item name="android:layout_weight">1</item>
+        <item name="android:gravity">end</item>
+        <item name="android:textColor">@color/liveness_accent</item>
+        <item name="android:textSize">15sp</item>
+        <item name="android:textStyle">bold</item>
+    </style>
+</resources>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/backup_rules.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/backup_rules.xml
new file mode 100644
index 000000000..26eaf574a
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/backup_rules.xml
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="utf-8"?>
+<full-backup-content>
+    <!-- Android 7–11 equivalent of data_extraction_rules.xml. -->
+    <exclude domain="file" path="face_hub/" />
+    <exclude domain="sharedpref" path="face_records_Pikachu.xml" />
+    <exclude domain="sharedpref" path="face_records_Megatron.xml" />
+    <exclude domain="sharedpref" path="camera_permission.xml" />
+</full-backup-content>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/data_extraction_rules.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/data_extraction_rules.xml
new file mode 100644
index 000000000..a1513d5f2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/data_extraction_rules.xml
@@ -0,0 +1,17 @@
+<?xml version="1.0" encoding="utf-8"?>
+<data-extraction-rules>
+    <!-- Face crops, native feature databases, identity metadata and the per-device
+         permission notice state must never leave this device through backup/transfer. -->
+    <cloud-backup disableIfNoEncryptionCapabilities="true">
+        <exclude domain="file" path="face_hub/" />
+        <exclude domain="sharedpref" path="face_records_Pikachu.xml" />
+        <exclude domain="sharedpref" path="face_records_Megatron.xml" />
+        <exclude domain="sharedpref" path="camera_permission.xml" />
+    </cloud-backup>
+    <device-transfer>
+        <exclude domain="file" path="face_hub/" />
+        <exclude domain="sharedpref" path="face_records_Pikachu.xml" />
+        <exclude domain="sharedpref" path="face_records_Megatron.xml" />
+        <exclude domain="sharedpref" path="camera_permission.xml" />
+    </device-transfer>
+</data-extraction-rules>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/locales_config.xml b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/locales_config.xml
new file mode 100644
index 000000000..5ae290080
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/main/res/xml/locales_config.xml
@@ -0,0 +1,5 @@
+<?xml version="1.0" encoding="utf-8"?>
+<locale-config xmlns:android="http://schemas.android.com/apk/res/android">
+    <locale android:name="en" />
+    <locale android:name="zh" />
+</locale-config>
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/DetectorDefaultsTest.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/DetectorDefaultsTest.java
new file mode 100644
index 000000000..253fef3ae
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/DetectorDefaultsTest.java
@@ -0,0 +1,17 @@
+package com.example.inspireface_example;
+
+import static org.junit.Assert.assertEquals;
+
+import org.junit.Test;
+
+public final class DetectorDefaultsTest {
+
+    @Test
+    public void allSessionDefaultsUse320InputPixels() {
+        assertEquals(320, DetectorDefaults.INPUT_PX);
+        assertEquals(DetectorDefaults.INPUT_PX,
+                StillImageSessionSettings.defaults().inputPx);
+        assertEquals(DetectorDefaults.INPUT_PX,
+                FaceTrackingSessionSettings.defaults().inputPx);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java
new file mode 100644
index 000000000..67bb0493e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java
@@ -0,0 +1,17 @@
+package com.example.inspireface_example;
+
+import org.junit.Test;
+
+import static org.junit.Assert.*;
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * @see <a href="http://d.android.com/tools/testing">Testing documentation</a>
+ */
+public class ExampleUnitTest {
+    @Test
+    public void addition_isCorrect() {
+        assertEquals(4, 2 + 2);
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/permission/CameraPermissionPolicyTest.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/permission/CameraPermissionPolicyTest.java
new file mode 100644
index 000000000..bb8b2b967
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/permission/CameraPermissionPolicyTest.java
@@ -0,0 +1,45 @@
+package com.example.inspireface_example.permission;
+
+import static com.example.inspireface_example.permission.CameraPermissionPolicy.Action.GRANTED;
+import static com.example.inspireface_example.permission.CameraPermissionPolicy.Action.OPEN_SETTINGS;
+import static com.example.inspireface_example.permission.CameraPermissionPolicy.Action.REQUEST_SYSTEM_PERMISSION;
+import static com.example.inspireface_example.permission.CameraPermissionPolicy.Action.SHOW_NOTICE;
+import static com.example.inspireface_example.permission.CameraPermissionPolicy.Action.SHOW_RATIONALE;
+import static org.junit.Assert.assertEquals;
+
+import org.junit.Test;
+
+public final class CameraPermissionPolicyTest {
+
+    @Test
+    public void firstCameraUseAlwaysShowsPrivacyNotice() {
+        assertEquals(SHOW_NOTICE,
+                CameraPermissionPolicy.nextAction(false, false, false, false));
+        assertEquals(SHOW_NOTICE,
+                CameraPermissionPolicy.nextAction(false, true, true, false));
+    }
+
+    @Test
+    public void acceptedNoticeAndGrantedPermissionStartsCamera() {
+        assertEquals(GRANTED,
+                CameraPermissionPolicy.nextAction(true, true, true, false));
+    }
+
+    @Test
+    public void firstAndroidPermissionRequestUsesSystemDialog() {
+        assertEquals(REQUEST_SYSTEM_PERMISSION,
+                CameraPermissionPolicy.nextAction(true, false, false, false));
+    }
+
+    @Test
+    public void retryableDenialShowsRationale() {
+        assertEquals(SHOW_RATIONALE,
+                CameraPermissionPolicy.nextAction(true, false, true, true));
+    }
+
+    @Test
+    public void permanentDenialRequiresApplicationSettings() {
+        assertEquals(OPEN_SETTINGS,
+                CameraPermissionPolicy.nextAction(true, false, true, false));
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/view/FaceStabilityGateTest.java b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/view/FaceStabilityGateTest.java
new file mode 100644
index 000000000..75e200217
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/app/src/test/java/com/example/inspireface_example/view/FaceStabilityGateTest.java
@@ -0,0 +1,53 @@
+package com.example.inspireface_example.view;
+
+import org.junit.Test;
+
+import static org.junit.Assert.assertEquals;
+
+public class FaceStabilityGateTest {
+
+    private static final float EPSILON = 0.0001f;
+
+    @Test
+    public void stableFaceWaitsOneSecondThenFillsForTwoSeconds() {
+        FaceStabilityGate gate = new FaceStabilityGate();
+
+        assertEquals(-1f, update(gate, 7, 0L), EPSILON);
+        assertEquals(-1f, update(gate, 7, 999L), EPSILON);
+        assertEquals(0f, update(gate, 7, 1_000L), EPSILON);
+        assertEquals(0.5f, update(gate, 7, 2_000L), EPSILON);
+        assertEquals(1f, update(gate, 7, 3_000L), EPSILON);
+    }
+
+    @Test
+    public void motionHidesRingAndStartsANewWarmup() {
+        FaceStabilityGate gate = new FaceStabilityGate();
+        update(gate, 7, 0L);
+        assertEquals(0.25f, update(gate, 7, 1_500L), EPSILON);
+
+        assertEquals(-1f, gate.update(7, 40f, 0f, 140f, 100f, 1_600L), EPSILON);
+        assertEquals(-1f, gate.update(7, 40f, 0f, 140f, 100f, 2_599L), EPSILON);
+        assertEquals(0f, gate.update(7, 40f, 0f, 140f, 100f, 2_600L), EPSILON);
+    }
+
+    @Test
+    public void switchingFirstFaceRestartsWarmup() {
+        FaceStabilityGate gate = new FaceStabilityGate();
+        update(gate, 7, 0L);
+        assertEquals(0.25f, update(gate, 7, 1_500L), EPSILON);
+        assertEquals(-1f, update(gate, 9, 1_600L), EPSILON);
+    }
+
+    @Test
+    public void configurableGateCanTriggerRecognitionAfterOneSecond() {
+        FaceStabilityGate gate = new FaceStabilityGate(1_000L, 1L);
+
+        assertEquals(-1f, update(gate, 7, 0L), EPSILON);
+        assertEquals(-1f, update(gate, 7, 999L), EPSILON);
+        assertEquals(0f, update(gate, 7, 1_000L), EPSILON);
+    }
+
+    private static float update(FaceStabilityGate gate, int id, long timeMs) {
+        return gate.update(id, 0f, 0f, 100f, 100f, timeMs);
+    }
+}
diff --git a/cpp-package/inspireface/android/InspireFaceExample/build.gradle b/cpp-package/inspireface/android/InspireFaceExample/build.gradle
new file mode 100644
index 000000000..99e73ad4c
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/build.gradle
@@ -0,0 +1,5 @@
+// Top-level build file where you can add configuration options common to all sub-projects/modules.
+plugins {
+alias(libs.plugins.android.application) apply false
+    alias(libs.plugins.android.library) apply false
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/docs/images/inspireface-android-example-app-download.png b/cpp-package/inspireface/android/InspireFaceExample/docs/images/inspireface-android-example-app-download.png
new file mode 100644
index 000000000..3bbc8fa02
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/docs/images/inspireface-android-example-app-download.png differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradle.properties b/cpp-package/inspireface/android/InspireFaceExample/gradle.properties
new file mode 100644
index 000000000..4387edc22
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/gradle.properties
@@ -0,0 +1,21 @@
+# Project-wide Gradle settings.
+# IDE (e.g. Android Studio) users:
+# Gradle settings configured through the IDE *will override*
+# any settings specified in this file.
+# For more details on how to configure your build environment visit
+# http://www.gradle.org/docs/current/userguide/build_environment.html
+# Specifies the JVM arguments used for the daemon process.
+# The setting is particularly useful for tweaking memory settings.
+org.gradle.jvmargs=-Xmx2048m -Dfile.encoding=UTF-8
+# When configured, Gradle will run in incubating parallel mode.
+# This option should only be used with decoupled projects. For more details, visit
+# https://developer.android.com/r/tools/gradle-multi-project-decoupled-projects
+# org.gradle.parallel=true
+# AndroidX package structure to make it clearer which packages are bundled with the
+# Android operating system, and which are packaged with your app's APK
+# https://developer.android.com/topic/libraries/support-library/androidx-rn
+android.useAndroidX=true
+# Enables namespacing of each library's R class so that its R class includes only the
+# resources declared in the library itself and none from the library's dependencies,
+# thereby reducing the size of the R class for that library
+android.nonTransitiveRClass=true
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradle/libs.versions.toml b/cpp-package/inspireface/android/InspireFaceExample/gradle/libs.versions.toml
new file mode 100644
index 000000000..13dd9f2fe
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/gradle/libs.versions.toml
@@ -0,0 +1,31 @@
+[versions]
+agp = "8.6.1"
+inspirefaceAndroidSdk = "1.2.0"
+junit = "4.13.2"
+junitVersion = "1.1.5"
+espressoCore = "3.5.1"
+appcompat = "1.6.1"
+material = "1.10.0"
+activity = "1.8.0"
+constraintlayout = "2.1.4"
+camerax = "1.5.3"
+exifinterface = "1.3.2"
+
+[libraries]
+inspireface-android-sdk = { module = "com.github.HyperInspire:inspireface-android-sdk", version.ref = "inspirefaceAndroidSdk" }
+junit = { group = "junit", name = "junit", version.ref = "junit" }
+ext-junit = { group = "androidx.test.ext", name = "junit", version.ref = "junitVersion" }
+espresso-core = { group = "androidx.test.espresso", name = "espresso-core", version.ref = "espressoCore" }
+appcompat = { group = "androidx.appcompat", name = "appcompat", version.ref = "appcompat" }
+material = { group = "com.google.android.material", name = "material", version.ref = "material" }
+activity = { group = "androidx.activity", name = "activity", version.ref = "activity" }
+constraintlayout = { group = "androidx.constraintlayout", name = "constraintlayout", version.ref = "constraintlayout" }
+camera-core = { group = "androidx.camera", name = "camera-core", version.ref = "camerax" }
+camera-camera2 = { group = "androidx.camera", name = "camera-camera2", version.ref = "camerax" }
+camera-lifecycle = { group = "androidx.camera", name = "camera-lifecycle", version.ref = "camerax" }
+camera-view = { group = "androidx.camera", name = "camera-view", version.ref = "camerax" }
+exifinterface = { group = "androidx.exifinterface", name = "exifinterface", version.ref = "exifinterface" }
+
+[plugins]
+android-application = { id = "com.android.application", version.ref = "agp" }
+android-library = { id = "com.android.library", version.ref = "agp" }
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.jar b/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.jar
new file mode 100644
index 000000000..e708b1c02
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.jar differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.properties b/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.properties
new file mode 100644
index 000000000..e846f349e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/gradle/wrapper/gradle-wrapper.properties
@@ -0,0 +1,6 @@
+#Tue Nov 26 15:12:19 CST 2024
+distributionBase=GRADLE_USER_HOME
+distributionPath=wrapper/dists
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.7-bin.zip
+zipStoreBase=GRADLE_USER_HOME
+zipStorePath=wrapper/dists
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradlew b/cpp-package/inspireface/android/InspireFaceExample/gradlew
new file mode 100755
index 000000000..4f906e0c8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/gradlew
@@ -0,0 +1,185 @@
+#!/usr/bin/env sh
+
+#
+# Copyright 2015 the original author or authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+##############################################################################
+##
+##  Gradle start up script for UN*X
+##
+##############################################################################
+
+# Attempt to set APP_HOME
+# Resolve links: $0 may be a link
+PRG="$0"
+# Need this for relative symlinks.
+while [ -h "$PRG" ] ; do
+    ls=`ls -ld "$PRG"`
+    link=`expr "$ls" : '.*-> \(.*\)$'`
+    if expr "$link" : '/.*' > /dev/null; then
+        PRG="$link"
+    else
+        PRG=`dirname "$PRG"`"/$link"
+    fi
+done
+SAVED="`pwd`"
+cd "`dirname \"$PRG\"`/" >/dev/null
+APP_HOME="`pwd -P`"
+cd "$SAVED" >/dev/null
+
+APP_NAME="Gradle"
+APP_BASE_NAME=`basename "$0"`
+
+# Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+DEFAULT_JVM_OPTS='"-Xmx64m" "-Xms64m"'
+
+# Use the maximum available, or set MAX_FD != -1 to use that value.
+MAX_FD="maximum"
+
+warn () {
+    echo "$*"
+}
+
+die () {
+    echo
+    echo "$*"
+    echo
+    exit 1
+}
+
+# OS specific support (must be 'true' or 'false').
+cygwin=false
+msys=false
+darwin=false
+nonstop=false
+case "`uname`" in
+  CYGWIN* )
+    cygwin=true
+    ;;
+  Darwin* )
+    darwin=true
+    ;;
+  MINGW* )
+    msys=true
+    ;;
+  NONSTOP* )
+    nonstop=true
+    ;;
+esac
+
+CLASSPATH=$APP_HOME/gradle/wrapper/gradle-wrapper.jar
+
+
+# Determine the Java command to use to start the JVM.
+if [ -n "$JAVA_HOME" ] ; then
+    if [ -x "$JAVA_HOME/jre/sh/java" ] ; then
+        # IBM's JDK on AIX uses strange locations for the executables
+        JAVACMD="$JAVA_HOME/jre/sh/java"
+    else
+        JAVACMD="$JAVA_HOME/bin/java"
+    fi
+    if [ ! -x "$JAVACMD" ] ; then
+        die "ERROR: JAVA_HOME is set to an invalid directory: $JAVA_HOME
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+    fi
+else
+    JAVACMD="java"
+    which java >/dev/null 2>&1 || die "ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+fi
+
+# Increase the maximum file descriptors if we can.
+if [ "$cygwin" = "false" -a "$darwin" = "false" -a "$nonstop" = "false" ] ; then
+    MAX_FD_LIMIT=`ulimit -H -n`
+    if [ $? -eq 0 ] ; then
+        if [ "$MAX_FD" = "maximum" -o "$MAX_FD" = "max" ] ; then
+            MAX_FD="$MAX_FD_LIMIT"
+        fi
+        ulimit -n $MAX_FD
+        if [ $? -ne 0 ] ; then
+            warn "Could not set maximum file descriptor limit: $MAX_FD"
+        fi
+    else
+        warn "Could not query maximum file descriptor limit: $MAX_FD_LIMIT"
+    fi
+fi
+
+# For Darwin, add options to specify how the application appears in the dock
+if $darwin; then
+    GRADLE_OPTS="$GRADLE_OPTS \"-Xdock:name=$APP_NAME\" \"-Xdock:icon=$APP_HOME/media/gradle.icns\""
+fi
+
+# For Cygwin or MSYS, switch paths to Windows format before running java
+if [ "$cygwin" = "true" -o "$msys" = "true" ] ; then
+    APP_HOME=`cygpath --path --mixed "$APP_HOME"`
+    CLASSPATH=`cygpath --path --mixed "$CLASSPATH"`
+
+    JAVACMD=`cygpath --unix "$JAVACMD"`
+
+    # We build the pattern for arguments to be converted via cygpath
+    ROOTDIRSRAW=`find -L / -maxdepth 1 -mindepth 1 -type d 2>/dev/null`
+    SEP=""
+    for dir in $ROOTDIRSRAW ; do
+        ROOTDIRS="$ROOTDIRS$SEP$dir"
+        SEP="|"
+    done
+    OURCYGPATTERN="(^($ROOTDIRS))"
+    # Add a user-defined pattern to the cygpath arguments
+    if [ "$GRADLE_CYGPATTERN" != "" ] ; then
+        OURCYGPATTERN="$OURCYGPATTERN|($GRADLE_CYGPATTERN)"
+    fi
+    # Now convert the arguments - kludge to limit ourselves to /bin/sh
+    i=0
+    for arg in "$@" ; do
+        CHECK=`echo "$arg"|egrep -c "$OURCYGPATTERN" -`
+        CHECK2=`echo "$arg"|egrep -c "^-"`                                 ### Determine if an option
+
+        if [ $CHECK -ne 0 ] && [ $CHECK2 -eq 0 ] ; then                    ### Added a condition
+            eval `echo args$i`=`cygpath --path --ignore --mixed "$arg"`
+        else
+            eval `echo args$i`="\"$arg\""
+        fi
+        i=`expr $i + 1`
+    done
+    case $i in
+        0) set -- ;;
+        1) set -- "$args0" ;;
+        2) set -- "$args0" "$args1" ;;
+        3) set -- "$args0" "$args1" "$args2" ;;
+        4) set -- "$args0" "$args1" "$args2" "$args3" ;;
+        5) set -- "$args0" "$args1" "$args2" "$args3" "$args4" ;;
+        6) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" ;;
+        7) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" ;;
+        8) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" "$args7" ;;
+        9) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" "$args7" "$args8" ;;
+    esac
+fi
+
+# Escape application args
+save () {
+    for i do printf %s\\n "$i" | sed "s/'/'\\\\''/g;1s/^/'/;\$s/\$/' \\\\/" ; done
+    echo " "
+}
+APP_ARGS=`save "$@"`
+
+# Collect all arguments for the java command, following the shell quoting and substitution rules
+eval set -- $DEFAULT_JVM_OPTS $JAVA_OPTS $GRADLE_OPTS "\"-Dorg.gradle.appname=$APP_BASE_NAME\"" -classpath "\"$CLASSPATH\"" org.gradle.wrapper.GradleWrapperMain "$APP_ARGS"
+
+exec "$JAVACMD" "$@"
diff --git a/cpp-package/inspireface/android/InspireFaceExample/gradlew.bat b/cpp-package/inspireface/android/InspireFaceExample/gradlew.bat
new file mode 100644
index 000000000..ac1b06f93
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/gradlew.bat
@@ -0,0 +1,89 @@
+@rem
+@rem Copyright 2015 the original author or authors.
+@rem
+@rem Licensed under the Apache License, Version 2.0 (the "License");
+@rem you may not use this file except in compliance with the License.
+@rem You may obtain a copy of the License at
+@rem
+@rem      https://www.apache.org/licenses/LICENSE-2.0
+@rem
+@rem Unless required by applicable law or agreed to in writing, software
+@rem distributed under the License is distributed on an "AS IS" BASIS,
+@rem WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+@rem See the License for the specific language governing permissions and
+@rem limitations under the License.
+@rem
+
+@if "%DEBUG%" == "" @echo off
+@rem ##########################################################################
+@rem
+@rem  Gradle startup script for Windows
+@rem
+@rem ##########################################################################
+
+@rem Set local scope for the variables with windows NT shell
+if "%OS%"=="Windows_NT" setlocal
+
+set DIRNAME=%~dp0
+if "%DIRNAME%" == "" set DIRNAME=.
+set APP_BASE_NAME=%~n0
+set APP_HOME=%DIRNAME%
+
+@rem Resolve any "." and ".." in APP_HOME to make it shorter.
+for %%i in ("%APP_HOME%") do set APP_HOME=%%~fi
+
+@rem Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+set DEFAULT_JVM_OPTS="-Xmx64m" "-Xms64m"
+
+@rem Find java.exe
+if defined JAVA_HOME goto findJavaFromJavaHome
+
+set JAVA_EXE=java.exe
+%JAVA_EXE% -version >NUL 2>&1
+if "%ERRORLEVEL%" == "0" goto execute
+
+echo.
+echo ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:findJavaFromJavaHome
+set JAVA_HOME=%JAVA_HOME:"=%
+set JAVA_EXE=%JAVA_HOME%/bin/java.exe
+
+if exist "%JAVA_EXE%" goto execute
+
+echo.
+echo ERROR: JAVA_HOME is set to an invalid directory: %JAVA_HOME%
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:execute
+@rem Setup the command line
+
+set CLASSPATH=%APP_HOME%\gradle\wrapper\gradle-wrapper.jar
+
+
+@rem Execute Gradle
+"%JAVA_EXE%" %DEFAULT_JVM_OPTS% %JAVA_OPTS% %GRADLE_OPTS% "-Dorg.gradle.appname=%APP_BASE_NAME%" -classpath "%CLASSPATH%" org.gradle.wrapper.GradleWrapperMain %*
+
+:end
+@rem End local scope for the variables with windows NT shell
+if "%ERRORLEVEL%"=="0" goto mainEnd
+
+:fail
+rem Set variable GRADLE_EXIT_CONSOLE if you need the _script_ return code instead of
+rem the _cmd.exe /c_ return code!
+if  not "" == "%GRADLE_EXIT_CONSOLE%" exit 1
+exit /b 1
+
+:mainEnd
+if "%OS%"=="Windows_NT" endlocal
+
+:omega
diff --git a/cpp-package/inspireface/android/InspireFaceExample/settings.gradle b/cpp-package/inspireface/android/InspireFaceExample/settings.gradle
new file mode 100644
index 000000000..5e4119ae9
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample/settings.gradle
@@ -0,0 +1,24 @@
+pluginManagement {
+    repositories {
+        google {
+            content {
+                includeGroupByRegex("com\\.android.*")
+                includeGroupByRegex("com\\.google.*")
+                includeGroupByRegex("androidx.*")
+            }
+        }
+        mavenCentral()
+        gradlePluginPortal()
+    }
+}
+dependencyResolutionManagement {
+    repositoriesMode.set(RepositoriesMode.FAIL_ON_PROJECT_REPOS)
+    repositories {
+        google()
+        mavenCentral()
+        maven { url 'https://jitpack.io' }
+    }
+}
+
+rootProject.name = "InspireFace-Example"
+include ':app'
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/.gitignore b/cpp-package/inspireface/android/InspireFaceExample_old/.gitignore
new file mode 100644
index 000000000..57042a8df
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/.gitignore
@@ -0,0 +1,16 @@
+.idea/
+*.iml
+.gradle
+/local.properties
+/.idea/caches
+/.idea/libraries
+/.idea/modules.xml
+/.idea/workspace.xml
+/.idea/navEditor.xml
+/.idea/assetWizardSettings.xml
+.DS_Store
+/build
+/captures
+.externalNativeBuild
+.cxx
+local.properties
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/.gitignore b/cpp-package/inspireface/android/InspireFaceExample_old/app/.gitignore
new file mode 100644
index 000000000..42afabfd2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/.gitignore
@@ -0,0 +1 @@
+/build
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/build.gradle b/cpp-package/inspireface/android/InspireFaceExample_old/app/build.gradle
new file mode 100644
index 000000000..d6f091816
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/build.gradle
@@ -0,0 +1,41 @@
+plugins {
+    alias(libs.plugins.android.application)
+}
+
+android {
+    namespace 'com.example.inspireface_example'
+    compileSdk 34
+
+    defaultConfig {
+        applicationId "com.example.inspireface_example"
+        minSdk 24
+        targetSdk 34
+        versionCode 1
+        versionName "1.0"
+
+        testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
+    }
+
+    buildTypes {
+        release {
+            minifyEnabled false
+            proguardFiles getDefaultProguardFile('proguard-android-optimize.txt'), 'proguard-rules.pro'
+        }
+    }
+    compileOptions {
+        sourceCompatibility JavaVersion.VERSION_1_8
+        targetCompatibility JavaVersion.VERSION_1_8
+    }
+}
+
+dependencies {
+
+    implementation libs.appcompat
+    implementation libs.material
+    implementation libs.activity
+    implementation libs.constraintlayout
+    testImplementation libs.junit
+    androidTestImplementation libs.ext.junit
+    androidTestImplementation libs.espresso.core
+    implementation libs.inspireface.android.sdk
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/proguard-rules.pro b/cpp-package/inspireface/android/InspireFaceExample_old/app/proguard-rules.pro
new file mode 100644
index 000000000..481bb4348
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/proguard-rules.pro
@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java
new file mode 100644
index 000000000..bfd1c0ded
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/androidTest/java/com/example/inspireface_example/ExampleInstrumentedTest.java
@@ -0,0 +1,26 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+
+import androidx.test.platform.app.InstrumentationRegistry;
+import androidx.test.ext.junit.runners.AndroidJUnit4;
+
+import org.junit.Test;
+import org.junit.runner.RunWith;
+
+import static org.junit.Assert.*;
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * @see <a href="http://d.android.com/tools/testing">Testing documentation</a>
+ */
+@RunWith(AndroidJUnit4.class)
+public class ExampleInstrumentedTest {
+    @Test
+    public void useAppContext() {
+        // Context of the app under test.
+        Context appContext = InstrumentationRegistry.getInstrumentation().getTargetContext();
+        assertEquals("com.example.inspireface_example", appContext.getPackageName());
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/AndroidManifest.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/AndroidManifest.xml
new file mode 100644
index 000000000..388f45c88
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/AndroidManifest.xml
@@ -0,0 +1,34 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:tools="http://schemas.android.com/tools">
+
+    <!-- Phone information -->
+    <uses-permission android:name="android.permission.READ_PHONE_STATE" />
+    <!-- ************************************* -->
+    <uses-permission android:name="android.permission.WRITE_EXTERNAL_STORAGE" />
+    <!-- Create and delete file permissions in SD card -->
+    <uses-permission android:name="android.permission.MOUNT_UNMOUNT_FILESYSTEMS"
+        tools:ignore="ProtectedPermissions" />
+
+    <application
+        android:allowBackup="true"
+        android:dataExtractionRules="@xml/data_extraction_rules"
+        android:fullBackupContent="@xml/backup_rules"
+        android:icon="@mipmap/ic_launcher"
+        android:label="@string/app_name"
+        android:roundIcon="@mipmap/ic_launcher_round"
+        android:supportsRtl="true"
+        android:theme="@style/Theme.InspireFaceExample"
+        tools:targetApi="31">
+        <activity
+            android:name=".MainActivity"
+            android:exported="true">
+            <intent-filter>
+                <action android:name="android.intent.action.MAIN" />
+
+                <category android:name="android.intent.category.LAUNCHER" />
+            </intent-filter>
+        </activity>
+    </application>
+
+</manifest>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/java/com/example/inspireface_example/MainActivity.java b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/java/com/example/inspireface_example/MainActivity.java
new file mode 100644
index 000000000..1414fcd5a
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/java/com/example/inspireface_example/MainActivity.java
@@ -0,0 +1,196 @@
+package com.example.inspireface_example;
+
+import android.content.Context;
+import android.content.res.AssetManager;
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.os.Bundle;
+import android.util.Log;
+
+import androidx.activity.EdgeToEdge;
+import androidx.appcompat.app.AppCompatActivity;
+import androidx.core.graphics.Insets;
+import androidx.core.view.ViewCompat;
+import androidx.core.view.WindowInsetsCompat;
+
+import com.insightface.sdk.inspireface.InspireFace;
+import com.insightface.sdk.inspireface.base.*;
+import com.insightface.sdk.inspireface.utils.SDKUtils;
+
+import java.io.IOException;
+
+public class MainActivity extends AppCompatActivity {
+
+    private final String TAG = "InspireFace";
+
+    void test() {
+        InspireFaceVersion version = InspireFace.QueryInspireFaceVersion();
+        Log.i(TAG, "InspireFace Version: " + version.major + "." + version.minor + "." + version.patch);
+        String dbPath = "/storage/emulated/0/Android/data/com.example.inspireface_example/files/f.db";
+        FeatureHubConfiguration configuration = InspireFace.CreateFeatureHubConfiguration()
+                .setEnablePersistence(false)
+                .setPersistenceDbPath(dbPath)
+                .setSearchThreshold(0.42f)
+                .setSearchMode(InspireFace.SEARCH_MODE_EXHAUSTIVE)
+                .setPrimaryKeyMode(InspireFace.PK_AUTO_INCREMENT);
+
+        boolean enableStatus = InspireFace.FeatureHubDataEnable(configuration);
+        Log.d(TAG, "Enable feature hub data status: " + enableStatus);
+        InspireFace.FeatureHubFaceSearchThresholdSetting(0.42f);
+
+        boolean launchStatus = InspireFace.GlobalLaunch(this, InspireFace.PIKACHU);
+        Log.d(TAG, "Launch status: " + launchStatus);
+        if (!launchStatus) {
+            Log.e(TAG, "Failed to launch InspireFace");
+            return;
+        }
+        CustomParameter parameter = InspireFace.CreateCustomParameter()
+                .enableRecognition(true)
+                .enableFaceQuality(true)
+                .enableFaceAttribute(true)
+                .enableInteractionLiveness(true)
+                .enableLiveness(true)
+                .enableMaskDetect(true);
+        Session session = InspireFace.CreateSession(parameter, InspireFace.DETECT_MODE_ALWAYS_DETECT, 10, -1, -1);
+        Log.i(TAG, "session handle: " + session.handle);
+        InspireFace.SetTrackPreviewSize(session, 320);
+        InspireFace.SetFaceDetectThreshold(session, 0.5f);
+        InspireFace.SetFilterMinimumFacePixelSize(session, 0);
+
+        Bitmap img = SDKUtils.getImageFromAssetsFile(this, "inspireface/kun.jpg");
+        ImageStream stream = InspireFace.CreateImageStreamFromBitmap(img, InspireFace.CAMERA_ROTATION_0);
+        Log.i(TAG, "stream handle: " + stream.handle);
+        InspireFace.WriteImageStreamToFile(stream, "/storage/emulated/0/Android/data/com.example.inspireface_example/files/out.jpg");
+
+        MultipleFaceData multipleFaceData = InspireFace.ExecuteFaceTrack(session, stream);
+        Log.i(TAG, "Face num: " + multipleFaceData.detectedNum);
+
+        if (multipleFaceData.detectedNum > 0) {
+            Point2f[] lmk = InspireFace.GetFaceDenseLandmarkFromFaceToken(multipleFaceData.tokens[0]);
+            for (Point2f p : lmk) {
+                Log.i(TAG, p.x + ", " + p.y);
+            }
+            FaceFeature feature = InspireFace.ExtractFaceFeature(session, stream, multipleFaceData.tokens[0]);
+            Log.i(TAG, "Feature size: " + feature.data.length);
+            String strFt = "";
+            for (int i = 0; i < feature.data.length; i++) {
+                strFt = strFt + feature.data[i] + ", ";
+            }
+            Log.i(TAG, strFt);
+
+            for (int i = 0; i < 10; i++) {
+                FaceFeatureIdentity identity = FaceFeatureIdentity.create(-1, feature);
+                boolean succ = InspireFace.FeatureHubInsertFeature(identity);
+                if (succ) {
+                    Log.i(TAG, "Allocation ID: " + identity.id);
+                }
+            }
+
+            FaceFeatureIdentity searched = InspireFace.FeatureHubFaceSearch(feature);
+            Log.i(TAG, "Searched id: " + searched.id + ", Confidence: " + searched.searchConfidence);
+
+            SearchTopKResults topKResults = InspireFace.FeatureHubFaceSearchTopK(feature, 10);
+            for (int i = 0; i < topKResults.num; i++) {
+                Log.i(TAG, "TopK id: " + topKResults.ids[i] + ", Confidence: " + topKResults.confidence[i]);
+            }
+
+            FaceFeature newFeature = new FaceFeature();
+            Log.i(TAG, "Feature length: " + InspireFace.GetFeatureLength());
+            newFeature.data = new float[InspireFace.GetFeatureLength()];
+            FaceFeatureIdentity identity = FaceFeatureIdentity.create(8, newFeature);
+            boolean updateSucc = InspireFace.FeatureHubFaceUpdate(identity);
+            if (updateSucc) {
+                Log.i(TAG, "Update feature success: " + 8);
+            }
+            boolean removeSucc = InspireFace.FeatureHubFaceRemove(4);
+            if (removeSucc) {
+                Log.i(TAG, "Remove feature success: " + 4);
+            }
+            SearchTopKResults topkAgn = InspireFace.FeatureHubFaceSearchTopK(feature, 10);
+            for (int i = 0; i < topkAgn.num; i++) {
+                Log.i(TAG, "Agn TopK id: " + topkAgn.ids[i] + ", Confidence: " + topKResults.confidence[i]);
+            }
+
+            FaceFeatureIdentity queryIdentity = InspireFace.FeatureHubGetFaceIdentity(4);
+            if (queryIdentity != null) {
+                Log.e(TAG, "query id: " + queryIdentity.id);
+            }
+            queryIdentity = InspireFace.FeatureHubGetFaceIdentity(2);
+            if (queryIdentity != null) {
+                strFt = "";
+                for (int i = 0; i < queryIdentity.feature.data.length; i++) {
+                    strFt = strFt + queryIdentity.feature.data[i] + ", ";
+                }
+                Log.i(TAG, "query id: " + queryIdentity.id);
+                Log.i(TAG, strFt);
+
+                float comp = InspireFace.FaceComparison(queryIdentity.feature, feature);
+                Log.i(TAG, "Comparison: " + comp);
+            }
+            CustomParameter pipelineNeedParam = InspireFace.CreateCustomParameter()
+                    .enableFaceQuality(true)
+                    .enableLiveness(true)
+                    .enableMaskDetect(true)
+                    .enableFaceAttribute(true)
+                    .enableInteractionLiveness(true);
+            boolean succPipe = InspireFace.MultipleFacePipelineProcess(session, stream, multipleFaceData, pipelineNeedParam);
+            if (succPipe) {
+                Log.i(TAG, "Exec pipeline success");
+                RGBLivenessConfidence rgbLivenessConfidence = InspireFace.GetRGBLivenessConfidence(session);
+                Log.i(TAG, "rgbLivenessConfidence: " + rgbLivenessConfidence.confidence[0]);
+                FaceQualityConfidence faceQualityConfidence = InspireFace.GetFaceQualityConfidence(session);
+                Log.i(TAG, "faceQualityConfidence: " + faceQualityConfidence.confidence[0]);
+                FaceMaskConfidence faceMaskConfidence = InspireFace.GetFaceMaskConfidence(session);
+                Log.i(TAG, "faceMaskConfidence: " + faceMaskConfidence.confidence[0]);
+                FaceInteractionState faceInteractionState = InspireFace.GetFaceInteractionStateResult(session);
+                Log.i(TAG, "Left eye status confidence: " + faceInteractionState.leftEyeStatusConfidence[0]);
+                Log.i(TAG, "Right eye status confidence: " + faceInteractionState.rightEyeStatusConfidence[0]);
+                FaceInteractionsActions faceInteractionsActions = InspireFace.GetFaceInteractionActionsResult(session);
+                Log.i(TAG, "Normal: " + faceInteractionsActions.normal[0]);
+                Log.i(TAG, "Shake: " + faceInteractionsActions.shake[0]);
+                Log.i(TAG, "Jaw open: " + faceInteractionsActions.jawOpen[0]);
+                Log.i(TAG, "Head raise: " + faceInteractionsActions.headRaise[0]);
+                Log.i(TAG, "Blink: " + faceInteractionsActions.blink[0]);
+                FaceAttributeResult faceAttributeResult = InspireFace.GetFaceAttributeResult(session);
+                Log.i(TAG, "Race: " + faceAttributeResult.race[0]);
+                Log.i(TAG, "Gender: " + faceAttributeResult.gender[0]);
+                Log.i(TAG, "Age bracket: " + faceAttributeResult.ageBracket[0]);
+            } else {
+
+                Log.e(TAG, "Exec pipeline fail");
+            }
+        }
+
+        int count = InspireFace.FeatureHubGetFaceCount();
+        Log.i(TAG, "Face count: " + count);
+
+        Bitmap crop = InspireFace.GetFaceAlignmentImage(session, stream, multipleFaceData.tokens[0]);
+        try {
+            SDKUtils.saveBitmap("/storage/emulated/0/Android/data/com.example.inspireface_example/files/", "crop", crop);
+        } catch (IOException e) {
+            throw new RuntimeException(e);
+        }
+
+        InspireFace.ReleaseImageStream(stream);
+        InspireFace.ReleaseSession(session);
+
+
+        InspireFace.FeatureHubDataDisable();
+    }
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        EdgeToEdge.enable(this);
+        setContentView(R.layout.activity_main);
+        ViewCompat.setOnApplyWindowInsetsListener(findViewById(R.id.main), (v, insets) -> {
+            Insets systemBars = insets.getInsets(WindowInsetsCompat.Type.systemBars());
+            v.setPadding(systemBars.left, systemBars.top, systemBars.right, systemBars.bottom);
+            return insets;
+        });
+
+        //
+
+        test();
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_background.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_background.xml
new file mode 100644
index 000000000..07d5da9cb
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_background.xml
@@ -0,0 +1,170 @@
+<?xml version="1.0" encoding="utf-8"?>
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    android:width="108dp"
+    android:height="108dp"
+    android:viewportWidth="108"
+    android:viewportHeight="108">
+    <path
+        android:fillColor="#3DDC84"
+        android:pathData="M0,0h108v108h-108z" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M9,0L9,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,0L19,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M29,0L29,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M39,0L39,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M49,0L49,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M59,0L59,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M69,0L69,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M79,0L79,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M89,0L89,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M99,0L99,108"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,9L108,9"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,19L108,19"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,29L108,29"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,39L108,39"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,49L108,49"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,59L108,59"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,69L108,69"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,79L108,79"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,89L108,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M0,99L108,99"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,29L89,29"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,39L89,39"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,49L89,49"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,59L89,59"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,69L89,69"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M19,79L89,79"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M29,19L29,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M39,19L39,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M49,19L49,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M59,19L59,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M69,19L69,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+    <path
+        android:fillColor="#00000000"
+        android:pathData="M79,19L79,89"
+        android:strokeWidth="0.8"
+        android:strokeColor="#33FFFFFF" />
+</vector>
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_foreground.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_foreground.xml
new file mode 100644
index 000000000..2b068d114
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/drawable/ic_launcher_foreground.xml
@@ -0,0 +1,30 @@
+<vector xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:aapt="http://schemas.android.com/aapt"
+    android:width="108dp"
+    android:height="108dp"
+    android:viewportWidth="108"
+    android:viewportHeight="108">
+    <path android:pathData="M31,63.928c0,0 6.4,-11 12.1,-13.1c7.2,-2.6 26,-1.4 26,-1.4l38.1,38.1L107,108.928l-32,-1L31,63.928z">
+        <aapt:attr name="android:fillColor">
+            <gradient
+                android:endX="85.84757"
+                android:endY="92.4963"
+                android:startX="42.9492"
+                android:startY="49.59793"
+                android:type="linear">
+                <item
+                    android:color="#44000000"
+                    android:offset="0.0" />
+                <item
+                    android:color="#00000000"
+                    android:offset="1.0" />
+            </gradient>
+        </aapt:attr>
+    </path>
+    <path
+        android:fillColor="#FFFFFF"
+        android:fillType="nonZero"
+        android:pathData="M65.3,45.828l3.8,-6.6c0.2,-0.4 0.1,-0.9 -0.3,-1.1c-0.4,-0.2 -0.9,-0.1 -1.1,0.3l-3.9,6.7c-6.3,-2.8 -13.4,-2.8 -19.7,0l-3.9,-6.7c-0.2,-0.4 -0.7,-0.5 -1.1,-0.3C38.8,38.328 38.7,38.828 38.9,39.228l3.8,6.6C36.2,49.428 31.7,56.028 31,63.928h46C76.3,56.028 71.8,49.428 65.3,45.828zM43.4,57.328c-0.8,0 -1.5,-0.5 -1.8,-1.2c-0.3,-0.7 -0.1,-1.5 0.4,-2.1c0.5,-0.5 1.4,-0.7 2.1,-0.4c0.7,0.3 1.2,1 1.2,1.8C45.3,56.528 44.5,57.328 43.4,57.328L43.4,57.328zM64.6,57.328c-0.8,0 -1.5,-0.5 -1.8,-1.2s-0.1,-1.5 0.4,-2.1c0.5,-0.5 1.4,-0.7 2.1,-0.4c0.7,0.3 1.2,1 1.2,1.8C66.5,56.528 65.6,57.328 64.6,57.328L64.6,57.328z"
+        android:strokeWidth="1"
+        android:strokeColor="#00000000" />
+</vector>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/layout/activity_main.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/layout/activity_main.xml
new file mode 100644
index 000000000..86a5d9779
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/layout/activity_main.xml
@@ -0,0 +1,19 @@
+<?xml version="1.0" encoding="utf-8"?>
+<androidx.constraintlayout.widget.ConstraintLayout xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:app="http://schemas.android.com/apk/res-auto"
+    xmlns:tools="http://schemas.android.com/tools"
+    android:id="@+id/main"
+    android:layout_width="match_parent"
+    android:layout_height="match_parent"
+    tools:context=".MainActivity">
+
+    <TextView
+        android:layout_width="wrap_content"
+        android:layout_height="wrap_content"
+        android:text="Hello World!"
+        app:layout_constraintBottom_toBottomOf="parent"
+        app:layout_constraintEnd_toEndOf="parent"
+        app:layout_constraintStart_toStartOf="parent"
+        app:layout_constraintTop_toTopOf="parent" />
+
+</androidx.constraintlayout.widget.ConstraintLayout>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml
new file mode 100644
index 000000000..6f3b755bf
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+    <monochrome android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml
new file mode 100644
index 000000000..6f3b755bf
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+    <monochrome android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher.webp
new file mode 100644
index 000000000..c209e78ec
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..b2dfe3d1b
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-hdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher.webp
new file mode 100644
index 000000000..4f0f1d64e
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..62b611da0
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-mdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher.webp
new file mode 100644
index 000000000..948a3070f
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..1b9a6956b
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp
new file mode 100644
index 000000000..28d4b77f9
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..9287f5083
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp
new file mode 100644
index 000000000..aa7d6427e
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp
new file mode 100644
index 000000000..9126ae37c
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/mipmap-xxxhdpi/ic_launcher_round.webp differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values-night/themes.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values-night/themes.xml
new file mode 100644
index 000000000..9274570a8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values-night/themes.xml
@@ -0,0 +1,7 @@
+<resources xmlns:tools="http://schemas.android.com/tools">
+    <!-- Base application theme. -->
+    <style name="Base.Theme.InspireFaceExample" parent="Theme.Material3.DayNight.NoActionBar">
+        <!-- Customize your dark theme here. -->
+        <!-- <item name="colorPrimary">@color/my_dark_primary</item> -->
+    </style>
+</resources>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/colors.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/colors.xml
new file mode 100644
index 000000000..c8524cd96
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/colors.xml
@@ -0,0 +1,5 @@
+<?xml version="1.0" encoding="utf-8"?>
+<resources>
+    <color name="black">#FF000000</color>
+    <color name="white">#FFFFFFFF</color>
+</resources>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/strings.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/strings.xml
new file mode 100644
index 000000000..b60058c75
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/strings.xml
@@ -0,0 +1,3 @@
+<resources>
+    <string name="app_name">InspireFace-Example</string>
+</resources>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/themes.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/themes.xml
new file mode 100644
index 000000000..2583ad73e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/values/themes.xml
@@ -0,0 +1,9 @@
+<resources xmlns:tools="http://schemas.android.com/tools">
+    <!-- Base application theme. -->
+    <style name="Base.Theme.InspireFaceExample" parent="Theme.Material3.DayNight.NoActionBar">
+        <!-- Customize your light theme here. -->
+        <!-- <item name="colorPrimary">@color/my_light_primary</item> -->
+    </style>
+
+    <style name="Theme.InspireFaceExample" parent="Base.Theme.InspireFaceExample" />
+</resources>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/backup_rules.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/backup_rules.xml
new file mode 100644
index 000000000..fa0f996d2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/backup_rules.xml
@@ -0,0 +1,13 @@
+<?xml version="1.0" encoding="utf-8"?><!--
+   Sample backup rules file; uncomment and customize as necessary.
+   See https://developer.android.com/guide/topics/data/autobackup
+   for details.
+   Note: This file is ignored for devices older that API 31
+   See https://developer.android.com/about/versions/12/backup-restore
+-->
+<full-backup-content>
+    <!--
+   <include domain="sharedpref" path="."/>
+   <exclude domain="sharedpref" path="device.xml"/>
+-->
+</full-backup-content>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/data_extraction_rules.xml b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/data_extraction_rules.xml
new file mode 100644
index 000000000..9ee9997b0
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/main/res/xml/data_extraction_rules.xml
@@ -0,0 +1,19 @@
+<?xml version="1.0" encoding="utf-8"?><!--
+   Sample data extraction rules file; uncomment and customize as necessary.
+   See https://developer.android.com/about/versions/12/backup-restore#xml-changes
+   for details.
+-->
+<data-extraction-rules>
+    <cloud-backup>
+        <!-- TODO: Use <include> and <exclude> to control what is backed up.
+        <include .../>
+        <exclude .../>
+        -->
+    </cloud-backup>
+    <!--
+    <device-transfer>
+        <include .../>
+        <exclude .../>
+    </device-transfer>
+    -->
+</data-extraction-rules>
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java
new file mode 100644
index 000000000..67bb0493e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/app/src/test/java/com/example/inspireface_example/ExampleUnitTest.java
@@ -0,0 +1,17 @@
+package com.example.inspireface_example;
+
+import org.junit.Test;
+
+import static org.junit.Assert.*;
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * @see <a href="http://d.android.com/tools/testing">Testing documentation</a>
+ */
+public class ExampleUnitTest {
+    @Test
+    public void addition_isCorrect() {
+        assertEquals(4, 2 + 2);
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/build.gradle b/cpp-package/inspireface/android/InspireFaceExample_old/build.gradle
new file mode 100644
index 000000000..99e73ad4c
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/build.gradle
@@ -0,0 +1,5 @@
+// Top-level build file where you can add configuration options common to all sub-projects/modules.
+plugins {
+alias(libs.plugins.android.application) apply false
+    alias(libs.plugins.android.library) apply false
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradle.properties b/cpp-package/inspireface/android/InspireFaceExample_old/gradle.properties
new file mode 100644
index 000000000..4387edc22
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/gradle.properties
@@ -0,0 +1,21 @@
+# Project-wide Gradle settings.
+# IDE (e.g. Android Studio) users:
+# Gradle settings configured through the IDE *will override*
+# any settings specified in this file.
+# For more details on how to configure your build environment visit
+# http://www.gradle.org/docs/current/userguide/build_environment.html
+# Specifies the JVM arguments used for the daemon process.
+# The setting is particularly useful for tweaking memory settings.
+org.gradle.jvmargs=-Xmx2048m -Dfile.encoding=UTF-8
+# When configured, Gradle will run in incubating parallel mode.
+# This option should only be used with decoupled projects. For more details, visit
+# https://developer.android.com/r/tools/gradle-multi-project-decoupled-projects
+# org.gradle.parallel=true
+# AndroidX package structure to make it clearer which packages are bundled with the
+# Android operating system, and which are packaged with your app's APK
+# https://developer.android.com/topic/libraries/support-library/androidx-rn
+android.useAndroidX=true
+# Enables namespacing of each library's R class so that its R class includes only the
+# resources declared in the library itself and none from the library's dependencies,
+# thereby reducing the size of the R class for that library
+android.nonTransitiveRClass=true
\ No newline at end of file
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradle/libs.versions.toml b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/libs.versions.toml
new file mode 100644
index 000000000..ba98484f2
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/libs.versions.toml
@@ -0,0 +1,25 @@
+[versions]
+agp = "8.5.1"
+inspirefaceAndroidSdk = "1.2.0"
+junit = "4.13.2"
+junitVersion = "1.1.5"
+espressoCore = "3.5.1"
+appcompat = "1.6.1"
+material = "1.10.0"
+activity = "1.8.0"
+constraintlayout = "2.1.4"
+
+[libraries]
+inspireface-android-sdk = { module = "com.github.HyperInspire:inspireface-android-sdk", version.ref = "inspirefaceAndroidSdk" }
+junit = { group = "junit", name = "junit", version.ref = "junit" }
+ext-junit = { group = "androidx.test.ext", name = "junit", version.ref = "junitVersion" }
+espresso-core = { group = "androidx.test.espresso", name = "espresso-core", version.ref = "espressoCore" }
+appcompat = { group = "androidx.appcompat", name = "appcompat", version.ref = "appcompat" }
+material = { group = "com.google.android.material", name = "material", version.ref = "material" }
+activity = { group = "androidx.activity", name = "activity", version.ref = "activity" }
+constraintlayout = { group = "androidx.constraintlayout", name = "constraintlayout", version.ref = "constraintlayout" }
+
+[plugins]
+android-application = { id = "com.android.application", version.ref = "agp" }
+android-library = { id = "com.android.library", version.ref = "agp" }
+
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.jar b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.jar
new file mode 100644
index 000000000..e708b1c02
Binary files /dev/null and b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.jar differ
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.properties b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.properties
new file mode 100644
index 000000000..e846f349e
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/gradle/wrapper/gradle-wrapper.properties
@@ -0,0 +1,6 @@
+#Tue Nov 26 15:12:19 CST 2024
+distributionBase=GRADLE_USER_HOME
+distributionPath=wrapper/dists
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.7-bin.zip
+zipStoreBase=GRADLE_USER_HOME
+zipStorePath=wrapper/dists
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradlew b/cpp-package/inspireface/android/InspireFaceExample_old/gradlew
new file mode 100755
index 000000000..4f906e0c8
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/gradlew
@@ -0,0 +1,185 @@
+#!/usr/bin/env sh
+
+#
+# Copyright 2015 the original author or authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+##############################################################################
+##
+##  Gradle start up script for UN*X
+##
+##############################################################################
+
+# Attempt to set APP_HOME
+# Resolve links: $0 may be a link
+PRG="$0"
+# Need this for relative symlinks.
+while [ -h "$PRG" ] ; do
+    ls=`ls -ld "$PRG"`
+    link=`expr "$ls" : '.*-> \(.*\)$'`
+    if expr "$link" : '/.*' > /dev/null; then
+        PRG="$link"
+    else
+        PRG=`dirname "$PRG"`"/$link"
+    fi
+done
+SAVED="`pwd`"
+cd "`dirname \"$PRG\"`/" >/dev/null
+APP_HOME="`pwd -P`"
+cd "$SAVED" >/dev/null
+
+APP_NAME="Gradle"
+APP_BASE_NAME=`basename "$0"`
+
+# Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+DEFAULT_JVM_OPTS='"-Xmx64m" "-Xms64m"'
+
+# Use the maximum available, or set MAX_FD != -1 to use that value.
+MAX_FD="maximum"
+
+warn () {
+    echo "$*"
+}
+
+die () {
+    echo
+    echo "$*"
+    echo
+    exit 1
+}
+
+# OS specific support (must be 'true' or 'false').
+cygwin=false
+msys=false
+darwin=false
+nonstop=false
+case "`uname`" in
+  CYGWIN* )
+    cygwin=true
+    ;;
+  Darwin* )
+    darwin=true
+    ;;
+  MINGW* )
+    msys=true
+    ;;
+  NONSTOP* )
+    nonstop=true
+    ;;
+esac
+
+CLASSPATH=$APP_HOME/gradle/wrapper/gradle-wrapper.jar
+
+
+# Determine the Java command to use to start the JVM.
+if [ -n "$JAVA_HOME" ] ; then
+    if [ -x "$JAVA_HOME/jre/sh/java" ] ; then
+        # IBM's JDK on AIX uses strange locations for the executables
+        JAVACMD="$JAVA_HOME/jre/sh/java"
+    else
+        JAVACMD="$JAVA_HOME/bin/java"
+    fi
+    if [ ! -x "$JAVACMD" ] ; then
+        die "ERROR: JAVA_HOME is set to an invalid directory: $JAVA_HOME
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+    fi
+else
+    JAVACMD="java"
+    which java >/dev/null 2>&1 || die "ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+fi
+
+# Increase the maximum file descriptors if we can.
+if [ "$cygwin" = "false" -a "$darwin" = "false" -a "$nonstop" = "false" ] ; then
+    MAX_FD_LIMIT=`ulimit -H -n`
+    if [ $? -eq 0 ] ; then
+        if [ "$MAX_FD" = "maximum" -o "$MAX_FD" = "max" ] ; then
+            MAX_FD="$MAX_FD_LIMIT"
+        fi
+        ulimit -n $MAX_FD
+        if [ $? -ne 0 ] ; then
+            warn "Could not set maximum file descriptor limit: $MAX_FD"
+        fi
+    else
+        warn "Could not query maximum file descriptor limit: $MAX_FD_LIMIT"
+    fi
+fi
+
+# For Darwin, add options to specify how the application appears in the dock
+if $darwin; then
+    GRADLE_OPTS="$GRADLE_OPTS \"-Xdock:name=$APP_NAME\" \"-Xdock:icon=$APP_HOME/media/gradle.icns\""
+fi
+
+# For Cygwin or MSYS, switch paths to Windows format before running java
+if [ "$cygwin" = "true" -o "$msys" = "true" ] ; then
+    APP_HOME=`cygpath --path --mixed "$APP_HOME"`
+    CLASSPATH=`cygpath --path --mixed "$CLASSPATH"`
+
+    JAVACMD=`cygpath --unix "$JAVACMD"`
+
+    # We build the pattern for arguments to be converted via cygpath
+    ROOTDIRSRAW=`find -L / -maxdepth 1 -mindepth 1 -type d 2>/dev/null`
+    SEP=""
+    for dir in $ROOTDIRSRAW ; do
+        ROOTDIRS="$ROOTDIRS$SEP$dir"
+        SEP="|"
+    done
+    OURCYGPATTERN="(^($ROOTDIRS))"
+    # Add a user-defined pattern to the cygpath arguments
+    if [ "$GRADLE_CYGPATTERN" != "" ] ; then
+        OURCYGPATTERN="$OURCYGPATTERN|($GRADLE_CYGPATTERN)"
+    fi
+    # Now convert the arguments - kludge to limit ourselves to /bin/sh
+    i=0
+    for arg in "$@" ; do
+        CHECK=`echo "$arg"|egrep -c "$OURCYGPATTERN" -`
+        CHECK2=`echo "$arg"|egrep -c "^-"`                                 ### Determine if an option
+
+        if [ $CHECK -ne 0 ] && [ $CHECK2 -eq 0 ] ; then                    ### Added a condition
+            eval `echo args$i`=`cygpath --path --ignore --mixed "$arg"`
+        else
+            eval `echo args$i`="\"$arg\""
+        fi
+        i=`expr $i + 1`
+    done
+    case $i in
+        0) set -- ;;
+        1) set -- "$args0" ;;
+        2) set -- "$args0" "$args1" ;;
+        3) set -- "$args0" "$args1" "$args2" ;;
+        4) set -- "$args0" "$args1" "$args2" "$args3" ;;
+        5) set -- "$args0" "$args1" "$args2" "$args3" "$args4" ;;
+        6) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" ;;
+        7) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" ;;
+        8) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" "$args7" ;;
+        9) set -- "$args0" "$args1" "$args2" "$args3" "$args4" "$args5" "$args6" "$args7" "$args8" ;;
+    esac
+fi
+
+# Escape application args
+save () {
+    for i do printf %s\\n "$i" | sed "s/'/'\\\\''/g;1s/^/'/;\$s/\$/' \\\\/" ; done
+    echo " "
+}
+APP_ARGS=`save "$@"`
+
+# Collect all arguments for the java command, following the shell quoting and substitution rules
+eval set -- $DEFAULT_JVM_OPTS $JAVA_OPTS $GRADLE_OPTS "\"-Dorg.gradle.appname=$APP_BASE_NAME\"" -classpath "\"$CLASSPATH\"" org.gradle.wrapper.GradleWrapperMain "$APP_ARGS"
+
+exec "$JAVACMD" "$@"
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/gradlew.bat b/cpp-package/inspireface/android/InspireFaceExample_old/gradlew.bat
new file mode 100644
index 000000000..ac1b06f93
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/gradlew.bat
@@ -0,0 +1,89 @@
+@rem
+@rem Copyright 2015 the original author or authors.
+@rem
+@rem Licensed under the Apache License, Version 2.0 (the "License");
+@rem you may not use this file except in compliance with the License.
+@rem You may obtain a copy of the License at
+@rem
+@rem      https://www.apache.org/licenses/LICENSE-2.0
+@rem
+@rem Unless required by applicable law or agreed to in writing, software
+@rem distributed under the License is distributed on an "AS IS" BASIS,
+@rem WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+@rem See the License for the specific language governing permissions and
+@rem limitations under the License.
+@rem
+
+@if "%DEBUG%" == "" @echo off
+@rem ##########################################################################
+@rem
+@rem  Gradle startup script for Windows
+@rem
+@rem ##########################################################################
+
+@rem Set local scope for the variables with windows NT shell
+if "%OS%"=="Windows_NT" setlocal
+
+set DIRNAME=%~dp0
+if "%DIRNAME%" == "" set DIRNAME=.
+set APP_BASE_NAME=%~n0
+set APP_HOME=%DIRNAME%
+
+@rem Resolve any "." and ".." in APP_HOME to make it shorter.
+for %%i in ("%APP_HOME%") do set APP_HOME=%%~fi
+
+@rem Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+set DEFAULT_JVM_OPTS="-Xmx64m" "-Xms64m"
+
+@rem Find java.exe
+if defined JAVA_HOME goto findJavaFromJavaHome
+
+set JAVA_EXE=java.exe
+%JAVA_EXE% -version >NUL 2>&1
+if "%ERRORLEVEL%" == "0" goto execute
+
+echo.
+echo ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:findJavaFromJavaHome
+set JAVA_HOME=%JAVA_HOME:"=%
+set JAVA_EXE=%JAVA_HOME%/bin/java.exe
+
+if exist "%JAVA_EXE%" goto execute
+
+echo.
+echo ERROR: JAVA_HOME is set to an invalid directory: %JAVA_HOME%
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:execute
+@rem Setup the command line
+
+set CLASSPATH=%APP_HOME%\gradle\wrapper\gradle-wrapper.jar
+
+
+@rem Execute Gradle
+"%JAVA_EXE%" %DEFAULT_JVM_OPTS% %JAVA_OPTS% %GRADLE_OPTS% "-Dorg.gradle.appname=%APP_BASE_NAME%" -classpath "%CLASSPATH%" org.gradle.wrapper.GradleWrapperMain %*
+
+:end
+@rem End local scope for the variables with windows NT shell
+if "%ERRORLEVEL%"=="0" goto mainEnd
+
+:fail
+rem Set variable GRADLE_EXIT_CONSOLE if you need the _script_ return code instead of
+rem the _cmd.exe /c_ return code!
+if  not "" == "%GRADLE_EXIT_CONSOLE%" exit 1
+exit /b 1
+
+:mainEnd
+if "%OS%"=="Windows_NT" endlocal
+
+:omega
diff --git a/cpp-package/inspireface/android/InspireFaceExample_old/settings.gradle b/cpp-package/inspireface/android/InspireFaceExample_old/settings.gradle
new file mode 100644
index 000000000..5e4119ae9
--- /dev/null
+++ b/cpp-package/inspireface/android/InspireFaceExample_old/settings.gradle
@@ -0,0 +1,24 @@
+pluginManagement {
+    repositories {
+        google {
+            content {
+                includeGroupByRegex("com\\.android.*")
+                includeGroupByRegex("com\\.google.*")
+                includeGroupByRegex("androidx.*")
+            }
+        }
+        mavenCentral()
+        gradlePluginPortal()
+    }
+}
+dependencyResolutionManagement {
+    repositoriesMode.set(RepositoriesMode.FAIL_ON_PROJECT_REPOS)
+    repositories {
+        google()
+        mavenCentral()
+        maven { url 'https://jitpack.io' }
+    }
+}
+
+rootProject.name = "InspireFace-Example"
+include ':app'
diff --git a/cpp-package/inspireface/ci/README.md b/cpp-package/inspireface/ci/README.md
new file mode 100644
index 000000000..7d9d10d00
--- /dev/null
+++ b/cpp-package/inspireface/ci/README.md
@@ -0,0 +1,4 @@
+# CI/CD
+
+- Linux-x86
+- ...
\ No newline at end of file
diff --git a/cpp-package/inspireface/ci/quick_test_linux_x86_usual.sh b/cpp-package/inspireface/ci/quick_test_linux_x86_usual.sh
new file mode 100644
index 000000000..d82e90807
--- /dev/null
+++ b/cpp-package/inspireface/ci/quick_test_linux_x86_usual.sh
@@ -0,0 +1,59 @@
+#!/bin/bash
+
+# Exit immediately if any command exits with a non-zero status
+set -e
+
+TARGET_DIR="test_res"
+BUILD_DIRNAME="ci_ubuntu18"
+TEST_DIR="./build/${BUILD_DIRNAME}/test"
+TEST_EXECUTABLE="./test/Test"
+
+# Make dir
+mkdir -p ${TARGET_DIR}/save/video_frames
+
+# Download models
+bash command/download_models_general.sh Pikachu
+
+# Get the absolute path of the target directory
+FULL_TEST_DIR="$(realpath ${TARGET_DIR})"
+
+# Create the build directory if it doesn't exist
+mkdir -p build/${BUILD_DIRNAME}/
+
+# Change directory to the build directory
+# Disable the shellcheck warning for potential directory changes
+# shellcheck disable=SC2164
+cd build/${BUILD_DIRNAME}/
+
+# Configure the CMake build system
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=ON \
+  -DISF_ENABLE_BENCHMARK=ON \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=OFF ../../
+
+# Compile the project using 4 parallel jobs
+make -j4
+
+# Check if the symbolic link or directory already exists
+if [ ! -e "$(basename ${FULL_TEST_DIR})" ]; then
+    # Create a symbolic link to the extracted test data directory
+    ln -s ${FULL_TEST_DIR} .
+    echo "Symbolic link to '${TARGET_DIR}' created."
+else
+    echo "Symbolic link or directory '$(basename ${FULL_TEST_DIR})' already exists. Skipping creation."
+fi
+
+# Check if the test executable file exists
+if [ ! -f "$TEST_EXECUTABLE" ]; then
+    # If not, print an error message and exit with a non-zero status code
+    echo "Error: Test executable '$TEST_EXECUTABLE' not found. Please ensure it is built correctly."
+    exit 1
+else
+    # If it exists, print a message and run the test executable
+    echo "Test executable found. Running tests..."
+    "$TEST_EXECUTABLE"
+fi
diff --git a/cpp-package/inspireface/ci/quick_test_linux_x86_usual_python_native_interface.sh b/cpp-package/inspireface/ci/quick_test_linux_x86_usual_python_native_interface.sh
new file mode 100644
index 000000000..7418a03ec
--- /dev/null
+++ b/cpp-package/inspireface/ci/quick_test_linux_x86_usual_python_native_interface.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+
+# Exit immediately if any command exits with a non-zero status
+set -e
+
+ROOT_DIR="$(pwd)"
+BUILD_DIRNAME="ubuntu18_shared"
+
+# Create the build directory if it doesn't exist
+mkdir -p build/${BUILD_DIRNAME}/
+
+# Change directory to the build directory
+# Disable the shellcheck warning for potential directory changes
+# shellcheck disable=SC2164
+cd build/${BUILD_DIRNAME}/
+
+# Configure the CMake build system
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON ../../
+
+# Compile the project using 4 parallel jobs
+make -j4
+
+# Come back to project root dir
+cd ${ROOT_DIR}
+
+# Important: You must copy the compiled dynamic library to this path!
+mkdir -p python/inspireface/modules/core/libs/linux/x64/
+cp build/${BUILD_DIRNAME}/lib/libInspireFace.so python/inspireface/modules/core/libs/linux/x64/
+
+# Install dependency
+pip install opencv-python
+pip install click
+pip install loguru
+pip install filelock
+pip install modelscope
+
+cd python/
+
+# Run sample
+python sample_face_detection.py ../test_res/data/bulk/woman.png
+
diff --git a/cpp-package/inspireface/ci/quick_test_local.sh b/cpp-package/inspireface/ci/quick_test_local.sh
new file mode 100644
index 000000000..8b975dfbb
--- /dev/null
+++ b/cpp-package/inspireface/ci/quick_test_local.sh
@@ -0,0 +1,71 @@
+#!/bin/bash
+
+# Exit immediately if any command exits with a non-zero status
+set -e
+
+TARGET_DIR="test_res"
+DOWNLOAD_URL="https://github.com/tunmx/inspireface-store/raw/main/resource/test_res-lite2.zip"
+ZIP_FILE="test_res-lite2.zip"
+BUILD_DIRNAME="quick_test_build"
+TEST_DIR="./build/${BUILD_DIRNAME}/test"
+TEST_EXECUTABLE="./test/Test"
+
+# Check if the target directory already exists
+if [ ! -d "$TARGET_DIR" ]; then
+    echo "Directory '$TARGET_DIR' does not exist. Downloading..."
+
+    # Download the dataset zip file
+    wget -q "$DOWNLOAD_URL" -O "$ZIP_FILE"
+
+    echo "Extracting '$ZIP_FILE' to '$TARGET_DIR'..."
+    # Unzip the downloaded file
+    unzip "$ZIP_FILE"
+
+    # Remove the downloaded zip file and unnecessary folders
+    rm "$ZIP_FILE"
+    rm -rf "__MACOSX"
+
+    echo "Download and extraction complete."
+else
+    echo "Directory '$TARGET_DIR' already exists. Skipping download."
+fi
+
+# Get the absolute path of the target directory
+FULL_TEST_DIR="$(realpath ${TARGET_DIR})"
+
+# Create the build directory if it doesn't exist
+mkdir -p build/${BUILD_DIRNAME}/
+
+# Change directory to the build directory
+# Disable the shellcheck warning for potential directory changes
+# shellcheck disable=SC2164
+cd build/${BUILD_DIRNAME}/
+
+# Configure the CMake build system
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=ON \
+  -DISF_ENABLE_BENCHMARK=ON \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=OFF ../../
+
+# Compile the project using 4 parallel jobs
+make -j4
+
+# Create a symbolic link to the extracted test data directory
+ln -s ${FULL_TEST_DIR} .
+
+# Check if the test executable file exists
+if [ ! -f "$TEST_EXECUTABLE" ]; then
+    # If not, print an error message and exit with a non-zero status code
+    echo "Error: Test executable '$TEST_EXECUTABLE' not found. Please ensure it is built correctly."
+    exit 1
+else
+    # If it exists, print a message and run the test executable
+    echo "Test executable found. Running tests..."
+    "$TEST_EXECUTABLE"
+fi
+
+# Executing python scripts
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build.sh b/cpp-package/inspireface/command/build.sh
new file mode 100644
index 000000000..f55c72433
--- /dev/null
+++ b/cpp-package/inspireface/command/build.sh
@@ -0,0 +1,93 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+# Detect the operating system
+OS_NAME=$(uname)
+BUILD_DIR="build"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+
+
+# Determine the appropriate build directory based on the OS
+case "$OS_NAME" in
+    Darwin)
+        # macOS system
+        BUILD_DIR="${BUILD_DIR}/inspireface-macos${TAG}"
+        ;;
+    Linux)
+        # Linux system, further identify the distribution if necessary
+        if [ -f /etc/os-release ]; then
+            . /etc/os-release
+            case "$ID" in
+                ubuntu)
+                    BUILD_DIR="${BUILD_DIR}/inspireface-linux-ubuntu${TAG}"
+                    ;;
+                centos)
+                    BUILD_DIR="${BUILD_DIR}/inspireface-linux-centos${TAG}"
+                    ;;
+                *)
+                    # If an unknown Linux distribution, default to generic 'linux'
+                    BUILD_DIR="${BUILD_DIR}/inspireface-linux${TAG}"
+                    ;;
+            esac
+        else
+            # If unable to detect Linux distribution, default to 'linux'
+            BUILD_DIR="${BUILD_DIR}/inspireface-linux${TAG}"
+        fi
+        ;;
+    *)
+        # If OS is not recognized, default to 'generic'
+        BUILD_DIR="${BUILD_DIR}/inspireface-generic${TAG}"
+        ;;
+esac
+
+# Create the build directory and navigate into it
+mkdir -p "$BUILD_DIR"
+cd "$BUILD_DIR" || exit 1
+
+# Run CMake configuration (adjust the options as needed)
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=ON \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  "$SCRIPT_DIR"
+
+# Compile and install
+make -j4
+make install
+
+# Move 'install' files to the build root directory using an absolute path
+move_install_files "$(pwd)"
diff --git a/cpp-package/inspireface/command/build_android.sh b/cpp-package/inspireface/command/build_android.sh
new file mode 100644
index 000000000..3dbf50778
--- /dev/null
+++ b/cpp-package/inspireface/command/build_android.sh
@@ -0,0 +1,155 @@
+#!/bin/bash
+
+reorganize_structure() {
+    local base_path=$1
+
+    # Define the new main directories
+    local main_dirs=("lib" "sample" "test")
+
+    # Check if the base path exists
+    if [[ ! -d "$base_path" ]]; then
+        echo "Error: The path '$base_path' does not exist."
+        return 1
+    fi
+
+    # Create new main directories at the base path
+    for dir in "${main_dirs[@]}"; do
+        mkdir -p "$base_path/$dir"
+    done
+
+    # Find all architecture directories (e.g., arm64-v8a, armeabi-v7a)
+    local arch_dirs=()
+    for d in "$base_path"/*; do
+        [[ -d "$d" ]] || continue
+        name=$(basename "$d")
+        if [[ "$name" != "lib" && "$name" != "sample" && "$name" != "test" && "$name" != "." && "$name" != ".." ]]; then
+            arch_dirs+=("$d")
+        fi
+    done
+
+    for arch_dir in "${arch_dirs[@]}"; do
+        # Get the architecture name (e.g., arm64-v8a)
+        local arch=$(basename "$arch_dir")
+
+        # Operate on each main directory
+        for main_dir in "${main_dirs[@]}"; do
+            # Create a specific directory for each architecture under the main directory
+            mkdir -p "$base_path/$main_dir/$arch"
+
+            # Selectively copy content based on the directory type
+            case "$main_dir" in
+                lib)
+                    # Copy the lib directory
+                    if [ -d "$arch_dir/InspireFace/lib" ]; then
+                        cp -r "$arch_dir/InspireFace/lib/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+                sample)
+                    # Copy the sample directory
+                    if [ -d "$arch_dir/InspireFace/sample" ]; then
+                        cp -r "$arch_dir/InspireFace/sample/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+                test)
+                    # Copy the test directory
+                    if [ -d "$arch_dir/InspireFace/test" ]; then
+                        cp -r "$arch_dir/InspireFace/test/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+            esac
+        done
+
+        # Copy version.txt file to the base path, ignoring duplicates
+        if [ -f "$arch_dir/version.txt" ]; then
+            cp -f "$arch_dir/version.txt" "$base_path/version.txt"
+        fi
+    done
+
+        # Copy include once from a valid arch
+    for arch_dir in "${arch_dirs[@]}"; do
+        if [ -d "$arch_dir/InspireFace/include" ]; then
+            mkdir -p "$base_path/include"
+            cp -r "$arch_dir/InspireFace/include/"* "$base_path/include/"
+            echo "Copied include from $arch_dir"
+            break
+        fi
+    done
+
+    # Delete the original architecture directories
+    for arch_dir in "${arch_dirs[@]}"; do
+        rm -rf "$arch_dir"
+    done
+
+
+    echo "Reorganization complete."
+}
+
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+build() {
+    arch=$1
+    NDK_API_LEVEL=$2
+    mkdir -p ${BUILD_FOLDER_PATH}/${arch}
+    pushd ${BUILD_FOLDER_PATH}/${arch}
+    cmake ${SCRIPT_DIR} \
+        -G "Unix Makefiles" \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+        -DCMAKE_C_FLAGS="-g0 ${CMAKE_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="-g0 ${CMAKE_CXX_FLAGS}" \
+        -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+        -DANDROID_TOOLCHAIN=clang \
+        -DANDROID_ABI=${arch} \
+        -DANDROID_NATIVE_API_LEVEL=${NDK_API_LEVEL} \
+        -DANDROID_STL=c++_static \
+        -DMNN_BUILD_FOR_ANDROID_COMMAND=true \
+        -DISF_BUILD_WITH_SAMPLE=OFF \
+        -DISF_BUILD_WITH_TEST=OFF \
+        -DISF_ENABLE_BENCHMARK=OFF \
+        -DISF_ENABLE_USE_LFW_DATA=OFF \
+        -DISF_ENABLE_TEST_EVALUATION=OFF \
+        -DISF_BUILD_SHARED_LIBS=ON \
+        -Wno-dev
+    make -j4
+    make install
+    popd
+    move_install_files "${BUILD_FOLDER_PATH}/${arch}"
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  # Project dir
+BUILD_FOLDER_PATH="build/inspireface-android${TAG}"
+
+build arm64-v8a 21
+build armeabi-v7a 21
+build x86_64 21
+
+reorganize_structure "${BUILD_FOLDER_PATH}"
+
diff --git a/cpp-package/inspireface/command/build_android_rk356x_rk3588.sh b/cpp-package/inspireface/command/build_android_rk356x_rk3588.sh
new file mode 100644
index 000000000..98726e690
--- /dev/null
+++ b/cpp-package/inspireface/command/build_android_rk356x_rk3588.sh
@@ -0,0 +1,142 @@
+#!/bin/bash
+
+reorganize_structure() {
+    local base_path=$1
+
+    # Define the new main directories
+    local main_dirs=("lib" "sample" "test")
+
+    # Check if the base path exists
+    if [[ ! -d "$base_path" ]]; then
+        echo "Error: The path '$base_path' does not exist."
+        return 1
+    fi
+
+    # Create new main directories at the base path
+    for dir in "${main_dirs[@]}"; do
+        mkdir -p "$base_path/$dir"
+    done
+
+    # Find all architecture directories (e.g., arm64-v8a, armeabi-v7a)
+    local arch_dirs=($(find "$base_path" -maxdepth 1 -type d -name "arm*"))
+
+    for arch_dir in "${arch_dirs[@]}"; do
+        # Get the architecture name (e.g., arm64-v8a)
+        local arch=$(basename "$arch_dir")
+
+        # Operate on each main directory
+        for main_dir in "${main_dirs[@]}"; do
+            # Create a specific directory for each architecture under the main directory
+            mkdir -p "$base_path/$main_dir/$arch"
+
+            # Selectively copy content based on the directory type
+            case "$main_dir" in
+                lib)
+                    # Copy the lib directory
+                    if [ -d "$arch_dir/InspireFace/lib" ]; then
+                        cp -r "$arch_dir/InspireFace/lib/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+                sample)
+                    # Copy the sample directory
+                    if [ -d "$arch_dir/InspireFace/sample" ]; then
+                        cp -r "$arch_dir/InspireFace/sample/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+                test)
+                    # Copy the test directory
+                    if [ -d "$arch_dir/InspireFace/test" ]; then
+                        cp -r "$arch_dir/InspireFace/test/"* "$base_path/$main_dir/$arch/"
+                    fi
+                    ;;
+            esac
+        done
+
+        # Copy version.txt file to the base path, ignoring duplicates
+        if [ -f "$arch_dir/version.txt" ]; then
+            cp -f "$arch_dir/version.txt" "$base_path/version.txt"
+        fi
+    done
+
+    # Delete the original architecture directories
+    for arch_dir in "${arch_dirs[@]}"; do
+        rm -rf "$arch_dir"
+    done
+
+    echo "Reorganization complete."
+}
+
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+
+build() {
+    arch=$1
+    NDK_API_LEVEL=$2
+    mkdir -p ${BUILD_FOLDER_PATH}/${arch}
+    pushd ${BUILD_FOLDER_PATH}/${arch}
+    cmake ${SCRIPT_DIR} \
+        -G "Unix Makefiles" \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+        -DCMAKE_C_FLAGS="-g0 ${CMAKE_C_FLAGS}" \
+        -DCMAKE_CXX_FLAGS="-g0 ${CMAKE_CXX_FLAGS}" \
+        -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+        -DANDROID_TOOLCHAIN=clang \
+        -DANDROID_ABI=${arch} \
+        -DANDROID_NATIVE_API_LEVEL=${NDK_API_LEVEL} \
+        -DANDROID_STL=c++_static \
+        -DMNN_BUILD_FOR_ANDROID_COMMAND=true \
+        -DISF_ENABLE_RKNN=ON \
+        -DISF_ENABLE_RGA=OFF \
+        -DISF_RK_DEVICE_TYPE=RK356X \
+        -DISF_RKNPU_MAJOR=rknpu2 \
+        -DISF_RK_COMPILER_TYPE=${arch} \
+        -DISF_BUILD_WITH_SAMPLE=OFF \
+        -DISF_BUILD_WITH_TEST=OFF \
+        -DISF_ENABLE_BENCHMARK=OFF \
+        -DISF_ENABLE_USE_LFW_DATA=OFF \
+        -DISF_ENABLE_TEST_EVALUATION=OFF \
+        -DISF_BUILD_SHARED_LIBS=ON \
+        -Wno-dev
+    make -j4
+    make install
+    popd
+    move_install_files "${BUILD_FOLDER_PATH}/${arch}"
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  # Project dir
+BUILD_FOLDER_PATH="build/inspireface-android-rk356x-rk3588${TAG}"
+
+build arm64-v8a 24
+build armeabi-v7a 24
+
+reorganize_structure "${BUILD_FOLDER_PATH}"
+
diff --git a/cpp-package/inspireface/command/build_cross_aarch64.sh b/cpp-package/inspireface/command/build_cross_aarch64.sh
new file mode 100644
index 000000000..014e3e0e1
--- /dev/null
+++ b/cpp-package/inspireface/command/build_cross_aarch64.sh
@@ -0,0 +1,61 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  # Project dir
+BUILD_FOLDER_PATH="build/inspireface-linux-aarch64${TAG}"
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+# export cross_compile_toolchain=/home/jingyuyan/software/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DCMAKE_SYSTEM_VERSION=1 \
+  -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+  -DCMAKE_C_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-gcc \
+  -DCMAKE_CXX_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-g++ \
+  -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -flax-vector-conversions" \
+  -DTARGET_PLATFORM=armlinux \
+  -DISF_BUILD_LINUX_AARCH64=ON \
+  -DISF_BUILD_LINUX_ARM7=OFF \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
diff --git a/cpp-package/inspireface/command/build_cross_armv7_armhf.sh b/cpp-package/inspireface/command/build_cross_armv7_armhf.sh
new file mode 100644
index 000000000..c6d75ca41
--- /dev/null
+++ b/cpp-package/inspireface/command/build_cross_armv7_armhf.sh
@@ -0,0 +1,60 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  # Project dir
+BUILD_FOLDER_PATH="build/inspireface-linux-armv7-armhf${TAG}"
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+# export cross_compile_toolchain=/home/jingyuyan/software/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DCMAKE_SYSTEM_VERSION=1 \
+  -DCMAKE_SYSTEM_PROCESSOR=armv7 \
+  -DCMAKE_C_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-linux-gnueabihf-gcc \
+  -DCMAKE_CXX_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-linux-gnueabihf-g++ \
+  -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -flax-vector-conversions" \
+  -DTARGET_PLATFORM=armlinux \
+  -DISF_BUILD_LINUX_ARM7=ON \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
diff --git a/cpp-package/inspireface/command/build_cross_rk356x_rk3588_aarch64.sh b/cpp-package/inspireface/command/build_cross_rk356x_rk3588_aarch64.sh
new file mode 100644
index 000000000..aa01bb132
--- /dev/null
+++ b/cpp-package/inspireface/command/build_cross_rk356x_rk3588_aarch64.sh
@@ -0,0 +1,76 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+
+SCRIPT_DIR=$(pwd)  # Project dir
+
+
+cd ${SCRIPT_DIR}
+
+
+BUILD_FOLDER_PATH="build/inspireface-linux-aarch64-rk356x-rk3588${TAG}"
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+# export ARM_CROSS_COMPILE_TOOLCHAIN=/host/software/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu
+
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DCMAKE_SYSTEM_VERSION=1 \
+  -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+  -DCMAKE_C_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-gcc \
+  -DCMAKE_CXX_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-g++ \
+  -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -flax-vector-conversions" \
+  -DTARGET_PLATFORM=armlinux \
+  -DISF_BUILD_LINUX_AARCH64=ON \
+  -DISF_BUILD_LINUX_ARM7=OFF \
+  -DMNN_SEP_BUILD=off \
+  -DISF_ENABLE_RKNN=ON \
+  -DISF_RK_DEVICE_TYPE=RK356X \
+  -DISF_RKNPU_MAJOR=rknpu2 \
+  -DISF_RK_COMPILER_TYPE=aarch64 \
+  -DISF_ENABLE_RGA=ON \
+  -DISF_ENABLE_COST_TIME=OFF \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_cross_rv1106_armhf_uclibc.sh b/cpp-package/inspireface/command/build_cross_rv1106_armhf_uclibc.sh
new file mode 100644
index 000000000..db858a70a
--- /dev/null
+++ b/cpp-package/inspireface/command/build_cross_rv1106_armhf_uclibc.sh
@@ -0,0 +1,108 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+
+SCRIPT_DIR=$(pwd)  # Project dir
+
+# Create .rknpu2_cache directory if it doesn't exist
+CACHE_DIR="$(pwd)/.rknpu2_cache"
+mkdir -p "$CACHE_DIR"
+
+# Check if MNN-2.3.0 directory already exists
+if [ ! -d "$CACHE_DIR/MNN-2.3.0" ]; then
+    echo "Downloading MNN 2.3.0..."
+    # Download MNN 2.3.0
+    if ! wget -P "$CACHE_DIR" https://github.com/alibaba/MNN/archive/refs/tags/2.3.0.zip; then
+        echo "Error: Failed to download MNN 2.3.0"
+        exit 1
+    fi
+    
+    # Extract the zip file
+    cd "$CACHE_DIR"
+    if ! unzip 2.3.0.zip; then
+        echo "Error: Failed to extract MNN 2.3.0"
+        exit 1
+    fi
+    
+    # Remove the zip file
+    rm 2.3.0.zip
+    
+    echo "MNN 2.3.0 downloaded and extracted"
+else
+    echo "MNN-2.3.0 already exists in cache"
+fi
+
+# Set absolute path to MNN source
+export ISF_MNN_CUSTOM_SOURCE="$CACHE_DIR/MNN-2.3.0"
+
+echo "ISF_MNN_CUSTOM_SOURCE: ${ISF_MNN_CUSTOM_SOURCE}"
+cd ${SCRIPT_DIR}
+
+# export ARM_CROSS_COMPILE_TOOLCHAIN=/root/arm-rockchip830-linux-uclibcgnueabihf/
+
+BUILD_FOLDER_PATH="build/inspireface-linux-armv7-rv1106-armhf-uclibc${TAG}"
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+# export cross_compile_toolchain=/home/jingyuyan/software/arm-rockchip830-linux-uclibcgnueabihf
+
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DCMAKE_SYSTEM_VERSION=1 \
+  -DCMAKE_SYSTEM_PROCESSOR=armv7 \
+  -DCMAKE_C_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-rockchip830-linux-uclibcgnueabihf-gcc \
+  -DCMAKE_CXX_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-rockchip830-linux-uclibcgnueabihf-g++ \
+  -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -flax-vector-conversions" \
+  -DTARGET_PLATFORM=armlinux \
+  -DISF_BUILD_LINUX_ARM7=ON \
+  -DISF_MNN_CUSTOM_SOURCE=${ISF_MNN_CUSTOM_SOURCE} \
+  -DMNN_SEP_BUILD=off \
+  -DISF_ENABLE_RKNN=ON \
+  -DISF_RK_DEVICE_TYPE=RV1106 \
+  -DISF_RKNPU_MAJOR=rknpu2 \
+  -DISF_RK_COMPILER_TYPE=armhf-uclibc \
+  -DISF_ENABLE_RGA=ON \
+  -DISF_ENABLE_COST_TIME=OFF \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_cross_rv1109rv1126_armhf.sh b/cpp-package/inspireface/command/build_cross_rv1109rv1126_armhf.sh
new file mode 100644
index 000000000..6f7973e22
--- /dev/null
+++ b/cpp-package/inspireface/command/build_cross_rv1109rv1126_armhf.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  # Project dir
+BUILD_FOLDER_PATH="build/inspireface-linux-armv7-rv1109rv1126-armhf${TAG}"
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+# export cross_compile_toolchain=/home/jingyuyan/software/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DCMAKE_SYSTEM_VERSION=1 \
+  -DCMAKE_SYSTEM_PROCESSOR=armv7 \
+  -DCMAKE_C_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-linux-gnueabihf-gcc \
+  -DCMAKE_CXX_COMPILER=$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-linux-gnueabihf-g++ \
+  -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -flax-vector-conversions" \
+  -DTARGET_PLATFORM=armlinux \
+  -DISF_BUILD_LINUX_ARM7=ON \
+  -DISF_ENABLE_RKNN=ON \
+  -DISF_RK_DEVICE_TYPE=RV1109RV1126 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_ios.sh b/cpp-package/inspireface/command/build_ios.sh
new file mode 100644
index 000000000..bb7e7833d
--- /dev/null
+++ b/cpp-package/inspireface/command/build_ios.sh
@@ -0,0 +1,157 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+# Define download URLs
+MNN_IOS_URL="https://github.com/alibaba/MNN/releases/download/2.8.1/mnn_2.8.1_ios_armv82_cpu_metal_coreml.zip"
+
+# Set the cache directory
+MACOS_CACHE="$PWD/.macos_cache/"
+
+# Create the directory if it does not exist
+mkdir -p "${MACOS_CACHE}"
+
+# Function to download and unzip a file if the required framework does not exist
+download_and_unzip() {
+    local url=$1
+    local dir=$2
+    local framework_name=$3  # Name of the framework directory to check
+
+    # Check if the framework already exists
+    if [ ! -d "${dir}${framework_name}" ]; then
+        local file_name=$(basename "$url")
+        local full_path="${dir}${file_name}"
+
+        # Check if the zip file already exists
+        if [ ! -f "$full_path" ]; then
+            echo "Downloading ${file_name}..."
+            # Download the file
+            curl -sL "$url" -o "$full_path"
+        else
+            echo "${file_name} already downloaded. Proceeding to unzip."
+        fi
+
+        # Unzip the file to a temporary directory
+        echo "Unzipping ${file_name}..."
+        unzip -q "$full_path" -d "${dir}"
+        rm "$full_path"
+
+        # Move the framework if it's in a subdirectory specific to the iOS build
+        if [ "${framework_name}" == "MNN.framework" ]; then
+            mv "${dir}ios_build/Release-iphoneos/${framework_name}" "${dir}"
+            rm -rf "${dir}ios_build"  # Clean up the subdirectory
+        fi
+
+        echo "${framework_name} has been set up."
+    else
+        echo "${framework_name} already exists in ${dir}. Skipping download and unzip."
+    fi
+}
+
+# Download and unzip MNN iOS package
+download_and_unzip "$MNN_IOS_URL" "$MACOS_CACHE" "MNN.framework"
+
+# Download and unzip OpenCV iOS package
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+
+TOOLCHAIN="$PWD/toolchain/ios.toolchain.cmake"
+
+BUILD_DIR="build/inspireface-ios$TAG"
+
+mkdir -p "$BUILD_DIR"
+
+cd "$BUILD_DIR"
+
+cmake \
+    -DCMAKE_MAKE_PROGRAM="$(xcrun --find make)" \
+    -DIOS_3RDPARTY="${MACOS_CACHE}" \
+    -DCMAKE_TOOLCHAIN_FILE=${TOOLCHAIN} \
+    -DCMAKE_OSX_ARCHITECTURES=arm64 \
+    -DENABLE_BITCODE=0 \
+    -DIOS_DEPLOYMENT_TARGET=11.0 \
+    -DISF_BUILD_WITH_SAMPLE=OFF \
+    -DISF_BUILD_WITH_TEST=OFF \
+    -DISF_BUILD_SHARED_LIBS=OFF \
+    ../..
+
+make -j8
+
+make install
+
+move_install_files "$(pwd)"
+
+# Set the framework name
+FRAMEWORK_NAME=InspireFace
+
+# Specify the version of the framework
+FRAMEWORK_VERSION=1.0.0
+
+# Root build directory
+BUILD_DIR="$(pwd)"
+
+BUILD_LIB_DIR="$BUILD_DIR/InspireFace"
+
+# Create the framework structure
+FRAMEWORK_DIR=$BUILD_DIR/$FRAMEWORK_NAME.framework
+mkdir -p $FRAMEWORK_DIR
+mkdir -p $FRAMEWORK_DIR/Headers
+mkdir -p $FRAMEWORK_DIR/Resources
+
+# Copy the static library to the framework directory
+cp $BUILD_LIB_DIR/lib/libInspireFace.a $FRAMEWORK_DIR/$FRAMEWORK_NAME
+
+# Copy header files to the framework's Headers directory
+cp $BUILD_LIB_DIR/include/*.h $FRAMEWORK_DIR/Headers/
+
+# Create Info.plist
+cat <<EOF >$FRAMEWORK_DIR/Resources/Info.plist
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>CFBundleExecutable</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleIdentifier</key>
+    <string>com.example.$FRAMEWORK_NAME</string>
+    <key>CFBundleName</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleVersion</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundleShortVersionString</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundlePackageType</key>
+    <string>FMWK</string>
+</dict>
+</plist>
+EOF
+
+echo "Framework $FRAMEWORK_NAME.framework has been created at $FRAMEWORK_DIR"
+
+cp -r $MACOS_CACHE/MNN.framework $BUILD_DIR/
diff --git a/cpp-package/inspireface/command/build_ios_coreml.sh b/cpp-package/inspireface/command/build_ios_coreml.sh
new file mode 100644
index 000000000..601615d7d
--- /dev/null
+++ b/cpp-package/inspireface/command/build_ios_coreml.sh
@@ -0,0 +1,157 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+# Define download URLs
+MNN_IOS_URL="https://github.com/alibaba/MNN/releases/download/2.8.1/mnn_2.8.1_ios_armv82_cpu_metal_coreml.zip"
+
+# Set the cache directory
+MACOS_CACHE="$PWD/.macos_cache/"
+
+# Create the directory if it does not exist
+mkdir -p "${MACOS_CACHE}"
+
+# Function to download and unzip a file if the required framework does not exist
+download_and_unzip() {
+    local url=$1
+    local dir=$2
+    local framework_name=$3  # Name of the framework directory to check
+
+    # Check if the framework already exists
+    if [ ! -d "${dir}${framework_name}" ]; then
+        local file_name=$(basename "$url")
+        local full_path="${dir}${file_name}"
+
+        # Check if the zip file already exists
+        if [ ! -f "$full_path" ]; then
+            echo "Downloading ${file_name}..."
+            # Download the file
+            curl -sL "$url" -o "$full_path"
+        else
+            echo "${file_name} already downloaded. Proceeding to unzip."
+        fi
+
+        # Unzip the file to a temporary directory
+        echo "Unzipping ${file_name}..."
+        unzip -q "$full_path" -d "${dir}"
+        rm "$full_path"
+
+        # Move the framework if it's in a subdirectory specific to the iOS build
+        if [ "${framework_name}" == "MNN.framework" ]; then
+            mv "${dir}ios_build/Release-iphoneos/${framework_name}" "${dir}"
+            rm -rf "${dir}ios_build"  # Clean up the subdirectory
+        fi
+
+        echo "${framework_name} has been set up."
+    else
+        echo "${framework_name} already exists in ${dir}. Skipping download and unzip."
+    fi
+}
+
+# Download and unzip MNN iOS package
+download_and_unzip "$MNN_IOS_URL" "$MACOS_CACHE" "MNN.framework"
+
+# Download and unzip OpenCV iOS package
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+
+TOOLCHAIN="$PWD/toolchain/ios.toolchain.cmake"
+
+BUILD_DIR="build/inspireface-ios-coreml-arm64$TAG"
+
+mkdir -p "$BUILD_DIR"
+
+cd "$BUILD_DIR"
+
+cmake \
+    -DIOS_3RDPARTY="${MACOS_CACHE}" \
+    -DCMAKE_TOOLCHAIN_FILE=${TOOLCHAIN} \
+    -DCMAKE_OSX_ARCHITECTURES=arm64 \
+    -DENABLE_BITCODE=0 \
+    -DIOS_DEPLOYMENT_TARGET=11.0 \
+    -DISF_ENABLE_APPLE_EXTENSION=ON \
+    -DISF_BUILD_WITH_SAMPLE=OFF \
+    -DISF_BUILD_WITH_TEST=OFF \
+    -DISF_BUILD_SHARED_LIBS=OFF \
+    ../..
+
+make -j8
+
+make install
+
+move_install_files "$(pwd)"
+
+# Set the framework name
+FRAMEWORK_NAME=InspireFace
+
+# Specify the version of the framework
+FRAMEWORK_VERSION=1.0.0
+
+# Root build directory
+BUILD_DIR="$(pwd)"
+
+BUILD_LIB_DIR="$BUILD_DIR/InspireFace"
+
+# Create the framework structure
+FRAMEWORK_DIR=$BUILD_DIR/$FRAMEWORK_NAME.framework
+mkdir -p $FRAMEWORK_DIR
+mkdir -p $FRAMEWORK_DIR/Headers
+mkdir -p $FRAMEWORK_DIR/Resources
+
+# Copy the static library to the framework directory
+cp $BUILD_LIB_DIR/lib/libInspireFace.a $FRAMEWORK_DIR/$FRAMEWORK_NAME
+
+# Copy header files to the framework's Headers directory
+cp $BUILD_LIB_DIR/include/*.h $FRAMEWORK_DIR/Headers/
+
+# Create Info.plist
+cat <<EOF >$FRAMEWORK_DIR/Resources/Info.plist
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>CFBundleExecutable</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleIdentifier</key>
+    <string>com.example.$FRAMEWORK_NAME</string>
+    <key>CFBundleName</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleVersion</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundleShortVersionString</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundlePackageType</key>
+    <string>FMWK</string>
+</dict>
+</plist>
+EOF
+
+echo "Framework $FRAMEWORK_NAME.framework has been created at $FRAMEWORK_DIR"
+
+cp -r $MACOS_CACHE/MNN.framework $BUILD_DIR/
diff --git a/cpp-package/inspireface/command/build_linux_cuda.sh b/cpp-package/inspireface/command/build_linux_cuda.sh
new file mode 100644
index 000000000..fd030f2c1
--- /dev/null
+++ b/cpp-package/inspireface/command/build_linux_cuda.sh
@@ -0,0 +1,54 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-linux-cuda${TAG}"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=ON \
+  -DISF_BUILD_WITH_TEST=ON \
+  -DISF_ENABLE_BENCHMARK=ON \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DMNN_CUDA=ON \
+  -DISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA=ON \
+  -Wno-dev \
+  -DISF_LINUX_MNN_CUDA=/home/tunm/softwate/MNN-2.7.2/build_cuda ${SCRIPT_DIR}
+
+make -j4
+
+move_install_files "$(pwd)"
diff --git a/cpp-package/inspireface/command/build_linux_cuda_dev.sh b/cpp-package/inspireface/command/build_linux_cuda_dev.sh
new file mode 100644
index 000000000..0295bb873
--- /dev/null
+++ b/cpp-package/inspireface/command/build_linux_cuda_dev.sh
@@ -0,0 +1,54 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-linux-cuda${TAG}"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_SYSTEM_NAME=Linux \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=ON \
+  -DISF_BUILD_WITH_TEST=ON \
+  -DISF_ENABLE_BENCHMARK=ON \
+  -DISF_ENABLE_USE_LFW_DATA=ON \
+  -DISF_ENABLE_TEST_EVALUATION=ON \
+  -DMNN_CUDA=ON \
+  -DISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA=ON \
+  -Wno-dev \
+  -DISF_LINUX_MNN_CUDA=/host/softwate/MNN-2.7.2/build_cuda ${SCRIPT_DIR}
+
+make -j4
+
+move_install_files "$(pwd)"
diff --git a/cpp-package/inspireface/command/build_linux_manylinux2014.sh b/cpp-package/inspireface/command/build_linux_manylinux2014.sh
new file mode 100644
index 000000000..c4ac8d693
--- /dev/null
+++ b/cpp-package/inspireface/command/build_linux_manylinux2014.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+# Build folder path
+BUILD_FOLDER_PATH="build/inspireface-linux-x86-manylinux2014${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -Wno-dev \
+  -DISF_BUILD_SHARED_LIBS=ON ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_linux_tensorrt.sh b/cpp-package/inspireface/command/build_linux_tensorrt.sh
new file mode 100644
index 000000000..3c30e7367
--- /dev/null
+++ b/cpp-package/inspireface/command/build_linux_tensorrt.sh
@@ -0,0 +1,121 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+get_cuda_ubuntu_tag() {
+    # If CUDA_TAG is set, use it
+    if [ -n "${CUDA_TAG}" ]; then
+        echo "${CUDA_TAG}"
+        return 0
+    fi
+    
+    # Get CUDA version
+    CUDA_VERSION="_none"
+    if command -v nvcc &> /dev/null; then
+        # Try to get version from nvcc
+        CUDA_VERSION=$(nvcc --version 2>/dev/null | grep "release" | awk '{print $6}' | cut -d',' -f1 | tr -d '.')
+        if [ -z "${CUDA_VERSION}" ]; then
+            CUDA_VERSION="_none"
+        else
+            CUDA_VERSION="${CUDA_VERSION}"
+        fi
+    elif [ -f "/usr/local/cuda/version.txt" ]; then
+        # Get version from CUDA installation directory
+        CUDA_VERSION=$(cat /usr/local/cuda/version.txt 2>/dev/null | grep "CUDA Version" | awk '{print $3}' | tr -d '.')
+        if [ -z "${CUDA_VERSION}" ]; then
+            CUDA_VERSION="_none"
+        fi
+    elif [ -d "/usr/local/cuda" ] && ls -l /usr/local/cuda 2>/dev/null | grep -q "cuda-"; then
+        # Get version from symbolic link
+        CUDA_LINK=$(ls -l /usr/local/cuda 2>/dev/null | grep -o "cuda-[0-9.]*" | head -n 1)
+        CUDA_VERSION=$(echo "${CUDA_LINK}" | cut -d'-' -f2 | tr -d '.')
+        if [ -z "${CUDA_VERSION}" ]; then
+            CUDA_VERSION="_none"
+        fi
+    fi
+    
+    # Get Ubuntu version
+    UBUNTU_VERSION="_none"
+    if [ -f "/etc/os-release" ]; then
+        # Check if it is Ubuntu
+        if grep -q "Ubuntu" /etc/os-release 2>/dev/null; then
+            UBUNTU_VERSION=$(grep "VERSION_ID" /etc/os-release 2>/dev/null | cut -d'"' -f2)
+            if [ -z "${UBUNTU_VERSION}" ]; then
+                UBUNTU_VERSION="_none"
+            fi
+        fi
+    elif [ -f "/etc/lsb-release" ]; then
+        # Get version from lsb-release
+        if grep -q "Ubuntu" /etc/lsb-release 2>/dev/null; then
+            UBUNTU_VERSION=$(grep "DISTRIB_RELEASE" /etc/lsb-release 2>/dev/null | cut -d'=' -f2)
+            if [ -z "${UBUNTU_VERSION}" ]; then
+                UBUNTU_VERSION="_none"
+            fi
+        fi
+    fi
+    
+    # Generate and return tag
+    echo "cuda${CUDA_VERSION}_ubuntu${UBUNTU_VERSION}"
+}
+
+CUDA_TAG=$(get_cuda_ubuntu_tag)
+echo "Cuda Tag: ${CUDA_TAG}"
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+SCRIPT_DIR=$(pwd)  
+BUILD_FOLDER_NAME="inspireface-linux-tensorrt-${CUDA_TAG}${TAG}"
+
+mkdir -p build/${BUILD_FOLDER_NAME}
+cd build/${BUILD_FOLDER_NAME}
+
+echo "TENSORRT_ROOT: ${TENSORRT_ROOT}"
+
+cmake  \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DISF_BUILD_WITH_SAMPLE=ON \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_TEST=ON \
+  -DISF_ENABLE_BENCHMARK=ON \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DTENSORRT_ROOT=${TENSORRT_ROOT} \
+  -DISF_ENABLE_TENSORRT=ON \
+  -Wno-dev \
+  ../..
+
+make -j4
+
+make install
+
+if [ $? -eq 0 ] && [ -d "$(pwd)/install" ]; then
+  move_install_files "$(pwd)"
+else
+  echo "Build failed or the installation directory does not exist"
+  exit 1
+fi
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_linux_ubuntu18.sh b/cpp-package/inspireface/command/build_linux_ubuntu18.sh
new file mode 100644
index 000000000..88862b80e
--- /dev/null
+++ b/cpp-package/inspireface/command/build_linux_ubuntu18.sh
@@ -0,0 +1,55 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-linux-x86-ubuntu18${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake --version  
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_macos_arm64.sh b/cpp-package/inspireface/command/build_macos_arm64.sh
new file mode 100644
index 000000000..b5734d1ad
--- /dev/null
+++ b/cpp-package/inspireface/command/build_macos_arm64.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-apple-silicon-arm64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_macos_coreml_arm64.sh b/cpp-package/inspireface/command/build_macos_coreml_arm64.sh
new file mode 100644
index 000000000..87491d4ab
--- /dev/null
+++ b/cpp-package/inspireface/command/build_macos_coreml_arm64.sh
@@ -0,0 +1,54 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-coreml-apple-silicon-arm64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_ENABLE_APPLE_EXTENSION=ON \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=OFF \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_macos_coreml_x86.sh b/cpp-package/inspireface/command/build_macos_coreml_x86.sh
new file mode 100644
index 000000000..675bfef14
--- /dev/null
+++ b/cpp-package/inspireface/command/build_macos_coreml_x86.sh
@@ -0,0 +1,54 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-coreml-intel-x86-64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_ENABLE_APPLE_EXTENSION=ON \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_macos_x86.sh b/cpp-package/inspireface/command/build_macos_x86.sh
new file mode 100644
index 000000000..d6215b18a
--- /dev/null
+++ b/cpp-package/inspireface/command/build_macos_x86.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-intel-x86-64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_wheel_macos_arm64.sh b/cpp-package/inspireface/command/build_wheel_macos_arm64.sh
new file mode 100644
index 000000000..ccbea3691
--- /dev/null
+++ b/cpp-package/inspireface/command/build_wheel_macos_arm64.sh
@@ -0,0 +1,67 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-apple-silicon-arm64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
+BUILD_DYLIB_PATH="$(pwd)/InspireFace/lib/libInspireFace.dylib"
+# Copy the library to the python directory
+
+DYLIB_DEST_PATH="${SCRIPT_DIR}/python/inspireface/modules/core/libs/darwin/arm64/"
+mkdir -p ${DYLIB_DEST_PATH}
+cp -r ${BUILD_DYLIB_PATH} ${DYLIB_DEST_PATH}
+
+pip3 install setuptools wheel twine
+
+PYTHON_PRJ_PATH=${SCRIPT_DIR}/python
+cd ${PYTHON_PRJ_PATH}/
+python3 setup.py bdist_wheel
+
+echo "Build wheel for MacOS Arm64, Well Done!"
diff --git a/cpp-package/inspireface/command/build_wheel_macos_x86.sh b/cpp-package/inspireface/command/build_wheel_macos_x86.sh
new file mode 100644
index 000000000..ae13d1b60
--- /dev/null
+++ b/cpp-package/inspireface/command/build_wheel_macos_x86.sh
@@ -0,0 +1,67 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-macos-intel-x86-64${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
+BUILD_DYLIB_PATH="$(pwd)/InspireFace/lib/libInspireFace.dylib"
+# Copy the library to the python directory
+
+DYLIB_DEST_PATH="${SCRIPT_DIR}/python/inspireface/modules/core/libs/darwin/x64/"
+mkdir -p ${DYLIB_DEST_PATH}
+cp -r ${BUILD_DYLIB_PATH} ${DYLIB_DEST_PATH}
+
+pip3 install setuptools wheel twine
+
+PYTHON_PRJ_PATH=${SCRIPT_DIR}/python
+cd ${PYTHON_PRJ_PATH}/
+python3 setup.py bdist_wheel
+
+echo "Build wheel for MacOS x86_64, Well Done!"
\ No newline at end of file
diff --git a/cpp-package/inspireface/command/build_wheel_manylinux2014_aarch64.sh b/cpp-package/inspireface/command/build_wheel_manylinux2014_aarch64.sh
new file mode 100644
index 000000000..27b35ae81
--- /dev/null
+++ b/cpp-package/inspireface/command/build_wheel_manylinux2014_aarch64.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-aarch64-manylinux2014${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
+BUILD_DYLIB_PATH="$(pwd)/InspireFace/lib/libInspireFace.so"
+# Copy the library to the python directory
+
+DYLIB_DEST_PATH="${SCRIPT_DIR}/python/inspireface/modules/core/libs/linux/arm64/"
+mkdir -p ${DYLIB_DEST_PATH}
+cp -r ${BUILD_DYLIB_PATH} ${DYLIB_DEST_PATH}
+
+
+PYTHON_PRJ_PATH=${SCRIPT_DIR}/python
+cd ${PYTHON_PRJ_PATH}/
+# Build wheels for Python 3.7-3.12
+for PYTHON_VERSION in python3.7 python3.8 python3.9 python3.10 python3.11 python3.12; do
+    if [[ "${PYTHON_VERSION}" == "python3.12" ]]; then
+        ${PYTHON_VERSION} -m pip install setuptools wheel twine
+    fi
+    ${PYTHON_VERSION} setup.py bdist_wheel
+done
+
+echo "Build wheel for Linux arm64, Well Done!"
diff --git a/cpp-package/inspireface/command/build_wheel_manylinux2014_x86.sh b/cpp-package/inspireface/command/build_wheel_manylinux2014_x86.sh
new file mode 100644
index 000000000..a19b65956
--- /dev/null
+++ b/cpp-package/inspireface/command/build_wheel_manylinux2014_x86.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-linux-x86-manylinux2014${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
+BUILD_DYLIB_PATH="$(pwd)/InspireFace/lib/libInspireFace.so"
+# Copy the library to the python directory
+
+DYLIB_DEST_PATH="${SCRIPT_DIR}/python/inspireface/modules/core/libs/linux/x64/"
+mkdir -p ${DYLIB_DEST_PATH}
+cp -r ${BUILD_DYLIB_PATH} ${DYLIB_DEST_PATH}
+
+
+PYTHON_PRJ_PATH=${SCRIPT_DIR}/python
+cd ${PYTHON_PRJ_PATH}/
+# Build wheels for Python 3.7-3.12
+for PYTHON_VERSION in python3.7 python3.8 python3.9 python3.10 python3.11 python3.12; do
+    if [[ "${PYTHON_VERSION}" == "python3.12" ]]; then
+        ${PYTHON_VERSION} -m pip install setuptools wheel twine
+    fi
+    ${PYTHON_VERSION} setup.py bdist_wheel
+done
+
+echo "Build wheel for Linux x86_64, Well Done!"
diff --git a/cpp-package/inspireface/command/build_wheel_manylinux_2_36_aarch64.sh b/cpp-package/inspireface/command/build_wheel_manylinux_2_36_aarch64.sh
new file mode 100644
index 000000000..ba112e215
--- /dev/null
+++ b/cpp-package/inspireface/command/build_wheel_manylinux_2_36_aarch64.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+
+# Reusable function to handle 'install' directory operations
+move_install_files() {
+    local root_dir="$1"
+    local install_dir="$root_dir/install"
+
+    # Step 1: Check if the 'install' directory exists
+    if [ ! -d "$install_dir" ]; then
+        echo "Error: 'install' directory does not exist in $root_dir"
+        exit 1
+    fi
+
+    # Step 2: Delete all other files/folders except 'install'
+    find "$root_dir" -mindepth 1 -maxdepth 1 -not -name "install" -exec rm -rf {} +
+
+    # Step 3: Move all files from 'install' to the root directory
+    mv "$install_dir"/* "$root_dir" 2>/dev/null
+
+    # Step 4: Remove the empty 'install' directory
+    rmdir "$install_dir"
+
+    echo "Files from 'install' moved to $root_dir, and 'install' directory deleted."
+}
+
+if [ -n "$VERSION" ]; then
+    TAG="-$VERSION"
+else
+    TAG=""
+fi
+
+BUILD_FOLDER_PATH="build/inspireface-aarch64-manylinux_2_36${TAG}/"
+SCRIPT_DIR=$(pwd)  # Project dir
+
+mkdir -p ${BUILD_FOLDER_PATH}
+# shellcheck disable=SC2164
+cd ${BUILD_FOLDER_PATH}
+
+cmake -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
+  -DISF_BUILD_WITH_SAMPLE=OFF \
+  -DISF_BUILD_WITH_TEST=OFF \
+  -DISF_ENABLE_BENCHMARK=OFF \
+  -DISF_ENABLE_USE_LFW_DATA=OFF \
+  -DISF_ENABLE_TEST_EVALUATION=OFF \
+  -DISF_BUILD_SHARED_LIBS=ON \
+  -Wno-dev \
+  ${SCRIPT_DIR}
+
+make -j4
+make install
+
+move_install_files "$(pwd)"
+BUILD_DYLIB_PATH="$(pwd)/InspireFace/lib/libInspireFace.so"
+# Copy the library to the python directory
+
+DYLIB_DEST_PATH="${SCRIPT_DIR}/python/inspireface/modules/core/libs/linux/arm64/"
+mkdir -p ${DYLIB_DEST_PATH}
+cp -r ${BUILD_DYLIB_PATH} ${DYLIB_DEST_PATH}
+
+
+PYTHON_PRJ_PATH=${SCRIPT_DIR}/python
+cd ${PYTHON_PRJ_PATH}/
+# Build wheels for Python 3.7-3.12
+for PYTHON_VERSION in python3.7 python3.8 python3.9 python3.10 python3.11 python3.12; do
+    if [[ "${PYTHON_VERSION}" == "python3.12" ]]; then
+        ${PYTHON_VERSION} -m pip install setuptools wheel twine
+    fi
+    ${PYTHON_VERSION} setup.py bdist_wheel
+done
+
+echo "Build wheel for Linux arm64, Well Done!"
diff --git a/cpp-package/inspireface/command/create_ios_framework.sh b/cpp-package/inspireface/command/create_ios_framework.sh
new file mode 100644
index 000000000..c5f2e9cfa
--- /dev/null
+++ b/cpp-package/inspireface/command/create_ios_framework.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+
+# Set the framework name
+FRAMEWORK_NAME=InspireFace
+
+# Specify the version of the framework
+FRAMEWORK_VERSION=1.0.0
+
+# Root build directory
+BUILD_DIR=build/inspireface-ios/install/InspireFace
+
+# Create the framework structure
+FRAMEWORK_DIR=$BUILD_DIR/$FRAMEWORK_NAME.framework
+mkdir -p $FRAMEWORK_DIR
+mkdir -p $FRAMEWORK_DIR/Headers
+mkdir -p $FRAMEWORK_DIR/Resources
+
+# Copy the static library to the framework directory
+cp $BUILD_DIR/lib/libInspireFace.a $FRAMEWORK_DIR/$FRAMEWORK_NAME
+
+# Copy header files to the framework's Headers directory
+cp $BUILD_DIR/include/*.h $FRAMEWORK_DIR/Headers/
+
+# Create Info.plist
+cat <<EOF >$FRAMEWORK_DIR/Resources/Info.plist
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>CFBundleExecutable</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleIdentifier</key>
+    <string>com.example.$FRAMEWORK_NAME</string>
+    <key>CFBundleName</key>
+    <string>$FRAMEWORK_NAME</string>
+    <key>CFBundleVersion</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundleShortVersionString</key>
+    <string>$FRAMEWORK_VERSION</string>
+    <key>CFBundlePackageType</key>
+    <string>FMWK</string>
+</dict>
+</plist>
+EOF
+
+echo "Framework $FRAMEWORK_NAME.framework has been created at $FRAMEWORK_DIR"
diff --git a/cpp-package/inspireface/command/download_models_general.sh b/cpp-package/inspireface/command/download_models_general.sh
new file mode 100644
index 000000000..794553aa4
--- /dev/null
+++ b/cpp-package/inspireface/command/download_models_general.sh
@@ -0,0 +1,97 @@
+#!/bin/bash
+
+# Target download folder
+DOWNLOAD_DIR="test_res/pack"
+
+# File URLs
+URL1="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Megatron"
+URL2="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Pikachu"
+URL3="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RV1109"
+URL4="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RV1106"
+URL5="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RK356X"
+URL6="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Megatron_TRT"
+URL7="https://github.com/HyperInspire/InspireFace/releases/download/v1.x/Gundam_RK3588"
+
+# Color codes
+YELLOW='\033[1;33m'
+NC='\033[0m' # No Color
+
+# Create download folder
+mkdir -p "$DOWNLOAD_DIR"
+
+# Function to download file
+download_file() {
+    local url=$1
+    if command -v wget > /dev/null 2>&1; then
+        echo "Using wget for download..."
+        wget --no-check-certificate -L -P "$DOWNLOAD_DIR" "$url"
+    else
+        echo "wget not found, using curl instead..."
+        cd "$DOWNLOAD_DIR"
+        curl -L -O "$url"
+        cd - > /dev/null
+    fi
+}
+
+# Function to print file path
+print_file_path() {
+    local filename=$1
+    echo -e "File downloaded to: ${YELLOW}$(cd "$DOWNLOAD_DIR" && pwd)/${filename}${NC}"
+}
+
+# Check if argument is provided
+if [ $# -eq 0 ]; then
+    echo "No argument provided, downloading all files..."
+    download_file "$URL1"
+    download_file "$URL2"
+    download_file "$URL3"
+    download_file "$URL4"
+    download_file "$URL5"
+    download_file "$URL6"
+    download_file "$URL7"
+    # Check all files
+    if [ -f "$DOWNLOAD_DIR/Megatron" ] && [ -f "$DOWNLOAD_DIR/Pikachu" ] && \
+       [ -f "$DOWNLOAD_DIR/Gundam_RV1109" ] && [ -f "$DOWNLOAD_DIR/Gundam_RV1106" ] && \
+       [ -f "$DOWNLOAD_DIR/Gundam_RK356X" ] && [ -f "$DOWNLOAD_DIR/Megatron_TRT" ] && \
+       [ -f "$DOWNLOAD_DIR/Gundam_RK3588" ]; then
+        echo "All downloads completed successfully!"
+        print_file_path "Megatron"
+        print_file_path "Pikachu"
+        print_file_path "Gundam_RV1109"
+        print_file_path "Gundam_RV1106"
+        print_file_path "Gundam_RK356X"
+        print_file_path "Megatron_TRT"
+        print_file_path "Gundam_RK3588"
+    else
+        echo "Download failed!"
+        exit 1
+    fi
+else
+    case "$1" in
+        "Megatron"|"Pikachu"|"Gundam_RV1109"|"Gundam_RV1106"|"Gundam_RK356X"|"Megatron_TRT"|"Gundam_RK3588")
+            echo "Downloading $1..."
+            case "$1" in
+                "Megatron") url="$URL1" ;;
+                "Pikachu") url="$URL2" ;;
+                "Gundam_RV1109") url="$URL3" ;;
+                "Gundam_RV1106") url="$URL4" ;;
+                "Gundam_RK356X") url="$URL5" ;;
+                "Megatron_TRT") url="$URL6" ;;
+                "Gundam_RK3588") url="$URL7" ;;
+            esac
+            download_file "$url"
+            # Check file
+            if [ -f "$DOWNLOAD_DIR/$1" ]; then
+                echo "$1 download completed successfully!"
+                print_file_path "$1"
+            else
+                echo "$1 download failed!"
+                exit 1
+            fi
+            ;;
+        *)
+            echo "Invalid argument. Please use 'Megatron', 'Pikachu', 'Gundam_RV1109', 'Gundam_RV1106', 'Gundam_RK356X', 'Megatron_TRT' or 'Gundam_RK3588'"
+            exit 1
+            ;;
+    esac
+fi
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/CMakeLists.txt b/cpp-package/inspireface/cpp/inspireface/CMakeLists.txt
new file mode 100644
index 000000000..bc044009e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/CMakeLists.txt
@@ -0,0 +1,270 @@
+cmake_minimum_required(VERSION 3.20)
+project(InspireFaceSDK)
+
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+
+option(ISF_BUILD_SHARED_LIBS "Build shared libraries (DLLs)." ON)
+
+file(GLOB_RECURSE SOURCE_FILES ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp)
+set(SOURCE_FILES ${SOURCE_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/c_api/inspireface.cc)        # Add C_API file
+
+
+if (ISF_ENABLE_RKNN)
+    if (ISF_RKNPU_MAJOR STREQUAL "rknpu1")
+        set(ISF_RKNN_API_INCLUDE_DIRS ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${ISF_RK_DEVICE_TYPE}/Linux/librknn_api/include)
+        set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${ISF_RK_DEVICE_TYPE}/Linux/librknn_api/${CPU_ARCH}/)
+        link_directories(${ISF_RKNN_API_LIB})
+        set(RKNN_LINKED rknn_api)
+    else()
+        if(ANDROID)
+            set(RK_PLATFORM "Android")
+            set(ISF_RKNN_API_INCLUDE_DIRS ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/include)
+            set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/${ANDROID_ABI}/)
+            link_directories(${ISF_RKNN_API_LIB})
+            set(RKNN_LINKED rknnrt)
+        else()
+            set(RK_PLATFORM "Linux")
+            set(ISF_RKNN_API_INCLUDE_DIRS ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/include)
+            if (ISF_RK_COMPILER_TYPE STREQUAL "aarch64")
+                set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/${ISF_RK_COMPILER_TYPE}/)
+                link_directories(${ISF_RKNN_API_LIB})
+            else()
+                # For rknpu2 with armv7, we recommend linking static libraries by default
+                set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/${ISF_RK_COMPILER_TYPE}/librknnmrt.a)
+                set(RKNN_LINKED ${ISF_RKNN_API_LIB})
+                set(RKNN_USE_STATIC_LIBS TRUE)
+            endif()
+        endif()
+        
+    endif()
+    
+endif()
+
+set(LINK_THIRD_LIBS ${MNN_LIBS})
+
+# OpenCV
+if(ISF_ENABLE_OPENCV)
+    set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${OpenCV_LIBS})
+endif()
+
+# SQLite3
+set(SOURCE_FILES ${SOURCE_FILES} 
+        ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/sqlite/sqlite3.c
+        ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/sqlite/sqlite-vec.c
+)   # Add SQLite3 C_API file
+set(SQLITE_INCLUDE ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/sqlite/)
+
+
+
+if (ISF_ENABLE_RKNN)
+    set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${RKNN_LINKED} dl)
+    # InferenceWrapper use RkNN
+    if (ISF_RKNPU_MAJOR STREQUAL "rknpu1")
+        add_definitions("-DINFERENCE_WRAPPER_ENABLE_RKNN")
+    elseif(ISF_RKNPU_MAJOR STREQUAL "rknpu2")
+        add_definitions("-DINFERENCE_WRAPPER_ENABLE_RKNN2")
+    endif()
+    if (ISF_RK_COMPILER_TYPE STREQUAL "aarch64")
+        set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} rknnrt)
+    endif()
+endif()
+
+# cpp yaml
+file(GLOB_RECURSE CPP_YAML_SRC ${ISF_THIRD_PARTY_DIR}/yaml-cpp/src/*.cpp)
+set(SOURCE_FILES ${SOURCE_FILES} ${CPP_YAML_SRC})
+set(CPP_YAML_INCLUDE ${ISF_THIRD_PARTY_DIR}/yaml-cpp/include)
+
+# archive
+set(SOURCE_FILES ${SOURCE_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/middleware/model_archive/core_archive/microtar/microtar.c)
+# CoreArchive source file
+set(CORE_ARCHIVE_SOURCE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/middleware/model_archive/core_archive/core_archive.cc CACHE PATH "")
+set(SOURCE_FILES ${SOURCE_FILES} ${CORE_ARCHIVE_SOURCE_FILE})
+
+if(ISF_ENABLE_APPLE_EXTENSION)
+    # link apple libs
+    find_library(FOUNDATION_LIBRARY Foundation)
+    find_library(COREML_LIBRARY CoreML)
+    find_library(ACCELERATE_LIBRARY Accelerate)
+    set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${FOUNDATION_LIBRARY} ${COREML_LIBRARY} ${ACCELERATE_LIBRARY})
+
+    # Add objective-c files
+    set(SOURCE_FILES ${SOURCE_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/middleware/inference_wrapper/coreml/CoreMLAdapter.mm)
+
+endif()
+
+# MNN
+link_directories(${MNN_LIBS})
+
+if(ISF_BUILD_SHARED_LIBS)
+    add_definitions("-DISF_BUILD_SHARED_LIBS")
+    add_library(InspireFace SHARED ${SOURCE_FILES} $<TARGET_OBJECTS:inspirecv>)
+else()
+    add_library(InspireFace STATIC ${SOURCE_FILES} $<TARGET_OBJECTS:inspirecv>)
+endif()
+
+target_compile_definitions(InspireFace PUBLIC INFERENCE_WRAPPER_ENABLE_MNN)
+target_compile_definitions(InspireFace PUBLIC FEATURE_BLOCK_ENABLE_OPENCV)
+
+# Include files
+set(NEED_INCLUDE . ${MNN_INCLUDE_DIRS})
+if (ISF_ENABLE_RKNN)
+    set(NEED_INCLUDE ${NEED_INCLUDE} ${ISF_RKNN_API_INCLUDE_DIRS})
+    if(ISF_ENABLE_RGA)
+        set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${RGA_LIBS})
+        set(NEED_INCLUDE ${NEED_INCLUDE} ${RGA_INCLUDE_DIRS})
+    endif()
+endif ()
+
+# Add include directory
+set(NEED_INCLUDE ${NEED_INCLUDE} ${CMAKE_CURRENT_SOURCE_DIR}/include/inspireface)
+
+if (ISF_BUILD_LINUX_ARM7 OR ANDROID)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon")
+endif()
+
+if (ISF_BUILD_LINUX_ARM7 OR ISF_BUILD_LINUX_AARCH64)
+    set(NEED_INCLUDE ${NEED_INCLUDE} ${OpenCV_STATIC_INCLUDE_DIR})
+endif ()
+
+if (ISF_ENABLE_TENSORRT)
+    set(NEED_INCLUDE ${NEED_INCLUDE} ${ISF_TENSORRT_INCLUDE_DIRS})
+    set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${ISF_TENSORRT_LIBRARIES})
+
+endif()
+
+# add cpp yaml header
+set(NEED_INCLUDE ${NEED_INCLUDE} ${CPP_YAML_INCLUDE} ${SQLITE_INCLUDE})
+
+if(PLAT STREQUAL "linux")
+    find_package(Threads REQUIRED)
+    set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${CMAKE_THREAD_LIBS_INIT} dl)
+endif()
+
+# set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} inspirecv)
+set(NEED_INCLUDE ${NEED_INCLUDE} ${INSPIRECV_INCLUDE_PATH})
+
+target_include_directories(InspireFace PUBLIC
+    ${NEED_INCLUDE}
+)
+
+if (ANDROID)
+    target_link_options(${CMAKE_PROJECT_NAME} PRIVATE "-Wl,-z,max-page-size=16384")
+    if(ANDROID_NDK_MAJOR LESS_EQUAL 22)
+        target_link_options(${CMAKE_PROJECT_NAME} PRIVATE "-Wl,-z,common-page-size=16384")
+    endif()
+    target_link_libraries(InspireFace PUBLIC ${LINK_THIRD_LIBS} jnigraphics log)
+
+    set_target_properties(InspireFace PROPERTIES
+            LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib/
+            ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib/
+            )
+elseif(IOS)
+#    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ObjC")
+#    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ObjC")
+    target_link_libraries(InspireFace
+        "-fobjc-arc"
+        "-framework Metal"
+        "-framework CoreML"
+        "-framework Foundation"
+        "-framework CoreVideo"
+        "-framework CoreMedia"
+        ${MNN_FRAMEWORK_PATH}
+    )
+    if(ISF_ENABLE_APPLE_EXTENSION)
+        set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} ${ACCELERATE_LIBRARY})
+    endif()
+    if(ISF_ENABLE_OPENCV)
+        set(LINK_THIRD_LIBS ${LINK_THIRD_LIBS} "-framework opencv2")
+    endif()
+    
+else()
+    target_link_libraries(InspireFace PUBLIC ${LINK_THIRD_LIBS})
+
+    set_target_properties(InspireFace PROPERTIES
+            LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib/
+            ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib/
+    )
+endif()
+
+# Print Message
+message(STATUS ">>>>>>>>>>>>>")
+message(STATUS "InspireFace Core:")
+message(STATUS "\t Version: ${INSPIRE_FACE_VERSION_MAJOR}.${INSPIRE_FACE_VERSION_MINOR}.${INSPIRE_FACE_VERSION_PATCH}")
+message(STATUS "\t MNN_INCLUDE_DIRS: ${MNN_INCLUDE_DIRS}")
+message(STATUS "\t MNN_LIBS: ${MNN_LIBS}")
+message(STATUS "\t ISF_BUILD_SHARED_LIBS: ${ISF_BUILD_SHARED_LIBS}")
+message(STATUS "\t ISF_ENABLE_RKNN: ${ISF_ENABLE_RKNN}")
+if (ISF_ENABLE_RKNN)
+    message(STATUS "\t ISF_RKNN_API_INCLUDE_DIRS: ${ISF_RKNN_API_INCLUDE_DIRS}")
+    message(STATUS "\t ISF_RKNN_API_LIB: ${ISF_RKNN_API_LIB}")
+endif ()
+if (ISF_ENABLE_RGA)
+    message(STATUS "\t ISF_ENABLE_RGA: ${ISF_ENABLE_RGA}")
+    message(STATUS "\t RGA_LIBS: ${RGA_LIBS}")
+    message(STATUS "\t RGA_INCLUDE_DIRS: ${RGA_INCLUDE_DIRS}")
+endif ()
+if (ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA)
+    message(STATUS "\t ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA: ${ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA}")
+endif ()
+
+message(STATUS "\t ISF_ENABLE_TENSORRT: ${ISF_ENABLE_TENSORRT}")
+if (ISF_ENABLE_TENSORRT)
+    message(STATUS "\t TENSORRT_INCLUDE_DIR: ${TENSORRT_INCLUDE_DIR}")
+    message(STATUS "\t TENSORRT_LIBRARY_INFER: ${TENSORRT_LIBRARY_INFER}")
+    message(STATUS "\t TENSORRT_LIBRARY_RUNTIME: ${TENSORRT_LIBRARY_RUNTIME}")
+    message(STATUS "\t CUDA_RUNTIME_LIBRARY: ${CUDA_RUNTIME_LIBRARY}")
+endif()
+
+# Install lib
+install(TARGETS InspireFace
+        LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib
+        ARCHIVE DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib
+)
+
+# Install header file
+install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/c_api/inspireface.h DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/include)
+
+install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/c_api/intypedef.h DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/include)
+
+install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/include/inspireface/herror.h DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/include)
+
+install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/version.txt DESTINATION ${CMAKE_INSTALL_PREFIX}/)
+
+if(ISF_INSTALL_CPP_HEADER)
+    # Install cpp api header file
+    install(DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/include/inspireface/ DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/include/inspireface/)
+
+    install(DIRECTORY ${INSPIRECV_INCLUDE_PATH}/inspirecv DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/include/)
+endif()
+
+if (ISF_ENABLE_RKNN AND ISF_RKNPU_MAJOR STREQUAL "rknpu1")
+    # Install rknn 3rd lib
+    install(FILES ${ISF_RKNN_API_LIB}/librknn_api.so DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+endif ()
+
+if(ISF_ENABLE_RKNN AND ISF_RKNPU_MAJOR STREQUAL "rknpu2" AND RK_PLATFORM STREQUAL "Android")
+    install(FILES ${ISF_RKNN_API_LIB}/librknnrt.so DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+endif()
+
+
+if (ISF_RK_COMPILER_TYPE STREQUAL "aarch64")
+    install(FILES ${ISF_RKNN_API_LIB}/librknnrt.so DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+endif()
+
+if (NOT IOS)
+    if (NOT ISF_BUILD_SHARED_LIBS)
+        if(MNN_BUILD_SHARED_LIBS)
+            install(FILES ${CMAKE_BINARY_DIR}/MNN/libMNN.so DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+        else()
+            install(FILES ${CMAKE_BINARY_DIR}/MNN/libMNN.a DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+        endif()
+        if(RKNN_USE_STATIC_LIBS)
+            # To be added: The compilation of the RK series needs to be added
+            install(FILES ${ISF_RKNN_API_LIB} DESTINATION ${CMAKE_INSTALL_PREFIX}/InspireFace/lib)
+        endif()
+        
+    endif()
+
+endif()
diff --git a/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.cc b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.cc
new file mode 100644
index 000000000..2ea8287b3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.cc
@@ -0,0 +1,1540 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "inspireface.h"
+#include "intypedef.h"
+#include "inspireface_internal.h"
+#include "information.h"
+#include "feature_hub_db.h"
+#include <launch.h>
+#include "runtime_module/resource_manage.h"
+#include "similarity_converter.h"
+#include "middleware/inference_wrapper/inference_wrapper.h"
+#if defined(ISF_ENABLE_TENSORRT)
+#include "cuda_toolkit.h"
+#endif
+#include <cstdarg>
+
+#define FACE_FEATURE_SIZE 512  ///< Temporary setup
+
+using namespace inspire;
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStream(PHFImageData data, PHFImageStream handle) {
+    if (data == nullptr || handle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+
+    auto stream = new HF_CameraStream();
+    switch (data->rotation) {
+        case HF_CAMERA_ROTATION_90:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_90);
+            break;
+        case HF_CAMERA_ROTATION_180:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_180);
+            break;
+        case HF_CAMERA_ROTATION_270:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_270);
+            break;
+        default:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_0);
+            break;
+    }
+    switch (data->format) {
+        case HF_STREAM_RGB:
+            stream->impl.SetDataFormat(inspirecv::RGB);
+            break;
+        case HF_STREAM_BGR:
+            stream->impl.SetDataFormat(inspirecv::BGR);
+            break;
+        case HF_STREAM_RGBA:
+            stream->impl.SetDataFormat(inspirecv::RGBA);
+            break;
+        case HF_STREAM_BGRA:
+            stream->impl.SetDataFormat(inspirecv::BGRA);
+            break;
+        case HF_STREAM_YUV_NV12:
+            stream->impl.SetDataFormat(inspirecv::NV12);
+            break;
+        case HF_STREAM_YUV_NV21:
+            stream->impl.SetDataFormat(inspirecv::NV21);
+            break;
+        case HF_STREAM_I420:
+            stream->impl.SetDataFormat(inspirecv::I420);
+            break;
+        case HF_STREAM_GRAY:
+            stream->impl.SetDataFormat(inspirecv::GRAY);
+            break;
+        default:
+            return HERR_INVALID_IMAGE_STREAM_PARAM;  // Assume there's a return code for unsupported
+                                                     // formats
+    }
+    stream->impl.SetDataBuffer(data->data, data->height, data->width);
+
+    *handle = (HFImageStream)stream;
+
+    // Record the creation of this stream in the ResourceManager
+    RESOURCE_MANAGE->createStream((long)*handle);
+
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStreamEmpty(PHFImageStream handle) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    auto stream = new HF_CameraStream();
+    *handle = (HFImageStream)stream;
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetBuffer(HFImageStream handle, HPUInt8 buffer, HInt32 width, HInt32 height) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    ((HF_CameraStream *)handle)->impl.SetDataBuffer(buffer, width, height);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetRotation(HFImageStream handle, HFRotation rotation) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    switch (rotation) {
+        case HF_CAMERA_ROTATION_90:
+            ((HF_CameraStream *)handle)->impl.SetRotationMode(inspirecv::ROTATION_90);
+            break;
+        case HF_CAMERA_ROTATION_180:
+            ((HF_CameraStream *)handle)->impl.SetRotationMode(inspirecv::ROTATION_180);
+            break;
+        case HF_CAMERA_ROTATION_270:
+            ((HF_CameraStream *)handle)->impl.SetRotationMode(inspirecv::ROTATION_270);
+            break;
+        default:
+            ((HF_CameraStream *)handle)->impl.SetRotationMode(inspirecv::ROTATION_0);
+            break;
+    }
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetFormat(HFImageStream handle, HFImageFormat format) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    switch (format) {
+        case HF_STREAM_RGB:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::RGB);
+            break;
+        case HF_STREAM_BGR:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::BGR);
+            break;
+        case HF_STREAM_RGBA:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::RGBA);
+            break;
+        case HF_STREAM_BGRA:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::BGRA);
+            break;
+        case HF_STREAM_YUV_NV12:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::NV12);
+            break;
+        case HF_STREAM_YUV_NV21:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::NV21);
+            break;
+        case HF_STREAM_I420:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::I420);
+            break;
+        case HF_STREAM_GRAY:
+            ((HF_CameraStream *)handle)->impl.SetDataFormat(inspirecv::GRAY);
+            break;
+        default:
+            return HERR_INVALID_IMAGE_STREAM_PARAM;  // Assume there's a return code for unsupported
+                                                     // formats
+    }
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFReleaseImageStream(HFImageStream streamHandle) {
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    // Check and mark this stream as released in the ResourceManager
+    if (!RESOURCE_MANAGE->releaseStream((long)streamHandle)) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;  // or other appropriate error code
+    }
+    delete (HF_CameraStream *)streamHandle;
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmap(PHFImageBitmapData data, PHFImageBitmap handle) {
+    if (data == nullptr || handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    auto bitmap = new HF_ImageBitmap();
+    bitmap->impl.Reset(data->width, data->height, data->channels, data->data);
+    *handle = (HFImageBitmap)bitmap;
+    // Record the creation of this image bitmap in the ResourceManager
+    RESOURCE_MANAGE->createImageBitmap((long)*handle);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmapFromFilePath(HPath filePath, HInt32 channels, PHFImageBitmap handle) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    auto image = inspirecv::Image::Create(filePath, channels);
+    auto bitmap = new HF_ImageBitmap();
+    bitmap->impl.Reset(image.Width(), image.Height(), image.Channels(), image.Data());
+    *handle = (HFImageBitmap)bitmap;
+    // Record the creation of this image bitmap in the ResourceManager
+    RESOURCE_MANAGE->createImageBitmap((long)*handle);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapCopy(HFImageBitmap handle, PHFImageBitmap copyHandle) {
+    if (handle == nullptr || copyHandle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    auto bitmap = new HF_ImageBitmap();
+    bitmap->impl.Reset(((HF_ImageBitmap *)handle)->impl.Width(), ((HF_ImageBitmap *)handle)->impl.Height(),
+                       ((HF_ImageBitmap *)handle)->impl.Channels(), ((HF_ImageBitmap *)handle)->impl.Data());
+    *copyHandle = (HFImageBitmap)bitmap;
+    // Record the creation of this image bitmap in the ResourceManager
+    RESOURCE_MANAGE->createImageBitmap((long)*copyHandle);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFReleaseImageBitmap(HFImageBitmap handle) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    // Check and mark this image bitmap as released in the ResourceManager
+    if (!RESOURCE_MANAGE->releaseImageBitmap((long)handle)) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;  // or other appropriate error code
+    }
+    delete (HF_ImageBitmap *)handle;
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStreamFromImageBitmap(HFImageBitmap handle, HFRotation rotation, PHFImageStream streamHandle) {
+    if (handle == nullptr || streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    auto stream = new HF_CameraStream();
+    switch (rotation) {
+        case HF_CAMERA_ROTATION_90:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_90);
+            break;
+        case HF_CAMERA_ROTATION_180:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_180);
+            break;
+        case HF_CAMERA_ROTATION_270:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_270);
+            break;
+        default:
+            stream->impl.SetRotationMode(inspirecv::ROTATION_0);
+            break;
+    }
+    if (((HF_ImageBitmap *)handle)->impl.Channels() == 1) {
+        stream->impl.SetDataFormat(inspirecv::GRAY);
+    } else {
+        stream->impl.SetDataFormat(inspirecv::BGR);
+    }
+    stream->impl.SetDataBuffer(((HF_ImageBitmap *)handle)->impl.Data(), ((HF_ImageBitmap *)handle)->impl.Height(),
+                               ((HF_ImageBitmap *)handle)->impl.Width());
+    *streamHandle = (HFImageStream)stream;
+
+    // Record the creation of this stream in the ResourceManager
+    RESOURCE_MANAGE->createStream((long)*streamHandle);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmapFromImageStreamProcess(HFImageStream streamHandle, PHFImageBitmap handle, HInt32 is_rotate,
+                                                                           HFloat scale) {
+    if (streamHandle == nullptr || handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    auto bitmap = new HF_ImageBitmap();
+    auto img = ((HF_CameraStream *)streamHandle)->impl.ExecuteImageScaleProcessing(scale, is_rotate);
+    bitmap->impl.Reset(img.Width(), img.Height(), img.Channels(), img.Data());
+    *handle = (HFImageBitmap)bitmap;
+    // Record the creation of this image bitmap in the ResourceManager
+    RESOURCE_MANAGE->createImageBitmap((long)*handle);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapWriteToFile(HFImageBitmap handle, HPath filePath) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    auto success = ((HF_ImageBitmap *)handle)->impl.Write(filePath);
+    if (success) {
+        return HSUCCEED;
+    } else {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawRect(HFImageBitmap handle, HFaceRect rect, HColor color, HInt32 thickness) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    inspirecv::Rect<int> rect_inner(rect.x, rect.y, rect.width, rect.height);
+    ((HF_ImageBitmap *)handle)->impl.DrawRect(rect_inner, {color.r, color.g, color.b}, thickness);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawCircle(HFImageBitmap handle, HPoint2i point, HInt32 radius, HColor color, HInt32 thickness) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    ((HF_ImageBitmap *)handle)->impl.DrawCircle({point.x, point.y}, radius, {color.r, color.g, color.b}, thickness);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawCircleF(HFImageBitmap handle, HPoint2f point, HInt32 radius, HColor color, HInt32 thickness) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    ((HF_ImageBitmap *)handle)->impl.DrawCircle({(int)point.x, (int)point.y}, radius, {color.r, color.g, color.b}, thickness);
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapGetData(HFImageBitmap handle, PHFImageBitmapData data) {
+    if (handle == nullptr || data == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    data->width = ((HF_ImageBitmap *)handle)->impl.Width();
+    data->height = ((HF_ImageBitmap *)handle)->impl.Height();
+    data->channels = ((HF_ImageBitmap *)handle)->impl.Channels();
+    data->data = (uint8_t *)((HF_ImageBitmap *)handle)->impl.Data();
+    return HSUCCEED;
+}
+
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapShow(HFImageBitmap handle, HString title, HInt32 delay) {
+    if (handle == nullptr) {
+        return HERR_INVALID_IMAGE_BITMAP_HANDLE;
+    }
+    ((HF_ImageBitmap *)handle)->impl.Show(title, delay);
+    return HSUCCEED;
+}
+
+void HFDeBugImageStreamImShow(HFImageStream streamHandle) {
+    if (streamHandle == nullptr) {
+        INSPIRE_LOGE("Handle error");
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        INSPIRE_LOGE("Image error");
+        return;
+    }
+    auto image = stream->impl.ExecuteImageScaleProcessing(1.0f, true);
+#ifdef DISABLE_GUI
+    image.Write("tmp.jpg");
+#else
+    image.Show();
+#endif
+}
+
+HResult HFDeBugImageStreamDecodeSave(HFImageStream streamHandle, HPath savePath) {
+    if (streamHandle == nullptr) {
+        INSPIRE_LOGE("Handle error");
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        INSPIRE_LOGE("Image error");
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    auto image = stream->impl.ExecuteImageScaleProcessing(1.0f, true);
+    auto ret = image.Write(savePath);
+    if (ret) {
+        INSPIRE_LOGI("Image saved successfully to %s", savePath);
+        return HSUCCEED;
+    } else {
+        INSPIRE_LOGE("Failed to save image to %s", savePath);
+        return HERR_IMAGE_STREAM_DECODE_FAILED;
+    }
+}
+
+HResult HFReleaseInspireFaceSession(HFSession handle) {
+    if (handle == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    // Check and mark this session as released in the ResourceManager
+    if (!RESOURCE_MANAGE->releaseSession((long)handle)) {
+        return HERR_INVALID_CONTEXT_HANDLE;  // or other appropriate error code
+    }
+    delete (HF_FaceAlgorithmSession *)handle;
+    return HSUCCEED;
+}
+
+HResult HFSessionClearTrackingFace(HFSession session) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    ctx->impl.ClearTrackingFace();
+    return HSUCCEED;
+}
+
+HResult HFSessionSetTrackLostRecoveryMode(HFSession session, HInt32 enable) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    ctx->impl.SetTrackLostRecoveryMode(enable);
+    return HSUCCEED;
+}
+
+HResult HFSessionSetLightTrackConfidenceThreshold(HFSession session, HFloat value) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    ctx->impl.SetLightTrackConfidenceThreshold(value);
+    return HSUCCEED;
+}
+
+HResult HFSwitchLandmarkEngine(HFSessionLandmarkEngine engine) {
+    inspire::Launch::LandmarkEngine type;
+    if (engine == HF_LANDMARK_HYPLMV2_0_25) {
+        type = inspire::Launch::LANDMARK_HYPLMV2_0_25;
+    } else if (engine == HF_LANDMARK_HYPLMV2_0_50) {
+        type = inspire::Launch::LANDMARK_HYPLMV2_0_50;
+    } else if (engine == HF_LANDMARK_INSIGHTFACE_2D106_TRACK) {
+        type = inspire::Launch::LANDMARK_INSIGHTFACE_2D106_TRACK;
+    } else {
+        INSPIRE_LOGE("Unsupported Landmark engine.");
+        return HERR_INVALID_PARAM;
+    }
+    INSPIREFACE_CONTEXT->SwitchLandmarkEngine(type);
+    return HSUCCEED;
+}
+
+HResult HFQuerySupportedPixelLevelsForFaceDetection(PHFFaceDetectPixelList pixel_levels) {
+    auto ret = INSPIREFACE_CONTEXT->GetFaceDetectPixelList();
+    pixel_levels->size = ret.size();
+    for (int i = 0; i < ret.size(); i++) {
+        pixel_levels->pixel_level[i] = ret[i];
+    }
+    return HSUCCEED;
+}
+
+HResult HFCreateInspireFaceSession(HFSessionCustomParameter parameter, HFDetectMode detectMode, HInt32 maxDetectFaceNum, HInt32 detectPixelLevel,
+                                   HInt32 trackByDetectModeFPS, PHFSession handle) {
+    inspire::ContextCustomParameter param;
+    param.enable_mask_detect = parameter.enable_mask_detect;
+    param.enable_liveness = parameter.enable_liveness;
+    param.enable_face_quality = parameter.enable_face_quality;
+    param.enable_interaction_liveness = parameter.enable_interaction_liveness;
+    param.enable_ir_liveness = parameter.enable_ir_liveness;
+    param.enable_recognition = parameter.enable_recognition;
+    param.enable_face_attribute = parameter.enable_face_attribute;
+    param.enable_face_pose = parameter.enable_face_pose;
+    param.enable_face_emotion = parameter.enable_face_emotion;
+    inspire::DetectModuleMode detMode = inspire::DETECT_MODE_ALWAYS_DETECT;
+    if (detectMode == HF_DETECT_MODE_LIGHT_TRACK) {
+        detMode = inspire::DETECT_MODE_LIGHT_TRACK;
+    } else if (detectMode == HF_DETECT_MODE_TRACK_BY_DETECTION) {
+        detMode = inspire::DETECT_MODE_TRACK_BY_DETECT;
+    }
+
+    HF_FaceAlgorithmSession *ctx = new HF_FaceAlgorithmSession();
+    auto ret = ctx->impl.Configuration(detMode, maxDetectFaceNum, param, detectPixelLevel, trackByDetectModeFPS);
+    if (ret != HSUCCEED) {
+        delete ctx;
+        *handle = nullptr;
+    } else {
+        *handle = ctx;
+        // Record the creation of this session in the ResourceManager
+        RESOURCE_MANAGE->createSession((long)*handle);
+    }
+
+    return ret;
+}
+
+HResult HFCreateInspireFaceSessionOptional(HOption customOption, HFDetectMode detectMode, HInt32 maxDetectFaceNum, HInt32 detectPixelLevel,
+                                           HInt32 trackByDetectModeFPS, PHFSession handle) {
+    inspire::ContextCustomParameter param;
+    if (customOption & HF_ENABLE_FACE_RECOGNITION) {
+        param.enable_recognition = true;
+    }
+    if (customOption & HF_ENABLE_LIVENESS) {
+        param.enable_liveness = true;
+    }
+    if (customOption & HF_ENABLE_IR_LIVENESS) {
+        param.enable_ir_liveness = true;
+    }
+    if (customOption & HF_ENABLE_FACE_ATTRIBUTE) {
+        param.enable_face_attribute = true;
+    }
+    if (customOption & HF_ENABLE_MASK_DETECT) {
+        param.enable_mask_detect = true;
+    }
+    if (customOption & HF_ENABLE_QUALITY) {
+        param.enable_face_quality = true;
+    }
+    if (customOption & HF_ENABLE_INTERACTION) {
+        param.enable_interaction_liveness = true;
+    }
+    if (customOption & HF_ENABLE_FACE_POSE) {
+        param.enable_face_pose = true;
+    }
+    if (customOption & HF_ENABLE_FACE_EMOTION) {
+        param.enable_face_emotion = true;
+    }
+    inspire::DetectModuleMode detMode = inspire::DETECT_MODE_ALWAYS_DETECT;
+    if (detectMode == HF_DETECT_MODE_LIGHT_TRACK) {
+        detMode = inspire::DETECT_MODE_LIGHT_TRACK;
+    } else if (detectMode == HF_DETECT_MODE_TRACK_BY_DETECTION) {
+        detMode = inspire::DETECT_MODE_TRACK_BY_DETECT;
+    }
+
+    HF_FaceAlgorithmSession *ctx = new HF_FaceAlgorithmSession();
+    auto ret = ctx->impl.Configuration(detMode, maxDetectFaceNum, param, detectPixelLevel, trackByDetectModeFPS);
+    if (ret != HSUCCEED) {
+        delete ctx;
+        *handle = nullptr;
+    } else {
+        *handle = ctx;
+        // Record the creation of this session in the ResourceManager
+        RESOURCE_MANAGE->createSession((long)*handle);
+    }
+
+    return ret;
+}
+
+HResult HFLaunchInspireFace(HPath resourcePath) {
+    std::string path(resourcePath);
+    return INSPIREFACE_CONTEXT->Load(resourcePath);
+}
+
+HResult HFReloadInspireFace(HPath resourcePath) {
+    std::string path(resourcePath);
+    return INSPIREFACE_CONTEXT->Reload(resourcePath);
+}
+
+HResult HFTerminateInspireFace() {
+    INSPIREFACE_CONTEXT->Unload();
+    return HSUCCEED;
+}
+
+HResult HFQueryInspireFaceLaunchStatus(HPInt32 status) {
+    *status = INSPIREFACE_CONTEXT->isMLoad();
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubDataDisable() {
+    return INSPIREFACE_FEATURE_HUB->DisableHub();
+}
+
+HResult HFQueryExpansiveHardwareRGACompileOption(HPInt32 enable) {
+#if defined(ISF_ENABLE_RGA)
+    INSPIRE_LOGI("RGA is enabled during compilation");
+    *enable = 1;
+#else
+    INSPIRE_LOGW("RGA is not enabled during compilation");
+    *enable = 0;
+#endif
+    return HSUCCEED;
+}
+
+HResult HFSetExpansiveHardwareRockchipDmaHeapPath(HPath path) {
+    INSPIREFACE_CONTEXT->SetRockchipDmaHeapPath(path);
+    return HSUCCEED;
+}
+
+HResult HFQueryExpansiveHardwareRockchipDmaHeapPath(HString path) {
+    strcpy(path, INSPIREFACE_CONTEXT->GetRockchipDmaHeapPath().c_str());
+    return HSUCCEED;
+}
+
+HResult HFSetAppleCoreMLInferenceMode(HFAppleCoreMLInferenceMode mode) {
+    if (mode == HF_APPLE_COREML_INFERENCE_MODE_CPU) {
+        INSPIREFACE_CONTEXT->SetGlobalCoreMLInferenceMode(inspire::Launch::NN_INFERENCE_CPU);
+    } else if (mode == HF_APPLE_COREML_INFERENCE_MODE_GPU) {
+        INSPIREFACE_CONTEXT->SetGlobalCoreMLInferenceMode(inspire::Launch::NN_INFERENCE_COREML_GPU);
+    } else if (mode == HF_APPLE_COREML_INFERENCE_MODE_ANE) {
+        INSPIREFACE_CONTEXT->SetGlobalCoreMLInferenceMode(inspire::Launch::NN_INFERENCE_COREML_ANE);
+    } else {
+        INSPIRE_LOGE("Unsupported Apple CoreML inference mode.");
+        return HERR_INVALID_PARAM;
+    }
+    return HSUCCEED;
+}
+
+HResult HFSwitchImageProcessingBackend(HFImageProcessingBackend backend) {
+    if (backend == HF_IMAGE_PROCESSING_CPU) {
+        INSPIREFACE_CONTEXT->SwitchImageProcessingBackend(inspire::Launch::IMAGE_PROCESSING_CPU);
+    } else if (backend == HF_IMAGE_PROCESSING_RGA) {
+        INSPIREFACE_CONTEXT->SwitchImageProcessingBackend(inspire::Launch::IMAGE_PROCESSING_RGA);
+    } else {
+        INSPIRE_LOGE("Unsupported image processing backend.");
+        return HERR_INVALID_PARAM;
+    }
+    return HSUCCEED;
+}
+
+HResult HFSetImageProcessAlignedWidth(HInt32 width) {
+    INSPIREFACE_CONTEXT->SetImageProcessAlignedWidth(width);
+    return HSUCCEED;
+}
+
+HResult HFSetCudaDeviceId(HInt32 device_id) {
+    INSPIREFACE_CONTEXT->SetCudaDeviceId(device_id);
+    return HSUCCEED;
+}
+
+HResult HFGetCudaDeviceId(HPInt32 device_id) {
+    *device_id = INSPIREFACE_CONTEXT->GetCudaDeviceId();
+    return HSUCCEED;
+}
+
+HResult HFPrintCudaDeviceInfo() {
+#if defined(ISF_ENABLE_TENSORRT)
+    return inspire::PrintCudaDeviceInfo();
+#else
+    INSPIRE_LOGW("CUDA is not supported, you need to enable the compile option that supports TensorRT");
+    return HERR_DEVICE_CUDA_DISABLE;
+#endif
+}
+
+HResult HFGetNumCudaDevices(HPInt32 num_devices) {
+#if defined(ISF_ENABLE_TENSORRT)
+    return inspire::GetCudaDeviceCount(num_devices);
+#else
+    INSPIRE_LOGW("CUDA is not supported, you need to enable the compile option that supports TensorRT");
+    return HERR_DEVICE_CUDA_DISABLE;
+#endif
+}
+
+HResult HFCheckCudaDeviceSupport(HPInt32 is_support) {
+#if defined(ISF_ENABLE_TENSORRT)
+    return inspire::CheckCudaUsability(is_support);
+#else
+    INSPIRE_LOGW("CUDA is not supported, you need to enable the compile option that supports TensorRT");
+    return HERR_DEVICE_CUDA_DISABLE;
+#endif
+}
+
+HResult HFFeatureHubDataEnable(HFFeatureHubConfiguration configuration) {
+    inspire::DatabaseConfiguration param;
+    if (configuration.primaryKeyMode != HF_PK_AUTO_INCREMENT && configuration.primaryKeyMode != HF_PK_MANUAL_INPUT) {
+        param.primary_key_mode = inspire::PrimaryKeyMode::AUTO_INCREMENT;
+    } else {
+        param.primary_key_mode = inspire::PrimaryKeyMode(configuration.primaryKeyMode);
+    }
+    if (configuration.persistenceDbPath == nullptr) {
+        INSPIRE_LOGE("persistenceDbPath is null, use default path");
+    }
+    // Add validation for persistenceDbPath
+    if (configuration.enablePersistence) {
+        if (configuration.persistenceDbPath == nullptr) {
+            param.persistence_db_path = std::string("");
+        } else {
+            param.persistence_db_path = std::string(configuration.persistenceDbPath);
+        }
+    } else {
+        param.persistence_db_path = std::string("");  // Empty string for in-memory mode
+    }
+    param.enable_persistence = configuration.enablePersistence;
+    param.recognition_threshold = configuration.searchThreshold;
+    param.search_mode = (inspire::SearchMode)configuration.searchMode;
+    auto ret = INSPIREFACE_FEATURE_HUB->EnableHub(param);
+    return ret;
+}
+
+HResult HFSessionSetTrackPreviewSize(HFSession session, HInt32 previewSize) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetTrackPreviewSize(previewSize);
+}
+
+HResult HFSessionGetTrackPreviewSize(HFSession session, HPInt32 previewSize) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    *previewSize = ctx->impl.GetTrackPreviewSize();
+    return HSUCCEED;
+}
+
+HResult HFSessionSetFilterMinimumFacePixelSize(HFSession session, HInt32 minSize) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetTrackFaceMinimumSize(minSize);
+}
+
+HResult HFSessionSetFaceTrackMode(HFSession session, HFDetectMode detectMode) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    inspire::DetectModuleMode detMode = inspire::DETECT_MODE_ALWAYS_DETECT;
+    if (detectMode == HF_DETECT_MODE_LIGHT_TRACK) {
+        detMode = inspire::DETECT_MODE_LIGHT_TRACK;
+    }
+    return ctx->impl.SetDetectMode(detMode);
+}
+
+HResult HFSessionSetFaceDetectThreshold(HFSession session, HFloat threshold) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetFaceDetectThreshold(threshold);
+}
+
+HResult HFSessionSetTrackModeSmoothRatio(HFSession session, HFloat ratio) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetTrackModeSmoothRatio(ratio);
+}
+
+HResult HFSessionSetTrackModeNumSmoothCacheFrame(HFSession session, HInt32 num) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetTrackModeNumSmoothCacheFrame(num);
+}
+
+HResult HFSessionSetTrackModeDetectInterval(HFSession session, HInt32 num) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    return ctx->impl.SetTrackModeDetectInterval(num);
+}
+
+HResult HFExecuteFaceTrack(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData results) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    auto ret = ctx->impl.FaceDetectAndTrack(stream->impl);
+    results->detectedNum = ctx->impl.GetNumberOfFacesCurrentlyDetected();
+    results->rects = (HFaceRect *)ctx->impl.GetFaceRectsCache().data();
+    results->trackIds = (HInt32 *)ctx->impl.GetTrackIDCache().data();
+    results->detConfidence = (HFloat *)ctx->impl.GetDetConfidenceCache().data();
+    results->trackCounts = (HInt32 *)ctx->impl.GetTrackCountCache().data();
+    results->angles.pitch = (HFloat *)ctx->impl.GetPitchResultsCache().data();
+    results->angles.roll = (HFloat *)ctx->impl.GetRollResultsCache().data();
+    results->angles.yaw = (HFloat *)ctx->impl.GetYawResultsCache().data();
+    results->tokens = (HFFaceBasicToken *)ctx->impl.GetFaceBasicDataCache().data();
+
+    return ret;
+}
+
+HResult HFSessionLastFaceDetectionGetDebugPreviewImageSize(HFSession session, HPInt32 size) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    *size = ctx->impl.GetDebugPreviewImageSize();
+    return HSUCCEED;
+}
+
+HResult HFCopyFaceBasicToken(HFFaceBasicToken token, HPBuffer buffer, HInt32 bufferSize) {
+    if (bufferSize < sizeof(inspire::FaceTrackWrap)) {
+        return HERR_INVALID_BUFFER_SIZE;
+    }
+    std::memcpy(buffer, token.data, sizeof(inspire::FaceTrackWrap));
+    return HSUCCEED;
+}
+
+HResult HFGetFaceBasicTokenSize(HPInt32 bufferSize) {
+    *bufferSize = sizeof(inspire::FaceTrackWrap);
+    return HSUCCEED;
+}
+
+HResult HFGetNumOfFaceDenseLandmark(HPInt32 num) {
+    *num = 106;
+    return HSUCCEED;
+}
+
+HResult HFGetFaceDenseLandmarkFromFaceToken(HFFaceBasicToken singleFace, PHPoint2f landmarks, HInt32 num) {
+    if (num != 106) {
+        return HERR_SESS_LANDMARK_NUM_NOT_MATCH;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    FaceTrackWrap face = {0};
+    HInt32 ret;
+    ret = RunDeserializeHyperFaceData((char *)data.data, data.dataSize, face);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    if (face.densityLandmarkEnable == HF_STATUS_DISABLE) {
+        INSPIRE_LOGW("To get dense landmarks in always-detect mode, you need to enable HF_ENABLE_DETECT_MODE_LANDMARK");
+        return HERR_SESS_LANDMARK_NOT_ENABLE;
+    }
+    for (size_t i = 0; i < num; i++) {
+        landmarks[i].x = face.densityLandmark[i].x;
+        landmarks[i].y = face.densityLandmark[i].y;
+    }
+    return HSUCCEED;
+}
+
+HResult HFGetFaceFiveKeyPointsFromFaceToken(HFFaceBasicToken singleFace, PHPoint2f landmarks, HInt32 num) {
+    if (num != 5) {
+        return HERR_SESS_KEY_POINT_NUM_NOT_MATCH;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    FaceTrackWrap face = {0};
+    HInt32 ret;
+    ret = RunDeserializeHyperFaceData((char *)data.data, data.dataSize, face);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    for (size_t i = 0; i < num; i++) {
+        landmarks[i].x = face.keyPoints[i].x;
+        landmarks[i].y = face.keyPoints[i].y;
+    }
+    return HSUCCEED;
+}
+
+HResult HFSessionSetEnableTrackCostSpend(HFSession session, HInt32 value) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    ctx->impl.SetEnableTrackCostSpend(value);
+    return HSUCCEED;
+}
+
+HResult HFSessionPrintTrackCostSpend(HFSession session) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    ctx->impl.PrintTrackCostSpend();
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubFaceSearchThresholdSetting(HFloat threshold) {
+    INSPIREFACE_FEATURE_HUB->SetRecognitionThreshold(threshold);
+    return HSUCCEED;
+}
+
+HResult HFFaceFeatureExtract(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace, PHFFaceFeature feature) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (singleFace.data == nullptr || singleFace.size <= 0) {
+        return HERR_INVALID_FACE_TOKEN;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    auto ret = ctx->impl.FaceFeatureExtract(stream->impl, data);
+    feature->size = ctx->impl.GetFaceFeatureCache().size();
+    feature->data = (HFloat *)ctx->impl.GetFaceFeatureCache().data();
+
+    return ret;
+}
+
+HResult HFFaceFeatureExtractTo(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace, HFFaceFeature feature) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (singleFace.data == nullptr || singleFace.size <= 0) {
+        return HERR_INVALID_FACE_TOKEN;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    auto ret = ctx->impl.FaceFeatureExtract(stream->impl, data);
+    for (int i = 0; i < ctx->impl.GetFaceFeatureCache().size(); ++i) {
+        feature.data[i] = ctx->impl.GetFaceFeatureCache()[i];
+    }
+
+    return HSUCCEED;
+}
+
+HResult HFFaceFeatureExtractCpy(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace, HPFloat feature) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (singleFace.data == nullptr || singleFace.size <= 0) {
+        return HERR_INVALID_FACE_TOKEN;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    auto ret = ctx->impl.FaceFeatureExtract(stream->impl, data);
+    for (int i = 0; i < ctx->impl.GetFaceFeatureCache().size(); ++i) {
+        feature[i] = ctx->impl.GetFaceFeatureCache()[i];
+    }
+
+    return ret;
+}
+
+HResult HFCreateFaceFeature(PHFFaceFeature feature) {
+    if (feature == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    feature->size = FACE_FEATURE_SIZE;
+    feature->data = new HFloat[FACE_FEATURE_SIZE];
+    RESOURCE_MANAGE->createFaceFeature((long)feature);
+    return HSUCCEED;
+}
+
+HResult HFReleaseFaceFeature(PHFFaceFeature feature) {
+    if (feature == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    delete[] feature->data;
+    RESOURCE_MANAGE->releaseFaceFeature((long)feature);
+    return HSUCCEED;
+}
+
+HResult HFFaceGetFaceAlignmentImage(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace, PHFImageBitmap handle) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (singleFace.data == nullptr || singleFace.size <= 0) {
+        return HERR_INVALID_FACE_TOKEN;
+    }
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+    auto bitmap = new HF_ImageBitmap();
+    auto ret = ctx->impl.FaceGetFaceAlignmentImage(stream->impl, data, bitmap->impl);
+    if (ret != HSUCCEED) {
+        delete bitmap;
+        return ret;
+    }
+    *handle = bitmap;
+    // Record the creation of this image bitmap in the ResourceManager
+    RESOURCE_MANAGE->createImageBitmap((long)*handle);
+    return HSUCCEED;
+}
+
+HResult HFFaceFeatureExtractWithAlignmentImage(HFSession session, HFImageStream streamHandle, HFFaceFeature feature) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    Embedded embedded;
+    float norm;
+    auto ret = ctx->impl.FaceRecognitionModule()->FaceExtractWithAlignmentImage(stream->impl, embedded, norm);
+    for (int i = 0; i < embedded.size(); ++i) {
+        feature.data[i] = embedded[i];
+    }
+    return ret;
+}
+
+HResult HFFaceComparison(HFFaceFeature feature1, HFFaceFeature feature2, HPFloat result) {
+    if (feature1.data == nullptr || feature2.data == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    if (feature1.size != feature2.size) {
+        INSPIRE_LOGE("feature1.size: %d, feature2.size: %d", feature1.size, feature2.size);
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    *result = 0.0f;
+    float res = -1.0f;
+    auto ret = INSPIREFACE_FEATURE_HUB->CosineSimilarity(feature1.data, feature2.data, feature1.size, res);
+    *result = res;
+
+    return ret;
+}
+
+HResult HFGetRecommendedCosineThreshold(HPFloat threshold) {
+    if (!INSPIREFACE_CONTEXT->isMLoad()) {
+        INSPIRE_LOGW("Inspireface is not launched, using default threshold 0.48");
+    }
+    *threshold = SIMILARITY_CONVERTER_GET_RECOMMENDED_COSINE_THRESHOLD();
+    return HSUCCEED;
+}
+
+HResult HFCosineSimilarityConvertToPercentage(HFloat similarity, HPFloat result) {
+    if (!INSPIREFACE_CONTEXT->isMLoad()) {
+        INSPIRE_LOGW("Inspireface is not launched.");
+    }
+    *result = SIMILARITY_CONVERTER_RUN(similarity);
+    return HSUCCEED;
+}
+
+HResult HFUpdateCosineSimilarityConverter(HFSimilarityConverterConfig config) {
+    if (!INSPIREFACE_CONTEXT->isMLoad()) {
+        INSPIRE_LOGW("Inspireface is not launched.");
+    }
+    inspire::SimilarityConverterConfig cfg;
+    cfg.threshold = config.threshold;
+    cfg.middleScore = config.middleScore;
+    cfg.steepness = config.steepness;
+    cfg.outputMin = config.outputMin;
+    cfg.outputMax = config.outputMax;
+    SIMILARITY_CONVERTER_UPDATE_CONFIG(cfg);
+    return HSUCCEED;
+}
+
+HResult HFGetCosineSimilarityConverter(PHFSimilarityConverterConfig config) {
+    if (!INSPIREFACE_CONTEXT->isMLoad()) {
+        INSPIRE_LOGW("Inspireface is not launched.");
+    }
+    inspire::SimilarityConverterConfig cfg = SIMILARITY_CONVERTER_GET_CONFIG();
+    config->threshold = cfg.threshold;
+    config->middleScore = cfg.middleScore;
+    config->steepness = cfg.steepness;
+    config->outputMin = cfg.outputMin;
+    config->outputMax = cfg.outputMax;
+    return HSUCCEED;
+}
+
+HResult HFGetFeatureLength(HPInt32 num) {
+    *num = FACE_FEATURE_SIZE;
+
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubInsertFeature(HFFaceFeatureIdentity featureIdentity, HPFaceId allocId) {
+    if (featureIdentity.feature->data == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    std::vector<float> feat;
+    feat.reserve(featureIdentity.feature->size);
+    for (int i = 0; i < featureIdentity.feature->size; ++i) {
+        feat.push_back(featureIdentity.feature->data[i]);
+    }
+    HInt32 ret = INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(feat, featureIdentity.id, *allocId);
+
+    return ret;
+}
+
+HResult HFFeatureHubFaceSearch(HFFaceFeature searchFeature, HPFloat confidence, PHFFaceFeatureIdentity mostSimilar) {
+    if (searchFeature.data == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    std::vector<float> feat;
+    feat.reserve(searchFeature.size);
+    for (int i = 0; i < searchFeature.size; ++i) {
+        feat.push_back(searchFeature.data[i]);
+    }
+    *confidence = -1.0f;
+    inspire::FaceSearchResult result;
+    HInt32 ret = INSPIREFACE_FEATURE_HUB->SearchFaceFeature(feat, result);
+    mostSimilar->feature = (HFFaceFeature *)INSPIREFACE_FEATURE_HUB->GetFaceFeaturePtrCache().get();
+    mostSimilar->feature->data = (HFloat *)INSPIREFACE_FEATURE_HUB->GetSearchFaceFeatureCache().data();
+    mostSimilar->feature->size = INSPIREFACE_FEATURE_HUB->GetSearchFaceFeatureCache().size();
+    mostSimilar->id = result.id;
+    if (mostSimilar->id != -1) {
+        *confidence = result.similarity;
+    }
+
+    return ret;
+}
+
+HResult HFFeatureHubFaceSearchTopK(HFFaceFeature searchFeature, HInt32 topK, PHFSearchTopKResults results) {
+    if (searchFeature.data == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    std::vector<float> feat;
+    feat.reserve(searchFeature.size);
+    for (int i = 0; i < searchFeature.size; ++i) {
+        feat.push_back(searchFeature.data[i]);
+    }
+    HInt32 ret = INSPIREFACE_FEATURE_HUB->SearchFaceFeatureTopKCache(feat, topK);
+    if (ret == HSUCCEED) {
+        results->size = INSPIREFACE_FEATURE_HUB->GetTopKConfidence().size();
+        results->confidence = INSPIREFACE_FEATURE_HUB->GetTopKConfidence().data();
+        results->ids = INSPIREFACE_FEATURE_HUB->GetTopKCustomIdsCache().data();
+    }
+
+    return ret;
+}
+
+HResult HFFeatureHubFaceRemove(HFaceId id) {
+    auto ret = INSPIREFACE_FEATURE_HUB->FaceFeatureRemove(id);
+    return ret;
+}
+
+HResult HFFeatureHubFaceUpdate(HFFaceFeatureIdentity featureIdentity) {
+    if (featureIdentity.feature->data == nullptr) {
+        return HERR_INVALID_FACE_FEATURE;
+    }
+    std::vector<float> feat;
+    feat.reserve(featureIdentity.feature->size);
+    for (int i = 0; i < featureIdentity.feature->size; ++i) {
+        feat.push_back(featureIdentity.feature->data[i]);
+    }
+
+    auto ret = INSPIREFACE_FEATURE_HUB->FaceFeatureUpdate(feat, featureIdentity.id);
+
+    return ret;
+}
+
+HResult HFFeatureHubGetFaceIdentity(HFaceId id, PHFFaceFeatureIdentity identity) {
+    auto ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(id);
+    if (ret == HSUCCEED) {
+        identity->id = id;
+        identity->feature = (HFFaceFeature *)INSPIREFACE_FEATURE_HUB->GetFaceFeaturePtrCache().get();
+        identity->feature->data = (HFloat *)INSPIREFACE_FEATURE_HUB->GetFaceFeaturePtrCache()->data;
+        identity->feature->size = INSPIREFACE_FEATURE_HUB->GetFaceFeaturePtrCache()->dataSize;
+    } else {
+        identity->id = -1;
+    }
+
+    return ret;
+}
+
+HResult HFMultipleFacePipelineProcess(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData faces, HFSessionCustomParameter parameter) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (faces->detectedNum == 0) {
+        return HSUCCEED;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (faces == nullptr || faces->tokens == nullptr || faces->tokens->data == nullptr) {
+        return HERR_INVALID_FACE_LIST;
+    }
+
+    inspire::ContextCustomParameter param;
+    param.enable_mask_detect = parameter.enable_mask_detect;
+    param.enable_face_attribute = parameter.enable_face_quality;
+    param.enable_liveness = parameter.enable_liveness;
+    param.enable_face_quality = parameter.enable_face_quality;
+    param.enable_interaction_liveness = parameter.enable_interaction_liveness;
+    param.enable_ir_liveness = parameter.enable_ir_liveness;
+    param.enable_recognition = parameter.enable_recognition;
+    param.enable_face_attribute = parameter.enable_face_attribute;
+    param.enable_face_emotion = parameter.enable_face_emotion;
+
+    HResult ret;
+    std::vector<inspire::FaceTrackWrap> data;
+    data.resize(faces->detectedNum);
+    for (int i = 0; i < faces->detectedNum; ++i) {
+        auto &face = data[i];
+        ret = RunDeserializeHyperFaceData((char *)faces->tokens[i].data, faces->tokens[i].size, face);
+        if (ret != HSUCCEED) {
+            return HERR_INVALID_FACE_TOKEN;
+        }
+    }
+
+    ret = ctx->impl.FacesProcess(stream->impl, data, param);
+
+    return ret;
+}
+
+HResult HFMultipleFacePipelineProcessOptional(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData faces, HInt32 customOption) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    if (streamHandle == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (faces->detectedNum == 0) {
+        return HSUCCEED;
+    }
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_CameraStream *stream = (HF_CameraStream *)streamHandle;
+    if (stream == nullptr) {
+        return HERR_INVALID_IMAGE_STREAM_HANDLE;
+    }
+    if (faces == nullptr || faces->tokens == nullptr || faces->tokens->data == nullptr) {
+        return HERR_INVALID_FACE_LIST;
+    }
+
+    inspire::ContextCustomParameter param;
+    if (customOption & HF_ENABLE_FACE_RECOGNITION) {
+        param.enable_recognition = true;
+    }
+    if (customOption & HF_ENABLE_LIVENESS) {
+        param.enable_liveness = true;
+    }
+    if (customOption & HF_ENABLE_IR_LIVENESS) {
+        param.enable_ir_liveness = true;
+    }
+    if (customOption & HF_ENABLE_FACE_ATTRIBUTE) {
+        param.enable_face_attribute = true;
+    }
+    if (customOption & HF_ENABLE_MASK_DETECT) {
+        param.enable_mask_detect = true;
+    }
+    if (customOption & HF_ENABLE_QUALITY) {
+        param.enable_face_quality = true;
+    }
+    if (customOption & HF_ENABLE_INTERACTION) {
+        param.enable_interaction_liveness = true;
+    }
+    if (customOption & HF_ENABLE_FACE_POSE) {
+        param.enable_face_pose = true;
+    }
+    if (customOption & HF_ENABLE_FACE_EMOTION) {
+        param.enable_face_emotion = true;
+    }
+
+    HResult ret;
+    std::vector<inspire::FaceTrackWrap> data;
+    data.resize(faces->detectedNum);
+    for (int i = 0; i < faces->detectedNum; ++i) {
+        auto &face = data[i];
+        ret = RunDeserializeHyperFaceData((char *)faces->tokens[i].data, faces->tokens[i].size, face);
+        if (ret != HSUCCEED) {
+            return HERR_INVALID_FACE_TOKEN;
+        }
+    }
+
+    ret = ctx->impl.FacesProcess(stream->impl, data, param);
+
+    return ret;
+}
+
+HResult HFGetRGBLivenessConfidence(HFSession session, PHFRGBLivenessConfidence confidence) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    confidence->num = ctx->impl.GetRgbLivenessResultsCache().size();
+    confidence->confidence = (HFloat *)ctx->impl.GetRgbLivenessResultsCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFGetFaceMaskConfidence(HFSession session, PHFFaceMaskConfidence confidence) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    confidence->num = ctx->impl.GetMaskResultsCache().size();
+    confidence->confidence = (HFloat *)ctx->impl.GetMaskResultsCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFGetFaceQualityConfidence(HFSession session, PHFFaceQualityConfidence confidence) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    confidence->num = ctx->impl.GetFaceQualityScoresResultsCache().size();
+    confidence->confidence = (HFloat *)ctx->impl.GetFaceQualityScoresResultsCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFFaceQualityDetect(HFSession session, HFFaceBasicToken singleFace, HPFloat confidence) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    inspire::FaceBasicData data;
+    data.dataSize = singleFace.size;
+    data.data = singleFace.data;
+
+    auto ret = inspire::FaceSession::FaceQualityDetect(data, *confidence);
+
+    return ret;
+}
+
+HResult HFGetFaceInteractionStateResult(HFSession session, PHFFaceInteractionState result) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    result->num = ctx->impl.GetFaceInteractionLeftEyeStatusCache().size();
+    result->leftEyeStatusConfidence = (HFloat *)ctx->impl.GetFaceInteractionLeftEyeStatusCache().data();
+    result->rightEyeStatusConfidence = (HFloat *)ctx->impl.GetFaceInteractionRightEyeStatusCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFGetFaceInteractionActionsResult(HFSession session, PHFFaceInteractionsActions actions) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    actions->num = ctx->impl.GetFaceNormalAactionsResultCache().size();
+    actions->normal = (HInt32 *)ctx->impl.GetFaceNormalAactionsResultCache().data();
+    actions->blink = (HInt32 *)ctx->impl.GetFaceBlinkAactionsResultCache().data();
+    actions->shake = (HInt32 *)ctx->impl.GetFaceShakeAactionsResultCache().data();
+    actions->headRaise = (HInt32 *)ctx->impl.GetFaceRaiseHeadAactionsResultCache().data();
+    actions->jawOpen = (HInt32 *)ctx->impl.GetFaceJawOpenAactionsResultCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFGetFaceAttributeResult(HFSession session, PHFFaceAttributeResult results) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    results->num = ctx->impl.GetFaceAgeBracketResultsCache().size();
+    results->race = (HPInt32)ctx->impl.GetFaceRaceResultsCache().data();
+    results->gender = (HPInt32)ctx->impl.GetFaceGenderResultsCache().data();
+    results->ageBracket = (HPInt32)ctx->impl.GetFaceAgeBracketResultsCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFGetFaceEmotionResult(HFSession session, PHFFaceEmotionResult result) {
+    if (session == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+    HF_FaceAlgorithmSession *ctx = (HF_FaceAlgorithmSession *)session;
+    if (ctx == nullptr) {
+        return HERR_INVALID_CONTEXT_HANDLE;
+    }
+
+    result->num = ctx->impl.GetFaceEmotionResultsCache().size();
+    result->emotion = (HPInt32)ctx->impl.GetFaceEmotionResultsCache().data();
+
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubGetFaceCount(HPInt32 count) {
+    *count = INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount();
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubViewDBTable() {
+    INSPIREFACE_FEATURE_HUB->ViewDBTable();
+    return HSUCCEED;
+}
+
+HResult HFFeatureHubGetExistingIds(PHFFeatureHubExistingIds ids) {
+    auto ret = INSPIREFACE_FEATURE_HUB->GetAllIds();
+    if (ret == HSUCCEED) {
+        ids->size = INSPIREFACE_FEATURE_HUB->GetExistingIds().size();
+        ids->ids = INSPIREFACE_FEATURE_HUB->GetExistingIds().data();
+    }
+    return ret;
+}
+
+HResult HFQueryInspireFaceVersion(PHFInspireFaceVersion version) {
+    version->major = atoi(INSPIRE_FACE_VERSION_MAJOR_STR);
+    version->minor = atoi(INSPIRE_FACE_VERSION_MINOR_STR);
+    version->patch = atoi(INSPIRE_FACE_VERSION_PATCH_STR);
+
+    return HSUCCEED;
+}
+
+HResult HFQueryInspireFaceExtendedInformation(PHFInspireFaceExtendedInformation information) {
+    strncpy(information->information, INSPIRE_FACE_EXTENDED_INFORMATION, strlen(INSPIRE_FACE_EXTENDED_INFORMATION));
+    return HSUCCEED;
+}
+
+HResult HFSetLogLevel(HFLogLevel level) {
+    INSPIRE_SET_LOG_LEVEL(LogLevel(level));
+    return HSUCCEED;
+}
+
+HResult HFLogDisable() {
+    INSPIRE_SET_LOG_LEVEL(inspire::ISF_LOG_NONE);
+
+    return HSUCCEED;
+}
+
+HResult HFLogPrint(HFLogLevel level, HFormat format, ...) {
+    inspire::LogLevel logLevel = static_cast<inspire::LogLevel>(level);
+    if (inspire::LogManager::getInstance()->getLogLevel() == inspire::ISF_LOG_NONE || logLevel < inspire::LogManager::getInstance()->getLogLevel()) {
+        return HSUCCEED;
+    }
+    char buffer[1024];
+    va_list args;
+    va_start(args, format);
+    vsnprintf(buffer, sizeof(buffer), format, args);
+    va_end(args);
+
+    switch (logLevel) {
+        case inspire::ISF_LOG_DEBUG:
+            INSPIRE_LOGD("%s", buffer);
+            break;
+        case inspire::ISF_LOG_INFO:
+            INSPIRE_LOGI("%s", buffer);
+            break;
+        case inspire::ISF_LOG_WARN:
+            INSPIRE_LOGW("%s", buffer);
+            break;
+        case inspire::ISF_LOG_ERROR:
+            INSPIRE_LOGE("%s", buffer);
+            break;
+        case inspire::ISF_LOG_FATAL:
+            INSPIRE_LOGF("%s", buffer);
+            break;
+        default:
+            break;
+    }
+
+    return HSUCCEED;
+}
+
+HResult HFDeBugShowResourceStatistics() {
+    RESOURCE_MANAGE->printResourceStatistics();
+    return HSUCCEED;
+}
+
+HResult HFDeBugGetUnreleasedSessionsCount(HPInt32 count) {
+    *count = RESOURCE_MANAGE->getUnreleasedSessions().size();
+    return HSUCCEED;
+}
+
+HResult HFDeBugGetUnreleasedSessions(PHFSession sessions, HInt32 count) {
+    std::vector<long> unreleasedSessions = RESOURCE_MANAGE->getUnreleasedSessions();
+    for (int i = 0; i < count; ++i) {
+        sessions[i] = (HFSession)unreleasedSessions[i];
+    }
+    return HSUCCEED;
+}
+
+HResult HFDeBugGetUnreleasedStreamsCount(HPInt32 count) {
+    *count = RESOURCE_MANAGE->getUnreleasedStreams().size();
+    return HSUCCEED;
+}
+
+HResult HFDeBugGetUnreleasedStreams(PHFImageStream streams, HInt32 count) {
+    std::vector<long> unreleasedStreams = RESOURCE_MANAGE->getUnreleasedStreams();
+    for (int i = 0; i < count; ++i) {
+        streams[i] = (HFImageStream)unreleasedStreams[i];
+    }
+    return HSUCCEED;
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.h b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.h
new file mode 100644
index 000000000..5038fb318
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface.h
@@ -0,0 +1,1499 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_H
+#define INSPIREFACE_H
+
+#include <stdint.h>
+#include "intypedef.h"
+#include "herror.h"
+
+#if defined(_WIN32)
+#ifdef ISF_BUILD_SHARED_LIBS
+#define HYPER_CAPI_EXPORT __declspec(dllexport)
+#else
+#define HYPER_CAPI_EXPORT
+#endif
+#else
+#define HYPER_CAPI_EXPORT __attribute__((visibility("default")))
+#endif  // _WIN32
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define HF_STATUS_ENABLE 1   ///< The status of the feature is enabled.
+#define HF_STATUS_DISABLE 0  ///< The status of the feature is disabled.
+
+#define HF_ENABLE_NONE 0x00000000              ///< Flag to enable no features.
+#define HF_ENABLE_FACE_RECOGNITION 0x00000002  ///< Flag to enable face recognition feature.
+#define HF_ENABLE_LIVENESS 0x00000004          ///< Flag to enable RGB liveness detection feature.
+#define HF_ENABLE_IR_LIVENESS 0x00000008       ///< Flag to enable IR (Infrared) liveness detection feature.
+#define HF_ENABLE_MASK_DETECT 0x00000010       ///< Flag to enable mask detection feature.
+#define HF_ENABLE_FACE_ATTRIBUTE 0x00000020    ///< Flag to enable face attribute prediction feature.
+#define HF_ENABLE_PLACEHOLDER_ 0x00000040      ///< -
+#define HF_ENABLE_QUALITY 0x00000080           ///< Flag to enable face quality assessment feature.
+#define HF_ENABLE_INTERACTION 0x00000100       ///< Flag to enable interaction feature.
+#define HF_ENABLE_FACE_POSE 0x00000200         ///< Flag to enable face pose estimation feature.
+#define HF_ENABLE_FACE_EMOTION 0x00000400      ///< Flag to enable face emotion recognition feature.
+
+/************************************************************************
+ * Image Stream Function
+ *
+ * ImageStream directly interacts with algorithm modules, providing image data streams for algorithm module input;
+ * ImageStream provides automatic transformation to adapt camera stream rotation angles and common image encoding/decoding format conversion;
+ * Camera picture rotation mode.
+ * To accommodate the rotation of certain devices, four image rotation modes are provided.
+ *
+ * 1. ROTATION_0 (No Rotation):
+ *    Original Image (w x h):               Scaled Image (w*s x h*s):
+ *    A(0,0) ----------- B(w-1,0)          A(0,0) ----------- B(w*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Scaled             |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        C(0,h*s-1) ---- D(w*s-1,h*s-1)
+ *    Point Mapping: A->A(0,0), B->B(w*s-1,0), C->C(0,h*s-1), D->D(w*s-1,h*s-1)
+ *
+ * 2. ROTATION_90 (90° Counter-Clockwise):
+ *    Original Image (w x h):               Rotated Image (h*s x w*s):
+ *    A(0,0) ----------- B(w-1,0)          B(0,0) ----------- A(h*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        D(0,w*s-1) ---- C(h*s-1,w*s-1)
+ *    Point Mapping: A->A(h*s-1,0), B->B(0,0), C->C(h*s-1,w*s-1), D->D(0,w*s-1)
+ *
+ * 3. ROTATION_180 (180° Rotation):
+ *    Original Image (w x h):               Rotated Image (w*s x h*s):
+ *    A(0,0) ----------- B(w-1,0)          D(0,0) ----------- C(w*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        B(0,h*s-1) ---- A(w*s-1,h*s-1)
+ *    Point Mapping: A->A(w*s-1,h*s-1), B->B(0,h*s-1), C->C(w*s-1,0), D->D(0,0)
+ *
+ * 4. ROTATION_270 (270° Counter-Clockwise):
+ *    Original Image (w x h):               Rotated Image (h*s x w*s):
+ *    A(0,0) ----------- B(w-1,0)          D(0,0) ----------- C(h*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        B(0,w*s-1) ---- A(h*s-1,w*s-1)
+ *    Point Mapping: A->A(h*s-1,w*s-1), B->B(0,w*s-1), C->C(h*s-1,0), D->D(0,0)
+ ************************************************************************/
+
+/**
+ * Camera stream format.
+ * Contains several common camera stream formats available in the market.
+ */
+typedef enum HFImageFormat {
+    HF_STREAM_RGB = 0,       ///< Image in RGB format.
+    HF_STREAM_BGR = 1,       ///< Image in BGR format (Opencv Mat default).
+    HF_STREAM_RGBA = 2,      ///< Image in RGB format with alpha channel.
+    HF_STREAM_BGRA = 3,      ///< Image in BGR format with alpha channel.
+    HF_STREAM_YUV_NV12 = 4,  ///< Image in YUV NV12 format.
+    HF_STREAM_YUV_NV21 = 5,  ///< Image in YUV NV21 format.
+    HF_STREAM_I420 = 6,      ///< Image in I420 format.
+    HF_STREAM_GRAY = 7,      ///< Image in GRAY format.
+} HFImageFormat;
+
+/**
+ * Camera picture rotation mode.
+ * To accommodate the rotation of certain devices, four image rotation modes are provided.
+ */
+typedef enum HFRotation {
+    HF_CAMERA_ROTATION_0 = 0,    ///< 0 degree rotation.
+    HF_CAMERA_ROTATION_90 = 1,   ///< 90 degree rotation.
+    HF_CAMERA_ROTATION_180 = 2,  ///< 180 degree rotation.
+    HF_CAMERA_ROTATION_270 = 3,  ///< 270 degree rotation.
+} HFRotation;
+
+/**
+ * Image Buffer Data structure.
+ * Defines the structure for image data stream.
+ */
+typedef struct HFImageData {
+    HPUInt8 data;          ///< Pointer to the image data stream.
+    HInt32 width;          ///< Width of the image.
+    HInt32 height;         ///< Height of the image.
+    HFImageFormat format;  ///< Format of the image, indicating the data stream format to be parsed.
+    HFRotation rotation;   ///< Rotation angle of the image.
+} HFImageData, *PHFImageData;
+
+/**
+ * @brief Create a data buffer stream instantiation object.
+ *
+ * This function is used to create an instance of a data buffer stream with the given image data.
+ *
+ * @param data Pointer to the image buffer data structure.
+ * @param handle Pointer to the stream handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStream(PHFImageData data, PHFImageStream handle);
+
+/**
+ * @brief Create an empty image stream instance.
+ *
+ * This function is used to create an instance of a data buffer stream with the given image data.
+ *
+ * @param handle Pointer to the stream handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStreamEmpty(PHFImageStream handle);
+
+/**
+ * @brief Set the buffer of the image stream.
+ *
+ * @param handle Pointer to the stream handle.
+ * @param buffer Pointer to the buffer.
+ * @param width Width of the image.
+ * @param height Height of the image.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetBuffer(HFImageStream handle, HPUInt8 buffer, HInt32 width, HInt32 height);
+
+/**
+ * @brief Set the rotation of the image stream.
+ *
+ * @param handle Pointer to the stream handle.
+ * @param rotation Rotation angle of the image.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetRotation(HFImageStream handle, HFRotation rotation);
+
+/**
+ * @brief Set the format of the image stream.
+ *
+ * @param handle Pointer to the stream handle.
+ * @param format Format of the image.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageStreamSetFormat(HFImageStream handle, HFImageFormat format);
+
+/**
+ * @brief Release the instantiated DataBuffer object.
+ *
+ * This function is used to release the DataBuffer object that has been previously instantiated.
+ *
+ * @param streamHandle Pointer to the DataBuffer handle representing the camera stream component.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFReleaseImageStream(HFImageStream streamHandle);
+
+/************************************************************************
+ * Image Bitmap Function
+ *
+ * Provides a simple Bitmap interface wrapper that copies image data when creating objects and requires manual release.
+ * Provides interfaces for copying, drawing, displaying (OpenCV-GUI), writing to files, and converting to/from ImageStream.
+ ************************************************************************/
+
+/**
+ * @brief Struct for image bitmap data.
+ */
+typedef struct HFImageBitmapData {
+    HPUInt8 data;     ///< Pointer to the image data.
+    HInt32 width;     ///< Width of the image.
+    HInt32 height;    ///< Height of the image.
+    HInt32 channels;  ///< Number of channels in the image, only support 3 channels or 1 channel.
+} HFImageBitmapData, *PHFImageBitmapData;
+
+/**
+ * @brief Create a image bitmap from data, default pixel format is BGR.
+ *
+ * @param data Pointer to the image bitmap data structure.
+ * @param handle Pointer to the image bitmap handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmap(PHFImageBitmapData data, PHFImageBitmap handle);
+
+/**
+ * @brief Create a image bitmap from file path, default pixel format is BGR.
+ *
+ * @param filePath The path to the image file.
+ * @param channels The number of channels in the image, only support 3 channels or 1 channel.
+ * @param handle Pointer to the image bitmap handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmapFromFilePath(HPath filePath, HInt32 channels, PHFImageBitmap handle);
+
+/**
+ * @brief Copy an image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param copyHandle Pointer to the image bitmap handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapCopy(HFImageBitmap handle, PHFImageBitmap copyHandle);
+
+/**
+ * @brief Release the image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFReleaseImageBitmap(HFImageBitmap handle);
+
+/**
+ * @brief Create a image stream from image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param rotation The rotation angle of the image.
+ * @param streamHandle Pointer to the image stream handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageStreamFromImageBitmap(HFImageBitmap handle, HFRotation rotation, PHFImageStream streamHandle);
+
+/**
+ * @brief Create a image bitmap from image stream.
+ *
+ * @param streamHandle Pointer to the image stream handle.
+ * @param handle Pointer to the image bitmap handle that will be returned.
+ * @param is_rotate Whether to rotate the image.
+ * @param scale The scale of the image.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateImageBitmapFromImageStreamProcess(HFImageStream streamHandle, PHFImageBitmap handle, HInt32 is_rotate,
+                                                                           HFloat scale);
+
+/**
+ * @brief Write the image bitmap to a file.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param filePath The path to the image file.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapWriteToFile(HFImageBitmap handle, HPath filePath);
+
+/**
+ * @brief Draw a rectangle on the image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param rect The rectangle to be drawn.
+ * @param color The color of the rectangle.
+ * @param thickness The thickness of the rectangle.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawRect(HFImageBitmap handle, HFaceRect rect, HColor color, HInt32 thickness);
+
+/**
+ * @brief Draw a circle on the image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param point The center point of the circle.
+ * @param radius The radius of the circle.
+ * @param color The color of the circle.
+ * @param thickness The thickness of the circle.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawCircleF(HFImageBitmap handle, HPoint2f point, HInt32 radius, HColor color, HInt32 thickness);
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapDrawCircle(HFImageBitmap handle, HPoint2i point, HInt32 radius, HColor color, HInt32 thickness);
+
+/**
+ * @brief Get the data of the image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle.
+ * @param data Pointer to the image bitmap data structure.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapGetData(HFImageBitmap handle, PHFImageBitmapData data);
+
+/**
+ * @brief Show the image bitmap.
+ *
+ * @param handle Pointer to the image bitmap handle, must rely on opencv's gui functionality
+ * @param title The title of the image.
+ * @param delay The delay time of the image.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFImageBitmapShow(HFImageBitmap handle, HString title, HInt32 delay);
+
+/************************************************************************
+ * Resource Function
+ *
+ * The resource module is a system-level module that manages the life cycle of all resources.
+ * It is responsible for loading and unloading resources, and managing the memory of resources.
+ ************************************************************************/
+
+/**
+ * @brief Launch InspireFace SDK
+ * Start the InspireFace SDK at the initialization stage of your program, as it is global and
+ * designed to be used only once. It serves as a prerequisite for other function interfaces, so it
+ * is essential to ensure it is initialized before calling any other APIs.
+ * @param resourcePath Initializes the path to the resource file that needs to be loaded
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFLaunchInspireFace(HPath resourcePath);
+
+/**
+ * @brief Reload InspireFace SDK
+ * Reload the InspireFace SDK, releasing all allocated resources.
+ * @param resourcePath Initializes the path to the resource file that needs to be loaded
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFReloadInspireFace(HPath resourcePath);
+
+/**
+ * @brief Terminate InspireFace SDK
+ * Terminate the InspireFace SDK, releasing all allocated resources.
+ * This should be called at the end of your program to ensure proper cleanup.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFTerminateInspireFace();
+
+/**
+ * @brief Query InspireFace SDK launch status
+ * Query the launch status of the InspireFace SDK.
+ * @param status Pointer to the status variable that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFQueryInspireFaceLaunchStatus(HPInt32 status);
+
+/************************************************************************
+ * Extended Interface Based on Third-party Hardware Devices
+ *
+ * According to different manufacturers' devices, manufacturers typically perform deep customization and optimization, such as neural network
+ * inference computation, geometric image acceleration computation, and deeply customized device interfaces, etc. These types of functionalities are
+ * usually difficult to abstract, so they are placed in extension module APIs, involving hybrid computing, heterogeneous computing, multi-device
+ * computing, and other features.
+ ************************************************************************/
+
+/**
+ * @brief Check whether RGA is enabled during compilation
+ * @param status Pointer to the status variable that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFQueryExpansiveHardwareRGACompileOption(HPInt32 enable);
+
+/**
+ * @brief Set the rockchip dma heap path
+ * By default, we have already configured the DMA Heap address used by RGA on RK devices.
+ * If you wish to customize this address, you can modify it through this API.
+ * @param path The path to the rockchip dma heap
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSetExpansiveHardwareRockchipDmaHeapPath(HPath path);
+
+/**
+ * @brief Query the rockchip dma heap path
+ * @param path Pointer to a pre-allocated character array that will store the returned path.
+ * The array should be at least 256 bytes in size.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFQueryExpansiveHardwareRockchipDmaHeapPath(HString path);
+
+/**
+ * @brief Enum for image processing backend.
+ */
+typedef enum HFImageProcessingBackend {
+    HF_IMAGE_PROCESSING_CPU = 0,  ///< CPU backend(Default)
+    HF_IMAGE_PROCESSING_RGA = 1,  ///< Rockchip RGA backend(Hardware support is mandatory)
+} HFImageProcessingBackend;
+
+/**
+ * @brief Switch the image processing backend, must be called before HFCreateInspireFaceSession.
+ * @param backend The image processing backend to be set.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSwitchImageProcessingBackend(HFImageProcessingBackend backend);
+
+/**
+ * @brief Set the image process aligned width, must be called before HFCreateInspireFaceSession.
+ * @param width The image process aligned width to be set.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSetImageProcessAlignedWidth(HInt32 width);
+
+/**
+ * @brief Enum for Apple CoreML inference mode.
+ */
+typedef enum HFAppleCoreMLInferenceMode {
+    HF_APPLE_COREML_INFERENCE_MODE_CPU = 0,  ///< CPU Only.
+    HF_APPLE_COREML_INFERENCE_MODE_GPU = 1,  ///< GPU first.
+    HF_APPLE_COREML_INFERENCE_MODE_ANE = 2,  ///< Automatic selection, ANE first.
+} HFAppleCoreMLInferenceMode;
+
+/**
+ * @brief Set the Apple CoreML inference mode, must be called before HFCreateInspireFaceSession.
+ * @param mode The inference mode to be set.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSetAppleCoreMLInferenceMode(HFAppleCoreMLInferenceMode mode);
+
+/**
+ * @brief Set the CUDA device id, must be called before HFCreateInspireFaceSession.
+ * @param device_id The device id to be set.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSetCudaDeviceId(HInt32 device_id);
+
+/**
+ * @brief Get the CUDA device id, must be called after HFCreateInspireFaceSession.
+ * @param device_id Pointer to the device id to be returned.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFGetCudaDeviceId(HPInt32 device_id);
+
+/**
+ * @brief Print the CUDA device information.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFPrintCudaDeviceInfo();
+
+/**
+ * @brief Get the number of CUDA devices.
+ * @param num_devices Pointer to the number of CUDA devices to be returned.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFGetNumCudaDevices(HPInt32 num_devices);
+
+/**
+ * @brief Check if the CUDA device is supported.
+ * @param support The support flag to be checked.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFCheckCudaDeviceSupport(HPInt32 is_support);
+
+/************************************************************************
+ * FaceSession Function
+ *
+ * FaceSession is responsible for all face image algorithm-related functions,
+ * including face detection, face alignment, face recognition, face quality detection, face attribute prediction, etc.
+ * FaceSession supports flexible configuration, allowing you to enable or disable certain functions, and also set parameters for certain functions.
+ * In concurrent scenarios, multiple sessions can be created, each session can run independently without interfering with each other.
+ ************************************************************************/
+
+/**
+ * @brief Struct for custom parameters in face recognition context.
+ *
+ * This struct holds various flags to enable or disable specific features
+ * in the face recognition context, such as face recognition, liveness detection,
+ * mask detection, age and gender prediction, etc.
+ */
+typedef struct HFSessionCustomParameter {
+    HInt32 enable_recognition;           ///< Enable face recognition feature.
+    HInt32 enable_liveness;              ///< Enable RGB liveness detection feature.
+    HInt32 enable_ir_liveness;           ///< Enable IR liveness detection feature.
+    HInt32 enable_mask_detect;           ///< Enable mask detection feature.
+    HInt32 enable_face_quality;          ///< Enable face quality detection feature.
+    HInt32 enable_face_attribute;        ///< Enable face attribute prediction feature.
+    HInt32 enable_interaction_liveness;  ///< Enable interaction for liveness detection feature.
+    HInt32 enable_detect_mode_landmark;  ///< Enable landmark detection in detection mode
+    HInt32 enable_face_pose;             ///< Enable face pose estimation feature.
+    HInt32 enable_face_emotion;          ///< Enable face emotion recognition feature.
+} HFSessionCustomParameter, *PHFSessionCustomParameter;
+
+/**
+ * @brief Enumeration for face detection modes.
+ */
+typedef enum HFDetectMode {
+    HF_DETECT_MODE_ALWAYS_DETECT,       ///< Image detection mode, always detect, applicable to images.
+    HF_DETECT_MODE_LIGHT_TRACK,         ///< Video detection mode, face tracking, applicable to video
+                                        ///< streaming, front camera.
+    HF_DETECT_MODE_TRACK_BY_DETECTION,  ///< Video detection mode, face tracking, applicable to high
+                                        ///< resolution, monitoring, capturing
+                                        //   (You need a specific option turned on at compile time
+                                        //   to use it).
+} HFDetectMode;
+
+/**
+ * @brief Enum for landmark engine.
+ */
+typedef enum HFSessionLandmarkEngine {
+    HF_LANDMARK_HYPLMV2_0_25 = 0,             ///< Hyplmkv2 0.25, default
+    HF_LANDMARK_HYPLMV2_0_50 = 1,             ///< Hyplmkv2 0.50
+    HF_LANDMARK_INSIGHTFACE_2D106_TRACK = 2,  ///< InsightFace 2d106 track
+} HFSessionLandmarkEngine;
+
+/**
+ * @brief Global switch the landmark engine. Set it globally before creating a session.
+ *  If it is changed, a new session needs to be created for it to be effective.
+ * @param engine The landmark engine to be set.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSwitchLandmarkEngine(HFSessionLandmarkEngine engine);
+
+/**
+ * @brief Enum for supported pixel levels for face detection.
+ */
+typedef struct HFFaceDetectPixelList {
+    HInt32 pixel_level[20];
+    HInt32 size;
+} HFFaceDetectPixelList, *PHFFaceDetectPixelList;
+
+/**
+ * @brief Query the supported pixel levels for face detection. It must be used before starting.
+ * @param pixel_levels Pointer to the array of supported pixel levels.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFQuerySupportedPixelLevelsForFaceDetection(PHFFaceDetectPixelList pixel_levels);
+
+/**
+ * @brief Create a session from a resource file.
+ *
+ * @param parameter Custom parameters for session.
+ * @param detectMode Detection mode to be used.
+ * @param maxDetectFaceNum Maximum number of faces to detect.
+ * @param detectPixelLevel Modify the input resolution level of the detector, the larger the better,
+ *          the need to input a multiple of 160, such as 160, 320, 640, the default value -1 is 320.
+ * @param trackByDetectModeFPS If you are using the MODE_TRACK_BY_DETECTION tracking mode,
+ *          this value is used to set the fps frame rate of your current incoming video stream,
+ * which defaults to -1 at 30fps.
+ * @param handle Pointer to the context handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateInspireFaceSession(HFSessionCustomParameter parameter, HFDetectMode detectMode, HInt32 maxDetectFaceNum,
+                                                            HInt32 detectPixelLevel, HInt32 trackByDetectModeFPS, PHFSession handle);
+
+/**
+ * @brief Create a session from a resource file with additional options.
+ *
+ * @param customOption Custom option for additional configuration.
+ * @param detectMode Detection mode to be used.
+ * @param maxDetectFaceNum Maximum number of faces to detect.
+ * @param detectPixelLevel Modify the input resolution level of the detector, the larger the better,
+ *          the need to input a multiple of 160, such as 160, 320, 640, the default value -1 is 320.
+ * @param trackByDetectModeFPS If you are using the MODE_TRACK_BY_DETECTION tracking mode,
+ *          this value is used to set the fps frame rate of your current incoming video stream,
+ * which defaults to -1 at 30fps.
+ * @param handle Pointer to the context handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateInspireFaceSessionOptional(HOption customOption, HFDetectMode detectMode, HInt32 maxDetectFaceNum,
+                                                                    HInt32 detectPixelLevel, HInt32 trackByDetectModeFPS, PHFSession handle);
+
+/**
+ * @brief Release the session.
+ *
+ * @param handle Handle to the session to be released.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFReleaseInspireFaceSession(HFSession handle);
+
+/************************************************************************
+ * FaceTrack Module
+ *
+ * FaceTrack provides the most basic face image algorithm functions, such as face detection, tracking, landmark detection, etc.
+ * FaceTrack is independent of FaceSession, and can be used independently.
+ ************************************************************************/
+
+/**
+ * @brief Struct representing a basic token for face data.
+ *
+ * This struct holds the size and data pointer for a basic token associated with face data.
+ */
+typedef struct HFFaceBasicToken {
+    HInt32 size;  ///< Size of the token.
+    HPVoid data;  ///< Pointer to the token data.
+} HFFaceBasicToken, *PHFFaceBasicToken;
+
+/**
+ * @brief Struct for face Euler angles.
+ *
+ * This struct represents the Euler angles (roll, yaw, pitch) for face orientation.
+ */
+typedef struct HFFaceEulerAngle {
+    HPFloat roll;   ///< Roll angle of the face.
+    HPFloat yaw;    ///< Yaw angle of the face.
+    HPFloat pitch;  ///< Pitch angle of the face.
+} HFFaceEulerAngle;
+
+/**
+ * @brief Struct for holding data of multiple detected faces.
+ *
+ * This struct stores the data related to multiple faces detected, including the number of faces,
+ * their bounding rectangles, track IDs, angles, and tokens.
+ */
+typedef struct HFMultipleFaceData {
+    HInt32 detectedNum;        ///< Number of faces detected.
+    PHFaceRect rects;          ///< Array of bounding rectangles for each face.
+    HPInt32 trackIds;          ///< Array of track IDs for each face.
+    HPInt32 trackCounts;       ///< Array of track counts for each face.
+    HPFloat detConfidence;     ///< Array of detection confidence for each face.
+    HFFaceEulerAngle angles;   ///< Euler angles for each face.
+    PHFFaceBasicToken tokens;  ///< Tokens associated with each face.
+} HFMultipleFaceData, *PHFMultipleFaceData;
+
+/**
+ * @brief Clear the tracking face
+ * @param session Handle to the session.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionClearTrackingFace(HFSession session);
+
+/**
+ * @brief Set the track lost recovery mode(only for LightTrack mode, default is false(0))
+ * @param session Handle to the session.
+ * @param enable The track lost recovery mode value (0: disable, 1: enable)
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetTrackLostRecoveryMode(HFSession session, HInt32 enable);
+
+/**
+ * @brief Set the light track confidence threshold(only for LightTrack mode, default is 0.1)
+ * @param session Handle to the session.
+ * @param value The light track confidence threshold value
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetLightTrackConfidenceThreshold(HFSession session, HFloat value);
+
+/**
+ * @brief Set the track preview size in the session, it works with face detection and tracking
+ * algorithms. Default preview size is 192(px).
+ *
+ * @param session Handle to the session.
+ * @param previewSize The size of the preview for tracking.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetTrackPreviewSize(HFSession session, HInt32 previewSize);
+
+/**
+ * @brief Get the track preview size in the session.
+ * @param session Handle to the session.
+ * @param previewSize The size of the preview for tracking.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionGetTrackPreviewSize(HFSession session, HPInt32 previewSize);
+
+/**
+ * @brief Set the minimum number of face pixels that the face detector can capture, and people below
+ * this number will be filtered.
+ *
+ * @param session Handle to the session.
+ * @param minSize The minimum pixel value, default value is 0.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetFilterMinimumFacePixelSize(HFSession session, HInt32 minSize);
+
+/**
+ * @brief Set the face detect threshold in the session.
+ *
+ * @param session Handle to the session.
+ * @param detectMode The mode of the detection mode for tracking.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetFaceDetectThreshold(HFSession session, HFloat threshold);
+
+/**
+ * @brief Set the track mode smooth ratio in the session. default value is  0.05
+ *
+ * @param session Handle to the session.
+ * @param ratio The smooth ratio value.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetTrackModeSmoothRatio(HFSession session, HFloat ratio);
+
+/**
+ * @brief Set the track mode num smooth cache frame in the session. default value is 5
+ *
+ * @param session Handle to the session.
+ * @param num The num smooth cache frame value.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetTrackModeNumSmoothCacheFrame(HFSession session, HInt32 num);
+
+/**
+ * @brief Set the track model detect interval in the session. default value is 20
+ *
+ * @param session Handle to the session.
+ * @param num The detect interval value.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetTrackModeDetectInterval(HFSession session, HInt32 num);
+
+/**
+ * @brief Run face tracking in the session.
+ *
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param results Pointer to the structure where the results will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFExecuteFaceTrack(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData results);
+
+/**
+ * @brief Gets the size of the debug preview image for the last face detection in the session.
+ * @param session Handle to the session.
+ * @param size The size of the preview for tracking.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFSessionLastFaceDetectionGetDebugPreviewImageSize(HFSession session, HPInt32 size);
+
+/**
+ * @brief Copies the data from a HF_FaceBasicToken to a specified buffer.
+ *
+ * This function copies the data pointed to by the HF_FaceBasicToken's data field
+ * into a user-provided buffer. The caller is responsible for ensuring that the buffer
+ * is large enough to hold the data being copied.
+ *
+ * @param token The HF_FaceBasicToken containing the data to be copied.
+ * @param buffer The buffer where the data will be copied to.
+ * @param bufferSize The size of the buffer provided by the caller. Must be large enough
+ *        to hold the data pointed to by the token's data field.
+ * @return HResult indicating the success or failure of the operation. Returns HSUCCEED
+ *         if the operation was successful, or an error code if the buffer was too small
+ *         or if any other error occurred.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCopyFaceBasicToken(HFFaceBasicToken token, HPBuffer buffer, HInt32 bufferSize);
+
+/**
+ * @brief Retrieves the size of the data contained in a HF_FaceBasicToken.
+ *
+ * This function is used to query the size of the data that a HF_FaceBasicToken is
+ * expected to contain. This is useful for allocating a buffer of appropriate size
+ * before copying data from a HF_FaceBasicToken.
+ *
+ * @param bufferSize Pointer to an integer where the size of the data will be stored.
+ *        On successful completion, this will contain the size of the data in bytes.
+ * @return HResult indicating the success or failure of the operation. Returns HSUCCEED
+ *         if the operation was successful, or an error code if it failed.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceBasicTokenSize(HPInt32 bufferSize);
+
+/**
+ * @brief Retrieve the number of dense facial landmarks.
+ * @param num Number of dense facial landmarks
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetNumOfFaceDenseLandmark(HPInt32 num);
+
+/**
+ * @brief When you pass in a valid facial token, you can retrieve a set of dense facial landmarks.
+ *          The memory for the dense landmarks must be allocated by you.
+ * @param singleFace Basic token representing a single face.
+ * @param landmarks Pre-allocated memory address of the array for 2D floating-point coordinates.
+ * @param num Number of landmark points
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceDenseLandmarkFromFaceToken(HFFaceBasicToken singleFace, PHPoint2f landmarks, HInt32 num);
+
+/**
+ * @brief Get the five key points from the face token.
+ * @param singleFace Basic token representing a single face.
+ * @param landmarks Pre-allocated memory address of the array for 2D floating-point coordinates.
+ * @param num Number of landmark points
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceFiveKeyPointsFromFaceToken(HFFaceBasicToken singleFace, PHPoint2f landmarks, HInt32 num);
+
+/**
+ * @brief Set the enable cost spend
+ * @param value The enable cost spend value
+ * @return HResult Status code of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSessionSetEnableTrackCostSpend(HFSession session, HInt32 value);
+
+/**
+ * @brief Print the cost spend
+ * @param session The session handle
+ * @return HResult Status code of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSessionPrintTrackCostSpend(HFSession session);
+
+/************************************************************************
+ * Face Recognition Module
+ *
+ * The interface of the face recognition module depends on FaceSession,
+ * providing face feature extraction, face alignment image processing, and face comparison interfaces.
+ ************************************************************************/
+
+/**
+ * @brief Struct representing a face feature.
+ *
+ * This struct holds the data related to a face feature, including size and actual feature data.
+ */
+typedef struct HFFaceFeature {
+    HInt32 size;   ///< Size of the feature data.
+    HPFloat data;  ///< Pointer to the feature data.
+} HFFaceFeature, *PHFFaceFeature;
+
+/**
+ * @brief Extract a face feature from a given face.
+ *
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param singleFace Basic token representing a single face.
+ * @param feature Pointer to the extracted face feature.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceFeatureExtract(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace,
+                                                      PHFFaceFeature feature);
+
+/**
+ * @brief Extract face features to the HFFaceFeature that has applied for memory in advance.
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param singleFace Basic token representing a single face.
+ * @param feature Pointer to the buffer where the extracted feature will be copied.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceFeatureExtractTo(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace,
+                                                        HFFaceFeature feature);
+
+/**
+ * @brief Extract a face feature from a given face and copy it to the provided feature buffer.
+ *
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param singleFace Basic token representing a single face.
+ * @param feature Pointer to the buffer where the extracted feature will be copied.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceFeatureExtractCpy(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace, HPFloat feature);
+
+/**
+ * @brief Create a face feature. Will allocate memory.
+ * @param feature Pointer to the face feature.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCreateFaceFeature(PHFFaceFeature feature);
+
+/**
+ * @brief Release a face feature. Only the features created through the HFCreateFaceFeature need to be processed.
+ * @param feature Pointer to the face feature.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFReleaseFaceFeature(PHFFaceFeature feature);
+
+/**
+ * @brief Get the face alignment image.
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param singleFace Basic token representing a single face.
+ * @param handle Pointer to the handle that will be returned.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceGetFaceAlignmentImage(HFSession session, HFImageStream streamHandle, HFFaceBasicToken singleFace,
+                                                             PHFImageBitmap handle);
+
+/**
+ * @brief Use the aligned face image to extract face features to the HFFaceFeature that has applied memory in advance.
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param feature Pointer to the buffer where the extracted feature will be copied.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceFeatureExtractWithAlignmentImage(HFSession session, HFImageStream streamHandle, HFFaceFeature feature);
+
+/************************************************************************
+ * Feature Hub
+ *
+ * FeatureHub is a built-in global lightweight face feature vector management functionality
+ * provided in the InspireFace-SDK. It supports basic face feature search, deletion, and
+ * modification functions, and offers two optional data storage modes: an in-memory model and a
+ * persistence model. If you have simple storage needs, you can enable it.
+ ************************************************************************/
+
+/**
+ * @brief Select the search mode in the process of face recognition search,
+ * and different modes will affect the execution efficiency and results
+ * */
+typedef enum HFSearchMode {
+    HF_SEARCH_MODE_EAGER = 0,   // Eager mode: Stops when a vector meets the threshold.
+    HF_SEARCH_MODE_EXHAUSTIVE,  // Exhaustive mode: Searches until the best match is found.
+} HFSearchMode;
+
+/**
+ * @brief Primary key mode for face feature management.
+ */
+typedef enum HFPKMode {
+    HF_PK_AUTO_INCREMENT = 0,  ///< Auto-increment mode for primary key.
+    HF_PK_MANUAL_INPUT,        ///< Manual input mode for primary key.
+} HFPKMode;
+
+/**
+ * @brief Struct for database configuration.
+ *
+ * This struct holds the configuration settings for using a database in the face recognition context.
+ */
+typedef struct HFFeatureHubConfiguration {
+    HFPKMode primaryKeyMode;    ///< Primary key mode(The id increment mode is recommended)
+    HInt32 enablePersistence;   ///< Flag to enable or disable the use of the database.
+    HString persistenceDbPath;  ///< Path to the database file.
+    HFloat searchThreshold;     ///< Threshold for face search
+    HFSearchMode searchMode;    ///< Mode of face search
+} HFFeatureHubConfiguration;
+
+/**
+ * @brief A lightweight face feature vector management.
+ * @details FeatureHub is a built-in global lightweight face feature vector management functionality
+ * provided in the InspireFace-SDK. It supports basic face feature search, deletion, and
+ * modification functions, and offers two optional data storage modes: an in-memory model and a
+ * persistence model. If you have simple storage needs, you can enable it.
+ *
+ * @param configuration FeatureHub configuration details.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubDataEnable(HFFeatureHubConfiguration configuration);
+
+/**
+ * @brief Disable the global FeatureHub feature, and you can enable it again if needed.
+ * @return HResult indicating the success or failure of the operation.
+ * */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubDataDisable();
+
+/**
+ * @brief Struct representing the identity of a face feature.
+ *
+ * This struct associates a custom identifier and a tag with a specific face feature.
+ */
+typedef struct HFFaceFeatureIdentity {
+    HFaceId id;              ///< If you use automatic assignment id mode when inserting, ignore it.
+    PHFFaceFeature feature;  ///< Pointer to the face feature.
+    // HString tag;                 ///< Not supported yet
+} HFFaceFeatureIdentity, *PHFFaceFeatureIdentity;
+
+/**
+ * Search structure for top-k mode
+ * */
+typedef struct HFSearchTopKResults {
+    HInt32 size;         ///< The number of faces searched
+    HPFloat confidence;  ///< Search confidence(it has already been filtered once by the threshold)
+    HPFaceId ids;        ///< Searched face ids
+} HFSearchTopKResults, *PHFSearchTopKResults;
+
+/**
+ * @brief Set the face recognition search threshold.
+ *
+ * This function sets the threshold for face recognition, which determines the sensitivity
+ * of the recognition process. A lower threshold may yield more matches but with less confidence.
+ *
+ * @param threshold The threshold value to set for face recognition (default is 0.48, suitable for
+ * access control scenarios).
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubFaceSearchThresholdSetting(HFloat threshold);
+
+/**
+ * @brief Perform a one-to-one comparison of two face features.
+ *  Result is a cosine similarity score, not a percentage similarity.
+ *
+ * @param session Handle to the session.
+ * @param feature1 The first face feature for comparison.
+ * @param feature2 The second face feature for comparison.
+ * @param result Pointer to the floating-point value where the comparison result will be stored.
+ *               The result is a cosine similarity score, not a percentage similarity.
+ *               The score ranges from -1 to 1, where 1 indicates identical features,
+ *               0 indicates orthogonal features, and -1 indicates opposite features.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceComparison(HFFaceFeature feature1, HFFaceFeature feature2, HPFloat result);
+
+/**
+ * @brief Get recommended cosine threshold from loaded resource.
+ *  Use it to determine face similarity. Note: it's just a reference and may not be optimal for your task.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetRecommendedCosineThreshold(HPFloat threshold);
+
+/**
+ * @brief Convert cosine similarity to percentage similarity.
+ *  This is a nonlinear transformation function. You can adjust curve parameters to map the similarity distribution you need.
+ * @note The conversion parameters are primarily read from the Resource file configuration, as different models
+ *       have different conversion parameters. The parameters provided in the Resource file are only reference
+ *       values. If they do not meet your specific use case requirements, you can implement your own conversion
+ *       function.
+ * @param similarity The cosine similarity score.
+ * @param result Pointer to the floating-point value where the percentage similarity will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFCosineSimilarityConvertToPercentage(HFloat similarity, HPFloat result);
+
+/**
+ * @brief Similarity converter configuration.
+ */
+typedef struct HFSimilarityConverterConfig {
+    HFloat threshold;    ///< If you think that the threshold for judging the same person using cosine is some value such as 0.42,
+                         // you need to convert him to a percentage of 0.6(pass), you can modify it.
+    HFloat middleScore;  ///< Cosine threshold converted to a percentage reference value,
+                         // usually set 0.6 or 0.5, greater than it indicates similar, pass
+    HFloat steepness;    ///< Steepness of the curve, usually set 8.0
+    HFloat outputMin;    ///< Minimum value of output range, usually set 0.01
+    HFloat outputMax;    ///< Maximum value of output range, usually set 1.0
+} HFSimilarityConverterConfig, *PHFSimilarityConverterConfig;
+
+/**
+ * @brief Update the similarity converter configuration.
+ * @note The default configuration is loaded from the resource file during initialization.
+ *       This function allows you to override those default settings if needed.
+ * @param config The new similarity converter configuration to apply.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFUpdateCosineSimilarityConverter(HFSimilarityConverterConfig config);
+
+/**
+ * @brief Get the similarity converter configuration.
+ * @param config Pointer to the similarity converter configuration to be filled.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetCosineSimilarityConverter(PHFSimilarityConverterConfig config);
+
+/**
+ * @brief Get the length of the face feature.
+ *
+ * @param num Pointer to an integer where the length of the feature will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFeatureLength(HPInt32 num);
+
+/**
+ * @brief Insert a face feature identity into the features group.
+ *
+ * @param featureIdentity The face feature identity to be inserted.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubInsertFeature(HFFaceFeatureIdentity featureIdentity, HPFaceId allocId);
+
+/**
+ * @brief Search for the most similar face feature in the features group.
+ *
+ * @param searchFeature The face feature to be searched.
+ * @param confidence Pointer to a floating-point value where the confidence level of the match will
+ * be stored.
+ * @param mostSimilar Pointer to the most similar face feature identity found.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubFaceSearch(HFFaceFeature searchFeature, HPFloat confidence, PHFFaceFeatureIdentity mostSimilar);
+
+/**
+ * @brief Search for the most similar k facial features in the feature group
+ *
+ * @param searchFeature The face feature to be searched.
+ * @param confidence topK Maximum number of searches
+ * @param PHFSearchTopKResults Output search result
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubFaceSearchTopK(HFFaceFeature searchFeature, HInt32 topK, PHFSearchTopKResults results);
+
+/**
+ * @brief Remove a face feature from the features group based on custom ID.
+ *
+ * @param ID The custom ID of the feature to be removed.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubFaceRemove(HFaceId id);
+
+/**
+ * @brief Update a face feature identity in the features group.
+ *
+ * @param featureIdentity The face feature identity to be updated.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubFaceUpdate(HFFaceFeatureIdentity featureIdentity);
+
+/**
+ * @brief Retrieve a face feature identity from the features group based on custom ID.
+ *
+ * @param customId The custom ID of the feature.
+ * @param identity Pointer to the face feature identity to be retrieved.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubGetFaceIdentity(HFaceId customId, PHFFaceFeatureIdentity identity);
+
+/**
+ * @brief Get the count of face features in the features group.
+ *
+ * @param count Pointer to an integer where the count of features will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubGetFaceCount(HPInt32 count);
+
+/**
+ * @brief View the face database table.
+ *
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubViewDBTable();
+
+/**
+ * @brief Struct representing the existing ids in the database.
+ */
+typedef struct HFFeatureHubExistingIds {
+    HInt32 size;   ///< The number of ids
+    HPFaceId ids;  ///< The ids
+} HFFeatureHubExistingIds, *PHFFeatureHubExistingIds;
+
+/**
+ * @brief Get all ids in the database.
+ * @param ids Output parameter to store the ids.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFeatureHubGetExistingIds(PHFFeatureHubExistingIds ids);
+
+/************************************************************************
+ * Face Pipeline Module
+ *
+ * FacePipeline depends on FaceSession, providing extended business for face image algorithms,
+ * supporting some face attributes, such as face mask detection, face quality detection, face attribute prediction, etc.
+ ************************************************************************/
+
+/**
+ * @brief Process multiple faces in a pipeline.
+ *
+ * This function processes multiple faces detected in an image or video frame, applying
+ * various face recognition and analysis features as specified in the parameters.
+ *
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param faces Pointer to the structure containing data of multiple detected faces.
+ * @param parameter Custom parameters for processing the faces.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFMultipleFacePipelineProcess(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData faces,
+                                                               HFSessionCustomParameter parameter);
+
+/**
+ * @brief Process multiple faces in a pipeline with an optional custom option.
+ *
+ * Similar to HFMultipleFacePipelineProcess, but allows for additional custom options
+ * to modify the face processing behavior.
+ *
+ * @param session Handle to the session.
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param faces Pointer to the structure containing data of multiple detected faces.
+ * @param customOption An integer representing a custom option for processing.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFMultipleFacePipelineProcessOptional(HFSession session, HFImageStream streamHandle, PHFMultipleFaceData faces,
+                                                                       HInt32 customOption);
+
+/**
+ * @brief Struct representing RGB liveness confidence.
+ *
+ * This struct holds the number of faces and the confidence level of liveness detection
+ * for each face, using RGB analysis.
+ */
+typedef struct HFRGBLivenessConfidence {
+    HInt32 num;          ///< Number of faces detected.
+    HPFloat confidence;  ///< Confidence level of RGB liveness detection for each face.
+} HFRGBLivenessConfidence, *PHFRGBLivenessConfidence;
+
+/**
+ * @brief Get the RGB liveness confidence.
+ *
+ * This function retrieves the confidence level of RGB liveness detection for faces detected
+ * in the current context.
+ *
+ * @param session Handle to the session.
+ * @param confidence Pointer to the structure where RGB liveness confidence data will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetRGBLivenessConfidence(HFSession session, PHFRGBLivenessConfidence confidence);
+
+/**
+ * @brief Struct representing face mask confidence.
+ *
+ * This struct holds the number of faces and the confidence level of mask detection
+ * for each face.
+ */
+typedef struct HFFaceMaskConfidence {
+    HInt32 num;          ///< Number of faces detected.
+    HPFloat confidence;  ///< Confidence level of mask detection for each face.
+} HFFaceMaskConfidence, *PHFFaceMaskConfidence;
+
+/**
+ * @brief Get the face mask confidence.
+ *
+ * This function retrieves the confidence level of mask detection for faces detected
+ * in the current context.
+ *
+ * @param session Handle to the session.
+ * @param confidence Pointer to the structure where face mask confidence data will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceMaskConfidence(HFSession session, PHFFaceMaskConfidence confidence);
+
+/**
+ * @brief Struct representing face quality predict confidence.
+ *
+ * This struct holds the number of faces and the confidence level of face quality predict
+ * for each face.
+ */
+typedef struct HFFaceQualityConfidence {
+    HInt32 num;          ///< Number of faces detected.
+    HPFloat confidence;  ///< Confidence level of face quality predict for each face.
+} HFFaceQualityConfidence, *PHFFaceQualityConfidence;
+
+/**
+ * @brief Get the face quality predict confidence.
+ *
+ * This function retrieves the confidence level of face quality predict for faces detected
+ * in the current context.
+ *
+ * @param session Handle to the session.
+ * @param confidence Pointer to the structure where face mask confidence data will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceQualityConfidence(HFSession session, PHFFaceQualityConfidence confidence);
+
+/**
+ * @brief Detect the quality of a face in an image.
+ *
+ * This function assesses the quality of a detected face, such as its clarity and visibility.
+ *
+ * @param session Handle to the session.
+ * @param singleFace A token representing a single face.
+ * @param confidence Pointer to a floating-point value where the quality confidence will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFFaceQualityDetect(HFSession session, HFFaceBasicToken singleFace, HPFloat confidence);
+
+/**
+ * @brief Facial states in the face interaction module.
+ */
+typedef struct HFFaceInteractionState {
+    HInt32 num;                        ///< Number of faces detected.
+    HPFloat leftEyeStatusConfidence;   ///< Left eye state: confidence close to 1 means open, close
+                                       ///< to 0 means closed.
+    HPFloat rightEyeStatusConfidence;  ///< Right eye state: confidence close to 1 means open, close
+                                       ///< to 0 means closed.
+} HFFaceInteractionState, *PHFFaceInteractionState;
+
+/**
+ * @brief Get the prediction results of face interaction.
+ * @param session Handle to the session.
+ * @param result Facial state prediction results in the face interaction module.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceInteractionStateResult(HFSession session, PHFFaceInteractionState result);
+
+/**
+ * @brief Actions detected in the face interaction module.
+ */
+typedef struct HFFaceInteractionsActions {
+    HInt32 num;         ///< Number of actions detected.
+    HPInt32 normal;     ///< Normal actions.
+    HPInt32 shake;      ///< Shake actions.
+    HPInt32 jawOpen;    ///< Jaw open actions.
+    HPInt32 headRaise;  ///< Head raise actions.
+    HPInt32 blink;      ///< Blink actions.
+} HFFaceInteractionsActions, *PHFFaceInteractionsActions;
+
+/**
+ * @brief Get the prediction results of face interaction actions.
+ * @param session Handle to the session.
+ * @param actions Facial action prediction results in the face interaction module.
+ * @return HResult indicating success or failure of the function call.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceInteractionActionsResult(HFSession session, PHFFaceInteractionsActions actions);
+
+/**
+ * @brief Struct representing face attribute results.
+ *
+ * This struct holds the race, gender, and age bracket attributes for a detected face.
+ */
+typedef struct HFFaceAttributeResult {
+    HInt32 num;          ///< Number of faces detected.
+    HPInt32 race;        ///< Race of the detected face.
+                         ///< 0: Black;
+                         ///< 1: Asian;
+                         ///< 2: Latino/Hispanic;
+                         ///< 3: Middle Eastern;
+                         ///< 4: White;
+    HPInt32 gender;      ///< Gender of the detected face.
+                         ///< 0: Female;
+                         ///< 1: Male;
+    HPInt32 ageBracket;  ///< Age bracket of the detected face.
+                         ///< 0: 0-2 years old;
+                         ///< 1: 3-9 years old;
+                         ///< 2: 10-19 years old;
+                         ///< 3: 20-29 years old;
+                         ///< 4: 30-39 years old;
+                         ///< 5: 40-49 years old;
+                         ///< 6: 50-59 years old;
+                         ///< 7: 60-69 years old;
+                         ///< 8: more than 70 years old;
+} HFFaceAttributeResult, *PHFFaceAttributeResult;
+
+/**
+ * @brief Get the face attribute results.
+ *
+ * This function retrieves the attribute results such as race, gender, and age bracket
+ * for faces detected in the current context.
+ *
+ * @param session Handle to the session.
+ * @param results Pointer to the structure where face attribute results will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceAttributeResult(HFSession session, PHFFaceAttributeResult results);
+
+/**
+ * @brief Struct representing face emotion results.
+ */
+typedef struct HFFaceEmotionResult {
+    HInt32 num;       ///< Number of faces detected.
+    HPInt32 emotion;  ///< Emotion of the detected face.
+                      ///< 0: Neutral;
+                      ///< 1: Happy;
+                      ///< 2: Sad;
+                      ///< 3: Surprise;
+                      ///< 4: Fear;
+                      ///< 5: Disgust;
+                      ///< 6: Anger;
+} HFFaceEmotionResult, *PHFFaceEmotionResult;
+
+/**
+ * @brief Get the face emotion result.
+ * @param session Handle to the session.
+ * @param result Pointer to the structure where face emotion results will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFGetFaceEmotionResult(HFSession session, PHFFaceEmotionResult result);
+
+/************************************************************************
+ * System Function
+ ************************************************************************/
+
+/**
+ * @brief Structure representing the version information of the InspireFace library.
+ */
+typedef struct HFInspireFaceVersion {
+    HInt32 major;  ///< Major version number.
+    HInt32 minor;  ///< Minor version number.
+    HInt32 patch;  ///< Patch version number.
+} HFInspireFaceVersion, *PHFInspireFaceVersion;
+
+/**
+ * @brief Function to query the version information of the InspireFace library.
+ *
+ * This function retrieves the version information of the InspireFace library.
+ *
+ * @param version Pointer to the structure where the version information will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFQueryInspireFaceVersion(PHFInspireFaceVersion version);
+
+/**
+ * @brief Struct representing the extended information of the InspireFace library.
+ */
+typedef struct HFInspireFaceExtendedInformation {
+    HChar information[256];
+    // TODO: Add more information
+} HFInspireFaceExtendedInformation, *PHFInspireFaceExtendedInformation;
+
+/**
+ * @brief Get the extended information of the InspireFace library.
+ *
+ * This function retrieves the extended information of the InspireFace library.
+ */
+HYPER_CAPI_EXPORT extern HResult HFQueryInspireFaceExtendedInformation(PHFInspireFaceExtendedInformation information);
+
+/**
+ * @brief SDK built-in log level mode
+ * */
+typedef enum HFLogLevel {
+    HF_LOG_NONE = 0,  // No logging, disables all log output
+    HF_LOG_DEBUG,     // Debug level for detailed system information mostly useful for developers
+    HF_LOG_INFO,      // Information level for general system information about operational status
+    HF_LOG_WARN,      // Warning level for non-critical issues that might need attention
+    HF_LOG_ERROR,     // Error level for error events that might still allow the application to continue running
+    HF_LOG_FATAL      // Fatal level for severe error events that will presumably lead the application to abort
+} HFLogLevel;
+
+/**
+ * @brief Set the log level built into the SDK.The default is HF LOG DEBUG
+ * */
+HYPER_CAPI_EXPORT extern HResult HFSetLogLevel(HFLogLevel level);
+
+/**
+ * @brief Disable the log function. Like HFSetLogLevel(HF_LOG_NONE)
+ * */
+HYPER_CAPI_EXPORT extern HResult HFLogDisable();
+
+/**
+ * @brief Print the log.
+ * @param level The log level.
+ * @param format The log format.
+ * @param ... The log arguments.
+ * @warning The maximum buffer size for log messages is 1024 bytes. Messages longer than this will be truncated.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFLogPrint(HFLogLevel level, HFormat format, ...);
+
+/********************************DEBUG Utils****************************************/
+
+/**
+ * @brief Display an image stream for debugging purposes.
+ *
+ * This function is used for debugging, allowing the visualization of the image stream
+ * as it is being processed. It can be useful to understand the data being received
+ * from the camera or image source.
+ *
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ */
+HYPER_CAPI_EXPORT extern void HFDeBugImageStreamImShow(HFImageStream streamHandle);
+
+/**
+ * @brief Decode the image from ImageStream and store it to a disk path.
+ *
+ * It is used to verify whether there is a problem with image codec, and can quickly perform bug
+ * analysis.
+ *
+ * @param streamHandle Handle to the data buffer representing the camera stream component.
+ * @param savePath The path to which the image is written.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugImageStreamDecodeSave(HFImageStream streamHandle, HPath savePath);
+
+/**
+ * @brief Display current resource management statistics.
+ *
+ * This function prints statistics about the resources managed by the ResourceManager,
+ * including the total number of created and released sessions and image streams, as well as
+ * the count of those that have not been released yet. This can be used for debugging purposes
+ * to ensure that resources are being properly managed and to identify potential resource leaks.
+ *
+ * @return HResult indicating the success or failure of the operation.
+ *         Returns HSUCCEED if the statistics were successfully displayed,
+ *         otherwise, it may return an error code if there is an issue accessing the resource
+ * manager.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugShowResourceStatistics();
+
+/**
+ * @brief Get the count of unreleased sessions.
+ *
+ * This function retrieves the count of sessions that have not been released yet.
+ *
+ * @param count Pointer to an integer where the count of unreleased sessions will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugGetUnreleasedSessionsCount(HPInt32 count);
+
+/**
+ * @brief Get the list of unreleased sessions.
+ *
+ * This function retrieves the list of sessions that have not been released yet.
+ *
+ * @param sessions Pointer to an array where the unreleased sessions will be stored.
+ * @param count The number of sessions to retrieve.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugGetUnreleasedSessions(PHFSession sessions, HInt32 count);
+
+/**
+ * @brief Get the count of unreleased image streams.
+ *
+ * This function retrieves the count of image streams that have not been released yet.
+ *
+ * @param count Pointer to an integer where the count of unreleased image streams will be stored.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugGetUnreleasedStreamsCount(HPInt32 count);
+
+/**
+ * @brief Get the list of unreleased image streams.
+ *
+ * This function retrieves the list of image streams that have not been released yet.
+ *
+ * @param streams Pointer to an array where the unreleased image streams will be stored.
+ * @param count The number of image streams to retrieve.
+ * @return HResult indicating the success or failure of the operation.
+ */
+HYPER_CAPI_EXPORT extern HResult HFDeBugGetUnreleasedStreams(PHFImageStream streams, HInt32 count);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // INSPIREFACE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/c_api/inspireface_internal.h b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface_internal.h
new file mode 100644
index 000000000..4b79100b1
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/c_api/inspireface_internal.h
@@ -0,0 +1,38 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_INTERNAL_H
+#define INSPIREFACE_INTERNAL_H
+
+#include "engine/face_session.h"
+
+/**
+ * @brief Struct for managing face algorithm session.
+ * 
+ * This struct holds the implementation of the face algorithm session.
+ */
+typedef struct HF_FaceAlgorithmSession {
+    inspire::FaceSession impl;  ///< Implementation of the face context.
+} HF_FaceAlgorithmSession;      ///< Handle for managing face context.
+
+/**
+ * @brief Struct for managing camera stream.
+ * 
+ * This struct holds the implementation of the camera stream.
+ */
+typedef struct HF_CameraStream {
+    inspirecv::FrameProcess impl;  ///< Implementation of the camera stream.
+} HF_CameraStream;                 ///< Handle for managing camera stream.
+
+/**
+ * @brief Struct for managing image bitmap.
+ * 
+ * This struct holds the implementation of the image bitmap.
+ */
+typedef struct HF_ImageBitmap {
+    inspirecv::Image impl;  ///< Implementation of the image bitmap.
+} HF_ImageBitmap;           ///< Handle for managing image bitmap.
+
+#endif  // INSPIREFACE_INTERNAL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/c_api/intypedef.h b/cpp-package/inspireface/cpp/inspireface/c_api/intypedef.h
new file mode 100644
index 000000000..ae80be1b3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/c_api/intypedef.h
@@ -0,0 +1,63 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_INTYPEDEF_H
+#define INSPIREFACE_INTYPEDEF_H
+#include <stdint.h>
+
+// clang-format off
+typedef void*               HFImageStream;                    ///< Handle for image.
+typedef void**              PHFImageStream;                   ///< Pointer to Handle for image.
+typedef void*               HFSession;                        ///< Handle for context.
+typedef void**              PHFSession;                       ///< Pointer to Handle for context.
+typedef void*               HFImageBitmap;                    ///< Handle for image bitmap.
+typedef void**              PHFImageBitmap;                   ///< Pointer to Handle for image bitmap.
+typedef void*               HPVoid;                           ///< Pointer to Void.
+typedef long                HLong;                            ///< Long integer.
+typedef float               HFloat;                           ///< Single-precision floating point.
+typedef float*              HPFloat;                          ///< Pointer to Single-precision floating point.
+typedef double              HDouble;                          ///< Double-precision floating point.
+typedef	unsigned char		HUInt8;                           ///< Unsigned 8-bit integer.
+typedef unsigned char*      HPUInt8;                          ///< Pointer to unsigned 8-bit integer.
+typedef signed int			HInt32;                           ///< Signed 32-bit integer.
+typedef signed int			HOption;                          ///< Signed 32-bit integer option.
+typedef signed int*			HPInt32;                          ///< Pointer to signed 32-bit integer.
+typedef int64_t             HFaceId;                          ///< Face ID type for non-Windows platforms
+typedef int64_t*            HPFaceId;                         ///< Pointer to Face ID type for non-Windows platforms
+typedef long                HResult;                          ///< Result code.
+typedef char*               HString;                          ///< String.
+typedef const char*         HPath;                            ///< Const String.
+typedef const char*         HFormat;                          ///< Const String.
+typedef char                HBuffer;                          ///< Character.
+typedef char                HChar;                            ///< Character.
+typedef char*               HPBuffer;                         ///< Pointer Character.
+typedef long                HSize;                            ///< Size
+typedef long*               HPSize;                           ///< Pointer Size
+// clang-format on  
+
+typedef struct HFaceRect {
+    HInt32 x;                                                 ///< X-coordinate of the top-left corner of the rectangle.
+    HInt32 y;                                                 ///< Y-coordinate of the top-left corner of the rectangle.
+    HInt32 width;                                             ///< Width of the rectangle.
+    HInt32 height;                                            ///< Height of the rectangle.
+} HFaceRect, *PHFaceRect;                                     ///< Rectangle representing a face region.
+
+typedef struct HPoint2f{
+    HFloat x;                                                 ///< X-coordinate
+    HFloat y;                                                 ///< Y-coordinate
+} HPoint2f, *PHPoint2f;
+
+typedef struct HPoint2i{
+    HInt32 x;                                                 ///< X-coordinate
+    HInt32 y;                                                 ///< Y-coordinate
+} HPoint2i, *PHPoint2i;
+
+typedef struct HColor {
+    HFloat r;                                                 ///< Red component
+    HFloat g;                                                 ///< Green component
+    HFloat b;                                                 ///< Blue component
+} HColor, *PHColor;
+
+#endif //INSPIREFACE_INTYPEDEF_H
diff --git a/cpp-package/inspireface/cpp/inspireface/common/face_data/face_serialize_tools.h b/cpp-package/inspireface/cpp/inspireface/common/face_data/face_serialize_tools.h
new file mode 100644
index 000000000..1ae2175dc
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/common/face_data/face_serialize_tools.h
@@ -0,0 +1,192 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIRE_FACE_SERIALIZE_TOOLS_H
+#define INSPIRE_FACE_SERIALIZE_TOOLS_H
+
+#include "face_wrapper.h"
+#include "../face_info/face_object_internal.h"
+#include "herror.h"
+#include "data_type.h"
+#include "track_module/landmark/all.h"
+#include <log.h>
+
+// Define the namespace "inspire" for encapsulation
+namespace inspire {
+
+/**
+ * @brief Print the transformation matrix.
+ * @param matrix The transformation matrix to print.
+ */
+inline void PrintTransformMatrix(const TransMatrix& matrix) {
+    INSPIRE_LOGI("Transformation Matrix:");
+    INSPIRE_LOGI("a: %f\tb: %f\ttx: %f", matrix.m00, matrix.m01, matrix.tx);
+    INSPIRE_LOGI("c: %f\td: %f\tty: %f", matrix.m10, matrix.m11, matrix.ty);
+}
+
+/**
+ * @brief Print FaceTrackWrap structure.
+ * @param data The FaceTrackWrap structure to print.
+ */
+inline void INSPIRE_API PrintHyperFaceDataDetail(const FaceTrackWrap& data) {
+    INSPIRE_LOGI("Track State: %d", data.trackState);
+    INSPIRE_LOGI("In Group Index: %d", data.inGroupIndex);
+    INSPIRE_LOGI("Track ID: %d", data.trackId);
+    INSPIRE_LOGI("Track Count: %d", data.trackCount);
+
+    INSPIRE_LOGI("Face Rectangle:");
+    INSPIRE_LOGI("x: %f\ty: %f\twidth: %f\theight: %f", data.rect.x, data.rect.y, data.rect.width, data.rect.height);
+
+    PrintTransformMatrix(data.trans);
+}
+
+/**
+ * @brief Convert a FaceObject to FaceTrackWrap.
+ * @param obj The FaceObject to convert.
+ * @param group_index The group index.
+ * @return The converted FaceTrackWrap structure.
+ */
+inline FaceTrackWrap INSPIRE_API FaceObjectInternalToHyperFaceData(const FaceObjectInternal& obj, int group_index = -1) {
+    FaceTrackWrap data;
+    // Face rect
+    data.rect.x = obj.bbox_.GetX();
+    data.rect.y = obj.bbox_.GetY();
+    data.rect.width = obj.bbox_.GetWidth();
+    data.rect.height = obj.bbox_.GetHeight();
+    // Trans matrix
+    data.trans.m00 = obj.getTransMatrix().Get(0, 0);
+    data.trans.m01 = obj.getTransMatrix().Get(0, 1);
+    data.trans.m10 = obj.getTransMatrix().Get(1, 0);
+    data.trans.m11 = obj.getTransMatrix().Get(1, 1);
+    data.trans.tx = obj.getTransMatrix().Get(0, 2);
+    data.trans.ty = obj.getTransMatrix().Get(1, 2);
+    // KetPoints five
+    if (!obj.high_result.lmk.empty()) {
+        for (int i = 0; i < obj.high_result.lmk.size(); ++i) {
+            data.keyPoints[i].x = obj.high_result.lmk[i].GetX();
+            data.keyPoints[i].y = obj.high_result.lmk[i].GetY();
+        }
+        for (int i = 0; i < 5; ++i) {
+            data.quality[i] = obj.high_result.lmk_quality[i];
+        }
+        //        LOGD("HIGHT");
+    } else {
+        for (int i = 0; i < obj.keyPointFive.size(); ++i) {
+            data.keyPoints[i].x = obj.keyPointFive[i].GetX();
+            data.keyPoints[i].y = obj.keyPointFive[i].GetY();
+        }
+        for (int i = 0; i < 5; ++i) {
+            data.quality[i] = -1.0f;
+        }
+    }
+    // Basic data
+    data.inGroupIndex = group_index;
+    data.trackCount = obj.tracking_count_;
+    data.trackId = obj.GetTrackingId();
+    data.trackState = obj.TrackingState();
+    // Face 3D Angle
+    data.face3DAngle.pitch = obj.high_result.pitch;
+    data.face3DAngle.roll = obj.high_result.roll;
+    data.face3DAngle.yaw = obj.high_result.yaw;
+    // Density Landmark
+    if (!obj.landmark_smooth_aux_.empty()) {
+        data.densityLandmarkEnable = 1;
+        const auto& lmk = obj.landmark_smooth_aux_.back();
+        for (size_t i = 0; i < FaceLandmarkAdapt::NUM_OF_LANDMARK; i++) {
+            data.densityLandmark[i].x = lmk[i].GetX();
+            data.densityLandmark[i].y = lmk[i].GetY();
+        }
+    } else {
+        data.densityLandmarkEnable = 0;
+    }
+
+    return data;
+}
+
+/**
+ * @brief Convert a TransMatrix to a cv::Mat.
+ * @param trans The TransMatrix to convert.
+ * @return The converted cv::Mat.
+ */
+inline inspirecv::TransformMatrix INSPIRE_API TransformMatrixToInternalMatrix(const TransMatrix& trans) {
+    inspirecv::TransformMatrix mat = inspirecv::TransformMatrix::Create(trans.m00, trans.m01, trans.tx, trans.m10, trans.m11, trans.ty);
+    return mat;
+}
+
+/**
+ * @brief Convert a FaceRect to cv::Rect.
+ * @param faceRect The FaceRect to convert.
+ * @return The converted cv::Rect.
+ */
+inline inspirecv::Rect2i INSPIRE_API FaceRectToInternalRect(const FaceRect& faceRect) {
+    return inspirecv::Rect2i(faceRect.x, faceRect.y, faceRect.width, faceRect.height);
+}
+
+/**
+ * @brief Convert a Point2F to cv::Point2f.
+ * @param point The Point2F to convert.
+ * @return The converted cv::Point2f.
+ */
+inline inspirecv::Point2f INSPIRE_API HPointToInternalPoint2f(const Point2F& point) {
+    return inspirecv::Point2f(point.x, point.y);
+}
+
+/**
+ * @brief Serialize FaceTrackWrap to a byte stream.
+ * @param data The FaceTrackWrap to serialize.
+ * @param byteArray The output byte stream.
+ * @return The result code.
+ */
+inline int32_t INSPIRE_API RunSerializeHyperFaceData(const FaceTrackWrap& data, ByteArray& byteArray) {
+    byteArray.reserve(sizeof(data));
+
+    // Serialize the FaceTrackWrap structure itself
+    const char* dataBytes = reinterpret_cast<const char*>(&data);
+    byteArray.insert(byteArray.end(), dataBytes, dataBytes + sizeof(data));
+
+    return HSUCCEED;
+}
+
+/**
+ * @brief Deserialize a byte stream to FaceTrackWrap.
+ * @param byteArray The input byte stream.
+ * @param data The output FaceTrackWrap structure.
+ * @return The result code.
+ */
+inline int32_t INSPIRE_API RunDeserializeHyperFaceData(const ByteArray& byteArray, FaceTrackWrap& data) {
+    // Check if the byte stream size is sufficient
+    if (byteArray.size() >= sizeof(data)) {
+        // Copy data from the byte stream to the FaceTrackWrap structure
+        std::memcpy(&data, byteArray.data(), sizeof(data));
+    } else {
+        INSPIRE_LOGE("The byte stream size is insufficient to restore FaceTrackWrap");
+        return HERR_SESS_FACE_DATA_ERROR;
+    }
+
+    return HSUCCEED;
+}
+
+/**
+ * @brief Deserialize a byte stream to FaceTrackWrap.
+ * @param byteArray The input byte stream as a character array.
+ * @param byteCount The size of the byte stream.
+ * @param data The output FaceTrackWrap structure.
+ * @return The result code.
+ */
+inline int32_t INSPIRE_API RunDeserializeHyperFaceData(const char* byteArray, size_t byteCount, FaceTrackWrap& data) {
+    // Check if the byte stream size is sufficient
+    if (byteCount >= sizeof(data)) {
+        // Copy data from the byte stream to the FaceTrackWrap structure
+        std::memcpy(&data, byteArray, sizeof(data));
+    } else {
+        INSPIRE_LOGE("The byte stream size is insufficient to restore FaceTrackWrap");
+        return HERR_SESS_FACE_DATA_ERROR;
+    }
+
+    return HSUCCEED;
+}
+
+}  // namespace inspire
+#endif  // INSPIRE_FACE_SERIALIZE_TOOLS_H
diff --git a/cpp-package/inspireface/cpp/inspireface/common/face_info/all.h b/cpp-package/inspireface/cpp/inspireface/common/face_info/all.h
new file mode 100644
index 000000000..53354d2bf
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/common/face_info/all.h
@@ -0,0 +1,13 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_FO_ALL_H
+#define INSPIRE_FACE_FO_ALL_H
+
+#include "face_object_internal.h"
+#include "face_action_data.h"
+#include "face_process.h"
+
+#endif  // INSPIRE_FACE_FO_ALL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/common/face_info/face_action_data.h b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_action_data.h
new file mode 100644
index 000000000..189a47c94
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_action_data.h
@@ -0,0 +1,143 @@
+#ifndef INSPIRSE_FACE_FACE_ACTION_DATA_H
+#define INSPIRSE_FACE_FACE_ACTION_DATA_H
+
+#include <iostream>
+#include <inspirecv/inspirecv.h>
+#include "middleware/utils.h"
+#include "data_type.h"
+#include "track_module/landmark/face_landmark_adapt.h"
+
+namespace inspire {
+
+enum FACE_ACTIONS { ACT_NORMAL = 0, ACT_SHAKE = 0, ACT_BLINK = 1, ACT_JAW_OPEN = 2, ACT_RAISE_HEAD = 3 };
+
+typedef struct FaceActionList {
+    int normal = 0;
+    int shake = 0;
+    int blink = 0;
+    int jawOpen = 0;
+    int raiseHead = 0;
+} FaceActionList;
+
+class INSPIRE_API FaceActionPredictor {
+public:
+    FaceActionPredictor(int record_list_length) {
+        record_list.resize(record_list_length);
+        record_list_euler.resize(record_list_length);
+        record_list_eyes.resize(record_list_length);
+        record_size = record_list_length;
+        index = 0;
+    }
+
+    void RecordActionFrame(const std::vector<inspirecv::Point2f> &landmark, const inspirecv::Vec3f &euler_angle,
+                           const inspirecv::Vec2f &eyes_status) {
+        MoveRecordList();
+        record_list[0] = landmark;
+        record_list_euler[0] = euler_angle;
+        record_list_eyes[0] = eyes_status;
+        index += 1;
+    }
+
+    void Reset() {
+        record_list.clear();
+        record_list.resize(record_size);
+        record_list_euler.clear();
+        record_list_euler.resize(record_size);
+        record_list_eyes.clear();
+        record_list_eyes.resize(record_size);
+        index = 0;
+    }
+
+    FaceActionList AnalysisFaceAction(const SemanticIndex& semantic_index) {
+        FaceActionList actionRecord;
+        actions.clear();
+        eye_state_list.clear();
+        if (index < record_list.size()) {
+            actions.push_back(ACT_NORMAL);
+            actionRecord.normal = 1;
+        } else {
+            for (int i = 0; i < record_list_eyes.size(); i++) {
+                const auto &eye = record_list_eyes[i];
+                std::pair<float, float> eye_state(eye[0], eye[1]);
+                eye_state_list.push_back(eye_state);
+            }
+
+            // count mouth aspect ratio
+
+            float mouth_widthwise_d = record_list[0][semantic_index.mouth_left_corner].Distance(record_list[0][semantic_index.mouth_right_corner]);
+            float mouth_heightwise_d = record_list[0][semantic_index.mouth_upper].Distance(record_list[0][semantic_index.mouth_lower]);
+            float mouth_aspect_ratio = mouth_heightwise_d / mouth_widthwise_d;
+            if (mouth_aspect_ratio > 0.3) {
+                actions.push_back(ACT_JAW_OPEN);
+                actionRecord.jawOpen = 1;
+            }
+
+            int counter_eye_open = 0;
+            int counter_eye_close = 0;
+            for (auto &e : eye_state_list) {
+                if (e.first < 0.5 || e.second < 0.5) {
+                    counter_eye_close += 1;
+                }
+                if (e.first > 0.5 || e.second > 0.5) {
+                    counter_eye_open += 1;
+                }
+            }
+            if (counter_eye_close > 0 && counter_eye_open > 2 && record_list_euler[0][1] > -6 && record_list_euler[0][0] < 6) {
+                actions.push_back(ACT_BLINK);
+                actionRecord.blink = 1;
+                Reset();
+            }
+
+            bool counter_head_shake_left = false;
+            bool counter_head_shake_right = false;
+            for (auto &e : record_list_euler) {
+                if (e[1] < -6) {
+                    counter_head_shake_left = true;
+                }
+                if (e[1] > 6) {
+                    counter_head_shake_right = true;
+                }
+            }
+            if (counter_head_shake_left && counter_head_shake_right) {
+                actions.push_back(ACT_SHAKE);
+                actionRecord.shake = 1;
+            }
+
+            if (record_list_euler[0][0] > 10) {
+                actions.push_back(ACT_RAISE_HEAD);
+                actionRecord.raiseHead = 1;
+            }
+        }
+        return actionRecord;
+    }
+
+    std::vector<FACE_ACTIONS> GetActions() const {
+        return actions;
+    }
+
+private:
+    void MoveRecordList() {
+        // for(int i = 0 ; i < record_list.size() - 1 ; i++){
+        //    record_list[i+1] = record_list[i];
+        //    record_list_euler[i+1] = record_list_euler[i];
+        //}
+        for (int i = record_list.size() - 1; i > 0; i--) {
+            record_list[i] = record_list[i - 1];
+            record_list_euler[i] = record_list_euler[i - 1];
+            record_list_eyes[i] = record_list_eyes[i - 1];
+        }
+    }
+
+    std::vector<std::vector<inspirecv::Point2f>> record_list;
+    std::vector<inspirecv::Vec3f> record_list_euler;
+    std::vector<inspirecv::Vec2f> record_list_eyes;
+    std::vector<std::pair<float, float>> eye_state_list;  // pair  left right
+    std::vector<float> mouth_state_list;
+    std::vector<FACE_ACTIONS> actions;
+    int record_size;
+    int index;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRSE_FACE_FACE_ACTION_DATA_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/common/face_info/face_object_internal.h b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_object_internal.h
new file mode 100755
index 000000000..4a9461aae
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_object_internal.h
@@ -0,0 +1,276 @@
+#ifndef INSPIRE_FACE_FACE_INFO_INTERNAL_H
+#define INSPIRE_FACE_FACE_INFO_INTERNAL_H
+
+#include <memory>
+#include <utility>
+#include <inspirecv/inspirecv.h>
+#include "middleware/utils.h"
+#include "data_type.h"
+#include "face_process.h"
+#include "face_action_data.h"
+#include "track_module/quality/face_pose_quality_adapt.h"
+#include "track_module/landmark/landmark_param.h"
+
+namespace inspire {
+
+enum ISF_TRACK_STATE { ISF_UNTRACKING = -1, ISF_DETECT = 0, ISF_READY = 1, ISF_TRACKING = 2 };
+
+class INSPIRE_API FaceObjectInternal {
+public:
+    FaceObjectInternal(int instance_id, inspirecv::Rect2i bbox, int num_landmark = 106) {
+        face_id_ = instance_id;
+        landmark_.resize(num_landmark);
+        bbox_ = std::move(bbox);
+        tracking_state_ = ISF_DETECT;
+        confidence_ = 1.0;
+        tracking_count_ = 0;
+        pose_euler_angle_.resize(3);
+        keyPointFive.resize(5);
+        face_action_ = std::make_shared<FaceActionPredictor>(10);
+        num_of_dense_landmark_ = num_landmark;
+    }
+
+    void SetLandmark(const std::vector<inspirecv::Point2f> &lmk, bool update_rect = true, bool update_matrix = true, float h = 0.06f, int n = 5,
+                     int num_of_lmk = 106 * 2) {
+        // if (lmk.size() != landmark_.size()) {
+        //     INSPIRE_LOGW("The SetLandmark function displays an exception indicating that the lmk number does not match");
+        //     return;
+        // }
+        std::copy(lmk.begin(), lmk.end(), landmark_.begin());
+        DynamicSmoothParamUpdate(landmark_, landmark_smooth_aux_, num_of_lmk, h, n);
+        // std::cout << "smooth ratio: " << h << " num smooth cache frame: " << n << std::endl;
+
+        // cv::Vec3d euler_angle;
+        // EstimateHeadPose(landmark_, euler_angle_);
+        // DynamicSmoothParamUpdate(landmark_, landmark_smooth_aux_, 106 * 2, 0.06);
+        if (update_rect)
+            bbox_ = inspirecv::MinBoundingRect(lmk).As<int>();
+        if (update_matrix && tracking_state_ == ISF_TRACKING) {
+            // pass
+        }
+
+        keyPointFive[0] = landmark_[55];
+        keyPointFive[1] = landmark_[105];
+        keyPointFive[2] = landmark_[69];
+        keyPointFive[3] = landmark_[45];
+        keyPointFive[4] = landmark_[50];
+    }
+
+    void setAlignMeanSquareError(const std::vector<inspirecv::Point2f> &lmk_5) {
+        float src_pts[] = {30.2946, 51.6963, 65.5318, 51.5014, 48.0252, 71.7366, 33.5493, 92.3655, 62.7299, 92.2041};
+        for (int i = 0; i < 5; i++) {
+            *(src_pts + 2 * i) += 8.0;
+        }
+        float sum = 0;
+        for (int i = 0; i < lmk_5.size(); i++) {
+            float l2 = L2norm(src_pts[i * 2 + 0], src_pts[i * 2 + 1], lmk_5[i].GetX(), lmk_5[i].GetY());
+            sum += l2;
+        }
+
+        align_mse_ = sum / 5.0f;
+    }
+
+    // Increment tracking count
+    void IncrementTrackingCount() {
+        tracking_count_++;
+    }
+
+    // Get tracking count
+    int GetTrackingCount() const {
+        return tracking_count_;
+    }
+
+    float GetAlignMSE() const {
+        return align_mse_;
+    }
+
+    std::vector<inspirecv::Point2f> GetLanmdark() const {
+        return landmark_;
+    }
+
+    inspirecv::Rect2i GetRect() const {
+        return bbox_;
+    }
+
+    inspirecv::Rect2i GetRectSquare(float padding_ratio = 0.0) const {
+        int cx = bbox_.GetX() + bbox_.GetWidth() / 2;
+        int cy = bbox_.GetY() + bbox_.GetHeight() / 2;
+        int R = std::max(bbox_.GetWidth(), bbox_.GetHeight()) / 2;
+        int R_padding = static_cast<int>(R * (1 + padding_ratio));
+        int x1 = cx - R_padding;
+        int y1 = cy - R_padding;
+        int x2 = cx + R_padding;
+        int y2 = cy + R_padding;
+        int width = x2 - x1;
+        int height = y2 - y1;
+        assert(width > 0);
+        assert(height > 0);
+        assert(height == width);
+        inspirecv::Rect2i box_square(x1, y1, width, height);
+        return box_square;
+    }
+
+    FaceActionList UpdateFaceAction(const SemanticIndex& semantic_index) {
+        inspirecv::Vec3f euler{high_result.pitch, high_result.yaw, high_result.roll};
+        inspirecv::Vec2f eyes{left_eye_status_.back(), right_eye_status_.back()};
+        face_action_->RecordActionFrame(landmark_, euler, eyes);
+        return face_action_->AnalysisFaceAction(semantic_index);
+    }
+
+    void DisableTracking() {
+        tracking_state_ = ISF_UNTRACKING;
+    }
+
+    void EnableTracking() {
+        tracking_state_ = ISF_TRACKING;
+    }
+
+    void ReadyTracking() {
+        tracking_state_ = ISF_READY;
+    }
+
+    ISF_TRACK_STATE TrackingState() const {
+        return tracking_state_;
+    }
+
+    float GetConfidence() const {
+        return confidence_;
+    }
+
+    void SetConfidence(float confidence) {
+        confidence_ = confidence;
+    }
+
+    int GetTrackingId() const {
+        return face_id_;
+    }
+
+    const inspirecv::TransformMatrix &getTransMatrix() const {
+        return trans_matrix_;
+    }
+
+    const inspirecv::TransformMatrix &getTransMatrixExtensive() const {
+        return trans_matrix_extensive_;
+    }
+
+    void setTransMatrix(const inspirecv::TransformMatrix &transMatrix) {
+        trans_matrix_ = transMatrix.Clone();
+    }
+
+    void setTransMatrixExtensive(const inspirecv::TransformMatrix &transMatrixExtensive) {
+        trans_matrix_extensive_ = transMatrixExtensive.Clone();
+    }
+
+    static float L2norm(float x0, float y0, float x1, float y1) {
+        return sqrt((x0 - x1) * (x0 - x1) + (y0 - y1) * (y0 - y1));
+    }
+
+    void DynamicSmoothParamUpdate(std::vector<inspirecv::Point2f> &landmarks, std::vector<std::vector<inspirecv::Point2f>> &landmarks_lastNframes,
+                                  int lm_length, float h = 0.06f, int n = 5) {
+        std::vector<inspirecv::Point2f> landmarks_temp;
+        landmarks_temp.assign(landmarks.begin(), landmarks.end());
+        if (landmarks_lastNframes.size() == n) {
+            for (int i = 0; i < lm_length / 2; i++) {
+                float sum_d = 1;
+                float max_d = 0;
+                for (int j = 0; j < n; j++) {
+                    float d = L2norm(landmarks_temp[i].GetX(), landmarks_temp[i].GetY(), landmarks_lastNframes[j][i].GetX(),
+                                     landmarks_lastNframes[j][i].GetY());
+                    if (d > max_d)
+                        max_d = d;
+                }
+                for (int j = 0; j < n; j++) {
+                    float d = exp(-max_d * (n - j) * h);
+                    sum_d += d;
+                    landmarks[i].SetX(landmarks[i].GetX() + d * landmarks_lastNframes[j][i].GetX());
+                    landmarks[i].SetY(landmarks[i].GetY() + d * landmarks_lastNframes[j][i].GetY());
+                }
+                landmarks[i].SetX(landmarks[i].GetX() / sum_d);
+                landmarks[i].SetY(landmarks[i].GetY() / sum_d);
+            }
+        }
+        std::vector<inspirecv::Point2f> landmarks_frame;
+        for (int i = 0; i < lm_length / 2; i++) {
+            landmarks_frame.push_back(inspirecv::Point2f(landmarks[i].GetX(), landmarks[i].GetY()));
+        }
+        landmarks_lastNframes.push_back(landmarks_frame);
+        if (landmarks_lastNframes.size() > n)
+            landmarks_lastNframes.erase(landmarks_lastNframes.begin());
+    }
+
+public:
+    std::vector<inspirecv::Point2f> landmark_;
+    std::vector<std::vector<inspirecv::Point2f>> landmark_smooth_aux_;
+    inspirecv::Rect2i bbox_;
+    inspirecv::Vec3f euler_angle_;
+    std::vector<float> pose_euler_angle_;
+
+    int num_of_dense_landmark_;
+
+    float align_mse_{};
+
+    const inspirecv::Vec3f &getEulerAngle() const {
+        return euler_angle_;
+    }
+
+    const std::vector<float> &getPoseEulerAngle() const {
+        return pose_euler_angle_;
+    }
+
+    void setPoseEulerAngle(const std::vector<float> &poseEulerAngle) {
+        pose_euler_angle_[0] = poseEulerAngle[0];
+        pose_euler_angle_[1] = poseEulerAngle[1];
+        pose_euler_angle_[2] = poseEulerAngle[2];
+
+        if (abs(pose_euler_angle_[0]) < 0.5 && abs(pose_euler_angle_[1]) < 0.48) {
+            is_standard_ = true;
+        }
+    }
+
+    bool isStandard() const {
+        return is_standard_;
+    }
+
+    const inspirecv::Rect2i &getBbox() const {
+        return bbox_;
+    }
+
+    void setBbox(const inspirecv::Rect2i &bbox) {
+        bbox_ = bbox;
+    }
+
+    std::vector<std::vector<float>> face_emotion_history_;
+    
+    inspirecv::TransformMatrix trans_matrix_;
+    inspirecv::TransformMatrix trans_matrix_extensive_;
+    float confidence_;
+    inspirecv::Rect2i detect_bbox_;
+    int tracking_count_;  // Tracking count
+
+    bool is_standard_;
+
+    FacePoseQualityAdaptResult high_result;
+
+    FaceProcess faceProcess;
+
+    std::vector<inspirecv::Point2f> keyPointFive;
+
+    void setId(int id) {
+        face_id_ = id;
+    }
+
+    std::vector<float> left_eye_status_;
+
+    std::vector<float> right_eye_status_;
+
+private:
+    ISF_TRACK_STATE tracking_state_;
+    std::shared_ptr<FaceActionPredictor> face_action_;
+    int face_id_;
+};
+
+typedef std::vector<FaceObjectInternal> FaceObjectInternalList;
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_FACE_INFO_INTERNAL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/common/face_info/face_process.h b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_process.h
new file mode 100644
index 000000000..ed46d3e2d
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/common/face_info/face_process.h
@@ -0,0 +1,53 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+// Include guard to prevent double inclusion of this header file
+#pragma once
+#ifndef INSPIRE_FACE_FACEPROCESS_H
+#define INSPIRE_FACE_FACEPROCESS_H
+
+// Include the necessary header file "data_type.h"
+#include "data_type.h"
+
+// Define the namespace "inspire" for encapsulation
+namespace inspire {
+
+/**
+ * Enumeration to represent different mask information.
+ */
+typedef enum MaskInfo {
+    UNKNOWN_MASK = -1,  ///< Unknown mask status
+    UNMASKED = 0,       ///< No mask
+    MASKED = 1,         ///< Wearing a mask
+} MaskInfo;
+
+/**
+ * Enumeration to represent different RGB liveness information.
+ */
+typedef enum RGBLivenessInfo {
+    UNKNOWN_RGB_LIVENESS = -1,  ///< Unknown RGB liveness status
+    LIVENESS_FAKE = 0,          ///< Fake liveness
+    LIVENESS_REAL = 1,          ///< Real liveness
+} RGBLivenessInfo;
+
+/**
+ * Class definition for FaceProcess.
+ */
+class INSPIRE_API FaceProcess {
+public:
+    /**
+     * Member variable to store mask information, initialized to UNKNOWN_MASK.
+     */
+    MaskInfo maskInfo = UNKNOWN_MASK;
+
+    /**
+     * Member variable to store RGB liveness information, initialized to UNKNOWN_RGB_LIVENESS.
+     */
+    RGBLivenessInfo rgbLivenessInfo = UNKNOWN_RGB_LIVENESS;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_FACEPROCESS_H
diff --git a/cpp-package/inspireface/cpp/inspireface/cost_time.h b/cpp-package/inspireface/cpp/inspireface/cost_time.h
new file mode 100644
index 000000000..4192c3f7b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/cost_time.h
@@ -0,0 +1,71 @@
+#ifndef COST_TIME_HPP
+#define COST_TIME_HPP
+
+#include <chrono>
+#include <iomanip>
+#include <sstream>
+#include "log.h"  // Assume log.h includes the INSPIRE_LOGI macro
+
+#ifndef INSPIRE_API
+#define INSPIRE_API
+#endif
+
+// Macro definition to control whether time cost calculations are enabled
+#ifdef ISF_ENABLE_COST_TIME
+
+// define the macro to calculate the time cost
+#define COST_TIME(id, precision) inspire::CostTime cost_time_##id(#id, __FILENAME__, __FUNCTION__, __LINE__, precision)
+
+// define the macro to calculate the time cost without precision
+#define COST_TIME_SIMPLE(id) inspire::CostTime cost_time_##id(#id, __FILENAME__, __FUNCTION__, __LINE__)
+
+#else
+
+// No operation, when ISF_ENABLE_COST_TIME is not defined
+#define COST_TIME(id, precision)
+
+// No operation, when ISF_ENABLE_COST_TIME is not defined
+#define COST_TIME_SIMPLE(id)
+
+#endif
+
+namespace inspire {
+
+/**
+ * @class CostTime
+ * @brief A class for measuring and logging the time taken by a block of code.
+ *
+ * This class is used to measure the time taken by a block of code and log the result.
+ * It uses the std::chrono library to measure the time taken.
+ */
+class INSPIRE_API CostTime {
+public:
+    // Constructor, records the start time
+    explicit CostTime(const char* id, const char* filename, const char* function, int line, int precision = 5)
+    : id_(id), filename_(filename), function_(function), line_(line), precision_(precision), start_time_(std::chrono::steady_clock::now()) {}
+
+    // Destructor, calculates and logs the time taken
+    ~CostTime() {
+        auto end_time = std::chrono::steady_clock::now();
+        std::chrono::duration<double> duration = end_time - start_time_;
+
+        // Use a string stream for formatted output, controlling decimal precision
+        std::ostringstream oss;
+        oss << std::fixed << std::setprecision(precision_) << duration.count();
+
+        // Log the time cost
+        INSPIRE_LOGI("CostTime: [%s] %s:%d %s took %s seconds", id_, filename_, line_, function_, oss.str().c_str());
+    }
+
+private:
+    const char* id_;
+    const char* filename_;
+    const char* function_;
+    int line_;
+    int precision_;
+    std::chrono::time_point<std::chrono::steady_clock> start_time_;
+};
+
+}  // namespace inspire
+
+#endif  // COST_TIME_HPP
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/engine/face_session.cpp b/cpp-package/inspireface/cpp/inspireface/engine/face_session.cpp
new file mode 100644
index 000000000..d49b82d24
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/engine/face_session.cpp
@@ -0,0 +1,518 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_session.h"
+#include <launch.h>
+#include <utility>
+#include "log.h"
+#include "herror.h"
+#include "middleware/utils.h"
+#include "recognition_module/dest_const.h"
+
+namespace inspire {
+
+FaceSession::FaceSession() = default;
+
+int32_t FaceSession::Configuration(DetectModuleMode detect_mode, int32_t max_detect_face, CustomPipelineParameter param, int32_t detect_level_px,
+                                   int32_t track_by_detect_mode_fps) {
+    m_detect_mode_ = detect_mode;
+    m_max_detect_face_ = max_detect_face;
+    m_parameter_ = param;
+    if (!INSPIREFACE_CONTEXT->isMLoad()) {
+        return HERR_ARCHIVE_NOT_LOAD;
+    }
+    if (INSPIREFACE_CONTEXT->getMArchive().QueryStatus() != SARC_SUCCESS) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+
+    m_face_track_ = std::make_shared<FaceTrackModule>(m_detect_mode_, m_max_detect_face_, 20, 192, detect_level_px, track_by_detect_mode_fps, true);
+    m_face_track_->Configuration(INSPIREFACE_CONTEXT->getMArchive(), "", m_parameter_.enable_face_pose || m_parameter_.enable_face_quality);
+    // SetDetectMode(m_detect_mode_);
+
+    m_face_recognition_ = std::make_shared<FeatureExtractionModule>(INSPIREFACE_CONTEXT->getMArchive(), m_parameter_.enable_recognition);
+    if (m_face_recognition_->QueryStatus() != HSUCCEED) {
+        return m_face_recognition_->QueryStatus();
+    }
+
+    m_face_pipeline_ = std::make_shared<FacePipelineModule>(INSPIREFACE_CONTEXT->getMArchive(), param.enable_liveness, param.enable_mask_detect,
+                                                            param.enable_face_attribute, param.enable_interaction_liveness, param.enable_face_emotion);
+    m_face_track_cost_ = std::make_shared<inspire::SpendTimer>("FaceTrack");
+
+    return HSUCCEED;
+}
+
+int32_t FaceSession::FaceDetectAndTrack(inspirecv::FrameProcess& process) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    if (m_enable_track_cost_spend_) {
+        m_face_track_cost_->Start();
+    }
+    m_detect_cache_.clear();
+    m_face_basic_data_cache_.clear();
+    m_face_rects_cache_.clear();
+    m_track_id_cache_.clear();
+    m_track_count_cache_.clear();
+    m_quality_results_cache_.clear();
+    m_roll_results_cache_.clear();
+    m_yaw_results_cache_.clear();
+    m_pitch_results_cache_.clear();
+    m_quality_score_results_cache_.clear();
+    m_react_left_eye_results_cache_.clear();
+    m_react_right_eye_results_cache_.clear();
+    m_face_emotion_results_cache_.clear();
+
+    m_action_normal_results_cache_.clear();
+    m_action_shake_results_cache_.clear();
+    m_action_blink_results_cache_.clear();
+    m_action_jaw_open_results_cache_.clear();
+    m_action_raise_head_results_cache_.clear();
+
+    m_quality_score_results_cache_.clear();
+    m_attribute_race_results_cache_.clear();
+    m_attribute_gender_results_cache_.clear();
+    m_det_confidence_cache_.clear();
+    if (m_face_track_ == nullptr) {
+        return HERR_SESS_TRACKER_FAILURE;
+    }
+    m_face_track_->UpdateStream(process);
+    for (int i = 0; i < m_face_track_->trackingFace.size(); ++i) {
+        auto& face = m_face_track_->trackingFace[i];
+        FaceTrackWrap data = FaceObjectInternalToHyperFaceData(face, i);
+        ByteArray byteArray;
+        auto ret = RunSerializeHyperFaceData(data, byteArray);
+        if (ret != HSUCCEED) {
+            return HERR_INVALID_SERIALIZATION_FAILED;
+        }
+        m_det_confidence_cache_.push_back(face.GetConfidence());
+        m_detect_cache_.push_back(byteArray);
+        m_track_id_cache_.push_back(face.GetTrackingId());
+        m_track_count_cache_.push_back(face.GetTrackingCount());
+        m_face_rects_cache_.push_back(data.rect);
+        m_quality_results_cache_.push_back(face.high_result);
+        m_roll_results_cache_.push_back(face.high_result.roll);
+        m_yaw_results_cache_.push_back(face.high_result.yaw);
+        m_pitch_results_cache_.push_back(face.high_result.pitch);
+        // Process quality scores
+        float avg = 0.0f;
+        for (int j = 0; j < 5; ++j) {
+            avg += data.quality[j];
+        }
+        avg /= 5.0f;
+        float quality_score = 1.0f - avg;  // reversal
+        m_quality_score_results_cache_.push_back(quality_score);
+    }
+    // ptr face_basic
+    m_face_basic_data_cache_.resize(m_face_track_->trackingFace.size());
+    for (int i = 0; i < m_face_basic_data_cache_.size(); ++i) {
+        auto& basic = m_face_basic_data_cache_[i];
+        basic.dataSize = m_detect_cache_[i].size();
+        basic.data = m_detect_cache_[i].data();
+    }
+    if (m_enable_track_cost_spend_) {
+        m_face_track_cost_->Stop();
+    }
+    //    LOGD("Track COST: %f", m_face_track_->GetTrackTotalUseTime());
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetLandmarkLoop(int32_t value) {
+    // TODO: implement this function
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetFaceDetectThreshold(float value) {
+    m_face_track_->SetDetectThreshold(value);
+    return HSUCCEED;
+}
+
+FaceObjectInternalList& FaceSession::GetTrackingFaceList() {
+    return m_face_track_->trackingFace;
+}
+
+const std::shared_ptr<FeatureExtractionModule>& FaceSession::FaceRecognitionModule() {
+    return m_face_recognition_;
+}
+
+const std::shared_ptr<FacePipelineModule>& FaceSession::PipelineModule() {
+    return m_face_pipeline_;
+}
+
+const int32_t FaceSession::GetNumberOfFacesCurrentlyDetected() const {
+    return m_face_track_->trackingFace.size();
+}
+
+int32_t FaceSession::FacesProcess(inspirecv::FrameProcess& process, const std::vector<FaceTrackWrap>& faces, const CustomPipelineParameter& param) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    m_mask_results_cache_.resize(faces.size(), -1.0f);
+    m_rgb_liveness_results_cache_.resize(faces.size(), -1.0f);
+    m_react_left_eye_results_cache_.resize(faces.size(), -1.0f);
+    m_react_right_eye_results_cache_.resize(faces.size(), -1.0f);
+    m_attribute_race_results_cache_.resize(faces.size(), -1);
+    m_attribute_gender_results_cache_.resize(faces.size(), -1);
+    m_attribute_age_results_cache_.resize(faces.size(), -1);
+    m_action_normal_results_cache_.resize(faces.size(), -1);
+    m_action_jaw_open_results_cache_.resize(faces.size(), -1);
+    m_action_blink_results_cache_.resize(faces.size(), -1);
+    m_action_raise_head_results_cache_.resize(faces.size(), -1);
+    m_action_shake_results_cache_.resize(faces.size(), -1);
+    m_face_emotion_results_cache_.resize(faces.size(), -1);
+    for (int i = 0; i < faces.size(); ++i) {
+        const auto& face = faces[i];
+        // RGB Liveness Detect
+        if (param.enable_liveness) {
+            auto ret = m_face_pipeline_->Process(process, face, PROCESS_RGB_LIVENESS);
+            if (ret != HSUCCEED) {
+                return ret;
+            }
+            m_rgb_liveness_results_cache_[i] = m_face_pipeline_->faceLivenessCache;
+        }
+        // Mask detection
+        if (param.enable_mask_detect) {
+            auto ret = m_face_pipeline_->Process(process, face, PROCESS_MASK);
+            if (ret != HSUCCEED) {
+                return ret;
+            }
+            m_mask_results_cache_[i] = m_face_pipeline_->faceMaskCache;
+        }
+        // Face attribute prediction
+        if (param.enable_face_attribute) {
+            auto ret = m_face_pipeline_->Process(process, face, PROCESS_ATTRIBUTE);
+            if (ret != HSUCCEED) {
+                return ret;
+            }
+            m_attribute_race_results_cache_[i] = m_face_pipeline_->faceAttributeCache[0];
+            m_attribute_gender_results_cache_[i] = m_face_pipeline_->faceAttributeCache[1];
+            m_attribute_age_results_cache_[i] = m_face_pipeline_->faceAttributeCache[2];
+        }
+
+        // Face interaction
+        if (param.enable_interaction_liveness) {
+            auto ret = m_face_pipeline_->Process(process, face, PROCESS_INTERACTION);
+            if (ret != HSUCCEED) {
+                return ret;
+            }
+            // Get eyes status
+            m_react_left_eye_results_cache_[i] = m_face_pipeline_->eyesStatusCache[0];
+            m_react_right_eye_results_cache_[i] = m_face_pipeline_->eyesStatusCache[1];
+            // Special handling:  ff it is a tracking state, it needs to be filtered
+            if (face.trackState > 0) {
+                auto idx = face.inGroupIndex;
+                if (idx < m_face_track_->trackingFace.size()) {
+                    auto& target = m_face_track_->trackingFace[idx];
+                    if (target.GetTrackingId() == face.trackId) {
+                        auto new_eye_left = EmaFilter(m_face_pipeline_->eyesStatusCache[0], target.left_eye_status_, 8, 0.2f);
+                        auto new_eye_right = EmaFilter(m_face_pipeline_->eyesStatusCache[1], target.right_eye_status_, 8, 0.2f);
+                        if (face.trackState > 1) {
+                            // The filtered value can be obtained only in the tracking state
+                            m_react_left_eye_results_cache_[i] = new_eye_left;
+                            m_react_right_eye_results_cache_[i] = new_eye_right;
+                        }
+                        const auto actions = target.UpdateFaceAction(INSPIREFACE_CONTEXT->getMArchive().GetLandmarkParam()->semantic_index);
+                        m_action_normal_results_cache_[i] = actions.normal;
+                        m_action_jaw_open_results_cache_[i] = actions.jawOpen;
+                        m_action_blink_results_cache_[i] = actions.blink;
+                        m_action_raise_head_results_cache_[i] = actions.raiseHead;
+                        m_action_shake_results_cache_[i] = actions.shake;
+                    } else {
+                        INSPIRE_LOGD(
+                          "Serialized objects cannot connect to trace objects in memory, and there may be some "
+                          "problems");
+                    }
+                } else {
+                    INSPIRE_LOGW(
+                      "The index of the trace object does not match the trace list in memory, and there may be some "
+                      "problems");
+                }
+            }
+        }
+        // Face emotion recognition
+        if (param.enable_face_emotion) {
+            auto ret = m_face_pipeline_->Process(process, face, PROCESS_FACE_EMOTION);
+            if (ret != HSUCCEED) {
+                return ret;
+            }
+            // Default mode
+            m_face_emotion_results_cache_[i] = argmax(m_face_pipeline_->faceEmotionCache.begin(), m_face_pipeline_->faceEmotionCache.end());
+            if (face.trackState > 0) {
+                // Tracking mode
+                auto idx = face.inGroupIndex;
+                if (idx < m_face_track_->trackingFace.size()) {
+                    auto& target = m_face_track_->trackingFace[idx];
+                    if (target.GetTrackingId() == face.trackId) {
+                        auto new_emotion = VectorEmaFilter(m_face_pipeline_->faceEmotionCache, target.face_emotion_history_, 8, 0.4f);
+                        m_face_emotion_results_cache_[i] = argmax(new_emotion.begin(), new_emotion.end());
+                    } else { 
+                        INSPIRE_LOGW(
+                          "Serialized objects cannot connect to trace objects in memory, and there may be some "
+                          "problems");
+                    }
+                } else {
+                    INSPIRE_LOGW(
+                      "The index of the trace object does not match the trace list in memory, and there may be some "
+                      "problems");
+                }
+            }
+        }
+    }
+
+    return 0;
+}
+
+const std::vector<ByteArray>& FaceSession::GetDetectCache() const {
+    return m_detect_cache_;
+}
+
+const std::vector<FaceBasicData>& FaceSession::GetFaceBasicDataCache() const {
+    return m_face_basic_data_cache_;
+}
+
+const std::vector<FaceRect>& FaceSession::GetFaceRectsCache() const {
+    return m_face_rects_cache_;
+}
+
+const std::vector<int32_t>& FaceSession::GetTrackIDCache() const {
+    return m_track_id_cache_;
+}
+
+const std::vector<int32_t>& FaceSession::GetTrackCountCache() const {
+    return m_track_count_cache_;
+}
+
+const std::vector<float>& FaceSession::GetRollResultsCache() const {
+    return m_roll_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetYawResultsCache() const {
+    return m_yaw_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetPitchResultsCache() const {
+    return m_pitch_results_cache_;
+}
+
+const std::vector<FacePoseQualityAdaptResult>& FaceSession::GetQualityResultsCache() const {
+    return m_quality_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetMaskResultsCache() const {
+    return m_mask_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetRgbLivenessResultsCache() const {
+    return m_rgb_liveness_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetFaceQualityScoresResultsCache() const {
+    return m_quality_score_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetFaceInteractionLeftEyeStatusCache() const {
+    return m_react_left_eye_results_cache_;
+}
+
+const std::vector<float>& FaceSession::GetFaceInteractionRightEyeStatusCache() const {
+    return m_react_right_eye_results_cache_;
+}
+
+const Embedded& FaceSession::GetFaceFeatureCache() const {
+    return m_face_feature_cache_;
+}
+
+const std::vector<float>& FaceSession::GetDetConfidenceCache() const {
+    return m_det_confidence_cache_;
+}
+
+const float FaceSession::GetFaceFeatureNormCache() const {
+    return m_face_feature_norm_;
+}
+
+const std::vector<int>& FaceSession::GetFaceRaceResultsCache() const {
+    return m_attribute_race_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceGenderResultsCache() const {
+    return m_attribute_gender_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceAgeBracketResultsCache() const {
+    return m_attribute_age_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceNormalAactionsResultCache() const {
+    return m_action_normal_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceJawOpenAactionsResultCache() const {
+    return m_action_jaw_open_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceBlinkAactionsResultCache() const {
+    return m_action_blink_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceShakeAactionsResultCache() const {
+    return m_action_shake_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceRaiseHeadAactionsResultCache() const {
+    return m_action_raise_head_results_cache_;
+}
+
+const std::vector<int>& FaceSession::GetFaceEmotionResultsCache() const {
+    return m_face_emotion_results_cache_;
+}
+
+int32_t FaceSession::FaceFeatureExtract(inspirecv::FrameProcess& process, FaceBasicData& data, bool normalize) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    int32_t ret;
+    FaceTrackWrap face = {0};
+    ret = RunDeserializeHyperFaceData((char*)data.data, data.dataSize, face);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    m_face_feature_cache_.clear();
+    ret = m_face_recognition_->FaceExtract(process, face, m_face_feature_cache_, m_face_feature_norm_, normalize);
+
+    return ret;
+}
+
+int32_t FaceSession::FaceFeatureExtract(inspirecv::FrameProcess& process, FaceTrackWrap& data, bool normalize) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    int32_t ret;
+    m_face_feature_cache_.clear();
+    ret = m_face_recognition_->FaceExtract(process, data, m_face_feature_cache_, m_face_feature_norm_, normalize);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+
+    return ret;
+}
+
+int32_t FaceSession::FaceFeatureExtractWithAlignmentImage(inspirecv::FrameProcess& process, Embedded& embedding, float& norm, bool normalize) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    int32_t ret;
+    m_face_feature_cache_.clear();
+    ret = m_face_recognition_->FaceExtractWithAlignmentImage(process, embedding, norm, normalize);
+
+    return ret;
+}
+
+int32_t FaceSession::FaceFeatureExtractWithAlignmentImage(const inspirecv::Image& wrapped, FaceEmbedding& embedding, float& norm, bool normalize) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    int32_t ret;
+    ret = m_face_recognition_->FaceExtractWithAlignmentImage(wrapped, embedding.embedding, norm, normalize);
+    return ret;
+}
+
+int32_t FaceSession::FaceGetFaceAlignmentImage(inspirecv::FrameProcess& process, FaceBasicData& data, inspirecv::Image& image) {
+    std::lock_guard<std::mutex> lock(m_mtx_);
+    int32_t ret;
+    FaceTrackWrap face = {0};
+    ret = RunDeserializeHyperFaceData((char*)data.data, data.dataSize, face);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    std::vector<inspirecv::Point2f> pointsFive;
+    for (const auto& p : face.keyPoints) {
+        pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+    }
+    auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+    image = process.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+    return ret;
+}
+
+const CustomPipelineParameter& FaceSession::getMParameter() const {
+    return m_parameter_;
+}
+
+int32_t FaceSession::FaceQualityDetect(FaceBasicData& data, float& result) {
+    int32_t ret;
+    FaceTrackWrap face = {0};
+    ret = RunDeserializeHyperFaceData((char*)data.data, data.dataSize, face);
+    //    PrintHyperFaceData(face);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    float avg = 0.0f;
+    for (int i = 0; i < 5; ++i) {
+        avg += face.quality[i];
+    }
+    avg /= 5.0f;
+    result = 1.0f - avg;  // reversal
+
+    return ret;
+}
+
+int32_t FaceSession::SetDetectMode(DetectModuleMode mode) {
+    m_detect_mode_ = mode;
+    if (m_detect_mode_ == DetectModuleMode::DETECT_MODE_ALWAYS_DETECT) {
+        m_always_detect_ = true;
+    } else {
+        m_always_detect_ = false;
+    }
+    return HSUCCEED;
+}
+
+bool FaceSession::IsDetectModeLandmark() const {
+    return m_face_track_->IsDetectModeLandmark();
+}
+
+void FaceSession::ClearTrackingFace() {
+    m_face_track_->ClearTrackingFace();
+}
+
+void FaceSession::SetTrackLostRecoveryMode(bool value) {
+    m_face_track_->SetTrackLostRecoveryMode(value);
+}
+
+void FaceSession::SetLightTrackConfidenceThreshold(float value) {
+    m_face_track_->SetLightTrackConfidenceThreshold(value);
+}
+
+int32_t FaceSession::SetTrackPreviewSize(const int32_t preview_size) {
+    m_face_track_->SetTrackPreviewSize(preview_size);
+    return HSUCCEED;
+}
+
+int32_t FaceSession::GetTrackPreviewSize() const {
+    return m_face_track_->GetTrackPreviewSize();
+}
+
+int32_t FaceSession::SetTrackFaceMinimumSize(int32_t minSize) {
+    m_face_track_->SetMinimumFacePxSize(minSize);
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetTrackModeSmoothRatio(float value) {
+    m_face_track_->SetTrackModeSmoothRatio(value);
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetTrackModeNumSmoothCacheFrame(int value) {
+    m_face_track_->SetTrackModeNumSmoothCacheFrame(value);
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetTrackModeDetectInterval(int value) {
+    m_face_track_->SetTrackModeDetectInterval(value);
+    return HSUCCEED;
+}
+
+int32_t FaceSession::SetEnableTrackCostSpend(int value) {
+    m_enable_track_cost_spend_ = value;
+    m_face_track_cost_->Reset();
+    return HSUCCEED;
+}
+
+void FaceSession::PrintTrackCostSpend() {
+    if (m_enable_track_cost_spend_) {
+        INSPIRE_LOGI("%s", m_face_track_cost_->Report().c_str());
+    }
+}
+
+int32_t FaceSession::GetDebugPreviewImageSize() const {
+    return m_face_track_->GetDebugPreviewImageSize();
+}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/engine/face_session.h b/cpp-package/inspireface/cpp/inspireface/engine/face_session.h
new file mode 100644
index 000000000..be5413c20
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/engine/face_session.h
@@ -0,0 +1,459 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#include <vector>
+#ifndef INSPIRE_FACE_CONTEXT_H
+#define INSPIRE_FACE_CONTEXT_H
+
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include "data_type.h"
+#include "track_module/face_track_module.h"
+#include "pipeline_module/face_pipeline_module.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "recognition_module/face_feature_extraction_module.h"
+#include "frame_process.h"
+#include "common/face_data/face_serialize_tools.h"
+#include "spend_timer.h"
+
+namespace inspire {
+
+/**
+ * @class FaceContext
+ * @brief Manages the context for face detection, tracking, and feature extraction in the HyperFaceRepo project.
+ *
+ * Provides interfaces to configure face detection modes, manage face tracking, perform recognition,
+ * and handle other face-related features. Integrates with various modules such as FaceTrack, FaceRecognition, and
+ * FacePipeline.
+ */
+class INSPIRE_API FaceSession {
+public:
+    /**
+     * @brief Constructor for the FaceContext class.
+     */
+    explicit FaceSession();
+
+    /**
+     * @brief Configures the face context with given parameters.
+     * @param model_file_path Path to the model file for face detection.
+     * @param detect_mode The detection mode to be used (image or video).
+     * @param max_detect_face The maximum number of faces to detect.
+     * @param param Custom parameters for the face pipeline.
+     * @return int32_t Returns 0 on success, non-zero for any error.
+     */
+    int32_t Configuration(DetectModuleMode detect_mode, int32_t max_detect_face, CustomPipelineParameter param, int32_t detect_level_px = -1,
+                          int32_t track_by_detect_mode_fps = -1);
+
+    /**
+     * @brief Performs face detection and tracking on a given image stream.
+     * @param image The camera stream to process for face detection and tracking.
+     * @return int32_t Returns the number of faces detected and tracked.
+     */// Method for face detection and tracking
+    int32_t FaceDetectAndTrack(inspirecv::FrameProcess& process);
+
+    /**
+     * @brief Set the face landmark loop
+     * @param value The landmark loop value
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetLandmarkLoop(int32_t value);
+
+    /**
+     * @brief Set the threshold of face detection function, which only acts on the detection model
+     * @param value threshold value
+     * @return int32_t Returns the number of faces detected and tracked.
+     * */
+    int32_t SetFaceDetectThreshold(float value);
+
+    /**
+     * @brief Retrieves the list of currently tracked faces.
+     * @return FaceObjectList A list of face objects currently being tracked.
+     */
+    FaceObjectInternalList& GetTrackingFaceList();
+
+    /**
+     * @brief Processes faces using the provided pipeline parameters.
+     * @param image Camera stream containing faces.
+     * @param faces Vector of FaceTrackWrap for detected faces.
+     * @param param Custom pipeline parameters.
+     * @return int32_t Status code of the processing.
+     */
+    int32_t FacesProcess(inspirecv::FrameProcess& process, const std::vector<FaceTrackWrap>& faces, const CustomPipelineParameter& param);
+
+    /**
+     * @brief Retrieves the face recognition module.
+     * @return std::shared_ptr<FaceRecognition> Shared pointer to the FaceRecognition module.
+     */
+    const std::shared_ptr<FeatureExtractionModule>& FaceRecognitionModule();
+
+    /**
+     * @brief Retrieves the face pipeline module.
+     * @return std::shared_ptr<FacePipeline> Shared pointer to the FacePipeline module.
+     */
+    const std::shared_ptr<FacePipelineModule>& PipelineModule();
+
+    /**
+     * @brief Gets the number of faces currently detected.
+     * @return int32_t Number of faces currently detected.
+     */
+    const int32_t GetNumberOfFacesCurrentlyDetected() const;
+
+    /**
+     * @brief Extracts features of a face from an image.
+     * @param image Camera stream containing the face.
+     * @param data FaceBasicData to store extracted features.
+     * @return int32_t Status code of the feature extraction.
+     */
+    int32_t FaceFeatureExtract(inspirecv::FrameProcess& process, FaceBasicData& data, bool normalize = true);
+
+    /**
+     * @brief Extracts features of a face from an image.
+     * @param image Camera stream containing the face.
+     * @param data FaceTrackWrap to store extracted features.
+     * @return int32_t Status code of the feature extraction.
+     */
+    int32_t FaceFeatureExtract(inspirecv::FrameProcess& process, FaceTrackWrap& data, bool normalize = true);
+
+    /**
+     * @brief Extracts features of a face from an image.
+     * @param image Camera stream containing the face.
+     * @param data FaceTrackWrap to store extracted features.
+     * @return int32_t Status code of the feature extraction.
+     */
+    int32_t FaceFeatureExtractWithAlignmentImage(inspirecv::FrameProcess& process, Embedded& embedding, float& norm, bool normalize = true);
+
+    /**
+     * @brief Extracts features of a face from an image.
+     * @param image Camera stream containing the face.
+     * @param data FaceTrackWrap to store extracted features.
+     * @return int32_t Status code of the feature extraction.
+     */
+    int32_t FaceFeatureExtractWithAlignmentImage(const inspirecv::Image& wrapped, FaceEmbedding& embedding, float& norm, bool normalize = true);
+
+    /**
+     * @brief Gets the face alignment image.
+     * @param process The image process object.
+     * @param data The face basic data.
+     * @param image The output image.
+     * @return int32_t The status code of the operation.
+     */
+    int32_t FaceGetFaceAlignmentImage(inspirecv::FrameProcess& process, FaceBasicData& data, inspirecv::Image& image);
+
+    /**
+     * @brief Retrieves the custom pipeline parameters.
+     * @return CustomPipelineParameter Current custom pipeline parameters.
+     */
+    const CustomPipelineParameter& getMParameter() const;
+
+    /**
+     * @brief Static method for detecting face quality.
+     * @param data FaceBasicData containing the face information.
+     * @param result Float to store the face quality result.
+     * @return int32_t Status code of the quality detection.
+     */
+    static int32_t FaceQualityDetect(FaceBasicData& data, float& result);
+
+    /**
+     * @brief Sets the preview size for face tracking.
+     * @param preview_size Integer specifying the new preview size.
+     * @return int32_t Status code of the operation.
+     */
+    int32_t SetTrackPreviewSize(int32_t preview_size);
+
+    /**
+     * @brief Gets the preview size for face tracking.
+     * @return int32_t The preview size.
+     */
+    int32_t GetTrackPreviewSize() const;
+
+    /**
+     * @brief Filter the minimum face pixel size.
+     * @param minSize The minimum pixel value.
+     * @return int32_t Status code of the operation.
+     */
+    int32_t SetTrackFaceMinimumSize(int32_t minSize);
+
+    /**
+     * @brief Sets the mode for face detection.
+     * @param mode You can select mode for track or detect.
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetDetectMode(DetectModuleMode mode);
+
+    /**
+     * @brief Check if landmark detection is enabled in detection mode.
+     * @return True if landmark detection is enabled, false otherwise.
+     */
+    bool IsDetectModeLandmark() const;
+
+    /**
+     * @brief Clear the tracking face
+     */
+    void ClearTrackingFace();
+
+    /**
+     * @brief Set the track lost recovery mode
+     * @param value Track lost recovery mode
+     */
+    void SetTrackLostRecoveryMode(bool value);
+
+    /**
+     * @brief Set the light track confidence threshold
+     * @param value Light track confidence threshold
+     */
+    void SetLightTrackConfidenceThreshold(float value);
+
+public:
+    // Accessor methods for various cached data
+    /**
+     * @brief Retrieves the cache of detected face data.
+     * @return std::vector<ByteArray> Cache of detected face data.
+     */
+    const std::vector<ByteArray>& GetDetectCache() const;
+
+    /**
+     * @brief Retrieves the cache of basic face data.
+     * @return std::vector<FaceBasicData> Cache of basic face data.
+     */
+    const std::vector<FaceBasicData>& GetFaceBasicDataCache() const;
+
+    /**
+     * @brief Retrieves the cache of face rectangles.
+     * @return std::vector<FaceRect> Cache of face rectangles.
+     */
+    const std::vector<FaceRect>& GetFaceRectsCache() const;
+
+    /**
+     * @brief Retrieves the cache of tracking IDs.
+     * @return std::vector<int32_t> Cache of tracking IDs.
+     */
+    const std::vector<int32_t>& GetTrackIDCache() const;
+
+    /**
+     * @brief Retrieves the cache of tracking count.
+     * @return std::vector<int32_t> Cache of tracking count.
+     */
+    const std::vector<int32_t>& GetTrackCountCache() const; 
+
+    /**
+     * @brief Retrieves the cache of roll results from face pose estimation.
+     * @return std::vector<float> Cache of roll results.
+     */
+    const std::vector<float>& GetRollResultsCache() const;
+
+    /**
+     * @brief Retrieves the cache of yaw results from face pose estimation.
+     * @return std::vector<float> Cache of yaw results.
+     */
+    const std::vector<float>& GetYawResultsCache() const;
+
+    /**
+     * @brief Gets the cache of pitch results from face pose estimation.
+     * @return A const reference to a vector containing pitch results.
+     */
+    const std::vector<float>& GetPitchResultsCache() const;
+
+    /**
+     * @brief Gets the cache of face pose quality results.
+     * @return A const reference to a vector of FacePoseQualityResult objects.
+     */
+    const std::vector<FacePoseQualityAdaptResult>& GetQualityResultsCache() const;
+
+    /**
+     * @brief Gets the cache of mask detection results.
+     * @return A const reference to a vector containing mask detection results.
+     */
+    const std::vector<float>& GetMaskResultsCache() const;
+
+    /**
+     * @brief Gets the cache of RGB liveness detection results.
+     * @return A const reference to a vector containing RGB liveness results.
+     */
+    const std::vector<float>& GetRgbLivenessResultsCache() const;
+
+    /**
+     * @brief Gets the cache of face quality predict results.
+     * @return A const reference to a vector containing face quality predict results.
+     */
+    const std::vector<float>& GetFaceQualityScoresResultsCache() const;
+
+    /**
+     * @brief Gets the cache of left eye status predict results.
+     * @return A const reference to a vector containing eye status predict results.
+     */
+    const std::vector<float>& GetFaceInteractionLeftEyeStatusCache() const;
+
+    /**
+     * @brief Gets the cache of right eye status predict results.
+     * @return A const reference to a vector containing eye status predict results.
+     */
+    const std::vector<float>& GetFaceInteractionRightEyeStatusCache() const;
+
+    /**
+     * @brief Gets the cache of face attribute rece results.
+     * @return A const reference to a vector containing face attribute rece results.
+     */
+    const std::vector<int>& GetFaceRaceResultsCache() const;
+
+    /**
+     * @brief Gets the cache of face attribute gender results.
+     * @return A const reference to a vector containing face attribute gender results.
+     */
+    const std::vector<int>& GetFaceGenderResultsCache() const;
+
+    /**
+     * @brief Gets the cache of face attribute age bracket results.
+     * @return A const reference to a vector containing face attribute age bracket results.
+     */
+    const std::vector<int>& GetFaceAgeBracketResultsCache() const;
+
+    /**
+     * @brief Gets the cache of face action normal results.
+     * @return A const reference to a vector containing face action normal results.
+     */
+    const std::vector<int>& GetFaceNormalAactionsResultCache() const;
+
+    /**
+     * @brief Gets the cache of face action jaw open results.
+     * @return A const reference to a vector containing face action jaw open results.
+     */
+    const std::vector<int>& GetFaceJawOpenAactionsResultCache() const;
+
+    /**
+     * @brief Gets the cache of face action blink results.
+     * @return A const reference to a vector containing face action blink results.
+     */
+    const std::vector<int>& GetFaceBlinkAactionsResultCache() const;
+
+    /**
+     * @brief Gets the cache of face action shake results.
+     * @return A const reference to a vector containing face action shake results.
+     */
+    const std::vector<int>& GetFaceShakeAactionsResultCache() const;
+
+    /**
+     * @brief Gets the cache of face action raise head results.
+     * @return A const reference to a vector containing face action raise head results.
+     */
+    const std::vector<int>& GetFaceRaiseHeadAactionsResultCache() const;
+
+    /**
+     * @brief Gets the cache of face emotion results.
+     * @return A const reference to a vector containing face emotion results.
+     */
+    const std::vector<int>& GetFaceEmotionResultsCache() const;
+
+    /**
+     * @brief Gets the cache of the current face features.
+     * @return A const reference to the Embedded object containing current face feature data.
+     */
+    const Embedded& GetFaceFeatureCache() const;
+
+    /**
+     * @brief Gets the cache of face detection confidence.
+     * @return A const reference to a vector containing face detection confidence.
+     */
+    const std::vector<float>& GetDetConfidenceCache() const;
+
+    /**
+     * @brief Gets the cache of face feature norm.
+     * @return A const reference to a float containing face feature norm.
+     */
+    const float GetFaceFeatureNormCache() const;
+
+    /**
+     * @brief Set the track mode smooth ratio
+     * @param value The smooth ratio value
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetTrackModeSmoothRatio(float value);
+
+    /**
+     * @brief Set the track mode num smooth cache frame
+     * @param value The num smooth cache frame value
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetTrackModeNumSmoothCacheFrame(int value);
+
+    /**
+     * @brief Set the track model detect interval
+     * @param value The detect interval value
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetTrackModeDetectInterval(int value);
+
+    /**
+     * @brief Set the enable cost spend
+     * @param value The enable cost spend value
+     * @return int32_t Status code of the operation.
+     * */
+    int32_t SetEnableTrackCostSpend(int value);
+
+    /**
+     * @brief Print the cost spend
+     * */
+    void PrintTrackCostSpend();
+
+    /**
+     * @brief Get the debug preview image size
+     * @return int32_t The debug preview image size
+     * */
+    int32_t GetDebugPreviewImageSize() const;
+
+private:
+    // Private member variables
+    CustomPipelineParameter m_parameter_;  ///< Stores custom parameters for the pipeline
+    int32_t m_max_detect_face_{};          ///< Maximum number of faces that can be detected
+    DetectModuleMode m_detect_mode_;       ///< Current detection mode (image or video)
+    bool m_always_detect_{};               ///< Flag to determine if detection should always occur
+
+    std::shared_ptr<FaceTrackModule> m_face_track_;                ///< Shared pointer to the FaceTrack object
+    std::shared_ptr<FeatureExtractionModule> m_face_recognition_;  ///< Shared pointer to the FaceRecognition object
+    std::shared_ptr<FacePipelineModule> m_face_pipeline_;          ///< Shared pointer to the FacePipeline object
+
+private:
+    // Cache data
+    std::vector<ByteArray> m_detect_cache_;                            ///< Cache for storing serialized detected face data
+    std::vector<FaceBasicData> m_face_basic_data_cache_;               ///< Cache for basic face data extracted from detection
+    std::vector<FaceRect> m_face_rects_cache_;                         ///< Cache for face rectangle data from detection
+    std::vector<int32_t> m_track_id_cache_;                            ///< Cache for tracking IDs of detected faces
+    std::vector<int32_t> m_track_count_cache_;                         ///< Cache for tracking count of detected faces
+    std::vector<float> m_det_confidence_cache_;                        ///< Cache for face detection confidence of detected faces
+    std::vector<float> m_roll_results_cache_;                          ///< Cache for storing roll results from face pose estimation
+    std::vector<float> m_yaw_results_cache_;                           ///< Cache for storing yaw results from face pose estimation
+    std::vector<float> m_pitch_results_cache_;                         ///< Cache for storing pitch results from face pose estimation
+    std::vector<FacePoseQualityAdaptResult> m_quality_results_cache_;  ///< Cache for face pose quality results
+    std::vector<float> m_mask_results_cache_;                          ///< Cache for mask detection results
+    std::vector<float> m_rgb_liveness_results_cache_;                  ///< Cache for RGB liveness detection results
+    std::vector<float> m_quality_score_results_cache_;                 ///< Cache for RGB face quality score results
+    std::vector<float> m_react_left_eye_results_cache_;                ///< Cache for Left eye state in face interaction
+    std::vector<float> m_react_right_eye_results_cache_;               ///< Cache for Right eye state in face interaction
+
+    std::vector<int> m_action_normal_results_cache_;      ///< Cache for normal action in face interaction
+    std::vector<int> m_action_shake_results_cache_;       ///< Cache for shake action in face interaction
+    std::vector<int> m_action_blink_results_cache_;       ///< Cache for blink action in face interaction
+    std::vector<int> m_action_jaw_open_results_cache_;    ///< Cache for jaw open action in face interaction
+    std::vector<int> m_action_raise_head_results_cache_;  ///< Cache for raise head action in face interaction
+
+    std::vector<int> m_attribute_race_results_cache_;    ///< Cache for face attribute race results
+    std::vector<int> m_attribute_gender_results_cache_;  ///< Cache for face attribute gender results
+    std::vector<int> m_attribute_age_results_cache_;     ///< Cache for face attribute age results
+
+    std::vector<int> m_face_emotion_results_cache_;      ///< Cache for face emotion classification results
+
+    Embedded m_face_feature_cache_;                      ///< Cache for current face feature data
+    float m_face_feature_norm_;                          ///< Cache for face feature norm
+
+    std::mutex m_mtx_;  ///< Mutex for thread safety.
+
+    // cost spend
+    std::shared_ptr<inspire::SpendTimer> m_face_track_cost_;
+
+    int m_enable_track_cost_spend_ = 0;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_CONTEXT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.cpp b/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.cpp
new file mode 100644
index 000000000..1abf4d6d2
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.cpp
@@ -0,0 +1,360 @@
+#include "embedding_db.h"
+#include "sqlite-vec.h"
+#include "isf_check.h"
+#include <algorithm>
+#if defined(__ANDROID__)
+#include <android/log.h>
+#endif
+
+namespace inspire {
+
+std::unique_ptr<EmbeddingDB> EmbeddingDB::instance_ = nullptr;
+std::mutex EmbeddingDB::instanceMutex_;
+
+EmbeddingDB &EmbeddingDB::GetInstance() {
+    std::lock_guard<std::mutex> lock(instanceMutex_);
+    INSPIREFACE_CHECK_MSG(instance_, "EmbeddingDB not initialized. Call Init() first.");
+    return *instance_;
+}
+
+void EmbeddingDB::Init(const std::string &dbPath, size_t vectorDim, IdMode idMode) {
+    std::lock_guard<std::mutex> lock(instanceMutex_);
+    if (instance_) {
+        INSPIRE_LOGW("EmbeddingDB already initialized, skipping duplicate initialization");
+        return;
+    }
+    instance_.reset(new EmbeddingDB(dbPath, vectorDim, "cosine", idMode));
+}
+
+EmbeddingDB::EmbeddingDB(const std::string &dbPath, size_t vectorDim, const std::string &distanceMetric, IdMode idMode)
+: vectorDim_(vectorDim), tableName_("vec_items"), idMode_(idMode) {
+    int rc = sqlite3_auto_extension((void (*)())sqlite3_vec_init);
+    CheckSQLiteError(rc, nullptr);
+
+    // Open database
+    rc = sqlite3_open(dbPath.c_str(), &db_);
+    CheckSQLiteError(rc, db_);
+
+    // Create vector table
+    std::string createTableSQL = "CREATE VIRTUAL TABLE IF NOT EXISTS " + tableName_ + " USING vec0(embedding float[" + std::to_string(vectorDim_) +
+                                 "] distance_metric=" + distanceMetric + ")";
+
+    ExecuteSQL(createTableSQL);
+    initialized_ = true;
+}
+
+EmbeddingDB::~EmbeddingDB() {
+    if (db_) {
+        sqlite3_close(db_);
+    }
+}
+
+bool EmbeddingDB::InsertVector(const std::vector<float> &vector, int64_t &allocId) {
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    return InsertVector(0, vector, allocId);  // In auto-increment mode, the passed ID is ignored
+}
+
+bool EmbeddingDB::InsertVector(int64_t id, const std::vector<float> &vector, int64_t &allocId) {
+    CheckVectorDimension(vector);
+
+    sqlite3_stmt *stmt;
+    std::string sql;
+
+    if (idMode_ == IdMode::AUTO_INCREMENT) {
+        sql = "INSERT INTO " + tableName_ + "(embedding) VALUES (?)";
+    } else {
+        sql = "INSERT INTO " + tableName_ + "(rowid, embedding) VALUES (?, ?)";
+    }
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    // CheckSQLiteError(rc, db_);
+    if (rc != SQLITE_OK) {
+        INSPIRE_LOGE("Failed to prepare statement: %s", sqlite3_errmsg(db_));
+        sqlite3_finalize(stmt);
+        return false;
+    }
+
+    if (idMode_ == IdMode::AUTO_INCREMENT) {
+        sqlite3_bind_blob(stmt, 1, vector.data(), vector.size() * sizeof(float), SQLITE_STATIC);
+    } else {
+        sqlite3_bind_int64(stmt, 1, id);
+        sqlite3_bind_blob(stmt, 2, vector.data(), vector.size() * sizeof(float), SQLITE_STATIC);
+    }
+
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    if (rc != SQLITE_DONE) {
+        INSPIRE_LOGE("Failed to insert vector: %s", sqlite3_errmsg(db_));
+        return false;
+    }
+    // CheckSQLiteError(rc == SQLITE_DONE ? SQLITE_OK : rc, db_);
+
+    allocId = idMode_ == IdMode::AUTO_INCREMENT ? GetLastInsertRowId() : id;
+    return true;
+}
+
+std::vector<float> EmbeddingDB::GetVector(int64_t id) const {
+    std::lock_guard<std::mutex> lock(dbMutex_);
+
+    sqlite3_stmt *stmt;
+    std::string sql = "SELECT embedding FROM " + tableName_ + " WHERE rowid = ?";
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    sqlite3_bind_int64(stmt, 1, id);
+
+    rc = sqlite3_step(stmt);
+    if (rc != SQLITE_ROW) {
+        sqlite3_finalize(stmt);
+        // throw std::runtime_error("Vector with id " + std::to_string(id) + " not found");
+        return {};
+    }
+
+    const float *blob_data = static_cast<const float *>(sqlite3_column_blob(stmt, 0));
+    size_t blob_size = sqlite3_column_bytes(stmt, 0) / sizeof(float);
+    std::vector<float> result(blob_data, blob_data + blob_size);
+
+    sqlite3_finalize(stmt);
+    return result;
+}
+
+std::vector<int64_t> EmbeddingDB::BatchInsertVectors(const std::vector<VectorData> &vectors) {
+    ExecuteSQL("BEGIN");
+    std::vector<int64_t> insertedIds;
+    insertedIds.reserve(vectors.size());
+
+    for (const auto &data : vectors) {
+        int64_t id = 0;
+        bool ret = InsertVector(data.id, data.vector, id);
+        INSPIREFACE_CHECK_MSG(ret, "Failed to insert vector");
+        insertedIds.push_back(id);
+    }
+    ExecuteSQL("COMMIT");
+
+    return insertedIds;
+}
+
+std::vector<int64_t> EmbeddingDB::BatchInsertVectors(const std::vector<std::vector<float>> &vectors) {
+    ExecuteSQL("BEGIN");
+    std::vector<int64_t> insertedIds;
+    insertedIds.reserve(vectors.size());
+
+    for (const auto &vector : vectors) {
+        int64_t id = 0;
+        bool ret = InsertVector(0, vector, id);
+        INSPIREFACE_CHECK_MSG(ret, "Failed to insert vector");
+        insertedIds.push_back(id);
+    }
+    ExecuteSQL("COMMIT");
+
+    return insertedIds;
+}
+
+int64_t EmbeddingDB::GetLastInsertRowId() const {
+    return sqlite3_last_insert_rowid(db_);
+}
+
+void EmbeddingDB::UpdateVector(int64_t id, const std::vector<float> &newVector) {
+    CheckVectorDimension(newVector);
+
+    sqlite3_stmt *stmt;
+    std::string sql = "UPDATE " + tableName_ + " SET embedding = ? WHERE rowid = ?";
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    sqlite3_bind_blob(stmt, 1, newVector.data(), newVector.size() * sizeof(float), SQLITE_STATIC);
+    sqlite3_bind_int64(stmt, 2, id);
+
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    INSPIREFACE_CHECK_MSG(rc == SQLITE_DONE, "Failed to update vector");
+    if (sqlite3_changes(db_) == 0) {
+        INSPIRE_LOGF("Vector with id %ld not found", id);
+    }
+}
+
+void EmbeddingDB::DeleteVector(int64_t id) {
+    sqlite3_stmt *stmt;
+    std::string sql = "DELETE FROM " + tableName_ + " WHERE rowid = ?";
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    sqlite3_bind_int64(stmt, 1, id);
+
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    CheckSQLiteError(rc == SQLITE_DONE ? SQLITE_OK : rc, db_);
+}
+
+std::vector<FaceSearchResult> EmbeddingDB::SearchSimilarVectors(const std::vector<float> &queryVector, size_t top_k, float keep_similar_threshold,
+                                                                bool return_feature) {
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    CheckVectorDimension(queryVector);
+
+    sqlite3_stmt *stmt;
+    std::string sql;
+    if (return_feature) {
+        sql =
+          "SELECT rowid, embedding, 1.0 - distance as similarity "
+          "FROM " +
+          tableName_ +
+          " "
+          "WHERE embedding MATCH ? "
+          "ORDER BY distance "
+          "LIMIT ?";
+    } else {
+        sql =
+          "SELECT rowid, 1.0 - distance as similarity "
+          "FROM " +
+          tableName_ +
+          " "
+          "WHERE embedding MATCH ? "
+          "ORDER BY distance "
+          "LIMIT ?";
+    }
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    sqlite3_bind_blob(stmt, 1, queryVector.data(), queryVector.size() * sizeof(float), SQLITE_STATIC);
+    sqlite3_bind_int64(stmt, 2, top_k);
+
+    std::vector<FaceSearchResult> results;
+    while ((rc = sqlite3_step(stmt)) == SQLITE_ROW) {
+        FaceSearchResult result;
+        result.id = sqlite3_column_int64(stmt, 0);
+        if (return_feature) {
+            const float *blob_data = static_cast<const float *>(sqlite3_column_blob(stmt, 1));
+            size_t blob_size = sqlite3_column_bytes(stmt, 1) / sizeof(float);
+            result.feature.assign(blob_data, blob_data + blob_size);
+            result.similarity = sqlite3_column_double(stmt, 2);
+        } else {
+            result.similarity = sqlite3_column_double(stmt, 1);
+        }
+        results.push_back(result);
+    }
+
+    sqlite3_finalize(stmt);
+    CheckSQLiteError(rc == SQLITE_DONE ? SQLITE_OK : rc, db_);
+
+    // Filter results whose similarity is below the threshold
+    results.erase(std::remove_if(results.begin(), results.end(),
+                                 [keep_similar_threshold](const FaceSearchResult &result) { return result.similarity < keep_similar_threshold; }),
+                  results.end());
+
+    return results;
+}
+
+int64_t EmbeddingDB::GetVectorCount() const {
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    sqlite3_stmt *stmt;
+    std::string sql = "SELECT COUNT(*) FROM " + tableName_;
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    rc = sqlite3_step(stmt);
+    CheckSQLiteError(rc == SQLITE_ROW ? SQLITE_OK : rc, db_);
+
+    int64_t count = sqlite3_column_int64(stmt, 0);
+    sqlite3_finalize(stmt);
+
+    return count;
+}
+
+void EmbeddingDB::CheckVectorDimension(const std::vector<float> &vector) const {
+    INSPIREFACE_CHECK_MSG(vector.size() == vectorDim_,
+                          ("Vector dimension mismatch. Expected: " + std::to_string(vectorDim_) + ", Got: " + std::to_string(vector.size())).c_str());
+}
+
+void EmbeddingDB::ExecuteSQL(const std::string &sql) {
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    char *errMsg = nullptr;
+    int rc = sqlite3_exec(db_, sql.c_str(), nullptr, nullptr, &errMsg);
+
+    if (errMsg) {
+        std::string error = errMsg;
+        sqlite3_free(errMsg);
+        INSPIREFACE_CHECK_MSG(false, ("SQL error: " + error).c_str());
+    }
+
+    CheckSQLiteError(rc, db_);
+}
+
+void EmbeddingDB::CheckSQLiteError(int rc, sqlite3 *db) {
+    std::string error = db ? sqlite3_errmsg(db) : "SQLite error";
+    INSPIREFACE_CHECK_MSG(rc == SQLITE_OK, error.c_str());
+}
+
+void EmbeddingDB::ShowTable() {
+    if (!initialized_) {
+        INSPIRE_LOGE("EmbeddingDB is not initialized");
+        return;
+    }
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    sqlite3_stmt *stmt;
+    std::string sql = "SELECT rowid, embedding FROM " + tableName_;
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    // Print header
+#ifdef __ANDROID__
+    __android_log_print(ANDROID_LOG_INFO, "EmbeddingDB", "=== Table Content ===");
+    __android_log_print(ANDROID_LOG_INFO, "EmbeddingDB", "ID | Vector (first 5 elements)");
+    __android_log_print(ANDROID_LOG_INFO, "EmbeddingDB", "------------------------");
+#else
+    printf("=== Table Content ===\n");
+    printf("ID | Vector (first 5 elements)\n");
+    printf("------------------------\n");
+#endif
+
+    while (sqlite3_step(stmt) == SQLITE_ROW) {
+        int64_t id = sqlite3_column_int64(stmt, 0);
+        const float *vector_data = static_cast<const float *>(sqlite3_column_blob(stmt, 1));
+        size_t vector_size = std::min(size_t(5), sqlite3_column_bytes(stmt, 1) / sizeof(float));
+
+        std::string vector_str;
+        for (size_t i = 0; i < vector_size; ++i) {
+            vector_str += std::to_string(vector_data[i]);
+            if (i < vector_size - 1)
+                vector_str += ", ";
+        }
+        vector_str += "...";
+
+#ifdef __ANDROID__
+        __android_log_print(ANDROID_LOG_INFO, "EmbeddingDB", "%lld | %s", id, vector_str.c_str());
+#else
+        printf("%lld | %s\n", id, vector_str.c_str());
+#endif
+    }
+
+    sqlite3_finalize(stmt);
+}
+
+std::vector<int64_t> EmbeddingDB::GetAllIds() {
+    if (!initialized_) {
+        INSPIRE_LOGE("EmbeddingDB is not initialized");
+        return {};
+    }
+    std::lock_guard<std::mutex> lock(dbMutex_);
+    std::vector<int64_t> ids;
+
+    sqlite3_stmt *stmt;
+    std::string sql = "SELECT rowid FROM " + tableName_;
+
+    int rc = sqlite3_prepare_v2(db_, sql.c_str(), -1, &stmt, nullptr);
+    CheckSQLiteError(rc, db_);
+
+    while (sqlite3_step(stmt) == SQLITE_ROW) {
+        ids.push_back(sqlite3_column_int64(stmt, 0));
+    }
+
+    sqlite3_finalize(stmt);
+    return ids;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.h b/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.h
new file mode 100644
index 000000000..e1605fd32
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/feature_hub/embedding_db/embedding_db.h
@@ -0,0 +1,125 @@
+#ifndef INSPIRE_EMBEDDING_DB_H
+#define INSPIRE_EMBEDDING_DB_H
+
+#ifndef SQLITE_CORE
+#define SQLITE_CORE
+#endif
+
+#ifndef SQLITE_VEC_STATIC
+#define SQLITE_VEC_STATIC
+#endif
+
+#ifndef SQLITE_VEC_ENABLE_AVX
+#define SQLITE_VEC_ENABLE_AVX
+#endif
+
+#include <sqlite3.h>
+#include <vector>
+#include <string>
+#include <memory>
+#include <stdexcept>
+#include <mutex>
+#include "data_type.h"
+
+#define EMBEDDING_DB inspire::EmbeddingDB
+
+namespace inspire {
+
+// Vector data structure
+struct VectorData {
+    int64_t id;  // This field is ignored in auto-increment mode
+    std::vector<float> vector;
+};
+
+// ID mode enumeration
+enum class IdMode {
+    AUTO_INCREMENT = 0,  // Auto-incrementing ID
+    MANUAL,              // Manually specify ID
+};
+
+class EmbeddingDB {
+public:
+    ~EmbeddingDB();
+
+    static EmbeddingDB &GetInstance();
+    static void Init(const std::string &dbPath = ":memory:", size_t vectorDim = 512, IdMode idMode = IdMode::AUTO_INCREMENT);
+
+    // Delete copy and move operations
+    EmbeddingDB(const EmbeddingDB &) = delete;
+    EmbeddingDB &operator=(const EmbeddingDB &) = delete;
+    EmbeddingDB(EmbeddingDB &&) = delete;
+    EmbeddingDB &operator=(EmbeddingDB &&) = delete;
+
+    // Insert a single vector
+    bool InsertVector(int64_t id, const std::vector<float> &vector, int64_t &allocId);
+    bool InsertVector(const std::vector<float> &vector, int64_t &allocId);  // For auto-increment mode
+
+    // Batch insert vectors
+    std::vector<int64_t> BatchInsertVectors(const std::vector<VectorData> &vectors);
+    std::vector<int64_t> BatchInsertVectors(const std::vector<std::vector<float>> &vectors);  // For auto-increment mode
+
+    // Update vector
+    void UpdateVector(int64_t id, const std::vector<float> &newVector);
+
+    // Delete vector
+    void DeleteVector(int64_t id);
+
+    std::vector<FaceSearchResult> SearchSimilarVectors(const std::vector<float> &queryVector, size_t top_k = 3, float keep_similar_threshold = 0.5f,
+                                                       bool return_feature = false);
+
+    // Get vector count
+    int64_t GetVectorCount() const;
+
+    // Get current ID mode
+    IdMode GetIdMode() const {
+        return idMode_;
+    }
+
+    bool IsInitialized() const {
+        return initialized_;
+    }
+
+    // De-initialize database
+    static void Deinit() {
+        std::lock_guard<std::mutex> lock(instanceMutex_);
+        if (instance_) {
+            instance_.reset();
+        }
+    }
+
+    std::vector<float> GetVector(int64_t id) const;
+
+    void ShowTable();
+
+    std::vector<int64_t> GetAllIds();
+
+private:
+    // Constructor: add ID mode parameter
+    explicit EmbeddingDB(const std::string &dbPath = ":memory:", size_t vectorDim = 4, const std::string &distanceMetric = "cosine",
+                         IdMode idMode = IdMode::AUTO_INCREMENT);
+
+private:
+    sqlite3 *db_;
+    size_t vectorDim_;
+    std::string tableName_;
+    IdMode idMode_;
+    bool initialized_ = false;
+
+    // Helper functions
+    void CheckVectorDimension(const std::vector<float> &vector) const;
+    void ExecuteSQL(const std::string &sql);
+    static void CheckSQLiteError(int rc, sqlite3 *db);
+    int64_t GetLastInsertRowId() const;
+
+private:
+    // Singleton related
+    static std::unique_ptr<EmbeddingDB> instance_;
+    static std::mutex instanceMutex_;
+
+    // Database operation mutex
+    mutable std::mutex dbMutex_;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_EMBEDDING_DB_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/feature_hub/feature_hub_db.cpp b/cpp-package/inspireface/cpp/inspireface/feature_hub/feature_hub_db.cpp
new file mode 100644
index 000000000..2c7625b97
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/feature_hub/feature_hub_db.cpp
@@ -0,0 +1,375 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "feature_hub_db.h"
+#include "simd.h"
+#include "herror.h"
+#include <thread>
+#include "middleware/utils.h"
+#include "middleware/system.h"
+#include "log.h"
+#include "feature_hub/embedding_db/embedding_db.h"
+
+#define DB_FILE_NAME ".feature_hub_db_v0"
+
+namespace inspire {
+
+class FeatureHubDB::Impl {
+public:
+    Impl() : m_enable_(false), m_recognition_threshold_(0.48f), m_search_mode_(SEARCH_MODE_EAGER) {}
+
+    Embedded m_search_face_feature_cache_;
+    Embedded m_getter_face_feature_cache_;
+    std::shared_ptr<FaceFeaturePtr> m_face_feature_ptr_cache_;
+
+    std::vector<FaceSearchResult> m_search_top_k_cache_;
+    std::vector<float> m_top_k_confidence_;
+    std::vector<int64_t> m_top_k_custom_ids_cache_;
+
+    std::vector<int64_t> m_all_ids_;
+
+    DatabaseConfiguration m_db_configuration_;
+    float m_recognition_threshold_;
+    SearchMode m_search_mode_;
+
+    bool m_enable_;
+
+    std::mutex m_res_mtx_;
+};
+
+std::mutex FeatureHubDB::mutex_;
+std::shared_ptr<FeatureHubDB> FeatureHubDB::instance_ = nullptr;
+
+FeatureHubDB::FeatureHubDB() : pImpl(new Impl()) {}
+
+FeatureHubDB::~FeatureHubDB() = default;
+
+std::shared_ptr<FeatureHubDB> FeatureHubDB::GetInstance() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!instance_) {
+        instance_ = std::shared_ptr<FeatureHubDB>(new FeatureHubDB());
+    }
+    return instance_;
+}
+
+int32_t FeatureHubDB::DisableHub() {
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGW("FeatureHub is already disabled.");
+        return HSUCCEED;
+    }
+
+    if (EMBEDDING_DB::GetInstance().IsInitialized()) {
+        EMBEDDING_DB::Deinit();
+    }
+
+    pImpl->m_search_face_feature_cache_.clear();
+
+    pImpl->m_db_configuration_ = DatabaseConfiguration();
+    pImpl->m_recognition_threshold_ = 0.0f;
+    pImpl->m_search_mode_ = SEARCH_MODE_EAGER;
+
+    pImpl->m_face_feature_ptr_cache_.reset();
+    pImpl->m_enable_ = false;
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::GetAllIds() {
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+    pImpl->m_all_ids_ = EMBEDDING_DB::GetInstance().GetAllIds();
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::EnableHub(const DatabaseConfiguration &configuration) {
+    if (pImpl->m_enable_) {
+        INSPIRE_LOGW("You have enabled the FeatureHub feature. It is not valid to do so again");
+        return HSUCCEED;
+    }
+
+    pImpl->m_db_configuration_ = configuration;
+    pImpl->m_recognition_threshold_ = pImpl->m_db_configuration_.recognition_threshold;
+    if (pImpl->m_recognition_threshold_ < -1.0f || pImpl->m_recognition_threshold_ > 1.0f) {
+        INSPIRE_LOGW("The search threshold entered does not fit the required range (-1.0f, 1.0f) and has been set to 0.5 by default");
+        pImpl->m_recognition_threshold_ = 0.5f;
+    }
+
+    std::string dbFile = ":memory:";
+    if (pImpl->m_db_configuration_.enable_persistence) {
+        if (IsDirectory(pImpl->m_db_configuration_.persistence_db_path)) {
+            dbFile = os::PathJoin(pImpl->m_db_configuration_.persistence_db_path, DB_FILE_NAME);
+        } else {
+            dbFile = pImpl->m_db_configuration_.persistence_db_path;
+        }
+    }
+
+    EMBEDDING_DB::Init(dbFile, 512, IdMode(configuration.primary_key_mode));
+    pImpl->m_enable_ = true;
+    pImpl->m_face_feature_ptr_cache_ = std::make_shared<FaceFeatureEntity>();
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::CosineSimilarity(const std::vector<float> &v1, const std::vector<float> &v2, float &res, bool normalize) {
+    if (v1.size() != v2.size() || v1.empty()) {
+        return HERR_SESS_REC_CONTRAST_FEAT_ERR;
+    }
+
+    if (normalize) {
+        std::vector<float> v1_norm = v1;
+        std::vector<float> v2_norm = v2;
+        float mse1 = 0.0f;
+        float mse2 = 0.0f;
+
+        for (const auto &one : v1_norm) {
+            mse1 += one * one;
+        }
+        mse1 = sqrt(mse1);
+        for (float &one : v1_norm) {
+            one /= mse1;
+        }
+
+        for (const auto &one : v2_norm) {
+            mse2 += one * one;
+        }
+        mse2 = sqrt(mse2);
+        for (float &one : v2_norm) {
+            one /= mse2;
+        }
+
+        res = simd_dot(v1_norm.data(), v2_norm.data(), v1_norm.size());
+    } else {
+        res = simd_dot(v1.data(), v2.data(), v1.size());
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::CosineSimilarity(const float *v1, const float *v2, int32_t size, float &res, bool normalize) {
+    if (normalize) {
+        std::vector<float> v1_norm(v1, v1 + size);
+        std::vector<float> v2_norm(v2, v2 + size);
+        float mse1 = 0.0f;
+        float mse2 = 0.0f;
+
+        for (const auto &one : v1_norm) {
+            mse1 += one * one;
+        }
+        mse1 = sqrt(mse1);
+        for (float &one : v1_norm) {
+            one /= mse1;
+        }
+
+        for (const auto &one : v2_norm) {
+            mse2 += one * one;
+        }
+        mse2 = sqrt(mse2);
+        for (float &one : v2_norm) {
+            one /= mse2;
+        }
+
+        res = simd_dot(v1_norm.data(), v2_norm.data(), v1_norm.size());
+    } else {
+        res = simd_dot(v1, v2, size);
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::GetFaceFeatureCount() {
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGW("FeatureHub is disabled, please enable it before it can be served");
+        return 0;
+    }
+
+    return EMBEDDING_DB::GetInstance().GetVectorCount();
+}
+
+int32_t FeatureHubDB::SearchFaceFeature(const Embedded &queryFeature, FaceSearchResult &searchResult, bool returnFeature) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HSUCCEED;
+    }
+
+    pImpl->m_search_face_feature_cache_.clear();
+    auto results = EMBEDDING_DB::GetInstance().SearchSimilarVectors(queryFeature, 1, pImpl->m_recognition_threshold_, returnFeature);
+    searchResult.id = -1;
+
+    if (!results.empty()) {
+        auto &searched = results[0];
+        searchResult.similarity = searched.similarity;
+        searchResult.id = searched.id;
+        if (returnFeature) {
+            searchResult.feature = searched.feature;
+            pImpl->m_search_face_feature_cache_ = searched.feature;
+            pImpl->m_face_feature_ptr_cache_->data = pImpl->m_search_face_feature_cache_.data();
+            pImpl->m_face_feature_ptr_cache_->dataSize = pImpl->m_search_face_feature_cache_.size();
+        }
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::SearchFaceFeatureTopKCache(const Embedded &queryFeature, size_t topK) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    pImpl->m_top_k_confidence_.clear();
+    pImpl->m_top_k_custom_ids_cache_.clear();
+    auto results = EMBEDDING_DB::GetInstance().SearchSimilarVectors(queryFeature, topK, pImpl->m_recognition_threshold_, false);
+
+    for (size_t i = 0; i < results.size(); i++) {
+        pImpl->m_top_k_custom_ids_cache_.push_back(results[i].id);
+        pImpl->m_top_k_confidence_.push_back(results[i].similarity);
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::SearchFaceFeatureTopK(const Embedded &queryFeature, std::vector<FaceSearchResult> &searchResult, size_t topK,
+                                            bool returnFeature) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGW("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    searchResult = EMBEDDING_DB::GetInstance().SearchSimilarVectors(queryFeature, topK, pImpl->m_recognition_threshold_, returnFeature);
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::FaceFeatureInsert(const std::vector<float> &feature, int32_t id, int64_t &result_id) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    bool ret = EMBEDDING_DB::GetInstance().InsertVector(id, feature, result_id);
+    if (!ret) {
+        result_id = -1;
+        return HERR_FT_HUB_INSERT_FAILURE;
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::FaceFeatureRemove(int32_t id) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    EMBEDDING_DB::GetInstance().DeleteVector(id);
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::FaceFeatureUpdate(const std::vector<float> &feature, int32_t customId) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    EMBEDDING_DB::GetInstance().UpdateVector(customId, feature);
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::GetFaceFeature(int32_t id) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    auto vec = EMBEDDING_DB::GetInstance().GetVector(id);
+    if (vec.empty()) {
+        return HERR_FT_HUB_NOT_FOUND_FEATURE;
+    }
+
+    pImpl->m_getter_face_feature_cache_ = vec;
+    pImpl->m_face_feature_ptr_cache_->data = pImpl->m_getter_face_feature_cache_.data();
+    pImpl->m_face_feature_ptr_cache_->dataSize = pImpl->m_getter_face_feature_cache_.size();
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::GetFaceFeature(int32_t id, std::vector<float> &feature) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGW("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    feature = EMBEDDING_DB::GetInstance().GetVector(id);
+    if (feature.empty()) {
+        return HERR_FT_HUB_NOT_FOUND_FEATURE;
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::GetFaceFeature(int32_t id, FaceEmbedding& feature) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGW("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+
+    feature.embedding = EMBEDDING_DB::GetInstance().GetVector(id);
+    if (feature.embedding.empty()) {
+        return HERR_FT_HUB_NOT_FOUND_FEATURE;
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FeatureHubDB::ViewDBTable() {
+    if (!pImpl->m_enable_) {
+        INSPIRE_LOGE("FeatureHub is disabled, please enable it before it can be served");
+        return HERR_FT_HUB_DISABLE;
+    }
+    EMBEDDING_DB::GetInstance().ShowTable();
+    return HSUCCEED;
+}
+
+void FeatureHubDB::SetRecognitionThreshold(float threshold) {
+    pImpl->m_recognition_threshold_ = threshold;
+}
+
+void FeatureHubDB::SetRecognitionSearchMode(SearchMode mode) {
+    pImpl->m_search_mode_ = mode;
+}
+
+// =========== Getter ===========
+
+const Embedded &FeatureHubDB::GetSearchFaceFeatureCache() const {
+    return pImpl->m_search_face_feature_cache_;
+}
+
+const std::shared_ptr<FaceFeaturePtr> &FeatureHubDB::GetFaceFeaturePtrCache() const {
+    return pImpl->m_face_feature_ptr_cache_;
+}
+
+std::vector<float> &FeatureHubDB::GetTopKConfidence() {
+    return pImpl->m_top_k_confidence_;
+}
+
+std::vector<int64_t> &FeatureHubDB::GetTopKCustomIdsCache() {
+    return pImpl->m_top_k_custom_ids_cache_;
+}
+
+std::vector<int64_t> &FeatureHubDB::GetExistingIds() {
+    return pImpl->m_all_ids_;
+}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/feature_hub/simd.h b/cpp-package/inspireface/cpp/inspireface/feature_hub/simd.h
new file mode 100755
index 000000000..0b4703861
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/feature_hub/simd.h
@@ -0,0 +1,66 @@
+#include <cstdint>
+
+#if defined(_MSC_VER)
+/* Microsoft C/C++-compatible compiler */
+#include <intrin.h>
+#elif (defined(__x86_64__) || defined(__i386__))
+/* GCC-compatible compiler, targeting x86/x86-64 */
+#include <x86intrin.h>
+#elif defined(__ARM_NEON__)
+/* GCC-compatible compiler, targeting ARM with NEON */
+#include <arm_neon.h>
+#endif
+
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__) || defined(_MSC_VER))
+inline float simd_dot(const float *x, const float *y, const long &len) {
+    // #pragma message("USE SSE")
+    float inner_prod = 0.0f;
+    __m128 X, Y, Z;                 // 128-bit values
+    __m128 acc = _mm_setzero_ps();  // set to (0, 0, 0, 0)
+    float temp[4];
+
+    long i;
+    for (i = 0; i + 4 < len; i += 4) {
+        X = _mm_loadu_ps(x + i);  // load chunk of 4 floats
+        Y = _mm_loadu_ps(y + i);
+        Z = _mm_mul_ps(X, Y);
+        acc = _mm_add_ps(acc, Z);
+    }
+    _mm_storeu_ps(&temp[0], acc);  // store acc into an array
+    inner_prod = temp[0] + temp[1] + temp[2] + temp[3];
+
+    // add the remaining values
+    for (; i < len; ++i) {
+        inner_prod += x[i] * y[i];
+    }
+    return inner_prod;
+}
+#elif defined(__ARM_NEON__)
+inline float simd_dot(const float *x, const float *y, const long &len) {
+    // #pragma message("USE NEON")
+    float inner_prod = 0.0f;
+    float32x4_t X, Y, Z;                  // 128-bit values
+    float32x4_t acc = vdupq_n_f32(0.0f);  // set to (0, 0, 0, 0)
+    long i;
+    for (i = 0; i + 4 < len; i += 4) {
+        X = vld1q_f32(x + i);  // load chunk of 4 floats
+        Y = vld1q_f32(y + i);
+        Z = vmulq_f32(X, Y);
+        acc = vaddq_f32(acc, Z);
+    }
+    inner_prod = vgetq_lane_f32(acc, 0) + vgetq_lane_f32(acc, 1) + vgetq_lane_f32(acc, 2) + vgetq_lane_f32(acc, 3);
+    for (; i < len; ++i) {
+        inner_prod += x[i] * y[i];
+    }
+    return inner_prod;
+}
+#else
+inline float simd_dot(const float *x, const float *y, const long &len) {
+    // #pragma message("USE SCALAR")
+    float inner_prod = 0.0f;
+    for (long i = 0; i < len; ++i) {
+        inner_prod += x[i] * y[i];
+    }
+    return inner_prod;
+}
+#endif
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/frame_process.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/frame_process.cpp
new file mode 100644
index 000000000..a61ab1bc9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/frame_process.cpp
@@ -0,0 +1,407 @@
+#include "frame_process.h"
+#include <vector>
+#include <MNN/ImageProcess.hpp>
+#include "isf_check.h"
+
+namespace inspirecv {
+
+class FrameProcess::Impl {
+public:
+    Impl() : buffer_(nullptr), height_(0), width_(0), preview_scale_(0), preview_size_(192), rotation_mode_(ROTATION_0) {
+        SetDataFormat(NV21);
+        SetDestFormat(BGR);
+        config_.filterType = MNN::CV::BILINEAR;
+        config_.wrap = MNN::CV::ZERO;
+    }
+
+    void SetDataFormat(DATA_FORMAT data_format) {
+        if (data_format == NV21) {
+            config_.sourceFormat = MNN::CV::YUV_NV21;
+        }
+        if (data_format == NV12) {
+            config_.sourceFormat = MNN::CV::YUV_NV12;
+        }
+        if (data_format == RGBA) {
+            config_.sourceFormat = MNN::CV::RGBA;
+        }
+        if (data_format == RGB) {
+            config_.sourceFormat = MNN::CV::RGB;
+        }
+        if (data_format == BGR) {
+            config_.sourceFormat = MNN::CV::BGR;
+        }
+        if (data_format == BGRA) {
+            config_.sourceFormat = MNN::CV::BGRA;
+        }
+        if (data_format == I420) {
+            config_.sourceFormat = MNN::CV::YUV_I420;
+        }
+        if (data_format == GRAY) {
+            config_.sourceFormat = MNN::CV::GRAY;
+        }
+    }
+
+    void SetDestFormat(DATA_FORMAT data_format) {
+        if (data_format == NV21) {
+            config_.destFormat = MNN::CV::YUV_NV21;
+        }
+        if (data_format == NV12) {
+            config_.destFormat = MNN::CV::YUV_NV12;
+        }
+        if (data_format == RGBA) {
+            config_.destFormat = MNN::CV::RGBA;
+        }
+        if (data_format == RGB) {
+            config_.destFormat = MNN::CV::RGB;
+        }
+        if (data_format == BGR) {
+            config_.destFormat = MNN::CV::BGR;
+        }
+        if (data_format == BGRA) {
+            config_.destFormat = MNN::CV::BGRA;
+        }
+        if (data_format == I420) {
+            config_.destFormat = MNN::CV::YUV_I420;
+        }
+        if (data_format == GRAY) {
+            config_.destFormat = MNN::CV::GRAY;
+        }
+    }
+
+    void UpdateTransformMatrix() {
+        float srcPoints[] = {0.0f, 0.0f, 0.0f, (float)(height_ - 1), (float)(width_ - 1), 0.0f, (float)(width_ - 1), (float)(height_ - 1)};
+
+        float dstPoints[8];
+        if (rotation_mode_ == ROTATION_270) {
+            float points[] = {(float)(height_ * preview_scale_ - 1),
+                              0.0f,
+                              0.0f,
+                              0.0f,
+                              (float)(height_ * preview_scale_ - 1),
+                              (float)(width_ * preview_scale_ - 1),
+                              0.0f,
+                              (float)(width_ * preview_scale_ - 1)};
+            memcpy(dstPoints, points, sizeof(points));
+        } else if (rotation_mode_ == ROTATION_90) {
+            float points[] = {0.0f,
+                              (float)(width_ * preview_scale_ - 1),
+                              (float)(height_ * preview_scale_ - 1),
+                              (float)(width_ * preview_scale_ - 1),
+                              0.0f,
+                              0.0f,
+                              (float)(height_ * preview_scale_ - 1),
+                              0.0f};
+            memcpy(dstPoints, points, sizeof(points));
+        } else if (rotation_mode_ == ROTATION_180) {
+            float points[] = {(float)(width_ * preview_scale_ - 1),
+                              (float)(height_ * preview_scale_ - 1),
+                              (float)(width_ * preview_scale_ - 1),
+                              0.0f,
+                              0.0f,
+                              (float)(height_ * preview_scale_ - 1),
+                              0.0f,
+                              0.0f};
+            memcpy(dstPoints, points, sizeof(points));
+        } else {  // ROTATION_0
+            float points[] = {0.0f,
+                              0.0f,
+                              0.0f,
+                              (float)(height_ * preview_scale_ - 1),
+                              (float)(width_ * preview_scale_ - 1),
+                              0.0f,
+                              (float)(width_ * preview_scale_ - 1),
+                              (float)(height_ * preview_scale_ - 1)};
+            memcpy(dstPoints, points, sizeof(points));
+        }
+
+        tr_.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+    }
+
+    const uint8_t *buffer_;                 // Pointer to the data buffer.
+    int height_;                            // Height of the camera stream image.
+    int width_;                             // Width of the camera stream image.
+    float preview_scale_;                   // Scaling factor for the preview image.
+    int preview_size_;                      // Size of the preview image.
+    MNN::CV::Matrix tr_;                    // Affine transformation matrix.
+    ROTATION_MODE rotation_mode_;           // Current rotation mode.
+    MNN::CV::ImageProcess::Config config_;  // Image processing configuration.
+};
+
+FrameProcess FrameProcess::Create(const uint8_t *data_buffer, int height, int width, DATA_FORMAT data_format, ROTATION_MODE rotation_mode) {
+    FrameProcess process;
+    process.SetDataBuffer(data_buffer, height, width);
+    process.SetDataFormat(data_format);
+    process.SetRotationMode(rotation_mode);
+    return process;
+}
+
+FrameProcess FrameProcess::Create(const inspirecv::Image &image, DATA_FORMAT data_format, ROTATION_MODE rotation_mode) {
+    return Create(image.Data(), image.Height(), image.Width(), data_format, rotation_mode);
+}
+
+FrameProcess::FrameProcess() : pImpl(std::make_unique<Impl>()) {
+    pImpl->UpdateTransformMatrix();
+}
+
+FrameProcess::~FrameProcess() = default;
+
+FrameProcess::FrameProcess(const FrameProcess &other) : pImpl(std::make_unique<Impl>(*other.pImpl)) {}
+
+FrameProcess::FrameProcess(FrameProcess &&other) noexcept = default;
+
+FrameProcess &FrameProcess::operator=(const FrameProcess &other) {
+    if (this != &other) {
+        *pImpl = *other.pImpl;
+    }
+    return *this;
+}
+
+FrameProcess &FrameProcess::operator=(FrameProcess &&other) noexcept = default;
+
+void FrameProcess::SetDataBuffer(const uint8_t *data_buffer, int height, int width) {
+    pImpl->buffer_ = data_buffer;
+    pImpl->height_ = height;
+    pImpl->width_ = width;
+    pImpl->preview_scale_ = pImpl->preview_size_ / static_cast<float>(std::max(height, width));
+    pImpl->UpdateTransformMatrix();
+}
+
+void FrameProcess::SetPreviewSize(const int size) {
+    pImpl->preview_size_ = size;
+    pImpl->preview_scale_ = pImpl->preview_size_ / static_cast<float>(std::max(pImpl->height_, pImpl->width_));
+    pImpl->UpdateTransformMatrix();
+}
+
+void FrameProcess::SetPreviewScale(const float scale) {
+    pImpl->preview_scale_ = scale;
+    pImpl->preview_size_ = static_cast<int>(pImpl->preview_scale_ * std::max(pImpl->height_, pImpl->width_));
+    pImpl->UpdateTransformMatrix();
+}
+
+void FrameProcess::SetRotationMode(ROTATION_MODE mode) {
+    pImpl->rotation_mode_ = mode;
+    pImpl->UpdateTransformMatrix();
+}
+
+void FrameProcess::SetDataFormat(DATA_FORMAT data_format) {
+    pImpl->SetDataFormat(data_format);
+}
+
+void FrameProcess::SetDestFormat(DATA_FORMAT data_format) {
+    pImpl->SetDestFormat(data_format);
+}
+
+float FrameProcess::GetPreviewScale() {
+    return pImpl->preview_scale_;
+}
+
+inspirecv::TransformMatrix FrameProcess::GetAffineMatrix() const {
+    auto affine_matrix = inspirecv::TransformMatrix::Create();
+    affine_matrix[0] = pImpl->tr_[0];
+    affine_matrix[1] = pImpl->tr_[1];
+    affine_matrix[2] = pImpl->tr_[2];
+    affine_matrix[3] = pImpl->tr_[3];
+    affine_matrix[4] = pImpl->tr_[4];
+    affine_matrix[5] = pImpl->tr_[5];
+    return affine_matrix;
+}
+
+int FrameProcess::GetHeight() const {
+    return pImpl->height_;
+}
+
+int FrameProcess::GetWidth() const {
+    return pImpl->width_;
+}
+
+ROTATION_MODE FrameProcess::getRotationMode() const {
+    return pImpl->rotation_mode_;
+}
+
+inspirecv::Image FrameProcess::ExecuteImageAffineProcessing(inspirecv::TransformMatrix &affine_matrix, const int width_out,
+                                                            const int height_out) const {
+    int sw = pImpl->width_;
+    int sh = pImpl->height_;
+    int rot_sw = sw;
+    int rot_sh = sh;
+    MNN::CV::Matrix tr;
+    std::vector<float> tr_cv({1, 0, 0, 0, 1, 0, 0, 0, 1});
+    memcpy(tr_cv.data(), affine_matrix.Squeeze().data(), sizeof(float) * 6);
+    tr.set9(tr_cv.data());
+    MNN::CV::Matrix tr_inv;
+    tr.invert(&tr_inv);
+    std::shared_ptr<MNN::CV::ImageProcess> process(MNN::CV::ImageProcess::create(pImpl->config_));
+    process->setMatrix(tr_inv);
+    auto img_out = inspirecv::Image::Create(width_out, height_out, 3);
+    std::shared_ptr<MNN::Tensor> tensor(MNN::Tensor::create<uint8_t>(std::vector<int>{1, height_out, width_out, 3}, (uint8_t *)img_out.Data()));
+    auto ret = process->convert(pImpl->buffer_, sw, sh, 0, tensor.get());
+    INSPIREFACE_CHECK_MSG(ret == MNN::ErrorCode::NO_ERROR, "ImageProcess::convert failed");
+    return img_out;
+}
+
+inspirecv::Image FrameProcess::ExecutePreviewImageProcessing(bool with_rotation) {
+    return ExecuteImageScaleProcessing(pImpl->preview_scale_, with_rotation);
+}
+
+inspirecv::Image FrameProcess::ExecuteImageScaleProcessing(const float scale, bool with_rotation) {
+    int sw = pImpl->width_;
+    int sh = pImpl->height_;
+    int rot_sw = sw;
+    int rot_sh = sh;
+    // MNN::CV::Matrix tr;
+    std::shared_ptr<MNN::CV::ImageProcess> process(MNN::CV::ImageProcess::create(pImpl->config_));
+    if (pImpl->rotation_mode_ == ROTATION_270 && with_rotation) {
+        float srcPoints[] = {
+          0.0f, 0.0f, 0.0f, (float)(pImpl->height_ - 1), (float)(pImpl->width_ - 1), 0.0f, (float)(pImpl->width_ - 1), (float)(pImpl->height_ - 1),
+        };
+        float dstPoints[] = {
+          (float)(pImpl->height_ * scale - 1), 0.0f, 0.0f, 0.0f, (float)(pImpl->height_ * scale - 1), (float)(pImpl->width_ * scale - 1), 0.0f,
+          (float)(pImpl->width_ * scale - 1)};
+
+        pImpl->tr_.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+        process->setMatrix(pImpl->tr_);
+        int scaled_height = static_cast<int>(pImpl->width_ * scale);
+        int scaled_width = static_cast<int>(pImpl->height_ * scale);
+        inspirecv::Image img_out(scaled_width, scaled_height, 3);
+        std::shared_ptr<MNN::Tensor> tensor(
+          MNN::Tensor::create<uint8_t>(std::vector<int>{1, scaled_height, scaled_width, 3}, (uint8_t *)img_out.Data()));
+        auto ret = process->convert(pImpl->buffer_, sw, sh, 0, tensor.get());
+        INSPIREFACE_CHECK_MSG(ret == MNN::ErrorCode::NO_ERROR, "ImageProcess::convert failed");
+        return img_out;
+    } else if (pImpl->rotation_mode_ == ROTATION_90 && with_rotation) {
+        float srcPoints[] = {
+          0.0f, 0.0f, 0.0f, (float)(pImpl->height_ - 1), (float)(pImpl->width_ - 1), 0.0f, (float)(pImpl->width_ - 1), (float)(pImpl->height_ - 1),
+        };
+        float dstPoints[] = {
+          0.0f,
+          (float)(pImpl->width_ * scale - 1),
+          (float)(pImpl->height_ * scale - 1),
+          (float)(pImpl->width_ * scale - 1),
+          0.0f,
+          0.0f,
+          (float)(pImpl->height_ * scale - 1),
+          0.0f,
+        };
+        pImpl->tr_.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+        process->setMatrix(pImpl->tr_);
+        int scaled_height = static_cast<int>(pImpl->width_ * scale);
+        int scaled_width = static_cast<int>(pImpl->height_ * scale);
+        inspirecv::Image img_out(scaled_width, scaled_height, 3);
+        std::shared_ptr<MNN::Tensor> tensor(
+          MNN::Tensor::create<uint8_t>(std::vector<int>{1, scaled_height, scaled_width, 3}, (uint8_t *)img_out.Data()));
+        auto ret = process->convert(pImpl->buffer_, sw, sh, 0, tensor.get());
+        INSPIREFACE_CHECK_MSG(ret == MNN::ErrorCode::NO_ERROR, "ImageProcess::convert failed");
+        return img_out;
+    } else if (pImpl->rotation_mode_ == ROTATION_180 && with_rotation) {
+        float srcPoints[] = {
+          0.0f, 0.0f, 0.0f, (float)(pImpl->height_ - 1), (float)(pImpl->width_ - 1), 0.0f, (float)(pImpl->width_ - 1), (float)(pImpl->height_ - 1),
+        };
+        float dstPoints[] = {
+          (float)(pImpl->width_ * scale - 1),
+          (float)(pImpl->height_ * scale - 1),
+          (float)(pImpl->width_ * scale - 1),
+          0.0f,
+          0.0f,
+          (float)(pImpl->height_ * scale - 1),
+          0.0f,
+          0.0f,
+        };
+        pImpl->tr_.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+        process->setMatrix(pImpl->tr_);
+        int scaled_height = static_cast<int>(pImpl->height_ * scale);
+        int scaled_width = static_cast<int>(pImpl->width_ * scale);
+        inspirecv::Image img_out(scaled_width, scaled_height, 3);
+        std::shared_ptr<MNN::Tensor> tensor(
+          MNN::Tensor::create<uint8_t>(std::vector<int>{1, scaled_height, scaled_width, 3}, (uint8_t *)img_out.Data()));
+        auto ret = process->convert(pImpl->buffer_, sw, sh, 0, tensor.get());
+        INSPIREFACE_CHECK_MSG(ret == MNN::ErrorCode::NO_ERROR, "ImageProcess::convert failed");
+        return img_out;
+    } else {
+        float srcPoints[] = {
+          0.0f, 0.0f, 0.0f, (float)(pImpl->height_ - 1), (float)(pImpl->width_ - 1), 0.0f, (float)(pImpl->width_ - 1), (float)(pImpl->height_ - 1),
+        };
+        float dstPoints[] = {
+          0.0f,
+          0.0f,
+          0.0f,
+          (float)(pImpl->height_ * scale - 1),
+          (float)(pImpl->width_ * scale - 1),
+          0.0f,
+          (float)(pImpl->width_ * scale - 1),
+          (float)(pImpl->height_ * scale - 1),
+        };
+        pImpl->tr_.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+        process->setMatrix(pImpl->tr_);
+        int scaled_height = static_cast<int>(pImpl->height_ * scale);
+        int scaled_width = static_cast<int>(pImpl->width_ * scale);
+
+        inspirecv::Image img_out(scaled_width, scaled_height, 3);
+        std::shared_ptr<MNN::Tensor> tensor(
+          MNN::Tensor::create<uint8_t>(std::vector<int>{1, scaled_height, scaled_width, 3}, (uint8_t *)img_out.Data()));
+        auto ret = process->convert(pImpl->buffer_, sw, sh, 0, tensor.get());
+        INSPIREFACE_CHECK_MSG(ret == MNN::ErrorCode::NO_ERROR, "ImageProcess::convert failed");
+        return img_out;
+    }
+}
+
+inspirecv::TransformMatrix FrameProcess::GetRotationModeAffineMatrix() const {
+    float srcPoints[] = {0.0f, 0.0f, 0.0f, (float)(pImpl->height_ - 1), (float)(pImpl->width_ - 1), 0.0f, (float)(pImpl->width_ - 1), (float)(pImpl->height_ - 1)};
+    float dstPoints[8];
+    
+    if (pImpl->rotation_mode_ == ROTATION_270) {
+        float points[] = {(float)(pImpl->height_ - 1),
+                         0.0f,
+                         0.0f,
+                         0.0f,
+                         (float)(pImpl->height_ - 1),
+                         (float)(pImpl->width_ - 1),
+                         0.0f,
+                         (float)(pImpl->width_ - 1)};
+        memcpy(dstPoints, points, sizeof(points));
+    } else if (pImpl->rotation_mode_ == ROTATION_90) {
+        float points[] = {0.0f,
+                         (float)(pImpl->width_ - 1),
+                         (float)(pImpl->height_ - 1),
+                         (float)(pImpl->width_ - 1),
+                         0.0f,
+                         0.0f,
+                         (float)(pImpl->height_ - 1),
+                         0.0f};
+        memcpy(dstPoints, points, sizeof(points));
+    } else if (pImpl->rotation_mode_ == ROTATION_180) {
+        float points[] = {(float)(pImpl->width_ - 1),
+                         (float)(pImpl->height_ - 1),
+                         (float)(pImpl->width_ - 1),
+                         0.0f,
+                         0.0f,
+                         (float)(pImpl->height_ - 1),
+                         0.0f,
+                         0.0f};
+        memcpy(dstPoints, points, sizeof(points));
+    } else {  // ROTATION_0
+        float points[] = {0.0f,
+                         0.0f,
+                         0.0f,
+                         (float)(pImpl->height_ - 1),
+                         (float)(pImpl->width_ - 1),
+                         0.0f,
+                         (float)(pImpl->width_ - 1),
+                         (float)(pImpl->height_ - 1)};
+        memcpy(dstPoints, points, sizeof(points));
+    }
+
+    MNN::CV::Matrix tr;
+    tr.setPolyToPoly((MNN::CV::Point *)dstPoints, (MNN::CV::Point *)srcPoints, 4);
+    
+    auto affine_matrix = inspirecv::TransformMatrix::Create();
+    affine_matrix[0] = tr[0];
+    affine_matrix[1] = tr[1];
+    affine_matrix[2] = tr[2];
+    affine_matrix[3] = tr[3];
+    affine_matrix[4] = tr[4];
+    affine_matrix[5] = tr[5];
+    
+    return affine_matrix;
+}
+
+}  // namespace inspirecv
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.cpp
new file mode 100644
index 000000000..6aa02e993
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.cpp
@@ -0,0 +1,34 @@
+#include "image_processor.h"
+#include "log.h"
+
+#if defined(ISF_ENABLE_RGA)
+#include "image_processor_rga.h"
+#endif
+#include "image_processor_general.h"
+
+namespace inspire {
+
+namespace nexus {
+
+ImageProcessor::~ImageProcessor() = default;
+
+std::unique_ptr<ImageProcessor> ImageProcessor::Create(inspire::Launch::ImageProcessingBackend backend) {
+    switch (backend) {
+        case inspire::Launch::IMAGE_PROCESSING_RGA:
+            #if defined(ISF_ENABLE_RGA)
+            return std::make_unique<RgaImageProcessor>();
+            #else
+            INSPIRE_LOGE("RGA backend is not enabled, using CPU backend instead");
+            return std::make_unique<GeneralImageProcessor>();
+            #endif
+        case inspire::Launch::IMAGE_PROCESSING_CPU:
+            return std::make_unique<GeneralImageProcessor>();
+        default:
+            // Default to CPU backend
+            return std::make_unique<GeneralImageProcessor>();
+    }
+}
+
+}  // namespace nexus
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.h b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.h
new file mode 100644
index 000000000..3cd029fd6
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor.h
@@ -0,0 +1,66 @@
+#ifndef INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_H
+#define INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_H
+
+#include <iostream>
+#include <vector>
+#include <inspirecv/inspirecv.h>
+#include <memory>
+#include "launch.h"
+#include "data_type.h"
+
+namespace inspire {
+
+namespace nexus {
+
+/**
+ * @brief Extensible image processing interface that supports hardware acceleration backends
+ *
+ * This interface provides common image processing operations like resize, color conversion,
+ * padding etc. It can be implemented by different backends based on compile options:
+ * - Default CPU-based implementation using InspireCV (always available)
+ * - Hardware accelerated implementation like Rockchip RGA (enabled with ISF_ENABLE_RGA) 
+ *
+ * The backend implementation is selected at runtime based on the backend parameter.
+ */
+class INSPIRE_API_EXPORT ImageProcessor {
+public:
+    static std::unique_ptr<ImageProcessor> Create(inspire::Launch::ImageProcessingBackend backend = inspire::Launch::IMAGE_PROCESSING_CPU);
+
+public:
+    // Virtual destructor
+    virtual ~ImageProcessor() = 0;
+
+    // Resize image to specified dimensions
+    virtual int32_t Resize(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data, int dst_width,
+                           int dst_height) = 0;
+
+    // Swap color channels of the image
+    virtual int32_t SwapColor(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data) = 0;
+
+    // Add padding to image borders
+    virtual int32_t Padding(const uint8_t* src_data, int src_width, int src_height, int channels, int top, int bottom, int left, int right,
+                            uint8_t** dst_data, int& dst_width, int& dst_height) = 0;
+
+    // Resize image and add padding as needed
+    virtual int32_t ResizeAndPadding(const uint8_t* src_data, int src_width, int src_height, int channels, int dst_width, int dst_height,
+                                     uint8_t** dst_data, float& scale) = 0;
+
+    // Mark processing as complete
+    virtual int32_t MarkDone() = 0;
+
+    // Display cache status information
+    virtual void DumpCacheStatus() const = 0;
+
+    // Get aligned width of the image
+    virtual int32_t GetAlignedWidth(int width) const = 0;
+
+    // Set aligned width of the image
+    virtual void SetAlignedWidth(int width) = 0;
+
+};  // class ImageProcessor
+
+}  // namespace nexus
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_H
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.cpp
new file mode 100644
index 000000000..481ef3fb7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.cpp
@@ -0,0 +1,71 @@
+#include "image_processor_general.h"
+#include "log.h"
+
+namespace inspire {
+
+namespace nexus {
+
+int32_t GeneralImageProcessor::Resize(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data, int dst_width,
+                                      int dst_height) {
+    inspirecv::Image src_img(src_width, src_height, channels, src_data, false);
+    last_buffer_.image = src_img.Resize(dst_width, dst_height);
+    *dst_data = last_buffer_.GetData();
+    return 0;
+}
+
+int32_t GeneralImageProcessor::SwapColor(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data) {
+    inspirecv::Image src_img(src_width, src_height, channels, src_data, false);
+    last_buffer_.image = src_img.SwapRB();
+    *dst_data = last_buffer_.GetData();
+    return 0;
+}
+
+int32_t GeneralImageProcessor::Padding(const uint8_t* src_data, int src_width, int src_height, int channels, int top, int bottom, int left, int right,
+                                       uint8_t** dst_data, int& dst_width, int& dst_height) {
+    inspirecv::Image src_img(src_width, src_height, channels, src_data, false);
+    dst_width = src_width + left + right;
+    dst_height = src_height + top + bottom;
+    last_buffer_.image = src_img.Pad(top, bottom, left, right, inspirecv::Color::Black);
+    *dst_data = last_buffer_.GetData();
+    return 0;
+}
+int32_t GeneralImageProcessor::ResizeAndPadding(const uint8_t* src_data, int src_width, int src_height, int channels, int dst_width, int dst_height,
+                                                uint8_t** dst_data, float& scale) {
+    inspirecv::Image src_img(src_width, src_height, channels, src_data, false);
+
+    scale = std::min(static_cast<float>(dst_width) / src_width, static_cast<float>(dst_height) / src_height);
+
+    int resized_w = static_cast<int>(src_width * scale);
+    int resized_h = static_cast<int>(src_height * scale);
+
+    int wpad = dst_width - resized_w;
+    int hpad = dst_height - resized_h;
+
+    inspirecv::Image resized_img = src_img.Resize(resized_w, resized_h);
+    last_buffer_.image = resized_img.Pad(0, hpad, 0, wpad, inspirecv::Color::Black);
+    *dst_data = last_buffer_.GetData();
+    return 0;
+}
+
+int32_t GeneralImageProcessor::MarkDone() {
+    return 0;
+}
+
+void GeneralImageProcessor::DumpCacheStatus() const {
+    INSPIRECV_LOG(INFO) << "GeneralImageProcessor has no cache to dump";
+}
+
+int32_t GeneralImageProcessor::GetAlignedWidth(int width) const {
+    // Not Supported
+    INSPIRE_LOGE("GeneralImageProcessor::GetAlignedWidth is not supported");
+    return 0;
+}
+
+void GeneralImageProcessor::SetAlignedWidth(int width) {
+    // Not Supported
+    INSPIRE_LOGE("GeneralImageProcessor::SetAlignedWidth is not supported");
+}
+
+}  // namespace nexus
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.h b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.h
new file mode 100644
index 000000000..1ba648298
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_general.h
@@ -0,0 +1,49 @@
+#ifndef INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_GENERAL_H
+#define INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_GENERAL_H
+
+#include "image_processor.h"
+#include <inspirecv/inspirecv.h>
+
+namespace inspire {
+
+namespace nexus {
+
+class INSPIRE_API_EXPORT GeneralImageProcessor : public ImageProcessor {
+public:
+    GeneralImageProcessor() = default;
+    ~GeneralImageProcessor() override = default;
+
+    int32_t Resize(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data, int dst_width, int dst_height) override;
+
+    int32_t SwapColor(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data) override;
+
+    int32_t Padding(const uint8_t* src_data, int src_width, int src_height, int channels, int top, int bottom, int left, int right,
+                    uint8_t** dst_data, int& dst_width, int& dst_height) override;
+
+    int32_t ResizeAndPadding(const uint8_t* src_data, int src_width, int src_height, int channels, int dst_width, int dst_height, uint8_t** dst_data,
+                             float& scale) override;
+
+    int32_t MarkDone() override;
+
+    void DumpCacheStatus() const override;
+
+    int32_t GetAlignedWidth(int width) const override;
+
+    void SetAlignedWidth(int width) override;
+
+private:
+    struct BufferWrapper {
+        inspirecv::Image image;
+        uint8_t* GetData() {
+            return const_cast<uint8_t*>(image.Data());
+        }
+    };
+    BufferWrapper last_buffer_;
+
+};  // GeneralImageProcessor
+
+}  // namespace nexus
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_GENERAL_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.cpp
new file mode 100644
index 000000000..10b485034
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.cpp
@@ -0,0 +1,256 @@
+#include "image_processor_rga.h"
+
+#if defined(ISF_ENABLE_RGA)
+
+namespace inspire {
+
+namespace nexus {
+
+RgaImageProcessor::RgaImageProcessor() {
+    aligned_width_ = 4;
+}
+
+RgaImageProcessor::~RgaImageProcessor() {}
+
+int32_t RgaImageProcessor::GetAlignedWidth(int width) const {
+    return aligned_width_;
+}
+
+void RgaImageProcessor::SetAlignedWidth(int width) {
+    aligned_width_ = width;
+}
+
+int32_t RgaImageProcessor::Resize(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data, int dst_width,
+                                  int dst_height) {
+    // Calculate width aligned to 4 bytes
+    int aligned_src_width = (src_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+    int aligned_dst_width = (dst_width + aligned_width_ - 1) & ~(aligned_width_ - 1);
+    // std::cout << "aligned_src_width: " << aligned_src_width << ", aligned_dst_width: " << aligned_dst_width << std::endl;
+
+    // 1. Get or create source buffer with aligned width
+    BufferKey src_key{aligned_src_width, src_height, channels};
+    auto& src_buffer = GetOrCreateBuffer(src_key);
+
+    // 2. Get or create destination buffer with aligned width
+    BufferKey dst_key{aligned_dst_width, dst_height, channels};
+    auto& dst_buffer = GetOrCreateBuffer(dst_key, false);
+
+    // 3. Copy source data to RGA buffer, handle padding row by row
+    for (int y = 0; y < src_height; y++) {
+        memcpy(static_cast<uint8_t*>(src_buffer.virtual_addr) + y * aligned_src_width * channels, src_data + y * src_width * channels,
+               src_width * channels);
+        // Padding area remains zero
+    }
+
+    dma_sync_cpu_to_device(src_buffer.dma_fd);
+    dma_sync_cpu_to_device(dst_buffer.dma_fd);
+
+    // 4. Execute RGA resize
+    int ret = imcheck(src_buffer.buffer, dst_buffer.buffer, {}, {});
+    if (IM_STATUS_NOERROR != ret) {
+        INSPIRECV_LOG(ERROR) << "RGA parameter check failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    ret = imresize(src_buffer.buffer, dst_buffer.buffer);
+
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA resize failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 5. Return pointer to destination buffer
+    *dst_data = static_cast<uint8_t*>(dst_buffer.virtual_addr);
+
+    return 0;
+}
+
+int32_t RgaImageProcessor::MarkDone() {
+    // Sync all buffers
+    for (const auto& pair : buffer_cache_) {
+        dma_sync_device_to_cpu(pair.second.dma_fd);
+    }
+
+    // // Print current cache status for debugging
+    // INSPIRECV_LOG(INFO) << "MarkDone: Current cache status:";
+    // INSPIRECV_LOG(INFO) << "Cache size: " << buffer_cache_.size();
+    // for (const auto& pair : buffer_cache_) {
+    //     INSPIRECV_LOG(INFO) << "Buffer: " << pair.second.width << "x" << pair.second.height << " dma_fd=" << pair.second.dma_fd;
+    // }
+    return 0;
+}
+
+int32_t RgaImageProcessor::SwapColor(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data) {
+    // Calculate width aligned to 4 bytes
+    int aligned_src_width = (src_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+    // 1. Get or create source buffer
+    BufferKey src_key{aligned_src_width, src_height, channels};
+    auto& src_buffer = GetOrCreateBuffer(src_key);
+
+    // 2. Get or create destination buffer
+    BufferKey dst_key{aligned_src_width, src_height, channels};
+    auto& dst_buffer = GetOrCreateBuffer(dst_key, false);
+
+    // 3. Copy source data to RGA buffer, handle padding row by row
+    for (int y = 0; y < src_height; y++) {
+        memcpy(static_cast<uint8_t*>(src_buffer.virtual_addr) + y * aligned_src_width * channels, src_data + y * src_width * channels,
+               src_width * channels);
+        // Padding area remains zero
+    }
+
+    dma_sync_cpu_to_device(src_buffer.dma_fd);
+    dma_sync_cpu_to_device(dst_buffer.dma_fd);
+
+    // 3. Execute RGA swap color
+    int ret = imcheck(src_buffer.buffer, dst_buffer.buffer, {}, {});
+    if (IM_STATUS_NOERROR != ret) {
+        INSPIRECV_LOG(ERROR) << "RGA parameter check failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 4. Execute RGA swap color
+    ret = imcvtcolor(src_buffer.buffer, dst_buffer.buffer, RK_FORMAT_RGB_888, RK_FORMAT_BGR_888);
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA color conversion failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    dma_sync_device_to_cpu(dst_buffer.dma_fd);
+
+    // 5. Return pointer to destination buffer
+    *dst_data = static_cast<uint8_t*>(dst_buffer.virtual_addr);
+
+    return 0;
+}
+
+int32_t RgaImageProcessor::Padding(const uint8_t* src_data, int src_width, int src_height, int channels, int top, int bottom, int left, int right,
+                                   uint8_t** dst_data, int& dst_width, int& dst_height) {
+    // Calculate final dimensions
+    dst_width = src_width + left + right;
+    dst_height = src_height + top + bottom;
+
+    // Calculate width aligned to 4 bytes
+    int aligned_src_width = (src_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+    int aligned_dst_width = (dst_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+
+    // 1. Get or create source buffer with aligned width
+    BufferKey src_key{aligned_src_width, src_height, channels};
+    auto& src_buffer = GetOrCreateBuffer(src_key, true);
+
+    // 2. Create destination buffer with padded dimensions
+    BufferKey dst_key{aligned_dst_width, dst_height, channels};
+    auto& dst_buffer = GetOrCreateBuffer(dst_key, false);
+
+    // 3. Copy source data to RGA buffer
+    for (int y = 0; y < src_height; y++) {
+        memcpy(static_cast<uint8_t*>(src_buffer.virtual_addr) + y * aligned_src_width * channels, src_data + y * src_width * channels,
+               src_width * channels);
+    }
+
+    dma_sync_cpu_to_device(src_buffer.dma_fd);
+    dma_sync_cpu_to_device(dst_buffer.dma_fd);
+
+    // 4. Execute padding operation
+    // Set source and destination regions
+    im_rect src_rect = {0, 0, src_width, src_height};
+    im_rect dst_rect = {left, top, src_width, src_height};  // Specify padding position
+
+    int ret = imcheck(src_buffer.buffer, dst_buffer.buffer, src_rect, dst_rect);
+    if (IM_STATUS_NOERROR != ret) {
+        INSPIRECV_LOG(ERROR) << "RGA parameter check failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 5. Fill entire destination area with black first
+    ret = imfill(dst_buffer.buffer, {0, 0, dst_width, dst_height}, 0x000000);
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA fill failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 6. Copy source image to specified position in destination
+    ret = improcess(src_buffer.buffer, dst_buffer.buffer, {}, src_rect, dst_rect, {}, IM_SYNC);
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA copy failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    dma_sync_device_to_cpu(dst_buffer.dma_fd);
+
+    *dst_data = static_cast<uint8_t*>(dst_buffer.virtual_addr);
+    return 0;
+}
+
+int32_t RgaImageProcessor::ResizeAndPadding(const uint8_t* src_data, int src_width, int src_height, int channels, int dst_width, int dst_height,
+                                            uint8_t** dst_data, float& scale) {
+    // Ensure target dimensions are multiples of 4
+    int aligned_dst_width = (dst_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+    int aligned_dst_height = (dst_height + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+
+    // Calculate scale (take minimum to fit target box)
+    scale = std::min(static_cast<float>(aligned_dst_width) / src_width, static_cast<float>(aligned_dst_height) / src_height);
+
+    // Calculate scaled dimensions
+    int resized_w = static_cast<int>(src_width * scale);
+    int resized_h = static_cast<int>(src_height * scale);
+
+    // Ensure scaled dimensions are multiples of 4
+    resized_w = (resized_w + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+    resized_h = (resized_h + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+
+    int aligned_src_width = (src_width + aligned_width_ - 1) & ~(aligned_width_ - 1);  // Round up to nearest multiple of aligned_width_
+
+    // 1. Get source buffer
+    BufferKey src_key{aligned_src_width, src_height, channels};
+    auto& src_buffer = GetOrCreateBuffer(src_key);
+
+    // 2. Get destination buffer
+    BufferKey dst_key{aligned_dst_width, aligned_dst_height, channels};
+    auto& dst_buffer = GetOrCreateBuffer(dst_key, false);
+
+    // 3. Copy source data to RGA buffer
+    for (int y = 0; y < src_height; y++) {
+        memcpy(static_cast<uint8_t*>(src_buffer.virtual_addr) + y * aligned_src_width * channels, src_data + y * src_width * channels,
+               src_width * channels);
+    }
+
+    dma_sync_cpu_to_device(src_buffer.dma_fd);
+    dma_sync_cpu_to_device(dst_buffer.dma_fd);
+
+    // 4. Set source and destination regions
+    im_rect src_rect = {0, 0, src_width, src_height};
+    im_rect dst_rect = {0, 0, resized_w, resized_h};  // Image placed in top-left corner
+
+    int ret = imcheck(src_buffer.buffer, dst_buffer.buffer, src_rect, dst_rect);
+    if (IM_STATUS_NOERROR != ret) {
+        INSPIRECV_LOG(ERROR) << "RGA parameter check failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 5. Fill entire destination area with black first
+    im_rect fill_rect = {0, 0, aligned_dst_width, aligned_dst_height};
+    ret = imfill(dst_buffer.buffer, fill_rect, 0x000000);  // Fill with black
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA fill failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    // 6. Execute resize operation, image will be placed in top-left corner
+    ret = improcess(src_buffer.buffer, dst_buffer.buffer, {}, src_rect, dst_rect, {}, IM_SYNC);
+    if (ret != IM_STATUS_SUCCESS) {
+        INSPIRECV_LOG(ERROR) << "RGA resize failed: " << imStrError((IM_STATUS)ret);
+        return -1;
+    }
+
+    dma_sync_device_to_cpu(dst_buffer.dma_fd);
+
+    // 7. Return processed data
+    *dst_data = static_cast<uint8_t*>(dst_buffer.virtual_addr);
+    return 0;
+}
+
+}  // namespace nexus
+
+}  // namespace inspire
+
+#endif  // ISF_ENABLE_RGA
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.h b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.h
new file mode 100644
index 000000000..c40b4a547
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.h
@@ -0,0 +1,216 @@
+#ifndef INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_RGA_H
+#define INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_RGA_H
+
+#if defined(ISF_ENABLE_RGA)
+
+#include "image_processor.h"
+#include <linux/stddef.h>
+#include <iostream>
+#include <stdint.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <errno.h>
+#include <time.h>
+#include <sys/types.h>
+#include <sys/time.h>
+#include <sys/mman.h>
+#include <fcntl.h>
+#include <signal.h>
+#include <unistd.h>
+#include <memory>
+#include <unordered_map>
+#include "im2d.hpp"
+#include "im2d_single.h"
+#include "RgaUtils.h"
+#include "rga/utils.h"
+#include "rga/dma_alloc.h"
+#include <launch.h>
+
+namespace inspire {
+
+namespace nexus {
+
+class INSPIRE_API_EXPORT RgaImageProcessor : public ImageProcessor {
+public:
+    RgaImageProcessor();
+    ~RgaImageProcessor() override;
+
+    int32_t Resize(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data, int dst_width, int dst_height) override;
+
+    int32_t SwapColor(const uint8_t* src_data, int src_width, int src_height, int channels, uint8_t** dst_data) override;
+
+    int32_t Padding(const uint8_t* src_data, int src_width, int src_height, int channels, int top, int bottom, int left, int right,
+                    uint8_t** dst_data, int& dst_width, int& dst_height) override;
+
+    int32_t ResizeAndPadding(const uint8_t* src_data, int src_width, int src_height, int channels, int dst_width, int dst_height, uint8_t** dst_data,
+                             float& scale) override;
+
+    int32_t MarkDone() override;
+
+    int32_t GetAlignedWidth(int width) const override;
+
+    void SetAlignedWidth(int width) override;
+
+public:
+    struct BufferInfo {
+        int dma_fd;
+        int width;
+        int height;
+        int channels;
+        size_t buffer_size;
+    };
+
+    BufferInfo GetCurrentSrcBufferInfo() const {
+        auto it = buffer_cache_.find(last_src_key_);
+        if (it != buffer_cache_.end()) {
+            const auto& buffer = it->second;
+            return {buffer.dma_fd, buffer.width, buffer.height, buffer.channels, buffer.buffer_size};
+        }
+        return {-1, 0, 0, 0, 0};  // Return invalid values if cache doesn't exist
+    }
+
+    BufferInfo GetCurrentDstBufferInfo() const {
+        auto it = buffer_cache_.find(last_dst_key_);
+        if (it != buffer_cache_.end()) {
+            const auto& buffer = it->second;
+            return {buffer.dma_fd, buffer.width, buffer.height, buffer.channels, buffer.buffer_size};
+        }
+        return {-1, 0, 0, 0, 0};  // Return invalid values if cache doesn't exist
+    }
+
+    size_t GetCacheSize() const {
+        return buffer_cache_.size();
+    }
+
+    void DumpCacheStatus() const override {
+        INSPIRECV_LOG(INFO) << "Current cache status:";
+        INSPIRECV_LOG(INFO) << "Cache size: " << buffer_cache_.size();
+
+        auto src_info = GetCurrentSrcBufferInfo();
+        INSPIRECV_LOG(INFO) << "Source buffer: "
+                            << "dma_fd=" << src_info.dma_fd << ", size=" << src_info.width << "x" << src_info.height << "x" << src_info.channels;
+
+        auto dst_info = GetCurrentDstBufferInfo();
+        INSPIRECV_LOG(INFO) << "Destination buffer: "
+                            << "dma_fd=" << dst_info.dma_fd << ", size=" << dst_info.width << "x" << dst_info.height << "x" << dst_info.channels;
+    }
+
+private:
+    struct RGABuffer {
+        int width{0};
+        int height{0};
+        int channels{0};
+        int dma_fd{-1};
+        void* virtual_addr{nullptr};
+        size_t buffer_size{0};
+        rga_buffer_handle_t handle{0};
+        rga_buffer_t buffer{};
+
+        bool Allocate(int w, int h, int c) {
+            width = w;
+            height = h;
+            channels = c;
+            buffer_size = width * height * channels;
+
+            int ret = dma_buf_alloc(INSPIREFACE_CONTEXT->GetRockchipDmaHeapPath().c_str(), buffer_size, &dma_fd, &virtual_addr);
+            if (ret < 0) {
+                INSPIRECV_LOG(ERROR) << "Failed to allocate DMA buffer: " << ret;
+                return false;
+            }
+
+            handle = importbuffer_fd(dma_fd, buffer_size);
+            if (handle == 0) {
+                INSPIRECV_LOG(ERROR) << "Failed to import buffer";
+                Release();
+                return false;
+            }
+
+            buffer = wrapbuffer_handle(handle, w, h, RK_FORMAT_RGB_888);
+
+            return true;
+        }
+
+        void Release() {
+            if (handle) {
+                releasebuffer_handle(handle);
+                handle = 0;
+            }
+            if (dma_fd >= 0) {
+                dma_buf_free(buffer_size, &dma_fd, virtual_addr);
+                dma_fd = -1;
+                virtual_addr = nullptr;
+            }
+        }
+
+        ~RGABuffer() {
+            Release();
+        }
+    };
+
+    struct BufferKey {
+        int width;
+        int height;
+        int channels;
+
+        bool operator==(const BufferKey& other) const {
+            return width == other.width && height == other.height && channels == other.channels;
+        }
+    };
+
+    struct BufferKeyHash {
+        std::size_t operator()(const BufferKey& key) const {
+            return std::hash<int>()(key.width) ^ (std::hash<int>()(key.height) << 1) ^ (std::hash<int>()(key.channels) << 2);
+        }
+    };
+
+    RGABuffer& GetOrCreateBuffer(const BufferKey& key, bool is_src = true) {
+        auto it = buffer_cache_.find(key);
+        if (it != buffer_cache_.end()) {
+            if (is_src) {
+                last_src_key_ = key;
+            } else {
+                last_dst_key_ = key;
+            }
+            return it->second;
+        }
+
+        if (buffer_cache_.size() >= 3) {  // Keep max 3 buffers in cache
+            for (auto it = buffer_cache_.begin(); it != buffer_cache_.end();) {
+                if (!(it->first == last_src_key_) && !(it->first == last_dst_key_)) {
+                    it = buffer_cache_.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+        }
+
+        auto& buffer = buffer_cache_[key];
+        if (!buffer.Allocate(key.width, key.height, key.channels)) {
+            INSPIRECV_LOG(ERROR) << "Failed to allocate RGA buffer";
+            throw std::runtime_error("RGA buffer allocation failed");
+        }
+
+        if (is_src) {
+            last_src_key_ = key;
+        } else {
+            last_dst_key_ = key;
+        }
+
+        return buffer;
+    }
+
+private:
+    std::unordered_map<BufferKey, RGABuffer, BufferKeyHash> buffer_cache_;
+    BufferKey last_src_key_{0, 0, 0};
+    BufferKey last_dst_key_{0, 0, 0};
+    int32_t aligned_width_{0};
+};
+
+}  // namespace nexus
+
+}  // namespace inspire
+
+#endif  // ISF_ENABLE_RGA
+
+#endif  // INSPIRE_FACE_NEXUS_IMAGE_PROCESSOR_RGA_H
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.cpp
new file mode 100644
index 000000000..4ef4e8f26
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.cpp
@@ -0,0 +1,133 @@
+#if defined(ISF_ENABLE_RGA)
+#include <getopt.h>
+#include <sys/mman.h>
+#include <sys/socket.h>
+#include <sys/stat.h>
+#include <time.h>
+#include <stdbool.h>
+#include <assert.h>
+#include <endian.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/poll.h>
+#include <unistd.h>
+#include <stdbool.h>
+#include <sys/eventfd.h>
+
+#include <sched.h>
+#include <pthread.h>
+
+#include <stdint.h>
+#include <math.h>
+#include <memory.h>
+#include <sys/time.h>
+
+#include "dma_alloc.h"
+#include "RgaUtils.h"
+
+typedef unsigned long long __u64;
+typedef unsigned int __u32;
+
+struct dma_heap_allocation_data {
+    __u64 len;
+    __u32 fd;
+    __u32 fd_flags;
+    __u64 heap_flags;
+};
+
+#define DMA_HEAP_IOC_MAGIC 'H'
+#define DMA_HEAP_IOCTL_ALLOC _IOWR(DMA_HEAP_IOC_MAGIC, 0x0, struct dma_heap_allocation_data)
+
+#define DMA_BUF_SYNC_READ (1 << 0)
+#define DMA_BUF_SYNC_WRITE (2 << 0)
+#define DMA_BUF_SYNC_RW (DMA_BUF_SYNC_READ | DMA_BUF_SYNC_WRITE)
+#define DMA_BUF_SYNC_START (0 << 2)
+#define DMA_BUF_SYNC_END (1 << 2)
+
+struct dma_buf_sync {
+    __u64 flags;
+};
+
+#define DMA_BUF_BASE 'b'
+#define DMA_BUF_IOCTL_SYNC _IOW(DMA_BUF_BASE, 0, struct dma_buf_sync)
+
+#define CMA_HEAP_SIZE 1024 * 1024
+
+int dma_sync_device_to_cpu(int fd) {
+    struct dma_buf_sync sync = {0};
+
+    sync.flags = DMA_BUF_SYNC_START | DMA_BUF_SYNC_RW;
+    return ioctl(fd, DMA_BUF_IOCTL_SYNC, &sync);
+}
+
+int dma_sync_cpu_to_device(int fd) {
+    struct dma_buf_sync sync = {0};
+
+    sync.flags = DMA_BUF_SYNC_END | DMA_BUF_SYNC_RW;
+    return ioctl(fd, DMA_BUF_IOCTL_SYNC, &sync);
+}
+
+int dma_buf_alloc(const char *path, size_t size, int *fd, void **va) {
+    int ret;
+    int prot;
+    void *mmap_va;
+    int dma_heap_fd = -1;
+    struct dma_heap_allocation_data buf_data;
+
+    /* open dma_heap fd */
+    dma_heap_fd = open(path, O_RDWR);
+    if (dma_heap_fd < 0) {
+        printf("open %s fail!\n", path);
+        return dma_heap_fd;
+    }
+
+    /* alloc buffer */
+    memset(&buf_data, 0x0, sizeof(struct dma_heap_allocation_data));
+
+    buf_data.len = size;
+    buf_data.fd_flags = O_CLOEXEC | O_RDWR;
+    ret = ioctl(dma_heap_fd, DMA_HEAP_IOCTL_ALLOC, &buf_data);
+    if (ret < 0) {
+        printf("RK_DMA_HEAP_ALLOC_BUFFER failed\n");
+        return ret;
+    }
+
+    /* mmap va */
+    if (fcntl(buf_data.fd, F_GETFL) & O_RDWR)
+        prot = PROT_READ | PROT_WRITE;
+    else
+        prot = PROT_READ;
+
+    /* mmap contiguors buffer to user */
+    mmap_va = (void *)mmap(NULL, buf_data.len, prot, MAP_SHARED, buf_data.fd, 0);
+    if (mmap_va == MAP_FAILED) {
+        printf("mmap failed: %s\n", strerror(errno));
+        return -errno;
+    }
+
+    *va = mmap_va;
+    *fd = buf_data.fd;
+
+    close(dma_heap_fd);
+
+    return 0;
+}
+
+void dma_buf_free(size_t size, int *fd, void *va) {
+    int len;
+
+    len = size;
+    munmap(va, len);
+
+    close(*fd);
+    *fd = -1;
+}
+
+#endif  // ISF_ENABLE_RGA
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.h b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.h
new file mode 100644
index 000000000..409009233
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/dma_alloc.h
@@ -0,0 +1,21 @@
+#ifndef RGA_ALLOCATOR_DMA_ALLOC_H__
+#define RGA_ALLOCATOR_DMA_ALLOC_H__
+
+#if defined(ISF_ENABLE_RGA)
+
+#define DMA_HEAP_UNCACHE_PATH "/dev/dma_heap/system-uncached"
+#define DMA_HEAP_PATH "/dev/dma_heap/system"
+#define DMA_HEAP_DMA32_UNCACHE_PATCH "/dev/dma_heap/system-uncached-dma32"
+#define DMA_HEAP_DMA32_PATCH "/dev/dma_heap/system-dma32"
+#define CMA_HEAP_UNCACHE_PATH "/dev/dma_heap/cma-uncached"
+#define RV1106_CMA_HEAP_PATH "/dev/rk_dma_heap/rk-dma-heap-cma"
+
+int dma_sync_device_to_cpu(int fd);
+int dma_sync_cpu_to_device(int fd);
+
+int dma_buf_alloc(const char *path, size_t size, int *fd, void **va);
+void dma_buf_free(size_t size, int *fd, void *va);
+
+#endif /* #ifndef RGA_ALLOCATOR_DMA_ALLOC_H__ */
+
+#endif  // ISF_ENABLE_RGA
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.cpp b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.cpp
new file mode 100644
index 000000000..b5e68bbf4
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.cpp
@@ -0,0 +1,183 @@
+#if defined(ISF_ENABLE_RGA)
+
+#include <iostream>
+#include <fstream>
+#include <sstream>
+#include <cstddef>
+#include <cmath>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/time.h>
+#include <unistd.h>
+
+#include "RgaUtils.h"
+
+int64_t get_cur_us() {
+    struct timeval tv;
+    gettimeofday(&tv, NULL);
+    return tv.tv_sec * 1000000 + tv.tv_usec;
+}
+
+int64_t get_cur_ms() {
+    struct timeval tv;
+    gettimeofday(&tv, NULL);
+    return tv.tv_sec * 1000 + tv.tv_usec / 1000;
+}
+
+void draw_rgba(char *buffer, int width, int height) {
+    for (int i = 0; i < height; i++) {
+        for (int j = 0; j < width / 4; j++) {
+            buffer[(i * width * 4) + j * 4 + 0] = 0xff;  // R
+            buffer[(i * width * 4) + j * 4 + 1] = 0x00;  // G
+            buffer[(i * width * 4) + j * 4 + 2] = 0x00;  // B
+            buffer[(i * width * 4) + j * 4 + 3] = 0xff;  // A
+        }
+        for (int j = width / 4; j < width / 4 * 2; j++) {
+            buffer[(i * width * 4) + j * 4 + 0] = 0x00;
+            buffer[(i * width * 4) + j * 4 + 1] = 0xff;
+            buffer[(i * width * 4) + j * 4 + 2] = 0x00;
+            buffer[(i * width * 4) + j * 4 + 3] = 0xff;
+        }
+        for (int j = width / 4 * 2; j < width / 4 * 3; j++) {
+            buffer[(i * width * 4) + j * 4 + 0] = 0x00;
+            buffer[(i * width * 4) + j * 4 + 1] = 0x00;
+            buffer[(i * width * 4) + j * 4 + 2] = 0xff;
+            buffer[(i * width * 4) + j * 4 + 3] = 0xff;
+        }
+        for (int j = width / 4 * 3; j < width; j++) {
+            buffer[(i * width * 4) + j * 4 + 0] = 0xff;
+            buffer[(i * width * 4) + j * 4 + 1] = 0xff;
+            buffer[(i * width * 4) + j * 4 + 2] = 0xff;
+            buffer[(i * width * 4) + j * 4 + 3] = 0xff;
+        }
+    }
+}
+
+void draw_YUV420(char *buffer, int width, int height) {
+    /* Y channel */
+    memset(buffer, 0xa8, width * height / 2);
+    memset(buffer + width * height / 2, 0x54, width * height / 2);
+    /* UV channel */
+    memset(buffer + width * height, 0x80, width * height / 4);
+    memset(buffer + (int)(width * height * 1.25), 0x30, width * height / 4);
+}
+
+void draw_YUV422(char *buffer, int width, int height) {
+    /* Y channel */
+    memset(buffer, 0xa8, width * height / 2);
+    memset(buffer + width * height / 2, 0x54, width * height / 2);
+    /* UV channel */
+    memset(buffer + width * height, 0x80, width * height / 2);
+    memset(buffer + (int)(width * height * 1.5), 0x30, width * height / 2);
+}
+
+void draw_gray256(char *buffer, int width, int height) {
+    for (int i = 0; i < height; i++) {
+        for (int j = 0; j < width / 4; j++) {
+            buffer[(i * width * 4) + j * 4] = 0xa8;
+        }
+        for (int j = width / 4; j < width / 4 * 2; j++) {
+            buffer[(i * width * 4) + j * 4] = 0x80;
+        }
+        for (int j = width / 4 * 2; j < width / 4 * 3; j++) {
+            buffer[(i * width * 4) + j * 4] = 0x54;
+        }
+        for (int j = width / 4 * 3; j < width; j++) {
+            buffer[(i * width * 4) + j * 4] = 0x30;
+        }
+    }
+}
+
+int read_image_from_fbc_file(void *buf, const char *path, int sw, int sh, int fmt, int index) {
+    int size;
+    char filePath[100];
+    const char *inputFbcFilePath = "%s/in%dw%d-h%d-%s-fbc.bin";
+
+    snprintf(filePath, 100, inputFbcFilePath, path, index, sw, sh, translate_format_str(fmt));
+
+    FILE *file = fopen(filePath, "rb");
+    if (!file) {
+        fprintf(stderr, "Could not open %s\n", filePath);
+        return -EINVAL;
+    }
+
+    size = sw * sh * get_bpp_from_format(fmt) * 1.5;
+
+    fread(buf, size, 1, file);
+
+    fclose(file);
+
+    return 0;
+}
+
+int read_image_from_file(void *buf, const char *path, int sw, int sh, int fmt, int index) {
+    int size;
+    char filePath[100];
+    const char *inputFilePath = "%s/in%dw%d-h%d-%s.bin";
+
+    snprintf(filePath, 100, inputFilePath, path, index, sw, sh, translate_format_str(fmt));
+
+    FILE *file = fopen(filePath, "rb");
+    if (!file) {
+        fprintf(stderr, "Could not open %s\n", filePath);
+        return -EINVAL;
+    }
+
+    size = sw * sh * get_bpp_from_format(fmt);
+
+    fread(buf, size, 1, file);
+
+    fclose(file);
+
+    return 0;
+}
+
+int write_image_to_fbc_file(void *buf, const char *path, int sw, int sh, int fmt, int index) {
+    int size;
+    char filePath[100];
+    const char *outputFbcFilePath = "%s/out%dw%d-h%d-%s-fbc.bin";
+
+    snprintf(filePath, 100, outputFbcFilePath, path, index, sw, sh, translate_format_str(fmt));
+
+    FILE *file = fopen(filePath, "wb+");
+    if (!file) {
+        fprintf(stderr, "Could not open %s\n", filePath);
+        return false;
+    } else {
+        fprintf(stderr, "open %s and write ok\n", filePath);
+    }
+
+    size = sw * sh * get_bpp_from_format(fmt) * 1.5;
+
+    fwrite(buf, size, 1, file);
+
+    fclose(file);
+
+    return 0;
+}
+
+int write_image_to_file(void *buf, const char *path, int sw, int sh, int fmt, int index) {
+    int size;
+    char filePath[100];
+    const char *outputFilePath = "%s/out%dw%d-h%d-%s.bin";
+
+    snprintf(filePath, 100, outputFilePath, path, index, sw, sh, translate_format_str(fmt));
+
+    FILE *file = fopen(filePath, "wb+");
+    if (!file) {
+        fprintf(stderr, "Could not open %s\n", filePath);
+        return false;
+    } else {
+        fprintf(stderr, "open %s and write ok\n", filePath);
+    }
+
+    size = sw * sh * get_bpp_from_format(fmt);
+
+    fwrite(buf, size, 1, file);
+
+    fclose(file);
+
+    return 0;
+}
+
+#endif  // ISF_ENABLE_RGA
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.h b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.h
new file mode 100644
index 000000000..4fa26d793
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/rga/utils.h
@@ -0,0 +1,21 @@
+#ifndef RGA_UTILS_H__
+#define RGA_UTILS_H__
+
+#include <stdlib.h>
+
+#if defined(ISF_ENABLE_RGA)
+
+int64_t get_cur_us();
+int64_t get_cur_ms();
+void draw_rgba(char *buffer, int width, int height);
+void draw_YUV420(char *buffer, int width, int height);
+void draw_YUV422(char *buffer, int width, int height);
+void draw_gray256(char *buffer, int width, int height);
+int read_image_from_fbc_file(void *buf, const char *path, int sw, int sh, int fmt, int index);
+int read_image_from_file(void *buf, const char *path, int sw, int sh, int fmt, int index);
+int write_image_to_fbc_file(void *buf, const char *path, int sw, int sh, int fmt, int index);
+int write_image_to_file(void *buf, const char *path, int sw, int sh, int fmt, int index);
+
+#endif  // ISF_ENABLE_RGA
+
+#endif /* #ifndef RGA_UTILS_H__ */
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/cuda_toolkit.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/cuda_toolkit.h
new file mode 100644
index 000000000..14cf649d8
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/cuda_toolkit.h
@@ -0,0 +1,22 @@
+#ifndef INSPIRE_CUDA_TOOLKIT_H
+#define INSPIRE_CUDA_TOOLKIT_H
+
+#include "data_type.h"
+
+namespace inspire {
+
+// Get the number of CUDA devices
+int32_t INSPIRE_API_EXPORT GetCudaDeviceCount(int32_t *device_count);
+
+// Check the availability of CUDA
+int32_t INSPIRE_API_EXPORT CheckCudaUsability(int32_t *is_support);
+
+// Internal function, print detailed information of CUDA devices
+int32_t INSPIRE_API_EXPORT _PrintCudaDeviceInfo();
+
+// Wrapper function to print CUDA device information
+int32_t INSPIRE_API_EXPORT PrintCudaDeviceInfo();
+
+}  // namespace inspire
+
+#endif  // INSPIRE_CUDA_TOOLKIT_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/data_type.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/data_type.h
new file mode 100644
index 000000000..172810c4d
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/data_type.h
@@ -0,0 +1,320 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_DATATYPE_H
+#define INSPIRE_FACE_DATATYPE_H
+
+#include <inspirecv/inspirecv.h>
+
+#include <cstdint>
+#if defined(_WIN32) && (defined(_DEBUG) || defined(DEBUG))
+#define _CRTDBG_MAP_ALLOC
+#include "crtdbg.h"
+#endif
+
+#ifndef INSPIRE_API
+#define INSPIRE_API
+#endif
+
+#if defined(_WIN32)
+#ifdef ISF_BUILD_SHARED_LIBS
+#define INSPIRE_API_EXPORT __declspec(dllexport)
+#else
+#define INSPIRE_API_EXPORT
+#endif
+#else
+#define INSPIRE_API_EXPORT __attribute__((visibility("default")))
+#endif  // _WIN32
+
+#ifndef M_PI
+#define M_PI 3.14159265358979323846264338327950288
+#endif
+
+namespace inspire {
+
+/**
+ * @defgroup DataType Definitions
+ * @brief Defines various data types used in the HyperFace project.
+ * @{
+ */
+
+#if !defined(int64)
+/** @typedef int64
+ *  @brief 64-bit integer type.
+ */
+typedef int64_t int64;
+#endif
+
+#if !defined(uint64)
+/** @typedef uint64
+ *  @brief 64-bit unsigned integer type.
+ */
+typedef uint64_t uint64;
+#endif
+
+#if !defined(int32)
+/** @typedef int32
+ *  @brief 32-bit integer type.
+ */
+typedef int32_t int32;
+#endif
+
+#if !defined(uint32)
+/** @typedef uint32
+ *  @brief 32-bit unsigned integer type.
+ */
+typedef uint32_t uint32;
+#endif
+
+#if !defined(int8)
+/** @typedef int8
+ *  @brief 8-bit integer type.
+ */
+typedef int8_t int8;
+#endif
+
+#if !defined(uint8)
+/** @typedef uint8
+ *  @brief 8-bit unsigned integer type.
+ */
+typedef uint8_t uint8;
+#endif
+
+/** @typedef ByteArray
+ *  @brief Type definition for a byte array (vector of chars).
+ */
+typedef std::vector<char> ByteArray;
+
+/** @typedef Point2i
+ *  @brief 2D coordinate point with integer precision.
+ */
+typedef inspirecv::Point2i Point2i;
+
+/** @typedef Point2f
+ *  @brief 2D coordinate point with float precision.
+ */
+typedef inspirecv::Point2f Point2f;
+
+/** @typedef PointsList2i
+ *  @brief List of 2D coordinate points with integer precision.
+ */
+typedef std::vector<Point2i> PointsList2i;
+
+/** @typedef PointsList2f
+ *  @brief List of 2D coordinate points with float precision.
+ */
+typedef std::vector<Point2f> PointsList2f;
+
+/** @typedef Contours2i
+ *  @brief Contours represented as a list of 2D integer points.
+ */
+typedef std::vector<PointsList2i> Contours2i;
+
+/** @typedef Contours2f
+ *  @brief Contours represented as a list of 2D float points.
+ */
+typedef std::vector<PointsList2f> Contours2f;
+
+/** @typedef Textures2i
+ *  @brief Texture lines represented as integer contours.
+ */
+typedef Contours2i Textures2i;
+
+/** @typedef AnyTensorFp32
+ *  @brief Generic tensor representation using a vector of floats.
+ */
+typedef std::vector<float> AnyTensorFp32;
+
+/** @typedef ImageBitmap
+ *  @brief Image bitmap representation.
+ */
+typedef inspirecv::Image ImageBitmap;
+
+/** @typedef Rectangle
+ *  @brief Rectangle representation using integer values.
+ */
+typedef inspirecv::Rect<int> Rectangle;
+
+/** @typedef Size
+ *  @brief Size representation using integer values.
+ */
+typedef inspirecv::Size<int> Size;
+
+/** @typedef Embedded
+ *  @brief Dense vector for feature embedding.
+ */
+typedef std::vector<float> Embedded;
+
+/** @typedef EmbeddedList
+ *  @brief List of dense vectors for feature embedding.
+ */
+typedef std::vector<Embedded> EmbeddedList;
+
+/** @typedef String
+ *  @brief String type definition.
+ */
+typedef std::string String;
+
+/** @typedef IndexList
+ *  @brief List of indices.
+ */
+typedef std::vector<int> IndexList;
+
+/**
+ * @enum DetectMode
+ * @brief Enumeration for different detection modes.
+ */
+enum DetectModuleMode {
+    DETECT_MODE_ALWAYS_DETECT = 0,  ///< Detection mode: Always detect
+    DETECT_MODE_LIGHT_TRACK,        ///< Detection mode: Light face track
+    DETECT_MODE_TRACK_BY_DETECT,    ///< Detection mode: Tracking by detection
+};
+
+/**
+ * @struct CustomPipelineParameter
+ * @brief Structure to hold custom parameters for the face detection and processing pipeline.
+ *
+ * Includes options for enabling various features such as recognition, liveness detection, and quality assessment.
+ */
+typedef struct CustomPipelineParameter {
+    bool enable_recognition = false;           ///< Enable face recognition feature
+    bool enable_liveness = false;              ///< Enable RGB liveness detection feature
+    bool enable_ir_liveness = false;           ///< Enable IR (Infrared) liveness detection feature
+    bool enable_mask_detect = false;           ///< Enable mask detection feature
+    bool enable_face_attribute = false;        ///< Enable face attribute prediction feature
+    bool enable_face_quality = false;          ///< Enable face quality assessment feature
+    bool enable_interaction_liveness = false;  ///< Enable interactive liveness detection feature
+    bool enable_face_pose = false;             ///< Enable face pose estimation feature
+    bool enable_face_emotion = false;          ///< Enable face emotion recognition feature
+} ContextCustomParameter;
+
+/** @struct FaceLoc
+ *  @brief Struct representing standardized face landmarks for detection.
+ *
+ *  Contains coordinates for the face, detection score, and landmarks.
+ */
+typedef struct FaceLoc {
+    float x1;
+    float y1;
+    float x2;
+    float y2;
+    float score;
+    float lmk[10];
+} FaceLoc;
+
+/** @typedef FaceLocList
+ *  @brief List of FaceLoc structures.
+ */
+typedef std::vector<FaceLoc> FaceLocList;
+
+/** @struct FaceBasicData
+ *  @brief Struct for basic face data.
+ *
+ *  Contains the size of the data and a pointer to the data.
+ */
+typedef struct FaceBasicData {
+    int32_t dataSize;
+    void* data;
+} FaceBasicData;
+
+/** @struct FaceFeatureEntity
+ *  @brief Struct for face feature data.
+ *
+ *  Contains the size of the feature data and a pointer to the feature array.
+ */
+typedef struct FaceFeatureEntity {
+    int32_t dataSize;
+    float* data;
+} FaceFeaturePtr;
+
+// Search for most similar vectors
+struct FaceSearchResult {
+    int64_t id;
+    double similarity;
+    std::vector<float> feature;
+};
+
+/** @struct FaceEmbedding
+ *  @brief Struct for face embedding data.
+ *
+ *  Contains the isNormal flag and the embedding vector.
+ */
+struct FaceEmbedding {
+    int32_t isNormal;
+    float norm;
+    Embedded embedding;
+};
+
+/** @struct FaceInteractionState
+ *  @brief Struct for face interaction state data.
+ *
+ *  Contains the confidence scores for face interaction.
+ */
+struct FaceInteractionState {
+    float left_eye_status_confidence;
+    float right_eye_status_confidence;
+};
+
+/** @struct FaceInteractionAction
+ *  @brief Struct for face interaction action data.
+ *
+ *  Contains the actions for face interaction.
+ */
+struct FaceInteractionAction {
+    int32_t normal;     ///< Normal action.
+    int32_t shake;      ///< Shake action.
+    int32_t jawOpen;    ///< Jaw open action.
+    int32_t headRaise;  ///< Head raise action.
+    int32_t blink;      ///< Blink action.
+};
+
+/** @struct FaceAttributeResult
+ *  @brief Struct for face attribute result data.
+ *
+ *  Contains the results for face attribute.
+ */
+struct FaceAttributeResult {
+    int32_t race;        ///< Race of the detected face.
+                         ///< 0: Black;
+                         ///< 1: Asian;
+                         ///< 2: Latino/Hispanic;
+                         ///< 3: Middle Eastern;
+                         ///< 4: White;
+    int32_t gender;      ///< Gender of the detected face.
+                         ///< 0: Female;
+                         ///< 1: Male;
+    int32_t ageBracket;  ///< Age bracket of the detected face.
+                         ///< 0: 0-2 years old;
+                         ///< 1: 3-9 years old;
+                         ///< 2: 10-19 years old;
+                         ///< 3: 20-29 years old;
+                         ///< 4: 30-39 years old;
+                         ///< 5: 40-49 years old;
+                         ///< 6: 50-59 years old;
+                         ///< 7: 60-69 years old;
+                         ///< 8: more than 70 years old;
+};
+
+/** @struct FaceEmotionResult
+ *  @brief Struct for face emotion result data.
+ *
+ *  Contains the results for face emotion.
+ */
+struct FaceEmotionResult {
+    int32_t emotion;     ///< Emotion of the detected face.
+                         ///< 0: Neutral;
+                         ///< 1: Happy;
+                         ///< 2: Sad;
+                         ///< 3: Surprise;
+                         ///< 4: Fear;
+                         ///< 5: Disgust;
+                         ///< 6: Anger;
+};
+
+/** @} */
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_DATATYPE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/face_wrapper.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/face_wrapper.h
new file mode 100644
index 000000000..96c36353a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/face_wrapper.h
@@ -0,0 +1,73 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+// Include guard to prevent double inclusion of this header file
+#pragma once
+#ifndef INSPIRE_FACE_FACEDATATYPE_H
+#define INSPIRE_FACE_FACEDATATYPE_H
+
+#include "data_type.h"
+
+namespace inspire {
+
+/**
+ * Struct to represent 3D angles of a face.
+ */
+typedef struct Face3DAngle {
+    float roll;   ///< Roll angle
+    float yaw;    ///< Yaw angle
+    float pitch;  ///< Pitch angle
+} Face3DAngle;
+
+/**
+ * Struct to represent the rectangle coordinates of a face.
+ */
+typedef struct FaceRect {
+    int x;       ///< X-coordinate of the top-left corner
+    int y;       ///< Y-coordinate of the top-left corner
+    int width;   ///< Width of the rectangle
+    int height;  ///< Height of the rectangle
+} FaceRect;
+
+/**
+ * Struct to represent 2D point coordinates.
+ */
+typedef struct Point2F {
+    float x;  ///< X-coordinate
+    float y;  ///< Y-coordinate
+} HPoint;
+
+/**
+ * Struct to represent a 2D transformation matrix.
+ */
+typedef struct TransMatrix {
+    double m00;  ///< Element (0,0) of the matrix
+    double m01;  ///< Element (0,1) of the matrix
+    double m10;  ///< Element (1,0) of the matrix
+    double m11;  ///< Element (1,1) of the matrix
+    double tx;   ///< Translation in the X-axis
+    double ty;   ///< Translation in the Y-axis
+} TransMatrix;
+
+/**
+ * Struct to represent basic face data.
+ */
+typedef struct FaceTrackWrap {
+    int trackState;                ///< Track state
+    int inGroupIndex;              ///< Index within a group
+    int trackId;                   ///< Track ID
+    int trackCount;                ///< Track count
+    FaceRect rect;                 ///< Face rectangle
+    TransMatrix trans;             ///< Transformation matrix
+    Point2F keyPoints[5];          ///< Key points (e.g., landmarks)
+    Face3DAngle face3DAngle;       ///< 3D face angles
+    float quality[5];              ///< Quality values for key points
+    Point2F densityLandmark[106];  ///< Face density landmark
+    int densityLandmarkEnable;     ///< Density landmark enable
+} FaceTrackWrap;
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_FACEDATATYPE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/feature_hub_db.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/feature_hub_db.h
new file mode 100644
index 000000000..9ec0b855b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/feature_hub_db.h
@@ -0,0 +1,251 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FEATURE_HUB_DB_H
+#define INSPIRE_FEATURE_HUB_DB_H
+
+#include <memory>
+#include <vector>
+#include <string>
+#include "data_type.h"
+#include <mutex>
+
+#define INSPIREFACE_FEATURE_HUB inspire::FeatureHubDB::GetInstance()
+#define INSPIRE_INVALID_ID -1
+
+namespace inspire {
+
+typedef enum SearchMode {
+    SEARCH_MODE_EAGER = 0,   // Eager mode: Stops when a vector meets the threshold.
+    SEARCH_MODE_EXHAUSTIVE,  // Exhaustive mode: Searches until the best match is found.
+} SearchMode;
+
+typedef enum PrimaryKeyMode {
+    AUTO_INCREMENT = 0,  // Auto-increment primary key
+    MANUAL_INPUT,        // Manual input primary key
+} PrimaryKeyMode;
+
+/**
+ * @struct DatabaseConfiguration
+ * @brief Structure to configure database settings for FaceRecognition.
+ */
+using DatabaseConfiguration = struct DatabaseConfiguration {
+    PrimaryKeyMode primary_key_mode = PrimaryKeyMode::AUTO_INCREMENT;  ///< Primary key mode
+    bool enable_persistence = false;                                   ///< Whether to enable data persistence.
+    std::string persistence_db_path;                                   ///< Path to the database file.
+    float recognition_threshold = 0.48f;                               ///< Face search threshold
+    SearchMode search_mode = SEARCH_MODE_EAGER;                        ///< Search mode (!!Temporarily unavailable!!)
+};
+
+/**
+ * @class FeatureHubDB
+ * @brief Service for internal feature vector storage.
+ *
+ * This class provides methods for face feature extraction, registration, update, search, and more.
+ * It uses the PIMPL (Pointer to Implementation) pattern to hide implementation details.
+ */
+class INSPIRE_API_EXPORT FeatureHubDB {
+public:
+    /**
+     * @brief Constructor for FeatureHubDB class.
+     */
+    FeatureHubDB();
+
+    /**
+     * @brief Destructor for FeatureHubDB class.
+     */
+    ~FeatureHubDB();
+
+    FeatureHubDB(const FeatureHubDB&) = delete;
+    FeatureHubDB& operator=(const FeatureHubDB&) = delete;
+
+    /**
+     * @brief Gets the singleton instance of FeatureHubDB.
+     * @return Shared pointer to the FeatureHubDB instance.
+     */
+    static std::shared_ptr<FeatureHubDB> GetInstance();
+
+    /**
+     * @brief Enables the feature hub with the specified configuration.
+     * @param configuration The database configuration settings.
+     * @return int32_t Returns a status code indicating success (0) or failure (non-zero).
+     */
+    int32_t EnableHub(const DatabaseConfiguration& configuration);
+
+    /**
+     * @brief Disables the feature hub, freeing all associated resources.
+     * @return int32_t Returns a status code indicating success (0) or failure (non-zero).
+     */
+    int32_t DisableHub();
+
+    /**
+     * @brief Get all ids in the database.
+     * @return int32_t Status code of the operation.
+     */
+    int32_t GetAllIds();
+
+    /**
+     * @brief Searches for a face feature within stored data.
+     * @param queryFeature Embedded feature to search for.
+     * @param searchResult SearchResult object to store search results.
+     * @param returnFeature Whether to return the feature data.
+     * @return int32_t Status code of the search operation.
+     */
+    int32_t SearchFaceFeature(const Embedded& queryFeature, FaceSearchResult& searchResult, bool returnFeature = true);
+
+    /**
+     * @brief Search the stored data for the top k facial features that are most similar.
+     * @param queryFeature Embedded feature to search for.
+     * @param topK Maximum number of results to return.
+     * @return int32_t Status code of the search operation.
+     */
+    int32_t SearchFaceFeatureTopKCache(const Embedded& queryFeature, size_t topK);
+
+    /**
+     * @brief Search the stored data for the top k facial features that are most similar.
+     * @param queryFeature Embedded feature to search for.
+     * @param searchResult Vector to store search results.
+     * @param topK Maximum number of results to return.
+     * @param returnFeature Whether to return the feature data.
+     * @return int32_t Status code of the search operation.
+     */
+    int32_t SearchFaceFeatureTopK(const Embedded& queryFeature, std::vector<FaceSearchResult>& searchResult, size_t topK, bool returnFeature = false);
+
+    /**
+     * @brief Inserts a face feature with a custom ID.
+     * @param feature Vector of floats representing the face feature.
+     * @param id ID for the feature.
+     * @param result_id Output parameter to store the resulting ID.
+     * @return int32_t Status code of the insertion operation.
+     */
+    int32_t FaceFeatureInsert(const std::vector<float>& feature, int32_t id, int64_t& result_id);
+
+    /**
+     * @brief Removes a face feature by its ID.
+     * @param id ID of the feature to remove.
+     * @return int32_t Status code of the removal operation.
+     */
+    int32_t FaceFeatureRemove(int32_t id);
+
+    /**
+     * @brief Updates a face feature by its ID.
+     * @param feature Vector of floats representing the new face feature.
+     * @param customId ID of the feature to update.
+     * @return int32_t Status code of the update operation.
+     */
+    int32_t FaceFeatureUpdate(const std::vector<float>& feature, int32_t customId);
+
+    /**
+     * @brief Retrieves a face feature by its ID.
+     * @param id ID of the feature to retrieve.
+     * @return int32_t Status code of the retrieval operation.
+     */
+    int32_t GetFaceFeature(int32_t id);
+
+    /**
+     * @brief Retrieves a face feature by its ID.
+     * @param id ID of the feature to retrieve.
+     * @param feature Vector to store the retrieved feature.
+     * @return int32_t Status code of the retrieval operation.
+     */
+    int32_t GetFaceFeature(int32_t id, std::vector<float>& feature);
+
+    /**
+     * @brief Retrieves a face feature by its ID.
+     * @param id ID of the feature to retrieve.
+     * @param feature Vector to store the retrieved feature.
+     * @return int32_t Status code of the retrieval operation.
+     */
+    int32_t GetFaceFeature(int32_t id, FaceEmbedding& feature);
+
+    /**
+     * @brief Views the database table containing face data.
+     * @return int32_t Status code of the operation.
+     */
+    int32_t ViewDBTable();
+
+    /**
+     * @brief Sets the recognition threshold for face recognition.
+     * @param threshold Float value of the new threshold.
+     */
+    void SetRecognitionThreshold(float threshold);
+
+    /**
+     * @brief Sets the search mode for face recognition.
+     * @param mode Search mode.
+     */
+    void SetRecognitionSearchMode(SearchMode mode);
+
+    /**
+     * @brief Computes the cosine similarity between two feature vectors.
+     * @param v1 First feature vector.
+     * @param v2 Second feature vector.
+     * @param res Output parameter to store the cosine similarity result.
+     * @param normalize Whether to normalize the vectors before computing similarity.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    static int32_t CosineSimilarity(const std::vector<float>& v1, const std::vector<float>& v2, float& res, bool normalize = false);
+
+    /**
+     * @brief Computes the cosine similarity between two feature vectors.
+     * @param v1 Pointer to the first feature vector.
+     * @param v2 Pointer to the second feature vector.
+     * @param size Size of the feature vectors.
+     * @param res Output parameter to store the cosine similarity result.
+     * @param normalize Whether to normalize the vectors before computing similarity.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    static int32_t CosineSimilarity(const float* v1, const float* v2, int32_t size, float& res, bool normalize = true);
+
+    // Getter methods
+
+    /**
+     * @brief Gets the cache used for search operations in face feature data.
+     * @return A const reference to the Embedded object containing face feature data for search.
+     */
+    const Embedded& GetSearchFaceFeatureCache() const;
+
+    /**
+     * @brief Gets the cache of face feature pointers.
+     * @return A shared pointer to the cache of face feature pointers.
+     */
+    const std::shared_ptr<FaceFeaturePtr>& GetFaceFeaturePtrCache() const;
+
+    /**
+     * @brief Retrieves the total number of facial features stored.
+     * @return int32_t Total number of facial features.
+     */
+    int32_t GetFaceFeatureCount();
+
+    /**
+     * @brief Retrieves the confidence scores for the top k facial features.
+     * @return A reference to the vector of confidence scores.
+     */
+    std::vector<float>& GetTopKConfidence();
+
+    /**
+     * @brief Retrieves the custom IDs for the top k facial features.
+     * @return A reference to the vector of custom IDs.
+     */
+    std::vector<int64_t>& GetTopKCustomIdsCache();
+
+    /**
+     * @brief Retrieves the existing ids in the database.
+     * @return A reference to the vector of existing ids.
+     */
+    std::vector<int64_t>& GetExistingIds();
+
+private:
+    class Impl;
+
+    std::unique_ptr<Impl> pImpl;
+
+    static std::mutex mutex_;
+    static std::shared_ptr<FeatureHubDB> instance_;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FEATURE_HUB_DB_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/frame_process.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/frame_process.h
new file mode 100644
index 000000000..ed9d8f345
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/frame_process.h
@@ -0,0 +1,255 @@
+#ifndef INSPIREFACE_FRAME_PROCESS_H
+#define INSPIREFACE_FRAME_PROCESS_H
+
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include "data_type.h"
+
+namespace inspirecv {
+
+/**
+ * @brief Enum to represent rotation modes.
+ */
+enum ROTATION_MODE { ROTATION_0 = 0, ROTATION_90 = 1, ROTATION_180 = 2, ROTATION_270 = 3 };
+
+/**
+ * @brief Enum to represent data formats.
+ */
+enum DATA_FORMAT { NV21 = 0, NV12 = 1, RGBA = 2, RGB = 3, BGR = 4, BGRA = 5 , I420 = 6, GRAY = 7};
+
+/**
+ * @brief A class to handle camera stream and image processing.
+ * 
+ * This class provides image processing capabilities including rotation, scaling, and format conversion.
+ * It supports four rotation modes with corresponding transformation matrices.
+ * 
+ * Rotation Transformation Matrix Diagrams:
+ * 
+ * 1. ROTATION_0 (No Rotation):
+ *    Original Image (w x h):               Scaled Image (w*s x h*s):
+ *    A(0,0) ----------- B(w-1,0)          A(0,0) ----------- B(w*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Scaled             |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        C(0,h*s-1) ---- D(w*s-1,h*s-1)
+ *    Point Mapping: A->A(0,0), B->B(w*s-1,0), C->C(0,h*s-1), D->D(w*s-1,h*s-1)
+ * 
+ * 2. ROTATION_90 (90° Counter-Clockwise):
+ *    Original Image (w x h):               Rotated Image (h*s x w*s):
+ *    A(0,0) ----------- B(w-1,0)          B(0,0) ----------- A(h*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        D(0,w*s-1) ---- C(h*s-1,w*s-1)
+ *    Point Mapping: A->A(h*s-1,0), B->B(0,0), C->C(h*s-1,w*s-1), D->D(0,w*s-1)
+ * 
+ * 3. ROTATION_180 (180° Rotation):
+ *    Original Image (w x h):               Rotated Image (w*s x h*s):
+ *    A(0,0) ----------- B(w-1,0)          D(0,0) ----------- C(w*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        B(0,h*s-1) ---- A(w*s-1,h*s-1)
+ *    Point Mapping: A->A(w*s-1,h*s-1), B->B(0,h*s-1), C->C(w*s-1,0), D->D(0,0)
+ * 
+ * 4. ROTATION_270 (270° Counter-Clockwise):
+ *    Original Image (w x h):               Rotated Image (h*s x w*s):
+ *    A(0,0) ----------- B(w-1,0)          D(0,0) ----------- C(h*s-1,0)
+ *    |                           |         |                           |
+ *    |        Original           |   =>    |        Rotated            |
+ *    |                           |         |                           |
+ *    C(0,h-1) --------- D(w-1,h-1)        B(0,w*s-1) ---- A(h*s-1,w*s-1)
+ *    Point Mapping: A->A(h*s-1,w*s-1), B->B(0,w*s-1), C->C(h*s-1,0), D->D(0,0)
+ * 
+ * Where:
+ * - w = original image width
+ * - h = original image height  
+ * - s = scale factor
+ * - A, B, C, D represent the four corner points of the image
+ * 
+ * The transformation matrices are computed using setPolyToPoly() with source and destination
+ * point arrays corresponding to the corner mappings shown above.
+ * 
+ * Usage Example:
+ * @code
+ * auto processor = FrameProcess::Create(buffer, height, width, BGR, ROTATION_90);
+ * processor.SetPreviewScale(0.5f);
+ * auto rotated_image = processor.ExecutePreviewImageProcessing(true);
+ * @endcode
+ */
+class INSPIRE_API_EXPORT FrameProcess {
+public:
+    /**
+     * @brief Create a FrameProcess instance.
+     *
+     * @param data_buffer Pointer to the data buffer.
+     * @param height Height of the image.
+     * @param width Width of the image.
+     * @param data_format Data format (e.g., NV21, RGBA).
+     * @param rotation_mode Rotation mode (e.g., ROTATION_0, ROTATION_90).
+     * @return FrameProcess instance.
+     */
+    static FrameProcess Create(const uint8_t* data_buffer, int height, int width, DATA_FORMAT data_format = BGR,
+                               ROTATION_MODE rotation_mode = ROTATION_0);
+
+    /**
+     * @brief Create a FrameProcess instance from an inspirecv::Image.
+     *
+     * @param image The image to process.
+     * @param data_format Data format (e.g., NV21, RGBA).
+     * @param rotation_mode Rotation mode (e.g., ROTATION_0, ROTATION_90).
+     * @return FrameProcess instance.
+     */
+    static FrameProcess Create(const inspirecv::Image& image, DATA_FORMAT data_format = BGR, ROTATION_MODE rotation_mode = ROTATION_0);
+
+    /**
+     * @brief Default constructor.
+     */
+    FrameProcess();
+
+    /**
+     * @brief Destructor.
+     */
+    ~FrameProcess();
+
+    /**
+     * @brief Copy constructor.
+     */
+    FrameProcess(const FrameProcess& other);
+
+    /**
+     * @brief Move constructor.
+     */
+    FrameProcess(FrameProcess&& other) noexcept;
+
+    /**
+     * @brief Copy assignment operator.
+     */
+    FrameProcess& operator=(const FrameProcess& other);
+
+    /**
+     * @brief Move assignment operator.
+     */
+    FrameProcess& operator=(FrameProcess&& other) noexcept;
+
+    /**
+     * @brief Set the data buffer, height, and width of the camera stream.
+     *
+     * @param data_buffer Pointer to the data buffer.
+     * @param height Height of the image.
+     * @param width Width of the image.
+     */
+    void SetDataBuffer(const uint8_t* data_buffer, int height, int width);
+
+    /**
+     * @brief Set the preview size.
+     *
+     * @param size Preview size.
+     */
+    void SetPreviewSize(const int size);
+
+    /**
+     * @brief Set the preview scale.
+     *
+     * @param scale Preview scale.
+     */
+    void SetPreviewScale(const float scale);
+
+    /**
+     * @brief Set the rotation mode.
+     *
+     * @param mode Rotation mode (e.g., ROTATION_0, ROTATION_90).
+     */
+    void SetRotationMode(ROTATION_MODE mode);
+
+    /**
+     * @brief Set the data format.
+     *
+     * @param data_format Data format (e.g., NV21, RGBA).
+     */
+    void SetDataFormat(DATA_FORMAT data_format);
+
+    /**
+     * @brief Set the destination format.
+     *
+     * @param data_format Data format (e.g., NV21, RGBA).
+     */
+    void SetDestFormat(DATA_FORMAT data_format);
+
+    /**
+     * @brief Get an affine-transformed image.
+     *
+     * @param affine_matrix Affine transformation matrix.
+     * @param width_out Width of the output image.
+     * @param height_out Height of the output image.
+     * @return inspirecv::Image Affine-transformed image.
+     */
+    inspirecv::Image ExecuteImageAffineProcessing(inspirecv::TransformMatrix& affine_matrix, const int width_out, const int height_out) const;
+
+    /**
+     * @brief Get a preview image with optional rotation.
+     *
+     * @param with_rotation True if rotation is applied, false otherwise.
+     * @return inspirecv::Image Preview image.
+     */
+    inspirecv::Image ExecutePreviewImageProcessing(bool with_rotation);
+
+    /**
+     * @brief Get the preview scale.
+     *
+     * @return float Preview scale.
+     */
+    float GetPreviewScale();
+
+    /**
+     * @brief Execute image scale processing.
+     *
+     * @param scale Scale factor.
+     * @param with_rotation True if rotation is applied, false otherwise.
+     * @return inspirecv::Image Scaled image.
+     */
+    inspirecv::Image ExecuteImageScaleProcessing(const float scale, bool with_rotation);
+
+    /**
+     * @brief Get the affine transformation matrix.
+     *
+     * @return inspirecv::TransformMatrix Affine transformation matrix.
+     */
+    inspirecv::TransformMatrix GetAffineMatrix() const;
+
+    /**
+     * @brief Get the rotation mode affine transformation matrix, scale coefficient is not included.
+     *
+     * @return inspirecv::TransformMatrix Rotation mode affine transformation matrix.
+     */
+    inspirecv::TransformMatrix GetRotationModeAffineMatrix() const;
+
+    /**
+     * @brief Get the height of the camera stream image.
+     *
+     * @return int Height.
+     */
+    int GetHeight() const;
+
+    /**
+     * @brief Get the width of the camera stream image.
+     *
+     * @return int Width.
+     */
+    int GetWidth() const;
+
+    /**
+     * @brief Get the current rotation mode.
+     *
+     * @return ROTATION_MODE Current rotation mode.
+     */
+    ROTATION_MODE getRotationMode() const;
+
+private:
+    class Impl;
+    std::unique_ptr<Impl> pImpl;
+};
+
+}  // namespace inspirecv
+
+#endif  // INSPIREFACE_FRAME_PROCESS_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/herror.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/herror.h
new file mode 100644
index 000000000..b087dd1d3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/herror.h
@@ -0,0 +1,75 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIRE_FACE_HERROR_H
+#define INSPIRE_FACE_HERROR_H
+
+// [Anchor-Begin]
+
+#define HSUCCEED (0)  // Success
+
+// Basic error types (1-99)
+#define HERR_BASIC_BASE 0x0001                                   // Basic error types
+#define HERR_UNKNOWN HERR_BASIC_BASE                             // Unknown error (1)
+#define HERR_INVALID_PARAM (HERR_BASIC_BASE + 1)                 // Invalid parameter (2)
+#define HERR_INVALID_IMAGE_STREAM_HANDLE (HERR_BASIC_BASE + 2)   // Invalid image stream handle (3)
+#define HERR_INVALID_CONTEXT_HANDLE (HERR_BASIC_BASE + 3)        // Invalid context handle (4)
+#define HERR_INVALID_FACE_TOKEN (HERR_BASIC_BASE + 4)            // Invalid face token (5)
+#define HERR_INVALID_FACE_FEATURE (HERR_BASIC_BASE + 5)          // Invalid face feature (6)
+#define HERR_INVALID_FACE_LIST (HERR_BASIC_BASE + 6)             // Invalid face feature list (7)
+#define HERR_INVALID_BUFFER_SIZE (HERR_BASIC_BASE + 7)           // Invalid copy token (8)
+#define HERR_INVALID_IMAGE_STREAM_PARAM (HERR_BASIC_BASE + 8)    // Invalid image param (9)
+#define HERR_INVALID_SERIALIZATION_FAILED (HERR_BASIC_BASE + 9)  // Invalid face serialization failed (10)
+#define HERR_INVALID_DETECTION_INPUT (HERR_BASIC_BASE + 10)      // Failed to modify detector input size (11)
+#define HERR_INVALID_IMAGE_BITMAP_HANDLE (HERR_BASIC_BASE + 11)  // Invalid image bitmap handle (12)
+#define HERR_IMAGE_STREAM_DECODE_FAILED (HERR_BASIC_BASE + 12)  // ImageStream failed to decode the image (13)
+
+// Session error types (100-199)
+#define HERR_SESS_BASE 0x0064                                   // Session error types (100)
+#define HERR_SESS_FUNCTION_UNUSABLE (HERR_SESS_BASE + 1)        // Function not usable (101)
+#define HERR_SESS_TRACKER_FAILURE (HERR_SESS_BASE + 2)          // Tracker module not initialized (102)
+#define HERR_SESS_PIPELINE_FAILURE (HERR_SESS_BASE + 3)         // Pipeline module not initialized (103)
+#define HERR_SESS_INVALID_RESOURCE (HERR_SESS_BASE + 4)         // Invalid static resource (104)
+#define HERR_SESS_LANDMARK_NUM_NOT_MATCH (HERR_SESS_BASE + 5)   // The number of input landmark points does not match (105)
+#define HERR_SESS_LANDMARK_NOT_ENABLE (HERR_SESS_BASE + 6)      // The landmark model is not enabled (106)
+#define HERR_SESS_KEY_POINT_NUM_NOT_MATCH (HERR_SESS_BASE + 7)  // The number of input key points does not match (107)
+#define HERR_SESS_REC_EXTRACT_FAILURE (HERR_SESS_BASE + 8)      // Face feature extraction not registered (108)
+#define HERR_SESS_REC_CONTRAST_FEAT_ERR (HERR_SESS_BASE + 9)    // Incorrect length of feature vector for comparison (109)
+#define HERR_SESS_FACE_DATA_ERROR (HERR_SESS_BASE + 10)         // Face data parsing (110)
+#define HERR_SESS_FACE_REC_OPTION_ERROR (HERR_SESS_BASE + 11)   // An optional parameter is incorrect (111)
+
+// FeatureHub error types (200-249)
+#define HERR_FT_HUB_BASE 0x00C8                               // FeatureHub error types (200)
+#define HERR_FT_HUB_DISABLE (HERR_FT_HUB_BASE + 1)            // FeatureHub is disabled (201)
+#define HERR_FT_HUB_INSERT_FAILURE (HERR_FT_HUB_BASE + 2)     // Data insertion error (202)
+#define HERR_FT_HUB_NOT_FOUND_FEATURE (HERR_FT_HUB_BASE + 3)  // Get face feature error (203)
+
+// Archive error types (250-299)
+#define HERR_ARCHIVE_BASE 0x00FA                                 // Archive error types (250)
+#define HERR_ARCHIVE_LOAD_FAILURE (HERR_ARCHIVE_BASE + 1)        // Archive load failure (251)
+#define HERR_ARCHIVE_LOAD_MODEL_FAILURE (HERR_ARCHIVE_BASE + 2)  // Model load failure (252)
+#define HERR_ARCHIVE_FILE_FORMAT_ERROR (HERR_ARCHIVE_BASE + 3)   // The archive format is incorrect (253)
+#define HERR_ARCHIVE_REPETITION_LOAD (HERR_ARCHIVE_BASE + 4)     // Do not reload the model (254)
+#define HERR_ARCHIVE_NOT_LOAD (HERR_ARCHIVE_BASE + 5)            // Model not loaded (255)
+
+// Device/Hardware error types (300-349)
+#define HERR_DEVICE_BASE 0x012C                                       // Hardware error types (300)
+#define HERR_DEVICE_CUDA_NOT_SUPPORT (HERR_DEVICE_BASE + 1)           // CUDA not supported (301)
+#define HERR_DEVICE_CUDA_TENSORRT_NOT_SUPPORT (HERR_DEVICE_BASE + 2)  // CUDA TensorRT not supported (302)
+#define HERR_DEVICE_CUDA_UNKNOWN_ERROR (HERR_DEVICE_BASE + 3)         // CUDA unknown error (303)
+#define HERR_DEVICE_CUDA_DISABLE (HERR_DEVICE_BASE + 4)               // CUDA support is disabled (304)
+
+// Extension module error types (350-549)
+#define HERR_EXTENSION_BASE 0x015E                                             // Extension module error types (350)
+#define HERR_EXTENSION_ERROR (HERR_EXTENSION_BASE + 1)                         // Extension module error (351)
+#define HERR_EXTENSION_MLMODEL_LOAD_FAILED (HERR_EXTENSION_BASE + 2)           // MLModel load failed (352)
+#define HERR_EXTENSION_HETERO_MODEL_TAG_ERROR (HERR_EXTENSION_BASE + 3)        // Incorrect heterogeneous model tag (353)
+#define HERR_EXTENSION_HETERO_REC_HEAD_CONFIG_ERROR (HERR_EXTENSION_BASE + 4)  // Rec head config error (354)
+#define HERR_EXTENSION_HETERO_MODEL_NOT_MATCH (HERR_EXTENSION_BASE + 5)        // Heterogeneous model dimensions do not match (355)
+#define HERR_EXTENSION_HETERO_MODEL_NOT_LOADED (HERR_EXTENSION_BASE + 6)       // Heterogeneous model dimensions not loaded (356)
+
+// [Anchor-End]
+
+#endif  // INSPIRE_FACE_HERROR_H
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/information.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/information.h
new file mode 100644
index 000000000..3e1f540ac
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/information.h
@@ -0,0 +1,30 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include "data_type.h"
+
+#ifndef INSPIRE_FACE_INFORMATION_H
+#define INSPIRE_FACE_INFORMATION_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// C-style function declarations, using API export macro
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionMajorStr();
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionMinorStr();
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionPatchStr();
+INSPIRE_API_EXPORT const char* GetInspireFaceExtendedInformation();
+
+#ifdef __cplusplus
+}
+#endif
+
+// C-style macro definitions
+#define INSPIRE_FACE_VERSION_MAJOR_STR GetInspireFaceVersionMajorStr()
+#define INSPIRE_FACE_VERSION_MINOR_STR GetInspireFaceVersionMinorStr()
+#define INSPIRE_FACE_VERSION_PATCH_STR GetInspireFaceVersionPatchStr()
+#define INSPIRE_FACE_EXTENDED_INFORMATION GetInspireFaceExtendedInformation()
+
+#endif  // INSPIRE_FACE_INFORMATION_H
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/inspireface.hpp b/cpp-package/inspireface/cpp/inspireface/include/inspireface/inspireface.hpp
new file mode 100644
index 000000000..5c368a644
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/inspireface.hpp
@@ -0,0 +1,14 @@
+#include "session.h"
+#include "cuda_toolkit.h"
+#include "data_type.h"
+#include "log.h"
+#include "herror.h"
+#include "feature_hub_db.h"
+#include "frame_process.h"
+#include "isf_check.h"
+#include "launch.h"
+#include "log.h"
+#include "similarity_converter.h"
+#include "spend_timer.h"
+#include "information.h"
+#include "face_wrapper.h"
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/isf_check.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/isf_check.h
new file mode 100644
index 000000000..44692d176
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/isf_check.h
@@ -0,0 +1,33 @@
+#ifndef INSPIRE_FACE_CHECK_H
+#define INSPIRE_FACE_CHECK_H
+#include "log.h"
+#include "herror.h"
+
+#define INSPIREFACE_RETURN_IF_ERROR(...)             \
+    do {                                             \
+        const int32_t _status = (__VA_ARGS__);       \
+        if (_status != HSUCCEED) {                   \
+            INSPIRE_LOGE("Error code: %d", _status); \
+            return _status;                          \
+        }                                            \
+    } while (0)
+
+#define INSPIREFACE_LOG_IF(severity, condition) \
+    if (condition)                              \
+    INSPIRE_LOG##severity
+
+#define INSPIREFACE_CHECK(condition)                        \
+    do {                                                    \
+        if (!(condition)) {                                 \
+            INSPIRE_LOGF("Check failed: (%s)", #condition); \
+        }                                                   \
+    } while (0)
+
+#define INSPIREFACE_CHECK_MSG(condition, message)                       \
+    do {                                                                \
+        if (!(condition)) {                                             \
+            INSPIRE_LOGF("Check failed: (%s) %s", #condition, message); \
+        }                                                               \
+    } while (0)
+
+#endif  // INSPIRE_FACE_CHECK_H
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/launch.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/launch.h
new file mode 100644
index 000000000..b5573ff95
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/launch.h
@@ -0,0 +1,137 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_LAUNCH_H
+#define INSPIREFACE_LAUNCH_H
+
+#include <memory>
+#include <string>
+#include <cstdint>
+#include "data_type.h"
+
+#define INSPIREFACE_CONTEXT inspire::Launch::GetInstance()
+
+namespace inspire {
+
+// Forward declarations
+class InspireArchive;
+
+// The Launch class acts as the main entry point for the InspireFace system.
+// It is responsible for loading static resources such as models, configurations, and parameters.
+class INSPIRE_API_EXPORT Launch {
+public:
+    // Special Backend enum for CoreML
+    enum NNInferenceBackend {
+        NN_INFERENCE_CPU = 0,
+        NN_INFERENCE_MMM_CUDA,
+        NN_INFERENCE_COREML_CPU,
+        NN_INFERENCE_COREML_GPU,
+        NN_INFERENCE_COREML_ANE,
+        NN_INFERENCE_TENSORRT_CUDA,
+    };
+
+    // Landmark engine enum
+    enum LandmarkEngine {
+        LANDMARK_HYPLMV2_0_25 = 0,
+        LANDMARK_HYPLMV2_0_50,
+        LANDMARK_INSIGHTFACE_2D106_TRACK,
+    };
+
+    // Image processing backend engine
+    enum ImageProcessingBackend {
+        IMAGE_PROCESSING_CPU = 0,  // CPU backend(Default)
+        IMAGE_PROCESSING_RGA,      // Rockchip RGA backend(Hardware support is mandatory)
+    };
+
+    Launch(const Launch&) = delete;             // Delete the copy constructor to prevent copying.
+    Launch& operator=(const Launch&) = delete;  // Delete the assignment operator to prevent assignment.
+    ~Launch();                                  // Destructor needs to be defined where the implementation is complete
+
+    // Retrieves the singleton instance of Launch, ensuring that only one instance exists.
+    static std::shared_ptr<Launch> GetInstance();
+
+    // Loads the necessary resources from a specified path.
+    // Returns an integer status code: 0 on success, non-zero on failure.
+    int32_t Load(const std::string& path);
+
+    // Reloads the resources from a specified path.
+    // Returns an integer status code: 0 on success, non-zero on failure.
+    int32_t Reload(const std::string& path);
+
+    // Provides access to the loaded InspireArchive instance.
+    InspireArchive& getMArchive();
+
+    // Checks if the resources have been successfully loaded.
+    bool isMLoad() const;
+
+    // Unloads the resources and resets the system to its initial state.
+    void Unload();
+
+    // Set the rockchip dma heap path
+    void SetRockchipDmaHeapPath(const std::string& path);
+
+    // Get the rockchip dma heap path
+    std::string GetRockchipDmaHeapPath() const;
+
+    // Set the extension path
+    void ConfigurationExtensionPath(const std::string& path);
+
+    // Get the extension path
+    std::string GetExtensionPath() const;
+
+    // Set the global coreml inference mode
+    void SetGlobalCoreMLInferenceMode(NNInferenceBackend mode);
+
+    // Get the global coreml inference mode
+    NNInferenceBackend GetGlobalCoreMLInferenceMode() const;
+
+    // Build the extension path
+    void BuildAppleExtensionPath(const std::string& resource_path);
+
+    // Set the cuda device id
+    void SetCudaDeviceId(int32_t device_id);
+
+    // Get the cuda device id
+    int32_t GetCudaDeviceId() const;
+
+    // Set the face detect pixel list
+    void SetFaceDetectPixelList(const std::vector<int32_t>& pixel_list);
+
+    // Get the face detect pixel list
+    std::vector<int32_t> GetFaceDetectPixelList() const;
+
+    // Set the face detect model list
+    void SetFaceDetectModelList(const std::vector<std::string>& model_list);
+
+    // Get the face detect model list
+    std::vector<std::string> GetFaceDetectModelList() const;
+
+    // Switch the landmark engine(It must be used before creating a session)
+    void SwitchLandmarkEngine(LandmarkEngine engine);
+
+    // Switch the image processing backend(It must be used before creating a session)
+    void SwitchImageProcessingBackend(ImageProcessingBackend backend);
+
+    // Get the image processing backend
+    ImageProcessingBackend GetImageProcessingBackend() const;
+
+    // Set the ImageProcess Aligned Width(It must be used before creating a session)
+    void SetImageProcessAlignedWidth(int32_t width);
+
+    // Get the ImageProcess Aligned Width
+    int32_t GetImageProcessAlignedWidth() const;
+
+private:
+    // Private constructor for the singleton pattern
+    Launch();
+
+    // Private implementation class
+    class Impl;
+    std::unique_ptr<Impl> pImpl;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_LAUNCH_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/log.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/log.h
new file mode 100755
index 000000000..1dcd4451a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/log.h
@@ -0,0 +1,89 @@
+#ifndef INSPIRE_FACE_LOG_H
+#define INSPIRE_FACE_LOG_H
+
+#include <memory>
+#include <string>
+#include <cstring>
+#include "data_type.h"
+
+// Macro to extract the filename from the full path
+#define __FILENAME__ (strrchr(__FILE__, '/') ? strrchr(__FILE__, '/') + 1 : __FILE__)
+
+#ifdef ANDROID
+// Android platform log macros
+#define INSPIRE_ANDROID_LOG_TAG "InspireFace"
+#define INSPIRE_LOGD(...) inspire::LogManager::getInstance()->logAndroid(inspire::LogLevel::ISF_LOG_DEBUG, INSPIRE_ANDROID_LOG_TAG, __VA_ARGS__)
+#define INSPIRE_LOGI(...) inspire::LogManager::getInstance()->logAndroid(inspire::LogLevel::ISF_LOG_INFO, INSPIRE_ANDROID_LOG_TAG, __VA_ARGS__)
+#define INSPIRE_LOGW(...) inspire::LogManager::getInstance()->logAndroid(inspire::LogLevel::ISF_LOG_WARN, INSPIRE_ANDROID_LOG_TAG, __VA_ARGS__)
+#define INSPIRE_LOGE(...) inspire::LogManager::getInstance()->logAndroid(inspire::LogLevel::ISF_LOG_ERROR, INSPIRE_ANDROID_LOG_TAG, __VA_ARGS__)
+#define INSPIRE_LOGF(...) inspire::LogManager::getInstance()->logAndroid(inspire::LogLevel::ISF_LOG_FATAL, INSPIRE_ANDROID_LOG_TAG, __VA_ARGS__)
+#else
+// Standard platform log macros
+#define INSPIRE_LOGD(...) \
+    inspire::LogManager::getInstance()->logStandard(inspire::LogLevel::ISF_LOG_DEBUG, __FILENAME__, __FUNCTION__, __LINE__, __VA_ARGS__)
+#define INSPIRE_LOGI(...) inspire::LogManager::getInstance()->logStandard(inspire::LogLevel::ISF_LOG_INFO, "", "", -1, __VA_ARGS__)
+#define INSPIRE_LOGW(...) inspire::LogManager::getInstance()->logStandard(inspire::LogLevel::ISF_LOG_WARN, "", "", -1, __VA_ARGS__)
+#define INSPIRE_LOGE(...) inspire::LogManager::getInstance()->logStandard(inspire::LogLevel::ISF_LOG_ERROR, "", "", -1, __VA_ARGS__)
+#define INSPIRE_LOGF(...) inspire::LogManager::getInstance()->logStandard(inspire::LogLevel::ISF_LOG_FATAL, "", "", -1, __VA_ARGS__)
+#endif
+
+// Macro to set the global log level
+#define INSPIRE_SET_LOG_LEVEL(level) inspire::LogManager::getInstance()->setLogLevel(level)
+
+namespace inspire {
+
+// Log levels
+enum LogLevel { ISF_LOG_NONE = 0, ISF_LOG_DEBUG, ISF_LOG_INFO, ISF_LOG_WARN, ISF_LOG_ERROR, ISF_LOG_FATAL };
+
+/**
+ * @class LogManager
+ * @brief A singleton class for logging messages to the console or Android logcat.
+ *
+ * This class provides methods to log messages of different severity levels (DEBUG, INFO, WARN, ERROR, FATAL)
+ * to the console or Android logcat based on the current log level setting.
+ *
+ * Implementation details are hidden using the PIMPL (Pointer to Implementation) pattern.
+ */
+class INSPIRE_API_EXPORT LogManager {
+public:
+    // Get the singleton instance
+    static LogManager* getInstance();
+
+    // Destructor
+    ~LogManager();
+
+    // Set the log level
+    void setLogLevel(LogLevel level);
+
+    // Get the current log level
+    LogLevel getLogLevel() const;
+
+#ifdef ANDROID
+    // Method for logging on the Android platform
+    void logAndroid(LogLevel level, const char* tag, const char* format, ...) const;
+#else
+    // Method for standard platform logging
+    void logStandard(LogLevel level, const char* filename, const char* function, int line, const char* format, ...) const;
+#endif
+
+private:
+    // Private constructor for singleton pattern
+    LogManager();
+
+    // Disable copy construction and assignment
+    LogManager(const LogManager&) = delete;
+    LogManager& operator=(const LogManager&) = delete;
+
+    // Forward declaration of the implementation class
+    class Impl;
+
+    // Pointer to implementation
+    std::unique_ptr<Impl> pImpl;
+
+    // Static instance for singleton pattern
+    static LogManager* instance;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_LOG_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/meta.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/meta.h
new file mode 100644
index 000000000..b10898260
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/meta.h
@@ -0,0 +1,54 @@
+#ifndef INSPIRE_FACE_META_H
+#define INSPIRE_FACE_META_H
+
+#include <iostream>
+#include <string>
+#include "data_type.h"
+
+namespace inspire {
+
+/**
+ * @brief SDK meta information
+ */
+struct INSPIRE_API_EXPORT SDKInfo {
+    // version
+    int version_major;
+    int version_minor;
+    int version_patch;
+
+    // series name
+    std::string series;
+
+    // build info
+    std::string build_date;
+    std::string build_time;
+    std::string compiler;
+    std::string platform;
+    std::string architecture;
+
+    // backend info
+    std::string inference_backend;
+    std::string inspirecv_backend;
+    bool rga_backend_enabled;
+
+    // description
+    std::string description;
+    
+    std::string GetFullVersionInfo() const;
+    
+    // auxiliary methods: return the string form of the version number
+    std::string GetVersionMajorStr() const;
+    std::string GetVersionMinorStr() const;
+    std::string GetVersionPatchStr() const;
+    std::string GetVersionString() const;
+};
+
+/**
+ * @brief Get the SDK info
+ * @return The constant reference of the SDK info
+ */
+INSPIRE_API_EXPORT const SDKInfo& GetSDKInfo();
+
+}   // namespace inspire
+
+#endif  // INSPIRE_FACE_META_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/session.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/session.h
new file mode 100644
index 000000000..5dec7d0d7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/session.h
@@ -0,0 +1,222 @@
+#ifndef INSPIRE_FACE_SESSION_H
+#define INSPIRE_FACE_SESSION_H
+#include <memory>
+#include "data_type.h"
+#include "frame_process.h"
+#include "face_wrapper.h"
+
+namespace inspire {
+
+/**
+ * @brief The face algorithm session class.
+ */
+class INSPIRE_API_EXPORT Session {
+public:
+    Session();
+    ~Session();
+
+    Session(Session&&) noexcept;
+    Session& operator=(Session&&) noexcept;
+
+    Session(const Session&) = delete;
+    Session& operator=(const Session&) = delete;
+
+    /**
+     * @brief Create a new session with the given parameters.
+     * @param detect_mode The mode of face detection.
+     * @param max_detect_face The maximum number of faces to detect.
+     * @param param The custom pipeline parameter.
+     * @param detect_level_px The detection level in pixels.
+     * @param track_by_detect_mode_fps The tracking frame rate.
+     * @return A new session.
+     */
+    static Session Create(DetectModuleMode detect_mode, int32_t max_detect_face, const CustomPipelineParameter& param, int32_t detect_level_px = -1,
+                          int32_t track_by_detect_mode_fps = -1);
+
+    /**
+     * @brief Create a new session pointer with the given parameters.
+     * @param detect_mode The mode of face detection.
+     * @param max_detect_face The maximum number of faces to detect.
+     * @param param The custom pipeline parameter.
+     * @param detect_level_px The detection level in pixels.
+     * @param track_by_detect_mode_fps The tracking frame rate.
+     * @return A raw pointer to new session. The caller is responsible for memory management.
+     */
+    static Session* CreatePtr(DetectModuleMode detect_mode, int32_t max_detect_face, const CustomPipelineParameter& param,
+                              int32_t detect_level_px = -1, int32_t track_by_detect_mode_fps = -1) {
+        return new Session(Create(detect_mode, max_detect_face, param, detect_level_px, track_by_detect_mode_fps));
+    }
+
+    /**
+     * @brief Clear the tracking face
+     */
+    void ClearTrackingFace();
+
+    /**
+     * @brief Set the track lost recovery mode(only for LightTrack mode, default is false)
+     * @param value The track lost recovery mode value
+     */
+    void SetTrackLostRecoveryMode(bool value);
+
+    /**
+     * @brief Set the light track confidence threshold
+     * @param value Light track confidence threshold
+     */
+    void SetLightTrackConfidenceThreshold(float value);
+
+    /**
+     * @brief Set the track preview size.
+     * @param preview_size The preview size.
+     */
+    void SetTrackPreviewSize(int32_t preview_size);
+
+    /**
+     * @brief Set the minimum face pixel size.
+     * @param min_face_pixel_size The minimum face pixel size.
+     */
+    void SetFilterMinimumFacePixelSize(int32_t min_face_pixel_size);
+
+    /**
+     * @brief Set the face detect threshold.
+     * @param threshold The face detect threshold.
+     */
+    void SetFaceDetectThreshold(float threshold);
+
+    /**
+     * @brief Set the track mode smooth ratio.
+     * @param smooth_ratio The track mode smooth ratio.
+     */
+    void SetTrackModeSmoothRatio(int32_t smooth_ratio);
+
+    /**
+     * @brief Set the track mode num smooth cache frame.
+     * @param num_smooth_cache_frame The track mode num smooth cache frame.
+     */
+    void SetTrackModeNumSmoothCacheFrame(int32_t num_smooth_cache_frame);
+
+    /**
+     * @brief Set the track mode detect interval.
+     * @param detect_interval The track mode detect interval.
+     */
+    void SetTrackModeDetectInterval(int32_t detect_interval);
+
+    /**
+     * @brief Detect and track the faces in the frame.
+     * @param process The frame process.
+     * @param results The detected faces.
+     */
+    int32_t FaceDetectAndTrack(inspirecv::FrameProcess& process, std::vector<FaceTrackWrap>& results);
+
+    /**
+     * @brief Get the face bounding box.
+     * @param face_data The face data.
+     * @return The face bounding box.
+     */
+    inspirecv::Rect2i GetFaceBoundingBox(const FaceTrackWrap& face_data);
+
+    /**
+     * @brief Get the face dense landmark.
+     * @param face_data The face data.
+     * @return The face dense landmark.
+     */
+    std::vector<inspirecv::Point2f> GetFaceDenseLandmark(const FaceTrackWrap& face_data);
+
+    /**
+     * @brief Get the face five key points.
+     * @param face_data The face data.
+     * @return The face five key points.
+     */
+    std::vector<inspirecv::Point2f> GetFaceFiveKeyPoints(const FaceTrackWrap& face_data);
+
+    /**
+     * @brief Extract the face feature.
+     * @param process The frame process.
+     * @param data The face data.
+     * @param embedding The face embedding.
+     * @param normalize The normalize flag.
+     */
+    int32_t FaceFeatureExtract(inspirecv::FrameProcess& process, FaceTrackWrap& data, FaceEmbedding& embedding, bool normalize = true);
+
+    /**
+     * @brief Get the face alignment image.
+     * @param process The frame process.
+     * @param data The face data.
+     * @param wrapped The wrapped image.
+     */
+    void GetFaceAlignmentImage(inspirecv::FrameProcess& process, FaceTrackWrap& data, inspirecv::Image& wrapped);
+
+    /**
+     * @brief Extract the face feature with alignment image.
+     * @param process The frame process.
+     * @param embedding The face embedding.
+     * @param normalize The normalize flag.
+     */
+    int32_t FaceFeatureExtractWithAlignmentImage(inspirecv::FrameProcess& process, FaceEmbedding& embedding, bool normalize = true);
+
+    /**
+     * @brief Extract the face feature with alignment image.
+     * @param wrapped The wrapped image.
+     * @param embedding The face embedding.
+     * @param normalize The normalize flag.
+     */
+    int32_t FaceFeatureExtractWithAlignmentImage(const inspirecv::Image& wrapped, FaceEmbedding& embedding, bool normalize = true);
+
+    /**
+     * @brief Multiple face pipeline process.
+     * @param process The frame process.
+     * @param param The custom pipeline parameter.
+     * @param face_data_list The face data list.
+     */
+    int32_t MultipleFacePipelineProcess(inspirecv::FrameProcess& process, const CustomPipelineParameter& param,
+                                        const std::vector<FaceTrackWrap>& face_data_list);
+
+    /**
+     * @brief Get the RGB liveness confidence.
+     * @return The RGB liveness confidence.
+     */
+    std::vector<float> GetRGBLivenessConfidence();
+
+    /**
+     * @brief Get the face mask confidence.
+     * @return The face mask confidence.
+     */
+    std::vector<float> GetFaceMaskConfidence();
+
+    /**
+     * @brief Get the face quality confidence.
+     * @return The face quality confidence.
+     */
+    std::vector<float> GetFaceQualityConfidence();
+
+    /**
+     * @brief Get the face interaction state.
+     * @return The face interaction state.
+     */
+    std::vector<FaceInteractionState> GetFaceInteractionState();
+
+    /**
+     * @brief Get the face interaction action.
+     * @return The face interaction action.
+     */
+    std::vector<FaceInteractionAction> GetFaceInteractionAction();
+
+    /**
+     * @brief Get the face attribute result.
+     * @return The face attribute result.
+     */
+    std::vector<FaceAttributeResult> GetFaceAttributeResult();
+
+    /**
+     * @brief Get the face emotion result.
+     * @return The face emotion result.
+     */
+    std::vector<FaceEmotionResult> GetFaceEmotionResult();
+
+private:
+    class Impl;
+    std::unique_ptr<Impl> pImpl;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_SESSION_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/similarity_converter.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/similarity_converter.h
new file mode 100644
index 000000000..f6dbae965
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/similarity_converter.h
@@ -0,0 +1,113 @@
+#ifndef SIMILARITY_CONVERTER_H
+#define SIMILARITY_CONVERTER_H
+
+#include <iostream>
+#include <cmath>
+#include <mutex>
+#include "data_type.h"
+
+#define SIMILARITY_CONVERTER_UPDATE_CONFIG(config) inspire::SimilarityConverter::getInstance().updateConfig(config)
+#define SIMILARITY_CONVERTER_RUN(cosine) inspire::SimilarityConverter::getInstance().convert(cosine)
+#define SIMILARITY_CONVERTER_GET_CONFIG() inspire::SimilarityConverter::getInstance().getConfig()
+#define SIMILARITY_CONVERTER_GET_RECOMMENDED_COSINE_THRESHOLD() inspire::SimilarityConverter::getInstance().getRecommendedCosineThreshold()
+#define SIMILARITY_CONVERTER_SET_RECOMMENDED_COSINE_THRESHOLD(threshold) \
+    inspire::SimilarityConverter::getInstance().setRecommendedCosineThreshold(threshold)
+
+namespace inspire {
+
+struct SimilarityConverterConfig {
+    double threshold = 0.48;   // Similarity threshold (e.g. 0.48 or 0.32)
+    double middleScore = 0.6;  // Target score at threshold (e.g. 0.6)
+    double steepness = 8.0;    // Steepness of the curve
+    double outputMin = 0.01;   // Minimum value of output range
+    double outputMax = 1.0;    // Maximum value of output range
+};
+
+class INSPIRE_API_EXPORT SimilarityConverter {
+private:
+    SimilarityConverterConfig config;
+    double outputScale;              // Scale of output range
+    double bias;                     // Sigmoid bias
+    mutable std::mutex configMutex;  // Mutex for protecting config updates
+
+    // Recommended cosine threshold
+    float recommendedCosineThreshold = 0.48;
+
+    static SimilarityConverter* instance;
+    static std::mutex instanceMutex;
+
+    // Update internal calculation parameters
+    void updateParameters() {
+        outputScale = config.outputMax - config.outputMin;
+        bias = -std::log((config.outputMax - config.middleScore) / (config.middleScore - config.outputMin));
+    }
+
+public:
+    // Get global singleton instance
+    static SimilarityConverter& getInstance() {
+        std::lock_guard<std::mutex> lock(instanceMutex);
+        if (instance == nullptr) {
+            instance = new SimilarityConverter();
+        }
+        return *instance;
+    }
+
+    // Allow external creation of new instances
+    explicit SimilarityConverter(const SimilarityConverterConfig& config = SimilarityConverterConfig()) : config(config) {
+        updateParameters();
+    }
+
+    // Prevent copying
+    SimilarityConverter(const SimilarityConverter&) = delete;
+    SimilarityConverter& operator=(const SimilarityConverter&) = delete;
+
+    // Update configuration (thread-safe)
+    void updateConfig(const SimilarityConverterConfig& newConfig) {
+        std::lock_guard<std::mutex> lock(configMutex);
+        config = newConfig;
+        updateParameters();
+    }
+
+    // Get current configuration (thread-safe)
+    SimilarityConverterConfig getConfig() const {
+        std::lock_guard<std::mutex> lock(configMutex);
+        return config;
+    }
+
+    // Convert similarity (thread-safe)
+    template <typename T>
+    double convert(T cosine) {
+        std::lock_guard<std::mutex> lock(configMutex);
+        // Calculate shifted input
+        double shiftedInput = config.steepness * (static_cast<double>(cosine) - config.threshold);
+        // Apply sigmoid function
+        double sigmoid = 1.0 / (1.0 + std::exp(-shiftedInput - bias));
+        // Map to output range
+        return sigmoid * outputScale + config.outputMin;
+    }
+
+    // Clean up singleton instance
+    static void destroyInstance() {
+        std::lock_guard<std::mutex> lock(instanceMutex);
+        if (instance != nullptr) {
+            delete instance;
+            instance = nullptr;
+        }
+    }
+
+    // Get recommended cosine threshold
+    float getRecommendedCosineThreshold() const {
+        return recommendedCosineThreshold;
+    }
+
+    // Set recommended cosine threshold
+    void setRecommendedCosineThreshold(float threshold) {
+        recommendedCosineThreshold = threshold;
+    }
+
+    ~SimilarityConverter() = default;
+};  // class SimilarityConverter
+
+}  // namespace inspire
+
+#endif  // SIMILARITY_CONVERTER_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/include/inspireface/spend_timer.h b/cpp-package/inspireface/cpp/inspireface/include/inspireface/spend_timer.h
new file mode 100644
index 000000000..cdc274606
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/include/inspireface/spend_timer.h
@@ -0,0 +1,51 @@
+#ifndef INSPIRE_FACE_TIMER_H
+#define INSPIRE_FACE_TIMER_H
+
+#include "data_type.h"
+
+namespace inspire {
+
+// Get the current time in microseconds.
+uint64_t INSPIRE_API_EXPORT _now();
+
+/**
+ * @brief A class to measure the cost of a block of code.
+ */
+class INSPIRE_API_EXPORT SpendTimer {
+public:
+    SpendTimer();
+    explicit SpendTimer(const std::string &name);
+
+    void Start();
+    void Stop();
+    void Reset();
+
+    uint64_t Get() const;
+    uint64_t Average() const;
+    uint64_t Total() const;
+    uint64_t Count() const;
+    uint64_t Min() const;
+    uint64_t Max() const;
+    const std::string &name() const;
+    std::string Report() const;
+
+    static void Disable();
+
+protected:
+    uint64_t start_;
+    uint64_t stop_;
+    uint64_t total_;
+    uint64_t count_;
+    uint64_t min_;
+    uint64_t max_;
+    std::string name_;
+
+    static int is_enable;
+};
+
+INSPIRE_API_EXPORT std::ostream &operator<<(std::ostream &os, const SpendTimer &timer);
+
+#define TIME_NOW inspirecv::_now()
+
+}  // namespace inspire
+#endif  // INSPIRE_FACE_TIMER_H
diff --git a/cpp-package/inspireface/cpp/inspireface/log.cpp b/cpp-package/inspireface/cpp/inspireface/log.cpp
new file mode 100644
index 000000000..4dd0af1a3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/log.cpp
@@ -0,0 +1,152 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include "log.h"
+#include <mutex>
+#include <cstdarg>
+#include <cstring>
+#include <iostream>
+
+#ifdef ANDROID
+#include <android/log.h>
+#endif
+
+namespace inspire {
+
+// Implementation class for LogManager
+class LogManager::Impl {
+public:
+    Impl() : currentLevel(LogLevel::ISF_LOG_INFO) {}
+
+    LogLevel currentLevel;
+    static std::mutex mutex;
+};
+
+// Static initialization
+std::mutex LogManager::Impl::mutex;
+LogManager* LogManager::instance = nullptr;
+
+// Constructor
+LogManager::LogManager() : pImpl(std::make_unique<Impl>()) {}
+
+// Destructor
+LogManager::~LogManager() = default;
+
+// Get singleton instance
+LogManager* LogManager::getInstance() {
+    std::lock_guard<std::mutex> lock(Impl::mutex);
+    if (instance == nullptr) {
+        instance = new LogManager();
+    }
+    return instance;
+}
+
+// Set log level
+void LogManager::setLogLevel(LogLevel level) {
+    pImpl->currentLevel = level;
+}
+
+// Get log level
+LogLevel LogManager::getLogLevel() const {
+    return pImpl->currentLevel;
+}
+
+#ifdef ANDROID
+// Android logging implementation
+void LogManager::logAndroid(LogLevel level, const char* tag, const char* format, ...) const {
+    if (pImpl->currentLevel == LogLevel::ISF_LOG_NONE || level < pImpl->currentLevel)
+        return;
+
+    int androidLevel;
+    switch (level) {
+        case LogLevel::ISF_LOG_DEBUG:
+            androidLevel = ANDROID_LOG_DEBUG;
+            break;
+        case LogLevel::ISF_LOG_INFO:
+            androidLevel = ANDROID_LOG_INFO;
+            break;
+        case LogLevel::ISF_LOG_WARN:
+            androidLevel = ANDROID_LOG_WARN;
+            break;
+        case LogLevel::ISF_LOG_ERROR:
+            androidLevel = ANDROID_LOG_ERROR;
+            break;
+        case LogLevel::ISF_LOG_FATAL:
+            androidLevel = ANDROID_LOG_FATAL;
+            break;
+        default:
+            androidLevel = ANDROID_LOG_DEFAULT;
+    }
+
+    va_list args;
+    va_start(args, format);
+    __android_log_vprint(androidLevel, tag, format, args);
+    va_end(args);
+
+    // If the log level is fatal, flush the error stream and abort the program
+    if (level == LogLevel::ISF_LOG_FATAL) {
+        std::flush(std::cerr);
+        abort();
+    }
+}
+#else
+// Standard logging implementation
+void LogManager::logStandard(LogLevel level, const char* filename, const char* function, int line, const char* format, ...) const {
+    // Check whether the current level is LOG NONE or the log level is not enough to log
+    if (pImpl->currentLevel == LogLevel::ISF_LOG_NONE || level < pImpl->currentLevel)
+        return;
+
+    // Build log prefix dynamically based on available data
+    bool hasPrintedPrefix = false;
+    if (filename && strlen(filename) > 0) {
+        printf("[%s]", filename);
+        hasPrintedPrefix = true;
+    }
+    if (function && strlen(function) > 0) {
+        printf("[%s]", function);
+        hasPrintedPrefix = true;
+    }
+    if (line != -1) {
+        printf("[%d]", line);
+        hasPrintedPrefix = true;
+    }
+
+    // Only add colon and space if any prefix was printed
+    if (hasPrintedPrefix) {
+        printf(": ");
+    }
+
+    // Set text color for different log levels, but only if not on iOS
+#ifndef TARGET_OS_IOS
+    if (level == LogLevel::ISF_LOG_ERROR || level == LogLevel::ISF_LOG_FATAL) {
+        printf("\033[1;31m");  // Red color for errors and fatal issues
+    } else if (level == LogLevel::ISF_LOG_WARN) {
+        printf("\033[1;33m");  // Yellow color for warnings
+    }
+#endif
+
+    // Print the actual log message
+    va_list args;
+    va_start(args, format);
+    vprintf(format, args);
+    va_end(args);
+
+    // Reset text color if needed, but only if not on iOS
+#ifndef TARGET_OS_IOS
+    if (level == LogLevel::ISF_LOG_ERROR || level == LogLevel::ISF_LOG_WARN || level == LogLevel::ISF_LOG_FATAL) {
+        printf("\033[0m");  // Reset color
+    }
+#endif
+
+    printf("\n");  // New line after log message
+
+    // If the log level is fatal, flush the error stream and abort the program
+    if (level == LogLevel::ISF_LOG_FATAL) {
+        std::flush(std::cerr);
+        abort();
+    }
+}
+#endif
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/meta.cpp.in b/cpp-package/inspireface/cpp/inspireface/meta.cpp.in
new file mode 100644
index 000000000..fb8b36257
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/meta.cpp.in
@@ -0,0 +1,137 @@
+#include "meta.h"
+#include <sstream>
+#include <iomanip>
+#include <cstdio>
+
+namespace inspire {
+
+// static global SDK info instance
+static const SDKInfo g_sdk_info = {
+    // version
+    @INSPIRE_FACE_VERSION_MAJOR_STR@,                          // version_major
+    @INSPIRE_FACE_VERSION_MINOR_STR@,                          // version_minor
+    @INSPIRE_FACE_VERSION_PATCH_STR@,                          // version_patch
+    
+    // series name
+    "@INSPIRE_FACE_SERIES_NAME@",                              // series
+    
+    // build info
+    __DATE__,                   // build_date
+    __TIME__,                   // build_time
+    
+#ifdef __GNUC__
+    "GCC " __VERSION__,         // compiler
+#elif defined(_MSC_VER)
+    "MSVC " + std::to_string(_MSC_VER),
+#elif defined(__clang__)
+    "Clang " __VERSION__,
+#else
+    "Unknown Compiler",
+#endif
+    
+#ifdef _WIN32
+    "Windows",                  // platform
+#elif defined(__APPLE__)
+    "MacOS",
+#elif defined(__linux__)
+    "Linux",
+#elif defined(__ANDROID__)
+    "Android",
+#else
+    "Unknown Platform",
+#endif
+
+#if defined(__x86_64__) || defined(_M_X64)
+    "x86_64",                   // architecture
+#elif defined(__i386) || defined(_M_IX86)
+    "x86",
+#elif defined(__aarch64__) || defined(_M_ARM64)
+    "arm64",
+#elif defined(__arm__) || defined(_M_ARM)
+    "arm",
+#else
+    "unknown",
+#endif
+
+    // backend info
+#if defined(ISF_ENABLE_TENSORRT)
+    "TensorRT(CUDA)",             // inference_backend
+#elif defined(ISF_ENABLE_APPLE_EXTENSION)
+    "CoreML(ANE)",
+#elif defined(ISF_ENABLE_RKNN)
+    "RKNN(RKNPU)",
+#else
+    "MNN(CPU-Only)",
+#endif
+
+#if defined(INSPIRECV_BACKEND_OPENCV)
+    "InspireCV-OpenCV",                   // inspirecv_backend
+#else
+    "InspireCV",
+#endif
+
+#if defined(ISF_ENABLE_RGA)
+    true,                       // rga_backend_enabled
+#else
+    false,
+#endif
+    
+    "InspireFace SDK - Advanced Face Recognition Library"  // description
+};
+
+std::string SDKInfo::GetFullVersionInfo() const {
+    std::stringstream ss;
+    ss << "InspireFace SDK " << "[" << series << "]" << " v" << version_major << "." << version_minor << "." << version_patch
+       << " (Built: " << build_date << " " << build_time << ")"
+       << " [" << platform << "/" << architecture << "/" << compiler << "]"
+       << " Backend: " << inference_backend << " " << inspirecv_backend << "[CPU" << (rga_backend_enabled ? ", RGA" : "") << "]";
+    return ss.str();
+}
+
+std::string SDKInfo::GetVersionMajorStr() const {
+    return std::to_string(version_major);
+}
+
+std::string SDKInfo::GetVersionMinorStr() const {
+    return std::to_string(version_minor);
+}
+
+std::string SDKInfo::GetVersionPatchStr() const {
+    return std::to_string(version_patch);
+}
+
+std::string SDKInfo::GetVersionString() const {
+    return std::to_string(version_major) + "." + std::to_string(version_minor) + "." + std::to_string(version_patch);
+}
+
+const SDKInfo& GetSDKInfo() {
+    return g_sdk_info;
+}
+
+}   // namespace inspire
+
+extern "C" {
+
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionMajorStr() {
+    static char buffer[16];
+    snprintf(buffer, sizeof(buffer), "%d", inspire::g_sdk_info.version_major);
+    return buffer;
+}
+
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionMinorStr() {
+    static char buffer[16];
+    snprintf(buffer, sizeof(buffer), "%d", inspire::g_sdk_info.version_minor);
+    return buffer;
+}
+
+INSPIRE_API_EXPORT const char* GetInspireFaceVersionPatchStr() {
+    static char buffer[16];
+    snprintf(buffer, sizeof(buffer), "%d", inspire::g_sdk_info.version_patch);
+    return buffer;
+}
+
+INSPIRE_API_EXPORT const char* GetInspireFaceExtendedInformation() {
+    return inspire::g_sdk_info.description.c_str();
+}
+
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/any_net_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/any_net_adapter.h
new file mode 100644
index 000000000..881505bfa
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/any_net_adapter.h
@@ -0,0 +1,321 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_ANYNETADAPTER_H
+#define INSPIREFACE_ANYNETADAPTER_H
+
+#include <utility>
+#include <inspirecv/inspirecv.h>
+#include "data_type.h"
+#include "inference_wrapper/inference_wrapper.h"
+#include "configurable.h"
+#include "log.h"
+#include "model_archive/inspire_archive.h"
+#include "image_process/nexus_processor/image_processor.h"
+#include <launch.h>
+#include "system.h"
+
+namespace inspire {
+
+using AnyTensorOutputs = std::vector<std::pair<std::string, std::vector<float>>>;
+
+/**
+ * @class AnyNet
+ * @brief Generic neural network class for various inference tasks.
+ *
+ * This class provides a general interface for different types of neural networks,
+ * facilitating loading parameters, initializing models, and executing forward passes.
+ */
+class INSPIRE_API AnyNetAdapter {
+    CONFIGURABLE_SUPPORT
+
+public:
+    /**
+     * @brief Constructor for AnyNet.
+     * @param name Name of the neural network.
+     */
+    explicit AnyNetAdapter(std::string name) : m_name_(std::move(name)) {
+        m_processor_ = nexus::ImageProcessor::Create(INSPIREFACE_CONTEXT->GetImageProcessingBackend());
+#if defined(ISF_ENABLE_RKNN)
+        m_processor_->SetAlignedWidth(INSPIREFACE_CONTEXT->GetImageProcessAlignedWidth());
+#endif
+    }
+
+    ~AnyNetAdapter() {
+        m_nn_inference_->Finalize();
+    }
+
+    /**
+     * @brief Loads parameters and initializes the model for inference.
+     * @param param Parameters for network configuration.
+     * @param model Pointer to the model.
+     * @param type Type of the inference helper (default: INFER_MNN).
+     * @return int32_t Status of the loading and initialization process.
+     */
+    int32_t LoadData(InspireModel &model, InferenceWrapper::EngineType type = InferenceWrapper::INFER_MNN, bool dynamic = false) {
+        m_infer_type_ = type;
+        // must
+        pushData<int>(model.Config(), "model_index", 0);
+        pushData<std::string>(model.Config(), "input_layer", "");
+        pushData<std::vector<std::string>>(model.Config(), "outputs_layers",
+                                           {
+                                             "",
+                                           });
+        pushData<std::vector<int>>(model.Config(), "input_size", {320, 320});
+        pushData<std::vector<float>>(model.Config(), "mean", {127.5f, 127.5f, 127.5f});
+        pushData<std::vector<float>>(model.Config(), "norm", {0.0078125f, 0.0078125f, 0.0078125f});
+        // rarely
+        pushData<int>(model.Config(), "input_channel", 3);
+        pushData<int>(model.Config(), "input_image_channel", 3);
+        pushData<bool>(model.Config(), "nchw", true);
+        pushData<bool>(model.Config(), "swap_color", false);
+        pushData<int>(model.Config(), "data_type", InputTensorInfo::InputTensorInfo::DataTypeImage);
+        pushData<int>(model.Config(), "input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeFp32);
+        pushData<int>(model.Config(), "output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeFp32);
+        pushData<int>(model.Config(), "infer_backend", 0);
+        pushData<int>(model.Config(), "threads", 1);
+
+        m_nn_inference_.reset(InferenceWrapper::Create(m_infer_type_));
+        m_nn_inference_->SetNumThreads(getData<int>("threads"));
+
+        if (m_infer_type_ == InferenceWrapper::INFER_TENSORRT) {
+            m_nn_inference_->SetDevice(INSPIREFACE_CONTEXT->GetCudaDeviceId());
+        }
+
+#if defined(ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA) && !defined(ISF_ENABLE_RKNN)
+        INSPIRE_LOGW("You have forced the global use of MNN_CUDA as the neural network inference backend");
+        m_nn_inference_->SetSpecialBackend(InferenceWrapper::MMM_CUDA);
+#endif
+
+#if defined(ISF_ENABLE_APPLE_EXTENSION)
+        if (INSPIREFACE_CONTEXT->GetGlobalCoreMLInferenceMode() == InferenceWrapper::COREML_CPU) {
+            m_nn_inference_->SetSpecialBackend(InferenceWrapper::COREML_CPU);
+        } else if (INSPIREFACE_CONTEXT->GetGlobalCoreMLInferenceMode() == InferenceWrapper::COREML_GPU) {
+            m_nn_inference_->SetSpecialBackend(InferenceWrapper::COREML_GPU);
+        } else if (INSPIREFACE_CONTEXT->GetGlobalCoreMLInferenceMode() == InferenceWrapper::COREML_ANE) {
+            m_nn_inference_->SetSpecialBackend(InferenceWrapper::COREML_ANE);
+        }
+#endif
+
+        m_output_tensor_info_list_.clear();
+        std::vector<std::string> outputs_layers = getData<std::vector<std::string>>("outputs_layers");
+        int tensor_type = getData<int>("input_tensor_type");
+        int out_tensor_type = getData<int>("output_tensor_type");
+        for (auto &name : outputs_layers) {
+            m_output_tensor_info_list_.push_back(OutputTensorInfo(name, out_tensor_type));
+        }
+        int32_t ret;
+        if (model.loadFilePath) {
+            auto extensionPath = INSPIREFACE_CONTEXT->GetExtensionPath();
+            if (extensionPath.empty()) {
+                INSPIRE_LOGE("Extension path is empty");
+                return InferenceWrapper::WrapperError;
+            }
+            std::string filePath = os::PathJoin(extensionPath, model.fullname);
+            ret = m_nn_inference_->Initialize(filePath, m_input_tensor_info_list_, m_output_tensor_info_list_);
+        } else {
+            ret = m_nn_inference_->Initialize(model.buffer, model.bufferSize, m_input_tensor_info_list_, m_output_tensor_info_list_);
+        }
+        if (ret != InferenceWrapper::WrapperOk) {
+            INSPIRE_LOGE("NN Initialize fail");
+            return ret;
+        }
+
+        if (ret != InferenceWrapper::WrapperOk) {
+            INSPIRE_LOGE("NN Initialize fail");
+            return ret;
+        }
+
+        m_input_tensor_info_list_.clear();
+        InputTensorInfo input_tensor_info(getData<std::string>("input_layer"), tensor_type, getData<bool>("nchw"));
+        std::vector<int> input_size = getData<std::vector<int>>("input_size");
+        int width = input_size[0];
+        int height = input_size[1];
+        m_input_image_size_ = {width, height};
+        int channel = getData<int>("input_channel");
+        if (getData<bool>("nchw")) {
+            input_tensor_info.tensor_dims = {1, channel, m_input_image_size_.GetHeight(), m_input_image_size_.GetWidth()};
+        } else {
+            input_tensor_info.tensor_dims = {1, m_input_image_size_.GetHeight(), m_input_image_size_.GetWidth(), channel};
+        }
+
+        input_tensor_info.data_type = getData<int>("data_type");
+        int image_channel = getData<int>("input_image_channel");
+        input_tensor_info.image_info.channel = image_channel;
+
+        std::vector<float> mean = getData<std::vector<float>>("mean");
+        std::vector<float> norm = getData<std::vector<float>>("norm");
+        input_tensor_info.normalize.mean[0] = mean[0];
+        input_tensor_info.normalize.mean[1] = mean[1];
+        input_tensor_info.normalize.mean[2] = mean[2];
+        input_tensor_info.normalize.norm[0] = norm[0];
+        input_tensor_info.normalize.norm[1] = norm[1];
+        input_tensor_info.normalize.norm[2] = norm[2];
+
+        input_tensor_info.image_info.width = width;
+        input_tensor_info.image_info.height = height;
+        input_tensor_info.image_info.channel = channel;
+        input_tensor_info.image_info.crop_x = 0;
+        input_tensor_info.image_info.crop_y = 0;
+        input_tensor_info.image_info.crop_width = width;
+        input_tensor_info.image_info.crop_height = height;
+        input_tensor_info.image_info.is_bgr = getData<bool>("nchw");
+        input_tensor_info.image_info.swap_color = getData<bool>("swap_color");
+
+        m_input_tensor_info_list_.push_back(input_tensor_info);
+
+        if (dynamic) {
+            m_nn_inference_->ResizeInput(m_input_tensor_info_list_);
+        }
+
+        return 0;
+    }
+
+    void Forward(const inspirecv::Image &image, AnyTensorOutputs &outputs) {
+        InputTensorInfo &input_tensor_info = getMInputTensorInfoList()[0];
+        if (m_infer_type_ == InferenceWrapper::INFER_RKNN) {
+            if (getData<bool>("swap_color")) {
+                m_cache_ = image.SwapRB();
+                input_tensor_info.data = (uint8_t *)m_cache_.Data();
+            } else {
+                input_tensor_info.data = (uint8_t *)image.Data();
+            }
+        } else {
+            input_tensor_info.data = (uint8_t *)image.Data();
+        }
+        Forward(outputs);
+    }
+
+    /**
+     * @brief Performs a forward pass of the network.
+     * @param outputs Outputs of the network (tensor outputs).
+     */
+    void Forward(AnyTensorOutputs &outputs) {
+        //        LOGD("ppPreProcess");
+        if (m_nn_inference_->PreProcess(m_input_tensor_info_list_) != InferenceWrapper::WrapperOk) {
+            INSPIRE_LOGD("PreProcess error");
+        }
+        //        LOGD("PreProcess");
+        if (m_nn_inference_->Process(m_output_tensor_info_list_) != InferenceWrapper::WrapperOk) {
+            INSPIRE_LOGD("Process error");
+        }
+        //        LOGD("Process");
+        for (int i = 0; i < m_output_tensor_info_list_.size(); ++i) {
+            std::vector<float> output_score_raw_list(m_output_tensor_info_list_[i].GetDataAsFloat(),
+                                                     m_output_tensor_info_list_[i].GetDataAsFloat() + m_output_tensor_info_list_[i].GetElementNum());
+            //            LOGE("m_output_tensor_info_list_[i].GetElementNum(): %d",m_output_tensor_info_list_[i].GetElementNum());
+            outputs.push_back(std::make_pair(m_output_tensor_info_list_[i].name, output_score_raw_list));
+        }
+    }
+
+public:
+    /**
+     * @brief Gets a reference to the input tensor information list.
+     * @return Reference to the vector of input tensor information.
+     */
+    std::vector<InputTensorInfo> &getMInputTensorInfoList() {
+        return m_input_tensor_info_list_;
+    }
+
+    /**
+     * @brief Gets a reference to the output tensor information list.
+     * @return Reference to the vector of output tensor information.
+     */
+    std::vector<OutputTensorInfo> &getMOutputTensorInfoList() {
+        return m_output_tensor_info_list_;
+    }
+
+    /**
+     * @brief Gets the size of the input image.
+     * @return Size of the input image.
+     */
+    inspirecv::Size<int> &getMInputImageSize() {
+        return m_input_image_size_;
+    }
+
+    /**
+     * @brief Softmax function.
+     *
+     * @param input The input vector.
+     * @return The softmax result.
+     */
+    static std::vector<float> Softmax(const std::vector<float> &input) {
+        std::vector<float> result;
+        float sum = 0.0;
+
+        // Calculate the exponentials and the sum of exponentials
+        for (float x : input) {
+            float exp_x = std::exp(x);
+            result.push_back(exp_x);
+            sum += exp_x;
+        }
+
+        // Normalize by dividing each element by the sum
+        for (float &value : result) {
+            value /= sum;
+        }
+
+        return result;
+    }
+
+protected:
+    std::string m_name_;  ///< Name of the neural network.
+
+    std::unique_ptr<nexus::ImageProcessor> m_processor_;  ///< Assign a nexus processor to each anynet object
+
+private:
+    InferenceWrapper::EngineType m_infer_type_;                ///< Inference engine type
+    std::shared_ptr<InferenceWrapper> m_nn_inference_;         ///< Shared pointer to the inference helper.
+    std::vector<InputTensorInfo> m_input_tensor_info_list_;    ///< List of input tensor information.
+    std::vector<OutputTensorInfo> m_output_tensor_info_list_;  ///< List of output tensor information.
+    inspirecv::Size<int> m_input_image_size_{};                ///< Size of the input image.
+    inspirecv::Image m_cache_;                                 ///< Cached matrix for image data.
+};
+
+template <typename ImageT, typename TensorT>
+AnyTensorOutputs ForwardService(std::shared_ptr<AnyNetAdapter> net, const ImageT &input, std::function<void(const ImageT &, TensorT &)> transform) {
+    InputTensorInfo &input_tensor_info = net->getMInputTensorInfoList()[0];
+    TensorT transform_tensor;
+    transform(input, transform_tensor);
+    input_tensor_info.data = transform_tensor.data;  // input tensor only support cv2::Mat
+
+    AnyTensorOutputs outputs;
+    net->Forward(outputs);
+
+    return outputs;
+}
+
+/**
+ * @brief Executes a forward pass through the neural network for a given input, with preprocessing.
+ * @tparam ImageT Type of the input image.
+ * @tparam TensorT Type of the transformed tensor.
+ * @tparam PreprocessCallbackT Type of the preprocessing callback function.
+ * @param net Shared pointer to the AnyNet neural network object.
+ * @param input The input image to be processed.
+ * @param callback Preprocessing callback function to be applied to the input.
+ * @param transform Transformation function to convert the input image to a tensor.
+ * @return AnyTensorOutputs Outputs of the network (tensor outputs).
+ *
+ * This template function handles the preprocessing of the input image, transformation to tensor,
+ * and then passes it through the neural network to get the output. The function is generic and
+ * can work with different types of images and tensors, as specified by the template parameters.
+ */
+template <typename ImageT, typename TensorT, typename PreprocessCallbackT>
+AnyTensorOutputs ForwardService(std::shared_ptr<AnyNetAdapter> net, const ImageT &input, PreprocessCallbackT &callback,
+                                std::function<void(const ImageT &, TensorT &, PreprocessCallbackT &)> transform) {
+    InputTensorInfo &input_tensor_info = net->getMInputTensorInfoList()[0];
+    TensorT transform_tensor;
+    transform(input, transform_tensor, callback);
+    input_tensor_info.data = transform_tensor.data;  // input tensor only support cv2::Mat
+
+    AnyTensorOutputs outputs;
+    net->Forward(outputs);
+
+    return outputs;
+}
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_ANYNETADAPTER_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/configurable.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/configurable.cpp
new file mode 100644
index 000000000..41593d701
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/configurable.cpp
@@ -0,0 +1,32 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "configurable.h"
+
+namespace inspire {
+
+std::string Configurable::toString(int indent) const {
+    if (indent != 0)
+        return m_configuration.dump(indent);
+    else
+        return m_configuration.dump();
+}
+
+std::vector<std::string> Configurable::getNameList() const {
+    std::vector<std::string> keys;
+    for (const auto &element : m_configuration.items()) {
+        keys.push_back(element.key());
+    }
+    return keys;
+}
+
+Configurable &Configurable::operator=(const Configurable &other) {
+    if (this != &other) {                         // Check the self-assignment
+        m_configuration = other.m_configuration;  // Deep copy using the assignment operator of nlohmann::json
+    }
+    return *this;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/configurable.h b/cpp-package/inspireface/cpp/inspireface/middleware/configurable.h
new file mode 100644
index 000000000..77cde3c54
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/configurable.h
@@ -0,0 +1,160 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#pragma once
+#ifndef INSPIREFACE_CONFIGURABLE_H
+#define INSPIREFACE_CONFIGURABLE_H
+#include "nlohmann/json.hpp"
+#include <iostream>
+
+using nlohmann::json;
+
+#ifndef INSPIRE_API
+#define INSPIRE_API
+#endif
+
+namespace inspire {
+
+/**
+ * @class Configurable
+ * @brief Class for managing parameters as JSON data.
+ *
+ * This class provides methods to set, get, and load parameters as JSON data.
+ */
+class INSPIRE_API Configurable {
+public:
+    Configurable() = default;
+    ;
+
+    Configurable(const Configurable& p) : m_configuration(p.m_configuration) {}
+
+    virtual ~Configurable() {}
+
+    Configurable& operator=(const Configurable& other);
+
+    /**
+     * @brief Get a list of parameter names.
+     * @return std::vector<std::string> A list of parameter names.
+     */
+    std::vector<std::string> getNameList() const;
+
+    /**
+     * @brief Convert parameters to a formatted JSON string.
+     * @param indent The indentation level for formatting.
+     * @return std::string The formatted JSON string representing the parameters.
+     */
+    std::string toString(int indent = 4) const;
+
+    /**
+     * @brief Check if a parameter exists by name.
+     * @param name The name of the parameter to check.
+     * @return bool True if the parameter exists, false otherwise.
+     */
+    bool has(const std::string& name) const noexcept {
+        return m_configuration.contains(name);
+    }
+
+    /**
+     * @brief Set a parameter with a specific name and value.
+     * @param name The name of the parameter.
+     * @param value The value to set for the parameter.
+     */
+    template <typename ValueType>
+    void set(const std::string& name, const ValueType& value) {
+        m_configuration[name] = value;
+    }
+
+    /**
+     * @brief Set a parameter with a specific name and value.
+     * @param name The name of the parameter.
+     * @param value The value to set for the parameter.
+     */
+    template <typename ValueType>
+    ValueType get(const std::string& name) const {
+        if (!has(name)) {
+            throw std::out_of_range("out_of_range in Parameter::get : " + name);
+        }
+        return m_configuration.at(name).get<ValueType>();
+    }
+
+    /**
+     * @brief Load parameters from a JSON object.
+     * @param j The JSON object containing parameters.
+     */
+    void load(const nlohmann::json& j) {
+        for (const auto& item : j.items()) {
+            const auto& key = item.key();
+            const auto& value = item.value();
+
+            if (value.is_boolean()) {
+                set<bool>(key, value.get<bool>());
+            } else if (value.is_number_integer()) {
+                set<int>(key, value.get<int>());
+            } else if (value.is_number_float()) {
+                set<float>(key, value.get<float>());
+            } else if (value.is_string()) {
+                set<std::string>(key, value.get<std::string>());
+            } else if (value.is_array()) {
+                if (!value.empty()) {
+                    if (value[0].is_number_integer()) {
+                        set<std::vector<int>>(key, value.get<std::vector<int>>());
+                    } else if (value[0].is_number_float()) {
+                        set<std::vector<float>>(key, value.get<std::vector<float>>());
+                    }  // add more types as needed
+                    // ...
+                }
+            }
+            // Add handling for other types as needed
+        }
+    }
+
+private:
+    json m_configuration;  ///< JSON object to store parameters.
+};
+
+#define CONFIGURABLE_SUPPORT                                                                                     \
+protected:                                                                                                       \
+    inspire::Configurable m_configuration;                                                                       \
+                                                                                                                 \
+public:                                                                                                          \
+    const inspire::Configurable& getConfiguration() const {                                                      \
+        return m_configuration;                                                                                  \
+    }                                                                                                            \
+                                                                                                                 \
+    void setConfiguration(const inspire::Configurable& param) {                                                  \
+        m_configuration = param;                                                                                 \
+    }                                                                                                            \
+                                                                                                                 \
+    bool hasData(const std::string& name) const noexcept {                                                       \
+        return m_configuration.has(name);                                                                        \
+    }                                                                                                            \
+                                                                                                                 \
+    template <typename ValueType>                                                                                \
+    void setData(const std::string& name, const ValueType& value) {                                              \
+        m_configuration.set<ValueType>(name, value);                                                             \
+    }                                                                                                            \
+                                                                                                                 \
+    template <typename ValueType>                                                                                \
+    ValueType getData(const std::string& name) const {                                                           \
+        return m_configuration.get<ValueType>(name);                                                             \
+    }                                                                                                            \
+                                                                                                                 \
+    template <typename ValueType>                                                                                \
+    void pushData(const inspire::Configurable& param, const std::string& name, const ValueType& default_value) { \
+        if (param.has(name)) {                                                                                   \
+            setData<ValueType>(name, param.get<ValueType>(name));                                                \
+        } else {                                                                                                 \
+            setData<ValueType>(name, default_value);                                                             \
+        }                                                                                                        \
+    }                                                                                                            \
+    void loadData(const nlohmann::json& j) {                                                                     \
+        m_configuration.load(j);                                                                                 \
+    }                                                                                                            \
+    std::string toStr(int indent = 4) {                                                                          \
+        return m_configuration.toString(indent);                                                                 \
+    }
+
+}  // namespace inspire
+#endif  // INSPIREFACE_CONFIGURABLE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/costman.h b/cpp-package/inspireface/cpp/inspireface/middleware/costman.h
new file mode 100644
index 000000000..788a1ce68
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/costman.h
@@ -0,0 +1,37 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_COSTMAN_H
+#define INSPIRE_FACE_COSTMAN_H
+#include <chrono>
+
+namespace inspire {
+
+class Timer {
+public:
+    Timer() {
+        current_time = std::chrono::high_resolution_clock::now();
+    }
+
+    double GetCostTime() const {
+        auto now = std::chrono::high_resolution_clock::now();
+        auto duration = std::chrono::duration_cast<std::chrono::nanoseconds>(now - current_time);
+        return duration.count() / 1000000.0;  // Convert to milliseconds
+    }
+
+    double GetCostTimeUpdate() {
+        auto now = std::chrono::high_resolution_clock::now();
+        auto duration = std::chrono::duration_cast<std::chrono::nanoseconds>(now - current_time);
+        current_time = now;
+        return duration.count() / 1000000.0;  // Convert to milliseconds
+    }
+
+private:
+    std::chrono::time_point<std::chrono::high_resolution_clock> current_time;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_COSTMAN_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/cuda_toolkit.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/cuda_toolkit.cpp
new file mode 100644
index 000000000..413fe1549
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/cuda_toolkit.cpp
@@ -0,0 +1,130 @@
+#ifdef ISF_ENABLE_TENSORRT
+#include <cuda_toolkit.h>
+#include <cuda_runtime_api.h>
+#include <NvInfer.h>
+#endif  // ISF_ENABLE_TENSORRT
+#include <log.h>
+#include <herror.h>
+
+namespace inspire {
+
+int32_t INSPIRE_API_EXPORT GetCudaDeviceCount(int32_t *device_count) {
+#ifdef ISF_ENABLE_TENSORRT
+    cudaError_t error = cudaGetDeviceCount(device_count);
+    if (error != cudaSuccess) {
+        INSPIRE_LOGE("CUDA error: %s", cudaGetErrorString(error));
+        return HERR_DEVICE_CUDA_UNKNOWN_ERROR;
+    }
+    return HSUCCEED;
+#else
+    *device_count = 0;
+    return HERR_DEVICE_CUDA_NOT_SUPPORT;
+#endif
+}
+
+int32_t INSPIRE_API_EXPORT CheckCudaUsability(int32_t *is_support) {
+#ifdef ISF_ENABLE_TENSORRT
+    int device_count;
+    auto ret = GetCudaDeviceCount(&device_count);
+    if (ret != HSUCCEED) {
+        return ret;
+    }
+    if (device_count == 0) {
+        *is_support = 0;
+        INSPIRE_LOGE("No CUDA devices found");
+        return HERR_DEVICE_CUDA_NOT_SUPPORT;
+    }
+    *is_support = device_count > 0;
+    return HSUCCEED;
+#else
+    *is_support = 0;
+    return HERR_DEVICE_CUDA_NOT_SUPPORT;
+#endif
+}
+
+int32_t INSPIRE_API_EXPORT _PrintCudaDeviceInfo() {
+#ifdef ISF_ENABLE_TENSORRT
+    try {
+        INSPIRE_LOGW("TensorRT version: %d.%d.%d", NV_TENSORRT_MAJOR, NV_TENSORRT_MINOR, NV_TENSORRT_PATCH);
+
+        // check if CUDA is available
+        int device_count;
+        cudaError_t error = cudaGetDeviceCount(&device_count);
+        if (error != cudaSuccess) {
+            INSPIRE_LOGE("CUDA error: %s", cudaGetErrorString(error));
+            return HERR_DEVICE_CUDA_UNKNOWN_ERROR;
+        }
+        INSPIRE_LOGW("available CUDA devices: %d", device_count);
+
+        // get current CUDA device
+        int currentDevice;
+        error = cudaGetDevice(&currentDevice);
+        if (error != cudaSuccess) {
+            INSPIRE_LOGE("[CUDA error] failed to get current CUDA device: %s", cudaGetErrorString(error));
+            return HERR_DEVICE_CUDA_UNKNOWN_ERROR;
+        }
+        INSPIRE_LOGW("current CUDA device ID: %d", currentDevice);
+
+        // get GPU device properties
+        cudaDeviceProp prop;
+        error = cudaGetDeviceProperties(&prop, currentDevice);
+        if (error != cudaSuccess) {
+            INSPIRE_LOGE("[CUDA error] failed to get CUDA device properties: %s", cudaGetErrorString(error));
+            return HERR_DEVICE_CUDA_UNKNOWN_ERROR;
+        }
+
+        // print device detailed information
+        INSPIRE_LOGW("\nCUDA device details:");
+        INSPIRE_LOGW("device name: %s", prop.name);
+        INSPIRE_LOGW("compute capability: %d.%d", prop.major, prop.minor);
+        INSPIRE_LOGW("global memory: %d MB", prop.totalGlobalMem / (1024 * 1024));
+        INSPIRE_LOGW("max shared memory/block: %d KB", prop.sharedMemPerBlock / 1024);
+        INSPIRE_LOGW("max threads/block: %d", prop.maxThreadsPerBlock);
+        INSPIRE_LOGW("max block dimensions: (%d, %d, %d)", prop.maxThreadsDim[0], prop.maxThreadsDim[1], prop.maxThreadsDim[2]);
+        INSPIRE_LOGW("max grid size: (%d, %d, %d)", prop.maxGridSize[0], prop.maxGridSize[1], prop.maxGridSize[2]);
+        INSPIRE_LOGW("total constant memory: %d KB", prop.totalConstMem / 1024);
+        INSPIRE_LOGW("multi-processor count: %d", prop.multiProcessorCount);
+        INSPIRE_LOGW("max blocks per multi-processor: %d", prop.maxBlocksPerMultiProcessor);
+        INSPIRE_LOGW("clock frequency: %d MHz", prop.clockRate / 1000);
+        INSPIRE_LOGW("memory frequency: %d MHz", prop.memoryClockRate / 1000);
+        INSPIRE_LOGW("memory bus width: %d bits", prop.memoryBusWidth);
+        INSPIRE_LOGW("L2 cache size: %d KB", prop.l2CacheSize / 1024);
+        INSPIRE_LOGW("theoretical memory bandwidth: %f GB/s", 2.0 * prop.memoryClockRate * (prop.memoryBusWidth / 8) / 1.0e6);
+
+        // check if FP16 is supported
+        bool supportsFP16 = prop.major >= 6 || (prop.major == 5 && prop.minor >= 3);
+        INSPIRE_LOGW("FP16 support: %s", supportsFP16 ? "yes" : "no");
+
+        // check if unified memory is supported
+        INSPIRE_LOGW("unified memory support: %s", prop.unifiedAddressing ? "yes" : "no");
+
+        // check if concurrent kernel execution is supported
+        INSPIRE_LOGW("concurrent kernel execution: %s", prop.concurrentKernels ? "yes" : "no");
+
+        // check if asynchronous engine is supported
+        INSPIRE_LOGW("asynchronous engine count: %d", prop.asyncEngineCount);
+
+        return HSUCCEED;
+    } catch (const std::exception &e) {
+        INSPIRE_LOGE("error when printing CUDA device info: %s", e.what());
+        return HERR_DEVICE_CUDA_UNKNOWN_ERROR;
+    }
+#else
+    INSPIRE_LOGE("CUDA/TensorRT support is not enabled");
+    return HERR_DEVICE_CUDA_NOT_SUPPORT;
+#endif
+}
+
+int32_t INSPIRE_API_EXPORT PrintCudaDeviceInfo() {
+#ifdef ISF_ENABLE_TENSORRT
+    INSPIRE_LOGW("================================================");
+    auto ret = _PrintCudaDeviceInfo();
+    INSPIRE_LOGW("================================================");
+    return ret;
+#else
+    INSPIRE_LOGE("CUDA/TensorRT support is not enabled");
+    return HERR_DEVICE_CUDA_NOT_SUPPORT;
+#endif
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/README.md b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/README.md
new file mode 100644
index 000000000..d4d6476d9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/README.md
@@ -0,0 +1 @@
+Plan, not implement.
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/implement/mnn_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/implement/mnn_adapter.h
new file mode 100644
index 000000000..4770a9c07
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/implement/mnn_adapter.h
@@ -0,0 +1,33 @@
+#ifndef MNN_ADAPTER_IMPL__
+#define MNN_ADAPTER_IMPL__
+#include "../inference_adapter.h"
+#include "opencv2/opencv.hpp"
+#include <MNN/ImageProcess.hpp>
+#include <MNN/Interpreter.hpp>
+#include <MNN/MNNDefine.h>
+#include <MNN/Tensor.hpp>
+#include <MNN/MNNForwardType.h>
+
+class MNNCVAdapter : public InferenceAdapter {
+public:
+
+    MNNCVAdapter() {};
+
+    ~MNNCVAdapter() override {};
+
+private:
+    std::shared_ptr<MNN::Interpreter> detect_model_;
+    MNN::Tensor *input_{};
+    std::vector<MNN::Tensor*> output_tensors_;
+    MNN::Session *sess{};
+    std::vector<int> tensor_shape_;
+    MNN::ScheduleConfig _config;
+    MNNForwardType backend_;
+    int width_{};
+    int height_{};
+    float mean[3]{};
+    float normal[3]{};
+
+};
+
+#endif  // MNN_ADAPTER_IMPL__
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/inference_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/inference_adapter.h
new file mode 100644
index 000000000..a99b64eee
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/inference_adapter.h
@@ -0,0 +1,98 @@
+#ifndef INSPIREFACE_OMNI_INFERENACE__
+#define INSPIREFACE_OMNI_INFERENACE__
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+
+class XOutputData {
+public:
+    XOutputData() : size(0), data(nullptr) {}
+
+    std::vector<float> CopyToFloatArray() {
+        if (!buffer.empty()) {
+            return buffer;
+        }
+
+        std::vector<float> floatArray;
+        floatArray.resize(size);
+        std::memcpy(floatArray.data(), data, size * sizeof(float));
+        return floatArray;
+    }
+
+public:
+    size_t size;
+    float* data;                // Use pointer
+    std::vector<float> buffer;  // Use copy
+};
+
+typedef std::vector<XOutputData> XOutputDataList;
+
+class XTransform {
+public:
+    XTransform() : swap_color(false) {
+        std::fill(std::begin(normalize.mean), std::end(normalize.mean), 0.0f);
+        std::fill(std::begin(normalize.norm), std::end(normalize.norm), 1.0f);
+    }
+
+    bool swap_color;
+
+    struct {
+        float mean[3];
+        float norm[3];
+    } normalize;
+};
+
+class XInputData {
+public:
+    XInputData() : nchw(false), bgr(false), height(0), width(0), channel(0), data(nullptr) {}
+
+public:
+    bool nchw;
+    bool bgr;
+    int32_t height;
+    int32_t width;
+    int32_t channel;
+    uint8_t* data;
+};
+
+typedef enum {
+    xEngineMNN,
+    xEngineRKNN,
+} EngineType;
+
+class InferenceAdapter {
+public:
+    enum {
+        xRetOk = 0,
+        xRetErr = -1,
+    };
+
+    typedef enum { xDefaultCPU, xMNNCuda, xCoreML } SpecialBackend;
+
+public:
+    virtual ~InferenceAdapter() {};
+    virtual int32_t SetNumThreads(const int32_t num_threads) = 0;
+    virtual int32_t Initialize(const std::string& model_filename, const XTransform& transform, const std::string& input_name,
+                               const std::vector<std::string>& outputs_name) = 0;
+    virtual int32_t Initialize(char* model_buffer, int model_size, const std::string& input_name, const XTransform& transform,
+                               const std::vector<std::string>& outputs_name) = 0;
+    virtual int32_t Finalize(void) = 0;
+    virtual int32_t SetInputsData(const std::vector<XInputData>& batch, ) = 0;
+    virtual int32_t Forward(std::vector<XOutputDataList>& batch_outputs) = 0;
+
+    virtual int32_t ResizeInputs() = 0;
+
+    virtual int32_t SetSpecialBackend(SpecialBackend backend) {
+        special_backend_ = backend;
+        return xRetOk;
+    };
+
+protected:
+    EngineType engine_type_;
+    SpecialBackend special_backend_ = xDefaultCPU;
+};
+
+#endif  // INSPIREFACE_OMNI_INFERENACE__
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/mnn_cv_adapter_inference.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/mnn_cv_adapter_inference.h
new file mode 100755
index 000000000..6a0efbcbe
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_adapter/mnn_cv_adapter_inference.h
@@ -0,0 +1,185 @@
+#ifndef MNN_SIMPLE_INFER_H
+#define MNN_SIMPLE_INFER_H
+
+#include "opencv2/opencv.hpp"
+#include <MNN/ImageProcess.hpp>
+#include <MNN/Interpreter.hpp>
+#include <MNN/MNNDefine.h>
+#include <MNN/Tensor.hpp>
+#include <MNN/MNNForwardType.h>
+
+class MNNCVAdapterInference {
+public:
+    MNNCVAdapterInference(const std::string &model, int thread, const float mean[], const float normal[],
+                          bool use_model_bin = false) {
+#ifdef FORWARD_CUDA
+        backend_ = MNN_FORWARD_CUDA;
+#else
+        backend_ = MNN_FORWARD_CPU;
+#endif
+
+        if (use_model_bin) {
+            detect_model_ = std::shared_ptr<MNN::Interpreter>(
+                MNN::Interpreter::createFromBuffer(model.c_str(), model.size()));
+        } else {
+            detect_model_ = std::shared_ptr<MNN::Interpreter>(
+                MNN::Interpreter::createFromFile(model.c_str()));
+        }
+
+        _config.type = backend_;
+        _config.numThread = thread;
+        MNN::BackendConfig backendConfig;
+        backendConfig.precision = MNN::BackendConfig::Precision_High;
+        backendConfig.power = MNN::BackendConfig::Power_High;
+        _config.backendConfig = &backendConfig;
+        for (int i = 0; i < 3; i++) {
+            this->mean[i] = mean[i];
+            this->normal[i] = normal[i];
+        }
+    }
+
+    ~MNNCVAdapterInference() {
+        detect_model_->releaseModel();
+        detect_model_->releaseSession(sess);
+    }
+
+    void Init(const std::string &input, const std::vector<std::string> &outputs, int width,
+              int height) {
+        sess = detect_model_->createSession(_config);
+        tensor_shape_ = {1, 3, height, width};
+        input_ = detect_model_->getSessionInput(sess, input.c_str());
+
+        // Resize input tensor
+        detect_model_->resizeTensor(input_, tensor_shape_);
+        
+        // Store output tensors and resize them
+        for (const auto& output_name : outputs) {
+            auto output_tensor = detect_model_->getSessionOutput(sess, output_name.c_str());
+            output_tensors_.emplace_back(output_tensor);
+        }
+
+        detect_model_->resizeSession(sess);
+
+        width_ = width;
+        height_ = height;
+    }
+
+    std::vector<std::vector<float>> Infer(const cv::Mat &mat) {
+        assert(mat.rows == height_);
+        assert(mat.cols == width_);
+        MNN::CV::ImageProcess::Config config;
+        config.destFormat = MNN::CV::ImageFormat::BGR;
+        config.sourceFormat = MNN::CV::BGR;
+        for (int i = 0; i < 3; i++) {
+            config.mean[i] = mean[i];
+            config.normal[i] = normal[i];
+        }
+
+        std::unique_ptr<MNN::CV::ImageProcess> process(
+                MNN::CV::ImageProcess::create(config));
+        process->convert(mat.data, mat.cols, mat.rows, (int) mat.step1(), input_);
+        detect_model_->runSession(sess);
+        
+        std::vector<std::vector<float>> results;
+
+        for (auto output : output_tensors_) {
+            auto dimType = input_->getDimensionType();
+
+            if (output->getType().code != halide_type_float) {
+                dimType = MNN::Tensor::TENSORFLOW;
+            }
+            std::shared_ptr<MNN::Tensor> outputUser(new MNN::Tensor(output, dimType));
+            output->copyToHostTensor(outputUser.get());
+            auto type = outputUser->getType();
+            auto size = outputUser->elementSize();
+            std::vector<float> tempValues(size);
+            if (type.code == halide_type_float) {
+                auto values = outputUser->host<float>();
+                for (int i = 0; i < size; ++i) {
+                    tempValues[i] = values[i];
+                }
+            }
+            results.push_back(tempValues);
+        }
+
+        return results;
+    }
+
+    std::vector<std::vector<std::vector<float>>> BatchInfer(const std::vector<cv::Mat> &images) {
+        int batch_size = images.size();
+        tensor_shape_[0] = batch_size;  // Update batch size
+        detect_model_->resizeTensor(input_, tensor_shape_);
+        detect_model_->resizeSession(sess);
+
+        MNN::CV::ImageProcess::Config config;
+        config.destFormat = MNN::CV::ImageFormat::BGR;
+        config.sourceFormat = MNN::CV::BGR;
+        for (int i = 0; i < 3; i++) {
+            config.mean[i] = mean[i];
+            config.normal[i] = normal[i];
+        }
+
+        std::unique_ptr<MNN::CV::ImageProcess> process(
+                MNN::CV::ImageProcess::create(config));
+        
+        std::shared_ptr<MNN::Tensor> inputUser(new MNN::Tensor(input_, MNN::Tensor::TENSORFLOW));
+        auto size_h = inputUser->height();
+        auto size_w = inputUser->width();
+        auto bpp = inputUser->channel();
+
+        for (int batch = 0; batch < batch_size; ++batch) {
+            const auto& mat = images[batch];
+            assert(mat.rows == height_);
+            assert(mat.cols == width_);
+
+            // No need to setScale since the images are already resized
+            process->convert(mat.data, mat.cols, mat.rows, (int)mat.step1(), inputUser->host<uint8_t>() + inputUser->stride(0) * batch * inputUser->getType().bytes(), size_w, size_h, bpp, 0, inputUser->getType());
+        }
+
+        input_->copyFromHostTensor(inputUser.get());
+        detect_model_->runSession(sess);
+        
+        std::vector<std::vector<std::vector<float>>> all_results(batch_size);
+
+        for (auto output : output_tensors_) {
+            auto dimType = input_->getDimensionType();
+
+            if (output->getType().code != halide_type_float) {
+                dimType = MNN::Tensor::TENSORFLOW;
+            }
+            std::shared_ptr<MNN::Tensor> outputUser(new MNN::Tensor(output, dimType));
+            output->copyToHostTensor(outputUser.get());
+            auto type = outputUser->getType();
+            auto size = outputUser->elementSize() / batch_size;
+
+            for (int batch = 0; batch < batch_size; ++batch) {
+                std::vector<float> tempValues(size);
+                if (type.code == halide_type_float) {
+                    auto values = outputUser->host<float>() + batch * outputUser->stride(0);
+                    for (int i = 0; i < size; ++i) {
+                        tempValues[i] = values[i];
+                    }
+                }
+                all_results[batch].push_back(tempValues);
+            }
+        }
+
+        return all_results;
+    }
+
+    float mean[3]{};
+    float normal[3]{};
+
+private:
+    std::shared_ptr<MNN::Interpreter> detect_model_;
+    MNN::Tensor *input_{};
+    std::vector<MNN::Tensor*> output_tensors_;
+    MNN::Session *sess{};
+    std::vector<int> tensor_shape_;
+    MNN::ScheduleConfig _config;
+    MNNForwardType backend_;
+    int width_{};
+    int height_{};
+};
+
+#endif // MNN_SIMPLE_INFER_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.h
new file mode 100644
index 000000000..2e8937c6e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.h
@@ -0,0 +1,84 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-02-01
+ */
+
+#ifndef INSPIRE_FACE_COREML_ADAPTER_H
+#define INSPIRE_FACE_COREML_ADAPTER_H
+#include <string>
+#include <vector>
+#include <map>
+#ifdef __OBJC__
+@class MLModel;
+@protocol MLFeatureProvider;
+#else
+typedef void MLModel;
+typedef void MLFeatureProvider;
+#endif
+
+#define COREML_HSUCCEED 0
+#define COREML_HFAIL -1
+#define COREML_FORWARD_FAILED -2
+
+/**
+ * @brief CoreML adapter for inference
+ */
+class CoreMLAdapter {
+public:
+    /**
+     * @brief Inference mode
+     */
+    enum class InferenceMode {
+        CPU,  ///< CPU inference
+        GPU,  ///< GPU inference
+        ANE   ///< Automatic selection, ANE first
+    };
+
+    /**
+     * @brief Output shapes map
+     */
+    typedef std::map<std::string, std::vector<int>> OutputShapesMap;
+
+    CoreMLAdapter();
+    ~CoreMLAdapter();
+
+    /**
+     * @brief Read model from file
+     * @param modelPath model path
+     * @return 0 if success, -1 if failed
+     */
+    int32_t readFromFile(const std::string &modelPath);
+
+    static CoreMLAdapter readNetFrom(const std::string &modelPath);
+
+    /**
+     * @brief Not implemented
+     */
+    static CoreMLAdapter readNetFromBin(const std::vector<char> &model_data);
+
+    std::vector<std::string> getInputNames() const;
+    std::vector<std::string> getOutputNames() const;
+    std::vector<int> getInputShapeByName(const std::string &name);
+    std::vector<int> getOutputShapeByName(const std::string &name);
+
+    void setInput(const char *inputName, const char *data);
+    int32_t forward();
+    const char *getOutput(const char *nodeName);
+    const std::vector<int> &getOutputShapeByName(const std::string &name) const;
+
+    /**
+     * @brief Set inference mode
+     * @param mode inference mode
+     */
+    void setInferenceMode(InferenceMode mode);
+
+    void printModelInfo() const;
+
+private:
+    class Impl;
+    Impl *pImpl;
+
+    OutputShapesMap m_outputShapes;
+};
+
+#endif  // INSPIRE_FACE_COREML_ADAPTER_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.mm b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.mm
new file mode 100644
index 000000000..ec23bc796
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/coreml/CoreMLAdapter.mm
@@ -0,0 +1,278 @@
+#import "CoreMLAdapter.h"
+#import <CoreML/CoreML.h>
+#import <Foundation/Foundation.h>
+#import <Accelerate/Accelerate.h>
+
+@interface CoreMLAdapterImpl : NSObject
+@property (nonatomic, strong) MLModel *model;
+@property (nonatomic, strong) id<MLFeatureProvider> inputFeatures;
+@property (nonatomic, strong) id<MLFeatureProvider> outputFeatures;
+@property (nonatomic, strong) NSURL *modelURL;
+@end
+
+@implementation CoreMLAdapterImpl
+@end // CoreMLAdapterImpl
+
+
+class CoreMLAdapter::Impl {
+public:
+    CoreMLAdapterImpl *impl;
+};
+
+
+CoreMLAdapter::CoreMLAdapter() : pImpl(new Impl) {
+    pImpl->impl = [[CoreMLAdapterImpl alloc] init];
+}
+
+CoreMLAdapter::~CoreMLAdapter() {
+    delete pImpl;
+}
+
+int32_t CoreMLAdapter::readFromFile(const std::string &modelPath) {
+    NSString *modelPathStr = [NSString stringWithUTF8String:modelPath.c_str()];
+    NSURL *modelURL = [NSURL fileURLWithPath:modelPathStr];    
+    NSError *error = nil;
+    pImpl->impl.model = [MLModel modelWithContentsOfURL:modelURL error:&error];
+    pImpl->impl.modelURL = modelURL;
+    if (error) {
+        NSLog(@"Error loading model: %@", error);
+        return COREML_HFAIL;
+    }
+    // printModelInfo();
+    return COREML_HSUCCEED;
+}
+
+
+CoreMLAdapter CoreMLAdapter::readNetFrom(const std::string& modelPath) {
+    CoreMLAdapter net;
+    NSString *modelPathStr = [NSString stringWithUTF8String:modelPath.c_str()];
+    NSURL *modelURL = [NSURL fileURLWithPath:modelPathStr];
+    NSError *error = nil;
+    net.pImpl->impl.model = [MLModel modelWithContentsOfURL:modelURL error:&error];
+    net.pImpl->impl.modelURL = modelURL;
+    if (error) {
+        NSLog(@"Error loading model: %@", error);
+    }
+    return net;
+}
+
+
+CoreMLAdapter CoreMLAdapter::readNetFromBin(const std::vector<char>& model_data) {
+    // Currently, reading binary files directly is not supported!
+    CoreMLAdapter net;
+    NSLog(@"readNetFromBin not fully implemented");
+    return net;
+}
+
+std::vector<std::string> CoreMLAdapter::getInputNames() const {
+    std::vector<std::string> names;
+    for (NSString *name in pImpl->impl.model.modelDescription.inputDescriptionsByName.allKeys) {
+        names.push_back([name UTF8String]);
+    }
+    return names;
+}
+
+
+std::vector<std::string> CoreMLAdapter::getOutputNames() const {
+    std::vector<std::string> names;
+    for (NSString *name in pImpl->impl.model.modelDescription.outputDescriptionsByName.allKeys) {
+        names.push_back([name UTF8String]);
+    }
+    return names;
+}
+
+std::vector<int> CoreMLAdapter::getInputShapeByName(const std::string &name) {
+    NSString *nsName = [NSString stringWithUTF8String:name.c_str()];
+    MLFeatureDescription *desc = pImpl->impl.model.modelDescription.inputDescriptionsByName[nsName];
+    
+    if (desc.type != MLFeatureTypeMultiArray) {
+        return {};
+    }
+    std::vector<int> shape;
+    for (NSNumber *dim in desc.multiArrayConstraint.shape) {
+        shape.push_back(dim.intValue);
+    }
+    return shape;
+}
+
+std::vector<int> CoreMLAdapter::getOutputShapeByName(const std::string &name) {
+    if (m_outputShapes.find(name) != m_outputShapes.end()) {
+        return m_outputShapes[name];
+    }
+
+    return {};
+}
+
+
+void CoreMLAdapter::setInput(const char* inputName, const char* data) {
+    NSString *nsInputName = [NSString stringWithUTF8String:inputName];
+    MLFeatureDescription *desc = pImpl->impl.model.modelDescription.inputDescriptionsByName[nsInputName];
+    if (desc.type != MLFeatureTypeMultiArray) {
+        NSLog(@"Input %s is not a MultiArray", inputName);
+        return;
+    }
+    
+    NSArray<NSNumber *> *shape = desc.multiArrayConstraint.shape;
+    NSMutableArray<NSNumber *> *strides = [NSMutableArray arrayWithCapacity:shape.count];
+    int stride = 1;
+    for (NSInteger i = shape.count - 1; i >= 0; i--) {
+        [strides insertObject:@(stride) atIndex:0];
+        stride *= [shape[i] intValue];
+    }
+    
+    NSError *error = nil;
+    MLMultiArray *multiArray = [[MLMultiArray alloc] initWithDataPointer:(void*)data
+                                                                   shape:shape
+                                                                dataType:MLMultiArrayDataTypeFloat32
+                                                                 strides:strides
+                                                             deallocator:nil
+                                                                   error:&error];
+    if (error) {
+        NSLog(@"Error creating MLMultiArray: %@", error);
+        return;
+    }
+    
+    MLFeatureValue *featureValue = [MLFeatureValue featureValueWithMultiArray:multiArray];
+    pImpl->impl.inputFeatures = [[MLDictionaryFeatureProvider alloc] initWithDictionary:@{nsInputName: featureValue} error:&error];
+    if (error) {
+        NSLog(@"Error creating MLDictionaryFeatureProvider: %@", error);
+    }
+}
+
+
+int32_t CoreMLAdapter::forward() {
+    @autoreleasepool {
+        if (!pImpl->impl.inputFeatures) {
+            NSLog(@"Input features not set");
+            return COREML_FORWARD_FAILED;
+        }
+        NSError *error = nil;
+        pImpl->impl.outputFeatures = [pImpl->impl.model predictionFromFeatures:pImpl->impl.inputFeatures
+                                                                    options:[[MLPredictionOptions alloc] init]
+                                                                        error:&error];
+        if (error) {
+            NSLog(@"Error in forward pass: %@", error);
+            return COREML_FORWARD_FAILED;
+        }
+
+        m_outputShapes.clear();
+        for (NSString *outputName in [pImpl->impl.outputFeatures featureNames]) {
+            MLFeatureValue *value = [pImpl->impl.outputFeatures featureValueForName:outputName];
+            if (value.multiArrayValue) {
+                NSArray<NSNumber *> *shapeArray = value.multiArrayValue.shape;
+                std::vector<int> shapeVector;
+                
+                for (NSNumber *dim in shapeArray) {
+                    shapeVector.push_back([dim intValue]);
+                }
+                
+                std::string outputNameStr = [outputName UTF8String];
+                m_outputShapes[outputNameStr] = shapeVector;
+            }
+        }
+        return COREML_HSUCCEED;
+    }
+}
+
+
+const char* CoreMLAdapter::getOutput(const char* nodeName) {
+    @autoreleasepool {
+        if (!pImpl->impl.outputFeatures) {
+            NSLog(@"No output features available. Did you call forward()?");
+            return nullptr;
+        }
+        NSString *nsNodeName = [NSString stringWithUTF8String:nodeName];
+        MLFeatureValue *featureValue = [pImpl->impl.outputFeatures featureValueForName:nsNodeName];
+        if (featureValue.type != MLFeatureTypeMultiArray) {
+            NSLog(@"Error getting output for node '%s': Output is not a MultiArray. Current type: %@", nodeName, @(featureValue.type));
+            return nullptr;
+        }
+        return (const char*)featureValue.multiArrayValue.dataPointer;
+    }
+}
+
+
+void CoreMLAdapter::setInferenceMode(InferenceMode mode) {
+    MLComputeUnits computeUnits;
+    switch (mode) {
+        case InferenceMode::CPU:
+            computeUnits = MLComputeUnitsCPUOnly;
+            break;
+        case InferenceMode::GPU:
+            computeUnits = MLComputeUnitsCPUAndGPU;
+            break;
+        case InferenceMode::ANE:
+            computeUnits = MLComputeUnitsAll;
+            break;
+    }
+    
+    MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
+    config.computeUnits = computeUnits;
+    
+    NSError *error = nil;
+    pImpl->impl.model = [MLModel modelWithContentsOfURL:pImpl->impl.modelURL
+                                          configuration:config
+                                                  error:&error];
+    if (error) {
+        NSLog(@"Error setting inference mode: %@", error);
+    }
+}
+
+
+void CoreMLAdapter::printModelInfo() const {
+    NSLog(@"Model Input Description:");
+    for (NSString *inputName in pImpl->impl.model.modelDescription.inputDescriptionsByName) {
+        MLFeatureDescription *desc = pImpl->impl.model.modelDescription.inputDescriptionsByName[inputName];
+        NSLog(@"Input Name: %@", inputName);
+        NSLog(@"Input Type: %@", @(desc.type));
+        
+        switch (desc.type) {
+            case MLFeatureTypeMultiArray:
+                NSLog(@"MLFeatureTypeMultiArray");
+                NSLog(@"  Shape: %@", desc.multiArrayConstraint.shape);
+                NSLog(@"  Data Type: %@", @(desc.multiArrayConstraint.dataType));
+                break;
+            case MLFeatureTypeImage:
+                NSLog(@"MLFeatureTypeImage");
+                NSLog(@"  Image Size: %dx%d", (int)desc.imageConstraint.pixelsWide, (int)desc.imageConstraint.pixelsHigh);
+                NSLog(@"  Color Space: %@", desc.imageConstraint.pixelFormatType == kCVPixelFormatType_32BGRA ? @"BGRA" : @"Other");
+                break;
+            case MLFeatureTypeDictionary:
+                NSLog(@"  Dictionary Key Type: %@", @(desc.dictionaryConstraint.keyType));
+                break;
+            case MLFeatureTypeSequence:
+                NSLog(@"  Sequence Constraint: %@", desc.sequenceConstraint);
+                break;
+            default:
+                NSLog(@"  Unknown type details");
+                break;
+        }
+    }
+    
+    NSLog(@"Model Output Description:");
+    for (NSString *outputName in pImpl->impl.model.modelDescription.outputDescriptionsByName) {
+        MLFeatureDescription *desc = pImpl->impl.model.modelDescription.outputDescriptionsByName[outputName];
+        NSLog(@"Output Name: %@", outputName);
+        NSLog(@"Output Type: %@", @(desc.type));
+        
+        switch (desc.type) {
+            case MLFeatureTypeMultiArray:
+                NSLog(@"  Shape: %@", desc.multiArrayConstraint.shape);
+                NSLog(@"  Data Type: %@", @(desc.multiArrayConstraint.dataType));
+                break;
+            case MLFeatureTypeImage:
+                NSLog(@"  Image Size: %dx%d", (int)desc.imageConstraint.pixelsWide, (int)desc.imageConstraint.pixelsHigh);
+                NSLog(@"  Color Space: %@", desc.imageConstraint.pixelFormatType == kCVPixelFormatType_32BGRA ? @"BGRA" : @"Other");
+                break;
+            case MLFeatureTypeDictionary:
+                NSLog(@"  Dictionary Key Type: %@", @(desc.dictionaryConstraint.keyType));
+                break;
+            case MLFeatureTypeSequence:
+                NSLog(@"  Sequence Constraint: %@", desc.sequenceConstraint);
+                break;
+            default:
+                NSLog(@"  Unknown type details");
+                break;
+        }
+    }
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter.h
new file mode 100644
index 000000000..844e9f40b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter.h
@@ -0,0 +1,498 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_RKNN_ADAPTER_RKNPU1_H
+#define INSPIREFACE_RKNN_ADAPTER_RKNPU1_H
+
+#include <iostream>
+#include "rknn_api.h"
+#include "data_type.h"
+#include "log.h"
+
+/**
+ * @brief Function to get RKNN data type string.
+ * @param type Data type
+ * @return String representation of the data type
+ */
+inline const char *get_type_string_(rknn_tensor_type type) {
+    switch (type) {
+        case RKNN_TENSOR_FLOAT32:
+            return "FP32";
+        case RKNN_TENSOR_FLOAT16:
+            return "FP16";
+        case RKNN_TENSOR_INT8:
+            return "INT8";
+        case RKNN_TENSOR_UINT8:
+            return "UINT8";
+        case RKNN_TENSOR_INT16:
+            return "INT16";
+        default:
+            return "UNKNOW";
+    }
+}
+
+/**
+ * @brief Function to get quantization type string.
+ * @param type Quantization type
+ * @return String representation of the quantization type
+ */
+inline const char *get_qnt_type_string_(rknn_tensor_qnt_type type) {
+    switch (type) {
+        case RKNN_TENSOR_QNT_NONE:
+            return "NONE";
+        case RKNN_TENSOR_QNT_DFP:
+            return "DFP";
+        case RKNN_TENSOR_QNT_AFFINE_ASYMMETRIC:
+            return "AFFINE";
+        default:
+            return "UNKNOW";
+    }
+}
+
+/**
+ * @brief Function to print tensor attributes.
+ * @param attr Tensor attributes
+ */
+inline void print_tensor_attr_(const rknn_tensor_attr &attr) {
+    printf("  n_dims:%d \n", attr.n_dims);
+    printf("  [ ");
+    for (int i = 0; i < attr.n_dims; i++) {
+        printf(" %d ", attr.dims[i]);
+    }
+    printf("] \n");
+    printf("  size:%d \n", attr.size);
+    printf("  n_elems:%d \n", attr.n_elems);
+    printf("  scale:%f \n", attr.scale);
+    printf("  name:%s \n", attr.name);
+}
+
+/**
+ * @brief Function to load data from a file.
+ * @param fp File pointer
+ * @param ofst Offset
+ * @param sz Size
+ * @return Pointer to loaded data
+ */
+inline unsigned char *load_data_(FILE *fp, size_t ofst, size_t sz) {
+    unsigned char *data;
+    int ret;
+
+    data = NULL;
+
+    if (NULL == fp) {
+        return NULL;
+    }
+
+    ret = fseek(fp, ofst, SEEK_SET);
+    if (ret != 0) {
+        printf("blob seek failure.\n");
+        return NULL;
+    }
+
+    data = (unsigned char *)malloc(sz);
+    if (data == NULL) {
+        printf("buffer malloc failure.\n");
+        return NULL;
+    }
+    ret = fread(data, 1, sz, fp);
+    return data;
+}
+
+/**
+ * @brief Function to load a model from a file.
+ * @param filename Model file name
+ * @param model_size Pointer to store model size
+ * @return Pointer to loaded model data
+ */
+inline unsigned char *load_model_(const char *filename, int *model_size) {
+    FILE *fp;
+    unsigned char *data;
+
+    fp = fopen(filename, "rb");
+    if (NULL == fp) {
+        printf("Open file %s failed.\n", filename);
+        return NULL;
+    }
+
+    fseek(fp, 0, SEEK_END);
+    int size = ftell(fp);
+
+    data = load_data_(fp, 0, size);
+
+    fclose(fp);
+
+    *model_size = size;
+    return data;
+}
+
+/**
+ * @brief Status of RKNN inference execution
+ * @ingroup NeuralNetwork
+ */
+enum Status {
+    SUCCESS = 0,            ///< Executed successfully
+    ERROR_SHAPE_MATCH = 1,  ///< Execution error. tensor shapes don't match
+    ERROR_DATA_ORDER = 2    ///< Execution error, tensor data sorting error
+};
+
+/**
+ * @brief RKNN Neural Network Inference Adapter
+ * @details Customizable general-purpose RKNN inference adapter wrapper class, suitable for use in tasks requiring neural network inference with RKNN
+ * @ingroup NeuralNetwork
+ */
+class RKNNAdapter {
+public:
+    RKNNAdapter(const RKNNAdapter &) = delete;
+    RKNNAdapter &operator=(const RKNNAdapter &) = delete;
+    RKNNAdapter() = default;
+
+    /**
+     * @brief Manually initialize
+     * @details Initialize the RKNN model and allocate memory for creating the inference engine session
+     * @param model_path Path to the RKNN model
+     * @return Initialization result
+     */
+    int Initialize(const char *model_path) {
+        /* Create the neural network */
+        int model_data_size = 0;
+        model_data = load_model_(model_path, &model_data_size);
+        load_ = true;
+        int ret = rknn_init(&rk_ctx_, model_data, model_data_size, 0);
+        //        INSPIRE_LOG_INFO("RKNN Init ok.");
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_init fail! ret=%d", ret);
+            return -1;
+        }
+        run_ = true;
+
+        return init_();
+    }
+
+    /**
+     * @brief Manually initialize
+     * @details Initialize the RKNN model using model data and its size, and allocate memory for creating the inference engine session
+     * @param model_data Pointer to the model data
+     * @param model_size Size of the model data
+     * @return Initialization result
+     */
+    int Initialize(const unsigned char *model_data, const unsigned int model_size) {
+        /* Create the neural network */
+        INSPIRE_LOGD("The neural network is being initialized...");
+        int ret = rknn_init(&rk_ctx_, (void *)model_data, model_size, 0);
+
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_init fail! ret=%d", ret);
+            return -1;
+        }
+        run_ = true;
+
+        return init_();
+    }
+
+    /**
+     * @brief Get the size of the input image Tensor
+     * @param index Index of the input layer
+     * @return Dimensions information composed of various sizes
+     */
+    std::vector<int> GetInputTensorSize(const int &index) {
+        std::vector<int> dims(input_attrs_[index].dims, input_attrs_[index].dims + input_attrs_[index].n_dims);
+        return dims;
+    }
+
+    /**
+     * @brief Get the size of the output image Tensor
+     * @param index Index of the output layer
+     * @return Dimensions information composed of various sizes
+     */
+    std::vector<unsigned long> GetOutputTensorSize(const int &index) {
+        //        std::cout << "output_attrs_[index].n_dims:" << output_attrs_[index].n_dims << std::endl;
+        std::vector<unsigned long> dims(output_attrs_[index].dims, output_attrs_[index].dims + output_attrs_[index].n_dims);
+        return dims;
+    }
+
+    /**
+     * @brief Get the length of the output Tensor
+     * @param index Index of the output layer
+     * @return Length information
+     */
+    int GetOutputTensorLen(const int &index) {
+        std::vector<unsigned long> tensor_size_out = GetOutputTensorSize(index);
+        int size = 1;
+        for (auto &one : tensor_size_out)
+            size *= one;
+        return size;
+    }
+
+    /**
+     * @brief Set the data stream for the input layer
+     * @param index Index of the input layer
+     * @param data Pointer to the input data
+     * @param width Width of the input data
+     * @param height Height of the input data
+     * @param channels Number of channels in the input data
+     * @param type Type of the input data (default: RKNN_TENSOR_UINT8)
+     * @param format Format of the input data (default: RKNN_TENSOR_NHWC)
+     * @return Input status
+     */
+    Status SetInputData(const int index, void *data, int width, int height, int channels, rknn_tensor_type type = RKNN_TENSOR_UINT8,
+                        rknn_tensor_format format = RKNN_TENSOR_NHWC) {
+        if (index < input_tensors_.size()) {
+            input_tensors_[index].index = 0;
+            input_tensors_[index].type = type;
+            input_tensors_[index].size = width * height * channels;
+            input_tensors_[index].fmt = format;
+            input_tensors_[index].buf = data;
+            input_tensors_[index].pass_through = 0;
+        } else {
+            INSPIRE_LOGE("error: assert index < len");
+        }
+        return SUCCESS;
+    }
+
+    /**
+     * @brief Execute neural network inference
+     * @details This step should be executed after input data has been provided to the input layer. This step is time-consuming.
+     * @return Inference status result
+     */
+    int RunModel() {
+        //        INSPIRE_LOGD("set input");
+        int ret = rknn_inputs_set(rk_ctx_, rk_io_num_.n_input, input_tensors_.data());
+        if (ret < 0)
+            INSPIRE_LOGE("rknn_input fail! ret=%d", ret);
+
+        for (int i = 0; i < rk_io_num_.n_output; i++) {
+            output_tensors_[i].want_float = outputs_want_float_;
+        }
+
+        //        INSPIRE_LOGD("rknn_run");
+        ret = rknn_run(rk_ctx_, nullptr);
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_run fail! ret=%d", ret);
+            return -1;
+        }
+
+        ret = rknn_outputs_get(rk_ctx_, rk_io_num_.n_output, output_tensors_.data(), NULL);
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_init fail! ret=%d", ret);
+            exit(0);
+        }
+        return ret;
+    }
+
+    /**
+     * @brief Get the data from the output layer
+     * @param index Index of the output layer
+     * @return Pointer to the output data
+     */
+    const float *GetOutputData(const int index) {
+        return (float *)(output_tensors_[index].buf);
+    }
+
+    /**
+     * @brief Get the output data buffer for a specific output layer
+     * @param index Index of the output layer
+     * @return Pointer to the output data buffer
+     */
+    void *GetOutputFlow(const int index) {
+        return output_tensors_[index].buf;
+    }
+
+    /**
+     * @brief Get the output layer data (UINT8)
+     * @details Returns the output layer UInt8 format data after inference, which can only be obtained by inference first
+     * @param index Output level index
+     * @return Returns a pointer to the output data
+     */
+    u_int8_t *GetOutputDataU8(const int index) {
+        return (uint8_t *)(output_tensors_[index].buf);
+    }
+
+    int32_t ReleaseOutputs() {
+        auto ret = rknn_outputs_release(rk_ctx_, rk_io_num_.n_output, output_tensors_.data());
+        return ret;
+    }
+
+    /**
+     * @brief Resize the input tensor with a specified name to a new shape
+     * @details This function is currently not implemented.
+     * @param index_name Name of the input tensor
+     * @param shape New shape for the input tensor
+     */
+    void ResizeInputTensor(const std::string &index_name, const std::vector<int> &shape) {
+        // No implementation
+    }
+
+    /**
+     * @brief Check the size (Placeholder)
+     * @details This function is currently not implemented.
+     */
+    void CheckSize() {
+        // No implementation
+    }
+
+    /**
+     * @brief Get the number of output layers
+     * @details This function is typically used in multi-task or multi-output neural networks.
+     * @return The number of output layers.
+     */
+    size_t GetOutputsNum() const {
+        return rk_io_num_.n_output;
+    }
+
+    /**
+     * @brief Get a reference to the vector of output tensors
+     * @return A reference to the vector of output tensors.
+     */
+    std::vector<rknn_output> &GetOutputTensors() {
+        return output_tensors_;
+    }
+
+    /**
+     * @brief Get a reference to a vector containing information about the attributes of the output tensors
+     * @details The attributes include output sizes, data types, and other relevant information.
+     * @return A reference to the vector containing output tensor attributes.
+     */
+    std::vector<rknn_tensor_attr> &GetOutputTensorAttr() {
+        return output_attrs_;
+    }
+
+    ~RKNNAdapter() {
+        Release();
+    }
+
+    /**
+     * @brief Release resources
+     * @details Release all resources in memory, typically called in the destructor
+     */
+    void Release() {
+        if (run_) {
+            rknn_destroy(rk_ctx_);
+            if (load_) {
+                free(model_data);
+            }
+            run_ = false;
+        }
+    }
+
+    /**
+     * @brief Set the output mode to support float format
+     * @details Depending on the encoding style, some post-processing may use UInt8 format for encoding and decoding
+     * @param outputsWantFloat 0 or 1
+     */
+    void setOutputsWantFloat(int outputsWantFloat) {
+        outputs_want_float_ = outputsWantFloat;
+    }
+
+private:
+    /**
+     * initial
+     * @return
+     */
+    int init_() {
+        rknn_sdk_version version;
+        int ret = rknn_query(rk_ctx_, RKNN_QUERY_SDK_VERSION, &version, sizeof(rknn_sdk_version));
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_init fail! ret=%d", ret);
+            return -1;
+        }
+        INSPIRE_LOGD("sdk version: %s driver version: %s", version.api_version, version.drv_version);
+
+        ret = rknn_query(rk_ctx_, RKNN_QUERY_IN_OUT_NUM, &rk_io_num_, sizeof(rk_io_num_));
+
+        if (ret != RKNN_SUCC) {
+            INSPIRE_LOGE("rknn_query ctx fail! ret=%d", ret);
+            return -1;
+        }
+
+        INSPIRE_LOGD("models input num: %d, output num: %d", rk_io_num_.n_input, rk_io_num_.n_output);
+
+        //        spdlog::trace("input tensors: ");
+        input_attrs_.resize(rk_io_num_.n_input);
+        output_attrs_.resize(rk_io_num_.n_output);
+        input_tensors_.resize(rk_io_num_.n_input);
+        output_tensors_.resize(rk_io_num_.n_output);
+
+        for (int i = 0; i < rk_io_num_.n_input; ++i) {
+            memset(&input_attrs_[i], 0, sizeof(input_attrs_[i]));
+            memset(&input_tensors_[i], 0, sizeof(input_tensors_[i]));
+            input_attrs_[i].index = i;
+            ret = rknn_query(rk_ctx_, RKNN_QUERY_INPUT_ATTR, &(input_attrs_[i]), sizeof(rknn_tensor_attr));
+
+            INSPIRE_LOGD("input node index %d", i);
+            int channel = 3;
+            int width = 0;
+            int height = 0;
+            if (input_attrs_[i].fmt == RKNN_TENSOR_NCHW) {
+                INSPIRE_LOGD("models is NCHW input fmt");
+                width = input_attrs_[i].dims[0];
+                height = input_attrs_[i].dims[1];
+            } else {
+                INSPIRE_LOGD("models is NHWC input fmt");
+                width = input_attrs_[i].dims[1];
+                height = input_attrs_[i].dims[2];
+            }
+            INSPIRE_LOGD("models input height=%d, width=%d, channel=%d", height, width, channel);
+            //            print_tensor_attr_(input_attrs_);
+            if (ret != RKNN_SUCC) {
+                INSPIRE_LOGE("rknn_query fail! ret=%d", ret);
+                return -1;
+            }
+        }
+
+        //        printf("[debug]models input num: %d, output num: %d\n", rk_io_num_.n_input, rk_io_num_.n_output);
+        for (int i = 0; i < rk_io_num_.n_output; ++i) {
+            memset(&output_attrs_[i], 0, sizeof(output_attrs_[i]));
+            memset(&output_tensors_[i], 0, sizeof(output_tensors_[i]));
+            output_attrs_[i].index = i;
+            ret = rknn_query(rk_ctx_, RKNN_QUERY_OUTPUT_ATTR, &(output_attrs_[i]), sizeof(rknn_tensor_attr));
+
+            if (output_attrs_[i].qnt_type != RKNN_TENSOR_QNT_AFFINE_ASYMMETRIC || output_attrs_[i].type != RKNN_TENSOR_UINT8) {
+                INSPIRE_LOGW(
+                  "The Demo required for a Affine asymmetric u8 quantized rknn models, but output quant type is %s, output "
+                  "data type is %s",
+                  get_qnt_type_string_(output_attrs_[i].qnt_type), get_type_string_(output_attrs_[i].type));
+                //                return -1;
+            }
+            //            print_tensor_attr_(output_attrs_[i]);
+
+            //            rknn_tensor_attr rknn_attr;
+            //            memset(&rknn_attr, 0, sizeof(rknn_tensor_attr));
+            //
+            //            ret = rknn_query(rk_ctx_, RKNN_QUERY_OUTPUT_ATTR, &rknn_attr,
+            //                             sizeof(rknn_tensor_attr));
+            //            printf("output node index %d \n", i);
+            //            print_tensor_attr_(rknn_attr);
+
+            if (ret != RKNN_SUCC) {
+                INSPIRE_LOGE("rknn_query fail! ret=%d", ret);
+                return -1;
+            }
+        }
+
+        return ret;
+    }
+
+private:
+    rknn_context rk_ctx_;              ///< The context manager for RKNN.
+    rknn_input_output_num rk_io_num_;  ///< The number of input and output streams in RKNN.
+
+    std::vector<rknn_tensor_attr> input_attrs_;   ///< Attributes of input tensors.
+    std::vector<rknn_tensor_attr> output_attrs_;  ///< Attributes of output tensors.
+    std::vector<rknn_input> input_tensors_;       ///< Input data for the neural network.
+    std::vector<rknn_output> output_tensors_;     ///< Output data from the neural network.
+
+    int outputs_want_float_ = 0;  ///< Flag to indicate support for floating-point output.
+
+    std::vector<int> tensor_shape_;  ///< The shape of input tensors.
+    int width_;                      ///< The width of input data (typically for images).
+    int height_;                     ///< The height of input data (typically for images).
+    bool run_status_;                ///< Flag to indicate the execution status of the neural network.
+
+    unsigned char *model_data;  ///< Pointer to the model's data stream.
+    bool load_;
+    bool run_;
+};
+
+#endif  // INSPIREFACE_RKNN_ADAPTER_RKNPU1_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter_nano.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter_nano.h
new file mode 100644
index 000000000..2d5b67639
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/customized/rknn_adapter_nano.h
@@ -0,0 +1,312 @@
+//
+// Created by Tunm-Air13 on 2023/11/3.
+//
+
+#ifndef SLEEPMONITORING_RKNN_ADAPTER_NANO_H
+#define SLEEPMONITORING_RKNN_ADAPTER_NANO_H
+
+#include "log.h"
+#include "memory"
+#include <float.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/time.h>
+#include "rknn_api.h"
+#include <algorithm>
+#include <iostream>
+#include <cmath>
+#include <vector>
+#include "log.h"
+
+inline std::vector<float> softmax(const std::vector<float> &input) {
+    std::vector<float> output;
+    output.reserve(input.size());
+
+    float max = *std::max_element(input.begin(), input.end());
+    float sum = 0.0;
+
+    for (float val : input) {
+        sum += std::exp(val - max);
+    }
+
+    for (float val : input) {
+        output.push_back(std::exp(val - max) / sum);
+    }
+
+    return output;
+}
+
+inline int NC1HWC2_int8_to_NCHW_float(const int8_t *src, float *dst, int *dims, int channel, int h, int w, int zp, float scale) {
+    int batch = dims[0];
+    int C1 = dims[1];
+    int C2 = dims[4];
+    int hw_src = dims[2] * dims[3];
+    int hw_dst = h * w;
+    for (int i = 0; i < batch; i++) {
+        src = src + i * C1 * hw_src * C2;
+        dst = dst + i * channel * hw_dst;
+        for (int c = 0; c < channel; ++c) {
+            int plane = c / C2;
+            const int8_t *src_c = plane * hw_src * C2 + src;
+            int offset = c % C2;
+            for (int cur_h = 0; cur_h < h; ++cur_h)
+                for (int cur_w = 0; cur_w < w; ++cur_w) {
+                    int cur_hw = cur_h * w + cur_w;
+                    dst[c * hw_dst + cur_h * w + cur_w] = (src_c[C2 * cur_hw + offset] - zp) * scale;  // int8-->float
+                }
+        }
+    }
+
+    return 0;
+}
+
+static void dump_tensor_attr(rknn_tensor_attr *attr) {
+    char dims[128] = {0};
+    for (int i = 0; i < attr->n_dims; ++i) {
+        int idx = strlen(dims);
+        sprintf(&dims[idx], "%d%s", attr->dims[i], (i == attr->n_dims - 1) ? "" : ", ");
+    }
+    INSPIRE_LOGD(
+      "  index=%d, name=%s, n_dims=%d, dims=[%s], n_elems=%d, size=%d, fmt=%s, type=%s, qnt_type=%s, "
+      "zp=%d, scale=%f",
+      attr->index, attr->name, attr->n_dims, dims, attr->n_elems, attr->size, get_format_string(attr->fmt), get_type_string(attr->type),
+      get_qnt_type_string(attr->qnt_type), attr->zp, attr->scale);
+}
+
+class RKNNAdapterNano {
+public:
+    RKNNAdapterNano(const RKNNAdapterNano &) = delete;
+    RKNNAdapterNano &operator=(const RKNNAdapterNano &) = delete;
+    RKNNAdapterNano() = default;
+
+    int32_t Initialize(void *model_data, unsigned int model_size) {
+        int ret = rknn_init(&m_rk_ctx_, model_data, model_size, 0, NULL);
+        if (ret < 0) {
+            INSPIRE_LOGE("rknn_init fail! ret = %d", ret);
+            return -1;
+        }
+
+        // Get sdk and driver version
+        rknn_sdk_version sdk_ver;
+        ret = rknn_query(m_rk_ctx_, RKNN_QUERY_SDK_VERSION, &sdk_ver, sizeof(sdk_ver));
+        if (ret != RKNN_SUCC) {
+            INSPIRE_LOGE("rknn_query fail! ret = %d", ret);
+            return -1;
+        }
+        INSPIRE_LOGD("rknn_api/rknnrt version: %s, driver version: %s", sdk_ver.api_version, sdk_ver.drv_version);
+
+        // Get Model Input Output Info
+        ret = rknn_query(m_rk_ctx_, RKNN_QUERY_IN_OUT_NUM, &m_rk_io_num_, sizeof(m_rk_io_num_));
+        if (ret != RKNN_SUCC) {
+            INSPIRE_LOGE("rknn_query fail! ret = %d", ret);
+            return -1;
+        }
+        INSPIRE_LOGD("model input num: %d, output num: %d", m_rk_io_num_.n_input, m_rk_io_num_.n_output);
+
+        INSPIRE_LOGD("input tensors:");
+        m_input_attrs_.resize(m_rk_io_num_.n_input);
+        for (uint32_t i = 0; i < m_rk_io_num_.n_input; i++) {
+            memset(&m_input_attrs_[i], 0, sizeof(m_input_attrs_[i]));
+            m_input_attrs_[i].index = i;
+            // query info
+            ret = rknn_query(m_rk_ctx_, RKNN_QUERY_INPUT_ATTR, &(m_input_attrs_[i]), sizeof(rknn_tensor_attr));
+            if (ret < 0) {
+                INSPIRE_LOGE("rknn_init error! ret = %d", ret);
+                return -1;
+            }
+            dump_tensor_attr(&m_input_attrs_[i]);
+        }
+
+        INSPIRE_LOGD("output tensors:");
+        m_output_attrs_.resize(m_rk_io_num_.n_output);
+        for (uint32_t i = 0; i < m_rk_io_num_.n_output; i++) {
+            memset(&m_output_attrs_[i], 0, sizeof(m_output_attrs_[i]));
+            m_output_attrs_[i].index = i;
+            // query info
+            ret = rknn_query(m_rk_ctx_, RKNN_QUERY_NATIVE_NHWC_OUTPUT_ATTR, &(m_output_attrs_[i]), sizeof(rknn_tensor_attr));
+            if (ret != RKNN_SUCC) {
+                INSPIRE_LOGE("rknn_query fail! ret = %d", ret);
+                return -1;
+            }
+            dump_tensor_attr(&m_output_attrs_[i]);
+        }
+
+        // Get custom string
+        rknn_custom_string custom_string;
+        ret = rknn_query(m_rk_ctx_, RKNN_QUERY_CUSTOM_STRING, &custom_string, sizeof(custom_string));
+        if (ret != RKNN_SUCC) {
+            INSPIRE_LOGE("rknn_query fail! ret = %d", ret);
+            return -1;
+        }
+        INSPIRE_LOGD("custom string: %s", custom_string.string);
+
+        // Create input tensor memory
+        m_input_mems_.resize(m_rk_io_num_.n_input);
+        for (int i = 0; i < m_rk_io_num_.n_input; ++i) {
+            m_input_mems_[i] = rknn_create_mem(m_rk_ctx_, m_input_attrs_[i].size_with_stride);
+        }
+
+        // Create output tensor memory
+        m_output_mems_.resize(m_rk_io_num_.n_output);
+        for (int i = 0; i < m_rk_io_num_.n_output; ++i) {
+            m_output_mems_[i] = rknn_create_mem(m_rk_ctx_, m_output_attrs_[i].size_with_stride);
+        }
+
+        INSPIRE_LOGD("output origin tensors:");
+        m_orig_output_attrs_.resize(m_rk_io_num_.n_output);
+        for (uint32_t i = 0; i < m_rk_io_num_.n_output; i++) {
+            memset(&m_orig_output_attrs_[i], 0, sizeof(m_orig_output_attrs_[i]));
+            m_orig_output_attrs_[i].index = i;
+            // query info
+            ret = rknn_query(m_rk_ctx_, RKNN_QUERY_OUTPUT_ATTR, &(m_orig_output_attrs_[i]), sizeof(rknn_tensor_attr));
+            if (ret != RKNN_SUCC) {
+                INSPIRE_LOGE("rknn_query fail! ret = %d", ret);
+                return -1;
+            }
+            dump_tensor_attr(&m_orig_output_attrs_[i]);
+        }
+
+        run_ = true;
+        return 0;
+    }
+
+    int32_t SetInputData(const int index, uint8_t *data, rknn_tensor_type type = RKNN_TENSOR_UINT8, rknn_tensor_format format = RKNN_TENSOR_NHWC) {
+        if (index < m_input_mems_.size()) {
+            m_input_attrs_[index].type = type;
+            m_input_attrs_[index].fmt = format;
+            // Copy input data to input tensor memory
+            int width = m_input_attrs_[index].dims[2];
+            int stride = m_input_attrs_[index].w_stride;
+            if (width == stride) {
+                memcpy(m_input_mems_[index]->virt_addr, data, width * m_input_attrs_[index].dims[1] * m_input_attrs_[index].dims[3]);
+            } else {
+                int height = m_input_attrs_[index].dims[1];
+                int channel = m_input_attrs_[index].dims[3];
+                // copy from src to dst with stride
+                uint8_t *src_ptr = data;
+                uint8_t *dst_ptr = (uint8_t *)m_input_mems_[index]->virt_addr;
+                // width-channel elements
+                int src_wc_elems = width * channel;
+                int dst_wc_elems = stride * channel;
+                for (int h = 0; h < height; ++h) {
+                    memcpy(dst_ptr, src_ptr, src_wc_elems);
+                    src_ptr += src_wc_elems;
+                    dst_ptr += dst_wc_elems;
+                }
+            }
+            // Set input tensor memory
+            auto ret = rknn_set_io_mem(m_rk_ctx_, m_input_mems_[index], &m_input_attrs_[index]);
+            if (ret < 0) {
+                INSPIRE_LOGE("rknn_set_io_mem fail! ret = %d", ret);
+                return -1;
+            }
+        } else {
+            INSPIRE_LOGE("error: assert index < input size");
+        }
+
+        return 0;
+    }
+
+    int32_t RunSession(bool use_raw_output = false) {
+        // Set output tensor memory
+        for (uint32_t i = 0; i < m_rk_io_num_.n_output; ++i) {
+            // set output memory and attribute
+            auto ret = rknn_set_io_mem(m_rk_ctx_, m_output_mems_[i], &m_output_attrs_[i]);
+            if (ret < 0) {
+                INSPIRE_LOGE("rknn_set_io_mem fail! ret = %d", ret);
+                return -1;
+            }
+        }
+
+        auto ret = rknn_run(m_rk_ctx_, NULL);
+        if (ret < 0) {
+            printf("rknn run error %d\n", ret);
+            return -1;
+        }
+
+        if (use_raw_output) {
+            m_output_nchw_.resize(m_rk_io_num_.n_output);
+            for (uint32_t i = 0; i < m_rk_io_num_.n_output; ++i) {
+                int num_elements = m_orig_output_attrs_[i].size_with_stride;
+                m_output_nchw_[i].resize(num_elements);
+            }
+
+            for (uint32_t i = 0; i < m_rk_io_num_.n_output; i++) {
+                if (m_output_attrs_[i].fmt == RKNN_TENSOR_NC1HWC2) {
+                    int channel = m_orig_output_attrs_[i].dims[1];
+                    int h = m_orig_output_attrs_[i].n_dims > 2 ? m_orig_output_attrs_[i].dims[2] : 1;
+                    int w = m_orig_output_attrs_[i].n_dims > 3 ? m_orig_output_attrs_[i].dims[3] : 1;
+                    int zp = m_output_attrs_[i].zp;
+                    float scale = m_output_attrs_[i].scale;
+                    NC1HWC2_int8_to_NCHW_float((int8_t *)m_output_mems_[i]->virt_addr, m_output_nchw_[i].data(), (int *)m_output_attrs_[i].dims,
+                                               channel, h, w, zp, scale);
+                } else {
+                    int8_t *src = (int8_t *)m_output_mems_[i]->virt_addr;
+                    float *dst = m_output_nchw_[i].data();
+                    for (int index = 0; index < m_output_attrs_[i].n_elems; index++) {
+                        dst[index] = (src[index] - m_output_attrs_[i].zp) * m_output_attrs_[i].scale;
+                    }
+                }
+            }
+        }
+
+        return 0;
+    }
+
+    std::vector<float> &GetOutputData(size_t index) {
+        return m_output_nchw_[index];
+    }
+
+    rknn_tensor_mem *GetOutputRawData(size_t index) {
+        return m_output_mems_[index];
+    }
+
+    std::vector<rknn_tensor_attr> &GetOutputAttrs() {
+        return m_output_attrs_;
+    }
+
+    const float *GetOutputDataPtr(const int index) {
+        return (float *)(m_output_nchw_[index].data());
+    }
+
+    std::vector<unsigned long> GetOutputTensorSize(const int &index) {
+        std::vector<unsigned long> dims(m_output_attrs_[index].dims, m_output_attrs_[index].dims + m_output_attrs_[index].n_dims);
+        return dims;
+    }
+
+    ~RKNNAdapterNano() {
+        Release();
+    }
+
+    void Release() {
+        if (run_) {
+            for (uint32_t i = 0; i < m_rk_io_num_.n_input; ++i) {
+                rknn_destroy_mem(m_rk_ctx_, m_input_mems_[i]);
+            }
+            for (uint32_t i = 0; i < m_rk_io_num_.n_output; ++i) {
+                rknn_destroy_mem(m_rk_ctx_, m_output_mems_[i]);
+            }
+            if (m_rk_ctx_) {
+                rknn_destroy(m_rk_ctx_);
+            }
+        }
+        run_ = false;
+    }
+
+private:
+    rknn_context m_rk_ctx_;
+
+    rknn_input_output_num m_rk_io_num_;
+    std::vector<rknn_tensor_attr> m_input_attrs_;
+    std::vector<rknn_tensor_attr> m_output_attrs_;
+    std::vector<rknn_tensor_attr> m_orig_output_attrs_;
+
+    std::vector<rknn_tensor_mem *> m_input_mems_;
+    std::vector<rknn_tensor_mem *> m_output_mems_;
+
+    std::vector<std::vector<float>> m_output_nchw_;
+    bool run_;
+};
+
+#endif  // SLEEPMONITORING_RKNN_ADAPTER_NANO_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.cpp
new file mode 100644
index 000000000..b0721999a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.cpp
@@ -0,0 +1,191 @@
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include "inference_wrapper_log.h"
+#include "inference_wrapper.h"
+
+#ifdef INFERENCE_WRAPPER_ENABLE_MNN
+#include "inference_wrapper_mnn.h"
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+// #include "inference_wrapper_rknn.h"
+#include "inference_wrapper_rknn_adapter.h"
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+#include "inference_wrapper_rknn_adapter_nano.h"
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_COREML
+#include "inference_wrapper_coreml.h"
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_TENSORRT
+#include "inference_wrapper_tensorrt.h"
+#endif
+
+#define TAG "InferenceWrapper"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+InferenceWrapper* InferenceWrapper::Create(const InferenceWrapper::EngineType helper_type) {
+    InferenceWrapper* p = nullptr;
+    switch (helper_type) {
+#ifdef INFERENCE_WRAPPER_ENABLE_MNN
+        case INFER_MNN:
+            //        PRINT("Use General Inference\n");
+            p = new InferenceWrapperMNN();
+            break;
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+        case INFER_RKNN:
+            //        PRINT("Use Rknn\n")
+            //        p = new InferenceWrapperRKNN();
+            p = new InferenceWrapperRKNNAdapter();
+            break;
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+        case INFER_RKNN:
+            // PRINT("Use Rknn2\n");
+            p = new InferenceWrapperRKNNAdapter();
+            break;
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_COREML
+        case INFER_COREML:
+            // PRINT("Use CoreML\n");
+            p = new InferenceWrapperCoreML();
+            break;
+#endif
+#ifdef INFERENCE_WRAPPER_ENABLE_TENSORRT
+        case INFER_TENSORRT:
+            // PRINT("Use TensorRT\n");
+            p = new InferenceWrapperTensorRT();
+            break;
+#endif
+        default:
+            PRINT_E("Unsupported inference helper type (%d)\n", helper_type)
+            break;
+    }
+    if (p == nullptr) {
+        PRINT_E("Failed to create inference helper\n")
+    } else {
+        p->helper_type_ = helper_type;
+    }
+    return p;
+}
+
+void InferenceWrapper::ConvertNormalizeParameters(InputTensorInfo& tensor_info) {
+    if (tensor_info.data_type != InputTensorInfo::DataTypeImage)
+        return;
+
+    /* Convert to speeden up normalization:  ((src / 255) - mean) / norm = (src  - (mean * 255))  * (1 / (255 * norm)) */
+    for (int32_t i = 0; i < 3; i++) {
+        tensor_info.normalize.mean[i] *= 255.0f;
+        tensor_info.normalize.norm[i] *= 255.0f;
+        tensor_info.normalize.norm[i] = 1.0f / tensor_info.normalize.norm[i];
+    }
+}
+
+void InferenceWrapper::PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, float* dst) {
+    const int32_t img_width = input_tensor_info.GetWidth();
+    const int32_t img_height = input_tensor_info.GetHeight();
+    const int32_t img_channel = input_tensor_info.GetChannel();
+    uint8_t* src = (uint8_t*)(input_tensor_info.data);
+    if (input_tensor_info.is_nchw == true) {
+        /* convert NHWC to NCHW */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t c = 0; c < img_channel; c++) {
+            for (int32_t i = 0; i < img_width * img_height; i++) {
+                dst[c * img_width * img_height + i] =
+                  (src[i * img_channel + c] - input_tensor_info.normalize.mean[c]) * input_tensor_info.normalize.norm[c];
+            }
+        }
+    } else {
+        /* convert NHWC to NHWC */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t i = 0; i < img_width * img_height; i++) {
+            for (int32_t c = 0; c < img_channel; c++) {
+                dst[i * img_channel + c] = (src[i * img_channel + c] - input_tensor_info.normalize.mean[c]) * input_tensor_info.normalize.norm[c];
+            }
+        }
+    }
+}
+
+void InferenceWrapper::PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, uint8_t* dst) {
+    const int32_t img_width = input_tensor_info.GetWidth();
+    const int32_t img_height = input_tensor_info.GetHeight();
+    const int32_t img_channel = input_tensor_info.GetChannel();
+    uint8_t* src = (uint8_t*)(input_tensor_info.data);
+    if (input_tensor_info.is_nchw == true) {
+        /* convert NHWC to NCHW */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t c = 0; c < img_channel; c++) {
+            for (int32_t i = 0; i < img_width * img_height; i++) {
+                dst[c * img_width * img_height + i] = src[i * img_channel + c];
+            }
+        }
+    } else {
+        /* convert NHWC to NHWC */
+        std::copy(src, src + input_tensor_info.GetElementNum(), dst);
+    }
+}
+
+void InferenceWrapper::PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, int8_t* dst) {
+    const int32_t img_width = input_tensor_info.GetWidth();
+    const int32_t img_height = input_tensor_info.GetHeight();
+    const int32_t img_channel = input_tensor_info.GetChannel();
+    uint8_t* src = (uint8_t*)(input_tensor_info.data);
+    if (input_tensor_info.is_nchw == true) {
+        /* convert NHWC to NCHW */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t c = 0; c < img_channel; c++) {
+            for (int32_t i = 0; i < img_width * img_height; i++) {
+                dst[c * img_width * img_height + i] = src[i * img_channel + c] - 128;
+            }
+        }
+    } else {
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t i = 0; i < img_width * img_height; i++) {
+            for (int32_t c = 0; c < img_channel; c++) {
+                dst[i * img_channel + c] = src[i * img_channel + c] - 128;
+            }
+        }
+    }
+}
+
+template <typename T>
+void InferenceWrapper::PreProcessBlob(int32_t num_thread, const InputTensorInfo& input_tensor_info, T* dst) {
+    const int32_t img_width = input_tensor_info.GetWidth();
+    const int32_t img_height = input_tensor_info.GetHeight();
+    const int32_t img_channel = input_tensor_info.GetChannel();
+    T* src = static_cast<T*>(input_tensor_info.data);
+    if ((input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNchw && input_tensor_info.is_nchw) ||
+        (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc && !input_tensor_info.is_nchw)) {
+        std::copy(src, src + input_tensor_info.GetElementNum(), dst);
+    } else if (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNchw) {
+        /* NCHW -> NHWC */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t i = 0; i < img_width * img_height; i++) {
+            for (int32_t c = 0; c < img_channel; c++) {
+                dst[i * img_channel + c] = src[c * (img_width * img_height) + i];
+            }
+        }
+    } else if (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) {
+        /* NHWC -> NCHW */
+#pragma omp parallel for num_threads(num_thread)
+        for (int32_t i = 0; i < img_width * img_height; i++) {
+            for (int32_t c = 0; c < img_channel; c++) {
+                dst[c * (img_width * img_height) + i] = src[i * img_channel + c];
+            }
+        }
+    }
+}
+
+template void InferenceWrapper::PreProcessBlob<float>(int32_t num_thread, const InputTensorInfo& input_tensor_info, float* dst);
+template void InferenceWrapper::PreProcessBlob<int32_t>(int32_t num_thread, const InputTensorInfo& input_tensor_info, int32_t* dst);
+template void InferenceWrapper::PreProcessBlob<int64_t>(int32_t num_thread, const InputTensorInfo& input_tensor_info, int64_t* dst);
+template void InferenceWrapper::PreProcessBlob<uint8_t>(int32_t num_thread, const InputTensorInfo& input_tensor_info, uint8_t* dst);
+template void InferenceWrapper::PreProcessBlob<int8_t>(int32_t num_thread, const InputTensorInfo& input_tensor_info, int8_t* dst);
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.h
new file mode 100644
index 000000000..3196243a3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper.h
@@ -0,0 +1,263 @@
+#ifndef INFERENCE_WRAPPER_
+#define INFERENCE_WRAPPER_
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+
+class TensorInfo {
+public:
+    enum {
+        TensorTypeNone,
+        TensorTypeUint8,
+        TensorTypeInt8,
+        TensorTypeFp32,
+        TensorTypeInt32,
+        TensorTypeInt64,
+    };
+
+public:
+    TensorInfo() : name(""), id(-1), tensor_type(TensorTypeNone), is_nchw(true) {}
+    ~TensorInfo() {}
+
+    int32_t GetElementNum() const {
+        int32_t element_num = 1;
+        for (const auto& dim : tensor_dims) {
+            element_num *= dim;
+        }
+        return element_num;
+    }
+
+    int32_t GetBatch() const {
+        if (tensor_dims.size() <= 0)
+            return -1;
+        return tensor_dims[0];
+    }
+
+    int32_t GetChannel() const {
+        if (is_nchw) {
+            if (tensor_dims.size() <= 1)
+                return -1;
+            return tensor_dims[1];
+        } else {
+            if (tensor_dims.size() <= 3)
+                return -1;
+            return tensor_dims[3];
+        }
+    }
+
+    int32_t GetHeight() const {
+        if (is_nchw) {
+            if (tensor_dims.size() <= 2)
+                return -1;
+            return tensor_dims[2];
+        } else {
+            if (tensor_dims.size() <= 1)
+                return -1;
+            return tensor_dims[1];
+        }
+    }
+
+    int32_t GetWidth() const {
+        if (is_nchw) {
+            if (tensor_dims.size() <= 3)
+                return -1;
+            return tensor_dims[3];
+        } else {
+            if (tensor_dims.size() <= 2)
+                return -1;
+            return tensor_dims[2];
+        }
+    }
+
+public:
+    std::string name;
+    int32_t id;
+    int32_t tensor_type;
+    std::vector<int32_t> tensor_dims;
+    bool is_nchw;
+};
+
+class InputTensorInfo : public TensorInfo {
+public:
+    enum {
+        DataTypeImage,
+        DataTypeBlobNhwc,  // data_ which already finished preprocess(color conversion, resize, normalize_, etc.)
+        DataTypeBlobNchw,
+    };
+
+public:
+    InputTensorInfo()
+    : data(nullptr),
+      data_type(DataTypeImage),
+      image_info({-1, -1, -1, -1, -1, -1, -1, true, false}),
+      normalize({0.0f, 0.0f, 0.0f, 1.0f, 1.0f, 1.0f}) {}
+
+    InputTensorInfo(std::string name_, int32_t tensor_type_, bool is_nchw_ = true) : InputTensorInfo() {
+        name = name_;
+        tensor_type = tensor_type_;
+        is_nchw = is_nchw_;
+    }
+
+    ~InputTensorInfo() {}
+
+public:
+    void* data;
+    int32_t data_type;
+
+    struct {
+        int32_t width;
+        int32_t height;
+        int32_t channel;
+        int32_t crop_x;
+        int32_t crop_y;
+        int32_t crop_width;
+        int32_t crop_height;
+        bool is_bgr;  // used when channel == 3 (true: BGR, false: RGB)
+        bool swap_color;
+    } image_info;
+
+    struct {
+        float mean[3];
+        float norm[3];
+    } normalize;
+};
+
+class OutputTensorInfo : public TensorInfo {
+public:
+    OutputTensorInfo() : data(nullptr), quant({1.0f, 0}), data_fp32_(nullptr) {}
+
+    OutputTensorInfo(std::string name_, int32_t tensor_type_, bool is_nchw_ = true) : OutputTensorInfo() {
+        name = name_;
+        tensor_type = tensor_type_;
+        is_nchw = is_nchw_;
+    }
+
+    ~OutputTensorInfo() {
+        if (data_fp32_ != nullptr) {
+            delete[] data_fp32_;
+        }
+    }
+
+    float* GetDataAsFloat() {
+        if (tensor_type == TensorTypeUint8 || tensor_type == TensorTypeInt8) {
+            if (data_fp32_ == nullptr) {
+                data_fp32_ = new float[GetElementNum()];
+            }
+            if (tensor_type == TensorTypeUint8) {
+#pragma omp parallel
+                for (int32_t i = 0; i < GetElementNum(); i++) {
+                    const uint8_t* val_uint8 = static_cast<const uint8_t*>(data);
+                    float val_float = (val_uint8[i] - quant.zero_point) * quant.scale;
+                    data_fp32_[i] = val_float;
+                }
+            } else {
+#pragma omp parallel
+                for (int32_t i = 0; i < GetElementNum(); i++) {
+                    const int8_t* val_int8 = static_cast<const int8_t*>(data);
+                    float val_float = (val_int8[i] - quant.zero_point) * quant.scale;
+                    data_fp32_[i] = val_float;
+                }
+            }
+            return data_fp32_;
+        } else if (tensor_type == TensorTypeFp32) {
+            return static_cast<float*>(data);
+        } else {
+            return nullptr;
+        }
+    }
+
+public:
+    void* data;
+    struct {
+        float scale;
+        int32_t zero_point;
+    } quant;
+
+private:
+    float* data_fp32_;
+};
+
+namespace cv {
+class Mat;
+};
+
+class InferenceWrapper {
+public:
+    enum {
+        WrapperOk = 0,
+        WrapperError = -1,
+    };
+
+    typedef enum {
+        DEFAULT_CPU,
+        MMM_CUDA,
+        COREML_CPU,
+        COREML_GPU,
+        COREML_ANE,
+        TENSORRT_CUDA,
+    } SpecialBackend;
+
+    typedef enum {
+        INFER_MNN,
+        INFER_RKNN,
+        INFER_COREML,
+        INFER_TENSORRT,
+    } EngineType;
+
+public:
+    static InferenceWrapper* Create(const EngineType helper_type);
+
+public:
+    virtual ~InferenceWrapper() {}
+    virtual int32_t SetNumThreads(const int32_t num_threads) = 0;
+    virtual int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                               std::vector<OutputTensorInfo>& output_tensor_info_list) = 0;
+    virtual int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                               std::vector<OutputTensorInfo>& output_tensor_info_list) = 0;
+    virtual int32_t Finalize(void) = 0;
+    virtual int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) = 0;
+    virtual int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) = 0;
+    virtual int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                            std::vector<OutputTensorInfo>& output_tensor_info_list) = 0;
+    
+#ifdef BATCH_FORWARD_IMPLEMENTED
+    virtual int32_t PreProcessBatch(const std::vector<std::vector<InputTensorInfo>>& input_tensor_info_list) = 0;
+    virtual int32_t ProcessBatch(std::vector<std::vector<OutputTensorInfo>>& output_tensor_info_list) = 0;
+    virtual int32_t PostProcessBatch(std::vector<std::vector<OutputTensorInfo>>& output_tensor_info_list) = 0;
+#endif
+
+    virtual int32_t SetSpecialBackend(SpecialBackend backend) {
+        special_backend_ = backend;
+        return WrapperOk;
+    };
+
+    virtual int32_t SetDevice(int32_t device_id) {
+        device_id_ = device_id;
+        return WrapperOk;
+    };
+
+    virtual int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) = 0;
+
+    virtual std::vector<std::string> GetInputNames() = 0;
+
+protected:
+    void ConvertNormalizeParameters(InputTensorInfo& tensor_info);
+
+    void PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, float* dst);
+    void PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, uint8_t* dst);
+    void PreProcessImage(int32_t num_thread, const InputTensorInfo& input_tensor_info, int8_t* dst);
+
+    template <typename T>
+    void PreProcessBlob(int32_t num_thread, const InputTensorInfo& input_tensor_info, T* dst);
+
+protected:
+    EngineType helper_type_;
+    SpecialBackend special_backend_ = DEFAULT_CPU;
+    int32_t device_id_ = 0;
+};
+
+#endif
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.cpp
new file mode 100644
index 000000000..71e16aea0
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.cpp
@@ -0,0 +1,215 @@
+#ifdef INFERENCE_WRAPPER_ENABLE_COREML
+
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include <MNN/ImageProcess.hpp>
+#include "inference_wrapper_log.h"
+#include "inference_wrapper_coreml.h"
+#include "log.h"
+#define TAG "InferenceWrapperCoreML"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+using namespace inspire;
+
+InferenceWrapperCoreML::InferenceWrapperCoreML() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperCoreML::~InferenceWrapperCoreML() {}
+
+int32_t InferenceWrapperCoreML::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperCoreML::ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                                        std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperCoreML::Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                           std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    PRINT_E("CoreML does not yet support buffer initialization of the model\n");
+    return WrapperError;
+}
+
+int32_t InferenceWrapperCoreML::Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                           std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    //    LOG_INFO("init MNN");
+    /*** Create network ***/
+    net_.reset(new CoreMLAdapter());
+    auto ret = net_->readFromFile(model_filename);
+    if (ret != WrapperOk) {
+        PRINT_E("Failed to load model file (%s)\n", model_filename.c_str());
+        return WrapperError;
+    }
+    if (special_backend_ == COREML_CPU || special_backend_ == DEFAULT_CPU) {
+        net_->setInferenceMode(CoreMLAdapter::InferenceMode::CPU);
+    } else if (special_backend_ == COREML_GPU) {
+        net_->setInferenceMode(CoreMLAdapter::InferenceMode::GPU);
+    } else if (special_backend_ == COREML_ANE) {
+        net_->setInferenceMode(CoreMLAdapter::InferenceMode::ANE);
+    } else {
+        PRINT_E("Unsupported backend (%d)\n, defaulting to ANE", special_backend_);
+        net_->setInferenceMode(CoreMLAdapter::InferenceMode::ANE);
+    }
+    if (!net_) {
+        PRINT_E("Failed to load model file (%s)\n", model_filename.c_str());
+        return WrapperError;
+    }
+
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+};
+
+int32_t InferenceWrapperCoreML::Finalize(void) {
+    net_.reset();
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperCoreML::PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    // Currently only single-input models are supported
+    for (const auto& input_tensor_info : input_tensor_info_list) {
+        input_tensor_.reset(MNN::Tensor::create<float>(
+          std::vector<int>{1, 3, input_tensor_info.image_info.height, input_tensor_info.image_info.width}, nullptr, MNN::Tensor::CAFFE));
+        if (input_tensor_ == nullptr) {
+            PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            INSPIRE_LOGE("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        if (input_tensor_info.data_type == InputTensorInfo::DataTypeImage) {
+            /* Crop */
+            if ((input_tensor_info.image_info.width != input_tensor_info.image_info.crop_width) ||
+                (input_tensor_info.image_info.height != input_tensor_info.image_info.crop_height)) {
+                PRINT_E("Crop is not supported\n");
+                return WrapperError;
+            }
+
+            MNN::CV::ImageProcess::Config image_processconfig;
+            /* Convert color type */
+            //            LOGD("input_tensor_info.image_info.channel: %d", input_tensor_info.image_info.channel);
+            //            LOGD("input_tensor_info.GetChannel(): %d", input_tensor_info.GetChannel());
+
+            // !!!!!! BUG !!!!!!!!!
+            // When initializing, setting the image channel to 3 and the tensor channel to 1,
+            // and configuring the processing to convert the color image to grayscale may cause some bugs.
+            // For example, the image channel might automatically change to 1.
+            // This issue has not been fully investigated,
+            // so it's necessary to manually convert the image to grayscale before input.
+            // !!!!!! BUG !!!!!!!!!
+
+            if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                if (input_tensor_info.image_info.swap_color) {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::RGB : MNN::CV::BGR;
+                } else {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                }
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+            } else if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+                //                LOGD("2gray");
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::BGR;
+            } else {
+                PRINT_E("Unsupported color conversion (%d, %d)\n", input_tensor_info.image_info.channel, input_tensor_info.GetChannel());
+                return WrapperError;
+            }
+
+            /* Normalize image */
+            std::memcpy(image_processconfig.mean, input_tensor_info.normalize.mean, sizeof(image_processconfig.mean));
+            std::memcpy(image_processconfig.normal, input_tensor_info.normalize.norm, sizeof(image_processconfig.normal));
+
+            /* Resize image */
+            image_processconfig.filterType = MNN::CV::BILINEAR;
+            MNN::CV::Matrix trans;
+            trans.setScale(static_cast<float>(input_tensor_info.image_info.crop_width) / input_tensor_info.GetWidth(),
+                           static_cast<float>(input_tensor_info.image_info.crop_height) / input_tensor_info.GetHeight());
+
+            /* Do pre-process */
+            std::shared_ptr<MNN::CV::ImageProcess> pretreat(MNN::CV::ImageProcess::create(image_processconfig));
+            pretreat->setMatrix(trans);
+            //            LOGD("k1");
+            pretreat->convert(static_cast<uint8_t*>(input_tensor_info.data), input_tensor_info.image_info.crop_width,
+                              input_tensor_info.image_info.crop_height, 0, input_tensor_.get());
+
+        } else if ((input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) ||
+                   (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNchw)) {
+            std::unique_ptr<MNN::Tensor> tensor;
+            if (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) {
+                tensor.reset(new MNN::Tensor(input_tensor_.get(), MNN::Tensor::TENSORFLOW));
+            } else {
+                tensor.reset(new MNN::Tensor(input_tensor_.get(), MNN::Tensor::CAFFE));
+            }
+            if (tensor->getType().code == halide_type_float) {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<float>()[i] = static_cast<float*>(input_tensor_info.data)[i];
+                }
+            } else {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<uint8_t>()[i] = static_cast<uint8_t*>(input_tensor_info.data)[i];
+                }
+            }
+            input_tensor_->copyFromHostTensor(tensor.get());
+        } else {
+            PRINT_E("Unsupported data type (%d)\n", input_tensor_info.data_type);
+            return WrapperError;
+        }
+        auto p = input_tensor_->host<float>();
+        net_->setInput(input_tensor_info.name.c_str(), reinterpret_cast<const char*>(p));
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperCoreML::Process(std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    auto ret = net_->forward();
+    if (ret != COREML_HSUCCEED) {
+        PRINT_E("Failed to forward\n");
+        return WrapperError;
+    }
+
+    // out_mat_list_.clear();
+    for (auto& output_tensor_info : output_tensor_info_list) {
+        // auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
+        const char* output_tensor = net_->getOutput(output_tensor_info.name.c_str());
+        if (output_tensor == nullptr) {
+            PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        output_tensor_info.data = (void*)output_tensor;
+
+        output_tensor_info.tensor_dims.clear();
+
+        const std::vector<int>& output_shape = net_->getOutputShapeByName(output_tensor_info.name);
+
+        int size = 1;
+        for (int32_t dim = 0; dim < output_shape.size(); dim++) {
+            output_tensor_info.tensor_dims.push_back(output_shape[dim]);
+            size *= output_shape[dim];
+        }
+    }
+
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperCoreML::GetInputNames() {
+    return input_names_;
+}
+
+int32_t InferenceWrapperCoreML::ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    PRINT_E("Currently, CoreML does not support input resizing\n");
+    return 0;
+}
+
+#endif  // INFERENCE_WRAPPER_ENABLE_COREML
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.h
new file mode 100644
index 000000000..d6f6ad5c7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_coreml.h
@@ -0,0 +1,43 @@
+#ifndef INFERENCE_WRAPPER_COREML_
+#define INFERENCE_WRAPPER_COREML_
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include "coreml/CoreMLAdapter.h"
+#include <MNN/ImageProcess.hpp>
+#include "inference_wrapper.h"
+
+class InferenceWrapperCoreML : public InferenceWrapper {
+public:
+    InferenceWrapperCoreML();
+    ~InferenceWrapperCoreML() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+    std::vector<std::string> GetInputNames() override;
+
+private:
+    std::unique_ptr<CoreMLAdapter> net_;
+    int32_t num_threads_;
+
+    std::vector<std::string> input_names_;
+
+    /** Using MNN imageprocess to do Image Preprocessing */
+    std::unique_ptr<MNN::Tensor> input_tensor_;
+};
+
+#endif  // INFERENCE_WRAPPER_COREML_
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_log.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_log.h
new file mode 100644
index 000000000..822718b41
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_log.h
@@ -0,0 +1,31 @@
+#ifndef INFERENCE_WRAPPER_LOG_
+#define INFERENCE_WRAPPER_LOG_
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+
+#if defined(ANDROID) || defined(__ANDROID__)
+#define CV_COLOR_IS_RGB
+#include <android/log.h>
+#define INFERENCE_WRAPPER_LOG_NDK_TAG "Inspireface-Native-Inference"
+#define INFERENCE_WRAPPER_LOG_PRINT_(...) __android_log_print(ANDROID_LOG_INFO, INFERENCE_WRAPPER_LOG_NDK_TAG, __VA_ARGS__)
+#else
+#define INFERENCE_WRAPPER_LOG_PRINT_(...) printf(__VA_ARGS__)
+#endif
+
+#define INFERENCE_WRAPPER_LOG_PRINT(INFERENCE_WRAPPER_LOG_PRINT_TAG, ...)                     \
+    do {                                                                                      \
+        INFERENCE_WRAPPER_LOG_PRINT_("[" INFERENCE_WRAPPER_LOG_PRINT_TAG "][%d] ", __LINE__); \
+        INFERENCE_WRAPPER_LOG_PRINT_(__VA_ARGS__);                                            \
+    } while (0);
+
+#define INFERENCE_WRAPPER_LOG_PRINT_E(INFERENCE_WRAPPER_LOG_PRINT_TAG, ...)                        \
+    do {                                                                                           \
+        INFERENCE_WRAPPER_LOG_PRINT_("[ERR: " INFERENCE_WRAPPER_LOG_PRINT_TAG "][%d] ", __LINE__); \
+        INFERENCE_WRAPPER_LOG_PRINT_(__VA_ARGS__);                                                 \
+    } while (0);
+
+#endif
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.cpp
new file mode 100644
index 000000000..9d80d4aa2
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.cpp
@@ -0,0 +1,320 @@
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include <MNN/ImageProcess.hpp>
+#include <MNN/Interpreter.hpp>
+#include <MNN/AutoTime.hpp>
+#include "inference_wrapper_log.h"
+#include "inference_wrapper_mnn.h"
+#include "log.h"
+#define TAG "InferenceWrapperMNN"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+using namespace inspire;
+
+InferenceWrapperMNN::InferenceWrapperMNN() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperMNN::~InferenceWrapperMNN() {}
+
+int32_t InferenceWrapperMNN::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperMNN::ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                                     std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    /* Check tensor info fits the info from model */
+    for (auto& input_tensor_info : input_tensor_info_list) {
+        auto input_tensor = net_->getSessionInput(session_, input_tensor_info.name.c_str());
+        if (input_tensor == nullptr) {
+            PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            //            LOGD("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        if ((input_tensor->getType().code == halide_type_float) && (input_tensor_info.tensor_type == TensorInfo::TensorTypeFp32)) {
+            /* OK */
+        } else if ((input_tensor->getType().code == halide_type_uint) && (input_tensor_info.tensor_type == TensorInfo::TensorTypeUint8)) {
+            /* OK */
+        } else {
+            PRINT_E("Incorrect input tensor type (%d, %d)\n", input_tensor->getType().code, input_tensor_info.tensor_type);
+            return WrapperError;
+        }
+        if ((input_tensor->channel() != -1) && (input_tensor->height() != -1) && (input_tensor->width() != -1)) {
+            if (input_tensor_info.GetChannel() != -1) {
+                if ((input_tensor->channel() == input_tensor_info.GetChannel()) && (input_tensor->height() == input_tensor_info.GetHeight()) &&
+                    (input_tensor->width() == input_tensor_info.GetWidth())) {
+                    /* OK */
+                } else {
+                    INSPIRE_LOGW("W: %d != %d", input_tensor->width(), input_tensor_info.GetWidth());
+                    INSPIRE_LOGW("H: %d != %d", input_tensor->height(), input_tensor_info.GetHeight());
+                    INSPIRE_LOGW("C: %d != %d", input_tensor->channel(), input_tensor_info.GetChannel());
+                    INSPIRE_LOGW("There may be some risk of input that is not used by model default");
+                    net_->resizeTensor(input_tensor,
+                                       {1, input_tensor_info.GetChannel(), input_tensor_info.GetHeight(), input_tensor_info.GetWidth()});
+                    net_->resizeSession(session_);
+                    return WrapperOk;
+                }
+            } else {
+                PRINT("Input tensor size is set from the model\n");
+                input_tensor_info.tensor_dims.clear();
+                for (int32_t dim = 0; dim < input_tensor->dimensions(); dim++) {
+                    input_tensor_info.tensor_dims.push_back(input_tensor->length(dim));
+                }
+            }
+        } else {
+            if (input_tensor_info.GetChannel() != -1) {
+                PRINT("Input tensor size is resized\n");
+                /* In case the input size  is not fixed */
+                net_->resizeTensor(input_tensor, {1, input_tensor_info.GetChannel(), input_tensor_info.GetHeight(), input_tensor_info.GetWidth()});
+                net_->resizeSession(session_);
+                INSPIRE_LOGE("GO RESIZE");
+            } else {
+                PRINT_E("Model input size is not set\n");
+                return WrapperError;
+            }
+        }
+    }
+    for (const auto& output_tensor_info : output_tensor_info_list) {
+        auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
+        if (output_tensor == nullptr) {
+            PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        /* Output size is set when run inference later */
+    }
+
+    /* Convert normalize parameter to speed up */
+    for (auto& input_tensor_info : input_tensor_info_list) {
+        ConvertNormalizeParameters(input_tensor_info);
+    }
+
+    /* Check if tensor info is set */
+    for (const auto& input_tensor_info : input_tensor_info_list) {
+        for (const auto& dim : input_tensor_info.tensor_dims) {
+            if (dim <= 0) {
+                PRINT_E("Invalid tensor size\n");
+                return WrapperError;
+            }
+        }
+    }
+
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperMNN::Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                        std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    net_.reset(MNN::Interpreter::createFromBuffer(model_buffer, model_size));
+    if (!net_) {
+        PRINT_E("Failed to load model model buffer\n");
+        return WrapperError;
+    }
+    MNN::ScheduleConfig scheduleConfig;
+    scheduleConfig.numThread = num_threads_;  // it seems, setting 1 has better performance on Android
+    MNN::BackendConfig bnconfig;
+    bnconfig.power = MNN::BackendConfig::Power_High;
+    bnconfig.precision = MNN::BackendConfig::Precision_Normal;
+    if (special_backend_ == MMM_CUDA) {
+        INSPIRE_LOGD("Enable CUDA");
+        scheduleConfig.type = MNN_FORWARD_CUDA;
+        bnconfig.power = MNN::BackendConfig::Power_Normal;
+        bnconfig.precision = MNN::BackendConfig::Precision_Normal;
+    } else {
+        scheduleConfig.type = MNN_FORWARD_CPU;
+    }
+    scheduleConfig.backendConfig = &bnconfig;
+
+    session_ = net_->createSession(scheduleConfig);
+    for (auto& item : net_->getSessionInputAll(session_)) {
+        input_names_.push_back(item.first.c_str());
+    }
+    if (!session_) {
+        PRINT_E("Failed to create session\n");
+        return WrapperError;
+    }
+
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+}
+
+int32_t InferenceWrapperMNN::Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                        std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    net_.reset(MNN::Interpreter::createFromFile(model_filename.c_str()));
+    if (!net_) {
+        PRINT_E("Failed to load model file (%s)\n", model_filename.c_str());
+        return WrapperError;
+    }
+
+    MNN::ScheduleConfig scheduleConfig;
+    scheduleConfig.type = MNN_FORWARD_CPU;
+    scheduleConfig.numThread = num_threads_;  // it seems, setting 1 has better performance on Android
+    // MNN::BackendConfig bnconfig;
+    // bnconfig.power = MNN::BackendConfig::Power_High;
+    // bnconfig.precision = MNN::BackendConfig::Precision_Low;
+    // scheduleConfig.backendConfig = &bnconfig;
+    session_ = net_->createSession(scheduleConfig);
+    if (!session_) {
+        PRINT_E("Failed to create session\n");
+        return WrapperError;
+    }
+
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+};
+
+int32_t InferenceWrapperMNN::Finalize(void) {
+    net_->releaseSession(session_);
+    net_->releaseModel();
+    net_.reset();
+    out_mat_list_.clear();
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperMNN::PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    for (const auto& input_tensor_info : input_tensor_info_list) {
+        auto input_tensor = net_->getSessionInput(session_, input_tensor_info.name.c_str());
+        if (input_tensor == nullptr) {
+            PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            INSPIRE_LOGE("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        if (input_tensor_info.data_type == InputTensorInfo::DataTypeImage) {
+            /* Crop */
+            if ((input_tensor_info.image_info.width != input_tensor_info.image_info.crop_width) ||
+                (input_tensor_info.image_info.height != input_tensor_info.image_info.crop_height)) {
+                PRINT_E("Crop is not supported\n");
+                return WrapperError;
+            }
+
+            MNN::CV::ImageProcess::Config image_processconfig;
+            /* Convert color type */
+            //            LOGD("input_tensor_info.image_info.channel: %d", input_tensor_info.image_info.channel);
+            //            LOGD("input_tensor_info.GetChannel(): %d", input_tensor_info.GetChannel());
+
+            // !!!!!! BUG !!!!!!!!!
+            // When initializing, setting the image channel to 3 and the tensor channel to 1,
+            // and configuring the processing to convert the color image to grayscale may cause some bugs.
+            // For example, the image channel might automatically change to 1.
+            // This issue has not been fully investigated,
+            // so it's necessary to manually convert the image to grayscale before input.
+            // !!!!!! BUG !!!!!!!!!
+
+            if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                if (input_tensor_info.image_info.swap_color) {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::RGB : MNN::CV::BGR;
+                } else {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                }
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+            } else if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+                //                LOGD("2gray");
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::BGR;
+            } else {
+                PRINT_E("Unsupported color conversion (%d, %d)\n", input_tensor_info.image_info.channel, input_tensor_info.GetChannel());
+                return WrapperError;
+            }
+
+            /* Normalize image */
+            std::memcpy(image_processconfig.mean, input_tensor_info.normalize.mean, sizeof(image_processconfig.mean));
+            std::memcpy(image_processconfig.normal, input_tensor_info.normalize.norm, sizeof(image_processconfig.normal));
+
+            /* Resize image */
+            image_processconfig.filterType = MNN::CV::BILINEAR;
+            MNN::CV::Matrix trans;
+            trans.setScale(static_cast<float>(input_tensor_info.image_info.crop_width) / input_tensor_info.GetWidth(),
+                           static_cast<float>(input_tensor_info.image_info.crop_height) / input_tensor_info.GetHeight());
+
+            /* Do pre-process */
+            std::shared_ptr<MNN::CV::ImageProcess> pretreat(MNN::CV::ImageProcess::create(image_processconfig));
+            pretreat->setMatrix(trans);
+            pretreat->convert(static_cast<uint8_t*>(input_tensor_info.data), input_tensor_info.image_info.crop_width,
+                              input_tensor_info.image_info.crop_height, 0, input_tensor);
+
+        } else if ((input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) ||
+                   (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNchw)) {
+            std::unique_ptr<MNN::Tensor> tensor;
+            if (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) {
+                tensor.reset(new MNN::Tensor(input_tensor, MNN::Tensor::TENSORFLOW));
+            } else {
+                tensor.reset(new MNN::Tensor(input_tensor, MNN::Tensor::CAFFE));
+            }
+            if (tensor->getType().code == halide_type_float) {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<float>()[i] = static_cast<float*>(input_tensor_info.data)[i];
+                }
+            } else {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<uint8_t>()[i] = static_cast<uint8_t*>(input_tensor_info.data)[i];
+                }
+            }
+            input_tensor->copyFromHostTensor(tensor.get());
+        } else {
+            PRINT_E("Unsupported data type (%d)\n", input_tensor_info.data_type);
+            return WrapperError;
+        }
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperMNN::Process(std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    net_->runSession(session_);
+
+    out_mat_list_.clear();
+    for (auto& output_tensor_info : output_tensor_info_list) {
+        auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
+        if (output_tensor == nullptr) {
+            PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
+            return WrapperError;
+        }
+
+        auto dimType = output_tensor->getDimensionType();
+        std::unique_ptr<MNN::Tensor> outputUser(new MNN::Tensor(output_tensor, dimType));
+        output_tensor->copyToHostTensor(outputUser.get());
+        auto type = outputUser->getType();
+        if (type.code == halide_type_float) {
+            output_tensor_info.tensor_type = TensorInfo::TensorTypeFp32;
+            output_tensor_info.data = outputUser->host<float>();
+        } else if (type.code == halide_type_uint && type.bytes() == 1) {
+            output_tensor_info.tensor_type = TensorInfo::TensorTypeUint8;
+            output_tensor_info.data = outputUser->host<uint8_t>();
+        } else {
+            PRINT_E("Unexpected data type\n");
+            return WrapperError;
+        }
+
+        output_tensor_info.tensor_dims.clear();
+        for (int32_t dim = 0; dim < outputUser->dimensions(); dim++) {
+            output_tensor_info.tensor_dims.push_back(outputUser->length(dim));
+        }
+
+        out_mat_list_.push_back(std::move(outputUser));  // store data in member variable so that data keep exist
+    }
+
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperMNN::GetInputNames() {
+    return input_names_;
+}
+
+int32_t InferenceWrapperMNN::ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    for (const auto& input_tensor_info : input_tensor_info_list) {
+        auto input_tensor = net_->getSessionInput(session_, input_tensor_info.name.c_str());
+        net_->resizeTensor(input_tensor, {1, input_tensor_info.GetChannel(), input_tensor_info.GetHeight(), input_tensor_info.GetWidth()});
+        net_->resizeSession(session_);
+    }
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.h
new file mode 100644
index 000000000..0e46925bb
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_mnn.h
@@ -0,0 +1,43 @@
+#ifndef INFERENCE_WRAPPER_MNN_
+#define INFERENCE_WRAPPER_MNN_
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include <MNN/ImageProcess.hpp>
+#include <MNN/Interpreter.hpp>
+#include <MNN/AutoTime.hpp>
+#include "inference_wrapper.h"
+
+class InferenceWrapperMNN : public InferenceWrapper {
+public:
+    InferenceWrapperMNN();
+    ~InferenceWrapperMNN() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+    std::vector<std::string> GetInputNames() override;
+
+private:
+    std::unique_ptr<MNN::Interpreter> net_;
+    MNN::Session* session_;
+    std::vector<std::unique_ptr<MNN::Tensor>> out_mat_list_;
+    int32_t num_threads_;
+
+    std::vector<std::string> input_names_;
+};
+
+#endif
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.cpp
new file mode 100644
index 000000000..5834ab2a0
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.cpp
@@ -0,0 +1,246 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+#include <iostream>
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include "inference_wrapper_rknn.h"
+#include "inference_wrapper_log.h"
+#include "log.h"
+
+static unsigned char* load_data(FILE* fp, size_t ofst, size_t sz) {
+    unsigned char* data;
+    int ret;
+
+    data = NULL;
+
+    if (NULL == fp) {
+        return NULL;
+    }
+
+    ret = fseek(fp, ofst, SEEK_SET);
+    if (ret != 0) {
+        printf("blob seek failure.\n");
+        return NULL;
+    }
+
+    data = (unsigned char*)malloc(sz);
+    if (data == NULL) {
+        printf("buffer malloc failure.\n");
+        return NULL;
+    }
+    ret = fread(data, 1, sz, fp);
+    return data;
+}
+
+static unsigned char* load_model(const char* filename, int* model_size) {
+    FILE* fp;
+    unsigned char* data;
+
+    fp = fopen(filename, "rb");
+    if (NULL == fp) {
+        printf("Open file %s failed.\n", filename);
+        return NULL;
+    }
+
+    fseek(fp, 0, SEEK_END);
+    int size = ftell(fp);
+
+    data = load_data(fp, 0, size);
+
+    fclose(fp);
+
+    *model_size = size;
+    return data;
+}
+
+/*** Macro ***/
+#define TAG "InferenceWrapperRKNN"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+InferenceWrapperRKNN::InferenceWrapperRKNN() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperRKNN::~InferenceWrapperRKNN() {}
+
+int32_t InferenceWrapperRKNN::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNN::ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                                      std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    auto ret = rknn_query(net_, RKNN_QUERY_IN_OUT_NUM, &rk_io_num_, sizeof(rk_io_num_));
+
+    if (ret != RKNN_SUCC) {
+        PRINT_E("rknn_query ctx fail! ret=%d\n", ret)
+        return WrapperError;
+    }
+
+    std::vector<rknn_tensor_attr> output_attrs_;
+    output_attrs_.resize(rk_io_num_.n_output);
+    for (int i = 0; i < rk_io_num_.n_output; ++i) {
+        //        memset(&output_attrs_[i], 0, sizeof(output_attrs_[i]));
+        //        memset(&output_tensors_[i], 0, sizeof(output_tensors_[i]));
+        output_attrs_[i].index = i;
+        ret = rknn_query(net_, RKNN_QUERY_OUTPUT_ATTR, &(output_attrs_[i]), sizeof(rknn_tensor_attr));
+        auto& output = output_tensor_info_list[i];
+        output.tensor_dims.clear();
+        for (int j = 0; j < output_attrs_[i].n_dims; ++j) {
+            output.tensor_dims.push_back(output_attrs_[i].dims[j]);
+            std::cout << "dim: " << output_attrs_[i].dims[j] << std::endl;
+        }
+        //        std::cout << output_attrs_[i].n_dims << std::endl;
+    }
+
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNN::Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                         std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    int model_data_size = 0;
+    unsigned char* model_data = load_model(model_filename.c_str(), &model_data_size);
+    int ret = rknn_init(&net_, model_data, model_data_size, 0);
+    if (ret < 0) {
+        PRINT_E("Failed to load model file (%s)\n", model_filename.c_str())
+        return WrapperError;
+    }
+    rknn_sdk_version version;
+    ret = rknn_query(net_, RKNN_QUERY_SDK_VERSION, &version, sizeof(rknn_sdk_version));
+    if (ret < 0) {
+        PRINT_E("rknn_init error ret=%d\n", ret)
+        return WrapperError;
+    }
+
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+}
+
+int32_t InferenceWrapperRKNN::Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                         std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    int ret = rknn_init(&net_, model_buffer, model_size, 0);
+    if (ret < 0) {
+        PRINT_E("rknn_init error ret=%d\n", ret)
+        return WrapperError;
+    }
+    rknn_sdk_version version;
+    ret = rknn_query(net_, RKNN_QUERY_SDK_VERSION, &version, sizeof(rknn_sdk_version));
+    if (ret < 0) {
+        PRINT_E("rknn_init error ret=%d\n", ret)
+        return WrapperError;
+    }
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+}
+
+int32_t InferenceWrapperRKNN::Finalize(void) {
+    rknn_destroy(net_);
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNN::PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    /* Check tensor info fits the info from model */
+    if (input_tensor_info_list.size() != rk_io_num_.n_input) {
+        PRINT_E("The inputs quantity is inconsistent: i: %d, m: %d", input_tensor_info_list.size(), rk_io_num_.n_input);
+        return WrapperError;
+    }
+
+    std::vector<rknn_input> input_tensors_;
+    for (size_t index = 0; index < input_tensor_info_list.size(); index++) {
+        auto& input_tensor_info = input_tensor_info_list[index];
+        if (input_tensor_info.data_type == InputTensorInfo::DataTypeImage) {
+            /* Crop */
+            // Not Implement
+
+            /* Convert color type */
+            // Not Implement
+            rknn_input input;
+            if (input_tensor_info.tensor_type == TensorInfo::TensorTypeUint8) {
+                input.type = RKNN_TENSOR_UINT8;
+            } else if (input_tensor_info.tensor_type == TensorInfo::TensorTypeFp32) {
+                input.type = RKNN_TENSOR_FLOAT32;
+            } else {
+                PRINT_E("Unsupported input type.")
+                return WrapperError;
+            }
+            if (input_tensor_info.is_nchw) {
+                input.fmt = RKNN_TENSOR_NCHW;
+            } else {
+                input.fmt = RKNN_TENSOR_NHWC;
+                std::cout << "NHWC" << std::endl;
+            }
+            input.index = index;
+            input.size = input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel();
+            input.buf = input_tensor_info.data;
+            input.pass_through = 0;
+            input_tensors_.push_back(input);
+        }
+    }
+    //    INSPIRE_LOGD("Prepare data!");
+    int ret = rknn_inputs_set(net_, input_tensor_info_list.size(), input_tensors_.data());
+    //    INSPIRE_LOGD("Set data!");
+    if (ret < 0) {
+        PRINT_E("rknn_run fail! ret=%d", ret)
+        return WrapperError;
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNN::Process(std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    if (output_tensor_info_list.size() != rk_io_num_.n_output) {
+        PRINT_E("The outputs quantity is inconsistent")
+        return WrapperError;
+    }
+    auto ret = rknn_run(net_, NULL);
+    if (ret < 0) {
+        PRINT_E("rknn_run fail! ret=%d", ret)
+        return WrapperError;
+    }
+
+    for (size_t index = 0; index < output_tensor_info_list.size(); index++) {
+        auto& output_tensor = output_tensor_info_list[index];
+        rknn_output output;
+        if (output_tensor.tensor_type == TensorInfo::TensorTypeFp32) {
+            output.want_float = 1;
+            INSPIRE_LOGD("want_float=1");
+        }
+        output.is_prealloc = 0;
+        output_tensors_.push_back(output);
+        //        output.want_float = 1;  // float
+    }
+
+    ret = rknn_outputs_get(net_, output_tensor_info_list.size(), output_tensors_.data(), NULL);
+    if (ret < 0) {
+        PRINT_E("rknn_run fail! ret=%d", ret)
+        return WrapperError;
+    }
+    for (size_t index = 0; index < output_tensor_info_list.size(); index++) {
+        auto& output_tensor = output_tensor_info_list[index];
+        auto& output = output_tensors_[index];
+        output_tensor.data = output.buf;
+        float* outBlob = (float*)output.buf;
+        //        output_tensor.
+    }
+
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperRKNN::GetInputNames() {
+    return input_names_;
+}
+
+int32_t InferenceWrapperRKNN::ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    // The function is not supported
+    return 0;
+}
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.h
new file mode 100644
index 000000000..1fafc1a39
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn.h
@@ -0,0 +1,47 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef RKNPU_PTOCR_INFERENCE_WRAPPER_RKNN_H
+#define RKNPU_PTOCR_INFERENCE_WRAPPER_RKNN_H
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include <rknn_api.h>
+#include "inference_wrapper.h"
+
+class InferenceWrapperRKNN : public InferenceWrapper {
+public:
+    InferenceWrapperRKNN();
+    ~InferenceWrapperRKNN() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    std::vector<std::string> GetInputNames() override;
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+private:
+    rknn_context net_;
+    int32_t num_threads_;
+    std::vector<rknn_output> output_tensors_;
+    rknn_input_output_num rk_io_num_;
+    std::vector<std::string> input_names_;
+
+private:
+};
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN
+#endif  // RKNPU_PTOCR_INFERENCE_WRAPPER_RKNN_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.cpp
new file mode 100644
index 000000000..5cd63bf6f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.cpp
@@ -0,0 +1,138 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+
+#include <iostream>
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include <cassert>
+#include "inference_wrapper_rknn_adapter.h"
+#include "inference_wrapper_log.h"
+#include "log.h"
+#include <cassert>
+
+#define TAG "InferenceWrapperRKNNAdapter"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+InferenceWrapperRKNNAdapter::InferenceWrapperRKNNAdapter() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperRKNNAdapter::~InferenceWrapperRKNNAdapter() {}
+
+int32_t InferenceWrapperRKNNAdapter::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::ParameterInitialization(std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                             std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Process(std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    if (output_tensor_info_list[0].tensor_type == TensorInfo::TensorTypeFp32) {
+        net_->setOutputsWantFloat(1);
+        //        INSPIRE_LOGD("WANT FLOAT!");
+    }
+
+    auto ret = net_->RunModel();
+    if (ret != 0) {
+        INSPIRE_LOGE("Run model error.");
+        return WrapperError;
+    }
+    auto outputs_size = net_->GetOutputsNum();
+
+    assert(outputs_size == output_tensor_info_list.size());
+
+    for (int index = 0; index < outputs_size; ++index) {
+        auto &output_tensor = output_tensor_info_list[index];
+        output_tensor.data = net_->GetOutputFlow(index);
+
+        auto dim = net_->GetOutputTensorSize(index);
+        output_tensor.tensor_dims.clear();
+        for (int i = 0; i < dim.size(); ++i) {
+            output_tensor.tensor_dims.push_back((int)dim[i]);
+            //            INSPIRE_LOGE("dim: %d", dim[i]);
+        }
+    }
+
+    net_->ReleaseOutputs();
+
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::PreProcess(const std::vector<InputTensorInfo> &input_tensor_info_list) {
+    for (int i = 0; i < input_tensor_info_list.size(); ++i) {
+        auto &input_tensor_info = input_tensor_info_list[i];
+
+        rknn_tensor_format fmt = RKNN_TENSOR_NHWC;
+        if (input_tensor_info.is_nchw) {
+            fmt = RKNN_TENSOR_NCHW;
+        } else {
+            fmt = RKNN_TENSOR_NHWC;
+            //            INSPIRE_LOGD("NHWC!");
+        }
+        rknn_tensor_type type = RKNN_TENSOR_UINT8;
+        if (input_tensor_info.tensor_type == InputTensorInfo::TensorInfo::TensorTypeFp32) {
+            type = RKNN_TENSOR_FLOAT32;
+        } else if (input_tensor_info.tensor_type == InputTensorInfo::TensorInfo::TensorTypeUint8) {
+            type = RKNN_TENSOR_UINT8;
+            //            INSPIRE_LOGD("UINT8!");
+        }
+        auto ret = net_->SetInputData(i, input_tensor_info.data, input_tensor_info.GetWidth(), input_tensor_info.GetHeight(),
+                                      input_tensor_info.GetChannel(), type, fmt);
+        if (ret != 0) {
+            INSPIRE_LOGE("Set data error.");
+            return ret;
+        }
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Initialize(const std::string &model_filename, std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    INSPIRE_LOGE("NOT IMPL");
+
+    return 0;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Initialize(char *model_buffer, int model_size, std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    net_ = std::make_shared<RKNNAdapter>();
+    auto ret = net_->Initialize((unsigned char *)model_buffer, model_size);
+    if (ret != 0) {
+        INSPIRE_LOGE("Rknn init error.");
+        return WrapperError;
+    }
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+}
+
+int32_t InferenceWrapperRKNNAdapter::Finalize(void) {
+    if (net_ != nullptr) {
+        net_->Release();
+    }
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperRKNNAdapter::GetInputNames() {
+    return std::vector<std::string>();
+}
+
+int32_t InferenceWrapperRKNNAdapter::ResizeInput(const std::vector<InputTensorInfo> &input_tensor_info_list) {
+    // The function is not supported
+    return 0;
+}
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.h
new file mode 100644
index 000000000..19c1c23ce
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter.h
@@ -0,0 +1,45 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_H
+#define INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_H
+
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include "inference_wrapper.h"
+#include "customized/rknn_adapter.h"
+
+class InferenceWrapperRKNNAdapter : public InferenceWrapper {
+public:
+    InferenceWrapperRKNNAdapter();
+    ~InferenceWrapperRKNNAdapter() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    std::vector<std::string> GetInputNames() override;
+
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+private:
+    std::shared_ptr<RKNNAdapter> net_;
+    int32_t num_threads_;
+};
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN
+
+#endif  // INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.cpp
new file mode 100644
index 000000000..8eee688cc
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.cpp
@@ -0,0 +1,134 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+
+#include <iostream>
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include <cassert>
+#include "inference_wrapper_rknn_adapter_nano.h"
+#include "inference_wrapper_log.h"
+#include "log.h"
+#include <cassert>
+
+#define TAG "InferenceWrapperRKNNAdapter"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+InferenceWrapperRKNNAdapter::InferenceWrapperRKNNAdapter() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperRKNNAdapter::~InferenceWrapperRKNNAdapter() {}
+
+int32_t InferenceWrapperRKNNAdapter::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::ParameterInitialization(std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                             std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Process(std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    if (output_tensor_info_list[0].tensor_type == TensorInfo::TensorTypeFp32) {
+        // net_->setOutputsWantFloat(1);
+        //        INSPIRE_LOGD("WANT FLOAT!");
+    }
+
+    auto ret = net_->RunSession(true);
+    if (ret != 0) {
+        INSPIRE_LOGE("Run model error.");
+        return WrapperError;
+    }
+    auto outputs_size = net_->GetOutputAttrs().size();
+
+    assert(outputs_size == output_tensor_info_list.size());
+
+    for (int index = 0; index < outputs_size; ++index) {
+        auto &output_tensor = output_tensor_info_list[index];
+        output_tensor.data = (void *)net_->GetOutputDataPtr(index);
+
+        auto dim = net_->GetOutputTensorSize(index);
+        output_tensor.tensor_dims.clear();
+        for (int i = 0; i < dim.size(); ++i) {
+            output_tensor.tensor_dims.push_back((int)dim[i]);
+            //            INSPIRE_LOGE("dim: %d", dim[i]);
+        }
+    }
+
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::PreProcess(const std::vector<InputTensorInfo> &input_tensor_info_list) {
+    for (int i = 0; i < input_tensor_info_list.size(); ++i) {
+        auto &input_tensor_info = input_tensor_info_list[i];
+        rknn_tensor_format fmt = RKNN_TENSOR_NHWC;
+        if (input_tensor_info.is_nchw) {
+            fmt = RKNN_TENSOR_NCHW;
+        } else {
+            fmt = RKNN_TENSOR_NHWC;
+            //            INSPIRE_LOGD("NHWC!");
+        }
+        rknn_tensor_type type = RKNN_TENSOR_UINT8;
+        if (input_tensor_info.tensor_type == InputTensorInfo::TensorInfo::TensorTypeFp32) {
+            type = RKNN_TENSOR_FLOAT32;
+        } else if (input_tensor_info.tensor_type == InputTensorInfo::TensorInfo::TensorTypeUint8) {
+            type = RKNN_TENSOR_UINT8;
+            //            INSPIRE_LOGD("UINT8!");
+        }
+        auto ret = net_->SetInputData(i, (uint8_t *)input_tensor_info.data, type, fmt);
+        if (ret != 0) {
+            INSPIRE_LOGE("Set data error.");
+            return ret;
+        }
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Initialize(const std::string &model_filename, std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    INSPIRE_LOGE("NOT IMPL");
+
+    return 0;
+}
+
+int32_t InferenceWrapperRKNNAdapter::Initialize(char *model_buffer, int model_size, std::vector<InputTensorInfo> &input_tensor_info_list,
+                                                std::vector<OutputTensorInfo> &output_tensor_info_list) {
+    net_ = std::make_shared<RKNNAdapterNano>();
+    auto ret = net_->Initialize((unsigned char *)model_buffer, model_size);
+    if (ret != 0) {
+        INSPIRE_LOGE("Rknn init error.");
+        return WrapperError;
+    }
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+}
+
+int32_t InferenceWrapperRKNNAdapter::Finalize(void) {
+    if (net_ != nullptr) {
+        net_->Release();
+    }
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperRKNNAdapter::GetInputNames() {
+    return std::vector<std::string>();
+}
+
+int32_t InferenceWrapperRKNNAdapter::ResizeInput(const std::vector<InputTensorInfo> &input_tensor_info_list) {
+    // The function is not supported
+    return 0;
+}
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN2
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.h
new file mode 100644
index 000000000..dc3557052
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_rknn_adapter_nano.h
@@ -0,0 +1,45 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_NANO_H
+#define INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_NANO_H
+
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include "inference_wrapper.h"
+#include "customized/rknn_adapter_nano.h"
+
+class InferenceWrapperRKNNAdapter : public InferenceWrapper {
+public:
+    InferenceWrapperRKNNAdapter();
+    ~InferenceWrapperRKNNAdapter() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    std::vector<std::string> GetInputNames() override;
+
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+private:
+    std::shared_ptr<RKNNAdapterNano> net_;
+    int32_t num_threads_;
+};
+
+#endif  // INFERENCE_WRAPPER_ENABLE_RKNN2
+
+#endif  // INSPIREFACE_INFERENCE_WRAPPER_RKNN_ADAPTER_NANO_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.cpp
new file mode 100644
index 000000000..ebedd235e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.cpp
@@ -0,0 +1,208 @@
+#if INFERENCE_WRAPPER_ENABLE_TENSORRT
+#include <iostream>
+#include <cstdint>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <chrono>
+#include <MNN/ImageProcess.hpp>
+#include "inference_wrapper_log.h"
+#include "inference_wrapper_tensorrt.h"
+#include "log.h"
+#define TAG "InferenceWrapperTensorRT"
+#define PRINT(...) INFERENCE_WRAPPER_LOG_PRINT(TAG, __VA_ARGS__)
+#define PRINT_E(...) INFERENCE_WRAPPER_LOG_PRINT_E(TAG, __VA_ARGS__)
+
+using namespace inspire;
+
+InferenceWrapperTensorRT::InferenceWrapperTensorRT() {
+    num_threads_ = 1;
+}
+
+InferenceWrapperTensorRT::~InferenceWrapperTensorRT() {}
+
+int32_t InferenceWrapperTensorRT::SetNumThreads(const int32_t num_threads) {
+    num_threads_ = num_threads;
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperTensorRT::ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                                          std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperTensorRT::Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                             std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    net_.reset(new TensorRTAdapter());
+    net_->setDevice(device_id_);
+    auto ret = net_->readFromBin(model_buffer, model_size);
+    if (ret != WrapperOk) {
+        PRINT_E("Failed to load TensorRT model\n");
+        return WrapperError;
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperTensorRT::Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                                             std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    //    LOG_INFO("init MNN");
+    /*** Create network ***/
+    net_.reset(new TensorRTAdapter());
+    net_->setDevice(device_id_);
+    auto ret = net_->readFromFile(model_filename);
+    if (ret != WrapperOk) {
+        PRINT_E("Failed to load model file (%s)\n", model_filename.c_str());
+        return WrapperError;
+    }
+
+    return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
+};
+
+int32_t InferenceWrapperTensorRT::Finalize(void) {
+    net_.reset();
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperTensorRT::PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    // Currently only single-input models are supported
+    for (const auto& input_tensor_info : input_tensor_info_list) {
+        input_tensor_.reset(MNN::Tensor::create<float>(
+          std::vector<int>{1, 3, input_tensor_info.image_info.height, input_tensor_info.image_info.width}, nullptr, MNN::Tensor::CAFFE));
+        if (input_tensor_ == nullptr) {
+            PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            INSPIRE_LOGE("Invalid input name (%s)\n", input_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        if (input_tensor_info.data_type == InputTensorInfo::DataTypeImage) {
+            /* Crop */
+            if ((input_tensor_info.image_info.width != input_tensor_info.image_info.crop_width) ||
+                (input_tensor_info.image_info.height != input_tensor_info.image_info.crop_height)) {
+                PRINT_E("Crop is not supported\n");
+                return WrapperError;
+            }
+
+            MNN::CV::ImageProcess::Config image_processconfig;
+            /* Convert color type */
+            //            LOGD("input_tensor_info.image_info.channel: %d", input_tensor_info.image_info.channel);
+            //            LOGD("input_tensor_info.GetChannel(): %d", input_tensor_info.GetChannel());
+
+            // !!!!!! BUG !!!!!!!!!
+            // When initializing, setting the image channel to 3 and the tensor channel to 1,
+            // and configuring the processing to convert the color image to grayscale may cause some bugs.
+            // For example, the image channel might automatically change to 1.
+            // This issue has not been fully investigated,
+            // so it's necessary to manually convert the image to grayscale before input.
+            // !!!!!! BUG !!!!!!!!!
+
+            if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                if (input_tensor_info.image_info.swap_color) {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::RGB : MNN::CV::BGR;
+                } else {
+                    image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                }
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+            } else if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 1)) {
+                image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
+                image_processconfig.destFormat = MNN::CV::GRAY;
+                //                LOGD("2gray");
+            } else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 3)) {
+                image_processconfig.sourceFormat = MNN::CV::GRAY;
+                image_processconfig.destFormat = MNN::CV::BGR;
+            } else {
+                PRINT_E("Unsupported color conversion (%d, %d)\n", input_tensor_info.image_info.channel, input_tensor_info.GetChannel());
+                return WrapperError;
+            }
+
+            /* Normalize image */
+            std::memcpy(image_processconfig.mean, input_tensor_info.normalize.mean, sizeof(image_processconfig.mean));
+            std::memcpy(image_processconfig.normal, input_tensor_info.normalize.norm, sizeof(image_processconfig.normal));
+
+            /* Resize image */
+            image_processconfig.filterType = MNN::CV::BILINEAR;
+            MNN::CV::Matrix trans;
+            trans.setScale(static_cast<float>(input_tensor_info.image_info.crop_width) / input_tensor_info.GetWidth(),
+                           static_cast<float>(input_tensor_info.image_info.crop_height) / input_tensor_info.GetHeight());
+
+            /* Do pre-process */
+            std::shared_ptr<MNN::CV::ImageProcess> pretreat(MNN::CV::ImageProcess::create(image_processconfig));
+            pretreat->setMatrix(trans);
+            //            LOGD("k1");
+            pretreat->convert(static_cast<uint8_t*>(input_tensor_info.data), input_tensor_info.image_info.crop_width,
+                              input_tensor_info.image_info.crop_height, 0, input_tensor_.get());
+
+        } else if ((input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) ||
+                   (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNchw)) {
+            std::unique_ptr<MNN::Tensor> tensor;
+            if (input_tensor_info.data_type == InputTensorInfo::DataTypeBlobNhwc) {
+                tensor.reset(new MNN::Tensor(input_tensor_.get(), MNN::Tensor::TENSORFLOW));
+            } else {
+                tensor.reset(new MNN::Tensor(input_tensor_.get(), MNN::Tensor::CAFFE));
+            }
+            if (tensor->getType().code == halide_type_float) {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<float>()[i] = static_cast<float*>(input_tensor_info.data)[i];
+                }
+            } else {
+                for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
+                    tensor->host<uint8_t>()[i] = static_cast<uint8_t*>(input_tensor_info.data)[i];
+                }
+            }
+            input_tensor_->copyFromHostTensor(tensor.get());
+        } else {
+            PRINT_E("Unsupported data type (%d)\n", input_tensor_info.data_type);
+            return WrapperError;
+        }
+        auto p = input_tensor_->host<float>();
+        net_->setInput(input_tensor_info.name.c_str(), reinterpret_cast<const char*>(p));
+    }
+    return WrapperOk;
+}
+
+int32_t InferenceWrapperTensorRT::Process(std::vector<OutputTensorInfo>& output_tensor_info_list) {
+    auto ret = net_->forward();
+    if (ret != TENSORRT_HSUCCEED) {
+        PRINT_E("Failed to forward\n");
+        return WrapperError;
+    }
+
+    // out_mat_list_.clear();
+    for (auto& output_tensor_info : output_tensor_info_list) {
+        // auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
+        const void* output_tensor = net_->getOutput(output_tensor_info.name.c_str());
+        if (output_tensor == nullptr) {
+            PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
+            return WrapperError;
+        }
+        output_tensor_info.data = (void*)output_tensor;
+
+        output_tensor_info.tensor_dims.clear();
+
+        const std::vector<int>& output_shape = net_->getOutputShapeByName(output_tensor_info.name);
+
+        int size = 1;
+        for (int32_t dim = 0; dim < output_shape.size(); dim++) {
+            output_tensor_info.tensor_dims.push_back(output_shape[dim]);
+            size *= output_shape[dim];
+        }
+    }
+
+    return WrapperOk;
+}
+
+std::vector<std::string> InferenceWrapperTensorRT::GetInputNames() {
+    return input_names_;
+}
+
+int32_t InferenceWrapperTensorRT::ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) {
+    PRINT_E("Currently, TensorRT does not support input resizing\n");
+    return 0;
+}
+
+#endif  // INFERENCE_WRAPPER_ENABLE_TENSORRT
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.h
new file mode 100644
index 000000000..340ef249f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/inference_wrapper_tensorrt.h
@@ -0,0 +1,44 @@
+#ifndef INFERENCE_WRAPPER_COREML_
+#define INFERENCE_WRAPPER_COREML_
+#if INFERENCE_WRAPPER_ENABLE_TENSORRT
+#include <cstdint>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <array>
+#include <memory>
+#include "tensorrt/tensorrt_adapter.h"
+#include <MNN/ImageProcess.hpp>
+#include "inference_wrapper.h"
+
+class InferenceWrapperTensorRT : public InferenceWrapper {
+public:
+    InferenceWrapperTensorRT();
+    ~InferenceWrapperTensorRT() override;
+    int32_t SetNumThreads(const int32_t num_threads) override;
+    int32_t Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list,
+                       std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t Finalize(void) override;
+    int32_t PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+    int32_t Process(std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+    int32_t ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list,
+                                    std::vector<OutputTensorInfo>& output_tensor_info_list) override;
+
+    int32_t ResizeInput(const std::vector<InputTensorInfo>& input_tensor_info_list) override;
+
+    std::vector<std::string> GetInputNames() override;
+
+private:
+    std::unique_ptr<TensorRTAdapter> net_;
+    int32_t num_threads_;
+
+    std::vector<std::string> input_names_;
+
+    /** Using MNN imageprocess to do Image Preprocessing */
+    std::unique_ptr<MNN::Tensor> input_tensor_;
+};
+
+#endif  // INFERENCE_WRAPPER_COREML_
+#endif  // INFERENCE_WRAPPER_ENABLE_TENSORRT
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.cpp
new file mode 100644
index 000000000..1da8a9ac3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.cpp
@@ -0,0 +1,657 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-16
+ */
+#if ISF_ENABLE_TENSORRT
+#include "tensorrt_adapter.h"
+#include <fstream>
+#include <iostream>
+#include <chrono>
+#include <memory>
+#include <cstring>
+#include <cuda_fp16.h>
+#include <NvInfer.h>
+#include <cuda_runtime_api.h>
+
+#include <log.h>
+#include <isf_check.h>
+
+// define specific deleters for TensorRT objects
+struct TRTRuntimeDeleter {
+    void operator()(nvinfer1::IRuntime *runtime) const {
+        if (runtime)
+            delete runtime;
+    }
+};
+
+struct TRTEngineDeleter {
+    void operator()(nvinfer1::ICudaEngine *engine) const {
+        if (engine)
+            delete engine;
+    }
+};
+
+struct TRTContextDeleter {
+    void operator()(nvinfer1::IExecutionContext *context) const {
+        if (context)
+            delete context;
+    }
+};
+
+// custom deleter for CUDA stream
+struct CUDAStreamDeleter {
+    void operator()(cudaStream_t *stream) const {
+        if (stream) {
+            cudaStreamDestroy(*stream);
+            delete stream;
+        }
+    }
+};
+
+// custom Logger class, inherit from TensorRT's ILogger
+class TRTLogger : public nvinfer1::ILogger {
+public:
+    void log(Severity severity, const char *msg) noexcept override {
+        if (severity <= Severity::kWARNING) {
+            INSPIRE_LOGI("[TensorRT] %s", msg);
+        }
+    }
+};
+
+// read model file to memory
+static std::vector<char> readModelFile(const std::string &filename) {
+    std::ifstream file(filename, std::ios::binary | std::ios::ate);
+    if (!file) {
+        INSPIRE_LOGE("failed to open model file: %s", filename.c_str());
+        return {};
+    }
+
+    size_t size = file.tellg();
+    file.seekg(0, std::ios::beg);
+
+    std::vector<char> buffer(size);
+    if (!file.read(buffer.data(), size)) {
+        INSPIRE_LOGE("failed to read model file: %s", filename.c_str());
+        return {};
+    }
+
+    return buffer;
+}
+
+// CUDA error check macro
+#define CHECK_CUDA(call)                                                \
+    do {                                                                \
+        cudaError_t error = call;                                       \
+        if (error != cudaSuccess) {                                     \
+            INSPIRE_LOGE("[CUDA error] %s", cudaGetErrorString(error)); \
+            return TENSORRT_HFAIL;                                      \
+        }                                                               \
+    } while (0)
+
+// TensorRT adapter implementation class
+class TensorRTAdapter::Impl {
+public:
+    Impl() : m_ownStream(false), m_inferenceMode(TensorRTAdapter::InferenceMode::FP32), m_deviceId(0) {
+        // create Logger with smart pointer
+        m_logger = std::make_unique<TRTLogger>();
+    }
+
+    int32_t initDevice() {
+        cudaError_t error = cudaSetDevice(m_deviceId);
+        if (error != cudaSuccess) {
+            INSPIRE_LOGE("[CUDA error] The device fails to use cuda:%d, %s", m_deviceId, cudaGetErrorString(error));
+            return TENSORRT_HFAIL;
+        }
+        return TENSORRT_HSUCCEED;
+    }
+
+    void setDevice(int32_t deviceId) {
+        m_deviceId = deviceId;
+    }
+
+    ~Impl() {
+        // release resources - device memory needs to be released manually
+        for (auto &pair : m_deviceBuffers) {
+            if (pair.second) {
+                cudaFree(pair.second);
+            }
+        }
+        m_deviceBuffers.clear();
+
+        // smart pointers will handle the release of other resources
+    }
+
+    int32_t readFromFile(const std::string &enginePath) {
+        // read serialized engine file
+        std::vector<char> modelData = readModelFile(enginePath);
+        if (modelData.empty()) {
+            return TENSORRT_HFAIL;
+        }
+
+        return deserializeEngine(modelData);
+    }
+
+    int32_t readFromBin(const std::vector<char> &model_data) {
+        if (model_data.empty()) {
+            return TENSORRT_HFAIL;
+        }
+
+        return deserializeEngine(model_data);
+    }
+
+    int32_t readFromBin(void *model_data, unsigned int model_size) {
+        if (!model_data || model_size == 0) {
+            INSPIRE_LOGE("[TensorRT error] invalid model data or size");
+            return TENSORRT_HFAIL;
+        }
+
+        // convert memory data to vector to reuse the existing deserializeEngine method
+        std::vector<char> modelBuffer(static_cast<char *>(model_data), static_cast<char *>(model_data) + model_size);
+
+        return deserializeEngine(modelBuffer);
+    }
+
+    // create and deserialize engine
+    int32_t deserializeEngine(const std::vector<char> &modelData) {
+        // init device
+        initDevice();
+        // create runtime
+        m_runtime.reset(nvinfer1::createInferRuntime(*m_logger));
+        if (!m_runtime) {
+            INSPIRE_LOGE("[TensorRT error] failed to create TensorRT runtime");
+            return TENSORRT_HFAIL;
+        }
+
+        // deserialize engine
+        m_engine.reset(m_runtime->deserializeCudaEngine(modelData.data(), modelData.size()));
+        if (!m_engine) {
+            INSPIRE_LOGE("[TensorRT error] failed to deserialize engine");
+            return TENSORRT_HFAIL;
+        }
+
+        // create execution context
+        m_context.reset(m_engine->createExecutionContext());
+        if (!m_context) {
+            INSPIRE_LOGE("[TensorRT error] failed to create execution context");
+            return TENSORRT_HFAIL;
+        }
+
+        // get all input and output tensor names
+        int numIoTensors = m_engine->getNbIOTensors();
+        for (int i = 0; i < numIoTensors; ++i) {
+            const char *name = m_engine->getIOTensorName(i);
+            nvinfer1::TensorIOMode mode = m_engine->getTensorIOMode(name);
+            if (mode == nvinfer1::TensorIOMode::kINPUT) {
+                m_inputNames.push_back(name);
+            } else {
+                m_outputNames.push_back(name);
+            }
+        }
+
+        // initialize CUDA stream
+        if (!m_stream) {
+            cudaStream_t *stream = new cudaStream_t;
+            CHECK_CUDA(cudaStreamCreate(stream));
+            m_stream.reset(stream);
+            m_ownStream = true;
+        }
+
+        // pre-allocate device memory
+        return allocateDeviceMemory();
+    }
+
+    // allocate device memory
+    int32_t allocateDeviceMemory() {
+        // allocate device memory for each input and output tensor
+        for (const auto &name : m_inputNames) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(name.c_str());
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(name.c_str());
+            size_t size = getMemorySize(dims, dtype);
+
+            void *buffer = nullptr;
+            CHECK_CUDA(cudaMalloc(&buffer, size));
+            m_deviceBuffers[name] = buffer;
+
+            // store shape information
+            m_inputShapes[name] = dimsToVector(dims);
+        }
+
+        for (const auto &name : m_outputNames) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(name.c_str());
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(name.c_str());
+            size_t size = getMemorySize(dims, dtype);
+
+            void *buffer = nullptr;
+            CHECK_CUDA(cudaMalloc(&buffer, size));
+            m_deviceBuffers[name] = buffer;
+
+            // Save shape information
+            m_outputShapes[name] = dimsToVector(dims);
+        }
+
+        return TENSORRT_HSUCCEED;
+    }
+
+    // set input data
+    void setInput(const char *inputName, const void *data) {
+        auto it = m_deviceBuffers.find(inputName);
+        if (it != m_deviceBuffers.end()) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(inputName);
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(inputName);
+            size_t size = getMemorySize(dims, dtype);
+
+            // copy data from host to device
+            cudaMemcpyAsync(it->second, data, size, cudaMemcpyHostToDevice, *m_stream.get());
+            cudaStreamSynchronize(*m_stream.get());  // add synchronization to ensure data is fully copied
+        }
+    }
+
+    // set batch size (only for models with dynamic shapes)
+    int32_t setBatchSize(int batchSize) {
+        if (m_inputNames.empty())
+            return TENSORRT_HFAIL;
+
+        for (const auto &name : m_inputNames) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(name.c_str());
+            if (dims.nbDims > 0) {
+                nvinfer1::Dims newDims = dims;
+                newDims.d[0] = batchSize;
+
+                if (!m_context->setInputShape(name.c_str(), newDims)) {
+                    INSPIRE_LOGE("[TensorRT error] failed to set input shape for %s", name.c_str());
+                    return TENSORRT_HFAIL;
+                }
+
+                // update shape information
+                m_inputShapes[name] = dimsToVector(newDims);
+            }
+        }
+
+        return TENSORRT_HSUCCEED;
+    }
+
+    // forward inference
+    int32_t forward() {
+        if (!m_context || !m_engine) {
+            return TENSORRT_HFAIL;
+        }
+
+        // check if all tensors are bound to addresses
+        for (const auto &name : m_inputNames) {
+            if (!m_context->setTensorAddress(name.c_str(), m_deviceBuffers[name])) {
+                INSPIRE_LOGE("[TensorRT error] failed to set input tensor %s address", name.c_str());
+                return TENSORRT_FORWARD_FAILED;
+            }
+        }
+
+        for (const auto &name : m_outputNames) {
+            if (!m_context->setTensorAddress(name.c_str(), m_deviceBuffers[name])) {
+                INSPIRE_LOGE("[TensorRT error] failed to set output tensor %s address", name.c_str());
+                return TENSORRT_FORWARD_FAILED;
+            }
+        }
+
+        // record start time - use high precision timing
+        auto start = std::chrono::high_resolution_clock::now();
+
+        // forward inference
+        bool status = m_context->enqueueV3(*m_stream.get());
+
+        // synchronize CUDA stream
+        cudaStreamSynchronize(*m_stream.get());
+
+        // record end time
+        auto end = std::chrono::high_resolution_clock::now();
+
+        // calculate duration (microseconds) then convert to milliseconds, keep high precision
+        auto duration_us = std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+        m_inferenceTime = duration_us.count() / 1000.0;
+
+        return status ? TENSORRT_HSUCCEED : TENSORRT_FORWARD_FAILED;
+    }
+
+    // get output data
+    const void *getOutput(const char *nodeName) {
+        auto it = m_deviceBuffers.find(nodeName);
+        if (it != m_deviceBuffers.end()) {
+            nvinfer1::Dims dims = m_context->getTensorShape(nodeName);
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(nodeName);
+            size_t size = getMemorySize(dims, dtype);
+
+            // copy output data from device to host
+            if (m_hostOutputBuffers.find(nodeName) == m_hostOutputBuffers.end()) {
+                m_hostOutputBuffers[nodeName].resize(size);
+            }
+
+            cudaMemcpyAsync(m_hostOutputBuffers[nodeName].data(), it->second, size, cudaMemcpyDeviceToHost, *m_stream.get());
+            cudaStreamSynchronize(*m_stream.get());
+
+            return m_hostOutputBuffers[nodeName].data();
+        }
+        return nullptr;
+    }
+
+    // get output data and convert to float type vector
+    std::vector<float> getOutputAsFloat(const char *nodeName) {
+        std::vector<float> result;
+        auto it = m_deviceBuffers.find(nodeName);
+        if (it != m_deviceBuffers.end()) {
+            nvinfer1::Dims dims = m_context->getTensorShape(nodeName);
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(nodeName);
+
+            // calculate total number of elements
+            size_t numElements = 1;
+            for (int i = 0; i < dims.nbDims; ++i) {
+                numElements *= dims.d[i];
+            }
+
+            // allocate buffer of appropriate size based on data type
+            size_t elementSize = 0;
+            switch (dtype) {
+                case nvinfer1::DataType::kFLOAT:
+                    elementSize = sizeof(float);
+                    break;
+                case nvinfer1::DataType::kHALF:
+                    elementSize = sizeof(half);
+                    break;
+                case nvinfer1::DataType::kINT8:
+                    elementSize = sizeof(int8_t);
+                    break;
+                case nvinfer1::DataType::kINT32:
+                    elementSize = sizeof(int32_t);
+                    break;
+                default:
+                    return result;
+            }
+
+            // allocate temporary buffer
+            std::vector<unsigned char> buffer(numElements * elementSize);
+
+            // copy data from device memory to host memory
+            cudaMemcpyAsync(buffer.data(), it->second, buffer.size(), cudaMemcpyDeviceToHost, *m_stream.get());
+            cudaStreamSynchronize(*m_stream.get());
+
+            // convert to float based on data type
+            result.resize(numElements);
+            switch (dtype) {
+                case nvinfer1::DataType::kFLOAT:
+                    std::memcpy(result.data(), buffer.data(), buffer.size());
+                    break;
+                case nvinfer1::DataType::kHALF: {
+                    const half *halfData = reinterpret_cast<const half *>(buffer.data());
+                    for (size_t i = 0; i < numElements; ++i) {
+                        result[i] = __half2float(halfData[i]);
+                    }
+                    break;
+                }
+                case nvinfer1::DataType::kINT8: {
+                    const int8_t *int8Data = reinterpret_cast<const int8_t *>(buffer.data());
+                    for (size_t i = 0; i < numElements; ++i) {
+                        result[i] = static_cast<float>(int8Data[i]);
+                    }
+                    break;
+                }
+                case nvinfer1::DataType::kINT32: {
+                    const int32_t *int32Data = reinterpret_cast<const int32_t *>(buffer.data());
+                    for (size_t i = 0; i < numElements; ++i) {
+                        result[i] = static_cast<float>(int32Data[i]);
+                    }
+                    break;
+                }
+            }
+        }
+        return result;
+    }
+
+    // set inference mode
+    void setInferenceMode(TensorRTAdapter::InferenceMode mode) {
+        m_inferenceMode = mode;
+        // apply this setting during actual inference
+    }
+
+    // set CUDA stream
+    void setCudaStream(void *streamPtr) {
+        if (m_ownStream) {
+            m_stream.reset();
+            m_ownStream = false;
+        }
+
+        // create a new smart pointer instead of using reset + lambda
+        cudaStream_t *streamPointer = static_cast<cudaStream_t *>(streamPtr);
+        // use empty deleter, because this stream is managed by external code
+        m_stream =
+          std::unique_ptr<cudaStream_t, CUDAStreamDeleter>(streamPointer,
+                                                           CUDAStreamDeleter()  // use default deleter, but not actually delete the external stream
+          );
+    }
+
+    // print model info
+    void printModelInfo() const {
+        INSPIRE_LOGI("================================================");
+        if (!m_engine) {
+            INSPIRE_LOGE("[TensorRT error] engine not initialized");
+            return;
+        }
+        INSPIRE_LOGI("\nengine info:");
+        INSPIRE_LOGI("engine layers: %d", m_engine->getNbLayers());
+        INSPIRE_LOGI("input/output tensors: %d", m_engine->getNbIOTensors());
+
+        INSPIRE_LOGI("\ninput tensors:");
+        for (const auto &name : m_inputNames) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(name.c_str());
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(name.c_str());
+
+            INSPIRE_LOGI("name: %s, shape: (", name.c_str());
+            for (int d = 0; d < dims.nbDims; ++d) {
+                INSPIRE_LOGI("%d", dims.d[d]);
+                if (d < dims.nbDims - 1)
+                    INSPIRE_LOGI(", ");
+            }
+            INSPIRE_LOGI("), type: %s", getDataTypeString(dtype).c_str());
+        }
+
+        INSPIRE_LOGI("\noutput tensors:");
+        for (const auto &name : m_outputNames) {
+            nvinfer1::Dims dims = m_engine->getTensorShape(name.c_str());
+            nvinfer1::DataType dtype = m_engine->getTensorDataType(name.c_str());
+
+            INSPIRE_LOGI("name: %s, shape: (", name.c_str());
+            for (int d = 0; d < dims.nbDims; ++d) {
+                INSPIRE_LOGI("%d", dims.d[d]);
+                if (d < dims.nbDims - 1)
+                    INSPIRE_LOGI(", ");
+            }
+            INSPIRE_LOGI("), type: %s", getDataTypeString(dtype).c_str());
+        }
+        INSPIRE_LOGI("================================================");
+    }
+
+    // get input tensor names list
+    const std::vector<std::string> &getInputNames() const {
+        return m_inputNames;
+    }
+
+    // get output tensor names list
+    const std::vector<std::string> &getOutputNames() const {
+        return m_outputNames;
+    }
+
+    // get input tensor shape by name
+    const std::vector<int> &getInputShapeByName(const std::string &name) const {
+        static std::vector<int> emptyShape;
+        auto it = m_inputShapes.find(name);
+        return (it != m_inputShapes.end()) ? it->second : emptyShape;
+    }
+
+    // get output tensor shape by name
+    const std::vector<int> &getOutputShapeByName(const std::string &name) const {
+        static std::vector<int> emptyShape;
+        auto it = m_outputShapes.find(name);
+        return (it != m_outputShapes.end()) ? it->second : emptyShape;
+    }
+
+    // get inference time
+    double getInferenceTime() const {
+        return m_inferenceTime;
+    }
+
+private:
+    // helper function: convert TensorRT's Dims to standard vector
+    std::vector<int> dimsToVector(const nvinfer1::Dims &dims) const {
+        std::vector<int> shape;
+        for (int i = 0; i < dims.nbDims; ++i) {
+            shape.push_back(dims.d[i]);
+        }
+        return shape;
+    }
+
+    // helper function: calculate memory size
+    size_t getMemorySize(const nvinfer1::Dims &dims, nvinfer1::DataType dtype) const {
+        size_t size = 1;
+        for (int i = 0; i < dims.nbDims; ++i) {
+            size *= dims.d[i];
+        }
+
+        switch (dtype) {
+            case nvinfer1::DataType::kFLOAT:
+                return size * 4;
+            case nvinfer1::DataType::kHALF:
+                return size * 2;
+            case nvinfer1::DataType::kINT8:
+                return size;
+            case nvinfer1::DataType::kINT32:
+                return size * 4;
+            case nvinfer1::DataType::kBOOL:
+                return size;
+            default:
+                return size;
+        }
+    }
+
+    // helper function: get data type string representation
+    std::string getDataTypeString(nvinfer1::DataType dtype) const {
+        switch (dtype) {
+            case nvinfer1::DataType::kFLOAT:
+                return "FLOAT";
+            case nvinfer1::DataType::kHALF:
+                return "HALF";
+            case nvinfer1::DataType::kINT8:
+                return "INT8";
+            case nvinfer1::DataType::kINT32:
+                return "INT32";
+            case nvinfer1::DataType::kBOOL:
+                return "BOOL";
+            default:
+                return "UNKNOWN";
+        }
+    }
+
+    // member variables - using smart pointers
+    std::unique_ptr<TRTLogger> m_logger;
+    std::unique_ptr<nvinfer1::IRuntime, TRTRuntimeDeleter> m_runtime;
+    std::unique_ptr<nvinfer1::ICudaEngine, TRTEngineDeleter> m_engine;
+    std::unique_ptr<nvinfer1::IExecutionContext, TRTContextDeleter> m_context;
+
+    bool m_ownStream;
+    std::unique_ptr<cudaStream_t, CUDAStreamDeleter> m_stream;
+
+    int32_t m_deviceId{0};
+
+    std::vector<std::string> m_inputNames;
+    std::vector<std::string> m_outputNames;
+    std::map<std::string, void *> m_deviceBuffers;
+    std::map<std::string, std::vector<unsigned char>> m_hostOutputBuffers;
+
+    std::map<std::string, std::vector<int>> m_inputShapes;
+    std::map<std::string, std::vector<int>> m_outputShapes;
+
+    TensorRTAdapter::InferenceMode m_inferenceMode;
+    double m_inferenceTime;
+};
+
+// implement TensorRTAdapter methods
+TensorRTAdapter::TensorRTAdapter() : pImpl(new Impl()) {}
+
+TensorRTAdapter::~TensorRTAdapter() {
+    if (pImpl) {
+        delete pImpl;
+        pImpl = nullptr;
+    }
+}
+
+int32_t TensorRTAdapter::readFromFile(const std::string &enginePath) {
+    return pImpl->readFromFile(enginePath);
+}
+
+int32_t TensorRTAdapter::readFromBin(void *model_data, unsigned int model_size) {
+    return pImpl->readFromBin(model_data, model_size);
+}
+
+TensorRTAdapter TensorRTAdapter::readNetFrom(const std::string &enginePath) {
+    TensorRTAdapter adapter;
+    adapter.readFromFile(enginePath);
+    return adapter;
+}
+
+TensorRTAdapter TensorRTAdapter::readNetFromBin(const std::vector<char> &model_data) {
+    TensorRTAdapter adapter;
+    adapter.pImpl->readFromBin(model_data);
+    return adapter;
+}
+
+std::vector<std::string> TensorRTAdapter::getInputNames() const {
+    return pImpl->getInputNames();
+}
+
+std::vector<std::string> TensorRTAdapter::getOutputNames() const {
+    return pImpl->getOutputNames();
+}
+
+std::vector<int> TensorRTAdapter::getInputShapeByName(const std::string &name) {
+    return pImpl->getInputShapeByName(name);
+}
+
+std::vector<int> TensorRTAdapter::getOutputShapeByName(const std::string &name) {
+    return pImpl->getOutputShapeByName(name);
+}
+
+void TensorRTAdapter::setInput(const char *inputName, const void *data) {
+    pImpl->setInput(inputName, data);
+}
+
+int32_t TensorRTAdapter::setBatchSize(int batchSize) {
+    return pImpl->setBatchSize(batchSize);
+}
+
+int32_t TensorRTAdapter::forward() {
+    return pImpl->forward();
+}
+
+const void *TensorRTAdapter::getOutput(const char *nodeName) {
+    return pImpl->getOutput(nodeName);
+}
+
+std::vector<float> TensorRTAdapter::getOutputAsFloat(const char *nodeName) {
+    return pImpl->getOutputAsFloat(nodeName);
+}
+
+double TensorRTAdapter::getInferenceTime() const {
+    return pImpl->getInferenceTime();
+}
+
+void TensorRTAdapter::setInferenceMode(InferenceMode mode) {
+    pImpl->setInferenceMode(mode);
+}
+
+void TensorRTAdapter::setCudaStream(void *streamPtr) {
+    pImpl->setCudaStream(streamPtr);
+}
+
+void TensorRTAdapter::printModelInfo() const {
+    pImpl->printModelInfo();
+}
+
+void TensorRTAdapter::setDevice(int32_t deviceId) {
+    pImpl->setDevice(deviceId);
+}
+#endif  // ISF_ENABLE_TENSORRT
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.h b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.h
new file mode 100644
index 000000000..3dcc0e983
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/inference_wrapper/tensorrt/tensorrt_adapter.h
@@ -0,0 +1,175 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-16
+ */
+#if ISF_ENABLE_TENSORRT
+#ifndef INSPIRE_TENSORRT_ADAPTER_H
+#define INSPIRE_TENSORRT_ADAPTER_H
+
+#include <string>
+#include <vector>
+#include <map>
+
+#define TENSORRT_HSUCCEED 0
+#define TENSORRT_HFAIL -1
+#define TENSORRT_FORWARD_FAILED -2
+
+/**
+ * @brief The TensorRT adapter is used for inference
+ */
+class TensorRTAdapter {
+public:
+    /**
+     * @brief inference mode (abeyance)
+     */
+    enum class InferenceMode {
+        FP32,  ///< FP32 precision inference
+        FP16,  ///< FP16 precision inference
+        INT8   ///< INT8 precision inference (requires calibration)
+    };
+
+    /**
+     * @brief output shape mapping
+     */
+    typedef std::map<std::string, std::vector<int>> TensorShapesMap;
+
+    /**
+     * @brief default constructor
+     */
+    TensorRTAdapter();
+
+    /**
+     * @brief destructor
+     */
+    ~TensorRTAdapter();
+
+    /**
+     * @brief read model from file
+     * @param enginePath engine file path
+     * @return 0 means success, -1 means failure
+     */
+    int32_t readFromFile(const std::string &enginePath);
+
+    /**
+     * @brief read model from binary data
+     * @param model_data binary model data
+     * @param model_size model data size
+     * @return 0 means success, -1 means failure
+     */
+    int32_t readFromBin(void *model_data, unsigned int model_size);
+
+    /**
+     * @brief read model from file static method
+     * @param enginePath engine file path
+     * @return TensorRTAdapter instance
+     */
+    static TensorRTAdapter readNetFrom(const std::string &enginePath);
+
+    /**
+     * @brief read model from binary data static method
+     * @param model_data binary model data
+     * @return TensorRTAdapter instance
+     */
+    static TensorRTAdapter readNetFromBin(const std::vector<char> &model_data);
+
+    /**
+     * @brief get all input tensor names
+     * @return input tensor names list
+     */
+    std::vector<std::string> getInputNames() const;
+
+    /**
+     * @brief get all output tensor names
+     * @return output tensor names list
+     */
+    std::vector<std::string> getOutputNames() const;
+
+    /**
+     * @brief get input tensor shape by name
+     * @param name tensor name
+     * @return shape vector
+     */
+    std::vector<int> getInputShapeByName(const std::string &name);
+
+    /**
+     * @brief get output tensor shape by name
+     * @param name tensor name
+     * @return shape vector
+     */
+    std::vector<int> getOutputShapeByName(const std::string &name);
+
+    /**
+     * @brief set input data
+     * @param inputName input tensor name
+     * @param data input data
+     */
+    void setInput(const char *inputName, const void *data);
+
+    /**
+     * @brief set dynamic batch size
+     * @param batchSize batch size
+     * @return 0 means success, -1 means failure
+     */
+    int32_t setBatchSize(int batchSize);
+
+    /**
+     * @brief forward inference
+     * @return 0 means success, -1 means failure
+     */
+    int32_t forward();
+
+    /**
+     * @brief get output data
+     * @param nodeName output tensor name
+     * @return output data pointer
+     */
+    const void *getOutput(const char *nodeName);
+
+    /**
+     * @brief get output data and convert to float type
+     * @param nodeName output tensor name
+     * @return output data pointer
+     */
+    std::vector<float> getOutputAsFloat(const char *nodeName);
+
+    /**
+     * @brief get inference time (ms)
+     * @return inference time
+     */
+    double getInferenceTime() const;
+
+    /**
+     * @brief set inference mode
+     * @param mode inference mode
+     */
+    void setInferenceMode(InferenceMode mode);
+
+    /**
+     * @brief set CUDA stream
+     * @param stream CUDA stream
+     */
+    void setCudaStream(void *streamPtr);
+
+    /**
+     * @brief print model info
+     */
+    void printModelInfo() const;
+
+    /**
+     * @brief set CUDA device
+     * @param deviceId CUDA device id
+     */
+    void setDevice(int32_t deviceId);
+
+private:
+    // use PIMPL pattern to hide implementation details
+    class Impl;
+    Impl *pImpl;
+
+    // output shape cache
+    TensorShapesMap m_inputShapes;
+    TensorShapesMap m_outputShapes;
+};
+
+#endif  // INSPIRE_TENSORRT_ADAPTER_H
+#endif  // ISF_ENABLE_TENSORRT
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.cc b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.cc
new file mode 100644
index 000000000..cb24d0b5a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.cc
@@ -0,0 +1,171 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-23
+ */
+#include "core_archive.h"
+#include "microtar/microtar.h"
+#include "log.h"
+#include <unordered_map>
+#include <iostream>
+
+namespace inspire {
+
+class CoreArchive::Impl {
+public:
+    Impl() : m_load_file_status_(SARC_NOT_LOAD), m_tar_(nullptr) {}
+
+    explicit Impl(const std::string& archiveFile) : m_load_file_status_(SARC_NOT_LOAD), m_tar_(nullptr) {
+        Reset(archiveFile);
+    }
+
+    ~Impl() {
+        Close();
+    }
+
+    int32_t Reset(const std::string& archiveFile) {
+        Close();
+        std::vector<char>().swap(m_empty_);
+        m_tar_ = std::make_shared<mtar_t>();
+        m_load_file_status_ = mtar_open(m_tar_.get(), archiveFile.c_str(), "r");
+        if (m_load_file_status_ != MTAR_ESUCCESS) {
+            INSPIRE_LOGE("Invalid archive file: %d", m_load_file_status_);
+            Close();
+            m_tar_.reset();
+            return m_load_file_status_;
+        }
+        mtar_header_t h;
+        m_load_file_status_ = mtar_read_header(m_tar_.get(), &h);
+        if (m_load_file_status_ != MTAR_ESUCCESS) {
+            INSPIRE_LOGE("Error reading root from archive.");
+            return m_load_file_status_;
+        }
+        //        m_file_archive_root_ = std::string(h.name);
+        size_t index = 0;
+        while ((mtar_read_header(m_tar_.get(), &h)) != MTAR_ENULLRECORD) {
+            m_load_file_status_ = mtar_next(m_tar_.get());
+            if (m_load_file_status_ != MTAR_ESUCCESS) {
+                INSPIRE_LOGE("Failed to scan the file: %d", m_load_file_status_);
+                Close();
+                return m_load_file_status_;
+            }
+            m_subfiles_names_.emplace_back(h.name);
+            index++;
+        }
+        return m_load_file_status_;
+    }
+
+    std::vector<char>& GetFileContent(const std::string& filename) {
+        auto index = filenameFuzzyMatching(filename);
+        if (index != std::string::npos) {
+            auto fullFilename = m_subfiles_names_[index];
+            auto ret = lazyReadFile(fullFilename);
+            if (ret != MTAR_ESUCCESS) {
+                INSPIRE_LOGE("Failed to load file");
+            }
+            return m_file_content_cache_map_[fullFilename];
+        }
+        return m_empty_;
+    }
+
+    int32_t QueryLoadStatus() const {
+        return m_load_file_status_;
+    }
+
+    const std::vector<std::string>& GetSubfilesNames() const {
+        return m_subfiles_names_;
+    }
+
+    void Close() {
+        if (m_tar_ && m_tar_.get() != nullptr) {
+            mtar_close(m_tar_.get());
+        }
+        m_tar_.reset();
+        m_load_file_status_ = SARC_NOT_LOAD;
+        m_subfiles_names_.clear();
+        m_file_content_cache_map_.clear();
+    }
+
+    void PrintSubFiles() {
+        std::cout << "Subfiles: " << m_subfiles_names_.size() << std::endl;
+        for (int i = 0; i < m_subfiles_names_.size(); ++i) {
+            std::cout << m_subfiles_names_[i] << std::endl;
+        }
+    }
+
+private:
+    size_t filenameFuzzyMatching(const std::string& filename) {
+        for (size_t i = 0; i < m_subfiles_names_.size(); ++i) {
+            if (m_subfiles_names_[i].find(filename) != std::string::npos) {
+                return i;
+            }
+        }
+        return std::string::npos;
+    }
+
+    int32_t lazyReadFile(const std::string& filename) {
+        if (m_file_content_cache_map_.find(filename) != m_file_content_cache_map_.end()) {
+            return MTAR_ESUCCESS;
+        }
+        mtar_header_t h;
+        auto ret = mtar_find(m_tar_.get(), filename.c_str(), &h);
+        if (ret == MTAR_ESUCCESS) {
+            std::vector<char> content(h.size);
+            ret = mtar_read_data(m_tar_.get(), content.data(), h.size);
+            if (ret == MTAR_ESUCCESS) {
+                m_file_content_cache_map_[filename] = std::move(content);  // Load and store the file contents
+                return MTAR_ESUCCESS;
+            } else {
+                INSPIRE_LOGE("Failed to load file: %d", ret);
+            }
+        } else {
+            INSPIRE_LOGE("Failed to find file: %d", ret);
+        }
+
+        return SARC_LOAD_FILE_FAIL;
+    }
+
+    std::string m_file_archive_root_;            ///< Archive file path
+    std::vector<std::string> m_subfiles_names_;  ///< Name list of subfiles
+    std::shared_ptr<mtar_t> m_tar_;              ///< mtar context
+    int32_t m_load_file_status_;                 ///< Initiation status code
+
+    std::vector<char> m_empty_;  ///< Const empty
+
+    std::unordered_map<std::string, std::vector<char>> m_file_content_cache_map_;  ///< File buffer cache
+};
+
+CoreArchive::CoreArchive() : m_pImpl(std::make_unique<Impl>()) {}
+
+CoreArchive::CoreArchive(const std::string& archiveFile) : m_pImpl(std::make_unique<Impl>(archiveFile)) {}
+
+CoreArchive::~CoreArchive() = default;
+
+CoreArchive::CoreArchive(CoreArchive&& other) noexcept = default;
+
+CoreArchive& CoreArchive::operator=(CoreArchive&& other) noexcept = default;
+
+int32_t CoreArchive::Reset(const std::string& archiveFile) {
+    return m_pImpl->Reset(archiveFile);
+}
+
+std::vector<char>& CoreArchive::GetFileContent(const std::string& filename) {
+    return m_pImpl->GetFileContent(filename);
+}
+
+int32_t CoreArchive::QueryLoadStatus() const {
+    return m_pImpl->QueryLoadStatus();
+}
+
+const std::vector<std::string>& CoreArchive::GetSubfilesNames() const {
+    return m_pImpl->GetSubfilesNames();
+}
+
+void CoreArchive::Close() {
+    m_pImpl->Close();
+}
+
+void CoreArchive::PrintSubFiles() {
+    m_pImpl->PrintSubFiles();
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.h b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.h
new file mode 100644
index 000000000..eeac7647c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/core_archive.h
@@ -0,0 +1,61 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-23
+ */
+#pragma once
+#ifndef MODELLOADERTAR_CPP_MICROTAR_H
+#define MODELLOADERTAR_CPP_MICROTAR_H
+
+#ifndef INSPIRE_API
+#define INSPIRE_API
+#endif
+
+#include <string>
+#include <vector>
+#include <memory>
+
+namespace inspire {
+
+enum {
+    SARC_SUCCESS = 0,
+    SARC_FAILURE = -1,
+    SARC_OPEN_FAIL = -2,
+    SARC_READ_FAIL = -3,
+    SARC_WRITE_FAIL = -4,
+    SARC_SEEK_FAIL = -5,
+    SARC_BAD_CHKSUM = -6,
+    SARC_NULL_RECORD = -7,
+    SARC_NOTFOUND = -8,
+    SARC_LOAD_FILE_FAIL = -9,
+    SARC_NOT_LOAD = -10,
+};
+
+class INSPIRE_API CoreArchive {
+public:
+    explicit CoreArchive(const std::string& archiveFile);
+    explicit CoreArchive();
+    ~CoreArchive();
+
+    // Copy construction and assignment operations are disabled
+    CoreArchive(const CoreArchive&) = delete;
+    CoreArchive& operator=(const CoreArchive&) = delete;
+
+    // Enable mobile construction and assignment operations
+    CoreArchive(CoreArchive&& other) noexcept;
+    CoreArchive& operator=(CoreArchive&& other) noexcept;
+
+    int32_t Reset(const std::string& archiveFile);
+    std::vector<char>& GetFileContent(const std::string& filename);
+    int32_t QueryLoadStatus() const;
+    const std::vector<std::string>& GetSubfilesNames() const;
+    void Close();
+    void PrintSubFiles();
+
+private:
+    class Impl;
+    std::unique_ptr<Impl> m_pImpl;
+
+};  // class CoreArchive
+}  // namespace inspire
+
+#endif  // MODELLOADERTAR_CPP_MICROTAR_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.c b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.c
new file mode 100644
index 000000000..6f76b0c30
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.c
@@ -0,0 +1,414 @@
+/*
+ * Copyright (c) 2017 rxi
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "microtar.h"
+
+typedef struct {
+  char name[100];
+  char mode[8];
+  char owner[8];
+  char group[8];
+  char size[12];
+  char mtime[12];
+  char checksum[8];
+  char type;
+  char linkname[100];
+  char _padding[255];
+} mtar_raw_header_t;
+
+
+static unsigned round_up(unsigned n, unsigned incr) {
+  return n + (incr - n % incr) % incr;
+}
+
+
+static unsigned checksum(const mtar_raw_header_t* rh) {
+  unsigned i;
+  unsigned char *p = (unsigned char*) rh;
+  unsigned res = 256;
+  for (i = 0; i < offsetof(mtar_raw_header_t, checksum); i++) {
+    res += p[i];
+  }
+  for (i = offsetof(mtar_raw_header_t, type); i < sizeof(*rh); i++) {
+    res += p[i];
+  }
+  return res;
+}
+
+
+static int tread(mtar_t *tar, void *data, unsigned size) {
+  int err = tar->read(tar, data, size);
+  tar->pos += size;
+  return err;
+}
+
+
+static int twrite(mtar_t *tar, const void *data, unsigned size) {
+  int err = tar->write(tar, data, size);
+  tar->pos += size;
+  return err;
+}
+
+
+static int write_null_bytes(mtar_t *tar, int n) {
+  int i, err;
+  char nul = '\0';
+  for (i = 0; i < n; i++) {
+    err = twrite(tar, &nul, 1);
+    if (err) {
+      return err;
+    }
+  }
+  return MTAR_ESUCCESS;
+}
+
+
+static int raw_to_header(mtar_header_t *h, const mtar_raw_header_t *rh) {
+  unsigned chksum1, chksum2;
+
+  /* If the checksum starts with a null byte we assume the record is NULL */
+  if (*rh->checksum == '\0') {
+    return MTAR_ENULLRECORD;
+  }
+
+  /* Build and compare checksum */
+  chksum1 = checksum(rh);
+  sscanf(rh->checksum, "%o", &chksum2);
+  if (chksum1 != chksum2) {
+    return MTAR_EBADCHKSUM;
+  }
+
+  /* Load raw header into header */
+  sscanf(rh->mode, "%o", &h->mode);
+  sscanf(rh->owner, "%o", &h->owner);
+  sscanf(rh->size, "%o", &h->size);
+  sscanf(rh->mtime, "%o", &h->mtime);
+  h->type = rh->type;
+  strcpy(h->name, rh->name);
+  strcpy(h->linkname, rh->linkname);
+
+  return MTAR_ESUCCESS;
+}
+
+
+static int header_to_raw(mtar_raw_header_t *rh, const mtar_header_t *h) {
+  unsigned chksum;
+
+  /* Load header into raw header */
+  memset(rh, 0, sizeof(*rh));
+  sprintf(rh->mode, "%o", h->mode);
+  sprintf(rh->owner, "%o", h->owner);
+  sprintf(rh->size, "%o", h->size);
+  sprintf(rh->mtime, "%o", h->mtime);
+  rh->type = h->type ? h->type : MTAR_TREG;
+  strcpy(rh->name, h->name);
+  strcpy(rh->linkname, h->linkname);
+
+  /* Calculate and write checksum */
+  chksum = checksum(rh);
+  sprintf(rh->checksum, "%06o", chksum);
+  rh->checksum[7] = ' ';
+
+  return MTAR_ESUCCESS;
+}
+
+
+const char* mtar_strerror(int err) {
+  switch (err) {
+    case MTAR_ESUCCESS     : return "success";
+    case MTAR_EFAILURE     : return "failure";
+    case MTAR_EOPENFAIL    : return "could not open";
+    case MTAR_EREADFAIL    : return "could not read";
+    case MTAR_EWRITEFAIL   : return "could not write";
+    case MTAR_ESEEKFAIL    : return "could not seek";
+    case MTAR_EBADCHKSUM   : return "bad checksum";
+    case MTAR_ENULLRECORD  : return "null record";
+    case MTAR_ENOTFOUND    : return "file not found";
+  }
+  return "unknown error";
+}
+
+
+static int file_write(mtar_t *tar, const void *data, unsigned size) {
+  unsigned res = fwrite(data, 1, size, tar->stream);
+  return (res == size) ? MTAR_ESUCCESS : MTAR_EWRITEFAIL;
+}
+
+static int file_read(mtar_t *tar, void *data, unsigned size) {
+  unsigned res = fread(data, 1, size, tar->stream);
+  return (res == size) ? MTAR_ESUCCESS : MTAR_EREADFAIL;
+}
+
+static int file_seek(mtar_t *tar, unsigned offset) {
+  int res = fseek(tar->stream, offset, SEEK_SET);
+  return (res == 0) ? MTAR_ESUCCESS : MTAR_ESEEKFAIL;
+}
+
+static int file_close(mtar_t *tar) {
+  if (tar->source_type == FROM_FILE) {
+    fclose(tar->stream);
+  }
+  return MTAR_ESUCCESS;
+}
+
+
+int mtar_open(mtar_t *tar, const char *filename, const char *mode) {
+  int err;
+  mtar_header_t h;
+
+  /* Init tar struct and functions */
+  memset(tar, 0, sizeof(*tar));
+  tar->write = file_write;
+  tar->read = file_read;
+  tar->seek = file_seek;
+  tar->close = file_close;
+  tar->source_type = FROM_FILE;
+  /* Assure mode is always binary */
+  if ( strchr(mode, 'r') ) mode = "rb";
+  if ( strchr(mode, 'w') ) mode = "wb";
+  if ( strchr(mode, 'a') ) mode = "ab";
+  /* Open file */
+  tar->stream = fopen(filename, mode);
+  if (!tar->stream) {
+    return MTAR_EOPENFAIL;
+  }
+  /* Read first header to check it is valid if mode is `r` */
+  if (*mode == 'r') {
+    err = mtar_read_header(tar, &h);
+    if (err != MTAR_ESUCCESS) {
+      mtar_close(tar);
+      return err;
+    }
+  }
+
+  /* Return ok */
+  return MTAR_ESUCCESS;
+}
+
+
+int mtar_close(mtar_t *tar) {
+  return tar->close(tar);
+}
+
+
+int mtar_seek(mtar_t *tar, unsigned pos) {
+  int err = tar->seek(tar, pos);
+  tar->pos = pos;
+  return err;
+}
+
+
+int mtar_rewind(mtar_t *tar) {
+  tar->remaining_data = 0;
+  tar->last_header = 0;
+  return mtar_seek(tar, 0);
+}
+
+
+int mtar_next(mtar_t *tar) {
+  int err, n;
+  mtar_header_t h;
+  /* Load header */
+  err = mtar_read_header(tar, &h);
+  if (err) {
+    return err;
+  }
+  /* Seek to next record */
+  n = round_up(h.size, 512) + sizeof(mtar_raw_header_t);
+  return mtar_seek(tar, tar->pos + n);
+}
+
+
+int mtar_find(mtar_t *tar, const char *name, mtar_header_t *h) {
+  int err;
+  mtar_header_t header;
+  /* Start at beginning */
+  err = mtar_rewind(tar);
+  if (err) {
+    return err;
+  }
+  /* Iterate all files until we hit an error or find the file */
+  while ( (err = mtar_read_header(tar, &header)) == MTAR_ESUCCESS ) {
+    if ( !strcmp(header.name, name) ) {
+      if (h) {
+        *h = header;
+      }
+      return MTAR_ESUCCESS;
+    }
+    mtar_next(tar);
+  }
+  /* Return error */
+  if (err == MTAR_ENULLRECORD) {
+    err = MTAR_ENOTFOUND;
+  }
+  return err;
+}
+
+
+int mtar_read_header(mtar_t *tar, mtar_header_t *h) {
+  int err;
+  mtar_raw_header_t rh;
+  /* Save header position */
+  tar->last_header = tar->pos;
+  /* Read raw header */
+  err = tread(tar, &rh, sizeof(rh));
+  if (err) {
+    return err;
+  }
+  /* Seek back to start of header */
+  err = mtar_seek(tar, tar->last_header);
+  if (err) {
+    return err;
+  }
+  /* Load raw header into header struct and return */
+  return raw_to_header(h, &rh);
+}
+
+
+int mtar_read_data(mtar_t *tar, void *ptr, unsigned size) {
+  int err;
+  /* If we have no remaining data then this is the first read, we get the size,
+   * set the remaining data and seek to the beginning of the data */
+  if (tar->remaining_data == 0) {
+    mtar_header_t h;
+    /* Read header */
+    err = mtar_read_header(tar, &h);
+    if (err) {
+      return err;
+    }
+    /* Seek past header and init remaining data */
+    err = mtar_seek(tar, tar->pos + sizeof(mtar_raw_header_t));
+    if (err) {
+      return err;
+    }
+    tar->remaining_data = h.size;
+  }
+  /* Read data */
+  err = tread(tar, ptr, size);
+  if (err) {
+    return err;
+  }
+  tar->remaining_data -= size;
+  /* If there is no remaining data we've finished reading and seek back to the
+   * header */
+  if (tar->remaining_data == 0) {
+    return mtar_seek(tar, tar->last_header);
+  }
+  return MTAR_ESUCCESS;
+}
+
+
+int mtar_write_header(mtar_t *tar, const mtar_header_t *h) {
+  mtar_raw_header_t rh;
+  /* Build raw header and write */
+  header_to_raw(&rh, h);
+  tar->remaining_data = h->size;
+  return twrite(tar, &rh, sizeof(rh));
+}
+
+
+int mtar_write_file_header(mtar_t *tar, const char *name, unsigned size) {
+  mtar_header_t h;
+  /* Build header */
+  memset(&h, 0, sizeof(h));
+  strcpy(h.name, name);
+  h.size = size;
+  h.type = MTAR_TREG;
+  h.mode = 0664;
+  /* Write header */
+  return mtar_write_header(tar, &h);
+}
+
+
+int mtar_write_dir_header(mtar_t *tar, const char *name) {
+  mtar_header_t h;
+  /* Build header */
+  memset(&h, 0, sizeof(h));
+  strcpy(h.name, name);
+  h.type = MTAR_TDIR;
+  h.mode = 0775;
+  /* Write header */
+  return mtar_write_header(tar, &h);
+}
+
+
+int mtar_write_data(mtar_t *tar, const void *data, unsigned size) {
+  int err;
+  /* Write data */
+  err = twrite(tar, data, size);
+  if (err) {
+    return err;
+  }
+  tar->remaining_data -= size;
+  /* Write padding if we've written all the data for this file */
+  if (tar->remaining_data == 0) {
+    return write_null_bytes(tar, round_up(tar->pos, 512) - tar->pos);
+  }
+  return MTAR_ESUCCESS;
+}
+
+
+int mtar_finalize(mtar_t *tar) {
+  /* Write two NULL records */
+  return write_null_bytes(tar, sizeof(mtar_raw_header_t) * 2);
+}
+
+
+static int memory_read(mtar_t *tar, void *data, unsigned size) {
+    if (tar->pos + size > tar->stream_size) {
+        return MTAR_EREADFAIL;
+    }
+    memcpy(data, (char *)tar->stream + tar->pos, size);
+    tar->pos += size;
+    return MTAR_ESUCCESS;
+}
+
+static int memory_seek(mtar_t *tar, unsigned pos) {
+    if (pos > tar->stream_size) {
+        return MTAR_ESEEKFAIL;
+    }
+    tar->pos = pos;
+    return MTAR_ESUCCESS;
+}
+
+int mtar_open_memory(mtar_t *tar, void *data, size_t size) {
+    memset(tar, 0, sizeof(*tar));
+    tar->read = memory_read;
+    tar->seek = memory_seek;
+    tar->stream = data;
+    tar->stream_size = size;  // Add a field to store the data size
+    tar->source_type = FROM_MEMORY;
+    tar->close = file_close;
+
+    // Read the first header to verify the data
+    mtar_header_t h;
+    int err = mtar_read_header(tar, &h);
+    if (err != MTAR_ESUCCESS) {
+        return err;
+    }
+    return mtar_rewind(tar);
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.h b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.h
new file mode 100644
index 000000000..3d84acff7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/microtar/microtar.h
@@ -0,0 +1,82 @@
+/**
+ * Copyright (c) 2017 rxi
+ *
+ * This library is free software; you can redistribute it and/or modify it
+ * under the terms of the MIT license. See `microtar.c` for details.
+ */
+
+#ifndef MICROTAR_H
+#define MICROTAR_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define MTAR_VERSION "0.2.0"
+
+enum {
+    MTAR_ESUCCESS = 0,
+    MTAR_EFAILURE = -1,
+    MTAR_EOPENFAIL = -2,
+    MTAR_EREADFAIL = -3,
+    MTAR_EWRITEFAIL = -4,
+    MTAR_ESEEKFAIL = -5,
+    MTAR_EBADCHKSUM = -6,
+    MTAR_ENULLRECORD = -7,
+    MTAR_ENOTFOUND = -8
+};
+
+enum { MTAR_TREG = '0', MTAR_TLNK = '1', MTAR_TSYM = '2', MTAR_TCHR = '3', MTAR_TBLK = '4', MTAR_TDIR = '5', MTAR_TFIFO = '6' };
+
+typedef struct {
+    unsigned mode;
+    unsigned owner;
+    unsigned size;
+    unsigned mtime;
+    unsigned type;
+    char name[100];
+    char linkname[100];
+} mtar_header_t;
+
+typedef struct mtar_t mtar_t;
+
+struct mtar_t {
+    int (*read)(mtar_t *tar, void *data, unsigned size);
+    int (*write)(mtar_t *tar, const void *data, unsigned size);
+    int (*seek)(mtar_t *tar, unsigned pos);
+    int (*close)(mtar_t *tar);
+    void *stream;
+    unsigned pos;
+    unsigned stream_size;  // Added a field to record the size of a memory stream
+    unsigned remaining_data;
+    unsigned last_header;
+    enum { FROM_FILE, FROM_MEMORY } source_type;
+};
+
+const char *mtar_strerror(int err);
+
+int mtar_open(mtar_t *tar, const char *filename, const char *mode);
+int mtar_close(mtar_t *tar);
+int mtar_open_memory(mtar_t *tar, void *data, size_t size);
+
+int mtar_seek(mtar_t *tar, unsigned pos);
+int mtar_rewind(mtar_t *tar);
+int mtar_next(mtar_t *tar);
+int mtar_find(mtar_t *tar, const char *name, mtar_header_t *h);
+int mtar_read_header(mtar_t *tar, mtar_header_t *h);
+int mtar_read_data(mtar_t *tar, void *ptr, unsigned size);
+
+int mtar_write_header(mtar_t *tar, const mtar_header_t *h);
+int mtar_write_file_header(mtar_t *tar, const char *name, unsigned size);
+int mtar_write_dir_header(mtar_t *tar, const char *name);
+int mtar_write_data(mtar_t *tar, const void *data, unsigned size);
+int mtar_finalize(mtar_t *tar);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/simple_archive.h b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/simple_archive.h
new file mode 100644
index 000000000..f0a4a3ce4
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/core_archive/simple_archive.h
@@ -0,0 +1,161 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef MODELLOADERTAR_CPP_MICROTAR_H
+#define MODELLOADERTAR_CPP_MICROTAR_H
+
+#include "memory"
+#include <unordered_map>
+#include <vector>
+#include <iostream>
+#include <utility>
+#include "microtar/microtar.h"
+#include "log.h"
+
+namespace inspire {
+
+enum {
+    SARC_SUCCESS = 0,
+    SARC_FAILURE = -1,
+    SARC_OPEN_FAIL = -2,
+    SARC_READ_FAIL = -3,
+    SARC_WRITE_FAIL = -4,
+    SARC_SEEK_FAIL = -5,
+    SARC_BAD_CHKSUM = -6,
+    SARC_NULL_RECORD = -7,
+    SARC_NOTFOUND = -8,
+    SARC_LOAD_FILE_FAIL = -9,
+    SARC_NOT_LOAD = -10,
+};
+
+class INSPIRE_API SimpleArchive {
+public:
+    explicit SimpleArchive(const std::string& archiveFile) {
+        this->Reset(archiveFile);
+    }
+
+    explicit SimpleArchive() {
+        m_load_file_status_ = SARC_NOT_LOAD;
+    }
+
+    int32_t Reset(const std::string& archiveFile) {
+        Close();
+        std::vector<char>().swap(m_empty_);
+        m_tar_ = std::make_shared<mtar_t>();
+        m_load_file_status_ = mtar_open(m_tar_.get(), archiveFile.c_str(), "r");
+        if (m_load_file_status_ != MTAR_ESUCCESS) {
+            INSPIRE_LOGE("Invalid archive file: %d", m_load_file_status_);
+            Close();
+            m_tar_.reset();
+            return m_load_file_status_;
+        }
+        mtar_header_t h;
+        m_load_file_status_ = mtar_read_header(m_tar_.get(), &h);
+        if (m_load_file_status_ != MTAR_ESUCCESS) {
+            INSPIRE_LOGE("Error reading root from archive.");
+            return m_load_file_status_;
+        }
+        //        m_file_archive_root_ = std::string(h.name);
+        size_t index = 0;
+        while ((mtar_read_header(m_tar_.get(), &h)) != MTAR_ENULLRECORD) {
+            m_load_file_status_ = mtar_next(m_tar_.get());
+            if (m_load_file_status_ != MTAR_ESUCCESS) {
+                INSPIRE_LOGE("Failed to scan the file: %d", m_load_file_status_);
+                Close();
+                return m_load_file_status_;
+            }
+            m_subfiles_names_.emplace_back(h.name);
+            index++;
+        }
+        return m_load_file_status_;
+    }
+
+    std::vector<char>& GetFileContent(const std::string& filename) {
+        auto index = filenameFuzzyMatching(filename);
+        if (index != std::string::npos) {
+            auto fullFilename = m_subfiles_names_[index];
+            auto ret = lazyReadFile(fullFilename);
+            if (ret != MTAR_ESUCCESS) {
+                INSPIRE_LOGE("Failed to load file");
+            }
+            return m_file_content_cache_map_[fullFilename];
+        }
+        return m_empty_;
+    }
+
+    int32_t QueryLoadStatus() const {
+        return m_load_file_status_;
+    }
+
+    const std::vector<std::string>& GetSubfilesNames() const {
+        return m_subfiles_names_;
+    }
+
+    ~SimpleArchive() {
+        Close();
+    }
+
+    void Close() {
+        if (m_tar_.get() != nullptr) {
+            mtar_close(m_tar_.get());
+        }
+        m_tar_.reset();
+        m_load_file_status_ = SARC_NOT_LOAD;
+        m_subfiles_names_.clear();
+    }
+
+    void PrintSubFiles() {
+        std::cout << "Subfiles: " << m_subfiles_names_.size() << std::endl;
+        for (int i = 0; i < m_subfiles_names_.size(); ++i) {
+            std::cout << m_subfiles_names_[i] << std::endl;
+        }
+    }
+
+protected:
+    size_t filenameFuzzyMatching(const std::string& filename) {
+        for (size_t i = 0; i < m_subfiles_names_.size(); ++i) {
+            if (m_subfiles_names_[i].find(filename) != std::string::npos) {
+                return i;
+            }
+        }
+        return std::string::npos;
+    }
+
+    int32_t lazyReadFile(const std::string& filename) {
+        if (m_file_content_cache_map_.find(filename) != m_file_content_cache_map_.end()) {
+            return MTAR_ESUCCESS;
+        }
+        mtar_header_t h;
+        auto ret = mtar_find(m_tar_.get(), filename.c_str(), &h);
+        if (ret == MTAR_ESUCCESS) {
+            std::vector<char> content(h.size);
+            ret = mtar_read_data(m_tar_.get(), content.data(), h.size);
+            if (ret == MTAR_ESUCCESS) {
+                m_file_content_cache_map_[filename] = std::move(content);  // Load and store the file contents
+                return MTAR_ESUCCESS;
+            } else {
+                INSPIRE_LOGE("Failed to load file: %d", ret);
+            }
+        } else {
+            INSPIRE_LOGE("Failed to find file: %d", ret);
+        }
+
+        return SARC_LOAD_FILE_FAIL;
+    }
+
+private:
+    std::string m_file_archive_root_;            ///< Archive file path
+    std::vector<std::string> m_subfiles_names_;  ///< Name list of subfiles
+    std::shared_ptr<mtar_t> m_tar_;              ///< mtar context
+    int32_t m_load_file_status_;                 ///< Initiation status code
+
+    std::vector<char> m_empty_;  ///< Const empty
+
+    std::unordered_map<std::string, std::vector<char>> m_file_content_cache_map_;  ///< File buffer cache
+};
+
+}  // namespace inspire
+
+#endif  // MODELLOADERTAR_CPP_MICROTAR_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_archive.h b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_archive.h
new file mode 100644
index 000000000..57e1684ba
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_archive.h
@@ -0,0 +1,216 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef MODELLOADERTAR_INSPIREARCHIVE_H
+#define MODELLOADERTAR_INSPIREARCHIVE_H
+
+#include "core_archive/core_archive.h"
+#include "inspire_model/inspire_model.h"
+#include "yaml-cpp/yaml.h"
+#include "fstream"
+#include "similarity_converter.h"
+#include "launch.h"
+#include "track_module/landmark/landmark_param.h"
+
+namespace inspire {
+
+enum {
+    MISS_MANIFEST = -11,
+    FORMAT_ERROR = -12,
+    NOT_MATCH_MODEL = -13,
+    ERROR_MODEL_BUFFER = -14,
+    NOT_READ = -15,
+};
+
+class INSPIRE_API InspireArchive {
+public:
+    InspireArchive() : m_archive_(std::make_shared<CoreArchive>()) {
+        m_status_ = NOT_READ;
+    }
+
+    explicit InspireArchive(const std::string& archiveFile) : m_archive_(std::make_shared<CoreArchive>(archiveFile)) {
+        m_status_ = m_archive_->QueryLoadStatus();
+        if (m_status_ == SARC_SUCCESS) {
+            m_status_ = loadManifestFile();
+        }
+    }
+
+    InspireArchive(const InspireArchive& other)
+    : m_archive_(other.m_archive_),
+      m_config_(other.m_config_),
+      m_status_(other.m_status_),
+      m_tag_(other.m_tag_),
+      m_version_(other.m_version_),
+      m_major_(other.m_major_),
+      m_release_time_(other.m_release_time_) {}
+
+    InspireArchive& operator=(const InspireArchive& other) {
+        if (this != &other) {
+            m_archive_ = other.m_archive_;
+            m_config_ = other.m_config_;
+            m_status_ = other.m_status_;
+            m_tag_ = other.m_tag_;
+            m_version_ = other.m_version_;
+            m_major_ = other.m_major_;
+            m_release_time_ = other.m_release_time_;
+        }
+        return *this;
+    }
+
+    int32_t ReLoad(const std::string& archiveFile) {
+        auto ret = m_archive_->Reset(archiveFile);
+        if (ret != SARC_SUCCESS) {
+            m_archive_->Close();
+            m_status_ = ret;
+            return ret;
+        }
+        m_status_ = loadManifestFile();
+        return m_status_;
+    }
+
+    int32_t QueryStatus() const {
+        return m_status_;
+    }
+
+    int32_t LoadModel(const std::string& name, InspireModel& model) {
+        if (m_config_[name]) {
+            auto ret = model.Reset(m_config_[name]);
+            if (ret != 0) {
+                return ret;
+            }
+            if (model.loadFilePath) {
+                // No model files are loaded, only configuration files are loaded for extension modules such as CoreML.
+                return SARC_SUCCESS;
+            }
+            auto& buffer = m_archive_->GetFileContent(model.name);
+            if (buffer.empty()) {
+                return ERROR_MODEL_BUFFER;
+            }
+            model.SetBuffer(buffer, buffer.size());
+            return SARC_SUCCESS;
+        } else {
+            return NOT_MATCH_MODEL;
+        }
+    }
+
+    void PrintSubFiles() {
+        m_archive_->PrintSubFiles();
+    }
+
+    const std::vector<std::string>& GetSubfilesNames() const {
+        return m_archive_->GetSubfilesNames();
+    }
+
+    void Release() {
+        m_status_ = NOT_READ;
+        m_archive_->Close();
+    }
+
+    std::vector<char>& GetFileContent(const std::string& filename) {
+        return m_archive_->GetFileContent(filename);
+    }
+
+    const std::vector<int>& GetFaceDetectPixelList() const {
+        return m_face_detect_pixel_list_;
+    }
+
+    const std::vector<std::string>& GetFaceDetectModelList() const {
+        return m_face_detect_model_list_;
+    }
+
+    const std::shared_ptr<LandmarkParam>& GetLandmarkParam() const {
+        return m_landmark_param_;
+    }
+private:
+    int32_t loadManifestFile() {
+        if (m_archive_->QueryLoadStatus() == SARC_SUCCESS) {
+            auto configBuffer = m_archive_->GetFileContent(MANIFEST_FILE);
+            configBuffer.push_back('\0');
+            if (configBuffer.empty()) {
+                return MISS_MANIFEST;
+            }
+            m_config_ = YAML::Load(configBuffer.data());
+            if (!m_config_["tag"] || !m_config_["version"]) {
+                return FORMAT_ERROR;
+            }
+            m_tag_ = m_config_["tag"].as<std::string>();
+            m_version_ = m_config_["version"].as<std::string>();
+            if (m_config_["major"]) {
+                m_major_ = m_config_["major"].as<std::string>();
+            } else {
+                m_major_ = "unknown";
+            }
+            if (m_config_["release"]) {
+                m_release_time_ = m_config_["release"].as<std::string>();
+            } else {
+                m_release_time_ = "unknown";
+            }
+            INSPIRE_LOGI("== Load %s-%s, Version: %s, Release: %s ==", m_tag_.c_str(), m_major_.c_str(), m_version_.c_str(), m_release_time_.c_str());
+            // Load similarity converter config
+            if (m_config_["similarity_converter"]) {
+                SimilarityConverterConfig config;
+                config.threshold = m_config_["similarity_converter"]["threshold"].as<double>();
+                config.middleScore = m_config_["similarity_converter"]["middle_score"].as<double>();
+                config.steepness = m_config_["similarity_converter"]["steepness"].as<double>();
+                config.outputMin = m_config_["similarity_converter"]["output_min"].as<double>();
+                config.outputMax = m_config_["similarity_converter"]["output_max"].as<double>();
+                SIMILARITY_CONVERTER_UPDATE_CONFIG(config);
+                INSPIRE_LOGI(
+                  "Successfully loaded similarity converter config: \n \t threshold: %f \n \t middle_score: %f \n \t steepness: %f \n \t output_min: "
+                  "%f \n \t output_max: %f",
+                  config.threshold, config.middleScore, config.steepness, config.outputMin, config.outputMax);
+                SIMILARITY_CONVERTER_SET_RECOMMENDED_COSINE_THRESHOLD(config.threshold);
+            } else {
+                INSPIRE_LOGW("No similarity converter config found, use default config: ");
+                auto config = SIMILARITY_CONVERTER_GET_CONFIG();
+                INSPIRE_LOGI("threshold: %f \n \t middle_score: %f \n \t steepness: %f \n \t output_min: %f \n \t output_max: %f", config.threshold,
+                             config.middleScore, config.steepness, config.outputMin, config.outputMax);
+                SIMILARITY_CONVERTER_SET_RECOMMENDED_COSINE_THRESHOLD(config.threshold);
+            }
+            // Load face detect model
+            if (m_config_["face_detect_pixel_list"] && m_config_["face_detect_model_list"]) {
+                auto node_face_detect_pixel_list = m_config_["face_detect_pixel_list"];
+                for (std::size_t i = 0; i < node_face_detect_pixel_list.size(); ++i) {
+                    m_face_detect_pixel_list_.push_back(node_face_detect_pixel_list[i].as<int>());
+                }
+                auto node_face_detect_model_list = m_config_["face_detect_model_list"];
+                for (std::size_t i = 0; i < node_face_detect_model_list.size(); ++i) {
+                    m_face_detect_model_list_.push_back(node_face_detect_model_list[i].as<std::string>());
+                }
+                if (m_face_detect_pixel_list_.size() != m_face_detect_model_list_.size()) {
+                    return FORMAT_ERROR;
+                }
+            } else {
+                m_face_detect_pixel_list_ = {160, 320, 640};
+                m_face_detect_model_list_ = {"face_detect_160", "face_detect_320", "face_detect_640"};
+            }
+            m_landmark_param_ = std::make_shared<LandmarkParam>(m_config_["landmark_table"]);
+        }
+        return 0;
+    }
+
+
+private:
+    std::shared_ptr<CoreArchive> m_archive_;
+    YAML::Node m_config_;
+
+    int32_t m_status_;
+
+    const std::string MANIFEST_FILE = "__inspire__";
+
+    std::string m_tag_;
+    std::string m_version_;
+    std::string m_major_;
+    std::string m_release_time_;
+
+    std::vector<int> m_face_detect_pixel_list_;
+    std::vector<std::string> m_face_detect_model_list_;
+
+    std::shared_ptr<LandmarkParam> m_landmark_param_;
+};
+
+}  // namespace inspire
+
+#endif  // MODELLOADERTAR_INSPIREARCHIVE_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_model/inspire_model.h b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_model/inspire_model.h
new file mode 100644
index 000000000..228fe2c2c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/model_archive/inspire_model/inspire_model.h
@@ -0,0 +1,236 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef MODELLOADERTAR_INSPIRE_MODEL_H
+#define MODELLOADERTAR_INSPIRE_MODEL_H
+#include <iostream>
+#include "vector"
+#include "yaml-cpp/yaml.h"
+#include "middleware/configurable.h"
+#include "log.h"
+#include "middleware/inference_wrapper/inference_wrapper.h"
+
+namespace inspire {
+
+typedef enum {
+    InspireInferBackendAuto = 10,
+    InspireInferBackendCPU = 0,
+    InspireInferBackendRKNPU = 1,
+    InspireInferBackendCUDA = 2,
+} InspireInferBackend;
+
+typedef enum {
+    InspireInferEngineMNN = 0,
+    InspireInferEngineRKNN = 1,
+    InspireInferEngineCoreML = 2,
+    InspireInferEngineTensorRT = 3,
+} InspireInferEngine;
+
+class INSPIRE_API InspireModel {
+    CONFIGURABLE_SUPPORT
+public:
+    explicit InspireModel(const YAML::Node &node) {
+        Reset(node);
+    }
+
+    explicit InspireModel() {};
+
+    int32_t Reset(const YAML::Node &node) {
+        buffer = nullptr;
+        bufferSize = 0;
+        if (node["name"]) {
+            name = node["name"].as<std::string>();
+        }
+        if (node["fullname"]) {
+            fullname = node["fullname"].as<std::string>();
+        }
+        if (node["version"]) {
+            version = node["version"].as<std::string>();
+        }
+        if (node["model_type"]) {
+            auto type = node["model_type"].as<std::string>();
+            if (type == "MNN") {
+                modelType = InferenceWrapper::INFER_MNN;
+            } else if (type == "RKNN") {
+                modelType = InferenceWrapper::INFER_RKNN;
+            } else if (type == "COREML") {
+                modelType = InferenceWrapper::INFER_COREML;
+                // Special handling, the binary model is not loaded by default
+                loadFilePath = 1;
+            } else if (type == "TensorRT") {
+                modelType = InferenceWrapper::INFER_TENSORRT;
+            }
+        }
+        if (node["infer_engine"]) {
+            auto type = node["infer_engine"].as<std::string>();
+            if (type == "MNN") {
+                inferEngine = InferenceWrapper::INFER_MNN;
+            } else if (type == "RKNN") {
+                inferEngine = InferenceWrapper::INFER_RKNN;
+            } else if (type == "COREML") {
+                inferEngine = InferenceWrapper::INFER_COREML;
+            } else if (type == "TensorRT") {
+                inferEngine = InferenceWrapper::INFER_TENSORRT;
+            }
+        }
+        if (node["infer_device"]) {
+            auto type = node["infer_device"].as<std::string>();
+            if (type == "MNN") {
+                inferDevice = InspireInferEngineMNN;
+            } else if (type == "RKNPU") {
+                inferDevice = InspireInferEngineRKNN;
+            } else if (type == "COREML") {
+                inferDevice = InspireInferEngineCoreML;
+            } else if (type == "CUDA") {
+                inferDevice = InspireInferEngineTensorRT;
+            }
+        }
+        if (node["infer_backend"]) {
+            auto type = node["infer_backend"].as<std::string>();
+            if (type == "CPU") {
+                inferBackend = InspireInferBackendCPU;
+            } else if (type == "RKNPU") {
+                inferBackend = InspireInferBackendRKNPU;
+            } else if (type == "AUTO") {
+                inferBackend = InspireInferBackendAuto;
+            } else if (type == "CUDA") {
+                inferBackend = InspireInferBackendCUDA;
+            }
+        }
+        return decode(node);
+    }
+
+    void print() {
+        INSPIRE_LOGD("%s", m_configuration.toString().c_str());
+    }
+
+    void SetBuffer(std::vector<char> &modelBuffer, size_t size) {
+        buffer = modelBuffer.data();
+        bufferSize = size;
+    }
+
+    Configurable &Config() {
+        return m_configuration;
+    }
+
+private:
+    int32_t decode(const YAML::Node &node) {
+        try {
+            if (node["input_channel"]) {
+                setData<int>("input_channel", node["input_channel"].as<int>());
+            }
+            if (node["input_image_channel"]) {
+                setData<int>("input_image_channel", node["input_image_channel"].as<int>());
+            }
+            if (node["nchw"]) {
+                setData<bool>("nchw", node["nchw"].as<bool>());
+            }
+            if (node["swap_color"]) {
+                setData<bool>("swap_color", node["swap_color"].as<bool>());
+            }
+            if (node["data_type"]) {
+                auto type = node["data_type"].as<std::string>();
+                if (type == "image") {
+                    setData<int>("data_type", InputTensorInfo::InputTensorInfo::DataTypeImage);
+                } else if (type == "data_nhwc") {
+                    setData<int>("data_type", InputTensorInfo::InputTensorInfo::DataTypeBlobNhwc);
+                } else if (type == "data_nchw") {
+                    setData<int>("data_type", InputTensorInfo::InputTensorInfo::DataTypeBlobNchw);
+                }
+            }
+            if (node["input_tensor_type"]) {
+                auto type = node["input_tensor_type"].as<std::string>();
+                if (type == "none") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeNone);
+                } else if (type == "uint8") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeUint8);
+                } else if (type == "int8") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt8);
+                } else if (type == "float32") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeFp32);
+                } else if (type == "int32") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt32);
+                } else if (type == "int64") {
+                    setData<int>("input_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt64);
+                }
+            }
+            if (node["output_tensor_type"]) {
+                auto type = node["output_tensor_type"].as<std::string>();
+                if (type == "none") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeNone);
+                } else if (type == "uint8") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeUint8);
+                } else if (type == "int8") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt8);
+                } else if (type == "float32") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeFp32);
+                } else if (type == "int32") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt32);
+                } else if (type == "int64") {
+                    setData<int>("output_tensor_type", InputTensorInfo::TensorInfo::TensorTypeInt64);
+                }
+            }
+            if (node["threads"]) {
+                setData<int>("threads", node["threads"].as<int>());
+            }
+            if (node["input_layer"]) {
+                setData<std::string>("input_layer", node["input_layer"].as<std::string>());
+            }
+            if (node["outputs_layers"]) {
+                auto n = node["outputs_layers"];
+                std::vector<std::string> names;
+                for (std::size_t i = 0; i < n.size(); ++i) {
+                    names.push_back(n[i].as<std::string>());
+                }
+                setData<std::vector<std::string>>("outputs_layers", names);
+            }
+            if (node["input_size"]) {
+                auto n = node["input_size"];
+                std::vector<int> size;
+                for (std::size_t i = 0; i < n.size(); ++i) {
+                    size.push_back(n[i].as<int>());
+                }
+                setData<std::vector<int>>("input_size", size);
+            }
+            if (node["mean"]) {
+                auto n = node["mean"];
+                std::vector<float> mean;
+                for (std::size_t i = 0; i < n.size(); ++i) {
+                    mean.push_back(n[i].as<float>());
+                }
+                setData<std::vector<float>>("mean", mean);
+            }
+            if (node["norm"]) {
+                auto n = node["norm"];
+                std::vector<float> norm;
+                for (std::size_t i = 0; i < n.size(); ++i) {
+                    norm.push_back(n[i].as<float>());
+                }
+                setData<std::vector<float>>("norm", norm);
+            }
+        } catch (const YAML::Exception &e) {
+            INSPIRE_LOGE("An error occurred parsing the interpretation file in archive: %s", e.what());
+            return -1;
+        }
+        return 0;
+    }
+
+public:
+    std::string name;
+    std::string fullname;
+    std::string version;
+    InferenceWrapper::EngineType modelType;
+    int inferEngine;
+    int inferDevice;
+    int inferBackend;
+    int loadFilePath{0};
+
+    char *buffer;
+    size_t bufferSize;
+};
+
+};  // namespace inspire
+
+#endif  // MODELLOADERTAR_INSPIRE_MODEL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/nlohmann/json.hpp b/cpp-package/inspireface/cpp/inspireface/middleware/nlohmann/json.hpp
new file mode 100644
index 000000000..e7ee0cb53
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/nlohmann/json.hpp
@@ -0,0 +1,22091 @@
+/*
+    __ _____ _____ _____
+ __|  |   __|     |   | |  JSON for Modern C++
+|  |  |__   |  |  | | | |  version 3.10.5
+|_____|_____|_____|_|___|  https://github.com/nlohmann/json
+
+Licensed under the MIT License <http://opensource.org/licenses/MIT>.
+SPDX-License-Identifier: MIT
+Copyright (c) 2013-2022 Niels Lohmann <http://nlohmann.me>.
+
+Permission is hereby  granted, free of charge, to any  person obtaining a copy
+of this software and associated  documentation files (the "Software"), to deal
+in the Software  without restriction, including without  limitation the rights
+to  use, copy,  modify, merge,  publish, distribute,  sublicense, and/or  sell
+copies  of  the Software,  and  to  permit persons  to  whom  the Software  is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE  IS PROVIDED "AS  IS", WITHOUT WARRANTY  OF ANY KIND,  EXPRESS OR
+IMPLIED,  INCLUDING BUT  NOT  LIMITED TO  THE  WARRANTIES OF  MERCHANTABILITY,
+FITNESS FOR  A PARTICULAR PURPOSE AND  NONINFRINGEMENT. IN NO EVENT  SHALL THE
+AUTHORS  OR COPYRIGHT  HOLDERS  BE  LIABLE FOR  ANY  CLAIM,  DAMAGES OR  OTHER
+LIABILITY, WHETHER IN AN ACTION OF  CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE  OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/
+
+/****************************************************************************\
+ * Note on documentation: The source files contain links to the online      *
+ * documentation of the public API at https://json.nlohmann.me. This URL    *
+ * contains the most recent documentation and should also be applicable to  *
+ * previous versions; documentation for deprecated functions is not         *
+ * removed, but marked deprecated. See "Generate documentation" section in  *
+ * file doc/README.md.                                                      *
+\****************************************************************************/
+
+#ifndef INCLUDE_NLOHMANN_JSON_HPP_
+#define INCLUDE_NLOHMANN_JSON_HPP_
+
+#define NLOHMANN_JSON_VERSION_MAJOR 3
+#define NLOHMANN_JSON_VERSION_MINOR 10
+#define NLOHMANN_JSON_VERSION_PATCH 5
+
+#include <algorithm> // all_of, find, for_each
+#include <cstddef> // nullptr_t, ptrdiff_t, size_t
+#include <functional> // hash, less
+#include <initializer_list> // initializer_list
+#ifndef JSON_NO_IO
+    #include <iosfwd> // istream, ostream
+#endif  // JSON_NO_IO
+#include <iterator> // random_access_iterator_tag
+#include <memory> // unique_ptr
+#include <numeric> // accumulate
+#include <string> // string, stoi, to_string
+#include <utility> // declval, forward, move, pair, swap
+#include <vector> // vector
+
+// #include <nlohmann/adl_serializer.hpp>
+
+
+#include <type_traits>
+#include <utility>
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+
+
+#include <algorithm> // transform
+#include <array> // array
+#include <forward_list> // forward_list
+#include <iterator> // inserter, front_inserter, end
+#include <map> // map
+#include <string> // string
+#include <tuple> // tuple, make_tuple
+#include <type_traits> // is_arithmetic, is_same, is_enum, underlying_type, is_convertible
+#include <unordered_map> // unordered_map
+#include <utility> // pair, declval
+#include <valarray> // valarray
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+
+#include <exception> // exception
+#include <stdexcept> // runtime_error
+#include <string> // to_string
+#include <vector> // vector
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t
+#include <string> // string
+
+namespace nlohmann
+{
+namespace detail
+{
+///////////////////////////
+// JSON type enumeration //
+///////////////////////////
+
+/*!
+@brief the JSON type enumeration
+
+This enumeration collects the different JSON types. It is internally used to
+distinguish the stored values, and the functions @ref basic_json::is_null(),
+@ref basic_json::is_object(), @ref basic_json::is_array(),
+@ref basic_json::is_string(), @ref basic_json::is_boolean(),
+@ref basic_json::is_number() (with @ref basic_json::is_number_integer(),
+@ref basic_json::is_number_unsigned(), and @ref basic_json::is_number_float()),
+@ref basic_json::is_discarded(), @ref basic_json::is_primitive(), and
+@ref basic_json::is_structured() rely on it.
+
+@note There are three enumeration entries (number_integer, number_unsigned, and
+number_float), because the library distinguishes these three types for numbers:
+@ref basic_json::number_unsigned_t is used for unsigned integers,
+@ref basic_json::number_integer_t is used for signed integers, and
+@ref basic_json::number_float_t is used for floating-point numbers or to
+approximate integers which do not fit in the limits of their respective type.
+
+@sa see @ref basic_json::basic_json(const value_t value_type) -- create a JSON
+value with the default value for a given type
+
+@since version 1.0.0
+*/
+enum class value_t : std::uint8_t
+{
+    null,             ///< null value
+    object,           ///< object (unordered set of name/value pairs)
+    array,            ///< array (ordered collection of values)
+    string,           ///< string value
+    boolean,          ///< boolean value
+    number_integer,   ///< number value (signed integer)
+    number_unsigned,  ///< number value (unsigned integer)
+    number_float,     ///< number value (floating-point)
+    binary,           ///< binary array (ordered collection of bytes)
+    discarded         ///< discarded by the parser callback function
+};
+
+/*!
+@brief comparison operator for JSON types
+
+Returns an ordering that is similar to Python:
+- order: null < boolean < number < object < array < string < binary
+- furthermore, each type is not smaller than itself
+- discarded values are not comparable
+- binary is represented as a b"" string in python and directly comparable to a
+  string; however, making a binary array directly comparable with a string would
+  be surprising behavior in a JSON file.
+
+@since version 1.0.0
+*/
+inline bool operator<(const value_t lhs, const value_t rhs) noexcept
+{
+    static constexpr std::array<std::uint8_t, 9> order = {{
+            0 /* null */, 3 /* object */, 4 /* array */, 5 /* string */,
+            1 /* boolean */, 2 /* integer */, 2 /* unsigned */, 2 /* float */,
+            6 /* binary */
+        }
+    };
+
+    const auto l_index = static_cast<std::size_t>(lhs);
+    const auto r_index = static_cast<std::size_t>(rhs);
+    return l_index < order.size() && r_index < order.size() && order[l_index] < order[r_index];
+}
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+
+#include <string>
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+#include <utility> // declval, pair
+// #include <nlohmann/thirdparty/hedley/hedley.hpp>
+
+
+/* Hedley - https://nemequ.github.io/hedley
+ * Created by Evan Nemerson <evan@nemerson.com>
+ *
+ * To the extent possible under law, the author(s) have dedicated all
+ * copyright and related and neighboring rights to this software to
+ * the public domain worldwide. This software is distributed without
+ * any warranty.
+ *
+ * For details, see <http://creativecommons.org/publicdomain/zero/1.0/>.
+ * SPDX-License-Identifier: CC0-1.0
+ */
+
+#if !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < 15)
+#if defined(JSON_HEDLEY_VERSION)
+    #undef JSON_HEDLEY_VERSION
+#endif
+#define JSON_HEDLEY_VERSION 15
+
+#if defined(JSON_HEDLEY_STRINGIFY_EX)
+    #undef JSON_HEDLEY_STRINGIFY_EX
+#endif
+#define JSON_HEDLEY_STRINGIFY_EX(x) #x
+
+#if defined(JSON_HEDLEY_STRINGIFY)
+    #undef JSON_HEDLEY_STRINGIFY
+#endif
+#define JSON_HEDLEY_STRINGIFY(x) JSON_HEDLEY_STRINGIFY_EX(x)
+
+#if defined(JSON_HEDLEY_CONCAT_EX)
+    #undef JSON_HEDLEY_CONCAT_EX
+#endif
+#define JSON_HEDLEY_CONCAT_EX(a,b) a##b
+
+#if defined(JSON_HEDLEY_CONCAT)
+    #undef JSON_HEDLEY_CONCAT
+#endif
+#define JSON_HEDLEY_CONCAT(a,b) JSON_HEDLEY_CONCAT_EX(a,b)
+
+#if defined(JSON_HEDLEY_CONCAT3_EX)
+    #undef JSON_HEDLEY_CONCAT3_EX
+#endif
+#define JSON_HEDLEY_CONCAT3_EX(a,b,c) a##b##c
+
+#if defined(JSON_HEDLEY_CONCAT3)
+    #undef JSON_HEDLEY_CONCAT3
+#endif
+#define JSON_HEDLEY_CONCAT3(a,b,c) JSON_HEDLEY_CONCAT3_EX(a,b,c)
+
+#if defined(JSON_HEDLEY_VERSION_ENCODE)
+    #undef JSON_HEDLEY_VERSION_ENCODE
+#endif
+#define JSON_HEDLEY_VERSION_ENCODE(major,minor,revision) (((major) * 1000000) + ((minor) * 1000) + (revision))
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MAJOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MAJOR(version) ((version) / 1000000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MINOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MINOR(version) (((version) % 1000000) / 1000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_REVISION)
+    #undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_REVISION(version) ((version) % 1000)
+
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #undef JSON_HEDLEY_GNUC_VERSION
+#endif
+#if defined(__GNUC__) && defined(__GNUC_PATCHLEVEL__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__)
+#elif defined(__GNUC__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, 0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GNUC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION)
+    #undef JSON_HEDLEY_MSVC_VERSION
+#endif
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 140000000) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 10000000, (_MSC_FULL_VER % 10000000) / 100000, (_MSC_FULL_VER % 100000) / 100)
+#elif defined(_MSC_FULL_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 1000000, (_MSC_FULL_VER % 1000000) / 10000, (_MSC_FULL_VER % 10000) / 10)
+#elif defined(_MSC_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_VER / 100, _MSC_VER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#endif
+#if !defined(JSON_HEDLEY_MSVC_VERSION)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (0)
+#elif defined(_MSC_VER) && (_MSC_VER >= 1400)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 10000000) + (minor * 100000) + (patch)))
+#elif defined(_MSC_VER) && (_MSC_VER >= 1200)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 1000000) + (minor * 10000) + (patch)))
+#else
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_VER >= ((major * 100) + (minor)))
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #undef JSON_HEDLEY_INTEL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, __INTEL_COMPILER_UPDATE)
+#elif defined(__INTEL_COMPILER) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && defined(__ICL)
+    #define JSON_HEDLEY_INTEL_CL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER, __INTEL_COMPILER_UPDATE, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_CL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #undef JSON_HEDLEY_PGI_VERSION
+#endif
+#if defined(__PGI) && defined(__PGIC__) && defined(__PGIC_MINOR__) && defined(__PGIC_PATCHLEVEL__)
+    #define JSON_HEDLEY_PGI_VERSION JSON_HEDLEY_VERSION_ENCODE(__PGIC__, __PGIC_MINOR__, __PGIC_PATCHLEVEL__)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION_CHECK)
+    #undef JSON_HEDLEY_PGI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PGI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #undef JSON_HEDLEY_SUNPRO_VERSION
+#endif
+#if defined(__SUNPRO_C) && (__SUNPRO_C > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_C >> 16) & 0xf) * 10) + ((__SUNPRO_C >> 12) & 0xf), (((__SUNPRO_C >> 8) & 0xf) * 10) + ((__SUNPRO_C >> 4) & 0xf), (__SUNPRO_C & 0xf) * 10)
+#elif defined(__SUNPRO_C)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_C >> 8) & 0xf, (__SUNPRO_C >> 4) & 0xf, (__SUNPRO_C) & 0xf)
+#elif defined(__SUNPRO_CC) && (__SUNPRO_CC > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_CC >> 16) & 0xf) * 10) + ((__SUNPRO_CC >> 12) & 0xf), (((__SUNPRO_CC >> 8) & 0xf) * 10) + ((__SUNPRO_CC >> 4) & 0xf), (__SUNPRO_CC & 0xf) * 10)
+#elif defined(__SUNPRO_CC)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_CC >> 8) & 0xf, (__SUNPRO_CC >> 4) & 0xf, (__SUNPRO_CC) & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION_CHECK)
+    #undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_SUNPRO_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#endif
+#if defined(__EMSCRIPTEN__)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION JSON_HEDLEY_VERSION_ENCODE(__EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_EMSCRIPTEN_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #undef JSON_HEDLEY_ARM_VERSION
+#endif
+#if defined(__CC_ARM) && defined(__ARMCOMPILER_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCOMPILER_VERSION / 1000000, (__ARMCOMPILER_VERSION % 1000000) / 10000, (__ARMCOMPILER_VERSION % 10000) / 100)
+#elif defined(__CC_ARM) && defined(__ARMCC_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCC_VERSION / 1000000, (__ARMCC_VERSION % 1000000) / 10000, (__ARMCC_VERSION % 10000) / 100)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION_CHECK)
+    #undef JSON_HEDLEY_ARM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_ARM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #undef JSON_HEDLEY_IBM_VERSION
+#endif
+#if defined(__ibmxl__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ibmxl_version__, __ibmxl_release__, __ibmxl_modification__)
+#elif defined(__xlC__) && defined(__xlC_ver__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, (__xlC_ver__ >> 8) & 0xff)
+#elif defined(__xlC__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, 0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION_CHECK)
+    #undef JSON_HEDLEY_IBM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IBM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #undef JSON_HEDLEY_TI_VERSION
+#endif
+#if \
+    defined(__TI_COMPILER_VERSION__) && \
+    ( \
+      defined(__TMS470__) || defined(__TI_ARM__) || \
+      defined(__MSP430__) || \
+      defined(__TMS320C2000__) \
+    )
+#if (__TI_COMPILER_VERSION__ >= 16000000)
+    #define JSON_HEDLEY_TI_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C2000__)
+    #define JSON_HEDLEY_TI_CL2000_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL2000_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #undef JSON_HEDLEY_TI_CL430_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__MSP430__)
+    #define JSON_HEDLEY_TI_CL430_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL430_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && (defined(__TMS470__) || defined(__TI_ARM__))
+    #define JSON_HEDLEY_TI_ARMCL_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_ARMCL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C6X__)
+    #define JSON_HEDLEY_TI_CL6X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL6X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__C7000__)
+    #define JSON_HEDLEY_TI_CL7X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL7X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__PRU__)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CLPRU_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #undef JSON_HEDLEY_CRAY_VERSION
+#endif
+#if defined(_CRAYC)
+    #if defined(_RELEASE_PATCHLEVEL)
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, _RELEASE_PATCHLEVEL)
+    #else
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION_CHECK)
+    #undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_CRAY_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #undef JSON_HEDLEY_IAR_VERSION
+#endif
+#if defined(__IAR_SYSTEMS_ICC__)
+    #if __VER__ > 1000
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE((__VER__ / 1000000), ((__VER__ / 1000) % 1000), (__VER__ % 1000))
+    #else
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE(__VER__ / 100, __VER__ % 100, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION_CHECK)
+    #undef JSON_HEDLEY_IAR_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IAR_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #undef JSON_HEDLEY_TINYC_VERSION
+#endif
+#if defined(__TINYC__)
+    #define JSON_HEDLEY_TINYC_VERSION JSON_HEDLEY_VERSION_ENCODE(__TINYC__ / 1000, (__TINYC__ / 100) % 10, __TINYC__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION_CHECK)
+    #undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TINYC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #undef JSON_HEDLEY_DMC_VERSION
+#endif
+#if defined(__DMC__)
+    #define JSON_HEDLEY_DMC_VERSION JSON_HEDLEY_VERSION_ENCODE(__DMC__ >> 8, (__DMC__ >> 4) & 0xf, __DMC__ & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION_CHECK)
+    #undef JSON_HEDLEY_DMC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_DMC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #undef JSON_HEDLEY_COMPCERT_VERSION
+#endif
+#if defined(__COMPCERT_VERSION__)
+    #define JSON_HEDLEY_COMPCERT_VERSION JSON_HEDLEY_VERSION_ENCODE(__COMPCERT_VERSION__ / 10000, (__COMPCERT_VERSION__ / 100) % 100, __COMPCERT_VERSION__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION_CHECK)
+    #undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_COMPCERT_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #undef JSON_HEDLEY_PELLES_VERSION
+#endif
+#if defined(__POCC__)
+    #define JSON_HEDLEY_PELLES_VERSION JSON_HEDLEY_VERSION_ENCODE(__POCC__ / 100, __POCC__ % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION_CHECK)
+    #undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PELLES_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION
+#endif
+#if defined(__LCC__) && defined(__LCC_MINOR__)
+    #define JSON_HEDLEY_MCST_LCC_VERSION JSON_HEDLEY_VERSION_ENCODE(__LCC__ / 100, __LCC__ % 100, __LCC_MINOR__)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_MCST_LCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #undef JSON_HEDLEY_GCC_VERSION
+#endif
+#if \
+    defined(JSON_HEDLEY_GNUC_VERSION) && \
+    !defined(__clang__) && \
+    !defined(JSON_HEDLEY_INTEL_VERSION) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_ARM_VERSION) && \
+    !defined(JSON_HEDLEY_CRAY_VERSION) && \
+    !defined(JSON_HEDLEY_TI_VERSION) && \
+    !defined(JSON_HEDLEY_TI_ARMCL_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL430_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL2000_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL6X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL7X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CLPRU_VERSION) && \
+    !defined(__COMPCERT__) && \
+    !defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION JSON_HEDLEY_GNUC_VERSION
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_ATTRIBUTE
+#endif
+#if \
+  defined(__has_attribute) && \
+  ( \
+    (!defined(JSON_HEDLEY_IAR_VERSION) || JSON_HEDLEY_IAR_VERSION_CHECK(8,5,9)) \
+  )
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) __has_attribute(attribute)
+#else
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#endif
+#if \
+    defined(__has_cpp_attribute) && \
+    defined(__cplusplus) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#endif
+#if !defined(__cplusplus) || !defined(__has_cpp_attribute)
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#elif \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_IAR_VERSION) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) && \
+    (!defined(JSON_HEDLEY_MSVC_VERSION) || JSON_HEDLEY_MSVC_VERSION_CHECK(19,20,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(ns::attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_BUILTIN)
+    #undef JSON_HEDLEY_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_FEATURE)
+    #undef JSON_HEDLEY_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_HAS_FEATURE(feature) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_HAS_FEATURE(feature) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GCC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_EXTENSION)
+    #undef JSON_HEDLEY_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_WARNING)
+    #undef JSON_HEDLEY_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_HAS_WARNING(warning) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_HAS_WARNING(warning) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_WARNING)
+    #undef JSON_HEDLEY_GNUC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_WARNING)
+    #undef JSON_HEDLEY_GCC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,17) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(8,0,0) || \
+    (JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) && defined(__C99_PRAGMA_OPERATOR))
+    #define JSON_HEDLEY_PRAGMA(value) _Pragma(#value)
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_PRAGMA(value) __pragma(value)
+#else
+    #define JSON_HEDLEY_PRAGMA(value)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_PUSH)
+    #undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#endif
+#if defined(JSON_HEDLEY_DIAGNOSTIC_POP)
+    #undef JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("clang diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("clang diagnostic pop")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("GCC diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("GCC diagnostic pop")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH __pragma(warning(push))
+    #define JSON_HEDLEY_DIAGNOSTIC_POP __pragma(warning(pop))
+#elif JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("pop")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,4,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("diag_push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("diag_pop")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH
+    #define JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+
+/* JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat")
+#    if JSON_HEDLEY_HAS_WARNING("-Wc++17-extensions")
+#      if JSON_HEDLEY_HAS_WARNING("-Wc++1z-extensions")
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++1z-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      else
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      endif
+#    else
+#      define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    endif
+#  endif
+#endif
+#if !defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(x) x
+#endif
+
+#if defined(JSON_HEDLEY_CONST_CAST)
+    #undef JSON_HEDLEY_CONST_CAST
+#endif
+#if defined(__cplusplus)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (const_cast<T>(expr))
+#elif \
+  JSON_HEDLEY_HAS_WARNING("-Wcast-qual") || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_CONST_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_REINTERPRET_CAST)
+    #undef JSON_HEDLEY_REINTERPRET_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) (reinterpret_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_CAST)
+    #undef JSON_HEDLEY_STATIC_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) (static_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_CPP_CAST)
+    #undef JSON_HEDLEY_CPP_CAST
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wold-style-cast")
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wold-style-cast\"") \
+    ((T) (expr)) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  elif JSON_HEDLEY_IAR_VERSION_CHECK(8,3,0)
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("diag_suppress=Pe137") \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_CPP_CAST(T, expr) ((T) (expr))
+#  endif
+#else
+#  define JSON_HEDLEY_CPP_CAST(T, expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wdeprecated-declarations")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warning(disable:1478 1786)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:1478 1786))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1216,1444,1445")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:4996))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1291,1718")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,E_DEPRECATED_ATT,E_DEPRECATED_ATT_MESS)")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,symdeprecated,symdeprecated2)")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress=Pe1444,Pe1215")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warn(disable:2241)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("clang diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("warning(disable:161)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:161))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 1675")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:4068))
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(16,9,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress=Pe161")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 161")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-attributes")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("clang diagnostic ignored \"-Wunknown-attributes\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("warning(disable:1292)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:1292))
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:5030))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097,1098")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("error_messages(off,attrskipunsup)")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1173")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress=Pe1097")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wcast-qual")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("clang diagnostic ignored \"-Wcast-qual\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("warning(disable:2203 2331)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("GCC diagnostic ignored \"-Wcast-qual\"")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunused-function")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("clang diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("GCC diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(1,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION __pragma(warning(disable:4505))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("diag_suppress 3142")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+
+#if defined(JSON_HEDLEY_DEPRECATED)
+    #undef JSON_HEDLEY_DEPRECATED
+#endif
+#if defined(JSON_HEDLEY_DEPRECATED_FOR)
+    #undef JSON_HEDLEY_DEPRECATED_FOR
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated("Since " # since))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated("Since " #since "; use " #replacement))
+#elif \
+    (JSON_HEDLEY_HAS_EXTENSION(attribute_deprecated_with_message) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__("Since " #since)))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__("Since " #since "; use " #replacement)))
+#elif defined(__cplusplus) && (__cplusplus >= 201402L)
+    #define JSON_HEDLEY_DEPRECATED(since) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since)]])
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since "; use " #replacement)]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(deprecated) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_PELLES_VERSION_CHECK(6,50,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DEPRECATED(since) _Pragma("deprecated")
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) _Pragma("deprecated")
+#else
+    #define JSON_HEDLEY_DEPRECATED(since)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement)
+#endif
+
+#if defined(JSON_HEDLEY_UNAVAILABLE)
+    #undef JSON_HEDLEY_UNAVAILABLE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warning) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNAVAILABLE(available_since) __attribute__((__warning__("Not available until " #available_since)))
+#else
+    #define JSON_HEDLEY_UNAVAILABLE(available_since)
+#endif
+
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#endif
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT_MSG)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warn_unused_result) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__))
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) __attribute__((__warn_unused_result__))
+#elif (JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) >= 201907L)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard(msg)]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+#elif defined(_Check_return_) /* SAL */
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT _Check_return_
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) _Check_return_
+#else
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg)
+#endif
+
+#if defined(JSON_HEDLEY_SENTINEL)
+    #undef JSON_HEDLEY_SENTINEL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(sentinel) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_SENTINEL(position) __attribute__((__sentinel__(position)))
+#else
+    #define JSON_HEDLEY_SENTINEL(position)
+#endif
+
+#if defined(JSON_HEDLEY_NO_RETURN)
+    #undef JSON_HEDLEY_NO_RETURN
+#endif
+#if JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NO_RETURN __noreturn
+#elif \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
+    #define JSON_HEDLEY_NO_RETURN _Noreturn
+#elif defined(__cplusplus) && (__cplusplus >= 201103L)
+    #define JSON_HEDLEY_NO_RETURN JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[noreturn]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noreturn) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,2,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("does_not_return")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("FUNC_NEVER_RETURNS;")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute((noreturn))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#else
+    #define JSON_HEDLEY_NO_RETURN
+#endif
+
+#if defined(JSON_HEDLEY_NO_ESCAPE)
+    #undef JSON_HEDLEY_NO_ESCAPE
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(noescape)
+    #define JSON_HEDLEY_NO_ESCAPE __attribute__((__noescape__))
+#else
+    #define JSON_HEDLEY_NO_ESCAPE
+#endif
+
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #undef JSON_HEDLEY_UNREACHABLE
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE_RETURN)
+    #undef JSON_HEDLEY_UNREACHABLE_RETURN
+#endif
+#if defined(JSON_HEDLEY_ASSUME)
+    #undef JSON_HEDLEY_ASSUME
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_ASSUME(expr) __assume(expr)
+#elif JSON_HEDLEY_HAS_BUILTIN(__builtin_assume)
+    #define JSON_HEDLEY_ASSUME(expr) __builtin_assume(expr)
+#elif \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+    #if defined(__cplusplus)
+        #define JSON_HEDLEY_ASSUME(expr) std::_nassert(expr)
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) _nassert(expr)
+    #endif
+#endif
+#if \
+    (JSON_HEDLEY_HAS_BUILTIN(__builtin_unreachable) && (!defined(JSON_HEDLEY_ARM_VERSION))) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,5) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(10,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNREACHABLE() __builtin_unreachable()
+#elif defined(JSON_HEDLEY_ASSUME)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+#if !defined(JSON_HEDLEY_ASSUME)
+    #if defined(JSON_HEDLEY_UNREACHABLE)
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, ((expr) ? 1 : (JSON_HEDLEY_UNREACHABLE(), 1)))
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, expr)
+    #endif
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #if  \
+        JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (JSON_HEDLEY_STATIC_CAST(void, JSON_HEDLEY_ASSUME(0)), (value))
+    #else
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) JSON_HEDLEY_UNREACHABLE()
+    #endif
+#else
+    #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (value)
+#endif
+#if !defined(JSON_HEDLEY_UNREACHABLE)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+
+JSON_HEDLEY_DIAGNOSTIC_PUSH
+#if JSON_HEDLEY_HAS_WARNING("-Wpedantic")
+    #pragma clang diagnostic ignored "-Wpedantic"
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat-pedantic") && defined(__cplusplus)
+    #pragma clang diagnostic ignored "-Wc++98-compat-pedantic"
+#endif
+#if JSON_HEDLEY_GCC_HAS_WARNING("-Wvariadic-macros",4,0,0)
+    #if defined(__clang__)
+        #pragma clang diagnostic ignored "-Wvariadic-macros"
+    #elif defined(JSON_HEDLEY_GCC_VERSION)
+        #pragma GCC diagnostic ignored "-Wvariadic-macros"
+    #endif
+#endif
+#if defined(JSON_HEDLEY_NON_NULL)
+    #undef JSON_HEDLEY_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NON_NULL(...) __attribute__((__nonnull__(__VA_ARGS__)))
+#else
+    #define JSON_HEDLEY_NON_NULL(...)
+#endif
+JSON_HEDLEY_DIAGNOSTIC_POP
+
+#if defined(JSON_HEDLEY_PRINTF_FORMAT)
+    #undef JSON_HEDLEY_PRINTF_FORMAT
+#endif
+#if defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && !defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(ms_printf, string_idx, first_to_check)))
+#elif defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(gnu_printf, string_idx, first_to_check)))
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(format) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(__printf__, string_idx, first_to_check)))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(6,0,0)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __declspec(vaformat(printf,string_idx,first_to_check))
+#else
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check)
+#endif
+
+#if defined(JSON_HEDLEY_CONSTEXPR)
+    #undef JSON_HEDLEY_CONSTEXPR
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_CONSTEXPR JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(constexpr)
+    #endif
+#endif
+#if !defined(JSON_HEDLEY_CONSTEXPR)
+    #define JSON_HEDLEY_CONSTEXPR
+#endif
+
+#if defined(JSON_HEDLEY_PREDICT)
+    #undef JSON_HEDLEY_PREDICT
+#endif
+#if defined(JSON_HEDLEY_LIKELY)
+    #undef JSON_HEDLEY_LIKELY
+#endif
+#if defined(JSON_HEDLEY_UNLIKELY)
+    #undef JSON_HEDLEY_UNLIKELY
+#endif
+#if defined(JSON_HEDLEY_UNPREDICTABLE)
+    #undef JSON_HEDLEY_UNPREDICTABLE
+#endif
+#if JSON_HEDLEY_HAS_BUILTIN(__builtin_unpredictable)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) __builtin_unpredictable((expr))
+#endif
+#if \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect_with_probability) && !defined(JSON_HEDLEY_PGI_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(9,0,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, value, probability) __builtin_expect_with_probability(  (expr), (value), (probability))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability)   __builtin_expect_with_probability(!!(expr),    1   , (probability))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability)  __builtin_expect_with_probability(!!(expr),    0   , (probability))
+#  define JSON_HEDLEY_LIKELY(expr)                      __builtin_expect                 (!!(expr),    1                  )
+#  define JSON_HEDLEY_UNLIKELY(expr)                    __builtin_expect                 (!!(expr),    0                  )
+#elif \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,27) || \
+  JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) \
+    (((probability) >= 0.9) ? __builtin_expect((expr), (expected)) : (JSON_HEDLEY_STATIC_CAST(void, expected), (expr)))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 1) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 0) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 0) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 1) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_LIKELY(expr)   __builtin_expect(!!(expr), 1)
+#  define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect(!!(expr), 0)
+#else
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) (JSON_HEDLEY_STATIC_CAST(void, expected), (expr))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_LIKELY(expr) (!!(expr))
+#  define JSON_HEDLEY_UNLIKELY(expr) (!!(expr))
+#endif
+#if !defined(JSON_HEDLEY_UNPREDICTABLE)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) JSON_HEDLEY_PREDICT(expr, 1, 0.5)
+#endif
+
+#if defined(JSON_HEDLEY_MALLOC)
+    #undef JSON_HEDLEY_MALLOC
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(malloc) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_MALLOC __attribute__((__malloc__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_MALLOC _Pragma("returns_new_memory")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_MALLOC __declspec(restrict)
+#else
+    #define JSON_HEDLEY_MALLOC
+#endif
+
+#if defined(JSON_HEDLEY_PURE)
+    #undef JSON_HEDLEY_PURE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(pure) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(2,96,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PURE __attribute__((__pure__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+#  define JSON_HEDLEY_PURE _Pragma("does_not_write_global_data")
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) \
+    )
+#  define JSON_HEDLEY_PURE _Pragma("FUNC_IS_PURE;")
+#else
+#  define JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_CONST)
+    #undef JSON_HEDLEY_CONST
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(const) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(2,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_CONST __attribute__((__const__))
+#elif \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_CONST _Pragma("no_side_effect")
+#else
+    #define JSON_HEDLEY_CONST JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_RESTRICT)
+    #undef JSON_HEDLEY_RESTRICT
+#endif
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT restrict
+#elif \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,4) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RESTRICT __restrict
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,3,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT _Restrict
+#else
+    #define JSON_HEDLEY_RESTRICT
+#endif
+
+#if defined(JSON_HEDLEY_INLINE)
+    #undef JSON_HEDLEY_INLINE
+#endif
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    (defined(__cplusplus) && (__cplusplus >= 199711L))
+    #define JSON_HEDLEY_INLINE inline
+#elif \
+    defined(JSON_HEDLEY_GCC_VERSION) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(6,2,0)
+    #define JSON_HEDLEY_INLINE __inline__
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,1,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_INLINE __inline
+#else
+    #define JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_ALWAYS_INLINE)
+    #undef JSON_HEDLEY_ALWAYS_INLINE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(always_inline) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+  JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __attribute__((__always_inline__)) JSON_HEDLEY_INLINE
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __forceinline
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+      JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+      JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) \
+    )
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("FUNC_ALWAYS_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("inline=forced")
+#else
+#  define JSON_HEDLEY_ALWAYS_INLINE JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_NEVER_INLINE)
+    #undef JSON_HEDLEY_NEVER_INLINE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noinline) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute__((__noinline__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(10,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("noinline")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("FUNC_CANNOT_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("inline=never")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute((noinline))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#else
+    #define JSON_HEDLEY_NEVER_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_PRIVATE)
+    #undef JSON_HEDLEY_PRIVATE
+#endif
+#if defined(JSON_HEDLEY_PUBLIC)
+    #undef JSON_HEDLEY_PUBLIC
+#endif
+#if defined(JSON_HEDLEY_IMPORT)
+    #undef JSON_HEDLEY_IMPORT
+#endif
+#if defined(_WIN32) || defined(__CYGWIN__)
+#  define JSON_HEDLEY_PRIVATE
+#  define JSON_HEDLEY_PUBLIC   __declspec(dllexport)
+#  define JSON_HEDLEY_IMPORT   __declspec(dllimport)
+#else
+#  if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(visibility) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    ( \
+      defined(__TI_EABI__) && \
+      ( \
+        (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) \
+      ) \
+    ) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#    define JSON_HEDLEY_PRIVATE __attribute__((__visibility__("hidden")))
+#    define JSON_HEDLEY_PUBLIC  __attribute__((__visibility__("default")))
+#  else
+#    define JSON_HEDLEY_PRIVATE
+#    define JSON_HEDLEY_PUBLIC
+#  endif
+#  define JSON_HEDLEY_IMPORT    extern
+#endif
+
+#if defined(JSON_HEDLEY_NO_THROW)
+    #undef JSON_HEDLEY_NO_THROW
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nothrow) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_THROW __attribute__((__nothrow__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NO_THROW __declspec(nothrow)
+#else
+    #define JSON_HEDLEY_NO_THROW
+#endif
+
+#if defined(JSON_HEDLEY_FALL_THROUGH)
+    #undef JSON_HEDLEY_FALL_THROUGH
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(fallthrough) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_FALL_THROUGH __attribute__((__fallthrough__))
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(clang,fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[clang::fallthrough]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[fallthrough]])
+#elif defined(__fallthrough) /* SAL */
+    #define JSON_HEDLEY_FALL_THROUGH __fallthrough
+#else
+    #define JSON_HEDLEY_FALL_THROUGH
+#endif
+
+#if defined(JSON_HEDLEY_RETURNS_NON_NULL)
+    #undef JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(returns_nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RETURNS_NON_NULL __attribute__((__returns_nonnull__))
+#elif defined(_Ret_notnull_) /* SAL */
+    #define JSON_HEDLEY_RETURNS_NON_NULL _Ret_notnull_
+#else
+    #define JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+
+#if defined(JSON_HEDLEY_ARRAY_PARAM)
+    #undef JSON_HEDLEY_ARRAY_PARAM
+#endif
+#if \
+    defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && \
+    !defined(__STDC_NO_VLA__) && \
+    !defined(__cplusplus) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_ARRAY_PARAM(name) (name)
+#else
+    #define JSON_HEDLEY_ARRAY_PARAM(name)
+#endif
+
+#if defined(JSON_HEDLEY_IS_CONSTANT)
+    #undef JSON_HEDLEY_IS_CONSTANT
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_CONSTEXPR)
+    #undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#endif
+/* JSON_HEDLEY_IS_CONSTEXPR_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #undef JSON_HEDLEY_IS_CONSTEXPR_
+#endif
+#if \
+    JSON_HEDLEY_HAS_BUILTIN(__builtin_constant_p) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,19) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) && !defined(__cplusplus)) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_IS_CONSTANT(expr) __builtin_constant_p(expr)
+#endif
+#if !defined(__cplusplus)
+#  if \
+       JSON_HEDLEY_HAS_BUILTIN(__builtin_types_compatible_p) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+       JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+       JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,24)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0)), int*)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((intptr_t) ((expr) * 0)) : (int*) 0)), int*)
+#endif
+#  elif \
+       ( \
+          defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) && \
+          !defined(JSON_HEDLEY_SUNPRO_VERSION) && \
+          !defined(JSON_HEDLEY_PGI_VERSION) && \
+          !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       (JSON_HEDLEY_HAS_EXTENSION(c_generic_selections) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,3,0)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0), int*: 1, void*: 0)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((intptr_t) * 0) : (int*) 0), int*: 1, void*: 0)
+#endif
+#  elif \
+       defined(JSON_HEDLEY_GCC_VERSION) || \
+       defined(JSON_HEDLEY_INTEL_VERSION) || \
+       defined(JSON_HEDLEY_TINYC_VERSION) || \
+       defined(JSON_HEDLEY_TI_ARMCL_VERSION) || \
+       JSON_HEDLEY_TI_CL430_VERSION_CHECK(18,12,0) || \
+       defined(JSON_HEDLEY_TI_CL2000_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL6X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL7X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CLPRU_VERSION) || \
+       defined(__clang__)
+#    define JSON_HEDLEY_IS_CONSTEXPR_(expr) ( \
+        sizeof(void) != \
+        sizeof(*( \
+                  1 ? \
+                  ((void*) ((expr) * 0L) ) : \
+((struct { char v[sizeof(void) * 2]; } *) 1) \
+                ) \
+              ) \
+                                            )
+#  endif
+#endif
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) JSON_HEDLEY_IS_CONSTEXPR_(expr)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (JSON_HEDLEY_IS_CONSTEXPR_(expr) ? (expr) : (-1))
+#else
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) (0)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_BEGIN_C_DECLS)
+    #undef JSON_HEDLEY_BEGIN_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_END_C_DECLS)
+    #undef JSON_HEDLEY_END_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_C_DECL)
+    #undef JSON_HEDLEY_C_DECL
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_BEGIN_C_DECLS extern "C" {
+    #define JSON_HEDLEY_END_C_DECLS }
+    #define JSON_HEDLEY_C_DECL extern "C"
+#else
+    #define JSON_HEDLEY_BEGIN_C_DECLS
+    #define JSON_HEDLEY_END_C_DECLS
+    #define JSON_HEDLEY_C_DECL
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_ASSERT)
+    #undef JSON_HEDLEY_STATIC_ASSERT
+#endif
+#if \
+  !defined(__cplusplus) && ( \
+      (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) || \
+      (JSON_HEDLEY_HAS_FEATURE(c_static_assert) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+      JSON_HEDLEY_GCC_VERSION_CHECK(6,0,0) || \
+      JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+      defined(_Static_assert) \
+    )
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) _Static_assert(expr, message)
+#elif \
+  (defined(__cplusplus) && (__cplusplus >= 201103L)) || \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(16,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(static_assert(expr, message))
+#else
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message)
+#endif
+
+#if defined(JSON_HEDLEY_NULL)
+    #undef JSON_HEDLEY_NULL
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(nullptr)
+    #elif defined(NULL)
+        #define JSON_HEDLEY_NULL NULL
+    #else
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_STATIC_CAST(void*, 0)
+    #endif
+#elif defined(NULL)
+    #define JSON_HEDLEY_NULL NULL
+#else
+    #define JSON_HEDLEY_NULL ((void*) 0)
+#endif
+
+#if defined(JSON_HEDLEY_MESSAGE)
+    #undef JSON_HEDLEY_MESSAGE
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_MESSAGE(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(message msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message msg)
+#elif JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(_CRI message msg)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_WARNING)
+    #undef JSON_HEDLEY_WARNING
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_WARNING(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(clang warning msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,8,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(GCC warning msg)
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_REQUIRE)
+    #undef JSON_HEDLEY_REQUIRE
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_MSG)
+    #undef JSON_HEDLEY_REQUIRE_MSG
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(diagnose_if)
+#  if JSON_HEDLEY_HAS_WARNING("-Wgcc-compat")
+#    define JSON_HEDLEY_REQUIRE(expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), #expr, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), msg, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_REQUIRE(expr) __attribute__((diagnose_if(!(expr), #expr, "error")))
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) __attribute__((diagnose_if(!(expr), msg, "error")))
+#  endif
+#else
+#  define JSON_HEDLEY_REQUIRE(expr)
+#  define JSON_HEDLEY_REQUIRE_MSG(expr,msg)
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS)
+    #undef JSON_HEDLEY_FLAGS
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(flag_enum) && (!defined(__cplusplus) || JSON_HEDLEY_HAS_WARNING("-Wbitfield-enum-conversion"))
+    #define JSON_HEDLEY_FLAGS __attribute__((__flag_enum__))
+#else
+    #define JSON_HEDLEY_FLAGS
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS_CAST)
+    #undef JSON_HEDLEY_FLAGS_CAST
+#endif
+#if JSON_HEDLEY_INTEL_VERSION_CHECK(19,0,0)
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        _Pragma("warning(disable:188)") \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) JSON_HEDLEY_STATIC_CAST(T, expr)
+#endif
+
+#if defined(JSON_HEDLEY_EMPTY_BASES)
+    #undef JSON_HEDLEY_EMPTY_BASES
+#endif
+#if \
+    (JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,23918) && !JSON_HEDLEY_MSVC_VERSION_CHECK(20,0,0)) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_EMPTY_BASES __declspec(empty_bases)
+#else
+    #define JSON_HEDLEY_EMPTY_BASES
+#endif
+
+/* Remaining macros are deprecated. */
+
+#if defined(JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) (0)
+#else
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_CLANG_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_BUILTIN)
+    #undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#endif
+#define JSON_HEDLEY_CLANG_HAS_BUILTIN(builtin) JSON_HEDLEY_HAS_BUILTIN(builtin)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_FEATURE)
+    #undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_FEATURE(feature) JSON_HEDLEY_HAS_FEATURE(feature)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_EXTENSION)
+    #undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#endif
+#define JSON_HEDLEY_CLANG_HAS_EXTENSION(extension) JSON_HEDLEY_HAS_EXTENSION(extension)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_DECLSPEC_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_WARNING)
+    #undef JSON_HEDLEY_CLANG_HAS_WARNING
+#endif
+#define JSON_HEDLEY_CLANG_HAS_WARNING(warning) JSON_HEDLEY_HAS_WARNING(warning)
+
+#endif /* !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < X) */
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+
+#include <type_traits>
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename ...Ts> struct make_void
+{
+    using type = void;
+};
+template<typename ...Ts> using void_t = typename make_void<Ts...>::type;
+} // namespace detail
+}  // namespace nlohmann
+
+
+// https://en.cppreference.com/w/cpp/experimental/is_detected
+namespace nlohmann
+{
+namespace detail
+{
+struct nonesuch
+{
+    nonesuch() = delete;
+    ~nonesuch() = delete;
+    nonesuch(nonesuch const&) = delete;
+    nonesuch(nonesuch const&&) = delete;
+    void operator=(nonesuch const&) = delete;
+    void operator=(nonesuch&&) = delete;
+};
+
+template<class Default,
+         class AlwaysVoid,
+         template<class...> class Op,
+         class... Args>
+struct detector
+{
+    using value_t = std::false_type;
+    using type = Default;
+};
+
+template<class Default, template<class...> class Op, class... Args>
+struct detector<Default, void_t<Op<Args...>>, Op, Args...>
+{
+    using value_t = std::true_type;
+    using type = Op<Args...>;
+};
+
+template<template<class...> class Op, class... Args>
+using is_detected = typename detector<nonesuch, void, Op, Args...>::value_t;
+
+template<template<class...> class Op, class... Args>
+struct is_detected_lazy : is_detected<Op, Args...> { };
+
+template<template<class...> class Op, class... Args>
+using detected_t = typename detector<nonesuch, void, Op, Args...>::type;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or = detector<Default, void, Op, Args...>;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or_t = typename detected_or<Default, Op, Args...>::type;
+
+template<class Expected, template<class...> class Op, class... Args>
+using is_detected_exact = std::is_same<Expected, detected_t<Op, Args...>>;
+
+template<class To, template<class...> class Op, class... Args>
+using is_detected_convertible =
+    std::is_convertible<detected_t<Op, Args...>, To>;
+}  // namespace detail
+}  // namespace nlohmann
+
+
+// This file contains all internal macro definitions
+// You MUST include macro_unscope.hpp at the end of json.hpp to undef all of them
+
+// exclude unsupported compilers
+#if !defined(JSON_SKIP_UNSUPPORTED_COMPILER_CHECK)
+    #if defined(__clang__)
+        #if (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) < 30400
+            #error "unsupported Clang version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #elif defined(__GNUC__) && !(defined(__ICC) || defined(__INTEL_COMPILER))
+        #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) < 40800
+            #error "unsupported GCC version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #endif
+#endif
+
+// C++ language standard detection
+// if the user manually specified the used c++ version this is skipped
+#if !defined(JSON_HAS_CPP_20) && !defined(JSON_HAS_CPP_17) && !defined(JSON_HAS_CPP_14) && !defined(JSON_HAS_CPP_11)
+    #if (defined(__cplusplus) && __cplusplus >= 202002L) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+        #define JSON_HAS_CPP_20
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201703L) || (defined(_HAS_CXX17) && _HAS_CXX17 == 1) // fix for issue #464
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201402L) || (defined(_HAS_CXX14) && _HAS_CXX14 == 1)
+        #define JSON_HAS_CPP_14
+    #endif
+    // the cpp 11 flag is always specified because it is the minimal required version
+    #define JSON_HAS_CPP_11
+#endif
+
+#if !defined(JSON_HAS_FILESYSTEM) && !defined(JSON_HAS_EXPERIMENTAL_FILESYSTEM)
+    #ifdef JSON_HAS_CPP_17
+        #if defined(__cpp_lib_filesystem)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif defined(__cpp_lib_experimental_filesystem)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif !defined(__has_include)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif __has_include(<filesystem>)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif __has_include(<experimental/filesystem>)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #endif
+
+        // std::filesystem does not work on MinGW GCC 8: https://sourceforge.net/p/mingw-w64/bugs/737/
+        #if defined(__MINGW32__) && defined(__GNUC__) && __GNUC__ == 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before GCC 8: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before Clang 7: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__clang_major__) && __clang_major__ < 7
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before MSVC 19.14: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(_MSC_VER) && _MSC_VER < 1940
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before iOS 13
+        #if defined(__IPHONE_OS_VERSION_MIN_REQUIRED) && __IPHONE_OS_VERSION_MIN_REQUIRED < 130000
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before macOS Catalina
+        #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 101500
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+    #endif
+#endif
+
+#ifndef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+    #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 0
+#endif
+
+#ifndef JSON_HAS_FILESYSTEM
+    #define JSON_HAS_FILESYSTEM 0
+#endif
+
+// disable documentation warnings on clang
+#if defined(__clang__)
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wdocumentation"
+    #pragma clang diagnostic ignored "-Wdocumentation-unknown-command"
+#endif
+
+// allow disabling exceptions
+#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)) && !defined(JSON_NOEXCEPTION)
+    #define JSON_THROW(exception) throw exception
+    #define JSON_TRY try
+    #define JSON_CATCH(exception) catch(exception)
+    #define JSON_INTERNAL_CATCH(exception) catch(exception)
+#else
+    #include <cstdlib>
+    #define JSON_THROW(exception) std::abort()
+    #define JSON_TRY if(true)
+    #define JSON_CATCH(exception) if(false)
+    #define JSON_INTERNAL_CATCH(exception) if(false)
+#endif
+
+// override exception macros
+#if defined(JSON_THROW_USER)
+    #undef JSON_THROW
+    #define JSON_THROW JSON_THROW_USER
+#endif
+#if defined(JSON_TRY_USER)
+    #undef JSON_TRY
+    #define JSON_TRY JSON_TRY_USER
+#endif
+#if defined(JSON_CATCH_USER)
+    #undef JSON_CATCH
+    #define JSON_CATCH JSON_CATCH_USER
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_CATCH_USER
+#endif
+#if defined(JSON_INTERNAL_CATCH_USER)
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_INTERNAL_CATCH_USER
+#endif
+
+// allow overriding assert
+#if !defined(JSON_ASSERT)
+    #include <cassert> // assert
+    #define JSON_ASSERT(x) assert(x)
+#endif
+
+// allow to access some private functions (needed by the test suite)
+#if defined(JSON_TESTS_PRIVATE)
+    #define JSON_PRIVATE_UNLESS_TESTED public
+#else
+    #define JSON_PRIVATE_UNLESS_TESTED private
+#endif
+
+/*!
+@brief macro to briefly define a mapping between an enum and JSON
+@def NLOHMANN_JSON_SERIALIZE_ENUM
+@since version 3.4.0
+*/
+#define NLOHMANN_JSON_SERIALIZE_ENUM(ENUM_TYPE, ...)                                            \
+    template<typename BasicJsonType>                                                            \
+    inline void to_json(BasicJsonType& j, const ENUM_TYPE& e)                                   \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [e](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool  \
+        {                                                                                       \
+            return ej_pair.first == e;                                                          \
+        });                                                                                     \
+        j = ((it != std::end(m)) ? it : std::begin(m))->second;                                 \
+    }                                                                                           \
+    template<typename BasicJsonType>                                                            \
+    inline void from_json(const BasicJsonType& j, ENUM_TYPE& e)                                 \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [&j](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool \
+        {                                                                                       \
+            return ej_pair.second == j;                                                         \
+        });                                                                                     \
+        e = ((it != std::end(m)) ? it : std::begin(m))->first;                                  \
+    }
+
+// Ugly macros to avoid uglier copy-paste when specializing basic_json. They
+// may be removed in the future once the class is split.
+
+#define NLOHMANN_BASIC_JSON_TPL_DECLARATION                                \
+    template<template<typename, typename, typename...> class ObjectType,   \
+             template<typename, typename...> class ArrayType,              \
+             class StringType, class BooleanType, class NumberIntegerType, \
+             class NumberUnsignedType, class NumberFloatType,              \
+             template<typename> class AllocatorType,                       \
+             template<typename, typename = void> class JSONSerializer,     \
+             class BinaryType>
+
+#define NLOHMANN_BASIC_JSON_TPL                                            \
+    basic_json<ObjectType, ArrayType, StringType, BooleanType,             \
+    NumberIntegerType, NumberUnsignedType, NumberFloatType,                \
+    AllocatorType, JSONSerializer, BinaryType>
+
+// Macros to simplify conversion from/to types
+
+#define NLOHMANN_JSON_EXPAND( x ) x
+#define NLOHMANN_JSON_GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, _61, _62, _63, _64, NAME,...) NAME
+#define NLOHMANN_JSON_PASTE(...) NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_GET_MACRO(__VA_ARGS__, \
+        NLOHMANN_JSON_PASTE64, \
+        NLOHMANN_JSON_PASTE63, \
+        NLOHMANN_JSON_PASTE62, \
+        NLOHMANN_JSON_PASTE61, \
+        NLOHMANN_JSON_PASTE60, \
+        NLOHMANN_JSON_PASTE59, \
+        NLOHMANN_JSON_PASTE58, \
+        NLOHMANN_JSON_PASTE57, \
+        NLOHMANN_JSON_PASTE56, \
+        NLOHMANN_JSON_PASTE55, \
+        NLOHMANN_JSON_PASTE54, \
+        NLOHMANN_JSON_PASTE53, \
+        NLOHMANN_JSON_PASTE52, \
+        NLOHMANN_JSON_PASTE51, \
+        NLOHMANN_JSON_PASTE50, \
+        NLOHMANN_JSON_PASTE49, \
+        NLOHMANN_JSON_PASTE48, \
+        NLOHMANN_JSON_PASTE47, \
+        NLOHMANN_JSON_PASTE46, \
+        NLOHMANN_JSON_PASTE45, \
+        NLOHMANN_JSON_PASTE44, \
+        NLOHMANN_JSON_PASTE43, \
+        NLOHMANN_JSON_PASTE42, \
+        NLOHMANN_JSON_PASTE41, \
+        NLOHMANN_JSON_PASTE40, \
+        NLOHMANN_JSON_PASTE39, \
+        NLOHMANN_JSON_PASTE38, \
+        NLOHMANN_JSON_PASTE37, \
+        NLOHMANN_JSON_PASTE36, \
+        NLOHMANN_JSON_PASTE35, \
+        NLOHMANN_JSON_PASTE34, \
+        NLOHMANN_JSON_PASTE33, \
+        NLOHMANN_JSON_PASTE32, \
+        NLOHMANN_JSON_PASTE31, \
+        NLOHMANN_JSON_PASTE30, \
+        NLOHMANN_JSON_PASTE29, \
+        NLOHMANN_JSON_PASTE28, \
+        NLOHMANN_JSON_PASTE27, \
+        NLOHMANN_JSON_PASTE26, \
+        NLOHMANN_JSON_PASTE25, \
+        NLOHMANN_JSON_PASTE24, \
+        NLOHMANN_JSON_PASTE23, \
+        NLOHMANN_JSON_PASTE22, \
+        NLOHMANN_JSON_PASTE21, \
+        NLOHMANN_JSON_PASTE20, \
+        NLOHMANN_JSON_PASTE19, \
+        NLOHMANN_JSON_PASTE18, \
+        NLOHMANN_JSON_PASTE17, \
+        NLOHMANN_JSON_PASTE16, \
+        NLOHMANN_JSON_PASTE15, \
+        NLOHMANN_JSON_PASTE14, \
+        NLOHMANN_JSON_PASTE13, \
+        NLOHMANN_JSON_PASTE12, \
+        NLOHMANN_JSON_PASTE11, \
+        NLOHMANN_JSON_PASTE10, \
+        NLOHMANN_JSON_PASTE9, \
+        NLOHMANN_JSON_PASTE8, \
+        NLOHMANN_JSON_PASTE7, \
+        NLOHMANN_JSON_PASTE6, \
+        NLOHMANN_JSON_PASTE5, \
+        NLOHMANN_JSON_PASTE4, \
+        NLOHMANN_JSON_PASTE3, \
+        NLOHMANN_JSON_PASTE2, \
+        NLOHMANN_JSON_PASTE1)(__VA_ARGS__))
+#define NLOHMANN_JSON_PASTE2(func, v1) func(v1)
+#define NLOHMANN_JSON_PASTE3(func, v1, v2) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE2(func, v2)
+#define NLOHMANN_JSON_PASTE4(func, v1, v2, v3) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE3(func, v2, v3)
+#define NLOHMANN_JSON_PASTE5(func, v1, v2, v3, v4) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE4(func, v2, v3, v4)
+#define NLOHMANN_JSON_PASTE6(func, v1, v2, v3, v4, v5) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE5(func, v2, v3, v4, v5)
+#define NLOHMANN_JSON_PASTE7(func, v1, v2, v3, v4, v5, v6) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE6(func, v2, v3, v4, v5, v6)
+#define NLOHMANN_JSON_PASTE8(func, v1, v2, v3, v4, v5, v6, v7) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE7(func, v2, v3, v4, v5, v6, v7)
+#define NLOHMANN_JSON_PASTE9(func, v1, v2, v3, v4, v5, v6, v7, v8) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE8(func, v2, v3, v4, v5, v6, v7, v8)
+#define NLOHMANN_JSON_PASTE10(func, v1, v2, v3, v4, v5, v6, v7, v8, v9) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE9(func, v2, v3, v4, v5, v6, v7, v8, v9)
+#define NLOHMANN_JSON_PASTE11(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE10(func, v2, v3, v4, v5, v6, v7, v8, v9, v10)
+#define NLOHMANN_JSON_PASTE12(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE11(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11)
+#define NLOHMANN_JSON_PASTE13(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE12(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12)
+#define NLOHMANN_JSON_PASTE14(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE13(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13)
+#define NLOHMANN_JSON_PASTE15(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE14(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14)
+#define NLOHMANN_JSON_PASTE16(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE15(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15)
+#define NLOHMANN_JSON_PASTE17(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE16(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16)
+#define NLOHMANN_JSON_PASTE18(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17)
+#define NLOHMANN_JSON_PASTE19(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18)
+#define NLOHMANN_JSON_PASTE20(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19)
+#define NLOHMANN_JSON_PASTE21(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20)
+#define NLOHMANN_JSON_PASTE22(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21)
+#define NLOHMANN_JSON_PASTE23(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22)
+#define NLOHMANN_JSON_PASTE24(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23)
+#define NLOHMANN_JSON_PASTE25(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24)
+#define NLOHMANN_JSON_PASTE26(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25)
+#define NLOHMANN_JSON_PASTE27(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE26(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26)
+#define NLOHMANN_JSON_PASTE28(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE27(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27)
+#define NLOHMANN_JSON_PASTE29(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE28(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28)
+#define NLOHMANN_JSON_PASTE30(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29)
+#define NLOHMANN_JSON_PASTE31(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30)
+#define NLOHMANN_JSON_PASTE32(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31)
+#define NLOHMANN_JSON_PASTE33(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32)
+#define NLOHMANN_JSON_PASTE34(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33)
+#define NLOHMANN_JSON_PASTE35(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34)
+#define NLOHMANN_JSON_PASTE36(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35)
+#define NLOHMANN_JSON_PASTE37(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36)
+#define NLOHMANN_JSON_PASTE38(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37)
+#define NLOHMANN_JSON_PASTE39(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE38(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38)
+#define NLOHMANN_JSON_PASTE40(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE39(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39)
+#define NLOHMANN_JSON_PASTE41(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE40(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40)
+#define NLOHMANN_JSON_PASTE42(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE41(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41)
+#define NLOHMANN_JSON_PASTE43(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE42(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42)
+#define NLOHMANN_JSON_PASTE44(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE43(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43)
+#define NLOHMANN_JSON_PASTE45(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE44(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44)
+#define NLOHMANN_JSON_PASTE46(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE45(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45)
+#define NLOHMANN_JSON_PASTE47(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE46(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46)
+#define NLOHMANN_JSON_PASTE48(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE47(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47)
+#define NLOHMANN_JSON_PASTE49(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE48(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48)
+#define NLOHMANN_JSON_PASTE50(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE49(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49)
+#define NLOHMANN_JSON_PASTE51(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE50(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50)
+#define NLOHMANN_JSON_PASTE52(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE51(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51)
+#define NLOHMANN_JSON_PASTE53(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE52(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52)
+#define NLOHMANN_JSON_PASTE54(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE53(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53)
+#define NLOHMANN_JSON_PASTE55(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE54(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54)
+#define NLOHMANN_JSON_PASTE56(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE55(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55)
+#define NLOHMANN_JSON_PASTE57(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE56(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56)
+#define NLOHMANN_JSON_PASTE58(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE57(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57)
+#define NLOHMANN_JSON_PASTE59(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE58(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58)
+#define NLOHMANN_JSON_PASTE60(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE59(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59)
+#define NLOHMANN_JSON_PASTE61(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE60(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60)
+#define NLOHMANN_JSON_PASTE62(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE61(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61)
+#define NLOHMANN_JSON_PASTE63(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE62(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62)
+#define NLOHMANN_JSON_PASTE64(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE63(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63)
+
+#define NLOHMANN_JSON_TO(v1) nlohmann_json_j[#v1] = nlohmann_json_t.v1;
+#define NLOHMANN_JSON_FROM(v1) nlohmann_json_j.at(#v1).get_to(nlohmann_json_t.v1);
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+
+// inspired from https://stackoverflow.com/a/26745591
+// allows to call any std function as if (e.g. with begin):
+// using std::begin; begin(x);
+//
+// it allows using the detected idiom to retrieve the return type
+// of such an expression
+#define NLOHMANN_CAN_CALL_STD_FUNC_IMPL(std_name)                                 \
+    namespace detail {                                                            \
+    using std::std_name;                                                          \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    }                                                                             \
+    \
+    namespace detail2 {                                                           \
+    struct std_name##_tag                                                         \
+    {                                                                             \
+    };                                                                            \
+    \
+    template<typename... T>                                                       \
+    std_name##_tag std_name(T&&...);                                              \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name                                              \
+    {                                                                             \
+        static constexpr auto const value = ::nlohmann::detail::                  \
+                                            is_detected_exact<std_name##_tag, result_of_##std_name, T...>::value; \
+    };                                                                            \
+    } /* namespace detail2 */ \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name : detail2::would_call_std_##std_name<T...>   \
+    {                                                                             \
+    }
+
+#ifndef JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_USE_IMPLICIT_CONVERSIONS 1
+#endif
+
+#if JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_EXPLICIT
+#else
+    #define JSON_EXPLICIT explicit
+#endif
+
+#ifndef JSON_DIAGNOSTICS
+    #define JSON_DIAGNOSTICS 0
+#endif
+
+
+namespace nlohmann
+{
+namespace detail
+{
+
+/*!
+@brief replace all occurrences of a substring by another string
+
+@param[in,out] s  the string to manipulate; changed so that all
+               occurrences of @a f are replaced with @a t
+@param[in]     f  the substring to replace with @a t
+@param[in]     t  the string to replace @a f
+
+@pre The search string @a f must not be empty. **This precondition is
+enforced with an assertion.**
+
+@since version 2.0.0
+*/
+inline void replace_substring(std::string& s, const std::string& f,
+                              const std::string& t)
+{
+    JSON_ASSERT(!f.empty());
+    for (auto pos = s.find(f);                // find first occurrence of f
+            pos != std::string::npos;         // make sure f was found
+            s.replace(pos, f.size(), t),      // replace with t, and
+            pos = s.find(f, pos + t.size()))  // find next occurrence of f
+    {}
+}
+
+/*!
+ * @brief string escaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to escape
+ * @return    escaped string
+ *
+ * Note the order of escaping "~" to "~0" and "/" to "~1" is important.
+ */
+inline std::string escape(std::string s)
+{
+    replace_substring(s, "~", "~0");
+    replace_substring(s, "/", "~1");
+    return s;
+}
+
+/*!
+ * @brief string unescaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to unescape
+ * @return    unescaped string
+ *
+ * Note the order of escaping "~1" to "/" and "~0" to "~" is important.
+ */
+static void unescape(std::string& s)
+{
+    replace_substring(s, "~1", "/");
+    replace_substring(s, "~0", "~");
+}
+
+} // namespace detail
+} // namespace nlohmann
+
+// #include <nlohmann/detail/input/position_t.hpp>
+
+
+#include <cstddef> // size_t
+
+namespace nlohmann
+{
+namespace detail
+{
+/// struct to capture the start position of the current token
+struct position_t
+{
+    /// the total number of characters read
+    std::size_t chars_read_total = 0;
+    /// the number of characters read in the current line
+    std::size_t chars_read_current_line = 0;
+    /// the number of lines read
+    std::size_t lines_read = 0;
+
+    /// conversion to size_t to preserve SAX interface
+    constexpr operator size_t() const
+    {
+        return chars_read_total;
+    }
+};
+
+} // namespace detail
+} // namespace nlohmann
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+////////////////
+// exceptions //
+////////////////
+
+/// @brief general exception of the @ref basic_json class
+/// @sa https://json.nlohmann.me/api/basic_json/exception/
+class exception : public std::exception
+{
+  public:
+    /// returns the explanatory string
+    const char* what() const noexcept override
+    {
+        return m.what();
+    }
+
+    /// the id of the exception
+    const int id; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
+
+  protected:
+    JSON_HEDLEY_NON_NULL(3)
+    exception(int id_, const char* what_arg) : id(id_), m(what_arg) {} // NOLINT(bugprone-throw-keyword-missing)
+
+    static std::string name(const std::string& ename, int id_)
+    {
+        return "[json.exception." + ename + "." + std::to_string(id_) + "] ";
+    }
+
+    template<typename BasicJsonType>
+    static std::string diagnostics(const BasicJsonType& leaf_element)
+    {
+#if JSON_DIAGNOSTICS
+        std::vector<std::string> tokens;
+        for (const auto* current = &leaf_element; current->m_parent != nullptr; current = current->m_parent)
+        {
+            switch (current->m_parent->type())
+            {
+                case value_t::array:
+                {
+                    for (std::size_t i = 0; i < current->m_parent->m_value.array->size(); ++i)
+                    {
+                        if (&current->m_parent->m_value.array->operator[](i) == current)
+                        {
+                            tokens.emplace_back(std::to_string(i));
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::object:
+                {
+                    for (const auto& element : *current->m_parent->m_value.object)
+                    {
+                        if (&element.second == current)
+                        {
+                            tokens.emplace_back(element.first.c_str());
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::null: // LCOV_EXCL_LINE
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:   // LCOV_EXCL_LINE
+                    break; // LCOV_EXCL_LINE
+            }
+        }
+
+        if (tokens.empty())
+        {
+            return "";
+        }
+
+        return "(" + std::accumulate(tokens.rbegin(), tokens.rend(), std::string{},
+                                     [](const std::string & a, const std::string & b)
+        {
+            return a + "/" + detail::escape(b);
+        }) + ") ";
+#else
+        static_cast<void>(leaf_element);
+        return "";
+#endif
+    }
+
+  private:
+    /// an exception object as storage for error messages
+    std::runtime_error m;
+};
+
+/// @brief exception indicating a parse error
+/// @sa https://json.nlohmann.me/api/basic_json/parse_error/
+class parse_error : public exception
+{
+  public:
+    /*!
+    @brief create a parse error exception
+    @param[in] id_       the id of the exception
+    @param[in] pos       the position where the error occurred (or with
+                         chars_read_total=0 if the position cannot be
+                         determined)
+    @param[in] what_arg  the explanatory string
+    @return parse_error object
+    */
+    template<typename BasicJsonType>
+    static parse_error create(int id_, const position_t& pos, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("parse_error", id_) + "parse error" +
+                        position_string(pos) + ": " + exception::diagnostics(context) + what_arg;
+        return {id_, pos.chars_read_total, w.c_str()};
+    }
+
+    template<typename BasicJsonType>
+    static parse_error create(int id_, std::size_t byte_, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("parse_error", id_) + "parse error" +
+                        (byte_ != 0 ? (" at byte " + std::to_string(byte_)) : "") +
+                        ": " + exception::diagnostics(context) + what_arg;
+        return {id_, byte_, w.c_str()};
+    }
+
+    /*!
+    @brief byte index of the parse error
+
+    The byte index of the last read character in the input file.
+
+    @note For an input with n bytes, 1 is the index of the first character and
+          n+1 is the index of the terminating null byte or the end of file.
+          This also holds true when reading a byte vector (CBOR or MessagePack).
+    */
+    const std::size_t byte;
+
+  private:
+    parse_error(int id_, std::size_t byte_, const char* what_arg)
+        : exception(id_, what_arg), byte(byte_) {}
+
+    static std::string position_string(const position_t& pos)
+    {
+        return " at line " + std::to_string(pos.lines_read + 1) +
+               ", column " + std::to_string(pos.chars_read_current_line);
+    }
+};
+
+/// @brief exception indicating errors with iterators
+/// @sa https://json.nlohmann.me/api/basic_json/invalid_iterator/
+class invalid_iterator : public exception
+{
+  public:
+    template<typename BasicJsonType>
+    static invalid_iterator create(int id_, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("invalid_iterator", id_) + exception::diagnostics(context) + what_arg;
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    invalid_iterator(int id_, const char* what_arg)
+        : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating executing a member function with a wrong type
+/// @sa https://json.nlohmann.me/api/basic_json/type_error/
+class type_error : public exception
+{
+  public:
+    template<typename BasicJsonType>
+    static type_error create(int id_, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("type_error", id_) + exception::diagnostics(context) + what_arg;
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    type_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating access out of the defined range
+/// @sa https://json.nlohmann.me/api/basic_json/out_of_range/
+class out_of_range : public exception
+{
+  public:
+    template<typename BasicJsonType>
+    static out_of_range create(int id_, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("out_of_range", id_) + exception::diagnostics(context) + what_arg;
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    out_of_range(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating other library errors
+/// @sa https://json.nlohmann.me/api/basic_json/other_error/
+class other_error : public exception
+{
+  public:
+    template<typename BasicJsonType>
+    static other_error create(int id_, const std::string& what_arg, const BasicJsonType& context)
+    {
+        std::string w = exception::name("other_error", id_) + exception::diagnostics(context) + what_arg;
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    other_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+
+#include <cstddef> // size_t
+#include <type_traits> // conditional, enable_if, false_type, integral_constant, is_constructible, is_integral, is_same, remove_cv, remove_reference, true_type
+#include <utility> // index_sequence, make_index_sequence, index_sequence_for
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+
+template<typename T>
+using uncvref_t = typename std::remove_cv<typename std::remove_reference<T>::type>::type;
+
+#ifdef JSON_HAS_CPP_14
+
+// the following utilities are natively available in C++14
+using std::enable_if_t;
+using std::index_sequence;
+using std::make_index_sequence;
+using std::index_sequence_for;
+
+#else
+
+// alias templates to reduce boilerplate
+template<bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+// The following code is taken from https://github.com/abseil/abseil-cpp/blob/10cb35e459f5ecca5b2ff107635da0bfa41011b4/absl/utility/utility.h
+// which is part of Google Abseil (https://github.com/abseil/abseil-cpp), licensed under the Apache License 2.0.
+
+//// START OF CODE FROM GOOGLE ABSEIL
+
+// integer_sequence
+//
+// Class template representing a compile-time integer sequence. An instantiation
+// of `integer_sequence<T, Ints...>` has a sequence of integers encoded in its
+// type through its template arguments (which is a common need when
+// working with C++11 variadic templates). `absl::integer_sequence` is designed
+// to be a drop-in replacement for C++14's `std::integer_sequence`.
+//
+// Example:
+//
+//   template< class T, T... Ints >
+//   void user_function(integer_sequence<T, Ints...>);
+//
+//   int main()
+//   {
+//     // user_function's `T` will be deduced to `int` and `Ints...`
+//     // will be deduced to `0, 1, 2, 3, 4`.
+//     user_function(make_integer_sequence<int, 5>());
+//   }
+template <typename T, T... Ints>
+struct integer_sequence
+{
+    using value_type = T;
+    static constexpr std::size_t size() noexcept
+    {
+        return sizeof...(Ints);
+    }
+};
+
+// index_sequence
+//
+// A helper template for an `integer_sequence` of `size_t`,
+// `absl::index_sequence` is designed to be a drop-in replacement for C++14's
+// `std::index_sequence`.
+template <size_t... Ints>
+using index_sequence = integer_sequence<size_t, Ints...>;
+
+namespace utility_internal
+{
+
+template <typename Seq, size_t SeqSize, size_t Rem>
+struct Extend;
+
+// Note that SeqSize == sizeof...(Ints). It's passed explicitly for efficiency.
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 0>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)... >;
+};
+
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 1>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)..., 2 * SeqSize >;
+};
+
+// Recursion helper for 'make_integer_sequence<T, N>'.
+// 'Gen<T, N>::type' is an alias for 'integer_sequence<T, 0, 1, ... N-1>'.
+template <typename T, size_t N>
+struct Gen
+{
+    using type =
+        typename Extend < typename Gen < T, N / 2 >::type, N / 2, N % 2 >::type;
+};
+
+template <typename T>
+struct Gen<T, 0>
+{
+    using type = integer_sequence<T>;
+};
+
+}  // namespace utility_internal
+
+// Compile-time sequences of integers
+
+// make_integer_sequence
+//
+// This template alias is equivalent to
+// `integer_sequence<int, 0, 1, ..., N-1>`, and is designed to be a drop-in
+// replacement for C++14's `std::make_integer_sequence`.
+template <typename T, T N>
+using make_integer_sequence = typename utility_internal::Gen<T, N>::type;
+
+// make_index_sequence
+//
+// This template alias is equivalent to `index_sequence<0, 1, ..., N-1>`,
+// and is designed to be a drop-in replacement for C++14's
+// `std::make_index_sequence`.
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
+
+// index_sequence_for
+//
+// Converts a typename pack into an index sequence of the same length, and
+// is designed to be a drop-in replacement for C++14's
+// `std::index_sequence_for()`
+template <typename... Ts>
+using index_sequence_for = make_index_sequence<sizeof...(Ts)>;
+
+//// END OF CODE FROM GOOGLE ABSEIL
+
+#endif
+
+// dispatch utility (taken from ranges-v3)
+template<unsigned N> struct priority_tag : priority_tag < N - 1 > {};
+template<> struct priority_tag<0> {};
+
+// taken from ranges-v3
+template<typename T>
+struct static_const
+{
+    static constexpr T value{};
+};
+
+template<typename T>
+constexpr T static_const<T>::value; // NOLINT(readability-redundant-declaration)
+
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+// dispatching helper struct
+template <class T> struct identity_tag {};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+#include <limits> // numeric_limits
+#include <type_traits> // false_type, is_constructible, is_integral, is_same, true_type
+#include <utility> // declval
+#include <tuple> // tuple
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+
+
+#include <iterator> // random_access_iterator_tag
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename It, typename = void>
+struct iterator_types {};
+
+template<typename It>
+struct iterator_types <
+    It,
+    void_t<typename It::difference_type, typename It::value_type, typename It::pointer,
+    typename It::reference, typename It::iterator_category >>
+{
+    using difference_type = typename It::difference_type;
+    using value_type = typename It::value_type;
+    using pointer = typename It::pointer;
+    using reference = typename It::reference;
+    using iterator_category = typename It::iterator_category;
+};
+
+// This is required as some compilers implement std::iterator_traits in a way that
+// doesn't work with SFINAE. See https://github.com/nlohmann/json/issues/1341.
+template<typename T, typename = void>
+struct iterator_traits
+{
+};
+
+template<typename T>
+struct iterator_traits < T, enable_if_t < !std::is_pointer<T>::value >>
+            : iterator_types<T>
+{
+};
+
+template<typename T>
+struct iterator_traits<T*, enable_if_t<std::is_object<T>::value>>
+{
+    using iterator_category = std::random_access_iterator_tag;
+    using value_type = T;
+    using difference_type = ptrdiff_t;
+    using pointer = T*;
+    using reference = T&;
+};
+} // namespace detail
+} // namespace nlohmann
+
+// #include <nlohmann/detail/meta/call_std/begin.hpp>
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(begin);
+} // namespace nlohmann
+
+// #include <nlohmann/detail/meta/call_std/end.hpp>
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(end);
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_
+#define INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+#include <cstdint> // int64_t, uint64_t
+#include <map> // map
+#include <memory> // allocator
+#include <string> // string
+#include <vector> // vector
+
+/*!
+@brief namespace for Niels Lohmann
+@see https://github.com/nlohmann
+@since version 1.0.0
+*/
+namespace nlohmann
+{
+/*!
+@brief default JSONSerializer template argument
+
+This serializer ignores the template arguments and uses ADL
+([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl))
+for serialization.
+*/
+template<typename T = void, typename SFINAE = void>
+struct adl_serializer;
+
+/// a class to store JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/
+template<template<typename U, typename V, typename... Args> class ObjectType =
+         std::map,
+         template<typename U, typename... Args> class ArrayType = std::vector,
+         class StringType = std::string, class BooleanType = bool,
+         class NumberIntegerType = std::int64_t,
+         class NumberUnsignedType = std::uint64_t,
+         class NumberFloatType = double,
+         template<typename U> class AllocatorType = std::allocator,
+         template<typename T, typename SFINAE = void> class JSONSerializer =
+         adl_serializer,
+         class BinaryType = std::vector<std::uint8_t>>
+class basic_json;
+
+/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+/// @sa https://json.nlohmann.me/api/json_pointer/
+template<typename BasicJsonType>
+class json_pointer;
+
+/*!
+@brief default specialization
+@sa https://json.nlohmann.me/api/json/
+*/
+using json = basic_json<>;
+
+/// @brief a minimal map-like container that preserves insertion order
+/// @sa https://json.nlohmann.me/api/ordered_map/
+template<class Key, class T, class IgnoredLess, class Allocator>
+struct ordered_map;
+
+/// @brief specialization that maintains the insertion order of object keys
+/// @sa https://json.nlohmann.me/api/ordered_json/
+using ordered_json = basic_json<nlohmann::ordered_map>;
+
+}  // namespace nlohmann
+
+#endif  // INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+
+namespace nlohmann
+{
+/*!
+@brief detail namespace with internal helper functions
+
+This namespace collects functions that should not be exposed,
+implementations of some @ref basic_json methods, and meta-programming helpers.
+
+@since version 2.1.0
+*/
+namespace detail
+{
+/////////////
+// helpers //
+/////////////
+
+// Note to maintainers:
+//
+// Every trait in this file expects a non CV-qualified type.
+// The only exceptions are in the 'aliases for detected' section
+// (i.e. those of the form: decltype(T::member_function(std::declval<T>())))
+//
+// In this case, T has to be properly CV-qualified to constraint the function arguments
+// (e.g. to_json(BasicJsonType&, const T&))
+
+template<typename> struct is_basic_json : std::false_type {};
+
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct is_basic_json<NLOHMANN_BASIC_JSON_TPL> : std::true_type {};
+
+//////////////////////
+// json_ref helpers //
+//////////////////////
+
+template<typename>
+class json_ref;
+
+template<typename>
+struct is_json_ref : std::false_type {};
+
+template<typename T>
+struct is_json_ref<json_ref<T>> : std::true_type {};
+
+//////////////////////////
+// aliases for detected //
+//////////////////////////
+
+template<typename T>
+using mapped_type_t = typename T::mapped_type;
+
+template<typename T>
+using key_type_t = typename T::key_type;
+
+template<typename T>
+using value_type_t = typename T::value_type;
+
+template<typename T>
+using difference_type_t = typename T::difference_type;
+
+template<typename T>
+using pointer_t = typename T::pointer;
+
+template<typename T>
+using reference_t = typename T::reference;
+
+template<typename T>
+using iterator_category_t = typename T::iterator_category;
+
+template<typename T, typename... Args>
+using to_json_function = decltype(T::to_json(std::declval<Args>()...));
+
+template<typename T, typename... Args>
+using from_json_function = decltype(T::from_json(std::declval<Args>()...));
+
+template<typename T, typename U>
+using get_template_function = decltype(std::declval<T>().template get<U>());
+
+// trait checking if JSONSerializer<T>::from_json(json const&, udt&) exists
+template<typename BasicJsonType, typename T, typename = void>
+struct has_from_json : std::false_type {};
+
+// trait checking if j.get<T> is valid
+// use this trait instead of std::is_constructible or std::is_convertible,
+// both rely on, or make use of implicit conversions, and thus fail when T
+// has several constructors/operator= (see https://github.com/nlohmann/json/issues/958)
+template <typename BasicJsonType, typename T>
+struct is_getable
+{
+    static constexpr bool value = is_detected<get_template_function, const BasicJsonType&, T>::value;
+};
+
+template<typename BasicJsonType, typename T>
+struct has_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, from_json_function, serializer,
+        const BasicJsonType&, T&>::value;
+};
+
+// This trait checks if JSONSerializer<T>::from_json(json const&) exists
+// this overload is used for non-default-constructible user-defined-types
+template<typename BasicJsonType, typename T, typename = void>
+struct has_non_default_from_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_non_default_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<T, from_json_function, serializer,
+        const BasicJsonType&>::value;
+};
+
+// This trait checks if BasicJsonType::json_serializer<T>::to_json exists
+// Do not evaluate the trait when T is a basic_json type, to avoid template instantiation infinite recursion.
+template<typename BasicJsonType, typename T, typename = void>
+struct has_to_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_to_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, to_json_function, serializer, BasicJsonType&,
+        T>::value;
+};
+
+
+///////////////////
+// is_ functions //
+///////////////////
+
+// https://en.cppreference.com/w/cpp/types/conjunction
+template<class...> struct conjunction : std::true_type { };
+template<class B1> struct conjunction<B1> : B1 { };
+template<class B1, class... Bn>
+struct conjunction<B1, Bn...>
+: std::conditional<bool(B1::value), conjunction<Bn...>, B1>::type {};
+
+// https://en.cppreference.com/w/cpp/types/negation
+template<class B> struct negation : std::integral_constant < bool, !B::value > { };
+
+// Reimplementation of is_constructible and is_default_constructible, due to them being broken for
+// std::pair and std::tuple until LWG 2367 fix (see https://cplusplus.github.io/LWG/lwg-defects.html#2367).
+// This causes compile errors in e.g. clang 3.5 or gcc 4.9.
+template <typename T>
+struct is_default_constructible : std::is_default_constructible<T> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<const std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename... Ts>
+struct is_default_constructible<std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+template <typename... Ts>
+struct is_default_constructible<const std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+
+template <typename T, typename... Args>
+struct is_constructible : std::is_constructible<T, Args...> {};
+
+template <typename T1, typename T2>
+struct is_constructible<std::pair<T1, T2>> : is_default_constructible<std::pair<T1, T2>> {};
+
+template <typename T1, typename T2>
+struct is_constructible<const std::pair<T1, T2>> : is_default_constructible<const std::pair<T1, T2>> {};
+
+template <typename... Ts>
+struct is_constructible<std::tuple<Ts...>> : is_default_constructible<std::tuple<Ts...>> {};
+
+template <typename... Ts>
+struct is_constructible<const std::tuple<Ts...>> : is_default_constructible<const std::tuple<Ts...>> {};
+
+
+template<typename T, typename = void>
+struct is_iterator_traits : std::false_type {};
+
+template<typename T>
+struct is_iterator_traits<iterator_traits<T>>
+{
+  private:
+    using traits = iterator_traits<T>;
+
+  public:
+    static constexpr auto value =
+        is_detected<value_type_t, traits>::value &&
+        is_detected<difference_type_t, traits>::value &&
+        is_detected<pointer_t, traits>::value &&
+        is_detected<iterator_category_t, traits>::value &&
+        is_detected<reference_t, traits>::value;
+};
+
+template<typename T>
+struct is_range
+{
+  private:
+    using t_ref = typename std::add_lvalue_reference<T>::type;
+
+    using iterator = detected_t<result_of_begin, t_ref>;
+    using sentinel = detected_t<result_of_end, t_ref>;
+
+    // to be 100% correct, it should use https://en.cppreference.com/w/cpp/iterator/input_or_output_iterator
+    // and https://en.cppreference.com/w/cpp/iterator/sentinel_for
+    // but reimplementing these would be too much work, as a lot of other concepts are used underneath
+    static constexpr auto is_iterator_begin =
+        is_iterator_traits<iterator_traits<iterator>>::value;
+
+  public:
+    static constexpr bool value = !std::is_same<iterator, nonesuch>::value && !std::is_same<sentinel, nonesuch>::value && is_iterator_begin;
+};
+
+template<typename R>
+using iterator_t = enable_if_t<is_range<R>::value, result_of_begin<decltype(std::declval<R&>())>>;
+
+template<typename T>
+using range_value_t = value_type_t<iterator_traits<iterator_t<T>>>;
+
+// The following implementation of is_complete_type is taken from
+// https://blogs.msdn.microsoft.com/vcblog/2015/12/02/partial-support-for-expression-sfinae-in-vs-2015-update-1/
+// and is written by Xiang Fan who agreed to using it in this library.
+
+template<typename T, typename = void>
+struct is_complete_type : std::false_type {};
+
+template<typename T>
+struct is_complete_type<T, decltype(void(sizeof(T)))> : std::true_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType,
+         typename = void>
+struct is_compatible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type_impl <
+    BasicJsonType, CompatibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, CompatibleObjectType>::value&&
+    is_detected<key_type_t, CompatibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    // macOS's is_constructible does not play well with nonesuch...
+    static constexpr bool value =
+        is_constructible<typename object_t::key_type,
+        typename CompatibleObjectType::key_type>::value &&
+        is_constructible<typename object_t::mapped_type,
+        typename CompatibleObjectType::mapped_type>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type
+    : is_compatible_object_type_impl<BasicJsonType, CompatibleObjectType> {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         typename = void>
+struct is_constructible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type_impl <
+    BasicJsonType, ConstructibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, ConstructibleObjectType>::value&&
+    is_detected<key_type_t, ConstructibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    static constexpr bool value =
+        (is_default_constructible<ConstructibleObjectType>::value &&
+         (std::is_move_assignable<ConstructibleObjectType>::value ||
+          std::is_copy_assignable<ConstructibleObjectType>::value) &&
+         (is_constructible<typename ConstructibleObjectType::key_type,
+          typename object_t::key_type>::value &&
+          std::is_same <
+          typename object_t::mapped_type,
+          typename ConstructibleObjectType::mapped_type >::value)) ||
+        (has_from_json<BasicJsonType,
+         typename ConstructibleObjectType::mapped_type>::value ||
+         has_non_default_from_json <
+         BasicJsonType,
+         typename ConstructibleObjectType::mapped_type >::value);
+};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type
+    : is_constructible_object_type_impl<BasicJsonType,
+      ConstructibleObjectType> {};
+
+template<typename BasicJsonType, typename CompatibleStringType>
+struct is_compatible_string_type
+{
+    static constexpr auto value =
+        is_constructible<typename BasicJsonType::string_t, CompatibleStringType>::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleStringType>
+struct is_constructible_string_type
+{
+    static constexpr auto value =
+        is_constructible<ConstructibleStringType,
+        typename BasicJsonType::string_t>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType, typename = void>
+struct is_compatible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type_impl <
+    BasicJsonType, CompatibleArrayType,
+    enable_if_t <
+    is_detected<iterator_t, CompatibleArrayType>::value&&
+    is_iterator_traits<iterator_traits<detected_t<iterator_t, CompatibleArrayType>>>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+    !std::is_same<CompatibleArrayType, detected_t<range_value_t, CompatibleArrayType>>::value >>
+{
+    static constexpr bool value =
+        is_constructible<BasicJsonType,
+        range_value_t<CompatibleArrayType>>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type
+    : is_compatible_array_type_impl<BasicJsonType, CompatibleArrayType> {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType, typename = void>
+struct is_constructible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t<std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value >>
+            : std::true_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t < !std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value&&
+    !is_compatible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+    is_default_constructible<ConstructibleArrayType>::value&&
+(std::is_move_assignable<ConstructibleArrayType>::value ||
+ std::is_copy_assignable<ConstructibleArrayType>::value)&&
+is_detected<iterator_t, ConstructibleArrayType>::value&&
+is_iterator_traits<iterator_traits<detected_t<iterator_t, ConstructibleArrayType>>>::value&&
+is_detected<range_value_t, ConstructibleArrayType>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+!std::is_same<ConstructibleArrayType, detected_t<range_value_t, ConstructibleArrayType>>::value&&
+        is_complete_type <
+        detected_t<range_value_t, ConstructibleArrayType >>::value >>
+{
+    using value_type = range_value_t<ConstructibleArrayType>;
+
+    static constexpr bool value =
+        std::is_same<value_type,
+        typename BasicJsonType::array_t::value_type>::value ||
+        has_from_json<BasicJsonType,
+        value_type>::value ||
+        has_non_default_from_json <
+        BasicJsonType,
+        value_type >::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type
+    : is_constructible_array_type_impl<BasicJsonType, ConstructibleArrayType> {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType,
+         typename = void>
+struct is_compatible_integer_type_impl : std::false_type {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type_impl <
+    RealIntegerType, CompatibleNumberIntegerType,
+    enable_if_t < std::is_integral<RealIntegerType>::value&&
+    std::is_integral<CompatibleNumberIntegerType>::value&&
+    !std::is_same<bool, CompatibleNumberIntegerType>::value >>
+{
+    // is there an assert somewhere on overflows?
+    using RealLimits = std::numeric_limits<RealIntegerType>;
+    using CompatibleLimits = std::numeric_limits<CompatibleNumberIntegerType>;
+
+    static constexpr auto value =
+        is_constructible<RealIntegerType,
+        CompatibleNumberIntegerType>::value &&
+        CompatibleLimits::is_integer &&
+        RealLimits::is_signed == CompatibleLimits::is_signed;
+};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type
+    : is_compatible_integer_type_impl<RealIntegerType,
+      CompatibleNumberIntegerType> {};
+
+template<typename BasicJsonType, typename CompatibleType, typename = void>
+struct is_compatible_type_impl: std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type_impl <
+    BasicJsonType, CompatibleType,
+    enable_if_t<is_complete_type<CompatibleType>::value >>
+{
+    static constexpr bool value =
+        has_to_json<BasicJsonType, CompatibleType>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type
+    : is_compatible_type_impl<BasicJsonType, CompatibleType> {};
+
+template<typename T1, typename T2>
+struct is_constructible_tuple : std::false_type {};
+
+template<typename T1, typename... Args>
+struct is_constructible_tuple<T1, std::tuple<Args...>> : conjunction<is_constructible<T1, Args>...> {};
+
+// a naive helper to check if a type is an ordered_map (exploits the fact that
+// ordered_map inherits capacity() from std::vector)
+template <typename T>
+struct is_ordered_map
+{
+    using one = char;
+
+    struct two
+    {
+        char x[2]; // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    };
+
+    template <typename C> static one test( decltype(&C::capacity) ) ;
+    template <typename C> static two test(...);
+
+    enum { value = sizeof(test<T>(nullptr)) == sizeof(char) }; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+};
+
+// to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
+template < typename T, typename U, enable_if_t < !std::is_same<T, U>::value, int > = 0 >
+T conditional_static_cast(U value)
+{
+    return static_cast<T>(value);
+}
+
+template<typename T, typename U, enable_if_t<std::is_same<T, U>::value, int> = 0>
+T conditional_static_cast(U value)
+{
+    return value;
+}
+
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+#if JSON_HAS_EXPERIMENTAL_FILESYSTEM
+#include <experimental/filesystem>
+namespace nlohmann::detail
+{
+namespace std_fs = std::experimental::filesystem;
+} // namespace nlohmann::detail
+#elif JSON_HAS_FILESYSTEM
+#include <filesystem>
+namespace nlohmann::detail
+{
+namespace std_fs = std::filesystem;
+} // namespace nlohmann::detail
+#endif
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename std::nullptr_t& n)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_null()))
+    {
+        JSON_THROW(type_error::create(302, "type must be null, but is " + std::string(j.type_name()), j));
+    }
+    n = nullptr;
+}
+
+// overloads for basic_json template parameters
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t < std::is_arithmetic<ArithmeticType>::value&&
+                         !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+                         int > = 0 >
+void get_arithmetic_value(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::boolean:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, "type must be number, but is " + std::string(j.type_name()), j));
+    }
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::boolean_t& b)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_boolean()))
+    {
+        JSON_THROW(type_error::create(302, "type must be boolean, but is " + std::string(j.type_name()), j));
+    }
+    b = *j.template get_ptr<const typename BasicJsonType::boolean_t*>();
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::string_t& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, "type must be string, but is " + std::string(j.type_name()), j));
+    }
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template <
+    typename BasicJsonType, typename ConstructibleStringType,
+    enable_if_t <
+        is_constructible_string_type<BasicJsonType, ConstructibleStringType>::value&&
+        !std::is_same<typename BasicJsonType::string_t,
+                      ConstructibleStringType>::value,
+        int > = 0 >
+void from_json(const BasicJsonType& j, ConstructibleStringType& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, "type must be string, but is " + std::string(j.type_name()), j));
+    }
+
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::number_float_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::number_unsigned_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::number_integer_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+void from_json(const BasicJsonType& j, EnumType& e)
+{
+    typename std::underlying_type<EnumType>::type val;
+    get_arithmetic_value(j, val);
+    e = static_cast<EnumType>(val);
+}
+
+// forward_list doesn't have an insert method
+template<typename BasicJsonType, typename T, typename Allocator,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+void from_json(const BasicJsonType& j, std::forward_list<T, Allocator>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+    l.clear();
+    std::transform(j.rbegin(), j.rend(),
+                   std::front_inserter(l), [](const BasicJsonType & i)
+    {
+        return i.template get<T>();
+    });
+}
+
+// valarray doesn't have an insert method
+template<typename BasicJsonType, typename T,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+void from_json(const BasicJsonType& j, std::valarray<T>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+    l.resize(j.size());
+    std::transform(j.begin(), j.end(), std::begin(l),
+                   [](const BasicJsonType & elem)
+    {
+        return elem.template get<T>();
+    });
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json(const BasicJsonType& j, T (&arr)[N])  // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType>
+void from_json_array_impl(const BasicJsonType& j, typename BasicJsonType::array_t& arr, priority_tag<3> /*unused*/)
+{
+    arr = *j.template get_ptr<const typename BasicJsonType::array_t*>();
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json_array_impl(const BasicJsonType& j, std::array<T, N>& arr,
+                          priority_tag<2> /*unused*/)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+auto from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr, priority_tag<1> /*unused*/)
+-> decltype(
+    arr.reserve(std::declval<typename ConstructibleArrayType::size_type>()),
+    j.template get<typename ConstructibleArrayType::value_type>(),
+    void())
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    ret.reserve(j.size());
+    std::transform(j.begin(), j.end(),
+                   std::inserter(ret, end(ret)), [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+void from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr,
+                          priority_tag<0> /*unused*/)
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    std::transform(
+        j.begin(), j.end(), std::inserter(ret, end(ret)),
+        [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template < typename BasicJsonType, typename ConstructibleArrayType,
+           enable_if_t <
+               is_constructible_array_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_object_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !std::is_same<ConstructibleArrayType, typename BasicJsonType::binary_t>::value&&
+               !is_basic_json<ConstructibleArrayType>::value,
+               int > = 0 >
+auto from_json(const BasicJsonType& j, ConstructibleArrayType& arr)
+-> decltype(from_json_array_impl(j, arr, priority_tag<3> {}),
+j.template get<typename ConstructibleArrayType::value_type>(),
+void())
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+
+    from_json_array_impl(j, arr, priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename T, std::size_t... Idx >
+std::array<T, sizeof...(Idx)> from_json_inplace_array_impl(BasicJsonType&& j,
+        identity_tag<std::array<T, sizeof...(Idx)>> /*unused*/, index_sequence<Idx...> /*unused*/)
+{
+    return { { std::forward<BasicJsonType>(j).at(Idx).template get<T>()... } };
+}
+
+template < typename BasicJsonType, typename T, std::size_t N >
+auto from_json(BasicJsonType&& j, identity_tag<std::array<T, N>> tag)
+-> decltype(from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+
+    return from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {});
+}
+
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, typename BasicJsonType::binary_t& bin)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_binary()))
+    {
+        JSON_THROW(type_error::create(302, "type must be binary, but is " + std::string(j.type_name()), j));
+    }
+
+    bin = *j.template get_ptr<const typename BasicJsonType::binary_t*>();
+}
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         enable_if_t<is_constructible_object_type<BasicJsonType, ConstructibleObjectType>::value, int> = 0>
+void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_object()))
+    {
+        JSON_THROW(type_error::create(302, "type must be object, but is " + std::string(j.type_name()), j));
+    }
+
+    ConstructibleObjectType ret;
+    const auto* inner_object = j.template get_ptr<const typename BasicJsonType::object_t*>();
+    using value_type = typename ConstructibleObjectType::value_type;
+    std::transform(
+        inner_object->begin(), inner_object->end(),
+        std::inserter(ret, ret.begin()),
+        [](typename BasicJsonType::object_t::value_type const & p)
+    {
+        return value_type(p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>());
+    });
+    obj = std::move(ret);
+}
+
+// overload for arithmetic types, not chosen for basic_json template arguments
+// (BooleanType, etc..); note: Is it really necessary to provide explicit
+// overloads for boolean_t etc. in case of a custom BooleanType which is not
+// an arithmetic type?
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t <
+               std::is_arithmetic<ArithmeticType>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_unsigned_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_integer_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_float_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+               int > = 0 >
+void from_json(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+        case value_t::boolean:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::boolean_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, "type must be number, but is " + std::string(j.type_name()), j));
+    }
+}
+
+template<typename BasicJsonType, typename... Args, std::size_t... Idx>
+std::tuple<Args...> from_json_tuple_impl_base(BasicJsonType&& j, index_sequence<Idx...> /*unused*/)
+{
+    return std::make_tuple(std::forward<BasicJsonType>(j).at(Idx).template get<Args>()...);
+}
+
+template < typename BasicJsonType, class A1, class A2 >
+std::pair<A1, A2> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::pair<A1, A2>> /*unused*/, priority_tag<0> /*unused*/)
+{
+    return {std::forward<BasicJsonType>(j).at(0).template get<A1>(),
+            std::forward<BasicJsonType>(j).at(1).template get<A2>()};
+}
+
+template<typename BasicJsonType, typename A1, typename A2>
+void from_json_tuple_impl(BasicJsonType&& j, std::pair<A1, A2>& p, priority_tag<1> /*unused*/)
+{
+    p = from_json_tuple_impl(std::forward<BasicJsonType>(j), identity_tag<std::pair<A1, A2>> {}, priority_tag<0> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+std::tuple<Args...> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::tuple<Args...>> /*unused*/, priority_tag<2> /*unused*/)
+{
+    return from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+void from_json_tuple_impl(BasicJsonType&& j, std::tuple<Args...>& t, priority_tag<3> /*unused*/)
+{
+    t = from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename TupleRelated>
+auto from_json(BasicJsonType&& j, TupleRelated&& t)
+-> decltype(from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+
+    return from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Compare, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+void from_json(const BasicJsonType& j, std::map<Key, Value, Compare, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(p.type_name()), j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Hash, typename KeyEqual, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+void from_json(const BasicJsonType& j, std::unordered_map<Key, Value, Hash, KeyEqual, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(j.type_name()), j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, "type must be array, but is " + std::string(p.type_name()), j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+void from_json(const BasicJsonType& j, std_fs::path& p)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, "type must be string, but is " + std::string(j.type_name()), j));
+    }
+    p = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+#endif
+
+struct from_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(const BasicJsonType& j, T&& val) const
+    noexcept(noexcept(from_json(j, std::forward<T>(val))))
+    -> decltype(from_json(j, std::forward<T>(val)))
+    {
+        return from_json(j, std::forward<T>(val));
+    }
+};
+}  // namespace detail
+
+/// namespace to hold default `from_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+constexpr const auto& from_json = detail::static_const<detail::from_json_fn>::value; // NOLINT(misc-definitions-in-headers)
+} // namespace
+} // namespace nlohmann
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+
+
+#include <algorithm> // copy
+#include <iterator> // begin, end
+#include <string> // string
+#include <tuple> // tuple, get
+#include <type_traits> // is_same, is_constructible, is_floating_point, is_enum, underlying_type
+#include <utility> // move, forward, declval, pair
+#include <valarray> // valarray
+#include <vector> // vector
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+
+
+#include <cstddef> // size_t
+#include <iterator> // input_iterator_tag
+#include <string> // string, to_string
+#include <tuple> // tuple_size, get, tuple_element
+#include <utility> // move
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename string_type>
+void int_to_string( string_type& target, std::size_t value )
+{
+    // For ADL
+    using std::to_string;
+    target = to_string(value);
+}
+template<typename IteratorType> class iteration_proxy_value
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    using value_type = iteration_proxy_value;
+    using pointer = value_type * ;
+    using reference = value_type & ;
+    using iterator_category = std::input_iterator_tag;
+    using string_type = typename std::remove_cv< typename std::remove_reference<decltype( std::declval<IteratorType>().key() ) >::type >::type;
+
+  private:
+    /// the iterator
+    IteratorType anchor;
+    /// an index for arrays (used to create key names)
+    std::size_t array_index = 0;
+    /// last stringified array index
+    mutable std::size_t array_index_last = 0;
+    /// a string representation of the array index
+    mutable string_type array_index_str = "0";
+    /// an empty string (to return a reference for primitive values)
+    const string_type empty_str{};
+
+  public:
+    explicit iteration_proxy_value(IteratorType it) noexcept
+        : anchor(std::move(it))
+    {}
+
+    /// dereference operator (needed for range-based for)
+    iteration_proxy_value& operator*()
+    {
+        return *this;
+    }
+
+    /// increment operator (needed for range-based for)
+    iteration_proxy_value& operator++()
+    {
+        ++anchor;
+        ++array_index;
+
+        return *this;
+    }
+
+    /// equality operator (needed for InputIterator)
+    bool operator==(const iteration_proxy_value& o) const
+    {
+        return anchor == o.anchor;
+    }
+
+    /// inequality operator (needed for range-based for)
+    bool operator!=(const iteration_proxy_value& o) const
+    {
+        return anchor != o.anchor;
+    }
+
+    /// return key of the iterator
+    const string_type& key() const
+    {
+        JSON_ASSERT(anchor.m_object != nullptr);
+
+        switch (anchor.m_object->type())
+        {
+            // use integer array index as key
+            case value_t::array:
+            {
+                if (array_index != array_index_last)
+                {
+                    int_to_string( array_index_str, array_index );
+                    array_index_last = array_index;
+                }
+                return array_index_str;
+            }
+
+            // use key from the object
+            case value_t::object:
+                return anchor.key();
+
+            // use an empty key for all primitive types
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return empty_str;
+        }
+    }
+
+    /// return value of the iterator
+    typename IteratorType::reference value() const
+    {
+        return anchor.value();
+    }
+};
+
+/// proxy class for the items() function
+template<typename IteratorType> class iteration_proxy
+{
+  private:
+    /// the container to iterate
+    typename IteratorType::reference container;
+
+  public:
+    /// construct iteration proxy from a container
+    explicit iteration_proxy(typename IteratorType::reference cont) noexcept
+        : container(cont) {}
+
+    /// return iterator begin (needed for range-based for)
+    iteration_proxy_value<IteratorType> begin() noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container.begin());
+    }
+
+    /// return iterator end (needed for range-based for)
+    iteration_proxy_value<IteratorType> end() noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container.end());
+    }
+};
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 0, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.key())
+{
+    return i.key();
+}
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 1, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.value())
+{
+    return i.value();
+}
+}  // namespace detail
+}  // namespace nlohmann
+
+// The Addition to the STD Namespace is required to add
+// Structured Bindings Support to the iteration_proxy_value class
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+namespace std
+{
+#if defined(__clang__)
+    // Fix: https://github.com/nlohmann/json/issues/1401
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wmismatched-tags"
+#endif
+template<typename IteratorType>
+class tuple_size<::nlohmann::detail::iteration_proxy_value<IteratorType>>
+            : public std::integral_constant<std::size_t, 2> {};
+
+template<std::size_t N, typename IteratorType>
+class tuple_element<N, ::nlohmann::detail::iteration_proxy_value<IteratorType >>
+{
+  public:
+    using type = decltype(
+                     get<N>(std::declval <
+                            ::nlohmann::detail::iteration_proxy_value<IteratorType >> ()));
+};
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+} // namespace std
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+#if JSON_HAS_EXPERIMENTAL_FILESYSTEM
+#include <experimental/filesystem>
+namespace nlohmann::detail
+{
+namespace std_fs = std::experimental::filesystem;
+} // namespace nlohmann::detail
+#elif JSON_HAS_FILESYSTEM
+#include <filesystem>
+namespace nlohmann::detail
+{
+namespace std_fs = std::filesystem;
+} // namespace nlohmann::detail
+#endif
+
+namespace nlohmann
+{
+namespace detail
+{
+//////////////////
+// constructors //
+//////////////////
+
+/*
+ * Note all external_constructor<>::construct functions need to call
+ * j.m_value.destroy(j.m_type) to avoid a memory leak in case j contains an
+ * allocated value (e.g., a string). See bug issue
+ * https://github.com/nlohmann/json/issues/2865 for more information.
+ */
+
+template<value_t> struct external_constructor;
+
+template<>
+struct external_constructor<value_t::boolean>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::boolean_t b) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::boolean;
+        j.m_value = b;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::string>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::string_t& s)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value = s;
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value = std::move(s);
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleStringType,
+               enable_if_t < !std::is_same<CompatibleStringType, typename BasicJsonType::string_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleStringType& str)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value.string = j.template create<typename BasicJsonType::string_t>(str);
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::binary>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::binary_t& b)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::binary;
+        j.m_value = typename BasicJsonType::binary_t(b);
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::binary_t&& b)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::binary;
+        j.m_value = typename BasicJsonType::binary_t(std::move(b));
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_float>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_float_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_float;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_unsigned>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_unsigned_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_unsigned;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_integer>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_integer_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_integer;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::array>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::array_t& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = arr;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = std::move(arr);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleArrayType,
+               enable_if_t < !std::is_same<CompatibleArrayType, typename BasicJsonType::array_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleArrayType& arr)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value.array = j.template create<typename BasicJsonType::array_t>(begin(arr), end(arr));
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const std::vector<bool>& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = value_t::array;
+        j.m_value.array->reserve(arr.size());
+        for (const bool x : arr)
+        {
+            j.m_value.array->push_back(x);
+            j.set_parent(j.m_value.array->back());
+        }
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType, typename T,
+             enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+    static void construct(BasicJsonType& j, const std::valarray<T>& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = value_t::array;
+        j.m_value.array->resize(arr.size());
+        if (arr.size() > 0)
+        {
+            std::copy(std::begin(arr), std::end(arr), j.m_value.array->begin());
+        }
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::object>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::object_t& obj)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value = obj;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value = std::move(obj);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleObjectType,
+               enable_if_t < !std::is_same<CompatibleObjectType, typename BasicJsonType::object_t>::value, int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleObjectType& obj)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value.object = j.template create<typename BasicJsonType::object_t>(begin(obj), end(obj));
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+/////////////
+// to_json //
+/////////////
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, typename BasicJsonType::boolean_t>::value, int> = 0>
+void to_json(BasicJsonType& j, T b) noexcept
+{
+    external_constructor<value_t::boolean>::construct(j, b);
+}
+
+template<typename BasicJsonType, typename CompatibleString,
+         enable_if_t<std::is_constructible<typename BasicJsonType::string_t, CompatibleString>::value, int> = 0>
+void to_json(BasicJsonType& j, const CompatibleString& s)
+{
+    external_constructor<value_t::string>::construct(j, s);
+}
+
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+{
+    external_constructor<value_t::string>::construct(j, std::move(s));
+}
+
+template<typename BasicJsonType, typename FloatType,
+         enable_if_t<std::is_floating_point<FloatType>::value, int> = 0>
+void to_json(BasicJsonType& j, FloatType val) noexcept
+{
+    external_constructor<value_t::number_float>::construct(j, static_cast<typename BasicJsonType::number_float_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberUnsignedType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_unsigned_t, CompatibleNumberUnsignedType>::value, int> = 0>
+void to_json(BasicJsonType& j, CompatibleNumberUnsignedType val) noexcept
+{
+    external_constructor<value_t::number_unsigned>::construct(j, static_cast<typename BasicJsonType::number_unsigned_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberIntegerType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType>::value, int> = 0>
+void to_json(BasicJsonType& j, CompatibleNumberIntegerType val) noexcept
+{
+    external_constructor<value_t::number_integer>::construct(j, static_cast<typename BasicJsonType::number_integer_t>(val));
+}
+
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+void to_json(BasicJsonType& j, EnumType e) noexcept
+{
+    using underlying_type = typename std::underlying_type<EnumType>::type;
+    external_constructor<value_t::number_integer>::construct(j, static_cast<underlying_type>(e));
+}
+
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, const std::vector<bool>& e)
+{
+    external_constructor<value_t::array>::construct(j, e);
+}
+
+template < typename BasicJsonType, typename CompatibleArrayType,
+           enable_if_t < is_compatible_array_type<BasicJsonType,
+                         CompatibleArrayType>::value&&
+                         !is_compatible_object_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !is_compatible_string_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !std::is_same<typename BasicJsonType::binary_t, CompatibleArrayType>::value&&
+                         !is_basic_json<CompatibleArrayType>::value,
+                         int > = 0 >
+void to_json(BasicJsonType& j, const CompatibleArrayType& arr)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, const typename BasicJsonType::binary_t& bin)
+{
+    external_constructor<value_t::binary>::construct(j, bin);
+}
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+void to_json(BasicJsonType& j, const std::valarray<T>& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template < typename BasicJsonType, typename CompatibleObjectType,
+           enable_if_t < is_compatible_object_type<BasicJsonType, CompatibleObjectType>::value&& !is_basic_json<CompatibleObjectType>::value, int > = 0 >
+void to_json(BasicJsonType& j, const CompatibleObjectType& obj)
+{
+    external_constructor<value_t::object>::construct(j, obj);
+}
+
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+{
+    external_constructor<value_t::object>::construct(j, std::move(obj));
+}
+
+template <
+    typename BasicJsonType, typename T, std::size_t N,
+    enable_if_t < !std::is_constructible<typename BasicJsonType::string_t,
+                  const T(&)[N]>::value, // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+                  int > = 0 >
+void to_json(BasicJsonType& j, const T(&arr)[N]) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template < typename BasicJsonType, typename T1, typename T2, enable_if_t < std::is_constructible<BasicJsonType, T1>::value&& std::is_constructible<BasicJsonType, T2>::value, int > = 0 >
+void to_json(BasicJsonType& j, const std::pair<T1, T2>& p)
+{
+    j = { p.first, p.second };
+}
+
+// for https://github.com/nlohmann/json/pull/1134
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, iteration_proxy_value<typename BasicJsonType::iterator>>::value, int> = 0>
+void to_json(BasicJsonType& j, const T& b)
+{
+    j = { {b.key(), b.value()} };
+}
+
+template<typename BasicJsonType, typename Tuple, std::size_t... Idx>
+void to_json_tuple_impl(BasicJsonType& j, const Tuple& t, index_sequence<Idx...> /*unused*/)
+{
+    j = { std::get<Idx>(t)... };
+}
+
+template<typename BasicJsonType, typename T, enable_if_t<is_constructible_tuple<BasicJsonType, T>::value, int > = 0>
+void to_json(BasicJsonType& j, const T& t)
+{
+    to_json_tuple_impl(j, t, make_index_sequence<std::tuple_size<T>::value> {});
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+void to_json(BasicJsonType& j, const std_fs::path& p)
+{
+    j = p.string();
+}
+#endif
+
+struct to_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(BasicJsonType& j, T&& val) const noexcept(noexcept(to_json(j, std::forward<T>(val))))
+    -> decltype(to_json(j, std::forward<T>(val)), void())
+    {
+        return to_json(j, std::forward<T>(val));
+    }
+};
+}  // namespace detail
+
+/// namespace to hold default `to_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+constexpr const auto& to_json = detail::static_const<detail::to_json_fn>::value; // NOLINT(misc-definitions-in-headers)
+} // namespace
+} // namespace nlohmann
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+namespace nlohmann
+{
+
+/// @sa https://json.nlohmann.me/api/adl_serializer/
+template<typename ValueType, typename>
+struct adl_serializer
+{
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j, TargetType& val) noexcept(
+        noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), val)))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), val), void())
+    {
+        ::nlohmann::from_json(std::forward<BasicJsonType>(j), val);
+    }
+
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j) noexcept(
+    noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {})))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {}))
+    {
+        return ::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {});
+    }
+
+    /// @brief convert any value type to a JSON value
+    /// @sa https://json.nlohmann.me/api/adl_serializer/to_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto to_json(BasicJsonType& j, TargetType && val) noexcept(
+        noexcept(::nlohmann::to_json(j, std::forward<TargetType>(val))))
+    -> decltype(::nlohmann::to_json(j, std::forward<TargetType>(val)), void())
+    {
+        ::nlohmann::to_json(j, std::forward<TargetType>(val));
+    }
+};
+}  // namespace nlohmann
+
+// #include <nlohmann/byte_container_with_subtype.hpp>
+
+
+#include <cstdint> // uint8_t, uint64_t
+#include <tuple> // tie
+#include <utility> // move
+
+namespace nlohmann
+{
+
+/// @brief an internal type for a backed binary type
+/// @sa https://json.nlohmann.me/api/byte_container_with_subtype/
+template<typename BinaryType>
+class byte_container_with_subtype : public BinaryType
+{
+  public:
+    using container_type = BinaryType;
+    using subtype_type = std::uint64_t;
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype() noexcept(noexcept(container_type()))
+        : container_type()
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b, subtype_type subtype_) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b, subtype_type subtype_) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    bool operator==(const byte_container_with_subtype& rhs) const
+    {
+        return std::tie(static_cast<const BinaryType&>(*this), m_subtype, m_has_subtype) ==
+               std::tie(static_cast<const BinaryType&>(rhs), rhs.m_subtype, rhs.m_has_subtype);
+    }
+
+    bool operator!=(const byte_container_with_subtype& rhs) const
+    {
+        return !(rhs == *this);
+    }
+
+    /// @brief sets the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/set_subtype/
+    void set_subtype(subtype_type subtype_) noexcept
+    {
+        m_subtype = subtype_;
+        m_has_subtype = true;
+    }
+
+    /// @brief return the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/subtype/
+    constexpr subtype_type subtype() const noexcept
+    {
+        return m_has_subtype ? m_subtype : static_cast<subtype_type>(-1);
+    }
+
+    /// @brief return whether the value has a subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/has_subtype/
+    constexpr bool has_subtype() const noexcept
+    {
+        return m_has_subtype;
+    }
+
+    /// @brief clears the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/clear_subtype/
+    void clear_subtype() noexcept
+    {
+        m_subtype = 0;
+        m_has_subtype = false;
+    }
+
+  private:
+    subtype_type m_subtype = 0;
+    bool m_has_subtype = false;
+};
+
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/hash.hpp>
+
+
+#include <cstdint> // uint8_t
+#include <cstddef> // size_t
+#include <functional> // hash
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+
+// boost::hash_combine
+inline std::size_t combine(std::size_t seed, std::size_t h) noexcept
+{
+    seed ^= h + 0x9e3779b9 + (seed << 6U) + (seed >> 2U);
+    return seed;
+}
+
+/*!
+@brief hash a JSON value
+
+The hash function tries to rely on std::hash where possible. Furthermore, the
+type of the JSON value is taken into account to has different hash values for
+null, 0, 0U, and false, etc.
+
+@tparam BasicJsonType basic_json specialization
+@param j JSON value to hash
+@return hash value of j
+*/
+template<typename BasicJsonType>
+std::size_t hash(const BasicJsonType& j)
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+    const auto type = static_cast<std::size_t>(j.type());
+    switch (j.type())
+    {
+        case BasicJsonType::value_t::null:
+        case BasicJsonType::value_t::discarded:
+        {
+            return combine(type, 0);
+        }
+
+        case BasicJsonType::value_t::object:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j.items())
+            {
+                const auto h = std::hash<string_t> {}(element.key());
+                seed = combine(seed, h);
+                seed = combine(seed, hash(element.value()));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::array:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j)
+            {
+                seed = combine(seed, hash(element));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::string:
+        {
+            const auto h = std::hash<string_t> {}(j.template get_ref<const string_t&>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::boolean:
+        {
+            const auto h = std::hash<bool> {}(j.template get<bool>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_integer:
+        {
+            const auto h = std::hash<number_integer_t> {}(j.template get<number_integer_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_unsigned:
+        {
+            const auto h = std::hash<number_unsigned_t> {}(j.template get<number_unsigned_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_float:
+        {
+            const auto h = std::hash<number_float_t> {}(j.template get<number_float_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::binary:
+        {
+            auto seed = combine(type, j.get_binary().size());
+            const auto h = std::hash<bool> {}(j.get_binary().has_subtype());
+            seed = combine(seed, h);
+            seed = combine(seed, static_cast<std::size_t>(j.get_binary().subtype()));
+            for (const auto byte : j.get_binary())
+            {
+                seed = combine(seed, std::hash<std::uint8_t> {}(byte));
+            }
+            return seed;
+        }
+
+        default:                   // LCOV_EXCL_LINE
+            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            return 0;              // LCOV_EXCL_LINE
+    }
+}
+
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+
+
+#include <algorithm> // generate_n
+#include <array> // array
+#include <cmath> // ldexp
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstdio> // snprintf
+#include <cstring> // memcpy
+#include <iterator> // back_inserter
+#include <limits> // numeric_limits
+#include <string> // char_traits, string
+#include <utility> // make_pair, move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstring> // strlen
+#include <iterator> // begin, end, iterator_traits, random_access_iterator_tag, distance, next
+#include <memory> // shared_ptr, make_shared, addressof
+#include <numeric> // accumulate
+#include <string> // string, char_traits
+#include <type_traits> // enable_if, is_base_of, is_pointer, is_integral, remove_pointer
+#include <utility> // pair, declval
+
+#ifndef JSON_NO_IO
+    #include <cstdio>   // FILE *
+    #include <istream>  // istream
+#endif                  // JSON_NO_IO
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+/// the supported input formats
+enum class input_format_t { json, cbor, msgpack, ubjson, bson };
+
+////////////////////
+// input adapters //
+////////////////////
+
+#ifndef JSON_NO_IO
+/*!
+Input adapter for stdio file access. This adapter read only 1 byte and do not use any
+ buffer. This adapter is a very low level adapter.
+*/
+class file_input_adapter
+{
+  public:
+    using char_type = char;
+
+    JSON_HEDLEY_NON_NULL(2)
+    explicit file_input_adapter(std::FILE* f) noexcept
+        : m_file(f)
+    {}
+
+    // make class move-only
+    file_input_adapter(const file_input_adapter&) = delete;
+    file_input_adapter(file_input_adapter&&) noexcept = default;
+    file_input_adapter& operator=(const file_input_adapter&) = delete;
+    file_input_adapter& operator=(file_input_adapter&&) = delete;
+    ~file_input_adapter() = default;
+
+    std::char_traits<char>::int_type get_character() noexcept
+    {
+        return std::fgetc(m_file);
+    }
+
+  private:
+    /// the file pointer to read from
+    std::FILE* m_file;
+};
+
+
+/*!
+Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at
+beginning of input. Does not support changing the underlying std::streambuf
+in mid-input. Maintains underlying std::istream and std::streambuf to support
+subsequent use of standard std::istream operations to process any input
+characters following those used in parsing the JSON input.  Clears the
+std::istream flags; any input errors (e.g., EOF) will be detected by the first
+subsequent call for input from the std::istream.
+*/
+class input_stream_adapter
+{
+  public:
+    using char_type = char;
+
+    ~input_stream_adapter()
+    {
+        // clear stream flags; we use underlying streambuf I/O, do not
+        // maintain ifstream flags, except eof
+        if (is != nullptr)
+        {
+            is->clear(is->rdstate() & std::ios::eofbit);
+        }
+    }
+
+    explicit input_stream_adapter(std::istream& i)
+        : is(&i), sb(i.rdbuf())
+    {}
+
+    // delete because of pointer members
+    input_stream_adapter(const input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&&) = delete;
+
+    input_stream_adapter(input_stream_adapter&& rhs) noexcept
+        : is(rhs.is), sb(rhs.sb)
+    {
+        rhs.is = nullptr;
+        rhs.sb = nullptr;
+    }
+
+    // std::istream/std::streambuf use std::char_traits<char>::to_int_type, to
+    // ensure that std::char_traits<char>::eof() and the character 0xFF do not
+    // end up as the same value, e.g. 0xFFFFFFFF.
+    std::char_traits<char>::int_type get_character()
+    {
+        auto res = sb->sbumpc();
+        // set eof manually, as we don't use the istream interface.
+        if (JSON_HEDLEY_UNLIKELY(res == std::char_traits<char>::eof()))
+        {
+            is->clear(is->rdstate() | std::ios::eofbit);
+        }
+        return res;
+    }
+
+  private:
+    /// the associated input stream
+    std::istream* is = nullptr;
+    std::streambuf* sb = nullptr;
+};
+#endif  // JSON_NO_IO
+
+// General-purpose iterator-based adapter. It might not be as fast as
+// theoretically possible for some containers, but it is extremely versatile.
+template<typename IteratorType>
+class iterator_input_adapter
+{
+  public:
+    using char_type = typename std::iterator_traits<IteratorType>::value_type;
+
+    iterator_input_adapter(IteratorType first, IteratorType last)
+        : current(std::move(first)), end(std::move(last))
+    {}
+
+    typename std::char_traits<char_type>::int_type get_character()
+    {
+        if (JSON_HEDLEY_LIKELY(current != end))
+        {
+            auto result = std::char_traits<char_type>::to_int_type(*current);
+            std::advance(current, 1);
+            return result;
+        }
+
+        return std::char_traits<char_type>::eof();
+    }
+
+  private:
+    IteratorType current;
+    IteratorType end;
+
+    template<typename BaseInputAdapter, size_t T>
+    friend struct wide_string_input_helper;
+
+    bool empty() const
+    {
+        return current == end;
+    }
+};
+
+
+template<typename BaseInputAdapter, size_t T>
+struct wide_string_input_helper;
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 4>
+{
+    // UTF-32
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-32 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u) & 0x1Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (wc <= 0xFFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u) & 0x0Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else if (wc <= 0x10FFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | ((static_cast<unsigned int>(wc) >> 18u) & 0x07u));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 12u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 4;
+            }
+            else
+            {
+                // unknown character
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+        }
+    }
+};
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 2>
+{
+    // UTF-16
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-16 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (0xD800 > wc || wc >= 0xE000)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else
+            {
+                if (JSON_HEDLEY_UNLIKELY(!input.empty()))
+                {
+                    const auto wc2 = static_cast<unsigned int>(input.get_character());
+                    const auto charcode = 0x10000u + (((static_cast<unsigned int>(wc) & 0x3FFu) << 10u) | (wc2 & 0x3FFu));
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | (charcode >> 18u));
+                    utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 12u) & 0x3Fu));
+                    utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 6u) & 0x3Fu));
+                    utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (charcode & 0x3Fu));
+                    utf8_bytes_filled = 4;
+                }
+                else
+                {
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                    utf8_bytes_filled = 1;
+                }
+            }
+        }
+    }
+};
+
+// Wraps another input apdater to convert wide character types into individual bytes.
+template<typename BaseInputAdapter, typename WideCharType>
+class wide_string_input_adapter
+{
+  public:
+    using char_type = char;
+
+    wide_string_input_adapter(BaseInputAdapter base)
+        : base_adapter(base) {}
+
+    typename std::char_traits<char>::int_type get_character() noexcept
+    {
+        // check if buffer needs to be filled
+        if (utf8_bytes_index == utf8_bytes_filled)
+        {
+            fill_buffer<sizeof(WideCharType)>();
+
+            JSON_ASSERT(utf8_bytes_filled > 0);
+            JSON_ASSERT(utf8_bytes_index == 0);
+        }
+
+        // use buffer
+        JSON_ASSERT(utf8_bytes_filled > 0);
+        JSON_ASSERT(utf8_bytes_index < utf8_bytes_filled);
+        return utf8_bytes[utf8_bytes_index++];
+    }
+
+  private:
+    BaseInputAdapter base_adapter;
+
+    template<size_t T>
+    void fill_buffer()
+    {
+        wide_string_input_helper<BaseInputAdapter, T>::fill_buffer(base_adapter, utf8_bytes, utf8_bytes_index, utf8_bytes_filled);
+    }
+
+    /// a buffer for UTF-8 bytes
+    std::array<std::char_traits<char>::int_type, 4> utf8_bytes = {{0, 0, 0, 0}};
+
+    /// index to the utf8_codes array for the next valid byte
+    std::size_t utf8_bytes_index = 0;
+    /// number of valid bytes in the utf8_codes array
+    std::size_t utf8_bytes_filled = 0;
+};
+
+
+template<typename IteratorType, typename Enable = void>
+struct iterator_input_adapter_factory
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using adapter_type = iterator_input_adapter<iterator_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(std::move(first), std::move(last));
+    }
+};
+
+template<typename T>
+struct is_iterator_of_multibyte
+{
+    using value_type = typename std::iterator_traits<T>::value_type;
+    enum
+    {
+        value = sizeof(value_type) > 1
+    };
+};
+
+template<typename IteratorType>
+struct iterator_input_adapter_factory<IteratorType, enable_if_t<is_iterator_of_multibyte<IteratorType>::value>>
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using base_adapter_type = iterator_input_adapter<iterator_type>;
+    using adapter_type = wide_string_input_adapter<base_adapter_type, char_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(base_adapter_type(std::move(first), std::move(last)));
+    }
+};
+
+// General purpose iterator-based input
+template<typename IteratorType>
+typename iterator_input_adapter_factory<IteratorType>::adapter_type input_adapter(IteratorType first, IteratorType last)
+{
+    using factory_type = iterator_input_adapter_factory<IteratorType>;
+    return factory_type::create(first, last);
+}
+
+// Convenience shorthand from container to iterator
+// Enables ADL on begin(container) and end(container)
+// Encloses the using declarations in namespace for not to leak them to outside scope
+
+namespace container_input_adapter_factory_impl
+{
+
+using std::begin;
+using std::end;
+
+template<typename ContainerType, typename Enable = void>
+struct container_input_adapter_factory {};
+
+template<typename ContainerType>
+struct container_input_adapter_factory< ContainerType,
+       void_t<decltype(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>()))>>
+       {
+           using adapter_type = decltype(input_adapter(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>())));
+
+           static adapter_type create(const ContainerType& container)
+{
+    return input_adapter(begin(container), end(container));
+}
+       };
+
+} // namespace container_input_adapter_factory_impl
+
+template<typename ContainerType>
+typename container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::adapter_type input_adapter(const ContainerType& container)
+{
+    return container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::create(container);
+}
+
+#ifndef JSON_NO_IO
+// Special cases with fast paths
+inline file_input_adapter input_adapter(std::FILE* file)
+{
+    return file_input_adapter(file);
+}
+
+inline input_stream_adapter input_adapter(std::istream& stream)
+{
+    return input_stream_adapter(stream);
+}
+
+inline input_stream_adapter input_adapter(std::istream&& stream)
+{
+    return input_stream_adapter(stream);
+}
+#endif  // JSON_NO_IO
+
+using contiguous_bytes_input_adapter = decltype(input_adapter(std::declval<const char*>(), std::declval<const char*>()));
+
+// Null-delimited strings, and the like.
+template < typename CharT,
+           typename std::enable_if <
+               std::is_pointer<CharT>::value&&
+               !std::is_array<CharT>::value&&
+               std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+               sizeof(typename std::remove_pointer<CharT>::type) == 1,
+               int >::type = 0 >
+contiguous_bytes_input_adapter input_adapter(CharT b)
+{
+    auto length = std::strlen(reinterpret_cast<const char*>(b));
+    const auto* ptr = reinterpret_cast<const char*>(b);
+    return input_adapter(ptr, ptr + length);
+}
+
+template<typename T, std::size_t N>
+auto input_adapter(T (&array)[N]) -> decltype(input_adapter(array, array + N)) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    return input_adapter(array, array + N);
+}
+
+// This class only handles inputs of input_buffer_adapter type.
+// It's required so that expressions like {ptr, len} can be implicitly cast
+// to the correct adapter.
+class span_input_adapter
+{
+  public:
+    template < typename CharT,
+               typename std::enable_if <
+                   std::is_pointer<CharT>::value&&
+                   std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+                   sizeof(typename std::remove_pointer<CharT>::type) == 1,
+                   int >::type = 0 >
+    span_input_adapter(CharT b, std::size_t l)
+        : ia(reinterpret_cast<const char*>(b), reinterpret_cast<const char*>(b) + l) {}
+
+    template<class IteratorType,
+             typename std::enable_if<
+                 std::is_same<typename iterator_traits<IteratorType>::iterator_category, std::random_access_iterator_tag>::value,
+                 int>::type = 0>
+    span_input_adapter(IteratorType first, IteratorType last)
+        : ia(input_adapter(first, last)) {}
+
+    contiguous_bytes_input_adapter&& get()
+    {
+        return std::move(ia); // NOLINT(hicpp-move-const-arg,performance-move-const-arg)
+    }
+
+  private:
+    contiguous_bytes_input_adapter ia;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+
+
+#include <cstddef>
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+
+/*!
+@brief SAX interface
+
+This class describes the SAX interface used by @ref nlohmann::json::sax_parse.
+Each function is called in different situations while the input is parsed. The
+boolean return value informs the parser whether to continue processing the
+input.
+*/
+template<typename BasicJsonType>
+struct json_sax
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @brief a null value was read
+    @return whether parsing should proceed
+    */
+    virtual bool null() = 0;
+
+    /*!
+    @brief a boolean value was read
+    @param[in] val  boolean value
+    @return whether parsing should proceed
+    */
+    virtual bool boolean(bool val) = 0;
+
+    /*!
+    @brief an integer number was read
+    @param[in] val  integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_integer(number_integer_t val) = 0;
+
+    /*!
+    @brief an unsigned integer number was read
+    @param[in] val  unsigned integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_unsigned(number_unsigned_t val) = 0;
+
+    /*!
+    @brief a floating-point number was read
+    @param[in] val  floating-point value
+    @param[in] s    raw token value
+    @return whether parsing should proceed
+    */
+    virtual bool number_float(number_float_t val, const string_t& s) = 0;
+
+    /*!
+    @brief a string value was read
+    @param[in] val  string value
+    @return whether parsing should proceed
+    @note It is safe to move the passed string value.
+    */
+    virtual bool string(string_t& val) = 0;
+
+    /*!
+    @brief a binary value was read
+    @param[in] val  binary value
+    @return whether parsing should proceed
+    @note It is safe to move the passed binary value.
+    */
+    virtual bool binary(binary_t& val) = 0;
+
+    /*!
+    @brief the beginning of an object was read
+    @param[in] elements  number of object elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_object(std::size_t elements) = 0;
+
+    /*!
+    @brief an object key was read
+    @param[in] val  object key
+    @return whether parsing should proceed
+    @note It is safe to move the passed string.
+    */
+    virtual bool key(string_t& val) = 0;
+
+    /*!
+    @brief the end of an object was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_object() = 0;
+
+    /*!
+    @brief the beginning of an array was read
+    @param[in] elements  number of array elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_array(std::size_t elements) = 0;
+
+    /*!
+    @brief the end of an array was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_array() = 0;
+
+    /*!
+    @brief a parse error occurred
+    @param[in] position    the position in the input where the error occurs
+    @param[in] last_token  the last read token
+    @param[in] ex          an exception object describing the error
+    @return whether parsing should proceed (must return false)
+    */
+    virtual bool parse_error(std::size_t position,
+                             const std::string& last_token,
+                             const detail::exception& ex) = 0;
+
+    json_sax() = default;
+    json_sax(const json_sax&) = default;
+    json_sax(json_sax&&) noexcept = default;
+    json_sax& operator=(const json_sax&) = default;
+    json_sax& operator=(json_sax&&) noexcept = default;
+    virtual ~json_sax() = default;
+};
+
+
+namespace detail
+{
+/*!
+@brief SAX implementation to create a JSON value from SAX events
+
+This class implements the @ref json_sax interface and processes the SAX events
+to create a JSON value which makes it basically a DOM parser. The structure or
+hierarchy of the JSON value is managed by the stack `ref_stack` which contains
+a pointer to the respective array or object for each recursion depth.
+
+After successful parsing, the value that is passed by reference to the
+constructor contains the parsed value.
+
+@tparam BasicJsonType  the JSON type
+*/
+template<typename BasicJsonType>
+class json_sax_dom_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @param[in,out] r  reference to a JSON value that is manipulated while
+                       parsing
+    @param[in] allow_exceptions_  whether parse errors yield exceptions
+    */
+    explicit json_sax_dom_parser(BasicJsonType& r, const bool allow_exceptions_ = true)
+        : root(r), allow_exceptions(allow_exceptions_)
+    {}
+
+    // make class move-only
+    json_sax_dom_parser(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_parser& operator=(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser& operator=(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::object));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, "excessive object size: " + std::to_string(len), *ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        // add null at given key and store the reference for later
+        object_element = &(ref_stack.back()->m_value.object->operator[](val));
+        return true;
+    }
+
+    bool end_object()
+    {
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::array));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, "excessive array size: " + std::to_string(len), *ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+    */
+    template<typename Value>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    BasicJsonType* handle_value(Value&& v)
+    {
+        if (ref_stack.empty())
+        {
+            root = BasicJsonType(std::forward<Value>(v));
+            return &root;
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->emplace_back(std::forward<Value>(v));
+            return &(ref_stack.back()->m_value.array->back());
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_object());
+        JSON_ASSERT(object_element);
+        *object_element = BasicJsonType(std::forward<Value>(v));
+        return object_element;
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+template<typename BasicJsonType>
+class json_sax_dom_callback_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using parser_callback_t = typename BasicJsonType::parser_callback_t;
+    using parse_event_t = typename BasicJsonType::parse_event_t;
+
+    json_sax_dom_callback_parser(BasicJsonType& r,
+                                 const parser_callback_t cb,
+                                 const bool allow_exceptions_ = true)
+        : root(r), callback(cb), allow_exceptions(allow_exceptions_)
+    {
+        keep_stack.push_back(true);
+    }
+
+    // make class move-only
+    json_sax_dom_callback_parser(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_callback_parser& operator=(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser& operator=(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_callback_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        // check callback for object start
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::object_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::object, true);
+        ref_stack.push_back(val.second);
+
+        // check object limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, "excessive object size: " + std::to_string(len), *ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        BasicJsonType k = BasicJsonType(val);
+
+        // check callback for key
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::key, k);
+        key_keep_stack.push_back(keep);
+
+        // add discarded value at given key and store the reference for later
+        if (keep && ref_stack.back())
+        {
+            object_element = &(ref_stack.back()->m_value.object->operator[](val) = discarded);
+        }
+
+        return true;
+    }
+
+    bool end_object()
+    {
+        if (ref_stack.back())
+        {
+            if (!callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::object_end, *ref_stack.back()))
+            {
+                // discard object
+                *ref_stack.back() = discarded;
+            }
+            else
+            {
+                ref_stack.back()->set_parents();
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        if (!ref_stack.empty() && ref_stack.back() && ref_stack.back()->is_structured())
+        {
+            // remove discarded value
+            for (auto it = ref_stack.back()->begin(); it != ref_stack.back()->end(); ++it)
+            {
+                if (it->is_discarded())
+                {
+                    ref_stack.back()->erase(it);
+                    break;
+                }
+            }
+        }
+
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::array_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::array, true);
+        ref_stack.push_back(val.second);
+
+        // check array limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, "excessive array size: " + std::to_string(len), *ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        bool keep = true;
+
+        if (ref_stack.back())
+        {
+            keep = callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::array_end, *ref_stack.back());
+            if (keep)
+            {
+                ref_stack.back()->set_parents();
+            }
+            else
+            {
+                // discard array
+                *ref_stack.back() = discarded;
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        // remove discarded value
+        if (!keep && !ref_stack.empty() && ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->pop_back();
+        }
+
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @param[in] v  value to add to the JSON value we build during parsing
+    @param[in] skip_callback  whether we should skip calling the callback
+               function; this is required after start_array() and
+               start_object() SAX events, because otherwise we would call the
+               callback function with an empty array or object, respectively.
+
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+
+    @return pair of boolean (whether value should be kept) and pointer (to the
+            passed value in the ref_stack hierarchy; nullptr if not kept)
+    */
+    template<typename Value>
+    std::pair<bool, BasicJsonType*> handle_value(Value&& v, const bool skip_callback = false)
+    {
+        JSON_ASSERT(!keep_stack.empty());
+
+        // do not handle this value if we know it would be added to a discarded
+        // container
+        if (!keep_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // create value
+        auto value = BasicJsonType(std::forward<Value>(v));
+
+        // check callback
+        const bool keep = skip_callback || callback(static_cast<int>(ref_stack.size()), parse_event_t::value, value);
+
+        // do not handle this value if we just learnt it shall be discarded
+        if (!keep)
+        {
+            return {false, nullptr};
+        }
+
+        if (ref_stack.empty())
+        {
+            root = std::move(value);
+            return {true, &root};
+        }
+
+        // skip this value if we already decided to skip the parent
+        // (https://github.com/nlohmann/json/issues/971#issuecomment-413678360)
+        if (!ref_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // we now only expect arrays and objects
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        // array
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->emplace_back(std::move(value));
+            return {true, &(ref_stack.back()->m_value.array->back())};
+        }
+
+        // object
+        JSON_ASSERT(ref_stack.back()->is_object());
+        // check if we should store an element for the current key
+        JSON_ASSERT(!key_keep_stack.empty());
+        const bool store_element = key_keep_stack.back();
+        key_keep_stack.pop_back();
+
+        if (!store_element)
+        {
+            return {false, nullptr};
+        }
+
+        JSON_ASSERT(object_element);
+        *object_element = std::move(value);
+        return {true, object_element};
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// stack to manage which values to keep
+    std::vector<bool> keep_stack {};
+    /// stack to manage which object keys to keep
+    std::vector<bool> key_keep_stack {};
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// callback function
+    const parser_callback_t callback = nullptr;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+    /// a discarded value for the callback
+    BasicJsonType discarded = BasicJsonType::value_t::discarded;
+};
+
+template<typename BasicJsonType>
+class json_sax_acceptor
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    bool null()
+    {
+        return true;
+    }
+
+    bool boolean(bool /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_integer(number_integer_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_float(number_float_t /*unused*/, const string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool string(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool binary(binary_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool start_object(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool key(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool end_object()
+    {
+        return true;
+    }
+
+    bool start_array(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool end_array()
+    {
+        return true;
+    }
+
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/, const detail::exception& /*unused*/)
+    {
+        return false;
+    }
+};
+}  // namespace detail
+
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+
+#include <array> // array
+#include <clocale> // localeconv
+#include <cstddef> // size_t
+#include <cstdio> // snprintf
+#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
+#include <initializer_list> // initializer_list
+#include <string> // char_traits, string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/position_t.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+///////////
+// lexer //
+///////////
+
+template<typename BasicJsonType>
+class lexer_base
+{
+  public:
+    /// token types for the parser
+    enum class token_type
+    {
+        uninitialized,    ///< indicating the scanner is uninitialized
+        literal_true,     ///< the `true` literal
+        literal_false,    ///< the `false` literal
+        literal_null,     ///< the `null` literal
+        value_string,     ///< a string -- use get_string() for actual value
+        value_unsigned,   ///< an unsigned integer -- use get_number_unsigned() for actual value
+        value_integer,    ///< a signed integer -- use get_number_integer() for actual value
+        value_float,      ///< an floating point number -- use get_number_float() for actual value
+        begin_array,      ///< the character for array begin `[`
+        begin_object,     ///< the character for object begin `{`
+        end_array,        ///< the character for array end `]`
+        end_object,       ///< the character for object end `}`
+        name_separator,   ///< the name separator `:`
+        value_separator,  ///< the value separator `,`
+        parse_error,      ///< indicating a parse error
+        end_of_input,     ///< indicating the end of the input buffer
+        literal_or_value  ///< a literal or the begin of a value (only for diagnostics)
+    };
+
+    /// return name of values of type token_type (only used for errors)
+    JSON_HEDLEY_RETURNS_NON_NULL
+    JSON_HEDLEY_CONST
+    static const char* token_type_name(const token_type t) noexcept
+    {
+        switch (t)
+        {
+            case token_type::uninitialized:
+                return "<uninitialized>";
+            case token_type::literal_true:
+                return "true literal";
+            case token_type::literal_false:
+                return "false literal";
+            case token_type::literal_null:
+                return "null literal";
+            case token_type::value_string:
+                return "string literal";
+            case token_type::value_unsigned:
+            case token_type::value_integer:
+            case token_type::value_float:
+                return "number literal";
+            case token_type::begin_array:
+                return "'['";
+            case token_type::begin_object:
+                return "'{'";
+            case token_type::end_array:
+                return "']'";
+            case token_type::end_object:
+                return "'}'";
+            case token_type::name_separator:
+                return "':'";
+            case token_type::value_separator:
+                return "','";
+            case token_type::parse_error:
+                return "<parse error>";
+            case token_type::end_of_input:
+                return "end of input";
+            case token_type::literal_or_value:
+                return "'[', '{', or a literal";
+            // LCOV_EXCL_START
+            default: // catch non-enum values
+                return "unknown token";
+                // LCOV_EXCL_STOP
+        }
+    }
+};
+/*!
+@brief lexical analysis
+
+This class organizes the lexical analysis during JSON deserialization.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class lexer : public lexer_base<BasicJsonType>
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename std::char_traits<char_type>::int_type;
+
+  public:
+    using token_type = typename lexer_base<BasicJsonType>::token_type;
+
+    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false) noexcept
+        : ia(std::move(adapter))
+        , ignore_comments(ignore_comments_)
+        , decimal_point_char(static_cast<char_int_type>(get_decimal_point()))
+    {}
+
+    // delete because of pointer members
+    lexer(const lexer&) = delete;
+    lexer(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    lexer& operator=(lexer&) = delete;
+    lexer& operator=(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~lexer() = default;
+
+  private:
+    /////////////////////
+    // locales
+    /////////////////////
+
+    /// return the locale-dependent decimal point
+    JSON_HEDLEY_PURE
+    static char get_decimal_point() noexcept
+    {
+        const auto* loc = localeconv();
+        JSON_ASSERT(loc != nullptr);
+        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
+    }
+
+    /////////////////////
+    // scan functions
+    /////////////////////
+
+    /*!
+    @brief get codepoint from 4 hex characters following `\u`
+
+    For input "\u c1 c2 c3 c4" the codepoint is:
+      (c1 * 0x1000) + (c2 * 0x0100) + (c3 * 0x0010) + c4
+    = (c1 << 12) + (c2 << 8) + (c3 << 4) + (c4 << 0)
+
+    Furthermore, the possible characters '0'..'9', 'A'..'F', and 'a'..'f'
+    must be converted to the integers 0x0..0x9, 0xA..0xF, 0xA..0xF, resp. The
+    conversion is done by subtracting the offset (0x30, 0x37, and 0x57)
+    between the ASCII value of the character and the desired integer value.
+
+    @return codepoint (0x0000..0xFFFF) or -1 in case of an error (e.g. EOF or
+            non-hex character)
+    */
+    int get_codepoint()
+    {
+        // this function only makes sense after reading `\u`
+        JSON_ASSERT(current == 'u');
+        int codepoint = 0;
+
+        const auto factors = { 12u, 8u, 4u, 0u };
+        for (const auto factor : factors)
+        {
+            get();
+
+            if (current >= '0' && current <= '9')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x30u) << factor);
+            }
+            else if (current >= 'A' && current <= 'F')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x37u) << factor);
+            }
+            else if (current >= 'a' && current <= 'f')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x57u) << factor);
+            }
+            else
+            {
+                return -1;
+            }
+        }
+
+        JSON_ASSERT(0x0000 <= codepoint && codepoint <= 0xFFFF);
+        return codepoint;
+    }
+
+    /*!
+    @brief check if the next byte(s) are inside a given range
+
+    Adds the current byte and, for each passed range, reads a new byte and
+    checks if it is inside the range. If a violation was detected, set up an
+    error message and return false. Otherwise, return true.
+
+    @param[in] ranges  list of integers; interpreted as list of pairs of
+                       inclusive lower and upper bound, respectively
+
+    @pre The passed list @a ranges must has 2, 4, or 6 elements; that is,
+         1, 2, or 3 pairs. This precondition is enforced by an assertion.
+
+    @return true if and only if no range violation was detected
+    */
+    bool next_byte_in_range(std::initializer_list<char_int_type> ranges)
+    {
+        JSON_ASSERT(ranges.size() == 2 || ranges.size() == 4 || ranges.size() == 6);
+        add(current);
+
+        for (auto range = ranges.begin(); range != ranges.end(); ++range)
+        {
+            get();
+            if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range)))
+            {
+                add(current);
+            }
+            else
+            {
+                error_message = "invalid string: ill-formed UTF-8 byte";
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief scan a string literal
+
+    This function scans a string according to Sect. 7 of RFC 8259. While
+    scanning, bytes are escaped and copied into buffer token_buffer. Then the
+    function returns successfully, token_buffer is *not* null-terminated (as it
+    may contain \0 bytes), and token_buffer.size() is the number of bytes in the
+    string.
+
+    @return token_type::value_string if string could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note In case of errors, variable error_message contains a textual
+          description.
+    */
+    token_type scan_string()
+    {
+        // reset token_buffer (ignore opening quote)
+        reset();
+
+        // we entered the function by reading an open quote
+        JSON_ASSERT(current == '\"');
+
+        while (true)
+        {
+            // get next character
+            switch (get())
+            {
+                // end of file while parsing string
+                case std::char_traits<char_type>::eof():
+                {
+                    error_message = "invalid string: missing closing quote";
+                    return token_type::parse_error;
+                }
+
+                // closing quote
+                case '\"':
+                {
+                    return token_type::value_string;
+                }
+
+                // escapes
+                case '\\':
+                {
+                    switch (get())
+                    {
+                        // quotation mark
+                        case '\"':
+                            add('\"');
+                            break;
+                        // reverse solidus
+                        case '\\':
+                            add('\\');
+                            break;
+                        // solidus
+                        case '/':
+                            add('/');
+                            break;
+                        // backspace
+                        case 'b':
+                            add('\b');
+                            break;
+                        // form feed
+                        case 'f':
+                            add('\f');
+                            break;
+                        // line feed
+                        case 'n':
+                            add('\n');
+                            break;
+                        // carriage return
+                        case 'r':
+                            add('\r');
+                            break;
+                        // tab
+                        case 't':
+                            add('\t');
+                            break;
+
+                        // unicode escapes
+                        case 'u':
+                        {
+                            const int codepoint1 = get_codepoint();
+                            int codepoint = codepoint1; // start with codepoint1
+
+                            if (JSON_HEDLEY_UNLIKELY(codepoint1 == -1))
+                            {
+                                error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                return token_type::parse_error;
+                            }
+
+                            // check if code point is a high surrogate
+                            if (0xD800 <= codepoint1 && codepoint1 <= 0xDBFF)
+                            {
+                                // expect next \uxxxx entry
+                                if (JSON_HEDLEY_LIKELY(get() == '\\' && get() == 'u'))
+                                {
+                                    const int codepoint2 = get_codepoint();
+
+                                    if (JSON_HEDLEY_UNLIKELY(codepoint2 == -1))
+                                    {
+                                        error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                        return token_type::parse_error;
+                                    }
+
+                                    // check if codepoint2 is a low surrogate
+                                    if (JSON_HEDLEY_LIKELY(0xDC00 <= codepoint2 && codepoint2 <= 0xDFFF))
+                                    {
+                                        // overwrite codepoint
+                                        codepoint = static_cast<int>(
+                                                        // high surrogate occupies the most significant 22 bits
+                                                        (static_cast<unsigned int>(codepoint1) << 10u)
+                                                        // low surrogate occupies the least significant 15 bits
+                                                        + static_cast<unsigned int>(codepoint2)
+                                                        // there is still the 0xD800, 0xDC00 and 0x10000 noise
+                                                        // in the result, so we have to subtract with:
+                                                        // (0xD800 << 10) + DC00 - 0x10000 = 0x35FDC00
+                                                        - 0x35FDC00u);
+                                    }
+                                    else
+                                    {
+                                        error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                        return token_type::parse_error;
+                                    }
+                                }
+                                else
+                                {
+                                    error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+                            else
+                            {
+                                if (JSON_HEDLEY_UNLIKELY(0xDC00 <= codepoint1 && codepoint1 <= 0xDFFF))
+                                {
+                                    error_message = "invalid string: surrogate U+DC00..U+DFFF must follow U+D800..U+DBFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+
+                            // result of the above calculation yields a proper codepoint
+                            JSON_ASSERT(0x00 <= codepoint && codepoint <= 0x10FFFF);
+
+                            // translate codepoint into bytes
+                            if (codepoint < 0x80)
+                            {
+                                // 1-byte characters: 0xxxxxxx (ASCII)
+                                add(static_cast<char_int_type>(codepoint));
+                            }
+                            else if (codepoint <= 0x7FF)
+                            {
+                                // 2-byte characters: 110xxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xC0u | (static_cast<unsigned int>(codepoint) >> 6u)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else if (codepoint <= 0xFFFF)
+                            {
+                                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xE0u | (static_cast<unsigned int>(codepoint) >> 12u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else
+                            {
+                                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xF0u | (static_cast<unsigned int>(codepoint) >> 18u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 12u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+
+                            break;
+                        }
+
+                        // other characters after escape
+                        default:
+                            error_message = "invalid string: forbidden character after backslash";
+                            return token_type::parse_error;
+                    }
+
+                    break;
+                }
+
+                // invalid control characters
+                case 0x00:
+                {
+                    error_message = "invalid string: control character U+0000 (NUL) must be escaped to \\u0000";
+                    return token_type::parse_error;
+                }
+
+                case 0x01:
+                {
+                    error_message = "invalid string: control character U+0001 (SOH) must be escaped to \\u0001";
+                    return token_type::parse_error;
+                }
+
+                case 0x02:
+                {
+                    error_message = "invalid string: control character U+0002 (STX) must be escaped to \\u0002";
+                    return token_type::parse_error;
+                }
+
+                case 0x03:
+                {
+                    error_message = "invalid string: control character U+0003 (ETX) must be escaped to \\u0003";
+                    return token_type::parse_error;
+                }
+
+                case 0x04:
+                {
+                    error_message = "invalid string: control character U+0004 (EOT) must be escaped to \\u0004";
+                    return token_type::parse_error;
+                }
+
+                case 0x05:
+                {
+                    error_message = "invalid string: control character U+0005 (ENQ) must be escaped to \\u0005";
+                    return token_type::parse_error;
+                }
+
+                case 0x06:
+                {
+                    error_message = "invalid string: control character U+0006 (ACK) must be escaped to \\u0006";
+                    return token_type::parse_error;
+                }
+
+                case 0x07:
+                {
+                    error_message = "invalid string: control character U+0007 (BEL) must be escaped to \\u0007";
+                    return token_type::parse_error;
+                }
+
+                case 0x08:
+                {
+                    error_message = "invalid string: control character U+0008 (BS) must be escaped to \\u0008 or \\b";
+                    return token_type::parse_error;
+                }
+
+                case 0x09:
+                {
+                    error_message = "invalid string: control character U+0009 (HT) must be escaped to \\u0009 or \\t";
+                    return token_type::parse_error;
+                }
+
+                case 0x0A:
+                {
+                    error_message = "invalid string: control character U+000A (LF) must be escaped to \\u000A or \\n";
+                    return token_type::parse_error;
+                }
+
+                case 0x0B:
+                {
+                    error_message = "invalid string: control character U+000B (VT) must be escaped to \\u000B";
+                    return token_type::parse_error;
+                }
+
+                case 0x0C:
+                {
+                    error_message = "invalid string: control character U+000C (FF) must be escaped to \\u000C or \\f";
+                    return token_type::parse_error;
+                }
+
+                case 0x0D:
+                {
+                    error_message = "invalid string: control character U+000D (CR) must be escaped to \\u000D or \\r";
+                    return token_type::parse_error;
+                }
+
+                case 0x0E:
+                {
+                    error_message = "invalid string: control character U+000E (SO) must be escaped to \\u000E";
+                    return token_type::parse_error;
+                }
+
+                case 0x0F:
+                {
+                    error_message = "invalid string: control character U+000F (SI) must be escaped to \\u000F";
+                    return token_type::parse_error;
+                }
+
+                case 0x10:
+                {
+                    error_message = "invalid string: control character U+0010 (DLE) must be escaped to \\u0010";
+                    return token_type::parse_error;
+                }
+
+                case 0x11:
+                {
+                    error_message = "invalid string: control character U+0011 (DC1) must be escaped to \\u0011";
+                    return token_type::parse_error;
+                }
+
+                case 0x12:
+                {
+                    error_message = "invalid string: control character U+0012 (DC2) must be escaped to \\u0012";
+                    return token_type::parse_error;
+                }
+
+                case 0x13:
+                {
+                    error_message = "invalid string: control character U+0013 (DC3) must be escaped to \\u0013";
+                    return token_type::parse_error;
+                }
+
+                case 0x14:
+                {
+                    error_message = "invalid string: control character U+0014 (DC4) must be escaped to \\u0014";
+                    return token_type::parse_error;
+                }
+
+                case 0x15:
+                {
+                    error_message = "invalid string: control character U+0015 (NAK) must be escaped to \\u0015";
+                    return token_type::parse_error;
+                }
+
+                case 0x16:
+                {
+                    error_message = "invalid string: control character U+0016 (SYN) must be escaped to \\u0016";
+                    return token_type::parse_error;
+                }
+
+                case 0x17:
+                {
+                    error_message = "invalid string: control character U+0017 (ETB) must be escaped to \\u0017";
+                    return token_type::parse_error;
+                }
+
+                case 0x18:
+                {
+                    error_message = "invalid string: control character U+0018 (CAN) must be escaped to \\u0018";
+                    return token_type::parse_error;
+                }
+
+                case 0x19:
+                {
+                    error_message = "invalid string: control character U+0019 (EM) must be escaped to \\u0019";
+                    return token_type::parse_error;
+                }
+
+                case 0x1A:
+                {
+                    error_message = "invalid string: control character U+001A (SUB) must be escaped to \\u001A";
+                    return token_type::parse_error;
+                }
+
+                case 0x1B:
+                {
+                    error_message = "invalid string: control character U+001B (ESC) must be escaped to \\u001B";
+                    return token_type::parse_error;
+                }
+
+                case 0x1C:
+                {
+                    error_message = "invalid string: control character U+001C (FS) must be escaped to \\u001C";
+                    return token_type::parse_error;
+                }
+
+                case 0x1D:
+                {
+                    error_message = "invalid string: control character U+001D (GS) must be escaped to \\u001D";
+                    return token_type::parse_error;
+                }
+
+                case 0x1E:
+                {
+                    error_message = "invalid string: control character U+001E (RS) must be escaped to \\u001E";
+                    return token_type::parse_error;
+                }
+
+                case 0x1F:
+                {
+                    error_message = "invalid string: control character U+001F (US) must be escaped to \\u001F";
+                    return token_type::parse_error;
+                }
+
+                // U+0020..U+007F (except U+0022 (quote) and U+005C (backspace))
+                case 0x20:
+                case 0x21:
+                case 0x23:
+                case 0x24:
+                case 0x25:
+                case 0x26:
+                case 0x27:
+                case 0x28:
+                case 0x29:
+                case 0x2A:
+                case 0x2B:
+                case 0x2C:
+                case 0x2D:
+                case 0x2E:
+                case 0x2F:
+                case 0x30:
+                case 0x31:
+                case 0x32:
+                case 0x33:
+                case 0x34:
+                case 0x35:
+                case 0x36:
+                case 0x37:
+                case 0x38:
+                case 0x39:
+                case 0x3A:
+                case 0x3B:
+                case 0x3C:
+                case 0x3D:
+                case 0x3E:
+                case 0x3F:
+                case 0x40:
+                case 0x41:
+                case 0x42:
+                case 0x43:
+                case 0x44:
+                case 0x45:
+                case 0x46:
+                case 0x47:
+                case 0x48:
+                case 0x49:
+                case 0x4A:
+                case 0x4B:
+                case 0x4C:
+                case 0x4D:
+                case 0x4E:
+                case 0x4F:
+                case 0x50:
+                case 0x51:
+                case 0x52:
+                case 0x53:
+                case 0x54:
+                case 0x55:
+                case 0x56:
+                case 0x57:
+                case 0x58:
+                case 0x59:
+                case 0x5A:
+                case 0x5B:
+                case 0x5D:
+                case 0x5E:
+                case 0x5F:
+                case 0x60:
+                case 0x61:
+                case 0x62:
+                case 0x63:
+                case 0x64:
+                case 0x65:
+                case 0x66:
+                case 0x67:
+                case 0x68:
+                case 0x69:
+                case 0x6A:
+                case 0x6B:
+                case 0x6C:
+                case 0x6D:
+                case 0x6E:
+                case 0x6F:
+                case 0x70:
+                case 0x71:
+                case 0x72:
+                case 0x73:
+                case 0x74:
+                case 0x75:
+                case 0x76:
+                case 0x77:
+                case 0x78:
+                case 0x79:
+                case 0x7A:
+                case 0x7B:
+                case 0x7C:
+                case 0x7D:
+                case 0x7E:
+                case 0x7F:
+                {
+                    add(current);
+                    break;
+                }
+
+                // U+0080..U+07FF: bytes C2..DF 80..BF
+                case 0xC2:
+                case 0xC3:
+                case 0xC4:
+                case 0xC5:
+                case 0xC6:
+                case 0xC7:
+                case 0xC8:
+                case 0xC9:
+                case 0xCA:
+                case 0xCB:
+                case 0xCC:
+                case 0xCD:
+                case 0xCE:
+                case 0xCF:
+                case 0xD0:
+                case 0xD1:
+                case 0xD2:
+                case 0xD3:
+                case 0xD4:
+                case 0xD5:
+                case 0xD6:
+                case 0xD7:
+                case 0xD8:
+                case 0xD9:
+                case 0xDA:
+                case 0xDB:
+                case 0xDC:
+                case 0xDD:
+                case 0xDE:
+                case 0xDF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!next_byte_in_range({0x80, 0xBF})))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+0800..U+0FFF: bytes E0 A0..BF 80..BF
+                case 0xE0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0xA0, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+1000..U+CFFF: bytes E1..EC 80..BF 80..BF
+                // U+E000..U+FFFF: bytes EE..EF 80..BF 80..BF
+                case 0xE1:
+                case 0xE2:
+                case 0xE3:
+                case 0xE4:
+                case 0xE5:
+                case 0xE6:
+                case 0xE7:
+                case 0xE8:
+                case 0xE9:
+                case 0xEA:
+                case 0xEB:
+                case 0xEC:
+                case 0xEE:
+                case 0xEF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+D000..U+D7FF: bytes ED 80..9F 80..BF
+                case 0xED:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x9F, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+10000..U+3FFFF F0 90..BF 80..BF 80..BF
+                case 0xF0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x90, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+40000..U+FFFFF F1..F3 80..BF 80..BF 80..BF
+                case 0xF1:
+                case 0xF2:
+                case 0xF3:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+100000..U+10FFFF F4 80..8F 80..BF 80..BF
+                case 0xF4:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x8F, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // remaining bytes (80..C1 and F5..FF) are ill-formed
+                default:
+                {
+                    error_message = "invalid string: ill-formed UTF-8 byte";
+                    return token_type::parse_error;
+                }
+            }
+        }
+    }
+
+    /*!
+     * @brief scan a comment
+     * @return whether comment could be scanned successfully
+     */
+    bool scan_comment()
+    {
+        switch (get())
+        {
+            // single-line comments skip input until a newline or EOF is read
+            case '/':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case '\n':
+                        case '\r':
+                        case std::char_traits<char_type>::eof():
+                        case '\0':
+                            return true;
+
+                        default:
+                            break;
+                    }
+                }
+            }
+
+            // multi-line comments skip input until */ is read
+            case '*':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case std::char_traits<char_type>::eof():
+                        case '\0':
+                        {
+                            error_message = "invalid comment; missing closing '*/'";
+                            return false;
+                        }
+
+                        case '*':
+                        {
+                            switch (get())
+                            {
+                                case '/':
+                                    return true;
+
+                                default:
+                                {
+                                    unget();
+                                    continue;
+                                }
+                            }
+                        }
+
+                        default:
+                            continue;
+                    }
+                }
+            }
+
+            // unexpected character after reading '/'
+            default:
+            {
+                error_message = "invalid comment; expecting '/' or '*' after '/'";
+                return false;
+            }
+        }
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(float& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtof(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtod(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(long double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtold(str, endptr);
+    }
+
+    /*!
+    @brief scan a number literal
+
+    This function scans a string according to Sect. 6 of RFC 8259.
+
+    The function is realized with a deterministic finite state machine derived
+    from the grammar described in RFC 8259. Starting in state "init", the
+    input is read and used to determined the next state. Only state "done"
+    accepts the number. State "error" is a trap state to model errors. In the
+    table below, "anything" means any character but the ones listed before.
+
+    state    | 0        | 1-9      | e E      | +       | -       | .        | anything
+    ---------|----------|----------|----------|---------|---------|----------|-----------
+    init     | zero     | any1     | [error]  | [error] | minus   | [error]  | [error]
+    minus    | zero     | any1     | [error]  | [error] | [error] | [error]  | [error]
+    zero     | done     | done     | exponent | done    | done    | decimal1 | done
+    any1     | any1     | any1     | exponent | done    | done    | decimal1 | done
+    decimal1 | decimal2 | decimal2 | [error]  | [error] | [error] | [error]  | [error]
+    decimal2 | decimal2 | decimal2 | exponent | done    | done    | done     | done
+    exponent | any2     | any2     | [error]  | sign    | sign    | [error]  | [error]
+    sign     | any2     | any2     | [error]  | [error] | [error] | [error]  | [error]
+    any2     | any2     | any2     | done     | done    | done    | done     | done
+
+    The state machine is realized with one label per state (prefixed with
+    "scan_number_") and `goto` statements between them. The state machine
+    contains cycles, but any cycle can be left when EOF is read. Therefore,
+    the function is guaranteed to terminate.
+
+    During scanning, the read bytes are stored in token_buffer. This string is
+    then converted to a signed integer, an unsigned integer, or a
+    floating-point number.
+
+    @return token_type::value_unsigned, token_type::value_integer, or
+            token_type::value_float if number could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note The scanner is independent of the current locale. Internally, the
+          locale's decimal point is used instead of `.` to work with the
+          locale-dependent converters.
+    */
+    token_type scan_number()  // lgtm [cpp/use-of-goto]
+    {
+        // reset token_buffer to store the number's bytes
+        reset();
+
+        // the type of the parsed number; initially set to unsigned; will be
+        // changed if minus sign, decimal point or exponent is read
+        token_type number_type = token_type::value_unsigned;
+
+        // state (init): we just found out we need to scan a number
+        switch (current)
+        {
+            case '-':
+            {
+                add(current);
+                goto scan_number_minus;
+            }
+
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            // all other characters are rejected outside scan_number()
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+scan_number_minus:
+        // state: we just parsed a leading minus sign
+        number_type = token_type::value_integer;
+        switch (get())
+        {
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '-'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_zero:
+        // state: we just parse a zero (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_any1:
+        // state: we just parsed a number 0-9 (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_decimal1:
+        // state: we just parsed a decimal point
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '.'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_decimal2:
+        // we just parsed at least one number after a decimal point
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_exponent:
+        // we just parsed an exponent
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '+':
+            case '-':
+            {
+                add(current);
+                goto scan_number_sign;
+            }
+
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message =
+                    "invalid number; expected '+', '-', or digit after exponent";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_sign:
+        // we just parsed an exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after exponent sign";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_any2:
+        // we just parsed a number after the exponent or exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_done:
+        // unget the character after the number (we only read it to know that
+        // we are done scanning a number)
+        unget();
+
+        char* endptr = nullptr; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        errno = 0;
+
+        // try to parse integers first and fall back to floats
+        if (number_type == token_type::value_unsigned)
+        {
+            const auto x = std::strtoull(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_unsigned = static_cast<number_unsigned_t>(x);
+                if (value_unsigned == x)
+                {
+                    return token_type::value_unsigned;
+                }
+            }
+        }
+        else if (number_type == token_type::value_integer)
+        {
+            const auto x = std::strtoll(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_integer = static_cast<number_integer_t>(x);
+                if (value_integer == x)
+                {
+                    return token_type::value_integer;
+                }
+            }
+        }
+
+        // this code is reached if we parse a floating-point number or if an
+        // integer conversion above failed
+        strtof(value_float, token_buffer.data(), &endptr);
+
+        // we checked the number format before
+        JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+        return token_type::value_float;
+    }
+
+    /*!
+    @param[in] literal_text  the literal text to expect
+    @param[in] length        the length of the passed literal text
+    @param[in] return_type   the token type to return on success
+    */
+    JSON_HEDLEY_NON_NULL(2)
+    token_type scan_literal(const char_type* literal_text, const std::size_t length,
+                            token_type return_type)
+    {
+        JSON_ASSERT(std::char_traits<char_type>::to_char_type(current) == literal_text[0]);
+        for (std::size_t i = 1; i < length; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(std::char_traits<char_type>::to_char_type(get()) != literal_text[i]))
+            {
+                error_message = "invalid literal";
+                return token_type::parse_error;
+            }
+        }
+        return return_type;
+    }
+
+    /////////////////////
+    // input management
+    /////////////////////
+
+    /// reset token_buffer; current character is beginning of token
+    void reset() noexcept
+    {
+        token_buffer.clear();
+        token_string.clear();
+        token_string.push_back(std::char_traits<char_type>::to_char_type(current));
+    }
+
+    /*
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a
+    `std::char_traits<char>::eof()` in that case.  Stores the scanned characters
+    for use in error messages.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++position.chars_read_total;
+        ++position.chars_read_current_line;
+
+        if (next_unget)
+        {
+            // just reset the next_unget variable and work with current
+            next_unget = false;
+        }
+        else
+        {
+            current = ia.get_character();
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
+        {
+            token_string.push_back(std::char_traits<char_type>::to_char_type(current));
+        }
+
+        if (current == '\n')
+        {
+            ++position.lines_read;
+            position.chars_read_current_line = 0;
+        }
+
+        return current;
+    }
+
+    /*!
+    @brief unget current character (read it again on next get)
+
+    We implement unget by setting variable next_unget to true. The input is not
+    changed - we just simulate ungetting by modifying chars_read_total,
+    chars_read_current_line, and token_string. The next call to get() will
+    behave as if the unget character is read again.
+    */
+    void unget()
+    {
+        next_unget = true;
+
+        --position.chars_read_total;
+
+        // in case we "unget" a newline, we have to also decrement the lines_read
+        if (position.chars_read_current_line == 0)
+        {
+            if (position.lines_read > 0)
+            {
+                --position.lines_read;
+            }
+        }
+        else
+        {
+            --position.chars_read_current_line;
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
+        {
+            JSON_ASSERT(!token_string.empty());
+            token_string.pop_back();
+        }
+    }
+
+    /// add a character to token_buffer
+    void add(char_int_type c)
+    {
+        token_buffer.push_back(static_cast<typename string_t::value_type>(c));
+    }
+
+  public:
+    /////////////////////
+    // value getters
+    /////////////////////
+
+    /// return integer value
+    constexpr number_integer_t get_number_integer() const noexcept
+    {
+        return value_integer;
+    }
+
+    /// return unsigned integer value
+    constexpr number_unsigned_t get_number_unsigned() const noexcept
+    {
+        return value_unsigned;
+    }
+
+    /// return floating-point value
+    constexpr number_float_t get_number_float() const noexcept
+    {
+        return value_float;
+    }
+
+    /// return current string value (implicitly resets the token; useful only once)
+    string_t& get_string()
+    {
+        return token_buffer;
+    }
+
+    /////////////////////
+    // diagnostics
+    /////////////////////
+
+    /// return position of last read token
+    constexpr position_t get_position() const noexcept
+    {
+        return position;
+    }
+
+    /// return the last read token (for errors only).  Will never contain EOF
+    /// (an arbitrary value that is not a valid char value, often -1), because
+    /// 255 may legitimately occur.  May contain NUL, which should be escaped.
+    std::string get_token_string() const
+    {
+        // escape control characters
+        std::string result;
+        for (const auto c : token_string)
+        {
+            if (static_cast<unsigned char>(c) <= '\x1F')
+            {
+                // escape control characters
+                std::array<char, 9> cs{{}};
+                static_cast<void>((std::snprintf)(cs.data(), cs.size(), "<U+%.4X>", static_cast<unsigned char>(c))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                result += cs.data();
+            }
+            else
+            {
+                // add character as is
+                result.push_back(static_cast<std::string::value_type>(c));
+            }
+        }
+
+        return result;
+    }
+
+    /// return syntax error message
+    JSON_HEDLEY_RETURNS_NON_NULL
+    constexpr const char* get_error_message() const noexcept
+    {
+        return error_message;
+    }
+
+    /////////////////////
+    // actual scanner
+    /////////////////////
+
+    /*!
+    @brief skip the UTF-8 byte order mark
+    @return true iff there is no BOM or the correct BOM has been skipped
+    */
+    bool skip_bom()
+    {
+        if (get() == 0xEF)
+        {
+            // check if we completely parse the BOM
+            return get() == 0xBB && get() == 0xBF;
+        }
+
+        // the first character is not the beginning of the BOM; unget it to
+        // process is later
+        unget();
+        return true;
+    }
+
+    void skip_whitespace()
+    {
+        do
+        {
+            get();
+        }
+        while (current == ' ' || current == '\t' || current == '\n' || current == '\r');
+    }
+
+    token_type scan()
+    {
+        // initially, skip the BOM
+        if (position.chars_read_total == 0 && !skip_bom())
+        {
+            error_message = "invalid BOM; must be 0xEF 0xBB 0xBF if given";
+            return token_type::parse_error;
+        }
+
+        // read next character and ignore whitespace
+        skip_whitespace();
+
+        // ignore comments
+        while (ignore_comments && current == '/')
+        {
+            if (!scan_comment())
+            {
+                return token_type::parse_error;
+            }
+
+            // skip following whitespace
+            skip_whitespace();
+        }
+
+        switch (current)
+        {
+            // structural characters
+            case '[':
+                return token_type::begin_array;
+            case ']':
+                return token_type::end_array;
+            case '{':
+                return token_type::begin_object;
+            case '}':
+                return token_type::end_object;
+            case ':':
+                return token_type::name_separator;
+            case ',':
+                return token_type::value_separator;
+
+            // literals
+            case 't':
+            {
+                std::array<char_type, 4> true_literal = {{static_cast<char_type>('t'), static_cast<char_type>('r'), static_cast<char_type>('u'), static_cast<char_type>('e')}};
+                return scan_literal(true_literal.data(), true_literal.size(), token_type::literal_true);
+            }
+            case 'f':
+            {
+                std::array<char_type, 5> false_literal = {{static_cast<char_type>('f'), static_cast<char_type>('a'), static_cast<char_type>('l'), static_cast<char_type>('s'), static_cast<char_type>('e')}};
+                return scan_literal(false_literal.data(), false_literal.size(), token_type::literal_false);
+            }
+            case 'n':
+            {
+                std::array<char_type, 4> null_literal = {{static_cast<char_type>('n'), static_cast<char_type>('u'), static_cast<char_type>('l'), static_cast<char_type>('l')}};
+                return scan_literal(null_literal.data(), null_literal.size(), token_type::literal_null);
+            }
+
+            // string
+            case '\"':
+                return scan_string();
+
+            // number
+            case '-':
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+                return scan_number();
+
+            // end of input (the null byte is needed when parsing from
+            // string literals)
+            case '\0':
+            case std::char_traits<char_type>::eof():
+                return token_type::end_of_input;
+
+            // error
+            default:
+                error_message = "invalid literal";
+                return token_type::parse_error;
+        }
+    }
+
+  private:
+    /// input adapter
+    InputAdapterType ia;
+
+    /// whether comments should be ignored (true) or signaled as errors (false)
+    const bool ignore_comments = false;
+
+    /// the current character
+    char_int_type current = std::char_traits<char_type>::eof();
+
+    /// whether the next get() call should just return current
+    bool next_unget = false;
+
+    /// the start position of the current token
+    position_t position {};
+
+    /// raw input token string (for error messages)
+    std::vector<char_type> token_string {};
+
+    /// buffer for variable-length tokens (numbers, strings)
+    string_t token_buffer {};
+
+    /// a description of occurred lexer errors
+    const char* error_message = "";
+
+    // number values
+    number_integer_t value_integer = 0;
+    number_unsigned_t value_unsigned = 0;
+    number_float_t value_float = 0;
+
+    /// the decimal point
+    const char_int_type decimal_point_char = '.';
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+
+
+#include <cstdint> // size_t
+#include <utility> // declval
+#include <string> // string
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename T>
+using null_function_t = decltype(std::declval<T&>().null());
+
+template<typename T>
+using boolean_function_t =
+    decltype(std::declval<T&>().boolean(std::declval<bool>()));
+
+template<typename T, typename Integer>
+using number_integer_function_t =
+    decltype(std::declval<T&>().number_integer(std::declval<Integer>()));
+
+template<typename T, typename Unsigned>
+using number_unsigned_function_t =
+    decltype(std::declval<T&>().number_unsigned(std::declval<Unsigned>()));
+
+template<typename T, typename Float, typename String>
+using number_float_function_t = decltype(std::declval<T&>().number_float(
+                                    std::declval<Float>(), std::declval<const String&>()));
+
+template<typename T, typename String>
+using string_function_t =
+    decltype(std::declval<T&>().string(std::declval<String&>()));
+
+template<typename T, typename Binary>
+using binary_function_t =
+    decltype(std::declval<T&>().binary(std::declval<Binary&>()));
+
+template<typename T>
+using start_object_function_t =
+    decltype(std::declval<T&>().start_object(std::declval<std::size_t>()));
+
+template<typename T, typename String>
+using key_function_t =
+    decltype(std::declval<T&>().key(std::declval<String&>()));
+
+template<typename T>
+using end_object_function_t = decltype(std::declval<T&>().end_object());
+
+template<typename T>
+using start_array_function_t =
+    decltype(std::declval<T&>().start_array(std::declval<std::size_t>()));
+
+template<typename T>
+using end_array_function_t = decltype(std::declval<T&>().end_array());
+
+template<typename T, typename Exception>
+using parse_error_function_t = decltype(std::declval<T&>().parse_error(
+        std::declval<std::size_t>(), std::declval<const std::string&>(),
+        std::declval<const Exception&>()));
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static constexpr bool value =
+        is_detected_exact<bool, null_function_t, SAX>::value &&
+        is_detected_exact<bool, boolean_function_t, SAX>::value &&
+        is_detected_exact<bool, number_integer_function_t, SAX, number_integer_t>::value &&
+        is_detected_exact<bool, number_unsigned_function_t, SAX, number_unsigned_t>::value &&
+        is_detected_exact<bool, number_float_function_t, SAX, number_float_t, string_t>::value &&
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value &&
+        is_detected_exact<bool, start_object_function_t, SAX>::value &&
+        is_detected_exact<bool, key_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, end_object_function_t, SAX>::value &&
+        is_detected_exact<bool, start_array_function_t, SAX>::value &&
+        is_detected_exact<bool, end_array_function_t, SAX>::value &&
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value;
+};
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax_static_asserts
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static_assert(is_detected_exact<bool, null_function_t, SAX>::value,
+                  "Missing/invalid function: bool null()");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(
+        is_detected_exact<bool, number_integer_function_t, SAX,
+        number_integer_t>::value,
+        "Missing/invalid function: bool number_integer(number_integer_t)");
+    static_assert(
+        is_detected_exact<bool, number_unsigned_function_t, SAX,
+        number_unsigned_t>::value,
+        "Missing/invalid function: bool number_unsigned(number_unsigned_t)");
+    static_assert(is_detected_exact<bool, number_float_function_t, SAX,
+                  number_float_t, string_t>::value,
+                  "Missing/invalid function: bool number_float(number_float_t, const string_t&)");
+    static_assert(
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value,
+        "Missing/invalid function: bool string(string_t&)");
+    static_assert(
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value,
+        "Missing/invalid function: bool binary(binary_t&)");
+    static_assert(is_detected_exact<bool, start_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_object(std::size_t)");
+    static_assert(is_detected_exact<bool, key_function_t, SAX, string_t>::value,
+                  "Missing/invalid function: bool key(string_t&)");
+    static_assert(is_detected_exact<bool, end_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_object()");
+    static_assert(is_detected_exact<bool, start_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_array(std::size_t)");
+    static_assert(is_detected_exact<bool, end_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_array()");
+    static_assert(
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value,
+        "Missing/invalid function: bool parse_error(std::size_t, const "
+        "std::string&, const exception&)");
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+
+/// how to treat CBOR tags
+enum class cbor_tag_handler_t
+{
+    error,   ///< throw a parse_error exception in case of a tag
+    ignore,  ///< ignore tags
+    store    ///< store tags as binary type
+};
+
+/*!
+@brief determine system byte order
+
+@return true if and only if system's byte order is little endian
+
+@note from https://stackoverflow.com/a/1001328/266378
+*/
+static inline bool little_endianness(int num = 1) noexcept
+{
+    return *reinterpret_cast<char*>(&num) == 1;
+}
+
+
+///////////////////
+// binary reader //
+///////////////////
+
+/*!
+@brief deserialization of CBOR, MessagePack, and UBJSON values
+*/
+template<typename BasicJsonType, typename InputAdapterType, typename SAX = json_sax_dom_parser<BasicJsonType>>
+class binary_reader
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using json_sax_t = SAX;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename std::char_traits<char_type>::int_type;
+
+  public:
+    /*!
+    @brief create a binary reader
+
+    @param[in] adapter  input adapter to read from
+    */
+    explicit binary_reader(InputAdapterType&& adapter) noexcept : ia(std::move(adapter))
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+    }
+
+    // make class move-only
+    binary_reader(const binary_reader&) = delete;
+    binary_reader(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    binary_reader& operator=(const binary_reader&) = delete;
+    binary_reader& operator=(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~binary_reader() = default;
+
+    /*!
+    @param[in] format  the binary format to parse
+    @param[in] sax_    a SAX event processor
+    @param[in] strict  whether to expect the input to be consumed completed
+    @param[in] tag_handler  how to treat CBOR tags
+
+    @return whether parsing was successful
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool sax_parse(const input_format_t format,
+                   json_sax_t* sax_,
+                   const bool strict = true,
+                   const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        sax = sax_;
+        bool result = false;
+
+        switch (format)
+        {
+            case input_format_t::bson:
+                result = parse_bson_internal();
+                break;
+
+            case input_format_t::cbor:
+                result = parse_cbor_internal(true, tag_handler);
+                break;
+
+            case input_format_t::msgpack:
+                result = parse_msgpack_internal();
+                break;
+
+            case input_format_t::ubjson:
+                result = parse_ubjson_internal();
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        // strict mode: next byte must be EOF
+        if (result && strict)
+        {
+            if (format == input_format_t::ubjson)
+            {
+                get_ignore_noop();
+            }
+            else
+            {
+                get();
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(current != std::char_traits<char_type>::eof()))
+            {
+                return sax->parse_error(chars_read, get_token_string(),
+                                        parse_error::create(110, chars_read, exception_message(format, "expected end of input; last byte: 0x" + get_token_string(), "value"), BasicJsonType()));
+            }
+        }
+
+        return result;
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @brief Reads in a BSON-object and passes it to the SAX-parser.
+    @return whether a valid BSON-value was passed to the SAX parser
+    */
+    bool parse_bson_internal()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/false)))
+        {
+            return false;
+        }
+
+        return sax->end_object();
+    }
+
+    /*!
+    @brief Parses a C-style string from the BSON input.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @return `true` if the \x00-byte indicating the end of the string was
+             encountered before the EOF; false` indicates an unexpected EOF.
+    */
+    bool get_bson_cstr(string_t& result)
+    {
+        auto out = std::back_inserter(result);
+        while (true)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "cstring")))
+            {
+                return false;
+            }
+            if (current == 0x00)
+            {
+                return true;
+            }
+            *out++ = static_cast<typename string_t::value_type>(current);
+        }
+    }
+
+    /*!
+    @brief Parses a zero-terminated string of length @a len from the BSON
+           input.
+    @param[in] len  The length (including the zero-byte at the end) of the
+                    string to be read.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 1
+    @return `true` if the string was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_string(const NumberType len, string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 1))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::bson, "string length must be at least 1, is " + std::to_string(len), "string"), BasicJsonType()));
+        }
+
+        return get_string(input_format_t::bson, len - static_cast<NumberType>(1), result) && get() != std::char_traits<char_type>::eof();
+    }
+
+    /*!
+    @brief Parses a byte array input of length @a len from the BSON input.
+    @param[in] len  The length of the byte array to be read.
+    @param[in,out] result  A reference to the binary variable where the read
+                            array is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 0
+    @return `true` if the byte array was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_binary(const NumberType len, binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 0))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::bson, "byte array length cannot be negative, is " + std::to_string(len), "binary"), BasicJsonType()));
+        }
+
+        // All BSON binary values have a subtype
+        std::uint8_t subtype{};
+        get_number<std::uint8_t>(input_format_t::bson, subtype);
+        result.set_subtype(subtype);
+
+        return get_binary(input_format_t::bson, len, result);
+    }
+
+    /*!
+    @brief Read a BSON document element of the given @a element_type.
+    @param[in] element_type The BSON element type, c.f. http://bsonspec.org/spec.html
+    @param[in] element_type_parse_position The position in the input stream,
+               where the `element_type` was read.
+    @warning Not all BSON element types are supported yet. An unsupported
+             @a element_type will give rise to a parse_error.114:
+             Unsupported BSON record type 0x...
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_internal(const char_int_type element_type,
+                                     const std::size_t element_type_parse_position)
+    {
+        switch (element_type)
+        {
+            case 0x01: // double
+            {
+                double number{};
+                return get_number<double, true>(input_format_t::bson, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0x02: // string
+            {
+                std::int32_t len{};
+                string_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_string(len, value) && sax->string(value);
+            }
+
+            case 0x03: // object
+            {
+                return parse_bson_internal();
+            }
+
+            case 0x04: // array
+            {
+                return parse_bson_array();
+            }
+
+            case 0x05: // binary
+            {
+                std::int32_t len{};
+                binary_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_binary(len, value) && sax->binary(value);
+            }
+
+            case 0x08: // boolean
+            {
+                return sax->boolean(get() != 0);
+            }
+
+            case 0x0A: // null
+            {
+                return sax->null();
+            }
+
+            case 0x10: // int32
+            {
+                std::int32_t value{};
+                return get_number<std::int32_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            case 0x12: // int64
+            {
+                std::int64_t value{};
+                return get_number<std::int64_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            default: // anything else not supported (yet)
+            {
+                std::array<char, 3> cr{{}};
+                static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(element_type))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                return sax->parse_error(element_type_parse_position, std::string(cr.data()), parse_error::create(114, element_type_parse_position, "Unsupported BSON record type 0x" + std::string(cr.data()), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief Read a BSON element list (as specified in the BSON-spec)
+
+    The same binary layout is used for objects and arrays, hence it must be
+    indicated with the argument @a is_array which one is expected
+    (true --> array, false --> object).
+
+    @param[in] is_array Determines if the element list being read is to be
+                        treated as an object (@a is_array == false), or as an
+                        array (@a is_array == true).
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_list(const bool is_array)
+    {
+        string_t key;
+
+        while (auto element_type = get())
+        {
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "element list")))
+            {
+                return false;
+            }
+
+            const std::size_t element_type_parse_position = chars_read;
+            if (JSON_HEDLEY_UNLIKELY(!get_bson_cstr(key)))
+            {
+                return false;
+            }
+
+            if (!is_array && !sax->key(key))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_internal(element_type, element_type_parse_position)))
+            {
+                return false;
+            }
+
+            // get_bson_cstr only appends
+            key.clear();
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief Reads an array from the BSON input and passes it to the SAX-parser.
+    @return whether a valid BSON-array was passed to the SAX parser
+    */
+    bool parse_bson_array()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/true)))
+        {
+            return false;
+        }
+
+        return sax->end_array();
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true) or whether the last read character should
+                         be considered instead (false)
+    @param[in] tag_handler how CBOR tags should be treated
+
+    @return whether a valid CBOR value was passed to the SAX parser
+    */
+    bool parse_cbor_internal(const bool get_char,
+                             const cbor_tag_handler_t tag_handler)
+    {
+        switch (get_char ? get() : current)
+        {
+            // EOF
+            case std::char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::cbor, "value");
+
+            // Integer 0x00..0x17 (0..23)
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            case 0x18: // Unsigned integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x19: // Unsigned integer (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1A: // Unsigned integer (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1B: // Unsigned integer (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            // Negative integer -1-0x00..-1-0x17 (-1..-24)
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+                return sax->number_integer(static_cast<std::int8_t>(0x20 - 1 - current));
+
+            case 0x38: // Negative integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x39: // Negative integer -1-n (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3A: // Negative integer -1-n (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3B: // Negative integer -1-n (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1)
+                        - static_cast<number_integer_t>(number));
+            }
+
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            case 0x5F: // Binary data (indefinite length)
+            {
+                binary_t b;
+                return get_cbor_binary(b) && sax->binary(b);
+            }
+
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                string_t s;
+                return get_cbor_string(s) && sax->string(s);
+            }
+
+            // array (0x00..0x17 data items follow)
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+                return get_cbor_array(static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0x98: // array (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x99: // array (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9A: // array (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9B: // array (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(detail::conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9F: // array (indefinite length)
+                return get_cbor_array(static_cast<std::size_t>(-1), tag_handler);
+
+            // map (0x00..0x17 pairs of data items follow)
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+                return get_cbor_object(static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0xB8: // map (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xB9: // map (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBA: // map (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBB: // map (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(detail::conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBF: // map (indefinite length)
+                return get_cbor_object(static_cast<std::size_t>(-1), tag_handler);
+
+            case 0xC6: // tagged item
+            case 0xC7:
+            case 0xC8:
+            case 0xC9:
+            case 0xCA:
+            case 0xCB:
+            case 0xCC:
+            case 0xCD:
+            case 0xCE:
+            case 0xCF:
+            case 0xD0:
+            case 0xD1:
+            case 0xD2:
+            case 0xD3:
+            case 0xD4:
+            case 0xD8: // tagged item (1 bytes follow)
+            case 0xD9: // tagged item (2 bytes follow)
+            case 0xDA: // tagged item (4 bytes follow)
+            case 0xDB: // tagged item (8 bytes follow)
+            {
+                switch (tag_handler)
+                {
+                    case cbor_tag_handler_t::error:
+                    {
+                        auto last_token = get_token_string();
+                        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::cbor, "invalid byte: 0x" + last_token, "value"), BasicJsonType()));
+                    }
+
+                    case cbor_tag_handler_t::ignore:
+                    {
+                        // ignore binary subtype
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            default:
+                                break;
+                        }
+                        return parse_cbor_internal(true, tag_handler);
+                    }
+
+                    case cbor_tag_handler_t::store:
+                    {
+                        binary_t b;
+                        // use binary subtype and store in binary container
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            default:
+                                return parse_cbor_internal(true, tag_handler);
+                        }
+                        get();
+                        return get_cbor_binary(b) && sax->binary(b);
+                    }
+
+                    default:                 // LCOV_EXCL_LINE
+                        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                        return false;        // LCOV_EXCL_LINE
+                }
+            }
+
+            case 0xF4: // false
+                return sax->boolean(false);
+
+            case 0xF5: // true
+                return sax->boolean(true);
+
+            case 0xF6: // null
+                return sax->null();
+
+            case 0xF9: // Half-Precision Float (two-byte IEEE 754)
+            {
+                const auto byte1_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+                const auto byte2_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+
+                const auto byte1 = static_cast<unsigned char>(byte1_raw);
+                const auto byte2 = static_cast<unsigned char>(byte2_raw);
+
+                // code from RFC 7049, Appendix D, Figure 3:
+                // As half-precision floating-point numbers were only added
+                // to IEEE 754 in 2008, today's programming platforms often
+                // still only have limited support for them. It is very
+                // easy to include at least decoding support for them even
+                // without such support. An example of a small decoder for
+                // half-precision floating-point numbers in the C language
+                // is shown in Fig. 3.
+                const auto half = static_cast<unsigned int>((byte1 << 8u) + byte2);
+                const double val = [&half]
+                {
+                    const int exp = (half >> 10u) & 0x1Fu;
+                    const unsigned int mant = half & 0x3FFu;
+                    JSON_ASSERT(0 <= exp&& exp <= 32);
+                    JSON_ASSERT(mant <= 1024);
+                    switch (exp)
+                    {
+                        case 0:
+                            return std::ldexp(mant, -24);
+                        case 31:
+                            return (mant == 0)
+                            ? std::numeric_limits<double>::infinity()
+                            : std::numeric_limits<double>::quiet_NaN();
+                        default:
+                            return std::ldexp(mant + 1024, exp - 25);
+                    }
+                }();
+                return sax->number_float((half & 0x8000u) != 0
+                                         ? static_cast<number_float_t>(-val)
+                                         : static_cast<number_float_t>(val), "");
+            }
+
+            case 0xFA: // Single-Precision Float (four-byte IEEE 754)
+            {
+                float number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xFB: // Double-Precision Float (eight-byte IEEE 754)
+            {
+                double number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            default: // anything else (0xFF is handled inside the other types)
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::cbor, "invalid byte: 0x" + last_token, "value"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+    Additionally, CBOR's strings with indefinite lengths are supported.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_cbor_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            {
+                return get_string(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    string_t chunk;
+                    if (!get_cbor_string(chunk))
+                    {
+                        return false;
+                    }
+                    result.append(chunk);
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::cbor, "expected length specification (0x60-0x7B) or indefinite string type (0x7F); last byte: 0x" + last_token, "string"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into the byte array.
+    Additionally, CBOR's byte arrays with indefinite lengths are supported.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_cbor_binary(binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "binary")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            {
+                return get_binary(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5F: // Binary data (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    binary_t chunk;
+                    if (!get_cbor_binary(chunk))
+                    {
+                        return false;
+                    }
+                    result.insert(result.end(), chunk.begin(), chunk.end());
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::cbor, "expected length specification (0x40-0x5B) or indefinite binary array type (0x5F); last byte: 0x" + last_token, "binary"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array or static_cast<std::size_t>(-1) for an
+                    array of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether array creation completed
+    */
+    bool get_cbor_array(const std::size_t len,
+                        const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        if (len != static_cast<std::size_t>(-1))
+        {
+            for (std::size_t i = 0; i < len; ++i)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+        else
+        {
+            while (get() != 0xFF)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(false, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object or static_cast<std::size_t>(-1) for an
+                    object of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether object creation completed
+    */
+    bool get_cbor_object(const std::size_t len,
+                         const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        if (len != 0)
+        {
+            string_t key;
+            if (len != static_cast<std::size_t>(-1))
+            {
+                for (std::size_t i = 0; i < len; ++i)
+                {
+                    get();
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                while (get() != 0xFF)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    /*!
+    @return whether a valid MessagePack value was passed to the SAX parser
+    */
+    bool parse_msgpack_internal()
+    {
+        switch (get())
+        {
+            // EOF
+            case std::char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::msgpack, "value");
+
+            // positive fixint
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+            case 0x18:
+            case 0x19:
+            case 0x1A:
+            case 0x1B:
+            case 0x1C:
+            case 0x1D:
+            case 0x1E:
+            case 0x1F:
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+            case 0x38:
+            case 0x39:
+            case 0x3A:
+            case 0x3B:
+            case 0x3C:
+            case 0x3D:
+            case 0x3E:
+            case 0x3F:
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58:
+            case 0x59:
+            case 0x5A:
+            case 0x5B:
+            case 0x5C:
+            case 0x5D:
+            case 0x5E:
+            case 0x5F:
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78:
+            case 0x79:
+            case 0x7A:
+            case 0x7B:
+            case 0x7C:
+            case 0x7D:
+            case 0x7E:
+            case 0x7F:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            // fixmap
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+                return get_msgpack_object(static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixarray
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+            case 0x98:
+            case 0x99:
+            case 0x9A:
+            case 0x9B:
+            case 0x9C:
+            case 0x9D:
+            case 0x9E:
+            case 0x9F:
+                return get_msgpack_array(static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            case 0xD9: // str 8
+            case 0xDA: // str 16
+            case 0xDB: // str 32
+            {
+                string_t s;
+                return get_msgpack_string(s) && sax->string(s);
+            }
+
+            case 0xC0: // nil
+                return sax->null();
+
+            case 0xC2: // false
+                return sax->boolean(false);
+
+            case 0xC3: // true
+                return sax->boolean(true);
+
+            case 0xC4: // bin 8
+            case 0xC5: // bin 16
+            case 0xC6: // bin 32
+            case 0xC7: // ext 8
+            case 0xC8: // ext 16
+            case 0xC9: // ext 32
+            case 0xD4: // fixext 1
+            case 0xD5: // fixext 2
+            case 0xD6: // fixext 4
+            case 0xD7: // fixext 8
+            case 0xD8: // fixext 16
+            {
+                binary_t b;
+                return get_msgpack_binary(b) && sax->binary(b);
+            }
+
+            case 0xCA: // float 32
+            {
+                float number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCB: // float 64
+            {
+                double number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCC: // uint 8
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCD: // uint 16
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCE: // uint 32
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCF: // uint 64
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xD0: // int 8
+            {
+                std::int8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD1: // int 16
+            {
+                std::int16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD2: // int 32
+            {
+                std::int32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD3: // int 64
+            {
+                std::int64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xDC: // array 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(static_cast<std::size_t>(len));
+            }
+
+            case 0xDD: // array 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(static_cast<std::size_t>(len));
+            }
+
+            case 0xDE: // map 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(static_cast<std::size_t>(len));
+            }
+
+            case 0xDF: // map 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(static_cast<std::size_t>(len));
+            }
+
+            // negative fixint
+            case 0xE0:
+            case 0xE1:
+            case 0xE2:
+            case 0xE3:
+            case 0xE4:
+            case 0xE5:
+            case 0xE6:
+            case 0xE7:
+            case 0xE8:
+            case 0xE9:
+            case 0xEA:
+            case 0xEB:
+            case 0xEC:
+            case 0xED:
+            case 0xEE:
+            case 0xEF:
+            case 0xF0:
+            case 0xF1:
+            case 0xF2:
+            case 0xF3:
+            case 0xF4:
+            case 0xF5:
+            case 0xF6:
+            case 0xF7:
+            case 0xF8:
+            case 0xF9:
+            case 0xFA:
+            case 0xFB:
+            case 0xFC:
+            case 0xFD:
+            case 0xFE:
+            case 0xFF:
+                return sax->number_integer(static_cast<std::int8_t>(current));
+
+            default: // anything else
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::msgpack, "invalid byte: 0x" + last_token, "value"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_msgpack_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::msgpack, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            {
+                return get_string(input_format_t::msgpack, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0xD9: // str 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDA: // str 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDB: // str 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::msgpack, "expected length specification (0xA0-0xBF, 0xD9-0xDB); last byte: 0x" + last_token, "string"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into a byte array.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_msgpack_binary(binary_t& result)
+    {
+        // helper function to set the subtype
+        auto assign_and_return_true = [&result](std::int8_t subtype)
+        {
+            result.set_subtype(static_cast<std::uint8_t>(subtype));
+            return true;
+        };
+
+        switch (current)
+        {
+            case 0xC4: // bin 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC5: // bin 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC6: // bin 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC7: // ext 8
+            {
+                std::uint8_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC8: // ext 16
+            {
+                std::uint16_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC9: // ext 32
+            {
+                std::uint32_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD4: // fixext 1
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 1, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD5: // fixext 2
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 2, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD6: // fixext 4
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 4, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD7: // fixext 8
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 8, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD8: // fixext 16
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 16, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            default:           // LCOV_EXCL_LINE
+                return false;  // LCOV_EXCL_LINE
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array
+    @return whether array creation completed
+    */
+    bool get_msgpack_array(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object
+    @return whether object creation completed
+    */
+    bool get_msgpack_object(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        string_t key;
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!get_msgpack_string(key) || !sax->key(key)))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+            key.clear();
+        }
+
+        return sax->end_object();
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether a valid UBJSON value was passed to the SAX parser
+    */
+    bool parse_ubjson_internal(const bool get_char = true)
+    {
+        return get_ubjson_value(get_char ? get_ignore_noop() : current);
+    }
+
+    /*!
+    @brief reads a UBJSON string
+
+    This function is either called after reading the 'S' byte explicitly
+    indicating a string, or in case of an object key where the 'S' byte can be
+    left out.
+
+    @param[out] result   created string
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether string creation completed
+    */
+    bool get_ubjson_string(string_t& result, const bool get_char = true)
+    {
+        if (get_char)
+        {
+            get();  // TODO(niels): may we ignore N here?
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::ubjson, "value")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            case 'U':
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::ubjson, len) && get_string(input_format_t::ubjson, len, result);
+            }
+
+            case 'i':
+            {
+                std::int8_t len{};
+                return get_number(input_format_t::ubjson, len) && get_string(input_format_t::ubjson, len, result);
+            }
+
+            case 'I':
+            {
+                std::int16_t len{};
+                return get_number(input_format_t::ubjson, len) && get_string(input_format_t::ubjson, len, result);
+            }
+
+            case 'l':
+            {
+                std::int32_t len{};
+                return get_number(input_format_t::ubjson, len) && get_string(input_format_t::ubjson, len, result);
+            }
+
+            case 'L':
+            {
+                std::int64_t len{};
+                return get_number(input_format_t::ubjson, len) && get_string(input_format_t::ubjson, len, result);
+            }
+
+            default:
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::ubjson, "expected length type specification (U, i, I, l, L); last byte: 0x" + last_token, "string"), BasicJsonType()));
+        }
+    }
+
+    /*!
+    @param[out] result  determined size
+    @return whether size determination completed
+    */
+    bool get_ubjson_size_value(std::size_t& result)
+    {
+        switch (get_ignore_noop())
+        {
+            case 'U':
+            {
+                std::uint8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format_t::ubjson, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format_t::ubjson, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number); // NOLINT(bugprone-signed-char-misuse,cert-str34-c): number is not a char
+                return true;
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format_t::ubjson, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format_t::ubjson, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format_t::ubjson, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::ubjson, "expected length type specification (U, i, I, l, L) after '#'; last byte: 0x" + last_token, "size"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @brief determine the type and size for a container
+
+    In the optimized UBJSON format, a type and a size can be provided to allow
+    for a more compact representation.
+
+    @param[out] result  pair of the size and the type
+
+    @return whether pair creation completed
+    */
+    bool get_ubjson_size_type(std::pair<std::size_t, char_int_type>& result)
+    {
+        result.first = string_t::npos; // size
+        result.second = 0; // type
+
+        get_ignore_noop();
+
+        if (current == '$')
+        {
+            result.second = get();  // must not ignore 'N', because 'N' maybe the type
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::ubjson, "type")))
+            {
+                return false;
+            }
+
+            get_ignore_noop();
+            if (JSON_HEDLEY_UNLIKELY(current != '#'))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::ubjson, "value")))
+                {
+                    return false;
+                }
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::ubjson, "expected '#' after type information; last byte: 0x" + last_token, "size"), BasicJsonType()));
+            }
+
+            return get_ubjson_size_value(result.first);
+        }
+
+        if (current == '#')
+        {
+            return get_ubjson_size_value(result.first);
+        }
+
+        return true;
+    }
+
+    /*!
+    @param prefix  the previously read or set type prefix
+    @return whether value creation completed
+    */
+    bool get_ubjson_value(const char_int_type prefix)
+    {
+        switch (prefix)
+        {
+            case std::char_traits<char_type>::eof():  // EOF
+                return unexpect_eof(input_format_t::ubjson, "value");
+
+            case 'T':  // true
+                return sax->boolean(true);
+            case 'F':  // false
+                return sax->boolean(false);
+
+            case 'Z':  // null
+                return sax->null();
+
+            case 'U':
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_unsigned(number);
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_integer(number);
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_integer(number);
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_integer(number);
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_integer(number);
+            }
+
+            case 'd':
+            {
+                float number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'D':
+            {
+                double number{};
+                return get_number(input_format_t::ubjson, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'H':
+            {
+                return get_ubjson_high_precision_number();
+            }
+
+            case 'C':  // char
+            {
+                get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::ubjson, "char")))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(current > 127))
+                {
+                    auto last_token = get_token_string();
+                    return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format_t::ubjson, "byte after 'C' must be in range 0x00..0x7F; last byte: 0x" + last_token, "char"), BasicJsonType()));
+                }
+                string_t s(1, static_cast<typename string_t::value_type>(current));
+                return sax->string(s);
+            }
+
+            case 'S':  // string
+            {
+                string_t s;
+                return get_ubjson_string(s) && sax->string(s);
+            }
+
+            case '[':  // array
+                return get_ubjson_array();
+
+            case '{':  // object
+                return get_ubjson_object();
+
+            default: // anything else
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format_t::ubjson, "invalid byte: 0x" + last_token, "value"), BasicJsonType()));
+            }
+        }
+    }
+
+    /*!
+    @return whether array creation completed
+    */
+    bool get_ubjson_array()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        if (size_and_type.first != string_t::npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                if (size_and_type.second != 'N')
+                {
+                    for (std::size_t i = 0; i < size_and_type.first; ++i)
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                        {
+                            return false;
+                        }
+                    }
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != ']')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal(false)))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @return whether object creation completed
+    */
+    bool get_ubjson_object()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        string_t key;
+        if (size_and_type.first != string_t::npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != '}')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key, false) || !sax->key(key)))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+                key.clear();
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    // Note, no reader for UBJSON binary types is implemented because they do
+    // not exist
+
+    bool get_ubjson_high_precision_number()
+    {
+        // get size of following number string
+        std::size_t size{};
+        auto res = get_ubjson_size_value(size);
+        if (JSON_HEDLEY_UNLIKELY(!res))
+        {
+            return res;
+        }
+
+        // get number string
+        std::vector<char> number_vector;
+        for (std::size_t i = 0; i < size; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::ubjson, "number")))
+            {
+                return false;
+            }
+            number_vector.push_back(static_cast<char>(current));
+        }
+
+        // parse number string
+        using ia_type = decltype(detail::input_adapter(number_vector));
+        auto number_lexer = detail::lexer<BasicJsonType, ia_type>(detail::input_adapter(number_vector), false);
+        const auto result_number = number_lexer.scan();
+        const auto number_string = number_lexer.get_token_string();
+        const auto result_remainder = number_lexer.scan();
+
+        using token_type = typename detail::lexer_base<BasicJsonType>::token_type;
+
+        if (JSON_HEDLEY_UNLIKELY(result_remainder != token_type::end_of_input))
+        {
+            return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read, exception_message(input_format_t::ubjson, "invalid number text: " + number_lexer.get_token_string(), "high-precision number"), BasicJsonType()));
+        }
+
+        switch (result_number)
+        {
+            case token_type::value_integer:
+                return sax->number_integer(number_lexer.get_number_integer());
+            case token_type::value_unsigned:
+                return sax->number_unsigned(number_lexer.get_number_unsigned());
+            case token_type::value_float:
+                return sax->number_float(number_lexer.get_number_float(), std::move(number_string));
+            case token_type::uninitialized:
+            case token_type::literal_true:
+            case token_type::literal_false:
+            case token_type::literal_null:
+            case token_type::value_string:
+            case token_type::begin_array:
+            case token_type::begin_object:
+            case token_type::end_array:
+            case token_type::end_object:
+            case token_type::name_separator:
+            case token_type::value_separator:
+            case token_type::parse_error:
+            case token_type::end_of_input:
+            case token_type::literal_or_value:
+            default:
+                return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read, exception_message(input_format_t::ubjson, "invalid number text: " + number_lexer.get_token_string(), "high-precision number"), BasicJsonType()));
+        }
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*!
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a -'ve valued
+    `std::char_traits<char_type>::eof()` in that case.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++chars_read;
+        return current = ia.get_character();
+    }
+
+    /*!
+    @return character read from the input after ignoring all 'N' entries
+    */
+    char_int_type get_ignore_noop()
+    {
+        do
+        {
+            get();
+        }
+        while (current == 'N');
+
+        return current;
+    }
+
+    /*
+    @brief read a number from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format   the current format (for diagnostics)
+    @param[out] result  number of type @a NumberType
+
+    @return whether conversion completed
+
+    @note This function needs to respect the system's endianness, because
+          bytes in CBOR, MessagePack, and UBJSON are stored in network order
+          (big endian) and therefore need reordering on little endian systems.
+    */
+    template<typename NumberType, bool InputIsLittleEndian = false>
+    bool get_number(const input_format_t format, NumberType& result)
+    {
+        // step 1: read input into array with system's byte order
+        std::array<std::uint8_t, sizeof(NumberType)> vec{};
+        for (std::size_t i = 0; i < sizeof(NumberType); ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "number")))
+            {
+                return false;
+            }
+
+            // reverse byte order prior to conversion if necessary
+            if (is_little_endian != InputIsLittleEndian)
+            {
+                vec[sizeof(NumberType) - i - 1] = static_cast<std::uint8_t>(current);
+            }
+            else
+            {
+                vec[i] = static_cast<std::uint8_t>(current); // LCOV_EXCL_LINE
+            }
+        }
+
+        // step 2: convert array into number of type T and return
+        std::memcpy(&result, vec.data(), sizeof(NumberType));
+        return true;
+    }
+
+    /*!
+    @brief create a string by reading characters from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of characters to read
+    @param[out] result string created by reading @a len bytes
+
+    @return whether string creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of string memory.
+    */
+    template<typename NumberType>
+    bool get_string(const input_format_t format,
+                    const NumberType len,
+                    string_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "string")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<typename string_t::value_type>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @brief create a byte array by reading bytes from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of bytes to read
+    @param[out] result byte array created by reading @a len bytes
+
+    @return whether byte array creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of memory.
+    */
+    template<typename NumberType>
+    bool get_binary(const input_format_t format,
+                    const NumberType len,
+                    binary_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "binary")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<std::uint8_t>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @param[in] format   the current format (for diagnostics)
+    @param[in] context  further context information (for diagnostics)
+    @return whether the last read character is not EOF
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool unexpect_eof(const input_format_t format, const char* context) const
+    {
+        if (JSON_HEDLEY_UNLIKELY(current == std::char_traits<char_type>::eof()))
+        {
+            return sax->parse_error(chars_read, "<end of file>",
+                                    parse_error::create(110, chars_read, exception_message(format, "unexpected end of input", context), BasicJsonType()));
+        }
+        return true;
+    }
+
+    /*!
+    @return a string representation of the last read byte
+    */
+    std::string get_token_string() const
+    {
+        std::array<char, 3> cr{{}};
+        static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(current))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        return std::string{cr.data()};
+    }
+
+    /*!
+    @param[in] format   the current format
+    @param[in] detail   a detailed error message
+    @param[in] context  further context information
+    @return a message string to use in the parse_error exceptions
+    */
+    std::string exception_message(const input_format_t format,
+                                  const std::string& detail,
+                                  const std::string& context) const
+    {
+        std::string error_msg = "syntax error while parsing ";
+
+        switch (format)
+        {
+            case input_format_t::cbor:
+                error_msg += "CBOR";
+                break;
+
+            case input_format_t::msgpack:
+                error_msg += "MessagePack";
+                break;
+
+            case input_format_t::ubjson:
+                error_msg += "UBJSON";
+                break;
+
+            case input_format_t::bson:
+                error_msg += "BSON";
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        return error_msg + " " + context + ": " + detail;
+    }
+
+  private:
+    /// input adapter
+    InputAdapterType ia;
+
+    /// the current character
+    char_int_type current = std::char_traits<char_type>::eof();
+
+    /// the number of characters read
+    std::size_t chars_read = 0;
+
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// the SAX parser
+    json_sax_t* sax = nullptr;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/input/parser.hpp>
+
+
+#include <cmath> // isfinite
+#include <cstdint> // uint8_t
+#include <functional> // function
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+////////////
+// parser //
+////////////
+
+enum class parse_event_t : std::uint8_t
+{
+    /// the parser read `{` and started to process a JSON object
+    object_start,
+    /// the parser read `}` and finished processing a JSON object
+    object_end,
+    /// the parser read `[` and started to process a JSON array
+    array_start,
+    /// the parser read `]` and finished processing a JSON array
+    array_end,
+    /// the parser read a key of a value in an object
+    key,
+    /// the parser finished reading a JSON value
+    value
+};
+
+template<typename BasicJsonType>
+using parser_callback_t =
+    std::function<bool(int /*depth*/, parse_event_t /*event*/, BasicJsonType& /*parsed*/)>;
+
+/*!
+@brief syntax analysis
+
+This class implements a recursive descent parser.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class parser
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using lexer_t = lexer<BasicJsonType, InputAdapterType>;
+    using token_type = typename lexer_t::token_type;
+
+  public:
+    /// a parser reading from an input adapter
+    explicit parser(InputAdapterType&& adapter,
+                    const parser_callback_t<BasicJsonType> cb = nullptr,
+                    const bool allow_exceptions_ = true,
+                    const bool skip_comments = false)
+        : callback(cb)
+        , m_lexer(std::move(adapter), skip_comments)
+        , allow_exceptions(allow_exceptions_)
+    {
+        // read first token
+        get_token();
+    }
+
+    /*!
+    @brief public parser interface
+
+    @param[in] strict      whether to expect the last token to be EOF
+    @param[in,out] result  parsed JSON value
+
+    @throw parse_error.101 in case of an unexpected token
+    @throw parse_error.102 if to_unicode fails or surrogate error
+    @throw parse_error.103 if to_unicode fails
+    */
+    void parse(const bool strict, BasicJsonType& result)
+    {
+        if (callback)
+        {
+            json_sax_dom_callback_parser<BasicJsonType> sdp(result, callback, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(),
+                                                    exception_message(token_type::end_of_input, "value"), BasicJsonType()));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+
+            // set top-level value to null if it was discarded by the callback
+            // function
+            if (result.is_discarded())
+            {
+                result = nullptr;
+            }
+        }
+        else
+        {
+            json_sax_dom_parser<BasicJsonType> sdp(result, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), BasicJsonType()));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+        }
+
+        result.assert_invariant();
+    }
+
+    /*!
+    @brief public accept interface
+
+    @param[in] strict  whether to expect the last token to be EOF
+    @return whether the input is a proper JSON text
+    */
+    bool accept(const bool strict = true)
+    {
+        json_sax_acceptor<BasicJsonType> sax_acceptor;
+        return sax_parse(&sax_acceptor, strict);
+    }
+
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse(SAX* sax, const bool strict = true)
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+        const bool result = sax_parse_internal(sax);
+
+        // strict mode: next byte must be EOF
+        if (result && strict && (get_token() != token_type::end_of_input))
+        {
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), BasicJsonType()));
+        }
+
+        return result;
+    }
+
+  private:
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse_internal(SAX* sax)
+    {
+        // stack to remember the hierarchy of structured values we are parsing
+        // true = array; false = object
+        std::vector<bool> states;
+        // value to avoid a goto (see comment where set to true)
+        bool skip_to_state_evaluation = false;
+
+        while (true)
+        {
+            if (!skip_to_state_evaluation)
+            {
+                // invariant: get_token() was called before each iteration
+                switch (last_token)
+                {
+                    case token_type::begin_object:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing } -> we are done
+                        if (get_token() == token_type::end_object)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // parse key
+                        if (JSON_HEDLEY_UNLIKELY(last_token != token_type::value_string))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), BasicJsonType()));
+                        }
+                        if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        // parse separator (:)
+                        if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), BasicJsonType()));
+                        }
+
+                        // remember we are now inside an object
+                        states.push_back(false);
+
+                        // parse values
+                        get_token();
+                        continue;
+                    }
+
+                    case token_type::begin_array:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing ] -> we are done
+                        if (get_token() == token_type::end_array)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // remember we are now inside an array
+                        states.push_back(true);
+
+                        // parse values (no need to call get_token)
+                        continue;
+                    }
+
+                    case token_type::value_float:
+                    {
+                        const auto res = m_lexer.get_number_float();
+
+                        if (JSON_HEDLEY_UNLIKELY(!std::isfinite(res)))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    out_of_range::create(406, "number overflow parsing '" + m_lexer.get_token_string() + "'", BasicJsonType()));
+                        }
+
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_float(res, m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        break;
+                    }
+
+                    case token_type::literal_false:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(false)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_null:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->null()))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_true:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(true)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_integer:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_integer(m_lexer.get_number_integer())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_string:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->string(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_unsigned:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(m_lexer.get_number_unsigned())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::parse_error:
+                    {
+                        // using "uninitialized" to avoid "expected" message
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::uninitialized, "value"), BasicJsonType()));
+                    }
+
+                    case token_type::uninitialized:
+                    case token_type::end_array:
+                    case token_type::end_object:
+                    case token_type::name_separator:
+                    case token_type::value_separator:
+                    case token_type::end_of_input:
+                    case token_type::literal_or_value:
+                    default: // the last token was unexpected
+                    {
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), BasicJsonType()));
+                    }
+                }
+            }
+            else
+            {
+                skip_to_state_evaluation = false;
+            }
+
+            // we reached this line after we successfully parsed a value
+            if (states.empty())
+            {
+                // empty stack: we reached the end of the hierarchy: done
+                return true;
+            }
+
+            if (states.back())  // array
+            {
+                // comma -> next value
+                if (get_token() == token_type::value_separator)
+                {
+                    // parse a new value
+                    get_token();
+                    continue;
+                }
+
+                // closing ]
+                if (JSON_HEDLEY_LIKELY(last_token == token_type::end_array))
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                    {
+                        return false;
+                    }
+
+                    // We are done with this array. Before we can parse a
+                    // new value, we need to evaluate the new state first.
+                    // By setting skip_to_state_evaluation to false, we
+                    // are effectively jumping to the beginning of this if.
+                    JSON_ASSERT(!states.empty());
+                    states.pop_back();
+                    skip_to_state_evaluation = true;
+                    continue;
+                }
+
+                return sax->parse_error(m_lexer.get_position(),
+                                        m_lexer.get_token_string(),
+                                        parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_array, "array"), BasicJsonType()));
+            }
+
+            // states.back() is false -> object
+
+            // comma -> next value
+            if (get_token() == token_type::value_separator)
+            {
+                // parse key
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::value_string))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), BasicJsonType()));
+                }
+
+                if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                {
+                    return false;
+                }
+
+                // parse separator (:)
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), BasicJsonType()));
+                }
+
+                // parse values
+                get_token();
+                continue;
+            }
+
+            // closing }
+            if (JSON_HEDLEY_LIKELY(last_token == token_type::end_object))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                {
+                    return false;
+                }
+
+                // We are done with this object. Before we can parse a
+                // new value, we need to evaluate the new state first.
+                // By setting skip_to_state_evaluation to false, we
+                // are effectively jumping to the beginning of this if.
+                JSON_ASSERT(!states.empty());
+                states.pop_back();
+                skip_to_state_evaluation = true;
+                continue;
+            }
+
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_object, "object"), BasicJsonType()));
+        }
+    }
+
+    /// get next token from lexer
+    token_type get_token()
+    {
+        return last_token = m_lexer.scan();
+    }
+
+    std::string exception_message(const token_type expected, const std::string& context)
+    {
+        std::string error_msg = "syntax error ";
+
+        if (!context.empty())
+        {
+            error_msg += "while parsing " + context + " ";
+        }
+
+        error_msg += "- ";
+
+        if (last_token == token_type::parse_error)
+        {
+            error_msg += std::string(m_lexer.get_error_message()) + "; last read: '" +
+                         m_lexer.get_token_string() + "'";
+        }
+        else
+        {
+            error_msg += "unexpected " + std::string(lexer_t::token_type_name(last_token));
+        }
+
+        if (expected != token_type::uninitialized)
+        {
+            error_msg += "; expected " + std::string(lexer_t::token_type_name(expected));
+        }
+
+        return error_msg;
+    }
+
+  private:
+    /// callback function
+    const parser_callback_t<BasicJsonType> callback = nullptr;
+    /// the type of the last read token
+    token_type last_token = token_type::uninitialized;
+    /// the lexer
+    lexer_t m_lexer;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+
+#include <cstddef> // ptrdiff_t
+#include <limits>  // numeric_limits
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+/*
+@brief an iterator for primitive JSON types
+
+This class models an iterator for primitive JSON types (boolean, number,
+string). It's only purpose is to allow the iterator/const_iterator classes
+to "iterate" over primitive values. Internally, the iterator is modeled by
+a `difference_type` variable. Value begin_value (`0`) models the begin,
+end_value (`1`) models past the end.
+*/
+class primitive_iterator_t
+{
+  private:
+    using difference_type = std::ptrdiff_t;
+    static constexpr difference_type begin_value = 0;
+    static constexpr difference_type end_value = begin_value + 1;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// iterator as signed integer type
+    difference_type m_it = (std::numeric_limits<std::ptrdiff_t>::min)();
+
+  public:
+    constexpr difference_type get_value() const noexcept
+    {
+        return m_it;
+    }
+
+    /// set iterator to a defined beginning
+    void set_begin() noexcept
+    {
+        m_it = begin_value;
+    }
+
+    /// set iterator to a defined past the end
+    void set_end() noexcept
+    {
+        m_it = end_value;
+    }
+
+    /// return whether the iterator can be dereferenced
+    constexpr bool is_begin() const noexcept
+    {
+        return m_it == begin_value;
+    }
+
+    /// return whether the iterator is at end
+    constexpr bool is_end() const noexcept
+    {
+        return m_it == end_value;
+    }
+
+    friend constexpr bool operator==(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it == rhs.m_it;
+    }
+
+    friend constexpr bool operator<(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it < rhs.m_it;
+    }
+
+    primitive_iterator_t operator+(difference_type n) noexcept
+    {
+        auto result = *this;
+        result += n;
+        return result;
+    }
+
+    friend constexpr difference_type operator-(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it - rhs.m_it;
+    }
+
+    primitive_iterator_t& operator++() noexcept
+    {
+        ++m_it;
+        return *this;
+    }
+
+    primitive_iterator_t const operator++(int) noexcept // NOLINT(readability-const-return-type)
+    {
+        auto result = *this;
+        ++m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator--() noexcept
+    {
+        --m_it;
+        return *this;
+    }
+
+    primitive_iterator_t const operator--(int) noexcept // NOLINT(readability-const-return-type)
+    {
+        auto result = *this;
+        --m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator+=(difference_type n) noexcept
+    {
+        m_it += n;
+        return *this;
+    }
+
+    primitive_iterator_t& operator-=(difference_type n) noexcept
+    {
+        m_it -= n;
+        return *this;
+    }
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+
+namespace nlohmann
+{
+namespace detail
+{
+/*!
+@brief an iterator value
+
+@note This structure could easily be a union, but MSVC currently does not allow
+unions members with complex constructors, see https://github.com/nlohmann/json/pull/105.
+*/
+template<typename BasicJsonType> struct internal_iterator
+{
+    /// iterator for JSON objects
+    typename BasicJsonType::object_t::iterator object_iterator {};
+    /// iterator for JSON arrays
+    typename BasicJsonType::array_t::iterator array_iterator {};
+    /// generic iterator for all other types
+    primitive_iterator_t primitive_iterator {};
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/iterators/iter_impl.hpp>
+
+
+#include <iterator> // iterator, random_access_iterator_tag, bidirectional_iterator_tag, advance, next
+#include <type_traits> // conditional, is_const, remove_const
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+// forward declare, to be able to friend it later on
+template<typename IteratorType> class iteration_proxy;
+template<typename IteratorType> class iteration_proxy_value;
+
+/*!
+@brief a template for a bidirectional iterator for the @ref basic_json class
+This class implements a both iterators (iterator and const_iterator) for the
+@ref basic_json class.
+@note An iterator is called *initialized* when a pointer to a JSON value has
+      been set (e.g., by a constructor or a copy assignment). If the iterator is
+      default-constructed, it is *uninitialized* and most methods are undefined.
+      **The library uses assertions to detect calls on uninitialized iterators.**
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+@since version 1.0.0, simplified in version 2.0.9, change to bidirectional
+       iterators in version 3.0.0 (see https://github.com/nlohmann/json/issues/593)
+*/
+template<typename BasicJsonType>
+class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+{
+    /// the iterator with BasicJsonType of different const-ness
+    using other_iter_impl = iter_impl<typename std::conditional<std::is_const<BasicJsonType>::value, typename std::remove_const<BasicJsonType>::type, const BasicJsonType>::type>;
+    /// allow basic_json to access private members
+    friend other_iter_impl;
+    friend BasicJsonType;
+    friend iteration_proxy<iter_impl>;
+    friend iteration_proxy_value<iter_impl>;
+
+    using object_t = typename BasicJsonType::object_t;
+    using array_t = typename BasicJsonType::array_t;
+    // make sure BasicJsonType is basic_json or const basic_json
+    static_assert(is_basic_json<typename std::remove_const<BasicJsonType>::type>::value,
+                  "iter_impl only accepts (const) basic_json");
+
+  public:
+
+    /// The std::iterator class template (used as a base class to provide typedefs) is deprecated in C++17.
+    /// The C++ Standard has never required user-defined iterators to derive from std::iterator.
+    /// A user-defined iterator should provide publicly accessible typedefs named
+    /// iterator_category, value_type, difference_type, pointer, and reference.
+    /// Note that value_type is required to be non-const, even for constant iterators.
+    using iterator_category = std::bidirectional_iterator_tag;
+
+    /// the type of the values when the iterator is dereferenced
+    using value_type = typename BasicJsonType::value_type;
+    /// a type to represent differences between iterators
+    using difference_type = typename BasicJsonType::difference_type;
+    /// defines a pointer to the type iterated over (value_type)
+    using pointer = typename std::conditional<std::is_const<BasicJsonType>::value,
+          typename BasicJsonType::const_pointer,
+          typename BasicJsonType::pointer>::type;
+    /// defines a reference to the type iterated over (value_type)
+    using reference =
+        typename std::conditional<std::is_const<BasicJsonType>::value,
+        typename BasicJsonType::const_reference,
+        typename BasicJsonType::reference>::type;
+
+    iter_impl() = default;
+    ~iter_impl() = default;
+    iter_impl(iter_impl&&) noexcept = default;
+    iter_impl& operator=(iter_impl&&) noexcept = default;
+
+    /*!
+    @brief constructor for a given JSON instance
+    @param[in] object  pointer to a JSON object for this iterator
+    @pre object != nullptr
+    @post The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    explicit iter_impl(pointer object) noexcept : m_object(object)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = typename object_t::iterator();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = typename array_t::iterator();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator = primitive_iterator_t();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @note The conventional copy constructor and copy assignment are implicitly
+          defined. Combined with the following converting constructor and
+          assignment, they support: (1) copy from iterator to iterator, (2)
+          copy from const iterator to const iterator, and (3) conversion from
+          iterator to const iterator. However conversion from const iterator
+          to iterator is not defined.
+    */
+
+    /*!
+    @brief const copy constructor
+    @param[in] other const iterator to copy from
+    @note This copy constructor had to be defined explicitly to circumvent a bug
+          occurring on msvc v19.0 compiler (VS 2015) debug build. For more
+          information refer to: https://github.com/nlohmann/json/issues/1608
+    */
+    iter_impl(const iter_impl<const BasicJsonType>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<const BasicJsonType>& other) noexcept
+    {
+        if (&other != this)
+        {
+            m_object = other.m_object;
+            m_it = other.m_it;
+        }
+        return *this;
+    }
+
+    /*!
+    @brief converting constructor
+    @param[in] other  non-const iterator to copy from
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other  non-const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept // NOLINT(cert-oop54-cpp)
+    {
+        m_object = other.m_object;
+        m_it = other.m_it;
+        return *this;
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief set the iterator to the first value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_begin() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_value.object->begin();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_value.array->begin();
+                break;
+            }
+
+            case value_t::null:
+            {
+                // set to end so begin()==end() is true: null is empty
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_begin();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @brief set the iterator past the last value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_end() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_value.object->end();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_value.array->end();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+        }
+    }
+
+  public:
+    /*!
+    @brief return a reference to the value pointed to by the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator*() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
+                return m_it.object_iterator->second;
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
+                return *m_it.array_iterator;
+            }
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", *m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", *m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief dereference the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    pointer operator->() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
+                return &(m_it.object_iterator->second);
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
+                return &*m_it.array_iterator;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", *m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief post-increment (it++)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl const operator++(int) // NOLINT(readability-const-return-type)
+    {
+        auto result = *this;
+        ++(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-increment (++it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator++()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, 1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, 1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                ++m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief post-decrement (it--)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl const operator--(int) // NOLINT(readability-const-return-type)
+    {
+        auto result = *this;
+        --(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-decrement (--it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator--()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, -1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, -1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                --m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief comparison: equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator==(const IterImpl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", *m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                return (m_it.object_iterator == other.m_it.object_iterator);
+
+            case value_t::array:
+                return (m_it.array_iterator == other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator == other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: not equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator!=(const IterImpl& other) const
+    {
+        return !operator==(other);
+    }
+
+    /*!
+    @brief comparison: smaller
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<(const iter_impl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", *m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(213, "cannot compare order of object iterators", *m_object));
+
+            case value_t::array:
+                return (m_it.array_iterator < other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator < other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: less than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<=(const iter_impl& other) const
+    {
+        return !other.operator < (*this);
+    }
+
+    /*!
+    @brief comparison: greater than
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>(const iter_impl& other) const
+    {
+        return !operator<=(other);
+    }
+
+    /*!
+    @brief comparison: greater than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>=(const iter_impl& other) const
+    {
+        return !operator<(other);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator+=(difference_type i)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", *m_object));
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, i);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator += i;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator-=(difference_type i)
+    {
+        return operator+=(-i);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator+(difference_type i) const
+    {
+        auto result = *this;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief addition of distance and iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    friend iter_impl operator+(difference_type i, const iter_impl& it)
+    {
+        auto result = it;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator-(difference_type i) const
+    {
+        auto result = *this;
+        result -= i;
+        return result;
+    }
+
+    /*!
+    @brief return difference
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    difference_type operator-(const iter_impl& other) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", *m_object));
+
+            case value_t::array:
+                return m_it.array_iterator - other.m_it.array_iterator;
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return m_it.primitive_iterator - other.m_it.primitive_iterator;
+        }
+    }
+
+    /*!
+    @brief access to successor
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator[](difference_type n) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(208, "cannot use operator[] for object iterators", *m_object));
+
+            case value_t::array:
+                return *std::next(m_it.array_iterator, n);
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", *m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.get_value() == -n))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", *m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief return the key of an object iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    const typename object_t::key_type& key() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        if (JSON_HEDLEY_LIKELY(m_object->is_object()))
+        {
+            return m_it.object_iterator->first;
+        }
+
+        JSON_THROW(invalid_iterator::create(207, "cannot use key() for non-object iterators", *m_object));
+    }
+
+    /*!
+    @brief return the value of an iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference value() const
+    {
+        return operator*();
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// associated JSON instance
+    pointer m_object = nullptr;
+    /// the actual iterator of the associated instance
+    internal_iterator<typename std::remove_const<BasicJsonType>::type> m_it {};
+};
+} // namespace detail
+} // namespace nlohmann
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+
+// #include <nlohmann/detail/iterators/json_reverse_iterator.hpp>
+
+
+#include <cstddef> // ptrdiff_t
+#include <iterator> // reverse_iterator
+#include <utility> // declval
+
+namespace nlohmann
+{
+namespace detail
+{
+//////////////////////
+// reverse_iterator //
+//////////////////////
+
+/*!
+@brief a template for a reverse iterator class
+
+@tparam Base the base iterator type to reverse. Valid types are @ref
+iterator (to create @ref reverse_iterator) and @ref const_iterator (to
+create @ref const_reverse_iterator).
+
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+- [OutputIterator](https://en.cppreference.com/w/cpp/named_req/OutputIterator):
+  It is possible to write to the pointed-to element (only if @a Base is
+  @ref iterator).
+
+@since version 1.0.0
+*/
+template<typename Base>
+class json_reverse_iterator : public std::reverse_iterator<Base>
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    /// shortcut to the reverse iterator adapter
+    using base_iterator = std::reverse_iterator<Base>;
+    /// the reference type for the pointed-to element
+    using reference = typename Base::reference;
+
+    /// create reverse iterator from iterator
+    explicit json_reverse_iterator(const typename base_iterator::iterator_type& it) noexcept
+        : base_iterator(it) {}
+
+    /// create reverse iterator from base class
+    explicit json_reverse_iterator(const base_iterator& it) noexcept : base_iterator(it) {}
+
+    /// post-increment (it++)
+    json_reverse_iterator const operator++(int) // NOLINT(readability-const-return-type)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator++(1));
+    }
+
+    /// pre-increment (++it)
+    json_reverse_iterator& operator++()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator++());
+    }
+
+    /// post-decrement (it--)
+    json_reverse_iterator const operator--(int) // NOLINT(readability-const-return-type)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator--(1));
+    }
+
+    /// pre-decrement (--it)
+    json_reverse_iterator& operator--()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator--());
+    }
+
+    /// add to iterator
+    json_reverse_iterator& operator+=(difference_type i)
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator+=(i));
+    }
+
+    /// add to iterator
+    json_reverse_iterator operator+(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator+(i));
+    }
+
+    /// subtract from iterator
+    json_reverse_iterator operator-(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator-(i));
+    }
+
+    /// return difference
+    difference_type operator-(const json_reverse_iterator& other) const
+    {
+        return base_iterator(*this) - base_iterator(other);
+    }
+
+    /// access to successor
+    reference operator[](difference_type n) const
+    {
+        return *(this->operator+(n));
+    }
+
+    /// return the key of an object iterator
+    auto key() const -> decltype(std::declval<Base>().key())
+    {
+        auto it = --this->base();
+        return it.key();
+    }
+
+    /// return the value of an iterator
+    reference value() const
+    {
+        auto it = --this->base();
+        return it.operator * ();
+    }
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/json_pointer.hpp>
+
+
+#include <algorithm> // all_of
+#include <cctype> // isdigit
+#include <limits> // max
+#include <numeric> // accumulate
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+
+/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+/// @sa https://json.nlohmann.me/api/json_pointer/
+template<typename BasicJsonType>
+class json_pointer
+{
+    // allow basic_json to access private members
+    NLOHMANN_BASIC_JSON_TPL_DECLARATION
+    friend class basic_json;
+
+  public:
+    /// @brief create JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/json_pointer/
+    explicit json_pointer(const std::string& s = "")
+        : reference_tokens(split(s))
+    {}
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/to_string/
+    std::string to_string() const
+    {
+        return std::accumulate(reference_tokens.begin(), reference_tokens.end(),
+                               std::string{},
+                               [](const std::string & a, const std::string & b)
+        {
+            return a + "/" + detail::escape(b);
+        });
+    }
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_string/
+    operator std::string() const
+    {
+        return to_string();
+    }
+
+    /// @brief append another JSON pointer at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(const json_pointer& ptr)
+    {
+        reference_tokens.insert(reference_tokens.end(),
+                                ptr.reference_tokens.begin(),
+                                ptr.reference_tokens.end());
+        return *this;
+    }
+
+    /// @brief append an unescaped reference token at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(std::string token)
+    {
+        push_back(std::move(token));
+        return *this;
+    }
+
+    /// @brief append an array index at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(std::size_t array_idx)
+    {
+        return *this /= std::to_string(array_idx);
+    }
+
+    /// @brief create a new JSON pointer by appending the right JSON pointer at the end of the left JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs,
+                                  const json_pointer& rhs)
+    {
+        return json_pointer(lhs) /= rhs;
+    }
+
+    /// @brief create a new JSON pointer by appending the unescaped token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, std::string token) // NOLINT(performance-unnecessary-value-param)
+    {
+        return json_pointer(lhs) /= std::move(token);
+    }
+
+    /// @brief create a new JSON pointer by appending the array-index-token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, std::size_t array_idx)
+    {
+        return json_pointer(lhs) /= array_idx;
+    }
+
+    /// @brief returns the parent of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/parent_pointer/
+    json_pointer parent_pointer() const
+    {
+        if (empty())
+        {
+            return *this;
+        }
+
+        json_pointer res = *this;
+        res.pop_back();
+        return res;
+    }
+
+    /// @brief remove last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/pop_back/
+    void pop_back()
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", BasicJsonType()));
+        }
+
+        reference_tokens.pop_back();
+    }
+
+    /// @brief return last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/back/
+    const std::string& back() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", BasicJsonType()));
+        }
+
+        return reference_tokens.back();
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(const std::string& token)
+    {
+        reference_tokens.push_back(token);
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(std::string&& token)
+    {
+        reference_tokens.push_back(std::move(token));
+    }
+
+    /// @brief return whether pointer points to the root document
+    /// @sa https://json.nlohmann.me/api/json_pointer/empty/
+    bool empty() const noexcept
+    {
+        return reference_tokens.empty();
+    }
+
+  private:
+    /*!
+    @param[in] s  reference token to be converted into an array index
+
+    @return integer representation of @a s
+
+    @throw parse_error.106  if an array index begins with '0'
+    @throw parse_error.109  if an array index begins not with a digit
+    @throw out_of_range.404 if string @a s could not be converted to an integer
+    @throw out_of_range.410 if an array index exceeds size_type
+    */
+    static typename BasicJsonType::size_type array_index(const std::string& s)
+    {
+        using size_type = typename BasicJsonType::size_type;
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && s[0] == '0'))
+        {
+            JSON_THROW(detail::parse_error::create(106, 0, "array index '" + s + "' must not begin with '0'", BasicJsonType()));
+        }
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && !(s[0] >= '1' && s[0] <= '9')))
+        {
+            JSON_THROW(detail::parse_error::create(109, 0, "array index '" + s + "' is not a number", BasicJsonType()));
+        }
+
+        std::size_t processed_chars = 0;
+        unsigned long long res = 0;  // NOLINT(runtime/int)
+        JSON_TRY
+        {
+            res = std::stoull(s, &processed_chars);
+        }
+        JSON_CATCH(std::out_of_range&)
+        {
+            JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + s + "'", BasicJsonType()));
+        }
+
+        // check if the string was completely read
+        if (JSON_HEDLEY_UNLIKELY(processed_chars != s.size()))
+        {
+            JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + s + "'", BasicJsonType()));
+        }
+
+        // only triggered on special platforms (like 32bit), see also
+        // https://github.com/nlohmann/json/pull/2203
+        if (res >= static_cast<unsigned long long>((std::numeric_limits<size_type>::max)()))  // NOLINT(runtime/int)
+        {
+            JSON_THROW(detail::out_of_range::create(410, "array index " + s + " exceeds size_type", BasicJsonType())); // LCOV_EXCL_LINE
+        }
+
+        return static_cast<size_type>(res);
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    json_pointer top() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", BasicJsonType()));
+        }
+
+        json_pointer result = *this;
+        result.reference_tokens = {reference_tokens[0]};
+        return result;
+    }
+
+  private:
+    /*!
+    @brief create and return a reference to the pointed to value
+
+    @complexity Linear in the number of reference tokens.
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.313 if value cannot be unflattened
+    */
+    BasicJsonType& get_and_create(BasicJsonType& j) const
+    {
+        auto* result = &j;
+
+        // in case no reference tokens exist, return a reference to the JSON value
+        // j which will be overwritten by a primitive value
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (result->type())
+            {
+                case detail::value_t::null:
+                {
+                    if (reference_token == "0")
+                    {
+                        // start a new array if reference token is 0
+                        result = &result->operator[](0);
+                    }
+                    else
+                    {
+                        // start a new object otherwise
+                        result = &result->operator[](reference_token);
+                    }
+                    break;
+                }
+
+                case detail::value_t::object:
+                {
+                    // create an entry in the object
+                    result = &result->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    // create an entry in the array
+                    result = &result->operator[](array_index(reference_token));
+                    break;
+                }
+
+                /*
+                The following code is only reached if there exists a reference
+                token _and_ the current value is primitive. In this case, we have
+                an error situation, because primitive values may only occur as
+                single value; that is, with an empty list of reference tokens.
+                */
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::type_error::create(313, "invalid value to unflatten", j));
+            }
+        }
+
+        return *result;
+    }
+
+    /*!
+    @brief return a reference to the pointed to value
+
+    @note This version does not throw if a value is not present, but tries to
+          create nested values instead. For instance, calling this function
+          with pointer `"/this/that"` on a null value is equivalent to calling
+          `operator[]("this").operator[]("that")` on that value, effectively
+          changing the null value to an object.
+
+    @param[in] ptr  a JSON value
+
+    @return reference to the JSON value pointed to by the JSON pointer
+
+    @complexity Linear in the length of the JSON pointer.
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    BasicJsonType& get_unchecked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            // convert null values to arrays or objects before continuing
+            if (ptr->is_null())
+            {
+                // check if reference token is a number
+                const bool nums =
+                    std::all_of(reference_token.begin(), reference_token.end(),
+                                [](const unsigned char x)
+                {
+                    return std::isdigit(x);
+                });
+
+                // change value to array for numbers or "-" or to object otherwise
+                *ptr = (nums || reference_token == "-")
+                       ? detail::value_t::array
+                       : detail::value_t::object;
+            }
+
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (reference_token == "-")
+                    {
+                        // explicitly treat "-" as index beyond the end
+                        ptr = &ptr->operator[](ptr->m_value.array->size());
+                    }
+                    else
+                    {
+                        // convert array index to number; unchecked access
+                        ptr = &ptr->operator[](array_index(reference_token));
+                    }
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + reference_token + "'", *ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    BasicJsonType& get_checked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402,
+                                                                "array index '-' (" + std::to_string(ptr->m_value.array->size()) +
+                                                                ") is out of range", *ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + reference_token + "'", *ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @brief return a const reference to the pointed to value
+
+    @param[in] ptr  a JSON value
+
+    @return const reference to the JSON value pointed to by the JSON
+    pointer
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    const BasicJsonType& get_unchecked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" cannot be used for const access
+                        JSON_THROW(detail::out_of_range::create(402, "array index '-' (" + std::to_string(ptr->m_value.array->size()) + ") is out of range", *ptr));
+                    }
+
+                    // use unchecked array access
+                    ptr = &ptr->operator[](array_index(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + reference_token + "'", *ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    const BasicJsonType& get_checked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402,
+                                                                "array index '-' (" + std::to_string(ptr->m_value.array->size()) +
+                                                                ") is out of range", *ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, "unresolved reference token '" + reference_token + "'", *ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    */
+    bool contains(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    if (!ptr->contains(reference_token))
+                    {
+                        // we did not find the key in the object
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() == 1 && !("0" <= reference_token && reference_token <= "9")))
+                    {
+                        // invalid char
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() > 1))
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!('1' <= reference_token[0] && reference_token[0] <= '9')))
+                        {
+                            // first char should be between '1' and '9'
+                            return false;
+                        }
+                        for (std::size_t i = 1; i < reference_token.size(); i++)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!('0' <= reference_token[i] && reference_token[i] <= '9')))
+                            {
+                                // other char should be between '0' and '9'
+                                return false;
+                            }
+                        }
+                    }
+
+                    const auto idx = array_index(reference_token);
+                    if (idx >= ptr->size())
+                    {
+                        // index out of range
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](idx);
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                {
+                    // we do not expect primitive values if there is still a
+                    // reference token to process
+                    return false;
+                }
+            }
+        }
+
+        // no reference token left means we found a primitive value
+        return true;
+    }
+
+    /*!
+    @brief split the string input to reference tokens
+
+    @note This function is only called by the json_pointer constructor.
+          All exceptions below are documented there.
+
+    @throw parse_error.107  if the pointer is not empty or begins with '/'
+    @throw parse_error.108  if character '~' is not followed by '0' or '1'
+    */
+    static std::vector<std::string> split(const std::string& reference_string)
+    {
+        std::vector<std::string> result;
+
+        // special case: empty reference string -> no reference tokens
+        if (reference_string.empty())
+        {
+            return result;
+        }
+
+        // check if nonempty reference string begins with slash
+        if (JSON_HEDLEY_UNLIKELY(reference_string[0] != '/'))
+        {
+            JSON_THROW(detail::parse_error::create(107, 1, "JSON pointer must be empty or begin with '/' - was: '" + reference_string + "'", BasicJsonType()));
+        }
+
+        // extract the reference tokens:
+        // - slash: position of the last read slash (or end of string)
+        // - start: position after the previous slash
+        for (
+            // search for the first slash after the first character
+            std::size_t slash = reference_string.find_first_of('/', 1),
+            // set the beginning of the first reference token
+            start = 1;
+            // we can stop if start == 0 (if slash == std::string::npos)
+            start != 0;
+            // set the beginning of the next reference token
+            // (will eventually be 0 if slash == std::string::npos)
+            start = (slash == std::string::npos) ? 0 : slash + 1,
+            // find next slash
+            slash = reference_string.find_first_of('/', start))
+        {
+            // use the text between the beginning of the reference token
+            // (start) and the last slash (slash).
+            auto reference_token = reference_string.substr(start, slash - start);
+
+            // check reference tokens are properly escaped
+            for (std::size_t pos = reference_token.find_first_of('~');
+                    pos != std::string::npos;
+                    pos = reference_token.find_first_of('~', pos + 1))
+            {
+                JSON_ASSERT(reference_token[pos] == '~');
+
+                // ~ must be followed by 0 or 1
+                if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
+                                         (reference_token[pos + 1] != '0' &&
+                                          reference_token[pos + 1] != '1')))
+                {
+                    JSON_THROW(detail::parse_error::create(108, 0, "escape character '~' must be followed with '0' or '1'", BasicJsonType()));
+                }
+            }
+
+            // finally, store the reference token
+            detail::unescape(reference_token);
+            result.push_back(reference_token);
+        }
+
+        return result;
+    }
+
+  private:
+    /*!
+    @param[in] reference_string  the reference string to the current value
+    @param[in] value             the value to consider
+    @param[in,out] result        the result object to insert values to
+
+    @note Empty objects or arrays are flattened to `null`.
+    */
+    static void flatten(const std::string& reference_string,
+                        const BasicJsonType& value,
+                        BasicJsonType& result)
+    {
+        switch (value.type())
+        {
+            case detail::value_t::array:
+            {
+                if (value.m_value.array->empty())
+                {
+                    // flatten empty array as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate array and use index as reference string
+                    for (std::size_t i = 0; i < value.m_value.array->size(); ++i)
+                    {
+                        flatten(reference_string + "/" + std::to_string(i),
+                                value.m_value.array->operator[](i), result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::object:
+            {
+                if (value.m_value.object->empty())
+                {
+                    // flatten empty object as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate object and use keys as reference string
+                    for (const auto& element : *value.m_value.object)
+                    {
+                        flatten(reference_string + "/" + detail::escape(element.first), element.second, result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::null:
+            case detail::value_t::string:
+            case detail::value_t::boolean:
+            case detail::value_t::number_integer:
+            case detail::value_t::number_unsigned:
+            case detail::value_t::number_float:
+            case detail::value_t::binary:
+            case detail::value_t::discarded:
+            default:
+            {
+                // add primitive value with its reference string
+                result[reference_string] = value;
+                break;
+            }
+        }
+    }
+
+    /*!
+    @param[in] value  flattened JSON
+
+    @return unflattened JSON
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.314  if value is not an object
+    @throw type_error.315  if object values are not primitive
+    @throw type_error.313  if value cannot be unflattened
+    */
+    static BasicJsonType
+    unflatten(const BasicJsonType& value)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!value.is_object()))
+        {
+            JSON_THROW(detail::type_error::create(314, "only objects can be unflattened", value));
+        }
+
+        BasicJsonType result;
+
+        // iterate the JSON object values
+        for (const auto& element : *value.m_value.object)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive()))
+            {
+                JSON_THROW(detail::type_error::create(315, "values in object must be primitive", element.second));
+            }
+
+            // assign value to reference pointed to by JSON pointer; Note that if
+            // the JSON pointer is "" (i.e., points to the whole value), function
+            // get_and_create returns a reference to result itself. An assignment
+            // will then create a primitive value.
+            json_pointer(element.first).get_and_create(result) = element.second;
+        }
+
+        return result;
+    }
+
+    /*!
+    @brief compares two JSON pointers for equality
+
+    @param[in] lhs  JSON pointer to compare
+    @param[in] rhs  JSON pointer to compare
+    @return whether @a lhs is equal to @a rhs
+
+    @complexity Linear in the length of the JSON pointer
+
+    @exceptionsafety No-throw guarantee: this function never throws exceptions.
+    */
+    friend bool operator==(json_pointer const& lhs,
+                           json_pointer const& rhs) noexcept
+    {
+        return lhs.reference_tokens == rhs.reference_tokens;
+    }
+
+    /*!
+    @brief compares two JSON pointers for inequality
+
+    @param[in] lhs  JSON pointer to compare
+    @param[in] rhs  JSON pointer to compare
+    @return whether @a lhs is not equal @a rhs
+
+    @complexity Linear in the length of the JSON pointer
+
+    @exceptionsafety No-throw guarantee: this function never throws exceptions.
+    */
+    friend bool operator!=(json_pointer const& lhs,
+                           json_pointer const& rhs) noexcept
+    {
+        return !(lhs == rhs);
+    }
+
+    /// the reference tokens
+    std::vector<std::string> reference_tokens;
+};
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/json_ref.hpp>
+
+
+#include <initializer_list>
+#include <utility>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+template<typename BasicJsonType>
+class json_ref
+{
+  public:
+    using value_type = BasicJsonType;
+
+    json_ref(value_type&& value)
+        : owned_value(std::move(value))
+    {}
+
+    json_ref(const value_type& value)
+        : value_ref(&value)
+    {}
+
+    json_ref(std::initializer_list<json_ref> init)
+        : owned_value(init)
+    {}
+
+    template <
+        class... Args,
+        enable_if_t<std::is_constructible<value_type, Args...>::value, int> = 0 >
+    json_ref(Args && ... args)
+        : owned_value(std::forward<Args>(args)...)
+    {}
+
+    // class should be movable only
+    json_ref(json_ref&&) noexcept = default;
+    json_ref(const json_ref&) = delete;
+    json_ref& operator=(const json_ref&) = delete;
+    json_ref& operator=(json_ref&&) = delete;
+    ~json_ref() = default;
+
+    value_type moved_or_copied() const
+    {
+        if (value_ref == nullptr)
+        {
+            return std::move(owned_value);
+        }
+        return *value_ref;
+    }
+
+    value_type const& operator*() const
+    {
+        return value_ref ? *value_ref : owned_value;
+    }
+
+    value_type const* operator->() const
+    {
+        return &** this;
+    }
+
+  private:
+    mutable value_type owned_value = nullptr;
+    value_type const* value_ref = nullptr;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+
+
+#include <algorithm> // reverse
+#include <array> // array
+#include <cmath> // isnan, isinf
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstring> // memcpy
+#include <limits> // numeric_limits
+#include <string> // string
+#include <utility> // move
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+
+#include <algorithm> // copy
+#include <cstddef> // size_t
+#include <iterator> // back_inserter
+#include <memory> // shared_ptr, make_shared
+#include <string> // basic_string
+#include <vector> // vector
+
+#ifndef JSON_NO_IO
+    #include <ios>      // streamsize
+    #include <ostream>  // basic_ostream
+#endif  // JSON_NO_IO
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+/// abstract output adapter interface
+template<typename CharType> struct output_adapter_protocol
+{
+    virtual void write_character(CharType c) = 0;
+    virtual void write_characters(const CharType* s, std::size_t length) = 0;
+    virtual ~output_adapter_protocol() = default;
+
+    output_adapter_protocol() = default;
+    output_adapter_protocol(const output_adapter_protocol&) = default;
+    output_adapter_protocol(output_adapter_protocol&&) noexcept = default;
+    output_adapter_protocol& operator=(const output_adapter_protocol&) = default;
+    output_adapter_protocol& operator=(output_adapter_protocol&&) noexcept = default;
+};
+
+/// a type to simplify interfaces
+template<typename CharType>
+using output_adapter_t = std::shared_ptr<output_adapter_protocol<CharType>>;
+
+/// output adapter for byte vectors
+template<typename CharType, typename AllocatorType = std::allocator<CharType>>
+class output_vector_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_vector_adapter(std::vector<CharType, AllocatorType>& vec) noexcept
+        : v(vec)
+    {}
+
+    void write_character(CharType c) override
+    {
+        v.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        std::copy(s, s + length, std::back_inserter(v));
+    }
+
+  private:
+    std::vector<CharType, AllocatorType>& v;
+};
+
+#ifndef JSON_NO_IO
+/// output adapter for output streams
+template<typename CharType>
+class output_stream_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_stream_adapter(std::basic_ostream<CharType>& s) noexcept
+        : stream(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        stream.put(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        stream.write(s, static_cast<std::streamsize>(length));
+    }
+
+  private:
+    std::basic_ostream<CharType>& stream;
+};
+#endif  // JSON_NO_IO
+
+/// output adapter for basic_string
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_string_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_string_adapter(StringType& s) noexcept
+        : str(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        str.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        str.append(s, length);
+    }
+
+  private:
+    StringType& str;
+};
+
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_adapter
+{
+  public:
+    template<typename AllocatorType = std::allocator<CharType>>
+    output_adapter(std::vector<CharType, AllocatorType>& vec)
+        : oa(std::make_shared<output_vector_adapter<CharType, AllocatorType>>(vec)) {}
+
+#ifndef JSON_NO_IO
+    output_adapter(std::basic_ostream<CharType>& s)
+        : oa(std::make_shared<output_stream_adapter<CharType>>(s)) {}
+#endif  // JSON_NO_IO
+
+    output_adapter(StringType& s)
+        : oa(std::make_shared<output_string_adapter<CharType, StringType>>(s)) {}
+
+    operator output_adapter_t<CharType>()
+    {
+        return oa;
+    }
+
+  private:
+    output_adapter_t<CharType> oa = nullptr;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+
+namespace nlohmann
+{
+namespace detail
+{
+///////////////////
+// binary writer //
+///////////////////
+
+/*!
+@brief serialization to CBOR and MessagePack values
+*/
+template<typename BasicJsonType, typename CharType>
+class binary_writer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+  public:
+    /*!
+    @brief create a binary writer
+
+    @param[in] adapter  output adapter to write to
+    */
+    explicit binary_writer(output_adapter_t<CharType> adapter) : oa(std::move(adapter))
+    {
+        JSON_ASSERT(oa);
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @pre       j.type() == value_t::object
+    */
+    void write_bson(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+            {
+                write_bson_object(*j.m_value.object);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::array:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                JSON_THROW(type_error::create(317, "to serialize to BSON, top-level type must be object, but is " + std::string(j.type_name()), j));
+            }
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_cbor(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                oa->write_character(to_char_type(0xF6));
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                oa->write_character(j.m_value.boolean
+                                    ? to_char_type(0xF5)
+                                    : to_char_type(0xF4));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_value.number_integer >= 0)
+                {
+                    // CBOR does not differentiate between positive signed
+                    // integers and unsigned integers. Therefore, we used the
+                    // code from the value_t::number_unsigned case here.
+                    if (j.m_value.number_integer <= 0x17)
+                    {
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x18));
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x19));
+                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x1A));
+                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x1B));
+                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    // The conversions below encode the sign in the first
+                    // byte, and the value is converted to a positive number.
+                    const auto positive_number = -1 - j.m_value.number_integer;
+                    if (j.m_value.number_integer >= -24)
+                    {
+                        write_number(static_cast<std::uint8_t>(0x20 + positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x38));
+                        write_number(static_cast<std::uint8_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x39));
+                        write_number(static_cast<std::uint16_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x3A));
+                        write_number(static_cast<std::uint32_t>(positive_number));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x3B));
+                        write_number(static_cast<std::uint64_t>(positive_number));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x18));
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x19));
+                    write_number(static_cast<std::uint16_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x1A));
+                    write_number(static_cast<std::uint32_t>(j.m_value.number_unsigned));
+                }
+                else
+                {
+                    oa->write_character(to_char_type(0x1B));
+                    write_number(static_cast<std::uint64_t>(j.m_value.number_unsigned));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                if (std::isnan(j.m_value.number_float))
+                {
+                    // NaN is 0xf97e00 in CBOR
+                    oa->write_character(to_char_type(0xF9));
+                    oa->write_character(to_char_type(0x7E));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else if (std::isinf(j.m_value.number_float))
+                {
+                    // Infinity is 0xf97c00, -Infinity is 0xf9fc00
+                    oa->write_character(to_char_type(0xf9));
+                    oa->write_character(j.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else
+                {
+                    write_compact_float(j.m_value.number_float, detail::input_format_t::cbor);
+                }
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_value.string->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x60 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x78));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x79));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_value.array->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x80 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x98));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x99));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_cbor(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (j.m_value.binary->has_subtype())
+                {
+                    if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd8));
+                        write_number(static_cast<std::uint8_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd9));
+                        write_number(static_cast<std::uint16_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xda));
+                        write_number(static_cast<std::uint32_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xdb));
+                        write_number(static_cast<std::uint64_t>(j.m_value.binary->subtype()));
+                    }
+                }
+
+                // step 1: write control byte and the binary array size
+                const auto N = j.m_value.binary->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x40 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x58));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x59));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_value.object->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0xA0 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB8));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB9));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBA));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBB));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_cbor(el.first);
+                    write_cbor(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_msgpack(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null: // nil
+            {
+                oa->write_character(to_char_type(0xC0));
+                break;
+            }
+
+            case value_t::boolean: // true and false
+            {
+                oa->write_character(j.m_value.boolean
+                                    ? to_char_type(0xC3)
+                                    : to_char_type(0xC2));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_value.number_integer >= 0)
+                {
+                    // MessagePack does not differentiate between positive
+                    // signed integers and unsigned integers. Therefore, we used
+                    // the code from the value_t::number_unsigned case here.
+                    if (j.m_value.number_unsigned < 128)
+                    {
+                        // positive fixnum
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        // uint 8
+                        oa->write_character(to_char_type(0xCC));
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        // uint 16
+                        oa->write_character(to_char_type(0xCD));
+                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        // uint 32
+                        oa->write_character(to_char_type(0xCE));
+                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        // uint 64
+                        oa->write_character(to_char_type(0xCF));
+                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    if (j.m_value.number_integer >= -32)
+                    {
+                        // negative fixnum
+                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int8_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                    {
+                        // int 8
+                        oa->write_character(to_char_type(0xD0));
+                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int16_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                    {
+                        // int 16
+                        oa->write_character(to_char_type(0xD1));
+                        write_number(static_cast<std::int16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int32_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                    {
+                        // int 32
+                        oa->write_character(to_char_type(0xD2));
+                        write_number(static_cast<std::int32_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int64_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                    {
+                        // int 64
+                        oa->write_character(to_char_type(0xD3));
+                        write_number(static_cast<std::int64_t>(j.m_value.number_integer));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned < 128)
+                {
+                    // positive fixnum
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // uint 8
+                    oa->write_character(to_char_type(0xCC));
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // uint 16
+                    oa->write_character(to_char_type(0xCD));
+                    write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // uint 32
+                    oa->write_character(to_char_type(0xCE));
+                    write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    // uint 64
+                    oa->write_character(to_char_type(0xCF));
+                    write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_compact_float(j.m_value.number_float, detail::input_format_t::msgpack);
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_value.string->size();
+                if (N <= 31)
+                {
+                    // fixstr
+                    write_number(static_cast<std::uint8_t>(0xA0 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // str 8
+                    oa->write_character(to_char_type(0xD9));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // str 16
+                    oa->write_character(to_char_type(0xDA));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // str 32
+                    oa->write_character(to_char_type(0xDB));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_value.array->size();
+                if (N <= 15)
+                {
+                    // fixarray
+                    write_number(static_cast<std::uint8_t>(0x90 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // array 16
+                    oa->write_character(to_char_type(0xDC));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // array 32
+                    oa->write_character(to_char_type(0xDD));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_msgpack(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                // step 0: determine if the binary type has a set subtype to
+                // determine whether or not to use the ext or fixext types
+                const bool use_ext = j.m_value.binary->has_subtype();
+
+                // step 1: write control byte and the byte string length
+                const auto N = j.m_value.binary->size();
+                if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    std::uint8_t output_type{};
+                    bool fixed = true;
+                    if (use_ext)
+                    {
+                        switch (N)
+                        {
+                            case 1:
+                                output_type = 0xD4; // fixext 1
+                                break;
+                            case 2:
+                                output_type = 0xD5; // fixext 2
+                                break;
+                            case 4:
+                                output_type = 0xD6; // fixext 4
+                                break;
+                            case 8:
+                                output_type = 0xD7; // fixext 8
+                                break;
+                            case 16:
+                                output_type = 0xD8; // fixext 16
+                                break;
+                            default:
+                                output_type = 0xC7; // ext 8
+                                fixed = false;
+                                break;
+                        }
+
+                    }
+                    else
+                    {
+                        output_type = 0xC4; // bin 8
+                        fixed = false;
+                    }
+
+                    oa->write_character(to_char_type(output_type));
+                    if (!fixed)
+                    {
+                        write_number(static_cast<std::uint8_t>(N));
+                    }
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    std::uint8_t output_type = use_ext
+                                               ? 0xC8 // ext 16
+                                               : 0xC5; // bin 16
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    std::uint8_t output_type = use_ext
+                                               ? 0xC9 // ext 32
+                                               : 0xC6; // bin 32
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 1.5: if this is an ext type, write the subtype
+                if (use_ext)
+                {
+                    write_number(static_cast<std::int8_t>(j.m_value.binary->subtype()));
+                }
+
+                // step 2: write the byte string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_value.object->size();
+                if (N <= 15)
+                {
+                    // fixmap
+                    write_number(static_cast<std::uint8_t>(0x80 | (N & 0xF)));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // map 16
+                    oa->write_character(to_char_type(0xDE));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // map 32
+                    oa->write_character(to_char_type(0xDF));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_msgpack(el.first);
+                    write_msgpack(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @param[in] use_count   whether to use '#' prefixes (optimized format)
+    @param[in] use_type    whether to use '$' prefixes (optimized format)
+    @param[in] add_prefix  whether prefixes need to be used for this value
+    */
+    void write_ubjson(const BasicJsonType& j, const bool use_count,
+                      const bool use_type, const bool add_prefix = true)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('Z'));
+                }
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(j.m_value.boolean
+                                        ? to_char_type('T')
+                                        : to_char_type('F'));
+                }
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_integer, add_prefix);
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_unsigned, add_prefix);
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_float, add_prefix);
+                break;
+            }
+
+            case value_t::string:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('S'));
+                }
+                write_number_with_ubjson_prefix(j.m_value.string->size(), true);
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_value.array->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front());
+                    const bool same_prefix = std::all_of(j.begin() + 1, j.end(),
+                                                         [this, first_prefix](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v) == first_prefix;
+                    });
+
+                    if (same_prefix)
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.array->size(), true);
+                }
+
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_ubjson(el, use_count, use_type, prefix_required);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                if (use_type && !j.m_value.binary->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    oa->write_character(to_char_type('$'));
+                    oa->write_character('U');
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.binary->size(), true);
+                }
+
+                if (use_type)
+                {
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                        j.m_value.binary->size());
+                }
+                else
+                {
+                    for (size_t i = 0; i < j.m_value.binary->size(); ++i)
+                    {
+                        oa->write_character(to_char_type('U'));
+                        oa->write_character(j.m_value.binary->data()[i]);
+                    }
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('{'));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_value.object->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front());
+                    const bool same_prefix = std::all_of(j.begin(), j.end(),
+                                                         [this, first_prefix](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v) == first_prefix;
+                    });
+
+                    if (same_prefix)
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.object->size(), true);
+                }
+
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_number_with_ubjson_prefix(el.first.size(), true);
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(el.first.c_str()),
+                        el.first.size());
+                    write_ubjson(el.second, use_count, use_type, prefix_required);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type('}'));
+                }
+
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @return The size of a BSON document entry header, including the id marker
+            and the entry name size (and its null-terminator).
+    */
+    static std::size_t calc_bson_entry_header_size(const string_t& name, const BasicJsonType& j)
+    {
+        const auto it = name.find(static_cast<typename string_t::value_type>(0));
+        if (JSON_HEDLEY_UNLIKELY(it != BasicJsonType::string_t::npos))
+        {
+            JSON_THROW(out_of_range::create(409, "BSON key cannot contain code point U+0000 (at byte " + std::to_string(it) + ")", j));
+            static_cast<void>(j);
+        }
+
+        return /*id*/ 1ul + name.size() + /*zero-terminator*/1u;
+    }
+
+    /*!
+    @brief Writes the given @a element_type and @a name to the output adapter
+    */
+    void write_bson_entry_header(const string_t& name,
+                                 const std::uint8_t element_type)
+    {
+        oa->write_character(to_char_type(element_type)); // boolean
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(name.c_str()),
+            name.size() + 1u);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and boolean value @a value
+    */
+    void write_bson_boolean(const string_t& name,
+                            const bool value)
+    {
+        write_bson_entry_header(name, 0x08);
+        oa->write_character(value ? to_char_type(0x01) : to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and double value @a value
+    */
+    void write_bson_double(const string_t& name,
+                           const double value)
+    {
+        write_bson_entry_header(name, 0x01);
+        write_number<double, true>(value);
+    }
+
+    /*!
+    @return The size of the BSON-encoded string in @a value
+    */
+    static std::size_t calc_bson_string_size(const string_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and string value @a value
+    */
+    void write_bson_string(const string_t& name,
+                           const string_t& value)
+    {
+        write_bson_entry_header(name, 0x02);
+
+        write_number<std::int32_t, true>(static_cast<std::int32_t>(value.size() + 1ul));
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(value.c_str()),
+            value.size() + 1);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and null value
+    */
+    void write_bson_null(const string_t& name)
+    {
+        write_bson_entry_header(name, 0x0A);
+    }
+
+    /*!
+    @return The size of the BSON-encoded integer @a value
+    */
+    static std::size_t calc_bson_integer_size(const std::int64_t value)
+    {
+        return (std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)()
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and integer @a value
+    */
+    void write_bson_integer(const string_t& name,
+                            const std::int64_t value)
+    {
+        if ((std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            write_bson_entry_header(name, 0x10); // int32
+            write_number<std::int32_t, true>(static_cast<std::int32_t>(value));
+        }
+        else
+        {
+            write_bson_entry_header(name, 0x12); // int64
+            write_number<std::int64_t, true>(static_cast<std::int64_t>(value));
+        }
+    }
+
+    /*!
+    @return The size of the BSON-encoded unsigned integer in @a j
+    */
+    static constexpr std::size_t calc_bson_unsigned_size(const std::uint64_t value) noexcept
+    {
+        return (value <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and unsigned @a value
+    */
+    void write_bson_unsigned(const string_t& name,
+                             const BasicJsonType& j)
+    {
+        if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x10 /* int32 */);
+            write_number<std::int32_t, true>(static_cast<std::int32_t>(j.m_value.number_unsigned));
+        }
+        else if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x12 /* int64 */);
+            write_number<std::int64_t, true>(static_cast<std::int64_t>(j.m_value.number_unsigned));
+        }
+        else
+        {
+            JSON_THROW(out_of_range::create(407, "integer number " + std::to_string(j.m_value.number_unsigned) + " cannot be represented by BSON as it does not fit int64", j));
+        }
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and object @a value
+    */
+    void write_bson_object_entry(const string_t& name,
+                                 const typename BasicJsonType::object_t& value)
+    {
+        write_bson_entry_header(name, 0x03); // object
+        write_bson_object(value);
+    }
+
+    /*!
+    @return The size of the BSON-encoded array @a value
+    */
+    static std::size_t calc_bson_array_size(const typename BasicJsonType::array_t& value)
+    {
+        std::size_t array_index = 0ul;
+
+        const std::size_t embedded_document_size = std::accumulate(std::begin(value), std::end(value), static_cast<std::size_t>(0), [&array_index](std::size_t result, const typename BasicJsonType::array_t::value_type & el)
+        {
+            return result + calc_bson_element_size(std::to_string(array_index++), el);
+        });
+
+        return sizeof(std::int32_t) + embedded_document_size + 1ul;
+    }
+
+    /*!
+    @return The size of the BSON-encoded binary array @a value
+    */
+    static std::size_t calc_bson_binary_size(const typename BasicJsonType::binary_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and array @a value
+    */
+    void write_bson_array(const string_t& name,
+                          const typename BasicJsonType::array_t& value)
+    {
+        write_bson_entry_header(name, 0x04); // array
+        write_number<std::int32_t, true>(static_cast<std::int32_t>(calc_bson_array_size(value)));
+
+        std::size_t array_index = 0ul;
+
+        for (const auto& el : value)
+        {
+            write_bson_element(std::to_string(array_index++), el);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and binary value @a value
+    */
+    void write_bson_binary(const string_t& name,
+                           const binary_t& value)
+    {
+        write_bson_entry_header(name, 0x05);
+
+        write_number<std::int32_t, true>(static_cast<std::int32_t>(value.size()));
+        write_number(value.has_subtype() ? static_cast<std::uint8_t>(value.subtype()) : static_cast<std::uint8_t>(0x00));
+
+        oa->write_characters(reinterpret_cast<const CharType*>(value.data()), value.size());
+    }
+
+    /*!
+    @brief Calculates the size necessary to serialize the JSON value @a j with its @a name
+    @return The calculated size for the BSON document entry for @a j with the given @a name.
+    */
+    static std::size_t calc_bson_element_size(const string_t& name,
+            const BasicJsonType& j)
+    {
+        const auto header_size = calc_bson_entry_header_size(name, j);
+        switch (j.type())
+        {
+            case value_t::object:
+                return header_size + calc_bson_object_size(*j.m_value.object);
+
+            case value_t::array:
+                return header_size + calc_bson_array_size(*j.m_value.array);
+
+            case value_t::binary:
+                return header_size + calc_bson_binary_size(*j.m_value.binary);
+
+            case value_t::boolean:
+                return header_size + 1ul;
+
+            case value_t::number_float:
+                return header_size + 8ul;
+
+            case value_t::number_integer:
+                return header_size + calc_bson_integer_size(j.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return header_size + calc_bson_unsigned_size(j.m_value.number_unsigned);
+
+            case value_t::string:
+                return header_size + calc_bson_string_size(*j.m_value.string);
+
+            case value_t::null:
+                return header_size + 0ul;
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return 0ul;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Serializes the JSON value @a j to BSON and associates it with the
+           key @a name.
+    @param name The name to associate with the JSON entity @a j within the
+                current BSON document
+    */
+    void write_bson_element(const string_t& name,
+                            const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+                return write_bson_object_entry(name, *j.m_value.object);
+
+            case value_t::array:
+                return write_bson_array(name, *j.m_value.array);
+
+            case value_t::binary:
+                return write_bson_binary(name, *j.m_value.binary);
+
+            case value_t::boolean:
+                return write_bson_boolean(name, j.m_value.boolean);
+
+            case value_t::number_float:
+                return write_bson_double(name, j.m_value.number_float);
+
+            case value_t::number_integer:
+                return write_bson_integer(name, j.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return write_bson_unsigned(name, j);
+
+            case value_t::string:
+                return write_bson_string(name, *j.m_value.string);
+
+            case value_t::null:
+                return write_bson_null(name);
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Calculates the size of the BSON serialization of the given
+           JSON-object @a j.
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    static std::size_t calc_bson_object_size(const typename BasicJsonType::object_t& value)
+    {
+        std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast<std::size_t>(0),
+                                    [](size_t result, const typename BasicJsonType::object_t::value_type & el)
+        {
+            return result += calc_bson_element_size(el.first, el.second);
+        });
+
+        return sizeof(std::int32_t) + document_size + 1ul;
+    }
+
+    /*!
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    void write_bson_object(const typename BasicJsonType::object_t& value)
+    {
+        write_number<std::int32_t, true>(static_cast<std::int32_t>(calc_bson_object_size(value)));
+
+        for (const auto& el : value)
+        {
+            write_bson_element(el.first, el.second);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    static constexpr CharType get_cbor_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xFA);  // Single-Precision Float
+    }
+
+    static constexpr CharType get_cbor_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xFB);  // Double-Precision Float
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    static constexpr CharType get_msgpack_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xCA);  // float 32
+    }
+
+    static constexpr CharType get_msgpack_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xCB);  // float 64
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    // UBJSON: write number (floating point)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_floating_point<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix)
+    {
+        if (add_prefix)
+        {
+            oa->write_character(get_ubjson_float_prefix(n));
+        }
+        write_number(n);
+    }
+
+    // UBJSON: write number (unsigned integer)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_unsigned<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix)
+    {
+        if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::uint8_t>(n));
+        }
+        else if (n <= (std::numeric_limits<std::uint8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n));
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n));
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n));
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n));
+        }
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+    }
+
+    // UBJSON: write number (signed integer)
+    template < typename NumberType, typename std::enable_if <
+                   std::is_signed<NumberType>::value&&
+                   !std::is_floating_point<NumberType>::value, int >::type = 0 >
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix)
+    {
+        if ((std::numeric_limits<std::int8_t>::min)() <= n && n <= (std::numeric_limits<std::int8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::int8_t>(n));
+        }
+        else if (static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n));
+        }
+        else if ((std::numeric_limits<std::int16_t>::min)() <= n && n <= (std::numeric_limits<std::int16_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n));
+        }
+        else if ((std::numeric_limits<std::int32_t>::min)() <= n && n <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n));
+        }
+        else if ((std::numeric_limits<std::int64_t>::min)() <= n && n <= (std::numeric_limits<std::int64_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n));
+        }
+        // LCOV_EXCL_START
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+        // LCOV_EXCL_STOP
+    }
+
+    /*!
+    @brief determine the type prefix of container values
+    */
+    CharType ubjson_prefix(const BasicJsonType& j) const noexcept
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+                return 'Z';
+
+            case value_t::boolean:
+                return j.m_value.boolean ? 'T' : 'F';
+
+            case value_t::number_integer:
+            {
+                if ((std::numeric_limits<std::int8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                {
+                    return 'i';
+                }
+                if ((std::numeric_limits<std::uint8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    return 'U';
+                }
+                if ((std::numeric_limits<std::int16_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                {
+                    return 'I';
+                }
+                if ((std::numeric_limits<std::int32_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                {
+                    return 'l';
+                }
+                if ((std::numeric_limits<std::int64_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                {
+                    return 'L';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+                {
+                    return 'i';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint8_t>::max)()))
+                {
+                    return 'U';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+                {
+                    return 'I';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+                {
+                    return 'l';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+                {
+                    return 'L';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_float:
+                return get_ubjson_float_prefix(j.m_value.number_float);
+
+            case value_t::string:
+                return 'S';
+
+            case value_t::array: // fallthrough
+            case value_t::binary:
+                return '[';
+
+            case value_t::object:
+                return '{';
+
+            case value_t::discarded:
+            default:  // discarded values
+                return 'N';
+        }
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(float /*unused*/)
+    {
+        return 'd';  // float 32
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(double /*unused*/)
+    {
+        return 'D';  // float 64
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*
+    @brief write a number to output input
+    @param[in] n number of type @a NumberType
+    @tparam NumberType the type of the number
+    @tparam OutputIsLittleEndian Set to true if output data is
+                                 required to be little endian
+
+    @note This function needs to respect the system's endianness, because bytes
+          in CBOR, MessagePack, and UBJSON are stored in network order (big
+          endian) and therefore need reordering on little endian systems.
+    */
+    template<typename NumberType, bool OutputIsLittleEndian = false>
+    void write_number(const NumberType n)
+    {
+        // step 1: write number to array of length NumberType
+        std::array<CharType, sizeof(NumberType)> vec{};
+        std::memcpy(vec.data(), &n, sizeof(NumberType));
+
+        // step 2: write array to output (with possible reordering)
+        if (is_little_endian != OutputIsLittleEndian)
+        {
+            // reverse byte order prior to conversion if necessary
+            std::reverse(vec.begin(), vec.end());
+        }
+
+        oa->write_characters(vec.data(), sizeof(NumberType));
+    }
+
+    void write_compact_float(const number_float_t n, detail::input_format_t format)
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        if (static_cast<double>(n) >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
+                static_cast<double>(n) <= static_cast<double>((std::numeric_limits<float>::max)()) &&
+                static_cast<double>(static_cast<float>(n)) == static_cast<double>(n))
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(static_cast<float>(n))
+                                : get_msgpack_float_prefix(static_cast<float>(n)));
+            write_number(static_cast<float>(n));
+        }
+        else
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(n)
+                                : get_msgpack_float_prefix(n));
+            write_number(n);
+        }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+  public:
+    // The following to_char_type functions are implement the conversion
+    // between uint8_t and CharType. In case CharType is not unsigned,
+    // such a conversion is required to allow values greater than 128.
+    // See <https://github.com/nlohmann/json/issues/1286> for a discussion.
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_signed<char>::value > * = nullptr >
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return *reinterpret_cast<char*>(&x);
+    }
+
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_unsigned<char>::value > * = nullptr >
+    static CharType to_char_type(std::uint8_t x) noexcept
+    {
+        static_assert(sizeof(std::uint8_t) == sizeof(CharType), "size of CharType must be equal to std::uint8_t");
+        static_assert(std::is_trivial<CharType>::value, "CharType must be trivial");
+        CharType result;
+        std::memcpy(&result, &x, sizeof(x));
+        return result;
+    }
+
+    template<typename C = CharType,
+             enable_if_t<std::is_unsigned<C>::value>* = nullptr>
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return x;
+    }
+
+    template < typename InputCharType, typename C = CharType,
+               enable_if_t <
+                   std::is_signed<C>::value &&
+                   std::is_signed<char>::value &&
+                   std::is_same<char, typename std::remove_cv<InputCharType>::type>::value
+                   > * = nullptr >
+    static constexpr CharType to_char_type(InputCharType x) noexcept
+    {
+        return x;
+    }
+
+  private:
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// the output
+    output_adapter_t<CharType> oa = nullptr;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/output/serializer.hpp>
+
+
+#include <algorithm> // reverse, remove, fill, find, none_of
+#include <array> // array
+#include <clocale> // localeconv, lconv
+#include <cmath> // labs, isfinite, isnan, signbit
+#include <cstddef> // size_t, ptrdiff_t
+#include <cstdint> // uint8_t
+#include <cstdio> // snprintf
+#include <limits> // numeric_limits
+#include <string> // string, char_traits
+#include <iomanip> // setfill, setw
+#include <sstream> // stringstream
+#include <type_traits> // is_same
+#include <utility> // move
+
+// #include <nlohmann/detail/conversions/to_chars.hpp>
+
+
+#include <array> // array
+#include <cmath>   // signbit, isfinite
+#include <cstdint> // intN_t, uintN_t
+#include <cstring> // memcpy, memmove
+#include <limits> // numeric_limits
+#include <type_traits> // conditional
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+
+/*!
+@brief implements the Grisu2 algorithm for binary to decimal floating-point
+conversion.
+
+This implementation is a slightly modified version of the reference
+implementation which may be obtained from
+http://florian.loitsch.com/publications (bench.tar.gz).
+
+The code is distributed under the MIT license, Copyright (c) 2009 Florian Loitsch.
+
+For a detailed description of the algorithm see:
+
+[1] Loitsch, "Printing Floating-Point Numbers Quickly and Accurately with
+    Integers", Proceedings of the ACM SIGPLAN 2010 Conference on Programming
+    Language Design and Implementation, PLDI 2010
+[2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and Accurately",
+    Proceedings of the ACM SIGPLAN 1996 Conference on Programming Language
+    Design and Implementation, PLDI 1996
+*/
+namespace dtoa_impl
+{
+
+template<typename Target, typename Source>
+Target reinterpret_bits(const Source source)
+{
+    static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
+
+    Target target;
+    std::memcpy(&target, &source, sizeof(Source));
+    return target;
+}
+
+struct diyfp // f * 2^e
+{
+    static constexpr int kPrecision = 64; // = q
+
+    std::uint64_t f = 0;
+    int e = 0;
+
+    constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
+
+    /*!
+    @brief returns x - y
+    @pre x.e == y.e and x.f >= y.f
+    */
+    static diyfp sub(const diyfp& x, const diyfp& y) noexcept
+    {
+        JSON_ASSERT(x.e == y.e);
+        JSON_ASSERT(x.f >= y.f);
+
+        return {x.f - y.f, x.e};
+    }
+
+    /*!
+    @brief returns x * y
+    @note The result is rounded. (Only the upper q bits are returned.)
+    */
+    static diyfp mul(const diyfp& x, const diyfp& y) noexcept
+    {
+        static_assert(kPrecision == 64, "internal error");
+
+        // Computes:
+        //  f = round((x.f * y.f) / 2^q)
+        //  e = x.e + y.e + q
+
+        // Emulate the 64-bit * 64-bit multiplication:
+        //
+        // p = u * v
+        //   = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
+        //   = (u_lo v_lo         ) + 2^32 ((u_lo v_hi         ) + (u_hi v_lo         )) + 2^64 (u_hi v_hi         )
+        //   = (p0                ) + 2^32 ((p1                ) + (p2                )) + 2^64 (p3                )
+        //   = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3                )
+        //   = (p0_lo             ) + 2^32 (p0_hi + p1_lo + p2_lo                      ) + 2^64 (p1_hi + p2_hi + p3)
+        //   = (p0_lo             ) + 2^32 (Q                                          ) + 2^64 (H                 )
+        //   = (p0_lo             ) + 2^32 (Q_lo + 2^32 Q_hi                           ) + 2^64 (H                 )
+        //
+        // (Since Q might be larger than 2^32 - 1)
+        //
+        //   = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
+        //
+        // (Q_hi + H does not overflow a 64-bit int)
+        //
+        //   = p_lo + 2^64 p_hi
+
+        const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
+        const std::uint64_t u_hi = x.f >> 32u;
+        const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
+        const std::uint64_t v_hi = y.f >> 32u;
+
+        const std::uint64_t p0 = u_lo * v_lo;
+        const std::uint64_t p1 = u_lo * v_hi;
+        const std::uint64_t p2 = u_hi * v_lo;
+        const std::uint64_t p3 = u_hi * v_hi;
+
+        const std::uint64_t p0_hi = p0 >> 32u;
+        const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
+        const std::uint64_t p1_hi = p1 >> 32u;
+        const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
+        const std::uint64_t p2_hi = p2 >> 32u;
+
+        std::uint64_t Q = p0_hi + p1_lo + p2_lo;
+
+        // The full product might now be computed as
+        //
+        // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
+        // p_lo = p0_lo + (Q << 32)
+        //
+        // But in this particular case here, the full p_lo is not required.
+        // Effectively we only need to add the highest bit in p_lo to p_hi (and
+        // Q_hi + 1 does not overflow).
+
+        Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
+
+        const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
+
+        return {h, x.e + y.e + 64};
+    }
+
+    /*!
+    @brief normalize x such that the significand is >= 2^(q-1)
+    @pre x.f != 0
+    */
+    static diyfp normalize(diyfp x) noexcept
+    {
+        JSON_ASSERT(x.f != 0);
+
+        while ((x.f >> 63u) == 0)
+        {
+            x.f <<= 1u;
+            x.e--;
+        }
+
+        return x;
+    }
+
+    /*!
+    @brief normalize x such that the result has the exponent E
+    @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
+    */
+    static diyfp normalize_to(const diyfp& x, const int target_exponent) noexcept
+    {
+        const int delta = x.e - target_exponent;
+
+        JSON_ASSERT(delta >= 0);
+        JSON_ASSERT(((x.f << delta) >> delta) == x.f);
+
+        return {x.f << delta, target_exponent};
+    }
+};
+
+struct boundaries
+{
+    diyfp w;
+    diyfp minus;
+    diyfp plus;
+};
+
+/*!
+Compute the (normalized) diyfp representing the input number 'value' and its
+boundaries.
+
+@pre value must be finite and positive
+*/
+template<typename FloatType>
+boundaries compute_boundaries(FloatType value)
+{
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // Convert the IEEE representation into a diyfp.
+    //
+    // If v is denormal:
+    //      value = 0.F * 2^(1 - bias) = (          F) * 2^(1 - bias - (p-1))
+    // If v is normalized:
+    //      value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
+
+    static_assert(std::numeric_limits<FloatType>::is_iec559,
+                  "internal error: dtoa_short requires an IEEE-754 floating-point implementation");
+
+    constexpr int      kPrecision = std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
+    constexpr int      kBias      = std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
+    constexpr int      kMinExp    = 1 - kBias;
+    constexpr std::uint64_t kHiddenBit = std::uint64_t{1} << (kPrecision - 1); // = 2^(p-1)
+
+    using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, std::uint64_t >::type;
+
+    const auto bits = static_cast<std::uint64_t>(reinterpret_bits<bits_type>(value));
+    const std::uint64_t E = bits >> (kPrecision - 1);
+    const std::uint64_t F = bits & (kHiddenBit - 1);
+
+    const bool is_denormal = E == 0;
+    const diyfp v = is_denormal
+                    ? diyfp(F, kMinExp)
+                    : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
+
+    // Compute the boundaries m- and m+ of the floating-point value
+    // v = f * 2^e.
+    //
+    // Determine v- and v+, the floating-point predecessor and successor if v,
+    // respectively.
+    //
+    //      v- = v - 2^e        if f != 2^(p-1) or e == e_min                (A)
+    //         = v - 2^(e-1)    if f == 2^(p-1) and e > e_min                (B)
+    //
+    //      v+ = v + 2^e
+    //
+    // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
+    // between m- and m+ round to v, regardless of how the input rounding
+    // algorithm breaks ties.
+    //
+    //      ---+-------------+-------------+-------------+-------------+---  (A)
+    //         v-            m-            v             m+            v+
+    //
+    //      -----------------+------+------+-------------+-------------+---  (B)
+    //                       v-     m-     v             m+            v+
+
+    const bool lower_boundary_is_closer = F == 0 && E > 1;
+    const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
+    const diyfp m_minus = lower_boundary_is_closer
+                          ? diyfp(4 * v.f - 1, v.e - 2)  // (B)
+                          : diyfp(2 * v.f - 1, v.e - 1); // (A)
+
+    // Determine the normalized w+ = m+.
+    const diyfp w_plus = diyfp::normalize(m_plus);
+
+    // Determine w- = m- such that e_(w-) = e_(w+).
+    const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
+
+    return {diyfp::normalize(v), w_minus, w_plus};
+}
+
+// Given normalized diyfp w, Grisu needs to find a (normalized) cached
+// power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
+// within a certain range [alpha, gamma] (Definition 3.2 from [1])
+//
+//      alpha <= e = e_c + e_w + q <= gamma
+//
+// or
+//
+//      f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
+//                          <= f_c * f_w * 2^gamma
+//
+// Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
+//
+//      2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
+//
+// or
+//
+//      2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
+//
+// The choice of (alpha,gamma) determines the size of the table and the form of
+// the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
+// in practice:
+//
+// The idea is to cut the number c * w = f * 2^e into two parts, which can be
+// processed independently: An integral part p1, and a fractional part p2:
+//
+//      f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
+//              = (f div 2^-e) + (f mod 2^-e) * 2^e
+//              = p1 + p2 * 2^e
+//
+// The conversion of p1 into decimal form requires a series of divisions and
+// modulos by (a power of) 10. These operations are faster for 32-bit than for
+// 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
+// achieved by choosing
+//
+//      -e >= 32   or   e <= -32 := gamma
+//
+// In order to convert the fractional part
+//
+//      p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
+//
+// into decimal form, the fraction is repeatedly multiplied by 10 and the digits
+// d[-i] are extracted in order:
+//
+//      (10 * p2) div 2^-e = d[-1]
+//      (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
+//
+// The multiplication by 10 must not overflow. It is sufficient to choose
+//
+//      10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
+//
+// Since p2 = f mod 2^-e < 2^-e,
+//
+//      -e <= 60   or   e >= -60 := alpha
+
+constexpr int kAlpha = -60;
+constexpr int kGamma = -32;
+
+struct cached_power // c = f * 2^e ~= 10^k
+{
+    std::uint64_t f;
+    int e;
+    int k;
+};
+
+/*!
+For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
+power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
+satisfies (Definition 3.2 from [1])
+
+     alpha <= e_c + e + q <= gamma.
+*/
+inline cached_power get_cached_power_for_binary_exponent(int e)
+{
+    // Now
+    //
+    //      alpha <= e_c + e + q <= gamma                                    (1)
+    //      ==> f_c * 2^alpha <= c * 2^e * 2^q
+    //
+    // and since the c's are normalized, 2^(q-1) <= f_c,
+    //
+    //      ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
+    //      ==> 2^(alpha - e - 1) <= c
+    //
+    // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
+    //
+    //      k = ceil( log_10( 2^(alpha - e - 1) ) )
+    //        = ceil( (alpha - e - 1) * log_10(2) )
+    //
+    // From the paper:
+    // "In theory the result of the procedure could be wrong since c is rounded,
+    //  and the computation itself is approximated [...]. In practice, however,
+    //  this simple function is sufficient."
+    //
+    // For IEEE double precision floating-point numbers converted into
+    // normalized diyfp's w = f * 2^e, with q = 64,
+    //
+    //      e >= -1022      (min IEEE exponent)
+    //           -52        (p - 1)
+    //           -52        (p - 1, possibly normalize denormal IEEE numbers)
+    //           -11        (normalize the diyfp)
+    //         = -1137
+    //
+    // and
+    //
+    //      e <= +1023      (max IEEE exponent)
+    //           -52        (p - 1)
+    //           -11        (normalize the diyfp)
+    //         = 960
+    //
+    // This binary exponent range [-1137,960] results in a decimal exponent
+    // range [-307,324]. One does not need to store a cached power for each
+    // k in this range. For each such k it suffices to find a cached power
+    // such that the exponent of the product lies in [alpha,gamma].
+    // This implies that the difference of the decimal exponents of adjacent
+    // table entries must be less than or equal to
+    //
+    //      floor( (gamma - alpha) * log_10(2) ) = 8.
+    //
+    // (A smaller distance gamma-alpha would require a larger table.)
+
+    // NB:
+    // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
+
+    constexpr int kCachedPowersMinDecExp = -300;
+    constexpr int kCachedPowersDecStep = 8;
+
+    static constexpr std::array<cached_power, 79> kCachedPowers =
+    {
+        {
+            { 0xAB70FE17C79AC6CA, -1060, -300 },
+            { 0xFF77B1FCBEBCDC4F, -1034, -292 },
+            { 0xBE5691EF416BD60C, -1007, -284 },
+            { 0x8DD01FAD907FFC3C,  -980, -276 },
+            { 0xD3515C2831559A83,  -954, -268 },
+            { 0x9D71AC8FADA6C9B5,  -927, -260 },
+            { 0xEA9C227723EE8BCB,  -901, -252 },
+            { 0xAECC49914078536D,  -874, -244 },
+            { 0x823C12795DB6CE57,  -847, -236 },
+            { 0xC21094364DFB5637,  -821, -228 },
+            { 0x9096EA6F3848984F,  -794, -220 },
+            { 0xD77485CB25823AC7,  -768, -212 },
+            { 0xA086CFCD97BF97F4,  -741, -204 },
+            { 0xEF340A98172AACE5,  -715, -196 },
+            { 0xB23867FB2A35B28E,  -688, -188 },
+            { 0x84C8D4DFD2C63F3B,  -661, -180 },
+            { 0xC5DD44271AD3CDBA,  -635, -172 },
+            { 0x936B9FCEBB25C996,  -608, -164 },
+            { 0xDBAC6C247D62A584,  -582, -156 },
+            { 0xA3AB66580D5FDAF6,  -555, -148 },
+            { 0xF3E2F893DEC3F126,  -529, -140 },
+            { 0xB5B5ADA8AAFF80B8,  -502, -132 },
+            { 0x87625F056C7C4A8B,  -475, -124 },
+            { 0xC9BCFF6034C13053,  -449, -116 },
+            { 0x964E858C91BA2655,  -422, -108 },
+            { 0xDFF9772470297EBD,  -396, -100 },
+            { 0xA6DFBD9FB8E5B88F,  -369,  -92 },
+            { 0xF8A95FCF88747D94,  -343,  -84 },
+            { 0xB94470938FA89BCF,  -316,  -76 },
+            { 0x8A08F0F8BF0F156B,  -289,  -68 },
+            { 0xCDB02555653131B6,  -263,  -60 },
+            { 0x993FE2C6D07B7FAC,  -236,  -52 },
+            { 0xE45C10C42A2B3B06,  -210,  -44 },
+            { 0xAA242499697392D3,  -183,  -36 },
+            { 0xFD87B5F28300CA0E,  -157,  -28 },
+            { 0xBCE5086492111AEB,  -130,  -20 },
+            { 0x8CBCCC096F5088CC,  -103,  -12 },
+            { 0xD1B71758E219652C,   -77,   -4 },
+            { 0x9C40000000000000,   -50,    4 },
+            { 0xE8D4A51000000000,   -24,   12 },
+            { 0xAD78EBC5AC620000,     3,   20 },
+            { 0x813F3978F8940984,    30,   28 },
+            { 0xC097CE7BC90715B3,    56,   36 },
+            { 0x8F7E32CE7BEA5C70,    83,   44 },
+            { 0xD5D238A4ABE98068,   109,   52 },
+            { 0x9F4F2726179A2245,   136,   60 },
+            { 0xED63A231D4C4FB27,   162,   68 },
+            { 0xB0DE65388CC8ADA8,   189,   76 },
+            { 0x83C7088E1AAB65DB,   216,   84 },
+            { 0xC45D1DF942711D9A,   242,   92 },
+            { 0x924D692CA61BE758,   269,  100 },
+            { 0xDA01EE641A708DEA,   295,  108 },
+            { 0xA26DA3999AEF774A,   322,  116 },
+            { 0xF209787BB47D6B85,   348,  124 },
+            { 0xB454E4A179DD1877,   375,  132 },
+            { 0x865B86925B9BC5C2,   402,  140 },
+            { 0xC83553C5C8965D3D,   428,  148 },
+            { 0x952AB45CFA97A0B3,   455,  156 },
+            { 0xDE469FBD99A05FE3,   481,  164 },
+            { 0xA59BC234DB398C25,   508,  172 },
+            { 0xF6C69A72A3989F5C,   534,  180 },
+            { 0xB7DCBF5354E9BECE,   561,  188 },
+            { 0x88FCF317F22241E2,   588,  196 },
+            { 0xCC20CE9BD35C78A5,   614,  204 },
+            { 0x98165AF37B2153DF,   641,  212 },
+            { 0xE2A0B5DC971F303A,   667,  220 },
+            { 0xA8D9D1535CE3B396,   694,  228 },
+            { 0xFB9B7CD9A4A7443C,   720,  236 },
+            { 0xBB764C4CA7A44410,   747,  244 },
+            { 0x8BAB8EEFB6409C1A,   774,  252 },
+            { 0xD01FEF10A657842C,   800,  260 },
+            { 0x9B10A4E5E9913129,   827,  268 },
+            { 0xE7109BFBA19C0C9D,   853,  276 },
+            { 0xAC2820D9623BF429,   880,  284 },
+            { 0x80444B5E7AA7CF85,   907,  292 },
+            { 0xBF21E44003ACDD2D,   933,  300 },
+            { 0x8E679C2F5E44FF8F,   960,  308 },
+            { 0xD433179D9C8CB841,   986,  316 },
+            { 0x9E19DB92B4E31BA9,  1013,  324 },
+        }
+    };
+
+    // This computation gives exactly the same results for k as
+    //      k = ceil((kAlpha - e - 1) * 0.30102999566398114)
+    // for |e| <= 1500, but doesn't require floating-point operations.
+    // NB: log_10(2) ~= 78913 / 2^18
+    JSON_ASSERT(e >= -1500);
+    JSON_ASSERT(e <=  1500);
+    const int f = kAlpha - e - 1;
+    const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
+
+    const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / kCachedPowersDecStep;
+    JSON_ASSERT(index >= 0);
+    JSON_ASSERT(static_cast<std::size_t>(index) < kCachedPowers.size());
+
+    const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
+    JSON_ASSERT(kAlpha <= cached.e + e + 64);
+    JSON_ASSERT(kGamma >= cached.e + e + 64);
+
+    return cached;
+}
+
+/*!
+For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
+For n == 0, returns 1 and sets pow10 := 1.
+*/
+inline int find_largest_pow10(const std::uint32_t n, std::uint32_t& pow10)
+{
+    // LCOV_EXCL_START
+    if (n >= 1000000000)
+    {
+        pow10 = 1000000000;
+        return 10;
+    }
+    // LCOV_EXCL_STOP
+    if (n >= 100000000)
+    {
+        pow10 = 100000000;
+        return  9;
+    }
+    if (n >= 10000000)
+    {
+        pow10 = 10000000;
+        return  8;
+    }
+    if (n >= 1000000)
+    {
+        pow10 = 1000000;
+        return  7;
+    }
+    if (n >= 100000)
+    {
+        pow10 = 100000;
+        return  6;
+    }
+    if (n >= 10000)
+    {
+        pow10 = 10000;
+        return  5;
+    }
+    if (n >= 1000)
+    {
+        pow10 = 1000;
+        return  4;
+    }
+    if (n >= 100)
+    {
+        pow10 = 100;
+        return  3;
+    }
+    if (n >= 10)
+    {
+        pow10 = 10;
+        return  2;
+    }
+
+    pow10 = 1;
+    return 1;
+}
+
+inline void grisu2_round(char* buf, int len, std::uint64_t dist, std::uint64_t delta,
+                         std::uint64_t rest, std::uint64_t ten_k)
+{
+    JSON_ASSERT(len >= 1);
+    JSON_ASSERT(dist <= delta);
+    JSON_ASSERT(rest <= delta);
+    JSON_ASSERT(ten_k > 0);
+
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    //                                  ten_k
+    //                                <------>
+    //                                       <---- rest ---->
+    // --------------[------------------+----+--------------]--------------
+    //                                  w    V
+    //                                       = buf * 10^k
+    //
+    // ten_k represents a unit-in-the-last-place in the decimal representation
+    // stored in buf.
+    // Decrement buf by ten_k while this takes buf closer to w.
+
+    // The tests are written in this order to avoid overflow in unsigned
+    // integer arithmetic.
+
+    while (rest < dist
+            && delta - rest >= ten_k
+            && (rest + ten_k < dist || dist - rest > rest + ten_k - dist))
+    {
+        JSON_ASSERT(buf[len - 1] != '0');
+        buf[len - 1]--;
+        rest += ten_k;
+    }
+}
+
+/*!
+Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
+M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
+*/
+inline void grisu2_digit_gen(char* buffer, int& length, int& decimal_exponent,
+                             diyfp M_minus, diyfp w, diyfp M_plus)
+{
+    static_assert(kAlpha >= -60, "internal error");
+    static_assert(kGamma <= -32, "internal error");
+
+    // Generates the digits (and the exponent) of a decimal floating-point
+    // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
+    // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= gamma.
+    //
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    // Grisu2 generates the digits of M+ from left to right and stops as soon as
+    // V is in [M-,M+].
+
+    JSON_ASSERT(M_plus.e >= kAlpha);
+    JSON_ASSERT(M_plus.e <= kGamma);
+
+    std::uint64_t delta = diyfp::sub(M_plus, M_minus).f; // (significand of (M+ - M-), implicit exponent is e)
+    std::uint64_t dist  = diyfp::sub(M_plus, w      ).f; // (significand of (M+ - w ), implicit exponent is e)
+
+    // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
+    //
+    //      M+ = f * 2^e
+    //         = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
+    //         = ((p1        ) * 2^-e + (p2        )) * 2^e
+    //         = p1 + p2 * 2^e
+
+    const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
+
+    auto p1 = static_cast<std::uint32_t>(M_plus.f >> -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
+    std::uint64_t p2 = M_plus.f & (one.f - 1);                    // p2 = f mod 2^-e
+
+    // 1)
+    //
+    // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
+
+    JSON_ASSERT(p1 > 0);
+
+    std::uint32_t pow10{};
+    const int k = find_largest_pow10(p1, pow10);
+
+    //      10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
+    //
+    //      p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
+    //         = (d[k-1]         ) * 10^(k-1) + (p1 mod 10^(k-1))
+    //
+    //      M+ = p1                                             + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1))          + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
+    //         = d[k-1] * 10^(k-1) + (                         rest) * 2^e
+    //
+    // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
+    //
+    //      p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
+    //
+    // but stop as soon as
+    //
+    //      rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
+
+    int n = k;
+    while (n > 0)
+    {
+        // Invariants:
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)    (buffer = 0 for n = k)
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        //
+        const std::uint32_t d = p1 / pow10;  // d = p1 div 10^(n-1)
+        const std::uint32_t r = p1 % pow10;  // r = p1 mod 10^(n-1)
+        //
+        //      M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
+        //         = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
+        //
+        p1 = r;
+        n--;
+        //
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)
+        //      pow10 = 10^n
+        //
+
+        // Now check if enough digits have been generated.
+        // Compute
+        //
+        //      p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
+        //
+        // Note:
+        // Since rest and delta share the same exponent e, it suffices to
+        // compare the significands.
+        const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
+        if (rest <= delta)
+        {
+            // V = buffer * 10^n, with M- <= V <= M+.
+
+            decimal_exponent += n;
+
+            // We may now just stop. But instead look if the buffer could be
+            // decremented to bring V closer to w.
+            //
+            // pow10 = 10^n is now 1 ulp in the decimal representation V.
+            // The rounding procedure works with diyfp's with an implicit
+            // exponent of e.
+            //
+            //      10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
+            //
+            const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
+            grisu2_round(buffer, length, dist, delta, rest, ten_n);
+
+            return;
+        }
+
+        pow10 /= 10;
+        //
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        // Invariants restored.
+    }
+
+    // 2)
+    //
+    // The digits of the integral part have been generated:
+    //
+    //      M+ = d[k-1]...d[1]d[0] + p2 * 2^e
+    //         = buffer            + p2 * 2^e
+    //
+    // Now generate the digits of the fractional part p2 * 2^e.
+    //
+    // Note:
+    // No decimal point is generated: the exponent is adjusted instead.
+    //
+    // p2 actually represents the fraction
+    //
+    //      p2 * 2^e
+    //          = p2 / 2^-e
+    //          = d[-1] / 10^1 + d[-2] / 10^2 + ...
+    //
+    // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
+    //
+    //      p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
+    //                      + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
+    //
+    // using
+    //
+    //      10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
+    //                = (                   d) * 2^-e + (                   r)
+    //
+    // or
+    //      10^m * p2 * 2^e = d + r * 2^e
+    //
+    // i.e.
+    //
+    //      M+ = buffer + p2 * 2^e
+    //         = buffer + 10^-m * (d + r * 2^e)
+    //         = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
+    //
+    // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
+
+    JSON_ASSERT(p2 > delta);
+
+    int m = 0;
+    for (;;)
+    {
+        // Invariant:
+        //      M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) * 2^e
+        //         = buffer * 10^-m + 10^-m * (p2                                 ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (10 * p2)                   ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + (10*p2 mod 2^-e)) * 2^e
+        //
+        JSON_ASSERT(p2 <= (std::numeric_limits<std::uint64_t>::max)() / 10);
+        p2 *= 10;
+        const std::uint64_t d = p2 >> -one.e;     // d = (10 * p2) div 2^-e
+        const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
+        //
+        //      M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
+        //         = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        p2 = r;
+        m++;
+        //
+        //      M+ = buffer * 10^-m + 10^-m * p2 * 2^e
+        // Invariant restored.
+
+        // Check if enough digits have been generated.
+        //
+        //      10^-m * p2 * 2^e <= delta * 2^e
+        //              p2 * 2^e <= 10^m * delta * 2^e
+        //                    p2 <= 10^m * delta
+        delta *= 10;
+        dist  *= 10;
+        if (p2 <= delta)
+        {
+            break;
+        }
+    }
+
+    // V = buffer * 10^-m, with M- <= V <= M+.
+
+    decimal_exponent -= m;
+
+    // 1 ulp in the decimal representation is now 10^-m.
+    // Since delta and dist are now scaled by 10^m, we need to do the
+    // same with ulp in order to keep the units in sync.
+    //
+    //      10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
+    //
+    const std::uint64_t ten_m = one.f;
+    grisu2_round(buffer, length, dist, delta, p2, ten_m);
+
+    // By construction this algorithm generates the shortest possible decimal
+    // number (Loitsch, Theorem 6.2) which rounds back to w.
+    // For an input number of precision p, at least
+    //
+    //      N = 1 + ceil(p * log_10(2))
+    //
+    // decimal digits are sufficient to identify all binary floating-point
+    // numbers (Matula, "In-and-Out conversions").
+    // This implies that the algorithm does not produce more than N decimal
+    // digits.
+    //
+    //      N = 17 for p = 53 (IEEE double precision)
+    //      N = 9  for p = 24 (IEEE single precision)
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void grisu2(char* buf, int& len, int& decimal_exponent,
+                   diyfp m_minus, diyfp v, diyfp m_plus)
+{
+    JSON_ASSERT(m_plus.e == m_minus.e);
+    JSON_ASSERT(m_plus.e == v.e);
+
+    //  --------(-----------------------+-----------------------)--------    (A)
+    //          m-                      v                       m+
+    //
+    //  --------------------(-----------+-----------------------)--------    (B)
+    //                      m-          v                       m+
+    //
+    // First scale v (and m- and m+) such that the exponent is in the range
+    // [alpha, gamma].
+
+    const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
+
+    const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
+
+    // The exponent of the products is = v.e + c_minus_k.e + q and is in the range [alpha,gamma]
+    const diyfp w       = diyfp::mul(v,       c_minus_k);
+    const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
+    const diyfp w_plus  = diyfp::mul(m_plus,  c_minus_k);
+
+    //  ----(---+---)---------------(---+---)---------------(---+---)----
+    //          w-                      w                       w+
+    //          = c*m-                  = c*v                   = c*m+
+    //
+    // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
+    // w+ are now off by a small amount.
+    // In fact:
+    //
+    //      w - v * 10^k < 1 ulp
+    //
+    // To account for this inaccuracy, add resp. subtract 1 ulp.
+    //
+    //  --------+---[---------------(---+---)---------------]---+--------
+    //          w-  M-                  w                   M+  w+
+    //
+    // Now any number in [M-, M+] (bounds included) will round to w when input,
+    // regardless of how the input rounding algorithm breaks ties.
+    //
+    // And digit_gen generates the shortest possible such number in [M-, M+].
+    // Note that this does not mean that Grisu2 always generates the shortest
+    // possible number in the interval (m-, m+).
+    const diyfp M_minus(w_minus.f + 1, w_minus.e);
+    const diyfp M_plus (w_plus.f  - 1, w_plus.e );
+
+    decimal_exponent = -cached.k; // = -(-k) = k
+
+    grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
+                  "internal error: not enough precision");
+
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // If the neighbors (and boundaries) of 'value' are always computed for double-precision
+    // numbers, all float's can be recovered using strtod (and strtof). However, the resulting
+    // decimal representations are not exactly "short".
+    //
+    // The documentation for 'std::to_chars' (https://en.cppreference.com/w/cpp/utility/to_chars)
+    // says "value is converted to a string as if by std::sprintf in the default ("C") locale"
+    // and since sprintf promotes floats to doubles, I think this is exactly what 'std::to_chars'
+    // does.
+    // On the other hand, the documentation for 'std::to_chars' requires that "parsing the
+    // representation using the corresponding std::from_chars function recovers value exactly". That
+    // indicates that single precision floating-point numbers should be recovered using
+    // 'std::strtof'.
+    //
+    // NB: If the neighbors are computed for single-precision numbers, there is a single float
+    //     (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision
+    //     value is off by 1 ulp.
+#if 0
+    const boundaries w = compute_boundaries(static_cast<double>(value));
+#else
+    const boundaries w = compute_boundaries(value);
+#endif
+
+    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
+}
+
+/*!
+@brief appends a decimal representation of e to buf
+@return a pointer to the element following the exponent.
+@pre -1000 < e < 1000
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* append_exponent(char* buf, int e)
+{
+    JSON_ASSERT(e > -1000);
+    JSON_ASSERT(e <  1000);
+
+    if (e < 0)
+    {
+        e = -e;
+        *buf++ = '-';
+    }
+    else
+    {
+        *buf++ = '+';
+    }
+
+    auto k = static_cast<std::uint32_t>(e);
+    if (k < 10)
+    {
+        // Always print at least two digits in the exponent.
+        // This is for compatibility with printf("%g").
+        *buf++ = '0';
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else if (k < 100)
+    {
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else
+    {
+        *buf++ = static_cast<char>('0' + k / 100);
+        k %= 100;
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+
+    return buf;
+}
+
+/*!
+@brief prettify v = buf * 10^decimal_exponent
+
+If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
+notation. Otherwise it will be printed in exponential notation.
+
+@pre min_exp < 0
+@pre max_exp > 0
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* format_buffer(char* buf, int len, int decimal_exponent,
+                           int min_exp, int max_exp)
+{
+    JSON_ASSERT(min_exp < 0);
+    JSON_ASSERT(max_exp > 0);
+
+    const int k = len;
+    const int n = len + decimal_exponent;
+
+    // v = buf * 10^(n-k)
+    // k is the length of the buffer (number of decimal digits)
+    // n is the position of the decimal point relative to the start of the buffer.
+
+    if (k <= n && n <= max_exp)
+    {
+        // digits[000]
+        // len <= max_exp + 2
+
+        std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
+        // Make it look like a floating-point number (#362, #378)
+        buf[n + 0] = '.';
+        buf[n + 1] = '0';
+        return buf + (static_cast<size_t>(n) + 2);
+    }
+
+    if (0 < n && n <= max_exp)
+    {
+        // dig.its
+        // len <= max_digits10 + 1
+
+        JSON_ASSERT(k > n);
+
+        std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n, static_cast<size_t>(k) - static_cast<size_t>(n));
+        buf[n] = '.';
+        return buf + (static_cast<size_t>(k) + 1U);
+    }
+
+    if (min_exp < n && n <= 0)
+    {
+        // 0.[000]digits
+        // len <= 2 + (-min_exp - 1) + max_digits10
+
+        std::memmove(buf + (2 + static_cast<size_t>(-n)), buf, static_cast<size_t>(k));
+        buf[0] = '0';
+        buf[1] = '.';
+        std::memset(buf + 2, '0', static_cast<size_t>(-n));
+        return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
+    }
+
+    if (k == 1)
+    {
+        // dE+123
+        // len <= 1 + 5
+
+        buf += 1;
+    }
+    else
+    {
+        // d.igitsE+123
+        // len <= max_digits10 + 1 + 5
+
+        std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
+        buf[1] = '.';
+        buf += 1 + static_cast<size_t>(k);
+    }
+
+    *buf++ = 'e';
+    return append_exponent(buf, n - 1);
+}
+
+} // namespace dtoa_impl
+
+/*!
+@brief generates a decimal representation of the floating-point number value in [first, last).
+
+The format of the resulting decimal representation is similar to printf's %g
+format. Returns an iterator pointing past-the-end of the decimal representation.
+
+@note The input number must be finite, i.e. NaN's and Inf's are not supported.
+@note The buffer must be large enough.
+@note The result is NOT null-terminated.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* to_chars(char* first, const char* last, FloatType value)
+{
+    static_cast<void>(last); // maybe unused - fix warning
+    JSON_ASSERT(std::isfinite(value));
+
+    // Use signbit(value) instead of (value < 0) since signbit works for -0.
+    if (std::signbit(value))
+    {
+        value = -value;
+        *first++ = '-';
+    }
+
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+    if (value == 0) // +-0
+    {
+        *first++ = '0';
+        // Make it look like a floating-point number (#362, #378)
+        *first++ = '.';
+        *first++ = '0';
+        return first;
+    }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e. the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    dtoa_impl::grisu2(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+} // namespace detail
+} // namespace nlohmann
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+namespace nlohmann
+{
+namespace detail
+{
+///////////////////
+// serialization //
+///////////////////
+
+/// how to treat decoding errors
+enum class error_handler_t
+{
+    strict,  ///< throw a type_error exception in case of invalid UTF-8
+    replace, ///< replace invalid UTF-8 sequences with U+FFFD
+    ignore   ///< ignore invalid UTF-8 sequences
+};
+
+template<typename BasicJsonType>
+class serializer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using binary_char_t = typename BasicJsonType::binary_t::value_type;
+    static constexpr std::uint8_t UTF8_ACCEPT = 0;
+    static constexpr std::uint8_t UTF8_REJECT = 1;
+
+  public:
+    /*!
+    @param[in] s  output stream to serialize to
+    @param[in] ichar  indentation character to use
+    @param[in] error_handler_  how to react on decoding errors
+    */
+    serializer(output_adapter_t<char> s, const char ichar,
+               error_handler_t error_handler_ = error_handler_t::strict)
+        : o(std::move(s))
+        , loc(std::localeconv())
+        , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
+        , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
+        , indent_char(ichar)
+        , indent_string(512, indent_char)
+        , error_handler(error_handler_)
+    {}
+
+    // delete because of pointer members
+    serializer(const serializer&) = delete;
+    serializer& operator=(const serializer&) = delete;
+    serializer(serializer&&) = delete;
+    serializer& operator=(serializer&&) = delete;
+    ~serializer() = default;
+
+    /*!
+    @brief internal implementation of the serialization function
+
+    This function is called by the public member function dump and organizes
+    the serialization internally. The indentation level is propagated as
+    additional parameter. In case of arrays and objects, the function is
+    called recursively.
+
+    - strings and object keys are escaped using `escape_string()`
+    - integer numbers are converted implicitly via `operator<<`
+    - floating-point numbers are converted to a string using `"%g"` format
+    - binary values are serialized as objects containing the subtype and the
+      byte array
+
+    @param[in] val               value to serialize
+    @param[in] pretty_print      whether the output shall be pretty-printed
+    @param[in] ensure_ascii If @a ensure_ascii is true, all non-ASCII characters
+    in the output are escaped with `\uXXXX` sequences, and the result consists
+    of ASCII characters only.
+    @param[in] indent_step       the indent level
+    @param[in] current_indent    the current indent level (only used internally)
+    */
+    void dump(const BasicJsonType& val,
+              const bool pretty_print,
+              const bool ensure_ascii,
+              const unsigned int indent_step,
+              const unsigned int current_indent = 0)
+    {
+        switch (val.m_type)
+        {
+            case value_t::object:
+            {
+                if (val.m_value.object->empty())
+                {
+                    o->write_characters("{}", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    auto i = val.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\": ", 3);
+                        dump(i->second, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\": ", 3);
+                    dump(i->second, true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_character('{');
+
+                    // first n-1 elements
+                    auto i = val.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\":", 2);
+                        dump(i->second, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\":", 2);
+                    dump(i->second, false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character('}');
+                }
+
+                return;
+            }
+
+            case value_t::array:
+            {
+                if (val.m_value.array->empty())
+                {
+                    o->write_characters("[]", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("[\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    for (auto i = val.m_value.array->cbegin();
+                            i != val.m_value.array->cend() - 1; ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        dump(*i, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_value.array->empty());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    dump(val.m_value.array->back(), true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character(']');
+                }
+                else
+                {
+                    o->write_character('[');
+
+                    // first n-1 elements
+                    for (auto i = val.m_value.array->cbegin();
+                            i != val.m_value.array->cend() - 1; ++i)
+                    {
+                        dump(*i, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_value.array->empty());
+                    dump(val.m_value.array->back(), false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character(']');
+                }
+
+                return;
+            }
+
+            case value_t::string:
+            {
+                o->write_character('\"');
+                dump_escaped(*val.m_value.string, ensure_ascii);
+                o->write_character('\"');
+                return;
+            }
+
+            case value_t::binary:
+            {
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"bytes\": [", 10);
+
+                    if (!val.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_value.binary->cbegin();
+                                i != val.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_characters(", ", 2);
+                        }
+                        dump_integer(val.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\n", 3);
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"subtype\": ", 11);
+                    if (val.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_value.binary->subtype());
+                    }
+                    else
+                    {
+                        o->write_characters("null", 4);
+                    }
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_characters("{\"bytes\":[", 10);
+
+                    if (!val.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_value.binary->cbegin();
+                                i != val.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_character(',');
+                        }
+                        dump_integer(val.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\"subtype\":", 12);
+                    if (val.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_value.binary->subtype());
+                        o->write_character('}');
+                    }
+                    else
+                    {
+                        o->write_characters("null}", 5);
+                    }
+                }
+                return;
+            }
+
+            case value_t::boolean:
+            {
+                if (val.m_value.boolean)
+                {
+                    o->write_characters("true", 4);
+                }
+                else
+                {
+                    o->write_characters("false", 5);
+                }
+                return;
+            }
+
+            case value_t::number_integer:
+            {
+                dump_integer(val.m_value.number_integer);
+                return;
+            }
+
+            case value_t::number_unsigned:
+            {
+                dump_integer(val.m_value.number_unsigned);
+                return;
+            }
+
+            case value_t::number_float:
+            {
+                dump_float(val.m_value.number_float);
+                return;
+            }
+
+            case value_t::discarded:
+            {
+                o->write_characters("<discarded>", 11);
+                return;
+            }
+
+            case value_t::null:
+            {
+                o->write_characters("null", 4);
+                return;
+            }
+
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief dump escaped string
+
+    Escape a string by replacing certain special characters by a sequence of an
+    escape character (backslash) and another character and other control
+    characters by a sequence of "\u" followed by a four-digit hex
+    representation. The escaped string is written to output stream @a o.
+
+    @param[in] s  the string to escape
+    @param[in] ensure_ascii  whether to escape non-ASCII characters with
+                             \uXXXX sequences
+
+    @complexity Linear in the length of string @a s.
+    */
+    void dump_escaped(const string_t& s, const bool ensure_ascii)
+    {
+        std::uint32_t codepoint{};
+        std::uint8_t state = UTF8_ACCEPT;
+        std::size_t bytes = 0;  // number of bytes written to string_buffer
+
+        // number of bytes written at the point of the last valid byte
+        std::size_t bytes_after_last_accept = 0;
+        std::size_t undumped_chars = 0;
+
+        for (std::size_t i = 0; i < s.size(); ++i)
+        {
+            const auto byte = static_cast<std::uint8_t>(s[i]);
+
+            switch (decode(state, codepoint, byte))
+            {
+                case UTF8_ACCEPT:  // decode found a new code point
+                {
+                    switch (codepoint)
+                    {
+                        case 0x08: // backspace
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'b';
+                            break;
+                        }
+
+                        case 0x09: // horizontal tab
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 't';
+                            break;
+                        }
+
+                        case 0x0A: // newline
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'n';
+                            break;
+                        }
+
+                        case 0x0C: // formfeed
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'f';
+                            break;
+                        }
+
+                        case 0x0D: // carriage return
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'r';
+                            break;
+                        }
+
+                        case 0x22: // quotation mark
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\"';
+                            break;
+                        }
+
+                        case 0x5C: // reverse solidus
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\\';
+                            break;
+                        }
+
+                        default:
+                        {
+                            // escape control characters (0x00..0x1F) or, if
+                            // ensure_ascii parameter is used, non-ASCII characters
+                            if ((codepoint <= 0x1F) || (ensure_ascii && (codepoint >= 0x7F)))
+                            {
+                                if (codepoint <= 0xFFFF)
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 7, "\\u%04x",
+                                                                      static_cast<std::uint16_t>(codepoint)));
+                                    bytes += 6;
+                                }
+                                else
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 13, "\\u%04x\\u%04x",
+                                                                      static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)),
+                                                                      static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu))));
+                                    bytes += 12;
+                                }
+                            }
+                            else
+                            {
+                                // copy byte to buffer (all previous bytes
+                                // been copied have in default case above)
+                                string_buffer[bytes++] = s[i];
+                            }
+                            break;
+                        }
+                    }
+
+                    // write buffer and reset index; there must be 13 bytes
+                    // left, as this is the maximal number of bytes to be
+                    // written ("\uxxxx\uxxxx\0") for one code point
+                    if (string_buffer.size() - bytes < 13)
+                    {
+                        o->write_characters(string_buffer.data(), bytes);
+                        bytes = 0;
+                    }
+
+                    // remember the byte position of this accept
+                    bytes_after_last_accept = bytes;
+                    undumped_chars = 0;
+                    break;
+                }
+
+                case UTF8_REJECT:  // decode found invalid UTF-8 byte
+                {
+                    switch (error_handler)
+                    {
+                        case error_handler_t::strict:
+                        {
+                            std::stringstream ss;
+                            ss << std::uppercase << std::setfill('0') << std::setw(2) << std::hex << (byte | 0);
+                            JSON_THROW(type_error::create(316, "invalid UTF-8 byte at index " + std::to_string(i) + ": 0x" + ss.str(), BasicJsonType()));
+                        }
+
+                        case error_handler_t::ignore:
+                        case error_handler_t::replace:
+                        {
+                            // in case we saw this character the first time, we
+                            // would like to read it again, because the byte
+                            // may be OK for itself, but just not OK for the
+                            // previous sequence
+                            if (undumped_chars > 0)
+                            {
+                                --i;
+                            }
+
+                            // reset length buffer to the last accepted index;
+                            // thus removing/ignoring the invalid characters
+                            bytes = bytes_after_last_accept;
+
+                            if (error_handler == error_handler_t::replace)
+                            {
+                                // add a replacement character
+                                if (ensure_ascii)
+                                {
+                                    string_buffer[bytes++] = '\\';
+                                    string_buffer[bytes++] = 'u';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'd';
+                                }
+                                else
+                                {
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xEF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBD');
+                                }
+
+                                // write buffer and reset index; there must be 13 bytes
+                                // left, as this is the maximal number of bytes to be
+                                // written ("\uxxxx\uxxxx\0") for one code point
+                                if (string_buffer.size() - bytes < 13)
+                                {
+                                    o->write_characters(string_buffer.data(), bytes);
+                                    bytes = 0;
+                                }
+
+                                bytes_after_last_accept = bytes;
+                            }
+
+                            undumped_chars = 0;
+
+                            // continue processing the string
+                            state = UTF8_ACCEPT;
+                            break;
+                        }
+
+                        default:            // LCOV_EXCL_LINE
+                            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+
+                default:  // decode found yet incomplete multi-byte code point
+                {
+                    if (!ensure_ascii)
+                    {
+                        // code point will not be escaped - copy byte to buffer
+                        string_buffer[bytes++] = s[i];
+                    }
+                    ++undumped_chars;
+                    break;
+                }
+            }
+        }
+
+        // we finished processing the string
+        if (JSON_HEDLEY_LIKELY(state == UTF8_ACCEPT))
+        {
+            // write buffer
+            if (bytes > 0)
+            {
+                o->write_characters(string_buffer.data(), bytes);
+            }
+        }
+        else
+        {
+            // we finish reading, but do not accept: string was incomplete
+            switch (error_handler)
+            {
+                case error_handler_t::strict:
+                {
+                    std::stringstream ss;
+                    ss << std::uppercase << std::setfill('0') << std::setw(2) << std::hex << (static_cast<std::uint8_t>(s.back()) | 0);
+                    JSON_THROW(type_error::create(316, "incomplete UTF-8 string; last byte: 0x" + ss.str(), BasicJsonType()));
+                }
+
+                case error_handler_t::ignore:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    break;
+                }
+
+                case error_handler_t::replace:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    // add a replacement character
+                    if (ensure_ascii)
+                    {
+                        o->write_characters("\\ufffd", 6);
+                    }
+                    else
+                    {
+                        o->write_characters("\xEF\xBF\xBD", 3);
+                    }
+                    break;
+                }
+
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        }
+    }
+
+  private:
+    /*!
+    @brief count digits
+
+    Count the number of decimal (base 10) digits for an input unsigned integer.
+
+    @param[in] x  unsigned integer number to count its digits
+    @return    number of decimal digits
+    */
+    inline unsigned int count_digits(number_unsigned_t x) noexcept
+    {
+        unsigned int n_digits = 1;
+        for (;;)
+        {
+            if (x < 10)
+            {
+                return n_digits;
+            }
+            if (x < 100)
+            {
+                return n_digits + 1;
+            }
+            if (x < 1000)
+            {
+                return n_digits + 2;
+            }
+            if (x < 10000)
+            {
+                return n_digits + 3;
+            }
+            x = x / 10000u;
+            n_digits += 4;
+        }
+    }
+
+    // templates to avoid warnings about useless casts
+    template <typename NumberType, enable_if_t<std::is_signed<NumberType>::value, int> = 0>
+    bool is_negative_number(NumberType x)
+    {
+        return x < 0;
+    }
+
+    template < typename NumberType, enable_if_t <std::is_unsigned<NumberType>::value, int > = 0 >
+    bool is_negative_number(NumberType /*unused*/)
+    {
+        return false;
+    }
+
+    /*!
+    @brief dump an integer
+
+    Dump a given integer to output stream @a o. Works internally with
+    @a number_buffer.
+
+    @param[in] x  integer number (signed or unsigned) to dump
+    @tparam NumberType either @a number_integer_t or @a number_unsigned_t
+    */
+    template < typename NumberType, detail::enable_if_t <
+                   std::is_integral<NumberType>::value ||
+                   std::is_same<NumberType, number_unsigned_t>::value ||
+                   std::is_same<NumberType, number_integer_t>::value ||
+                   std::is_same<NumberType, binary_char_t>::value,
+                   int > = 0 >
+    void dump_integer(NumberType x)
+    {
+        static constexpr std::array<std::array<char, 2>, 100> digits_to_99
+        {
+            {
+                {{'0', '0'}}, {{'0', '1'}}, {{'0', '2'}}, {{'0', '3'}}, {{'0', '4'}}, {{'0', '5'}}, {{'0', '6'}}, {{'0', '7'}}, {{'0', '8'}}, {{'0', '9'}},
+                {{'1', '0'}}, {{'1', '1'}}, {{'1', '2'}}, {{'1', '3'}}, {{'1', '4'}}, {{'1', '5'}}, {{'1', '6'}}, {{'1', '7'}}, {{'1', '8'}}, {{'1', '9'}},
+                {{'2', '0'}}, {{'2', '1'}}, {{'2', '2'}}, {{'2', '3'}}, {{'2', '4'}}, {{'2', '5'}}, {{'2', '6'}}, {{'2', '7'}}, {{'2', '8'}}, {{'2', '9'}},
+                {{'3', '0'}}, {{'3', '1'}}, {{'3', '2'}}, {{'3', '3'}}, {{'3', '4'}}, {{'3', '5'}}, {{'3', '6'}}, {{'3', '7'}}, {{'3', '8'}}, {{'3', '9'}},
+                {{'4', '0'}}, {{'4', '1'}}, {{'4', '2'}}, {{'4', '3'}}, {{'4', '4'}}, {{'4', '5'}}, {{'4', '6'}}, {{'4', '7'}}, {{'4', '8'}}, {{'4', '9'}},
+                {{'5', '0'}}, {{'5', '1'}}, {{'5', '2'}}, {{'5', '3'}}, {{'5', '4'}}, {{'5', '5'}}, {{'5', '6'}}, {{'5', '7'}}, {{'5', '8'}}, {{'5', '9'}},
+                {{'6', '0'}}, {{'6', '1'}}, {{'6', '2'}}, {{'6', '3'}}, {{'6', '4'}}, {{'6', '5'}}, {{'6', '6'}}, {{'6', '7'}}, {{'6', '8'}}, {{'6', '9'}},
+                {{'7', '0'}}, {{'7', '1'}}, {{'7', '2'}}, {{'7', '3'}}, {{'7', '4'}}, {{'7', '5'}}, {{'7', '6'}}, {{'7', '7'}}, {{'7', '8'}}, {{'7', '9'}},
+                {{'8', '0'}}, {{'8', '1'}}, {{'8', '2'}}, {{'8', '3'}}, {{'8', '4'}}, {{'8', '5'}}, {{'8', '6'}}, {{'8', '7'}}, {{'8', '8'}}, {{'8', '9'}},
+                {{'9', '0'}}, {{'9', '1'}}, {{'9', '2'}}, {{'9', '3'}}, {{'9', '4'}}, {{'9', '5'}}, {{'9', '6'}}, {{'9', '7'}}, {{'9', '8'}}, {{'9', '9'}},
+            }
+        };
+
+        // special case for "0"
+        if (x == 0)
+        {
+            o->write_character('0');
+            return;
+        }
+
+        // use a pointer to fill the buffer
+        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+
+        number_unsigned_t abs_value;
+
+        unsigned int n_chars{};
+
+        if (is_negative_number(x))
+        {
+            *buffer_ptr = '-';
+            abs_value = remove_sign(static_cast<number_integer_t>(x));
+
+            // account one more byte for the minus sign
+            n_chars = 1 + count_digits(abs_value);
+        }
+        else
+        {
+            abs_value = static_cast<number_unsigned_t>(x);
+            n_chars = count_digits(abs_value);
+        }
+
+        // spare 1 byte for '\0'
+        JSON_ASSERT(n_chars < number_buffer.size() - 1);
+
+        // jump to the end to generate the string from backward,
+        // so we later avoid reversing the result
+        buffer_ptr += n_chars;
+
+        // Fast int2ascii implementation inspired by "Fastware" talk by Andrei Alexandrescu
+        // See: https://www.youtube.com/watch?v=o4-CwDo2zpg
+        while (abs_value >= 100)
+        {
+            const auto digits_index = static_cast<unsigned>((abs_value % 100));
+            abs_value /= 100;
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+
+        if (abs_value >= 10)
+        {
+            const auto digits_index = static_cast<unsigned>(abs_value);
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+        else
+        {
+            *(--buffer_ptr) = static_cast<char>('0' + abs_value);
+        }
+
+        o->write_characters(number_buffer.data(), n_chars);
+    }
+
+    /*!
+    @brief dump a floating-point number
+
+    Dump a given floating-point number to output stream @a o. Works internally
+    with @a number_buffer.
+
+    @param[in] x  floating-point number to dump
+    */
+    void dump_float(number_float_t x)
+    {
+        // NaN / inf
+        if (!std::isfinite(x))
+        {
+            o->write_characters("null", 4);
+            return;
+        }
+
+        // If number_float_t is an IEEE-754 single or double precision number,
+        // use the Grisu2 algorithm to produce short numbers which are
+        // guaranteed to round-trip, using strtof and strtod, resp.
+        //
+        // NB: The test below works if <long double> == <double>.
+        static constexpr bool is_ieee_single_or_double
+            = (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 24 && std::numeric_limits<number_float_t>::max_exponent == 128) ||
+              (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 53 && std::numeric_limits<number_float_t>::max_exponent == 1024);
+
+        dump_float(x, std::integral_constant<bool, is_ieee_single_or_double>());
+    }
+
+    void dump_float(number_float_t x, std::true_type /*is_ieee_single_or_double*/)
+    {
+        auto* begin = number_buffer.data();
+        auto* end = ::nlohmann::detail::to_chars(begin, begin + number_buffer.size(), x);
+
+        o->write_characters(begin, static_cast<size_t>(end - begin));
+    }
+
+    void dump_float(number_float_t x, std::false_type /*is_ieee_single_or_double*/)
+    {
+        // get number of digits for a float -> text -> float round-trip
+        static constexpr auto d = std::numeric_limits<number_float_t>::max_digits10;
+
+        // the actual conversion
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        std::ptrdiff_t len = (std::snprintf)(number_buffer.data(), number_buffer.size(), "%.*g", d, x);
+
+        // negative value indicates an error
+        JSON_ASSERT(len > 0);
+        // check if buffer was large enough
+        JSON_ASSERT(static_cast<std::size_t>(len) < number_buffer.size());
+
+        // erase thousands separator
+        if (thousands_sep != '\0')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::remove returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto end = std::remove(number_buffer.begin(), number_buffer.begin() + len, thousands_sep);
+            std::fill(end, number_buffer.end(), '\0');
+            JSON_ASSERT((end - number_buffer.begin()) <= len);
+            len = (end - number_buffer.begin());
+        }
+
+        // convert decimal point to '.'
+        if (decimal_point != '\0' && decimal_point != '.')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::find returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto dec_pos = std::find(number_buffer.begin(), number_buffer.end(), decimal_point);
+            if (dec_pos != number_buffer.end())
+            {
+                *dec_pos = '.';
+            }
+        }
+
+        o->write_characters(number_buffer.data(), static_cast<std::size_t>(len));
+
+        // determine if we need to append ".0"
+        const bool value_is_int_like =
+            std::none_of(number_buffer.begin(), number_buffer.begin() + len + 1,
+                         [](char c)
+        {
+            return c == '.' || c == 'e';
+        });
+
+        if (value_is_int_like)
+        {
+            o->write_characters(".0", 2);
+        }
+    }
+
+    /*!
+    @brief check whether a string is UTF-8 encoded
+
+    The function checks each byte of a string whether it is UTF-8 encoded. The
+    result of the check is stored in the @a state parameter. The function must
+    be called initially with state 0 (accept). State 1 means the string must
+    be rejected, because the current byte is not allowed. If the string is
+    completely processed, but the state is non-zero, the string ended
+    prematurely; that is, the last byte indicated more bytes should has
+    followed.
+
+    @param[in,out] state  the state of the decoding
+    @param[in,out] codep  codepoint (valid only if resulting state is UTF8_ACCEPT)
+    @param[in] byte       next byte to decode
+    @return               new state
+
+    @note The function has been edited: a std::array is used.
+
+    @copyright Copyright (c) 2008-2009 Bjoern Hoehrmann <bjoern@hoehrmann.de>
+    @sa http://bjoern.hoehrmann.de/utf-8/decoder/dfa/
+    */
+    static std::uint8_t decode(std::uint8_t& state, std::uint32_t& codep, const std::uint8_t byte) noexcept
+    {
+        static const std::array<std::uint8_t, 400> utf8d =
+        {
+            {
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 00..1F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 20..3F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 40..5F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 60..7F
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, // 80..9F
+                7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // A0..BF
+                8, 8, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // C0..DF
+                0xA, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x4, 0x3, 0x3, // E0..EF
+                0xB, 0x6, 0x6, 0x6, 0x5, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, // F0..FF
+                0x0, 0x1, 0x2, 0x3, 0x5, 0x8, 0x7, 0x1, 0x1, 0x1, 0x4, 0x6, 0x1, 0x1, 0x1, 0x1, // s0..s0
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, // s1..s2
+                1, 2, 1, 1, 1, 1, 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, // s3..s4
+                1, 2, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, // s5..s6
+                1, 3, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 // s7..s8
+            }
+        };
+
+        JSON_ASSERT(byte < utf8d.size());
+        const std::uint8_t type = utf8d[byte];
+
+        codep = (state != UTF8_ACCEPT)
+                ? (byte & 0x3fu) | (codep << 6u)
+                : (0xFFu >> type) & (byte);
+
+        std::size_t index = 256u + static_cast<size_t>(state) * 16u + static_cast<size_t>(type);
+        JSON_ASSERT(index < 400);
+        state = utf8d[index];
+        return state;
+    }
+
+    /*
+     * Overload to make the compiler happy while it is instantiating
+     * dump_integer for number_unsigned_t.
+     * Must never be called.
+     */
+    number_unsigned_t remove_sign(number_unsigned_t x)
+    {
+        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        return x; // LCOV_EXCL_LINE
+    }
+
+    /*
+     * Helper function for dump_integer
+     *
+     * This function takes a negative signed integer and returns its absolute
+     * value as unsigned integer. The plus/minus shuffling is necessary as we can
+     * not directly remove the sign of an arbitrary signed integer as the
+     * absolute values of INT_MIN and INT_MAX are usually not the same. See
+     * #1708 for details.
+     */
+    inline number_unsigned_t remove_sign(number_integer_t x) noexcept
+    {
+        JSON_ASSERT(x < 0 && x < (std::numeric_limits<number_integer_t>::max)()); // NOLINT(misc-redundant-expression)
+        return static_cast<number_unsigned_t>(-(x + 1)) + 1;
+    }
+
+  private:
+    /// the output of the serializer
+    output_adapter_t<char> o = nullptr;
+
+    /// a (hopefully) large enough character buffer
+    std::array<char, 64> number_buffer{{}};
+
+    /// the locale
+    const std::lconv* loc = nullptr;
+    /// the locale's thousand separator character
+    const char thousands_sep = '\0';
+    /// the locale's decimal point character
+    const char decimal_point = '\0';
+
+    /// string buffer
+    std::array<char, 512> string_buffer{{}};
+
+    /// the indentation character
+    const char indent_char;
+    /// the indentation string
+    string_t indent_string;
+
+    /// error_handler how to react on decoding errors
+    const error_handler_t error_handler;
+};
+}  // namespace detail
+}  // namespace nlohmann
+
+// #include <nlohmann/detail/value_t.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+
+// #include <nlohmann/ordered_map.hpp>
+
+
+#include <functional> // less
+#include <initializer_list> // initializer_list
+#include <iterator> // input_iterator_tag, iterator_traits
+#include <memory> // allocator
+#include <stdexcept> // for out_of_range
+#include <type_traits> // enable_if, is_convertible
+#include <utility> // pair
+#include <vector> // vector
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+namespace nlohmann
+{
+
+/// ordered_map: a minimal map-like container that preserves insertion order
+/// for use within nlohmann::basic_json<ordered_map>
+template <class Key, class T, class IgnoredLess = std::less<Key>,
+          class Allocator = std::allocator<std::pair<const Key, T>>>
+                  struct ordered_map : std::vector<std::pair<const Key, T>, Allocator>
+{
+    using key_type = Key;
+    using mapped_type = T;
+    using Container = std::vector<std::pair<const Key, T>, Allocator>;
+    using iterator = typename Container::iterator;
+    using const_iterator = typename Container::const_iterator;
+    using size_type = typename Container::size_type;
+    using value_type = typename Container::value_type;
+
+    // Explicit constructors instead of `using Container::Container`
+    // otherwise older compilers choke on it (GCC <= 5.5, xcode <= 9.4)
+    ordered_map(const Allocator& alloc = Allocator()) : Container{alloc} {}
+    template <class It>
+    ordered_map(It first, It last, const Allocator& alloc = Allocator())
+        : Container{first, last, alloc} {}
+    ordered_map(std::initializer_list<T> init, const Allocator& alloc = Allocator() )
+        : Container{init, alloc} {}
+
+    std::pair<iterator, bool> emplace(const key_type& key, T&& t)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return {it, false};
+            }
+        }
+        Container::emplace_back(key, t);
+        return {--this->end(), true};
+    }
+
+    T& operator[](const Key& key)
+    {
+        return emplace(key, T{}).first->second;
+    }
+
+    const T& operator[](const Key& key) const
+    {
+        return at(key);
+    }
+
+    T& at(const Key& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    const T& at(const Key& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    size_type erase(const Key& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                // Since we cannot move const Keys, re-construct them in place
+                for (auto next = it; ++next != this->end(); ++it)
+                {
+                    it->~value_type(); // Destroy but keep allocation
+                    new (&*it) value_type{std::move(*next)};
+                }
+                Container::pop_back();
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator erase(iterator pos)
+    {
+        return erase(pos, std::next(pos));
+    }
+
+    iterator erase(iterator first, iterator last)
+    {
+        const auto elements_affected = std::distance(first, last);
+        const auto offset = std::distance(Container::begin(), first);
+
+        // This is the start situation. We need to delete elements_affected
+        // elements (3 in this example: e, f, g), and need to return an
+        // iterator past the last deleted element (h in this example).
+        // Note that offset is the distance from the start of the vector
+        // to first. We will need this later.
+
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // Since we cannot move const Keys, we re-construct them in place.
+        // We start at first and re-construct (viz. copy) the elements from
+        // the back of the vector. Example for first iteration:
+
+        //               ,--------.
+        //               v        |   destroy e and re-construct with h
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //               it       it + elements_affected
+
+        for (auto it = first; std::next(it, elements_affected) != Container::end(); ++it)
+        {
+            it->~value_type(); // destroy but keep allocation
+            new (&*it) value_type{std::move(*std::next(it, elements_affected))}; // "move" next element to it
+        }
+
+        // [ a, b, c, d, h, i, j, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // remove the unneeded elements at the end of the vector
+        Container::resize(this->size() - static_cast<size_type>(elements_affected));
+
+        // [ a, b, c, d, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // first is now pointing past the last deleted element, but we cannot
+        // use this iterator, because it may have been invalidated by the
+        // resize call. Instead, we can return begin() + offset.
+        return Container::begin() + offset;
+    }
+
+    size_type count(const Key& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator find(const Key& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    const_iterator find(const Key& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == key)
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    std::pair<iterator, bool> insert( value_type&& value )
+    {
+        return emplace(value.first, std::move(value.second));
+    }
+
+    std::pair<iterator, bool> insert( const value_type& value )
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (it->first == value.first)
+            {
+                return {it, false};
+            }
+        }
+        Container::push_back(value);
+        return {--this->end(), true};
+    }
+
+    template<typename InputIt>
+    using require_input_iter = typename std::enable_if<std::is_convertible<typename std::iterator_traits<InputIt>::iterator_category,
+            std::input_iterator_tag>::value>::type;
+
+    template<typename InputIt, typename = require_input_iter<InputIt>>
+    void insert(InputIt first, InputIt last)
+    {
+        for (auto it = first; it != last; ++it)
+        {
+            insert(*it);
+        }
+    }
+};
+
+}  // namespace nlohmann
+
+
+#if defined(JSON_HAS_CPP_17)
+    #include <string_view>
+#endif
+
+/*!
+@brief namespace for Niels Lohmann
+@see https://github.com/nlohmann
+@since version 1.0.0
+*/
+namespace nlohmann
+{
+
+/*!
+@brief a class to store JSON values
+
+@internal
+@invariant The member variables @a m_value and @a m_type has the following
+relationship:
+- If `m_type == value_t::object`, then `m_value.object != nullptr`.
+- If `m_type == value_t::array`, then `m_value.array != nullptr`.
+- If `m_type == value_t::string`, then `m_value.string != nullptr`.
+The invariants are checked by member function assert_invariant().
+
+@note ObjectType trick from https://stackoverflow.com/a/9860911
+@endinternal
+
+@since version 1.0.0
+
+@nosubgrouping
+*/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+{
+  private:
+    template<detail::value_t> friend struct detail::external_constructor;
+    friend ::nlohmann::json_pointer<basic_json>;
+
+    template<typename BasicJsonType, typename InputType>
+    friend class ::nlohmann::detail::parser;
+    friend ::nlohmann::detail::serializer<basic_json>;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::iter_impl;
+    template<typename BasicJsonType, typename CharType>
+    friend class ::nlohmann::detail::binary_writer;
+    template<typename BasicJsonType, typename InputType, typename SAX>
+    friend class ::nlohmann::detail::binary_reader;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_parser;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_callback_parser;
+    friend class ::nlohmann::detail::exception;
+
+    /// workaround type for MSVC
+    using basic_json_t = NLOHMANN_BASIC_JSON_TPL;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // convenience aliases for types residing in namespace detail;
+    using lexer = ::nlohmann::detail::lexer_base<basic_json>;
+
+    template<typename InputAdapterType>
+    static ::nlohmann::detail::parser<basic_json, InputAdapterType> parser(
+        InputAdapterType adapter,
+        detail::parser_callback_t<basic_json>cb = nullptr,
+        const bool allow_exceptions = true,
+        const bool ignore_comments = false
+                                 )
+    {
+        return ::nlohmann::detail::parser<basic_json, InputAdapterType>(std::move(adapter),
+                std::move(cb), allow_exceptions, ignore_comments);
+    }
+
+  private:
+    using primitive_iterator_t = ::nlohmann::detail::primitive_iterator_t;
+    template<typename BasicJsonType>
+    using internal_iterator = ::nlohmann::detail::internal_iterator<BasicJsonType>;
+    template<typename BasicJsonType>
+    using iter_impl = ::nlohmann::detail::iter_impl<BasicJsonType>;
+    template<typename Iterator>
+    using iteration_proxy = ::nlohmann::detail::iteration_proxy<Iterator>;
+    template<typename Base> using json_reverse_iterator = ::nlohmann::detail::json_reverse_iterator<Base>;
+
+    template<typename CharType>
+    using output_adapter_t = ::nlohmann::detail::output_adapter_t<CharType>;
+
+    template<typename InputType>
+    using binary_reader = ::nlohmann::detail::binary_reader<basic_json, InputType>;
+    template<typename CharType> using binary_writer = ::nlohmann::detail::binary_writer<basic_json, CharType>;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    using serializer = ::nlohmann::detail::serializer<basic_json>;
+
+  public:
+    using value_t = detail::value_t;
+    /// JSON Pointer, see @ref nlohmann::json_pointer
+    using json_pointer = ::nlohmann::json_pointer<basic_json>;
+    template<typename T, typename SFINAE>
+    using json_serializer = JSONSerializer<T, SFINAE>;
+    /// how to treat decoding errors
+    using error_handler_t = detail::error_handler_t;
+    /// how to treat CBOR tags
+    using cbor_tag_handler_t = detail::cbor_tag_handler_t;
+    /// helper type for initializer lists of basic_json values
+    using initializer_list_t = std::initializer_list<detail::json_ref<basic_json>>;
+
+    using input_format_t = detail::input_format_t;
+    /// SAX interface type, see @ref nlohmann::json_sax
+    using json_sax_t = json_sax<basic_json>;
+
+    ////////////////
+    // exceptions //
+    ////////////////
+
+    /// @name exceptions
+    /// Classes to implement user-defined exceptions.
+    /// @{
+
+    using exception = detail::exception;
+    using parse_error = detail::parse_error;
+    using invalid_iterator = detail::invalid_iterator;
+    using type_error = detail::type_error;
+    using out_of_range = detail::out_of_range;
+    using other_error = detail::other_error;
+
+    /// @}
+
+
+    /////////////////////
+    // container types //
+    /////////////////////
+
+    /// @name container types
+    /// The canonic container types to use @ref basic_json like any other STL
+    /// container.
+    /// @{
+
+    /// the type of elements in a basic_json container
+    using value_type = basic_json;
+
+    /// the type of an element reference
+    using reference = value_type&;
+    /// the type of an element const reference
+    using const_reference = const value_type&;
+
+    /// a type to represent differences between iterators
+    using difference_type = std::ptrdiff_t;
+    /// a type to represent container sizes
+    using size_type = std::size_t;
+
+    /// the allocator type
+    using allocator_type = AllocatorType<basic_json>;
+
+    /// the type of an element pointer
+    using pointer = typename std::allocator_traits<allocator_type>::pointer;
+    /// the type of an element const pointer
+    using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
+
+    /// an iterator for a basic_json container
+    using iterator = iter_impl<basic_json>;
+    /// a const iterator for a basic_json container
+    using const_iterator = iter_impl<const basic_json>;
+    /// a reverse iterator for a basic_json container
+    using reverse_iterator = json_reverse_iterator<typename basic_json::iterator>;
+    /// a const reverse iterator for a basic_json container
+    using const_reverse_iterator = json_reverse_iterator<typename basic_json::const_iterator>;
+
+    /// @}
+
+
+    /// @brief returns the allocator associated with the container
+    /// @sa https://json.nlohmann.me/api/basic_json/get_allocator/
+    static allocator_type get_allocator()
+    {
+        return allocator_type();
+    }
+
+    /// @brief returns version information on the library
+    /// @sa https://json.nlohmann.me/api/basic_json/meta/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json meta()
+    {
+        basic_json result;
+
+        result["copyright"] = "(C) 2013-2022 Niels Lohmann";
+        result["name"] = "JSON for Modern C++";
+        result["url"] = "https://github.com/nlohmann/json";
+        result["version"]["string"] =
+            std::to_string(NLOHMANN_JSON_VERSION_MAJOR) + "." +
+            std::to_string(NLOHMANN_JSON_VERSION_MINOR) + "." +
+            std::to_string(NLOHMANN_JSON_VERSION_PATCH);
+        result["version"]["major"] = NLOHMANN_JSON_VERSION_MAJOR;
+        result["version"]["minor"] = NLOHMANN_JSON_VERSION_MINOR;
+        result["version"]["patch"] = NLOHMANN_JSON_VERSION_PATCH;
+
+#ifdef _WIN32
+        result["platform"] = "win32";
+#elif defined __linux__
+        result["platform"] = "linux";
+#elif defined __APPLE__
+        result["platform"] = "apple";
+#elif defined __unix__
+        result["platform"] = "unix";
+#else
+        result["platform"] = "unknown";
+#endif
+
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+        result["compiler"] = {{"family", "icc"}, {"version", __INTEL_COMPILER}};
+#elif defined(__clang__)
+        result["compiler"] = {{"family", "clang"}, {"version", __clang_version__}};
+#elif defined(__GNUC__) || defined(__GNUG__)
+        result["compiler"] = {{"family", "gcc"}, {"version", std::to_string(__GNUC__) + "." + std::to_string(__GNUC_MINOR__) + "." + std::to_string(__GNUC_PATCHLEVEL__)}};
+#elif defined(__HP_cc) || defined(__HP_aCC)
+        result["compiler"] = "hp"
+#elif defined(__IBMCPP__)
+        result["compiler"] = {{"family", "ilecpp"}, {"version", __IBMCPP__}};
+#elif defined(_MSC_VER)
+        result["compiler"] = {{"family", "msvc"}, {"version", _MSC_VER}};
+#elif defined(__PGI)
+        result["compiler"] = {{"family", "pgcpp"}, {"version", __PGI}};
+#elif defined(__SUNPRO_CC)
+        result["compiler"] = {{"family", "sunpro"}, {"version", __SUNPRO_CC}};
+#else
+        result["compiler"] = {{"family", "unknown"}, {"version", "unknown"}};
+#endif
+
+#ifdef __cplusplus
+        result["compiler"]["c++"] = std::to_string(__cplusplus);
+#else
+        result["compiler"]["c++"] = "unknown";
+#endif
+        return result;
+    }
+
+
+    ///////////////////////////
+    // JSON value data types //
+    ///////////////////////////
+
+    /// @name JSON value data types
+    /// The data types to store a JSON value. These types are derived from
+    /// the template arguments passed to class @ref basic_json.
+    /// @{
+
+    /// @brief object key comparator type
+    /// @sa https://json.nlohmann.me/api/basic_json/object_comparator_t/
+#if defined(JSON_HAS_CPP_14)
+    // Use transparent comparator if possible, combined with perfect forwarding
+    // on find() and count() calls prevents unnecessary string construction.
+    using object_comparator_t = std::less<>;
+#else
+    using object_comparator_t = std::less<StringType>;
+#endif
+
+    /// @brief a type for an object
+    /// @sa https://json.nlohmann.me/api/basic_json/object_t/
+    using object_t = ObjectType<StringType,
+          basic_json,
+          object_comparator_t,
+          AllocatorType<std::pair<const StringType,
+          basic_json>>>;
+
+    /// @brief a type for an array
+    /// @sa https://json.nlohmann.me/api/basic_json/array_t/
+    using array_t = ArrayType<basic_json, AllocatorType<basic_json>>;
+
+    /// @brief a type for a string
+    /// @sa https://json.nlohmann.me/api/basic_json/string_t/
+    using string_t = StringType;
+
+    /// @brief a type for a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/boolean_t/
+    using boolean_t = BooleanType;
+
+    /// @brief a type for a number (integer)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_integer_t/
+    using number_integer_t = NumberIntegerType;
+
+    /// @brief a type for a number (unsigned)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_unsigned_t/
+    using number_unsigned_t = NumberUnsignedType;
+
+    /// @brief a type for a number (floating-point)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_float_t/
+    using number_float_t = NumberFloatType;
+
+    /// @brief a type for a packed binary type
+    /// @sa https://json.nlohmann.me/api/basic_json/binary_t/
+    using binary_t = nlohmann::byte_container_with_subtype<BinaryType>;
+
+    /// @}
+
+  private:
+
+    /// helper for exception-safe object creation
+    template<typename T, typename... Args>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    static T* create(Args&& ... args)
+    {
+        AllocatorType<T> alloc;
+        using AllocatorTraits = std::allocator_traits<AllocatorType<T>>;
+
+        auto deleter = [&](T * obj)
+        {
+            AllocatorTraits::deallocate(alloc, obj, 1);
+        };
+        std::unique_ptr<T, decltype(deleter)> obj(AllocatorTraits::allocate(alloc, 1), deleter);
+        AllocatorTraits::construct(alloc, obj.get(), std::forward<Args>(args)...);
+        JSON_ASSERT(obj != nullptr);
+        return obj.release();
+    }
+
+    ////////////////////////
+    // JSON value storage //
+    ////////////////////////
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief a JSON value
+
+    The actual storage for a JSON value of the @ref basic_json class. This
+    union combines the different storage types for the JSON value types
+    defined in @ref value_t.
+
+    JSON type | value_t type    | used type
+    --------- | --------------- | ------------------------
+    object    | object          | pointer to @ref object_t
+    array     | array           | pointer to @ref array_t
+    string    | string          | pointer to @ref string_t
+    boolean   | boolean         | @ref boolean_t
+    number    | number_integer  | @ref number_integer_t
+    number    | number_unsigned | @ref number_unsigned_t
+    number    | number_float    | @ref number_float_t
+    binary    | binary          | pointer to @ref binary_t
+    null      | null            | *no value is stored*
+
+    @note Variable-length types (objects, arrays, and strings) are stored as
+    pointers. The size of the union should not exceed 64 bits if the default
+    value types are used.
+
+    @since version 1.0.0
+    */
+    union json_value
+    {
+        /// object (stored with pointer to save storage)
+        object_t* object;
+        /// array (stored with pointer to save storage)
+        array_t* array;
+        /// string (stored with pointer to save storage)
+        string_t* string;
+        /// binary (stored with pointer to save storage)
+        binary_t* binary;
+        /// boolean
+        boolean_t boolean;
+        /// number (integer)
+        number_integer_t number_integer;
+        /// number (unsigned integer)
+        number_unsigned_t number_unsigned;
+        /// number (floating-point)
+        number_float_t number_float;
+
+        /// default constructor (for null values)
+        json_value() = default;
+        /// constructor for booleans
+        json_value(boolean_t v) noexcept : boolean(v) {}
+        /// constructor for numbers (integer)
+        json_value(number_integer_t v) noexcept : number_integer(v) {}
+        /// constructor for numbers (unsigned)
+        json_value(number_unsigned_t v) noexcept : number_unsigned(v) {}
+        /// constructor for numbers (floating-point)
+        json_value(number_float_t v) noexcept : number_float(v) {}
+        /// constructor for empty values of a given type
+        json_value(value_t t)
+        {
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    object = create<object_t>();
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    array = create<array_t>();
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    string = create<string_t>("");
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    binary = create<binary_t>();
+                    break;
+                }
+
+                case value_t::boolean:
+                {
+                    boolean = static_cast<boolean_t>(false);
+                    break;
+                }
+
+                case value_t::number_integer:
+                {
+                    number_integer = static_cast<number_integer_t>(0);
+                    break;
+                }
+
+                case value_t::number_unsigned:
+                {
+                    number_unsigned = static_cast<number_unsigned_t>(0);
+                    break;
+                }
+
+                case value_t::number_float:
+                {
+                    number_float = static_cast<number_float_t>(0.0);
+                    break;
+                }
+
+                case value_t::null:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    break;
+                }
+
+                case value_t::discarded:
+                default:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    if (JSON_HEDLEY_UNLIKELY(t == value_t::null))
+                    {
+                        JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.10.5", basic_json())); // LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+            }
+        }
+
+        /// constructor for strings
+        json_value(const string_t& value) : string(create<string_t>(value)) {}
+
+        /// constructor for rvalue strings
+        json_value(string_t&& value) : string(create<string_t>(std::move(value))) {}
+
+        /// constructor for objects
+        json_value(const object_t& value) : object(create<object_t>(value)) {}
+
+        /// constructor for rvalue objects
+        json_value(object_t&& value) : object(create<object_t>(std::move(value))) {}
+
+        /// constructor for arrays
+        json_value(const array_t& value) : array(create<array_t>(value)) {}
+
+        /// constructor for rvalue arrays
+        json_value(array_t&& value) : array(create<array_t>(std::move(value))) {}
+
+        /// constructor for binary arrays
+        json_value(const typename binary_t::container_type& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays
+        json_value(typename binary_t::container_type&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        /// constructor for binary arrays (internal type)
+        json_value(const binary_t& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays (internal type)
+        json_value(binary_t&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        void destroy(value_t t)
+        {
+            if (t == value_t::array || t == value_t::object)
+            {
+                // flatten the current json_value to a heap-allocated stack
+                std::vector<basic_json> stack;
+
+                // move the top-level items to stack
+                if (t == value_t::array)
+                {
+                    stack.reserve(array->size());
+                    std::move(array->begin(), array->end(), std::back_inserter(stack));
+                }
+                else
+                {
+                    stack.reserve(object->size());
+                    for (auto&& it : *object)
+                    {
+                        stack.push_back(std::move(it.second));
+                    }
+                }
+
+                while (!stack.empty())
+                {
+                    // move the last item to local variable to be processed
+                    basic_json current_item(std::move(stack.back()));
+                    stack.pop_back();
+
+                    // if current_item is array/object, move
+                    // its children to the stack to be processed later
+                    if (current_item.is_array())
+                    {
+                        std::move(current_item.m_value.array->begin(), current_item.m_value.array->end(), std::back_inserter(stack));
+
+                        current_item.m_value.array->clear();
+                    }
+                    else if (current_item.is_object())
+                    {
+                        for (auto&& it : *current_item.m_value.object)
+                        {
+                            stack.push_back(std::move(it.second));
+                        }
+
+                        current_item.m_value.object->clear();
+                    }
+
+                    // it's now safe that current_item get destructed
+                    // since it doesn't have any children
+                }
+            }
+
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    AllocatorType<object_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, object);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, object, 1);
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    AllocatorType<array_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, array);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, array, 1);
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, string, 1);
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, binary, 1);
+                    break;
+                }
+
+                case value_t::null:
+                case value_t::boolean:
+                case value_t::number_integer:
+                case value_t::number_unsigned:
+                case value_t::number_float:
+                case value_t::discarded:
+                default:
+                {
+                    break;
+                }
+            }
+        }
+    };
+
+  private:
+    /*!
+    @brief checks the class invariants
+
+    This function asserts the class invariants. It needs to be called at the
+    end of every constructor to make sure that created objects respect the
+    invariant. Furthermore, it has to be called each time the type of a JSON
+    value is changed, because the invariant expresses a relationship between
+    @a m_type and @a m_value.
+
+    Furthermore, the parent relation is checked for arrays and objects: If
+    @a check_parents true and the value is an array or object, then the
+    container's elements must has the current value as parent.
+
+    @param[in] check_parents  whether the parent relation should be checked.
+               The value is true by default and should only be set to false
+               during destruction of objects when the invariant does not
+               need to hold.
+    */
+    void assert_invariant(bool check_parents = true) const noexcept
+    {
+        JSON_ASSERT(m_type != value_t::object || m_value.object != nullptr);
+        JSON_ASSERT(m_type != value_t::array || m_value.array != nullptr);
+        JSON_ASSERT(m_type != value_t::string || m_value.string != nullptr);
+        JSON_ASSERT(m_type != value_t::binary || m_value.binary != nullptr);
+
+#if JSON_DIAGNOSTICS
+        JSON_TRY
+        {
+            // cppcheck-suppress assertWithSideEffect
+            JSON_ASSERT(!check_parents || !is_structured() || std::all_of(begin(), end(), [this](const basic_json & j)
+            {
+                return j.m_parent == this;
+            }));
+        }
+        JSON_CATCH(...) {} // LCOV_EXCL_LINE
+#endif
+        static_cast<void>(check_parents);
+    }
+
+    void set_parents()
+    {
+#if JSON_DIAGNOSTICS
+        switch (m_type)
+        {
+            case value_t::array:
+            {
+                for (auto& element : *m_value.array)
+                {
+                    element.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::object:
+            {
+                for (auto& element : *m_value.object)
+                {
+                    element.second.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+#endif
+    }
+
+    iterator set_parents(iterator it, typename iterator::difference_type count_set_parents)
+    {
+#if JSON_DIAGNOSTICS
+        for (typename iterator::difference_type i = 0; i < count_set_parents; ++i)
+        {
+            (it + i)->m_parent = this;
+        }
+#else
+        static_cast<void>(count_set_parents);
+#endif
+        return it;
+    }
+
+    reference set_parent(reference j, std::size_t old_capacity = static_cast<std::size_t>(-1))
+    {
+#if JSON_DIAGNOSTICS
+        if (old_capacity != static_cast<std::size_t>(-1))
+        {
+            // see https://github.com/nlohmann/json/issues/2838
+            JSON_ASSERT(type() == value_t::array);
+            if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
+            {
+                // capacity has changed: update all parents
+                set_parents();
+                return j;
+            }
+        }
+
+        // ordered_json uses a vector internally, so pointers could have
+        // been invalidated; see https://github.com/nlohmann/json/issues/2962
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning(push )
+#pragma warning(disable : 4127) // ignore warning to replace if with if constexpr
+#endif
+        if (detail::is_ordered_map<object_t>::value)
+        {
+            set_parents();
+            return j;
+        }
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning( pop )
+#endif
+
+        j.m_parent = this;
+#else
+        static_cast<void>(j);
+        static_cast<void>(old_capacity);
+#endif
+        return j;
+    }
+
+  public:
+    //////////////////////////
+    // JSON parser callback //
+    //////////////////////////
+
+    /// @brief parser event types
+    /// @sa https://json.nlohmann.me/api/basic_json/parse_event_t/
+    using parse_event_t = detail::parse_event_t;
+
+    /// @brief per-element parser callback type
+    /// @sa https://json.nlohmann.me/api/basic_json/parser_callback_t/
+    using parser_callback_t = detail::parser_callback_t<basic_json>;
+
+    //////////////////
+    // constructors //
+    //////////////////
+
+    /// @name constructors and destructors
+    /// Constructors of class @ref basic_json, copy/move constructor, copy
+    /// assignment, static functions creating objects, and the destructor.
+    /// @{
+
+    /// @brief create an empty value with a given type
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const value_t v)
+        : m_type(v), m_value(v)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a null object
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(std::nullptr_t = nullptr) noexcept
+        : basic_json(value_t::null)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from compatible types
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename CompatibleType,
+               typename U = detail::uncvref_t<CompatibleType>,
+               detail::enable_if_t <
+                   !detail::is_basic_json<U>::value && detail::is_compatible_type<basic_json_t, U>::value, int > = 0 >
+    basic_json(CompatibleType && val) noexcept(noexcept( // NOLINT(bugprone-forwarding-reference-overload,bugprone-exception-escape)
+                JSONSerializer<U>::to_json(std::declval<basic_json_t&>(),
+                                           std::forward<CompatibleType>(val))))
+    {
+        JSONSerializer<U>::to_json(*this, std::forward<CompatibleType>(val));
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from an existing one
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value&& !std::is_same<basic_json, BasicJsonType>::value, int > = 0 >
+    basic_json(const BasicJsonType& val)
+    {
+        using other_boolean_t = typename BasicJsonType::boolean_t;
+        using other_number_float_t = typename BasicJsonType::number_float_t;
+        using other_number_integer_t = typename BasicJsonType::number_integer_t;
+        using other_number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+        using other_string_t = typename BasicJsonType::string_t;
+        using other_object_t = typename BasicJsonType::object_t;
+        using other_array_t = typename BasicJsonType::array_t;
+        using other_binary_t = typename BasicJsonType::binary_t;
+
+        switch (val.type())
+        {
+            case value_t::boolean:
+                JSONSerializer<other_boolean_t>::to_json(*this, val.template get<other_boolean_t>());
+                break;
+            case value_t::number_float:
+                JSONSerializer<other_number_float_t>::to_json(*this, val.template get<other_number_float_t>());
+                break;
+            case value_t::number_integer:
+                JSONSerializer<other_number_integer_t>::to_json(*this, val.template get<other_number_integer_t>());
+                break;
+            case value_t::number_unsigned:
+                JSONSerializer<other_number_unsigned_t>::to_json(*this, val.template get<other_number_unsigned_t>());
+                break;
+            case value_t::string:
+                JSONSerializer<other_string_t>::to_json(*this, val.template get_ref<const other_string_t&>());
+                break;
+            case value_t::object:
+                JSONSerializer<other_object_t>::to_json(*this, val.template get_ref<const other_object_t&>());
+                break;
+            case value_t::array:
+                JSONSerializer<other_array_t>::to_json(*this, val.template get_ref<const other_array_t&>());
+                break;
+            case value_t::binary:
+                JSONSerializer<other_binary_t>::to_json(*this, val.template get_ref<const other_binary_t&>());
+                break;
+            case value_t::null:
+                *this = nullptr;
+                break;
+            case value_t::discarded:
+                m_type = value_t::discarded;
+                break;
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a container (array or object) from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(initializer_list_t init,
+               bool type_deduction = true,
+               value_t manual_type = value_t::array)
+    {
+        // check if each element is an array with two elements whose first
+        // element is a string
+        bool is_an_object = std::all_of(init.begin(), init.end(),
+                                        [](const detail::json_ref<basic_json>& element_ref)
+        {
+            return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[0].is_string();
+        });
+
+        // adjust type if type deduction is not wanted
+        if (!type_deduction)
+        {
+            // if array is wanted, do not create an object though possible
+            if (manual_type == value_t::array)
+            {
+                is_an_object = false;
+            }
+
+            // if object is wanted but impossible, throw an exception
+            if (JSON_HEDLEY_UNLIKELY(manual_type == value_t::object && !is_an_object))
+            {
+                JSON_THROW(type_error::create(301, "cannot create object from initializer list", basic_json()));
+            }
+        }
+
+        if (is_an_object)
+        {
+            // the initializer list is a list of pairs -> create object
+            m_type = value_t::object;
+            m_value = value_t::object;
+
+            for (auto& element_ref : init)
+            {
+                auto element = element_ref.moved_or_copied();
+                m_value.object->emplace(
+                    std::move(*((*element.m_value.array)[0].m_value.string)),
+                    std::move((*element.m_value.array)[1]));
+            }
+        }
+        else
+        {
+            // the initializer list describes an array -> create array
+            m_type = value_t::array;
+            m_value.array = create<array_t>(init.begin(), init.end());
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief explicitly create a binary array (without subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = init;
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = binary_t(init, subtype);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = std::move(init);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = binary_t(std::move(init), subtype);
+        return res;
+    }
+
+    /// @brief explicitly create an array from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/array/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json array(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::array);
+    }
+
+    /// @brief explicitly create an object from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/object/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json object(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::object);
+    }
+
+    /// @brief construct an array with count copies of given value
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(size_type cnt, const basic_json& val)
+        : m_type(value_t::array)
+    {
+        m_value.array = create<array_t>(cnt, val);
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief construct a JSON container given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < class InputIT, typename std::enable_if <
+                   std::is_same<InputIT, typename basic_json_t::iterator>::value ||
+                   std::is_same<InputIT, typename basic_json_t::const_iterator>::value, int >::type = 0 >
+    basic_json(InputIT first, InputIT last)
+    {
+        JSON_ASSERT(first.m_object != nullptr);
+        JSON_ASSERT(last.m_object != nullptr);
+
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(201, "iterators are not compatible", basic_json()));
+        }
+
+        // copy type from first iterator
+        m_type = first.m_object->m_type;
+
+        // check if iterator range is complete for primitive values
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!first.m_it.primitive_iterator.is_begin()
+                                         || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", *first.m_object));
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::object:
+            case value_t::array:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        switch (m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_value.number_integer = first.m_object->m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value.number_unsigned = first.m_object->m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value.number_float = first.m_object->m_value.number_float;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value.boolean = first.m_object->m_value.boolean;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value = *first.m_object->m_value.string;
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_value.object = create<object_t>(first.m_it.object_iterator,
+                                                  last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value.array = create<array_t>(first.m_it.array_iterator,
+                                                last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value = *first.m_object->m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(invalid_iterator::create(206, "cannot construct with iterators from " + std::string(first.m_object->type_name()), *first.m_object));
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+
+    ///////////////////////////////////////
+    // other constructors and destructor //
+    ///////////////////////////////////////
+
+    template<typename JsonRef,
+             detail::enable_if_t<detail::conjunction<detail::is_json_ref<JsonRef>,
+                                 std::is_same<typename JsonRef::value_type, basic_json>>::value, int> = 0 >
+    basic_json(const JsonRef& ref) : basic_json(ref.moved_or_copied()) {}
+
+    /// @brief copy constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const basic_json& other)
+        : m_type(other.m_type)
+    {
+        // check of passed value is valid
+        other.assert_invariant();
+
+        switch (m_type)
+        {
+            case value_t::object:
+            {
+                m_value = *other.m_value.object;
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value = *other.m_value.array;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value = *other.m_value.string;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value = other.m_value.boolean;
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                m_value = other.m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value = other.m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value = other.m_value.number_float;
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value = *other.m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief move constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(basic_json&& other) noexcept
+        : m_type(std::move(other.m_type)),
+          m_value(std::move(other.m_value))
+    {
+        // check that passed value is valid
+        other.assert_invariant(false);
+
+        // invalidate payload
+        other.m_type = value_t::null;
+        other.m_value = {};
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief copy assignment
+    /// @sa https://json.nlohmann.me/api/basic_json/operator=/
+    basic_json& operator=(basic_json other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        // check that passed value is valid
+        other.assert_invariant();
+
+        using std::swap;
+        swap(m_type, other.m_type);
+        swap(m_value, other.m_value);
+
+        set_parents();
+        assert_invariant();
+        return *this;
+    }
+
+    /// @brief destructor
+    /// @sa https://json.nlohmann.me/api/basic_json/~basic_json/
+    ~basic_json() noexcept
+    {
+        assert_invariant(false);
+        m_value.destroy(m_type);
+    }
+
+    /// @}
+
+  public:
+    ///////////////////////
+    // object inspection //
+    ///////////////////////
+
+    /// @name object inspection
+    /// Functions to inspect the type of a JSON value.
+    /// @{
+
+    /// @brief serialization
+    /// @sa https://json.nlohmann.me/api/basic_json/dump/
+    string_t dump(const int indent = -1,
+                  const char indent_char = ' ',
+                  const bool ensure_ascii = false,
+                  const error_handler_t error_handler = error_handler_t::strict) const
+    {
+        string_t result;
+        serializer s(detail::output_adapter<char, string_t>(result), indent_char, error_handler);
+
+        if (indent >= 0)
+        {
+            s.dump(*this, true, ensure_ascii, static_cast<unsigned int>(indent));
+        }
+        else
+        {
+            s.dump(*this, false, ensure_ascii, 0);
+        }
+
+        return result;
+    }
+
+    /// @brief return the type of the JSON value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/type/
+    constexpr value_t type() const noexcept
+    {
+        return m_type;
+    }
+
+    /// @brief return whether type is primitive
+    /// @sa https://json.nlohmann.me/api/basic_json/is_primitive/
+    constexpr bool is_primitive() const noexcept
+    {
+        return is_null() || is_string() || is_boolean() || is_number() || is_binary();
+    }
+
+    /// @brief return whether type is structured
+    /// @sa https://json.nlohmann.me/api/basic_json/is_structured/
+    constexpr bool is_structured() const noexcept
+    {
+        return is_array() || is_object();
+    }
+
+    /// @brief return whether value is null
+    /// @sa https://json.nlohmann.me/api/basic_json/is_null/
+    constexpr bool is_null() const noexcept
+    {
+        return m_type == value_t::null;
+    }
+
+    /// @brief return whether value is a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/is_boolean/
+    constexpr bool is_boolean() const noexcept
+    {
+        return m_type == value_t::boolean;
+    }
+
+    /// @brief return whether value is a number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number/
+    constexpr bool is_number() const noexcept
+    {
+        return is_number_integer() || is_number_float();
+    }
+
+    /// @brief return whether value is an integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_integer/
+    constexpr bool is_number_integer() const noexcept
+    {
+        return m_type == value_t::number_integer || m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is an unsigned integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_unsigned/
+    constexpr bool is_number_unsigned() const noexcept
+    {
+        return m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is a floating-point number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_float/
+    constexpr bool is_number_float() const noexcept
+    {
+        return m_type == value_t::number_float;
+    }
+
+    /// @brief return whether value is an object
+    /// @sa https://json.nlohmann.me/api/basic_json/is_object/
+    constexpr bool is_object() const noexcept
+    {
+        return m_type == value_t::object;
+    }
+
+    /// @brief return whether value is an array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_array/
+    constexpr bool is_array() const noexcept
+    {
+        return m_type == value_t::array;
+    }
+
+    /// @brief return whether value is a string
+    /// @sa https://json.nlohmann.me/api/basic_json/is_string/
+    constexpr bool is_string() const noexcept
+    {
+        return m_type == value_t::string;
+    }
+
+    /// @brief return whether value is a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_binary/
+    constexpr bool is_binary() const noexcept
+    {
+        return m_type == value_t::binary;
+    }
+
+    /// @brief return whether value is discarded
+    /// @sa https://json.nlohmann.me/api/basic_json/is_discarded/
+    constexpr bool is_discarded() const noexcept
+    {
+        return m_type == value_t::discarded;
+    }
+
+    /// @brief return the type of the JSON value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_value_t/
+    constexpr operator value_t() const noexcept
+    {
+        return m_type;
+    }
+
+    /// @}
+
+  private:
+    //////////////////
+    // value access //
+    //////////////////
+
+    /// get a boolean (explicit)
+    boolean_t get_impl(boolean_t* /*unused*/) const
+    {
+        if (JSON_HEDLEY_LIKELY(is_boolean()))
+        {
+            return m_value.boolean;
+        }
+
+        JSON_THROW(type_error::create(302, "type must be boolean, but is " + std::string(type_name()), *this));
+    }
+
+    /// get a pointer to the value (object)
+    object_t* get_impl_ptr(object_t* /*unused*/) noexcept
+    {
+        return is_object() ? m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (object)
+    constexpr const object_t* get_impl_ptr(const object_t* /*unused*/) const noexcept
+    {
+        return is_object() ? m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    array_t* get_impl_ptr(array_t* /*unused*/) noexcept
+    {
+        return is_array() ? m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    constexpr const array_t* get_impl_ptr(const array_t* /*unused*/) const noexcept
+    {
+        return is_array() ? m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    string_t* get_impl_ptr(string_t* /*unused*/) noexcept
+    {
+        return is_string() ? m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    constexpr const string_t* get_impl_ptr(const string_t* /*unused*/) const noexcept
+    {
+        return is_string() ? m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    boolean_t* get_impl_ptr(boolean_t* /*unused*/) noexcept
+    {
+        return is_boolean() ? &m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    constexpr const boolean_t* get_impl_ptr(const boolean_t* /*unused*/) const noexcept
+    {
+        return is_boolean() ? &m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    number_integer_t* get_impl_ptr(number_integer_t* /*unused*/) noexcept
+    {
+        return is_number_integer() ? &m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    constexpr const number_integer_t* get_impl_ptr(const number_integer_t* /*unused*/) const noexcept
+    {
+        return is_number_integer() ? &m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    number_unsigned_t* get_impl_ptr(number_unsigned_t* /*unused*/) noexcept
+    {
+        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    constexpr const number_unsigned_t* get_impl_ptr(const number_unsigned_t* /*unused*/) const noexcept
+    {
+        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    number_float_t* get_impl_ptr(number_float_t* /*unused*/) noexcept
+    {
+        return is_number_float() ? &m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    constexpr const number_float_t* get_impl_ptr(const number_float_t* /*unused*/) const noexcept
+    {
+        return is_number_float() ? &m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    binary_t* get_impl_ptr(binary_t* /*unused*/) noexcept
+    {
+        return is_binary() ? m_value.binary : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    constexpr const binary_t* get_impl_ptr(const binary_t* /*unused*/) const noexcept
+    {
+        return is_binary() ? m_value.binary : nullptr;
+    }
+
+    /*!
+    @brief helper function to implement get_ref()
+
+    This function helps to implement get_ref() without code duplication for
+    const and non-const overloads
+
+    @tparam ThisType will be deduced as `basic_json` or `const basic_json`
+
+    @throw type_error.303 if ReferenceType does not match underlying value
+    type of the current JSON
+    */
+    template<typename ReferenceType, typename ThisType>
+    static ReferenceType get_ref_impl(ThisType& obj)
+    {
+        // delegate the call to get_ptr<>()
+        auto* ptr = obj.template get_ptr<typename std::add_pointer<ReferenceType>::type>();
+
+        if (JSON_HEDLEY_LIKELY(ptr != nullptr))
+        {
+            return *ptr;
+        }
+
+        JSON_THROW(type_error::create(303, "incompatible ReferenceType for get_ref, actual type is " + std::string(obj.type_name()), obj));
+    }
+
+  public:
+    /// @name value access
+    /// Direct access to the stored value of a JSON value.
+    /// @{
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get_ptr() noexcept -> decltype(std::declval<basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>()
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template < typename PointerType, typename std::enable_if <
+                   std::is_pointer<PointerType>::value&&
+                   std::is_const<typename std::remove_pointer<PointerType>::type>::value, int >::type = 0 >
+    constexpr auto get_ptr() const noexcept -> decltype(std::declval<const basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>() const
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+  private:
+    /*!
+    @brief get a value (explicit)
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    ValueType ret;
+    JSONSerializer<ValueType>::from_json(*this, ret);
+    return ret;
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json,
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `void from_json(const basic_json&, ValueType&)`, and
+    - @ref json_serializer<ValueType> does not has a `from_json()` method of
+      the form `ValueType from_json(const basic_json&)`
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,get__ValueType_const}
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::is_default_constructible<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<0> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), std::declval<ValueType&>())))
+    {
+        auto ret = ValueType();
+        JSONSerializer<ValueType>::from_json(*this, ret);
+        return ret;
+    }
+
+    /*!
+    @brief get a value (explicit); special case
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is **not** [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and **not** [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    return JSONSerializer<ValueType>::from_json(*this);
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json and
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `ValueType from_json(const basic_json&)`
+
+    @note If @ref json_serializer<ValueType> has both overloads of
+    `from_json()`, this one is chosen.
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::has_non_default_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<1> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>())))
+    {
+        return JSONSerializer<ValueType>::from_json(*this);
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads converts the current @ref basic_json in a different
+    @ref basic_json type
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this, converted into @a BasicJsonType
+
+    @complexity Depending on the implementation of the called `from_json()`
+                method.
+
+    @since version 3.2.0
+    */
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value,
+                   int > = 0 >
+    BasicJsonType get_impl(detail::priority_tag<2> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads avoids a lot of template boilerplate, it can be seen as the
+    identity method
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this
+
+    @complexity Constant.
+
+    @since version 2.1.0
+    */
+    template<typename BasicJsonType,
+             detail::enable_if_t<
+                 std::is_same<BasicJsonType, basic_json_t>::value,
+                 int> = 0>
+    basic_json get_impl(detail::priority_tag<3> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+    @copydoc get()
+    */
+    template<typename PointerType,
+             detail::enable_if_t<
+                 std::is_pointer<PointerType>::value,
+                 int> = 0>
+    constexpr auto get_impl(detail::priority_tag<4> /*unused*/) const noexcept
+    -> decltype(std::declval<const basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+  public:
+    /*!
+    @brief get a (pointer) value (explicit)
+
+    Performs explicit type conversion between the JSON value and a compatible value if required.
+
+    - If the requested type is a pointer to the internally stored JSON value that pointer is returned.
+    No copies are made.
+
+    - If the requested type is the current @ref basic_json, or a different @ref basic_json convertible
+    from the current @ref basic_json.
+
+    - Otherwise the value is converted by calling the @ref json_serializer<ValueType> `from_json()`
+    method.
+
+    @tparam ValueTypeCV the provided value type
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @tparam ValueType if necessary
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws if conversion is required
+
+    @since version 2.1.0
+    */
+    template < typename ValueTypeCV, typename ValueType = detail::uncvref_t<ValueTypeCV>>
+#if defined(JSON_HAS_CPP_14)
+    constexpr
+#endif
+    auto get() const noexcept(
+    noexcept(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {})))
+    -> decltype(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {}))
+    {
+        // we cannot static_assert on ValueTypeCV being non-const, because
+        // there is support for get<const basic_json_t>(), which is why we
+        // still need the uncvref
+        static_assert(!std::is_reference<ValueTypeCV>::value,
+                      "get() cannot be used with reference types, you might want to use get_ref()");
+        return get_impl<ValueType>(detail::priority_tag<4> {});
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+
+    Explicit pointer access to the internally stored JSON value. No copies are
+    made.
+
+    @warning The pointer becomes invalid if the underlying JSON object
+    changes.
+
+    @tparam PointerType pointer type; must be a pointer to @ref array_t, @ref
+    object_t, @ref string_t, @ref boolean_t, @ref number_integer_t,
+    @ref number_unsigned_t, or @ref number_float_t.
+
+    @return pointer to the internally stored JSON value if the requested
+    pointer type @a PointerType fits to the JSON value; `nullptr` otherwise
+
+    @complexity Constant.
+
+    @liveexample{The example below shows how pointers to internal values of a
+    JSON value can be requested. Note that no type conversions are made and a
+    `nullptr` is returned if the value and the requested pointer type does not
+    match.,get__PointerType}
+
+    @sa see @ref get_ptr() for explicit pointer-member access
+
+    @since version 1.0.0
+    */
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get() noexcept -> decltype(std::declval<basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+    /// @brief get a value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_to/
+    template < typename ValueType,
+               detail::enable_if_t <
+                   !detail::is_basic_json<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType & get_to(ValueType& v) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<ValueType>::from_json(*this, v);
+        return v;
+    }
+
+    // specialization to allow calling get_to with a basic_json value
+    // see https://github.com/nlohmann/json/issues/2175
+    template<typename ValueType,
+             detail::enable_if_t <
+                 detail::is_basic_json<ValueType>::value,
+                 int> = 0>
+    ValueType & get_to(ValueType& v) const
+    {
+        v = *this;
+        return v;
+    }
+
+    template <
+        typename T, std::size_t N,
+        typename Array = T (&)[N], // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+        detail::enable_if_t <
+            detail::has_from_json<basic_json_t, Array>::value, int > = 0 >
+    Array get_to(T (&v)[N]) const // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    noexcept(noexcept(JSONSerializer<Array>::from_json(
+                          std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<Array>::from_json(*this, v);
+        return v;
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template<typename ReferenceType, typename std::enable_if<
+                 std::is_reference<ReferenceType>::value, int>::type = 0>
+    ReferenceType get_ref()
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template < typename ReferenceType, typename std::enable_if <
+                   std::is_reference<ReferenceType>::value&&
+                   std::is_const<typename std::remove_reference<ReferenceType>::type>::value, int >::type = 0 >
+    ReferenceType get_ref() const
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /*!
+    @brief get a value (implicit)
+
+    Implicit type conversion between the JSON value and a compatible value.
+    The call is realized by calling @ref get() const.
+
+    @tparam ValueType non-pointer type compatible to the JSON value, for
+    instance `int` for JSON integer numbers, `bool` for JSON booleans, or
+    `std::vector` types for JSON arrays. The character type of @ref string_t
+    as well as an initializer list of this type is excluded to avoid
+    ambiguities as these types implicitly convert to `std::string`.
+
+    @return copy of the JSON value, converted to type @a ValueType
+
+    @throw type_error.302 in case passed type @a ValueType is incompatible
+    to the JSON value type (e.g., the JSON value is of type boolean, but a
+    string is requested); see example below
+
+    @complexity Linear in the size of the JSON value.
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,operator__ValueType}
+
+    @since version 1.0.0
+    */
+    template < typename ValueType, typename std::enable_if <
+                   detail::conjunction <
+                       detail::negation<std::is_pointer<ValueType>>,
+                       detail::negation<std::is_same<ValueType, detail::json_ref<basic_json>>>,
+                                        detail::negation<std::is_same<ValueType, typename string_t::value_type>>,
+                                        detail::negation<detail::is_basic_json<ValueType>>,
+                                        detail::negation<std::is_same<ValueType, std::initializer_list<typename string_t::value_type>>>,
+
+#if defined(JSON_HAS_CPP_17) && (defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1910 && _MSC_VER <= 1914))
+                                                detail::negation<std::is_same<ValueType, std::string_view>>,
+#endif
+                                                detail::is_detected_lazy<detail::get_template_function, const basic_json_t&, ValueType>
+                                                >::value, int >::type = 0 >
+                                        JSON_EXPLICIT operator ValueType() const
+    {
+        // delegate the call to get<>() const
+        return get<ValueType>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    binary_t& get_binary()
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, "type must be binary, but is " + std::string(type_name()), *this));
+        }
+
+        return *get_ptr<binary_t*>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    const binary_t& get_binary() const
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, "type must be binary, but is " + std::string(type_name()), *this));
+        }
+
+        return *get_ptr<const binary_t*>();
+    }
+
+    /// @}
+
+
+    ////////////////////
+    // element access //
+    ////////////////////
+
+    /// @name element access
+    /// Access to the JSON value.
+    /// @{
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(size_type idx)
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return set_parent(m_value.array->at(idx));
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, "array index " + std::to_string(idx) + " is out of range", *this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, "cannot use at() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(size_type idx) const
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return m_value.array->at(idx);
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, "array index " + std::to_string(idx) + " is out of range", *this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, "cannot use at() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const typename object_t::key_type& key)
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            JSON_TRY
+            {
+                return set_parent(m_value.object->at(key));
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(403, "key '" + key + "' not found", *this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, "cannot use at() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const typename object_t::key_type& key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            JSON_TRY
+            {
+                return m_value.object->at(key);
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(403, "key '" + key + "' not found", *this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, "cannot use at() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](size_type idx)
+    {
+        // implicitly convert null value to an empty array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value.array = create<array_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // fill up array with null values if given idx is outside range
+            if (idx >= m_value.array->size())
+            {
+#if JSON_DIAGNOSTICS
+                // remember array size & capacity before resizing
+                const auto old_size = m_value.array->size();
+                const auto old_capacity = m_value.array->capacity();
+#endif
+                m_value.array->resize(idx + 1);
+
+#if JSON_DIAGNOSTICS
+                if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
+                {
+                    // capacity has changed: update all parents
+                    set_parents();
+                }
+                else
+                {
+                    // set parent for values added above
+                    set_parents(begin() + static_cast<typename iterator::difference_type>(old_size), static_cast<typename iterator::difference_type>(idx + 1 - old_size));
+                }
+#endif
+                assert_invariant();
+            }
+
+            return m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a numeric argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](size_type idx) const
+    {
+        // const operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            return m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a numeric argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](const typename object_t::key_type& key)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            return set_parent(m_value.object->operator[](key));
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a string argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const typename object_t::key_type& key) const
+    {
+        // const operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            JSON_ASSERT(m_value.object->find(key) != m_value.object->end());
+            return m_value.object->find(key)->second;
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a string argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<typename T>
+    JSON_HEDLEY_NON_NULL(2)
+    reference operator[](T* key)
+    {
+        // implicitly convert null to object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            return set_parent(m_value.object->operator[](key));
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a string argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<typename T>
+    JSON_HEDLEY_NON_NULL(2)
+    const_reference operator[](T* key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            JSON_ASSERT(m_value.object->find(key) != m_value.object->end());
+            return m_value.object->find(key)->second;
+        }
+
+        JSON_THROW(type_error::create(305, "cannot use operator[] with a string argument with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    /// using std::is_convertible in a std::enable_if will fail when using explicit conversions
+    template < class ValueType, typename std::enable_if <
+                   detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, ValueType>::value, int >::type = 0 >
+    ValueType value(const typename object_t::key_type& key, const ValueType& default_value) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(key);
+            if (it != end())
+            {
+                return it->template get<ValueType>();
+            }
+
+            return default_value;
+        }
+
+        JSON_THROW(type_error::create(306, "cannot use value() with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    /// overload for a default value of type const char*
+    string_t value(const typename object_t::key_type& key, const char* default_value) const
+    {
+        return value(key, string_t(default_value));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template<class ValueType, typename std::enable_if<
+                 detail::is_getable<basic_json_t, ValueType>::value, int>::type = 0>
+    ValueType value(const json_pointer& ptr, const ValueType& default_value) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if pointer resolves a value, return it or use default value
+            JSON_TRY
+            {
+                return ptr.get_checked(this).template get<ValueType>();
+            }
+            JSON_INTERNAL_CATCH (out_of_range&)
+            {
+                return default_value;
+            }
+        }
+
+        JSON_THROW(type_error::create(306, "cannot use value() with " + std::string(type_name()), *this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    /// overload for a default value of type const char*
+    JSON_HEDLEY_NON_NULL(3)
+    string_t value(const json_pointer& ptr, const char* default_value) const
+    {
+        return value(ptr, string_t(default_value));
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    reference front()
+    {
+        return *begin();
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    const_reference front() const
+    {
+        return *cbegin();
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    reference back()
+    {
+        auto tmp = end();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    const_reference back() const
+    {
+        auto tmp = cend();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief remove element given an iterator
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, typename std::enable_if <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int >::type
+               = 0 >
+    IteratorType erase(IteratorType pos)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != pos.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", *this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!pos.m_it.primitive_iterator.is_begin()))
+                {
+                    JSON_THROW(invalid_iterator::create(205, "iterator out of range", *this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
+                    m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
+                    m_value.binary = nullptr;
+                }
+
+                m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_value.object->erase(pos.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_value.array->erase(pos.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, "cannot use erase() with " + std::string(type_name()), *this));
+        }
+
+        return result;
+    }
+
+    /// @brief remove elements given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, typename std::enable_if <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int >::type
+               = 0 >
+    IteratorType erase(IteratorType first, IteratorType last)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != first.m_object || this != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(203, "iterators do not fit current value", *this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_LIKELY(!first.m_it.primitive_iterator.is_begin()
+                                       || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", *this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
+                    m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
+                    m_value.binary = nullptr;
+                }
+
+                m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_value.object->erase(first.m_it.object_iterator,
+                                              last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_value.array->erase(first.m_it.array_iterator,
+                                             last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, "cannot use erase() with " + std::string(type_name()), *this));
+        }
+
+        return result;
+    }
+
+    /// @brief remove element from a JSON object given a key
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    size_type erase(const typename object_t::key_type& key)
+    {
+        // this erase only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            return m_value.object->erase(key);
+        }
+
+        JSON_THROW(type_error::create(307, "cannot use erase() with " + std::string(type_name()), *this));
+    }
+
+    /// @brief remove element from a JSON array given an index
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    void erase(const size_type idx)
+    {
+        // this erase only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            if (JSON_HEDLEY_UNLIKELY(idx >= size()))
+            {
+                JSON_THROW(out_of_range::create(401, "array index " + std::to_string(idx) + " is out of range", *this));
+            }
+
+            m_value.array->erase(m_value.array->begin() + static_cast<difference_type>(idx));
+        }
+        else
+        {
+            JSON_THROW(type_error::create(307, "cannot use erase() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @}
+
+
+    ////////////
+    // lookup //
+    ////////////
+
+    /// @name lookup
+    /// @{
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<typename KeyT>
+    iterator find(KeyT&& key)
+    {
+        auto result = end();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(std::forward<KeyT>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<typename KeyT>
+    const_iterator find(KeyT&& key) const
+    {
+        auto result = cend();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(std::forward<KeyT>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief returns the number of occurrences of a key in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/count/
+    template<typename KeyT>
+    size_type count(KeyT&& key) const
+    {
+        // return 0 for all nonobject types
+        return is_object() ? m_value.object->count(std::forward<KeyT>(key)) : 0;
+    }
+
+    /// @brief check the existence of an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    template < typename KeyT, typename std::enable_if <
+                   !std::is_same<typename std::decay<KeyT>::type, json_pointer>::value, int >::type = 0 >
+    bool contains(KeyT && key) const
+    {
+        return is_object() && m_value.object->find(std::forward<KeyT>(key)) != m_value.object->end();
+    }
+
+    /// @brief check the existence of an element in a JSON object given a JSON pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    bool contains(const json_pointer& ptr) const
+    {
+        return ptr.contains(this);
+    }
+
+    /// @}
+
+
+    ///////////////
+    // iterators //
+    ///////////////
+
+    /// @name iterators
+    /// @{
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    iterator begin() noexcept
+    {
+        iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    const_iterator begin() const noexcept
+    {
+        return cbegin();
+    }
+
+    /// @brief returns a const iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/cbegin/
+    const_iterator cbegin() const noexcept
+    {
+        const_iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    iterator end() noexcept
+    {
+        iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    const_iterator end() const noexcept
+    {
+        return cend();
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/cend/
+    const_iterator cend() const noexcept
+    {
+        const_iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    reverse_iterator rbegin() noexcept
+    {
+        return reverse_iterator(end());
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    const_reverse_iterator rbegin() const noexcept
+    {
+        return crbegin();
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    reverse_iterator rend() noexcept
+    {
+        return reverse_iterator(begin());
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    const_reverse_iterator rend() const noexcept
+    {
+        return crend();
+    }
+
+    /// @brief returns a const reverse iterator to the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/crbegin/
+    const_reverse_iterator crbegin() const noexcept
+    {
+        return const_reverse_iterator(cend());
+    }
+
+    /// @brief returns a const reverse iterator to one before the first
+    /// @sa https://json.nlohmann.me/api/basic_json/crend/
+    const_reverse_iterator crend() const noexcept
+    {
+        return const_reverse_iterator(cbegin());
+    }
+
+  public:
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use @ref items() instead;
+    ///             that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<iterator> iterator_wrapper(reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///         version 4.0.0 of the library. Please use @ref items() instead;
+    ///         that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<const_iterator> iterator_wrapper(const_reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<iterator> items() noexcept
+    {
+        return iteration_proxy<iterator>(*this);
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<const_iterator> items() const noexcept
+    {
+        return iteration_proxy<const_iterator>(*this);
+    }
+
+    /// @}
+
+
+    //////////////
+    // capacity //
+    //////////////
+
+    /// @name capacity
+    /// @{
+
+    /// @brief checks whether the container is empty.
+    /// @sa https://json.nlohmann.me/api/basic_json/empty/
+    bool empty() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return true;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::empty()
+                return m_value.array->empty();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::empty()
+                return m_value.object->empty();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types are nonempty
+                return false;
+            }
+        }
+    }
+
+    /// @brief returns the number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/size/
+    size_type size() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return 0;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::size()
+                return m_value.array->size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::size()
+                return m_value.object->size();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have size 1
+                return 1;
+            }
+        }
+    }
+
+    /// @brief returns the maximum possible number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/max_size/
+    size_type max_size() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::array:
+            {
+                // delegate call to array_t::max_size()
+                return m_value.array->max_size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::max_size()
+                return m_value.object->max_size();
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have max_size() == size()
+                return size();
+            }
+        }
+    }
+
+    /// @}
+
+
+    ///////////////
+    // modifiers //
+    ///////////////
+
+    /// @name modifiers
+    /// @{
+
+    /// @brief clears the contents
+    /// @sa https://json.nlohmann.me/api/basic_json/clear/
+    void clear() noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_value.number_integer = 0;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value.number_unsigned = 0;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value.number_float = 0.0;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value.boolean = false;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value.string->clear();
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value.binary->clear();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value.array->clear();
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_value.object->clear();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(basic_json&& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, "cannot use push_back() with " + std::string(type_name()), *this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (move semantics)
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->push_back(std::move(val));
+        set_parent(m_value.array->back(), old_capacity);
+        // if val is moved from, basic_json move constructor marks it null, so we do not call the destructor
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(basic_json&& val)
+    {
+        push_back(std::move(val));
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const basic_json& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, "cannot use push_back() with " + std::string(type_name()), *this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->push_back(val);
+        set_parent(m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const basic_json& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const typename object_t::value_type& val)
+    {
+        // push_back only works for null objects or objects
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(308, "cannot use push_back() with " + std::string(type_name()), *this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to object
+        auto res = m_value.object->insert(val);
+        set_parent(res.first->second);
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const typename object_t::value_type& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(initializer_list_t init)
+    {
+        if (is_object() && init.size() == 2 && (*init.begin())->is_string())
+        {
+            basic_json&& key = init.begin()->moved_or_copied();
+            push_back(typename object_t::value_type(
+                          std::move(key.get_ref<string_t&>()), (init.begin() + 1)->moved_or_copied()));
+        }
+        else
+        {
+            push_back(basic_json(init));
+        }
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(initializer_list_t init)
+    {
+        push_back(init);
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace_back/
+    template<class... Args>
+    reference emplace_back(Args&& ... args)
+    {
+        // emplace_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(311, "cannot use emplace_back() with " + std::string(type_name()), *this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->emplace_back(std::forward<Args>(args)...);
+        return set_parent(m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an object if key does not exist
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace/
+    template<class... Args>
+    std::pair<iterator, bool> emplace(Args&& ... args)
+    {
+        // emplace only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(311, "cannot use emplace() with " + std::string(type_name()), *this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        auto res = m_value.object->emplace(std::forward<Args>(args)...);
+        set_parent(res.first->second);
+
+        // create result iterator and set iterator to the result of emplace
+        auto it = begin();
+        it.m_it.object_iterator = res.first;
+
+        // return pair of iterator and boolean
+        return {it, res.second};
+    }
+
+    /// Helper for insertion of an iterator
+    /// @note: This uses std::distance to support GCC 4.8,
+    ///        see https://github.com/nlohmann/json/pull/1257
+    template<typename... Args>
+    iterator insert_iterator(const_iterator pos, Args&& ... args)
+    {
+        iterator result(this);
+        JSON_ASSERT(m_value.array != nullptr);
+
+        auto insert_pos = std::distance(m_value.array->begin(), pos.m_it.array_iterator);
+        m_value.array->insert(pos.m_it.array_iterator, std::forward<Args>(args)...);
+        result.m_it.array_iterator = m_value.array->begin() + insert_pos;
+
+        // This could have been written as:
+        // result.m_it.array_iterator = m_value.array->insert(pos.m_it.array_iterator, cnt, val);
+        // but the return value of insert is missing in GCC 4.8, so it is written this way instead.
+
+        set_parents();
+        return result;
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", *this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, val);
+        }
+
+        JSON_THROW(type_error::create(309, "cannot use insert() with " + std::string(type_name()), *this));
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, basic_json&& val)
+    {
+        return insert(pos, val);
+    }
+
+    /// @brief inserts copies of element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, size_type cnt, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", *this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, cnt, val);
+        }
+
+        JSON_THROW(type_error::create(309, "cannot use insert() with " + std::string(type_name()), *this));
+    }
+
+    /// @brief inserts range of elements into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const_iterator first, const_iterator last)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, "cannot use insert() with " + std::string(type_name()), *this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", *this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", *this));
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(first.m_object == this))
+        {
+            JSON_THROW(invalid_iterator::create(211, "passed iterators may not belong to container", *this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, first.m_it.array_iterator, last.m_it.array_iterator);
+    }
+
+    /// @brief inserts elements from initializer list into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, initializer_list_t ilist)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, "cannot use insert() with " + std::string(type_name()), *this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", *this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, ilist.begin(), ilist.end());
+    }
+
+    /// @brief inserts range of elements into object
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    void insert(const_iterator first, const_iterator last)
+    {
+        // insert only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(309, "cannot use insert() with " + std::string(type_name()), *this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", *this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterators first and last must point to objects", *this));
+        }
+
+        m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_reference j, bool merge_objects = false)
+    {
+        update(j.begin(), j.end(), merge_objects);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_iterator first, const_iterator last, bool merge_objects = false)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(312, "cannot use update() with " + std::string(type_name()), *this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", *this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(type_error::create(312, "cannot use update() with " + std::string(first.m_object->type_name()), *first.m_object));
+        }
+
+        for (auto it = first; it != last; ++it)
+        {
+            if (merge_objects && it.value().is_object())
+            {
+                auto it2 = m_value.object->find(it.key());
+                if (it2 != m_value.object->end())
+                {
+                    it2->second.update(it.value(), true);
+                    continue;
+                }
+            }
+            m_value.object->operator[](it.key()) = it.value();
+#if JSON_DIAGNOSTICS
+            m_value.object->operator[](it.key()).m_parent = this;
+#endif
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(reference other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        std::swap(m_type, other.m_type);
+        std::swap(m_value, other.m_value);
+
+        set_parents();
+        other.set_parents();
+        assert_invariant();
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    friend void swap(reference left, reference right) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        left.swap(right);
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(array_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            std::swap(*(m_value.array), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, "cannot use swap() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(object_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            std::swap(*(m_value.object), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, "cannot use swap() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(string_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_string()))
+        {
+            std::swap(*(m_value.string), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, "cannot use swap() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(binary_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            std::swap(*(m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, "cannot use swap() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(typename binary_t::container_type& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            std::swap(*(m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, "cannot use swap() with " + std::string(type_name()), *this));
+        }
+    }
+
+    /// @}
+
+  public:
+    //////////////////////////////////////////
+    // lexicographical comparison operators //
+    //////////////////////////////////////////
+
+    /// @name lexicographical comparison operators
+    /// @{
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    friend bool operator==(const_reference lhs, const_reference rhs) noexcept
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        const auto lhs_type = lhs.type();
+        const auto rhs_type = rhs.type();
+
+        if (lhs_type == rhs_type)
+        {
+            switch (lhs_type)
+            {
+                case value_t::array:
+                    return *lhs.m_value.array == *rhs.m_value.array;
+
+                case value_t::object:
+                    return *lhs.m_value.object == *rhs.m_value.object;
+
+                case value_t::null:
+                    return true;
+
+                case value_t::string:
+                    return *lhs.m_value.string == *rhs.m_value.string;
+
+                case value_t::boolean:
+                    return lhs.m_value.boolean == rhs.m_value.boolean;
+
+                case value_t::number_integer:
+                    return lhs.m_value.number_integer == rhs.m_value.number_integer;
+
+                case value_t::number_unsigned:
+                    return lhs.m_value.number_unsigned == rhs.m_value.number_unsigned;
+
+                case value_t::number_float:
+                    return lhs.m_value.number_float == rhs.m_value.number_float;
+
+                case value_t::binary:
+                    return *lhs.m_value.binary == *rhs.m_value.binary;
+
+                case value_t::discarded:
+                default:
+                    return false;
+            }
+        }
+        else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float)
+        {
+            return static_cast<number_float_t>(lhs.m_value.number_integer) == rhs.m_value.number_float;
+        }
+        else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer)
+        {
+            return lhs.m_value.number_float == static_cast<number_float_t>(rhs.m_value.number_integer);
+        }
+        else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float)
+        {
+            return static_cast<number_float_t>(lhs.m_value.number_unsigned) == rhs.m_value.number_float;
+        }
+        else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned)
+        {
+            return lhs.m_value.number_float == static_cast<number_float_t>(rhs.m_value.number_unsigned);
+        }
+        else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer)
+        {
+            return static_cast<number_integer_t>(lhs.m_value.number_unsigned) == rhs.m_value.number_integer;
+        }
+        else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned)
+        {
+            return lhs.m_value.number_integer == static_cast<number_integer_t>(rhs.m_value.number_unsigned);
+        }
+
+        return false;
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs == basic_json(rhs);
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) == rhs;
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    friend bool operator!=(const_reference lhs, const_reference rhs) noexcept
+    {
+        return !(lhs == rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs != basic_json(rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) != rhs;
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    friend bool operator<(const_reference lhs, const_reference rhs) noexcept
+    {
+        const auto lhs_type = lhs.type();
+        const auto rhs_type = rhs.type();
+
+        if (lhs_type == rhs_type)
+        {
+            switch (lhs_type)
+            {
+                case value_t::array:
+                    // note parentheses are necessary, see
+                    // https://github.com/nlohmann/json/issues/1530
+                    return (*lhs.m_value.array) < (*rhs.m_value.array);
+
+                case value_t::object:
+                    return (*lhs.m_value.object) < (*rhs.m_value.object);
+
+                case value_t::null:
+                    return false;
+
+                case value_t::string:
+                    return (*lhs.m_value.string) < (*rhs.m_value.string);
+
+                case value_t::boolean:
+                    return (lhs.m_value.boolean) < (rhs.m_value.boolean);
+
+                case value_t::number_integer:
+                    return (lhs.m_value.number_integer) < (rhs.m_value.number_integer);
+
+                case value_t::number_unsigned:
+                    return (lhs.m_value.number_unsigned) < (rhs.m_value.number_unsigned);
+
+                case value_t::number_float:
+                    return (lhs.m_value.number_float) < (rhs.m_value.number_float);
+
+                case value_t::binary:
+                    return (*lhs.m_value.binary) < (*rhs.m_value.binary);
+
+                case value_t::discarded:
+                default:
+                    return false;
+            }
+        }
+        else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float)
+        {
+            return static_cast<number_float_t>(lhs.m_value.number_integer) < rhs.m_value.number_float;
+        }
+        else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer)
+        {
+            return lhs.m_value.number_float < static_cast<number_float_t>(rhs.m_value.number_integer);
+        }
+        else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float)
+        {
+            return static_cast<number_float_t>(lhs.m_value.number_unsigned) < rhs.m_value.number_float;
+        }
+        else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned)
+        {
+            return lhs.m_value.number_float < static_cast<number_float_t>(rhs.m_value.number_unsigned);
+        }
+        else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned)
+        {
+            return lhs.m_value.number_integer < static_cast<number_integer_t>(rhs.m_value.number_unsigned);
+        }
+        else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer)
+        {
+            return static_cast<number_integer_t>(lhs.m_value.number_unsigned) < rhs.m_value.number_integer;
+        }
+
+        // We only reach this line if we cannot compare values. In that case,
+        // we compare types. Note we have to call the operator explicitly,
+        // because MSVC has problems otherwise.
+        return operator<(lhs_type, rhs_type);
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs < basic_json(rhs);
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) < rhs;
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    friend bool operator<=(const_reference lhs, const_reference rhs) noexcept
+    {
+        return !(rhs < lhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs <= basic_json(rhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) <= rhs;
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    friend bool operator>(const_reference lhs, const_reference rhs) noexcept
+    {
+        return !(lhs <= rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs > basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) > rhs;
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    friend bool operator>=(const_reference lhs, const_reference rhs) noexcept
+    {
+        return !(lhs < rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs >= basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) >= rhs;
+    }
+
+    /// @}
+
+    ///////////////////
+    // serialization //
+    ///////////////////
+
+    /// @name serialization
+    /// @{
+#ifndef JSON_NO_IO
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    friend std::ostream& operator<<(std::ostream& o, const basic_json& j)
+    {
+        // read width member and use it as indentation parameter if nonzero
+        const bool pretty_print = o.width() > 0;
+        const auto indentation = pretty_print ? o.width() : 0;
+
+        // reset width to 0 for subsequent calls to this stream
+        o.width(0);
+
+        // do the actual serialization
+        serializer s(detail::output_adapter<char>(o), o.fill());
+        s.dump(j, pretty_print, false, static_cast<unsigned int>(indentation));
+        return o;
+    }
+
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    /// @deprecated This function is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator<<(std::ostream&, const basic_json&) instead; that is,
+    ///             replace calls like `j >> o;` with `o << j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator<<(std::ostream&, const basic_json&))
+    friend std::ostream& operator>>(const basic_json& j, std::ostream& o)
+    {
+        return o << j;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+
+    /////////////////////
+    // deserialization //
+    /////////////////////
+
+    /// @name deserialization
+    /// @{
+
+    /// @brief deserialize from a compatible input
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(InputType&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::forward<InputType>(i)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief deserialize from a pair of character iterators
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(IteratorType first,
+                            IteratorType last,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::move(first), std::move(last)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, parse(ptr, ptr + len))
+    static basic_json parse(detail::span_input_adapter&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(i.get(), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename InputType>
+    static bool accept(InputType&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::forward<InputType>(i)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename IteratorType>
+    static bool accept(IteratorType first, IteratorType last,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::move(first), std::move(last)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, accept(ptr, ptr + len))
+    static bool accept(detail::span_input_adapter&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(i.get(), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template <typename InputType, typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(InputType&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia)).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template<class IteratorType, class SAX>
+    JSON_HEDLEY_NON_NULL(3)
+    static bool sax_parse(IteratorType first, IteratorType last, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia)).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    /// @deprecated This function is deprecated since 3.8.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             sax_parse(ptr, ptr + len) instead.
+    template <typename SAX>
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, sax_parse(ptr, ptr + len, ...))
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(detail::span_input_adapter&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = i.get();
+        return format == input_format_t::json
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia)).sax_parse(format, sax, strict);
+    }
+#ifndef JSON_NO_IO
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    /// @deprecated This stream operator is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator>>(std::istream&, basic_json&) instead; that is,
+    ///             replace calls like `j << i;` with `i >> j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator>>(std::istream&, basic_json&))
+    friend std::istream& operator<<(basic_json& j, std::istream& i)
+    {
+        return operator>>(i, j);
+    }
+
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    friend std::istream& operator>>(std::istream& i, basic_json& j)
+    {
+        parser(detail::input_adapter(i)).parse(false, j);
+        return i;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+    ///////////////////////////
+    // convenience functions //
+    ///////////////////////////
+
+    /// @brief return the type as string
+    /// @sa https://json.nlohmann.me/api/basic_json/type_name/
+    JSON_HEDLEY_RETURNS_NON_NULL
+    const char* type_name() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+                return "null";
+            case value_t::object:
+                return "object";
+            case value_t::array:
+                return "array";
+            case value_t::string:
+                return "string";
+            case value_t::boolean:
+                return "boolean";
+            case value_t::binary:
+                return "binary";
+            case value_t::discarded:
+                return "discarded";
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            default:
+                return "number";
+        }
+    }
+
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    //////////////////////
+    // member variables //
+    //////////////////////
+
+    /// the type of the current element
+    value_t m_type = value_t::null;
+
+    /// the value of the current element
+    json_value m_value = {};
+
+#if JSON_DIAGNOSTICS
+    /// a pointer to a parent value (for debugging purposes)
+    basic_json* m_parent = nullptr;
+#endif
+
+    //////////////////////////////////////////
+    // binary serialization/deserialization //
+    //////////////////////////////////////////
+
+    /// @name binary serialization/deserialization support
+    /// @{
+
+  public:
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static std::vector<std::uint8_t> to_cbor(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_cbor(j, result);
+        return result;
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_cbor(j);
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_cbor(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static std::vector<std::uint8_t> to_msgpack(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_msgpack(j, result);
+        return result;
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_msgpack(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_msgpack(j);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static std::vector<std::uint8_t> to_ubjson(const basic_json& j,
+            const bool use_size = false,
+            const bool use_type = false)
+    {
+        std::vector<std::uint8_t> result;
+        to_ubjson(j, result, use_size, use_type);
+        return result;
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<std::uint8_t> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<char> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<char>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static std::vector<std::uint8_t> to_bson(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_bson(j, result);
+        return result;
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_bson(j);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_bson(j);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        return from_cbor(ptr, ptr + len, strict, allow_exceptions, tag_handler);
+    }
+
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(InputType&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(IteratorType first, IteratorType last,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(const T* ptr, std::size_t len,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        return from_msgpack(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(detail::span_input_adapter&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(InputType&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(IteratorType first, IteratorType last,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(const T* ptr, std::size_t len,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        return from_ubjson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(detail::span_input_adapter&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        return from_bson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia)).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+    /// @}
+
+    //////////////////////////
+    // JSON Pointer support //
+    //////////////////////////
+
+    /// @name JSON Pointer functions
+    /// @{
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](const json_pointer& ptr)
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const json_pointer& ptr) const
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const json_pointer& ptr)
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const json_pointer& ptr) const
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief return flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/flatten/
+    basic_json flatten() const
+    {
+        basic_json result(value_t::object);
+        json_pointer::flatten("", *this, result);
+        return result;
+    }
+
+    /// @brief unflatten a previously flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/unflatten/
+    basic_json unflatten() const
+    {
+        return json_pointer::unflatten(*this);
+    }
+
+    /// @}
+
+    //////////////////////////
+    // JSON Patch functions //
+    //////////////////////////
+
+    /// @name JSON Patch functions
+    /// @{
+
+    /// @brief applies a JSON patch
+    /// @sa https://json.nlohmann.me/api/basic_json/patch/
+    basic_json patch(const basic_json& json_patch) const
+    {
+        // make a working copy to apply the patch to
+        basic_json result = *this;
+
+        // the valid JSON Patch operations
+        enum class patch_operations {add, remove, replace, move, copy, test, invalid};
+
+        const auto get_op = [](const std::string & op)
+        {
+            if (op == "add")
+            {
+                return patch_operations::add;
+            }
+            if (op == "remove")
+            {
+                return patch_operations::remove;
+            }
+            if (op == "replace")
+            {
+                return patch_operations::replace;
+            }
+            if (op == "move")
+            {
+                return patch_operations::move;
+            }
+            if (op == "copy")
+            {
+                return patch_operations::copy;
+            }
+            if (op == "test")
+            {
+                return patch_operations::test;
+            }
+
+            return patch_operations::invalid;
+        };
+
+        // wrapper for "add" operation; add value at ptr
+        const auto operation_add = [&result](json_pointer & ptr, basic_json val)
+        {
+            // adding to the root of the target document means replacing it
+            if (ptr.empty())
+            {
+                result = val;
+                return;
+            }
+
+            // make sure the top element of the pointer exists
+            json_pointer top_pointer = ptr.top();
+            if (top_pointer != ptr)
+            {
+                result.at(top_pointer);
+            }
+
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            basic_json& parent = result[ptr];
+
+            switch (parent.m_type)
+            {
+                case value_t::null:
+                case value_t::object:
+                {
+                    // use operator[] to add value
+                    parent[last_path] = val;
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    if (last_path == "-")
+                    {
+                        // special case: append to back
+                        parent.push_back(val);
+                    }
+                    else
+                    {
+                        const auto idx = json_pointer::array_index(last_path);
+                        if (JSON_HEDLEY_UNLIKELY(idx > parent.size()))
+                        {
+                            // avoid undefined behavior
+                            JSON_THROW(out_of_range::create(401, "array index " + std::to_string(idx) + " is out of range", parent));
+                        }
+
+                        // default case: insert add offset
+                        parent.insert(parent.begin() + static_cast<difference_type>(idx), val);
+                    }
+                    break;
+                }
+
+                // if there exists a parent it cannot be primitive
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        };
+
+        // wrapper for "remove" operation; remove value at ptr
+        const auto operation_remove = [this, &result](json_pointer & ptr)
+        {
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            basic_json& parent = result.at(ptr);
+
+            // remove child
+            if (parent.is_object())
+            {
+                // perform range check
+                auto it = parent.find(last_path);
+                if (JSON_HEDLEY_LIKELY(it != parent.end()))
+                {
+                    parent.erase(it);
+                }
+                else
+                {
+                    JSON_THROW(out_of_range::create(403, "key '" + last_path + "' not found", *this));
+                }
+            }
+            else if (parent.is_array())
+            {
+                // note erase performs range check
+                parent.erase(json_pointer::array_index(last_path));
+            }
+        };
+
+        // type check: top level value must be an array
+        if (JSON_HEDLEY_UNLIKELY(!json_patch.is_array()))
+        {
+            JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", json_patch));
+        }
+
+        // iterate and apply the operations
+        for (const auto& val : json_patch)
+        {
+            // wrapper to get a value for an operation
+            const auto get_value = [&val](const std::string & op,
+                                          const std::string & member,
+                                          bool string_type) -> basic_json &
+            {
+                // find value
+                auto it = val.m_value.object->find(member);
+
+                // context-sensitive error message
+                const auto error_msg = (op == "op") ? "operation" : "operation '" + op + "'";
+
+                // check if desired value is present
+                if (JSON_HEDLEY_UNLIKELY(it == val.m_value.object->end()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, error_msg + " must has member '" + member + "'", val));
+                }
+
+                // check if result is of type string
+                if (JSON_HEDLEY_UNLIKELY(string_type && !it->second.is_string()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, error_msg + " must has string member '" + member + "'", val));
+                }
+
+                // no error: return value
+                return it->second;
+            };
+
+            // type check: every element of the array must be an object
+            if (JSON_HEDLEY_UNLIKELY(!val.is_object()))
+            {
+                JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", val));
+            }
+
+            // collect mandatory members
+            const auto op = get_value("op", "op", true).template get<std::string>();
+            const auto path = get_value(op, "path", true).template get<std::string>();
+            json_pointer ptr(path);
+
+            switch (get_op(op))
+            {
+                case patch_operations::add:
+                {
+                    operation_add(ptr, get_value("add", "value", false));
+                    break;
+                }
+
+                case patch_operations::remove:
+                {
+                    operation_remove(ptr);
+                    break;
+                }
+
+                case patch_operations::replace:
+                {
+                    // the "path" location must exist - use at()
+                    result.at(ptr) = get_value("replace", "value", false);
+                    break;
+                }
+
+                case patch_operations::move:
+                {
+                    const auto from_path = get_value("move", "from", true).template get<std::string>();
+                    json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json v = result.at(from_ptr);
+
+                    // The move operation is functionally identical to a
+                    // "remove" operation on the "from" location, followed
+                    // immediately by an "add" operation at the target
+                    // location with the value that was just removed.
+                    operation_remove(from_ptr);
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::copy:
+                {
+                    const auto from_path = get_value("copy", "from", true).template get<std::string>();
+                    const json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json v = result.at(from_ptr);
+
+                    // The copy is functionally identical to an "add"
+                    // operation at the target location using the value
+                    // specified in the "from" member.
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::test:
+                {
+                    bool success = false;
+                    JSON_TRY
+                    {
+                        // check if "value" matches the one at "path"
+                        // the "path" location must exist - use at()
+                        success = (result.at(ptr) == get_value("test", "value", false));
+                    }
+                    JSON_INTERNAL_CATCH (out_of_range&)
+                    {
+                        // ignore out of range errors: success remains false
+                    }
+
+                    // throw an exception if test fails
+                    if (JSON_HEDLEY_UNLIKELY(!success))
+                    {
+                        JSON_THROW(other_error::create(501, "unsuccessful: " + val.dump(), val));
+                    }
+
+                    break;
+                }
+
+                case patch_operations::invalid:
+                default:
+                {
+                    // op must be "add", "remove", "replace", "move", "copy", or
+                    // "test"
+                    JSON_THROW(parse_error::create(105, 0, "operation value '" + op + "' is invalid", val));
+                }
+            }
+        }
+
+        return result;
+    }
+
+    /// @brief creates a diff as a JSON patch
+    /// @sa https://json.nlohmann.me/api/basic_json/diff/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json diff(const basic_json& source, const basic_json& target,
+                           const std::string& path = "")
+    {
+        // the patch
+        basic_json result(value_t::array);
+
+        // if the values are the same, return empty patch
+        if (source == target)
+        {
+            return result;
+        }
+
+        if (source.type() != target.type())
+        {
+            // different types: replace value
+            result.push_back(
+            {
+                {"op", "replace"}, {"path", path}, {"value", target}
+            });
+            return result;
+        }
+
+        switch (source.type())
+        {
+            case value_t::array:
+            {
+                // first pass: traverse common elements
+                std::size_t i = 0;
+                while (i < source.size() && i < target.size())
+                {
+                    // recursive call to compare array values at index i
+                    auto temp_diff = diff(source[i], target[i], path + "/" + std::to_string(i));
+                    result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    ++i;
+                }
+
+                // We now reached the end of at least one array
+                // in a second pass, traverse the remaining elements
+
+                // remove my remaining elements
+                const auto end_index = static_cast<difference_type>(result.size());
+                while (i < source.size())
+                {
+                    // add operations in reverse order to avoid invalid
+                    // indices
+                    result.insert(result.begin() + end_index, object(
+                    {
+                        {"op", "remove"},
+                        {"path", path + "/" + std::to_string(i)}
+                    }));
+                    ++i;
+                }
+
+                // add other remaining elements
+                while (i < target.size())
+                {
+                    result.push_back(
+                    {
+                        {"op", "add"},
+                        {"path", path + "/-"},
+                        {"value", target[i]}
+                    });
+                    ++i;
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // first pass: traverse this object's elements
+                for (auto it = source.cbegin(); it != source.cend(); ++it)
+                {
+                    // escape the key name to be used in a JSON patch
+                    const auto path_key = path + "/" + detail::escape(it.key());
+
+                    if (target.find(it.key()) != target.end())
+                    {
+                        // recursive call to compare object values at key it
+                        auto temp_diff = diff(it.value(), target[it.key()], path_key);
+                        result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    }
+                    else
+                    {
+                        // found a key that is not in o -> remove it
+                        result.push_back(object(
+                        {
+                            {"op", "remove"}, {"path", path_key}
+                        }));
+                    }
+                }
+
+                // second pass: traverse other object's elements
+                for (auto it = target.cbegin(); it != target.cend(); ++it)
+                {
+                    if (source.find(it.key()) == source.end())
+                    {
+                        // found a key that is not in this -> add it
+                        const auto path_key = path + "/" + detail::escape(it.key());
+                        result.push_back(
+                        {
+                            {"op", "add"}, {"path", path_key},
+                            {"value", it.value()}
+                        });
+                    }
+                }
+
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // both primitive type: replace value
+                result.push_back(
+                {
+                    {"op", "replace"}, {"path", path}, {"value", target}
+                });
+                break;
+            }
+        }
+
+        return result;
+    }
+
+    /// @}
+
+    ////////////////////////////////
+    // JSON Merge Patch functions //
+    ////////////////////////////////
+
+    /// @name JSON Merge Patch functions
+    /// @{
+
+    /// @brief applies a JSON Merge Patch
+    /// @sa https://json.nlohmann.me/api/basic_json/merge_patch/
+    void merge_patch(const basic_json& apply_patch)
+    {
+        if (apply_patch.is_object())
+        {
+            if (!is_object())
+            {
+                *this = object();
+            }
+            for (auto it = apply_patch.begin(); it != apply_patch.end(); ++it)
+            {
+                if (it.value().is_null())
+                {
+                    erase(it.key());
+                }
+                else
+                {
+                    operator[](it.key()).merge_patch(it.value());
+                }
+            }
+        }
+        else
+        {
+            *this = apply_patch;
+        }
+    }
+
+    /// @}
+};
+
+/// @brief user-defined to_string function for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/to_string/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+std::string to_string(const NLOHMANN_BASIC_JSON_TPL& j)
+{
+    return j.dump();
+}
+
+} // namespace nlohmann
+
+///////////////////////
+// nonmember support //
+///////////////////////
+
+namespace std // NOLINT(cert-dcl58-cpp)
+{
+
+/// @brief hash value for JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_hash/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct hash<nlohmann::NLOHMANN_BASIC_JSON_TPL>
+{
+    std::size_t operator()(const nlohmann::NLOHMANN_BASIC_JSON_TPL& j) const
+    {
+        return nlohmann::detail::hash(j);
+    }
+};
+
+// specialization for std::less<value_t>
+template<>
+struct less< ::nlohmann::detail::value_t> // do not remove the space after '<', see https://github.com/nlohmann/json/pull/679
+{
+    /*!
+    @brief compare two value_t enum values
+    @since version 3.0.0
+    */
+    bool operator()(nlohmann::detail::value_t lhs,
+                    nlohmann::detail::value_t rhs) const noexcept
+    {
+        return nlohmann::detail::operator<(lhs, rhs);
+    }
+};
+
+// C++20 prohibit function specialization in the std namespace.
+#ifndef JSON_HAS_CPP_20
+
+/// @brief exchanges the values of two JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_swap/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept(  // NOLINT(readability-inconsistent-declaration-parameter-name)
+    is_nothrow_move_constructible<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value&&                          // NOLINT(misc-redundant-expression)
+    is_nothrow_move_assignable<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value)
+{
+    j1.swap(j2);
+}
+
+#endif
+
+} // namespace std
+
+/// @brief user-defined string literal for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/
+JSON_HEDLEY_NON_NULL(1)
+inline nlohmann::json operator "" _json(const char* s, std::size_t n)
+{
+    return nlohmann::json::parse(s, s + n);
+}
+
+/// @brief user-defined string literal for JSON pointer
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/
+JSON_HEDLEY_NON_NULL(1)
+inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n)
+{
+    return nlohmann::json::json_pointer(std::string(s, n));
+}
+
+// #include <nlohmann/detail/macro_unscope.hpp>
+
+
+// restore clang diagnostic settings
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+
+// clean up
+#undef JSON_ASSERT
+#undef JSON_INTERNAL_CATCH
+#undef JSON_CATCH
+#undef JSON_THROW
+#undef JSON_TRY
+#undef JSON_PRIVATE_UNLESS_TESTED
+#undef JSON_HAS_CPP_11
+#undef JSON_HAS_CPP_14
+#undef JSON_HAS_CPP_17
+#undef JSON_HAS_CPP_20
+#undef JSON_HAS_FILESYSTEM
+#undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+#undef NLOHMANN_BASIC_JSON_TPL_DECLARATION
+#undef NLOHMANN_BASIC_JSON_TPL
+#undef JSON_EXPLICIT
+#undef NLOHMANN_CAN_CALL_STD_FUNC_IMPL
+
+// #include <nlohmann/thirdparty/hedley/hedley_undef.hpp>
+
+
+#undef JSON_HEDLEY_ALWAYS_INLINE
+#undef JSON_HEDLEY_ARM_VERSION
+#undef JSON_HEDLEY_ARM_VERSION_CHECK
+#undef JSON_HEDLEY_ARRAY_PARAM
+#undef JSON_HEDLEY_ASSUME
+#undef JSON_HEDLEY_BEGIN_C_DECLS
+#undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#undef JSON_HEDLEY_CLANG_HAS_WARNING
+#undef JSON_HEDLEY_COMPCERT_VERSION
+#undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#undef JSON_HEDLEY_CONCAT
+#undef JSON_HEDLEY_CONCAT3
+#undef JSON_HEDLEY_CONCAT3_EX
+#undef JSON_HEDLEY_CONCAT_EX
+#undef JSON_HEDLEY_CONST
+#undef JSON_HEDLEY_CONSTEXPR
+#undef JSON_HEDLEY_CONST_CAST
+#undef JSON_HEDLEY_CPP_CAST
+#undef JSON_HEDLEY_CRAY_VERSION
+#undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#undef JSON_HEDLEY_C_DECL
+#undef JSON_HEDLEY_DEPRECATED
+#undef JSON_HEDLEY_DEPRECATED_FOR
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#undef JSON_HEDLEY_DIAGNOSTIC_POP
+#undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#undef JSON_HEDLEY_DMC_VERSION
+#undef JSON_HEDLEY_DMC_VERSION_CHECK
+#undef JSON_HEDLEY_EMPTY_BASES
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#undef JSON_HEDLEY_END_C_DECLS
+#undef JSON_HEDLEY_FLAGS
+#undef JSON_HEDLEY_FLAGS_CAST
+#undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#undef JSON_HEDLEY_GCC_HAS_FEATURE
+#undef JSON_HEDLEY_GCC_HAS_WARNING
+#undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#undef JSON_HEDLEY_GCC_VERSION
+#undef JSON_HEDLEY_GCC_VERSION_CHECK
+#undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#undef JSON_HEDLEY_GNUC_HAS_WARNING
+#undef JSON_HEDLEY_GNUC_VERSION
+#undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#undef JSON_HEDLEY_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_BUILTIN
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_EXTENSION
+#undef JSON_HEDLEY_HAS_FEATURE
+#undef JSON_HEDLEY_HAS_WARNING
+#undef JSON_HEDLEY_IAR_VERSION
+#undef JSON_HEDLEY_IAR_VERSION_CHECK
+#undef JSON_HEDLEY_IBM_VERSION
+#undef JSON_HEDLEY_IBM_VERSION_CHECK
+#undef JSON_HEDLEY_IMPORT
+#undef JSON_HEDLEY_INLINE
+#undef JSON_HEDLEY_INTEL_CL_VERSION
+#undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#undef JSON_HEDLEY_INTEL_VERSION
+#undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#undef JSON_HEDLEY_IS_CONSTANT
+#undef JSON_HEDLEY_IS_CONSTEXPR_
+#undef JSON_HEDLEY_LIKELY
+#undef JSON_HEDLEY_MALLOC
+#undef JSON_HEDLEY_MCST_LCC_VERSION
+#undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#undef JSON_HEDLEY_MESSAGE
+#undef JSON_HEDLEY_MSVC_VERSION
+#undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#undef JSON_HEDLEY_NEVER_INLINE
+#undef JSON_HEDLEY_NON_NULL
+#undef JSON_HEDLEY_NO_ESCAPE
+#undef JSON_HEDLEY_NO_RETURN
+#undef JSON_HEDLEY_NO_THROW
+#undef JSON_HEDLEY_NULL
+#undef JSON_HEDLEY_PELLES_VERSION
+#undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#undef JSON_HEDLEY_PGI_VERSION
+#undef JSON_HEDLEY_PGI_VERSION_CHECK
+#undef JSON_HEDLEY_PREDICT
+#undef JSON_HEDLEY_PRINTF_FORMAT
+#undef JSON_HEDLEY_PRIVATE
+#undef JSON_HEDLEY_PUBLIC
+#undef JSON_HEDLEY_PURE
+#undef JSON_HEDLEY_REINTERPRET_CAST
+#undef JSON_HEDLEY_REQUIRE
+#undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#undef JSON_HEDLEY_REQUIRE_MSG
+#undef JSON_HEDLEY_RESTRICT
+#undef JSON_HEDLEY_RETURNS_NON_NULL
+#undef JSON_HEDLEY_SENTINEL
+#undef JSON_HEDLEY_STATIC_ASSERT
+#undef JSON_HEDLEY_STATIC_CAST
+#undef JSON_HEDLEY_STRINGIFY
+#undef JSON_HEDLEY_STRINGIFY_EX
+#undef JSON_HEDLEY_SUNPRO_VERSION
+#undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#undef JSON_HEDLEY_TINYC_VERSION
+#undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#undef JSON_HEDLEY_TI_ARMCL_VERSION
+#undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL2000_VERSION
+#undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL430_VERSION
+#undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL6X_VERSION
+#undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL7X_VERSION
+#undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CLPRU_VERSION
+#undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#undef JSON_HEDLEY_TI_VERSION
+#undef JSON_HEDLEY_TI_VERSION_CHECK
+#undef JSON_HEDLEY_UNAVAILABLE
+#undef JSON_HEDLEY_UNLIKELY
+#undef JSON_HEDLEY_UNPREDICTABLE
+#undef JSON_HEDLEY_UNREACHABLE
+#undef JSON_HEDLEY_UNREACHABLE_RETURN
+#undef JSON_HEDLEY_VERSION
+#undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#undef JSON_HEDLEY_VERSION_ENCODE
+#undef JSON_HEDLEY_WARNING
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#undef JSON_HEDLEY_FALL_THROUGH
+
+
+
+#endif  // INCLUDE_NLOHMANN_JSON_HPP_
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/system.h b/cpp-package/inspireface/cpp/inspireface/middleware/system.h
new file mode 100644
index 000000000..65efe1e36
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/system.h
@@ -0,0 +1,141 @@
+#ifndef INSPIRE_FACE_SYSTEM_H
+#define INSPIRE_FACE_SYSTEM_H
+
+#include <string>
+#include <utility>
+#include <vector>
+#include <memory>
+#include <string>
+#include <initializer_list>
+#if defined(_WIN32)
+#include <windows.h>
+#else
+#include <dirent.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#endif
+
+namespace inspire {
+namespace os {
+
+template <typename... Args>
+inline std::string PathJoin(Args... args) {
+    std::string result;
+#if defined(_WIN32)
+    const char sep = '\\';
+#else
+    const char sep = '/';
+#endif
+    std::vector<std::string> paths = {args...};
+    bool first = true;
+    for (const auto& p : paths) {
+        if (p.empty())
+            continue;
+
+        if (first) {
+            result = p;
+            first = false;
+            continue;
+        }
+
+        if (result.back() != sep)
+            result += sep;
+        result += p;
+    }
+
+    return result;
+}
+
+inline std::pair<std::string, std::string> PathSplit(const std::string& path) {
+    std::string directory, filename;
+    size_t pos = path.rfind('/');
+
+#ifdef _WIN32
+    size_t backslash_pos = path.rfind('\\');
+    if ((pos == std::string::npos) || (backslash_pos != std::string::npos && backslash_pos > pos)) {
+        pos = backslash_pos;
+    }
+#endif
+
+    if (pos == std::string::npos) {
+        directory = "";
+        filename = path;
+    } else if (pos == 0) {
+        directory = path.substr(0, 1);
+        filename = path.substr(1);
+    } else {
+        directory = path.substr(0, pos);
+        filename = path.substr(pos + 1);
+    }
+
+    return {directory, filename};
+}
+
+inline std::pair<std::string, std::string> SplitExt(const std::string& path) {
+    std::string basename, extension;
+
+    size_t pos = path.rfind('.');
+
+    if (pos == std::string::npos) {
+        basename = path;
+        extension = "";
+    } else {
+        basename = path.substr(0, pos);
+        extension = path.substr(pos);
+    }
+
+    return {basename, extension};
+}
+
+inline std::string Dirname(const std::string& path) {
+    return PathSplit(path).first;
+}
+
+inline std::string Basename(const std::string& path) {
+    return PathSplit(path).second;
+}
+
+#if defined(_WIN32)
+inline std::wstring Utf8ToWideChar(const std::string& utf8str) {
+    int size_required = MultiByteToWideChar(CP_UTF8, 0, utf8str.c_str(), utf8str.size(), NULL, 0);
+    std::wstring ws_translated_str(size_required, 0);
+    MultiByteToWideChar(CP_UTF8, 0, utf8str.c_str(), utf8str.size(), &ws_translated_str[0], size_required);
+    return ws_translated_str;
+}
+#endif
+
+inline bool IsExists(const std::string& path) {
+#if defined(_WIN32)
+    std::wstring ws_path = Utf8ToWideChar(path);
+    return _waccess(ws_path.c_str(), 0) == 0;
+#else
+    return access(path.c_str(), F_OK) == 0;
+#endif
+}
+
+inline bool IsDir(const std::string& path) {
+#if defined(_WIN32)
+    return GetFileAttributesA(path.c_str()) == FILE_ATTRIBUTE_DIRECTORY;
+#else
+    struct stat s;
+    if (stat(path.c_str(), &s) != 0)
+        return false;
+    return S_ISDIR(s.st_mode);
+#endif
+}
+
+inline bool IsFile(const std::string& path) {
+#if defined(_WIN32)
+    return GetFileAttributesA(path.c_str()) == FILE_ATTRIBUTE_ARCHIVE;
+#else
+    struct stat s;
+    if (stat(path.c_str(), &s) != 0)
+        return false;
+    return S_ISREG(s.st_mode);
+#endif
+}
+
+}  // namespace os
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_SYSTEM_H
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/thread/resource_pool.h b/cpp-package/inspireface/cpp/inspireface/middleware/thread/resource_pool.h
new file mode 100644
index 000000000..7a4d3611a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/thread/resource_pool.h
@@ -0,0 +1,148 @@
+#ifndef INSPIRE_RESOURCE_POOL_H
+#define INSPIRE_RESOURCE_POOL_H
+
+#include <iostream>
+#include <mutex>
+#include <queue>
+#include <condition_variable>
+#include <memory>
+#include <functional>
+
+namespace inspire {
+namespace parallel {
+
+/**
+ * @brief ResourcePool is a thread-safe resource pool that can be used to manage resources in a multi-threaded environment.
+ * @tparam Resource The type of the resource to be managed.
+ */
+template <typename Resource>
+class ResourcePool {
+public:
+    using ResourceDeleter = std::function<void(Resource&)>;
+
+    class ResourceGuard {
+    public:
+        ResourceGuard(Resource& resource, ResourcePool& pool) : m_resource(resource), m_pool(pool), m_valid(true) {}
+
+        // Move constructor
+        ResourceGuard(ResourceGuard&& other) noexcept : m_resource(other.m_resource), m_pool(other.m_pool), m_valid(other.m_valid) {
+            other.m_valid = false;
+        }
+
+        // Disable copy
+        ResourceGuard(const ResourceGuard&) = delete;
+        ResourceGuard& operator=(const ResourceGuard&) = delete;
+
+        ~ResourceGuard() {
+            if (m_valid) {
+                m_pool.ReturnResource(std::move(m_resource));
+            }
+        }
+
+        Resource* operator->() {
+            return &m_resource;
+        }
+
+        Resource& operator*() {
+            return m_resource;
+        }
+
+    private:
+        Resource& m_resource;
+        ResourcePool& m_pool;
+        bool m_valid;
+    };
+
+    explicit ResourcePool(size_t size, ResourceDeleter deleter = nullptr) : m_deleter(deleter) {
+        m_resources.reserve(size);
+    }
+
+    ~ResourcePool() {
+        if (m_deleter) {
+            std::lock_guard<std::mutex> lock(m_mutex);
+            for (auto& resource : m_resources) {
+                m_deleter(resource);
+            }
+        }
+    }
+
+    void AddResource(Resource&& resource) {
+        std::lock_guard<std::mutex> lock(m_mutex);
+        m_resources.push_back(std::move(resource));
+        m_available_resources.push(&m_resources.back());
+        m_cv.notify_one();
+    }
+
+    // Acquire resource (blocking mode)
+    ResourceGuard AcquireResource() {
+        std::unique_lock<std::mutex> lock(m_mutex);
+        m_cv.wait(lock, [this] { return !m_available_resources.empty(); });
+
+        Resource* resource = m_available_resources.front();
+        m_available_resources.pop();
+
+        return ResourceGuard(*resource, *this);
+    }
+
+    // Try to acquire resource (non-blocking mode), returns nullptr if no resource available
+    std::unique_ptr<ResourceGuard> TryAcquireResource() {
+        std::lock_guard<std::mutex> lock(m_mutex);
+        if (m_available_resources.empty()) {
+            return nullptr;
+        }
+
+        Resource* resource = m_available_resources.front();
+        m_available_resources.pop();
+
+        return std::unique_ptr<ResourceGuard>(new ResourceGuard(*resource, *this));
+    }
+
+    // Acquire resource with timeout, returns nullptr if timeout
+    std::unique_ptr<ResourceGuard> AcquireResource(std::chrono::milliseconds timeout) {
+        std::unique_lock<std::mutex> lock(m_mutex);
+        if (!m_cv.wait_for(lock, timeout, [this] { return !m_available_resources.empty(); })) {
+            return nullptr;
+        }
+
+        Resource* resource = m_available_resources.front();
+        m_available_resources.pop();
+
+        return std::unique_ptr<ResourceGuard>(new ResourceGuard(*resource, *this));
+    }
+
+    size_t AvailableCount() const {
+        std::lock_guard<std::mutex> lock(m_mutex);
+        return m_available_resources.size();
+    }
+
+    size_t TotalCount() const {
+        std::lock_guard<std::mutex> lock(m_mutex);
+        return m_resources.size();
+    }
+
+private:
+    void ReturnResource(Resource&& resource) {
+        std::lock_guard<std::mutex> lock(m_mutex);
+        for (auto& stored_resource : m_resources) {
+            if (&stored_resource == &resource) {
+                m_available_resources.push(&stored_resource);
+                m_cv.notify_one();
+                break;
+            }
+        }
+    }
+
+private:
+    mutable std::mutex m_mutex;
+    std::condition_variable m_cv;
+    std::vector<Resource> m_resources;            // Store actual resources
+    std::queue<Resource*> m_available_resources;  // Queue of available resources
+    ResourceDeleter m_deleter;                    // Resource cleanup callback
+
+    friend class ResourceGuard;
+};
+
+}  // namespace parallel
+}  // namespace inspire
+
+#endif  // INSPIRE_RESOURCE_POOL_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/timer.cpp b/cpp-package/inspireface/cpp/inspireface/middleware/timer.cpp
new file mode 100644
index 000000000..693d34cf2
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/timer.cpp
@@ -0,0 +1,111 @@
+#include "spend_timer.h"
+#include <ostream>
+#include <sstream>
+#include "log.h"
+
+#if defined(_MSC_VER)
+#include <chrono>  // NOLINT
+#else
+#include <sys/time.h>
+#endif
+
+namespace inspire {
+
+#if defined(_MSC_VER)
+
+uint64_t INSPIRE_API_EXPORT _now() {
+    return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
+}
+
+#else
+
+uint64_t INSPIRE_API_EXPORT _now() {
+    struct timeval tv;
+    gettimeofday(&tv, nullptr);
+    return static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
+}
+
+#endif  // defined(_MSC_VER)
+
+int SpendTimer::is_enable = true;
+
+SpendTimer::SpendTimer() {
+    Reset();
+}
+
+SpendTimer::SpendTimer(const std::string &name) {
+    name_ = name;
+    Reset();
+}
+
+void SpendTimer::Start() {
+    start_ = _now();
+}
+
+void SpendTimer::Stop() {
+    stop_ = _now();
+    uint64_t d = stop_ - start_;
+    total_ += d;
+    ++count_;
+    min_ = std::min(min_, d);
+    max_ = std::max(max_, d);
+}
+
+void SpendTimer::Reset() {
+    start_ = 0;
+    stop_ = 0;
+    total_ = 0;
+    count_ = 0;
+    min_ = UINT64_MAX;
+    max_ = 0;
+}
+
+uint64_t SpendTimer::Get() const {
+    return stop_ - start_;
+}
+
+uint64_t SpendTimer::Average() const {
+    return count_ == 0 ? 0 : total_ / count_;
+}
+
+uint64_t SpendTimer::Total() const {
+    return total_;
+}
+
+uint64_t SpendTimer::Count() const {
+    return count_;
+}
+
+uint64_t SpendTimer::Min() const {
+    return count_ == 0 ? 0 : min_;
+}
+
+uint64_t SpendTimer::Max() const {
+    return max_;
+}
+
+const std::string &SpendTimer::name() const {
+    return name_;
+}
+
+std::string SpendTimer::Report() const {
+    std::stringstream ss;
+    if (is_enable) {
+        ss << "[Time(us) Total:" << Total() << " Ave:" << Average() << " Min:" << Min() << " Max:" << Max() << " Count:" << Count() << " " << name_
+           << "]";
+    } else {
+        ss << "Timer Disabled.";
+    }
+    return ss.str();
+}
+
+void SpendTimer::Disable() {
+    is_enable = false;
+}
+
+std::ostream &operator<<(std::ostream &os, const SpendTimer &timer) {
+    os << timer.Report();
+    return os;
+}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/middleware/utils.h b/cpp-package/inspireface/cpp/inspireface/middleware/utils.h
new file mode 100755
index 000000000..081f3bc3f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/middleware/utils.h
@@ -0,0 +1,227 @@
+
+#ifndef INSPIRE_FACE_UTILS_H
+#define INSPIRE_FACE_UTILS_H
+
+#include <cmath>
+#include <iostream>
+#include <string>
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#endif
+
+#include <inspirecv/inspirecv.h>
+
+namespace inspire {
+
+
+inline bool IsDirectory(const std::string& path) {
+#ifdef _WIN32
+    DWORD dwAttrib = GetFileAttributes(path.c_str());
+    return (dwAttrib != INVALID_FILE_ATTRIBUTES && (dwAttrib & FILE_ATTRIBUTE_DIRECTORY));
+#else
+    struct stat st;
+    if (stat(path.c_str(), &st) == 0) {
+        return S_ISDIR(st.st_mode);
+    } else {
+        return false;
+    }
+#endif
+}
+
+
+inline inspirecv::TransformMatrix SquareToSquare(inspirecv::Rect2f src, inspirecv::Rect2f dst,
+                              float win_size = 112) {
+    float src_x1 = static_cast<float>(src.GetX());
+    float src_y1 = static_cast<float>(src.GetY());
+    float src_x2 = static_cast<float>(src.GetX() + src.GetWidth());
+    float src_y2 = static_cast<float>(src.GetY() + src.GetHeight());
+
+    float dst_x1 = static_cast<float>(dst.GetX());
+    float dst_y1 = static_cast<float>(dst.GetY());
+    float dst_x2 = static_cast<float>(dst.GetX() + dst.GetWidth());
+    float dst_y2 = static_cast<float>(dst.GetY() + dst.GetHeight());
+
+    std::vector<inspirecv::Point2f> src_pts = {
+            {src_x1, src_y1},
+            {src_x2, src_y1},
+            {src_x2, src_y2},
+            {src_x1, src_y2}};
+    std::vector<inspirecv::Point2f> dst_pts = {
+            {dst_x1, dst_y1},
+            {dst_x2, dst_y1},
+            {dst_x2, dst_y2},
+            {dst_x1, dst_y2}};
+    inspirecv::TransformMatrix m = inspirecv::SimilarityTransformEstimate(src_pts, dst_pts);
+    return m;
+}
+
+inline std::vector<inspirecv::Point2f>
+FixPointsMeanshape(std::vector<inspirecv::Point2f> &points,
+                   const std::vector<inspirecv::Point2f> &mean_shape) {
+
+    inspirecv::Rect2f bbox = inspirecv::MinBoundingRect(points);
+    int R = std::max(bbox.GetHeight(), bbox.GetWidth());
+    int cx = bbox.GetX() + bbox.GetWidth() / 2;
+    int cy = bbox.GetY() + bbox.GetHeight() / 2;
+    inspirecv::Rect2f old(cx - R / 2, cy - R / 2, R, R);
+
+    inspirecv::Rect2f mean_shape_box = inspirecv::MinBoundingRect(mean_shape);
+    int m_R = std::max(mean_shape_box.GetHeight(), mean_shape_box.GetWidth());
+    int m_cx = mean_shape_box.GetX() + mean_shape_box.GetWidth() / 2;
+    int m_cy = mean_shape_box.GetY() + mean_shape_box.GetHeight() / 2;
+    inspirecv::Rect2f _new(m_cx - m_R / 2, m_cy - m_R / 2, m_R, m_R);
+    inspirecv::TransformMatrix affine = SquareToSquare(old, _new);
+    std::vector<inspirecv::Point2f> new_pts = ApplyTransformToPoints(points, affine);
+    return new_pts;
+}
+
+template<class ForwardIterator>
+inline size_t argmax(ForwardIterator first, ForwardIterator last) {
+    return std::distance(first, std::max_element(first, last));
+}
+
+
+// Structure to hold bounding box coordinates
+struct BoundingBox {
+    int left_top_x;
+    int left_top_y;
+    int right_bottom_x;
+    int right_bottom_y;
+};
+
+inline inspirecv::Rect2i AlignmentBoxToStrideSquareBox(const inspirecv::Rect2i &bbox, int stride) {
+    // 1. Convert xywh to cxcywh (center point coordinates and width/height)
+    int center_x = bbox.GetX() + bbox.GetWidth() / 2;
+    int center_y = bbox.GetY() + bbox.GetHeight() / 2;
+    int width = bbox.GetWidth();
+    int height = bbox.GetHeight();
+    
+    // 2. Get the shortest side of the width and height
+    int min_side = std::min(width, height);
+    
+    // 3. Align the shortest side to the stride
+    int aligned_side = (min_side / stride) * stride;
+    
+    // 4. Create a square box (keep the center point unchanged)
+    int half_side = aligned_side / 2;
+    int new_x = center_x - half_side;
+    int new_y = center_y - half_side;
+    
+    // 5. Convert cxcywh back to xywh
+    return inspirecv::Rect2i(new_x, new_y, aligned_side, aligned_side);
+}
+
+inline inspirecv::Rect2i GetNewBox(int src_w, int src_h, inspirecv::Rect2i bbox, float scale) {
+    // Convert cv::Rect to BoundingBox
+    BoundingBox box;
+    box.left_top_x = bbox.GetX();
+    box.left_top_y = bbox.GetY();
+    box.right_bottom_x = bbox.GetX() + bbox.GetWidth();
+    box.right_bottom_y = bbox.GetY() + bbox.GetHeight();
+
+    // Compute new bounding box
+    scale = std::min({static_cast<float>(src_h - 1) / bbox.GetHeight(), static_cast<float>(src_w - 1) / bbox.GetWidth(), scale});
+
+    float new_width = bbox.GetWidth() * scale;
+    float new_height = bbox.GetHeight() * scale;
+    float center_x = bbox.GetWidth() / 2.0f + bbox.GetX();
+    float center_y = bbox.GetHeight() / 2.0f + bbox.GetY();
+
+    float left_top_x = center_x - new_width / 2.0f;
+    float left_top_y = center_y - new_height / 2.0f;
+    float right_bottom_x = center_x + new_width / 2.0f;
+    float right_bottom_y = center_y + new_height / 2.0f;
+
+    if (left_top_x < 0) {
+        right_bottom_x -= left_top_x;
+        left_top_x = 0;
+    }
+
+    if (left_top_y < 0) {
+        right_bottom_y -= left_top_y;
+        left_top_y = 0;
+    }
+
+    if (right_bottom_x > src_w - 1) {
+        left_top_x -= right_bottom_x - src_w + 1;
+        right_bottom_x = src_w - 1;
+    }
+
+    if (right_bottom_y > src_h - 1) {
+        left_top_y -= right_bottom_y - src_h + 1;
+        right_bottom_y = src_h - 1;
+    }
+
+    // Convert back to cv::Rect for output
+    inspirecv::Rect2i new_bbox(static_cast<int>(left_top_x), static_cast<int>(left_top_y),
+                      static_cast<int>(right_bottom_x - left_top_x), static_cast<int>(right_bottom_y - left_top_y));
+    return new_bbox;
+}
+
+
+template<typename T>
+inline bool isShortestSideGreaterThan(const inspirecv::Rect<T>& rect, T value, float scale) {
+    // Find the shortest edge
+    T shortestSide = std::min(static_cast<float>(rect.GetWidth()) / scale, static_cast<float>(rect.GetHeight()) / scale);
+    // Determines whether the shortest edge is greater than the given value
+    return shortestSide > value;
+}
+
+// Exponential Moving Average (EMA) filter function
+inline float EmaFilter(float currentProb, std::vector<float> &history, int max, float alpha = 0.2f) {
+    // Add current probability to history
+    history.push_back(currentProb);
+
+    // Trim history if it exceeds max size
+    if (history.size() > max) {
+        history.erase(history.begin(), history.begin() + (history.size() - max));
+    }
+
+    // Compute EMA
+    float ema = history[0];  // Initial value
+    for (size_t i = 1; i < history.size(); ++i) {
+        ema = alpha * history[i] + (1 - alpha) * ema;
+    }
+
+    return ema;
+}
+
+
+// Vector EMA filter function
+inline std::vector<float> VectorEmaFilter(const std::vector<float>& currentProbs, 
+                                         std::vector<std::vector<float>>& history, 
+                                         int max, 
+                                         float alpha = 0.2f) {
+    // Add current probability vector to history
+    history.push_back(currentProbs);
+    
+    // Trim history if it exceeds max size
+    if (history.size() > max) {
+        history.erase(history.begin(), history.begin() + (history.size() - max));
+    }
+    
+    // If only one sample, return it directly
+    if (history.size() == 1) {
+        return history[0];
+    }
+    
+    // Compute EMA for each dimension
+    std::vector<float> ema = history[0];  // Initial values
+    
+    for (size_t i = 1; i < history.size(); ++i) {
+        for (size_t j = 0; j < ema.size(); ++j) {
+            ema[j] = alpha * history[i][j] + (1 - alpha) * ema[j];
+        }
+    }
+    
+    return ema;
+}
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_UTILS_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/all.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/all.h
new file mode 100644
index 000000000..3bb7f2786
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/all.h
@@ -0,0 +1,13 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef HYPERFACEREPO_ATTRIBUTE_ALL_H
+#define HYPERFACEREPO_ATTRIBUTE_ALL_H
+
+#include "mask_predict_adapt.h"
+#include "face_attribute_adapt.h"
+#include "face_emotion_adapt.h"
+
+#endif  // HYPERFACEREPO_ATTRIBUTE_ALL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.cpp
new file mode 100644
index 000000000..84cff11c0
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.cpp
@@ -0,0 +1,43 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_attribute_adapt.h"
+#include "middleware/utils.h"
+
+namespace inspire {
+
+FaceAttributePredictAdapt::FaceAttributePredictAdapt() : AnyNetAdapter("FaceAttributePredictAdapt") {}
+
+std::vector<int> FaceAttributePredictAdapt::operator()(const inspirecv::Image &bgr_affine) {
+    AnyTensorOutputs outputs;
+    if (bgr_affine.Width() != INPUT_WIDTH || bgr_affine.Height() != INPUT_HEIGHT) {
+        auto resized = bgr_affine.Resize(INPUT_WIDTH, INPUT_HEIGHT);
+        Forward(resized, outputs);
+    } else {
+        Forward(bgr_affine, outputs);
+    }
+
+    // cv::imshow("w", bgr_affine);
+    // cv::waitKey(0);
+
+    std::vector<float> &raceOut = outputs[0].second;
+    std::vector<float> &genderOut = outputs[1].second;
+    std::vector<float> &ageOut = outputs[2].second;
+
+    auto raceIdx = argmax(raceOut.begin(), raceOut.end());
+    auto genderIdx = argmax(genderOut.begin(), genderOut.end());
+    auto ageIdx = argmax(ageOut.begin(), ageOut.end());
+
+    std::string raceLabel = m_original_labels_[raceIdx];
+    std::string simplifiedLabel = m_label_map_.at(raceLabel);
+    int simplifiedRaceIdx = m_simplified_label_index_.at(simplifiedLabel);
+
+    // std::cout << raceLabel << std::endl;
+    // std::cout << simplifiedLabel << std::endl;
+
+    return {simplifiedRaceIdx, 1 - (int)genderIdx, (int)ageIdx};
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.h
new file mode 100644
index 000000000..d6065950d
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_attribute_adapt.h
@@ -0,0 +1,61 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_ATTRIBUTE_ADAPT_H
+#define INSPIRE_FACE_ATTRIBUTE_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class FaceAttributePredict
+ * @brief According to the face image, three classification information of age, gender and race were extracted.
+ *
+ * This class inherits from AnyNet and provides methods for performing face attribute prediction.
+ */
+class INSPIRE_API FaceAttributePredictAdapt : public AnyNetAdapter {
+public:
+    int32_t INPUT_WIDTH = 112;
+    int32_t INPUT_HEIGHT = 112;
+
+    /**
+     * @brief Constructor for FaceAttributePredict class.
+     */
+    FaceAttributePredictAdapt();
+
+    /**
+     * @brief Exec infer.
+     *
+     * @param bgr_affine The BGR affine matrix to perform mask prediction on.
+     * @return The multi-list attribute prediction result.
+     */
+    std::vector<int> operator()(const inspirecv::Image& bgr_affine);
+
+private:
+    // Define primitive tag
+    const std::vector<std::string> m_original_labels_ = {"Black",          "East Asian",      "Indian", "Latino_Hispanic",
+                                                         "Middle Eastern", "Southeast Asian", "White"};
+
+    // Define simplified labels
+    const std::vector<std::string> m_simplified_labels_ = {"Black", "Asian", "Latino/Hispanic", "Middle Eastern", "White"};
+
+    // Define the mapping from the original tag to the simplified tag
+    const std::unordered_map<std::string, std::string> m_label_map_ = {{"Black", "Black"},
+                                                                       {"East Asian", "Asian"},
+                                                                       {"Indian", "Asian"},
+                                                                       {"Latino_Hispanic", "Latino/Hispanic"},
+                                                                       {"Middle Eastern", "Middle Eastern"},
+                                                                       {"Southeast Asian", "Asian"},
+                                                                       {"White", "White"}};
+
+    // Define index maps for simplified labels
+    const std::unordered_map<std::string, int> m_simplified_label_index_ = {
+      {"Black", 0}, {"Asian", 1}, {"Latino/Hispanic", 2}, {"Middle Eastern", 3}, {"White", 4}};
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_ATTRIBUTE_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.cpp
new file mode 100644
index 000000000..d11e4bbc5
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.cpp
@@ -0,0 +1,24 @@
+#include "face_emotion_adapt.h"
+
+namespace inspire {
+
+FaceEmotionAdapt::FaceEmotionAdapt() : AnyNetAdapter("FaceEmotionAdapt") {}
+
+FaceEmotionAdapt::~FaceEmotionAdapt() {}
+
+std::vector<float> FaceEmotionAdapt::operator()(const inspirecv::Image& bgr_affine) {
+    AnyTensorOutputs outputs;
+    if (bgr_affine.Width() != INPUT_WIDTH || bgr_affine.Height() != INPUT_HEIGHT) {
+        auto resized = bgr_affine.Resize(INPUT_WIDTH, INPUT_HEIGHT);
+        Forward(resized, outputs);
+    } else {
+        Forward(bgr_affine, outputs);
+    }
+
+    std::vector<float> &emotionOut = outputs[0].second;
+    auto sm = Softmax(emotionOut);
+
+    return sm;
+}
+
+}   // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.h
new file mode 100644
index 000000000..1be5c674d
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/face_emotion_adapt.h
@@ -0,0 +1,26 @@
+#ifndef INSPIREFACE_PIPELINE_MODULE_ATTRIBUTE_FACE_EMOTION_ADAPT_H
+#define INSPIREFACE_PIPELINE_MODULE_ATTRIBUTE_FACE_EMOTION_ADAPT_H
+
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+    
+class INSPIRE_API FaceEmotionAdapt : public AnyNetAdapter {
+public:
+
+    const std::vector<std::string> EMOTION_LABELS = {"Neutral", "Happy", "Sad", "Surprise", "Fear", "Disgust", "Anger"};
+    const int32_t INPUT_WIDTH = 112;
+    const int32_t INPUT_HEIGHT = 112;
+    const int32_t OUTPUT_SIZE = 7;
+public:
+    FaceEmotionAdapt();
+    ~FaceEmotionAdapt();
+
+    std::vector<float> operator()(const inspirecv::Image& bgr_affine);
+    
+};  // class FaceEmotionAdapt
+
+}   // namespace inspire
+
+#endif // INSPIREFACE_PIPELINE_MODULE_ATTRIBUTE_FACE_EMOTION_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.cpp
new file mode 100644
index 000000000..0097012df
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.cpp
@@ -0,0 +1,34 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "mask_predict_adapt.h"
+
+namespace inspire {
+
+MaskPredictAdapt::MaskPredictAdapt() : AnyNetAdapter("MaskPredictAdapt") {}
+
+float MaskPredictAdapt::operator()(const inspirecv::Image& bgr_affine) {
+    AnyTensorOutputs outputs;
+    if (bgr_affine.Height() == m_input_size_ && bgr_affine.Width() == m_input_size_) {
+        Forward(bgr_affine, outputs);
+
+    } else {
+        // auto resized = bgr_affine.Resize(m_input_size_, m_input_size_);
+        uint8_t* resized_data = nullptr;
+        m_processor_->Resize(bgr_affine.Data(), bgr_affine.Width(), bgr_affine.Height(), bgr_affine.Channels(), &resized_data, m_input_size_,
+                             m_input_size_);
+        auto resized = inspirecv::Image::Create(m_input_size_, m_input_size_, bgr_affine.Channels(), resized_data, false);
+        Forward(resized, outputs);
+    }
+    m_processor_->MarkDone();
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+    auto sm = Softmax(outputs[0].second);
+    return sm[0];
+#else
+    return outputs[0].second[0];
+#endif
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.h
new file mode 100644
index 000000000..668efa952
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/attribute/mask_predict_adapt.h
@@ -0,0 +1,40 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_MASKPREDICT_ADAPT_H
+#define INSPIRE_FACE_MASKPREDICT_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class MaskPredict
+ * @brief Class for performing mask prediction.
+ *
+ * This class inherits from AnyNet and provides methods for performing mask prediction.
+ */
+class INSPIRE_API MaskPredictAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for MaskPredict class.
+     */
+    MaskPredictAdapt();
+
+    /**
+     * @brief Operator for performing mask prediction on a BGR affine matrix.
+     *
+     * @param bgr_affine The BGR affine matrix to perform mask prediction on.
+     * @return float The mask prediction result.
+     */
+    float operator()(const inspirecv::Image& bgr_affine);
+
+private:
+    const int m_input_size_ = 96;  ///< The input size for the model.
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_MASKPREDICT_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.cpp
new file mode 100644
index 000000000..31e4c8aae
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.cpp
@@ -0,0 +1,320 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_pipeline_module.h"
+
+#include "log.h"
+#include "track_module/landmark/face_landmark_adapt.h"
+#include "track_module/landmark/landmark_param.h"
+#include "recognition_module/dest_const.h"
+#include "herror.h"
+
+namespace inspire {
+
+FacePipelineModule::FacePipelineModule(InspireArchive &archive, bool enableLiveness, bool enableMaskDetect, bool enableAttribute,
+                                       bool enableInteractionLiveness, bool enableFaceEmotion)
+: m_enable_liveness_(enableLiveness),
+  m_enable_mask_detect_(enableMaskDetect),
+  m_enable_attribute_(enableAttribute),
+  m_enable_interaction_liveness_(enableInteractionLiveness),
+  m_enable_face_emotion_(enableFaceEmotion) {
+    if (m_enable_attribute_) {
+        InspireModel attrModel;
+        auto ret = archive.LoadModel("face_attribute", attrModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("Load Face attribute model: %d", ret);
+        }
+
+        ret = InitFaceAttributePredict(attrModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("InitAgePredict error.");
+        }
+    }
+    m_landmark_param_ = archive.GetLandmarkParam();
+    // Initialize the mask detection model
+    if (m_enable_mask_detect_) {
+        InspireModel maskModel;
+        auto ret = archive.LoadModel("mask_detect", maskModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("Load Mask model: %d", ret);
+        }
+        ret = InitMaskPredict(maskModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("InitMaskPredict error.");
+        }
+    }
+
+    // Initializing the RGB live detection model
+    if (m_enable_liveness_) {
+        InspireModel livenessModel;
+        auto ret = archive.LoadModel("rgb_anti_spoofing", livenessModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("Load anti-spoofing model.");
+        }
+        ret = InitRBGAntiSpoofing(livenessModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("InitRBGAntiSpoofing error.");
+        }
+    }
+
+    // There may be a combination of algorithms for facial interaction
+    if (m_enable_interaction_liveness_) {
+        // Blink model
+        InspireModel blinkModel;
+        auto ret = archive.LoadModel("blink_predict", blinkModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("Load Blink model error.");
+        }
+        ret = InitBlinkFromLivenessInteraction(blinkModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("InitBlinkFromLivenessInteraction error.");
+        }
+    }
+
+    // Initialize the face emotion model
+    if (m_enable_face_emotion_) {
+        InspireModel faceEmotionModel;
+        auto ret = archive.LoadModel("face_emotion", faceEmotionModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("Load Face emotion model error.");
+        }
+        ret = InitFaceEmotion(faceEmotionModel);
+        if (ret != 0) {
+            INSPIRE_LOGE("InitFaceEmotion error.");
+        }
+    }
+}
+
+int32_t FacePipelineModule::Process(inspirecv::FrameProcess &processor, const FaceTrackWrap &face, FaceProcessFunctionOption proc) {
+    // Original image
+    inspirecv::Image originImage;
+    std::vector<inspirecv::Point2f> stand_lmk;
+    switch (proc) {
+        case PROCESS_MASK: {
+            if (m_mask_predict_ == nullptr) {
+                INSPIRE_LOGE("Mask detection disabled");
+                return HERR_SESS_PIPELINE_FAILURE;  // uninitialized
+            }
+            std::vector<inspirecv::Point2f> pointsFive;
+            for (const auto &p : face.keyPoints) {
+                pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+            }
+
+            auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+            auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+            auto mask_score = (*m_mask_predict_)(crop);
+            // crop.Show();
+            faceMaskCache = mask_score;
+            break;
+        }
+        case PROCESS_RGB_LIVENESS: {
+            if (m_rgb_anti_spoofing_ == nullptr) {
+                INSPIRE_LOGE("RGB liveness detection disabled");
+                return HERR_SESS_PIPELINE_FAILURE;  // uninitialized
+            }
+            // New scheme: padding differences cause errors in inference results
+            // inspirecv::TransformMatrix rotation_mode_affine = processor.GetAffineMatrix();
+            // if (stand_lmk.empty()) {
+            //     std::vector<inspirecv::Point2f> lmk;
+            //     for (const auto &p : face.densityLandmark) {
+            //         lmk.emplace_back(p.x, p.y);
+            //     }
+            //     stand_lmk = inspirecv::ApplyTransformToPoints(lmk, rotation_mode_affine.GetInverse());
+            // }
+
+            // auto rect_face = inspirecv::MinBoundingRect(stand_lmk);
+            // auto rect_pts = rect_face.Square(2.7f).As<float>().ToFourVertices();
+            // std::vector<inspirecv::Point2f> dst_pts = {{0, 0}, {112, 0}, {112, 112}, {0, 112}};
+            // std::vector<inspirecv::Point2f> camera_pts = inspirecv::ApplyTransformToPoints(rect_pts, rotation_mode_affine);
+
+            // auto affine = inspirecv::SimilarityTransformEstimate(camera_pts, dst_pts);
+            // auto image_affine = processor.ExecuteImageAffineProcessing(affine, 112, 112);
+            // image_affine.Write("liveness_affine.jpg");
+
+            if (originImage.Empty()) {
+                // This is a poor approach that impacts performance, 
+                // but in order to capture clearer images and improve liveness detection accuracy, 
+                // we have to keep it.
+                originImage = processor.ExecuteImageScaleProcessing(1.0, true);
+            }
+            inspirecv::Rect2i oriRect(face.rect.x, face.rect.y, face.rect.width, face.rect.height);
+            auto rect = GetNewBox(originImage.Width(), originImage.Height(), oriRect, 2.7f);
+            if (Launch::GetInstance()->GetImageProcessingBackend() == Launch::IMAGE_PROCESSING_RGA) {
+                // RKRGA must be aligned to 16
+                rect = AlignmentBoxToStrideSquareBox(rect, 16);
+            }
+            auto crop = originImage.Crop(rect);
+            auto score = (*m_rgb_anti_spoofing_)(crop);
+            // crop.Show();
+            // crop.Resize(112, 112).Write("liveness.jpg");
+            faceLivenessCache = score;
+            break;
+        }
+        case PROCESS_INTERACTION: {
+            if (m_blink_predict_ == nullptr) {
+                INSPIRE_LOGE("Interaction action detection disabled");
+                return HERR_SESS_PIPELINE_FAILURE;  // uninitialized
+            }
+            std::vector<std::vector<int>> order_list = {m_landmark_param_->semantic_index.left_eye_region, m_landmark_param_->semantic_index.right_eye_region};
+            eyesStatusCache = {0, 0};
+            inspirecv::Point2f left_eye = inspirecv::Point2f(face.keyPoints[0].x, face.keyPoints[0].y);
+            inspirecv::Point2f right_eye = inspirecv::Point2f(face.keyPoints[1].x, face.keyPoints[1].y);
+            std::vector<inspirecv::Point2f> eyes = {left_eye, right_eye};
+            // Get affine matrix
+            inspirecv::TransformMatrix rotation_mode_affine = processor.GetAffineMatrix();
+            // Get stand landmark
+            if (stand_lmk.empty()) {
+                std::vector<inspirecv::Point2f> lmk;
+                for (const auto &p : face.densityLandmark) {
+                    lmk.emplace_back(p.x, p.y);
+                }
+                stand_lmk = inspirecv::ApplyTransformToPoints(lmk, rotation_mode_affine.GetInverse());
+            }
+            for (size_t i = 0; i < order_list.size(); i++) {
+                const auto &index = order_list[i];
+                std::vector<inspirecv::Point2i> points;
+                for (const auto &idx : index) {
+                    points.emplace_back(stand_lmk[idx].GetX(), stand_lmk[idx].GetY());
+                }
+                auto rect_eye = inspirecv::MinBoundingRect(points).Square(1.5f);
+                auto rect_pts_eye = rect_eye.As<float>().ToFourVertices();
+                std::vector<inspirecv::Point2f> dst_pts_eye = {{0, 0}, {64, 0}, {64, 64}, {0, 64}};
+                std::vector<inspirecv::Point2f> camera_pts_eye = inspirecv::ApplyTransformToPoints(rect_pts_eye, rotation_mode_affine);
+
+                auto affine_eye = inspirecv::SimilarityTransformEstimate(camera_pts_eye, dst_pts_eye);
+                auto eye_affine = processor.ExecuteImageAffineProcessing(affine_eye, 64, 64);
+                // eye_affine.Write("eye_"+std::to_string(i)+".jpg");
+                // auto crop = originImage.Crop(new_rect);
+                auto score = (*m_blink_predict_)(eye_affine);
+                eyesStatusCache[i] = score;
+            }
+            break;
+        }
+        case PROCESS_ATTRIBUTE: {
+            if (m_attribute_predict_ == nullptr) {
+                INSPIRE_LOGE("Face attribute detection disabled");
+                return HERR_SESS_PIPELINE_FAILURE;  // uninitialized
+            }
+            std::vector<inspirecv::Point2f> pointsFive;
+            for (const auto &p : face.keyPoints) {
+                pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+            }
+            auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+            auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+            auto outputs = (*m_attribute_predict_)(crop);
+            faceAttributeCache = inspirecv::Vec3i{outputs[0], outputs[1], outputs[2]};
+            break;
+        }
+        case PROCESS_FACE_EMOTION: {
+            if (m_face_emotion_ == nullptr) {
+                INSPIRE_LOGE("Face emotion detection disabled");
+                return HERR_SESS_PIPELINE_FAILURE;  // uninitialized
+            }
+            std::vector<inspirecv::Point2f> pointsFive;
+            for (const auto &p : face.keyPoints) {
+                pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+            }
+            auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+            auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+            // crop.Show();
+            faceEmotionCache = (*m_face_emotion_)(crop);
+            break;
+        }
+    }
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::Process(inspirecv::FrameProcess &processor, FaceObjectInternal &face) {
+    // In the tracking state, the count meets the requirements or the pipeline is executed in the detection state
+    auto lmk = face.keyPointFive;
+    std::vector<inspirecv::Point2f> lmk_5 = {lmk[m_landmark_param_->semantic_index.left_eye_center], lmk[m_landmark_param_->semantic_index.right_eye_center],
+                                             lmk[m_landmark_param_->semantic_index.nose_corner], lmk[m_landmark_param_->semantic_index.mouth_left_corner],
+                                             lmk[m_landmark_param_->semantic_index.mouth_right_corner]};
+    auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, lmk_5);
+    auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+    if (m_mask_predict_ != nullptr) {
+        auto mask_score = (*m_mask_predict_)(crop);
+        if (mask_score > 0.95) {
+            face.faceProcess.maskInfo = MaskInfo::MASKED;
+        } else {
+            face.faceProcess.maskInfo = MaskInfo::UNMASKED;
+        }
+    }
+
+    if (m_rgb_anti_spoofing_ != nullptr) {
+        auto img = processor.ExecuteImageScaleProcessing(1.0, true);
+        inspirecv::Rect2i oriRect(face.detect_bbox_.GetX(), face.detect_bbox_.GetY(), face.detect_bbox_.GetWidth(), face.detect_bbox_.GetHeight());
+        auto rect = oriRect.Square(2.7f);
+        auto crop = img.Crop(rect);
+        auto score = (*m_rgb_anti_spoofing_)(crop);
+        if (score > 0.88) {
+            face.faceProcess.rgbLivenessInfo = RGBLivenessInfo::LIVENESS_REAL;
+        } else {
+            face.faceProcess.rgbLivenessInfo = RGBLivenessInfo::LIVENESS_FAKE;
+        }
+    }
+
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::InitFaceAttributePredict(InspireModel &model) {
+    m_attribute_predict_ = std::make_shared<FaceAttributePredictAdapt>();
+    auto ret = m_attribute_predict_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::InitMaskPredict(InspireModel &model) {
+    m_mask_predict_ = std::make_shared<MaskPredictAdapt>();
+    auto ret = m_mask_predict_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::InitRBGAntiSpoofing(InspireModel &model) {
+    auto input_size = model.Config().get<std::vector<int>>("input_size");
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+    m_rgb_anti_spoofing_ = std::make_shared<RBGAntiSpoofingAdapt>(input_size[0], true);
+#else
+    m_rgb_anti_spoofing_ = std::make_shared<RBGAntiSpoofingAdapt>(input_size[0]);
+#endif
+    auto ret = m_rgb_anti_spoofing_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::InitBlinkFromLivenessInteraction(InspireModel &model) {
+    m_blink_predict_ = std::make_shared<BlinkPredictAdapt>();
+    auto ret = m_blink_predict_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int32_t FacePipelineModule::InitLivenessInteraction(InspireModel &model) {
+    return 0;
+}
+
+int32_t FacePipelineModule::InitFaceEmotion(InspireModel &model) {
+    m_face_emotion_ = std::make_shared<FaceEmotionAdapt>();
+    auto ret = m_face_emotion_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+const std::shared_ptr<RBGAntiSpoofingAdapt> &FacePipelineModule::getMRgbAntiSpoofing() const {
+    return m_rgb_anti_spoofing_;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.h
new file mode 100644
index 000000000..96598b645
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/face_pipeline_module.h
@@ -0,0 +1,154 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIRE_FACE_PIPELINE_MODULE_H
+#define INSPIRE_FACE_PIPELINE_MODULE_H
+
+#include "frame_process.h"
+#include "common/face_info/face_object_internal.h"
+#include "attribute/face_attribute_adapt.h"
+#include "attribute/mask_predict_adapt.h"
+#include "liveness/rgb_anti_spoofing_adapt.h"
+#include "liveness/blink_predict_adapt.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "face_wrapper.h"
+#include "track_module/landmark/landmark_param.h"
+#include "attribute/face_emotion_adapt.h"
+
+namespace inspire {
+
+/**
+ * @enum FaceProcessFunction
+ * @brief Enumeration for different face processing functions in the FacePipeline.
+ */
+typedef enum FaceProcessFunctionOption {
+    PROCESS_MASK = 0,      ///< Mask detection.
+    PROCESS_RGB_LIVENESS,  ///< RGB liveness detection.
+    PROCESS_ATTRIBUTE,     ///< Face attribute estimation.
+    PROCESS_INTERACTION,   ///< Face interaction.
+    PROCESS_FACE_EMOTION,  ///< Face emotion recognition.
+} FaceProcessFunctionOption;
+
+/**
+ * @class FacePipeline
+ * @brief Class for performing face processing tasks in a pipeline.
+ *
+ * This class provides methods for processing faces in a pipeline, including mask detection, liveness detection,
+ * age estimation, and gender prediction.
+ */
+class FacePipelineModule {
+public:
+    /**
+     * @brief Constructor for FacePipeline class.
+     *
+     * @param archive Model archive instance for model loading.
+     * @param enableLiveness Whether RGB liveness detection is enabled.
+     * @param enableMaskDetect Whether mask detection is enabled.
+     * @param enableAttributee Whether face attribute estimation is enabled.
+     * @param enableInteractionLiveness Whether interaction liveness detection is enabled.
+     * @param enableFaceEmotion Whether interaction emotion recognition is enabled.
+     */
+    explicit FacePipelineModule(InspireArchive &archive, bool enableLiveness, bool enableMaskDetect, bool enableAttribute,
+                                bool enableInteractionLiveness, bool enableFaceEmotion);
+
+    /**
+     * @brief Processes a face using the specified FaceProcessFunction.
+     *
+     * @param image CameraStream instance containing the image.
+     * @param face FaceObject representing the detected face.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t Process(inspirecv::FrameProcess &processor, FaceObjectInternal &face);
+
+    /**
+     * @brief Processes a face using the specified FaceProcessFunction.
+     *
+     * @param image CameraStream instance containing the image.
+     * @param face FaceTrackWrap representing the detected face.
+     * @param proc The FaceProcessFunction to apply to the face.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t Process(inspirecv::FrameProcess &processor, const FaceTrackWrap &face, FaceProcessFunctionOption proc);
+
+    /**
+     * @brief Get Rgb AntiSpoofing module
+     * @return AntiSpoofing module
+     */
+    const std::shared_ptr<RBGAntiSpoofingAdapt> &getMRgbAntiSpoofing() const;
+
+private:
+    /**
+     * @brief Initializes the FaceAttributePredict model.
+     *
+     * @param model Pointer to the FaceAttributePredict model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitFaceAttributePredict(InspireModel &model);
+
+    /**
+     * @brief Initializes the MaskPredict model.
+     *
+     * @param model Pointer to the MaskPredict model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitMaskPredict(InspireModel &model);
+
+    /**
+     * @brief Initializes the RBGAntiSpoofing model.
+     *
+     * @param model Pointer to the RBGAntiSpoofing model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitRBGAntiSpoofing(InspireModel &model);
+
+    /**
+     * @brief Initializes the LivenessInteraction model.
+     *
+     * @param model Pointer to the LivenessInteraction model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitLivenessInteraction(InspireModel &model);
+
+    /**
+     * @brief Initializes the Blink predict model.
+     *
+     * @param model Pointer to the Blink predict model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitBlinkFromLivenessInteraction(InspireModel &model);
+
+    /**
+     * @brief Initializes the FaceEmotion model.
+     *
+     * @param model Pointer to the FaceEmotion model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitFaceEmotion(InspireModel &model);
+
+private:
+    const bool m_enable_liveness_ = false;              ///< Whether RGB liveness detection is enabled.
+    const bool m_enable_mask_detect_ = false;           ///< Whether mask detection is enabled.
+    const bool m_enable_attribute_ = false;             ///< Whether face attribute is enabled.
+    const bool m_enable_interaction_liveness_ = false;  ///< Whether interaction liveness detection is enabled.
+    const bool m_enable_face_emotion_ = false;          ///< Whether face emotion is enabled.
+    
+    std::shared_ptr<FaceAttributePredictAdapt> m_attribute_predict_;  ///< Pointer to AgePredict instance.
+    std::shared_ptr<MaskPredictAdapt> m_mask_predict_;                ///< Pointer to MaskPredict instance.
+    std::shared_ptr<RBGAntiSpoofingAdapt> m_rgb_anti_spoofing_;       ///< Pointer to RBGAntiSpoofing instance.
+    std::shared_ptr<BlinkPredictAdapt> m_blink_predict_;              ///< Pointer to Blink predict instance.
+    std::shared_ptr<FaceEmotionAdapt> m_face_emotion_;                ///< Pointer to FaceEmotion instance.
+    std::shared_ptr<LandmarkParam> m_landmark_param_;                ///< Pointer to LandmarkParam instance.
+
+public:
+    float faceMaskCache;                  ///< Cache for face mask detection result.
+    float faceLivenessCache;              ///< Cache for face liveness detection result.
+    inspirecv::Vec2f eyesStatusCache;     ///< Cache for blink predict result.
+    inspirecv::Vec3i faceAttributeCache;  ///< Cache for face attribute predict result.
+    std::vector<float> faceEmotionCache;   ///< Cache for face emotion recognition result.
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_PIPELINE_MODULE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/all.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/all.h
new file mode 100644
index 000000000..fa0ef9fcf
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/all.h
@@ -0,0 +1,13 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_LIVENESS_ALL_H
+#define INSPIREFACE_LIVENESS_ALL_H
+
+#include "blink_predict_adapt.h"
+#include "rgb_anti_spoofing_adapt.h"
+#include "order_of_hyper_landmark.h"
+
+#endif  // INSPIREFACE_LIVENESS_ALL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.cpp
new file mode 100644
index 000000000..e53a16f8c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.cpp
@@ -0,0 +1,28 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "blink_predict_adapt.h"
+#include "middleware/utils.h"
+
+namespace inspire {
+
+BlinkPredictAdapt::BlinkPredictAdapt() : AnyNetAdapter("BlinkPredictAdapt") {}
+
+float BlinkPredictAdapt::operator()(const inspirecv::Image &bgr_affine) {
+    AnyTensorOutputs outputs;
+    if (bgr_affine.Width() == BLINK_EYE_INPUT_SIZE && bgr_affine.Height() == BLINK_EYE_INPUT_SIZE) {
+        auto input = bgr_affine.ToGray();
+        Forward(input, outputs);
+    } else {
+        auto input = bgr_affine.ToGray();
+        input = input.Resize(BLINK_EYE_INPUT_SIZE, BLINK_EYE_INPUT_SIZE);
+        Forward(input, outputs);
+    }
+    auto &map = outputs[0].second;
+
+    return map[1];
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.h
new file mode 100644
index 000000000..2f46473d1
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/blink_predict_adapt.h
@@ -0,0 +1,40 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_BLINK_PREDICT_H
+#define INSPIREFACE_BLINK_PREDICT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class BlinkPredict
+ * @brief Prediction whether the eyes are open or closed.
+ *
+ * This class inherits from AnyNet and provides methods for performing blink prediction.
+ */
+class INSPIRE_API BlinkPredictAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for MaskPredict class.
+     */
+    BlinkPredictAdapt();
+
+    /**
+     * @brief Operator for performing blink prediction on a BGR affine matrix.
+     *
+     * @param bgr_affine The BGR affine matrix to perform mask prediction on.
+     * @return Blink prediction result.
+     */
+    float operator()(const inspirecv::Image& bgr_affine);
+
+public:
+    static const int BLINK_EYE_INPUT_SIZE = 64;  ///< Input size
+};
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_BLINK_PREDICT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.cpp
new file mode 100644
index 000000000..271173038
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.cpp
@@ -0,0 +1,36 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "rgb_anti_spoofing_adapt.h"
+
+namespace inspire {
+
+RBGAntiSpoofingAdapt::RBGAntiSpoofingAdapt(int input_size, bool use_softmax) : AnyNetAdapter("RBGAntiSpoofingAdapt") {
+    m_input_size_ = input_size;
+    m_softmax_ = use_softmax;
+}
+
+float RBGAntiSpoofingAdapt::operator()(const inspirecv::Image& bgr_affine27) {
+    AnyTensorOutputs outputs;
+    if (bgr_affine27.Width() != m_input_size_ || bgr_affine27.Height() != m_input_size_) {
+        // auto resized = bgr_affine27.Resize(m_input_size_, m_input_size_);
+        uint8_t* resized_data = nullptr;
+        float scale;
+        m_processor_->Resize(bgr_affine27.Data(), bgr_affine27.Width(), bgr_affine27.Height(), bgr_affine27.Channels(), &resized_data, m_input_size_,
+                             m_input_size_);
+        auto resized = inspirecv::Image::Create(m_input_size_, m_input_size_, bgr_affine27.Channels(), resized_data, false);
+        Forward(resized, outputs);
+    } else {
+        Forward(bgr_affine27, outputs);
+    }
+    if (m_softmax_) {
+        auto sm = Softmax(outputs[0].second);
+        return sm[1];
+    } else {
+        return outputs[0].second[1];
+    }
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.h b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.h
new file mode 100644
index 000000000..9aff8784a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/pipeline_module/liveness/rgb_anti_spoofing_adapt.h
@@ -0,0 +1,44 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_RGB_ANTI_SPOOFING_ADAPT_H
+#define INSPIRE_RGB_ANTI_SPOOFING_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class RBGAntiSpoofing
+ * @brief Class for performing RGB anti-spoofing.
+ *
+ * This class inherits from AnyNet and provides methods for performing RGB anti-spoofing.
+ */
+class INSPIRE_API RBGAntiSpoofingAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for RBGAntiSpoofing class.
+     *
+     * @param input_size The input size for the model (default is 112).
+     * @param use_softmax Whether to use softmax activation (default is false).
+     */
+    RBGAntiSpoofingAdapt(int input_size = 112, bool use_softmax = false);
+
+    /**
+     * @brief Operator for performing RGB anti-spoofing on a BGR affine27 matrix.
+     *
+     * @param bgr_affine27 The BGR affine27 matrix to perform anti-spoofing on.
+     * @return float The anti-spoofing result.
+     */
+    float operator()(const inspirecv::Image& bgr_affine27);
+
+private:
+    int m_input_size_;        ///< The input size for the model.
+    bool m_softmax_ = false;  ///< Whether to use softmax activation.
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_RGB_ANTI_SPOOFING_ADAPT_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/platform/jni/android/inspireface_jni.cpp b/cpp-package/inspireface/cpp/inspireface/platform/jni/android/inspireface_jni.cpp
new file mode 100644
index 000000000..3ac1aa93b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/platform/jni/android/inspireface_jni.cpp
@@ -0,0 +1,1891 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-11-26
+ */
+
+#ifdef ANDROID
+
+#include <jni.h>
+#include <string>
+#include <android/log.h>
+#include <stdlib.h>
+#include <android/bitmap.h>
+#include "../common/common.h"
+#include "c_api/inspireface.h"
+#include "log.h"
+#include "herror.h"
+
+/**
+ * Java Native Interface (JNI) macro for generating JNI function names.
+ */
+#define INSPIRE_FACE_JNI(sig) Java_com_insightface_sdk_inspireface_##sig
+
+extern "C" {
+
+/**
+ * @brief Launch InspireFace.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param resourcePath The resource path.
+ * @return True if the launch is successful, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_GlobalLaunch)(JNIEnv *env, jobject thiz, jstring resourcePath) {
+    std::string path = jstring2str(env, resourcePath);
+    auto result = HFLaunchInspireFace(path.c_str());
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to launch InspireFace, error code: %d", result);
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Terminate InspireFace.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return True if the termination is successful, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_GlobalTerminate)(JNIEnv *env, jobject thiz) {
+    auto result = HFTerminateInspireFace();
+    if (result != 0) {
+        INSPIRE_LOGE("Failed to terminate InspireFace, error code: %d", result);
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Query InspireFace launch status.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return True if the launch status is successful, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_QueryLaunchStatus)(JNIEnv *env, jobject thiz) {
+    HInt32 status;
+    auto result = HFQueryInspireFaceLaunchStatus(&status);
+    if (result != 0) {
+        INSPIRE_LOGE("Failed to query InspireFace launch status, error code: %d", result);
+        return false;
+    }
+    return status;
+}
+
+/**
+ * @brief Create a new InspireFace session.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param customParameter The custom parameter object.
+ * @param detectMode The detection mode.
+ * @param maxDetectFaceNum The maximum number of detected faces.
+ * @param detectPixelLevel The detection pixel level.
+ * @param trackByDetectModeFPS The tracking frame rate.
+ * @return The new InspireFace session object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_CreateSession)(JNIEnv *env, jobject thiz, jobject customParameter, jint detectMode,
+                                                              jint maxDetectFaceNum, jint detectPixelLevel, jint trackByDetectModeFPS) {
+    // Get CustomParameter class and fields
+    jclass customParamClass = env->GetObjectClass(customParameter);
+    jfieldID enableRecognitionField = env->GetFieldID(customParamClass, "enableRecognition", "I");
+    jfieldID enableLivenessField = env->GetFieldID(customParamClass, "enableLiveness", "I");
+    jfieldID enableIrLivenessField = env->GetFieldID(customParamClass, "enableIrLiveness", "I");
+    jfieldID enableMaskDetectField = env->GetFieldID(customParamClass, "enableMaskDetect", "I");
+    jfieldID enableFaceQualityField = env->GetFieldID(customParamClass, "enableFaceQuality", "I");
+    jfieldID enableFaceAttributeField = env->GetFieldID(customParamClass, "enableFaceAttribute", "I");
+    jfieldID enableInteractionLivenessField = env->GetFieldID(customParamClass, "enableInteractionLiveness", "I");
+    jfieldID enableFacePoseField = env->GetFieldID(customParamClass, "enableFacePose", "I");
+    jfieldID enableFaceEmotionField = env->GetFieldID(customParamClass, "enableFaceEmotion", "I");
+
+    // Create HFSessionCustomParameter struct
+    HFSessionCustomParameter parameter;
+    parameter.enable_recognition = env->GetIntField(customParameter, enableRecognitionField);
+    parameter.enable_liveness = env->GetIntField(customParameter, enableLivenessField);
+    parameter.enable_ir_liveness = env->GetIntField(customParameter, enableIrLivenessField);
+    parameter.enable_mask_detect = env->GetIntField(customParameter, enableMaskDetectField);
+    parameter.enable_face_quality = env->GetIntField(customParameter, enableFaceQualityField);
+    parameter.enable_face_attribute = env->GetIntField(customParameter, enableFaceAttributeField);
+    parameter.enable_interaction_liveness = env->GetIntField(customParameter, enableInteractionLivenessField);
+    parameter.enable_face_pose = env->GetIntField(customParameter, enableFacePoseField);
+    parameter.enable_face_emotion = env->GetIntField(customParameter, enableFaceEmotionField);
+
+    // Create session
+    HFSession handle;
+    HResult result =
+      HFCreateInspireFaceSession(parameter, (HFDetectMode)detectMode, maxDetectFaceNum, detectPixelLevel, trackByDetectModeFPS, &handle);
+
+    if (result != 0) {
+        INSPIRE_LOGE("Failed to create session, error code: %d", result);
+        return nullptr;
+    }
+
+    // Create Session object
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jmethodID constructor = env->GetMethodID(sessionClass, "<init>", "()V");
+    jobject session = env->NewObject(sessionClass, constructor);
+
+    // Set handle
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    env->SetLongField(session, handleField, (jlong)handle);
+
+    return session;
+}
+
+/**
+ * @brief Release an InspireFace session.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_ReleaseSession)(JNIEnv *env, jobject thiz, jobject session) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    HFSession handle = (HFSession)env->GetLongField(session, handleField);
+    auto result = HFReleaseInspireFaceSession(handle);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to release session, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Create an InspireFace image stream from a bitmap.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param bitmap The bitmap object.
+ * @param rotation The rotation.
+ * @return The InspireFace image stream object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_CreateImageStreamFromBitmap)(JNIEnv *env, jobject thiz, jobject bitmap, jint rotation) {
+    AndroidBitmapInfo info;
+    void *pixels = nullptr;
+    HFImageFormat format = HF_STREAM_RGB;
+    if (AndroidBitmap_getInfo(env, bitmap, &info) < 0) {
+        INSPIRE_LOGE("Failed to get bitmap info");
+        return nullptr;
+    }
+    if (AndroidBitmap_lockPixels(env, bitmap, &pixels) < 0) {
+        AndroidBitmap_unlockPixels(env, bitmap);
+        INSPIRE_LOGE("Failed to lock pixels");
+        return nullptr;
+    }
+    if (info.format == ANDROID_BITMAP_FORMAT_RGBA_8888) {
+        format = HF_STREAM_RGBA;
+    } else if (info.format == ANDROID_BITMAP_FORMAT_RGB_565) {
+        format = HF_STREAM_RGB;
+    } else {
+        AndroidBitmap_unlockPixels(env, bitmap);
+        INSPIRE_LOGE("Unsupported bitmap format: %d", info.format);
+        return nullptr;
+    }
+
+    HFImageData imageData;
+    imageData.data = (uint8_t *)pixels;
+    imageData.width = info.width;
+    imageData.height = info.height;
+    imageData.format = (HFImageFormat)format;
+    imageData.rotation = (HFRotation)rotation;
+
+    HFImageStream streamHandle;
+    auto result = HFCreateImageStream(&imageData, &streamHandle);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to create image stream, error code: %d", result);
+        return nullptr;
+    }
+    AndroidBitmap_unlockPixels(env, bitmap);
+
+    jclass streamClass = env->FindClass("com/insightface/sdk/inspireface/base/ImageStream");
+    jmethodID constructor = env->GetMethodID(streamClass, "<init>", "()V");
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    jobject imageStreamObj = env->NewObject(streamClass, constructor);
+    env->SetLongField(imageStreamObj, streamHandleField, (jlong)streamHandle);
+
+    return imageStreamObj;
+}
+
+/**
+ * @brief Create an InspireFace image stream from a byte buffer.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param data The byte buffer.
+ * @param width The width.
+ * @param height The height.
+ * @param format The format.
+ * @param rotation The rotation.
+ * @return The InspireFace image stream object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_CreateImageStreamFromByteBuffer)(JNIEnv *env, jobject thiz, jbyteArray data, jint width, jint height,
+                                                                                jint format, jint rotation) {
+    // Convert jbyteArray to byte*
+    uint8_t *buf = (uint8_t *)env->GetByteArrayElements(data, 0);
+    HFImageData imageData;
+    imageData.data = buf;
+    imageData.width = width;
+    imageData.height = height;
+    imageData.format = (HFImageFormat)format;
+    imageData.rotation = (HFRotation)rotation;
+
+    HFImageStream streamHandle;
+    auto result = HFCreateImageStream(&imageData, &streamHandle);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to create image stream, error code: %d", result);
+        return nullptr;
+    }
+
+    jclass streamClass = env->FindClass("com/insightface/sdk/inspireface/base/ImageStream");
+    jmethodID constructor = env->GetMethodID(streamClass, "<init>", "()V");
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    jobject imageStreamObj = env->NewObject(streamClass, constructor);
+    env->SetLongField(imageStreamObj, streamHandleField, (jlong)streamHandle);
+    env->ReleaseByteArrayElements(data, (jbyte *)buf, JNI_ABORT);
+
+    return imageStreamObj;
+}
+
+/**
+ * @brief Write an InspireFace image stream to a file.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param imageStream The InspireFace image stream object.
+ * @param filePath The file path.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_WriteImageStreamToFile)(JNIEnv *env, jobject thiz, jobject imageStream, jstring filePath) {
+    jclass streamClass = env->GetObjectClass(imageStream);
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    HFImageStream streamHandle = (HFImageStream)env->GetLongField(imageStream, streamHandleField);
+
+    std::string path = jstring2str(env, filePath);
+    HFDeBugImageStreamDecodeSave(streamHandle, path.c_str());
+}
+
+/**
+ * @brief Release an InspireFace image stream.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param imageStream The InspireFace image stream object.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_ReleaseImageStream)(JNIEnv *env, jobject thiz, jobject imageStream) {
+    jclass streamClass = env->GetObjectClass(imageStream);
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    HFImageStream streamHandle = (HFImageStream)env->GetLongField(imageStream, streamHandleField);
+    auto result = HFReleaseImageStream(streamHandle);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to release image stream, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Execute face track.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param streamHandle The InspireFace image stream object.
+ * @return The face track results.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_ExecuteFaceTrack)(JNIEnv *env, jobject thiz, jobject session, jobject streamHandle) {
+    // Get session handle
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    HFSession sessionHandle = (HFSession)env->GetLongField(session, handleField);
+
+    // Get stream handle
+    jclass streamClass = env->GetObjectClass(streamHandle);
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    HFImageStream imageStreamHandle = (HFImageStream)env->GetLongField(streamHandle, streamHandleField);
+
+    // Execute face track
+    HFMultipleFaceData results;
+    HResult result = HFExecuteFaceTrack(sessionHandle, imageStreamHandle, &results);
+    if (result != 0) {
+        INSPIRE_LOGE("Failed to execute face track, error code: %d", result);
+        return nullptr;
+    }
+
+    // Create MultipleFaceData object
+    jclass multipleFaceDataClass = env->FindClass("com/insightface/sdk/inspireface/base/MultipleFaceData");
+    jmethodID constructor = env->GetMethodID(multipleFaceDataClass, "<init>", "()V");
+    jobject multipleFaceData = env->NewObject(multipleFaceDataClass, constructor);
+
+    // Set detected number
+    jfieldID detectedNumField = env->GetFieldID(multipleFaceDataClass, "detectedNum", "I");
+    env->SetIntField(multipleFaceData, detectedNumField, results.detectedNum);
+
+    if (results.detectedNum > 0) {
+        // Create and set face rects array
+        jclass faceRectClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceRect");
+        jobjectArray rectArray = env->NewObjectArray(results.detectedNum, faceRectClass, nullptr);
+        jmethodID rectConstructor = env->GetMethodID(faceRectClass, "<init>", "()V");
+        jfieldID xField = env->GetFieldID(faceRectClass, "x", "I");
+        jfieldID yField = env->GetFieldID(faceRectClass, "y", "I");
+        jfieldID widthField = env->GetFieldID(faceRectClass, "width", "I");
+        jfieldID heightField = env->GetFieldID(faceRectClass, "height", "I");
+
+        // Create and set track IDs array
+        jintArray trackIdsArray = env->NewIntArray(results.detectedNum);
+        env->SetIntArrayRegion(trackIdsArray, 0, results.detectedNum, results.trackIds);
+
+        // Create and set detection confidence array
+        jfloatArray detConfArray = env->NewFloatArray(results.detectedNum);
+        env->SetFloatArrayRegion(detConfArray, 0, results.detectedNum, results.detConfidence);
+
+        // Create and set angles array
+        jclass angleClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceEulerAngle");
+        jobjectArray angleArray = env->NewObjectArray(results.detectedNum, angleClass, nullptr);
+        jmethodID angleConstructor = env->GetMethodID(angleClass, "<init>", "()V");
+        jfieldID rollField = env->GetFieldID(angleClass, "roll", "F");
+        jfieldID yawField = env->GetFieldID(angleClass, "yaw", "F");
+        jfieldID pitchField = env->GetFieldID(angleClass, "pitch", "F");
+
+        // Create and set tokens array
+        jclass tokenClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceBasicToken");
+        jobjectArray tokenArray = env->NewObjectArray(results.detectedNum, tokenClass, nullptr);
+        jmethodID tokenConstructor = env->GetMethodID(tokenClass, "<init>", "()V");
+        jfieldID tokenDataField = env->GetFieldID(tokenClass, "data", "[B");
+        jfieldID sizeField = env->GetFieldID(tokenClass, "size", "I");
+
+        // Get token size first
+        HInt32 tokenSize = 0;
+        HFGetFaceBasicTokenSize(&tokenSize);
+
+        for (int i = 0; i < results.detectedNum; i++) {
+            // Set face rect
+            jobject rect = env->NewObject(faceRectClass, rectConstructor);
+            env->SetIntField(rect, xField, results.rects[i].x);
+            env->SetIntField(rect, yField, results.rects[i].y);
+            env->SetIntField(rect, widthField, results.rects[i].width);
+            env->SetIntField(rect, heightField, results.rects[i].height);
+            env->SetObjectArrayElement(rectArray, i, rect);
+
+            // Set angle
+            jobject angle = env->NewObject(angleClass, angleConstructor);
+            env->SetFloatField(angle, rollField, *results.angles.roll);
+            env->SetFloatField(angle, yawField, *results.angles.yaw);
+            env->SetFloatField(angle, pitchField, *results.angles.pitch);
+            env->SetObjectArrayElement(angleArray, i, angle);
+
+            // Create token object
+            jobject token = env->NewObject(tokenClass, tokenConstructor);
+            
+            // Create byte array to hold token data
+            jbyteArray dataArray = env->NewByteArray(tokenSize);
+            jbyte* buffer = env->GetByteArrayElements(dataArray, nullptr);
+            
+            // Copy token data using HFCopyFaceBasicToken
+            HResult copyResult = HFCopyFaceBasicToken(results.tokens[i], reinterpret_cast<char*>(buffer), tokenSize);
+            if (copyResult == HSUCCEED) {
+                // Set data and size fields
+                env->SetObjectField(token, tokenDataField, dataArray);
+                env->SetIntField(token, sizeField, tokenSize);
+                env->ReleaseByteArrayElements(dataArray, buffer, 0);
+            } else {
+                INSPIRE_LOGE("Failed to copy token data for face %d, error code: %d", i, copyResult);
+                env->ReleaseByteArrayElements(dataArray, buffer, JNI_ABORT);
+                env->DeleteLocalRef(dataArray);
+            }
+            env->SetObjectArrayElement(tokenArray, i, token);
+
+            // Release local references
+            env->DeleteLocalRef(rect);
+            env->DeleteLocalRef(angle);
+            env->DeleteLocalRef(token);
+        }
+
+        // Set arrays to MultipleFaceData
+        jfieldID rectsField = env->GetFieldID(multipleFaceDataClass, "rects", "[Lcom/insightface/sdk/inspireface/base/FaceRect;");
+        jfieldID trackIdsField = env->GetFieldID(multipleFaceDataClass, "trackIds", "[I");
+        jfieldID detConfidenceField = env->GetFieldID(multipleFaceDataClass, "detConfidence", "[F");
+        jfieldID anglesField = env->GetFieldID(multipleFaceDataClass, "angles", "[Lcom/insightface/sdk/inspireface/base/FaceEulerAngle;");
+        jfieldID tokensField = env->GetFieldID(multipleFaceDataClass, "tokens", "[Lcom/insightface/sdk/inspireface/base/FaceBasicToken;");
+
+        env->SetObjectField(multipleFaceData, rectsField, rectArray);
+        env->SetObjectField(multipleFaceData, trackIdsField, trackIdsArray);
+        env->SetObjectField(multipleFaceData, detConfidenceField, detConfArray);
+        env->SetObjectField(multipleFaceData, anglesField, angleArray);
+        env->SetObjectField(multipleFaceData, tokensField, tokenArray);
+    }
+
+    return multipleFaceData;
+}
+
+/**
+ * @brief Get face dense landmark from face token.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param token The face token object.
+ * @return The face dense landmark array.
+ */
+JNIEXPORT jobjectArray INSPIRE_FACE_JNI(InspireFace_GetFaceDenseLandmarkFromFaceToken)(JNIEnv *env, jobject thiz, jobject token) {
+    // Get token data and size from FaceBasicToken object
+    jclass tokenClass = env->GetObjectClass(token);
+    jfieldID tokenDataField = env->GetFieldID(tokenClass, "data", "[B");
+    jfieldID tokenSizeField = env->GetFieldID(tokenClass, "size", "I");
+    jbyteArray tokenDataArray = (jbyteArray)env->GetObjectField(token, tokenDataField);
+    jint tokenSize = env->GetIntField(token, tokenSizeField);
+    
+    if (tokenDataArray == nullptr) {
+        INSPIRE_LOGE("Token data array is null");
+        return nullptr;
+    }
+
+    // Get number of landmarks
+    int32_t numLandmarks = 0;
+    HFGetNumOfFaceDenseLandmark(&numLandmarks);
+
+    // Allocate memory for landmarks
+    HPoint2f *landmarks = new HPoint2f[numLandmarks];
+
+    // Create face token struct from byte array data
+    HFFaceBasicToken faceToken;
+    faceToken.size = tokenSize;
+    faceToken.data = env->GetByteArrayElements(tokenDataArray, nullptr);
+
+    // Get landmarks from token
+    HResult result = HFGetFaceDenseLandmarkFromFaceToken(faceToken, landmarks, numLandmarks);
+
+    // Release byte array elements
+    env->ReleaseByteArrayElements(tokenDataArray, (jbyte*)faceToken.data, JNI_ABORT);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face dense landmark from face token, error code: %d", result);
+        delete[] landmarks;
+        return nullptr;
+    }
+
+    // Create Point2f array to return
+    jclass point2fClass = env->FindClass("com/insightface/sdk/inspireface/base/Point2f");
+    jobjectArray landmarkArray = env->NewObjectArray(numLandmarks, point2fClass, nullptr);
+    jmethodID constructor = env->GetMethodID(point2fClass, "<init>", "()V");
+    jfieldID xField = env->GetFieldID(point2fClass, "x", "F");
+    jfieldID yField = env->GetFieldID(point2fClass, "y", "F");
+
+    // Fill array with landmark points
+    for (int i = 0; i < numLandmarks; i++) {
+        jobject point = env->NewObject(point2fClass, constructor);
+        env->SetFloatField(point, xField, landmarks[i].x);
+        env->SetFloatField(point, yField, landmarks[i].y);
+        env->SetObjectArrayElement(landmarkArray, i, point);
+    }
+
+    delete[] landmarks;
+    return landmarkArray;
+}
+
+/**
+ * @brief Extract face feature from a face token.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param streamHandle The InspireFace image stream object.
+ * @param token The face token object.
+ * @return The face feature object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_ExtractFaceFeature)(JNIEnv *env, jobject thiz, jobject session, jobject streamHandle, jobject token) {
+    // Get session handle
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+
+    // Get stream handle
+    jclass streamClass = env->GetObjectClass(streamHandle);
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    jlong streamHandleValue = env->GetLongField(streamHandle, streamHandleField);
+
+    // Get token data and size from FaceBasicToken object
+    jclass tokenClass = env->GetObjectClass(token);
+    jfieldID tokenDataField = env->GetFieldID(tokenClass, "data", "[B");
+    jfieldID tokenSizeField = env->GetFieldID(tokenClass, "size", "I");
+    jbyteArray tokenDataArray = (jbyteArray)env->GetObjectField(token, tokenDataField);
+    jint tokenSize = env->GetIntField(token, tokenSizeField);
+    
+    if (tokenDataArray == nullptr) {
+        INSPIRE_LOGE("Token data array is null");
+        return nullptr;
+    }
+
+    // Create face token struct from byte array data
+    HFFaceBasicToken faceToken;
+    faceToken.size = tokenSize;
+    faceToken.data = env->GetByteArrayElements(tokenDataArray, nullptr);
+
+    // Extract face feature
+    HFFaceFeature feature;
+    HResult result =
+      HFFaceFeatureExtract(reinterpret_cast<HFSession>(sessionHandle), reinterpret_cast<HFImageStream>(streamHandleValue), faceToken, &feature);
+
+    // Release byte array elements
+    env->ReleaseByteArrayElements(tokenDataArray, (jbyte*)faceToken.data, JNI_ABORT);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to extract face feature, error code: %d", result);
+        return nullptr;
+    }
+
+    // Create FaceFeature object to return
+    jclass faceFeatureClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeature");
+    jobject faceFeature = env->NewObject(faceFeatureClass, env->GetMethodID(faceFeatureClass, "<init>", "()V"));
+
+    // Create float array and copy feature data
+    jfloatArray dataArray = env->NewFloatArray(feature.size);
+    env->SetFloatArrayRegion(dataArray, 0, feature.size, reinterpret_cast<jfloat *>(feature.data));
+
+    // Set data field in FaceFeature object
+    jfieldID dataField = env->GetFieldID(faceFeatureClass, "data", "[F");
+    env->SetObjectField(faceFeature, dataField, dataArray);
+
+    return faceFeature;
+}
+
+/**
+ * @brief Get face alignment image.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param streamHandle The InspireFace image stream object.
+ * @param token The face token object.
+ * @return The face alignment image object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceAlignmentImage)(JNIEnv *env, jobject thiz, jobject session, jobject streamHandle,
+                                                                      jobject token) {
+    // Get session handle
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+
+    // Get stream handle
+    jclass streamClass = env->GetObjectClass(streamHandle);
+    jfieldID streamHandleField = env->GetFieldID(streamClass, "handle", "J");
+    jlong streamHandleValue = env->GetLongField(streamHandle, streamHandleField);
+
+    // Get token data and size from FaceBasicToken object
+    jclass tokenClass = env->GetObjectClass(token);
+    jfieldID tokenDataField = env->GetFieldID(tokenClass, "data", "[B");
+    jfieldID tokenSizeField = env->GetFieldID(tokenClass, "size", "I");
+    jbyteArray tokenDataArray = (jbyteArray)env->GetObjectField(token, tokenDataField);
+    jint tokenSize = env->GetIntField(token, tokenSizeField);
+    
+    if (tokenDataArray == nullptr) {
+        INSPIRE_LOGE("Token data array is null");
+        return nullptr;
+    }
+
+    // Create face token struct from byte array data
+    HFFaceBasicToken faceToken;
+    faceToken.size = tokenSize;
+    faceToken.data = env->GetByteArrayElements(tokenDataArray, nullptr);
+
+    // Get face alignment image
+    HFImageBitmap imageBitmap;
+    HResult result = HFFaceGetFaceAlignmentImage(reinterpret_cast<HFSession>(sessionHandle), reinterpret_cast<HFImageStream>(streamHandleValue),
+                                                 faceToken, &imageBitmap);
+
+    // Release byte array elements
+    env->ReleaseByteArrayElements(tokenDataArray, (jbyte*)faceToken.data, JNI_ABORT);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face alignment image, error code: %d", result);
+        return nullptr;
+    }
+
+    // Get image bitmap data
+    HFImageBitmapData bitmapData;
+    result = HFImageBitmapGetData(imageBitmap, &bitmapData);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get image bitmap data, error code: %d", result);
+        HFReleaseImageBitmap(imageBitmap);
+        return nullptr;
+    }
+
+    // Create Android Bitmap
+    jclass bitmapClass = env->FindClass("android/graphics/Bitmap");
+    jmethodID createBitmapMethod =
+      env->GetStaticMethodID(bitmapClass, "createBitmap", "(IILandroid/graphics/Bitmap$Config;)Landroid/graphics/Bitmap;");
+
+    // Get Bitmap.Config.ARGB_8888
+    jclass bitmapConfigClass = env->FindClass("android/graphics/Bitmap$Config");
+    jfieldID argb8888FieldID = env->GetStaticFieldID(bitmapConfigClass, "ARGB_8888", "Landroid/graphics/Bitmap$Config;");
+    jobject argb8888Obj = env->GetStaticObjectField(bitmapConfigClass, argb8888FieldID);
+
+    // Create bitmap with alignment image dimensions
+    jobject bitmap = env->CallStaticObjectMethod(bitmapClass, createBitmapMethod, bitmapData.width, bitmapData.height, argb8888Obj);
+
+    // Copy pixels to bitmap
+    AndroidBitmapInfo bitmapInfo;
+    void *pixels;
+    if (AndroidBitmap_getInfo(env, bitmap, &bitmapInfo) < 0 || AndroidBitmap_lockPixels(env, bitmap, &pixels) < 0) {
+        INSPIRE_LOGE("Failed to get bitmap info or lock pixels, error code: %d", result);
+        HFReleaseImageBitmap(imageBitmap);
+        return nullptr;
+    }
+
+    // Convert BGR to ARGB
+    uint8_t *src = (uint8_t *)bitmapData.data;
+    uint32_t *dst = (uint32_t *)pixels;
+    for (int i = 0; i < bitmapData.width * bitmapData.height; i++) {
+        // BGR to ARGB conversion
+        uint8_t r = src[i * 3 + 2];
+        uint8_t g = src[i * 3 + 1];
+        uint8_t b = src[i * 3];
+        dst[i] = (0xFF << 24) | (b << 16) | (g << 8) | r;
+    }
+
+    AndroidBitmap_unlockPixels(env, bitmap);
+    HFReleaseImageBitmap(imageBitmap);
+
+    return bitmap;
+}
+
+/**
+ * @brief Set track preview size.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param previewSize The preview size.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetTrackPreviewSize)(JNIEnv *env, jobject thiz, jobject session, jint previewSize) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetTrackPreviewSize((HFSession)sessionHandle, previewSize);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set track preview size, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Set filter minimum face pixel size.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param minSize The minimum face pixel size.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetFilterMinimumFacePixelSize)(JNIEnv *env, jobject thiz, jobject session, jint minSize) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetFilterMinimumFacePixelSize((HFSession)sessionHandle, minSize);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set filter minimum face pixel size, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Set face detect threshold.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param threshold The face detect threshold.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetFaceDetectThreshold)(JNIEnv *env, jobject thiz, jobject session, jfloat threshold) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetFaceDetectThreshold((HFSession)sessionHandle, threshold);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set face detect threshold, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Set track mode smooth ratio.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param ratio The track mode smooth ratio.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetTrackModeSmoothRatio)(JNIEnv *env, jobject thiz, jobject session, jfloat ratio) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetTrackModeSmoothRatio((HFSession)sessionHandle, ratio);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set track mode smooth ratio, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Set track mode num smooth cache frame.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param num The track mode num smooth cache frame.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetTrackModeNumSmoothCacheFrame)(JNIEnv *env, jobject thiz, jobject session, jint num) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetTrackModeNumSmoothCacheFrame((HFSession)sessionHandle, num);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set track mode num smooth cache frame, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Set track mode detect interval.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The InspireFace session object.
+ * @param interval The track mode detect interval.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetTrackModeDetectInterval)(JNIEnv *env, jobject thiz, jobject session, jint interval) {
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID sessionHandleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, sessionHandleField);
+    auto result = HFSessionSetTrackModeDetectInterval((HFSession)sessionHandle, interval);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set track mode detect interval, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Enable feature hub data.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param configuration The configuration object.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_FeatureHubDataEnable)(JNIEnv *env, jobject thiz, jobject configuration) {
+    jclass configClass = env->GetObjectClass(configuration);
+
+    jfieldID primaryKeyModeField = env->GetFieldID(configClass, "primaryKeyMode", "I");
+    jfieldID enablePersistenceField = env->GetFieldID(configClass, "enablePersistence", "I");
+    jfieldID persistenceDbPathField = env->GetFieldID(configClass, "persistenceDbPath", "Ljava/lang/String;");
+    jfieldID searchThresholdField = env->GetFieldID(configClass, "searchThreshold", "F");
+    jfieldID searchModeField = env->GetFieldID(configClass, "searchMode", "I");
+    HFFeatureHubConfiguration config;
+    config.primaryKeyMode = (HFPKMode)env->GetIntField(configuration, primaryKeyModeField);
+    config.enablePersistence = env->GetIntField(configuration, enablePersistenceField);
+
+    // Add null check for dbPath
+    jstring dbPath = (jstring)env->GetObjectField(configuration, persistenceDbPathField);
+    if (dbPath != nullptr) {
+        const char *nativeDbPath = env->GetStringUTFChars(dbPath, nullptr);
+        if (nativeDbPath != nullptr) {
+            config.persistenceDbPath = const_cast<char *>(nativeDbPath);
+        } else {
+            config.persistenceDbPath[0] = '\0';
+        }
+    } else {
+        config.persistenceDbPath[0] = '\0';
+    }
+
+    config.searchThreshold = env->GetFloatField(configuration, searchThresholdField);
+    config.searchMode = (HFSearchMode)env->GetIntField(configuration, searchModeField);
+
+    // Remove debug logs that might interfere with error handling
+    auto result = HFFeatureHubDataEnable(config);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to enable feature hub data, error code: %d", result);
+        return false;
+    }
+
+    return true;
+}
+
+/**
+ * @brief Disable feature hub data.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return True if the feature hub data is disabled successfully, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_FeatureHubDataDisable)(JNIEnv *env, jobject thiz) {
+    auto result = HFFeatureHubDataDisable();
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to disable feature hub data, error code: %d", result);
+        return false;
+    }
+
+    return true;
+}
+
+/**
+ * @brief Insert a feature into the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param newIdentity The new identity object.
+ * @return True if the feature is inserted successfully, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_FeatureHubInsertFeature)(JNIEnv *env, jobject thiz, jobject newIdentity) {
+    // Get FaceFeatureIdentity class and fields
+    jclass identityClass = env->GetObjectClass(newIdentity);
+    jfieldID idField = env->GetFieldID(identityClass, "id", "J");
+    jfieldID featureField = env->GetFieldID(identityClass, "feature", "Lcom/insightface/sdk/inspireface/base/FaceFeature;");
+
+    // Get feature object
+    jobject featureObj = env->GetObjectField(newIdentity, featureField);
+    if (featureObj == nullptr) {
+        INSPIRE_LOGE("Feature object is null");
+        return false;
+    }
+
+    // Get FaceFeature class and fields
+    jclass featureClass = env->GetObjectClass(featureObj);
+    jfieldID dataField = env->GetFieldID(featureClass, "data", "[F");
+
+    // Get feature data array
+    jfloatArray dataArray = (jfloatArray)env->GetObjectField(featureObj, dataField);
+    if (dataArray == nullptr) {
+        INSPIRE_LOGE("Feature data array is null");
+        return false;
+    }
+
+    // Get feature data
+    jsize length = env->GetArrayLength(dataArray);
+    jfloat *data = env->GetFloatArrayElements(dataArray, nullptr);
+
+    // Create HFFaceFeature
+    HFFaceFeature feature;
+    feature.data = data;
+    feature.size = length;
+
+    // Create HFFaceFeatureIdentity
+    HFFaceFeatureIdentity identity;
+    identity.id = env->GetLongField(newIdentity, idField);
+    identity.feature = &feature;
+
+    // Insert feature and get allocated ID
+    HFaceId allocId = -1;
+    auto result = HFFeatureHubInsertFeature(identity, &allocId);
+
+    // Release array elements
+    env->ReleaseFloatArrayElements(dataArray, data, 0);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to insert feature, error code: %d", result);
+        return false;
+    }
+
+    // Update ID field with allocated ID
+    env->SetLongField(newIdentity, idField, allocId);
+    return true;
+}
+
+/**
+ * @brief Search for a feature in the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param searchFeature The search feature object.
+ * @param confidence The confidence.
+ * @return The search result object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_FeatureHubFaceSearch)(JNIEnv *env, jobject thiz, jobject searchFeature, jfloat confidence) {
+    // Get FaceFeature class and fields
+    jclass featureClass = env->GetObjectClass(searchFeature);
+    jfieldID dataField = env->GetFieldID(featureClass, "data", "[F");
+
+    // Get feature data array
+    jfloatArray dataArray = (jfloatArray)env->GetObjectField(searchFeature, dataField);
+    if (dataArray == nullptr) {
+        INSPIRE_LOGE("Feature data array is null");
+        return nullptr;
+    }
+
+    // Get feature data
+    jsize length = env->GetArrayLength(dataArray);
+    jfloat *data = env->GetFloatArrayElements(dataArray, nullptr);
+
+    // Create HFFaceFeature
+    HFFaceFeature feature;
+    feature.data = data;
+    feature.size = length;
+
+    // Create HFFaceFeatureIdentity to store search result
+    float searchConfidence = 0.0f;
+    HFFaceFeatureIdentity mostSimilar;
+    auto result = HFFeatureHubFaceSearch(feature, &searchConfidence, &mostSimilar);
+
+    // Release array elements
+    env->ReleaseFloatArrayElements(dataArray, data, 0);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to search feature, error code: %d", result);
+        return nullptr;
+    }
+
+    // Create FaceFeatureIdentity object to return
+    jclass identityClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeatureIdentity");
+    jmethodID constructor = env->GetMethodID(identityClass, "<init>", "()V");
+    jobject identityObj = env->NewObject(identityClass, constructor);
+
+    // Set id field
+    jfieldID idField = env->GetFieldID(identityClass, "id", "J");
+    env->SetLongField(identityObj, idField, mostSimilar.id);
+
+    // Create FaceFeature object and set data
+    jclass faceFeatureClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeature");
+    jmethodID featureConstructor = env->GetMethodID(faceFeatureClass, "<init>", "()V");
+    jobject featureObj = env->NewObject(faceFeatureClass, featureConstructor);
+
+    // Create and set float array for feature data
+    jfloatArray featureArray = env->NewFloatArray(mostSimilar.feature->size);
+    env->SetFloatArrayRegion(featureArray, 0, mostSimilar.feature->size, mostSimilar.feature->data);
+    env->SetObjectField(featureObj, dataField, featureArray);
+
+    // Set feature field in identity object
+    jfieldID featureField = env->GetFieldID(identityClass, "feature", "Lcom/insightface/sdk/inspireface/base/FaceFeature;");
+    env->SetObjectField(identityObj, featureField, featureObj);
+
+    // Set searchConfidence field
+    jfieldID searchConfidenceField = env->GetFieldID(identityClass, "searchConfidence", "F");
+    env->SetFloatField(identityObj, searchConfidenceField, searchConfidence);
+
+    return identityObj;
+}
+
+/**
+ * @brief Search for top-k features in the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param searchFeature The search feature object.
+ * @param topK The top-k.
+ * @return The search results object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_FeatureHubFaceSearchTopK)(JNIEnv *env, jobject thiz, jobject searchFeature, jint topK) {
+    // Get feature data array
+    jclass faceFeatureClass = env->GetObjectClass(searchFeature);
+    jfieldID dataField = env->GetFieldID(faceFeatureClass, "data", "[F");
+    jfloatArray dataArray = (jfloatArray)env->GetObjectField(searchFeature, dataField);
+    jfloat *data = env->GetFloatArrayElements(dataArray, 0);
+    jsize length = env->GetArrayLength(dataArray);
+
+    // Create HFFaceFeature
+    HFFaceFeature feature;
+    feature.data = data;
+    feature.size = length;
+
+    // Create HFSearchTopKResults to store search results
+    HFSearchTopKResults results;
+    auto result = HFFeatureHubFaceSearchTopK(feature, topK, &results);
+
+    // Release array elements
+    env->ReleaseFloatArrayElements(dataArray, data, 0);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to search top-k features, error code: %d", result);
+        return nullptr;
+    }
+
+    // Create SearchTopKResults object to return
+    jclass searchResultsClass = env->FindClass("com/insightface/sdk/inspireface/base/SearchTopKResults");
+    jmethodID constructor = env->GetMethodID(searchResultsClass, "<init>", "()V");
+    jobject searchResultsObj = env->NewObject(searchResultsClass, constructor);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(searchResultsClass, "num", "I");
+    env->SetIntField(searchResultsObj, numField, results.size);
+
+    // Create and set confidence array
+    jfloatArray confidenceArray = env->NewFloatArray(results.size);
+    env->SetFloatArrayRegion(confidenceArray, 0, results.size, results.confidence);
+    jfieldID confidenceField = env->GetFieldID(searchResultsClass, "confidence", "[F");
+    env->SetObjectField(searchResultsObj, confidenceField, confidenceArray);
+
+    // Create and set ids array
+    jlongArray idsArray = env->NewLongArray(results.size);
+    env->SetLongArrayRegion(idsArray, 0, results.size, results.ids);
+    jfieldID idsField = env->GetFieldID(searchResultsClass, "ids", "[J");
+    env->SetObjectField(searchResultsObj, idsField, idsArray);
+
+    return searchResultsObj;
+}
+
+/**
+ * @brief Remove a feature from the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param id The id of the feature to remove.
+ * @return True if the feature is removed successfully, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_FeatureHubFaceRemove)(JNIEnv *env, jobject thiz, jlong id) {
+    auto result = HFFeatureHubFaceRemove(id);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to remove feature, error code: %d", result);
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Update a feature in the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param featureIdentity The feature identity object.
+ * @return True if the feature is updated successfully, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_FeatureHubFaceUpdate)(JNIEnv *env, jobject thiz, jobject featureIdentity) {
+    // Get FaceFeatureIdentity class and fields
+    jclass identityClass = env->GetObjectClass(featureIdentity);
+    jfieldID idField = env->GetFieldID(identityClass, "id", "J");
+    jfieldID featureField = env->GetFieldID(identityClass, "feature", "Lcom/insightface/sdk/inspireface/base/FaceFeature;");
+
+    // Get id value
+    jlong id = env->GetLongField(featureIdentity, idField);
+
+    // Get feature object
+    jobject featureObj = env->GetObjectField(featureIdentity, featureField);
+    jclass featureClass = env->GetObjectClass(featureObj);
+    jfieldID dataField = env->GetFieldID(featureClass, "data", "[F");
+
+    // Get feature data array
+    jfloatArray dataArray = (jfloatArray)env->GetObjectField(featureObj, dataField);
+    jsize length = env->GetArrayLength(dataArray);
+    jfloat *data = env->GetFloatArrayElements(dataArray, nullptr);
+
+    // Create HFFaceFeature
+    HFFaceFeature feature;
+    feature.data = data;
+    feature.size = length;
+
+    // Create HFFaceFeatureIdentity
+    HFFaceFeatureIdentity identity;
+    identity.id = id;
+    identity.feature = &feature;
+
+    // Update feature
+    auto result = HFFeatureHubFaceUpdate(identity);
+
+    // Release array elements
+    env->ReleaseFloatArrayElements(dataArray, data, 0);
+
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to update feature, error code: %d", result);
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Get a face identity from the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param id The id of the face identity to get.
+ * @return The face identity object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_FeatureHubGetFaceIdentity)(JNIEnv *env, jobject thiz, jlong id) {
+    // Create HFFaceFeatureIdentity
+    HFFaceFeatureIdentity identity;
+    auto result = HFFeatureHubGetFaceIdentity(id, &identity);
+    if (result != HSUCCEED) {
+        return nullptr;
+    }
+
+    // Create FaceFeature object
+    jclass featureClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeature");
+    jmethodID featureConstructor = env->GetMethodID(featureClass, "<init>", "()V");
+    jobject featureObj = env->NewObject(featureClass, featureConstructor);
+
+    // Set feature data array
+    jfieldID dataField = env->GetFieldID(featureClass, "data", "[F");
+    jfloatArray dataArray = env->NewFloatArray(identity.feature->size);
+    env->SetFloatArrayRegion(dataArray, 0, identity.feature->size, identity.feature->data);
+    env->SetObjectField(featureObj, dataField, dataArray);
+
+    // Create FaceFeatureIdentity object
+    jclass identityClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeatureIdentity");
+    jmethodID identityConstructor = env->GetMethodID(identityClass, "<init>", "()V");
+    jobject identityObj = env->NewObject(identityClass, identityConstructor);
+
+    // Set id and feature fields
+    jfieldID idField = env->GetFieldID(identityClass, "id", "J");
+    jfieldID featureField = env->GetFieldID(identityClass, "feature", "Lcom/insightface/sdk/inspireface/base/FaceFeature;");
+    env->SetLongField(identityObj, idField, identity.id);
+    env->SetObjectField(identityObj, featureField, featureObj);
+
+    return identityObj;
+}
+
+/**
+ * @brief Get the number of faces in the feature hub.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return The number of faces in the feature hub.
+ */
+JNIEXPORT jint INSPIRE_FACE_JNI(InspireFace_FeatureHubGetFaceCount)(JNIEnv *env, jobject thiz) {
+    HInt32 count;
+    auto result = HFFeatureHubGetFaceCount(&count);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face count, error code: %d", result);
+        return -1;
+    }
+    return count;
+}
+
+/**
+ * @brief Set the face search threshold.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param threshold The face search threshold.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_FeatureHubFaceSearchThresholdSetting)(JNIEnv *env, jobject thiz, jfloat threshold) {
+    auto result = HFFeatureHubFaceSearchThresholdSetting(threshold);
+    if (result != HSUCCEED) {
+        INSPIRE_LOGE("Failed to set face search threshold, error code: %d", result);
+    }
+}
+
+/**
+ * @brief Get the length of the feature.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return The length of the feature.
+ */
+JNIEXPORT jint INSPIRE_FACE_JNI(InspireFace_GetFeatureLength)(JNIEnv *env, jobject thiz) {
+    HInt32 length;
+    auto result = HFGetFeatureLength(&length);
+    return length;
+}
+
+/**
+ * @brief Update the cosine similarity converter.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param config The cosine similarity converter configuration object.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_UpdateCosineSimilarityConverter)(JNIEnv *env, jobject thiz, jobject config) {
+    jclass configClass = env->GetObjectClass(config);
+    jfieldID thresholdField = env->GetFieldID(configClass, "threshold", "F");
+    jfieldID middleScoreField = env->GetFieldID(configClass, "middleScore", "F");
+    jfieldID steepnessField = env->GetFieldID(configClass, "steepness", "F");
+    jfieldID outputMinField = env->GetFieldID(configClass, "outputMin", "F");
+    jfieldID outputMaxField = env->GetFieldID(configClass, "outputMax", "F");
+
+    HFSimilarityConverterConfig converterConfig;
+    converterConfig.threshold = env->GetFloatField(config, thresholdField);
+    converterConfig.middleScore = env->GetFloatField(config, middleScoreField);
+    converterConfig.steepness = env->GetFloatField(config, steepnessField);
+    converterConfig.outputMin = env->GetFloatField(config, outputMinField);
+    converterConfig.outputMax = env->GetFloatField(config, outputMaxField);
+    HFUpdateCosineSimilarityConverter(converterConfig);
+}
+
+/**
+ * @brief Get the cosine similarity converter.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return The cosine similarity converter object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetCosineSimilarityConverter)(JNIEnv *env, jobject thiz) {
+    HFSimilarityConverterConfig converterConfig;
+    HFGetCosineSimilarityConverter(&converterConfig);
+    jclass configClass = env->FindClass("com/insightface/sdk/inspireface/base/SimilarityConverterConfig");
+    jmethodID constructor = env->GetMethodID(configClass, "<init>", "()V");
+    jobject configObj = env->NewObject(configClass, constructor);
+
+    // Get field IDs
+    jfieldID thresholdField = env->GetFieldID(configClass, "threshold", "F");
+    jfieldID middleScoreField = env->GetFieldID(configClass, "middleScore", "F");
+    jfieldID steepnessField = env->GetFieldID(configClass, "steepness", "F");
+    jfieldID outputMinField = env->GetFieldID(configClass, "outputMin", "F");
+    jfieldID outputMaxField = env->GetFieldID(configClass, "outputMax", "F");
+
+    // Set fields
+    env->SetFloatField(configObj, thresholdField, converterConfig.threshold);
+    env->SetFloatField(configObj, middleScoreField, converterConfig.middleScore);
+    env->SetFloatField(configObj, steepnessField, converterConfig.steepness);
+    env->SetFloatField(configObj, outputMinField, converterConfig.outputMin);
+    env->SetFloatField(configObj, outputMaxField, converterConfig.outputMax);
+
+    return configObj;
+}
+
+/**
+ * @brief Convert cosine similarity to percentage.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param similarity The cosine similarity.
+ * @return The percentage.
+ */
+JNIEXPORT jfloat INSPIRE_FACE_JNI(InspireFace_CosineSimilarityConvertToPercentage)(JNIEnv *env, jobject thiz, jfloat similarity) {
+    HFloat result;
+    HFCosineSimilarityConvertToPercentage(similarity, &result);
+    return result;
+}
+
+/**
+ * @brief Compare two face features.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param feature1 The first face feature.
+ * @param feature2 The second face feature.
+ * @return The similarity between the two face features.
+ */
+JNIEXPORT jfloat INSPIRE_FACE_JNI(InspireFace_FaceComparison)(JNIEnv *env, jobject thiz, jobject feature1, jobject feature2) {
+    if (feature1 == nullptr || feature2 == nullptr) {
+        return -1.0f;
+    }
+
+    // Get FaceFeature class and data field
+    jclass featureClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceFeature");
+    jfieldID dataField = env->GetFieldID(featureClass, "data", "[F");
+
+    // Get float arrays from FaceFeature objects
+    jfloatArray data1Array = (jfloatArray)env->GetObjectField(feature1, dataField);
+    jfloatArray data2Array = (jfloatArray)env->GetObjectField(feature2, dataField);
+
+    if (data1Array == nullptr || data2Array == nullptr) {
+        return -1.0f;
+    }
+
+    // Get array lengths
+    jsize len1 = env->GetArrayLength(data1Array);
+    jsize len2 = env->GetArrayLength(data2Array);
+
+    if (len1 != len2) {
+        return -1.0f;
+    }
+
+    // Get float data
+    jfloat *data1 = env->GetFloatArrayElements(data1Array, nullptr);
+    jfloat *data2 = env->GetFloatArrayElements(data2Array, nullptr);
+
+    // Create HFFaceFeature structs
+    HFFaceFeature ft1, ft2;
+    ft1.data = data1;
+    ft1.size = len1;
+    ft2.data = data2;
+    ft2.size = len2;
+
+    // Compare features
+    float compareResult;
+    HResult ret = HFFaceComparison(ft1, ft2, &compareResult);
+
+    // Release arrays
+    env->ReleaseFloatArrayElements(data1Array, data1, 0);
+    env->ReleaseFloatArrayElements(data2Array, data2, 0);
+
+    if (ret != HSUCCEED) {
+        return -1.0f;
+    }
+
+    return compareResult;
+}
+
+/**
+ * @brief Get the recommended cosine threshold.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return The recommended cosine threshold.
+ */
+JNIEXPORT jfloat INSPIRE_FACE_JNI(InspireFace_GetRecommendedCosineThreshold)(JNIEnv *env, jobject thiz) {
+    HFloat threshold;
+    HFGetRecommendedCosineThreshold(&threshold);
+    return threshold;
+}
+
+/**
+ * @brief Process multiple faces in the pipeline.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @param streamHandle The stream handle object.
+ * @param faces The faces object.
+ * @param parameter The parameter object.
+ * @return True if the multiple faces are processed successfully, false otherwise.
+ */
+JNIEXPORT jboolean INSPIRE_FACE_JNI(InspireFace_MultipleFacePipelineProcess)(JNIEnv *env, jobject thiz, jobject session, jobject streamHandle,
+                                                                             jobject faces, jobject parameter) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Get stream handle
+    jclass streamClass = env->FindClass("com/insightface/sdk/inspireface/base/ImageStream");
+    handleField = env->GetFieldID(streamClass, "handle", "J");
+    jlong streamHandleValue = env->GetLongField(streamHandle, handleField);
+
+    // Get faces data
+    jclass facesClass = env->FindClass("com/insightface/sdk/inspireface/base/MultipleFaceData");
+    jfieldID numField = env->GetFieldID(facesClass, "detectedNum", "I");
+    jfieldID tokensField = env->GetFieldID(facesClass, "tokens", "[Lcom/insightface/sdk/inspireface/base/FaceBasicToken;");
+
+    jint detectedNum = env->GetIntField(faces, numField);
+    jobjectArray tokenArray = (jobjectArray)env->GetObjectField(faces, tokensField);
+
+    // Create HFMultipleFaceData struct
+    HFMultipleFaceData faceData;
+    faceData.detectedNum = detectedNum;
+
+    // Get token data
+    HFFaceBasicToken *tokens = nullptr;
+    std::vector<jbyteArray> tokenDataArrays;
+    std::vector<jbyte*> tokenBuffers;
+    
+    if (detectedNum > 0) {
+        jclass tokenClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceBasicToken");
+        jfieldID tokenDataField = env->GetFieldID(tokenClass, "data", "[B");
+        jfieldID sizeField = env->GetFieldID(tokenClass, "size", "I");
+
+        tokens = new HFFaceBasicToken[detectedNum];
+        tokenDataArrays.resize(detectedNum);
+        tokenBuffers.resize(detectedNum);
+        
+        for (int i = 0; i < detectedNum; i++) {
+            jobject token = env->GetObjectArrayElement(tokenArray, i);
+            jbyteArray dataArray = (jbyteArray)env->GetObjectField(token, tokenDataField);
+            jint size = env->GetIntField(token, sizeField);
+            
+            if (dataArray != nullptr) {
+                tokens[i].size = size;
+                tokens[i].data = env->GetByteArrayElements(dataArray, nullptr);
+                tokenDataArrays[i] = dataArray;
+                tokenBuffers[i] = (jbyte*)tokens[i].data;
+            } else {
+                tokens[i].size = 0;
+                tokens[i].data = nullptr;
+                tokenDataArrays[i] = nullptr;
+                tokenBuffers[i] = nullptr;
+            }
+            env->DeleteLocalRef(token);
+        }
+        faceData.tokens = tokens;
+    } else {
+        faceData.tokens = nullptr;
+    }
+
+    // Get custom parameter fields
+    jclass paramClass = env->FindClass("com/insightface/sdk/inspireface/base/CustomParameter");
+    jfieldID enableRecognitionField = env->GetFieldID(paramClass, "enableRecognition", "I");
+    jfieldID enableLivenessField = env->GetFieldID(paramClass, "enableLiveness", "I");
+    jfieldID enableIrLivenessField = env->GetFieldID(paramClass, "enableIrLiveness", "I");
+    jfieldID enableMaskDetectField = env->GetFieldID(paramClass, "enableMaskDetect", "I");
+    jfieldID enableFaceQualityField = env->GetFieldID(paramClass, "enableFaceQuality", "I");
+    jfieldID enableFaceAttributeField = env->GetFieldID(paramClass, "enableFaceAttribute", "I");
+    jfieldID enableInteractionLivenessField = env->GetFieldID(paramClass, "enableInteractionLiveness", "I");
+    jfieldID enableFacePoseField = env->GetFieldID(paramClass, "enableFacePose", "I");
+    jfieldID enableFaceEmotionField = env->GetFieldID(paramClass, "enableFaceEmotion", "I");
+    // Get parameter values
+    HFSessionCustomParameter customParam;
+    customParam.enable_recognition = env->GetIntField(parameter, enableRecognitionField);
+    customParam.enable_liveness = env->GetIntField(parameter, enableLivenessField);
+    customParam.enable_ir_liveness = env->GetIntField(parameter, enableIrLivenessField);
+    customParam.enable_mask_detect = env->GetIntField(parameter, enableMaskDetectField);
+    customParam.enable_face_quality = env->GetIntField(parameter, enableFaceQualityField);
+    customParam.enable_face_attribute = env->GetIntField(parameter, enableFaceAttributeField);
+    customParam.enable_interaction_liveness = env->GetIntField(parameter, enableInteractionLivenessField);
+    customParam.enable_face_pose = env->GetIntField(parameter, enableFacePoseField);
+    customParam.enable_face_emotion = env->GetIntField(parameter, enableFaceEmotionField);
+    // Call native function
+    HResult ret = HFMultipleFacePipelineProcess((HFSession)sessionHandle, (HFImageStream)streamHandleValue, &faceData, customParam);
+
+    // Clean up allocated memory
+    if (tokens != nullptr) {
+        for (int i = 0; i < detectedNum; i++) {
+            if (tokenDataArrays[i] != nullptr && tokenBuffers[i] != nullptr) {
+                env->ReleaseByteArrayElements(tokenDataArrays[i], tokenBuffers[i], JNI_ABORT);
+            }
+        }
+        delete[] tokens;
+    }
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to process multiple faces, error code: %d", ret);
+        return false;
+    }
+
+    return true;
+}
+
+/**
+ * @brief Get the RGB liveness confidence.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The RGB liveness confidence object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetRGBLivenessConfidence)(JNIEnv *env, jobject thiz, jobject session) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Create native confidence struct
+    HFRGBLivenessConfidence confidence;
+    HResult ret = HFGetRGBLivenessConfidence((HFSession)sessionHandle, &confidence);
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get RGB liveness confidence, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java RGBLivenessConfidence object
+    jclass confidenceClass = env->FindClass("com/insightface/sdk/inspireface/base/RGBLivenessConfidence");
+    jobject confidenceObj = env->AllocObject(confidenceClass);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(confidenceClass, "num", "I");
+    env->SetIntField(confidenceObj, numField, confidence.num);
+
+    // Set confidence array field
+    jfieldID confidenceField = env->GetFieldID(confidenceClass, "confidence", "[F");
+    jfloatArray confidenceArray = env->NewFloatArray(confidence.num);
+    float *confidencePtr = confidence.confidence;
+    env->SetFloatArrayRegion(confidenceArray, 0, confidence.num, confidencePtr);
+    env->SetObjectField(confidenceObj, confidenceField, confidenceArray);
+
+    return confidenceObj;
+}
+
+/**
+ * @brief Get the face quality confidence.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face quality confidence object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceQualityConfidence)(JNIEnv *env, jobject thiz, jobject session) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Create native confidence struct
+    HFFaceQualityConfidence confidence;
+    HResult ret = HFGetFaceQualityConfidence((HFSession)sessionHandle, &confidence);
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face quality confidence, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceQualityConfidence object
+    jclass confidenceClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceQualityConfidence");
+    jobject confidenceObj = env->AllocObject(confidenceClass);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(confidenceClass, "num", "I");
+    env->SetIntField(confidenceObj, numField, confidence.num);
+
+    // Set confidence array field
+    jfieldID confidenceField = env->GetFieldID(confidenceClass, "confidence", "[F");
+    jfloatArray confidenceArray = env->NewFloatArray(confidence.num);
+    float *confidencePtr = confidence.confidence;
+    env->SetFloatArrayRegion(confidenceArray, 0, confidence.num, confidencePtr);
+    env->SetObjectField(confidenceObj, confidenceField, confidenceArray);
+
+    return confidenceObj;
+}
+
+/**
+ * @brief Get the face mask confidence.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face mask confidence object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceMaskConfidence)(JNIEnv *env, jobject thiz, jobject session) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Create native confidence struct
+    HFFaceMaskConfidence confidence;
+    HResult ret = HFGetFaceMaskConfidence((HFSession)sessionHandle, &confidence);
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face mask confidence, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceMaskConfidence object
+    jclass confidenceClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceMaskConfidence");
+    jobject confidenceObj = env->AllocObject(confidenceClass);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(confidenceClass, "num", "I");
+    env->SetIntField(confidenceObj, numField, confidence.num);
+
+    // Set confidence array field
+    jfieldID confidenceField = env->GetFieldID(confidenceClass, "confidence", "[F");
+    jfloatArray confidenceArray = env->NewFloatArray(confidence.num);
+    float *confidencePtr = confidence.confidence;
+    env->SetFloatArrayRegion(confidenceArray, 0, confidence.num, confidencePtr);
+    env->SetObjectField(confidenceObj, confidenceField, confidenceArray);
+
+    return confidenceObj;
+}
+
+/**
+ * @brief Get the face interaction state result.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face interaction state result object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceInteractionStateResult)(JNIEnv *env, jobject thiz, jobject session) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Create native state struct
+    HFFaceInteractionState state;
+    HResult ret = HFGetFaceInteractionStateResult((HFSession)sessionHandle, &state);
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face interaction state result, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceInteractionState object
+    jclass stateClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceInteractionState");
+    jobject stateObj = env->AllocObject(stateClass);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(stateClass, "num", "I");
+    env->SetIntField(stateObj, numField, state.num);
+
+    // Set leftEyeStatusConfidence array field
+    jfieldID leftEyeField = env->GetFieldID(stateClass, "leftEyeStatusConfidence", "[F");
+    jfloatArray leftEyeArray = env->NewFloatArray(state.num);
+    float *leftEyePtr = state.leftEyeStatusConfidence;
+    env->SetFloatArrayRegion(leftEyeArray, 0, state.num, leftEyePtr);
+    env->SetObjectField(stateObj, leftEyeField, leftEyeArray);
+
+    // Set rightEyeStatusConfidence array field
+    jfieldID rightEyeField = env->GetFieldID(stateClass, "rightEyeStatusConfidence", "[F");
+    jfloatArray rightEyeArray = env->NewFloatArray(state.num);
+    float *rightEyePtr = state.rightEyeStatusConfidence;
+    env->SetFloatArrayRegion(rightEyeArray, 0, state.num, rightEyePtr);
+    env->SetObjectField(stateObj, rightEyeField, rightEyeArray);
+
+    return stateObj;
+}
+
+/**
+ * @brief Get the face interaction actions result.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face interaction actions result object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceInteractionActionsResult)(JNIEnv *env, jobject thiz, jobject session) {
+    // Get session handle
+    jclass sessionClass = env->FindClass("com/insightface/sdk/inspireface/base/Session");
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+
+    // Create native actions struct
+    HFFaceInteractionsActions actions;
+    HResult ret = HFGetFaceInteractionActionsResult((HFSession)sessionHandle, &actions);
+
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face interaction actions result, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceInteractionsActions object
+    jclass actionsClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceInteractionsActions");
+    jobject actionsObj = env->AllocObject(actionsClass);
+
+    // Set num field
+    jfieldID numField = env->GetFieldID(actionsClass, "num", "I");
+    env->SetIntField(actionsObj, numField, actions.num);
+
+    // Set normal array field
+    jfieldID normalField = env->GetFieldID(actionsClass, "normal", "[I");
+    jintArray normalArray = env->NewIntArray(actions.num);
+    int *normalPtr = actions.normal;
+    env->SetIntArrayRegion(normalArray, 0, actions.num, normalPtr);
+    env->SetObjectField(actionsObj, normalField, normalArray);
+
+    // Set shake array field
+    jfieldID shakeField = env->GetFieldID(actionsClass, "shake", "[I");
+    jintArray shakeArray = env->NewIntArray(actions.num);
+    int *shakePtr = actions.shake;
+    env->SetIntArrayRegion(shakeArray, 0, actions.num, shakePtr);
+    env->SetObjectField(actionsObj, shakeField, shakeArray);
+
+    // Set jawOpen array field
+    jfieldID jawOpenField = env->GetFieldID(actionsClass, "jawOpen", "[I");
+    jintArray jawOpenArray = env->NewIntArray(actions.num);
+    int *jawOpenPtr = actions.jawOpen;
+    env->SetIntArrayRegion(jawOpenArray, 0, actions.num, jawOpenPtr);
+    env->SetObjectField(actionsObj, jawOpenField, jawOpenArray);
+
+    // Set headRaise array field
+    jfieldID headRaiseField = env->GetFieldID(actionsClass, "headRaise", "[I");
+    jintArray headRaiseArray = env->NewIntArray(actions.num);
+    int *headRaisePtr = actions.headRaise;
+    env->SetIntArrayRegion(headRaiseArray, 0, actions.num, headRaisePtr);
+    env->SetObjectField(actionsObj, headRaiseField, headRaiseArray);
+
+    // Set blink array field
+    jfieldID blinkField = env->GetFieldID(actionsClass, "blink", "[I");
+    jintArray blinkArray = env->NewIntArray(actions.num);
+    int *blinkPtr = actions.blink;
+    env->SetIntArrayRegion(blinkArray, 0, actions.num, blinkPtr);
+    env->SetObjectField(actionsObj, blinkField, blinkArray);
+
+    return actionsObj;
+}
+
+/**
+ * @brief Get the face attribute result.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face attribute result object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceAttributeResult)(JNIEnv *env, jobject thiz, jobject session) {
+    // Validate input parameters
+    if (!env || !session) {
+        INSPIRE_LOGE("Invalid input parameters");
+        return nullptr;
+    }
+
+    // Get session handle
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+    if (!sessionHandle) {
+        INSPIRE_LOGE("Invalid session handle");
+        return nullptr;
+    }
+
+    // Get face attribute results
+    HFFaceAttributeResult results = {};
+    HResult ret = HFGetFaceAttributeResult((HFSession)sessionHandle, &results);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face attribute result, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceAttributeResult object
+    jclass attributeClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceAttributeResult");
+    if (!attributeClass) {
+        INSPIRE_LOGE("Failed to find FaceAttributeResult class");
+        return nullptr;
+    }
+
+    jmethodID constructor = env->GetMethodID(attributeClass, "<init>", "()V");
+    jobject attributeObj = env->NewObject(attributeClass, constructor);
+    if (!attributeObj) {
+        INSPIRE_LOGE("Failed to create FaceAttributeResult object");
+        return nullptr;
+    }
+
+    // Set fields
+    jfieldID numField = env->GetFieldID(attributeClass, "num", "I");
+    jfieldID raceField = env->GetFieldID(attributeClass, "race", "[I");
+    jfieldID genderField = env->GetFieldID(attributeClass, "gender", "[I");
+    jfieldID ageBracketField = env->GetFieldID(attributeClass, "ageBracket", "[I");
+
+    if (!numField || !raceField || !genderField || !ageBracketField) {
+        INSPIRE_LOGE("Failed to get field IDs");
+        return nullptr;
+    }
+
+    // Set num
+    env->SetIntField(attributeObj, numField, results.num);
+
+    // Set arrays
+    jintArray raceArray = env->NewIntArray(results.num);
+    jintArray genderArray = env->NewIntArray(results.num);
+    jintArray ageBracketArray = env->NewIntArray(results.num);
+
+    if (!raceArray || !genderArray || !ageBracketArray) {
+        INSPIRE_LOGE("Failed to create arrays");
+        if (raceArray)
+            env->DeleteLocalRef(raceArray);
+        if (genderArray)
+            env->DeleteLocalRef(genderArray);
+        if (ageBracketArray)
+            env->DeleteLocalRef(ageBracketArray);
+        return nullptr;
+    }
+
+    env->SetIntArrayRegion(raceArray, 0, results.num, results.race);
+    env->SetIntArrayRegion(genderArray, 0, results.num, results.gender);
+    env->SetIntArrayRegion(ageBracketArray, 0, results.num, results.ageBracket);
+
+    env->SetObjectField(attributeObj, raceField, raceArray);
+    env->SetObjectField(attributeObj, genderField, genderArray);
+    env->SetObjectField(attributeObj, ageBracketField, ageBracketArray);
+
+    env->DeleteLocalRef(raceArray);
+    env->DeleteLocalRef(genderArray);
+    env->DeleteLocalRef(ageBracketArray);
+
+    return attributeObj;
+}
+
+/**
+ * @brief Get the face emotion result.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param session The session object.
+ * @return The face emotion result object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_GetFaceEmotionResult)(JNIEnv *env, jobject thiz, jobject session) {
+    // Validate input parameters
+    if (!env || !session) {
+        INSPIRE_LOGE("Invalid input parameters");
+        return nullptr;
+    }
+
+    // Get session handle
+    jclass sessionClass = env->GetObjectClass(session);
+    jfieldID handleField = env->GetFieldID(sessionClass, "handle", "J");
+    jlong sessionHandle = env->GetLongField(session, handleField);
+    if (!sessionHandle) {
+        INSPIRE_LOGE("Invalid session handle");
+        return nullptr;
+    }
+
+    // Get face emotion results
+    HFFaceEmotionResult results = {};
+    HResult ret = HFGetFaceEmotionResult((HFSession)sessionHandle, &results);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to get face emotion result, error code: %d", ret);
+        return nullptr;
+    }
+
+    // Create Java FaceEmotionResult object
+    jclass emotionClass = env->FindClass("com/insightface/sdk/inspireface/base/FaceEmotionResult");
+    if (!emotionClass) {
+        INSPIRE_LOGE("Failed to find FaceEmotionResult class");
+        return nullptr;
+    }
+
+    jmethodID constructor = env->GetMethodID(emotionClass, "<init>", "()V");
+    jobject emotionObj = env->NewObject(emotionClass, constructor);
+    if (!emotionObj) {
+        INSPIRE_LOGE("Failed to create FaceEmotionResult object");
+        return nullptr;
+    }
+
+    // Set fields
+    jfieldID numField = env->GetFieldID(emotionClass, "num", "I");
+    jfieldID emotionField = env->GetFieldID(emotionClass, "emotion", "[I");
+
+    if (!numField || !emotionField) {
+        INSPIRE_LOGE("Failed to get field IDs");
+        return nullptr;
+    }
+
+    // Set num
+    env->SetIntField(emotionObj, numField, results.num);
+
+    // Set emotion array
+    jintArray emotionArray = env->NewIntArray(results.num);
+    if (!emotionArray) {
+        INSPIRE_LOGE("Failed to create emotion array");
+        return nullptr;
+    }
+
+    env->SetIntArrayRegion(emotionArray, 0, results.num, results.emotion);
+    env->SetObjectField(emotionObj, emotionField, emotionArray);
+
+    env->DeleteLocalRef(emotionArray);
+
+    return emotionObj;
+}
+
+/**
+ * @brief Query the InspireFace version.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @return The InspireFace version object.
+ */
+JNIEXPORT jobject INSPIRE_FACE_JNI(InspireFace_QueryInspireFaceVersion)(JNIEnv *env, jobject thiz) {
+    // Get version info
+    HFInspireFaceVersion versionInfo;
+    HFQueryInspireFaceVersion(&versionInfo);
+
+    // Create new InspireFaceVersion object
+    jclass versionClass = env->FindClass("com/insightface/sdk/inspireface/base/InspireFaceVersion");
+    if (!versionClass) {
+        INSPIRE_LOGE("Failed to find InspireFaceVersion class");
+        return nullptr;
+    }
+
+    jmethodID constructor = env->GetMethodID(versionClass, "<init>", "()V");
+    jobject version = env->NewObject(versionClass, constructor);
+    if (!version) {
+        INSPIRE_LOGE("Failed to create InspireFaceVersion object");
+        return nullptr;
+    }
+
+    // Get field IDs
+    jfieldID majorField = env->GetFieldID(versionClass, "major", "I");
+    jfieldID minorField = env->GetFieldID(versionClass, "minor", "I");
+    jfieldID patchField = env->GetFieldID(versionClass, "patch", "I");
+
+    if (!majorField || !minorField || !patchField) {
+        INSPIRE_LOGE("Failed to get InspireFaceVersion field IDs");
+        return nullptr;
+    }
+
+    // Set version fields
+    env->SetIntField(version, majorField, versionInfo.major);
+    env->SetIntField(version, minorField, versionInfo.minor);
+    env->SetIntField(version, patchField, versionInfo.patch);
+
+    return version;
+}
+
+/**
+ * @brief Set the log level.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ * @param level The log level.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_SetLogLevel)(JNIEnv *env, jobject thiz, jint level) {
+    HFSetLogLevel((HFLogLevel)level);
+}
+
+/**
+ * @brief Disable the log.
+ *
+ * @param env The JNI environment.
+ * @param thiz The Java object.
+ */
+JNIEXPORT void INSPIRE_FACE_JNI(InspireFace_LogDisable)(JNIEnv *env, jobject thiz) {
+    HFLogDisable();
+}
+
+}  // extern "C"
+
+#endif  // ANDROID
diff --git a/cpp-package/inspireface/cpp/inspireface/platform/jni/common/common.h b/cpp-package/inspireface/cpp/inspireface/platform/jni/common/common.h
new file mode 100644
index 000000000..7fb8e14cf
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/platform/jni/common/common.h
@@ -0,0 +1,37 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-11-26
+ */
+
+#ifndef INSPIRE_FACE_JNI_COMMON_H
+#define INSPIRE_FACE_JNI_COMMON_H
+
+#include <jni.h>
+#include <string>
+
+/**
+ * @brief Convert jstring to std::string
+ * @param env JNIEnv pointer
+ * @param jstr jstring object
+ * @return std::string
+ */
+inline std::string jstring2str(JNIEnv *env, jstring jstr) {
+    char *rtn = NULL;
+    jclass clsstring = env->FindClass("java/lang/String");
+    jstring strencode = env->NewStringUTF("GB2312");
+    jmethodID mid = env->GetMethodID(clsstring, "getBytes", "(Ljava/lang/String;)[B");
+    jbyteArray barr = (jbyteArray)env->CallObjectMethod(jstr, mid, strencode);
+    jsize alen = env->GetArrayLength(barr);
+    jbyte *ba = env->GetByteArrayElements(barr, JNI_FALSE);
+    if (alen > 0) {
+        rtn = (char *)malloc(alen + 1);
+        memcpy(rtn, ba, alen);
+        rtn[alen] = 0;
+    }
+    env->ReleaseByteArrayElements(barr, ba, 0);
+    std::string stemp(rtn);
+    free(rtn);
+    return stemp;
+}
+
+#endif  // INSPIRE_FACE_JNI_COMMON_H
diff --git a/cpp-package/inspireface/cpp/inspireface/platform/jni/java/java.h b/cpp-package/inspireface/cpp/inspireface/platform/jni/java/java.h
new file mode 100644
index 000000000..843aea086
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/platform/jni/java/java.h
@@ -0,0 +1,9 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-11-26
+ */
+
+#ifndef INSPIRE_FACE_JNI_JAVA_H
+#define INSPIRE_FACE_JNI_JAVA_H
+
+#endif  // INSPIRE_FACE_JNI_JAVA_H
diff --git a/cpp-package/inspireface/cpp/inspireface/platform/platform.h b/cpp-package/inspireface/cpp/inspireface/platform/platform.h
new file mode 100644
index 000000000..66403bb00
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/platform/platform.h
@@ -0,0 +1,9 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-11-26
+ */
+
+#ifndef INSPIRE_FACE_PLATFORM_H
+#define INSPIRE_FACE_PLATFORM_H
+
+#endif  // INSPIRE_FACE_PLATFORM_H
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/dest_const.h b/cpp-package/inspireface/cpp/inspireface/recognition_module/dest_const.h
new file mode 100644
index 000000000..d8b554666
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/dest_const.h
@@ -0,0 +1,21 @@
+#ifndef INSPIREFACE_DEST_CONST_H
+#define INSPIREFACE_DEST_CONST_H
+
+#include <inspirecv/inspirecv.h>
+
+namespace inspire {
+
+// face feature dimension
+const int32_t FACE_FEATURE_DIM = 512;
+// face crop size
+const int32_t FACE_CROP_SIZE = 112;  
+// similarity transform destination points
+const std::vector<inspirecv::Point2f> SIMILARITY_TRANSFORM_DEST = {{38.2946, 51.6963},
+            {73.5318, 51.5014},
+            {56.0252, 71.7366},
+            {41.5493, 92.3655},
+            {70.7299, 92.2041}};
+
+} // namespace inspire
+
+#endif //INSPIREFACE_DEST_CONST_H
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.cpp
new file mode 100644
index 000000000..c560b8d4e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.cpp
@@ -0,0 +1,151 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "extract_adapt.h"
+#include <array>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+
+namespace {
+
+bool IsDebugRawFeatureEnabled() {
+    const char *env = std::getenv("ISF_DEBUG_RAW_FEATURE");
+    return env != nullptr && env[0] != '\0' && std::strcmp(env, "0") != 0;
+}
+
+float CalcSegmentNorm(const std::vector<float> &embedding, size_t begin, size_t end) {
+    float sum = 0.0f;
+    for (size_t i = begin; i < end; ++i) {
+        sum += embedding[i] * embedding[i];
+    }
+    return std::sqrt(sum);
+}
+
+float CalcSegmentCosine(const std::vector<float> &embedding, size_t lhs_begin, size_t rhs_begin, size_t length) {
+    float lhs_norm = 0.0f;
+    float rhs_norm = 0.0f;
+    float dot = 0.0f;
+    for (size_t i = 0; i < length; ++i) {
+        const float lhs = embedding[lhs_begin + i];
+        const float rhs = embedding[rhs_begin + i];
+        lhs_norm += lhs * lhs;
+        rhs_norm += rhs * rhs;
+        dot += lhs * rhs;
+    }
+    if (lhs_norm <= 0.0f || rhs_norm <= 0.0f) {
+        return 0.0f;
+    }
+    return dot / std::sqrt(lhs_norm * rhs_norm);
+}
+
+float CalcSegmentMaxAbsDiff(const std::vector<float> &embedding, size_t lhs_begin, size_t rhs_begin, size_t length) {
+    float max_abs_diff = 0.0f;
+    for (size_t i = 0; i < length; ++i) {
+        const float diff = std::fabs(embedding[lhs_begin + i] - embedding[rhs_begin + i]);
+        if (diff > max_abs_diff) {
+            max_abs_diff = diff;
+        }
+    }
+    return max_abs_diff;
+}
+
+void DumpRawFeature(const inspire::AnyTensorOutputs &outputs, const std::vector<float> &embedding, float norm) {
+    INSPIRE_LOGI("[ISF_DEBUG_RAW_FEATURE] outputs=%zu primary_name=%s primary_size=%zu raw_l2_norm=%.9f",
+                 outputs.size(),
+                 outputs.empty() ? "<none>" : outputs[0].first.c_str(),
+                 embedding.size(),
+                 norm);
+
+    for (const auto &output : outputs) {
+        INSPIRE_LOGI("[ISF_DEBUG_RAW_FEATURE] tensor name=%s size=%zu", output.first.c_str(), output.second.size());
+    }
+
+    if (embedding.size() == 512) {
+        constexpr size_t kChunkSize = 128;
+        std::array<float, 4> chunk_norms{};
+        for (size_t i = 0; i < chunk_norms.size(); ++i) {
+            const size_t begin = i * kChunkSize;
+            chunk_norms[i] = CalcSegmentNorm(embedding, begin, begin + kChunkSize);
+        }
+        INSPIRE_LOGI("[ISF_DEBUG_RAW_FEATURE] chunk_norms[128x4]=[%.9f, %.9f, %.9f, %.9f]",
+                     chunk_norms[0], chunk_norms[1], chunk_norms[2], chunk_norms[3]);
+        INSPIRE_LOGI("[ISF_DEBUG_RAW_FEATURE] chunk_cosine c01=%.9f c02=%.9f c03=%.9f c12=%.9f c13=%.9f c23=%.9f",
+                     CalcSegmentCosine(embedding, 0, 128, kChunkSize),
+                     CalcSegmentCosine(embedding, 0, 256, kChunkSize),
+                     CalcSegmentCosine(embedding, 0, 384, kChunkSize),
+                     CalcSegmentCosine(embedding, 128, 256, kChunkSize),
+                     CalcSegmentCosine(embedding, 128, 384, kChunkSize),
+                     CalcSegmentCosine(embedding, 256, 384, kChunkSize));
+        INSPIRE_LOGI("[ISF_DEBUG_RAW_FEATURE] chunk_max_abs_diff d01=%.9f d02=%.9f d03=%.9f",
+                     CalcSegmentMaxAbsDiff(embedding, 0, 128, kChunkSize),
+                     CalcSegmentMaxAbsDiff(embedding, 0, 256, kChunkSize),
+                     CalcSegmentMaxAbsDiff(embedding, 0, 384, kChunkSize));
+    }
+
+    const char *dump_path = std::getenv("ISF_DEBUG_RAW_FEATURE_PATH");
+    if (dump_path == nullptr || dump_path[0] == '\0') {
+        return;
+    }
+
+    FILE *fp = std::fopen(dump_path, "a");
+    if (fp == nullptr) {
+        INSPIRE_LOGW("[ISF_DEBUG_RAW_FEATURE] failed to open dump path: %s", dump_path);
+        return;
+    }
+
+    std::fprintf(fp, "primary_name=%s size=%zu raw_l2_norm=%.9f\n",
+                 outputs.empty() ? "<none>" : outputs[0].first.c_str(),
+                 embedding.size(),
+                 norm);
+    std::fprintf(fp, "values=");
+    for (size_t i = 0; i < embedding.size(); ++i) {
+        std::fprintf(fp, i == 0 ? "%.9f" : ",%.9f", embedding[i]);
+    }
+    std::fprintf(fp, "\n");
+    std::fclose(fp);
+}
+
+}  // namespace
+
+namespace inspire {
+
+Embedded ExtractAdapt::GetFaceFeature(const inspirecv::Image &bgr_affine) {
+    AnyTensorOutputs outputs;
+    Forward(bgr_affine, outputs);
+
+    return outputs[0].second;
+}
+
+Embedded ExtractAdapt::operator()(const inspirecv::Image &bgr_affine, float &norm, bool normalize) {
+    AnyTensorOutputs outputs;
+    Forward(bgr_affine, outputs);
+
+    auto &embedded = outputs[0].second;
+    float mse = 0.0f;
+    for (const auto &one : embedded) {
+        mse += one * one;
+    }
+    mse = sqrt(mse);
+    norm = mse;
+
+    if (IsDebugRawFeatureEnabled()) {
+        DumpRawFeature(outputs, embedded, norm);
+    }
+
+    if (normalize) {
+        for (float &one : embedded) {
+            one /= mse;
+        }
+    }
+
+    return embedded;
+}
+
+ExtractAdapt::ExtractAdapt() : AnyNetAdapter("ExtractAdapt") {}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.h b/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.h
new file mode 100644
index 000000000..9e283dbcc
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/extract/extract_adapt.h
@@ -0,0 +1,46 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_EXTRACT_ADAPT_H
+#define INSPIREFACE_EXTRACT_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class Extract
+ * @brief Class for extracting features from faces, inheriting from AnyNet.
+ *
+ * This class specializes in processing face images to extract embedded facial features,
+ * which can be used for further analysis like face recognition or verification.
+ */
+class INSPIRE_API ExtractAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for the Extract class.
+     */
+    ExtractAdapt();
+
+    /**
+     * @brief Operator to process an affine-transformed face image and return the extracted features.
+     * @param bgr_affine Affine-transformed face image in BGR format.
+     * @param norm The L2 norm of a vector.
+     * @param normalize Whether the obtained features are normalized.
+     * @return Embedded Vector of extracted features.
+     */
+    Embedded operator()(const inspirecv::Image& bgr_affine, float& norm, bool normalize = true);
+
+    /**
+     * @brief Gets the facial features from an affine-transformed face image.
+     * @param bgr_affine Affine-transformed face image in BGR format.
+     * @return Embedded Vector of extracted facial features.
+     */
+    Embedded GetFaceFeature(const inspirecv::Image& bgr_affine);
+};
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_EXTRACT_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.cpp b/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.cpp
new file mode 100644
index 000000000..7e8413884
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.cpp
@@ -0,0 +1,112 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_feature_extraction_module.h"
+#include "feature_hub/simd.h"
+#include "track_module/landmark/face_landmark_adapt.h"
+#include "herror.h"
+#include "dest_const.h"
+
+namespace inspire {
+
+FeatureExtractionModule::FeatureExtractionModule(InspireArchive &archive, bool enable_recognition) : m_status_code_(SARC_SUCCESS) {
+    if (enable_recognition) {
+        InspireModel model;
+        m_status_code_ = archive.LoadModel("feature", model);
+        if (m_status_code_ != SARC_SUCCESS) {
+            INSPIRE_LOGE("Load rec model error.");
+        }
+        m_status_code_ = InitExtractInteraction(model);
+        if (m_status_code_ != 0) {
+            INSPIRE_LOGE("FaceRecognition error.");
+        }
+    }
+    m_landmark_param_ = archive.GetLandmarkParam();
+}
+
+int32_t FeatureExtractionModule::InitExtractInteraction(InspireModel &model) {
+    try {
+        auto input_size = model.Config().get<std::vector<int>>("input_size");
+        m_extract_ = std::make_shared<ExtractAdapt>();
+        auto ret = m_extract_->LoadData(model, model.modelType);
+        if (ret != InferenceWrapper::WrapperOk) {
+            return HERR_ARCHIVE_LOAD_FAILURE;
+        }
+        return HSUCCEED;
+
+    } catch (const std::runtime_error &e) {
+        INSPIRE_LOGE("%s", e.what());
+        return HERR_SESS_FACE_REC_OPTION_ERROR;
+    }
+}
+
+int32_t FeatureExtractionModule::QueryStatus() const {
+    return m_status_code_;
+}
+
+int32_t FeatureExtractionModule::FaceExtract(inspirecv::FrameProcess &processor, const FaceTrackWrap &face, Embedded &embedded, float &norm,
+                                             bool normalize) {
+    if (m_extract_ == nullptr) {
+        return HERR_SESS_REC_EXTRACT_FAILURE;
+    }
+
+    std::vector<inspirecv::Point2f> pointsFive;
+    for (const auto &p : face.keyPoints) {
+        pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+    }
+    auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+    auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+    //    cv::imshow("w", crop);
+    //    cv::waitKey(0);
+    embedded = (*m_extract_)(crop, norm, normalize);
+
+    return 0;
+}
+
+int32_t FeatureExtractionModule::FaceExtractWithAlignmentImage(inspirecv::FrameProcess &processor, Embedded &embedded, float &norm,
+                                             bool normalize) {
+    if (m_extract_ == nullptr) {
+        return HERR_SESS_REC_EXTRACT_FAILURE;
+    }
+    auto crop = processor.ExecuteImageScaleProcessing(1.0f, false);
+    embedded = (*m_extract_)(crop, norm, normalize);
+
+    return 0;
+}
+
+int32_t FeatureExtractionModule::FaceExtractWithAlignmentImage(const inspirecv::Image& wrapped, Embedded &embedded, float &norm,
+                                             bool normalize) {
+    if (m_extract_ == nullptr) {
+        return HERR_SESS_REC_EXTRACT_FAILURE;
+    }
+    embedded = (*m_extract_)(wrapped, norm, normalize);
+
+    return 0;
+}
+
+int32_t FeatureExtractionModule::FaceExtract(inspirecv::FrameProcess &processor, const FaceObjectInternal &face, Embedded &embedded, float &norm,
+                                             bool normalize) {
+    if (m_extract_ == nullptr) {
+        return HERR_SESS_REC_EXTRACT_FAILURE;
+    }
+
+    auto lmk = face.landmark_;
+    std::vector<inspirecv::Point2f> lmk_5 = {lmk[m_landmark_param_->semantic_index.left_eye_center], lmk[m_landmark_param_->semantic_index.right_eye_center],
+                                             lmk[m_landmark_param_->semantic_index.nose_corner], lmk[m_landmark_param_->semantic_index.mouth_left_corner],
+                                             lmk[m_landmark_param_->semantic_index.mouth_right_corner]};
+
+    auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, lmk_5);
+    auto crop = processor.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+
+    embedded = (*m_extract_)(crop, norm, normalize);
+
+    return 0;
+}
+
+const std::shared_ptr<ExtractAdapt> &FeatureExtractionModule::getMExtract() const {
+    return m_extract_;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.h b/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.h
new file mode 100644
index 000000000..af64a9d48
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/face_feature_extraction_module.h
@@ -0,0 +1,98 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_FEATURE_EXTRACTION_MODULE_H
+#define INSPIRE_FACE_FEATURE_EXTRACTION_MODULE_H
+#include <mutex>
+#include "extract/extract_adapt.h"
+#include "common/face_info/face_object_internal.h"
+#include "face_wrapper.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "frame_process.h"
+
+namespace inspire {
+
+class INSPIRE_API FeatureExtractionModule {
+public:
+    /**
+     * @brief Constructor for FaceRecognition class.
+     *
+     * @param archive Model active instance for model loading.
+     * @param enable_recognition Whether face recognition is enabled.
+     * @param core Type of matrix core to use for feature extraction.
+     * @param feature_block_num Number of feature blocks to use.
+     */
+    FeatureExtractionModule(InspireArchive &archive, bool enable_recognition);
+
+    /**
+     * @brief Example Query the module loading status
+     * @return Status code
+     * */
+    int32_t QueryStatus() const;
+
+    /**
+     * @brief Extracts a facial feature from an image and stores it in the provided 'embedded'.
+     *
+     * @param processor inspirecv::FrameProcess instance containing the image.
+     * @param face FaceObject representing the detected face.
+     * @param embedded Output parameter to store the extracted facial feature.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t FaceExtract(inspirecv::FrameProcess &processor, const FaceObjectInternal &face, Embedded &embedded, float &norm, bool normalize = false);
+
+    /**
+     * @brief Extracts a facial feature from an image and stores it in the provided 'embedded'.
+     *
+     * @param processor inspirecv::FrameProcess instance containing the image.
+     * @param face FaceTrackWrap representing the detected face.
+     * @param embedded Output parameter to store the extracted facial feature.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t FaceExtract(inspirecv::FrameProcess &processor, const FaceTrackWrap &face, Embedded &embedded, float &norm, bool normalize = true);
+
+    /**
+     * @brief Extracts a facial feature from an image and stores it in the provided 'embedded'.
+     *
+     * @param processor inspirecv::FrameProcess instance containing the image.
+     * @param embedded Output parameter to store the extracted facial feature.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t FaceExtractWithAlignmentImage(inspirecv::FrameProcess &processor, Embedded &embedding, float &norm, bool normalize = true);
+
+    /**
+     * @brief Extracts a facial feature from an image and stores it in the provided 'embedding'.
+     *
+     * @param wrapped inspirecv::Image instance containing the image.
+     * @param embedding Output parameter to store the extracted facial feature.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t FaceExtractWithAlignmentImage(const inspirecv::Image& wrapped, Embedded &embedding, float &norm, bool normalize = true);
+
+    /**
+     * @brief Gets the Extract instance associated with this FaceRecognition.
+     *
+     * @return const std::shared_ptr<Extract>& Pointer to the Extract instance.
+     */
+    const std::shared_ptr<ExtractAdapt> &getMExtract() const;
+
+private:
+    /**
+     * @brief Initializes the interaction with the Extract model.
+     *
+     * @param model Pointer to the loaded model.
+     * @return int32_t Status code indicating success (0) or failure.
+     */
+    int32_t InitExtractInteraction(InspireModel &model);
+
+private:
+    std::shared_ptr<ExtractAdapt> m_extract_;  ///< Pointer to the Extract instance.
+    std::shared_ptr<LandmarkParam> m_landmark_param_;  ///< Pointer to the LandmarkParam instance.
+
+    int32_t m_status_code_;  ///< Status code
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_FEATURE_EXTRACTION_MODULE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/recognition_module/similarity_converter.cpp b/cpp-package/inspireface/cpp/inspireface/recognition_module/similarity_converter.cpp
new file mode 100644
index 000000000..d1f972ee5
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/recognition_module/similarity_converter.cpp
@@ -0,0 +1,9 @@
+#include <mutex>
+#include "similarity_converter.h"
+
+namespace inspire {
+
+SimilarityConverter* SimilarityConverter::instance = nullptr;
+std::mutex SimilarityConverter::instanceMutex;
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/runtime_module/launch.cpp b/cpp-package/inspireface/cpp/inspireface/runtime_module/launch.cpp
new file mode 100644
index 000000000..2ad404656
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/runtime_module/launch.cpp
@@ -0,0 +1,334 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "launch.h"
+#include "log.h"
+#include "herror.h"
+#include "isf_check.h"
+
+// Include the implementation details here, hidden from public header
+#include "middleware/model_archive/inspire_archive.h"
+#if defined(ISF_ENABLE_RGA)
+#include "image_process/nexus_processor/rga/dma_alloc.h"
+#endif
+#include <mutex>
+#include "middleware/inference_wrapper/inference_wrapper.h"
+#include "middleware/system.h"
+#if defined(ISF_ENABLE_TENSORRT)
+#include "cuda_toolkit.h"
+#endif
+#include "meta.h"
+
+#define APPLE_EXTENSION_SUFFIX ".bundle"
+
+namespace inspire {
+
+// Implementation class definition
+class Launch::Impl {
+public:
+    Impl()
+    : m_load_(false),
+      m_archive_(nullptr),
+      m_cuda_device_id_(0),
+      m_global_coreml_inference_mode_(InferenceWrapper::COREML_ANE),
+      m_image_processing_backend_(IMAGE_PROCESSING_CPU) {
+#if defined(ISF_ENABLE_RGA)
+        m_image_processing_backend_ = IMAGE_PROCESSING_RGA;
+        INSPIRE_LOGW("Default image processing backend is RGA.");
+#if defined(ISF_RKNPU_RV1106)
+        m_rockchip_dma_heap_path_ = RV1106_CMA_HEAP_PATH;
+#else
+        m_rockchip_dma_heap_path_ = DMA_HEAP_DMA32_UNCACHE_PATCH;
+#endif
+        INSPIRE_LOGW("Rockchip dma heap configured path: %s", m_rockchip_dma_heap_path_.c_str());
+#endif
+        m_face_detect_pixel_list_ = {160, 320, 640};
+        m_face_detect_model_list_ = {"face_detect_160", "face_detect_320", "face_detect_640"};
+        INSPIRE_LOGI("%s", GetSDKInfo().GetFullVersionInfo().c_str());
+    }
+    // Face Detection pixel size
+    std::vector<int32_t> m_face_detect_pixel_list_;
+
+    // Face Detection model list
+    std::vector<std::string> m_face_detect_model_list_;
+
+    // Static members
+    static std::mutex mutex_;
+    static std::shared_ptr<Launch> instance_;
+
+    // Data members
+    std::string m_rockchip_dma_heap_path_;
+    std::string m_extension_path_;
+    std::unique_ptr<InspireArchive> m_archive_;
+    bool m_load_;
+    int32_t m_cuda_device_id_;
+    InferenceWrapper::SpecialBackend m_global_coreml_inference_mode_;
+    Launch::ImageProcessingBackend m_image_processing_backend_;
+    int32_t m_image_process_aligned_width_{4};
+};
+
+// Initialize static members
+std::mutex Launch::Impl::mutex_;
+std::shared_ptr<Launch> Launch::Impl::instance_ = nullptr;
+
+// Constructor implementation
+Launch::Launch() : pImpl(std::make_unique<Impl>()) {}
+
+// Destructor implementation
+Launch::~Launch() = default;
+
+std::shared_ptr<Launch> Launch::GetInstance() {
+    std::lock_guard<std::mutex> lock(Impl::mutex_);
+    if (!Impl::instance_) {
+        Impl::instance_ = std::shared_ptr<Launch>(new Launch());
+    }
+    return Impl::instance_;
+}
+
+InspireArchive& Launch::getMArchive() {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (!pImpl->m_archive_) {
+        throw std::runtime_error("Archive not initialized");
+    }
+    return *(pImpl->m_archive_);
+}
+
+int32_t Launch::Load(const std::string& path) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+#if defined(ISF_ENABLE_TENSORRT)
+    int32_t support_cuda;
+    auto ret = CheckCudaUsability(&support_cuda);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("An error occurred while checking CUDA device support. Please ensure that your environment supports CUDA!");
+        return ret;
+    }
+    if (!support_cuda) {
+        INSPIRE_LOGE("Your environment does not support CUDA! Please ensure that your environment supports CUDA!");
+        return HERR_DEVICE_CUDA_NOT_SUPPORT;
+    }
+#endif
+    INSPIREFACE_CHECK_MSG(os::IsExists(path), "The package path does not exist because the launch failed.");
+#if defined(ISF_ENABLE_APPLE_EXTENSION)
+    BuildAppleExtensionPath(path);
+#endif
+    if (!pImpl->m_load_) {
+        try {
+            pImpl->m_archive_ = std::make_unique<InspireArchive>();
+            pImpl->m_archive_->ReLoad(path);
+
+            // Update face detect pixel list and model list
+            pImpl->m_face_detect_pixel_list_ = pImpl->m_archive_->GetFaceDetectPixelList();
+            pImpl->m_face_detect_model_list_ = pImpl->m_archive_->GetFaceDetectModelList();
+
+            if (pImpl->m_archive_->QueryStatus() == SARC_SUCCESS) {
+                pImpl->m_load_ = true;
+                INSPIRE_LOGI("Successfully loaded resources");
+                return HSUCCEED;
+            } else {
+                pImpl->m_archive_.reset();
+                INSPIRE_LOGE("Failed to load resources");
+                return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+            }
+        } catch (const std::exception& e) {
+            pImpl->m_archive_.reset();
+            INSPIRE_LOGE("Exception during resource loading: %s", e.what());
+            return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+        }
+    } else {
+        INSPIRE_LOGW("There is no need to call launch more than once, as subsequent calls will not affect the initialization.");
+        return HSUCCEED;
+    }
+}
+
+int32_t Launch::Reload(const std::string& path) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    INSPIREFACE_CHECK_MSG(os::IsExists(path), "The package path does not exist because the launch failed.");
+#if defined(ISF_ENABLE_APPLE_EXTENSION)
+    BuildAppleExtensionPath(path);
+#endif
+    try {
+        // Clean up existing archive if it exists
+        if (pImpl->m_archive_) {
+            pImpl->m_archive_.reset();
+            pImpl->m_load_ = false;
+        }
+
+        // Create and load new archive
+        pImpl->m_archive_ = std::make_unique<InspireArchive>();
+        pImpl->m_archive_->ReLoad(path);
+
+        if (pImpl->m_archive_->QueryStatus() == SARC_SUCCESS) {
+            pImpl->m_load_ = true;
+            INSPIRE_LOGI("Successfully reloaded resources");
+            return HSUCCEED;
+        } else {
+            pImpl->m_archive_.reset();
+            INSPIRE_LOGE("Failed to reload resources");
+            return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+        }
+    } catch (const std::exception& e) {
+        pImpl->m_archive_.reset();
+        INSPIRE_LOGE("Exception during resource reloading: %s", e.what());
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+}
+
+bool Launch::isMLoad() const {
+    return pImpl->m_load_;
+}
+
+void Launch::Unload() {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (pImpl->m_load_) {
+        pImpl->m_archive_.reset();
+        pImpl->m_load_ = false;
+        INSPIRE_LOGI("All resources have been successfully unloaded and system is reset.");
+    } else {
+        INSPIRE_LOGW("Unload called but system was not loaded.");
+    }
+}
+
+void Launch::SetRockchipDmaHeapPath(const std::string& path) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    pImpl->m_rockchip_dma_heap_path_ = path;
+}
+
+std::string Launch::GetRockchipDmaHeapPath() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_rockchip_dma_heap_path_;
+}
+
+void Launch::ConfigurationExtensionPath(const std::string& path) {
+#if defined(ISF_ENABLE_APPLE_EXTENSION)
+    INSPIREFACE_CHECK_MSG(os::IsDir(path), "The apple extension path is not a directory, please check.");
+#endif
+    INSPIREFACE_CHECK_MSG(os::IsExists(path), "The extension path is not exists, please check.");
+    pImpl->m_extension_path_ = path;
+}
+
+std::string Launch::GetExtensionPath() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_extension_path_;
+}
+
+void Launch::SetGlobalCoreMLInferenceMode(NNInferenceBackend mode) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (mode == NN_INFERENCE_CPU) {
+        pImpl->m_global_coreml_inference_mode_ = InferenceWrapper::COREML_CPU;
+    } else if (mode == NN_INFERENCE_COREML_GPU) {
+        pImpl->m_global_coreml_inference_mode_ = InferenceWrapper::COREML_GPU;
+    } else if (mode == NN_INFERENCE_COREML_ANE) {
+        pImpl->m_global_coreml_inference_mode_ = InferenceWrapper::COREML_ANE;
+    } else {
+        INSPIRE_LOGE("Invalid CoreML inference mode");
+    }
+    if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_CPU) {
+        INSPIRE_LOGW("Global CoreML Compute Units set to CPU Only.");
+    } else if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_GPU) {
+        INSPIRE_LOGW("Global CoreML Compute Units set to CPU and GPU.");
+    } else if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_ANE) {
+        INSPIRE_LOGW("Global CoreML Compute Units set to Auto Switch (ANE, GPU, CPU).");
+    }
+}
+
+Launch::NNInferenceBackend Launch::GetGlobalCoreMLInferenceMode() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_CPU) {
+        return NN_INFERENCE_CPU;
+    } else if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_GPU) {
+        return NN_INFERENCE_COREML_GPU;
+    } else if (pImpl->m_global_coreml_inference_mode_ == InferenceWrapper::COREML_ANE) {
+        return NN_INFERENCE_COREML_ANE;
+    } else {
+        INSPIRE_LOGE("Invalid CoreML inference mode");
+        return NN_INFERENCE_CPU;
+    }
+}
+
+void Launch::BuildAppleExtensionPath(const std::string& resource_path) {
+    std::string basename = os::Basename(resource_path);
+    pImpl->m_extension_path_ = os::PathJoin(os::Dirname(resource_path), basename + APPLE_EXTENSION_SUFFIX);
+    INSPIREFACE_CHECK_MSG(os::IsExists(pImpl->m_extension_path_), "The apple extension path is not exists, please check.");
+    INSPIREFACE_CHECK_MSG(os::IsDir(pImpl->m_extension_path_), "The apple extension path is not a directory, please check.");
+}
+
+void Launch::SetCudaDeviceId(int32_t device_id) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    pImpl->m_cuda_device_id_ = device_id;
+}
+
+int32_t Launch::GetCudaDeviceId() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_cuda_device_id_;
+}
+
+void Launch::SetFaceDetectPixelList(const std::vector<int32_t>& pixel_list) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    pImpl->m_face_detect_pixel_list_ = pixel_list;
+}
+
+std::vector<int32_t> Launch::GetFaceDetectPixelList() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_face_detect_pixel_list_;
+}
+
+void Launch::SetFaceDetectModelList(const std::vector<std::string>& model_list) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    pImpl->m_face_detect_model_list_ = model_list;
+}
+
+std::vector<std::string> Launch::GetFaceDetectModelList() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_face_detect_model_list_;
+}
+
+void Launch::SwitchLandmarkEngine(LandmarkEngine engine) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (pImpl->m_archive_->QueryStatus() != SARC_SUCCESS) {
+        INSPIRE_LOGE("The InspireFace is not initialized, please call launch first.");
+        return;
+    }
+    auto landmark_param = pImpl->m_archive_->GetLandmarkParam();
+    bool ret = false;
+    if (engine == LANDMARK_HYPLMV2_0_25) {
+        ret = landmark_param->ReLoad("landmark");
+    } else if (engine == LANDMARK_HYPLMV2_0_50) {
+        ret = landmark_param->ReLoad("landmark_0_50");
+    } else if (engine == LANDMARK_INSIGHTFACE_2D106_TRACK) {
+        ret = landmark_param->ReLoad("landmark_insightface_2d106");
+    }
+    INSPIREFACE_CHECK_MSG(ret, "Failed to switch landmark engine");
+}
+
+void Launch::SwitchImageProcessingBackend(ImageProcessingBackend backend) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    if (backend == IMAGE_PROCESSING_RGA) {
+#if defined(ISF_ENABLE_RGA)
+        pImpl->m_image_processing_backend_ = backend;
+#else
+        INSPIRE_LOGE("RKRGA is not enabled, please check the build configuration.");
+#endif  // ISF_ENABLE_RGA
+        return;
+    } else {
+        pImpl->m_image_processing_backend_ = backend;
+    }
+}
+
+Launch::ImageProcessingBackend Launch::GetImageProcessingBackend() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_image_processing_backend_;
+}
+
+void Launch::SetImageProcessAlignedWidth(int32_t width) {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    pImpl->m_image_process_aligned_width_ = width;
+}
+
+int32_t Launch::GetImageProcessAlignedWidth() const {
+    std::lock_guard<std::mutex> lock(pImpl->mutex_);
+    return pImpl->m_image_process_aligned_width_;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.cpp b/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.cpp
new file mode 100644
index 000000000..6ad58c264
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.cpp
@@ -0,0 +1,4 @@
+#include "resource_manage.h"
+
+std::unique_ptr<inspire::ResourceManager> inspire::ResourceManager::instance;
+std::mutex inspire::ResourceManager::mutex;
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.h b/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.h
new file mode 100644
index 000000000..0897ebefe
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/runtime_module/resource_manage.h
@@ -0,0 +1,235 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_RESOURCE_MANAGE_H
+#define INSPIREFACE_RESOURCE_MANAGE_H
+#include <iostream>
+#include <mutex>
+#include <unordered_map>
+#include <memory>
+#include <iomanip>  // For std::setw and std::left
+#include <vector>
+#include "log.h"
+#ifndef INSPIRE_API
+#define INSPIRE_API
+#endif
+
+#define RESOURCE_MANAGE inspire::ResourceManager::getInstance()
+
+namespace inspire {
+
+/**
+ * @brief ResourceManager is a singleton class that manages the creation and release of sessions and image streams.
+ * It uses hash tables to store session and image stream handles, and provides methods to create, release, and query these resources.
+ * The ResourceManager class is designed to be used in a multi-threaded environment, and it uses a mutex to synchronize access to its data structures.
+ */
+class INSPIRE_API ResourceManager {
+private:
+    // Private static instance pointer
+    static std::unique_ptr<ResourceManager> instance;
+    static std::mutex mutex;
+
+    // Use hash tables to store session and image stream handles
+    std::unordered_map<long, bool> sessionMap;
+    std::unordered_map<long, bool> streamMap;
+    std::unordered_map<long, bool> imageBitmapMap;
+    std::unordered_map<long, bool> faceFeatureMap;
+
+    // The private constructor guarantees singletons
+    ResourceManager() {}
+
+public:
+    // Remove copy constructors and assignment operators
+    ResourceManager(const ResourceManager&) = delete;
+    ResourceManager& operator=(const ResourceManager&) = delete;
+
+    // Method of obtaining singleton instance
+    static ResourceManager* getInstance() {
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!instance) {
+            instance.reset(new ResourceManager());
+        }
+        return instance.get();
+    }
+
+    // Method of obtaining singleton instance
+    void createSession(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        sessionMap[handle] = false;  // false indicates that it is not released
+    }
+
+    // Release session
+    bool releaseSession(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        auto it = sessionMap.find(handle);
+        if (it != sessionMap.end() && !it->second) {
+            it->second = true;  // Mark as released
+            return true;
+        }
+        return false;  // Release failed, possibly because the handle could not be found or was
+                       // released
+    }
+
+    // Create and record image streams
+    void createStream(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        streamMap[handle] = false;  // false indicates that it is not released
+    }
+
+    // Release image stream
+    bool releaseStream(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        auto it = streamMap.find(handle);
+        if (it != streamMap.end() && !it->second) {
+            it->second = true;  // Mark as released
+            return true;
+        }
+        return false;  // Release failed, possibly because the handle could not be found or was
+                       // released
+    }
+
+    // Create and record image bitmaps
+    void createImageBitmap(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        imageBitmapMap[handle] = false;  // false indicates that it is not released
+    }
+
+    // Release image bitmap
+    bool releaseImageBitmap(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        auto it = imageBitmapMap.find(handle);
+        if (it != imageBitmapMap.end() && !it->second) {
+            it->second = true;  // Mark as released
+            return true;
+        }
+        return false;  // Release failed, possibly because the handle could not be found or was released
+    }
+
+    // Create and record face features
+    void createFaceFeature(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        faceFeatureMap[handle] = false;  // false indicates that it is not released
+    }
+
+    // Release face feature
+    bool releaseFaceFeature(long handle) {
+        std::lock_guard<std::mutex> lock(mutex);
+        auto it = faceFeatureMap.find(handle);
+        if (it != faceFeatureMap.end() && !it->second) {
+            it->second = true;  // Mark as released
+            return true;
+        }
+        return false;  // Release failed, possibly because the handle could not be found or was released
+    }
+
+    // Gets a list of unreleased session handles
+    std::vector<long> getUnreleasedSessions() {
+        std::lock_guard<std::mutex> lock(mutex);
+        std::vector<long> unreleasedSessions;
+        for (const auto& entry : sessionMap) {
+            if (!entry.second) {
+                unreleasedSessions.push_back(entry.first);
+            }
+        }
+        return unreleasedSessions;
+    }
+
+    // Gets a list of unreleased image stream handles
+    std::vector<long> getUnreleasedStreams() {
+        std::lock_guard<std::mutex> lock(mutex);
+        std::vector<long> unreleasedStreams;
+        for (const auto& entry : streamMap) {
+            if (!entry.second) {
+                unreleasedStreams.push_back(entry.first);
+            }
+        }
+        return unreleasedStreams;
+    }
+
+    // Gets a list of unreleased image bitmap handles
+    std::vector<long> getUnreleasedImageBitmaps() {
+        std::lock_guard<std::mutex> lock(mutex);
+        std::vector<long> unreleasedImageBitmaps;
+        for (const auto& entry : imageBitmapMap) {
+            if (!entry.second) {
+                unreleasedImageBitmaps.push_back(entry.first);
+            }
+        }
+        return unreleasedImageBitmaps;
+    }
+
+    // Gets a list of unreleased face feature handles
+    std::vector<long> getUnreleasedFaceFeatures() {
+        std::lock_guard<std::mutex> lock(mutex);
+        std::vector<long> unreleasedFaceFeatures;
+        for (const auto& entry : faceFeatureMap) {
+            if (!entry.second) {
+                unreleasedFaceFeatures.push_back(entry.first);
+            }
+        }
+        return unreleasedFaceFeatures;
+    }
+
+    // Method to print resource management statistics
+    void printResourceStatistics() {
+        std::lock_guard<std::mutex> lock(mutex);
+        INSPIRE_LOGI("================================================================");
+        INSPIRE_LOGI("%-15s%-15s%-15s%-15s", "Resource Name", "Total Created", "Total Released", "Not Released");
+        INSPIRE_LOGI("----------------------------------------------------------------");
+
+        // Print session statistics
+        int totalSessionsCreated = sessionMap.size();
+        int totalSessionsReleased = 0;
+        int sessionsNotReleased = 0;
+        for (const auto& entry : sessionMap) {
+            if (entry.second)
+                ++totalSessionsReleased;
+            if (!entry.second)
+                ++sessionsNotReleased;
+        }
+        INSPIRE_LOGI("%-15s%-15d%-15d%-15d", "Session", totalSessionsCreated, totalSessionsReleased, sessionsNotReleased);
+
+        // Print stream statistics
+        int totalStreamsCreated = streamMap.size();
+        int totalStreamsReleased = 0;
+        int streamsNotReleased = 0;
+        for (const auto& entry : streamMap) {
+            if (entry.second)
+                ++totalStreamsReleased;
+            if (!entry.second)
+                ++streamsNotReleased;
+        }
+        INSPIRE_LOGI("%-15s%-15d%-15d%-15d", "Stream", totalStreamsCreated, totalStreamsReleased, streamsNotReleased);
+
+        // Print bitmap statistics
+        int totalBitmapsCreated = imageBitmapMap.size();
+        int totalBitmapsReleased = 0;
+        int bitmapsNotReleased = 0;
+        for (const auto& entry : imageBitmapMap) {
+            if (entry.second)
+                ++totalBitmapsReleased;
+            if (!entry.second)
+                ++bitmapsNotReleased;
+        }
+        INSPIRE_LOGI("%-15s%-15d%-15d%-15d", "Bitmap", totalBitmapsCreated, totalBitmapsReleased, bitmapsNotReleased);
+
+        // Print face feature statistics
+        int totalFeaturesCreated = faceFeatureMap.size();
+        int totalFeaturesReleased = 0;
+        int featuresNotReleased = 0;
+        for (const auto& entry : faceFeatureMap) {
+            if (entry.second)
+                ++totalFeaturesReleased;
+            if (!entry.second)
+                ++featuresNotReleased;
+        }
+        INSPIRE_LOGI("%-15s%-15d%-15d%-15d", "FaceFeature", totalFeaturesCreated, totalFeaturesReleased, featuresNotReleased);
+        INSPIRE_LOGI("================================================================");
+    }
+};
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_RESOURCE_MANAGE_H
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/session.cpp b/cpp-package/inspireface/cpp/inspireface/session.cpp
new file mode 100644
index 000000000..ba1ab4430
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/session.cpp
@@ -0,0 +1,316 @@
+#include <memory>
+#include "session.h"
+#include "engine/face_session.h"
+#include "recognition_module/dest_const.h"
+
+namespace inspire {
+
+class Session::Impl {
+public:
+    Impl() : m_face_session_(std::make_unique<FaceSession>()) {}
+
+    int32_t Configure(DetectModuleMode detect_mode, int32_t max_detect_face, CustomPipelineParameter param, int32_t detect_level_px,
+                      int32_t track_by_detect_mode_fps) {
+        return m_face_session_->Configuration(detect_mode, max_detect_face, param, detect_level_px, track_by_detect_mode_fps);
+    }
+
+    void ClearTrackingFace() {
+        m_face_session_->ClearTrackingFace();
+    }
+
+    ~Impl() = default;
+
+    void SetTrackPreviewSize(int32_t preview_size) {
+        m_face_session_->SetTrackPreviewSize(preview_size);
+    }
+
+    void SetFilterMinimumFacePixelSize(int32_t min_face_pixel_size) {
+        m_face_session_->SetTrackFaceMinimumSize(min_face_pixel_size);
+    }
+
+    void SetFaceDetectThreshold(float threshold) {
+        m_face_session_->SetFaceDetectThreshold(threshold);
+    }
+
+    void SetTrackModeSmoothRatio(int32_t smooth_ratio) {
+        m_face_session_->SetTrackModeSmoothRatio(smooth_ratio);
+    }
+
+    void SetTrackModeNumSmoothCacheFrame(int32_t num_smooth_cache_frame) {
+        m_face_session_->SetTrackModeNumSmoothCacheFrame(num_smooth_cache_frame);
+    }
+
+    void SetTrackModeDetectInterval(int32_t detect_interval) {
+        m_face_session_->SetTrackModeDetectInterval(detect_interval);
+    }
+
+    void SetTrackLostRecoveryMode(bool value) {
+        m_face_session_->SetTrackLostRecoveryMode(value);
+    }
+
+    void SetLightTrackConfidenceThreshold(float value) {
+        m_face_session_->SetLightTrackConfidenceThreshold(value);
+    }
+
+    int32_t FaceDetectAndTrack(inspirecv::FrameProcess& process, std::vector<FaceTrackWrap>& results) {
+        int32_t ret = m_face_session_->FaceDetectAndTrack(process);
+        if (ret < 0) {
+            return ret;
+        }
+        results.clear();
+        const auto& face_data = m_face_session_->GetDetectCache();
+        for (const auto& data : face_data) {
+            FaceTrackWrap hyper_face_data;
+            RunDeserializeHyperFaceData(data, hyper_face_data);
+            results.emplace_back(hyper_face_data);
+        }
+
+        return ret;
+    }
+
+    inspirecv::Rect2i GetFaceBoundingBox(const FaceTrackWrap& face_data) {
+        return inspirecv::Rect2i{face_data.rect.x, face_data.rect.y, face_data.rect.width, face_data.rect.height};
+    }
+
+    std::vector<inspirecv::Point2f> GetFaceDenseLandmark(const FaceTrackWrap& face_data) {
+        std::vector<inspirecv::Point2f> points;
+        for (const auto& p : face_data.densityLandmark) {
+            points.emplace_back(inspirecv::Point2f(p.x, p.y));
+        }
+        return points;
+    }
+
+    std::vector<inspirecv::Point2f> GetFaceFiveKeyPoints(const FaceTrackWrap& face_data) {
+        std::vector<inspirecv::Point2f> points;
+        for (const auto& p : face_data.keyPoints) {
+            points.emplace_back(inspirecv::Point2f(p.x, p.y));
+        }
+        return points;
+    }
+
+    int32_t FaceFeatureExtract(inspirecv::FrameProcess& process, FaceTrackWrap& data, FaceEmbedding& embedding, bool normalize) {
+        int32_t ret = m_face_session_->FaceFeatureExtract(process, data, normalize);
+        if (ret < 0) {
+            return ret;
+        }
+        embedding.isNormal = normalize;
+        embedding.embedding = m_face_session_->GetFaceFeatureCache();
+        embedding.norm = m_face_session_->GetFaceFeatureNormCache();
+
+        return ret;
+    }
+
+    int32_t FaceFeatureExtractWithAlignmentImage(inspirecv::FrameProcess& process, FaceEmbedding& embedding, bool normalize) {
+        int32_t ret = m_face_session_->FaceFeatureExtractWithAlignmentImage(process, embedding.embedding, embedding.norm, normalize);
+        if (ret < 0) {
+            return ret;
+        }
+        embedding.isNormal = normalize;
+        embedding.norm = embedding.norm;
+        return ret;
+    }
+
+    int32_t FaceFeatureExtractWithAlignmentImage(const inspirecv::Image& wrapped, FaceEmbedding& embedding, bool normalize) {
+        int32_t ret = m_face_session_->FaceFeatureExtractWithAlignmentImage(wrapped, embedding, embedding.norm, normalize);
+        if (ret < 0) {
+            return ret;
+        }
+        embedding.isNormal = normalize;
+        embedding.norm = embedding.norm;
+        return ret;
+    }
+
+    void GetFaceAlignmentImage(inspirecv::FrameProcess& process, FaceTrackWrap& data, inspirecv::Image& wrapped) {
+        std::vector<inspirecv::Point2f> pointsFive;
+        for (const auto& p : data.keyPoints) {
+            pointsFive.push_back(inspirecv::Point2f(p.x, p.y));
+        }
+        auto trans = inspirecv::SimilarityTransformEstimateUmeyama(SIMILARITY_TRANSFORM_DEST, pointsFive);
+        wrapped = process.ExecuteImageAffineProcessing(trans, FACE_CROP_SIZE, FACE_CROP_SIZE);
+    }
+
+    int32_t MultipleFacePipelineProcess(inspirecv::FrameProcess& process, const CustomPipelineParameter& param,
+                                        const std::vector<FaceTrackWrap>& face_data_list) {
+        int32_t ret = m_face_session_->FacesProcess(process, face_data_list, param);
+        return ret;
+    }
+
+    std::vector<float> GetRGBLivenessConfidence() {
+        return m_face_session_->GetDetConfidenceCache();
+    }
+
+    std::vector<float> GetFaceMaskConfidence() {
+        return m_face_session_->GetMaskResultsCache();
+    }
+
+    std::vector<float> GetFaceQualityConfidence() {
+        return m_face_session_->GetFaceQualityScoresResultsCache();
+    }
+
+    std::vector<FaceInteractionState> GetFaceInteractionState() {
+        auto left_eyes_confidence = m_face_session_->GetFaceInteractionLeftEyeStatusCache();
+        auto right_eyes_confidence = m_face_session_->GetFaceInteractionRightEyeStatusCache();
+        std::vector<FaceInteractionState> face_interaction_state;
+        for (size_t i = 0; i < left_eyes_confidence.size(); ++i) {
+            face_interaction_state.emplace_back(FaceInteractionState{left_eyes_confidence[i], right_eyes_confidence[i]});
+        }
+        return face_interaction_state;
+    }
+
+    std::vector<FaceInteractionAction> GetFaceInteractionAction() {
+        auto num = m_face_session_->GetFaceNormalAactionsResultCache().size();
+        std::vector<FaceInteractionAction> face_interaction_action;
+        face_interaction_action.resize(num);
+        for (size_t i = 0; i < num; ++i) {
+            face_interaction_action[i].normal = m_face_session_->GetFaceNormalAactionsResultCache()[i];
+            face_interaction_action[i].shake = m_face_session_->GetFaceShakeAactionsResultCache()[i];
+            face_interaction_action[i].jawOpen = m_face_session_->GetFaceJawOpenAactionsResultCache()[i];
+            face_interaction_action[i].headRaise = m_face_session_->GetFaceRaiseHeadAactionsResultCache()[i];
+            face_interaction_action[i].blink = m_face_session_->GetFaceBlinkAactionsResultCache()[i];
+        }
+        return face_interaction_action;
+    }
+
+    std::vector<FaceAttributeResult> GetFaceAttributeResult() {
+        auto num = m_face_session_->GetFaceNormalAactionsResultCache().size();
+        std::vector<FaceAttributeResult> face_attribute_result;
+        face_attribute_result.resize(num);
+        for (size_t i = 0; i < num; ++i) {
+            face_attribute_result[i].race = m_face_session_->GetFaceRaceResultsCache()[i];
+            face_attribute_result[i].gender = m_face_session_->GetFaceGenderResultsCache()[i];
+            face_attribute_result[i].ageBracket = m_face_session_->GetFaceAgeBracketResultsCache()[i];
+        }
+        return face_attribute_result;
+    }
+
+    std::vector<FaceEmotionResult> GetFaceEmotionResult() {
+        auto num = m_face_session_->GetFaceEmotionResultsCache().size();
+        std::vector<FaceEmotionResult> face_emotion_result;
+        face_emotion_result.resize(num);
+        for (size_t i = 0; i < num; ++i) {
+            face_emotion_result[i].emotion = m_face_session_->GetFaceEmotionResultsCache()[i];
+        }
+        return face_emotion_result;
+    }
+
+
+    std::unique_ptr<FaceSession> m_face_session_;
+};
+
+Session::Session() : pImpl(std::make_unique<Impl>()) {}
+
+Session::~Session() = default;
+
+Session::Session(Session&&) noexcept = default;
+
+Session& Session::operator=(Session&&) noexcept = default;
+
+Session Session::Create(DetectModuleMode detect_mode, int32_t max_detect_face, const CustomPipelineParameter& param, int32_t detect_level_px,
+                        int32_t track_by_detect_mode_fps) {
+    Session session;
+    session.pImpl->Configure(detect_mode, max_detect_face, param, detect_level_px, track_by_detect_mode_fps);
+    return session;
+}
+
+void Session::ClearTrackingFace() {
+    pImpl->ClearTrackingFace();
+}
+
+void Session::SetTrackLostRecoveryMode(bool value) {
+    pImpl->SetTrackLostRecoveryMode(value);
+}
+
+void Session::SetLightTrackConfidenceThreshold(float value) {
+    pImpl->SetLightTrackConfidenceThreshold(value);
+}
+
+void Session::SetTrackPreviewSize(int32_t preview_size) {
+    pImpl->SetTrackPreviewSize(preview_size);
+}
+
+void Session::SetFilterMinimumFacePixelSize(int32_t min_face_pixel_size) {
+    pImpl->SetFilterMinimumFacePixelSize(min_face_pixel_size);
+}
+
+void Session::SetFaceDetectThreshold(float threshold) {
+    pImpl->SetFaceDetectThreshold(threshold);
+}
+
+void Session::SetTrackModeSmoothRatio(int32_t smooth_ratio) {
+    pImpl->SetTrackModeSmoothRatio(smooth_ratio);
+}
+
+void Session::SetTrackModeNumSmoothCacheFrame(int32_t num_smooth_cache_frame) {
+    pImpl->SetTrackModeNumSmoothCacheFrame(num_smooth_cache_frame);
+}
+
+void Session::SetTrackModeDetectInterval(int32_t detect_interval) {
+    pImpl->SetTrackModeDetectInterval(detect_interval);
+}
+
+int32_t Session::FaceDetectAndTrack(inspirecv::FrameProcess& process, std::vector<FaceTrackWrap>& results) {
+    return pImpl->FaceDetectAndTrack(process, results);
+}
+
+inspirecv::Rect2i Session::GetFaceBoundingBox(const FaceTrackWrap& face_data) {
+    return pImpl->GetFaceBoundingBox(face_data);
+}
+
+std::vector<inspirecv::Point2f> Session::GetFaceDenseLandmark(const FaceTrackWrap& face_data) {
+    return pImpl->GetFaceDenseLandmark(face_data);
+}
+
+std::vector<inspirecv::Point2f> Session::GetFaceFiveKeyPoints(const FaceTrackWrap& face_data) {
+    return pImpl->GetFaceFiveKeyPoints(face_data);
+}
+
+int32_t Session::FaceFeatureExtract(inspirecv::FrameProcess& process, FaceTrackWrap& data, FaceEmbedding& embedding, bool normalize) {
+    return pImpl->FaceFeatureExtract(process, data, embedding, normalize);
+}
+
+int32_t Session::FaceFeatureExtractWithAlignmentImage(inspirecv::FrameProcess& process, FaceEmbedding& embedding, bool normalize) {
+    return pImpl->FaceFeatureExtractWithAlignmentImage(process, embedding, normalize);
+}
+
+int32_t Session::FaceFeatureExtractWithAlignmentImage(const inspirecv::Image& wrapped, FaceEmbedding& embedding, bool normalize) {
+    return pImpl->FaceFeatureExtractWithAlignmentImage(wrapped, embedding, normalize);
+}
+
+void Session::GetFaceAlignmentImage(inspirecv::FrameProcess& process, FaceTrackWrap& data, inspirecv::Image& wrapped) {
+    pImpl->GetFaceAlignmentImage(process, data, wrapped);
+}
+
+int32_t Session::MultipleFacePipelineProcess(inspirecv::FrameProcess& process, const CustomPipelineParameter& param,
+                                             const std::vector<FaceTrackWrap>& face_data_list) {
+    return pImpl->MultipleFacePipelineProcess(process, param, face_data_list);
+}
+
+std::vector<float> Session::GetRGBLivenessConfidence() {
+    return pImpl->GetRGBLivenessConfidence();
+}
+
+std::vector<float> Session::GetFaceMaskConfidence() {
+    return pImpl->GetFaceMaskConfidence();
+}
+
+std::vector<float> Session::GetFaceQualityConfidence() {
+    return pImpl->GetFaceQualityConfidence();
+}
+
+std::vector<FaceInteractionState> Session::GetFaceInteractionState() {
+    return pImpl->GetFaceInteractionState();
+}
+
+std::vector<FaceInteractionAction> Session::GetFaceInteractionAction() {
+    return pImpl->GetFaceInteractionAction();
+}
+
+std::vector<FaceAttributeResult> Session::GetFaceAttributeResult() {
+    return pImpl->GetFaceAttributeResult();
+}
+
+std::vector<FaceEmotionResult> Session::GetFaceEmotionResult() {
+    return pImpl->GetFaceEmotionResult();
+}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/all.h b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/all.h
new file mode 100644
index 000000000..18a846f82
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/all.h
@@ -0,0 +1,12 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIRE_FACE_FD_ALL_H
+#define INSPIRE_FACE_FD_ALL_H
+
+#include "rnet_adapt.h"
+#include "face_detect_adapt.h"
+
+#endif  // INSPIRE_FACE_FD_ALL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.cpp
new file mode 100644
index 000000000..793bf1bba
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.cpp
@@ -0,0 +1,166 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_detect_adapt.h"
+#include "cost_time.h"
+#include "spend_timer.h"
+
+namespace inspire {
+
+FaceDetectAdapt::FaceDetectAdapt(int input_size, float nms_threshold, float cls_threshold)
+: AnyNetAdapter("FaceDetectAdapt"), m_nms_threshold_(nms_threshold), m_cls_threshold_(cls_threshold), m_input_size_(input_size) {}
+
+FaceLocList FaceDetectAdapt::operator()(const inspirecv::Image &bgr) {
+    inspire::SpendTimer time_image_process("Image process");
+    time_image_process.Start();
+    int ori_w = bgr.Width();
+    int ori_h = bgr.Height();
+    float scale;
+
+    inspirecv::Image pad;
+
+    uint8_t *resized_data = nullptr;
+    if (ori_w == m_input_size_ && ori_h == m_input_size_) {
+        scale = 1.0f;
+        resized_data = (uint8_t *)bgr.Data();
+    } else {
+        m_processor_->ResizeAndPadding(bgr.Data(), bgr.Width(), bgr.Height(), bgr.Channels(), m_input_size_, m_input_size_, &resized_data, scale);
+    }
+
+    pad = inspirecv::Image::Create(m_input_size_, m_input_size_, bgr.Channels(), resized_data, false);
+
+    time_image_process.Stop();
+    // std::cout << time_image_process << std::endl;
+    // pad.Write("pad.jpg");
+    //    LOGD("Prepare");
+    AnyTensorOutputs outputs;
+    inspire::SpendTimer time_forward("Forward");
+    time_forward.Start();
+    Forward(pad, outputs);
+    time_forward.Stop();
+    // std::cout << time_forward << std::endl;
+    //    LOGD("Forward");
+
+    inspire::SpendTimer time_decode("Decode");
+    time_decode.Start();
+    std::vector<FaceLoc> results;
+    std::vector<int> strides = {8, 16, 32};
+    for (int i = 0; i < strides.size(); ++i) {
+        const std::vector<float> &tensor_cls = outputs[i].second;
+        const std::vector<float> &tensor_box = outputs[i + 3].second;
+        const std::vector<float> &tensor_lmk = outputs[i + 6].second;
+        _decode(tensor_cls, tensor_box, tensor_lmk, strides[i], results);
+    }
+    time_decode.Stop();
+    // std::cout << time_decode << std::endl;
+
+    _nms(results, m_nms_threshold_);
+    std::sort(results.begin(), results.end(), [](FaceLoc a, FaceLoc b) { return (a.y2 - a.y1) * (a.x2 - a.x1) > (b.y2 - b.y1) * (b.x2 - b.x1); });
+    for (auto &face : results) {
+        face.x1 = face.x1 / scale;
+        face.y1 = face.y1 / scale;
+        face.x2 = face.x2 / scale;
+        face.y2 = face.y2 / scale;
+        for (int i = 0; i < 5; ++i) {
+            face.lmk[i * 2 + 0] = face.lmk[i * 2 + 0] / scale;
+            face.lmk[i * 2 + 1] = face.lmk[i * 2 + 1] / scale;
+        }
+    }
+    m_processor_->MarkDone();
+
+    return results;
+}
+
+void FaceDetectAdapt::_nms(std::vector<FaceLoc> &input_faces, float nms_threshold) {
+    std::sort(input_faces.begin(), input_faces.end(), [](FaceLoc a, FaceLoc b) { return a.score > b.score; });
+    std::vector<float> area(input_faces.size());
+    for (int i = 0; i < int(input_faces.size()); ++i) {
+        area[i] = (input_faces.at(i).x2 - input_faces.at(i).x1 + 1) * (input_faces.at(i).y2 - input_faces.at(i).y1 + 1);
+    }
+    for (int i = 0; i < int(input_faces.size()); ++i) {
+        for (int j = i + 1; j < int(input_faces.size());) {
+            float xx1 = (std::max)(input_faces[i].x1, input_faces[j].x1);
+            float yy1 = (std::max)(input_faces[i].y1, input_faces[j].y1);
+            float xx2 = (std::min)(input_faces[i].x2, input_faces[j].x2);
+            float yy2 = (std::min)(input_faces[i].y2, input_faces[j].y2);
+            float w = (std::max)(float(0), xx2 - xx1 + 1);
+            float h = (std::max)(float(0), yy2 - yy1 + 1);
+            float inter = w * h;
+            float ovr = inter / (area[i] + area[j] - inter);
+            if (ovr >= nms_threshold) {
+                input_faces.erase(input_faces.begin() + j);
+                area.erase(area.begin() + j);
+            } else {
+                j++;
+            }
+        }
+    }
+}
+
+void FaceDetectAdapt::_generate_anchors(int stride, int input_size, int num_anchors, std::vector<float> &anchors) {
+    int height = ceil(input_size / stride);
+    int width = ceil(input_size / stride);
+    for (int j = 0; j < height; ++j) {
+        for (int i = 0; i < width; ++i) {
+            for (int k = 0; k < num_anchors; ++k) {
+                anchors.push_back(i * stride);
+                anchors.push_back(j * stride);
+            }
+        }
+    }
+}
+
+void FaceDetectAdapt::_decode(const std::vector<float> &cls_pred, const std::vector<float> &box_pred, const std::vector<float> &lmk_pred, int stride,
+                              std::vector<FaceLoc> &results) {
+    std::vector<float> anchors_center;
+    _generate_anchors(stride, m_input_size_, 2, anchors_center);
+
+    for (int i = 0; i < anchors_center.size() / 2; ++i) {
+        if (cls_pred[i] > m_cls_threshold_) {
+            FaceLoc faceInfo;
+            float cx = anchors_center[i * 2 + 0];
+            float cy = anchors_center[i * 2 + 1];
+            float x1 = cx - box_pred[i * 4 + 0] * stride;
+            float y1 = cy - box_pred[i * 4 + 1] * stride;
+            float x2 = cx + box_pred[i * 4 + 2] * stride;
+            float y2 = cy + box_pred[i * 4 + 3] * stride;
+            faceInfo.x1 = x1;
+            faceInfo.y1 = y1;
+            faceInfo.x2 = x2;
+            faceInfo.y2 = y2;
+            faceInfo.score = cls_pred[i];
+            //            if (use_kps_) {
+            for (int j = 0; j < 5; ++j) {
+                float px = cx + lmk_pred[i * 10 + j * 2 + 0] * stride;
+                float py = cy + lmk_pred[i * 10 + j * 2 + 1] * stride;
+                faceInfo.lmk[j * 2 + 0] = px;
+                faceInfo.lmk[j * 2 + 1] = py;
+            }
+            //            }
+            results.push_back(faceInfo);
+        }
+        std::sort(results.begin(), results.end(), SortBoxSizeAdapt);
+    }
+}
+
+void FaceDetectAdapt::SetNmsThreshold(float mNmsThreshold) {
+    m_nms_threshold_ = mNmsThreshold;
+}
+
+void FaceDetectAdapt::SetClsThreshold(float mClsThreshold) {
+    m_cls_threshold_ = mClsThreshold;
+}
+
+bool SortBoxSizeAdapt(const FaceLoc &a, const FaceLoc &b) {
+    int sq_a = (a.y2 - a.y1) * (a.x2 - a.x1);
+    int sq_b = (b.y2 - b.y1) * (b.x2 - b.x1);
+    return sq_a > sq_b;
+}
+
+int FaceDetectAdapt::GetInputSize() const {
+    return m_input_size_;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.h b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.h
new file mode 100644
index 000000000..6c8d91cf7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/face_detect_adapt.h
@@ -0,0 +1,94 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_FACE_DETECT_ADAPT_H
+#define INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_FACE_DETECT_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+#include "image_process/nexus_processor/image_processor.h"
+
+namespace inspire {
+
+/**
+ * @class FaceDetect
+ * @brief Class for face detection, inheriting from AnyNet.
+ *
+ * This class provides functionalities to detect faces in images using neural network models.
+ */
+class INSPIRE_API FaceDetectAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for the FaceDetect class.
+     * @param input_size The size of the input image for the neural network.
+     * @param nms_threshold The threshold for non-maximum suppression.
+     * @param cls_threshold The threshold for classification score.
+     */
+    explicit FaceDetectAdapt(int input_size = 160, float nms_threshold = 0.4f, float cls_threshold = 0.5f);
+
+    /**
+     * @brief Detects faces in a given image.
+     * @param bgr The input image in BGR format.
+     * @return FaceLocList List of detected faces with location and landmarks.
+     */
+    FaceLocList operator()(const inspirecv::Image &bgr);
+
+    /** @brief Set non-maximum suppression threshold */
+    void SetNmsThreshold(float mNmsThreshold);
+
+    /** @brief Set face classification threshold */
+    void SetClsThreshold(float mClsThreshold);
+
+    /**
+     * @brief Get the input size
+     * @return int The input size
+     */
+    int GetInputSize() const;
+
+private:
+    /**
+     * @brief Applies non-maximum suppression to reduce overlapping detected faces.
+     * @param input_faces List of detected faces to be filtered.
+     * @param nms_threshold The threshold for non-maximum suppression.
+     */
+    static void _nms(FaceLocList &input_faces, float nms_threshold);
+
+    /**
+     * @brief Generates detection anchors based on stride.
+     * @param stride The stride of the detection.
+     * @param input_size The size of the input image.
+     * @param num_anchors The number of anchors.
+     * @param anchors The generated anchors.
+     */
+    void _generate_anchors(int stride, int input_size, int num_anchors, std::vector<float> &anchors);
+
+    /**
+     * @brief Decodes network outputs to face locations.
+     * @param cls_pred Classification predictions.
+     * @param box_pred Bounding box predictions.
+     * @param lmk_pred Landmark predictions.
+     * @param stride The stride of the detection.
+     * @param results Decoded face locations.
+     */
+    void _decode(const std::vector<float> &cls_pred, const std::vector<float> &box_pred, const std::vector<float> &lmk_pred, int stride,
+                 std::vector<FaceLoc> &results);
+
+private:
+    float m_nms_threshold_;  ///< Threshold for non-maximum suppression.
+    float m_cls_threshold_;  ///< Threshold for classification score.
+    int m_input_size_;       ///< Input size for the neural network model.
+};
+
+/**
+ * @brief Sorts FaceLoc objects in descending order of area.
+ * @param a The first FaceLoc object.
+ * @param b The second FaceLoc object.
+ * @return bool True if 'a' is larger than 'b'.
+ */
+bool SortBoxSizeAdapt(const FaceLoc &a, const FaceLoc &b);
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_FACE_DETECT_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.cpp
new file mode 100644
index 000000000..cc1bbe811
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.cpp
@@ -0,0 +1,38 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "rnet_adapt.h"
+
+namespace inspire {
+
+float RNetAdapt::operator()(const inspirecv::Image &bgr_affine) {
+    // auto resized = bgr_affine.Resize(24, 24);
+    uint8_t *resized_data = nullptr;
+    float scale;
+    auto ret = m_processor_->Resize(bgr_affine.Data(), bgr_affine.Width(), bgr_affine.Height(), bgr_affine.Channels(), &resized_data, 24, 24);
+    inspirecv::Image resized;
+    if (ret == -1) {
+        // Some RK devices seem unable to resize to 24x24, fallback to CPU processing
+        resized = bgr_affine.Resize(24, 24);
+    } else {
+        // RGA resize success
+        resized = inspirecv::Image::Create(24, 24, bgr_affine.Channels(), resized_data, false);
+    }
+
+    AnyTensorOutputs outputs;
+    Forward(resized, outputs);
+    m_processor_->MarkDone();
+#ifdef INFERENCE_WRAPPER_ENABLE_RKNN2
+    auto sm = Softmax(outputs[0].second);
+    return sm[1];
+#else
+    return outputs[0].second[1];
+    // std::cout << outputs[0].second[0] << ", " << outputs[0].second[1] << std ::endl;
+#endif
+}
+
+RNetAdapt::RNetAdapt() : AnyNetAdapter("RNetAdapt") {}
+
+}  //  namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.h b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.h
new file mode 100644
index 000000000..efe3cad98
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_detect/rnet_adapt.h
@@ -0,0 +1,37 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_RNET_ADAPT_H
+#define INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_RNET_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class RNet
+ * @brief Class representing the RNet (Refinement Network), inheriting from AnyNet.
+ *
+ * This class is used for refining face detection results, typically as part of a cascaded
+ * network system for facial recognition or detection tasks.
+ */
+class INSPIRE_API RNetAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Constructor for the RNet class.
+     */
+    RNetAdapt();
+
+    /**
+     * @brief Operator to process an affine-transformed face image and return a score indicating the quality of the refinement.
+     * @param bgr_affine Affine-transformed face image in BGR format.
+     * @return float Score representing the quality or confidence of the refinement.
+     */
+    float operator()(const inspirecv::Image& bgr_affine);
+};
+
+}  //  namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_FACE_DETECT_RNET_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.cpp
new file mode 100644
index 000000000..4d487c2c6
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.cpp
@@ -0,0 +1,594 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_track_module.h"
+#include "log.h"
+#include <algorithm>
+#include <cstddef>
+#include "middleware/costman.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "middleware/utils.h"
+#include "herror.h"
+#include "middleware/costman.h"
+#include "cost_time.h"
+#include "spend_timer.h"
+#include "launch.h"
+
+namespace inspire {
+
+FaceTrackModule::FaceTrackModule(DetectModuleMode mode, int max_detected_faces, int detection_interval, int track_preview_size,
+                                 int dynamic_detection_input_level, int TbD_mode_fps, bool detect_mode_landmark)
+: m_mode_(mode),
+  max_detected_faces_(max_detected_faces),
+  detection_interval_(detection_interval),
+  track_preview_size_(track_preview_size),
+  m_dynamic_detection_input_level_(dynamic_detection_input_level),
+  m_detect_mode_landmark_(detect_mode_landmark) {
+    detection_index_ = -1;
+    tracking_idx_ = 0;
+    if (TbD_mode_fps < 0) {
+        TbD_mode_fps = 30;
+    }
+    if (mode == DETECT_MODE_LIGHT_TRACK) {
+        // In lightweight tracking mode, landmark detection is always required
+        m_detect_mode_landmark_ = true;
+    } else {
+        // This version uses lmk106 to replace five key points, so lmk must be forcibly enabled!
+        m_detect_mode_landmark_ = true;
+    }
+    if (m_mode_ == DETECT_MODE_TRACK_BY_DETECT) {
+        m_TbD_tracker_ = std::make_shared<BYTETracker>(TbD_mode_fps, 30);
+    }
+}
+
+void FaceTrackModule::SparseLandmarkPredict(const inspirecv::Image &raw_face_crop, std::vector<inspirecv::Point2f> &landmarks_output, float &score,
+                                            float size) {
+    COST_TIME_SIMPLE(SparseLandmarkPredict);
+    landmarks_output.resize(m_landmark_param_->num_of_landmark);
+    std::vector<float> lmk_out = (*m_landmark_predictor_)(raw_face_crop);
+    for (int i = 0; i < m_landmark_param_->num_of_landmark; ++i) {
+        float x = lmk_out[i * 2 + 0] * size;
+        float y = lmk_out[i * 2 + 1] * size;
+        landmarks_output[i] = inspirecv::Point<float>(x, y);
+    }
+}
+
+float FaceTrackModule::PredictTrackScore(const inspirecv::Image &raw_face_crop) {
+    return (*m_refine_net_)(raw_face_crop);
+}
+
+bool FaceTrackModule::TrackFace(inspirecv::FrameProcess &image, FaceObjectInternal &face) {
+    COST_TIME_SIMPLE(TrackFace);
+    // If the track lost recovery mode is enabled,  the lag information of the previous frame will not be used in the current frame
+    if (face.GetConfidence() < m_light_track_confidence_threshold_ && !m_track_lost_recovery_mode_) {
+        // If the face confidence level is below the threshold, disable tracking
+        face.DisableTracking();
+        return false;
+    }
+
+    inspirecv::TransformMatrix affine;
+    std::vector<inspirecv::Point2f> landmark_back;
+
+    // Increase track count
+    face.IncrementTrackingCount();
+
+    float score;
+    // If it is a detection state, calculate the affine transformation matrix
+    if (face.TrackingState() == ISF_DETECT) {
+        COST_TIME_SIMPLE(GetRectSquare);
+        inspirecv::Rect2i rect_square = face.GetRectSquare(0);
+
+        std::vector<inspirecv::Point2f> rect_pts = rect_square.As<float>().ToFourVertices();
+        inspirecv::TransformMatrix rotation_mode_affine = image.GetAffineMatrix();
+        std::vector<inspirecv::Point2f> camera_pts = ApplyTransformToPoints(rect_pts, rotation_mode_affine);
+        // camera_pts.erase(camera_pts.end() - 1);
+        std::vector<inspirecv::Point2f> dst_pts = {{0, 0},
+                                                   {(float)m_landmark_param_->input_size, 0},
+                                                   {(float)m_landmark_param_->input_size, (float)m_landmark_param_->input_size},
+                                                   {0, (float)m_landmark_param_->input_size}};
+        affine = inspirecv::SimilarityTransformEstimate(camera_pts, dst_pts);
+        face.setTransMatrix(affine);
+
+        if (!m_detect_mode_landmark_) {
+            /*If landmark is not extracted, the detection frame of the preview image needs to be changed
+            back to the coordinate system of the original image */
+            std::vector<inspirecv::Point2f> restore_rect_pts = inspirecv::ApplyTransformToPoints(rect_pts, rotation_mode_affine);
+            inspirecv::Rect2f restore_rect = inspirecv::MinBoundingRect(restore_rect_pts);
+            face.bbox_ = restore_rect.As<int>();
+        }
+    }
+
+    if (m_face_quality_ != nullptr) {
+        COST_TIME_SIMPLE(FaceQuality);
+        auto affine_extensive = face.getTransMatrix();
+        auto trans_e = ScaleAffineMatrixPreserveCenter(affine_extensive, m_crop_extensive_ratio_, m_landmark_param_->input_size);
+        auto pre_crop = image.ExecuteImageAffineProcessing(trans_e, m_landmark_param_->input_size, m_landmark_param_->input_size);
+        auto res = (*m_face_quality_)(pre_crop);
+        // pre_crop.Show("pre_crop");
+
+        auto affine_extensive_inv = affine_extensive.GetInverse();
+        std::vector<inspirecv::Point2f> lmk_extensive = ApplyTransformToPoints(res.lmk, affine_extensive_inv);
+        res.lmk = lmk_extensive;
+        face.high_result = res;
+    } else {
+        // If face pose and quality model is not initialized, set the default value
+        FacePoseQualityAdaptResult empty_result;
+        empty_result.lmk = std::vector<inspirecv::Point2f>(5, inspirecv::Point2f(0, 0));
+        empty_result.lmk_quality = std::vector<float>(5, 2.0f);
+        empty_result.pitch = 0.0f;
+        empty_result.yaw = 0.0f;
+        empty_result.roll = 0.0f;
+        face.high_result = empty_result;
+    }
+
+    if (m_detect_mode_landmark_) {
+        // If Landmark need to be extracted in detection mode,
+        // Landmark must be detected when fast tracing is enabled
+        affine = face.getTransMatrix();
+        inspirecv::TransformMatrix affine_inv = affine.GetInverse();
+        std::vector<inspirecv::Point2f> landmark_rawout;
+        std::vector<std::vector<inspirecv::Point2f>> multiscale_landmark_back;
+
+        auto track_crop = image.ExecuteImageAffineProcessing(affine, m_landmark_param_->input_size, m_landmark_param_->input_size);
+        score = PredictTrackScore(track_crop);
+        // track_crop.Show("track_crop");
+
+        // If the track lost recovery mode is enabled, 
+        // it will determine whether to discard the invalid face that has been tracked in the current frame
+        if (score < m_light_track_confidence_threshold_ && m_track_lost_recovery_mode_) {
+            face.DisableTracking();
+            return false;
+        }
+
+        for (int i = 0; i < m_multiscale_landmark_scales_.size(); i++) {
+            inspirecv::Image crop;
+            // Get the RGB image after affine transformation
+            auto affine_scale = ScaleAffineMatrixPreserveCenter(affine, m_multiscale_landmark_scales_[i], m_landmark_param_->input_size);
+            crop = image.ExecuteImageAffineProcessing(affine_scale, m_landmark_param_->input_size, m_landmark_param_->input_size);
+
+            std::vector<inspirecv::Point2f> lmk_predict;
+
+            // Predicted sparse key point
+            SparseLandmarkPredict(crop, lmk_predict, score, m_landmark_param_->input_size);
+
+            // Save the first scale landmark
+            if (i == 0) {
+                landmark_rawout = lmk_predict;
+            }
+
+            std::vector<inspirecv::Point2f> lmk_back;
+            // Convert key points back to the original coordinate system
+            lmk_back.resize(lmk_predict.size());
+            lmk_back = inspirecv::ApplyTransformToPoints(lmk_predict, affine_scale.GetInverse());
+
+            multiscale_landmark_back.push_back(lmk_back);
+        }
+
+        landmark_back = MultiFrameLandmarkMean(multiscale_landmark_back);
+
+        // Extract 5 key points
+        std::vector<inspirecv::Point2f> lmk_5 = {
+          landmark_rawout[m_landmark_param_->semantic_index.left_eye_center], landmark_rawout[m_landmark_param_->semantic_index.right_eye_center],
+          landmark_rawout[m_landmark_param_->semantic_index.nose_corner], landmark_rawout[m_landmark_param_->semantic_index.mouth_left_corner],
+          landmark_rawout[m_landmark_param_->semantic_index.mouth_right_corner]};
+        face.setAlignMeanSquareError(lmk_5);
+
+        int MODE = 1;
+
+        if (MODE > 0) {
+            if (face.TrackingState() == ISF_DETECT) {
+                face.ReadyTracking();
+            } else if (face.TrackingState() == ISF_READY || face.TrackingState() == ISF_TRACKING) {
+                COST_TIME_SIMPLE(LandmarkBack);
+                inspirecv::TransformMatrix trans_m;
+                // inspirecv::TransformMatrix tmp = face.getTransMatrix();
+                std::vector<inspirecv::Point2f> inside_points;
+                if (m_landmark_param_->input_size == 112) {
+                    inside_points = landmark_rawout;
+                } else {
+                    inside_points = LandmarkCropped(landmark_rawout);
+                }
+
+                auto &mean_shape_ = m_landmark_param_->mean_shape_points;
+
+                auto _affine = inspirecv::SimilarityTransformEstimate(inside_points, mean_shape_);
+                auto mid_inside_points = ApplyTransformToPoints(inside_points, _affine);
+                inside_points = FixPointsMeanshape(mid_inside_points, mean_shape_);
+
+                trans_m = inspirecv::SimilarityTransformEstimate(landmark_back, inside_points);
+                face.setTransMatrix(trans_m);
+                face.EnableTracking();
+            }
+        }
+        // Update face key points
+        face.SetLandmark(landmark_back, true, true, m_track_mode_smooth_ratio_, m_track_mode_num_smooth_cache_frame_,
+                         m_landmark_param_->num_of_landmark * 2);
+        // Get the smoothed landmark
+        auto &landmark_smooth = face.landmark_smooth_aux_.back();
+        // Update the face key points
+        face.high_result.lmk[0] = landmark_smooth[m_landmark_param_->semantic_index.left_eye_center];
+        face.high_result.lmk[1] = landmark_smooth[m_landmark_param_->semantic_index.right_eye_center];
+        face.high_result.lmk[2] = landmark_smooth[m_landmark_param_->semantic_index.nose_corner];
+        face.high_result.lmk[3] = landmark_smooth[m_landmark_param_->semantic_index.mouth_left_corner];
+        face.high_result.lmk[4] = landmark_smooth[m_landmark_param_->semantic_index.mouth_right_corner];
+    }
+
+    // If tracking status, update the confidence level
+    if (face.TrackingState() == ISF_TRACKING) {
+        face.SetConfidence(score);
+    }
+
+    return true;
+}
+
+void FaceTrackModule::UpdateStream(inspirecv::FrameProcess &image) {
+    inspire::SpendTimer total("UpdateStream");
+    total.Start();
+    COST_TIME_SIMPLE(FaceTrackUpdateStream);
+    detection_index_ += 1;
+    if (m_mode_ == DETECT_MODE_ALWAYS_DETECT || m_mode_ == DETECT_MODE_TRACK_BY_DETECT)
+        trackingFace.clear();
+
+    // Record whether the detection has been performed in this frame
+    bool detection_executed = false;
+    if (trackingFace.empty() || (detection_interval_ > 0 && detection_index_ % detection_interval_ == 0) || m_mode_ == DETECT_MODE_ALWAYS_DETECT ||
+        m_mode_ == DETECT_MODE_TRACK_BY_DETECT) {
+        image.SetPreviewSize(track_preview_size_);
+        inspirecv::Image image_detect = image.ExecutePreviewImageProcessing(true);
+        m_debug_preview_image_size_ = image_detect.Width();
+
+        nms();
+        for (auto const &face : trackingFace) {
+            inspirecv::Rect2i m_mask_rect = face.GetRect();
+            std::vector<inspirecv::Point2f> pts = m_mask_rect.As<float>().ToFourVertices();
+            inspirecv::TransformMatrix rotation_mode_affine = image.GetAffineMatrix();
+            auto rotation_mode_affine_inv = rotation_mode_affine.GetInverse();
+            std::vector<inspirecv::Point2f> affine_pts = inspirecv::ApplyTransformToPoints(pts, rotation_mode_affine_inv);
+            inspirecv::Rect2f mask_rect = inspirecv::MinBoundingRect(affine_pts);
+            BlackingTrackingRegion(image_detect, mask_rect);
+        }
+
+        Timer det_cost_time;
+        DetectFace(image_detect, image.GetPreviewScale());
+        detection_executed = true;
+    }
+
+    if (!candidate_faces_.empty()) {
+        for (int i = 0; i < candidate_faces_.size(); i++) {
+            trackingFace.push_back(candidate_faces_[i]);
+        }
+        candidate_faces_.clear();
+    }
+
+    // Record the number of faces before tracking
+    size_t faces_before_tracking = trackingFace.size();
+
+    for (std::vector<FaceObjectInternal>::iterator iter = trackingFace.begin(); iter != trackingFace.end();) {
+        if (!TrackFace(image, *iter)) {
+            iter = trackingFace.erase(iter);
+        } else {
+            iter++;
+        }
+    }
+
+    // In the track lost recovery mode, if all faces are triggered to be lost, detection will be executed immediately 
+    if (m_track_lost_recovery_mode_ && !detection_executed && faces_before_tracking > 0 && trackingFace.empty()) {
+        image.SetPreviewSize(track_preview_size_);
+        inspirecv::Image image_detect = image.ExecutePreviewImageProcessing(true);
+        m_debug_preview_image_size_ = image_detect.Width();
+        DetectFace(image_detect, image.GetPreviewScale());
+
+        // Reset the detection index to 0
+        detection_index_ = 0;
+
+        // Add the detected faces to the tracking face list
+        if (!candidate_faces_.empty()) {
+            for (int i = 0; i < candidate_faces_.size(); i++) {
+                trackingFace.push_back(candidate_faces_[i]);
+            }
+            candidate_faces_.clear();
+        }
+
+        // Track the faces  
+        for (std::vector<FaceObjectInternal>::iterator iter = trackingFace.begin(); iter != trackingFace.end();) {
+            if (!TrackFace(image, *iter)) {
+                iter = trackingFace.erase(iter);
+            } else {
+                iter++;
+            }
+        }
+    }
+
+    total.Stop();
+    // std::cout << total << std::endl;
+}
+
+void FaceTrackModule::nms(float th) {
+    std::sort(trackingFace.begin(), trackingFace.end(), [](FaceObjectInternal a, FaceObjectInternal b) { return a.confidence_ > b.confidence_; });
+    std::vector<float> area(trackingFace.size());
+    for (int i = 0; i < int(trackingFace.size()); ++i) {
+        area[i] = trackingFace.at(i).getBbox().Area();
+    }
+    for (int i = 0; i < int(trackingFace.size()); ++i) {
+        for (int j = i + 1; j < int(trackingFace.size());) {
+            float xx1 = (std::max)(trackingFace[i].getBbox().GetX(), trackingFace[j].getBbox().GetX());
+            float yy1 = (std::max)(trackingFace[i].getBbox().GetY(), trackingFace[j].getBbox().GetY());
+            float xx2 = (std::min)(trackingFace[i].getBbox().GetX() + trackingFace[i].getBbox().GetWidth(),
+                                   trackingFace[j].getBbox().GetX() + trackingFace[j].getBbox().GetWidth());
+            float yy2 = (std::min)(trackingFace[i].getBbox().GetY() + trackingFace[i].getBbox().GetHeight(),
+                                   trackingFace[j].getBbox().GetY() + trackingFace[j].getBbox().GetHeight());
+            float w = (std::max)(float(0), xx2 - xx1 + 1);
+            float h = (std::max)(float(0), yy2 - yy1 + 1);
+            float inter = w * h;
+            float ovr = inter / (area[i] + area[j] - inter);
+            if (ovr >= th) {
+                trackingFace.erase(trackingFace.begin() + j);
+                area.erase(area.begin() + j);
+            } else {
+                j++;
+            }
+        }
+    }
+}
+
+void FaceTrackModule::BlackingTrackingRegion(inspirecv::Image &image, inspirecv::Rect2f &rect_mask) {
+    COST_TIME_SIMPLE(BlackingTrackingRegion);
+    int height = image.Height();
+    int width = image.Width();
+    auto ext = rect_mask.Square(1.5f);
+    inspirecv::Rect2i safe_rect = ext.SafeRect(width, height).As<int>();
+    image.Fill(safe_rect, {0, 0, 0});
+}
+
+void FaceTrackModule::DetectFace(const inspirecv::Image &input, float scale) {
+    std::vector<FaceLoc> boxes = (*m_face_detector_)(input);
+
+    if (m_mode_ == DETECT_MODE_TRACK_BY_DETECT) {
+        std::vector<Object> objects;
+        auto num_of_effective = std::min(boxes.size(), (size_t)max_detected_faces_);
+        for (size_t i = 0; i < num_of_effective; i++) {
+            Object obj;
+            const auto box = boxes[i];
+            obj.rect = inspirecv::Rect<int>(box.x1, box.y1, box.x2 - box.x1, box.y2 - box.y1);
+            if (!isShortestSideGreaterThan<int>(obj.rect, filter_minimum_face_px_size, scale)) {
+                // Filter too small face detection box
+                continue;
+            }
+            obj.label = 0;  // assuming all detections are faces
+            obj.prob = box.score;
+            objects.push_back(obj);
+        }
+        std::vector<STrack> output_stracks = m_TbD_tracker_->update(objects);
+        for (const auto &st_track : output_stracks) {
+            inspirecv::Rect<int> rect = inspirecv::Rect<int>(st_track.tlwh[0], st_track.tlwh[1], st_track.tlwh[2], st_track.tlwh[3]);
+            FaceObjectInternal faceinfo(st_track.track_id, rect, FaceLandmarkAdapt::NUM_OF_LANDMARK + 10);
+            faceinfo.detect_bbox_ = rect.As<int>();
+            candidate_faces_.push_back(faceinfo);
+        }
+    } else {
+        std::vector<inspirecv::Rect2i> bbox;
+        bbox.resize(boxes.size());
+        for (int i = 0; i < boxes.size(); i++) {
+            bbox[i] = inspirecv::Rect<int>::Create(boxes[i].x1, boxes[i].y1, boxes[i].x2 - boxes[i].x1, boxes[i].y2 - boxes[i].y1);
+
+            if (!isShortestSideGreaterThan<int>(bbox[i], filter_minimum_face_px_size, scale)) {
+                // Filter too small face detection box
+                continue;
+            }
+            if (m_mode_ == DETECT_MODE_ALWAYS_DETECT) {
+                // Always detect mode without assigning an id
+                tracking_idx_ = -1;
+            } else {
+                tracking_idx_ = tracking_idx_ + 1;
+            }
+
+            FaceObjectInternal faceinfo(tracking_idx_, bbox[i], m_landmark_param_->num_of_landmark + 10);
+            faceinfo.detect_bbox_ = bbox[i];
+            faceinfo.SetConfidence(boxes[i].score);
+
+            // Control that the number of faces detected does not exceed the maximum limit
+            if (candidate_faces_.size() >= max_detected_faces_) {
+                continue;
+            }
+
+            candidate_faces_.push_back(faceinfo);
+        }
+    }
+}
+
+int FaceTrackModule::Configuration(inspire::InspireArchive &archive, const std::string &expansion_path, bool enable_face_pose_and_quality) {
+    // Initialize the detection model
+    m_landmark_param_ = archive.GetLandmarkParam();
+    m_expansion_path_ = std::move(expansion_path);
+    InspireModel detModel;
+    auto scheme = ChoiceMultiLevelDetectModel(m_dynamic_detection_input_level_, track_preview_size_);
+    auto ret = archive.LoadModel(scheme, detModel);
+    if (ret != SARC_SUCCESS) {
+        INSPIRE_LOGE("Load %s error: %d", scheme.c_str(), ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+    InitDetectModel(detModel);
+
+    // Initialize the landmark model
+    InspireModel lmkModel;
+    ret = archive.LoadModel(m_landmark_param_->landmark_engine_name, lmkModel);
+    if (ret != SARC_SUCCESS) {
+        INSPIRE_LOGE("Load %s error: %d", m_landmark_param_->landmark_engine_name.c_str(), ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+    InitLandmarkModel(lmkModel);
+
+    // Initialize the r-net model
+    InspireModel rnetModel;
+    ret = archive.LoadModel("refine_net", rnetModel);
+    if (ret != SARC_SUCCESS) {
+        INSPIRE_LOGE("Load %s error: %d", "refine_net", ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+    InitRNetModel(rnetModel);
+    if (enable_face_pose_and_quality) {
+        // Initialize the pose quality model
+        InspireModel pquModel;
+        ret = archive.LoadModel("pose_quality", pquModel);
+        if (ret != SARC_SUCCESS) {
+            INSPIRE_LOGE("Load %s error: %d", "pose_quality", ret);
+            return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+        }
+        InitFacePoseAndQualityModel(pquModel);
+    }
+    m_landmark_crop_ratio_ = m_landmark_param_->expansion_scale;
+    m_multiscale_landmark_scales_ = GenerateCropScales(m_landmark_crop_ratio_, m_multiscale_landmark_loop_num_);
+    return 0;
+}
+
+int FaceTrackModule::InitLandmarkModel(InspireModel &model) {
+    m_landmark_predictor_ =
+      std::make_shared<FaceLandmarkAdapt>(m_landmark_param_->input_size, m_landmark_param_->normalization_mode == "CenterScaling");
+    auto ret = m_landmark_predictor_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int FaceTrackModule::InitDetectModel(InspireModel &model) {
+    std::vector<int> input_size;
+    input_size = model.Config().get<std::vector<int>>("input_size");
+
+    m_face_detector_ = std::make_shared<FaceDetectAdapt>(input_size[0]);
+    auto ret = m_face_detector_->LoadData(model, model.modelType, false);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int FaceTrackModule::InitRNetModel(InspireModel &model) {
+    m_refine_net_ = std::make_shared<RNetAdapt>();
+    auto ret = m_refine_net_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+int FaceTrackModule::InitFacePoseAndQualityModel(InspireModel &model) {
+    m_face_quality_ = std::make_shared<FacePoseQualityAdapt>();
+    auto ret = m_face_quality_->LoadData(model, model.modelType);
+    if (ret != InferenceWrapper::WrapperOk) {
+        return HERR_ARCHIVE_LOAD_FAILURE;
+    }
+    return HSUCCEED;
+}
+
+void FaceTrackModule::SetDetectThreshold(float value) {
+    m_face_detector_->SetClsThreshold(value);
+}
+
+void FaceTrackModule::SetMinimumFacePxSize(float value) {
+    filter_minimum_face_px_size = value;
+}
+
+void FaceTrackModule::SetTrackPreviewSize(int preview_size) {
+    track_preview_size_ = preview_size;
+    if (track_preview_size_ == -1) {
+        track_preview_size_ = m_face_detector_->GetInputSize();
+    } else if (track_preview_size_ < 160) {
+        INSPIRE_LOGW("Track preview size %d is less than the minimum input size %d", track_preview_size_, 160);
+        track_preview_size_ = 160;
+    }
+}
+
+int32_t FaceTrackModule::GetTrackPreviewSize() const {
+    return track_preview_size_;
+}
+
+std::string FaceTrackModule::ChoiceMultiLevelDetectModel(const int32_t pixel_size, int32_t &final_size) {
+    const auto face_detect_pixel_list = Launch::GetInstance()->GetFaceDetectPixelList();
+    const auto face_detect_model_list = Launch::GetInstance()->GetFaceDetectModelList();
+    const int32_t num_sizes = face_detect_pixel_list.size();
+    if (pixel_size == -1) {
+        // Find index with value 320, use index 1 as fallback
+        int index = 1;
+        for (int i = 0; i < face_detect_pixel_list.size(); ++i) {
+            if (face_detect_pixel_list[i] == 320) {
+                index = i;
+                break;
+            }
+        }
+        final_size = face_detect_pixel_list[index];
+        return face_detect_model_list[index];
+    }
+
+    // Check for exact match
+    for (int i = 0; i < num_sizes; ++i) {
+        if (pixel_size == face_detect_pixel_list[i]) {
+            final_size = face_detect_pixel_list[i];
+            return face_detect_model_list[i];
+        }
+    }
+
+    // Find the closest match
+    int32_t closest_size = face_detect_pixel_list[0];
+    std::string closest_scheme = face_detect_model_list[0];
+    int32_t min_diff = std::abs(pixel_size - face_detect_pixel_list[0]);
+
+    for (int i = 1; i < num_sizes; ++i) {
+        int32_t diff = std::abs(pixel_size - face_detect_pixel_list[i]);
+        if (diff < min_diff) {
+            min_diff = diff;
+            closest_size = face_detect_pixel_list[i];
+            closest_scheme = face_detect_model_list[i];
+        }
+    }
+
+    INSPIRE_LOGW(
+      "Input pixel size %d is not supported. Choosing the closest scheme: %s closest_scheme for "
+      "size %d.",
+      pixel_size, closest_scheme.c_str(), closest_size);
+    final_size = closest_size;
+
+    return closest_scheme;
+}
+
+bool FaceTrackModule::IsDetectModeLandmark() const {
+    return m_detect_mode_landmark_;
+}
+
+void FaceTrackModule::SetTrackModeSmoothRatio(float value) {
+    m_track_mode_smooth_ratio_ = value;
+}
+
+void FaceTrackModule::SetTrackModeNumSmoothCacheFrame(int value) {
+    m_track_mode_num_smooth_cache_frame_ = value;
+}
+
+void FaceTrackModule::SetTrackModeDetectInterval(int value) {
+    detection_interval_ = value;
+}
+
+void FaceTrackModule::SetMultiscaleLandmarkLoop(int value) {
+    m_multiscale_landmark_loop_num_ = value;
+    m_multiscale_landmark_scales_ = GenerateCropScales(m_landmark_crop_ratio_, m_multiscale_landmark_loop_num_);
+}
+
+void FaceTrackModule::SetTrackLostRecoveryMode(bool value) {
+    m_track_lost_recovery_mode_ = value;
+}
+
+void FaceTrackModule::SetLightTrackConfidenceThreshold(float value) {
+    m_light_track_confidence_threshold_ = value;
+}
+
+void FaceTrackModule::ClearTrackingFace() {
+    trackingFace.clear();
+    candidate_faces_.clear();
+}
+
+int32_t FaceTrackModule::GetDebugPreviewImageSize() const {
+    return m_debug_preview_image_size_;
+}
+
+}  // namespace inspire
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.h b/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.h
new file mode 100644
index 000000000..e6e98e197
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/face_track_module.h
@@ -0,0 +1,265 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_FACE_TRACK_MODULE_H
+#define INSPIRE_FACE_TRACK_MODULE_FACE_TRACK_MODULE_H
+#include <iostream>
+#include "face_detect/face_detect_adapt.h"
+#include "face_detect/rnet_adapt.h"
+#include "landmark/all.h"
+#include "common/face_info/face_object_internal.h"
+#include "frame_process.h"
+#include "quality/face_pose_quality_adapt.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "tracker_optional/bytetrack/BYTETracker.h"
+#include <data_type.h>
+#include "landmark/landmark_param.h"
+
+namespace inspire {
+
+/**
+ * @class FaceTrack
+ * @brief Class for tracking faces in video streams.
+ *
+ * This class provides functionalities to track faces in real-time video streams. It integrates
+ * face detection, landmark prediction, and face pose quality assessment to track and analyze faces.
+ */
+class INSPIRE_API FaceTrackModule {
+public:
+    /**
+     * @brief Constructor for FaceTrack.
+     * @param max_detected_faces Maximum number of faces to be detected.
+     * @param detection_interval Interval between detections to track faces.
+     * @param track_preview_size Size of the preview for tracking.
+     * @param dynamic_detection_input_level Change the detector input size.
+     */
+    explicit FaceTrackModule(DetectModuleMode mode, int max_detected_faces = 1, int detection_interval = 20, int track_preview_size = -1,
+                             int dynamic_detection_input_level = -1, int TbD_mode_fps = 30, bool detect_mode_landmark = true);
+
+    /**
+     * @brief Configures the face tracking with models.
+     * @param archive Model archive for loading the required modes.
+     * @param expansion_path Expand the path if you need it.
+     * @return int Status of the configuration.
+     */
+    int Configuration(InspireArchive &archive, const std::string &expansion_path = "", bool enable_face_pose_and_quality = false);
+
+    /**
+     * @brief Updates the video stream for face tracking.
+     * @param image Camera stream to process.
+     * @param is_detect Flag to enable/disable face detection.
+     */
+    void UpdateStream(inspirecv::FrameProcess &image);
+
+    /**
+     * @brief Sets the preview size for tracking.
+     * @param preview_size Size of the preview for tracking.
+     */
+    void SetTrackPreviewSize(int preview_size = -1);
+
+    /**
+     * @brief Gets the preview size for tracking.
+     * @return Size of the preview for tracking.
+     */
+    int32_t GetTrackPreviewSize() const;
+
+private:
+    /**
+     * @brief Predicts sparse landmarks for a cropped face image.
+     * @param raw_face_crop Cropped face image.
+     * @param landmarks_output Output vector for predicted landmarks.
+     * @param score Confidence score for the landmarks prediction.
+     * @param size Size for normalizing the face crop.
+     */
+    void SparseLandmarkPredict(const inspirecv::Image &raw_face_crop, std::vector<inspirecv::Point2f> &landmarks_output, float &score,
+                               float size = 112.0);
+
+    /**
+     * @brief Predicts the tracking score for a cropped face image.
+     * @param raw_face_crop Cropped face image.
+     * @return float Tracking score.
+     */
+    float PredictTrackScore(const inspirecv::Image &raw_face_crop);
+
+    /**
+     * @brief Tracks a face in the given image stream.
+     * @param image Camera stream containing the face.
+     * @param face FaceObject to be tracked.
+     * @return bool Status of face tracking.
+     */
+    bool TrackFace(inspirecv::FrameProcess &image, FaceObjectInternal &face);
+
+    /**
+     * @brief Blacks out the region specified in the image for tracking.
+     * @param image Image in which the region needs to be blacked out.
+     * @param rect_mask Rectangle specifying the region to black out.
+     */
+    static void BlackingTrackingRegion(inspirecv::Image &image, inspirecv::Rect2f &rect_mask);
+
+    /**
+     * @brief Performs non-maximum suppression for face detection.
+     * @param th Threshold for suppression.
+     */
+    void nms(float th = 0.5);
+
+    /**
+     * @brief Detects faces in the given image.
+     * @param input Image in which faces are to be detected.
+     * @param scale Scale factor for image processing.
+     */
+    void DetectFace(const inspirecv::Image &input, float scale);
+
+    /**
+     * @brief Initializes the landmark model.
+     * @param model Pointer to the landmark model.
+     * @return int Status of initialization.
+     */
+    int InitLandmarkModel(InspireModel &model);
+
+    /**
+     * @brief Initializes the detection model.
+     * @param model Pointer to the detection model to be initialized.
+     * @return int Status of the initialization process. Returns 0 for success.
+     */
+    int InitDetectModel(InspireModel &model);
+
+    /**
+     * @brief Initializes the RNet (Refinement Network) model.
+     * @param model Pointer to the RNet model to be initialized.
+     * @return int Status of the initialization process. Returns 0 for success.
+     */
+    int InitRNetModel(InspireModel &model);
+
+    /**
+     * @brief Initializes the face pose and quality estimation model.
+     * @param model Pointer to the face pose and quality model to be initialized.
+     * @return int Status of the initialization process. Returns 0 for success.
+     */
+    int InitFacePoseAndQualityModel(InspireModel &model);
+
+    /**
+     * @brief Select the detection model scheme to be used according to the input pixel level.
+     * @param pixel_size Currently, only 160, 320, and 640 pixel sizes are supported.
+     * @return Return the corresponding scheme name, only ”face_detect_160”, ”face_detect_320”, ”face_detect_640” are supported.
+     */
+    std::string ChoiceMultiLevelDetectModel(const int32_t pixel_size, int32_t &final_size);
+
+public:
+    /**
+     * @brief Set the detect threshold
+     * */
+    void SetDetectThreshold(float value);
+
+    /**
+     * @brief Set the minimum face size
+     * */
+    void SetMinimumFacePxSize(float value);
+
+    /**
+     * @brief Check if landmark detection is enabled in detection mode.
+     * @return True if landmark detection is enabled, false otherwise.
+     */
+    bool IsDetectModeLandmark() const;
+
+    /**
+     * @brief Sets the smoothing ratio for landmark tracking
+     * @param value Smoothing ratio between 0 and 1, smaller values mean stronger smoothing
+     */
+    void SetTrackModeSmoothRatio(float value);
+
+    /**
+     * @brief Set the number of smooth cache frame
+     * @param value Number of frames to cache for smoothing
+     */
+    void SetTrackModeNumSmoothCacheFrame(int value);
+
+    /**
+     * @brief Set the detect interval
+     * @param value Interval between detections
+     */
+    void SetTrackModeDetectInterval(int value);
+
+    /**
+     * @brief Set the multiscale landmark loop num
+     * @param value Multiscale landmark loop num
+     */
+    void SetMultiscaleLandmarkLoop(int value);
+
+    /**
+     * @brief Set the track lost recovery mode
+     * @param value Track lost recovery mode
+     */
+    void SetTrackLostRecoveryMode(bool value);
+
+    /**
+     * @brief Set the light track confidence threshold
+     * @param value Light track confidence threshold
+     */
+    void SetLightTrackConfidenceThreshold(float value);
+
+    /**
+     * @brief Clear the tracking face
+     */
+    void ClearTrackingFace();
+
+public:
+    std::vector<FaceObjectInternal> trackingFace;  ///< Vector of FaceObjects currently being tracked.
+
+public:
+    int32_t GetDebugPreviewImageSize() const;
+
+private:
+    const int max_detected_faces_;                     ///< Maximum number of faces to detect.
+    std::vector<FaceObjectInternal> candidate_faces_;  ///< Vector of candidate FaceObjects for tracking.
+    int detection_index_;                              ///< Current detection index.
+    int detection_interval_;                           ///< Interval between detections.
+    int tracking_idx_;                                 ///< Current tracking index.
+    int track_preview_size_;                           ///< Size of the tracking preview.
+    int filter_minimum_face_px_size = 0;               ///< Minimum face pixel allowed to be retained (take the edge with the smallest Rect).
+
+private:
+    // Debug cache
+    int32_t m_debug_preview_image_size_{0};  ///< Debug preview image size
+
+private:
+    std::shared_ptr<FaceDetectAdapt> m_face_detector_;         ///< Shared pointer to the face detector.
+    std::shared_ptr<FaceLandmarkAdapt> m_landmark_predictor_;  ///< Shared pointer to the landmark predictor.
+    std::shared_ptr<RNetAdapt> m_refine_net_;                  ///< Shared pointer to the RNet model.
+
+    std::shared_ptr<FacePoseQualityAdapt> m_face_quality_;  ///< Shared pointer to the face pose quality assessor.
+
+    std::shared_ptr<BYTETracker> m_TbD_tracker_;  ///< Shared pointer to the Bytetrack.
+    int m_dynamic_detection_input_level_ = -1;    ///< Detector size class for dynamic input.
+
+    float m_crop_extensive_ratio_ = 1.8f;  ///< Crop extensive ratio
+    // float m_crop_extensive_ratio_ = 1.5f;  ///< Crop extensive ratio
+    int m_crop_extensive_size_ = 96;  ///< Crop extensive size
+
+    DetectModuleMode m_mode_;  ///< Detect mode
+
+    std::string m_expansion_path_{""};  ///< Expand the path if you need it.
+
+    bool m_detect_mode_landmark_{true};  ///< Detect mode landmark
+
+    int m_track_mode_num_smooth_cache_frame_ = 5;  ///< Track mode number of smooth cache frame
+
+    float m_track_mode_smooth_ratio_ = 0.05;  ///< Track mode smooth ratio
+
+    int m_multiscale_landmark_loop_num_ = 1;  ///< Multiscale landmark loop num
+
+    float m_landmark_crop_ratio_ = 1.1f;
+
+    float m_light_track_confidence_threshold_ = 0.1;  ///< Light track confidence threshold
+
+    std::vector<float> m_multiscale_landmark_scales_;
+
+    bool m_track_lost_recovery_mode_{false};  ///< Track lost recovery mode(only for LightTrack mode)
+
+    std::shared_ptr<LandmarkParam> m_landmark_param_;
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_FACE_TRACK_MODULE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/all.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/all.h
new file mode 100644
index 000000000..c733fb508
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/all.h
@@ -0,0 +1,15 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_LMK_ALL_H
+#define INSPIREFACE_LMK_ALL_H
+
+#include "face_landmark_adapt.h"
+#include "mean_shape.h"
+#include "order_of_hyper_landmark.h"
+#include "face_landmark_adapt.h"
+#include "landmark_tools.h"
+
+#endif  // INSPIREFACE_LMK_ALL_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.cpp
new file mode 100644
index 000000000..46dd6dd54
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.cpp
@@ -0,0 +1,32 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_landmark_adapt.h"
+#include "cost_time.h"
+
+namespace inspire {
+
+std::vector<float> FaceLandmarkAdapt::operator()(const inspirecv::Image& bgr_affine) {
+    COST_TIME_SIMPLE(FaceLandmarkAdapt);
+    AnyTensorOutputs outputs;
+    Forward(bgr_affine, outputs);
+    auto& out = outputs[0].second;
+    if (m_is_center_scaling_) {
+        for (int i = 0; i < out.size(); ++i) {
+            out[i] = (out[i] + 1) / 2;
+        }
+    }
+
+    return out;
+}
+
+FaceLandmarkAdapt::FaceLandmarkAdapt(int input_size, bool is_center_scaling) 
+    : AnyNetAdapter("FaceLandmarkAdapt"), m_input_size_(input_size), m_is_center_scaling_(is_center_scaling) {}
+
+int FaceLandmarkAdapt::getInputSize() const {
+    return m_input_size_;
+}
+
+}  //  namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.h
new file mode 100644
index 000000000..8eca207b8
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/face_landmark_adapt.h
@@ -0,0 +1,51 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_LANDMARK_FACE_LANDMARK_ADAPT_H
+#define INSPIRE_FACE_TRACK_MODULE_LANDMARK_FACE_LANDMARK_ADAPT_H
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @class FaceLandmark
+ * @brief Class for facial landmark detection, inheriting from AnyNet.
+ *
+ * This class specializes in detecting facial landmarks from images using neural network models.
+ */
+class INSPIRE_API FaceLandmarkAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Operator to process an affine-transformed face image and return facial landmarks.
+     * @param bgr_affine Affine-transformed face image in BGR format.
+     * @return std::vector<float> Vector containing the coordinates of facial landmarks.
+     */
+    std::vector<float> operator()(const inspirecv::Image& bgr_affine);
+
+    /**
+     * @brief Constructor for the FaceLandmark class.
+     * @param input_size The size of the input image for the neural network.
+     */
+    explicit FaceLandmarkAdapt(int input_size = 112, bool is_center_scaling = false);
+
+    /**
+     * @brief Gets the input size for the neural network model.
+     * @return int The input size.
+     */
+    int getInputSize() const;
+
+public:
+
+    const static int NUM_OF_LANDMARK = 106;  ///< Total number of landmarks detected.
+
+private:
+    const int m_input_size_;  ///< The input size for the neural network model.
+    bool m_is_center_scaling_;  ///< Whether to use center scaling.
+};
+
+}  //  namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_LANDMARK_FACE_LANDMARK_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_param.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_param.h
new file mode 100644
index 000000000..20ed52b44
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_param.h
@@ -0,0 +1,127 @@
+#ifndef INSPIRE_LANDMARK_PARAM_H
+#define INSPIRE_LANDMARK_PARAM_H
+
+#include "data_type.h"
+#include "yaml-cpp/yaml.h"
+#include "mean_shape.h"
+#include "log.h"
+#include "landmark_tools.h"
+#include "order_of_hyper_landmark.h"
+
+namespace inspire {
+
+typedef struct {
+    int32_t left_eye_center = 67;
+    int32_t right_eye_center = 68;
+    int32_t nose_corner = 100;
+    int32_t mouth_left_corner = 104;
+    int32_t mouth_right_corner = 105;
+    int32_t mouth_lower = 84;
+    int32_t mouth_upper = 87;
+    std::vector<int32_t> left_eye_region = HLMK_LEFT_EYE_POINTS_INDEX;
+    std::vector<int32_t> right_eye_region = HLMK_RIGHT_EYE_POINTS_INDEX;
+} SemanticIndex;
+
+class INSPIRE_API LandmarkParam {
+public:
+    LandmarkParam(const YAML::Node &config) {
+        LoadDefaultMeshShape();
+        if (!config) {
+        } else {
+            // TODO: parse config
+            m_is_available_ = true;
+        }
+        m_table_ = config;
+    }
+
+    void LoadDefaultMeshShape() {
+        mean_shape_points.clear();
+        mean_shape_points.resize(num_of_landmark);
+        for (int k = 0; k < num_of_landmark; k++) {
+            mean_shape_points[k].SetX(HYPLMK_MESH_SHAPE[k * 2]);
+            mean_shape_points[k].SetY(HYPLMK_MESH_SHAPE[k * 2 + 1]);
+        }
+    }
+
+    bool ReLoad(const std::string &name) {
+        if (!m_is_available_) {
+            landmark_engine_name = name;
+            return true;
+        }
+        // parse config
+        auto landmark_table = m_table_[name];
+        if (!landmark_table) {
+            INSPIRE_LOGE("landmark config not found: %s", name.c_str());
+            return false;
+        }
+        num_of_landmark = landmark_table["num_of_landmark"].as<int>();
+        expansion_scale = landmark_table["expansion_scale"].as<float>();
+        input_size = landmark_table["input_size"].as<int>();
+        auto semanic_index = landmark_table["semantic_index"];
+        if (!semanic_index) {
+            INSPIRE_LOGE("semantic_index not found: %s", name.c_str());
+            return false;
+        }
+        semantic_index.left_eye_center = semanic_index["left_eye_center"].as<int>();
+        semantic_index.right_eye_center = semanic_index["right_eye_center"].as<int>();
+        semantic_index.nose_corner = semanic_index["nose_corner"].as<int>();
+        semantic_index.mouth_left_corner = semanic_index["mouth_left_corner"].as<int>();
+        semantic_index.mouth_right_corner = semanic_index["mouth_right_corner"].as<int>();
+        semantic_index.mouth_lower = semanic_index["mouth_lower"].as<int>();
+        semantic_index.mouth_upper = semanic_index["mouth_upper"].as<int>();
+        auto left_eye_region = semanic_index["left_eye_region"];
+        if (left_eye_region) {
+            semantic_index.left_eye_region = left_eye_region.as<std::vector<int>>();
+        }
+        auto right_eye_region = semanic_index["right_eye_region"];
+        if (right_eye_region) {
+            semantic_index.right_eye_region = right_eye_region.as<std::vector<int>>();
+        }
+        auto mesh_shape = landmark_table["mesh_shape"];
+        if (mesh_shape.size() > 0) {
+            std::vector<float> mesh_shape_data = mesh_shape.as<std::vector<float>>();
+            mean_shape_points.clear();
+            mean_shape_points.resize(num_of_landmark);
+            if (mesh_shape_data.size() == num_of_landmark * 2) {
+                for (int i = 0; i < num_of_landmark; i++) {
+                    mean_shape_points[i].SetX(mesh_shape_data[i * 2]);
+                    mean_shape_points[i].SetY(mesh_shape_data[i * 2 + 1]);
+                }
+                mean_shape_points = LandmarkCropped(mean_shape_points);
+                // auto img = inspirecv::Image::Create(192, 192, 3);
+                // img.Fill(0);
+                // for (int i = 0; i < num_of_landmark; i++) {
+                //     auto point = mean_shape_points[i];
+                //     img.DrawCircle(inspirecv::Point2i(point.GetX(), point.GetY()), 2, inspirecv::Color::Red);
+                // }
+                // img.Show("mean_shape");
+            } else {
+                INSPIRE_LOGE("norm_track_index_from_112x size is not equal to num_of_landmark: %s", name.c_str());
+                return false;
+            }
+        } else {
+            LoadDefaultMeshShape();
+        }
+        normalization_mode = landmark_table["normalization_mode"].as<std::string>();
+        landmark_engine_name = name;
+
+        return true;
+    }
+
+public:
+    int num_of_landmark{106};
+    float expansion_scale{1.1f};
+    int input_size{112};
+    std::vector<inspirecv::Point2f> mean_shape_points;
+    SemanticIndex semantic_index;
+    std::string landmark_engine_name{"landmark"};
+    std::string normalization_mode{"MinMax"};
+
+private:
+    YAML::Node m_table_;
+    bool m_is_available_{false};
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_LANDMARK_PARAM_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_tools.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_tools.h
new file mode 100644
index 000000000..fa5ed7e1f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/landmark_tools.h
@@ -0,0 +1,126 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-04-26
+ */
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_LANDMARK_TOOLS_H
+#define INSPIRE_FACE_TRACK_MODULE_LANDMARK_TOOLS_H
+
+#include <vector>
+#include <cmath>
+#include <algorithm>
+#include <inspirecv/inspirecv.h>
+
+namespace inspire {
+
+// Generate crop scales
+inline std::vector<float> GenerateCropScales(float start_scale, int N) {
+    std::vector<float> result;
+    if (N <= 0)
+        return result;
+
+    result.push_back(start_scale);
+
+    float delta_step = 0.02f;  // Initial step size
+    int direction = -1;        // First decrease
+    int expand_count = 1;      // Current expansion count
+
+    for (int i = 1; i < N; ++i) {
+        // The step size increases slightly every two expansions, e.g. 0.02 -> 0.04 -> 0.06
+        float current_delta = delta_step * (static_cast<float>((expand_count + 1) / 2));
+
+        float new_scale = start_scale + direction * current_delta;
+
+        // Keep 5 decimal places
+        new_scale = std::round(new_scale * 100000.0f) / 100000.0f;
+
+        result.push_back(new_scale);
+
+        direction *= -1;  // Alternating positive and negative
+        expand_count++;
+    }
+
+    return result;
+}
+
+inline inspirecv::TransformMatrix ScaleAffineMatrixPreserveCenter(const inspirecv::TransformMatrix& affine, float scale, int output_size = 112) {
+    // The center point in the output image is (cx, cy).
+    float cx = output_size / 2.0f;
+    float cy = output_size / 2.0f;
+
+    // 1. Obtain the position of the current center point in the original image (inverse affine transformation)
+    inspirecv::TransformMatrix inv_affine = affine.GetInverse();
+    float center_x = inv_affine.Get(0, 0) * cx + inv_affine.Get(0, 1) * cy + inv_affine.Get(0, 2);
+    float center_y = inv_affine.Get(1, 0) * cx + inv_affine.Get(1, 1) * cy + inv_affine.Get(1, 2);
+
+    // 2. Create a new scaling matrix (note that the scale value should be reduced → to "expand the cropping area")
+    float inv_scale = 1.0f / scale;
+    float new_a11 = affine.Get(0, 0) * inv_scale;
+    float new_a12 = affine.Get(0, 1) * inv_scale;
+    float new_a21 = affine.Get(1, 0) * inv_scale;
+    float new_a22 = affine.Get(1, 1) * inv_scale;
+
+    // 3. Calculate the new offset to ensure that the center of the scaled image is still mapped to (cx, cy)
+    float new_b1 = cx - (new_a11 * center_x + new_a12 * center_y);
+    float new_b2 = cy - (new_a21 * center_x + new_a22 * center_y);
+
+    return inspirecv::TransformMatrix::Create(new_a11, new_a12, new_b1, new_a21, new_a22, new_b2);
+}
+
+inline std::vector<inspirecv::Point2f> MultiFrameLandmarkMean(const std::vector<std::vector<inspirecv::Point2f>>& points) {
+    std::vector<inspirecv::Point2f> mean_points;
+
+    if (points.empty()) {
+        return mean_points;
+    }
+
+    if (points.size() == 1) {
+        return points[0];
+    }
+
+    size_t num_frames = points.size();
+    size_t num_points = points[0].size();
+
+    // Initialize to 0
+    mean_points.resize(num_points, inspirecv::Point2f(0.0f, 0.0f));
+
+    // Accumulate the coordinates of each point
+    for (const auto& frame : points) {
+        if (frame.size() != num_points)
+            continue;  // skip invalid frame size
+
+        for (size_t i = 0; i < num_points; ++i) {
+            mean_points[i].SetX(mean_points[i].GetX() + frame[i].GetX());
+            mean_points[i].SetY(mean_points[i].GetY() + frame[i].GetY());
+        }
+    }
+
+    // Calculate the average
+    for (auto& pt : mean_points) {
+        pt.SetX(pt.GetX() / static_cast<float>(num_frames));
+        pt.SetY(pt.GetY() / static_cast<float>(num_frames));
+    }
+
+    return mean_points;
+}
+
+inline std::vector<inspirecv::Point2f> LandmarkCropped(const std::vector<inspirecv::Point2f>& points, int output_size = 192) {
+    inspirecv::Rect2f rect = inspirecv::MinBoundingRect(points);
+
+    // Calculate the scale
+    float scale = output_size / std::max(rect.GetWidth(), rect.GetHeight());
+
+    // Create the result vector and scale each point
+    std::vector<inspirecv::Point2f> result;
+    result.reserve(points.size());
+    for (const auto& pt : points) {
+        // Translate to the origin (subtract the top-left corner coordinates), then scale
+        result.push_back(inspirecv::Point2f((pt.GetX() - rect.GetX()) * scale, (pt.GetY() - rect.GetY()) * scale));
+    }
+
+    return result;
+}
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_LANDMARK_TOOLS_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/mean_shape.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/mean_shape.h
new file mode 100644
index 000000000..b1f8331f5
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/mean_shape.h
@@ -0,0 +1,44 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-02-19
+ */
+#pragma once
+#ifndef INSPIREFACE_MEAN_SHAPE_H
+#define INSPIREFACE_MEAN_SHAPE_H
+
+// The mean shape of the hyperlandmarkv2 model
+static const float HYPLMK_MESH_SHAPE[212] = {
+  108.38947968990857, 25.872507850440652, 107.83061691599185, 32.740667888401575, 107.39614893771035, 39.61981246234094,  106.91122964874452,
+  46.49403692042678,  106.2010288700007,  53.34741525135486,  105.09197874965457, 60.15582154225913,  103.42464830472807, 66.83258235371402,
+  101.06766895276914, 73.25011372147983,  97.9868709520959,   79.30544663628336,  94.26571910204586,  84.93503057741032,  90.00562918536812,
+  90.1122202354656,   85.32941316890714,  94.87185192884652,  80.35160378755248,  99.24044649063815,  75.077051388358,    103.10840866085465,
+  69.44018790369117,  106.28129102952379, 63.371791205040466, 108.56111173383506, 56.80262016834233,  109.7499051441407,  50.10379888471675,
+  108.6992507479294,  43.90788570414131,  106.50137834485915, 38.14868418493501,  103.37055610413495, 32.759996698129285, 99.52105297615987,
+  27.680053930061195, 95.16092929087473,  22.917297681524964, 90.39786780114727,  18.575706289743966, 85.20751276433494,  14.768101387475808,
+  79.56267473535601,  11.595841658352757, 73.49136080613575,  9.138881618312933,  67.05404266869571,  7.365347520342082,  60.353342780036044,
+  6.170510409991614,  53.51182231740061,  5.413249465696331,  46.61201199457635,  4.919113585044149,  39.68424569622725,  4.508607756550588,
+  32.75335234242255,  4.0021445204797725, 25.844283881249446, 20.10333643336724,  11.931315790409663, 24.650849944400893, 7.726547826795972,
+  30.408737910838845, 5.739425286951978,  36.63709825167494,  5.7990141975975735, 42.5960288130762,   7.734380599398077,  47.54591990037575,
+  11.374600124045191, 40.6701702686607,   11.684019226473147, 33.713706506191905, 11.418272465358037, 26.812784451308616, 11.269863943170142,
+  93.07932357431682,  11.92899217593448,  88.67053327362375,  7.753896364911641,  83.02397300977135,  5.806492520461124,  76.89097637117361,
+  5.898778431596352,  71.0228769585926,   7.842751904864862,  66.17079582408657,  11.45044902205504,  72.93048624555809,  11.673920920047593,
+  79.77352047369524,  11.42746145492823,  86.54226883746315,  11.312143719439764, 26.190494438734053, 24.325682223561593, 30.366775767967628,
+  21.29695831806158,  35.6070030786359,   20.300764805001755, 40.84233026971642,  21.32760476469489,  44.56320847446876,  24.863093217233942,
+  40.15727240006,     26.299853886839397, 35.25656863819322,  26.745706888418976, 30.437323925963305, 26.09743522459145,  87.12625473052692,
+  24.281244176246002, 83.10628533147951,  21.29726688067164,  77.9629162209028,   20.330772078573286, 72.7718115646307,   21.34368525497503,
+  69.02681716843543,  24.8625797889175,   73.45785066951291,  26.264874015322185, 78.3288579738154,   26.70498208986356,  83.03448398701677,
+  26.040749214404023, 35.98974721945182,  23.296285437719114, 77.6104064849568,   23.28902786261666,  35.42762213737087,  23.907137413710117,
+  78.1019097047022,   23.890769184322778, 73.64132732704913,  75.09299946772475,  68.33809472897121,  71.1431381125012,   61.45715432795517,
+  68.5255727062771,   56.88494393202458,  69.54893655228194,  52.23001938967138,  68.53591299060278,  45.1946880963144,   71.15044515105585,
+  39.531841299807645, 75.07763099178555,  43.88786976326114,  79.83475120739998,  49.515567201234155, 83.26256423798866,  56.81551845731522,
+  84.56820600073104,  63.98463972206808,  83.17902074650735,  69.46501161287217,  79.75710932436276,  65.0011806029609,   76.95904955102614,
+  56.80326229443618,  77.9217550745768,   48.61522852368347,  76.98346163901918,  48.66939742480621,  73.65781445636614,  56.856195386353555,
+  73.62930599992023,  64.98468095979847,  73.66401999131368,  63.45803902648853,  25.14206608791347,  65.950699532727,    38.42145308657457,
+  71.15236199888838,  51.7461970201108,   68.35896800514699,  56.90136467351805,  63.49285164236126,  55.15790554274934,  56.91186565985643,
+  58.48273034863865,  50.38652736379088,  55.156095329634354, 45.490615257275394, 57.00613038907359,  42.508457049067346, 51.765068617716196,
+  47.67302806353165,  38.441738574750175, 50.168926238237994, 25.17549045412592,  56.97274142951366,  48.885497519339204, 56.70557388413479,
+  21.19000556177852,  56.89002821964203,  39.52139553527144,  56.85088564345258,  30.32367503455182,  39.531841299807645, 75.07763099178555,
+  73.64132732704913,  75.09299946772475,
+};
+
+#endif  // INSPIREFACE_MEAN_SHAPE_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/landmark/order_of_hyper_landmark.h b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/order_of_hyper_landmark.h
new file mode 100644
index 000000000..501a51ebe
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/landmark/order_of_hyper_landmark.h
@@ -0,0 +1,21 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_ORDER_HYPERLANDMARK_H
+#define INSPIREFACE_ORDER_HYPERLANDMARK_H
+#include <iostream>
+#include <vector>
+
+namespace inspire {
+
+// HyperLandmarkV2 left eye contour points sequence of dense facial landmarks.
+const std::vector<int> HLMK_LEFT_EYE_POINTS_INDEX = {51, 52, 53, 54, 55, 56, 57, 58};
+
+// HyperLandmarkV2 right eye contour points sequence of dense facial landmarks.
+const std::vector<int> HLMK_RIGHT_EYE_POINTS_INDEX = {59, 60, 61, 62, 63, 64, 65, 66};
+
+}  // namespace inspire
+
+#endif  // INSPIREFACE_ORDER_HYPERLANDMARK_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.cpp
new file mode 100644
index 000000000..889efad27
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.cpp
@@ -0,0 +1,49 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "face_pose_quality_adapt.h"
+#include "middleware/utils.h"
+
+namespace inspire {
+
+FacePoseQualityAdapt::FacePoseQualityAdapt() : AnyNetAdapter("FacePoseQuality") {}
+
+FacePoseQualityAdaptResult FacePoseQualityAdapt::operator()(const inspirecv::Image &img) {
+    FacePoseQualityAdaptResult res;
+    AnyTensorOutputs outputs;
+    if (img.Width() != INPUT_WIDTH || img.Height() != INPUT_HEIGHT) {
+        uint8_t* resized_data = nullptr;
+        m_processor_->Resize(img.Data(), img.Width(), img.Height(), img.Channels(), &resized_data, INPUT_WIDTH, INPUT_HEIGHT);
+        auto resized = inspirecv::Image::Create(INPUT_WIDTH, INPUT_HEIGHT, img.Channels(), resized_data, false);
+        Forward(resized, outputs);
+    } else {
+        Forward(img, outputs);
+    }
+    const auto &output = outputs[0].second;
+    res.pitch = output[0] * 90;
+    res.yaw = output[1] * 90;
+    res.roll = output[2] * 90;
+    std::vector<float> quality(output.begin() + 13, output.end());
+    res.lmk_quality = quality;
+    std::vector<float> face_pts5(output.begin() + 3, output.begin() + 13);
+    res.lmk.resize(5);
+    for (int i = 0; i < 5; i++) {
+        res.lmk[i].SetX((face_pts5[i * 2] + 1) * (INPUT_WIDTH / 2));
+        res.lmk[i].SetY((face_pts5[i * 2 + 1] + 1) * (INPUT_HEIGHT / 2));
+    }
+
+    return res;
+}
+
+inspirecv::TransformMatrix FacePoseQualityAdapt::ComputeCropMatrix(const inspirecv::Rect2i &rect) {
+    auto padding_rect = rect.Square(1.5);
+    std::vector<inspirecv::Point2i> rect_pts = padding_rect.ToFourVertices();
+    std::vector<inspirecv::Point2i> dst_pts = {{0, 0}, {INPUT_WIDTH, 0}, {INPUT_WIDTH, INPUT_HEIGHT}, {0, INPUT_HEIGHT}};
+    inspirecv::TransformMatrix m = inspirecv::SimilarityTransformEstimate(rect_pts, dst_pts);
+
+    return m;
+}
+
+}  // namespace inspire
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.h b/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.h
new file mode 100644
index 000000000..54dd3c331
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/quality/face_pose_quality_adapt.h
@@ -0,0 +1,62 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIRE_FACE_TRACK_MODULE_QUALITY_FACE_POSE_QUALITY_ADAPT_H
+#define INSPIRE_FACE_TRACK_MODULE_QUALITY_FACE_POSE_QUALITY_ADAPT_H
+
+#include "data_type.h"
+#include "middleware/any_net_adapter.h"
+
+namespace inspire {
+
+/**
+ * @struct FacePoseQualityResult
+ * @brief Structure to store the results of face pose quality analysis.
+ *
+ * This includes pitch, yaw, roll angles, landmarks, and their corresponding quality scores.
+ */
+struct FacePoseQualityAdaptResult {
+    float pitch;                          ///< Pitch angle of the face.
+    float yaw;                            ///< Yaw angle of the face.
+    float roll;                           ///< Roll angle of the face.
+    std::vector<inspirecv::Point2f> lmk;  ///< Landmarks of the face.
+    std::vector<float> lmk_quality;       ///< Quality scores for each landmark.
+};
+
+/**
+ * @class FacePoseQuality
+ * @brief Class for assessing the quality of face pose using neural networks.
+ *
+ * Inherits from AnyNet and provides functionalities to compute face pose quality metrics.
+ */
+class INSPIRE_API FacePoseQualityAdapt : public AnyNetAdapter {
+public:
+    /**
+     * @brief Default constructor for FacePoseQuality.
+     */
+    FacePoseQualityAdapt();
+
+    /**
+     * @brief Computes face pose quality for a given affine-transformed face image.
+     * @param bgr_affine Affine-transformed face image in BGR format.
+     * @return FacePoseQualityResult The computed face pose quality metrics.
+     */
+    FacePoseQualityAdaptResult operator()(const inspirecv::Image& img);
+
+    /**
+     * @brief Computes the affine transformation matrix for face cropping.
+     * @param rect Rectangle representing the face in the image.
+     * @return cv::Mat The computed affine transformation matrix.
+     */
+    static inspirecv::TransformMatrix ComputeCropMatrix(const inspirecv::Rect2i& rect);
+
+public:
+    const static int INPUT_WIDTH = 96;   ///< Width of the input image for the network.
+    const static int INPUT_HEIGHT = 96;  ///< Height of the input image for the network.
+};
+
+}  // namespace inspire
+
+#endif  // INSPIRE_FACE_TRACK_MODULE_QUALITY_FACE_POSE_QUALITY_ADAPT_H
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/README.md b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/README.md
new file mode 100644
index 000000000..17b8dbe11
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/README.md
@@ -0,0 +1,10 @@
+## Tracker Optional
+
+### Tracking by Landmark
+
+Tracking by Landmark is a basic tracking mode with high speed but average accuracy. It is recommended to be used in the portrait mode of the front camera of the mobile phone, which is suitable for the case of tracking 1-5 faces.
+
+### Tracking by Detection
+
+Tracking by Detection has high accuracy but slow speed. Currently, only bytetrack is implemented. Source code is in [DeepSORT](https://github.com/shaoshengsong/DeepSORT/tree/master/tracker/bytetrack/).
+
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.cpp
new file mode 100644
index 000000000..efe66eb7b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.cpp
@@ -0,0 +1,206 @@
+#include "BYTETracker.h"
+#include <fstream>
+#include <iostream>
+
+BYTETracker::BYTETracker(int frame_rate, int track_buffer) {
+    track_thresh = 0.5;
+    high_thresh = 0.6;
+    match_thresh = 0.8;
+
+    frame_id = 0;
+    max_time_lost = int(frame_rate / 30.0 * track_buffer);
+    std::cout << "Init ByteTrack!" << std::endl;
+}
+
+BYTETracker::~BYTETracker() {}
+
+vector<STrack> BYTETracker::update(const vector<Object> &objects) {
+    ////////////////// Step 1: Get detections //////////////////
+    this->frame_id++;
+    vector<STrack> activated_stracks;
+    vector<STrack> refind_stracks;
+    vector<STrack> removed_stracks;
+    vector<STrack> lost_stracks;
+    vector<STrack> detections;
+    vector<STrack> detections_low;
+
+    vector<STrack> detections_cp;
+    vector<STrack> tracked_stracks_swap;
+    vector<STrack> resa, resb;
+    vector<STrack> output_stracks;
+
+    vector<STrack *> unconfirmed;
+    vector<STrack *> tracked_stracks;
+    vector<STrack *> strack_pool;
+    vector<STrack *> r_tracked_stracks;
+
+    if (objects.size() > 0) {
+        for (int i = 0; i < objects.size(); i++) {
+            vector<float> tlbr_;
+            tlbr_.resize(4);
+            tlbr_[0] = objects[i].rect.GetX();
+            tlbr_[1] = objects[i].rect.GetY();
+            tlbr_[2] = objects[i].rect.GetX() + objects[i].rect.GetWidth();
+            tlbr_[3] = objects[i].rect.GetY() + objects[i].rect.GetHeight();
+
+            float score = objects[i].prob;
+
+            STrack strack(STrack::tlbr_to_tlwh(tlbr_), score);
+            if (score >= track_thresh) {
+                detections.push_back(strack);
+            } else {
+                detections_low.push_back(strack);
+            }
+        }
+    }
+
+    // Add newly detected tracklets to tracked_stracks
+    for (int i = 0; i < this->tracked_stracks.size(); i++) {
+        if (!this->tracked_stracks[i].is_activated)
+            unconfirmed.push_back(&this->tracked_stracks[i]);
+        else
+            tracked_stracks.push_back(&this->tracked_stracks[i]);
+    }
+
+    ////////////////// Step 2: First association, with IoU //////////////////
+    strack_pool = joint_stracks(tracked_stracks, this->lost_stracks);
+    STrack::multi_predict(strack_pool, this->kalman_filter);
+
+    vector<vector<float> > dists;
+    int dist_size = 0, dist_size_size = 0;
+    dists = iou_distance(strack_pool, detections, dist_size, dist_size_size);
+
+    vector<vector<int> > matches;
+    vector<int> u_track, u_detection;
+    linear_assignment(dists, dist_size, dist_size_size, match_thresh, matches, u_track, u_detection);
+
+    for (int i = 0; i < matches.size(); i++) {
+        STrack *track = strack_pool[matches[i][0]];
+        STrack *det = &detections[matches[i][1]];
+        if (track->state == TrackState::Tracked) {
+            track->update(*det, this->frame_id);
+            activated_stracks.push_back(*track);
+        } else {
+            track->re_activate(*det, this->frame_id, false);
+            refind_stracks.push_back(*track);
+        }
+    }
+
+    ////////////////// Step 3: Second association, using low score dets //////////////////
+    for (int i = 0; i < u_detection.size(); i++) {
+        detections_cp.push_back(detections[u_detection[i]]);
+    }
+    detections.clear();
+    detections.assign(detections_low.begin(), detections_low.end());
+
+    for (int i = 0; i < u_track.size(); i++) {
+        if (strack_pool[u_track[i]]->state == TrackState::Tracked) {
+            r_tracked_stracks.push_back(strack_pool[u_track[i]]);
+        }
+    }
+
+    dists.clear();
+    dists = iou_distance(r_tracked_stracks, detections, dist_size, dist_size_size);
+
+    matches.clear();
+    u_track.clear();
+    u_detection.clear();
+    linear_assignment(dists, dist_size, dist_size_size, 0.5, matches, u_track, u_detection);
+
+    for (int i = 0; i < matches.size(); i++) {
+        STrack *track = r_tracked_stracks[matches[i][0]];
+        STrack *det = &detections[matches[i][1]];
+        if (track->state == TrackState::Tracked) {
+            track->update(*det, this->frame_id);
+            activated_stracks.push_back(*track);
+        } else {
+            track->re_activate(*det, this->frame_id, false);
+            refind_stracks.push_back(*track);
+        }
+    }
+
+    for (int i = 0; i < u_track.size(); i++) {
+        STrack *track = r_tracked_stracks[u_track[i]];
+        if (track->state != TrackState::Lost) {
+            track->mark_lost();
+            lost_stracks.push_back(*track);
+        }
+    }
+
+    // Deal with unconfirmed tracks, usually tracks with only one beginning frame
+    detections.clear();
+    detections.assign(detections_cp.begin(), detections_cp.end());
+
+    dists.clear();
+    dists = iou_distance(unconfirmed, detections, dist_size, dist_size_size);
+
+    matches.clear();
+    vector<int> u_unconfirmed;
+    u_detection.clear();
+    linear_assignment(dists, dist_size, dist_size_size, 0.7, matches, u_unconfirmed, u_detection);
+
+    for (int i = 0; i < matches.size(); i++) {
+        unconfirmed[matches[i][0]]->update(detections[matches[i][1]], this->frame_id);
+        activated_stracks.push_back(*unconfirmed[matches[i][0]]);
+    }
+
+    for (int i = 0; i < u_unconfirmed.size(); i++) {
+        STrack *track = unconfirmed[u_unconfirmed[i]];
+        track->mark_removed();
+        removed_stracks.push_back(*track);
+    }
+
+    ////////////////// Step 4: Init new stracks //////////////////
+    for (int i = 0; i < u_detection.size(); i++) {
+        STrack *track = &detections[u_detection[i]];
+        if (track->score < this->high_thresh)
+            continue;
+        track->activate(this->kalman_filter, this->frame_id);
+        activated_stracks.push_back(*track);
+    }
+
+    ////////////////// Step 5: Update state //////////////////
+    for (int i = 0; i < this->lost_stracks.size(); i++) {
+        if (this->frame_id - this->lost_stracks[i].end_frame() > this->max_time_lost) {
+            this->lost_stracks[i].mark_removed();
+            removed_stracks.push_back(this->lost_stracks[i]);
+        }
+    }
+
+    for (int i = 0; i < this->tracked_stracks.size(); i++) {
+        if (this->tracked_stracks[i].state == TrackState::Tracked) {
+            tracked_stracks_swap.push_back(this->tracked_stracks[i]);
+        }
+    }
+    this->tracked_stracks.clear();
+    this->tracked_stracks.assign(tracked_stracks_swap.begin(), tracked_stracks_swap.end());
+
+    this->tracked_stracks = joint_stracks(this->tracked_stracks, activated_stracks);
+    this->tracked_stracks = joint_stracks(this->tracked_stracks, refind_stracks);
+
+    // std::cout << activated_stracks.size() << std::endl;
+
+    this->lost_stracks = sub_stracks(this->lost_stracks, this->tracked_stracks);
+    for (int i = 0; i < lost_stracks.size(); i++) {
+        this->lost_stracks.push_back(lost_stracks[i]);
+    }
+
+    this->lost_stracks = sub_stracks(this->lost_stracks, this->removed_stracks);
+    for (int i = 0; i < removed_stracks.size(); i++) {
+        this->removed_stracks.push_back(removed_stracks[i]);
+    }
+
+    remove_duplicate_stracks(resa, resb, this->tracked_stracks, this->lost_stracks);
+
+    this->tracked_stracks.clear();
+    this->tracked_stracks.assign(resa.begin(), resa.end());
+    this->lost_stracks.clear();
+    this->lost_stracks.assign(resb.begin(), resb.end());
+
+    for (int i = 0; i < this->tracked_stracks.size(); i++) {
+        if (this->tracked_stracks[i].is_activated) {
+            output_stracks.push_back(this->tracked_stracks[i]);
+        }
+    }
+    return output_stracks;
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.h b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.h
new file mode 100644
index 000000000..25acdb1b9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/BYTETracker.h
@@ -0,0 +1,47 @@
+#pragma once
+
+#include "STrack.h"
+#include "inspirecv/inspirecv.h"
+
+struct Object {
+    inspirecv::Rect<int> rect;
+    int label;
+    float prob;
+};
+
+class BYTETracker {
+public:
+    BYTETracker(int frame_rate = 30, int track_buffer = 30);
+    ~BYTETracker();
+
+    vector<STrack> update(const vector<Object> &objects);
+    inspirecv::Vec3i get_color(int idx);
+
+private:
+    vector<STrack *> joint_stracks(vector<STrack *> &tlista, vector<STrack> &tlistb);
+    vector<STrack> joint_stracks(vector<STrack> &tlista, vector<STrack> &tlistb);
+
+    vector<STrack> sub_stracks(vector<STrack> &tlista, vector<STrack> &tlistb);
+    void remove_duplicate_stracks(vector<STrack> &resa, vector<STrack> &resb, vector<STrack> &stracksa, vector<STrack> &stracksb);
+
+    void linear_assignment(vector<vector<float> > &cost_matrix, int cost_matrix_size, int cost_matrix_size_size, float thresh,
+                           vector<vector<int> > &matches, vector<int> &unmatched_a, vector<int> &unmatched_b);
+    vector<vector<float> > iou_distance(vector<STrack *> &atracks, vector<STrack> &btracks, int &dist_size, int &dist_size_size);
+    vector<vector<float> > iou_distance(vector<STrack> &atracks, vector<STrack> &btracks);
+    vector<vector<float> > ious(vector<vector<float> > &atlbrs, vector<vector<float> > &btlbrs);
+
+    double lapjv(const vector<vector<float> > &cost, vector<int> &rowsol, vector<int> &colsol, bool extend_cost = false, float cost_limit = LONG_MAX,
+                 bool return_cost = true);
+
+private:
+    float track_thresh;
+    float high_thresh;
+    float match_thresh;
+    int frame_id;
+    int max_time_lost;
+
+    vector<STrack> tracked_stracks;
+    vector<STrack> lost_stracks;
+    vector<STrack> removed_stracks;
+    byte_kalman::KalmanFilter kalman_filter;
+};
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.cpp
new file mode 100644
index 000000000..c7512db29
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.cpp
@@ -0,0 +1,172 @@
+#include "STrack.h"
+
+STrack::STrack(vector<float> tlwh_, float score) {
+    _tlwh.resize(4);
+    _tlwh.assign(tlwh_.begin(), tlwh_.end());
+
+    is_activated = false;
+    track_id = 0;
+    state = TrackState::New;
+
+    tlwh.resize(4);
+    tlbr.resize(4);
+
+    static_tlwh();
+    static_tlbr();
+    frame_id = 0;
+    tracklet_len = 0;
+    this->score = score;
+    start_frame = 0;
+}
+
+STrack::~STrack() {}
+
+void STrack::activate(byte_kalman::KalmanFilter &kalman_filter, int frame_id) {
+    this->kalman_filter = kalman_filter;
+    this->track_id = this->next_id();
+
+    vector<float> _tlwh_tmp(4);
+    _tlwh_tmp[0] = this->_tlwh[0];
+    _tlwh_tmp[1] = this->_tlwh[1];
+    _tlwh_tmp[2] = this->_tlwh[2];
+    _tlwh_tmp[3] = this->_tlwh[3];
+    vector<float> xyah = tlwh_to_xyah(_tlwh_tmp);
+    DETECTBOX xyah_box;
+    xyah_box[0] = xyah[0];
+    xyah_box[1] = xyah[1];
+    xyah_box[2] = xyah[2];
+    xyah_box[3] = xyah[3];
+    auto mc = this->kalman_filter.initiate(xyah_box);
+    this->mean = mc.first;
+    this->covariance = mc.second;
+
+    static_tlwh();
+    static_tlbr();
+
+    this->tracklet_len = 0;
+    this->state = TrackState::Tracked;
+    if (frame_id == 1) {
+        this->is_activated = true;
+    }
+    // this->is_activated = true;
+    this->frame_id = frame_id;
+    this->start_frame = frame_id;
+}
+
+void STrack::re_activate(STrack &new_track, int frame_id, bool new_id) {
+    vector<float> xyah = tlwh_to_xyah(new_track.tlwh);
+    DETECTBOX xyah_box;
+    xyah_box[0] = xyah[0];
+    xyah_box[1] = xyah[1];
+    xyah_box[2] = xyah[2];
+    xyah_box[3] = xyah[3];
+    auto mc = this->kalman_filter.update(this->mean, this->covariance, xyah_box);
+    this->mean = mc.first;
+    this->covariance = mc.second;
+
+    static_tlwh();
+    static_tlbr();
+
+    this->tracklet_len = 0;
+    this->state = TrackState::Tracked;
+    this->is_activated = true;
+    this->frame_id = frame_id;
+    this->score = new_track.score;
+    if (new_id)
+        this->track_id = next_id();
+}
+
+void STrack::update(STrack &new_track, int frame_id) {
+    this->frame_id = frame_id;
+    this->tracklet_len++;
+
+    vector<float> xyah = tlwh_to_xyah(new_track.tlwh);
+    DETECTBOX xyah_box;
+    xyah_box[0] = xyah[0];
+    xyah_box[1] = xyah[1];
+    xyah_box[2] = xyah[2];
+    xyah_box[3] = xyah[3];
+
+    auto mc = this->kalman_filter.update(this->mean, this->covariance, xyah_box);
+    this->mean = mc.first;
+    this->covariance = mc.second;
+
+    static_tlwh();
+    static_tlbr();
+
+    this->state = TrackState::Tracked;
+    this->is_activated = true;
+
+    this->score = new_track.score;
+}
+
+void STrack::static_tlwh() {
+    if (this->state == TrackState::New) {
+        tlwh[0] = _tlwh[0];
+        tlwh[1] = _tlwh[1];
+        tlwh[2] = _tlwh[2];
+        tlwh[3] = _tlwh[3];
+        return;
+    }
+
+    tlwh[0] = mean[0];
+    tlwh[1] = mean[1];
+    tlwh[2] = mean[2];
+    tlwh[3] = mean[3];
+
+    tlwh[2] *= tlwh[3];
+    tlwh[0] -= tlwh[2] / 2;
+    tlwh[1] -= tlwh[3] / 2;
+}
+
+void STrack::static_tlbr() {
+    tlbr.clear();
+    tlbr.assign(tlwh.begin(), tlwh.end());
+    tlbr[2] += tlbr[0];
+    tlbr[3] += tlbr[1];
+}
+
+vector<float> STrack::tlwh_to_xyah(vector<float> tlwh_tmp) {
+    vector<float> tlwh_output = tlwh_tmp;
+    tlwh_output[0] += tlwh_output[2] / 2;
+    tlwh_output[1] += tlwh_output[3] / 2;
+    tlwh_output[2] /= tlwh_output[3];
+    return tlwh_output;
+}
+
+vector<float> STrack::to_xyah() {
+    return tlwh_to_xyah(tlwh);
+}
+
+vector<float> STrack::tlbr_to_tlwh(vector<float> &tlbr) {
+    tlbr[2] -= tlbr[0];
+    tlbr[3] -= tlbr[1];
+    return tlbr;
+}
+
+void STrack::mark_lost() {
+    state = TrackState::Lost;
+}
+
+void STrack::mark_removed() {
+    state = TrackState::Removed;
+}
+
+int STrack::next_id() {
+    static int _count = 0;
+    _count++;
+    return _count;
+}
+
+int STrack::end_frame() {
+    return this->frame_id;
+}
+
+void STrack::multi_predict(vector<STrack *> &stracks, byte_kalman::KalmanFilter &kalman_filter) {
+    for (int i = 0; i < stracks.size(); i++) {
+        if (stracks[i]->state != TrackState::Tracked) {
+            stracks[i]->mean[7] = 0;
+        }
+        kalman_filter.predict(stracks[i]->mean, stracks[i]->covariance);
+    }
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.h b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.h
new file mode 100644
index 000000000..123c73b97
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/STrack.h
@@ -0,0 +1,46 @@
+#pragma once
+#include "kalmanFilter.h"
+
+using namespace std;
+
+enum TrackState { New = 0, Tracked, Lost, Removed };
+
+class STrack {
+public:
+    STrack(vector<float> tlwh_, float score);
+    ~STrack();
+
+    vector<float> static tlbr_to_tlwh(vector<float> &tlbr);
+    void static multi_predict(vector<STrack *> &stracks, byte_kalman::KalmanFilter &kalman_filter);
+    void static_tlwh();
+    void static_tlbr();
+    vector<float> tlwh_to_xyah(vector<float> tlwh_tmp);
+    vector<float> to_xyah();
+    void mark_lost();
+    void mark_removed();
+    int next_id();
+    int end_frame();
+
+    void activate(byte_kalman::KalmanFilter &kalman_filter, int frame_id);
+    void re_activate(STrack &new_track, int frame_id, bool new_id = false);
+    void update(STrack &new_track, int frame_id);
+
+public:
+    bool is_activated;
+    int track_id;
+    int state;
+
+    vector<float> _tlwh;
+    vector<float> tlwh;
+    vector<float> tlbr;
+    int frame_id;
+    int tracklet_len;
+    int start_frame;
+
+    KAL_MEAN mean;
+    KAL_COVA covariance;
+    float score;
+
+private:
+    byte_kalman::KalmanFilter kalman_filter;
+};
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/dataType.h b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/dataType.h
new file mode 100644
index 000000000..58b730bea
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/dataType.h
@@ -0,0 +1,36 @@
+#pragma once
+
+#include <cstddef>
+#include <vector>
+
+#include <Eigen/Core>
+#include <Eigen/Dense>
+typedef Eigen::Matrix<float, 1, 4, Eigen::RowMajor> DETECTBOX;
+typedef Eigen::Matrix<float, -1, 4, Eigen::RowMajor> DETECTBOXSS;
+typedef Eigen::Matrix<float, 1, 128, Eigen::RowMajor> FEATURE;
+typedef Eigen::Matrix<float, Eigen::Dynamic, 128, Eigen::RowMajor> FEATURESS;
+// typedef std::vector<FEATURE> FEATURESS;
+
+// Kalmanfilter
+// typedef Eigen::Matrix<float, 8, 8, Eigen::RowMajor> KAL_FILTER;
+typedef Eigen::Matrix<float, 1, 8, Eigen::RowMajor> KAL_MEAN;
+typedef Eigen::Matrix<float, 8, 8, Eigen::RowMajor> KAL_COVA;
+typedef Eigen::Matrix<float, 1, 4, Eigen::RowMajor> KAL_HMEAN;
+typedef Eigen::Matrix<float, 4, 4, Eigen::RowMajor> KAL_HCOVA;
+using KAL_DATA = std::pair<KAL_MEAN, KAL_COVA>;
+using KAL_HDATA = std::pair<KAL_HMEAN, KAL_HCOVA>;
+
+// main
+using RESULT_DATA = std::pair<int, DETECTBOX>;
+
+// tracker:
+using TRACKER_DATA = std::pair<int, FEATURESS>;
+using MATCH_DATA = std::pair<int, int>;
+typedef struct t {
+    std::vector<MATCH_DATA> matches;
+    std::vector<int> unmatched_tracks;
+    std::vector<int> unmatched_detections;
+} TRACHER_MATCHD;
+
+// linear_assignment:
+typedef Eigen::Matrix<float, -1, -1, Eigen::RowMajor> DYNAMICM;
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.cpp
new file mode 100644
index 000000000..b8bd40e50
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.cpp
@@ -0,0 +1,123 @@
+#include "kalmanFilter.h"
+#include <Eigen/Cholesky>
+
+namespace byte_kalman {
+const double KalmanFilter::chi2inv95[10] = {0, 3.8415, 5.9915, 7.8147, 9.4877, 11.070, 12.592, 14.067, 15.507, 16.919};
+KalmanFilter::KalmanFilter() {
+    int ndim = 4;
+    double dt = 1.;
+
+    _motion_mat = Eigen::MatrixXf::Identity(8, 8);
+    for (int i = 0; i < ndim; i++) {
+        _motion_mat(i, ndim + i) = dt;
+    }
+    _update_mat = Eigen::MatrixXf::Identity(4, 8);
+
+    this->_std_weight_position = 1. / 20;
+    this->_std_weight_velocity = 1. / 160;
+}
+
+KAL_DATA KalmanFilter::initiate(const DETECTBOX &measurement) {
+    DETECTBOX mean_pos = measurement;
+    DETECTBOX mean_vel;
+    for (int i = 0; i < 4; i++)
+        mean_vel(i) = 0;
+
+    KAL_MEAN mean;
+    for (int i = 0; i < 8; i++) {
+        if (i < 4)
+            mean(i) = mean_pos(i);
+        else
+            mean(i) = mean_vel(i - 4);
+    }
+
+    KAL_MEAN std;
+    std(0) = 2 * _std_weight_position * measurement[3];
+    std(1) = 2 * _std_weight_position * measurement[3];
+    std(2) = 1e-2;
+    std(3) = 2 * _std_weight_position * measurement[3];
+    std(4) = 10 * _std_weight_velocity * measurement[3];
+    std(5) = 10 * _std_weight_velocity * measurement[3];
+    std(6) = 1e-5;
+    std(7) = 10 * _std_weight_velocity * measurement[3];
+
+    KAL_MEAN tmp = std.array().square();
+    KAL_COVA var = tmp.asDiagonal();
+    return std::make_pair(mean, var);
+}
+
+void KalmanFilter::predict(KAL_MEAN &mean, KAL_COVA &covariance) {
+    // revise the data;
+    DETECTBOX std_pos;
+    std_pos << _std_weight_position * mean(3), _std_weight_position * mean(3), 1e-2, _std_weight_position * mean(3);
+    DETECTBOX std_vel;
+    std_vel << _std_weight_velocity * mean(3), _std_weight_velocity * mean(3), 1e-5, _std_weight_velocity * mean(3);
+    KAL_MEAN tmp;
+    tmp.block<1, 4>(0, 0) = std_pos;
+    tmp.block<1, 4>(0, 4) = std_vel;
+    tmp = tmp.array().square();
+    KAL_COVA motion_cov = tmp.asDiagonal();
+    KAL_MEAN mean1 = this->_motion_mat * mean.transpose();
+    KAL_COVA covariance1 = this->_motion_mat * covariance * (_motion_mat.transpose());
+    covariance1 += motion_cov;
+
+    mean = mean1;
+    covariance = covariance1;
+}
+
+KAL_HDATA KalmanFilter::project(const KAL_MEAN &mean, const KAL_COVA &covariance) {
+    DETECTBOX std;
+    std << _std_weight_position * mean(3), _std_weight_position * mean(3), 1e-1, _std_weight_position * mean(3);
+    KAL_HMEAN mean1 = _update_mat * mean.transpose();
+    KAL_HCOVA covariance1 = _update_mat * covariance * (_update_mat.transpose());
+    Eigen::Matrix<float, 4, 4> diag = std.asDiagonal();
+    diag = diag.array().square().matrix();
+    covariance1 += diag;
+    //    covariance1.diagonal() << diag;
+    return std::make_pair(mean1, covariance1);
+}
+
+KAL_DATA
+KalmanFilter::update(const KAL_MEAN &mean, const KAL_COVA &covariance, const DETECTBOX &measurement) {
+    KAL_HDATA pa = project(mean, covariance);
+    KAL_HMEAN projected_mean = pa.first;
+    KAL_HCOVA projected_cov = pa.second;
+
+    // chol_factor, lower =
+    // scipy.linalg.cho_factor(projected_cov, lower=True, check_finite=False)
+    // kalmain_gain =
+    // scipy.linalg.cho_solve((cho_factor, lower),
+    // np.dot(covariance, self._upadte_mat.T).T,
+    // check_finite=False).T
+    Eigen::Matrix<float, 4, 8> B = (covariance * (_update_mat.transpose())).transpose();
+    Eigen::Matrix<float, 8, 4> kalman_gain = (projected_cov.llt().solve(B)).transpose();  // eg.8x4
+    Eigen::Matrix<float, 1, 4> innovation = measurement - projected_mean;                 // eg.1x4
+    auto tmp = innovation * (kalman_gain.transpose());
+    KAL_MEAN new_mean = (mean.array() + tmp.array()).matrix();
+    KAL_COVA new_covariance = covariance - kalman_gain * projected_cov * (kalman_gain.transpose());
+    return std::make_pair(new_mean, new_covariance);
+}
+
+Eigen::Matrix<float, 1, -1> KalmanFilter::gating_distance(const KAL_MEAN &mean, const KAL_COVA &covariance,
+                                                          const std::vector<DETECTBOX> &measurements, bool only_position) {
+    KAL_HDATA pa = this->project(mean, covariance);
+    if (only_position) {
+        printf("not implement!");
+        exit(0);
+    }
+    KAL_HMEAN mean1 = pa.first;
+    KAL_HCOVA covariance1 = pa.second;
+
+    //    Eigen::Matrix<float, -1, 4, Eigen::RowMajor> d(size, 4);
+    DETECTBOXSS d(measurements.size(), 4);
+    int pos = 0;
+    for (DETECTBOX box : measurements) {
+        d.row(pos++) = box - mean1;
+    }
+    Eigen::Matrix<float, -1, -1, Eigen::RowMajor> factor = covariance1.llt().matrixL();
+    Eigen::Matrix<float, -1, -1> z = factor.triangularView<Eigen::Lower>().solve<Eigen::OnTheRight>(d).transpose();
+    auto zz = ((z.array()) * (z.array())).matrix();
+    auto square_maha = zz.colwise().sum();
+    return square_maha;
+}
+}  // namespace byte_kalman
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.h b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.h
new file mode 100644
index 000000000..8cf0cdd66
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/kalmanFilter.h
@@ -0,0 +1,24 @@
+#pragma once
+
+#include "dataType.h"
+
+namespace byte_kalman {
+class KalmanFilter {
+public:
+    static const double chi2inv95[10];
+    KalmanFilter();
+    KAL_DATA initiate(const DETECTBOX& measurement);
+    void predict(KAL_MEAN& mean, KAL_COVA& covariance);
+    KAL_HDATA project(const KAL_MEAN& mean, const KAL_COVA& covariance);
+    KAL_DATA update(const KAL_MEAN& mean, const KAL_COVA& covariance, const DETECTBOX& measurement);
+
+    Eigen::Matrix<float, 1, -1> gating_distance(const KAL_MEAN& mean, const KAL_COVA& covariance, const std::vector<DETECTBOX>& measurements,
+                                                bool only_position = false);
+
+private:
+    Eigen::Matrix<float, 8, 8, Eigen::RowMajor> _motion_mat;
+    Eigen::Matrix<float, 4, 8, Eigen::RowMajor> _update_mat;
+    float _std_weight_position;
+    float _std_weight_velocity;
+};
+}  // namespace byte_kalman
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.cpp
new file mode 100644
index 000000000..373b69e0b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.cpp
@@ -0,0 +1,306 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "lapjv.h"
+
+/** Column-reduction and reduction transfer for a dense cost matrix.
+ */
+int_t _ccrrt_dense(const uint_t n, cost_t *cost[], int_t *free_rows, int_t *x, int_t *y, cost_t *v) {
+    int_t n_free_rows;
+    boolean *unique;
+
+    for (uint_t i = 0; i < n; i++) {
+        x[i] = -1;
+        v[i] = LARGE;
+        y[i] = 0;
+    }
+    for (uint_t i = 0; i < n; i++) {
+        for (uint_t j = 0; j < n; j++) {
+            const cost_t c = cost[i][j];
+            if (c < v[j]) {
+                v[j] = c;
+                y[j] = i;
+            }
+            PRINTF("i=%d, j=%d, c[i,j]=%f, v[j]=%f y[j]=%d\n", i, j, c, v[j], y[j]);
+        }
+    }
+    PRINT_COST_ARRAY(v, n);
+    PRINT_INDEX_ARRAY(y, n);
+    NEW(unique, boolean, n);
+    memset(unique, TRUE, n);
+    {
+        int_t j = n;
+        do {
+            j--;
+            const int_t i = y[j];
+            if (x[i] < 0) {
+                x[i] = j;
+            } else {
+                unique[i] = FALSE;
+                y[j] = -1;
+            }
+        } while (j > 0);
+    }
+    n_free_rows = 0;
+    for (uint_t i = 0; i < n; i++) {
+        if (x[i] < 0) {
+            free_rows[n_free_rows++] = i;
+        } else if (unique[i]) {
+            const int_t j = x[i];
+            cost_t min = LARGE;
+            for (uint_t j2 = 0; j2 < n; j2++) {
+                if (j2 == (uint_t)j) {
+                    continue;
+                }
+                const cost_t c = cost[i][j2] - v[j2];
+                if (c < min) {
+                    min = c;
+                }
+            }
+            PRINTF("v[%d] = %f - %f\n", j, v[j], min);
+            v[j] -= min;
+        }
+    }
+    FREE(unique);
+    return n_free_rows;
+}
+
+/** Augmenting row reduction for a dense cost matrix.
+ */
+int_t _carr_dense(const uint_t n, cost_t *cost[], const uint_t n_free_rows, int_t *free_rows, int_t *x, int_t *y, cost_t *v) {
+    uint_t current = 0;
+    int_t new_free_rows = 0;
+    uint_t rr_cnt = 0;
+    PRINT_INDEX_ARRAY(x, n);
+    PRINT_INDEX_ARRAY(y, n);
+    PRINT_COST_ARRAY(v, n);
+    PRINT_INDEX_ARRAY(free_rows, n_free_rows);
+    while (current < n_free_rows) {
+        int_t i0;
+        int_t j1, j2;
+        cost_t v1, v2, v1_new;
+        boolean v1_lowers;
+
+        rr_cnt++;
+        PRINTF("current = %d rr_cnt = %d\n", current, rr_cnt);
+        const int_t free_i = free_rows[current++];
+        j1 = 0;
+        v1 = cost[free_i][0] - v[0];
+        j2 = -1;
+        v2 = LARGE;
+        for (uint_t j = 1; j < n; j++) {
+            PRINTF("%d = %f %d = %f\n", j1, v1, j2, v2);
+            const cost_t c = cost[free_i][j] - v[j];
+            if (c < v2) {
+                if (c >= v1) {
+                    v2 = c;
+                    j2 = j;
+                } else {
+                    v2 = v1;
+                    v1 = c;
+                    j2 = j1;
+                    j1 = j;
+                }
+            }
+        }
+        i0 = y[j1];
+        v1_new = v[j1] - (v2 - v1);
+        v1_lowers = v1_new < v[j1];
+        PRINTF("%d %d 1=%d,%f 2=%d,%f v1'=%f(%d,%g) \n", free_i, i0, j1, v1, j2, v2, v1_new, v1_lowers, v[j1] - v1_new);
+        if (rr_cnt < current * n) {
+            if (v1_lowers) {
+                v[j1] = v1_new;
+            } else if (i0 >= 0 && j2 >= 0) {
+                j1 = j2;
+                i0 = y[j2];
+            }
+            if (i0 >= 0) {
+                if (v1_lowers) {
+                    free_rows[--current] = i0;
+                } else {
+                    free_rows[new_free_rows++] = i0;
+                }
+            }
+        } else {
+            PRINTF("rr_cnt=%d >= %d (current=%d * n=%d)\n", rr_cnt, current * n, current, n);
+            if (i0 >= 0) {
+                free_rows[new_free_rows++] = i0;
+            }
+        }
+        x[free_i] = j1;
+        y[j1] = free_i;
+    }
+    return new_free_rows;
+}
+
+/** Find columns with minimum d[j] and put them on the SCAN list.
+ */
+uint_t _find_dense(const uint_t n, uint_t lo, cost_t *d, int_t *cols, int_t *y) {
+    uint_t hi = lo + 1;
+    cost_t mind = d[cols[lo]];
+    for (uint_t k = hi; k < n; k++) {
+        int_t j = cols[k];
+        if (d[j] <= mind) {
+            if (d[j] < mind) {
+                hi = lo;
+                mind = d[j];
+            }
+            cols[k] = cols[hi];
+            cols[hi++] = j;
+        }
+    }
+    return hi;
+}
+
+// Scan all columns in TODO starting from arbitrary column in SCAN
+// and try to decrease d of the TODO columns using the SCAN column.
+int_t _scan_dense(const uint_t n, cost_t *cost[], uint_t *plo, uint_t *phi, cost_t *d, int_t *cols, int_t *pred, int_t *y, cost_t *v) {
+    uint_t lo = *plo;
+    uint_t hi = *phi;
+    cost_t h, cred_ij;
+
+    while (lo != hi) {
+        int_t j = cols[lo++];
+        const int_t i = y[j];
+        const cost_t mind = d[j];
+        h = cost[i][j] - v[j] - mind;
+        PRINTF("i=%d j=%d h=%f\n", i, j, h);
+        // For all columns in TODO
+        for (uint_t k = hi; k < n; k++) {
+            j = cols[k];
+            cred_ij = cost[i][j] - v[j] - h;
+            if (cred_ij < d[j]) {
+                d[j] = cred_ij;
+                pred[j] = i;
+                if (cred_ij == mind) {
+                    if (y[j] < 0) {
+                        return j;
+                    }
+                    cols[k] = cols[hi];
+                    cols[hi++] = j;
+                }
+            }
+        }
+    }
+    *plo = lo;
+    *phi = hi;
+    return -1;
+}
+
+/** Single iteration of modified Dijkstra shortest path algorithm as explained in the JV paper.
+ *
+ * This is a dense matrix version.
+ *
+ * \return The closest free column index.
+ */
+int_t find_path_dense(const uint_t n, cost_t *cost[], const int_t start_i, int_t *y, cost_t *v, int_t *pred) {
+    uint_t lo = 0, hi = 0;
+    int_t final_j = -1;
+    uint_t n_ready = 0;
+    int_t *cols;
+    cost_t *d;
+
+    NEW(cols, int_t, n);
+    NEW(d, cost_t, n);
+
+    for (uint_t i = 0; i < n; i++) {
+        cols[i] = i;
+        pred[i] = start_i;
+        d[i] = cost[start_i][i] - v[i];
+    }
+    PRINT_COST_ARRAY(d, n);
+    while (final_j == -1) {
+        // No columns left on the SCAN list.
+        if (lo == hi) {
+            PRINTF("%d..%d -> find\n", lo, hi);
+            n_ready = lo;
+            hi = _find_dense(n, lo, d, cols, y);
+            PRINTF("check %d..%d\n", lo, hi);
+            PRINT_INDEX_ARRAY(cols, n);
+            for (uint_t k = lo; k < hi; k++) {
+                const int_t j = cols[k];
+                if (y[j] < 0) {
+                    final_j = j;
+                }
+            }
+        }
+        if (final_j == -1) {
+            PRINTF("%d..%d -> scan\n", lo, hi);
+            final_j = _scan_dense(n, cost, &lo, &hi, d, cols, pred, y, v);
+            PRINT_COST_ARRAY(d, n);
+            PRINT_INDEX_ARRAY(cols, n);
+            PRINT_INDEX_ARRAY(pred, n);
+        }
+    }
+
+    PRINTF("found final_j=%d\n", final_j);
+    PRINT_INDEX_ARRAY(cols, n);
+    {
+        const cost_t mind = d[cols[lo]];
+        for (uint_t k = 0; k < n_ready; k++) {
+            const int_t j = cols[k];
+            v[j] += d[j] - mind;
+        }
+    }
+
+    FREE(cols);
+    FREE(d);
+
+    return final_j;
+}
+
+/** Augment for a dense cost matrix.
+ */
+int_t _ca_dense(const uint_t n, cost_t *cost[], const uint_t n_free_rows, int_t *free_rows, int_t *x, int_t *y, cost_t *v) {
+    int_t *pred;
+
+    NEW(pred, int_t, n);
+
+    for (int_t *pfree_i = free_rows; pfree_i < free_rows + n_free_rows; pfree_i++) {
+        int_t i = -1, j;
+        uint_t k = 0;
+
+        PRINTF("looking at free_i=%d\n", *pfree_i);
+        j = find_path_dense(n, cost, *pfree_i, y, v, pred);
+        ASSERT(j >= 0);
+        ASSERT(j < n);
+        while (i != *pfree_i) {
+            PRINTF("augment %d\n", j);
+            PRINT_INDEX_ARRAY(pred, n);
+            i = pred[j];
+            PRINTF("y[%d]=%d -> %d\n", j, y[j], i);
+            y[j] = i;
+            PRINT_INDEX_ARRAY(x, n);
+            SWAP_INDICES(j, x[i]);
+            k++;
+            if (k >= n) {
+                ASSERT(FALSE);
+            }
+        }
+    }
+    FREE(pred);
+    return 0;
+}
+
+/** Solve dense sparse LAP.
+ */
+int lapjv_internal(const uint_t n, cost_t *cost[], int_t *x, int_t *y) {
+    int ret;
+    int_t *free_rows;
+    cost_t *v;
+
+    NEW(free_rows, int_t, n);
+    NEW(v, cost_t, n);
+    ret = _ccrrt_dense(n, cost, free_rows, x, y, v);
+    int i = 0;
+    while (ret > 0 && i < 2) {
+        ret = _carr_dense(n, cost, ret, free_rows, x, y, v);
+        i++;
+    }
+    if (ret > 0) {
+        ret = _ca_dense(n, cost, ret, free_rows, x, y, v);
+    }
+    FREE(v);
+    FREE(free_rows);
+    return ret;
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.h b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.h
new file mode 100644
index 000000000..38c5d263f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/lapjv.h
@@ -0,0 +1,72 @@
+#ifndef LAPJV_H
+#define LAPJV_H
+
+#define LARGE 1000000
+
+#if !defined TRUE
+#define TRUE 1
+#endif
+#if !defined FALSE
+#define FALSE 0
+#endif
+
+#define NEW(x, t, n)                               \
+    if ((x = (t *)malloc(sizeof(t) * (n))) == 0) { \
+        return -1;                                 \
+    }
+#define FREE(x)   \
+    if (x != 0) { \
+        free(x);  \
+        x = 0;    \
+    }
+#define SWAP_INDICES(a, b)     \
+    {                          \
+        int_t _temp_index = a; \
+        a = b;                 \
+        b = _temp_index;       \
+    }
+
+#if 0
+#include <assert.h>
+#define ASSERT(cond) assert(cond)
+#define PRINTF(fmt, ...) printf(fmt, ##__VA_ARGS__)
+#define PRINT_COST_ARRAY(a, n)               \
+    while (1) {                              \
+        printf(#a " = [");                   \
+        if ((n) > 0) {                       \
+            printf("%f", (a)[0]);            \
+            for (uint_t j = 1; j < n; j++) { \
+                printf(", %f", (a)[j]);      \
+            }                                \
+        }                                    \
+        printf("]\n");                       \
+        break;                               \
+    }
+#define PRINT_INDEX_ARRAY(a, n)              \
+    while (1) {                              \
+        printf(#a " = [");                   \
+        if ((n) > 0) {                       \
+            printf("%d", (a)[0]);            \
+            for (uint_t j = 1; j < n; j++) { \
+                printf(", %d", (a)[j]);      \
+            }                                \
+        }                                    \
+        printf("]\n");                       \
+        break;                               \
+    }
+#else
+#define ASSERT(cond)
+#define PRINTF(fmt, ...)
+#define PRINT_COST_ARRAY(a, n)
+#define PRINT_INDEX_ARRAY(a, n)
+#endif
+
+typedef signed int int_t;
+typedef unsigned int uint_t;
+typedef double cost_t;
+typedef char boolean;
+typedef enum fp_t { FP_1 = 1, FP_2 = 2, FP_DYNAMIC = 3 } fp_t;
+
+extern int_t lapjv_internal(const uint_t n, cost_t *cost[], int_t *x, int_t *y);
+
+#endif
diff --git a/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/tracker_utils.cpp b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/tracker_utils.cpp
new file mode 100644
index 000000000..9253efbf3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/track_module/tracker_optional/bytetrack/tracker_utils.cpp
@@ -0,0 +1,346 @@
+#include "BYTETracker.h"
+#include "lapjv.h"
+#include <map>
+#include <iostream>
+#include <algorithm>
+
+vector<STrack *> BYTETracker::joint_stracks(vector<STrack *> &tlista, vector<STrack> &tlistb) {
+    std::map<int, int> exists;
+    vector<STrack *> res;
+    for (int i = 0; i < tlista.size(); i++) {
+        exists.insert(pair<int, int>(tlista[i]->track_id, 1));
+        res.push_back(tlista[i]);
+    }
+    for (int i = 0; i < tlistb.size(); i++) {
+        int tid = tlistb[i].track_id;
+        if (!exists[tid] || exists.count(tid) == 0) {
+            exists[tid] = 1;
+            res.push_back(&tlistb[i]);
+        }
+    }
+    return res;
+}
+
+vector<STrack> BYTETracker::joint_stracks(vector<STrack> &tlista, vector<STrack> &tlistb) {
+    map<int, int> exists;
+    vector<STrack> res;
+    for (int i = 0; i < tlista.size(); i++) {
+        exists.insert(pair<int, int>(tlista[i].track_id, 1));
+        res.push_back(tlista[i]);
+    }
+    for (int i = 0; i < tlistb.size(); i++) {
+        int tid = tlistb[i].track_id;
+        if (!exists[tid] || exists.count(tid) == 0) {
+            exists[tid] = 1;
+            res.push_back(tlistb[i]);
+        }
+    }
+    return res;
+}
+
+vector<STrack> BYTETracker::sub_stracks(vector<STrack> &tlista, vector<STrack> &tlistb) {
+    map<int, STrack> stracks;
+    for (int i = 0; i < tlista.size(); i++) {
+        stracks.insert(pair<int, STrack>(tlista[i].track_id, tlista[i]));
+    }
+    for (int i = 0; i < tlistb.size(); i++) {
+        int tid = tlistb[i].track_id;
+        if (stracks.count(tid) != 0) {
+            stracks.erase(tid);
+        }
+    }
+
+    vector<STrack> res;
+    std::map<int, STrack>::iterator it;
+    for (it = stracks.begin(); it != stracks.end(); ++it) {
+        res.push_back(it->second);
+    }
+
+    return res;
+}
+
+void BYTETracker::remove_duplicate_stracks(vector<STrack> &resa, vector<STrack> &resb, vector<STrack> &stracksa, vector<STrack> &stracksb) {
+    vector<vector<float> > pdist = iou_distance(stracksa, stracksb);
+    vector<pair<int, int> > pairs;
+    for (int i = 0; i < pdist.size(); i++) {
+        for (int j = 0; j < pdist[i].size(); j++) {
+            if (pdist[i][j] < 0.15) {
+                pairs.push_back(pair<int, int>(i, j));
+            }
+        }
+    }
+
+    vector<int> dupa, dupb;
+    for (int i = 0; i < pairs.size(); i++) {
+        int timep = stracksa[pairs[i].first].frame_id - stracksa[pairs[i].first].start_frame;
+        int timeq = stracksb[pairs[i].second].frame_id - stracksb[pairs[i].second].start_frame;
+        if (timep > timeq)
+            dupb.push_back(pairs[i].second);
+        else
+            dupa.push_back(pairs[i].first);
+    }
+
+    for (int i = 0; i < stracksa.size(); i++) {
+        vector<int>::iterator iter = find(dupa.begin(), dupa.end(), i);
+        if (iter == dupa.end()) {
+            resa.push_back(stracksa[i]);
+        }
+    }
+
+    for (int i = 0; i < stracksb.size(); i++) {
+        vector<int>::iterator iter = find(dupb.begin(), dupb.end(), i);
+        if (iter == dupb.end()) {
+            resb.push_back(stracksb[i]);
+        }
+    }
+}
+
+void BYTETracker::linear_assignment(vector<vector<float> > &cost_matrix, int cost_matrix_size, int cost_matrix_size_size, float thresh,
+                                    vector<vector<int> > &matches, vector<int> &unmatched_a, vector<int> &unmatched_b) {
+    if (cost_matrix.size() == 0) {
+        for (int i = 0; i < cost_matrix_size; i++) {
+            unmatched_a.push_back(i);
+        }
+        for (int i = 0; i < cost_matrix_size_size; i++) {
+            unmatched_b.push_back(i);
+        }
+        return;
+    }
+
+    vector<int> rowsol;
+    vector<int> colsol;
+    float c = lapjv(cost_matrix, rowsol, colsol, true, thresh);
+    for (int i = 0; i < rowsol.size(); i++) {
+        if (rowsol[i] >= 0) {
+            vector<int> match;
+            match.push_back(i);
+            match.push_back(rowsol[i]);
+            matches.push_back(match);
+        } else {
+            unmatched_a.push_back(i);
+        }
+    }
+
+    for (int i = 0; i < colsol.size(); i++) {
+        if (colsol[i] < 0) {
+            unmatched_b.push_back(i);
+        }
+    }
+}
+
+vector<vector<float> > BYTETracker::ious(vector<vector<float> > &atlbrs, vector<vector<float> > &btlbrs) {
+    vector<vector<float> > ious;
+    if (atlbrs.size() * btlbrs.size() == 0)
+        return ious;
+
+    ious.resize(atlbrs.size());
+    for (int i = 0; i < ious.size(); i++) {
+        ious[i].resize(btlbrs.size());
+    }
+
+    // bbox_ious
+    for (int k = 0; k < btlbrs.size(); k++) {
+        vector<float> ious_tmp;
+        float box_area = (btlbrs[k][2] - btlbrs[k][0] + 1) * (btlbrs[k][3] - btlbrs[k][1] + 1);
+        for (int n = 0; n < atlbrs.size(); n++) {
+            float iw = min(atlbrs[n][2], btlbrs[k][2]) - max(atlbrs[n][0], btlbrs[k][0]) + 1;
+            if (iw > 0) {
+                float ih = min(atlbrs[n][3], btlbrs[k][3]) - max(atlbrs[n][1], btlbrs[k][1]) + 1;
+                if (ih > 0) {
+                    float ua = (atlbrs[n][2] - atlbrs[n][0] + 1) * (atlbrs[n][3] - atlbrs[n][1] + 1) + box_area - iw * ih;
+                    ious[n][k] = iw * ih / ua;
+                } else {
+                    ious[n][k] = 0.0;
+                }
+            } else {
+                ious[n][k] = 0.0;
+            }
+        }
+    }
+
+    return ious;
+}
+
+vector<vector<float> > BYTETracker::iou_distance(vector<STrack *> &atracks, vector<STrack> &btracks, int &dist_size, int &dist_size_size) {
+    vector<vector<float> > cost_matrix;
+    if (atracks.size() * btracks.size() == 0) {
+        dist_size = atracks.size();
+        dist_size_size = btracks.size();
+        return cost_matrix;
+    }
+    vector<vector<float> > atlbrs, btlbrs;
+    for (int i = 0; i < atracks.size(); i++) {
+        atlbrs.push_back(atracks[i]->tlbr);
+    }
+    for (int i = 0; i < btracks.size(); i++) {
+        btlbrs.push_back(btracks[i].tlbr);
+    }
+
+    dist_size = atracks.size();
+    dist_size_size = btracks.size();
+
+    vector<vector<float> > _ious = ious(atlbrs, btlbrs);
+
+    for (int i = 0; i < _ious.size(); i++) {
+        vector<float> _iou;
+        for (int j = 0; j < _ious[i].size(); j++) {
+            _iou.push_back(1 - _ious[i][j]);
+        }
+        cost_matrix.push_back(_iou);
+    }
+
+    return cost_matrix;
+}
+
+vector<vector<float> > BYTETracker::iou_distance(vector<STrack> &atracks, vector<STrack> &btracks) {
+    vector<vector<float> > atlbrs, btlbrs;
+    for (int i = 0; i < atracks.size(); i++) {
+        atlbrs.push_back(atracks[i].tlbr);
+    }
+    for (int i = 0; i < btracks.size(); i++) {
+        btlbrs.push_back(btracks[i].tlbr);
+    }
+
+    vector<vector<float> > _ious = ious(atlbrs, btlbrs);
+    vector<vector<float> > cost_matrix;
+    for (int i = 0; i < _ious.size(); i++) {
+        vector<float> _iou;
+        for (int j = 0; j < _ious[i].size(); j++) {
+            _iou.push_back(1 - _ious[i][j]);
+        }
+        cost_matrix.push_back(_iou);
+    }
+
+    return cost_matrix;
+}
+
+double BYTETracker::lapjv(const vector<vector<float> > &cost, vector<int> &rowsol, vector<int> &colsol, bool extend_cost, float cost_limit,
+                          bool return_cost) {
+    vector<vector<float> > cost_c;
+    cost_c.assign(cost.begin(), cost.end());
+
+    vector<vector<float> > cost_c_extended;
+
+    int n_rows = cost.size();
+    int n_cols = cost[0].size();
+    rowsol.resize(n_rows);
+    colsol.resize(n_cols);
+
+    int n = 0;
+    if (n_rows == n_cols) {
+        n = n_rows;
+    } else {
+        if (!extend_cost) {
+            std::cout << "set extend_cost=True" << std::endl;
+            // system("pause");
+            exit(0);
+        }
+    }
+
+    if (extend_cost || cost_limit < LONG_MAX) {
+        n = n_rows + n_cols;
+        cost_c_extended.resize(n);
+        for (int i = 0; i < cost_c_extended.size(); i++)
+            cost_c_extended[i].resize(n);
+
+        if (cost_limit < LONG_MAX) {
+            for (int i = 0; i < cost_c_extended.size(); i++) {
+                for (int j = 0; j < cost_c_extended[i].size(); j++) {
+                    cost_c_extended[i][j] = cost_limit / 2.0;
+                }
+            }
+        } else {
+            float cost_max = -1;
+            for (int i = 0; i < cost_c.size(); i++) {
+                for (int j = 0; j < cost_c[i].size(); j++) {
+                    if (cost_c[i][j] > cost_max)
+                        cost_max = cost_c[i][j];
+                }
+            }
+            for (int i = 0; i < cost_c_extended.size(); i++) {
+                for (int j = 0; j < cost_c_extended[i].size(); j++) {
+                    cost_c_extended[i][j] = cost_max + 1;
+                }
+            }
+        }
+
+        for (int i = n_rows; i < cost_c_extended.size(); i++) {
+            for (int j = n_cols; j < cost_c_extended[i].size(); j++) {
+                cost_c_extended[i][j] = 0;
+            }
+        }
+        for (int i = 0; i < n_rows; i++) {
+            for (int j = 0; j < n_cols; j++) {
+                cost_c_extended[i][j] = cost_c[i][j];
+            }
+        }
+
+        cost_c.clear();
+        cost_c.assign(cost_c_extended.begin(), cost_c_extended.end());
+    }
+
+    double **cost_ptr;
+    cost_ptr = new double *[sizeof(double *) * n];
+    for (int i = 0; i < n; i++)
+        cost_ptr[i] = new double[sizeof(double) * n];
+
+    for (int i = 0; i < n; i++) {
+        for (int j = 0; j < n; j++) {
+            cost_ptr[i][j] = cost_c[i][j];
+        }
+    }
+
+    int *x_c = new int[sizeof(int) * n];
+    int *y_c = new int[sizeof(int) * n];
+
+    int ret = lapjv_internal(n, cost_ptr, x_c, y_c);
+    if (ret != 0) {
+        cout << "Calculate Wrong!" << endl;
+        // system("pause");
+        exit(0);
+    }
+
+    double opt = 0.0;
+
+    if (n != n_rows) {
+        for (int i = 0; i < n; i++) {
+            if (x_c[i] >= n_cols)
+                x_c[i] = -1;
+            if (y_c[i] >= n_rows)
+                y_c[i] = -1;
+        }
+        for (int i = 0; i < n_rows; i++) {
+            rowsol[i] = x_c[i];
+        }
+        for (int i = 0; i < n_cols; i++) {
+            colsol[i] = y_c[i];
+        }
+
+        if (return_cost) {
+            for (int i = 0; i < rowsol.size(); i++) {
+                if (rowsol[i] != -1) {
+                    // cout << i << "\t" << rowsol[i] << "\t" << cost_ptr[i][rowsol[i]] << endl;
+                    opt += cost_ptr[i][rowsol[i]];
+                }
+            }
+        }
+    } else if (return_cost) {
+        for (int i = 0; i < rowsol.size(); i++) {
+            opt += cost_ptr[i][rowsol[i]];
+        }
+    }
+
+    for (int i = 0; i < n; i++) {
+        delete[] cost_ptr[i];
+    }
+    delete[] cost_ptr;
+    delete[] x_c;
+    delete[] y_c;
+
+    return opt;
+}
+
+inspirecv::Vec3i BYTETracker::get_color(int idx) {
+    idx += 3;
+    return inspirecv::Vec3i{37 * idx % 255, 17 * idx % 255, 29 * idx % 255};
+}
diff --git a/cpp-package/inspireface/cpp/inspireface/version.txt b/cpp-package/inspireface/cpp/inspireface/version.txt
new file mode 100644
index 000000000..717914c6e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/version.txt
@@ -0,0 +1 @@
+InspireFace Version: 1.2.3
diff --git a/cpp-package/inspireface/cpp/inspireface/version.txt.in b/cpp-package/inspireface/cpp/inspireface/version.txt.in
new file mode 100644
index 000000000..3178c81c9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/inspireface/version.txt.in
@@ -0,0 +1 @@
+InspireFace Version: @INSPIRE_FACE_VERSION_MAJOR@.@INSPIRE_FACE_VERSION_MINOR@.@INSPIRE_FACE_VERSION_PATCH@
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/CMakeLists.txt b/cpp-package/inspireface/cpp/sample/CMakeLists.txt
new file mode 100644
index 000000000..e66e1b783
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/CMakeLists.txt
@@ -0,0 +1,399 @@
+cmake_minimum_required(VERSION 3.20)
+project(InspireFaceSample)
+
+option(ISF_BUILD_SAMPLE_CLUTTERED "Whether to compile the cluttered sample program (debug code during development)" OFF)
+option(ISF_BUILD_SAMPLE_INTERNAL "Whether to compile the internal sample program (debug code during development)" OFF)
+
+include_directories(${SRC_DIR})
+include_directories(${SRC_DIR}/inspireface/c_api)
+include_directories(${SRC_DIR}/inspireface/include)
+
+if (ISF_ENABLE_RKNN AND ISF_RKNPU_MAJOR STREQUAL "rknpu1")
+        set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${ISF_RK_DEVICE_TYPE}/Linux/librknn_api/${CPU_ARCH}/)
+    link_directories(${ISF_RKNN_API_LIB})
+    set(ext rknn_api dl)
+endif ()
+
+if (ISF_ENABLE_RKNN AND ISF_RKNPU_MAJOR STREQUAL "rknpu2" AND ISF_RK_COMPILER_TYPE STREQUAL "aarch64")
+    if(ANDROID)
+        set(RK_PLATFORM "Android")
+    else()
+        set(RK_PLATFORM "Linux")
+    endif()
+    set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/${ISF_RK_COMPILER_TYPE}/)
+    message("ISF_RKNN_API_LIB: ${ISF_RKNN_API_LIB}")
+    link_directories(${ISF_RKNN_API_LIB})
+    set(ext rknnrt dl)
+endif ()
+
+add_executable(Leak api/leak.c)
+target_link_libraries(Leak InspireFace ${ext})
+set_target_properties(Leak PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+# # Examples of face detection and tracking
+add_executable(FaceTrackSample api/sample_face_track.c)
+target_link_libraries(FaceTrackSample InspireFace ${ext})
+set_target_properties(FaceTrackSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(FaceTrackBenchmarkSample api/sample_face_track_benchmark.c)
+target_link_libraries(FaceTrackBenchmarkSample InspireFace ${ext})
+set_target_properties(FaceTrackBenchmarkSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+# Examples of face recognition
+add_executable(FaceComparisonSample api/sample_face_comparison.c)
+target_link_libraries(FaceComparisonSample InspireFace ${ext})
+set_target_properties(FaceComparisonSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(FaceFeatureHubSample api/sample_feature_hub.c)
+target_link_libraries(FaceFeatureHubSample InspireFace ${ext})
+set_target_properties(FaceFeatureHubSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+
+add_executable(FaceLoadReloadSample api/sample_load_reload.c)
+target_link_libraries(FaceLoadReloadSample InspireFace ${ext})
+set_target_properties(FaceLoadReloadSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(CppResourcePoolSample api/sample_cpp_resource_pool.cpp)
+target_link_libraries(CppResourcePoolSample InspireFace ${ext})
+set_target_properties(CppResourcePoolSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(FaceCrudSample api/sample_face_crud.c)
+target_link_libraries(FaceCrudSample InspireFace ${ext})
+set_target_properties(FaceCrudSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(FeatureHubPersistenceSample api/sample_feature_hub_persistence.c)
+target_link_libraries(FeatureHubPersistenceSample InspireFace ${ext})
+set_target_properties(FeatureHubPersistenceSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(CPUWatchingSample api/sample_cpu_watching.c)
+target_link_libraries(CPUWatchingSample InspireFace ${ext})
+set_target_properties(CPUWatchingSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(CreateSessionSample api/sample_create_session.c)
+target_link_libraries(CreateSessionSample InspireFace ${ext})
+set_target_properties(CreateSessionSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(GrayInput api/sample_gray_input.c)
+target_link_libraries(GrayInput InspireFace ${ext})
+set_target_properties(GrayInput PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(SearchFaceSample api/sample_search_face.cpp)
+target_link_libraries(SearchFaceSample InspireFace ${ext})
+set_target_properties(SearchFaceSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+add_executable(CppMultithreadingSample api/sample_cpp_multithreading.cpp)
+target_link_libraries(CppMultithreadingSample InspireFace ${ext})
+set_target_properties(CppMultithreadingSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/api/"
+)
+
+# --- C++ API ---
+
+add_executable(CppSessionSample cpp_api/cpp_sample_face_track.cpp)
+target_link_libraries(CppSessionSample InspireFace ${ext})
+set_target_properties(CppSessionSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+add_executable(CppFaceComparisonSample cpp_api/cpp_sample_face_comparison.cpp)
+target_link_libraries(CppFaceComparisonSample InspireFace ${ext})
+set_target_properties(CppFaceComparisonSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+add_executable(CppFaceCrudSample cpp_api/cpp_sample_face_crud.cpp)
+target_link_libraries(CppFaceCrudSample InspireFace ${ext})
+set_target_properties(CppFaceCrudSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+add_executable(CppSampleAffine cpp_api/cpp_sample_affine.cpp)
+target_link_libraries(CppSampleAffine InspireFace ${ext})
+set_target_properties(CppSampleAffine PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+add_executable(CppSampleInspireCV cpp_api/cpp_sample_inspirecv.cpp)
+target_link_libraries(CppSampleInspireCV InspireFace ${ext})
+set_target_properties(CppSampleInspireCV PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+
+add_executable(CppSampleGrayInput cpp_api/cpp_gray_input.cpp)
+target_link_libraries(CppSampleGrayInput InspireFace ${ext})
+set_target_properties(CppSampleGrayInput PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+add_executable(CppNexusProcessSample cpp_api/cpp_nexus_process.cpp)
+target_link_libraries(CppNexusProcessSample InspireFace ${ext})
+set_target_properties(CppNexusProcessSample PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cpp_api/"
+)
+
+
+if(ISF_BUILD_SAMPLE_INTERNAL)
+        add_executable(FaceTrackerSample source/tracker_sample.cpp)
+        target_link_libraries(FaceTrackerSample InspireFace ${ext})
+        set_target_properties(FaceTrackerSample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/"
+        )
+
+        add_executable(ExpansionLoadSample source/expansion_load.cpp)
+        target_link_libraries(ExpansionLoadSample InspireFace ${ext})
+        set_target_properties(ExpansionLoadSample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/"
+        )
+
+        add_executable(FaceTrackPipelineSample source/tracker_pipeline.cpp)
+        target_link_libraries(FaceTrackPipelineSample InspireFace ${ext})
+        set_target_properties(FaceTrackPipelineSample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/"
+        )
+
+        add_executable(FeatureHubSample source/feature_hub_sample.cpp)
+        target_link_libraries(FeatureHubSample InspireFace ${ext})
+        set_target_properties(FeatureHubSample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/"
+        )
+
+        add_executable(LandmarkSample source/landmark_sample.cpp)
+        target_link_libraries(LandmarkSample InspireFace ${ext})
+        set_target_properties(LandmarkSample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/"
+        )
+
+endif()
+
+# Platform watershed
+if (ISF_BUILD_LINUX_ARM7 OR ISF_BUILD_LINUX_AARCH64)
+    # Typically this is an embedded system or development board scenario where some GUI-related functions are not supported
+elseif(ANDROID)
+    # The executable program on the Android platform generally refers to the running on the shell
+else()
+    # Usually x86 linux or macos
+
+endif ()
+
+if(ISF_RK_DEVICE_TYPE STREQUAL "RV1106")
+    add_executable(FaceTrackSampleRV1106 rv1106/face_detect.cpp)
+    target_link_libraries(FaceTrackSampleRV1106 InspireFace ${ext})
+    set_target_properties(FaceTrackSampleRV1106 PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/internal/"
+            )
+
+    add_executable(FaceAttributeSampleRV1106 rv1106/face_attribute.cpp)
+    target_link_libraries(FaceAttributeSampleRV1106 InspireFace ${ext})
+    set_target_properties(FaceAttributeSampleRV1106 PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/internal/"
+            )
+
+add_executable(NexusImageSample rv1106/rga_image.cpp)
+target_link_libraries(NexusImageSample InspireFace ${ext})
+    set_target_properties(NexusImageSample PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/internal/"
+            )
+
+        # These sample programs are debugging and testing code left behind by developers during the development process.
+        # They are cluttered and have not been organized, or similar functionalities have already been organized in the standard samples.
+        # You can ignore them.
+        if (ISF_BUILD_SAMPLE_CLUTTERED)
+        if (NOT ISF_BUILD_LINUX_ARM7 AND NOT ISF_BUILD_LINUX_AARCH64)
+
+                # =======================InspireFace Sample===========================
+                add_executable(TrackerSample cluttered/standard/tracker_sample.cpp)
+                target_link_libraries(TrackerSample InspireFace)
+
+                set_target_properties(TrackerSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+                add_executable(ContextSample cluttered/standard/context_sample.cpp)
+                target_link_libraries(ContextSample InspireFace)
+
+                set_target_properties(ContextSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+                add_executable(TestSample cluttered/standard/test_sample.cpp)
+                target_link_libraries(TestSample InspireFace)
+
+                set_target_properties(TestSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                add_executable(NetSample cluttered/standard/net_sample.cpp)
+                target_link_libraries(NetSample InspireFace)
+
+                set_target_properties(NetSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+                add_executable(RecSample cluttered/standard/rec_sample.cpp)
+                target_link_libraries(RecSample InspireFace)
+
+                set_target_properties(RecSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+                add_executable(BMSample cluttered/standard/bm_sample.cpp)
+                target_link_libraries(BMSample InspireFace)
+
+                set_target_properties(BMSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+        else()
+
+                # =======================RK Temporary test category===========================
+
+                if (ISF_ENABLE_RKNN)
+                set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/${ISF_RKNPU_MAJOR}/runtime/${ISF_RK_DEVICE_TYPE}/Linux/librknn_api/${CPU_ARCH}/)
+                message("Enable RKNN Inference")
+                link_directories(${ISF_RKNN_API_LIB})
+
+                # Face detection
+                add_executable(RKFaceDetSample cluttered/rk_sample/rk_face_det_sample.cpp)
+                target_link_libraries(RKFaceDetSample InspireFace rknn_api dl)
+
+                set_target_properties(RKFaceDetSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                # Simple network test
+                add_executable(RKSimpleNetSample cluttered/rk_sample/rk_simple_net_sample.cpp)
+                target_link_libraries(RKSimpleNetSample InspireFace rknn_api dl)
+
+                set_target_properties(RKSimpleNetSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                # Face recognize
+                add_executable(RKFaceRecSample cluttered/rk_sample/rk_face_recognize_sample.cpp)
+                target_link_libraries(RKFaceRecSample InspireFace rknn_api dl)
+
+                set_target_properties(RKFaceRecSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                # Tracking module
+                add_executable(RKTrackerSample cluttered/rk_sample/rk_tracker_sample.cpp)
+                target_link_libraries(RKTrackerSample InspireFace rknn_api dl)
+
+                set_target_properties(RKTrackerSample PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                # Debug
+                add_executable(DebugRKRec cluttered/rk_sample/debug_rk_rec.cpp)
+                target_link_libraries(DebugRKRec InspireFace rknn_api dl)
+
+                set_target_properties(DebugRKRec PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+
+                add_executable(ArchTest cluttered/standard/archive_test.cpp)
+                target_link_libraries(ArchTest InspireFace)
+
+                set_target_properties(ArchTest PROPERTIES
+                        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                        )
+
+                endif()
+
+        endif()
+
+
+        # Tracking module
+        add_executable(SQLiteTest cluttered/standard/test_sqlite_sample.cpp)
+        target_link_libraries(SQLiteTest InspireFace)
+
+        set_target_properties(SQLiteTest PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                )
+
+        if (ISF_ENABLE_RKNN)
+                set(DEPEND rknn_api dl)
+        endif ()
+
+        # C_API Demo
+        add_executable(CAPISample cluttered/standard/c_api_sample.cpp)
+        target_link_libraries(CAPISample InspireFace ${DEPEND})
+
+        set_target_properties(CAPISample PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                )
+
+        # C_API Demo
+        add_executable(LoopTracker cluttered/standard/loop_tracker.cpp)
+        target_link_libraries(LoopTracker InspireFace ${DEPEND})
+
+        set_target_properties(LoopTracker PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                )
+
+        add_executable(ArchTracker cluttered/standard/archive_tracker.cpp)
+        target_link_libraries(ArchTracker InspireFace)
+
+        set_target_properties(ArchTracker PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                )
+
+        add_executable(ErrorTest cluttered/standard/error_test.cpp)
+        target_link_libraries(ErrorTest InspireFace)
+
+        set_target_properties(ErrorTest PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/sample/cluttered/"
+                )
+        endif ()
+
+endif()
+
+# Print Message
+message(STATUS ">>>>>>>>>>>>>")
+message(STATUS "InspireFace Sample:")
+message(STATUS "\t ISF_BUILD_SAMPLE_CLUTTERED: ${ISF_BUILD_SAMPLE_CLUTTERED}")
+
+# Install c bin
+install(TARGETS Leak RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+install(TARGETS FaceTrackSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+install(TARGETS FaceComparisonSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+install(TARGETS FaceCrudSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+
+# Install cpp bin
+install(TARGETS CppSessionSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+install(TARGETS CppFaceComparisonSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
+install(TARGETS CppFaceCrudSample RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/sample)
diff --git a/cpp-package/inspireface/cpp/sample/api/leak.c b/cpp-package/inspireface/cpp/sample/api/leak.c
new file mode 100644
index 000000000..efc01d148
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/leak.c
@@ -0,0 +1,11 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <stdlib.h>
+
+int main() {
+    char *n = malloc(1024);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_cpp_multithreading.cpp b/cpp-package/inspireface/cpp/sample/api/sample_cpp_multithreading.cpp
new file mode 100644
index 000000000..f8f840007
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_cpp_multithreading.cpp
@@ -0,0 +1,293 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-07-13
+ * 
+ * Multi-threading example: One thread performs face tracking and stores tokens,
+ * another thread monitors tokens and extracts face features.
+ */
+#include <iostream>
+#include <vector>
+#include <thread>
+#include <mutex>
+#include <condition_variable>
+#include <atomic>
+#include <chrono>
+#include <memory>
+#include <inspireface.h>
+
+// Thread-safe token storage using C++ containers
+class ThreadSafeTokenStorage {
+private:
+    std::vector<HFFaceBasicToken> tokens_;
+    mutable std::mutex mutex_;  // Make mutex mutable for const member functions
+    std::condition_variable cv_;
+    std::atomic<bool> stop_flag_{false};
+
+public:
+    // Add a token to the storage
+    void addToken(const HFFaceBasicToken& token) {
+        std::lock_guard<std::mutex> lock(mutex_);
+        
+        // Copy token data
+        HInt32 tokenSize;
+        HFGetFaceBasicTokenSize(&tokenSize);
+        char* tokenBuffer = new char[tokenSize];
+        HFCopyFaceBasicToken(token, tokenBuffer, tokenSize);
+        
+        // Create copied token
+        HFFaceBasicToken copiedToken;
+        copiedToken.size = tokenSize;
+        copiedToken.data = tokenBuffer;
+        
+        tokens_.push_back(copiedToken);
+        cv_.notify_one();
+    }
+    
+    // Get the last token and remove it
+    bool getLastToken(HFFaceBasicToken& token) {
+        std::unique_lock<std::mutex> lock(mutex_);
+        
+        if (tokens_.empty()) {
+            return false;
+        }
+        
+        token = tokens_.back();
+        tokens_.pop_back();
+        return true;
+    }
+    
+    // Wait for a token with timeout
+    bool waitForToken(HFFaceBasicToken& token, int timeout_ms = 1000) {
+        std::unique_lock<std::mutex> lock(mutex_);
+        
+        if (cv_.wait_for(lock, std::chrono::milliseconds(timeout_ms), 
+                         [this] { return !tokens_.empty() || stop_flag_.load(); })) {
+            if (!tokens_.empty()) {
+                token = tokens_.back();
+                tokens_.pop_back();
+                return true;
+            }
+        }
+        return false;
+    }
+    
+    // Check if there are tokens available
+    bool hasTokens() const {
+        std::lock_guard<std::mutex> lock(mutex_);
+        return !tokens_.empty();
+    }
+    
+    // Get token count
+    size_t getTokenCount() const {
+        std::lock_guard<std::mutex> lock(mutex_);
+        return tokens_.size();
+    }
+    
+    // Stop the threads
+    void stop() {
+        stop_flag_.store(true);
+        cv_.notify_all();
+    }
+    
+    // Check if should stop
+    bool shouldStop() const {
+        return stop_flag_.load();
+    }
+    
+    // Cleanup allocated memory
+    void cleanup() {
+        std::lock_guard<std::mutex> lock(mutex_);
+        for (auto& token : tokens_) {
+            if (token.data != nullptr) {
+                delete[] static_cast<char*>(token.data);
+            }
+        }
+        tokens_.clear();
+    }
+    
+    ~ThreadSafeTokenStorage() {
+        cleanup();
+    }
+};
+
+// Face tracking thread function
+void faceTrackingThread(ThreadSafeTokenStorage& tokenStorage, 
+                       HFSession session,
+                       const std::string& imagePath,
+                       int maxFrames = 100) {
+    std::cout << "Face tracking thread started" << std::endl;
+    
+    // Load image using C API
+    HFImageBitmap imageBitmap = {0};
+    HResult ret = HFCreateImageBitmapFromFilePath(imagePath.c_str(), 3, &imageBitmap);
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to create image bitmap: " << ret << std::endl;
+        return;
+    }
+    
+    // Create image stream
+    HFImageStream stream;
+    ret = HFCreateImageStreamFromImageBitmap(imageBitmap, HF_CAMERA_ROTATION_0, &stream);
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to create image stream: " << ret << std::endl;
+        HFReleaseImageBitmap(imageBitmap);
+        return;
+    }
+    
+    int frameCount = 0;
+    while (!tokenStorage.shouldStop() && frameCount < maxFrames) {
+        // Perform face detection and tracking using C API
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, stream, &multipleFaceData);
+        
+        if (ret == HSUCCEED && multipleFaceData.detectedNum > 0) {
+            // Add tokens to storage
+            for (int i = 0; i < multipleFaceData.detectedNum; i++) {
+                tokenStorage.addToken(multipleFaceData.tokens[i]);
+                
+                std::cout << "Frame " << frameCount << ": Added token for face " 
+                          << multipleFaceData.trackIds[i] << std::endl;
+            }
+        }
+        
+        frameCount++;
+        std::this_thread::sleep_for(std::chrono::milliseconds(100)); // Simulate frame processing
+    }
+    
+    // Cleanup
+    HFReleaseImageStream(stream);
+    HFReleaseImageBitmap(imageBitmap);
+    
+    std::cout << "Face tracking thread finished after " << frameCount << " frames" << std::endl;
+}
+
+// Face feature extraction thread function
+void featureExtractionThread(ThreadSafeTokenStorage& tokenStorage,
+                           HFSession session,
+                           const std::string& imagePath) {
+    std::cout << "Feature extraction thread started" << std::endl;
+    
+    // Load image using C API (same as tracking thread)
+    HFImageBitmap imageBitmap = {0};
+    HResult ret = HFCreateImageBitmapFromFilePath(imagePath.c_str(), 3, &imageBitmap);
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to create image bitmap: " << ret << std::endl;
+        return;
+    }
+    
+    // Create image stream
+    HFImageStream stream;
+    ret = HFCreateImageStreamFromImageBitmap(imageBitmap, HF_CAMERA_ROTATION_0, &stream);
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to create image stream: " << ret << std::endl;
+        HFReleaseImageBitmap(imageBitmap);
+        return;
+    }
+    
+    int extractedCount = 0;
+    while (!tokenStorage.shouldStop()) {
+        HFFaceBasicToken token;
+        
+        // Wait for token with timeout
+        if (tokenStorage.waitForToken(token, 2000)) {
+            // Extract face feature using C API
+            HFFaceFeature feature;
+            ret = HFCreateFaceFeature(&feature);
+            if (ret == HSUCCEED) {
+                ret = HFFaceFeatureExtractTo(session, stream, token, feature);
+                if (ret == HSUCCEED) {
+                    extractedCount++;
+                    std::cout << "Extracted feature " << extractedCount 
+                              << " (feature size: " << feature.size << ")" << std::endl;
+                    
+                    // Print first few feature values as example
+                    std::cout << "Feature values: ";
+                    for (int i = 0; i < std::min(5, feature.size); i++) {
+                        std::cout << feature.data[i] << " ";
+                    }
+                    std::cout << "..." << std::endl;
+                } else {
+                    std::cerr << "Feature extraction failed with error: " << ret << std::endl;
+                }
+                
+                HFReleaseFaceFeature(&feature);
+            }
+            
+            // Clean up token memory
+            if (token.data != nullptr) {
+                delete[] static_cast<char*>(token.data);
+            }
+        } else {
+            std::cout << "No tokens available, waiting..." << std::endl;
+        }
+    }
+    
+    // Cleanup
+    HFReleaseImageStream(stream);
+    HFReleaseImageBitmap(imageBitmap);
+    
+    std::cout << "Feature extraction thread finished, extracted " 
+              << extractedCount << " features" << std::endl;
+}
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cout << "Usage: " << argv[0] << " <model_path> <image_path>" << std::endl;
+        return -1;
+    }
+    
+    std::string modelPath = argv[1];
+    std::string imagePath = argv[2];
+    
+    // Initialize InspireFace using C API
+    HResult ret = HFLaunchInspireFace(modelPath.c_str());
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to launch InspireFace: " << ret << std::endl;
+        return -1;
+    }
+    
+    // Create session using C API
+    HOption option = HF_ENABLE_FACE_RECOGNITION;
+    HFSession session;
+    ret = HFCreateInspireFaceSessionOptional(option, HF_DETECT_MODE_LIGHT_TRACK, 10, -1, -1, &session);
+    if (ret != HSUCCEED) {
+        std::cerr << "Failed to create session: " << ret << std::endl;
+        return -1;
+    }
+    
+    // Set session parameters using C API
+    HFSessionSetTrackPreviewSize(session, 640);
+    
+    // Create thread-safe token storage
+    ThreadSafeTokenStorage tokenStorage;
+    
+    // Start face tracking thread
+    std::thread trackingThread(faceTrackingThread, std::ref(tokenStorage), session, imagePath, 50);
+    
+    // Start feature extraction thread
+    std::thread extractionThread(featureExtractionThread, std::ref(tokenStorage), session, imagePath);
+    
+    // Main thread: monitor and print statistics
+    std::cout << "Main thread: Monitoring token storage..." << std::endl;
+    for (int i = 0; i < 30; i++) {
+        std::this_thread::sleep_for(std::chrono::milliseconds(500));
+        
+        size_t tokenCount = tokenStorage.getTokenCount();
+        std::cout << "Token storage status: " << tokenCount << " tokens available" << std::endl;
+    }
+    
+    // Stop threads
+    std::cout << "Stopping threads..." << std::endl;
+    tokenStorage.stop();
+    
+    // Wait for threads to finish
+    trackingThread.join();
+    extractionThread.join();
+    
+    // Cleanup using C API
+    HFReleaseInspireFaceSession(session);
+    
+    std::cout << "All threads finished successfully" << std::endl;
+    
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_cpp_resource_pool.cpp b/cpp-package/inspireface/cpp/sample/api/sample_cpp_resource_pool.cpp
new file mode 100644
index 000000000..5041cc27a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_cpp_resource_pool.cpp
@@ -0,0 +1,100 @@
+#include <iostream>
+#include <vector>
+#include <string>
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/include/inspireface/session.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/middleware/thread/resource_pool.h>
+#include <inspireface/include/inspireface/spend_timer.h>
+#include <thread>
+
+int main(int argc, char** argv) {
+    if (argc != 5) {
+        std::cerr << "Usage: " << argv[0] << " <model_path> <image_path> <loop_count> <thread_num>" << std::endl;
+        return -1;
+    }
+
+    std::string model_path = argv[1];
+    std::string image_path = argv[2];
+    int loop = std::stoi(argv[3]);
+    int thread_num = std::stoi(argv[4]);
+
+    if (thread_num > 10) {
+        std::cerr << "Error: thread_num cannot be greater than 10" << std::endl;
+        return -1;
+    }
+    if (loop < 1000) {
+        std::cerr << "Error: loop count must be at least 1000" << std::endl;
+        return -1;
+    }
+
+    INSPIREFACE_CONTEXT->Load(model_path);
+    inspirecv::Image image = inspirecv::Image::Create(image_path);
+    inspirecv::FrameProcess process =
+      inspirecv::FrameProcess::Create(image.Data(), image.Height(), image.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    inspire::parallel::ResourcePool<inspire::Session> sessionPool(thread_num, [](inspire::Session& session) {
+
+    });
+
+    for (int i = 0; i < thread_num; ++i) {
+        inspire::CustomPipelineParameter param;
+        param.enable_recognition = true;
+        param.enable_liveness = true;
+        param.enable_mask_detect = true;
+        param.enable_face_attribute = true;
+        param.enable_face_quality = true;
+        inspire::Session session = inspire::Session::Create(inspire::DetectModuleMode::DETECT_MODE_ALWAYS_DETECT, 1, param);
+        sessionPool.AddResource(std::move(session));
+    }
+
+    std::vector<std::thread> threads;
+    int tasksPerThread = loop / thread_num;
+    int remainingTasks = loop % thread_num;
+
+    // Run the task in parallel
+    for (int i = 0; i < thread_num; ++i) {
+        int taskCount = tasksPerThread + (i < remainingTasks ? 1 : 0);
+        threads.emplace_back([&, taskCount]() {
+            for (int j = 0; j < taskCount; ++j) {
+                auto sessionGuard = sessionPool.AcquireResource();
+                std::vector<inspire::FaceTrackWrap> results;
+                int32_t ret;
+                ret = sessionGuard->FaceDetectAndTrack(process, results);
+                if (ret != 0) {
+                    std::cerr << "FaceDetectAndTrack failed" << std::endl;
+                    break;
+                }
+                if (results.size() == 0) {
+                    std::cerr << "Not found face" << std::endl;
+                    break;
+                }
+            }
+        });
+    }
+
+    // Print basic information before starting
+    std::cout << "\n=== Configuration Information ===" << std::endl;
+    std::cout << "Model Path: " << model_path << std::endl;
+    std::cout << "Image Path: " << image_path << std::endl;
+    std::cout << "Total Loop Count: " << loop << std::endl;
+    std::cout << "Number of Threads: " << thread_num << std::endl;
+    std::cout << "Tasks per Thread: " << tasksPerThread << std::endl;
+    std::cout << "Remaining Tasks: " << remainingTasks << std::endl;
+    std::cout << "==============================\n" << std::endl;
+
+    inspire::SpendTimer timer("Number of threads: " + std::to_string(thread_num) + ", Number of tasks: " + std::to_string(loop));
+    timer.Start();
+    for (auto& thread : threads) {
+        thread.join();
+    }
+    timer.Stop();
+    std::cout << timer << std::endl;
+
+    // Convert microseconds to milliseconds and print
+    double milliseconds = timer.Total() / 1000.0;
+    std::cout << "Total execution time: " << milliseconds << " ms" << std::endl;
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_cpu_watching.c b/cpp-package/inspireface/cpp/sample/api/sample_cpu_watching.c
new file mode 100644
index 000000000..fac31c2d6
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_cpu_watching.c
@@ -0,0 +1,47 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <inspireface.h>
+#include <unistd.h>
+
+int main(int argc, char* argv[]) {
+    const char* resourcePath = argv[1];
+    HResult ret = HFReloadInspireFace(resourcePath);
+    if (ret != HSUCCEED)
+    {
+        HFLogPrint(HF_LOG_ERROR, "Failed to launch InspireFace: %d", ret);
+        return 1;
+    }
+
+    // Enable the functions in the pipeline: mask detection, live detection, and face quality
+    // detection
+    HOption option = HF_ENABLE_QUALITY | HF_ENABLE_MASK_DETECT | HF_ENABLE_LIVENESS | HF_ENABLE_IR_LIVENESS | HF_ENABLE_FACE_RECOGNITION;
+    // Non-video or frame sequence mode uses IMAGE-MODE, which is always face detection without
+    // tracking
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    // Maximum number of faces detected
+    HInt32 maxDetectNum = 20;
+    // Face detection image input level
+    HInt32 detectPixelLevel = 160;
+    // Handle of the current face SDK algorithm context
+    HFSession session = {0};
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create FaceContext error: %d", ret);
+        return ret;
+    }
+
+    // Wait for 2000 seconds, watching the cpu usage
+    sleep(2000);
+
+    ret = HFReleaseInspireFaceSession(session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release InspireFaceSession error: %d", ret);
+        return ret;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_create_session.c b/cpp-package/inspireface/cpp/sample/api/sample_create_session.c
new file mode 100644
index 000000000..ed839d5b1
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_create_session.c
@@ -0,0 +1,131 @@
+/**
+ * Created by InspireFace
+ * @date 2025-06-22
+ * Sample code for creating InspireFace session with all features enabled
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <inspireface.h>
+
+int main(int argc, char* argv[]) {
+    HResult ret;
+    const char* packPath;
+    HOption option;
+    HFDetectMode detMode;
+    HInt32 maxDetectNum;
+    HInt32 detectPixelLevel;
+    HFSession session;
+    HFFaceDetectPixelList pixelLevels;
+
+    packPath = argv[1];
+    HFLogPrint(HF_LOG_INFO, "Pack file Path: %s", packPath);
+
+    /* Set log level to debug for detailed output */
+    HFSetLogLevel(HF_LOG_DEBUG);
+
+    /* The resource file must be loaded before it can be used */
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Resource loaded successfully");
+
+    /* Query supported pixel levels for face detection */
+    ret = HFQuerySupportedPixelLevelsForFaceDetection(&pixelLevels);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "HFQuerySupportedPixelLevelsForFaceDetection error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Supported pixel levels for face detection: %d", pixelLevels.size);
+    for (int i = 0; i < pixelLevels.size; i++) {
+        HFLogPrint(HF_LOG_INFO, "Supported pixel level %d: %d", i + 1, pixelLevels.pixel_level[i]);
+    }
+
+    /* Enable ALL available functions in the pipeline */
+    option = HF_ENABLE_FACE_RECOGNITION |    // Face recognition
+             HF_ENABLE_LIVENESS |            // RGB liveness detection
+             HF_ENABLE_IR_LIVENESS |         // IR liveness detection
+             HF_ENABLE_MASK_DETECT |         // Mask detection
+             HF_ENABLE_FACE_ATTRIBUTE |      // Face attribute prediction
+             HF_ENABLE_QUALITY |             // Face quality assessment
+             HF_ENABLE_INTERACTION |         // Interaction feature
+             HF_ENABLE_FACE_POSE |           // Face pose estimation
+             HF_ENABLE_FACE_EMOTION;         // Face emotion recognition
+
+    HFLogPrint(HF_LOG_INFO, "Enabled features:");
+    HFLogPrint(HF_LOG_INFO, "- Face Recognition: YES");
+    HFLogPrint(HF_LOG_INFO, "- RGB Liveness Detection: YES");
+    HFLogPrint(HF_LOG_INFO, "- IR Liveness Detection: YES");
+    HFLogPrint(HF_LOG_INFO, "- Mask Detection: YES");
+    HFLogPrint(HF_LOG_INFO, "- Face Attributes: YES");
+    HFLogPrint(HF_LOG_INFO, "- Face Quality: YES");
+    HFLogPrint(HF_LOG_INFO, "- Interaction: YES");
+    HFLogPrint(HF_LOG_INFO, "- Face Pose: YES");
+    HFLogPrint(HF_LOG_INFO, "- Face Emotion: YES");
+
+    /* Set detection mode - use light track for general purpose */
+    detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    HFLogPrint(HF_LOG_INFO, "Detection mode: HF_DETECT_MODE_LIGHT_TRACK");
+
+    /* Maximum number of faces to detect */
+    maxDetectNum = 20;
+    HFLogPrint(HF_LOG_INFO, "Maximum faces to detect: %d", maxDetectNum);
+
+    /* Face detection image input level */
+    detectPixelLevel = 320;
+    HFLogPrint(HF_LOG_INFO, "Detection pixel level: %d", detectPixelLevel);
+
+    /* Create InspireFace session with all features enabled */
+    session = NULL;
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create InspireFace session error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "InspireFace session created successfully");
+
+    /* Configure session parameters */
+    ret = HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Set track preview size error: %d", ret);
+    } else {
+        HFLogPrint(HF_LOG_INFO, "Track preview size set to: %d", detectPixelLevel);
+    }
+
+    ret = HFSessionSetFilterMinimumFacePixelSize(session, 4);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Set minimum face pixel size error: %d", ret);
+    } else {
+        HFLogPrint(HF_LOG_INFO, "Minimum face pixel size set to: 4");
+    }
+
+    /* Session is now ready for use */
+    HFLogPrint(HF_LOG_INFO, "========================================");
+    HFLogPrint(HF_LOG_INFO, "Session created successfully with ALL features enabled!");
+    HFLogPrint(HF_LOG_INFO, "You can now use this session for:");
+    HFLogPrint(HF_LOG_INFO, "- Face detection and tracking");
+    HFLogPrint(HF_LOG_INFO, "- Face recognition");
+    HFLogPrint(HF_LOG_INFO, "- Liveness detection (RGB & IR)");
+    HFLogPrint(HF_LOG_INFO, "- Mask detection");
+    HFLogPrint(HF_LOG_INFO, "- Face attribute analysis");
+    HFLogPrint(HF_LOG_INFO, "- Face quality assessment");
+    HFLogPrint(HF_LOG_INFO, "- Interaction detection");
+    HFLogPrint(HF_LOG_INFO, "- Face pose estimation");
+    HFLogPrint(HF_LOG_INFO, "- Emotion recognition");
+    HFLogPrint(HF_LOG_INFO, "========================================");
+
+    /* Clean up resources */
+    ret = HFReleaseInspireFaceSession(session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release session error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Session released successfully");
+
+    /* Show resource statistics */
+    HFLogPrint(HF_LOG_INFO, "");
+    HFDeBugShowResourceStatistics();
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_face_comparison.c b/cpp-package/inspireface/cpp/sample/api/sample_face_comparison.c
new file mode 100644
index 000000000..2c79dad9e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_face_comparison.c
@@ -0,0 +1,166 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <inspireface.h>
+
+#define NUM_IMAGES 2
+
+int main(int argc, char* argv[]) {
+    HResult ret;
+    const char* packPath;
+    const char* imgPath1;
+    const char* imgPath2;
+    HOption option;
+    HFSession session;
+    HFFaceFeature features[NUM_IMAGES];
+    const char* imgPaths[NUM_IMAGES];
+    int i;
+    HFloat similarity;
+    HFloat recommended_cosine_threshold;
+    HFloat percentage;
+
+    /* Check whether the number of parameters is correct */
+    if (argc != 4) {
+        HFLogPrint(HF_LOG_ERROR, "Usage: %s <pack_path> <img1_path> <img2_path>", argv[0]);
+        return 1;
+    }
+
+    packPath = argv[1];
+    imgPath1 = argv[2];
+    imgPath2 = argv[3];
+
+    /* Initialize features array to NULL */
+    memset(features, 0, sizeof(features));
+
+    /* Allocate memory for feature vectors */
+    for (i = 0; i < NUM_IMAGES; i++) {
+        ret = HFCreateFaceFeature(&features[i]);
+        if (ret != HSUCCEED) {
+            HFLogPrint(HF_LOG_ERROR, "Create face feature error: %d", ret);
+            goto cleanup;
+        }
+    }
+
+    /* Set the image path array */
+    imgPaths[0] = imgPath1;
+    imgPaths[1] = imgPath2;
+
+    HFLogPrint(HF_LOG_INFO, "Pack file Path: %s", packPath);
+    HFLogPrint(HF_LOG_INFO, "Source file Path 1: %s", imgPath1);
+    HFLogPrint(HF_LOG_INFO, "Source file Path 2: %s", imgPath2);
+
+    /* The resource file must be loaded before it can be used */
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        goto cleanup;
+    }
+
+    /* Create a session for face recognition */
+    option = HF_ENABLE_FACE_RECOGNITION;
+    ret = HFCreateInspireFaceSessionOptional(option, HF_DETECT_MODE_ALWAYS_DETECT, 1, -1, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create session error: %d", ret);
+        goto cleanup;
+    }
+
+    /* Process two images */
+    for (i = 0; i < NUM_IMAGES; i++) {
+        HFImageBitmap imageBitmap = {0};
+        HFImageStream stream;
+        HFMultipleFaceData multipleFaceData = {0};
+
+        ret = HFCreateImageBitmapFromFilePath(imgPaths[i], 3, &imageBitmap);
+        if (ret != HSUCCEED) {
+            HFReleaseImageBitmap(imageBitmap);
+            HFLogPrint(HF_LOG_ERROR, "Create image bitmap error: %d", ret);
+            goto cleanup;
+        }
+
+        ret = HFCreateImageStreamFromImageBitmap(imageBitmap, HF_CAMERA_ROTATION_0, &stream);
+        if (ret != HSUCCEED) {
+            HFReleaseImageStream(stream);
+            HFReleaseImageBitmap(imageBitmap);
+            HFLogPrint(HF_LOG_ERROR, "Create stream error: %d", ret);
+            goto cleanup;
+        }
+
+        ret = HFExecuteFaceTrack(session, stream, &multipleFaceData);
+        if (ret != HSUCCEED) {
+            HFReleaseImageStream(stream);
+            HFReleaseImageBitmap(imageBitmap);
+            HFLogPrint(HF_LOG_ERROR, "Run face track error: %d", ret);
+            goto cleanup;
+        }
+
+        if (multipleFaceData.detectedNum == 0) {
+            HFReleaseImageStream(stream);
+            HFReleaseImageBitmap(imageBitmap);
+            HFLogPrint(HF_LOG_ERROR, "No face was detected: %s", imgPaths[i]);
+            goto cleanup;
+        }
+
+        ret = HFFaceFeatureExtractTo(session, stream, multipleFaceData.tokens[0], features[i]);
+        if (ret != HSUCCEED) {
+            HFReleaseImageStream(stream);
+            HFReleaseImageBitmap(imageBitmap);
+            HFLogPrint(HF_LOG_ERROR, "Extract feature error: %d", ret);
+            goto cleanup;
+        }
+
+        HFReleaseImageStream(stream);
+        HFReleaseImageBitmap(imageBitmap);
+    }
+
+    HFFaceFeature feature1 = features[0];
+    HFFaceFeature feature2 = features[1];
+
+    /* Run comparison */
+    ret = HFFaceComparison(feature1, feature2, &similarity);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Feature comparison error: %d", ret);
+        goto cleanup;
+    }
+
+    ret = HFGetRecommendedCosineThreshold(&recommended_cosine_threshold);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Get recommended cosine threshold error: %d", ret);
+        goto cleanup;
+    }
+
+    if (similarity > recommended_cosine_threshold) {
+        HFLogPrint(HF_LOG_INFO, "%.3f > %.3f ✓ Same face", similarity, recommended_cosine_threshold);
+    } else {
+        HFLogPrint(HF_LOG_WARN, "%.3f < %.3f ✗ Different face", similarity, recommended_cosine_threshold);
+    }
+    HFLogPrint(HF_LOG_INFO, "Similarity score: %.3f", similarity);
+
+    ret = HFCosineSimilarityConvertToPercentage(similarity, &percentage);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Convert similarity to percentage error: %d", ret);
+        goto cleanup;
+    }
+    HFLogPrint(HF_LOG_INFO, "Percentage similarity: %f", percentage);
+
+    /* Clean up resources */
+    ret = HFReleaseInspireFaceSession(session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release session error: %d", ret);
+    }
+    
+cleanup:
+    /* Release the feature vector memory */
+    for (i = 0; i < NUM_IMAGES; i++) {
+        if (features[i].data != NULL) {  // Only release features that were successfully created
+            HFReleaseFaceFeature(&features[i]);
+        }
+    }
+    
+    HFDeBugShowResourceStatistics();
+    
+    return ret;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_face_crud.c b/cpp-package/inspireface/cpp/sample/api/sample_face_crud.c
new file mode 100644
index 000000000..404601837
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_face_crud.c
@@ -0,0 +1,178 @@
+#include <inspireface.h>
+#include <stdio.h>
+
+int main() {
+    HResult ret;
+    // The resource file must be loaded before it can be used
+    ret = HFLaunchInspireFace("test_res/pack/Pikachu");
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+
+    char *db_path = "case_crud.db";
+    if (remove(db_path) != 0) {
+        HFLogPrint(HF_LOG_ERROR, "Remove database file error: %d", ret);
+        return ret;
+    }
+    
+    HFFeatureHubConfiguration configuration;
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = db_path;
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    ret = HFFeatureHubDataEnable(configuration);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Enable feature hub error: %d", ret);
+        return ret;
+    }
+
+    // Create a session
+    HFSession session;
+    ret = HFCreateInspireFaceSessionOptional(HF_ENABLE_FACE_RECOGNITION, HF_DETECT_MODE_ALWAYS_DETECT, 1, 320, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create session error: %d", ret);
+        return ret;
+    }
+    
+    // Prepare an image for insertion into the hub
+    HFImageBitmap image;
+    ret = HFCreateImageBitmapFromFilePath("test_res/data/bulk/kun.jpg", 3, &image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image bitmap error: %d", ret);
+        return ret;
+    }
+    
+    // Create an image stream
+    HFImageStream imageHandle;
+    ret = HFCreateImageStreamFromImageBitmap(image, HF_CAMERA_ROTATION_0, &imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image stream error: %d", ret);
+        return ret;
+    }
+
+    // Detect and track
+    HFMultipleFaceData multipleFaceData;
+    ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Execute face track error: %d", ret);
+        return ret;
+    }
+
+    if (multipleFaceData.detectedNum > 0) {
+        HFLogPrint(HF_LOG_INFO, "Face detected: %d", multipleFaceData.detectedNum);
+    }
+
+    HFFaceFeature feature;
+    ret = HFCreateFaceFeature(&feature);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create face feature error: %d", ret);
+        return ret;
+    }
+
+    ret = HFFaceFeatureExtractCpy(session, imageHandle, multipleFaceData.tokens[0], feature.data);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Extract feature error: %d", ret);
+        return ret;
+    }
+
+    // Insert face feature into the hub
+    HFFaceFeatureIdentity featureIdentity;
+    featureIdentity.feature = &feature;
+    featureIdentity.id = -1;
+    HFaceId result_id;
+    ret = HFFeatureHubInsertFeature(featureIdentity, &result_id);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Insert feature error: %d", ret);
+        return ret;
+    }
+
+    // Prepare a photo of the same person for the query
+    HFImageBitmap query_image;
+    ret = HFCreateImageBitmapFromFilePath("test_res/data/bulk/jntm.jpg", 3, &query_image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image bitmap error: %d", ret);
+        return ret;
+    }
+
+    // Create an image stream
+    HFImageStream query_imageHandle;
+    ret = HFCreateImageStreamFromImageBitmap(query_image, HF_CAMERA_ROTATION_0, &query_imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image stream error: %d", ret);
+        return ret;
+    }
+
+    // Detect and track
+    ret = HFExecuteFaceTrack(session, query_imageHandle, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Execute face track error: %d", ret);
+        return ret;
+    }
+
+    if (multipleFaceData.detectedNum > 0) {
+        HFLogPrint(HF_LOG_INFO, "Face detected: %d", multipleFaceData.detectedNum);
+    }
+
+    HFFaceFeature query_feature;
+    ret = HFCreateFaceFeature(&query_feature);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create face feature error: %d", ret);
+        return ret;
+    }
+    
+    // Extract face feature
+    ret = HFFaceFeatureExtractTo(session, query_imageHandle, multipleFaceData.tokens[0], query_feature);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Extract feature error: %d", ret);
+        return ret;
+    }
+
+    // Search face feature
+    HFFaceFeatureIdentity query_featureIdentity;
+    query_featureIdentity.feature = &query_feature;
+    query_featureIdentity.id = -1;
+    HFloat confidence;
+    ret = HFFeatureHubFaceSearch(query_feature, &confidence, &query_featureIdentity);
+
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Search feature error: %d", ret);
+        return ret;
+    }
+
+    HFLogPrint(HF_LOG_INFO, "Search feature result: %d", query_featureIdentity.id);
+    HFLogPrint(HF_LOG_INFO, "Search feature confidence: %f", confidence);
+
+    // Remove face feature
+    ret = HFFeatureHubFaceRemove(result_id);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Remove feature error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Remove feature result: %d", result_id);    
+
+    // Query again
+    ret = HFFeatureHubFaceSearch(query_feature, &confidence, &query_featureIdentity);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Search feature error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Query again, search feature result: %d", query_featureIdentity.id);
+    if (query_featureIdentity.id != -1) {
+        HFLogPrint(HF_LOG_INFO, "Remove feature failed");
+    }
+
+    // Release resources
+    HFReleaseFaceFeature(&feature);
+    HFReleaseFaceFeature(&query_feature);
+    HFReleaseImageStream(imageHandle);
+    HFReleaseImageStream(query_imageHandle);
+    HFReleaseImageBitmap(image);
+    HFReleaseImageBitmap(query_image);
+    HFReleaseInspireFaceSession(session);
+
+    HFDeBugShowResourceStatistics();
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_face_track.c b/cpp-package/inspireface/cpp/sample/api/sample_face_track.c
new file mode 100644
index 000000000..17bddb676
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_face_track.c
@@ -0,0 +1,277 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <inspireface.h>
+
+int main(int argc, char* argv[]) {
+    HResult ret;
+    const char* packPath;
+    const char* sourcePath;
+    int rotation;
+    HFRotation rotation_enum;
+    HOption option;
+    HFDetectMode detMode;
+    HInt32 maxDetectNum;
+    HInt32 detectPixelLevel;
+    HFSession session;
+    HFImageBitmap image;
+    HFImageStream imageHandle;
+    HFMultipleFaceData multipleFaceData;
+    int faceNum;
+    HFImageBitmap drawImage;
+    HFImageBitmapData data;
+    int index;
+    HFFaceMaskConfidence maskConfidence;
+    HFFaceQualityConfidence qualityConfidence;
+    HOption pipelineOption;
+    HFFaceDetectPixelList pixelLevels;
+    HFFaceEmotionResult faceEmotionResult;
+
+    /* Check whether the number of parameters is correct */
+    if (argc < 3 || argc > 4) {
+        HFLogPrint(HF_LOG_ERROR, "Usage: %s <pack_path> <source_path> [rotation]", argv[0]);
+        return 1;
+    }
+
+    packPath = argv[1];
+    sourcePath = argv[2];
+    rotation = 0;
+
+    /* If rotation is provided, check and set the value */
+    if (argc == 4) {
+        rotation = atoi(argv[3]);
+        if (rotation != 0 && rotation != 90 && rotation != 180 && rotation != 270) {
+            HFLogPrint(HF_LOG_ERROR, "Invalid rotation value. Allowed values are 0, 90, 180, 270.");
+            return 1;
+        }
+    }
+
+    /* Set rotation based on input parameter */
+    switch (rotation) {
+        case 90:
+            rotation_enum = HF_CAMERA_ROTATION_90;
+            break;
+        case 180:
+            rotation_enum = HF_CAMERA_ROTATION_180;
+            break;
+        case 270:
+            rotation_enum = HF_CAMERA_ROTATION_270;
+            break;
+        case 0:
+        default:
+            rotation_enum = HF_CAMERA_ROTATION_0;
+            break;
+    }
+
+    HFLogPrint(HF_LOG_INFO, "Pack file Path: %s", packPath);
+    HFLogPrint(HF_LOG_INFO, "Source file Path: %s", sourcePath);
+    HFLogPrint(HF_LOG_INFO, "Rotation: %d", rotation);
+
+    HFSetLogLevel(HF_LOG_DEBUG);
+
+    /* The resource file must be loaded before it can be used */
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+
+    ret = HFQuerySupportedPixelLevelsForFaceDetection(&pixelLevels);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "HFQuerySupportedPixelLevelsForFaceDetection error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Supported pixel levels for face detection: %d", pixelLevels.size);
+    for (int i = 0; i < pixelLevels.size; i++) {
+        HFLogPrint(HF_LOG_INFO, "Supported pixel level %d: %d", i + 1, pixelLevels.pixel_level[i]);
+    }
+
+    /* Enable the functions in the pipeline: mask detection, live detection, and face quality
+     * detection */
+    option = HF_ENABLE_QUALITY | HF_ENABLE_MASK_DETECT | HF_ENABLE_LIVENESS | HF_ENABLE_FACE_EMOTION;
+    /* Non-video or frame sequence mode uses IMAGE-MODE, which is always face detection without
+     * tracking */
+    detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    /* Maximum number of faces detected */
+    maxDetectNum = 20;
+    /* Face detection image input level */
+    detectPixelLevel = 320;
+    /* Handle of the current face SDK algorithm context */
+    session = NULL;
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create FaceContext error: %d", ret);
+        return ret;
+    }
+
+    HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+    HFSessionSetFilterMinimumFacePixelSize(session, 4);
+
+    /* Load a image */
+    ret = HFCreateImageBitmapFromFilePath(sourcePath, 3, &image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "The source entered is not a picture or read error.");
+        return ret;
+    }
+    /* Prepare an image parameter structure for configuration */
+    ret = HFCreateImageStreamFromImageBitmap(image, rotation_enum, &imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create ImageStream error: %d", ret);
+        return ret;
+    }
+
+    /* Execute HF_FaceContextRunFaceTrack captures face information in an image */
+    ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Execute HFExecuteFaceTrack error: %d", ret);
+        return ret;
+    }
+
+    /* Print the number of faces detected */
+    faceNum = multipleFaceData.detectedNum;
+    HFLogPrint(HF_LOG_INFO, "Num of face: %d", faceNum);
+
+    /* Copy a new image to draw */
+    ret = HFImageBitmapCopy(image, &drawImage);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Copy ImageBitmap error: %d", ret);
+        return ret;
+    }
+    ret = HFImageBitmapGetData(drawImage, &data);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Get ImageBitmap data error: %d", ret);
+        return ret;
+    }
+    for (index = 0; index < faceNum; ++index) {
+        HInt32 numOfLmk;
+        HPoint2f* denseLandmarkPoints;
+        HPoint2f fiveKeyPoints[5];
+        float area;
+        size_t i;
+
+        HFLogPrint(HF_LOG_INFO, "========================================");
+        HFLogPrint(HF_LOG_INFO, "Token size: %d", multipleFaceData.tokens[index].size);
+        HFLogPrint(HF_LOG_INFO, "Process face index: %d", index);
+        HFLogPrint(HF_LOG_INFO, "DetConfidence: %f", multipleFaceData.detConfidence[index]);
+        HFLogPrint(HF_LOG_INFO, "TrackCount: %d", multipleFaceData.trackCounts[index]);
+        
+        HFImageBitmapDrawRect(drawImage, multipleFaceData.rects[index], (HColor){0, 100, 255}, 4);
+
+        /* Get the number of dense landmark points */
+        HFGetNumOfFaceDenseLandmark(&numOfLmk);
+        denseLandmarkPoints = (HPoint2f*)malloc(sizeof(HPoint2f) * numOfLmk);
+        if (denseLandmarkPoints == NULL) {
+            HFLogPrint(HF_LOG_ERROR, "Memory allocation failed!");
+            return -1;
+        }
+
+        ret = HFGetFaceDenseLandmarkFromFaceToken(multipleFaceData.tokens[index], denseLandmarkPoints, numOfLmk);
+        if (ret != HSUCCEED) {
+            free(denseLandmarkPoints);
+            HFLogPrint(HF_LOG_ERROR, "HFGetFaceDenseLandmarkFromFaceToken error!!");
+            return -1;
+        }
+
+        /* Draw dense landmark points */
+        for (i = 0; i < numOfLmk; i++) {
+            HFImageBitmapDrawCircleF(drawImage, 
+                                   (HPoint2f){denseLandmarkPoints[i].x, denseLandmarkPoints[i].y}, 
+                                   0, 
+                                   (HColor){100, 100, 0}, 
+                                   2);
+        }
+        free(denseLandmarkPoints);
+
+        HFaceRect rt = multipleFaceData.rects[index];
+        area = ((float)(rt.height * rt.width)) / (data.width * data.height);
+        HFLogPrint(HF_LOG_INFO, "area: %f", area);
+
+        ret = HFGetFaceFiveKeyPointsFromFaceToken(multipleFaceData.tokens[index], fiveKeyPoints, 5);
+        if (ret != HSUCCEED) {
+            HFLogPrint(HF_LOG_ERROR, "HFGetFaceFiveKeyPointsFromFaceToken error!!");
+            return -1;
+        }
+        for (i = 0; i < 5; i++) {
+            HFImageBitmapDrawCircleF(drawImage, (HPoint2f){fiveKeyPoints[i].x, fiveKeyPoints[i].y}, 0, (HColor){0, 0, 232}, 2);
+        }
+    }
+    HFImageBitmapWriteToFile(drawImage, "draw_detected.jpg");
+    HFLogPrint(HF_LOG_WARN, "Write to file success: %s", "draw_detected.jpg");
+
+    /* Run pipeline function */
+    /* Select the pipeline function that you want to execute, provided that it is already enabled
+     * when FaceContext is created! */
+    pipelineOption = HF_ENABLE_QUALITY | HF_ENABLE_MASK_DETECT | HF_ENABLE_LIVENESS | HF_ENABLE_FACE_EMOTION;
+    /* In this loop, all faces are processed */
+    ret = HFMultipleFacePipelineProcessOptional(session, imageHandle, &multipleFaceData, pipelineOption);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Execute Pipeline error: %d", ret);
+        return ret;
+    }
+
+    /* Get mask detection results from the pipeline cache */
+    ret = HFGetFaceMaskConfidence(session, &maskConfidence);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Get mask detect result error: %d", ret);
+        return -1;
+    }
+
+    /* Get face quality results from the pipeline cache */
+    ret = HFGetFaceQualityConfidence(session, &qualityConfidence);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Get face quality result error: %d", ret);
+        return -1;
+    }
+
+    ret = HFGetFaceEmotionResult(session, &faceEmotionResult);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Get face emotion result error: %d", ret);
+        return -1;
+    }
+
+    for (index = 0; index < faceNum; ++index) {
+        HFLogPrint(HF_LOG_INFO, "========================================");
+        HFLogPrint(HF_LOG_INFO, "Process face index from pipeline: %d", index);
+        HFLogPrint(HF_LOG_INFO, "Mask detect result: %f", maskConfidence.confidence[index]);
+        HFLogPrint(HF_LOG_INFO, "Quality predict result: %f", qualityConfidence.confidence[index]);
+        HFLogPrint(HF_LOG_INFO, "Emotion result: %d", faceEmotionResult.emotion[index]);
+        /* We set the threshold of wearing a mask as 0.85. If it exceeds the threshold, it will be
+         * judged as wearing a mask. The threshold can be adjusted according to the scene */
+        if (maskConfidence.confidence[index] > 0.85) {
+            HFLogPrint(HF_LOG_INFO, "Mask");
+        } else {
+            HFLogPrint(HF_LOG_INFO, "Non Mask");
+        }
+    }
+
+    ret = HFReleaseImageStream(imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release image stream error: %d", ret);
+    }
+    /* The memory must be freed at the end of the program */
+    ret = HFReleaseInspireFaceSession(session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release session error: %d", ret);
+        return ret;
+    }
+
+    ret = HFReleaseImageBitmap(image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release image bitmap error: %d", ret);
+        return ret;
+    }
+
+    ret = HFReleaseImageBitmap(drawImage);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release draw image bitmap error: %d", ret);
+        return ret;
+    }
+
+    HFLogPrint(HF_LOG_INFO, "");
+    HFDeBugShowResourceStatistics();
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_face_track_benchmark.c b/cpp-package/inspireface/cpp/sample/api/sample_face_track_benchmark.c
new file mode 100644
index 000000000..dcab37c2b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_face_track_benchmark.c
@@ -0,0 +1,145 @@
+/*
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <inspireface.h>
+
+int main(int argc, char* argv[]) {
+    /* Check whether the number of parameters is correct */
+    if (argc < 3 || argc > 4) {
+        HFLogPrint(HF_LOG_ERROR, "Usage: %s <pack_path> <source_path> [rotation]", argv[0]);
+        return 1;
+    }
+
+    const char* packPath = argv[1];
+    const char* sourcePath = argv[2];
+    int rotation = 0;
+
+    /* If rotation is provided, check and set the value */
+    if (argc == 4) {
+        rotation = atoi(argv[3]);
+        if (rotation != 0 && rotation != 90 && rotation != 180 && rotation != 270) {
+            HFLogPrint(HF_LOG_ERROR, "Invalid rotation value. Allowed values are 0, 90, 180, 270.");
+            return 1;
+        }
+    }
+    
+    HFRotation rotation_enum;
+    /* Set rotation based on input parameter */
+    switch (rotation) {
+        case 90:
+            rotation_enum = HF_CAMERA_ROTATION_90;
+            break;
+        case 180:
+            rotation_enum = HF_CAMERA_ROTATION_180;
+            break;
+        case 270:
+            rotation_enum = HF_CAMERA_ROTATION_270;
+            break;
+        case 0:
+        default:
+            rotation_enum = HF_CAMERA_ROTATION_0;
+            break;
+    }
+
+    HFLogPrint(HF_LOG_INFO, "Pack file Path: %s", packPath);
+    HFLogPrint(HF_LOG_INFO, "Source file Path: %s", sourcePath);
+    HFLogPrint(HF_LOG_INFO, "Rotation: %d", rotation);
+
+    HFSetLogLevel(HF_LOG_INFO);
+
+    HResult ret;
+    /* The resource file must be loaded before it can be used */
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+
+    /* Enable the functions in the pipeline: mask detection, live detection, and face quality
+     * detection */
+    HOption option = HF_ENABLE_QUALITY | HF_ENABLE_MASK_DETECT | HF_ENABLE_LIVENESS;
+    /* Non-video or frame sequence mode uses IMAGE-MODE, which is always face detection without
+     * tracking */
+    HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    /* Maximum number of faces detected */
+    HInt32 maxDetectNum = 20;
+    /* Face detection image input level */
+    HInt32 detectPixelLevel = 160;
+    /* Handle of the current face SDK algorithm context */
+    HFSession session = {0};
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create FaceContext error: %d", ret);
+        return ret;
+    }
+
+    HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+    HFSessionSetFilterMinimumFacePixelSize(session, 4);
+
+    /* Load a image */
+    HFImageBitmap image;
+    ret = HFCreateImageBitmapFromFilePath(sourcePath, 3, &image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "The source entered is not a picture or read error.");
+        return ret;
+    }
+    /* Prepare an image parameter structure for configuration */
+    HFImageStream imageHandle = {0};
+    ret = HFCreateImageStreamFromImageBitmap(image, rotation_enum, &imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create ImageStream error: %d", ret);
+        return ret;
+    }
+
+    int loop = 100;
+
+    /* Enable the cost spend */
+    HFSessionSetEnableTrackCostSpend(session, 1);
+
+    int i;
+    /* Execute HF_FaceContextRunFaceTrack captures face information in an image */
+    HFMultipleFaceData multipleFaceData = {0};
+    for (i = 0; i < loop; i++) {
+        ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+        if (ret != HSUCCEED) {
+            HFLogPrint(HF_LOG_ERROR, "Execute HFExecuteFaceTrack error: %d", ret);
+            return ret;
+        }
+    }
+    HFLogPrint(HF_LOG_INFO, "Number of Detection: %d", multipleFaceData.detectedNum);
+    HFSessionPrintTrackCostSpend(session);
+
+    if (multipleFaceData.detectedNum > 0) {
+        HFLogPrint(HF_LOG_INFO, "========================================");
+        for (i = 0; i < multipleFaceData.detectedNum; i++) {
+            HFLogPrint(HF_LOG_INFO, "TrackId: %d", multipleFaceData.trackIds[i]);
+            HFLogPrint(HF_LOG_INFO, "TrackCount: %d", multipleFaceData.trackCounts[i]);
+        }
+    } else {
+        HFLogPrint(HF_LOG_WARN, "The face cannot be detected, and the tracking test results may be invalid!");
+    }
+
+    ret = HFReleaseImageStream(imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release image stream error: %d", ret);
+    }
+
+
+    ret = HFReleaseImageBitmap(image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release image bitmap error: %d", ret);
+        return ret;
+    }    
+    
+    /* The memory must be freed at the end of the program */
+    ret = HFReleaseInspireFaceSession(session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Release session error: %d", ret);
+        return ret;
+    }
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_feature_hub.c b/cpp-package/inspireface/cpp/sample/api/sample_feature_hub.c
new file mode 100644
index 000000000..76aa71845
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_feature_hub.c
@@ -0,0 +1,99 @@
+#include <inspireface.h>
+
+static float FT[] = {
+  0.0706566,   0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,
+  -0.00683422, -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,
+  0.0164787,   -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,
+  -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,
+  -0.00625798, 0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,
+  0.0465634,   0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,
+  -0.102007,   0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,
+  -0.00694243, 0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,
+  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,
+  0.0511028,   -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,
+  -0.0225608,  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,
+  0.056139,    -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,   0.0706566,   0.00640248,  0.0418103,   -0.00597861,
+  0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,   -0.00683422, -0.0338589,  0.0533989,   -0.0371725,
+  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,   0.0164787,   -0.00732871, -0.0458209,  -0.0100137,
+  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,
+  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,   -0.00625798, 0.00748682,  0.0533557,   0.0314002,
+  -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,   0.0627244,   0.0449248,   -0.037054,
+  -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,   -0.102007,   0.0649219,   0.0630833,   0.0421069,
+  0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,   -0.00694243, 0.0366775,   0.0672882,   0.0122419,
+  -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,
+  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,   0.0511028,   -0.0310699,  -0.0322141,  0.00996936,
+  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,  -0.0225608,  -0.0332198,  0.00102291,  0.108814,
+  -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,   0.056139,    -0.0166783,  -0.0803042,  -0.014245,
+  -0.0476374,  0.048495,    0.0378856,   0.0706566,   0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,
+  -0.0080779,  -0.0550556,  0.0229963,   -0.00683422, -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743,
+  -0.0128782,  0.0935529,   0.0588179,   0.0164787,   -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,
+  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,
+  0.00192449,  -0.0593105,  0.0191942,   -0.00625798, 0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,
+  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,   0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,
+  -0.0161973,  0.0319588,   0.0112792,   -0.102007,   0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,
+  0.071994,    -0.0272229,  0.0167103,   -0.00694243, 0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,
+  0.000983093, -0.00776073, -0.0268638,  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,
+  0.000502882, 0.0496892,   0.0126823,   0.0511028,   -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,
+  0.0419618,   -0.00358901, -0.0309569,  -0.0225608,  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,
+  0.0721224,   -0.0795082,  0.0340047,   0.056139,    -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,   0.0706566,
+  0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,   -0.00683422,
+  -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,   0.0164787,
+  -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,
+  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,   -0.00625798,
+  0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,
+  0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,   -0.102007,
+  0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,   -0.00694243,
+  0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,  0.00682446,
+  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,   0.0511028,
+  -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,  -0.0225608,
+  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,   0.056139,
+  -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,
+};
+
+int main() {
+    HResult ret;
+    // The resource file must be loaded before it can be used
+    ret = HFLaunchInspireFace("test_res/pack/Pikachu");
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+
+    HFFeatureHubConfiguration configuration;
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = "test.db";
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    ret = HFFeatureHubDataEnable(configuration);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Enable feature hub error: %d", ret);
+        return ret;
+    }
+
+    int64_t result_id = 0;
+    HFFaceFeature feature = {0};
+    feature.data = FT;
+    feature.size = sizeof(FT) / sizeof(FT[0]);
+    HFFaceFeatureIdentity identity = {0};
+    identity.feature = &feature;
+    ret = HFFeatureHubInsertFeature(identity, &result_id);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Insert feature error: %d", ret);
+        return ret;
+    }
+
+    HFFaceFeature query_feature = {0};
+    query_feature.data = FT;
+    query_feature.size = sizeof(FT) / sizeof(FT[0]);
+    HFloat confidence;
+    HFFaceFeatureIdentity search_result = {0};
+    ret = HFFeatureHubFaceSearch(query_feature, &confidence, &search_result);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Search feature error: %d", ret);
+    } else {
+        HFLogPrint(HF_LOG_INFO, "Search face feature success, result_id: %d", search_result.id);
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_feature_hub_persistence.c b/cpp-package/inspireface/cpp/sample/api/sample_feature_hub_persistence.c
new file mode 100644
index 000000000..ce3361cd9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_feature_hub_persistence.c
@@ -0,0 +1,52 @@
+#include <inspireface.h>
+#include <unistd.h>
+#include <stdio.h>
+
+int main(int argc, char* argv[]) {
+    if (argc != 2) {
+        HFLogPrint(HF_LOG_ERROR, "Usage: %s <pack_path>", argv[0]);
+        return -1;
+    }
+    
+    const char* packPath = argv[1];
+    HResult ret;
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+    const char* DBFilePath = "feature.db";
+    
+    // remove old db file
+    if (access(DBFilePath, F_OK) == 0) {
+        if (remove(DBFilePath) != 0) {
+            HFLogPrint(HF_LOG_ERROR, "Failed to remove old db file: %s", DBFilePath);
+            return -1;
+        }
+        HFLogPrint(HF_LOG_INFO, "Remove old db file: %s", DBFilePath);
+    }
+
+    HFFeatureHubConfiguration featureHubConfiguration;
+    featureHubConfiguration.primaryKeyMode = HF_PK_AUTO_INCREMENT; 
+    featureHubConfiguration.enablePersistence = 1;                 
+    featureHubConfiguration.persistenceDbPath = DBFilePath;      
+    featureHubConfiguration.searchMode = HF_SEARCH_MODE_EAGER;     
+    featureHubConfiguration.searchThreshold = 0.48f;             
+
+    ret = HFFeatureHubDataEnable(featureHubConfiguration);
+    if (ret != HSUCCEED)
+    {
+        HFLogPrint(HF_LOG_ERROR, "Enable FeatureHub failed: %d\n", ret);
+        return ret;
+    }
+    if (access(DBFilePath, F_OK) != 0) {
+        HFLogPrint(HF_LOG_ERROR, "DB file not found: %s", DBFilePath);
+        return -1;
+    }
+    HFLogPrint(HF_LOG_INFO, "DB file found: %s", DBFilePath);
+
+    // ....
+
+    HFTerminateInspireFace();
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_gray_input.c b/cpp-package/inspireface/cpp/sample/api/sample_gray_input.c
new file mode 100644
index 000000000..70ae63ef7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_gray_input.c
@@ -0,0 +1,113 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-06-29
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <inspireface.h>
+
+uint8_t* read_binary_file(const char* filename, size_t* file_size) {
+    FILE* file = fopen(filename, "rb");
+    if (!file) {
+        printf("Error: Cannot open file %s\n", filename);
+        return NULL;
+    }
+    
+    fseek(file, 0, SEEK_END);
+    *file_size = ftell(file);
+    fseek(file, 0, SEEK_SET);
+    
+    uint8_t* buffer = (uint8_t*)malloc(*file_size);
+    if (!buffer) {
+        printf("Error: Memory allocation failed\n");
+        fclose(file);
+        return NULL;
+    }
+    
+    size_t bytes_read = fread(buffer, 1, *file_size, file);
+    fclose(file);
+    
+    if (bytes_read != *file_size) {
+        printf("Error: File read incomplete. Expected %zu bytes, got %zu bytes\n", 
+               *file_size, bytes_read);
+        free(buffer);
+        return NULL;
+    }
+    
+    return buffer;
+}
+
+void free_binary_data(uint8_t* data) {
+    if (data) {
+        free(data);
+    }
+}
+
+int main() {
+    HResult ret;
+    ret = HFLaunchInspireFace("test_res/pack/Pikachu");
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+    HOption option = HF_ENABLE_FACE_RECOGNITION | HF_ENABLE_MASK_DETECT | HF_ENABLE_QUALITY;
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HInt32 maxDetectNum = 1;
+    HInt32 detectPixelLevel = 320;
+    HFSession session = {0};
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, maxDetectNum, detectPixelLevel, -1, &session);
+    if (ret != HSUCCEED)
+    {
+        HFLogPrint(HF_LOG_ERROR, "Create FaceContext error: %d", ret);
+        return ret;
+    }
+
+    // HFImageBitmap image;
+    HFImageStream imageHandle;
+    /* Load a image */
+    // ret = HFCreateImageBitmapFromFilePath("/Users/tunm/Downloads/Desktop/outputbw.jpg", 1, &image);
+    // if (ret != HSUCCEED) {
+    //     HFLogPrint(HF_LOG_ERROR, "The source entered is not a picture or read error.");
+    //     return ret;
+    // }
+    // /* Prepare an image parameter structure for configuration */
+    // ret = HFCreateImageStreamFromImageBitmap(image, HF_CAMERA_ROTATION_0, &imageHandle);
+    // if (ret != HSUCCEED) {
+    //     HFLogPrint(HF_LOG_ERROR, "Create ImageStream error: %d", ret);
+    //     return ret;
+    // }
+
+
+    size_t file_size;
+    uint8_t* buffer = read_binary_file("/Users/tunm/Downloads/Desktop/outputbw.byte", &file_size);
+    if (buffer == NULL) {
+        HFLogPrint(HF_LOG_ERROR, "Read file error.");
+        return -1;
+    }
+    HFImageData imageData;
+    imageData.data = buffer;
+    imageData.width = 640;
+    imageData.height = 480;
+    imageData.format = HF_STREAM_GRAY;
+    imageData.rotation = HF_CAMERA_ROTATION_0;
+
+    ret = HFCreateImageStream(&imageData, &imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create ImageStream error: %d", ret);
+        return ret;
+    }
+
+    HFDeBugImageStreamDecodeSave(imageHandle, "2.jpg");
+
+    HFMultipleFaceData multipleFaceData;
+    ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Detect Face error: %d", ret);
+        return ret;
+    }
+    HFLogPrint(HF_LOG_INFO, "Face num: %d", multipleFaceData.detectedNum);
+
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_load_reload.c b/cpp-package/inspireface/cpp/sample/api/sample_load_reload.c
new file mode 100644
index 000000000..607170f41
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_load_reload.c
@@ -0,0 +1,19 @@
+#include <inspireface.h>
+
+int main() {
+    const char* resourcePath = "test_res/pack/Pikachu";
+    HResult ret = HFReloadInspireFace(resourcePath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Failed to launch InspireFace: %d", ret);
+        return 1;
+    }
+
+    // Switch to another resource
+    ret = HFReloadInspireFace("test_res/pack/Megatron");
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Failed to reload InspireFace: %d", ret);
+        return 1;
+    }
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/api/sample_search_face.cpp b/cpp-package/inspireface/cpp/sample/api/sample_search_face.cpp
new file mode 100644
index 000000000..c10bd5f57
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/api/sample_search_face.cpp
@@ -0,0 +1,130 @@
+#include <inspireface.h>
+#include <unistd.h>
+#include <stdio.h>
+
+
+int main(int argc, char* argv[]) {
+    if (argc != 4) {
+        HFLogPrint(HF_LOG_ERROR, "Usage: %s <pack_path> <feature_path> <image_path>", argv[0]);
+        return -1;
+    }
+    
+    char* packPath = argv[1];
+    char* featurePath = argv[2];
+    char* imagePath = argv[3];
+    
+    HResult ret;
+    ret = HFLaunchInspireFace(packPath);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Load Resource error: %d", ret);
+        return ret;
+    }
+    
+    // 1. Enable feature hub
+    HFFeatureHubConfiguration featureHubConfiguration;
+    featureHubConfiguration.primaryKeyMode = HF_PK_AUTO_INCREMENT; 
+    featureHubConfiguration.enablePersistence = 1;
+    featureHubConfiguration.persistenceDbPath = featurePath;
+    featureHubConfiguration.searchThreshold = 0.32f;
+    featureHubConfiguration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    ret = HFFeatureHubDataEnable(featureHubConfiguration);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Enable feature hub error: %d", ret);
+        return ret;
+    }
+
+    // 2. Create session
+    HFSession session;
+    ret = HFCreateInspireFaceSessionOptional(HF_ENABLE_FACE_RECOGNITION, HF_DETECT_MODE_ALWAYS_DETECT, 1, 320, -1, &session);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create session error: %d", ret);
+        return ret;
+    }
+    
+    // 3. Load image
+    HFImageBitmap image;
+    ret = HFCreateImageBitmapFromFilePath(imagePath, 3, &image);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image bitmap error: %d", ret);
+        HFReleaseInspireFaceSession(session);
+        return ret;
+    }
+    
+    // 4. Create image stream
+    HFImageStream imageHandle;
+    ret = HFCreateImageStreamFromImageBitmap(image, HF_CAMERA_ROTATION_0, &imageHandle);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create image stream error: %d", ret);
+        HFReleaseImageBitmap(image);
+        HFReleaseInspireFaceSession(session);
+        return ret;
+    }
+
+    // 5. Detect face
+    HFMultipleFaceData multipleFaceData;
+    ret = HFExecuteFaceTrack(session, imageHandle, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Execute face track error: %d", ret);
+        HFReleaseImageStream(imageHandle);
+        HFReleaseImageBitmap(image);
+        HFReleaseInspireFaceSession(session);
+        return ret;
+    }
+
+    if (multipleFaceData.detectedNum == 0) {
+        HFLogPrint(HF_LOG_WARN, "No face detected in image: %s", imagePath);
+        HFReleaseImageStream(imageHandle);
+        HFReleaseImageBitmap(image);
+        HFReleaseInspireFaceSession(session);
+        return -1;
+    }
+
+    HFLogPrint(HF_LOG_INFO, "Face detected: %d", multipleFaceData.detectedNum);
+
+    // 6. Extract feature
+    HFFaceFeature feature;
+    ret = HFCreateFaceFeature(&feature);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Create face feature error: %d", ret);
+        HFReleaseImageStream(imageHandle);
+        HFReleaseImageBitmap(image);
+        HFReleaseInspireFaceSession(session);
+        return ret;
+    }
+
+    ret = HFFaceFeatureExtractTo(session, imageHandle, multipleFaceData.tokens[0], feature);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Extract feature error: %d", ret);
+        HFReleaseFaceFeature(&feature);
+        HFReleaseImageStream(imageHandle);
+        HFReleaseImageBitmap(image);
+        HFReleaseInspireFaceSession(session);
+        return ret;
+    }
+
+    // 7. Search face
+    HFloat confidence;
+    HFFaceFeatureIdentity searchResult = {0};
+    ret = HFFeatureHubFaceSearch(feature, &confidence, &searchResult);
+    if (ret != HSUCCEED) {
+        HFLogPrint(HF_LOG_ERROR, "Search face feature error: %d", ret);
+    } else {
+        if (searchResult.id != -1) {
+            HFLogPrint(HF_LOG_INFO, "Search face feature success!");
+            HFLogPrint(HF_LOG_INFO, "Found matching face ID: %d", searchResult.id);
+            HFLogPrint(HF_LOG_INFO, "Confidence score: %.4f", confidence);
+        } else {
+            HFLogPrint(HF_LOG_INFO, "No matching face found in database");
+        }
+    }
+
+    // 8. Clean up
+    HFReleaseFaceFeature(&feature);
+    HFReleaseImageStream(imageHandle);
+    HFReleaseImageBitmap(image);
+    HFReleaseInspireFaceSession(session);
+
+    HFDeBugShowResourceStatistics();
+
+    return ret;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/debug_rk_rec.cpp b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/debug_rk_rec.cpp
new file mode 100644
index 000000000..ff761f62f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/debug_rk_rec.cpp
@@ -0,0 +1,109 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "opencv2/opencv.hpp"
+
+#include "inspireface/middleware/costman.h"
+#include "middleware/inference_wrapper/customized/rknn_adapter.h"
+#include "inspireface/feature_hub/simd.h"
+#include <memory>
+#include "inspireface/recognition_module/extract/extract.h"
+#include "middleware/model_archive/inspire_archive.h"
+
+using namespace inspire;
+
+int main() {
+    std::vector<std::string> names = {
+      "test_res/images/test_data/0.jpg",
+      "test_res/images/test_data/1.jpg",
+      "test_res/images/test_data/2.jpg",
+    };
+    InspireArchive loader("test_res/pack/test_zip_rec");
+    {
+        InspireModel model;
+        loader.LoadModel("feature", model);
+        auto net = std::make_shared<RKNNAdapter>();
+        net->Initialize((unsigned char *)model.buffer, model.bufferSize);
+        net->setOutputsWantFloat(1);
+
+        EmbeddedList list;
+        for (int i = 0; i < names.size(); ++i) {
+            cv::Mat image = cv::imread(names[i]);
+            cv::Mat rgb;
+            cv::cvtColor(image, rgb, cv::COLOR_BGR2RGB);
+            net->SetInputData(0, rgb);
+            net->RunModel();
+
+            auto out = net->GetOutputData(0);
+            auto dims = net->GetOutputTensorSize(0);
+            //        for (int i = 0; i < dims.size(); ++i) {
+            //            LOGD("%lu", dims[i]);
+            //        }
+            //
+            for (int i = 0; i < 512; ++i) {
+                std::cout << out[i] << ", ";
+            }
+            std::cout << std::endl;
+
+            Embedded emb;
+            for (int j = 0; j < 512; ++j) {
+                emb.push_back(out[j]);
+            }
+            list.push_back(emb);
+        }
+
+        for (int i = 0; i < list.size(); ++i) {
+            auto &embedded = list[i];
+            float mse = 0.0f;
+            for (const auto &one : embedded) {
+                mse += one * one;
+            }
+            mse = sqrt(mse);
+            for (float &one : embedded) {
+                one /= mse;
+            }
+        }
+
+        auto cos = simd_dot(list[0].data(), list[1].data(), 512);
+        LOGD("COS: %f", cos);
+    }
+
+    {
+        std::shared_ptr<Extract> m_extract_;
+
+        Configurable param;
+        param.set<int>("model_index", 0);
+        param.set<std::string>("input_layer", "input");
+        param.set<std::vector<std::string>>("outputs_layers", {
+                                                                "267",
+                                                              });
+        param.set<std::vector<int>>("input_size", {112, 112});
+        param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+        param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+        param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+        param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+        param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+        param.set<bool>("nchw", false);
+        param.set<bool>("swap_color", true);  // RK requires rgb input
+
+        m_extract_ = std::make_shared<Extract>();
+        InspireModel model;
+        loader.LoadModel("feature", model);
+        m_extract_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+        cv::Mat image = cv::imread(names[0]);
+        //        cv::Mat rgb;
+        //        cv::cvtColor(image, rgb, cv::COLOR_BGR2RGB);
+        auto feat = m_extract_->GetFaceFeature(image);
+        for (int i = 0; i < 512; ++i) {
+            std::cout << feat[i] << ", ";
+        }
+        std::cout << std::endl;
+    }
+
+    LOGD("End");
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_det_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_det_sample.cpp
new file mode 100644
index 000000000..c06decf17
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_det_sample.cpp
@@ -0,0 +1,51 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include "opencv2/opencv.hpp"
+// #include "inspireface/middleware/model_loader/model_loader.h"
+#include "inspireface/track_module/face_detect/all.h"
+
+#include "inspireface/middleware/costman.h"
+#include "middleware/model_archive/inspire_archive.h"
+#include "log.h"
+
+using namespace inspire;
+
+int main() {
+    auto detModel = "test_res/pack/Gundam_RV1109";
+    InspireArchive inspireArchive;
+    auto ret = inspireArchive.ReLoad(detModel);
+    if (ret != SARC_SUCCESS) {
+        LOGE("Error load");
+        return ret;
+    }
+    InspireModel model;
+    ret = inspireArchive.LoadModel("face_detect", model);
+    if (ret != SARC_SUCCESS) {
+        LOGE("Error model");
+        return ret;
+    }
+
+    std::cout << model.Config().toString() << std::endl;
+
+    std::shared_ptr<FaceDetect> m_face_detector_;
+    m_face_detector_ = std::make_shared<FaceDetect>(320);
+    m_face_detector_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    // Load a image
+    cv::Mat image = cv::imread("test_res/images/face_sample.png");
+
+    Timer timer;
+    FaceLocList locs = (*m_face_detector_)(image);
+    LOGD("cost: %f", timer.GetCostTimeUpdate());
+
+    LOGD("Faces: %ld", locs.size());
+
+    for (auto &loc : locs) {
+        cv::rectangle(image, cv::Point2f(loc.x1, loc.y1), cv::Point2f(loc.x2, loc.y2), cv::Scalar(0, 0, 255), 3);
+    }
+    cv::imwrite("det.jpg", image);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_recognize_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_recognize_sample.cpp
new file mode 100644
index 000000000..8276f849b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_face_recognize_sample.cpp
@@ -0,0 +1,85 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "opencv2/opencv.hpp"
+
+#include "inspireface/middleware/costman.h"
+#include "inspireface/feature_hub/face_recognition.h"
+#include "inspireface/feature_hub/simd.h"
+#include "inspireface/middleware/model_archive/inspire_archive.h"
+
+using namespace inspire;
+
+std::shared_ptr<InspireArchive> loader;
+
+void rec_function() {
+    std::shared_ptr<Extract> m_extract_;
+
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_03_extract);
+    param.set<std::string>("input_layer", "input");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "267",
+                                                          });
+    param.set<std::vector<int>>("input_size", {112, 112});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+    param.set<bool>("swap_color", true);  // RK requires rgb input
+
+    m_extract_ = std::make_shared<Extract>();
+    InspireModel model;
+    loader->LoadModel("feature", model);
+    m_extract_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    loader.reset();
+
+    std::vector<std::string> files = {
+      "test_res/images/test_data/0.jpg",
+      "test_res/images/test_data/1.jpg",
+      "test_res/images/test_data/2.jpg",
+    };
+    EmbeddedList embedded_list;
+    for (int i = 0; i < files.size(); ++i) {
+        auto warped = cv::imread(files[i]);
+        Timer timer;
+        auto emb = (*m_extract_)(warped);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+        embedded_list.push_back(emb);
+        LOGD("%lu", emb.size());
+    }
+
+    float _0v1;
+    float _0v2;
+    float _1v2;
+    FaceRecognition::CosineSimilarity(embedded_list[0], embedded_list[1], _0v1);
+    FaceRecognition::CosineSimilarity(embedded_list[0], embedded_list[2], _0v2);
+    FaceRecognition::CosineSimilarity(embedded_list[1], embedded_list[2], _1v2);
+    LOGD("0 vs 1 : %f", _0v1);
+    LOGD("0 vs 2 : %f", _0v2);
+    LOGD("1 vs 2 : %f", _1v2);
+
+    //    LOGD("size: %lu", embedded_list.size());
+    //    LOGD("num of vector: %lu", embedded_list[2].size());
+    //
+    //    float _0v1 = simd_dot(embedded_list[0].data(), embedded_list[1].data(), 512);
+    //    float _0v2 = simd_dot(embedded_list[0].data(), embedded_list[2].data(), 512);
+    //    float _1v2 = simd_dot(embedded_list[1].data(), embedded_list[2].data(), 512);
+    //    LOGD("0 vs 1 : %f", _0v1);
+    //    LOGD("0 vs 2 : %f", _0v2);
+    //    LOGD("1 vs 2 : %f", _1v2);
+}
+
+int main() {
+    loader = std::make_shared<InspireArchive>();
+    loader->ReLoad("test_res/pack/Gundam_RV1109");
+
+    rec_function();
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_simple_net_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_simple_net_sample.cpp
new file mode 100644
index 000000000..fae87aa9b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_simple_net_sample.cpp
@@ -0,0 +1,303 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "opencv2/opencv.hpp"
+#include "inspireface/track_module/face_detect/all.h"
+#include "inspireface/pipeline_module/attribute/mask_predict.h"
+
+#include "inspireface/middleware/costman.h"
+#include "inspireface/track_module/quality/face_pose_quality.h"
+#include "inspireface/track_module/landmark/face_landmark.h"
+#include "inspireface/pipeline_module/liveness/rgb_anti_spoofing.h"
+#include "inspireface/face_context.h"
+
+using namespace inspire;
+
+InspireArchive loader;
+
+void test_rnet() {
+    std::shared_ptr<RNet> m_rnet_;
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_04_refine_net);
+    param.set<std::string>("input_layer", "input_1");
+    param.set<std::vector<std::string>>("outputs_layers", {"conv5-1/Softmax", "conv5-2/BiasAdd"});
+    param.set<std::vector<int>>("input_size", {24, 24});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<bool>("swap_color", true);  // RGB mode
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+
+    InspireModel model;
+    loader.LoadModel("refine_net", model);
+    m_rnet_ = std::make_shared<RNet>();
+    m_rnet_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    {
+        // Load a image
+        cv::Mat image = cv::imread("test_res/images/test_data/hasface.jpg");
+
+        Timer timer;
+        auto score = (*m_rnet_)(image);
+        LOGD("RNETcost: %f", timer.GetCostTimeUpdate());
+        LOGD("has face: %f", score);
+    }
+
+    {
+        // Load a image
+        cv::Mat image = cv::imread("test_res/images/test_data/noface.jpg");
+
+        Timer timer;
+        auto score = (*m_rnet_)(image);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+        LOGD("non face: %f", score);
+    }
+}
+
+void test_mask() {
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_05_mask);
+    param.set<std::string>("input_layer", "input_1");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "activation_1/Softmax",
+                                                          });
+    param.set<std::vector<int>>("input_size", {96, 96});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<bool>("swap_color", true);  // RGB mode
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+
+    std::shared_ptr<MaskPredict> m_mask_predict_;
+    m_mask_predict_ = std::make_shared<MaskPredict>();
+    InspireModel model;
+    loader.LoadModel("mask_detect", model);
+    m_mask_predict_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    {
+        // Load a image
+        cv::Mat image = cv::imread("test_res/images/test_data/mask.jpg");
+
+        Timer timer;
+        auto score = (*m_mask_predict_)(image);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+        LOGD("mask: %f", score);
+    }
+
+    {
+        // Load a image
+        cv::Mat image = cv::imread("test_res/images/test_data/nomask.jpg");
+
+        Timer timer;
+        auto score = (*m_mask_predict_)(image);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+        LOGD("maskless: %f", score);
+    }
+}
+
+void test_quality() {
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_07_pose_q_fp16);
+    param.set<std::string>("input_layer", "data");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "fc1",
+                                                          });
+    param.set<std::vector<int>>("input_size", {96, 96});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<bool>("swap_color", true);  // RGB mode
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+    std::shared_ptr<FacePoseQuality> m_face_quality_;
+    m_face_quality_ = std::make_shared<FacePoseQuality>();
+    InspireModel model;
+    loader.LoadModel("pose_quality", model);
+    m_face_quality_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    {
+        std::vector<std::string> names = {
+          "test_res/images/test_data/p3.jpg",
+          //                "test_res/images/test_data/p1.jpg",
+        };
+        for (int i = 0; i < names.size(); ++i) {
+            LOGD("Image: %s", names[i].c_str());
+            cv::Mat image = cv::imread(names[i]);
+
+            Timer timer;
+            auto pose_res = (*m_face_quality_)(image);
+            LOGD("quality cost: %f", timer.GetCostTimeUpdate());
+
+            for (auto &p : pose_res.lmk) {
+                cv::circle(image, p, 0, cv::Scalar(0, 0, 255), 2);
+            }
+            cv::imwrite("pose.jpg", image);
+            LOGD("pitch: %f", pose_res.pitch);
+            LOGD("yam: %f", pose_res.yaw);
+            LOGD("roll: %f", pose_res.roll);
+
+            for (auto q : pose_res.lmk_quality) {
+                std::cout << q << ", ";
+            }
+            std::cout << std::endl;
+        }
+    }
+}
+
+void test_landmark_mnn() {
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_01_lmk);
+    param.set<std::string>("input_layer", "input_1");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "prelu1/add",
+                                                          });
+    param.set<std::vector<int>>("input_size", {112, 112});
+    param.set<std::vector<float>>("mean", {127.5f, 127.5f, 127.5f});
+    param.set<std::vector<float>>("norm", {0.0078125f, 0.0078125f, 0.0078125f});
+
+    std::shared_ptr<FaceLandmark> m_landmark_predictor_;
+    m_landmark_predictor_ = std::make_shared<FaceLandmark>(112);
+    InspireModel model;
+    loader.LoadModel("landmark", model);
+    m_landmark_predictor_->LoadData(model);
+
+    cv::Mat image = cv::imread("test_res/images/test_data/crop.png");
+    cv::resize(image, image, cv::Size(112, 112));
+
+    std::vector<float> lmk;
+    Timer timer;
+    for (int i = 0; i < 50; ++i) {
+        lmk = (*m_landmark_predictor_)(image);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+    }
+
+    for (int i = 0; i < FaceLandmark::NUM_OF_LANDMARK; ++i) {
+        float x = lmk[i * 2 + 0] * 112;
+        float y = lmk[i * 2 + 1] * 112;
+        cv::circle(image, cv::Point2f(x, y), 0, cv::Scalar(0, 0, 255), 1);
+    }
+
+    cv::imwrite("lmk.jpg", image);
+}
+
+void test_landmark() {
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_01_lmk);
+    param.set<std::string>("input_layer", "input_1");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "prelu1/add",
+                                                          });
+    param.set<std::vector<int>>("input_size", {112, 112});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+
+    std::shared_ptr<FaceLandmark> m_landmark_predictor_;
+    m_landmark_predictor_ = std::make_shared<FaceLandmark>(112);
+    InspireModel model;
+    loader.LoadModel("landmark", model);
+    m_landmark_predictor_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    cv::Mat image = cv::imread("test_res/images/test_data/0.jpg");
+    cv::resize(image, image, cv::Size(112, 112));
+
+    std::vector<float> lmk;
+    Timer timer;
+    for (int i = 0; i < 50; ++i) {
+        lmk = (*m_landmark_predictor_)(image);
+        LOGD("LMKcost: %f", timer.GetCostTimeUpdate());
+    }
+
+    for (int i = 0; i < FaceLandmark::NUM_OF_LANDMARK; ++i) {
+        float x = lmk[i * 2 + 0] * 112;
+        float y = lmk[i * 2 + 1] * 112;
+        cv::circle(image, cv::Point2f(x, y), 0, cv::Scalar(0, 0, 255), 1);
+    }
+
+    cv::imwrite("lmk.jpg", image);
+}
+
+void test_liveness() {
+    Configurable param;
+    //    param.set<int>("model_index", ModelIndex::_06_msafa27);
+    param.set<std::string>("input_layer", "data");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "556",
+                                                          });
+    param.set<std::vector<int>>("input_size", {80, 80});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<bool>("swap_color", false);  // RGB mode
+    param.set<int>("data_type", InputTensorInfo::DataTypeImage);
+    param.set<int>("input_tensor_type", InputTensorInfo::TensorTypeUint8);
+    param.set<int>("output_tensor_type", InputTensorInfo::TensorTypeFp32);
+    param.set<bool>("nchw", false);
+
+    std::shared_ptr<RBGAntiSpoofing> m_rgb_anti_spoofing_;
+
+    InspireModel model;
+    loader.LoadModel("rgb_anti_spoofing", model);
+    m_rgb_anti_spoofing_ = std::make_shared<RBGAntiSpoofing>(80, true);
+    m_rgb_anti_spoofing_->LoadData(model, InferenceWrapper::INFER_RKNN);
+
+    std::vector<std::string> names = {
+      "test_res/images/test_data/real.jpg", "test_res/images/test_data/fake.jpg", "test_res/images/test_data/live.jpg",
+      "test_res/images/test_data/ttt.jpg",  "test_res/images/test_data/w.jpg",    "test_res/images/test_data/w2.jpg",
+    };
+
+    for (int i = 0; i < names.size(); ++i) {
+        auto image = cv::imread(names[i]);
+        Timer timer;
+        auto score = (*m_rgb_anti_spoofing_)(image);
+        LOGD("cost: %f", timer.GetCostTimeUpdate());
+        LOGD("%s : %f", names[i].c_str(), score);
+    }
+}
+
+int test_liveness_ctx() {
+    CustomPipelineParameter parameter;
+    parameter.enable_liveness = true;
+    FaceContext ctx;
+    ctx.Configuration("test_res/pack/Gundam_RV1109", inspire::DETECT_MODE_IMAGE, 3, parameter);
+    std::vector<std::string> names = {
+      "test_res/images/test_data/real.jpg", "test_res/images/test_data/fake.jpg", "test_res/images/test_data/live.jpg",
+      "test_res/images/test_data/ttt.jpg",  "test_res/images/test_data/w.jpg",    "test_res/images/test_data/w2.jpg",
+      "test_res/images/test_data/bb.png",
+    };
+
+    for (int i = 0; i < names.size(); ++i) {
+        auto image = cv::imread(names[i]);
+        auto score = (*ctx.FacePipelineModule()->getMRgbAntiSpoofing())(image);
+        LOGD("%s : %f", names[i].c_str(), score);
+    }
+
+    return 0;
+}
+
+int main() {
+    loader.ReLoad("test_res/pack/Gundam_RV1109");
+
+    //    test_rnet();
+
+    //    test_mask();
+
+    //    test_quality();
+
+    //    test_landmark_mnn();
+
+    //    test_landmark();
+
+    test_liveness();
+    test_liveness_ctx();
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_tracker_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_tracker_sample.cpp
new file mode 100644
index 000000000..9f88d4a74
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/rk_sample/rk_tracker_sample.cpp
@@ -0,0 +1,72 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "opencv2/opencv.hpp"
+
+#include "inspireface/middleware/costman.h"
+#include "inspireface/face_context.h"
+
+using namespace inspire;
+
+int main() {
+    FaceContext ctx;
+    CustomPipelineParameter param;
+    int32_t ret = ctx.Configuration("test_res/pack/Gundam_RV1109", DetectMode::DETECT_MODE_VIDEO, 3, param);
+    if (ret != HSUCCEED) {
+        LOGE("Initiate error");
+    }
+    cv::Mat frame;
+    std::string imageFolder = "test_res/video_frames/";
+
+    //    auto video_frame_num = 10;
+    auto video_frame_num = 288;
+    for (int i = 0; i < video_frame_num; ++i) {
+        auto index = i + 1;
+        std::stringstream frameFileName;
+        frameFileName << imageFolder << "frame-" << std::setw(4) << std::setfill('0') << index << ".jpg";
+
+        frame = cv::imread(frameFileName.str());
+
+        CameraStream stream;
+        stream.SetRotationMode(ROTATION_0);
+        stream.SetDataFormat(BGR);
+        stream.SetDataBuffer(frame.data, frame.rows, frame.cols);
+
+        Timer timer;
+        ctx.FaceDetectAndTrack(stream);
+        LOGD("Cost: %f", timer.GetCostTimeUpdate());
+        LOGD("faces: %d", ctx.GetNumberOfFacesCurrentlyDetected());
+
+        LOGD("track id: %d", ctx.GetTrackingFaceList()[0].GetTrackingId());
+
+        auto &face = ctx.GetTrackingFaceList()[0];
+        for (auto &p : face.landmark_) {
+            cv::circle(frame, p, 0, cv::Scalar(0, 0, 255), 3);
+        }
+
+        auto rect = face.GetRect();
+        int track_id = face.GetTrackingId();
+        int track_count = face.GetTrackingCount();
+
+        cv::rectangle(frame, rect, cv::Scalar(0, 0, 255), 2, 1);
+
+        std::string text =
+          "ID: " + std::to_string(track_id) + " Count: " + std::to_string(track_count) + " Cf: " + std::to_string(face.GetConfidence());
+
+        cv::Point text_position(rect.x, rect.y - 10);
+        int font_face = cv::FONT_HERSHEY_SIMPLEX;
+        double font_scale = 0.5;
+        int font_thickness = 1;
+        cv::Scalar font_color(255, 255, 255);
+
+        cv::putText(frame, text, text_position, font_face, font_scale, font_color, font_thickness);
+
+        std::stringstream saveFile;
+        saveFile << "track_frames/" << "result-" << std::setw(4) << std::setfill('0') << index << ".jpg";
+        cv::imwrite(saveFile.str(), frame);
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_test.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_test.cpp
new file mode 100644
index 000000000..2ba6f899f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_test.cpp
@@ -0,0 +1,38 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "track_module/face_track.h"
+#include "inspireface/feature_hub/face_recognition.h"
+#include "log.h"
+#include "track_module/face_track.h"
+#include "pipeline_module/face_pipeline.h"
+#include "inspireface/feature_hub/face_recognition.h"
+#include "middleware/inference_wrapper/customized/rknn_adapter.h"
+
+using namespace inspire;
+
+int main() {
+    InspireArchive archive;
+    auto ret = archive.ReLoad("test_res/pack/Gundam_RV1109");
+    LOGD("ReLoad %d", ret);
+    //    InspireModel model;
+    //    ret = archive.LoadModel("mask_detect", model);
+    //    LOGD("LoadModel %d", ret);
+
+    FaceTrack track;
+    ret = track.Configuration(archive);
+    LOGD("Configuration %d", ret);
+
+    FacePipeline pipeline(archive, true, true, true, true, true);
+
+    FaceRecognition recognition(archive, true);
+
+    //    std::shared_ptr<RKNNAdapter> rknet = std::make_shared<RKNNAdapter>();
+    //    ret = rknet->Initialize((unsigned char* )model.buffer, model.bufferSize);
+    //
+    //    LOGD("LoadModel %d", ret);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_tracker.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_tracker.cpp
new file mode 100644
index 000000000..51f793823
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/archive_tracker.cpp
@@ -0,0 +1,48 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "track_module/face_track.h"
+#include "inspireface/recognition_module/face_feature_extraction.h"
+#include "log.h"
+
+using namespace inspire;
+
+int main() {
+    InspireArchive archive;
+    archive.ReLoad("test_res/pack/Gundam_RV1109");
+
+    FaceTrack track;
+    //    FaceRecognition recognition(archive, true);
+
+    auto ret = track.Configuration(archive);
+    INSPIRE_LOGD("ret=%d", ret);
+    if (ret != 0) {
+        return -1;
+    }
+
+    auto image = cv::imread("test_res/data/bulk/kun.jpg");
+    CameraStream stream;
+    stream.SetDataBuffer(image.data, image.rows, image.cols);
+    stream.SetDataFormat(BGR);
+    stream.SetRotationMode(ROTATION_0);
+
+    track.UpdateStream(stream, true);
+
+    //    if (!track.trackingFace.empty()) {
+    //        auto const &face = track.trackingFace[0];
+    //        cv::rectangle(image, face.GetRectSquare(), cv::Scalar(200, 0, 20), 2);
+    //    }
+    //
+    //    cv::imshow("w", image);
+    //    cv::waitKey(0);
+
+    InspireModel model;
+    ret = archive.LoadModel("mask_detect", model);
+    std::cout << ret << std::endl;
+
+    archive.PublicPrintSubFiles();
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/bm_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/bm_sample.cpp
new file mode 100644
index 000000000..92bb06f3a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/bm_sample.cpp
@@ -0,0 +1,82 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "opencv2/opencv.hpp"
+#include "log.h"
+#include "inspireface/feature_hub/simd.h"
+// #include <Eigen/Dense>
+
+using namespace inspire;
+
+int main() {
+    int N = 512;
+    int vectorSize = 512;  // Vector length
+    {
+        // Create an Nx512 matrix of type CV_32F and fill it with random numbers
+        cv::Mat mat(N, vectorSize, CV_32F);
+        cv::randu(mat, cv::Scalar(0), cv::Scalar(1));
+
+        // Create a 512x1 CV_32F matrix and fill it with random numbers
+        cv::Mat one(vectorSize, 1, CV_32F);
+        cv::randu(one, cv::Scalar(0), cv::Scalar(1));
+
+        std::cout << mat.size << std::endl;
+        std::cout << one.size << std::endl;
+
+        auto timeStart = (double)cv::getTickCount();
+
+        cv::Mat cosineSimilarities;
+        cv::gemm(mat, one, 1, cv::Mat(), 0, cosineSimilarities);
+
+        double cost = ((double)cv::getTickCount() - timeStart) / cv::getTickFrequency() * 1000;
+        INSPIRE_LOGD("Matrix COST: %f", cost);
+    }
+
+    {
+        std::srand(static_cast<unsigned int>(std::time(nullptr)));
+
+        std::vector<std::vector<float>> matrix(N, std::vector<float>(vectorSize));
+        for (int i = 0; i < N; ++i) {
+            for (int j = 0; j < vectorSize; ++j) {
+                matrix[i][j] = static_cast<float>(std::rand()) / RAND_MAX;
+            }
+        }
+
+        std::vector<float> vectorOne(vectorSize);
+        for (int i = 0; i < vectorSize; ++i) {
+            vectorOne[i] = static_cast<float>(std::rand()) / RAND_MAX;
+        }
+
+        auto timeStart = (double)cv::getTickCount();
+        // dot
+        for (const auto &v : matrix) {
+            simd_dot(v.data(), vectorOne.data(), vectorSize);
+        }
+
+        double cost = ((double)cv::getTickCount() - timeStart) / cv::getTickFrequency() * 1000;
+        INSPIRE_LOGD("Vector COST: %f", cost);
+    }
+
+    //    {
+    //        Eigen::initParallel();
+    //        Eigen::MatrixXf mat(N, vectorSize);
+    //        mat = Eigen::MatrixXf::Random(N, vectorSize);
+    //
+    //        std::cout << mat.rows() << " x " << mat.cols() << std::endl;
+    //
+    //
+    //        Eigen::VectorXf one(vectorSize);
+    //        one = Eigen::VectorXf::Random(vectorSize);
+    //
+    //        auto timeStart = (double) cv::getTickCount();
+    //        Eigen::VectorXf result = mat * one;
+    //
+    //        double cost = ((double) cv::getTickCount() - timeStart) / cv::getTickFrequency() * 1000;
+    //        LOGD("Eigen COST: %f", cost);
+    //    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/c_api_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/c_api_sample.cpp
new file mode 100644
index 000000000..cd038208c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/c_api_sample.cpp
@@ -0,0 +1,390 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "inspireface/c_api/inspireface.h"
+#include "opencv2/opencv.hpp"
+#include "inspireface/log.h"
+
+using namespace inspire;
+
+std::string basename(const std::string &path) {
+    size_t lastSlash = path.find_last_of("/\\");  // Take into account the cross-platform separator
+    if (lastSlash == std::string::npos) {
+        return path;  // Without the slash, the entire path is the base name
+    } else {
+        return path.substr(lastSlash + 1);  // Returns the part after the last slash
+    }
+}
+
+int compare() {
+    HResult ret;
+    // Initialize context
+#ifdef ISF_ENABLE_RKNN
+    HPath path = "test_res/pack/Gundam_RV1109";
+#else
+    HPath path = "test_res/pack/Pikachu";
+#endif
+    HF_ContextCustomParameter parameter = {0};
+    parameter.enable_liveness = 1;
+    parameter.enable_mask_detect = 1;
+    parameter.enable_recognition = 1;
+    parameter.enable_face_quality = 1;
+    HF_DetectMode detMode = HF_DETECT_MODE_IMAGE;  // Selecting the image mode is always detection
+    HContextHandle session;
+    ret = HF_CreateFaceContextFromResourceFile(path, parameter, detMode, 3, &session);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGD("An error occurred while creating ctx: %ld", ret);
+    }
+
+    std::vector<std::string> names = {
+      "/Users/tunm/datasets/lfw_funneled/Abel_Pacheco/Abel_Pacheco_0001.jpg",
+      "/Users/tunm/datasets/lfw_funneled/Abel_Pacheco/Abel_Pacheco_0004.jpg",
+    };
+    HInt32 featureNum;
+    HF_GetFeatureLength(&featureNum);
+    INSPIRE_LOGD("Feature length: %d", featureNum);
+    HFloat featuresCache[names.size()][featureNum];  // Store the cached vector
+
+    for (int i = 0; i < names.size(); ++i) {
+        auto &name = names[i];
+        cv::Mat image = cv::imread(name);
+        if (image.empty()) {
+            INSPIRE_LOGD("%s is empty!", name.c_str());
+            return -1;
+        }
+        HF_ImageData imageData = {0};
+        imageData.data = image.data;
+        imageData.height = image.rows;
+        imageData.width = image.cols;
+        imageData.rotation = CAMERA_ROTATION_0;
+        imageData.format = STREAM_BGR;
+
+        HImageHandle imageSteamHandle;
+        ret = HF_CreateImageStream(&imageData, &imageSteamHandle);
+        if (ret == HSUCCEED) {
+            INSPIRE_LOGD("image handle: %ld", (long)imageSteamHandle);
+        }
+
+        HF_MultipleFaceData multipleFaceData = {0};
+        HF_FaceContextRunFaceTrack(session, imageSteamHandle, &multipleFaceData);
+        INSPIRE_LOGD("Number of faces detected: %d", multipleFaceData.detectedNum);
+
+        for (int i = 0; i < multipleFaceData.detectedNum; ++i) {
+            cv::Rect rect =
+              cv::Rect(multipleFaceData.rects[i].x, multipleFaceData.rects[i].y, multipleFaceData.rects[i].width, multipleFaceData.rects[i].height);
+            cv::rectangle(image, rect, cv::Scalar(0, 255, 200), 2);
+            INSPIRE_LOGD("%d, track_id: %d, pitch: %f, yaw: %f, roll: %f", i, multipleFaceData.trackIds[i], multipleFaceData.angles.pitch[i],
+                         multipleFaceData.angles.yaw[i], multipleFaceData.angles.roll[i]);
+            INSPIRE_LOGD("token size: %d", multipleFaceData.tokens->size);
+        }
+#ifndef DISABLE_GUI
+//        cv::imshow("wq", image);
+//        cv::waitKey(0);
+#endif
+
+        ret = HF_FaceFeatureExtractCpy(session, imageSteamHandle, multipleFaceData.tokens[0], featuresCache[i]);
+
+        std::cout << "wtg" << std::endl;
+        if (ret != HSUCCEED) {
+            INSPIRE_LOGE("Abnormal feature extraction: %d", ret);
+            return -1;
+        }
+
+        //        for (int j = 0; j < 512; ++j) {
+        //            std::cout << featuresCache[0][j] << ", ";
+        //        }
+        //        std::cout << std::endl;
+
+        //        HSize size;
+        //        HF_GetFaceBasicTokenSize(&size);
+        //        LOGD("in size: %ld", size);
+        //
+        //        LOGD("o size %d", multipleFaceData.tokens[0].size);
+
+        HBuffer buffer[multipleFaceData.tokens[0].size];
+        HF_CopyFaceBasicToken(multipleFaceData.tokens[0], buffer, multipleFaceData.tokens[0].size);
+
+        HF_FaceBasicToken token = {0};
+        token.size = multipleFaceData.tokens[0].size;
+        token.data = buffer;
+
+        HFloat quality;
+        //        ret = HF_FaceQualityDetect(session, multipleFaceData.tokens[0], &quality);
+        ret = HF_FaceQualityDetect(session, token, &quality);
+        INSPIRE_LOGD("RET : %d", ret);
+        INSPIRE_LOGD("Q: %f", quality);
+
+        ret = HF_ReleaseImageStream(imageSteamHandle);
+        if (ret == HSUCCEED) {
+            imageSteamHandle = nullptr;
+            INSPIRE_LOGD("image released");
+        } else {
+            INSPIRE_LOGE("image release error: %ld", ret);
+        }
+    }
+
+    HFloat compResult;
+    HF_FaceFeature compFeature1 = {0};
+    HF_FaceFeature compFeature2 = {0};
+    compFeature1.size = featureNum;
+    compFeature1.data = featuresCache[0];
+    compFeature2.size = featureNum;
+    compFeature2.data = featuresCache[1];
+    ret = HF_FaceComparison1v1(compFeature1, compFeature2, &compResult);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Contrast failure: %d", ret);
+        return -1;
+    }
+    INSPIRE_LOGD("similarity: %f", compResult);
+
+    ret = HF_ReleaseFaceContext(session);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGD("Release error");
+    }
+
+    return 0;
+}
+
+int search() {
+    HResult ret;
+    HString path = "test_res/pack/Pikachu";
+    HF_ContextCustomParameter parameter = {0};
+    parameter.enable_liveness = 1;
+    parameter.enable_mask_detect = 1;
+    parameter.enable_recognition = 1;
+    HF_DetectMode detMode = HF_DETECT_MODE_IMAGE;
+    HContextHandle session;
+    ret = HF_CreateFaceContextFromResourceFile(path, parameter, detMode, 3, &session);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGD("An error occurred while creating ctx: %ld", ret);
+    }
+    HF_FeatureHubConfiguration databaseConfiguration = {0};
+    databaseConfiguration.enablePersistence = 1;
+    databaseConfiguration.dbPath = "./";
+    ret = HF_FeatureHubDataEnable(databaseConfiguration);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Database configuration failure: %ld", ret);
+        return -1;
+    }
+
+    std::vector<std::string> files_list = {
+
+    };
+
+    for (int i = 0; i < files_list.size(); ++i) {
+        auto &name = files_list[i];
+        cv::Mat image = cv::imread(name);
+        HF_ImageData imageData = {0};
+        imageData.data = image.data;
+        imageData.height = image.rows;
+        imageData.width = image.cols;
+        imageData.rotation = CAMERA_ROTATION_0;
+        imageData.format = STREAM_BGR;
+
+        HImageHandle imageSteamHandle;
+        ret = HF_CreateImageStream(&imageData, &imageSteamHandle);
+        if (ret != HSUCCEED) {
+            INSPIRE_LOGE("image handle error: %ld", (long)imageSteamHandle);
+            return -1;
+        }
+
+        HF_MultipleFaceData multipleFaceData = {0};
+        HF_FaceContextRunFaceTrack(session, imageSteamHandle, &multipleFaceData);
+
+        if (multipleFaceData.detectedNum <= 0) {
+            INSPIRE_LOGE("%s No face detected", name.c_str());
+            return -1;
+        }
+
+        HF_FaceFeature feature = {0};
+        ret = HF_FaceFeatureExtract(session, imageSteamHandle, multipleFaceData.tokens[0], &feature);
+        if (ret != HSUCCEED) {
+            INSPIRE_LOGE("Feature extraction error: %ld", ret);
+            return -1;
+        }
+
+        auto tag = basename(name);
+        char *tagName = new char[tag.size() + 1];
+        std::strcpy(tagName, tag.c_str());
+        HF_FaceFeatureIdentity identity = {0};
+        identity.feature = &feature;
+        identity.customId = i;
+        identity.tag = tagName;
+
+        ret = HF_FeatureHubInsertFeature(identity);
+        if (ret != HSUCCEED) {
+            INSPIRE_LOGE("Insert failed: %ld", ret);
+            return -1;
+        }
+
+        //        // Test duplicate insertion operation
+        //        ret = HF_FeaturesGroupInsertFeature(session, identity);
+        //        if (ret != HSUCCEED) {
+        //            INSPIRE_LOGE("Cannot insert duplicate ID: %ld", ret);
+        //        }
+
+        delete[] tagName;
+
+        ret = HF_ReleaseImageStream(imageSteamHandle);
+        if (ret == HSUCCEED) {
+            imageSteamHandle = nullptr;
+            INSPIRE_LOGD("image released");
+        } else {
+            INSPIRE_LOGE("image release error: %ld", ret);
+        }
+    }
+
+    cv::Mat image = cv::imread("test_res/images/kun.jpg");
+    HF_ImageData imageData = {0};
+    imageData.data = image.data;
+    imageData.height = image.rows;
+    imageData.width = image.cols;
+    imageData.rotation = CAMERA_ROTATION_0;
+    imageData.format = STREAM_BGR;
+
+    HImageHandle imageSteamHandle;
+    ret = HF_CreateImageStream(&imageData, &imageSteamHandle);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("image handle error: %ld", (long)imageSteamHandle);
+        return -1;
+    }
+    HF_MultipleFaceData multipleFaceData = {0};
+    HF_FaceContextRunFaceTrack(session, imageSteamHandle, &multipleFaceData);
+
+    if (multipleFaceData.detectedNum <= 0) {
+        INSPIRE_LOGE("No face detected");
+        return -1;
+    }
+
+    HF_FaceFeature feature = {0};
+    ret = HF_FaceFeatureExtract(session, imageSteamHandle, multipleFaceData.tokens[0], &feature);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Feature extraction error: %ld", ret);
+        return -1;
+    }
+
+    //    ret = HF_FaceContextFeatureRemove(session, 3);
+    //    if (ret != HSUCCEED) {
+    //        LOGE("delete failed: %ld", ret);
+    //    }
+
+    std::string newName = "Six";
+    char *newTagName = new char[newName.size() + 1];
+    std::strcpy(newTagName, newName.c_str());
+    HF_FaceFeatureIdentity updateIdentity = {0};
+    updateIdentity.customId = 1;
+    updateIdentity.tag = newTagName;
+    updateIdentity.feature = &feature;
+    ret = HF_FeatureHubFaceUpdate(updateIdentity);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Update failure: %ld", ret);
+    }
+    delete[] newTagName;
+
+    HF_FaceFeatureIdentity searchIdentity = {0};
+    //    HF_FaceFeature featureSearched = {0};
+    //    searchIdentity.feature = &featureSearched;
+    HFloat confidence;
+    ret = HF_FeatureHubFaceSearch(feature, &confidence, &searchIdentity);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Search failure: %ld", ret);
+        return -1;
+    }
+
+    INSPIRE_LOGD("Search for confidence: %f", confidence);
+    INSPIRE_LOGD("The matched tag: %s", searchIdentity.tag);
+    INSPIRE_LOGD("The matched customId: %d", searchIdentity.customId);
+
+    // Face Pipeline
+    ret = HF_MultipleFacePipelineProcess(session, imageSteamHandle, &multipleFaceData, parameter);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("pipeline execution failed: %ld", ret);
+        return -1;
+    }
+
+    HF_RGBLivenessConfidence livenessConfidence = {0};
+    ret = HF_GetRGBLivenessConfidence(session, &livenessConfidence);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to obtain live data");
+        return -1;
+    }
+    INSPIRE_LOGD("Failed to obtain live data: %f", livenessConfidence.confidence[0]);
+
+    HF_FaceMaskConfidence maskConfidence = {0};
+    ret = HF_GetFaceMaskConfidence(session, &maskConfidence);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to obtain live data");
+        return -1;
+    }
+    INSPIRE_LOGD("Mask wearing confidence: %f", maskConfidence.confidence[0]);
+
+    HInt32 faceNum;
+    ret = HF_FeatureHubGetFaceCount(&faceNum);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("fail to get");
+    }
+    INSPIRE_LOGD("Number of facial features: %d", faceNum);
+
+    HF_FeatureHubViewDBTable();
+
+    HF_FaceFeatureIdentity identity;
+    ret = HF_FeatureHubGetFaceIdentity(100, &identity);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Feature acquisition failure");
+    }
+
+    ret = HF_ReleaseImageStream(imageSteamHandle);
+    if (ret == HSUCCEED) {
+        imageSteamHandle = nullptr;
+        INSPIRE_LOGD("image released");
+    } else {
+        INSPIRE_LOGE("image release error: %ld", ret);
+    }
+
+    return 0;
+}
+
+int opiton() {
+    //    HInt32 mask = HF_ENABLE_FACE_RECOGNITION | HF_ENABLE_LIVENESS;
+
+    return 0;
+}
+
+int main() {
+    HResult ret;
+
+    //    {
+    //        // TestImageStream
+    //        cv::Mat image = cv::imread("test_res/images/kun.jpg");
+    //        HF_ImageData imageData = {0};
+    //        imageData.data = image.data;
+    //        imageData.height = image.rows;
+    //        imageData.width = image.cols;
+    //        imageData.rotation = CAMERA_ROTATION_0;
+    //        imageData.format = STREAM_BGR;
+    //
+    //        HImageHandle imageSteamHandle;
+    //        ret = HF_CreateImageStream(&imageData, &imageSteamHandle);
+    //        if (ret == HSUCCEED) {
+    //            LOGD("image handle: %ld", (long )imageSteamHandle);
+    //        }
+    //        HF_DeBugImageStreamImShow(imageSteamHandle);
+    //
+    //        ret = HF_ReleaseImageStream(imageSteamHandle);
+    //        if (ret == HSUCCEED) {
+    //            imageSteamHandle = nullptr;
+    //            LOGD("image released");
+    //        } else {
+    //            LOGE("image release error: %ld", ret);
+    //        }
+    //
+    //    }
+
+    //    compare();
+
+    search();
+
+    opiton();
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/context_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/context_sample.cpp
new file mode 100644
index 000000000..bb9bb4366
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/context_sample.cpp
@@ -0,0 +1,95 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "face_context.h"
+#include "opencv2/opencv.hpp"
+#include "sample/utils/test_helper.h"
+
+using namespace inspire;
+
+int main() {
+#ifndef ISF_USE_MOBILE_OPENCV_IN_LOCAL
+    FaceContext ctx;
+    CustomPipelineParameter param;
+    param.enable_liveness = true;
+    param.enable_face_quality = true;
+    int32_t ret = ctx.Configuration("test_res/pack/Pikachu-t1", DetectMode::DETECT_MODE_VIDEO, 1, param);
+    if (ret != 0) {
+        INSPIRE_LOGE("Initialization error");
+        return -1;
+    }
+    cv::VideoCapture cap(0);
+    if (!cap.isOpened()) {
+        std::cerr << "Unable to open the camera." << std::endl;
+        return -1;
+    }
+
+    cv::namedWindow("Webcam", cv::WINDOW_NORMAL);
+
+    while (true) {
+        cv::Mat frame;
+
+        cap >> frame;
+
+        if (frame.empty()) {
+            std::cerr << "Unable to obtain images from the camera." << std::endl;
+            break;
+        }
+
+        CameraStream stream;
+        stream.SetDataBuffer(frame.data, frame.rows, frame.cols);
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+
+        ctx.FaceDetectAndTrack(stream);
+
+        //        LOGD("Track Cost: %f", ctx.GetTrackTotalUseTime());
+
+        auto &faces = ctx.GetTrackingFaceList();
+        for (auto &face : faces) {
+            auto rect = face.GetRect();
+            int track_id = face.GetTrackingId();
+            int track_count = face.GetTrackingCount();
+
+            cv::rectangle(frame, rect, cv::Scalar(0, 0, 255), 2, 1);
+
+            std::string text = "ID: " + std::to_string(track_id) + " Count: " + std::to_string(track_count);
+
+            cv::Point text_position(rect.x, rect.y - 10);
+
+            const auto &pose_and_quality = face.high_result;
+            float mean_quality = 0.0f;
+            for (int i = 0; i < pose_and_quality.lmk_quality.size(); ++i) {
+                mean_quality += pose_and_quality.lmk_quality[i];
+            }
+            mean_quality /= pose_and_quality.lmk_quality.size();
+            mean_quality = 1 - mean_quality;
+            std::string pose_text = "pitch: " + std::to_string(pose_and_quality.pitch) + ",Yaw: " + std::to_string(pose_and_quality.yaw) +
+                                    ",roll:" + std::to_string(pose_and_quality.roll) + ", q: " + std::to_string(mean_quality);
+
+            cv::Point pose_position(rect.x, rect.y + rect.height + 20);
+
+            int font_face = cv::FONT_HERSHEY_SIMPLEX;
+            double font_scale = 0.5;
+            int font_thickness = 1;
+            cv::Scalar font_color(255, 255, 255);
+
+            cv::putText(frame, text, text_position, font_face, font_scale, font_color, font_thickness);
+            cv::putText(frame, pose_text, pose_position, font_face, font_scale, font_color, font_thickness);
+        }
+
+        cv::imshow("Webcam", frame);
+
+        if (cv::waitKey(1) == 27) {
+            break;
+        }
+    }
+    cap.release();
+
+    cv::destroyAllWindows();
+#endif
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/error_test.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/error_test.cpp
new file mode 100644
index 000000000..96e776545
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/error_test.cpp
@@ -0,0 +1,35 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "inspireface/c_api/inspireface.h"
+#include "inspireface/middleware/camera_stream/camera_stream.h"
+
+void non_file_test() {
+    HResult ret;
+    HPath path = "test_res/pack/abc";  // Use error path
+    HF_ContextCustomParameter parameter = {0};
+    HF_DetectMode detMode = HF_DETECT_MODE_IMAGE;
+    HContextHandle session;
+    ret = HF_CreateFaceContextFromResourceFile(path, parameter, detMode, 3, &session);
+    if (ret != 0) {
+        std::cout << "wrong" << std::endl;
+    }
+
+    HF_ReleaseFaceContext(session);
+}
+
+void camera_test() {
+    cv::Mat image = cv::imread("test_res/data/bulk/jntm.jpg");
+    inspire::CameraStream stream;
+    stream.SetRotationMode(inspire::ROTATION_0);
+    stream.SetDataFormat(inspire::NV12);
+    stream.SetDataBuffer(image.data, image.rows, image.cols);
+    auto decode = stream.GetScaledImage(1.0f, true);
+}
+
+int main() {
+    camera_test();
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/loop_tracker.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/loop_tracker.cpp
new file mode 100644
index 000000000..8317bab80
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/loop_tracker.cpp
@@ -0,0 +1,38 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "track_module/face_track.h"
+#include "inspireface/recognition_module/face_feature_extraction.h"
+#include "log.h"
+
+using namespace inspire;
+
+int main() {
+    InspireArchive archive("test_res/pack/Pikachu");
+
+    FaceTrack track;
+    //    FaceRecognition recognition(archive, true);
+
+    auto ret = track.Configuration(archive);
+    INSPIRE_LOGD("ret=%d", ret);
+
+    auto image = cv::imread("test_res/data/bulk/kun.jpg");
+    for (int i = 0; i < 10000000; ++i) {
+        CameraStream stream;
+        stream.SetDataBuffer(image.data, image.rows, image.cols);
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+
+        track.UpdateStream(stream, true);
+    }
+
+    //    InspireModel model;
+    //    ret = archive.LoadModel("mask_detect", model);
+    //    std::cout << ret << std::endl;
+    //
+    //    archive.PublicPrintSubFiles();
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/net_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/net_sample.cpp
new file mode 100644
index 000000000..a7f82e28c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/net_sample.cpp
@@ -0,0 +1,43 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "track_module/face_detect/face_pose.h"
+
+#include "middleware/model_archive/inspire_archive.h"
+
+using namespace inspire;
+
+int main(int argc, char** argv) {
+    InspireArchive loader;
+    loader.ReLoad("resource/pack/Pikachu");
+
+    Configurable param;
+    param.set<std::string>("input_layer", "data");
+    param.set<std::vector<std::string>>("outputs_layers", {
+                                                            "ip3_pose",
+                                                          });
+    param.set<std::vector<int>>("input_size", {112, 112});
+    param.set<std::vector<float>>("mean", {0.0f, 0.0f, 0.0f});
+    param.set<std::vector<float>>("norm", {1.0f, 1.0f, 1.0f});
+    param.set<int>("input_channel", 1);        // Input Gray
+    param.set<int>("input_image_channel", 1);  // BGR 2 Gray
+
+    auto m_pose_net_ = std::make_shared<FacePose>();
+    InspireModel model;
+    loader.LoadModel("", model);
+    m_pose_net_->LoadData(model);
+
+    auto image = cv::imread("resource/images/crop.png");
+
+    cv::Mat gray;
+    cv::resize(image, gray, cv::Size(112, 112));
+
+    auto res = (*m_pose_net_)(gray);
+    INSPIRE_LOGD("%f", res[0]);
+    INSPIRE_LOGD("%f", res[1]);
+    INSPIRE_LOGD("%f", res[2]);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/rec_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/rec_sample.cpp
new file mode 100644
index 000000000..5f35217f3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/rec_sample.cpp
@@ -0,0 +1,174 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "face_context.h"
+#include "sample/utils/test_helper.h"
+#include "inspireface/recognition_module/extract/alignment.h"
+#include "recognition_module/face_feature_extraction.h"
+#include "feature_hub/feature_hub.h"
+
+using namespace inspire;
+
+std::string GetFileNameWithoutExtension(const std::string &filePath) {
+    size_t slashPos = filePath.find_last_of("/\\");
+    if (slashPos != std::string::npos) {
+        std::string fileName = filePath.substr(slashPos + 1);
+
+        size_t dotPos = fileName.find_last_of('.');
+        if (dotPos != std::string::npos) {
+            return fileName.substr(0, dotPos);
+        } else {
+            return fileName;
+        }
+    }
+
+    size_t dotPos = filePath.find_last_of('.');
+    if (dotPos != std::string::npos) {
+        return filePath.substr(0, dotPos);
+    }
+
+    return filePath;
+}
+
+int comparison1v1(FaceContext &ctx) {
+    Embedded feature_1;
+    Embedded feature_2;
+
+    {
+        auto image = cv::imread("");
+        cv::Mat rot90;
+        TestUtils::rotate(image, rot90, ROTATION_90);
+
+        CameraStream stream;
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_90);
+        stream.SetDataBuffer(rot90.data, rot90.rows, rot90.cols);
+        ctx.FaceDetectAndTrack(stream);
+        const auto &faces = ctx.GetTrackingFaceList();
+        if (faces.empty()) {
+            INSPIRE_LOGD("image1 not face");
+            return -1;
+        }
+        ctx.FaceRecognitionModule()->FaceExtract(stream, faces[0], feature_1);
+    }
+
+    {
+        auto image = cv::imread("");
+        CameraStream stream;
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+        stream.SetDataBuffer(image.data, image.rows, image.cols);
+        ctx.FaceDetectAndTrack(stream);
+        const auto &faces = ctx.GetTrackingFaceList();
+        if (faces.empty()) {
+            INSPIRE_LOGD("image1 not face");
+            return -1;
+        }
+        ctx.FaceRecognitionModule()->FaceExtract(stream, faces[0], feature_2);
+    }
+
+    float rec;
+    auto ret = FEATURE_HUB->CosineSimilarity(feature_1, feature_2, rec);
+    INSPIRE_LOGD("rec: %f", rec);
+
+    return 0;
+}
+
+int search(FaceContext &ctx) {
+    //    std::shared_ptr<FeatureBlock> block;
+    //    block.reset(FeatureBlock::Create(hyper::MC_OPENCV));
+
+    std::vector<String> files_list = {};
+    for (int i = 0; i < files_list.size(); ++i) {
+        auto image = cv::imread(files_list[i]);
+        CameraStream stream;
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+        stream.SetDataBuffer(image.data, image.rows, image.cols);
+        ctx.FaceDetectAndTrack(stream);
+        const auto &faces = ctx.GetTrackingFaceList();
+        if (faces.empty()) {
+            INSPIRE_LOGD("image1 not face");
+            return -1;
+        }
+        Embedded feature;
+        ctx.FaceRecognitionModule()->FaceExtract(stream, faces[0], feature);
+        FEATURE_HUB->RegisterFaceFeature(feature, i, GetFileNameWithoutExtension(files_list[i]), 1000 + i);
+    }
+
+    //    ctx.FaceRecognitionModule()->PrintMatrix();
+
+    //    auto ret = block->DeleteFeature(3);
+    //    LOGD("DEL: %d", ret);
+    //    block->PrintMatrix();
+
+    FEATURE_HUB->DeleteFaceFeature(2);
+
+    INSPIRE_LOGD("Number of faces in the library: %d", FEATURE_HUB->GetFaceFeatureCount());
+
+    // Update or insert a face
+    {
+        Embedded feature;
+        auto image = cv::imread("");
+        CameraStream stream;
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+        stream.SetDataBuffer(image.data, image.rows, image.cols);
+        ctx.FaceDetectAndTrack(stream);
+        const auto &faces = ctx.GetTrackingFaceList();
+        if (faces.empty()) {
+            INSPIRE_LOGD("image1 not face");
+            return -1;
+        }
+        ctx.FaceRecognitionModule()->FaceExtract(stream, faces[0], feature);
+
+        //        block->UpdateFeature(4, feature);
+        //        block->AddFeature(feature);
+    }
+
+    // Prepare an image to search
+    {
+        Embedded feature;
+        auto image = cv::imread("");
+        CameraStream stream;
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+        stream.SetDataBuffer(image.data, image.rows, image.cols);
+        ctx.FaceDetectAndTrack(stream);
+        const auto &faces = ctx.GetTrackingFaceList();
+        if (faces.empty()) {
+            INSPIRE_LOGD("image1 not face");
+            return -1;
+        }
+        ctx.FaceRecognitionModule()->FaceExtract(stream, faces[0], feature);
+
+        SearchResult result;
+        auto timeStart = (double)cv::getTickCount();
+        FEATURE_HUB->SearchFaceFeature(feature, result);
+        double cost = ((double)cv::getTickCount() - timeStart) / cv::getTickFrequency() * 1000;
+        INSPIRE_LOGD("Search time: %f", cost);
+        INSPIRE_LOGD("Top1: %d, %f, %s %d", result.index, result.score, result.tag.c_str(), result.customId);
+    }
+
+    return 0;
+}
+
+int main(int argc, char **argv) {
+    FaceContext ctx;
+    CustomPipelineParameter param;
+    param.enable_recognition = true;
+    int32_t ret = ctx.Configuration("test_res/pack/Pikachu", DetectMode::DETECT_MODE_IMAGE, 1, param);
+    if (ret != 0) {
+        INSPIRE_LOGE("Initialization error");
+        return -1;
+    }
+
+    comparison1v1(ctx);
+
+    //    search(ctx);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sample.cpp
new file mode 100644
index 000000000..6b08b9b99
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sample.cpp
@@ -0,0 +1,72 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "face_context.h"
+#include "opencv2/opencv.hpp"
+#include "sample/utils/test_helper.h"
+
+using namespace inspire;
+
+int main(int argc, char** argv) {
+    FaceContext ctx;
+    CustomPipelineParameter param;
+    param.enable_liveness = true;
+    param.enable_face_quality = true;
+    int32_t ret = ctx.Configuration("test_res/pack/Pikachu", DetectMode::DETECT_MODE_IMAGE, 1, param);
+    if (ret != 0) {
+        INSPIRE_LOGE("Initialization error");
+        return -1;
+    }
+    auto image = cv::imread("test_res/images/kun.jpg");
+    cv::Mat rot90;
+    TestUtils::rotate(image, rot90, ROTATION_90);
+
+    CameraStream stream;
+    stream.SetDataFormat(BGR);
+    stream.SetRotationMode(ROTATION_90);
+    stream.SetDataBuffer(rot90.data, rot90.rows, rot90.cols);
+    ctx.FaceDetectAndTrack(stream);
+
+    std::vector<FaceTrackWrap> faces;
+    for (int i = 0; i < ctx.GetNumberOfFacesCurrentlyDetected(); ++i) {
+        //        const ByteArray &byteArray = ctx.GetDetectCache()[i];
+        FaceTrackWrap face = {0};
+        //        ret = DeserializeHyperFaceData(byteArray, face);
+
+        const FaceBasicData& faceBasic = ctx.GetFaceBasicDataCache()[i];
+        ret = DeserializeHyperFaceData((char*)faceBasic.data, faceBasic.dataSize, face);
+        INSPIRE_LOGD("OK!");
+
+        if (ret != HSUCCEED) {
+            return -1;
+        }
+        faces.push_back(face);
+
+        cv::Rect rect(face.rect.x, face.rect.y, face.rect.width, face.rect.height);
+        std::cout << rect << std::endl;
+        cv::rectangle(rot90, rect, cv::Scalar(0, 0, 233), 2);
+
+        for (auto& p : face.keyPoints) {
+            cv::Point2f point(p.x, p.y);
+            cv::circle(rot90, point, 0, cv::Scalar(0, 0, 255), 5);
+        }
+    }
+
+    //    cv::imshow("wq", rot90);
+    //    cv::waitKey(0);
+    cv::imwrite("wq.png", rot90);
+
+    ret = ctx.FacesProcess(stream, faces, param);
+    if (ret != HSUCCEED) {
+        return -1;
+    }
+
+    // view
+    int32_t index = 0;
+    INSPIRE_LOGD("liveness: %f", ctx.GetRgbLivenessResultsCache()[index]);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sqlite_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sqlite_sample.cpp
new file mode 100644
index 000000000..66275a193
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/test_sqlite_sample.cpp
@@ -0,0 +1,19 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "inspireface/feature_hub/persistence/sqlite_faces_manage.h"
+
+using namespace inspire;
+
+int main() {
+    SQLiteFaceManage db;
+
+    db.OpenDatabase("t.db");
+
+    db.ViewTotal();
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cluttered/standard/tracker_sample.cpp b/cpp-package/inspireface/cpp/sample/cluttered/standard/tracker_sample.cpp
new file mode 100644
index 000000000..401b097e6
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cluttered/standard/tracker_sample.cpp
@@ -0,0 +1,199 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "inspireface/track_module/face_track.h"
+#include "opencv2/opencv.hpp"
+
+using namespace inspire;
+
+int video_test(FaceTrack& ctx, int cam_id) {
+#ifndef ISF_USE_MOBILE_OPENCV_IN_LOCAL
+    cv::VideoCapture cap(cam_id);
+
+    if (!cap.isOpened()) {
+        std::cerr << "Unable to open the camera." << std::endl;
+        return -1;
+    }
+
+    cv::namedWindow("Webcam", cv::WINDOW_NORMAL);
+
+    while (true) {
+        cv::Mat frame;
+
+        cap >> frame;
+
+        if (frame.empty()) {
+            std::cerr << "Unable to obtain images from the camera." << std::endl;
+            break;
+        }
+
+        CameraStream stream;
+        stream.SetDataBuffer(frame.data, frame.rows, frame.cols);
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+
+        ctx.UpdateStream(stream, false);
+
+        INSPIRE_LOGD("Track Cost: %f", ctx.GetTrackTotalUseTime());
+
+        auto const& faces = ctx.trackingFace;
+        for (auto const& face : faces) {
+            auto rect = face.GetRect();
+            int track_id = face.GetTrackingId();
+            int track_count = face.GetTrackingCount();
+
+            cv::rectangle(frame, rect, cv::Scalar(0, 0, 255), 2, 1);
+
+            std::string text = "ID: " + std::to_string(track_id) + " Count: " + std::to_string(track_count);
+
+            cv::Point text_position(rect.x, rect.y - 10);
+
+            const auto& pose_and_quality = face.high_result;
+            std::vector<float> euler = {pose_and_quality.yaw, pose_and_quality.roll, pose_and_quality.pitch};
+            std::string pose_text = "P: " + std::to_string(euler[0]) + ",Yaw: " + std::to_string(euler[1]) + ",roll:" + std::to_string(euler[2]);
+
+            cv::Point pose_position(rect.x, rect.y + rect.height + 20);
+
+            int font_face = cv::FONT_HERSHEY_SIMPLEX;
+            double font_scale = 0.5;
+            int font_thickness = 1;
+            cv::Scalar font_color(255, 255, 255);
+
+            cv::putText(frame, text, text_position, font_face, font_scale, font_color, font_thickness);
+            cv::putText(frame, pose_text, pose_position, font_face, font_scale, font_color, font_thickness);
+        }
+
+        cv::imshow("Webcam", frame);
+
+        if (cv::waitKey(1) == 27) {
+            break;
+        }
+    }
+    cap.release();
+
+    cv::destroyAllWindows();
+
+#endif
+    return 0;
+}
+
+void video_file_test(FaceTrack& ctx, const std::string& video_filename) {
+#ifndef ISF_USE_MOBILE_OPENCV_IN_LOCAL
+    cv::VideoCapture cap(video_filename);
+
+    if (!cap.isOpened()) {
+        std::cerr << "Unable to open the video file: " << video_filename << std::endl;
+        return;
+    }
+
+    cv::namedWindow("Video", cv::WINDOW_NORMAL);
+
+    while (true) {
+        cv::Mat frame;
+        cap >> frame;
+
+        if (frame.empty()) {
+            std::cerr << "Unable to get frames from the video file." << std::endl;
+            break;
+        }
+
+        CameraStream stream;
+        stream.SetDataBuffer(frame.data, frame.rows, frame.cols);
+        stream.SetDataFormat(BGR);
+        stream.SetRotationMode(ROTATION_0);
+
+        ctx.UpdateStream(stream, false);
+        INSPIRE_LOGD("Track Cost: %f", ctx.GetTrackTotalUseTime());
+
+        auto const& faces = ctx.trackingFace;
+        for (auto const& face : faces) {
+            auto rect = face.GetRect();
+            int track_id = face.GetTrackingId();
+            int track_count = face.GetTrackingCount();
+
+            cv::rectangle(frame, rect, cv::Scalar(0, 0, 255), 2, 1);
+
+            auto lmk = face.GetLanmdark();
+            for (auto& p : lmk) {
+                cv::circle(frame, p, 0, cv::Scalar(0, 0, 242), 2);
+            }
+
+            std::string text = "ID: " + std::to_string(track_id) + " Count: " + std::to_string(track_count);
+
+            cv::Point text_position(rect.x, rect.y - 10);
+
+            const auto& euler = face.high_result;
+            std::string pose_text =
+              "pitch: " + std::to_string(euler.pitch) + ",Yaw: " + std::to_string(euler.yaw) + ",roll:" + std::to_string(euler.roll);
+
+            cv::Point pose_position(rect.x, rect.y + rect.height + 20);
+
+            int font_face = cv::FONT_HERSHEY_SIMPLEX;
+            double font_scale = 0.5;
+            int font_thickness = 1;
+            cv::Scalar font_color(255, 255, 255);
+
+            cv::putText(frame, text, text_position, font_face, font_scale, font_color, font_thickness);
+            cv::putText(frame, pose_text, pose_position, font_face, font_scale, font_color, font_thickness);
+        }
+
+        cv::imshow("Video", frame);
+
+        if (cv::waitKey(1) == 27) {
+            break;
+        }
+    }
+
+    cap.release();
+    cv::destroyAllWindows();
+#endif
+}
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cerr << "Usage: " << argv[0] << " <source> <input>" << std::endl;
+        return 1;
+    }
+
+    INSPIRE_SET_LOG_LEVEL(LogLevel::LOG_NONE);
+
+    const std::string source = argv[1];
+    const std::string input = argv[2];
+
+    const std::string folder = "test_res/pack/Pikachu";
+    INSPIRE_LOGD("%s", folder.c_str());
+    //    ModelLoader loader;
+    //    loader.Reset(folder);
+
+    InspireArchive archive;
+    archive.ReLoad(folder);
+    std::cout << archive.QueryStatus() << std::endl;
+    if (archive.QueryStatus() != SARC_SUCCESS) {
+        INSPIRE_LOGE("error archive");
+        return -1;
+    }
+
+    FaceTrack ctx;
+    ctx.Configuration(archive);
+
+    if (source == "webcam") {
+        int cam_id = std::stoi(input);
+        video_test(ctx, cam_id);
+    } else if (source == "image") {
+        cv::Mat image = cv::imread(input);
+        if (!image.empty()) {
+            //            image_test(ctx, image);
+        } else {
+            std::cerr << "Unable to open the image file." << std::endl;
+        }
+    } else if (source == "video") {
+        video_file_test(ctx, input);
+    } else {
+        std::cerr << "Invalid input source: " << source << std::endl;
+        return 1;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_gray_input.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_gray_input.cpp
new file mode 100644
index 000000000..72a9930c9
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_gray_input.cpp
@@ -0,0 +1,38 @@
+#include <iostream>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+
+int main() {
+    INSPIREFACE_CONTEXT->Reload("test_res/pack/Pikachu");
+    // Create session
+    inspire::CustomPipelineParameter param;
+    param.enable_recognition = true;
+    param.enable_liveness = true;
+    param.enable_mask_detect = true;
+    param.enable_face_attribute = true;
+    param.enable_face_quality = true;
+    std::shared_ptr<inspire::Session> session(inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, 1, param, 320));
+    // Prepare image
+    inspirecv::Image img = inspirecv::Image::Create("data.jpg", 3);
+    inspirecv::Image gray_img = img.ToGray();
+
+    // Create image and frame process
+    inspirecv::FrameProcess process =
+      inspirecv::FrameProcess::Create(gray_img.Data(), gray_img.Height(), gray_img.Width(), inspirecv::GRAY, inspirecv::ROTATION_0);
+    auto decode = process.ExecutePreviewImageProcessing(true);
+    decode.Write("decode.jpg");
+
+    // Detect
+    std::vector<inspire::FaceTrackWrap> results;
+    int32_t ret;
+    ret = session->FaceDetectAndTrack(process, results);
+    std::cout << "Face size: " << results.size() << std::endl;
+    for (const auto& result : results) {
+        // Draw face
+        inspirecv::Rect2i rect = inspirecv::Rect2i::Create(result.rect.x, result.rect.y, result.rect.width, result.rect.height);
+        img.DrawRect(rect, inspirecv::Color::Red);
+    }
+    img.Write("result.jpg");
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_nexus_process.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_nexus_process.cpp
new file mode 100644
index 000000000..e3538cb9b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_nexus_process.cpp
@@ -0,0 +1,29 @@
+#include <iostream>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+#include <inspireface/image_process/nexus_processor/image_processor.h>
+
+void test_resize(std::unique_ptr<inspire::nexus::ImageProcessor>& processor, int aligned_width) {
+    processor->SetAlignedWidth(aligned_width);
+    // Create image
+    inspirecv::Image img = inspirecv::Image::Create("kun.jpg", 3);
+    // Resize image
+    uint8_t* dst_data = nullptr;
+    int dst_width = 112;
+    int dst_height = 112;
+    processor->Resize(img.Data(), img.Width(), img.Height(), 3, &dst_data, dst_width, dst_height);
+    inspirecv::Image dst_img(dst_width, dst_height, 3, dst_data, false);
+    dst_img.Write("dst_w" + std::to_string(aligned_width) + ".jpg");
+    std::cout << "Save dst image to dst_w" << aligned_width << ".jpg" << std::endl;
+}
+
+int main() {
+    // Create image processor
+    auto processor = inspire::nexus::ImageProcessor::Create(inspire::Launch::IMAGE_PROCESSING_RGA);
+    test_resize(processor, 4);
+    test_resize(processor, 16);
+    // wrong aligned width
+    test_resize(processor, 7);
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_affine.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_affine.cpp
new file mode 100644
index 000000000..4bdcf6b9f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_affine.cpp
@@ -0,0 +1,84 @@
+#include <iostream>
+#include <vector>
+#include <string>
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+#include "inspireface/track_module/landmark/order_of_hyper_landmark.h"
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cout << "Usage: " << argv[0] << " <model_path> <image_path>" << std::endl;
+        return -1;
+    }
+
+    std::string model_path = argv[1];
+    std::string image_path = argv[2];
+
+    // Global init(only once)
+    INSPIREFACE_CONTEXT->Reload(model_path);
+
+    // Create image and frame process
+    inspirecv::Image image = inspirecv::Image::Create(image_path);
+    inspirecv::FrameProcess process =
+      inspirecv::FrameProcess::Create(image.Data(), image.Height(), image.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    // Create session
+    inspire::CustomPipelineParameter param;
+    param.enable_recognition = true;
+    param.enable_liveness = true;
+    param.enable_mask_detect = true;
+    param.enable_face_attribute = true;
+    param.enable_face_quality = true;
+    param.enable_interaction_liveness = true;
+    std::shared_ptr<inspire::Session> session(inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, 1, param, 320));
+
+    INSPIREFACE_CHECK_MSG(session != nullptr, "Session is not valid");
+
+    // Detect and track
+    std::vector<inspire::FaceTrackWrap> results;
+    int32_t ret;
+    ret = session->FaceDetectAndTrack(process, results);
+    INSPIREFACE_CHECK_MSG(ret == 0, "FaceDetectAndTrack failed");
+
+    auto first = results[0];
+    auto lmk = session->GetFaceDenseLandmark(first);
+    std::cout << "lmk: " << lmk.size() << std::endl;
+    for (size_t i = 0; i < lmk.size(); i++) {
+        image.DrawCircle(lmk[i].As<int>(), 5, inspirecv::Color::Red);
+    }
+
+    inspirecv::TransformMatrix rotation_mode_affine = process.GetAffineMatrix();
+
+    std::vector<inspirecv::Point2f> stand_lmk = ApplyTransformToPoints(lmk, rotation_mode_affine.GetInverse());
+
+    // Use total lmk
+    auto rect = inspirecv::MinBoundingRect(stand_lmk);
+    auto rect_pts = rect.As<float>().ToFourVertices();
+    std::vector<inspirecv::Point2f> dst_pts = {{0, 0}, {112, 0}, {112, 112}, {0, 112}};
+    std::vector<inspirecv::Point2f> camera_pts = ApplyTransformToPoints(rect_pts, rotation_mode_affine);
+
+    auto affine = inspirecv::SimilarityTransformEstimate(camera_pts, dst_pts);
+    auto image_affine = process.ExecuteImageAffineProcessing(affine, 112, 112);
+    image_affine.Write("affine.jpg");
+
+    // image.DrawRect(rect.As<int>(), inspirecv::Color::Red);
+    // image.Write("lmk.jpg");
+
+    std::vector<inspirecv::Point2i> points;
+    for (const auto& idx : inspire::HLMK_LEFT_EYE_POINTS_INDEX) {
+        points.emplace_back(stand_lmk[idx].GetX(), stand_lmk[idx].GetY());
+    }
+    std::cout << "points: " << points.size() << std::endl;
+    auto rect_eye = inspirecv::MinBoundingRect(points).Square(1.4f);
+    // draw debug
+    image.DrawRect(rect_eye.As<int>(), inspirecv::Color::Red);
+    auto rect_pts_eye = rect_eye.As<float>().ToFourVertices();
+    std::vector<inspirecv::Point2f> dst_pts_eye = {{0, 0}, {64, 0}, {64, 64}, {0, 64}};
+    std::vector<inspirecv::Point2f> camera_pts_eye = ApplyTransformToPoints(rect_pts_eye, rotation_mode_affine);
+
+    auto affine_eye = inspirecv::SimilarityTransformEstimate(camera_pts_eye, dst_pts_eye);
+    auto eye_affine = process.ExecuteImageAffineProcessing(affine_eye, 64, 64);
+    eye_affine.Write("eye.jpg");
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_comparison.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_comparison.cpp
new file mode 100644
index 000000000..e629fe449
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_comparison.cpp
@@ -0,0 +1,78 @@
+#include <iostream>
+#include <vector>
+#include <string>
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+
+int main(int argc, char** argv) {
+    if (argc != 4) {
+        std::cout << "Usage: " << argv[0] << " <model_path> <image_path1> <image_path2>" << std::endl;
+        return -1;
+    }
+
+    std::string model_path = argv[1];
+    std::string image_path1 = argv[2];
+    std::string image_path2 = argv[3];
+
+    // Global init(only once)
+    INSPIREFACE_CONTEXT->Reload(model_path);
+
+    // Create image and frame process
+    inspirecv::Image image1 = inspirecv::Image::Create(image_path1);
+    inspirecv::Image image2 = inspirecv::Image::Create(image_path2);
+    inspirecv::FrameProcess process1 = inspirecv::FrameProcess::Create(image1.Data(), image1.Height(), image1.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+    inspirecv::FrameProcess process2 = inspirecv::FrameProcess::Create(image2.Data(), image2.Height(), image2.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    // Create session
+    inspire::CustomPipelineParameter param;
+    param.enable_recognition = true;
+
+    // Create session
+    std::shared_ptr<inspire::Session> session(
+        inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, 1, param, 320));
+
+    INSPIREFACE_CHECK_MSG(session != nullptr, "Session is not valid");
+
+    // Detect and track
+    std::vector<inspire::FaceTrackWrap> results1;
+    std::vector<inspire::FaceTrackWrap> results2;
+
+    // Detect and track
+    session->FaceDetectAndTrack(process1, results1);
+    session->FaceDetectAndTrack(process2, results2);
+
+    INSPIREFACE_CHECK_MSG(!results1.empty() && !results2.empty(), "No face detected");
+
+    // Get feature
+    inspire::FaceEmbedding feature1;
+    inspire::FaceEmbedding feature2;
+    session->FaceFeatureExtract(process1, results1[0], feature1);
+    session->FaceFeatureExtract(process2, results2[0], feature2);
+
+    // Compare
+    float similarity;
+    INSPIREFACE_FEATURE_HUB->CosineSimilarity(feature1.embedding, feature2.embedding, similarity);
+    std::cout << "cosine of similarity: " << similarity << std::endl;
+    std::cout << "percentage of similarity: " << SIMILARITY_CONVERTER_RUN(similarity) << std::endl;
+
+    std::cout << "== using alignment image ==" << std::endl;
+
+    // Get face alignment image
+    inspirecv::Image wrapped1;
+    inspirecv::Image wrapped2;
+    session->GetFaceAlignmentImage(process1, results1[0], wrapped1);
+    session->GetFaceAlignmentImage(process2, results2[0], wrapped2);
+    wrapped1.Write("wrapped1.jpg");
+    wrapped2.Write("wrapped2.jpg");
+
+    inspire::FaceEmbedding feature1_alignment;
+    inspire::FaceEmbedding feature2_alignment;
+    session->FaceFeatureExtractWithAlignmentImage(wrapped1, feature1_alignment);
+    session->FaceFeatureExtractWithAlignmentImage(wrapped2, feature2_alignment);
+
+    INSPIREFACE_FEATURE_HUB->CosineSimilarity(feature1_alignment.embedding, feature2_alignment.embedding, similarity);
+    std::cout << "cosine of similarity: " << similarity << std::endl;
+    std::cout << "percentage of similarity: " << SIMILARITY_CONVERTER_RUN(similarity) << std::endl;
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_crud.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_crud.cpp
new file mode 100644
index 000000000..45a3066c2
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_crud.cpp
@@ -0,0 +1,94 @@
+#include <iostream>
+#include <inspireface/inspireface.hpp>
+
+int main() {
+    // Launch InspireFace
+    std::string model_path = "test_res/pack/Pikachu";
+    INSPIREFACE_CONTEXT->Reload(model_path);
+    INSPIREFACE_CHECK_MSG(INSPIREFACE_CONTEXT->isMLoad(), "InspireFace is not loaded");
+
+    // Enable feature hub
+    std::string db_path = "case_crud.db";
+    // Remove the database file if it exists
+    if (std::remove(db_path.c_str()) != 0) {
+        std::cerr << "Error removing database file: " << db_path << std::endl;
+    }
+    inspire::DatabaseConfiguration db_config;
+    db_config.enable_persistence = true;
+    db_config.persistence_db_path = db_path;
+    db_config.search_mode = inspire::SEARCH_MODE_EXHAUSTIVE;
+    db_config.recognition_threshold = 0.48f;
+    db_config.primary_key_mode = inspire::AUTO_INCREMENT;
+    auto ret = INSPIREFACE_FEATURE_HUB->EnableHub(db_config);
+    INSPIREFACE_CHECK_MSG(ret == HSUCCEED, "EnableHub failed");
+
+    // Create a session
+    auto param = inspire::CustomPipelineParameter();
+    param.enable_recognition = true;
+    auto session = inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, 1, param, 320);
+    INSPIREFACE_CHECK_MSG(session != nullptr, "Session is not created");
+
+    // Prepare an image for insertion into the hub
+    auto image = inspirecv::Image::Create("test_res/data/bulk/kun.jpg");
+    auto image_process = inspirecv::FrameProcess::Create(image.Data(), image.Height(), image.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    // Detect and track
+    std::vector<inspire::FaceTrackWrap> results;
+    session->FaceDetectAndTrack(image_process, results);
+    INSPIREFACE_CHECK_MSG(results.size() > 0, "No face detected");
+
+    // Extract face feature
+    inspire::FaceEmbedding feature;
+    session->FaceFeatureExtract(image_process, results[0], feature);
+
+    // Insert face feature into the hub, because the id is INSPIRE_INVALID_ID, so input id is ignored
+    int64_t result_id;
+    INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(feature.embedding, INSPIRE_INVALID_ID, result_id);
+
+    inspire::FaceEmbedding face_feature;
+    INSPIREFACE_FEATURE_HUB->GetFaceFeature(result_id, face_feature);
+    
+    // Prepare a photo of the same person for the query
+    auto query_image = inspirecv::Image::Create("test_res/data/bulk/jntm.jpg");
+    auto query_image_process = inspirecv::FrameProcess::Create(query_image.Data(), query_image.Height(), query_image.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    // Detect and track
+    std::vector<inspire::FaceTrackWrap> query_results;
+    session->FaceDetectAndTrack(query_image_process, query_results);
+    INSPIREFACE_CHECK_MSG(query_results.size() > 0, "No face detected");
+
+    // Extract face feature
+    inspire::FaceEmbedding query_feature;
+    session->FaceFeatureExtract(query_image_process, query_results[0], query_feature);
+
+    // Search face feature
+    inspire::FaceSearchResult search_result;
+    INSPIREFACE_FEATURE_HUB->SearchFaceFeature(query_feature.embedding, search_result, true);
+    std::cout << "Search face feature result: " << search_result.id << std::endl;
+    std::cout << "Search face feature similarity: " << search_result.similarity << std::endl;
+
+    INSPIREFACE_CHECK_MSG(search_result.id == result_id, "Search face feature result id is not equal to the inserted id");
+    
+    // Update the face feature
+    INSPIREFACE_FEATURE_HUB->FaceFeatureUpdate(query_feature.embedding, result_id);
+
+    // Remove the face feature
+    INSPIREFACE_FEATURE_HUB->FaceFeatureRemove(result_id);
+    INSPIREFACE_CHECK_MSG(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == 0, "Face feature is not removed");
+
+    
+    std::cout << "Remove face feature successfully" << std::endl;
+
+    // Query again
+    INSPIREFACE_FEATURE_HUB->SearchFaceFeature(query_feature.embedding, search_result, true);
+    INSPIREFACE_CHECK_MSG(search_result.id == INSPIRE_INVALID_ID, "Search face feature result id is not equal to the inserted id");
+    std::cout << "Query again, search face feature result: " << search_result.id << std::endl;
+
+    
+    // Top-k query
+    std::vector<inspire::FaceSearchResult> top_k_results;
+    INSPIREFACE_FEATURE_HUB->SearchFaceFeatureTopK(query_feature.embedding, top_k_results, 10, true);
+    std::cout << "Top-k query result: " << top_k_results.size() << std::endl;
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track.cpp
new file mode 100644
index 000000000..a302195dc
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track.cpp
@@ -0,0 +1,65 @@
+#include <iostream>
+#include <vector>
+#include <string>
+#include <memory>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cout << "Usage: " << argv[0] << " <model_path> <image_path>" << std::endl;
+        return -1;
+    }
+
+    std::string model_path = argv[1];
+    std::string image_path = argv[2];
+
+    // Global init(only once)
+    INSPIREFACE_CONTEXT->Reload(model_path);
+
+    // Create image and frame process
+    inspirecv::Image image = inspirecv::Image::Create(image_path);
+    inspirecv::FrameProcess process =
+      inspirecv::FrameProcess::Create(image.Data(), image.Height(), image.Width(), inspirecv::BGR, inspirecv::ROTATION_90);
+
+    // Create session
+    inspire::CustomPipelineParameter param;
+    param.enable_recognition = true;
+    param.enable_liveness = true;
+    param.enable_mask_detect = true;
+    param.enable_face_attribute = true;
+    param.enable_face_quality = true;
+    std::shared_ptr<inspire::Session> session(inspire::Session::CreatePtr(inspire::DETECT_MODE_ALWAYS_DETECT, 100, param, 640));
+    session->SetTrackPreviewSize(640);
+
+    INSPIREFACE_CHECK_MSG(session != nullptr, "Session is not valid");
+
+    // Detect and track
+    std::vector<inspire::FaceTrackWrap> results;
+    int32_t ret;
+    ret = session->FaceDetectAndTrack(process, results);
+    INSPIREFACE_CHECK_MSG(ret == 0, "FaceDetectAndTrack failed");
+
+    // Run pipeline for each face
+    ret = session->MultipleFacePipelineProcess(process, param, results);
+    INSPIREFACE_CHECK_MSG(ret == 0, "MultipleFacePipelineProcess failed");
+
+    for (auto& result : results) {
+        std::cout << "result: " << result.trackId << std::endl;
+        std::cout << "quality: " << result.quality[0] << ", " << result.quality[1] << ", " << result.quality[2] << ", " << result.quality[3] << ", "
+                  << result.quality[4] << std::endl;
+        inspirecv::Rect2i rect = inspirecv::Rect2i::Create(result.rect.x, result.rect.y, result.rect.width, result.rect.height);
+        std::cout << rect << std::endl;
+        image.DrawRect(rect, inspirecv::Color::Red);
+        inspirecv::TransformMatrix trans = inspirecv::TransformMatrix::Create(result.trans.m00, result.trans.m01, result.trans.tx, result.trans.m10, result.trans.m11, result.trans.ty);
+        std::cout << "trans: " << trans.GetInverse() << std::endl;
+
+        std::vector<inspirecv::Point2f> landmark = session->GetFaceDenseLandmark(result);
+        for (auto& point : landmark) {
+            image.DrawCircle(point.As<int>(), 2, inspirecv::Color::Green);
+        }
+    }
+    image.Write("result.jpg");
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track_time.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track_time.cpp
new file mode 100644
index 000000000..b9bc0f695
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_face_track_time.cpp
@@ -0,0 +1,58 @@
+#include <iostream>
+#include <vector>
+#include <string>
+#include <memory>
+#include <chrono>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cout << "Usage: " << argv[0] << " <model_path> <image_path>" << std::endl;
+        return -1;
+    }
+
+    std::string model_path = argv[1];
+    std::string image_path = argv[2];
+
+    // Global init(only once)
+    INSPIREFACE_CONTEXT->Reload(model_path);
+
+    // Create image and frame process
+    inspirecv::Image image = inspirecv::Image::Create(image_path);
+    inspirecv::FrameProcess process =
+      inspirecv::FrameProcess::Create(image.Data(), image.Height(), image.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+
+    // Create session
+    inspire::CustomPipelineParameter param;
+    param.enable_recognition = true;
+    param.enable_liveness = true;
+    param.enable_mask_detect = true;
+    param.enable_face_attribute = true;
+    param.enable_face_quality = true;
+    std::shared_ptr<inspire::Session> session(inspire::Session::CreatePtr(inspire::DETECT_MODE_LIGHT_TRACK, 100, param, 640));
+    session->SetTrackPreviewSize(640);
+    session->SetTrackModeDetectInterval(10);
+
+    INSPIREFACE_CHECK_MSG(session != nullptr, "Session is not valid");
+
+
+    for (int i = 0; i < 100; i++) {
+        
+        // Detect and track
+        std::vector<inspire::FaceTrackWrap> results;
+        int32_t ret;
+        // Start time
+        auto start = std::chrono::high_resolution_clock::now();
+        ret = session->FaceDetectAndTrack(process, results);
+        INSPIREFACE_CHECK_MSG(ret == 0, "FaceDetectAndTrack failed");
+        auto end = std::chrono::high_resolution_clock::now();
+        // End time
+        auto duration = std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+        std::cout << i << " MultipleFacePipelineProcess: " << duration.count() / 1000.0 << " ms" << std::endl;
+
+    }
+    
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_inspirecv.cpp b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_inspirecv.cpp
new file mode 100644
index 000000000..32927bc0b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/cpp_api/cpp_sample_inspirecv.cpp
@@ -0,0 +1,200 @@
+#include <iostream>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/inspireface.hpp>
+#ifdef _WIN32
+#include <direct.h>
+#define CREATE_DIR(dir) _mkdir(dir)
+#else
+#include <sys/stat.h>
+#define CREATE_DIR(dir) mkdir(dir, 0777)
+#endif
+
+int main() {
+    // Make directory
+    if (CREATE_DIR("cv") == -1) {
+        // If the directory already exists, it is not an error
+        if (errno != EEXIST) {
+            std::cerr << "Error creating directory" << std::endl;
+            return 1;
+        }
+    }
+
+    /* Image I/O */
+
+    // Load image from file
+    // Load with 3 channels (BGR, like opencv)
+    inspirecv::Image img = inspirecv::Image::Create("test_res/data/bulk/kun_cartoon_crop.jpg", 3);
+
+    // Load image from buffer
+    // uint8_t* buffer = ...;  // buffer is a pointer to the image data
+    // bool is_alloc_mem = false;  // if true, will allocate memory for the image data,
+    //                             // false is recommended to point to the original data to avoid copying
+    // inspirecv::Image img_buffer = inspirecv::Image::Create(width, height, channel, buffer, is_alloc_mem);
+
+    // Save image to file
+    img.Write("cv/output.jpg");
+
+    // Show image, warning: it must depend on opencv
+    // img.Show("input");
+
+    // Get pointer to image data
+    const uint8_t* ptr = img.Data();
+
+    /* Image Processing */
+    // Convert to grayscale
+    inspirecv::Image gray = img.ToGray();
+    gray.Write("cv/gray.jpg");
+
+    // Apply Gaussian blur
+    inspirecv::Image blurred = img.GaussianBlur(3, 1.0);
+    blurred.Write("cv/blurred.jpg");
+
+    // Geometric transformations
+    auto scale = 0.35;
+    bool use_bilinear = true;
+    inspirecv::Image resized = img.Resize(img.Width() * scale, img.Height() * scale, use_bilinear);  // Resize image
+    resized.Write("cv/resized.jpg");
+
+    // Rotate 90 degrees clockwise
+    inspirecv::Image rotated = img.Rotate90();
+    rotated.Write("cv/rotated.jpg");
+
+    // Flip vertically
+    inspirecv::Image flipped_vertical = img.FlipVertical();
+    flipped_vertical.Write("cv/flipped_vertical.jpg");
+
+    // Flip horizontally
+    inspirecv::Image flipped_horizontal = img.FlipHorizontal();
+    flipped_horizontal.Write("cv/flipped_horizontal.jpg");
+
+    // Crop for rectangle
+    inspirecv::Rect<int> rect = inspirecv::Rect<int>::Create(78, 41, 171, 171);
+    inspirecv::Image cropped = img.Crop(rect);
+    cropped.Write("cv/cropped.jpg");
+
+    // Image padding
+    int top = 50, bottom = 50, left = 50, right = 50;
+    inspirecv::Image padded = img.Pad(top, bottom, left, right, inspirecv::Color::Black);
+    padded.Write("cv/padded.jpg");
+
+    // Swap red and blue channels
+    inspirecv::Image swapped = img.SwapRB();
+    swapped.Write("cv/swapped.jpg");
+
+    // Multiply image by scale factor
+    double scale_factor = 0.5;
+    inspirecv::Image scaled = img.Mul(scale_factor);
+    scaled.Write("cv/scaled.jpg");
+
+    // Add value to image
+    double value = -175;
+    inspirecv::Image added = img.Add(value);
+    added.Write("cv/added.jpg");
+
+    // Rotate 90 degrees clockwise(also support 270 and 180)
+    inspirecv::Image rotated_90 = img.Rotate90();
+    rotated_90.Write("cv/rotated_90.jpg");
+
+    // Affine transform
+    /**
+     * Create a transform matrix from the following matrix
+     * [[a11, a12, tx],
+     *  [a21, a22, ty]]
+     *
+     * Face crop transform matrix
+     * [[0.0, -1.37626, 261.127],
+     *  [1.37626, 0.0, 85.1831]]
+     */
+    float a11 = 0.0f;
+    float a12 = -1.37626f;
+    float a21 = 1.37626f;
+    float a22 = 0.0f;
+    float b1 = 261.127f;
+    float b2 = 85.1831f;
+
+    inspirecv::TransformMatrix trans = inspirecv::TransformMatrix::Create(a11, a12, b1, a21, a22, b2);
+    int dst_width = 112;
+    int dst_height = 112;
+    inspirecv::Image affine = rotated_90.WarpAffine(trans, dst_width, dst_height);
+    affine.Write("cv/affine.jpg");
+
+    /* Image Draw */
+    inspirecv::Image draw_img = img.Clone();
+
+    // Draw a rectangle
+    inspirecv::Rect<int> new_rect = rect.Square(1.1f);  // Square and expand the rect
+    int thickness = 3;
+    draw_img.DrawRect(new_rect, inspirecv::Color::Green, thickness);
+    draw_img.Write("cv/draw_rect.jpg");
+
+    // Draw a circle
+    draw_img = img.Clone();
+    std::vector<inspirecv::Point<int>> points = new_rect.As<int>().ToFourVertices();
+    for (auto& point : points) {
+        draw_img.DrawCircle(point, 1, inspirecv::Color::Red, 5);
+    }
+    draw_img.Write("cv/draw_circle.jpg");
+
+    // Draw a line
+    draw_img = img.Clone();
+    draw_img.DrawLine(points[0], points[1], inspirecv::Color::Cyan, 2);
+    draw_img.DrawLine(points[1], points[2], inspirecv::Color::Magenta, 2);
+    draw_img.DrawLine(points[2], points[3], inspirecv::Color::Pink, 2);
+    draw_img.DrawLine(points[3], points[0], inspirecv::Color::Yellow, 2);
+    draw_img.Write("cv/draw_line.jpg");
+
+    // Fill a rectangle
+    draw_img = img.Clone();
+    draw_img.Fill(new_rect, inspirecv::Color::Purple);
+    draw_img.Write("cv/fill_rect.jpg");
+
+    // Reset
+    std::vector<uint8_t> gray_color(img.Width() * img.Height() * 3, 128);
+    img.Reset(img.Width(), img.Height(), 3, gray_color.data());
+    img.Write("cv/reset.jpg");
+
+    /** FrameProcess */
+
+    // BGR888 as raw data
+    inspirecv::Image raw = inspirecv::Image::Create("test_res/data/bulk/kun_cartoon_crop_r90.jpg", 3);
+    const uint8_t* buffer = raw.Data();
+
+    // You can also use other image format, like NV21, NV12, RGBA, RGB, BGR, BGRA
+    // const uint8_t* buffer = ...;
+
+    // Create frame process
+    auto width = raw.Width();
+    auto height = raw.Height();
+    auto rotation_mode = inspirecv::ROTATION_90;
+    auto data_format = inspirecv::BGR;
+    inspirecv::FrameProcess frame_process = inspirecv::FrameProcess::Create(buffer, height, width, data_format, rotation_mode);
+
+    // Set preview size
+    frame_process.SetPreviewSize(160);
+
+    // Set preview scale
+    // frame_process.SetPreviewScale(0.5f);
+
+    // Get transform image
+    inspirecv::Image transform_img = frame_process.ExecutePreviewImageProcessing(true);
+    transform_img.Write("cv/transform_img.jpg");
+
+    // ExecuteImageAffineProcessing
+
+    // Face crop transform matrix
+    // [[0.0, 0.726607, -61.8946],
+    //  [-0.726607, 0.0, 189.737]]
+    a11 = 0.0f;
+    a12 = 0.726607f;
+    a21 = -0.726607;
+    a22 = 0.0f;
+    b1 = -61.8946f;
+    b2 = 189.737f;
+    inspirecv::TransformMatrix affine_matrix = inspirecv::TransformMatrix::Create(a11, a12, b1, a21, a22, b2);
+    dst_width = 112;
+    dst_height = 112;
+    inspirecv::Image affine_img = frame_process.ExecuteImageAffineProcessing(affine_matrix, dst_width, dst_height);
+    affine_img.Write("cv/affine_img.jpg");
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/rv1106/face_attribute.cpp b/cpp-package/inspireface/cpp/sample/rv1106/face_attribute.cpp
new file mode 100644
index 000000000..8929b467f
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/rv1106/face_attribute.cpp
@@ -0,0 +1,31 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/pipeline_module/attribute/face_attribute_adapt.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+#include <inspireface/include/inspireface/spend_timer.h>
+#include <inspireface/include/inspireface/herror.h>
+#include <log.h>
+
+using namespace inspire;
+
+int main() {
+    INSPIRE_SET_LOG_LEVEL(ISF_LOG_DEBUG);
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Gundam_RV1106");
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    InspireModel detModel;
+    auto ret = archive.LoadModel("face_attribute", detModel);
+    if (ret != SARC_SUCCESS) {
+        INSPIRE_LOGE("Load %s error: %d", "face_detect_160", ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+
+    FaceAttributePredictAdapt face_attribute;
+    face_attribute.LoadData(detModel, detModel.modelType, false);
+
+    auto img = inspirecv::Image::Create("test_res/data/crop/crop.png");
+    auto result = face_attribute(img);
+    std::cout << "result: " << result[0] << ", " << result[1] << ", " << result[2] << std::endl;
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/rv1106/face_detect.cpp b/cpp-package/inspireface/cpp/sample/rv1106/face_detect.cpp
new file mode 100644
index 000000000..778f9b48e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/rv1106/face_detect.cpp
@@ -0,0 +1,50 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/track_module/face_track_module.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+#include <inspireface/include/inspireface/spend_timer.h>
+#include <inspireface/include/inspireface/herror.h>
+
+using namespace inspire;
+
+int main() {
+    INSPIRE_SET_LOG_LEVEL(ISF_LOG_DEBUG);
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Gundam_RV1106");
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    InspireModel detModel;
+    auto ret = archive.LoadModel("face_detect_160", detModel);
+    if (ret != SARC_SUCCESS) {
+        INSPIRE_LOGE("Load %s error: %d", "face_detect_160", ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+    FaceDetectAdapt face_detect(160);
+    std::vector<int> input_size;
+    input_size = detModel.Config().get<std::vector<int>>("input_size");
+
+    ret = face_detect.LoadData(detModel, detModel.modelType, false);
+    if (ret != 0) {
+        INSPIRE_LOGE("Load %s error: %d", "face_detect_160", ret);
+        return HERR_ARCHIVE_LOAD_MODEL_FAILURE;
+    }
+
+    std::cout << "Load model success" << std::endl;
+
+    auto img = inspirecv::Image::Create("data/bulk/kun.jpg");
+
+    inspire::SpendTimer time_spend("Detect");
+    FaceLocList results;
+    for (int i = 0; i < 10; i++) {
+        time_spend.Start();
+        results = face_detect(img);
+        time_spend.Stop();
+        std::cout << "================" << std::endl;
+    }
+    std::cout << time_spend << std::endl;
+    std::cout << "Face detect success:" << results.size() << std::endl;
+    for (auto &face : results) {
+        std::cout << "Face detect success:" << face.x1 << " " << face.y1 << " " << face.x2 << " " << face.y2 << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/rv1106/rga_image.cpp b/cpp-package/inspireface/cpp/sample/rv1106/rga_image.cpp
new file mode 100644
index 000000000..e4f265a41
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/rv1106/rga_image.cpp
@@ -0,0 +1,130 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/image_process/nexus_processor/image_processor.h>
+#include "log.h"
+#include <inspireface/include/inspireface/spend_timer.h>
+#include <inspireface/include/inspireface/herror.h>
+
+using namespace inspire;
+
+int main() {
+#if defined(ISF_ENABLE_RGA)
+    std::cout << "ISF_ENABLE_RGA is open" << std::endl;
+#else
+    std::cout << "ISF_ENABLE_RGA is close" << std::endl;
+#endif
+
+    auto img = inspirecv::Image::Create("data/bulk/r0.jpg");
+    auto processor = nexus::ImageProcessor::Create();
+
+    uint8_t* resized_data = nullptr;
+    int resized_width = 100;
+    int resized_height = 100;
+    inspire::SpendTimer time_spend("RGA resize");
+    for (int i = 0; i < 10; i++) {
+        time_spend.Start();
+        auto ret = processor->Resize(img.Data(), img.Width(), img.Height(), img.Channels(), &resized_data, resized_width, resized_height);
+        time_spend.Stop();
+        if (ret != 0) {
+            INSPIRE_LOGE("RGA resize failed: %d", ret);
+            return -1;
+        }
+    }
+    processor->DumpCacheStatus();
+    std::cout << time_spend << std::endl;
+
+    auto resized_img = inspirecv::Image::Create(resized_width, resized_height, img.Channels(), resized_data);
+    resized_img.Write("save_dir/kun_resized.jpg");
+
+    processor->MarkDone();
+
+    uint8_t* swapped_data = nullptr;
+    inspire::SpendTimer swap_time_spend("RGA swap color");
+    for (int i = 0; i < 10; i++) {
+        swap_time_spend.Start();
+        auto ret = processor->SwapColor(resized_img.Data(), resized_img.Width(), resized_img.Height(), resized_img.Channels(), &swapped_data);
+        swap_time_spend.Stop();
+    }
+    std::cout << swap_time_spend << std::endl;
+    processor->DumpCacheStatus();
+
+    auto swapped_img = inspirecv::Image::Create(resized_img.Width(), resized_img.Height(), resized_img.Channels(), swapped_data);
+    swapped_img.Write("save_dir/kun_swapped.jpg");
+
+    processor->MarkDone();
+
+    // padding
+    uint8_t* padded_data = nullptr;
+    int top = 10;
+    int bottom = 10;
+    int left = 10;
+    int right = 10;
+    inspire::SpendTimer padding_time_spend("RGA padding");
+    int padded_width = 0;
+    int padded_height = 0;
+    for (int i = 0; i < 10; i++) {
+        padding_time_spend.Start();
+        auto ret = processor->Padding(swapped_img.Data(), swapped_img.Width(), swapped_img.Height(), swapped_img.Channels(), top, bottom, left, right,
+                                      &padded_data, padded_width, padded_height);
+        padding_time_spend.Stop();
+    }
+    processor->DumpCacheStatus();
+    std::cout << padding_time_spend << std::endl;
+
+    auto padded_img = inspirecv::Image::Create(padded_width, padded_height, swapped_img.Channels(), padded_data);
+    padded_img.Write("save_dir/kun_padded.jpg");
+
+    processor->MarkDone();
+
+    // inspirecv crop
+    inspirecv::Rect2i rect(30, 30, 70, 70);
+    inspire::SpendTimer inspirecv_crop_time_spend("InspireCV crop");
+    inspirecv::Image inspirecv_cropped_img;
+    for (int i = 0; i < 10; i++) {
+        inspirecv_crop_time_spend.Start();
+        inspirecv_cropped_img = padded_img.Crop(rect);
+        inspirecv_crop_time_spend.Stop();
+    }
+    std::cout << inspirecv_crop_time_spend << std::endl;
+    inspirecv_cropped_img.Write("save_dir/kun_cropped_inspirecv.jpg");
+    // Padding and crop
+    inspirecv::Image image = inspirecv::Image::Create("data/bulk/r90.jpg");
+    uint8_t* padded_cropped_data = nullptr;
+    int dst_width = 320;
+    int dst_height = 320;
+    float scale = 0.0f;
+    inspire::SpendTimer padded_crop_time_spend("RGA padded and cropped");
+    for (int i = 0; i < 10; i++) {
+        padded_crop_time_spend.Start();
+        auto ret = processor->ResizeAndPadding(image.Data(), image.Width(), image.Height(), image.Channels(), dst_width, dst_height,
+                                               &padded_cropped_data, scale);
+        padded_crop_time_spend.Stop();
+    }
+    processor->DumpCacheStatus();
+    std::cout << padded_crop_time_spend << std::endl;
+
+    auto padded_cropped_img = inspirecv::Image::Create(dst_width, dst_height, img.Channels(), padded_cropped_data);
+    padded_cropped_img.Write("save_dir/image_padded_cropped.jpg");
+
+    processor->MarkDone();
+
+    // resize 2
+    uint8_t* resized_data_2 = nullptr;
+    int resized_width_2 = 512;
+    int resized_height_2 = 512;
+    inspire::SpendTimer time_spend_2("RGA resize 2");
+    for (int i = 0; i < 10; i++) {
+        time_spend_2.Start();
+        auto ret = processor->Resize(padded_cropped_img.Data(), padded_cropped_img.Width(), padded_cropped_img.Height(),
+                                     padded_cropped_img.Channels(), &resized_data_2, resized_width_2, resized_height_2);
+        time_spend_2.Stop();
+    }
+    std::cout << time_spend_2 << std::endl;
+    processor->DumpCacheStatus();
+
+    auto resized_img_2 = inspirecv::Image::Create(resized_width_2, resized_height_2, padded_cropped_img.Channels(), resized_data_2);
+    resized_img_2.Write("save_dir/image_padded_cropped_resized_2.jpg");
+
+    processor->MarkDone();
+
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/source/expansion_load.cpp b/cpp-package/inspireface/cpp/sample/source/expansion_load.cpp
new file mode 100644
index 000000000..84b937b1e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/source/expansion_load.cpp
@@ -0,0 +1,121 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/include/inspireface/launch.h>
+#include "inspireface/middleware/model_archive/inspire_archive.h"
+#include "inspireface/track_module/face_detect/face_detect_adapt.h"
+#include "inspireface/track_module/landmark/face_landmark_adapt.h"
+#include "inspireface/track_module/quality/face_pose_quality_adapt.h"
+#include "inspireface/recognition_module/extract/extract_adapt.h"
+#include "inspireface/include/inspireface/spend_timer.h"
+
+void test_face_detect() {
+    inspire::InspireModel model;
+    INSPIREFACE_CONTEXT->getMArchive().LoadModel("face_detect_160", model);
+    auto input_size = 160;
+    inspire::FaceDetectAdapt faceDetectAdapt(input_size);
+    faceDetectAdapt.LoadData(model, model.modelType);
+    inspirecv::Image image = inspirecv::Image::Create("test_res/data/bulk/kun.jpg");
+    inspire::FaceLocList faces;
+    inspire::SpendTimer timeSpend("Face Detect@" + std::to_string(input_size));
+    for (int i = 0; i < 1000; i++) {
+        timeSpend.Start();
+        faces = faceDetectAdapt(image);
+        timeSpend.Stop();
+    }
+    std::cout << timeSpend << std::endl;
+    ;
+    std::cout << "faces size: " << faces.size() << std::endl;
+    for (auto &face : faces) {
+        inspirecv::Rect2i rect = inspirecv::Rect2i::Create(face.x1, face.y1, face.x2 - face.x1, face.y2 - face.y1);
+        image.DrawRect(rect, {0, 0, 255});
+    }
+    image.Write("im.jpg");
+}
+
+void test_landmark() {
+    inspire::InspireModel model;
+    INSPIREFACE_CONTEXT->getMArchive().LoadModel("landmark", model);
+    auto input_size = 112;
+    inspire::FaceLandmarkAdapt landmarkAdapt(input_size);
+    landmarkAdapt.LoadData(model, model.modelType);
+    inspirecv::Image image = inspirecv::Image::Create("test_res/data/crop/crop.png");
+    image = image.Resize(input_size, input_size);
+    std::vector<float> lmk;
+    inspire::SpendTimer timeSpend("Landmark@" + std::to_string(input_size));
+    timeSpend.Start();
+    for (int i = 0; i < 10; i++) {
+        lmk = landmarkAdapt(image);
+    }
+    timeSpend.Stop();
+    std::cout << timeSpend << std::endl;
+    ;
+    for (int i = 0; i < inspire::FaceLandmarkAdapt::NUM_OF_LANDMARK; i++) {
+        auto p = inspirecv::Point2i::Create(lmk[i * 2] * input_size, lmk[i * 2 + 1] * input_size);
+        image.DrawCircle(p, 5, {0, 0, 255});
+    }
+    image.Write("lm.jpg");
+}
+
+void test_quality() {
+    inspire::InspireModel model;
+    INSPIREFACE_CONTEXT->getMArchive().LoadModel("pose_quality", model);
+    auto input_size = 96;
+    inspire::FacePoseQualityAdapt poseQualityAdapt;
+    poseQualityAdapt.LoadData(model, model.modelType);
+    inspirecv::Image image = inspirecv::Image::Create("test_res/data/crop/crop.png");
+    image = image.Resize(input_size, input_size);
+    inspire::FacePoseQualityAdaptResult quality;
+    inspire::SpendTimer timeSpend("Pose Quality@" + std::to_string(input_size));
+    timeSpend.Start();
+    for (int i = 0; i < 10; i++) {
+        quality = poseQualityAdapt(image);
+    }
+    timeSpend.Stop();
+    std::cout << timeSpend << std::endl;
+    ;
+    std::cout << "quality: " << quality.pitch << ", " << quality.yaw << ", " << quality.roll << std::endl;
+    for (int i = 0; i < quality.lmk.size(); i++) {
+        std::cout << "lmk: " << quality.lmk[i].GetX() << ", " << quality.lmk[i].GetY() << std::endl;
+        auto p = inspirecv::Point2i::Create(quality.lmk[i].GetX(), quality.lmk[i].GetY());
+        image.DrawCircle(p, 3, {0, 0, 255});
+    }
+    image.Write("qu.jpg");
+}
+
+void test_feature() {
+    inspire::InspireModel model;
+    INSPIREFACE_CONTEXT->getMArchive().LoadModel("feature", model);
+    auto input_size = 112;
+    inspire::ExtractAdapt extractAdapt;
+    extractAdapt.LoadData(model, model.modelType);
+    inspirecv::Image image = inspirecv::Image::Create("test_res/data/crop/crop.png");
+    image = image.Resize(input_size, input_size);
+    float norm;
+    bool normalize = true;
+    inspire::SpendTimer timeSpend("Extract@" + std::to_string(input_size));
+    timeSpend.Start();
+    inspire::Embedded feature;
+    for (int i = 0; i < 10; i++) {
+        feature = extractAdapt(image, norm, normalize);
+    }
+    timeSpend.Stop();
+    std::cout << timeSpend << std::endl;
+    ;
+    std::cout << "feature: " << feature.size() << std::endl;
+}
+
+int main() {
+    std::string archivePath = "test_res/pack/Pikachu_Apple";
+    INSPIREFACE_CONTEXT->Load(archivePath);
+    // Test face detect
+    test_face_detect();
+
+    // Test landmark
+    // test_landmark();
+
+    // Test quality
+    // test_quality();
+
+    // Test feature
+    // test_feature();
+    return 0;
+}
diff --git a/cpp-package/inspireface/cpp/sample/source/feature_hub_sample.cpp b/cpp-package/inspireface/cpp/sample/source/feature_hub_sample.cpp
new file mode 100644
index 000000000..029c68bbc
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/source/feature_hub_sample.cpp
@@ -0,0 +1,93 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/track_module/face_track_module.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+#include <inspireface/pipeline_module/face_pipeline_module.h>
+#include <inspireface/common/face_data/face_serialize_tools.h>
+#include <inspireface/include/inspireface/feature_hub_db.h>
+
+using namespace inspire;
+
+static std::vector<float> FT = {
+  0.0706566,   0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,
+  -0.00683422, -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,
+  0.0164787,   -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,
+  -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,
+  -0.00625798, 0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,
+  0.0465634,   0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,
+  -0.102007,   0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,
+  -0.00694243, 0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,
+  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,
+  0.0511028,   -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,
+  -0.0225608,  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,
+  0.056139,    -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,   0.0706566,   0.00640248,  0.0418103,   -0.00597861,
+  0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,   -0.00683422, -0.0338589,  0.0533989,   -0.0371725,
+  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,   0.0164787,   -0.00732871, -0.0458209,  -0.0100137,
+  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,
+  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,   -0.00625798, 0.00748682,  0.0533557,   0.0314002,
+  -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,   0.0627244,   0.0449248,   -0.037054,
+  -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,   -0.102007,   0.0649219,   0.0630833,   0.0421069,
+  0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,   -0.00694243, 0.0366775,   0.0672882,   0.0122419,
+  -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,
+  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,   0.0511028,   -0.0310699,  -0.0322141,  0.00996936,
+  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,  -0.0225608,  -0.0332198,  0.00102291,  0.108814,
+  -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,   0.056139,    -0.0166783,  -0.0803042,  -0.014245,
+  -0.0476374,  0.048495,    0.0378856,   0.0706566,   0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,
+  -0.0080779,  -0.0550556,  0.0229963,   -0.00683422, -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743,
+  -0.0128782,  0.0935529,   0.0588179,   0.0164787,   -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,
+  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,
+  0.00192449,  -0.0593105,  0.0191942,   -0.00625798, 0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,
+  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,   0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,
+  -0.0161973,  0.0319588,   0.0112792,   -0.102007,   0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,
+  0.071994,    -0.0272229,  0.0167103,   -0.00694243, 0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,
+  0.000983093, -0.00776073, -0.0268638,  0.00682446,  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,
+  0.000502882, 0.0496892,   0.0126823,   0.0511028,   -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,
+  0.0419618,   -0.00358901, -0.0309569,  -0.0225608,  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,
+  0.0721224,   -0.0795082,  0.0340047,   0.056139,    -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,   0.0706566,
+  0.00640248,  0.0418103,   -0.00597861, 0.0269879,   0.0187478,   0.0486305,   0.0349162,   -0.0080779,  -0.0550556,  0.0229963,   -0.00683422,
+  -0.0338589,  0.0533989,   -0.0371725,  0.000972469, 0.0612415,   0.0389846,   -0.00126743, -0.0128782,  0.0935529,   0.0588179,   0.0164787,
+  -0.00732871, -0.0458209,  -0.0100137,  -0.0372892,  0.000871123, 0.0245121,   -0.0811471,  -0.00481095, 0.0266868,   0.0712961,   -0.0675362,
+  -0.0117453,  0.0658745,   -0.0694139,  -0.00704822, -0.0237313,  0.0209365,   0.0131902,   0.00192449,  -0.0593105,  0.0191942,   -0.00625798,
+  0.00748682,  0.0533557,   0.0314002,   -0.0627113,  0.0827862,   0.00336722,  -0.0191575,  -0.0180252,  0.0150318,   -0.0686462,  0.0465634,
+  0.0627244,   0.0449248,   -0.037054,   -0.0486668,  0.040752,    0.0143315,   -0.0763842,  -0.0161973,  0.0319588,   0.0112792,   -0.102007,
+  0.0649219,   0.0630833,   0.0421069,   0.0519043,   -0.084082,   0.0249516,   0.023046,    0.071994,    -0.0272229,  0.0167103,   -0.00694243,
+  0.0366775,   0.0672882,   0.0122419,   -0.0233413,  -0.0144258,  -0.012853,   -0.0202025,  0.000983093, -0.00776073, -0.0268638,  0.00682446,
+  0.0262906,   -0.0407654,  -0.0144264,  -0.0310807,  0.0596711,   0.0238081,   -0.0138019,  0.000502882, 0.0496892,   0.0126823,   0.0511028,
+  -0.0310699,  -0.0322141,  0.00996936,  0.0675392,   -0.0164277,  0.0930009,   -0.037467,   0.0419618,   -0.00358901, -0.0309569,  -0.0225608,
+  -0.0332198,  0.00102291,  0.108814,    -0.0831313,  0.048208,    -0.0277542,  -0.061584,   0.0721224,   -0.0795082,  0.0340047,   0.056139,
+  -0.0166783,  -0.0803042,  -0.014245,   -0.0476374,  0.048495,    0.0378856,
+};
+
+int main() {
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Pikachu");
+
+    DatabaseConfiguration configuration;
+    configuration.primary_key_mode = PrimaryKeyMode::MANUAL_INPUT;
+    configuration.enable_persistence = false;
+    configuration.recognition_threshold = 0.48f;
+
+    INSPIREFACE_FEATURE_HUB->EnableHub(configuration);
+
+    // std::vector<float> feature(512, 0.0f);
+
+    int64_t result_id = 0;
+    auto ret = INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(FT, 10086, result_id);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to insert face feature");
+        INSPIRE_LOGI("result id: %lld", result_id);
+    } else {
+        INSPIRE_LOGI("Insert face feature success, result_id: %lld", result_id);
+    }
+
+    // std::vector<float> query_feature(512, 20.0f);
+    FaceSearchResult search_result;
+    ret = INSPIREFACE_FEATURE_HUB->SearchFaceFeature(FT, search_result, true);
+    if (ret != HSUCCEED) {
+        INSPIRE_LOGE("Failed to search face feature");
+    } else {
+        INSPIRE_LOGI("Search face feature success, result_id: %lld", search_result.id);
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/source/landmark_sample.cpp b/cpp-package/inspireface/cpp/sample/source/landmark_sample.cpp
new file mode 100644
index 000000000..e04aa37fe
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/source/landmark_sample.cpp
@@ -0,0 +1,35 @@
+#include <iostream>
+#include <inspirecv/inspirecv.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+#include "inspireface/track_module/landmark/face_landmark_adapt.h"
+int main() {
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Pikachu-t4");
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+
+    inspire::InspireModel lmkModel;
+    auto ret = archive.LoadModel("landmark", lmkModel);
+    if (ret != 0) {
+        INSPIRE_LOGE("Load %s error: %d", "landmark", ret);
+        return -1;
+    }
+
+    inspire::FaceLandmarkAdapt lmk;
+    lmk.LoadData(lmkModel, lmkModel.modelType);
+
+    auto image = inspirecv::Image::Create("test_res/data/crop/crop.png");
+    auto data = image.Resize(112, 112);
+    auto lmk_out = lmk(data);
+    std::vector<inspirecv::Point2i> landmarks_output(inspire::FaceLandmarkAdapt::NUM_OF_LANDMARK);
+    for (int i = 0; i < inspire::FaceLandmarkAdapt::NUM_OF_LANDMARK; ++i) {
+        float x = lmk_out[i * 2 + 0] * image.Width();
+        float y = lmk_out[i * 2 + 1] * image.Height();
+        landmarks_output[i] = inspirecv::Point<int>(x, y);
+    }
+
+    for (int i = 0; i < landmarks_output.size(); ++i) {
+        image.DrawCircle(landmarks_output[i], 5, {0, 0, 255});
+    }
+    image.Write("crop_lmk.png");
+}
diff --git a/cpp-package/inspireface/cpp/sample/source/tracker_pipeline.cpp b/cpp-package/inspireface/cpp/sample/source/tracker_pipeline.cpp
new file mode 100644
index 000000000..f449b9551
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/source/tracker_pipeline.cpp
@@ -0,0 +1,48 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/track_module/face_track_module.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+#include <inspireface/pipeline_module/face_pipeline_module.h>
+#include <inspireface/common/face_data/face_serialize_tools.h>
+
+using namespace inspire;
+
+int main() {
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Pikachu");
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    auto mode = inspire::DetectModuleMode::DETECT_MODE_LIGHT_TRACK;
+    FaceTrackModule tracker(mode, 10, 20, 320, -1);
+    tracker.Configuration(archive, expansion_path);
+
+    FacePipelineModule pipe(archive, true, true, true, true, true);
+
+    auto image = inspirecv::Image::Create("test_res/data/bulk/r90.jpg");
+    inspirecv::FrameProcess processor;
+    processor.SetDataBuffer(image.Data(), image.Height(), image.Width());
+    processor.SetDataFormat(inspirecv::DATA_FORMAT::BGR);
+    processor.SetRotationMode(inspirecv::ROTATION_MODE::ROTATION_90);
+    std::vector<FaceProcessFunctionOption> methods = {PROCESS_MASK, PROCESS_RGB_LIVENESS, PROCESS_ATTRIBUTE, PROCESS_INTERACTION};
+    for (int i = 0; i < 1; i++) {
+        auto show = image.Clone();
+        tracker.UpdateStream(processor);
+        auto faces = tracker.trackingFace;
+        int index = 0;
+        if (faces.size() > 0) {
+            auto &face = faces[index];
+            auto hyper_face = FaceObjectInternalToHyperFaceData(face);
+            PrintHyperFaceDataDetail(hyper_face);
+
+            std::cout << face.getTransMatrix() << std::endl;
+
+            for (auto method : methods) {
+                pipe.Process(processor, hyper_face, method);
+            }
+            std::cout << "eyes status: " << pipe.eyesStatusCache[0] << " " << pipe.eyesStatusCache[1] << std::endl;
+        }
+
+        // show.Show("faces", );
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/source/tracker_sample.cpp b/cpp-package/inspireface/cpp/sample/source/tracker_sample.cpp
new file mode 100644
index 000000000..78ed2918b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/source/tracker_sample.cpp
@@ -0,0 +1,39 @@
+#include <inspirecv/inspirecv.h>
+#include <inspireface/track_module/face_track_module.h>
+#include <inspireface/include/inspireface/launch.h>
+#include <inspireface/include/inspireface/frame_process.h>
+
+using namespace inspire;
+
+int main() {
+    std::string expansion_path = "";
+    INSPIREFACE_CONTEXT->Load("test_res/pack/Pikachu");
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    auto mode = inspire::DetectModuleMode::DETECT_MODE_ALWAYS_DETECT;
+    FaceTrackModule tracker(mode, 10, 20, 320, -1);
+    tracker.Configuration(archive, expansion_path);
+
+    auto image = inspirecv::Image::Create("test_res/data/bulk/r0.jpg");
+    inspirecv::FrameProcess processor;
+    processor.SetDataBuffer(image.Data(), image.Height(), image.Width());
+    processor.SetDataFormat(inspirecv::DATA_FORMAT::BGR);
+    processor.SetRotationMode(inspirecv::ROTATION_MODE::ROTATION_0);
+    for (int i = 0; i < 100; i++) {
+        auto show = image.Clone();
+        tracker.UpdateStream(processor);
+        auto faces = tracker.trackingFace;
+        int index = 0;
+        if (faces.size() > 0) {
+            auto &face = faces[index];
+            for (auto &p : face.high_result.lmk) {
+                show.DrawCircle(p.As<int>(), 5, {0, 255, 0});
+            }
+            for (auto &p : face.landmark_) {
+                show.DrawCircle(p.As<int>(), 5, {0, 0, 255});
+            }
+        }
+        show.Show("faces", 0);
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/sample/utils/test_folder_helper.h b/cpp-package/inspireface/cpp/sample/utils/test_folder_helper.h
new file mode 100755
index 000000000..6963ad88d
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/utils/test_folder_helper.h
@@ -0,0 +1,176 @@
+//
+// Created by YH-Mac on 2020/9/12.
+//
+
+#ifndef LMKTRACKING_LIB_TEST_FOLDER_HELPER_H
+#define LMKTRACKING_LIB_TEST_FOLDER_HELPER_H
+
+#include <dirent.h>
+#include <iostream>
+#include <fstream>
+#include <stdio.h>
+#include <stack>
+#include <stdlib.h>
+#include <string.h>
+#include <dirent.h>
+#include <unistd.h>
+#include <vector>
+#include <sstream>
+#include "sys/types.h"
+#include "sys/stat.h"
+#include <chrono>
+//using namespace std;
+#define MODE (S_IRWXU | S_IRWXG | S_IRWXO)
+
+using namespace std::chrono;
+class Timer
+{
+public:
+    std::stack<high_resolution_clock::time_point> tictoc_stack;
+
+    void tic()
+    {
+        high_resolution_clock::time_point t1 = high_resolution_clock::now();
+        tictoc_stack.push(t1);
+    }
+
+    double toc(std::string msg = "", bool flag = true)
+    {
+        double diff = duration_cast<milliseconds>(high_resolution_clock::now() - tictoc_stack.top()).count();
+        if(msg.size() > 0){
+            if (flag)
+                printf("%s time elapsed: %f ms\n", msg.c_str(), diff);
+        }
+        if(msg == "ir")
+          return diff/3;
+
+        tictoc_stack.pop();
+        return diff;
+    }
+    void reset()
+    {
+        tictoc_stack = std::stack<high_resolution_clock::time_point>();
+    }
+};
+
+std::string dePrefix(std::string file_name, std::string prefix){
+    int n = file_name.find_last_of(prefix);
+    std::string de = file_name.substr(n + 1, file_name.size());
+    return de;
+}
+
+std::string deSuffix(std::string file_name, std::string suffix){
+    int n = file_name.find_last_of(suffix);
+    std::string de = file_name.substr(0, n);
+    return de;
+}
+
+bool check_folder(const std::string folder_path, bool is_create) {
+    int res = access(folder_path.c_str(), 0);
+//    std::cout << "res " << res << std::endl;
+    if (res != -1) {
+        std::cout << folder_path << " exists." << std::endl;
+        return true;
+    } else if(is_create){
+        std::cout << folder_path << " not exists." << std::endl;
+        int mkdir_res = mkdir(folder_path.c_str(), MODE);
+        if (mkdir_res != -1) {
+            std::cout << "mkdir successful." << std::endl;
+        } else {
+            std::cout << "mkdir fail." << std::endl;
+        }
+        return true;
+    } else{
+        std::cout << folder_path << " not exists." << std::endl;
+        return false;
+    }
+}
+
+std::vector<float> read_feat_txt(const std::string feat_path){
+    std::ifstream infile;
+    infile.open(feat_path);
+    if (!infile)
+        std::cout << "file error: " << feat_path << std::endl;
+    std::vector<float> feat;
+    float t1;
+    while (infile >> t1) {
+        feat.push_back(t1);
+    }
+    infile.close();
+    return feat;
+}
+
+void extract_feat_to_txt(const std::string feat_path, std::vector<float> feat) {
+    std::ofstream txt(feat_path);
+    for (int i = 0; i < feat.size(); ++i) {
+        txt << feat[i] << " ";
+    }
+    txt.close();
+//    std::cout << "export feature to " << feat_path << std::endl;
+}
+
+bool decide_ext(const char *gname, const char *nsuff)
+// gname: name of the given file
+// nsuff: suffix  you need
+{
+    char dot = '.';
+    char suff[10] = {0};
+    int c;
+    int j = 0;
+    c = strlen(gname);
+//    std::cout<< c << '\n';
+    for (int i = 0; i < c; i++) {
+        if (gname[i] == dot)
+            j = i;
+    }
+    int k = j;
+    j = c - j - 1;
+    for (int i = 0; i < j; i++) {
+        suff[i] = gname[k + i + 1];
+    }
+    if (0 == strcmp(suff, nsuff))
+        return true;
+    else
+        return false;
+}
+
+
+std::vector<std::string>
+readFileListSuffix(const char *basePath, const char *suffix, std::vector<std::string>& file_names, bool recursive) {
+    DIR *dir;
+    struct dirent *ptr;
+    char base[1000];
+    std::vector<std::string> file_list;
+    if ((dir = opendir(basePath)) == NULL) {
+        perror("Open dir error...");
+        exit(1);
+    }
+
+    while ((ptr = readdir(dir)) != NULL) {
+        if (strcmp(ptr->d_name, ".") == 0 || strcmp(ptr->d_name, "..") == 0)    ///current dir OR parrent dir
+            continue;
+        else if (ptr->d_type == 8) {    ///file
+//            printf("d_name:%s/%s\n", basePath, ptr->d_name);
+            if (decide_ext(ptr->d_name, suffix)) {
+                std::stringstream ss;
+                ss << basePath << "/" << ptr->d_name;
+                std::string path = ss.str();
+                file_names.push_back(ptr->d_name);
+                file_list.push_back(path);
+            }
+        } else if (ptr->d_type == 10) {    ///link file
+//            printf("d_name:%s/%s\n", basePath, ptr->d_name);
+        } else if (ptr->d_type == 4 and recursive)    ///dir
+        {
+            memset(base, '\0', sizeof(base));
+            strcpy(base, basePath);
+            strcat(base, "/");
+            strcat(base, ptr->d_name);
+            readFileListSuffix(base, suffix, file_names, recursive);
+        }
+    }
+    closedir(dir);
+    return file_list;
+}
+
+#endif //LMKTRACKING_LIB_TEST_FOLDER_HELPER_H
diff --git a/cpp-package/inspireface/cpp/sample/utils/test_helper.h b/cpp-package/inspireface/cpp/sample/utils/test_helper.h
new file mode 100755
index 000000000..8bf554f0b
--- /dev/null
+++ b/cpp-package/inspireface/cpp/sample/utils/test_helper.h
@@ -0,0 +1,122 @@
+//
+// Created by Jack YU on 2020/6/11.
+//
+
+#ifndef ZEUSEESTRACKING_LIB_TEST_HELPER_H
+#define ZEUSEESTRACKING_LIB_TEST_HELPER_H
+
+#include <dirent.h>
+#include <sys/stat.h>
+
+#include "middleware/camera_stream/camera_stream.h"
+
+namespace TestUtils {
+
+inline uint8_t *rgb2nv21(const cv::Mat &Img) {
+  if (Img.empty()) {
+    exit(0);
+  }
+  int cols = Img.cols;
+  int rows = Img.rows;
+
+  int Yindex = 0;
+  int UVindex = rows * cols;
+
+  unsigned char *yuvbuff =
+      new unsigned char[static_cast<int>(1.5 * rows * cols)];
+
+  cv::Mat NV21(rows + rows / 2, cols, CV_8UC1);
+  cv::Mat OpencvYUV;
+  cv::Mat OpencvImg;
+  cv::cvtColor(Img, OpencvYUV, cv::COLOR_BGR2YUV_YV12);
+  for (int i = 0; i < rows; i++) {
+    for (int j = 0; j < cols; j++) {
+      uchar *YPointer = NV21.ptr<uchar>(i);
+      int B = Img.at<cv::Vec3b>(i, j)[0];
+      int G = Img.at<cv::Vec3b>(i, j)[1];
+      int R = Img.at<cv::Vec3b>(i, j)[2];
+      int Y = (77 * R + 150 * G + 29 * B) >> 8;
+      YPointer[j] = Y;
+      yuvbuff[Yindex++] = (Y < 0) ? 0 : ((Y > 255) ? 255 : Y);
+      uchar *UVPointer = NV21.ptr<uchar>(rows + i / 2);
+      if (i % 2 == 0 && (j) % 2 == 0) {
+        int U = ((-44 * R - 87 * G + 131 * B) >> 8) + 128;
+        int V = ((131 * R - 110 * G - 21 * B) >> 8) + 128;
+        UVPointer[j] = V;
+        UVPointer[j + 1] = U;
+        yuvbuff[UVindex++] = (V < 0) ? 0 : ((V > 255) ? 255 : V);
+        yuvbuff[UVindex++] = (U < 0) ? 0 : ((U > 255) ? 255 : U);
+      }
+    }
+  }
+  return yuvbuff;
+}
+
+inline void rotate(const cv::Mat &image, cv::Mat &out,
+                   inspire::ROTATION_MODE mode) {
+  if (mode == inspire::ROTATION_90) {
+    cv::transpose(image, out);
+    cv::flip(out, out, 2);
+  } else if (mode == inspire::ROTATION_180) {
+    cv::flip(out, out, -1);
+  } else if (mode == inspire::ROTATION_270) {
+    cv::transpose(image, out);
+    cv::flip(out, out, 0);
+  }
+}
+
+
+    void GetFilesInDirectory(std::vector<std::string> &out, const std::string &directory)
+    {
+#ifdef WINDOWS
+        HANDLE dir;
+    WIN32_FIND_DATA file_data;
+
+    if ((dir = FindFirstFile((directory + "/*").c_str(), &file_data)) == INVALID_HANDLE_VALUE)
+        return; /* No files found */
+
+    do {
+        const string file_name = file_data.cFileName;
+        const string full_file_name = directory + "/" + file_name;
+        const bool is_directory = (file_data.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY) != 0;
+
+        if (file_name[0] == '.')
+            continue;
+
+        if (is_directory)
+            continue;
+
+        out.push_back(full_file_name);
+    } while (FindNextFile(dir, &file_data));
+
+    FindClose(dir);
+#else
+        DIR *dir;
+        class dirent *ent;
+        class stat st;
+
+        dir = opendir(directory.c_str());
+        while ((ent = readdir(dir)) != NULL) {
+            const std::string file_name = ent->d_name;
+            const std::string full_file_name = directory + "/" + file_name;
+
+            if (file_name[0] == '.')
+                continue;
+
+            if (stat(full_file_name.c_str(), &st) == -1)
+                continue;
+
+            const bool is_directory = (st.st_mode & S_IFDIR) != 0;
+
+            if (is_directory)
+                continue;
+
+            out.push_back(full_file_name);
+        }
+        closedir(dir);
+#endif
+    }
+
+} // namespace TestUtils
+
+#endif // ZEUSEESTRACKING_LIB_TEST_HELPER_H
diff --git a/cpp-package/inspireface/cpp/test/CMakeLists.txt b/cpp-package/inspireface/cpp/test/CMakeLists.txt
new file mode 100644
index 000000000..a6391d3ef
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/CMakeLists.txt
@@ -0,0 +1,112 @@
+cmake_minimum_required(VERSION 3.20)
+project(InspireFaceTest)
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+# If you want to test the benchmark, you need to set this to ON
+option(ISF_ENABLE_BENCHMARK "Enable the benchmark test cases." ON)
+# If you want to test the LFW data set, you need to set this to ON
+option(ISF_ENABLE_USE_LFW_DATA "Enable test cases for LFW data sets." OFF)
+# If you want to test the evaluation function, you need to set this to ON, need LFW data set.
+option(ISF_ENABLE_TEST_EVALUATION "Enable evaluation function test cases." OFF)
+# If you want to test the base test cases, you need to set this to ON, if you want to hide the symbols, you need to set ISF_ENABLE_SYMBOL_HIDING to ON
+option(ISF_ENABLE_TEST_INTERNAL "Enable internal test cases." OFF)
+
+if (ISF_ENABLE_BENCHMARK)
+    add_definitions("-DISF_ENABLE_BENCHMARK")
+endif ()
+
+if (ISF_ENABLE_USE_LFW_DATA)
+    add_definitions("-DISF_ENABLE_USE_LFW_DATA")
+endif ()
+
+if (ISF_ENABLE_TEST_EVALUATION)
+    add_definitions("-DISF_ENABLE_TEST_EVALUATION")
+endif ()
+
+if (ISF_ENABLE_RKNN)
+    if (ISF_RKNPU_MAJOR STREQUAL "rknpu1")
+    set(DEPEND rknn_api dl)
+    set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${ISF_RK_DEVICE_TYPE}/Linux/librknn_api/${CPU_ARCH}/)
+        message("Enable RKNN Inference")
+        link_directories(${ISF_RKNN_API_LIB})
+        set(DEPEND rknn_api dl)
+    elseif(ISF_RKNPU_MAJOR STREQUAL "rknpu2" AND ISF_RK_COMPILER_TYPE STREQUAL "aarch64")
+        set(DEPEND rknnrt dl)
+        if(ANDROID)
+        set(RK_PLATFORM "Android")
+        else()
+            set(RK_PLATFORM "Linux")
+        endif()
+        set(ISF_RKNN_API_LIB ${ISF_THIRD_PARTY_DIR}/inspireface-precompile-lite/rknn/${ISF_RKNPU_MAJOR}/runtime/${RK_PLATFORM}/librknn_api/${ISF_RK_COMPILER_TYPE}/)
+        message("ISF_RKNN_API_LIB: ${ISF_RKNN_API_LIB}")
+        link_directories(${ISF_RKNN_API_LIB})
+        set(DEPEND rknnrt dl)
+    else()
+        set(DEPEND dl)
+    endif()
+endif ()
+
+include_directories(${SRC_DIR})
+
+# =======================Common File Configuration===========================
+set(TEST_COMMON_FILES ${CMAKE_CURRENT_SOURCE_DIR}/settings/test_settings.cpp)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR}/settings)
+
+# =======================Internal Base Import Tests===========================
+if (ISF_ENABLE_TEST_INTERNAL)
+    file(GLOB_RECURSE INTERNAL_TEST_INTERNAL_FILES unit/base/*.cpp)
+    add_executable(TestInternal ${CMAKE_CURRENT_SOURCE_DIR}/test_base.cpp ${INTERNAL_TEST_INTERNAL_FILES} ${TEST_COMMON_FILES})
+    target_link_libraries(TestInternal InspireFace ${DEPEND})
+    target_include_directories(TestInternal PUBLIC
+            ${CMAKE_CURRENT_SOURCE_DIR}/
+            ${ISF_THIRD_PARTY_DIR}/spdlog/include
+            ${ISF_THIRD_PARTY_DIR}/Catch2/single_include/
+            ${ISF_THIRD_PARTY_DIR}/indicators/include/
+            ${SRC_DIR}
+            ${INSPIRECV_INCLUDE_PATH}
+    )
+
+    set_target_properties(TestInternal PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/test/")
+endif()
+
+# =======================External API Testing===========================
+file(GLOB_RECURSE TEST_INTERNAL_FILES unit/api/*.cpp)
+add_executable(Test ${CMAKE_CURRENT_SOURCE_DIR}/test.cpp ${TEST_INTERNAL_FILES} ${TEST_COMMON_FILES})
+target_link_libraries(Test InspireFace ${DEPEND})
+target_include_directories(Test PUBLIC
+        ${CMAKE_CURRENT_SOURCE_DIR}/
+        ${ISF_THIRD_PARTY_DIR}/spdlog/include
+        ${ISF_THIRD_PARTY_DIR}/Catch2/single_include/
+        ${ISF_THIRD_PARTY_DIR}/indicators/include/
+        ${SRC_DIR}
+        )
+
+set_target_properties(Test PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/test/")
+
+# =======================CPP API Testing===========================
+file(GLOB_RECURSE TEST_CPP_FILES unit/cpp_api/*.cpp)
+add_executable(TestCPP ${CMAKE_CURRENT_SOURCE_DIR}/test_cpp.cpp ${TEST_CPP_FILES} ${TEST_COMMON_FILES})
+target_link_libraries(TestCPP InspireFace ${DEPEND})
+target_include_directories(TestCPP PUBLIC
+        ${CMAKE_CURRENT_SOURCE_DIR}/
+        ${ISF_THIRD_PARTY_DIR}/spdlog/include
+        ${ISF_THIRD_PARTY_DIR}/Catch2/single_include/
+        ${ISF_THIRD_PARTY_DIR}/indicators/include/
+        ${SRC_DIR}
+        )
+
+set_target_properties(TestCPP PROPERTIES
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/test/")
+
+# Print Message
+message(STATUS ">>>>>>>>>>>>>")
+message(STATUS "InspireFace Test:")
+message(STATUS "\t ISF_ENABLE_BENCHMARK: ${ISF_ENABLE_BENCHMARK}")
+message(STATUS "\t ISF_ENABLE_USE_LFW_DATA: ${ISF_ENABLE_USE_LFW_DATA}")
+message(STATUS "\t ISF_ENABLE_TEST_EVALUATION: ${ISF_ENABLE_TEST_EVALUATION}")
+
+# Install bin
+install(TARGETS Test RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/test)
+# install(TARGETS TestSource RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/test)
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/settings/check.h b/cpp-package/inspireface/cpp/test/settings/check.h
new file mode 100644
index 000000000..e7663506a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/settings/check.h
@@ -0,0 +1,56 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-14
+ */
+#pragma once
+#ifndef INSPIREFACE_TEST_CHECK_
+#define INSPIREFACE_TEST_CHECK_
+
+#include <cstdint>  // for uint8_t
+#include <limits>   // for std::numeric_limits
+#include <sstream>
+
+#define REQUIRE_EQ_IMAGE(a, b, h, w, c)                                                                                                \
+    do {                                                                                                                               \
+        double eps = 0.01;                                                                                                             \
+        double mse = CalculateImageMSE(a, b, h, w, c);                                                                                 \
+        REQUIRE(mse <= eps);                                                                                                           \
+        if (mse > eps) {                                                                                                               \
+            std::stringstream ss;                                                                                                      \
+            ss << "Image comparison failed! MSE: " << mse << " (threshold: " << eps << "), dimensions: " << h << "x" << w << "x" << c; \
+            INFO(ss.str());                                                                                                            \
+        }                                                                                                                              \
+    } while (0)
+
+#define REQUIRE_EQ_IMAGE_WITH_EPS(a, b, h, w, c, eps)                                                                                  \
+    do {                                                                                                                               \
+        double mse = CalculateImageMSE(a, b, h, w, c);                                                                                 \
+        REQUIRE(mse <= eps);                                                                                                           \
+        if (mse > eps) {                                                                                                               \
+            std::stringstream ss;                                                                                                      \
+            ss << "Image comparison failed! MSE: " << mse << " (threshold: " << eps << "), dimensions: " << h << "x" << w << "x" << c; \
+            INFO(ss.str());                                                                                                            \
+        }                                                                                                                              \
+    } while (0)
+
+inline double CalculateImageMSE(const uint8_t* a, const uint8_t* b, int h, int w, int c) {
+    if (a == nullptr || b == nullptr || h <= 0 || w <= 0 || c <= 0) {
+        return std::numeric_limits<double>::infinity();
+    }
+
+    double sum_squared_diff = 0.0;
+    size_t total_pixels = static_cast<size_t>(h) * w * c;
+    const double normalize_factor = 255.0;
+
+    for (size_t i = 0; i < total_pixels; ++i) {
+        double a_normalized = static_cast<double>(a[i]) / normalize_factor;
+        double b_normalized = static_cast<double>(b[i]) / normalize_factor;
+
+        double diff = a_normalized - b_normalized;
+        sum_squared_diff += diff * diff;
+    }
+
+    return sum_squared_diff / total_pixels;
+}
+
+#endif  // INSPIREFACE_TEST_CHECK_
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/settings/enviro.h b/cpp-package/inspireface/cpp/test/settings/enviro.h
new file mode 100644
index 000000000..339f97840
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/settings/enviro.h
@@ -0,0 +1,60 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef INSPIREFACE_ENVIRO_H
+#define INSPIREFACE_ENVIRO_H
+
+#include <string>
+
+/**
+ * @class Enviro
+ * @brief A singleton class for managing environment-specific configurations.
+ *
+ * This class provides methods to set and get various environment-specific configurations
+ * such as the resource pack name, test result directory, and runtime full path.
+ */
+class Enviro {
+public:
+    static Enviro &getInstance() {
+        static Enviro instance;
+        return instance;
+    }
+
+    Enviro(Enviro const &) = delete;
+    void operator=(Enviro const &) = delete;
+
+    std::string getPackName() const {
+        return packName;
+    }
+
+    void setPackName(const std::string &name) {
+        packName = name;
+    }
+
+    const std::string &getTestResDir() const {
+        return testResDir;
+    }
+
+    void setTestResDir(const std::string &dir) {
+        Enviro::testResDir = dir;
+    }
+
+    const std::string &getTestRuntimeFullPath() const {
+        return runtimeFullPath;
+    }
+
+    void setTestRuntimeFullPath(const std::string &path) {
+        Enviro::runtimeFullPath = path;
+    }
+
+private:
+    Enviro() {}
+
+    std::string packName{"Pikachu"};
+    std::string testResDir{"test_res"};
+    std::string runtimeFullPath;
+};
+
+#endif  // INSPIREFACE_ENVIRO_H
diff --git a/cpp-package/inspireface/cpp/test/settings/test_settings.cpp b/cpp-package/inspireface/cpp/test/settings/test_settings.cpp
new file mode 100644
index 000000000..e1d1a61f4
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/settings/test_settings.cpp
@@ -0,0 +1,61 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include "test_settings.h"
+
+std::string getTestDataDir() {
+    return Enviro::getInstance().getTestResDir();
+}
+
+std::string getTestData(const std::string& name) {
+    return getTestDataDir() + "/" + name;
+}
+
+std::string getTestSaveDir() {
+    return getTestData("save");
+}
+
+std::string getTestSaveData(const std::string& name) {
+    return getTestSaveDir() + "/" + name;
+}
+
+std::string getTestModelsFile() {
+    std::string path = "pack/";
+    path = path + TEST_MODEL_FILE;
+    return getTestData(path);
+}
+
+std::string getTestLFWFunneledTxt() {
+    return getTestData(TEST_LFW_FUNNELED_TXT);
+}
+
+std::string getTestLFWFunneledEvaTxt() {
+    return getLFWFunneledDir() + "/" + TEST_LFW_EVALUATION_TXT;
+}
+
+std::string getLFWFunneledDir() {
+    const char* testLfwFunneledDir = LFW_FUNNELED_DIR;
+    std::string lfwFunneledPath;
+    if (testLfwFunneledDir != nullptr && std::string(testLfwFunneledDir) != "") {
+        lfwFunneledPath = testLfwFunneledDir;
+    } else {
+        const char* lfwFunneled = std::getenv("LFW_FUNNELED_DIR");
+        if (lfwFunneled != nullptr) {
+            lfwFunneledPath = lfwFunneled;
+        }
+    }
+    if (lfwFunneledPath.empty()) {
+        spdlog::warn("lfw funneled dir is empty!");
+    }
+    return lfwFunneledPath;
+}
+
+std::string getBenchmarkRecordFile() {
+    return getTestSaveData(TEST_BENCHMARK_RECORD);
+}
+
+std::string getEvaluationRecordFile() {
+    return getTestSaveData(TEST_EVALUATION_RECORD);
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/settings/test_settings.h b/cpp-package/inspireface/cpp/test/settings/test_settings.h
new file mode 100644
index 000000000..28548b2af
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/settings/test_settings.h
@@ -0,0 +1,158 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma once
+#ifndef BIGGUYSMAIN_TEST_SETTINGS_H
+#define BIGGUYSMAIN_TEST_SETTINGS_H
+#include <catch2/catch.hpp>
+#include <spdlog/spdlog.h>
+#include <iostream>
+#include "enviro.h"
+#include "check.h"
+#include "inspireface/middleware/system.h"
+#include "inspireface/include/inspireface/spend_timer.h"
+
+// Define the test model file
+#define TEST_MODEL_FILE Enviro::getInstance().getPackName()
+
+// Set the pack name
+#define SET_PACK_NAME(name) Enviro::getInstance().setPackName(name)
+
+// Set the test directory
+#define SET_TEST_DIR(dir) Enviro::getInstance().setTestResDir(dir)
+// Set the runtime full path
+#define SET_RUNTIME_FULLPATH_NAME(name) Enviro::getInstance().setTestRuntimeFullPath(name)
+// Get the runtime full path
+#define GET_RUNTIME_FULLPATH_NAME Enviro::getInstance().getTestRuntimeFullPath()
+
+// Define the LFW funneled index txt file
+#define TEST_LFW_FUNNELED_TXT "valid_lfw_funneled.txt"
+// Define the LFW funneled data directory
+#define LFW_FUNNELED_DIR ""
+// Define the LFW evaluation txt file
+#define TEST_LFW_EVALUATION_TXT "pairs.txt"
+// Define the benchmark record file
+#define TEST_BENCHMARK_RECORD "benchmark.csv"
+// Define the evaluation record file
+#define TEST_EVALUATION_RECORD "evaluation.csv"
+
+using namespace Catch::Detail;
+
+// Print test message
+#define TEST_PRINT(...) SPDLOG_LOGGER_CALL(spdlog::get("TEST"), spdlog::level::trace, __VA_ARGS__)
+// Print test message output
+#define TEST_PRINT_OUTPUT(open) TestMessageBroadcast test_msg_broadcast_##open(open)
+// Set the log output level
+#define LOG_OUTPUT_LEVEL(level) LogLevelBroadcast log_level_broadcast_##level(level);
+// Print test error message
+#define TEST_ERROR_PRINT(...) SPDLOG_LOGGER_CALL(spdlog::get("TEST"), spdlog::level::err, __VA_ARGS__)
+
+// Get the test data directory
+#define GET_DIR getTestDataDir()
+// Get the test data
+#define GET_DATA(filename) getTestData(filename)
+// Get the test models file
+#define GET_MODEL_FILE() getTestModelsFile()
+// Get the LFW funneled index txt file
+#define GET_LFW_FUNNELED_TXT() getTestLFWFunneledTxt()
+// Get the LFW funneled evaluation txt file
+#define GET_LFW_FUNNELED_EVA_TXT() getTestLFWFunneledEvaTxt()
+// Get the LFW funneled directory
+#define GET_LFW_FUNNELED_DIR() getLFWFunneledDir()
+// Get the benchmark record file
+#define GET_BENCHMARK_RECORD_FILE() getBenchmarkRecordFile()
+// Get the evaluation record file
+#define GET_SAVE_DIR getTestSaveDir()
+// Get the test save data
+#define GET_SAVE_DATA(filename) getTestSaveData(filename)
+
+std::string getTestDataDir();
+
+std::string getTestData(const std::string &name);
+
+std::string getTestSaveDir();
+
+std::string getTestSaveData(const std::string &name);
+
+std::string getTestModelsFile();
+
+std::string getTestLFWFunneledTxt();
+
+std::string getTestLFWFunneledEvaTxt();
+
+std::string getLFWFunneledDir();
+
+std::string getBenchmarkRecordFile();
+
+std::string getEvaluationRecordFile();
+
+/** Logger level */
+enum LOG_LEVEL {
+    TRACE = 0,  ///< trace
+    DEBUG = 1,  ///< debug
+    INFO = 2,   ///< information
+    WARN = 3,   ///< warn
+    ERROR = 4,  ///< error
+    FATAL = 5,  ///< fatal
+    OFF = 6,    ///< off
+};
+
+class TestMessageBroadcast {
+public:
+    explicit TestMessageBroadcast(bool open) : m_old_level(spdlog::level::trace) {
+        auto logger = spdlog::get("TEST");
+        m_old_level = logger->level();
+        if (open) {
+            if (m_old_level != spdlog::level::trace) {
+                logger->set_level(spdlog::level::trace);
+            }
+        } else {
+            if (m_old_level == spdlog::level::trace) {
+                logger->set_level(spdlog::level::info);
+            }
+        }
+    }
+
+    ~TestMessageBroadcast() {
+        spdlog::get("TEST")->set_level(m_old_level);
+    }
+
+private:
+    spdlog::level::level_enum m_old_level;
+};
+
+/**
+ * @class LogLevelBroadcast
+ * @brief A class for broadcasting log level changes.
+ *
+ * This class is used to set the log level for the logger.
+ */
+class LogLevelBroadcast {
+public:
+    /**
+     * @brief Constructor for LogLevelBroadcast.
+     * @param level The log level to set.
+     */
+    explicit LogLevelBroadcast(LOG_LEVEL level) {
+        m_oldLevel = level;
+        spdlog::set_level(spdlog::level::level_enum(level));
+    }
+
+    ~LogLevelBroadcast() {
+        spdlog::set_level(spdlog::level::level_enum(spdlog::level::trace));
+    }
+
+private:
+    LOG_LEVEL m_oldLevel;
+};
+
+struct test_case_split {
+    ~test_case_split() {
+        std::cout << "===============================================================================" << std::endl;
+    }
+
+#define DRAW_SPLIT_LINE test_case_split split_line_x;
+};
+
+#endif  // BIGGUYSMAIN_TEST_SETTINGS_H
diff --git a/cpp-package/inspireface/cpp/test/test.cpp b/cpp-package/inspireface/cpp/test/test.cpp
new file mode 100644
index 000000000..2f0c14b0e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/test.cpp
@@ -0,0 +1,135 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <string>
+#define CATCH_CONFIG_RUNNER
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include "spdlog/spdlog.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/simple_csv_writer.h"
+
+int init_test_logger() {
+    std::string name("TEST");
+    auto stdout_sink = std::make_shared<spdlog::sinks::stdout_color_sink_mt>();
+    auto logger = std::make_shared<spdlog::logger>(name, stdout_sink);
+#if ENABLE_TEST_MSG
+    logger->set_level(spdlog::level::trace);
+#else
+    logger->set_level(spdlog::level::off);
+#endif
+    logger->set_pattern("%Y-%m-%d %H:%M:%S.%e [Test Message] ===> %v");
+    spdlog::register_logger(logger);
+    return 0;
+}
+
+int init_test_benchmark_record() {
+#ifdef ISF_ENABLE_BENCHMARK
+    if (std::remove(getBenchmarkRecordFile().c_str()) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    BenchmarkRecord record(getBenchmarkRecordFile(), TEST_MODEL_FILE);
+#endif
+    return 0;
+}
+
+int init_test_evaluation_record() {
+#ifdef ISF_ENABLE_TEST_EVALUATION
+    if (std::remove(getEvaluationRecordFile().c_str()) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    EvaluationRecord record(getEvaluationRecordFile());
+#endif
+    return 0;
+}
+
+int main(int argc, char* argv[]) {
+    init_test_logger();
+    init_test_benchmark_record();
+    init_test_evaluation_record();
+    TEST_PRINT_OUTPUT(true);
+
+    Catch::Session session;
+    // Pack file name and test directory
+    std::string pack;
+    std::string testDir;
+    std::string packPath;
+
+    HInt32 ret;
+
+    // Add command line options
+    auto cli = session.cli() | Catch::clara::Opt(pack, "value")["--pack"]("Resource pack filename") |
+               Catch::clara::Opt(testDir, "value")["--test_dir"]("Test dir resource") |
+               Catch::clara::Opt(packPath, "value")["--pack_path"]("The specified path to the pack file");
+
+    // Set combined CLI to the session
+    session.cli(cli);
+
+    // Parse command line arguments
+    int returnCode = session.applyCommandLine(argc, argv);
+    if (returnCode != 0)  // Indicate an error
+        return returnCode;
+
+    if (!testDir.empty()) {
+        SET_TEST_DIR(testDir);
+        TEST_PRINT("Updated test dir to: {}", getTestDataDir());
+    } else {
+        TEST_PRINT("Using default test dir: {}", getTestDataDir());
+    }
+
+#if defined(ISF_ENABLE_TENSORRT)
+    HInt32 support_cuda;
+    ret = HFCheckCudaDeviceSupport(&support_cuda);
+    if (ret != HSUCCEED) {
+        TEST_ERROR_PRINT("An error occurred while checking CUDA device support: {}", ret);
+        return ret;
+    }
+    if (!support_cuda) {
+        TEST_ERROR_PRINT("CUDA device support is not available");
+        return HERR_DEVICE_CUDA_NOT_SUPPORT;
+    }
+
+    HFPrintCudaDeviceInfo();
+#endif
+
+    std::string fullPath;
+    // Check whether custom parameters are set
+    if (!pack.empty()) {
+        SET_PACK_NAME(pack);
+        fullPath = GET_MODEL_FILE();
+        TEST_PRINT("Updated global Pack to: {}", TEST_MODEL_FILE);
+        SET_RUNTIME_FULLPATH_NAME(fullPath);
+    } else if (!packPath.empty()) {
+        fullPath = packPath;
+        TEST_PRINT("Updated global Pack File to: {}", packPath);
+        SET_RUNTIME_FULLPATH_NAME(packPath);
+    } else {
+        fullPath = GET_MODEL_FILE();
+        TEST_PRINT("Using default global Pack: {}", TEST_MODEL_FILE);
+        SET_RUNTIME_FULLPATH_NAME(fullPath);
+    }
+
+    TEST_PRINT("Launching InspireFace with path: {}", fullPath);
+    ret = HFLaunchInspireFace(fullPath.c_str());
+    if (ret != HSUCCEED) {
+        TEST_ERROR_PRINT("An error occurred while starting InspireFace: {}", ret);
+        return ret;
+    }
+    
+    // Set log level
+    HFSetLogLevel(HF_LOG_INFO);
+
+    // Run the test
+    ret = session.run();
+
+    // Terminate the InspireFace instance
+    HFTerminateInspireFace();
+
+    // Show resource statistics
+    HFDeBugShowResourceStatistics();
+
+    return ret;
+}
diff --git a/cpp-package/inspireface/cpp/test/test_base.cpp b/cpp-package/inspireface/cpp/test/test_base.cpp
new file mode 100644
index 000000000..2fc8ccc4e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/test_base.cpp
@@ -0,0 +1,40 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#define CATCH_CONFIG_RUNNER
+#include <iostream>
+#include "settings/test_settings.h"
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include "spdlog/spdlog.h"
+#include <inspireface/include/inspireface/launch.h>
+
+#define ENABLE_DRAW_SPLIT_LINE 1  // Whether dividers are printed during the test
+#define ENABLE_TEST_MSG 1         // TEST PRINT output
+
+int init_test_logger() {
+    std::string name("TEST");
+    auto stdout_sink = std::make_shared<spdlog::sinks::stdout_color_sink_mt>();
+    auto logger = std::make_shared<spdlog::logger>(name, stdout_sink);
+#if ENABLE_TEST_MSG
+    logger->set_level(spdlog::level::trace);
+#else
+    logger->set_level(spdlog::level::off);
+#endif
+    logger->set_pattern("%Y-%m-%d %H:%M:%S.%e [test message] =====> %v");
+    spdlog::register_logger(logger);
+    return 0;
+}
+
+int main(int argc, char* argv[]) {
+    init_test_logger();
+
+    auto ret = INSPIREFACE_CONTEXT->Load("test_res/pack/Pikachu");
+    if (ret != 0) {
+        std::cerr << "Load error" << std::endl;
+        return -1;
+    }
+
+    return Catch::Session().run(argc, argv);
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/test_cpp.cpp b/cpp-package/inspireface/cpp/test/test_cpp.cpp
new file mode 100644
index 000000000..316f69d89
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/test_cpp.cpp
@@ -0,0 +1,135 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <string>
+#define CATCH_CONFIG_RUNNER
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include "spdlog/spdlog.h"
+#include <inspireface/include/inspireface/inspireface.hpp>
+#include "unit/test_helper/simple_csv_writer.h"
+
+int init_test_logger() {
+    std::string name("TEST");
+    auto stdout_sink = std::make_shared<spdlog::sinks::stdout_color_sink_mt>();
+    auto logger = std::make_shared<spdlog::logger>(name, stdout_sink);
+#if ENABLE_TEST_MSG
+    logger->set_level(spdlog::level::trace);
+#else
+    logger->set_level(spdlog::level::off);
+#endif
+    logger->set_pattern("%Y-%m-%d %H:%M:%S.%e [Test Message] ===> %v");
+    spdlog::register_logger(logger);
+    return 0;
+}
+
+int init_test_benchmark_record() {
+#ifdef ISF_ENABLE_BENCHMARK
+    if (std::remove(getBenchmarkRecordFile().c_str()) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    BenchmarkRecord record(getBenchmarkRecordFile(), TEST_MODEL_FILE);
+#endif
+    return 0;
+}
+
+int init_test_evaluation_record() {
+#ifdef ISF_ENABLE_TEST_EVALUATION
+    if (std::remove(getEvaluationRecordFile().c_str()) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    EvaluationRecord record(getEvaluationRecordFile());
+#endif
+    return 0;
+}
+
+int main(int argc, char* argv[]) {
+    init_test_logger();
+    init_test_benchmark_record();
+    init_test_evaluation_record();
+    TEST_PRINT_OUTPUT(true);
+
+    TEST_PRINT("InspireFace Version: v{}.{}.{}", INSPIRE_FACE_VERSION_MAJOR_STR, INSPIRE_FACE_VERSION_MINOR_STR, INSPIRE_FACE_VERSION_PATCH_STR);
+    TEST_PRINT("Extended Information: {}", INSPIRE_FACE_EXTENDED_INFORMATION);
+
+    Catch::Session session;
+    // Pack file name and test directory
+    std::string pack;
+    std::string testDir;
+    std::string packPath;
+
+    int32_t ret;
+
+    // Add command line options
+    auto cli = session.cli() | Catch::clara::Opt(pack, "value")["--pack"]("Resource pack filename") |
+               Catch::clara::Opt(testDir, "value")["--test_dir"]("Test dir resource") |
+               Catch::clara::Opt(packPath, "value")["--pack_path"]("The specified path to the pack file");
+
+    // Set combined CLI to the session
+    session.cli(cli);
+
+    // Parse command line arguments
+    int returnCode = session.applyCommandLine(argc, argv);
+    if (returnCode != 0)  // Indicate an error
+        return returnCode;
+
+    if (!testDir.empty()) {
+        SET_TEST_DIR(testDir);
+        TEST_PRINT("Updated test dir to: {}", getTestDataDir());
+    } else {
+        TEST_PRINT("Using default test dir: {}", getTestDataDir());
+    }
+
+#if defined(ISF_ENABLE_TENSORRT)
+    int32_t cuda_device_count;
+    ret = inspire::GetCudaDeviceCount(&cuda_device_count);
+    if (ret != HSUCCEED) {
+        TEST_ERROR_PRINT("An error occurred while checking CUDA device support: {}", ret);
+        return ret;
+    }
+    if (cuda_device_count == 0) {
+        TEST_ERROR_PRINT("CUDA device support is not available");
+        return HERR_DEVICE_CUDA_NOT_SUPPORT;
+    }
+
+    inspire::PrintCudaDeviceInfo();
+#endif
+
+    std::string fullPath;
+    // Check whether custom parameters are set
+    if (!pack.empty()) {
+        SET_PACK_NAME(pack);
+        fullPath = GET_MODEL_FILE();
+        TEST_PRINT("Updated global Pack to: {}", TEST_MODEL_FILE);
+        SET_RUNTIME_FULLPATH_NAME(fullPath);
+    } else if (!packPath.empty()) {
+        fullPath = packPath;
+        TEST_PRINT("Updated global Pack File to: {}", packPath);
+        SET_RUNTIME_FULLPATH_NAME(packPath);
+    } else {
+        fullPath = GET_MODEL_FILE();
+        TEST_PRINT("Using default global Pack: {}", TEST_MODEL_FILE);
+        SET_RUNTIME_FULLPATH_NAME(fullPath);
+    }
+
+    TEST_PRINT("Launching InspireFace with path: {}", fullPath);
+    ret = INSPIREFACE_CONTEXT->Load(fullPath.c_str());
+    if (ret != HSUCCEED) {
+        TEST_ERROR_PRINT("An error occurred while starting InspireFace: {}", ret);
+        return ret;
+    }
+
+    // Set log level
+    INSPIRE_SET_LOG_LEVEL(inspire::LogLevel::ISF_LOG_INFO);
+
+    // Run the test
+    ret = session.run();
+
+    // Terminate the InspireFace instance
+    INSPIREFACE_CONTEXT->Unload();
+
+    return ret;
+}
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_benchmark.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_benchmark.cpp
new file mode 100644
index 000000000..83d063173
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_benchmark.cpp
@@ -0,0 +1,670 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-01-12
+ */
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/simple_csv_writer.h"
+#include "unit/test_helper/test_help.h"
+#include "unit/test_helper/test_tools.h"
+#include "middleware/costman.h"
+
+#ifdef ISF_ENABLE_BENCHMARK
+
+TEST_CASE("test_BenchmarkFaceDetect", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    const int loop = 1000;
+
+    SECTION("Benchmark face detection@160") {
+        auto pixLevel = 160;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, pixLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        inspire::SpendTimer timeSpend("Face Detect@160");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            // Extract basic face information from photos
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum == 1);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Benchmark face detection@320") {
+        auto pixLevel = 320;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, pixLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        inspire::SpendTimer timeSpend("Face Detect@320");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            // Extract basic face information from photos
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum == 1);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Benchmark face detection@640") {
+        auto pixLevel = 640;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, pixLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        inspire::SpendTimer timeSpend("Face Detect@640");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            // Extract basic face information from photos
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum == 1);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+TEST_CASE("test_BenchmarkFaceTrack", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    const int loop = 1000;
+    auto pixLevel = 160;
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    HFSession session;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 3, pixLevel, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    // Get a face picture
+    HFImageStream imgHandle;
+    auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+    ret = CVImageToImageStream(image, imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    inspire::SpendTimer timeSpend("Face Track");
+    for (size_t i = 0; i < loop; i++) {
+        timeSpend.Start();
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        timeSpend.Stop();
+    }
+    std::cout << timeSpend << std::endl;
+
+    ret = HFReleaseImageStream(imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_BenchmarkFaceExtractWithAlign", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    const int loop = 1000;
+    auto pixLevel = 160;
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    HFSession session;
+
+    ret = HFCreateInspireFaceSessionOptional(HF_ENABLE_FACE_RECOGNITION, detMode, 3, -1, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    // Face track
+    auto dstImage = inspirecv::Image::Create(GET_DATA("data/search/Teresa_Williams_0001_1k.jpg"));
+
+    // Get a face picture
+    HFImageStream imgHandle;
+    ret = CVImageToImageStream(dstImage, imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    HFMultipleFaceData multipleFaceData = {0};
+    ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+    REQUIRE(ret == HSUCCEED);
+
+    inspire::SpendTimer timeSpend("Face Extract With Align");
+    for (size_t i = 0; i < loop; i++) {
+        timeSpend.Start();
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        timeSpend.Stop();
+    }
+    std::cout << timeSpend << std::endl;
+
+    ret = HFReleaseImageStream(imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_BenchmarkFaceComparison", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    int loop = 1000;
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    parameter.enable_recognition = 1;
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    auto image = inspirecv::Image::Create(GET_DATA("data/bulk/woman.png"));
+    HFImageStream imgHandle;
+    ret = CVImageToImageStream(image, imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    // Extract basic face information from photos
+    HFMultipleFaceData multipleFaceDataZy = {0};
+    ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceDataZy);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(multipleFaceDataZy.detectedNum > 0);
+
+    HInt32 featureNum;
+    HFGetFeatureLength(&featureNum);
+
+    // Extract face feature
+    HFloat featureCacheZy[featureNum];
+    ret = HFFaceFeatureExtractCpy(session, imgHandle, multipleFaceDataZy.tokens[0], featureCacheZy);
+    HFFaceFeature featureZy = {0};
+    featureZy.size = featureNum;
+    featureZy.data = featureCacheZy;
+    REQUIRE(ret == HSUCCEED);
+
+    auto imageQuery = inspirecv::Image::Create(GET_DATA("data/bulk/woman_search.jpeg"));
+    HFImageStream imgHandleQuery;
+    ret = CVImageToImageStream(imageQuery, imgHandleQuery);
+    REQUIRE(ret == HSUCCEED);
+
+    HFMultipleFaceData multipleFaceDataQuery = {0};
+    ret = HFExecuteFaceTrack(session, imgHandleQuery, &multipleFaceDataQuery);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(multipleFaceDataQuery.detectedNum > 0);
+
+    // Extract face feature
+    HFloat featureCacheZyQuery[featureNum];
+    ret = HFFaceFeatureExtractCpy(session, imgHandleQuery, multipleFaceDataQuery.tokens[0], featureCacheZyQuery);
+    HFFaceFeature featureZyQuery = {0};
+    featureZyQuery.data = featureCacheZyQuery;
+    featureZyQuery.size = featureNum;
+    REQUIRE(ret == HSUCCEED);
+
+    inspire::SpendTimer timeSpend("Face Comparison");
+    for (int i = 0; i < loop; ++i) {
+        timeSpend.Start();
+        HFloat compRes;
+        ret = HFFaceComparison(featureZy, featureZyQuery, &compRes);
+        REQUIRE(ret == HSUCCEED);
+        timeSpend.Stop();
+    }
+    std::cout << timeSpend << std::endl;
+
+    ret = HFReleaseImageStream(imgHandle);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseImageStream(imgHandleQuery);
+    REQUIRE(ret == HSUCCEED);
+
+    // Finish
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_BenchmarkFaceHubSearchPersistence", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Benchmark search 1k@Persistence") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 1000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 1k@Persistence");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+
+    SECTION("Benchmark search 5k@Persistence") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 5000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 4800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 5k@Persistence");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+
+    SECTION("Benchmark search 10k@Persistence") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 10000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 9800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 10k@Persistence");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+}
+
+TEST_CASE("test_BenchmarkFaceHubSearchMemory", "[benchmark]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Benchmark search 1k@Memory") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 0;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 1000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 1k@Memory");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Benchmark search 5k@Persistence") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 0;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 5000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 4800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 5k@Memory");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Benchmark search 10k@Persistence") {
+        const int loop = 1000;
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 0;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 10000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        HInt32 targetId = 9800;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFloat confidence = 0.0f;
+        HFFaceFeatureIdentity mostSimilar = {0};
+        inspire::SpendTimer timeSpend("Face Search 10k@Memory");
+        for (size_t i = 0; i < loop; i++) {
+            timeSpend.Start();
+            ret = HFFeatureHubFaceSearch(searchFeature, &confidence, &mostSimilar);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(mostSimilar.id == targetId);
+            REQUIRE(confidence > 0.88f);
+            timeSpend.Stop();
+        }
+        std::cout << timeSpend << std::endl;
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+#endif
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_evaluation.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_evaluation.cpp
new file mode 100644
index 000000000..b11aad4b3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_evaluation.cpp
@@ -0,0 +1,118 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/simple_csv_writer.h"
+#include "unit/test_helper/test_help.h"
+#include "unit/test_helper/test_tools.h"
+
+TEST_CASE("test_Evaluation", "[face_evaluation") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Test compare tools") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 5, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        float mostSim = -1.0f;
+        auto succ = FindMostSimilarScoreFromTwoPic(session, GET_DATA("data/bulk/jntm.jpg"), GET_DATA("data/bulk/kun.jpg"), mostSim);
+        CHECK(succ);
+        TEST_PRINT("kun v kun :{}", mostSim);
+
+        succ = FindMostSimilarScoreFromTwoPic(session, GET_DATA("data/bulk/jntm.jpg"), GET_DATA("data/bulk/Rob_Lowe_0001.jpg"), mostSim);
+        CHECK(succ);
+        TEST_PRINT("kun v other :{}", mostSim);
+
+        succ = FindMostSimilarScoreFromTwoPic(session, GET_DATA("data/bulk/kun.jpg"), GET_DATA("data/bulk/view.jpg"), mostSim);
+        CHECK(!succ);
+        TEST_PRINT("kun v other :{}", mostSim);
+
+        // finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Test LFW evaluation") {
+#ifdef ISF_ENABLE_TEST_EVALUATION
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 5, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        std::vector<int> labels;
+        std::vector<float> confidences;
+        auto pairs = ReadPairs(getTestLFWFunneledEvaTxt());
+        // Hide cursor
+        show_console_cursor(false);
+        BlockProgressBar bar{option::BarWidth{60},
+                             option::Start{"["},
+                             option::End{"]"},
+                             option::PostfixText{"Extracting face features"},
+                             option::ForegroundColor{Color::white},
+                             option::FontStyles{std::vector<FontStyle>{FontStyle::bold}}};
+        auto progress = 0.0f;
+
+        for (int i = 0; i < pairs.size(); ++i) {
+            bar.set_progress(progress);
+
+            auto &pair = pairs[i];
+            std::string person1, person2;
+            int imgNum1, imgNum2;
+            std::string imgPath1, imgPath2;
+            int match;
+            if (pair.size() == 3) {
+                person1 = pair[0];
+                imgNum1 = std::stoi(pair[1]);
+                imgNum2 = std::stoi(pair[2]);
+                imgPath1 = PathJoin(PathJoin(getLFWFunneledDir(), person1), person1 + "_" + zfill(imgNum1, 4) + ".jpg");
+                imgPath2 = PathJoin(PathJoin(getLFWFunneledDir(), person1), person1 + "_" + zfill(imgNum2, 4) + ".jpg");
+                match = 1;
+            } else {
+                person1 = pair[0];
+                imgNum1 = std::stoi(pair[1]);
+                person2 = pair[2];
+                imgNum2 = std::stoi(pair[3]);
+                imgPath1 = PathJoin(PathJoin(getLFWFunneledDir(), person1), person1 + "_" + zfill(imgNum1, 4) + ".jpg");
+                imgPath2 = PathJoin(PathJoin(getLFWFunneledDir(), person2), person2 + "_" + zfill(imgNum2, 4) + ".jpg");
+                match = 0;
+            }
+
+            float mostSim;
+            auto succ = FindMostSimilarScoreFromTwoPic(session, imgPath1, imgPath2, mostSim);
+            if (!succ) {
+                continue;
+            }
+
+            labels.push_back(match);
+            confidences.push_back(mostSim);
+
+            // Update progress
+            progress = 100.0f * (float)(i + 1) / pairs.size();
+        }
+        // Show cursor
+        show_console_cursor(true);
+
+        REQUIRE(labels.size() == confidences.size());
+        TEST_PRINT("scan pair: {}", labels.size());
+        bar.set_progress(100.0f);
+
+        auto result = FindBestThreshold(confidences, labels);
+        TEST_PRINT("Best Threshold: {}, Best Accuracy: {}", result.first, result.second);
+        EvaluationRecord record(getEvaluationRecordFile());
+        record.insertEvaluationData(TEST_MODEL_FILE, "LFW", result.second, result.first);
+        // finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+#endif
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_face_context.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_face_context.cpp
new file mode 100644
index 000000000..ef5b08703
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_face_context.cpp
@@ -0,0 +1,25 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include <cstdio>
+
+TEST_CASE("test_FeatureContext", "[face_context]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Test the new context positive process") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_face_pipeline.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_face_pipeline.cpp
new file mode 100644
index 000000000..6ade3842a
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_face_pipeline.cpp
@@ -0,0 +1,554 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "../test_helper/test_tools.h"
+#include "../test_helper/test_help.h"
+
+TEST_CASE("test_FaceEmotion", "[face_emotion]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    enum EMOTION {
+        NEUTRAL = 0,   ///< Emotion: neutral
+        HAPPY = 1,     ///< Emotion: happy
+        SAD = 2,       ///< Emotion: sad
+        SURPRISE = 3,  ///< Emotion: surprise
+        FEAR = 4,      ///< Emotion: fear
+        DISGUST = 5,   ///< Emotion: disgust
+        ANGER = 6,     ///< Emotion: anger
+    };
+
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    HInt32 faceDetectPixelLevel = 320;
+    HInt32 option = HF_ENABLE_FACE_EMOTION;
+    ret = HFCreateInspireFaceSessionOptional(option, detMode, 5, faceDetectPixelLevel, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    std::vector<std::string> test_images = {
+      "data/emotion/anger.png",
+      "data/emotion/sad.png",
+      "data/emotion/happy.png",
+    };
+
+    std::vector<EMOTION> expected_emotions = {
+      ANGER,
+      SAD,
+      HAPPY,
+    };
+    REQUIRE(test_images.size() == expected_emotions.size());
+
+    for (size_t i = 0; i < test_images.size(); i++) {
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA(test_images[i]));
+        REQUIRE(!img.Empty());
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+
+        ret = HFMultipleFacePipelineProcessOptional(session, imgHandle, &multipleFaceData, option);
+        REQUIRE(ret == HSUCCEED);
+
+        HFFaceEmotionResult result = {0};
+        ret = HFGetFaceEmotionResult(session, &result);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(result.num == 1);
+        CHECK(result.emotion[0] == (HInt32)expected_emotions[i]);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+        imgHandle = nullptr;
+    }
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+    session = nullptr;
+}
+
+TEST_CASE("test_FacePipelineAttribute", "[face_pipeline_attribute]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    enum AGE_BRACKED {
+        AGE_0_2 = 0,   ///< Age 0-2 years old
+        AGE_3_9,       ///< Age 3-9 years old
+        AGE_10_19,     ///< Age 10-19 years old
+        AGE_20_29,     ///< Age 20-29 years old
+        AGE_30_39,     ///< Age 30-39 years old
+        AGE_40_49,     ///< Age 40-49 years old
+        AGE_50_59,     ///< Age 50-59 years old
+        AGE_60_69,     ///< Age 60-69 years old
+        MORE_THAN_70,  ///< Age more than 70 years old
+    };
+    enum GENDER {
+        FEMALE = 0,  ///< Female
+        MALE,        ///< Male
+    };
+    enum RACE {
+        BLACK = 0,        ///< Black
+        ASIAN,            ///< Asian
+        LATINO_HISPANIC,  ///< Latino/Hispanic
+        MIDDLE_EASTERN,   ///< Middle Eastern
+        WHITE,            ///< White
+    };
+
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    parameter.enable_face_attribute = 1;
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    HInt32 faceDetectPixelLevel = 320;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 5, faceDetectPixelLevel, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    SECTION("a black girl") {
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA("data/attribute/1423.jpg"));
+        REQUIRE(!img.Empty());
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+
+        // Run pipeline
+        ret = HFMultipleFacePipelineProcessOptional(session, imgHandle, &multipleFaceData, HF_ENABLE_FACE_ATTRIBUTE);
+        REQUIRE(ret == HSUCCEED);
+
+        HFFaceAttributeResult result = {0};
+        ret = HFGetFaceAttributeResult(session, &result);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(result.num == 1);
+
+        // Check attribute
+        CHECK(result.race[0] == BLACK);
+        CHECK(result.ageBracket[0] == AGE_10_19);
+        CHECK(result.gender[0] == FEMALE);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+        imgHandle = nullptr;
+    }
+
+    SECTION("two young white women") {
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA("data/attribute/7242.jpg"));
+        REQUIRE(!img.Empty());
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 2);
+
+        // Run pipeline
+        ret = HFMultipleFacePipelineProcessOptional(session, imgHandle, &multipleFaceData, HF_ENABLE_FACE_ATTRIBUTE);
+        REQUIRE(ret == HSUCCEED);
+
+        HFFaceAttributeResult result = {0};
+        ret = HFGetFaceAttributeResult(session, &result);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(result.num == 2);
+
+        // Check attribute
+        for (size_t i = 0; i < result.num; i++) {
+            CHECK(result.race[i] == WHITE);
+            CHECK(result.ageBracket[i] == AGE_20_29);
+            CHECK(result.gender[i] == FEMALE);
+        }
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+        imgHandle = nullptr;
+    }
+
+    ret = HFReleaseInspireFaceSession(session);
+    session = nullptr;
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_FacePipeline", "[face_pipeline]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("rgb liveness detect") {
+#ifndef INFERENCE_WRAPPER_ENABLE_RKNN2
+        /** The anti spoofing model based on RGB faces seems to have some problems with quantization under RKNPU2, so it is not started yet */
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_liveness = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, 320, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream img1Handle;
+        auto img1 = inspirecv::Image::Create(GET_DATA("data/bulk/image_T1.jpeg"));
+        ret = CVImageToImageStream(img1, img1Handle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, img1Handle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        ret = HFMultipleFacePipelineProcess(session, img1Handle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+        HFRGBLivenessConfidence confidence;
+        ret = HFGetRGBLivenessConfidence(session, &confidence);
+        TEST_PRINT("{}", confidence.confidence[0]);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(confidence.num > 0);
+        // CHECK(confidence.confidence[0] > 0.70f);
+
+        ret = HFReleaseImageStream(img1Handle);
+        REQUIRE(ret == HSUCCEED);
+        img1Handle = nullptr;
+
+        // fake face
+        HFImageStream img2Handle;
+        auto img2 = inspirecv::Image::Create(GET_DATA("data/bulk/rgb_fake.jpg"));
+        ret = CVImageToImageStream(img2, img2Handle);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFExecuteFaceTrack(session, img2Handle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFMultipleFacePipelineProcess(session, img2Handle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFGetRGBLivenessConfidence(session, &confidence);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(confidence.num > 0);
+        CHECK(confidence.confidence[0] < 0.9);
+
+        ret = HFReleaseImageStream(img2Handle);
+        REQUIRE(ret == HSUCCEED);
+        img2Handle = nullptr;
+
+        ret = HFReleaseInspireFaceSession(session);
+        session = nullptr;
+        REQUIRE(ret == HSUCCEED);
+#else
+        TEST_PRINT("The anti spoofing model based on RGB faces seems to have some problems with quantization under RKNPU2, so we skip this test.");
+#endif
+    }
+
+    SECTION("face mask detect") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_mask_detect = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream img1Handle;
+        auto img1 = inspirecv::Image::Create(GET_DATA("data/bulk/mask2.jpg"));
+        ret = CVImageToImageStream(img1, img1Handle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, img1Handle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        ret = HFMultipleFacePipelineProcess(session, img1Handle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+        HFFaceMaskConfidence confidence;
+        ret = HFGetFaceMaskConfidence(session, &confidence);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(confidence.num > 0);
+        CHECK(confidence.confidence[0] > 0.9);
+
+        ret = HFReleaseImageStream(img1Handle);
+        REQUIRE(ret == HSUCCEED);
+        img1Handle = nullptr;
+
+        // no mask face
+        HFImageStream img2Handle;
+        auto img2 = inspirecv::Image::Create(GET_DATA("data/bulk/face_sample.png"));
+        ret = CVImageToImageStream(img2, img2Handle);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFExecuteFaceTrack(session, img2Handle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFMultipleFacePipelineProcess(session, img2Handle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFGetFaceMaskConfidence(session, &confidence);
+        REQUIRE(ret == HSUCCEED);
+        //        spdlog::info("mask {}", confidence.confidence[0]);
+        CHECK(confidence.num > 0);
+        CHECK(confidence.confidence[0] < 0.1);
+
+        ret = HFReleaseImageStream(img2Handle);
+        REQUIRE(ret == HSUCCEED);
+        img2Handle = nullptr;
+
+        ret = HFReleaseInspireFaceSession(session);
+        session = nullptr;
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("face quality") {
+        HResult ret;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HInt32 option = HF_ENABLE_QUALITY;
+        HFSession session;
+        ret = HFCreateInspireFaceSessionOptional(option, detMode, 3, 320, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream superiorHandle;
+        auto superior = inspirecv::Image::Create(GET_DATA("data/bulk/yifei.jpg"));
+        ret = CVImageToImageStream(superior, superiorHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, superiorHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        ret = HFMultipleFacePipelineProcessOptional(session, superiorHandle, &multipleFaceData, option);
+        REQUIRE(ret == HSUCCEED);
+
+        HFloat quality;
+        ret = HFFaceQualityDetect(session, multipleFaceData.tokens[0], &quality);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(quality > 0.8);
+
+        // blur image
+        HFImageStream blurHandle;
+        auto blur = inspirecv::Image::Create(GET_DATA("data/bulk/blur.jpg"));
+        ret = CVImageToImageStream(blur, blurHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        ret = HFExecuteFaceTrack(session, blurHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        ret = HFMultipleFacePipelineProcessOptional(session, blurHandle, &multipleFaceData, option);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFaceQualityDetect(session, multipleFaceData.tokens[0], &quality);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(quality < 0.85);
+
+        ret = HFReleaseImageStream(superiorHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(blurHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+TEST_CASE("test_FaceReaction", "[face_reaction]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    parameter.enable_interaction_liveness = 1;
+    parameter.enable_liveness = 1;
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+    SECTION("open eyes") {
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA("data/reaction/open_eyes.png"));
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Predict eyes status
+        ret = HFMultipleFacePipelineProcess(session, imgHandle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get results
+        HFFaceInteractionState result;
+        ret = HFGetFaceInteractionStateResult(session, &result);
+        REQUIRE(multipleFaceData.detectedNum == result.num);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check
+        CHECK(result.leftEyeStatusConfidence[0] > 0.5f);
+        CHECK(result.rightEyeStatusConfidence[0] > 0.5f);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("close eyes") {
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA("data/reaction/close_eyes.jpeg"));
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Predict eyes status
+        ret = HFMultipleFacePipelineProcess(session, imgHandle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get results
+        HFFaceInteractionState result;
+        ret = HFGetFaceInteractionStateResult(session, &result);
+        REQUIRE(multipleFaceData.detectedNum == result.num);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check
+        CHECK(result.leftEyeStatusConfidence[0] < 0.5f);
+        CHECK(result.rightEyeStatusConfidence[0] < 0.5f);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Close one eye and open the other") {
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto img = inspirecv::Image::Create(GET_DATA("data/reaction/close_open_eyes.jpeg"));
+        ret = CVImageToImageStream(img, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Predict eyes status
+        ret = HFMultipleFacePipelineProcess(session, imgHandle, &multipleFaceData, parameter);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get results
+        HFFaceInteractionState result;
+        ret = HFGetFaceInteractionStateResult(session, &result);
+        REQUIRE(multipleFaceData.detectedNum == result.num);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check
+        CHECK(result.leftEyeStatusConfidence[0] < 0.5f);
+        CHECK(result.rightEyeStatusConfidence[0] > 0.5f);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_TrackModeFaceAction", "[face_action]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFSessionCustomParameter parameter = {0};
+    parameter.enable_interaction_liveness = 1;
+    HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+    HFSession session;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+
+#if 0  // Temporarily shut down the processing of frame related cases to prevent excessive data from occupying the github memory
+    SECTION("Action Blink") {
+        auto start = 130, end = 150;
+        std::vector<std::string> filenames = generateFilenames("frame-%04d.jpg", start, end);
+        int count = 0;
+        for (size_t i = 0; i < filenames.size(); i++) {
+            auto filename = filenames[i];
+            HFImageStream imgHandle;
+            auto image = inspirecv::Image::Create(GET_DATA("data/video_frames/" + filename));
+            ret = CVImageToImageStream(image, imgHandle);
+            REQUIRE(ret == HSUCCEED);
+
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum > 0);
+
+            ret = HFMultipleFacePipelineProcessOptional(session, imgHandle, &multipleFaceData, HF_ENABLE_INTERACTION);
+            REQUIRE(ret == HSUCCEED);
+
+            HFFaceInteractionsActions result;
+            ret = HFGetFaceInteractionActionsResult(session, &result);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum == result.num);
+
+            count += result.blink[0];
+            ret = HFReleaseImageStream(imgHandle);
+            REQUIRE(ret == HSUCCEED);
+        }
+        // Blink at least once
+        REQUIRE(count > 0);
+    }
+#endif
+
+#if 0  // Temporarily shut down the processing of frame related cases to prevent excessive data from occupying the github memory
+    SECTION("Action Jaw Open") {
+        auto start = 110, end = 150;
+        std::vector<std::string> filenames = generateFilenames("frame-%04d.jpg", start, end);
+        int count = 0;
+        for (size_t i = 0; i < filenames.size(); i++) {
+            auto filename = filenames[i];
+            HFImageStream imgHandle;
+            auto image = inspirecv::Image::Create(GET_DATA("data/video_frames/" + filename));
+            ret = CVImageToImageStream(image, imgHandle);
+            REQUIRE(ret == HSUCCEED);
+
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum > 0);
+
+            ret = HFMultipleFacePipelineProcessOptional(session, imgHandle, &multipleFaceData, HF_ENABLE_INTERACTION);
+            REQUIRE(ret == HSUCCEED);
+
+            HFFaceInteractionsActions result;
+            ret = HFGetFaceInteractionActionsResult(session, &result);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum == result.num);
+
+            count += result.jawOpen[0];
+            ret = HFReleaseImageStream(imgHandle);
+            REQUIRE(ret == HSUCCEED);
+        }
+        // Jaw open at least once
+        REQUIRE(count > 0);
+    }
+#endif
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+}
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_face_track.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_face_track.cpp
new file mode 100644
index 000000000..9b2fd33a7
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_face_track.cpp
@@ -0,0 +1,555 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/simple_csv_writer.h"
+#include "unit/test_helper/test_help.h"
+#include "unit/test_helper/test_tools.h"
+#include "middleware/costman.h"
+
+TEST_CASE("test_FaceTrack", "[face_track]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Face detection from image") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        TEST_ERROR_PRINT("error ret :{}", ret);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+
+        // Detect face position
+        auto rect = multipleFaceData.rects[0];
+        HFaceRect expect = {0};
+        expect.x = 79;
+        expect.y = 104;
+        expect.width = 168;
+        expect.height = 167;
+
+        auto iou = CalculateOverlap(rect, expect);
+        auto cvRect = inspirecv::Rect<int>::Create(rect.x, rect.y, rect.width, rect.height);
+        image.DrawRect(cvRect, {0, 0, 255}, 2);
+        image.Write("ww.jpg");
+        // The iou is allowed to have an error of 50%
+        CHECK(iou == Approx(1.0f).epsilon(0.5));
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Prepare non-face images
+        HFImageStream viewHandle;
+        auto view = inspirecv::Image::Create(GET_DATA("data/bulk/view.jpg"));
+        ret = CVImageToImageStream(view, viewHandle);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFExecuteFaceTrack(session, viewHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 0);
+
+        ret = HFReleaseImageStream(viewHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+// Temporarily shut down the processing of frame related cases to prevent excessive data from occupying the github memory
+#if 0
+    SECTION("Face tracking stability from frames") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        auto expectedId = 1;
+        int start = 1, end = 288;
+        std::vector<std::string> filenames = generateFilenames("frame-%04d.jpg", start, end);
+        auto count_loss = 0;
+        for (int i = 0; i < filenames.size(); ++i) {
+            auto filename = filenames[i];
+            HFImageStream imgHandle;
+            auto image = inspirecv::Image::Create(GET_DATA("data/video_frames/" + filename));
+            ret = CVImageToImageStream(image, imgHandle);
+            REQUIRE(ret == HSUCCEED);
+
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            //            CHECK(multipleFaceData.detectedNum == 1);
+            if (multipleFaceData.detectedNum != 1) {
+                count_loss++;
+                continue;
+            }
+            auto rect = multipleFaceData.rects[0];
+            auto cvRect = inspirecv::Rect<int>::Create(rect.x, rect.y, rect.width, rect.height);
+            image.DrawRect(cvRect, {0, 0, 255}, 2);
+            std::string save = GET_SAVE_DATA("video_frames") + "/" + std::to_string(i) + ".jpg";
+            image.Write(save);
+            auto id = multipleFaceData.trackIds[0];
+            //            TEST_PRINT("{}", id);
+            if (id != expectedId) {
+                count_loss++;
+            }
+
+            ret = HFReleaseImageStream(imgHandle);
+            REQUIRE(ret == HSUCCEED);
+        }
+        float loss = (float)count_loss / filenames.size();
+        // The face track loss is allowed to have an error of 5%
+        //        CHECK(loss == Approx(0.0f).epsilon(0.05));
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+#endif
+
+    SECTION("Head pose estimation") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_face_pose = true;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+
+        // Left side face
+        HFImageStream leftHandle;
+        auto left = inspirecv::Image::Create(GET_DATA("data/pose/left_face.jpeg"));
+        ret = CVImageToImageStream(left, leftHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, leftHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+
+        HFloat yaw, pitch, roll;
+        bool checked;
+
+        // Left-handed rotation
+        yaw = multipleFaceData.angles.yaw[0];
+        checked = (yaw > -90 && yaw < -10);
+        CHECK(checked);
+
+        HFReleaseImageStream(leftHandle);
+
+        // Right-handed rotation
+        HFImageStream rightHandle;
+        auto right = inspirecv::Image::Create(GET_DATA("data/pose/right_face.png"));
+        ret = CVImageToImageStream(right, rightHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, rightHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        yaw = multipleFaceData.angles.yaw[0];
+        checked = (yaw > 10 && yaw < 90);
+        CHECK(checked);
+
+        HFReleaseImageStream(rightHandle);
+
+        // Rise head
+        HFImageStream riseHandle;
+        auto rise = inspirecv::Image::Create(GET_DATA("data/pose/rise_face.jpeg"));
+        ret = CVImageToImageStream(rise, riseHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, riseHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        pitch = multipleFaceData.angles.pitch[0];
+        CHECK(pitch > 3);
+        HFReleaseImageStream(riseHandle);
+
+        // Lower head
+        HFImageStream lowerHandle;
+        auto lower = inspirecv::Image::Create(GET_DATA("data/pose/lower_face.jpeg"));
+        ret = CVImageToImageStream(lower, lowerHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, lowerHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        pitch = multipleFaceData.angles.pitch[0];
+        CHECK(pitch < -10);
+        HFReleaseImageStream(lowerHandle);
+
+        // Roll head
+        HFImageStream leftWryneckHandle;
+        auto leftWryneck = inspirecv::Image::Create(GET_DATA("data/pose/left_wryneck.png"));
+        ret = CVImageToImageStream(leftWryneck, leftWryneckHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, leftWryneckHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        roll = multipleFaceData.angles.roll[0];
+        CHECK(roll < -30);
+        HFReleaseImageStream(leftWryneckHandle);
+
+        // Roll head
+        HFImageStream rightWryneckHandle;
+        auto rightWryneck = inspirecv::Image::Create(GET_DATA("data/pose/right_wryneck.png"));
+        ret = CVImageToImageStream(rightWryneck, rightWryneckHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFExecuteFaceTrack(session, rightWryneckHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        roll = multipleFaceData.angles.roll[0];
+        CHECK(roll > 25);
+        HFReleaseImageStream(rightWryneckHandle);
+
+        //  finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Face detection benchmark") {
+#ifdef ISF_ENABLE_BENCHMARK
+        int loop = 1000;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Prepare an image
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+        BenchmarkRecord record(getBenchmarkRecordFile());
+
+        REQUIRE(ret == HSUCCEED);
+        HFMultipleFaceData multipleFaceData = {0};
+        auto timer = inspire::Timer();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        }
+        auto cost = timer.GetCostTime();
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum == 1);
+        TEST_PRINT("<Benchmark> Face Detect -> Loop: {}, Total Time: {:.5f}ms, Average Time: {:.5f}ms", loop, cost, cost / loop);
+        record.insertBenchmarkData("Face Detect", loop, cost, cost / loop);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+#else
+        TEST_PRINT("Skip the face detection benchmark test. To run it, you need to turn on the benchmark test.");
+#endif
+    }
+
+    SECTION("Face light track benchmark") {
+#ifdef ISF_ENABLE_BENCHMARK
+        int loop = 1000;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_LIGHT_TRACK;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Prepare an image
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+        BenchmarkRecord record(getBenchmarkRecordFile());
+
+        // Case: Execute the benchmark using the VIDEO mode(Track)
+        REQUIRE(ret == HSUCCEED);
+        HFMultipleFaceData multipleFaceData = {0};
+        auto timer = inspire::Timer();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        }
+        auto cost = timer.GetCostTime();
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum >= 1);
+        TEST_PRINT("<Benchmark> Face Track -> Loop: {}, Total Time: {:.5f}ms, Average Time: {:.5f}ms", loop, cost, cost / loop);
+        record.insertBenchmarkData("Face Track", loop, cost, cost / loop);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+#else
+        TEST_PRINT("Skip the face light track benchmark test. To run it, you need to turn on the benchmark test.");
+#endif
+    }
+}
+
+TEST_CASE("test_MultipleLevelFaceDetect", "[face_detect]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Detect input 192px") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 detectPixelLevel = 192;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 20, detectPixelLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+        HFSessionSetFilterMinimumFacePixelSize(session, 0);
+        
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(previewSize == detectPixelLevel);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 0);
+        CHECK(multipleFaceData.detectedNum < 7);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Detect input 320px") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 detectPixelLevel = 320;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 20, detectPixelLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+        HFSessionSetFilterMinimumFacePixelSize(session, 0);
+
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(previewSize == detectPixelLevel);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 9);
+        CHECK(multipleFaceData.detectedNum < 12);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Detect input 640px") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 detectPixelLevel = 640;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 25, detectPixelLevel, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFSessionSetTrackPreviewSize(session, detectPixelLevel);
+        HFSessionSetFilterMinimumFacePixelSize(session, 0);
+
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(previewSize == detectPixelLevel);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 15);
+        CHECK(multipleFaceData.detectedNum < 21);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+TEST_CASE("test_FaceTrackPreviewSizeSetting", "[face_track]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Default preview size and detection level size") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 levelSize = -1;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 20, levelSize, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        
+        CHECK(previewSize == 320);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 0);
+        // Check the preview size
+        HInt32 debugPreviewSize;
+        ret = HFSessionLastFaceDetectionGetDebugPreviewImageSize(session, &debugPreviewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(debugPreviewSize == 320);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Set preview size to 320px") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 levelSize = 320;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 20, levelSize, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(previewSize == levelSize);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Set the detect level to an invalid value") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        HInt32 levelSize = 1000;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 20, levelSize, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check the preview size
+        HInt32 previewSize;
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        // If the detect level value is invalid, the value will be automatically adjusted to the nearest legal size. 
+        // If the default value of preview_size is -1, the value will also be adjusted
+        CHECK(previewSize == 640);
+
+        // Get a face picture
+        HFImageStream imgHandle;
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+        ret = CVImageToImageStream(image, imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 0);
+        
+        // Check the preview size
+        HInt32 debugPreviewSize;
+        ret = HFSessionLastFaceDetectionGetDebugPreviewImageSize(session, &debugPreviewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(debugPreviewSize == 640);
+
+        // Set a value manually
+        ret = HFSessionSetTrackPreviewSize(session, 192);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check the preview size
+        ret = HFSessionGetTrackPreviewSize(session, &previewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(previewSize == 192);
+        
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+
+        CHECK(multipleFaceData.detectedNum > 0);
+
+        // Check the preview size
+        ret = HFSessionLastFaceDetectionGetDebugPreviewImageSize(session, &debugPreviewSize);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(debugPreviewSize == 192);
+
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+    
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_feature_hub.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_feature_hub.cpp
new file mode 100644
index 000000000..770ef8d84
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_feature_hub.cpp
@@ -0,0 +1,552 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/test_help.h"
+#include <thread>
+
+TEST_CASE("test_FeatureHubBase", "[FeatureHub][BasicFunction]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("FeatureHub basic function") {
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+
+    SECTION("FeatureHub search top-k") {
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<std::vector<HFloat>> baseFeatures;
+        size_t genSizeOfBase = 2000;
+        HInt32 featureLength;
+        HFGetFeatureLength(&featureLength);
+        REQUIRE(featureLength > 0);
+        for (int i = 0; i < genSizeOfBase; ++i) {
+            auto feat = GenerateRandomFeature(featureLength);
+            baseFeatures.push_back(feat);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            HFaceId allocId;
+            ret = HFFeatureHubInsertFeature(identity, &allocId);
+            REQUIRE(ret == HSUCCEED);
+        }
+        HInt32 totalFace;
+        ret = HFFeatureHubGetFaceCount(&totalFace);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(totalFace == genSizeOfBase);
+
+        // 2000 data was imported
+        HInt32 targetId = 524;
+        auto targetFeature = baseFeatures[targetId - 1];
+
+        std::vector<std::vector<HFloat>> similarVectors;
+        std::vector<HInt32> coverIds = {2, 300, 524, 789, 1024, 1995};
+        for (int i = 0; i < coverIds.size(); ++i) {
+            auto feat = SimulateSimilarVector(targetFeature);
+            // Construct face feature
+            HFFaceFeature feature = {0};
+            feature.size = feat.size();
+            feature.data = feat.data();
+            HFFaceFeatureIdentity identity = {0};
+            identity.feature = &feature;
+            identity.id = coverIds[i];
+            ret = HFFeatureHubFaceUpdate(identity);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        // Generate a new similar feature for search
+        auto topK = 10;
+        auto searchFeat = SimulateSimilarVector(targetFeature);
+        HFFaceFeature searchFeature = {0};
+        searchFeature.size = searchFeat.size();
+        searchFeature.data = searchFeat.data();
+        HFSearchTopKResults results = {0};
+        ret = HFFeatureHubFaceSearchTopK(searchFeature, topK, &results);
+        REQUIRE(ret == HSUCCEED);
+
+        REQUIRE(coverIds.size() == results.size);
+        for (int i = 0; i < results.size; ++i) {
+            REQUIRE(std::find(coverIds.begin(), coverIds.end(), results.ids[i]) != coverIds.end());
+        }
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+
+    SECTION("Repeat the enable and disable tests") {
+        HResult ret;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        HFFeatureHubConfiguration configuration;
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 0;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+    }
+
+    SECTION("Only memory storage is used") {
+        HResult ret;
+        HFFeatureHubConfiguration configuration;
+        configuration.enablePersistence = 0;
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        // TODO
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+TEST_CASE("test_ConcurrencyInsertion", "[FeatureHub][Concurrency]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFFeatureHubConfiguration configuration;
+    auto dbPath = GET_SAVE_DATA(".test");
+    HString dbPathStr = new char[dbPath.size() + 1];
+    std::strcpy(dbPathStr, dbPath.c_str());
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = dbPathStr;
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    // Delete the previous data before testing
+    if (std::remove(configuration.persistenceDbPath) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    ret = HFFeatureHubDataEnable(configuration);
+    REQUIRE(ret == HSUCCEED);
+
+    HInt32 baseNum;
+    ret = HFFeatureHubGetFaceCount(&baseNum);
+    REQUIRE(ret == HSUCCEED);
+
+    HInt32 featureLength;
+    HFGetFeatureLength(&featureLength);
+
+    const int numThreads = 4;
+    const int insertsPerThread = 50;
+    std::vector<std::thread> threads;
+    auto beginGenId = 2000;
+
+    for (int i = 0; i < numThreads; ++i) {
+        threads.emplace_back([=]() {
+            for (int j = 0; j < insertsPerThread; ++j) {
+                auto feat = GenerateRandomFeature(featureLength);
+                auto name = std::to_string(beginGenId + j + i * insertsPerThread);
+                std::vector<char> nameBuffer(name.begin(), name.end());
+                nameBuffer.push_back('\0');
+                HFFaceFeature feature = {0};
+                feature.size = feat.size();
+                feature.data = feat.data();
+                HFFaceFeatureIdentity featureIdentity = {0};
+                featureIdentity.feature = &feature;
+                // featureIdentity.customId = beginGenId + j + i * insertsPerThread;
+                // featureIdentity.tag = nameBuffer.data();
+                HFaceId allocId;
+                auto ret = HFFeatureHubInsertFeature(featureIdentity, &allocId);
+                REQUIRE(ret == HSUCCEED);
+            }
+        });
+    }
+
+    for (auto &th : threads) {
+        th.join();
+    }
+
+    HInt32 count;
+    ret = HFFeatureHubGetFaceCount(&count);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(count == baseNum + numThreads * insertsPerThread);  // Ensure that the previous base data is added to the newly inserted data
+
+    ret = HFFeatureHubDataDisable();
+    REQUIRE(ret == HSUCCEED);
+
+    delete[] dbPathStr;
+}
+
+TEST_CASE("test_ConcurrencyRemove", "[FeatureHub][Concurrency]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFFeatureHubConfiguration configuration;
+    auto dbPath = GET_SAVE_DATA(".test");
+    HString dbPathStr = new char[dbPath.size() + 1];
+    std::strcpy(dbPathStr, dbPath.c_str());
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = dbPathStr;
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    // Delete the previous data before testing
+    if (std::remove(configuration.persistenceDbPath) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    ret = HFFeatureHubDataEnable(configuration);
+    REQUIRE(ret == HSUCCEED);
+
+    std::vector<std::vector<HFloat>> baseFeatures;
+    size_t genSizeOfBase = 1000;
+    HInt32 featureLength;
+    HFGetFeatureLength(&featureLength);
+
+    REQUIRE(featureLength > 0);
+    for (int i = 0; i < genSizeOfBase; ++i) {
+        auto feat = GenerateRandomFeature(featureLength);
+        baseFeatures.push_back(feat);
+        auto name = std::to_string(i);
+        // Establish a security buffer
+        std::vector<char> nameBuffer(name.begin(), name.end());
+        nameBuffer.push_back('\0');
+        // Construct face feature
+        HFFaceFeature feature = {0};
+        feature.size = feat.size();
+        feature.data = feat.data();
+        HFFaceFeatureIdentity identity = {0};
+        identity.feature = &feature;
+        // identity.customId = i;
+        // identity.tag = nameBuffer.data();
+        HFaceId allocId;
+        ret = HFFeatureHubInsertFeature(identity, &allocId);
+        REQUIRE(ret == HSUCCEED);
+    }
+    HInt32 totalFace;
+    ret = HFFeatureHubGetFaceCount(&totalFace);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(totalFace == genSizeOfBase);
+
+    const int numThreads = 4;
+    const int removePerThread = genSizeOfBase / 5;
+    std::vector<std::thread> threads;
+    for (int t = 0; t < numThreads; ++t) {
+        threads.emplace_back([&, t]() {
+            for (int j = 0; j < removePerThread; ++j) {
+                int idToRemove = t * removePerThread + j;
+                auto ret = HFFeatureHubFaceRemove(idToRemove);
+                REQUIRE(ret == HSUCCEED);
+            }
+        });
+    }
+    // Wait for all threads to complete
+    for (auto &th : threads) {
+        th.join();
+    }
+    HInt32 remainingCount;
+    ret = HFFeatureHubGetFaceCount(&remainingCount);
+    REQUIRE(ret == HSUCCEED);
+    // need exclude id=0
+    REQUIRE(remainingCount - 1 == genSizeOfBase - numThreads * removePerThread);
+    TEST_PRINT("Remaining Count: {}", remainingCount);
+
+    ret = HFFeatureHubDataDisable();
+    REQUIRE(ret == HSUCCEED);
+
+    delete[] dbPathStr;
+}
+
+TEST_CASE("test_ConcurrencySearch", "[FeatureHub][Concurrency]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFFeatureHubConfiguration configuration;
+    auto dbPath = GET_SAVE_DATA(".test");
+    HString dbPathStr = new char[dbPath.size() + 1];
+    std::strcpy(dbPathStr, dbPath.c_str());
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = dbPathStr;
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    // Delete the previous data before testing
+    if (std::remove(configuration.persistenceDbPath) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    ret = HFFeatureHubDataEnable(configuration);
+    REQUIRE(ret == HSUCCEED);
+
+    std::vector<std::vector<HFloat>> baseFeatures;
+    size_t genSizeOfBase = 1000;
+    HInt32 featureLength;
+    HFGetFeatureLength(&featureLength);
+    REQUIRE(featureLength > 0);
+    for (int i = 0; i < genSizeOfBase; ++i) {
+        auto feat = GenerateRandomFeature(featureLength);
+        baseFeatures.push_back(feat);
+        auto name = std::to_string(i);
+        // Establish a security buffer
+        std::vector<char> nameBuffer(name.begin(), name.end());
+        nameBuffer.push_back('\0');
+        // Construct face feature
+        HFFaceFeature feature = {0};
+        feature.size = feat.size();
+        feature.data = feat.data();
+        HFFaceFeatureIdentity identity = {0};
+        identity.feature = &feature;
+        // identity.customId = i;
+        // identity.tag = nameBuffer.data();
+        HFaceId allocId;
+        ret = HFFeatureHubInsertFeature(identity, &allocId);
+        REQUIRE(ret == HSUCCEED);
+    }
+    HInt32 totalFace;
+    ret = HFFeatureHubGetFaceCount(&totalFace);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(totalFace == genSizeOfBase);
+
+    auto preDataSample = 200;
+
+    // Generate some feature vectors that are similar to those of the existing database
+    auto numberOfSimilar = preDataSample;
+    auto targetIds = GenerateRandomNumbers(numberOfSimilar, 0, genSizeOfBase - 1);
+    std::vector<std::vector<HFloat>> similarFeatures;
+    for (int i = 0; i < numberOfSimilar; ++i) {
+        auto index = targetIds[i];
+        HFFaceFeatureIdentity identity = {0};
+        ret = HFFeatureHubGetFaceIdentity(index + 1, &identity);
+        REQUIRE(ret == HSUCCEED);
+        std::vector<HFloat> feature(identity.feature->data, identity.feature->data + identity.feature->size);
+        auto simFeat = SimulateSimilarVector(feature);
+        HFFaceFeature simFeature = {0};
+        simFeature.data = simFeat.data();
+        simFeature.size = simFeat.size();
+        HFFaceFeature target = {0};
+        target.data = identity.feature->data;
+        target.size = identity.feature->size;
+        HFloat cosine;
+        ret = HFFaceComparison(target, simFeature, &cosine);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(cosine > 0.80f);
+        similarFeatures.push_back(feature);
+    }
+    REQUIRE(similarFeatures.size() == numberOfSimilar);
+
+    auto numberOfNotSimilar = preDataSample;
+    std::vector<std::vector<HFloat>> notSimilarFeatures;
+    // Generate some feature vectors that are not similar to the existing database
+    for (int i = 0; i < numberOfNotSimilar; ++i) {
+        auto feat = GenerateRandomFeature(featureLength);
+        HFFaceFeature feature = {0};
+        feature.size = feat.size();
+        feature.data = feat.data();
+        HFFaceFeatureIdentity mostSim = {0};
+        HFloat cosine;
+        HFFeatureHubFaceSearch(feature, &cosine, &mostSim);
+        REQUIRE(cosine < 0.3f);
+        notSimilarFeatures.push_back(feat);
+    }
+    REQUIRE(notSimilarFeatures.size() == numberOfNotSimilar);
+
+    // Multithreaded search simulation
+    const int numThreads = 5;
+    std::vector<std::thread> threads;
+    std::mutex mutex;
+
+    // Start threads for concurrent searching
+    for (int t = 0; t < numThreads; ++t) {
+        threads.emplace_back([&]() {
+            std::random_device rd;
+            std::mt19937 gen(rd());
+            std::uniform_int_distribution<> dis(0, preDataSample - 1);
+            for (int j = 0; j < 50; ++j) {  // Each thread performs 50 similar searches
+                int idx = dis(gen);
+                auto targetId = targetIds[idx];
+                HFFaceFeature feature = {0};
+                feature.data = similarFeatures[idx].data();
+                feature.size = similarFeatures[idx].size();
+                HFloat score;
+                HFFaceFeatureIdentity identity = {0};
+                HFFeatureHubFaceSearch(feature, &score, &identity);
+                REQUIRE(identity.id == targetId + 1);
+            }
+            for (int j = 0; j < 50; ++j) {
+                int idx = dis(gen);
+                HFFaceFeature feature = {0};
+                feature.data = notSimilarFeatures[idx].data();
+                feature.size = notSimilarFeatures[idx].size();
+                HFloat score;
+                HFFaceFeatureIdentity identity = {0};
+                HFFeatureHubFaceSearch(feature, &score, &identity);
+                REQUIRE(identity.id == -1);
+            }
+        });
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+
+    ret = HFFeatureHubDataDisable();
+    REQUIRE(ret == HSUCCEED);
+
+    delete[] dbPathStr;
+}
+
+TEST_CASE("test_FeatureCache", "[FeatureHub][Concurrency]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HFFeatureHubConfiguration configuration;
+    auto dbPath = GET_SAVE_DATA(".test");
+    HString dbPathStr = new char[dbPath.size() + 1];
+    std::strcpy(dbPathStr, dbPath.c_str());
+    configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+    configuration.enablePersistence = 1;
+    configuration.persistenceDbPath = dbPathStr;
+    configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+    configuration.searchThreshold = 0.48f;
+    // Delete the previous data before testing
+    if (std::remove(configuration.persistenceDbPath) != 0) {
+        spdlog::trace("Error deleting file");
+    }
+    ret = HFFeatureHubDataEnable(configuration);
+    REQUIRE(ret == HSUCCEED);
+
+    auto randomVec = GenerateRandomFeature(512);
+    HFFaceFeature feature = {0};
+    feature.data = randomVec.data();
+    feature.size = randomVec.size();
+    HFFaceFeatureIdentity identity = {0};
+    identity.feature = &feature;
+    HFaceId allocId;
+    ret = HFFeatureHubInsertFeature(identity, &allocId);
+    REQUIRE(ret == HSUCCEED);
+
+    auto simVec = SimulateSimilarVector(randomVec);
+    HFFaceFeature simFeature = {0};
+    simFeature.data = simVec.data();
+    simFeature.size = simVec.size();
+
+    for (int i = 0; i < 10; ++i) {
+        HFFaceFeatureIdentity capture = {0};
+        ret = HFFeatureHubGetFaceIdentity(allocId, &capture);
+        REQUIRE(ret == HSUCCEED);
+
+        HFFaceFeature target = {0};
+        target.data = capture.feature->data;
+        target.size = capture.feature->size;
+
+        HFloat cosine;
+        ret = HFFaceComparison(target, simFeature, &cosine);
+        REQUIRE(cosine > 0.8f);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    ret = HFFeatureHubDataDisable();
+    REQUIRE(ret == HSUCCEED);
+
+    delete[] dbPathStr;
+}
+
+TEST_CASE("test_FeatureHubManualInput", "[FeatureHub][ManualInput]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    HResult ret;
+    HFFeatureHubConfiguration configuration;
+    configuration.primaryKeyMode = HF_PK_MANUAL_INPUT;
+    configuration.enablePersistence = 0;
+    TEST_PRINT("Start enable feature hub");
+    ret = HFFeatureHubDataEnable(configuration);
+    REQUIRE(ret == HSUCCEED);
+    TEST_PRINT("Enable feature hub success");
+
+    std::vector<HFaceId> ids = {10086, 23541, 2124, 24, 204};
+
+    for (auto id : ids) {
+        auto randomVec = GenerateRandomFeature(512);
+        HFFaceFeature feature = {0};
+        feature.data = randomVec.data();
+        feature.size = randomVec.size();
+        HFFaceFeatureIdentity identity = {0};
+        identity.feature = &feature;
+        identity.id = id;
+        HFaceId allocId;
+        ret = HFFeatureHubInsertFeature(identity, &allocId);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    HFFeatureHubExistingIds existingIds = {0};
+    ret = HFFeatureHubGetExistingIds(&existingIds);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(existingIds.size == ids.size());
+    for (int i = 0; i < existingIds.size; ++i) {
+        TEST_PRINT("Existing ID: {}", existingIds.ids[i]);
+        REQUIRE(existingIds.ids[i] == ids[i]);
+    }
+
+    ret = HFFeatureHubViewDBTable();
+    REQUIRE(ret == HSUCCEED);
+
+    // query
+    for (auto id : ids) {
+        HFFaceFeatureIdentity query = {0};
+        ret = HFFeatureHubGetFaceIdentity(id, &query);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(query.id == id);
+    }
+
+    ret = HFFeatureHubDataDisable();
+    REQUIRE(ret == HSUCCEED);
+}
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_feature_manage.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_feature_manage.cpp
new file mode 100644
index 000000000..5c741f6cd
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_feature_manage.cpp
@@ -0,0 +1,982 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#if 0
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "opencv2/opencv.hpp"
+#include "unit/test_helper/simple_csv_writer.h"
+#include "unit/test_helper/test_help.h"
+
+TEST_CASE("test_FeatureManage", "[feature_manage]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Face feature management basic functions") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        // Get a face picture
+        cv::Mat kunImage = cv::imread(GET_DATA("data/bulk/kun.jpg"));
+        HFImageData imageData = {0};
+        imageData.data = kunImage.data;
+        imageData.height = kunImage.rows;
+        imageData.width = kunImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        // Insert data into feature management
+        HFFaceFeatureIdentity identity = {0};
+        identity.feature = &feature;
+        identity.tag = "chicken";
+        identity.customId = 1234;
+        ret = HFFeatureHubInsertFeature(identity);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check number
+        HInt32 num;
+        ret = HFFeatureHubGetFaceCount(&num);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(num == 1);
+
+        // Update Face info
+        HFFaceFeatureIdentity updatedIdentity = {0};
+        updatedIdentity.feature = identity.feature;
+        updatedIdentity.customId = identity.customId;
+        updatedIdentity.tag = "iKun";
+        ret = HFFeatureHubFaceUpdate(updatedIdentity);
+        REQUIRE(ret == HSUCCEED);
+
+        // Trying to update an identity that doesn't exist
+        HFFaceFeatureIdentity nonIdentity = {0};
+        nonIdentity.customId = 234;
+        nonIdentity.tag = "no";
+        nonIdentity.feature = &feature;
+        ret = HFFeatureHubFaceUpdate(nonIdentity);
+        REQUIRE(ret != HSUCCEED);
+
+        // Trying to delete an identity that doesn't exist
+        ret = HFFeatureHubFaceRemove(nonIdentity.customId);
+        REQUIRE(ret != HSUCCEED);
+
+        // Delete kunkun
+        ret = HFFeatureHubFaceRemove(identity.customId);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubGetFaceCount(&num);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(num == 0);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+    }
+
+    SECTION("Import a large faces data") {
+#ifdef ISF_ENABLE_USE_LFW_DATA
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        size_t numOfNeedImport = 1000;
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+
+#else
+        TEST_PRINT("The test case that uses LFW is not enabled, so it will be skipped.");
+#endif
+    }
+
+    SECTION("Faces feature CURD") {
+#ifdef ISF_ENABLE_USE_LFW_DATA
+        // This section needs to be connected to the "Import a large faces data" section before it
+        // can be executed
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        // Face track
+        cv::Mat dstImage = cv::imread(GET_DATA("data/bulk/Nathalie_Baye_0002.jpg"));
+        HFImageData imageData = {0};
+        imageData.data = dstImage.data;
+        imageData.height = dstImage.rows;
+        imageData.width = dstImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search for a face
+        HFloat confidence;
+        HFFaceFeatureIdentity searchedIdentity = {0};
+        ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedIdentity);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(searchedIdentity.customId == 898);
+        CHECK(std::string(searchedIdentity.tag) == "Nathalie_Baye");
+
+        // Delete kunkun and search
+        ret = HFFeatureHubFaceRemove(searchedIdentity.customId);
+        REQUIRE(ret == HSUCCEED);
+        // Search again
+        ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedIdentity);
+        //        spdlog::info("{}", confidence);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(searchedIdentity.customId == -1);
+
+        // Insert again
+        HFFaceFeatureIdentity againIdentity = {0};
+        againIdentity.customId = 898;
+        againIdentity.tag = "Cover";
+        againIdentity.feature = &feature;
+        ret = HFFeatureHubInsertFeature(againIdentity);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search again
+        HFFaceFeatureIdentity searchedAgainIdentity = {0};
+        ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedAgainIdentity);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(searchedAgainIdentity.customId == 898);
+
+        // Update any feature
+        HInt32 updateId = 909;
+        cv::Mat zyImage = cv::imread(GET_DATA("data/bulk/woman.png"));
+        HFImageData imageDataZy = {0};
+        imageDataZy.data = zyImage.data;
+        imageDataZy.height = zyImage.rows;
+        imageDataZy.width = zyImage.cols;
+        imageDataZy.format = HF_STREAM_BGR;
+        imageDataZy.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandleZy;
+        ret = HFCreateImageStream(&imageDataZy, &imgHandleZy);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceDataZy = {0};
+        ret = HFExecuteFaceTrack(session, imgHandleZy, &multipleFaceDataZy);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceDataZy.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature featureZy = {0};
+        ret = HFFaceFeatureExtract(session, imgHandleZy, multipleFaceDataZy.tokens[0], &featureZy);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandleZy);
+        REQUIRE(ret == HSUCCEED);
+
+        // Update id: 11297
+        HFFaceFeatureIdentity updateIdentity = {0};
+        updateIdentity.customId = updateId;
+        updateIdentity.tag = "ZY";
+        updateIdentity.feature = &featureZy;
+        ret = HFFeatureHubFaceUpdate(updateIdentity);
+        REQUIRE(ret == HSUCCEED);
+
+        //
+        // Prepare a zy query image
+        cv::Mat zyImageQuery = cv::imread(GET_DATA("data/bulk/woman_search.jpeg"));
+        HFImageData imageDataZyQuery = {0};
+        imageDataZyQuery.data = zyImageQuery.data;
+        imageDataZyQuery.height = zyImageQuery.rows;
+        imageDataZyQuery.width = zyImageQuery.cols;
+        imageDataZyQuery.format = HF_STREAM_BGR;
+        imageDataZyQuery.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandleZyQuery;
+        ret = HFCreateImageStream(&imageDataZyQuery, &imgHandleZyQuery);
+        REQUIRE(ret == HSUCCEED);
+        //
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceDataZyQuery = {0};
+        ret = HFExecuteFaceTrack(session, imgHandleZyQuery, &multipleFaceDataZyQuery);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceDataZyQuery.detectedNum > 0);
+        //
+        // Extract face feature
+        HFFaceFeature featureZyQuery = {0};
+        ret = HFFaceFeatureExtract(session, imgHandleZyQuery, multipleFaceDataZyQuery.tokens[0],
+                                   &featureZyQuery);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandleZyQuery);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search
+        HFloat confidenceQuery;
+        HFFaceFeatureIdentity searchedIdentityQuery = {0};
+        ret = HFFeatureHubFaceSearch(featureZyQuery, &confidenceQuery, &searchedIdentityQuery);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(searchedIdentityQuery.customId == updateId);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+#else
+        TEST_PRINT("The test case that uses LFW is not enabled, so it will be skipped.");
+#endif
+    }
+}
+
+TEST_CASE("test_SearchTopK", "[feature_search_top_k]") {
+#ifdef ISF_ENABLE_USE_LFW_DATA
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Face feature management basic functions") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.46f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        // Import 1k faces
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        size_t numOfNeedImport = 1000;
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        // Prepare multiple photos of a person
+        std::vector<std::string> photos = {
+          GET_DATA("data/RD/d1.jpeg"),
+          GET_DATA("data/RD/d2.jpeg"),
+          GET_DATA("data/RD/d3.jpeg"),
+          GET_DATA("data/RD/d4.jpeg"),
+        };
+        std::vector<std::string> tags = {
+          "d1",
+          "d2",
+          "d3",
+          "d4",
+        };
+        std::vector<HInt32> updateIds = {
+          5,
+          163,
+          670,
+          971,
+        };
+        REQUIRE(photos.size() == tags.size());
+        REQUIRE(updateIds.size() == tags.size());
+
+        // Replace the face features in the photo with each target in FeatureHub
+        for (int i = 0; i < photos.size(); ++i) {
+            // Face track
+            cv::Mat dstImage = cv::imread(photos[i]);
+            HFImageData imageData = {0};
+            imageData.data = dstImage.data;
+            imageData.height = dstImage.rows;
+            imageData.width = dstImage.cols;
+            imageData.format = HF_STREAM_BGR;
+            imageData.rotation = HF_CAMERA_ROTATION_0;
+            HFImageStream imgHandle;
+            ret = HFCreateImageStream(&imageData, &imgHandle);
+            REQUIRE(ret == HSUCCEED);
+
+            // Extract basic face information from photos
+            HFMultipleFaceData multipleFaceData = {0};
+            ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+            REQUIRE(ret == HSUCCEED);
+            REQUIRE(multipleFaceData.detectedNum > 0);
+
+            // Extract face feature
+            HFFaceFeature feature = {0};
+            ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+            REQUIRE(ret == HSUCCEED);
+
+            char* cstr = new char[tags[i].size() + 1];  // Dynamically allocate memory for the name
+            strcpy(cstr, tags[i].c_str());              // Copy the name into the allocated memory
+
+            // Create identity
+            HFFaceFeatureIdentity identity = {0};
+            identity.customId = updateIds[i];
+            identity.feature = &feature;
+            identity.tag = cstr;
+
+            // Update
+            ret = HFFeatureHubFaceUpdate(identity);
+            REQUIRE(ret == HSUCCEED);
+
+            ret = HFReleaseImageStream(imgHandle);
+            REQUIRE(ret == HSUCCEED);
+            delete[] cstr;  // Clean up the dynamically allocated memory
+        }
+
+        // Prepare a target photo for a face top-k search
+        cv::Mat image = cv::imread(GET_DATA("data/RD/d5.jpeg"));
+        HFImageData imageData = {0};
+        imageData.data = image.data;
+        imageData.height = image.rows;
+        imageData.width = image.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        // Run the top-k search
+        HFSearchTopKResults topk;
+        ret = HFFeatureHubFaceSearchTopK(feature, 10, &topk);
+        REQUIRE(ret == HSUCCEED);
+
+        // Check whether the top-k result is consistent with the expectation
+        CHECK(topk.size == photos.size());
+        for (int i = 0; i < topk.size; ++i) {
+            TEST_PRINT("Top-{} -> id: {}, {}", i + 1, topk.customIds[i], topk.confidence[i]);
+            CHECK(std::find(updateIds.begin(), updateIds.end(), topk.customIds[i]) !=
+                  updateIds.end());
+        }
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+    }
+#endif
+}
+
+TEST_CASE("test_FeatureBenchmark", "[feature_benchmark]") {
+    // Test the search time at 1k, 5k and 10k of the face library (the target face is at the back).
+    SECTION("Search face benchmark from 1k") {
+#if defined(ISF_ENABLE_BENCHMARK) && defined(ISF_ENABLE_USE_LFW_DATA)
+        size_t loop = 1000;
+        size_t numOfNeedImport = 1000;
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        //        TEST_PRINT("{}", dataList.size());
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        // Face track
+        cv::Mat dstImage = cv::imread(GET_DATA("data/search/Teresa_Williams_0001_1k.jpg"));
+        HFImageData imageData = {0};
+        imageData.data = dstImage.data;
+        imageData.height = dstImage.rows;
+        imageData.width = dstImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search for a face
+        HFloat confidence;
+        HFFaceFeatureIdentity searchedIdentity = {0};
+        auto start = (double)cv::getTickCount();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedIdentity);
+        }
+        auto cost = ((double)cv::getTickCount() - start) / cv::getTickFrequency() * 1000;
+
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(searchedIdentity.customId == 999);
+        REQUIRE(std::string(searchedIdentity.tag) == "Teresa_Williams");
+
+        TEST_PRINT(
+          "<Benchmark> Search Face from 1k -> Loop: {}, Total Time: {:.5f}ms, Average Time: "
+          "{:.5f}ms",
+          loop, cost, cost / loop);
+
+        BenchmarkRecord record(getBenchmarkRecordFile());
+        record.insertBenchmarkData("Search Face from 1k", loop, cost, cost / loop);
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+#else
+        TEST_PRINT(
+          "Skip face search benchmark test, you need to enable both lfw and benchmark test.");
+#endif
+    }
+
+    SECTION("Search face benchmark from 5k") {
+#if defined(ISF_ENABLE_BENCHMARK) && defined(ISF_ENABLE_USE_LFW_DATA)
+        size_t loop = 1000;
+        size_t numOfNeedImport = 5000;
+        HResult ret;
+        std::string modelPath = GET_MODEL_FILE();
+        HPath path = modelPath.c_str();
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        // Face track
+        cv::Mat dstImage = cv::imread(GET_DATA("data/search/Mary_Katherine_Smart_0001_5k.jpg"));
+        HFImageData imageData = {0};
+        imageData.data = dstImage.data;
+        imageData.height = dstImage.rows;
+        imageData.width = dstImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search for a face
+        HFloat confidence;
+        HFFaceFeatureIdentity searchedIdentity = {0};
+        auto start = (double)cv::getTickCount();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedIdentity);
+        }
+        auto cost = ((double)cv::getTickCount() - start) / cv::getTickFrequency() * 1000;
+
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(searchedIdentity.customId == 4998);
+        REQUIRE(std::string(searchedIdentity.tag) == "Mary_Katherine_Smart");
+
+        TEST_PRINT(
+          "<Benchmark> Search Face from 5k -> Loop: {}, Total Time: {:.5f}ms, Average Time: "
+          "{:.5f}ms",
+          loop, cost, cost / loop);
+
+        BenchmarkRecord record(getBenchmarkRecordFile());
+        record.insertBenchmarkData("Search Face from 5k", loop, cost, cost / loop);
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+#else
+        TEST_PRINT(
+          "Skip face search benchmark test, you need to enable both lfw and benchmark test.");
+#endif
+    }
+
+    SECTION("Search face benchmark from 10k") {
+#if defined(ISF_ENABLE_BENCHMARK) && defined(ISF_ENABLE_USE_LFW_DATA)
+        size_t loop = 1000;
+        size_t numOfNeedImport = 10000;
+        HResult ret;
+        std::string modelPath = GET_MODEL_FILE();
+        HPath path = modelPath.c_str();
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        //        TEST_PRINT("{}", dataList.size());
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        // Update any feature
+        HInt32 updateId = numOfNeedImport - 1;
+        cv::Mat zyImage = cv::imread(GET_DATA("data/bulk/woman.png"));
+        HFImageData imageDataZy = {0};
+        imageDataZy.data = zyImage.data;
+        imageDataZy.height = zyImage.rows;
+        imageDataZy.width = zyImage.cols;
+        imageDataZy.format = HF_STREAM_BGR;
+        imageDataZy.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandleZy;
+        ret = HFCreateImageStream(&imageDataZy, &imgHandleZy);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceDataZy = {0};
+        ret = HFExecuteFaceTrack(session, imgHandleZy, &multipleFaceDataZy);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceDataZy.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature featureZy = {0};
+        ret = HFFaceFeatureExtract(session, imgHandleZy, multipleFaceDataZy.tokens[0], &featureZy);
+        REQUIRE(ret == HSUCCEED);
+
+        // Update id: 11297
+        HFFaceFeatureIdentity updateIdentity = {0};
+        updateIdentity.customId = updateId;
+        updateIdentity.tag = "ZY";
+        updateIdentity.feature = &featureZy;
+        ret = HFFeatureHubFaceUpdate(updateIdentity);
+        REQUIRE(ret == HSUCCEED);
+
+        HFReleaseImageStream(imgHandleZy);
+
+        // Face track
+        cv::Mat dstImage = cv::imread(GET_DATA("data/bulk/woman_search.jpeg"));
+        HFImageData imageData = {0};
+        imageData.data = dstImage.data;
+        imageData.height = dstImage.rows;
+        imageData.width = dstImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        REQUIRE(ret == HSUCCEED);
+
+        // Search for a face
+        HFloat confidence;
+        HFFaceFeatureIdentity searchedIdentity = {0};
+        auto start = (double)cv::getTickCount();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFFeatureHubFaceSearch(feature, &confidence, &searchedIdentity);
+        }
+        auto cost = ((double)cv::getTickCount() - start) / cv::getTickFrequency() * 1000;
+
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(searchedIdentity.customId == updateId);
+        REQUIRE(std::string(searchedIdentity.tag) == "ZY");
+
+        TEST_PRINT(
+          "<Benchmark> Search Face from 10k -> Loop: {}, Total Time: {:.5f}ms, Average Time: "
+          "{:.5f}ms",
+          loop, cost, cost / loop);
+
+        BenchmarkRecord record(getBenchmarkRecordFile());
+        record.insertBenchmarkData("Search Face from 10k", loop, cost, cost / loop);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+#else
+        TEST_PRINT(
+          "Skip face search benchmark test, you need to enable both lfw and benchmark test.");
+#endif
+    }
+
+    SECTION("Face comparison benchmark") {
+#ifdef ISF_ENABLE_BENCHMARK
+        int loop = 1000;
+        HResult ret;
+        std::string modelPath = GET_MODEL_FILE();
+        HPath path = modelPath.c_str();
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        cv::Mat zyImage = cv::imread(GET_DATA("data/bulk/woman.png"));
+        HFImageData imageDataZy = {0};
+        imageDataZy.data = zyImage.data;
+        imageDataZy.height = zyImage.rows;
+        imageDataZy.width = zyImage.cols;
+        imageDataZy.format = HF_STREAM_BGR;
+        imageDataZy.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandleZy;
+        ret = HFCreateImageStream(&imageDataZy, &imgHandleZy);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceDataZy = {0};
+        ret = HFExecuteFaceTrack(session, imgHandleZy, &multipleFaceDataZy);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceDataZy.detectedNum > 0);
+
+        HInt32 featureNum;
+        HFGetFeatureLength(&featureNum);
+
+        // Extract face feature
+        HFloat featureCacheZy[featureNum];
+        ret = HFFaceFeatureExtractCpy(session, imgHandleZy, multipleFaceDataZy.tokens[0],
+                                      featureCacheZy);
+        HFFaceFeature featureZy = {0};
+        featureZy.size = featureNum;
+        featureZy.data = featureCacheZy;
+        REQUIRE(ret == HSUCCEED);
+
+        cv::Mat zyImageQuery = cv::imread(GET_DATA("data/bulk/woman_search.jpeg"));
+        HFImageData imageDataZyQuery = {0};
+        imageDataZyQuery.data = zyImageQuery.data;
+        imageDataZyQuery.height = zyImageQuery.rows;
+        imageDataZyQuery.width = zyImageQuery.cols;
+        imageDataZyQuery.format = HF_STREAM_BGR;
+        imageDataZyQuery.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandleZyQuery;
+        ret = HFCreateImageStream(&imageDataZyQuery, &imgHandleZyQuery);
+        REQUIRE(ret == HSUCCEED);
+        //
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceDataZyQuery = {0};
+        ret = HFExecuteFaceTrack(session, imgHandleZyQuery, &multipleFaceDataZyQuery);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceDataZyQuery.detectedNum > 0);
+        //
+        // Extract face feature
+        HFloat featureCacheZyQuery[featureNum];
+        ret = HFFaceFeatureExtractCpy(session, imgHandleZyQuery, multipleFaceDataZyQuery.tokens[0],
+                                      featureCacheZyQuery);
+        HFFaceFeature featureZyQuery = {0};
+        featureZyQuery.data = featureCacheZyQuery;
+        featureZyQuery.size = featureNum;
+        REQUIRE(ret == HSUCCEED);
+
+        auto start = (double)cv::getTickCount();
+        for (int i = 0; i < loop; ++i) {
+            HFloat compRes;
+            ret = HFFaceComparison(featureZy, featureZyQuery, &compRes);
+        }
+        auto cost = ((double)cv::getTickCount() - start) / cv::getTickFrequency() * 1000;
+        REQUIRE(ret == HSUCCEED);
+        TEST_PRINT(
+          "<Benchmark> Face Comparison -> Loop: {}, Total Time: {:.5f}ms, Average Time: {:.5f}ms",
+          loop, cost, cost / loop);
+
+        BenchmarkRecord record(getBenchmarkRecordFile());
+        record.insertBenchmarkData("Face Comparison", loop, cost, cost / loop);
+
+        HFReleaseImageStream(imgHandleZy);
+        HFReleaseImageStream(imgHandleZyQuery);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+#else
+        TEST_PRINT("The benchmark is not enabled, so all relevant test cases are skipped.");
+#endif
+    }
+
+    SECTION("Face feature extract benchmark") {
+#ifdef ISF_ENABLE_BENCHMARK
+        int loop = 1000;
+        HResult ret;
+        std::string modelPath = GET_MODEL_FILE();
+        HPath path = modelPath.c_str();
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration = {0};
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.enablePersistence = 1;
+        configuration.dbPath = dbPathStr;
+        configuration.featureBlockNum = 20;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.dbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        // Face track
+        cv::Mat dstImage = cv::imread(GET_DATA("data/search/Teresa_Williams_0001_1k.jpg"));
+        HFImageData imageData = {0};
+        imageData.data = dstImage.data;
+        imageData.height = dstImage.rows;
+        imageData.width = dstImage.cols;
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        ret = HFCreateImageStream(&imageData, &imgHandle);
+        REQUIRE(ret == HSUCCEED);
+
+        // Extract basic face information from photos
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(session, imgHandle, &multipleFaceData);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(multipleFaceData.detectedNum > 0);
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        auto start = (double)cv::getTickCount();
+        for (int i = 0; i < loop; ++i) {
+            ret = HFFaceFeatureExtract(session, imgHandle, multipleFaceData.tokens[0], &feature);
+        }
+        auto cost = ((double)cv::getTickCount() - start) / cv::getTickFrequency() * 1000;
+        REQUIRE(ret == HSUCCEED);
+        TEST_PRINT(
+          "<Benchmark> Face Extract -> Loop: {}, Total Time: {:.5f}ms, Average Time: {:.5f}ms",
+          loop, cost, cost / loop);
+
+        BenchmarkRecord record(getBenchmarkRecordFile());
+        record.insertBenchmarkData("Face Extract", loop, cost, cost / loop);
+
+        HFReleaseImageStream(imgHandle);
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+        delete[] dbPathStr;
+#else
+        TEST_PRINT(
+          "Skip the face feature extraction benchmark test. To run it, you need to turn on the "
+          "benchmark test.");
+#endif
+    }
+}
+#endif
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_help_tools.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_help_tools.cpp
new file mode 100644
index 000000000..cc2596766
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_help_tools.cpp
@@ -0,0 +1,66 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "../test_helper/test_help.h"
+
+TEST_CASE("test_HelpTools", "[help_tools]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Load lfw funneled data") {
+#ifdef ISF_ENABLE_USE_LFW_DATA
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        HFFeatureHubConfiguration configuration;
+        auto dbPath = GET_SAVE_DATA(".test");
+        HString dbPathStr = new char[dbPath.size() + 1];
+        std::strcpy(dbPathStr, dbPath.c_str());
+        configuration.primaryKeyMode = HF_PK_AUTO_INCREMENT;
+        configuration.enablePersistence = 1;
+        configuration.persistenceDbPath = dbPathStr;
+        configuration.searchMode = HF_SEARCH_MODE_EXHAUSTIVE;
+        configuration.searchThreshold = 0.48f;
+        // Delete the previous data before testing
+        if (std::remove(configuration.persistenceDbPath) != 0) {
+            spdlog::trace("Error deleting file");
+        }
+        ret = HFFeatureHubDataEnable(configuration);
+        REQUIRE(ret == HSUCCEED);
+
+        auto lfwDir = getLFWFunneledDir();
+        auto dataList = LoadLFWFunneledValidData(lfwDir, getTestLFWFunneledTxt());
+        size_t numOfNeedImport = 100;
+        auto importStatus = ImportLFWFunneledValidData(session, dataList, numOfNeedImport);
+        HFFeatureHubViewDBTable();
+        REQUIRE(importStatus);
+        HInt32 count;
+        ret = HFFeatureHubGetFaceCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        CHECK(count == numOfNeedImport);
+
+        //        ret = HF_ViewFaceDBTable(session);
+        //        REQUIRE(ret == HSUCCEED);
+
+        // Finish
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFFeatureHubDataDisable();
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] dbPathStr;
+
+#else
+        TEST_PRINT("The test case that uses LFW is not enabled, so it will be skipped.");
+#endif
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_image_bitmap.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_image_bitmap.cpp
new file mode 100644
index 000000000..f6516ae64
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_image_bitmap.cpp
@@ -0,0 +1,45 @@
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include <cstdio>
+
+TEST_CASE("test_ImageBitmap", "[image_bitmap]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HFImageBitmap handle;
+    HResult ret = HFCreateImageBitmapFromFilePath(GET_DATA("data/bulk/r90.jpg").c_str(), 3, &handle);
+    REQUIRE(ret == HSUCCEED);
+
+    HFImageStream stream;
+    ret = HFCreateImageStreamFromImageBitmap(handle, HF_CAMERA_ROTATION_90, &stream);
+    REQUIRE(ret == HSUCCEED);
+
+    HFSessionCustomParameter parameter = {0};
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+    TEST_ERROR_PRINT("error ret :{}", ret);
+    REQUIRE(ret == HSUCCEED);
+
+    // Extract basic face information from photos
+    HFMultipleFaceData multipleFaceData = {0};
+    ret = HFExecuteFaceTrack(session, stream, &multipleFaceData);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(multipleFaceData.detectedNum == 1);
+
+    auto rect = multipleFaceData.rects[0];
+    HColor color = {0, 0, 255};
+    HFImageBitmapDrawRect(handle, rect, color, 2);
+    HFImageBitmapWriteToFile(handle, "bitmap_draw_test.jpg");
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseImageStream(stream);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseImageBitmap(handle);
+    REQUIRE(ret == HSUCCEED);
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_image_process.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_image_process.cpp
new file mode 100644
index 000000000..af306af9e
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_image_process.cpp
@@ -0,0 +1,215 @@
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include <cstdio>
+
+uint8_t* ReadNV21File(const char* filepath, size_t* fileSize) {
+    FILE* fp = fopen(filepath, "rb");
+    if (!fp) {
+        if (fileSize)
+            *fileSize = 0;
+        return nullptr;
+    }
+
+    fseek(fp, 0, SEEK_END);
+    size_t size = ftell(fp);
+    fseek(fp, 0, SEEK_SET);
+
+    uint8_t* data = new uint8_t[size];
+
+    size_t read_size = fread(data, 1, size, fp);
+    fclose(fp);
+
+    if (read_size != size) {
+        delete[] data;
+        if (fileSize)
+            *fileSize = 0;
+        return nullptr;
+    }
+
+    if (fileSize)
+        *fileSize = size;
+    return data;
+}
+
+TEST_CASE("test_ImageProcessRotateNV21", "[image_process]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HFImageBitmap originBmp;
+    size_t fileSize;
+    uint8_t* data = ReadNV21File(GET_DATA("data/bulk/r0_w330_h409_c3.nv21").c_str(), &fileSize);
+    REQUIRE(data != nullptr);
+
+    HFImageData imageData;
+    imageData.data = data;
+    imageData.width = 330;
+    imageData.height = 409;
+    imageData.rotation = HF_CAMERA_ROTATION_0;
+    imageData.format = HF_STREAM_YUV_NV21;
+
+    HFImageStream stream;
+    HResult ret = HFCreateImageStream(&imageData, &stream);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFCreateImageBitmapFromImageStreamProcess(stream, &originBmp, 1, 1.0f);
+    REQUIRE(ret == HSUCCEED);
+
+    HFImageBitmapData originData;
+    ret = HFImageBitmapGetData(originBmp, &originData);
+    REQUIRE(ret == HSUCCEED);
+
+    // compare with eps(0~1)
+    float eps = 0.01;
+
+    SECTION("rotate 90") {
+        size_t fileSize;
+        uint8_t* r90nv21 = ReadNV21File(GET_DATA("data/bulk/r90_w409_h330_c3.nv21").c_str(), &fileSize);
+        REQUIRE(r90nv21 != nullptr);
+
+        HFImageData imageData;
+        imageData.data = r90nv21;
+        imageData.width = 409;
+        imageData.height = 330;
+        imageData.rotation = HF_CAMERA_ROTATION_90;
+        imageData.format = HF_STREAM_YUV_NV21;
+
+        HFImageStream stream;
+        ret = HFCreateImageStream(&imageData, &stream);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmap rot90;
+        ret = HFCreateImageBitmapFromImageStreamProcess(stream, &rot90, 1, 1.0f);
+        REQUIRE(ret == HSUCCEED);
+
+        // HFImageBitmapShow(rot90, "w", 0);
+
+        HFImageBitmapData rot90Data;
+        ret = HFImageBitmapGetData(rot90, &rot90Data);
+        REQUIRE(ret == HSUCCEED);
+
+        REQUIRE_EQ_IMAGE_WITH_EPS(originData.data, rot90Data.data, originData.height, originData.width, originData.channels, eps);
+
+        ret = HFReleaseImageBitmap(rot90);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(stream);
+        REQUIRE(ret == HSUCCEED);
+
+        delete[] r90nv21;
+    }
+
+    ret = HFReleaseImageStream(stream);
+    REQUIRE(ret == HSUCCEED);
+
+    ret = HFReleaseImageBitmap(originBmp);
+    REQUIRE(ret == HSUCCEED);
+
+    delete[] data;
+}
+
+TEST_CASE("test_ImageProcessRotate", "[image_process]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HFImageBitmap originBmp;
+    HResult ret = HFCreateImageBitmapFromFilePath(GET_DATA("data/bulk/r0.jpg").c_str(), 3, &originBmp);
+    REQUIRE(ret == HSUCCEED);
+
+    HFImageBitmapData originData;
+    ret = HFImageBitmapGetData(originBmp, &originData);
+    REQUIRE(ret == HSUCCEED);
+
+    // compare with eps(0~1)
+    float eps = 0.001;
+
+    SECTION("rotate 90") {
+        HFImageBitmap bitmap;
+        HResult ret = HFCreateImageBitmapFromFilePath(GET_DATA("data/bulk/r90.jpg").c_str(), 3, &bitmap);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageStream stream;
+        ret = HFCreateImageStreamFromImageBitmap(bitmap, HF_CAMERA_ROTATION_90, &stream);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmap rot90;
+        ret = HFCreateImageBitmapFromImageStreamProcess(stream, &rot90, 1, 1.0f);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmapData rot90Data;
+        ret = HFImageBitmapGetData(rot90, &rot90Data);
+        REQUIRE(ret == HSUCCEED);
+
+        REQUIRE_EQ_IMAGE_WITH_EPS(originData.data, rot90Data.data, originData.height, originData.width, originData.channels, eps);
+
+        ret = HFReleaseImageBitmap(rot90);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(stream);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageBitmap(bitmap);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("rotate 180") {
+        HFImageBitmap bitmap;
+        HResult ret = HFCreateImageBitmapFromFilePath(GET_DATA("data/bulk/r180.jpg").c_str(), 3, &bitmap);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageStream stream;
+        ret = HFCreateImageStreamFromImageBitmap(bitmap, HF_CAMERA_ROTATION_180, &stream);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmap rot180;
+        ret = HFCreateImageBitmapFromImageStreamProcess(stream, &rot180, 1, 1.0f);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmapData rot180Data;
+        ret = HFImageBitmapGetData(rot180, &rot180Data);
+        REQUIRE(ret == HSUCCEED);
+
+        REQUIRE_EQ_IMAGE_WITH_EPS(originData.data, rot180Data.data, originData.height, originData.width, originData.channels, eps);
+
+        ret = HFReleaseImageBitmap(rot180);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(stream);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageBitmap(bitmap);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("rotate 270") {
+        HFImageBitmap bitmap;
+        HResult ret = HFCreateImageBitmapFromFilePath(GET_DATA("data/bulk/r270.jpg").c_str(), 3, &bitmap);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageStream stream;
+        ret = HFCreateImageStreamFromImageBitmap(bitmap, HF_CAMERA_ROTATION_270, &stream);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmap rot270;
+        ret = HFCreateImageBitmapFromImageStreamProcess(stream, &rot270, 1, 1.0f);
+        REQUIRE(ret == HSUCCEED);
+
+        HFImageBitmapData rot270Data;
+        ret = HFImageBitmapGetData(rot270, &rot270Data);
+        REQUIRE(ret == HSUCCEED);
+
+        REQUIRE_EQ_IMAGE_WITH_EPS(originData.data, rot270Data.data, originData.height, originData.width, originData.channels, eps);
+
+        ret = HFReleaseImageBitmap(rot270);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageStream(stream);
+        REQUIRE(ret == HSUCCEED);
+
+        ret = HFReleaseImageBitmap(bitmap);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    ret = HFReleaseImageBitmap(originBmp);
+    REQUIRE(ret == HSUCCEED);
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_session_parallel.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_session_parallel.cpp
new file mode 100644
index 000000000..4f60e6cb1
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_session_parallel.cpp
@@ -0,0 +1,155 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/test_help.h"
+#include "inspireface/middleware/thread/resource_pool.h"
+#include <thread>
+
+TEST_CASE("test_SessionParallel", "[Session][Parallel]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    auto image1 = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+    auto image2 = inspirecv::Image::Create(GET_DATA("data/bulk/jntm.jpg"));
+
+    int loop = 100;
+
+    // Run it once to make sure the similarity is stable
+    HFSessionCustomParameter parameter = {0};
+    parameter.enable_recognition = 1;
+    HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+    HFSession session;
+    HResult ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+    REQUIRE(ret == HSUCCEED);
+    float expectedSimilarity = GenerateRandomNumbers(1, 0, 100)[0] / 100.0f;
+    ret = CompareTwoFaces(session, image1, image2, expectedSimilarity);
+    REQUIRE(ret);
+    TEST_PRINT("Expected similarity: {}", expectedSimilarity);
+
+    ret = HFReleaseInspireFaceSession(session);
+    REQUIRE(ret == HSUCCEED);
+
+    SECTION("Serial") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+
+        float similarity = 0.0f;
+        inspire::SpendTimer timeSpend("Serial loop: " + std::to_string(loop));
+        timeSpend.Start();
+        for (int i = 0; i < loop; ++i) {
+            ret = CompareTwoFaces(session, image1, image2, similarity);
+            REQUIRE(ret);
+            REQUIRE(similarity == Approx(expectedSimilarity).epsilon(0.01));
+        }
+        timeSpend.Stop();
+        std::cout << timeSpend << std::endl;
+        ;
+
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+
+    SECTION("Parallel") {
+        int N = 4;  // Use 4 sessions in parallel
+#ifdef ISF_RKNPU_RV1106
+        N = 1;  // Use 1 session in parallel
+#endif
+        inspire::parallel::ResourcePool<HFSession> sessionPool(N, [](HFSession& session) {
+            auto ret = HFReleaseInspireFaceSession(session);
+            if (ret != HSUCCEED) {
+                TEST_ERROR_PRINT("Failed to release session: {}", ret);
+            }
+        });
+
+        // Example Initialize N sessions to the resource pool
+        for (int i = 0; i < N; ++i) {
+            HFSessionCustomParameter parameter = {0};
+            parameter.enable_recognition = 1;
+            HFSession session;
+            HResult ret = HFCreateInspireFaceSession(parameter, HF_DETECT_MODE_ALWAYS_DETECT, 3, -1, -1, &session);
+            REQUIRE(ret == HSUCCEED);
+            sessionPool.AddResource(std::move(session));
+        }
+
+        // Create a thread pool to execute a task
+        std::vector<std::thread> threads;
+        std::atomic<int> completed(0);
+        float similaritySum = 0.0f;
+        std::mutex similarityMutex;
+
+        inspire::SpendTimer timeSpend("Parallel loop: " + std::to_string(loop) + ", thread: " + std::to_string(N));
+        timeSpend.Start();
+
+        // Start worker thread
+        int tasksPerThread = loop / N;
+        int remainingTasks = loop % N;
+
+        for (int i = 0; i < N; ++i) {
+            int taskCount = tasksPerThread + (i < remainingTasks ? 1 : 0);
+            threads.emplace_back([&, taskCount]() {
+                for (int j = 0; j < taskCount; ++j) {
+                    auto sessionGuard = sessionPool.AcquireResource();
+                    float similarity = 0.0f;
+                    HResult ret = CompareTwoFaces(*sessionGuard, image1, image2, similarity);
+                    REQUIRE(ret);
+
+                    REQUIRE(similarity == Approx(expectedSimilarity).epsilon(0.01));
+                    {
+                        std::lock_guard<std::mutex> lock(similarityMutex);
+                        similaritySum += similarity;
+                    }
+
+                    completed++;
+                }
+            });
+        }
+
+        // Wait for all threads to complete
+        for (auto& thread : threads) {
+            thread.join();
+        }
+
+        timeSpend.Stop();
+        std::cout << timeSpend << std::endl;
+        ;
+
+        // Optional: Output average similarity(stability)
+        TEST_PRINT("Average similarity: {}", (similaritySum / loop));
+    }
+}
+
+TEST_CASE("test_SessionParallel_Memory", "[Session][Parallel][Memory]") {
+    size_t memoryUsage = getCurrentMemoryUsage();
+    TEST_PRINT("Current memory usage: {}MB", memoryUsage);
+    int loop = 4;
+#ifdef ISF_RKNPU_RV1106
+    loop = 1;
+#endif
+    std::vector<HFSession> sessions;
+    for (int i = 0; i < loop; ++i) {
+        HFSessionCustomParameter parameter = {0};
+        parameter.enable_recognition = 1;
+        HFSession session;
+        HResult ret = HFCreateInspireFaceSession(parameter, HF_DETECT_MODE_ALWAYS_DETECT, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        sessions.push_back(session);
+        size_t memoryUsage = getCurrentMemoryUsage();
+        TEST_PRINT("[alloc{}] Current memory usage: {}MB", i + 1, memoryUsage);
+    }
+    // Release all sessions
+    for (int i = 0; i < loop; ++i) {
+        auto ret = HFReleaseInspireFaceSession(sessions[i]);
+        REQUIRE(ret == HSUCCEED);
+        size_t memoryUsage = getCurrentMemoryUsage();
+        TEST_PRINT("[free{}] Current memory usage: {}MB", i + 1, memoryUsage);
+    }
+}
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_similarity_converter.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_similarity_converter.cpp
new file mode 100644
index 000000000..5fa30f5c1
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_similarity_converter.cpp
@@ -0,0 +1,51 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-01-20
+ */
+#include <iostream>
+#include "settings/test_settings.h"
+#include "unit/test_helper/test_help.h"
+#include "inspireface/include/inspireface/similarity_converter.h"
+
+TEST_CASE("test_similarity_converter", "[similarity_converter]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("test_similarity_converter_0.42") {
+        inspire::SimilarityConverterConfig config;
+        config.threshold = 0.42;
+        config.middleScore = 0.6;
+        config.steepness = 8.0;
+        config.outputMin = 0.01;
+        config.outputMax = 1.0;
+        inspire::SimilarityConverter similarity_converter(config);
+        std::vector<double> test_points = {-0.80, -0.20, 0.02, 0.10, 0.25, 0.30, 0.48, 0.70, 0.80, 0.90, 1.00};
+        std::vector<double> expected_scores = {0.0101, 0.0201, 0.0661, 0.1113, 0.2819, 0.3673, 0.7074, 0.9334, 0.9689, 0.9858, 0.9936};
+        REQUIRE(test_points.size() == expected_scores.size());
+
+        for (size_t i = 0; i < test_points.size(); ++i) {
+            double cosine = test_points[i];
+            double similarity = similarity_converter.convert(cosine);
+            REQUIRE(similarity == Approx(expected_scores[i]).epsilon(0.01));
+        }
+    }
+
+    SECTION("test_similarity_converter_0.32") {
+        inspire::SimilarityConverterConfig config;
+        config.threshold = 0.32;
+        config.middleScore = 0.6;
+        config.steepness = 10.0;
+        config.outputMin = 0.02;
+        config.outputMax = 1.0;
+        inspire::SimilarityConverter similarity_converter(config);
+        std::vector<double> test_points = {-0.80, -0.20, 0.02, 0.10, 0.25, 0.32, 0.50, 0.70, 0.80, 0.90, 1.00};
+        std::vector<double> expected_scores = {0.0200, 0.0278, 0.0860, 0.1557, 0.4302, 0.6000, 0.8997, 0.9851, 0.9945, 0.9980, 0.9992};
+        REQUIRE(test_points.size() == expected_scores.size());
+
+        for (size_t i = 0; i < test_points.size(); ++i) {
+            double cosine = test_points[i];
+            double similarity = similarity_converter.convert(cosine);
+            REQUIRE(similarity == Approx(expected_scores[i]).epsilon(0.01));
+        }
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/api/test_system.cpp b/cpp-package/inspireface/cpp/test/unit/api/test_system.cpp
new file mode 100644
index 000000000..229ebcf01
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/api/test_system.cpp
@@ -0,0 +1,272 @@
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "inspireface/include/inspireface/herror.h"
+#include "unit/test_helper/test_tools.h"
+#include <cstdio>
+
+TEST_CASE("test_System", "[system]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    HResult ret;
+    HInt32 status;
+    ret = HFQueryInspireFaceLaunchStatus(&status);
+    REQUIRE(ret == HSUCCEED);
+    if (status == HF_STATUS_ENABLE) {
+        // The global TEST environment has been started, so this side needs to be temporarily closed
+        // before testing
+        HFTerminateInspireFace();
+    }
+
+    SECTION("Create a session test when it is not loaded") {
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HERR_ARCHIVE_NOT_LOAD);
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HERR_INVALID_CONTEXT_HANDLE);
+    }
+
+    // Restart and start InspireFace
+    ret = HFLaunchInspireFace(GET_RUNTIME_FULLPATH_NAME.c_str());
+    REQUIRE(ret == HSUCCEED);
+
+    SECTION("Create a session test when it is reloaded") {
+        HResult ret;
+        HFSessionCustomParameter parameter = {0};
+        HFDetectMode detMode = HF_DETECT_MODE_ALWAYS_DETECT;
+        HFSession session;
+        ret = HFCreateInspireFaceSession(parameter, detMode, 3, -1, -1, &session);
+        REQUIRE(ret == HSUCCEED);
+        ret = HFReleaseInspireFaceSession(session);
+        REQUIRE(ret == HSUCCEED);
+    }
+}
+
+TEST_CASE("test_SystemSessionReleaseCase", "[system]") {
+    /**
+     * @brief Test the release of sessions
+     * @details Test the release of sessions and check the unreleased sessions count
+     */
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    HResult ret;
+
+    SECTION("CreateSessions") {
+        /**
+         * @brief Create sessions
+         * @details Create 10 sessions and check the unreleased sessions count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 0);
+
+        HInt32 createCount = 10;
+        HFSession sessions[createCount];
+
+        for (int i = 0; i < createCount; ++i) {
+            ret = HFCreateInspireFaceSessionOptional(HF_ENABLE_NONE, HF_DETECT_MODE_ALWAYS_DETECT, 3, -1, -1, &sessions[i]);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == createCount);
+
+        HFSession sessionsGet[createCount];
+        ret = HFDeBugGetUnreleasedSessions(sessionsGet, createCount);
+        REQUIRE(ret == HSUCCEED);
+        // The session list obtained from the api is also unordered because it is sorted internally using an unordered dictionary
+        for (int i = 0; i < createCount; ++i) {
+            bool found = false;
+            for (int j = 0; j < createCount; ++j) {
+                if (sessions[i] == sessionsGet[j]) {
+                    found = true;
+                    break;
+                }
+            }
+            REQUIRE(found);
+        }
+    }
+
+    SECTION("ReleaseSomeSessions") {
+        /**
+         * @brief Release some sessions
+         * @details Release some sessions and check the unreleased sessions count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 10);
+        auto createCount = count;
+
+        HFSession sessionsGet[createCount];
+        ret = HFDeBugGetUnreleasedSessions(sessionsGet, createCount);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<int32_t> releaseIndex = {0, 2, 4, 6, 8};
+        for (int i = 0; i < releaseIndex.size(); ++i) {
+            ret = HFReleaseInspireFaceSession(sessionsGet[releaseIndex[i]]);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == createCount - releaseIndex.size());
+
+        HFSession sessionsGet2[count];
+        ret = HFDeBugGetUnreleasedSessions(sessionsGet2, count);
+        REQUIRE(ret == HSUCCEED);
+        for (int i = 0; i < count; ++i) {
+            bool found = false;
+            for (int j = 0; j < releaseIndex.size(); ++j) {
+                if (sessionsGet2[i] == sessionsGet[releaseIndex[j]]) {
+                    found = true;
+                    break;
+                }
+            }
+            REQUIRE(!found);
+        }
+    }
+
+    SECTION("ReleaseAllSessions") {
+        /**
+         * @brief Release all sessions
+         * @details Release all sessions and check the unreleased sessions count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 5);
+
+        HFSession sessionsGet[count];
+        ret = HFDeBugGetUnreleasedSessions(sessionsGet, count);
+        REQUIRE(ret == HSUCCEED);
+
+        for (int i = 0; i < count; ++i) {
+            ret = HFReleaseInspireFaceSession(sessionsGet[i]);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        ret = HFDeBugGetUnreleasedSessionsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 0);
+    }
+}
+
+TEST_CASE("test_SystemStreamReleaseCase", "[system]") {
+    /**
+     * @brief Test the release of streams
+     * @details Test the release of streams and check the unreleased streams count
+     */
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    HResult ret;
+
+    SECTION("CreateStreams") {
+        /**
+         * @brief Create streams
+         * @details Create 10 streams and check the unreleased streams count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 0);
+
+        HInt32 createCount = 10;
+        HFImageStream streams[createCount];
+
+        for (int i = 0; i < createCount; ++i) {
+            HFImageStream imgHandle;
+            auto image = inspirecv::Image::Create(GET_DATA("data/bulk/pedestrian.png"));
+            ret = CVImageToImageStream(image, imgHandle);
+            REQUIRE(ret == HSUCCEED);
+            streams[i] = imgHandle;
+        }
+
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == createCount);
+
+        HFImageStream streamsGet[createCount];
+        ret = HFDeBugGetUnreleasedStreams(streamsGet, createCount);
+        REQUIRE(ret == HSUCCEED);
+        for (int i = 0; i < createCount; ++i) {
+            bool found = false;
+            for (int j = 0; j < createCount; ++j) {
+                if (streams[i] == streamsGet[j]) {
+                    found = true;
+                    break;
+                }
+            }
+            REQUIRE(found);
+        }
+    }
+
+    SECTION("ReleaseSomeStreams") {
+        /**
+         * @brief Release some streams
+         * @details Release some streams and check the unreleased streams count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 10);
+        auto createCount = count;
+
+        HFImageStream streamsGet[count];
+        ret = HFDeBugGetUnreleasedStreams(streamsGet, count);
+        REQUIRE(ret == HSUCCEED);
+
+        std::vector<int32_t> releaseIndex = {0, 2, 4, 6, 8};
+        for (int i = 0; i < releaseIndex.size(); ++i) {
+            ret = HFReleaseImageStream(streamsGet[releaseIndex[i]]);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == createCount - releaseIndex.size());
+
+        HFImageStream streamsGet2[count];
+        ret = HFDeBugGetUnreleasedStreams(streamsGet2, count);
+        REQUIRE(ret == HSUCCEED);
+        for (int i = 0; i < count; ++i) {
+            bool found = false;
+            for (int j = 0; j < releaseIndex.size(); ++j) {
+                if (streamsGet2[i] == streamsGet[releaseIndex[j]]) {
+                    found = true;
+                    break;
+                }
+            }
+            REQUIRE(!found);
+        }
+    }
+
+    SECTION("ReleaseAllStreams") {
+        /**
+         * @brief Release all streams
+         * @details Release all streams and check the unreleased streams count
+         */
+        HInt32 count;
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 5);
+
+        HFImageStream streamsGet[count];
+        ret = HFDeBugGetUnreleasedStreams(streamsGet, count);
+        REQUIRE(ret == HSUCCEED);
+
+        for (int i = 0; i < count; ++i) {
+            ret = HFReleaseImageStream(streamsGet[i]);
+            REQUIRE(ret == HSUCCEED);
+        }
+
+        ret = HFDeBugGetUnreleasedStreamsCount(&count);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(count == 0);
+    }
+}
diff --git a/cpp-package/inspireface/cpp/test/unit/base/test_face_session.cpp b/cpp-package/inspireface/cpp/test/unit/base/test_face_session.cpp
new file mode 100644
index 000000000..429771981
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/base/test_face_session.cpp
@@ -0,0 +1,83 @@
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "unit/test_helper/help.h"
+#include "middleware/costman.h"
+#include <inspireface/include/inspireface/launch.h>
+#include "frame_process.h"
+#include "inspireface/engine/face_session.h"
+#include <inspireface/include/inspireface/feature_hub_db.h>
+
+using namespace inspire;
+
+TEST_CASE("test_FaceSession", "[face_session") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    int32_t ret;
+    CustomPipelineParameter param;
+    param.enable_recognition = true;
+    param.enable_liveness = true;
+    param.enable_mask_detect = true;
+    param.enable_face_attribute = true;
+    param.enable_face_quality = true;
+
+    FaceSession session;
+    ret = session.Configuration(DetectModuleMode::DETECT_MODE_ALWAYS_DETECT, 1, param);
+    REQUIRE(ret == HSUCCEED);
+
+    inspirecv::Image kun1 = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+    inspirecv::Image kun2 = inspirecv::Image::Create(GET_DATA("data/bulk/jntm.jpg"));
+    inspirecv::FrameProcess proc1 = inspirecv::FrameProcess::Create(kun1.Data(), kun1.Height(), kun1.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+    inspirecv::FrameProcess proc2 = inspirecv::FrameProcess::Create(kun2.Data(), kun2.Height(), kun2.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+    std::vector<std::vector<float>> features;
+    std::vector<inspirecv::FrameProcess> processes = {proc1, proc2};
+    for (auto &process : processes) {
+        ret = session.FaceDetectAndTrack(process);
+        REQUIRE(ret == HSUCCEED);
+        if (session.GetDetectCache().size() > 0) {
+            FaceBasicData data = session.GetFaceBasicDataCache()[0];
+            ret = session.FaceFeatureExtract(process, data);
+            REQUIRE(ret == HSUCCEED);
+            const auto &faces = session.GetTrackingFaceList();
+            REQUIRE(faces.size() > 0);
+            Embedded feature;
+            FaceTrackWrap hyper_face_data = FaceObjectInternalToHyperFaceData(faces[0]);
+            float norm;
+            ret = session.FaceRecognitionModule()->FaceExtract(process, hyper_face_data, feature, norm);
+            REQUIRE(ret == HSUCCEED);
+            features.push_back(feature);
+        }
+    }
+    REQUIRE(features.size() == 2);
+    float res;
+    ret = FeatureHubDB::CosineSimilarity(features[0].data(), features[1].data(), features[0].size(), res);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(res > 0.5f);
+
+    inspirecv::Image other = inspirecv::Image::Create(GET_DATA("data/bulk/woman.png"));
+    inspirecv::FrameProcess proc3 =
+      inspirecv::FrameProcess::Create(other.Data(), other.Height(), other.Width(), inspirecv::BGR, inspirecv::ROTATION_0);
+    ret = session.FaceDetectAndTrack(proc3);
+    REQUIRE(ret == HSUCCEED);
+    if (session.GetDetectCache().size() > 0) {
+        FaceBasicData data = session.GetFaceBasicDataCache()[0];
+        ret = session.FaceFeatureExtract(proc3, data);
+        auto faces = session.GetTrackingFaceList();
+        REQUIRE(ret == HSUCCEED);
+        Embedded feature;
+        FaceTrackWrap hyper_face_data = FaceObjectInternalToHyperFaceData(faces[0]);
+        float norm;
+        ret = session.FaceRecognitionModule()->FaceExtract(proc3, hyper_face_data, feature, norm);
+        REQUIRE(ret == HSUCCEED);
+        features.push_back(feature);
+    }
+    REQUIRE(features.size() == 3);
+    float other_v_kun1, other_v_kun2;
+    ret = FeatureHubDB::CosineSimilarity(features[0].data(), features[2].data(), features[0].size(), other_v_kun1);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(other_v_kun1 < 0.5f);
+    ret = FeatureHubDB::CosineSimilarity(features[1].data(), features[2].data(), features[0].size(), other_v_kun2);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(other_v_kun2 < 0.5f);
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/base/test_module_feature_hub.cpp b/cpp-package/inspireface/cpp/test/unit/base/test_module_feature_hub.cpp
new file mode 100644
index 000000000..d5cb82562
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/base/test_module_feature_hub.cpp
@@ -0,0 +1,198 @@
+#include <iostream>
+#include "settings/test_settings.h"
+#include "inspireface/c_api/inspireface.h"
+#include "unit/test_helper/help.h"
+#include <inspireface/include/inspireface/feature_hub_db.h>
+#include "middleware/costman.h"
+
+using namespace inspire;
+
+TEST_CASE("test_FeatureHubBasic", "[feature_hub") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    // Enable feature hub
+    DatabaseConfiguration config;
+    config.primary_key_mode = PrimaryKeyMode::AUTO_INCREMENT;
+    config.enable_persistence = false;  // memory mode
+    int32_t ret;
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+
+    // Check if feature hub is enabled
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+
+    // Check number of features
+    int32_t count = 1000;
+    std::vector<int64_t> ids;
+    std::vector<int64_t> expected_ids;
+    for (int32_t i = 0; i < count; i++) {
+        auto vec = GenerateRandomFeature(512, false);
+        int64_t alloc_id;
+        ret = INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(vec, -1, alloc_id);
+        REQUIRE(ret == HSUCCEED);
+        ids.push_back(alloc_id);
+        expected_ids.push_back(i + 1);
+    }
+    REQUIRE(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == ids.size());
+    REQUIRE(ids == expected_ids);
+
+    // Delete data
+    std::vector<int64_t> delete_ids = {5, 20, 100};
+    for (auto id : delete_ids) {
+        INSPIREFACE_FEATURE_HUB->FaceFeatureRemove(id);
+    }
+    REQUIRE(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == ids.size() - delete_ids.size());
+
+    // Check if the deleted data can be found
+    std::vector<float> feature;
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(5, feature);
+    REQUIRE(ret == HERR_FT_HUB_NOT_FOUND_FEATURE);
+
+    // Check if the data can be found
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1, feature);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(feature.size() == 512);
+
+    // Check if the cached data is correct
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1);
+    REQUIRE(ret == HSUCCEED);
+    auto cached_feature = INSPIREFACE_FEATURE_HUB->GetFaceFeaturePtrCache();
+    for (size_t i = 0; i < cached_feature->dataSize; i++) {
+        REQUIRE(feature[i] == cached_feature->data[i]);
+    }
+
+    // Update data
+    auto update_feature = GenerateRandomFeature(512, false);
+    ret = INSPIREFACE_FEATURE_HUB->FaceFeatureUpdate(update_feature, 1);
+    REQUIRE(ret == HSUCCEED);
+
+    // Check if the updated data is correct
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1, feature);
+    REQUIRE(ret == HSUCCEED);
+    for (size_t i = 0; i < feature.size(); i++) {
+        REQUIRE(feature[i] == Approx(update_feature[i]).epsilon(0.0001));
+    }
+
+    // Update removed data
+    ret = INSPIREFACE_FEATURE_HUB->FaceFeatureUpdate(update_feature, 5);
+    REQUIRE(ret == HERR_FT_HUB_NOT_FOUND_FEATURE);
+
+    // Disable feature hub
+    INSPIREFACE_FEATURE_HUB->DisableHub();
+    REQUIRE(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == 0);
+
+    // Check if the data can be found
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1, feature);
+    REQUIRE(ret == HERR_FT_HUB_DISABLE);
+
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+    // Because the memory mode is turned on, once the data is turned off, it goes back to empty
+    REQUIRE(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == 0);
+
+    ret = INSPIREFACE_FEATURE_HUB->DisableHub();
+    REQUIRE(ret == HSUCCEED);
+}
+
+TEST_CASE("test_PerformanceMemoryMode", "[feature_hub") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    DatabaseConfiguration config;
+    config.primary_key_mode = PrimaryKeyMode::AUTO_INCREMENT;
+    config.enable_persistence = false;  // memory mode
+    int32_t ret;
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+
+    Timer t1;
+    int num = 10000;
+    for (int i = 0; i < num; i++) {
+        auto vec = GenerateRandomFeature(512, false);
+        int64_t alloc_id;
+        ret = INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(vec, -1, alloc_id);
+        REQUIRE(ret == HSUCCEED);
+    }
+    TEST_PRINT("[Memory Mode]Insert 10000 features cost: {:.2f} ms", t1.GetCostTime());
+
+    Timer t2;
+    std::vector<float> feature;
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1, feature);
+    TEST_PRINT("[Memory Mode]Get feature from id cost: {:.2f} ms", t2.GetCostTime());
+    REQUIRE(ret == HSUCCEED);
+
+    Timer t3;
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(9998, feature);
+    TEST_PRINT("[Memory Mode]Get feature from id cost: {:.2f} ms", t3.GetCostTime());
+    REQUIRE(ret == HSUCCEED);
+    auto sim_vec = SimulateSimilarVector(feature, false);
+    FaceSearchResult search_result;
+    Timer t4;
+    INSPIREFACE_FEATURE_HUB->SearchFaceFeature(sim_vec, search_result, true);
+    TEST_PRINT("[Memory Mode]Search feature cost: {:.2f} ms", t4.GetCostTime());
+    REQUIRE(search_result.id == 9998);
+
+    ret = INSPIREFACE_FEATURE_HUB->FaceFeatureRemove(9998);
+    REQUIRE(ret == HSUCCEED);
+
+    INSPIREFACE_FEATURE_HUB->DisableHub();
+}
+
+TEST_CASE("test_PerformancePersistentMode", "[feature_hub") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    std::string db_path = ".test_db";
+    std::remove(db_path.c_str());
+
+    DatabaseConfiguration config;
+    config.primary_key_mode = PrimaryKeyMode::AUTO_INCREMENT;
+    config.enable_persistence = true;  // persistent mode
+    config.persistence_db_path = db_path;
+
+    int32_t ret;
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+
+    Timer t1;
+    int num = 10000;
+    for (int i = 0; i < num; i++) {
+        auto vec = GenerateRandomFeature(512, false);
+        int64_t alloc_id;
+        ret = INSPIREFACE_FEATURE_HUB->FaceFeatureInsert(vec, -1, alloc_id);
+        REQUIRE(ret == HSUCCEED);
+    }
+    TEST_PRINT("[Persistent Mode]Insert 10000 features cost: {:.2f} ms", t1.GetCostTime());
+
+    Timer t2;
+    std::vector<float> feature;
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(1, feature);
+    TEST_PRINT("[Persistent Mode]Get feature from id cost: {:.2f} ms", t2.GetCostTime());
+    REQUIRE(ret == HSUCCEED);
+
+    Timer t3;
+    ret = INSPIREFACE_FEATURE_HUB->GetFaceFeature(9998, feature);
+    TEST_PRINT("[Persistent Mode]Get feature from id cost: {:.2f} ms", t3.GetCostTime());
+    REQUIRE(ret == HSUCCEED);
+    auto sim_vec = SimulateSimilarVector(feature, false);
+    FaceSearchResult search_result;
+    Timer t4;
+    INSPIREFACE_FEATURE_HUB->SearchFaceFeature(sim_vec, search_result, true);
+    TEST_PRINT("[Persistent Mode]Search feature cost: {:.2f} ms", t4.GetCostTime());
+    REQUIRE(search_result.id == 9998);
+
+    ret = INSPIREFACE_FEATURE_HUB->FaceFeatureRemove(9998);
+    REQUIRE(ret == HSUCCEED);
+
+    auto remark_num = INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount();
+    REQUIRE(remark_num == num - 1);
+
+    // Verify important of persistence test
+    ret = INSPIREFACE_FEATURE_HUB->EnableHub(config);
+    REQUIRE(ret == HSUCCEED);
+    REQUIRE(INSPIREFACE_FEATURE_HUB->GetFaceFeatureCount() == remark_num);
+
+    INSPIREFACE_FEATURE_HUB->DisableHub();
+}
diff --git a/RetinaFace/rcnn/PY_OP/__init__.py b/cpp-package/inspireface/cpp/test/unit/base/test_module_pipeline.cpp
old mode 100755
new mode 100644
similarity index 100%
rename from RetinaFace/rcnn/PY_OP/__init__.py
rename to cpp-package/inspireface/cpp/test/unit/base/test_module_pipeline.cpp
diff --git a/RetinaFace/rcnn/__init__.py b/cpp-package/inspireface/cpp/test/unit/base/test_module_recognition.cpp
similarity index 100%
rename from RetinaFace/rcnn/__init__.py
rename to cpp-package/inspireface/cpp/test/unit/base/test_module_recognition.cpp
diff --git a/cpp-package/inspireface/cpp/test/unit/base/test_module_track.cpp b/cpp-package/inspireface/cpp/test/unit/base/test_module_track.cpp
new file mode 100644
index 000000000..db4bb8cb3
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/base/test_module_track.cpp
@@ -0,0 +1,111 @@
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "unit/test_helper/help.h"
+#include <inspireface/include/inspireface/feature_hub_db.h>
+#include "middleware/costman.h"
+#include "track_module/face_detect/all.h"
+#include "track_module/face_track_module.h"
+#include <inspireface/include/inspireface/frame_process.h>
+
+using namespace inspire;
+
+TEST_CASE("test_FaceDetect", "[track_module") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    const std::vector<int32_t> supported_sizes = {160, 320, 640};
+    const std::vector<std::string> scheme_names = {"face_detect_160", "face_detect_320", "face_detect_640"};
+    for (size_t i = 0; i < scheme_names.size(); i++) {
+        InspireModel model;
+        auto ret = archive.LoadModel(scheme_names[i], model);
+        REQUIRE(ret == 0);
+        FaceDetectAdapt face_detector(supported_sizes[i]);
+        face_detector.LoadData(model, model.modelType, false);
+
+        inspirecv::Image img = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        auto result = face_detector(img);
+        REQUIRE(result.size() == 1);
+    }
+}
+
+TEST_CASE("test_RefineNet", "[track_module") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    InspireModel model;
+    auto ret = archive.LoadModel("refine_net", model);
+    REQUIRE(ret == 0);
+    RNetAdapt rnet;
+    rnet.LoadData(model, model.modelType, false);
+
+    inspirecv::Image face = inspirecv::Image::Create(GET_DATA("data/crop/crop.png"));
+    auto result1 = rnet(face);
+    REQUIRE(result1 > 0.5f);
+
+    inspirecv::Image no_face = inspirecv::Image::Create(GET_DATA("data/crop/no_face.png"));
+    auto result2 = rnet(no_face);
+    REQUIRE(result2 < 0.5f);
+}
+
+TEST_CASE("test_Landmark", "[track_module") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    InspireModel model;
+    auto ret = archive.LoadModel("landmark", model);
+    REQUIRE(ret == 0);
+
+    FaceLandmarkAdapt face_landmark(112);
+    face_landmark.LoadData(model, model.modelType);
+
+    inspirecv::Image img = inspirecv::Image::Create(GET_DATA("data/crop/crop.png"));
+    auto result = face_landmark(img);
+    REQUIRE(result.size() == 106 * 2);
+}
+
+TEST_CASE("test_Quality", "[track_module") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+    InspireModel model;
+    auto ret = archive.LoadModel("pose_quality", model);
+    REQUIRE(ret == 0);
+    FacePoseQualityAdapt quality;
+    ret = quality.LoadData(model, model.modelType);
+    REQUIRE(ret == 0);
+
+    inspirecv::Image img = inspirecv::Image::Create(GET_DATA("data/crop/crop.png"));
+    auto result = quality(img);
+    REQUIRE(result.lmk.size() == 5);
+    REQUIRE(result.lmk_quality.size() == 5);
+}
+
+TEST_CASE("test_FaceTrackModule", "[track_module") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+    auto archive = INSPIREFACE_CONTEXT->getMArchive();
+
+    SECTION("Test face detect rotate 0") {
+        auto mode = DetectModuleMode::DETECT_MODE_ALWAYS_DETECT;
+        int max_detected_faces = 10;
+        FaceTrackModule face_track(mode, max_detected_faces);
+        face_track.Configuration(archive);
+        inspirecv::Image img = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        inspirecv::FrameProcess image = inspirecv::FrameProcess::Create(img.Data(), img.Height(), img.Width(), inspirecv::BGR);
+        face_track.UpdateStream(image);
+        REQUIRE(face_track.trackingFace.size() == 1);
+    }
+
+    SECTION("Test face detect rotate 90") {
+        auto mode = DetectModuleMode::DETECT_MODE_ALWAYS_DETECT;
+        int max_detected_faces = 10;
+        FaceTrackModule face_track(mode, max_detected_faces);
+        face_track.Configuration(archive);
+        inspirecv::Image img = inspirecv::Image::Create(GET_DATA("data/bulk/r90.jpg"));
+        inspirecv::FrameProcess image =
+          inspirecv::FrameProcess::Create(img.Data(), img.Height(), img.Width(), inspirecv::BGR, inspirecv::ROTATION_90);
+        face_track.UpdateStream(image);
+        REQUIRE(face_track.trackingFace.size() == 1);
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/cpp_api/test_session_face_track.cpp b/cpp-package/inspireface/cpp/test/unit/cpp_api/test_session_face_track.cpp
new file mode 100644
index 000000000..496fbc49c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/cpp_api/test_session_face_track.cpp
@@ -0,0 +1,118 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2025-03-29
+ */
+
+#include <iostream>
+#include "settings/test_settings.h"
+#include "unit/test_helper/simple_csv_writer.h"
+#include "unit/test_helper/test_help.h"
+#include "unit/test_helper/test_tools.h"
+#include <inspireface/include/inspireface/inspireface.hpp>
+
+using namespace inspire;
+
+TEST_CASE("test_SessionFaceTrack", "[session_face_track]") {
+    DRAW_SPLIT_LINE
+    TEST_PRINT_OUTPUT(true);
+
+    SECTION("Face detection from image") {
+        CustomPipelineParameter param;
+        int32_t ret;
+
+        std::shared_ptr<Session> session = std::shared_ptr<Session>(Session::CreatePtr(DetectModuleMode::DETECT_MODE_ALWAYS_DETECT, 3, param));
+        REQUIRE(session != nullptr);
+
+        auto image = inspirecv::Image::Create(GET_DATA("data/bulk/kun.jpg"));
+        auto process = inspirecv::FrameProcess::Create(image, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        std::vector<FaceTrackWrap> results;
+        ret = session->FaceDetectAndTrack(process, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+    }
+
+    SECTION("Head pose estimation") {
+        CustomPipelineParameter param;
+        int32_t ret;
+
+        std::shared_ptr<Session> session = std::shared_ptr<Session>(Session::CreatePtr(DetectModuleMode::DETECT_MODE_ALWAYS_DETECT, 3, param));
+        REQUIRE(session != nullptr);
+
+        HFMultipleFaceData multipleFaceData = {0};
+        std::vector<FaceTrackWrap> results;
+
+        // Left side face
+        HFImageStream leftHandle;
+        auto left = inspirecv::Image::Create(GET_DATA("data/pose/left_face.jpeg"));
+        auto leftProcess = inspirecv::FrameProcess::Create(left, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(leftProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        HFloat yaw, pitch, roll;
+        bool checked;
+
+        // Left-handed rotation
+        yaw = results[0].face3DAngle.yaw;
+        checked = (yaw > -90 && yaw < -10);
+        CHECK(checked);
+        
+        // Right-handed rotation
+        auto right = inspirecv::Image::Create(GET_DATA("data/pose/right_face.png"));
+        auto rightProcess = inspirecv::FrameProcess::Create(right, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(rightProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        yaw = results[0].face3DAngle.yaw;
+        checked = (yaw > 10 && yaw < 90);
+        CHECK(checked);
+
+        // Rise head
+        auto rise = inspirecv::Image::Create(GET_DATA("data/pose/rise_face.jpeg"));
+        auto riseProcess = inspirecv::FrameProcess::Create(rise, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(riseProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        pitch = results[0].face3DAngle.pitch;
+        CHECK(pitch > 3);
+
+        // Lower head
+        auto lower = inspirecv::Image::Create(GET_DATA("data/pose/lower_face.jpeg"));
+        auto lowerProcess = inspirecv::FrameProcess::Create(lower, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(lowerProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        pitch = results[0].face3DAngle.pitch;
+        CHECK(pitch < -10);
+
+        // Roll head
+        auto leftWryneck = inspirecv::Image::Create(GET_DATA("data/pose/left_wryneck.png"));
+        auto leftWryneckProcess = inspirecv::FrameProcess::Create(leftWryneck, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(leftWryneckProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        roll = results[0].face3DAngle.roll;
+        CHECK(roll < -30);
+
+        // Roll head
+        auto rightWryneck = inspirecv::Image::Create(GET_DATA("data/pose/right_wryneck.png"));
+        auto rightWryneckProcess = inspirecv::FrameProcess::Create(rightWryneck, inspirecv::DATA_FORMAT::BGR, inspirecv::ROTATION_MODE::ROTATION_0);
+
+        ret = session->FaceDetectAndTrack(rightWryneckProcess, results);
+        REQUIRE(ret == HSUCCEED);
+        REQUIRE(results.size() == 1);
+
+        roll = results[0].face3DAngle.roll;
+        CHECK(roll > 25);
+    }
+}
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/test_helper/help.h b/cpp-package/inspireface/cpp/test/unit/test_helper/help.h
new file mode 100644
index 000000000..50df97310
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/test_helper/help.h
@@ -0,0 +1,3 @@
+#include "test_tools.h"
+#include "test_help.h"
+#include "simple_csv_writer.h"
\ No newline at end of file
diff --git a/cpp-package/inspireface/cpp/test/unit/test_helper/simple_csv_writer.h b/cpp-package/inspireface/cpp/test/unit/test_helper/simple_csv_writer.h
new file mode 100644
index 000000000..366b1fb79
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/test_helper/simple_csv_writer.h
@@ -0,0 +1,100 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_SIMPLE_CSV_WRITER_H
+#define INSPIREFACE_SIMPLE_CSV_WRITER_H
+
+#include <fstream>
+#include <vector>
+#include <string>
+#include <iostream>
+#include <iomanip>  // Used to set the output format
+
+class SimpleCSVWriter {
+public:
+    SimpleCSVWriter(const std::string& filepath) {
+        // Check whether the file exists
+        std::ifstream file(filepath);
+        if (!file.good()) {
+            // The file does not exist. Create a new csv file
+            std::ofstream outfile(filepath);
+            if (!outfile.is_open()) {
+                std::cerr << "Failed to create file: " << filepath << std::endl;
+            }
+            outfile.close();
+        }
+        // Save the file path for later use
+        this->filepath = filepath;
+    }
+
+    virtual ~SimpleCSVWriter() {}  // Add a virtual destructor to ensure correct destructor behavior
+
+protected:
+    std::string filepath;
+
+    void insertData(const std::vector<std::string>& data) {
+        std::ofstream file(this->filepath, std::ios_base::app);  // Open the file in append mode
+        if (!file.is_open()) {
+            std::cerr << "Failed to open file: " << this->filepath << std::endl;
+            return;
+        }
+        for (const auto& datum : data) {
+            file << datum;
+            if (&datum != &data.back()) {  // If it is not the last element, add a comma separation
+                file << ",";
+            }
+        }
+        file << "\n";  // Add a newline character after each inserted row of data
+        file.close();
+    }
+};
+
+class BenchmarkRecord : public SimpleCSVWriter {
+public:
+    BenchmarkRecord(const std::string& filepath, const std::string& name = "Benchmark") : SimpleCSVWriter(filepath) {
+        std::ifstream file(this->filepath);
+        if (file.peek() == std::ifstream::traits_type::eof()) {  // If the file is empty, insert header data
+            std::vector<std::string> header = {name, "Loops", "Total Time(ms)", "Average Time(ms)"};
+            SimpleCSVWriter::insertData(header);
+        }
+    }
+
+    void insertBenchmarkData(const std::string& caseName, int loops, double totalCost, double avgCost) {
+        std::ofstream file(this->filepath, std::ios_base::app);
+        if (!file.is_open()) {
+            std::cerr << "Failed to open file: " << this->filepath << std::endl;
+            return;
+        }
+        // Format output
+        file << std::fixed << std::setprecision(5);
+        file << caseName << "," << loops << "," << totalCost << "," << avgCost << "\n";
+        file.close();
+    }
+};
+
+class EvaluationRecord : public SimpleCSVWriter {
+public:
+    EvaluationRecord(const std::string& filepath) : SimpleCSVWriter(filepath) {
+        std::ifstream file(this->filepath);
+        if (file.peek() == std::ifstream::traits_type::eof()) {  // If the file is empty, insert header data
+            std::vector<std::string> header = {"Resource Version", "Dataset", "Accuracy", "Best Threshold"};
+            SimpleCSVWriter::insertData(header);
+        }
+    }
+
+    void insertEvaluationData(const std::string& modelName, const std::string& dataset, double accuracy, double bestThreshold) {
+        std::ofstream file(this->filepath, std::ios_base::app);
+        if (!file.is_open()) {
+            std::cerr << "Failed to open file: " << this->filepath << std::endl;
+            return;
+        }
+        // Format output
+        file << std::fixed << std::setprecision(5);
+        file << modelName << "," << dataset << "," << accuracy << "," << bestThreshold << "\n";
+        file.close();
+    }
+};
+
+#endif  // INSPIREFACE_SIMPLE_CSV_WRITER_H
diff --git a/cpp-package/inspireface/cpp/test/unit/test_helper/test_help.h b/cpp-package/inspireface/cpp/test/unit/test_helper/test_help.h
new file mode 100644
index 000000000..df9f62b2c
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/test_helper/test_help.h
@@ -0,0 +1,508 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+
+#ifndef INSPIREFACE_TEST_TEST_HELP_H
+#define INSPIREFACE_TEST_TEST_HELP_H
+
+#include <iostream>
+#include <vector>
+#include <string>
+#include <fstream>
+#include <cassert>
+#include <sstream>
+#include <iomanip>
+#include <indicators/block_progress_bar.hpp>
+#include <indicators/cursor_control.hpp>
+#include "inspireface/c_api/inspireface.h"
+#include <inspirecv/inspirecv.h>
+#include <iomanip>
+#include "test_tools.h"
+#include <random>
+#include <fstream>
+#ifdef _WIN32
+#include <windows.h>
+#include <psapi.h>
+#else
+#include <sys/resource.h>
+#endif
+
+using namespace indicators;
+
+typedef std::vector<std::pair<std::string, std::string>> FaceImageDataList;
+
+inline void Split(const std::string& src, std::vector<std::string>& res, const std::string& pattern, size_t maxsplit = std::string::npos) {
+    res.clear();
+    size_t Start = 0;
+    size_t end = 0;
+    std::string sub;
+    while (Start < src.size()) {
+        end = src.find_first_of(pattern, Start);
+        if (std::string::npos == end || res.size() >= maxsplit) {
+            sub = src.substr(Start);
+            res.push_back(sub);
+            return;
+        }
+        sub = src.substr(Start, end - Start);
+        res.push_back(sub);
+        Start = end + 1;
+    }
+    return;
+}
+
+inline std::vector<std::string> Split(const std::string& src, const std::string& pattern, size_t maxsplit = std::string::npos) {
+    std::vector<std::string> res;
+    Split(src, res, pattern, maxsplit);
+    return res;
+}
+
+inline bool EndsWith(const std::string& str, const std::string& suffix) {
+    if (suffix.length() > str.length()) {
+        return false;
+    }
+    return 0 == str.compare(str.length() - suffix.length(), suffix.length(), suffix);
+}
+
+inline std::string PathJoin(const std::string& path1, const std::string& path2) {
+    if (EndsWith(path1, "/")) {
+        return path1 + path2;
+    }
+    return path1 + "/" + path2;
+}
+
+inline FaceImageDataList LoadLFWFunneledValidData(const std::string& dir, const std::string& txtPath) {
+    FaceImageDataList list;
+    std::ifstream file(txtPath);
+    std::string line;
+
+    while (std::getline(file, line)) {
+        std::vector<std::string> parts;
+        Split(line, parts, "/");
+        if (parts.size() >= 2) {
+            std::string name = parts[0];
+            std::string fullPath = dir + "/" + line;
+
+            list.push_back({name, fullPath});
+        }
+    }
+
+    return list;
+}
+
+inline bool ImportLFWFunneledValidData(HFSession handle, FaceImageDataList& data, size_t importNum) {
+    auto dataSize = data.size();
+    std::string title = "Import " + std::to_string(importNum) + " face data...";
+    // Hide cursor
+    show_console_cursor(false);
+    BlockProgressBar bar{option::BarWidth{60},
+                         option::Start{"["},
+                         option::End{"]"},
+                         option::PostfixText{title},
+                         option::ForegroundColor{Color::white},
+                         option::FontStyles{std::vector<FontStyle>{FontStyle::bold}}};
+
+    auto progress = 0.0f;
+    for (size_t i = 0; i < importNum; ++i) {
+        bar.set_progress(progress);
+        size_t index = i % dataSize;
+        // Data processing
+        auto item = data[index];
+        inspirecv::Image image = inspirecv::Image::Create(item.second);
+        HFImageData imageData = {0};
+        imageData.data = (uint8_t*)image.Data();
+        imageData.height = image.Height();
+        imageData.width = image.Width();
+        imageData.format = HF_STREAM_BGR;
+        imageData.rotation = HF_CAMERA_ROTATION_0;
+        HFImageStream imgHandle;
+        auto ret = HFCreateImageStream(&imageData, &imgHandle);
+        if (ret != HSUCCEED || image.Empty()) {
+            std::cerr << "Error image: " << std::to_string(ret) << " , " << item.second << std::endl;
+            return false;
+        }
+        // Face tracked
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(handle, imgHandle, &multipleFaceData);
+
+        if (ret != HSUCCEED) {
+            std::cerr << "Error Track: " << std::to_string(ret) << " , " << item.second << std::endl;
+            return false;
+        }
+
+        if (multipleFaceData.detectedNum == 0) {
+            std::cerr << "Not Detected face: " << item.second << std::endl;
+            return false;
+        }
+
+        // Extract face feature
+        HFFaceFeature feature = {0};
+        ret = HFFaceFeatureExtract(handle, imgHandle, multipleFaceData.tokens[0], &feature);
+        if (ret != HSUCCEED) {
+            std::cerr << "Error extract: " << std::to_string(ret) << " , " << item.second << std::endl;
+            return false;
+        }
+        char* newTagName = new char[item.first.size() + 1];
+        std::strcpy(newTagName, item.first.c_str());
+        HFFaceFeatureIdentity identity = {0};
+        identity.id = i;
+        // identity.tag = newTagName;
+        identity.feature = &feature;
+        HFaceId allocId;
+        ret = HFFeatureHubInsertFeature(identity, &allocId);
+        if (ret != HSUCCEED) {
+            std::cerr << "Error insert feature: " << std::to_string(ret) << " , " << item.second << std::endl;
+            return false;
+        }
+
+        delete[] newTagName;
+        HFReleaseImageStream(imgHandle);
+        // Update progress
+        progress = 100.0f * (float)(i + 1) / importNum;
+    }
+    bar.set_progress(100.0f);
+    // Show cursor
+    show_console_cursor(true);
+    std::cout << "\033[0m\n";  // ANSI resets the color code
+
+    return true;
+}
+
+inline double CalculateOverlap(const HFaceRect& box1, const HFaceRect& box2) {
+    // Calculate the coordinates of the intersection rectangle
+    HInt32 x_overlap = std::max(0, std::min(box1.x + box1.width, box2.x + box2.width) - std::max(box1.x, box2.x));
+    HInt32 y_overlap = std::max(0, std::min(box1.y + box1.height, box2.y + box2.height) - std::max(box1.y, box2.y));
+
+    // Calculate the area of the intersection
+    HInt32 overlap_area = x_overlap * y_overlap;
+
+    // Calculate the area of each rectangle
+    HInt32 box1_area = box1.width * box1.height;
+    HInt32 box2_area = box2.width * box2.height;
+
+    // Calculate the total area
+    HInt32 total_area = box1_area + box2_area - overlap_area;
+
+    // Calculate the overlap ratio
+    double overlap_ratio = total_area > 0 ? static_cast<double>(overlap_area) / total_area : 0;
+
+    return overlap_ratio;
+}
+
+inline std::vector<std::string> generateFilenames(const std::string& templateStr, int start, int end) {
+    std::vector<std::string> filenames;
+    for (int i = start; i <= end; ++i) {
+        std::ostringstream oss;
+        oss << "frame-" << std::setw(4) << std::setfill('0') << i << ".jpg";
+        filenames.push_back(oss.str());
+    }
+    return filenames;
+}
+
+inline bool FindMostSimilarScoreFromTwoPic(HFSession handle, const std::string& img1, const std::string& img2, float& mostSimilar) {
+    mostSimilar = -1.0f;
+    std::vector<std::vector<std::vector<float>>> features(2);
+    std::vector<std::string> images = {img1, img2};
+    for (int i = 0; i < 2; ++i) {
+        HFImageStream img;
+        //        auto ret = ReadImageToImageStream(images[i].c_str(), img);
+        auto cvMat = inspirecv::Image::Create(images[i]);
+        auto ret = CVImageToImageStream(cvMat, img);
+        if (ret != 0) {
+            std::cerr << "Image is not found: " << ret << std::endl;
+            return false;
+        }
+        HFMultipleFaceData multipleFaceData = {0};
+        ret = HFExecuteFaceTrack(handle, img, &multipleFaceData);
+        if (ret != 0) {
+            std::cerr << "Error track: " << ret << std::endl;
+            HFReleaseImageStream(img);
+            return false;
+        }
+        HInt32 featureNum;
+        HFGetFeatureLength(&featureNum);
+        for (int j = 0; j < multipleFaceData.detectedNum; ++j) {
+            std::vector<float> feature(featureNum, 0.0f);
+            ret = HFFaceFeatureExtractCpy(handle, img, multipleFaceData.tokens[j], feature.data());
+            if (ret != 0) {
+                std::cerr << "Error extract: " << ret << std::endl;
+                HFReleaseImageStream(img);
+                return false;
+            }
+            features[i].push_back(feature);
+            float qu;
+            HFFaceQualityDetect(handle, multipleFaceData.tokens[j], &qu);
+            std::ofstream file("tem.csv", std::ios::app);
+
+            if (file.is_open()) {
+                file << qu << ",\n";
+                file.close();
+            } else {
+                std::cerr << "Failed to open file tem.csv" << std::endl;
+            }
+        }
+        HFReleaseImageStream(img);
+    }
+
+    if (features[0].empty() || features[1].empty()) {
+        //        std::cerr << "Not detected " << std::endl;
+        return false;
+    }
+    auto& features1 = features[0];
+    auto& features2 = features[1];
+    for (auto& feat1 : features1) {
+        for (auto& feat2 : features2) {
+            float comp;
+            HFFaceFeature faceFeature1 = {0};
+            faceFeature1.size = feat1.size();
+            faceFeature1.data = feat1.data();
+            HFFaceFeature faceFeature2 = {0};
+            faceFeature2.size = feat2.size();
+            faceFeature2.data = feat2.data();
+
+            HFFaceComparison(faceFeature1, faceFeature2, &comp);
+            if (comp > mostSimilar) {
+                mostSimilar = comp;
+            }
+        }
+    }
+
+    return true;
+}
+
+inline std::vector<std::vector<std::string>> ReadPairs(const std::string& pairs_filename) {
+    std::vector<std::vector<std::string>> pairs;
+    std::ifstream file(pairs_filename);  // Open the file
+    std::string line;
+
+    if (!file.is_open()) {
+        std::cerr << "Unable to open file: " << pairs_filename << std::endl;
+        return pairs;  // If the file cannot be opened, an empty list is returned
+    }
+
+    std::getline(file, line);  // Skip the first line
+    while (std::getline(file, line)) {
+        std::istringstream iss(line);
+        std::vector<std::string> pair;
+        std::string element;
+
+        while (iss >> element) {
+            pair.push_back(element);
+        }
+        if (!pair.empty()) {
+            pairs.push_back(pair);
+        }
+    }
+
+    return pairs;
+}
+
+inline std::string zfill(int number, int width) {
+    std::ostringstream oss;
+    // Set padding to '0' and define the string width
+    oss << std::setfill('0') << std::setw(width) << number;
+    return oss.str();
+}
+
+inline std::pair<float, float> FindBestThreshold(const std::vector<float>& similarities, const std::vector<int>& labels) {
+    std::vector<float> thresholds;
+    for (float i = 0.0f; i < 1.0f; i += 0.01f) {
+        thresholds.push_back(i);
+    }
+
+    float best_threshold = 0.0f;
+    float best_accuracy = 0.0f;
+
+    for (auto& threshold : thresholds) {
+        std::vector<int> predictions;
+        for (auto& similarity : similarities) {
+            predictions.push_back(similarity > threshold ? 1 : 0);
+        }
+
+        int correct = 0;
+        for (size_t i = 0; i < labels.size(); ++i) {
+            if (predictions[i] == labels[i]) {
+                ++correct;
+            }
+        }
+
+        float accuracy = static_cast<float>(correct) / static_cast<float>(labels.size());
+
+        if (accuracy > best_accuracy) {
+            best_accuracy = accuracy;
+            best_threshold = threshold;
+        }
+    }
+
+    return {best_threshold, best_accuracy};
+}
+
+/** Generate random eigenvectors of the specified length */
+inline std::vector<float> GenerateRandomFeature(size_t length, bool normalize = true) {
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_real_distribution<> dis(-1.0, 1.0);
+
+    std::vector<float> featureVector(length);
+    float norm = 0.0;
+
+    for (float& value : featureVector) {
+        value = dis(gen);
+        norm += value * value;
+    }
+
+    norm = std::sqrt(norm);
+
+    if (norm > 0 && normalize) {
+        for (float& value : featureVector) {
+            value /= norm;
+        }
+    }
+
+    return featureVector;
+}
+
+inline std::vector<float> SimulateSimilarVector(const std::vector<float>& original, bool normalize = true) {
+    std::vector<float> similar(original.size());
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::normal_distribution<float> dis(0.0, 0.02);
+
+    for (size_t i = 0; i < original.size(); ++i) {
+        similar[i] = original[i] + dis(gen);
+    }
+
+    float norm = 0.0f;
+    for (auto& value : similar) {
+        norm += value * value;
+    }
+    norm = std::sqrt(norm);
+    if (norm > 0 && normalize) {
+        for (auto& value : similar) {
+            value /= norm;
+        }
+    }
+
+    return similar;
+}
+
+inline std::vector<int> GenerateRandomNumbers(int n, int min, int max) {
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_int_distribution<> distrib(min, max);
+
+    std::vector<int> numbers;
+    numbers.reserve(n);
+
+    for (int i = 0; i < n; ++i) {
+        numbers.push_back(distrib(gen));
+    }
+
+    return numbers;
+}
+
+inline std::string ReplaceFileExtension(const std::string& filePath, const std::string& newExtension) {
+    size_t lastDotPos = filePath.find_last_of(".");
+    if (lastDotPos == std::string::npos) {
+        // If the 'dot' is not found, return to the original path
+        return filePath;
+    }
+
+    return filePath.substr(0, lastDotPos) + newExtension;
+}
+
+inline bool CompareTwoFaces(HFSession handle, const inspirecv::Image& img1, const inspirecv::Image& img2, float& similarity) {
+    HFImageStream img;
+    auto ret = CVImageToImageStream(img1, img);
+    if (ret != 0) {
+        std::cerr << "Image is not found: " << ret << std::endl;
+        return false;
+    }
+    HFMultipleFaceData multipleFaceData = {0};
+    ret = HFExecuteFaceTrack(handle, img, &multipleFaceData);
+    if (ret != 0) {
+        std::cerr << "Error track: " << ret << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+
+    if (multipleFaceData.detectedNum == 0) {
+        std::cerr << "No face detected in first image" << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+
+    // Extract features from first image
+    HInt32 featureLength;
+    HFGetFeatureLength(&featureLength);
+    std::vector<float> feature1(featureLength);
+    ret = HFFaceFeatureExtractCpy(handle, img, multipleFaceData.tokens[0], feature1.data());
+    if (ret != HSUCCEED) {
+        std::cerr << "Feature extraction failed for first image" << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+    HFReleaseImageStream(img);
+
+    // Process second image
+    ret = CVImageToImageStream(img2, img);
+    if (ret != HSUCCEED) {
+        std::cerr << "Second image not found" << std::endl;
+        return false;
+    }
+
+    ret = HFExecuteFaceTrack(handle, img, &multipleFaceData);
+    if (ret != HSUCCEED) {
+        std::cerr << "Face tracking failed on second image" << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+
+    if (multipleFaceData.detectedNum == 0) {
+        std::cerr << "No face detected in second image" << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+
+    // Extract features from second image
+    std::vector<float> feature2(featureLength);
+    ret = HFFaceFeatureExtractCpy(handle, img, multipleFaceData.tokens[0], feature2.data());
+    if (ret != HSUCCEED) {
+        std::cerr << "Feature extraction failed for second image" << std::endl;
+        HFReleaseImageStream(img);
+        return false;
+    }
+    HFReleaseImageStream(img);
+
+    // Compare features
+    HFFaceFeature faceFeature1 = {0};
+    faceFeature1.data = feature1.data();
+    faceFeature1.size = feature1.size();
+
+    HFFaceFeature faceFeature2 = {0};
+    faceFeature2.data = feature2.data();
+    faceFeature2.size = feature2.size();
+
+    ret = HFFaceComparison(faceFeature1, faceFeature2, &similarity);
+    if (ret != HSUCCEED) {
+        std::cerr << "Face comparison failed" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+// Get the current memory usage in MB
+inline size_t getCurrentMemoryUsage() {
+#ifdef _WIN32
+    PROCESS_MEMORY_COUNTERS_EX pmc;
+    GetProcessMemoryInfo(GetCurrentProcess(), (PROCESS_MEMORY_COUNTERS*)&pmc, sizeof(pmc));
+    return pmc.WorkingSetSize / (1024 * 1024);  // Convert bytes to MB
+#else
+    struct rusage rusage;
+    getrusage(RUSAGE_SELF, &rusage);
+    return (size_t)rusage.ru_maxrss / 1024;  // ru_maxrss is in KB, convert to MB
+#endif
+}
+
+#endif  // INSPIREFACE_TEST_TEST_HELP_H
diff --git a/cpp-package/inspireface/cpp/test/unit/test_helper/test_tools.h b/cpp-package/inspireface/cpp/test/unit/test_helper/test_tools.h
new file mode 100644
index 000000000..0f6b76c60
--- /dev/null
+++ b/cpp-package/inspireface/cpp/test/unit/test_helper/test_tools.h
@@ -0,0 +1,60 @@
+/**
+ * Created by Jingyu Yan
+ * @date 2024-10-01
+ */
+#pragma
+#ifndef INSPIREFACE_TEST_TOOLS_H
+#define INSPIREFACE_TEST_TOOLS_H
+
+#include "inspireface/c_api/inspireface.h"
+#include <fstream>
+#include <cstdint>  // For uint8_t
+#include <inspirecv/inspirecv.h>
+
+inline HResult CVImageToImageStream(const inspirecv::Image &image, HFImageStream &handle, HFImageFormat format = HF_STREAM_BGR,
+                                    HFRotation rot = HF_CAMERA_ROTATION_0) {
+    if (image.Empty()) {
+        return -1;
+    }
+    HFImageData imageData = {0};
+    imageData.data = (uint8_t *)image.Data();
+    imageData.height = image.Height();
+    imageData.width = image.Width();
+    imageData.format = format;
+    imageData.rotation = rot;
+
+    auto ret = HFCreateImageStream(&imageData, &handle);
+
+    return ret;
+}
+
+inline uint8_t *ReadNV21Data(const char *filePath, int width, int height) {
+    const int nv21Size = width * height * 3 / 2;  // Calculate the NV21 data size
+
+    // Memory is allocated dynamically to store NV21 data
+    uint8_t *nv21Data = new uint8_t[nv21Size];
+
+    std::ifstream file(filePath, std::ios::binary);
+    if (!file.is_open()) {
+        std::cerr << "Unable to open the file " << filePath << std::endl;
+        delete[] nv21Data;
+        return nullptr;
+    }
+
+    // Read data
+    file.read(reinterpret_cast<char *>(nv21Data), nv21Size);
+    if (!file) {
+        std::cerr << "Read error or incomplete file" << std::endl;
+        file.close();
+        delete[] nv21Data;
+        return nullptr;
+    }
+
+    // Open file
+    file.close();
+
+    // Returns a pointer to NV21 data
+    return nv21Data;
+}
+
+#endif  // INSPIREFACE_TEST_TOOLS_H
diff --git a/cpp-package/inspireface/doc/Benchmark-Remark(Updating).md b/cpp-package/inspireface/doc/Benchmark-Remark(Updating).md
new file mode 100644
index 000000000..d7ff57a62
--- /dev/null
+++ b/cpp-package/inspireface/doc/Benchmark-Remark(Updating).md
@@ -0,0 +1,105 @@
+# Benchmark Remark(Updating)
+
+The benchmark tests will be continuously updated.
+
+## Pikachu_Apple(ANE, Apple Neural Engine)
+
+### Device: Mac mini 2023 , Apple M2[FP32]
+| **Benchmark**   | **Loops** | **Total Time** | **Average Time** |
+| --------------- | --------- | -------------- | ---------------- |
+| Face Detect@160 | 1000      | 553 ms         | **0.53 ms**      |
+| Face Detect@320 | 1000      | 1635ms         | **1.64ms**       |
+| Face Detect@640 | 1000      | 5903ms         | **5.90ms**       |
+| Face Comparison | 1000      | 1ms            | **0.001ms**      |
+
+**Note**: The above data inference backend uses CoreML.
+
+
+## Megatron_Apple(ANE, Apple Neural Engine)
+
+### Device: iPhone13, Apple A15[FP32]
+| **Benchmark**          | **Loops** | **Total Time** | **Average Time** |
+| ---------------------- | --------- | -------------- | ---------------- |
+| Face Detect@160        | 1000      | 711 ms         | **0.71 ms**      |
+| Face Detect@192        | 1000      | 832ms          | **0.83ms**       |
+| Face Detect@256        | 1000      | 931ms          | **0.92ms**       |
+| Face Detect@320        | 1000      | 1324ms         | **1.32ms**       |
+| Face Detect@640        | 1000      | 3881ms         | **3.88ms**       |
+| Face Extract(**MNet**) | 1000      | 853ms          | **0.85ms**       |
+| Face Extract(**R50**)  | 1000      | 3856ms         | **3.86ms**       |
+
+### Device: Mac mini 2023 , Apple M2[FP32]
+| **Benchmark**          | **Loops** | **Total Time** | **Average Time** |
+| ---------------------- | --------- | -------------- | ---------------- |
+| Face Detect@160        | 1000      | 414 ms         | **0.45 ms**      |
+| Face Detect@192        | 1000      | 574ms          | **0.57ms**       |
+| Face Detect@256        | 1000      | 769ms          | **0.77ms**       |
+| Face Detect@320        | 1000      | 1073ms         | **1.07ms**       |
+| Face Detect@640        | 1000      | 3743ms         | **3.74ms**       |
+| Face Extract(**MNet**) | 1000      | 573ms          | **0.57ms**       |
+| Face Extract(**R50**)  | 1000      | 3527ms         | **3.53ms**       |
+
+**Note**: The above data inference backend uses CoreML.
+
+## Pikachu(CPU)
+### Device: Macbook pro 16-inch, 2019 2.6 GHz Intel Core i7[FP32]
+| **Benchmark** | **Loops** | **Total Time** | **Average Time** |
+| --- | --- | --- | --- |
+| Face Detect@160          | 1000      | 4170.91578 ms  | **4.1709 ms**    |
+| Face Detect@320          | 1000      | 8493.06583 ms  | **8.4893 ms**    |
+| Face Detect@640          | 1000      | 25808.39749 ms | **25.808 ms**    |
+| Face Light-Track | 1000 | 1957.73 ms | **1.9577 ms** |
+| Face alignment & Extract | 1000 | 6139.67 ms | **6.1397 ms** |
+| Face Comparison | 1000 | 0.24ms  | **0.0002ms** |
+| Search Face from 1k | 1000 | 72.39ms | **0.07ms** |
+| Search Face from 5k | 1000 | 364.21ms | **0.36ms** |
+| Search Face from 10k | 1000 | 1193.26ms | **1.19ms** |
+
+## Gundam_RV1109(RKNPU)
+### Device: RV1126[INT8]
+| **Benchmark** | **Loops** | **Total Time** | **Average Time** |
+| --- | --- | --- | --- |
+| Face Detect@160          | 1000      | 17342.88616ms  | **17.34289ms**   |
+| Face Detect@320          | 1000      | 22638.11865ms  | **22.63812ms**   |
+| Face Detect@640          | 1000      | 39745.28562ms  | **39.74529ms**   |
+| Face Light-Track | 1000 | 8858.03802ms | **8.85804ms** |
+| Face alignment & Extract | 1000 | 42352.03367ms | **42.35203ms** |
+| Face Comparison | 1000 | 1.30754ms  | **0.00131ms** |
+| Search Face from 1k | 1000 | 3198.13874ms | **3.19814ms** |
+| Search Face from 5k | 1000 | 15745.00533ms | **15.74501ms** |
+| Search Face from 10k | 1000 | 31267.2301ms | **31.26723ms** |
+
+## Gundam_RV1106(RKNPU2)
+### Device: RV1106[INT8]
+| **Benchmark** | **Loops** | **Total Time** | **Average Time** |
+| --- | --- | --- | --- |
+| Face Detect@160          | 1000      | 23776ms         | **23.78ms**      |
+| Face Detect@320          | 1000      | 33310ms         | **33.31ms**      |
+| Face Detect@640          | 1000      | 58631ms         | **58.63ms**      |
+| Face Light-Track         | 1000      | 15642ms         | **15.64ms**      |
+| Face alignment & Extract | 1000      | 15178ms         | **15.18ms**      |
+| Face Comparison          | 1000      | 23ms          | **0.023us**      |
+
+## Gundam_RK356X(RKNPU2)
+### Device: RK3568[INT8]
+| **Benchmark** | **Loops** | **Total Time** | **Average Time** |
+| --- | --- | --- | --- |
+| Face Detect@160          | 1000      | 16946ms         | **16.95ms**      |
+| Face Detect@320          | 1000      | 25108ms         | **25.11ms**      |
+| Face Detect@640          | 1000      | 68778ms         | **68.78ms**      |
+| Face Light-Track         | 1000      | 11215ms         | **11.22ms**      |
+| Face alignment & Extract | 1000      | 9070ms          | **9.07ms**       |
+| Face Comparison          | 1000      | 9ms           | **0.009ms**      |
+
+## Megatron_TRT(CUDA)
+### Device: NVIDIA RTX3060 12G[FP16]
+| **Benchmark** | **Loops** | **Total Time** | **Average Time** |
+| --- | --- | --- | --- |
+| Face Detect@160          | 1000      | 911ms           | **0.91ms**      |
+| Face Detect@320          | 1000      | 2374ms          | **2.37ms**      |
+| Face Detect@640          | 1000      | 8685ms          | **8.685ms**      |
+| Face Light-Track         | 1000      | 621ms           | **0.62ms**      |
+| Face alignment & Extract | 1000      | 1009ms          | **1.01ms**       |
+| Face Comparison          | 1000      | 1ms           | **0.001ms**      |
+
+**Note**: The test results are all calculated by the test programs in the '**cpp/test**' subproject.
diff --git a/cpp-package/inspireface/doc/CMake-Option.md b/cpp-package/inspireface/doc/CMake-Option.md
new file mode 100644
index 000000000..e96e4da1e
--- /dev/null
+++ b/cpp-package/inspireface/doc/CMake-Option.md
@@ -0,0 +1,37 @@
+# Overview of CMake Option for Compilation
+
+Here are the translation details for the compilation parameters as per your requirement:
+
+| **Parameter**                         | **Default Value** | **Description**                                                                                                                                            |
+|---------------------------------------| :---: |------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| ISF_THIRD_PARTY_DIR                   | 3rdparty | Path for required third-party libraries                                                                                                                    |
+| ISF_SANITIZE_ADDRESS                      | OFF | Enable AddressSanitizer for memory error detection                                                                                                         |
+| ISF_SANITIZE_LEAK                         | OFF | Enable LeakSanitizer to detect memory leaks                                                                                                                |
+| ISF_ENABLE_SYMBOL_HIDING | ON | Enable symbol hiding by default for better security and performance. Only symbols explicitly marked for export will be visible. |
+| ISF_INSTALL_CPP_HEADER | OFF | Whether to install the headers file for **CPP-API** (the default **C-API** with wider compatibility is recommended) |
+| ISF_ENABLE_RKNN                           | OFF | Enable RKNN for Rockchip embedded devices                                                                                                                  |
+| ISF_RK_DEVICE_TYPE                        | RV1109RV1126 | Target device model for Rockchip(Supports RV1109RV1126, RV1106, RV356X) |
+| ISF_RK_COMPILER_TYPE | armhf | The **armhf**, **armhf-uclibc** and aarch64 compilers are supported. Select one based on the actual situation |
+| ISF_ENABLE_RGA | OFF | Enable **RGA** image acceleration on Rockchip devices (currently only supported on devices using **RKNPU2**) |
+| ISF_BUILD_LINUX_ARM7                      | OFF | Compile for ARM7 architecture                                                                                                                              |
+| ISF_BUILD_LINUX_AARCH64                   | OFF | Compile for AARCH64 architecture                                                                                                                           |
+| ISF_BUILD_WITH_TEST                       | OFF | Compile test case programs                                                                                                                                 |
+| ISF_BUILD_WITH_SAMPLE                     | ON | Compile sample programs                                                                                                                                    |
+| ISF_BUILD_SHARED_LIBS                     | ON | Compile shared libraries                                                                                                                                   |
+| ISF_ENABLE_BENCHMARK                      | OFF | Enable Benchmark tests for test cases                                                                                                                      |
+| ISF_ENABLE_USE_LFW_DATA                   | OFF | Enable using LFW data for test cases                                                                                                                       |
+| ISF_ENABLE_TEST_EVALUATION                | OFF | Enable evaluation functionality for test cases, must be used together with ISF_ENABLE_USE_LFW_DATA                                                         |
+| ISF_ENABLE_TEST_INTERNAL | OFF | Enable test cases for some internal functions, requires disabling **ISF_ENABLE_SYMBOL_HIDING** for compilation. |
+| ISF_BUILD_SAMPLE_INTERNAL | OFF | Enable executable examples for some internal functions, requires disabling **ISF_ENABLE_SYMBOL_HIDING** for compilation. |
+| ISF_ENABLE_TENSORRT | OFF | Enable the backend of inference using TensorRT, Linux must be on **NVIDIA devices**, and **CUDA**, **TensorRT-10** must be installed |
+| TENSORRT_ROOT | /usr/local/TensorRT | **TensorRT-10** installation path |
+| ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA | OFF | Enable global MNN_CUDA inference mode, requires device support for CUDA                                                                                    |
+| ISF_LINUX_MNN_CUDA                        | "" | Specific MNN library path, requires pre-compiled MNN library supporting MNN_CUDA, only effective when ISF_GLOBAL_INFERENCE_BACKEND_USE_MNN_CUDA is enabled |
+| ISF_ENABLE_APPLE_EXTENSION | OFF | If you are using an **Apple device (MacOS/iOS)**, you can enable it to allow some parts of the SDK's models to switch to certain backend neural network acceleration inference on Apple devices, such as Metal and ANE |
+| ISF_MNN_CUSTOM_SOURCE | "" | Using this option to replace different versions of MNN, the input must be the root directory of the MNN source code |
+| INSPIRECV_BACKEND_OPENCV | OFF | The image processing backend relies on OpenCV and is **not recommended** |
+| ISF_ENABLE_COST_TIME | OFF | This parameter can be used to print the time of some important compute nodes in the Debug state |
+| ISF_NEVER_USE_OPENCV | ON | When you need to use OpenCV as the image processing engine, you need to turn this option off |
+
+
+
diff --git a/cpp-package/inspireface/doc/Error-Feedback-Codes.md b/cpp-package/inspireface/doc/Error-Feedback-Codes.md
new file mode 100644
index 000000000..159b9229a
--- /dev/null
+++ b/cpp-package/inspireface/doc/Error-Feedback-Codes.md
@@ -0,0 +1,51 @@
+# Error Feedback Codes
+
+During the use of InspireFace, some error feedback codes may be generated. Here is a table of error feedback codes.
+
+- As of **June 15, 2025**, the error code definitions have been restructured. Some legacy codes from historical versions have been removed, and a more streamlined version has been reorganized and consolidated.
+
+ | Index | Name | Code | Comment | 
+ | --- | --- | --- | --- | 
+ | 1 | HSUCCEED | 0 | Success | 
+ | 2 | HERR_UNKNOWN | 1 | Unknown error (1) | 
+ | 3 | HERR_INVALID_PARAM | 2 | Invalid parameter (2) | 
+ | 4 | HERR_INVALID_IMAGE_STREAM_HANDLE | 3 | Invalid image stream handle (3) | 
+ | 5 | HERR_INVALID_CONTEXT_HANDLE | 4 | Invalid context handle (4) | 
+ | 6 | HERR_INVALID_FACE_TOKEN | 5 | Invalid face token (5) | 
+ | 7 | HERR_INVALID_FACE_FEATURE | 6 | Invalid face feature (6) | 
+ | 8 | HERR_INVALID_FACE_LIST | 7 | Invalid face feature list (7) | 
+ | 9 | HERR_INVALID_BUFFER_SIZE | 8 | Invalid copy token (8) | 
+ | 10 | HERR_INVALID_IMAGE_STREAM_PARAM | 9 | Invalid image param (9) | 
+ | 11 | HERR_INVALID_SERIALIZATION_FAILED | 10 | Invalid face serialization failed (10) | 
+ | 12 | HERR_INVALID_DETECTION_INPUT | 11 | Failed to modify detector input size (11) | 
+ | 13 | HERR_INVALID_IMAGE_BITMAP_HANDLE | 12 | Invalid image bitmap handle (12) | 
+ | 14 | HERR_IMAGE_STREAM_DECODE_FAILED | 13 | ImageStream failed to decode the image (13) | 
+ | 15 | HERR_SESS_FUNCTION_UNUSABLE | 101 | Function not usable (101) | 
+ | 16 | HERR_SESS_TRACKER_FAILURE | 102 | Tracker module not initialized (102) | 
+ | 17 | HERR_SESS_PIPELINE_FAILURE | 103 | Pipeline module not initialized (103) | 
+ | 18 | HERR_SESS_INVALID_RESOURCE | 104 | Invalid static resource (104) | 
+ | 19 | HERR_SESS_LANDMARK_NUM_NOT_MATCH | 105 | The number of input landmark points does not match (105) | 
+ | 20 | HERR_SESS_LANDMARK_NOT_ENABLE | 106 | The landmark model is not enabled (106) | 
+ | 21 | HERR_SESS_KEY_POINT_NUM_NOT_MATCH | 107 | The number of input key points does not match (107) | 
+ | 22 | HERR_SESS_REC_EXTRACT_FAILURE | 108 | Face feature extraction not registered (108) | 
+ | 23 | HERR_SESS_REC_CONTRAST_FEAT_ERR | 109 | Incorrect length of feature vector for comparison (109) | 
+ | 24 | HERR_SESS_FACE_DATA_ERROR | 110 | Face data parsing (110) | 
+ | 25 | HERR_SESS_FACE_REC_OPTION_ERROR | 111 | An optional parameter is incorrect (111) | 
+ | 26 | HERR_FT_HUB_DISABLE | 201 | FeatureHub is disabled (201) | 
+ | 27 | HERR_FT_HUB_INSERT_FAILURE | 202 | Data insertion error (202) | 
+ | 28 | HERR_FT_HUB_NOT_FOUND_FEATURE | 203 | Get face feature error (203) | 
+ | 29 | HERR_ARCHIVE_LOAD_FAILURE | 251 | Archive load failure (251) | 
+ | 30 | HERR_ARCHIVE_LOAD_MODEL_FAILURE | 252 | Model load failure (252) | 
+ | 31 | HERR_ARCHIVE_FILE_FORMAT_ERROR | 253 | The archive format is incorrect (253) | 
+ | 32 | HERR_ARCHIVE_REPETITION_LOAD | 254 | Do not reload the model (254) | 
+ | 33 | HERR_ARCHIVE_NOT_LOAD | 255 | Model not loaded (255) | 
+ | 34 | HERR_DEVICE_CUDA_NOT_SUPPORT | 301 | CUDA not supported (301) | 
+ | 35 | HERR_DEVICE_CUDA_TENSORRT_NOT_SUPPORT | 302 | CUDA TensorRT not supported (302) | 
+ | 36 | HERR_DEVICE_CUDA_UNKNOWN_ERROR | 303 | CUDA unknown error (303) | 
+ | 37 | HERR_DEVICE_CUDA_DISABLE | 304 | CUDA support is disabled (304) | 
+ | 38 | HERR_EXTENSION_ERROR | 351 | Extension module error (351) | 
+ | 39 | HERR_EXTENSION_MLMODEL_LOAD_FAILED | 352 | MLModel load failed (352) | 
+ | 40 | HERR_EXTENSION_HETERO_MODEL_TAG_ERROR | 353 | Incorrect heterogeneous model tag (353) | 
+ | 41 | HERR_EXTENSION_HETERO_REC_HEAD_CONFIG_ERROR | 354 | Rec head config error (354) | 
+ | 42 | HERR_EXTENSION_HETERO_MODEL_NOT_MATCH | 355 | Heterogeneous model dimensions do not match (355) | 
+ | 43 | HERR_EXTENSION_HETERO_MODEL_NOT_LOADED | 356 | Heterogeneous model dimensions not loaded (356) | 
diff --git a/cpp-package/inspireface/doc/diagrams/communication_sequence_diagram.puml b/cpp-package/inspireface/doc/diagrams/communication_sequence_diagram.puml
new file mode 100644
index 000000000..946d21f0b
--- /dev/null
+++ b/cpp-package/inspireface/doc/diagrams/communication_sequence_diagram.puml
@@ -0,0 +1,88 @@
+@startuml
+title Communication
+actor User
+participant "Client Application" as Client
+participant "InspireFace SDK" as SDK
+database "FeatureHub" as Database
+
+User -> Client : Start application
+activate Client
+
+Client -> SDK : Initialize SDK\nHFLaunchInspireFace()
+activate SDK
+SDK -> Client : Return initialization result
+deactivate SDK
+
+Client -> SDK : Create session\nHFCreateInspireFaceSession()
+activate SDK
+SDK -> Client : Return session handle
+deactivate SDK
+
+User -> Client : Send image or video data stream
+activate Client
+
+Client -> Client : Process data stream
+Client -> SDK : Create image stream\nHFCreateImageStream()
+activate SDK
+SDK -> Client : Return image stream handle
+deactivate SDK
+
+Client -> SDK : Request face detection\nHFExecuteFaceTrack()
+activate SDK
+SDK -> Client : Return detection results\n(detectedNum, faceRect[])
+deactivate SDK
+
+Client -> SDK : Perform face pipeline processing\nHFMultipleFacePipelineProcess()
+activate SDK
+SDK -> Client : Return processing results
+deactivate SDK
+
+Client -> SDK : Get RGB liveness confidence\nHFGetRGBLivenessConfidence()
+activate SDK
+SDK -> Client : Return liveness detection results
+deactivate SDK
+
+Client -> SDK : Get mask detection confidence\nHFGetFaceMaskConfidence()
+activate SDK
+SDK -> Client : Return mask detection results
+deactivate SDK
+
+Client -> SDK : Get face quality prediction confidence\nHFGetFaceQualityConfidence()
+activate SDK
+SDK -> Client : Return face quality results
+deactivate SDK
+
+alt Liveness confirmed and no mask and good quality
+    loop Each qualified face
+        Client -> SDK : Request face feature extraction\nHFFaceFeatureExtract()
+        activate SDK
+        SDK -> Client : Return feature data
+        deactivate SDK
+
+        Client -> SDK : Search for matching face features\nHFFeatureHubFaceSearch()
+        activate SDK
+        SDK -> Database : Request search matching
+        activate Database
+        Database -> SDK : Return match results\n(customId, tag, confidence)
+        deactivate Database
+        SDK -> Client : Send match results
+        deactivate SDK
+    end
+else
+    Client -> User : Display faces do not meet search criteria
+end
+
+Client -> SDK : Release image stream\nHFReleaseImageStream()
+activate SDK
+SDK -> Client : Confirm image stream release
+deactivate SDK
+
+Client -> User : Display face recognition results
+deactivate Client
+
+Client -> SDK : Release session\nHFReleaseInspireFaceSession()
+activate SDK
+SDK -> Client : Confirm session release
+deactivate SDK
+
+@enduml
diff --git a/cpp-package/inspireface/doc/diagrams/component diagram.puml b/cpp-package/inspireface/doc/diagrams/component diagram.puml
new file mode 100644
index 000000000..ab4a9204a
--- /dev/null
+++ b/cpp-package/inspireface/doc/diagrams/component diagram.puml	
@@ -0,0 +1,79 @@
+@startuml
+package "InspireFace SDK" {
+    [face_context] as FC
+
+    package "Algorithm Modules" {
+        package "Track Module" as TrackModule {
+            [face_detect] as FD
+            [face_pose] as FP
+            [face_track] as FT
+            [landmark] as LM
+            FT --> FD : uses
+            FT --> LM : uses
+            FT --> FP : uses
+        }
+
+        package "Recognition Module" as RecognitionModule {
+            [feature_extract] as FE
+            [alignment] as AL
+        }
+
+        package "Pipeline Module" as PipelineModule {
+                    [mask_predict] as MA
+                    [face_liveness] as FL
+                    [face_quality] as QU
+                }
+    }
+
+    package "AnyNet" as AnyNet {
+        [MNN Engine] as MNN
+        [RKNN Engine] as RKNN
+        [OpenCV-DNN Engine] as OCV_DNN
+        [AnyNet Module] as ANY
+        ANY -down-> MNN : supports
+        ANY -down-> RKNN : supports
+        ANY -down-> OCV_DNN : supports
+    }
+
+    package "FeatureHub" as FHM{
+        package "Feature Blocks" {
+            [None Engine] as NE
+            [OpenCV Engine] as OCV
+            [Eigen Engine] as EIG
+            [Feature Block] as FB
+            FB -up-> NE : implements
+            FB -up-> OCV : implements
+            FB -up-> EIG : implements
+        }
+
+        package "Persistence Module" as PM{
+            [SQLite Engine] as SQL
+            [Persistence] as PS
+            PS -right-> SQL : uses
+        }
+
+        [feature_hub] as FH
+        FH -down-> FB : includes
+        FH -down-> PM : includes
+    }
+
+    package "Initialization Module" {
+        [inspire_archive] as IA
+        [resource_decode] as RD
+        [Launch] as MH
+        MH -down-> IA : contains
+        MH -down-> RD : contains
+    }
+
+    FC --> TrackModule : uses
+    FC --> RecognitionModule : uses
+    FC --> PipelineModule : uses
+    FC --> AnyNet : depends on
+    FC --> FH : uses
+    FC --> MH : initializes
+
+    [log] as LG
+    FC ..> LG : logs activities
+}
+
+@enduml
diff --git a/cpp-package/inspireface/doc/diagrams/mem_model.drawio b/cpp-package/inspireface/doc/diagrams/mem_model.drawio
new file mode 100644
index 000000000..17cc35dd6
--- /dev/null
+++ b/cpp-package/inspireface/doc/diagrams/mem_model.drawio
@@ -0,0 +1,115 @@
+<mxfile host="Electron" agent="Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) draw.io/24.7.5 Chrome/126.0.6478.183 Electron/31.3.0 Safari/537.36" version="24.7.5">
+  <diagram name="page 1" id="FP0FvKHnDjYPDM2ZKx7P">
+    <mxGraphModel dx="781" dy="688" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="827" pageHeight="1169" math="0" shadow="0">
+      <root>
+        <mxCell id="0" />
+        <mxCell id="1" parent="0" />
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-44" value="" style="rounded=0;whiteSpace=wrap;html=1;fillColor=none;strokeWidth=2;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="30" y="140" width="670" height="530" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-39" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="460" y="230" width="220" height="100" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-2" value="&lt;b&gt;&lt;font style=&quot;font-size: 14px;&quot;&gt;Client&lt;/font&gt;&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#fad9d5;strokeColor=#ae4132;" vertex="1" parent="1">
+          <mxGeometry x="50" y="150" width="630" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-3" value="&lt;b&gt;Session 1&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="60" y="281" width="90" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-5" value="...." style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="280" y="281" width="50" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-9" value="&lt;b&gt;Session 2&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="170" y="281" width="90" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-10" value="&lt;b&gt;Session N&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="350" y="281" width="90" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-11" value="&lt;b&gt;Image Stream&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1">
+          <mxGeometry x="470" y="281" width="90" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-12" value="&lt;b&gt;Image Bitmap&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1">
+          <mxGeometry x="570" y="281" width="100" height="40" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-13" value="&lt;b&gt;&lt;font style=&quot;font-size: 14px;&quot;&gt;App Context&lt;/font&gt;&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffe6cc;strokeColor=#d79b00;" vertex="1" parent="1">
+          <mxGeometry x="60" y="382" width="380" height="60" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-14" value="&lt;b&gt;&lt;font style=&quot;font-size: 14px;&quot;&gt;Feature Hub&lt;/font&gt;&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
+          <mxGeometry x="470" y="382" width="200" height="60" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-17" value="Face Recognition Algorithm Function" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="60" y="241" width="380" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-18" value="Image CODEC and Processing" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1">
+          <mxGeometry x="470" y="241" width="200" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-20" value="Global&amp;nbsp;Launch" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffe6cc;strokeColor=#d79b00;" vertex="1" parent="1">
+          <mxGeometry x="60" y="342" width="110" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-21" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;&lt;font style=&quot;font-size: 12px;&quot; face=&quot;Helvetica&quot;&gt;Performance Configuration&lt;/font&gt;&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffe6cc;strokeColor=#d79b00;" vertex="1" parent="1">
+          <mxGeometry x="180" y="342" width="150" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-22" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;Common Api&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffe6cc;strokeColor=#d79b00;" vertex="1" parent="1">
+          <mxGeometry x="344" y="342" width="96" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-23" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;Storage Mode&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
+          <mxGeometry x="470" y="342" width="110" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-24" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;CRUD&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
+          <mxGeometry x="587" y="342" width="83" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-25" value="&lt;span style=&quot;font-size: 14px;&quot;&gt;&lt;b&gt;Underlying&amp;nbsp;&lt;/b&gt;&lt;/span&gt;&lt;b style=&quot;font-size: 14px;&quot;&gt;Engine&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="60" y="540" width="310" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-27" value="Heterogeneous Multi-Inference Engine" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="60" y="460" width="230" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-28" value="Multi-Backend Image Processing" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="60" y="500" width="190" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-29" value="&lt;span style=&quot;font-size: 14px;&quot;&gt;&lt;b&gt;Hardware Support&lt;/b&gt;&lt;/span&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="400" y="540" width="270" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-30" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;Image Processing Acceleration Unit&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="400" y="500" width="270" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-32" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; color: rgba(0, 0, 0, 0.85); text-align: start;&quot; class=&quot;p1&quot;&gt;Tensor Acceleration Unit&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="520" y="460" width="150" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-33" value="&lt;p style=&quot;margin: 0px; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-kerning: auto; font-optical-sizing: auto; font-feature-settings: normal; font-variation-settings: normal; font-variant-position: normal; font-stretch: normal; line-height: normal; text-align: start;&quot; class=&quot;p1&quot;&gt;&lt;font face=&quot;Helvetica&quot; style=&quot;font-size: 12px;&quot;&gt;Memory O&lt;/font&gt;&lt;span style=&quot;background-color: initial; text-align: center;&quot;&gt;ptimize&lt;/span&gt;&lt;/p&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="400" y="460" width="110" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-34" value="Vector Database" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="260" y="500" width="110" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-36" value="Abstract" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="300" y="460" width="70" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-37" value="&lt;span style=&quot;font-size: 14px;&quot;&gt;&lt;b&gt;Platform&lt;/b&gt;&lt;/span&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#fad7ac;strokeColor=#b46504;" vertex="1" parent="1">
+          <mxGeometry x="50" y="630" width="630" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-38" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="50" y="230" width="400" height="100" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-40" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="460" y="337" width="220" height="110" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-41" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="50" y="335" width="400" height="113" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-42" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="50" y="452" width="330" height="128" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-43" value="" style="rounded=1;whiteSpace=wrap;html=1;fillColor=none;dashed=1;dashPattern=1 2;" vertex="1" parent="1">
+          <mxGeometry x="390" y="453" width="290" height="127" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-46" value="&lt;b&gt;&lt;font style=&quot;font-size: 14px;&quot;&gt;API Gateway(C/C++/Java/Objective-C/JS)&lt;/font&gt;&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#bac8d3;strokeColor=#23445d;" vertex="1" parent="1">
+          <mxGeometry x="50" y="190" width="630" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="ycdcpBqM2tXTCb_d-7C_-47" value="&lt;b&gt;&lt;font style=&quot;font-size: 14px;&quot;&gt;System(Linux/Android/iOS/MacOS)&lt;/font&gt;&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#bac8d3;strokeColor=#23445d;" vertex="1" parent="1">
+          <mxGeometry x="50" y="590" width="630" height="30" as="geometry" />
+        </mxCell>
+      </root>
+    </mxGraphModel>
+  </diagram>
+</mxfile>
diff --git a/cpp-package/inspireface/doc/diagrams/sample_pipeline.puml b/cpp-package/inspireface/doc/diagrams/sample_pipeline.puml
new file mode 100644
index 000000000..053b0a6fe
--- /dev/null
+++ b/cpp-package/inspireface/doc/diagrams/sample_pipeline.puml
@@ -0,0 +1,56 @@
+@startuml InspireFace SDK Usage Sequence
+!theme plain
+
+title Execution Sequence
+footer : 20250615 create by Jingyu Yan
+header : @InspireFace
+
+participant "Application" as App
+participant "InspireFace SDK" as SDK
+participant "FaceSession" as Session
+
+note over SDK, Session
+  Sessions support multiple and are allocated across multiple threads, 
+	but cache data between sessions cannot be mixed
+end note
+
+== Initialization Phase ==
+App -> SDK: HFLaunchInspireFace
+SDK --> App: Return initialization result
+
+== Session Creation & Configuration ==
+App -> SDK: HFCreateInspireFaceSessionOptional
+SDK -> Session: Create session instance
+SDK --> App: Return session handle
+
+== Image Preparation ==
+App -> SDK: HFCreateImageStream
+SDK --> App: Return image handle
+
+== Face Detection & Processing ==
+App -> Session: HFExecuteFaceTrack
+activate Session
+Session --> App: Return MultipleFaceData
+
+App -> Session: HFMultipleFacePipelineProcessOptional
+Session --> App: Pipeline processing completed
+
+== Result Retrieval Phase ==
+App -> Session: HFGetFaceMaskConfidence
+Session --> App: Mask detection result
+
+App -> Session: HFGetFaceQualityConfidence
+Session --> App: Quality assessment result
+
+App -> Session: Other APIs
+Session --> App: ....
+
+== Resource Release ==
+App -> SDK: HFReleaseImageStream
+
+== Program Termination ==
+
+App -> Session: HFReleaseInspireFaceSession
+App -> SDK: HFTerminateInspireFace
+
+@enduml
\ No newline at end of file
diff --git a/cpp-package/inspireface/doc/modelscope_readme.md b/cpp-package/inspireface/doc/modelscope_readme.md
new file mode 100644
index 000000000..a819000d2
--- /dev/null
+++ b/cpp-package/inspireface/doc/modelscope_readme.md
@@ -0,0 +1,84 @@
+# InspireFace
+
+[![pypi](https://img.shields.io/pypi/v/inspireface.svg?style=for-the-badge&color=orange&label=PYPI+release&logo=python)](https://pypi.org/project/inspireface/)
+[![GitHub release](https://img.shields.io/github/v/release/HyperInspire/InspireFace.svg?style=for-the-badge&color=blue&label=Github+release&logo=github)](https://github.com/HyperInspire/InspireFace/releases/latest)
+
+InspireFace is a cross-platform face recognition SDK developed in C/C++, supporting multiple operating systems and various backend types for inference, such as CPU, GPU, and NPU.
+
+## Key Features
+
+- **Face Detection & Tracking** - High-precision face detection and real-time tracking
+- **Face Recognition** - Face feature extraction and comparison
+- **Landmark Detection** - Accurate facial landmark localization
+- **Face Attributes** - Age, gender, and emotion recognition
+- **Liveness Detection** - Silent and cooperative liveness detection
+- **Mask Detection** - Face mask detection capability
+- **Quality Assessment** - Face quality evaluation
+- **Pose Estimation** - Face pose angle estimation
+
+<img src="https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/blogs_box/o-10.gif" width="200" height="200"> 
+
+## Project Origin
+
+InspireFace is derived from the InsightFace project, specifically from the C++ package implementation. 
+
+**Original Repository**: [https://github.com/deepinsight/insightface/tree/master/cpp-package/inspireface](https://github.com/deepinsight/insightface/tree/master/cpp-package/inspireface)
+
+## Supported Platforms and Architectures
+
+| No. | Platform | Architecture | Device | **Supported** | Passed Tests |
+| ------- | -------------------- | --------------------- | -------------------------- | :-----------: | :----------------: |
+| 1       | **Linux**(CPU)      | ARMv7                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 2       |                      | ARMv8                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 3       |                      | x86/x86_64            | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 4       | **Linux**(Rockchip) | ARMv7                 | RV1109/RV1126              | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 5 | | ARMv7 | RV1103/RV1106 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 6 | | ARMv8 | RK3566/RK3568 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 7 | | ARMv8 | RK3588 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - |
+| 8      | **Linux**(MNN_CUDA) | x86/x86_64            | NVIDIA-GPU          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 9      | **Linux**(CUDA) | x86/x86_64            | NVIDIA-GPU          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 10      | **MacOS**           | Intel       | CPU/Metal/**ANE** | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 11   |                      | Apple Silicon         | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 12     | **iOS**              | ARM                   | CPU/Metal/**ANE**         | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 13     | **Android**          | ARMv7                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 14     |                      | ARMv8                 | -                          | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 15 | | x86_64 | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 16 | **Android**(Rockchip) | ARMv8 | RK3566/RK3568 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| 17 |  | ARMv8 | RK3588 | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+
+## Features Support Matrix
+
+| Feature | CPU | RKNPU(RV1109/1126) | RKNPU(RV1103/1106) | RKNPU(RK3566/3568/3588) | ANE(MacOS/iOS) | GPU(TensorRT) |
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| Face Detection | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Landmark | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Embeddings | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Comparison | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+| Face Recognition | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Alignment | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+| Tracking | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Mask Detection | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - |
+| Silent Liveness | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - |
+| Face Quality | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Pose Estimation | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Attribute | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Cooperative Liveness | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Face Emotion | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) |
+| Embedding Management | [![](https://img.shields.io/badge/%E2%9C%93-green)](#) | - | - | - | - | - |
+
+Note: Some models and features that do not support NPU or GPU will automatically use CPU for computation.
+
+## Model Packages
+
+For different scenarios, we provide several model packages, each containing multiple models and configurations:
+
+| Name | Supported Devices | Note | Last Update |
+| --- | --- | --- | --- |
+| Pikachu | CPU | Lightweight edge-side models | Jun 22, 2025 |
+| Megatron | CPU, GPU | Mobile and server models | Jun 15, 2025 |
+| Megatron_TRT | GPU | CUDA-based server models | Jun 15, 2025 |
+| Gundam-RV1109 | RKNPU | Supports RK1109 and RK1126 | Jun 15, 2025 |
+| Gundam-RV1106 | RKNPU | Supports RV1103 and RV1106 | Jul 6, 2025 |
+| Gundam-RK356X | RKNPU | Supports RK3566 and RK3568 | Jun 15, 2025 |
+| Gundam-RK3588 | RKNPU | Supports RK3588 | Jun 15, 2025 |
+
diff --git a/cpp-package/inspireface/docker-compose.yml b/cpp-package/inspireface/docker-compose.yml
new file mode 100644
index 000000000..f94617f2a
--- /dev/null
+++ b/cpp-package/inspireface/docker-compose.yml
@@ -0,0 +1,133 @@
+version: '3'
+services:
+  build-ubuntu18:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.ubuntu18 # Use the Ubuntu18.04
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_linux_ubuntu18.sh
+  build-cross-rv1109rv1126-armhf:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.arm-linux-gnueabihf # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_rv1109rv1126_armhf.sh
+  build-cross-armv7-armhf:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.arm-linux-gnueabihf # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_armv7_armhf.sh
+  build-cross-rv1106-armhf-uclibc:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.arm-linux-rockchip830-armhf-uclibc
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_rv1106_armhf_uclibc.sh
+  build-cross-aarch64:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.arm-linux-aarch64 # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_aarch64.sh
+  build-cross-rk356x-rk3588-aarch64:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.arm-linux-aarch64 # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_rk356x_rk3588_aarch64.sh
+  build-cross-android:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.android # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_android.sh
+  build-cross-android-rk356x-rk3588:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.android # Use the arm-linux-gnueabihf tool chain
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_cross_rk356x_rk3588_aarch64.sh
+  build-cuda-ubuntu20:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.cuda.ubuntu20
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_linux_cuda.sh
+  build-manylinux2014-x86:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.manylinux2014-x86
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_wheel_manylinux2014_x86.sh
+  build-manylinux2014-aarch64:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.manylinux2014_aarch64
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_wheel_manylinux2014_aarch64.sh
+  build-manylinux_2_36-aarch64:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.manylinux_2_36_aarch64
+    environment:
+      - VERSION=${VERSION}
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_wheel_manylinux_2_36_aarch64.sh
+  build-tensorrt-cuda12-ubuntu22:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.cuda12_ubuntu22
+    environment:
+      - VERSION=${VERSION}
+      - CUDA_TAG=cuda12.2_ubuntu22.04
+    working_dir: /workspace
+    volumes:
+      - .:/workspace # Mount the project root directory to the container
+    command: bash command/build_linux_tensorrt.sh
diff --git a/cpp-package/inspireface/docker/Dockerfile.android b/cpp-package/inspireface/docker/Dockerfile.android
new file mode 100644
index 000000000..638408539
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.android
@@ -0,0 +1,56 @@
+# Use Ubuntu 18.04 as the base image
+FROM ubuntu:18.04
+
+# Update the package list and install basic development tools
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        software-properties-common \
+        wget \
+        curl \
+        git \
+        unzip
+
+# Download and install CMake
+ARG CMAKE_URL="https://github.com/Kitware/CMake/releases/download/v3.17.0-rc3/cmake-3.17.0-rc3-Linux-x86_64.sh"
+RUN mkdir /opt/cmake && \
+    wget -qO /opt/cmake/install-cmake.sh ${CMAKE_URL} && \
+    chmod +x /opt/cmake/install-cmake.sh && \
+    /opt/cmake/install-cmake.sh --skip-license --prefix=/usr/local && \
+    rm /opt/cmake/install-cmake.sh
+
+# Set the URL for the Android NDK and OpenCV Android SDK
+ARG ANDROID_NDK_URL="https://dl.google.com/android/repository/android-ndk-r18b-linux-x86_64.zip"
+ARG OPENCV_URL="https://github.com/opencv/opencv/releases/download/4.5.1/opencv-4.5.1-android-sdk.zip"
+
+# Set the installation path for the Android NDK and OpenCV Android SDK
+ARG ANDROID_NDK_PATH="/opt/android-ndk-r18b"
+ARG OPENCV_PATH="/opt/opencv-android-sdk"
+
+# Download and extract the Android NDK
+RUN mkdir -p ${ANDROID_NDK_PATH} && \
+    wget -qO /tmp/android-ndk.zip ${ANDROID_NDK_URL} && \
+    unzip /tmp/android-ndk.zip -d /opt && \
+    rm /tmp/android-ndk.zip
+
+# Set environment variable to point to the NDK directory
+ENV ANDROID_NDK=${ANDROID_NDK_PATH}
+
+# Download and extract the OpenCV Android SDK
+RUN mkdir -p ${OPENCV_PATH} && \
+    wget -qO /tmp/opencv-android-sdk.zip ${OPENCV_URL} && \
+    unzip /tmp/opencv-android-sdk.zip -d ${OPENCV_PATH} && \
+    rm /tmp/opencv-android-sdk.zip
+
+# Set environment variable to point to the OpenCV SDK directory
+ENV OPENCV_DIR=${OPENCV_PATH}/OpenCV-android-sdk/sdk/native/jni
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.arm-linux-aarch64 b/cpp-package/inspireface/docker/Dockerfile.arm-linux-aarch64
new file mode 100644
index 000000000..0d7281718
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.arm-linux-aarch64
@@ -0,0 +1,55 @@
+# Use Ubuntu 18.04 as the base image
+FROM ubuntu:18.04
+
+# Update the package list and install basic development tools
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        software-properties-common \
+        wget \
+        curl \
+        git
+
+# Install CMake with architecture detection
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends cmake
+
+# Alternative: Install specific CMake version with architecture detection
+# RUN apt-get update && \
+#     apt-get install -y --no-install-recommends wget && \
+#     cd /opt && \
+#     if [ "$(uname -m)" = "x86_64" ]; then \
+#         wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+#         tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+#         ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+#         rm cmake-3.20.6-linux-x86_64.tar.gz; \
+#     elif [ "$(uname -m)" = "aarch64" ]; then \
+#         wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-aarch64.tar.gz && \
+#         tar -zxvf cmake-3.20.6-linux-aarch64.tar.gz && \
+#         ln -s /opt/cmake-3.20.6-linux-aarch64/bin/* /usr/local/bin/ && \
+#         rm cmake-3.20.6-linux-aarch64.tar.gz; \
+#     else \
+#         apt-get install -y cmake; \
+#     fi
+
+# Set the URL and installation path for the Linaro toolchain
+ARG LINARO_URL="https://developer.arm.com/-/cdn-downloads/permalink/legacy-linaro-gnu-toolchains/6.3-2017.05/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu.tar.xz"
+ARG TOOLCHAIN_PATH="/opt/linaro-toolchain"
+
+# Create the installation path, download, and extract the Linaro toolchain
+RUN mkdir -p ${TOOLCHAIN_PATH} && \
+    wget -qO- ${LINARO_URL} | tar -xJ -C ${TOOLCHAIN_PATH} --strip-components=1
+
+# Set environment variables to point to the toolchain directory
+ENV ARM_CROSS_COMPILE_TOOLCHAIN=${TOOLCHAIN_PATH}
+ENV PATH="${TOOLCHAIN_PATH}/bin:${PATH}"
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.arm-linux-gnueabihf b/cpp-package/inspireface/docker/Dockerfile.arm-linux-gnueabihf
new file mode 100644
index 000000000..0d062dffd
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.arm-linux-gnueabihf
@@ -0,0 +1,44 @@
+# Use Ubuntu 18.04 as the base image
+FROM ubuntu:18.04
+
+# Update the package list and install basic development tools
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        software-properties-common \
+        wget \
+        curl \
+        git \
+        vim
+
+# Install CMake
+RUN apt-get install -y --no-install-recommends
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Set the URL and installation path for the Linaro toolchain
+ARG LINARO_URL="https://developer.arm.com/-/cdn-downloads/permalink/legacy-linaro-gnu-toolchains/6.3-2017.05/gcc-linaro-6.3.1-2017.05-x86_64_arm-linux-gnueabihf.tar.xz"
+ARG TOOLCHAIN_PATH="/opt/linaro-toolchain"
+
+# Create the installation path, download, and extract the Linaro toolchain
+RUN mkdir -p ${TOOLCHAIN_PATH} && \
+    wget -qO- ${LINARO_URL} | tar -xJ -C ${TOOLCHAIN_PATH} --strip-components=1
+
+# Set environment variables to point to the toolchain directory
+ENV ARM_CROSS_COMPILE_TOOLCHAIN=${TOOLCHAIN_PATH}
+ENV PATH="${TOOLCHAIN_PATH}/bin:${PATH}"
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.arm-linux-rockchip830-armhf-uclibc b/cpp-package/inspireface/docker/Dockerfile.arm-linux-rockchip830-armhf-uclibc
new file mode 100644
index 000000000..8202bd708
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.arm-linux-rockchip830-armhf-uclibc
@@ -0,0 +1,40 @@
+# Use Ubuntu 18.04 as the base image
+FROM ubuntu:18.04
+
+# Update the package list and install basic development tools
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        software-properties-common \
+        wget \
+        curl \
+        git \
+        vim \
+        unzip
+
+# Install CMake
+RUN apt-get install -y --no-install-recommends
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Clone RV1106 toolchain repository
+RUN git clone https://github.com/tunmx/arm-rockchip830-linux-uclibcgnueabihf.git /opt/toolchain
+
+# Set environment variables to point to the toolchain directory
+ENV ARM_CROSS_COMPILE_TOOLCHAIN=/opt/toolchain
+ENV PATH="/opt/toolchain/bin:${PATH}"
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu18 b/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu18
new file mode 100644
index 000000000..597e70779
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu18
@@ -0,0 +1,37 @@
+# Use Ubuntu 18.04 as the base image
+FROM nvidia/cuda:11.1.1-cudnn8-devel-ubuntu20.04
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+RUN apt-get update
+
+
+# Update the package list and install basic development tools
+RUN apt-get install -y --no-install-recommends build-essential
+RUN apt-get install -y --no-install-recommends software-properties-common
+RUN apt-get install -y --no-install-recommends wget
+RUN apt-get install -y --no-install-recommends curl
+RUN apt-get install -y --no-install-recommends git
+RUN apt-get install -y --no-install-recommends vim
+
+# Install CMake
+RUN apt-get install -y --no-install-recommends
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu20 b/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu20
new file mode 100644
index 000000000..247a5fa28
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.cuda.ubuntu20
@@ -0,0 +1,36 @@
+# Use Ubuntu 18.04 as the base image
+FROM nvidia/cuda:11.2.2-devel-ubuntu20.04
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+RUN apt-get update
+
+# Update the package list and install basic development tools
+RUN apt-get install -y --no-install-recommends build-essential
+RUN apt-get install -y --no-install-recommends software-properties-common
+RUN apt-get install -y --no-install-recommends wget
+RUN apt-get install -y --no-install-recommends curl
+RUN apt-get install -y --no-install-recommends git
+RUN apt-get install -y --no-install-recommends vim
+
+# Install CMake
+RUN apt-get install -y --no-install-recommends
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.cuda12_ubuntu22 b/cpp-package/inspireface/docker/Dockerfile.cuda12_ubuntu22
new file mode 100644
index 000000000..9270b52ff
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.cuda12_ubuntu22
@@ -0,0 +1,48 @@
+# Use Ubuntu 18.04 as the base image
+FROM nvidia/cuda:12.0.0-devel-ubuntu22.04
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+RUN apt-get update 
+
+# Update the package list and install basic development tools
+RUN apt-get install -y --no-install-recommends build-essential
+RUN apt-get install -y --no-install-recommends software-properties-common
+RUN apt-get install -y --no-install-recommends curl
+RUN apt-get install -y --no-install-recommends wget
+
+# Install CMake
+RUN apt-get install -y --no-install-recommends
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Clean temporary files to reduce image size
+RUN apt-get clean && \
+    rm -rf /var/lib/apt/lists/* 
+
+# Download TensorRT-10.8
+WORKDIR /
+RUN curl -L https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.8.0/tars/TensorRT-10.8.0.43.Linux.x86_64-gnu.cuda-12.8.tar.gz -o TensorRT.tar.gz && \
+    tar -zxvf TensorRT.tar.gz && \
+    rm TensorRT.tar.gz
+
+# Set environment variables
+RUN echo 'export TENSORRT_ROOT=/TensorRT-10.8.0.43/' >> /root/.bashrc && \
+    echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$TENSORRT_ROOT/lib' >> /root/.bashrc
+
+# Set environment variables
+ENV TENSORRT_ROOT=/TensorRT-10.8.0.43/
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$TENSORRT_ROOT/lib
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
\ No newline at end of file
diff --git a/cpp-package/inspireface/docker/Dockerfile.manylinux2014-x86 b/cpp-package/inspireface/docker/Dockerfile.manylinux2014-x86
new file mode 100644
index 000000000..0e637d40a
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.manylinux2014-x86
@@ -0,0 +1,13 @@
+# Use manylinux2014_x86 as the base image
+FROM quay.io/pypa/manylinux2014_x86_64
+# FROM skybro/ubuntu-cn:18
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.manylinux2014_aarch64 b/cpp-package/inspireface/docker/Dockerfile.manylinux2014_aarch64
new file mode 100644
index 000000000..e6960cd70
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.manylinux2014_aarch64
@@ -0,0 +1,12 @@
+# Use manylinux2014_aarch64 as the base image
+FROM quay.io/pypa/manylinux2014_aarch64
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.manylinux_2_36_aarch64 b/cpp-package/inspireface/docker/Dockerfile.manylinux_2_36_aarch64
new file mode 100644
index 000000000..487dec467
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.manylinux_2_36_aarch64
@@ -0,0 +1,12 @@
+# Use manylinux_2_36_aarch64 as the base image
+FROM quay.io/pypa/manylinux_2_36_aarch64
+
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/docker/Dockerfile.ubuntu18 b/cpp-package/inspireface/docker/Dockerfile.ubuntu18
new file mode 100644
index 000000000..120e1eb40
--- /dev/null
+++ b/cpp-package/inspireface/docker/Dockerfile.ubuntu18
@@ -0,0 +1,34 @@
+# Use Ubuntu 18.04 as the base image
+FROM ubuntu:18.04
+
+# Proxy settings
+ARG https_proxy
+ARG http_proxy
+ARG all_proxy
+
+# Base packages
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        software-properties-common \
+        wget \
+        curl \
+        git \
+        vim \
+        ca-certificates \
+        openssl && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install CMake 3.20.6
+RUN cd /opt && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.6/cmake-3.20.6-linux-x86_64.tar.gz && \
+    tar -zxvf cmake-3.20.6-linux-x86_64.tar.gz && \
+    ln -s /opt/cmake-3.20.6-linux-x86_64/bin/* /usr/local/bin/ && \
+    rm cmake-3.20.6-linux-x86_64.tar.gz
+
+# Set the working directory
+WORKDIR /workspace
+
+# Default to running Bash
+CMD ["/bin/bash"]
diff --git a/cpp-package/inspireface/images/banner.jpg b/cpp-package/inspireface/images/banner.jpg
new file mode 100644
index 000000000..b7ea664fc
Binary files /dev/null and b/cpp-package/inspireface/images/banner.jpg differ
diff --git a/cpp-package/inspireface/images/inspireface-android-example-app-download.png b/cpp-package/inspireface/images/inspireface-android-example-app-download.png
new file mode 100644
index 000000000..3bbc8fa02
Binary files /dev/null and b/cpp-package/inspireface/images/inspireface-android-example-app-download.png differ
diff --git a/cpp-package/inspireface/python/.gitignore b/cpp-package/inspireface/python/.gitignore
new file mode 100644
index 000000000..d5fd46c02
--- /dev/null
+++ b/cpp-package/inspireface/python/.gitignore
@@ -0,0 +1,29 @@
+tmp/
+*.pyc
+*.pyo
+*.pyd
+__pycache__/
+*.so
+*.dylib
+venv/
+env/
+.build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+.pytest_cache/
+.idea/
+*.db
+version.txt
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/README.md b/cpp-package/inspireface/python/README.md
new file mode 100644
index 000000000..124027231
--- /dev/null
+++ b/cpp-package/inspireface/python/README.md
@@ -0,0 +1,103 @@
+# InspireFace Python API
+
+InspireFace provides an easy-to-use Python API that wraps the underlying dynamic link library through ctypes. You can install the latest release version via pip or configure it using the project's self-compiled dynamic library.
+
+## Quick Installation
+
+### Install via pip (Recommended)
+
+```bash
+pip install inspireface
+```
+
+### Manual Installation
+
+1. First install the necessary dependencies:
+```bash
+pip install loguru tqdm opencv-python
+```
+
+2. Copy the compiled dynamic library to the specified directory:
+```bash
+# Copy the compiled dynamic library to the corresponding system architecture directory
+cp YOUR_BUILD_DIR/libInspireFace.so inspireface/modules/core/SYSTEM/CORE_ARCH/
+```
+
+3. Install the Python package:
+```bash
+python setup.py install
+```
+
+## Quick Start
+
+Here's a simple example showing how to use InspireFace for face detection and landmark drawing:
+
+```python
+import cv2
+import inspireface as isf
+
+# Create session with required features enabled
+session = isf.InspireFaceSession(
+    opt=isf.HF_ENABLE_NONE,  # Optional features
+    detect_mode=isf.HF_DETECT_MODE_ALWAYS_DETECT  # Detection mode
+)
+
+# Set detection confidence threshold
+session.set_detection_confidence_threshold(0.5)
+
+# Read image
+image = cv2.imread("path/to/your/image.jpg")
+assert image is not None, "Please check if the image path is correct"
+
+# Perform face detection
+faces = session.face_detection(image)
+print(f"Detected {len(faces)} faces")
+
+# Draw detection results on image
+draw = image.copy()
+for idx, face in enumerate(faces):
+    # Get face bounding box coordinates
+    x1, y1, x2, y2 = face.location
+    
+    # Calculate rotated box parameters
+    center = ((x1 + x2) / 2, (y1 + y2) / 2)
+    size = (x2 - x1, y2 - y1)
+    angle = face.roll
+    
+    # Draw rotated box
+    rect = ((center[0], center[1]), (size[0], size[1]), angle)
+    box = cv2.boxPoints(rect)
+    box = box.astype(int)
+    cv2.drawContours(draw, [box], 0, (100, 180, 29), 2)
+    
+    # Draw landmarks
+    landmarks = session.get_face_dense_landmark(face)
+    for x, y in landmarks.astype(int):
+        cv2.circle(draw, (x, y), 0, (220, 100, 0), 2)
+```
+
+## More Examples
+
+The project provides multiple example files demonstrating different features:
+
+- `sample_face_detection.py`: Basic face detection
+- `sample_face_track_from_video.py`: Video face tracking
+- `sample_face_recognition.py`: Face recognition
+- `sample_face_comparison.py`: Face comparison
+- `sample_feature_hub.py`: Feature extraction
+- `sample_system_resource_statistics.py`: System resource statistics
+
+## Running Tests
+
+The project includes unit tests. You can adjust test content by modifying parameters in `test/test_settings.py`:
+
+```bash
+python -m unittest discover -s test
+```
+
+## Notes
+
+1. Ensure that OpenCV and other necessary dependencies are installed on your system
+2. Make sure the dynamic library is correctly installed before use
+3. Python 3.7 or higher is recommended
+4. The default version is CPU, if you want to use the GPU, CoreML, or NPU backend version, you can refer to the [documentation](https://doc.inspireface.online/guides/python-rockchip-device.html) to replace the so and make a Python installation package
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/inspireface/__init__.py b/cpp-package/inspireface/python/inspireface/__init__.py
new file mode 100644
index 000000000..9c140c054
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/__init__.py
@@ -0,0 +1,5 @@
+from .modules import *
+from .param import *
+
+__version__ = version()
+
diff --git a/cpp-package/inspireface/python/inspireface/modules/__init__.py b/cpp-package/inspireface/python/inspireface/modules/__init__.py
new file mode 100644
index 000000000..b37715997
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/__init__.py
@@ -0,0 +1,11 @@
+from .inspireface import ImageStream, FaceExtended, FaceInformation, SessionCustomParameter, InspireFaceSession, \
+    launch, terminate, FeatureHubConfiguration, feature_hub_enable, feature_hub_disable, feature_comparison, \
+    FaceIdentity, feature_hub_set_search_threshold, feature_hub_face_insert, SearchResult, \
+    feature_hub_face_search, feature_hub_face_search_top_k, feature_hub_face_update, feature_hub_face_remove, \
+    feature_hub_get_face_identity, feature_hub_get_face_count, feature_hub_get_face_id_list, view_table_in_terminal, version, query_launch_status, reload, \
+    set_logging_level, disable_logging, show_system_resource_statistics, get_recommended_cosine_threshold, cosine_similarity_convert_to_percentage, \
+    get_similarity_converter_config, set_similarity_converter_config, pull_latest_model, switch_landmark_engine, \
+    HF_PK_AUTO_INCREMENT, HF_PK_MANUAL_INPUT, HF_SEARCH_MODE_EAGER, HF_SEARCH_MODE_EXHAUSTIVE, \
+    ignore_check_latest_model, set_cuda_device_id, get_cuda_device_id, print_cuda_device_info, get_num_cuda_devices, check_cuda_device_support, terminate, \
+    InspireFaceError, InvalidInputError, SystemNotReadyError, ProcessingError, ResourceError, HardwareError, FeatureHubError, \
+    switch_image_processing_backend, HF_IMAGE_PROCESSING_CPU, HF_IMAGE_PROCESSING_RGA, set_image_process_aligned_width, use_oss_download
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/inspireface/modules/core/__init__.py b/cpp-package/inspireface/python/inspireface/modules/core/__init__.py
new file mode 100644
index 000000000..f1ac2f346
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/core/__init__.py
@@ -0,0 +1 @@
+from .native import *
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/inspireface/modules/core/native.py b/cpp-package/inspireface/python/inspireface/modules/core/native.py
new file mode 100644
index 000000000..1c4359056
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/core/native.py
@@ -0,0 +1,2349 @@
+__docformat__ = "restructuredtext"
+
+# Begin preamble for Python
+
+import ctypes
+import sys
+from ctypes import *  # noqa: F401, F403
+import platform
+from pathlib import Path
+import subprocess
+import os
+
+def get_lib_path():
+    """
+    Get the appropriate library path based on the current platform and architecture.
+    
+    Returns:
+        str: The path to the platform-specific library
+        
+    Raises:
+        RuntimeError: If the platform/architecture is unsupported or library not found
+    """
+    package_dir = Path(__file__).parent
+    
+    # Get basic platform information
+    system = platform.system().lower()
+    machine = platform.machine().lower()
+    
+    # Initialize variables
+    platform_dir = None
+    lib_name = None
+    arch = None
+    
+    if system == 'windows':
+        platform_dir = 'windows'
+        lib_name = 'libInspireFace.dll'
+        # Windows architecture detection
+        arch = 'x64' if machine == 'amd64' or machine == 'x86_64' else 'arm64'
+        
+    elif system == 'linux':
+        platform_dir = 'linux'
+        lib_name = 'libInspireFace.so'
+        # Linux architecture detection
+        if machine == 'x86_64':
+            arch = 'x64'
+        elif machine in ['aarch64', 'arm64']:
+            arch = 'arm64'
+        elif machine.startswith('arm'):
+            arch = 'arm64'  # Might need more specific ARM version distinction
+            
+    elif system == 'darwin':  # macOS
+        platform_dir = 'darwin'
+        lib_name = 'libInspireFace.dylib'
+        
+        # macOS architecture detection
+        if machine == 'x86_64':
+            # Check if running under Rosetta 2
+            try:
+                # Use sysctl to detect Rosetta 2
+                is_rosetta = bool(int(subprocess.check_output(
+                    ['sysctl', '-n', 'sysctl.proc_translated']).decode().strip()))
+                # If running under Rosetta, it's actually an ARM machine
+                if is_rosetta:
+                    arch = 'arm64'
+                else:
+                    arch = 'x64'
+            except:
+                # If detection fails, assume native x64
+                arch = 'x64'
+        elif machine == 'arm64':
+            arch = 'arm64'
+            
+    # Validate that all necessary parameters were set
+    if not all([platform_dir, lib_name, arch]):
+        raise RuntimeError(
+            f"Unsupported platform: system={system}, machine={machine}")
+    
+    # Construct the full library path
+    dir_path = package_dir / 'libs' / platform_dir / arch
+    os.makedirs(dir_path, exist_ok=True)
+    lib_path = dir_path / lib_name
+    
+    # Verify that the library file exists
+    if not lib_path.exists():
+        raise RuntimeError(
+            f"Library not found at {lib_path}. "
+            f"System: {system}, Architecture: {arch}")
+    
+    return str(lib_path)
+
+try:    
+    _LIBRARY_FILENAME = get_lib_path()
+except Exception as e:
+    print(e)
+
+_int_types = (ctypes.c_int16, ctypes.c_int32)
+if hasattr(ctypes, "c_int64"):
+    # Some builds of ctypes apparently do not have ctypes.c_int64
+    # defined; it's a pretty good bet that these builds do not
+    # have 64-bit pointers.
+    _int_types += (ctypes.c_int64,)
+for t in _int_types:
+    if ctypes.sizeof(t) == ctypes.sizeof(ctypes.c_size_t):
+        c_ptrdiff_t = t
+del t
+del _int_types
+
+
+
+class UserString:
+    def __init__(self, seq):
+        if isinstance(seq, bytes):
+            self.data = seq
+        elif isinstance(seq, UserString):
+            self.data = seq.data[:]
+        else:
+            self.data = str(seq).encode()
+
+    def __bytes__(self):
+        return self.data
+
+    def __str__(self):
+        return self.data.decode()
+
+    def __repr__(self):
+        return repr(self.data)
+
+    def __int__(self):
+        return int(self.data.decode())
+
+    def __long__(self):
+        return int(self.data.decode())
+
+    def __float__(self):
+        return float(self.data.decode())
+
+    def __complex__(self):
+        return complex(self.data.decode())
+
+    def __hash__(self):
+        return hash(self.data)
+
+    def __le__(self, string):
+        if isinstance(string, UserString):
+            return self.data <= string.data
+        else:
+            return self.data <= string
+
+    def __lt__(self, string):
+        if isinstance(string, UserString):
+            return self.data < string.data
+        else:
+            return self.data < string
+
+    def __ge__(self, string):
+        if isinstance(string, UserString):
+            return self.data >= string.data
+        else:
+            return self.data >= string
+
+    def __gt__(self, string):
+        if isinstance(string, UserString):
+            return self.data > string.data
+        else:
+            return self.data > string
+
+    def __eq__(self, string):
+        if isinstance(string, UserString):
+            return self.data == string.data
+        else:
+            return self.data == string
+
+    def __ne__(self, string):
+        if isinstance(string, UserString):
+            return self.data != string.data
+        else:
+            return self.data != string
+
+    def __contains__(self, char):
+        return char in self.data
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, index):
+        return self.__class__(self.data[index])
+
+    def __getslice__(self, start, end):
+        start = max(start, 0)
+        end = max(end, 0)
+        return self.__class__(self.data[start:end])
+
+    def __add__(self, other):
+        if isinstance(other, UserString):
+            return self.__class__(self.data + other.data)
+        elif isinstance(other, bytes):
+            return self.__class__(self.data + other)
+        else:
+            return self.__class__(self.data + str(other).encode())
+
+    def __radd__(self, other):
+        if isinstance(other, bytes):
+            return self.__class__(other + self.data)
+        else:
+            return self.__class__(str(other).encode() + self.data)
+
+    def __mul__(self, n):
+        return self.__class__(self.data * n)
+
+    __rmul__ = __mul__
+
+    def __mod__(self, args):
+        return self.__class__(self.data % args)
+
+    # the following methods are defined in alphabetical order:
+    def capitalize(self):
+        return self.__class__(self.data.capitalize())
+
+    def center(self, width, *args):
+        return self.__class__(self.data.center(width, *args))
+
+    def count(self, sub, start=0, end=sys.maxsize):
+        return self.data.count(sub, start, end)
+
+    def decode(self, encoding=None, errors=None):  # XXX improve this?
+        if encoding:
+            if errors:
+                return self.__class__(self.data.decode(encoding, errors))
+            else:
+                return self.__class__(self.data.decode(encoding))
+        else:
+            return self.__class__(self.data.decode())
+
+    def encode(self, encoding=None, errors=None):  # XXX improve this?
+        if encoding:
+            if errors:
+                return self.__class__(self.data.encode(encoding, errors))
+            else:
+                return self.__class__(self.data.encode(encoding))
+        else:
+            return self.__class__(self.data.encode())
+
+    def endswith(self, suffix, start=0, end=sys.maxsize):
+        return self.data.endswith(suffix, start, end)
+
+    def expandtabs(self, tabsize=8):
+        return self.__class__(self.data.expandtabs(tabsize))
+
+    def find(self, sub, start=0, end=sys.maxsize):
+        return self.data.find(sub, start, end)
+
+    def index(self, sub, start=0, end=sys.maxsize):
+        return self.data.index(sub, start, end)
+
+    def isalpha(self):
+        return self.data.isalpha()
+
+    def isalnum(self):
+        return self.data.isalnum()
+
+    def isdecimal(self):
+        return self.data.isdecimal()
+
+    def isdigit(self):
+        return self.data.isdigit()
+
+    def islower(self):
+        return self.data.islower()
+
+    def isnumeric(self):
+        return self.data.isnumeric()
+
+    def isspace(self):
+        return self.data.isspace()
+
+    def istitle(self):
+        return self.data.istitle()
+
+    def isupper(self):
+        return self.data.isupper()
+
+    def join(self, seq):
+        return self.data.join(seq)
+
+    def ljust(self, width, *args):
+        return self.__class__(self.data.ljust(width, *args))
+
+    def lower(self):
+        return self.__class__(self.data.lower())
+
+    def lstrip(self, chars=None):
+        return self.__class__(self.data.lstrip(chars))
+
+    def partition(self, sep):
+        return self.data.partition(sep)
+
+    def replace(self, old, new, maxsplit=-1):
+        return self.__class__(self.data.replace(old, new, maxsplit))
+
+    def rfind(self, sub, start=0, end=sys.maxsize):
+        return self.data.rfind(sub, start, end)
+
+    def rindex(self, sub, start=0, end=sys.maxsize):
+        return self.data.rindex(sub, start, end)
+
+    def rjust(self, width, *args):
+        return self.__class__(self.data.rjust(width, *args))
+
+    def rpartition(self, sep):
+        return self.data.rpartition(sep)
+
+    def rstrip(self, chars=None):
+        return self.__class__(self.data.rstrip(chars))
+
+    def split(self, sep=None, maxsplit=-1):
+        return self.data.split(sep, maxsplit)
+
+    def rsplit(self, sep=None, maxsplit=-1):
+        return self.data.rsplit(sep, maxsplit)
+
+    def splitlines(self, keepends=0):
+        return self.data.splitlines(keepends)
+
+    def startswith(self, prefix, start=0, end=sys.maxsize):
+        return self.data.startswith(prefix, start, end)
+
+    def strip(self, chars=None):
+        return self.__class__(self.data.strip(chars))
+
+    def swapcase(self):
+        return self.__class__(self.data.swapcase())
+
+    def title(self):
+        return self.__class__(self.data.title())
+
+    def translate(self, *args):
+        return self.__class__(self.data.translate(*args))
+
+    def upper(self):
+        return self.__class__(self.data.upper())
+
+    def zfill(self, width):
+        return self.__class__(self.data.zfill(width))
+
+
+class MutableString(UserString):
+    """mutable string objects
+
+    Python strings are immutable objects.  This has the advantage, that
+    strings may be used as dictionary keys.  If this property isn't needed
+    and you insist on changing string values in place instead, you may cheat
+    and use MutableString.
+
+    But the purpose of this class is an educational one: to prevent
+    people from inventing their own mutable string class derived
+    from UserString and than forget thereby to remove (override) the
+    __hash__ method inherited from UserString.  This would lead to
+    errors that would be very hard to track down.
+
+    A faster and better solution is to rewrite your program using lists."""
+
+    def __init__(self, string=""):
+        self.data = string
+
+    def __hash__(self):
+        raise TypeError("unhashable type (it is mutable)")
+
+    def __setitem__(self, index, sub):
+        if index < 0:
+            index += len(self.data)
+        if index < 0 or index >= len(self.data):
+            raise IndexError
+        self.data = self.data[:index] + sub + self.data[index + 1 :]
+
+    def __delitem__(self, index):
+        if index < 0:
+            index += len(self.data)
+        if index < 0 or index >= len(self.data):
+            raise IndexError
+        self.data = self.data[:index] + self.data[index + 1 :]
+
+    def __setslice__(self, start, end, sub):
+        start = max(start, 0)
+        end = max(end, 0)
+        if isinstance(sub, UserString):
+            self.data = self.data[:start] + sub.data + self.data[end:]
+        elif isinstance(sub, bytes):
+            self.data = self.data[:start] + sub + self.data[end:]
+        else:
+            self.data = self.data[:start] + str(sub).encode() + self.data[end:]
+
+    def __delslice__(self, start, end):
+        start = max(start, 0)
+        end = max(end, 0)
+        self.data = self.data[:start] + self.data[end:]
+
+    def immutable(self):
+        return UserString(self.data)
+
+    def __iadd__(self, other):
+        if isinstance(other, UserString):
+            self.data += other.data
+        elif isinstance(other, bytes):
+            self.data += other
+        else:
+            self.data += str(other).encode()
+        return self
+
+    def __imul__(self, n):
+        self.data *= n
+        return self
+
+
+class String(MutableString, ctypes.Union):
+
+    _fields_ = [("raw", ctypes.POINTER(ctypes.c_char)), ("data", ctypes.c_char_p)]
+
+    def __init__(self, obj=b""):
+        if isinstance(obj, (bytes, UserString)):
+            self.data = bytes(obj)
+        else:
+            self.raw = obj
+
+    def __len__(self):
+        return self.data and len(self.data) or 0
+
+    def from_param(cls, obj):
+        # Convert None or 0
+        if obj is None or obj == 0:
+            return cls(ctypes.POINTER(ctypes.c_char)())
+
+        # Convert from String
+        elif isinstance(obj, String):
+            return obj
+
+        # Convert from bytes
+        elif isinstance(obj, bytes):
+            return cls(obj)
+
+        # Convert from str
+        elif isinstance(obj, str):
+            return cls(obj.encode())
+
+        # Convert from c_char_p
+        elif isinstance(obj, ctypes.c_char_p):
+            return obj
+
+        # Convert from POINTER(ctypes.c_char)
+        elif isinstance(obj, ctypes.POINTER(ctypes.c_char)):
+            return obj
+
+        # Convert from raw pointer
+        elif isinstance(obj, int):
+            return cls(ctypes.cast(obj, ctypes.POINTER(ctypes.c_char)))
+
+        # Convert from ctypes.c_char array
+        elif isinstance(obj, ctypes.c_char * len(obj)):
+            return obj
+
+        # Convert from object
+        else:
+            return String.from_param(obj._as_parameter_)
+
+    from_param = classmethod(from_param)
+
+
+def ReturnString(obj, func=None, arguments=None):
+    return String.from_param(obj)
+
+
+# As of ctypes 1.0, ctypes does not support custom error-checking
+# functions on callbacks, nor does it support custom datatypes on
+# callbacks, so we must ensure that all callbacks return
+# primitive datatypes.
+#
+# Non-primitive return values wrapped with UNCHECKED won't be
+# typechecked, and will be converted to ctypes.c_void_p.
+def UNCHECKED(type):
+    if hasattr(type, "_type_") and isinstance(type._type_, str) and type._type_ != "P":
+        return type
+    else:
+        return ctypes.c_void_p
+
+
+# ctypes doesn't have direct support for variadic functions, so we have to write
+# our own wrapper class
+class _variadic_function(object):
+    def __init__(self, func, restype, argtypes, errcheck):
+        self.func = func
+        self.func.restype = restype
+        self.argtypes = argtypes
+        if errcheck:
+            self.func.errcheck = errcheck
+
+    def _as_parameter_(self):
+        # So we can pass this variadic function as a function pointer
+        return self.func
+
+    def __call__(self, *args):
+        fixed_args = []
+        i = 0
+        for argtype in self.argtypes:
+            # Typecheck what we can
+            fixed_args.append(argtype.from_param(args[i]))
+            i += 1
+        return self.func(*fixed_args + list(args[i:]))
+
+
+def ord_if_char(value):
+    """
+    Simple helper used for casts to simple builtin types:  if the argument is a
+    string type, it will be converted to it's ordinal value.
+
+    This function will raise an exception if the argument is string with more
+    than one characters.
+    """
+    return ord(value) if (isinstance(value, bytes) or isinstance(value, str)) else value
+
+# End preamble
+
+_libs = {}
+_libdirs = []
+
+# Begin loader
+
+"""
+Load libraries - appropriately for all our supported platforms
+"""
+# ----------------------------------------------------------------------------
+# Copyright (c) 2008 David James
+# Copyright (c) 2006-2008 Alex Holkner
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in
+#    the documentation and/or other materials provided with the
+#    distribution.
+#  * Neither the name of pyglet nor the names of its
+#    contributors may be used to endorse or promote products
+#    derived from this software without specific prior written
+#    permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+# FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+# COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+# INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+# BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+# LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+# ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+# POSSIBILITY OF SUCH DAMAGE.
+# ----------------------------------------------------------------------------
+
+import ctypes
+import ctypes.util
+import glob
+import os.path
+import platform
+import re
+import sys
+
+
+def _environ_path(name):
+    """Split an environment variable into a path-like list elements"""
+    if name in os.environ:
+        return os.environ[name].split(":")
+    return []
+
+
+class LibraryLoader:
+    """
+    A base class For loading of libraries ;-)
+    Subclasses load libraries for specific platforms.
+    """
+
+    # library names formatted specifically for platforms
+    name_formats = ["%s"]
+
+    class Lookup:
+        """Looking up calling conventions for a platform"""
+
+        mode = ctypes.DEFAULT_MODE
+
+        def __init__(self, path):
+            super(LibraryLoader.Lookup, self).__init__()
+            self.access = dict(cdecl=ctypes.CDLL(path, self.mode))
+
+        def get(self, name, calling_convention="cdecl"):
+            """Return the given name according to the selected calling convention"""
+            if calling_convention not in self.access:
+                raise LookupError(
+                    "Unknown calling convention '{}' for function '{}'".format(
+                        calling_convention, name
+                    )
+                )
+            return getattr(self.access[calling_convention], name)
+
+        def has(self, name, calling_convention="cdecl"):
+            """Return True if this given calling convention finds the given 'name'"""
+            if calling_convention not in self.access:
+                return False
+            return hasattr(self.access[calling_convention], name)
+
+        def __getattr__(self, name):
+            return getattr(self.access["cdecl"], name)
+
+    def __init__(self):
+        self.other_dirs = []
+
+    def __call__(self, libname):
+        """Given the name of a library, load it."""
+        paths = self.getpaths(libname)
+
+        for path in paths:
+            # noinspection PyBroadException
+            try:
+                return self.Lookup(path)
+            except Exception as err:  # pylint: disable=broad-except
+                print(err)
+
+        raise ImportError("Could not load %s." % libname)
+
+    def getpaths(self, libname):
+        """Return a list of paths where the library might be found."""
+        if os.path.isabs(libname):
+            yield libname
+        else:
+            # search through a prioritized series of locations for the library
+
+            # we first search any specific directories identified by user
+            for dir_i in self.other_dirs:
+                for fmt in self.name_formats:
+                    # dir_i should be absolute already
+                    yield os.path.join(dir_i, fmt % libname)
+
+            # check if this code is even stored in a physical file
+            try:
+                this_file = __file__
+            except NameError:
+                this_file = None
+
+            # then we search the directory where the generated python interface is stored
+            if this_file is not None:
+                for fmt in self.name_formats:
+                    yield os.path.abspath(os.path.join(os.path.dirname(__file__), fmt % libname))
+
+            # now, use the ctypes tools to try to find the library
+            for fmt in self.name_formats:
+                path = ctypes.util.find_library(fmt % libname)
+                if path:
+                    yield path
+
+            # then we search all paths identified as platform-specific lib paths
+            for path in self.getplatformpaths(libname):
+                yield path
+
+            # Finally, we'll try the users current working directory
+            for fmt in self.name_formats:
+                yield os.path.abspath(os.path.join(os.path.curdir, fmt % libname))
+
+    def getplatformpaths(self, _libname):  # pylint: disable=no-self-use
+        """Return all the library paths available in this platform"""
+        return []
+
+
+# Darwin (Mac OS X)
+
+
+class DarwinLibraryLoader(LibraryLoader):
+    """Library loader for MacOS"""
+
+    name_formats = [
+        "lib%s.dylib",
+        "lib%s.so",
+        "lib%s.bundle",
+        "%s.dylib",
+        "%s.so",
+        "%s.bundle",
+        "%s",
+    ]
+
+    class Lookup(LibraryLoader.Lookup):
+        """
+        Looking up library files for this platform (Darwin aka MacOS)
+        """
+
+        # Darwin requires dlopen to be called with mode RTLD_GLOBAL instead
+        # of the default RTLD_LOCAL.  Without this, you end up with
+        # libraries not being loadable, resulting in "Symbol not found"
+        # errors
+        mode = ctypes.RTLD_GLOBAL
+
+    def getplatformpaths(self, libname):
+        if os.path.pathsep in libname:
+            names = [libname]
+        else:
+            names = [fmt % libname for fmt in self.name_formats]
+
+        for directory in self.getdirs(libname):
+            for name in names:
+                yield os.path.join(directory, name)
+
+    @staticmethod
+    def getdirs(libname):
+        """Implements the dylib search as specified in Apple documentation:
+
+        http://developer.apple.com/documentation/DeveloperTools/Conceptual/
+            DynamicLibraries/Articles/DynamicLibraryUsageGuidelines.html
+
+        Before commencing the standard search, the method first checks
+        the bundle's ``Frameworks`` directory if the application is running
+        within a bundle (OS X .app).
+        """
+
+        dyld_fallback_library_path = _environ_path("DYLD_FALLBACK_LIBRARY_PATH")
+        if not dyld_fallback_library_path:
+            dyld_fallback_library_path = [
+                os.path.expanduser("~/lib"),
+                "/usr/local/lib",
+                "/usr/lib",
+            ]
+
+        dirs = []
+
+        if "/" in libname:
+            dirs.extend(_environ_path("DYLD_LIBRARY_PATH"))
+        else:
+            dirs.extend(_environ_path("LD_LIBRARY_PATH"))
+            dirs.extend(_environ_path("DYLD_LIBRARY_PATH"))
+            dirs.extend(_environ_path("LD_RUN_PATH"))
+
+        if hasattr(sys, "frozen") and getattr(sys, "frozen") == "macosx_app":
+            dirs.append(os.path.join(os.environ["RESOURCEPATH"], "../../..", "Frameworks"))
+
+        dirs.extend(dyld_fallback_library_path)
+
+        return dirs
+
+
+# Posix
+
+
+class PosixLibraryLoader(LibraryLoader):
+    """Library loader for POSIX-like systems (including Linux)"""
+
+    _ld_so_cache = None
+
+    _include = re.compile(r"^\s*include\s+(?P<pattern>.*)")
+
+    name_formats = ["lib%s.so", "%s.so", "%s"]
+
+    class _Directories(dict):
+        """Deal with directories"""
+
+        def __init__(self):
+            dict.__init__(self)
+            self.order = 0
+
+        def add(self, directory):
+            """Add a directory to our current set of directories"""
+            if len(directory) > 1:
+                directory = directory.rstrip(os.path.sep)
+            # only adds and updates order if exists and not already in set
+            if not os.path.exists(directory):
+                return
+            order = self.setdefault(directory, self.order)
+            if order == self.order:
+                self.order += 1
+
+        def extend(self, directories):
+            """Add a list of directories to our set"""
+            for a_dir in directories:
+                self.add(a_dir)
+
+        def ordered(self):
+            """Sort the list of directories"""
+            return (i[0] for i in sorted(self.items(), key=lambda d: d[1]))
+
+    def _get_ld_so_conf_dirs(self, conf, dirs):
+        """
+        Recursive function to help parse all ld.so.conf files, including proper
+        handling of the `include` directive.
+        """
+
+        try:
+            with open(conf) as fileobj:
+                for dirname in fileobj:
+                    dirname = dirname.strip()
+                    if not dirname:
+                        continue
+
+                    match = self._include.match(dirname)
+                    if not match:
+                        dirs.add(dirname)
+                    else:
+                        for dir2 in glob.glob(match.group("pattern")):
+                            self._get_ld_so_conf_dirs(dir2, dirs)
+        except IOError:
+            pass
+
+    def _create_ld_so_cache(self):
+        # Recreate search path followed by ld.so.  This is going to be
+        # slow to build, and incorrect (ld.so uses ld.so.cache, which may
+        # not be up-to-date).  Used only as fallback for distros without
+        # /sbin/ldconfig.
+        #
+        # We assume the DT_RPATH and DT_RUNPATH binary sections are omitted.
+
+        directories = self._Directories()
+        for name in (
+            "LD_LIBRARY_PATH",
+            "SHLIB_PATH",  # HP-UX
+            "LIBPATH",  # OS/2, AIX
+            "LIBRARY_PATH",  # BE/OS
+        ):
+            if name in os.environ:
+                directories.extend(os.environ[name].split(os.pathsep))
+
+        self._get_ld_so_conf_dirs("/etc/ld.so.conf", directories)
+
+        bitage = platform.architecture()[0]
+
+        unix_lib_dirs_list = []
+        if bitage.startswith("64"):
+            # prefer 64 bit if that is our arch
+            unix_lib_dirs_list += ["/lib64", "/usr/lib64"]
+
+        # must include standard libs, since those paths are also used by 64 bit
+        # installs
+        unix_lib_dirs_list += ["/lib", "/usr/lib"]
+        if sys.platform.startswith("linux"):
+            # Try and support multiarch work in Ubuntu
+            # https://wiki.ubuntu.com/MultiarchSpec
+            if bitage.startswith("32"):
+                # Assume Intel/AMD x86 compat
+                unix_lib_dirs_list += ["/lib/i386-linux-gnu", "/usr/lib/i386-linux-gnu"]
+            elif bitage.startswith("64"):
+                # Assume Intel/AMD x86 compatible
+                unix_lib_dirs_list += [
+                    "/lib/x86_64-linux-gnu",
+                    "/usr/lib/x86_64-linux-gnu",
+                ]
+            else:
+                # guess...
+                unix_lib_dirs_list += glob.glob("/lib/*linux-gnu")
+        directories.extend(unix_lib_dirs_list)
+
+        cache = {}
+        lib_re = re.compile(r"lib(.*)\.s[ol]")
+        # ext_re = re.compile(r"\.s[ol]$")
+        for our_dir in directories.ordered():
+            try:
+                for path in glob.glob("%s/*.s[ol]*" % our_dir):
+                    file = os.path.basename(path)
+
+                    # Index by filename
+                    cache_i = cache.setdefault(file, set())
+                    cache_i.add(path)
+
+                    # Index by library name
+                    match = lib_re.match(file)
+                    if match:
+                        library = match.group(1)
+                        cache_i = cache.setdefault(library, set())
+                        cache_i.add(path)
+            except OSError:
+                pass
+
+        self._ld_so_cache = cache
+
+    def getplatformpaths(self, libname):
+        if self._ld_so_cache is None:
+            self._create_ld_so_cache()
+
+        result = self._ld_so_cache.get(libname, set())
+        for i in result:
+            # we iterate through all found paths for library, since we may have
+            # actually found multiple architectures or other library types that
+            # may not load
+            yield i
+
+
+# Windows
+
+
+class WindowsLibraryLoader(LibraryLoader):
+    """Library loader for Microsoft Windows"""
+
+    name_formats = ["%s.dll", "lib%s.dll", "%slib.dll", "%s"]
+
+    class Lookup(LibraryLoader.Lookup):
+        """Lookup class for Windows libraries..."""
+
+        def __init__(self, path):
+            super(WindowsLibraryLoader.Lookup, self).__init__(path)
+            self.access["stdcall"] = ctypes.windll.LoadLibrary(path)
+
+
+# Platform switching
+
+# If your value of sys.platform does not appear in this dict, please contact
+# the Ctypesgen maintainers.
+
+loaderclass = {
+    "darwin": DarwinLibraryLoader,
+    "cygwin": WindowsLibraryLoader,
+    "win32": WindowsLibraryLoader,
+    "msys": WindowsLibraryLoader,
+}
+
+load_library = loaderclass.get(sys.platform, PosixLibraryLoader)()
+
+
+def add_library_search_dirs(other_dirs):
+    """
+    Add libraries to search paths.
+    If library paths are relative, convert them to absolute with respect to this
+    file's directory
+    """
+    for path in other_dirs:
+        if not os.path.isabs(path):
+            path = os.path.abspath(path)
+        load_library.other_dirs.append(path)
+
+
+del loaderclass
+
+# End loader
+
+add_library_search_dirs([])
+
+# Begin libraries
+_libs[_LIBRARY_FILENAME] = load_library(_LIBRARY_FILENAME)
+# 1 libraries
+# End libraries
+
+# No modules
+
+HFImageStream = POINTER(None)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 11
+
+PHFImageStream = POINTER(POINTER(None))# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 12
+
+HFSession = POINTER(None)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 13
+
+PHFSession = POINTER(POINTER(None))# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 14
+
+HFImageBitmap = POINTER(None)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 15
+
+PHFImageBitmap = POINTER(POINTER(None))# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 16
+
+HPVoid = POINTER(None)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 17
+
+HFloat = c_float# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 19
+
+HPFloat = POINTER(c_float)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 20
+
+HPUInt8 = POINTER(c_ubyte)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 23
+
+HInt32 = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 24
+
+HOption = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 25
+
+HPInt32 = POINTER(c_int)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 26
+
+HFaceId = c_int64# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 27
+
+HPFaceId = POINTER(c_int64)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 28
+
+HResult = c_long# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 29
+
+HString = String# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 30
+
+HPath = String# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 31
+
+HFormat = String# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 32
+
+HChar = c_char# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 34
+
+HPBuffer = String# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 35
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 45
+class struct_HFaceRect(Structure):
+    pass
+
+struct_HFaceRect.__slots__ = [
+    'x',
+    'y',
+    'width',
+    'height',
+]
+struct_HFaceRect._fields_ = [
+    ('x', HInt32),
+    ('y', HInt32),
+    ('width', HInt32),
+    ('height', HInt32),
+]
+
+HFaceRect = struct_HFaceRect# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 45
+
+PHFaceRect = POINTER(struct_HFaceRect)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 45
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 50
+class struct_HPoint2f(Structure):
+    pass
+
+struct_HPoint2f.__slots__ = [
+    'x',
+    'y',
+]
+struct_HPoint2f._fields_ = [
+    ('x', HFloat),
+    ('y', HFloat),
+]
+
+HPoint2f = struct_HPoint2f# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 50
+
+PHPoint2f = POINTER(struct_HPoint2f)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 50
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 55
+class struct_HPoint2i(Structure):
+    pass
+
+struct_HPoint2i.__slots__ = [
+    'x',
+    'y',
+]
+struct_HPoint2i._fields_ = [
+    ('x', HInt32),
+    ('y', HInt32),
+]
+
+HPoint2i = struct_HPoint2i# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 55
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 61
+class struct_HColor(Structure):
+    pass
+
+struct_HColor.__slots__ = [
+    'r',
+    'g',
+    'b',
+]
+struct_HColor._fields_ = [
+    ('r', HFloat),
+    ('g', HFloat),
+    ('b', HFloat),
+]
+
+HColor = struct_HColor# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/intypedef.h: 61
+
+enum_HFImageFormat = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_RGB = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_BGR = 1# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_RGBA = 2# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_BGRA = 3# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_YUV_NV12 = 4# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_YUV_NV21 = 5# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_I420 = 6# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HF_STREAM_GRAY = 7# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+HFImageFormat = enum_HFImageFormat# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 100
+
+enum_HFRotation = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+HF_CAMERA_ROTATION_0 = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+HF_CAMERA_ROTATION_90 = 1# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+HF_CAMERA_ROTATION_180 = 2# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+HF_CAMERA_ROTATION_270 = 3# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+HFRotation = enum_HFRotation# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 111
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 123
+class struct_HFImageData(Structure):
+    pass
+
+struct_HFImageData.__slots__ = [
+    'data',
+    'width',
+    'height',
+    'format',
+    'rotation',
+]
+struct_HFImageData._fields_ = [
+    ('data', HPUInt8),
+    ('width', HInt32),
+    ('height', HInt32),
+    ('format', HFImageFormat),
+    ('rotation', HFRotation),
+]
+
+HFImageData = struct_HFImageData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 123
+
+PHFImageData = POINTER(struct_HFImageData)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 123
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 134
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageStream", "cdecl"):
+    HFCreateImageStream = _libs[_LIBRARY_FILENAME].get("HFCreateImageStream", "cdecl")
+    HFCreateImageStream.argtypes = [PHFImageData, PHFImageStream]
+    HFCreateImageStream.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 144
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageStreamEmpty", "cdecl"):
+    HFCreateImageStreamEmpty = _libs[_LIBRARY_FILENAME].get("HFCreateImageStreamEmpty", "cdecl")
+    HFCreateImageStreamEmpty.argtypes = [PHFImageStream]
+    HFCreateImageStreamEmpty.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 155
+if _libs[_LIBRARY_FILENAME].has("HFImageStreamSetBuffer", "cdecl"):
+    HFImageStreamSetBuffer = _libs[_LIBRARY_FILENAME].get("HFImageStreamSetBuffer", "cdecl")
+    HFImageStreamSetBuffer.argtypes = [HFImageStream, HPUInt8, HInt32, HInt32]
+    HFImageStreamSetBuffer.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 164
+if _libs[_LIBRARY_FILENAME].has("HFImageStreamSetRotation", "cdecl"):
+    HFImageStreamSetRotation = _libs[_LIBRARY_FILENAME].get("HFImageStreamSetRotation", "cdecl")
+    HFImageStreamSetRotation.argtypes = [HFImageStream, HFRotation]
+    HFImageStreamSetRotation.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 173
+if _libs[_LIBRARY_FILENAME].has("HFImageStreamSetFormat", "cdecl"):
+    HFImageStreamSetFormat = _libs[_LIBRARY_FILENAME].get("HFImageStreamSetFormat", "cdecl")
+    HFImageStreamSetFormat.argtypes = [HFImageStream, HFImageFormat]
+    HFImageStreamSetFormat.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 183
+if _libs[_LIBRARY_FILENAME].has("HFReleaseImageStream", "cdecl"):
+    HFReleaseImageStream = _libs[_LIBRARY_FILENAME].get("HFReleaseImageStream", "cdecl")
+    HFReleaseImageStream.argtypes = [HFImageStream]
+    HFReleaseImageStream.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 200
+class struct_HFImageBitmapData(Structure):
+    pass
+
+struct_HFImageBitmapData.__slots__ = [
+    'data',
+    'width',
+    'height',
+    'channels',
+]
+struct_HFImageBitmapData._fields_ = [
+    ('data', HPUInt8),
+    ('width', HInt32),
+    ('height', HInt32),
+    ('channels', HInt32),
+]
+
+HFImageBitmapData = struct_HFImageBitmapData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 200
+
+PHFImageBitmapData = POINTER(struct_HFImageBitmapData)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 200
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 209
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageBitmap", "cdecl"):
+    HFCreateImageBitmap = _libs[_LIBRARY_FILENAME].get("HFCreateImageBitmap", "cdecl")
+    HFCreateImageBitmap.argtypes = [PHFImageBitmapData, PHFImageBitmap]
+    HFCreateImageBitmap.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 219
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageBitmapFromFilePath", "cdecl"):
+    HFCreateImageBitmapFromFilePath = _libs[_LIBRARY_FILENAME].get("HFCreateImageBitmapFromFilePath", "cdecl")
+    HFCreateImageBitmapFromFilePath.argtypes = [HPath, HInt32, PHFImageBitmap]
+    HFCreateImageBitmapFromFilePath.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 228
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapCopy", "cdecl"):
+    HFImageBitmapCopy = _libs[_LIBRARY_FILENAME].get("HFImageBitmapCopy", "cdecl")
+    HFImageBitmapCopy.argtypes = [HFImageBitmap, PHFImageBitmap]
+    HFImageBitmapCopy.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 236
+if _libs[_LIBRARY_FILENAME].has("HFReleaseImageBitmap", "cdecl"):
+    HFReleaseImageBitmap = _libs[_LIBRARY_FILENAME].get("HFReleaseImageBitmap", "cdecl")
+    HFReleaseImageBitmap.argtypes = [HFImageBitmap]
+    HFReleaseImageBitmap.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 246
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageStreamFromImageBitmap", "cdecl"):
+    HFCreateImageStreamFromImageBitmap = _libs[_LIBRARY_FILENAME].get("HFCreateImageStreamFromImageBitmap", "cdecl")
+    HFCreateImageStreamFromImageBitmap.argtypes = [HFImageBitmap, HFRotation, PHFImageStream]
+    HFCreateImageStreamFromImageBitmap.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 257
+if _libs[_LIBRARY_FILENAME].has("HFCreateImageBitmapFromImageStreamProcess", "cdecl"):
+    HFCreateImageBitmapFromImageStreamProcess = _libs[_LIBRARY_FILENAME].get("HFCreateImageBitmapFromImageStreamProcess", "cdecl")
+    HFCreateImageBitmapFromImageStreamProcess.argtypes = [HFImageStream, PHFImageBitmap, HInt32, HFloat]
+    HFCreateImageBitmapFromImageStreamProcess.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 267
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapWriteToFile", "cdecl"):
+    HFImageBitmapWriteToFile = _libs[_LIBRARY_FILENAME].get("HFImageBitmapWriteToFile", "cdecl")
+    HFImageBitmapWriteToFile.argtypes = [HFImageBitmap, HPath]
+    HFImageBitmapWriteToFile.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 278
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapDrawRect", "cdecl"):
+    HFImageBitmapDrawRect = _libs[_LIBRARY_FILENAME].get("HFImageBitmapDrawRect", "cdecl")
+    HFImageBitmapDrawRect.argtypes = [HFImageBitmap, HFaceRect, HColor, HInt32]
+    HFImageBitmapDrawRect.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 290
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapDrawCircleF", "cdecl"):
+    HFImageBitmapDrawCircleF = _libs[_LIBRARY_FILENAME].get("HFImageBitmapDrawCircleF", "cdecl")
+    HFImageBitmapDrawCircleF.argtypes = [HFImageBitmap, HPoint2f, HInt32, HColor, HInt32]
+    HFImageBitmapDrawCircleF.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 291
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapDrawCircle", "cdecl"):
+    HFImageBitmapDrawCircle = _libs[_LIBRARY_FILENAME].get("HFImageBitmapDrawCircle", "cdecl")
+    HFImageBitmapDrawCircle.argtypes = [HFImageBitmap, HPoint2i, HInt32, HColor, HInt32]
+    HFImageBitmapDrawCircle.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 300
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapGetData", "cdecl"):
+    HFImageBitmapGetData = _libs[_LIBRARY_FILENAME].get("HFImageBitmapGetData", "cdecl")
+    HFImageBitmapGetData.argtypes = [HFImageBitmap, PHFImageBitmapData]
+    HFImageBitmapGetData.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 310
+if _libs[_LIBRARY_FILENAME].has("HFImageBitmapShow", "cdecl"):
+    HFImageBitmapShow = _libs[_LIBRARY_FILENAME].get("HFImageBitmapShow", "cdecl")
+    HFImageBitmapShow.argtypes = [HFImageBitmap, HString, HInt32]
+    HFImageBitmapShow.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 327
+if _libs[_LIBRARY_FILENAME].has("HFLaunchInspireFace", "cdecl"):
+    HFLaunchInspireFace = _libs[_LIBRARY_FILENAME].get("HFLaunchInspireFace", "cdecl")
+    HFLaunchInspireFace.argtypes = [HPath]
+    HFLaunchInspireFace.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 335
+if _libs[_LIBRARY_FILENAME].has("HFReloadInspireFace", "cdecl"):
+    HFReloadInspireFace = _libs[_LIBRARY_FILENAME].get("HFReloadInspireFace", "cdecl")
+    HFReloadInspireFace.argtypes = [HPath]
+    HFReloadInspireFace.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 343
+if _libs[_LIBRARY_FILENAME].has("HFTerminateInspireFace", "cdecl"):
+    HFTerminateInspireFace = _libs[_LIBRARY_FILENAME].get("HFTerminateInspireFace", "cdecl")
+    HFTerminateInspireFace.argtypes = []
+    HFTerminateInspireFace.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 351
+if _libs[_LIBRARY_FILENAME].has("HFQueryInspireFaceLaunchStatus", "cdecl"):
+    HFQueryInspireFaceLaunchStatus = _libs[_LIBRARY_FILENAME].get("HFQueryInspireFaceLaunchStatus", "cdecl")
+    HFQueryInspireFaceLaunchStatus.argtypes = [HPInt32]
+    HFQueryInspireFaceLaunchStatus.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 367
+if _libs[_LIBRARY_FILENAME].has("HFQueryExpansiveHardwareRGACompileOption", "cdecl"):
+    HFQueryExpansiveHardwareRGACompileOption = _libs[_LIBRARY_FILENAME].get("HFQueryExpansiveHardwareRGACompileOption", "cdecl")
+    HFQueryExpansiveHardwareRGACompileOption.argtypes = [HPInt32]
+    HFQueryExpansiveHardwareRGACompileOption.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 376
+if _libs[_LIBRARY_FILENAME].has("HFSetExpansiveHardwareRockchipDmaHeapPath", "cdecl"):
+    HFSetExpansiveHardwareRockchipDmaHeapPath = _libs[_LIBRARY_FILENAME].get("HFSetExpansiveHardwareRockchipDmaHeapPath", "cdecl")
+    HFSetExpansiveHardwareRockchipDmaHeapPath.argtypes = [HPath]
+    HFSetExpansiveHardwareRockchipDmaHeapPath.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 384
+if _libs[_LIBRARY_FILENAME].has("HFQueryExpansiveHardwareRockchipDmaHeapPath", "cdecl"):
+    HFQueryExpansiveHardwareRockchipDmaHeapPath = _libs[_LIBRARY_FILENAME].get("HFQueryExpansiveHardwareRockchipDmaHeapPath", "cdecl")
+    HFQueryExpansiveHardwareRockchipDmaHeapPath.argtypes = [HString]
+    HFQueryExpansiveHardwareRockchipDmaHeapPath.restype = HResult
+
+enum_HFImageProcessingBackend = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 392
+
+HF_IMAGE_PROCESSING_CPU = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 392
+
+HF_IMAGE_PROCESSING_RGA = 1# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 392
+
+HFImageProcessingBackend = enum_HFImageProcessingBackend# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 392
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 399
+if _libs[_LIBRARY_FILENAME].has("HFSwitchImageProcessingBackend", "cdecl"):
+    HFSwitchImageProcessingBackend = _libs[_LIBRARY_FILENAME].get("HFSwitchImageProcessingBackend", "cdecl")
+    HFSwitchImageProcessingBackend.argtypes = [HFImageProcessingBackend]
+    HFSwitchImageProcessingBackend.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 406
+if _libs[_LIBRARY_FILENAME].has("HFSetImageProcessAlignedWidth", "cdecl"):
+    HFSetImageProcessAlignedWidth = _libs[_LIBRARY_FILENAME].get("HFSetImageProcessAlignedWidth", "cdecl")
+    HFSetImageProcessAlignedWidth.argtypes = [HInt32]
+    HFSetImageProcessAlignedWidth.restype = HResult
+
+enum_HFAppleCoreMLInferenceMode = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 415
+
+HF_APPLE_COREML_INFERENCE_MODE_CPU = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 415
+
+HF_APPLE_COREML_INFERENCE_MODE_GPU = 1# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 415
+
+HF_APPLE_COREML_INFERENCE_MODE_ANE = 2# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 415
+
+HFAppleCoreMLInferenceMode = enum_HFAppleCoreMLInferenceMode# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 415
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 422
+if _libs[_LIBRARY_FILENAME].has("HFSetAppleCoreMLInferenceMode", "cdecl"):
+    HFSetAppleCoreMLInferenceMode = _libs[_LIBRARY_FILENAME].get("HFSetAppleCoreMLInferenceMode", "cdecl")
+    HFSetAppleCoreMLInferenceMode.argtypes = [HFAppleCoreMLInferenceMode]
+    HFSetAppleCoreMLInferenceMode.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 429
+if _libs[_LIBRARY_FILENAME].has("HFSetCudaDeviceId", "cdecl"):
+    HFSetCudaDeviceId = _libs[_LIBRARY_FILENAME].get("HFSetCudaDeviceId", "cdecl")
+    HFSetCudaDeviceId.argtypes = [HInt32]
+    HFSetCudaDeviceId.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 436
+if _libs[_LIBRARY_FILENAME].has("HFGetCudaDeviceId", "cdecl"):
+    HFGetCudaDeviceId = _libs[_LIBRARY_FILENAME].get("HFGetCudaDeviceId", "cdecl")
+    HFGetCudaDeviceId.argtypes = [HPInt32]
+    HFGetCudaDeviceId.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 442
+if _libs[_LIBRARY_FILENAME].has("HFPrintCudaDeviceInfo", "cdecl"):
+    HFPrintCudaDeviceInfo = _libs[_LIBRARY_FILENAME].get("HFPrintCudaDeviceInfo", "cdecl")
+    HFPrintCudaDeviceInfo.argtypes = []
+    HFPrintCudaDeviceInfo.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 449
+if _libs[_LIBRARY_FILENAME].has("HFGetNumCudaDevices", "cdecl"):
+    HFGetNumCudaDevices = _libs[_LIBRARY_FILENAME].get("HFGetNumCudaDevices", "cdecl")
+    HFGetNumCudaDevices.argtypes = [HPInt32]
+    HFGetNumCudaDevices.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 456
+if _libs[_LIBRARY_FILENAME].has("HFCheckCudaDeviceSupport", "cdecl"):
+    HFCheckCudaDeviceSupport = _libs[_LIBRARY_FILENAME].get("HFCheckCudaDeviceSupport", "cdecl")
+    HFCheckCudaDeviceSupport.argtypes = [HPInt32]
+    HFCheckCudaDeviceSupport.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 485
+class struct_HFSessionCustomParameter(Structure):
+    pass
+
+struct_HFSessionCustomParameter.__slots__ = [
+    'enable_recognition',
+    'enable_liveness',
+    'enable_ir_liveness',
+    'enable_mask_detect',
+    'enable_face_quality',
+    'enable_face_attribute',
+    'enable_interaction_liveness',
+    'enable_detect_mode_landmark',
+    'enable_face_pose',
+    'enable_face_emotion',
+]
+struct_HFSessionCustomParameter._fields_ = [
+    ('enable_recognition', HInt32),
+    ('enable_liveness', HInt32),
+    ('enable_ir_liveness', HInt32),
+    ('enable_mask_detect', HInt32),
+    ('enable_face_quality', HInt32),
+    ('enable_face_attribute', HInt32),
+    ('enable_interaction_liveness', HInt32),
+    ('enable_detect_mode_landmark', HInt32),
+    ('enable_face_pose', HInt32),
+    ('enable_face_emotion', HInt32),
+]
+
+HFSessionCustomParameter = struct_HFSessionCustomParameter# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 485
+
+PHFSessionCustomParameter = POINTER(struct_HFSessionCustomParameter)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 485
+
+enum_HFDetectMode = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 498
+
+HF_DETECT_MODE_ALWAYS_DETECT = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 498
+
+HF_DETECT_MODE_LIGHT_TRACK = (HF_DETECT_MODE_ALWAYS_DETECT + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 498
+
+HF_DETECT_MODE_TRACK_BY_DETECTION = (HF_DETECT_MODE_LIGHT_TRACK + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 498
+
+HFDetectMode = enum_HFDetectMode# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 498
+
+enum_HFSessionLandmarkEngine = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 507
+
+HF_LANDMARK_HYPLMV2_0_25 = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 507
+
+HF_LANDMARK_HYPLMV2_0_50 = 1# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 507
+
+HF_LANDMARK_INSIGHTFACE_2D106_TRACK = 2# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 507
+
+HFSessionLandmarkEngine = enum_HFSessionLandmarkEngine# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 507
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 515
+if _libs[_LIBRARY_FILENAME].has("HFSwitchLandmarkEngine", "cdecl"):
+    HFSwitchLandmarkEngine = _libs[_LIBRARY_FILENAME].get("HFSwitchLandmarkEngine", "cdecl")
+    HFSwitchLandmarkEngine.argtypes = [HFSessionLandmarkEngine]
+    HFSwitchLandmarkEngine.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 523
+class struct_HFFaceDetectPixelList(Structure):
+    pass
+
+struct_HFFaceDetectPixelList.__slots__ = [
+    'pixel_level',
+    'size',
+]
+struct_HFFaceDetectPixelList._fields_ = [
+    ('pixel_level', HInt32 * int(20)),
+    ('size', HInt32),
+]
+
+HFFaceDetectPixelList = struct_HFFaceDetectPixelList# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 523
+
+PHFFaceDetectPixelList = POINTER(struct_HFFaceDetectPixelList)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 523
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 530
+if _libs[_LIBRARY_FILENAME].has("HFQuerySupportedPixelLevelsForFaceDetection", "cdecl"):
+    HFQuerySupportedPixelLevelsForFaceDetection = _libs[_LIBRARY_FILENAME].get("HFQuerySupportedPixelLevelsForFaceDetection", "cdecl")
+    HFQuerySupportedPixelLevelsForFaceDetection.argtypes = [PHFFaceDetectPixelList]
+    HFQuerySupportedPixelLevelsForFaceDetection.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 546
+if _libs[_LIBRARY_FILENAME].has("HFCreateInspireFaceSession", "cdecl"):
+    HFCreateInspireFaceSession = _libs[_LIBRARY_FILENAME].get("HFCreateInspireFaceSession", "cdecl")
+    HFCreateInspireFaceSession.argtypes = [HFSessionCustomParameter, HFDetectMode, HInt32, HInt32, HInt32, PHFSession]
+    HFCreateInspireFaceSession.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 563
+if _libs[_LIBRARY_FILENAME].has("HFCreateInspireFaceSessionOptional", "cdecl"):
+    HFCreateInspireFaceSessionOptional = _libs[_LIBRARY_FILENAME].get("HFCreateInspireFaceSessionOptional", "cdecl")
+    HFCreateInspireFaceSessionOptional.argtypes = [HOption, HFDetectMode, HInt32, HInt32, HInt32, PHFSession]
+    HFCreateInspireFaceSessionOptional.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 572
+if _libs[_LIBRARY_FILENAME].has("HFReleaseInspireFaceSession", "cdecl"):
+    HFReleaseInspireFaceSession = _libs[_LIBRARY_FILENAME].get("HFReleaseInspireFaceSession", "cdecl")
+    HFReleaseInspireFaceSession.argtypes = [HFSession]
+    HFReleaseInspireFaceSession.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 589
+class struct_HFFaceBasicToken(Structure):
+    pass
+
+struct_HFFaceBasicToken.__slots__ = [
+    'size',
+    'data',
+]
+struct_HFFaceBasicToken._fields_ = [
+    ('size', HInt32),
+    ('data', HPVoid),
+]
+
+HFFaceBasicToken = struct_HFFaceBasicToken# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 589
+
+PHFFaceBasicToken = POINTER(struct_HFFaceBasicToken)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 589
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 600
+class struct_HFFaceEulerAngle(Structure):
+    pass
+
+struct_HFFaceEulerAngle.__slots__ = [
+    'roll',
+    'yaw',
+    'pitch',
+]
+struct_HFFaceEulerAngle._fields_ = [
+    ('roll', HPFloat),
+    ('yaw', HPFloat),
+    ('pitch', HPFloat),
+]
+
+HFFaceEulerAngle = struct_HFFaceEulerAngle# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 600
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 616
+class struct_HFMultipleFaceData(Structure):
+    pass
+
+struct_HFMultipleFaceData.__slots__ = [
+    'detectedNum',
+    'rects',
+    'trackIds',
+    'trackCounts',
+    'detConfidence',
+    'angles',
+    'tokens',
+]
+struct_HFMultipleFaceData._fields_ = [
+    ('detectedNum', HInt32),
+    ('rects', PHFaceRect),
+    ('trackIds', HPInt32),
+    ('trackCounts', HPInt32),
+    ('detConfidence', HPFloat),
+    ('angles', HFFaceEulerAngle),
+    ('tokens', PHFFaceBasicToken),
+]
+
+HFMultipleFaceData = struct_HFMultipleFaceData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 616
+
+PHFMultipleFaceData = POINTER(struct_HFMultipleFaceData)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 616
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 623
+if _libs[_LIBRARY_FILENAME].has("HFSessionClearTrackingFace", "cdecl"):
+    HFSessionClearTrackingFace = _libs[_LIBRARY_FILENAME].get("HFSessionClearTrackingFace", "cdecl")
+    HFSessionClearTrackingFace.argtypes = [HFSession]
+    HFSessionClearTrackingFace.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 631
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetTrackLostRecoveryMode", "cdecl"):
+    HFSessionSetTrackLostRecoveryMode = _libs[_LIBRARY_FILENAME].get("HFSessionSetTrackLostRecoveryMode", "cdecl")
+    HFSessionSetTrackLostRecoveryMode.argtypes = [HFSession, HInt32]
+    HFSessionSetTrackLostRecoveryMode.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 639
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetLightTrackConfidenceThreshold", "cdecl"):
+    HFSessionSetLightTrackConfidenceThreshold = _libs[_LIBRARY_FILENAME].get("HFSessionSetLightTrackConfidenceThreshold", "cdecl")
+    HFSessionSetLightTrackConfidenceThreshold.argtypes = [HFSession, HFloat]
+    HFSessionSetLightTrackConfidenceThreshold.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 649
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetTrackPreviewSize", "cdecl"):
+    HFSessionSetTrackPreviewSize = _libs[_LIBRARY_FILENAME].get("HFSessionSetTrackPreviewSize", "cdecl")
+    HFSessionSetTrackPreviewSize.argtypes = [HFSession, HInt32]
+    HFSessionSetTrackPreviewSize.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 657
+if _libs[_LIBRARY_FILENAME].has("HFSessionGetTrackPreviewSize", "cdecl"):
+    HFSessionGetTrackPreviewSize = _libs[_LIBRARY_FILENAME].get("HFSessionGetTrackPreviewSize", "cdecl")
+    HFSessionGetTrackPreviewSize.argtypes = [HFSession, HPInt32]
+    HFSessionGetTrackPreviewSize.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 667
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetFilterMinimumFacePixelSize", "cdecl"):
+    HFSessionSetFilterMinimumFacePixelSize = _libs[_LIBRARY_FILENAME].get("HFSessionSetFilterMinimumFacePixelSize", "cdecl")
+    HFSessionSetFilterMinimumFacePixelSize.argtypes = [HFSession, HInt32]
+    HFSessionSetFilterMinimumFacePixelSize.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 676
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetFaceDetectThreshold", "cdecl"):
+    HFSessionSetFaceDetectThreshold = _libs[_LIBRARY_FILENAME].get("HFSessionSetFaceDetectThreshold", "cdecl")
+    HFSessionSetFaceDetectThreshold.argtypes = [HFSession, HFloat]
+    HFSessionSetFaceDetectThreshold.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 685
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetTrackModeSmoothRatio", "cdecl"):
+    HFSessionSetTrackModeSmoothRatio = _libs[_LIBRARY_FILENAME].get("HFSessionSetTrackModeSmoothRatio", "cdecl")
+    HFSessionSetTrackModeSmoothRatio.argtypes = [HFSession, HFloat]
+    HFSessionSetTrackModeSmoothRatio.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 694
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetTrackModeNumSmoothCacheFrame", "cdecl"):
+    HFSessionSetTrackModeNumSmoothCacheFrame = _libs[_LIBRARY_FILENAME].get("HFSessionSetTrackModeNumSmoothCacheFrame", "cdecl")
+    HFSessionSetTrackModeNumSmoothCacheFrame.argtypes = [HFSession, HInt32]
+    HFSessionSetTrackModeNumSmoothCacheFrame.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 703
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetTrackModeDetectInterval", "cdecl"):
+    HFSessionSetTrackModeDetectInterval = _libs[_LIBRARY_FILENAME].get("HFSessionSetTrackModeDetectInterval", "cdecl")
+    HFSessionSetTrackModeDetectInterval.argtypes = [HFSession, HInt32]
+    HFSessionSetTrackModeDetectInterval.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 713
+if _libs[_LIBRARY_FILENAME].has("HFExecuteFaceTrack", "cdecl"):
+    HFExecuteFaceTrack = _libs[_LIBRARY_FILENAME].get("HFExecuteFaceTrack", "cdecl")
+    HFExecuteFaceTrack.argtypes = [HFSession, HFImageStream, PHFMultipleFaceData]
+    HFExecuteFaceTrack.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 721
+if _libs[_LIBRARY_FILENAME].has("HFSessionLastFaceDetectionGetDebugPreviewImageSize", "cdecl"):
+    HFSessionLastFaceDetectionGetDebugPreviewImageSize = _libs[_LIBRARY_FILENAME].get("HFSessionLastFaceDetectionGetDebugPreviewImageSize", "cdecl")
+    HFSessionLastFaceDetectionGetDebugPreviewImageSize.argtypes = [HFSession, HPInt32]
+    HFSessionLastFaceDetectionGetDebugPreviewImageSize.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 738
+if _libs[_LIBRARY_FILENAME].has("HFCopyFaceBasicToken", "cdecl"):
+    HFCopyFaceBasicToken = _libs[_LIBRARY_FILENAME].get("HFCopyFaceBasicToken", "cdecl")
+    HFCopyFaceBasicToken.argtypes = [HFFaceBasicToken, HPBuffer, HInt32]
+    HFCopyFaceBasicToken.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 752
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceBasicTokenSize", "cdecl"):
+    HFGetFaceBasicTokenSize = _libs[_LIBRARY_FILENAME].get("HFGetFaceBasicTokenSize", "cdecl")
+    HFGetFaceBasicTokenSize.argtypes = [HPInt32]
+    HFGetFaceBasicTokenSize.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 759
+if _libs[_LIBRARY_FILENAME].has("HFGetNumOfFaceDenseLandmark", "cdecl"):
+    HFGetNumOfFaceDenseLandmark = _libs[_LIBRARY_FILENAME].get("HFGetNumOfFaceDenseLandmark", "cdecl")
+    HFGetNumOfFaceDenseLandmark.argtypes = [HPInt32]
+    HFGetNumOfFaceDenseLandmark.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 769
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceDenseLandmarkFromFaceToken", "cdecl"):
+    HFGetFaceDenseLandmarkFromFaceToken = _libs[_LIBRARY_FILENAME].get("HFGetFaceDenseLandmarkFromFaceToken", "cdecl")
+    HFGetFaceDenseLandmarkFromFaceToken.argtypes = [HFFaceBasicToken, PHPoint2f, HInt32]
+    HFGetFaceDenseLandmarkFromFaceToken.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 778
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceFiveKeyPointsFromFaceToken", "cdecl"):
+    HFGetFaceFiveKeyPointsFromFaceToken = _libs[_LIBRARY_FILENAME].get("HFGetFaceFiveKeyPointsFromFaceToken", "cdecl")
+    HFGetFaceFiveKeyPointsFromFaceToken.argtypes = [HFFaceBasicToken, PHPoint2f, HInt32]
+    HFGetFaceFiveKeyPointsFromFaceToken.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 785
+if _libs[_LIBRARY_FILENAME].has("HFSessionSetEnableTrackCostSpend", "cdecl"):
+    HFSessionSetEnableTrackCostSpend = _libs[_LIBRARY_FILENAME].get("HFSessionSetEnableTrackCostSpend", "cdecl")
+    HFSessionSetEnableTrackCostSpend.argtypes = [HFSession, HInt32]
+    HFSessionSetEnableTrackCostSpend.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 792
+if _libs[_LIBRARY_FILENAME].has("HFSessionPrintTrackCostSpend", "cdecl"):
+    HFSessionPrintTrackCostSpend = _libs[_LIBRARY_FILENAME].get("HFSessionPrintTrackCostSpend", "cdecl")
+    HFSessionPrintTrackCostSpend.argtypes = [HFSession]
+    HFSessionPrintTrackCostSpend.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 809
+class struct_HFFaceFeature(Structure):
+    pass
+
+struct_HFFaceFeature.__slots__ = [
+    'size',
+    'data',
+]
+struct_HFFaceFeature._fields_ = [
+    ('size', HInt32),
+    ('data', HPFloat),
+]
+
+HFFaceFeature = struct_HFFaceFeature# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 809
+
+PHFFaceFeature = POINTER(struct_HFFaceFeature)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 809
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 820
+if _libs[_LIBRARY_FILENAME].has("HFFaceFeatureExtract", "cdecl"):
+    HFFaceFeatureExtract = _libs[_LIBRARY_FILENAME].get("HFFaceFeatureExtract", "cdecl")
+    HFFaceFeatureExtract.argtypes = [HFSession, HFImageStream, HFFaceBasicToken, PHFFaceFeature]
+    HFFaceFeatureExtract.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 831
+if _libs[_LIBRARY_FILENAME].has("HFFaceFeatureExtractTo", "cdecl"):
+    HFFaceFeatureExtractTo = _libs[_LIBRARY_FILENAME].get("HFFaceFeatureExtractTo", "cdecl")
+    HFFaceFeatureExtractTo.argtypes = [HFSession, HFImageStream, HFFaceBasicToken, HFFaceFeature]
+    HFFaceFeatureExtractTo.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 843
+if _libs[_LIBRARY_FILENAME].has("HFFaceFeatureExtractCpy", "cdecl"):
+    HFFaceFeatureExtractCpy = _libs[_LIBRARY_FILENAME].get("HFFaceFeatureExtractCpy", "cdecl")
+    HFFaceFeatureExtractCpy.argtypes = [HFSession, HFImageStream, HFFaceBasicToken, HPFloat]
+    HFFaceFeatureExtractCpy.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 850
+if _libs[_LIBRARY_FILENAME].has("HFCreateFaceFeature", "cdecl"):
+    HFCreateFaceFeature = _libs[_LIBRARY_FILENAME].get("HFCreateFaceFeature", "cdecl")
+    HFCreateFaceFeature.argtypes = [PHFFaceFeature]
+    HFCreateFaceFeature.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 857
+if _libs[_LIBRARY_FILENAME].has("HFReleaseFaceFeature", "cdecl"):
+    HFReleaseFaceFeature = _libs[_LIBRARY_FILENAME].get("HFReleaseFaceFeature", "cdecl")
+    HFReleaseFaceFeature.argtypes = [PHFFaceFeature]
+    HFReleaseFaceFeature.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 867
+if _libs[_LIBRARY_FILENAME].has("HFFaceGetFaceAlignmentImage", "cdecl"):
+    HFFaceGetFaceAlignmentImage = _libs[_LIBRARY_FILENAME].get("HFFaceGetFaceAlignmentImage", "cdecl")
+    HFFaceGetFaceAlignmentImage.argtypes = [HFSession, HFImageStream, HFFaceBasicToken, PHFImageBitmap]
+    HFFaceGetFaceAlignmentImage.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 877
+if _libs[_LIBRARY_FILENAME].has("HFFaceFeatureExtractWithAlignmentImage", "cdecl"):
+    HFFaceFeatureExtractWithAlignmentImage = _libs[_LIBRARY_FILENAME].get("HFFaceFeatureExtractWithAlignmentImage", "cdecl")
+    HFFaceFeatureExtractWithAlignmentImage.argtypes = [HFSession, HFImageStream, HFFaceFeature]
+    HFFaceFeatureExtractWithAlignmentImage.restype = HResult
+
+enum_HFSearchMode = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 895
+
+HF_SEARCH_MODE_EAGER = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 895
+
+HF_SEARCH_MODE_EXHAUSTIVE = (HF_SEARCH_MODE_EAGER + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 895
+
+HFSearchMode = enum_HFSearchMode# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 895
+
+enum_HFPKMode = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 903
+
+HF_PK_AUTO_INCREMENT = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 903
+
+HF_PK_MANUAL_INPUT = (HF_PK_AUTO_INCREMENT + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 903
+
+HFPKMode = enum_HFPKMode# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 903
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 916
+class struct_HFFeatureHubConfiguration(Structure):
+    pass
+
+struct_HFFeatureHubConfiguration.__slots__ = [
+    'primaryKeyMode',
+    'enablePersistence',
+    'persistenceDbPath',
+    'searchThreshold',
+    'searchMode',
+]
+struct_HFFeatureHubConfiguration._fields_ = [
+    ('primaryKeyMode', HFPKMode),
+    ('enablePersistence', HInt32),
+    ('persistenceDbPath', HString),
+    ('searchThreshold', HFloat),
+    ('searchMode', HFSearchMode),
+]
+
+HFFeatureHubConfiguration = struct_HFFeatureHubConfiguration# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 916
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 928
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubDataEnable", "cdecl"):
+    HFFeatureHubDataEnable = _libs[_LIBRARY_FILENAME].get("HFFeatureHubDataEnable", "cdecl")
+    HFFeatureHubDataEnable.argtypes = [HFFeatureHubConfiguration]
+    HFFeatureHubDataEnable.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 934
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubDataDisable", "cdecl"):
+    HFFeatureHubDataDisable = _libs[_LIBRARY_FILENAME].get("HFFeatureHubDataDisable", "cdecl")
+    HFFeatureHubDataDisable.argtypes = []
+    HFFeatureHubDataDisable.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 945
+class struct_HFFaceFeatureIdentity(Structure):
+    pass
+
+struct_HFFaceFeatureIdentity.__slots__ = [
+    'id',
+    'feature',
+]
+struct_HFFaceFeatureIdentity._fields_ = [
+    ('id', HFaceId),
+    ('feature', PHFFaceFeature),
+]
+
+HFFaceFeatureIdentity = struct_HFFaceFeatureIdentity# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 945
+
+PHFFaceFeatureIdentity = POINTER(struct_HFFaceFeatureIdentity)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 945
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 954
+class struct_HFSearchTopKResults(Structure):
+    pass
+
+struct_HFSearchTopKResults.__slots__ = [
+    'size',
+    'confidence',
+    'ids',
+]
+struct_HFSearchTopKResults._fields_ = [
+    ('size', HInt32),
+    ('confidence', HPFloat),
+    ('ids', HPFaceId),
+]
+
+HFSearchTopKResults = struct_HFSearchTopKResults# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 954
+
+PHFSearchTopKResults = POINTER(struct_HFSearchTopKResults)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 954
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 966
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubFaceSearchThresholdSetting", "cdecl"):
+    HFFeatureHubFaceSearchThresholdSetting = _libs[_LIBRARY_FILENAME].get("HFFeatureHubFaceSearchThresholdSetting", "cdecl")
+    HFFeatureHubFaceSearchThresholdSetting.argtypes = [HFloat]
+    HFFeatureHubFaceSearchThresholdSetting.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 981
+if _libs[_LIBRARY_FILENAME].has("HFFaceComparison", "cdecl"):
+    HFFaceComparison = _libs[_LIBRARY_FILENAME].get("HFFaceComparison", "cdecl")
+    HFFaceComparison.argtypes = [HFFaceFeature, HFFaceFeature, HPFloat]
+    HFFaceComparison.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 988
+if _libs[_LIBRARY_FILENAME].has("HFGetRecommendedCosineThreshold", "cdecl"):
+    HFGetRecommendedCosineThreshold = _libs[_LIBRARY_FILENAME].get("HFGetRecommendedCosineThreshold", "cdecl")
+    HFGetRecommendedCosineThreshold.argtypes = [HPFloat]
+    HFGetRecommendedCosineThreshold.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1001
+if _libs[_LIBRARY_FILENAME].has("HFCosineSimilarityConvertToPercentage", "cdecl"):
+    HFCosineSimilarityConvertToPercentage = _libs[_LIBRARY_FILENAME].get("HFCosineSimilarityConvertToPercentage", "cdecl")
+    HFCosineSimilarityConvertToPercentage.argtypes = [HFloat, HPFloat]
+    HFCosineSimilarityConvertToPercentage.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1014
+class struct_HFSimilarityConverterConfig(Structure):
+    pass
+
+struct_HFSimilarityConverterConfig.__slots__ = [
+    'threshold',
+    'middleScore',
+    'steepness',
+    'outputMin',
+    'outputMax',
+]
+struct_HFSimilarityConverterConfig._fields_ = [
+    ('threshold', HFloat),
+    ('middleScore', HFloat),
+    ('steepness', HFloat),
+    ('outputMin', HFloat),
+    ('outputMax', HFloat),
+]
+
+HFSimilarityConverterConfig = struct_HFSimilarityConverterConfig# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1014
+
+PHFSimilarityConverterConfig = POINTER(struct_HFSimilarityConverterConfig)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1014
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1023
+if _libs[_LIBRARY_FILENAME].has("HFUpdateCosineSimilarityConverter", "cdecl"):
+    HFUpdateCosineSimilarityConverter = _libs[_LIBRARY_FILENAME].get("HFUpdateCosineSimilarityConverter", "cdecl")
+    HFUpdateCosineSimilarityConverter.argtypes = [HFSimilarityConverterConfig]
+    HFUpdateCosineSimilarityConverter.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1030
+if _libs[_LIBRARY_FILENAME].has("HFGetCosineSimilarityConverter", "cdecl"):
+    HFGetCosineSimilarityConverter = _libs[_LIBRARY_FILENAME].get("HFGetCosineSimilarityConverter", "cdecl")
+    HFGetCosineSimilarityConverter.argtypes = [PHFSimilarityConverterConfig]
+    HFGetCosineSimilarityConverter.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1038
+if _libs[_LIBRARY_FILENAME].has("HFGetFeatureLength", "cdecl"):
+    HFGetFeatureLength = _libs[_LIBRARY_FILENAME].get("HFGetFeatureLength", "cdecl")
+    HFGetFeatureLength.argtypes = [HPInt32]
+    HFGetFeatureLength.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1046
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubInsertFeature", "cdecl"):
+    HFFeatureHubInsertFeature = _libs[_LIBRARY_FILENAME].get("HFFeatureHubInsertFeature", "cdecl")
+    HFFeatureHubInsertFeature.argtypes = [HFFaceFeatureIdentity, HPFaceId]
+    HFFeatureHubInsertFeature.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1057
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubFaceSearch", "cdecl"):
+    HFFeatureHubFaceSearch = _libs[_LIBRARY_FILENAME].get("HFFeatureHubFaceSearch", "cdecl")
+    HFFeatureHubFaceSearch.argtypes = [HFFaceFeature, HPFloat, PHFFaceFeatureIdentity]
+    HFFeatureHubFaceSearch.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1067
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubFaceSearchTopK", "cdecl"):
+    HFFeatureHubFaceSearchTopK = _libs[_LIBRARY_FILENAME].get("HFFeatureHubFaceSearchTopK", "cdecl")
+    HFFeatureHubFaceSearchTopK.argtypes = [HFFaceFeature, HInt32, PHFSearchTopKResults]
+    HFFeatureHubFaceSearchTopK.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1075
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubFaceRemove", "cdecl"):
+    HFFeatureHubFaceRemove = _libs[_LIBRARY_FILENAME].get("HFFeatureHubFaceRemove", "cdecl")
+    HFFeatureHubFaceRemove.argtypes = [HFaceId]
+    HFFeatureHubFaceRemove.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1083
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubFaceUpdate", "cdecl"):
+    HFFeatureHubFaceUpdate = _libs[_LIBRARY_FILENAME].get("HFFeatureHubFaceUpdate", "cdecl")
+    HFFeatureHubFaceUpdate.argtypes = [HFFaceFeatureIdentity]
+    HFFeatureHubFaceUpdate.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1092
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubGetFaceIdentity", "cdecl"):
+    HFFeatureHubGetFaceIdentity = _libs[_LIBRARY_FILENAME].get("HFFeatureHubGetFaceIdentity", "cdecl")
+    HFFeatureHubGetFaceIdentity.argtypes = [HFaceId, PHFFaceFeatureIdentity]
+    HFFeatureHubGetFaceIdentity.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1100
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubGetFaceCount", "cdecl"):
+    HFFeatureHubGetFaceCount = _libs[_LIBRARY_FILENAME].get("HFFeatureHubGetFaceCount", "cdecl")
+    HFFeatureHubGetFaceCount.argtypes = [HPInt32]
+    HFFeatureHubGetFaceCount.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1107
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubViewDBTable", "cdecl"):
+    HFFeatureHubViewDBTable = _libs[_LIBRARY_FILENAME].get("HFFeatureHubViewDBTable", "cdecl")
+    HFFeatureHubViewDBTable.argtypes = []
+    HFFeatureHubViewDBTable.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1115
+class struct_HFFeatureHubExistingIds(Structure):
+    pass
+
+struct_HFFeatureHubExistingIds.__slots__ = [
+    'size',
+    'ids',
+]
+struct_HFFeatureHubExistingIds._fields_ = [
+    ('size', HInt32),
+    ('ids', HPFaceId),
+]
+
+HFFeatureHubExistingIds = struct_HFFeatureHubExistingIds# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1115
+
+PHFFeatureHubExistingIds = POINTER(struct_HFFeatureHubExistingIds)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1115
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1122
+if _libs[_LIBRARY_FILENAME].has("HFFeatureHubGetExistingIds", "cdecl"):
+    HFFeatureHubGetExistingIds = _libs[_LIBRARY_FILENAME].get("HFFeatureHubGetExistingIds", "cdecl")
+    HFFeatureHubGetExistingIds.argtypes = [PHFFeatureHubExistingIds]
+    HFFeatureHubGetExistingIds.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1143
+if _libs[_LIBRARY_FILENAME].has("HFMultipleFacePipelineProcess", "cdecl"):
+    HFMultipleFacePipelineProcess = _libs[_LIBRARY_FILENAME].get("HFMultipleFacePipelineProcess", "cdecl")
+    HFMultipleFacePipelineProcess.argtypes = [HFSession, HFImageStream, PHFMultipleFaceData, HFSessionCustomParameter]
+    HFMultipleFacePipelineProcess.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1158
+if _libs[_LIBRARY_FILENAME].has("HFMultipleFacePipelineProcessOptional", "cdecl"):
+    HFMultipleFacePipelineProcessOptional = _libs[_LIBRARY_FILENAME].get("HFMultipleFacePipelineProcessOptional", "cdecl")
+    HFMultipleFacePipelineProcessOptional.argtypes = [HFSession, HFImageStream, PHFMultipleFaceData, HInt32]
+    HFMultipleFacePipelineProcessOptional.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1170
+class struct_HFRGBLivenessConfidence(Structure):
+    pass
+
+struct_HFRGBLivenessConfidence.__slots__ = [
+    'num',
+    'confidence',
+]
+struct_HFRGBLivenessConfidence._fields_ = [
+    ('num', HInt32),
+    ('confidence', HPFloat),
+]
+
+HFRGBLivenessConfidence = struct_HFRGBLivenessConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1170
+
+PHFRGBLivenessConfidence = POINTER(struct_HFRGBLivenessConfidence)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1170
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1182
+if _libs[_LIBRARY_FILENAME].has("HFGetRGBLivenessConfidence", "cdecl"):
+    HFGetRGBLivenessConfidence = _libs[_LIBRARY_FILENAME].get("HFGetRGBLivenessConfidence", "cdecl")
+    HFGetRGBLivenessConfidence.argtypes = [HFSession, PHFRGBLivenessConfidence]
+    HFGetRGBLivenessConfidence.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1193
+class struct_HFFaceMaskConfidence(Structure):
+    pass
+
+struct_HFFaceMaskConfidence.__slots__ = [
+    'num',
+    'confidence',
+]
+struct_HFFaceMaskConfidence._fields_ = [
+    ('num', HInt32),
+    ('confidence', HPFloat),
+]
+
+HFFaceMaskConfidence = struct_HFFaceMaskConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1193
+
+PHFFaceMaskConfidence = POINTER(struct_HFFaceMaskConfidence)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1193
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1205
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceMaskConfidence", "cdecl"):
+    HFGetFaceMaskConfidence = _libs[_LIBRARY_FILENAME].get("HFGetFaceMaskConfidence", "cdecl")
+    HFGetFaceMaskConfidence.argtypes = [HFSession, PHFFaceMaskConfidence]
+    HFGetFaceMaskConfidence.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1216
+class struct_HFFaceQualityConfidence(Structure):
+    pass
+
+struct_HFFaceQualityConfidence.__slots__ = [
+    'num',
+    'confidence',
+]
+struct_HFFaceQualityConfidence._fields_ = [
+    ('num', HInt32),
+    ('confidence', HPFloat),
+]
+
+HFFaceQualityConfidence = struct_HFFaceQualityConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1216
+
+PHFFaceQualityConfidence = POINTER(struct_HFFaceQualityConfidence)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1216
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1228
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceQualityConfidence", "cdecl"):
+    HFGetFaceQualityConfidence = _libs[_LIBRARY_FILENAME].get("HFGetFaceQualityConfidence", "cdecl")
+    HFGetFaceQualityConfidence.argtypes = [HFSession, PHFFaceQualityConfidence]
+    HFGetFaceQualityConfidence.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1240
+if _libs[_LIBRARY_FILENAME].has("HFFaceQualityDetect", "cdecl"):
+    HFFaceQualityDetect = _libs[_LIBRARY_FILENAME].get("HFFaceQualityDetect", "cdecl")
+    HFFaceQualityDetect.argtypes = [HFSession, HFFaceBasicToken, HPFloat]
+    HFFaceQualityDetect.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1251
+class struct_HFFaceInteractionState(Structure):
+    pass
+
+struct_HFFaceInteractionState.__slots__ = [
+    'num',
+    'leftEyeStatusConfidence',
+    'rightEyeStatusConfidence',
+]
+struct_HFFaceInteractionState._fields_ = [
+    ('num', HInt32),
+    ('leftEyeStatusConfidence', HPFloat),
+    ('rightEyeStatusConfidence', HPFloat),
+]
+
+HFFaceInteractionState = struct_HFFaceInteractionState# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1251
+
+PHFFaceInteractionState = POINTER(struct_HFFaceInteractionState)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1251
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1258
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceInteractionStateResult", "cdecl"):
+    HFGetFaceInteractionStateResult = _libs[_LIBRARY_FILENAME].get("HFGetFaceInteractionStateResult", "cdecl")
+    HFGetFaceInteractionStateResult.argtypes = [HFSession, PHFFaceInteractionState]
+    HFGetFaceInteractionStateResult.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1270
+class struct_HFFaceInteractionsActions(Structure):
+    pass
+
+struct_HFFaceInteractionsActions.__slots__ = [
+    'num',
+    'normal',
+    'shake',
+    'jawOpen',
+    'headRaise',
+    'blink',
+]
+struct_HFFaceInteractionsActions._fields_ = [
+    ('num', HInt32),
+    ('normal', HPInt32),
+    ('shake', HPInt32),
+    ('jawOpen', HPInt32),
+    ('headRaise', HPInt32),
+    ('blink', HPInt32),
+]
+
+HFFaceInteractionsActions = struct_HFFaceInteractionsActions# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1270
+
+PHFFaceInteractionsActions = POINTER(struct_HFFaceInteractionsActions)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1270
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1278
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceInteractionActionsResult", "cdecl"):
+    HFGetFaceInteractionActionsResult = _libs[_LIBRARY_FILENAME].get("HFGetFaceInteractionActionsResult", "cdecl")
+    HFGetFaceInteractionActionsResult.argtypes = [HFSession, PHFFaceInteractionsActions]
+    HFGetFaceInteractionActionsResult.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1306
+class struct_HFFaceAttributeResult(Structure):
+    pass
+
+struct_HFFaceAttributeResult.__slots__ = [
+    'num',
+    'race',
+    'gender',
+    'ageBracket',
+]
+struct_HFFaceAttributeResult._fields_ = [
+    ('num', HInt32),
+    ('race', HPInt32),
+    ('gender', HPInt32),
+    ('ageBracket', HPInt32),
+]
+
+HFFaceAttributeResult = struct_HFFaceAttributeResult# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1306
+
+PHFFaceAttributeResult = POINTER(struct_HFFaceAttributeResult)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1306
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1318
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceAttributeResult", "cdecl"):
+    HFGetFaceAttributeResult = _libs[_LIBRARY_FILENAME].get("HFGetFaceAttributeResult", "cdecl")
+    HFGetFaceAttributeResult.argtypes = [HFSession, PHFFaceAttributeResult]
+    HFGetFaceAttributeResult.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1333
+class struct_HFFaceEmotionResult(Structure):
+    pass
+
+struct_HFFaceEmotionResult.__slots__ = [
+    'num',
+    'emotion',
+]
+struct_HFFaceEmotionResult._fields_ = [
+    ('num', HInt32),
+    ('emotion', HPInt32),
+]
+
+HFFaceEmotionResult = struct_HFFaceEmotionResult# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1333
+
+PHFFaceEmotionResult = POINTER(struct_HFFaceEmotionResult)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1333
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1341
+if _libs[_LIBRARY_FILENAME].has("HFGetFaceEmotionResult", "cdecl"):
+    HFGetFaceEmotionResult = _libs[_LIBRARY_FILENAME].get("HFGetFaceEmotionResult", "cdecl")
+    HFGetFaceEmotionResult.argtypes = [HFSession, PHFFaceEmotionResult]
+    HFGetFaceEmotionResult.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1354
+class struct_HFInspireFaceVersion(Structure):
+    pass
+
+struct_HFInspireFaceVersion.__slots__ = [
+    'major',
+    'minor',
+    'patch',
+]
+struct_HFInspireFaceVersion._fields_ = [
+    ('major', HInt32),
+    ('minor', HInt32),
+    ('patch', HInt32),
+]
+
+HFInspireFaceVersion = struct_HFInspireFaceVersion# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1354
+
+PHFInspireFaceVersion = POINTER(struct_HFInspireFaceVersion)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1354
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1364
+if _libs[_LIBRARY_FILENAME].has("HFQueryInspireFaceVersion", "cdecl"):
+    HFQueryInspireFaceVersion = _libs[_LIBRARY_FILENAME].get("HFQueryInspireFaceVersion", "cdecl")
+    HFQueryInspireFaceVersion.argtypes = [PHFInspireFaceVersion]
+    HFQueryInspireFaceVersion.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1372
+class struct_HFInspireFaceExtendedInformation(Structure):
+    pass
+
+struct_HFInspireFaceExtendedInformation.__slots__ = [
+    'information',
+]
+struct_HFInspireFaceExtendedInformation._fields_ = [
+    ('information', HChar * int(256)),
+]
+
+HFInspireFaceExtendedInformation = struct_HFInspireFaceExtendedInformation# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1372
+
+PHFInspireFaceExtendedInformation = POINTER(struct_HFInspireFaceExtendedInformation)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1372
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1379
+if _libs[_LIBRARY_FILENAME].has("HFQueryInspireFaceExtendedInformation", "cdecl"):
+    HFQueryInspireFaceExtendedInformation = _libs[_LIBRARY_FILENAME].get("HFQueryInspireFaceExtendedInformation", "cdecl")
+    HFQueryInspireFaceExtendedInformation.argtypes = [PHFInspireFaceExtendedInformation]
+    HFQueryInspireFaceExtendedInformation.restype = HResult
+
+enum_HFLogLevel = c_int# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_NONE = 0# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_DEBUG = (HF_LOG_NONE + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_INFO = (HF_LOG_DEBUG + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_WARN = (HF_LOG_INFO + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_ERROR = (HF_LOG_WARN + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HF_LOG_FATAL = (HF_LOG_ERROR + 1)# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+HFLogLevel = enum_HFLogLevel# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1391
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1396
+if _libs[_LIBRARY_FILENAME].has("HFSetLogLevel", "cdecl"):
+    HFSetLogLevel = _libs[_LIBRARY_FILENAME].get("HFSetLogLevel", "cdecl")
+    HFSetLogLevel.argtypes = [HFLogLevel]
+    HFSetLogLevel.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1401
+if _libs[_LIBRARY_FILENAME].has("HFLogDisable", "cdecl"):
+    HFLogDisable = _libs[_LIBRARY_FILENAME].get("HFLogDisable", "cdecl")
+    HFLogDisable.argtypes = []
+    HFLogDisable.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1411
+if _libs[_LIBRARY_FILENAME].has("HFLogPrint", "cdecl"):
+    _func = _libs[_LIBRARY_FILENAME].get("HFLogPrint", "cdecl")
+    _restype = HResult
+    _errcheck = None
+    _argtypes = [HFLogLevel, HFormat]
+    HFLogPrint = _variadic_function(_func,_restype,_argtypes,_errcheck)
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1424
+if _libs[_LIBRARY_FILENAME].has("HFDeBugImageStreamImShow", "cdecl"):
+    HFDeBugImageStreamImShow = _libs[_LIBRARY_FILENAME].get("HFDeBugImageStreamImShow", "cdecl")
+    HFDeBugImageStreamImShow.argtypes = [HFImageStream]
+    HFDeBugImageStreamImShow.restype = None
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1436
+if _libs[_LIBRARY_FILENAME].has("HFDeBugImageStreamDecodeSave", "cdecl"):
+    HFDeBugImageStreamDecodeSave = _libs[_LIBRARY_FILENAME].get("HFDeBugImageStreamDecodeSave", "cdecl")
+    HFDeBugImageStreamDecodeSave.argtypes = [HFImageStream, HPath]
+    HFDeBugImageStreamDecodeSave.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1451
+if _libs[_LIBRARY_FILENAME].has("HFDeBugShowResourceStatistics", "cdecl"):
+    HFDeBugShowResourceStatistics = _libs[_LIBRARY_FILENAME].get("HFDeBugShowResourceStatistics", "cdecl")
+    HFDeBugShowResourceStatistics.argtypes = []
+    HFDeBugShowResourceStatistics.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1461
+if _libs[_LIBRARY_FILENAME].has("HFDeBugGetUnreleasedSessionsCount", "cdecl"):
+    HFDeBugGetUnreleasedSessionsCount = _libs[_LIBRARY_FILENAME].get("HFDeBugGetUnreleasedSessionsCount", "cdecl")
+    HFDeBugGetUnreleasedSessionsCount.argtypes = [HPInt32]
+    HFDeBugGetUnreleasedSessionsCount.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1472
+if _libs[_LIBRARY_FILENAME].has("HFDeBugGetUnreleasedSessions", "cdecl"):
+    HFDeBugGetUnreleasedSessions = _libs[_LIBRARY_FILENAME].get("HFDeBugGetUnreleasedSessions", "cdecl")
+    HFDeBugGetUnreleasedSessions.argtypes = [PHFSession, HInt32]
+    HFDeBugGetUnreleasedSessions.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1482
+if _libs[_LIBRARY_FILENAME].has("HFDeBugGetUnreleasedStreamsCount", "cdecl"):
+    HFDeBugGetUnreleasedStreamsCount = _libs[_LIBRARY_FILENAME].get("HFDeBugGetUnreleasedStreamsCount", "cdecl")
+    HFDeBugGetUnreleasedStreamsCount.argtypes = [HPInt32]
+    HFDeBugGetUnreleasedStreamsCount.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1493
+if _libs[_LIBRARY_FILENAME].has("HFDeBugGetUnreleasedStreams", "cdecl"):
+    HFDeBugGetUnreleasedStreams = _libs[_LIBRARY_FILENAME].get("HFDeBugGetUnreleasedStreams", "cdecl")
+    HFDeBugGetUnreleasedStreams.argtypes = [PHFImageStream, HInt32]
+    HFDeBugGetUnreleasedStreams.restype = HResult
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 27
+try:
+    HF_STATUS_ENABLE = 1
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 28
+try:
+    HF_STATUS_DISABLE = 0
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 30
+try:
+    HF_ENABLE_NONE = 0x00000000
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 31
+try:
+    HF_ENABLE_FACE_RECOGNITION = 0x00000002
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 32
+try:
+    HF_ENABLE_LIVENESS = 0x00000004
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 33
+try:
+    HF_ENABLE_IR_LIVENESS = 0x00000008
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 34
+try:
+    HF_ENABLE_MASK_DETECT = 0x00000010
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 35
+try:
+    HF_ENABLE_FACE_ATTRIBUTE = 0x00000020
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 36
+try:
+    HF_ENABLE_PLACEHOLDER_ = 0x00000040
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 37
+try:
+    HF_ENABLE_QUALITY = 0x00000080
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 38
+try:
+    HF_ENABLE_INTERACTION = 0x00000100
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 39
+try:
+    HF_ENABLE_FACE_POSE = 0x00000200
+except:
+    pass
+
+# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 40
+try:
+    HF_ENABLE_FACE_EMOTION = 0x00000400
+except:
+    pass
+
+HFImageData = struct_HFImageData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 123
+
+HFImageBitmapData = struct_HFImageBitmapData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 200
+
+HFSessionCustomParameter = struct_HFSessionCustomParameter# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 485
+
+HFFaceDetectPixelList = struct_HFFaceDetectPixelList# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 523
+
+HFFaceBasicToken = struct_HFFaceBasicToken# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 589
+
+HFFaceEulerAngle = struct_HFFaceEulerAngle# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 600
+
+HFMultipleFaceData = struct_HFMultipleFaceData# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 616
+
+HFFaceFeature = struct_HFFaceFeature# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 809
+
+HFFeatureHubConfiguration = struct_HFFeatureHubConfiguration# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 916
+
+HFFaceFeatureIdentity = struct_HFFaceFeatureIdentity# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 945
+
+HFSearchTopKResults = struct_HFSearchTopKResults# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 954
+
+HFSimilarityConverterConfig = struct_HFSimilarityConverterConfig# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1014
+
+HFFeatureHubExistingIds = struct_HFFeatureHubExistingIds# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1115
+
+HFRGBLivenessConfidence = struct_HFRGBLivenessConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1170
+
+HFFaceMaskConfidence = struct_HFFaceMaskConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1193
+
+HFFaceQualityConfidence = struct_HFFaceQualityConfidence# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1216
+
+HFFaceInteractionState = struct_HFFaceInteractionState# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1251
+
+HFFaceInteractionsActions = struct_HFFaceInteractionsActions# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1270
+
+HFFaceAttributeResult = struct_HFFaceAttributeResult# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1306
+
+HFFaceEmotionResult = struct_HFFaceEmotionResult# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1333
+
+HFInspireFaceVersion = struct_HFInspireFaceVersion# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1354
+
+HFInspireFaceExtendedInformation = struct_HFInspireFaceExtendedInformation# /Users/tunm/work/InspireFace/cpp/inspireface/c_api/inspireface.h: 1372
+
+# No inserted files
+
+# No prefix-stripping
+
diff --git a/cpp-package/inspireface/python/inspireface/modules/exception.py b/cpp-package/inspireface/python/inspireface/modules/exception.py
new file mode 100644
index 000000000..3dbd3fe8e
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/exception.py
@@ -0,0 +1,242 @@
+from . import herror as errcode
+from typing import Optional, Dict, Any
+
+
+class InspireFaceError(Exception):
+    """Base class for all InspireFace exceptions"""
+    
+    def __init__(self, message: str, error_code: Optional[int] = None, **context):
+        super().__init__(message)
+        self.error_code = error_code
+        self.context = context
+        self._error_name = self._get_error_name(error_code) if error_code else None
+    
+    def _get_error_name(self, error_code: int) -> str:
+        """Get error name corresponding to error code"""
+        for name, value in errcode.__dict__.items():
+            if isinstance(value, int) and value == error_code:
+                return name
+        return f"UNKNOWN_ERROR"
+    
+    def __str__(self):
+        base_msg = super().__str__()
+        if self.error_code is not None:
+            return f"[{self._error_name}({self.error_code})] {base_msg}"
+        return base_msg
+    
+    @property
+    def error_name(self) -> Optional[str]:
+        """Get error name"""
+        return self._error_name
+
+
+class InvalidInputError(InspireFaceError, ValueError):
+    """Input parameter/data format error"""
+    pass
+
+
+class SystemNotReadyError(InspireFaceError, RuntimeError):
+    """System not initialized or resources not ready"""
+    pass
+
+
+class ProcessingError(InspireFaceError):
+    """Business logic error during processing"""
+    pass
+
+
+class ResourceError(InspireFaceError, OSError):
+    """Resource-related errors (handles, memory, files, etc.)"""
+    pass
+
+
+class HardwareError(InspireFaceError):
+    """Hardware-related errors (CUDA, devices, etc.)"""
+    pass
+
+
+class FeatureHubError(InspireFaceError):
+    """Feature hub related errors"""
+    pass
+
+
+# === Error code mapping table ===
+ERROR_CODE_MAPPING = {
+    # Input parameter errors
+    'invalid_input': [
+        errcode.HERR_INVALID_PARAM,
+        errcode.HERR_INVALID_IMAGE_STREAM_PARAM,
+        errcode.HERR_INVALID_BUFFER_SIZE,
+        errcode.HERR_INVALID_DETECTION_INPUT,
+    ],
+    
+    # System not ready
+    'system_not_ready': [
+        errcode.HERR_ARCHIVE_NOT_LOAD,
+        errcode.HERR_SESS_INVALID_RESOURCE,
+        errcode.HERR_ARCHIVE_LOAD_MODEL_FAILURE,
+    ],
+    
+    # Processing errors
+    'processing': [
+        errcode.HERR_SESS_FUNCTION_UNUSABLE,
+        errcode.HERR_SESS_TRACKER_FAILURE,
+        errcode.HERR_SESS_PIPELINE_FAILURE,
+        errcode.HERR_SESS_REC_EXTRACT_FAILURE,
+        errcode.HERR_SESS_LANDMARK_NOT_ENABLE,
+        errcode.HERR_IMAGE_STREAM_DECODE_FAILED,
+    ],
+    
+    # Resource errors
+    'resource': [
+        errcode.HERR_INVALID_IMAGE_STREAM_HANDLE,
+        errcode.HERR_INVALID_CONTEXT_HANDLE,
+        errcode.HERR_INVALID_FACE_TOKEN,
+        errcode.HERR_INVALID_FACE_FEATURE,
+        errcode.HERR_INVALID_FACE_LIST,
+        errcode.HERR_INVALID_IMAGE_BITMAP_HANDLE,
+    ],
+    
+    # Hardware errors
+    'hardware': [
+        errcode.HERR_DEVICE_CUDA_NOT_SUPPORT,
+        errcode.HERR_DEVICE_CUDA_UNKNOWN_ERROR,
+    ],
+    
+    # Feature hub errors
+    'feature_hub': [
+        errcode.HERR_FT_HUB_DISABLE,
+        errcode.HERR_FT_HUB_INSERT_FAILURE,
+        errcode.HERR_FT_HUB_NOT_FOUND_FEATURE,
+    ],
+}
+
+
+def check_error(error_code: int, operation: str = "", **context):
+    """
+    Check error code and raise corresponding exception
+    
+    Args:
+        error_code: Error code returned by C library
+        operation: Operation description for building error message
+        **context: Additional context information
+    
+    Raises:
+        Corresponding InspireFaceError subclass exception
+    """
+    if error_code == errcode.HSUCCEED:
+        return
+    
+    # Get error name
+    error_name = None
+    for name, value in errcode.__dict__.items():
+        if isinstance(value, int) and value == error_code:
+            error_name = name
+            break
+    
+    # Build basic error message
+    if operation:
+        message = f"{operation} failed"
+        if error_name:
+            message += f": {error_name}"
+    else:
+        message = error_name or f"Unknown error (code: {error_code})"
+    
+    # Select exception type based on error code
+    exception_class = InspireFaceError  # Default exception type
+    
+    for category, codes in ERROR_CODE_MAPPING.items():
+        if error_code in codes:
+            if category == 'invalid_input':
+                exception_class = InvalidInputError
+            elif category == 'system_not_ready':
+                exception_class = SystemNotReadyError
+            elif category == 'processing':
+                exception_class = ProcessingError
+            elif category == 'resource':
+                exception_class = ResourceError
+            elif category == 'hardware':
+                exception_class = HardwareError
+            elif category == 'feature_hub':
+                exception_class = FeatureHubError
+            break
+    
+    # Raise corresponding exception
+    raise exception_class(message, error_code, **context)
+
+
+# === Convenient validation functions ===
+
+def validate_image_format(image, operation: str = "Image validation"):
+    """Validate image format"""
+    import numpy as np
+    
+    if not isinstance(image, np.ndarray):
+        raise InvalidInputError(
+            f"{operation}: Input must be a numpy array",
+            errcode.HERR_INVALID_PARAM,
+            input_type=type(image).__name__
+        )
+    
+    if len(image.shape) != 3:
+        raise InvalidInputError(
+            f"{operation}: Image must be 3-dimensional (H, W, C)",
+            errcode.HERR_INVALID_IMAGE_STREAM_PARAM,
+            actual_shape=image.shape
+        )
+    
+    h, w, c = image.shape
+    if c not in [3, 4]:
+        raise InvalidInputError(
+            f"{operation}: Image must have 3 or 4 channels",
+            errcode.HERR_INVALID_IMAGE_STREAM_PARAM,
+            actual_channels=c
+        )
+
+
+def validate_feature_data(data, operation: str = "Feature validation"):
+    """Validate feature data format"""
+    import numpy as np
+    
+    if not isinstance(data, np.ndarray):
+        raise InvalidInputError(
+            f"{operation}: Feature data must be a numpy array",
+            errcode.HERR_INVALID_FACE_FEATURE,
+            input_type=type(data).__name__
+        )
+    
+    if data.dtype != np.float32:
+        raise InvalidInputError(
+            f"{operation}: Feature data must be in float32 format",
+            errcode.HERR_INVALID_FACE_FEATURE,
+            actual_dtype=str(data.dtype)
+        )
+
+
+def validate_session_initialized(session, operation: str = "Session operation"):
+    """Validate if session is initialized"""
+    if session is None or session._sess is None:
+        raise ResourceError(
+            f"{operation}: Session not initialized",
+            errcode.HERR_INVALID_CONTEXT_HANDLE
+        )
+
+
+# === Exception handling decorators for special scenarios ===
+
+def handle_c_api_errors(operation_name: str):
+    """Decorator for wrapping C API calls"""
+    def decorator(func):
+        def wrapper(*args, **kwargs):
+            try:
+                return func(*args, **kwargs)
+            except Exception as e:
+                if not isinstance(e, InspireFaceError):
+                    # Wrap non-InspireFace exceptions as ProcessingError
+                    raise ProcessingError(
+                        f"{operation_name}: {str(e)}",
+                        context={'original_exception': type(e).__name__}
+                    ) from e
+                raise
+        return wrapper
+    return decorator
diff --git a/cpp-package/inspireface/python/inspireface/modules/herror.py b/cpp-package/inspireface/python/inspireface/modules/herror.py
new file mode 100644
index 000000000..b4fdeae83
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/herror.py
@@ -0,0 +1,50 @@
+"""
+InspireFace Error Codes
+
+Auto-generated error code definitions from cpp/inspireface/include/inspireface/herror.h
+Generated on: 2025-07-01
+"""
+
+HSUCCEED = 0  # Success
+HERR_UNKNOWN = 1  # Unknown error (1)
+HERR_INVALID_PARAM = 2  # Invalid parameter (2)
+HERR_INVALID_IMAGE_STREAM_HANDLE = 3  # Invalid image stream handle (3)
+HERR_INVALID_CONTEXT_HANDLE = 4  # Invalid context handle (4)
+HERR_INVALID_FACE_TOKEN = 5  # Invalid face token (5)
+HERR_INVALID_FACE_FEATURE = 6  # Invalid face feature (6)
+HERR_INVALID_FACE_LIST = 7  # Invalid face feature list (7)
+HERR_INVALID_BUFFER_SIZE = 8  # Invalid copy token (8)
+HERR_INVALID_IMAGE_STREAM_PARAM = 9  # Invalid image param (9)
+HERR_INVALID_SERIALIZATION_FAILED = 10  # Invalid face serialization failed (10)
+HERR_INVALID_DETECTION_INPUT = 11  # Failed to modify detector input size (11)
+HERR_INVALID_IMAGE_BITMAP_HANDLE = 12  # Invalid image bitmap handle (12)
+HERR_IMAGE_STREAM_DECODE_FAILED = 13  # ImageStream failed to decode the image (13)
+HERR_SESS_FUNCTION_UNUSABLE = 101  # Function not usable (101)
+HERR_SESS_TRACKER_FAILURE = 102  # Tracker module not initialized (102)
+HERR_SESS_PIPELINE_FAILURE = 103  # Pipeline module not initialized (103)
+HERR_SESS_INVALID_RESOURCE = 104  # Invalid static resource (104)
+HERR_SESS_LANDMARK_NUM_NOT_MATCH = 105  # The number of input landmark points does not match (105)
+HERR_SESS_LANDMARK_NOT_ENABLE = 106  # The landmark model is not enabled (106)
+HERR_SESS_KEY_POINT_NUM_NOT_MATCH = 107  # The number of input key points does not match (107)
+HERR_SESS_REC_EXTRACT_FAILURE = 108  # Face feature extraction not registered (108)
+HERR_SESS_REC_CONTRAST_FEAT_ERR = 109  # Incorrect length of feature vector for comparison (109)
+HERR_SESS_FACE_DATA_ERROR = 110  # Face data parsing (110)
+HERR_SESS_FACE_REC_OPTION_ERROR = 111  # An optional parameter is incorrect (111)
+HERR_FT_HUB_DISABLE = 201  # FeatureHub is disabled (201)
+HERR_FT_HUB_INSERT_FAILURE = 202  # Data insertion error (202)
+HERR_FT_HUB_NOT_FOUND_FEATURE = 203  # Get face feature error (203)
+HERR_ARCHIVE_LOAD_FAILURE = 251  # Archive load failure (251)
+HERR_ARCHIVE_LOAD_MODEL_FAILURE = 252  # Model load failure (252)
+HERR_ARCHIVE_FILE_FORMAT_ERROR = 253  # The archive format is incorrect (253)
+HERR_ARCHIVE_REPETITION_LOAD = 254  # Do not reload the model (254)
+HERR_ARCHIVE_NOT_LOAD = 255  # Model not loaded (255)
+HERR_DEVICE_CUDA_NOT_SUPPORT = 301  # CUDA not supported (301)
+HERR_DEVICE_CUDA_TENSORRT_NOT_SUPPORT = 302  # CUDA TensorRT not supported (302)
+HERR_DEVICE_CUDA_UNKNOWN_ERROR = 303  # CUDA unknown error (303)
+HERR_DEVICE_CUDA_DISABLE = 304  # CUDA support is disabled (304)
+HERR_EXTENSION_ERROR = 351  # Extension module error (351)
+HERR_EXTENSION_MLMODEL_LOAD_FAILED = 352  # MLModel load failed (352)
+HERR_EXTENSION_HETERO_MODEL_TAG_ERROR = 353  # Incorrect heterogeneous model tag (353)
+HERR_EXTENSION_HETERO_REC_HEAD_CONFIG_ERROR = 354  # Rec head config error (354)
+HERR_EXTENSION_HETERO_MODEL_NOT_MATCH = 355  # Heterogeneous model dimensions do not match (355)
+HERR_EXTENSION_HETERO_MODEL_NOT_LOADED = 356  # Heterogeneous model dimensions not loaded (356)
diff --git a/cpp-package/inspireface/python/inspireface/modules/inspireface.py b/cpp-package/inspireface/python/inspireface/modules/inspireface.py
new file mode 100644
index 000000000..3e6d0639a
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/inspireface.py
@@ -0,0 +1,1358 @@
+import ctypes
+
+import numpy as np
+from .core import *
+from typing import Tuple, List
+from dataclasses import dataclass
+from loguru import logger
+from .utils import ResourceManager
+from .utils.resource import set_use_oss_download
+from . import herror as errcode
+# Exception system
+from .exception import (
+    check_error, validate_image_format, validate_feature_data, 
+    validate_session_initialized, handle_c_api_errors,
+    InspireFaceError, InvalidInputError, SystemNotReadyError, 
+    ProcessingError, ResourceError, HardwareError, FeatureHubError
+)
+
+# If True, the latest model will not be verified
+IGNORE_VERIFICATION_OF_THE_LATEST_MODEL = False
+
+def ignore_check_latest_model(ignore: bool):
+    global IGNORE_VERIFICATION_OF_THE_LATEST_MODEL
+    IGNORE_VERIFICATION_OF_THE_LATEST_MODEL = ignore
+
+def use_oss_download(use_oss: bool = True):
+    """Enable OSS download instead of ModelScope (for backward compatibility)
+    
+    Args:
+        use_oss (bool): If True, use OSS download; if False, use ModelScope (default)
+    """
+    set_use_oss_download(use_oss)
+
+class ImageStream(object):
+    """
+    ImageStream class handles the conversion of image data from various sources into a format compatible with the InspireFace library.
+    It allows loading image data from numpy arrays, buffer objects, and directly from OpenCV images.
+    """
+
+    @staticmethod
+    def load_from_cv_image(image: np.ndarray, stream_format=HF_STREAM_BGR, rotation=HF_CAMERA_ROTATION_0):
+        """
+        Load image data from an OpenCV image (numpy ndarray).
+
+        Args:
+            image (np.ndarray): The image data as a numpy array.
+            stream_format (int): The format of the image data (e.g., BGR, RGB).
+            rotation (int): The rotation angle to be applied to the image data.
+
+        Returns:
+            ImageStream: An instance of the ImageStream class initialized with the provided image data.
+
+        Raises:
+            InvalidInputError: If the image does not have 3 or 4 channels.
+        """
+        validate_image_format(image, "Load from CV image")
+        h, w, c = image.shape
+        return ImageStream(image, w, h, stream_format, rotation)
+
+    @staticmethod
+    def load_from_ndarray(data: np.ndarray, width: int, height: int, stream_format: int, rotation: int):
+        """
+        Load image data from a numpy array specifying width and height explicitly.
+
+        Args:
+            data (np.ndarray): The raw image data.
+            width (int): The width of the image.
+            height (int): The height of the image.
+            stream_format (int): The format of the image data.
+            rotation (int): The rotation angle to be applied to the image data.
+
+        Returns:
+            ImageStream: An instance of the ImageStream class.
+        """
+        return ImageStream(data, width, height, stream_format, rotation)
+
+    @staticmethod
+    def load_from_buffer(data, width: int, height: int, stream_format: int, rotation: int):
+        """
+        Load image data from a buffer (like bytes or bytearray).
+
+        Args:
+            data: The buffer containing the image data.
+            width (int): The width of the image.
+            height (int): The height of the image.
+            stream_format (int): The format of the image data.
+            rotation (int): The rotation angle to be applied to the image data.
+
+        Returns:
+            ImageStream: An instance of the ImageStream class.
+        """
+        return ImageStream(data, width, height, stream_format, rotation)
+
+    def __init__(self, data, width: int, height: int, stream_format: int, rotation: int):
+        """
+        Initialize the ImageStream object with provided data and configuration.
+
+        Args:
+            data: The image data (numpy array or buffer).
+            width (int): The width of the image.
+            height (int): The height of the image.
+            stream_format (int): The format of the image data.
+            rotation (int): The rotation applied to the image.
+
+        Raises:
+            ResourceError: If there is an error in creating the image stream.
+        """
+        self.rotate = rotation
+        self.data_format = stream_format
+        if isinstance(data, np.ndarray):
+            data_ptr = ctypes.cast(data.ctypes.data, ctypes.POINTER(ctypes.c_uint8))
+        else:
+            data_ptr = ctypes.cast(data, ctypes.POINTER(ctypes.c_uint8))
+        image_struct = HFImageData()
+        image_struct.data = data_ptr
+        image_struct.width = width
+        image_struct.height = height
+        image_struct.format = self.data_format
+        image_struct.rotation = self.rotate
+        self._handle = HFImageStream()
+        ret = HFCreateImageStream(PHFImageData(image_struct), self._handle)
+        check_error(ret, "Create ImageStream", width=width, height=height, format=stream_format)
+
+    def write_to_file(self, file_path: str):
+        """
+        Write the image stream to a file. Like PATH/image.jpg
+        """
+        ret = HFDeBugImageStreamDecodeSave(self._handle, file_path)
+        check_error(ret, "Write ImageStream to file", file_path=file_path)
+
+    def release(self):
+        """
+        Release the resources associated with the ImageStream.
+
+        Logs an error if the release fails.
+        """
+        if self._handle is not None:
+            ret = HFReleaseImageStream(self._handle)
+            if ret != 0:
+                logger.warning(f"Failed to release ImageStream: error code {ret}")
+
+    def __del__(self):
+        """
+        Ensure that resources are released when the ImageStream object is garbage collected.
+        """
+        self.release()
+
+    def debug_show(self):
+        """
+        Display the image using a debug function provided by the library.
+        """
+        HFDeBugImageStreamImShow(self._handle)
+
+    @property
+    def handle(self):
+        """
+        Return the internal handle of the image stream.
+        Returns:
+        The handle to the internal image stream, used for interfacing with the underlying C/C++ library.
+        """
+        return self._handle
+
+
+# == Session API ==
+
+@dataclass
+class FaceExtended:
+    """
+    A data class to hold extended face information with confidence levels for various attributes.
+
+    Attributes:
+        rgb_liveness_confidence (float): Confidence level of RGB-based liveness detection.
+        mask_confidence (float): Confidence level of mask detection on the face.
+        quality_confidence (float): Confidence level of the overall quality of the face capture.
+    """
+    rgb_liveness_confidence: float
+    mask_confidence: float
+    quality_confidence: float
+    left_eye_status_confidence: float
+    right_eye_status_confidence: float
+    action_normal: int
+    action_jaw_open: int
+    action_shake: int
+    action_blink: int
+    action_head_raise: int
+    race: int
+    gender: int
+    age_bracket: int
+    emotion: int
+
+    def __repr__(self) -> str:
+        return f"FaceExtended(rgb_liveness_confidence={self.rgb_liveness_confidence}, mask_confidence={self.mask_confidence}, quality_confidence={self.quality_confidence}, left_eye_status_confidence={self.left_eye_status_confidence}, right_eye_status_confidence={self.right_eye_status_confidence}, action_normal={self.action_normal}, action_jaw_open={self.action_jaw_open}, action_shake={self.action_shake}, action_blink={self.action_blink}, action_head_raise={self.action_head_raise}, race={self.race}, gender={self.gender}, age_bracket={self.age_bracket}, emotion={self.emotion})"
+
+
+class FaceInformation:
+    """
+    Holds detailed information about a detected face including location and orientation.
+
+    Attributes:
+        track_id (int): Unique identifier for tracking the face across frames.
+        location (Tuple): Coordinates of the face in the form (x, y, width, height).
+        roll (float): Roll angle of the face.
+        yaw (float): Yaw angle of the face.
+        pitch (float): Pitch angle of the face.
+        _token (HFFaceBasicToken): A token containing low-level details about the face.
+        _feature (np.array, optional): An optional numpy array holding the facial feature data.
+
+    Methods:
+        __init__: Initializes a new instance of FaceInformation.
+    """
+
+    def __init__(self,
+                 track_id: int,
+                 track_count: int,
+                 detection_confidence: float,
+                 location: Tuple,
+                 roll: float,
+                 yaw: float,
+                 pitch: float,
+                 _token: HFFaceBasicToken,
+                 _feature: np.array = None):
+        self.track_id = track_id
+        self.track_count = track_count
+        self.detection_confidence = detection_confidence
+        self.location = location
+        self.roll = roll
+        self.yaw = yaw
+        self.pitch = pitch
+
+        # Calculate the required buffer size for the face token and copy it.
+        token_size = HInt32()
+        HFGetFaceBasicTokenSize(HPInt32(token_size))
+        buffer_size = token_size.value
+        self.buffer = create_string_buffer(buffer_size)
+        ret = HFCopyFaceBasicToken(_token, self.buffer, token_size)
+        check_error(ret, "Copy face basic token", track_id=track_id)
+
+        # Store the copied token.
+        self._token = HFFaceBasicToken()
+        self._token.size = buffer_size
+        self._token.data = cast(addressof(self.buffer), c_void_p)
+
+    def __repr__(self) -> str:
+        return f"FaceInformation(track_id={self.track_id}, track_count={self.track_count}, detection_confidence={self.detection_confidence}, location={self.location}, roll={self.roll}, yaw={self.yaw}, pitch={self.pitch})"
+
+
+@dataclass
+class SessionCustomParameter:
+    """
+    A data class for configuring the optional parameters in a face recognition session.
+
+    Attributes are set to False by default and can be enabled as needed.
+
+    Methods:
+        _c_struct: Converts the Python attributes to a C-compatible structure for session configuration.
+    """
+    enable_recognition: bool = False
+    enable_liveness: bool = False
+    enable_ir_liveness: bool = False
+    enable_mask_detect: bool = False
+    enable_face_attribute: bool = False
+    enable_face_quality: bool = False
+    enable_interaction_liveness: bool = False
+    enable_face_emotion: bool = False
+
+    def _c_struct(self):
+        """
+        Creates a C structure from the current state of the instance.
+
+        Returns:
+            HFSessionCustomParameter: The corresponding C structure with proper type conversions.
+        """
+        custom_param = HFSessionCustomParameter(
+            enable_recognition=int(self.enable_recognition),
+            enable_liveness=int(self.enable_liveness),
+            enable_ir_liveness=int(self.enable_ir_liveness),
+            enable_mask_detect=int(self.enable_mask_detect),
+            enable_face_attribute=int(self.enable_face_attribute),
+            enable_face_quality=int(self.enable_face_quality),
+            enable_interaction_liveness=int(self.enable_interaction_liveness),
+            enable_face_emotion=int(self.enable_face_emotion)
+        )
+
+        return custom_param
+
+    def __repr__(self) -> str:
+        return f"SessionCustomParameter(enable_recognition={self.enable_recognition}, enable_liveness={self.enable_liveness}, enable_ir_liveness={self.enable_ir_liveness}, enable_mask_detect={self.enable_mask_detect}, enable_face_attribute={self.enable_face_attribute}, enable_face_quality={self.enable_face_quality}, enable_interaction_liveness={self.enable_interaction_liveness}, enable_face_emotion={self.enable_face_emotion})"
+
+
+class InspireFaceSession(object):
+    """
+    Manages a session for face detection and recognition processes using the InspireFace library.
+
+    Attributes:
+        multiple_faces (HFMultipleFaceData): Stores data about multiple detected faces during the session.
+        _sess (HFSession): The handle to the underlying library session.
+        param (int or SessionCustomParameter): Configuration parameters or flags for the session.
+
+    """
+
+    def __init__(self, param, detect_mode: int = HF_DETECT_MODE_ALWAYS_DETECT,
+                 max_detect_num: int = 10, detect_pixel_level=-1, track_by_detect_mode_fps=-1):
+        """
+        Initializes a new session with the provided configuration parameters.
+        
+        Args:
+            param (int or SessionCustomParameter): Configuration parameters or flags.
+            detect_mode (int): Detection mode to be used (e.g., image-based detection).
+            max_detect_num (int): Maximum number of faces to detect.
+            
+        Raises:
+            SystemNotReadyError: If InspireFace is not launched.
+            ProcessingError: If session creation fails.
+        """
+        # Initialize _sess to None first to prevent AttributeError in __del__
+        self._sess = None
+        self.multiple_faces = None
+        self.param = param
+        
+        # If InspireFace is not initialized, run launch() use Pikachu model
+        if not query_launch_status():
+            ret = launch()
+            if not ret:
+                raise SystemNotReadyError("Failed to launch InspireFace automatically")
+
+        self._sess = HFSession()
+        
+        if isinstance(self.param, SessionCustomParameter):
+            ret = HFCreateInspireFaceSession(self.param._c_struct(), detect_mode, max_detect_num, detect_pixel_level,
+                                             track_by_detect_mode_fps, self._sess)
+        elif isinstance(self.param, int):
+            ret = HFCreateInspireFaceSessionOptional(self.param, detect_mode, max_detect_num, detect_pixel_level,
+                                                     track_by_detect_mode_fps, self._sess)
+        else:
+            raise InvalidInputError("Session parameter must be SessionCustomParameter or int", 
+                                   context={'param_type': type(self.param).__name__})
+        
+        check_error(ret, "Create InspireFace session", 
+                   detect_mode=detect_mode, max_detect_num=max_detect_num)
+
+    @handle_c_api_errors("Face detection")
+    def face_detection(self, image) -> List[FaceInformation]:
+        """
+        Detects faces in the given image and returns a list of FaceInformation objects containing detailed face data.
+        
+        Args:
+            image (np.ndarray or ImageStream): The image in which to detect faces.
+            
+        Returns:
+            List[FaceInformation]: A list of detected face information.
+            
+        Raises:
+            ResourceError: If session is not initialized.
+            ProcessingError: If face detection fails.
+        """
+        validate_session_initialized(self, "Face detection")
+        stream = self._get_image_stream(image)
+        self.multiple_faces = HFMultipleFaceData()
+        ret = HFExecuteFaceTrack(self._sess, stream.handle,
+                                PHFMultipleFaceData(self.multiple_faces))
+        check_error(ret, "Execute face tracking")
+
+        if self.multiple_faces.detectedNum > 0:
+            boxes = self._get_faces_boundary_boxes()
+            track_ids = self._get_faces_track_ids()
+            euler_angle = self._get_faces_euler_angle()
+            tokens = self._get_faces_tokens()
+            track_counts = self._get_faces_track_counts()
+
+            infos = list()
+            for idx in range(self.multiple_faces.detectedNum):
+                top_left = (boxes[idx][0], boxes[idx][1])
+                bottom_right = (boxes[idx][0] + boxes[idx][2], boxes[idx][1] + boxes[idx][3])
+                roll = euler_angle[idx][0]
+                yaw = euler_angle[idx][1]
+                pitch = euler_angle[idx][2]
+                track_id = track_ids[idx]
+                _token = tokens[idx]
+                detection_confidence = self.multiple_faces.detConfidence[idx]
+                track_count = track_counts[idx]
+
+                info = FaceInformation(
+                    location=(top_left[0], top_left[1], bottom_right[0], bottom_right[1]),
+                    roll=roll,
+                    yaw=yaw,
+                    pitch=pitch,
+                    track_id=track_id,
+                    track_count=track_count,
+                    _token=_token,
+                    detection_confidence=detection_confidence,
+                )
+                infos.append(info)
+
+            return infos
+        else:
+            return []
+        
+    def get_face_five_key_points(self, single_face: FaceInformation):
+        """Get five key points for a detected face"""
+        validate_session_initialized(self, "Get face five key points")
+        num_landmarks = 5
+        landmarks_array = (HPoint2f * num_landmarks)()
+        ret = HFGetFaceFiveKeyPointsFromFaceToken(single_face._token, landmarks_array, num_landmarks)
+        check_error(ret, "Get face five key points", track_id=single_face.track_id)
+
+        landmark = []
+        for point in landmarks_array:
+            landmark.append(point.x)
+            landmark.append(point.y)
+
+        return np.asarray(landmark).reshape(-1, 2)
+
+    def get_face_dense_landmark(self, single_face: FaceInformation):
+        """Get dense landmarks for a detected face"""
+        validate_session_initialized(self, "Get face dense landmark")
+        num_landmarks = HInt32()
+        HFGetNumOfFaceDenseLandmark(byref(num_landmarks))
+        landmarks_array = (HPoint2f * num_landmarks.value)()
+        ret = HFGetFaceDenseLandmarkFromFaceToken(single_face._token, landmarks_array, num_landmarks)
+        check_error(ret, "Get face dense landmark", track_id=single_face.track_id)
+
+        landmark = []
+        for point in landmarks_array:
+            landmark.append(point.x)
+            landmark.append(point.y)
+
+        return np.asarray(landmark).reshape(-1, 2)
+    
+    def print_track_cost_spend(self):
+        """Print tracking cost statistics"""
+        validate_session_initialized(self, "Print track cost spend")
+        ret = HFSessionPrintTrackCostSpend(self._sess)
+        check_error(ret, "Print track cost spend")
+
+    def set_enable_track_cost_spend(self, enable: bool):
+        """Enable or disable track cost spend monitoring"""
+        validate_session_initialized(self, "Set enable track cost spend")
+        ret = HFSessionSetEnableTrackCostSpend(self._sess, enable)
+        check_error(ret, "Set enable track cost spend", enable=enable)
+    
+    def set_detection_confidence_threshold(self, threshold: float):
+        """
+        Sets the detection confidence threshold for the face detection session.
+
+        Args:
+            threshold (float): The confidence threshold for face detection.
+        """
+        validate_session_initialized(self, "Set detection confidence threshold")
+        ret = HFSessionSetFaceDetectThreshold(self._sess, threshold)
+        check_error(ret, "Set detection confidence threshold", threshold=threshold)
+
+    def set_track_preview_size(self, size=192):
+        """
+        Sets the preview size for the face tracking session.
+
+        Args:
+            size (int, optional): The size of the preview area for face tracking. Default is 192.
+        """
+        validate_session_initialized(self, "Set track preview size")
+        ret = HFSessionSetTrackPreviewSize(self._sess, size)
+        check_error(ret, "Set track preview size", size=size)
+
+    def set_filter_minimum_face_pixel_size(self, min_size=32):
+        """Set minimum face pixel size filter"""
+        validate_session_initialized(self, "Set filter minimum face pixel size")
+        ret = HFSessionSetFilterMinimumFacePixelSize(self._sess, min_size)
+        check_error(ret, "Set filter minimum face pixel size", min_size=min_size)
+
+    def set_track_mode_smooth_ratio(self, ratio=0.025):
+        """Set track mode smooth ratio"""
+        validate_session_initialized(self, "Set track mode smooth ratio")
+        ret = HFSessionSetTrackModeSmoothRatio(self._sess, ratio)
+        check_error(ret, "Set track mode smooth ratio", ratio=ratio)
+
+    def set_track_mode_num_smooth_cache_frame(self, num=15):
+        """Set track mode number of smooth cache frames"""
+        validate_session_initialized(self, "Set track mode num smooth cache frame")
+        ret = HFSessionSetTrackModeNumSmoothCacheFrame(self._sess, num)
+        check_error(ret, "Set track mode num smooth cache frame", num=num)
+
+    def set_track_model_detect_interval(self, num=20):
+        """Set track model detect interval"""
+        validate_session_initialized(self, "Set track model detect interval")
+        ret = HFSessionSetTrackModeDetectInterval(self._sess, num)
+        check_error(ret, "Set track model detect interval", num=num)
+
+    def set_landmark_augmentation_num(self, num=1):
+        """Set landmark augmentation number"""
+        validate_session_initialized(self, "Set landmark augmentation num")
+        ret = HFSessionSetLandmarkAugmentationNum(self._sess, num)
+        check_error(ret, "Set landmark augmentation num", num=num)
+
+    def set_track_lost_recovery_mode(self, value=False):
+        """Set track lost recovery mode"""
+        validate_session_initialized(self, "Set track lost recovery mode")
+        ret = HFSessionSetTrackLostRecoveryMode(self._sess, value)
+        check_error(ret, "Set track lost recovery mode", value=value)
+
+    @handle_c_api_errors("Face pipeline processing")
+    def face_pipeline(self, image, faces: List[FaceInformation], exec_param) -> List[FaceExtended]:
+        """
+        Processes detected faces to extract additional attributes based on the provided execution parameters.
+
+        Args:
+            image (np.ndarray or ImageStream): The image from which faces are detected.
+            faces (List[FaceInformation]): A list of FaceInformation objects containing detected face data.
+            exec_param (SessionCustomParameter or int): Custom parameters for processing faces.
+
+        Returns:
+            List[FaceExtended]: A list of FaceExtended objects with updated attributes like mask confidence, liveness, etc.
+        """
+        validate_session_initialized(self, "Face pipeline processing")
+        stream = self._get_image_stream(image)
+        fn, pm, flag = self._get_processing_function_and_param(exec_param)
+        tokens = [face._token for face in faces]
+        tokens_array = (HFFaceBasicToken * len(tokens))(*tokens)
+        tokens_ptr = cast(tokens_array, PHFFaceBasicToken)
+
+        multi_faces = HFMultipleFaceData()
+        multi_faces.detectedNum = len(tokens)
+        multi_faces.tokens = tokens_ptr
+        ret = fn(self._sess, stream.handle, PHFMultipleFaceData(multi_faces), pm)
+
+        check_error(ret, "Face pipeline processing", num_faces=len(faces))
+
+        extends = [FaceExtended(-1.0, -1.0, -1.0, -1.0, -1.0, 0, 0, 0, 0, 0, -1, -1, -1, -1) for _ in range(len(faces))]
+        self._update_mask_confidence(exec_param, flag, extends)
+        self._update_rgb_liveness_confidence(exec_param, flag, extends)
+        self._update_face_quality_confidence(exec_param, flag, extends)
+        self._update_face_attribute_confidence(exec_param, flag, extends)
+        self._update_face_interact_confidence(exec_param, flag, extends)
+        self._update_face_emotion_confidence(exec_param, flag, extends)
+
+        return extends
+
+    @handle_c_api_errors("Face feature extraction")
+    def face_feature_extract(self, image, face_information: FaceInformation):
+        """
+        Extracts facial features from a specified face within an image for recognition or comparison purposes.
+
+        Args:
+            image (np.ndarray or ImageStream): The image from which the face features are to be extracted.
+            face_information (FaceInformation): The FaceInformation object containing the details of the face.
+
+        Returns:
+            np.ndarray: A numpy array containing the extracted facial features, or None if the extraction fails.
+        """
+        validate_session_initialized(self, "Face feature extraction")
+        stream = self._get_image_stream(image)
+        feature_length = HInt32()
+        HFGetFeatureLength(byref(feature_length))
+
+        feature = np.zeros((feature_length.value,), dtype=np.float32)
+        ret = HFFaceFeatureExtractCpy(self._sess, stream.handle, face_information._token,
+                                      feature.ctypes.data_as(ctypes.POINTER(HFloat)))
+
+        check_error(ret, "Face feature extraction", track_id=face_information.track_id)
+        return feature
+
+    @staticmethod
+    def _get_image_stream(image):
+        """Convert image to ImageStream if needed"""
+        if isinstance(image, np.ndarray):
+            return ImageStream.load_from_cv_image(image)
+        elif isinstance(image, ImageStream):
+            return image
+        else:
+            raise InvalidInputError("Image must be numpy.ndarray or ImageStream", 
+                                   context={'input_type': type(image).__name__})
+
+    @staticmethod
+    def _get_processing_function_and_param(exec_param):
+        """Get processing function and parameters"""
+        if isinstance(exec_param, SessionCustomParameter):
+            return HFMultipleFacePipelineProcess, exec_param._c_struct(), "object"
+        elif isinstance(exec_param, int):
+            return HFMultipleFacePipelineProcessOptional, exec_param, "bitmask"
+        else:
+            raise InvalidInputError("exec_param must be SessionCustomParameter or int",
+                                   context={'param_type': type(exec_param).__name__})
+
+    def _update_mask_confidence(self, exec_param, flag, extends):
+        """Update mask confidence in extends list"""
+        if (flag == "object" and exec_param.enable_mask_detect) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_MASK_DETECT):
+            mask_results = HFFaceMaskConfidence()
+            ret = HFGetFaceMaskConfidence(self._sess, PHFFaceMaskConfidence(mask_results))
+            if ret == errcode.HSUCCEED:
+                for i in range(mask_results.num):
+                    extends[i].mask_confidence = mask_results.confidence[i]
+            else:
+                logger.warning(f"Failed to get mask confidence: error code {ret}")
+
+    def _update_face_interact_confidence(self, exec_param, flag, extends):
+        """Update face interaction confidence in extends list"""
+        if (flag == "object" and exec_param.enable_interaction_liveness) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_INTERACTION):
+            results = HFFaceInteractionState()
+            ret = HFGetFaceInteractionStateResult(self._sess, PHFFaceInteractionState(results))
+            if ret == errcode.HSUCCEED:
+                for i in range(results.num):
+                    extends[i].left_eye_status_confidence = results.leftEyeStatusConfidence[i]
+                    extends[i].right_eye_status_confidence = results.rightEyeStatusConfidence[i]
+            else:
+                logger.warning(f"Failed to get face interaction state: error code {ret}")
+                
+            actions = HFFaceInteractionsActions()
+            ret = HFGetFaceInteractionActionsResult(self._sess, PHFFaceInteractionsActions(actions))
+            if ret == errcode.HSUCCEED:
+                for i in range(results.num):
+                    extends[i].action_normal = actions.normal[i]
+                    extends[i].action_shake = actions.shake[i]
+                    extends[i].action_jaw_open = actions.jawOpen[i]
+                    extends[i].action_head_raise = actions.headRaise[i]
+                    extends[i].action_blink = actions.blink[i]
+            else:
+                logger.warning(f"Failed to get face interaction actions: error code {ret}")
+
+    def _update_face_emotion_confidence(self, exec_param, flag, extends):
+        """Update face emotion confidence in extends list"""
+        if (flag == "object" and exec_param.enable_face_emotion) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_FACE_EMOTION):
+            emotion_results = HFFaceEmotionResult()
+            ret = HFGetFaceEmotionResult(self._sess, PHFFaceEmotionResult(emotion_results))
+            if ret == errcode.HSUCCEED:
+                for i in range(emotion_results.num):
+                    extends[i].emotion = emotion_results.emotion[i]
+            else:
+                logger.warning(f"Failed to get face emotion result: error code {ret}")
+
+    def _update_rgb_liveness_confidence(self, exec_param, flag, extends: List[FaceExtended]):
+        """Update RGB liveness confidence in extends list"""
+        if (flag == "object" and exec_param.enable_liveness) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_LIVENESS):
+            liveness_results = HFRGBLivenessConfidence()
+            ret = HFGetRGBLivenessConfidence(self._sess, PHFRGBLivenessConfidence(liveness_results))
+            if ret == errcode.HSUCCEED:
+                for i in range(liveness_results.num):
+                    extends[i].rgb_liveness_confidence = liveness_results.confidence[i]
+            else:
+                logger.warning(f"Failed to get RGB liveness confidence: error code {ret}")
+
+    def _update_face_attribute_confidence(self, exec_param, flag, extends: List[FaceExtended]):
+        """Update face attribute confidence in extends list"""
+        if (flag == "object" and exec_param.enable_face_attribute) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_FACE_ATTRIBUTE):
+            attribute_results = HFFaceAttributeResult()
+            ret = HFGetFaceAttributeResult(self._sess, PHFFaceAttributeResult(attribute_results))
+            if ret == errcode.HSUCCEED:
+                for i in range(attribute_results.num):
+                    extends[i].gender = attribute_results.gender[i]
+                    extends[i].age_bracket = attribute_results.ageBracket[i]
+                    extends[i].race = attribute_results.race[i]
+            else:
+                logger.warning(f"Failed to get face attribute result: error code {ret}")
+
+    def _update_face_quality_confidence(self, exec_param, flag, extends: List[FaceExtended]):
+        """Update face quality confidence in extends list"""
+        if (flag == "object" and exec_param.enable_face_quality) or (
+                flag == "bitmask" and exec_param & HF_ENABLE_QUALITY):
+            quality_results = HFFaceQualityConfidence()
+            ret = HFGetFaceQualityConfidence(self._sess, PHFFaceQualityConfidence(quality_results))
+            if ret == errcode.HSUCCEED:
+                for i in range(quality_results.num):
+                    extends[i].quality_confidence = quality_results.confidence[i]
+            else:
+                logger.warning(f"Failed to get face quality confidence: error code {ret}")
+
+    def _get_faces_boundary_boxes(self) -> List:
+        """Get face boundary boxes from detection results"""
+        num_of_faces = self.multiple_faces.detectedNum
+        rects_ptr = self.multiple_faces.rects
+        rects = [(rects_ptr[i].x, rects_ptr[i].y, rects_ptr[i].width, rects_ptr[i].height) for i in range(num_of_faces)]
+        return rects
+
+    def _get_faces_track_ids(self) -> List:
+        """Get face track IDs from detection results"""
+        num_of_faces = self.multiple_faces.detectedNum
+        track_ids_ptr = self.multiple_faces.trackIds
+        track_ids = [track_ids_ptr[i] for i in range(num_of_faces)]
+        return track_ids
+
+    def _get_faces_euler_angle(self) -> List:
+        """Get face euler angles from detection results"""
+        num_of_faces = self.multiple_faces.detectedNum
+        euler_angle = self.multiple_faces.angles
+        angles = [(euler_angle.roll[i], euler_angle.yaw[i], euler_angle.pitch[i]) for i in range(num_of_faces)]
+        return angles
+     
+    def _get_faces_track_counts(self) -> List:
+        """Get face track counts from detection results"""
+        num_of_faces = self.multiple_faces.detectedNum
+        track_counts_ptr = self.multiple_faces.trackCounts
+        track_counts = [track_counts_ptr[i] for i in range(num_of_faces)]
+        return track_counts
+
+    def _get_faces_tokens(self) -> List[HFFaceBasicToken]:
+        """Get face tokens from detection results"""
+        num_of_faces = self.multiple_faces.detectedNum
+        tokens_ptr = self.multiple_faces.tokens
+        tokens = [tokens_ptr[i] for i in range(num_of_faces)]
+        return tokens
+
+    def release(self):
+        """Release session resources"""
+        if self._sess is not None:
+            HFReleaseInspireFaceSession(self._sess)
+            self._sess = None
+
+    def __del__(self):
+        self.release()
+
+
+# == Global API ==
+
+def _check_modelscope_availability():
+    """
+    Check if ModelScope is available when needed and provide helpful error message if not.
+    
+    Exits the program if ModelScope is needed but not available and OSS is not enabled.
+    """
+    import sys
+    from .utils.resource import USE_OSS_DOWNLOAD
+    
+    # Dynamic check for ModelScope availability (don't rely on cached MODELSCOPE_AVAILABLE)
+    modelscope_available = True
+    try:
+        from modelscope.hub.snapshot_download import snapshot_download
+        print("ModelScope import successful")
+    except Exception as e:
+        modelscope_available = False
+        print(f"ModelScope import failed: {e}")
+    
+    if not USE_OSS_DOWNLOAD and not modelscope_available:
+        print("ModelScope is not available, cannot download models!")
+        print("\nPlease choose one of the following solutions:")
+        print("1. Reinstall ModelScope with all dependencies:")
+        print("   pip install --upgrade modelscope")
+        print("\n2. Install missing dependencies manually:")
+        print("   pip install filelock")
+        print("\n3. Switch to OSS download mode:")
+        print("   import inspireface as isf")
+        print("   isf.use_oss_download(True)  # Execute before calling launch()")
+        print("\nNote: OSS download requires stable international network connection")
+        sys.exit(1)
+
+def launch(model_name: str = "Pikachu", resource_path: str = None) -> bool:
+    """
+    Launches the InspireFace system with the specified resource directory.
+
+    Args:
+        model_name (str): the name of the model to use.
+        resource_path (str): if None, use the default model path.
+
+    Returns:
+        bool: True if the system was successfully launched, False otherwise.
+
+    Raises:
+        SystemNotReadyError: If launch fails due to resource issues.
+    """
+    if resource_path is None:
+        from .utils.resource import USE_OSS_DOWNLOAD
+        
+        # Check if ModelScope is available when needed
+        _check_modelscope_availability()
+        
+        # Use ModelScope by default unless OSS is forced
+        sm = ResourceManager(use_modelscope=not USE_OSS_DOWNLOAD)
+        resource_path = sm.get_model(model_name, ignore_verification=IGNORE_VERIFICATION_OF_THE_LATEST_MODEL)
+    path_c = String(bytes(resource_path, encoding="utf8"))
+    ret = HFLaunchInspireFace(path_c)
+    if ret != 0:
+        if ret == errcode.HERR_ARCHIVE_REPETITION_LOAD:
+            logger.warning("Duplicate loading was found")
+            return True
+        else:
+            check_error(ret, "Launch InspireFace", model_name=model_name, resource_path=resource_path)
+    return True
+
+def pull_latest_model(model_name: str = "Pikachu") -> str:
+    """
+    Pulls the latest model from the resource manager.
+
+    Args:
+        model_name (str): the name of the model to use.
+
+    Returns:
+        str: Path to the downloaded model.
+    """
+    from .utils.resource import USE_OSS_DOWNLOAD
+    
+    # Check if ModelScope is available when needed
+    _check_modelscope_availability()
+    
+    sm = ResourceManager(use_modelscope=not USE_OSS_DOWNLOAD)
+    resource_path = sm.get_model(model_name, re_download=True)
+    return resource_path
+
+def reload(model_name: str = "Pikachu", resource_path: str = None) -> bool:
+    """
+    Reloads the InspireFace system with the specified resource directory.
+
+    Args:
+        model_name (str): the name of the model to use.
+        resource_path (str): if None, use the default model path.
+
+    Returns:
+        bool: True if reload was successful.
+    """
+    if resource_path is None:
+        from .utils.resource import USE_OSS_DOWNLOAD
+        
+        # Check if ModelScope is available when needed
+        _check_modelscope_availability()
+        
+        sm = ResourceManager(use_modelscope=not USE_OSS_DOWNLOAD)
+        resource_path = sm.get_model(model_name, ignore_verification=IGNORE_VERIFICATION_OF_THE_LATEST_MODEL)
+    path_c = String(bytes(resource_path, encoding="utf8"))
+    ret = HFReloadInspireFace(path_c)
+    if ret != 0:
+        if ret == errcode.HERR_ARCHIVE_REPETITION_LOAD:
+            logger.warning("Duplicate loading was found")
+            return True
+        else:
+            check_error(ret, "Reload InspireFace", model_name=model_name, resource_path=resource_path)
+    return True
+
+def terminate() -> bool:
+    """
+    Terminates the InspireFace system.
+
+    Returns:
+        bool: True if the system was successfully terminated, False otherwise.
+    """
+    ret = HFTerminateInspireFace()
+    check_error(ret, "Terminate InspireFace")
+    return True
+
+def query_launch_status() -> bool:
+    """
+    Queries the launch status of the InspireFace SDK.
+
+    Returns:
+        bool: True if InspireFace is launched, False otherwise.
+    """
+    status = HInt32()
+    ret = HFQueryInspireFaceLaunchStatus(byref(status))
+    check_error(ret, "Query launch status")
+    return status.value == 1
+
+def switch_landmark_engine(engine: int):
+    """Switch landmark engine"""
+    ret = HFSwitchLandmarkEngine(engine)
+    check_error(ret, "Switch landmark engine", engine=engine)
+    return True
+
+def switch_image_processing_backend(backend: int):
+    """Switch image processing backend"""
+    ret = HFSwitchImageProcessingBackend(backend)
+    check_error(ret, "Switch image processing backend", backend=backend)
+    return True
+
+def set_image_process_aligned_width(width: int):
+    """Set the image process aligned width"""
+    ret = HFSetImageProcessAlignedWidth(width)
+    check_error(ret, "Set image process aligned width", width=width)
+    return True
+
+@dataclass
+class FeatureHubConfiguration:
+    """
+    Configuration settings for managing the feature hub, including database and search settings.
+
+    Attributes:
+        primary_key_mode (int): Primary key mode for the database.
+        enable_persistence (bool): Flag to indicate if the database should be used.
+        persistence_db_path (str): Path to the database file.
+        search_threshold (float): The threshold value for considering a match.
+        search_mode (int): The mode of searching in the database.
+    """
+    primary_key_mode: int
+    enable_persistence: bool
+    persistence_db_path: str
+    search_threshold: float
+    search_mode: int
+
+    def _c_struct(self):
+        """
+        Converts the data class attributes to a C-compatible structure for use in the InspireFace SDK.
+
+        Returns:
+            HFFeatureHubConfiguration: A C-structure for feature hub configuration.
+        """
+        return HFFeatureHubConfiguration(
+            primaryKeyMode=self.primary_key_mode,
+            enablePersistence=int(self.enable_persistence),
+            persistenceDbPath=String(bytes(self.persistence_db_path, encoding="utf8")),
+            searchThreshold=self.search_threshold,
+            searchMode=self.search_mode
+        )
+
+
+def feature_hub_enable(config: FeatureHubConfiguration) -> bool:
+    """
+    Enables the feature hub with the specified configuration.
+
+    Args:
+        config (FeatureHubConfiguration): Configuration settings for the feature hub.
+
+    Returns:
+        bool: True if successfully enabled, False otherwise.
+    """
+    ret = HFFeatureHubDataEnable(config._c_struct())
+    check_error(ret, "Enable FeatureHub")
+    return True
+
+
+def feature_hub_disable() -> bool:
+    """
+    Disables the feature hub.
+
+    Returns:
+        bool: True if successfully disabled, False otherwise.
+    """
+    ret = HFFeatureHubDataDisable()
+    check_error(ret, "Disable FeatureHub")
+    return True
+
+
+def feature_comparison(feature1: np.ndarray, feature2: np.ndarray) -> float:
+    """
+    Compares two facial feature arrays to determine their similarity.
+
+    Args:
+        feature1 (np.ndarray): The first feature array.
+        feature2 (np.ndarray): The second feature array.
+
+    Returns:
+        float: A similarity score, where -1.0 indicates an error during comparison.
+    """
+    validate_feature_data(feature1, "Feature comparison")
+    validate_feature_data(feature2, "Feature comparison")
+    faces = [feature1, feature2]
+    feats = []
+    for face in faces:
+        feature = HFFaceFeature()
+        data_ptr = face.ctypes.data_as(HPFloat)
+        feature.size = HInt32(face.size)
+        feature.data = data_ptr
+        feats.append(feature)
+
+    comparison_result = HFloat()
+    ret = HFFaceComparison(feats[0], feats[1], HPFloat(comparison_result))
+    check_error(ret, "Face feature comparison")
+    return float(comparison_result.value)
+
+
+class FaceIdentity(object):
+    """
+    Represents an identity based on facial features, associating the features with a custom ID and a tag.
+
+    Attributes:
+        feature (np.ndarray): The facial features as a numpy array.
+        id (int): A custom identifier for the face identity.
+
+    Methods:
+        __init__: Initializes a new instance of FaceIdentity.
+        from_ctypes: Converts a C structure to a FaceIdentity instance.
+        _c_struct: Converts the instance back to a compatible C structure.
+    """
+
+    def __init__(self, data: np.ndarray, id: int):
+        """
+        Initializes a new FaceIdentity instance with facial feature data, a custom identifier, and a tag.
+
+        Args:
+            data (np.ndarray): The facial feature data.
+            id (int): A custom identifier for tracking or referencing the face identity.
+        """
+        validate_feature_data(data, "FaceIdentity initialization")
+        self.feature = data
+        self.id = id
+
+    def __repr__(self) -> str:
+        return f"FaceIdentity(id={self.id}, feature={self.feature})"
+
+    @staticmethod
+    def from_ctypes(raw_identity: HFFaceFeatureIdentity):
+        """
+        Converts a ctypes structure representing a face identity into a FaceIdentity object.
+
+        Args:
+            raw_identity (HFFaceFeatureIdentity): The ctypes structure containing the face identity data.
+
+        Returns:
+            FaceIdentity: An instance of FaceIdentity with data extracted from the ctypes structure.
+        """
+        feature_size = raw_identity.feature.contents.size
+        feature_data_ptr = raw_identity.feature.contents.data
+        feature_data = np.ctypeslib.as_array(cast(feature_data_ptr, HPFloat), (feature_size,))
+        id_ = raw_identity.id
+
+        return FaceIdentity(data=feature_data, id=id_)
+
+    def _c_struct(self):
+        """
+        Converts this FaceIdentity instance into a C-compatible structure for use with InspireFace APIs.
+
+        Returns:
+            HFFaceFeatureIdentity: A C structure representing this face identity.
+        """
+        feature = HFFaceFeature()
+        data_ptr = self.feature.ctypes.data_as(HPFloat)
+        feature.size = HInt32(self.feature.size)
+        feature.data = data_ptr
+        return HFFaceFeatureIdentity(
+            id=HFaceId(self.id),
+            feature=PHFFaceFeature(feature)
+        )
+
+
+def feature_hub_set_search_threshold(threshold: float):
+    """
+    Sets the search threshold for face matching in the FeatureHub.
+
+    Args:
+        threshold (float): The similarity threshold for determining a match.
+    """
+    HFFeatureHubFaceSearchThresholdSetting(threshold)
+
+
+def feature_hub_face_insert(face_identity: FaceIdentity) -> Tuple[bool, int]:
+    """
+    Inserts a face identity into the FeatureHub database.
+
+    Args:
+        face_identity (FaceIdentity): The face identity to insert.
+
+    Returns:
+        Tuple[bool, int]: (True, allocated_id) if the face identity was successfully inserted.
+    """
+    alloc_id = HFaceId()
+    ret = HFFeatureHubInsertFeature(face_identity._c_struct(), HPFaceId(alloc_id))
+    check_error(ret, "Insert face feature into FeatureHub", identity_id=face_identity.id)
+    return True, int(alloc_id.value)
+
+
+@dataclass
+class SearchResult:
+    """
+    Represents the result of a face search operation with confidence level and the most similar face identity found.
+
+    Attributes:
+        confidence (float): The confidence score of the search result, indicating the similarity.
+        similar_identity (FaceIdentity): The face identity that most closely matches the search query.
+    """
+    confidence: float
+    similar_identity: FaceIdentity
+
+    def __repr__(self) -> str:
+        return f"SearchResult(confidence={self.confidence}, similar_identity={self.similar_identity})"
+
+def feature_hub_face_search(data: np.ndarray) -> SearchResult:
+    """
+    Searches for the most similar face identity in the feature hub based on provided facial features.
+
+    Args:
+        data (np.ndarray): The facial feature data to search for.
+
+    Returns:
+        SearchResult: The search result containing the confidence and the most similar identity found.
+    """
+    validate_feature_data(data, "FeatureHub face search")
+    feature = HFFaceFeature(size=HInt32(data.size), data=data.ctypes.data_as(HPFloat))
+    confidence = HFloat()
+    most_similar = HFFaceFeatureIdentity()
+    ret = HFFeatureHubFaceSearch(feature, HPFloat(confidence), PHFFaceFeatureIdentity(most_similar))
+    check_error(ret, "Search face in FeatureHub")
+    
+    if most_similar.id != -1:
+        search_identity = FaceIdentity.from_ctypes(most_similar)
+        return SearchResult(confidence=confidence.value, similar_identity=search_identity)
+    else:
+        none = FaceIdentity(np.zeros(0, dtype=np.float32), most_similar.id)
+        return SearchResult(confidence=confidence.value, similar_identity=none)
+
+
+def feature_hub_face_search_top_k(data: np.ndarray, top_k: int) -> List[Tuple]:
+    """
+    Searches for the top 'k' most similar face identities in the feature hub based on provided facial features.
+
+    Args:
+        data (np.ndarray): The facial feature data to search for.
+        top_k (int): The number of top results to retrieve.
+
+    Returns:
+        List[Tuple]: A list of tuples, each containing the confidence and custom ID of the top results.
+    """
+    validate_feature_data(data, "FeatureHub face search top k")
+    feature = HFFaceFeature(size=HInt32(data.size), data=data.ctypes.data_as(HPFloat))
+    results = HFSearchTopKResults()
+    ret = HFFeatureHubFaceSearchTopK(feature, top_k, PHFSearchTopKResults(results))
+    outputs = []
+    if ret == errcode.HSUCCEED:
+        for idx in range(results.size):
+            confidence = results.confidence[idx]
+            id_ = results.ids[idx]
+            outputs.append((confidence, id_))
+    return outputs
+
+
+def feature_hub_face_update(face_identity: FaceIdentity) -> bool:
+    """
+    Updates an existing face identity in the feature hub.
+
+    Args:
+        face_identity (FaceIdentity): The face identity to update.
+
+    Returns:
+        bool: True if the update was successful, False otherwise.
+
+    Notes:
+        Logs an error if the update operation fails.
+    """
+    ret = HFFeatureHubFaceUpdate(face_identity._c_struct())
+    if ret != 0:
+        logger.error(f"Failed to update face feature data in FeatureHub: {ret}")
+        return False
+    return True
+
+
+def feature_hub_face_remove(custom_id: int) -> bool:
+    """
+    Removes a face identity from the feature hub using its custom ID.
+
+    Args:
+        custom_id (int): The custom ID of the face identity to remove.
+
+    Returns:
+        bool: True if the face was successfully removed, False otherwise.
+
+    Notes:
+        Logs an error if the removal operation fails.
+    """
+    ret = HFFeatureHubFaceRemove(HFaceId(custom_id))
+    if ret != 0:
+        logger.error(f"Failed to remove face feature data from FeatureHub: {ret}")
+        return False
+    return True
+
+
+def feature_hub_get_face_identity(custom_id: int):
+    """
+    Retrieves a face identity from the feature hub using its custom ID.
+
+    Args:
+        custom_id (int): The custom ID of the face identity to retrieve.
+
+    Returns:
+        FaceIdentity: The face identity retrieved, or None if the operation fails.
+
+    Notes:
+        Logs an error if retrieving the face identity fails.
+    """
+    identify = HFFaceFeatureIdentity()
+    ret = HFFeatureHubGetFaceIdentity(HFaceId(custom_id), PHFFaceFeatureIdentity(identify))
+    check_error(ret, "Get face identity from FeatureHub", custom_id=custom_id)
+
+    return FaceIdentity.from_ctypes(identify)
+
+
+def feature_hub_get_face_count() -> int:
+    """
+    Retrieves the total count of face identities stored in the feature hub.
+
+    Returns:
+        int: The count of face identities.
+
+    Notes:
+        Logs an error if the operation to retrieve the count fails.
+    """
+    count = HInt32()
+    ret = HFFeatureHubGetFaceCount(HPInt32(count))
+    check_error(ret, "Get face count")
+
+    return int(count.value)
+
+
+def feature_hub_get_face_id_list() -> List[int]:
+    """
+    Retrieves a list of face IDs from the feature hub.
+
+    Returns:
+        List[int]: A list of face IDs.
+    """
+    ids = HFFeatureHubExistingIds()
+    ptr = PHFFeatureHubExistingIds(ids)
+    ret = HFFeatureHubGetExistingIds(ptr)
+    if ret != 0:
+        logger.error(f"Failed to get face id list: {ret}")
+    return [int(ids.ids[i]) for i in range(ids.size)]
+
+def view_table_in_terminal():
+    """
+    Displays the database table of face identities in the terminal.
+
+    Notes:
+        Logs an error if the operation to view the table fails.
+    """
+    ret = HFFeatureHubViewDBTable()
+    check_error(ret, "View DB table")
+
+def get_recommended_cosine_threshold() -> float:
+    """
+    Retrieves the recommended cosine threshold.
+    """
+    threshold = HFloat()
+    HFGetRecommendedCosineThreshold(threshold)
+    return float(threshold.value)
+
+def get_similarity_converter_config() -> dict:
+    """
+    Retrieves the similarity converter configuration.
+    """
+    config = HFSimilarityConverterConfig()
+    ret = HFGetCosineSimilarityConverter(PHFSimilarityConverterConfig(config))
+    if ret != 0:
+        logger.error(f"Failed to get cosine similarity converter config: {ret}")
+    cfg = { 
+        "threshold": config.threshold,
+        "middleScore": config.middleScore,
+        "steepness": config.steepness,
+        "outputMin": config.outputMin,
+        "outputMax": config.outputMax
+    }
+    return cfg
+
+def set_similarity_converter_config(cfg: dict):
+    """
+    Sets the similarity converter configuration.
+    """
+    config = HFSimilarityConverterConfig()
+    config.threshold = cfg["threshold"]
+    config.middleScore = cfg["middleScore"]
+    config.steepness = cfg["steepness"]
+    config.outputMin = cfg["outputMin"]
+    config.outputMax = cfg["outputMax"]
+    HFUpdateCosineSimilarityConverter(config)
+
+def cosine_similarity_convert_to_percentage(similarity: float) -> float:
+    """
+    Converts a cosine similarity score to a percentage similarity score.
+    """
+    result = HFloat()
+    ret = HFCosineSimilarityConvertToPercentage(HFloat(similarity), HPFloat(result))
+    if ret != 0:
+        logger.error(f"Failed to convert cosine similarity to percentage: {ret}")
+    return float(result.value)
+
+def version() -> str:
+    """
+    Retrieves the version of the InspireFace library.
+
+    Returns:
+        str: The version string of the library.
+    """
+    ver = HFInspireFaceVersion()
+    HFQueryInspireFaceVersion(PHFInspireFaceVersion(ver))
+    return f"{ver.major}.{ver.minor}.{ver.patch}"
+
+
+def set_logging_level(level: int) -> None:
+    """
+    Sets the logging level of the InspireFace library.
+
+    Args:
+        level (int): The level to set the logging to.
+    """
+    HFSetLogLevel(level)
+
+def disable_logging() -> None:
+    """
+    Disables all logging from the InspireFace library.
+    """
+    HFLogDisable()
+
+def show_system_resource_statistics():
+    """
+    Displays the system resource information.
+    """
+    HFDeBugShowResourceStatistics()
+
+def switch_apple_coreml_inference_mode(mode: int):
+    """
+    Switches the Apple CoreML inference mode.
+    """
+    ret = HFSetAppleCoreMLInferenceMode(mode)
+    if ret != 0:
+        logger.error(f"Failed to set Apple CoreML inference mode: {ret}")
+        return False
+    return True
+
+def set_expansive_hardware_rockchip_dma_heap_path(path: str):
+    """
+    Sets the path to the expansive hardware Rockchip DMA heap.
+    """
+    ret = HFSetExpansiveHardwareRockchipDmaHeapPath(path)
+    check_error(ret, "Set expansive hardware Rockchip DMA heap path", path=path)
+
+def query_expansive_hardware_rockchip_dma_heap_path() -> str:
+    """
+    Queries the path to the expansive hardware Rockchip DMA heap.
+    """
+    path = HString()
+    ret = HFQueryExpansiveHardwareRockchipDmaHeapPath(path)
+    check_error(ret, "Query expansive hardware Rockchip DMA heap path")
+    return str(path.value)
+
+
+def set_cuda_device_id(device_id: int):
+    """
+    Sets the CUDA device ID.
+    """
+    ret = HFSetCudaDeviceId(device_id)
+    check_error(ret, "Set CUDA device ID", device_id=device_id)
+
+def get_cuda_device_id() -> int:
+    """
+    Gets the CUDA device ID.
+    """
+    id = HInt32()
+    ret = HFGetCudaDeviceId(id)
+    check_error(ret, "Get CUDA device ID")
+    return int(id.value)
+
+def print_cuda_device_info():
+    """
+    Prints the CUDA device information.
+    """
+    HFPrintCudaDeviceInfo()
+    
+def get_num_cuda_devices() -> int:
+    """
+    Gets the number of CUDA devices.
+    """
+    num = HInt32()
+    ret = HFGetNumCudaDevices(num)
+    check_error(ret, "Get number of CUDA devices")
+    return int(num.value)
+
+def check_cuda_device_support() -> bool:
+    """
+    Checks if the CUDA device is supported.
+    """
+    is_support = HInt32()
+    ret = HFCheckCudaDeviceSupport(is_support)
+    check_error(ret, "Check CUDA device support")
+    return bool(is_support.value)
diff --git a/cpp-package/inspireface/python/inspireface/modules/utils/__init__.py b/cpp-package/inspireface/python/inspireface/modules/utils/__init__.py
new file mode 100644
index 000000000..7e30d71d3
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/utils/__init__.py
@@ -0,0 +1 @@
+from .resource import *
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/inspireface/modules/utils/resource.py b/cpp-package/inspireface/python/inspireface/modules/utils/resource.py
new file mode 100644
index 000000000..7cff1df61
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/modules/utils/resource.py
@@ -0,0 +1,269 @@
+"""
+InspireFace Resource Manager
+
+This module provides model downloading functionality with two modes:
+1. Original mode: Download models from COS (Tencent Cloud Object Storage)
+2. ModelScope mode: Download models from ModelScope platform
+
+ModelScope mode usage:
+    rm = ResourceManager(use_modelscope=True, modelscope_model_id="tunmxy/InspireFace")
+    model_path = rm.get_model("Gundam_RV1106")
+
+Requirements for ModelScope mode:
+    pip install modelscope
+"""
+
+import os
+import sys
+from pathlib import Path
+import urllib.request
+import ssl
+import hashlib
+
+try:
+    from modelscope.hub.snapshot_download import snapshot_download
+    MODELSCOPE_AVAILABLE = True
+except ImportError:
+    MODELSCOPE_AVAILABLE = False
+
+# Global configuration for resource downloading
+USE_OSS_DOWNLOAD = False  # If True, force use OSS download instead of ModelScope
+
+def set_use_oss_download(use_oss: bool):
+    """Set whether to use OSS download instead of ModelScope
+    
+    Args:
+        use_oss (bool): If True, use OSS download; if False, use ModelScope (default)
+    """
+    global USE_OSS_DOWNLOAD
+    USE_OSS_DOWNLOAD = use_oss
+
+def get_file_hash_sha256(file_path):
+    sha256 = hashlib.sha256()
+    with open(file_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096), b''):
+            sha256.update(chunk)
+    return sha256.hexdigest()
+
+class ResourceManager:
+    def __init__(self, use_modelscope: bool = True, modelscope_model_id: str = "tunmxy/InspireFace"):
+        """Initialize resource manager and create necessary directories
+        
+        Args:
+            use_modelscope: Whether to download models from ModelScope platform
+            modelscope_model_id: ModelScope model ID (default: tunmxy/InspireFace)
+        """
+        self.user_home = Path.home()
+        self.base_dir = self.user_home / '.inspireface'
+        self.models_dir = self.base_dir / 'models'
+        
+        # ModelScope configuration
+        self.use_modelscope = use_modelscope
+        self.modelscope_model_id = modelscope_model_id
+        self.modelscope_cache_dir = self.base_dir / 'ms'
+        
+        # Create directories
+        self.base_dir.mkdir(exist_ok=True)
+        self.models_dir.mkdir(exist_ok=True)
+        if self.use_modelscope:
+            self.modelscope_cache_dir.mkdir(exist_ok=True)
+            
+        # Check ModelScope availability
+        if self.use_modelscope and not MODELSCOPE_AVAILABLE:
+            raise ImportError(
+                "ModelScope is not available. You have two options:\n"
+                "1. Install ModelScope: pip install modelscope\n" 
+                "2. Switch to OSS download mode by calling: inspireface.use_oss_download(True) before using InspireFace"
+            )
+        
+        # Model URLs
+        self._MODEL_LIST = {
+            "Pikachu": {
+                "url": "https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/Pikachu",
+                "filename": "Pikachu",
+                "md5": "5037ba1f49905b783a1c973d5d58b834a645922cc2814c8e3ca630a38dc24431"
+            },
+            "Megatron": {
+                "url": "https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/Megatron",
+                "filename": "Megatron",
+                "md5": "709fddf024d9f34ec034d8ef79a4779e1543b867b05e428c1d4b766f69287050"
+            },
+            "Megatron_TRT": {
+                "url": "https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/Megatron_TRT",
+                "filename": "Megatron_TRT",
+                "md5": "bc9123bdc510954b28d703b8ffe6023f469fb81123fd0b0b27fd452dfa369bab"
+            },
+            "Gundam_RK356X": {
+                "url": "https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/Gundam_RK356X",
+                "filename": "Gundam_RK356X",
+                "md5": "0fa12a425337ed98bd82610768a50de71cf93ef42a0929ba06cc94c86f4bd415"
+            },
+            "Gundam_RK3588": {
+                "url": "https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/Gundam_RK3588",
+                "filename": "Gundam_RK3588",
+                "md5": "66070e8d654408b666a2210bd498a976bbad8b33aef138c623e652f8d956641e"
+            }
+        }
+
+    def _download_from_modelscope(self, model_name: str) -> str:
+        """Download model from ModelScope platform
+        
+        Args:
+            model_name: Name of the model to download
+            
+        Returns:
+            str: Path to the downloaded model file
+        """
+        if not MODELSCOPE_AVAILABLE:
+            raise ImportError("ModelScope is not available. Please install it with: pip install modelscope")
+            
+        print(f"Downloading model '{model_name}' from ModelScope...")
+        
+        try:
+            # Download specific model file from ModelScope
+            cache_dir = snapshot_download(
+                model_id=self.modelscope_model_id,
+                cache_dir=str(self.modelscope_cache_dir),
+                allow_file_pattern=[model_name]  # Only download the specific model file
+            )
+            
+            model_file_path = Path(cache_dir) / model_name
+            
+            if not model_file_path.exists():
+                raise FileNotFoundError(f"Model file '{model_name}' not found in downloaded repository")
+                
+            print(f"ModelScope download completed: {model_file_path}")
+            return str(model_file_path)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to download model from ModelScope: {e}")
+
+    def get_model(self, name: str, re_download: bool = False, ignore_verification: bool = False) -> str:
+        """
+        Get model path. Download if not exists or re_download is True.
+        
+        Args:
+            name: Model name
+            re_download: Force re-download if True
+            ignore_verification: Skip model hash verification if True
+            
+        Returns:
+            str: Full path to model file
+        """
+        # Check global OSS setting first, then instance setting
+        use_oss = USE_OSS_DOWNLOAD
+        use_modelscope_actual = self.use_modelscope and not use_oss
+        
+        # Use ModelScope download if enabled and OSS is not forced
+        if use_modelscope_actual:
+            return self._download_from_modelscope_with_cache(name, re_download)
+            
+        # Original download logic for backwards compatibility
+        if name not in self._MODEL_LIST:
+            raise ValueError(f"Model '{name}' not found. Available models: {list(self._MODEL_LIST.keys())}")
+
+        model_info = self._MODEL_LIST[name]
+        model_file = self.models_dir / model_info["filename"]
+        downloading_flag = model_file.with_suffix('.downloading')
+        
+        # Check if model exists and is complete
+        if model_file.exists() and not downloading_flag.exists() and not re_download:
+            if ignore_verification:
+                print(f"Warning: Model verification skipped for '{name}' as requested.")
+                return str(model_file)
+                
+            current_hash = get_file_hash_sha256(model_file)
+            if current_hash == model_info["md5"]:
+                return str(model_file)
+            else:
+                print(f"Model file hash mismatch for '{name}'. Re-downloading...")
+
+        # Start download
+        try:
+            print(f"Downloading model '{name}'...")
+            downloading_flag.touch()
+
+            # Create SSL context and headers
+            ssl_context = ssl.create_default_context()
+            ssl_context.check_hostname = False
+            ssl_context.verify_mode = ssl.CERT_NONE
+            headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'}
+            
+            req = urllib.request.Request(model_info["url"], headers=headers)
+            
+            with urllib.request.urlopen(req, context=ssl_context) as response:
+                total_size = int(response.headers.get('content-length', 0))
+                block_size = 8192
+                downloaded_size = 0
+                
+                with open(model_file, 'wb') as f:
+                    while True:
+                        buffer = response.read(block_size)
+                        if not buffer:
+                            break
+                        downloaded_size += len(buffer)
+                        f.write(buffer)
+                        
+                        if total_size > 0:
+                            percent = (downloaded_size / total_size) * 100
+                            sys.stdout.write(f"\rDownloading {name}: {percent:.1f}%")
+                            sys.stdout.flush()
+            
+            print("\nDownload completed")
+            downloading_flag.unlink()  # Remove the downloading flag
+            return str(model_file)
+
+        except Exception as e:
+            if model_file.exists():
+                model_file.unlink()
+            if downloading_flag.exists():
+                downloading_flag.unlink()
+            raise RuntimeError(f"Failed to download model: {e}")
+
+    def _download_from_modelscope_with_cache(self, name: str, re_download: bool = False) -> str:
+        """Download model from ModelScope with local caching logic
+        
+        Args:
+            name: Model name
+            re_download: Force re-download if True
+            
+        Returns:
+            str: Path to the model file
+        """
+        # Check if model exists in ModelScope cache
+        model_file_path = self.modelscope_cache_dir / name
+        
+        if model_file_path.exists() and not re_download:
+            print(f"Using cached model '{name}' from ModelScope")
+            return str(model_file_path)
+            
+        # Download from ModelScope
+        return self._download_from_modelscope(name)
+        
+# Usage examples
+if __name__ == "__main__":
+    try:
+        # Example 1: Default mode (ModelScope)
+        print("=== Default mode (ModelScope) ===")
+        rm = ResourceManager()
+        model_path = rm.get_model("Gundam_RV1106")
+        print(f"ModelScope model path: {model_path}")
+        
+        # Example 2: Force OSS mode using global setting
+        print("\n=== OSS mode (global setting) ===")
+        set_use_oss_download(True)
+        rm_oss = ResourceManager()
+        model_path_oss = rm_oss.get_model("Pikachu")
+        print(f"OSS model path: {model_path_oss}")
+        
+        # Reset to default
+        set_use_oss_download(False)
+        
+        # Example 3: Explicit ModelScope mode
+        print("\n=== Explicit ModelScope mode ===")
+        rm_ms = ResourceManager(use_modelscope=True, modelscope_model_id="tunmxy/InspireFace")
+        model_path_ms = rm_ms.get_model("Gundam_RV1106")
+        print(f"Explicit ModelScope model path: {model_path_ms}")
+        
+    except Exception as e:
+        print(f"Error: {e}")
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/inspireface/param.py b/cpp-package/inspireface/python/inspireface/param.py
new file mode 100644
index 000000000..7ed5fdf9d
--- /dev/null
+++ b/cpp-package/inspireface/python/inspireface/param.py
@@ -0,0 +1,21 @@
+
+
+# Session option
+from inspireface.modules.core.native import HF_ENABLE_NONE, HF_ENABLE_FACE_RECOGNITION, HF_ENABLE_LIVENESS,  HF_ENABLE_IR_LIVENESS, \
+    HF_ENABLE_MASK_DETECT, HF_ENABLE_FACE_ATTRIBUTE, HF_ENABLE_QUALITY, HF_ENABLE_INTERACTION, HF_ENABLE_FACE_POSE, HF_ENABLE_FACE_EMOTION, HF_PK_AUTO_INCREMENT, HF_PK_MANUAL_INPUT, \
+        HF_LANDMARK_HYPLMV2_0_25, HF_LANDMARK_HYPLMV2_0_50, HF_LANDMARK_INSIGHTFACE_2D106_TRACK
+
+# Face track mode
+from inspireface.modules.core.native import HF_DETECT_MODE_ALWAYS_DETECT, HF_DETECT_MODE_LIGHT_TRACK, HF_DETECT_MODE_TRACK_BY_DETECTION
+
+# Image format
+from inspireface.modules.core.native import HF_STREAM_RGB, HF_STREAM_BGR, HF_STREAM_RGBA, HF_STREAM_BGRA, HF_STREAM_YUV_NV12, HF_STREAM_YUV_NV21
+
+# Image rotation
+from inspireface.modules.core.native import HF_CAMERA_ROTATION_0, HF_CAMERA_ROTATION_90, HF_CAMERA_ROTATION_180, HF_CAMERA_ROTATION_270
+
+# Search mode
+from inspireface.modules.core.native import HF_SEARCH_MODE_EAGER, HF_SEARCH_MODE_EXHAUSTIVE
+
+# Logger level
+from inspireface.modules.core.native import HF_LOG_NONE, HF_LOG_DEBUG, HF_LOG_INFO, HF_LOG_WARN, HF_LOG_ERROR, HF_LOG_FATAL
diff --git a/RetinaFace/rcnn/core/__init__.py b/cpp-package/inspireface/python/post
similarity index 100%
rename from RetinaFace/rcnn/core/__init__.py
rename to cpp-package/inspireface/python/post
diff --git a/cpp-package/inspireface/python/pull_models.py b/cpp-package/inspireface/python/pull_models.py
new file mode 100644
index 000000000..5e9e4e688
--- /dev/null
+++ b/cpp-package/inspireface/python/pull_models.py
@@ -0,0 +1,4 @@
+import inspireface
+
+for model in ["Pikachu", "Megatron"]:
+    inspireface.pull_latest_model(model)
diff --git a/cpp-package/inspireface/python/read_nv21.py b/cpp-package/inspireface/python/read_nv21.py
new file mode 100644
index 000000000..cc3da64d1
--- /dev/null
+++ b/cpp-package/inspireface/python/read_nv21.py
@@ -0,0 +1,71 @@
+import cv2
+import numpy as np
+from inspireface import ImageStream
+import inspireface as isf
+
+
+def read_nv21(file_path, width, height, rotate=0):
+    with open(file_path, 'rb') as f:
+        nv21_data = f.read()
+    
+    yuv = np.frombuffer(nv21_data, dtype=np.uint8)
+    
+    expected_size = width * height * 3 // 2
+    if len(yuv) < expected_size:
+        raise ValueError(f"NV21 data size is not enough: expected {expected_size} bytes, actual {len(yuv)} bytes")
+    
+    yuv_mat = np.zeros((height * 3 // 2, width), dtype=np.uint8)
+    yuv_mat[:] = yuv[:height * width * 3 // 2].reshape(height * 3 // 2, width)
+    
+    bgr_mat = cv2.cvtColor(yuv_mat, cv2.COLOR_YUV2BGR_NV21)
+    
+    # add reverse rotate
+    if rotate != 0:
+        # calculate reverse rotate angle
+        reverse_angle = (360 - rotate) % 360
+        
+        # select rotate method by angle
+        if reverse_angle == 90:
+            bgr_mat = cv2.rotate(bgr_mat, cv2.ROTATE_90_CLOCKWISE)
+        elif reverse_angle == 180:
+            bgr_mat = cv2.rotate(bgr_mat, cv2.ROTATE_180)
+        elif reverse_angle == 270:
+            bgr_mat = cv2.rotate(bgr_mat, cv2.ROTATE_90_COUNTERCLOCKWISE)
+        else:
+            # for non 90 degree angle, use affine transform
+            center = (bgr_mat.shape[1] // 2, bgr_mat.shape[0] // 2)
+            rotation_matrix = cv2.getRotationMatrix2D(center, reverse_angle, 1.0)
+            bgr_mat = cv2.warpAffine(bgr_mat, rotation_matrix, (bgr_mat.shape[1], bgr_mat.shape[0]))
+    
+    return bgr_mat
+
+# example usage
+if __name__ == "__main__":
+    image_width = 640
+    image_height = 480
+    nv21_file = ""
+    pre_rotate = 90
+    try:
+        # read and reverse rotate image
+        bgr_image = read_nv21(nv21_file, image_width, image_height, pre_rotate)
+        
+        cv2.imshow("origin", bgr_image)
+        
+        # show original image for comparison
+        bgr_image_original = read_nv21(nv21_file, image_width, image_height)
+        cv2.imshow("rotate ", bgr_image_original)
+        
+        cv2.waitKey(0)
+        cv2.destroyAllWindows()
+        
+    except Exception as e:
+        print(f"error: {e}")
+
+    with open(nv21_file, 'rb') as f:
+        nv21_data = f.read()
+
+    stream = ImageStream.load_from_buffer(nv21_data, image_width, image_height, isf.HF_STREAM_YUV_NV21, isf.HF_CAMERA_ROTATION_90)
+
+    stream.debug_show()
+
+    
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/sample_face_comparison.py b/cpp-package/inspireface/python/sample_face_comparison.py
new file mode 100644
index 000000000..b7c1f9dde
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_face_comparison.py
@@ -0,0 +1,52 @@
+import cv2
+import inspireface as isf
+import click
+
+@click.command()
+@click.argument('image_path1')
+@click.argument('image_path2') 
+def case_face_comparison(image_path1, image_path2):
+    """
+    This is a sample application for comparing two face images.
+    Args:
+        image_path1 (str): Path to the first face image
+        image_path2 (str): Path to the second face image
+    """
+    # Enable face recognition features
+    opt = isf.HF_ENABLE_FACE_RECOGNITION
+    session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+
+    # Load and check the first image
+    image1 = cv2.imread(image_path1)
+    assert image1 is not None, "Failed to load first image"
+    
+    # Load and check the second image  
+    image2 = cv2.imread(image_path2)
+    assert image2 is not None, "Failed to load second image"
+
+    # Detect faces in first image
+    faces1 = session.face_detection(image1)
+    assert faces1, "No face detected in first image"
+    face1 = faces1[0]  # Use the first detected face
+
+    # Detect faces in second image
+    faces2 = session.face_detection(image2)
+    assert faces2, "No face detected in second image"
+    face2 = faces2[0]  # Use the first detected face
+
+    # Extract features
+    feature1 = session.face_feature_extract(image1, face1)
+    feature2 = session.face_feature_extract(image2, face2)
+
+    # Calculate similarity score between the two faces
+    similarity = isf.feature_comparison(feature1, feature2)
+    
+    print(f"The cosine similarity score: {similarity:.4f}")
+    print(f"{'Same person' if similarity > isf.get_recommended_cosine_threshold() else 'Different person'}")
+
+    percentage = isf.cosine_similarity_convert_to_percentage(similarity)
+    print(f"The percentage similarity: {percentage:.4f}")
+
+
+if __name__ == '__main__':
+    case_face_comparison()
diff --git a/cpp-package/inspireface/python/sample_face_detection.py b/cpp-package/inspireface/python/sample_face_detection.py
new file mode 100644
index 000000000..9f1ae1094
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_face_detection.py
@@ -0,0 +1,97 @@
+import os
+import cv2
+import inspireface as isf
+import click
+
+race_tags = ["Black", "Asian", "Latino/Hispanic", "Middle Eastern", "White"]
+gender_tags = ["Female", "Male"]
+age_bracket_tags = [
+    "0-2 years old", "3-9 years old", "10-19 years old", "20-29 years old", "30-39 years old",
+    "40-49 years old", "50-59 years old", "60-69 years old", "more than 70 years old"
+]
+emotion_tags = ["Neutral", "Happy", "Sad", "Surprise", "Fear", "Disgust", "Anger"]
+
+@click.command()
+@click.argument('image_path')
+@click.option('--show', is_flag=True, help='Display the image with detected faces.')
+def case_face_detection_image(image_path, show):
+    """
+    This is a sample application for face detection and tracking using an image.
+    It also includes pipeline extensions such as RGB liveness, mask detection, and face quality evaluation.
+    """
+    isf.switch_image_processing_backend(isf.HF_IMAGE_PROCESSING_CPU)
+    opt = isf.HF_ENABLE_FACE_RECOGNITION | isf.HF_ENABLE_QUALITY | isf.HF_ENABLE_MASK_DETECT | \
+          isf.HF_ENABLE_LIVENESS | isf.HF_ENABLE_INTERACTION | isf.HF_ENABLE_FACE_ATTRIBUTE | isf.HF_ENABLE_FACE_EMOTION
+    session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+    session.set_detection_confidence_threshold(0.5)
+    # Load image
+    image = cv2.imread(image_path)
+    assert image is not None, "Please check that the image path is correct."
+
+    # Dynamic drawing parameters (adjusted to image size)
+    h, w = image.shape[:2]
+    scale = max(w, h) / 1000.0
+    line_thickness = max(1, int(2 * scale))
+    circle_radius = max(1, int(1.5 * scale))
+    font_scale = 0.5 * scale
+
+    # Detect faces
+    faces = session.face_detection(image)
+    print(faces)
+    print(f"face detection: {len(faces)} found")
+
+    draw = image.copy()
+    for idx, face in enumerate(faces):
+        print(f"{'==' * 20}")
+        print(f"idx: {idx}")
+        print(f"detection confidence: {face.detection_confidence}")
+        print(f"roll: {face.roll}, yaw: {face.yaw}, pitch: {face.pitch}")
+
+        x1, y1, x2, y2 = face.location
+        center = ((x1 + x2) / 2, (y1 + y2) / 2)
+        size = (x2 - x1, y2 - y1)
+        angle = face.roll
+
+        rect = ((center[0], center[1]), (size[0], size[1]), angle)
+        box = cv2.boxPoints(rect).astype(int)
+        cv2.drawContours(draw, [box], 0, (100, 180, 29), line_thickness)
+
+        # Draw landmark
+        lmk = session.get_face_dense_landmark(face)
+        for x, y in lmk.astype(int):
+            cv2.circle(draw, (x, y), circle_radius, (220, 100, 0), -1)
+
+        # Optional: Add detection confidence (text) on the face box
+        # label = f"{face.detection_confidence:.2f}"
+        # cv2.putText(draw, label, (x1, max(y1 - 10, 0)), cv2.FONT_HERSHEY_SIMPLEX,
+        #             font_scale, (255, 255, 255), line_thickness)
+
+    # Execute extended functions (optional modules)
+    select_exec_func = isf.HF_ENABLE_QUALITY | isf.HF_ENABLE_MASK_DETECT | \
+                       isf.HF_ENABLE_LIVENESS | isf.HF_ENABLE_INTERACTION | isf.HF_ENABLE_FACE_ATTRIBUTE | isf.HF_ENABLE_FACE_EMOTION
+    extends = session.face_pipeline(image, faces, select_exec_func)
+    print(extends)
+    for idx, ext in enumerate(extends):
+        print(f"{'==' * 20}")
+        print(f"idx: {idx}")
+        print(f"quality: {ext.quality_confidence}")
+        print(f"rgb liveness: {ext.rgb_liveness_confidence}")
+        print(f"face mask: {ext.mask_confidence}")
+        print(f"face eyes status: left eye: {ext.left_eye_status_confidence} right eye: {ext.right_eye_status_confidence}")
+        print(f"gender: {gender_tags[ext.gender]}")
+        print(f"race: {race_tags[ext.race]}")
+        print(f"age: {age_bracket_tags[ext.age_bracket]}")
+        print(f"emotion: {emotion_tags[ext.emotion]}")
+
+    # Save the annotated image
+    save_path = os.path.join("tmp", "det.jpg")
+    os.makedirs("tmp", exist_ok=True)
+    cv2.imwrite(save_path, draw)
+    if show:
+        cv2.imshow("Face Detection", draw)
+        cv2.waitKey(0)
+        cv2.destroyAllWindows()
+    print(f"\nSave annotated image to {save_path}")
+
+if __name__ == '__main__':
+    case_face_detection_image()
diff --git a/cpp-package/inspireface/python/sample_face_recognition.py b/cpp-package/inspireface/python/sample_face_recognition.py
new file mode 100644
index 000000000..991b787f2
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_face_recognition.py
@@ -0,0 +1,97 @@
+import os
+import cv2
+import inspireface as isf
+import click
+
+@click.command()
+@click.argument('test_data_folder')
+def case_face_recognition(test_data_folder):
+    """
+    Launches the face recognition system, inserts face features into a database, and performs searches.
+    Args:
+        resource_path (str): Path to the resource directory for face recognition algorithms.
+        test_data_folder (str): Path to the test data containing images for insertion and recognition tests.
+    """
+    # If you need to switch from the default Pikachu model to another model like Megatron, you can use reload
+    ret = isf.reload("Megatron")
+    assert ret, "Launch failure. Please ensure the resource path is correct."
+
+    # Enable face recognition features.
+    opt = isf.HF_ENABLE_FACE_RECOGNITION
+    session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+
+    # Configure the feature management system.
+    feature_hub_config = isf.FeatureHubConfiguration(
+        primary_key_mode=isf.HF_PK_AUTO_INCREMENT,
+        enable_persistence=False,
+        persistence_db_path="",
+        search_threshold=0.48,
+        search_mode=isf.HF_SEARCH_MODE_EAGER,
+    )
+    ret = isf.feature_hub_enable(feature_hub_config)
+    assert ret, "Failed to enable FeatureHub."
+
+    # Insert face features from 'bulk' directory.
+    bulk_path = os.path.join(test_data_folder, "bulk")
+    assert os.path.exists(bulk_path), "Bulk directory does not exist."
+
+    insert_images = [os.path.join(bulk_path, path) for path in os.listdir(bulk_path) if path.endswith(".jpg")]
+    for idx, image_path in enumerate(insert_images):
+        name = os.path.basename(image_path).replace(".jpg", "")
+        image = cv2.imread(image_path)
+        assert image is not None, f"Failed to load image {image_path}"
+        faces = session.face_detection(image)
+        if faces:
+            face = faces[0]  # Assume the most prominent face is what we want.
+            feature = session.face_feature_extract(image, face)
+            identity = isf.FaceIdentity(feature, id=idx)
+            ret, alloc_id = isf.feature_hub_face_insert(identity)
+            assert ret, "Failed to insert face."
+
+    count = isf.feature_hub_get_face_count()
+    print(f"Number of faces inserted: {count}")
+
+    # Process faces from 'RD' directory and insert them.
+    RD = os.path.join(test_data_folder, "RD")
+    assert os.path.exists(RD), "RD directory does not exist."
+    RD_images = [os.path.join(RD, path) for path in os.listdir(RD) if path.endswith(".jpeg")]
+
+    for idx, image_path in enumerate(RD_images[:-1]):
+        name = os.path.basename(image_path).replace(".jpeg", "")
+        image = cv2.imread(image_path)
+        assert image is not None, f"Failed to load image {image_path}"
+        faces = session.face_detection(image)
+        if faces:
+            face = faces[0]
+            feature = session.face_feature_extract(image, face)
+            identity = isf.FaceIdentity(feature, id=idx+count+1)
+            ret, alloc_id = isf.feature_hub_face_insert(identity)
+            assert ret, "Failed to insert face."
+
+    count = isf.feature_hub_get_face_count()
+    print(f"Total number of faces after insertion: {count}")
+
+    # Search for a similar face using the last image in RD directory.
+    remain = cv2.imread(RD_images[-1])
+    assert remain is not None, f"Failed to load image {RD_images[-1]}"
+    faces = session.face_detection(remain)
+    assert faces, "No faces detected."
+    face = faces[0]
+    feature = session.face_feature_extract(remain, face)
+
+    search = isf.feature_hub_face_search(feature)
+    if search.similar_identity.id != -1:
+        print(f"Found similar identity with ID: {search.similar_identity.id}, Confidence: {search.confidence:.2f}")
+    else:
+        print("No similar identity found.")
+
+    # Display top-k similar face identities.
+    print("Top-k similar identities:")
+    search_top_k = isf.feature_hub_face_search_top_k(feature, 10)
+    for idx, (conf, _id) in enumerate(search_top_k):
+        identity = isf.feature_hub_get_face_identity(_id)
+        print(f"Top-{idx + 1}: ID: {_id}, Confidence: {conf:.2f}")
+
+
+if __name__ == '__main__':
+    case_face_recognition()
diff --git a/cpp-package/inspireface/python/sample_face_track_from_video.py b/cpp-package/inspireface/python/sample_face_track_from_video.py
new file mode 100644
index 000000000..93b716e96
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_face_track_from_video.py
@@ -0,0 +1,164 @@
+import time
+
+import click
+import cv2
+import inspireface as isf
+import numpy as np
+import time
+
+def generate_color(id):
+    """
+    Generate a bright color based on the given integer ID. Ensures 50 unique colors.
+
+    Args:
+        id (int): The ID for which to generate a color.
+
+    Returns:
+        tuple: A tuple representing the color in BGR format.
+    """
+    # Handle invalid ID (-1)
+    if id < 0:
+        return (128, 128, 128)  # Return gray color for invalid ID
+        
+    max_id = 50  # Number of unique colors
+    id = id % max_id
+
+    # Generate HSV color
+    hue = int((id * 360 / max_id) % 360)  # Distribute hue values equally
+    saturation = 200 + (55 * id) % 55  # High saturation for bright colors
+    value = 200 + (55 * id) % 55  # High value for bright colors
+
+    hsv_color = np.uint8([[[hue, saturation, value]]])
+    rgb_color = cv2.cvtColor(hsv_color, cv2.COLOR_HSV2BGR)[0][0]
+
+    return (int(rgb_color[0]), int(rgb_color[1]), int(rgb_color[2]))
+
+@click.command()
+@click.argument('source')
+@click.option('--show', is_flag=True, help='Display the video stream or video file in a window.')
+@click.option('--out', type=str, default=None, help='Path to save the processed video.')
+def case_face_tracker_from_video(source, show, out):
+    """
+    Launch a face tracking process from a video source. The 'source' can either be a webcam index (0, 1, ...)
+    or a path to a video file. Use the --show option to display the video.
+
+    Args:
+        resource_path (str): Path to the resource directory for face tracking algorithms.
+        source (str): Webcam index or path to the video file.
+        show (bool): If set, the video will be displayed in a window.
+        out (str): Path to save the processed video.
+    """
+    # Optional features, loaded during session creation based on the modules specified.
+    opt = isf.HF_ENABLE_NONE | isf.HF_ENABLE_INTERACTION
+    session = isf.InspireFaceSession(opt, isf.HF_DETECT_MODE_LIGHT_TRACK, max_detect_num=25, detect_pixel_level=320)    # Use video mode
+    session.set_track_mode_smooth_ratio(0.06)
+    session.set_track_mode_num_smooth_cache_frame(15)
+    session.set_filter_minimum_face_pixel_size(0)
+    session.set_track_model_detect_interval(0)
+    session.set_track_lost_recovery_mode(True)
+    session.set_enable_track_cost_spend(True)
+    # Determine if the source is a digital webcam index or a video file path.
+    try:
+        source_index = int(source)  # Try to convert source to an integer.
+        print(f"Using webcam at index {source_index}.")
+        cap = cv2.VideoCapture(source_index)
+    except ValueError:
+        # If conversion fails, treat source as a file path.
+        print(f"Opening video file at {source}.")
+        cap = cv2.VideoCapture(source)
+
+    if not cap.isOpened():
+        print("Error: Could not open video source.")
+        return
+
+    # VideoWriter to save the processed video if out is provided.
+    if out:
+        fourcc = cv2.VideoWriter_fourcc(*'XVID')
+        fps = cap.get(cv2.CAP_PROP_FPS) if cap.get(cv2.CAP_PROP_FPS) > 0 else 30
+        frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        out_video = cv2.VideoWriter(out, fourcc, fps, (frame_width, frame_height))
+        print(f"Saving video to: {out}")
+
+    # Main loop to process video frames.
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break  # Exit loop if no more frames or error occurs.
+
+        # Process frame here (e.g., face detection/tracking).
+        t1 = time.time()
+        faces = session.face_detection(frame)
+        t2 = time.time()
+        # print(f"Face detection time: {t2 - t1} seconds")
+        session.print_track_cost_spend()
+        exts = session.face_pipeline(frame, faces, isf.HF_ENABLE_INTERACTION)
+
+        for idx, face in enumerate(faces):
+            # Get face bounding box
+            x1, y1, x2, y2 = face.location
+
+            # Calculate center, size, and angle
+            center = ((x1 + x2) / 2, (y1 + y2) / 2)
+            size = (x2 - x1, y2 - y1)
+            angle = face.roll
+
+            # Apply rotation to the bounding box corners
+            rect = ((center[0], center[1]), (size[0], size[1]), angle)
+            box = cv2.boxPoints(rect)
+            box = box.astype(int)
+
+            actions = []
+            if exts[idx].action_normal:
+                actions.append("Normal")
+            if exts[idx].action_jaw_open:
+                actions.append("Jaw Open") 
+            if exts[idx].action_shake:
+                actions.append("Shake")
+            if exts[idx].action_blink:
+                actions.append("Blink")
+            if exts[idx].action_head_raise:
+                actions.append("Head Raise")
+            print("Actions:", actions)
+            
+            color = generate_color(face.track_id)
+
+            # Draw the rotated bounding box
+            cv2.drawContours(frame, [box], 0, color, 4)
+
+            # Draw landmarks
+            lmk = session.get_face_dense_landmark(face)
+            for x, y in lmk.astype(int):
+                cv2.circle(frame, (x, y), 0, color, 4)
+
+            five_key_points = session.get_face_five_key_points(face)
+            for x, y in five_key_points.astype(int):
+                cv2.circle(frame, (x, y), 0, (255-color[0], 255-color[1], 255-color[2]), 6)
+
+            # Draw track ID at the top of the bounding box
+            text = f"ID: {face.track_id}, Count: {face.track_count}"
+            text_size, _ = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
+            text_x = min(box[:, 0])
+            text_y = min(box[:, 1]) - 10
+            if text_y < 0:
+                text_y = min(box[:, 1]) + text_size[1] + 10
+            cv2.putText(frame, text, (text_x, text_y), cv2.FONT_HERSHEY_SIMPLEX, 0.6, color, 2)
+
+        if show:
+            cv2.imshow("Face Tracker", frame)
+            if cv2.waitKey(25) & 0xFF == ord('q'):
+                break  # Exit loop if 'q' is pressed.
+
+        if out:
+            out_video.write(frame)
+
+    # Cleanup: release video capture and close any open windows.
+    cap.release()
+    if out:
+        out_video.release()
+    cv2.destroyAllWindows()
+    print("Released all resources and closed windows.")
+
+
+if __name__ == '__main__':
+    case_face_tracker_from_video()
diff --git a/cpp-package/inspireface/python/sample_feature_hub.py b/cpp-package/inspireface/python/sample_feature_hub.py
new file mode 100644
index 000000000..3c43b3d74
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_feature_hub.py
@@ -0,0 +1,38 @@
+import os
+import inspireface as isf
+import numpy as np
+import os
+
+def case_feature_hub():
+    # Gen a random feature
+    gen = np.random.rand(512).astype(np.float32)
+    # Set db path
+    db_path = "test.db"
+    # Configure the feature management system.
+    feature_hub_config = isf.FeatureHubConfiguration(
+        primary_key_mode=isf.HF_PK_AUTO_INCREMENT,
+        enable_persistence=True,
+        persistence_db_path=db_path,
+        search_threshold=0.48,
+        search_mode=isf.HF_SEARCH_MODE_EAGER,
+    )
+    ret = isf.feature_hub_enable(feature_hub_config)
+    assert ret, "Failed to enable FeatureHub."
+    print('T1, face count:', isf.feature_hub_get_face_count())
+    for i in range(10):
+        v = np.random.rand(512).astype(np.float32)
+        feature = isf.FaceIdentity(v, i)
+        ret, _ = isf.feature_hub_face_insert(feature)
+        assert ret, "Failed to insert face feature data into FeatureHub."
+    feature = isf.FaceIdentity(gen, -1)
+    isf.feature_hub_face_insert(feature)
+    result = isf.feature_hub_face_search(gen)
+    print(f"result: {result}")
+    assert os.path.exists(db_path), "FeatureHub database file not found."
+    ids = isf.feature_hub_get_face_id_list()
+    print(f"ids: {ids}")
+
+
+
+if __name__ == "__main__":
+    case_feature_hub()
diff --git a/cpp-package/inspireface/python/sample_feature_hub_crud.py b/cpp-package/inspireface/python/sample_feature_hub_crud.py
new file mode 100644
index 000000000..9beb294cf
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_feature_hub_crud.py
@@ -0,0 +1,123 @@
+import os
+import inspireface as isf
+import numpy as np
+import os
+
+import random
+
+random.seed(43)
+
+def gen_feature():
+    # Generate a random vector of length 512 and normalize it
+    vector = np.random.uniform(-1, 1, 512).astype(np.float32)
+    normalized_vector = vector / np.linalg.norm(vector)
+    return normalized_vector
+
+def gen_similar_feature(feature):
+    noise_strength = 0.3
+    
+    noise = np.random.uniform(-1, 1, len(feature)).astype(np.float32)
+    noise = noise / np.linalg.norm(noise)  
+    
+    similar_vector = feature + noise_strength * noise
+    
+    similar_vector = similar_vector / np.linalg.norm(similar_vector)
+    
+    return similar_vector
+
+def case_feature_hub():
+    # Set db path
+    db_path = "test.db"
+    if os.path.exists(db_path):
+        os.remove(db_path)
+    # Configure the feature management system.
+    feature_hub_config = isf.FeatureHubConfiguration(
+        primary_key_mode=isf.HF_PK_AUTO_INCREMENT,
+        enable_persistence=True,
+        persistence_db_path=db_path,
+        search_threshold=0.48,
+        search_mode=isf.HF_SEARCH_MODE_EAGER,
+    )
+    ret = isf.feature_hub_enable(feature_hub_config)
+    assert ret, "Failed to enable FeatureHub."
+    
+    embedding_list = []
+    for i in range(10):
+        # Generate a random embedding
+        embedding = gen_feature()
+        face_identity = isf.FaceIdentity(embedding, -1)
+        # Insert the face identity into the feature hub
+        ret, id = isf.feature_hub_face_insert(face_identity)
+        assert ret, "Failed to insert face identity."
+        print(f"Inserted face id: {id}")
+        embedding_list.append(embedding)
+
+    print(f"face count: {isf.feature_hub_get_face_count()}")
+    assert isf.feature_hub_get_face_count() == 10, "Failed to insert face identity."
+
+    # Search for the face identity
+    query_embedding = embedding_list[3]
+    search_result = isf.feature_hub_face_search(query_embedding)
+    assert search_result, "Failed to search for face identity."
+    # The auto-increment id is calculated starting from 1
+    assert search_result.similar_identity.id == 3 + 1, "Failed to search for face identity."
+    print(f"search confidence: {search_result.confidence}")
+
+    # Update the face identity
+    update_embedding = gen_feature()
+    update_face_identity = isf.FaceIdentity(update_embedding, search_result.similar_identity.id)
+    ret = isf.feature_hub_face_update(update_face_identity)
+    assert ret, "Failed to update face identity."
+    print(f"Updated face id: {search_result.similar_identity.id}")
+    
+    # Search for the face identity again
+    search_result = isf.feature_hub_face_search(query_embedding)
+    assert search_result, "Failed to search for face identity."
+    assert search_result.similar_identity.id == -1, "Failed to update face identity."
+    print(f"search confidence: {search_result.confidence}, id: {search_result.similar_identity.id}")
+
+    # Delete the face identity
+    ret = isf.feature_hub_face_remove(4)
+    assert ret, "Failed to delete face identity."
+    print(f"Deleted face id: {4}")
+    
+    # Search for the face identity again
+    search_result = isf.feature_hub_face_search(embedding_list[3])
+    assert search_result.similar_identity.id == -1, "Failed to delete face identity."
+    print(f"search confidence: {search_result.confidence}, id: {search_result.similar_identity.id}")
+    print(f"face count: {isf.feature_hub_get_face_count()}")
+    assert isf.feature_hub_get_face_count() == 9, "Failed to delete face identity."
+
+
+    # Top-k sample
+    # Set target face id = 9
+    target_face_id = 9
+    # Set k = 3
+    k = 4
+    # Gen k similar features
+    similar_features = []
+    for i in range(k):
+        similar_features.append(gen_similar_feature(embedding_list[target_face_id - 1]))
+    # Insert the similar features into the feature hub
+    expect_ids = []
+    for similar_feature in similar_features:
+        similar_face_identity = isf.FaceIdentity(similar_feature, -1)
+        ret, id = isf.feature_hub_face_insert(similar_face_identity)
+        assert ret, "Failed to insert similar face identity."
+        print(f"Inserted similar face id: {id}")
+        expect_ids.append(id)
+    # Insert the target face id
+    expect_ids.append(target_face_id)
+    print(f"expect ids: {expect_ids}")
+    # Search the top-k similar features
+    top_k_search_result = isf.feature_hub_face_search_top_k(embedding_list[target_face_id - 1], 1000)
+    assert len(top_k_search_result) == k + 1, "Failed to search for top-k similar features."
+    for result in top_k_search_result:
+        confidence, _id = result
+        print(f"search confidence: {confidence}, id: {_id}")
+        assert _id in expect_ids, "Failed to search for top-k similar features."
+        print(f"search confidence: {confidence}, id: {_id}")
+
+
+if __name__ == "__main__":
+    case_feature_hub()
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/sample_system_resource_statistics.py b/cpp-package/inspireface/python/sample_system_resource_statistics.py
new file mode 100644
index 000000000..73e7221fb
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_system_resource_statistics.py
@@ -0,0 +1,36 @@
+
+import inspireface as isf
+import click
+    
+@click.command()
+@click.argument("resource_path")
+def case_show_system_resource_statistics(resource_path):
+    """
+    This case is used to test the system resource statistics.
+    """
+    ret = isf.launch(resource_path)
+    assert ret, "Launch failure. Please ensure the resource path is correct."
+    print("-" * 100)
+    print("Initialization state")
+    print("-" * 100)
+    isf.show_system_resource_statistics()
+    print("-" * 100)
+    print("Create 10 sessions")
+    print("-" * 100)
+    print("")
+    num_created_sessions = 10
+    sessions = []
+    for i in range(num_created_sessions):
+        session = isf.InspireFaceSession(isf.HF_ENABLE_FACE_RECOGNITION, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+        sessions.append(session)
+    isf.show_system_resource_statistics()
+    print("-" * 100)
+    print("Release 10 sessions")
+    print("-" * 100)
+    print()
+    for session in sessions:
+        session.release()
+    isf.show_system_resource_statistics()
+
+if __name__ == "__main__":
+    case_show_system_resource_statistics()
diff --git a/cpp-package/inspireface/python/sample_video.py b/cpp-package/inspireface/python/sample_video.py
new file mode 100644
index 000000000..fd34ca643
--- /dev/null
+++ b/cpp-package/inspireface/python/sample_video.py
@@ -0,0 +1,51 @@
+import os
+import cv2
+import inspireface as isf
+import numpy as np
+import os
+import cv2
+
+
+def get_quality(image, session: isf.InspireFaceSession) -> float:
+    select_exec_func = isf.HF_ENABLE_QUALITY | isf.HF_ENABLE_MASK_DETECT | isf.HF_ENABLE_LIVENESS | isf.HF_ENABLE_INTERACTION | isf.HF_ENABLE_FACE_ATTRIBUTE
+    faces = session.face_detection(image)
+    if len(faces) > 0:
+        extends = session.face_pipeline(image, faces, select_exec_func)
+        if len(faces) == 0:
+            return 0
+        for idx, ext in enumerate(extends):
+            print(f"{'==' * 20}")
+            print(f"idx: {idx}")
+            # For these pipeline results, you can set thresholds based on the specific scenario to make judgments.
+            print(f"quality: {ext.quality_confidence}")
+            print(f"rgb liveness: {ext.rgb_liveness_confidence}")
+            print(f"face mask: {ext.mask_confidence}")
+            print(
+                f"face eyes status: left eye: {ext.left_eye_status_confidence} right eye: {ext.right_eye_status_confidence}")
+            print(f"gender: {ext.gender}")
+            print(f"race: {ext.race}")
+            print(f"age: {ext.age_bracket}")
+
+
+if __name__ == "__main__":
+    register_exec_func = isf.HF_ENABLE_QUALITY | isf.HF_ENABLE_MASK_DETECT | isf.HF_ENABLE_LIVENESS | isf.HF_ENABLE_INTERACTION | isf.HF_ENABLE_FACE_ATTRIBUTE
+    session = isf.InspireFaceSession(register_exec_func, isf.HF_DETECT_MODE_ALWAYS_DETECT)
+    cam = cv2.VideoCapture(0)
+    while True:
+        ret, frame = cam.read()
+        if not ret:
+            break
+            
+        quality = get_quality(frame, session)
+        
+        faces = session.face_detection(frame)
+        for face in faces:
+            x1, y1, x2, y2 = face.location
+            cv2.rectangle(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2)
+            
+        cv2.imshow('Video', frame)
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break
+            
+    cam.release()
+    cv2.destroyAllWindows()
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/setup.py b/cpp-package/inspireface/python/setup.py
new file mode 100644
index 000000000..a3fd3542d
--- /dev/null
+++ b/cpp-package/inspireface/python/setup.py
@@ -0,0 +1,152 @@
+from setuptools import setup, find_packages
+from wheel.bdist_wheel import bdist_wheel
+import platform
+import subprocess
+import os
+
+def get_version():
+    """Get version number"""
+    version_path = os.path.join(os.path.dirname(__file__), 'version.txt')
+    try:
+        with open(version_path, 'r') as f:
+            return f.read().strip()
+    except FileNotFoundError:
+        return "0.0.0"
+
+def get_post_version():
+    """Get post version number"""
+    post_path = os.path.join(os.path.dirname(__file__), 'post')
+    try:
+        with open(post_path, 'r') as f:
+            return f.read().strip()
+    except FileNotFoundError:
+        return ""
+
+def get_wheel_platform_tag():
+    """Get wheel package platform tag"""
+    system = platform.system().lower()
+    machine = platform.machine().lower()
+
+    arch_mapping = {
+        'x86_64': {
+            'windows': 'win_amd64',
+            'linux': 'manylinux2014_x86_64',
+            'darwin': 'macosx_12_0_x86_64'
+        },
+        'amd64': {
+            'windows': 'win_amd64',
+            'linux': 'manylinux2014_x86_64',
+            'darwin': 'macosx_12_0_x86_64'
+        },
+        'arm64': {
+            'windows': 'win_arm64',
+            'linux': 'manylinux2014_aarch64',
+            'darwin': 'macosx_11_0_arm64'
+        },
+        'aarch64': {
+            'windows': 'win_arm64',
+            'linux': 'manylinux2014_aarch64',
+            'darwin': 'macosx_11_0_arm64'
+        }
+    }
+    if os.getenv('INSPIRE_FACE_TARGET_AARCH_MAPPING'):
+        platform_arch = os.getenv('INSPIRE_FACE_TARGET_AARCH_MAPPING')
+    else:
+        platform_arch = arch_mapping.get(machine, {}).get(system)
+    if not platform_arch:
+        print("Unsupported platform: {} {}".format(system, machine))
+        raise RuntimeError("Unsupported platform: {} {}".format(system, machine))
+    
+    return platform_arch
+
+def get_lib_path_info():
+    """Get library file path information"""
+    system = platform.system().lower()
+    machine = platform.machine().lower()
+    
+    if system == 'windows':
+        arch = 'x64' if machine in ['amd64', 'x86_64'] else 'arm64'
+    elif system == 'linux':
+        arch = 'x64' if machine == 'x86_64' else 'arm64'
+    elif system == 'darwin':
+        if machine == 'x86_64':
+            try:
+                is_rosetta = bool(int(subprocess.check_output(
+                    ['sysctl', '-n', 'sysctl.proc_translated']).decode().strip()))
+                arch = 'arm64' if is_rosetta else 'x64'
+            except:
+                arch = 'x64'
+        else:
+            arch = 'arm64'
+    else:
+        raise RuntimeError(f"Unsupported system: {system}")
+    
+    return system, arch
+
+class BinaryDistWheel(bdist_wheel):
+    def finalize_options(self):
+        super().finalize_options()
+        # Mark this is not a pure Python package
+        self.root_is_pure = False
+        # Set platform tag
+        self.plat_name = get_wheel_platform_tag()
+        self.universal = False
+
+
+def get_target_platform_for_envs():
+    """Get target platform for environments"""
+    system = os.environ.get('INSPIRE_FACE_TARGET_PLATFORM')
+    if system is None:
+        system = get_lib_path_info()[0]
+    
+    machine = os.environ.get('INSPIRE_FACE_TARGET_ARCH')
+    if machine is None:
+        machine = get_lib_path_info()[1]
+    
+    return system, machine
+
+# Get current platform information
+system, arch = get_target_platform_for_envs()
+print(f"Building for system: {system}, arch: {arch}")
+
+# Build library file path relative to package
+lib_path = os.path.join('modules', 'core', 'libs', system, arch, '*')
+
+setup(
+    name='inspireface',
+    version=get_version() + get_post_version(),
+    packages=find_packages(),
+    # package_data path should be relative to package directory
+    package_data={
+        'inspireface': [lib_path]
+    },
+    install_requires=[
+        'numpy',
+        'loguru',
+        'filelock',
+        'modelscope'
+    ],
+    author='Jingyu Yan',
+    author_email='tunmxy@163.com',
+    description='InspireFace Python SDK',
+    long_description=open(os.path.join(os.path.dirname(os.path.dirname(__file__)), 'README.md')).read(),
+    long_description_content_type='text/markdown',
+    url='https://github.com/HyperInspire/InspireFace',
+    classifiers=[
+        'Development Status :: 5 - Production/Stable',
+        'Intended Audience :: Developers',
+        'License :: OSI Approved :: MIT License',
+        'Programming Language :: Python :: 3',
+        'Programming Language :: Python :: 3.7',
+        'Programming Language :: Python :: 3.8',
+        'Programming Language :: Python :: 3.9',
+        'Programming Language :: Python :: 3.10',
+        'Operating System :: POSIX :: Linux',
+        'Operating System :: Microsoft :: Windows',
+        'Operating System :: MacOS :: MacOS X',
+    ],
+    python_requires='>=3.7',
+    cmdclass={
+        'bdist_wheel': BinaryDistWheel
+    }
+)
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/test.py b/cpp-package/inspireface/python/test.py
new file mode 100644
index 000000000..27151eca3
--- /dev/null
+++ b/cpp-package/inspireface/python/test.py
@@ -0,0 +1,6 @@
+import os
+import cv2
+import inspireface as ifac
+from inspireface.param import *
+import click
+import numpy as np
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/test.sh b/cpp-package/inspireface/python/test.sh
new file mode 100644
index 000000000..2bb07545c
--- /dev/null
+++ b/cpp-package/inspireface/python/test.sh
@@ -0,0 +1 @@
+python -m unittest discover -s test
diff --git a/cpp-package/inspireface/python/test/__init__.py b/cpp-package/inspireface/python/test/__init__.py
new file mode 100644
index 000000000..f1ae1249a
--- /dev/null
+++ b/cpp-package/inspireface/python/test/__init__.py
@@ -0,0 +1,7 @@
+from .test_settings import *
+from .test_utilis import *
+
+# Unit module
+from .unit import *
+from .performance import *
+
diff --git a/RetinaFace/rcnn/cython/__init__.py b/cpp-package/inspireface/python/test/integration/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/cython/__init__.py
rename to cpp-package/inspireface/python/test/integration/__init__.py
diff --git a/RetinaFace/rcnn/io/__init__.py b/cpp-package/inspireface/python/test/performance/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/io/__init__.py
rename to cpp-package/inspireface/python/test/performance/__init__.py
diff --git a/cpp-package/inspireface/python/test/performance/test_lfw_precision.py b/cpp-package/inspireface/python/test/performance/test_lfw_precision.py
new file mode 100644
index 000000000..be91d39d7
--- /dev/null
+++ b/cpp-package/inspireface/python/test/performance/test_lfw_precision.py
@@ -0,0 +1,59 @@
+from test import *
+import unittest
+import cv2
+
+
+@optional(ENABLE_LFW_PRECISION_TEST, "LFW dataset precision tests have been closed.")
+class LFWPrecisionTestCase(unittest.TestCase):
+
+    def setUp(self) -> None:
+        self.quick = QuickComparison()
+
+    def test_lfw_precision(self):
+        pairs_path = os.path.join(LFW_FUNNELED_DIR_PATH, 'pairs.txt')
+        pairs = read_pairs(pairs_path)
+        self.assertEqual(True, len(pairs) > 0)
+        if os.path.exists(LFW_PREDICT_DATA_CACHE_PATH):
+            print("Loading results from cache")
+            cache = np.load(LFW_PREDICT_DATA_CACHE_PATH, allow_pickle=True)
+            similarities = cache[0]
+            labels = cache[1]
+        else:
+            similarities = []
+            labels = []
+
+            for pair in tqdm(pairs):
+                if len(pair) == 3:
+                    person, img_num1, img_num2 = pair
+                    img_path1 = os.path.join(LFW_FUNNELED_DIR_PATH, person, f"{person}_{img_num1.zfill(4)}.jpg")
+                    img_path2 = os.path.join(LFW_FUNNELED_DIR_PATH, person, f"{person}_{img_num2.zfill(4)}.jpg")
+                    match = True
+                else:
+                    person1, img_num1, person2, img_num2 = pair
+                    img_path1 = os.path.join(LFW_FUNNELED_DIR_PATH, person1, f"{person1}_{img_num1.zfill(4)}.jpg")
+                    img_path2 = os.path.join(LFW_FUNNELED_DIR_PATH, person2, f"{person2}_{img_num2.zfill(4)}.jpg")
+                    match = False
+
+                img1 = cv2.imread(img_path1)
+                img2 = cv2.imread(img_path2)
+
+                if not self.quick.setup(img1, img2):
+                    print("not detect face")
+                    continue
+
+                cosine_similarity = self.quick.comp()
+                similarities.append(cosine_similarity)
+                labels.append(match)
+
+            similarities = np.array(similarities)
+            labels = np.array(labels)
+            # save cache file
+            np.save(LFW_PREDICT_DATA_CACHE_PATH, [similarities, labels])
+
+        # find best threshold
+        best_threshold, best_accuracy = find_best_threshold(similarities, labels)
+        print(f"Best Threshold: {best_threshold:.2f}, Best Accuracy: {best_accuracy:.3f}")
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/cpp-package/inspireface/python/test/test_settings.py b/cpp-package/inspireface/python/test/test_settings.py
new file mode 100644
index 000000000..22cd279fe
--- /dev/null
+++ b/cpp-package/inspireface/python/test/test_settings.py
@@ -0,0 +1,68 @@
+import os
+import sys
+import inspireface as ifac
+
+# ++ OPTIONAL ++
+
+# Enabling will run all the benchmark tests, which takes time
+ENABLE_BENCHMARK_TEST = True
+
+# Enabling will run all the CRUD tests, which will take time
+ENABLE_CRUD_TEST = False
+
+# Enabling will run the face search benchmark, which takes time and must be configured with the correct
+# 'LFW_FUNNELED_DIR_PATH' parameter
+ENABLE_SEARCH_BENCHMARK_TEST = True
+
+# Enabling will run the LFW dataset precision test, which will take time
+ENABLE_LFW_PRECISION_TEST = False
+
+# Testing model name
+TEST_MODEL_NAME = "Pikachu"
+# TEST_MODEL_NAME = "Megatron"
+
+# Testing length of face feature
+TEST_MODEL_FACE_FEATURE_LENGTH = 512
+
+# Testing face comparison image threshold
+TEST_FACE_COMPARISON_IMAGE_THRESHOLD = 0.45
+
+# ++ END OPTIONAL ++
+
+# Current project path
+TEST_PROJECT_PATH = os.path.dirname(os.path.abspath(__file__))
+
+# Current project path
+CURRENT_PROJECT_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+# Main project path
+MAIN_PROJECT_PATH = os.path.dirname(CURRENT_PROJECT_PATH)
+
+# Model zip path
+MODEL_ZIP_PATH = os.path.join(MAIN_PROJECT_PATH, "test_res/pack/")
+
+# Testing model full path
+TEST_MODEL_PATH = os.path.join(MODEL_ZIP_PATH, TEST_MODEL_NAME)
+
+# Python test data folder
+PYTHON_TEST_DATA_FOLDER = os.path.join(TEST_PROJECT_PATH, "data/")
+
+# Stores some temporary file data generated during testing
+TMP_FOLDER = os.path.join(CURRENT_PROJECT_PATH, "tmp")
+
+# Default db file path
+DEFAULT_DB_PATH = os.path.join(TMP_FOLDER, ".E63520A95DD5B3892C56DA38C3B28E551D8173FD")
+
+# Create tmp if not exist
+os.makedirs(TMP_FOLDER, exist_ok=True)
+
+# lfw_funneled Dataset dir path
+LFW_FUNNELED_DIR_PATH = "/Users/tunm/datasets/lfw_funneled/"
+
+# The LFW data predicted by the algorithm is used and cached to save time in the next prediction, and it can be
+# re-predicted by manually deleting it
+LFW_PREDICT_DATA_CACHE_PATH = os.path.join(TMP_FOLDER, "LFW_PRED.npy")
+
+assert os.path.exists(LFW_FUNNELED_DIR_PATH), "'LFW_FUNNELED_DIR_PATH' is not found."
+
+ifac.launch(TEST_MODEL_PATH)
\ No newline at end of file
diff --git a/cpp-package/inspireface/python/test/test_utilis.py b/cpp-package/inspireface/python/test/test_utilis.py
new file mode 100644
index 000000000..ce916d127
--- /dev/null
+++ b/cpp-package/inspireface/python/test/test_utilis.py
@@ -0,0 +1,279 @@
+from test.test_settings import *
+import inspireface as ifac
+from inspireface.param import *
+import numpy as np
+import time
+from functools import wraps
+import cv2
+from itertools import cycle
+from tqdm import tqdm
+
+from unittest import skipUnless as optional
+
+
+def title(name: str = None):
+    print("--" * 35)
+    print(f" InspireFace Version: {ifac.__version__}")
+    if name is not None:
+        print(f" {name}")
+    print("--" * 35)
+
+
+def get_test_data(path: str) -> str:
+    return os.path.join(PYTHON_TEST_DATA_FOLDER, path)
+
+
+def calculate_overlap(box1, box2):
+    """
+    Calculate the overlap ratio between two rectangular boxes.
+    Parameters:
+    - box1: The first rectangle, format ((x1, y1), (x2, y2)), where (x1, y1) is the top left coordinate, and (x2, y2) is the bottom right coordinate.
+    - box2: The second rectangle, format the same as box1.
+
+    Returns:
+    - The overlap ratio, 0 if the rectangles do not overlap.
+    """
+    # Unpack rectangle coordinates
+    x1_box1, y1_box1, x2_box1, y2_box1 = box1
+    x1_box2, y1_box2, x2_box2, y2_box2 = box2
+
+    # Calculate the coordinates of the intersection rectangle
+    x_overlap = max(0, min(x2_box1, x2_box2) - max(x1_box1, x1_box2))
+    y_overlap = max(0, min(y2_box1, y2_box2) - max(y1_box1, y1_box2))
+
+    # Calculate the area of the intersection
+    overlap_area = x_overlap * y_overlap
+
+    # Calculate the area of each rectangle
+    box1_area = (x2_box1 - x1_box1) * (y2_box1 - y1_box1)
+    box2_area = (x2_box2 - x1_box2) * (y2_box2 - y1_box2)
+
+    # Calculate the total area
+    total_area = box1_area + box2_area - overlap_area
+
+    # Calculate the overlap ratio
+    overlap_ratio = overlap_area / total_area if total_area > 0 else 0
+
+    return overlap_ratio
+
+
+def restore_rotated_box(original_width, original_height, box, rotation):
+    """
+    Restore the coordinates of a rotated face box based on the original image width, height, and rotation angle.
+
+    Parameters:
+    - original_width: The width of the original image.
+    - original_height: The height of the original image.
+    - box: The coordinates of the rotated box, format ((x1, y1), (x2, y2)).
+    - rotation: The rotation angle, represented by 0, 1, 2, 3 for 0, 90, 180, 270 degrees respectively.
+
+    Returns:
+    - The restored box coordinates, format same as box.
+    """
+    # For 90 or 270 degrees rotation, the image width and height are swapped
+    if rotation == 1 or rotation == 3:
+        width, height = original_height, original_width
+    else:
+        width, height = original_width, original_height
+
+    (x1, y1, x2, y2) = box
+
+    if rotation == 0:  # No transformation needed for 0 degrees
+        restored_box = box
+    elif rotation == 1:  # 90 degrees rotation
+        restored_box = (y1, width - x2, y2, width - x1)
+    elif rotation == 2:  # 180 degrees rotation
+        restored_box = (width - x2, height - y2, width - x1, height - y1)
+    elif rotation == 3:  # 270 degrees rotation
+        restored_box = (height - y2, x1, height - y1, x2)
+    else:
+        raise ValueError("Rotation must be 0, 1, 2, or 3 representing 0, 90, 180, 270 degrees.")
+
+    return restored_box
+
+
+def read_binary_file_to_ndarray(file_path, width, height):
+    nv21_size = width * height * 3 // 2  # NV21 size calculation
+
+    try:
+        with open(file_path, 'rb') as file:
+            file_data = file.read()  # Read the entire file
+
+            if len(file_data) != nv21_size:
+                print(f"Expected file size is {nv21_size}, but got {len(file_data)}")
+                return None
+
+            # Assuming the file data is a complete NV21 frame
+            data = np.frombuffer(file_data, dtype=np.uint8)
+            return data
+    except FileNotFoundError:
+        print(f"File '{file_path}' not found.")
+        return None
+    except Exception as e:
+        print(f"An error occurred while reading the file: {str(e)}")
+        return None
+
+
+def print_benchmark_table(benchmark_results):
+    print("\n")
+    header_format = "{:<20} | {:<10} | {:<15} | {:<15}"
+    row_format = "{:<20} | {:<10} | {:>10.2f} ms | {:>10.4f} ms"
+    print(header_format.format('Benchmark', 'Loops', 'Total Time', 'Avg Time'))
+    print("-" * 70)  
+
+    for name, loops, total_time in benchmark_results:
+        avg_time = total_time / loops
+        print(row_format.format(name, loops, total_time * 1000, avg_time * 1000))
+
+
+def benchmark(test_name, loop):
+    def benchmark_decorator(func):
+        @wraps(func)
+        def wrapper(self, *args, **kwargs):
+            # Set the loop property on the test object
+            setattr(self, 'loop', loop)
+
+            start_time = time.time()
+            try:
+                result = func(self, *args, **kwargs)
+            finally:
+                end_time = time.time()
+                cost_total = end_time - start_time
+                self.__class__.benchmark_results.append((test_name, loop, cost_total))
+
+            # After the test is complete, delete the loop property to prevent other tests from being affected
+            delattr(self, 'loop')
+            return result
+
+        return wrapper
+
+    return benchmark_decorator
+
+
+def read_video_generator(video_path):
+    cap = cv2.VideoCapture(video_path)
+
+    if not cap.isOpened():
+        raise IOError(f"Cannot open video {video_path}")
+
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+        yield frame
+
+    cap.release()
+
+
+def lfw_generator(directory_path):
+    while True:
+        for root, dirs, files in os.walk(directory_path):
+            for file_name in files:
+                # Be sure to only process JPG images that end in '0001.jpg'
+                if file_name.endswith('0001.jpg'):
+                    # Extract the name of the person as the last part of the directory name
+                    name = os.path.basename(root)
+                    image_path = os.path.join(root, file_name)
+                    image = cv2.imread(image_path)
+                    assert image is not None, "Error of image data."
+
+                    yield image, name
+
+
+def batch_import_lfw_faces(lfw_path, engine: ifac.InspireFaceSession, num_of_faces: int):
+    generator = lfw_generator(lfw_path)
+    registered_faces = 0
+
+    # With the tqdm wrapper generator, unknown totals are used with total=None, and tqdm will run in unknown total mode
+    for image, name in tqdm(generator, total=num_of_faces, desc="Registering faces"):
+        faces_info = engine.face_detection(image)
+        if len(faces_info) == 0:
+            continue
+
+        # Extract features from the first face detected
+        first_face_info = faces_info[0]
+        feature = engine.face_feature_extract(image, first_face_info)
+
+        # The extracted features are used for face registration
+        if feature is not None:
+            face_identity = ifac.FaceIdentity(data=feature, tag=name, custom_id=registered_faces)
+            ifac.feature_hub_face_insert(face_identity)
+            registered_faces += 1
+            if registered_faces >= num_of_faces:
+                break
+
+    print(f"Completed. Total faces registered: {registered_faces}")
+
+
+class QuickComparison(object):
+
+    def __init__(self):
+        param = ifac.SessionCustomParameter()
+        param.enable_recognition = True
+        self.engine = ifac.InspireFaceSession(param)
+        self.faces_set_1 = None
+        self.faces_set_2 = None
+
+    def setup(self, image1: np.ndarray, image2: np.ndarray) -> bool:
+        images = [image1, image2]
+        self.faces_set_1 = list()
+        self.faces_set_2 = list()
+        for idx, img in enumerate(images):
+            results = self.engine.face_detection(img)
+            vector_list = list()
+            if len(results) > 0:
+                for info in results:
+                    feature = self.engine.face_feature_extract(img, info)
+                    vector_list.append(feature)
+            else:
+                return False
+
+            if idx == 0:
+                self.faces_set_1 = vector_list
+            else:
+                self.faces_set_2 = vector_list
+
+        return True
+
+    def comp(self) -> float:
+        """
+        Cross-compare one by one, keep the value with the highest score and return it, calling self.recognition.face_comparison1v1(info1, info2)
+        :return: Maximum matching score
+        """
+        max_score = 0.0
+
+        # Each face in faces_set_1 is traversed and compared with each face in faces_set_2
+        for face1 in self.faces_set_1:
+            for face2 in self.faces_set_2:
+                score = ifac.feature_comparison(face1, face2)
+                if score > max_score:
+                    max_score = score
+
+        return max_score
+
+    def match(self, threshold) -> bool:
+        return self.comp() > threshold
+
+
+def find_best_threshold(similarities, labels):
+    thresholds = np.arange(0, 1, 0.01)
+    best_threshold = best_accuracy = 0
+
+    for threshold in thresholds:
+        predictions = (similarities > threshold)
+        accuracy = np.mean((predictions == labels).astype(int))
+        if accuracy > best_accuracy:
+            best_accuracy = accuracy
+            best_threshold = threshold
+
+    return best_threshold, best_accuracy
+
+
+def read_pairs(pairs_filename):
+    """Read the pairs.txt file and return a list of image pairs"""
+    pairs = []
+    with open(pairs_filename, 'r') as f:
+        for line in f.readlines()[1:]:
+            pair = line.strip().split()
+            pairs.append(pair)
+    return pairs
diff --git a/cpp-package/inspireface/python/test/unit/__init__.py b/cpp-package/inspireface/python/test/unit/__init__.py
new file mode 100644
index 000000000..8b1378917
--- /dev/null
+++ b/cpp-package/inspireface/python/test/unit/__init__.py
@@ -0,0 +1 @@
+
diff --git a/cpp-package/inspireface/python/test/unit/test_base_module.py b/cpp-package/inspireface/python/test/unit/test_base_module.py
new file mode 100644
index 000000000..f22b475d6
--- /dev/null
+++ b/cpp-package/inspireface/python/test/unit/test_base_module.py
@@ -0,0 +1,51 @@
+from test import *
+import unittest
+import inspireface as ifac
+from inspireface.param import *
+import cv2
+
+
+class CameraStreamCase(unittest.TestCase):
+    def setUp(self) -> None:
+        """Shared area for priority execution"""
+        pass
+
+    def test_image_codec(self) -> None:
+        image = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.assertIsNotNone(image)
+
+    def test_stream_rotation(self) -> None:
+        # Prepare material
+        engine = ifac.InspireFaceSession(HF_ENABLE_NONE, HF_DETECT_MODE_ALWAYS_DETECT)
+        # Prepare rotation images
+        rotation_images_filenames = ["rotate/rot_0.jpg", "rotate/rot_90.jpg", "rotate/rot_180.jpg","rotate/rot_270.jpg"]
+        rotation_images = [cv2.imread(get_test_data(path)) for path in rotation_images_filenames]
+        self.assertEqual(True, all(isinstance(item, np.ndarray) for item in rotation_images))
+
+        # Detecting face images without rotation
+        rot_0 = rotation_images[0]
+        h, w, _ = rot_0.shape
+        self.assertIsNotNone(rot_0, "Image is empty")
+        rot_0_faces = engine.face_detection(image=rot_0)
+        self.assertEqual(True, len(rot_0_faces) > 0)
+        rot_0_face_box = rot_0_faces[0].location
+        num_of_faces = len(rot_0_faces)
+
+        # Detect images with other rotation angles
+        rotation_tags = [HF_CAMERA_ROTATION_90, HF_CAMERA_ROTATION_180, HF_CAMERA_ROTATION_270]
+        streams = [ifac.ImageStream.load_from_cv_image(img, rotation=rotation_tags[idx]) for idx, img in enumerate(rotation_images[1:])]
+        results = [engine.face_detection(stream) for stream in streams]
+        # No matter how many degrees the image is rotated, the same number of faces should be detected
+        self.assertEqual(True, all(len(item) == num_of_faces for item in results))
+        # Select all the first face box
+        rot_other_faces_boxes = [face[0].location for face in results]
+        # We need to restore the rotated face box
+        restored_boxes = [restore_rotated_box(w, h, rot_other_faces_boxes[idx], rotation_tags[idx]) for idx, box in enumerate(rot_other_faces_boxes)]
+        # IoU is performed with the face box of the original image to calculate the overlap
+        iou_results = [calculate_overlap(box, rot_0_face_box) for box in restored_boxes]
+        # The face box position of all rotated images is detected to be consistent with that of the original image
+        self.assertEqual(all(0.95 < iou < 1.0 for iou in iou_results), True)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/cpp-package/inspireface/python/test/unit/test_recognition_module.py b/cpp-package/inspireface/python/test/unit/test_recognition_module.py
new file mode 100644
index 000000000..7b1e340cf
--- /dev/null
+++ b/cpp-package/inspireface/python/test/unit/test_recognition_module.py
@@ -0,0 +1,223 @@
+import unittest
+from test import *
+import inspireface as ifac
+from inspireface.param import *
+import cv2
+
+
+class FaceRecognitionBaseCase(unittest.TestCase):
+    """
+    This case is mainly used to test the basic functions of face recognition.
+    """
+
+    def setUp(self) -> None:
+        # Prepare material
+        track_mode = HF_DETECT_MODE_ALWAYS_DETECT
+        param = ifac.SessionCustomParameter()
+        param.enable_recognition = True
+        self.engine = ifac.InspireFaceSession(param, track_mode, 10)
+
+    def test_face_feature_extraction(self):
+        # Prepare a image
+        image = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.assertIsNotNone(image)
+        # Face detection
+        faces = self.engine.face_detection(image)
+        # "kun.jpg" has only one face
+        self.assertEqual(len(faces), 1)
+        face = faces[0]
+        box = face.location
+        expect_box = (98, 146, 233, 272)
+        # Calculate the location of the detected box and the expected box
+        iou = calculate_overlap(box, expect_box)
+        self.assertAlmostEqual(iou, 1.0, places=3)
+
+        # Extract feature
+        feature = self.engine.face_feature_extract(image, face)
+        self.assertIsNotNone(feature)
+#
+    def test_face_comparison(self):
+        # Prepare two pictures of someone
+        images_path_list = [get_test_data("bulk/kun.jpg"), get_test_data("bulk/jntm.jpg")]
+        self.assertEqual(len(images_path_list), 2, "Only 2 photos can be used for the 1v1 scene.")
+        images = [cv2.imread(pth) for pth in images_path_list]
+        faces_list = [self.engine.face_detection(img) for img in images]
+        # Check num of faces detection
+        self.assertEqual(len(faces_list[0]), 1)
+        self.assertEqual(len(faces_list[1]), 1)
+        # Extract features
+        features = [self.engine.face_feature_extract(images[idx], faces[0]) for idx, faces in enumerate(faces_list)]
+        self.assertEqual(features[0].size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        self.assertEqual(features[1].size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        # Comparison
+        similarity = ifac.feature_comparison(features[0], features[1])
+        self.assertEqual(True, similarity > TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+
+        # Prepare a picture of a different person
+        woman = cv2.imread(get_test_data("bulk/woman.png"))
+        self.assertIsNotNone(woman)
+        woman_faces = self.engine.face_detection(woman)
+        self.assertEqual(len(woman_faces), 1)
+        face_3 = woman_faces[0]
+        feature = self.engine.face_feature_extract(woman, face_3)
+        self.assertEqual(feature.size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        # Comparison
+        similarity = ifac.feature_comparison(features[0], feature)
+        self.assertEqual(True, similarity < TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+        similarity = ifac.feature_comparison(features[1], feature)
+        self.assertEqual(True, similarity < TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+
+
+@optional(ENABLE_CRUD_TEST, "All CRUD related tests have been closed.")
+class FaceRecognitionCRUDMemoryCase(unittest.TestCase):
+    """
+    This case is mainly used to test the CRUD functions of face recognition.
+    """
+
+    engine = None
+    default_faces_num = 10000
+
+    @classmethod
+    def setUpClass(cls):
+        config = ifac.FeatureHubConfiguration(
+            feature_block_num=20,
+            enable_use_db=False,
+            db_path="",
+            search_mode=HF_SEARCH_MODE_EAGER,
+            search_threshold=TEST_FACE_COMPARISON_IMAGE_THRESHOLD,
+        )
+        ifac.feature_hub_enable(config)
+        track_mode = HF_DETECT_MODE_ALWAYS_DETECT
+        param = ifac.SessionCustomParameter()
+        param.enable_recognition = True
+        cls.engine = ifac.InspireFaceSession(param, track_mode)
+        batch_import_lfw_faces(LFW_FUNNELED_DIR_PATH, cls.engine, cls.default_faces_num)
+
+
+    def test_face_search(self):
+        num_current = ifac.feature_hub_get_face_count()
+        registered = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.assertIsNotNone(registered)
+        faces = self.engine.face_detection(registered)
+        self.assertEqual(len(faces), 1)
+        face = faces[0]
+        feature = self.engine.face_feature_extract(registered, face)
+        self.assertEqual(feature.size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        # Insert a new face
+        registered_identity = ifac.FaceIdentity(feature, custom_id=num_current + 1, tag="Kun")
+        ret = ifac.feature_hub_face_insert(registered_identity)
+        self.assertEqual(ret, True)
+
+        # Prepare a picture of searched face
+        searched = cv2.imread(get_test_data("bulk/jntm.jpg"))
+        self.assertIsNotNone(searched)
+        faces = self.engine.face_detection(searched)
+        self.assertEqual(len(faces), 1)
+        searched_face = faces[0]
+        feature = self.engine.face_feature_extract(searched, searched_face)
+        self.assertEqual(feature.size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        searched_result = ifac.feature_hub_face_search(feature)
+        self.assertEqual(True, searched_result.confidence > TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+        self.assertEqual(searched_result.similar_identity.tag, registered_identity.tag)
+        self.assertEqual(searched_result.similar_identity.custom_id, registered_identity.custom_id)
+
+        # Prepare a picture of a stranger's face
+        stranger = cv2.imread(get_test_data("bulk/woman.png"))
+        self.assertIsNotNone(stranger)
+        faces = self.engine.face_detection(stranger)
+        self.assertEqual(len(faces), 1)
+        stranger_face = faces[0]
+        feature = self.engine.face_feature_extract(stranger, stranger_face)
+        self.assertEqual(feature.size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        stranger_result = ifac.feature_hub_face_search(feature)
+        self.assertEqual(True, stranger_result.confidence < TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+        self.assertEqual(stranger_result.similar_identity.custom_id, -1)
+#
+    def test_face_remove(self):
+        query_image = cv2.imread(get_test_data("bulk/Nathalie_Baye_0002.jpg"))
+        self.assertIsNotNone(query_image)
+        faces = self.engine.face_detection(query_image)
+        self.assertEqual(len(faces), 1)
+        query_face = faces[0]
+        feature = self.engine.face_feature_extract(query_image, query_face)
+        self.assertEqual(feature.size, TEST_MODEL_FACE_FEATURE_LENGTH)
+        # First search
+        result = ifac.feature_hub_face_search(feature)
+        self.assertEqual(True, result.confidence > TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+        self.assertEqual("Nathalie_Baye", result.similar_identity.tag)
+
+        # Remove that
+        remove_id = result.similar_identity.custom_id
+        ret = ifac.feature_hub_face_remove(remove_id)
+        self.assertEqual(ret, True)
+
+        # Second search
+        result = ifac.feature_hub_face_search(feature)
+        self.assertEqual(True, result.confidence < TEST_FACE_COMPARISON_IMAGE_THRESHOLD)
+        self.assertEqual(result.similar_identity.custom_id, -1)
+
+        # Reusability testing
+        new_face_image = cv2.imread(get_test_data("bulk/yifei.jpg"))
+        self.assertIsNotNone(new_face_image)
+        faces = self.engine.face_detection(new_face_image)
+        self.assertEqual(len(faces), 1)
+        new_face = faces[0]
+        feature = self.engine.face_feature_extract(new_face_image, new_face)
+        # Insert that
+        registered_identity = ifac.FaceIdentity(feature, custom_id=remove_id, tag="YF")
+        ifac.feature_hub_face_insert(registered_identity)
+
+    def test_face_update(self):
+        pass
+
+
+@optional(ENABLE_BENCHMARK_TEST, "All benchmark related tests have been closed.")
+class FaceRecognitionFeatureExtractCase(unittest.TestCase):
+    benchmark_results = list()
+    loop = 1
+
+    @classmethod
+    def setUpClass(cls):
+        cls.benchmark_results = []
+
+    def setUp(self) -> None:
+        # Prepare image
+        image = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.stream = ifac.ImageStream.load_from_cv_image(image)
+        self.assertIsNotNone(self.stream)
+        # Prepare material
+        track_mode = HF_DETECT_MODE_ALWAYS_DETECT
+        param = ifac.SessionCustomParameter()
+        param.enable_recognition = True
+        self.engine = ifac.InspireFaceSession(param, track_mode)
+        # Prepare a face
+        faces = self.engine.face_detection(self.stream)
+        # "kun.jpg" has only one face
+        self.assertEqual(len(faces), 1)
+        self.face = faces[0]
+        box = self.face.location
+        expect_box = (98, 146, 233, 272)
+        # Calculate the location of the detected box and the expected box
+        iou = calculate_overlap(box, expect_box)
+        self.assertAlmostEqual(iou, 1.0, places=3)
+        self.feature = self.engine.face_feature_extract(self.stream, self.face)
+
+
+    @benchmark(test_name="Feature Extract", loop=1000)
+    def test_benchmark_feature_extract(self):
+        for _ in range(self.loop):
+            feature = self.engine.face_feature_extract(self.stream, self.face)
+            self.assertEqual(TEST_MODEL_FACE_FEATURE_LENGTH, feature.size)
+
+    @benchmark(test_name="Face comparison 1v1", loop=1000)
+    def test_benchmark_face_comparison1v1(self):
+        for _ in range(self.loop):
+            ifac.feature_comparison(self.feature, self.feature)
+
+    @classmethod
+    def tearDownClass(cls):
+        print_benchmark_table(cls.benchmark_results)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/cpp-package/inspireface/python/test/unit/test_tracker_module.py b/cpp-package/inspireface/python/test/unit/test_tracker_module.py
new file mode 100644
index 000000000..0d85e98fa
--- /dev/null
+++ b/cpp-package/inspireface/python/test/unit/test_tracker_module.py
@@ -0,0 +1,130 @@
+import unittest
+from test import *
+import inspireface as ifac
+from inspireface.param import *
+import cv2
+
+
+class FaceTrackerCase(unittest.TestCase):
+
+    def setUp(self) -> None:
+        # Prepare material
+        track_mode = HF_DETECT_MODE_ALWAYS_DETECT
+        self.engine = ifac.InspireFaceSession(param=ifac.SessionCustomParameter(),
+                                        detect_mode=track_mode)
+
+        self.engine_tk = ifac.InspireFaceSession(param=ifac.SessionCustomParameter(),
+                                        detect_mode=HF_DETECT_MODE_LIGHT_TRACK)
+
+    def test_face_detection_from_image(self):
+        image = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.assertIsNotNone(image)
+
+        # Detection
+        faces = self.engine.face_detection(image)
+        # "kun.jpg" has only one face
+        self.assertEqual(len(faces), 1)
+        face = faces[0]
+        expect_box = (98, 146, 233, 272)
+        # Calculate the location of the detected box and the expected box
+        iou = calculate_overlap(face.location, expect_box)
+        self.assertAlmostEqual(iou, 1.0, places=3)
+
+        # Prepare non-face images
+        any_image = cv2.imread(get_test_data("bulk/view.jpg"))
+        self.assertIsNotNone(any_image)
+        self.assertEqual(len(self.engine.face_detection(any_image)), 0)
+
+    def test_face_pose(self):
+        # Test yaw (shake one's head)
+        left_face = cv2.imread(get_test_data("pose/left_face.jpeg"))
+        self.assertIsNotNone(left_face)
+        faces = self.engine.face_detection(left_face)
+        self.assertEqual(len(faces), 1)
+        left_face_yaw = faces[0].yaw
+        # The expected value is not completely accurate, it is only a rough estimate
+        expect_left_shake_range = (-90, -10)
+        self.assertEqual(True, expect_left_shake_range[0] < left_face_yaw < expect_left_shake_range[1])
+
+        right_face = cv2.imread(get_test_data("pose/right_face.png"))
+        self.assertIsNotNone(right_face)
+        faces = self.engine.face_detection(right_face)
+        self.assertEqual(len(faces), 1)
+        right_face_yaw = faces[0].yaw
+        expect_right_shake_range = (10, 90)
+        self.assertEqual(True, expect_right_shake_range[0] < right_face_yaw < expect_right_shake_range[1])
+
+        # Test pitch (nod head)
+        rise_face = cv2.imread(get_test_data("pose/rise_face.jpeg"))
+        self.assertIsNotNone(rise_face)
+        faces = self.engine.face_detection(rise_face)
+        self.assertEqual(len(faces), 1)
+        left_face_pitch = faces[0].pitch
+        self.assertEqual(True, left_face_pitch > 5)
+
+        lower_face = cv2.imread(get_test_data("pose/lower_face.jpeg"))
+        self.assertIsNotNone(lower_face)
+        faces = self.engine.face_detection(lower_face)
+        self.assertEqual(len(faces), 1)
+        lower_face_pitch = faces[0].pitch
+        self.assertEqual(True, lower_face_pitch < -10)
+
+        # Test roll (wryneck head)
+        left_wryneck_face = cv2.imread(get_test_data("pose/left_wryneck.png"))
+        self.assertIsNotNone(left_wryneck_face)
+        faces = self.engine.face_detection(left_wryneck_face)
+        self.assertEqual(len(faces), 1)
+        left_face_roll = faces[0].roll
+        self.assertEqual(True, left_face_roll < -30)
+
+        right_wryneck_face = cv2.imread(get_test_data("pose/right_wryneck.png"))
+        self.assertIsNotNone(right_wryneck_face)
+        faces = self.engine.face_detection(right_wryneck_face)
+        self.assertEqual(len(faces), 1)
+        right_face_roll = faces[0].roll
+        self.assertEqual(True, right_face_roll > 30)
+
+@optional(ENABLE_BENCHMARK_TEST, "All benchmark related tests have been closed.")
+class FaceTrackerBenchmarkCase(unittest.TestCase):
+    benchmark_results = list()
+    loop = 1
+
+    @classmethod
+    def setUpClass(cls):
+        cls.benchmark_results = []
+
+    def setUp(self) -> None:
+        # Prepare image
+        self.image = cv2.imread(get_test_data("bulk/kun.jpg"))
+        self.assertIsNotNone(self.image)
+        # Prepare material
+        self.engine = ifac.InspireFaceSession(HF_ENABLE_NONE, HF_DETECT_MODE_ALWAYS_DETECT, )
+        self.engine_tk = ifac.InspireFaceSession(HF_ENABLE_NONE, HF_DETECT_MODE_LIGHT_TRACK, )
+        # Prepare video data
+        self.video_gen = read_video_generator(get_test_data("video/810_1684206192.mp4"))
+
+    @benchmark(test_name="Face Detect", loop=1000)
+    def test_benchmark_face_detect(self):
+        for _ in range(self.loop):
+            faces = self.engine.face_detection(self.image)
+            self.assertEqual(len(faces), 1, "No face detected may have an error, please check.")
+
+    @benchmark(test_name="Face Track", loop=1000)
+    def test_benchmark_face_track(self):
+        for _ in range(self.loop):
+            faces = self.engine_tk.face_detection(self.image)
+            self.assertEqual(len(faces), 1, "No face detected may have an error, please check.")
+
+    @benchmark(test_name="Face Track(Video)", loop=345)
+    def test_benchmark_face_track_video(self):
+        for frame in self.video_gen:
+            faces = self.engine_tk.face_detection(frame)
+            self.assertEqual(len(faces), 1, "No face detected may have an error, please check.")
+
+    @classmethod
+    def tearDownClass(cls):
+        print_benchmark_table(cls.benchmark_results)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/cpp-package/inspireface/python/version.txt.in b/cpp-package/inspireface/python/version.txt.in
new file mode 100644
index 000000000..98f829538
--- /dev/null
+++ b/cpp-package/inspireface/python/version.txt.in
@@ -0,0 +1 @@
+@INSPIRE_FACE_VERSION_MAJOR@.@INSPIRE_FACE_VERSION_MINOR@.@INSPIRE_FACE_VERSION_PATCH@
\ No newline at end of file
diff --git a/cpp-package/inspireface/test_res/data/RD/d1.jpeg b/cpp-package/inspireface/test_res/data/RD/d1.jpeg
new file mode 100644
index 000000000..b6f841b31
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/RD/d1.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/RD/d2.jpeg b/cpp-package/inspireface/test_res/data/RD/d2.jpeg
new file mode 100644
index 000000000..c212929b8
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/RD/d2.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/RD/d3.jpeg b/cpp-package/inspireface/test_res/data/RD/d3.jpeg
new file mode 100644
index 000000000..6030d471e
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/RD/d3.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/RD/d4.jpeg b/cpp-package/inspireface/test_res/data/RD/d4.jpeg
new file mode 100644
index 000000000..f5863d84b
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/RD/d4.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/RD/d5.jpeg b/cpp-package/inspireface/test_res/data/RD/d5.jpeg
new file mode 100644
index 000000000..56078d20e
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/RD/d5.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/attribute/1423.jpg b/cpp-package/inspireface/test_res/data/attribute/1423.jpg
new file mode 100755
index 000000000..3bf9a0bdc
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/attribute/1423.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/attribute/1561.jpg b/cpp-package/inspireface/test_res/data/attribute/1561.jpg
new file mode 100755
index 000000000..f7adf7ff8
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/attribute/1561.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/attribute/1996.jpg b/cpp-package/inspireface/test_res/data/attribute/1996.jpg
new file mode 100755
index 000000000..5319219bf
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/attribute/1996.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/attribute/2963.jpg b/cpp-package/inspireface/test_res/data/attribute/2963.jpg
new file mode 100755
index 000000000..91186b9ea
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/attribute/2963.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/attribute/7242.jpg b/cpp-package/inspireface/test_res/data/attribute/7242.jpg
new file mode 100755
index 000000000..7da773dcc
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/attribute/7242.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002.jpg b/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002.jpg
new file mode 100644
index 000000000..70ffc90a0
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002_w250_h250_c3.nv21 b/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002_w250_h250_c3.nv21
new file mode 100644
index 000000000..df2e48a69
--- /dev/null
+++ b/cpp-package/inspireface/test_res/data/bulk/Nathalie_Baye_0002_w250_h250_c3.nv21
@@ -0,0 +1,246 @@
+PUajeXLE:3+.09DOKC:-$#"!!#&&"	
+
#*058:;:70*%  $*7EPWYYWRMD<3*%!YZdqq^I?FHE@8:;><>>3)!"   #%&"	
+
#*058::960*&!$+7EOVYYWRMD;2*$ "#"  





\]domWFEIPOKFF:.+68+&"   "$$!		"*/47::960+'"#*7DMTXYWRLD91("")-*+--+)(((&#!



SYcmhOAFONGFHK<,&20"%'#"%$!
		")/4799861,(#"*6CLSWXUQLD90(!"/;<71,('&%%&'())'&%$" 
+
   
QV]gdPAEOE746<7300& #"!$&%!
+	!(/3799862-)#!)5AKRUVSOJB91("!'/5-+($!   "#""!"##! 	  

NOU_`SHIOB314::;7.%")&%%!%%$ 
+	!(.3699873/)# (5AJQSTRNI@80("$'( ""!!  !!! 

GGP_bWNMQF99=D?:8227@<5-!$%#
+		
!(.36999740)$(5AIOQSQMH?5-&!!#%%# !%('&$$%%$&&%$#####"!



OGL\b[VVXSIHNXTLA;99DID5'!%&$ 
+	
+")/36999851+% (6BHNPRQMG>4,%!"$%"   $)*+*(')**))***)(($$#" ! 


[NMYebYZ\[PMS]lsbI2.:A<3)$#&#
+#(.2587641-($ (4?KNPTSNI?4*&$"  %   ! "&*+-*'(,0.,,,,,+*))(%$%$"!!! "! 
j^X^c`VWad]VNHIIQKFB>93-''%%$ !%'% 
+!'-0466540,(%!(3>IMORPKG@4,)'$#!"$$'!!""$'+-..,+.21-+)()***))(&&'''&$!%""!
pidhhh]]^`ZQIEHHCKUTF92,,.//-(##$#&(%!
+	
 %,0367651-*&"'2=GLNOLHE?8/+'%% !%&&  $''&%&')+.2650.0/,,.01/./1431/...,&'%! mnmtqqbaec^XQLIDDMYWH=83541.,($$##$%#
+		
%,04788740*&!(3=HLNOKGE=8-'#!!!(01-'$ "'.31-*)*))*1681*)+-25:<;:<@AB@;8740*.0+&! "!_fmyxxjge^[YWUQKPSWRGA@>?8-%$"    "$"
+		
 &-0489973/)$(3>IMOOLIE<4+&!!1??6+'&(,27:83/.-+('*,.,('*/--/37<AEHJIEA?;537952/+&!"! OWcu{~uup^RLKQX\_\VRQQOMLC4)'&$#""$&$!
			
 &-1599862-&! (3>HLNNLIE;0)&!"))%!%0982,+39<@ABB?;963/.+)*,.--.&%$&)09AHKLIGD@;9>A>:9:92.)% "! 

EJXlx{w{u[JBCMYbdb\[bb]]^XMC>820,(''%"		
+!&+1599861,'#!(3<EJLLJHE;0)&#!&4DOLE?;:8669GOOOLLQQOJC=:86435750+*,-+(*2:ADGHHHFC>CFC@@CEKHB:.#!  
;>Nbsvv{vWC=@LV^`feejfbhmle]UKA<5-'%" 	
+!%)/478751,)&!!(2;DJLLJHF=9.++.;NaptpaSMKINRbgb]WW]`_WLD@>?@@?>:3+/133/,-07;@DHKLKDEEBACHLMNNI@5+%#  6;I_mmmsmQ>8CMTX]bfhjiffijkmi^QI@60/*!
+		
+
 #(.467752+($!!!$*5=EKOMIGFD@947I[dq���sbZ]entznaUUY[VNH>9=>=;::71,/--/-**.0147;AEE??@@@@BDGLPROF:0,'"  8<I]jigjcG52?HMPU[_dikihecbfoqi]OB60(!				"'-247862*(%!!# $$&/;EIMOMJGEDE?DSchhr���sc^do|z~{oaSKFEA9226642110.,,,+++)(*-))+,.4894686459=FJOUWRG>5/'"!" 7;FXdd__Y>01?FIJQX\]afkoc^Y[dlnl_N<.% 
+	
+"&-247972*&#&*##*+,3?HIJKLJHHHHIZq|vkktzvh]WV[`^^_^_[OC741.480*'&&%#&*+-/0.,***+-.-022.///03577:AMUWRNC:.$ "  !! 37AR^^YVS>68DHJLUZ[XZ_eiaZVWXX\caXJ;,$	
+
+
 "',136751*)&# !$$)31(#)1446=EGGILMNQU[`lwzvnidbZPIC=<:?DO^t}kT=14795)$#! "&'+/010.+..0/-./-,**,--./..1:CJOSOJB5*'$ 
15>MYYSQNBADKJLNRTSU[`a`^[\_\TNNRUUK9+#	
+
 "'*047632-(# !)-39:8-.39:7:@CEGNSW_fsw{xqjc[TOF><8645=J]q��|_M<9<:1*$%$""$&&$%&(,266211/...,)(+.*$%+,+*-19DNQSRH:0( 
36?MWVQPLGKLMJJMPQPRX[\]_ejiaVJBDFHC8.&"'+37:85421)#!%+=HQOF<871,-4736<CEKT\cltz}|viZPJFDB?BHMNMTbnnmleYYL>90$!#$#!!$$"$-3655899:84.,.1/(',--///15:CJSTK@3("%$! 




47@MUSPQOKNJIFIKRSRQQPTZfouobVNIID>81,& 	
!',48;9441' #.6?CIJC:440**0326=HJPX^dillnfa\UUWZXVW]fmmoq{ygTLO^e`SJ?3.(&# "&)/29>BEFC>:78840/2345641.2:IVZTF8-)$ ! 25>KSQNQRMMEBCGKJLMPQQSYjpush\VVYSK?2*&"
	
"'+26961/,'$$',5>A6/-263/.0--267<CHJMQTTSRQWQQYakqxwxz||sg``hfWF:>CFFEFB;773-)%" ### "&*-39@IQSQNJEA=964214<A@<427ERXTM@2&#"
36<DKPQQOIGC=AGFBBHNQYcghgfhid][\_XK>0#
+	

"&+066465-&(/9BD@:0+(+0233248<?BEGHHFCBA>:9=GSct���}}xpf_WONPMF@<3967<:4431/-,)'&!$*((*/6<DOW[]^YQH>855326>DEA826?FMRME4+&&!




8<BHMPPOJEEE@@FHDDHOV_gka\]hlcZYVTQMG:* 	
 "',27511/*+4>DFFA;4115:<<<<<>BFJKJIFA=;;9528EWiv~��}wutsqpoaQHD@=;>582/220013443/,'#!!!&)&##&+06?HPV]`^YSMIC:1-.49=;;;::?ELRL>/'  ! ! 




EFGHJLNNHEHKE@CFMORVYZ\__^`jneXQRMD>==3%
+	
+ #)-14661..5=EIKKF@::>DHGFFDGJKJIHHE@954665:@HR\bkv�~{zytnnkaVPI?859260,,-/369;<;61/---,,*&   $'+29DKTZ_dec`VJB;40/16<?<89;CKRNC1'$$ !!"" 



UNGEDDGJJJPSLCBDLQTSRNOV`hkknl^OIB=;;:1"
	
+
#)--/7;416@@ENTQH@;@GMMJHIIJJGB?>=;71/15:=DMSWYVV^gkqy|p`ZVQMLG=753.31/-.1248<@@><88;=<93.,)(((')-5>IRZadd]YVVTMC<027<>?>=>@JRXH8-!!"" 




YLCBA?AFIMTWSLIHKLIGGHO[clonruogTE?DA5)"
	
+

"(,--1568?GELXYK=7:BKPOKIIOJC<97511.,-06=CLTY]`[RRZ]ajqiZRGDEHGA<;60201--//1369<>>>BHNQNHC=:851-*+/5=EMU[]Y\`cc^VOB>:89<@CADHJWWK<( !##"QF@EHFGKIOUXXVROLIFGLQYbdbdlrqt{u`MC:/&"
	

!&)01-1=FJGKUZO<56AJSUTSQOQOIB;61,*)*,04<CNTX]ddbccdgnriWIDBCFHEB@>73.-()*.011258:AELU\[TQNMKGB<743336<EOV]afhjg_WSOG?86;ADGGGOUXXG3!##!QJHNTUVXSVXYZXSPIFJRW[_b]VV`eacnxq_I;1%	
!$(/329IQLIQXQA7:BOV[ZYZVPKOOG>6/+&%',/39@CGP[chlpumd`chgaXOB;9:;;;84/-''(*.23359;=AFOWWQOX[[WSOID<8457;CLUW]elmhb^XQKD=:;BDFLOSZf`O5"!" 

]WTV[^`ba`\ZZUNJHGOXYXYYYUTX[XVZekeSC5'	  !#)+2:ERQFLWXK>=FLW\]YXZTJCFE?94/+$"%+/38>?;@LV\aeof_]ew��vfN;322103421)**-2786579=?CKTTNMQW\\]_]XND:547ALMQYdknoqg\SRQJ@::?BKOTSVhgV<)!   

hb^ZZ[]aefd^WRLHGHLPOLMQQUX\a`_be`\XNB3&		


 !$',3=JPMIT^[JCHNTXYXX[VMID<54552,! !&/4;EMH:2?T^\URMJGFP_infTH@72/,-0001149:86569=?BGNSSPVZ[Z]cedbYQJ@66=GIRZ`ipqoeYRUXND??GGITPGOdm[=#!$"!! 

ljjcYSQV^aa\UPNNKJHFEFILSX[_cdejgaYRLF8)			


 "&-9EKJLT[YUMLPTYV\^^_YNHC;0+1<:.!$*1;CKMPND8:EPSTSNE>=EOTRNJB<2-++*,13686442124:@EIMRVYYXVVXYXZbih`SC88<HRTYdkpoeVQW[]TJEBEMQOFQ]aW?&!&&%# 

motodYSVZ[\YTQQRPPNLNSUUUYZ[\^dkkcVHA?7*
+		
 "(1?JKGMZ[RQX\\\[^cc_\UKD=:2*,8<7'&276:DRXO@:@NWVOF;5.06;>?=940,(%'+-.///23100146;BLSTTW\ab_\XQGHYmpgQJJPRSWZenoh^WZeg^PGDGRYOHJVaZA(#&&#! ljljfa]^a]ZYXVUUVWWWZ]]ZZ]_][Z^cfcYI<6/'
	

  &,7CJNOSXZSWa_\cfgf_UOMF>653-).:A7*'+*08I_jbVFC><@A@@5+''(+..0/.*&$$"&*,/2200010.2<GMWXY[^adecdc_^bdda^]]VOOS\iswp`XZej`THCLQRMKQ\bS:'!$$#!qha]^acghc^]]\YXXVVY\]^^`^ZSNPW]accZI;1) 
 &1<FMTXWTX]in^Vbhb_VMIJD:3/-)%)8EIB/"$')--.7CNU[RF?>>AF:/)&"!##'),-*'$ !(.011266:9439?AHPUUU\ejihhhgecaaejibXRRZhnsvpg\_hgcYMJGLQUSSZYO<+ "$# qid`_bdigeca_][ZWTUZ\ZZ]]ZUNHLW]bbc^O>0)!%7ALSVUST^gqscZ`]YWTPNNE831-% $3?NO<'$$!$*//'!#2BZgbJ;<<7730,$!'+,,+('.210/26:<;8:?A>?FKMPY`biifb``cegihikf[STZ_lvyxn`_aikbVLMQUUSV[]O8$  """"



iimnkhcecefda_][XUUXWRNOUY__[\abb^ZSD4'
+	


"*9FQTQMR[kmigaae[UUWURPG;576-"!+3CI?.&" " .::1*0=UjhM@E@1453.("$&),-05642/.0657=FMMJABABHT_d`aa`^]]^cfedfebaWRTgty�{k_[`hld[UTV\_\Z\YF3)$""""!ejuzwofdcgihgea\WTQQPLJJQWac_bhic^XO?/#	
+

!(28HSQJKWflmda`di[UUWUPNI@69<7+#%*5=>3(#$%!)9HG915J_h\NE>7/0-*,*$"!&)-135763323:DKLIHQQOJFJU^b`ciib^`addcdccgi^Wblw�|kaZ^mnia^[^b^Z][PB4'""$$#ijtyztgbebccaff_WRNPTUSOLMJLWahoig`TF9)			
%-9;EPIEWgmhecffecYU[ZTGAD>55?D9*#$*/82&"#&#'FOA3,9K`h[G@D9('))+()" !&$+..29<:;:@JJC@CHLRXZYWV^`dilkfbcbcffffdhngibiw��|cVYcjolf`^````[UQD/$$"$$!okpuvoede_\\]ab^WPLNQROMMSTTZ`dijhbWI=0%
+
+
#)09?>BAKbmkhbcghf_WRXZS@49<32=GB2&#'-43("()$/ELF;19EPVWTO@+%$&-,,()*)&$$%&(**-5<?@@BFIJJJFBBKV]^\`befhihfdfijkmh_\_j{ldgo~�sc]`eouqg^\`cd`VQL<-'#""unotskdfje_\[ZWTSRSVWVWY^cb_`dgif`ZPA7.%
+	"'*.149BGPdprjbcegf[URW\V?,/9848DI=-%")15*$%(+;FPWOE63BUZSA0' &),,0442/..-,*))-26;<?BHOVZWWSKDFR^cfgecdeeeghghnpldbYgqpdeu��wjd^epyueYYcpj]VSE3($! {usuqgacjhb^[XTRWY\`bdefffcabegh`UMF9/)#
+		
 &'()*/7EP\ksrldfceh^ZUW][E+'1:;9AKF8-$&/5/&$!+"%//,=YfZF65@HH?3*! !#%',//.0344431/.-/2=JPQV^_][WSNHEOX`ccdee``cedgpwwk\^amssit�rb_gv|rd[bqvh^VJ;/&" �}zwod``bc`\[ZZ\ZXVW^dc_][\_`acd\QHC9-&" 
+		#''&()/9HYhsvrljkdekdc[WZ\M2$%4>>AJMH=/(/346-*1*%'#&<UYXJ7,2=>5,# !#%$(.15:>@?:555<EGDFMSZcig`WQLRWXX]dieejnmlmnstuxphebafy���sb^m|�uc_juoe[RH9+$ �|xphgh`aba^[YXZVQPT[\ZYX[][Z]_XTLC7)	
+		
"&'&$'+2AP`ltxvnpoedieddZWYT?("2=:9CQVJ8/511:5689/# $3@PSE3/484-&"!!"#   !$&+3:@CB?;;=BGKOSXYZ]`ceedc]VRU]dhfhlryxrjoiiqqkhb`n����sccn}~pdaippg\P?0($





}~{ztmpqfadid[TOOPQQQQUX]^a_ZXWSTZTB3%				!%&%$#%,<IUesyxqpsneceddgYONTK2%5?939JWM;8<3/7777@:.,'"3FOH:101-&!$),)$!!"!!%+047:DINORVXWWVSRSX^accba^]^adbaciryzxvspox}qd^bk����}f[fv{q`dovtk[G8/)!




|~{|worsh]_jg]VQRTZ`_ZY]adigb_WJRa[C2'		
 $#!##%.?P^mxyvmorjcbcfibSEBMN8)5CA82>OI>==75<:15<;<A6#,AJB5-.($"!%)53,&'+(#! "$*1;AFFHNW]`_^\\\[ZUVZ_cehlpnnja^dlrtrrusttkea^u����u_[bsyojnszzjQ@4,#
�~woqvnabkida\TMS\`fgfhijprl`PMSP?8,#	 "$*4FX_pysslsled_alqiQA<DH8*6JK:2<@NIB78EPL:489?KG2"!")6@;4/*+,'%(363--2.$ #&'),00135;GS]acddccd]ZVX^adjotvrj^[ckhjptrppqog`cn����r`]issoxtz�zcI9/#
����zrsxuhekifgf\RQSW`inqrrtrmh`MF?780&
 %0=N^htwqsorja_X[jum`H69EA-8FG939DTXH65I`eT<<:=IQD/"&#+540.056/('-12//1/*""$'++*(+*(%%*3;BJPSW\aehe`^]Z\bljiklheeebdiljhilprqjcm����m[\ek|}qs��rV@1&!

����{uvyxojllijmkaXRQYcoruvvnd`ZQLNLA. 
+!"! #,;JWbmutqspndZZQUbousW;4BK<4<=87:HRXN>8GfvkNE?<?LL:*%#%*),-18:3+)(*++**-1,+**+,+**,*$!#&'+158@JU\`dggebchqiddgkljea_``^]_dltyrdal����kZ[g~�ss{�}eK4(%

��~{xxyxtopqmkouoi_YW[ft{~}ti`UJBFI>-#"#!$0AOZ^ipqrpjf_VYQRY`ek`I8:HK<??=@@HEJNKAEazz\JDC>GUS9'""#'%%++,/1/,..---)$(0112441//.00+((('%$"$,7CNQW[]`dinjjmmhggdeb_][XWW]dimplfet����l^ev~{|z|�r[;)'
�}yz{{zyxvrrurossstog_\bltxxvriZOFFG8,#	
+"%$!%0BQW[gnopi`]\XYOOUXX]dZE;>IGJKKPRVLJNNJGXu�lRMRGBR_L7,'&'%(*)&%'(,154670%"&))1=B=52231-)(*+*'$#$&.9HSXSPT]emospgcca\\\\[XUTTY[^hmjhjw����ndmy��}w��lH.$

�{vw{|zvwwsqtttypqvvpicefmqtwyrcLJNL6)"	!&'&(0CUYbnrokaYVYZTJHU][RSNF@:>AEJNU[`\[TNNIKe�g^aR==LXK9)&(((()&%&'+0558;7.'&%!#/9;877873.*'%()**&&9GU[[VNJSYcklic\UUUUUUTTQVZ^ba`d_al����ian���tz�z]<&


�zuwz|yrrtsqqrv{tortrponjlmprrmaVOKG6+!
	
+!# $&(-6I[couslc[VTVYOFBXeeSFAACGKLJOY_cadeZOQOJZr�xmh^F9?SP>/..,()**)*++-1148<<73.'!!'.369=?<95-%!%)--(""*;MUPJP[]^`chgb`][WSPOMLOTY_`XSVYYbr���x`d}��zu��vV3!�yvx{|xrnrtspquyzonpqvxwvsnnkhf`NG@?2)
		"'*+&&'-7CTemxwpg]WTUSWMGBZhh[H?67HUUNRbheZacYNT\Y\gt{qfbO=>EK@994/,*,,,..,+)*-3<B>864.%"%+.6<?=?@9.')*.31&"+>T_ZSS]^]^a``djhc\TME??ELX`\UTMNX_i���o_l}��wy��mB* ywy|znloqqooty||uquutw||zpdcd^RC71)* 	
+(0/)*,8EL\m{{xm^VTSWYZOII[pm^RD75ARWU\igfb]`cYX][^cnx|wjTA<9A<<<551/,*,./.-)()*3=;78<;0*(&).6:8:BGE83*,030'! .FY\VWZadbadgikklh^VPIGHFHWgg]UORUVk���ocs���o|�tT0# 
|xxy|}wkflrtpmpu}���zqrxtx�oecaWH;1&%


'(%09DKRcu��xgVQTWZ\]SNOatrcWI=8APXZcpokdaammd[W]fjt��{hRA897>;/2631.-----,/0+-6769::653,(,0335;CHC<1./45.""1F\ed`bhiegosmkmnhda^]YPIQ_eb\UQRO[p��|hg{��vuzyc=%utvwxxpdaisupkms|�����y{{z��vnkh^L;-!3>JMZk|��ucTPVZZ]_YVYl|uf[OB<BOX`kyxna__t�wbX\fdeozzr\C;88@9+/5430-**+-/6:2076340269:2*,,04339AGD=8788/(!(7Ujpgaflinxzwqnnkf`ZZ\ZX]fllpi^]Z[aq��qbm��smtlO2"
mpsuuqh^^fosokmrx}{����~���~wvtn_K7(
+

!#39GPfw��~scWV[^]]`^\\n}th_TD=BQ]jw��tbaav��ob_de^\dnpaLC>=@8.02553.)&')+16339857/059<83-.4;:425CHJGC>;1,#%).>Wmsliiinss{xwvtl`VUWXX[clquuoqkd]bq~yhep~�wjrpa?)
imqssmd]]bjqqnoruxx}��������~wrkaN</  
+
+	
")+48JYt��wk^UW\]WUY__^q~ukeZJAGWft���vefjy��}j_^ccYT`lfZLD?>94558983-)('*+/039<>A4137;96249?B>61=FPOLDB:2*')+/@]rupghlpuz~���zma[XSOPXdmkqxvi]]huzqheq�wovvqM/ 

gkorrkca_agptsqrvu~|���~���wkc\OD8&!	"(.5=Tf|��vqh^WY[YPNT^``u�skhbULSboz���qdiq{���q]T_h]KQbheYKBA>7698:97420-,,2568:AE8237:86;;;<BE@78AMPRJJA;3++103D^osljls��}��{of_TKJP\ejnw}pc^_kvvndgz|wyvy]9 

ehkpqjdfdbdmtusrvv||z�������th]NF?7'$ (17Gcw��~usng`_]XVUX^^_szoghg_X[gs{��qks{����}eT`e^NKXcf`SIGC93745668:72-,37755;D=879:79@?:6=HHA:<FLUOOEA91065.1I\mttor}���}���~wl\MHKT[jqso}ymb[`jomd`j~yzuzkH#efiprjekhbaiqtsstzvs{�����~yqngZLA91$#
+!.7<Qr���wvria^[UZZ[^\]qumdfid\[dmu}�~us}������ta`\[VNPY]^WOJF:0122138:72/,03688=A@>97529==72:INIB=@FTRSJG=8777339Gb{�yux���������whZRQTVgwtcz�ykaYZbhe]bmoywzx[1
bcdvnlkkkgdglnnquvvqsxz~��wrngaWI;2* 
+		
",9F`|����~yphaZYW[^[]^\huoccnslaftwz�~ot��������ra[YYTORXXUTQG<879867;=70---18>?@AA?;777699878>GNJFDGQ][NHC@:1068<>Nh}��}|���������xhYV[V\s|pm{�ylZV[`b^_ejnst|n="
_acyqokjkjfdefhknprnmsux��volgaS?1+(" 
+	$-?Oj������yne^^`aa_[^^\iyvfbm{{qv��~|t}��������|j`\][TQRWWXSJCA@988:>@;4/0038=AEGEB@>=:766:<839CPMJKTac\OD=9.*4@C@CTm��������������wg]_[Scy�co�rcZWY[Z[_cgnpxsS.
]_a{xplghlib]\^behnkgkmluzurpkaM7($%"!
+%/DXv������zoc^dhmla[_\Yi|wd]j��yx���}~~����������yh]\\XTQWZ\XOKHE;:9;?@=7554479>DJIGEC@<9426<=759HNOKNXeo[I:5/*-5<CGL]s����z���������vid_OTp�ki{}vh`^[ZZZ[_hlrvkC_`^y~mmadjkbYTUY[`kjdfgdhoqtsnbL6$! 
+
+'7Kc�������|tgbkpvtd]^WTf{zf]l���wz�~x{������������sb\[[ZWVY\[VSNGB?<<?@=;<:6578<BGKKGBA?=744:@@;7=GKHEJ]sj[H:42/*1ERSXbn{��x���������zpcTRe~kryuoj_ZWWWXahltvX&__[u�glZ`gidYOLPPVce_ab`_dgnniaM5#
+
+,CXq�������~ymjstwtd^[SRcxzjap���x~��z}������������~pe`_^YTSVZ\\WPNG@?@?>?C?:9;<>ADJLFBDD@?<9:<>>=ADFHIHSeml_H831.5CNT[_cky�����������|hYTXh�nksv|zzte]XWSPW`gpxg3^^\p}afX_dfbYOIJJNY[XY[\[\`ihb[I4"
+		6Qd|��������}rrwtsnc_\SQ_u�zu����������������������~pe`^[WTTY^b`ZVKBBA==@JGA?BB@@CIJEDHGAD@<;;<AGFCEMQPRYfon]F502;>CO]egjhq~�����������}jYPKYwvkjnuy{vja\VOHKV`kvnE_^akp]^]adc^YRKHGKTWVUUZ]^bie^WE4#	&C^n���������~tx|vpjcc`WOYp����������}����������������|k`^^_\YX\bb\XJBDD=;AMMJGHHEBEJKIGHEAGEA>>AFKHEHOTVXYaion`J;9<:>JXahmfip~��������~uj^RIQm�tlnsywpkd^WMEDKVborY.
c_ehe\Ydcea[XVOIEISXZXW^edehb]YG6%
+.Nfx���������}s{�zrkefdZMRk�����ø���v|����������������ve``ba^YZab[YJBHH?<CKONLMNKIHLPOJECBLOMD?DHINIIMS[_[\aiqo\IC>:<ENYdkmibr���������ypmi]RLg��{wz�xppibYNFCEMYgtg;

d_cfe]X`b\`UVYQMDIOSVX]ekica`_YK6$
+	
%=Vn����������|q{~|smggkZJNc~�����ɮ��uv�����������������wk`_`a^[]ab]QJMNFABGMQPMLLJHMQRNE@DNTOEAABGJMMKQajg][cupj[GA<>EQT_kqsiju�������}spojWK\x���|{ttli`WOE?CN\lgI#

_[]__YTYZWYQTVPKCHLNOPXbjjebcbVA-!
+	,Jax�����������uz|xpkckjXKL]z�����̾��|t�����������������yhbdfd__``\YTRSSNGIMQQNLJIHIOVTIACIOLECA@CFJLKRcoqh]VbpylO?;FPW[fiq~wmlv��������wtqoaMIb����|xvtlpldWG>?GSdbK(
[WXY[YTWUVUORQOKDKOLFEMXdhebcaQ6" 		 2Pby�����������vuzyuiaonXKKXr������Ⱥ��vv������������������vliebabc`[]ZW[`ZOLMOQQNLLHBHUWMDADIIJMLHIIMPQS[elj`UYn�}fM<DOS^ppu�qnv��������{uqohYFRx���|uwrourm`L@>ALXZN2

[XUUYYUTVZVQQOQOJNOLIGMU[`_ZXVH2
+		!1Of������������yu||zg_kuZIJVj����������t�������������������xla^cihc]\YZbhcYOMNSTRPQH>AOUQKGBAAGPRNMOPUXVPS[cfe`hs��eG?DKYpuw����|���������vpnmiRKf|��~wupwvpofTF>=EMRS?#

`\UTYZWRV\VSONUVOMJILPTXZ_\WQLC4!
		+=Vt�������������}�{ve]`x\CFUe�������Ǹ��v�������������������ueaeklgc_Z\fmi`TPQWYUQQJ@@JTXWSKC<@JPOMNRVYWQPV`ksnfc|�}cMBEQeov����}y��������ystsmWFVo���}uq�vooi[L@:?FMVJ+
[XPPWYVOUZVSJJSUROIDFMV^diig^RB0!	 !5J\}���������������yqd[VvaBETg����������{|��������������������shdhlmjgb_fomdZVV\]WQNKECIRZ]\WJ?=DLPPRYZWXXWY`jtxpft���jLEJ\lw�����~~�������|txthYLQf}���zw�ytmh`RD<<AGRM2
TQILUWTLTUUSFHPRSRNKP\n{����{gJ-   
+)! 1FV���������������|tncYTniIGVh}�������ɷ���~��������������������vjgknnnjdeklga\[_a[TQPMHHMSX\[QG@AGMPW^\WXYXZ^enyyvwz���dULWjs~�����{�{yx��yxsf_YVbv����~|yiedZL@=<BJH6"
XVNRZ[WNPOSSEJQQUSQWi���������]8 ! 
+	!+#!0CS���������������xohcXUhoPKXe|��������ò��|z��������������������wmjipqniefhjf`]`b_ZXVSKFGJOW^YRIDFMOWZXXZURW]dhty��zx��~mTRbjn�����y{�ppz��yvjc_Zeo}����v|~hdh`QC<:>ED:+

ZTT[c]NJJJORHTTRW[Xf|���������tI* ! 
		!" +>Nf���������������ypf_UZ_lYT\fx������������|{|��������������������yrllmnnmkd[^dgda_\YZUPJFFLR]`YOKIKRVWVVUQQW^bimv���{v���e[ZWfz�����~z}xjp}|{tsmaYY^kv����{stkleZRG>:9@GA3 WQRW[TFBBDKPITQMS^k�����������\0 # 
+	#4JZt���������������xpaUOV_g\\]_t���������ǵ�}|��������������������{sllosqjaY^adc`a`[^[VOHFKQX[\YUVXYTSVWSPQPVY^bhv���|���~uhW^q������{~}oluxuomg\TRW`o����iqopcZRLB86=FE8%VLGGJIEEAEMRMUOGP`z������~{����f9&# 	&:N`���������������xpZMJRa_[`[Xq���������;��~|�|~������������������|tponligddcedbef`]]ZTMJMQPOTWY_a[UOTZVUVRPRUY]fy��������w]Vf�����z{|snrrojg`YQLPZi{����fkrtd]TRF94:FH>)
MFDGLOIECIOSOSMDTa~������|wz���pC!%"
+&:La����������������{nZIDKaY^h]To��������������{�|��������������������~ujdhophcddbhmg`_]WRPQRQNOQSZbdcYX]ZVWWSSSUX[hy���}����s_bz����wrpotumfdb__YONUct����mcopgbYTH>7:EH?*PMNS\]NACIOSQUTObk���������}��yO'#		
%:Og�����������������laOAD_Waj[Rl�����������~���|��������������������uhgnrhbaa_fmiihe_[XTPTVUQMMWglkhc]VTXYXTSVW\ew��������zow�����ynegrunb`^^b_SNPZk����j^gehc\TKD<<DF>,ca\[efP==CHOSYaaxz�������������~Z/"
+
+
'<Rn�����������������lh\GD[R_eVQi����������Ƴ�������~�����������������}��vnpvnhea]`dalnnlhcZQLRSOH>>JWirnie`^\[ZXWVUWdz���zv~���}~�����wiekrtle^XYZRNMQ]y���t_^[e`\SQJA=AD?0logallQ;6<AJS[hm����������������b3! 

-CXw�����������������smbOLYLY`YTk����������ɹ���������������������������}}��{{vqlgcb`\eikjhgb\Z]][TI@<@Thpsvtmeb`^\ZXT_hv���zrw���}~����rjintumdXRRMIFKVm����eZXa]YSVNC=>DD7#otj`moUB9=BMX`ou����������������f3!#!3Mb������������������}si\[\FOZ[[p����������ɹ���������~������������������|����xvrnlkjeaab`ZZahjkge_VQKA@EUenx}wsib``bb]WXh}��ofo{����|��vmkknpmkaVQJB?J[i|��~jZ[`[WTXQD<=EH;&
wzl`bn`I==HT\enw����������������i4!#!7Wl���������������qy��|ob\ZDGQ\_r����������˸���������}|�����������������������|unjjjiifgc[Y]bghghhb[VSIFP\^gtwuoigfeeggcgow��|shmx���}y��uljgknke[YRNDBOYgv��ve\[[\UUTQI>;BF;+
vpifeh[N><EUbgn{����������������m9 " 
$<\r���������������npy��va\WDJNW]n����������˺����������������������������������|rmideiikjffecaacjpnhb^SKJOU\fmpnjfefhhjklqy��njmx��}~~woe_djkg]VRRHBNXcr}th\VVXTRNLI?8>F=,
xmlkgeUNFAEVioq|����������������g6
+ 9[r���������������idk{�{b^ZJQLORd�����������������������~|�����������������������zqg^[]^beefhhhccgmmjhf[QHMWYZ_dgheegihgnqv{yx�urlhit}}~zrdY[chhaTVWKBMV_kx}uj[ROSPLEEIB7:E?+
ysqhinVMNKIUmwuz����������������d6
+":Zr��������������{b^bmx{mgbVTIHFW�����������ź����������~{�����������������������umic\YVVX\``^`[Y[^afhaZTUYXVX\_a``cfgfggjooostu~wg[`isz��yl]UX_dcV^^NBKV^gt{xk]SLNKHA@IG97D@-wwo`kwZORQOWmxw����������������b6 (?[r��������������xa\^cgtsid\N=75A{�������������������}����yuz���������������������~xuk]TTXYVVXXVPMNQV^c_\ZUPS[`bcba`adfdaaaciljgm�uc[X`kw~~wgVQV[_YddQ@HWagpwujaYPMHGB>GI<6AA2utj\itXPSUU[lty�����������������Z1 -D]s��������������}f\]]]_d[]dXE5-9{�������������������������xrty���������������������|m_QRWXSSZ^YTPNJKQUZVRLHOY]beggd`^^]]adcilgbew~}rf[X]cmstn`VSUZW`aP<BVcfknmhb]TNGFC<@D<4:=4!toiafkVRWZ[]irz�������ú��������R- !1I`v���������������m`^][QRMO\c_TAI�������ü�����������������}xwz�������������~zxy{}|vmfdXQQPSWV]\\WLEFJUTOJIIIKV_hjgb]ZZZ[\^cikeaimtuoe]][^ahnh^UTUTVYN<=Qacfffe_]VPGEB97<=5462$
smmhdgZW[_^Zesy����������������zJ(!2I`u���������������reca]WQL>>Rcg`]����¿���������������������}yy}}������������ywvsomptxzwj`YONTVTZcbUIGIIQPMMFAHS`lmkjhea^WTWWY`j^[\hqpfbfc]X_kg`USTRPUO?<N^adefd][VRHD@829@921.(pkiifba`_cgfjuu{���������������q@" 
+4J[u����~���������|yjedcZYR;+/=P[o����������������������������|xxx}����������}wyuqmknv����}ocTNUWUZbaWLFFKRXXRKGPZdfhoqrodYVVQMO_ebYV^ddfehe]YZ^\WTSRRQOCEPZbffgg_YXTNF?837A>3-.+%

mhec_^abckvvt~}���������������g6	8N_t��������������~}oedc\WR@+$&6Lj�����������������������������|xy~������~��ywz{zzxuw�������r[R[Z]bb_[VUW[][VPKMT\^dnrttmbZVTUXW_fhc_ad__bb]XXZ\XVVTSPMJLSX\^[[[URSROG?638??82/.*#pjc^]bikir��~������������������S)  
#?Sdw����������������shfcZTQA.$#/Ei����������������}}������������}yy�����yx}}yz����{uz�������xb`\Z\_cca`aa_^[UOOSWW_kqutsne[TTYUU[cfa]]_]]^[UTVYWWXWTQMHKQUX[YXYUSUWVND757<?=6--+'xoe`bltvrw��������������������s?  " *KVdx���������������}pklf`ZR>-()3Qx����������������zy}������������}wz����~xuxz{�������xvy��������obZ]bddcaba_`b_XUXXU[gmttvwsfVQVZRMPY^^[`]Z\[USTTSUWVSOKHKQTX[XWZVUW[[SH=868=@9-(&"yrjcepz}|{�������������û�����b- !

+
+MPXp���������������|pmnje]P:.++5Y�����������������|z|��������������}|~��yttux�������{xz��������pacfeee``]\_ed^Z^\UYbhqxy{{tf\[]WPKMT\a[][`aXUTRQTVURNJILSW[^[YZVUX\^XNE<44;A<1("
troihpz~}��������������������N#
%CIQl����������������zpifbZL;71+7[����Ĵ����������������������������}{����}zwwxwvtuy|ytqw|������zf_\[`efeb^`dda_c_WY]_hqvxz}znd^[WQKJS_^c_bc[YXWUVXVTPKJLSY_c`^\XWZ^`ZRLA627@?6+"
qqrpou}�~�������������������v; 
;LUn�����������������tfcd^L:;5*5\����ŷ�������������������������������|}}~~~}}}vtqomligjjnrv���~qcWTY^hjjhhiigehc\^_\aemqrx}ysjc[UOKOUahba_XYZXVVXWURMNMRW]a^]^YY]a`[USJ=34>@:/#
qqsux}������{����������������`+
+8PVh���~������������zvkhdbL477+2Z���º����|w|xqstx��������ý����������|vqlklmmmnlf_^`ga`dfksyw��vbZYW[billlmmhjf`fg`b`gjils|�zn_VSRRS[f`_]TUUTQRTTSQMKILRY]\]`\^cgfb\ZTF74>B;0$
qlqsw|������}����������������G	7OY_u��z�������������umdd`O376,1W��Ⱥ���xrkjwtsllnu~���������Ƕ�������������zrsb[LJLJGFRTVQNXbdiwvxypc\]Z]__afjkhdabdfhjeeceloqv{yrg^YXYYU_aYSQSQKLOMOOJGCFRX[U\`_agllgb_]L>;>>72%
rjjnssv���������������������q8	3JU^u��{������������~woccaP876.2X�ú��oXV``bqyswpq}���������̽���������������yaN<11,+0-19BINS^iilqtpgba^`_[[`gjfa[[_bb`dddfgfeggnvwrjb]WS[]WSQQNJJKIKNKEBDS^^Wa__cimomgbaUHB><73*!qhemtpp���������������������[+	
+-BO\u����������������wpifdSB:614Z���]F7@F@BPZd\npw����������˾���������������uV>786,%!9RgodPLP[ekjhhgdeeda^_cgihebab`]_^^^]^bc`chmpnha[X\ZTRQPPLLKJNSQIGFWfbXc^^bgkstlaaZSJ?972.&nihrvvw���������������������M%
+
+	
+'9K]u����������������xrohcUN@529_���W7*$,0/-,2=HT^ny����������Ƽ��������������nO:=H:/(&b����t[LO[b_`hjfdbcgfa_alqsniiihfa`]Y[abc`_djlhbccd^VSRPSRROMQTRMMKYkdW`][_cfrwnb_ZZTD:71.&qppur|���������������������m?!  	
+
+$3H_rs���������������}voh_SVI61@k��n?(%$+=L>&):db^bm{���������Ͽ�������������}jQAEK6,, ,j�����rP=@FMYdgea`afihfeilnmnqqojfec^_babaacghebcfgd^XSLNRUQKKLLKNMXjdX[YZ`cdmtmiaZ^^O@833*
wwuwn����������������}z���V0! 	
+!.D^mf���������������~wqj_UZT8-D|��zG/.2>LZG,7Y���wikw��������̿�����������~tdRGJI7(+-&Na����tVB64EYchlcedbfnqndbbirunfeb_^^bdda^]`cb]XW\_bc]RFEMSPJFGIJNPYgb[XSYejgjonqeZ^bWF97;3#
urnvt�����~������������~��xD% 	';Wid|���������������zvk_W[]<+F����\>9>CB@5,Ac�����khpy~������·����������{xk\PLJE:/4<3APs�tkmgXG?M^eimjooifinoh`]eoqkfhc[WZ^`a_[[`eaWOOUV\`[TJGLOPNKHHIKQZd`_WQXenonpqtiZ[b]N?>C:)
kf`t|�~}x~���������������e9 
+
+!3Pgi}����������������vh\VYb@+I����]FGS[[LBBU_�����oY[iv������������������xrcUMMGJKS[[^cltrpsuxqhchqutsvz}xmegmpe^afintoj^X[^]]^[Z]`^WSOUTWZXXUQOMOTQIEEDNZa^bXSWanutrstk\[cbVGEG=,

^]`p�~rsvw���������������yM//Lbi����������������|l\UPK6(U��¦�jgqyynjgnn�����hUTXl�������������������zob[Z[bo����������������������|njmoib`adjoure]Y[a__\YZ]`_]Y\\XSRUXSNELUQIFDDO^ec^XTU\dsoxxpia_dh`QKE=2$

^`_l{~olqx���������������d=$1M]b~�����������������rd[MA72c��Dz��������������zsmlox�����������������������~����������������������������uia`addabgnrlfaaecca]YWZ^ad`[WTRQPONJPVQJFDERbgc\UTVZamlw|tkcbhlg\PE=9,
]behtwmgo}��������������sL-" 

/GX_}�����������������yldN97:p���µ�����������������������������¾�������������������������������������������xighmlfbdiqsoiffb`a_XRRYaca^[WQLIMQTZ]XRKEFRbea[VZ[[cllu�vlcchljcVF@?3"
Xalckfb^k~��������������Y4! #"
!,?P^}�����������������td_L449r����ſ���ż�������������������ǽ��¾��������������������������������������������umlnkgegkopnhc_]^^ZTRW]dgihd^VPLOU^`\XRKIR]`^^]__\fmlnvlcaflke\MHE5&
XbpfbYRQ_w�������������}@# #!",7DZ�����������������lUPE54?z���������ø�������������������º��������������������������������������������������slnqokhcdimjb]\ZWTRQRTUY\]ac_XPNQ\a^^[VQV__]_a`^Zgokeutmfbfnoh`WRI7+#
\clg]XIJUr������������z],!
'-:[������������������nNC=9:N����������¹������������������ƿ����������������������������������������������������ywxwtnc_ciia]aYQNOOONNPQORX\\WRPYabeg^Y^eb]]^baZerobotqkeelmg`[WM=1&
dci`ZVKKYv�����������z^4
+
#6`������������������oN;4;A^������������������������������¼������������������������������������������������������~zxshb`dgb`f]TQRSQPQRPLHHKNQPMT^bgkb^bhd]\]ff]_qsemutpidgig`ZWRH9(
riqZ]UUSb}����������~aA
	

4c������������������pP:2AJm�����������ľ����������������ʿ��������������������������������������������������������}{skf_`fefd\UTUTQOQOOPMHFFHLKQ]bgld_bie^^afi]Xmsgkvusmfehjc[WYTA+
|wz^d[X]p�����������qD0!	0a������������������mT=6@Ku�����������¹����������������Ƚ���������������������������������������������������������xvnd`bca\VVX\[RMMNRTTPJJJIKN\_jlecfcaacdee]Vfnbgrvtie_dki_VZ]N5&
�}yjmb]bx�����������[3&+]������������������jO:3;K{�����������������������������ù�����������������������������������������������������������}tjddca[XZZ\^WNJGGJLKIIFJLQ\ajnhdd^\^cgge`[fk`dq{|mc[_fj^QQ[WA/"
~�{wtiacu����������iA#%Y������������������iL:25M��������������������������������������������������������������������������������������������zsomlje_\WX_[RI?89<?ABCJKOW_ipmfcYXY`fgcbaikcdq��ueZ]cj_NGQVF3%
z��zqi^_n���������yK-"X������������������iK<20K���������Ľ���������������������������������������������������������������������������������|zywvre\WX_^ZQA3028@BEKJISYhspgaVWU[bd_adkpjgp��yh_`cheVEDJB3&
�umfXYgt��������]2"Y������������������fJ=0)D���������Ŀ�������������������ý������¿�������������������������������������������������������|ztf``^]]`]N924;GKNQPGSVivoe_RVSZ`a]_dkrpim�uiaaaei`LACA9+
uyrqgUS]h~�����e;!X������������������cK>.&;|��������Ľ�������������������Ŀ��������������¿������������������������������������������������}{thcc_VU`cYA66<HNV\[KVVkung_MSU]b`]_dhprmkw{pjc`_`b^QFDD@3#
|qqrvn[RUg|���{lU:
 V������������������dQB/&7t�����������������������������Ŀ�����������������������������������������������������������������~yofb_VR\c`I;:<GN\jhPWYkrppgKPWbc^]`dekstpuwqogcb^YVSKFFE=, 
�xqqurdXShy��xgJ2


V������������������gVF2(5p�����������������������������¾������������������������������������������������������������������|tice_UZbfQCBCLT_soTU\kptypMPYdbZ[^ccfr{vwwuukfgbWTVOGFHE3$
��yz~{rkkknfZI-$U������������������f[D7(4l����������ľ����������������ɿ�������������������������������������������������������������������}ymfif[T]dZMBEOW\miUR]gmt~tRNVbaWWY[]epvqtwywmdje\TVULE?=5'

��~nmrpl]SI@;3$%T������������������j]F7(0e����������ƿ����������������Ž��������������������������������������������������������������������|nink\OT\^WFFOTSccQN]iknzuUNS_^VUWYZ`jqmoqtumbdf\RTVMD<:3(

���ohic]ND<44.

"N����������������i[F7'+[����������������������������»��������������������������������������������������������������������|onrpbRJJV[JGJNRbbRO^gfiuu[PR]]WUWZ[_ipjkmptoc`eZPQQJC=:4*!

��rgdZM>83130#





B���}������������}cWG7((S�����������½�������������������������������������������������������������������������������������~usroeXK?DQLHEIVfdWV_ceiqs`SQ[^YVX\\^gnhiklpmb\aYNLJDBA?8-#

x|}tkh]J>95753. 

+
+
+;}���������������v_VK7(%L�����������Ŀ��������������¾���������������������������������������������������������������������|uvspkaVH=CJJDIRdbY^a`gmoocVQY^ZWWZXY`fghjiif]X_XMIFAAA?7,#
zwz{tnaM>837641
	
+:|�����������{}���u`XO7*"D��������������������������¿����������������������������������������������������������������������yruvuqib^D;BHHLOaa\bb_hpplbXRU[XUTUUW^cggjgdaZV]VLIF@?=;3)" 
}vuvsk_QD7,+040'!		
+
;v��|�������|xy���ta[T:,#=�������������������������ƿ�����������������������������������������������������������������������~tvwsmfikL:;BJOTbhddddgloi^VRQTSONOQW`efegeba[WYUMIE>=<91& 
|tljkga]N<-)38/.#

	!:n�wu�������zwv���pa\X>1&:}������������������������ž��������Ŀ�������������������������������������������������������������uvvqjfmnM>:<KRYeolegifelfZSPLNMJHILU_ddaccba\XUTNIB:<=<2%
~ykdb`a^R@34:;5-"
		#:l�vu�������zvw�nfc]H;(:s����������ÿ��������������������������������������������������������������������������������������xwtomjonVB:=DO^lnjggifee\OIFGKOHFDFU_ac^_bbb^VRTOHA<<981)%#

��tjc]^\O@54684-#


+
#9f�}{������~xxz��~mlleUC19l�����������������������Ͼ�������������������������������������������������������������������������{ytnjhopWB;@EOfmkebbfefe\OE@BKTHBAHU\]^[]__`]VTXTNF?<9431/(!


��xlebb_PB83255-  
+
+$7]���������wtw{��zjlrlaH76d��������ľ�������������˽�������������������������������������������������������������������������|vnihqsW>:CKVmojea`bbcc\QF>CPTF<<HVXWXWXXY[ZUY^[UMD@=72-*'"

�~peefc]UNG<014/
+

+
+
+'4Pw��������rosx}�ughrriL94_��������ú�������������ɿ��������������������������������������������������������������������������|vpjgqtW<9EQ^mmmjea`^_a]RG?DPOC87GVWTUTRRTVVRZ`]VOEA>;0&!

~wkaabXNVZ[J4-/-!
+
+	
+,2Dj��������smpuz}sggrxpV=6Z�������¾��������������ļ��������������������������������������������������������������������������zvqkgprW=:FSadfkjd`^]]a[NC?BJIB85CTUQQPPQUVSOX_[TMB<:60)"


srogb\OGPZ]SD9/%				&23Af��������xnnsy|rfeoztcG;R����������������������ʽ�������¾�����������������������~�����������������������������������������zwqkitpT=9FT_Y\dc\\]\]`WG??CHE?95>MPNLMOTYWPMYa\UNA964/*$


goundZPPSXZ\`V<%
+	"!
+
+ 073Ad~�������~pmrxyl_\ixvkUDQ����������������������˺��������������������������������������������������������������������������}{xoilujN>=HRXOU`]WYZVZ[QB>AEJC:64:GLLHILTYTMLY`ZRL@755,# 



alwpcZUZ^`_hwoK-
+! 



(67.>`z����}���rmqvufXSdvvn`NX����������������������ø��������������������������������������������������������������������������zzwkfnm`IBCKNLJR`]VYWPVVL@=@EIC6359DKMEFIQVOJLT[SJF;432'


bhrp^RZXeoqprlS3
 +40+7Zu����w���wprtqe^X^p�xgVZz�����˿����������������������������������������������������������������������������������������tsrjdh^MBCKNJIKV]VUULINND98=AF=523:CFJF@EOOIIKTTMIC81--%
mlmm[MS[gnmmpm[A'

+!(+)(4Ur����x���vmoplef]\m��tfhv�����Ȼ�����������������������������������������������������}~���������������������������������nklg`]MCAGOOGDNTUMMMECFF<24=AB81/2;CEGE?BJIEGKQPJID8/*+'"�siYRV_ejikopfT9&
+"$#$)4Ss�������qfjjecha\j}��xok�����Ƕ����������������í������������������������������������}{����������������������������������kcdaWPA>DMRNC?HMLFJI?>CA706CE@71/3;BAAD>?CBAEJMKFFD9/(%))$��xe[VUY_ehmrsofM6#
+	!!(3Qw��������lafib_a[Yer���q\o����ų�����������������Ƹ������������������������������������|��������������������������������~lc]ULE??EKLG?>BHFDHF=>F@62=JI@82029>>?B>=??>BGJH@==8/& '-+"mh^R[WY_[cjpttuq^G2("&2Tz��������pbhmg_YWYcm���rY`����Ǹ��������������������������������������������������������~�������������������������������wpi[LC>?=ACA>;>FKFAA?<@D:14@JF=50/05:<?>:8:;;<ADC:465-"#+/(eaZIUTX^^flqrrttiUA51/.,# $3Y|��������yhlqmeZWZdq���wb\����Ƚ���������������������Ÿ����������������������������������|~����������������������������|qnom]LC>@=><977=FKHB>;=><2-4?EA84111378<9403876:>?6122+ &+'ghcPVTX`flpsrqttm[F83673-*'"*-*6\~��������xilpmi_Z`j{���~lcw���ż���������������������ĸ����������������������������������~�����������������������������teabbYQI@@=<8545:@FFD?;=:5..7=@=74379::8:70*.6647;;4120*!""UW[UVNO_nrtvututk[C0*./*/00/)!"!(7HD7=_}����}���qeikjjcepw���}yury���Ǽ���������������������÷����������������������������������������������������������������l]URRQQL@=9842249<ACB<:=60-29<=<735<@BA<<8/',6658:9202.)$
YRQUQKSbrwuxwwvvkYA/$!"!(),..,'# !&&)Cdz���|rutth`fighflx~~��xpnsv���ʾ���������������������ƺ���������������������������������������������������������������wiXNFEING96530/379==;88:95*.5;=<82,08GJA<95+%'/44871043.'#XONSROXgqzwvwwvrjYA-!#$&(**)'&!
+%)7Uit���zoljia`jmfffhnoow}tga_^u�������������������������ȿ���������������������������������������������������������������peUH@@CE?6431..268:9547973)/7;;80(&+6HPI?73+%%+0272./32+$[SQWYW^krztswuvphYA+ "$&()*&#"*3@\iu���uiefd\]kqkjia\Y^iqk_UPO`�������������������������Ƚ��������������������������{qov��������������������������������xgbRF>>=:755530/2576425862//5::62+$$-=R[SG;2*$#(.03-+-0.)#

[VTX\_gouxqsvoqmgY@( !$'&'**%!9@CUdz���sdadd][goopocXU\ba^YTSV_z��������������������ƿ����������������~zxvyutrnlic\YUQ]t�������������������������������j]`PE>;72379984356621266.+/7;:52/("$2F[e^PB4+&$',,,*()++)%

XXWX]fntv{xwqijgfX=# "#$(+)$!!
 &3ABHScy���te_^b`_elptqiaae`YXYX[`aq�������������������������|y~~�{ustqppjjnorsqpnmjc^W^i|���������������������������������_V^TG:420047875566620./+"&6DA7022) "/AYjiYE5.*'&((')'''))&

X[^_cillr~�}kjkfdW; !"(,,)&"$6A@HRey���sf`\_a`enutiadhf^Z\__`_\e�������������z|���������qfbeffihehki`aabfdfedceedcdj|����������������������������������vZV``N:120,,-.//022230(" '<JA1(.0' (4Mfo_G4,+(#%'('&%%')%

Z^dhlnliq{�}gomgcX;#*,+*(" 1A@EQdy��}nc^Y[\\bmup_V_b^[]bfhg`[_~�������������x��������na[X[XYXYbf[QXYV[\]\\_egipy������������������������������������rZT[`N9240(&')))*,-,-,# "(5?;.%('  '.>Uc^M7*&$!%((%%%#%(%

aciqx|||utzycpkfcY<!%*/.+& *<?FVes|�|l]WSTVW\emiZX]^[\^bhqofa`|���������������������{peZUVZVTWbv~umqfW[^^`cjsw{��������������������������������������s[ONQC524-%%***()**)#'!&*'(1860,$).4APXS@,! !&&$$%$""'&jjlosw}�vhilcrh][T<#!(-3.!
&4<>BShxzxqjcXOMPOTZ\]]abda^[^jx|oedp~������������������td^[U__oijw������r^cgmuz�����������������������������������������r`C=</+21*&(++($"$$" $'%#(3<:1)$"# (-;DSTD2$!!&'""&'"!$#
ptwvrpswshhjfndYUO:$%/1)"!*7BEERhzvjbkeZUTWW]cdghkiljbYYftxfZVZh~������������������ycY\l����������vvpkty�������������������������������������������pbG>8334/(&(*+)%#"!!""!*3640,)#$(7?KKB2'$%))""&%pv|}zxxyxqlkji`WSL9$%++)% #)3AJKTdtod]feZNILNWbhptrnsqaONZdkXGAD[|���������������ù��~mjx���������xorw|��}�������������������������������������������rgQICA>8/)'%!#%%&'&$%'&$#   $+254/*% !#38>?=2*((+)!#$ ! "%$#% 
jmquyyvqsrjfjb`]YP=& #'++'%+3@MTYbeiid_gaLGHGNZckookqnWDDP_k`OGJ\������������������»���vvz~}}urx�������w{������������������������������������������~tiXTROH</)&#!#$%'&%(*)$ "'&&)/44-'&"!"/25693/-,-*#"$$! $$%.554, ssrsx{wqnqkeh]__\VF+ "(/-.3:BKOU^gstjYb]MLLIP\flljhkbL@DPYab\[XRk������������������ļ���|�u~z��������}�������������������������������������z}������~{siZSQPMB3)&&      ###'*)*055-%"!!#,-03665311/*('&$$%# $'(.<DD</"

|}zututqmqoig\][VWO4%&%$&/17:=CC@IZp�|kXTGBAFIQantrlhcTC?EPTSPRZVDO������ĵ�����������������|�y~�����wqu�����������������������������������zvz������xqkZLFDDD<2,)" ! "$''(-45.$!&++-12466655441+%&))&&).3=JOK<-!

}��{rnljfijhd`_ZSUVB )//-).16669::FYjiZNB=;EKSamtqmeXH@@CIX\OINQHGd��������������Ƽ������xvnyopu}�{srx������������������������������������~wu{�����}|tmhYI>99?A<4.*%%*&!    $.51)"!$)+***,/379::<<8-"!)//26?GOUQG6'

r}�|qjgdhgigbc`XUV[M,,464,,-))&(/9GWdy�iZPL=AOUYalroh]NAADBDLeXJHPTJRy��������������Ļ����ymrpunjhpy||��������������������������������������zuu{���~|trmh`TH;66;><83/)"!$&'$ '0+ (671)&(-+)&#'+06<=>BGIE=:?A?AFPY]ZM>-$

kqy}tlkhhffccfb\XX]VD<FK>93+&$4IZau~rVKIJPVX[bkrodWIB@BB8=PVSLHLEBa��������Ŀ����������yz|{unox����������������������������������������{vw|}~{tpljc\WMD<7432465.+)*.0/.)$.)#%/54/-/21*# &).5=@@BEHIJNSQLKSX[[R@2&" 

oosxtnmihddbcheb_[ZZS,"N]^RG:3) 1Jads|sUGFQZ^^^clrmaQ@:?B=7>IMSSLF96Ip��������ü����ƻ��������~|~�����������������������������������������zuw{|zvokfc[SNE>731101451*&)3;>=5*)+%! "+0011472(" &(+19<;;:>ACGLNLJLMLKD8/+)(&"


yqmplhjhgcccejif`XTXZ6*]upcP?7-*E`hu�zVEBTabacgopeSE95=A:>KLHMSSM?55N|��������¾��������������������������������������������������������zxvtvyyvrmjc_VKC<9643411474+%'1;@AA:4-%"!  (--028;3'!#))).5852049=BHJIEE@=>><;;:51+


ypklgbfhdadehmkg`YUYY?'-a��q\MA5% $>_q}�zWJCO_^^dkqkWD99;CE=DIH?<ET]XK9:\��������������þ�������������������������������������������������xqqstvwurolha\RF>::;9974356/($(-27=BD?3)# '--+09<2%#(,+*,240-+05;CIJHGF@;>DIKLH;0*


nijokehicbefiolga[[[VI5)W��{k`Q@-%&" 7[v�qUPGHSTV]df]J5+/38>=CFG@88AMQJ;0:q�������������ſ�������������������������������������������������rjipvywsnjg_XTLB=<??=:51331+%$*+(*07@A7+#%+,)/9;1%$+/.,,01.+*/37=DIKNMGBGNNJKE4&!lhipmggeeegegomha^`^WR@#G|�}ujYG5--'(EcsznXOJFFLLNPNKC93;>::;>CDD@6-,044,%G��������������û������������������������������������������������rd^hswsolfbXROH@=>CB?93.//+'""()# !'284.($$"((*088.$%+//+)+,*()-0006AHKLIHNQJA>6'miinjdecgijefnmgabca[ZG"%@m}~}oaO@74- .F[mp`KJHBHD?><?EPTZWLA:8537<:/%$,03,1[���������������������������������������������������������������wdW]ikfef`ZRNLF@@BFF@<72-*&"!"%&%" %+.,)&$$%'-21)"$),+'#$&$#$),++09@BFIKLHB>8, ggjpkeffilmffnle^bb`\YG* '<`r}�rhXK?93' (7G\lfJKIDI@876=JVXND>857:205;8,%%*=E39g�������������ǿ������������������������������������������������hVV``WX\XQMIID@BEIIB@?8.'$'**'&'%"#'(% !#"'+)# #')("! "(,05:=?FMMD:9?;+
gfiljdejiknjomg`^ab^\\K1!!7Ogy~wkZMB<3**.5;L]gMFCGGC:57>CB;79==<;9841240)$#AV?)>h�������������Ž�����������������������������������������������m\RSSSSTRPKLLHDDHMKLIB:'"-1.%!"$&#!!#"!!! "('"!"'-5?EEEPWO>47>;+

gfhjgcejedkjmkd`_bc^\\N7&*9Pgvyo`TG>513548BOZMGCEE@9347:6329AFED?=72131,)&?SC&(Fn��������������ÿ���������������������������������������������rdVSPRVYXVVUSPLHFEJKKF9!&*)"!$'&" #$ #((!%-**.6>DFIRTJ>9=A;.


ffgigdfkc^hgihca`cc^\\Q>+ 4Mfxqg]QD;:;:65:BNPHCCA<4//12..1;HRSPID:1/120-)8MH+.T����������������ſ������������������������������������������|ri\XTVZ]]]\]^]XOHEMVXK7 ## "&*'"!#!$'&  $)2.,++.7@EIG=7;?@;3&


efhklkkme^hddgfeefd^]]TE0#8Tlib]VLDA>;846=HRIDD@8/)),/./3<KWYVPI=2/2553-6GD/ !:h����������������������������������������������������������{uplfeca_`bb]`cc\USU]gfS7"""##%'%! ! !%%%*+)(,021/29A@:45<?>:5*" 

efhkoppngdledkliije___XM7-CZ[USROIDA?945<BLCA?<4+%%+0224:FRUSOH=415882,4@:.%'Dp�������������������������������������������������������|xtqqqqtsmhggf`acb__elrpiV6#%%$$%""""!  "" (.00./.18::;?:425:;:72' 
ggfgjnnlgjoggnqkijf`aa[U>$#*8JQNKJID@??<6487:651/,&#'-48769@IMMGB:425892,070*+$)By�����������������������������������������������������ytsrrtvyzyurqpnkjjlptwx|vhR/%(&#$%!"$"!#"  %(0:B;.(.458:52367896- 
jjgdeiiffnnghlqjgjgaaa\ZB'!$(3@MPIC>9788:622,*+*''&%%(07<<::=AEFA=71/28;=514.&'##G�����������ž�������������������������������������{tqtuvxy|zz{{{|}�~}����}|wgL+!$'#!##%"#)*'# !+9A=3,+))1:76899<?9*nnkfdffcfokeghohgkicb`]]E(!#&/9LRI>5003232..'#)((((''*17>=<::;>?A=6.+/8?@:375' 'T�����������º������������������������������������|zxqpvyz|}}z|�������������|sbL,"$'! #$!&02/*#'011-**&)6<<<;99=A;*jjjgcabbfjkmjcikhiklbZ^aO1!%+8GQG8131..001/*((((%%&&)07=?;7667<@A=4,.5;A91;3#-p���������ſ�����������������������������������|xuvusu{}{|�~}��������������s\F,"$&"$% "(054-%!#!"#"&*"!%#%1<E?;978<>7)

fhhdcedabdirmdeligipj^^`S7$)4@NH;22313..00-+*(''())*.4;><74237>CC9/.5;;607+ M��������������������������������������������}xsquwwz~~{|�����������������tYA*!"$"!"""&.21.)$ "%*+("&+""1AG@:66:><6%
eihdfjhcbbgogcajnjfkme^[P8
!'-3CGC96557630.+++*(+/20-.29==94/04:=<8579940+/#'+w��������������������������������������������~xrrxz|~||�������������������w[=&!" #*153/)#  "*04/%"& !$'3CG>514=@<5&dihdfihfgehkbe`dsqfbhjcXN9"#%2AIC=855<72-))++).5984236;=:4/0473/19@=70,$$"'!V��������������������������������������������~xssx{||ywz�������������������za=#!  *4885.'#$.86' *226AF:0/5?A>4)

`edabbbflghkdjccsuiaejh_WA##6DFE@955442.----18>?;6447;92./33/+-5<;5,*!$%" 6p�����Ǿ������������������������������������{vuux}}{xy������������������}xc@$
 ")4:=80*%"$*.+  %/669BE8./5<>=2(

_ccbdaaipfchdjjikomkighg]I*)5>GJFB78:9645558=BC@:522650,-//-**/53.'' !!!J����������������������������������������|{uruwy{z{}��������������������s_D+
+


'08840/.*)(&#&()*,/8D?4,-0244."


bfeeieckpgcecjrofipvqfdd\K.!%.=GJNFEEB>=????ACC?83.-02.,--..,,/1-'!# 1j�������������������������������������~wtpqppuwzzz}����������������������o[H8#
+
+
+


$*,+,.00*% )'&" %1:4.*+++-.*


chfglidhmljgfnxoehq{viaZZL2#  -8CNQOMJEDEECCBA?:3.+(,0/./.145420+% !$Av���������������������������������}wsqnolptw{}������������������������n[KB2
+
+
+
!%'("$!!)*,+++)+./(!


fjiimkedhnrmhlqrkfisuofa]S9!'%$,6DHJHDA@@?><;82-+)&'.331269:72-)&"!""#+Cv����������������������������|xwtolmquww{������������������������������p^LH>
+
+					 $*(,,*'%(&"
higjqofddkqniilljcagnpjb^V>#   &.37;>>=;8688510-+*'&,343102430,'#!    58,9`����������~�����~�������ypljlnoqw���������������������������������}thVPD#
+				



 %(,*'#!$"


khdisriebipqligefc_^bijeb[F) "&,28:61/5840011/+(,241-*++,,*&"! 
""4KB&1\��������}|}~~}}}}��viddbbelqv{�����������������������������������~yk^H#




+		 #&%#!!$"ojdgquoidhotrmgabgg_X\dfe`M,")263/*/452/1430--252,('%%&&$"!!  /PT90Piklpwvrqnljkmnmha[TSVbpwz~~�������������������������������������������hC
+
+		
+
!$'#rngelutoihlrtrjcchkcWV\aecQ0
!(.00+(,571044421464.(&$#" ?cR/
(LeYLOTQMNOQNOPSX_fjuy}�������������������������������������������������d8
+		
+


"'("
rpkdgqurlhhmrsoihfeb]YXXacS3
 (/30''394026764452-'&$#!#OiN&%S}~h]XRLJKPXamu{������������������������������������������������������zU,
+
	
+
+



+$))"
pqmgelqqnjghlrtqkeacca]X^bS4
#-31+(-0//1887310-($#$%$ 
!;]bB
I����}ztmggjs��������������������������������������������������������rN(#
+
		
+
+
+
+
+	
+!  !&**#ppojfinoplhfiqxxmgdfgfd`]bR4
+(0230+(*055531/-*&"!"&'$ 
*LdT03t�������|yv{���������������������������������������������������������sR*%)#
+
+				
+	
+
 $##%(' ononpfdflqmeeiovqmgcdiha^_P4"(4@?1(+36310,)))&" !$%&$ !3TiO&Y��������ysu{��������������������������������������������������������tY5!"&


+	
+
+
+		
+  #%&&%&'#

klnprkedmssmfdipsmga`cd`bbR5

 %3CG<1/13/,($"#&&!"%&%#(:V_?
?~�������|{{��������������������������������������������������������v]B(#
+
+


+	
+
 "$%&''&&%!

hjlrurjfknttlefjokgd`_``deU7
+
++:C?:8763/)%####$! &*("+;TS0
+&d��������|}��������������������������������������������������������s\L3#	
+
+	
+	
!#$%&%'(&$"

eghnqtmjdcluslhfffkojcabb`O2

$08AFGFC>732/)%"#"(+&"+4FaX, N|��������|~�������������������������������������������������������nXNA&	


	
 "###&'$!

ijhmmspna]dqupjdadoxthbb^VA'
&2?GIJIG@<??:80/*" !!  &.6Ki],  #$9l�����������������������������������������������������������������|o[PN)
+
+


 "#$&%"pnklkopnebenroieccmvqd^^YN6%.568<EID<;=?B>83.& &*0FcS$.0.')(&R�����������������������������������������������������������������{rbTN*
+
+
+
+	
+		"$&&&#
rkhffhmlmjginokjgcfkeZVXSD,
 ')%'.CQRA4./632695-'# $-G\E%:>5-.+C�����������������������������������������������������������������xmbV?&
+	
+
+

			
 "#$&'%#

ukgbdgoopmedkpnnkcac]TRUO:"!"=V_M6('/0221254.(!! (@P6-?@/)/+G�����������������������������������������������������������������scZT. 			
+
	

		
+ $%$$$%"

rrmfgqvromggjnpiga\__VTVN4
2Q^YE0042.*'*0335(# /<+,;2&"&$\�����������������������������������������������������������������o]VK&

+
+			
+
+		
+
!  



stpiglrtppkjhggd`\Z_c_[UE. 
'DV]UFA;2-(&&)-13-,' *)%)2<5# !!1u����������������������������������������������������������������yf^T= 

+
+
+	




mnlhgehovunic[Z\^]\bfeZG0&#'%2AOTPJ?6/(&%$%(-.3/($!*269@?,(%%A�����������������������������������������������������������������nb]M0	
+
+	






jmlmnhejqngfd]\````fggU8&(03'#*5@HLD9/%!""!!%-552/*% )17<:AD4..*L����������������������������������������������������������������wh`XA$
+	





mrrrskfiid_fkfef^__feeQ5-284% '0883+ !"#*05982*%(;6.26;:0)1+N���������������������������������������������������������������se[N4	
+
+	
+		
+




ptusoheila]dhfhh]^\c^ZH418<3$
 &%! %'%+5:6/#3F7&)231-&2*Iq��������������������������������������������������������������ytjV>' 	









qpppmhhki_^eccjk]_]e[P>4298."

 %$"/3/'!;C3(-,+,.*50Jb|�������������������������������������������������������������xvlM0" 




 #%%$ 

ngfjkhikd^dld`hhZ_`k_O<9;:0$
!##$0.#0A8'%+'"&,-79S]kz������������������������������������������������������������{weA+$!!



!(,.--&




rkfghikkedkmffhcX`dh^G>@>9*!(30&&@A+ "&-05HXadq�������������������������������������������������������������yR2*'#$ ##"

%),/1



ytojfhllihmmffhdY^cfSB@C<4'
"1=/
.A9$#*--6N^ejk�������������������������������������������������������������iE-+'!&%
 ### 
 #+-/


uuskeglmljlibcd_Y_ijK?BA7+"

&6=( 29+",,*;Wbfod���ĵ�������������������������������������������������������xW<24,#'%		
+
+!


!%).34
loojehkkjhhe`ab\ZangA<B;/"


!+32#,60  (()A]b`nbw�����������������������������������������������������������iB345)$&"


! 
%+1;=
jkljhikigffeeggb^biW38A7)

&30&-CD6$ &%)C_b]kcl���ľ�����������������������������������������������������Z;46.$)(
+
+
! 
$-58

kjihgjljhfeghjhdacdK4;@3$
 -7-=[WG3%!').D^eej`^|����Ļ���������������������������������������������������qH<97*+/,!%'),lljfeknkjfcehhfdbc`G?@<-
+
+&02'!9Te_Q?+!)3<K_gkj]Vk�����Ÿ��������������������������������������������������[<94/$ #.2.


!%)+*,-npngelpkidadhhgfda[CDA5'



#.-(!7See]QA-*;IS`gjj]Vb�����¼������������������������������������������������kF;@4)  (/3/
!&''''()+.//4:9


sqoiiomlgc`fkjgea\M=E=1



)11(#%2K_d_UM@+(<GVfcdebSYu������������������������������������������������������wVIGI9$!(0/(
"$$$$&,155560-+)''& 

spmhjqpngdbfihgebUD:C8,

'040(&2G]e`ZRI;(
-<Qfigf_MWy�}������Ļ�������������������������������������������~ePSXP;%$./*!#'+/01/,**)'&)#
+

smjgjqqofffhhhgfcRA<D6)
+

#/570,2DZhf\RLC4%	 1G^jlkdQUv��|��������������������������������������������������ub]_WF3#(30'


#%()&" 

+

niggjpqnfgjihhghdTB=B5'
+
+ )38848CTdiaVJD<."
.?UjnidUSj���r{����ǹ�����������������������������������������vjchaK5)#" +53'

 
+	


	jfgkknqldejhfgehgUC><.#

+
+%)/598=FTaifYND=6+ 	&/Caof_VSc}��sn����������������������������������������������hSKHPRI:1,($%.42'
!$%	

kikpmlnhdcgfegfjjXHE=)
+	
+
+%+/258:CSbhf]PIB:2)
+
+!%7Xjg_VVdz���ww����ɾ��������������������������������������iB*$"&4BJB8.)+01-#


 !!!!!+-&

khkqnkmfjdgegkkro^MH:&#',026=CN_id[QIF@7/%
+	
*Jag`VYh{����tm���¾�������������������������������������kI'
)?D@6111/*!


 !!&-(
+
fcfpnmqiqhigjpqyscN@/ 




			





%(('),/39DO]gf[SLEB=5,!
+&D_g_W`q�����xou����������������������������������������oN0		%9C>972/*!
!!$,(
+

cbjpqrokngefnt���jK0"




!)001-,,/5@P^ef`TKGB<50(
+	&&9[i_V`v�����xsw��������������������������������������zS6*
+
+	,=C<>52+!"*&


eelqrojidhedn{����oK-



		
+

!.453-++,.3<K\hibXPKF>84.'
+	
*)5Vd\U_r�����|tu{�������������������������������������dA/%		
	
1;?A81)
 ("



jjnpojeihpdk}����}X7"




+
+
+		 )-43/,()*.16CWgodZQMH@963/)	$0-0R_ZXcv������~tqt������������������������������������pO3-&
+		
+%4CF<2+!


 &
nllkjffopr^x�����`6








(,1.,&$%%&*06<L_kmcYQLE;543-&
(2.-P[XZfx������}upq�����������������������������������bA.,
+	
+0AJ@71%

!'	
lkljifgrqoVu���gE$
	

+
+
)-.,(&'%$%""'1<GWgkg_WMGA;753+!
+		&)1--NWTYev�����{vrrx����������������������������������y\;*"	0=MB:4'

$'llpokfdknhNaz�oH6&




+$%)/-)&  #"#%%&.;JVbjje\YQG?;;;71(!(.+/.0LUU\gx�����{xtqru~���������������������������������b=(
(
!5?PA60$$#	nnsrkfdficSVv�h,
+



	
+		
(/3//'%&" " &,2=FNRTWYZWZYOD@BEB?5&#(-0..12IUYbl{����{uqpqtz�����������������������������������c=(Hj_8
+%8DR@1) 
#
nmrqjhhilhf\}�m*

+





$-59:5/ !"-39@B@9229@DJNMKKNPOG;.')*&)).-0-32FT[dmz~{ztmkms{�����������������������������������Y8)"3���e4#5GR@.%
!

hkkjiknkgmk_|�S%





	
+!$*041,'#")8@7/%%.9@FLPRSB:1,.0*$')*/'41EY^fnuzzzvqllpx}�����������������������������������}Q4&'Q����L2IK:+!
+


feiljhjmp`[i�j-

	




	
+
+	"$$'(*,+$")7@;(
$*07=9551-+,,"$%).)43F^fmsxz{zxwutwz}����������������������������������tN4',]����])$8KI5&




njmmjkmuykt��g#!&	
+






 %$#"$%&&''''$(;D<)	
+



+
&43,((&%!(+..-2/Khry~��|}}}}}~����������������������������������lL5'-\����m3,?OH1"

 



rnrnpvniim|�u> 




"$%$" !#$#(%#!4IG8
+	-?6(%&#"(1/.,355Rpz������������������������������������������������gG5*,Q����r7!0CQH0!
!

rllhnvi]govhA"

 ""!"  ""#
$@OH>
+8KB+%%%)-1.//77>[u|������������������������������������������������b@4/.:u���l-/CQG/		  

qlmjjmhed]VB&
	
+
 ! 

	-CRPF 
3RW2$(++*)+.+-0Dgy~�����������������������������������������������|_@62-0K���W'.EPC+
!


mnsuqlig\E0
+





!5FU`O'
+

"Lf=((,**(-136:Sq}�����������������������������������������������mYD:1,.-BPK)4LN=(			
!
+

qnmqspjbK5


%>OWlZ5
+


<fH0$'-*-=KZ_Xbu}~�~�������������������������������������������z]NE</-+$#$$!<RM9&
+




!
+

ookjpldW<	

(ER\eK4
+
+%TRC1*)3F]t~safu|}���������������������������������������������{bRK@72/*$
+*?XM6"
+
" 

+


ppljpkbU4  

+HQQN2$
+	
+
+;GOD5*5Tr~q[Pa{�������������������������������������������������kZRG?:4/)
+
)>WK4
+
+" 
+


�g�g�i�k�m�o�o�p�r�t�v�w�x�w�v�t�s�s�u�x�{�}�~�~�~�|�y�x�w�u�p�n�i�h�i�k�o�s�v�w�w�w�x�z�z�u�m�e�_�\�\�`�f�m�r�t�v�w�w�w�w�w�w�w�w�x�w�w�w�v�t�s�u�v�w�w�w�w�x�z�y�{�}�}�|�z�y�x�x�x�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�v�v�w�w�y�{�}�~�}�|�z�y�x�x�x�w�w�w�w�w�g�g�h�j�k�m�n�o�q�s�u�w�w�w�v�u�s�s�u�w�z�}�~�~�~�|�y�x�w�u�p�n�i�g�h�j�o�r�u�w�w�w�x�z�z�u�m�f�`�]�^�b�h�o�t�w�w�w�w�w�w�w�x�x�x�w�w�v�u�t�s�s�s�t�u�u�u�u�v�x�z�{�}�|�z�x�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�w�v�v�w�w�y�{�}�~�}�{�z�y�x�x�x�w�w�w�w�w�i�h�g�h�h�j�k�m�o�q�s�u�w�w�w�v�u�t�t�w�y�|�~�~�~�{�y�x�w�t�q�n�i�g�h�j�o�r�u�w�x�x�x�z�z�v�n�g�_�]�_�b�i�o�t�v�w�w�w�w�w�w�w�w�w�w�v�t�s�t�s�s�r�r�s�s�s�t�v�w�z�{�|�{�y�w�w�w�w�w�w�w�w�x�x�x�x�x�x�w�w�w�w�w�w�v�v�w�x�z�{�}�}�|�{�z�y�x�x�x�w�w�w�w�w�k�j�h�g�g�g�i�j�m�o�p�s�u�v�v�v�u�t�t�v�x�|�~��~�{�y�x�w�t�p�n�i�h�h�k�p�s�v�w�x�x�x�z�y�v�o�g�_�]�^�b�i�o�r�t�w�w�v�u�u�v�v�v�w�v�t�s�r�r�r�s�s�s�t�t�s�t�u�w�x�z�{�z�y�w�w�x�w�w�w�w�w�w�w�w�x�x�x�w�w�w�w�w�w�v�v�w�x�z�{�}�}�|�{�z�y�x�x�x�w�w�w�w�w�n�m�j�h�g�g�g�h�k�l�n�o�q�s�t�u�u�t�s�v�x�|�~��~�{�y�x�w�u�p�n�j�i�j�l�p�u�w�x�y�x�x�y�y�u�o�g�`�_�`�e�j�o�r�t�w�w�v�u�t�t�t�u�u�t�r�q�q�q�r�s�q�r�r�r�q�q�s�t�v�x�y�y�x�x�y�y�x�x�x�x�w�w�w�w�w�w�w�w�w�w�w�w�v�v�w�w�x�z�|�|�}�|�{�z�y�x�x�x�w�w�w�w�w�q�o�n�k�j�h�g�g�h�i�j�l�n�p�q�r�s�s�s�v�x�|�~�~�~�|�z�x�w�u�q�n�k�i�j�m�q�u�x�y�y�x�x�x�x�u�o�h�c�b�c�g�m�p�s�u�v�v�t�s�q�q�q�q�p�o�o�o�o�o�p�q�o�p�o�n�n�m�n�o�s�t�u�v�v�v�w�x�x�x�x�x�w�w�v�v�w�w�w�w�w�w�w�w�v�v�w�w�y�z�|�|�|�|�{�y�x�x�x�x�w�w�w�w�w�t�s�q�o�m�j�i�h�g�g�h�h�j�m�n�p�q�q�s�v�y�}�~�~�~�|�z�x�w�u�q�n�j�i�i�l�p�u�w�x�y�x�w�w�w�t�n�h�c�c�d�g�l�o�r�q�r�p�m�l�l�j�j�k�j�j�k�l�m�n�o�o�o�o�o�n�l�k�l�m�o�p�p�p�q�q�s�u�v�w�w�w�w�w�w�w�w�w�w�v�v�v�v�v�v�v�w�w�y�z�|�|�|�{�z�y�x�x�x�x�w�w�w�w�w�u�u�s�r�o�m�j�i�g�g�g�g�h�j�l�m�o�p�r�v�z�}�~�~�~�|�z�x�w�v�q�n�i�h�i�l�p�t�w�x�y�w�w�w�w�s�n�h�b�a�b�e�i�m�m�k�i�h�g�e�d�c�c�d�e�g�h�k�l�m�m�n�o�p�o�o�n�m�m�m�m�n�n�m�l�m�n�p�r�t�v�w�x�y�x�x�w�w�v�v�v�v�v�u�u�v�w�x�y�{�|�|�|�{�z�y�x�x�x�x�w�w�w�w�w�v�v�v�u�s�p�n�l�h�g�g�g�g�g�i�k�m�o�r�v�x�{�~���}�z�y�w�v�r�o�j�i�i�l�p�t�v�w�w�v�v�v�t�q�l�h�b�a�`�`�b�c�c�c�b�c�d�d�c�a�a�a�b�e�i�l�n�o�o�p�o�o�p�q�p�o�o�n�o�m�l�j�i�i�j�j�m�o�r�v�z�|�{�y�y�x�w�w�w�v�u�t�t�u�w�x�z�{�|�|�|�{�z�y�x�x�x�x�w�w�w�w�w�w�w�w�w�t�q�o�l�i�i�i�g�g�g�h�i�k�n�q�u�x�{�~��~�}�z�x�w�v�r�o�j�h�i�l�o�r�u�v�u�u�t�t�q�o�k�g�_�]�]�]�]�^�^�^�a�b�c�d�c�b�b�b�d�h�k�n�o�q�r�r�r�q�p�p�o�o�o�o�o�n�n�m�k�j�i�h�j�l�o�r�w�z�z�y�y�x�w�w�w�v�u�s�t�u�w�x�z�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�x�x�x�w�u�r�p�n�k�l�l�j�h�g�g�h�j�m�p�t�x�|�~��~�|�z�x�w�v�r�o�k�i�i�l�o�q�s�t�s�r�q�p�n�l�h�e�_�]�\�\�\�\�\�[�^�a�b�d�d�c�c�d�h�j�m�o�q�s�t�v�u�t�q�p�o�o�o�o�n�o�o�o�n�l�j�i�h�i�j�m�r�u�w�x�y�y�y�x�w�v�u�s�t�u�w�x�z�|�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�x�x�x�w�u�s�p�o�n�n�o�m�k�h�g�g�j�l�o�s�x�}����~�|�z�x�w�v�r�p�l�j�j�l�n�q�q�q�p�n�m�k�k�j�g�e�c�b�a�`�`�^�\�Z�[�]�_�a�a�b�d�f�j�l�n�o�p�r�t�v�u�s�r�p�o�n�n�n�m�m�n�n�m�l�l�k�i�h�h�i�m�p�s�v�w�x�y�y�w�v�t�s�t�u�w�x�z�|�}�}�|�{�z�y�x�x�x�x�w�w�w�w�w�w�w�w�v�u�s�q�p�p�p�p�o�l�i�g�g�j�k�n�s�x�}�����~�|�z�x�w�v�r�p�m�l�k�k�m�n�o�n�k�j�h�g�g�g�g�g�f�f�f�e�d�a�^�\�X�Z�\�^�`�b�d�f�j�l�m�n�o�o�q�r�q�q�q�q�o�n�l�l�k�j�i�i�i�j�m�m�k�j�h�h�i�k�o�r�t�w�y�y�w�u�t�s�t�u�w�y�{�|�}�}�|�{�z�y�x�x�x�x�w�w�w�w�w�u�u�t�t�s�r�q�q�q�r�q�o�m�j�h�h�j�k�m�r�y�~�����~�{�z�x�w�v�s�p�n�l�j�i�j�j�k�j�g�e�d�c�d�g�g�h�i�h�h�g�f�c�_�\�Y�Z�\�^�`�b�d�g�j�l�m�m�m�n�o�o�n�n�o�o�o�n�l�k�j�h�g�f�g�g�j�k�k�j�i�h�g�h�j�l�q�t�w�x�w�u�s�s�t�u�w�y�{�|�}�~�|�{�z�y�x�x�x�x�w�w�w�w�w�q�q�q�r�q�q�q�q�s�s�r�p�m�j�i�i�k�k�m�r�y�~�����~�{�z�x�x�w�s�p�m�k�h�g�g�g�f�f�c�b�a�a�c�g�i�k�l�k�k�j�h�e�`�]�\�\�^�`�a�d�g�i�k�m�n�o�n�m�m�m�l�l�l�l�l�l�k�k�j�h�g�f�e�f�f�g�h�i�i�h�g�g�g�h�n�s�v�w�w�u�s�s�t�u�w�y�{�|�~�~�}�{�z�y�x�x�x�x�w�w�w�w�w�o�o�o�p�p�p�q�q�s�s�r�p�m�k�k�k�k�k�m�r�y������~�{�z�x�x�w�t�q�m�i�g�e�d�c�c�b�`�_�_�`�c�g�j�m�o�o�n�m�k�g�c�`�^�_�a�b�d�f�i�k�l�n�o�o�o�m�l�l�k�j�h�h�i�j�k�l�j�i�h�g�f�e�c�c�f�g�h�g�g�e�e�f�j�o�t�w�w�u�s�t�t�u�w�y�{�}�~�~�}�{�z�y�x�x�x�x�w�w�w�w�w�n�n�n�n�n�o�p�r�r�r�q�p�o�n�m�l�l�l�n�s�y�~�����~�{�z�x�x�w�t�q�l�i�f�b�a�`�_�_�]�^�_�b�f�i�m�o�r�r�p�o�m�j�g�e�b�b�c�e�g�i�k�m�o�o�o�o�o�n�m�l�j�h�g�g�g�h�i�j�h�h�h�g�f�e�c�b�d�e�f�f�d�c�b�b�g�j�n�s�v�w�u�r�t�v�w�y�{�{�}�~�}�{�z�y�x�x�x�x�w�w�w�w�w�l�l�l�l�l�m�o�p�q�q�q�p�o�n�m�l�l�l�o�t�y�~�����~�{�z�x�x�w�t�q�k�g�c�`�^�\�\�\�\�]�`�c�g�j�n�p�s�s�q�p�o�m�i�g�e�e�f�g�j�l�n�p�q�q�q�p�o�o�n�m�k�i�h�g�h�h�h�h�h�h�g�g�e�d�b�b�c�d�e�e�d�c�a�a�c�e�h�m�r�t�s�r�t�v�x�z�{�{�}�~�|�{�z�y�x�x�x�x�w�w�w�w�w�k�k�k�j�k�k�m�o�o�p�p�q�p�o�m�l�l�l�o�t�y�~�����~�|�z�x�w�w�s�p�j�f�a�]�[�Y�X�W�[�]�a�e�h�l�o�p�r�r�r�s�r�o�m�i�g�h�h�j�l�o�q�s�t�t�s�r�q�o�o�n�m�k�i�h�h�h�h�g�g�g�g�e�c�b�a�a�a�b�c�d�d�b�a�`�_�`�b�g�l�q�q�r�s�u�x�z�{�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�k�k�k�j�i�j�l�n�o�o�p�q�p�o�n�m�l�l�o�t�y�~�����~�|�y�x�w�v�s�p�j�e�_�[�Y�W�U�U�Z�]�b�g�i�l�o�o�p�q�s�t�u�s�o�l�i�i�j�k�n�p�s�u�u�u�t�s�q�p�o�n�l�k�i�i�i�i�h�g�g�f�e�c�a�`�_�_�_�`�a�c�c�a�`�^�\�\�]�`�e�k�o�r�r�u�x�z�{�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�k�k�k�j�j�j�l�n�n�o�p�q�p�o�o�n�m�n�o�t�y�~�����~�|�y�w�w�u�r�o�j�e�^�[�X�W�U�U�Z�^�d�i�k�l�m�o�o�o�r�t�u�t�q�n�k�j�j�k�n�o�r�t�u�t�t�r�q�p�o�m�k�i�h�i�i�i�h�g�f�e�b�a�_�]�]�]�]�]�_�a�a�`�_�]�Z�Z�[�]�`�f�l�p�q�t�w�y�z�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�k�k�k�k�j�k�m�o�o�o�p�p�p�o�o�o�o�o�p�t�z�~�����~�{�x�w�v�t�q�n�i�c�]�X�W�X�V�V�Z�_�h�k�l�k�k�l�n�n�p�r�t�t�q�o�m�l�l�l�n�p�r�s�t�t�s�r�q�p�o�m�j�i�i�j�j�j�i�g�e�c�a�^�\�[�[�[�[�\�\�]�]�]�]�]�[�Z�Y�Z�]�a�g�m�p�s�v�y�z�{�|�|�|�{�z�y�x�x�x�x�w�w�w�w�w�j�j�j�j�j�k�m�o�p�p�o�o�o�o�o�o�p�p�q�u�z�~���~�{�x�w�v�s�p�m�f�`�Z�V�V�W�W�V�Z�`�h�n�m�j�j�j�k�l�n�o�r�r�r�p�o�o�n�n�n�o�q�s�t�t�t�s�q�p�o�n�l�k�k�l�l�k�i�g�d�a�_�\�Z�Y�Y�Y�Y�Z�Z�Z�[�\�\�]�\�[�Y�Y�Z�]�b�h�n�s�u�x�z�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�i�i�i�i�i�j�l�n�q�q�o�o�o�o�p�p�q�q�r�v�z�~���~�{�x�v�v�s�p�l�d�]�V�T�U�V�V�V�[�a�i�m�l�i�h�h�i�j�l�n�o�p�p�q�r�q�o�o�o�p�r�t�u�t�t�s�r�p�o�o�o�n�n�n�m�k�h�g�c�a�^�\�Z�X�W�W�Y�X�X�X�Z�\�\�\�\�[�Z�X�W�[�_�d�l�q�u�x�z�{�|�}�|�{�z�y�x�x�x�x�w�w�w�w�w�g�g�g�g�g�i�k�n�o�p�p�p�o�o�o�o�p�r�t�w�z�}�~���{�w�v�w�t�o�i�`�Z�U�T�U�W�X�X�]�c�i�l�k�g�g�e�f�g�h�h�g�h�l�o�r�r�p�o�o�q�s�t�s�r�r�r�r�q�q�p�o�o�p�p�o�m�i�f�c�a�^�\�Z�X�W�W�X�X�X�Y�[�\�[�[�Z�Z�Z�X�W�Y�\�`�j�p�t�x�z�{�|�|�|�|�z�y�x�x�x�x�w�w�w�w�x�h�h�h�h�h�i�l�n�o�p�p�p�o�o�o�o�p�r�t�w�z�}�~��~�{�w�w�w�s�m�f�^�Y�U�T�U�X�Y�Z�]�b�h�j�h�d�c�b�b�c�d�c�b�c�g�l�p�p�p�o�o�p�q�r�r�r�r�r�q�q�q�p�p�p�p�q�q�o�l�j�f�c�`�]�Z�Y�X�Y�Y�Y�Y�[�\�\�[�Z�Y�Z�Y�X�W�W�Z�]�f�m�t�x�z�z�{�}�}�|�{�z�y�y�x�x�x�w�w�w�x�h�i�i�j�j�k�n�o�o�p�p�o�o�n�n�o�o�q�s�w�y�}�~��~�{�x�w�v�q�k�b�[�X�U�T�W�Z�[�\�]�a�f�g�d�`�]�\�^�_�_�]�\�^�b�g�l�l�n�o�o�o�p�p�p�p�p�p�p�p�p�p�q�p�p�q�r�r�q�o�j�g�c�_�[�Z�Z�Z�Z�[�[�\�]�]�\�Z�X�Y�Y�X�W�V�X�Z�b�k�t�x�y�y�{�}�}�|�|�{�z�y�y�y�x�w�w�w�w�j�j�k�k�k�l�o�p�p�p�p�o�o�m�n�o�o�q�s�w�y�}�~��}�{�y�w�t�n�f�^�X�V�U�V�Y�\�_�\�]�`�e�d�`�\�Y�X�[�\�\�Z�X�Z�_�d�g�i�l�n�o�o�o�o�o�o�o�o�o�p�o�o�p�o�o�p�r�s�t�s�o�k�f�a�]�\�\�\�\�\�\�_�_�^�\�Z�Y�X�X�X�W�V�V�W�^�f�q�w�x�y�{�~�}�}�|�{�z�z�y�y�x�w�w�w�w�l�l�m�m�m�m�o�p�p�q�p�o�n�m�m�n�o�p�s�w�z�}����}�{�y�v�r�j�a�Z�U�U�V�X�Z�[�\�\�\�_�d�a�]�X�U�U�Y�Z�Y�X�V�Y�]�b�f�g�k�n�o�o�o�o�n�n�n�n�n�n�o�o�n�n�n�n�o�r�s�r�q�n�i�e�a�_�]�]�\�\�]�^�_�_�]�\�Y�Y�X�X�W�V�U�U�Z�b�m�u�w�y�{�}�}�}�|�{�{�z�y�y�x�w�w�w�x�o�p�p�o�o�o�o�p�q�q�p�o�n�l�l�m�o�p�s�w�z�}����~�|�y�u�o�f�]�V�T�U�X�Y�Z�Z�Y�Y�Z�\�a�_�Z�U�S�T�W�X�Y�X�W�X�]�a�c�f�i�l�o�o�o�n�m�l�l�l�l�l�m�m�l�l�k�l�m�n�o�o�o�n�l�i�f�b�`�^�]�\�\�]�_�`�_�^�[�Z�X�W�W�W�U�U�W�^�h�q�v�y�{�|�|�|�|�{�z�z�y�x�x�w�w�w�z�s�s�s�r�p�p�p�q�q�q�p�o�m�l�l�m�n�p�s�w�z�~�����~�|�x�t�l�c�Z�S�S�U�X�Z�Z�X�V�U�W�Y�]�[�V�S�Q�R�U�W�Y�Y�X�Z�]�a�b�d�g�j�l�m�m�m�l�k�k�j�j�j�k�k�k�j�i�j�j�j�j�j�l�m�n�m�j�f�b�_�]�\�\�\�^�`�a�`�]�[�Y�W�X�X�W�U�V�[�d�m�w�z�|�{�{�{�{�{�z�y�x�w�w�w�w�x�|�v�v�v�t�r�r�r�r�q�q�p�o�m�k�l�m�n�p�s�v�z�~������|�x�r�j�`�W�R�S�U�X�Z�Y�W�U�T�T�V�X�X�U�R�P�R�T�V�Y�Z�Y�Z�]�`�a�b�e�h�j�k�l�l�l�k�j�i�i�i�i�j�k�j�i�i�i�h�g�f�i�l�o�o�l�g�b�_�]�\�[�\�]�_�a�a�^�\�Y�W�X�Y�X�W�V�Z�`�j�v�z�|�|�{�{�{�{�z�y�x�w�w�w�w�y�|�z�y�x�w�v�t�s�s�s�s�r�o�m�k�l�m�o�p�s�u�z�����~�|�x�o�e�\�T�Q�S�U�X�Y�X�W�U�T�S�T�T�T�S�Q�P�Q�T�V�Y�[�\�\�]�^�]�_�b�f�h�j�k�k�k�j�i�h�g�h�h�i�j�k�k�k�i�g�d�c�g�k�n�n�l�h�c�a�_�\�Z�[�\�\�`�a�_�\�Y�W�X�Y�Y�Y�Y�W�[�e�r�y�}�~�}�|�|�{�z�y�x�x�w�x�w�x�|�{�z�y�x�w�w�v�u�u�u�s�p�m�k�l�n�o�q�s�u�z�����|�{�u�m�c�Y�S�P�R�T�V�W�W�W�W�V�T�S�S�R�Q�P�P�R�U�W�Z�\�\�]�]�^�\�^�a�d�g�h�i�i�j�i�h�g�g�g�h�h�j�k�k�k�i�g�d�c�f�i�l�n�l�h�d�b�`�\�Z�Z�Z�Z�]�a�_�]�[�X�X�X�Y�Z�[�Y�[�c�n�x�|�~�}�|�|�{�z�y�x�x�w�x�w�x�|�|�{�{�z�z�y�x�w�w�w�u�q�m�k�m�o�p�s�u�v�z�����{�x�t�i�_�W�Q�N�P�Q�S�T�V�W�Y�X�T�S�Q�P�O�O�Q�R�T�W�[�]�^�^�^�^�\�]�_�a�d�f�g�g�h�g�g�f�g�g�g�h�j�j�k�j�h�g�d�c�d�g�i�l�k�h�e�c�`�\�Z�Z�Y�Y�Z�^�_�_�]�Z�X�W�Y�\�]�\�\�`�j�u�{�}�|�|�|�{�z�y�x�w�w�x�w�x�|�{�{�{�{�{�z�z�y�y�x�v�q�m�k�m�o�q�t�v�w�z�����{�v�p�f�]�U�O�M�M�N�P�S�U�Y�[�[�V�S�Q�O�N�O�Q�R�T�V�[�_�`�`�_�_�]�]�^�_�a�b�c�d�f�e�e�e�e�f�g�g�i�j�j�i�h�f�d�b�b�d�g�i�j�i�g�c�`�\�Z�Y�X�W�W�[�_�`�_�[�X�W�Z�]�_�_�]�^�f�q�x�{�|�|�|�{�z�y�x�w�w�w�w�x�|�z�z�{�{�{�{�z�z�y�x�v�p�m�k�l�o�r�u�v�w�z�����z�u�l�c�Z�T�O�M�L�L�M�Q�U�Z�\�\�X�T�Q�P�P�Q�R�R�S�V�[�_�a�a�a�a�_�^�]�]�^�`�a�b�c�c�d�d�e�f�g�g�i�i�i�h�g�f�d�b�a�b�d�g�h�h�g�d�_�\�Z�Z�X�V�V�Y�^�`�_�]�Z�Y�[�_�`�a�a�^�c�n�w�z�|�|�|�{�z�y�x�w�w�w�w�x�|�y�y�z�{�{�{�z�y�x�w�t�o�k�i�k�n�r�t�v�w�z����~�y�s�i�_�X�S�O�M�L�K�L�Q�V�[�]�]�Y�U�S�R�S�S�R�Q�R�U�Z�^�a�a�b�b�a�`�^�]�\�]�_�`�a�b�b�c�d�f�g�g�g�h�h�h�g�f�d�c�a�a�b�e�g�g�f�c�_�\�[�[�Y�W�V�X�\�]�^�]�\�\�\�_�a�a�a�a�d�l�u�y�{�{�|�{�z�y�x�w�w�w�w�x�|�x�y�z~{~{�{�z�z�x�v�r�m�i�g�i�m�r�t�u�w�y�~��}�x�p�f�]�V�R�O�M�K�J�L�Q�X�\�^�]�Z�W�U�U�V�U�R�P�Q�T�Y�]�`�b�d�f�d�b�_�]�\�\�^�_�a�a�a�c�d�f�g�g�g�g�g�g�g�f�e�c�b�a�a�c�f�g�d�b�_�]�[�\�[�W�W�X�Z�[�\�]�^�^�^�^�`�`�a�c�f�k�r�y�{�{�|�{�z�y�x�w�w�w�w�w�|�x�y�{}}}}�|�z�z�w�t�p�k�g�f�h�k�q�s�u�v�y�~��}�v�n�d�Z�T�Q�O�M�J�J�L�R�Y�]�]�\�Z�X�X�X�Y�W�S�O�Q�T�X�[�_�b�e�g�f�c�`�^�\�\�]�_�`�a�a�c�d�f�g�g�g�g�g�g�g�f�e�d�c�b�a�c�e�e�c�a�_�]�\�\�\�Y�W�X�X�Y�\�]�`�`�_�^�_�_�b�e�g�j�q�y�{�|�|�{�z�y�x�w�w�w�w�w�|�x�{~{{~~�{�y�v�q�k�g�d�b�e�i�o�q�s�u�x�}�~�{�t�m�b�W�R�R�O�L�K�K�M�S�Z�]�]�\�\�Z�Y�Z�\�Z�U�R�P�S�V�Z�]�a�f�h�g�f�b�`�]�]�]�^�^�_�`�a�c�d�e�f�e�e�f�f�f�f�f�f�e�d�d�e�e�e�b�a�`�^�\�\�]�\�Y�U�U�X�]�_�`�^�\�\�]�^�a�d�f�h�o�w�z�}�}�z�y�x�x�w�w�w�w�x�}�x�z~|�{�~�z�x�u�o�h�e�b�a�d�h�o�r�t�v�y�}�~�z�s�k�a�V�R�R�O�K�J�K�N�T�\�^�_�]�\�Z�Z�[�\�\�W�T�Q�R�T�X�\�`�e�h�h�g�d�a�`�_�_�_�^�_�`�a�b�d�e�e�f�g�g�g�g�g�g�g�f�f�e�f�f�e�b�a�a�_�]�]�^�\�Y�U�T�X�]�`�`�]�\�[�\�]�`�c�e�g�n�u�z�}�~�z�y�y�x�w�w�w�w�x�}�w�z�~|�|��}�y�v�r�m�g�d�a�`�c�h�n�r�u�w�z�~�}�y�q�i�^�U�R�Q�N�K�I�K�O�U�]�`�`�^�\�Z�Z�\�^�]�[�V�R�S�T�V�Z�^�c�g�h�g�f�d�b�a�a�`�_�_�_�`�a�d�e�g�h�h�h�h�h�h�h�h�g�g�f�f�f�d�b�a�`�_�^�^�^�\�X�U�U�Y�]�`�_�]�\�[�\�\�_�b�d�f�m�t�y�}�}�z�y�y�x�w�w�w�w�x�~�w�y�}}}�}�x�u�p�k�g�c�a�`�c�h�m�r�v�x�{�}�|�x�o�e�\�T�R�P�M�J�I�K�P�V�\�_�_�^�\�Z�Y�\�^�_�\�Y�T�T�T�U�X�\�a�e�g�g�g�g�f�d�b�a�a�`�`�`�a�d�f�g�h�h�h�h�h�h�g�g�g�g�f�e�e�c�a�_�]�]�]�]�]�\�X�V�W�Y�]�_�`�^�\�[�\�\�^�b�d�f�k�q�x�|�}�z�y�y�x�w�w�x�w�x��u�x�|~~~~�{�w�s�o�k�h�d�a�`�d�i�m�r�v�x�|�}�{�w�m�c�Y�T�R�P�L�J�I�L�Q�V�[�]�_�^�\�Z�X�Z�]�^�]�[�W�V�V�V�X�[�_�a�f�g�h�i�i�g�e�c�b�a�a�a�b�d�f�g�g�g�g�g�g�f�f�e�e�d�d�c�c�a�_�\�Z�Y�Z�\�\�[�Y�Y�Y�[�]�_�`�_�]�\�\�\�^�b�d�f�j�p�w�{�|�z�y�y�y�w�w�x�w�y���t�w�{�~�~�z�u�q�n�k�h�e�a�`�d�j�n�r�v�x�{�}�|�w�l�a�W�T�R�P�L�J�K�N�R�V�Z�]�_�`�^�[�Y�Y�\�]�]�\�[�Z�Y�X�X�Z�\�_�d�f�i�k�l�j�g�e�c�c�c�c�d�e�f�g�f�f�f�e�d�d�c�c�c�b�b�a�a�`�^�\�X�W�X�Y�Z�Z�Z�Z�[�\�]�`�a�a�_�]�]�]�^�b�d�f�j�o�v�z�{�z�y�y�y�w�w�x�w�z���s�v�z�~�}�x�s�p�m�j�h�e�a�_�c�j�o�r�v�w�z�}�{�v�k�_�V�S�S�P�L�J�L�N�S�V�[�_�b�e�e�]�Z�X�Z�]�^�^�]�]�\�Z�Y�Y�Z�\�b�e�i�l�n�l�h�f�e�e�e�e�f�f�f�f�e�e�e�d�c�b�a�a�a�a�a�a�a�`�^�\�Y�Y�Y�X�X�Y�Z�\�\�\�^�a�b�c�a�_�^�^�_�b�f�g�j�o�u�y�{�z�y�z�y�w�w�x�w�z���r�u�z�|�|�x�s�o�m�j�h�d�`�_�c�j�o�r�v�w�y�|�z�u�j�_�U�S�S�P�L�J�L�M�R�V�\�a�f�h�e�`�[�X�Z�\�^�_�`�_�_�\�[�Y�Z�[�`�d�h�l�n�l�i�g�f�g�g�g�g�f�f�f�e�e�e�d�c�b�a�a�a�a�a�a�a�a�_�]�[�[�Z�Y�Y�Y�Z�\�]�]�_�a�d�d�b�`�`�_�`�c�g�h�k�o�u�x�z�y�y�z�y�w�w�x�x�z���q�t�x�{�y�w�s�q�n�k�g�c�a�a�e�j�p�r�w�w�x�z�{�v�j�_�V�T�R�M�J�K�J�M�S�V�]�d�l�l�g�_�\�[�Z�[�]�^�`�b�c�a�_�]�\�]�_�b�e�h�j�j�j�i�h�h�i�h�g�f�f�f�g�g�h�g�f�e�c�c�c�b�b�b�c�d�b�a�`�^�]�\�\�\�\�\�]�_�a�b�c�b�a�`�a�a�b�d�f�h�k�n�t�x�z�y�x�y�y�w�w�x�x�y���q�t�x�z�y�w�t�r�o�m�h�c�a�b�g�k�r�s�w�w�x�z�{�u�j�_�V�T�R�L�I�J�H�M�Q�U�\�e�k�m�g�_�\�]�\�\�_�`�c�d�d�c�a�`�_�`�a�b�c�e�g�i�k�l�j�j�j�i�h�h�h�h�g�g�h�g�f�e�d�d�c�b�b�c�d�d�c�b�a�_�^�]�\�\�\�\�]�^�`�a�b�a�a�`�`�a�b�c�e�h�k�n�s�w�z�y�y�z�y�w�w�w�x�y���p�s�w�z�y�w�u�s�q�n�i�d�a�c�h�m�t�t�w�w�x�{�|�w�j�_�V�T�Q�L�I�H�F�J�O�T�\�e�k�m�g�`�^�`�`�b�e�e�f�f�d�c�a�a�b�b�d�d�d�d�g�j�m�o�m�m�m�m�m�l�k�j�i�h�g�f�e�d�d�d�b�b�b�b�c�c�b�a�`�_�]�\�\�\�]�]�]�^�_�`�a�a�a�`�_�`�a�b�d�g�j�m�q�v�y�y�y�z�y�w�w�w�x�y���p�r�w�y�y�w�u�s�q�o�j�e�b�d�i�o�t�t�v�w�x�|�|�w�j�`�W�T�Q�L�H�G�F�I�M�S�\�e�m�n�h�b�b�e�g�i�m�m�j�h�e�c�b�c�d�e�g�g�f�f�h�l�o�r�p�p�q�r�r�q�p�o�l�i�g�e�d�d�d�d�b�a�a�a�b�b�a�`�^�]�\�[�\�\�\�]�]�]�^�^�_�`�`�`�^�_�`�a�b�e�g�j�o�w�y�y�z�{�z�w�w�w�x�y���p�r�w�x�x�w�t�s�p�n�j�f�d�f�k�s�t�t�v�w�x�}�~�x�l�b�Y�T�Q�M�I�G�G�J�M�R�\�g�n�o�i�e�f�j�m�p�s�s�p�m�j�g�e�e�f�g�h�i�i�k�m�o�r�s�u�t�u�v�x�w�v�r�n�k�g�e�d�d�d�c�b�a�a�a�a�a�`�^�\�[�Z�Y�Z�[�\�]�^�^�_�_�_�_�_�_�]�^�_�`�a�b�e�h�l�s�w�y�z�{�z�w�w�w�x�y���p�s�w�x�w�u�s�q�o�m�j�g�g�h�n�s�u�t�u�w�x�}�~�x�n�d�Z�S�P�M�J�H�H�J�L�R�\�g�o�p�j�f�h�m�p�s�v�w�v�u�r�l�h�g�h�i�j�k�m�o�q�s�t�u�w�w�w�y�|�|�y�v�q�m�i�f�f�f�e�c�b�b�b�b�b�a�`�]�\�Z�Y�X�Y�[�\�]�_�`�`�a�`�`�_�^�\�]�]�^�_�`�b�f�j�q�v�y�z�|�z�w�w�w�x�y���q�t�w�x�w�t�q�o�n�m�l�i�i�l�p�u�v�u�u�v�x�}�~�x�p�g�\�S�P�N�L�I�I�J�L�P�Z�f�n�n�h�d�h�n�p�s�u�v�x�y�w�p�i�e�g�j�l�m�n�o�r�t�v�w�w�w�x�{�}�~�z�w�t�p�l�i�i�i�g�e�d�d�d�d�c�a�`�]�\�Z�Y�X�Y�[�\�^�`�a�a�b�a�`�_�]�\�\�\�\�]�^�a�c�h�o�v�x�z�|�z�w�w�w�x�y���q�t�w�w�w�s�p�o�n�m�m�l�l�n�s�v�w�u�u�v�x�}�~�x�r�i�]�T�P�N�L�J�I�I�J�O�Y�e�m�m�f�b�e�k�n�p�r�s�w�y�w�p�f�a�c�g�m�n�n�o�p�t�w�w�w�w�w�y�}�}�z�w�v�r�o�m�l�k�i�g�f�f�e�e�e�c�a�^�\�Z�Y�Y�Z�\�]�_�a�a�b�c�b�a�_�]�\�\�\�\�\�]�_�a�g�m�u�x�z�{�z�w�w�x�x�{���p�t�w�x�v�r�o�n�n�m�m�m�o�q�u�w�w�v�v�w�z�}�~�|�t�l�`�U�O�N�M�J�I�J�I�L�V�d�j�i�b�a�b�e�h�k�o�p�u�v�t�m�e�a�b�e�i�k�l�m�p�s�v�w�w�v�u�v�v�w�v�u�t�r�q�o�o�m�i�g�f�e�e�e�c�b�a�_�\�[�[�[�\�\�]�^�`�`�a�b�b�a�`�^�]�\�\�\�\�\�]�_�e�j�q�w�z�{�z�x�x�w�x�}���p�t�w�x�v�q�o�m�m�m�m�n�o�s�u�w�w�w�v�x�z�}�~�|�v�n�c�W�Q�O�M�J�I�J�I�K�V�d�j�h�b�a�a�b�d�g�k�n�r�s�p�k�d�a�b�e�h�i�j�l�o�q�t�v�u�t�s�s�t�t�t�s�q�p�o�o�o�m�j�h�g�f�f�e�c�b�a�`�]�\�\�\�\�\�\�]�^�_�_�`�a�a�_�]�\�\�\�[�[�\�]�_�c�h�n�u�y�{�z�y�y�w�x�}���p�s�w�w�u�p�o�m�l�m�n�p�r�t�v�w�w�w�v�x�z�}�}�|�x�q�e�Y�R�P�N�K�K�K�J�L�W�d�i�g�b�a�`�a�b�e�h�k�n�n�m�h�c�a�b�e�h�i�i�j�m�o�q�r�q�p�p�p�p�q�p�p�o�n�n�n�n�m�k�i�g�g�g�e�d�b�a�`�]�]�]�]�]�]�^�^�]�]�]�^�_�_�]�\�\�\�[�Z�[�\�\�^�a�e�k�r�w�z�z�z�y�w�x�}���p�r�v�w�t�p�o�n�m�n�o�r�u�w�w�w�w�w�v�x�z�}�}�|�y�s�g�Z�S�P�M�L�L�L�K�M�X�e�i�e�`�`�`�`�a�c�g�j�l�k�i�f�c�c�e�g�h�h�h�h�j�l�o�o�o�n�n�n�o�o�o�o�n�m�l�l�m�l�j�i�h�g�g�f�d�b�a�`�_�_�_�_�`�`�a�a�_�]�]�]�^�^�]�\�[�[�Z�Z�Z�[�\�\�_�c�i�q�v�y�z�{�y�w�x�}���o�r�u�v�s�p�o�o�o�o�r�u�w�w�w�w�w�w�v�x�z�}�|�{�y�u�i�\�S�O�M�K�L�L�L�N�Y�e�h�d�_�`�a�a�b�c�h�k�j�h�g�d�d�e�g�h�g�g�g�g�g�j�l�m�m�l�l�m�n�o�o�o�o�n�m�l�l�k�i�g�g�g�g�e�d�c�a�a�a�a�a�b�d�e�f�e�a�`�^�]�^�^�]�\�[�Z�Y�Y�Z�[�\�\�^�a�g�n�t�w�y�z�y�w�x�}���o�q�t�u�s�q�p�p�q�s�u�w�x�x�w�w�w�v�v�x�{�}�|�{�z�v�k�]�S�O�M�K�L�L�L�O�Z�f�h�c�_�`�c�c�c�d�g�j�i�g�d�c�e�g�h�i�g�g�f�f�g�h�j�k�m�l�k�l�n�o�o�o�o�o�m�l�k�i�g�f�f�f�e�e�d�c�c�b�b�b�d�f�g�h�i�h�d�b�`�^�^�^�]�\�[�Z�Y�Y�Z�[�\�\�]�`�f�l�q�v�x�y�y�w�x�}���n�p�t�u�s�q�q�r�t�v�w�y�y�y�x�w�w�v�v�x�{�}�}�{�{�w�l�^�T�O�M�L�M�L�L�O�[�g�j�c�_�a�e�f�d�c�f�h�h�e�b�a�d�g�h�h�g�g�e�e�f�g�i�k�l�k�k�l�n�o�o�o�o�m�k�j�i�h�g�e�d�d�c�d�d�d�d�e�e�e�g�g�h�h�i�i�g�e�a�_�^�]�\�\�[�Z�Z�Y�Z�[�\�]�^�`�e�j�p�t�w�w�x�w�x�}���n�p�s�t�s�s�s�s�w�x�y�z�z�y�x�w�w�u�v�x�{�}�}�|�|�x�m�_�U�P�M�L�L�L�L�O�[�g�k�c�_�b�f�g�e�c�d�g�g�c�a�a�c�g�g�g�g�f�e�d�f�g�i�j�l�l�k�l�n�o�o�o�n�l�j�i�h�g�g�e�c�b�b�c�c�d�e�f�g�g�g�h�h�h�g�g�g�f�b�_�]�\�\�\�[�Z�Z�Z�Z�[�\�]�^�a�e�j�n�s�v�w�x�w�x�}���o�q�u�v�v�u�v�w�y�z�{�{�{�{�z�y�w�u�u�w�z�}�~�}�}�x�l�_�V�P�M�J�L�M�L�O�Y�f�l�g�a�b�e�g�e�e�g�g�g�c�a�a�c�f�g�f�e�d�d�d�f�h�k�m�n�o�o�o�o�o�o�p�o�n�l�j�h�g�f�d�b�a�a�a�a�c�e�g�h�i�i�i�i�h�g�g�f�d�b�^�\�\�\�]�\�\�[�Z�Z�[�\�\�^�`�d�h�m�q�u�w�x�x�x�}���q�s�w�w�w�w�w�x�y�y�z�{�{�{�z�y�w�v�u�w�z�}�~�}�}�x�m�_�V�P�M�J�L�M�L�M�X�e�l�i�b�a�e�f�e�f�g�g�g�e�b�c�d�f�g�g�e�d�c�d�f�h�k�n�o�p�p�p�o�o�p�q�p�n�l�j�h�g�e�c�a�a�`�`�a�b�e�g�j�l�m�m�m�k�i�g�e�c�a�^�\�\�\�^�\�\�[�Z�Z�Z�[�\�^�a�d�h�m�p�u�w�x�x�x�}���s�u�w�x�x�w�x�y�y�y�z�{�{�{�z�y�w�v�u�w�z�}�~�}�}�y�o�_�V�Q�M�K�L�L�K�L�V�c�m�j�b�a�d�f�e�f�h�i�i�g�d�d�e�g�g�g�f�e�c�c�d�h�k�n�q�r�r�r�q�p�q�q�p�n�k�i�g�f�d�b�a�`�_�_�a�b�e�g�k�m�o�p�o�l�h�f�b�`�^�\�\�\�^�_�]�\�\�Z�Z�Z�[�\�^�a�e�i�m�q�u�w�x�x�x�}���t�v�x�x�w�w�x�y�x�y�z�{�{�z�y�x�w�v�u�w�z�}�~�~�|�x�n�`�V�Q�M�K�L�L�K�K�S�`�l�k�b�`�b�e�e�g�j�j�k�h�f�f�g�h�h�h�g�e�c�b�c�f�j�m�q�r�s�s�r�p�p�q�o�n�j�h�g�e�d�b�a�_�^�^�`�b�e�g�j�m�o�q�o�k�f�c�_�]�\�\�\�]�_�`�_�^�\�[�[�[�[�\�^�a�e�i�n�q�u�w�x�x�x�~���t�v�w�w�v�v�w�x�w�x�z�{�{�z�y�x�w�v�u�w�y�}�~�~�|�w�m�`�W�R�N�K�L�L�K�L�R�]�g�i�b�`�a�d�e�h�k�l�l�j�g�g�g�i�j�i�g�e�c�b�c�f�i�l�p�q�s�s�r�p�o�o�o�m�j�h�f�e�d�b�a�_�^�]�`�b�d�g�j�n�q�s�q�l�f�a�\�Z�Y�Z�\�_�`�a�a�`�^�\�\�\�\�\�^�a�e�i�n�q�v�w�x�x�x����u�v�w�v�t�t�t�v�w�y�z�{�{�z�y�x�w�v�u�w�y�}�~�~�{�w�m�a�X�R�N�L�L�L�M�M�R�Y�b�f�b�_�`�b�d�h�l�l�m�j�h�g�h�j�k�j�h�f�d�d�d�f�i�l�o�q�r�s�r�p�o�o�o�m�i�g�f�e�d�b�a�_�^�]�`�b�d�g�k�o�t�v�s�n�h�c�[�Y�X�Z�]�a�b�b�d�c�a�_�]�\�\�]�^�a�e�i�n�q�v�w�x�x�x�����v�w�w�u�r�r�r�t�w�x�|�|�|�{�y�x�w�v�u�w�y�}�~�~�{�w�m�a�X�S�N�L�M�M�O�P�R�V�]�c�a�^�_�a�d�i�m�l�m�j�h�g�i�k�k�k�i�h�g�f�g�h�j�l�o�q�s�s�r�p�o�o�o�l�h�g�f�e�d�c�a�`�^�^�`�b�d�g�k�o�t�v�s�n�h�c�\�Y�Y�[�`�c�e�e�g�f�c�a�`�^�^�^�_�a�e�i�n�q�u�w�w�x�x�����w�w�w�t�q�p�q�s�w�x�{�|�|�{�z�x�w�v�t�v�y�}�~�~�{�w�m�b�X�S�N�L�M�N�Q�S�T�T�Y�^�_�]�_�a�d�i�m�m�l�j�g�g�i�k�l�k�k�j�i�h�i�j�l�m�o�q�s�s�s�q�p�p�o�l�h�g�e�e�d�d�b�a�_�^�`�b�d�f�j�n�q�s�p�l�f�a�\�[�[�]�b�e�g�g�h�g�e�b�a�`�`�_�_�a�f�j�n�q�u�w�w�x�y�����y�x�w�s�p�o�o�q�v�x�z�z�{�z�z�y�w�v�t�v�y�}�~�~�z�v�p�e�[�S�N�L�M�P�T�U�U�T�U�V�V�[�_�a�c�h�l�m�j�h�f�f�g�h�k�l�k�k�j�j�k�l�n�o�o�p�q�r�r�r�q�q�o�m�h�g�e�e�e�e�b�`�_�^�`�b�d�f�j�n�p�p�m�h�d�`�^�_�a�c�d�f�g�h�i�h�f�e�d�c�a�a�a�d�h�l�o�s�v�x�w�x�x�����y�x�w�s�p�o�o�p�u�w�x�y�y�y�x�x�w�u�t�v�x�|�~�~�y�v�p�f�\�S�M�L�M�Q�T�U�T�S�S�T�U�Z�^�a�c�h�l�m�j�h�e�e�e�g�j�k�k�k�k�k�l�m�n�n�o�o�o�p�q�q�q�q�p�m�i�g�f�e�e�e�b�`�_�]�_�a�c�e�i�l�o�o�l�h�e�a�`�a�c�e�f�g�g�h�j�h�g�e�e�e�c�c�c�f�j�n�q�t�w�w�w�y�x�����y�x�w�s�p�o�o�o�r�t�v�w�x�x�x�w�v�u�s�u�x�|�~�}�y�v�p�f�\�S�M�L�O�R�T�T�U�R�R�T�U�Z�^�a�c�h�l�l�i�g�d�b�c�f�h�j�l�l�l�l�l�m�m�n�m�m�n�o�o�p�p�p�p�n�j�h�f�f�e�e�b�`�_�]�^�`�a�c�f�i�l�m�l�h�f�d�b�d�f�g�g�h�h�i�j�i�g�f�g�g�f�e�e�h�m�p�s�u�w�w�x�y�x�����y�x�w�s�p�o�o�o�q�s�u�w�w�w�w�w�v�u�s�u�x�{�}�|�x�u�o�e�[�S�M�M�O�Q�S�T�S�S�S�T�U�Z�^�`�d�i�l�l�i�f�b�`�a�d�g�i�l�l�m�m�m�m�m�m�j�k�k�l�m�n�o�p�q�o�k�h�g�g�f�f�c�a�_�]�^�`�a�c�e�g�j�k�j�h�g�g�f�g�h�h�i�h�h�j�l�i�g�g�h�k�i�h�g�j�o�s�v�w�w�w�x�y�x�����w�w�u�s�p�o�o�p�q�s�u�w�y�y�x�w�w�u�t�u�x�{�|�{�x�t�m�c�Y�S�O�O�P�Q�R�S�S�T�T�U�V�Z�]�`�e�j�l�k�h�e�`�]�_�b�f�h�l�m�n�n�n�m�l�k�h�h�h�i�j�m�o�p�q�o�l�i�h�g�g�g�d�b�_�_�_�a�b�d�e�g�i�j�j�i�i�h�h�i�j�j�k�i�j�k�n�k�h�g�j�n�l�k�j�m�q�u�w�w�w�w�x�y�x�����u�u�t�r�p�o�p�q�s�v�x�z�{�{�z�z�w�v�u�v�x�z�{�z�v�r�j�`�W�R�Q�Q�S�S�S�S�T�U�V�V�X�Z�]�`�e�j�l�k�h�d�_�\�]�a�e�h�l�m�n�o�o�m�k�j�g�g�g�g�h�l�o�p�q�p�l�j�i�i�h�g�e�c�a�a�a�c�e�f�g�h�j�k�k�k�k�l�k�l�m�m�l�k�l�m�o�l�h�i�l�o�o�m�m�o�s�v�w�w�w�w�x�z�y�����s�s�r�q�o�o�q�s�v�x�{�}�~�~�}�|�x�w�v�w�x�z�z�y�u�p�g�^�U�Q�R�T�U�T�T�T�T�U�V�V�X�Z�]�`�e�j�m�k�h�d�^�[�\�_�d�h�l�m�o�o�o�m�j�h�g�e�e�f�h�k�o�q�q�p�m�k�j�j�i�h�f�d�b�b�c�f�h�i�k�l�m�m�m�m�n�o�o�o�p�o�o�n�n�o�p�m�i�i�m�p�p�o�o�q�t�v�v�w�w�w�x�z�y�����q�q�q�p�p�p�r�t�x�z�}������~�}�y�x�w�w�x�z�z�y�u�o�f�\�T�R�S�U�Y�W�U�T�U�U�V�V�W�Z�]�`�e�j�l�j�h�c�]�Z�[�^�c�g�k�m�o�o�o�m�j�h�g�e�d�f�h�k�o�p�q�o�m�l�k�k�j�i�g�e�c�d�e�h�j�l�n�o�o�o�o�o�p�q�q�q�r�r�p�o�o�q�q�m�i�i�m�p�p�o�p�r�t�v�v�v�w�w�x�y�z�����r�r�q�q�q�q�r�s�x�z�}�������~�z�{�{�z�y�x�x�x�v�o�f�]�W�U�W�Y�]�Z�U�T�T�V�Y�Y�Y�[�]�`�c�f�h�h�f�b�]�[�\�^�b�f�j�l�n�o�o�l�i�g�f�e�e�f�h�k�n�o�o�o�n�l�k�k�k�j�g�f�e�f�g�i�l�n�o�o�p�p�q�r�r�s�r�r�s�r�q�p�p�p�r�n�j�j�l�o�p�o�p�r�u�u�t�s�t�v�v�x�|����r�r�q�q�q�r�s�t�w�y�}�������~}}}�~�|�y�w�w�w�u�n�e�]�W�V�Y�\�^�[�U�S�U�X�\�]�\�]�`�a�b�c�e�f�d�`�]�[�\�]�a�d�g�i�k�m�m�j�h�g�f�e�e�g�h�k�m�o�o�n�m�l�k�j�k�k�h�h�g�h�i�k�m�n�o�o�p�q�q�r�r�s�r�s�s�s�q�p�p�q�q�p�l�k�l�n�o�o�p�r�u�v�s�q�p�q�s�w�|�����r�r�r�r�r�s�t�u�w�y�}�~�����|}���}�x�u�s�s�s�n�f�^�X�X�\�`�_�Z�U�S�U�[�`�a�a�a�a�a�a�b�b�b�a�_�\�[�\�]�`�b�d�f�h�j�j�i�g�f�f�f�f�g�h�j�l�m�m�l�k�j�j�j�j�j�j�j�j�k�l�m�n�o�o�o�o�p�q�r�r�r�r�s�t�t�s�q�q�q�q�p�n�m�l�l�m�m�o�q�s�s�q�n�l�l�p�u�|����r�r�s�s�t�t�u�u�w�y�}~�}������~�����}�v�q�n�n�s�p�h�_�Y�X�\�b�`�Z�U�T�W�]�b�d�d�d�c�c�b�a�`�_�]�\�\�[�\�]�`�b�d�f�h�j�j�i�g�f�g�g�g�g�h�i�k�k�j�j�i�i�i�i�i�i�k�l�m�n�o�o�o�n�m�m�n�o�p�q�q�q�r�s�u�u�t�s�s�r�p�p�p�o�m�l�k�l�m�n�n�m�j�h�h�h�m�t�{����s�s�t�t�u�u�v�v�w�y�}~�}�������������|�s�l�h�i�q�s�l�c�[�X�\�`�^�[�W�V�Z�`�d�f�f�f�e�e�e�c�`�]�[�[�\�\�\�^�a�d�g�i�k�m�l�j�h�g�g�g�g�h�h�i�i�i�h�h�h�h�h�h�h�h�j�l�m�n�o�n�m�m�k�k�m�n�o�o�o�p�r�t�v�w�v�u�t�t�q�q�r�p�n�l�l�l�k�j�h�f�d�e�g�h�m�r�{�����t�t�u�v�v�v�v�v�w�y�}~�}�~������������z�p�h�d�e�o�r�p�g�^�Y�\�_�_�]�[�\�`�e�f�g�f�f�g�i�j�g�c�_�[�\�]�\�]�^�a�d�i�l�n�o�o�m�j�h�h�h�h�h�h�h�h�g�g�g�g�g�g�g�g�g�h�i�k�l�l�l�k�j�i�k�m�n�o�o�o�p�r�u�w�w�w�v�u�u�s�r�r�q�p�o�n�m�k�h�e�b�a�d�g�i�m�r�{�����u�u�v�w�w�w�v�v�w�y�}�}�������������y�n�e�a�b�k�p�q�k�a�[�[�^�a�a�a�b�h�h�g�g�f�g�i�n�o�m�g�b�]�^�_�]�\�^�a�d�i�k�n�o�o�o�l�j�h�i�i�i�i�h�g�g�f�f�f�f�f�f�e�f�g�g�h�h�h�h�h�i�j�k�m�o�o�p�p�p�r�u�w�x�x�w�v�v�v�s�q�q�r�r�p�o�m�j�g�e�d�f�g�i�n�s�{�����v�v�w�w�w�w�v�v�w�y�}��~�����������y�n�e�`�`�h�n�q�n�c�\�\�_�c�c�e�f�h�i�h�g�g�h�l�p�t�q�l�f�a�a�`�^�\�]�`�c�g�i�m�o�o�o�m�l�j�j�i�i�i�g�g�f�f�f�f�f�f�e�e�d�e�e�e�f�f�f�g�g�j�l�n�p�r�r�r�r�s�u�w�y�y�w�w�w�w�t�q�r�s�t�r�p�o�m�j�h�h�h�i�j�n�t�|�����w�w�w�w�w�w�v�v�w�y�{�~���~�~�~�����|�t�k�c�`�f�j�m�k�f�b�`�a�e�g�h�i�h�g�g�e�e�g�k�p�u�t�p�l�f�c�a�_�]�^�`�b�e�g�j�m�o�o�n�m�l�k�k�i�h�g�g�f�f�f�e�e�e�d�c�c�c�b�a�a�a�a�c�e�i�n�s�v�w�u�u�u�u�v�w�x�y�y�x�w�w�u�t�t�t�t�s�r�o�o�m�k�j�j�k�l�o�u�|�����w�w�w�w�w�w�v�v�w�y�{�~���}�}}����}�x�p�g�b�e�h�l�l�i�f�e�f�i�i�k�j�i�g�g�e�e�g�k�p�t�t�q�p�j�g�c�`�^�_�a�b�e�g�j�m�o�o�n�m�k�k�j�i�h�h�g�g�g�g�g�g�f�e�e�d�d�c�b�b�b�c�e�g�j�n�s�v�v�u�t�u�v�w�w�y�y�y�x�w�w�v�u�t�u�t�s�r�o�o�n�l�k�i�j�l�o�t�|�����w�w�w�w�w�w�v�v�w�x�z�}��}�}~~����{�t�k�e�d�g�j�k�k�j�j�l�n�n�n�k�i�g�f�e�f�h�j�n�r�t�s�r�o�m�g�c�`�_�a�c�e�h�j�m�o�o�n�n�k�j�i�h�h�h�h�h�g�g�g�h�h�h�h�h�g�f�f�e�f�g�h�i�l�o�s�u�u�t�t�u�w�w�x�y�y�x�x�w�w�w�v�v�v�u�t�r�p�o�n�l�k�i�i�k�n�t�{�����w�w�w�w�w�w�v�v�w�w�x�{�~}}|�}�����y�p�i�e�f�h�j�k�m�n�p�o�o�m�j�h�g�f�e�g�h�j�m�o�r�t�u�u�q�l�e�a�`�b�d�f�h�j�m�o�o�o�n�k�i�h�g�g�g�g�g�g�g�h�i�i�j�j�j�j�i�i�i�j�l�l�l�n�q�t�v�v�v�v�w�w�w�x�x�x�w�w�w�x�w�w�w�w�v�t�s�q�p�o�m�k�i�j�k�n�t�{�����w�w�w�w�w�w�v�v�w�w�w�z�}~}�}�|�}�~�����|�u�n�g�g�g�i�j�l�o�p�o�n�l�j�h�g�g�g�g�h�j�l�n�p�t�w�x�v�p�h�b�a�b�e�g�h�j�m�n�o�n�n�k�j�h�f�e�e�e�f�g�g�h�j�k�k�l�m�n�m�l�m�n�p�o�o�p�s�v�w�x�x�x�z�y�x�x�w�w�w�w�w�x�w�w�w�w�w�u�t�r�r�q�o�m�k�l�l�o�u�|�����w�w�w�w�w�w�v�v�w�v�w�x�|~�}��~�}~����y�t�l�i�h�h�h�j�l�m�l�k�j�i�i�h�h�h�i�j�k�l�n�q�u�x�z�x�t�l�e�b�c�f�h�i�k�m�n�n�n�n�l�j�h�f�d�c�c�e�g�h�i�k�l�m�n�o�o�o�o�o�r�s�q�o�q�t�x�z�z�{�|�}�{�z�y�x�w�w�w�x�x�w�w�w�x�w�v�u�u�u�t�r�p�n�n�n�p�u�~�����w�w�w�w�w�w�v�v�v�u�v�x�{�}����~}~~���|�x�r�o�k�h�g�h�h�i�h�h�i�j�j�k�k�k�k�m�n�o�o�s�v�z�|�{�w�p�i�f�f�h�i�j�l�m�n�n�m�m�l�j�h�f�d�d�e�f�i�j�k�l�n�o�o�o�q�p�o�q�t�t�r�p�q�t�y�|�}�}�~��~�}�|�z�y�y�y�y�x�w�w�x�x�x�w�v�w�w�w�u�r�p�p�p�q�v������w�w�w�w�w�w�v�v�v�u�v�w�{�~����}}}~��~�{�w�s�o�j�h�g�g�g�g�g�i�k�l�m�n�n�n�o�o�p�q�u�x�|�|�}�y�s�m�i�h�j�j�k�l�m�n�n�m�m�l�j�h�f�f�f�h�i�j�l�m�n�o�o�p�q�r�p�p�r�t�t�r�o�o�s�y�|�~�~�~�����~�|�{�{�{�{�x�w�x�x�y�x�w�w�x�y�x�w�t�s�r�r�s�w������w�w�x�x�w�w�w�w�v�u�u�w�y�}�������~~}~~�~���|�x�u�q�m�j�h�g�h�h�j�l�n�n�o�o�p�o�o�o�s�w�z�|�{�}�y�s�r�p�l�j�i�k�l�n�n�m�n�n�l�k�j�j�j�l�m�n�n�o�o�o�p�p�q�r�r�q�p�q�q�p�n�k�m�o�u�{�~�~}|�}�}��~�~�~�}�{�z�z�y�y�x�x�w�z�{�{�y�w�v�u�u�u�x�������w�w�w�w�w�w�w�w�v�u�u�w�x�|�������}}�����|�z�x�u�r�p�m�l�k�k�l�m�n�o�o�p�r�p�o�p�t�x�{�|�z�|�x�t�v�u�o�l�k�k�l�m�m�m�m�n�m�m�m�m�m�n�o�p�o�o�o�o�p�q�q�r�r�q�o�p�p�o�l�i�j�m�s�y�~�~}|�{�{�|�}�~�~~�|�|�{�{�z�y�y�y�z�{�{�z�x�w�v�v�w�y�������w�w�w�w�w�w�w�w�v�u�u�v�x�{�~���}�}��������~�}�|�{�y�v�t�r�o�n�m�m�n�o�q�r�r�q�q�s�u�x�z�{�y�y�v�u�x�y�t�o�m�l�l�m�m�m�m�m�n�o�o�o�o�q�r�r�p�p�p�p�p�q�r�r�q�p�o�o�o�n�j�g�h�j�q�x�}�}~{�y�z�z�{�|�}�}�~��~�~�}�|�{�{�z�z�z�{�{�z�x�w�w�w�w�z�������u�u�u�v�v�v�w�w�v�v�v�v�x�z�|�~�~�}�}�������}�����~�|�y�w�r�p�n�m�m�o�p�r�s�s�s�t�w�x�z�{�x�v�s�u�y�{�w�s�n�m�l�l�l�l�l�m�o�o�o�p�q�r�s�s�q�p�p�p�p�q�r�r�q�o�n�n�n�l�h�f�f�h�o�w�|�}�{�x�y�z�{�{�|�}�~��~�~�~�}�|�{�z�z�z�{�{�z�y�w�x�x�x�{�������s�s�s�t�t�t�u�v�v�v�v�w�x�z�|�~�~�}�}��������}���������~�{�x�u�r�o�m�l�m�o�p�r�s�t�v�w�w�y�y�v�r�q�u�y�{�x�t�o�l�j�j�k�l�l�l�n�o�p�q�r�s�s�s�q�p�o�o�o�p�p�q�o�o�m�m�m�k�h�e�e�h�n�v�{�|�~|�z�z�z�{�|�}����~�~�}�}�|�{�{�z�z�{�{�z�y�x�x�y�y�|������r�r�r�r�r�s�t�u�v�v�w�w�y�{�}�~��}�}��������~���������|�y�w�t�p�m�k�l�n�n�o�r�u�v�w�w�w�x�v�q�q�v�y�y�w�t�n�k�h�i�j�l�m�m�n�o�p�q�r�r�r�r�p�p�o�n�o�o�o�o�o�n�l�m�m�k�h�f�f�h�o�v�{�{�}~�{�z�{�|�}������~�}�}�}�|�|�{�{�{�{�{�z�y�x�y�z�z�|������t�s�s�s�s�t�u�u�v�w�w�y�z�|�~�~���}�|��������~�����~���}�|�y�v�r�n�k�k�k�l�n�p�t�v�v�v�w�w�v�q�s�x�w�w�w�t�n�j�g�g�j�m�n�m�n�o�p�r�r�r�q�q�o�o�n�m�m�m�n�o�n�m�l�m�m�l�i�g�g�j�p�w�{�{�}�{�z�z�{�}������~�~�~�~�}�}�}�}�{�{�{�z�y�x�x�y�y�|�������u�u�t�t�t�u�v�v�v�w�w�y�{�}�~����}�}�~������}�����~����~�{�x�t�p�m�l�k�l�m�p�t�v�v�u�v�w�v�s�u�x�v�t�u�t�m�i�f�g�i�l�n�n�o�o�q�r�s�r�q�p�o�o�m�l�k�l�m�n�m�l�l�m�n�m�j�h�h�l�r�x�{�{�}�~z�z�z�{�}������~�~�~�~�~�~�~�~�|�|�{�z�y�x�x�y�y�{�������u�u�u�t�t�t�u�u�t�w�x�y�z�{�}��������������}�}�������������~�|�x�u�q�o�m�l�m�q�u�w�v�u�u�w�v�w�w�v�r�p�q�p�i�e�c�f�h�i�j�k�k�m�o�p�q�q�p�o�o�o�m�l�l�l�m�n�m�l�l�l�m�m�l�k�k�o�t�x�y�z�|�~{z�z�|�}��~�}�}�}�}�~������}�|�{�z�z�z�y�y�z�}������u�u�u�u�u�u�u�u�v�w�w�x�x�y�|�~���������}�}�}���������������~�|�y�u�r�o�n�p�s�w�w�v�u�u�v�w�w�w�u�r�p�o�n�h�d�b�c�g�g�i�k�k�l�o�p�q�q�p�o�o�o�m�l�l�l�m�m�m�l�l�m�n�n�m�m�n�q�v�x�x�y�{�~|{�{�|����~}|�|�}�������~�|�{�z�{�{�z�y�y�|�����v�v�u�u�u�u�v�w�x�w�v�t�u�x�|�~�}�����}�}�|�}�}����������~�����}�z�w�s�r�t�w�z�z�x�w�w�w�w�w�w�v�u�r�p�m�h�c�`�a�d�f�h�l�k�l�o�p�q�q�p�o�o�o�n�m�l�l�m�m�l�l�l�n�o�o�o�o�p�t�x�y�x�y�{�~~|�|�~������~~~}~}�}�}�����~�}�|�{�{�{�z�y�x�|�����v�v�v�v�v�v�v�w�z�x�u�r�s�w�|��}�}�����}�|�|�|�}����������|�}������~�{�x�w�x�z�|�}�{�y�x�x�x�w�w�w�w�v�q�m�h�c�_�`�a�d�h�k�k�l�o�p�q�p�o�o�p�o�o�n�m�m�m�m�l�l�m�o�p�q�p�o�s�v�y�y�x�x�z�}�}�}������~�~}}~}}�}����~�~�}�{�{�{�{�y�x�w�|������v�v�v�u�v�v�v�w�z�w�t�r�t�x�}�|�}�������}�|�|�|�}����������|�|�~�����~�}�|�{�}�~�~�}�{�z�y�w�v�w�x�z�x�t�p�k�d�`�`�a�c�g�j�k�l�n�o�p�p�o�o�p�p�o�o�n�m�m�m�l�m�n�p�r�r�q�p�s�w�x�x�w�x�z�~���}������~�~}~~}�}����~�}�|�{�z�z�y�x�w�w�|������u�u�u�u�u�v�v�v�x�v�t�t�w�{�~�|�}�������}�|�|�|�}�}�����}�}}~~~�}~~~�|�z�x�w�r�s�v�y�z�y�v�s�o�h�b�a�a�c�f�i�k�l�n�o�p�p�o�o�p�p�o�o�o�n�n�n�m�n�o�q�s�s�q�p�r�u�w�w�v�w�z�~���}�������~�~�~�~����~�~�}�|�{�z�z�y�x�w�w�|�������u�t�t�t�u�u�v�v�u�t�t�v�z�~��}�|�~���������~�|�|�|�|�}�}�}�}�}���~~~}}�~�~�}{�y�w�w�r�l�o�u�x�z�x�w�v�r�k�e�b�b�c�e�h�k�l�n�o�p�p�o�o�p�p�o�o�o�o�o�o�n�o�p�s�t�s�q�o�p�s�u�u�u�v�z���}�|�}�����~�~����~�~�~�}�|�{�{�z�y�x�w�w�|�������t�t�t�t�t�u�u�v�s�s�u�x�}���}�|�����������|�|�|�|�}�}�}�|�}������~�~~~~~�~�|�y�w�t�q�k�f�k�s�x�y�x�w�w�u�n�h�e�d�c�e�g�k�l�n�o�p�p�o�o�p�o�o�o�o�o�o�o�o�o�q�s�t�s�p�o�p�q�s�t�t�v�z���}�|�}������������~�~�~�~�~�~�}�|�|�{�z�y�x�x�}�������t�s�s�t�u�v�v�u�q�q�u�{��|�}��}�����������}�~�}�|�{�{�{|�|�}���~�~|�{�}�~�}�{�x�v�s�n�g�b�a�i�s�z�{�y�w�w�v�s�o�h�c�b�d�g�j�l�n�o�o�o�o�o�o�o�o�p�p�q�r�r�q�q�r�r�q�p�p�o�p�r�r�r�s�v�{�~�}�~�������������~�~�~�}�}�}�}�}�~�~�~�}�}�|�}��������t�s�s�t�u�v�u�s�o�q�v�|�|�|��}������������}�}�|�{�z�z�z�{�{�|���~�~|�|�}�}�z�w�s�p�n�i�c�_�a�i�t�{�|�y�w�v�v�t�q�l�f�d�e�g�i�k�m�o�o�o�o�o�p�p�p�p�q�r�s�s�s�s�r�r�r�r�q�q�r�r�r�q�q�v�{�}�}��������������~�~�~�}�}�}�~�~�~����~����������s�r�s�t�u�t�r�p�n�p�v�}�|�{�}�}�����������}�{{�z�y�y�y�z�{�|����~~�~�|�w�r�m�h�f�d�a�_�b�k�v�}�~�z�v�t�t�t�r�o�j�g�g�g�h�k�m�n�o�o�o�o�p�p�p�p�q�r�s�s�t�s�r�r�s�t�t�t�t�r�o�o�q�v�|�}�|�}������������~�~~~�~�~�~����~����������������r�r�s�t�t�s�p�n�l�o�w�~�|�{�|�}�����������}�{�z�y�y�y�y�z�{�}����~�~�~�}�{�y�t�n�h�b�`�`�`�a�e�m�y���{�v�t�s�t�t�r�n�j�g�g�h�k�m�n�n�o�o�p�q�q�p�p�q�r�r�s�r�q�p�r�u�y�{�y�v�r�n�m�p�v�|�}�|�}����������~~~~~~��~�}�|�|�|�}�~������������r�r�s�t�t�q�n�l�l�q�y�}�{�{�|�}�}�����~�}�{�z�z�z�{�{�||~�~�}�|�{�y�x�v�p�j�c�^�\�]�`�b�h�q�z���z�v�s�r�s�t�s�p�m�i�g�i�k�m�n�o�o�p�q�r�q�q�p�p�q�q�q�o�n�o�r�x�|���y�r�m�l�p�w�}�}�|�}���������~~~~~}|||�|�|�}���������s�s�s�s�s�p�n�l�o�u�{�{�{�{�{�|�|�|�}�}�|�|�|�|�|�|�|�}�}~~}}�}�}�|�z�x�w�u�s�n�h�b�^�\�]�a�f�k�r�{�~�|�x�u�s�q�q�r�r�p�n�k�j�j�l�n�o�o�o�p�r�s�r�q�p�o�o�o�o�m�l�m�r�y�������{�s�m�l�q�x�}�}�}���������~~~}~}}�}}~}~}~}~}~~~��������t�s�r�r�r�p�n�n�t�y~{�z�z�{�{�{�{�{�{�{�{�{�|}�~�~��~�|�{�|�|�|�|�z�x�v�u�s�o�l�g�b�a�c�g�j�n�s�z�{�x�u�s�r�p�p�p�p�o�o�m�l�l�n�o�o�o�p�q�s�s�s�q�p�o�o�n�n�k�j�l�r�{������|�s�m�m�r�y�~�~}�����������~~~}~}}�}}~~~~�~�~�~��������t�s�r�r�q�o�o�o�x~z�y�y�z�{�{�{�{�z�z�z�z�{|~~�~�~�}�{�z�z�{�|�|�|�{�x�v�u�u�t�q�n�j�i�j�n�p�p�s�v�w�u�r�q�q�p�o�o�o�o�o�n�m�n�o�o�p�p�p�r�s�t�s�q�o�o�n�m�m�j�i�k�r�{������{�t�n�n�s�{�~�~����������~~}~}}}~~�~�~�~�~�~�~���������s�s�r�q�o�o�r�u�|�x�w�y�z�z�z�z�z�z�z�{�|~~~�|�|�|�{�z�y�y�y�x�{�}�{�x�v�w�y�x�x�w�v�t�t�t�u�r�q�r�r�q�p�o�o�o�n�n�m�n�n�n�n�n�n�o�o�p�q�r�s�s�r�q�o�o�n�n�n�j�h�i�q�{�������z�s�o�p�u�z��������������~~~}}��~�~�}�}�}�~��������r�s�s�q�q�r�u�y�{�x�w�y�z�z�y�{�{�{�|�|�}~~�}�|�{�{�{�z�y�y�y�z�|�}�{�w�u�w�y�|�}�}�}�|�{�z�x�s�q�p�o�o�o�n�n�n�m�m�m�m�m�n�n�n�n�o�o�p�q�r�r�r�r�p�o�o�n�n�n�k�h�i�o�x�����y�r�o�q�v�{��������������~~~}}��~�~�}�|�}�~��������r�s�t�s�t�v�y~}�{�y�x�y�z�z�z�|�}�}�}~~~�~�}�|�{�{�{�{�{�{�{�}�~�~�y�u�s�u�y�~���~�~�~�|�{�u�r�p�o�m�m�m�n�m�l�l�l�m�m�n�n�o�o�o�o�p�p�q�q�q�q�p�o�o�o�o�o�m�j�i�m�t�|�~�~�w�q�o�r�x�}��������~~����~~~}~}���~�}�|�~�~��������r�s�u�v�w�y|~z�|�{�z�y�y�z�z�{���~�~�~�~�}�}�}�|�|�}�}�}�}�}�}~��}�x�s�q�t�x�}�~��}�|�~~�|�{�w�t�q�o�m�l�m�n�m�m�m�m�n�o�o�o�o�o�o�o�o�p�p�p�p�p�o�o�o�o�o�o�o�l�i�j�p�x�|�|�u�p�o�t�z�~������~~����~~~}~}���~�}�}�~�~��������r�t�v�w�x}|x�v�|�{�{�z�y�y�z�|���~�~�}�}�}�}�}�}�~�~�~���~�~}�~��}�x�q�o�r�w�|�~��}�|�~�|�{�y�v�r�o�m�l�m�n�n�n�n�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�o�n�i�g�m�v�|�{�u�o�o�t�|�������}�}}~~����~~~}~}����~�}�}�~�������
\ No newline at end of file
diff --git a/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0001.jpg b/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0001.jpg
new file mode 100644
index 000000000..39e94a04a
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0001.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0002.jpg b/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0002.jpg
new file mode 100644
index 000000000..91012435d
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/Rob_Lowe_0002.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/blur.jpg b/cpp-package/inspireface/test_res/data/bulk/blur.jpg
new file mode 100644
index 000000000..a84afd597
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/blur.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/face_sample.png b/cpp-package/inspireface/test_res/data/bulk/face_sample.png
new file mode 100644
index 000000000..dd2870924
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/face_sample.png differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/image_T1.jpeg b/cpp-package/inspireface/test_res/data/bulk/image_T1.jpeg
new file mode 100644
index 000000000..11e7f4c71
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/image_T1.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/jntm.jpg b/cpp-package/inspireface/test_res/data/bulk/jntm.jpg
new file mode 100644
index 000000000..2d1f8238b
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/jntm.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/jntm.png b/cpp-package/inspireface/test_res/data/bulk/jntm.png
new file mode 100644
index 000000000..2d1f8238b
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/jntm.png differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/kun.jpg b/cpp-package/inspireface/test_res/data/bulk/kun.jpg
new file mode 100644
index 000000000..a6a34b46c
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/kun.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop.jpg b/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop.jpg
new file mode 100644
index 000000000..058b91c82
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop_r90.jpg b/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop_r90.jpg
new file mode 100644
index 000000000..9c4157e15
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/kun_cartoon_crop_r90.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/kun_crop.jpg b/cpp-package/inspireface/test_res/data/bulk/kun_crop.jpg
new file mode 100644
index 000000000..bae5c2408
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/kun_crop.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/mask.png b/cpp-package/inspireface/test_res/data/bulk/mask.png
new file mode 100644
index 000000000..70420699e
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/mask.png differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/mask2.jpg b/cpp-package/inspireface/test_res/data/bulk/mask2.jpg
new file mode 100644
index 000000000..0ffa843bd
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/mask2.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/pedestrian.png b/cpp-package/inspireface/test_res/data/bulk/pedestrian.png
new file mode 100644
index 000000000..02c50baf9
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/pedestrian.png differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r0.jpg b/cpp-package/inspireface/test_res/data/bulk/r0.jpg
new file mode 100644
index 000000000..303d131ae
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r0.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r0_w330_h409_c3.nv21 b/cpp-package/inspireface/test_res/data/bulk/r0_w330_h409_c3.nv21
new file mode 100644
index 000000000..130c703a4
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r0_w330_h409_c3.nv21 differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r180.jpg b/cpp-package/inspireface/test_res/data/bulk/r180.jpg
new file mode 100644
index 000000000..0f2b3a705
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r180.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r180_w330_h409_c3.nv21 b/cpp-package/inspireface/test_res/data/bulk/r180_w330_h409_c3.nv21
new file mode 100644
index 000000000..63b79c34d
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r180_w330_h409_c3.nv21 differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r270.jpg b/cpp-package/inspireface/test_res/data/bulk/r270.jpg
new file mode 100644
index 000000000..8903ff3db
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r270.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r270_w409_h330_c3.nv21 b/cpp-package/inspireface/test_res/data/bulk/r270_w409_h330_c3.nv21
new file mode 100644
index 000000000..4c662fb5e
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r270_w409_h330_c3.nv21 differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r90.jpg b/cpp-package/inspireface/test_res/data/bulk/r90.jpg
new file mode 100644
index 000000000..a5faefdcd
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r90.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/r90_w409_h330_c3.nv21 b/cpp-package/inspireface/test_res/data/bulk/r90_w409_h330_c3.nv21
new file mode 100644
index 000000000..bfe37d203
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/r90_w409_h330_c3.nv21 differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/rgb_fake.jpg b/cpp-package/inspireface/test_res/data/bulk/rgb_fake.jpg
new file mode 100644
index 000000000..c557f150c
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/rgb_fake.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/view.jpg b/cpp-package/inspireface/test_res/data/bulk/view.jpg
new file mode 100644
index 000000000..3a97ae25e
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/view.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/woman.png b/cpp-package/inspireface/test_res/data/bulk/woman.png
new file mode 100644
index 000000000..dd2870924
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/woman.png differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/woman_search.jpeg b/cpp-package/inspireface/test_res/data/bulk/woman_search.jpeg
new file mode 100644
index 000000000..75a383a2d
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/woman_search.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/bulk/yifei.jpg b/cpp-package/inspireface/test_res/data/bulk/yifei.jpg
new file mode 100644
index 000000000..79c804d94
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/bulk/yifei.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/crop/crop.png b/cpp-package/inspireface/test_res/data/crop/crop.png
new file mode 100644
index 000000000..02fb97440
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/crop/crop.png differ
diff --git a/cpp-package/inspireface/test_res/data/crop/crop_mask.png b/cpp-package/inspireface/test_res/data/crop/crop_mask.png
new file mode 100644
index 000000000..f992ca6af
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/crop/crop_mask.png differ
diff --git a/cpp-package/inspireface/test_res/data/crop/no_face.png b/cpp-package/inspireface/test_res/data/crop/no_face.png
new file mode 100644
index 000000000..ba0cd0275
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/crop/no_face.png differ
diff --git a/cpp-package/inspireface/test_res/data/emotion/anger.png b/cpp-package/inspireface/test_res/data/emotion/anger.png
new file mode 100644
index 000000000..f8c31362b
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/emotion/anger.png differ
diff --git a/cpp-package/inspireface/test_res/data/emotion/happy.png b/cpp-package/inspireface/test_res/data/emotion/happy.png
new file mode 100644
index 000000000..665149d82
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/emotion/happy.png differ
diff --git a/cpp-package/inspireface/test_res/data/emotion/sad.png b/cpp-package/inspireface/test_res/data/emotion/sad.png
new file mode 100644
index 000000000..b66ca16ab
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/emotion/sad.png differ
diff --git a/cpp-package/inspireface/test_res/data/pose/left_face.jpeg b/cpp-package/inspireface/test_res/data/pose/left_face.jpeg
new file mode 100644
index 000000000..1b9b78538
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/left_face.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/pose/left_wryneck.png b/cpp-package/inspireface/test_res/data/pose/left_wryneck.png
new file mode 100644
index 000000000..f7f1651d0
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/left_wryneck.png differ
diff --git a/cpp-package/inspireface/test_res/data/pose/lower_face.jpeg b/cpp-package/inspireface/test_res/data/pose/lower_face.jpeg
new file mode 100644
index 000000000..cfd98f680
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/lower_face.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/pose/right_face.png b/cpp-package/inspireface/test_res/data/pose/right_face.png
new file mode 100644
index 000000000..2150073da
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/right_face.png differ
diff --git a/cpp-package/inspireface/test_res/data/pose/right_wryneck.png b/cpp-package/inspireface/test_res/data/pose/right_wryneck.png
new file mode 100644
index 000000000..aec02564d
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/right_wryneck.png differ
diff --git a/cpp-package/inspireface/test_res/data/pose/rise_face.jpeg b/cpp-package/inspireface/test_res/data/pose/rise_face.jpeg
new file mode 100644
index 000000000..b1ea0b229
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/pose/rise_face.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/reaction/close_eyes.jpeg b/cpp-package/inspireface/test_res/data/reaction/close_eyes.jpeg
new file mode 100644
index 000000000..1e365ce51
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/reaction/close_eyes.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/reaction/close_open_eyes.jpeg b/cpp-package/inspireface/test_res/data/reaction/close_open_eyes.jpeg
new file mode 100644
index 000000000..cd2e91697
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/reaction/close_open_eyes.jpeg differ
diff --git a/cpp-package/inspireface/test_res/data/reaction/open_eyes.png b/cpp-package/inspireface/test_res/data/reaction/open_eyes.png
new file mode 100644
index 000000000..3c9b987e7
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/reaction/open_eyes.png differ
diff --git a/cpp-package/inspireface/test_res/data/rotate/rot_0.jpg b/cpp-package/inspireface/test_res/data/rotate/rot_0.jpg
new file mode 100644
index 000000000..2fd96c40f
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/rotate/rot_0.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/rotate/rot_180.jpg b/cpp-package/inspireface/test_res/data/rotate/rot_180.jpg
new file mode 100644
index 000000000..cc519ccd6
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/rotate/rot_180.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/rotate/rot_270.jpg b/cpp-package/inspireface/test_res/data/rotate/rot_270.jpg
new file mode 100644
index 000000000..d14c45e0c
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/rotate/rot_270.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/rotate/rot_90.jpg b/cpp-package/inspireface/test_res/data/rotate/rot_90.jpg
new file mode 100644
index 000000000..478e875a3
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/rotate/rot_90.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/search/Mary_Katherine_Smart_0001_5k.jpg b/cpp-package/inspireface/test_res/data/search/Mary_Katherine_Smart_0001_5k.jpg
new file mode 100644
index 000000000..ec232451c
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/search/Mary_Katherine_Smart_0001_5k.jpg differ
diff --git a/cpp-package/inspireface/test_res/data/search/Teresa_Williams_0001_1k.jpg b/cpp-package/inspireface/test_res/data/search/Teresa_Williams_0001_1k.jpg
new file mode 100644
index 000000000..d6a9913cd
Binary files /dev/null and b/cpp-package/inspireface/test_res/data/search/Teresa_Williams_0001_1k.jpg differ
diff --git a/RetinaFace/rcnn/processing/__init__.py b/cpp-package/inspireface/test_res/save/.gitkeep
similarity index 100%
rename from RetinaFace/rcnn/processing/__init__.py
rename to cpp-package/inspireface/test_res/save/.gitkeep
diff --git a/cpp-package/inspireface/test_res/valid_lfw_funneled.txt b/cpp-package/inspireface/test_res/valid_lfw_funneled.txt
new file mode 100644
index 000000000..1e982f8c7
--- /dev/null
+++ b/cpp-package/inspireface/test_res/valid_lfw_funneled.txt
@@ -0,0 +1,5749 @@
+German_Khan/German_Khan_0001.jpg
+Stefano_Gabbana/Stefano_Gabbana_0001.jpg
+Dragan_Covic/Dragan_Covic_0001.jpg
+Jeff_Hornacek/Jeff_Hornacek_0001.jpg
+Sureyya_Ayhan/Sureyya_Ayhan_0001.jpg
+Deb_Santos/Deb_Santos_0001.jpg
+Bob_Newhart/Bob_Newhart_0001.jpg
+Wang_Hailan/Wang_Hailan_0001.jpg
+Paul_McNulty/Paul_McNulty_0002.jpg
+Jimmy_Iovine/Jimmy_Iovine_0001.jpg
+Claudia_Pechstein/Claudia_Pechstein_0001.jpg
+Ranil_Wickremasinghe/Ranil_Wickremasinghe_0002.jpg
+Ben_Chandler/Ben_Chandler_0001.jpg
+Mark_Komara/Mark_Komara_0001.jpg
+Rand_Beers/Rand_Beers_0001.jpg
+Joanne_Woodward/Joanne_Woodward_0001.jpg
+John_Bond/John_Bond_0001.jpg
+Reginald_Hudlin/Reginald_Hudlin_0001.jpg
+Lee_Baca/Lee_Baca_0001.jpg
+Mary-Kate_Olsen/Mary-Kate_Olsen_0002.jpg
+Emily_Stevens/Emily_Stevens_0001.jpg
+Xiang_Huaicheng/Xiang_Huaicheng_0001.jpg
+Phil_Mickelson/Phil_Mickelson_0002.jpg
+Gerry_Kelly/Gerry_Kelly_0001.jpg
+Salma_Hayek/Salma_Hayek_0012.jpg
+Jim_Edmonds/Jim_Edmonds_0001.jpg
+Martina_McBride/Martina_McBride_0003.jpg
+Anthony_Pico/Anthony_Pico_0001.jpg
+Jose_Theodore/Jose_Theodore_0002.jpg
+Heidi_Fleiss/Heidi_Fleiss_0002.jpg
+Mark_Richt/Mark_Richt_0001.jpg
+Mike_Smith/Mike_Smith_0001.jpg
+Paul_ONeill/Paul_ONeill_0008.jpg
+Horace_Newcomb/Horace_Newcomb_0001.jpg
+Mohammad_Fares/Mohammad_Fares_0001.jpg
+JK_Rowling/JK_Rowling_0004.jpg
+Shinya_Taniguchi/Shinya_Taniguchi_0001.jpg
+Tony_Curtis/Tony_Curtis_0002.jpg
+Laurie_Pirtle/Laurie_Pirtle_0001.jpg
+Angel_Maza/Angel_Maza_0001.jpg
+Ray_Sherman/Ray_Sherman_0001.jpg
+Nancy_Powell/Nancy_Powell_0001.jpg
+Kathryn_Tucker/Kathryn_Tucker_0001.jpg
+Gro_Harlem_Brundtland/Gro_Harlem_Brundtland_0002.jpg
+Jacqueline_Marris/Jacqueline_Marris_0001.jpg
+Mikhail_Gorbachev/Mikhail_Gorbachev_0002.jpg
+Takuma_Sato/Takuma_Sato_0001.jpg
+Ray_Lewis/Ray_Lewis_0001.jpg
+Hank_Aaron/Hank_Aaron_0001.jpg
+Chip_Knight/Chip_Knight_0001.jpg
+Tracee_Treadwell/Tracee_Treadwell_0001.jpg
+Ronald_Young_Jr/Ronald_Young_Jr_0001.jpg
+Gustavo_Cisneros/Gustavo_Cisneros_0001.jpg
+Jack_Osbourne/Jack_Osbourne_0001.jpg
+Isidro_Pastor/Isidro_Pastor_0001.jpg
+Mike_Duke/Mike_Duke_0001.jpg
+Natalie_Williams/Natalie_Williams_0001.jpg
+Angela_Lansbury/Angela_Lansbury_0002.jpg
+Cathy_Freeman/Cathy_Freeman_0002.jpg
+Milt_Palacio/Milt_Palacio_0001.jpg
+Wilbert_Elki_Meza_Majino/Wilbert_Elki_Meza_Majino_0001.jpg
+Ekke_Hard_Forberg/Ekke_Hard_Forberg_0001.jpg
+Jack_Welch/Jack_Welch_0001.jpg
+Nizar_Trabelsi/Nizar_Trabelsi_0001.jpg
+Gonzalo_Barrientos/Gonzalo_Barrientos_0001.jpg
+Jason_Biggs/Jason_Biggs_0001.jpg
+Lisa_Leslie/Lisa_Leslie_0001.jpg
+Tayyeb_Abdel_Rahim/Tayyeb_Abdel_Rahim_0001.jpg
+Tyron_Garner/Tyron_Garner_0001.jpg
+John_Sidgmore/John_Sidgmore_0001.jpg
+Alexander_Payne/Alexander_Payne_0001.jpg
+Brad_Brownell/Brad_Brownell_0001.jpg
+John_Duprey/John_Duprey_0001.jpg
+Ainsworth_Dyer/Ainsworth_Dyer_0001.jpg
+Christine_Ebersole/Christine_Ebersole_0002.jpg
+Amer_al-Saadi/Amer_al-Saadi_0004.jpg
+Alison_Lohman/Alison_Lohman_0001.jpg
+Duane_Lee_Chapman/Duane_Lee_Chapman_0001.jpg
+Andre_Smith/Andre_Smith_0001.jpg
+Iban_Mayo/Iban_Mayo_0002.jpg
+Jeong_Se-hyun/Jeong_Se-hyun_0009.jpg
+Charles_Rogers/Charles_Rogers_0001.jpg
+Eric_Rosser/Eric_Rosser_0001.jpg
+Ziwang_Xu/Ziwang_Xu_0001.jpg
+Larry_Pleau/Larry_Pleau_0001.jpg
+Bill_Byrne/Bill_Byrne_0001.jpg
+Peter_Gabriel/Peter_Gabriel_0001.jpg
+Paul_LeClerc/Paul_LeClerc_0001.jpg
+Maria_Bello/Maria_Bello_0001.jpg
+Shawn_Kemp/Shawn_Kemp_0001.jpg
+Liza_Minnelli/Liza_Minnelli_0001.jpg
+Toby_Keith/Toby_Keith_0001.jpg
+Horace_Donovan_Reid/Horace_Donovan_Reid_0001.jpg
+David_Wells/David_Wells_0001.jpg
+Jose_Alencar/Jose_Alencar_0001.jpg
+Raul_Gonzalez/Raul_Gonzalez_0001.jpg
+Alex_Popov/Alex_Popov_0001.jpg
+Anne_Donovan/Anne_Donovan_0001.jpg
+Vin_Diesel/Vin_Diesel_0001.jpg
+Stuart_Whitman/Stuart_Whitman_0001.jpg
+Jesse_Jackson/Jesse_Jackson_0003.jpg
+Valorie_Brabazon/Valorie_Brabazon_0001.jpg
+Rob_Niedermayer/Rob_Niedermayer_0001.jpg
+Lynne_Cheney/Lynne_Cheney_0002.jpg
+Minnie_Mendoza/Minnie_Mendoza_0001.jpg
+Mark_Rosenbaum/Mark_Rosenbaum_0001.jpg
+Stephane_Rousseau/Stephane_Rousseau_0001.jpg
+Rita_Wilson/Rita_Wilson_0004.jpg
+Terence_Newman/Terence_Newman_0001.jpg
+Joseph_Biden/Joseph_Biden_0004.jpg
+Tommy_Tubberville/Tommy_Tubberville_0001.jpg
+Meles_Zenawi/Meles_Zenawi_0001.jpg
+Natasha_Henstridge/Natasha_Henstridge_0001.jpg
+Trevor_McDonald/Trevor_McDonald_0001.jpg
+Pat_Burns/Pat_Burns_0002.jpg
+Caroline_Link/Caroline_Link_0001.jpg
+Clay_Campbell/Clay_Campbell_0001.jpg
+Holly_Hunter/Holly_Hunter_0003.jpg
+Sophia_Loren/Sophia_Loren_0006.jpg
+Nicolas_Massu/Nicolas_Massu_0001.jpg
+Kaisser_Bazan/Kaisser_Bazan_0001.jpg
+Bryan_Murray/Bryan_Murray_0001.jpg
+Francois_Ozon/Francois_Ozon_0001.jpg
+Mathias_Reichhold/Mathias_Reichhold_0001.jpg
+Whoopi_Goldberg/Whoopi_Goldberg_0001.jpg
+Bill_Cartwright/Bill_Cartwright_0001.jpg
+Victor_Garber/Victor_Garber_0001.jpg
+Jeane_Kirkpatrick/Jeane_Kirkpatrick_0001.jpg
+Sue_Grafton/Sue_Grafton_0001.jpg
+Manuel_Jesus/Manuel_Jesus_0001.jpg
+Giovanny_Cordoba/Giovanny_Cordoba_0001.jpg
+James_Sensenbrenner/James_Sensenbrenner_0001.jpg
+Jennie_Finch/Jennie_Finch_0001.jpg
+David_Westerfield/David_Westerfield_0001.jpg
+Costas_Simitis/Costas_Simitis_0004.jpg
+Carlos_Iturgaitz/Carlos_Iturgaitz_0001.jpg
+Mark_Shapiro/Mark_Shapiro_0001.jpg
+George_Lopez/George_Lopez_0002.jpg
+Taylor_Twellman/Taylor_Twellman_0001.jpg
+Peter_Hartz/Peter_Hartz_0001.jpg
+Gianna_Angelopoulos-Daskalaki/Gianna_Angelopoulos-Daskalaki_0003.jpg
+Beth_Jones/Beth_Jones_0002.jpg
+Brock_Berlin/Brock_Berlin_0001.jpg
+James_McPherson/James_McPherson_0001.jpg
+Geoff_Hoon/Geoff_Hoon_0003.jpg
+Elizabeth_Hurley/Elizabeth_Hurley_0001.jpg
+Hipolito_Mejia/Hipolito_Mejia_0004.jpg
+Bart_Hendricks/Bart_Hendricks_0001.jpg
+David_Brown/David_Brown_0001.jpg
+Craig_Burley/Craig_Burley_0001.jpg
+Hermogenes_Ebdane_Jr/Hermogenes_Ebdane_Jr_0001.jpg
+Joey_Harrington/Joey_Harrington_0001.jpg
+Wolfgang_Schwarz/Wolfgang_Schwarz_0001.jpg
+Saoud_Al_Faisal/Saoud_Al_Faisal_0001.jpg
+Sebastian_Saja/Sebastian_Saja_0002.jpg
+Rick_Wagoner/Rick_Wagoner_0001.jpg
+Don_Henley/Don_Henley_0001.jpg
+John_Baldacci/John_Baldacci_0001.jpg
+Dustin_Brown/Dustin_Brown_0001.jpg
+Dominic_Monaghan/Dominic_Monaghan_0001.jpg
+Chris_Byrd/Chris_Byrd_0002.jpg
+Niall_Connolly/Niall_Connolly_0001.jpg
+Richard_Palmer/Richard_Palmer_0001.jpg
+Fazal-ur-Rehman/Fazal-ur-Rehman_0001.jpg
+Boris_Jordan/Boris_Jordan_0001.jpg
+Charlotte_Rampling/Charlotte_Rampling_0001.jpg
+Hutomo_Mandala_Putra/Hutomo_Mandala_Putra_0001.jpg
+Lars_Burgsmuller/Lars_Burgsmuller_0001.jpg
+Win_Aung/Win_Aung_0002.jpg
+Billy_Bob_Thornton/Billy_Bob_Thornton_0001.jpg
+Brett_Perry/Brett_Perry_0001.jpg
+Linda_Lingle/Linda_Lingle_0001.jpg
+Robert_Ehrlich/Robert_Ehrlich_0002.jpg
+Glenn_Rivers/Glenn_Rivers_0001.jpg
+David_Howard/David_Howard_0001.jpg
+Ian_Wilmut/Ian_Wilmut_0001.jpg
+Queen_Elizabeth_II/Queen_Elizabeth_II_0009.jpg
+Ingrid_Betancourt/Ingrid_Betancourt_0001.jpg
+Grady_Irvin_Jr/Grady_Irvin_Jr_0002.jpg
+Peter_Care/Peter_Care_0001.jpg
+Prince_Edward/Prince_Edward_0001.jpg
+Jennifer_Aniston/Jennifer_Aniston_0010.jpg
+Eric_Schacht/Eric_Schacht_0001.jpg
+Steve_Stirling/Steve_Stirling_0001.jpg
+Alberto_Acosta/Alberto_Acosta_0001.jpg
+Jim_Zorn/Jim_Zorn_0001.jpg
+Billy_Tibbets/Billy_Tibbets_0001.jpg
+Justin_Leonard/Justin_Leonard_0001.jpg
+Steve_Alford/Steve_Alford_0001.jpg
+Osmond_Smith/Osmond_Smith_0001.jpg
+Miranda_Gaddis/Miranda_Gaddis_0001.jpg
+Geovani_Lapentti/Geovani_Lapentti_0001.jpg
+Alma_Powell/Alma_Powell_0001.jpg
+Jimmy_Lee/Jimmy_Lee_0001.jpg
+Shobha_De/Shobha_De_0001.jpg
+Anil_Ramsook/Anil_Ramsook_0001.jpg
+Peter_Mullan/Peter_Mullan_0001.jpg
+Herman_Moore/Herman_Moore_0001.jpg
+Mark_Martin/Mark_Martin_0001.jpg
+Din_Samsudin/Din_Samsudin_0001.jpg
+Jason_Statham/Jason_Statham_0001.jpg
+Mike_Bair/Mike_Bair_0001.jpg
+Yana_Klochkova/Yana_Klochkova_0001.jpg
+Doug_Racine/Doug_Racine_0001.jpg
+Chan_Ho_Park/Chan_Ho_Park_0001.jpg
+Elena_Likhovtseva/Elena_Likhovtseva_0001.jpg
+Paul_Ebert/Paul_Ebert_0001.jpg
+Anita_DeFrantz/Anita_DeFrantz_0001.jpg
+Lee_Nam-shin/Lee_Nam-shin_0001.jpg
+Crispin_Glover/Crispin_Glover_0001.jpg
+Kevin_Marshall/Kevin_Marshall_0001.jpg
+Sven_Ottke/Sven_Ottke_0001.jpg
+Raf_Vallone/Raf_Vallone_0001.jpg
+Nick_Reilly/Nick_Reilly_0001.jpg
+Bob_Huggins/Bob_Huggins_0003.jpg
+David_Zeplowitz/David_Zeplowitz_0001.jpg
+Tayshaun_Prince/Tayshaun_Prince_0001.jpg
+Emmanuelle_Jagodsinski/Emmanuelle_Jagodsinski_0001.jpg
+Lyle_Vanclief/Lyle_Vanclief_0001.jpg
+Paula_Locke/Paula_Locke_0001.jpg
+James_Mathis/James_Mathis_0001.jpg
+Lou_Piniella/Lou_Piniella_0001.jpg
+Soenarno/Soenarno_0001.jpg
+Tara_Reid/Tara_Reid_0001.jpg
+Christian_Von_Wernich/Christian_Von_Wernich_0001.jpg
+Jacques_Rogge/Jacques_Rogge_0008.jpg
+Nawabzada_Nasrullah_Khan/Nawabzada_Nasrullah_Khan_0001.jpg
+Catherine_Donkers/Catherine_Donkers_0001.jpg
+Agnelo_Queiroz/Agnelo_Queiroz_0001.jpg
+Henry_Hilow/Henry_Hilow_0001.jpg
+Eric_Hinske/Eric_Hinske_0001.jpg
+Chris_Cornell/Chris_Cornell_0001.jpg
+Bonnie_Fuller/Bonnie_Fuller_0001.jpg
+Matt_Morris/Matt_Morris_0001.jpg
+Harbhajan_Singh/Harbhajan_Singh_0001.jpg
+Kim_Dong-hwa/Kim_Dong-hwa_0001.jpg
+Tim_Matheson/Tim_Matheson_0001.jpg
+Dick_Smothers/Dick_Smothers_0001.jpg
+Jia_Qinglin/Jia_Qinglin_0002.jpg
+Gerald_Calabrese/Gerald_Calabrese_0001.jpg
+Rowan_Williams/Rowan_Williams_0001.jpg
+Peter_Goldmark/Peter_Goldmark_0001.jpg
+Brian_Van_Dusen/Brian_Van_Dusen_0001.jpg
+Osama_bin_Laden/Osama_bin_Laden_0002.jpg
+Stanley_McChrystal/Stanley_McChrystal_0002.jpg
+Carlos_Ghosn/Carlos_Ghosn_0001.jpg
+Jean_Todt/Jean_Todt_0001.jpg
+Wally_Szczerbiak/Wally_Szczerbiak_0001.jpg
+Zara_Akhmadova/Zara_Akhmadova_0001.jpg
+Donald_Keck/Donald_Keck_0001.jpg
+Andy_Roddick/Andy_Roddick_0004.jpg
+John_Paul_II/John_Paul_II_0003.jpg
+Paul_Tracy/Paul_Tracy_0001.jpg
+Ismail_Khan/Ismail_Khan_0001.jpg
+Svetislav_Pesic/Svetislav_Pesic_0001.jpg
+Sven_Goran_Eriksson/Sven_Goran_Eriksson_0001.jpg
+Aline_Chretien/Aline_Chretien_0001.jpg
+Jeffery_Hendren/Jeffery_Hendren_0001.jpg
+Ian_Huntley/Ian_Huntley_0001.jpg
+Andrew_Luster/Andrew_Luster_0001.jpg
+Emilio_Azcarraga_Jean/Emilio_Azcarraga_Jean_0001.jpg
+Wilbert_Foy/Wilbert_Foy_0001.jpg
+Pham_Thi_Mai_Phuong/Pham_Thi_Mai_Phuong_0001.jpg
+William_McDonough/William_McDonough_0001.jpg
+Linda_Franklin/Linda_Franklin_0001.jpg
+Liane_Janda/Liane_Janda_0001.jpg
+Jana_Pittman/Jana_Pittman_0001.jpg
+Leuris_Pupo/Leuris_Pupo_0001.jpg
+Martin_Sheen/Martin_Sheen_0001.jpg
+Lisa_Girman/Lisa_Girman_0001.jpg
+Brian_Clemens/Brian_Clemens_0001.jpg
+Larry_Donald/Larry_Donald_0001.jpg
+Colleen_Ryan/Colleen_Ryan_0001.jpg
+Max_Baucus/Max_Baucus_0001.jpg
+Bill_Sizemore/Bill_Sizemore_0001.jpg
+Sandra_Day_OConner/Sandra_Day_OConner_0001.jpg
+Mirela_Manjani/Mirela_Manjani_0001.jpg
+Estella_Warren/Estella_Warren_0001.jpg
+Zoran_Djindjic/Zoran_Djindjic_0001.jpg
+Nate_Hybl/Nate_Hybl_0001.jpg
+Ed_Case/Ed_Case_0001.jpg
+Stephen_Funk/Stephen_Funk_0001.jpg
+John_McEnroe/John_McEnroe_0002.jpg
+Walter_Mondale/Walter_Mondale_0008.jpg
+Diana_Krall/Diana_Krall_0001.jpg
+Heather_Willson/Heather_Willson_0001.jpg
+Felipe_De_Borbon/Felipe_De_Borbon_0001.jpg
+Carolina_Moraes/Carolina_Moraes_0002.jpg
+Ali_Mohammed_Maher/Ali_Mohammed_Maher_0001.jpg
+Rick_Husband/Rick_Husband_0001.jpg
+Stephanie_Zimbalist/Stephanie_Zimbalist_0001.jpg
+Shanna_Zolman/Shanna_Zolman_0001.jpg
+Isaiah_Washington/Isaiah_Washington_0002.jpg
+Edgar_Savisaar/Edgar_Savisaar_0001.jpg
+Dave_Tucker/Dave_Tucker_0001.jpg
+Brandon_Larson/Brandon_Larson_0001.jpg
+Phoenix_Chang/Phoenix_Chang_0001.jpg
+Penelope_Ann_Miller/Penelope_Ann_Miller_0002.jpg
+Natalie_Juniardi/Natalie_Juniardi_0001.jpg
+Ted_Maher/Ted_Maher_0002.jpg
+Noor_Mohammed/Noor_Mohammed_0001.jpg
+Walt_Harris/Walt_Harris_0001.jpg
+Olivera_Labus/Olivera_Labus_0001.jpg
+Stephen_Covey/Stephen_Covey_0001.jpg
+Woodrow_Stanley/Woodrow_Stanley_0001.jpg
+Yuvraj_Singh/Yuvraj_Singh_0001.jpg
+Bobby_Bowden/Bobby_Bowden_0001.jpg
+Hussein_Malik/Hussein_Malik_0001.jpg
+Tom_Hanks/Tom_Hanks_0009.jpg
+Gary_Gitnick/Gary_Gitnick_0001.jpg
+Mike_Richter/Mike_Richter_0001.jpg
+Luke_Walton/Luke_Walton_0001.jpg
+Denys_Arcand/Denys_Arcand_0001.jpg
+Bill_Butler/Bill_Butler_0001.jpg
+Michael_Goldrich/Michael_Goldrich_0001.jpg
+Yekaterina_Guseva/Yekaterina_Guseva_0001.jpg
+Darren_Clarke/Darren_Clarke_0001.jpg
+Intisar_Ajouri/Intisar_Ajouri_0002.jpg
+Cristiano_da_Matta/Cristiano_da_Matta_0001.jpg
+Hugh_Carey/Hugh_Carey_0001.jpg
+Danny_Glover/Danny_Glover_0001.jpg
+Peter_Hunt/Peter_Hunt_0001.jpg
+Steven_Feldman/Steven_Feldman_0001.jpg
+Dinah_Turner/Dinah_Turner_0001.jpg
+Rosny_Desroches/Rosny_Desroches_0001.jpg
+Cosmo_Iacavazzi/Cosmo_Iacavazzi_0001.jpg
+Daniel_Day-Lewis/Daniel_Day-Lewis_0001.jpg
+Robert_Wiener/Robert_Wiener_0001.jpg
+Derrick_Battie/Derrick_Battie_0001.jpg
+Candice_Beatty/Candice_Beatty_0001.jpg
+Gerald_Riley/Gerald_Riley_0001.jpg
+Julie_Andrews/Julie_Andrews_0001.jpg
+Jean-Marc_de_La_Sabliere/Jean-Marc_de_La_Sabliere_0002.jpg
+Zhang_Ziyi/Zhang_Ziyi_0004.jpg
+Kurt_Busch/Kurt_Busch_0002.jpg
+Roberto_Lavagna/Roberto_Lavagna_0001.jpg
+Dick_Cheney/Dick_Cheney_0009.jpg
+Hamza_Atiya_Muhsen/Hamza_Atiya_Muhsen_0001.jpg
+David_Hilt/David_Hilt_0001.jpg
+Susan_Walvius/Susan_Walvius_0001.jpg
+Kim_Weeks/Kim_Weeks_0001.jpg
+Darryl_Stingley/Darryl_Stingley_0001.jpg
+Colin_Prescot/Colin_Prescot_0001.jpg
+Felix_Mantilla/Felix_Mantilla_0001.jpg
+Frank_Van_Ecke/Frank_Van_Ecke_0001.jpg
+Mike_Sherman/Mike_Sherman_0001.jpg
+David_Myers/David_Myers_0002.jpg
+Guillermo_Ruiz_Polanco/Guillermo_Ruiz_Polanco_0001.jpg
+Ken_Loach/Ken_Loach_0001.jpg
+Robert_Flodquist/Robert_Flodquist_0001.jpg
+Ben_Lee/Ben_Lee_0001.jpg
+Aung_San_Suu_Kyi/Aung_San_Suu_Kyi_0001.jpg
+Keith_Brown/Keith_Brown_0001.jpg
+Thomas_Gottschalk/Thomas_Gottschalk_0001.jpg
+Shae-Lynn_Bourne/Shae-Lynn_Bourne_0001.jpg
+Bob_Sulkin/Bob_Sulkin_0001.jpg
+Chakib_Khelil/Chakib_Khelil_0001.jpg
+Robert_Wagner/Robert_Wagner_0001.jpg
+William_Cocksedge/William_Cocksedge_0001.jpg
+Giselle_Estefania_Tavarelli/Giselle_Estefania_Tavarelli_0001.jpg
+Lorne_Michaels/Lorne_Michaels_0001.jpg
+Ricky_Barnes/Ricky_Barnes_0002.jpg
+Laura_Ziskin/Laura_Ziskin_0001.jpg
+Ray_Evernham/Ray_Evernham_0001.jpg
+Pedro_Velasquez/Pedro_Velasquez_0001.jpg
+Marion_Fahnestock/Marion_Fahnestock_0001.jpg
+Jane_Pauley/Jane_Pauley_0002.jpg
+Bernardo_Segura/Bernardo_Segura_0002.jpg
+Padraig_Harrington/Padraig_Harrington_0001.jpg
+Luke_Smith/Luke_Smith_0001.jpg
+Jeffrey_Pfeffer/Jeffrey_Pfeffer_0001.jpg
+Brandon_Boyd/Brandon_Boyd_0001.jpg
+Jim_Jeffords/Jim_Jeffords_0001.jpg
+Claudio_Abbado/Claudio_Abbado_0001.jpg
+Chen_Liang_Yu/Chen_Liang_Yu_0001.jpg
+Maura_Tierney/Maura_Tierney_0001.jpg
+John_Danforth/John_Danforth_0001.jpg
+Cheryl_Hines/Cheryl_Hines_0001.jpg
+Timothy_Goebel/Timothy_Goebel_0001.jpg
+Frank_Sinatra/Frank_Sinatra_0001.jpg
+Marissa_Jaret_Winokur/Marissa_Jaret_Winokur_0001.jpg
+Masamori_Tokuyama/Masamori_Tokuyama_0001.jpg
+Bill_Rainer/Bill_Rainer_0001.jpg
+Ken_Watanabe/Ken_Watanabe_0002.jpg
+Kimora_Lee/Kimora_Lee_0001.jpg
+Sanjay_Gupta/Sanjay_Gupta_0001.jpg
+Gary_Locke/Gary_Locke_0001.jpg
+Sinead_OConnor/Sinead_OConnor_0001.jpg
+James_W_Kennedy/James_W_Kennedy_0001.jpg
+Kurt_Suzuki/Kurt_Suzuki_0001.jpg
+Frederick_Madden/Frederick_Madden_0001.jpg
+Priscilla_Owen/Priscilla_Owen_0002.jpg
+Malcolm_Jamal_Warner/Malcolm_Jamal_Warner_0001.jpg
+Naoto_Kan/Naoto_Kan_0003.jpg
+Sue_Guevara/Sue_Guevara_0001.jpg
+John_Sweeney/John_Sweeney_0001.jpg
+Brian_Cashman/Brian_Cashman_0001.jpg
+Marco_Pantani/Marco_Pantani_0001.jpg
+Theodore_Tweed_Roosevelt/Theodore_Tweed_Roosevelt_0002.jpg
+Se_Hyuk_Joo/Se_Hyuk_Joo_0001.jpg
+Ian_Moran/Ian_Moran_0001.jpg
+Michael_Olowokandi/Michael_Olowokandi_0001.jpg
+James_Brosnahan/James_Brosnahan_0001.jpg
+Kevin_Nealon/Kevin_Nealon_0001.jpg
+Miguel_Estrada/Miguel_Estrada_0001.jpg
+Alison_Krauss/Alison_Krauss_0001.jpg
+Bill_Callahan/Bill_Callahan_0001.jpg
+Allison_Searing/Allison_Searing_0001.jpg
+Torri_Edwards/Torri_Edwards_0002.jpg
+Bryan_Thomas/Bryan_Thomas_0001.jpg
+Kathy_Baker/Kathy_Baker_0001.jpg
+Martha_Stewart/Martha_Stewart_0005.jpg
+William_Hurt/William_Hurt_0001.jpg
+Serge_Tchuruk/Serge_Tchuruk_0001.jpg
+Michael_Brandon/Michael_Brandon_0001.jpg
+Adam_Sandler/Adam_Sandler_0001.jpg
+Sam_Gerald/Sam_Gerald_0001.jpg
+Laura_Linney/Laura_Linney_0001.jpg
+Aitor_Gonzalez/Aitor_Gonzalez_0002.jpg
+Hamzah_Haz/Hamzah_Haz_0001.jpg
+James_Young/James_Young_0001.jpg
+Misty_Dawn_Clymer/Misty_Dawn_Clymer_0001.jpg
+Bronson_Arroyo/Bronson_Arroyo_0001.jpg
+Curtis_Rodriguez/Curtis_Rodriguez_0001.jpg
+Paul_Burrell/Paul_Burrell_0008.jpg
+Jennie_Garth/Jennie_Garth_0001.jpg
+Irina_Yatchenko/Irina_Yatchenko_0001.jpg
+Don_Matthews/Don_Matthews_0001.jpg
+David_Eldon/David_Eldon_0001.jpg
+Christian_Wulff/Christian_Wulff_0001.jpg
+Paul_Byrd/Paul_Byrd_0001.jpg
+Derek_Bond/Derek_Bond_0001.jpg
+Wolfgang_Clement/Wolfgang_Clement_0001.jpg
+Carol_Williams/Carol_Williams_0001.jpg
+John_Goold/John_Goold_0001.jpg
+Mesut_Yilmaz/Mesut_Yilmaz_0002.jpg
+John_Dallager/John_Dallager_0001.jpg
+Cecile_de_France/Cecile_de_France_0001.jpg
+Adrian_Fernandez/Adrian_Fernandez_0001.jpg
+Alan_Greer/Alan_Greer_0001.jpg
+Mike_Babcock/Mike_Babcock_0001.jpg
+Casy_Preslar/Casy_Preslar_0001.jpg
+Giuseppe_Gibilisco/Giuseppe_Gibilisco_0003.jpg
+Wesley_Clark/Wesley_Clark_0002.jpg
+Tom_Osborne/Tom_Osborne_0001.jpg
+Robert_Mueller/Robert_Mueller_0002.jpg
+Keith_Urban/Keith_Urban_0001.jpg
+Steve_Cox/Steve_Cox_0001.jpg
+Carly_Fiorina/Carly_Fiorina_0001.jpg
+Irv_Nathan/Irv_Nathan_0001.jpg
+Goh_Kun/Goh_Kun_0002.jpg
+Imam_Samudra/Imam_Samudra_0001.jpg
+Leon_Barmore/Leon_Barmore_0001.jpg
+Martha_Beatriz_Roque/Martha_Beatriz_Roque_0001.jpg
+Lope_Mendoza/Lope_Mendoza_0001.jpg
+Roberto_Benigni/Roberto_Benigni_0002.jpg
+Vincent_Sombrotto/Vincent_Sombrotto_0001.jpg
+Scott_Rudin/Scott_Rudin_0001.jpg
+Jesus_Cardenal/Jesus_Cardenal_0001.jpg
+Kimi_Raikkonen/Kimi_Raikkonen_0001.jpg
+Warren_Beatty/Warren_Beatty_0002.jpg
+Sherri_Coale/Sherri_Coale_0001.jpg
+Matt_Roney/Matt_Roney_0001.jpg
+Michelle_Branch/Michelle_Branch_0002.jpg
+Teresa_Graves/Teresa_Graves_0001.jpg
+Michel_Temer/Michel_Temer_0002.jpg
+Benazir_Bhutto/Benazir_Bhutto_0001.jpg
+Uday_Hussein/Uday_Hussein_0001.jpg
+Brad_Russ/Brad_Russ_0001.jpg
+Gracia_Burnham/Gracia_Burnham_0001.jpg
+James_Watt/James_Watt_0001.jpg
+Reggie_Sanders/Reggie_Sanders_0001.jpg
+Orrin_Hatch/Orrin_Hatch_0001.jpg
+Howard_Schultz/Howard_Schultz_0002.jpg
+Piers_Sellers/Piers_Sellers_0001.jpg
+Doug_Christie/Doug_Christie_0001.jpg
+Larry_Harris/Larry_Harris_0001.jpg
+Lin_Yung_Hsi/Lin_Yung_Hsi_0001.jpg
+Abdullah_al-Attiyah/Abdullah_al-Attiyah_0001.jpg
+Vladimir_Ustinov/Vladimir_Ustinov_0001.jpg
+Dustan_Mohr/Dustan_Mohr_0001.jpg
+Farouk_Kaddoumi/Farouk_Kaddoumi_0001.jpg
+Shigeru_Ishiba/Shigeru_Ishiba_0001.jpg
+Peter_Sejna/Peter_Sejna_0001.jpg
+Robert_Tyrrell/Robert_Tyrrell_0001.jpg
+Beyonce_Knowles/Beyonce_Knowles_0001.jpg
+Brandon_Inge/Brandon_Inge_0001.jpg
+Ignatius_Wang/Ignatius_Wang_0001.jpg
+Patsy_Kensit/Patsy_Kensit_0001.jpg
+Keith_Lockhart/Keith_Lockhart_0001.jpg
+Spencer_Abraham/Spencer_Abraham_0009.jpg
+Ronde_Barber/Ronde_Barber_0001.jpg
+Antonio_Palocci/Antonio_Palocci_0001.jpg
+Desmon_Farmer/Desmon_Farmer_0001.jpg
+Dieter_Holzer/Dieter_Holzer_0001.jpg
+Stacey_Yamaguchi/Stacey_Yamaguchi_0001.jpg
+Viara_Vike-Freiberga/Viara_Vike-Freiberga_0001.jpg
+Lokendra_Bahadur_Chand/Lokendra_Bahadur_Chand_0001.jpg
+Ivan_Shvedoff/Ivan_Shvedoff_0001.jpg
+Dana_Vollmer/Dana_Vollmer_0001.jpg
+Bob_Curtis/Bob_Curtis_0001.jpg
+John_Burnett/John_Burnett_0001.jpg
+Thomas_Cloyd/Thomas_Cloyd_0001.jpg
+Katrin_Cartlidge/Katrin_Cartlidge_0001.jpg
+Horst_Koehler/Horst_Koehler_0001.jpg
+Julian_Battle/Julian_Battle_0001.jpg
+Billy_Crawford/Billy_Crawford_0001.jpg
+Carolyn_Kuhl/Carolyn_Kuhl_0001.jpg
+Marc_Gold/Marc_Gold_0001.jpg
+Jerry_Pauley/Jerry_Pauley_0001.jpg
+Li_Ka-shing/Li_Ka-shing_0001.jpg
+Vaclav_Havel/Vaclav_Havel_0008.jpg
+Jorge_Quiroga/Jorge_Quiroga_0001.jpg
+Manijeh_Hekmat/Manijeh_Hekmat_0001.jpg
+Graciano_Rocchigiani/Graciano_Rocchigiani_0001.jpg
+Marquis_Estill/Marquis_Estill_0001.jpg
+Clive_Lloyd/Clive_Lloyd_0001.jpg
+Sandra_Bullock/Sandra_Bullock_0004.jpg
+Buck_Rodgers/Buck_Rodgers_0001.jpg
+Natalie_Stewart/Natalie_Stewart_0001.jpg
+Adisai_Bodharamik/Adisai_Bodharamik_0001.jpg
+Jim_Flaherty/Jim_Flaherty_0001.jpg
+Erik_Morales/Erik_Morales_0001.jpg
+Cynthia_Rowley/Cynthia_Rowley_0001.jpg
+Stanislas_Wawrinka/Stanislas_Wawrinka_0001.jpg
+Brendan_Fraser/Brendan_Fraser_0001.jpg
+Carson_Daly/Carson_Daly_0002.jpg
+Natalia_Dmitrieva/Natalia_Dmitrieva_0001.jpg
+Sarah_Jessica_Parker/Sarah_Jessica_Parker_0004.jpg
+Fidel_Castro/Fidel_Castro_0014.jpg
+Terry_Semel/Terry_Semel_0001.jpg
+Trista_Rehn/Trista_Rehn_0001.jpg
+Carolina_Kluft/Carolina_Kluft_0001.jpg
+Katie_Couric/Katie_Couric_0001.jpg
+Carlos_Beltran/Carlos_Beltran_0001.jpg
+Peter_Shaw/Peter_Shaw_0001.jpg
+Shinzo_Abe/Shinzo_Abe_0001.jpg
+Candace_Sutton/Candace_Sutton_0001.jpg
+Dave_Matthews/Dave_Matthews_0001.jpg
+Bob_Cantrell/Bob_Cantrell_0001.jpg
+Danielle_Spencer/Danielle_Spencer_0001.jpg
+Kiki_Vandeweghe/Kiki_Vandeweghe_0001.jpg
+Scott_Hubbard/Scott_Hubbard_0001.jpg
+Michael_Frayn/Michael_Frayn_0001.jpg
+James_Robertson_Jr/James_Robertson_Jr_0001.jpg
+Kieran_Prendergast/Kieran_Prendergast_0001.jpg
+Patti_Balgojevich/Patti_Balgojevich_0001.jpg
+Scott_Wittman/Scott_Wittman_0001.jpg
+Richard_Lennon/Richard_Lennon_0001.jpg
+Nestor_Gonzalez/Nestor_Gonzalez_0001.jpg
+Tom_McClintock/Tom_McClintock_0001.jpg
+Iain_Anderson/Iain_Anderson_0001.jpg
+Michael_Powell/Michael_Powell_0004.jpg
+Lela_Rochon/Lela_Rochon_0001.jpg
+Shane_Warne/Shane_Warne_0002.jpg
+Lauren_Killian/Lauren_Killian_0002.jpg
+Perry_Gibbs/Perry_Gibbs_0001.jpg
+Curt_Weldon/Curt_Weldon_0001.jpg
+Tony_Elias/Tony_Elias_0001.jpg
+Bill_McBride/Bill_McBride_0009.jpg
+Nick_Markakis/Nick_Markakis_0001.jpg
+Mikhail_Kalashnikov/Mikhail_Kalashnikov_0001.jpg
+Jonathan_Karsh/Jonathan_Karsh_0001.jpg
+Igor_Trunov/Igor_Trunov_0001.jpg
+Renee_Zellweger/Renee_Zellweger_0007.jpg
+Joaquin_Phoenix/Joaquin_Phoenix_0001.jpg
+Magda_Kertasz/Magda_Kertasz_0001.jpg
+Mel_Brooks/Mel_Brooks_0001.jpg
+Todd_MacCulloch/Todd_MacCulloch_0001.jpg
+Keiko_Sofia_Fujimori/Keiko_Sofia_Fujimori_0001.jpg
+Jeffrey_Archer/Jeffrey_Archer_0001.jpg
+Richard_Dreyfuss/Richard_Dreyfuss_0001.jpg
+Frank_Shea/Frank_Shea_0001.jpg
+Dan_Monson/Dan_Monson_0001.jpg
+Ruth_Stubbs/Ruth_Stubbs_0001.jpg
+James_McMahon/James_McMahon_0001.jpg
+Lennox_Lewis/Lennox_Lewis_0002.jpg
+Tian_Liang/Tian_Liang_0001.jpg
+Timbul_Silaen/Timbul_Silaen_0001.jpg
+BJ_Habibie/BJ_Habibie_0001.jpg
+George_Gregan/George_Gregan_0001.jpg
+Tom_Curley/Tom_Curley_0001.jpg
+Justin_Timberlake/Justin_Timberlake_0004.jpg
+Frank_Wycheck/Frank_Wycheck_0001.jpg
+Nicola_Bono/Nicola_Bono_0001.jpg
+Nikolay_Davydenko/Nikolay_Davydenko_0001.jpg
+John_Petty/John_Petty_0001.jpg
+Paul_Hogan/Paul_Hogan_0002.jpg
+Paul_Cerjan/Paul_Cerjan_0001.jpg
+Ali_Ahmeti/Ali_Ahmeti_0001.jpg
+Rachel_Griffiths/Rachel_Griffiths_0001.jpg
+Nate_Blackwell/Nate_Blackwell_0001.jpg
+Yves_Brodeur/Yves_Brodeur_0001.jpg
+Yory_Boy_Campas/Yory_Boy_Campas_0001.jpg
+Patricia_Clarkson/Patricia_Clarkson_0004.jpg
+Jan_Petersen/Jan_Petersen_0001.jpg
+Tatiana_Shchegoleva/Tatiana_Shchegoleva_0001.jpg
+George_Foreman/George_Foreman_0002.jpg
+Peter_Chan/Peter_Chan_0001.jpg
+Baburam_Bhattari/Baburam_Bhattari_0001.jpg
+Stacy_Dragila/Stacy_Dragila_0002.jpg
+George_Tenet/George_Tenet_0001.jpg
+Sam_Mendes/Sam_Mendes_0001.jpg
+Humberto_Espinoza/Humberto_Espinoza_0001.jpg
+Alexa_Vega/Alexa_Vega_0001.jpg
+Vaclav_Klaus/Vaclav_Klaus_0002.jpg
+Judy_Vassar/Judy_Vassar_0001.jpg
+Aram_Adler/Aram_Adler_0001.jpg
+Andy_Hebb/Andy_Hebb_0001.jpg
+Sheldon_Silver/Sheldon_Silver_0001.jpg
+Ilham_Aliev/Ilham_Aliev_0001.jpg
+Maria_Soledad_Alvear_Valenzuela/Maria_Soledad_Alvear_Valenzuela_0003.jpg
+Luis_Horna/Luis_Horna_0004.jpg
+Tom_Daschle/Tom_Daschle_0001.jpg
+Kevin_Keegan/Kevin_Keegan_0001.jpg
+Taia_Balk/Taia_Balk_0001.jpg
+Kwon_Young-gil/Kwon_Young-gil_0001.jpg
+Tab_Baldwin/Tab_Baldwin_0001.jpg
+Ramon_Cardenas/Ramon_Cardenas_0001.jpg
+Mary_Tyler_Moore/Mary_Tyler_Moore_0001.jpg
+Audrey_Sauret/Audrey_Sauret_0001.jpg
+Bill_Frist/Bill_Frist_0008.jpg
+Scott_Wallach/Scott_Wallach_0001.jpg
+Larry_Wilmore/Larry_Wilmore_0001.jpg
+Rosa_Haywa_de_Condori/Rosa_Haywa_de_Condori_0001.jpg
+Clara_Harris/Clara_Harris_0003.jpg
+Alvaro_Uribe/Alvaro_Uribe_0022.jpg
+Mark_Philippoussis/Mark_Philippoussis_0009.jpg
+Boris_Berezovsky/Boris_Berezovsky_0001.jpg
+Petro_Symonenko/Petro_Symonenko_0001.jpg
+Tammy_Lynn_Michaels/Tammy_Lynn_Michaels_0001.jpg
+Angelina_Jolie/Angelina_Jolie_0019.jpg
+Juanes/Juanes_0001.jpg
+BB_King/BB_King_0001.jpg
+Wayne_Newton/Wayne_Newton_0001.jpg
+JT_Snow/JT_Snow_0001.jpg
+Stefano_Accorsi/Stefano_Accorsi_0001.jpg
+Francis_Mer/Francis_Mer_0002.jpg
+William_Donaldson/William_Donaldson_0004.jpg
+Monica_Gabrielle/Monica_Gabrielle_0001.jpg
+Woody_Allen/Woody_Allen_0004.jpg
+Will_Ofenheusle/Will_Ofenheusle_0001.jpg
+Jim_Beattie/Jim_Beattie_0001.jpg
+Mitchell_Potter/Mitchell_Potter_0001.jpg
+Stacey_Jones/Stacey_Jones_0001.jpg
+Debra_Messing/Debra_Messing_0001.jpg
+Dan_Boyle/Dan_Boyle_0001.jpg
+Lisa_Murkowski/Lisa_Murkowski_0001.jpg
+Damarius_Bilbo/Damarius_Bilbo_0001.jpg
+Steny_Hoyer/Steny_Hoyer_0001.jpg
+Luciano_Pavarotti/Luciano_Pavarotti_0001.jpg
+Jason_Petty/Jason_Petty_0001.jpg
+Brian_Pavlich/Brian_Pavlich_0001.jpg
+Ismail_Cem/Ismail_Cem_0001.jpg
+Ana_Claudia_Talancon/Ana_Claudia_Talancon_0001.jpg
+Charlene_Barshefsky/Charlene_Barshefsky_0001.jpg
+Mitchell_Swartz/Mitchell_Swartz_0001.jpg
+Steve_Cutler/Steve_Cutler_0001.jpg
+Venus_Williams/Venus_Williams_0009.jpg
+Kwame_Kilpatrick/Kwame_Kilpatrick_0001.jpg
+Eddie_Sutton/Eddie_Sutton_0002.jpg
+Enrique_Oliu/Enrique_Oliu_0001.jpg
+Tom_Sizemore/Tom_Sizemore_0001.jpg
+Vincent_Brooks/Vincent_Brooks_0007.jpg
+Ricky_Quick/Ricky_Quick_0001.jpg
+Slobodan_Milosevic/Slobodan_Milosevic_0001.jpg
+Mary_Landrieu/Mary_Landrieu_0001.jpg
+Christopher_Reeve/Christopher_Reeve_0002.jpg
+Daryl_Hannah/Daryl_Hannah_0001.jpg
+Adel_Al-Jubeir/Adel_Al-Jubeir_0001.jpg
+Mario_Austin/Mario_Austin_0001.jpg
+Pak_Gil_Yon/Pak_Gil_Yon_0001.jpg
+Kimberly_Stewart/Kimberly_Stewart_0001.jpg
+Robert_Hanssen/Robert_Hanssen_0001.jpg
+Mary_Matalin/Mary_Matalin_0001.jpg
+Mary_Hill/Mary_Hill_0001.jpg
+Cesar_Gaviria/Cesar_Gaviria_0001.jpg
+Michael_Kostelnik/Michael_Kostelnik_0002.jpg
+Romain_Duris/Romain_Duris_0001.jpg
+Mohammad_Khatami/Mohammad_Khatami_0002.jpg
+Peter_Caruana/Peter_Caruana_0001.jpg
+Lynn_Abraham/Lynn_Abraham_0002.jpg
+Kelly_Osbourne/Kelly_Osbourne_0001.jpg
+Condoleezza_Rice/Condoleezza_Rice_0007.jpg
+Julia_Tymoshenko/Julia_Tymoshenko_0003.jpg
+Eminem/Eminem_0001.jpg
+Tom_Watson/Tom_Watson_0003.jpg
+Thomas_Klestil/Thomas_Klestil_0001.jpg
+Caio_Blat/Caio_Blat_0001.jpg
+Bob_Ferguson/Bob_Ferguson_0001.jpg
+Randy_Jackson/Randy_Jackson_0001.jpg
+Eunice_Barber/Eunice_Barber_0002.jpg
+Ozzy_Osbourne/Ozzy_Osbourne_0001.jpg
+William_Jackson/William_Jackson_0001.jpg
+Osama_Al_Baz/Osama_Al_Baz_0001.jpg
+Gus_Van_Sant/Gus_Van_Sant_0001.jpg
+Bryce_Carmine/Bryce_Carmine_0001.jpg
+Edward_Johnson/Edward_Johnson_0001.jpg
+Stephen_Cooper/Stephen_Cooper_0001.jpg
+Richard_Harris/Richard_Harris_0001.jpg
+Sadie_Frost/Sadie_Frost_0001.jpg
+Rod_Jong-il/Rod_Jong-il_0001.jpg
+Marco_Irizarry/Marco_Irizarry_0001.jpg
+Bulent_Ecevit/Bulent_Ecevit_0006.jpg
+Rick_Dinse/Rick_Dinse_0001.jpg
+Hubie_Brown/Hubie_Brown_0001.jpg
+Javier_Camara/Javier_Camara_0001.jpg
+Kurt_Warner/Kurt_Warner_0001.jpg
+Rick_Pitino/Rick_Pitino_0001.jpg
+Toni_Jennings/Toni_Jennings_0001.jpg
+Branko_Crvenkovski/Branko_Crvenkovski_0002.jpg
+Chip_Ganassi/Chip_Ganassi_0001.jpg
+Jason_Mewes/Jason_Mewes_0001.jpg
+Zalmay_Khalilzad/Zalmay_Khalilzad_0001.jpg
+Ken_Macha/Ken_Macha_0001.jpg
+Ali_Adbul_Karim_Madani/Ali_Adbul_Karim_Madani_0001.jpg
+Raul_Chacon/Raul_Chacon_0001.jpg
+Luis_Guzman/Luis_Guzman_0001.jpg
+Rena_Sofer/Rena_Sofer_0001.jpg
+Vince_Dooley/Vince_Dooley_0001.jpg
+Janet_Ecker/Janet_Ecker_0001.jpg
+Carlos_Fasciolo/Carlos_Fasciolo_0001.jpg
+Massoud_Barzani/Massoud_Barzani_0001.jpg
+Candice_Bergen/Candice_Bergen_0001.jpg
+Ralph_Firman/Ralph_Firman_0002.jpg
+Rick_Santorum/Rick_Santorum_0001.jpg
+Javier_Bardem/Javier_Bardem_0001.jpg
+Luis_Rosario_Huertas/Luis_Rosario_Huertas_0001.jpg
+Sandra_Milo/Sandra_Milo_0001.jpg
+Julius_Barnes/Julius_Barnes_0001.jpg
+Taha_Yassin_Ramadan/Taha_Yassin_Ramadan_0009.jpg
+El_Hadji_Diouf/El_Hadji_Diouf_0001.jpg
+Felix_Doh/Felix_Doh_0001.jpg
+John_Stallworth/John_Stallworth_0001.jpg
+Francis_Ford_Coppola/Francis_Ford_Coppola_0001.jpg
+Peter_Hillary/Peter_Hillary_0001.jpg
+Kristen_Breitweiser/Kristen_Breitweiser_0001.jpg
+Gregor_Gysi/Gregor_Gysi_0001.jpg
+Eglis_Yaima_Cruz/Eglis_Yaima_Cruz_0001.jpg
+Sandra_Shamas/Sandra_Shamas_0001.jpg
+Helen_Clark/Helen_Clark_0002.jpg
+Ted_Christopher/Ted_Christopher_0001.jpg
+Takeo_Hiranuma/Takeo_Hiranuma_0001.jpg
+Marcos_Milinkovic/Marcos_Milinkovic_0001.jpg
+Jim_Carrey/Jim_Carrey_0001.jpg
+Steve-O/Steve-O_0001.jpg
+Edward_James_Olmos/Edward_James_Olmos_0001.jpg
+John_McCain/John_McCain_0006.jpg
+Gustavo_Noboa/Gustavo_Noboa_0001.jpg
+Ms_Dynamite/Ms_Dynamite_0001.jpg
+Chuck_Yeager/Chuck_Yeager_0002.jpg
+Steve_Backley/Steve_Backley_0002.jpg
+Jacqueline_Edwards/Jacqueline_Edwards_0001.jpg
+Lena_Olin/Lena_Olin_0001.jpg
+Kaspar_Villiger/Kaspar_Villiger_0001.jpg
+Edwina_Currie/Edwina_Currie_0004.jpg
+Mark_Hogan/Mark_Hogan_0001.jpg
+Ferenc_Madl/Ferenc_Madl_0002.jpg
+Jake_Brace/Jake_Brace_0001.jpg
+Gavyn_Davies/Gavyn_Davies_0001.jpg
+Janine_Pietsch/Janine_Pietsch_0001.jpg
+Dagmar_Dunlevy/Dagmar_Dunlevy_0001.jpg
+Zico/Zico_0002.jpg
+Tatiana_Panova/Tatiana_Panova_0001.jpg
+James_Spalding/James_Spalding_0001.jpg
+Asif_Hanif/Asif_Hanif_0001.jpg
+Larry_Ellison/Larry_Ellison_0001.jpg
+Marc-Andre_Fleury/Marc-Andre_Fleury_0001.jpg
+Andrew_Weissmann/Andrew_Weissmann_0001.jpg
+Jose_Vicente_Rangel/Jose_Vicente_Rangel_0001.jpg
+Marc_Bulger/Marc_Bulger_0001.jpg
+Grant_Hackett/Grant_Hackett_0005.jpg
+Victor_Kraatz/Victor_Kraatz_0001.jpg
+David_Chase/David_Chase_0001.jpg
+Javier_Weber/Javier_Weber_0001.jpg
+Marina_Hands/Marina_Hands_0001.jpg
+Rogelio_Ramos/Rogelio_Ramos_0001.jpg
+Khalid_Qazi/Khalid_Qazi_0001.jpg
+Andres_Manuel_Lopez_Obrador/Andres_Manuel_Lopez_Obrador_0001.jpg
+Assad_Ahmadi/Assad_Ahmadi_0001.jpg
+Bill_Maher/Bill_Maher_0001.jpg
+Hitoshi_Oshitani/Hitoshi_Oshitani_0001.jpg
+Queen_Silvia/Queen_Silvia_0001.jpg
+Julio_Cesar_Franco/Julio_Cesar_Franco_0001.jpg
+Richard_Naughton/Richard_Naughton_0001.jpg
+Paddy_Long/Paddy_Long_0001.jpg
+Mikhail_Youzhny/Mikhail_Youzhny_0001.jpg
+David_Carradine/David_Carradine_0001.jpg
+Peter_Struck/Peter_Struck_0001.jpg
+Terry_Lynn_Barton/Terry_Lynn_Barton_0001.jpg
+Gerald_Fitch/Gerald_Fitch_0001.jpg
+Cemil_Cicek/Cemil_Cicek_0001.jpg
+Frank_Lautenberg/Frank_Lautenberg_0002.jpg
+Georgia_Giddings/Georgia_Giddings_0001.jpg
+Glafcos_Clerides/Glafcos_Clerides_0003.jpg
+Filippo_Inzaghi/Filippo_Inzaghi_0003.jpg
+Edina_Batar/Edina_Batar_0001.jpg
+Malak_Habbak/Malak_Habbak_0001.jpg
+David_McCallum/David_McCallum_0001.jpg
+Garry_Alejano/Garry_Alejano_0001.jpg
+Peter_Medgyessy/Peter_Medgyessy_0001.jpg
+JJ_Redick/JJ_Redick_0001.jpg
+Phil_Gramm/Phil_Gramm_0002.jpg
+Vanessa_Redgrave/Vanessa_Redgrave_0004.jpg
+Mike_Johanns/Mike_Johanns_0001.jpg
+Mark_Salter/Mark_Salter_0001.jpg
+Jimmy_Carter/Jimmy_Carter_0008.jpg
+Wei_Wu/Wei_Wu_0001.jpg
+John_Blaney/John_Blaney_0002.jpg
+Dave_Lewis/Dave_Lewis_0001.jpg
+Hiroki_Gomi/Hiroki_Gomi_0001.jpg
+Richard_Myers/Richard_Myers_0018.jpg
+MC_Hammer/MC_Hammer_0001.jpg
+John_Leguizamo/John_Leguizamo_0001.jpg
+Sonja_Kesselschlager/Sonja_Kesselschlager_0001.jpg
+Jessica_Capshaw/Jessica_Capshaw_0001.jpg
+Robert_De_Niro/Robert_De_Niro_0004.jpg
+Anja_Paerson/Anja_Paerson_0001.jpg
+Andres_Pastrana/Andres_Pastrana_0001.jpg
+Vytas_Danelius/Vytas_Danelius_0001.jpg
+Alexander_Rumyantsev/Alexander_Rumyantsev_0001.jpg
+Scott_Blum/Scott_Blum_0001.jpg
+Hermes_Gamonal/Hermes_Gamonal_0001.jpg
+Karen_Mok/Karen_Mok_0001.jpg
+Stan_Heath/Stan_Heath_0002.jpg
+Greg_Kinnear/Greg_Kinnear_0001.jpg
+Ascencion_Barajas/Ascencion_Barajas_0001.jpg
+John_Barnett/John_Barnett_0001.jpg
+Esteban_Cordoba-Velazquez/Esteban_Cordoba-Velazquez_0001.jpg
+Judy_Locy/Judy_Locy_0001.jpg
+Randy_Travis/Randy_Travis_0001.jpg
+Rahul_Dravid/Rahul_Dravid_0001.jpg
+Joseph_Hoy/Joseph_Hoy_0001.jpg
+Roger_Lyons/Roger_Lyons_0001.jpg
+Daryl_Smith/Daryl_Smith_0001.jpg
+Oleksandr_Moroz/Oleksandr_Moroz_0002.jpg
+Barbara_Walters/Barbara_Walters_0004.jpg
+William_Murabito/William_Murabito_0001.jpg
+Ozzie_Smith/Ozzie_Smith_0001.jpg
+Matt_Braker/Matt_Braker_0001.jpg
+Steven_Seagal/Steven_Seagal_0002.jpg
+Kevin_Sorbo/Kevin_Sorbo_0001.jpg
+Suk_Chung_Hong/Suk_Chung_Hong_0001.jpg
+James_Hoffa/James_Hoffa_0001.jpg
+Lucio_Angulo/Lucio_Angulo_0001.jpg
+Larry_Lindsey/Larry_Lindsey_0001.jpg
+Donald_Keyser/Donald_Keyser_0001.jpg
+Adrian_Nastase/Adrian_Nastase_0002.jpg
+Jane_Leeves/Jane_Leeves_0001.jpg
+Katie_Holmes/Katie_Holmes_0001.jpg
+Lois_Smart/Lois_Smart_0001.jpg
+Rosie_Perez/Rosie_Perez_0001.jpg
+Vince_Vaughan/Vince_Vaughan_0001.jpg
+Ricardo_Lopez_Murphy/Ricardo_Lopez_Murphy_0002.jpg
+Daniel_Barenboim/Daniel_Barenboim_0001.jpg
+Rachel_Kempson/Rachel_Kempson_0001.jpg
+Liliana_Cavani/Liliana_Cavani_0001.jpg
+Mark_Redman/Mark_Redman_0001.jpg
+LeRoy_Millette_Jr/LeRoy_Millette_Jr_0001.jpg
+Celso_Amorim/Celso_Amorim_0001.jpg
+Jacky_Cheung/Jacky_Cheung_0001.jpg
+Bo_Ryan/Bo_Ryan_0002.jpg
+Agnes_Bruckner/Agnes_Bruckner_0001.jpg
+Ali_Abbas/Ali_Abbas_0002.jpg
+Dirk_Kempthorne/Dirk_Kempthorne_0001.jpg
+Jennifer_Pena/Jennifer_Pena_0001.jpg
+Stellan_Skarsgard/Stellan_Skarsgard_0002.jpg
+Jung_Bong/Jung_Bong_0001.jpg
+Charlie_Deane/Charlie_Deane_0001.jpg
+Eliza_Dushku/Eliza_Dushku_0001.jpg
+Bilal_Erdogan/Bilal_Erdogan_0001.jpg
+Etta_James/Etta_James_0001.jpg
+Vecdi_Gonul/Vecdi_Gonul_0001.jpg
+Larry_Hagman/Larry_Hagman_0001.jpg
+Gil_de_Ferran/Gil_de_Ferran_0004.jpg
+Ken_Dorsey/Ken_Dorsey_0001.jpg
+Matt_Welsh/Matt_Welsh_0001.jpg
+Claudine_Farrell/Claudine_Farrell_0001.jpg
+Jen_Bice/Jen_Bice_0001.jpg
+Alexis_Dennisoff/Alexis_Dennisoff_0001.jpg
+Karen_Allen/Karen_Allen_0001.jpg
+E_Clay_Shaw/E_Clay_Shaw_0001.jpg
+Jose_Lina/Jose_Lina_0001.jpg
+Fujio_Mitarai/Fujio_Mitarai_0001.jpg
+Leon_Silver/Leon_Silver_0001.jpg
+Emmanuel_Milingo/Emmanuel_Milingo_0001.jpg
+David_Dorfman/David_Dorfman_0001.jpg
+Henning_Scherf/Henning_Scherf_0001.jpg
+Lloyd_Mudiwa/Lloyd_Mudiwa_0001.jpg
+Barbara_Brezigar/Barbara_Brezigar_0002.jpg
+Robert_Fico/Robert_Fico_0002.jpg
+Steve_Wariner/Steve_Wariner_0001.jpg
+Emily_Mortimer/Emily_Mortimer_0001.jpg
+Don_Boudria/Don_Boudria_0001.jpg
+Diane_Ladd/Diane_Ladd_0001.jpg
+Piotr_Anderszewski/Piotr_Anderszewski_0001.jpg
+Robin_Cook/Robin_Cook_0002.jpg
+Eduardo_Duhalde/Eduardo_Duhalde_0008.jpg
+Nathalie_Baye/Nathalie_Baye_0001.jpg
+Kurt_Thomas/Kurt_Thomas_0001.jpg
+Hong_Myung/Hong_Myung_0001.jpg
+Tonga/Tonga_0001.jpg
+Caroline_Kennedy/Caroline_Kennedy_0002.jpg
+Allan_Kemakeza/Allan_Kemakeza_0001.jpg
+Elisabeth_Schumacher/Elisabeth_Schumacher_0002.jpg
+Sanjay_Chawla/Sanjay_Chawla_0001.jpg
+Douglas_Faneuil/Douglas_Faneuil_0001.jpg
+Picabo_Street/Picabo_Street_0001.jpg
+Dan_LaCoutre/Dan_LaCoutre_0001.jpg
+Andy_Madikians/Andy_Madikians_0001.jpg
+Namuddu_Florence/Namuddu_Florence_0001.jpg
+Nathalie_Dechy/Nathalie_Dechy_0001.jpg
+Chen_Tsai-chin/Chen_Tsai-chin_0001.jpg
+Pascal_Lamy/Pascal_Lamy_0002.jpg
+Leni_Bjorklund/Leni_Bjorklund_0001.jpg
+Herbert_Haupt/Herbert_Haupt_0001.jpg
+Anna_Chicherova/Anna_Chicherova_0001.jpg
+Bernadette_Peters/Bernadette_Peters_0001.jpg
+Sue_Slavec/Sue_Slavec_0001.jpg
+Lloyd_Ward/Lloyd_Ward_0002.jpg
+Lim_Dong-won/Lim_Dong-won_0002.jpg
+Chris_Claiborne/Chris_Claiborne_0001.jpg
+Catherine_Woodard/Catherine_Woodard_0001.jpg
+Ian_Campbell/Ian_Campbell_0001.jpg
+Dave_McGinnis/Dave_McGinnis_0001.jpg
+Tom_Christerson/Tom_Christerson_0001.jpg
+Rosemarie_Stack/Rosemarie_Stack_0001.jpg
+Tommy_Franks/Tommy_Franks_0006.jpg
+Rosalyn_Carter/Rosalyn_Carter_0001.jpg
+Kostya_Tszyu/Kostya_Tszyu_0001.jpg
+Masatoshi_Koshiba/Masatoshi_Koshiba_0001.jpg
+Hannah_Stockbauer/Hannah_Stockbauer_0001.jpg
+Stephanie_Moore/Stephanie_Moore_0001.jpg
+Saparmurat_Niyazov/Saparmurat_Niyazov_0001.jpg
+Suh_Young-hoon/Suh_Young-hoon_0001.jpg
+Nanni_Moretti/Nanni_Moretti_0001.jpg
+Omar_Vizquel/Omar_Vizquel_0001.jpg
+Marie_Haghal/Marie_Haghal_0001.jpg
+Tom_Harkin/Tom_Harkin_0005.jpg
+Jenna_Elfman/Jenna_Elfman_0001.jpg
+Joseph_LePore/Joseph_LePore_0001.jpg
+Astrid_Betancourt/Astrid_Betancourt_0001.jpg
+Jennifer_Murray/Jennifer_Murray_0001.jpg
+Shaukat_Aziz/Shaukat_Aziz_0001.jpg
+Lachlan_Murdoch/Lachlan_Murdoch_0001.jpg
+Nora_Ephron/Nora_Ephron_0001.jpg
+Jamie_Carey/Jamie_Carey_0001.jpg
+Antonio_Catania/Antonio_Catania_0001.jpg
+Francois_Botha/Francois_Botha_0001.jpg
+Jim_Freudenberg/Jim_Freudenberg_0001.jpg
+Mohammad_Mustapha_Miro/Mohammad_Mustapha_Miro_0001.jpg
+Marcelo_Bielsa/Marcelo_Bielsa_0001.jpg
+Miguel_Angel_Rodriguez/Miguel_Angel_Rodriguez_0001.jpg
+Teri_Files/Teri_Files_0001.jpg
+Andrew_Shutley/Andrew_Shutley_0001.jpg
+Ted_Nolan/Ted_Nolan_0001.jpg
+Nikki_McKibbin/Nikki_McKibbin_0001.jpg
+Debbie_Reynolds/Debbie_Reynolds_0004.jpg
+Billy_Graham/Billy_Graham_0002.jpg
+Kirsten_Dunst/Kirsten_Dunst_0001.jpg
+Bruce_Springsteen/Bruce_Springsteen_0002.jpg
+Maurice_Cheeks/Maurice_Cheeks_0001.jpg
+Jose_Miguel_Aleman/Jose_Miguel_Aleman_0001.jpg
+Dave_Odom/Dave_Odom_0001.jpg
+Joe_Finley/Joe_Finley_0001.jpg
+Eddy_Hartenstein/Eddy_Hartenstein_0001.jpg
+Bruce_Weber/Bruce_Weber_0002.jpg
+Adolfo_Aguilar_Zinser/Adolfo_Aguilar_Zinser_0001.jpg
+Farouk_al-Sharaa/Farouk_al-Sharaa_0001.jpg
+Peter_Harrison/Peter_Harrison_0001.jpg
+Mayumi_Moriyama/Mayumi_Moriyama_0001.jpg
+Sally_Clark/Sally_Clark_0001.jpg
+Matthew_McConaughey/Matthew_McConaughey_0001.jpg
+Gideon_Yago/Gideon_Yago_0001.jpg
+Natalya_Sazanovich/Natalya_Sazanovich_0001.jpg
+Elizabeth_Regan/Elizabeth_Regan_0001.jpg
+Abel_Aguilar/Abel_Aguilar_0001.jpg
+Sultan_Qaboos/Sultan_Qaboos_0001.jpg
+Lucia_Kenny_Anthony/Lucia_Kenny_Anthony_0001.jpg
+Raza_Rabbani/Raza_Rabbani_0001.jpg
+Bill_Guerin/Bill_Guerin_0001.jpg
+Francis_George/Francis_George_0001.jpg
+Mario_Kreutzberger/Mario_Kreutzberger_0002.jpg
+Hana_Makhmalbaf/Hana_Makhmalbaf_0001.jpg
+Robert_Beck/Robert_Beck_0001.jpg
+Chris_Moore/Chris_Moore_0001.jpg
+Yolanda_King/Yolanda_King_0001.jpg
+William_Macy/William_Macy_0002.jpg
+Harriet_Lessy/Harriet_Lessy_0001.jpg
+Kathryn_Morris/Kathryn_Morris_0001.jpg
+Lucrecia_Orozco/Lucrecia_Orozco_0001.jpg
+Dave_Ragone/Dave_Ragone_0001.jpg
+Melanie_Griffith/Melanie_Griffith_0001.jpg
+Joe_Glover/Joe_Glover_0001.jpg
+Thomas_OBrien/Thomas_OBrien_0008.jpg
+Mike_Martz/Mike_Martz_0007.jpg
+Buddy_Ryan/Buddy_Ryan_0001.jpg
+Dennis_Archer/Dennis_Archer_0001.jpg
+Marat_Safin/Marat_Safin_0003.jpg
+Teresa_Williams/Teresa_Williams_0001.jpg
+Jerry_McEntee/Jerry_McEntee_0001.jpg
+Debra_Brown/Debra_Brown_0002.jpg
+Leo_Mullin/Leo_Mullin_0001.jpg
+Brian_Scalabrine/Brian_Scalabrine_0001.jpg
+Judy_Spreckels/Judy_Spreckels_0001.jpg
+Dan_Prinster/Dan_Prinster_0001.jpg
+Albert_Montanes/Albert_Montanes_0001.jpg
+James_Carville/James_Carville_0001.jpg
+Leszek_Miller/Leszek_Miller_0003.jpg
+Michael_Guiler/Michael_Guiler_0001.jpg
+Cecilia_Cheung/Cecilia_Cheung_0001.jpg
+Andres_DAlessandro/Andres_DAlessandro_0001.jpg
+Claudia_Cardinale/Claudia_Cardinale_0001.jpg
+Hans_Eichel/Hans_Eichel_0001.jpg
+Robert_Bullock/Robert_Bullock_0001.jpg
+Luis_Berrondo/Luis_Berrondo_0001.jpg
+Mahmoud_Abbas/Mahmoud_Abbas_0007.jpg
+Clark_Randt/Clark_Randt_0001.jpg
+Paul_Henderson/Paul_Henderson_0001.jpg
+Katja_Riemann/Katja_Riemann_0001.jpg
+Marcus_Gronholm/Marcus_Gronholm_0001.jpg
+Jamir_Miller/Jamir_Miller_0001.jpg
+Milo_Maestrecampo/Milo_Maestrecampo_0001.jpg
+Tanya_Lindenmuth/Tanya_Lindenmuth_0001.jpg
+Peter_Schultz/Peter_Schultz_0001.jpg
+Javier_Solana/Javier_Solana_0009.jpg
+William_Ford_Jr/William_Ford_Jr_0001.jpg
+Liu_Xiaoqing/Liu_Xiaoqing_0001.jpg
+Suzanne_Torrance/Suzanne_Torrance_0001.jpg
+Imran_Khan/Imran_Khan_0001.jpg
+Igor_Ivanov/Igor_Ivanov_0006.jpg
+Glenn_Plummer/Glenn_Plummer_0001.jpg
+JC_Chasez/JC_Chasez_0001.jpg
+John_McKay/John_McKay_0001.jpg
+Ian_Smith/Ian_Smith_0001.jpg
+Afton_Smith/Afton_Smith_0001.jpg
+Geno_Auriemma/Geno_Auriemma_0002.jpg
+Milt_Heflin/Milt_Heflin_0001.jpg
+Robert_Towne/Robert_Towne_0001.jpg
+David_Heymann/David_Heymann_0005.jpg
+Wayne_Gretzky/Wayne_Gretzky_0004.jpg
+Manuel_Pellegrini/Manuel_Pellegrini_0001.jpg
+Diego_Diego_Lerman/Diego_Diego_Lerman_0001.jpg
+Aaron_Eckhart/Aaron_Eckhart_0001.jpg
+Roberto_Marinho/Roberto_Marinho_0002.jpg
+Jerry_Colangelo/Jerry_Colangelo_0001.jpg
+Paul_McCartney/Paul_McCartney_0001.jpg
+Abel_Pacheco/Abel_Pacheco_0004.jpg
+Jacques_Villeneuve/Jacques_Villeneuve_0001.jpg
+Michael_Bouchard/Michael_Bouchard_0001.jpg
+Mariano_Zabaleta/Mariano_Zabaleta_0001.jpg
+Theresa_May/Theresa_May_0001.jpg
+Duane_Barber/Duane_Barber_0001.jpg
+Roger_Mahony/Roger_Mahony_0001.jpg
+Ahmet_Demir/Ahmet_Demir_0001.jpg
+Christy_Turlington/Christy_Turlington_0001.jpg
+Momir_Nikolic/Momir_Nikolic_0001.jpg
+Edward_Norton/Edward_Norton_0001.jpg
+Georgina_Bardach/Georgina_Bardach_0001.jpg
+Chris_Whitney/Chris_Whitney_0001.jpg
+Frank_Stallone/Frank_Stallone_0002.jpg
+Adam_Rich/Adam_Rich_0001.jpg
+Anderson_Varejao/Anderson_Varejao_0001.jpg
+Gennifer_Flowers/Gennifer_Flowers_0001.jpg
+Larry_Lucchino/Larry_Lucchino_0002.jpg
+John_Hartson/John_Hartson_0001.jpg
+Christopher_Matero/Christopher_Matero_0001.jpg
+Anne_Cavers/Anne_Cavers_0001.jpg
+Marion_Barry/Marion_Barry_0001.jpg
+Inam-ul-Haq/Inam-ul-Haq_0001.jpg
+Ryan_Leaf/Ryan_Leaf_0001.jpg
+Patricia_Russo/Patricia_Russo_0001.jpg
+William_Bratton/William_Bratton_0001.jpg
+Ed_Mekertichian/Ed_Mekertichian_0001.jpg
+George_Allen/George_Allen_0001.jpg
+Lennart_Johansson/Lennart_Johansson_0001.jpg
+Eliott_Spitzer/Eliott_Spitzer_0001.jpg
+Todd_Reid/Todd_Reid_0001.jpg
+Josh_Childress/Josh_Childress_0001.jpg
+Pat_Cox/Pat_Cox_0001.jpg
+Giulietta_Masina/Giulietta_Masina_0001.jpg
+Edmund_Stoiber/Edmund_Stoiber_0010.jpg
+Carlos_Barragan/Carlos_Barragan_0001.jpg
+Jiang_Zemin/Jiang_Zemin_0019.jpg
+Brigitte_Boisselier/Brigitte_Boisselier_0002.jpg
+LeAnn_Rimes/LeAnn_Rimes_0002.jpg
+Alan_Dreher/Alan_Dreher_0001.jpg
+Rodolfo_Abalos/Rodolfo_Abalos_0001.jpg
+Alex_Ferguson/Alex_Ferguson_0001.jpg
+Wendy_Selig/Wendy_Selig_0001.jpg
+Julian_Fantino/Julian_Fantino_0001.jpg
+Robin_McLaurin_Williams/Robin_McLaurin_Williams_0001.jpg
+Robert_Durst/Robert_Durst_0001.jpg
+Natalia_Vodonova/Natalia_Vodonova_0001.jpg
+John_Taylor/John_Taylor_0002.jpg
+Angela_Mascia-Frye/Angela_Mascia-Frye_0001.jpg
+Troy_Jenkins/Troy_Jenkins_0001.jpg
+Gerhard_Boekel/Gerhard_Boekel_0001.jpg
+Michelle_Hofland/Michelle_Hofland_0001.jpg
+Hal_McCoy/Hal_McCoy_0001.jpg
+Carla_Myers/Carla_Myers_0001.jpg
+Gene_Robinson/Gene_Robinson_0001.jpg
+Rodrigo_Borja/Rodrigo_Borja_0001.jpg
+John_Lawrence/John_Lawrence_0001.jpg
+Teresa_Worbis/Teresa_Worbis_0001.jpg
+Joel_Todd/Joel_Todd_0001.jpg
+Erick_Barkley/Erick_Barkley_0001.jpg
+Kyle_Shewfelt/Kyle_Shewfelt_0001.jpg
+James_Gibson/James_Gibson_0001.jpg
+Sergei_Ivanov/Sergei_Ivanov_0003.jpg
+Mike_Flanagan/Mike_Flanagan_0001.jpg
+Fernando_Hierro/Fernando_Hierro_0001.jpg
+Jean-David_Levitte/Jean-David_Levitte_0001.jpg
+Antje_Buschschulte/Antje_Buschschulte_0001.jpg
+Elena_de_Chavez/Elena_de_Chavez_0001.jpg
+Dino_de_Laurentis/Dino_de_Laurentis_0001.jpg
+Tatiana_Gratcheva/Tatiana_Gratcheva_0001.jpg
+Queen_Beatrix/Queen_Beatrix_0004.jpg
+Bill_Parcells/Bill_Parcells_0001.jpg
+Carlos_Mesa/Carlos_Mesa_0001.jpg
+Ruben_Sierra/Ruben_Sierra_0001.jpg
+Cathy_Cunningham/Cathy_Cunningham_0001.jpg
+Shaul_Mofaz/Shaul_Mofaz_0001.jpg
+Charlotte_Chambers/Charlotte_Chambers_0001.jpg
+Nikki_Reed/Nikki_Reed_0001.jpg
+Muhammad_Saeed_al-Sahhaf/Muhammad_Saeed_al-Sahhaf_0005.jpg
+Martin_Scorsese/Martin_Scorsese_0004.jpg
+Jim_Cantalupo/Jim_Cantalupo_0001.jpg
+Olivier_Rochus/Olivier_Rochus_0001.jpg
+Billy_Donovan/Billy_Donovan_0001.jpg
+Jessica_Lynch/Jessica_Lynch_0001.jpg
+Quincy_Jones/Quincy_Jones_0001.jpg
+Mile_Mrksic/Mile_Mrksic_0001.jpg
+Jim_Haslett/Jim_Haslett_0001.jpg
+Lana_Clarkson/Lana_Clarkson_0002.jpg
+Jose_Rosado/Jose_Rosado_0001.jpg
+George_Papandreou/George_Papandreou_0004.jpg
+Jim_Bollman/Jim_Bollman_0001.jpg
+Heath_Ledger/Heath_Ledger_0002.jpg
+Mike_Bryan/Mike_Bryan_0001.jpg
+John_Negroponte/John_Negroponte_0006.jpg
+Rollie_Massimino/Rollie_Massimino_0001.jpg
+Kurt_Schottenheimer/Kurt_Schottenheimer_0001.jpg
+Martin_McGuinness/Martin_McGuinness_0001.jpg
+Tonino_Guerra/Tonino_Guerra_0001.jpg
+Patrick_Dennehy/Patrick_Dennehy_0001.jpg
+Tatyana_Tomashova/Tatyana_Tomashova_0001.jpg
+Graeme_Lloyd/Graeme_Lloyd_0001.jpg
+Adam_Freier/Adam_Freier_0001.jpg
+Kristin_Chenoweth/Kristin_Chenoweth_0001.jpg
+Michel_Charles_Chretien/Michel_Charles_Chretien_0001.jpg
+Ben_Wallace/Ben_Wallace_0001.jpg
+John_Velazquez/John_Velazquez_0001.jpg
+John_Burkett/John_Burkett_0001.jpg
+Giancarlo_Fisichella/Giancarlo_Fisichella_0001.jpg
+Harrison_Ford/Harrison_Ford_0009.jpg
+Herb_Dhaliwal/Herb_Dhaliwal_0001.jpg
+Gabi_Zimmer/Gabi_Zimmer_0001.jpg
+Bill_King/Bill_King_0001.jpg
+Clint_Eastwood/Clint_Eastwood_0001.jpg
+Melissa_Gilbert/Melissa_Gilbert_0001.jpg
+Stephen_Ambrose/Stephen_Ambrose_0001.jpg
+Danny_Green/Danny_Green_0001.jpg
+Wang_Yi/Wang_Yi_0002.jpg
+Mike_Tyson/Mike_Tyson_0004.jpg
+Quin_Snyder/Quin_Snyder_0001.jpg
+Roy_Halladay/Roy_Halladay_0001.jpg
+Moshe_Katsav/Moshe_Katsav_0004.jpg
+Edith_Masai/Edith_Masai_0001.jpg
+Mercedes_Amor/Mercedes_Amor_0001.jpg
+Claudette_Robinson/Claudette_Robinson_0001.jpg
+Eric_Robert_Rudolph/Eric_Robert_Rudolph_0002.jpg
+James_Butts/James_Butts_0001.jpg
+Paul_Shanley/Paul_Shanley_0002.jpg
+Nathan_Doudney/Nathan_Doudney_0001.jpg
+Miguel_Contreras/Miguel_Contreras_0002.jpg
+Fernando_Alonso/Fernando_Alonso_0001.jpg
+Monica_Vergara/Monica_Vergara_0001.jpg
+Miguel_Aldana_Ibarra/Miguel_Aldana_Ibarra_0001.jpg
+Pauley_Perrette/Pauley_Perrette_0001.jpg
+Sim_Yong/Sim_Yong_0001.jpg
+Dexter_Jackson/Dexter_Jackson_0001.jpg
+Jim_Leach/Jim_Leach_0001.jpg
+Anwar_Ibrahim/Anwar_Ibrahim_0002.jpg
+Anthony_Hopkins/Anthony_Hopkins_0001.jpg
+Donna_Brazile/Donna_Brazile_0001.jpg
+Elsa_Zylberstein/Elsa_Zylberstein_0006.jpg
+Ambrose_Lee/Ambrose_Lee_0001.jpg
+Inga_Hall/Inga_Hall_0001.jpg
+David_Anderson/David_Anderson_0001.jpg
+Ralph_Goodale/Ralph_Goodale_0001.jpg
+John_Warner/John_Warner_0004.jpg
+Richard_Regenhard/Richard_Regenhard_0001.jpg
+Kathleen_Glynn/Kathleen_Glynn_0001.jpg
+Bill_Walton/Bill_Walton_0001.jpg
+Patti_Smith/Patti_Smith_0001.jpg
+Perry_Farrell/Perry_Farrell_0001.jpg
+Joseph_Estrada/Joseph_Estrada_0001.jpg
+Eurico_Guterres/Eurico_Guterres_0001.jpg
+Patsy_Hardy/Patsy_Hardy_0001.jpg
+Vanessa_Williams/Vanessa_Williams_0003.jpg
+Gretchen_Mol/Gretchen_Mol_0001.jpg
+Rainer_Gut/Rainer_Gut_0001.jpg
+Emilio_Botin/Emilio_Botin_0001.jpg
+Raymond_Odierno/Raymond_Odierno_0001.jpg
+Kim_Jong-Il/Kim_Jong-Il_0004.jpg
+Atom_Egoyan/Atom_Egoyan_0001.jpg
+Ben_Cohen/Ben_Cohen_0001.jpg
+Nina_Jacobson/Nina_Jacobson_0001.jpg
+Chung_Mong-hun/Chung_Mong-hun_0001.jpg
+Terry_Bradshaw/Terry_Bradshaw_0001.jpg
+Michael_Donovan/Michael_Donovan_0001.jpg
+Brett_Hull/Brett_Hull_0001.jpg
+Dan_Bartlett/Dan_Bartlett_0001.jpg
+Roberto_Tovar/Roberto_Tovar_0001.jpg
+Roman_Coppola/Roman_Coppola_0001.jpg
+Ricky_Ponting/Ricky_Ponting_0001.jpg
+Eugene_Teslovic/Eugene_Teslovic_0001.jpg
+Tuncay_Sanli/Tuncay_Sanli_0001.jpg
+Naji_Sabri/Naji_Sabri_0008.jpg
+Jim_Abbott/Jim_Abbott_0001.jpg
+Hasan_Wirayuda/Hasan_Wirayuda_0001.jpg
+Andrew_Jarecki/Andrew_Jarecki_0001.jpg
+Mark_Swartz/Mark_Swartz_0001.jpg
+Andy_Bryant/Andy_Bryant_0001.jpg
+Cristina_Torrens_Valero/Cristina_Torrens_Valero_0001.jpg
+Scott_McNealy/Scott_McNealy_0002.jpg
+Steven_Tyler/Steven_Tyler_0001.jpg
+Hartmut_Mehdorn/Hartmut_Mehdorn_0001.jpg
+Queen_Noor/Queen_Noor_0001.jpg
+Jamie_Martin/Jamie_Martin_0001.jpg
+Steven_Briggs/Steven_Briggs_0001.jpg
+Donald_Carty/Donald_Carty_0001.jpg
+Michael_Arif/Michael_Arif_0001.jpg
+Mireya_Moscoso/Mireya_Moscoso_0004.jpg
+GL_Peiris/GL_Peiris_0004.jpg
+Terry_McAuliffe/Terry_McAuliffe_0001.jpg
+Greg_Rusedski/Greg_Rusedski_0004.jpg
+Jean-Claude_Juncker/Jean-Claude_Juncker_0002.jpg
+Susilo_Bambang_Yudhoyono/Susilo_Bambang_Yudhoyono_0004.jpg
+Alex_King/Alex_King_0001.jpg
+Mona_Ayoub/Mona_Ayoub_0001.jpg
+Anna_Kournikova/Anna_Kournikova_0008.jpg
+Wolfgang_Becker/Wolfgang_Becker_0001.jpg
+Nate_Huffman/Nate_Huffman_0001.jpg
+Paul_Schrader/Paul_Schrader_0001.jpg
+David_Nalbandian/David_Nalbandian_0014.jpg
+Jeanne_Anne_Schroeder/Jeanne_Anne_Schroeder_0001.jpg
+Harvey_Fierstein/Harvey_Fierstein_0001.jpg
+Annie_Chaplin/Annie_Chaplin_0001.jpg
+Sourav_Ganguly/Sourav_Ganguly_0005.jpg
+Mark_Everson/Mark_Everson_0001.jpg
+Alexandra_Rozovskaya/Alexandra_Rozovskaya_0001.jpg
+Ismail_Merchant/Ismail_Merchant_0001.jpg
+Craig_Doblin/Craig_Doblin_0001.jpg
+Marsah_Ambrosius/Marsah_Ambrosius_0001.jpg
+Richard_Jewell/Richard_Jewell_0001.jpg
+Richard_Parsons/Richard_Parsons_0001.jpg
+Carlos_Moya/Carlos_Moya_0004.jpg
+Chuck_Bednarik/Chuck_Bednarik_0001.jpg
+Carson_Palmer/Carson_Palmer_0003.jpg
+Sigourney_Weaver/Sigourney_Weaver_0001.jpg
+Tim_Welsh/Tim_Welsh_0001.jpg
+Mitt_Romney/Mitt_Romney_0001.jpg
+Hitomi_Soga/Hitomi_Soga_0005.jpg
+Joe_Dicaro/Joe_Dicaro_0001.jpg
+Park_Jung_Sung/Park_Jung_Sung_0001.jpg
+Yogi_Berra/Yogi_Berra_0001.jpg
+Robert_Horan/Robert_Horan_0001.jpg
+Laura_Pausini/Laura_Pausini_0001.jpg
+Jay_Garner/Jay_Garner_0003.jpg
+George_Brumley/George_Brumley_0001.jpg
+Bill_Kollar/Bill_Kollar_0001.jpg
+Monte_Kiffin/Monte_Kiffin_0001.jpg
+Jayne_Yarris/Jayne_Yarris_0001.jpg
+Barbra_Streisand/Barbra_Streisand_0003.jpg
+Dan_Reeves/Dan_Reeves_0001.jpg
+Vince_Carter/Vince_Carter_0001.jpg
+Pinar_del_Rio/Pinar_del_Rio_0001.jpg
+Scott_McClellan/Scott_McClellan_0001.jpg
+Jacques_Chirac/Jacques_Chirac_0043.jpg
+Emelie_Loit/Emelie_Loit_0001.jpg
+Princess_Aiko/Princess_Aiko_0001.jpg
+Pedro_Duque/Pedro_Duque_0001.jpg
+Guy_Verhofstadt/Guy_Verhofstadt_0001.jpg
+Uzi_Even/Uzi_Even_0001.jpg
+Alberto_Ruiz_Gallardon/Alberto_Ruiz_Gallardon_0001.jpg
+Lew_Rywin/Lew_Rywin_0001.jpg
+Peter_Fisher/Peter_Fisher_0001.jpg
+Alexandra_Stevenson/Alexandra_Stevenson_0001.jpg
+Doug_Moe/Doug_Moe_0001.jpg
+James_McGreevey/James_McGreevey_0003.jpg
+Mohammed_Al-Douri/Mohammed_Al-Douri_0009.jpg
+Richard_Rodriguez/Richard_Rodriguez_0001.jpg
+Zelma_Novelo/Zelma_Novelo_0001.jpg
+Tom_Poston/Tom_Poston_0001.jpg
+Edouard_Michelin/Edouard_Michelin_0001.jpg
+Lidija_Djukanovic/Lidija_Djukanovic_0001.jpg
+Richard_Branson/Richard_Branson_0001.jpg
+Roland_Koch/Roland_Koch_0001.jpg
+Ismail_Abu_Shanab/Ismail_Abu_Shanab_0001.jpg
+David_Glenn/David_Glenn_0001.jpg
+Enos_Slaughter/Enos_Slaughter_0001.jpg
+Robert_Witt/Robert_Witt_0002.jpg
+Gong_Li/Gong_Li_0001.jpg
+Ben_Affleck/Ben_Affleck_0001.jpg
+Ed_Wade/Ed_Wade_0001.jpg
+Steve_Patterson/Steve_Patterson_0001.jpg
+Carla_Gay_Balingit/Carla_Gay_Balingit_0001.jpg
+Hugo_Conte/Hugo_Conte_0001.jpg
+Sivan_Klein/Sivan_Klein_0001.jpg
+Simon_Yam/Simon_Yam_0001.jpg
+Steve_Pagliuca/Steve_Pagliuca_0001.jpg
+Andy_Garcia/Andy_Garcia_0001.jpg
+Justine_Pasek/Justine_Pasek_0001.jpg
+Laurent_Woulzy/Laurent_Woulzy_0001.jpg
+Qian_Qichen/Qian_Qichen_0001.jpg
+Ibrahim_Haddad/Ibrahim_Haddad_0001.jpg
+Brian_StPierre/Brian_StPierre_0001.jpg
+Guillaume_Soro/Guillaume_Soro_0001.jpg
+Denise_Johnson/Denise_Johnson_0002.jpg
+Justine_Henin/Justine_Henin_0003.jpg
+Rod_Stewart/Rod_Stewart_0001.jpg
+Edward_Belvin/Edward_Belvin_0001.jpg
+Joel_Gallen/Joel_Gallen_0001.jpg
+Lubomir_Zaoralek/Lubomir_Zaoralek_0001.jpg
+Ahmed_Ahmed/Ahmed_Ahmed_0001.jpg
+Kristin_Scott/Kristin_Scott_0001.jpg
+Shamai_Leibowitz/Shamai_Leibowitz_0001.jpg
+Ted_Washington/Ted_Washington_0001.jpg
+Yoon_Won-Sik/Yoon_Won-Sik_0001.jpg
+Lonnie_Donegan/Lonnie_Donegan_0001.jpg
+Naomi_Bronstein/Naomi_Bronstein_0001.jpg
+Herb_Ritts/Herb_Ritts_0001.jpg
+David_Kelley/David_Kelley_0001.jpg
+Al_Leiter/Al_Leiter_0001.jpg
+Hector_Grullon/Hector_Grullon_0001.jpg
+Maha_Habib/Maha_Habib_0001.jpg
+John_Mayer/John_Mayer_0003.jpg
+Terri_Clark/Terri_Clark_0001.jpg
+Dionigi_Tettamanzi/Dionigi_Tettamanzi_0001.jpg
+James_Phelps/James_Phelps_0001.jpg
+Jose_Luis_Chilavert/Jose_Luis_Chilavert_0001.jpg
+Bob_Stoops/Bob_Stoops_0007.jpg
+Gala_Leon_Garcia/Gala_Leon_Garcia_0001.jpg
+Jan_Paul_Miller/Jan_Paul_Miller_0001.jpg
+Anthony_Mazur/Anthony_Mazur_0001.jpg
+Tom_Jones/Tom_Jones_0002.jpg
+David_Braley/David_Braley_0001.jpg
+Dean_Barker/Dean_Barker_0001.jpg
+Peter_Harvey/Peter_Harvey_0001.jpg
+Bob_Krueger/Bob_Krueger_0001.jpg
+Billy_Gilman/Billy_Gilman_0001.jpg
+Frank_Coraci/Frank_Coraci_0001.jpg
+Warren_Truss/Warren_Truss_0001.jpg
+Chyung_Dai-chul/Chyung_Dai-chul_0001.jpg
+Randy_Johnson/Randy_Johnson_0001.jpg
+Pa_Kou_Hang/Pa_Kou_Hang_0001.jpg
+Alberto_Fujimori/Alberto_Fujimori_0002.jpg
+Aileen_Riggin_Soule/Aileen_Riggin_Soule_0001.jpg
+Nancy_Reagan/Nancy_Reagan_0002.jpg
+Tommy_Amaker/Tommy_Amaker_0001.jpg
+Shannyn_Sossamon/Shannyn_Sossamon_0001.jpg
+Mike_Scioscia/Mike_Scioscia_0002.jpg
+Vojislav_Seselj/Vojislav_Seselj_0001.jpg
+Earl_Fritts/Earl_Fritts_0001.jpg
+Jeffrey_Donaldson/Jeffrey_Donaldson_0001.jpg
+Rina_Lazo/Rina_Lazo_0001.jpg
+Herb_Sendek/Herb_Sendek_0004.jpg
+Kenneth_Evans/Kenneth_Evans_0001.jpg
+Kang_Gum-sil/Kang_Gum-sil_0001.jpg
+Mary_Elizabeth_Mastrantonio/Mary_Elizabeth_Mastrantonio_0001.jpg
+Grace_Dodd/Grace_Dodd_0001.jpg
+Kelvin_Sampson/Kelvin_Sampson_0001.jpg
+Ruano_Pascual/Ruano_Pascual_0001.jpg
+Leandrinho_Barbosa/Leandrinho_Barbosa_0001.jpg
+Marvan_Atapattu/Marvan_Atapattu_0001.jpg
+Eric_Fehr/Eric_Fehr_0001.jpg
+John_Sununu/John_Sununu_0001.jpg
+Otto_Schily/Otto_Schily_0001.jpg
+David_Bowie/David_Bowie_0001.jpg
+Angie_Arzola/Angie_Arzola_0001.jpg
+Rudolph_Giuliani/Rudolph_Giuliani_0021.jpg
+Emily_Mason/Emily_Mason_0001.jpg
+Alex_Cejka/Alex_Cejka_0001.jpg
+Blythe_Danner/Blythe_Danner_0001.jpg
+Jerry_Rice/Jerry_Rice_0001.jpg
+Karen_Sharpe_Kramer/Karen_Sharpe_Kramer_0001.jpg
+Jorge_Arce/Jorge_Arce_0002.jpg
+Jesse_Harris/Jesse_Harris_0003.jpg
+Davey_Johnson/Davey_Johnson_0001.jpg
+Eloy_Gutierrez/Eloy_Gutierrez_0001.jpg
+Garry_Kasparov/Garry_Kasparov_0002.jpg
+Gerhard_Schroeder/Gerhard_Schroeder_0095.jpg
+Talisa_Bratt/Talisa_Bratt_0001.jpg
+Jean-Rene_Fourtou/Jean-Rene_Fourtou_0001.jpg
+Dennis_Powell/Dennis_Powell_0002.jpg
+Rafael_Bielsa/Rafael_Bielsa_0001.jpg
+Bill_Hughes/Bill_Hughes_0001.jpg
+Miguel_Juarez_Perez/Miguel_Juarez_Perez_0001.jpg
+Paula_Dobriansky/Paula_Dobriansky_0001.jpg
+Gary_Sinise/Gary_Sinise_0001.jpg
+Alberto_Gonzales/Alberto_Gonzales_0001.jpg
+Pauline_Landers/Pauline_Landers_0001.jpg
+Ricky_Cottrill/Ricky_Cottrill_0001.jpg
+Richard_Perle/Richard_Perle_0001.jpg
+Sidney_Poitier/Sidney_Poitier_0001.jpg
+Diane_Lane/Diane_Lane_0001.jpg
+Zeng_Qinghong/Zeng_Qinghong_0001.jpg
+Mark_Andrew/Mark_Andrew_0001.jpg
+King_Gyanendra/King_Gyanendra_0001.jpg
+Johnnie_Lynn/Johnnie_Lynn_0001.jpg
+Millicent_Martin/Millicent_Martin_0001.jpg
+David_Przybyszewski/David_Przybyszewski_0001.jpg
+Jeff_George/Jeff_George_0001.jpg
+Djabir_Said-Guerni/Djabir_Said-Guerni_0001.jpg
+Columba_Bush/Columba_Bush_0001.jpg
+Ataollah_Mohajerani/Ataollah_Mohajerani_0001.jpg
+Hermando_Harton/Hermando_Harton_0001.jpg
+Sandra_Ceccarelli/Sandra_Ceccarelli_0001.jpg
+Paul_Wollnough/Paul_Wollnough_0001.jpg
+Tomoko_Hagiwara/Tomoko_Hagiwara_0002.jpg
+Ben_Stein/Ben_Stein_0001.jpg
+John_Perrota/John_Perrota_0001.jpg
+Michelle_Lecky/Michelle_Lecky_0001.jpg
+Martin_Rodriguez/Martin_Rodriguez_0001.jpg
+Robin_Williams/Robin_Williams_0001.jpg
+Stuart_Knoll/Stuart_Knoll_0001.jpg
+Enola_Rice/Enola_Rice_0001.jpg
+Filippo_Volandri/Filippo_Volandri_0001.jpg
+Jorge_Valdano/Jorge_Valdano_0001.jpg
+Michalis_Chrisohoides/Michalis_Chrisohoides_0001.jpg
+Jim_Letten/Jim_Letten_0001.jpg
+Jean-Pierre_Raffarin/Jean-Pierre_Raffarin_0003.jpg
+Charles_Cope/Charles_Cope_0001.jpg
+Rick_Bragg/Rick_Bragg_0001.jpg
+Raag_Singhal/Raag_Singhal_0001.jpg
+Frank_Abagnale_Jr/Frank_Abagnale_Jr_0001.jpg
+Helen_Darling/Helen_Darling_0001.jpg
+Steffi_Graf/Steffi_Graf_0005.jpg
+Linda_Baboolal/Linda_Baboolal_0001.jpg
+Craig_Morgan/Craig_Morgan_0001.jpg
+Emma_Thompson/Emma_Thompson_0001.jpg
+Edward_Seaga/Edward_Seaga_0001.jpg
+Eric_Christian_Olsen/Eric_Christian_Olsen_0001.jpg
+Jennifer_Thompson/Jennifer_Thompson_0002.jpg
+Cheryl_Little/Cheryl_Little_0001.jpg
+Ryan_Nyquist/Ryan_Nyquist_0001.jpg
+Carl_Pope/Carl_Pope_0001.jpg
+Joseph_Blatter/Joseph_Blatter_0002.jpg
+Tori_Amos/Tori_Amos_0001.jpg
+Turner_Gill/Turner_Gill_0001.jpg
+Azra_Akin/Azra_Akin_0002.jpg
+William_Overlin/William_Overlin_0001.jpg
+Luke_Ridnour/Luke_Ridnour_0001.jpg
+Shingo_Katayama/Shingo_Katayama_0001.jpg
+Thierry_Mariani/Thierry_Mariani_0001.jpg
+Frank_Zappa/Frank_Zappa_0001.jpg
+Christian_Malcolm/Christian_Malcolm_0001.jpg
+Garry_Trudeau/Garry_Trudeau_0001.jpg
+Rudi_Voeller/Rudi_Voeller_0001.jpg
+Reina_Hayes/Reina_Hayes_0001.jpg
+Brian_Williams/Brian_Williams_0001.jpg
+Peter_Gilmour/Peter_Gilmour_0001.jpg
+Jack_Knowlton/Jack_Knowlton_0001.jpg
+Jerelle_Kraus/Jerelle_Kraus_0001.jpg
+Mike_Myers/Mike_Myers_0003.jpg
+John_Lynch/John_Lynch_0001.jpg
+Paul_Lockhart/Paul_Lockhart_0001.jpg
+Beatriz_Merino/Beatriz_Merino_0001.jpg
+Stefan_Holm/Stefan_Holm_0001.jpg
+Chuanyun_Li/Chuanyun_Li_0001.jpg
+Linda_Amicangioli/Linda_Amicangioli_0001.jpg
+Barry_Alvarez/Barry_Alvarez_0002.jpg
+Paul_Newman/Paul_Newman_0001.jpg
+Rick_Perry/Rick_Perry_0006.jpg
+Tom_OBrien/Tom_OBrien_0001.jpg
+Margaret_Thatcher/Margaret_Thatcher_0002.jpg
+Al_Gore/Al_Gore_0005.jpg
+Hassan_Wirajuda/Hassan_Wirajuda_0002.jpg
+Zulfiqar_Ahmed/Zulfiqar_Ahmed_0001.jpg
+Kelsey_Grammer/Kelsey_Grammer_0001.jpg
+Jose_Maria_Aznar/Jose_Maria_Aznar_0019.jpg
+Martin_Landau/Martin_Landau_0001.jpg
+Chris_Cookson/Chris_Cookson_0001.jpg
+Sam_Brownback/Sam_Brownback_0001.jpg
+Roh_Moo-hyun/Roh_Moo-hyun_0011.jpg
+Lloyd_Novick/Lloyd_Novick_0001.jpg
+Troy_Hudson/Troy_Hudson_0001.jpg
+Prince_Willem-Alexander/Prince_Willem-Alexander_0001.jpg
+Chris_Kolanas/Chris_Kolanas_0001.jpg
+Ray_Romano/Ray_Romano_0006.jpg
+Kit_Bond/Kit_Bond_0001.jpg
+Jean-Francois_Pontal/Jean-Francois_Pontal_0001.jpg
+Billy_Joel/Billy_Joel_0001.jpg
+Danny_Ainge/Danny_Ainge_0001.jpg
+Adam_Ant/Adam_Ant_0001.jpg
+Hamid_Karzai/Hamid_Karzai_0018.jpg
+William_Joppy/William_Joppy_0001.jpg
+Doris_Roberts/Doris_Roberts_0002.jpg
+John_Scarlett/John_Scarlett_0001.jpg
+Huan_Chung_Yi/Huan_Chung_Yi_0001.jpg
+David_McCullough/David_McCullough_0001.jpg
+Tim_Blake_Nelson/Tim_Blake_Nelson_0001.jpg
+Drew_Bledsoe/Drew_Bledsoe_0001.jpg
+Brad_Smith/Brad_Smith_0001.jpg
+Steve_Allan/Steve_Allan_0001.jpg
+Roberto_Canessa/Roberto_Canessa_0001.jpg
+Yukiko_Okudo/Yukiko_Okudo_0001.jpg
+Mitzi_Gaynor/Mitzi_Gaynor_0001.jpg
+Diana_Ross/Diana_Ross_0001.jpg
+Tristan_Gretzky/Tristan_Gretzky_0001.jpg
+Bill_Nelson/Bill_Nelson_0002.jpg
+Jamling_Norgay/Jamling_Norgay_0001.jpg
+Alicia_Hollowell/Alicia_Hollowell_0001.jpg
+Jimmy_Szymanski/Jimmy_Szymanski_0001.jpg
+Lane_Odom/Lane_Odom_0001.jpg
+Taufik_Hidayat/Taufik_Hidayat_0001.jpg
+Joe_Torre/Joe_Torre_0004.jpg
+Tyler_Grillo/Tyler_Grillo_0001.jpg
+Art_Lopez/Art_Lopez_0001.jpg
+Alonzo_Mourning/Alonzo_Mourning_0001.jpg
+Doug_Duncan/Doug_Duncan_0002.jpg
+Jose_Santos/Jose_Santos_0001.jpg
+Jean-Claude_Trichet/Jean-Claude_Trichet_0001.jpg
+Jerry_Jones/Jerry_Jones_0001.jpg
+Nicoletta_Braschi/Nicoletta_Braschi_0001.jpg
+Steve_Lavin/Steve_Lavin_0006.jpg
+Natalia_Motuziuk/Natalia_Motuziuk_0001.jpg
+Vladimir_Putin/Vladimir_Putin_0016.jpg
+Jack_Grubman/Jack_Grubman_0001.jpg
+Paul_Lo_Duca/Paul_Lo_Duca_0001.jpg
+Jennifer_Lopez/Jennifer_Lopez_0011.jpg
+Rob_Morrow/Rob_Morrow_0001.jpg
+Julie_Gerberding/Julie_Gerberding_0014.jpg
+Abdulaziz_Kamilov/Abdulaziz_Kamilov_0001.jpg
+Federico_Castelan_Sayre/Federico_Castelan_Sayre_0001.jpg
+Andre_Techine/Andre_Techine_0001.jpg
+Joan_Jett/Joan_Jett_0001.jpg
+Norio_Ohga/Norio_Ohga_0001.jpg
+Chick_Hearn/Chick_Hearn_0003.jpg
+Richard_Hellfant/Richard_Hellfant_0001.jpg
+Izzat_Ibrahim/Izzat_Ibrahim_0001.jpg
+Tyrone_Medley/Tyrone_Medley_0001.jpg
+Alessandra_Cerna/Alessandra_Cerna_0001.jpg
+Sharon_Osbourne/Sharon_Osbourne_0001.jpg
+Corey_Maggette/Corey_Maggette_0001.jpg
+Glenn_Tilton/Glenn_Tilton_0001.jpg
+Leonard_Glick/Leonard_Glick_0001.jpg
+Mike_Davis/Mike_Davis_0001.jpg
+Guenter_Verheugen/Guenter_Verheugen_0001.jpg
+Akhmed_Zakayev/Akhmed_Zakayev_0003.jpg
+Diana_Silvius/Diana_Silvius_0001.jpg
+Tim_Conway/Tim_Conway_0002.jpg
+Don_Hewitt/Don_Hewitt_0001.jpg
+Tony_Blair/Tony_Blair_0002.jpg
+Dion_Glover/Dion_Glover_0001.jpg
+Eric_Taino/Eric_Taino_0001.jpg
+David_Scott_Morris/David_Scott_Morris_0001.jpg
+Scott_Weiland/Scott_Weiland_0001.jpg
+Gorden_Tallis/Gorden_Tallis_0001.jpg
+Erika_Reyes/Erika_Reyes_0001.jpg
+Christian_Longo/Christian_Longo_0001.jpg
+Victoria_Clarke/Victoria_Clarke_0001.jpg
+Gael_Garcia_Bermal/Gael_Garcia_Bermal_0001.jpg
+Hernan_Crespo/Hernan_Crespo_0001.jpg
+Felipe_Perez_Roque/Felipe_Perez_Roque_0004.jpg
+Wu_Peng/Wu_Peng_0001.jpg
+Horacio_de_Jesus_Montoya/Horacio_de_Jesus_Montoya_0001.jpg
+Tomas_Enge/Tomas_Enge_0001.jpg
+Dave_Wannstedt/Dave_Wannstedt_0001.jpg
+Brian_Henson/Brian_Henson_0001.jpg
+Omar_el-Heib/Omar_el-Heib_0001.jpg
+Andreas_Vinciguerra/Andreas_Vinciguerra_0001.jpg
+Luis_Figo/Luis_Figo_0002.jpg
+Kim_Su_Nam/Kim_Su_Nam_0001.jpg
+Diana_Taurasi/Diana_Taurasi_0002.jpg
+Parthiv_Patel/Parthiv_Patel_0001.jpg
+George_Lucas/George_Lucas_0001.jpg
+Kate_Capshaw/Kate_Capshaw_0001.jpg
+Mike_Tice/Mike_Tice_0001.jpg
+Aaron_Peirsol/Aaron_Peirsol_0002.jpg
+Marcus_Allen/Marcus_Allen_0001.jpg
+Mitch_Kupchak/Mitch_Kupchak_0001.jpg
+Gore_Verbinski/Gore_Verbinski_0001.jpg
+Brian_Florence/Brian_Florence_0001.jpg
+Sarah_Wynter/Sarah_Wynter_0001.jpg
+Michael_Jackson/Michael_Jackson_0008.jpg
+Jim_Kelly/Jim_Kelly_0001.jpg
+Bartosz_Kizierowski/Bartosz_Kizierowski_0001.jpg
+Carmen_Electra/Carmen_Electra_0005.jpg
+Michael_Deutsch/Michael_Deutsch_0001.jpg
+Ernie_Eves/Ernie_Eves_0002.jpg
+Asif_Ali_Zardari/Asif_Ali_Zardari_0001.jpg
+Laura_Bozzo/Laura_Bozzo_0001.jpg
+Norodom_Sihanouk/Norodom_Sihanouk_0001.jpg
+Luciano_Bovicelli/Luciano_Bovicelli_0001.jpg
+Michael_Chiklis/Michael_Chiklis_0001.jpg
+Eva_Amurri/Eva_Amurri_0001.jpg
+Stephane_Rochon/Stephane_Rochon_0001.jpg
+Mick_McCarthy/Mick_McCarthy_0001.jpg
+Cherie_Blair/Cherie_Blair_0004.jpg
+Craig_MacTavish/Craig_MacTavish_0001.jpg
+Elvis_Presley/Elvis_Presley_0002.jpg
+Dick_Devine/Dick_Devine_0001.jpg
+Lindy_Ruff/Lindy_Ruff_0001.jpg
+Kelly_Ripa/Kelly_Ripa_0001.jpg
+Daniel_Darnell/Daniel_Darnell_0001.jpg
+Pauline_Phillips/Pauline_Phillips_0001.jpg
+Jean-Pierre_Bemba/Jean-Pierre_Bemba_0001.jpg
+Martha_Sahagun_de_Fox/Martha_Sahagun_de_Fox_0001.jpg
+Sam_Bith/Sam_Bith_0002.jpg
+Oliver_Neuville/Oliver_Neuville_0001.jpg
+Edward_Lu/Edward_Lu_0002.jpg
+Dick_Jauron/Dick_Jauron_0001.jpg
+Tony_Fernandes/Tony_Fernandes_0001.jpg
+Ronnie_Jagday/Ronnie_Jagday_0001.jpg
+Nestor_Santillan/Nestor_Santillan_0001.jpg
+Juergen_Chrobog/Juergen_Chrobog_0001.jpg
+Vassilis_Xiros/Vassilis_Xiros_0001.jpg
+Muhammad_Ali/Muhammad_Ali_0006.jpg
+Karol_Kucera/Karol_Kucera_0001.jpg
+Charlton_Heston/Charlton_Heston_0006.jpg
+Freddy_Garcia/Freddy_Garcia_0002.jpg
+David_Caraway/David_Caraway_0001.jpg
+Gerald_Ford/Gerald_Ford_0001.jpg
+Catherine_Deneuve/Catherine_Deneuve_0004.jpg
+Tim_Lobinger/Tim_Lobinger_0001.jpg
+Carlos_Arroyo/Carlos_Arroyo_0001.jpg
+Muammar_Gaddafi/Muammar_Gaddafi_0001.jpg
+Urmila_Matondkar/Urmila_Matondkar_0001.jpg
+Jack_Straw/Jack_Straw_0002.jpg
+Peter_Greenspun/Peter_Greenspun_0001.jpg
+Ivan_Stambolic/Ivan_Stambolic_0001.jpg
+Derek_Abney/Derek_Abney_0001.jpg
+Scott_Verplank/Scott_Verplank_0001.jpg
+Pete_Sampras/Pete_Sampras_0018.jpg
+Hwang_Doo-yun/Hwang_Doo-yun_0001.jpg
+Stella_Tennant/Stella_Tennant_0001.jpg
+Avril_Lavigne/Avril_Lavigne_0001.jpg
+Ted_Williams/Ted_Williams_0001.jpg
+Jorge_Alberto_Galindo/Jorge_Alberto_Galindo_0001.jpg
+Tang_Jiaxuan/Tang_Jiaxuan_0008.jpg
+Elizabeth_Taylor/Elizabeth_Taylor_0001.jpg
+Jim_Greenwood/Jim_Greenwood_0001.jpg
+Michael_Lopez-Alegria/Michael_Lopez-Alegria_0001.jpg
+Andrea_Kiser/Andrea_Kiser_0001.jpg
+Eddie_Lewis/Eddie_Lewis_0001.jpg
+Aishwarya_Rai/Aishwarya_Rai_0001.jpg
+Bashar_Assad/Bashar_Assad_0004.jpg
+Edward_Lohn/Edward_Lohn_0001.jpg
+Cesar_Maia/Cesar_Maia_0001.jpg
+Svend_Robinson/Svend_Robinson_0001.jpg
+Philippe_Noiret/Philippe_Noiret_0001.jpg
+Roberto_Guaterroma/Roberto_Guaterroma_0001.jpg
+John_Eder/John_Eder_0001.jpg
+Floyd_Keith/Floyd_Keith_0001.jpg
+Dai_Chul_Chyung/Dai_Chul_Chyung_0001.jpg
+Alan_Jackson/Alan_Jackson_0001.jpg
+Calista_Flockhart/Calista_Flockhart_0005.jpg
+Jorge_Batlle/Jorge_Batlle_0001.jpg
+Robert_Pollack/Robert_Pollack_0001.jpg
+Steve_Nash/Steve_Nash_0005.jpg
+Gregory_Peck/Gregory_Peck_0001.jpg
+Dinora_Rosales/Dinora_Rosales_0001.jpg
+Spike_Helmick/Spike_Helmick_0001.jpg
+James_Williams/James_Williams_0001.jpg
+Faye_Wong/Faye_Wong_0001.jpg
+Marcos_Cafu/Marcos_Cafu_0001.jpg
+Henrique_Meirelles/Henrique_Meirelles_0002.jpg
+Frank_Murkowski/Frank_Murkowski_0001.jpg
+Tony_Stewart/Tony_Stewart_0001.jpg
+Mike_Samp/Mike_Samp_0001.jpg
+Colin_Phillips/Colin_Phillips_0001.jpg
+Scott_Dalton/Scott_Dalton_0001.jpg
+Karl-Heinz_Rummenigge/Karl-Heinz_Rummenigge_0001.jpg
+Tom_Tunney/Tom_Tunney_0001.jpg
+Rachel_Leigh_Cook/Rachel_Leigh_Cook_0001.jpg
+Michael_Wayne/Michael_Wayne_0001.jpg
+Chadha_Gurinder/Chadha_Gurinder_0001.jpg
+Martin_Torrijos/Martin_Torrijos_0001.jpg
+Henry_Castellanos/Henry_Castellanos_0001.jpg
+John_Rowe/John_Rowe_0001.jpg
+Mohammed_Abulhasan/Mohammed_Abulhasan_0001.jpg
+Ty_Votaw/Ty_Votaw_0001.jpg
+John_Moe/John_Moe_0001.jpg
+Chris_Tucker/Chris_Tucker_0002.jpg
+Franz_Beckenbauer/Franz_Beckenbauer_0002.jpg
+Raja_Ibrahim/Raja_Ibrahim_0001.jpg
+Maria_Burks/Maria_Burks_0001.jpg
+Kate_Moss/Kate_Moss_0001.jpg
+Amelie_Mauresmo/Amelie_Mauresmo_0015.jpg
+Raul_Mondesi/Raul_Mondesi_0001.jpg
+Erin_Brockovich/Erin_Brockovich_0001.jpg
+Bill_Clinton/Bill_Clinton_0024.jpg
+Brooke_Gordon/Brooke_Gordon_0001.jpg
+Max_Mosley/Max_Mosley_0001.jpg
+Wayne_Brady/Wayne_Brady_0001.jpg
+Erin_Runnion/Erin_Runnion_0004.jpg
+Emmanuel_Filiberto/Emmanuel_Filiberto_0001.jpg
+Gholamreza_Aghazadeh/Gholamreza_Aghazadeh_0001.jpg
+Wen_Ho_Lee/Wen_Ho_Lee_0001.jpg
+Alberto_Sordi/Alberto_Sordi_0001.jpg
+Masao_Azuma/Masao_Azuma_0001.jpg
+Fernando_Velardez/Fernando_Velardez_0001.jpg
+Adam_Mair/Adam_Mair_0001.jpg
+Kjell_Magne_Bondevik/Kjell_Magne_Bondevik_0003.jpg
+Carlos_De_Abreu/Carlos_De_Abreu_0001.jpg
+James_Gandolfini/James_Gandolfini_0001.jpg
+Drew_Gooden/Drew_Gooden_0001.jpg
+Alessandro_Nesta/Alessandro_Nesta_0001.jpg
+Fredric_Seaman/Fredric_Seaman_0001.jpg
+Bernice_Wong/Bernice_Wong_0001.jpg
+Dany_Heatley/Dany_Heatley_0001.jpg
+Tom_Miller/Tom_Miller_0001.jpg
+Sonia_Gandhi/Sonia_Gandhi_0004.jpg
+Kelly_Santos/Kelly_Santos_0001.jpg
+John_Banko/John_Banko_0002.jpg
+Robert_Gordon_Card/Robert_Gordon_Card_0001.jpg
+Martha_Burk/Martha_Burk_0001.jpg
+Sarah_Price/Sarah_Price_0001.jpg
+John_Eastman/John_Eastman_0001.jpg
+Aron_Ralston/Aron_Ralston_0001.jpg
+Jose_Jose/Jose_Jose_0001.jpg
+Gene_Hackman/Gene_Hackman_0001.jpg
+Viktor_Yushchenko/Viktor_Yushchenko_0001.jpg
+Martin_Gecht/Martin_Gecht_0001.jpg
+Tina_Fey/Tina_Fey_0001.jpg
+Aaron_Sorkin/Aaron_Sorkin_0002.jpg
+Dianne_Feinstein/Dianne_Feinstein_0001.jpg
+William_Hyde/William_Hyde_0001.jpg
+Mary_McCarty/Mary_McCarty_0001.jpg
+George_Roy_Hill/George_Roy_Hill_0002.jpg
+Hikmat_al-Azzawi/Hikmat_al-Azzawi_0001.jpg
+Dustin_Hoffman/Dustin_Hoffman_0001.jpg
+Jose_Viegas_Filho/Jose_Viegas_Filho_0002.jpg
+Hushiar_Zebari/Hushiar_Zebari_0001.jpg
+Richard_Norton-Taylor/Richard_Norton-Taylor_0002.jpg
+Nicolas_Sarkozy/Nicolas_Sarkozy_0001.jpg
+Bustam_A_Zedan_Aljanabi/Bustam_A_Zedan_Aljanabi_0001.jpg
+Romario_Farias/Romario_Farias_0001.jpg
+Mara_Georges/Mara_Georges_0001.jpg
+Brian_Heidik/Brian_Heidik_0001.jpg
+Robert_Gallo/Robert_Gallo_0001.jpg
+Ren_Qingjin/Ren_Qingjin_0001.jpg
+Dennis_Franchione/Dennis_Franchione_0001.jpg
+Pieter_Bouw/Pieter_Bouw_0001.jpg
+Marie-Josee_Croze/Marie-Josee_Croze_0001.jpg
+Peri_Gilpin/Peri_Gilpin_0001.jpg
+Bob_Alper/Bob_Alper_0001.jpg
+Robert_Woody_Johnson/Robert_Woody_Johnson_0001.jpg
+Mark_Kelly/Mark_Kelly_0001.jpg
+Michael_Chang/Michael_Chang_0003.jpg
+Andy_Warhol/Andy_Warhol_0001.jpg
+Yoon_Jin-Sik/Yoon_Jin-Sik_0001.jpg
+Pete_Gillen/Pete_Gillen_0001.jpg
+Bison_Dele/Bison_Dele_0001.jpg
+Bill_Doba/Bill_Doba_0001.jpg
+Sharon_Frey/Sharon_Frey_0002.jpg
+Adriana_Perez_Navarro/Adriana_Perez_Navarro_0001.jpg
+Tracy_McGrady/Tracy_McGrady_0001.jpg
+Joe_Metz/Joe_Metz_0001.jpg
+Will_Young/Will_Young_0001.jpg
+Jamie_King/Jamie_King_0001.jpg
+Zdravko_Mucic/Zdravko_Mucic_0001.jpg
+Mohammed_Abu_Sharia/Mohammed_Abu_Sharia_0001.jpg
+Phil_Vassar/Phil_Vassar_0001.jpg
+Carlos_Paternina/Carlos_Paternina_0001.jpg
+Ally_Sheedy/Ally_Sheedy_0001.jpg
+Reese_Witherspoon/Reese_Witherspoon_0004.jpg
+Craig_Fitzgibbon/Craig_Fitzgibbon_0001.jpg
+Tzipora_Obziler/Tzipora_Obziler_0001.jpg
+Jessica_Alba/Jessica_Alba_0001.jpg
+Charles_Bell/Charles_Bell_0001.jpg
+Ed_Smart/Ed_Smart_0001.jpg
+Brian_Kerr/Brian_Kerr_0001.jpg
+Gordana_Grubin/Gordana_Grubin_0001.jpg
+John_Reilly/John_Reilly_0001.jpg
+Tom_Amstutz/Tom_Amstutz_0001.jpg
+Nigel_Redden/Nigel_Redden_0001.jpg
+Yusuf_Misbac/Yusuf_Misbac_0001.jpg
+Saeed_Anwar/Saeed_Anwar_0001.jpg
+Carol_Niedermayer/Carol_Niedermayer_0001.jpg
+Caroline_Dhavernas/Caroline_Dhavernas_0001.jpg
+Max_Mayfield/Max_Mayfield_0002.jpg
+Daniel_Comisso_Urdaneta/Daniel_Comisso_Urdaneta_0001.jpg
+Hisashi_Owada/Hisashi_Owada_0001.jpg
+Julie_Goodenough/Julie_Goodenough_0001.jpg
+Geoff_Dixon/Geoff_Dixon_0001.jpg
+Hugh_Hefner/Hugh_Hefner_0001.jpg
+Koichi_Haraguchi/Koichi_Haraguchi_0001.jpg
+Dolma_Tsering/Dolma_Tsering_0002.jpg
+Ralph_Friedgen/Ralph_Friedgen_0001.jpg
+Ishaq_Shahryar/Ishaq_Shahryar_0001.jpg
+Biljana_Plavsic/Biljana_Plavsic_0002.jpg
+John_Howard/John_Howard_0001.jpg
+Young_Kim/Young_Kim_0001.jpg
+Farida_Ragoonanan/Farida_Ragoonanan_0001.jpg
+Kevin_Stallings/Kevin_Stallings_0001.jpg
+Danny_Avalon/Danny_Avalon_0001.jpg
+Joseph_Nacchio/Joseph_Nacchio_0001.jpg
+Phil_McGraw/Phil_McGraw_0001.jpg
+Steven_Curtis_Chapman/Steven_Curtis_Chapman_0001.jpg
+Ham_Pong-sil/Ham_Pong-sil_0001.jpg
+Johannes_Rau/Johannes_Rau_0001.jpg
+Lee_Soo-hyuck/Lee_Soo-hyuck_0002.jpg
+Luis_Gonzalez_Macchi/Luis_Gonzalez_Macchi_0002.jpg
+Jerry_Lewis/Jerry_Lewis_0001.jpg
+Ilan_Goldfajn/Ilan_Goldfajn_0001.jpg
+Juliette_Lewis/Juliette_Lewis_0001.jpg
+Gary_Paer/Gary_Paer_0001.jpg
+Santiago_Botero/Santiago_Botero_0001.jpg
+Christopher_Russell/Christopher_Russell_0001.jpg
+Rebecca_Romijn-Stamos/Rebecca_Romijn-Stamos_0003.jpg
+David_Heyman/David_Heyman_0001.jpg
+Antonio_Bernardo/Antonio_Bernardo_0001.jpg
+Shawn_Marion/Shawn_Marion_0001.jpg
+Ron_Gonzales/Ron_Gonzales_0001.jpg
+Zhang_Wenkang/Zhang_Wenkang_0001.jpg
+Michelle_Rodriguez/Michelle_Rodriguez_0002.jpg
+Jeb_Bush/Jeb_Bush_0001.jpg
+David_Kelly/David_Kelly_0001.jpg
+Linda_Sanchez/Linda_Sanchez_0001.jpg
+Mike_Cunning/Mike_Cunning_0001.jpg
+Janet_Leigh/Janet_Leigh_0001.jpg
+Tatiana_Kennedy_Schlossberg/Tatiana_Kennedy_Schlossberg_0001.jpg
+Candie_Kung/Candie_Kung_0004.jpg
+Betty_Garrison/Betty_Garrison_0001.jpg
+Jerry_Springer/Jerry_Springer_0004.jpg
+Keira_Knightley/Keira_Knightley_0002.jpg
+Isabelle_Huppert/Isabelle_Huppert_0002.jpg
+Nestor_Kirchner/Nestor_Kirchner_0023.jpg
+Willie_Nelson/Willie_Nelson_0001.jpg
+Rudy_Tomjanovich/Rudy_Tomjanovich_0001.jpg
+Isabel_Orellana/Isabel_Orellana_0001.jpg
+Nancy_Demme/Nancy_Demme_0002.jpg
+Ellen_Pompeo/Ellen_Pompeo_0001.jpg
+Al_Cardenas/Al_Cardenas_0001.jpg
+Pamela_Melroy/Pamela_Melroy_0001.jpg
+Gabriel_Valdes/Gabriel_Valdes_0001.jpg
+Ruth_Christofferson/Ruth_Christofferson_0001.jpg
+Tatsuya_Fuji/Tatsuya_Fuji_0001.jpg
+Dick_Armey/Dick_Armey_0001.jpg
+Richard_Paul_Evans/Richard_Paul_Evans_0001.jpg
+Prakash_Hinduja/Prakash_Hinduja_0001.jpg
+Richard_Langille/Richard_Langille_0001.jpg
+Mikhail_Wehbe/Mikhail_Wehbe_0003.jpg
+Colin_Campbell/Colin_Campbell_0001.jpg
+James_Hallock/James_Hallock_0001.jpg
+Strom_Thurmond/Strom_Thurmond_0001.jpg
+Hilmi_Ozkok/Hilmi_Ozkok_0002.jpg
+Dominick_Dunne/Dominick_Dunne_0001.jpg
+Pedro_Pauleta/Pedro_Pauleta_0001.jpg
+Terrell_Suggs/Terrell_Suggs_0001.jpg
+Sananda_Maitreya/Sananda_Maitreya_0001.jpg
+Michael_Kirby/Michael_Kirby_0001.jpg
+Amporn_Falise/Amporn_Falise_0001.jpg
+Anibal_Ibarra/Anibal_Ibarra_0002.jpg
+Ellen_Martin/Ellen_Martin_0001.jpg
+Jeff_Feldman/Jeff_Feldman_0001.jpg
+Chris_Klein/Chris_Klein_0001.jpg
+Jan-Michael_Gambill/Jan-Michael_Gambill_0001.jpg
+Boris_Becker/Boris_Becker_0001.jpg
+Roseanne_Barr/Roseanne_Barr_0002.jpg
+Monica_Bellucci/Monica_Bellucci_0003.jpg
+Neil_Goldman/Neil_Goldman_0001.jpg
+Ralph_Lauren/Ralph_Lauren_0001.jpg
+Fujio_Cho/Fujio_Cho_0005.jpg
+Ben_Curtis/Ben_Curtis_0002.jpg
+Carlos_Salinas/Carlos_Salinas_0001.jpg
+Laura_Flessel/Laura_Flessel_0001.jpg
+Brenda_Wilson/Brenda_Wilson_0001.jpg
+Paddy_Torsney/Paddy_Torsney_0001.jpg
+Raja_Ramani/Raja_Ramani_0001.jpg
+Jafar_Umar_Thalib/Jafar_Umar_Thalib_0001.jpg
+Don_Meredith/Don_Meredith_0001.jpg
+Calbert_Cheaney/Calbert_Cheaney_0001.jpg
+Ornella_Muti/Ornella_Muti_0001.jpg
+Kirk_Johnson/Kirk_Johnson_0003.jpg
+OJ_Simpson/OJ_Simpson_0002.jpg
+Eric_Dubin/Eric_Dubin_0001.jpg
+Paulie_Ayala/Paulie_Ayala_0001.jpg
+Pat_Summerall/Pat_Summerall_0001.jpg
+Brenda_van_Dam/Brenda_van_Dam_0001.jpg
+Gus_Frerotte/Gus_Frerotte_0001.jpg
+Nastassia_Kinski/Nastassia_Kinski_0001.jpg
+Raymond_Arthurs/Raymond_Arthurs_0001.jpg
+Francesco_Totti/Francesco_Totti_0001.jpg
+Timothy_Coughlin/Timothy_Coughlin_0001.jpg
+Zumrati_Juma/Zumrati_Juma_0001.jpg
+Yoon_Jeong_Cho/Yoon_Jeong_Cho_0001.jpg
+Ahmed_Ghazi/Ahmed_Ghazi_0001.jpg
+Daniell_Sunjata/Daniell_Sunjata_0001.jpg
+William_Rehnquist/William_Rehnquist_0001.jpg
+Dean_Jacek/Dean_Jacek_0001.jpg
+Charlie_Zaa/Charlie_Zaa_0002.jpg
+Virgina_Ruano_Pascal/Virgina_Ruano_Pascal_0001.jpg
+Michael_Leavitt/Michael_Leavitt_0001.jpg
+Brendan_Stai/Brendan_Stai_0001.jpg
+Hestrie_Cloette/Hestrie_Cloette_0001.jpg
+Tom_Scully/Tom_Scully_0001.jpg
+Ana_Guevara/Ana_Guevara_0002.jpg
+Benjamin_Bratt/Benjamin_Bratt_0001.jpg
+Lili_Marinho/Lili_Marinho_0001.jpg
+Kent_McCord/Kent_McCord_0001.jpg
+Will_Ferrell/Will_Ferrell_0001.jpg
+Andy_North/Andy_North_0001.jpg
+Eduard_Limonov/Eduard_Limonov_0001.jpg
+Roberto_Arguelles/Roberto_Arguelles_0001.jpg
+Jonathan_Tiomkin/Jonathan_Tiomkin_0001.jpg
+Lina_Krasnoroutskaya/Lina_Krasnoroutskaya_0001.jpg
+Tom_Reilly/Tom_Reilly_0001.jpg
+Patrick_Eaves/Patrick_Eaves_0001.jpg
+Robert_Douglas/Robert_Douglas_0001.jpg
+Hilda_Fortune/Hilda_Fortune_0001.jpg
+Alan_Mulally/Alan_Mulally_0002.jpg
+Michel_Duclos/Michel_Duclos_0001.jpg
+Mikulas_Dzurinda/Mikulas_Dzurinda_0002.jpg
+Brad_Gushue/Brad_Gushue_0001.jpg
+Svetoslav_Todorov/Svetoslav_Todorov_0001.jpg
+Bob_Taft/Bob_Taft_0001.jpg
+Pete_Aldridge/Pete_Aldridge_0001.jpg
+Anthony_Ervin/Anthony_Ervin_0001.jpg
+Pablo_Latras/Pablo_Latras_0001.jpg
+Hassan_Nasrallah/Hassan_Nasrallah_0001.jpg
+Khader_Rashid_Rahim/Khader_Rashid_Rahim_0001.jpg
+Georgina_Papin/Georgina_Papin_0001.jpg
+George_Harrison/George_Harrison_0001.jpg
+Barbara_Becker/Barbara_Becker_0001.jpg
+Rudolph_Holton/Rudolph_Holton_0001.jpg
+Robinson_Stevenin/Robinson_Stevenin_0001.jpg
+Larry_Beinfest/Larry_Beinfest_0001.jpg
+Stephanie_Cohen_Aloro/Stephanie_Cohen_Aloro_0001.jpg
+Paul_Coppin/Paul_Coppin_0002.jpg
+John_Rusnak/John_Rusnak_0001.jpg
+Jose_Cevallos/Jose_Cevallos_0001.jpg
+Earl_Campbell/Earl_Campbell_0001.jpg
+Susan_Whelan/Susan_Whelan_0001.jpg
+Jorge_Rodolfo_Canicoba_Corral/Jorge_Rodolfo_Canicoba_Corral_0002.jpg
+Paul_Otellini/Paul_Otellini_0001.jpg
+Mary_Jo_Myers/Mary_Jo_Myers_0001.jpg
+Nicolas_Latorre/Nicolas_Latorre_0001.jpg
+Klaus_Schwab/Klaus_Schwab_0001.jpg
+Rod_Thorn/Rod_Thorn_0001.jpg
+Kathleen_Kennedy_Townsend/Kathleen_Kennedy_Townsend_0001.jpg
+Angel_Lockward/Angel_Lockward_0001.jpg
+Janela_Jara/Janela_Jara_0001.jpg
+Dawna_LoPiccolo/Dawna_LoPiccolo_0001.jpg
+Irene_Kahn/Irene_Kahn_0001.jpg
+Constance_Marie/Constance_Marie_0002.jpg
+Stephan_Eberharter/Stephan_Eberharter_0001.jpg
+Bill_Bradley/Bill_Bradley_0001.jpg
+Regina_Ip/Regina_Ip_0001.jpg
+Miles_Stewart/Miles_Stewart_0001.jpg
+Geoffrey_Rush/Geoffrey_Rush_0001.jpg
+Hiroyuki_Yoshino/Hiroyuki_Yoshino_0001.jpg
+Mike_Fisher/Mike_Fisher_0001.jpg
+Irina_Framtsova/Irina_Framtsova_0001.jpg
+Ed_Rosenthal/Ed_Rosenthal_0002.jpg
+Jane_Kaczmarek/Jane_Kaczmarek_0002.jpg
+Isabella_Rossellini/Isabella_Rossellini_0001.jpg
+Paul_Celluci/Paul_Celluci_0001.jpg
+Ahmed_Lopez/Ahmed_Lopez_0001.jpg
+Jeffrey_Immelt/Jeffrey_Immelt_0002.jpg
+Steve_Coterill/Steve_Coterill_0001.jpg
+Thomas_Rupprath/Thomas_Rupprath_0003.jpg
+Bob_Geldof/Bob_Geldof_0002.jpg
+Henri_Proglio/Henri_Proglio_0001.jpg
+Michael_Smith_Foster/Michael_Smith_Foster_0001.jpg
+Lindsay_Lohan/Lindsay_Lohan_0001.jpg
+Luis_Gonzalez/Luis_Gonzalez_0001.jpg
+Steffeny_Holtz/Steffeny_Holtz_0001.jpg
+Robert_Bonner/Robert_Bonner_0003.jpg
+Eugene_Melnyk/Eugene_Melnyk_0001.jpg
+Noel_Niell/Noel_Niell_0001.jpg
+David_Coulthard/David_Coulthard_0001.jpg
+Laurence_Tribe/Laurence_Tribe_0001.jpg
+Jonathan_Edwards/Jonathan_Edwards_0005.jpg
+Taufik_Kiemas/Taufik_Kiemas_0001.jpg
+Jan_Pronk/Jan_Pronk_0001.jpg
+David_Duke/David_Duke_0001.jpg
+Dusty_Baker/Dusty_Baker_0001.jpg
+Phil_Johnson/Phil_Johnson_0001.jpg
+Hermann_Maier/Hermann_Maier_0001.jpg
+Lynne_Slepian/Lynne_Slepian_0001.jpg
+Alain_Cervantes/Alain_Cervantes_0001.jpg
+James_Meredeth/James_Meredeth_0001.jpg
+Lisa_Gottsegen/Lisa_Gottsegen_0001.jpg
+Manfred_Stolpe/Manfred_Stolpe_0001.jpg
+Serena_Karlan/Serena_Karlan_0001.jpg
+Ashlea_Talbot/Ashlea_Talbot_0001.jpg
+Claire_Leger/Claire_Leger_0002.jpg
+Hunter_Kemper/Hunter_Kemper_0001.jpg
+Andrei_Nikolishin/Andrei_Nikolishin_0001.jpg
+Leon_LaPorte/Leon_LaPorte_0001.jpg
+Ashanti/Ashanti_0001.jpg
+Brandon_Spann/Brandon_Spann_0001.jpg
+Deepa_Mehta/Deepa_Mehta_0001.jpg
+Tubby_Smith/Tubby_Smith_0001.jpg
+Jan_Bjoerklund/Jan_Bjoerklund_0001.jpg
+Tom_Lantos/Tom_Lantos_0001.jpg
+Gustavo_Terrazas/Gustavo_Terrazas_0001.jpg
+Andrei_Mikhnevich/Andrei_Mikhnevich_0001.jpg
+Marc_Racicot/Marc_Racicot_0001.jpg
+Andrew_Bunner/Andrew_Bunner_0002.jpg
+Robert_Nillson/Robert_Nillson_0001.jpg
+Kenny_Brack/Kenny_Brack_0001.jpg
+Lisa_Ling/Lisa_Ling_0002.jpg
+Infanta_Cristina/Infanta_Cristina_0001.jpg
+Darko_Milicic/Darko_Milicic_0001.jpg
+Laura_Romero/Laura_Romero_0001.jpg
+David_Leahy/David_Leahy_0002.jpg
+Charles_Bronson/Charles_Bronson_0001.jpg
+Richard_Greenberg/Richard_Greenberg_0001.jpg
+Joan_Collins/Joan_Collins_0001.jpg
+Chance_Mock/Chance_Mock_0001.jpg
+Jonathan_Fine/Jonathan_Fine_0001.jpg
+Jim_Harrick/Jim_Harrick_0002.jpg
+Tom_Cruise/Tom_Cruise_0008.jpg
+Allen_Iverson/Allen_Iverson_0001.jpg
+Judi_Dench/Judi_Dench_0001.jpg
+Thabo_Mbeki/Thabo_Mbeki_0005.jpg
+Naomi_Watts/Naomi_Watts_0022.jpg
+Peter_Mugyeni/Peter_Mugyeni_0001.jpg
+Robbie_Coltrane/Robbie_Coltrane_0001.jpg
+Alice_Fisher/Alice_Fisher_0001.jpg
+Kristy_Curry/Kristy_Curry_0001.jpg
+Chawki_Armali/Chawki_Armali_0001.jpg
+Luo_Linquan/Luo_Linquan_0001.jpg
+Cheryl_Ford/Cheryl_Ford_0001.jpg
+Robin_Wagner/Robin_Wagner_0001.jpg
+Khaled_Sawalhi/Khaled_Sawalhi_0001.jpg
+Donna_Shalala/Donna_Shalala_0001.jpg
+Ernie_Harwell/Ernie_Harwell_0001.jpg
+Rafidah_Aziz/Rafidah_Aziz_0001.jpg
+Cora_Cambell/Cora_Cambell_0001.jpg
+Xiang_Liu/Xiang_Liu_0001.jpg
+Eladio_Larez/Eladio_Larez_0001.jpg
+Pierre_Lacroix/Pierre_Lacroix_0001.jpg
+Gian_Marco/Gian_Marco_0002.jpg
+Clive_Woodward/Clive_Woodward_0001.jpg
+Grace_Brinell/Grace_Brinell_0001.jpg
+Stan_Kroenke/Stan_Kroenke_0001.jpg
+James_Hughes/James_Hughes_0001.jpg
+Eric_Daze/Eric_Daze_0001.jpg
+Jonathan_Woodgate/Jonathan_Woodgate_0001.jpg
+Yang_Pao-yu/Yang_Pao-yu_0001.jpg
+Eric_Staal/Eric_Staal_0001.jpg
+Anders_Ebbeson/Anders_Ebbeson_0002.jpg
+Ellen_DeGeneres/Ellen_DeGeneres_0002.jpg
+Laura_Schlessinger/Laura_Schlessinger_0001.jpg
+Matt_Siebrandt/Matt_Siebrandt_0001.jpg
+Celso_Lafer/Celso_Lafer_0001.jpg
+Scott_Sullivan/Scott_Sullivan_0002.jpg
+Cristina_Fernandez/Cristina_Fernandez_0002.jpg
+Vanessa_Incontrada/Vanessa_Incontrada_0004.jpg
+Kenenisa_Bekele/Kenenisa_Bekele_0001.jpg
+Dewayne_White/Dewayne_White_0001.jpg
+Teruaki_Masumoto/Teruaki_Masumoto_0001.jpg
+James_Roberts/James_Roberts_0001.jpg
+John_Gruden/John_Gruden_0001.jpg
+Fernando_Vargas/Fernando_Vargas_0004.jpg
+Walid_Al-Awadi/Walid_Al-Awadi_0001.jpg
+Sidney_Kimmel/Sidney_Kimmel_0001.jpg
+Begum_Khaleda_Zia/Begum_Khaleda_Zia_0001.jpg
+Daryl_Sabara/Daryl_Sabara_0001.jpg
+Steve_Largent/Steve_Largent_0001.jpg
+Bill_Self/Bill_Self_0001.jpg
+Terry_Hoeppner/Terry_Hoeppner_0001.jpg
+Dick_Bennett/Dick_Bennett_0001.jpg
+Phil_Cline/Phil_Cline_0001.jpg
+Mary_Sue_Coleman/Mary_Sue_Coleman_0001.jpg
+Chita_Rivera/Chita_Rivera_0001.jpg
+Hrithik_Roshan/Hrithik_Roshan_0001.jpg
+Michael_Schumacher/Michael_Schumacher_0001.jpg
+Marwan_Muasher/Marwan_Muasher_0001.jpg
+Harold_Brown/Harold_Brown_0001.jpg
+Gonzalo_Sanchez_de_Lozada/Gonzalo_Sanchez_de_Lozada_0009.jpg
+Alex_Corretja/Alex_Corretja_0001.jpg
+Nick_Rahall/Nick_Rahall_0001.jpg
+Carlos_Bianchi/Carlos_Bianchi_0002.jpg
+Diane_Green/Diane_Green_0002.jpg
+Dereck_Whittenburg/Dereck_Whittenburg_0001.jpg
+Tariq_Aziz/Tariq_Aziz_0001.jpg
+Roger_Staubach/Roger_Staubach_0001.jpg
+Jim_Otto/Jim_Otto_0001.jpg
+Frank_Griswold/Frank_Griswold_0001.jpg
+Andre_Lange/Andre_Lange_0001.jpg
+Renato_Soru/Renato_Soru_0001.jpg
+Khin_Nyunt/Khin_Nyunt_0001.jpg
+Kate_Winslet/Kate_Winslet_0001.jpg
+Sheila_Wellstone/Sheila_Wellstone_0002.jpg
+Jim_Ahern/Jim_Ahern_0001.jpg
+Shane_Phillips/Shane_Phillips_0001.jpg
+Ana_Isabel_Sanchez/Ana_Isabel_Sanchez_0001.jpg
+Greg_Frers/Greg_Frers_0001.jpg
+Erika_Harold/Erika_Harold_0005.jpg
+Chea_Sophara/Chea_Sophara_0001.jpg
+Thomas_Fargo/Thomas_Fargo_0004.jpg
+Willis_Roberts/Willis_Roberts_0001.jpg
+Ilan_Ramon/Ilan_Ramon_0001.jpg
+Cindy_Moll/Cindy_Moll_0001.jpg
+Elias_Attallah/Elias_Attallah_0001.jpg
+Richard_Jefferson/Richard_Jefferson_0001.jpg
+Natalie_Cole/Natalie_Cole_0002.jpg
+Ali_Abdullah_Saleh/Ali_Abdullah_Saleh_0001.jpg
+Peter_Fonda/Peter_Fonda_0001.jpg
+Peter_Camejo/Peter_Camejo_0001.jpg
+Steve_McManaman/Steve_McManaman_0001.jpg
+Joseph_Galante/Joseph_Galante_0001.jpg
+Mike_Slive/Mike_Slive_0001.jpg
+Peter_Max/Peter_Max_0001.jpg
+Javier_Zanetti/Javier_Zanetti_0001.jpg
+Kenneth_Carlsen/Kenneth_Carlsen_0001.jpg
+Paul_Gascoigne/Paul_Gascoigne_0001.jpg
+Lyle_Lovett/Lyle_Lovett_0001.jpg
+Irina_Lobacheva/Irina_Lobacheva_0001.jpg
+Peter_Fitzgerald/Peter_Fitzgerald_0001.jpg
+Kurt_Hellstrom/Kurt_Hellstrom_0001.jpg
+Justin_Gatlin/Justin_Gatlin_0001.jpg
+Derek_Jeter/Derek_Jeter_0004.jpg
+Octavio_Lara/Octavio_Lara_0001.jpg
+Dario_Camuffo/Dario_Camuffo_0001.jpg
+Prospero_Pichay/Prospero_Pichay_0001.jpg
+Teresa_Heinz_Kerry/Teresa_Heinz_Kerry_0001.jpg
+Adriana_Lima/Adriana_Lima_0001.jpg
+Mark_Mariscal/Mark_Mariscal_0001.jpg
+Curtis_Joseph/Curtis_Joseph_0001.jpg
+Amy_Yasbeck/Amy_Yasbeck_0001.jpg
+Stella_Keitel/Stella_Keitel_0001.jpg
+Mark_Hurlbert/Mark_Hurlbert_0004.jpg
+Kyra_Sedgwick/Kyra_Sedgwick_0001.jpg
+Jean_Charest/Jean_Charest_0008.jpg
+Simon_Chalk/Simon_Chalk_0001.jpg
+Cameron_Diaz/Cameron_Diaz_0004.jpg
+Thomas_Daily/Thomas_Daily_0001.jpg
+Lou_Lang/Lou_Lang_0001.jpg
+JP_Suarez/JP_Suarez_0001.jpg
+Fred_Funk/Fred_Funk_0001.jpg
+Robert_Altman/Robert_Altman_0001.jpg
+Andrew_Caldecott/Andrew_Caldecott_0001.jpg
+Eva_Marie_Saint/Eva_Marie_Saint_0001.jpg
+Mark_Butcher/Mark_Butcher_0001.jpg
+Ludivine_Sagnier/Ludivine_Sagnier_0004.jpg
+Bob_Dole/Bob_Dole_0001.jpg
+Mikhail_Khodorkovsky/Mikhail_Khodorkovsky_0001.jpg
+Will_Self/Will_Self_0001.jpg
+Hal_Sutton/Hal_Sutton_0001.jpg
+Alejandro_Lembo/Alejandro_Lembo_0001.jpg
+Christina_Aguilera/Christina_Aguilera_0004.jpg
+David_Bell/David_Bell_0001.jpg
+Chan_Choi/Chan_Choi_0001.jpg
+Marcio_de_Souza/Marcio_de_Souza_0001.jpg
+Bo_Pelini/Bo_Pelini_0001.jpg
+Hadley_Bilger/Hadley_Bilger_0001.jpg
+Pham_Sy_Chien/Pham_Sy_Chien_0001.jpg
+Jessica_Brungo/Jessica_Brungo_0001.jpg
+Mariam_Ali_Hassan/Mariam_Ali_Hassan_0001.jpg
+Emile_Lahoud/Emile_Lahoud_0001.jpg
+Joe_Lieberman/Joe_Lieberman_0001.jpg
+Barry_Bonds/Barry_Bonds_0001.jpg
+Evander_Holyfield/Evander_Holyfield_0001.jpg
+Eileen_Spina/Eileen_Spina_0001.jpg
+Jose_Bove/Jose_Bove_0001.jpg
+Gene_Sauers/Gene_Sauers_0001.jpg
+Ben_Howland/Ben_Howland_0002.jpg
+Charles_Kartman/Charles_Kartman_0001.jpg
+Ester_Canadas/Ester_Canadas_0001.jpg
+Cindy_Margolis/Cindy_Margolis_0001.jpg
+Alfredo_Moreno/Alfredo_Moreno_0001.jpg
+Aaron_Guiel/Aaron_Guiel_0001.jpg
+Micheal_Jourdain_Jr/Micheal_Jourdain_Jr_0001.jpg
+Jude_Law/Jude_Law_0001.jpg
+Vijay_Nambiar/Vijay_Nambiar_0001.jpg
+Leo_Ramirez/Leo_Ramirez_0001.jpg
+Richard_Cohen/Richard_Cohen_0001.jpg
+Richard_Sterner/Richard_Sterner_0001.jpg
+Tamara_Mowry/Tamara_Mowry_0001.jpg
+Mary_Maddux/Mary_Maddux_0001.jpg
+Daniel_Patrick_Moynihan/Daniel_Patrick_Moynihan_0001.jpg
+Dorothy_Lamour/Dorothy_Lamour_0001.jpg
+Eddie_Lucio/Eddie_Lucio_0001.jpg
+Ali_Naimi/Ali_Naimi_0008.jpg
+Prince_Naruhito/Prince_Naruhito_0001.jpg
+James_Lockhart/James_Lockhart_0001.jpg
+Amy_Redford/Amy_Redford_0001.jpg
+Jay_Rasulo/Jay_Rasulo_0002.jpg
+Nicolas_Cage/Nicolas_Cage_0004.jpg
+Carl_Levin/Carl_Levin_0001.jpg
+Zinedine_Zidane/Zinedine_Zidane_0002.jpg
+Jim_Paxson/Jim_Paxson_0001.jpg
+Bud_Selig/Bud_Selig_0001.jpg
+Juergen_Peters/Juergen_Peters_0001.jpg
+Missy_Crider/Missy_Crider_0001.jpg
+Martin_Burnham/Martin_Burnham_0001.jpg
+Jim_Hardin/Jim_Hardin_0001.jpg
+Philippe_Gagnon/Philippe_Gagnon_0001.jpg
+Perri_Shaw/Perri_Shaw_0001.jpg
+Barbara_Esbin/Barbara_Esbin_0001.jpg
+Gabriel_Farhi/Gabriel_Farhi_0001.jpg
+Mira_Sorvino/Mira_Sorvino_0001.jpg
+Scott_Yates/Scott_Yates_0001.jpg
+Charlie_Williams/Charlie_Williams_0001.jpg
+Jeff_Van_Gundy/Jeff_Van_Gundy_0002.jpg
+Laurence_Fishburne/Laurence_Fishburne_0001.jpg
+Colleen_Donovan/Colleen_Donovan_0001.jpg
+William_Pryor_Jr/William_Pryor_Jr_0001.jpg
+Annette_Lu/Annette_Lu_0002.jpg
+Brad_Miller/Brad_Miller_0001.jpg
+Ximena_Bohorquez/Ximena_Bohorquez_0001.jpg
+Giannina_Facio/Giannina_Facio_0001.jpg
+Na_Na_Keum/Na_Na_Keum_0001.jpg
+Dan_Dickau/Dan_Dickau_0001.jpg
+Lutz_Freitag/Lutz_Freitag_0001.jpg
+Kim_Hong-up/Kim_Hong-up_0001.jpg
+Aparna_Pillai/Aparna_Pillai_0001.jpg
+Christoph_Daum/Christoph_Daum_0001.jpg
+Paul_Johnson/Paul_Johnson_0001.jpg
+Bob_Hayes/Bob_Hayes_0001.jpg
+George_Murphy/George_Murphy_0001.jpg
+Kelly_Clarkson/Kelly_Clarkson_0002.jpg
+Kofi_Annan/Kofi_Annan_0022.jpg
+Jackie_Sherrill/Jackie_Sherrill_0001.jpg
+Ian_Thorpe/Ian_Thorpe_0009.jpg
+Jon_Stewart/Jon_Stewart_0001.jpg
+Mohamed_Seineldin/Mohamed_Seineldin_0001.jpg
+Chip_Burrus/Chip_Burrus_0001.jpg
+Hana_Urushima/Hana_Urushima_0001.jpg
+Mukesh_Ambani/Mukesh_Ambani_0001.jpg
+Antwun_Echols/Antwun_Echols_0001.jpg
+Fabrice_Santoro/Fabrice_Santoro_0002.jpg
+Princess_Elisabeth/Princess_Elisabeth_0001.jpg
+Nina_Pecari/Nina_Pecari_0001.jpg
+Bob_Petrino/Bob_Petrino_0001.jpg
+Ed_Book/Ed_Book_0001.jpg
+Luiz_Inacio_Lula_da_Silva/Luiz_Inacio_Lula_da_Silva_0033.jpg
+Matt_Dillon/Matt_Dillon_0002.jpg
+Vicente_Fernandez/Vicente_Fernandez_0002.jpg
+Tim_Floyd/Tim_Floyd_0001.jpg
+Franko_Simatovic/Franko_Simatovic_0001.jpg
+Leonardo_DiCaprio/Leonardo_DiCaprio_0008.jpg
+James_Brazelton/James_Brazelton_0001.jpg
+Ewan_McGregor/Ewan_McGregor_0001.jpg
+Patricia_Heaton/Patricia_Heaton_0001.jpg
+Paul_Walker/Paul_Walker_0001.jpg
+Jean-Francois_Lemounier/Jean-Francois_Lemounier_0001.jpg
+Adelina_Avila/Adelina_Avila_0001.jpg
+Eric_Vigouroux/Eric_Vigouroux_0001.jpg
+Anthony_Rackauckas/Anthony_Rackauckas_0001.jpg
+Lemuel_Montulo/Lemuel_Montulo_0001.jpg
+Jane_Rooney/Jane_Rooney_0001.jpg
+Karin_Viard/Karin_Viard_0001.jpg
+Anna_Faris/Anna_Faris_0001.jpg
+Judd_Davies/Judd_Davies_0001.jpg
+Charles_Richardson/Charles_Richardson_0001.jpg
+James_Maguire/James_Maguire_0001.jpg
+Pascal_Rheaume/Pascal_Rheaume_0001.jpg
+Barbara_Boxer/Barbara_Boxer_0001.jpg
+Lee_Tae-sik/Lee_Tae-sik_0001.jpg
+Coco_dEste/Coco_dEste_0001.jpg
+Sarah_Weddington/Sarah_Weddington_0001.jpg
+Marty_Mornhinweg/Marty_Mornhinweg_0002.jpg
+Michelle_Chiklis/Michelle_Chiklis_0001.jpg
+Lorraine_Fenton/Lorraine_Fenton_0001.jpg
+Carroll_Weimer/Carroll_Weimer_0001.jpg
+Sohail_Abbas/Sohail_Abbas_0001.jpg
+Mark_Broxmeyer/Mark_Broxmeyer_0001.jpg
+Martin_Keown/Martin_Keown_0001.jpg
+Diana_Taylor/Diana_Taylor_0001.jpg
+Sarah_Canale/Sarah_Canale_0001.jpg
+Julio_De_Brun/Julio_De_Brun_0001.jpg
+Gloria_Gaynor/Gloria_Gaynor_0001.jpg
+David_Stern/David_Stern_0002.jpg
+John_Kerr/John_Kerr_0001.jpg
+Kim_Clijsters/Kim_Clijsters_0012.jpg
+Gabrielle_Union/Gabrielle_Union_0001.jpg
+James_Franco/James_Franco_0001.jpg
+Larry_Templeton/Larry_Templeton_0001.jpg
+Chelsea_Clinton/Chelsea_Clinton_0001.jpg
+Julie_Taymor/Julie_Taymor_0001.jpg
+Steve_Lenard/Steve_Lenard_0001.jpg
+Charlie_Coles/Charlie_Coles_0001.jpg
+Valentino_Rossi/Valentino_Rossi_0001.jpg
+Amanda_Bynes/Amanda_Bynes_0004.jpg
+Thierry_Falise/Thierry_Falise_0003.jpg
+Gary_Barnett/Gary_Barnett_0001.jpg
+Gavyn_Arthur/Gavyn_Arthur_0001.jpg
+Debbie_Allen/Debbie_Allen_0001.jpg
+Claudia_Coslovich/Claudia_Coslovich_0001.jpg
+Wanda_Ilene_Barzee/Wanda_Ilene_Barzee_0001.jpg
+Patrick_Roy/Patrick_Roy_0001.jpg
+Bernard_Kerik/Bernard_Kerik_0001.jpg
+Joaquim_Levy/Joaquim_Levy_0001.jpg
+Elliott_Mincberg/Elliott_Mincberg_0001.jpg
+Momcilo_Perisic/Momcilo_Perisic_0001.jpg
+David_Caruso/David_Caruso_0003.jpg
+Tony_Clement/Tony_Clement_0001.jpg
+Kevin_Garnett/Kevin_Garnett_0001.jpg
+Larry_Anderson/Larry_Anderson_0001.jpg
+Maria_Simon/Maria_Simon_0001.jpg
+Winston_Churchill/Winston_Churchill_0001.jpg
+Yoelbi_Quesada/Yoelbi_Quesada_0001.jpg
+Bob_Crippen/Bob_Crippen_0001.jpg
+Aiysha_Smith/Aiysha_Smith_0001.jpg
+Elijan_Ingram/Elijan_Ingram_0001.jpg
+Catriona_Le_May_Doan/Catriona_Le_May_Doan_0001.jpg
+Thanongsak_Tuvinan/Thanongsak_Tuvinan_0001.jpg
+Dan_Duquette/Dan_Duquette_0001.jpg
+Selma_Phoenix/Selma_Phoenix_0001.jpg
+Jules_Asner/Jules_Asner_0001.jpg
+Jamie_Kellner/Jamie_Kellner_0001.jpg
+Hayley_Tullett/Hayley_Tullett_0002.jpg
+Phillips_Idowu/Phillips_Idowu_0001.jpg
+Melissa_Mulloy/Melissa_Mulloy_0001.jpg
+Dan_Wheldon/Dan_Wheldon_0002.jpg
+Christopher_Patten/Christopher_Patten_0001.jpg
+Rick_Bland/Rick_Bland_0001.jpg
+Duncan_Fletcher/Duncan_Fletcher_0001.jpg
+Abbas_Kiarostami/Abbas_Kiarostami_0001.jpg
+Ciaran_Hinds/Ciaran_Hinds_0001.jpg
+Mohammed_Ashraf_Hafiz/Mohammed_Ashraf_Hafiz_0001.jpg
+Bijan_Darvish/Bijan_Darvish_0003.jpg
+John_Moxley/John_Moxley_0001.jpg
+Paul_Bremer/Paul_Bremer_0002.jpg
+Suzanne_Mubarak/Suzanne_Mubarak_0001.jpg
+Roy_Williams/Roy_Williams_0003.jpg
+Jana_Henke/Jana_Henke_0001.jpg
+Luis_Fonsi/Luis_Fonsi_0001.jpg
+Tippi_Hedren/Tippi_Hedren_0002.jpg
+Tim_Jones/Tim_Jones_0001.jpg
+Kenneth_Branagh/Kenneth_Branagh_0001.jpg
+Sue_Wicks/Sue_Wicks_0002.jpg
+Jody_Richards/Jody_Richards_0001.jpg
+Carla_Tricoli/Carla_Tricoli_0001.jpg
+Princess_Anne/Princess_Anne_0001.jpg
+Mitsou_Gelinas/Mitsou_Gelinas_0001.jpg
+Chloe_Sevigny/Chloe_Sevigny_0001.jpg
+Paradorn_Srichaphan/Paradorn_Srichaphan_0008.jpg
+Gary_Carter/Gary_Carter_0001.jpg
+Nicole_Parker/Nicole_Parker_0001.jpg
+Tony_Shalhoub/Tony_Shalhoub_0001.jpg
+Brady_Rodgers/Brady_Rodgers_0001.jpg
+John_Salazar/John_Salazar_0001.jpg
+Paulo_Cesar_Pinheiro/Paulo_Cesar_Pinheiro_0002.jpg
+Bill_OReilly/Bill_OReilly_0001.jpg
+Juergen_Schrempp/Juergen_Schrempp_0001.jpg
+John_Franco/John_Franco_0001.jpg
+Tom_Kelly/Tom_Kelly_0001.jpg
+Alex_Wallau/Alex_Wallau_0001.jpg
+Hideki_Sato/Hideki_Sato_0001.jpg
+Robbie_Fowler/Robbie_Fowler_0002.jpg
+Beecher_Ray_Kirby/Beecher_Ray_Kirby_0001.jpg
+Robert_Kocharian/Robert_Kocharian_0005.jpg
+Ruth_Dreifuss/Ruth_Dreifuss_0001.jpg
+Chen_Kaige/Chen_Kaige_0001.jpg
+Jim_Wessling/Jim_Wessling_0001.jpg
+Abdel_Nasser_Assidi/Abdel_Nasser_Assidi_0001.jpg
+Azmi_Bishara/Azmi_Bishara_0001.jpg
+Guy_Ritchie/Guy_Ritchie_0002.jpg
+Michael_Patrick_King/Michael_Patrick_King_0002.jpg
+Sergei_Yushenkov/Sergei_Yushenkov_0001.jpg
+Tanya_Holyk/Tanya_Holyk_0001.jpg
+Hedayat_Amin_Arsala/Hedayat_Amin_Arsala_0001.jpg
+Laura_Gobai/Laura_Gobai_0001.jpg
+Milan_Milutinovic/Milan_Milutinovic_0001.jpg
+Kamel_Morjane/Kamel_Morjane_0001.jpg
+Saburo_Kawabuchi/Saburo_Kawabuchi_0001.jpg
+Charles_Holzner/Charles_Holzner_0001.jpg
+Michael_Boyce/Michael_Boyce_0001.jpg
+Edward_Greenspan/Edward_Greenspan_0001.jpg
+George_Bovell/George_Bovell_0001.jpg
+Eliza_Manningham-Buller/Eliza_Manningham-Buller_0001.jpg
+Vince_Gill/Vince_Gill_0001.jpg
+Kosuke_Kitajima/Kosuke_Kitajima_0002.jpg
+Elena_Dementieva/Elena_Dementieva_0001.jpg
+Penny_Dupuie/Penny_Dupuie_0001.jpg
+Johnny_Benson/Johnny_Benson_0001.jpg
+Gary_Stevens/Gary_Stevens_0001.jpg
+Chris_Hernandez/Chris_Hernandez_0001.jpg
+Ernie_Stewart/Ernie_Stewart_0001.jpg
+Rashid_Qureshi/Rashid_Qureshi_0001.jpg
+Daja_Bedanova/Daja_Bedanova_0001.jpg
+Matt_LeBlanc/Matt_LeBlanc_0001.jpg
+Miyako_Miyazaki/Miyako_Miyazaki_0001.jpg
+Elizabeth_Dole/Elizabeth_Dole_0001.jpg
+Bobo_Balde/Bobo_Balde_0001.jpg
+Hamid_Reza_Asefi/Hamid_Reza_Asefi_0001.jpg
+Keith_Foulke/Keith_Foulke_0001.jpg
+Henry_Suazo/Henry_Suazo_0001.jpg
+Ralph_Sampson/Ralph_Sampson_0001.jpg
+Rosario_Dawson/Rosario_Dawson_0001.jpg
+TA_McLendon/TA_McLendon_0001.jpg
+Sebastian_Cuattrin/Sebastian_Cuattrin_0001.jpg
+Mstislav_Rostropovich/Mstislav_Rostropovich_0001.jpg
+Wang_Nan/Wang_Nan_0001.jpg
+Eileen_Coparropa/Eileen_Coparropa_0002.jpg
+Charles_Schumer/Charles_Schumer_0002.jpg
+Fred_Huff/Fred_Huff_0001.jpg
+Mario_Gallegos/Mario_Gallegos_0001.jpg
+Guido_Westerwelle/Guido_Westerwelle_0001.jpg
+Seydou_Diarra/Seydou_Diarra_0001.jpg
+Jeff_Dederian/Jeff_Dederian_0001.jpg
+Ray_Morrough/Ray_Morrough_0001.jpg
+Tony_Bennett/Tony_Bennett_0004.jpg
+Lars_Von_Trier/Lars_Von_Trier_0001.jpg
+Gwyneth_Paltrow/Gwyneth_Paltrow_0001.jpg
+Qazi_Hussain_Ahmed/Qazi_Hussain_Ahmed_0001.jpg
+Peter_Greenaway/Peter_Greenaway_0001.jpg
+Fidel_Castro_Daiz-Balart/Fidel_Castro_Daiz-Balart_0001.jpg
+Aldo_Paredes/Aldo_Paredes_0001.jpg
+Charley_Armey/Charley_Armey_0001.jpg
+Oscar_DLeon/Oscar_DLeon_0001.jpg
+Larry_Campbell/Larry_Campbell_0001.jpg
+Roger_Suarez/Roger_Suarez_0001.jpg
+Trevor_Watson/Trevor_Watson_0001.jpg
+Bernard_Law/Bernard_Law_0001.jpg
+Prince_Felipe/Prince_Felipe_0001.jpg
+Ismael_Miranda/Ismael_Miranda_0001.jpg
+Lee_Hong-ki/Lee_Hong-ki_0001.jpg
+Gerard_Kleisterlee/Gerard_Kleisterlee_0001.jpg
+Cedric_Benson/Cedric_Benson_0001.jpg
+Ben_Glisan/Ben_Glisan_0001.jpg
+Florencia_Macri/Florencia_Macri_0001.jpg
+Oxana_Fedorova/Oxana_Fedorova_0003.jpg
+Mario_Lemieux/Mario_Lemieux_0001.jpg
+John_F_Kennedy_Jr/John_F_Kennedy_Jr_0002.jpg
+Danny_Morgan/Danny_Morgan_0001.jpg
+Marina_Canetti/Marina_Canetti_0001.jpg
+Valery_Giscard_dEstaing/Valery_Giscard_dEstaing_0005.jpg
+Ghassan_Elashi/Ghassan_Elashi_0001.jpg
+Saied_Hadi_al_Mudarissi/Saied_Hadi_al_Mudarissi_0001.jpg
+Brad_Garrett/Brad_Garrett_0001.jpg
+Eduardo_Romero/Eduardo_Romero_0001.jpg
+Kurt_Russell/Kurt_Russell_0002.jpg
+Eve_Ensler/Eve_Ensler_0001.jpg
+Andy_Griffith/Andy_Griffith_0001.jpg
+Ralph_Nader/Ralph_Nader_0001.jpg
+Carlos_Menem/Carlos_Menem_0021.jpg
+Cristian_Barros/Cristian_Barros_0001.jpg
+Allan_Wagner/Allan_Wagner_0001.jpg
+Tom_Coverdale/Tom_Coverdale_0002.jpg
+Carlos_Lordkipanitse/Carlos_Lordkipanitse_0001.jpg
+Yingfan_Wang/Yingfan_Wang_0001.jpg
+Dean_Sheremet/Dean_Sheremet_0001.jpg
+Mike_Leach/Mike_Leach_0001.jpg
+Baz_Luhrmann/Baz_Luhrmann_0001.jpg
+Derrick_Taylor/Derrick_Taylor_0001.jpg
+Kamal_Kharrazi/Kamal_Kharrazi_0002.jpg
+Joshua_Perper/Joshua_Perper_0001.jpg
+Sonia_Lopez/Sonia_Lopez_0001.jpg
+Mark_Sacco/Mark_Sacco_0001.jpg
+Cecilia_Chang/Cecilia_Chang_0001.jpg
+Jim_Ryan/Jim_Ryan_0001.jpg
+Catherine_Zeta-Jones/Catherine_Zeta-Jones_0009.jpg
+Mike_Alden/Mike_Alden_0001.jpg
+Laura_Elena_Harring/Laura_Elena_Harring_0001.jpg
+Cho_Myung-kyun/Cho_Myung-kyun_0001.jpg
+David_Wolf/David_Wolf_0002.jpg
+Devin_Harris/Devin_Harris_0001.jpg
+Sergei_Yastrzhembsky/Sergei_Yastrzhembsky_0001.jpg
+Steve_Ballmer/Steve_Ballmer_0001.jpg
+Chistian_Stahl/Chistian_Stahl_0001.jpg
+Camille_Colvin/Camille_Colvin_0001.jpg
+Judith_Nathan/Judith_Nathan_0001.jpg
+Anthony_Carter/Anthony_Carter_0001.jpg
+Joy_Bryant/Joy_Bryant_0001.jpg
+Stephen_Push/Stephen_Push_0001.jpg
+Tia_Mowry/Tia_Mowry_0001.jpg
+Abraham_Foxman/Abraham_Foxman_0001.jpg
+Rick_Romley/Rick_Romley_0001.jpg
+Martin_Cauchon/Martin_Cauchon_0002.jpg
+Nobuyuki_Idei/Nobuyuki_Idei_0001.jpg
+Bill_Herrion/Bill_Herrion_0001.jpg
+Jeff_Bzdelik/Jeff_Bzdelik_0001.jpg
+Toutai_Kefu/Toutai_Kefu_0001.jpg
+Harald_Ringstorff/Harald_Ringstorff_0001.jpg
+Patrick_Coleman/Patrick_Coleman_0001.jpg
+Ray_Liotta/Ray_Liotta_0001.jpg
+Katie_Harman/Katie_Harman_0002.jpg
+Jean-Sebastien_Giguere/Jean-Sebastien_Giguere_0002.jpg
+Robert_Nardelli/Robert_Nardelli_0001.jpg
+Nick_Nolte/Nick_Nolte_0003.jpg
+Matthew_Broderick/Matthew_Broderick_0001.jpg
+Valerie_Harper/Valerie_Harper_0002.jpg
+Ann_Godbehere/Ann_Godbehere_0001.jpg
+Donatella_Versace/Donatella_Versace_0002.jpg
+Richard_Crenna/Richard_Crenna_0002.jpg
+Raul_Cubas/Raul_Cubas_0001.jpg
+Jim_Schwarz/Jim_Schwarz_0001.jpg
+Collis_Temple_III/Collis_Temple_III_0001.jpg
+Mary_Steenburgen/Mary_Steenburgen_0001.jpg
+Robert_Lange/Robert_Lange_0001.jpg
+Trudi_Lacey/Trudi_Lacey_0001.jpg
+Goran_Persson/Goran_Persson_0002.jpg
+Eric_Ryan_Donnelly/Eric_Ryan_Donnelly_0001.jpg
+Alexandra_Spann/Alexandra_Spann_0001.jpg
+Juan_Ignacio_Chela/Juan_Ignacio_Chela_0001.jpg
+Buzz_Hargrove/Buzz_Hargrove_0001.jpg
+Robert_Torricelli/Robert_Torricelli_0003.jpg
+Andrew_Sabey/Andrew_Sabey_0001.jpg
+Anna_Jones/Anna_Jones_0001.jpg
+Jeffrey_Jones/Jeffrey_Jones_0001.jpg
+Nancy_Kerrigan/Nancy_Kerrigan_0001.jpg
+Iveta_Benesova/Iveta_Benesova_0001.jpg
+Paul_Wellstone/Paul_Wellstone_0003.jpg
+Mahdi_Al_Bassam/Mahdi_Al_Bassam_0001.jpg
+Marcella_Anderson/Marcella_Anderson_0001.jpg
+Madeleine_Webber/Madeleine_Webber_0001.jpg
+Mario_Dumont/Mario_Dumont_0001.jpg
+Daryl_Jones/Daryl_Jones_0001.jpg
+Juan_Carlos/Juan_Carlos_0001.jpg
+Linda_Dano/Linda_Dano_0001.jpg
+Leigh_Winchell/Leigh_Winchell_0001.jpg
+Anthony_Lee_Johnson/Anthony_Lee_Johnson_0001.jpg
+Brian_Grazier/Brian_Grazier_0001.jpg
+Yossi_Beilin/Yossi_Beilin_0002.jpg
+Darlene_Garrettson/Darlene_Garrettson_0001.jpg
+Alex_Holmes/Alex_Holmes_0001.jpg
+Alicia_Keys/Alicia_Keys_0001.jpg
+Marina_Anissina/Marina_Anissina_0001.jpg
+Larenz_Tate/Larenz_Tate_0001.jpg
+Brian_Schneider/Brian_Schneider_0001.jpg
+Porter_Goss/Porter_Goss_0001.jpg
+Oswaldo_Paya/Oswaldo_Paya_0005.jpg
+Mary_Anne_Souza/Mary_Anne_Souza_0001.jpg
+Charles_Ingram/Charles_Ingram_0001.jpg
+Abba_Eban/Abba_Eban_0001.jpg
+Ramon_Santana/Ramon_Santana_0001.jpg
+Max_Biaggi/Max_Biaggi_0001.jpg
+Maggie_Cheung/Maggie_Cheung_0001.jpg
+Ricardo_Mayorga/Ricardo_Mayorga_0001.jpg
+Bruna_Colosio/Bruna_Colosio_0001.jpg
+Mario_Alfaro-Lopez/Mario_Alfaro-Lopez_0001.jpg
+Guillermo_Canas/Guillermo_Canas_0001.jpg
+Danis_Tanovic/Danis_Tanovic_0001.jpg
+Jeremy_Fogel/Jeremy_Fogel_0001.jpg
+Marricia_Tate/Marricia_Tate_0001.jpg
+Valerie_Thwaites/Valerie_Thwaites_0001.jpg
+Keith_Tyson/Keith_Tyson_0001.jpg
+LK_Advani/LK_Advani_0002.jpg
+Wilson_Alvarez/Wilson_Alvarez_0001.jpg
+Marwan_Barghouthi/Marwan_Barghouthi_0001.jpg
+Alek_Wek/Alek_Wek_0001.jpg
+Tina_Andrews/Tina_Andrews_0001.jpg
+Vicente_Fox_de_la_Concha/Vicente_Fox_de_la_Concha_0001.jpg
+Billy_Andrade/Billy_Andrade_0001.jpg
+Noah_Wyle/Noah_Wyle_0001.jpg
+Dariusz_Michalczewski/Dariusz_Michalczewski_0001.jpg
+Max_von_Sydow/Max_von_Sydow_0001.jpg
+Erin_Hershey_Presley/Erin_Hershey_Presley_0001.jpg
+Prem_Kumar_Nair/Prem_Kumar_Nair_0001.jpg
+Bob_Menendez/Bob_Menendez_0001.jpg
+Nathan_Smith/Nathan_Smith_0001.jpg
+Scott_Ritter/Scott_Ritter_0002.jpg
+Victoria_Beckham/Victoria_Beckham_0001.jpg
+Thomas_Birmingham/Thomas_Birmingham_0001.jpg
+Takeo_Fukui/Takeo_Fukui_0001.jpg
+Barrett_Jackman/Barrett_Jackman_0001.jpg
+Jacob_Frenkel/Jacob_Frenkel_0001.jpg
+Benicio_Del_Toro/Benicio_Del_Toro_0001.jpg
+Barry_Hinson/Barry_Hinson_0001.jpg
+Todd_Wike/Todd_Wike_0001.jpg
+Kate_Richardson/Kate_Richardson_0001.jpg
+Ann_Veneman/Ann_Veneman_0005.jpg
+James_Wallack/James_Wallack_0001.jpg
+Edward_Kennedy/Edward_Kennedy_0001.jpg
+Chanda_Rubin/Chanda_Rubin_0005.jpg
+Chhouk_Rin/Chhouk_Rin_0001.jpg
+Richard_Gere/Richard_Gere_0003.jpg
+Frank_Bell/Frank_Bell_0001.jpg
+David_Modell/David_Modell_0001.jpg
+Jennifer_Reilly/Jennifer_Reilly_0001.jpg
+Aicha_El_Ouafi/Aicha_El_Ouafi_0003.jpg
+Robin_McGraw/Robin_McGraw_0001.jpg
+Habib_Hisham/Habib_Hisham_0001.jpg
+Cabas/Cabas_0001.jpg
+Dan_Morales/Dan_Morales_0003.jpg
+David_McKiernan/David_McKiernan_0001.jpg
+Jakob_Kellenberger/Jakob_Kellenberger_0001.jpg
+Lili_Taylor/Lili_Taylor_0002.jpg
+Patty_Sheehan/Patty_Sheehan_0001.jpg
+Mother_Teresa/Mother_Teresa_0001.jpg
+Douglas_Paal/Douglas_Paal_0001.jpg
+Mehdi_Baala/Mehdi_Baala_0001.jpg
+Wallace_Capel/Wallace_Capel_0001.jpg
+Suzanne_Gaudet/Suzanne_Gaudet_0001.jpg
+Jim_Tressel/Jim_Tressel_0004.jpg
+Robert_Blake/Robert_Blake_0003.jpg
+Abdullah_Nasseef/Abdullah_Nasseef_0001.jpg
+Carol_Carmody/Carol_Carmody_0001.jpg
+Ben_Cahoon/Ben_Cahoon_0001.jpg
+Carla_Gugino/Carla_Gugino_0001.jpg
+Bobby_Goldwater/Bobby_Goldwater_0002.jpg
+Jason_Gardner/Jason_Gardner_0001.jpg
+Mary_Lou_Retton/Mary_Lou_Retton_0001.jpg
+Mark_Polansky/Mark_Polansky_0001.jpg
+John_Lisowski/John_Lisowski_0001.jpg
+Mukhtar_Alytnbayev/Mukhtar_Alytnbayev_0001.jpg
+Steve_Case/Steve_Case_0001.jpg
+Vincent_Spadea/Vincent_Spadea_0001.jpg
+Rhett_Warrener/Rhett_Warrener_0001.jpg
+Wolfgang_Schuessel/Wolfgang_Schuessel_0001.jpg
+Hunter_Bates/Hunter_Bates_0001.jpg
+Harry_Belafonte/Harry_Belafonte_0002.jpg
+John_Marburger/John_Marburger_0001.jpg
+Romeo_Gigli/Romeo_Gigli_0001.jpg
+Michael_Pfleger/Michael_Pfleger_0001.jpg
+Tamara_Brooks/Tamara_Brooks_0002.jpg
+Jane_Krakowski/Jane_Krakowski_0001.jpg
+Brandon_Jones/Brandon_Jones_0001.jpg
+Tom_Koenigs/Tom_Koenigs_0001.jpg
+Stephen_Swindal/Stephen_Swindal_0001.jpg
+Yasein_Taher/Yasein_Taher_0001.jpg
+Gordon_Campbell/Gordon_Campbell_0001.jpg
+Nila_Ferran/Nila_Ferran_0001.jpg
+David_Surrett/David_Surrett_0001.jpg
+Gary_Bauer/Gary_Bauer_0001.jpg
+Morgan_Freeman/Morgan_Freeman_0001.jpg
+Imre_Kertasz/Imre_Kertasz_0001.jpg
+Lee_Chang-dong/Lee_Chang-dong_0001.jpg
+Yasushi_Chimura/Yasushi_Chimura_0001.jpg
+Mike_Krzyzewski/Mike_Krzyzewski_0005.jpg
+Dave_Barr/Dave_Barr_0001.jpg
+Alan_Stonecipher/Alan_Stonecipher_0001.jpg
+Brian_Billick/Brian_Billick_0001.jpg
+Sergio_Garcia/Sergio_Garcia_0001.jpg
+Vladimir_Spidla/Vladimir_Spidla_0002.jpg
+Helio_Castroneves/Helio_Castroneves_0001.jpg
+Queen_Rania/Queen_Rania_0005.jpg
+Holly_Robinson_Peete/Holly_Robinson_Peete_0001.jpg
+Jamie_Dimon/Jamie_Dimon_0001.jpg
+Brian_Campbell/Brian_Campbell_0001.jpg
+Douglas_Meester/Douglas_Meester_0001.jpg
+Demi_Moore/Demi_Moore_0004.jpg
+Ringo_Starr/Ringo_Starr_0001.jpg
+Takaloo/Takaloo_0001.jpg
+Ciro_Gomes/Ciro_Gomes_0005.jpg
+Peter_Ueberroth/Peter_Ueberroth_0001.jpg
+Festus_Mogae/Festus_Mogae_0001.jpg
+Megan_Mullally/Megan_Mullally_0003.jpg
+Sharon_Stone/Sharon_Stone_0001.jpg
+Gina_Gershon/Gina_Gershon_0001.jpg
+Jimmy_Gobble/Jimmy_Gobble_0001.jpg
+Tom_Coughlin/Tom_Coughlin_0001.jpg
+Petria_Thomas/Petria_Thomas_0002.jpg
+Gordon_Cooper/Gordon_Cooper_0001.jpg
+Kelli_White/Kelli_White_0002.jpg
+Bonnie_Hunt/Bonnie_Hunt_0001.jpg
+Annie-Jeanne_Reynaud/Annie-Jeanne_Reynaud_0001.jpg
+Odilia_Collazo/Odilia_Collazo_0001.jpg
+Maryn_McKenna/Maryn_McKenna_0001.jpg
+Helene_Eksterowicz/Helene_Eksterowicz_0001.jpg
+Fiona_Milne/Fiona_Milne_0001.jpg
+Armando_Carrillo/Armando_Carrillo_0001.jpg
+Odai_Hussein/Odai_Hussein_0001.jpg
+Wilton_Gregory/Wilton_Gregory_0001.jpg
+Thomas_Kelly/Thomas_Kelly_0001.jpg
+Suh_Chung-won/Suh_Chung-won_0001.jpg
+Barry_Nakell/Barry_Nakell_0001.jpg
+Kevin_Tarrant/Kevin_Tarrant_0001.jpg
+Artieas_Shanks/Artieas_Shanks_0001.jpg
+Ben_Davis/Ben_Davis_0001.jpg
+Toshihiko_Fukui/Toshihiko_Fukui_0002.jpg
+Roger_Machado/Roger_Machado_0001.jpg
+Dominik_Garcia-Lorido/Dominik_Garcia-Lorido_0002.jpg
+Monica_Seles/Monica_Seles_0003.jpg
+Atiabet_Ijan_Amabel/Atiabet_Ijan_Amabel_0001.jpg
+Adam_Herbert/Adam_Herbert_0001.jpg
+Victor_Hanescu/Victor_Hanescu_0001.jpg
+David_Millar/David_Millar_0001.jpg
+Richard_Tubb/Richard_Tubb_0001.jpg
+Perry_Compton/Perry_Compton_0001.jpg
+Kevin_Harvick/Kevin_Harvick_0001.jpg
+Faisal_Saleh_Hayat/Faisal_Saleh_Hayat_0001.jpg
+Barry_Ford/Barry_Ford_0001.jpg
+Michelle_Collins/Michelle_Collins_0002.jpg
+Robert_Kipkoech_Cheruiyot/Robert_Kipkoech_Cheruiyot_0001.jpg
+James_Ivory/James_Ivory_0001.jpg
+Hichiro_Naemura/Hichiro_Naemura_0001.jpg
+Sylvie_Guillem/Sylvie_Guillem_0001.jpg
+Karin_Pilsaeter/Karin_Pilsaeter_0001.jpg
+Johnny_Carson/Johnny_Carson_0002.jpg
+Mark_Hanson/Mark_Hanson_0001.jpg
+Samira_Makhmalbaf/Samira_Makhmalbaf_0002.jpg
+Mitchell_Crooks/Mitchell_Crooks_0001.jpg
+Yang_Hee_Kim/Yang_Hee_Kim_0001.jpg
+Lisa_Raymond/Lisa_Raymond_0001.jpg
+Pierce_Brosnan/Pierce_Brosnan_0004.jpg
+Stanley_Tong/Stanley_Tong_0002.jpg
+Hugh_Campbell/Hugh_Campbell_0001.jpg
+Pharrell_Williams/Pharrell_Williams_0001.jpg
+Ion_Iliescu/Ion_Iliescu_0001.jpg
+John_Cruz/John_Cruz_0001.jpg
+Lindsay_Benko/Lindsay_Benko_0002.jpg
+Raquel_Welch/Raquel_Welch_0002.jpg
+Michael_Bolton/Michael_Bolton_0001.jpg
+Bixente_LIzarazu/Bixente_LIzarazu_0001.jpg
+Anthony_Fauci/Anthony_Fauci_0001.jpg
+Razali_Ismail/Razali_Ismail_0001.jpg
+James_Wattana/James_Wattana_0001.jpg
+Michelle_Bachelet/Michelle_Bachelet_0001.jpg
+Joaquim_Rodriguez/Joaquim_Rodriguez_0001.jpg
+Iain_Duncan_Smith/Iain_Duncan_Smith_0001.jpg
+Jaap_de_Hoop_Scheffer/Jaap_de_Hoop_Scheffer_0001.jpg
+Yukio_Hatoyama/Yukio_Hatoyama_0001.jpg
+Juan_Jose_Lucas/Juan_Jose_Lucas_0001.jpg
+Laura_Marlow/Laura_Marlow_0001.jpg
+Ben_Broussard/Ben_Broussard_0001.jpg
+Mike_Montgomery/Mike_Montgomery_0002.jpg
+Bob_Holden/Bob_Holden_0001.jpg
+Keith_Olbermann/Keith_Olbermann_0001.jpg
+Joxel_Garcia/Joxel_Garcia_0001.jpg
+Elizabeth_Hill/Elizabeth_Hill_0001.jpg
+Monique_Garbrecht-Enfeldt/Monique_Garbrecht-Enfeldt_0001.jpg
+Denise_Locke/Denise_Locke_0001.jpg
+Prince_Charles/Prince_Charles_0002.jpg
+Saeed_Mortazavi/Saeed_Mortazavi_0001.jpg
+Alina_Kabaeva/Alina_Kabaeva_0001.jpg
+Ben_Betts/Ben_Betts_0001.jpg
+Alex_Barros/Alex_Barros_0001.jpg
+Mike_Thibault/Mike_Thibault_0001.jpg
+Mohammad_Al-Sharief/Mohammad_Al-Sharief_0001.jpg
+Maria_Luisa_Mendonca/Maria_Luisa_Mendonca_0002.jpg
+Anzori_Kikalishvili/Anzori_Kikalishvili_0001.jpg
+James_Becker/James_Becker_0001.jpg
+Darcy_Regier/Darcy_Regier_0001.jpg
+Hamad_Bin_Jassim/Hamad_Bin_Jassim_0001.jpg
+Yao_Ming/Yao_Ming_0006.jpg
+Jason_White/Jason_White_0001.jpg
+Lily_Safra/Lily_Safra_0001.jpg
+Barzan_al-Tikriti/Barzan_al-Tikriti_0001.jpg
+Ron_Kirk/Ron_Kirk_0001.jpg
+Oscar_De_La_Hoya/Oscar_De_La_Hoya_0003.jpg
+Sarah_Hughes/Sarah_Hughes_0002.jpg
+John_Mabry/John_Mabry_0001.jpg
+Diana_Renee_Valdivieso_Dubon/Diana_Renee_Valdivieso_Dubon_0001.jpg
+Juan_Antonio_Samaranch/Juan_Antonio_Samaranch_0001.jpg
+Ruth_Bader_Ginsburg/Ruth_Bader_Ginsburg_0001.jpg
+Eric_Lindros/Eric_Lindros_0001.jpg
+Martie_Maguire/Martie_Maguire_0001.jpg
+Gina_Lollobrigida/Gina_Lollobrigida_0001.jpg
+Stella_McCartney/Stella_McCartney_0001.jpg
+Edward_Burns/Edward_Burns_0001.jpg
+Jimmy_Jimenez/Jimmy_Jimenez_0001.jpg
+Rand_Miller/Rand_Miller_0001.jpg
+Sharess_Harrell/Sharess_Harrell_0001.jpg
+Tangra_Riggle/Tangra_Riggle_0001.jpg
+Ernie_Preate/Ernie_Preate_0001.jpg
+Soon_Yi/Soon_Yi_0001.jpg
+Fabian_Vargas/Fabian_Vargas_0001.jpg
+Carlos_Ruckauf/Carlos_Ruckauf_0001.jpg
+Jackie_Chan/Jackie_Chan_0009.jpg
+Leonardo_Fernandez/Leonardo_Fernandez_0001.jpg
+Richard_Carl/Richard_Carl_0001.jpg
+Lech_Walesa/Lech_Walesa_0002.jpg
+Jose_Sarney/Jose_Sarney_0003.jpg
+Denis_Coderre/Denis_Coderre_0001.jpg
+Keith_Fotta/Keith_Fotta_0001.jpg
+Tim_Curry/Tim_Curry_0002.jpg
+Qusai_Hussein/Qusai_Hussein_0001.jpg
+Anthony_Garotinho/Anthony_Garotinho_0001.jpg
+Larry_Ralston/Larry_Ralston_0001.jpg
+Jim_Fassel/Jim_Fassel_0001.jpg
+Joanne_Duquette/Joanne_Duquette_0001.jpg
+Kim_Yun-kyu/Kim_Yun-kyu_0001.jpg
+Luiz_Felipe_Scolari/Luiz_Felipe_Scolari_0002.jpg
+Phil_Jackson/Phil_Jackson_0001.jpg
+Saman_Shali/Saman_Shali_0001.jpg
+Iva_Majoli/Iva_Majoli_0001.jpg
+Tracy_Wyle/Tracy_Wyle_0001.jpg
+Mohammed_Al_Hindi/Mohammed_Al_Hindi_0001.jpg
+Christine_Baumgartner/Christine_Baumgartner_0005.jpg
+Wu_Yi/Wu_Yi_0002.jpg
+Gina_Torres/Gina_Torres_0001.jpg
+Anne_McLellan/Anne_McLellan_0003.jpg
+Tony_LaRussa/Tony_LaRussa_0001.jpg
+Milton_Berle/Milton_Berle_0001.jpg
+Edward_Egan/Edward_Egan_0001.jpg
+Eric_Idle/Eric_Idle_0001.jpg
+Tom_Crean/Tom_Crean_0001.jpg
+Monica_Serra/Monica_Serra_0001.jpg
+Hank_Stram/Hank_Stram_0001.jpg
+George_Galloway/George_Galloway_0004.jpg
+Christine_Arron/Christine_Arron_0001.jpg
+Kong_Quan/Kong_Quan_0001.jpg
+Chris_Crocker/Chris_Crocker_0001.jpg
+Aleksander_Kwasniewski/Aleksander_Kwasniewski_0001.jpg
+Emmit_Smith/Emmit_Smith_0001.jpg
+Nicklas_Lidstrom/Nicklas_Lidstrom_0001.jpg
+Yoo_Jay-Kun/Yoo_Jay-Kun_0001.jpg
+Mohammed_Salmane/Mohammed_Salmane_0001.jpg
+Mike_Szymanczyk/Mike_Szymanczyk_0001.jpg
+Michael_Chertoff/Michael_Chertoff_0001.jpg
+Chris_Terry/Chris_Terry_0001.jpg
+Hubert_Green/Hubert_Green_0001.jpg
+Sean_Townsend/Sean_Townsend_0001.jpg
+Ted_Turner/Ted_Turner_0001.jpg
+Dave_McNally/Dave_McNally_0001.jpg
+James_Meeks/James_Meeks_0001.jpg
+Freda_Black/Freda_Black_0001.jpg
+Milo_Djukanovic/Milo_Djukanovic_0001.jpg
+Ronald_Post/Ronald_Post_0001.jpg
+Hilary_Duff/Hilary_Duff_0003.jpg
+Ramon_Delgado/Ramon_Delgado_0001.jpg
+Fred_Swan/Fred_Swan_0001.jpg
+Ryan_Goodman/Ryan_Goodman_0001.jpg
+Mike_Farrar/Mike_Farrar_0001.jpg
+Stephen_Friedman/Stephen_Friedman_0001.jpg
+Sami_Al-Arian/Sami_Al-Arian_0001.jpg
+Kathy_Bates/Kathy_Bates_0001.jpg
+Eliane_Karp/Eliane_Karp_0003.jpg
+Rolf_Zimmermann/Rolf_Zimmermann_0001.jpg
+Christopher_Walken/Christopher_Walken_0004.jpg
+Paul_Patton/Paul_Patton_0002.jpg
+Jodie_Foster/Jodie_Foster_0001.jpg
+Thaksin_Shinawatra/Thaksin_Shinawatra_0003.jpg
+Michael_Kahn/Michael_Kahn_0001.jpg
+Rachel_Wheatley/Rachel_Wheatley_0001.jpg
+Abdel_Aziz_Al-Hakim/Abdel_Aziz_Al-Hakim_0001.jpg
+Arnaud_Clement/Arnaud_Clement_0002.jpg
+Natasa_Micic/Natasa_Micic_0001.jpg
+Peter_Albertsen/Peter_Albertsen_0001.jpg
+Erika_Christensen/Erika_Christensen_0001.jpg
+Danny_Elfman/Danny_Elfman_0001.jpg
+Rita_Grande/Rita_Grande_0002.jpg
+Gary_Winnick/Gary_Winnick_0001.jpg
+Laurel_Clark/Laurel_Clark_0001.jpg
+Jean-Luc_Bideau/Jean-Luc_Bideau_0001.jpg
+Randall_Terry/Randall_Terry_0001.jpg
+Miranda_Otto/Miranda_Otto_0001.jpg
+Park_Na-kyong/Park_Na-kyong_0001.jpg
+Hideki_Matsui/Hideki_Matsui_0002.jpg
+Donna_Barrera/Donna_Barrera_0001.jpg
+Alexandre_Vinokourov/Alexandre_Vinokourov_0001.jpg
+Tim_Pawlenty/Tim_Pawlenty_0001.jpg
+Luca_Cordero_di_Montezemolo/Luca_Cordero_di_Montezemolo_0001.jpg
+Vladimir_Meciar/Vladimir_Meciar_0001.jpg
+John_Swofford/John_Swofford_0001.jpg
+Muhammad_Ibrahim_Bilal/Muhammad_Ibrahim_Bilal_0001.jpg
+Winona_Ryder/Winona_Ryder_0024.jpg
+Micky_Arison/Micky_Arison_0001.jpg
+Tom_Brady/Tom_Brady_0001.jpg
+Darrell_Issa/Darrell_Issa_0001.jpg
+Mark_Dacey/Mark_Dacey_0001.jpg
+John_White/John_White_0001.jpg
+Dave_Potter/Dave_Potter_0001.jpg
+Richard_Gephardt/Richard_Gephardt_0001.jpg
+Phil_Morris/Phil_Morris_0001.jpg
+Miguel_Cotto/Miguel_Cotto_0001.jpg
+Kitin_Munoz/Kitin_Munoz_0001.jpg
+Susan_Sarandon/Susan_Sarandon_0003.jpg
+Nadine_Vinzens/Nadine_Vinzens_0001.jpg
+Emma_Watson/Emma_Watson_0001.jpg
+Tom_Ridge/Tom_Ridge_0010.jpg
+Robert_DeFraites/Robert_DeFraites_0001.jpg
+Ronald_Perelman/Ronald_Perelman_0001.jpg
+Dennis_Kucinich/Dennis_Kucinich_0006.jpg
+Carlo_Ancelotti/Carlo_Ancelotti_0002.jpg
+Bill_Carmody/Bill_Carmody_0001.jpg
+Gordon_Lightfoot/Gordon_Lightfoot_0001.jpg
+Najib_al-Salhi/Najib_al-Salhi_0001.jpg
+Liam_Neeson/Liam_Neeson_0001.jpg
+AnFernce_Negron/AnFernce_Negron_0001.jpg
+Daisy_Fuentes/Daisy_Fuentes_0001.jpg
+David_Alpay/David_Alpay_0001.jpg
+Rodney_Dangerfield/Rodney_Dangerfield_0001.jpg
+Mitchell_Daniels/Mitchell_Daniels_0003.jpg
+John_Paul_DeJoria/John_Paul_DeJoria_0001.jpg
+Mohammed_Mehdi_Saleh/Mohammed_Mehdi_Saleh_0001.jpg
+Luis_Sanchez/Luis_Sanchez_0001.jpg
+Greg_Ostertag/Greg_Ostertag_0002.jpg
+Zavad_Zarif/Zavad_Zarif_0001.jpg
+Khalid_Khannouchi/Khalid_Khannouchi_0001.jpg
+Omar_Sharif/Omar_Sharif_0001.jpg
+Jean_Chretien/Jean_Chretien_0026.jpg
+Carly_Gullickson/Carly_Gullickson_0001.jpg
+Ron_Lantz/Ron_Lantz_0001.jpg
+Robert_Morvillo/Robert_Morvillo_0001.jpg
+Amy_Pascal/Amy_Pascal_0001.jpg
+Mitchell_McLaughlin/Mitchell_McLaughlin_0001.jpg
+Marieta_Chrousala/Marieta_Chrousala_0001.jpg
+Patsy_Mink/Patsy_Mink_0001.jpg
+Brian_Meadors/Brian_Meadors_0001.jpg
+Mahima_Chaudhari/Mahima_Chaudhari_0001.jpg
+Daniele_Nardello/Daniele_Nardello_0001.jpg
+David_Trimble/David_Trimble_0005.jpg
+Paul_Michael_Daniels/Paul_Michael_Daniels_0001.jpg
+Roger_Daltrey/Roger_Daltrey_0001.jpg
+Robert_Evans/Robert_Evans_0003.jpg
+Julio_Rossi/Julio_Rossi_0001.jpg
+Melissa_Stark/Melissa_Stark_0001.jpg
+Jesse_James/Jesse_James_0001.jpg
+Gong_Ruina/Gong_Ruina_0001.jpg
+Ray_Price/Ray_Price_0001.jpg
+Jim_Hahn/Jim_Hahn_0004.jpg
+Habib_Rizieq/Habib_Rizieq_0002.jpg
+Budd_Schulberg/Budd_Schulberg_0001.jpg
+Evgeni_Plushenko/Evgeni_Plushenko_0001.jpg
+Jeff_Schiffner/Jeff_Schiffner_0001.jpg
+Martin_Lawrence/Martin_Lawrence_0001.jpg
+Kim_Cattrall/Kim_Cattrall_0001.jpg
+Nathirah_Hussein/Nathirah_Hussein_0001.jpg
+Maria_Shriver/Maria_Shriver_0001.jpg
+Henk_Bekedam/Henk_Bekedam_0001.jpg
+Hanan_Ashrawi/Hanan_Ashrawi_0001.jpg
+Manuel_Llorente/Manuel_Llorente_0001.jpg
+Salman_Rushdie/Salman_Rushdie_0003.jpg
+Iain_Richmond/Iain_Richmond_0001.jpg
+Tommy_Haas/Tommy_Haas_0001.jpg
+Bryan_Chui/Bryan_Chui_0001.jpg
+Ivan_Lee/Ivan_Lee_0001.jpg
+Hu_Jintao/Hu_Jintao_0007.jpg
+Peter_Holmberg/Peter_Holmberg_0001.jpg
+Bob_Eskridge/Bob_Eskridge_0001.jpg
+Vagit_Alekperov/Vagit_Alekperov_0001.jpg
+Andy_Griggs/Andy_Griggs_0001.jpg
+Koji_Uehara/Koji_Uehara_0001.jpg
+Brennon_Leighton/Brennon_Leighton_0001.jpg
+William_Delahunt/William_Delahunt_0001.jpg
+Nabil_Shaath/Nabil_Shaath_0001.jpg
+Michael_Rolinee/Michael_Rolinee_0001.jpg
+Leon_Lai/Leon_Lai_0001.jpg
+Malik_Mahmud/Malik_Mahmud_0001.jpg
+Matt_Anderson/Matt_Anderson_0001.jpg
+Greg_Kinsey/Greg_Kinsey_0001.jpg
+Oswald_Gruebel/Oswald_Gruebel_0001.jpg
+David_Welch/David_Welch_0001.jpg
+Penelope_Cruz/Penelope_Cruz_0003.jpg
+Gary_Bald/Gary_Bald_0001.jpg
+Roger_Etchegaray/Roger_Etchegaray_0001.jpg
+John_Engler/John_Engler_0001.jpg
+Bobby_Jackson/Bobby_Jackson_0001.jpg
+Rick_Stansbury/Rick_Stansbury_0001.jpg
+Larry_Tanenbaum/Larry_Tanenbaum_0001.jpg
+Andre_Agassi/Andre_Agassi_0014.jpg
+Zach_Safrin/Zach_Safrin_0001.jpg
+Sanja_Papic/Sanja_Papic_0001.jpg
+Lincoln_Chafee/Lincoln_Chafee_0001.jpg
+Jerry_Oliver/Jerry_Oliver_0001.jpg
+David_Siegel/David_Siegel_0001.jpg
+Armand_Sargen/Armand_Sargen_0001.jpg
+Greg_Hodge/Greg_Hodge_0001.jpg
+Mohamed_ElBaradei/Mohamed_ElBaradei_0008.jpg
+Ben_Braun/Ben_Braun_0001.jpg
+Michael_Sheehan/Michael_Sheehan_0001.jpg
+Clay_Aiken/Clay_Aiken_0006.jpg
+Jim_Bunning/Jim_Bunning_0001.jpg
+Peter_Bacanovic/Peter_Bacanovic_0002.jpg
+Martin_McCauley/Martin_McCauley_0002.jpg
+Chris_Swecker/Chris_Swecker_0001.jpg
+Cindy_Crawford/Cindy_Crawford_0001.jpg
+Luis_Pujols/Luis_Pujols_0001.jpg
+Heather_Chinnock/Heather_Chinnock_0001.jpg
+Billy_Edelin/Billy_Edelin_0001.jpg
+Mauricio_Macri/Mauricio_Macri_0001.jpg
+Ellen_Saracini/Ellen_Saracini_0001.jpg
+Chuck_Amato/Chuck_Amato_0001.jpg
+Paul_Wilson/Paul_Wilson_0001.jpg
+Billy_Boyd/Billy_Boyd_0001.jpg
+Pupi_Avati/Pupi_Avati_0001.jpg
+Cuba_Gooding_Jr/Cuba_Gooding_Jr_0001.jpg
+Nuon_Chea/Nuon_Chea_0001.jpg
+Mahendra_Chaudhry/Mahendra_Chaudhry_0001.jpg
+Randy_Dryer/Randy_Dryer_0001.jpg
+Maria_Shkolnikova/Maria_Shkolnikova_0001.jpg
+Thomas_Wilkens/Thomas_Wilkens_0001.jpg
+Franco_Cangele/Franco_Cangele_0001.jpg
+Rose_Marie/Rose_Marie_0001.jpg
+Jong_Thae_Hwa/Jong_Thae_Hwa_0002.jpg
+Carlos_Manuel_Pruneda/Carlos_Manuel_Pruneda_0001.jpg
+Angelica_Romero/Angelica_Romero_0001.jpg
+Reggie_Miller/Reggie_Miller_0002.jpg
+Alejandro_Lopez/Alejandro_Lopez_0001.jpg
+Choi_Yun-yong/Choi_Yun-yong_0001.jpg
+Steven_Craig/Steven_Craig_0001.jpg
+Jean_Brumley/Jean_Brumley_0002.jpg
+Leah_Remini/Leah_Remini_0001.jpg
+Janice_Abreu/Janice_Abreu_0001.jpg
+Frank_Schmoekel/Frank_Schmoekel_0001.jpg
+Sadam_Hassan/Sadam_Hassan_0001.jpg
+David_Dewhurst/David_Dewhurst_0001.jpg
+Eric_Wedge/Eric_Wedge_0001.jpg
+Claude_Jorda/Claude_Jorda_0001.jpg
+Tirunesh_Dibaba/Tirunesh_Dibaba_0001.jpg
+Ron_Howard/Ron_Howard_0002.jpg
+Mehmet_Ali_Sahin/Mehmet_Ali_Sahin_0001.jpg
+Jewel_Howard-Taylor/Jewel_Howard-Taylor_0001.jpg
+Bruce_Gebhardt/Bruce_Gebhardt_0001.jpg
+Pat_DAmuro/Pat_DAmuro_0001.jpg
+Lisa_Marie_Presley/Lisa_Marie_Presley_0004.jpg
+Matthew_During/Matthew_During_0001.jpg
+Jason_Campbell/Jason_Campbell_0001.jpg
+Terry_Gilliam/Terry_Gilliam_0001.jpg
+Hamad_Bin_Isa_al-Khalifa/Hamad_Bin_Isa_al-Khalifa_0001.jpg
+Ramon_Ponce_de_Leon/Ramon_Ponce_de_Leon_0001.jpg
+Mathilda_Karel_Spak/Mathilda_Karel_Spak_0001.jpg
+Glen_Sather/Glen_Sather_0001.jpg
+Jorge_Castaneda/Jorge_Castaneda_0002.jpg
+John_Snow/John_Snow_0008.jpg
+Myung_Yang/Myung_Yang_0001.jpg
+Patricia_Wartusch/Patricia_Wartusch_0001.jpg
+Eddy_Merckx/Eddy_Merckx_0001.jpg
+Scott_Fawell/Scott_Fawell_0001.jpg
+Tiago_Splitter/Tiago_Splitter_0001.jpg
+Derrick_Rodgers/Derrick_Rodgers_0001.jpg
+Miroljub/Miroljub_0001.jpg
+Parris_Glendening/Parris_Glendening_0002.jpg
+Kimberly_Bruckner/Kimberly_Bruckner_0001.jpg
+Jim_Piper/Jim_Piper_0001.jpg
+Feliciano_Lopez/Feliciano_Lopez_0001.jpg
+Mike_Sweeney/Mike_Sweeney_0001.jpg
+Makiko_Tanaka/Makiko_Tanaka_0001.jpg
+Michael_Broad/Michael_Broad_0001.jpg
+Boutros_Boutros_Ghali/Boutros_Boutros_Ghali_0001.jpg
+David_Beckham/David_Beckham_0002.jpg
+Donald_Rumsfeld/Donald_Rumsfeld_0029.jpg
+Geoffrey_Davis/Geoffrey_Davis_0001.jpg
+Charlotte_Church/Charlotte_Church_0001.jpg
+Robin_Tunney/Robin_Tunney_0001.jpg
+Chris_Neil/Chris_Neil_0001.jpg
+Anastasia_Kelesidou/Anastasia_Kelesidou_0001.jpg
+James_Cameron/James_Cameron_0002.jpg
+Chris_Bell/Chris_Bell_0001.jpg
+James_Morris/James_Morris_0002.jpg
+Gary_Gero/Gary_Gero_0001.jpg
+Derek_Parra/Derek_Parra_0001.jpg
+John_Ferguson/John_Ferguson_0001.jpg
+Stephen_Glassroth/Stephen_Glassroth_0001.jpg
+Anthony_Principi/Anthony_Principi_0001.jpg
+Mike_Stefanik/Mike_Stefanik_0001.jpg
+Yusaku_Miyazato/Yusaku_Miyazato_0001.jpg
+Darla_Moore/Darla_Moore_0001.jpg
+Mark_McClellan/Mark_McClellan_0001.jpg
+Li_Ruihuan/Li_Ruihuan_0001.jpg
+Mauricio_Pochetino/Mauricio_Pochetino_0001.jpg
+Kobe_Bryant/Kobe_Bryant_0002.jpg
+George_Plimpton/George_Plimpton_0001.jpg
+George_Brumley_III/George_Brumley_III_0001.jpg
+Shigeo_Nagashima/Shigeo_Nagashima_0001.jpg
+Raghad_Saddam_Hussein/Raghad_Saddam_Hussein_0001.jpg
+Ali_Hammoud/Ali_Hammoud_0001.jpg
+Darren_Campel/Darren_Campel_0001.jpg
+Anthony_Corso/Anthony_Corso_0001.jpg
+Svetlana_Belousova/Svetlana_Belousova_0001.jpg
+Dawn_Staley/Dawn_Staley_0001.jpg
+Denis_Fassou-Nguesso/Denis_Fassou-Nguesso_0001.jpg
+Juergen_Braehmer/Juergen_Braehmer_0001.jpg
+Amanda_Beard/Amanda_Beard_0001.jpg
+Martin_ONeill/Martin_ONeill_0001.jpg
+Juan_Francisco_Palencia/Juan_Francisco_Palencia_0001.jpg
+Anneli_Jaatteenmaki/Anneli_Jaatteenmaki_0001.jpg
+Joe_Plumeri/Joe_Plumeri_0001.jpg
+Margaret_Caruso/Margaret_Caruso_0001.jpg
+Rani_Mukherjee/Rani_Mukherjee_0001.jpg
+Steve_Avery/Steve_Avery_0001.jpg
+Uthai_Pimchaichon/Uthai_Pimchaichon_0001.jpg
+James_Traficant/James_Traficant_0002.jpg
+Harry_Schmidt/Harry_Schmidt_0001.jpg
+Talisa_Soto/Talisa_Soto_0001.jpg
+Norman_Jewison/Norman_Jewison_0001.jpg
+Laura_Bush/Laura_Bush_0032.jpg
+Arnold_Schwarzenegger/Arnold_Schwarzenegger_0025.jpg
+Dale_Earnhardt/Dale_Earnhardt_0001.jpg
+John_Thune/John_Thune_0001.jpg
+Jim_Calhoun/Jim_Calhoun_0001.jpg
+Lawrence_Di_Rita/Lawrence_Di_Rita_0001.jpg
+Damon_van_Dam/Damon_van_Dam_0002.jpg
+James_May/James_May_0001.jpg
+Manuela_Montebrun/Manuela_Montebrun_0001.jpg
+Derian_Hatcher/Derian_Hatcher_0001.jpg
+Robert_McKee/Robert_McKee_0001.jpg
+Andrzej_Tyszkiewicz/Andrzej_Tyszkiewicz_0001.jpg
+Joseph_Lopez/Joseph_Lopez_0001.jpg
+Rohinton_Mistry/Rohinton_Mistry_0001.jpg
+Roberto_Cavalli/Roberto_Cavalli_0001.jpg
+Tung_Chee-hwa/Tung_Chee-hwa_0009.jpg
+Scott_Wolf/Scott_Wolf_0002.jpg
+Bryant_Young/Bryant_Young_0001.jpg
+Tony_Cummo/Tony_Cummo_0001.jpg
+Angelo_Genova/Angelo_Genova_0001.jpg
+Paul_Wolfowitz/Paul_Wolfowitz_0009.jpg
+Tom_Izzo/Tom_Izzo_0001.jpg
+Sylvester_Stallone/Sylvester_Stallone_0008.jpg
+James_Kirtley/James_Kirtley_0001.jpg
+George_HW_Bush/George_HW_Bush_0009.jpg
+Rudolf_Schuster/Rudolf_Schuster_0001.jpg
+Eric_Clapton/Eric_Clapton_0002.jpg
+Charlize_Theron/Charlize_Theron_0001.jpg
+Erwin_Abdullah/Erwin_Abdullah_0001.jpg
+Paul_Pierce/Paul_Pierce_0002.jpg
+Katerina_Smrzova/Katerina_Smrzova_0001.jpg
+Junko_Tabei/Junko_Tabei_0001.jpg
+Courtney_Cox/Courtney_Cox_0001.jpg
+Platon_Lebedev/Platon_Lebedev_0001.jpg
+Du_Qinglin/Du_Qinglin_0001.jpg
+Jean-Claude_Van_Damme/Jean-Claude_Van_Damme_0001.jpg
+Christine_Gregoire/Christine_Gregoire_0001.jpg
+James_Collinson/James_Collinson_0001.jpg
+Edward_Albee/Edward_Albee_0001.jpg
+Al_Sharpton/Al_Sharpton_0007.jpg
+Jan_van_Breda_Kolff/Jan_van_Breda_Kolff_0001.jpg
+Edward_Arsenault/Edward_Arsenault_0001.jpg
+David_Montoya/David_Montoya_0001.jpg
+Jan_De_Bont/Jan_De_Bont_0001.jpg
+Irwan_Fadzi_Idris/Irwan_Fadzi_Idris_0001.jpg
+Lazaro_Castro/Lazaro_Castro_0001.jpg
+Joseph_Kabila/Joseph_Kabila_0001.jpg
+Art_Howe/Art_Howe_0004.jpg
+Magdalena_Maleeva/Magdalena_Maleeva_0001.jpg
+Carlos_Barra/Carlos_Barra_0001.jpg
+Rob_Lowe/Rob_Lowe_0001.jpg
+Kevin_James/Kevin_James_0001.jpg
+Shoshana_Johnson/Shoshana_Johnson_0001.jpg
+Sila_Calderon/Sila_Calderon_0004.jpg
+Gary_Williams/Gary_Williams_0002.jpg
+Oscar_Bolanos/Oscar_Bolanos_0001.jpg
+Allan_Houston/Allan_Houston_0001.jpg
+Kathy_Winters/Kathy_Winters_0001.jpg
+Gore_Vidal/Gore_Vidal_0001.jpg
+Ana_Sebastiao/Ana_Sebastiao_0001.jpg
+Helo_Pinheiro/Helo_Pinheiro_0001.jpg
+Jeanette_Gray/Jeanette_Gray_0001.jpg
+Michael_Haneke/Michael_Haneke_0001.jpg
+Marsha_Sharp/Marsha_Sharp_0001.jpg
+Mariah_Carey/Mariah_Carey_0001.jpg
+David_Dodge/David_Dodge_0001.jpg
+Portia_de_Rossi/Portia_de_Rossi_0001.jpg
+Elvis_Costello/Elvis_Costello_0001.jpg
+Bill_Duffey/Bill_Duffey_0001.jpg
+Jennifer_Rodriguez/Jennifer_Rodriguez_0002.jpg
+Juan_Manuel_Marquez/Juan_Manuel_Marquez_0002.jpg
+Olympia_Dukakis/Olympia_Dukakis_0001.jpg
+Guangdong_Ou_Guangyuan/Guangdong_Ou_Guangyuan_0001.jpg
+John_Edwards/John_Edwards_0008.jpg
+Jesse_James_Leija/Jesse_James_Leija_0001.jpg
+Eva_Dimas/Eva_Dimas_0002.jpg
+Bela_Karolyi/Bela_Karolyi_0001.jpg
+Enrica_Fico/Enrica_Fico_0001.jpg
+Spike_Jonze/Spike_Jonze_0001.jpg
+Sandra_Banning/Sandra_Banning_0001.jpg
+Norman_Mailer/Norman_Mailer_0001.jpg
+Jaqueline_Godoy/Jaqueline_Godoy_0001.jpg
+Lucie_Lapovsky/Lucie_Lapovsky_0001.jpg
+Tom_Hanusik/Tom_Hanusik_0001.jpg
+Jonathan_Schroeder/Jonathan_Schroeder_0001.jpg
+Brawley_King/Brawley_King_0001.jpg
+Gary_Marshall/Gary_Marshall_0001.jpg
+Jason_Lezak/Jason_Lezak_0002.jpg
+Heather_Locklear/Heather_Locklear_0001.jpg
+Franck_Cerutti/Franck_Cerutti_0001.jpg
+Bode_Miller/Bode_Miller_0002.jpg
+Masahiko_Nagasawa/Masahiko_Nagasawa_0002.jpg
+Thomas_Wyman/Thomas_Wyman_0001.jpg
+John_Lithgow/John_Lithgow_0001.jpg
+Frank_Hsieh/Frank_Hsieh_0001.jpg
+Shaun_Pollock/Shaun_Pollock_0001.jpg
+Elodie_Bouchez/Elodie_Bouchez_0001.jpg
+Gabrielle_Rose/Gabrielle_Rose_0001.jpg
+Chin-Feng_Chen/Chin-Feng_Chen_0001.jpg
+Arsinee_Khanjian/Arsinee_Khanjian_0002.jpg
+Thor_Pedersen/Thor_Pedersen_0001.jpg
+Rob_Marshall/Rob_Marshall_0001.jpg
+Elijah_Wood/Elijah_Wood_0002.jpg
+Ed_Sullivan/Ed_Sullivan_0001.jpg
+Emily_Robison/Emily_Robison_0001.jpg
+Maggie_Smith/Maggie_Smith_0002.jpg
+Muwafak_al-Ani/Muwafak_al-Ani_0001.jpg
+Eve_Pelletier/Eve_Pelletier_0001.jpg
+Joerg_Haider/Joerg_Haider_0001.jpg
+Carey_Lowell/Carey_Lowell_0001.jpg
+Michael_Shane_Jolly/Michael_Shane_Jolly_0001.jpg
+Marian_Dolan/Marian_Dolan_0001.jpg
+Norm_Macdonald/Norm_Macdonald_0001.jpg
+Troy_Aikman/Troy_Aikman_0001.jpg
+Rocco_Buttiglione/Rocco_Buttiglione_0001.jpg
+Vincent_Cianci_Jr/Vincent_Cianci_Jr_0001.jpg
+Bijan_Namdar_Zangeneh/Bijan_Namdar_Zangeneh_0001.jpg
+James_Parker/James_Parker_0002.jpg
+John_Madden/John_Madden_0001.jpg
+Leander_Paes/Leander_Paes_0002.jpg
+Jim_Thome/Jim_Thome_0001.jpg
+Keith_Snyder/Keith_Snyder_0001.jpg
+Brook_Robinson/Brook_Robinson_0001.jpg
+Phillip_Fulmer/Phillip_Fulmer_0001.jpg
+Zorica_Radovic/Zorica_Radovic_0001.jpg
+Stephen_Joseph/Stephen_Joseph_0001.jpg
+Wilfredo_Moreno/Wilfredo_Moreno_0001.jpg
+Anne_ONeil/Anne_ONeil_0001.jpg
+Rafiq_Hariri/Rafiq_Hariri_0001.jpg
+Sam_Rockwell/Sam_Rockwell_0001.jpg
+Joseph_Fiennes/Joseph_Fiennes_0001.jpg
+Jo_Joong-hyon/Jo_Joong-hyon_0001.jpg
+Yoko_Ono/Yoko_Ono_0006.jpg
+Alec_Baldwin/Alec_Baldwin_0001.jpg
+Camryn_Manheim/Camryn_Manheim_0001.jpg
+Michael_Peat/Michael_Peat_0001.jpg
+Elgin_Baylor/Elgin_Baylor_0001.jpg
+Alistair_MacDonald/Alistair_MacDonald_0001.jpg
+Chris_Simon/Chris_Simon_0001.jpg
+Severino_Antinori/Severino_Antinori_0001.jpg
+Alicia_Witt/Alicia_Witt_0001.jpg
+Lenny_Kravitz/Lenny_Kravitz_0001.jpg
+Zakia_Hakki/Zakia_Hakki_0001.jpg
+Werner_Schlager/Werner_Schlager_0001.jpg
+Silvia_Farina_Elia/Silvia_Farina_Elia_0001.jpg
+Ashraf_Alasmar/Ashraf_Alasmar_0001.jpg
+Prince_Rainier_III/Prince_Rainier_III_0001.jpg
+Elizabeth_Pena/Elizabeth_Pena_0001.jpg
+Alan_Greenspan/Alan_Greenspan_0005.jpg
+Deniz_Baykal/Deniz_Baykal_0001.jpg
+Kirsten_Gilham/Kirsten_Gilham_0001.jpg
+Dominique_Perben/Dominique_Perben_0001.jpg
+Peter_Rasch/Peter_Rasch_0001.jpg
+Akbar_Hashemi_Rafsanjani/Akbar_Hashemi_Rafsanjani_0002.jpg
+Conan_OBrien/Conan_OBrien_0001.jpg
+George_P_Bush/George_P_Bush_0002.jpg
+Richard_Krajicek/Richard_Krajicek_0003.jpg
+Tyler_Hamilton/Tyler_Hamilton_0001.jpg
+Sarah_Michelle_Gellar/Sarah_Michelle_Gellar_0001.jpg
+Askar_Akayev/Askar_Akayev_0001.jpg
+Daryl_Parks/Daryl_Parks_0001.jpg
+Miguel_Rosseto/Miguel_Rosseto_0001.jpg
+Rosalie_Perkov/Rosalie_Perkov_0001.jpg
+Kalid_Kaid/Kalid_Kaid_0001.jpg
+David_Ballantyne/David_Ballantyne_0001.jpg
+Hernan_Diaz/Hernan_Diaz_0001.jpg
+Jim_Anderson/Jim_Anderson_0001.jpg
+Justin_Guarini/Justin_Guarini_0003.jpg
+Roger_Cook/Roger_Cook_0001.jpg
+Juan_Carlos_Ferrero/Juan_Carlos_Ferrero_0003.jpg
+Jacqueline_Gold/Jacqueline_Gold_0001.jpg
+Hung_Wan-ting/Hung_Wan-ting_0001.jpg
+Steve_Waugh/Steve_Waugh_0001.jpg
+Enrik_Vendt/Enrik_Vendt_0001.jpg
+Darrell_Dickey/Darrell_Dickey_0001.jpg
+Wang_Yingfan/Wang_Yingfan_0001.jpg
+Maritza_Macias_Furano/Maritza_Macias_Furano_0001.jpg
+Alexandre_Daigle/Alexandre_Daigle_0001.jpg
+Mark_Leno/Mark_Leno_0001.jpg
+Chris_Rock/Chris_Rock_0001.jpg
+Aby_Har-Even/Aby_Har-Even_0001.jpg
+Landon_Donovan/Landon_Donovan_0001.jpg
+Cindy_Zagorski/Cindy_Zagorski_0001.jpg
+Janet_Napolitano/Janet_Napolitano_0003.jpg
+Bill_Stapleton/Bill_Stapleton_0001.jpg
+Elizabeth_Shue/Elizabeth_Shue_0001.jpg
+Stephen_Oake/Stephen_Oake_0001.jpg
+Michelangelo_Antonioni/Michelangelo_Antonioni_0001.jpg
+James_Barksdale/James_Barksdale_0001.jpg
+Roberto_Carlos/Roberto_Carlos_0001.jpg
+Courtney_Love/Courtney_Love_0001.jpg
+Alexandra_Pelosi/Alexandra_Pelosi_0001.jpg
+Kristin_Scott_Thomas/Kristin_Scott_Thomas_0001.jpg
+Joseph_Safra/Joseph_Safra_0001.jpg
+Tom_Vilsack/Tom_Vilsack_0001.jpg
+John_Richardson/John_Richardson_0001.jpg
+Joe_Strummer/Joe_Strummer_0001.jpg
+Ira_Einhorn/Ira_Einhorn_0001.jpg
+Ronaldo_Luis_Nazario_de_Lima/Ronaldo_Luis_Nazario_de_Lima_0004.jpg
+Roy_Blunt/Roy_Blunt_0001.jpg
+Dianne_Reeves/Dianne_Reeves_0001.jpg
+Ulrich_Kueperkoch/Ulrich_Kueperkoch_0001.jpg
+Scott_Gorelick/Scott_Gorelick_0001.jpg
+Talal_Keenaan/Talal_Keenaan_0001.jpg
+Ryan_Newman/Ryan_Newman_0001.jpg
+Juan_Fernandez/Juan_Fernandez_0001.jpg
+Uri_Lopolianski/Uri_Lopolianski_0001.jpg
+John_Starks/John_Starks_0001.jpg
+Nathalie_Gagnon/Nathalie_Gagnon_0001.jpg
+Patrick_Ewing/Patrick_Ewing_0002.jpg
+Barry_Collier/Barry_Collier_0001.jpg
+Debra_Shank/Debra_Shank_0001.jpg
+Colin_Jackson/Colin_Jackson_0002.jpg
+Saadi_Gadhafi/Saadi_Gadhafi_0001.jpg
+Nikki_Teasley/Nikki_Teasley_0001.jpg
+Dimitar_Berbatov/Dimitar_Berbatov_0001.jpg
+Lucio_Stanca/Lucio_Stanca_0001.jpg
+Brian_Wells/Brian_Wells_0002.jpg
+Randy_Ferbey/Randy_Ferbey_0001.jpg
+Coretta_Scott_King/Coretta_Scott_King_0001.jpg
+Craig_David/Craig_David_0001.jpg
+Patrick_Leahy/Patrick_Leahy_0001.jpg
+Brad_Pitt/Brad_Pitt_0001.jpg
+Dan_Ackroyd/Dan_Ackroyd_0001.jpg
+Nancy_Smith/Nancy_Smith_0001.jpg
+Ari_Bousbib/Ari_Bousbib_0001.jpg
+Chris_Cirino/Chris_Cirino_0001.jpg
+Martin_Kristof/Martin_Kristof_0001.jpg
+Gary_Dellaverson/Gary_Dellaverson_0001.jpg
+Fabricio_Oberto/Fabricio_Oberto_0001.jpg
+Laurent_Gbagbo/Laurent_Gbagbo_0002.jpg
+Mian_Khursheed_Mehmood_Kasuri/Mian_Khursheed_Mehmood_Kasuri_0004.jpg
+Ruth_Pearce/Ruth_Pearce_0001.jpg
+Raja_Zafar-ul-Haq/Raja_Zafar-ul-Haq_0001.jpg
+Cindy_Taylor/Cindy_Taylor_0001.jpg
+Emmy_Rossum/Emmy_Rossum_0001.jpg
+Manuel_Gehring/Manuel_Gehring_0001.jpg
+Humberto_Coelho/Humberto_Coelho_0001.jpg
+Faye_Dunaway/Faye_Dunaway_0003.jpg
+Wayne_Allard/Wayne_Allard_0001.jpg
+Ion_Tiriac/Ion_Tiriac_0001.jpg
+Art_Hoffmann/Art_Hoffmann_0002.jpg
+Jack_Nicholson/Jack_Nicholson_0003.jpg
+Delphine_Chuillot/Delphine_Chuillot_0001.jpg
+Nick_Turner/Nick_Turner_0001.jpg
+Shawn_Bradley/Shawn_Bradley_0001.jpg
+Yale_Kamisar/Yale_Kamisar_0001.jpg
+Christian_Gimenez/Christian_Gimenez_0001.jpg
+Martin_Hoellwarth/Martin_Hoellwarth_0002.jpg
+Mike_Webster/Mike_Webster_0001.jpg
+Rubens_Barrichello/Rubens_Barrichello_0008.jpg
+Gary_Coleman/Gary_Coleman_0001.jpg
+Heidi_Klum/Heidi_Klum_0003.jpg
+Esther_Macklin/Esther_Macklin_0001.jpg
+Ronald_Ito/Ronald_Ito_0001.jpg
+Scott_Hamilton/Scott_Hamilton_0001.jpg
+Jacques_Kallis/Jacques_Kallis_0001.jpg
+William_Perry/William_Perry_0001.jpg
+Takashi_Sorimachi/Takashi_Sorimachi_0001.jpg
+Kate_Starbird/Kate_Starbird_0001.jpg
+Morris_Watts/Morris_Watts_0001.jpg
+Bettina_Rheims/Bettina_Rheims_0001.jpg
+Yuri_Malenchenko/Yuri_Malenchenko_0002.jpg
+Michael_Jasny/Michael_Jasny_0001.jpg
+Jim_Wall/Jim_Wall_0001.jpg
+Julio_Iglesias_Jr/Julio_Iglesias_Jr_0002.jpg
+Jim_Doyle/Jim_Doyle_0001.jpg
+John_Fox/John_Fox_0001.jpg
+Albaro_Recoba/Albaro_Recoba_0001.jpg
+Chris_Cooper/Chris_Cooper_0002.jpg
+Thomas_Watjen/Thomas_Watjen_0001.jpg
+Adolfo_Rodriguez_Saa/Adolfo_Rodriguez_Saa_0001.jpg
+Laurie_Chan/Laurie_Chan_0001.jpg
+Michael_Moore/Michael_Moore_0001.jpg
+Sebastian_Porto/Sebastian_Porto_0001.jpg
+Betty_Williams/Betty_Williams_0001.jpg
+Wang_Fei/Wang_Fei_0001.jpg
+Sepp_Blatter/Sepp_Blatter_0004.jpg
+Paul_Vathis/Paul_Vathis_0001.jpg
+Richard_Armitage/Richard_Armitage_0005.jpg
+Katherine_Harris/Katherine_Harris_0003.jpg
+Elena_Bereznaya/Elena_Bereznaya_0001.jpg
+Princess_Masako/Princess_Masako_0002.jpg
+David_Ho/David_Ho_0001.jpg
+Bob_Graham/Bob_Graham_0004.jpg
+Andy_Graves/Andy_Graves_0001.jpg
+Mary_Descenza/Mary_Descenza_0001.jpg
+Raoul_Ruiz/Raoul_Ruiz_0004.jpg
+Richard_Daley/Richard_Daley_0001.jpg
+Katalin_Kollat/Katalin_Kollat_0001.jpg
+Nia_Vardalos/Nia_Vardalos_0005.jpg
+Leland_Chapman/Leland_Chapman_0001.jpg
+William_Bulger/William_Bulger_0001.jpg
+Earl_Scruggs/Earl_Scruggs_0001.jpg
+Thomas_Van_Essen/Thomas_Van_Essen_0001.jpg
+Nathan_Lane/Nathan_Lane_0001.jpg
+Jennifer_Love_Hewitt/Jennifer_Love_Hewitt_0001.jpg
+Kirsten_Clark/Kirsten_Clark_0001.jpg
+Wanda_de_la_Jesus/Wanda_de_la_Jesus_0001.jpg
+Kate_Hudson/Kate_Hudson_0009.jpg
+Kevin_Millwood/Kevin_Millwood_0001.jpg
+Chandrika_Kumaratunga/Chandrika_Kumaratunga_0001.jpg
+Julia_Ormond/Julia_Ormond_0001.jpg
+Richard_Fine/Richard_Fine_0001.jpg
+Todd_Parrott/Todd_Parrott_0001.jpg
+Justin_Marshall/Justin_Marshall_0001.jpg
+Sally_Field/Sally_Field_0001.jpg
+Frederique_van_der_Wal/Frederique_van_der_Wal_0001.jpg
+Barbara_Bodine/Barbara_Bodine_0001.jpg
+Dave_McNealey/Dave_McNealey_0001.jpg
+Enrique_Haroldo_Gorriaran_Merlo/Enrique_Haroldo_Gorriaran_Merlo_0001.jpg
+Tara_Kirk/Tara_Kirk_0001.jpg
+Rio_Ferdinand/Rio_Ferdinand_0002.jpg
+Freddy_Vasques_Kinchokpe/Freddy_Vasques_Kinchokpe_0001.jpg
+Steve_Blake/Steve_Blake_0001.jpg
+Rolf_Eckrodt/Rolf_Eckrodt_0002.jpg
+Brooke_Adams/Brooke_Adams_0001.jpg
+Xanana_Gusmao/Xanana_Gusmao_0001.jpg
+Ahmed_Ibrahim_Bilal/Ahmed_Ibrahim_Bilal_0001.jpg
+Wan_Yanhai/Wan_Yanhai_0001.jpg
+James_Coviello/James_Coviello_0001.jpg
+Shavon_Earp/Shavon_Earp_0001.jpg
+Hayden_Panettiere/Hayden_Panettiere_0001.jpg
+Derek_Lowe/Derek_Lowe_0002.jpg
+Pernilla_Bjorn/Pernilla_Bjorn_0001.jpg
+Claire_Hentzen/Claire_Hentzen_0002.jpg
+Tamika_Catchings/Tamika_Catchings_0001.jpg
+Deece_Eckstein/Deece_Eckstein_0001.jpg
+Dunn_Lampton/Dunn_Lampton_0001.jpg
+Ryan_Drese/Ryan_Drese_0001.jpg
+Mary_Blige/Mary_Blige_0001.jpg
+Dwayne_Johnson/Dwayne_Johnson_0002.jpg
+Leslie_Ann_Woodward/Leslie_Ann_Woodward_0002.jpg
+Mario_Lobo_Zagallo/Mario_Lobo_Zagallo_0001.jpg
+Michel_Therrien/Michel_Therrien_0001.jpg
+Steve_Blankenship/Steve_Blankenship_0001.jpg
+Chris_Pronger/Chris_Pronger_0001.jpg
+Mohammad_Aktar/Mohammad_Aktar_0001.jpg
+Horacio_Julio_Pina/Horacio_Julio_Pina_0001.jpg
+Steve_Redgrave/Steve_Redgrave_0001.jpg
+Julien_Varlet/Julien_Varlet_0001.jpg
+Maria_Guida/Maria_Guida_0001.jpg
+Marc_Grossman/Marc_Grossman_0002.jpg
+Stanley_Nelson/Stanley_Nelson_0001.jpg
+Catherine_Ndereba/Catherine_Ndereba_0001.jpg
+Joe_Paterno/Joe_Paterno_0001.jpg
+Alfonso_Cuaron/Alfonso_Cuaron_0001.jpg
+Antonio_Banderas/Antonio_Banderas_0002.jpg
+Wayne_Ferreira/Wayne_Ferreira_0004.jpg
+Rene_Portland/Rene_Portland_0001.jpg
+Carol_Moseley_Braun/Carol_Moseley_Braun_0001.jpg
+Paek_Nam_Sun/Paek_Nam_Sun_0001.jpg
+Abdoulaye_Wade/Abdoulaye_Wade_0003.jpg
+Roberto_Laratro/Roberto_Laratro_0001.jpg
+Gary_Leon_Ridgway/Gary_Leon_Ridgway_0001.jpg
+Jaime_Pressly/Jaime_Pressly_0001.jpg
+Rulon_Gardner/Rulon_Gardner_0001.jpg
+Martin_Frost/Martin_Frost_0001.jpg
+Anne_Krueger/Anne_Krueger_0002.jpg
+Conrad_Black/Conrad_Black_0001.jpg
+Alastair_Johnston/Alastair_Johnston_0001.jpg
+Deena_Burnett/Deena_Burnett_0001.jpg
+Bill_Stein/Bill_Stein_0001.jpg
+Alecos_Markides/Alecos_Markides_0001.jpg
+Alan_Trammell/Alan_Trammell_0001.jpg
+Doug_Collins/Doug_Collins_0001.jpg
+Boris_Trajkovski/Boris_Trajkovski_0001.jpg
+Felix_Sanchez/Felix_Sanchez_0001.jpg
+Jorge_Enrique_Jimenez/Jorge_Enrique_Jimenez_0001.jpg
+Jalal_Talabani/Jalal_Talabani_0001.jpg
+John_Timoney/John_Timoney_0001.jpg
+Pierre_Boulanger/Pierre_Boulanger_0001.jpg
+Martha_Martinez_Flores/Martha_Martinez_Flores_0001.jpg
+Todd_Haynes/Todd_Haynes_0004.jpg
+Amber_Frey/Amber_Frey_0001.jpg
+Hun_Sen/Hun_Sen_0004.jpg
+John_Fenn/John_Fenn_0001.jpg
+Brandon_Fails/Brandon_Fails_0001.jpg
+Martin_Howard/Martin_Howard_0001.jpg
+Francis_Ricciardone/Francis_Ricciardone_0001.jpg
+Matt_Damon/Matt_Damon_0004.jpg
+Sargis_Sargsian/Sargis_Sargsian_0001.jpg
+Jason_Vale/Jason_Vale_0001.jpg
+Colleen_Atwood/Colleen_Atwood_0001.jpg
+Lee_Ann_Terlaji/Lee_Ann_Terlaji_0001.jpg
+Rod_Paige/Rod_Paige_0001.jpg
+Taylyn_Solomon/Taylyn_Solomon_0001.jpg
+Cristina_Kirchner/Cristina_Kirchner_0001.jpg
+Bob_Hope/Bob_Hope_0002.jpg
+Doug_Melvin/Doug_Melvin_0001.jpg
+Yuri_Fedotov/Yuri_Fedotov_0001.jpg
+Charles_Taylor/Charles_Taylor_0008.jpg
+Sofyan_Dawood/Sofyan_Dawood_0001.jpg
+Dennis_Hastert/Dennis_Hastert_0005.jpg
+Paris_Hilton/Paris_Hilton_0001.jpg
+Brendan_Hansen/Brendan_Hansen_0002.jpg
+Luis_Ernesto_Derbez_Bautista/Luis_Ernesto_Derbez_Bautista_0002.jpg
+Blythe_Hartley/Blythe_Hartley_0002.jpg
+Pierre_Van_Hooijdonk/Pierre_Van_Hooijdonk_0001.jpg
+Polona_Bas/Polona_Bas_0001.jpg
+Corliss_Williamson/Corliss_Williamson_0001.jpg
+Rhina_Villatoro/Rhina_Villatoro_0001.jpg
+Leonard_Hamilton/Leonard_Hamilton_0001.jpg
+Juan_Pablo_Montoya/Juan_Pablo_Montoya_0008.jpg
+Tommy_Thompson/Tommy_Thompson_0009.jpg
+Stephen_Webster/Stephen_Webster_0001.jpg
+Donald_Pettit/Donald_Pettit_0001.jpg
+Donna_Walker/Donna_Walker_0001.jpg
+Kate_Lee/Kate_Lee_0001.jpg
+Shafal_Mosed/Shafal_Mosed_0001.jpg
+Mariana_Ohata/Mariana_Ohata_0001.jpg
+Jefferson_Perez/Jefferson_Perez_0002.jpg
+Evo_Morales/Evo_Morales_0001.jpg
+Steve_Valentine/Steve_Valentine_0001.jpg
+Steve_Kerr/Steve_Kerr_0001.jpg
+Vivica_Fox/Vivica_Fox_0001.jpg
+Ahmad_Jbarah/Ahmad_Jbarah_0001.jpg
+Ellen_MacArthur/Ellen_MacArthur_0001.jpg
+Nelson_Mandela/Nelson_Mandela_0004.jpg
+Steve_Austin/Steve_Austin_0001.jpg
+Roger_Grimes/Roger_Grimes_0001.jpg
+Sterling_Hitchcock/Sterling_Hitchcock_0001.jpg
+Rainer_Schuettler/Rainer_Schuettler_0005.jpg
+Julio_Toro/Julio_Toro_0001.jpg
+Carrie-Anne_Moss/Carrie-Anne_Moss_0004.jpg
+Charles_Mathews/Charles_Mathews_0002.jpg
+Arminio_Fraga/Arminio_Fraga_0005.jpg
+Jan_Peter_Balkenende/Jan_Peter_Balkenende_0001.jpg
+Christine_Todd_Whitman/Christine_Todd_Whitman_0001.jpg
+Nona_Gaye/Nona_Gaye_0001.jpg
+Mladen_Naletilic/Mladen_Naletilic_0001.jpg
+Tassos_Papadopoulos/Tassos_Papadopoulos_0003.jpg
+Daniel_Radcliffe/Daniel_Radcliffe_0002.jpg
+Tim_Salmon/Tim_Salmon_0001.jpg
+Orlando_Bloom/Orlando_Bloom_0003.jpg
+Aaron_Patterson/Aaron_Patterson_0001.jpg
+Gary_Sayler/Gary_Sayler_0001.jpg
+Brandon_Webb/Brandon_Webb_0001.jpg
+David_Dewayne_Williams/David_Dewayne_Williams_0001.jpg
+Lauren_Hutton/Lauren_Hutton_0001.jpg
+Herbie_Hancock/Herbie_Hancock_0001.jpg
+Janica_Kostelic/Janica_Kostelic_0001.jpg
+Elena_Tihomirova/Elena_Tihomirova_0001.jpg
+Theo_Epstein/Theo_Epstein_0001.jpg
+Wilma_McNabb/Wilma_McNabb_0001.jpg
+John_Gordnick/John_Gordnick_0001.jpg
+Alan_Tang_Kwong-wing/Alan_Tang_Kwong-wing_0001.jpg
+Jean-Claude_Braquet/Jean-Claude_Braquet_0001.jpg
+Oprah_Winfrey/Oprah_Winfrey_0001.jpg
+Jessica_Simpson/Jessica_Simpson_0001.jpg
+Dorothy_Loudon/Dorothy_Loudon_0001.jpg
+Valentina_Tereshkova/Valentina_Tereshkova_0001.jpg
+Svetlana_Koroleva/Svetlana_Koroleva_0001.jpg
+Jon_Kitna/Jon_Kitna_0001.jpg
+Lyudmila_Putin/Lyudmila_Putin_0001.jpg
+Greg_Gilbert/Greg_Gilbert_0001.jpg
+Sara_Elisabeth_Ahmad/Sara_Elisabeth_Ahmad_0001.jpg
+Sandy_Wise/Sandy_Wise_0001.jpg
+Denzel_Washington/Denzel_Washington_0001.jpg
+Joshua_Harapko/Joshua_Harapko_0001.jpg
+Ana_Palacio/Ana_Palacio_0002.jpg
+Joe_Cravens/Joe_Cravens_0001.jpg
+Dolly_Parton/Dolly_Parton_0002.jpg
+Damon_Dash/Damon_Dash_0001.jpg
+Tabare_Vazquez/Tabare_Vazquez_0001.jpg
+Nicolas_Macrozonaris/Nicolas_Macrozonaris_0001.jpg
+Christiane_Wulff/Christiane_Wulff_0001.jpg
+Tim_Norbeck/Tim_Norbeck_0001.jpg
+John_Robbins/John_Robbins_0001.jpg
+Susan_Collins/Susan_Collins_0001.jpg
+Yoon_Young-kwan/Yoon_Young-kwan_0001.jpg
+Laurent_Jalabert/Laurent_Jalabert_0002.jpg
+Shingo_Suetsugu/Shingo_Suetsugu_0001.jpg
+Poala_Suarez/Poala_Suarez_0001.jpg
+Mary_Lou_Markakis/Mary_Lou_Markakis_0001.jpg
+Mikhail_Kasyanov/Mikhail_Kasyanov_0004.jpg
+Chris_Andrews/Chris_Andrews_0001.jpg
+Don_King/Don_King_0001.jpg
+Joey_Mantia/Joey_Mantia_0001.jpg
+Pedro_Solbes/Pedro_Solbes_0001.jpg
+Kurt_Budke/Kurt_Budke_0001.jpg
+Christian_Fittipaldi/Christian_Fittipaldi_0001.jpg
+Colin_Powell/Colin_Powell_0195.jpg
+Jessica_Biel/Jessica_Biel_0001.jpg
+Peter_OToole/Peter_OToole_0001.jpg
+Sergio_Vieira_De_Mello/Sergio_Vieira_De_Mello_0006.jpg
+Billy_Sollie/Billy_Sollie_0001.jpg
+Uma_Thurman/Uma_Thurman_0001.jpg
+Andrea_De_Cruz/Andrea_De_Cruz_0001.jpg
+Brandon_Lloyd/Brandon_Lloyd_0001.jpg
+Betsy_Coffin/Betsy_Coffin_0001.jpg
+Daniel_Pearl/Daniel_Pearl_0001.jpg
+Randy_Brown/Randy_Brown_0001.jpg
+Jose_Manuel_Durao_Barroso/Jose_Manuel_Durao_Barroso_0003.jpg
+Ignacio_Antonio_Velasco/Ignacio_Antonio_Velasco_0001.jpg
+Joseph_Salgado/Joseph_Salgado_0001.jpg
+Terry_Stotts/Terry_Stotts_0002.jpg
+Dennis_Erickson/Dennis_Erickson_0002.jpg
+Charles_Lebois/Charles_Lebois_0001.jpg
+John_Cornyn/John_Cornyn_0001.jpg
+Dave_Johnson/Dave_Johnson_0001.jpg
+Frank_Beamer/Frank_Beamer_0001.jpg
+Tyra_Banks/Tyra_Banks_0001.jpg
+Dorothy_Wilson/Dorothy_Wilson_0001.jpg
+Patrice_Chereau/Patrice_Chereau_0002.jpg
+Thomas_Enqvist/Thomas_Enqvist_0001.jpg
+Alexa_Loren/Alexa_Loren_0001.jpg
+Warren_Granados/Warren_Granados_0001.jpg
+Dennis_Oswald/Dennis_Oswald_0001.jpg
+Riek_Blanjaar/Riek_Blanjaar_0001.jpg
+Mary_Bono/Mary_Bono_0001.jpg
+Claire_Tomalin/Claire_Tomalin_0001.jpg
+Makiya_Ali_Hassan/Makiya_Ali_Hassan_0001.jpg
+Mark_Bellhorn/Mark_Bellhorn_0001.jpg
+Carlos_Vives/Carlos_Vives_0001.jpg
+Ricardo_Monasterio/Ricardo_Monasterio_0002.jpg
+Lea_Fastow/Lea_Fastow_0001.jpg
+David_Hasselhoff/David_Hasselhoff_0001.jpg
+Shania_Twain/Shania_Twain_0001.jpg
+Zach_Pillar/Zach_Pillar_0001.jpg
+Maurice_Papon/Maurice_Papon_0001.jpg
+Robert_Duvall/Robert_Duvall_0008.jpg
+Martin_Luther_King_III/Martin_Luther_King_III_0001.jpg
+Mahmoud_Al_Zhar/Mahmoud_Al_Zhar_0001.jpg
+Miguel_Hakim/Miguel_Hakim_0001.jpg
+Tina_Conner/Tina_Conner_0001.jpg
+Valentina_Cervi/Valentina_Cervi_0001.jpg
+Brian_Olson/Brian_Olson_0001.jpg
+Fernando_Gonzalez/Fernando_Gonzalez_0007.jpg
+Tommy_Maddox/Tommy_Maddox_0001.jpg
+Grant_Rossenmeyer/Grant_Rossenmeyer_0001.jpg
+Nova_Esther_Guthrie/Nova_Esther_Guthrie_0001.jpg
+Alex_Sink/Alex_Sink_0003.jpg
+Roger_Clemens/Roger_Clemens_0002.jpg
+William_Shatner/William_Shatner_0001.jpg
+Lucas_Wysocki/Lucas_Wysocki_0001.jpg
+Steve_Phillips/Steve_Phillips_0001.jpg
+Pierre_Pettigrew/Pierre_Pettigrew_0002.jpg
+Elinor_Caplan/Elinor_Caplan_0001.jpg
+Rainer_Geulen/Rainer_Geulen_0001.jpg
+Rey_Sanchez/Rey_Sanchez_0001.jpg
+Patricia_Johnson/Patricia_Johnson_0001.jpg
+Brian_Cowen/Brian_Cowen_0002.jpg
+Michael_Denzel/Michael_Denzel_0001.jpg
+Chuck_Finley/Chuck_Finley_0001.jpg
+Christopher_Amolsch/Christopher_Amolsch_0001.jpg
+Donald_Hays/Donald_Hays_0001.jpg
+Kweisi_Mfume/Kweisi_Mfume_0001.jpg
+Rick_Rickert/Rick_Rickert_0001.jpg
+Stefano_Basalini/Stefano_Basalini_0001.jpg
+Lesley_Flood/Lesley_Flood_0001.jpg
+Lesley_Coppin/Lesley_Coppin_0001.jpg
+Colleen_OClair/Colleen_OClair_0001.jpg
+Robert_F_Kennedy_Jr/Robert_F_Kennedy_Jr_0001.jpg
+Rob_Ramsay/Rob_Ramsay_0001.jpg
+Joe_Mendes/Joe_Mendes_0001.jpg
+Lori_Berenson/Lori_Berenson_0001.jpg
+Elaine_Chao/Elaine_Chao_0001.jpg
+Larry_Flynt/Larry_Flynt_0001.jpg
+Emmanuelle_Beart/Emmanuelle_Beart_0001.jpg
+James_Blake/James_Blake_0009.jpg
+Nicole_Hiltz/Nicole_Hiltz_0001.jpg
+Aleksander_Voloshin/Aleksander_Voloshin_0001.jpg
+Otto_Reich/Otto_Reich_0001.jpg
+Joseph_Deiss/Joseph_Deiss_0001.jpg
+Russell_Crowe/Russell_Crowe_0002.jpg
+Jose_Serra/Jose_Serra_0009.jpg
+Janis_Ruth_Coulter/Janis_Ruth_Coulter_0001.jpg
+Ernie_Grunfeld/Ernie_Grunfeld_0001.jpg
+Mona_Rishmawi/Mona_Rishmawi_0001.jpg
+Tamara_Stokes/Tamara_Stokes_0001.jpg
+Sean_Combs/Sean_Combs_0001.jpg
+Eli_Rosenbaum/Eli_Rosenbaum_0001.jpg
+Mary_Carey/Mary_Carey_0001.jpg
+Saddam_Hussein/Saddam_Hussein_0019.jpg
+Aaron_Pena/Aaron_Pena_0001.jpg
+Barbara_Felt-Miller/Barbara_Felt-Miller_0001.jpg
+Blas_Ople/Blas_Ople_0001.jpg
+Markus_Beyer/Markus_Beyer_0001.jpg
+Shimon_Peres/Shimon_Peres_0007.jpg
+Imad_Moustapha/Imad_Moustapha_0001.jpg
+Mohammed_Dahlan/Mohammed_Dahlan_0001.jpg
+Koichi_Tanaka/Koichi_Tanaka_0001.jpg
+Damon_Stoudamire/Damon_Stoudamire_0001.jpg
+Joe_Pantoliano/Joe_Pantoliano_0001.jpg
+Tavis_Smiley/Tavis_Smiley_0001.jpg
+Matthias_Sammer/Matthias_Sammer_0001.jpg
+Eli_Broad/Eli_Broad_0001.jpg
+Thalia/Thalia_0001.jpg
+San_Lan/San_Lan_0001.jpg
+Johan_Bruyneel/Johan_Bruyneel_0001.jpg
+Tom_Rouen/Tom_Rouen_0001.jpg
+Morgan_Hentzen/Morgan_Hentzen_0001.jpg
+Michael_Stark/Michael_Stark_0001.jpg
+Owen_Wilson/Owen_Wilson_0002.jpg
+Eric_Shinseki/Eric_Shinseki_0001.jpg
+Stepan_Demirchian/Stepan_Demirchian_0001.jpg
+Isabela_Moraes/Isabela_Moraes_0001.jpg
+Toshi_Izawa/Toshi_Izawa_0001.jpg
+Todd_Robbins/Todd_Robbins_0001.jpg
+Nick_Cassavetes/Nick_Cassavetes_0001.jpg
+Eric_Lloyd/Eric_Lloyd_0001.jpg
+Dale_Bosworth/Dale_Bosworth_0001.jpg
+Jennifer_Garner/Jennifer_Garner_0009.jpg
+Anton_Balasingham/Anton_Balasingham_0001.jpg
+Sada_Jacobson/Sada_Jacobson_0001.jpg
+Rogelio_Montemayor/Rogelio_Montemayor_0001.jpg
+Benjamin_Martinez/Benjamin_Martinez_0001.jpg
+Chan_Gailey/Chan_Gailey_0003.jpg
+Ontario_Lett/Ontario_Lett_0001.jpg
+Kim_Chinn/Kim_Chinn_0001.jpg
+Carol_Burnett/Carol_Burnett_0001.jpg
+Eli_Stutsman/Eli_Stutsman_0001.jpg
+Mel_Gibson/Mel_Gibson_0001.jpg
+Henry_Kissinger/Henry_Kissinger_0001.jpg
+Paul_Greengrass/Paul_Greengrass_0001.jpg
+Alfredo_di_Stefano/Alfredo_di_Stefano_0001.jpg
+Leisel_Jones/Leisel_Jones_0001.jpg
+Jiri_Novak/Jiri_Novak_0002.jpg
+Jaouad_Gharib/Jaouad_Gharib_0001.jpg
+Arthur_Johnson/Arthur_Johnson_0001.jpg
+Ashley_Judd/Ashley_Judd_0001.jpg
+Peter_Hollingworth/Peter_Hollingworth_0001.jpg
+Susie_Castillo/Susie_Castillo_0001.jpg
+Marc_Anthony/Marc_Anthony_0001.jpg
+Mark_Schweiker/Mark_Schweiker_0002.jpg
+Madonna/Madonna_0004.jpg
+Adam_Kennedy/Adam_Kennedy_0001.jpg
+Martina_Hingis/Martina_Hingis_0001.jpg
+Dennis_Kozlowski/Dennis_Kozlowski_0001.jpg
+Queen_Sofia/Queen_Sofia_0001.jpg
+Michelle_Yeoh/Michelle_Yeoh_0003.jpg
+Gaston_Gaudio/Gaston_Gaudio_0001.jpg
+Jennifer_McCoy/Jennifer_McCoy_0001.jpg
+Jorge_Moreno/Jorge_Moreno_0001.jpg
+Cori_Enghusen/Cori_Enghusen_0001.jpg
+Ilie_Nastase/Ilie_Nastase_0001.jpg
+Robby_Ginepri/Robby_Ginepri_0001.jpg
+John_Jones/John_Jones_0001.jpg
+Karin_Stoiber/Karin_Stoiber_0002.jpg
+Mona_Locke/Mona_Locke_0001.jpg
+Diego_Armando_Maradona/Diego_Armando_Maradona_0001.jpg
+Jesse_Helms/Jesse_Helms_0001.jpg
+Michael_DeMinico/Michael_DeMinico_0001.jpg
+Dwayne_Wade/Dwayne_Wade_0001.jpg
+Arif_Mardin/Arif_Mardin_0001.jpg
+Ricardo_Lagos/Ricardo_Lagos_0005.jpg
+Penelope_Taylor/Penelope_Taylor_0001.jpg
+Akiko_Morigami/Akiko_Morigami_0001.jpg
+Edward_Flynn/Edward_Flynn_0001.jpg
+Toshimitsu_Motegi/Toshimitsu_Motegi_0001.jpg
+Dick_Clark/Dick_Clark_0002.jpg
+Howard_Stringer/Howard_Stringer_0001.jpg
+Raul_Castaneda/Raul_Castaneda_0001.jpg
+Larry_Hahn/Larry_Hahn_0001.jpg
+Margaret_Okayo/Margaret_Okayo_0001.jpg
+Yann_Martel/Yann_Martel_0001.jpg
+Christian_Bale/Christian_Bale_0001.jpg
+Herb_Brooks/Herb_Brooks_0001.jpg
+Richard_Hamilton/Richard_Hamilton_0001.jpg
+Audrey_Lacroix/Audrey_Lacroix_0001.jpg
+Edwin_Edwards/Edwin_Edwards_0001.jpg
+Yuri_Luzhkov/Yuri_Luzhkov_0001.jpg
+Mika_Hakkinen/Mika_Hakkinen_0001.jpg
+Bill_Kong/Bill_Kong_0001.jpg
+Antanas_Valionis/Antanas_Valionis_0001.jpg
+Fernando_Leon_de_Aranoa/Fernando_Leon_de_Aranoa_0001.jpg
+Nelson_Shanks/Nelson_Shanks_0001.jpg
+Alan_Ball/Alan_Ball_0001.jpg
+Tiffany_Limos/Tiffany_Limos_0001.jpg
+Elvis_Stojko/Elvis_Stojko_0001.jpg
+Nicholas_Byron/Nicholas_Byron_0002.jpg
+Clifford_Robinson/Clifford_Robinson_0001.jpg
+Frank_Taylor/Frank_Taylor_0001.jpg
+Charles_Moose/Charles_Moose_0001.jpg
+Mark_Geragos/Mark_Geragos_0002.jpg
+Cyndi_Thompson/Cyndi_Thompson_0002.jpg
+Dario_Franchitti/Dario_Franchitti_0001.jpg
+Paul_Martin/Paul_Martin_0008.jpg
+Arlen_Specter/Arlen_Specter_0001.jpg
+Narendra_Modi/Narendra_Modi_0001.jpg
+Chang_Tso/Chang_Tso_0001.jpg
+William_Rosenberg/William_Rosenberg_0001.jpg
+Ray_Young/Ray_Young_0001.jpg
+Phil_Cullen/Phil_Cullen_0001.jpg
+Joshua_Gracin/Joshua_Gracin_0001.jpg
+Neri_Marcore/Neri_Marcore_0001.jpg
+Stefaan_Declerk/Stefaan_Declerk_0001.jpg
+Randall_Tobias/Randall_Tobias_0001.jpg
+Alvaro_Silva_Calderon/Alvaro_Silva_Calderon_0001.jpg
+Amelia_Vega/Amelia_Vega_0001.jpg
+Takashi_Yamamoto/Takashi_Yamamoto_0001.jpg
+Margaret_Hoelzer/Margaret_Hoelzer_0001.jpg
+Ridley_Scott/Ridley_Scott_0001.jpg
+Jerry_Bruckheimer/Jerry_Bruckheimer_0001.jpg
+William_Morrow/William_Morrow_0001.jpg
+Dyab_Abou_Jahjah/Dyab_Abou_Jahjah_0001.jpg
+Robin_Johansen/Robin_Johansen_0001.jpg
+Elisabeth_Welch/Elisabeth_Welch_0001.jpg
+Mekhi_Phifer/Mekhi_Phifer_0001.jpg
+Julia_Glass/Julia_Glass_0001.jpg
+Julianna_Margulies/Julianna_Margulies_0002.jpg
+James_Comey/James_Comey_0001.jpg
+Martha_Smith/Martha_Smith_0001.jpg
+Larry_Johnson/Larry_Johnson_0002.jpg
+Alfred_Ford/Alfred_Ford_0001.jpg
+Thomas_Bjorn/Thomas_Bjorn_0001.jpg
+Kultida_Woods/Kultida_Woods_0001.jpg
+Keith_Van_Horn/Keith_Van_Horn_0001.jpg
+Carlos_Alberto_Parreira/Carlos_Alberto_Parreira_0001.jpg
+Bill_Simon/Bill_Simon_0009.jpg
+Janet_Thorpe/Janet_Thorpe_0002.jpg
+Clint_Howard/Clint_Howard_0001.jpg
+Lino_Oviedo/Lino_Oviedo_0002.jpg
+Jackie_Dennis/Jackie_Dennis_0001.jpg
+Cynthia_Nixon/Cynthia_Nixon_0001.jpg
+Desiree_McKenzie/Desiree_McKenzie_0001.jpg
+Richard_Shelby/Richard_Shelby_0002.jpg
+Martin_Short/Martin_Short_0001.jpg
+Martha_Lucia_Ramirez/Martha_Lucia_Ramirez_0004.jpg
+Elton_John/Elton_John_0003.jpg
+Elisha_Cuthbert/Elisha_Cuthbert_0001.jpg
+Franz_Muentefering/Franz_Muentefering_0004.jpg
+Nadia_Petrova/Nadia_Petrova_0004.jpg
+John_Wayne/John_Wayne_0001.jpg
+Gene_Orza/Gene_Orza_0001.jpg
+Diana_Munz/Diana_Munz_0001.jpg
+Ashton_Kutcher/Ashton_Kutcher_0002.jpg
+Bart_Freundlich/Bart_Freundlich_0001.jpg
+Celine_Dion/Celine_Dion_0004.jpg
+Andrew_Firestone/Andrew_Firestone_0001.jpg
+Leslie_Caldwell/Leslie_Caldwell_0001.jpg
+Priyanka_Chopra/Priyanka_Chopra_0001.jpg
+Sean_Astin/Sean_Astin_0002.jpg
+Juljia_Vysotskij/Juljia_Vysotskij_0001.jpg
+Red_Auerbach/Red_Auerbach_0001.jpg
+Maurice_Strong/Maurice_Strong_0001.jpg
+Stephen_Thompson/Stephen_Thompson_0001.jpg
+Pierre_Gagnon/Pierre_Gagnon_0001.jpg
+Eddie_Compass/Eddie_Compass_0001.jpg
+Rachel_Roy/Rachel_Roy_0001.jpg
+Veronica_Lake/Veronica_Lake_0001.jpg
+Sushma_Swaraj/Sushma_Swaraj_0001.jpg
+Michelle_Kwan/Michelle_Kwan_0001.jpg
+Valdas_Adamkus/Valdas_Adamkus_0001.jpg
+Stefan_Tafrov/Stefan_Tafrov_0001.jpg
+Ali_Fallahian/Ali_Fallahian_0001.jpg
+Philip_Cummings/Philip_Cummings_0001.jpg
+Jeremy_Greenstock/Jeremy_Greenstock_0022.jpg
+Ed_Rendell/Ed_Rendell_0001.jpg
+Steven_Hatfill/Steven_Hatfill_0002.jpg
+Tom_Smothers/Tom_Smothers_0001.jpg
+Lionel_Hampton/Lionel_Hampton_0001.jpg
+Mahmoud_Diyab_al-Ahmed/Mahmoud_Diyab_al-Ahmed_0001.jpg
+Ivana_Trump/Ivana_Trump_0001.jpg
+Kareena_Kapoor/Kareena_Kapoor_0001.jpg
+Thomas_Day/Thomas_Day_0001.jpg
+Sammy_Sosa/Sammy_Sosa_0001.jpg
+Jeff_Weaver/Jeff_Weaver_0001.jpg
+Roger_Toussaint/Roger_Toussaint_0001.jpg
+Tessa_Jowell/Tessa_Jowell_0001.jpg
+Thomas_Scavone/Thomas_Scavone_0001.jpg
+Emma_Nicholson/Emma_Nicholson_0001.jpg
+Sheila_Copps/Sheila_Copps_0004.jpg
+Vicki_Zhao_Wei/Vicki_Zhao_Wei_0001.jpg
+George_Karl/George_Karl_0002.jpg
+Marilyn_Monroe/Marilyn_Monroe_0002.jpg
+Hector_Babenco/Hector_Babenco_0001.jpg
+Pamela_Anderson/Pamela_Anderson_0005.jpg
+Griffin_Colvin/Griffin_Colvin_0001.jpg
+David_Obey/David_Obey_0001.jpg
+Jelena_Dokic/Jelena_Dokic_0006.jpg
+Vidar_Helgesen/Vidar_Helgesen_0002.jpg
+Gunter_Pleuger/Gunter_Pleuger_0001.jpg
+Darvis_Patton/Darvis_Patton_0001.jpg
+Jose_Canseco/Jose_Canseco_0003.jpg
+Jennette_Bradley/Jennette_Bradley_0001.jpg
+Graeme_Smith/Graeme_Smith_0001.jpg
+James_Jones/James_Jones_0002.jpg
+Prince_Claus/Prince_Claus_0004.jpg
+James_Brown/James_Brown_0001.jpg
+Jose_Carlo_Fernandez/Jose_Carlo_Fernandez_0001.jpg
+Brandon_Robinson/Brandon_Robinson_0001.jpg
+Kathryn_Bigelow/Kathryn_Bigelow_0001.jpg
+Sharon_Robinson/Sharon_Robinson_0001.jpg
+Moby/Moby_0001.jpg
+Elizabeth_Berkeley/Elizabeth_Berkeley_0001.jpg
+Patricia_Hearst/Patricia_Hearst_0001.jpg
+Park_Jie-won/Park_Jie-won_0001.jpg
+Gordon_McDonald/Gordon_McDonald_0001.jpg
+Jeremy_Wotherspoon/Jeremy_Wotherspoon_0001.jpg
+Paul_Krueger/Paul_Krueger_0001.jpg
+Sergey_Lavrov/Sergey_Lavrov_0009.jpg
+Katie_Wagner/Katie_Wagner_0001.jpg
+Ann_Morgan/Ann_Morgan_0001.jpg
+Bertrand_Delanoe/Bertrand_Delanoe_0001.jpg
+Janet_Crawford/Janet_Crawford_0001.jpg
+Jaromir_Jagr/Jaromir_Jagr_0001.jpg
+Roel_Campos/Roel_Campos_0001.jpg
+Michael_Friedman/Michael_Friedman_0001.jpg
+Erskine_Bowles/Erskine_Bowles_0001.jpg
+Franco_Dragone/Franco_Dragone_0001.jpg
+Keith_Osik/Keith_Osik_0001.jpg
+Derek_King/Derek_King_0001.jpg
+Brandon_Knight/Brandon_Knight_0001.jpg
+Steve_Allee/Steve_Allee_0001.jpg
+Gene_Keady/Gene_Keady_0001.jpg
+Carlos_Quintanilla_Schmidt/Carlos_Quintanilla_Schmidt_0001.jpg
+Keanu_Reeves/Keanu_Reeves_0008.jpg
+Kirk_Ferentz/Kirk_Ferentz_0001.jpg
+Sandy_Smith/Sandy_Smith_0001.jpg
+Nino_DAngelo/Nino_DAngelo_0001.jpg
+Stockard_Channing/Stockard_Channing_0001.jpg
+Mireille_Jospin-Dandieu/Mireille_Jospin-Dandieu_0001.jpg
+Alastair_Campbell/Alastair_Campbell_0001.jpg
+Ricky_Ray/Ricky_Ray_0001.jpg
+James_Coburn/James_Coburn_0001.jpg
+Michael_Kors/Michael_Kors_0001.jpg
+Paul_Bettany/Paul_Bettany_0001.jpg
+Janez_Drnovsek/Janez_Drnovsek_0001.jpg
+Jade_Jagger/Jade_Jagger_0001.jpg
+Jens_Lehmann/Jens_Lehmann_0001.jpg
+Rustu_Recber/Rustu_Recber_0001.jpg
+Michael_Linscott/Michael_Linscott_0001.jpg
+Ellen_Engleman/Ellen_Engleman_0002.jpg
+George_McCloud/George_McCloud_0001.jpg
+Kevin_Costner/Kevin_Costner_0008.jpg
+David_Gest/David_Gest_0001.jpg
+Sasha_Alexander/Sasha_Alexander_0001.jpg
+Mike_Weir/Mike_Weir_0001.jpg
+Andrea_Yates/Andrea_Yates_0001.jpg
+Ahmet_Necdet_Sezer/Ahmet_Necdet_Sezer_0001.jpg
+Arianna_Huffington/Arianna_Huffington_0002.jpg
+Pringe_Ernst_August/Pringe_Ernst_August_0001.jpg
+Gerard_Butler/Gerard_Butler_0001.jpg
+Michael_Keaton/Michael_Keaton_0002.jpg
+Troy_Polamalu/Troy_Polamalu_0001.jpg
+Don_Siegelman/Don_Siegelman_0004.jpg
+Jada_Pinkett_Smith/Jada_Pinkett_Smith_0002.jpg
+Carlton_Baugh/Carlton_Baugh_0001.jpg
+Jose_Carreras/Jose_Carreras_0001.jpg
+Jim_Talent/Jim_Talent_0001.jpg
+Priscilla_Presley/Priscilla_Presley_0001.jpg
+Gil_Cates/Gil_Cates_0001.jpg
+Yang_Jianli/Yang_Jianli_0001.jpg
+Eva_Herzigova/Eva_Herzigova_0001.jpg
+Thomas_Ulrich/Thomas_Ulrich_0001.jpg
+Louisa_Baileche/Louisa_Baileche_0001.jpg
+Davis_Love_III/Davis_Love_III_0001.jpg
+Larry_Thompson/Larry_Thompson_0003.jpg
+Dan_Quayle/Dan_Quayle_0001.jpg
+Thomas_Mesereau_Jr/Thomas_Mesereau_Jr_0001.jpg
+Sean_OKeefe/Sean_OKeefe_0003.jpg
+Vincent_Gallo/Vincent_Gallo_0001.jpg
+Kevin_Hearn/Kevin_Hearn_0001.jpg
+Lance_Armstrong/Lance_Armstrong_0003.jpg
+Sheila_Fraser/Sheila_Fraser_0002.jpg
+Stan_Kasten/Stan_Kasten_0001.jpg
+Gary_Condit/Gary_Condit_0001.jpg
+Mufti_Mohammad_Syed/Mufti_Mohammad_Syed_0001.jpg
+Kristin_Davis/Kristin_Davis_0003.jpg
+Tara_Dawn_Christensen/Tara_Dawn_Christensen_0001.jpg
+Pedro_Malan/Pedro_Malan_0004.jpg
+Jose_Mourinho/Jose_Mourinho_0001.jpg
+Joy_Lee_Sadler/Joy_Lee_Sadler_0001.jpg
+Anastasia_Myskina/Anastasia_Myskina_0003.jpg
+Carin_Koch/Carin_Koch_0001.jpg
+Guillermo_Ortiz/Guillermo_Ortiz_0002.jpg
+Fatmir_Limaj/Fatmir_Limaj_0001.jpg
+Val_Ackerman/Val_Ackerman_0001.jpg
+Vitali_Klitschko/Vitali_Klitschko_0004.jpg
+Shane_Loux/Shane_Loux_0001.jpg
+Melissa_Etheridge/Melissa_Etheridge_0001.jpg
+Bill_Belichick/Bill_Belichick_0002.jpg
+Jean_Nagel/Jean_Nagel_0001.jpg
+Steve_Coogan/Steve_Coogan_0001.jpg
+Rachel_Wadsworth/Rachel_Wadsworth_0001.jpg
+Nur_Jaafar/Nur_Jaafar_0001.jpg
+Maureen_Kanka/Maureen_Kanka_0001.jpg
+Job_Cohen/Job_Cohen_0002.jpg
+Sergio_Castellitto/Sergio_Castellitto_0001.jpg
+Meghann_Shaughnessy/Meghann_Shaughnessy_0002.jpg
+Chuck_Woolery/Chuck_Woolery_0001.jpg
+Patrick_McEnroe/Patrick_McEnroe_0002.jpg
+Fabiola_Zuluaga/Fabiola_Zuluaga_0002.jpg
+Franz_Gsell/Franz_Gsell_0001.jpg
+Chris_Penn/Chris_Penn_0001.jpg
+Joshua_Davey/Joshua_Davey_0001.jpg
+Yevgeny_Kafelnikov/Yevgeny_Kafelnikov_0001.jpg
+Nicolas_Kiefer/Nicolas_Kiefer_0001.jpg
+Kajsa_Bergqvist/Kajsa_Bergqvist_0001.jpg
+Simon_Cowell/Simon_Cowell_0001.jpg
+Mae_Jemison/Mae_Jemison_0001.jpg
+Larry_Coker/Larry_Coker_0004.jpg
+Omar_Khan_Sharif/Omar_Khan_Sharif_0001.jpg
+Norbert_van_Heyst/Norbert_van_Heyst_0001.jpg
+Lynne_Thigpen/Lynne_Thigpen_0001.jpg
+Sammy_Knight/Sammy_Knight_0001.jpg
+Larry_Greene/Larry_Greene_0001.jpg
+Garth_Drabinsky/Garth_Drabinsky_0001.jpg
+Chin-Hui_Tsao/Chin-Hui_Tsao_0001.jpg
+John_Geoghan/John_Geoghan_0001.jpg
+King_Bhumibol_Adulyadej/King_Bhumibol_Adulyadej_0001.jpg
+Frank_Keating/Frank_Keating_0001.jpg
+Zaini_Abdullah/Zaini_Abdullah_0001.jpg
+Gao_Qiang/Gao_Qiang_0001.jpg
+Yasushi_Akashi/Yasushi_Akashi_0001.jpg
+Joan_Laporta/Joan_Laporta_0009.jpg
+Paula_Abdul/Paula_Abdul_0001.jpg
+Anne_Heche/Anne_Heche_0001.jpg
+Brian_Weaver/Brian_Weaver_0001.jpg
+Andre_Bucher/Andre_Bucher_0001.jpg
+Joe_Gatti/Joe_Gatti_0001.jpg
+Eriko_Tamura/Eriko_Tamura_0001.jpg
+Diego_Colorado/Diego_Colorado_0001.jpg
+Lene_Espersen/Lene_Espersen_0001.jpg
+Felix_Trinidad/Felix_Trinidad_0001.jpg
+David_Suazo/David_Suazo_0001.jpg
+Jose_Lopez_Beltran/Jose_Lopez_Beltran_0001.jpg
+Jodie_Kidd/Jodie_Kidd_0001.jpg
+Joanna_Poitier/Joanna_Poitier_0001.jpg
+Cindy_Klassen/Cindy_Klassen_0001.jpg
+Demetrin_Veal/Demetrin_Veal_0001.jpg
+Madeleine_Albright/Madeleine_Albright_0001.jpg
+Hussam_Mohammed_Amin/Hussam_Mohammed_Amin_0001.jpg
+Hans_Leistritz/Hans_Leistritz_0001.jpg
+Itzhak_Perlman/Itzhak_Perlman_0001.jpg
+Edie_Falco/Edie_Falco_0002.jpg
+Allison_Janney/Allison_Janney_0001.jpg
+Lindsay_Davenport/Lindsay_Davenport_0009.jpg
+Adrian_McPherson/Adrian_McPherson_0002.jpg
+Elena_Bovina/Elena_Bovina_0001.jpg
+Nicholoas_DiMarzio/Nicholoas_DiMarzio_0001.jpg
+Mike_Helton/Mike_Helton_0001.jpg
+Nicolas_Eyzaguirre/Nicolas_Eyzaguirre_0001.jpg
+Marlene_Weingartner/Marlene_Weingartner_0001.jpg
+David_Donohue/David_Donohue_0001.jpg
+Monica_Lewinsky/Monica_Lewinsky_0002.jpg
+Boris_Henry/Boris_Henry_0001.jpg
+Marta_Dominguz/Marta_Dominguz_0001.jpg
+Jason_Sehorn/Jason_Sehorn_0001.jpg
+Peter_Mackay/Peter_Mackay_0001.jpg
+Debra_Yang/Debra_Yang_0001.jpg
+Lawrence_Foley/Lawrence_Foley_0001.jpg
+Hootie_Johnson/Hootie_Johnson_0001.jpg
+Al_Davis/Al_Davis_0001.jpg
+Melinda_Czink/Melinda_Czink_0001.jpg
+Anjum_Hussain/Anjum_Hussain_0001.jpg
+Oracene_Williams/Oracene_Williams_0001.jpg
+Ronald_Harwood/Ronald_Harwood_0001.jpg
+Ibrahim_Jaafari/Ibrahim_Jaafari_0001.jpg
+Brian_Mulroney/Brian_Mulroney_0002.jpg
+Mike_Maroth/Mike_Maroth_0001.jpg
+Dora_Bakoyianni/Dora_Bakoyianni_0001.jpg
+Conchita_Martinez/Conchita_Martinez_0002.jpg
+Arantxa_Sanchez-Vicario/Arantxa_Sanchez-Vicario_0002.jpg
+Paul_Wals/Paul_Wals_0001.jpg
+Carlos_Ruiz/Carlos_Ruiz_0001.jpg
+Naomi_Hayashi/Naomi_Hayashi_0001.jpg
+Gunilla_Backman/Gunilla_Backman_0001.jpg
+Krishna_Bhadur_Mahara/Krishna_Bhadur_Mahara_0001.jpg
+John_Nash/John_Nash_0001.jpg
+Amber_Tamblyn/Amber_Tamblyn_0002.jpg
+Clifford_Etienne/Clifford_Etienne_0001.jpg
+Bob_Colvin/Bob_Colvin_0002.jpg
+Ray_Nagin/Ray_Nagin_0001.jpg
+Nida_Blanca/Nida_Blanca_0001.jpg
+Dieter_Zetsche/Dieter_Zetsche_0001.jpg
+Tomomi_Morita/Tomomi_Morita_0001.jpg
+Rob_Schneider/Rob_Schneider_0001.jpg
+Roger_Moore/Roger_Moore_0003.jpg
+Joseph_Ganim/Joseph_Ganim_0001.jpg
+Steve_Nesbitt/Steve_Nesbitt_0001.jpg
+Jon_Voight/Jon_Voight_0001.jpg
+Felicity_Huffman/Felicity_Huffman_0001.jpg
+Anthony_Scott_Miller/Anthony_Scott_Miller_0001.jpg
+Howard_Wilkinson/Howard_Wilkinson_0001.jpg
+John_Rosa/John_Rosa_0001.jpg
+Janet_Chandler/Janet_Chandler_0001.jpg
+Jennifer_Granholm/Jennifer_Granholm_0001.jpg
+Princess_Victoria/Princess_Victoria_0001.jpg
+Thomas_Franklin/Thomas_Franklin_0001.jpg
+Robert_Weitzel/Robert_Weitzel_0001.jpg
+Juan_Roman_Carrasco/Juan_Roman_Carrasco_0001.jpg
+Carolyn_Dawn_Johnson/Carolyn_Dawn_Johnson_0001.jpg
+Stephen_Ebberharter/Stephen_Ebberharter_0001.jpg
+Phillipe_Comtois/Phillipe_Comtois_0001.jpg
+David_Canary/David_Canary_0001.jpg
+Alex_Penelas/Alex_Penelas_0001.jpg
+David_Rivkin_Jr/David_Rivkin_Jr_0001.jpg
+Daniel_Rouse/Daniel_Rouse_0001.jpg
+AJ_Cook/AJ_Cook_0001.jpg
+Richard_Barry/Richard_Barry_0001.jpg
+Toni_Braxton/Toni_Braxton_0002.jpg
+Johnny_Unitas/Johnny_Unitas_0001.jpg
+Jason_Jennings/Jason_Jennings_0002.jpg
+James_Hakett/James_Hakett_0001.jpg
+Robert_Stack/Robert_Stack_0002.jpg
+Mike_OConnell/Mike_OConnell_0001.jpg
+Aaron_Tippin/Aaron_Tippin_0001.jpg
+Roberta_Combs/Roberta_Combs_0001.jpg
+David_Collenette/David_Collenette_0001.jpg
+Michael_Andretti/Michael_Andretti_0001.jpg
+Francis_Crick/Francis_Crick_0001.jpg
+Nan_Wang/Nan_Wang_0003.jpg
+Roberto_Robaina/Roberto_Robaina_0001.jpg
+Oscar_Elias_Biscet/Oscar_Elias_Biscet_0001.jpg
+Calvin_Harrison/Calvin_Harrison_0001.jpg
+Lesia_Burlak/Lesia_Burlak_0001.jpg
+Michael_Shelby/Michael_Shelby_0001.jpg
+Ashley_Postell/Ashley_Postell_0001.jpg
+Marlon_Devonish/Marlon_Devonish_0001.jpg
+Donald_Regan/Donald_Regan_0001.jpg
+Svend_Aage_Jensby/Svend_Aage_Jensby_0001.jpg
+Tim_Curley/Tim_Curley_0001.jpg
+Barbara_Roberts/Barbara_Roberts_0001.jpg
+Osrat_Iosef/Osrat_Iosef_0001.jpg
+Robert_Blackwill/Robert_Blackwill_0001.jpg
+John_Elway/John_Elway_0001.jpg
+Michael_Killeen/Michael_Killeen_0001.jpg
+Kent_Rominger/Kent_Rominger_0001.jpg
+Dennis_Miller/Dennis_Miller_0001.jpg
+Alexander_Downer/Alexander_Downer_0003.jpg
+Milton_Wynants/Milton_Wynants_0001.jpg
+Gisele_Bundchen/Gisele_Bundchen_0001.jpg
+Tatiana_Paus/Tatiana_Paus_0001.jpg
+Chris_Noth/Chris_Noth_0001.jpg
+Lee_Ann_Knight/Lee_Ann_Knight_0001.jpg
+Patricia_Phillips/Patricia_Phillips_0001.jpg
+Jason_Keep/Jason_Keep_0001.jpg
+Bernard_Ebbers/Bernard_Ebbers_0001.jpg
+Jonathan_Byrd/Jonathan_Byrd_0001.jpg
+Heather_Whitestone_McCallum/Heather_Whitestone_McCallum_0001.jpg
+Bill_Mauldin/Bill_Mauldin_0001.jpg
+Ruben_Studdard/Ruben_Studdard_0001.jpg
+Benjamin_Netanyahu/Benjamin_Netanyahu_0005.jpg
+Placido_Domingo/Placido_Domingo_0001.jpg
+Pat_Summitt/Pat_Summitt_0001.jpg
+Johnny_Tapia/Johnny_Tapia_0002.jpg
+Stephen_Silas/Stephen_Silas_0001.jpg
+Angie_Martinez/Angie_Martinez_0001.jpg
+Casey_Mears/Casey_Mears_0001.jpg
+Gianni_Agnelli/Gianni_Agnelli_0001.jpg
+Clare_Latimer/Clare_Latimer_0001.jpg
+Alejandro_Avila/Alejandro_Avila_0001.jpg
+Calvin_Joseph_Coleman/Calvin_Joseph_Coleman_0001.jpg
+Lenny_Wilkens/Lenny_Wilkens_0001.jpg
+Carl_Reiner/Carl_Reiner_0001.jpg
+Daniel_Montgomery/Daniel_Montgomery_0001.jpg
+Hans_Corell/Hans_Corell_0001.jpg
+Jodie_Henry/Jodie_Henry_0001.jpg
+Dalai_Lama/Dalai_Lama_0002.jpg
+Wen_Jiabao/Wen_Jiabao_0008.jpg
+Michael_Milton/Michael_Milton_0001.jpg
+Chok_Tong_Goh/Chok_Tong_Goh_0001.jpg
+Joey_Buttafuoco/Joey_Buttafuoco_0001.jpg
+Tony_Parker/Tony_Parker_0001.jpg
+Brian_Lara/Brian_Lara_0001.jpg
+Queen_Latifah/Queen_Latifah_0002.jpg
+Gerard_Depardieu/Gerard_Depardieu_0002.jpg
+Peter_Mansbridge/Peter_Mansbridge_0001.jpg
+Daniel_Scioli/Daniel_Scioli_0001.jpg
+Brian_Griese/Brian_Griese_0001.jpg
+Ross_Verba/Ross_Verba_0001.jpg
+Stuart_Townsend/Stuart_Townsend_0001.jpg
+Jesper_Parnevik/Jesper_Parnevik_0001.jpg
+Larry_Bowa/Larry_Bowa_0002.jpg
+Dave_Campo/Dave_Campo_0001.jpg
+Lionel_Chalmers/Lionel_Chalmers_0001.jpg
+Rose_Linkins/Rose_Linkins_0001.jpg
+Khum_Bahadur_Khadka/Khum_Bahadur_Khadka_0001.jpg
+Prince_Philippe/Prince_Philippe_0001.jpg
+Francisco_Maturana/Francisco_Maturana_0001.jpg
+Jonathan_Horton/Jonathan_Horton_0001.jpg
+Takahiro_Mori/Takahiro_Mori_0001.jpg
+Will_Smith/Will_Smith_0002.jpg
+Li_Changchun/Li_Changchun_0001.jpg
+Cecilia_Bolocco/Cecilia_Bolocco_0002.jpg
+Richie_Adubato/Richie_Adubato_0002.jpg
+Masaru_Hayami/Masaru_Hayami_0001.jpg
+Skip_Prosser/Skip_Prosser_0001.jpg
+Melvin_Talbert/Melvin_Talbert_0001.jpg
+Syed_Abdul_Rahman_Geelani/Syed_Abdul_Rahman_Geelani_0001.jpg
+Felipe_Fernandez/Felipe_Fernandez_0001.jpg
+Stipe_Mesic/Stipe_Mesic_0001.jpg
+Joaquin_Sanchez/Joaquin_Sanchez_0001.jpg
+Robert_Marshall/Robert_Marshall_0001.jpg
+Tara_VanDerveer/Tara_VanDerveer_0001.jpg
+Mike_Carona/Mike_Carona_0001.jpg
+Claudia_Schiffer/Claudia_Schiffer_0003.jpg
+Jamie_Villafane/Jamie_Villafane_0001.jpg
+Viola_Davis/Viola_Davis_0001.jpg
+Charles_Pouty/Charles_Pouty_0001.jpg
+Laszlo_Kovacs/Laszlo_Kovacs_0001.jpg
+Howard_Ross/Howard_Ross_0001.jpg
+Tom_Welch/Tom_Welch_0001.jpg
+Bob_Melvin/Bob_Melvin_0001.jpg
+John_Allen_Muhammad/John_Allen_Muhammad_0001.jpg
+Jerome_Golmard/Jerome_Golmard_0001.jpg
+Carolina_Barco/Carolina_Barco_0001.jpg
+Craig_OClair/Craig_OClair_0001.jpg
+Stephen_Keener/Stephen_Keener_0001.jpg
+Byron_Scott/Byron_Scott_0001.jpg
+Ken_Balk/Ken_Balk_0001.jpg
+Joe_Dumars/Joe_Dumars_0002.jpg
+Lynn_Redgrave/Lynn_Redgrave_0002.jpg
+Ahmad_Masood/Ahmad_Masood_0002.jpg
+Hidetoshi_Nakata/Hidetoshi_Nakata_0001.jpg
+Kim_Ryong-sung/Kim_Ryong-sung_0004.jpg
+William_Genego/William_Genego_0001.jpg
+Vinnie_Jones/Vinnie_Jones_0001.jpg
+Amram_Mitzna/Amram_Mitzna_0001.jpg
+Chen_Shui-bian/Chen_Shui-bian_0002.jpg
+Marina_Kuptsova/Marina_Kuptsova_0001.jpg
+Tom_Brennan/Tom_Brennan_0001.jpg
+Javier_Vargas/Javier_Vargas_0001.jpg
+Timothy_McVeigh/Timothy_McVeigh_0001.jpg
+Francisco_Santos/Francisco_Santos_0001.jpg
+Graham_Bentley/Graham_Bentley_0001.jpg
+Jeanne_Moreau/Jeanne_Moreau_0001.jpg
+Karen_Pereiras/Karen_Pereiras_0001.jpg
+Ray_Allen/Ray_Allen_0002.jpg
+Albert_Costa/Albert_Costa_0004.jpg
+Grace_Kelly/Grace_Kelly_0001.jpg
+Marianne_Stanley/Marianne_Stanley_0001.jpg
+Willie_Wilson/Willie_Wilson_0001.jpg
+Sam_Torrance/Sam_Torrance_0002.jpg
+Claudio_Lopez/Claudio_Lopez_0001.jpg
+Mahathir_Mohamad/Mahathir_Mohamad_0004.jpg
+Mike_Holmgren/Mike_Holmgren_0001.jpg
+Jim_Taylor/Jim_Taylor_0001.jpg
+Anna_Nicole_Smith/Anna_Nicole_Smith_0001.jpg
+Geraldo_Rivera/Geraldo_Rivera_0001.jpg
+Gerardo_Gambala/Gerardo_Gambala_0002.jpg
+Michael_Munoz/Michael_Munoz_0001.jpg
+Chris_Dodd/Chris_Dodd_0001.jpg
+Steve_Spurrier/Steve_Spurrier_0002.jpg
+Mitoji_Yabunaka/Mitoji_Yabunaka_0002.jpg
+Tim_Robbins/Tim_Robbins_0005.jpg
+Jean-Marc_Olive/Jean-Marc_Olive_0001.jpg
+Jason_Clermont/Jason_Clermont_0001.jpg
+Gilberto_Simoni/Gilberto_Simoni_0001.jpg
+Himmler_Rebu/Himmler_Rebu_0001.jpg
+Ralph_Klein/Ralph_Klein_0001.jpg
+Florecita_Cobian/Florecita_Cobian_0001.jpg
+Enrique_Iglesias/Enrique_Iglesias_0001.jpg
+SJ_Twu/SJ_Twu_0001.jpg
+Taoufik_Mathlouthi/Taoufik_Mathlouthi_0001.jpg
+Kim_Jin-sun/Kim_Jin-sun_0001.jpg
+Francis_Collins/Francis_Collins_0001.jpg
+Ronald_Kessler/Ronald_Kessler_0001.jpg
+David_Spade/David_Spade_0002.jpg
+Christy_Ferer/Christy_Ferer_0001.jpg
+Matt_Doherty/Matt_Doherty_0002.jpg
+George_Clooney/George_Clooney_0007.jpg
+Tommy_Robredo/Tommy_Robredo_0001.jpg
+Liv_Tyler/Liv_Tyler_0001.jpg
+Morris_Dees/Morris_Dees_0001.jpg
+Jaime_Orti/Jaime_Orti_0001.jpg
+Mohammaed_Ahmad_Al_Jarallah/Mohammaed_Ahmad_Al_Jarallah_0001.jpg
+Curtis_Strange/Curtis_Strange_0001.jpg
+Ruben_Wolkowyski/Ruben_Wolkowyski_0001.jpg
+Sean_Patrick_Thomas/Sean_Patrick_Thomas_0001.jpg
+Marisol_Breton/Marisol_Breton_0001.jpg
+Chang_Saio-yue/Chang_Saio-yue_0001.jpg
+Larry_Brown/Larry_Brown_0004.jpg
+Mickey_Gilley/Mickey_Gilley_0001.jpg
+Michael_Ballack/Michael_Ballack_0003.jpg
+Coleen_Rowley/Coleen_Rowley_0001.jpg
+Louis_Van_Gaal/Louis_Van_Gaal_0001.jpg
+Richard_Butler/Richard_Butler_0001.jpg
+Kalpana_Chawla/Kalpana_Chawla_0002.jpg
+Michael_Doleac/Michael_Doleac_0001.jpg
+Aretha_Franklin/Aretha_Franklin_0001.jpg
+Maria_Garcia/Maria_Garcia_0001.jpg
+Hugo_Colace/Hugo_Colace_0001.jpg
+Robert_Redford/Robert_Redford_0008.jpg
+Sofia_Milos/Sofia_Milos_0001.jpg
+Riccardo_Muti/Riccardo_Muti_0001.jpg
+Faisal_Iqbal/Faisal_Iqbal_0001.jpg
+Sara_Silverman/Sara_Silverman_0001.jpg
+Chris_Columbus/Chris_Columbus_0001.jpg
+Harry_Kalas/Harry_Kalas_0001.jpg
+Melina_Kanakaredes/Melina_Kanakaredes_0001.jpg
+Tom_Craddick/Tom_Craddick_0003.jpg
+Laura_Morante/Laura_Morante_0001.jpg
+Sheikh_Ahmed_Yassin/Sheikh_Ahmed_Yassin_0001.jpg
+Szu_Yu_Chen/Szu_Yu_Chen_0001.jpg
+Harvey_Weinstein/Harvey_Weinstein_0002.jpg
+Thomas_Stewart/Thomas_Stewart_0001.jpg
+Della_Clara/Della_Clara_0001.jpg
+Mike_Matheny/Mike_Matheny_0001.jpg
+Jose_Dirceu/Jose_Dirceu_0001.jpg
+Jamie_Lee_Curtis/Jamie_Lee_Curtis_0001.jpg
+Ron_Dittemore/Ron_Dittemore_0008.jpg
+Gideon_Black/Gideon_Black_0001.jpg
+John_Bolton/John_Bolton_0008.jpg
+Jonathan_Mostow/Jonathan_Mostow_0001.jpg
+William_Harrison/William_Harrison_0001.jpg
+Kenny_Chesney/Kenny_Chesney_0001.jpg
+Abdullah/Abdullah_0004.jpg
+Takeshi_Kitano/Takeshi_Kitano_0001.jpg
+Kim_Hong-gul/Kim_Hong-gul_0001.jpg
+Andy_Rooney/Andy_Rooney_0001.jpg
+Chris_Gratton/Chris_Gratton_0001.jpg
+Angela_Merkel/Angela_Merkel_0005.jpg
+Eric_Snow/Eric_Snow_0001.jpg
+Eric_Bana/Eric_Bana_0001.jpg
+John_Rigas/John_Rigas_0001.jpg
+Juliette_Binoche/Juliette_Binoche_0001.jpg
+Barbora_Strycova/Barbora_Strycova_0001.jpg
+Donna_Morrissey/Donna_Morrissey_0001.jpg
+Gen_Meredith/Gen_Meredith_0001.jpg
+Makhdoom_Amin_Fahim/Makhdoom_Amin_Fahim_0001.jpg
+Hugo_Chavez/Hugo_Chavez_0003.jpg
+Jeffrey_Katzenberg/Jeffrey_Katzenberg_0001.jpg
+Peter_Rasmussen/Peter_Rasmussen_0001.jpg
+Howard_Smith/Howard_Smith_0001.jpg
+Lisa_Stone/Lisa_Stone_0001.jpg
+Jeremy_Gompertz/Jeremy_Gompertz_0001.jpg
+Helena_Schneider/Helena_Schneider_0001.jpg
+Alejandro_Fernandez/Alejandro_Fernandez_0001.jpg
+Paul_Luvera/Paul_Luvera_0001.jpg
+Lane_Bryant/Lane_Bryant_0001.jpg
+Paul_Gannon/Paul_Gannon_0001.jpg
+Christine_Rau/Christine_Rau_0001.jpg
+Jane_Russell/Jane_Russell_0001.jpg
+Madge_Overhouse/Madge_Overhouse_0001.jpg
+Alyse_Beaupre/Alyse_Beaupre_0001.jpg
+Jack_LaLanne/Jack_LaLanne_0001.jpg
+Stanley_Ho/Stanley_Ho_0001.jpg
+Alisha_Richman/Alisha_Richman_0001.jpg
+Nicole_Kidman/Nicole_Kidman_0013.jpg
+Ernie_Els/Ernie_Els_0001.jpg
+Monique_Gagnon-Tremblay/Monique_Gagnon-Tremblay_0001.jpg
+Josh_Kronfeld/Josh_Kronfeld_0001.jpg
+David_Brinkley/David_Brinkley_0001.jpg
+Carla_Sullivan/Carla_Sullivan_0001.jpg
+Julianne_Moore/Julianne_Moore_0009.jpg
+Lena_Katina/Lena_Katina_0001.jpg
+Chang_Jae_On/Chang_Jae_On_0001.jpg
+John_Spencer/John_Spencer_0001.jpg
+Kathryn_Grayson/Kathryn_Grayson_0001.jpg
+Dudley_Rogers/Dudley_Rogers_0001.jpg
+Alicia_Silverstone/Alicia_Silverstone_0001.jpg
+Yoshiyuki_Kamei/Yoshiyuki_Kamei_0001.jpg
+Shia_LaBeouf/Shia_LaBeouf_0001.jpg
+Pat_Wharton/Pat_Wharton_0001.jpg
+George_Blaney/George_Blaney_0001.jpg
+Larry_Eustachy/Larry_Eustachy_0001.jpg
+Choi_Sung-hong/Choi_Sung-hong_0005.jpg
+Jennifer_Tilly/Jennifer_Tilly_0001.jpg
+Abdul_Majeed_Shobokshi/Abdul_Majeed_Shobokshi_0001.jpg
+Philip_Murtaugh/Philip_Murtaugh_0001.jpg
+Chang_Sang/Chang_Sang_0001.jpg
+John_Lennon/John_Lennon_0001.jpg
+Zafarullah_Khan_Jamali/Zafarullah_Khan_Jamali_0001.jpg
+Ravan_AG_Farhadi/Ravan_AG_Farhadi_0001.jpg
+David_Hanson/David_Hanson_0001.jpg
+Bryan_Cooley/Bryan_Cooley_0001.jpg
+Dionne_Warwick/Dionne_Warwick_0001.jpg
+Manfred_Reyes_Villa/Manfred_Reyes_Villa_0001.jpg
+Gary_Doer/Gary_Doer_0001.jpg
+Meg_Mallon/Meg_Mallon_0001.jpg
+Alfonso_Portillo/Alfonso_Portillo_0001.jpg
+Ronald_White/Ronald_White_0001.jpg
+Donna_Ralston/Donna_Ralston_0001.jpg
+Frank_Solich/Frank_Solich_0004.jpg
+Mark_Sisk/Mark_Sisk_0001.jpg
+Charles_Grassley/Charles_Grassley_0002.jpg
+Eduardo_Chillida/Eduardo_Chillida_0001.jpg
+Anthony_LaPaglia/Anthony_LaPaglia_0002.jpg
+Allen_Rock/Allen_Rock_0001.jpg
+Katharine_Hepburn/Katharine_Hepburn_0001.jpg
+Ain_Seppik/Ain_Seppik_0001.jpg
+Charla_Moye/Charla_Moye_0001.jpg
+Einars_Repse/Einars_Repse_0001.jpg
+Kevin_Satterfield/Kevin_Satterfield_0001.jpg
+Marquier_Montano_Contreras/Marquier_Montano_Contreras_0001.jpg
+Alan_Zemaitis/Alan_Zemaitis_0001.jpg
+Yishan_Zhang/Yishan_Zhang_0001.jpg
+Reyyan_Uzuner/Reyyan_Uzuner_0001.jpg
+Kay_Behrensmeyer/Kay_Behrensmeyer_0001.jpg
+Flavia_Pennetta/Flavia_Pennetta_0001.jpg
+Mario_Puzo/Mario_Puzo_0001.jpg
+Zhong_Nanshan/Zhong_Nanshan_0001.jpg
+Todd_Petit/Todd_Petit_0001.jpg
+Ashraf_Ghani/Ashraf_Ghani_0001.jpg
+Douglas_Gansler/Douglas_Gansler_0001.jpg
+Nebojsa_Pavkovic/Nebojsa_Pavkovic_0001.jpg
+Howard_Dean/Howard_Dean_0008.jpg
+Prince_Harry/Prince_Harry_0002.jpg
+David_Sibleyk/David_Sibleyk_0001.jpg
+Bill_Curry/Bill_Curry_0001.jpg
+Hisham_Halawi/Hisham_Halawi_0001.jpg
+Timothy_Wirth/Timothy_Wirth_0001.jpg
+Laurie_Laychak/Laurie_Laychak_0001.jpg
+Matthew_Perry/Matthew_Perry_0006.jpg
+Chris_Forsyth/Chris_Forsyth_0001.jpg
+Paul_Murphy/Paul_Murphy_0001.jpg
+Brittany_Snow/Brittany_Snow_0001.jpg
+Rupert_Murdoch/Rupert_Murdoch_0001.jpg
+Joe_Mantello/Joe_Mantello_0001.jpg
+Lee_Hoi-chang/Lee_Hoi-chang_0001.jpg
+Jennifer_Capriati/Jennifer_Capriati_0016.jpg
+John_Travolta/John_Travolta_0001.jpg
+Dan_Guerrero/Dan_Guerrero_0001.jpg
+Marcelo_Rios/Marcelo_Rios_0003.jpg
+Michael_Hagee/Michael_Hagee_0001.jpg
+Natalia_Verbeke/Natalia_Verbeke_0001.jpg
+Jon_Corzine/Jon_Corzine_0002.jpg
+Iran_Brown/Iran_Brown_0001.jpg
+Pete_Carroll/Pete_Carroll_0003.jpg
+Joe_Carnahan/Joe_Carnahan_0001.jpg
+Jim_Nochols/Jim_Nochols_0001.jpg
+Charles_Chandler_IV/Charles_Chandler_IV_0001.jpg
+John_Prescott/John_Prescott_0001.jpg
+Lord_Hutton/Lord_Hutton_0002.jpg
+Ken_Kutaragi/Ken_Kutaragi_0001.jpg
+Patti_Labelle/Patti_Labelle_0002.jpg
+Didier_Defago/Didier_Defago_0001.jpg
+Halle_Berry/Halle_Berry_0014.jpg
+Simona_Hradil/Simona_Hradil_0001.jpg
+James_Ballenger/James_Ballenger_0001.jpg
+Lewis_Booth/Lewis_Booth_0001.jpg
+Marie-Reine_Le_Gougne/Marie-Reine_Le_Gougne_0001.jpg
+Paul_Kariya/Paul_Kariya_0001.jpg
+Russell_Coutts/Russell_Coutts_0001.jpg
+John_Wolf/John_Wolf_0001.jpg
+Surakait_Sathirathai/Surakait_Sathirathai_0001.jpg
+Abid_Hamid_Mahmud_Al-Tikriti/Abid_Hamid_Mahmud_Al-Tikriti_0003.jpg
+Moon-So-ri/Moon-So-ri_0001.jpg
+Stefan_Koubek/Stefan_Koubek_0001.jpg
+Garry_Witherall/Garry_Witherall_0001.jpg
+Bridget_Fonda/Bridget_Fonda_0001.jpg
+Amy_Gale/Amy_Gale_0001.jpg
+Jose_Canseco_Sr/Jose_Canseco_Sr_0001.jpg
+Philip_Zalewski/Philip_Zalewski_0001.jpg
+Jason_Kidd/Jason_Kidd_0004.jpg
+Chante_Jawan_Mallard/Chante_Jawan_Mallard_0001.jpg
+Percy_Gibson/Percy_Gibson_0001.jpg
+Jean_Carnahan/Jean_Carnahan_0001.jpg
+Hu_Maoyuan/Hu_Maoyuan_0001.jpg
+Laurie_Hobbs/Laurie_Hobbs_0001.jpg
+Jong_Wook_Lee/Jong_Wook_Lee_0004.jpg
+Robert_Mugabe/Robert_Mugabe_0001.jpg
+Akmal_Taher/Akmal_Taher_0001.jpg
+Sharon_Davis/Sharon_Davis_0001.jpg
+Liu_Mingkang/Liu_Mingkang_0002.jpg
+John_Darby/John_Darby_0001.jpg
+Ibrahim_Rugova/Ibrahim_Rugova_0001.jpg
+Frances_Fisher/Frances_Fisher_0002.jpg
+Sachin_Tendulkar/Sachin_Tendulkar_0001.jpg
+Hashim_Thaci/Hashim_Thaci_0002.jpg
+Karen_Lynn_Gorney/Karen_Lynn_Gorney_0001.jpg
+Alimzhan_Tokhtakhounov/Alimzhan_Tokhtakhounov_0002.jpg
+Alex_Zanardi/Alex_Zanardi_0001.jpg
+Linus_Roache/Linus_Roache_0001.jpg
+Joe_Leonard/Joe_Leonard_0001.jpg
+Steve_Mariucci/Steve_Mariucci_0002.jpg
+Alejandro_Lerner/Alejandro_Lerner_0001.jpg
+John_Norquist/John_Norquist_0001.jpg
+Oliver_Phelps/Oliver_Phelps_0001.jpg
+Babe_Ruth/Babe_Ruth_0001.jpg
+Huang_Suey-Sheng/Huang_Suey-Sheng_0001.jpg
+Hitoshi_Tanaka/Hitoshi_Tanaka_0001.jpg
+Roy_Chaderton/Roy_Chaderton_0001.jpg
+Jessica_Lange/Jessica_Lange_0001.jpg
+Frank_Cassell/Frank_Cassell_0003.jpg
+Sasha_Cohen/Sasha_Cohen_0001.jpg
+Micky_Ward/Micky_Ward_0001.jpg
+Lara_Logan/Lara_Logan_0001.jpg
+Astrid_Eyzaguirre/Astrid_Eyzaguirre_0001.jpg
+Jeffery_Strelzin/Jeffery_Strelzin_0001.jpg
+Jaymon_Crabb/Jaymon_Crabb_0001.jpg
+Arturo_Gatti/Arturo_Gatti_0001.jpg
+Gabriel_Batistuta/Gabriel_Batistuta_0002.jpg
+Tatjana_Gsell/Tatjana_Gsell_0001.jpg
+Ali_Bin_Hussein/Ali_Bin_Hussein_0001.jpg
+Carlton_Dotson/Carlton_Dotson_0001.jpg
+Judy_Dean/Judy_Dean_0001.jpg
+Kenneth_Dam/Kenneth_Dam_0001.jpg
+Paul_Kelleher/Paul_Kelleher_0001.jpg
+Heather_Mills/Heather_Mills_0001.jpg
+Donald_Evans/Donald_Evans_0002.jpg
+Don_Nickles/Don_Nickles_0001.jpg
+Samuel_Waksal/Samuel_Waksal_0003.jpg
+Scott_OGrady/Scott_OGrady_0001.jpg
+George_Voinovich/George_Voinovich_0002.jpg
+Mitchell_Garabedian/Mitchell_Garabedian_0001.jpg
+David_Tornberg/David_Tornberg_0001.jpg
+James_Harris/James_Harris_0001.jpg
+Lindsey_Graham/Lindsey_Graham_0002.jpg
+Doug_Wilson/Doug_Wilson_0001.jpg
+Marisol_Martinez_Sambran/Marisol_Martinez_Sambran_0001.jpg
+Joschka_Fischer/Joschka_Fischer_0010.jpg
+Maria_Sanchez_Lorenzo/Maria_Sanchez_Lorenzo_0001.jpg
+Hana_Sadiq/Hana_Sadiq_0001.jpg
+Colin_Farrell/Colin_Farrell_0005.jpg
+Nastia_Liukin/Nastia_Liukin_0001.jpg
+Darrell_Porter/Darrell_Porter_0002.jpg
+Billy_Beane/Billy_Beane_0001.jpg
+Ronald_Brower/Ronald_Brower_0001.jpg
+Imelda_Marcos/Imelda_Marcos_0001.jpg
+Pete_Beaudrault/Pete_Beaudrault_0001.jpg
+Gloria_Allred/Gloria_Allred_0001.jpg
+Jack_Valenti/Jack_Valenti_0001.jpg
+Tiger_Woods/Tiger_Woods_0010.jpg
+Dai_Bachtiar/Dai_Bachtiar_0002.jpg
+Syed_Ibrahim/Syed_Ibrahim_0001.jpg
+Binyamin_Ben-Eliezer/Binyamin_Ben-Eliezer_0006.jpg
+Daniele_Hypolito/Daniele_Hypolito_0001.jpg
+Michael_Piuze/Michael_Piuze_0001.jpg
+Matt_Herden/Matt_Herden_0001.jpg
+Brent_Coles/Brent_Coles_0001.jpg
+Linda_Ham/Linda_Ham_0001.jpg
+Patrik_Kristiansson/Patrik_Kristiansson_0001.jpg
+Julio_Cesar_Chavez/Julio_Cesar_Chavez_0001.jpg
+John_Ruiz/John_Ruiz_0001.jpg
+Gerald_Barbarito/Gerald_Barbarito_0001.jpg
+Anatoliy_Kinakh/Anatoliy_Kinakh_0001.jpg
+Paul_Desmarais/Paul_Desmarais_0001.jpg
+Abdullah_Ahmad_Badawi/Abdullah_Ahmad_Badawi_0001.jpg
+Augusto_Pinochet/Augusto_Pinochet_0001.jpg
+Cate_Blanchett/Cate_Blanchett_0002.jpg
+Mick_Jagger/Mick_Jagger_0004.jpg
+Penny_Lancaster/Penny_Lancaster_0001.jpg
+Angela_Alvarado_Rosa/Angela_Alvarado_Rosa_0001.jpg
+Helen_Alvare/Helen_Alvare_0001.jpg
+Masja_Juel/Masja_Juel_0001.jpg
+Brian_Gregory/Brian_Gregory_0001.jpg
+Michael_Weiss/Michael_Weiss_0001.jpg
+Goldie_Hawn/Goldie_Hawn_0003.jpg
+Maria_Conchita_Alonso/Maria_Conchita_Alonso_0001.jpg
+Vladimir_Voltchkov/Vladimir_Voltchkov_0002.jpg
+Olesya_Bonabarenko/Olesya_Bonabarenko_0001.jpg
+Mariana_Gonzalez/Mariana_Gonzalez_0001.jpg
+Norman_Mineta/Norman_Mineta_0001.jpg
+Pascal_Affi_Nguessan/Pascal_Affi_Nguessan_0001.jpg
+Alyson_Hannigan/Alyson_Hannigan_0001.jpg
+Alexis_Bledel/Alexis_Bledel_0001.jpg
+Nick_Rimando/Nick_Rimando_0001.jpg
+Tim_Allen/Tim_Allen_0004.jpg
+Shannon_OBrien/Shannon_OBrien_0002.jpg
+Robert_Lee_Yates_Jr/Robert_Lee_Yates_Jr_0001.jpg
+Brian_McIntyre/Brian_McIntyre_0001.jpg
+Fann_Wong/Fann_Wong_0001.jpg
+Kifah_Ajouri/Kifah_Ajouri_0001.jpg
+Rick_Carlisle/Rick_Carlisle_0004.jpg
+Catherine_Bell/Catherine_Bell_0001.jpg
+Butch_Davis/Butch_Davis_0002.jpg
+Patricia_Garone/Patricia_Garone_0001.jpg
+Mike_Miller/Mike_Miller_0002.jpg
+Jennifer_Keller/Jennifer_Keller_0001.jpg
+Mia_Mottley/Mia_Mottley_0001.jpg
+Nicholas_Tse/Nicholas_Tse_0001.jpg
+Sid_Caesar/Sid_Caesar_0001.jpg
+Leonard_Schrank/Leonard_Schrank_0001.jpg
+Sheryl_Crow/Sheryl_Crow_0008.jpg
+Ellen_Barkin/Ellen_Barkin_0001.jpg
+Gwen_Stefani/Gwen_Stefani_0001.jpg
+S_Jayakumar/S_Jayakumar_0001.jpg
+Jane_Clayson/Jane_Clayson_0001.jpg
+Gustavo_Kuerten/Gustavo_Kuerten_0001.jpg
+Andrew_Fastow/Andrew_Fastow_0001.jpg
+Carla_Del_Ponte/Carla_Del_Ponte_0001.jpg
+Flavia_Delaroli/Flavia_Delaroli_0001.jpg
+Gerry_Adams/Gerry_Adams_0008.jpg
+Brett_Boone/Brett_Boone_0001.jpg
+Corinna_Harfouch/Corinna_Harfouch_0001.jpg
+George_Maxwell_Richards/George_Maxwell_Richards_0001.jpg
+Billy_Crystal/Billy_Crystal_0002.jpg
+Jeff_Roehm/Jeff_Roehm_0001.jpg
+Teri_ORourke/Teri_ORourke_0001.jpg
+Loretta_Lynn_Harper/Loretta_Lynn_Harper_0001.jpg
+Martin_Brooke/Martin_Brooke_0001.jpg
+Ray_Lucas/Ray_Lucas_0001.jpg
+Richard_Virenque/Richard_Virenque_0008.jpg
+David_Sousa/David_Sousa_0001.jpg
+Michael_Capellas/Michael_Capellas_0002.jpg
+Frank_Hilldrup/Frank_Hilldrup_0001.jpg
+Gary_Bettman/Gary_Bettman_0001.jpg
+Michael_Winterbottom/Michael_Winterbottom_0001.jpg
+Ralf_Schumacher/Ralf_Schumacher_0008.jpg
+Phillip_Seymor_Hoffmann/Phillip_Seymor_Hoffmann_0001.jpg
+Terrence_Kiel/Terrence_Kiel_0001.jpg
+Javier_Vazquez/Javier_Vazquez_0001.jpg
+Cari_Davis/Cari_Davis_0001.jpg
+John_McCallum/John_McCallum_0002.jpg
+Annette_Bening/Annette_Bening_0002.jpg
+Brian_Jordan/Brian_Jordan_0001.jpg
+Spike_Lee/Spike_Lee_0001.jpg
+Betsy_Smith/Betsy_Smith_0001.jpg
+John_Rowland/John_Rowland_0002.jpg
+David_Shayler/David_Shayler_0001.jpg
+Shane_Hmiel/Shane_Hmiel_0001.jpg
+Valeri_Bure/Valeri_Bure_0001.jpg
+Rick_Caruso/Rick_Caruso_0001.jpg
+Tina_Sinatra/Tina_Sinatra_0001.jpg
+Fatma_Kusibeh/Fatma_Kusibeh_0001.jpg
+Juan_Roman_Riquelme/Juan_Roman_Riquelme_0001.jpg
+Steven_Kinlock/Steven_Kinlock_0001.jpg
+Robert_Hyatt/Robert_Hyatt_0001.jpg
+John_Henry/John_Henry_0001.jpg
+Rodrigo_Rato/Rodrigo_Rato_0001.jpg
+Barry_Diller/Barry_Diller_0001.jpg
+Jose_Luis_Santiago_Vasconcelos/Jose_Luis_Santiago_Vasconcelos_0001.jpg
+Lawrence_MacAulay/Lawrence_MacAulay_0001.jpg
+James_Cunningham/James_Cunningham_0001.jpg
+Kellie_Coffey/Kellie_Coffey_0001.jpg
+Jeri_Ryan/Jeri_Ryan_0001.jpg
+Jim_Furyk/Jim_Furyk_0003.jpg
+William_Ragland/William_Ragland_0001.jpg
+McGuire_Gibson/McGuire_Gibson_0001.jpg
+Raja_Qureshi/Raja_Qureshi_0001.jpg
+Elaine_Stritch/Elaine_Stritch_0001.jpg
+Sun_Myung_Moon/Sun_Myung_Moon_0001.jpg
+Roger_Winter/Roger_Winter_0001.jpg
+Jo_Dee_Messina/Jo_Dee_Messina_0001.jpg
+Hanns_Schumacher/Hanns_Schumacher_0001.jpg
+Dionyssis_Georgiadis/Dionyssis_Georgiadis_0001.jpg
+Rafael_Ramirez/Rafael_Ramirez_0002.jpg
+Daniel_Coats/Daniel_Coats_0001.jpg
+Michael_Hoffa/Michael_Hoffa_0001.jpg
+Hal_Gehman/Hal_Gehman_0004.jpg
+Abdullah_Gul/Abdullah_Gul_0019.jpg
+Rick_Reed/Rick_Reed_0001.jpg
+Kirby_Puckett/Kirby_Puckett_0001.jpg
+Dean_Barkley/Dean_Barkley_0002.jpg
+Lawrence_Vito/Lawrence_Vito_0001.jpg
+Lesley_McCulloch/Lesley_McCulloch_0002.jpg
+Britney_Spears/Britney_Spears_0008.jpg
+Leticia_Van_de_Putte/Leticia_Van_de_Putte_0001.jpg
+Brenda_Magana/Brenda_Magana_0001.jpg
+Jack_Smith/Jack_Smith_0001.jpg
+Kevin_Spacey/Kevin_Spacey_0003.jpg
+Mohamed_Benaissa/Mohamed_Benaissa_0002.jpg
+Barbara_De_Brun/Barbara_De_Brun_0001.jpg
+Jose_Luis_Rodriguez_Zapatero/Jose_Luis_Rodriguez_Zapatero_0001.jpg
+Efrain_Rios_Montt/Efrain_Rios_Montt_0001.jpg
+Seymour_Cassell/Seymour_Cassell_0001.jpg
+Marcelo_Ebrard/Marcelo_Ebrard_0002.jpg
+Robbie_Mc_Ewen/Robbie_Mc_Ewen_0001.jpg
+Frank_Dunham_Jr/Frank_Dunham_Jr_0001.jpg
+Simon_Larose/Simon_Larose_0001.jpg
+Michael_J_Sheehan/Michael_J_Sheehan_0002.jpg
+Malcolm_Wild/Malcolm_Wild_0001.jpg
+Fred_Thompson/Fred_Thompson_0001.jpg
+Michael_Clarke_Duncan/Michael_Clarke_Duncan_0001.jpg
+Jerry_Seinfeld/Jerry_Seinfeld_0001.jpg
+Karen_Clarkson/Karen_Clarkson_0001.jpg
+Manuel_Poggiali/Manuel_Poggiali_0002.jpg
+Gregorio_Honasan/Gregorio_Honasan_0001.jpg
+Fernando_Henrique_Cardoso/Fernando_Henrique_Cardoso_0008.jpg
+Rich_Gannon/Rich_Gannon_0002.jpg
+Ricardo_Sanchez/Ricardo_Sanchez_0006.jpg
+Tom_Glavine/Tom_Glavine_0001.jpg
+Ray_Halbritter/Ray_Halbritter_0001.jpg
+Zach_Parise/Zach_Parise_0001.jpg
+Joe_Nichols/Joe_Nichols_0001.jpg
+Cristina_Saralegui/Cristina_Saralegui_0002.jpg
+Sean_Patrick_OMalley/Sean_Patrick_OMalley_0001.jpg
+Pedro_Alvarez/Pedro_Alvarez_0001.jpg
+Itamar_Franco/Itamar_Franco_0001.jpg
+Bill_Paxton/Bill_Paxton_0002.jpg
+Junichiro_Koizumi/Junichiro_Koizumi_0045.jpg
+Dorthy_Moxley/Dorthy_Moxley_0001.jpg
+Alejandro_Gonzalez_Inarritu/Alejandro_Gonzalez_Inarritu_0001.jpg
+Armando_Avila_Panchame/Armando_Avila_Panchame_0001.jpg
+Donny_Osmond/Donny_Osmond_0001.jpg
+Allyson_Felix/Allyson_Felix_0001.jpg
+Christopher_Conyers/Christopher_Conyers_0001.jpg
+Mickey_Loomis/Mickey_Loomis_0001.jpg
+Rick_Barnes/Rick_Barnes_0003.jpg
+Franz_Fischler/Franz_Fischler_0004.jpg
+Ian_McKellen/Ian_McKellen_0001.jpg
+Fran_Drescher/Fran_Drescher_0002.jpg
+Micah_Knorr/Micah_Knorr_0001.jpg
+Joe_Garner/Joe_Garner_0001.jpg
+Norodom_Chakrapong/Norodom_Chakrapong_0001.jpg
+Bob_Bowlsby/Bob_Bowlsby_0001.jpg
+Alan_Dershowitz/Alan_Dershowitz_0001.jpg
+Donald_Anderson/Donald_Anderson_0001.jpg
+Takenori_Kanzaki/Takenori_Kanzaki_0001.jpg
+Vanessa_Laine/Vanessa_Laine_0001.jpg
+Anthony_Pisciotti/Anthony_Pisciotti_0001.jpg
+Abner_Martinez/Abner_Martinez_0001.jpg
+Kenneth_Bowersox/Kenneth_Bowersox_0003.jpg
+Nicole/Nicole_0001.jpg
+Fred_Wilpon/Fred_Wilpon_0001.jpg
+Pablo_Khulental/Pablo_Khulental_0001.jpg
+Natasha_McElhone/Natasha_McElhone_0001.jpg
+Antony_Leung/Antony_Leung_0003.jpg
+Oliver_Stone/Oliver_Stone_0001.jpg
+Rene_Antonio_Leon_Rodriguez/Rene_Antonio_Leon_Rodriguez_0001.jpg
+Mark_Gangloff/Mark_Gangloff_0001.jpg
+Natalie_Maines/Natalie_Maines_0004.jpg
+Mark_Stuart/Mark_Stuart_0001.jpg
+Wendell_Bryant/Wendell_Bryant_0001.jpg
+Fruit_Chan/Fruit_Chan_0001.jpg
+James_Murdoch/James_Murdoch_0001.jpg
+Gavin_Degraw/Gavin_Degraw_0001.jpg
+Earl_Counter/Earl_Counter_0001.jpg
+Warren_Buffett/Warren_Buffett_0002.jpg
+Claudio_Ranieri/Claudio_Ranieri_0001.jpg
+Julius_Erving/Julius_Erving_0001.jpg
+Harland_Braun/Harland_Braun_0001.jpg
+Mark_Wahlberg/Mark_Wahlberg_0004.jpg
+Johnny_Hallyday/Johnny_Hallyday_0001.jpg
+Arye_Mekel/Arye_Mekel_0002.jpg
+Maribel_Dominguez/Maribel_Dominguez_0001.jpg
+Islam_Karimov/Islam_Karimov_0001.jpg
+Pedro_Mahecha/Pedro_Mahecha_0001.jpg
+Bill_Graham/Bill_Graham_0009.jpg
+Dule_Hill/Dule_Hill_0001.jpg
+Yashwant_Sinha/Yashwant_Sinha_0007.jpg
+Edward_Said/Edward_Said_0001.jpg
+Nong_Duc_Manh/Nong_Duc_Manh_0001.jpg
+Kyoko_Nakayama/Kyoko_Nakayama_0001.jpg
+Christian_Patino/Christian_Patino_0001.jpg
+Alanna_Ubach/Alanna_Ubach_0001.jpg
+Pete_Rose/Pete_Rose_0002.jpg
+Qazi_Afzal/Qazi_Afzal_0001.jpg
+Frank_Pallone/Frank_Pallone_0001.jpg
+Gloria_Macapagal_Arroyo/Gloria_Macapagal_Arroyo_0017.jpg
+Paul_Kagame/Paul_Kagame_0001.jpg
+Lou_Ye/Lou_Ye_0001.jpg
+Natalie_Imbruglia/Natalie_Imbruglia_0001.jpg
+Javier_Delgado/Javier_Delgado_0001.jpg
+Darin_Erstad/Darin_Erstad_0001.jpg
+George_Pataki/George_Pataki_0004.jpg
+King_Abdullah_II/King_Abdullah_II_0001.jpg
+John_Manley/John_Manley_0001.jpg
+Rich_Brooks/Rich_Brooks_0001.jpg
+Helio_Rubens_Garcia/Helio_Rubens_Garcia_0001.jpg
+Gabriel_Hughes/Gabriel_Hughes_0001.jpg
+Maria_Callas/Maria_Callas_0001.jpg
+Retief_Goosen/Retief_Goosen_0001.jpg
+Steven_Spielberg/Steven_Spielberg_0003.jpg
+Natasha_Lyonne/Natasha_Lyonne_0001.jpg
+Stephen_Crampton/Stephen_Crampton_0001.jpg
+Bianca_Jagger/Bianca_Jagger_0001.jpg
+Joe_Cocker/Joe_Cocker_0001.jpg
+Jason_Bentley/Jason_Bentley_0001.jpg
+Arnie_Boehm/Arnie_Boehm_0001.jpg
+Rodney_Rempt/Rodney_Rempt_0001.jpg
+David_Blaine/David_Blaine_0001.jpg
+John_Malkovich/John_Malkovich_0002.jpg
+Cheryl_James/Cheryl_James_0001.jpg
+Hashan_Tillakaratne/Hashan_Tillakaratne_0001.jpg
+Joe_Crede/Joe_Crede_0001.jpg
+Ann_Landers/Ann_Landers_0001.jpg
+Martin_Verkerk/Martin_Verkerk_0001.jpg
+Sheila_Taormina/Sheila_Taormina_0001.jpg
+Greg_Hennigar/Greg_Hennigar_0001.jpg
+Brad_Wilk/Brad_Wilk_0001.jpg
+AJ_Lamas/AJ_Lamas_0001.jpg
+Jeffrey_Ashby/Jeffrey_Ashby_0001.jpg
+Guennadi_Chipouline/Guennadi_Chipouline_0001.jpg
+Bill_Gates/Bill_Gates_0008.jpg
+Ricky_Martin/Ricky_Martin_0001.jpg
+Melissa_Joan_Hart/Melissa_Joan_Hart_0001.jpg
+Mike_Matthews/Mike_Matthews_0001.jpg
+Pier_Ferdinando_Casini/Pier_Ferdinando_Casini_0001.jpg
+John_Wright/John_Wright_0001.jpg
+Fred_Eckhard/Fred_Eckhard_0003.jpg
+Nathalia_Gillot/Nathalia_Gillot_0001.jpg
+Jeffrey_Scott_Postell/Jeffrey_Scott_Postell_0001.jpg
+Joao_Rocha/Joao_Rocha_0001.jpg
+Thad_Matta/Thad_Matta_0001.jpg
+Christopher_Whittle/Christopher_Whittle_0001.jpg
+Christian_Lacroix/Christian_Lacroix_0001.jpg
+Bing_Crosby/Bing_Crosby_0001.jpg
+Eddie_Fenech_Adami/Eddie_Fenech_Adami_0001.jpg
+Keith_Bishop_Jr/Keith_Bishop_Jr_0001.jpg
+Scott_Peterson/Scott_Peterson_0002.jpg
+Hugh_Miller/Hugh_Miller_0001.jpg
+Albert_Pujols/Albert_Pujols_0001.jpg
+William_Martin/William_Martin_0001.jpg
+Guus_Hiddink/Guus_Hiddink_0001.jpg
+Franco_Frattini/Franco_Frattini_0001.jpg
+Rohman_al-Ghozi/Rohman_al-Ghozi_0001.jpg
+Rita_Moreno/Rita_Moreno_0002.jpg
+Vladimiro_Montesinos/Vladimiro_Montesinos_0001.jpg
+Jerome_Jenkins/Jerome_Jenkins_0001.jpg
+Samantha_Daniels/Samantha_Daniels_0001.jpg
+Gregg_Popovich/Gregg_Popovich_0004.jpg
+Jim_OBrien/Jim_OBrien_0001.jpg
+Bill_Readdy/Bill_Readdy_0001.jpg
+Lucio_Gutierrez/Lucio_Gutierrez_0008.jpg
+Erwin_Mapasseng/Erwin_Mapasseng_0001.jpg
+Steve_Karsay/Steve_Karsay_0001.jpg
+Stacey_Dales-Schuman/Stacey_Dales-Schuman_0001.jpg
+Ernest_Hollings/Ernest_Hollings_0001.jpg
+Beatrice_Dalle/Beatrice_Dalle_0001.jpg
+Jenny_Romero/Jenny_Romero_0001.jpg
+Newton_Carlton_Slawson/Newton_Carlton_Slawson_0001.jpg
+Shaun_Rusling/Shaun_Rusling_0001.jpg
+Melana_Scantlin/Melana_Scantlin_0001.jpg
+Alejandro_Toledo/Alejandro_Toledo_0029.jpg
+DAngelo_Jimenez/DAngelo_Jimenez_0001.jpg
+Mario_Dominguez/Mario_Dominguez_0001.jpg
+Qais_al-Kazali/Qais_al-Kazali_0001.jpg
+Jon_Gruden/Jon_Gruden_0006.jpg
+Jeannette_Biedermann/Jeannette_Biedermann_0001.jpg
+Bernard_Landry/Bernard_Landry_0004.jpg
+Narayan_Singh_Pun/Narayan_Singh_Pun_0001.jpg
+Christian_Olsson/Christian_Olsson_0001.jpg
+Eddie_Murray/Eddie_Murray_0001.jpg
+Mark_Warner/Mark_Warner_0001.jpg
+TJ_Ford/TJ_Ford_0001.jpg
+Ekaterina_Dmitriev/Ekaterina_Dmitriev_0001.jpg
+William_Swor/William_Swor_0001.jpg
+Tono_Suratman/Tono_Suratman_0001.jpg
+Juan_Carlos_Ortega/Juan_Carlos_Ortega_0001.jpg
+Bono/Bono_0001.jpg
+Dan_Bylsma/Dan_Bylsma_0001.jpg
+Stephane_Delajoux/Stephane_Delajoux_0001.jpg
+Fernando_Sanz/Fernando_Sanz_0001.jpg
+Zahir_Shah/Zahir_Shah_0001.jpg
+Ian_Knop/Ian_Knop_0001.jpg
+Clint_Lamebear/Clint_Lamebear_0001.jpg
+Bertrand_Bonello/Bertrand_Bonello_0001.jpg
+Mark_Mishkin/Mark_Mishkin_0001.jpg
+Gene_Autry/Gene_Autry_0001.jpg
+Dyana_Calub/Dyana_Calub_0001.jpg
+Robert_Vowler/Robert_Vowler_0001.jpg
+Jolanta_Kwasniewski/Jolanta_Kwasniewski_0002.jpg
+Amy_Cotton/Amy_Cotton_0001.jpg
+Hisao_Oguchi/Hisao_Oguchi_0002.jpg
+Greg_Owen/Greg_Owen_0001.jpg
+Mohamed_Hammam/Mohamed_Hammam_0001.jpg
+Markus_Naslund/Markus_Naslund_0002.jpg
+Donald_Fehr/Donald_Fehr_0001.jpg
+Chris_Matthews/Chris_Matthews_0001.jpg
+Brian_De_Palma/Brian_De_Palma_0001.jpg
+Hugh_Grant/Hugh_Grant_0008.jpg
+Benjamin_McKenzie/Benjamin_McKenzie_0001.jpg
+Pyar_Jung_Thapa/Pyar_Jung_Thapa_0001.jpg
+Jalen_Rose/Jalen_Rose_0001.jpg
+Nancy_Sinatra/Nancy_Sinatra_0002.jpg
+Hugh_Jessiman/Hugh_Jessiman_0001.jpg
+Zhang_Yimou/Zhang_Yimou_0001.jpg
+Dave_Robertson/Dave_Robertson_0001.jpg
+Hans_Blix/Hans_Blix_0033.jpg
+Richard_Pennington/Richard_Pennington_0001.jpg
+Phil_Donahue/Phil_Donahue_0001.jpg
+James_Hill/James_Hill_0001.jpg
+Ahmed_Chalabi/Ahmed_Chalabi_0004.jpg
+Carlos_Juarez/Carlos_Juarez_0001.jpg
+Lester_Holt/Lester_Holt_0001.jpg
+Kate_Burton/Kate_Burton_0001.jpg
+Cass_Ballenger/Cass_Ballenger_0001.jpg
+Charlotte_Casiraghi/Charlotte_Casiraghi_0001.jpg
+Glen_Clark/Glen_Clark_0001.jpg
+Wolfgang_Schneiderhan/Wolfgang_Schneiderhan_0001.jpg
+Kaoru_Hasuike/Kaoru_Hasuike_0001.jpg
+Timothy_Rigas/Timothy_Rigas_0001.jpg
+Zydrunas_Ilgauskas/Zydrunas_Ilgauskas_0001.jpg
+Magui_Serna/Magui_Serna_0002.jpg
+Scott_Rolen/Scott_Rolen_0001.jpg
+Teddy_Kollek/Teddy_Kollek_0001.jpg
+Margie_Puente/Margie_Puente_0001.jpg
+John_Abizaid/John_Abizaid_0003.jpg
+Flor_Montulo/Flor_Montulo_0001.jpg
+Lin_Yi-fu/Lin_Yi-fu_0001.jpg
+Suzanne_Fox/Suzanne_Fox_0001.jpg
+Donald_Trump/Donald_Trump_0001.jpg
+Reggie_Lewis/Reggie_Lewis_0001.jpg
+Sophie/Sophie_0001.jpg
+Wes_Craven/Wes_Craven_0001.jpg
+Goran_Zivkovic/Goran_Zivkovic_0001.jpg
+Mamdouh_Habib/Mamdouh_Habib_0001.jpg
+Jane_Riley/Jane_Riley_0001.jpg
+Travis_Rudolph/Travis_Rudolph_0001.jpg
+John_Tyson/John_Tyson_0001.jpg
+Pierre_Png/Pierre_Png_0001.jpg
+Pilar_Montenegro/Pilar_Montenegro_0001.jpg
+Keith_Lowen/Keith_Lowen_0001.jpg
+Ramiro_Goben_Reducindo/Ramiro_Goben_Reducindo_0001.jpg
+Paul_Brandt/Paul_Brandt_0001.jpg
+Tim_Henman/Tim_Henman_0003.jpg
+Kirk_Douglas/Kirk_Douglas_0001.jpg
+Steven_Van_Zandt/Steven_Van_Zandt_0001.jpg
+Rupert_Grint/Rupert_Grint_0003.jpg
+Silvio_Fernandez/Silvio_Fernandez_0001.jpg
+Stacy_Nelson/Stacy_Nelson_0001.jpg
+Beth_Blough/Beth_Blough_0001.jpg
+Kurt_Tanabe/Kurt_Tanabe_0001.jpg
+David_Brent/David_Brent_0001.jpg
+Ariel_Sharon/Ariel_Sharon_0050.jpg
+Alexandre_Herchcovitch/Alexandre_Herchcovitch_0001.jpg
+John_Herrington/John_Herrington_0001.jpg
+Martha_Bowen/Martha_Bowen_0002.jpg
+Janette_Husarova/Janette_Husarova_0001.jpg
+Dale_Earnhardt_Jr/Dale_Earnhardt_Jr_0001.jpg
+Christian_Lirette/Christian_Lirette_0001.jpg
+Guillaume_Depardieu/Guillaume_Depardieu_0001.jpg
+Hector_Mitelman/Hector_Mitelman_0001.jpg
+Idi_Amin/Idi_Amin_0001.jpg
+Tomas_Malik/Tomas_Malik_0001.jpg
+Nursultan_Nazarbayev/Nursultan_Nazarbayev_0002.jpg
+Recep_Tayyip_Erdogan/Recep_Tayyip_Erdogan_0027.jpg
+Austin_Kearns/Austin_Kearns_0001.jpg
+Suzanne_Somers/Suzanne_Somers_0001.jpg
+Sue_Johnston/Sue_Johnston_0001.jpg
+Bill_Richardson/Bill_Richardson_0001.jpg
+Mario_Cipollini/Mario_Cipollini_0001.jpg
+Eduardo_Fischer/Eduardo_Fischer_0001.jpg
+Alfred_Sant/Alfred_Sant_0001.jpg
+Tim_Chapman/Tim_Chapman_0002.jpg
+Michael_Richards/Michael_Richards_0001.jpg
+Adrian_Murrell/Adrian_Murrell_0001.jpg
+Mikhail_Shvydkoi/Mikhail_Shvydkoi_0001.jpg
+Yasser_Arafat/Yasser_Arafat_0008.jpg
+Ajit_Agarkar/Ajit_Agarkar_0001.jpg
+John_Anderson/John_Anderson_0001.jpg
+Bernard_Siegel/Bernard_Siegel_0001.jpg
+Jesse_Ventura/Jesse_Ventura_0003.jpg
+Vadim_Devyatovskiy/Vadim_Devyatovskiy_0001.jpg
+Paul-Henri_Mathieu/Paul-Henri_Mathieu_0001.jpg
+Lee_Hyung-taik/Lee_Hyung-taik_0001.jpg
+Ray_Bradbury/Ray_Bradbury_0001.jpg
+Corinne_Coman/Corinne_Coman_0002.jpg
+Emyr_Jones_Parry/Emyr_Jones_Parry_0001.jpg
+Princess_Diana/Princess_Diana_0001.jpg
+Don_Carcieri/Don_Carcieri_0001.jpg
+Benjamin_Franklin/Benjamin_Franklin_0001.jpg
+Asa_Hutchinson/Asa_Hutchinson_0002.jpg
+Patrick_Bourrat/Patrick_Bourrat_0001.jpg
+Matt_Walters/Matt_Walters_0001.jpg
+Roy_Rogers/Roy_Rogers_0001.jpg
+Roman_Tam/Roman_Tam_0001.jpg
+Pat_Rochester/Pat_Rochester_0001.jpg
+Evelyn_Lauder/Evelyn_Lauder_0001.jpg
+Annie_Machon/Annie_Machon_0001.jpg
+Cole_Chapman/Cole_Chapman_0001.jpg
+Ruth_Harlow/Ruth_Harlow_0002.jpg
+Juan_Sanchez/Juan_Sanchez_0001.jpg
+Dave_Williams/Dave_Williams_0001.jpg
+Theresa_Gattung/Theresa_Gattung_0001.jpg
+Nicolas_Escude/Nicolas_Escude_0002.jpg
+Lloyd_Richards/Lloyd_Richards_0001.jpg
+Kellie_Greene/Kellie_Greene_0001.jpg
+Kathleen_Abernathy/Kathleen_Abernathy_0001.jpg
+Casey_Crowder/Casey_Crowder_0001.jpg
+Jerry_Sexton/Jerry_Sexton_0001.jpg
+Jorge_Marquez-Ruarte/Jorge_Marquez-Ruarte_0001.jpg
+Michael_Diekmann/Michael_Diekmann_0001.jpg
+William_Hochul/William_Hochul_0002.jpg
+Yoriko_Kawaguchi/Yoriko_Kawaguchi_0009.jpg
+Hilmi_Akin_Zorlu/Hilmi_Akin_Zorlu_0001.jpg
+George_Robertson/George_Robertson_0001.jpg
+Leslie_Moonves/Leslie_Moonves_0002.jpg
+Kevin_Crane/Kevin_Crane_0001.jpg
+Gary_Bergeron/Gary_Bergeron_0001.jpg
+Jeanette_Stauffer/Jeanette_Stauffer_0001.jpg
+Haydar_Aliyev/Haydar_Aliyev_0001.jpg
+Carina_Lau_Ka-ling/Carina_Lau_Ka-ling_0001.jpg
+Josh_Evans/Josh_Evans_0001.jpg
+Lance_Bass/Lance_Bass_0001.jpg
+Sahim_Alwan/Sahim_Alwan_0001.jpg
+Meg_Wakeman/Meg_Wakeman_0001.jpg
+Keizo_Yamada/Keizo_Yamada_0001.jpg
+Ian_Gillan/Ian_Gillan_0001.jpg
+Fayssal_Mekdad/Fayssal_Mekdad_0004.jpg
+Monique_Ferreira/Monique_Ferreira_0001.jpg
+Armando_Calderon_Sol/Armando_Calderon_Sol_0001.jpg
+David_Hannay/David_Hannay_0001.jpg
+Meryl_Streep/Meryl_Streep_0011.jpg
+Federico_Fellini/Federico_Fellini_0001.jpg
+Tina_Brown/Tina_Brown_0001.jpg
+Muffet_McGraw/Muffet_McGraw_0001.jpg
+Mike_Price/Mike_Price_0002.jpg
+Jonathan_Arden/Jonathan_Arden_0001.jpg
+Gloria_Trevi/Gloria_Trevi_0004.jpg
+Mel_Karmazin/Mel_Karmazin_0001.jpg
+Dimitri_Perricos/Dimitri_Perricos_0001.jpg
+Marc_Shaiman/Marc_Shaiman_0001.jpg
+Ronald_Kadish/Ronald_Kadish_0001.jpg
+Luther_Htu/Luther_Htu_0001.jpg
+Paola_Espinoza/Paola_Espinoza_0001.jpg
+Harold_Scott/Harold_Scott_0001.jpg
+Katie_Boone/Katie_Boone_0001.jpg
+Mark_Lazarus/Mark_Lazarus_0001.jpg
+Jake_Gyllenhaal/Jake_Gyllenhaal_0001.jpg
+Albert_Brooks/Albert_Brooks_0001.jpg
+Mohammed_Baqir_al-Hakim/Mohammed_Baqir_al-Hakim_0001.jpg
+Roman_Polanski/Roman_Polanski_0001.jpg
+Kara_Lynn_Joyce/Kara_Lynn_Joyce_0001.jpg
+Arnold_Palmer/Arnold_Palmer_0001.jpg
+Andrew_Gilligan/Andrew_Gilligan_0001.jpg
+LeBron_James/LeBron_James_0005.jpg
+Leonardo_Del_Vecchio/Leonardo_Del_Vecchio_0001.jpg
+Kai-Uwe_Ricke/Kai-Uwe_Ricke_0001.jpg
+Tonya_Payne/Tonya_Payne_0001.jpg
+Supachai_Panitchpakdi/Supachai_Panitchpakdi_0001.jpg
+James_Caan/James_Caan_0001.jpg
+Evie_Lazarou/Evie_Lazarou_0001.jpg
+Nora_Bendijo/Nora_Bendijo_0002.jpg
+Dalia_Rabin-Pelosoff/Dalia_Rabin-Pelosoff_0001.jpg
+Anthony_Hazen/Anthony_Hazen_0001.jpg
+Andrew_Bernard/Andrew_Bernard_0001.jpg
+Dominik_Hrbaty/Dominik_Hrbaty_0001.jpg
+Adrianna_Zuzic/Adrianna_Zuzic_0001.jpg
+Vicente_Fox/Vicente_Fox_0005.jpg
+Surya_Bahadur_Thapa/Surya_Bahadur_Thapa_0001.jpg
+Peter_Lundgren/Peter_Lundgren_0001.jpg
+Michael_Adams/Michael_Adams_0001.jpg
+Cruz_Bustamante/Cruz_Bustamante_0001.jpg
+Rod_Blagojevich/Rod_Blagojevich_0001.jpg
+Claire_Danes/Claire_Danes_0001.jpg
+Gordon_Brown/Gordon_Brown_0008.jpg
+Juan_Carlos_Morales/Juan_Carlos_Morales_0001.jpg
+Arthur_Martinez/Arthur_Martinez_0001.jpg
+Justin_Wilson/Justin_Wilson_0001.jpg
+Paulina_Rodriguez_Davila/Paulina_Rodriguez_Davila_0001.jpg
+Larry_Nichols/Larry_Nichols_0001.jpg
+Christina_Sawaya/Christina_Sawaya_0001.jpg
+Michael_Fitzgerald/Michael_Fitzgerald_0001.jpg
+Linn_Thornton/Linn_Thornton_0001.jpg
+Harvey_Wachsman/Harvey_Wachsman_0001.jpg
+Silvie_Cabero/Silvie_Cabero_0001.jpg
+George_W_Bush/George_W_Bush_0387.jpg
+Daniel_Ortega/Daniel_Ortega_0001.jpg
+Kristen_Rivera/Kristen_Rivera_0001.jpg
+Angela_Bassett/Angela_Bassett_0001.jpg
+Colleen_Jones/Colleen_Jones_0001.jpg
+Hans-Christian_Schmid/Hans-Christian_Schmid_0001.jpg
+Nadia_Forte/Nadia_Forte_0001.jpg
+Heizo_Takenaka/Heizo_Takenaka_0001.jpg
+Morgan_Fairchild/Morgan_Fairchild_0001.jpg
+Robbie_Naish/Robbie_Naish_0001.jpg
+Elin_Nordegren/Elin_Nordegren_0001.jpg
+Stephen_Frears/Stephen_Frears_0001.jpg
+Richard_Levin/Richard_Levin_0001.jpg
+Zoe_Ball/Zoe_Ball_0001.jpg
+Brett_Hawke/Brett_Hawke_0001.jpg
+Mary_Katherine_Smart/Mary_Katherine_Smart_0001.jpg
+Roger_Corbett/Roger_Corbett_0001.jpg
+Joe_Mantegna/Joe_Mantegna_0001.jpg
+Robert_Korzeniowski/Robert_Korzeniowski_0001.jpg
+Irfan_Ahmed/Irfan_Ahmed_0001.jpg
+Connie_Chung/Connie_Chung_0001.jpg
+Tommy_Shane_Steiner/Tommy_Shane_Steiner_0001.jpg
+Jacqueline_Obradors/Jacqueline_Obradors_0002.jpg
+Manuel_Rosendo/Manuel_Rosendo_0001.jpg
+Mario_Jardel/Mario_Jardel_0001.jpg
+Richard_Ward/Richard_Ward_0001.jpg
+Dick_Vermeil/Dick_Vermeil_0001.jpg
+Mark_Mulder/Mark_Mulder_0001.jpg
+Reinhard_Buetikofer/Reinhard_Buetikofer_0001.jpg
+Enrique_Medina_Gomez/Enrique_Medina_Gomez_0001.jpg
+Carla_Moreno/Carla_Moreno_0001.jpg
+Juergen_Trittin/Juergen_Trittin_0001.jpg
+Denise_van_Outen/Denise_van_Outen_0001.jpg
+Brajesh_Mishra/Brajesh_Mishra_0001.jpg
+Mark_Gottfried/Mark_Gottfried_0002.jpg
+Terrence_Trammell/Terrence_Trammell_0001.jpg
+Lee_Jun/Lee_Jun_0002.jpg
+Katrin_Susi/Katrin_Susi_0001.jpg
+Kaye_Young/Kaye_Young_0001.jpg
+Jose_Acasuso/Jose_Acasuso_0001.jpg
+Patrick_Clawsen/Patrick_Clawsen_0001.jpg
+Ted_Costa/Ted_Costa_0001.jpg
+Jimmy_Gurule/Jimmy_Gurule_0001.jpg
+Ramona_Rispton/Ramona_Rispton_0001.jpg
+Raul_Rivero/Raul_Rivero_0001.jpg
+Hee-Won_Han/Hee-Won_Han_0002.jpg
+Halbert_Fillinger/Halbert_Fillinger_0001.jpg
+Marc_Leger/Marc_Leger_0001.jpg
+Rafael_Vinoly/Rafael_Vinoly_0001.jpg
+Guillermo_Monroy/Guillermo_Monroy_0001.jpg
+Mohammad_Hasanein/Mohammad_Hasanein_0001.jpg
+Rafeeuddin_Ahmed/Rafeeuddin_Ahmed_0001.jpg
+Charlie_Garner/Charlie_Garner_0001.jpg
+Olivia_Newton-John/Olivia_Newton-John_0002.jpg
+Martin_Boryczewski/Martin_Boryczewski_0001.jpg
+Nicola_Wells/Nicola_Wells_0001.jpg
+Connie_Freydell/Connie_Freydell_0001.jpg
+John_Williams/John_Williams_0001.jpg
+Patrick_Kron/Patrick_Kron_0001.jpg
+Ralph_Fiennes/Ralph_Fiennes_0001.jpg
+Alejandro_Atchugarry/Alejandro_Atchugarry_0002.jpg
+Gerrit_Zalm/Gerrit_Zalm_0001.jpg
+Silvio_Berlusconi/Silvio_Berlusconi_0017.jpg
+Robert_Zoellick/Robert_Zoellick_0004.jpg
+Eddie_Jordan/Eddie_Jordan_0001.jpg
+Norah_Jones/Norah_Jones_0003.jpg
+Stephen_Daldry/Stephen_Daldry_0001.jpg
+Pedro_Martinez/Pedro_Martinez_0001.jpg
+Ahmed_Qureia/Ahmed_Qureia_0001.jpg
+Paul_Clark/Paul_Clark_0001.jpg
+Eva_Mendes/Eva_Mendes_0001.jpg
+Judy_Genshaft/Judy_Genshaft_0001.jpg
+Tom_DeLay/Tom_DeLay_0001.jpg
+Van_Hilley/Van_Hilley_0001.jpg
+Olene_Walker/Olene_Walker_0001.jpg
+Thomas_Malchow/Thomas_Malchow_0002.jpg
+Lorraine_Bracco/Lorraine_Bracco_0001.jpg
+Michael_Bloomberg/Michael_Bloomberg_0005.jpg
+Julien_Boutter/Julien_Boutter_0001.jpg
+Carlos_Alberto/Carlos_Alberto_0001.jpg
+Jason_Sorens/Jason_Sorens_0001.jpg
+Maxim_Afinogenov/Maxim_Afinogenov_0001.jpg
+Linda_Mason/Linda_Mason_0001.jpg
+Gary_Forsee/Gary_Forsee_0002.jpg
+Bob_Guccione/Bob_Guccione_0001.jpg
+Alvaro_Noboa/Alvaro_Noboa_0001.jpg
+Sergei_Alexandrovitch_Ordzhonikidze/Sergei_Alexandrovitch_Ordzhonikidze_0001.jpg
+Naomi_Campbell/Naomi_Campbell_0002.jpg
+Tom_Schnackenberg/Tom_Schnackenberg_0001.jpg
+Tocker_Pudwill/Tocker_Pudwill_0001.jpg
+Amy_Smart/Amy_Smart_0001.jpg
+Peter_Arnett/Peter_Arnett_0002.jpg
+Abdul_Rahman/Abdul_Rahman_0001.jpg
+Filip_De_Winter/Filip_De_Winter_0001.jpg
+Katie_Smith/Katie_Smith_0001.jpg
+Jerry_Hall/Jerry_Hall_0001.jpg
+Tim_Howard/Tim_Howard_0001.jpg
+Daniel_Montenegro/Daniel_Montenegro_0001.jpg
+Malcolm_Glazer/Malcolm_Glazer_0001.jpg
+Melissa_Manchester/Melissa_Manchester_0001.jpg
+Estelle_Morris/Estelle_Morris_0001.jpg
+Solomon_Passy/Solomon_Passy_0001.jpg
+Faye_Alibocus/Faye_Alibocus_0001.jpg
+Bill_Lerach/Bill_Lerach_0001.jpg
+Hamid_Efendi/Hamid_Efendi_0001.jpg
+Lon_Kruger/Lon_Kruger_0002.jpg
+Bernard_Lord/Bernard_Lord_0001.jpg
+Fernando_Valenzuela/Fernando_Valenzuela_0001.jpg
+John_Brady/John_Brady_0001.jpg
+Hassanal_Bolkiah/Hassanal_Bolkiah_0001.jpg
+Nelson_Acosta/Nelson_Acosta_0001.jpg
+Mo_Elleithee/Mo_Elleithee_0001.jpg
+Klaus_Zwickel/Klaus_Zwickel_0002.jpg
+Kristanna_Loken/Kristanna_Loken_0004.jpg
+Johnny_Htu/Johnny_Htu_0001.jpg
+Francisco_Flores/Francisco_Flores_0004.jpg
+Adrien_Brody/Adrien_Brody_0012.jpg
+Hosni_Mubarak/Hosni_Mubarak_0007.jpg
+Ward_Cuff/Ward_Cuff_0001.jpg
+Jason_Priestley/Jason_Priestley_0001.jpg
+Adrian_Annus/Adrian_Annus_0001.jpg
+Sean_Penn/Sean_Penn_0002.jpg
+Mark_Heller/Mark_Heller_0002.jpg
+Donnie_Brennan/Donnie_Brennan_0001.jpg
+Michael_Michele/Michael_Michele_0001.jpg
+Martin_Brodeur/Martin_Brodeur_0001.jpg
+Kirk_Franklin/Kirk_Franklin_0001.jpg
+Atal_Bihari_Vajpayee/Atal_Bihari_Vajpayee_0019.jpg
+Paul_Li_Calsi/Paul_Li_Calsi_0001.jpg
+William_Nessen/William_Nessen_0001.jpg
+Bob_Goldman/Bob_Goldman_0001.jpg
+Scott_Dickson/Scott_Dickson_0001.jpg
+Grady_Little/Grady_Little_0001.jpg
+John_Philip_Elkann/John_Philip_Elkann_0001.jpg
+Shane_Mosley/Shane_Mosley_0001.jpg
+Cathryn_Crawford/Cathryn_Crawford_0001.jpg
+Bill_Pryor/Bill_Pryor_0001.jpg
+Camilla_Parker_Bowles/Camilla_Parker_Bowles_0001.jpg
+Zhu_Rongji/Zhu_Rongji_0008.jpg
+Margaret_Hasley/Margaret_Hasley_0001.jpg
+Lee_Byung-woong/Lee_Byung-woong_0001.jpg
+Serge_Klarsfeld/Serge_Klarsfeld_0001.jpg
+Nikki_Cascone/Nikki_Cascone_0001.jpg
+Alberta_Lee/Alberta_Lee_0001.jpg
+Nick_Price/Nick_Price_0001.jpg
+Shane_Reynolds/Shane_Reynolds_0001.jpg
+Russ_Ortiz/Russ_Ortiz_0001.jpg
+Sonya_Walger/Sonya_Walger_0001.jpg
+Johnson_Panjaitan/Johnson_Panjaitan_0002.jpg
+Joan_Claybrook/Joan_Claybrook_0002.jpg
+Andrea_Bocelli/Andrea_Bocelli_0001.jpg
+Michael_McNeely/Michael_McNeely_0001.jpg
+Andy_Lau/Andy_Lau_0001.jpg
+Jimmy_Kimmel/Jimmy_Kimmel_0001.jpg
+Albrecht_Mentz/Albrecht_Mentz_0002.jpg
+Hank_Azaria/Hank_Azaria_0001.jpg
+Doris_Schroeder/Doris_Schroeder_0003.jpg
+Sedigh_Barmak/Sedigh_Barmak_0001.jpg
+Frank_Marshall/Frank_Marshall_0001.jpg
+Sybille_Schmid/Sybille_Schmid_0001.jpg
+Jerry_Sloan/Jerry_Sloan_0001.jpg
+Dot_Helms/Dot_Helms_0001.jpg
+Imad_Khadduri/Imad_Khadduri_0001.jpg
+Hilary_McKay/Hilary_McKay_0001.jpg
+Matthew_Vaughan/Matthew_Vaughan_0001.jpg
+Jim_Parque/Jim_Parque_0001.jpg
+Wycliffe_Grousbeck/Wycliffe_Grousbeck_0001.jpg
+Buford_Blount/Buford_Blount_0001.jpg
+Mike_Brey/Mike_Brey_0001.jpg
+Bill_Clancy/Bill_Clancy_0001.jpg
+Michael_Jordan/Michael_Jordan_0001.jpg
+Bobby_Kielty/Bobby_Kielty_0001.jpg
+Patti_Lank/Patti_Lank_0001.jpg
+David_Arquette/David_Arquette_0001.jpg
+Tracee_Ellis_Ross/Tracee_Ellis_Ross_0001.jpg
+Alain_Ducasse/Alain_Ducasse_0001.jpg
+Mack_Brown/Mack_Brown_0002.jpg
+Aly_Wagner/Aly_Wagner_0001.jpg
+Phil_Bredesen/Phil_Bredesen_0001.jpg
+Natalie_Coughlin/Natalie_Coughlin_0001.jpg
+Marcus_Garrettson/Marcus_Garrettson_0001.jpg
+Darryl_McDaniels/Darryl_McDaniels_0001.jpg
+John_Jumper/John_Jumper_0002.jpg
+Bobby_Robson/Bobby_Robson_0001.jpg
+Patty_Duke/Patty_Duke_0001.jpg
+Mike_Eskew/Mike_Eskew_0001.jpg
+James_Smith/James_Smith_0002.jpg
+David_Bisbal/David_Bisbal_0001.jpg
+Guy_Hemmings/Guy_Hemmings_0001.jpg
+Akbar_Al_Baker/Akbar_Al_Baker_0001.jpg
+Leslie_Wiser_Jr/Leslie_Wiser_Jr_0001.jpg
+Hank_Bass/Hank_Bass_0001.jpg
+Bruce_Paltrow/Bruce_Paltrow_0001.jpg
+Amanda_Coetzer/Amanda_Coetzer_0002.jpg
+Andrew_Niccol/Andrew_Niccol_0001.jpg
+Mireya_Elisa_Moscoso_Rodriguez/Mireya_Elisa_Moscoso_Rodriguez_0001.jpg
+Kent_Robinson/Kent_Robinson_0001.jpg
+Lisa_Stansfield/Lisa_Stansfield_0001.jpg
+Annika_Sorenstam/Annika_Sorenstam_0001.jpg
+Hans_Peter_Briegel/Hans_Peter_Briegel_0001.jpg
+Owen_Nolan/Owen_Nolan_0001.jpg
+Zurab_Tsereteli/Zurab_Tsereteli_0001.jpg
+Nicanor_Duarte_Frutos/Nicanor_Duarte_Frutos_0006.jpg
+Jayson_Williams/Jayson_Williams_0001.jpg
+Esad_Landzo/Esad_Landzo_0001.jpg
+John_Stockton/John_Stockton_0001.jpg
+Cha_Yung-gu/Cha_Yung-gu_0001.jpg
+Thomas_Weston/Thomas_Weston_0001.jpg
+Gustavo_Franco/Gustavo_Franco_0001.jpg
+Tammy_Helm/Tammy_Helm_0001.jpg
+Serena_Williams/Serena_Williams_0012.jpg
+Augusto_Roa_Bastos/Augusto_Roa_Bastos_0002.jpg
+Alexandra_Vodjanikova/Alexandra_Vodjanikova_0002.jpg
+Brad_Banks/Brad_Banks_0001.jpg
+Marsha_Thomason/Marsha_Thomason_0001.jpg
+Charlie_Sheen/Charlie_Sheen_0001.jpg
+Mehmet_Okur/Mehmet_Okur_0001.jpg
+Gilberto_Rodriguez_Orejuela/Gilberto_Rodriguez_Orejuela_0001.jpg
+Johnny_Depp/Johnny_Depp_0001.jpg
+Evan_Marriott/Evan_Marriott_0001.jpg
+Mariana_Pollack/Mariana_Pollack_0001.jpg
+Hillary_Clinton/Hillary_Clinton_0003.jpg
+Ronald_Reagan/Ronald_Reagan_0002.jpg
+Jerry_Falwell/Jerry_Falwell_0002.jpg
+Cherry_Jones/Cherry_Jones_0001.jpg
+Suzanne_Haik_Terrell/Suzanne_Haik_Terrell_0001.jpg
+Salman_Khan/Salman_Khan_0001.jpg
+Tora_Takagi/Tora_Takagi_0001.jpg
+Shoshannah_Stern/Shoshannah_Stern_0001.jpg
+Romano_Prodi/Romano_Prodi_0001.jpg
+Sally_Kirkland/Sally_Kirkland_0004.jpg
+Princess_Caroline/Princess_Caroline_0001.jpg
+Al_Pacino/Al_Pacino_0001.jpg
+Lima_Azimi/Lima_Azimi_0001.jpg
+Jean-Patrick_Nazon/Jean-Patrick_Nazon_0001.jpg
+Mark_Foley/Mark_Foley_0001.jpg
+T_Boone_Pickens/T_Boone_Pickens_0001.jpg
+Pedro_Almodovar/Pedro_Almodovar_0006.jpg
+Jan_Ullrich/Jan_Ullrich_0004.jpg
+Alicia_Molik/Alicia_Molik_0001.jpg
+Kenneth_Brill/Kenneth_Brill_0001.jpg
+Joseph_Ralston/Joseph_Ralston_0001.jpg
+Daniele_Bergamin/Daniele_Bergamin_0001.jpg
+Peter_Costello/Peter_Costello_0001.jpg
+Joe_Vandever/Joe_Vandever_0001.jpg
+Leticia_Dolera/Leticia_Dolera_0001.jpg
+Daniel_Osorno/Daniel_Osorno_0001.jpg
+Shireen_Amir_Begum/Shireen_Amir_Begum_0001.jpg
+Paige_Fitzgerald/Paige_Fitzgerald_0001.jpg
+Charles_Tannok/Charles_Tannok_0001.jpg
+Hama_Arba_Diallo/Hama_Arba_Diallo_0001.jpg
+Gerard_de_Cortanze/Gerard_de_Cortanze_0001.jpg
+Abdel_Madi_Shabneh/Abdel_Madi_Shabneh_0001.jpg
+Tsutomu_Takebe/Tsutomu_Takebe_0001.jpg
+Bridgette_Wilson-Sampras/Bridgette_Wilson-Sampras_0001.jpg
+Andy_Wisecarver/Andy_Wisecarver_0001.jpg
+Javier_Saviola/Javier_Saviola_0001.jpg
+Alfonso_Soriano/Alfonso_Soriano_0001.jpg
+Yasar_Yakis/Yasar_Yakis_0004.jpg
+Barry_Williams/Barry_Williams_0001.jpg
+Boris_Yeltsin/Boris_Yeltsin_0002.jpg
+Helmut_Panke/Helmut_Panke_0001.jpg
+Nasser_al-Kidwa/Nasser_al-Kidwa_0001.jpg
+George_Ryan/George_Ryan_0004.jpg
+Paula_Zahn/Paula_Zahn_0001.jpg
+Ashley_Olsen/Ashley_Olsen_0001.jpg
+Minnie_Driver/Minnie_Driver_0001.jpg
+Darrell_Royal/Darrell_Royal_0001.jpg
+Ratna_Sari_Dewi_Sukarno/Ratna_Sari_Dewi_Sukarno_0001.jpg
+Samantha_Ledster/Samantha_Ledster_0001.jpg
+Benito_Santiago/Benito_Santiago_0001.jpg
+Jose_Genoino/Jose_Genoino_0001.jpg
+Marcelo_Salas/Marcelo_Salas_0001.jpg
+Nicolas_Lapentti/Nicolas_Lapentti_0002.jpg
+Astou_Ndiaye-Diatta/Astou_Ndiaye-Diatta_0001.jpg
+Paul_Tagliabue/Paul_Tagliabue_0001.jpg
+David_Hyde_Pierce/David_Hyde_Pierce_0002.jpg
+Molly_Sims/Molly_Sims_0001.jpg
+Neil_Moritz/Neil_Moritz_0001.jpg
+Megawati_Sukarnoputri/Megawati_Sukarnoputri_0006.jpg
+Hal_Sellers/Hal_Sellers_0001.jpg
+Mario_Vasquez_Rana/Mario_Vasquez_Rana_0001.jpg
+James_Layug/James_Layug_0001.jpg
+Rachel_Corrie/Rachel_Corrie_0001.jpg
+Ibrahim_Hilal/Ibrahim_Hilal_0001.jpg
+Carlos_Savedra/Carlos_Savedra_0001.jpg
+Howard_Stern/Howard_Stern_0001.jpg
+Robin_Wright_Penn/Robin_Wright_Penn_0001.jpg
+John_Walsh/John_Walsh_0001.jpg
+Roy_Jones_Jr/Roy_Jones_Jr_0002.jpg
+Miguel_Jimenez/Miguel_Jimenez_0001.jpg
+Arnoldo_Aleman/Arnoldo_Aleman_0001.jpg
+Noer_Moeis/Noer_Moeis_0001.jpg
+Ben_Kingsley/Ben_Kingsley_0001.jpg
+James_Kopp/James_Kopp_0001.jpg
+Dennis_Johnson/Dennis_Johnson_0001.jpg
+Paul_Crake/Paul_Crake_0001.jpg
+Daniela_Cicarelli/Daniela_Cicarelli_0001.jpg
+Michele_Placido/Michele_Placido_0001.jpg
+Roger_Penske/Roger_Penske_0001.jpg
+Tab_Turner/Tab_Turner_0001.jpg
+Ernesto_Zedillo/Ernesto_Zedillo_0001.jpg
+David_Duval/David_Duval_0001.jpg
+Laura_Hernandez/Laura_Hernandez_0001.jpg
+James_Schultz/James_Schultz_0001.jpg
+Franklin_Damann/Franklin_Damann_0001.jpg
+Nathan_Powell/Nathan_Powell_0001.jpg
+Charles_Pickering/Charles_Pickering_0001.jpg
+Juan_Valencia_Osorio/Juan_Valencia_Osorio_0002.jpg
+Michael_Dell/Michael_Dell_0001.jpg
+Richard_Sambrook/Richard_Sambrook_0002.jpg
+Claire_De_Gryse/Claire_De_Gryse_0001.jpg
+Daniel_Kurtzer/Daniel_Kurtzer_0001.jpg
+Bruno_Junquiera/Bruno_Junquiera_0001.jpg
+Junichi_Inamoto/Junichi_Inamoto_0001.jpg
+Khatol_Mohammad_Zai/Khatol_Mohammad_Zai_0001.jpg
+Jim_Sterk/Jim_Sterk_0001.jpg
+Masum_Turker/Masum_Turker_0003.jpg
+Keith_Bogans/Keith_Bogans_0001.jpg
+Heinrich_Wolfgang/Heinrich_Wolfgang_0001.jpg
+John_Kerry/John_Kerry_0008.jpg
+Marcos_Daniel_Jimenez/Marcos_Daniel_Jimenez_0001.jpg
+Clare_Short/Clare_Short_0004.jpg
+Steve_Shiver/Steve_Shiver_0001.jpg
+Bruce_Van_De_Velde/Bruce_Van_De_Velde_0002.jpg
+Barry_Switzer/Barry_Switzer_0001.jpg
+Gina_Centrello/Gina_Centrello_0001.jpg
+Gerard_Tronche/Gerard_Tronche_0001.jpg
+Ernie_Fletcher/Ernie_Fletcher_0001.jpg
+Michael_Douglas/Michael_Douglas_0006.jpg
+Stephen_Arigbabu/Stephen_Arigbabu_0001.jpg
+Alexander_Losyukov/Alexander_Losyukov_0004.jpg
+Celia_Cruz/Celia_Cruz_0001.jpg
+Daniel_Zelman/Daniel_Zelman_0001.jpg
+Joe_DeLamielleure/Joe_DeLamielleure_0001.jpg
+Jeremy_Shockey/Jeremy_Shockey_0001.jpg
+Kirk_Doerger/Kirk_Doerger_0001.jpg
+Michael_Lechner/Michael_Lechner_0001.jpg
+Matthew_Ouimet/Matthew_Ouimet_0001.jpg
+Edmund_Hillary/Edmund_Hillary_0001.jpg
+Robbie_Williams/Robbie_Williams_0001.jpg
+Noer_Muis/Noer_Muis_0001.jpg
+Michel_Kratochvil/Michel_Kratochvil_0001.jpg
+Rachel_Hunter/Rachel_Hunter_0004.jpg
+Teri_Garr/Teri_Garr_0001.jpg
+Taku_Yamasaki/Taku_Yamasaki_0001.jpg
+Chris_Thomas/Chris_Thomas_0001.jpg
+Mike_Easley/Mike_Easley_0001.jpg
+Bob_Iger/Bob_Iger_0001.jpg
+Steve_Peace/Steve_Peace_0001.jpg
+Mickey_Sherman/Mickey_Sherman_0001.jpg
+Princess_Stephanie/Princess_Stephanie_0001.jpg
+Daniel_Chin/Daniel_Chin_0001.jpg
+Steve_Rush/Steve_Rush_0001.jpg
+Janet_Horvath/Janet_Horvath_0001.jpg
+Joan_Dangerfield/Joan_Dangerfield_0001.jpg
+Silvan_Shalom/Silvan_Shalom_0003.jpg
+Gregorio_Rosal/Gregorio_Rosal_0001.jpg
+Lucio_Cecchinello/Lucio_Cecchinello_0001.jpg
+Theo_Angelopoulos/Theo_Angelopoulos_0001.jpg
+Atsushi_Sato/Atsushi_Sato_0001.jpg
+Mike_Gable/Mike_Gable_0001.jpg
+Jerry_Regier/Jerry_Regier_0004.jpg
+Ivan_Helguera/Ivan_Helguera_0001.jpg
+Turner_Stevenson/Turner_Stevenson_0001.jpg
+Normand_Legault/Normand_Legault_0001.jpg
+Evan_Rachel_Wood/Evan_Rachel_Wood_0002.jpg
+Roy_Romanow/Roy_Romanow_0001.jpg
+Bob_Beauprez/Bob_Beauprez_0001.jpg
+Gregory_Hines/Gregory_Hines_0002.jpg
+Lily_Tomlin/Lily_Tomlin_0001.jpg
+Seth_Gorney/Seth_Gorney_0001.jpg
+Federico_Trillo/Federico_Trillo_0003.jpg
+Maureen_Fanning/Maureen_Fanning_0001.jpg
+Jawad_Boulus/Jawad_Boulus_0001.jpg
+Melchor_Cob_Castro/Melchor_Cob_Castro_0001.jpg
+Suzie_McConnell_Serio/Suzie_McConnell_Serio_0001.jpg
+Cheryl_Tiegs/Cheryl_Tiegs_0001.jpg
+Lee_Yuan-tseh/Lee_Yuan-tseh_0001.jpg
+Vyacheslav_Fetisov/Vyacheslav_Fetisov_0001.jpg
+Chuck_Eidson/Chuck_Eidson_0001.jpg
+Ray_Wasden/Ray_Wasden_0001.jpg
+Mark_Podlesny/Mark_Podlesny_0001.jpg
+Luc_Montagnier/Luc_Montagnier_0001.jpg
+Ali_Khamenei/Ali_Khamenei_0001.jpg
+Basdeo_Panday/Basdeo_Panday_0001.jpg
+Andy_Benes/Andy_Benes_0001.jpg
+Galen_Rowell/Galen_Rowell_0001.jpg
+Gregg_Berhalter/Gregg_Berhalter_0001.jpg
+Stefanie_De_Roux/Stefanie_De_Roux_0001.jpg
+Paul_Farley/Paul_Farley_0001.jpg
+Kevin_Gil/Kevin_Gil_0001.jpg
+Kevin_Borseth/Kevin_Borseth_0001.jpg
+Richard_Chamberlain/Richard_Chamberlain_0001.jpg
+Benedita_da_Silva/Benedita_da_Silva_0001.jpg
+Tom_Moss/Tom_Moss_0001.jpg
+Wim_Duisenberg/Wim_Duisenberg_0001.jpg
+Chang_Dae-whan/Chang_Dae-whan_0002.jpg
+Michael_J_Fox/Michael_J_Fox_0001.jpg
+Oleg_Romantsev/Oleg_Romantsev_0001.jpg
+Terunobu_Maeda/Terunobu_Maeda_0001.jpg
+Ron_Zook/Ron_Zook_0001.jpg
+Bruce_Willis/Bruce_Willis_0001.jpg
+Jason_Kapono/Jason_Kapono_0001.jpg
+Bertie_Ahern/Bertie_Ahern_0004.jpg
+Almeida_Baptista/Almeida_Baptista_0001.jpg
+Bryan_Adams/Bryan_Adams_0001.jpg
+Alexandra_Jackson/Alexandra_Jackson_0001.jpg
+Camille_Lewis/Camille_Lewis_0001.jpg
+Nancy_Greenlaw/Nancy_Greenlaw_0001.jpg
+John_McCormack/John_McCormack_0001.jpg
+Michael_Phelps/Michael_Phelps_0004.jpg
+Andrew_Cuomo/Andrew_Cuomo_0001.jpg
+John_Daly_Jr/John_Daly_Jr_0001.jpg
+Rogerio_Romero/Rogerio_Romero_0002.jpg
+Arnold_Scott/Arnold_Scott_0001.jpg
+John_Connolly/John_Connolly_0001.jpg
+Billy_Rork/Billy_Rork_0001.jpg
+Oscar_de_la_Renta/Oscar_de_la_Renta_0001.jpg
+Anders_Fogh_Rasmussen/Anders_Fogh_Rasmussen_0003.jpg
+Gray_Davis/Gray_Davis_0004.jpg
+Sally_Ride/Sally_Ride_0001.jpg
+Kim_Yong-il/Kim_Yong-il_0001.jpg
+Jon_Constance/Jon_Constance_0001.jpg
+Patrick_Rafter/Patrick_Rafter_0001.jpg
+Christopher_Speer/Christopher_Speer_0001.jpg
+Antonio_Cassano/Antonio_Cassano_0001.jpg
+Eduard_Shevardnadze/Eduard_Shevardnadze_0005.jpg
+Paul_Reiser/Paul_Reiser_0001.jpg
+Jose_Woldenberg/Jose_Woldenberg_0001.jpg
+Steve_Zahn/Steve_Zahn_0001.jpg
+Carlos_Queiroz/Carlos_Queiroz_0001.jpg
+Kaio_Almeida/Kaio_Almeida_0001.jpg
+Henry_Hyde/Henry_Hyde_0001.jpg
+Vladimir_Golovlyov/Vladimir_Golovlyov_0001.jpg
+Bill_Elliott/Bill_Elliott_0001.jpg
+Glen_DaSilva/Glen_DaSilva_0001.jpg
+Vernon_Forrest/Vernon_Forrest_0001.jpg
+Thomas_Haeggstroem/Thomas_Haeggstroem_0001.jpg
+Gerry_Parsky/Gerry_Parsky_0001.jpg
+Russell_Simmons/Russell_Simmons_0004.jpg
+Floyd_Mayweather/Floyd_Mayweather_0001.jpg
+Tex_Ritter/Tex_Ritter_0001.jpg
+Ken_Wharfe/Ken_Wharfe_0001.jpg
+Lee_Yeo-jin/Lee_Yeo-jin_0001.jpg
+Kathy_Gannon/Kathy_Gannon_0001.jpg
+Laila_Ali/Laila_Ali_0003.jpg
+Janice_Goldfinger/Janice_Goldfinger_0001.jpg
+Emanuel_Ginobili/Emanuel_Ginobili_0002.jpg
+Phil_Bennett/Phil_Bennett_0001.jpg
+Des_Brown/Des_Brown_0001.jpg
+Angelo_Reyes/Angelo_Reyes_0003.jpg
+Mariangel_Ruiz_Torrealba/Mariangel_Ruiz_Torrealba_0002.jpg
+Sachiko_Yamada/Sachiko_Yamada_0004.jpg
+Francisco_Urenda/Francisco_Urenda_0001.jpg
+Shi_Guangsheng/Shi_Guangsheng_0001.jpg
+Dick_Latessa/Dick_Latessa_0002.jpg
+William_Burns/William_Burns_0001.jpg
+Gabriel_Jorge_Ferreia/Gabriel_Jorge_Ferreia_0001.jpg
+Robert_Downey_Jr/Robert_Downey_Jr_0001.jpg
+David_Oh/David_Oh_0001.jpg
+Demetrius_Ferraciu/Demetrius_Ferraciu_0002.jpg
+Liu_Ye/Liu_Ye_0001.jpg
+Georgi_Parvanov/Georgi_Parvanov_0002.jpg
+David_Provost/David_Provost_0001.jpg
+Lionel_Richie/Lionel_Richie_0002.jpg
+Roy_Moore/Roy_Moore_0005.jpg
+Roman_Abramovich/Roman_Abramovich_0001.jpg
+Compay_Segundo/Compay_Segundo_0001.jpg
+Maria_Wetterstrand/Maria_Wetterstrand_0001.jpg
+Pele/Pele_0001.jpg
+Lee_Ann_Womack/Lee_Ann_Womack_0001.jpg
+Anette_Hosoi/Anette_Hosoi_0001.jpg
+Richard_Haass/Richard_Haass_0001.jpg
+Jim_Wong/Jim_Wong_0001.jpg
+Giuseppe_Morchio/Giuseppe_Morchio_0001.jpg
+Ronnie_Musgrove/Ronnie_Musgrove_0001.jpg
+Ari_Fleischer/Ari_Fleischer_0009.jpg
+Nancy_Humbert/Nancy_Humbert_0001.jpg
+John_Nimmo/John_Nimmo_0001.jpg
+Hank_McKinnell/Hank_McKinnell_0001.jpg
+Daniela_Hantuchova/Daniela_Hantuchova_0001.jpg
+Francois_Pienaar/Francois_Pienaar_0001.jpg
+Adoor_Gopalakarishnan/Adoor_Gopalakarishnan_0001.jpg
+Thomas_Ferguson/Thomas_Ferguson_0001.jpg
+Natanaela_Barnova/Natanaela_Barnova_0001.jpg
+Asmaa_Assad/Asmaa_Assad_0001.jpg
+Jeff_Bridges/Jeff_Bridges_0001.jpg
+Guillermo_Coria/Guillermo_Coria_0022.jpg
+Brad_Johnson/Brad_Johnson_0001.jpg
+Carlos_Ortega/Carlos_Ortega_0001.jpg
+Chuck_Hagel/Chuck_Hagel_0001.jpg
+Tommy_Lasorda/Tommy_Lasorda_0001.jpg
+John_Coomber/John_Coomber_0001.jpg
+Raul_Ibanez/Raul_Ibanez_0001.jpg
+Augustin_Calleri/Augustin_Calleri_0002.jpg
+Ivo_Dubs/Ivo_Dubs_0001.jpg
+Mitar_Rasevic/Mitar_Rasevic_0001.jpg
+Margerry_Bakley/Margerry_Bakley_0001.jpg
+Antonio_Elias_Saca/Antonio_Elias_Saca_0001.jpg
+Hatsui_Hasuike/Hatsui_Hasuike_0001.jpg
+Barbara_Bach/Barbara_Bach_0001.jpg
+Clemente_de_la_Vega/Clemente_de_la_Vega_0001.jpg
+Hoda_Asfor/Hoda_Asfor_0001.jpg
+Sean_Hayes/Sean_Hayes_0002.jpg
+Paul_William_Hurley/Paul_William_Hurley_0001.jpg
+Chris_Reitsma/Chris_Reitsma_0001.jpg
+James_Wolfensohn/James_Wolfensohn_0005.jpg
+Rod_Bryden/Rod_Bryden_0001.jpg
+Kay_Bailey_Hutchison/Kay_Bailey_Hutchison_0001.jpg
+Robert_Schuller/Robert_Schuller_0001.jpg
+Bob_Hartley/Bob_Hartley_0001.jpg
+Kim_Gandy/Kim_Gandy_0001.jpg
+Dick_Posthumus/Dick_Posthumus_0001.jpg
+Paula_Radcliffe/Paula_Radcliffe_0004.jpg
+Sebastien_Grosjean/Sebastien_Grosjean_0004.jpg
+Patricia_Medina/Patricia_Medina_0001.jpg
+Tali_Imani/Tali_Imani_0001.jpg
+Barry_Zito/Barry_Zito_0002.jpg
+Tom_Foy/Tom_Foy_0001.jpg
+Rien_Long/Rien_Long_0001.jpg
+Alanis_Morissette/Alanis_Morissette_0001.jpg
+Cathy_Chisholm/Cathy_Chisholm_0001.jpg
+Elizabeth_Smart/Elizabeth_Smart_0004.jpg
+Gillian_Anderson/Gillian_Anderson_0002.jpg
+Michael_Taylor/Michael_Taylor_0001.jpg
+Franklin_Brown/Franklin_Brown_0001.jpg
+Vadim_Strogalev/Vadim_Strogalev_0001.jpg
+Andy_Perez/Andy_Perez_0001.jpg
+Mary_Catherine_Correll/Mary_Catherine_Correll_0001.jpg
+Ai_Sugiyama/Ai_Sugiyama_0004.jpg
+Tian_Zhuang_Zhuang/Tian_Zhuang_Zhuang_0001.jpg
+Len_Jenoff/Len_Jenoff_0001.jpg
+Don_Flanagan/Don_Flanagan_0001.jpg
+Marisa_Tomei/Marisa_Tomei_0002.jpg
+Bill_Parsons/Bill_Parsons_0001.jpg
+Gwendal_Peizerat/Gwendal_Peizerat_0001.jpg
+Jane_Fonda/Jane_Fonda_0002.jpg
+Ethan_Hawke/Ethan_Hawke_0001.jpg
+Colin_Montgomerie/Colin_Montgomerie_0005.jpg
+Kieran_Culkin/Kieran_Culkin_0001.jpg
+Debra_Rose/Debra_Rose_0001.jpg
+Aidan_Quinn/Aidan_Quinn_0001.jpg
+Dita_Von_Tesse/Dita_Von_Tesse_0001.jpg
+Jim_Hendry/Jim_Hendry_0001.jpg
+William_Webster/William_Webster_0001.jpg
+Radovan_Karadzic/Radovan_Karadzic_0001.jpg
+Nancy_Pelosi/Nancy_Pelosi_0004.jpg
+Mark_Hamister/Mark_Hamister_0001.jpg
+Mauro_Viza/Mauro_Viza_0001.jpg
+Ricardo_Maduro/Ricardo_Maduro_0001.jpg
+Brandon_Hammond/Brandon_Hammond_0001.jpg
+Fred_Rogers/Fred_Rogers_0001.jpg
+Munir_Akram/Munir_Akram_0001.jpg
+Brooke_Shields/Brooke_Shields_0002.jpg
+Chung_Mong-joon/Chung_Mong-joon_0001.jpg
+Terje_Roed-Larsen/Terje_Roed-Larsen_0002.jpg
+Kwon_Yang-sook/Kwon_Yang-sook_0003.jpg
+Bruce_Lunsford/Bruce_Lunsford_0001.jpg
+Amy_Brenneman/Amy_Brenneman_0001.jpg
+Michelle_Pfeiffer/Michelle_Pfeiffer_0001.jpg
+Yannos_Papantoniou/Yannos_Papantoniou_0001.jpg
+Noelle_Bush/Noelle_Bush_0004.jpg
+Doc_Rivers/Doc_Rivers_0001.jpg
+Rob_Moore/Rob_Moore_0001.jpg
+Herman_Edwards/Herman_Edwards_0001.jpg
+Cassandra_Heise/Cassandra_Heise_0001.jpg
+Milan_Kucan/Milan_Kucan_0001.jpg
+Wendy_Kennedy/Wendy_Kennedy_0001.jpg
+Koichiro_Matsuura/Koichiro_Matsuura_0001.jpg
+Han_Sung_Joo/Han_Sung_Joo_0001.jpg
+Eric_Benet/Eric_Benet_0001.jpg
+Shirley_Jones/Shirley_Jones_0001.jpg
+Walter_Woods/Walter_Woods_0001.jpg
+Jamie_Olis/Jamie_Olis_0001.jpg
+William_Umbach/William_Umbach_0001.jpg
+Phan_Van_Khai/Phan_Van_Khai_0001.jpg
+Xiang_Xu/Xiang_Xu_0001.jpg
+Mary_Frances_Seiter/Mary_Frances_Seiter_0001.jpg
+Ibrahim_Al-Marashi/Ibrahim_Al-Marashi_0001.jpg
+Elmar_Brok/Elmar_Brok_0001.jpg
+Brian_Cook/Brian_Cook_0001.jpg
+Keith_Rodriguez/Keith_Rodriguez_0001.jpg
+Benjamin_Neulander/Benjamin_Neulander_0001.jpg
+Gabriella_Bo/Gabriella_Bo_0001.jpg
+Alfredo_Pena/Alfredo_Pena_0001.jpg
+Tim_Lopes/Tim_Lopes_0001.jpg
+Trisha_Meili/Trisha_Meili_0001.jpg
+Martin_Bandier/Martin_Bandier_0001.jpg
+Pervez_Musharraf/Pervez_Musharraf_0018.jpg
+Serge_Melac/Serge_Melac_0001.jpg
+Li_Peng/Li_Peng_0004.jpg
+Dan_Snyder/Dan_Snyder_0001.jpg
+Inocencio_Arias/Inocencio_Arias_0001.jpg
+Charmaine_Crooks/Charmaine_Crooks_0001.jpg
+Yu_Shyi-kun/Yu_Shyi-kun_0004.jpg
+Saeb_Erekat/Saeb_Erekat_0001.jpg
+Lou_Reed/Lou_Reed_0001.jpg
+Richard_Reid/Richard_Reid_0001.jpg
+Judi_Patton/Judi_Patton_0001.jpg
+Prince_William/Prince_William_0001.jpg
+James_Baker/James_Baker_0001.jpg
+John_Cusack/John_Cusack_0001.jpg
+Zarai_Toledo/Zarai_Toledo_0001.jpg
+Kelly_Leigh/Kelly_Leigh_0001.jpg
+Art_Cooper/Art_Cooper_0001.jpg
+Florencia_Kirchner/Florencia_Kirchner_0001.jpg
+Jimmy_Smits/Jimmy_Smits_0001.jpg
+Peter_Ahearn/Peter_Ahearn_0001.jpg
+Pat_Riley/Pat_Riley_0001.jpg
+John_Reid/John_Reid_0001.jpg
+Leandro_Andrade/Leandro_Andrade_0001.jpg
+Julie_Infante/Julie_Infante_0001.jpg
+Mary_Zorn/Mary_Zorn_0001.jpg
+Bob_Riley/Bob_Riley_0001.jpg
+James_Kelly/James_Kelly_0009.jpg
+Lydia_Shum/Lydia_Shum_0001.jpg
+Bak_Chang-Ryun/Bak_Chang-Ryun_0001.jpg
+Sereyvuth_Kem/Sereyvuth_Kem_0001.jpg
+Gilles_Panizzi/Gilles_Panizzi_0001.jpg
+Patty_Schnyder/Patty_Schnyder_0004.jpg
+Patrick_Stewart/Patrick_Stewart_0001.jpg
+Sherry_Irving/Sherry_Irving_0001.jpg
+Dwayne_Williams/Dwayne_Williams_0001.jpg
+Daniel_Bruehl/Daniel_Bruehl_0001.jpg
+Jay_Leno/Jay_Leno_0003.jpg
+James_Dingemans/James_Dingemans_0001.jpg
+Tino_Martinez/Tino_Martinez_0001.jpg
+Jennifer_Connelly/Jennifer_Connelly_0004.jpg
+Peggy_McGuinness/Peggy_McGuinness_0001.jpg
+Meirion_Evans/Meirion_Evans_0001.jpg
+Steve_Fehr/Steve_Fehr_0001.jpg
+Jamie_Cooke/Jamie_Cooke_0001.jpg
+Kemal_Dervis/Kemal_Dervis_0001.jpg
+Bob_Wright/Bob_Wright_0001.jpg
+Arie_Haan/Arie_Haan_0001.jpg
+Ana_Paula_Gerard/Ana_Paula_Gerard_0001.jpg
+Dwain_Kyles/Dwain_Kyles_0001.jpg
+Scott_Hoch/Scott_Hoch_0001.jpg
+Carlo_Azeglio_Ciampi/Carlo_Azeglio_Ciampi_0001.jpg
+Rick_Lu/Rick_Lu_0001.jpg
+Uzi_Landau/Uzi_Landau_0001.jpg
+Jack_Goodman/Jack_Goodman_0001.jpg
+Charlie_Hunnam/Charlie_Hunnam_0001.jpg
+Jen_Schefft/Jen_Schefft_0001.jpg
+Newt_Gingrich/Newt_Gingrich_0001.jpg
+Mark_Cuban/Mark_Cuban_0001.jpg
+Princess_Hisako/Princess_Hisako_0001.jpg
+Princess_Maxima/Princess_Maxima_0001.jpg
+Rolandas_Paksas/Rolandas_Paksas_0002.jpg
+Bill_Fennelly/Bill_Fennelly_0001.jpg
+Brad_Alexander_Smith/Brad_Alexander_Smith_0001.jpg
+Olivier_Boulay/Olivier_Boulay_0001.jpg
+Dan_Kellner/Dan_Kellner_0001.jpg
+Lucy_Liu/Lucy_Liu_0002.jpg
+Fred_Durst/Fred_Durst_0001.jpg
+Trent_Lott/Trent_Lott_0008.jpg
+Rebekah_Chantay_Revels/Rebekah_Chantay_Revels_0004.jpg
+Jerry_Angelo/Jerry_Angelo_0001.jpg
+Marco_Archer_Cardoso_Moreira/Marco_Archer_Cardoso_Moreira_0001.jpg
+Amr_Moussa/Amr_Moussa_0001.jpg
+Cliff_Ellis/Cliff_Ellis_0001.jpg
+Paul_Sarbanes/Paul_Sarbanes_0001.jpg
+Sylvia_Plachy/Sylvia_Plachy_0001.jpg
+Alex_Gonzalez/Alex_Gonzalez_0001.jpg
+Crandall_Bowles/Crandall_Bowles_0001.jpg
+Jane_Menelaus/Jane_Menelaus_0001.jpg
+Paula_Prentiss/Paula_Prentiss_0001.jpg
+Marco_Antonio_Barrera/Marco_Antonio_Barrera_0003.jpg
+Don_Lake/Don_Lake_0001.jpg
+Garry_McCoy/Garry_McCoy_0001.jpg
+Jim_Spinoza/Jim_Spinoza_0001.jpg
+Kenneth_Reichert/Kenneth_Reichert_0001.jpg
+Leonid_Kuchma/Leonid_Kuchma_0006.jpg
+Gerhard_Schmid/Gerhard_Schmid_0001.jpg
+Raaf_Schefter/Raaf_Schefter_0001.jpg
+Lleyton_Hewitt/Lleyton_Hewitt_0041.jpg
+Walter_Annenberg/Walter_Annenberg_0001.jpg
+Kim_Dae-jung/Kim_Dae-jung_0002.jpg
+Bruce_Arena/Bruce_Arena_0001.jpg
+Roberto_Cercelletta/Roberto_Cercelletta_0001.jpg
+Elva_Hsiao/Elva_Hsiao_0001.jpg
+Vojislav_Kostunica/Vojislav_Kostunica_0006.jpg
+Satnarine_Sharma/Satnarine_Sharma_0001.jpg
+Brendan_Gaughan/Brendan_Gaughan_0001.jpg
+John_Ashcroft/John_Ashcroft_0038.jpg
+Kathie_Louise_Saunders/Kathie_Louise_Saunders_0001.jpg
+Tina_Pisnik/Tina_Pisnik_0001.jpg
+Geraldine_Chaplin/Geraldine_Chaplin_0002.jpg
+Amanda_Marsh/Amanda_Marsh_0001.jpg
+Jennifer_Renee_Short/Jennifer_Renee_Short_0001.jpg
+Troy_Garity/Troy_Garity_0001.jpg
+John_Belushi/John_Belushi_0001.jpg
+Jennifer_Gratz/Jennifer_Gratz_0001.jpg
+Andrew_Wetzler/Andrew_Wetzler_0001.jpg
+Amanda_Plumer/Amanda_Plumer_0001.jpg
+Colin_Cowie/Colin_Cowie_0001.jpg
+Joe_Darrell/Joe_Darrell_0001.jpg
+Dino_Risi/Dino_Risi_0001.jpg
+Mickey_Rooney/Mickey_Rooney_0001.jpg
+Bob_Herz/Bob_Herz_0001.jpg
+Xavier_Malisse/Xavier_Malisse_0005.jpg
+Ludwig_Ovalle/Ludwig_Ovalle_0001.jpg
+Michael_Caine/Michael_Caine_0003.jpg
+Agbani_Darego/Agbani_Darego_0001.jpg
+Abdullatif_Sener/Abdullatif_Sener_0002.jpg
+Zeljko_Rebraca/Zeljko_Rebraca_0001.jpg
+Tom_Gamboa/Tom_Gamboa_0001.jpg
+Marina_Silva/Marina_Silva_0001.jpg
+Tommy_Lewis/Tommy_Lewis_0001.jpg
+Bernard_Giraudeau/Bernard_Giraudeau_0001.jpg
+Giulio_Andreotti/Giulio_Andreotti_0001.jpg
+Jake_Plummer/Jake_Plummer_0001.jpg
+Andy_Dick/Andy_Dick_0001.jpg
+Latrell_Sprewell/Latrell_Sprewell_0001.jpg
+Sung_Hong_Choi/Sung_Hong_Choi_0001.jpg
+Dominique_de_Villepin/Dominique_de_Villepin_0014.jpg
+Kyle_McLaren/Kyle_McLaren_0001.jpg
+Alex_Cabrera/Alex_Cabrera_0001.jpg
+Lawrence_Roberts/Lawrence_Roberts_0001.jpg
+Heinz_Feldmann/Heinz_Feldmann_0001.jpg
+Sok_An/Sok_An_0001.jpg
+Arnaud_Lagardere/Arnaud_Lagardere_0001.jpg
+Kim_Dong-tae/Kim_Dong-tae_0001.jpg
+Richard_Penniman/Richard_Penniman_0001.jpg
+Leandro_Garcia/Leandro_Garcia_0001.jpg
+Jane_Walker_Wood/Jane_Walker_Wood_0001.jpg
+Antonio_Trillanes/Antonio_Trillanes_0003.jpg
+Pascal_Quignard/Pascal_Quignard_0003.jpg
+Herta_Daeubler-Gmelin/Herta_Daeubler-Gmelin_0001.jpg
+Francisco_Garcia/Francisco_Garcia_0001.jpg
+Drew_Barrymore/Drew_Barrymore_0001.jpg
+Adam_Scott/Adam_Scott_0001.jpg
+Jennifer_Furminger/Jennifer_Furminger_0001.jpg
+Thomas_Manger/Thomas_Manger_0001.jpg
+Guillaume_Cannet/Guillaume_Cannet_0001.jpg
+Rodrigo_de_la_Cerna/Rodrigo_de_la_Cerna_0001.jpg
+Noel_Forgeard/Noel_Forgeard_0001.jpg
+Michael_Sullivan/Michael_Sullivan_0002.jpg
+Joe_Friedberg/Joe_Friedberg_0001.jpg
+Pio_Laghi/Pio_Laghi_0001.jpg
+Janusz_Kaminski/Janusz_Kaminski_0001.jpg
+Desiree_Lemosi/Desiree_Lemosi_0001.jpg
+Steve_Park/Steve_Park_0001.jpg
+Mary_Robinson/Mary_Robinson_0001.jpg
+Craig_Wilson/Craig_Wilson_0001.jpg
+Roger_Federer/Roger_Federer_0001.jpg
+Jason_Alexander/Jason_Alexander_0001.jpg
+Kenneth_Cooper/Kenneth_Cooper_0001.jpg
+Joe_Calzaghe/Joe_Calzaghe_0001.jpg
+Sabah_Al-Ahmad_Al-Jaber_Al-Sabah/Sabah_Al-Ahmad_Al-Jaber_Al-Sabah_0001.jpg
+Norm_Coleman/Norm_Coleman_0002.jpg
+Jerry_Tarkanian/Jerry_Tarkanian_0001.jpg
+Jorma_Huhtala/Jorma_Huhtala_0001.jpg
+Andrei_Konchalovsky/Andrei_Konchalovsky_0001.jpg
+Alexandre_Despatie/Alexandre_Despatie_0001.jpg
+Tim_Duncan/Tim_Duncan_0001.jpg
+Richard_Lugar/Richard_Lugar_0001.jpg
+Alexander_Lukashenko/Alexander_Lukashenko_0001.jpg
+Enrique_Bolanos/Enrique_Bolanos_0002.jpg
+Gregory_Geoffroy/Gregory_Geoffroy_0002.jpg
+Li_Zhaoxing/Li_Zhaoxing_0008.jpg
+Sherry_Fisher/Sherry_Fisher_0001.jpg
+John_Garamendi/John_Garamendi_0001.jpg
+Roger_King/Roger_King_0001.jpg
+Michel_Minard/Michel_Minard_0001.jpg
diff --git a/cpp-package/inspireface/toolchain/FindTensorRT.cmake b/cpp-package/inspireface/toolchain/FindTensorRT.cmake
new file mode 100644
index 000000000..834783a24
--- /dev/null
+++ b/cpp-package/inspireface/toolchain/FindTensorRT.cmake
@@ -0,0 +1,35 @@
+# FindTensorRT.cmake - Simple Version
+# Contains basic functionality for finding TensorRT libraries
+
+# Find CUDA
+find_package(CUDA REQUIRED)
+include_directories(${CUDA_INCLUDE_DIRS})
+link_directories(${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+
+# Find TensorRT header files
+find_path(TENSORRT_INCLUDE_DIR NvInfer.h
+    HINTS ${TENSORRT_ROOT}
+    PATH_SUFFIXES include)
+
+# Find TensorRT libraries
+find_library(TENSORRT_LIBRARY_INFER nvinfer
+    HINTS ${TENSORRT_ROOT}
+    PATH_SUFFIXES lib lib64)
+
+find_library(TENSORRT_LIBRARY_RUNTIME nvinfer_plugin
+    HINTS ${TENSORRT_ROOT}
+    PATH_SUFFIXES lib lib64)
+
+# Find CUDA runtime library
+find_library(CUDA_RUNTIME_LIBRARY cudart
+    HINTS ${CUDA_TOOLKIT_ROOT_DIR}
+    PATH_SUFFIXES lib64 lib lib64/stubs lib/stubs)
+
+# Set result variables, can be used in projects that include this module
+set(ISF_TENSORRT_INCLUDE_DIRS ${TENSORRT_INCLUDE_DIR})
+set(ISF_TENSORRT_LIBRARIES ${TENSORRT_LIBRARY_INFER} ${TENSORRT_LIBRARY_RUNTIME} ${CUDA_RUNTIME_LIBRARY})
+
+# Output status messages
+message(STATUS "Found TensorRT include: ${TENSORRT_INCLUDE_DIR}")
+message(STATUS "Found TensorRT libraries: ${TENSORRT_LIBRARY_INFER} ${TENSORRT_LIBRARY_RUNTIME}")
+message(STATUS "Found CUDA runtime library: ${CUDA_RUNTIME_LIBRARY}")
\ No newline at end of file
diff --git a/cpp-package/inspireface/toolchain/ios.toolchain.cmake b/cpp-package/inspireface/toolchain/ios.toolchain.cmake
new file mode 100644
index 000000000..4afc1db27
--- /dev/null
+++ b/cpp-package/inspireface/toolchain/ios.toolchain.cmake
@@ -0,0 +1,486 @@
+# This file is part of the ios-cmake project. It was retrieved from
+# https://github.com/cristeab/ios-cmake.git, which is a fork of
+# https://code.google.com/p/ios-cmake/. Which in turn is based off of
+# the Platform/Darwin.cmake and Platform/UnixPaths.cmake files which
+# are included with CMake 2.8.4
+#
+# The ios-cmake project is licensed under the new BSD license.
+#
+# Copyright (c) 2014, Bogdan Cristea and LTE Engineering Software,
+# Kitware, Inc., Insight Software Consortium.  All rights reserved.
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+# 1. Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#
+# 2. Redistributions in binary form must reproduce the above copyright
+# notice, this list of conditions and the following disclaimer in the
+# documentation and/or other materials provided with the distribution.
+#
+# 3. Neither the name of the copyright holder nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+# FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+# COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+# INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+# BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+# LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+# ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+# POSSIBILITY OF SUCH DAMAGE.
+#
+# This file is based off of the Platform/Darwin.cmake and
+# Platform/UnixPaths.cmake files which are included with CMake 2.8.4
+# It has been altered for iOS development.
+#
+# Updated by Alex Stewart (alexs.mac@gmail.com)
+#
+# *****************************************************************************
+#      Now maintained by Alexander Widerberg (widerbergaren [at] gmail.com)
+#                      under the BSD-3-Clause license
+#                   https://github.com/leetal/ios-cmake
+# *****************************************************************************
+#
+#                           INFORMATION / HELP
+#
+# The following variables control the behaviour of this toolchain:
+#
+# IOS_PLATFORM: OS (default) or SIMULATOR or SIMULATOR64 or TVOS or SIMULATOR_TVOS or WATCHOS or SIMULATOR_WATCHOS
+#    OS = Build for iPhoneOS.
+#    OS64 = Build for arm64 arm64e iPhoneOS.
+#    SIMULATOR = Build for x86 i386 iPhone Simulator.
+#    SIMULATOR64 = Build for x86_64 iPhone Simulator.
+#    TVOS = Build for AppleTVOS.
+#    SIMULATOR_TVOS = Build for x86_64 AppleTV Simulator.
+#    WATCHOS = Build for armv7k arm64_32 for WatchOS.
+#    SIMULATOR_WATCHOS = Build for x86_64 for Watch Simulator.
+# CMAKE_OSX_SYSROOT: Path to the iOS SDK to use.  By default this is
+#    automatically determined from IOS_PLATFORM and xcodebuild, but
+#    can also be manually specified (although this should not be required).
+# CMAKE_IOS_DEVELOPER_ROOT: Path to the Developer directory for the iOS platform
+#    being compiled for.  By default this is automatically determined from
+#    CMAKE_OSX_SYSROOT, but can also be manually specified (although this should
+#    not be required).
+# ENABLE_BITCODE: (1|0) Enables or disables bitcode support. Default 1 (true)
+# ENABLE_ARC: (1|0) Enables or disables ARC support. Default 1 (true, ARC enabled by default)
+# ENABLE_VISIBILITY: (1|0) Enables or disables symbol visibility support. Default 0 (false, visibility hidden by default)
+# IOS_ARCH: (armv7 armv7s armv7k arm64 arm64e arm64_32 i386 x86_64) If specified, will override the default architectures for the given IOS_PLATFORM
+#    OS = armv7 armv7s arm64 arm64e (if applicable)
+#    OS64 = arm64 arm64e (if applicable)
+#    SIMULATOR = i386 x86_64
+#    SIMULATOR64 = x86_64
+#    TVOS = arm64
+#    SIMULATOR_TVOS = x86_64 (i386 has since long been deprecated)
+#    WATCHOS = armv7k arm64_32 (if applicable)
+#    SIMULATOR_WATCHOS = x86_64 (i386 has since long been deprecated)
+#
+# This toolchain defines the following variables for use externally:
+#
+# XCODE_VERSION: Version number (not including Build version) of Xcode detected.
+# IOS_SDK_VERSION: Version of iOS SDK being used.
+# CMAKE_OSX_ARCHITECTURES: Architectures being compiled for (generated from
+#    IOS_PLATFORM).
+#
+# This toolchain defines the following macros for use externally:
+#
+# set_xcode_property (TARGET XCODE_PROPERTY XCODE_VALUE XCODE_VARIANT)
+#   A convenience macro for setting xcode specific properties on targets.
+#   Available variants are: All, Release, RelWithDebInfo, Debug, MinSizeRel
+#   example: set_xcode_property (myioslib IPHONEOS_DEPLOYMENT_TARGET "3.1" "all").
+#
+# find_host_package (PROGRAM ARGS)
+#   A macro used to find executable programs on the host system, not within the
+#   iOS environment.  Thanks to the android-cmake project for providing the
+#   command.
+
+# Fix for PThread library not in path
+set(CMAKE_THREAD_LIBS_INIT "-lpthread")
+set(CMAKE_HAVE_THREADS_LIBRARY 1)
+set(CMAKE_USE_WIN32_THREADS_INIT 0)
+set(CMAKE_USE_PTHREADS_INIT 1)
+
+# Get the Xcode version being used.
+execute_process(COMMAND xcodebuild -version
+  OUTPUT_VARIABLE XCODE_VERSION
+  ERROR_QUIET
+  OUTPUT_STRIP_TRAILING_WHITESPACE)
+string(REGEX MATCH "Xcode [0-9\\.]+" XCODE_VERSION "${XCODE_VERSION}")
+string(REGEX REPLACE "Xcode ([0-9\\.]+)" "\\1" XCODE_VERSION "${XCODE_VERSION}")
+message(STATUS "Building with Xcode version: ${XCODE_VERSION}")
+# Default to building for iPhoneOS if not specified otherwise, and we cannot
+# determine the platform from the CMAKE_OSX_ARCHITECTURES variable. The use
+# of CMAKE_OSX_ARCHITECTURES is such that try_compile() projects can correctly
+# determine the value of IOS_PLATFORM from the root project, as
+# CMAKE_OSX_ARCHITECTURES is propagated to them by CMake.
+if (NOT DEFINED IOS_PLATFORM)
+  if (CMAKE_OSX_ARCHITECTURES)
+    if (CMAKE_OSX_ARCHITECTURES MATCHES ".*arm.*")
+      set(IOS_PLATFORM "OS")
+    elseif (CMAKE_OSX_ARCHITECTURES MATCHES "i386")
+      set(IOS_PLATFORM "SIMULATOR")
+    elseif (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+      set(IOS_PLATFORM "SIMULATOR64")
+    elseif (CMAKE_OSX_ARCHITECTURES MATCHES "armv7k")
+      set(IOS_PLATFORM "WATCHOS")
+    endif()
+  endif()
+  if (NOT IOS_PLATFORM)
+    set(IOS_PLATFORM "OS")
+  endif()
+endif()
+set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING
+  "Type of iOS platform for which to build.")
+# Determine the platform name and architectures for use in xcodebuild commands
+# from the specified IOS_PLATFORM name.
+if (IOS_PLATFORM STREQUAL "OS")
+  set(XCODE_IOS_PLATFORM iphoneos)
+  if(NOT IOS_ARCH)
+    if (XCODE_VERSION VERSION_GREATER 10.0)
+      set(IOS_ARCH armv7 armv7s arm64 arm64e)
+    else()
+      set(IOS_ARCH armv7 armv7s arm64)
+    endif()
+  endif()
+ elseif (IOS_PLATFORM STREQUAL "OS64")
+  set(XCODE_IOS_PLATFORM iphoneos)
+  if(NOT IOS_ARCH)
+    if (XCODE_VERSION VERSION_GREATER 10.0)
+      set(IOS_ARCH arm64 arm64e)
+    else()
+      set(IOS_ARCH arm64)
+    endif()
+  endif()
+elseif (IOS_PLATFORM STREQUAL "SIMULATOR")
+  set(XCODE_IOS_PLATFORM iphonesimulator)
+  if(NOT IOS_ARCH)
+    set(IOS_ARCH i386 x86_64)
+  endif()
+elseif(IOS_PLATFORM STREQUAL "SIMULATOR64")
+  set(XCODE_IOS_PLATFORM iphonesimulator)
+  if(NOT IOS_ARCH)
+    set(IOS_ARCH x86_64)
+  endif()
+elseif (IOS_PLATFORM STREQUAL "TVOS")
+  set(XCODE_IOS_PLATFORM appletvos)
+  if(NOT IOS_ARCH)
+    set(IOS_ARCH arm64)
+  endif()
+elseif (IOS_PLATFORM STREQUAL "SIMULATOR_TVOS")
+  set(XCODE_IOS_PLATFORM appletvsimulator)
+  if(NOT IOS_ARCH)
+    set(IOS_ARCH x86_64)
+  endif()
+elseif (IOS_PLATFORM STREQUAL "WATCHOS")
+  set(XCODE_IOS_PLATFORM watchos)
+  if(NOT IOS_ARCH)
+    if (XCODE_VERSION VERSION_GREATER 10.0)
+      set(IOS_ARCH armv7k arm64_32)
+    else()
+      set(IOS_ARCH armv7k)
+    endif()
+  endif()
+elseif (IOS_PLATFORM STREQUAL "SIMULATOR_WATCHOS")
+  set(XCODE_IOS_PLATFORM watchsimulator)
+  if(NOT IOS_ARCH)
+    set(IOS_ARCH x86_64)
+  endif()
+else()
+  message(FATAL_ERROR "Invalid IOS_PLATFORM: ${IOS_PLATFORM}")
+endif()
+message(STATUS "Configuring iOS build for platform: ${IOS_PLATFORM}, "
+  "architecture(s): ${IOS_ARCH}")
+# If user did not specify the SDK root to use, then query xcodebuild for it.
+execute_process(COMMAND xcodebuild -version -sdk ${XCODE_IOS_PLATFORM} Path
+    OUTPUT_VARIABLE CMAKE_OSX_SYSROOT_INT
+     OUTPUT_QUIET ERROR_QUIET
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+# If user did not specify the SDK root to use, then query xcodebuild for it.
+if (NOT DEFINED CMAKE_OSX_SYSROOT OR (NOT CMAKE_OSX_SYSROOT STREQUAL CMAKE_OSX_SYSROOT_INT))
+  execute_process(COMMAND xcodebuild -version -sdk ${XCODE_IOS_PLATFORM} Path
+    OUTPUT_VARIABLE CMAKE_OSX_SYSROOT
+    ERROR_QUIET
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+endif()
+if (NOT EXISTS ${CMAKE_OSX_SYSROOT})
+  message(SEND_ERROR "Please make sure that Xcode is installed and that the toolchain"
+  "is pointing to the correct path. Please run:"
+  "sudo xcode-select -s /Applications/Xcode.app/Contents/Developer"
+  "and see if that fixes the problem for you.")
+  message(FATAL_ERROR "Invalid CMAKE_OSX_SYSROOT: ${CMAKE_OSX_SYSROOT} "
+  "does not exist.")
+elseif(DEFINED CMAKE_OSX_SYSROOT)
+  message(STATUS "Using manually set SDK path: ${CMAKE_OSX_SYSROOT} for platform: ${IOS_PLATFORM}")
+else()
+   message(STATUS "Using SDK: ${CMAKE_OSX_SYSROOT} for platform: ${IOS_PLATFORM}")
+endif()
+# Specify minimum version of deployment target.
+if (NOT DEFINED IOS_DEPLOYMENT_TARGET)
+  if (IOS_PLATFORM STREQUAL "WATCHOS" OR IOS_PLATFORM STREQUAL "SIMULATOR_WATCHOS")
+    # Unless specified, SDK version 2.0 is used by default as minimum target version (watchOS).
+    set(IOS_DEPLOYMENT_TARGET "2.0"
+            CACHE STRING "Minimum iOS version to build for." )
+  else()
+    # Unless specified, SDK version 8.0 is used by default as minimum target version (iOS, tvOS).
+    set(IOS_DEPLOYMENT_TARGET "8.0"
+            CACHE STRING "Minimum iOS version to build for." )
+  endif()
+  message(STATUS "Using the default min-version since IOS_DEPLOYMENT_TARGET not provided!")
+endif()
+# Use bitcode or not
+if (NOT DEFINED ENABLE_BITCODE AND NOT IOS_ARCH MATCHES "((^|, )(i386|x86_64))+")
+  # Unless specified, enable bitcode support by default
+  set(ENABLE_BITCODE TRUE CACHE BOOL "Whether or not to enable bitcode")
+  message(STATUS "Enabling bitcode support by default. ENABLE_BITCODE not provided!")
+endif()
+if (NOT DEFINED ENABLE_BITCODE)
+  message(STATUS "Disabling bitcode support by default on simulators. ENABLE_BITCODE not provided for override!")
+endif()
+# Use ARC or not
+if (NOT DEFINED ENABLE_ARC)
+  # Unless specified, enable ARC support by default
+  set(ENABLE_ARC TRUE CACHE BOOL "Whether or not to enable ARC")
+  message(STATUS "Enabling ARC support by default. ENABLE_ARC not provided!")
+endif()
+# Use hidden visibility or not
+if (NOT DEFINED ENABLE_VISIBILITY)
+  # Unless specified, disable symbols visibility by default
+  set(ENABLE_VISIBILITY FALSE CACHE BOOL "Whether or not to hide symbols (-fvisibility=hidden)")
+  message(STATUS "Hiding symbols visibility by default. ENABLE_VISIBILITY not provided!")
+endif()
+# Get the SDK version information.
+execute_process(COMMAND xcodebuild -sdk ${CMAKE_OSX_SYSROOT} -version SDKVersion
+  OUTPUT_VARIABLE IOS_SDK_VERSION
+  ERROR_QUIET
+  OUTPUT_STRIP_TRAILING_WHITESPACE)
+# Find the Developer root for the specific iOS platform being compiled for
+# from CMAKE_OSX_SYSROOT.  Should be ../../ from SDK specified in
+# CMAKE_OSX_SYSROOT.  There does not appear to be a direct way to obtain
+# this information from xcrun or xcodebuild.
+if (NOT CMAKE_IOS_DEVELOPER_ROOT)
+  get_filename_component(IOS_PLATFORM_SDK_DIR ${CMAKE_OSX_SYSROOT} PATH)
+  get_filename_component(CMAKE_IOS_DEVELOPER_ROOT ${IOS_PLATFORM_SDK_DIR} PATH)
+endif()
+if (NOT EXISTS ${CMAKE_IOS_DEVELOPER_ROOT})
+  message(FATAL_ERROR "Invalid CMAKE_IOS_DEVELOPER_ROOT: "
+    "${CMAKE_IOS_DEVELOPER_ROOT} does not exist.")
+endif()
+# Find the C & C++ compilers for the specified SDK.
+if (NOT CMAKE_C_COMPILER)
+  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang
+    OUTPUT_VARIABLE CMAKE_C_COMPILER
+    ERROR_QUIET
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+  message(STATUS "Using C compiler: ${CMAKE_C_COMPILER}")
+endif()
+if (NOT CMAKE_CXX_COMPILER)
+  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang++
+    OUTPUT_VARIABLE CMAKE_CXX_COMPILER
+    ERROR_QUIET
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+  message(STATUS "Using CXX compiler: ${CMAKE_CXX_COMPILER}")
+endif()
+# Find (Apple's) libtool.
+execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find libtool
+  OUTPUT_VARIABLE IOS_LIBTOOL
+  ERROR_QUIET
+  OUTPUT_STRIP_TRAILING_WHITESPACE)
+message(STATUS "Using libtool: ${IOS_LIBTOOL}")
+# Configure libtool to be used instead of ar + ranlib to build static libraries.
+# This is required on Xcode 7+, but should also work on previous versions of
+# Xcode.
+set(CMAKE_C_CREATE_STATIC_LIBRARY
+  "${IOS_LIBTOOL} -static -o <TARGET> <LINK_FLAGS> <OBJECTS> ")
+set(CMAKE_CXX_CREATE_STATIC_LIBRARY
+  "${IOS_LIBTOOL} -static -o <TARGET> <LINK_FLAGS> <OBJECTS> ")
+# Get the version of Darwin (OS X) of the host.
+execute_process(COMMAND uname -r
+  OUTPUT_VARIABLE CMAKE_HOST_SYSTEM_VERSION
+  ERROR_QUIET
+  OUTPUT_STRIP_TRAILING_WHITESPACE)
+# Standard settings.
+set(CMAKE_SYSTEM_NAME Darwin CACHE INTERNAL "")
+set(CMAKE_SYSTEM_VERSION ${IOS_SDK_VERSION} CACHE INTERNAL "")
+set(UNIX TRUE CACHE BOOL "")
+set(APPLE TRUE CACHE BOOL "")
+set(IOS TRUE CACHE BOOL "")
+set(CMAKE_AR ar CACHE FILEPATH "" FORCE)
+set(CMAKE_RANLIB ranlib CACHE FILEPATH "" FORCE)
+# Force unset of OS X-specific deployment target (otherwise autopopulated),
+# required as of cmake 2.8.10.
+set(CMAKE_OSX_DEPLOYMENT_TARGET "" CACHE STRING
+  "Must be empty for iOS builds." FORCE)
+# Set the architectures for which to build.
+set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE STRING "Build architecture for iOS")
+# Change the type of target generated for try_compile() so it'll work when cross-compiling
+set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
+# All iOS/Darwin specific settings - some may be redundant.
+set(CMAKE_SHARED_LIBRARY_PREFIX "lib")
+set(CMAKE_SHARED_LIBRARY_SUFFIX ".dylib")
+set(CMAKE_SHARED_MODULE_PREFIX "lib")
+set(CMAKE_SHARED_MODULE_SUFFIX ".so")
+set(CMAKE_C_COMPILER_ABI ELF)
+set(CMAKE_CXX_COMPILER_ABI ELF)
+set(CMAKE_C_HAS_ISYSROOT 1)
+set(CMAKE_CXX_HAS_ISYSROOT 1)
+set(CMAKE_MODULE_EXISTS 1)
+set(CMAKE_DL_LIBS "")
+set(CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG "-compatibility_version ")
+set(CMAKE_C_OSX_CURRENT_VERSION_FLAG "-current_version ")
+set(CMAKE_CXX_OSX_COMPATIBILITY_VERSION_FLAG "${CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG}")
+set(CMAKE_CXX_OSX_CURRENT_VERSION_FLAG "${CMAKE_C_OSX_CURRENT_VERSION_FLAG}")
+
+if(IOS_ARCH MATCHES "((^|, )(arm64|arm64e|x86_64))+")
+  set(CMAKE_C_SIZEOF_DATA_PTR 8)
+  set(CMAKE_CXX_SIZEOF_DATA_PTR 8)
+  message(STATUS "Using a data_ptr size of 8")
+else()
+  set(CMAKE_C_SIZEOF_DATA_PTR 4)
+  set(CMAKE_CXX_SIZEOF_DATA_PTR 4)
+  message(STATUS "Using a data_ptr size of 4")
+endif()
+
+message(STATUS "Building for minimum iOS version: ${IOS_DEPLOYMENT_TARGET}"
+               " (SDK version: ${IOS_SDK_VERSION})")
+# Note that only Xcode 7+ supports the newer more specific:
+# -m${XCODE_IOS_PLATFORM}-version-min flags, older versions of Xcode use:
+# -m(ios/ios-simulator)-version-min instead.
+if (IOS_PLATFORM STREQUAL "OS" OR IOS_PLATFORM STREQUAL "OS64")
+  if (XCODE_VERSION VERSION_LESS 7.0)
+    set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+      "-mios-version-min=${IOS_DEPLOYMENT_TARGET}")
+  else()
+    # Xcode 7.0+ uses flags we can build directly from XCODE_IOS_PLATFORM.
+    set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+      "-m${XCODE_IOS_PLATFORM}-version-min=${IOS_DEPLOYMENT_TARGET}")
+  endif()
+elseif (IOS_PLATFORM STREQUAL "TVOS")
+  set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+    "-mtvos-version-min=${IOS_DEPLOYMENT_TARGET}")
+elseif (IOS_PLATFORM STREQUAL "SIMULATOR_TVOS")
+  set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+    "-mtvos-simulator-version-min=${IOS_DEPLOYMENT_TARGET}")
+elseif (IOS_PLATFORM STREQUAL "WATCHOS")
+  set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+    "-mwatchos-version-min=${IOS_DEPLOYMENT_TARGET}")
+elseif (IOS_PLATFORM STREQUAL "SIMULATOR_WATCHOS")
+  set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+    "-mwatchos-simulator-version-min=${IOS_DEPLOYMENT_TARGET}")
+else()
+  # SIMULATOR or SIMULATOR64 both use -mios-simulator-version-min.
+  set(XCODE_IOS_PLATFORM_VERSION_FLAGS
+    "-mios-simulator-version-min=${IOS_DEPLOYMENT_TARGET}")
+endif()
+message(STATUS "Version flags set to: ${XCODE_IOS_PLATFORM_VERSION_FLAGS}")
+
+if (ENABLE_BITCODE)
+  set(BITCODE "-fembed-bitcode")
+  set(HEADER_PAD "")
+  message(STATUS "Enabling bitcode support.")
+else()
+  set(BITCODE "")
+  set(HEADER_PAD "-headerpad_max_install_names")
+  message(STATUS "Disabling bitcode support.")
+endif()
+
+if (ENABLE_ARC)
+  set(FOBJC_ARC "-fobjc-arc")
+  message(STATUS "Enabling ARC support.")
+else()
+  set(FOBJC_ARC "-fno-objc-arc")
+  message(STATUS "Disabling ARC support.")
+endif()
+
+if (NOT ENABLE_VISIBILITY)
+  set(VISIBILITY "-fvisibility=hidden")
+  message(STATUS "Hiding symbols (-fvisibility=hidden).")
+else()
+  set(VISIBILITY "")
+endif()
+
+set(CMAKE_C_FLAGS
+"${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${BITCODE} -fobjc-abi-version=2 ${FOBJC_ARC} ${CMAKE_C_FLAGS}")
+# Hidden visibilty is required for C++ on iOS.
+set(CMAKE_CXX_FLAGS
+"${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${BITCODE} ${VISIBILITY} -fvisibility-inlines-hidden -fobjc-abi-version=2 ${FOBJC_ARC} ${CMAKE_CXX_FLAGS}")
+set(CMAKE_CXX_FLAGS_MINSIZEREL "${CMAKE_CXX_FLAGS} -DNDEBUG -Os -ffast-math ${BITCODE} ${CMAKE_CXX_FLAGS_MINSIZEREL}")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS} -DNDEBUG -O2 -g -ffast-math ${BITCODE} ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} -DNDEBUG -O3 -ffast-math ${BITCODE} ${CMAKE_CXX_FLAGS_RELEASE}")
+set(CMAKE_C_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first ${CMAKE_C_LINK_FLAGS}")
+set(CMAKE_CXX_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS}  -Wl,-search_paths_first ${CMAKE_CXX_LINK_FLAGS}")
+
+# In order to ensure that the updated compiler flags are used in try_compile()
+# tests, we have to forcibly set them in the CMake cache, not merely set them
+# in the local scope.
+list(APPEND VARS_TO_FORCE_IN_CACHE
+  CMAKE_C_FLAGS
+  CMAKE_CXX_FLAGS
+  CMAKE_CXX_FLAGS_RELWITHDEBINFO
+  CMAKE_CXX_FLAGS_MINSIZEREL
+  CMAKE_CXX_FLAGS_RELEASE
+  CMAKE_C_LINK_FLAGS
+  CMAKE_CXX_LINK_FLAGS)
+foreach(VAR_TO_FORCE ${VARS_TO_FORCE_IN_CACHE})
+  set(${VAR_TO_FORCE} "${${VAR_TO_FORCE}}" CACHE STRING "")
+endforeach()
+
+set(CMAKE_PLATFORM_HAS_INSTALLNAME 1)
+set (CMAKE_SHARED_LINKER_FLAGS "-rpath @executable_path/Frameworks -rpath @loader_path/Frameworks")
+set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib ${HEADER_PAD}")
+set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle ${HEADER_PAD}")
+set(CMAKE_SHARED_MODULE_LOADER_C_FLAG "-Wl,-bundle_loader,")
+set(CMAKE_SHARED_MODULE_LOADER_CXX_FLAG "-Wl,-bundle_loader,")
+set(CMAKE_FIND_LIBRARY_SUFFIXES ".dylib" ".so" ".a")
+
+# Hack: if a new cmake (which uses CMAKE_INSTALL_NAME_TOOL) runs on an old
+# build tree (where install_name_tool was hardcoded) and where
+# CMAKE_INSTALL_NAME_TOOL isn't in the cache and still cmake didn't fail in
+# CMakeFindBinUtils.cmake (because it isn't rerun) hardcode
+# CMAKE_INSTALL_NAME_TOOL here to install_name_tool, so it behaves as it did
+# before, Alex.
+if (NOT DEFINED CMAKE_INSTALL_NAME_TOOL)
+  find_program(CMAKE_INSTALL_NAME_TOOL install_name_tool)
+endif (NOT DEFINED CMAKE_INSTALL_NAME_TOOL)
+
+# Set the find root to the iOS developer roots and to user defined paths.
+set(CMAKE_FIND_ROOT_PATH ${CMAKE_IOS_DEVELOPER_ROOT} ${CMAKE_OSX_SYSROOT}
+  ${CMAKE_PREFIX_PATH} CACHE string  "iOS find search path root" FORCE)
+# Default to searching for frameworks first.
+set(CMAKE_FIND_FRAMEWORK FIRST)
+# Set up the default search directories for frameworks.
+set(CMAKE_SYSTEM_FRAMEWORK_PATH
+  ${CMAKE_OSX_SYSROOT}/System/Library/Frameworks
+  ${CMAKE_OSX_SYSROOT}/System/Library/PrivateFrameworks
+  ${CMAKE_OSX_SYSROOT}/Developer/Library/Frameworks)
+# Only search the specified iOS SDK, not the remainder of the host filesystem.
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+# This little macro lets you set any XCode specific property.
+macro(set_xcode_property TARGET XCODE_PROPERTY XCODE_VALUE XCODE_RELVERSION)
+  set(XCODE_RELVERSION_I "${XCODE_RELVERSION}")
+  if (XCODE_RELVERSION_I STREQUAL "All")
+    set_property(TARGET ${TARGET} PROPERTY
+    XCODE_ATTRIBUTE_${XCODE_PROPERTY} "${XCODE_VALUE}")
+  else()
+    set_property(TARGET ${TARGET} PROPERTY
+    XCODE_ATTRIBUTE_${XCODE_PROPERTY}[variant=${XCODE_RELVERSION_I}] "${XCODE_VALUE}")
+  endif()
+endmacro(set_xcode_property)
+# This macro lets you find executable programs on the host system.
+macro(find_host_package)
+  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER)
+  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER)
+  set(IOS FALSE)
+  find_package(${ARGN})
+  set(IOS TRUE)
+  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY)
+  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+endmacro(find_host_package)
diff --git a/cpp-package/inspireface/tools/generate_error_tabel_to_python.py b/cpp-package/inspireface/tools/generate_error_tabel_to_python.py
new file mode 100644
index 000000000..42e9aa418
--- /dev/null
+++ b/cpp-package/inspireface/tools/generate_error_tabel_to_python.py
@@ -0,0 +1,109 @@
+import click
+import re
+from datetime import datetime
+
+# Function to calculate the actual error code value based on the expressions
+def calculate_error_code_value(error_code_str, error_definitions):
+    try:
+        # Replace the hex values and error definitions with actual values
+        error_code_str = re.sub(r'0X([0-9A-F]+)', lambda m: str(int(m.group(1), 16)), error_code_str)
+        for error_name, error_value in error_definitions.items():
+            error_code_str = error_code_str.replace(error_name, str(error_value))
+        # Evaluate the expression to get the actual error code value
+        return eval(error_code_str)
+    except Exception as e:
+        # If there is an error in evaluation (like undefined reference), return the original string
+        return error_code_str
+
+# Define a function to parse the error codes between anchors in a given C header file content
+def parse_and_calculate_error_codes(header_content):
+    # Define the start and end anchor tags
+    start_anchor = '// [Anchor-Begin]'
+    end_anchor = '// [Anchor-End]'
+
+    # Find the index of start and end anchor tags
+    start_index = header_content.find(start_anchor)
+    end_index = header_content.find(end_anchor, start_index)
+
+    # Extract the content between anchors
+    anchor_content = header_content[start_index+len(start_anchor):end_index].strip()
+
+    # Split the content into lines
+    lines = anchor_content.split('\n')
+
+    # Dictionary to hold the error names and their evaluated values
+    error_definitions = {}
+    # List to hold the parsed error codes
+    error_codes = []
+
+    # Process each line
+    for line in lines:
+        line = line.strip()
+        # Skip empty lines and lines that don't start with #define
+        if not line or not line.startswith('#define'):
+            continue
+            
+        # Split by comment marker and handle cases with no comment
+        parts = line.split('//')
+        define_part = parts[0].strip()
+        comment_part = parts[1].strip() if len(parts) > 1 else ''
+
+        # Extract the error name and code, handling parentheses
+        define_parts = define_part.split(None, 2)  # Split into max 3 parts
+        if len(define_parts) >= 2:  # Changed from == 3
+            error_name = define_parts[1]
+            error_code_str = define_parts[2].strip('()') if len(define_parts) > 2 else ''
+
+            # Calculate the actual error code value
+            error_code_value = calculate_error_code_value(error_code_str, error_definitions)
+
+            # Store the calculated error code value for later reference
+            error_definitions[error_name] = error_code_value
+
+            # Skip base offset definitions (ending with _BASE)
+            if error_name.endswith('_BASE'):
+                continue
+
+            # Append the extracted information to the error_codes list
+            error_codes.append((error_name, error_code_value, comment_part))
+
+    return error_codes
+
+# Click command for processing the header file and outputting Python error table
+@click.command()
+@click.argument('header_path', type=click.Path(exists=True, file_okay=True, dir_okay=False, readable=True))
+@click.argument('output_path', type=click.Path(file_okay=True, dir_okay=False, writable=True))
+def process_header(header_path, output_path):
+    # Read the header file content
+    with open(header_path, 'r', encoding='utf-8') as file:
+        header_content = file.read()
+
+    # Parse and calculate the error codes from the header content
+    parsed_error_codes = parse_and_calculate_error_codes(header_content)
+
+    # Generate Python file content
+    python_content = f'''"""
+InspireFace Error Codes
+
+Auto-generated error code definitions from {header_path}
+Generated on: {datetime.now().strftime("%Y-%m-%d")}
+"""
+
+'''
+
+    # Generate simple error code definitions
+    for name, code, comment in parsed_error_codes:
+        if comment:
+            python_content += f"{name} = {code}  # {comment}\n"
+        else:
+            python_content += f"{name} = {code}\n"
+
+    # Write the Python file
+    with open(output_path, 'w', encoding='utf-8') as py_file:
+        py_file.write(python_content)
+
+    click.echo(f"Python error table has been written to {output_path}")
+    click.echo(f"Generated {len(parsed_error_codes)} error code definitions")
+
+if __name__ == '__main__':
+    process_header()
diff --git a/cpp-package/inspireface/tools/generate_release_models_info.py b/cpp-package/inspireface/tools/generate_release_models_info.py
new file mode 100644
index 000000000..738aebf29
--- /dev/null
+++ b/cpp-package/inspireface/tools/generate_release_models_info.py
@@ -0,0 +1,49 @@
+import hashlib
+import os
+import json
+import click
+
+need_models = [
+    "Pikachu",
+    "Megatron",
+    "Megatron_TRT",
+    "Gundam_RK356X",
+    "Gundam_RK3588",
+]
+
+def get_file_hash_sha256(file_path):
+    sha256 = hashlib.sha256()
+    with open(file_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096), b''):
+            sha256.update(chunk)
+    return sha256.hexdigest()
+
+@click.command()
+@click.argument('model_dir', default="test_res/pack/")
+@click.option('--models', '-m', multiple=True, help='Specify the model name to process')
+def main(model_dir, models):
+    model_info = {}
+    
+    # If no model is specified, process all predefined models
+    models_to_process = models if models else need_models
+    
+    for file in models_to_process:
+        file_path = os.path.join(model_dir, file)
+        if os.path.exists(file_path):
+            md5 = get_file_hash_sha256(file_path)
+            model_info[file] = {
+                "url": f"https://inspireface-1259028827.cos.ap-singapore.myqcloud.com/inspireface_modelzoo/t4/{file}",
+                "filename": file,
+                "md5": md5
+            }
+        else:
+            print(f"Warning: File {file_path} does not exist")
+    
+    print("\033[33mNeed to modify the python/inspireface/modules/utils/resource.py, changes to the information release\033[0m")
+    # Output the result
+    print(json.dumps(model_info, indent=4, ensure_ascii=False))
+
+if __name__ == "__main__":
+    main()
+
+
diff --git a/cpp-package/inspireface/tools/get_model_md5.py b/cpp-package/inspireface/tools/get_model_md5.py
new file mode 100644
index 000000000..57045df44
--- /dev/null
+++ b/cpp-package/inspireface/tools/get_model_md5.py
@@ -0,0 +1,27 @@
+import hashlib
+import os
+import click
+
+file_list = [
+    "Pikachu",
+    "Megatron",
+    "Megatron_TRT",
+    "Gundam_RK356X",
+    "Gundam_RK3588",
+]
+
+def get_file_hash_sha256(file_path):
+    sha256 = hashlib.sha256()
+    with open(file_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096), b''):
+            sha256.update(chunk)
+    return sha256.hexdigest()
+
+@click.command()
+@click.argument('model_dir', default="test_res/pack/")
+def main(model_dir):
+    for file in file_list:
+        print(f"{file}: {get_file_hash_sha256(os.path.join(model_dir, file))}")
+
+if __name__ == "__main__":
+    main()
diff --git a/cpp-package/inspireface/tools/inspire_archive/archive_packing.py b/cpp-package/inspireface/tools/inspire_archive/archive_packing.py
new file mode 100644
index 000000000..8da8d9031
--- /dev/null
+++ b/cpp-package/inspireface/tools/inspire_archive/archive_packing.py
@@ -0,0 +1,47 @@
+import os
+import tarfile
+import click
+import tqdm
+
+def remove_suffix(filename):
+    """Remove the file suffix."""
+    return os.path.splitext(filename)[0]
+
+@click.command()
+@click.argument('folder_path')
+@click.argument('output_filename')
+@click.option('--rm-suffix', is_flag=True, default=True, help='Remove file suffixes in the archive.')
+def make_tar(folder_path, output_filename, rm_suffix):
+    """
+    Package the specified folder into a tar file. FOLDER_PATH is the path to the folder to be packaged,
+    and OUTPUT_FILENAME is the name of the output tar file.
+    """
+    # Collect all files to be packed for printing their info before packaging
+    files_to_pack = []
+    for root, dirs, files in os.walk(folder_path):
+        for file in files:
+            file_path = os.path.join(root, file)
+            arcname = os.path.relpath(file_path, start=folder_path)
+            if rm_suffix:
+                # Split the file name and directory name
+                dirs, filename = os.path.split(arcname)
+                # Remove the file suffix
+                filename = remove_suffix(filename)
+                # Reassemble the directory and the processed file name
+                arcname = os.path.join(dirs, filename)
+            files_to_pack.append((file_path, arcname))
+
+    # Print file information before packing
+    print("Printing file information before packaging...")
+    for file_path, arcname in files_to_pack:
+        file_size_kb = os.path.getsize(file_path) / 1024  # Convert size to kilobytes
+        print(f"File: {arcname}, Size: {file_size_kb:.2f} KB")
+
+    # Package the files
+    with tarfile.open(output_filename, "w") as tar:
+        print("Packing....")
+        for file_path, arcname in tqdm.tqdm(files_to_pack):
+            tar.add(file_path, arcname=arcname)
+
+if __name__ == '__main__':
+    make_tar()
diff --git a/cpp-package/inspireface/tools/output_error_table.py b/cpp-package/inspireface/tools/output_error_table.py
new file mode 100644
index 000000000..b33574ba0
--- /dev/null
+++ b/cpp-package/inspireface/tools/output_error_table.py
@@ -0,0 +1,107 @@
+import click
+import re
+
+# Function to calculate the actual error code value based on the expressions
+def calculate_error_code_value(error_code_str, error_definitions):
+    try:
+        # Replace the hex values and error definitions with actual values
+        error_code_str = re.sub(r'0X([0-9A-F]+)', lambda m: str(int(m.group(1), 16)), error_code_str)
+        for error_name, error_value in error_definitions.items():
+            error_code_str = error_code_str.replace(error_name, str(error_value))
+        # Evaluate the expression to get the actual error code value
+        return eval(error_code_str)
+    except Exception as e:
+        # If there is an error in evaluation (like undefined reference), return the original string
+        return error_code_str
+
+# Define a function to parse the error codes between anchors in a given C header file content
+def parse_and_calculate_error_codes(header_content):
+    # Define the start and end anchor tags
+    start_anchor = '// [Anchor-Begin]'
+    end_anchor = '// [Anchor-End]'
+
+    # Find the index of start and end anchor tags
+    start_index = header_content.find(start_anchor)
+    end_index = header_content.find(end_anchor, start_index)
+
+    # Extract the content between anchors
+    anchor_content = header_content[start_index+len(start_anchor):end_index].strip()
+
+    # Split the content into lines
+    lines = anchor_content.split('\n')
+
+    # Dictionary to hold the error names and their evaluated values
+    error_definitions = {}
+    # List to hold the parsed error codes
+    error_codes = []
+
+    # Process each line
+    for line in lines:
+        line = line.strip()
+        # Skip empty lines and lines that don't start with #define
+        if not line or not line.startswith('#define'):
+            continue
+            
+        # Split by comment marker and handle cases with no comment
+        parts = line.split('//')
+        define_part = parts[0].strip()
+        comment_part = parts[1].strip() if len(parts) > 1 else ''
+
+        # Extract the error name and code, handling parentheses
+        define_parts = define_part.split(None, 2)  # Split into max 3 parts
+        if len(define_parts) >= 2:  # Changed from == 3
+            error_name = define_parts[1]
+            error_code_str = define_parts[2].strip('()') if len(define_parts) > 2 else ''
+
+            # Calculate the actual error code value
+            error_code_value = calculate_error_code_value(error_code_str, error_definitions)
+
+            # Store the calculated error code value for later reference
+            error_definitions[error_name] = error_code_value
+
+            # Skip base offset definitions (ending with _BASE)
+            if error_name.endswith('_BASE'):
+                continue
+
+            # Append the extracted information to the error_codes list
+            error_codes.append((error_name, error_code_value, comment_part))
+
+    return error_codes
+
+# Click command for processing the header file and outputting Markdown table
+@click.command()
+@click.argument('header_path', type=click.Path(exists=True, file_okay=True, dir_okay=False, readable=True))
+@click.argument('output_path', type=click.Path(file_okay=True, dir_okay=False, writable=True))
+def process_header(header_path, output_path):
+    # Read the header file content
+    with open(header_path, 'r', encoding='utf-8') as file:
+        header_content = file.read()
+
+    # Parse and calculate the error codes from the header content
+    parsed_error_codes = parse_and_calculate_error_codes(header_content)
+
+    md_table = """# Error Feedback Codes
+
+During the use of InspireFace, some error feedback codes may be generated. Here is a table of error feedback codes.
+
+- As of **June 15, 2025**, the error code definitions have been restructured. Some legacy codes from historical versions have been removed, and a more streamlined version has been reorganized and consolidated.
+
+"""
+
+    # Prepare the Markdown table header
+    md_table += " | Index | Name | Code | Comment | \n"
+    md_table += " | --- | --- | --- | --- | \n"
+
+    # Fill the Markdown table with parsed error codes
+    for index, (name, code, comment) in enumerate(parsed_error_codes, start=1):
+        md_table += f" | {index} | {name} | {code} | {comment} | \n"
+
+    # Write the Markdown table to the output file
+    with open(output_path, 'w', encoding='utf-8') as md_file:
+        print(md_table)
+        md_file.write(md_table)
+
+    click.echo(f"Markdown table of error codes has been written to {output_path}")
+
+if __name__ == '__main__':
+    process_header()
diff --git a/cpp-package/inspireface/tools/release_modelscope.py b/cpp-package/inspireface/tools/release_modelscope.py
new file mode 100644
index 000000000..ec00a1f4b
--- /dev/null
+++ b/cpp-package/inspireface/tools/release_modelscope.py
@@ -0,0 +1,53 @@
+import os
+import click
+from modelscope.hub.api import HubApi
+
+
+@click.command()
+@click.option(
+    '--model-id', 
+    default='tunmxy/InspireFace',
+    help='ModelScope model ID'
+)
+@click.option(
+    '--model-dir',
+    required=True,
+    type=click.Path(exists=True, file_okay=False, dir_okay=True),
+    help='Local model directory path'
+)
+@click.option(
+    '--token',
+    help='ModelScope access token, if not provided will get from MODELSCOPE_TOKEN environment variable'
+)
+def upload_model(model_id, model_dir, token):
+    """Upload model to ModelScope"""
+    
+    # Get token
+    if not token:
+        token = os.getenv('MODELSCOPE_TOKEN')
+        if not token:
+            click.echo("Error: No token provided and MODELSCOPE_TOKEN environment variable not set", err=True)
+            raise click.Abort()
+    
+    try:
+        # Login and upload
+        api = HubApi()
+        api.login(token)
+        
+        click.echo(f"Starting to upload model to {model_id}...")
+        click.echo(f"Local directory: {model_dir}")
+        
+        api.push_model(
+            model_id=model_id,
+            model_dir=model_dir
+        )
+        
+        click.echo("✅ Model uploaded successfully!")
+        
+    except Exception as e:
+        click.echo(f"❌ Upload failed: {e}", err=True)
+        raise click.Abort()
+
+
+if __name__ == '__main__':
+    upload_model()
\ No newline at end of file
diff --git a/cpp-package/inspireface/tools/transform_similarity_testing.py b/cpp-package/inspireface/tools/transform_similarity_testing.py
new file mode 100644
index 000000000..a38162a02
--- /dev/null
+++ b/cpp-package/inspireface/tools/transform_similarity_testing.py
@@ -0,0 +1,104 @@
+import numpy as np
+import matplotlib.pyplot as plt
+
+
+class SimilarityConverter:
+    def __init__(self, 
+                 threshold=0.48, 
+                 middle_score=0.6, 
+                 steepness=8.0, 
+                 output_range=(0.01, 0.99)):
+        self.threshold = threshold
+        self.middle_score = middle_score
+        self.steepness = steepness
+        self.output_min, self.output_max = output_range
+        self.output_scale = self.output_max - self.output_min
+        
+        self.bias = -np.log((self.output_max - self.middle_score)/(self.middle_score - self.output_min))
+        
+    def convert(self, cosine):
+        shifted_input = self.steepness * (cosine - self.threshold)
+        sigmoid = 1 / (1 + np.exp(-shifted_input - self.bias))
+        return sigmoid * self.output_scale + self.output_min
+    
+    def plot(self, test_points=None):
+        x = np.linspace(-1, 1, 200)
+        y = self.convert(x)
+        
+        plt.figure(figsize=(12, 6))
+        plt.plot(x, y, '-', label='Conversion Function', color='blue')
+        
+        if test_points is not None:
+            test_scores = self.convert(test_points)
+            plt.plot(test_points, test_scores, 'ro', label='Test Points')
+            
+            for point, score in zip(test_points, test_scores):
+                plt.annotate(f'({point:.2f}, {score:.3f})', 
+                            xy=(point, score), 
+                            xytext=(10, 10),
+                            textcoords='offset points')
+        
+        plt.axvline(x=self.threshold, color='gray', linestyle='--', alpha=0.5, 
+                   label=f'Threshold ({self.threshold:.2f})')
+        plt.axhline(y=self.middle_score, color='gray', linestyle='--', alpha=0.5)
+        
+        plt.plot(self.threshold, self.middle_score, 'r*', markersize=15, 
+                label='Passing Point', color='red', zorder=5)
+        
+        plt.annotate(f'Passing Point\n({self.threshold:.2f}, {self.middle_score:.2f})',
+                    xy=(self.threshold, self.middle_score),
+                    xytext=(30, 20),
+                    textcoords='offset points',
+                    bbox=dict(boxstyle='round,pad=0.5', fc='yellow', alpha=0.5),
+                    arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0', 
+                                  color='red', lw=2))
+        
+        
+        plt.fill_between(x, y, 0, 
+                        where=(x >= self.threshold),
+                        alpha=0.1, color='green',
+                        label='Passing Region')
+        
+        plt.grid(True, alpha=0.3)
+        plt.xlabel('Cosine Similarity')
+        plt.ylabel('Converted Score')
+        plt.title('Similarity Score Conversion')
+        plt.legend()
+        plt.xlim(-1.1, 1.1)
+        plt.ylim(-0.1, 1.1)
+        plt.tight_layout()
+        plt.show()
+
+if __name__ == "__main__":
+    # Set the threshold to 0.42
+    converter1 = SimilarityConverter(
+        threshold=0.42,
+        middle_score=0.6,
+        steepness=8.0,
+        output_range=(0.01, 1.0)
+    )
+    test_points1 = np.array([-0.8, -0.2, 0.02, 0.1, 0.25, 0.3, 0.48, 0.7, 0.8, 0.9, 1.0])
+    print("\nModel 1 (Threshold = 0.48):")
+    converter1.plot(test_points1)
+    
+    # Set the threshold to 0.32
+    converter2 = SimilarityConverter(
+        threshold=0.32,
+        middle_score=0.6,
+        steepness=10.0,
+        output_range=(0.02, 1.0)
+    )
+    test_points2 = np.array([-0.8, -0.2, 0.02, 0.1, 0.25,0.32, 0.5, 0.7, 0.8, 0.9, 1.0])
+    print("\nModel 2 (Threshold = 0.32):")
+    converter2.plot(test_points2)
+    
+    # Print the results
+    print("\nTest Results for Model 1 (threshold=0.48):")
+    for point in test_points1:
+        score = converter1.convert(point)
+        print(f"Cosine: {point:6.2f} -> Score: {score:.4f}")
+        
+    print("\nTest Results for Model 2 (threshold=0.32):")
+    for point in test_points2:
+        score = converter2.convert(point)
+        print(f"Cosine: {point:6.2f} -> Score: {score:.4f}")
\ No newline at end of file
diff --git a/datasets/README.md b/datasets/README.md
deleted file mode 100644
index 828aedd8e..000000000
--- a/datasets/README.md
+++ /dev/null
@@ -1 +0,0 @@
-Put datasets here.
diff --git a/deploy/benchmark.py b/deploy/benchmark.py
deleted file mode 100644
index 42773b07f..000000000
--- a/deploy/benchmark.py
+++ /dev/null
@@ -1,26 +0,0 @@
-import face_embedding
-import argparse
-import cv2
-import numpy as np
-import datetime
-
-parser = argparse.ArgumentParser(description='face model test')
-# general
-parser.add_argument('--image-size', default='112,112', help='')
-parser.add_argument('--model', default='../models/model-r34-amf/model,0', help='path to load model.')
-parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-parser.add_argument('--det', default=2, type=int, help='mtcnn option, 2 means using R+O, else using O')
-parser.add_argument('--flip', default=0, type=int, help='whether do lr flip aug')
-parser.add_argument('--threshold', default=1.24, type=float, help='ver dist threshold')
-args = parser.parse_args()
-
-model = face_embedding.FaceModel(args)
-#img = cv2.imread('/raid5data/dplearn/lfw/Jude_Law/Jude_Law_0001.jpg')
-img = cv2.imread('/raid5data/dplearn/megaface/facescrubr/112x112/Tom_Hanks/Tom_Hanks_54745.png')
-
-time_now = datetime.datetime.now()
-for i in range(3000):
-  f1 = model.get_feature(img)
-time_now2 = datetime.datetime.now()
-diff = time_now2 - time_now
-print(diff.total_seconds()/3000)
diff --git a/deploy/convert_onnx.py b/deploy/convert_onnx.py
deleted file mode 100644
index 75d4f674f..000000000
--- a/deploy/convert_onnx.py
+++ /dev/null
@@ -1,31 +0,0 @@
-import sys
-import os
-import argparse
-import onnx
-import mxnet as mx
-
-print('mxnet version:', mx.__version__)
-print('onnx version:', onnx.__version__)
-#make sure to install onnx-1.2.1
-#pip uninstall onnx
-#pip install onnx==1.2.1
-assert onnx.__version__=='1.2.1'
-import numpy as np
-from mxnet.contrib import onnx as onnx_mxnet
-
-parser = argparse.ArgumentParser(description='convert insightface models to onnx')
-# general
-parser.add_argument('--prefix', default='./r100-arcface/model', help='prefix to load model.')
-parser.add_argument('--epoch', default=0, type=int, help='epoch number to load model.')
-parser.add_argument('--input-shape', default='3,112,112', help='input shape.')
-parser.add_argument('--output-onnx', default='./r100.onnx', help='path to write onnx model.')
-args = parser.parse_args()
-input_shape = (1,) + tuple( [int(x) for x in args.input_shape.split(',')] )
-print('input-shape:', input_shape)
-
-sym_file = "%s-symbol.json"%args.prefix
-params_file = "%s-%04d.params"%(args.prefix, args.epoch)
-assert os.path.exists(sym_file)
-assert os.path.exists(params_file)
-converted_model_path = onnx_mxnet.export_model(sym_file, params_file, [input_shape], np.float32, args.output_onnx)
-
diff --git a/deploy/face_embedding.py b/deploy/face_embedding.py
deleted file mode 100644
index ae2dd1c07..000000000
--- a/deploy/face_embedding.py
+++ /dev/null
@@ -1,94 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-import mxnet as mx
-import random
-import cv2
-import sklearn
-from sklearn.decomposition import PCA
-from time import sleep
-from easydict import EasyDict as edict
-from mtcnn_detector import MtcnnDetector
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'src', 'common'))
-import face_image
-import face_preprocess
-
-
-def do_flip(data):
-  for idx in range(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-class FaceModel:
-  def __init__(self, args):
-    self.args = args
-    model = edict()
-
-    self.threshold = args.threshold
-    self.det_minsize = 50
-    self.det_threshold = [0.4,0.6,0.6]
-    self.det_factor = 0.9
-    _vec = args.image_size.split(',')
-    assert len(_vec)==2
-    image_size = (int(_vec[0]), int(_vec[1]))
-    self.image_size = image_size
-    _vec = args.model.split(',')
-    assert len(_vec)==2
-    prefix = _vec[0]
-    epoch = int(_vec[1])
-    print('loading',prefix, epoch)
-    ctx = mx.gpu(args.gpu)
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    self.model = model
-    mtcnn_path = os.path.join(os.path.dirname(__file__), 'mtcnn-model')
-    detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=[0.0,0.0,0.2])
-    self.detector = detector
-
-
-  def get_feature(self, face_img):
-    #face_img is bgr image
-    ret = self.detector.detect_face_limited(face_img, det_type = self.args.det)
-    if ret is None:
-      return None
-    bbox, points = ret
-    if bbox.shape[0]==0:
-      return None
-    bbox = bbox[0,0:4]
-    points = points[0,:].reshape((2,5)).T
-    #print(bbox)
-    #print(points)
-    nimg = face_preprocess.preprocess(face_img, bbox, points, image_size='112,112')
-    nimg = cv2.cvtColor(nimg, cv2.COLOR_BGR2RGB)
-    aligned = np.transpose(nimg, (2,0,1))
-    #print(nimg.shape)
-    embedding = None
-    for flipid in [0,1]:
-      if flipid==1:
-        if self.args.flip==0:
-          break
-        do_flip(aligned)
-      input_blob = np.expand_dims(aligned, axis=0)
-      data = mx.nd.array(input_blob)
-      db = mx.io.DataBatch(data=(data,))
-      self.model.forward(db, is_train=False)
-      _embedding = self.model.get_outputs()[0].asnumpy()
-      #print(_embedding.shape)
-      if embedding is None:
-        embedding = _embedding
-      else:
-        embedding += _embedding
-    embedding = sklearn.preprocessing.normalize(embedding).flatten()
-    return embedding
-
diff --git a/deploy/face_model.py b/deploy/face_model.py
deleted file mode 100644
index 6aee02f64..000000000
--- a/deploy/face_model.py
+++ /dev/null
@@ -1,108 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-#import tensorflow as tf
-import numpy as np
-import mxnet as mx
-import random
-import cv2
-import sklearn
-from sklearn.decomposition import PCA
-from time import sleep
-from easydict import EasyDict as edict
-from mtcnn_detector import MtcnnDetector
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'src', 'common'))
-import face_image
-import face_preprocess
-
-
-def do_flip(data):
-  for idx in range(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_model(ctx, image_size, model_str, layer):
-  _vec = model_str.split(',')
-  assert len(_vec)==2
-  prefix = _vec[0]
-  epoch = int(_vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers[layer+'_output']
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-  model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  return model
-
-class FaceModel:
-  def __init__(self, args):
-    self.args = args
-    ctx = mx.gpu(args.gpu)
-    _vec = args.image_size.split(',')
-    assert len(_vec)==2
-    image_size = (int(_vec[0]), int(_vec[1]))
-    self.model = None
-    self.ga_model = None
-    if len(args.model)>0:
-      self.model = get_model(ctx, image_size, args.model, 'fc1')
-    if len(args.ga_model)>0:
-      self.ga_model = get_model(ctx, image_size, args.ga_model, 'fc1')
-
-    self.threshold = args.threshold
-    self.det_minsize = 50
-    self.det_threshold = [0.6,0.7,0.8]
-    #self.det_factor = 0.9
-    self.image_size = image_size
-    mtcnn_path = os.path.join(os.path.dirname(__file__), 'mtcnn-model')
-    if args.det==0:
-      detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=self.det_threshold)
-    else:
-      detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=[0.0,0.0,0.2])
-    self.detector = detector
-
-
-  def get_input(self, face_img):
-    ret = self.detector.detect_face(face_img, det_type = self.args.det)
-    if ret is None:
-      return None
-    bbox, points = ret
-    if bbox.shape[0]==0:
-      return None
-    bbox = bbox[0,0:4]
-    points = points[0,:].reshape((2,5)).T
-    #print(bbox)
-    #print(points)
-    nimg = face_preprocess.preprocess(face_img, bbox, points, image_size='112,112')
-    nimg = cv2.cvtColor(nimg, cv2.COLOR_BGR2RGB)
-    aligned = np.transpose(nimg, (2,0,1))
-    return aligned
-
-  def get_feature(self, aligned):
-    input_blob = np.expand_dims(aligned, axis=0)
-    data = mx.nd.array(input_blob)
-    db = mx.io.DataBatch(data=(data,))
-    self.model.forward(db, is_train=False)
-    embedding = self.model.get_outputs()[0].asnumpy()
-    embedding = sklearn.preprocessing.normalize(embedding).flatten()
-    return embedding
-
-  def get_ga(self, aligned):
-    input_blob = np.expand_dims(aligned, axis=0)
-    data = mx.nd.array(input_blob)
-    db = mx.io.DataBatch(data=(data,))
-    self.ga_model.forward(db, is_train=False)
-    ret = self.ga_model.get_outputs()[0].asnumpy()
-    g = ret[:,0:2].flatten()
-    gender = np.argmax(g)
-    a = ret[:,2:202].reshape( (100,2) )
-    a = np.argmax(a, axis=1)
-    age = int(sum(a))
-
-    return gender, age
-
diff --git a/deploy/ga_merge.py b/deploy/ga_merge.py
deleted file mode 100644
index d6397f0df..000000000
--- a/deploy/ga_merge.py
+++ /dev/null
@@ -1,52 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import sys
-import os
-import argparse
-import numpy as np
-import mxnet as mx
-
-parser = argparse.ArgumentParser(description='merge age and gender models')
-# general
-parser.add_argument('--age-model', default='', help='path to load age model.')
-parser.add_argument('--gender-model', default='', help='path to load gender model.')
-parser.add_argument('--prefix', default='', help='path to save model.')
-args = parser.parse_args()
-
-i = 0
-tsym = None
-targ = {}
-taux = {}
-for model in [args.age_model, args.gender_model]:
-  _vec = model.split(',')
-  assert len(_vec)==2
-  prefix = _vec[0]
-  epoch = int(_vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  if tsym is None:
-    all_layers = sym.get_internals()
-    tsym = all_layers['fc1_output']
-  if i==0:
-    prefix = 'age'
-  else:
-    prefix = 'gender'
-  for k,v in arg_params.iteritems():
-    if k.startswith(prefix):
-      print('arg', i, k)
-      targ[k] = v
-  for k,v in aux_params.iteritems():
-    if k.startswith(prefix):
-      print('aux', i, k)
-      taux[k] = v
-  i+=1
-dellist = []
-#for k,v in arg_params.iteritems():
-#  if k.startswith('fc7'):
-#    dellist.append(k)
-for d in dellist:
-  del targ[d]
-mx.model.save_checkpoint(args.prefix, 0, tsym, targ, taux)
-
diff --git a/deploy/helper.py b/deploy/helper.py
deleted file mode 100644
index b82c4b77a..000000000
--- a/deploy/helper.py
+++ /dev/null
@@ -1,168 +0,0 @@
-# coding: utf-8
-# YuanYang
-import math
-import cv2
-import numpy as np
-
-
-def nms(boxes, overlap_threshold, mode='Union'):
-    """
-        non max suppression
-
-    Parameters:
-    ----------
-        box: numpy array n x 5
-            input bbox array
-        overlap_threshold: float number
-            threshold of overlap
-        mode: float number
-            how to compute overlap ratio, 'Union' or 'Min'
-    Returns:
-    -------
-        index array of the selected bbox
-    """
-    # if there are no boxes, return an empty list
-    if len(boxes) == 0:
-        return []
-
-    # if the bounding boxes integers, convert them to floats
-    if boxes.dtype.kind == "i":
-        boxes = boxes.astype("float")
-
-    # initialize the list of picked indexes
-    pick = []
-
-    # grab the coordinates of the bounding boxes
-    x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)]
-
-    area = (x2 - x1 + 1) * (y2 - y1 + 1)
-    idxs = np.argsort(score)
-
-    # keep looping while some indexes still remain in the indexes list
-    while len(idxs) > 0:
-        # grab the last index in the indexes list and add the index value to the list of picked indexes
-        last = len(idxs) - 1
-        i = idxs[last]
-        pick.append(i)
-
-        xx1 = np.maximum(x1[i], x1[idxs[:last]])
-        yy1 = np.maximum(y1[i], y1[idxs[:last]])
-        xx2 = np.minimum(x2[i], x2[idxs[:last]])
-        yy2 = np.minimum(y2[i], y2[idxs[:last]])
-
-        # compute the width and height of the bounding box
-        w = np.maximum(0, xx2 - xx1 + 1)
-        h = np.maximum(0, yy2 - yy1 + 1)
-
-        inter = w * h
-        if mode == 'Min':
-            overlap = inter / np.minimum(area[i], area[idxs[:last]])
-        else:
-            overlap = inter / (area[i] + area[idxs[:last]] - inter)
-
-        # delete all indexes from the index list that have
-        idxs = np.delete(idxs, np.concatenate(([last],
-                                               np.where(overlap > overlap_threshold)[0])))
-
-    return pick
-
-def adjust_input(in_data):
-    """
-        adjust the input from (h, w, c) to ( 1, c, h, w) for network input
-
-    Parameters:
-    ----------
-        in_data: numpy array of shape (h, w, c)
-            input data
-    Returns:
-    -------
-        out_data: numpy array of shape (1, c, h, w)
-            reshaped array
-    """
-    if in_data.dtype is not np.dtype('float32'):
-        out_data = in_data.astype(np.float32)
-    else:
-        out_data = in_data
-
-    out_data = out_data.transpose((2,0,1))
-    out_data = np.expand_dims(out_data, 0)
-    out_data = (out_data - 127.5)*0.0078125
-    return out_data
-
-def generate_bbox(map, reg, scale, threshold):
-     """
-         generate bbox from feature map
-     Parameters:
-     ----------
-         map: numpy array , n x m x 1
-             detect score for each position
-         reg: numpy array , n x m x 4
-             bbox
-         scale: float number
-             scale of this detection
-         threshold: float number
-             detect threshold
-     Returns:
-     -------
-         bbox array
-     """
-     stride = 2
-     cellsize = 12
-
-     t_index = np.where(map>threshold)
-
-     # find nothing
-     if t_index[0].size == 0:
-         return np.array([])
-
-     dx1, dy1, dx2, dy2 = [reg[0, i, t_index[0], t_index[1]] for i in range(4)]
-
-     reg = np.array([dx1, dy1, dx2, dy2])
-     score = map[t_index[0], t_index[1]]
-     boundingbox = np.vstack([np.round((stride*t_index[1]+1)/scale),
-                              np.round((stride*t_index[0]+1)/scale),
-                              np.round((stride*t_index[1]+1+cellsize)/scale),
-                              np.round((stride*t_index[0]+1+cellsize)/scale),
-                              score,
-                              reg])
-
-     return boundingbox.T
-
-
-def detect_first_stage(img, net, scale, threshold):
-    """
-        run PNet for first stage
-    
-    Parameters:
-    ----------
-        img: numpy array, bgr order
-            input image
-        scale: float number
-            how much should the input image scale
-        net: PNet
-            worker
-    Returns:
-    -------
-        total_boxes : bboxes
-    """
-    height, width, _ = img.shape
-    hs = int(math.ceil(height * scale))
-    ws = int(math.ceil(width * scale))
-    
-    im_data = cv2.resize(img, (ws,hs))
-    
-    # adjust for the network input
-    input_buf = adjust_input(im_data)
-    output = net.predict(input_buf)
-    boxes = generate_bbox(output[1][0,1,:,:], output[0], scale, threshold)
-
-    if boxes.size == 0:
-        return None
-
-    # nms
-    pick = nms(boxes[:,0:5], 0.5, mode='Union')
-    boxes = boxes[pick]
-    return boxes
-
-def detect_first_stage_warpper( args ):
-    return detect_first_stage(*args)
diff --git a/deploy/model_slim.py b/deploy/model_slim.py
deleted file mode 100644
index 327adde40..000000000
--- a/deploy/model_slim.py
+++ /dev/null
@@ -1,31 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import sys
-import os
-import argparse
-import numpy as np
-import mxnet as mx
-
-parser = argparse.ArgumentParser(description='face model slim')
-# general
-parser.add_argument('--model', default='../models/model-r34-amf/model,60', help='path to load model.')
-args = parser.parse_args()
-
-_vec = args.model.split(',')
-assert len(_vec)==2
-prefix = _vec[0]
-epoch = int(_vec[1])
-print('loading',prefix, epoch)
-sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-all_layers = sym.get_internals()
-sym = all_layers['fc1_output']
-dellist = []
-for k,v in arg_params.iteritems():
-  if k.startswith('fc7'):
-    dellist.append(k)
-for d in dellist:
-  del arg_params[d]
-mx.model.save_checkpoint(prefix+"s", 0, sym, arg_params, aux_params)
-
diff --git a/deploy/mtcnn-model/det1-0001.params b/deploy/mtcnn-model/det1-0001.params
deleted file mode 100644
index e4b04aa90..000000000
Binary files a/deploy/mtcnn-model/det1-0001.params and /dev/null differ
diff --git a/deploy/mtcnn-model/det1-symbol.json b/deploy/mtcnn-model/det1-symbol.json
deleted file mode 100644
index bd9b7720b..000000000
--- a/deploy/mtcnn-model/det1-symbol.json
+++ /dev/null
@@ -1,266 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "10", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(2,2)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "16", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[11, 0], [12, 0], [13, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[14, 0], [15, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(1,1)", 
-        "no_bias": "False", 
-        "num_filter": "4", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4_2", 
-      "inputs": [[16, 0], [17, 0], [18, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(1,1)", 
-        "no_bias": "False", 
-        "num_filter": "2", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4_1", 
-      "inputs": [[16, 0], [20, 0], [21, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxActivation", 
-      "param": {"mode": "channel"}, 
-      "name": "prob1", 
-      "inputs": [[22, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    12, 
-    13, 
-    15, 
-    17, 
-    18, 
-    20, 
-    21
-  ], 
-  "heads": [[19, 0], [23, 0]]
-}
\ No newline at end of file
diff --git a/deploy/mtcnn-model/det1.caffemodel b/deploy/mtcnn-model/det1.caffemodel
deleted file mode 100644
index 79e93b488..000000000
Binary files a/deploy/mtcnn-model/det1.caffemodel and /dev/null differ
diff --git a/deploy/mtcnn-model/det1.prototxt b/deploy/mtcnn-model/det1.prototxt
deleted file mode 100644
index c5c1657c1..000000000
--- a/deploy/mtcnn-model/det1.prototxt
+++ /dev/null
@@ -1,177 +0,0 @@
-name: "PNet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 12
-input_dim: 12
-
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 10
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 2
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 16
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-}
-
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "conv2"
-  top: "conv3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 32
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-	  type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-}
-
-
-layer {
-  name: "conv4-1"
-  type: "Convolution"
-  bottom: "conv3"
-  top: "conv4-1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 2
-    kernel_size: 1
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "conv4-2"
-  type: "Convolution"
-  bottom: "conv3"
-  top: "conv4-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 4
-    kernel_size: 1
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv4-1"
-  top: "prob1"
-}
diff --git a/deploy/mtcnn-model/det2-0001.params b/deploy/mtcnn-model/det2-0001.params
deleted file mode 100644
index a14a478e1..000000000
Binary files a/deploy/mtcnn-model/det2-0001.params and /dev/null differ
diff --git a/deploy/mtcnn-model/det2-symbol.json b/deploy/mtcnn-model/det2-symbol.json
deleted file mode 100644
index a13246a7a..000000000
--- a/deploy/mtcnn-model/det2-symbol.json
+++ /dev/null
@@ -1,324 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2", 
-      "inputs": [[11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[12, 0], [13, 0], [14, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[15, 0], [16, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "128"
-      }, 
-      "name": "conv4", 
-      "inputs": [[17, 0], [18, 0], [19, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[20, 0], [21, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "4"
-      }, 
-      "name": "conv5_2", 
-      "inputs": [[22, 0], [23, 0], [24, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "conv5_1", 
-      "inputs": [[22, 0], [26, 0], [27, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prob1_label", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxOutput", 
-      "param": {
-        "grad_scale": "1", 
-        "ignore_label": "-1", 
-        "multi_output": "False", 
-        "normalization": "null", 
-        "use_ignore": "False"
-      }, 
-      "name": "prob1", 
-      "inputs": [[28, 0], [29, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    13, 
-    14, 
-    16, 
-    18, 
-    19, 
-    21, 
-    23, 
-    24, 
-    26, 
-    27, 
-    29
-  ], 
-  "heads": [[25, 0], [30, 0]]
-}
\ No newline at end of file
diff --git a/deploy/mtcnn-model/det2.caffemodel b/deploy/mtcnn-model/det2.caffemodel
deleted file mode 100644
index a5a540c00..000000000
Binary files a/deploy/mtcnn-model/det2.caffemodel and /dev/null differ
diff --git a/deploy/mtcnn-model/det2.prototxt b/deploy/mtcnn-model/det2.prototxt
deleted file mode 100644
index 51093e6d9..000000000
--- a/deploy/mtcnn-model/det2.prototxt
+++ /dev/null
@@ -1,228 +0,0 @@
-name: "RNet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 24
-input_dim: 24
-
-
-##########################
-######################
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-  propagate_down: true
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-  propagate_down: true
-}
-layer {
-  name: "pool2"
-  type: "Pooling"
-  bottom: "conv2"
-  top: "pool2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-####################################
-
-##################################
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "pool2"
-  top: "conv3"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-  propagate_down: true
-}
-###############################
-
-###############################
-
-layer {
-  name: "conv4"
-  type: "InnerProduct"
-  bottom: "conv3"
-  top: "conv4"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  inner_product_param {
-    num_output: 128
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "conv4"
-  top: "conv4"
-}
-
-layer {
-  name: "conv5-1"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5-1"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  inner_product_param {
-    num_output: 2
-    #kernel_size: 1
-    #stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv5-2"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 4
-    #kernel_size: 1
-    #stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv5-1"
-  top: "prob1"
-}
\ No newline at end of file
diff --git a/deploy/mtcnn-model/det3-0001.params b/deploy/mtcnn-model/det3-0001.params
deleted file mode 100644
index cae898b8a..000000000
Binary files a/deploy/mtcnn-model/det3-0001.params and /dev/null differ
diff --git a/deploy/mtcnn-model/det3-symbol.json b/deploy/mtcnn-model/det3-symbol.json
deleted file mode 100644
index 00061edfd..000000000
--- a/deploy/mtcnn-model/det3-symbol.json
+++ /dev/null
@@ -1,418 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2", 
-      "inputs": [[11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[12, 0], [13, 0], [14, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[15, 0], [16, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(2,2)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool3", 
-      "inputs": [[17, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4", 
-      "inputs": [[18, 0], [19, 0], [20, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[21, 0], [22, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "256"
-      }, 
-      "name": "conv5", 
-      "inputs": [[23, 0], [24, 0], [25, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu5", 
-      "inputs": [[26, 0], [27, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "10"
-      }, 
-      "name": "conv6_3", 
-      "inputs": [[28, 0], [29, 0], [30, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "4"
-      }, 
-      "name": "conv6_2", 
-      "inputs": [[28, 0], [32, 0], [33, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "conv6_1", 
-      "inputs": [[28, 0], [35, 0], [36, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prob1_label", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxOutput", 
-      "param": {
-        "grad_scale": "1", 
-        "ignore_label": "-1", 
-        "multi_output": "False", 
-        "normalization": "null", 
-        "use_ignore": "False"
-      }, 
-      "name": "prob1", 
-      "inputs": [[37, 0], [38, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    13, 
-    14, 
-    16, 
-    19, 
-    20, 
-    22, 
-    24, 
-    25, 
-    27, 
-    29, 
-    30, 
-    32, 
-    33, 
-    35, 
-    36, 
-    38
-  ], 
-  "heads": [[31, 0], [34, 0], [39, 0]]
-}
\ No newline at end of file
diff --git a/deploy/mtcnn-model/det3.caffemodel b/deploy/mtcnn-model/det3.caffemodel
deleted file mode 100644
index 7b4b8a4af..000000000
Binary files a/deploy/mtcnn-model/det3.caffemodel and /dev/null differ
diff --git a/deploy/mtcnn-model/det3.prototxt b/deploy/mtcnn-model/det3.prototxt
deleted file mode 100644
index a19230786..000000000
--- a/deploy/mtcnn-model/det3.prototxt
+++ /dev/null
@@ -1,294 +0,0 @@
-name: "ONet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 48
-input_dim: 48
-##################################
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 32
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "prelu2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-}
-layer {
-  name: "pool2"
-  type: "Pooling"
-  bottom: "conv2"
-  top: "pool2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "pool2"
-  top: "conv3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-	num_output: 64
-	kernel_size: 3
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-}
-layer {
-  name: "pool3"
-  type: "Pooling"
-  bottom: "conv3"
-  top: "pool3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 2
-    stride: 2
-  }
-}
-layer {
-  name: "conv4"
-  type: "Convolution"
-  bottom: "pool3"
-  top: "conv4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-	num_output: 128
-	kernel_size: 2
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "conv4"
-  top: "conv4"
-}
-
-
-layer {
-  name: "conv5"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-	#kernel_size: 3
-	num_output: 256
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "drop5"
-  type: "Dropout"
-  bottom: "conv5"
-  top: "conv5"
-  dropout_param {
-    dropout_ratio: 0.25
-  }
-}
-layer {
-  name: "prelu5"
-  type: "PReLU"
-  bottom: "conv5"
-  top: "conv5"
-}
-
-
-layer {
-  name: "conv6-1"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    #kernel_size: 1
-	num_output: 2
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv6-2"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-  	#kernel_size: 1
-	num_output: 4
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv6-3"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-  	#kernel_size: 1
-	num_output: 10
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv6-1"
-  top: "prob1"
-}
diff --git a/deploy/mtcnn-model/det4-0001.params b/deploy/mtcnn-model/det4-0001.params
deleted file mode 100644
index efca9a959..000000000
Binary files a/deploy/mtcnn-model/det4-0001.params and /dev/null differ
diff --git a/deploy/mtcnn-model/det4-symbol.json b/deploy/mtcnn-model/det4-symbol.json
deleted file mode 100644
index aa90e2a09..000000000
--- a/deploy/mtcnn-model/det4-symbol.json
+++ /dev/null
@@ -1,1392 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SliceChannel", 
-      "param": {
-        "axis": "1", 
-        "num_outputs": "5", 
-        "squeeze_axis": "False"
-      }, 
-      "name": "slice", 
-      "inputs": [[0, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_1", 
-      "inputs": [[1, 0], [2, 0], [3, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_1", 
-      "inputs": [[4, 0], [5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_1", 
-      "inputs": [[6, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_1", 
-      "inputs": [[7, 0], [8, 0], [9, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_1", 
-      "inputs": [[10, 0], [11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_1", 
-      "inputs": [[12, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_1", 
-      "inputs": [[13, 0], [14, 0], [15, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_1", 
-      "inputs": [[16, 0], [17, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_2", 
-      "inputs": [[1, 1], [19, 0], [20, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_2", 
-      "inputs": [[21, 0], [22, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_2", 
-      "inputs": [[23, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_2", 
-      "inputs": [[24, 0], [25, 0], [26, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_2", 
-      "inputs": [[27, 0], [28, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_2", 
-      "inputs": [[29, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_2", 
-      "inputs": [[30, 0], [31, 0], [32, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_2", 
-      "inputs": [[33, 0], [34, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_3", 
-      "inputs": [[1, 2], [36, 0], [37, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_3", 
-      "inputs": [[38, 0], [39, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_3", 
-      "inputs": [[40, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_3", 
-      "inputs": [[41, 0], [42, 0], [43, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_3", 
-      "inputs": [[44, 0], [45, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_3", 
-      "inputs": [[46, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_3", 
-      "inputs": [[47, 0], [48, 0], [49, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_3", 
-      "inputs": [[50, 0], [51, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_4", 
-      "inputs": [[1, 3], [53, 0], [54, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_4", 
-      "inputs": [[55, 0], [56, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_4", 
-      "inputs": [[57, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_4", 
-      "inputs": [[58, 0], [59, 0], [60, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_4", 
-      "inputs": [[61, 0], [62, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_4", 
-      "inputs": [[63, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_4", 
-      "inputs": [[64, 0], [65, 0], [66, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_4", 
-      "inputs": [[67, 0], [68, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_5", 
-      "inputs": [[1, 4], [70, 0], [71, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_5", 
-      "inputs": [[72, 0], [73, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_5", 
-      "inputs": [[74, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_5", 
-      "inputs": [[75, 0], [76, 0], [77, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_5", 
-      "inputs": [[78, 0], [79, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_5", 
-      "inputs": [[80, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_5", 
-      "inputs": [[81, 0], [82, 0], [83, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_5", 
-      "inputs": [[84, 0], [85, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Concat", 
-      "param": {
-        "dim": "1", 
-        "num_args": "5"
-      }, 
-      "name": "concat", 
-      "inputs": [[18, 0], [35, 0], [52, 0], [69, 0], [86, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "256"
-      }, 
-      "name": "fc4", 
-      "inputs": [[87, 0], [88, 0], [89, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[90, 0], [91, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_1", 
-      "inputs": [[92, 0], [93, 0], [94, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_1", 
-      "inputs": [[95, 0], [96, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_1", 
-      "inputs": [[97, 0], [98, 0], [99, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_2", 
-      "inputs": [[92, 0], [101, 0], [102, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_2", 
-      "inputs": [[103, 0], [104, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_2", 
-      "inputs": [[105, 0], [106, 0], [107, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_3", 
-      "inputs": [[92, 0], [109, 0], [110, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_3", 
-      "inputs": [[111, 0], [112, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_3", 
-      "inputs": [[113, 0], [114, 0], [115, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_4", 
-      "inputs": [[92, 0], [117, 0], [118, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_4", 
-      "inputs": [[119, 0], [120, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_4", 
-      "inputs": [[121, 0], [122, 0], [123, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_5", 
-      "inputs": [[92, 0], [125, 0], [126, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_5", 
-      "inputs": [[127, 0], [128, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_5", 
-      "inputs": [[129, 0], [130, 0], [131, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    2, 
-    3, 
-    5, 
-    8, 
-    9, 
-    11, 
-    14, 
-    15, 
-    17, 
-    19, 
-    20, 
-    22, 
-    25, 
-    26, 
-    28, 
-    31, 
-    32, 
-    34, 
-    36, 
-    37, 
-    39, 
-    42, 
-    43, 
-    45, 
-    48, 
-    49, 
-    51, 
-    53, 
-    54, 
-    56, 
-    59, 
-    60, 
-    62, 
-    65, 
-    66, 
-    68, 
-    70, 
-    71, 
-    73, 
-    76, 
-    77, 
-    79, 
-    82, 
-    83, 
-    85, 
-    88, 
-    89, 
-    91, 
-    93, 
-    94, 
-    96, 
-    98, 
-    99, 
-    101, 
-    102, 
-    104, 
-    106, 
-    107, 
-    109, 
-    110, 
-    112, 
-    114, 
-    115, 
-    117, 
-    118, 
-    120, 
-    122, 
-    123, 
-    125, 
-    126, 
-    128, 
-    130, 
-    131
-  ], 
-  "heads": [[100, 0], [108, 0], [116, 0], [124, 0], [132, 0]]
-}
\ No newline at end of file
diff --git a/deploy/mtcnn-model/det4.caffemodel b/deploy/mtcnn-model/det4.caffemodel
deleted file mode 100644
index 38353c4e3..000000000
Binary files a/deploy/mtcnn-model/det4.caffemodel and /dev/null differ
diff --git a/deploy/mtcnn-model/det4.prototxt b/deploy/mtcnn-model/det4.prototxt
deleted file mode 100644
index 4cdc329d0..000000000
--- a/deploy/mtcnn-model/det4.prototxt
+++ /dev/null
@@ -1,995 +0,0 @@
-name: "LNet"
-input: "data"
-input_dim: 1
-input_dim: 15
-input_dim: 24
-input_dim: 24
-
-layer {
-  name: "slicer_data"
-  type: "Slice"
-  bottom: "data"
-  top: "data241"
-  top: "data242"
-  top: "data243"
-  top: "data244"
-  top: "data245"
-  slice_param {
-    axis: 1
-    slice_point: 3
-    slice_point: 6
-    slice_point: 9
-    slice_point: 12
-  }
-}
-layer {
-  name: "conv1_1"
-  type: "Convolution"
-  bottom: "data241"
-  top: "conv1_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_1"
-  type: "PReLU"
-  bottom: "conv1_1"
-  top: "conv1_1"
-
-}
-layer {
-  name: "pool1_1"
-  type: "Pooling"
-  bottom: "conv1_1"
-  top: "pool1_1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_1"
-  type: "Convolution"
-  bottom: "pool1_1"
-  top: "conv2_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_1"
-  type: "PReLU"
-  bottom: "conv2_1"
-  top: "conv2_1"
-}
-layer {
-  name: "pool2_1"
-  type: "Pooling"
-  bottom: "conv2_1"
-  top: "pool2_1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_1"
-  type: "Convolution"
-  bottom: "pool2_1"
-  top: "conv3_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_1"
-  type: "PReLU"
-  bottom: "conv3_1"
-  top: "conv3_1"
-}
-##########################
-layer {
-  name: "conv1_2"
-  type: "Convolution"
-  bottom: "data242"
-  top: "conv1_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_2"
-  type: "PReLU"
-  bottom: "conv1_2"
-  top: "conv1_2"
-
-}
-layer {
-  name: "pool1_2"
-  type: "Pooling"
-  bottom: "conv1_2"
-  top: "pool1_2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_2"
-  type: "Convolution"
-  bottom: "pool1_2"
-  top: "conv2_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_2"
-  type: "PReLU"
-  bottom: "conv2_2"
-  top: "conv2_2"
-}
-layer {
-  name: "pool2_2"
-  type: "Pooling"
-  bottom: "conv2_2"
-  top: "pool2_2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_2"
-  type: "Convolution"
-  bottom: "pool2_2"
-  top: "conv3_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_2"
-  type: "PReLU"
-  bottom: "conv3_2"
-  top: "conv3_2"
-}
-##########################
-##########################
-layer {
-  name: "conv1_3"
-  type: "Convolution"
-  bottom: "data243"
-  top: "conv1_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_3"
-  type: "PReLU"
-  bottom: "conv1_3"
-  top: "conv1_3"
-
-}
-layer {
-  name: "pool1_3"
-  type: "Pooling"
-  bottom: "conv1_3"
-  top: "pool1_3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_3"
-  type: "Convolution"
-  bottom: "pool1_3"
-  top: "conv2_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_3"
-  type: "PReLU"
-  bottom: "conv2_3"
-  top: "conv2_3"
-}
-layer {
-  name: "pool2_3"
-  type: "Pooling"
-  bottom: "conv2_3"
-  top: "pool2_3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_3"
-  type: "Convolution"
-  bottom: "pool2_3"
-  top: "conv3_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_3"
-  type: "PReLU"
-  bottom: "conv3_3"
-  top: "conv3_3"
-}
-##########################
-##########################
-layer {
-  name: "conv1_4"
-  type: "Convolution"
-  bottom: "data244"
-  top: "conv1_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_4"
-  type: "PReLU"
-  bottom: "conv1_4"
-  top: "conv1_4"
-
-}
-layer {
-  name: "pool1_4"
-  type: "Pooling"
-  bottom: "conv1_4"
-  top: "pool1_4"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_4"
-  type: "Convolution"
-  bottom: "pool1_4"
-  top: "conv2_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_4"
-  type: "PReLU"
-  bottom: "conv2_4"
-  top: "conv2_4"
-}
-layer {
-  name: "pool2_4"
-  type: "Pooling"
-  bottom: "conv2_4"
-  top: "pool2_4"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_4"
-  type: "Convolution"
-  bottom: "pool2_4"
-  top: "conv3_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_4"
-  type: "PReLU"
-  bottom: "conv3_4"
-  top: "conv3_4"
-}
-##########################
-##########################
-layer {
-  name: "conv1_5"
-  type: "Convolution"
-  bottom: "data245"
-  top: "conv1_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_5"
-  type: "PReLU"
-  bottom: "conv1_5"
-  top: "conv1_5"
-
-}
-layer {
-  name: "pool1_5"
-  type: "Pooling"
-  bottom: "conv1_5"
-  top: "pool1_5"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_5"
-  type: "Convolution"
-  bottom: "pool1_5"
-  top: "conv2_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_5"
-  type: "PReLU"
-  bottom: "conv2_5"
-  top: "conv2_5"
-}
-layer {
-  name: "pool2_5"
-  type: "Pooling"
-  bottom: "conv2_5"
-  top: "pool2_5"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_5"
-  type: "Convolution"
-  bottom: "pool2_5"
-  top: "conv3_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_5"
-  type: "PReLU"
-  bottom: "conv3_5"
-  top: "conv3_5"
-}
-##########################
-layer { 
-	name: "concat" 
-	bottom: "conv3_1" 
-	bottom: "conv3_2" 
-	bottom: "conv3_3" 
-	bottom: "conv3_4" 
-	bottom: "conv3_5" 
-	top: "conv3" 
-	type: "Concat" 
-	concat_param { 
-		axis: 1 
-	} 
-}
-##########################
-layer {
-  name: "fc4"
-  type: "InnerProduct"
-  bottom: "conv3"
-  top: "fc4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 256
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "fc4"
-  top: "fc4"
-}
-############################
-layer {
-  name: "fc4_1"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_1"
-  type: "PReLU"
-  bottom: "fc4_1"
-  top: "fc4_1"
-}
-layer {
-  name: "fc5_1"
-  type: "InnerProduct"
-  bottom: "fc4_1"
-  top: "fc5_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-
-#########################
-layer {
-  name: "fc4_2"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_2"
-  type: "PReLU"
-  bottom: "fc4_2"
-  top: "fc4_2"
-}
-layer {
-  name: "fc5_2"
-  type: "InnerProduct"
-  bottom: "fc4_2"
-  top: "fc5_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_3"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_3"
-  type: "PReLU"
-  bottom: "fc4_3"
-  top: "fc4_3"
-}
-layer {
-  name: "fc5_3"
-  type: "InnerProduct"
-  bottom: "fc4_3"
-  top: "fc5_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_4"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_4"
-  type: "PReLU"
-  bottom: "fc4_4"
-  top: "fc4_4"
-}
-layer {
-  name: "fc5_4"
-  type: "InnerProduct"
-  bottom: "fc4_4"
-  top: "fc5_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_5"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_5"
-  type: "PReLU"
-  bottom: "fc4_5"
-  top: "fc4_5"
-}
-layer {
-  name: "fc5_5"
-  type: "InnerProduct"
-  bottom: "fc4_5"
-  top: "fc5_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-
diff --git a/deploy/mtcnn_detector.py b/deploy/mtcnn_detector.py
deleted file mode 100644
index c7332a5cf..000000000
--- a/deploy/mtcnn_detector.py
+++ /dev/null
@@ -1,659 +0,0 @@
-# coding: utf-8
-import os
-import mxnet as mx
-import numpy as np
-import math
-import cv2
-from multiprocessing import Pool
-from itertools import repeat
-try:
-    from itertools import izip
-except ImportError:
-    izip = zip
-
-from helper import nms, adjust_input, generate_bbox, detect_first_stage_warpper
-
-class MtcnnDetector(object):
-    """
-        Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks
-        see https://github.com/kpzhang93/MTCNN_face_detection_alignment
-        this is a mxnet version
-    """
-    def __init__(self,
-                 model_folder='.',
-                 minsize = 20,
-                 threshold = [0.6, 0.7, 0.8],
-                 factor = 0.709,
-                 num_worker = 1,
-                 accurate_landmark = False,
-                 ctx=mx.cpu()):
-        """
-            Initialize the detector
-
-            Parameters:
-            ----------
-                model_folder : string
-                    path for the models
-                minsize : float number
-                    minimal face to detect
-                threshold : float number
-                    detect threshold for 3 stages
-                factor: float number
-                    scale factor for image pyramid
-                num_worker: int number
-                    number of processes we use for first stage
-                accurate_landmark: bool
-                    use accurate landmark localization or not
-
-        """
-        self.num_worker = num_worker
-        self.accurate_landmark = accurate_landmark
-
-        # load 4 models from folder
-        models = ['det1', 'det2', 'det3','det4']
-        models = [ os.path.join(model_folder, f) for f in models]
-        
-        self.PNets = []
-        for i in range(num_worker):
-            workner_net = mx.model.FeedForward.load(models[0], 1, ctx=ctx)
-            self.PNets.append(workner_net)
-
-        #self.Pool = Pool(num_worker)
-
-        self.RNet = mx.model.FeedForward.load(models[1], 1, ctx=ctx)
-        self.ONet = mx.model.FeedForward.load(models[2], 1, ctx=ctx)
-        self.LNet = mx.model.FeedForward.load(models[3], 1, ctx=ctx)
-
-        self.minsize   = float(minsize)
-        self.factor    = float(factor)
-        self.threshold = threshold
-
-
-    def convert_to_square(self, bbox):
-        """
-            convert bbox to square
-
-        Parameters:
-        ----------
-            bbox: numpy array , shape n x 5
-                input bbox
-
-        Returns:
-        -------
-            square bbox
-        """
-        square_bbox = bbox.copy()
-
-        h = bbox[:, 3] - bbox[:, 1] + 1
-        w = bbox[:, 2] - bbox[:, 0] + 1
-        max_side = np.maximum(h,w)
-        square_bbox[:, 0] = bbox[:, 0] + w*0.5 - max_side*0.5
-        square_bbox[:, 1] = bbox[:, 1] + h*0.5 - max_side*0.5
-        square_bbox[:, 2] = square_bbox[:, 0] + max_side - 1
-        square_bbox[:, 3] = square_bbox[:, 1] + max_side - 1
-        return square_bbox
-
-    def calibrate_box(self, bbox, reg):
-        """
-            calibrate bboxes
-
-        Parameters:
-        ----------
-            bbox: numpy array, shape n x 5
-                input bboxes
-            reg:  numpy array, shape n x 4
-                bboxex adjustment
-
-        Returns:
-        -------
-            bboxes after refinement
-
-        """
-        w = bbox[:, 2] - bbox[:, 0] + 1
-        w = np.expand_dims(w, 1)
-        h = bbox[:, 3] - bbox[:, 1] + 1
-        h = np.expand_dims(h, 1)
-        reg_m = np.hstack([w, h, w, h])
-        aug = reg_m * reg
-        bbox[:, 0:4] = bbox[:, 0:4] + aug
-        return bbox
-
- 
-    def pad(self, bboxes, w, h):
-        """
-            pad the the bboxes, alse restrict the size of it
-
-        Parameters:
-        ----------
-            bboxes: numpy array, n x 5
-                input bboxes
-            w: float number
-                width of the input image
-            h: float number
-                height of the input image
-        Returns :
-        ------s
-            dy, dx : numpy array, n x 1
-                start point of the bbox in target image
-            edy, edx : numpy array, n x 1
-                end point of the bbox in target image
-            y, x : numpy array, n x 1
-                start point of the bbox in original image
-            ex, ex : numpy array, n x 1
-                end point of the bbox in original image
-            tmph, tmpw: numpy array, n x 1
-                height and width of the bbox
-
-        """
-        tmpw, tmph = bboxes[:, 2] - bboxes[:, 0] + 1,  bboxes[:, 3] - bboxes[:, 1] + 1
-        num_box = bboxes.shape[0]
-
-        dx , dy= np.zeros((num_box, )), np.zeros((num_box, ))
-        edx, edy  = tmpw.copy()-1, tmph.copy()-1
-
-        x, y, ex, ey = bboxes[:, 0], bboxes[:, 1], bboxes[:, 2], bboxes[:, 3]
-
-        tmp_index = np.where(ex > w-1)
-        edx[tmp_index] = tmpw[tmp_index] + w - 2 - ex[tmp_index]
-        ex[tmp_index] = w - 1
-
-        tmp_index = np.where(ey > h-1)
-        edy[tmp_index] = tmph[tmp_index] + h - 2 - ey[tmp_index]
-        ey[tmp_index] = h - 1
-
-        tmp_index = np.where(x < 0)
-        dx[tmp_index] = 0 - x[tmp_index]
-        x[tmp_index] = 0
-
-        tmp_index = np.where(y < 0)
-        dy[tmp_index] = 0 - y[tmp_index]
-        y[tmp_index] = 0
-
-        return_list = [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph]
-        return_list = [item.astype(np.int32) for item in return_list]
-
-        return  return_list
-
-    def slice_index(self, number):
-        """
-            slice the index into (n,n,m), m < n
-        Parameters:
-        ----------
-            number: int number
-                number
-        """
-        def chunks(l, n):
-            """Yield successive n-sized chunks from l."""
-            for i in range(0, len(l), n):
-                yield l[i:i + n]
-        num_list = range(number)
-        return list(chunks(num_list, self.num_worker))
-        
-    def detect_face_limited(self, img, det_type=2):
-        height, width, _ = img.shape
-        if det_type>=2:
-          total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-          num_box = total_boxes.shape[0]
-
-          # pad the bbox
-          [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-          # (3, 24, 24) is the input shape for RNet
-          input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)
-
-          for i in range(num_box):
-              tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)
-              tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-              input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))
-
-          output = self.RNet.predict(input_buf)
-
-          # filter the total_boxes with threshold
-          passed = np.where(output[1][:, 1] > self.threshold[1])
-          total_boxes = total_boxes[passed]
-
-          if total_boxes.size == 0:
-              return None
-
-          total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))
-          reg = output[0][passed]
-
-          # nms
-          pick = nms(total_boxes, 0.7, 'Union')
-          total_boxes = total_boxes[pick]
-          total_boxes = self.calibrate_box(total_boxes, reg[pick])
-          total_boxes = self.convert_to_square(total_boxes)
-          total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-        else:
-          total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-        num_box = total_boxes.shape[0]
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 48, 48) is the input shape for ONet
-        input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))
-
-        output = self.ONet.predict(input_buf)
-        #print(output[2])
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[2][:, 1] > self.threshold[2])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))
-        reg = output[1][passed]
-        points = output[0][passed]
-
-        # compute landmark points
-        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-        points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]
-        points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]
-
-        # nms
-        total_boxes = self.calibrate_box(total_boxes, reg)
-        pick = nms(total_boxes, 0.7, 'Min')
-        total_boxes = total_boxes[pick]
-        points = points[pick]
-        
-        if not self.accurate_landmark:
-            return total_boxes, points
-
-        #############################################
-        # extended stage
-        #############################################
-        num_box = total_boxes.shape[0]
-        patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)
-        patchw = np.round(patchw*0.25)
-
-        # make it even
-        patchw[np.where(np.mod(patchw,2) == 1)] += 1
-
-        input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)
-        for i in range(5):
-            x, y = points[:, i], points[:, i+5]
-            x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)
-            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,
-                                                                    width,
-                                                                    height)
-            for j in range(num_box):
-                tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)
-                tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]
-                input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))
-
-        output = self.LNet.predict(input_buf)
-
-        pointx = np.zeros((num_box, 5))
-        pointy = np.zeros((num_box, 5))
-
-        for k in range(5):
-            # do not make a large movement
-            tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)
-            output[k][tmp_index[0]] = 0.5
-
-            pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw
-            pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw
-
-        points = np.hstack([pointx, pointy])
-        points = points.astype(np.int32)
-
-        return total_boxes, points
-
-    def detect_face(self, img, det_type=0):
-        """
-            detect face over img
-        Parameters:
-        ----------
-            img: numpy array, bgr order of shape (1, 3, n, m)
-                input image
-        Retures:
-        -------
-            bboxes: numpy array, n x 5 (x1,y2,x2,y2,score)
-                bboxes
-            points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
-                landmarks
-        """
-
-        # check input
-        height, width, _ = img.shape
-        if det_type==0:
-            MIN_DET_SIZE = 12
-
-            if img is None:
-                return None
-
-            # only works for color image
-            if len(img.shape) != 3:
-                return None
-
-            # detected boxes
-            total_boxes = []
-
-            minl = min( height, width)
-
-            # get all the valid scales
-            scales = []
-            m = MIN_DET_SIZE/self.minsize
-            minl *= m
-            factor_count = 0
-            while minl > MIN_DET_SIZE:
-                scales.append(m*self.factor**factor_count)
-                minl *= self.factor
-                factor_count += 1
-
-            #############################################
-            # first stage
-            #############################################
-            #for scale in scales:
-            #    return_boxes = self.detect_first_stage(img, scale, 0)
-            #    if return_boxes is not None:
-            #        total_boxes.append(return_boxes)
-            
-            sliced_index = self.slice_index(len(scales))
-            total_boxes = []
-            for batch in sliced_index:
-                #local_boxes = self.Pool.map( detect_first_stage_warpper, \
-                #        izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )
-                local_boxes = map( detect_first_stage_warpper, \
-                        izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )
-                total_boxes.extend(local_boxes)
-            
-            # remove the Nones 
-            total_boxes = [ i for i in total_boxes if i is not None]
-
-            if len(total_boxes) == 0:
-                return None
-            
-            total_boxes = np.vstack(total_boxes)
-
-            if total_boxes.size == 0:
-                return None
-
-            # merge the detection from first stage
-            pick = nms(total_boxes[:, 0:5], 0.7, 'Union')
-            total_boxes = total_boxes[pick]
-
-            bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-            bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-
-            # refine the bboxes
-            total_boxes = np.vstack([total_boxes[:, 0]+total_boxes[:, 5] * bbw,
-                                     total_boxes[:, 1]+total_boxes[:, 6] * bbh,
-                                     total_boxes[:, 2]+total_boxes[:, 7] * bbw,
-                                     total_boxes[:, 3]+total_boxes[:, 8] * bbh,
-                                     total_boxes[:, 4]
-                                     ])
-
-            total_boxes = total_boxes.T
-            total_boxes = self.convert_to_square(total_boxes)
-            total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-        else:
-            total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-
-        #############################################
-        # second stage
-        #############################################
-        num_box = total_boxes.shape[0]
-
-        # pad the bbox
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 24, 24) is the input shape for RNet
-        input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))
-
-        output = self.RNet.predict(input_buf)
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[1][:, 1] > self.threshold[1])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))
-        reg = output[0][passed]
-
-        # nms
-        pick = nms(total_boxes, 0.7, 'Union')
-        total_boxes = total_boxes[pick]
-        total_boxes = self.calibrate_box(total_boxes, reg[pick])
-        total_boxes = self.convert_to_square(total_boxes)
-        total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-
-        #############################################
-        # third stage
-        #############################################
-        num_box = total_boxes.shape[0]
-
-        # pad the bbox
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 48, 48) is the input shape for ONet
-        input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))
-
-        output = self.ONet.predict(input_buf)
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[2][:, 1] > self.threshold[2])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))
-        reg = output[1][passed]
-        points = output[0][passed]
-
-        # compute landmark points
-        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-        points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]
-        points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]
-
-        # nms
-        total_boxes = self.calibrate_box(total_boxes, reg)
-        pick = nms(total_boxes, 0.7, 'Min')
-        total_boxes = total_boxes[pick]
-        points = points[pick]
-        
-        if not self.accurate_landmark:
-            return total_boxes, points
-
-        #############################################
-        # extended stage
-        #############################################
-        num_box = total_boxes.shape[0]
-        patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)
-        patchw = np.round(patchw*0.25)
-
-        # make it even
-        patchw[np.where(np.mod(patchw,2) == 1)] += 1
-
-        input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)
-        for i in range(5):
-            x, y = points[:, i], points[:, i+5]
-            x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)
-            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,
-                                                                    width,
-                                                                    height)
-            for j in range(num_box):
-                tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)
-                tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]
-                input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))
-
-        output = self.LNet.predict(input_buf)
-
-        pointx = np.zeros((num_box, 5))
-        pointy = np.zeros((num_box, 5))
-
-        for k in range(5):
-            # do not make a large movement
-            tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)
-            output[k][tmp_index[0]] = 0.5
-
-            pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw
-            pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw
-
-        points = np.hstack([pointx, pointy])
-        points = points.astype(np.int32)
-
-        return total_boxes, points
-
-
-
-    def list2colmatrix(self, pts_list):
-        """
-            convert list to column matrix
-        Parameters:
-        ----------
-            pts_list:
-                input list
-        Retures:
-        -------
-            colMat: 
-
-        """
-        assert len(pts_list) > 0
-        colMat = []
-        for i in range(len(pts_list)):
-            colMat.append(pts_list[i][0])
-            colMat.append(pts_list[i][1])
-        colMat = np.matrix(colMat).transpose()
-        return colMat
-
-    def find_tfrom_between_shapes(self, from_shape, to_shape):
-        """
-            find transform between shapes
-        Parameters:
-        ----------
-            from_shape: 
-            to_shape: 
-        Retures:
-        -------
-            tran_m:
-            tran_b:
-        """
-        assert from_shape.shape[0] == to_shape.shape[0] and from_shape.shape[0] % 2 == 0
-
-        sigma_from = 0.0
-        sigma_to = 0.0
-        cov = np.matrix([[0.0, 0.0], [0.0, 0.0]])
-
-        # compute the mean and cov
-        from_shape_points = from_shape.reshape(from_shape.shape[0]/2, 2)
-        to_shape_points = to_shape.reshape(to_shape.shape[0]/2, 2)
-        mean_from = from_shape_points.mean(axis=0)
-        mean_to = to_shape_points.mean(axis=0)
-
-        for i in range(from_shape_points.shape[0]):
-            temp_dis = np.linalg.norm(from_shape_points[i] - mean_from)
-            sigma_from += temp_dis * temp_dis
-            temp_dis = np.linalg.norm(to_shape_points[i] - mean_to)
-            sigma_to += temp_dis * temp_dis
-            cov += (to_shape_points[i].transpose() - mean_to.transpose()) * (from_shape_points[i] - mean_from)
-
-        sigma_from = sigma_from / to_shape_points.shape[0]
-        sigma_to = sigma_to / to_shape_points.shape[0]
-        cov = cov / to_shape_points.shape[0]
-
-        # compute the affine matrix
-        s = np.matrix([[1.0, 0.0], [0.0, 1.0]])
-        u, d, vt = np.linalg.svd(cov)
-
-        if np.linalg.det(cov) < 0:
-            if d[1] < d[0]:
-                s[1, 1] = -1
-            else:
-                s[0, 0] = -1
-        r = u * s * vt
-        c = 1.0
-        if sigma_from != 0:
-            c = 1.0 / sigma_from * np.trace(np.diag(d) * s)
-
-        tran_b = mean_to.transpose() - c * r * mean_from.transpose()
-        tran_m = c * r
-
-        return tran_m, tran_b
-
-    def extract_image_chips(self, img, points, desired_size=256, padding=0):
-        """
-            crop and align face
-        Parameters:
-        ----------
-            img: numpy array, bgr order of shape (1, 3, n, m)
-                input image
-            points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
-            desired_size: default 256
-            padding: default 0
-        Retures:
-        -------
-            crop_imgs: list, n
-                cropped and aligned faces 
-        """
-        crop_imgs = []
-        for p in points:
-            shape  =[]
-            for k in range(len(p)/2):
-                shape.append(p[k])
-                shape.append(p[k+5])
-
-            if padding > 0:
-                padding = padding
-            else:
-                padding = 0
-            # average positions of face points
-            mean_face_shape_x = [0.224152, 0.75610125, 0.490127, 0.254149, 0.726104]
-            mean_face_shape_y = [0.2119465, 0.2119465, 0.628106, 0.780233, 0.780233]
-
-            from_points = []
-            to_points = []
-
-            for i in range(len(shape)/2):
-                x = (padding + mean_face_shape_x[i]) / (2 * padding + 1) * desired_size
-                y = (padding + mean_face_shape_y[i]) / (2 * padding + 1) * desired_size
-                to_points.append([x, y])
-                from_points.append([shape[2*i], shape[2*i+1]])
-
-            # convert the points to Mat
-            from_mat = self.list2colmatrix(from_points)
-            to_mat = self.list2colmatrix(to_points)
-
-            # compute the similar transfrom
-            tran_m, tran_b = self.find_tfrom_between_shapes(from_mat, to_mat)
-
-            probe_vec = np.matrix([1.0, 0.0]).transpose()
-            probe_vec = tran_m * probe_vec
-
-            scale = np.linalg.norm(probe_vec)
-            angle = 180.0 / math.pi * math.atan2(probe_vec[1, 0], probe_vec[0, 0])
-
-            from_center = [(shape[0]+shape[2])/2.0, (shape[1]+shape[3])/2.0]
-            to_center = [0, 0]
-            to_center[1] = desired_size * 0.4
-            to_center[0] = desired_size * 0.5
-
-            ex = to_center[0] - from_center[0]
-            ey = to_center[1] - from_center[1]
-
-            rot_mat = cv2.getRotationMatrix2D((from_center[0], from_center[1]), -1*angle, scale)
-            rot_mat[0][2] += ex
-            rot_mat[1][2] += ey
-
-            chips = cv2.warpAffine(img, rot_mat, (desired_size, desired_size))
-            crop_imgs.append(chips)
-
-        return crop_imgs
-
diff --git a/deploy/test.py b/deploy/test.py
deleted file mode 100644
index 6150a08b1..000000000
--- a/deploy/test.py
+++ /dev/null
@@ -1,34 +0,0 @@
-import face_model
-import argparse
-import cv2
-import sys
-import numpy as np
-
-parser = argparse.ArgumentParser(description='face model test')
-# general
-parser.add_argument('--image-size', default='112,112', help='')
-parser.add_argument('--model', default='', help='path to load model.')
-parser.add_argument('--ga-model', default='', help='path to load model.')
-parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-parser.add_argument('--det', default=0, type=int, help='mtcnn option, 1 means using R+O, 0 means detect from begining')
-parser.add_argument('--flip', default=0, type=int, help='whether do lr flip aug')
-parser.add_argument('--threshold', default=1.24, type=float, help='ver dist threshold')
-args = parser.parse_args()
-
-model = face_model.FaceModel(args)
-img = cv2.imread('Tom_Hanks_54745.png')
-img = model.get_input(img)
-#f1 = model.get_feature(img)
-#print(f1[0:10])
-gender, age = model.get_ga(img)
-print(gender)
-print(age)
-sys.exit(0)
-img = cv2.imread('/raid5data/dplearn/megaface/facescrubr/112x112/Tom_Hanks/Tom_Hanks_54733.png')
-f2 = model.get_feature(img)
-dist = np.sum(np.square(f1-f2))
-print(dist)
-sim = np.dot(f1, f2.T)
-print(sim)
-#diff = np.subtract(source_feature, target_feature)
-#dist = np.sum(np.square(diff),1)
diff --git a/detection/README.md b/detection/README.md
new file mode 100644
index 000000000..8419ec15f
--- /dev/null
+++ b/detection/README.md
@@ -0,0 +1,42 @@
+## Face Detection
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/logo3.jpg" width="240"/>
+</div>
+
+
+## Introduction
+
+These are the face detection methods of [InsightFace](https://insightface.ai)
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/11513D05.jpg" width="800"/>
+</div>
+
+
+### Datasets
+
+  Please refer to [datasets](_datasets_) page for the details of face detection datasets used for training and evaluation.
+
+### Evaluation
+
+  Please refer to [evaluation](_evaluation_) page for the details of face recognition evaluation.
+
+
+## Methods
+
+
+Supported methods:
+
+- [x] [RetinaFace (CVPR'2020)](retinaface)
+- [x] [SCRFD (Arxiv'2021)](scrfd)
+- [x] [blazeface_paddle](blazeface_paddle)
+
+
+## Contributing
+
+We appreciate all contributions to improve the face detection model zoo of InsightFace. 
+
+
diff --git a/detection/_datasets_/README.md b/detection/_datasets_/README.md
new file mode 100644
index 000000000..618227009
--- /dev/null
+++ b/detection/_datasets_/README.md
@@ -0,0 +1,31 @@
+# Face Detection Datasets
+
+(Updating)
+
+## Training Datasets
+
+### WiderFace
+
+http://shuoyang1213.me/WIDERFACE/
+
+
+
+## Test Datasets
+
+### WiderFace
+
+http://shuoyang1213.me/WIDERFACE/
+
+### FDDB
+
+http://vis-www.cs.umass.edu/fddb/
+
+### AFW
+
+
+### PASCAL FACE
+
+
+### MALF
+
+http://www.cbsr.ia.ac.cn/faceevaluation/
diff --git a/detection/blazeface_paddle/README.md b/detection/blazeface_paddle/README.md
new file mode 120000
index 000000000..13c4f964b
--- /dev/null
+++ b/detection/blazeface_paddle/README.md
@@ -0,0 +1 @@
+README_en.md
\ No newline at end of file
diff --git a/detection/blazeface_paddle/README_cn.md b/detection/blazeface_paddle/README_cn.md
new file mode 100644
index 000000000..0762058e5
--- /dev/null
+++ b/detection/blazeface_paddle/README_cn.md
@@ -0,0 +1,355 @@
+简体中文 | [English](README_en.md)
+
+# 人脸检测模型
+
+* [1. 简介](#简介)
+* [2. 模型库](#模型库)
+* [3. 安装](#安装)
+* [4. 数据准备](#数据准备)
+* [5. 参数配置](#参数配置)
+* [6. 训练与评估](#训练与评估)
+  * [6.1 训练](#训练)
+  * [6.2 在WIDER-FACE数据集上评估](#评估)
+  * [6.3 推理部署](#推理部署)
+  * [6.4 推理速度提升](#推理速度提升)
+  * [6.5 人脸检测demo](#人脸检测demo)
+* [7. 参考文献](#参考文献)
+
+<a name="简介"></a>
+
+## 1. 简介
+
+`Arcface-Paddle`是基于PaddlePaddle实现的，开源深度人脸检测、识别工具。`Arcface-Paddle`目前提供了三个预训练模型，包括用于人脸检测的 `BlazeFace`、用于人脸识别的 `ArcFace` 和 `MobileFace`。
+
+- 本部分内容为人脸检测部分，基于PaddleDetection进行开发。
+- 人脸识别相关内容可以参考：[人脸识别](../../recognition/arcface_paddle/README_cn.md)。
+- 基于PaddleInference的Whl包预测部署内容可以参考：[Whl包预测部署](https://github.com/littletomatodonkey/insight-face-paddle)。
+
+
+<a name="模型库"></a>
+
+## 2. 模型库
+
+### WIDER-FACE数据集上的mAP
+
+| 网络结构 | 输入尺寸 | 图片个数/GPU | epoch数量 | Easy/Medium/Hard Set  | CPU预测时延 | GPU 预测时延 | 模型大小(MB) | 预训练模型地址 | inference模型地址 |  配置文件 |
+|:------------:|:--------:|:----:|:-------:|:-------:|:-------:|:---------:|:----------:|:---------:|:---------:|:--------:|
+| BlazeFace-FPN-SSH  | 640  |    8    | 1000    | 0.9187 / 0.8979 / 0.8168 | 31.7ms  |  5.6ms | 0.646 |[下载链接](https://paddledet.bj.bcebos.com/models/blazeface_fpn_ssh_1000e.pdparams) | [下载链接](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/blazeface_fpn_ssh_1000e_v1.0_infer.tar) |  [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.1/configs/face_detection/blazeface_fpn_ssh_1000e.yml) |
+| RetinaFace  | 480x640  |    -    | -     | - / - / 0.8250 | 182.0ms  |  17.4ms | 1.680 | -  |  - | - |
+
+
+**注意:**  
+- 我们使用多尺度评估策略得到`Easy/Medium/Hard Set`里的mAP。具体细节请参考[在WIDER-FACE数据集上评估](#评估)。
+- 测量速度时我们使用640*640的分辨，在 Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz cpu，CPU线程数设置为5，更多细节请参考[推理速度提升](#推理速度提升)。
+- `RetinaFace`的速度测试代码参考自：[../retinaface/README.md](../retinaface/README.md).
+- 测试环境为
+  - CPU: Intel(R) Xeon(R) Gold 6184 CPU @ 2.40GHz
+  - GPU: a single NVIDIA Tesla V100
+
+
+<a name="安装"></a>
+
+## 3. 安装
+
+请参考[安装教程](../../recognition/arcface_paddle/install_ch.md)安装PaddlePaddle以及PaddleDetection。
+
+<a name="数据准备"></a>
+
+## 4. 数据准备
+我们使用[WIDER-FACE数据集](http://shuoyang1213.me/WIDERFACE/)进行训练和模型测试，官方网站提供了详细的数据介绍。
+- WIDER-Face数据源:  
+使用如下目录结构加载`wider_face`类型的数据集：
+
+  ```
+  dataset/wider_face/
+  ├── wider_face_split
+  │   ├── wider_face_train_bbx_gt.txt
+  │   ├── wider_face_val_bbx_gt.txt
+  ├── WIDER_train
+  │   ├── images
+  │   │   ├── 0--Parade
+  │   │   │   ├── 0_Parade_marchingband_1_100.jpg
+  │   │   │   ├── 0_Parade_marchingband_1_381.jpg
+  │   │   │   │   ...
+  │   │   ├── 10--People_Marching
+  │   │   │   ...
+  ├── WIDER_val
+  │   ├── images
+  │   │   ├── 0--Parade
+  │   │   │   ├── 0_Parade_marchingband_1_1004.jpg
+  │   │   │   ├── 0_Parade_marchingband_1_1045.jpg
+  │   │   │   │   ...
+  │   │   ├── 10--People_Marching
+  │   │   │   ...
+  ```
+
+- 手动下载数据集：
+要下载WIDER-FACE数据集，请运行以下命令：
+```
+cd dataset/wider_face && ./download_wider_face.sh
+```
+
+<a name="参数配置"></a>
+
+## 5. 参数配置
+
+我们使用 `configs/face_detection/blazeface_fpn_ssh_1000e.yml`配置进行训练，配置文件摘要如下：
+
+```yaml
+
+_BASE_: [
+  '../datasets/wider_face.yml',
+  '../runtime.yml',
+  '_base_/optimizer_1000e.yml',
+  '_base_/blazeface_fpn.yml',
+  '_base_/face_reader.yml',
+]
+weights: output/blazeface_fpn_ssh_1000e/model_final
+multi_scale_eval: True
+
+```
+
+`blazeface_fpn_ssh_1000e.yml` 配置需要依赖其他的配置文件，在该例子中需要依赖:
+
+```
+wider_face.yml：主要说明了训练数据和验证数据的路径
+
+runtime.yml：主要说明了公共的运行参数，比如是否使用GPU、每多少个epoch存储checkpoint等
+
+optimizer_1000e.yml：主要说明了学习率和优化器的配置
+
+blazeface_fpn.yml：主要说明模型和主干网络的情况
+
+face_reader.yml：主要说明数据读取器配置，如batch size，并发加载子进程数等，同时包含读取后预处理操作，如resize、数据增强等等
+```
+
+根据实际情况，修改上述文件，比如数据集路径、batch size等。
+
+基础模型的配置可以参考`configs/face_detection/_base_/blazeface.yml`；
+改进模型增加FPN和SSH的neck结构，配置文件可以参考`configs/face_detection/_base_/blazeface_fpn.yml`，可以根据需求配置FPN和SSH，具体如下：
+```yaml
+BlazeNet:
+   blaze_filters: [[24, 24], [24, 24], [24, 48, 2], [48, 48], [48, 48]]
+   double_blaze_filters: [[48, 24, 96, 2], [96, 24, 96], [96, 24, 96],
+                           [96, 24, 96, 2], [96, 24, 96], [96, 24, 96]]
+   act: hard_swish # 配置backbone中BlazeBlock的激活函数，基础模型为relu，增加FPN和SSH时需使用hard_swish
+
+BlazeNeck:
+   neck_type : fpn_ssh # 可选only_fpn、only_ssh和fpn_ssh
+   in_channel: [96,96]
+```
+
+<a name="训练与评估"></a>
+
+## 6. 训练与评估
+
+<a name="训练"></a>
+
+### 6.1 训练
+首先，下载预训练模型文件：
+```bash
+wget https://paddledet.bj.bcebos.com/models/pretrained/blazenet_pretrain.pdparams
+```
+PaddleDetection提供了单卡/多卡训练模式，满足用户多种训练需求
+* GPU单卡训练
+```bash
+export CUDA_VISIBLE_DEVICES=0 #windows和Mac下不需要执行该命令
+python tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -o pretrain_weight=blazenet_pretrain
+```
+
+* GPU多卡训练
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3 #windows和Mac下不需要执行该命令
+python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -o pretrain_weight=blazenet_pretrain
+```
+* 模型恢复训练
+
+  在日常训练过程中，有的用户由于一些原因导致训练中断，用户可以使用-r的命令恢复训练
+
+```bash
+export CUDA_VISIBLE_DEVICES=0 #windows和Mac下不需要执行该命令
+python tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -r output/blazeface_fan_ssh_1000e/100
+ ```
+* 训练策略
+
+`BlazeFace`训练是以每卡`batch_size=32`在4卡GPU上进行训练(总`batch_size`是128),学习率为0.002，并且训练1000epoch。
+
+
+**注意:** 人脸检测模型目前不支持边训练边评估。
+
+<a name="评估"></a>
+
+### 6.2 在WIDER-FACE数据集上评估
+- 步骤一：评估并生成结果文件：
+```shell
+python -u tools/eval.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml \
+       -o weights=output/blazeface_fpn_ssh_1000e/model_final \
+       multi_scale_eval=True BBoxPostProcess.nms.score_threshold=0.1
+```
+设置`multi_scale_eval=True`进行多尺度评估，评估完成后，将在`output/pred`中生成txt格式的测试结果。
+
+- 步骤二：下载官方评估脚本和Ground Truth文件：
+```
+wget http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/support/eval_script/eval_tools.zip
+unzip eval_tools.zip && rm -f eval_tools.zip
+```
+
+- 步骤三：开始评估
+
+方法一：python评估。
+
+```bash
+git clone https://github.com/wondervictor/WiderFace-Evaluation.git
+cd WiderFace-Evaluation
+# 编译
+python3 setup.py build_ext --inplace
+# 开始评估
+python3 evaluation.py -p /path/to/PaddleDetection/output/pred -g /path/to/eval_tools/ground_truth
+```
+
+方法二：MatLab评估。
+
+```bash
+# 在`eval_tools/wider_eval.m`中修改保存结果路径和绘制曲线的名称：
+pred_dir = './pred';  
+legend_name = 'Paddle-BlazeFace';
+
+`wider_eval.m` 是评估模块的主要执行程序。运行命令如下：
+matlab -nodesktop -nosplash -nojvm -r "run wider_eval.m;quit;"
+```
+<a name="推理部署"></a>
+
+### 6.3 推理部署
+
+在模型训练过程中保存的模型文件是包含前向预测和反向传播的过程，在实际的工业部署则不需要反向传播，因此需要将模型进行导成部署需要的模型格式。
+在PaddleDetection中提供了 `tools/export_model.py`脚本来导出模型
+
+```bash
+python tools/export_model.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml --output_dir=./inference_model \
+ -o weights=output/blazeface_fpn_ssh_1000e/best_model BBoxPostProcess.nms.score_threshold=0.1
+```
+
+预测模型会导出到`inference_model/blazeface_fpn_ssh_1000e`目录下，分别为`infer_cfg.yml`, `model.pdiparams`, `model.pdiparams.info`,`model.pdmodel` 如果不指定文件夹，模型则会导出在`output_inference`
+
+* 这里将nms后处理`score_threshold`修改为0.1，因为mAP基本没有影响的情况下，GPU预测速度能够大幅提升。更多关于模型导出的文档，请参考[模型导出文档](https://github.com/PaddlePaddle/PaddleDetection/deploy/EXPORT_MODEL.md)
+
+ PaddleDetection提供了PaddleInference、PaddleServing、PaddleLite多种部署形式，支持服务端、移动端、嵌入式等多种平台，提供了完善的Python和C++部署方案。
+* 在这里，我们以Python为例，说明如何使用PaddleInference进行模型部署
+
+```bash
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_file=demo/road554.png --use_gpu=True
+```
+* 同时`infer.py`提供了丰富的接口，用户进行接入视频文件、摄像头进行预测，更多内容请参考[Python端预测部署](https://github.com/PaddlePaddle/PaddleDetection/deploy/python.md)
+
+* 更多关于预测部署的文档，请参考[预测部署文档](https://github.com/PaddlePaddle/PaddleDetection/deploy/README.md) 。
+
+<a name="推理速度提升"></a>
+
+### 6.4 推理速度提升
+如果想要复现我们提供的速度指标，请修改预测模型配置文件`./inference_model/blazeface_fpn_ssh_1000e/infer_cfg.yml`中的输入尺寸，如下所示:
+```yaml
+mode: fluid
+draw_threshold: 0.5
+metric: WiderFace
+arch: Face
+min_subgraph_size: 3
+Preprocess:
+- is_scale: false
+  mean:
+  - 123
+  - 117
+  - 104
+  std:
+  - 127.502231
+  - 127.502231
+  - 127.502231
+  type: NormalizeImage
+- interp: 1
+  keep_ratio: false
+  target_size:
+  - 640
+  - 640
+  type: Resize
+- type: Permute
+label_list:
+- face
+```
+如果希望模型在cpu环境下更快推理，可安装[paddlepaddle_gpu-0.0.0](https://paddle-wheel.bj.bcebos.com/develop-cpu-mkl/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl) （mkldnn的依赖）可开启mkldnn加速推理。
+
+```bash
+# 使用GPU测速：
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_dir=./path/images --run_benchmark=True --use_gpu=True
+
+# 使用cpu测速：
+# 下载paddle whl包
+wget https://paddle-wheel.bj.bcebos.com/develop-cpu-mkl/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl
+# 安装paddlepaddle_gpu-0.0.0
+pip install paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl
+# 推理
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_dir=./path/images --enable_mkldnn=True --run_benchmark=True --cpu_threads=5
+```
+
+<a name="人脸检测demo"></a>
+
+### 6.5 人脸检测demo
+
+本节介绍基于提供的BlazeFace模型进行人脸检测。
+
+先下载待检测图像与字体文件。
+
+```bash
+# 下载用于人脸检测的示例图像
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends1.jpg
+# 下载字体，用于可视化
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/SourceHanSansCN-Medium.otf
+```
+
+示例图像如下所示。
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends1.jpg"  width = "800" />
+</div>
+
+
+检测的示例命令如下。
+
+```shell
+python3.7 test_blazeface.py --input=friends1.jpg  --output="./output"
+```
+
+最终可视化结果保存在`output`目录下，可视化结果如下所示。
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/output/friends1.jpg"  width = "800" />
+</div>
+
+
+更多关于参数解释，索引库构建、人脸识别、whl包预测部署的内容可以参考：[Whl包预测部署](https://github.com/littletomatodonkey/insight-face-paddle)。
+
+<a name="参考文献"></a>
+
+## 7. 参考文献
+
+```
+@misc{long2020ppyolo,
+title={PP-YOLO: An Effective and Efficient Implementation of Object Detector},
+author={Xiang Long and Kaipeng Deng and Guanzhong Wang and Yang Zhang and Qingqing Dang and Yuan Gao and Hui Shen and Jianguo Ren and Shumin Han and Errui Ding and Shilei Wen},
+year={2020},
+eprint={2007.12099},
+archivePrefix={arXiv},
+primaryClass={cs.CV}
+}
+@misc{ppdet2019,
+title={PaddleDetection, Object detection and instance segmentation toolkit based on PaddlePaddle.},
+author={PaddlePaddle Authors},
+howpublished = {\url{https://github.com/PaddlePaddle/PaddleDetection}},
+year={2019}
+}
+@article{bazarevsky2019blazeface,
+title={BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs},
+author={Valentin Bazarevsky and Yury Kartynnik and Andrey Vakunov and Karthik Raveendran and Matthias Grundmann},
+year={2019},
+eprint={1907.05047},
+ archivePrefix={arXiv}
+}
+```
diff --git a/detection/blazeface_paddle/README_en.md b/detection/blazeface_paddle/README_en.md
new file mode 100644
index 000000000..24fd17863
--- /dev/null
+++ b/detection/blazeface_paddle/README_en.md
@@ -0,0 +1,354 @@
+[简体中文](README_cn.md) | English
+
+# FaceDetection
+
+* [1. Introduction](#Introduction)
+* [2. Model Zoo](#Model_Zoo)
+* [3. Installation](#Installation)
+* [4. Data Pipline](#Data_Pipline)
+* [5. Configuration File](#Configuration_File)
+* [6. Training and Inference](#Training_and_Inference)
+  * [6.1 Training](#Training)
+  * [6.2 Evaluate on the WIDER FACE](#Evaluation)
+  * [6.3 Inference deployment](#Inference_deployment)
+  * [6.4 Improvement of inference speed](#Increase_in_inference_speed)
+  * [6.4 Face detection demo](#Face_detection_demo)
+* [7. Citations](#Citations)
+
+<a name="Introduction"></a>
+
+## 1. Introduction
+
+`Arcface-Paddle` is an open source deep face detection and recognition toolkit, powered by PaddlePaddle. `Arcface-Paddle` provides three related pretrained models now, include `BlazeFace` for face detection, `ArcFace` and `MobileFace` for face recognition.
+
+- This tutorial is mainly about face detection based on `PaddleDetection`.
+- For face recognition task, please refer to: [Face recognition tuturial](../../recognition/arcface_paddle/README_en.md).
+- For Whl package inference using PaddleInference, please refer to [whl package inference](https://github.com/littletomatodonkey/insight-face-paddle).
+
+<a name="Model_Zoo"></a>
+
+## 2. Model Zoo
+
+### mAP in WIDER FACE
+
+| Model | input size | images/GPU | epochs | Easy/Medium/Hard Set  | CPU time cost | GPU time cost| Model Size(MB) | Pretrained model | Inference model | Config |
+|:------------:|:--------:|:----:|:-------:|:-------:|:---------:|:---------:|:----------:|:---------:|:--------:|:--------:|
+| BlazeFace-FPN-SSH  | 640×640  |    8    | 1000     | 0.9187 / 0.8979 / 0.8168 | 31.7ms  |  5.6ms | 0.646 |[download link](https://paddledet.bj.bcebos.com/models/blazeface_fpn_ssh_1000e.pdparams)  |  [download link](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/blazeface_fpn_ssh_1000e_v1.0_infer.tar) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.1/configs/face_detection/blazeface_fpn_ssh_1000e.yml) |
+| RetinaFace  | 480x640  |    -    | -     | - / - / 0.8250 | 182.0ms  |  17.4ms | 1.680 | -  |  - | - |
+
+
+**NOTE:**  
+- Get mAP in `Easy/Medium/Hard Set` by multi-scale evaluation. For details can refer to [Evaluation](#Evaluate-on-the-WIDER-FACE).
+- Measuring the speed, we use the resolution of `640×640`, in Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz environment, cpu-threads are set as 5. For more details, you can refer to [Improvement of inference speed](#Increase_in_inference_speed).
+- Benchmark code for `RetinaFace` is from: [../retinaface/README.md](../retinaface/README.md).
+- The benchmark environment is
+  - CPU: Intel(R) Xeon(R) Gold 6184 CPU @ 2.40GHz
+  - GPU: a single NVIDIA Tesla V100
+
+<a name="Installation"></a>
+
+## 3. Installation
+
+Please refer to [installation tutorial](../../recognition/arcface_paddle/install_en.md) to install PaddlePaddle and PaddleDetection.
+
+
+<a name="Data_Pipline"></a>
+
+## 4. Data Pipline
+We use the [WIDER FACE dataset](http://shuoyang1213.me/WIDERFACE/) to carry out the training
+and testing of the model, the official website gives detailed data introduction.
+- WIDER Face data source:  
+Loads `wider_face` type dataset with directory structures like this:
+
+  ```
+  dataset/wider_face/
+  ├── wider_face_split
+  │   ├── wider_face_train_bbx_gt.txt
+  │   ├── wider_face_val_bbx_gt.txt
+  ├── WIDER_train
+  │   ├── images
+  │   │   ├── 0--Parade
+  │   │   │   ├── 0_Parade_marchingband_1_100.jpg
+  │   │   │   ├── 0_Parade_marchingband_1_381.jpg
+  │   │   │   │   ...
+  │   │   ├── 10--People_Marching
+  │   │   │   ...
+  ├── WIDER_val
+  │   ├── images
+  │   │   ├── 0--Parade
+  │   │   │   ├── 0_Parade_marchingband_1_1004.jpg
+  │   │   │   ├── 0_Parade_marchingband_1_1045.jpg
+  │   │   │   │   ...
+  │   │   ├── 10--People_Marching
+  │   │   │   ...
+  ```
+
+- Download dataset manually:  
+To download the WIDER FACE dataset, run the following commands:
+```
+cd dataset/wider_face && ./download_wider_face.sh
+```
+
+<a name="Configuration_file"></a>
+
+## 5. Configuration file
+
+We use the `configs/face_detection/blazeface_fpn_ssh_1000e.yml` configuration for training. The summary of the configuration file is as follows:
+
+```yaml
+_BASE_: [
+  '../datasets/wider_face.yml',
+  '../runtime.yml',
+  '_base_/optimizer_1000e.yml',
+  '_base_/blazeface_fpn.yml',
+  '_base_/face_reader.yml',
+]
+weights: output/blazeface_fpn_ssh_1000e/model_final
+multi_scale_eval: True
+```
+
+`blazeface_fpn_ssh_1000e.yml` The configuration needs to rely on other configuration files, in this example it needs to rely on:
+
+```
+wider_face.yml：Mainly explains the path of training data and verification data
+
+runtime.yml：Mainly describes the common operating parameters, such as whether to use GPU, how many epochs to store checkpoints, etc.
+
+optimizer_1000e.yml：Mainly explains the configuration of learning rate and optimizer
+
+blazeface_fpn.yml：Mainly explain the situation of the model and the backbone network
+
+face_reader.yml：It mainly describes the configuration of the data reader, such as batch size, the number of concurrent loading subprocesses, etc., and also includes post-reading preprocessing operations, such as resize, data enhancement, etc.
+```
+
+According to the actual situation, modify the above files, such as the data set path, batch size, etc.
+
+For the configuration of the base model, please refer to `configs/face_detection/_base_/blazeface.yml`.
+The improved model adds the neck structure of FPN and SSH. For the configuration file, please refer to `configs/face_detection/_base_/blazeface_fpn.yml`. You can configure FPN and SSH if needed, which is as follows:
+
+```yaml
+BlazeNet:
+   blaze_filters: [[24, 24], [24, 24], [24, 48, 2], [48, 48], [48, 48]]
+   double_blaze_filters: [[48, 24, 96, 2], [96, 24, 96], [96, 24, 96],
+                           [96, 24, 96, 2], [96, 24, 96], [96, 24, 96]]
+   act: hard_swish # Configure the activation function of BlazeBlock in backbone, the basic model is relu, hard_swish is required when adding FPN and SSH
+
+BlazeNeck:
+   neck_type : fpn_ssh # Optional only_fpn, only_ssh and fpn_ssh
+   in_channel: [96,96]
+```
+
+<a name="Training_and_Inference"></a>
+
+## 6. Training_and_Inference
+
+<a name="Training"></a>
+
+### 6.1 Training
+Firstly, download the pretrained model.
+```bash
+wget https://paddledet.bj.bcebos.com/models/pretrained/blazenet_pretrain.pdparams
+```
+PaddleDetection provides a single-GPU/multi-GPU training mode to meet the various training needs of users.
+* single-GPU training
+```bash
+export CUDA_VISIBLE_DEVICES=0 # Do not need to execute this command under windows and Mac
+python tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -o pretrain_weight=blazenet_pretrain
+```
+
+* multi-GPU training
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3 # Do not need to execute this command under windows and Mac
+python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -o pretrain_weight=blazenet_pretrain
+```
+* Resume training from Checkpoint
+
+  In the daily training process, if the training was be interrupted, using the -r command to resume training:
+
+```bash
+export CUDA_VISIBLE_DEVICES=0 # Do not need to execute this command under windows and Mac
+python tools/train.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml -r output/blazeface_fan_ssh_1000e/100
+ ```
+* Training hyperparameters
+
+`BlazeFace` training is based on each GPU `batch_size=32` training on 4 GPUs (total `batch_size` is 128), the learning rate is 0.002, and the total training epoch is set as 1000.
+
+
+**NOTE:** Not support evaluation during train.
+
+<a name="Evaluation"></a>
+
+### 6.2 Evaluate on the WIDER FACE
+- Evaluate and generate results files:
+```shell
+python -u tools/eval.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml \
+       -o weights=output/blazeface_fpn_ssh_1000e/model_final \
+       multi_scale_eval=True BBoxPostProcess.nms.score_threshold=0.1
+```
+Set `multi_scale_eval=True` for multi-scale evaluation，after the evaluation is completed, the test result in txt format will be generated in `output/pred`.
+
+- Download the official evaluation script to evaluate the AP metrics:
+
+```bash
+wget http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/support/eval_script/eval_tools.zip
+unzip eval_tools.zip && rm -f eval_tools.zip
+```
+
+- Start evaluation:
+
+Method One: Python evaluation:
+
+```bash
+git clone https://github.com/wondervictor/WiderFace-Evaluation.git
+cd WiderFace-Evaluation
+# Compile
+python3 setup.py build_ext --inplace
+# Start evaluation
+python3 evaluation.py -p /path/to/PaddleDetection/output/pred -g /path/to/eval_tools/ground_truth
+```
+
+Method Two: MatLab evaluation:
+
+```bash
+# Modify the result path and the name of the curve to be drawn in `eval_tools/wider_eval.m`:
+pred_dir = './pred';  
+legend_name = 'Paddle-BlazeFace';
+
+`wider_eval.m` is the main execution program of the evaluation module. The run command is as follows:
+matlab -nodesktop -nosplash -nojvm -r "run wider_eval.m;quit;"
+```
+<a name="Inference_deployment"></a>
+
+### 6.3 Inference deployment
+
+The model file saved in the model training process includes forward prediction and back propagation. In actual industrial deployment, back propagation is not required. Therefore, the model needs to be exported into the model format required for deployment.
+The `tools/export_model.py` script is provided in PaddleDetection to export the model:
+
+```bash
+python tools/export_model.py -c configs/face_detection/blazeface_fpn_ssh_1000e.yml --output_dir=./inference_model \
+ -o weights=output/blazeface_fpn_ssh_1000e/best_model BBoxPostProcess.nms.score_threshold=0.1
+```
+The inference model will be exported to the `inference_model/blazeface_fpn_ssh_1000e` directory, which are `infer_cfg.yml`, `model.pdiparams`, `model.pdiparams.info`, `model.pdmodel` If no folder is specified, the model will be exported In `output_inference`.
+
+* `score_threshold` for nms is modified as 0.1 for inference, because it takes great speed performance improvement while has little effect on mAP. For more documentation about model export, please refer to: [export doc](https://github.com/PaddlePaddle/PaddleDetection/deploy/EXPORT_MODEL.md)
+
+ PaddleDetection provides multiple deployment forms of PaddleInference, PaddleServing, and PaddleLite, supports multiple platforms such as server, mobile, and embedded, and provides a complete deployment plan for Python and C++.
+* Here, we take Python as an example to illustrate how to use PaddleInference for model deployment:
+```bash
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_file=demo/road554.png --use_gpu=True
+```
+* `infer.py` provides a rich interface for users to access video files and cameras for prediction. For more information, please refer to: [Python deployment](https://github.com/PaddlePaddle/PaddleDetection/deploy/python.md).
+
+* For more documentation on deployment, please refer to: [deploy doc](https://github.com/PaddlePaddle/PaddleDetection/deploy/README.md).
+
+<a name="Increase_in_inference_speed"></a>
+
+### 6.4 Improvement of inference speed
+
+If you want to reproduce our speed indicators, you need to modify the input size of inference model in the `./inference_model/blazeface_fpn_ssh_1000e/infer_cfg.yml` configuration file. As follows:
+```yaml
+mode: fluid
+draw_threshold: 0.5
+metric: WiderFace
+arch: Face
+min_subgraph_size: 3
+Preprocess:
+- is_scale: false
+  mean:
+  - 123
+  - 117
+  - 104
+  std:
+  - 127.502231
+  - 127.502231
+  - 127.502231
+  type: NormalizeImage
+- interp: 1
+  keep_ratio: false
+  target_size:
+  - 640
+  - 640
+  type: Resize
+- type: Permute
+label_list:
+- face
+```
+
+If you want the model to be inferred faster in the CPU environment, install [paddlepaddle_gpu-0.0.0](https://paddle-wheel.bj.bcebos.com/develop-cpu-mkl/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl) (dependency of mkldnn) and enable_mkldnn is set to True, when predicting acceleration. 
+
+```bash
+# use GPU:
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_dir=./path/images --run_benchmark=True --use_gpu=True
+
+# inference with mkldnn use CPU
+# downdoad whl package
+wget https://paddle-wheel.bj.bcebos.com/develop-cpu-mkl/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl
+#install paddlepaddle_gpu-0.0.0
+pip install paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl
+python deploy/python/infer.py --model_dir=./inference_model/blazeface_fpn_ssh_1000e --image_dir=./path/images --enable_mkldnn=True --run_benchmark=True --cpu_threads=5
+```
+
+<a name="Face_detection_demo"></a>
+
+### 6.5 Face detection demo
+
+This part talks about how to detect faces using BlazeFace model.
+
+Firstly, use the following commands to download the demo image and font file for visualization.
+
+
+```bash
+# Demo image
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends1.jpg
+# Font file for visualization
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/SourceHanSansCN-Medium.otf
+```
+
+The demo image is shown as follows.
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends1.jpg"  width = "800" />
+</div>
+
+
+Use the following command to run the face detection process.
+
+```shell
+python3.7 test_blazeface.py --input=friends1.jpg  --output="./output"
+```
+
+The final result is save in folder `output/`, which is shown as follows.
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/output/friends1.jpg"  width = "800" />
+</div>
+
+
+For more details about parameter explanations, face recognition, index gallery construction and whl package inference, please refer to [Whl package inference tutorial](https://github.com/littletomatodonkey/insight-face-paddle).
+
+
+## 7. Citations
+
+```
+@misc{long2020ppyolo,
+title={PP-YOLO: An Effective and Efficient Implementation of Object Detector},
+author={Xiang Long and Kaipeng Deng and Guanzhong Wang and Yang Zhang and Qingqing Dang and Yuan Gao and Hui Shen and Jianguo Ren and Shumin Han and Errui Ding and Shilei Wen},
+year={2020},
+eprint={2007.12099},
+archivePrefix={arXiv},
+primaryClass={cs.CV}
+}
+@misc{ppdet2019,
+title={PaddleDetection, Object detection and instance segmentation toolkit based on PaddlePaddle.},
+author={PaddlePaddle Authors},
+howpublished = {\url{https://github.com/PaddlePaddle/PaddleDetection}},
+year={2019}
+}
+@article{bazarevsky2019blazeface,
+title={BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs},
+author={Valentin Bazarevsky and Yury Kartynnik and Andrey Vakunov and Karthik Raveendran and Matthias Grundmann},
+year={2019},
+eprint={1907.05047},
+ archivePrefix={arXiv}
+}
+```
diff --git a/detection/blazeface_paddle/test_blazeface.py b/detection/blazeface_paddle/test_blazeface.py
new file mode 100644
index 000000000..fa8a8f103
--- /dev/null
+++ b/detection/blazeface_paddle/test_blazeface.py
@@ -0,0 +1,593 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import argparse
+import requests
+import logging
+import imghdr
+import pickle
+import tarfile
+from functools import partial
+
+import cv2
+import numpy as np
+from sklearn.metrics.pairwise import cosine_similarity
+from tqdm import tqdm
+from prettytable import PrettyTable
+from PIL import Image, ImageDraw, ImageFont
+import paddle
+from paddle.inference import Config
+from paddle.inference import create_predictor
+
+__all__ = ["parser"]
+BASE_INFERENCE_MODEL_DIR = os.path.expanduser("~/.insightface/ppmodels/")
+BASE_DOWNLOAD_URL = "https://paddle-model-ecology.bj.bcebos.com/model/insight-face/{}.tar"
+
+
+def parser(add_help=True):
+    def str2bool(v):
+        return v.lower() in ("true", "t", "1")
+
+    parser = argparse.ArgumentParser(add_help=add_help)
+    
+    parser.add_argument(
+        "--det_model",
+        type=str,
+        default="BlazeFace",
+        help="The detection model.")
+    parser.add_argument(
+        "--use_gpu",
+        type=str2bool,
+        default=True,
+        help="Whether use GPU to predict. Default by True.")
+    parser.add_argument(
+        "--enable_mkldnn",
+        type=str2bool,
+        default=True,
+        help="Whether use MKLDNN to predict, valid only when --use_gpu is False. Default by False."
+    )
+    parser.add_argument(
+        "--cpu_threads",
+        type=int,
+        default=1,
+        help="The num of threads with CPU, valid only when --use_gpu is False. Default by 1."
+    )
+    parser.add_argument(
+        "--input",
+        type=str,
+        help="The path or directory of image(s) or video to be predicted.")
+    parser.add_argument(
+        "--output", type=str, default="./output/", help="The directory of prediction result.")
+    parser.add_argument(
+        "--det_thresh",
+        type=float,
+        default=0.8,
+        help="The threshold of detection postprocess. Default by 0.8.")
+    return parser
+
+
+def print_config(args):
+    args = vars(args)
+    table = PrettyTable(['Param', 'Value'])
+    for param in args:
+        table.add_row([param, args[param]])
+    width = len(str(table).split("\n")[0])
+    print("{}".format("-" * width))
+    print("PaddleFace".center(width))
+    print(table)
+    print("Powered by PaddlePaddle!".rjust(width))
+    print("{}".format("-" * width))
+
+
+def download_with_progressbar(url, save_path):
+    """Download from url with progressbar.
+    """
+    if os.path.isfile(save_path):
+        os.remove(save_path)
+    response = requests.get(url, stream=True)
+    total_size_in_bytes = int(response.headers.get("content-length", 0))
+    block_size = 1024  # 1 Kibibyte
+    progress_bar = tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True)
+    with open(save_path, "wb") as file:
+        for data in response.iter_content(block_size):
+            progress_bar.update(len(data))
+            file.write(data)
+    progress_bar.close()
+    if total_size_in_bytes == 0 or progress_bar.n != total_size_in_bytes or not os.path.isfile(
+            save_path):
+        raise Exception(
+            f"Something went wrong while downloading model/image from {url}")
+
+
+def check_model_file(model):
+    """Check the model files exist and download and untar when no exist.
+    """
+    model_map = {
+        "ArcFace": "arcface_iresnet50_v1.0_infer",
+        "BlazeFace": "blazeface_fpn_ssh_1000e_v1.0_infer",
+        "MobileFace": "mobileface_v1.0_infer"
+    }
+
+    if os.path.isdir(model):
+        model_file_path = os.path.join(model, "inference.pdmodel")
+        params_file_path = os.path.join(model, "inference.pdiparams")
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            raise Exception(
+                f"The specifed model directory error. The drectory must include 'inference.pdmodel' and 'inference.pdiparams'."
+            )
+
+    elif model in model_map:
+        storage_directory = partial(os.path.join, BASE_INFERENCE_MODEL_DIR,
+                                    model)
+        url = BASE_DOWNLOAD_URL.format(model_map[model])
+
+        tar_file_name_list = [
+            "inference.pdiparams", "inference.pdiparams.info",
+            "inference.pdmodel"
+        ]
+        model_file_path = storage_directory("inference.pdmodel")
+        params_file_path = storage_directory("inference.pdiparams")
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            tmp_path = storage_directory(url.split("/")[-1])
+            logging.info(f"Download {url} to {tmp_path}")
+            os.makedirs(storage_directory(), exist_ok=True)
+            download_with_progressbar(url, tmp_path)
+            with tarfile.open(tmp_path, "r") as tarObj:
+                for member in tarObj.getmembers():
+                    filename = None
+                    for tar_file_name in tar_file_name_list:
+                        if tar_file_name in member.name:
+                            filename = tar_file_name
+                    if filename is None:
+                        continue
+                    file = tarObj.extractfile(member)
+                    with open(storage_directory(filename), "wb") as f:
+                        f.write(file.read())
+            os.remove(tmp_path)
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            raise Exception(
+                f"Something went wrong while downloading and unzip the model[{model}] files!"
+            )
+    else:
+        raise Exception(
+            f"The specifed model name error. Support 'BlazeFace' for detection. And support local directory that include model files ('inference.pdmodel' and 'inference.pdiparams')."
+        )
+
+    return model_file_path, params_file_path
+
+
+def normalize_image(img, scale=None, mean=None, std=None, order='chw'):
+    if isinstance(scale, str):
+        scale = eval(scale)
+    scale = np.float32(scale if scale is not None else 1.0 / 255.0)
+    mean = mean if mean is not None else [0.485, 0.456, 0.406]
+    std = std if std is not None else [0.229, 0.224, 0.225]
+
+    shape = (3, 1, 1) if order == 'chw' else (1, 1, 3)
+    mean = np.array(mean).reshape(shape).astype('float32')
+    std = np.array(std).reshape(shape).astype('float32')
+
+    if isinstance(img, Image.Image):
+        img = np.array(img)
+
+    assert isinstance(img, np.ndarray), "invalid input 'img' in NormalizeImage"
+    return (img.astype('float32') * scale - mean) / std
+
+
+def to_CHW_image(img):
+    if isinstance(img, Image.Image):
+        img = np.array(img)
+    return img.transpose((2, 0, 1))
+
+
+class ColorMap(object):
+    def __init__(self, num):
+        super().__init__()
+        self.get_color_map_list(num)
+        self.color_map = {}
+        self.ptr = 0
+
+    def __getitem__(self, key):
+        return self.color_map[key]
+
+    def update(self, keys):
+        for key in keys:
+            if key not in self.color_map:
+                i = self.ptr % len(self.color_list)
+                self.color_map[key] = self.color_list[i]
+                self.ptr += 1
+
+    def get_color_map_list(self, num_classes):
+        color_map = num_classes * [0, 0, 0]
+        for i in range(0, num_classes):
+            j = 0
+            lab = i
+            while lab:
+                color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
+                color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
+                color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
+                j += 1
+                lab >>= 3
+        self.color_list = [
+            color_map[i:i + 3] for i in range(0, len(color_map), 3)
+        ]
+
+
+class ImageReader(object):
+    def __init__(self, inputs):
+        super().__init__()
+        self.idx = 0
+        if isinstance(inputs, np.ndarray):
+            self.image_list = [inputs]
+        else:
+            imgtype_list = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff'}
+            self.image_list = []
+            if os.path.isfile(inputs):
+                if imghdr.what(inputs) not in imgtype_list:
+                    raise Exception(
+                        f"Error type of input path, only support: {imgtype_list}"
+                    )
+                self.image_list.append(inputs)
+            elif os.path.isdir(inputs):
+                tmp_file_list = os.listdir(inputs)
+                warn_tag = False
+                for file_name in tmp_file_list:
+                    file_path = os.path.join(inputs, file_name)
+                    if not os.path.isfile(file_path):
+                        warn_tag = True
+                        continue
+                    if imghdr.what(file_path) in imgtype_list:
+                        self.image_list.append(file_path)
+                    else:
+                        warn_tag = True
+                if warn_tag:
+                    logging.warning(
+                        f"The directory of input contine directory or not supported file type, only support: {imgtype_list}"
+                    )
+            else:
+                raise Exception(
+                    f"The file of input path not exist! Please check input: {inputs}"
+                )
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        if self.idx >= len(self.image_list):
+            raise StopIteration
+
+        data = self.image_list[self.idx]
+        if isinstance(data, np.ndarray):
+            self.idx += 1
+            return data, "tmp.png"
+        path = data
+        _, file_name = os.path.split(path)
+        img = cv2.imread(path)
+        if img is None:
+            logging.warning(f"Error in reading image: {path}! Ignored.")
+            self.idx += 1
+            return self.__next__()
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        self.idx += 1
+        return img, file_name
+
+    def __len__(self):
+        return len(self.image_list)
+
+
+class VideoReader(object):
+    def __init__(self, inputs):
+        super().__init__()
+        videotype_list = {"mp4"}
+        if os.path.splitext(inputs)[-1][1:] not in videotype_list:
+            raise Exception(
+                f"The input file is not supported, only support: {videotype_list}"
+            )
+        if not os.path.isfile(inputs):
+            raise Exception(
+                f"The file of input path not exist! Please check input: {inputs}"
+            )
+        self.capture = cv2.VideoCapture(inputs)
+        self.file_name = os.path.split(inputs)[-1]
+
+    def get_info(self):
+        info = {}
+        width = int(self.capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(self.capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        fourcc = cv2.VideoWriter_fourcc(* 'mp4v')
+        info["file_name"] = self.file_name
+        info["fps"] = 30
+        info["shape"] = (width, height)
+        info["fourcc"] = cv2.VideoWriter_fourcc(* 'mp4v')
+        return info
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        ret, frame = self.capture.read()
+        if not ret:
+            raise StopIteration
+        return frame, self.file_name
+
+
+class ImageWriter(object):
+    def __init__(self, output_dir):
+        super().__init__()
+        if output_dir is None:
+            raise Exception(
+                "Please specify the directory of saving prediction results by --output."
+            )
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+        self.output_dir = output_dir
+
+    def write(self, image, file_name):
+        path = os.path.join(self.output_dir, file_name)
+        cv2.imwrite(path, cv2.cvtColor(image, cv2.COLOR_RGB2BGR))
+
+
+class VideoWriter(object):
+    def __init__(self, output_dir, video_info):
+        super().__init__()
+        if output_dir is None:
+            raise Exception(
+                "Please specify the directory of saving prediction results by --output."
+            )
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+        output_path = os.path.join(output_dir, video_info["file_name"])
+        fourcc = cv2.VideoWriter_fourcc(* 'mp4v')
+        self.writer = cv2.VideoWriter(output_path, video_info["fourcc"],
+                                      video_info["fps"], video_info["shape"])
+
+    def write(self, frame, file_name):
+        self.writer.write(frame)
+
+    def __del__(self):
+        if hasattr(self, "writer"):
+            self.writer.release()
+
+
+class BasePredictor(object):
+    def __init__(self, predictor_config):
+        super().__init__()
+        self.predictor_config = predictor_config
+        self.predictor, self.input_names, self.output_names = self.load_predictor(
+            predictor_config["model_file"], predictor_config["params_file"])
+
+    def load_predictor(self, model_file, params_file):
+        config = Config(model_file, params_file)
+        if self.predictor_config["use_gpu"]:
+            config.enable_use_gpu(200, 0)
+            config.switch_ir_optim(True)
+        else:
+            config.disable_gpu()
+            config.set_cpu_math_library_num_threads(self.predictor_config[
+                "cpu_threads"])
+
+            if self.predictor_config["enable_mkldnn"]:
+                try:
+                    # cache 10 different shapes for mkldnn to avoid memory leak
+                    config.set_mkldnn_cache_capacity(10)
+                    config.enable_mkldnn()
+                except Exception as e:
+                    logging.error(
+                        "The current environment does not support `mkldnn`, so disable mkldnn."
+                    )
+        config.disable_glog_info()
+        config.enable_memory_optim()
+        # use zero copy
+        config.switch_use_feed_fetch_ops(False)
+        predictor = create_predictor(config)
+        input_names = predictor.get_input_names()
+        output_names = predictor.get_output_names()
+        return predictor, input_names, output_names
+
+    def preprocess(self):
+        raise NotImplementedError
+
+    def postprocess(self):
+        raise NotImplementedError
+
+    def predict(self, img):
+        raise NotImplementedError
+
+
+class Detector(BasePredictor):
+    def __init__(self, det_config, predictor_config):
+        super().__init__(predictor_config)
+        self.det_config = det_config
+        self.target_size = self.det_config["target_size"]
+        self.thresh = self.det_config["thresh"]
+
+    def preprocess(self, img):
+        resize_h, resize_w = self.target_size
+        img_shape = img.shape
+        img_scale_x = resize_w / img_shape[1]
+        img_scale_y = resize_h / img_shape[0]
+        img = cv2.resize(
+            img, None, None, fx=img_scale_x, fy=img_scale_y, interpolation=1)
+        img = normalize_image(
+            img,
+            scale=1. / 255.,
+            mean=[0.485, 0.456, 0.406],
+            std=[0.229, 0.224, 0.225],
+            order='hwc')
+        img_info = {}
+        img_info["im_shape"] = np.array(
+            img.shape[:2], dtype=np.float32)[np.newaxis, :]
+        img_info["scale_factor"] = np.array(
+            [img_scale_y, img_scale_x], dtype=np.float32)[np.newaxis, :]
+
+        img = img.transpose((2, 0, 1)).copy()
+        img_info["image"] = img[np.newaxis, :, :, :]
+        return img_info
+
+    def postprocess(self, np_boxes):
+        expect_boxes = (np_boxes[:, 1] > self.thresh) & (np_boxes[:, 0] > -1)
+        return np_boxes[expect_boxes, :]
+
+    def predict(self, img):
+        inputs = self.preprocess(img)
+        for input_name in self.input_names:
+            input_tensor = self.predictor.get_input_handle(input_name)
+            input_tensor.copy_from_cpu(inputs[input_name])
+        self.predictor.run()
+        output_tensor = self.predictor.get_output_handle(self.output_names[0])
+        np_boxes = output_tensor.copy_to_cpu()
+        # boxes_num = self.detector.get_output_handle(self.detector_output_names[1])
+        # np_boxes_num = boxes_num.copy_to_cpu()
+        box_list = self.postprocess(np_boxes)
+        return box_list
+
+class FaceDetector(object):
+    def __init__(self, args, print_info=True):
+        super().__init__()
+        if print_info:
+            print_config(args)
+
+        self.font_path = os.path.join(
+            os.path.abspath(os.path.dirname(__file__)),
+            "SourceHanSansCN-Medium.otf")
+        self.args = args
+
+        predictor_config = {
+            "use_gpu": args.use_gpu,
+            "enable_mkldnn": args.enable_mkldnn,
+            "cpu_threads": args.cpu_threads
+        }
+
+        model_file_path, params_file_path = check_model_file(
+            args.det_model)
+        det_config = {"thresh": args.det_thresh, "target_size": [640, 640]}
+        predictor_config["model_file"] = model_file_path
+        predictor_config["params_file"] = params_file_path
+        self.det_predictor = Detector(det_config, predictor_config)
+        self.color_map = ColorMap(100)
+
+    def preprocess(self, img):
+        img = img.astype(np.float32, copy=False)
+        return img
+
+    def draw(self, img, box_list, labels):
+        self.color_map.update(labels)
+        im = Image.fromarray(img)
+        draw = ImageDraw.Draw(im)
+
+        for i, dt in enumerate(box_list):
+            bbox, score = dt[2:], dt[1]
+            label = labels[i]
+            color = tuple(self.color_map[label])
+
+            xmin, ymin, xmax, ymax = bbox
+
+            font_size = max(int((xmax - xmin) // 6), 10)
+            font = ImageFont.truetype(self.font_path, font_size)
+
+            text = "{} {:.4f}".format(label, score)
+            th = sum(font.getmetrics())
+            tw = font.getsize(text)[0]
+            start_y = max(0, ymin - th)
+
+            draw.rectangle(
+                [(xmin, start_y), (xmin + tw + 1, start_y + th)], fill=color)
+            draw.text(
+                (xmin + 1, start_y),
+                text,
+                fill=(255, 255, 255),
+                font=font,
+                anchor="la")
+            draw.rectangle(
+                [(xmin, ymin), (xmax, ymax)], width=2, outline=color)
+        return np.array(im)
+
+    def predict_np_img(self, img):
+        input_img = self.preprocess(img)
+        box_list = None
+        np_feature = None
+        if hasattr(self, "det_predictor"):
+            box_list = self.det_predictor.predict(input_img)
+        return box_list, np_feature
+
+    def init_reader_writer(self, input_data):
+        if isinstance(input_data, np.ndarray):
+            self.input_reader = ImageReader(input_data)
+            if hasattr(self, "det_predictor"):
+                self.output_writer = ImageWriter(self.args.output)
+        elif isinstance(input_data, str):
+            if input_data.endswith('mp4'):
+                self.input_reader = VideoReader(input_data)
+                info = self.input_reader.get_info()
+                self.output_writer = VideoWriter(self.args.output, info)
+            else:
+                self.input_reader = ImageReader(input_data)
+                if hasattr(self, "det_predictor"):
+                    self.output_writer = ImageWriter(self.args.output)
+        else:
+            raise Exception(
+                f"The input data error. Only support path of image or video(.mp4) and dirctory that include images."
+            )
+
+    def predict(self, input_data, print_info=False):
+        """Predict input_data.
+
+        Args:
+            input_data (str | NumPy.array): The path of image, or the derectory including images, or the image data in NumPy.array format.
+            print_info (bool, optional): Wheather to print the prediction results. Defaults to False.
+
+        Yields:
+            dict: {
+                "box_list": The prediction results of detection.
+                "features": The output of recognition.
+                "labels": The results of retrieval.
+                }
+        """
+        self.init_reader_writer(input_data)
+        for img, file_name in self.input_reader:
+            if img is None:
+                logging.warning(f"Error in reading img {file_name}! Ignored.")
+                continue
+            box_list, np_feature = self.predict_np_img(img)
+            labels = ["face"] * len(box_list)
+            if box_list is not None:
+                result = self.draw(img, box_list, labels=labels)
+                self.output_writer.write(result, file_name)
+            if print_info:
+                logging.info(f"File: {file_name}, predict label(s): {labels}")
+            yield {
+                "box_list": box_list,
+                "features": np_feature,
+                "labels": labels
+            }
+        logging.info(f"Predict complete!")
+
+
+# for CLI
+def main(args=None):
+    logging.basicConfig(level=logging.INFO)
+
+    args = parser().parse_args()
+    predictor = FaceDetector(args)
+    res = predictor.predict(args.input, print_info=True)
+    for _ in res:
+        pass
+
+
+if __name__ == "__main__":
+    main()
diff --git a/RetinaFace/Makefile b/detection/retinaface/Makefile
similarity index 100%
rename from RetinaFace/Makefile
rename to detection/retinaface/Makefile
diff --git a/detection/retinaface/README.md b/detection/retinaface/README.md
new file mode 100644
index 000000000..6665d3b12
--- /dev/null
+++ b/detection/retinaface/README.md
@@ -0,0 +1,86 @@
+# RetinaFace Face Detector
+
+## Introduction
+
+RetinaFace is a practical single-stage [SOTA](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html) face detector which is initially introduced in [arXiv technical report](https://arxiv.org/abs/1905.00641) and then accepted by [CVPR 2020](https://openaccess.thecvf.com/content_CVPR_2020/html/Deng_RetinaFace_Single-Shot_Multi-Level_Face_Localisation_in_the_Wild_CVPR_2020_paper.html).
+
+![demoimg1](https://insightface.ai/assets/img/github/11513D05.jpg)
+
+![demoimg2](https://insightface.ai/assets/img/github/widerfacevaltest.png)
+
+## Data
+
+1. Download our annotations (face bounding boxes & five facial landmarks) from [baidu cloud](https://pan.baidu.com/s/1Laby0EctfuJGgGMgRRgykA) or [gdrive](https://drive.google.com/file/d/1BbXxIiY-F74SumCNG6iwmJJ5K3heoemT/view?usp=sharing)
+
+2. Download the [WIDERFACE](http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html) dataset.
+
+3. Organise the dataset directory under ``insightface/RetinaFace/`` as follows:
+
+```Shell
+  data/retinaface/
+    train/
+      images/
+      label.txt
+    val/
+      images/
+      label.txt
+    test/
+      images/
+      label.txt
+```
+
+## Install
+
+1. Install MXNet with GPU support.
+2. Install Deformable Convolution V2 operator from [Deformable-ConvNets](https://github.com/msracver/Deformable-ConvNets) if you use the DCN based backbone.
+3. Type ``make`` to build cxx tools.
+
+## Training
+
+Please check ``train.py`` for training.
+
+1. Copy ``rcnn/sample_config.py`` to ``rcnn/config.py``
+2. Download ImageNet pretrained models and put them into ``model/``(these models are not for detection testing/inferencing but training and parameters initialization). 
+
+    ImageNet ResNet50 ([baidu cloud](https://pan.baidu.com/s/1WAkU9ZA_j-OmzO-sdk9whA) and [googledrive](https://drive.google.com/file/d/1ibQOCG4eJyTrlKAJdnioQ3tyGlnbSHjy/view?usp=sharing)). 
+
+    ImageNet ResNet152 ([baidu cloud](https://pan.baidu.com/s/1nzQ6CzmdKFzg8bM8ChZFQg) and [googledrive](https://drive.google.com/file/d/1FEjeiIB4u-XBYdASgkyx78pFybrlKUA4/view?usp=sharing)).
+
+3. Start training with ``CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --prefix ./model/retina --network resnet``.  
+Before training, you can check the ``resnet`` network configuration (e.g. pretrained model path, anchor setting and learning rate policy etc..) in ``rcnn/config.py``.
+4. We have two predefined network settings named ``resnet``(for medium and large models) and ``mnet``(for lightweight models).
+
+## Testing
+
+Please check ``test.py`` for testing.
+
+## RetinaFace Pretrained Models
+
+Pretrained Model: RetinaFace-R50 ([baidu cloud](https://pan.baidu.com/s/1C6nKq122gJxRhb37vK0_LQ) or [googledrive](https://drive.google.com/file/d/1_DKgGxQWqlTqe78pw0KavId9BIMNUWfu/view?usp=sharing)) is a medium size model with ResNet50 backbone.
+It can output face bounding boxes and five facial landmarks in a single forward pass.
+
+WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4. 
+
+To avoid the confliction with the WiderFace Challenge (ICCV 2019), we postpone the release time of our best model.
+
+## Third-party
+
+[yangfly](https://github.com/yangfly): RetinaFace-MobileNet0.25 ([baidu cloud](https://pan.baidu.com/s/1P1ypO7VYUbNAezdvLm2m9w):nzof).
+WiderFace validation mAP: Hard 82.5. (model size: 1.68Mb) 
+
+[clancylian](https://github.com/clancylian/retinaface): C++ version
+
+RetinaFace in [modelscope](https://modelscope.cn/models/damo/cv_resnet50_face-detection_retinaface/summary)
+
+## References
+
+```  
+@inproceedings{Deng2020CVPR,
+title = {RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild},
+author = {Deng, Jiankang and Guo, Jia and Ververas, Evangelos and Kotsia, Irene and Zafeiriou, Stefanos},
+booktitle = {CVPR},
+year = {2020}
+}
+```
+
+
diff --git a/RetinaFace/rcnn/tools/__init__.py b/detection/retinaface/rcnn/PY_OP/__init__.py
old mode 100644
new mode 100755
similarity index 100%
rename from RetinaFace/rcnn/tools/__init__.py
rename to detection/retinaface/rcnn/PY_OP/__init__.py
diff --git a/detection/retinaface/rcnn/PY_OP/cascade_refine.py b/detection/retinaface/rcnn/PY_OP/cascade_refine.py
new file mode 100644
index 000000000..e3c6556fa
--- /dev/null
+++ b/detection/retinaface/rcnn/PY_OP/cascade_refine.py
@@ -0,0 +1,518 @@
+from __future__ import print_function
+import sys
+import mxnet as mx
+import numpy as np
+import datetime
+from distutils.util import strtobool
+from ..config import config, generate_config
+from ..processing.generate_anchor import generate_anchors, anchors_plane
+from ..processing.bbox_transform import bbox_overlaps, bbox_transform, landmark_transform
+
+STAT = {0: 0}
+STEP = 28800
+
+
+class CascadeRefineOperator(mx.operator.CustomOp):
+    def __init__(self, stride=0, network='', dataset='', prefix=''):
+        super(CascadeRefineOperator, self).__init__()
+        self.stride = int(stride)
+        self.prefix = prefix
+        generate_config(network, dataset)
+        self.mode = config.TRAIN.OHEM_MODE  #0 for random 10:245, 1 for 10:246, 2 for 10:30, mode 1 for default
+        stride = self.stride
+        sstride = str(stride)
+        base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
+        allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
+        ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
+        scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
+        base_anchors = generate_anchors(base_size=base_size,
+                                        ratios=list(ratios),
+                                        scales=np.array(scales,
+                                                        dtype=np.float32),
+                                        stride=stride,
+                                        dense_anchor=config.DENSE_ANCHOR)
+        num_anchors = base_anchors.shape[0]
+        feat_height, feat_width = config.SCALES[0][
+            0] // self.stride, config.SCALES[0][0] // self.stride
+        feat_stride = self.stride
+
+        A = num_anchors
+        K = feat_height * feat_width
+        self.A = A
+
+        all_anchors = anchors_plane(feat_height, feat_width, feat_stride,
+                                    base_anchors)
+        all_anchors = all_anchors.reshape((K * A, 4))
+        self.ori_anchors = all_anchors
+        self.nbatch = 0
+        global STAT
+        for k in config.RPN_FEAT_STRIDE:
+            STAT[k] = [0, 0, 0]
+
+    def apply_bbox_pred(self, bbox_pred, ind=None):
+        box_deltas = bbox_pred
+        box_deltas[:, 0::4] = box_deltas[:, 0::4] * config.TRAIN.BBOX_STDS[0]
+        box_deltas[:, 1::4] = box_deltas[:, 1::4] * config.TRAIN.BBOX_STDS[1]
+        box_deltas[:, 2::4] = box_deltas[:, 2::4] * config.TRAIN.BBOX_STDS[2]
+        box_deltas[:, 3::4] = box_deltas[:, 3::4] * config.TRAIN.BBOX_STDS[3]
+        if ind is None:
+            boxes = self.ori_anchors
+        else:
+            boxes = self.ori_anchors[ind]
+        #print('in apply',self.stride, box_deltas.shape, boxes.shape)
+
+        widths = boxes[:, 2] - boxes[:, 0] + 1.0
+        heights = boxes[:, 3] - boxes[:, 1] + 1.0
+        ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+        ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+
+        dx = box_deltas[:, 0:1]
+        dy = box_deltas[:, 1:2]
+        dw = box_deltas[:, 2:3]
+        dh = box_deltas[:, 3:4]
+
+        pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
+        pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
+        pred_w = np.exp(dw) * widths[:, np.newaxis]
+        pred_h = np.exp(dh) * heights[:, np.newaxis]
+
+        pred_boxes = np.zeros(box_deltas.shape)
+        # x1
+        pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
+        # y1
+        pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
+        # x2
+        pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
+        # y2
+        pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
+        return pred_boxes
+
+    def assign_anchor_fpn(self,
+                          gt_label,
+                          anchors,
+                          landmark=False,
+                          prefix='face'):
+        IOU = config.TRAIN.CASCADE_OVERLAP
+
+        gt_boxes = gt_label['gt_boxes']
+        #_label = gt_label['gt_label']
+        # clean up boxes
+        #nonneg = np.where(_label[:] != -1)[0]
+        #gt_boxes = gt_boxes[nonneg]
+        if landmark:
+            gt_landmarks = gt_label['gt_landmarks']
+            #gt_landmarks = gt_landmarks[nonneg]
+            assert gt_boxes.shape[0] == gt_landmarks.shape[0]
+        #scales = np.array(scales, dtype=np.float32)
+        feat_strides = config.RPN_FEAT_STRIDE
+        bbox_pred_len = 4
+        landmark_pred_len = 10
+        num_anchors = anchors.shape[0]
+        A = self.A
+        total_anchors = num_anchors
+        feat_height, feat_width = config.SCALES[0][
+            0] // self.stride, config.SCALES[0][0] // self.stride
+
+        #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
+
+        # label: 1 is positive, 0 is negative, -1 is dont care
+        labels = np.empty((num_anchors, ), dtype=np.float32)
+        labels.fill(-1)
+        #print('BB', anchors.shape, len(inds_inside))
+        #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
+        #tb = datetime.datetime.now()
+        #self._times[0] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+
+        if gt_boxes.size > 0:
+            # overlap between the anchors and the gt boxes
+            # overlaps (ex, gt)
+            overlaps = bbox_overlaps(anchors.astype(np.float),
+                                     gt_boxes.astype(np.float))
+            argmax_overlaps = overlaps.argmax(axis=1)
+            #print('AAA', argmax_overlaps.shape)
+            max_overlaps = overlaps[np.arange(num_anchors), argmax_overlaps]
+            gt_argmax_overlaps = overlaps.argmax(axis=0)
+            gt_max_overlaps = overlaps[gt_argmax_overlaps,
+                                       np.arange(overlaps.shape[1])]
+            gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
+
+            if not config.TRAIN.RPN_CLOBBER_POSITIVES:
+                # assign bg labels first so that positive labels can clobber them
+                labels[max_overlaps < IOU[0]] = 0
+
+            # fg label: for each gt, anchor with highest overlap
+            if config.TRAIN.RPN_FORCE_POSITIVE:
+                labels[gt_argmax_overlaps] = 1
+
+            # fg label: above threshold IoU
+            labels[max_overlaps >= IOU[1]] = 1
+
+            if config.TRAIN.RPN_CLOBBER_POSITIVES:
+                # assign bg labels last so that negative labels can clobber positives
+                labels[max_overlaps < IOU[0]] = 0
+        else:
+            labels[:] = 0
+        fg_inds = np.where(labels == 1)[0]
+        #print('fg count', len(fg_inds))
+
+        # subsample positive labels if we have too many
+        if config.TRAIN.RPN_ENABLE_OHEM == 0:
+            fg_inds = np.where(labels == 1)[0]
+            num_fg = int(config.TRAIN.RPN_FG_FRACTION *
+                         config.TRAIN.RPN_BATCH_SIZE)
+            if len(fg_inds) > num_fg:
+                disable_inds = npr.choice(fg_inds,
+                                          size=(len(fg_inds) - num_fg),
+                                          replace=False)
+                if DEBUG:
+                    disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
+                labels[disable_inds] = -1
+
+            # subsample negative labels if we have too many
+            num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
+            bg_inds = np.where(labels == 0)[0]
+            if len(bg_inds) > num_bg:
+                disable_inds = npr.choice(bg_inds,
+                                          size=(len(bg_inds) - num_bg),
+                                          replace=False)
+                if DEBUG:
+                    disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+                labels[disable_inds] = -1
+
+            #fg_inds = np.where(labels == 1)[0]
+            #num_fg = len(fg_inds)
+            #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
+
+            #bg_inds = np.where(labels == 0)[0]
+            #if len(bg_inds) > num_bg:
+            #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
+            #    if DEBUG:
+            #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+            #    labels[disable_inds] = -1
+        else:
+            fg_inds = np.where(labels == 1)[0]
+            num_fg = len(fg_inds)
+            bg_inds = np.where(labels == 0)[0]
+            num_bg = len(bg_inds)
+
+        #print('anchor stat', num_fg, num_bg)
+
+        bbox_targets = np.zeros((num_anchors, bbox_pred_len), dtype=np.float32)
+        if gt_boxes.size > 0:
+            #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
+            bbox_targets[:, :] = bbox_transform(anchors,
+                                                gt_boxes[argmax_overlaps, :])
+            #bbox_targets[:,4] = gt_blur
+        #tb = datetime.datetime.now()
+        #self._times[1] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+
+        bbox_weights = np.zeros((num_anchors, bbox_pred_len), dtype=np.float32)
+        #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
+        bbox_weights[labels == 1, 0:4] = 1.0
+        if bbox_pred_len > 4:
+            bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
+
+        if landmark:
+            landmark_targets = np.zeros((num_anchors, landmark_pred_len),
+                                        dtype=np.float32)
+            landmark_weights = np.zeros((num_anchors, landmark_pred_len),
+                                        dtype=np.float32)
+            #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
+            if landmark_pred_len == 10:
+                landmark_weights[labels == 1, :] = 1.0
+            elif landmark_pred_len == 15:
+                v = [1.0, 1.0, 0.1] * 5
+                assert len(v) == 15
+                landmark_weights[labels == 1, :] = np.array(v)
+            else:
+                assert False
+            #TODO here
+            if gt_landmarks.size > 0:
+                #print('AAA',argmax_overlaps)
+                a_landmarks = gt_landmarks[argmax_overlaps, :, :]
+                landmark_targets[:] = landmark_transform(anchors, a_landmarks)
+                invalid = np.where(a_landmarks[:, 0, 2] < 0.0)[0]
+                #assert len(invalid)==0
+                #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
+                landmark_weights[invalid, :] = 0.0
+        #tb = datetime.datetime.now()
+        #self._times[2] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+        bbox_targets[:,
+                     0::4] = bbox_targets[:, 0::4] / config.TRAIN.BBOX_STDS[0]
+        bbox_targets[:,
+                     1::4] = bbox_targets[:, 1::4] / config.TRAIN.BBOX_STDS[1]
+        bbox_targets[:,
+                     2::4] = bbox_targets[:, 2::4] / config.TRAIN.BBOX_STDS[2]
+        bbox_targets[:,
+                     3::4] = bbox_targets[:, 3::4] / config.TRAIN.BBOX_STDS[3]
+
+        #print('CC', anchors.shape, len(inds_inside))
+        label = {}
+        _label = labels.reshape(
+            (1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
+        _label = _label.reshape((1, A * feat_height * feat_width))
+        bbox_target = bbox_targets.reshape(
+            (1, feat_height * feat_width,
+             A * bbox_pred_len)).transpose(0, 2, 1)
+        bbox_weight = bbox_weights.reshape(
+            (1, feat_height * feat_width, A * bbox_pred_len)).transpose(
+                (0, 2, 1))
+        label['%s_label' % prefix] = _label[0]
+        label['%s_bbox_target' % prefix] = bbox_target[0]
+        label['%s_bbox_weight' % prefix] = bbox_weight[0]
+        if landmark:
+            landmark_target = landmark_target.reshape(
+                (1, feat_height * feat_width,
+                 A * landmark_pred_len)).transpose(0, 2, 1)
+            landmark_target /= config.TRAIN.LANDMARK_STD
+            landmark_weight = landmark_weight.reshape(
+                (1, feat_height * feat_width,
+                 A * landmark_pred_len)).transpose((0, 2, 1))
+            label['%s_landmark_target' % prefix] = landmark_target[0]
+            label['%s_landmark_weight' % prefix] = landmark_weight[0]
+
+        return label
+
+    def forward(self, is_train, req, in_data, out_data, aux):
+        self.nbatch += 1
+        ta = datetime.datetime.now()
+        global STAT
+        A = config.NUM_ANCHORS
+
+        cls_label_t0 = in_data[0].asnumpy()  #BS, AHW
+        cls_score_t0 = in_data[1].asnumpy()  #BS, C, AHW
+        cls_score = in_data[2].asnumpy()  #BS, C, AHW
+        #labels_raw = in_data[1].asnumpy() #BS, ANCHORS
+        bbox_pred_t0 = in_data[3].asnumpy()  #BS, AC, HW
+        bbox_target_t0 = in_data[4].asnumpy()  #BS, AC, HW
+        cls_label_raw = in_data[5].asnumpy()  #BS, AHW
+        gt_boxes = in_data[6].asnumpy()  #BS, N, C=4+1
+        #imgs = in_data[7].asnumpy().astype(np.uint8)
+
+        batch_size = cls_score.shape[0]
+        num_anchors = cls_score.shape[2]
+        #print('in cas', cls_score.shape, bbox_target.shape)
+
+        labels_out = np.zeros(shape=(batch_size, num_anchors),
+                              dtype=np.float32)
+        bbox_target_out = np.zeros(shape=bbox_target_t0.shape,
+                                   dtype=np.float32)
+        anchor_weight = np.zeros((batch_size, num_anchors, 1),
+                                 dtype=np.float32)
+        valid_count = np.zeros((batch_size, 1), dtype=np.float32)
+
+        bbox_pred_t0 = bbox_pred_t0.transpose((0, 2, 1))
+        bbox_pred_t0 = bbox_pred_t0.reshape(
+            (bbox_pred_t0.shape[0], -1, 4))  #BS, H*W*A, C
+        bbox_target_t0 = bbox_target_t0.transpose((0, 2, 1))
+        bbox_target_t0 = bbox_target_t0.reshape(
+            (bbox_target_t0.shape[0], -1, 4))
+
+        #print('anchor_weight', anchor_weight.shape)
+
+        #assert labels.shape[0]==1
+        #assert cls_score.shape[0]==1
+        #assert bbox_weight.shape[0]==1
+        #print('shape', cls_score.shape, labels.shape, file=sys.stderr)
+        #print('bbox_weight 0', bbox_weight.shape, file=sys.stderr)
+        #bbox_weight = np.zeros( (labels_raw.shape[0], labels_raw.shape[1], 4), dtype=np.float32)
+        _stat = [0, 0, 0]
+        SEL_TOPK = config.TRAIN.RPN_BATCH_SIZE
+        FAST = False
+        for ibatch in range(batch_size):
+            #bgr = imgs[ibatch].transpose( (1,2,0) )[:,:,::-1]
+
+            if not FAST:
+                _gt_boxes = gt_boxes[ibatch]  #N, 4+1
+                _gtind = len(np.where(_gt_boxes[:, 4] >= 0)[0])
+                #print('gt num', _gtind)
+                _gt_boxes = _gt_boxes[0:_gtind, :]
+
+                #anchors_t1 = self.ori_anchors.copy()
+                #_cls_label_raw = cls_label_raw[ibatch] #AHW
+                #_cls_label_raw = _cls_label_raw.reshape( (A, -1) ).transpose( (1,0) ).reshape( (-1,) ) #HWA
+                #fg_ind_raw = np.where(_cls_label_raw>0)[0]
+                #_bbox_target_t0 = bbox_target_t0[ibatch][fg_ind_raw]
+                #_bbox_pred_t0 = bbox_pred_t0[ibatch][fg_ind_raw]
+                #anchors_t1_pos = self.apply_bbox_pred(_bbox_pred_t0, ind=fg_ind_raw)
+                #anchors_t1[fg_ind_raw,:] = anchors_t1_pos
+
+                anchors_t1 = self.apply_bbox_pred(bbox_pred_t0[ibatch])
+                assert anchors_t1.shape[0] == self.ori_anchors.shape[0]
+
+                #for i in range(_gt_boxes.shape[0]):
+                #  box = _gt_boxes[i].astype(np.int)
+                #  print('%d: gt%d'%(self.nbatch, i), box)
+                #  #color = (0,0,255)
+                #  #cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
+                #for i in range(anchors_t1.shape[0]):
+                #  box1 = self.ori_anchors[i].astype(np.int)
+                #  box2 = anchors_t1[i].astype(np.int)
+                #  print('%d %d: anchorscompare %d'%(self.nbatch, self.stride, i), box1, box2)
+                #color = (255,255,0)
+                #cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
+                #filename = "./debug/%d_%d_%d.jpg"%(self.nbatch, ibatch, stride)
+                #cv2.imwrite(filename, img)
+                #print(filename)
+                #gt_label = {'gt_boxes': gt_anchors, 'gt_label' : labels_raw[ibatch]}
+                gt_label = {'gt_boxes': _gt_boxes}
+                new_label_dict = self.assign_anchor_fpn(gt_label,
+                                                        anchors_t1,
+                                                        False,
+                                                        prefix=self.prefix)
+                labels = new_label_dict['%s_label' % self.prefix]  #AHW
+                new_bbox_target = new_label_dict['%s_bbox_target' %
+                                                 self.prefix]  #AC,HW
+                #print('assign ret', labels.shape, new_bbox_target.shape)
+                _anchor_weight = np.zeros((num_anchors, 1), dtype=np.float32)
+                fg_score = cls_score[ibatch, 1, :] - cls_score[ibatch, 0, :]
+                fg_inds = np.where(labels > 0)[0]
+                num_fg = int(config.TRAIN.RPN_FG_FRACTION *
+                             config.TRAIN.RPN_BATCH_SIZE)
+                origin_num_fg = len(fg_inds)
+                #continue
+                #print('cas fg', len(fg_inds), num_fg, file=sys.stderr)
+                if len(fg_inds) > num_fg:
+                    if self.mode == 0:
+                        disable_inds = np.random.choice(fg_inds,
+                                                        size=(len(fg_inds) -
+                                                              num_fg),
+                                                        replace=False)
+                        labels[disable_inds] = -1
+                    else:
+                        pos_ohem_scores = fg_score[fg_inds]
+                        order_pos_ohem_scores = pos_ohem_scores.ravel(
+                        ).argsort()
+                        sampled_inds = fg_inds[order_pos_ohem_scores[:num_fg]]
+                        labels[fg_inds] = -1
+                        labels[sampled_inds] = 1
+
+                n_fg = np.sum(labels > 0)
+                fg_inds = np.where(labels > 0)[0]
+                num_bg = config.TRAIN.RPN_BATCH_SIZE - n_fg
+                if self.mode == 2:
+                    num_bg = max(
+                        48, n_fg * int(1.0 / config.TRAIN.RPN_FG_FRACTION - 1))
+
+                bg_inds = np.where(labels == 0)[0]
+                origin_num_bg = len(bg_inds)
+                if num_bg == 0:
+                    labels[bg_inds] = -1
+                elif len(bg_inds) > num_bg:
+                    # sort ohem scores
+
+                    if self.mode == 0:
+                        disable_inds = np.random.choice(bg_inds,
+                                                        size=(len(bg_inds) -
+                                                              num_bg),
+                                                        replace=False)
+                        labels[disable_inds] = -1
+                    else:
+                        neg_ohem_scores = fg_score[bg_inds]
+                        order_neg_ohem_scores = neg_ohem_scores.ravel(
+                        ).argsort()[::-1]
+                        sampled_inds = bg_inds[order_neg_ohem_scores[:num_bg]]
+                        #print('sampled_inds_bg', sampled_inds, file=sys.stderr)
+                        labels[bg_inds] = -1
+                        labels[sampled_inds] = 0
+
+                if n_fg > 0:
+                    order0_labels = labels.reshape((1, A, -1)).transpose(
+                        (0, 2, 1)).reshape((-1, ))
+                    bbox_fg_inds = np.where(order0_labels > 0)[0]
+                    #print('bbox_fg_inds, order0 ', bbox_fg_inds, file=sys.stderr)
+                    _anchor_weight[bbox_fg_inds, :] = 1.0
+                anchor_weight[ibatch] = _anchor_weight
+                valid_count[ibatch][0] = n_fg
+                labels_out[ibatch] = labels
+                #print('labels_out', self.stride, ibatch, labels)
+                bbox_target_out[ibatch] = new_bbox_target
+                #print('cascade stat', self.stride, ibatch, len(labels), len(np.where(labels==1)[0]), len(np.where(labels==0)[0]))
+            else:  #FAST MODE
+                fg_score_t0 = cls_score_t0[ibatch, 1, :] - cls_score_t0[ibatch,
+                                                                        0, :]
+                sort_idx_t0 = np.argsort(
+                    fg_score_t0.flatten())[::-1][0:SEL_TOPK]
+                _bbox_pred_t0 = bbox_pred_t0[ibatch][sort_idx_t0]
+                _bbox_target_t0 = bbox_target_t0[ibatch][sort_idx_t0]
+                #print('SEL fg score:', fg_score_t0[sort_idx[-1]], fg_score_t0[sort_idx[0]])
+                anchors_t0 = self.apply_bbox_pred(_bbox_pred_t0)
+                gt_anchors = self.apply_bbox_pred(_bbox_target_t0)
+                #gt_label = {'gt_boxes': gt_anchors, 'gt_label' : labels_raw[ibatch]}
+                gt_label = {'gt_boxes': gt_anchors}
+                new_label_dict = self.assign_anchor_fpn(gt_label,
+                                                        anchors_t0,
+                                                        False,
+                                                        prefix=self.prefix)
+                labels = new_label_dict['%s_label' % self.prefix]
+                new_bbox_target = new_label_dict['%s_bbox_target' %
+                                                 self.prefix]
+                #print('assign ret', labels.shape, new_bbox_target.shape)
+                _anchor_weight = np.zeros((num_anchors, 1), dtype=np.float32)
+                fg_score = cls_score[ibatch, 1, :] - cls_score[ibatch, 0, :]
+                fg_inds = np.where(labels > 0)[0]
+                _labels = np.empty(shape=labels.shape, dtype=np.float32)
+                _labels.fill(-1)
+                _labels[sort_idx_idx] = labels
+
+                anchor_weight[ibatch] = _anchor_weight
+                valid_count[ibatch][0] = len(fg_inds)
+                labels_out[ibatch] = _labels
+                #print('labels_out', self.stride, ibatch, labels)
+                bbox_target_out[ibatch] = new_bbox_target
+
+        #print('cascade pos stat', self.stride, batch_size, np.sum(valid_count))
+        for ind, val in enumerate(
+            [labels_out, bbox_target_out, anchor_weight, valid_count]):
+            val = mx.nd.array(val)
+            self.assign(out_data[ind], req[ind], val)
+        tb = datetime.datetime.now()
+        #print('cascade forward cost', self.stride, (tb-ta).total_seconds())
+
+    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
+        for i in range(len(in_grad)):
+            self.assign(in_grad[i], req[i], 0)
+
+
+@mx.operator.register('cascade_refine')
+class CascadeRefineProp(mx.operator.CustomOpProp):
+    def __init__(self, stride=0, network='', dataset='', prefix=''):
+        super(CascadeRefineProp, self).__init__(need_top_grad=False)
+        self.stride = stride
+        self.network = network
+        self.dataset = dataset
+        self.prefix = prefix
+
+    def list_arguments(self):
+        #return ['cls_label_t0', 'cls_pred_t0', 'cls_pred', 'bbox_pred_t0', 'bbox_label_t0', 'cls_label_raw', 'cas_gt_boxes', 'cas_img']
+        return [
+            'cls_label_t0', 'cls_pred_t0', 'cls_pred', 'bbox_pred_t0',
+            'bbox_label_t0', 'cls_label_raw', 'cas_gt_boxes'
+        ]
+
+    def list_outputs(self):
+        return [
+            'cls_label_out', 'bbox_label_out', 'anchor_weight_out',
+            'pos_count_out'
+        ]
+
+    def infer_shape(self, in_shape):
+        cls_pred_shape = in_shape[1]
+        bs = cls_pred_shape[0]
+        num_anchors = cls_pred_shape[2]
+        #print('in_rpn_ohem', in_shape[0], in_shape[1], in_shape[2], file=sys.stderr)
+        #print('in_rpn_ohem', labels_shape, anchor_weight_shape)
+        cls_label_shape = [bs, num_anchors]
+
+        return in_shape, \
+               [cls_label_shape, in_shape[4], [bs,num_anchors,1], [bs,1]]
+
+    def create_operator(self, ctx, shapes, dtypes):
+        return CascadeRefineOperator(self.stride, self.network, self.dataset,
+                                     self.prefix)
+
+    def declare_backward_dependency(self, out_grad, in_data, out_data):
+        return []
diff --git a/detection/retinaface/rcnn/PY_OP/rpn_fpn_ohem3.py b/detection/retinaface/rcnn/PY_OP/rpn_fpn_ohem3.py
new file mode 100644
index 000000000..b8f7d462e
--- /dev/null
+++ b/detection/retinaface/rcnn/PY_OP/rpn_fpn_ohem3.py
@@ -0,0 +1,175 @@
+from __future__ import print_function
+import sys
+import mxnet as mx
+import numpy as np
+from distutils.util import strtobool
+from ..config import config, generate_config
+
+STAT = {0: 0}
+STEP = 28800
+
+
+class RPNFPNOHEM3Operator(mx.operator.CustomOp):
+    def __init__(self, stride=0, network='', dataset='', prefix=''):
+        super(RPNFPNOHEM3Operator, self).__init__()
+        self.stride = int(stride)
+        self.prefix = prefix
+        generate_config(network, dataset)
+        self.mode = config.TRAIN.OHEM_MODE  #0 for random 10:245, 1 for 10:246, 2 for 10:30, mode 1 for default
+        global STAT
+        for k in config.RPN_FEAT_STRIDE:
+            STAT[k] = [0, 0, 0]
+
+    def forward(self, is_train, req, in_data, out_data, aux):
+        global STAT
+
+        cls_score = in_data[0].asnumpy()  #BS, 2, ANCHORS
+        labels_raw = in_data[1].asnumpy()  # BS, ANCHORS
+
+        A = config.NUM_ANCHORS
+        anchor_weight = np.zeros((labels_raw.shape[0], labels_raw.shape[1], 1),
+                                 dtype=np.float32)
+        valid_count = np.zeros((labels_raw.shape[0], 1), dtype=np.float32)
+        #print('anchor_weight', anchor_weight.shape)
+
+        #assert labels.shape[0]==1
+        #assert cls_score.shape[0]==1
+        #assert bbox_weight.shape[0]==1
+        #print('shape', cls_score.shape, labels.shape, file=sys.stderr)
+        #print('bbox_weight 0', bbox_weight.shape, file=sys.stderr)
+        #bbox_weight = np.zeros( (labels_raw.shape[0], labels_raw.shape[1], 4), dtype=np.float32)
+        _stat = [0, 0, 0]
+        for ibatch in range(labels_raw.shape[0]):
+            _anchor_weight = np.zeros((labels_raw.shape[1], 1),
+                                      dtype=np.float32)
+            labels = labels_raw[ibatch]
+            fg_score = cls_score[ibatch, 1, :] - cls_score[ibatch, 0, :]
+
+            fg_inds = np.where(labels > 0)[0]
+            num_fg = int(config.TRAIN.RPN_FG_FRACTION *
+                         config.TRAIN.RPN_BATCH_SIZE)
+            origin_num_fg = len(fg_inds)
+            #print(len(fg_inds), num_fg, file=sys.stderr)
+            if len(fg_inds) > num_fg:
+                if self.mode == 0:
+                    disable_inds = np.random.choice(fg_inds,
+                                                    size=(len(fg_inds) -
+                                                          num_fg),
+                                                    replace=False)
+                    labels[disable_inds] = -1
+                else:
+                    pos_ohem_scores = fg_score[fg_inds]
+                    order_pos_ohem_scores = pos_ohem_scores.ravel().argsort()
+                    sampled_inds = fg_inds[order_pos_ohem_scores[:num_fg]]
+                    labels[fg_inds] = -1
+                    labels[sampled_inds] = 1
+
+            n_fg = np.sum(labels > 0)
+            fg_inds = np.where(labels > 0)[0]
+            num_bg = config.TRAIN.RPN_BATCH_SIZE - n_fg
+            if self.mode == 2:
+                num_bg = max(
+                    48, n_fg * int(1.0 / config.TRAIN.RPN_FG_FRACTION - 1))
+
+            bg_inds = np.where(labels == 0)[0]
+            origin_num_bg = len(bg_inds)
+            if num_bg == 0:
+                labels[bg_inds] = -1
+            elif len(bg_inds) > num_bg:
+                # sort ohem scores
+
+                if self.mode == 0:
+                    disable_inds = np.random.choice(bg_inds,
+                                                    size=(len(bg_inds) -
+                                                          num_bg),
+                                                    replace=False)
+                    labels[disable_inds] = -1
+                else:
+                    neg_ohem_scores = fg_score[bg_inds]
+                    order_neg_ohem_scores = neg_ohem_scores.ravel().argsort(
+                    )[::-1]
+                    sampled_inds = bg_inds[order_neg_ohem_scores[:num_bg]]
+                    #print('sampled_inds_bg', sampled_inds, file=sys.stderr)
+                    labels[bg_inds] = -1
+                    labels[sampled_inds] = 0
+
+            if n_fg > 0:
+                order0_labels = labels.reshape((1, A, -1)).transpose(
+                    (0, 2, 1)).reshape((-1, ))
+                bbox_fg_inds = np.where(order0_labels > 0)[0]
+                #print('bbox_fg_inds, order0 ', bbox_fg_inds, file=sys.stderr)
+                _anchor_weight[bbox_fg_inds, :] = 1.0
+            anchor_weight[ibatch] = _anchor_weight
+            valid_count[ibatch][0] = n_fg
+
+            #if self.prefix=='face':
+            #  #print('fg-bg', self.stride, n_fg, num_bg)
+            #  STAT[0]+=1
+            #  STAT[self.stride][0] += config.TRAIN.RPN_BATCH_SIZE
+            #  STAT[self.stride][1] += n_fg
+            #  STAT[self.stride][2] += np.sum(fg_score[fg_inds]>=0)
+            #  #_stat[0] += config.TRAIN.RPN_BATCH_SIZE
+            #  #_stat[1] += n_fg
+            #  #_stat[2] += np.sum(fg_score[fg_inds]>=0)
+            #  #print('stride num_fg', self.stride, n_fg, file=sys.stderr)
+            #  #ACC[self.stride] += np.sum(fg_score[fg_inds]>=0)
+            #  #x = float(labels_raw.shape[0]*len(config.RPN_FEAT_STRIDE))
+            #  x = 1.0
+            #  if STAT[0]%STEP==0:
+            #    _str = ['STAT']
+            #    STAT[0] = 0
+            #    for k in config.RPN_FEAT_STRIDE:
+            #      acc = float(STAT[k][2])/STAT[k][1]
+            #      acc0 = float(STAT[k][1])/STAT[k][0]
+            #      #_str.append("%d: all-fg(%d, %d, %.4f), fg-fgcorrect(%d, %d, %.4f)"%(k,STAT[k][0], STAT[k][1], acc0, STAT[k][1], STAT[k][2], acc))
+            #      _str.append("%d: (%d, %d, %.4f)"%(k, STAT[k][1], STAT[k][2], acc))
+            #      STAT[k] = [0,0,0]
+            #    _str = ' | '.join(_str)
+            #    print(_str, file=sys.stderr)
+            #if self.stride==4 and num_fg>0:
+            #  print('_stat_', self.stride, num_fg, num_bg, file=sys.stderr)
+
+        #labels_ohem = mx.nd.array(labels_raw)
+        #anchor_weight = mx.nd.array(anchor_weight)
+        #print('valid_count', self.stride, np.sum(valid_count))
+        #print('_stat', _stat, valid_count)
+
+        for ind, val in enumerate([labels_raw, anchor_weight, valid_count]):
+            val = mx.nd.array(val)
+            self.assign(out_data[ind], req[ind], val)
+
+    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
+        for i in range(len(in_grad)):
+            self.assign(in_grad[i], req[i], 0)
+
+
+@mx.operator.register('rpn_fpn_ohem3')
+class RPNFPNOHEM3Prop(mx.operator.CustomOpProp):
+    def __init__(self, stride=0, network='', dataset='', prefix=''):
+        super(RPNFPNOHEM3Prop, self).__init__(need_top_grad=False)
+        self.stride = stride
+        self.network = network
+        self.dataset = dataset
+        self.prefix = prefix
+
+    def list_arguments(self):
+        return ['cls_score', 'labels']
+
+    def list_outputs(self):
+        return ['labels_ohem', 'anchor_weight', 'valid_count']
+
+    def infer_shape(self, in_shape):
+        labels_shape = in_shape[1]
+        #print('in_rpn_ohem', in_shape[0], in_shape[1], in_shape[2], file=sys.stderr)
+        anchor_weight_shape = [labels_shape[0], labels_shape[1], 1]
+        #print('in_rpn_ohem', labels_shape, anchor_weight_shape)
+
+        return in_shape, \
+               [labels_shape, anchor_weight_shape, [labels_shape[0], 1]]
+
+    def create_operator(self, ctx, shapes, dtypes):
+        return RPNFPNOHEM3Operator(self.stride, self.network, self.dataset,
+                                   self.prefix)
+
+    def declare_backward_dependency(self, out_grad, in_data, out_data):
+        return []
diff --git a/RetinaFace/rcnn/utils/__init__.py b/detection/retinaface/rcnn/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/utils/__init__.py
rename to detection/retinaface/rcnn/__init__.py
diff --git a/RetinaFaceAntiCov/rcnn/processing/__init__.py b/detection/retinaface/rcnn/core/__init__.py
similarity index 100%
rename from RetinaFaceAntiCov/rcnn/processing/__init__.py
rename to detection/retinaface/rcnn/core/__init__.py
diff --git a/detection/retinaface/rcnn/core/callback.py b/detection/retinaface/rcnn/core/callback.py
new file mode 100644
index 000000000..729a3920e
--- /dev/null
+++ b/detection/retinaface/rcnn/core/callback.py
@@ -0,0 +1,16 @@
+import mxnet as mx
+
+
+def do_checkpoint(prefix, means, stds):
+    def _callback(iter_no, sym, arg, aux):
+        if 'bbox_pred_weight' in arg:
+            arg['bbox_pred_weight_test'] = (arg['bbox_pred_weight'].T *
+                                            mx.nd.array(stds)).T
+            arg['bbox_pred_bias_test'] = arg['bbox_pred_bias'] * mx.nd.array(
+                stds) + mx.nd.array(means)
+        mx.model.save_checkpoint(prefix, iter_no + 1, sym, arg, aux)
+        if 'bbox_pred_weight' in arg:
+            arg.pop('bbox_pred_weight_test')
+            arg.pop('bbox_pred_bias_test')
+
+    return _callback
diff --git a/detection/retinaface/rcnn/core/loader.py b/detection/retinaface/rcnn/core/loader.py
new file mode 100644
index 000000000..1d34d2eb3
--- /dev/null
+++ b/detection/retinaface/rcnn/core/loader.py
@@ -0,0 +1,549 @@
+from __future__ import print_function
+import sys
+import mxnet as mx
+import numpy as np
+import random
+import datetime
+import multiprocessing
+import cv2
+from mxnet.executor_manager import _split_input_slice
+
+from rcnn.config import config
+from rcnn.io.image import tensor_vstack
+from rcnn.io.rpn import get_rpn_testbatch, get_rpn_batch, assign_anchor_fpn, get_crop_batch, AA
+
+
+class CropLoader(mx.io.DataIter):
+    def __init__(self,
+                 feat_sym,
+                 roidb,
+                 batch_size=1,
+                 shuffle=False,
+                 ctx=None,
+                 work_load_list=None,
+                 aspect_grouping=False):
+        """
+        This Iter will provide roi data to Fast R-CNN network
+        :param feat_sym: to infer shape of assign_output
+        :param roidb: must be preprocessed
+        :param batch_size: must divide BATCH_SIZE(128)
+        :param shuffle: bool
+        :param ctx: list of contexts
+        :param work_load_list: list of work load
+        :param aspect_grouping: group images with similar aspects
+        :return: AnchorLoader
+        """
+        super(CropLoader, self).__init__()
+
+        # save parameters as properties
+        self.feat_sym = feat_sym
+        self.roidb = roidb
+        self.batch_size = batch_size
+        self.shuffle = shuffle
+        self.ctx = ctx
+        if self.ctx is None:
+            self.ctx = [mx.cpu()]
+        self.work_load_list = work_load_list
+        #self.feat_stride = feat_stride
+        #self.anchor_scales = anchor_scales
+        #self.anchor_ratios = anchor_ratios
+        #self.allowed_border = allowed_border
+        self.aspect_grouping = aspect_grouping
+        self.feat_stride = config.RPN_FEAT_STRIDE
+
+        # infer properties from roidb
+        self.size = len(roidb)
+        self.index = np.arange(self.size)
+
+        # decide data and label names
+        #self.data_name = ['data']
+        #self.label_name = []
+        #self.label_name.append('label')
+        #self.label_name.append('bbox_target')
+        #self.label_name.append('bbox_weight')
+
+        self.data_name = ['data']
+        #self.label_name = ['label', 'bbox_target', 'bbox_weight']
+        self.label_name = []
+        prefixes = ['face']
+        if config.HEAD_BOX:
+            prefixes.append('head')
+        names = []
+        for prefix in prefixes:
+            names += [
+                prefix + '_label', prefix + '_bbox_target',
+                prefix + '_bbox_weight'
+            ]
+            if prefix == 'face' and config.FACE_LANDMARK:
+                names += [
+                    prefix + '_landmark_target', prefix + '_landmark_weight'
+                ]
+        #names = ['label', 'bbox_weight']
+        for stride in self.feat_stride:
+            for n in names:
+                k = "%s_stride%d" % (n, stride)
+                self.label_name.append(k)
+        if config.CASCADE > 0:
+            self.label_name.append('gt_boxes')
+
+        # status variable for synchronization between get_data and get_label
+        self.cur = 0
+        self.batch = None
+        self.data = None
+        self.label = None
+        # infer shape
+        feat_shape_list = []
+        _data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]),
+                                 max([v[1] for v in config.SCALES])))]
+        _data_shape = dict(_data_shape)
+        for i in range(len(self.feat_stride)):
+            _, feat_shape, _ = self.feat_sym[i].infer_shape(**_data_shape)
+            feat_shape = [int(i) for i in feat_shape[0]]
+            feat_shape_list.append(feat_shape)
+        self.aa = AA(feat_shape_list)
+
+        self._debug = False
+        self._debug_id = 0
+        self._times = [0.0, 0.0, 0.0, 0.0]
+
+        # get first batch to fill in provide_data and provide_label
+        self.reset()
+        self.get_batch()
+
+    @property
+    def provide_data(self):
+        return [(k, v.shape) for k, v in zip(self.data_name, self.data)]
+
+    @property
+    def provide_label(self):
+        return [(k, v.shape) for k, v in zip(self.label_name, self.label)]
+
+    def reset(self):
+        self.cur = 0
+        if self.shuffle:
+            np.random.shuffle(self.index)
+
+    def iter_next(self):
+        return self.cur + self.batch_size <= self.size
+
+    def next(self):
+        if self.iter_next():
+            self.get_batch()
+            self.cur += self.batch_size
+            return mx.io.DataBatch(data=self.data,
+                                   label=self.label,
+                                   pad=self.getpad(),
+                                   index=self.getindex(),
+                                   provide_data=self.provide_data,
+                                   provide_label=self.provide_label)
+        else:
+            raise StopIteration
+
+    def getindex(self):
+        return self.cur / self.batch_size
+
+    def getpad(self):
+        if self.cur + self.batch_size > self.size:
+            return self.cur + self.batch_size - self.size
+        else:
+            return 0
+
+    def infer_shape(self, max_data_shape=None, max_label_shape=None):
+        """ Return maximum data and label shape for single gpu """
+        if max_data_shape is None:
+            max_data_shape = []
+        if max_label_shape is None:
+            max_label_shape = []
+        max_shapes = dict(max_data_shape + max_label_shape)
+        input_batch_size = max_shapes['data'][0]
+        dummy_boxes = np.zeros((0, 5))
+        dummy_info = [[max_shapes['data'][2], max_shapes['data'][3], 1.0]]
+        dummy_label = {'gt_boxes': dummy_boxes}
+        dummy_blur = np.zeros((0, ))
+        dummy_label['gt_blur'] = dummy_blur
+
+        label_dict = {}
+        if config.HEAD_BOX:
+            head_label_dict = self.aa.assign_anchor_fpn(dummy_label,
+                                                        dummy_info,
+                                                        False,
+                                                        prefix='head')
+            label_dict.update(head_label_dict)
+
+        if config.FACE_LANDMARK:
+            dummy_landmarks = np.zeros((0, 5, 3))
+            dummy_label['gt_landmarks'] = dummy_landmarks
+        face_label_dict = self.aa.assign_anchor_fpn(dummy_label,
+                                                    dummy_info,
+                                                    config.FACE_LANDMARK,
+                                                    prefix='face')
+        label_dict.update(face_label_dict)
+        if config.CASCADE > 0:
+            label_dict['gt_boxes'] = np.zeros(
+                (0, config.TRAIN.MAX_BBOX_PER_IMAGE, 5), dtype=np.float32)
+
+        label_list = []
+        for k in self.label_name:
+            label_list.append(label_dict[k])
+        label_shape = [(k, tuple([input_batch_size] + list(v.shape[1:])))
+                       for k, v in zip(self.label_name, label_list)]
+        return max_data_shape, label_shape
+
+    def get_batch(self):
+        # slice roidb
+        cur_from = self.cur
+        cur_to = min(cur_from + self.batch_size, self.size)
+        assert cur_to == cur_from + self.batch_size
+        roidb = [self.roidb[self.index[i]] for i in range(cur_from, cur_to)]
+
+        # decide multi device slice
+        work_load_list = self.work_load_list
+        ctx = self.ctx
+        if work_load_list is None:
+            work_load_list = [1] * len(ctx)
+        assert isinstance(work_load_list, list) and len(work_load_list) == len(ctx), \
+            "Invalid settings for work load. "
+        slices = _split_input_slice(self.batch_size, work_load_list)
+
+        # get testing data for multigpu
+        data_list = []
+        label_list = []
+        for islice in slices:
+            iroidb = [roidb[i] for i in range(islice.start, islice.stop)]
+            data, label = get_crop_batch(iroidb)
+            data_list += data
+            label_list += label
+            #data_list.append(data)
+            #label_list.append(label)
+
+        # pad data first and then assign anchor (read label)
+        #data_tensor = tensor_vstack([batch['data'] for batch in data_list])
+        #for i_card in range(len(data_list)):
+        #    data_list[i_card]['data'] = data_tensor[
+        #                                i_card * config.TRAIN.BATCH_IMAGES:(1 + i_card) * config.TRAIN.BATCH_IMAGES]
+
+        #iiddxx = 0
+        select_stride = 0
+        if config.RANDOM_FEAT_STRIDE:
+            select_stride = random.choice(config.RPN_FEAT_STRIDE)
+
+        for data, label in zip(data_list, label_list):
+            data_shape = {k: v.shape for k, v in data.items()}
+            del data_shape['im_info']
+            feat_shape_list = []
+            for s in range(len(self.feat_stride)):
+                _, feat_shape, _ = self.feat_sym[s].infer_shape(**data_shape)
+                feat_shape = [int(i) for i in feat_shape[0]]
+                feat_shape_list.append(feat_shape)
+            im_info = data['im_info']
+            gt_boxes = label['gt_boxes']
+            gt_label = {'gt_boxes': gt_boxes}
+            if config.USE_BLUR:
+                gt_blur = label['gt_blur']
+                gt_label['gt_blur'] = gt_blur
+            if self._debug:
+                img = data['data'].copy()[0].transpose(
+                    (1, 2, 0))[:, :, ::-1].copy()
+                print('DEBUG SHAPE', data['data'].shape,
+                      label['gt_boxes'].shape)
+
+                box = label['gt_boxes'].copy()[0][0:4].astype(np.int)
+                cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]),
+                              (0, 255, 0), 2)
+                filename = './debugout/%d.png' % (self._debug_id)
+                print('debug write', filename)
+                cv2.imwrite(filename, img)
+                self._debug_id += 1
+                #print('DEBUG', img.shape, bbox.shape)
+            label_dict = {}
+            if config.HEAD_BOX:
+                head_label_dict = self.aa.assign_anchor_fpn(
+                    gt_label,
+                    im_info,
+                    False,
+                    prefix='head',
+                    select_stride=select_stride)
+                label_dict.update(head_label_dict)
+            if config.FACE_LANDMARK:
+                gt_landmarks = label['gt_landmarks']
+                gt_label['gt_landmarks'] = gt_landmarks
+            #ta = datetime.datetime.now()
+            #face_label_dict = assign_anchor_fpn(feat_shape_list, gt_label, im_info, config.FACE_LANDMARK, prefix='face', select_stride = select_stride)
+            face_label_dict = self.aa.assign_anchor_fpn(
+                gt_label,
+                im_info,
+                config.FACE_LANDMARK,
+                prefix='face',
+                select_stride=select_stride)
+            #tb = datetime.datetime.now()
+            #self._times[0] += (tb-ta).total_seconds()
+            label_dict.update(face_label_dict)
+            #for k in label_dict:
+            #  print(k, label_dict[k].shape)
+
+            if config.CASCADE > 0:
+                pad_gt_boxes = np.empty(
+                    (1, config.TRAIN.MAX_BBOX_PER_IMAGE, 5), dtype=np.float32)
+                pad_gt_boxes.fill(-1)
+                pad_gt_boxes[0, 0:gt_boxes.shape[0], :] = gt_boxes
+                label_dict['gt_boxes'] = pad_gt_boxes
+            #print('im_info', im_info.shape)
+            #print(gt_boxes.shape)
+            for k in self.label_name:
+                label[k] = label_dict[k]
+
+        all_data = dict()
+        for key in self.data_name:
+            all_data[key] = tensor_vstack([batch[key] for batch in data_list])
+
+        all_label = dict()
+        for key in self.label_name:
+            pad = 0 if key.startswith('bbox_') else -1
+            #print('label vstack', key, pad, len(label_list), file=sys.stderr)
+            all_label[key] = tensor_vstack(
+                [batch[key] for batch in label_list], pad=pad)
+
+        self.data = [mx.nd.array(all_data[key]) for key in self.data_name]
+        self.label = [mx.nd.array(all_label[key]) for key in self.label_name]
+        #for _label in self.label:
+        #  print('LABEL SHAPE', _label.shape)
+        #print(self._times)
+
+
+class CropLoader2(mx.io.DataIter):
+    def __init__(self,
+                 feat_sym,
+                 roidb,
+                 batch_size=1,
+                 shuffle=False,
+                 ctx=None,
+                 work_load_list=None,
+                 aspect_grouping=False):
+        """
+        This Iter will provide roi data to Fast R-CNN network
+        :param feat_sym: to infer shape of assign_output
+        :param roidb: must be preprocessed
+        :param batch_size: must divide BATCH_SIZE(128)
+        :param shuffle: bool
+        :param ctx: list of contexts
+        :param work_load_list: list of work load
+        :param aspect_grouping: group images with similar aspects
+        :return: AnchorLoader
+        """
+        super(CropLoader2, self).__init__()
+
+        # save parameters as properties
+        self.feat_sym = feat_sym
+        self.roidb = roidb
+        self.batch_size = batch_size
+        self.shuffle = shuffle
+        self.ctx = ctx
+        if self.ctx is None:
+            self.ctx = [mx.cpu()]
+        self.work_load_list = work_load_list
+        #self.feat_stride = feat_stride
+        #self.anchor_scales = anchor_scales
+        #self.anchor_ratios = anchor_ratios
+        #self.allowed_border = allowed_border
+        self.aspect_grouping = aspect_grouping
+        self.feat_stride = config.RPN_FEAT_STRIDE
+
+        # infer properties from roidb
+        self.size = len(roidb)
+
+        # decide data and label names
+        #self.data_name = ['data']
+        #self.label_name = []
+        #self.label_name.append('label')
+        #self.label_name.append('bbox_target')
+        #self.label_name.append('bbox_weight')
+
+        self.data_name = ['data']
+        #self.label_name = ['label', 'bbox_target', 'bbox_weight']
+        self.label_name = []
+        prefixes = ['face']
+        if config.HEAD_BOX:
+            prefixes.append('head')
+        names = []
+        for prefix in prefixes:
+            names += [
+                prefix + '_label', prefix + '_bbox_target',
+                prefix + '_bbox_weight'
+            ]
+            if prefix == 'face' and config.FACE_LANDMARK:
+                names += [
+                    prefix + '_landmark_target', prefix + '_landmark_weight'
+                ]
+        #names = ['label', 'bbox_weight']
+        for stride in self.feat_stride:
+            for n in names:
+                k = "%s_stride%d" % (n, stride)
+                self.label_name.append(k)
+        # status variable for synchronization between get_data and get_label
+        self.cur = 0
+        self.batch = None
+        self.data = None
+        self.label = None
+
+        # get first batch to fill in provide_data and provide_label
+        self.reset()
+        self.q_in = [
+            multiprocessing.Queue(1024) for i in range(config.NUM_CPU)
+        ]
+        #self.q_in = multiprocessing.Queue(1024)
+        self.q_out = multiprocessing.Queue(1024)
+        self.start()
+        self.get_batch()
+
+    @property
+    def provide_data(self):
+        return [(k, v.shape) for k, v in zip(self.data_name, self.data)]
+
+    @property
+    def provide_label(self):
+        return [(k, v.shape) for k, v in zip(self.label_name, self.label)]
+
+    def reset(self):
+        pass
+
+    @staticmethod
+    def input_worker(q_in, roidb, batch_size):
+        index = np.arange(len(roidb))
+        np.random.shuffle(index)
+        cur_from = 0
+        while True:
+            cur_to = cur_from + batch_size
+            if cur_to > len(roidb):
+                np.random.shuffle(index)
+                cur_from = 0
+                continue
+            _roidb = [roidb[index[i]] for i in range(cur_from, cur_to)]
+            istart = index[cur_from]
+            q_in[istart % len(q_in)].put(_roidb)
+            cur_from = cur_to
+
+    @staticmethod
+    def gen_worker(q_in, q_out):
+        while True:
+            deq = q_in.get()
+            if deq is None:
+                break
+            _roidb = deq
+            data, label = get_crop_batch(_roidb)
+            print('generated')
+            q_out.put((data, label))
+
+    def start(self):
+        input_process = multiprocessing.Process(
+            target=CropLoader2.input_worker,
+            args=(self.q_in, self.roidb, self.batch_size))
+        #gen_process = multiprocessing.Process(target=gen_worker, args=(q_in, q_out))
+        gen_process = [multiprocessing.Process(target=CropLoader2.gen_worker, args=(self.q_in[i], self.q_out)) \
+                  for i in range(config.NUM_CPU)]
+        input_process.start()
+        for p in gen_process:
+            p.start()
+
+    def next(self):
+        self.get_batch()
+        return mx.io.DataBatch(data=self.data,
+                               label=self.label,
+                               provide_data=self.provide_data,
+                               provide_label=self.provide_label)
+
+    def infer_shape(self, max_data_shape=None, max_label_shape=None):
+        """ Return maximum data and label shape for single gpu """
+        if max_data_shape is None:
+            max_data_shape = []
+        if max_label_shape is None:
+            max_label_shape = []
+        max_shapes = dict(max_data_shape + max_label_shape)
+        input_batch_size = max_shapes['data'][0]
+        dummy_boxes = np.zeros((0, 5))
+        dummy_info = [[max_shapes['data'][2], max_shapes['data'][3], 1.0]]
+        dummy_label = {'gt_boxes': dummy_boxes}
+
+        # infer shape
+        feat_shape_list = []
+        for i in range(len(self.feat_stride)):
+            _, feat_shape, _ = self.feat_sym[i].infer_shape(**max_shapes)
+            feat_shape = [int(i) for i in feat_shape[0]]
+            feat_shape_list.append(feat_shape)
+
+        label_dict = {}
+        if config.HEAD_BOX:
+            head_label_dict = assign_anchor_fpn(feat_shape_list,
+                                                dummy_label,
+                                                dummy_info,
+                                                False,
+                                                prefix='head')
+            label_dict.update(head_label_dict)
+
+        if config.FACE_LANDMARK:
+            dummy_landmarks = np.zeros((0, 11))
+            dummy_label['gt_landmarks'] = dummy_landmarks
+        face_label_dict = assign_anchor_fpn(feat_shape_list,
+                                            dummy_label,
+                                            dummy_info,
+                                            config.FACE_LANDMARK,
+                                            prefix='face')
+        label_dict.update(face_label_dict)
+
+        label_list = []
+        for k in self.label_name:
+            label_list.append(label_dict[k])
+        label_shape = [(k, tuple([input_batch_size] + list(v.shape[1:])))
+                       for k, v in zip(self.label_name, label_list)]
+        return max_data_shape, label_shape
+
+    def get_batch(self):
+        deq = self.q_out.get()
+        print('q_out got')
+        data_list, label_list = deq
+
+        for data, label in zip(data_list, label_list):
+            data_shape = {k: v.shape for k, v in data.items()}
+            del data_shape['im_info']
+            feat_shape_list = []
+            for s in range(len(self.feat_stride)):
+                _, feat_shape, _ = self.feat_sym[s].infer_shape(**data_shape)
+                feat_shape = [int(i) for i in feat_shape[0]]
+                feat_shape_list.append(feat_shape)
+            #for k in self.label_name:
+            #  label[k] = [0 for i in range(config.TRAIN.BATCH_IMAGES)]
+            im_info = data['im_info']
+            gt_boxes = label['gt_boxes']
+            gt_label = {'gt_boxes': gt_boxes}
+            label_dict = {}
+            head_label_dict = assign_anchor_fpn(feat_shape_list,
+                                                gt_label,
+                                                im_info,
+                                                False,
+                                                prefix='head')
+            label_dict.update(head_label_dict)
+            if config.FACE_LANDMARK:
+                gt_landmarks = label['gt_landmarks']
+                gt_label['gt_landmarks'] = gt_landmarks
+            face_label_dict = assign_anchor_fpn(feat_shape_list,
+                                                gt_label,
+                                                im_info,
+                                                config.FACE_LANDMARK,
+                                                prefix='face')
+            label_dict.update(face_label_dict)
+            #print('im_info', im_info.shape)
+            #print(gt_boxes.shape)
+            for k in self.label_name:
+                label[k] = label_dict[k]
+
+        all_data = dict()
+        for key in self.data_name:
+            all_data[key] = tensor_vstack([batch[key] for batch in data_list])
+
+        all_label = dict()
+        for key in self.label_name:
+            pad = 0 if key.startswith('bbox_') else -1
+            #print('label vstack', key, pad, len(label_list), file=sys.stderr)
+            all_label[key] = tensor_vstack(
+                [batch[key] for batch in label_list], pad=pad)
+        self.data = [mx.nd.array(all_data[key]) for key in self.data_name]
+        self.label = [mx.nd.array(all_label[key]) for key in self.label_name]
diff --git a/detection/retinaface/rcnn/core/metric.py b/detection/retinaface/rcnn/core/metric.py
new file mode 100644
index 000000000..afdc92522
--- /dev/null
+++ b/detection/retinaface/rcnn/core/metric.py
@@ -0,0 +1,166 @@
+from __future__ import print_function
+import sys
+import mxnet as mx
+import numpy as np
+
+from rcnn.config import config
+
+
+def get_rpn_names():
+    pred = ['rpn_cls_prob', 'rpn_bbox_loss', 'rpn_label', 'rpn_bbox_weight']
+    label = ['rpn_label', 'rpn_bbox_target', 'rpn_bbox_weight']
+    return pred, label
+
+
+class RPNAccMetric(mx.metric.EvalMetric):
+    def __init__(self, pred_idx=-1, label_idx=-1, name='RPNAcc'):
+        super(RPNAccMetric, self).__init__(name)
+        self.pred, self.label = get_rpn_names()
+        #self.name = 'RPNAcc'
+        self.name = [name, name + '_BG', name + '_FG']
+        self.pred_idx = pred_idx
+        self.label_idx = label_idx
+        self.STAT = [0, 0, 0]
+
+    def reset(self):
+        """Clear the internal statistics to initial state."""
+        if isinstance(self.name, str):
+            self.num_inst = 0
+            self.sum_metric = 0.0
+        else:
+            #print('reset to ',len(self.name), self.name, file=sys.stderr)
+            self.num_inst = [0] * len(self.name)
+            self.sum_metric = [0.0] * len(self.name)
+
+    def get(self):
+        if isinstance(self.name, str):
+            if self.num_inst == 0:
+                return (self.name, float('nan'))
+            else:
+                return (self.name, self.sum_metric / self.num_inst)
+        else:
+            names = ['%s' % (self.name[i]) for i in range(len(self.name))]
+            values = [x / y if y != 0 else float('nan') \
+                for x, y in zip(self.sum_metric, self.num_inst)]
+            return (names, values)
+
+    def update(self, labels, preds):
+        if self.pred_idx >= 0 and self.label_idx >= 0:
+            pred = preds[self.pred_idx]
+            label = preds[self.label_idx]
+        else:
+            pred = preds[self.pred.index('rpn_cls_prob')]
+            label = labels[self.label.index('rpn_label')]
+            #label = preds[self.pred.index('rpn_label')]
+
+        num_images = pred.shape[0]
+        #print(pred.shape, label.shape, file=sys.stderr)
+        # pred (b, c, p) or (b, c, h, w)
+        pred_label = mx.ndarray.argmax_channel(pred).asnumpy().astype('int32')
+        #pred_label = pred_label.reshape((pred_label.shape[0], -1))
+        pred_label = pred_label.reshape(-1, )
+        # label (b, p)
+        label = label.asnumpy().astype('int32').reshape(-1, )
+        #print(pred_label.shape, label.shape)
+
+        # filter with keep_inds
+        keep_inds = np.where(label != -1)[0]
+        #print('in_metric acc', pred_label.shape, label.shape, len(keep_inds), file=sys.stderr)
+        #print(keep_inds, file=sys.stderr)
+        _pred_label = pred_label[keep_inds]
+        _label = label[keep_inds]
+        #print('in_metric2', pred_label.shape, label.shape, len(keep_inds), file=sys.stderr)
+        if isinstance(self.name, str):
+            self.sum_metric += np.sum(_pred_label.flat == _label.flat)
+            self.num_inst += len(_pred_label.flat)
+        else:
+            self.sum_metric[0] += np.sum(_pred_label.flat == _label.flat)
+            self.num_inst[0] += len(_pred_label.flat)
+
+            keep_inds = np.where(label == 0)[0]
+            _pred_label = pred_label[keep_inds]
+            _label = label[keep_inds]
+            self.sum_metric[1] += np.sum(_pred_label.flat == _label.flat)
+            self.num_inst[1] += len(_pred_label.flat)
+
+            keep_inds = np.where(label == 1)[0]
+            _pred_label = pred_label[keep_inds]
+            _label = label[keep_inds]
+            a = np.sum(_pred_label.flat == _label.flat)
+            b = len(_pred_label.flat)
+            self.sum_metric[2] += a
+            self.num_inst[2] += b
+
+            #self.STAT[0]+=a
+            #self.STAT[1]+=b
+            #self.STAT[2]+=num_images
+            #if self.STAT[2]%400==0:
+            #  print('FG_ACC', self.pred_idx, self.STAT[2], self.STAT[0], self.STAT[1], float(self.STAT[0])/self.STAT[1], file=sys.stderr)
+            #  self.STAT = [0,0,0]
+
+
+class RPNLogLossMetric(mx.metric.EvalMetric):
+    def __init__(self, pred_idx=-1, label_idx=-1):
+        super(RPNLogLossMetric, self).__init__('RPNLogLoss')
+        self.pred, self.label = get_rpn_names()
+        self.pred_idx = pred_idx
+        self.label_idx = label_idx
+
+    def update(self, labels, preds):
+        if self.pred_idx >= 0 and self.label_idx >= 0:
+            pred = preds[self.pred_idx]
+            label = preds[self.label_idx]
+        else:
+            pred = preds[self.pred.index('rpn_cls_prob')]
+            label = labels[self.label.index('rpn_label')]
+            #label = preds[self.pred.index('rpn_label')]
+
+        # label (b, p)
+        label = label.asnumpy().astype('int32').reshape((-1))
+        # pred (b, c, p) or (b, c, h, w) --> (b, p, c) --> (b*p, c)
+        pred = pred.asnumpy().reshape(
+            (pred.shape[0], pred.shape[1], -1)).transpose((0, 2, 1))
+        pred = pred.reshape((label.shape[0], -1))
+
+        # filter with keep_inds
+        keep_inds = np.where(label != -1)[0]
+        label = label[keep_inds]
+        cls = pred[keep_inds, label]
+        #print('in_metric log', label.shape, cls.shape, file=sys.stderr)
+
+        cls += 1e-14
+        cls_loss = -1 * np.log(cls)
+        cls_loss = np.sum(cls_loss)
+        self.sum_metric += cls_loss
+        self.num_inst += label.shape[0]
+
+
+class RPNL1LossMetric(mx.metric.EvalMetric):
+    def __init__(self, loss_idx=-1, weight_idx=-1, name='RPNL1Loss'):
+        super(RPNL1LossMetric, self).__init__(name)
+        self.pred, self.label = get_rpn_names()
+        self.loss_idx = loss_idx
+        self.weight_idx = weight_idx
+        self.name = name
+
+    def update(self, labels, preds):
+        if self.loss_idx >= 0 and self.weight_idx >= 0:
+            bbox_loss = preds[self.loss_idx].asnumpy()
+            bbox_weight = preds[self.weight_idx].asnumpy()
+        else:
+            bbox_loss = preds[self.pred.index('rpn_bbox_loss')].asnumpy()
+            bbox_weight = labels[self.label.index('rpn_bbox_weight')].asnumpy()
+            #bbox_weight = preds[self.pred.index('rpn_bbox_weight')].asnumpy()
+
+        #print('in_metric', self.name, bbox_weight.shape, bbox_loss.shape)
+
+        # calculate num_inst (average on those fg anchors)
+        if config.LR_MODE == 0:
+            num_inst = np.sum(bbox_weight > 0) / (bbox_weight.shape[1] /
+                                                  config.NUM_ANCHORS)
+        else:
+            num_inst = 1
+        #print('in_metric log', bbox_loss.shape, num_inst, file=sys.stderr)
+
+        self.sum_metric += np.sum(bbox_loss)
+        self.num_inst += num_inst
diff --git a/detection/retinaface/rcnn/core/module.py b/detection/retinaface/rcnn/core/module.py
new file mode 100644
index 000000000..731a4ff1b
--- /dev/null
+++ b/detection/retinaface/rcnn/core/module.py
@@ -0,0 +1,259 @@
+"""A `MutableModule` implement the `BaseModule` API, and allows input shape
+varying with training iterations. If shapes vary, executors will rebind,
+using shared arrays from the initial module binded with maximum shape.
+"""
+
+import logging
+
+from mxnet import context as ctx
+from mxnet.initializer import Uniform
+from mxnet.module.base_module import BaseModule
+from mxnet.module.module import Module
+
+
+class MutableModule(BaseModule):
+    """A mutable module is a module that supports variable input data.
+
+    Parameters
+    ----------
+    symbol : Symbol
+    data_names : list of str
+    label_names : list of str
+    logger : Logger
+    context : Context or list of Context
+    work_load_list : list of number
+    max_data_shapes : list of (name, shape) tuple, designating inputs whose shape vary
+    max_label_shapes : list of (name, shape) tuple, designating inputs whose shape vary
+    fixed_param_prefix : list of str, indicating fixed parameters
+    """
+    def __init__(self,
+                 symbol,
+                 data_names,
+                 label_names,
+                 logger=logging,
+                 context=ctx.cpu(),
+                 work_load_list=None,
+                 max_data_shapes=None,
+                 max_label_shapes=None,
+                 fixed_param_prefix=None):
+        super(MutableModule, self).__init__(logger=logger)
+        self._symbol = symbol
+        self._data_names = data_names
+        self._label_names = label_names
+        self._context = context
+        self._work_load_list = work_load_list
+
+        self._curr_module = None
+        self._max_data_shapes = max_data_shapes
+        self._max_label_shapes = max_label_shapes
+        self._fixed_param_prefix = fixed_param_prefix
+
+        fixed_param_names = list()
+        if fixed_param_prefix is not None:
+            for name in self._symbol.list_arguments():
+                for prefix in self._fixed_param_prefix:
+                    if prefix in name:
+                        fixed_param_names.append(name)
+        self._fixed_param_names = fixed_param_names
+
+    def _reset_bind(self):
+        self.binded = False
+        self._curr_module = None
+
+    @property
+    def data_names(self):
+        return self._data_names
+
+    @property
+    def output_names(self):
+        return self._symbol.list_outputs()
+
+    @property
+    def data_shapes(self):
+        assert self.binded
+        return self._curr_module.data_shapes
+
+    @property
+    def label_shapes(self):
+        assert self.binded
+        return self._curr_module.label_shapes
+
+    @property
+    def output_shapes(self):
+        assert self.binded
+        return self._curr_module.output_shapes
+
+    def get_params(self):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_params()
+
+    def init_params(self,
+                    initializer=Uniform(0.01),
+                    arg_params=None,
+                    aux_params=None,
+                    allow_missing=False,
+                    force_init=False,
+                    allow_extra=False):
+        if self.params_initialized and not force_init:
+            return
+        assert self.binded, 'call bind before initializing the parameters'
+        self._curr_module.init_params(initializer=initializer,
+                                      arg_params=arg_params,
+                                      aux_params=aux_params,
+                                      allow_missing=allow_missing,
+                                      force_init=force_init,
+                                      allow_extra=allow_extra)
+        self.params_initialized = True
+
+    def bind(self,
+             data_shapes,
+             label_shapes=None,
+             for_training=True,
+             inputs_need_grad=False,
+             force_rebind=False,
+             shared_module=None):
+        # in case we already initialized params, keep it
+        if self.params_initialized:
+            arg_params, aux_params = self.get_params()
+
+        # force rebinding is typically used when one want to switch from
+        # training to prediction phase.
+        if force_rebind:
+            self._reset_bind()
+
+        if self.binded:
+            self.logger.warning('Already binded, ignoring bind()')
+            return
+
+        assert shared_module is None, 'shared_module for MutableModule is not supported'
+
+        self.for_training = for_training
+        self.inputs_need_grad = inputs_need_grad
+        self.binded = True
+
+        max_shapes_dict = dict()
+        if self._max_data_shapes is not None:
+            max_shapes_dict.update(dict(self._max_data_shapes))
+        if self._max_label_shapes is not None:
+            max_shapes_dict.update(dict(self._max_label_shapes))
+
+        max_data_shapes = list()
+        for name, shape in data_shapes:
+            if name in max_shapes_dict:
+                max_data_shapes.append((name, max_shapes_dict[name]))
+            else:
+                max_data_shapes.append((name, shape))
+
+        max_label_shapes = list()
+        if label_shapes is not None:
+            for name, shape in label_shapes:
+                if name in max_shapes_dict:
+                    max_label_shapes.append((name, max_shapes_dict[name]))
+                else:
+                    max_label_shapes.append((name, shape))
+
+        if len(max_label_shapes) == 0:
+            max_label_shapes = None
+
+        module = Module(self._symbol,
+                        self._data_names,
+                        self._label_names,
+                        logger=self.logger,
+                        context=self._context,
+                        work_load_list=self._work_load_list,
+                        fixed_param_names=self._fixed_param_names)
+        module.bind(max_data_shapes,
+                    max_label_shapes,
+                    for_training,
+                    inputs_need_grad,
+                    force_rebind=False,
+                    shared_module=None)
+        self._curr_module = module
+
+        # copy back saved params, if already initialized
+        if self.params_initialized:
+            self.set_params(arg_params, aux_params)
+
+    def init_optimizer(self,
+                       kvstore='local',
+                       optimizer='sgd',
+                       optimizer_params=(('learning_rate', 0.01), ),
+                       force_init=False):
+        assert self.binded and self.params_initialized
+        if self.optimizer_initialized and not force_init:
+            self.logger.warning('optimizer already initialized, ignoring.')
+            return
+
+        self._curr_module.init_optimizer(kvstore,
+                                         optimizer,
+                                         optimizer_params,
+                                         force_init=force_init)
+        self.optimizer_initialized = True
+
+    def forward(self, data_batch, is_train=None):
+        assert self.binded and self.params_initialized
+
+        # get current_shapes
+        if self._curr_module.label_shapes is not None:
+            current_shapes = dict(self._curr_module.data_shapes +
+                                  self._curr_module.label_shapes)
+        else:
+            current_shapes = dict(self._curr_module.data_shapes)
+
+        # get input_shapes
+        if data_batch.provide_label is not None:
+            input_shapes = dict(data_batch.provide_data +
+                                data_batch.provide_label)
+        else:
+            input_shapes = dict(data_batch.provide_data)
+
+        # decide if shape changed
+        shape_changed = False
+        for k, v in current_shapes.items():
+            if v != input_shapes[k]:
+                shape_changed = True
+
+        if shape_changed:
+            module = Module(self._symbol,
+                            self._data_names,
+                            self._label_names,
+                            logger=self.logger,
+                            context=self._context,
+                            work_load_list=self._work_load_list,
+                            fixed_param_names=self._fixed_param_names)
+            module.bind(data_batch.provide_data,
+                        data_batch.provide_label,
+                        self._curr_module.for_training,
+                        self._curr_module.inputs_need_grad,
+                        force_rebind=False,
+                        shared_module=self._curr_module)
+            self._curr_module = module
+
+        self._curr_module.forward(data_batch, is_train=is_train)
+
+    def backward(self, out_grads=None):
+        assert self.binded and self.params_initialized
+        self._curr_module.backward(out_grads=out_grads)
+
+    def update(self):
+        assert self.binded and self.params_initialized and self.optimizer_initialized
+        self._curr_module.update()
+
+    def get_outputs(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_outputs(
+            merge_multi_context=merge_multi_context)
+
+    def get_input_grads(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized and self.inputs_need_grad
+        return self._curr_module.get_input_grads(
+            merge_multi_context=merge_multi_context)
+
+    def update_metric(self, eval_metric, labels):
+        assert self.binded and self.params_initialized
+        self._curr_module.update_metric(eval_metric, labels)
+
+    def install_monitor(self, mon):
+        """ Install monitor on all executors """
+        assert self.binded
+        self._curr_module.install_monitor(mon)
diff --git a/detection/retinaface/rcnn/core/module_bak.py b/detection/retinaface/rcnn/core/module_bak.py
new file mode 100644
index 000000000..2c8190251
--- /dev/null
+++ b/detection/retinaface/rcnn/core/module_bak.py
@@ -0,0 +1,260 @@
+"""A `MutableModule` implement the `BaseModule` API, and allows input shape
+varying with training iterations. If shapes vary, executors will rebind,
+using shared arrays from the initial module binded with maximum shape.
+"""
+
+import logging
+
+from mxnet import context as ctx
+from mxnet.initializer import Uniform
+from mxnet.module.base_module import BaseModule
+from mxnet.module.module import Module
+
+
+class MutableModule(BaseModule):
+    """A mutable module is a module that supports variable input data.
+
+    Parameters
+    ----------
+    symbol : Symbol
+    data_names : list of str
+    label_names : list of str
+    logger : Logger
+    context : Context or list of Context
+    work_load_list : list of number
+    max_data_shapes : list of (name, shape) tuple, designating inputs whose shape vary
+    max_label_shapes : list of (name, shape) tuple, designating inputs whose shape vary
+    fixed_param_prefix : list of str, indicating fixed parameters
+    """
+    def __init__(self,
+                 symbol,
+                 data_names,
+                 label_names,
+                 logger=logging,
+                 context=ctx.cpu(),
+                 work_load_list=None,
+                 max_data_shapes=None,
+                 max_label_shapes=None,
+                 fixed_param_prefix=None):
+        super(MutableModule, self).__init__(logger=logger)
+        self._symbol = symbol
+        self._data_names = data_names
+        self._label_names = label_names
+        self._context = context
+        self._work_load_list = work_load_list
+
+        self._curr_module = None
+        self._max_data_shapes = max_data_shapes
+        self._max_label_shapes = max_label_shapes
+        self._fixed_param_prefix = fixed_param_prefix
+
+        fixed_param_names = list()
+        if fixed_param_prefix is not None:
+            for name in self._symbol.list_arguments():
+                for prefix in self._fixed_param_prefix:
+                    if prefix in name:
+                        fixed_param_names.append(name)
+        self._fixed_param_names = fixed_param_names
+
+    def _reset_bind(self):
+        self.binded = False
+        self._curr_module = None
+
+    @property
+    def data_names(self):
+        return self._data_names
+
+    @property
+    def output_names(self):
+        return self._symbol.list_outputs()
+
+    @property
+    def data_shapes(self):
+        assert self.binded
+        return self._curr_module.data_shapes
+
+    @property
+    def label_shapes(self):
+        assert self.binded
+        return self._curr_module.label_shapes
+
+    @property
+    def output_shapes(self):
+        assert self.binded
+        return self._curr_module.output_shapes
+
+    def get_params(self):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_params()
+
+    def init_params(self,
+                    initializer=Uniform(0.01),
+                    arg_params=None,
+                    aux_params=None,
+                    allow_missing=False,
+                    force_init=False,
+                    allow_extra=False):
+        if self.params_initialized and not force_init:
+            return
+        assert self.binded, 'call bind before initializing the parameters'
+        self._curr_module.init_params(initializer=initializer,
+                                      arg_params=arg_params,
+                                      aux_params=aux_params,
+                                      allow_missing=allow_missing,
+                                      force_init=force_init,
+                                      allow_extra=allow_extra)
+        self.params_initialized = True
+
+    def bind(self,
+             data_shapes,
+             label_shapes=None,
+             for_training=True,
+             inputs_need_grad=False,
+             force_rebind=False,
+             shared_module=None,
+             grad_req='write'):
+        # in case we already initialized params, keep it
+        if self.params_initialized:
+            arg_params, aux_params = self.get_params()
+
+        # force rebinding is typically used when one want to switch from
+        # training to prediction phase.
+        if force_rebind:
+            self._reset_bind()
+
+        if self.binded:
+            self.logger.warning('Already binded, ignoring bind()')
+            return
+
+        assert shared_module is None, 'shared_module for MutableModule is not supported'
+
+        self.for_training = for_training
+        self.inputs_need_grad = inputs_need_grad
+        self.binded = True
+
+        max_shapes_dict = dict()
+        if self._max_data_shapes is not None:
+            max_shapes_dict.update(dict(self._max_data_shapes))
+        if self._max_label_shapes is not None:
+            max_shapes_dict.update(dict(self._max_label_shapes))
+
+        max_data_shapes = list()
+        for name, shape in data_shapes:
+            if name in max_shapes_dict:
+                max_data_shapes.append((name, max_shapes_dict[name]))
+            else:
+                max_data_shapes.append((name, shape))
+
+        max_label_shapes = list()
+        if label_shapes is not None:
+            for name, shape in label_shapes:
+                if name in max_shapes_dict:
+                    max_label_shapes.append((name, max_shapes_dict[name]))
+                else:
+                    max_label_shapes.append((name, shape))
+
+        if len(max_label_shapes) == 0:
+            max_label_shapes = None
+
+        module = Module(self._symbol,
+                        self._data_names,
+                        self._label_names,
+                        logger=self.logger,
+                        context=self._context,
+                        work_load_list=self._work_load_list,
+                        fixed_param_names=self._fixed_param_names)
+        module.bind(max_data_shapes,
+                    max_label_shapes,
+                    for_training,
+                    inputs_need_grad,
+                    force_rebind=False,
+                    shared_module=None)
+        self._curr_module = module
+
+        # copy back saved params, if already initialized
+        if self.params_initialized:
+            self.set_params(arg_params, aux_params)
+
+    def init_optimizer(self,
+                       kvstore='local',
+                       optimizer='sgd',
+                       optimizer_params=(('learning_rate', 0.01), ),
+                       force_init=False):
+        assert self.binded and self.params_initialized
+        if self.optimizer_initialized and not force_init:
+            self.logger.warning('optimizer already initialized, ignoring.')
+            return
+
+        self._curr_module.init_optimizer(kvstore,
+                                         optimizer,
+                                         optimizer_params,
+                                         force_init=force_init)
+        self.optimizer_initialized = True
+
+    def forward(self, data_batch, is_train=None):
+        assert self.binded and self.params_initialized
+
+        # get current_shapes
+        if self._curr_module.label_shapes is not None:
+            current_shapes = dict(self._curr_module.data_shapes +
+                                  self._curr_module.label_shapes)
+        else:
+            current_shapes = dict(self._curr_module.data_shapes)
+
+        # get input_shapes
+        if data_batch.provide_label is not None:
+            input_shapes = dict(data_batch.provide_data +
+                                data_batch.provide_label)
+        else:
+            input_shapes = dict(data_batch.provide_data)
+
+        # decide if shape changed
+        shape_changed = False
+        for k, v in current_shapes.items():
+            if v != input_shapes[k]:
+                shape_changed = True
+
+        if shape_changed:
+            module = Module(self._symbol,
+                            self._data_names,
+                            self._label_names,
+                            logger=self.logger,
+                            context=self._context,
+                            work_load_list=self._work_load_list,
+                            fixed_param_names=self._fixed_param_names)
+            module.bind(data_batch.provide_data,
+                        data_batch.provide_label,
+                        self._curr_module.for_training,
+                        self._curr_module.inputs_need_grad,
+                        force_rebind=False,
+                        shared_module=self._curr_module)
+            self._curr_module = module
+
+        self._curr_module.forward(data_batch, is_train=is_train)
+
+    def backward(self, out_grads=None):
+        assert self.binded and self.params_initialized
+        self._curr_module.backward(out_grads=out_grads)
+
+    def update(self):
+        assert self.binded and self.params_initialized and self.optimizer_initialized
+        self._curr_module.update()
+
+    def get_outputs(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_outputs(
+            merge_multi_context=merge_multi_context)
+
+    def get_input_grads(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized and self.inputs_need_grad
+        return self._curr_module.get_input_grads(
+            merge_multi_context=merge_multi_context)
+
+    def update_metric(self, eval_metric, labels):
+        assert self.binded and self.params_initialized
+        self._curr_module.update_metric(eval_metric, labels)
+
+    def install_monitor(self, mon):
+        """ Install monitor on all executors """
+        assert self.binded
+        self._curr_module.install_monitor(mon)
diff --git a/detection/retinaface/rcnn/core/tester.py b/detection/retinaface/rcnn/core/tester.py
new file mode 100644
index 000000000..d9d5d32b1
--- /dev/null
+++ b/detection/retinaface/rcnn/core/tester.py
@@ -0,0 +1,527 @@
+from __future__ import print_function
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+import os
+import sys
+import time
+import mxnet as mx
+import numpy as np
+from builtins import range
+
+from mxnet.module import Module
+from .module import MutableModule
+from rcnn.logger import logger
+from rcnn.config import config
+from rcnn.io import image
+from rcnn.processing.bbox_transform import bbox_pred, clip_boxes
+from rcnn.processing.nms import py_nms_wrapper, cpu_nms_wrapper, gpu_nms_wrapper
+from rcnn.processing.bbox_transform import bbox_overlaps
+
+
+def IOU(Reframe, GTframe):
+    x1 = Reframe[0]
+    y1 = Reframe[1]
+    width1 = Reframe[2] - Reframe[0]
+    height1 = Reframe[3] - Reframe[1]
+
+    x2 = GTframe[0]
+    y2 = GTframe[1]
+    width2 = GTframe[2] - GTframe[0]
+    height2 = GTframe[3] - GTframe[1]
+
+    endx = max(x1 + width1, x2 + width2)
+    startx = min(x1, x2)
+    width = width1 + width2 - (endx - startx)
+
+    endy = max(y1 + height1, y2 + height2)
+    starty = min(y1, y2)
+    height = height1 + height2 - (endy - starty)
+
+    if width <= 0 or height <= 0:
+        ratio = 0
+    else:
+        Area = width * height
+        Area1 = width1 * height1
+        Area2 = width2 * height2
+        ratio = Area * 1. / (Area1 + Area2 - Area)
+    return ratio
+
+
+class Predictor(object):
+    def __init__(self,
+                 symbol,
+                 data_names,
+                 label_names,
+                 context=mx.cpu(),
+                 max_data_shapes=None,
+                 provide_data=None,
+                 provide_label=None,
+                 arg_params=None,
+                 aux_params=None):
+        #self._mod = MutableModule(symbol, data_names, label_names,
+        #                          context=context, max_data_shapes=max_data_shapes)
+        self._mod = Module(symbol, data_names, label_names, context=context)
+        self._mod.bind(provide_data, provide_label, for_training=False)
+        self._mod.init_params(arg_params=arg_params, aux_params=aux_params)
+
+    def predict(self, data_batch):
+        self._mod.forward(data_batch)
+        return dict(zip(self._mod.output_names,
+                        self._mod.get_outputs()))  #TODO
+        #return self._mod.get_outputs()
+
+
+def im_proposal(predictor, data_batch, data_names, scale):
+    data_dict = dict(zip(data_names, data_batch.data))
+    output = predictor.predict(data_batch)
+
+    # drop the batch index
+    boxes = output['rois_output'].asnumpy()[:, 1:]
+    scores = output['rois_score'].asnumpy()
+
+    # transform to original scale
+    boxes = boxes / scale
+
+    return scores, boxes, data_dict
+
+
+def _im_proposal(predictor, data_batch, data_names, scale):
+    data_dict = dict(zip(data_names, data_batch.data))
+    output = predictor.predict(data_batch)
+    print('output', output)
+
+    # drop the batch index
+    boxes = output['rois_output'].asnumpy()[:, 1:]
+    scores = output['rois_score'].asnumpy()
+
+    # transform to original scale
+    boxes = boxes / scale
+
+    return scores, boxes, data_dict
+
+
+def generate_proposals(predictor, test_data, imdb, vis=False, thresh=0.):
+    """
+    Generate detections results using RPN.
+    :param predictor: Predictor
+    :param test_data: data iterator, must be non-shuffled
+    :param imdb: image database
+    :param vis: controls visualization
+    :param thresh: thresh for valid detections
+    :return: list of detected boxes
+    """
+    assert vis or not test_data.shuffle
+    data_names = [k[0] for k in test_data.provide_data]
+
+    i = 0
+    t = time.time()
+    imdb_boxes = list()
+    original_boxes = list()
+    for im_info, data_batch in test_data:
+        t1 = time.time() - t
+        t = time.time()
+
+        scale = im_info[0, 2]
+        scores, boxes, data_dict = im_proposal(predictor, data_batch,
+                                               data_names, scale)
+        print(scores.shape, boxes.shape, file=sys.stderr)
+        t2 = time.time() - t
+        t = time.time()
+
+        # assemble proposals
+        dets = np.hstack((boxes, scores))
+        original_boxes.append(dets)
+
+        # filter proposals
+        keep = np.where(dets[:, 4:] > thresh)[0]
+        dets = dets[keep, :]
+        imdb_boxes.append(dets)
+
+        if vis:
+            vis_all_detection(data_dict['data'].asnumpy(), [dets], ['obj'],
+                              scale)
+
+        logger.info('generating %d/%d ' % (i + 1, imdb.num_images) +
+                    'proposal %d ' % (dets.shape[0]) + 'data %.4fs net %.4fs' %
+                    (t1, t2))
+        i += 1
+
+    assert len(imdb_boxes) == imdb.num_images, 'calculations not complete'
+
+    # save results
+    rpn_folder = os.path.join(imdb.root_path, 'rpn_data')
+    if not os.path.exists(rpn_folder):
+        os.mkdir(rpn_folder)
+
+    rpn_file = os.path.join(rpn_folder, imdb.name + '_rpn.pkl')
+    with open(rpn_file, 'wb') as f:
+        pickle.dump(imdb_boxes, f, pickle.HIGHEST_PROTOCOL)
+
+    if thresh > 0:
+        full_rpn_file = os.path.join(rpn_folder, imdb.name + '_full_rpn.pkl')
+        with open(full_rpn_file, 'wb') as f:
+            pickle.dump(original_boxes, f, pickle.HIGHEST_PROTOCOL)
+
+    logger.info('wrote rpn proposals to %s' % rpn_file)
+    return imdb_boxes
+
+
+def test_proposals(predictor, test_data, imdb, roidb, vis=False):
+    """
+    Test detections results using RPN.
+    :param predictor: Predictor
+    :param test_data: data iterator, must be non-shuffled
+    :param imdb: image database
+    :param roidb: roidb 
+    :param vis: controls visualization
+    :return: recall, mAP
+    """
+    assert vis or not test_data.shuffle
+    data_names = [k[0] for k in test_data.provide_data]
+
+    #bbox_file = os.path.join(rpn_folder, imdb.name + '_bbox.txt')
+    #bbox_f = open(bbox_file, 'w')
+
+    i = 0
+    t = time.time()
+    output_folder = os.path.join(imdb.root_path, 'output')
+    if not os.path.exists(output_folder):
+        os.mkdir(output_folder)
+    imdb_boxes = list()
+    original_boxes = list()
+    gt_overlaps = np.zeros(0)
+    overall = [0.0, 0.0]
+    gt_max = np.array((0.0, 0.0))
+    num_pos = 0
+    #apply scale, for SSH
+    #_, roidb = image.get_image(roidb)
+    for im_info, data_batch in test_data:
+        t1 = time.time() - t
+        t = time.time()
+
+        oscale = im_info[0, 2]
+        #print('scale', scale, file=sys.stderr)
+        scale = 1.0  #fix scale=1.0 for SSH face detector
+        scores, boxes, data_dict = im_proposal(predictor, data_batch,
+                                               data_names, scale)
+        #print(scores.shape, boxes.shape, file=sys.stderr)
+        t2 = time.time() - t
+        t = time.time()
+
+        # assemble proposals
+        dets = np.hstack((boxes, scores))
+        original_boxes.append(dets)
+
+        # filter proposals
+        keep = np.where(dets[:, 4:] > config.TEST.SCORE_THRESH)[0]
+        dets = dets[keep, :]
+        imdb_boxes.append(dets)
+
+        logger.info('generating %d/%d ' % (i + 1, imdb.num_images) +
+                    'proposal %d ' % (dets.shape[0]) + 'data %.4fs net %.4fs' %
+                    (t1, t2))
+
+        #if dets.shape[0]==0:
+        #  continue
+        if vis:
+            vis_all_detection(data_dict['data'].asnumpy(), [dets], ['obj'],
+                              scale)
+        boxes = dets
+        #max_gt_overlaps = roidb[i]['gt_overlaps'].max(axis=1)
+        #gt_inds = np.where((roidb[i]['gt_classes'] > 0) & (max_gt_overlaps == 1))[0]
+        #gt_boxes = roidb[i]['boxes'][gt_inds, :]
+        gt_boxes = roidb[i]['boxes'].copy(
+        ) * oscale  # as roidb is the original one, need to scale GT for SSH
+        gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] + 1) * (gt_boxes[:, 3] -
+                                                            gt_boxes[:, 1] + 1)
+        num_pos += gt_boxes.shape[0]
+
+        overlaps = bbox_overlaps(boxes.astype(np.float),
+                                 gt_boxes.astype(np.float))
+        #print(im_info, gt_boxes.shape, boxes.shape, overlaps.shape, file=sys.stderr)
+
+        _gt_overlaps = np.zeros((gt_boxes.shape[0]))
+        # choose whatever is smaller to iterate
+
+        #for j in range(gt_boxes.shape[0]):
+        #  print('gt %d,%d,%d,%d'% (gt_boxes[j][0], gt_boxes[j][1], gt_boxes[j][2]-gt_boxes[j][0], gt_boxes[j][3]-gt_boxes[j][1]), file=sys.stderr)
+        #  gt_max = np.maximum( gt_max, np.array( (gt_boxes[j][2], gt_boxes[j][3]) ) )
+        #print('gt max', gt_max, file=sys.stderr)
+        #for j in range(boxes.shape[0]):
+        #  print('anchor_box %.2f,%.2f,%.2f,%.2f'% (boxes[j][0], boxes[j][1], boxes[j][2]-boxes[j][0], boxes[j][3]-boxes[j][1]), file=sys.stderr)
+
+        #rounds = min(boxes.shape[0], gt_boxes.shape[0])
+        #for j in range(rounds):
+        #    # find which proposal maximally covers each gt box
+        #    argmax_overlaps = overlaps.argmax(axis=0)
+        #    print(j, 'argmax_overlaps', argmax_overlaps, file=sys.stderr)
+        #    # get the IoU amount of coverage for each gt box
+        #    max_overlaps = overlaps.max(axis=0)
+        #    print(j, 'max_overlaps', max_overlaps, file=sys.stderr)
+        #    # find which gt box is covered by most IoU
+        #    gt_ind = max_overlaps.argmax()
+        #    gt_ovr = max_overlaps.max()
+        #    assert (gt_ovr >= 0), '%s\n%s\n%s' % (boxes, gt_boxes, overlaps)
+        #    # find the proposal box that covers the best covered gt box
+        #    box_ind = argmax_overlaps[gt_ind]
+        #    print('max box', gt_ind, box_ind, (boxes[box_ind][0], boxes[box_ind][1], boxes[box_ind][2]-boxes[box_ind][0], boxes[box_ind][3]-boxes[box_ind][1], boxes[box_ind][4]), file=sys.stderr)
+        #    # record the IoU coverage of this gt box
+        #    _gt_overlaps[j] = overlaps[box_ind, gt_ind]
+        #    assert (_gt_overlaps[j] == gt_ovr)
+        #    # mark the proposal box and the gt box as used
+        #    overlaps[box_ind, :] = -1
+        #    overlaps[:, gt_ind] = -1
+
+        if boxes.shape[0] > 0:
+            _gt_overlaps = overlaps.max(axis=0)
+            #print('max_overlaps', _gt_overlaps, file=sys.stderr)
+            for j in range(len(_gt_overlaps)):
+                if _gt_overlaps[j] > config.TEST.IOU_THRESH:
+                    continue
+                print(j,
+                      'failed',
+                      gt_boxes[j],
+                      'max_overlap:',
+                      _gt_overlaps[j],
+                      file=sys.stderr)
+                #_idx = np.where(overlaps[:,j]>0.4)[0]
+                #print(j, _idx, file=sys.stderr)
+                #print(overlaps[_idx,j], file=sys.stderr)
+                #for __idx in _idx:
+                #  print(gt_boxes[j], boxes[__idx], overlaps[__idx,j], IOU(gt_boxes[j], boxes[__idx,0:4]), file=sys.stderr)
+
+            # append recorded IoU coverage level
+            found = (_gt_overlaps > config.TEST.IOU_THRESH).sum()
+            _recall = found / float(gt_boxes.shape[0])
+            print('recall',
+                  _recall,
+                  gt_boxes.shape[0],
+                  boxes.shape[0],
+                  gt_areas,
+                  file=sys.stderr)
+            overall[0] += found
+            overall[1] += gt_boxes.shape[0]
+            #gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
+            #_recall = (gt_overlaps >= threshold).sum() / float(num_pos)
+            _recall = float(overall[0]) / overall[1]
+            print('recall_all', _recall, file=sys.stderr)
+
+        boxes[:, 0:4] /= oscale
+        _vec = roidb[i]['image'].split('/')
+        out_dir = os.path.join(output_folder, _vec[-2])
+        if not os.path.exists(out_dir):
+            os.mkdir(out_dir)
+        out_file = os.path.join(out_dir, _vec[-1].replace('jpg', 'txt'))
+        with open(out_file, 'w') as f:
+            name = '/'.join(roidb[i]['image'].split('/')[-2:])
+            f.write("%s\n" % (name))
+            f.write("%d\n" % (boxes.shape[0]))
+            for b in range(boxes.shape[0]):
+                box = boxes[b]
+                f.write(
+                    "%d %d %d %d %g \n" %
+                    (box[0], box[1], box[2] - box[0], box[3] - box[1], box[4]))
+        i += 1
+
+    #bbox_f.close()
+    return
+    gt_overlaps = np.sort(gt_overlaps)
+    recalls = np.zeros_like(thresholds)
+
+    # compute recall for each IoU threshold
+    for i, t in enumerate(thresholds):
+        recalls[i] = (gt_overlaps >= t).sum() / float(num_pos)
+    ar = recalls.mean()
+
+    # print results
+    print('average recall for {}: {:.3f}'.format(area_name, ar))
+    for threshold, recall in zip(thresholds, recalls):
+        print('recall @{:.2f}: {:.3f}'.format(threshold, recall))
+
+    assert len(imdb_boxes) == imdb.num_images, 'calculations not complete'
+
+    # save results
+
+    rpn_file = os.path.join(rpn_folder, imdb.name + '_rpn.pkl')
+    with open(rpn_file, 'wb') as f:
+        pickle.dump(imdb_boxes, f, pickle.HIGHEST_PROTOCOL)
+
+    logger.info('wrote rpn proposals to %s' % rpn_file)
+    return imdb_boxes
+
+
+def im_detect(predictor, data_batch, data_names, scale):
+    output = predictor.predict(data_batch)
+
+    data_dict = dict(zip(data_names, data_batch.data))
+    if config.TEST.HAS_RPN:
+        rois = output['rois_output'].asnumpy()[:, 1:]
+    else:
+        rois = data_dict['rois'].asnumpy().reshape((-1, 5))[:, 1:]
+    im_shape = data_dict['data'].shape
+
+    # save output
+    scores = output['cls_prob_reshape_output'].asnumpy()[0]
+    bbox_deltas = output['bbox_pred_reshape_output'].asnumpy()[0]
+
+    # post processing
+    pred_boxes = bbox_pred(rois, bbox_deltas)
+    pred_boxes = clip_boxes(pred_boxes, im_shape[-2:])
+
+    # we used scaled image & roi to train, so it is necessary to transform them back
+    pred_boxes = pred_boxes / scale
+
+    return scores, pred_boxes, data_dict
+
+
+def pred_eval(predictor, test_data, imdb, vis=False, thresh=1e-3):
+    """
+    wrapper for calculating offline validation for faster data analysis
+    in this example, all threshold are set by hand
+    :param predictor: Predictor
+    :param test_data: data iterator, must be non-shuffle
+    :param imdb: image database
+    :param vis: controls visualization
+    :param thresh: valid detection threshold
+    :return:
+    """
+    assert vis or not test_data.shuffle
+    data_names = [k[0] for k in test_data.provide_data]
+
+    nms = py_nms_wrapper(config.TEST.NMS)
+
+    # limit detections to max_per_image over all classes
+    max_per_image = -1
+
+    num_images = imdb.num_images
+    # all detections are collected into:
+    #    all_boxes[cls][image] = N x 5 array of detections in
+    #    (x1, y1, x2, y2, score)
+    all_boxes = [[[] for _ in range(num_images)]
+                 for _ in range(imdb.num_classes)]
+
+    i = 0
+    t = time.time()
+    for im_info, data_batch in test_data:
+        t1 = time.time() - t
+        t = time.time()
+
+        scale = im_info[0, 2]
+        scores, boxes, data_dict = im_detect(predictor, data_batch, data_names,
+                                             scale)
+
+        t2 = time.time() - t
+        t = time.time()
+
+        for j in range(1, imdb.num_classes):
+            indexes = np.where(scores[:, j] > thresh)[0]
+            cls_scores = scores[indexes, j, np.newaxis]
+            cls_boxes = boxes[indexes, j * 4:(j + 1) * 4]
+            cls_dets = np.hstack((cls_boxes, cls_scores))
+            keep = nms(cls_dets)
+            all_boxes[j][i] = cls_dets[keep, :]
+
+        if max_per_image > 0:
+            image_scores = np.hstack(
+                [all_boxes[j][i][:, -1] for j in range(1, imdb.num_classes)])
+            if len(image_scores) > max_per_image:
+                image_thresh = np.sort(image_scores)[-max_per_image]
+                for j in range(1, imdb.num_classes):
+                    keep = np.where(all_boxes[j][i][:, -1] >= image_thresh)[0]
+                    all_boxes[j][i] = all_boxes[j][i][keep, :]
+
+        if vis:
+            boxes_this_image = [[]] + [
+                all_boxes[j][i] for j in range(1, imdb.num_classes)
+            ]
+            vis_all_detection(data_dict['data'].asnumpy(), boxes_this_image,
+                              imdb.classes, scale)
+
+        t3 = time.time() - t
+        t = time.time()
+        logger.info('testing %d/%d data %.4fs net %.4fs post %.4fs' %
+                    (i, imdb.num_images, t1, t2, t3))
+        i += 1
+
+    det_file = os.path.join(imdb.cache_path, imdb.name + '_detections.pkl')
+    with open(det_file, 'wb') as f:
+        pickle.dump(all_boxes, f, protocol=pickle.HIGHEST_PROTOCOL)
+
+    imdb.evaluate_detections(all_boxes)
+
+
+def vis_all_detection(im_array, detections, class_names, scale):
+    """
+    visualize all detections in one image
+    :param im_array: [b=1 c h w] in rgb
+    :param detections: [ numpy.ndarray([[x1 y1 x2 y2 score]]) for j in classes ]
+    :param class_names: list of names in imdb
+    :param scale: visualize the scaled image
+    :return:
+    """
+    import matplotlib.pyplot as plt
+    import random
+    im = image.transform_inverse(im_array, config.PIXEL_MEANS)
+    plt.imshow(im)
+    for j, name in enumerate(class_names):
+        if name == '__background__':
+            continue
+        color = (random.random(), random.random(), random.random()
+                 )  # generate a random color
+        dets = detections[j]
+        for det in dets:
+            bbox = det[:4] * scale
+            score = det[-1]
+            rect = plt.Rectangle((bbox[0], bbox[1]),
+                                 bbox[2] - bbox[0],
+                                 bbox[3] - bbox[1],
+                                 fill=False,
+                                 edgecolor=color,
+                                 linewidth=3.5)
+            plt.gca().add_patch(rect)
+            plt.gca().text(bbox[0],
+                           bbox[1] - 2,
+                           '{:s} {:.3f}'.format(name, score),
+                           bbox=dict(facecolor=color, alpha=0.5),
+                           fontsize=12,
+                           color='white')
+    plt.show()
+
+
+def draw_all_detection(im_array, detections, class_names, scale):
+    """
+    visualize all detections in one image
+    :param im_array: [b=1 c h w] in rgb
+    :param detections: [ numpy.ndarray([[x1 y1 x2 y2 score]]) for j in classes ]
+    :param class_names: list of names in imdb
+    :param scale: visualize the scaled image
+    :return:
+    """
+    import cv2
+    import random
+    color_white = (255, 255, 255)
+    im = image.transform_inverse(im_array, config.PIXEL_MEANS)
+    # change to bgr
+    im = cv2.cvtColor(im, cv2.cv.CV_RGB2BGR)
+    for j, name in enumerate(class_names):
+        if name == '__background__':
+            continue
+        color = (random.randint(0, 256), random.randint(0, 256),
+                 random.randint(0, 256))  # generate a random color
+        dets = detections[j]
+        for det in dets:
+            bbox = det[:4] * scale
+            score = det[-1]
+            bbox = map(int, bbox)
+            cv2.rectangle(im, (bbox[0], bbox[1]), (bbox[2], bbox[3]),
+                          color=color,
+                          thickness=2)
+            cv2.putText(im,
+                        '%s %.3f' % (class_names[j], score),
+                        (bbox[0], bbox[1] + 10),
+                        color=color_white,
+                        fontFace=cv2.FONT_HERSHEY_COMPLEX,
+                        fontScale=0.5)
+    return im
diff --git a/RetinaFace/rcnn/cython/.gitignore b/detection/retinaface/rcnn/cython/.gitignore
similarity index 100%
rename from RetinaFace/rcnn/cython/.gitignore
rename to detection/retinaface/rcnn/cython/.gitignore
diff --git a/recognition/partial_fc/pytorch/__init__.py b/detection/retinaface/rcnn/cython/__init__.py
similarity index 100%
rename from recognition/partial_fc/pytorch/__init__.py
rename to detection/retinaface/rcnn/cython/__init__.py
diff --git a/RetinaFace/rcnn/cython/anchors.pyx b/detection/retinaface/rcnn/cython/anchors.pyx
similarity index 100%
rename from RetinaFace/rcnn/cython/anchors.pyx
rename to detection/retinaface/rcnn/cython/anchors.pyx
diff --git a/RetinaFace/rcnn/cython/bbox.pyx b/detection/retinaface/rcnn/cython/bbox.pyx
similarity index 100%
rename from RetinaFace/rcnn/cython/bbox.pyx
rename to detection/retinaface/rcnn/cython/bbox.pyx
diff --git a/RetinaFace/rcnn/cython/cpu_nms.pyx b/detection/retinaface/rcnn/cython/cpu_nms.pyx
similarity index 100%
rename from RetinaFace/rcnn/cython/cpu_nms.pyx
rename to detection/retinaface/rcnn/cython/cpu_nms.pyx
diff --git a/RetinaFace/rcnn/cython/gpu_nms.hpp b/detection/retinaface/rcnn/cython/gpu_nms.hpp
similarity index 100%
rename from RetinaFace/rcnn/cython/gpu_nms.hpp
rename to detection/retinaface/rcnn/cython/gpu_nms.hpp
diff --git a/RetinaFace/rcnn/cython/gpu_nms.pyx b/detection/retinaface/rcnn/cython/gpu_nms.pyx
similarity index 100%
rename from RetinaFace/rcnn/cython/gpu_nms.pyx
rename to detection/retinaface/rcnn/cython/gpu_nms.pyx
diff --git a/RetinaFace/rcnn/cython/nms_kernel.cu b/detection/retinaface/rcnn/cython/nms_kernel.cu
similarity index 100%
rename from RetinaFace/rcnn/cython/nms_kernel.cu
rename to detection/retinaface/rcnn/cython/nms_kernel.cu
diff --git a/detection/retinaface/rcnn/cython/setup.py b/detection/retinaface/rcnn/cython/setup.py
new file mode 100644
index 000000000..1af1a1aed
--- /dev/null
+++ b/detection/retinaface/rcnn/cython/setup.py
@@ -0,0 +1,165 @@
+# --------------------------------------------------------
+# Fast R-CNN
+# Copyright (c) 2015 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+import os
+from os.path import join as pjoin
+from setuptools import setup
+from distutils.extension import Extension
+from Cython.Distutils import build_ext
+import numpy as np
+
+
+def find_in_path(name, path):
+    "Find a file in a search path"
+    # Adapted fom
+    # http://code.activestate.com/recipes/52224-find-a-file-given-a-search-path/
+    for dir in path.split(os.pathsep):
+        binpath = pjoin(dir, name)
+        if os.path.exists(binpath):
+            return os.path.abspath(binpath)
+    return None
+
+
+def locate_cuda():
+    """Locate the CUDA environment on the system
+
+    Returns a dict with keys 'home', 'nvcc', 'include', and 'lib64'
+    and values giving the absolute path to each directory.
+
+    Starts by looking for the CUDAHOME env variable. If not found, everything
+    is based on finding 'nvcc' in the PATH.
+    """
+
+    # first check if the CUDAHOME env variable is in use
+    if 'CUDAHOME' in os.environ:
+        home = os.environ['CUDAHOME']
+        nvcc = pjoin(home, 'bin', 'nvcc')
+    else:
+        # otherwise, search the PATH for NVCC
+        default_path = pjoin(os.sep, 'usr', 'local', 'cuda', 'bin')
+        nvcc = find_in_path('nvcc',
+                            os.environ['PATH'] + os.pathsep + default_path)
+        if nvcc is None:
+            raise EnvironmentError(
+                'The nvcc binary could not be '
+                'located in your $PATH. Either add it to your path, or set $CUDAHOME'
+            )
+        home = os.path.dirname(os.path.dirname(nvcc))
+
+    cudaconfig = {
+        'home': home,
+        'nvcc': nvcc,
+        'include': pjoin(home, 'include'),
+        'lib64': pjoin(home, 'lib64')
+    }
+    for k, v in cudaconfig.items():
+        if not os.path.exists(v):
+            raise EnvironmentError(
+                'The CUDA %s path could not be located in %s' % (k, v))
+
+    return cudaconfig
+
+
+# Test if cuda could be foun
+try:
+    CUDA = locate_cuda()
+except EnvironmentError:
+    CUDA = None
+
+# Obtain the numpy include directory.  This logic works across numpy versions.
+try:
+    numpy_include = np.get_include()
+except AttributeError:
+    numpy_include = np.get_numpy_include()
+
+
+def customize_compiler_for_nvcc(self):
+    """inject deep into distutils to customize how the dispatch
+    to gcc/nvcc works.
+
+    If you subclass UnixCCompiler, it's not trivial to get your subclass
+    injected in, and still have the right customizations (i.e.
+    distutils.sysconfig.customize_compiler) run on it. So instead of going
+    the OO route, I have this. Note, it's kindof like a wierd functional
+    subclassing going on."""
+
+    # tell the compiler it can processes .cu
+    self.src_extensions.append('.cu')
+
+    # save references to the default compiler_so and _comple methods
+    default_compiler_so = self.compiler_so
+    super = self._compile
+
+    # now redefine the _compile method. This gets executed for each
+    # object but distutils doesn't have the ability to change compilers
+    # based on source extension: we add it.
+    def _compile(obj, src, ext, cc_args, extra_postargs, pp_opts):
+        if os.path.splitext(src)[1] == '.cu':
+            # use the cuda for .cu files
+            self.set_executable('compiler_so', CUDA['nvcc'])
+            # use only a subset of the extra_postargs, which are 1-1 translated
+            # from the extra_compile_args in the Extension class
+            postargs = extra_postargs['nvcc']
+        else:
+            postargs = extra_postargs['gcc']
+
+        super(obj, src, ext, cc_args, postargs, pp_opts)
+        # reset the default compiler_so, which we might have changed for cuda
+        self.compiler_so = default_compiler_so
+
+    # inject our redefined _compile method into the class
+    self._compile = _compile
+
+
+# run the customize_compiler
+class custom_build_ext(build_ext):
+    def build_extensions(self):
+        customize_compiler_for_nvcc(self.compiler)
+        build_ext.build_extensions(self)
+
+
+ext_modules = [
+    Extension("bbox", ["bbox.pyx"],
+              extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
+              include_dirs=[numpy_include]),
+    Extension("anchors", ["anchors.pyx"],
+              extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
+              include_dirs=[numpy_include]),
+    Extension("cpu_nms", ["cpu_nms.pyx"],
+              extra_compile_args={'gcc': ["-Wno-cpp", "-Wno-unused-function"]},
+              include_dirs=[numpy_include]),
+]
+
+if CUDA is not None:
+    ext_modules.append(
+        Extension(
+            'gpu_nms',
+            ['nms_kernel.cu', 'gpu_nms.pyx'],
+            library_dirs=[CUDA['lib64']],
+            libraries=['cudart'],
+            language='c++',
+            runtime_library_dirs=[CUDA['lib64']],
+            # this syntax is specific to this build system
+            # we're only going to use certain compiler args with nvcc and not with
+            # gcc the implementation of this trick is in customize_compiler() below
+            extra_compile_args={
+                'gcc': ["-Wno-unused-function"],
+                'nvcc': [
+                    '-arch=sm_35', '--ptxas-options=-v', '-c',
+                    '--compiler-options', "'-fPIC'"
+                ]
+            },
+            include_dirs=[numpy_include, CUDA['include']]))
+else:
+    print('Skipping GPU_NMS')
+
+setup(
+    name='frcnn_cython',
+    ext_modules=ext_modules,
+    # inject our custom trigger
+    cmdclass={'build_ext': custom_build_ext},
+)
diff --git a/RetinaFace/rcnn/dataset/__init__.py b/detection/retinaface/rcnn/dataset/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/dataset/__init__.py
rename to detection/retinaface/rcnn/dataset/__init__.py
diff --git a/RetinaFace/rcnn/dataset/ds_utils.py b/detection/retinaface/rcnn/dataset/ds_utils.py
similarity index 100%
rename from RetinaFace/rcnn/dataset/ds_utils.py
rename to detection/retinaface/rcnn/dataset/ds_utils.py
diff --git a/detection/retinaface/rcnn/dataset/imdb.py b/detection/retinaface/rcnn/dataset/imdb.py
new file mode 100644
index 000000000..3c19817a8
--- /dev/null
+++ b/detection/retinaface/rcnn/dataset/imdb.py
@@ -0,0 +1,351 @@
+"""
+General image database
+An image database creates a list of relative image path called image_set_index and
+transform index to absolute image path. As to training, it is necessary that ground
+truth and proposals are mixed together for training.
+roidb
+basic format [image_index]
+['image', 'height', 'width', 'flipped',
+'boxes', 'gt_classes', 'gt_overlaps', 'max_classes', 'max_overlaps', 'bbox_targets']
+"""
+
+from ..logger import logger
+import os
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+import numpy as np
+from ..processing.bbox_transform import bbox_overlaps
+
+
+class IMDB(object):
+    def __init__(self, name, image_set, root_path, dataset_path):
+        """
+        basic information about an image database
+        :param name: name of image database will be used for any output
+        :param root_path: root path store cache and proposal data
+        :param dataset_path: dataset path store images and image lists
+        """
+        self.name = name + '_' + image_set
+        self.image_set = image_set
+        self.root_path = root_path
+        self.data_path = dataset_path
+
+        # abstract attributes
+        self.classes = []
+        self.num_classes = 0
+        self.image_set_index = []
+        self.num_images = 0
+
+        self.config = {}
+
+    def image_path_from_index(self, index):
+        raise NotImplementedError
+
+    def gt_roidb(self):
+        raise NotImplementedError
+
+    def evaluate_detections(self, detections):
+        raise NotImplementedError
+
+    @property
+    def cache_path(self):
+        """
+        make a directory to store all caches
+        :return: cache path
+        """
+        cache_path = os.path.join(self.root_path, 'cache')
+        if not os.path.exists(cache_path):
+            os.mkdir(cache_path)
+        return cache_path
+
+    def image_path_at(self, index):
+        """
+        access image at index in image database
+        :param index: image index in image database
+        :return: image path
+        """
+        return self.image_path_from_index(self.image_set_index[index])
+
+    def load_rpn_data(self, full=False):
+        if full:
+            rpn_file = os.path.join(self.root_path, 'rpn_data',
+                                    self.name + '_full_rpn.pkl')
+        else:
+            rpn_file = os.path.join(self.root_path, 'rpn_data',
+                                    self.name + '_rpn.pkl')
+        assert os.path.exists(
+            rpn_file), '%s rpn data not found at %s' % (self.name, rpn_file)
+        logger.info('%s loading rpn data from %s' % (self.name, rpn_file))
+        with open(rpn_file, 'rb') as f:
+            box_list = pickle.load(f)
+        return box_list
+
+    def load_rpn_roidb(self, gt_roidb):
+        """
+        turn rpn detection boxes into roidb
+        :param gt_roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        """
+        box_list = self.load_rpn_data()
+        return self.create_roidb_from_box_list(box_list, gt_roidb)
+
+    def rpn_roidb(self, gt_roidb, append_gt=False):
+        """
+        get rpn roidb and ground truth roidb
+        :param gt_roidb: ground truth roidb
+        :param append_gt: append ground truth
+        :return: roidb of rpn
+        """
+        if append_gt:
+            logger.info('%s appending ground truth annotations' % self.name)
+            rpn_roidb = self.load_rpn_roidb(gt_roidb)
+            roidb = IMDB.merge_roidbs(gt_roidb, rpn_roidb)
+        else:
+            roidb = self.load_rpn_roidb(gt_roidb)
+        return roidb
+
+    def create_roidb_from_box_list(self, box_list, gt_roidb):
+        """
+        given ground truth, prepare roidb
+        :param box_list: [image_index] ndarray of [box_index][x1, x2, y1, y2]
+        :param gt_roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        """
+        assert len(
+            box_list
+        ) == self.num_images, 'number of boxes matrix must match number of images'
+        roidb = []
+        for i in range(self.num_images):
+            roi_rec = dict()
+            roi_rec['image'] = gt_roidb[i]['image']
+            roi_rec['height'] = gt_roidb[i]['height']
+            roi_rec['width'] = gt_roidb[i]['width']
+
+            boxes = box_list[i]
+            if boxes.shape[1] == 5:
+                boxes = boxes[:, :4]
+            num_boxes = boxes.shape[0]
+            overlaps = np.zeros((num_boxes, self.num_classes),
+                                dtype=np.float32)
+            if gt_roidb is not None and gt_roidb[i]['boxes'].size > 0:
+                gt_boxes = gt_roidb[i]['boxes']
+                gt_classes = gt_roidb[i]['gt_classes']
+                # n boxes and k gt_boxes => n * k overlap
+                gt_overlaps = bbox_overlaps(boxes.astype(np.float),
+                                            gt_boxes.astype(np.float))
+                # for each box in n boxes, select only maximum overlap (must be greater than zero)
+                argmaxes = gt_overlaps.argmax(axis=1)
+                maxes = gt_overlaps.max(axis=1)
+                I = np.where(maxes > 0)[0]
+                overlaps[I, gt_classes[argmaxes[I]]] = maxes[I]
+
+            roi_rec.update({
+                'boxes':
+                boxes,
+                'gt_classes':
+                np.zeros((num_boxes, ), dtype=np.int32),
+                'gt_overlaps':
+                overlaps,
+                'max_classes':
+                overlaps.argmax(axis=1),
+                'max_overlaps':
+                overlaps.max(axis=1),
+                'flipped':
+                False
+            })
+
+            # background roi => background class
+            zero_indexes = np.where(roi_rec['max_overlaps'] == 0)[0]
+            assert all(roi_rec['max_classes'][zero_indexes] == 0)
+            # foreground roi => foreground class
+            nonzero_indexes = np.where(roi_rec['max_overlaps'] > 0)[0]
+            assert all(roi_rec['max_classes'][nonzero_indexes] != 0)
+
+            roidb.append(roi_rec)
+
+        return roidb
+
+    def append_flipped_images(self, roidb):
+        """
+        append flipped images to an roidb
+        flip boxes coordinates, images will be actually flipped when loading into network
+        :param roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        :return: roidb: [image_index]['boxes', 'gt_classes', 'gt_overlaps', 'flipped']
+        """
+        logger.info('%s append flipped images to roidb' % self.name)
+        assert self.num_images == len(roidb)
+        for i in range(self.num_images):
+            roi_rec = roidb[i]
+            entry = {
+                'image': roi_rec['image'],
+                'stream': roi_rec['stream'],
+                'height': roi_rec['height'],
+                'width': roi_rec['width'],
+                #'boxes': boxes,
+                'gt_classes': roidb[i]['gt_classes'],
+                'gt_overlaps': roidb[i]['gt_overlaps'],
+                'max_classes': roidb[i]['max_classes'],
+                'max_overlaps': roidb[i]['max_overlaps'],
+                'flipped': True
+            }
+            for k in roi_rec:
+                if not k.startswith('boxes'):
+                    continue
+                boxes = roi_rec[k].copy()
+                oldx1 = boxes[:, 0].copy()
+                oldx2 = boxes[:, 2].copy()
+                boxes[:, 0] = roi_rec['width'] - oldx2 - 1
+                boxes[:, 2] = roi_rec['width'] - oldx1 - 1
+                assert (boxes[:, 2] >= boxes[:, 0]).all()
+                entry[k] = boxes
+            if 'landmarks' in roi_rec:
+                k = 'landmarks'
+                landmarks = roi_rec[k].copy()
+                landmarks[:, :, 0] *= -1
+                landmarks[:, :, 0] += (roi_rec['width'] - 1)
+                #for a in range(0,10,2):
+                #  oldx1 = landmarks[:, a].copy()
+                #  landmarks[:,a] = roi_rec['width'] - oldx1 - 1
+                order = [1, 0, 2, 4, 3]
+                flandmarks = landmarks.copy()
+                for idx, a in enumerate(order):
+                    flandmarks[:, idx, :] = landmarks[:, a, :]
+
+                entry[k] = flandmarks
+            if 'blur' in roi_rec:
+                entry['blur'] = roi_rec['blur']
+            roidb.append(entry)
+
+        self.image_set_index *= 2
+        return roidb
+
+    def evaluate_recall(self, roidb, candidate_boxes=None, thresholds=None):
+        """
+        evaluate detection proposal recall metrics
+        record max overlap value for each gt box; return vector of overlap values
+        :param roidb: used to evaluate
+        :param candidate_boxes: if not given, use roidb's non-gt boxes
+        :param thresholds: array-like recall threshold
+        :return: None
+        ar: average recall, recalls: vector recalls at each IoU overlap threshold
+        thresholds: vector of IoU overlap threshold, gt_overlaps: vector of all ground-truth overlaps
+        """
+        area_names = [
+            'all', '0-25', '25-50', '50-100', '100-200', '200-300', '300-inf'
+        ]
+        area_ranges = [[0**2, 1e5**2], [0**2, 25**2], [25**2, 50**2],
+                       [50**2, 100**2], [100**2, 200**2], [200**2, 300**2],
+                       [300**2, 1e5**2]]
+        area_counts = []
+        for area_name, area_range in zip(area_names[1:], area_ranges[1:]):
+            area_count = 0
+            for i in range(self.num_images):
+                if candidate_boxes is None:
+                    # default is use the non-gt boxes from roidb
+                    non_gt_inds = np.where(roidb[i]['gt_classes'] == 0)[0]
+                    boxes = roidb[i]['boxes'][non_gt_inds, :]
+                else:
+                    boxes = candidate_boxes[i]
+                boxes_areas = (boxes[:, 2] - boxes[:, 0] +
+                               1) * (boxes[:, 3] - boxes[:, 1] + 1)
+                valid_range_inds = np.where((boxes_areas >= area_range[0])
+                                            & (boxes_areas < area_range[1]))[0]
+                area_count += len(valid_range_inds)
+            area_counts.append(area_count)
+        total_counts = float(sum(area_counts))
+        for area_name, area_count in zip(area_names[1:], area_counts):
+            logger.info('percentage of %s is %f' %
+                        (area_name, area_count / total_counts))
+        logger.info('average number of proposal is %f' %
+                    (total_counts / self.num_images))
+        for area_name, area_range in zip(area_names, area_ranges):
+            gt_overlaps = np.zeros(0)
+            num_pos = 0
+            for i in range(self.num_images):
+                # check for max_overlaps == 1 avoids including crowd annotations
+                max_gt_overlaps = roidb[i]['gt_overlaps'].max(axis=1)
+                gt_inds = np.where((roidb[i]['gt_classes'] > 0)
+                                   & (max_gt_overlaps == 1))[0]
+                gt_boxes = roidb[i]['boxes'][gt_inds, :]
+                gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] +
+                            1) * (gt_boxes[:, 3] - gt_boxes[:, 1] + 1)
+                valid_gt_inds = np.where((gt_areas >= area_range[0])
+                                         & (gt_areas < area_range[1]))[0]
+                gt_boxes = gt_boxes[valid_gt_inds, :]
+                num_pos += len(valid_gt_inds)
+
+                if candidate_boxes is None:
+                    # default is use the non-gt boxes from roidb
+                    non_gt_inds = np.where(roidb[i]['gt_classes'] == 0)[0]
+                    boxes = roidb[i]['boxes'][non_gt_inds, :]
+                else:
+                    boxes = candidate_boxes[i]
+                if boxes.shape[0] == 0:
+                    continue
+
+                overlaps = bbox_overlaps(boxes.astype(np.float),
+                                         gt_boxes.astype(np.float))
+
+                _gt_overlaps = np.zeros((gt_boxes.shape[0]))
+                # choose whatever is smaller to iterate
+                rounds = min(boxes.shape[0], gt_boxes.shape[0])
+                for j in range(rounds):
+                    # find which proposal maximally covers each gt box
+                    argmax_overlaps = overlaps.argmax(axis=0)
+                    # get the IoU amount of coverage for each gt box
+                    max_overlaps = overlaps.max(axis=0)
+                    # find which gt box is covered by most IoU
+                    gt_ind = max_overlaps.argmax()
+                    gt_ovr = max_overlaps.max()
+                    assert (gt_ovr >=
+                            0), '%s\n%s\n%s' % (boxes, gt_boxes, overlaps)
+                    # find the proposal box that covers the best covered gt box
+                    box_ind = argmax_overlaps[gt_ind]
+                    # record the IoU coverage of this gt box
+                    _gt_overlaps[j] = overlaps[box_ind, gt_ind]
+                    assert (_gt_overlaps[j] == gt_ovr)
+                    # mark the proposal box and the gt box as used
+                    overlaps[box_ind, :] = -1
+                    overlaps[:, gt_ind] = -1
+                # append recorded IoU coverage level
+                gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
+
+            gt_overlaps = np.sort(gt_overlaps)
+            if thresholds is None:
+                step = 0.05
+                thresholds = np.arange(0.5, 0.95 + 1e-5, step)
+            recalls = np.zeros_like(thresholds)
+
+            # compute recall for each IoU threshold
+            for i, t in enumerate(thresholds):
+                recalls[i] = (gt_overlaps >= t).sum() / float(num_pos)
+            ar = recalls.mean()
+
+            # print results
+            print('average recall for {}: {:.3f}, number:{}'.format(
+                area_name, ar, num_pos))
+            for threshold, recall in zip(thresholds, recalls):
+                print('recall @{:.2f}: {:.3f}'.format(threshold, recall))
+
+    @staticmethod
+    def merge_roidbs(a, b):
+        """
+        merge roidbs into one
+        :param a: roidb to be merged into
+        :param b: roidb to be merged
+        :return: merged imdb
+        """
+        assert len(a) == len(b)
+        for i in range(len(a)):
+            a[i]['boxes'] = np.vstack((a[i]['boxes'], b[i]['boxes']))
+            a[i]['gt_classes'] = np.hstack(
+                (a[i]['gt_classes'], b[i]['gt_classes']))
+            a[i]['gt_overlaps'] = np.vstack(
+                (a[i]['gt_overlaps'], b[i]['gt_overlaps']))
+            a[i]['max_classes'] = np.hstack(
+                (a[i]['max_classes'], b[i]['max_classes']))
+            a[i]['max_overlaps'] = np.hstack(
+                (a[i]['max_overlaps'], b[i]['max_overlaps']))
+        return a
diff --git a/detection/retinaface/rcnn/dataset/retinaface.py b/detection/retinaface/rcnn/dataset/retinaface.py
new file mode 100644
index 000000000..6e3b85689
--- /dev/null
+++ b/detection/retinaface/rcnn/dataset/retinaface.py
@@ -0,0 +1,197 @@
+from __future__ import print_function
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+import cv2
+import os
+import numpy as np
+import json
+#from PIL import Image
+
+from ..logger import logger
+from .imdb import IMDB
+from .ds_utils import unique_boxes, filter_small_boxes
+from ..config import config
+
+
+class retinaface(IMDB):
+    def __init__(self, image_set, root_path, data_path):
+        super(retinaface, self).__init__('retinaface', image_set, root_path,
+                                         data_path)
+        #assert image_set=='train'
+
+        split = image_set
+        self._split = image_set
+        self._image_set = image_set
+
+        self.root_path = root_path
+        self.data_path = data_path
+
+        self._dataset_path = self.data_path
+        self._imgs_path = os.path.join(self._dataset_path, image_set, 'images')
+        self._fp_bbox_map = {}
+        label_file = os.path.join(self._dataset_path, image_set, 'label.txt')
+        name = None
+        for line in open(label_file, 'r'):
+            line = line.strip()
+            if line.startswith('#'):
+                name = line[1:].strip()
+                self._fp_bbox_map[name] = []
+                continue
+            assert name is not None
+            assert name in self._fp_bbox_map
+            self._fp_bbox_map[name].append(line)
+        print('origin image size', len(self._fp_bbox_map))
+
+        #self.num_images = len(self._image_paths)
+        #self._image_index = range(len(self._image_paths))
+        self.classes = ['bg', 'face']
+        self.num_classes = len(self.classes)
+
+    def gt_roidb(self):
+        cache_file = os.path.join(
+            self.cache_path,
+            '{}_{}_gt_roidb.pkl'.format(self.name, self._split))
+        if os.path.exists(cache_file):
+            with open(cache_file, 'rb') as fid:
+                roidb = pickle.load(fid)
+            print('{} gt roidb loaded from {}'.format(self.name, cache_file))
+            self.num_images = len(roidb)
+            return roidb
+
+        roidb = []
+        max_num_boxes = 0
+        nonattr_box_num = 0
+        landmark_num = 0
+
+        pp = 0
+        for fp in self._fp_bbox_map:
+            pp += 1
+            if pp % 1000 == 0:
+                print('loading', pp)
+            if self._split == 'test':
+                image_path = os.path.join(self._imgs_path, fp)
+                roi = {'image': image_path}
+                roidb.append(roi)
+                continue
+            boxes = np.zeros([len(self._fp_bbox_map[fp]), 4], np.float)
+            landmarks = np.zeros([len(self._fp_bbox_map[fp]), 5, 3], np.float)
+            blur = np.zeros((len(self._fp_bbox_map[fp]), ), np.float)
+            boxes_mask = []
+
+            gt_classes = np.ones([len(self._fp_bbox_map[fp])], np.int32)
+            overlaps = np.zeros([len(self._fp_bbox_map[fp]), 2], np.float)
+
+            imsize = cv2.imread(os.path.join(self._imgs_path,
+                                             fp)).shape[0:2][::-1]
+            ix = 0
+
+            for aline in self._fp_bbox_map[fp]:
+                #imsize = Image.open(os.path.join(self._imgs_path, fp)).size
+                values = [float(x) for x in aline.strip().split()]
+                bbox = [
+                    values[0], values[1], values[0] + values[2],
+                    values[1] + values[3]
+                ]
+
+                x1 = bbox[0]
+                y1 = bbox[1]
+                x2 = min(imsize[0], bbox[2])
+                y2 = min(imsize[1], bbox[3])
+                if x1 >= x2 or y1 >= y2:
+                    continue
+
+                if config.BBOX_MASK_THRESH > 0:
+                    if (
+                            x2 - x1
+                    ) < config.BBOX_MASK_THRESH or y2 - y1 < config.BBOX_MASK_THRESH:
+                        boxes_mask.append(np.array([x1, y1, x2, y2], np.float))
+                        continue
+                if (
+                        x2 - x1
+                ) < config.TRAIN.MIN_BOX_SIZE or y2 - y1 < config.TRAIN.MIN_BOX_SIZE:
+                    continue
+
+                boxes[ix, :] = np.array([x1, y1, x2, y2], np.float)
+                if self._split == 'train':
+                    landmark = np.array(values[4:19],
+                                        dtype=np.float32).reshape((5, 3))
+                    for li in range(5):
+                        #print(landmark)
+                        if landmark[li][0] == -1. and landmark[li][
+                                1] == -1.:  #missing landmark
+                            assert landmark[li][2] == -1
+                        else:
+                            assert landmark[li][2] >= 0
+                            if li == 0:
+                                landmark_num += 1
+                            if landmark[li][2] == 0.0:  #visible
+                                landmark[li][2] = 1.0
+                            else:
+                                landmark[li][2] = 0.0
+
+                    landmarks[ix] = landmark
+
+                    blur[ix] = values[19]
+                    #print(aline, blur[ix])
+                    if blur[ix] < 0.0:
+                        blur[ix] = 0.3
+                        nonattr_box_num += 1
+
+                cls = int(1)
+                gt_classes[ix] = cls
+                overlaps[ix, cls] = 1.0
+                ix += 1
+            max_num_boxes = max(max_num_boxes, ix)
+            #overlaps = scipy.sparse.csr_matrix(overlaps)
+            if self._split == 'train' and ix == 0:
+                continue
+            boxes = boxes[:ix, :]
+            landmarks = landmarks[:ix, :, :]
+            blur = blur[:ix]
+            gt_classes = gt_classes[:ix]
+            overlaps = overlaps[:ix, :]
+            image_path = os.path.join(self._imgs_path, fp)
+            with open(image_path, 'rb') as fin:
+                stream = fin.read()
+            stream = np.fromstring(stream, dtype=np.uint8)
+
+            roi = {
+                'image': image_path,
+                'stream': stream,
+                'height': imsize[1],
+                'width': imsize[0],
+                'boxes': boxes,
+                'landmarks': landmarks,
+                'blur': blur,
+                'gt_classes': gt_classes,
+                'gt_overlaps': overlaps,
+                'max_classes': overlaps.argmax(axis=1),
+                'max_overlaps': overlaps.max(axis=1),
+                'flipped': False,
+            }
+            if len(boxes_mask) > 0:
+                boxes_mask = np.array(boxes_mask)
+                roi['boxes_mask'] = boxes_mask
+            roidb.append(roi)
+        for roi in roidb:
+            roi['max_num_boxes'] = max_num_boxes
+        self.num_images = len(roidb)
+        print('roidb size', len(roidb))
+        print('non attr box num', nonattr_box_num)
+        print('landmark num', landmark_num)
+        with open(cache_file, 'wb') as fid:
+            pickle.dump(roidb, fid, pickle.HIGHEST_PROTOCOL)
+        print('wrote gt roidb to {}'.format(cache_file))
+
+        return roidb
+
+    def write_detections(self, all_boxes, output_dir='./output/'):
+        pass
+
+    def evaluate_detections(self,
+                            all_boxes,
+                            output_dir='./output/',
+                            method_name='insightdetection'):
+        pass
diff --git a/src/align/__init__.py b/detection/retinaface/rcnn/io/__init__.py
similarity index 100%
rename from src/align/__init__.py
rename to detection/retinaface/rcnn/io/__init__.py
diff --git a/detection/retinaface/rcnn/io/image.py b/detection/retinaface/rcnn/io/image.py
new file mode 100644
index 000000000..0296fb4de
--- /dev/null
+++ b/detection/retinaface/rcnn/io/image.py
@@ -0,0 +1,886 @@
+from __future__ import print_function
+import numpy as np
+import cv2
+import os
+import math
+import sys
+import random
+from ..config import config
+
+
+def brightness_aug(src, x):
+    alpha = 1.0 + random.uniform(-x, x)
+    src *= alpha
+    return src
+
+
+def contrast_aug(src, x):
+    alpha = 1.0 + random.uniform(-x, x)
+    coef = np.array([[[0.299, 0.587, 0.114]]])
+    gray = src * coef
+    gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
+    src *= alpha
+    src += gray
+    return src
+
+
+def saturation_aug(src, x):
+    alpha = 1.0 + random.uniform(-x, x)
+    coef = np.array([[[0.299, 0.587, 0.114]]])
+    gray = src * coef
+    gray = np.sum(gray, axis=2, keepdims=True)
+    gray *= (1.0 - alpha)
+    src *= alpha
+    src += gray
+    return src
+
+
+def color_aug(img, x):
+    if config.COLOR_MODE > 1:
+        augs = [brightness_aug, contrast_aug, saturation_aug]
+        random.shuffle(augs)
+    else:
+        augs = [brightness_aug]
+    for aug in augs:
+        #print(img.shape)
+        img = aug(img, x)
+        #print(img.shape)
+    return img
+
+
+def get_image(roidb, scale=False):
+    """
+    preprocess image and return processed roidb
+    :param roidb: a list of roidb
+    :return: list of img as in mxnet format
+    roidb add new item['im_info']
+    0 --- x (width, second dim of im)
+    |
+    y (height, first dim of im)
+    """
+    num_images = len(roidb)
+    processed_ims = []
+    processed_roidb = []
+    for i in range(num_images):
+        roi_rec = roidb[i]
+        if 'stream' in roi_rec:
+            im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
+        else:
+            assert os.path.exists(
+                roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
+            im = cv2.imread(roi_rec['image'])
+        if roidb[i]['flipped']:
+            im = im[:, ::-1, :]
+        new_rec = roi_rec.copy()
+        if scale:
+            scale_range = config.TRAIN.SCALE_RANGE
+            im_scale = np.random.uniform(scale_range[0], scale_range[1])
+            im = cv2.resize(im,
+                            None,
+                            None,
+                            fx=im_scale,
+                            fy=im_scale,
+                            interpolation=cv2.INTER_LINEAR)
+        elif not config.ORIGIN_SCALE:
+            scale_ind = random.randrange(len(config.SCALES))
+            target_size = config.SCALES[scale_ind][0]
+            max_size = config.SCALES[scale_ind][1]
+            im, im_scale = resize(im,
+                                  target_size,
+                                  max_size,
+                                  stride=config.IMAGE_STRIDE)
+        else:
+            im_scale = 1.0
+        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS)
+        if 'boxes_mask' in roi_rec:
+            im = im.astype(np.float32)
+            boxes_mask = roi_rec['boxes_mask'].copy() * im_scale
+            boxes_mask = boxes_mask.astype(np.int)
+            for j in range(boxes_mask.shape[0]):
+                m = boxes_mask[j]
+                im_tensor[:, :, m[1]:m[3], m[0]:m[2]] = 0.0
+                #print('find mask', m, file=sys.stderr)
+        processed_ims.append(im_tensor)
+        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
+        if config.TRAIN.IMAGE_ALIGN > 0:
+            if im_tensor.shape[
+                    2] % config.TRAIN.IMAGE_ALIGN != 0 or im_tensor.shape[
+                        3] % config.TRAIN.IMAGE_ALIGN != 0:
+                new_height = math.ceil(
+                    float(im_tensor.shape[2]) /
+                    config.TRAIN.IMAGE_ALIGN) * config.TRAIN.IMAGE_ALIGN
+                new_width = math.ceil(
+                    float(im_tensor.shape[3]) /
+                    config.TRAIN.IMAGE_ALIGN) * config.TRAIN.IMAGE_ALIGN
+                new_im_tensor = np.zeros(
+                    (1, 3, int(new_height), int(new_width)))
+                new_im_tensor[:, :, 0:im_tensor.shape[2],
+                              0:im_tensor.shape[3]] = im_tensor
+                print(im_tensor.shape, new_im_tensor.shape, file=sys.stderr)
+                im_tensor = new_im_tensor
+        #print('boxes', new_rec['boxes'], file=sys.stderr)
+        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
+        new_rec['im_info'] = im_info
+        processed_roidb.append(new_rec)
+    return processed_ims, processed_roidb
+
+
+TMP_ID = -1
+
+
+#bakup method
+def __get_crop_image(roidb):
+    """
+    preprocess image and return processed roidb
+    :param roidb: a list of roidb
+    :return: list of img as in mxnet format
+    roidb add new item['im_info']
+    0 --- x (width, second dim of im)
+    |
+    y (height, first dim of im)
+    """
+    #roidb and each roi_rec can not be changed as it will be reused in next epoch
+    num_images = len(roidb)
+    processed_ims = []
+    processed_roidb = []
+    for i in range(num_images):
+        roi_rec = roidb[i]
+        if 'stream' in roi_rec:
+            im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
+        else:
+            assert os.path.exists(
+                roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
+            im = cv2.imread(roi_rec['image'])
+        if roidb[i]['flipped']:
+            im = im[:, ::-1, :]
+        if 'boxes_mask' in roi_rec:
+            #im = im.astype(np.float32)
+            boxes_mask = roi_rec['boxes_mask'].copy()
+            boxes_mask = boxes_mask.astype(np.int)
+            for j in range(boxes_mask.shape[0]):
+                m = boxes_mask[j]
+                im[m[1]:m[3], m[0]:m[2], :] = 0
+                #print('find mask', m, file=sys.stderr)
+        new_rec = roi_rec.copy()
+
+        #choose one gt randomly
+        SIZE = config.SCALES[0][0]
+        TARGET_BOX_SCALES = np.array([16, 32, 64, 128, 256, 512])
+        assert roi_rec['boxes'].shape[0] > 0
+        candidates = []
+        for i in range(roi_rec['boxes'].shape[0]):
+            box = roi_rec['boxes'][i]
+            box_size = max(box[2] - box[0], box[3] - box[1])
+            if box_size < config.TRAIN.MIN_BOX_SIZE:
+                continue
+            #if box[0]<0 or box[1]<0:
+            #  continue
+            #if box[2]>im.shape[1] or box[3]>im.shape[0]:
+            #  continue;
+            candidates.append(i)
+        assert len(candidates) > 0
+        box_ind = random.choice(candidates)
+        box = roi_rec['boxes'][box_ind]
+        box_size = max(box[2] - box[0], box[3] - box[1])
+        dist = np.abs(TARGET_BOX_SCALES - box_size)
+        nearest = np.argmin(dist)
+        target_ind = random.randrange(min(len(TARGET_BOX_SCALES), nearest + 2))
+        target_box_size = TARGET_BOX_SCALES[target_ind]
+        im_scale = float(target_box_size) / box_size
+        #min_scale = float(SIZE)/np.min(im.shape[0:2])
+        #if im_scale<min_scale:
+        #  im_scale = min_scale
+        im = cv2.resize(im,
+                        None,
+                        None,
+                        fx=im_scale,
+                        fy=im_scale,
+                        interpolation=cv2.INTER_LINEAR)
+        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
+        box_scale = new_rec['boxes'][box_ind].copy().astype(np.int)
+        ul_min = box_scale[2:4] - SIZE
+        ul_max = box_scale[0:2]
+        assert ul_min[0] <= ul_max[0]
+        assert ul_min[1] <= ul_max[1]
+        #print('ul', ul_min, ul_max, box)
+        up, left = np.random.randint(ul_min[1],
+                                     ul_max[1] + 1), np.random.randint(
+                                         ul_min[0], ul_max[0] + 1)
+        #print('box', box, up, left)
+        M = [
+            [1.0, 0.0, -left],
+            [0.0, 1.0, -up],
+        ]
+        M = np.array(M)
+        im = cv2.warpAffine(im,
+                            M, (SIZE, SIZE),
+                            borderValue=tuple(config.PIXEL_MEANS))
+        #tbox = np.array([left, left+SIZE, up, up+SIZE], dtype=np.int)
+        #im_new = np.zeros( (SIZE, SIZE,3), dtype=im.dtype)
+        #for i in range(3):
+        #  im_new[:,:,i] = config.PIXEL_MEANS[i]
+        new_rec['boxes'][:, 0] -= left
+        new_rec['boxes'][:, 2] -= left
+        new_rec['boxes'][:, 1] -= up
+        new_rec['boxes'][:, 3] -= up
+        box_trans = new_rec['boxes'][box_ind].copy().astype(np.int)
+        #print('sel box', im_scale, box, box_scale, box_trans, file=sys.stderr)
+        #print('before', new_rec['boxes'].shape[0])
+        boxes_new = []
+        classes_new = []
+        for i in range(new_rec['boxes'].shape[0]):
+            box = new_rec['boxes'][i]
+            box_size = max(box[2] - box[0], box[3] - box[1])
+            center = np.array(([box[0], box[1]] + [box[2], box[3]])) / 2
+            if center[0] < 0 or center[1] < 0 or center[0] >= im.shape[
+                    1] or center[1] >= im.shape[0]:
+                continue
+            if box_size < config.TRAIN.MIN_BOX_SIZE:
+                continue
+            boxes_new.append(box)
+            classes_new.append(new_rec['gt_classes'][i])
+        new_rec['boxes'] = np.array(boxes_new)
+        new_rec['gt_classes'] = np.array(classes_new)
+        #print('after', new_rec['boxes'].shape[0])
+        #assert new_rec['boxes'].shape[0]>0
+        DEBUG = True
+        if DEBUG:
+            global TMP_ID
+            if TMP_ID < 10:
+                tim = im.copy()
+                for i in range(new_rec['boxes'].shape[0]):
+                    box = new_rec['boxes'][i].copy().astype(np.int)
+                    cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]),
+                                  (255, 0, 0), 1)
+                filename = './trainimages/train%d.png' % TMP_ID
+                TMP_ID += 1
+                cv2.imwrite(filename, tim)
+
+        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS,
+                              config.PIXEL_SCALE)
+
+        processed_ims.append(im_tensor)
+        #print('boxes', new_rec['boxes'], file=sys.stderr)
+        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
+        new_rec['im_info'] = im_info
+        processed_roidb.append(new_rec)
+    return processed_ims, processed_roidb
+
+
+def expand_bboxes(bboxes,
+                  image_width,
+                  image_height,
+                  expand_left=2.,
+                  expand_up=2.,
+                  expand_right=2.,
+                  expand_down=2.):
+    """
+    Expand bboxes, expand 2 times by defalut.
+    """
+    expand_boxes = []
+    for bbox in bboxes:
+        xmin = bbox[0]
+        ymin = bbox[1]
+        xmax = bbox[2]
+        ymax = bbox[3]
+        w = xmax - xmin
+        h = ymax - ymin
+        ex_xmin = max(xmin - w / expand_left, 0.)
+        ex_ymin = max(ymin - h / expand_up, 0.)
+        ex_xmax = min(xmax + w / expand_right, image_width)
+        ex_ymax = min(ymax + h / expand_down, image_height)
+        expand_boxes.append([ex_xmin, ex_ymin, ex_xmax, ex_ymax])
+    return expand_boxes
+
+
+def get_crop_image1(roidb):
+    """
+    preprocess image and return processed roidb
+    :param roidb: a list of roidb
+    :return: list of img as in mxnet format
+    roidb add new item['im_info']
+    0 --- x (width, second dim of im)
+    |
+    y (height, first dim of im)
+    """
+    #roidb and each roi_rec can not be changed as it will be reused in next epoch
+    num_images = len(roidb)
+    processed_ims = []
+    processed_roidb = []
+    for i in range(num_images):
+        roi_rec = roidb[i]
+        if 'stream' in roi_rec:
+            im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
+        else:
+            assert os.path.exists(
+                roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
+            im = cv2.imread(roi_rec['image'])
+        if roidb[i]['flipped']:
+            im = im[:, ::-1, :]
+        if 'boxes_mask' in roi_rec:
+            #im = im.astype(np.float32)
+            boxes_mask = roi_rec['boxes_mask'].copy()
+            boxes_mask = boxes_mask.astype(np.int)
+            for j in range(boxes_mask.shape[0]):
+                m = boxes_mask[j]
+                im[m[1]:m[3], m[0]:m[2], :] = 127
+                #print('find mask', m, file=sys.stderr)
+        SIZE = config.SCALES[0][0]
+        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
+        #PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0, 0.8, 1.0, 0.8, 1.0]
+        _scale = random.choice(PRE_SCALES)
+        #_scale = np.random.uniform(PRE_SCALES[0], PRE_SCALES[-1])
+        size = int(np.min(im.shape[0:2]) * _scale)
+        #size = int(np.round(_scale*np.min(im.shape[0:2])))
+        im_scale = float(SIZE) / size
+        #origin_im_scale = im_scale
+        #size = np.round(np.min(im.shape[0:2])*im_scale)
+        #im_scale *= (float(SIZE)/size)
+        origin_shape = im.shape
+        if _scale > 10.0:  #avoid im.size<SIZE, never?
+            sizex = int(np.round(im.shape[1] * im_scale))
+            sizey = int(np.round(im.shape[0] * im_scale))
+            if sizex < SIZE:
+                sizex = SIZE
+                print('keepx', sizex)
+            if sizey < SIZE:
+                sizey = SIZE
+                print('keepy', sizex)
+            im = cv2.resize(im, (sizex, sizey), interpolation=cv2.INTER_LINEAR)
+        else:
+            im = cv2.resize(im,
+                            None,
+                            None,
+                            fx=im_scale,
+                            fy=im_scale,
+                            interpolation=cv2.INTER_LINEAR)
+
+        assert im.shape[0] >= SIZE and im.shape[1] >= SIZE
+        #print('image size', origin_shape, _scale, SIZE, size, im_scale)
+
+        new_rec = roi_rec.copy()
+        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
+        if config.FACE_LANDMARK:
+            new_rec['landmarks'] = roi_rec['landmarks'].copy()
+            new_rec['landmarks'][:, :, 0:2] *= im_scale
+        retry = 0
+        LIMIT = 25
+        size = SIZE
+        while retry < LIMIT:
+            up, left = (np.random.randint(0, im.shape[0] - size + 1),
+                        np.random.randint(0, im.shape[1] - size + 1))
+            boxes_new = new_rec['boxes'].copy()
+            im_new = im[up:(up + size), left:(left + size), :]
+            #print('crop', up, left, size, im_scale)
+            boxes_new[:, 0] -= left
+            boxes_new[:, 2] -= left
+            boxes_new[:, 1] -= up
+            boxes_new[:, 3] -= up
+            if config.FACE_LANDMARK:
+                landmarks_new = new_rec['landmarks'].copy()
+                landmarks_new[:, :, 0] -= left
+                landmarks_new[:, :, 1] -= up
+                #for i in range(0,10,2):
+                #  landmarks_new[:,i] -= left
+                #for i in range(1,10,2):
+                #  landmarks_new[:,i] -= up
+                valid_landmarks = []
+            #im_new = cv2.resize(im_new, (SIZE, SIZE), interpolation=cv2.INTER_LINEAR)
+            #boxes_new *= im_scale
+            #print(origin_shape, im_new.shape, im_scale)
+            valid = []
+            valid_boxes = []
+            for i in range(boxes_new.shape[0]):
+                box = boxes_new[i]
+                #center = np.array(([box[0], box[1]]+[box[2], box[3]]))/2
+                centerx = (box[0] + box[2]) / 2
+                centery = (box[1] + box[3]) / 2
+
+                #box[0] = max(0, box[0])
+                #box[1] = max(0, box[1])
+                #box[2] = min(im_new.shape[1], box[2])
+                #box[3] = min(im_new.shape[0], box[3])
+                box_size = max(box[2] - box[0], box[3] - box[1])
+
+                if centerx < 0 or centery < 0 or centerx >= im_new.shape[
+                        1] or centery >= im_new.shape[0]:
+                    continue
+                if box_size < config.TRAIN.MIN_BOX_SIZE:
+                    continue
+                #filter by landmarks? TODO
+                valid.append(i)
+                valid_boxes.append(box)
+                if config.FACE_LANDMARK:
+                    valid_landmarks.append(landmarks_new[i])
+            if len(valid) > 0 or retry == LIMIT - 1:
+                im = im_new
+                new_rec['boxes'] = np.array(valid_boxes)
+                new_rec['gt_classes'] = new_rec['gt_classes'][valid]
+                if config.FACE_LANDMARK:
+                    new_rec['landmarks'] = np.array(valid_landmarks)
+                if config.HEAD_BOX:
+                    face_box = new_rec['boxes']
+                    head_box = expand_bboxes(face_box,
+                                             image_width=im.shape[1],
+                                             image_height=im.shape[0])
+                    new_rec['boxes_head'] = np.array(head_box)
+                break
+
+            retry += 1
+
+        if config.COLOR_MODE > 0 and config.COLOR_JITTERING > 0.0:
+            im = im.astype(np.float32)
+            im = color_aug(im, config.COLOR_JITTERING)
+
+        #assert np.all(new_rec['landmarks'][:,10]>0.0)
+        global TMP_ID
+        if TMP_ID >= 0 and TMP_ID < 10:
+            tim = im.copy().astype(np.uint8)
+            for i in range(new_rec['boxes'].shape[0]):
+                box = new_rec['boxes'][i].copy().astype(np.int)
+                cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]),
+                              (255, 0, 0), 1)
+                print('draw box:', box)
+            if config.FACE_LANDMARK:
+                for i in range(new_rec['landmarks'].shape[0]):
+                    landmark = new_rec['landmarks'][i].copy()
+                    if landmark[0][2] < 0:
+                        print('zero', landmark)
+                        continue
+                    landmark = landmark.astype(np.int)
+                    print('draw landmark', landmark)
+                    for k in range(5):
+                        color = (0, 0, 255)
+                        if k == 0 or k == 3:
+                            color = (0, 255, 0)
+                        pp = (landmark[k][0], landmark[k][1])
+                        cv2.circle(tim, (pp[0], pp[1]), 1, color, 2)
+            filename = './trainimages/train%d.png' % TMP_ID
+            print('write', filename)
+            cv2.imwrite(filename, tim)
+            TMP_ID += 1
+
+        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS,
+                              config.PIXEL_SCALE)
+
+        processed_ims.append(im_tensor)
+        #print('boxes', new_rec['boxes'], file=sys.stderr)
+        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
+        new_rec['im_info'] = np.array(im_info, dtype=np.float32)
+        processed_roidb.append(new_rec)
+    return processed_ims, processed_roidb
+
+
+def get_crop_image2(roidb):
+    """
+    preprocess image and return processed roidb
+    :param roidb: a list of roidb
+    :return: list of img as in mxnet format
+    roidb add new item['im_info']
+    0 --- x (width, second dim of im)
+    |
+    y (height, first dim of im)
+    """
+    #roidb and each roi_rec can not be changed as it will be reused in next epoch
+    num_images = len(roidb)
+    processed_ims = []
+    processed_roidb = []
+    for i in range(num_images):
+        roi_rec = roidb[i]
+        if 'stream' in roi_rec:
+            im = cv2.imdecode(roi_rec['stream'], cv2.IMREAD_COLOR)
+        else:
+            assert os.path.exists(
+                roi_rec['image']), '{} does not exist'.format(roi_rec['image'])
+            im = cv2.imread(roi_rec['image'])
+        if roidb[i]['flipped']:
+            im = im[:, ::-1, :]
+        if 'boxes_mask' in roi_rec:
+            #im = im.astype(np.float32)
+            boxes_mask = roi_rec['boxes_mask'].copy()
+            boxes_mask = boxes_mask.astype(np.int)
+            for j in range(boxes_mask.shape[0]):
+                m = boxes_mask[j]
+                im[m[1]:m[3], m[0]:m[2], :] = 0
+                #print('find mask', m, file=sys.stderr)
+        SIZE = config.SCALES[0][0]
+        scale_array = np.array([16, 32, 64, 128, 256, 512], dtype=np.float32)
+        candidates = []
+        for i in range(roi_rec['boxes'].shape[0]):
+            box = roi_rec['boxes'][i]
+            box_size = max(box[2] - box[0], box[3] - box[1])
+            if box_size < config.TRAIN.MIN_BOX_SIZE:
+                continue
+            #if box[0]<0 or box[1]<0:
+            #  continue
+            #if box[2]>im.shape[1] or box[3]>im.shape[0]:
+            #  continue;
+            candidates.append(i)
+        assert len(candidates) > 0
+        box_ind = random.choice(candidates)
+        box = roi_rec['boxes'][box_ind]
+        width = box[2] - box[0]
+        height = box[3] - box[1]
+        wid = width
+        hei = height
+        resize_width, resize_height = config.SCALES[0]
+        image_width = im.shape[0]
+        image_height = im.shape[1]
+        area = width * height
+        range_size = 0
+        for scale_ind in range(0, len(scale_array) - 1):
+            if area > scale_array[scale_ind] ** 2 and area < \
+                    scale_array[scale_ind + 1] ** 2:
+                range_size = scale_ind + 1
+                break
+
+        if area > scale_array[len(scale_array) - 2]**2:
+            range_size = len(scale_array) - 2
+        scale_choose = 0.0
+        if range_size == 0:
+            rand_idx_size = 0
+        else:
+            # np.random.randint range: [low, high)
+            rng_rand_size = np.random.randint(0, range_size + 1)
+            rand_idx_size = rng_rand_size % (range_size + 1)
+
+        if rand_idx_size == range_size:
+            min_resize_val = scale_array[rand_idx_size] / 2.0
+            max_resize_val = min(2.0 * scale_array[rand_idx_size],
+                                 2 * math.sqrt(wid * hei))
+            scale_choose = random.uniform(min_resize_val, max_resize_val)
+        else:
+            min_resize_val = scale_array[rand_idx_size] / 2.0
+            max_resize_val = 2.0 * scale_array[rand_idx_size]
+            scale_choose = random.uniform(min_resize_val, max_resize_val)
+
+        sample_bbox_size = wid * resize_width / scale_choose
+
+        w_off_orig = 0.0
+        h_off_orig = 0.0
+        if sample_bbox_size < max(image_height, image_width):
+            if wid <= sample_bbox_size:
+                w_off_orig = np.random.uniform(xmin + wid - sample_bbox_size,
+                                               xmin)
+            else:
+                w_off_orig = np.random.uniform(xmin,
+                                               xmin + wid - sample_bbox_size)
+
+            if hei <= sample_bbox_size:
+                h_off_orig = np.random.uniform(ymin + hei - sample_bbox_size,
+                                               ymin)
+            else:
+                h_off_orig = np.random.uniform(ymin,
+                                               ymin + hei - sample_bbox_size)
+
+        else:
+            w_off_orig = np.random.uniform(image_width - sample_bbox_size, 0.0)
+            h_off_orig = np.random.uniform(image_height - sample_bbox_size,
+                                           0.0)
+
+        w_off_orig = math.floor(w_off_orig)
+        h_off_orig = math.floor(h_off_orig)
+
+        # Figure out top left coordinates.
+        w_off = 0.0
+        h_off = 0.0
+        w_off = float(w_off_orig / image_width)
+        h_off = float(h_off_orig / image_height)
+        im_new = im[up:(up + size), left:(left + size), :]
+
+        sampled_bbox = bbox(w_off, h_off,
+                            w_off + float(sample_bbox_size / image_width),
+                            h_off + float(sample_bbox_size / image_height))
+        return sampled_bbox
+
+        box_size = max(box[2] - box[0], box[3] - box[1])
+        dist = np.abs(TARGET_BOX_SCALES - box_size)
+        nearest = np.argmin(dist)
+        target_ind = random.randrange(min(len(TARGET_BOX_SCALES), nearest + 2))
+        target_box_size = TARGET_BOX_SCALES[target_ind]
+        im_scale = float(target_box_size) / box_size
+        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
+        _scale = random.choice(PRE_SCALES)
+        #_scale = np.random.uniform(PRE_SCALES[0], PRE_SCALES[-1])
+        size = int(np.round(_scale * np.min(im.shape[0:2])))
+        im_scale = float(SIZE) / size
+        #origin_im_scale = im_scale
+        #size = np.round(np.min(im.shape[0:2])*im_scale)
+        #im_scale *= (float(SIZE)/size)
+        origin_shape = im.shape
+        if _scale > 10.0:  #avoid im.size<SIZE, never?
+            sizex = int(np.round(im.shape[1] * im_scale))
+            sizey = int(np.round(im.shape[0] * im_scale))
+            if sizex < SIZE:
+                sizex = SIZE
+                print('keepx', sizex)
+            if sizey < SIZE:
+                sizey = SIZE
+                print('keepy', sizex)
+            im = cv2.resize(im, (sizex, sizey), interpolation=cv2.INTER_LINEAR)
+        else:
+            im = cv2.resize(im,
+                            None,
+                            None,
+                            fx=im_scale,
+                            fy=im_scale,
+                            interpolation=cv2.INTER_LINEAR)
+        assert im.shape[0] >= SIZE and im.shape[1] >= SIZE
+
+        new_rec = roi_rec.copy()
+        new_rec['boxes'] = roi_rec['boxes'].copy() * im_scale
+        if config.FACE_LANDMARK:
+            new_rec['landmarks'] = roi_rec['landmarks'].copy() * im_scale
+        retry = 0
+        LIMIT = 25
+        size = SIZE
+        while retry < LIMIT:
+            up, left = (np.random.randint(0, im.shape[0] - size + 1),
+                        np.random.randint(0, im.shape[1] - size + 1))
+            boxes_new = new_rec['boxes'].copy()
+            im_new = im[up:(up + size), left:(left + size), :]
+            #print('crop', up, left, size, im_scale)
+            boxes_new[:, 0] -= left
+            boxes_new[:, 2] -= left
+            boxes_new[:, 1] -= up
+            boxes_new[:, 3] -= up
+            if config.FACE_LANDMARK:
+                landmarks_new = new_rec['landmarks'].copy()
+                for i in range(0, 10, 2):
+                    landmarks_new[:, i] -= left
+                for i in range(1, 10, 2):
+                    landmarks_new[:, i] -= up
+                valid_landmarks = []
+            #im_new = cv2.resize(im_new, (SIZE, SIZE), interpolation=cv2.INTER_LINEAR)
+            #boxes_new *= im_scale
+            #print(origin_shape, im_new.shape, im_scale)
+            valid = []
+            valid_boxes = []
+            for i in range(boxes_new.shape[0]):
+                box = boxes_new[i]
+                #center = np.array(([box[0], box[1]]+[box[2], box[3]]))/2
+                centerx = (box[0] + box[2]) / 2
+                centery = (box[1] + box[3]) / 2
+
+                #box[0] = max(0, box[0])
+                #box[1] = max(0, box[1])
+                #box[2] = min(im_new.shape[1], box[2])
+                #box[3] = min(im_new.shape[0], box[3])
+                box_size = max(box[2] - box[0], box[3] - box[1])
+
+                if centerx < 0 or centery < 0 or centerx >= im_new.shape[
+                        1] or centery >= im_new.shape[0]:
+                    continue
+                if box_size < config.TRAIN.MIN_BOX_SIZE:
+                    continue
+                #filter by landmarks? TODO
+                valid.append(i)
+                valid_boxes.append(box)
+                if config.FACE_LANDMARK:
+                    valid_landmarks.append(landmarks_new[i])
+            if len(valid) > 0 or retry == LIMIT - 1:
+                im = im_new
+                new_rec['boxes'] = np.array(valid_boxes)
+                new_rec['gt_classes'] = new_rec['gt_classes'][valid]
+                if config.FACE_LANDMARK:
+                    new_rec['landmarks'] = np.array(valid_landmarks)
+                break
+
+            retry += 1
+
+        if config.COLOR_JITTERING > 0.0:
+            im = im.astype(np.float32)
+            im = color_aug(im, config.COLOR_JITTERING)
+
+        #assert np.all(new_rec['landmarks'][:,10]>0.0)
+        global TMP_ID
+        if TMP_ID >= 0 and TMP_ID < 10:
+            tim = im.copy().astype(np.uint8)
+            for i in range(new_rec['boxes'].shape[0]):
+                box = new_rec['boxes'][i].copy().astype(np.int)
+                cv2.rectangle(tim, (box[0], box[1]), (box[2], box[3]),
+                              (255, 0, 0), 1)
+                print('draw box:', box)
+            if config.FACE_LANDMARK:
+                for i in range(new_rec['landmarks'].shape[0]):
+                    landmark = new_rec['landmarks'][i].copy()
+                    if landmark[10] == 0.0:
+                        print('zero', landmark)
+                        continue
+                    landmark = landmark.astype(np.int)
+                    print('draw landmark', landmark)
+                    for k in range(5):
+                        color = (0, 0, 255)
+                        if k == 0 or k == 3:
+                            color = (0, 255, 0)
+                        pp = (landmark[k * 2], landmark[1 + k * 2])
+                        cv2.circle(tim, (pp[0], pp[1]), 1, color, 2)
+            filename = './trainimages/train%d.png' % TMP_ID
+            print('write', filename)
+            cv2.imwrite(filename, tim)
+            TMP_ID += 1
+
+        im_tensor = transform(im, config.PIXEL_MEANS, config.PIXEL_STDS,
+                              config.PIXEL_SCALE)
+
+        processed_ims.append(im_tensor)
+        #print('boxes', new_rec['boxes'], file=sys.stderr)
+        im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
+        new_rec['im_info'] = np.array(im_info, dtype=np.float32)
+        processed_roidb.append(new_rec)
+    return processed_ims, processed_roidb
+
+
+def do_mixup(im1, roidb1, im2, roidb2):
+    im = (im1 + im2) / 2.0
+    roidb = {}
+    #print(roidb1.keys())
+    #for k in roidb1:
+    for k in ['boxes', 'landmarks', 'gt_classes', 'im_info']:
+        v1 = roidb1[k]
+        v2 = roidb2[k]
+        if k != 'im_info':
+            #print('try', k, v1.shape, v2.shape)
+            if v1.shape[0] > 0 and v2.shape[0] > 0:
+                v = np.concatenate((v1, v2), axis=0)
+            else:
+                v = v1
+        else:
+            v = v1
+        #print(k, v1.shape, v2.shape, v.shape)
+        roidb[k] = v
+    return im, roidb
+
+
+def get_crop_image(roidb):
+    ims, roidbs = get_crop_image1(roidb)
+    if config.MIXUP > 0.0 and np.random.random() < config.MIXUP:
+        for i in range(len(ims)):
+            im = ims[i]
+            roidb = roidbs[i]
+            j = np.random.randint(0, len(ims) - 1)
+            if j >= i:
+                j += 1
+            im, roidb = do_mixup(im, roidb, ims[j], roidbs[j])
+            ims[i] = im
+            roidbs[i] = roidb
+    return ims, roidbs
+
+
+def resize(im, target_size, max_size, stride=0, min_size=0):
+    """
+    only resize input image to target size and return scale
+    :param im: BGR image input by opencv
+    :param target_size: one dimensional size (the short side)
+    :param max_size: one dimensional max size (the long side)
+    :param stride: if given, pad the image to designated stride
+    :return:
+    """
+    im_shape = im.shape
+    im_size_min = np.min(im_shape[0:2])
+    im_size_max = np.max(im_shape[0:2])
+    im_scale = float(target_size) / float(im_size_min)
+    # prevent bigger axis from being more than max_size:
+    if np.round(im_scale * im_size_max) > max_size:
+        im_scale = float(max_size) / float(im_size_max)
+        if min_size > 0 and np.round(im_scale * im_size_min) < min_size:
+            im_scale = float(min_size) / float(im_size_min)
+    im = cv2.resize(im,
+                    None,
+                    None,
+                    fx=im_scale,
+                    fy=im_scale,
+                    interpolation=cv2.INTER_LINEAR)
+
+    if stride == 0:
+        return im, im_scale
+    else:
+        # pad to product of stride
+        im_height = int(np.ceil(im.shape[0] / float(stride)) * stride)
+        im_width = int(np.ceil(im.shape[1] / float(stride)) * stride)
+        im_channel = im.shape[2]
+        padded_im = np.zeros((im_height, im_width, im_channel))
+        padded_im[:im.shape[0], :im.shape[1], :] = im
+        return padded_im, im_scale
+
+
+def transform(im, pixel_means, pixel_stds, pixel_scale):
+    """
+    transform into mxnet tensor,
+    subtract pixel size and transform to correct format
+    :param im: [height, width, channel] in BGR
+    :param pixel_means: [B, G, R pixel means]
+    :return: [batch, channel, height, width]
+    """
+    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
+    for i in range(3):
+        im_tensor[0, i, :, :] = (im[:, :, 2 - i] / pixel_scale -
+                                 pixel_means[2 - i]) / pixel_stds[2 - i]
+    return im_tensor
+
+
+def transform_inverse(im_tensor, pixel_means):
+    """
+    transform from mxnet im_tensor to ordinary RGB image
+    im_tensor is limited to one image
+    :param im_tensor: [batch, channel, height, width]
+    :param pixel_means: [B, G, R pixel means]
+    :return: im [height, width, channel(RGB)]
+    """
+    assert im_tensor.shape[0] == 1
+    im_tensor = im_tensor.copy()
+    # put channel back
+    channel_swap = (0, 2, 3, 1)
+    im_tensor = im_tensor.transpose(channel_swap)
+    im = im_tensor[0]
+    assert im.shape[2] == 3
+    im += pixel_means[[2, 1, 0]]
+    im = im.astype(np.uint8)
+    return im
+
+
+def tensor_vstack(tensor_list, pad=0):
+    """
+    vertically stack tensors
+    :param tensor_list: list of tensor to be stacked vertically
+    :param pad: label to pad with
+    :return: tensor with max shape
+    """
+    ndim = len(tensor_list[0].shape)
+    dtype = tensor_list[0].dtype
+    islice = tensor_list[0].shape[0]
+    dimensions = []
+    first_dim = sum([tensor.shape[0] for tensor in tensor_list])
+    dimensions.append(first_dim)
+    for dim in range(1, ndim):
+        dimensions.append(max([tensor.shape[dim] for tensor in tensor_list]))
+    if pad == 0:
+        all_tensor = np.zeros(tuple(dimensions), dtype=dtype)
+    elif pad == 1:
+        all_tensor = np.ones(tuple(dimensions), dtype=dtype)
+    else:
+        all_tensor = np.full(tuple(dimensions), pad, dtype=dtype)
+    if ndim == 1:
+        for ind, tensor in enumerate(tensor_list):
+            all_tensor[ind * islice:(ind + 1) * islice] = tensor
+    elif ndim == 2:
+        for ind, tensor in enumerate(tensor_list):
+            all_tensor[ind * islice:(ind + 1) *
+                       islice, :tensor.shape[1]] = tensor
+    elif ndim == 3:
+        for ind, tensor in enumerate(tensor_list):
+            all_tensor[ind * islice:(ind + 1) *
+                       islice, :tensor.shape[1], :tensor.shape[2]] = tensor
+    elif ndim == 4:
+        for ind, tensor in enumerate(tensor_list):
+            all_tensor[ind * islice:(ind + 1) * islice, :tensor.
+                       shape[1], :tensor.shape[2], :tensor.shape[3]] = tensor
+    elif ndim == 5:
+        for ind, tensor in enumerate(tensor_list):
+            all_tensor[ind * islice:(ind + 1) *
+                       islice, :tensor.shape[1], :tensor.shape[2], :tensor.
+                       shape[3], :tensor.shape[4]] = tensor
+    else:
+        print(tensor_list[0].shape)
+        raise Exception('Sorry, unimplemented.')
+    return all_tensor
diff --git a/detection/retinaface/rcnn/io/rcnn.py b/detection/retinaface/rcnn/io/rcnn.py
new file mode 100644
index 000000000..1b3a443f6
--- /dev/null
+++ b/detection/retinaface/rcnn/io/rcnn.py
@@ -0,0 +1,661 @@
+"""
+Fast R-CNN:
+data =
+    {'data': [num_images, c, h, w],
+    'rois': [num_rois, 5]}
+label =
+    {'label': [num_rois],
+    'bbox_target': [num_rois, 4 * num_classes],
+    'bbox_weight': [num_rois, 4 * num_classes]}
+roidb extended format [image_index]
+    ['image', 'height', 'width', 'flipped',
+     'boxes', 'gt_classes', 'gt_overlaps', 'max_classes', 'max_overlaps', 'bbox_targets']
+"""
+
+import numpy as np
+import numpy.random as npr
+
+from ..config import config
+from ..io.image import get_image, tensor_vstack
+from ..processing.bbox_transform import bbox_overlaps, bbox_transform
+from ..processing.bbox_regression import expand_bbox_regression_targets
+
+
+def get_rcnn_testbatch(roidb):
+    """
+    return a dict of testbatch
+    :param roidb: ['image', 'flipped'] + ['boxes']
+    :return: data, label, im_info
+    """
+    assert len(roidb) == 1, 'Single batch only'
+    imgs, roidb = get_image(roidb)
+    im_array = imgs[0]
+    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
+
+    im_rois = roidb[0]['boxes']
+    rois = im_rois
+    batch_index = 0 * np.ones((rois.shape[0], 1))
+    rois_array = np.hstack((batch_index, rois))[np.newaxis, :]
+
+    data = {'data': im_array, 'rois': rois_array, 'im_info': im_info}
+    label = {}
+
+    return data, label
+
+
+def get_rcnn_batch(roidb):
+    """
+    return a dict of multiple images
+    :param roidb: a list of dict, whose length controls batch size
+    ['images', 'flipped'] + ['gt_boxes', 'boxes', 'gt_overlap'] => ['bbox_targets']
+    :return: data, label
+    """
+    num_images = len(roidb)
+    imgs, roidb = get_image(roidb)
+    im_array = tensor_vstack(imgs)
+
+    assert config.TRAIN.BATCH_ROIS % config.TRAIN.BATCH_IMAGES == 0, \
+        'BATCHIMAGES {} must divide BATCH_ROIS {}'.format(config.TRAIN.BATCH_IMAGES, config.TRAIN.BATCH_ROIS)
+    rois_per_image = int(config.TRAIN.BATCH_ROIS / config.TRAIN.BATCH_IMAGES)
+    fg_rois_per_image = int(round(config.TRAIN.FG_FRACTION * rois_per_image))
+
+    rois_array = list()
+    labels_array = list()
+    bbox_targets_array = list()
+    bbox_weights_array = list()
+
+    for im_i in range(num_images):
+        roi_rec = roidb[im_i]
+
+        # infer num_classes from gt_overlaps
+        num_classes = roi_rec['gt_overlaps'].shape[1]
+
+        # label = class RoI has max overlap with
+        rois = roi_rec['boxes']
+        labels = roi_rec['max_classes']
+        overlaps = roi_rec['max_overlaps']
+        bbox_targets = roi_rec['bbox_targets']
+
+        im_rois, labels, bbox_targets, bbox_weights = \
+            sample_rois(rois, fg_rois_per_image, rois_per_image, num_classes,
+                        labels, overlaps, bbox_targets)
+
+        # project im_rois
+        # do not round roi
+        rois = im_rois
+        batch_index = im_i * np.ones((rois.shape[0], 1))
+        rois_array_this_image = np.hstack((batch_index, rois))
+        rois_array.append(rois_array_this_image)
+
+        # add labels
+        labels_array.append(labels)
+        bbox_targets_array.append(bbox_targets)
+        bbox_weights_array.append(bbox_weights)
+
+    rois_array = np.array(rois_array)
+    labels_array = np.array(labels_array)
+    bbox_targets_array = np.array(bbox_targets_array)
+    bbox_weights_array = np.array(bbox_weights_array)
+
+    data = {'data': im_array, 'rois': rois_array}
+    label = {
+        'label': labels_array,
+        'bbox_target': bbox_targets_array,
+        'bbox_weight': bbox_weights_array
+    }
+
+    return data, label
+
+
+def sample_rois(rois,
+                fg_rois_per_image,
+                rois_per_image,
+                num_classes,
+                labels=None,
+                overlaps=None,
+                bbox_targets=None,
+                gt_boxes=None):
+    """
+    generate random sample of ROIs comprising foreground and background examples
+    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
+    :param fg_rois_per_image: foreground roi number
+    :param rois_per_image: total roi number
+    :param num_classes: number of classes
+    :param labels: maybe precomputed
+    :param overlaps: maybe precomputed (max_overlaps)
+    :param bbox_targets: maybe precomputed
+    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
+    :return: (labels, rois, bbox_targets, bbox_weights)
+    """
+    if labels is None:
+        overlaps = bbox_overlaps(rois[:, 1:].astype(np.float),
+                                 gt_boxes[:, :4].astype(np.float))
+        gt_assignment = overlaps.argmax(axis=1)
+        overlaps = overlaps.max(axis=1)
+        labels = gt_boxes[gt_assignment, 4]
+
+    # foreground RoI with FG_THRESH overlap
+    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
+    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
+    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
+    # Sample foreground regions without replacement
+    if len(fg_indexes) > fg_rois_per_this_image:
+        fg_indexes = npr.choice(fg_indexes,
+                                size=fg_rois_per_this_image,
+                                replace=False)
+
+    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
+    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI)
+                          & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
+    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
+    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
+    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image,
+                                        bg_indexes.size)
+    # Sample foreground regions without replacement
+    if len(bg_indexes) > bg_rois_per_this_image:
+        bg_indexes = npr.choice(bg_indexes,
+                                size=bg_rois_per_this_image,
+                                replace=False)
+
+    # indexes selected
+    keep_indexes = np.append(fg_indexes, bg_indexes)
+    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
+    neg_rois = rois[neg_idx]
+    # pad more to ensure a fixed minibatch size
+    while keep_indexes.shape[0] < rois_per_image:
+        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
+        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
+        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
+
+    # select labels
+    labels = labels[keep_indexes]
+    # set labels of bg_rois to be 0
+    labels[fg_rois_per_this_image:] = 0
+    rois = rois[keep_indexes]
+
+    # load or compute bbox_target
+    if bbox_targets is not None:
+        bbox_target_data = bbox_targets[keep_indexes, :]
+    else:
+        targets = bbox_transform(rois[:, 1:],
+                                 gt_boxes[gt_assignment[keep_indexes], :4])
+        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS)) /
+                       np.array(config.TRAIN.BBOX_STDS))
+        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
+
+    bbox_targets, bbox_weights = \
+        expand_bbox_regression_targets(bbox_target_data, num_classes)
+
+    return rois, labels, bbox_targets, bbox_weights
+
+
+def get_fpn_rcnn_testbatch(roidb):
+    """
+    return a dict of testbatch
+    :param roidb: ['image', 'flipped'] + ['boxes']
+    :return: data, label, im_info
+    """
+    assert len(roidb) == 1, 'Single batch only'
+    imgs, roidb = get_image(roidb)
+    im_array = imgs[0]
+    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
+
+    im_rois = roidb[0]['boxes']
+    rois = im_rois
+
+    # assign rois
+    rois_area = np.sqrt((rois[:, 2] - rois[:, 0]) * (rois[:, 3] - rois[:, 1]))
+    area_threshold = {'P5': 448, 'P4': 224, 'P3': 112}
+    rois_p5 = rois[area_threshold['P5'] <= rois_area]
+    rois_p4 = rois[np.logical_and(area_threshold['P4'] <= rois_area,
+                                  rois_area < area_threshold['P5'])]
+    rois_p3 = rois[np.logical_and(area_threshold['P3'] <= rois_area,
+                                  rois_area < area_threshold['P4'])]
+    rois_p2 = rois[np.logical_and(0 < rois_area,
+                                  rois_area < area_threshold['P3'])]
+
+    # pad a virtual rois if on rois assigned
+    if rois_p5.size == 0:
+        rois_p5 = np.array([[12, 34, 56, 78]])
+    if rois_p4.size == 0:
+        rois_p4 = np.array([[12, 34, 56, 78]])
+    if rois_p3.size == 0:
+        rois_p3 = np.array([[12, 34, 56, 78]])
+    if rois_p2.size == 0:
+        rois_p2 = np.array([[12, 34, 56, 78]])
+
+    p5_batch_index = 0 * np.ones((rois_p5.shape[0], 1))
+    rois_p5_array = np.hstack((p5_batch_index, rois_p5))[np.newaxis, :]
+
+    p4_batch_index = 0 * np.ones((rois_p4.shape[0], 1))
+    rois_p4_array = np.hstack((p4_batch_index, rois_p4))[np.newaxis, :]
+
+    p3_batch_index = 0 * np.ones((rois_p3.shape[0], 1))
+    rois_p3_array = np.hstack((p3_batch_index, rois_p3))[np.newaxis, :]
+
+    p2_batch_index = 0 * np.ones((rois_p2.shape[0], 1))
+    rois_p2_array = np.hstack((p2_batch_index, rois_p2))[np.newaxis, :]
+
+    data = {
+        'data': im_array,
+        'rois_stride32': rois_p5_array,
+        'rois_stride16': rois_p4_array,
+        'rois_stride8': rois_p3_array,
+        'rois_stride4': rois_p2_array
+    }
+    label = {}
+
+    return data, label, im_info
+
+
+def get_fpn_maskrcnn_batch(roidb):
+    """
+    return a dictionary that contains raw data.
+    """
+    num_images = len(roidb)
+    imgs, roidb = get_image(roidb, scale=config.TRAIN.SCALE)  #TODO
+    #imgs, roidb = get_image(roidb)
+    im_array = tensor_vstack(imgs)
+
+    assert config.TRAIN.BATCH_ROIS % config.TRAIN.BATCH_IMAGES == 0, \
+        'BATCHIMAGES {} must divide BATCH_ROIS {}'.format(config.TRAIN.BATCH_IMAGES, config.TRAIN.BATCH_ROIS)
+    rois_per_image = config.TRAIN.BATCH_ROIS / config.TRAIN.BATCH_IMAGES
+    fg_rois_per_image = np.round(config.TRAIN.FG_FRACTION *
+                                 rois_per_image).astype(int)
+
+    rois_on_imgs = dict()
+    labels_on_imgs = dict()
+    bbox_targets_on_imgs = dict()
+    bbox_weights_on_imgs = dict()
+    mask_targets_on_imgs = dict()
+    mask_weights_on_imgs = dict()
+    for s in config.RCNN_FEAT_STRIDE:
+        rois_on_imgs.update({'stride%s' % s: list()})
+        labels_on_imgs.update({'stride%s' % s: list()})
+        bbox_targets_on_imgs.update({'stride%s' % s: list()})
+        bbox_weights_on_imgs.update({'stride%s' % s: list()})
+        mask_targets_on_imgs.update({'stride%s' % s: list()})
+        mask_weights_on_imgs.update({'stride%s' % s: list()})
+
+    # Sample rois
+    level_related_data_on_imgs = {}
+    for im_i in range(num_images):
+        roi_rec = roidb[im_i]
+        # infer num_classes from gt_overlaps
+        num_classes = roi_rec['gt_overlaps'].shape[1]
+        # label = class RoI has max overlap with
+        rois = roi_rec['boxes']
+        labels = roi_rec['max_classes']
+        overlaps = roi_rec['max_overlaps']
+        bbox_targets = roi_rec['bbox_targets']
+        im_info = roi_rec['im_info']
+
+        mask_targets = roi_rec['mask_targets']
+        mask_labels = roi_rec['mask_labels']
+        mask_inds = roi_rec['mask_inds']
+
+        assign_levels = roi_rec['assign_levels']
+
+        im_rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels, mask_targets_on_levels, mask_weights_on_levels = \
+            sample_rois_fpn(rois, assign_levels, fg_rois_per_image, rois_per_image, num_classes,
+                            labels, overlaps, bbox_targets, mask_targets=mask_targets, mask_labels=mask_labels, mask_inds=mask_inds, im_info=im_info)
+
+        level_related_data_on_imgs.update({
+            'img_%s' % im_i: {
+                'rois_on_levels': im_rois_on_levels,
+                'labels_on_levels': labels_on_levels,
+                'bbox_targets_on_levels': bbox_targets_on_levels,
+                'bbox_weights_on_levels': bbox_weights_on_levels,
+                'mask_targets_on_levels': mask_targets_on_levels,
+                'mask_weights_on_levels': mask_weights_on_levels,
+            }
+        })
+
+    return im_array, level_related_data_on_imgs
+
+
+def sample_rois(rois,
+                fg_rois_per_image,
+                rois_per_image,
+                num_classes,
+                labels=None,
+                overlaps=None,
+                bbox_targets=None,
+                gt_boxes=None,
+                mask_targets=None,
+                mask_labels=None,
+                mask_inds=None):
+    """
+    generate random sample of ROIs comprising foreground and background examples
+    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
+    :param fg_rois_per_image: foreground roi number
+    :param rois_per_image: total roi number
+    :param num_classes: number of classes
+    :param labels: maybe precomputed
+    :param overlaps: maybe precomputed (max_overlaps)
+    :param bbox_targets: maybe precomputed
+    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
+    :return: (rois, labels, bbox_targets, bbox_weights)
+    """
+    if labels is None:
+        if len(gt_boxes) == 0:
+            gt_boxes = np.zeros((1, 5))
+            gt_assignment = np.zeros((len(rois), ), dtype=np.int32)
+            overlaps = np.zeros((len(rois), ))
+            labels = np.zeros((len(rois), ))
+        else:
+            overlaps = bbox_overlaps(rois[:, 1:].astype(np.float),
+                                     gt_boxes[:, :4].astype(np.float))
+            gt_assignment = overlaps.argmax(axis=1)
+            overlaps = overlaps.max(axis=1)
+            labels = gt_boxes[gt_assignment, 4]
+
+    num_rois = rois.shape[0]
+    # foreground RoI with FG_THRESH overlap
+    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
+    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
+    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
+    # Sample foreground regions without replacement
+    if len(fg_indexes) > fg_rois_per_this_image:
+        fg_indexes = npr.choice(fg_indexes,
+                                size=fg_rois_per_this_image,
+                                replace=False)
+
+    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
+    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI)
+                          & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
+    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
+    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
+    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image,
+                                        bg_indexes.size)
+    # Sample foreground regions without replacement
+    if len(bg_indexes) > bg_rois_per_this_image:
+        bg_indexes = npr.choice(bg_indexes,
+                                size=bg_rois_per_this_image,
+                                replace=False)
+
+    # indexes selected
+    keep_indexes = np.append(fg_indexes, bg_indexes)
+
+    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
+    neg_rois = rois[neg_idx]
+
+    # pad more to ensure a fixed minibatch size
+    while keep_indexes.shape[0] < rois_per_image:
+        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
+        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
+        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
+
+    # select labels
+    labels = labels[keep_indexes]
+    # set labels of bg_rois to be 0
+    labels[fg_rois_per_this_image:] = 0
+    rois = rois[keep_indexes]
+    if mask_targets is not None:
+        assert mask_labels is not None
+        assert mask_inds is not None
+
+        def _mask_umap(mask_targets, mask_labels, mask_inds):
+            _mask_targets = np.zeros((num_rois, num_classes, 28, 28),
+                                     dtype=np.int8)
+            _mask_weights = np.zeros((num_rois, num_classes, 28, 28),
+                                     dtype=np.int8)
+            _mask_targets[mask_inds, mask_labels] = mask_targets
+            _mask_weights[mask_inds, mask_labels] = 1
+            _mask_weights[:, 0] = 0  # set background mask weight to zeros
+            return _mask_targets, _mask_weights  # [num_rois, num_classes, 28, 28]
+
+        mask_targets, mask_weights = _mask_umap(mask_targets, mask_labels,
+                                                mask_inds)
+        mask_targets = mask_targets[keep_indexes]
+        mask_weights = mask_weights[keep_indexes]
+
+    # load or compute bbox_target
+    if bbox_targets is not None:
+        bbox_target_data = bbox_targets[keep_indexes, :]
+    else:
+        targets = bbox_transform(rois[:, 1:],
+                                 gt_boxes[gt_assignment[keep_indexes], :4])
+        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS)) /
+                       np.array(config.TRAIN.BBOX_STDS))
+        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
+
+    bbox_targets, bbox_weights = \
+        expand_bbox_regression_targets(bbox_target_data, num_classes)
+
+    if mask_targets is not None:
+        return rois, labels, bbox_targets, bbox_weights, mask_targets, mask_weights
+    else:
+        return rois, labels, bbox_targets, bbox_weights
+
+
+def sample_rois_fpn(rois,
+                    assign_levels,
+                    fg_rois_per_image,
+                    rois_per_image,
+                    num_classes,
+                    labels=None,
+                    overlaps=None,
+                    bbox_targets=None,
+                    mask_targets=None,
+                    mask_labels=None,
+                    mask_inds=None,
+                    gt_boxes=None,
+                    im_info=None):
+    """
+    generate random sample of ROIs comprising foreground and background examples
+    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
+    :param assign_levels: [n]
+    :param fg_rois_per_image: foreground roi number
+    :param rois_per_image: total roi number
+    :param num_classes: number of classes
+    :param labels: maybe precomputed
+    :param overlaps: maybe precomputed (max_overlaps)
+    :param bbox_targets: maybe precomputed
+    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
+    :return: (rois, labels, bbox_targets, bbox_weights)
+    """
+    DEBUG = False
+    if labels is None:
+        if len(gt_boxes) == 0:
+            gt_boxes = np.zeros((1, 5))
+            gt_assignment = np.zeros((len(rois), ), dtype=np.int32)
+            overlaps = np.zeros((len(rois), ))
+            labels = np.zeros((len(rois), ))
+        else:
+            overlaps = bbox_overlaps(rois[:, 1:].astype(np.float),
+                                     gt_boxes[:, :4].astype(np.float))
+            gt_assignment = overlaps.argmax(axis=1)
+            overlaps = overlaps.max(axis=1)
+            labels = gt_boxes[gt_assignment, 4]
+
+    num_rois = rois.shape[0]
+    # foreground RoI with FG_THRESH overlap
+    fg_indexes = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
+    # guard against the case when an image has fewer than fg_rois_per_image foreground RoIs
+    fg_rois_per_this_image = np.minimum(fg_rois_per_image, fg_indexes.size)
+
+    if DEBUG:
+        print 'fg total num:', len(fg_indexes)
+
+    # Sample foreground regions without replacement
+    if len(fg_indexes) > fg_rois_per_this_image:
+        fg_indexes = npr.choice(fg_indexes,
+                                size=fg_rois_per_this_image,
+                                replace=False)
+
+    # Select background RoIs as those within [BG_THRESH_LO, BG_THRESH_HI)
+    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI)
+                          & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
+    if DEBUG:
+        print 'bg total num:', len(bg_indexes)
+    # Compute number of background RoIs to take from this image (guarding against there being fewer than desired)
+    bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
+    bg_rois_per_this_image = np.minimum(bg_rois_per_this_image,
+                                        bg_indexes.size)
+    # Sample foreground regions without replacement
+    if len(bg_indexes) > bg_rois_per_this_image:
+        bg_indexes = npr.choice(bg_indexes,
+                                size=bg_rois_per_this_image,
+                                replace=False)
+    if DEBUG:
+        print 'fg num:', len(fg_indexes)
+        print 'bg num:', len(bg_indexes)
+
+    # bg rois statistics
+    if DEBUG:
+        bg_assign = assign_levels[bg_indexes]
+        bg_rois_on_levels = dict()
+        for i, s in enumerate(config.RCNN_FEAT_STRIDE):
+            bg_rois_on_levels.update(
+                {'stride%s' % s: len(np.where(bg_assign == s)[0])})
+        print bg_rois_on_levels
+
+    # indexes selected
+    keep_indexes = np.append(fg_indexes, bg_indexes)
+
+    neg_idx = np.where(overlaps < config.TRAIN.FG_THRESH)[0]
+    neg_rois = rois[neg_idx]
+
+    # pad more to ensure a fixed minibatch size
+    while keep_indexes.shape[0] < rois_per_image:
+        gap = np.minimum(len(neg_rois), rois_per_image - keep_indexes.shape[0])
+        gap_indexes = npr.choice(range(len(neg_rois)), size=gap, replace=False)
+        keep_indexes = np.append(keep_indexes, neg_idx[gap_indexes])
+
+    # select labels
+    labels = labels[keep_indexes]
+    # set labels of bg_rois to be 0
+    labels[fg_rois_per_this_image:] = 0
+    rois = rois[keep_indexes]
+    assign_levels = assign_levels[keep_indexes]
+
+    if mask_targets is not None:
+        assert mask_labels is not None
+        assert mask_inds is not None
+
+        def _mask_umap(mask_targets, mask_labels, mask_inds):
+            _mask_targets = np.zeros((num_rois, num_classes, 28, 28),
+                                     dtype=np.int8)
+            _mask_weights = np.zeros((num_rois, num_classes, 1, 1),
+                                     dtype=np.int8)
+            _mask_targets[mask_inds, mask_labels] = mask_targets
+            _mask_weights[mask_inds, mask_labels] = 1
+            return _mask_targets, _mask_weights  # [num_rois, num_classes, 28, 28]
+
+        mask_targets, mask_weights = _mask_umap(mask_targets, mask_labels,
+                                                mask_inds)
+        mask_targets = mask_targets[keep_indexes]
+        mask_weights = mask_weights[keep_indexes]
+
+    # load or compute bbox_target
+    if bbox_targets is not None:
+        bbox_target_data = bbox_targets[keep_indexes, :]
+    else:
+        targets = bbox_transform(rois[:, 1:],
+                                 gt_boxes[gt_assignment[keep_indexes], :4])
+        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS)) /
+                       np.array(config.TRAIN.BBOX_STDS))
+        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
+
+    bbox_targets, bbox_weights = \
+        expand_bbox_regression_targets(bbox_target_data, num_classes)
+
+    # Assign to levels
+    rois_on_levels = dict()
+    labels_on_levels = dict()
+    bbox_targets_on_levels = dict()
+    bbox_weights_on_levels = dict()
+    if mask_targets is not None:
+        mask_targets_on_levels = dict()
+        mask_weights_on_levels = dict()
+    for i, s in enumerate(config.RCNN_FEAT_STRIDE):
+        index = np.where(assign_levels == s)
+        _rois = rois[index]
+        _labels = labels[index]
+        _bbox_targets = bbox_targets[index]
+        _bbox_weights = bbox_weights[index]
+        if mask_targets is not None:
+            _mask_targets = mask_targets[index]
+            _mask_weights = mask_weights[index]
+
+        rois_on_levels.update({'stride%s' % s: _rois})
+        labels_on_levels.update({'stride%s' % s: _labels})
+        bbox_targets_on_levels.update({'stride%s' % s: _bbox_targets})
+        bbox_weights_on_levels.update({'stride%s' % s: _bbox_weights})
+        if mask_targets is not None:
+            mask_targets_on_levels.update({'stride%s' % s: _mask_targets})
+            mask_weights_on_levels.update({'stride%s' % s: _mask_weights})
+
+    if mask_targets is not None:
+        return rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels, mask_targets_on_levels, mask_weights_on_levels
+    else:
+        return rois_on_levels, labels_on_levels, bbox_targets_on_levels, bbox_weights_on_levels
+
+
+def get_rois(rois,
+             rois_per_image,
+             num_classes,
+             labels=None,
+             overlaps=None,
+             bbox_targets=None,
+             gt_boxes=None):
+    """
+    get top N ROIs, used in online hard example mining
+    :param rois: all_rois [n, 4]; e2e: [n, 5] with batch_index
+    :param rois_per_image: total roi number
+    :param num_classes: number of classes
+    :param labels: maybe precomputed
+    :param overlaps: maybe precomputed (max_overlaps)
+    :param bbox_targets: maybe precomputed
+    :param gt_boxes: optional for e2e [n, 5] (x1, y1, x2, y2, cls)
+    :return: (rois, labels, bbox_targets, bbox_weights)
+    """
+    if labels is None:
+        if len(gt_boxes) == 0:
+            gt_boxes = np.array([[1, 1, 1, 1, 0]])
+        overlaps = bbox_overlaps(rois[:, 1:].astype(np.float),
+                                 gt_boxes[:, :4].astype(np.float))
+        gt_assignment = overlaps.argmax(axis=1)
+        overlaps = overlaps.max(axis=1)
+        labels = gt_boxes[gt_assignment, 4]
+
+    # select indices
+    keep_indexes = np.arange(rois.shape[0])
+    if keep_indexes.shape[0] > rois_per_image:
+        keep_indexes = npr.choice(keep_indexes,
+                                  size=rois_per_image,
+                                  replace=False)
+
+    # if not enough, pad until rois_per_image is satisfied
+    while keep_indexes.shape[0] < rois_per_image:
+        gap = np.minimum(rois_per_image - keep_indexes.shape[0], len(rois))
+        gap_indexes = npr.choice(range(len(rois)), size=gap, replace=False)
+        keep_indexes = np.append(keep_indexes, gap_indexes)
+
+    # suppress any bg defined by overlap
+    bg_indexes = np.where((overlaps < config.TRAIN.BG_THRESH_HI)
+                          & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
+    labels[bg_indexes] = 0
+
+    labels = labels[keep_indexes]
+    rois = rois[keep_indexes]
+
+    # load or compute bbox_target
+    if bbox_targets is not None:
+        bbox_target_data = bbox_targets[keep_indexes, :]
+    else:
+        targets = bbox_transform(rois[:, 1:],
+                                 gt_boxes[gt_assignment[keep_indexes], :4])
+        if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+            targets = ((targets - np.array(config.TRAIN.BBOX_MEANS)) /
+                       np.array(config.TRAIN.BBOX_STDS))
+        bbox_target_data = np.hstack((labels[:, np.newaxis], targets))
+
+    bbox_targets, bbox_weights = \
+        expand_bbox_regression_targets(bbox_target_data, num_classes)
+
+    return rois, labels, bbox_targets, bbox_weights
diff --git a/detection/retinaface/rcnn/io/rpn.py b/detection/retinaface/rcnn/io/rpn.py
new file mode 100644
index 000000000..998fad8c8
--- /dev/null
+++ b/detection/retinaface/rcnn/io/rpn.py
@@ -0,0 +1,887 @@
+"""
+RPN:
+data =
+    {'data': [num_images, c, h, w],
+     'im_info': [num_images, 4] (optional)}
+label =
+    {'gt_boxes': [num_boxes, 5] (optional),
+     'label': [batch_size, 1] <- [batch_size, num_anchors, feat_height, feat_width],
+     'bbox_target': [batch_size, num_anchors, feat_height, feat_width],
+     'bbox_weight': [batch_size, num_anchors, feat_height, feat_width]}
+"""
+
+from __future__ import print_function
+import sys
+import logging
+import datetime
+import numpy as np
+import numpy.random as npr
+
+from ..logger import logger
+from ..config import config
+from .image import get_image, tensor_vstack, get_crop_image
+from ..processing.generate_anchor import generate_anchors, anchors_plane
+from ..processing.bbox_transform import bbox_overlaps, bbox_transform, landmark_transform
+
+STAT = {0: 0, 8: 0, 16: 0, 32: 0}
+
+
+def get_rpn_testbatch(roidb):
+    """
+    return a dict of testbatch
+    :param roidb: ['image', 'flipped']
+    :return: data, label, im_info
+    """
+    assert len(roidb) == 1, 'Single batch only'
+    imgs, roidb = get_image(roidb)
+    im_array = imgs[0]
+    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
+
+    data = {'data': im_array, 'im_info': im_info}
+    label = {}
+
+    return data, label, im_info
+
+
+def get_rpn_batch(roidb):
+    """
+    prototype for rpn batch: data, im_info, gt_boxes
+    :param roidb: ['image', 'flipped'] + ['gt_boxes', 'boxes', 'gt_classes']
+    :return: data, label
+    """
+    assert len(roidb) == 1, 'Single batch only'
+    imgs, roidb = get_image(roidb)
+    im_array = imgs[0]
+    im_info = np.array([roidb[0]['im_info']], dtype=np.float32)
+
+    # gt boxes: (x1, y1, x2, y2, cls)
+    if roidb[0]['gt_classes'].size > 0:
+        gt_inds = np.where(roidb[0]['gt_classes'] != 0)[0]
+        gt_boxes = np.empty((roidb[0]['boxes'].shape[0], 5), dtype=np.float32)
+        gt_boxes[:, 0:4] = roidb[0]['boxes'][gt_inds, :]
+        gt_boxes[:, 4] = roidb[0]['gt_classes'][gt_inds]
+    else:
+        gt_boxes = np.empty((0, 5), dtype=np.float32)
+
+    data = {'data': im_array, 'im_info': im_info}
+    label = {'gt_boxes': gt_boxes}
+
+    return data, label
+
+
+def get_crop_batch(roidb):
+    """
+    prototype for rpn batch: data, im_info, gt_boxes
+    :param roidb: ['image', 'flipped'] + ['gt_boxes', 'boxes', 'gt_classes']
+    :return: data, label
+    """
+    #assert len(roidb) == 1, 'Single batch only'
+    data_list = []
+    label_list = []
+    imgs, roidb = get_crop_image(roidb)
+    assert len(imgs) == len(roidb)
+    for i in range(len(imgs)):
+        im_array = imgs[i]
+        im_info = np.array([roidb[i]['im_info']], dtype=np.float32)
+
+        # gt boxes: (x1, y1, x2, y2, cls)
+        if roidb[i]['gt_classes'].size > 0:
+            gt_inds = np.where(roidb[i]['gt_classes'] != 0)[0]
+            gt_boxes = np.empty((roidb[i]['boxes'].shape[0], 5),
+                                dtype=np.float32)
+            gt_boxes[:, 0:4] = roidb[i]['boxes'][gt_inds, :]
+            gt_boxes[:, 4] = roidb[i]['gt_classes'][gt_inds]
+            if config.USE_BLUR:
+                gt_blur = roidb[i]['blur']
+            if config.FACE_LANDMARK:
+                #gt_landmarks = np.empty((roidb[i]['landmarks'].shape[0], 11), dtype=np.float32)
+                gt_landmarks = roidb[i]['landmarks'][gt_inds, :, :]
+            if config.HEAD_BOX:
+                gt_boxes_head = np.empty((roidb[i]['boxes_head'].shape[0], 5),
+                                         dtype=np.float32)
+                gt_boxes_head[:, 0:4] = roidb[i]['boxes_head'][gt_inds, :]
+                gt_boxes_head[:, 4] = roidb[i]['gt_classes'][gt_inds]
+        else:
+            gt_boxes = np.empty((0, 5), dtype=np.float32)
+            if config.USE_BLUR:
+                gt_blur = np.empty((0, ), dtype=np.float32)
+            if config.FACE_LANDMARK:
+                gt_landmarks = np.empty((0, 5, 3), dtype=np.float32)
+            if config.HEAD_BOX:
+                gt_boxes_head = np.empty((0, 5), dtype=np.float32)
+
+        data = {'data': im_array, 'im_info': im_info}
+        label = {'gt_boxes': gt_boxes}
+        if config.USE_BLUR:
+            label['gt_blur'] = gt_blur
+        if config.FACE_LANDMARK:
+            label['gt_landmarks'] = gt_landmarks
+        if config.HEAD_BOX:
+            label['gt_boxes_head'] = gt_boxes_head
+        data_list.append(data)
+        label_list.append(label)
+
+    return data_list, label_list
+
+
+def assign_anchor_fpn(feat_shape,
+                      gt_label,
+                      im_info,
+                      landmark=False,
+                      prefix='face',
+                      select_stride=0):
+    """
+    assign ground truth boxes to anchor positions
+    :param feat_shape: infer output shape
+    :param gt_boxes: assign ground truth
+    :param im_info: filter out anchors overlapped with edges
+    :return: tuple
+    labels: of shape (batch_size, 1) <- (batch_size, num_anchors, feat_height, feat_width)
+    bbox_targets: of shape (batch_size, num_anchors * 4, feat_height, feat_width)
+    bbox_weights: mark the assigned anchors
+    """
+    def _unmap(data, count, inds, fill=0):
+        """" unmap a subset inds of data into original data of size count """
+        if len(data.shape) == 1:
+            ret = np.empty((count, ), dtype=np.float32)
+            ret.fill(fill)
+            ret[inds] = data
+        else:
+            ret = np.empty((count, ) + data.shape[1:], dtype=np.float32)
+            ret.fill(fill)
+            ret[inds, :] = data
+        return ret
+
+    global STAT
+    DEBUG = False
+
+    im_info = im_info[0]
+    gt_boxes = gt_label['gt_boxes']
+    # clean up boxes
+    nonneg = np.where(gt_boxes[:, 4] != -1)[0]
+    gt_boxes = gt_boxes[nonneg]
+    if config.USE_BLUR:
+        gt_blur = gt_label['gt_blur']
+        gt_blur = gt_blur[nonneg]
+    if landmark:
+        gt_landmarks = gt_label['gt_landmarks']
+        gt_landmarks = gt_landmarks[nonneg]
+        assert gt_boxes.shape[0] == gt_landmarks.shape[0]
+    #scales = np.array(scales, dtype=np.float32)
+    feat_strides = config.RPN_FEAT_STRIDE
+    bbox_pred_len = 4
+    landmark_pred_len = 10
+    if config.USE_BLUR:
+        gt_boxes[:, 4] = gt_blur
+        bbox_pred_len = 5
+    if config.USE_OCCLUSION:
+        landmark_pred_len = 15
+
+    anchors_list = []
+    anchors_num_list = []
+    inds_inside_list = []
+    feat_infos = []
+    A_list = []
+    for i in range(len(feat_strides)):
+        stride = feat_strides[i]
+        sstride = str(stride)
+        base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
+        allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
+        ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
+        scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
+        base_anchors = generate_anchors(base_size=base_size,
+                                        ratios=list(ratios),
+                                        scales=np.array(scales,
+                                                        dtype=np.float32),
+                                        stride=stride,
+                                        dense_anchor=config.DENSE_ANCHOR)
+        num_anchors = base_anchors.shape[0]
+        feat_height, feat_width = feat_shape[i][-2:]
+        feat_stride = feat_strides[i]
+        feat_infos.append([feat_height, feat_width])
+
+        A = num_anchors
+        A_list.append(A)
+        K = feat_height * feat_width
+
+        all_anchors = anchors_plane(feat_height, feat_width, feat_stride,
+                                    base_anchors)
+        all_anchors = all_anchors.reshape((K * A, 4))
+        #print('anchor0', stride, all_anchors[0])
+
+        total_anchors = int(K * A)
+        anchors_num_list.append(total_anchors)
+        # only keep anchors inside the image
+        inds_inside = np.where(
+            (all_anchors[:, 0] >= -allowed_border)
+            & (all_anchors[:, 1] >= -allowed_border)
+            & (all_anchors[:, 2] < im_info[1] + allowed_border)
+            & (all_anchors[:, 3] < im_info[0] + allowed_border))[0]
+        if DEBUG:
+            print('total_anchors', total_anchors)
+            print('inds_inside', len(inds_inside))
+
+        # keep only inside anchors
+        anchors = all_anchors[inds_inside, :]
+        #print('AA', anchors.shape, len(inds_inside))
+
+        anchors_list.append(anchors)
+        inds_inside_list.append(inds_inside)
+
+    # Concat anchors from each level
+    anchors = np.concatenate(anchors_list)
+    for i in range(1, len(inds_inside_list)):
+        inds_inside_list[i] = inds_inside_list[i] + sum(anchors_num_list[:i])
+    inds_inside = np.concatenate(inds_inside_list)
+    total_anchors = sum(anchors_num_list)
+    #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
+
+    # label: 1 is positive, 0 is negative, -1 is dont care
+    labels = np.empty((len(inds_inside), ), dtype=np.float32)
+    labels.fill(-1)
+    #print('BB', anchors.shape, len(inds_inside))
+    #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
+
+    if gt_boxes.size > 0:
+        # overlap between the anchors and the gt boxes
+        # overlaps (ex, gt)
+        overlaps = bbox_overlaps(anchors.astype(np.float),
+                                 gt_boxes.astype(np.float))
+        argmax_overlaps = overlaps.argmax(axis=1)
+        #print('AAA', argmax_overlaps.shape)
+        max_overlaps = overlaps[np.arange(len(inds_inside)), argmax_overlaps]
+        gt_argmax_overlaps = overlaps.argmax(axis=0)
+        gt_max_overlaps = overlaps[gt_argmax_overlaps,
+                                   np.arange(overlaps.shape[1])]
+        gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
+
+        if not config.TRAIN.RPN_CLOBBER_POSITIVES:
+            # assign bg labels first so that positive labels can clobber them
+            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
+
+        # fg label: for each gt, anchor with highest overlap
+        if config.TRAIN.RPN_FORCE_POSITIVE:
+            labels[gt_argmax_overlaps] = 1
+
+        # fg label: above threshold IoU
+        labels[max_overlaps >= config.TRAIN.RPN_POSITIVE_OVERLAP] = 1
+
+        if config.TRAIN.RPN_CLOBBER_POSITIVES:
+            # assign bg labels last so that negative labels can clobber positives
+            labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
+    else:
+        labels[:] = 0
+    fg_inds = np.where(labels == 1)[0]
+    #print('fg count', len(fg_inds))
+
+    # subsample positive labels if we have too many
+    if config.TRAIN.RPN_ENABLE_OHEM == 0:
+        fg_inds = np.where(labels == 1)[0]
+        num_fg = int(config.TRAIN.RPN_FG_FRACTION *
+                     config.TRAIN.RPN_BATCH_SIZE)
+        if len(fg_inds) > num_fg:
+            disable_inds = npr.choice(fg_inds,
+                                      size=(len(fg_inds) - num_fg),
+                                      replace=False)
+            if DEBUG:
+                disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
+            labels[disable_inds] = -1
+
+        # subsample negative labels if we have too many
+        num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
+        bg_inds = np.where(labels == 0)[0]
+        if len(bg_inds) > num_bg:
+            disable_inds = npr.choice(bg_inds,
+                                      size=(len(bg_inds) - num_bg),
+                                      replace=False)
+            if DEBUG:
+                disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+            labels[disable_inds] = -1
+
+        #fg_inds = np.where(labels == 1)[0]
+        #num_fg = len(fg_inds)
+        #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
+
+        #bg_inds = np.where(labels == 0)[0]
+        #if len(bg_inds) > num_bg:
+        #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
+        #    if DEBUG:
+        #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+        #    labels[disable_inds] = -1
+    else:
+        fg_inds = np.where(labels == 1)[0]
+        num_fg = len(fg_inds)
+        bg_inds = np.where(labels == 0)[0]
+        num_bg = len(bg_inds)
+
+    #print('anchor stat', num_fg, num_bg)
+
+    bbox_targets = np.zeros((len(inds_inside), bbox_pred_len),
+                            dtype=np.float32)
+    if gt_boxes.size > 0:
+        #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
+        bbox_targets[:, :] = bbox_transform(anchors,
+                                            gt_boxes[argmax_overlaps, :])
+        #bbox_targets[:,4] = gt_blur
+
+    bbox_weights = np.zeros((len(inds_inside), bbox_pred_len),
+                            dtype=np.float32)
+    #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
+    bbox_weights[labels == 1, 0:4] = 1.0
+    if bbox_pred_len > 4:
+        bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
+
+    if landmark:
+        landmark_targets = np.zeros((len(inds_inside), landmark_pred_len),
+                                    dtype=np.float32)
+        #landmark_weights = np.zeros((len(inds_inside), 10), dtype=np.float32)
+        landmark_weights = np.zeros((len(inds_inside), landmark_pred_len),
+                                    dtype=np.float32)
+        #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
+        if landmark_pred_len == 10:
+            landmark_weights[labels == 1, :] = 1.0
+        elif landmark_pred_len == 15:
+            v = [1.0, 1.0, 0.1] * 5
+            assert len(v) == 15
+            landmark_weights[labels == 1, :] = np.array(v)
+        else:
+            assert False
+        #TODO here
+        if gt_landmarks.size > 0:
+            #print('AAA',argmax_overlaps)
+            a_landmarks = gt_landmarks[argmax_overlaps, :, :]
+            landmark_targets[:] = landmark_transform(anchors, a_landmarks)
+            invalid = np.where(a_landmarks[:, 0, 2] < 0.0)[0]
+            #assert len(invalid)==0
+            #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
+            landmark_weights[invalid, :] = 0.0
+
+    #if DEBUG:
+    #    _sums = bbox_targets[labels == 1, :].sum(axis=0)
+    #    _squared_sums = (bbox_targets[labels == 1, :] ** 2).sum(axis=0)
+    #    _counts = np.sum(labels == 1)
+    #    means = _sums / (_counts + 1e-14)
+    #    stds = np.sqrt(_squared_sums / _counts - means ** 2)
+    #    print 'means', means
+    #    print 'stdevs', stds
+    # map up to original set of anchors
+    #print(labels.shape, total_anchors, inds_inside.shape, inds_inside[0], inds_inside[-1])
+    labels = _unmap(labels, total_anchors, inds_inside, fill=-1)
+    bbox_targets = _unmap(bbox_targets, total_anchors, inds_inside, fill=0)
+    bbox_weights = _unmap(bbox_weights, total_anchors, inds_inside, fill=0)
+    if landmark:
+        landmark_targets = _unmap(landmark_targets,
+                                  total_anchors,
+                                  inds_inside,
+                                  fill=0)
+        landmark_weights = _unmap(landmark_weights,
+                                  total_anchors,
+                                  inds_inside,
+                                  fill=0)
+    #print('CC', anchors.shape, len(inds_inside))
+
+    #if DEBUG:
+    #    if gt_boxes.size > 0:
+    #        print 'rpn: max max_overlaps', np.max(max_overlaps)
+    #    print 'rpn: num_positives', np.sum(labels == 1)
+    #    print 'rpn: num_negatives', np.sum(labels == 0)
+    #    _fg_sum = np.sum(labels == 1)
+    #    _bg_sum = np.sum(labels == 0)
+    #    _count = 1
+    #    print 'rpn: num_positive avg', _fg_sum / _count
+    #    print 'rpn: num_negative avg', _bg_sum / _count
+
+    # resahpe
+    label_list = list()
+    bbox_target_list = list()
+    bbox_weight_list = list()
+    if landmark:
+        landmark_target_list = list()
+        landmark_weight_list = list()
+    anchors_num_range = [0] + anchors_num_list
+    label = {}
+    for i in range(len(feat_strides)):
+        stride = feat_strides[i]
+        feat_height, feat_width = feat_infos[i]
+        A = A_list[i]
+        _label = labels[sum(anchors_num_range[:i +
+                                              1]):sum(anchors_num_range[:i +
+                                                                        1]) +
+                        anchors_num_range[i + 1]]
+        if select_stride > 0 and stride != select_stride:
+            #print('set', stride, select_stride)
+            _label[:] = -1
+        #print('_label', _label.shape, select_stride)
+        #_fg_inds = np.where(_label == 1)[0]
+        #n_fg = len(_fg_inds)
+        #STAT[0]+=1
+        #STAT[stride]+=n_fg
+        #if STAT[0]%100==0:
+        #  print('rpn_stat', STAT, file=sys.stderr)
+        bbox_target = bbox_targets[sum(anchors_num_range[:i + 1]
+                                       ):sum(anchors_num_range[:i + 1]) +
+                                   anchors_num_range[i + 1]]
+        bbox_weight = bbox_weights[sum(anchors_num_range[:i + 1]
+                                       ):sum(anchors_num_range[:i + 1]) +
+                                   anchors_num_range[i + 1]]
+        if landmark:
+            landmark_target = landmark_targets[
+                sum(anchors_num_range[:i + 1]):sum(anchors_num_range[:i + 1]) +
+                anchors_num_range[i + 1]]
+            landmark_weight = landmark_weights[
+                sum(anchors_num_range[:i + 1]):sum(anchors_num_range[:i + 1]) +
+                anchors_num_range[i + 1]]
+
+        _label = _label.reshape(
+            (1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
+        _label = _label.reshape((1, A * feat_height * feat_width))
+        bbox_target = bbox_target.reshape(
+            (1, feat_height * feat_width,
+             A * bbox_pred_len)).transpose(0, 2, 1)
+        bbox_weight = bbox_weight.reshape(
+            (1, feat_height * feat_width, A * bbox_pred_len)).transpose(
+                (0, 2, 1))
+        label['%s_label_stride%d' % (prefix, stride)] = _label
+        label['%s_bbox_target_stride%d' % (prefix, stride)] = bbox_target
+        label['%s_bbox_weight_stride%d' % (prefix, stride)] = bbox_weight
+        if landmark:
+            landmark_target = landmark_target.reshape(
+                (1, feat_height * feat_width,
+                 A * landmark_pred_len)).transpose(0, 2, 1)
+            landmark_weight = landmark_weight.reshape(
+                (1, feat_height * feat_width,
+                 A * landmark_pred_len)).transpose((0, 2, 1))
+            label['%s_landmark_target_stride%d' %
+                  (prefix, stride)] = landmark_target
+            label['%s_landmark_weight_stride%d' %
+                  (prefix, stride)] = landmark_weight
+        #print('in_rpn', stride,_label.shape, bbox_target.shape, bbox_weight.shape, file=sys.stderr)
+        label_list.append(_label)
+        #print('DD', _label.shape)
+        bbox_target_list.append(bbox_target)
+        bbox_weight_list.append(bbox_weight)
+        if landmark:
+            landmark_target_list.append(landmark_target)
+            landmark_weight_list.append(landmark_weight)
+
+    label_concat = np.concatenate(label_list, axis=1)
+    bbox_target_concat = np.concatenate(bbox_target_list, axis=2)
+    bbox_weight_concat = np.concatenate(bbox_weight_list, axis=2)
+    #fg_inds = np.where(label_concat[0] == 1)[0]
+    #print('fg_inds_in_rpn2', fg_inds, file=sys.stderr)
+
+    label.update({
+        '%s_label' % prefix: label_concat,
+        '%s_bbox_target' % prefix: bbox_target_concat,
+        '%s_bbox_weight' % prefix: bbox_weight_concat
+    })
+    if landmark:
+        landmark_target_concat = np.concatenate(landmark_target_list, axis=2)
+        landmark_weight_concat = np.concatenate(landmark_weight_list, axis=2)
+        label['%s_landmark_target' % prefix] = landmark_target_concat
+        label['%s_landmark_weight' % prefix] = landmark_weight_concat
+    return label
+
+
+class AA:
+    def __init__(self, feat_shape):
+        self.feat_shape = feat_shape
+        feat_strides = config.RPN_FEAT_STRIDE
+        anchors_list = []
+        anchors_num_list = []
+        inds_inside_list = []
+        feat_infos = []
+        A_list = []
+        DEBUG = False
+        for i in range(len(feat_strides)):
+            stride = feat_strides[i]
+            sstride = str(stride)
+            base_size = config.RPN_ANCHOR_CFG[sstride]['BASE_SIZE']
+            allowed_border = config.RPN_ANCHOR_CFG[sstride]['ALLOWED_BORDER']
+            ratios = config.RPN_ANCHOR_CFG[sstride]['RATIOS']
+            scales = config.RPN_ANCHOR_CFG[sstride]['SCALES']
+            base_anchors = generate_anchors(base_size=base_size,
+                                            ratios=list(ratios),
+                                            scales=np.array(scales,
+                                                            dtype=np.float32),
+                                            stride=stride,
+                                            dense_anchor=config.DENSE_ANCHOR)
+            num_anchors = base_anchors.shape[0]
+            feat_height, feat_width = feat_shape[i][-2:]
+            feat_stride = feat_strides[i]
+            feat_infos.append([feat_height, feat_width])
+
+            A = num_anchors
+            A_list.append(A)
+            K = feat_height * feat_width
+
+            all_anchors = anchors_plane(feat_height, feat_width, feat_stride,
+                                        base_anchors)
+            all_anchors = all_anchors.reshape((K * A, 4))
+            #print('anchor0', stride, all_anchors[0])
+
+            total_anchors = int(K * A)
+            anchors_num_list.append(total_anchors)
+            # only keep anchors inside the image
+            inds_inside = np.where(
+                (all_anchors[:, 0] >= -allowed_border)
+                & (all_anchors[:, 1] >= -allowed_border)
+                & (all_anchors[:, 2] < config.SCALES[0][1] + allowed_border) &
+                (all_anchors[:, 3] < config.SCALES[0][1] + allowed_border))[0]
+            if DEBUG:
+                print('total_anchors', total_anchors)
+                print('inds_inside', len(inds_inside))
+
+            # keep only inside anchors
+            anchors = all_anchors[inds_inside, :]
+            #print('AA', anchors.shape, len(inds_inside))
+
+            anchors_list.append(anchors)
+            inds_inside_list.append(inds_inside)
+        anchors = np.concatenate(anchors_list)
+        for i in range(1, len(inds_inside_list)):
+            inds_inside_list[i] = inds_inside_list[i] + sum(
+                anchors_num_list[:i])
+        inds_inside = np.concatenate(inds_inside_list)
+        #self.anchors_list = anchors_list
+        #self.inds_inside_list = inds_inside_list
+        self.anchors = anchors
+        self.inds_inside = inds_inside
+        self.anchors_num_list = anchors_num_list
+        self.feat_infos = feat_infos
+        self.A_list = A_list
+        self._times = [0.0, 0.0, 0.0, 0.0]
+
+    @staticmethod
+    def _unmap(data, count, inds, fill=0):
+        """" unmap a subset inds of data into original data of size count """
+        if len(data.shape) == 1:
+            ret = np.empty((count, ), dtype=np.float32)
+            ret.fill(fill)
+            ret[inds] = data
+        else:
+            ret = np.empty((count, ) + data.shape[1:], dtype=np.float32)
+            ret.fill(fill)
+            ret[inds, :] = data
+        return ret
+
+    def assign_anchor_fpn(self,
+                          gt_label,
+                          im_info,
+                          landmark=False,
+                          prefix='face',
+                          select_stride=0):
+
+        #ta = datetime.datetime.now()
+        im_info = im_info[0]
+        gt_boxes = gt_label['gt_boxes']
+        # clean up boxes
+        nonneg = np.where(gt_boxes[:, 4] != -1)[0]
+        gt_boxes = gt_boxes[nonneg]
+        if config.USE_BLUR:
+            gt_blur = gt_label['gt_blur']
+            gt_blur = gt_blur[nonneg]
+        if landmark:
+            gt_landmarks = gt_label['gt_landmarks']
+            gt_landmarks = gt_landmarks[nonneg]
+            assert gt_boxes.shape[0] == gt_landmarks.shape[0]
+        #scales = np.array(scales, dtype=np.float32)
+        feat_strides = config.RPN_FEAT_STRIDE
+        bbox_pred_len = 4
+        landmark_pred_len = 10
+        if config.USE_BLUR:
+            gt_boxes[:, 4] = gt_blur
+            bbox_pred_len = 5
+        if config.USE_OCCLUSION:
+            landmark_pred_len = 15
+
+        #anchors_list = self.anchors_list
+        #inds_inside_list = self.inds_inside_list
+        anchors = self.anchors
+        inds_inside = self.inds_inside
+        anchors_num_list = self.anchors_num_list
+        feat_infos = self.feat_infos
+        A_list = self.A_list
+
+        total_anchors = sum(anchors_num_list)
+        #print('total_anchors', anchors.shape[0], len(inds_inside), file=sys.stderr)
+
+        # label: 1 is positive, 0 is negative, -1 is dont care
+        labels = np.empty((len(inds_inside), ), dtype=np.float32)
+        labels.fill(-1)
+        #print('BB', anchors.shape, len(inds_inside))
+        #print('gt_boxes', gt_boxes.shape, file=sys.stderr)
+        #tb = datetime.datetime.now()
+        #self._times[0] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+
+        if gt_boxes.size > 0:
+            # overlap between the anchors and the gt boxes
+            # overlaps (ex, gt)
+            overlaps = bbox_overlaps(anchors.astype(np.float),
+                                     gt_boxes.astype(np.float))
+            argmax_overlaps = overlaps.argmax(axis=1)
+            #print('AAA', argmax_overlaps.shape)
+            max_overlaps = overlaps[np.arange(len(inds_inside)),
+                                    argmax_overlaps]
+            gt_argmax_overlaps = overlaps.argmax(axis=0)
+            gt_max_overlaps = overlaps[gt_argmax_overlaps,
+                                       np.arange(overlaps.shape[1])]
+            gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0]
+
+            if not config.TRAIN.RPN_CLOBBER_POSITIVES:
+                # assign bg labels first so that positive labels can clobber them
+                labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
+
+            # fg label: for each gt, anchor with highest overlap
+            if config.TRAIN.RPN_FORCE_POSITIVE:
+                labels[gt_argmax_overlaps] = 1
+
+            # fg label: above threshold IoU
+            labels[max_overlaps >= config.TRAIN.RPN_POSITIVE_OVERLAP] = 1
+
+            if config.TRAIN.RPN_CLOBBER_POSITIVES:
+                # assign bg labels last so that negative labels can clobber positives
+                labels[max_overlaps < config.TRAIN.RPN_NEGATIVE_OVERLAP] = 0
+        else:
+            labels[:] = 0
+        fg_inds = np.where(labels == 1)[0]
+        #print('fg count', len(fg_inds))
+
+        # subsample positive labels if we have too many
+        if config.TRAIN.RPN_ENABLE_OHEM == 0:
+            fg_inds = np.where(labels == 1)[0]
+            num_fg = int(config.TRAIN.RPN_FG_FRACTION *
+                         config.TRAIN.RPN_BATCH_SIZE)
+            if len(fg_inds) > num_fg:
+                disable_inds = npr.choice(fg_inds,
+                                          size=(len(fg_inds) - num_fg),
+                                          replace=False)
+                if DEBUG:
+                    disable_inds = fg_inds[:(len(fg_inds) - num_fg)]
+                labels[disable_inds] = -1
+
+            # subsample negative labels if we have too many
+            num_bg = config.TRAIN.RPN_BATCH_SIZE - np.sum(labels == 1)
+            bg_inds = np.where(labels == 0)[0]
+            if len(bg_inds) > num_bg:
+                disable_inds = npr.choice(bg_inds,
+                                          size=(len(bg_inds) - num_bg),
+                                          replace=False)
+                if DEBUG:
+                    disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+                labels[disable_inds] = -1
+
+            #fg_inds = np.where(labels == 1)[0]
+            #num_fg = len(fg_inds)
+            #num_bg = num_fg*int(1.0/config.TRAIN.RPN_FG_FRACTION-1)
+
+            #bg_inds = np.where(labels == 0)[0]
+            #if len(bg_inds) > num_bg:
+            #    disable_inds = npr.choice(bg_inds, size=(len(bg_inds) - num_bg), replace=False)
+            #    if DEBUG:
+            #        disable_inds = bg_inds[:(len(bg_inds) - num_bg)]
+            #    labels[disable_inds] = -1
+        else:
+            fg_inds = np.where(labels == 1)[0]
+            num_fg = len(fg_inds)
+            bg_inds = np.where(labels == 0)[0]
+            num_bg = len(bg_inds)
+
+        #print('anchor stat', num_fg, num_bg)
+
+        bbox_targets = np.zeros((len(inds_inside), bbox_pred_len),
+                                dtype=np.float32)
+        if gt_boxes.size > 0:
+            #print('GT', gt_boxes.shape, gt_boxes[argmax_overlaps, :4].shape)
+            bbox_targets[:, :] = bbox_transform(anchors,
+                                                gt_boxes[argmax_overlaps, :])
+            #bbox_targets[:,4] = gt_blur
+        #tb = datetime.datetime.now()
+        #self._times[1] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+
+        bbox_weights = np.zeros((len(inds_inside), bbox_pred_len),
+                                dtype=np.float32)
+        #bbox_weights[labels == 1, :] = np.array(config.TRAIN.RPN_BBOX_WEIGHTS)
+        bbox_weights[labels == 1, 0:4] = 1.0
+        if bbox_pred_len > 4:
+            bbox_weights[labels == 1, 4:bbox_pred_len] = 0.1
+
+        if landmark:
+            landmark_targets = np.zeros((len(inds_inside), landmark_pred_len),
+                                        dtype=np.float32)
+            #landmark_weights = np.zeros((len(inds_inside), 10), dtype=np.float32)
+            landmark_weights = np.zeros((len(inds_inside), landmark_pred_len),
+                                        dtype=np.float32)
+            #landmark_weights[labels == 1, :] = np.array(config.TRAIN.RPN_LANDMARK_WEIGHTS)
+            if landmark_pred_len == 10:
+                landmark_weights[labels == 1, :] = 1.0
+            elif landmark_pred_len == 15:
+                v = [1.0, 1.0, 0.1] * 5
+                assert len(v) == 15
+                landmark_weights[labels == 1, :] = np.array(v)
+            else:
+                assert False
+            #TODO here
+            if gt_landmarks.size > 0:
+                #print('AAA',argmax_overlaps)
+                a_landmarks = gt_landmarks[argmax_overlaps, :, :]
+                landmark_targets[:] = landmark_transform(anchors, a_landmarks)
+                invalid = np.where(a_landmarks[:, 0, 2] < 0.0)[0]
+                #assert len(invalid)==0
+                #landmark_weights[invalid, :] = np.array(config.TRAIN.RPN_INVALID_LANDMARK_WEIGHTS)
+                landmark_weights[invalid, :] = 0.0
+        #tb = datetime.datetime.now()
+        #self._times[2] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+
+        #if DEBUG:
+        #    _sums = bbox_targets[labels == 1, :].sum(axis=0)
+        #    _squared_sums = (bbox_targets[labels == 1, :] ** 2).sum(axis=0)
+        #    _counts = np.sum(labels == 1)
+        #    means = _sums / (_counts + 1e-14)
+        #    stds = np.sqrt(_squared_sums / _counts - means ** 2)
+        #    print 'means', means
+        #    print 'stdevs', stds
+        # map up to original set of anchors
+        #print(labels.shape, total_anchors, inds_inside.shape, inds_inside[0], inds_inside[-1])
+        labels = AA._unmap(labels, total_anchors, inds_inside, fill=-1)
+        bbox_targets = AA._unmap(bbox_targets,
+                                 total_anchors,
+                                 inds_inside,
+                                 fill=0)
+        bbox_weights = AA._unmap(bbox_weights,
+                                 total_anchors,
+                                 inds_inside,
+                                 fill=0)
+        if landmark:
+            landmark_targets = AA._unmap(landmark_targets,
+                                         total_anchors,
+                                         inds_inside,
+                                         fill=0)
+            landmark_weights = AA._unmap(landmark_weights,
+                                         total_anchors,
+                                         inds_inside,
+                                         fill=0)
+        #print('CC', anchors.shape, len(inds_inside))
+
+        bbox_targets[:,
+                     0::4] = bbox_targets[:, 0::4] / config.TRAIN.BBOX_STDS[0]
+        bbox_targets[:,
+                     1::4] = bbox_targets[:, 1::4] / config.TRAIN.BBOX_STDS[1]
+        bbox_targets[:,
+                     2::4] = bbox_targets[:, 2::4] / config.TRAIN.BBOX_STDS[2]
+        bbox_targets[:,
+                     3::4] = bbox_targets[:, 3::4] / config.TRAIN.BBOX_STDS[3]
+        landmark_targets /= config.TRAIN.LANDMARK_STD
+        #print('applied STD')
+
+        #if DEBUG:
+        #    if gt_boxes.size > 0:
+        #        print 'rpn: max max_overlaps', np.max(max_overlaps)
+        #    print 'rpn: num_positives', np.sum(labels == 1)
+        #    print 'rpn: num_negatives', np.sum(labels == 0)
+        #    _fg_sum = np.sum(labels == 1)
+        #    _bg_sum = np.sum(labels == 0)
+        #    _count = 1
+        #    print 'rpn: num_positive avg', _fg_sum / _count
+        #    print 'rpn: num_negative avg', _bg_sum / _count
+
+        # resahpe
+        label_list = list()
+        bbox_target_list = list()
+        bbox_weight_list = list()
+        if landmark:
+            landmark_target_list = list()
+            landmark_weight_list = list()
+        anchors_num_range = [0] + anchors_num_list
+        label = {}
+        for i in range(len(feat_strides)):
+            stride = feat_strides[i]
+            feat_height, feat_width = feat_infos[i]
+            A = A_list[i]
+            _label = labels[sum(anchors_num_range[:i + 1]
+                                ):sum(anchors_num_range[:i + 1]) +
+                            anchors_num_range[i + 1]]
+            if select_stride > 0 and stride != select_stride:
+                #print('set', stride, select_stride)
+                _label[:] = -1
+            #print('_label', _label.shape, select_stride)
+            #_fg_inds = np.where(_label == 1)[0]
+            #n_fg = len(_fg_inds)
+            #STAT[0]+=1
+            #STAT[stride]+=n_fg
+            #if STAT[0]%100==0:
+            #  print('rpn_stat', STAT, file=sys.stderr)
+            bbox_target = bbox_targets[sum(anchors_num_range[:i + 1]
+                                           ):sum(anchors_num_range[:i + 1]) +
+                                       anchors_num_range[i + 1]]
+            bbox_weight = bbox_weights[sum(anchors_num_range[:i + 1]
+                                           ):sum(anchors_num_range[:i + 1]) +
+                                       anchors_num_range[i + 1]]
+            if landmark:
+                landmark_target = landmark_targets[
+                    sum(anchors_num_range[:i +
+                                          1]):sum(anchors_num_range[:i + 1]) +
+                    anchors_num_range[i + 1]]
+                landmark_weight = landmark_weights[
+                    sum(anchors_num_range[:i +
+                                          1]):sum(anchors_num_range[:i + 1]) +
+                    anchors_num_range[i + 1]]
+
+            _label = _label.reshape(
+                (1, feat_height, feat_width, A)).transpose(0, 3, 1, 2)
+            _label = _label.reshape((1, A * feat_height * feat_width))
+            bbox_target = bbox_target.reshape(
+                (1, feat_height * feat_width,
+                 A * bbox_pred_len)).transpose(0, 2, 1)
+            bbox_weight = bbox_weight.reshape(
+                (1, feat_height * feat_width, A * bbox_pred_len)).transpose(
+                    (0, 2, 1))
+            label['%s_label_stride%d' % (prefix, stride)] = _label
+            label['%s_bbox_target_stride%d' % (prefix, stride)] = bbox_target
+            label['%s_bbox_weight_stride%d' % (prefix, stride)] = bbox_weight
+            if landmark:
+                landmark_target = landmark_target.reshape(
+                    (1, feat_height * feat_width,
+                     A * landmark_pred_len)).transpose(0, 2, 1)
+                landmark_weight = landmark_weight.reshape(
+                    (1, feat_height * feat_width,
+                     A * landmark_pred_len)).transpose((0, 2, 1))
+                label['%s_landmark_target_stride%d' %
+                      (prefix, stride)] = landmark_target
+                label['%s_landmark_weight_stride%d' %
+                      (prefix, stride)] = landmark_weight
+            #print('in_rpn', stride,_label.shape, bbox_target.shape, bbox_weight.shape, file=sys.stderr)
+            label_list.append(_label)
+            #print('DD', _label.shape)
+            bbox_target_list.append(bbox_target)
+            bbox_weight_list.append(bbox_weight)
+            if landmark:
+                landmark_target_list.append(landmark_target)
+                landmark_weight_list.append(landmark_weight)
+
+        label_concat = np.concatenate(label_list, axis=1)
+        bbox_target_concat = np.concatenate(bbox_target_list, axis=2)
+        bbox_weight_concat = np.concatenate(bbox_weight_list, axis=2)
+        #fg_inds = np.where(label_concat[0] == 1)[0]
+        #print('fg_inds_in_rpn2', fg_inds, file=sys.stderr)
+
+        label.update({
+            '%s_label' % prefix: label_concat,
+            '%s_bbox_target' % prefix: bbox_target_concat,
+            '%s_bbox_weight' % prefix: bbox_weight_concat
+        })
+        if landmark:
+            landmark_target_concat = np.concatenate(landmark_target_list,
+                                                    axis=2)
+            landmark_weight_concat = np.concatenate(landmark_weight_list,
+                                                    axis=2)
+            label['%s_landmark_target' % prefix] = landmark_target_concat
+            label['%s_landmark_weight' % prefix] = landmark_weight_concat
+        #tb = datetime.datetime.now()
+        #self._times[3] += (tb-ta).total_seconds()
+        #ta = datetime.datetime.now()
+        #print(self._times)
+        return label
diff --git a/RetinaFace/rcnn/logger.py b/detection/retinaface/rcnn/logger.py
similarity index 100%
rename from RetinaFace/rcnn/logger.py
rename to detection/retinaface/rcnn/logger.py
diff --git a/src/common/__init__.py b/detection/retinaface/rcnn/processing/__init__.py
similarity index 100%
rename from src/common/__init__.py
rename to detection/retinaface/rcnn/processing/__init__.py
diff --git a/detection/retinaface/rcnn/processing/assign_levels.py b/detection/retinaface/rcnn/processing/assign_levels.py
new file mode 100755
index 000000000..012d73d21
--- /dev/null
+++ b/detection/retinaface/rcnn/processing/assign_levels.py
@@ -0,0 +1,36 @@
+from rcnn.config import config
+import numpy as np
+
+
+def compute_assign_targets(rois, threshold):
+    rois_area = np.sqrt(
+        (rois[:, 2] - rois[:, 0] + 1) * (rois[:, 3] - rois[:, 1] + 1))
+    num_rois = np.shape(rois)[0]
+    assign_levels = np.zeros(num_rois, dtype=np.uint8)
+    for i, stride in enumerate(config.RCNN_FEAT_STRIDE):
+        thd = threshold[i]
+        idx = np.logical_and(thd[1] <= rois_area, rois_area < thd[0])
+        assign_levels[idx] = stride
+
+    assert 0 not in assign_levels, "All rois should assign to specify levels."
+    return assign_levels
+
+
+def add_assign_targets(roidb):
+    """
+    given roidb, add ['assign_level']
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    """
+    print 'add assign targets'
+    assert len(roidb) > 0
+    assert 'boxes' in roidb[0]
+
+    area_threshold = [[np.inf, 448], [448, 224], [224, 112], [112, 0]]
+
+    assert len(config.RCNN_FEAT_STRIDE) == len(area_threshold)
+
+    num_images = len(roidb)
+    for im_i in range(num_images):
+        rois = roidb[im_i]['boxes']
+        roidb[im_i]['assign_levels'] = compute_assign_targets(
+            rois, area_threshold)
diff --git a/detection/retinaface/rcnn/processing/bbox_regression.py b/detection/retinaface/rcnn/processing/bbox_regression.py
new file mode 100644
index 000000000..0eaf917a6
--- /dev/null
+++ b/detection/retinaface/rcnn/processing/bbox_regression.py
@@ -0,0 +1,263 @@
+"""
+This file has functions about generating bounding box regression targets
+"""
+
+from ..pycocotools.mask import encode
+import numpy as np
+
+from ..logger import logger
+from .bbox_transform import bbox_overlaps, bbox_transform
+from rcnn.config import config
+import math
+import cv2
+import PIL.Image as Image
+import threading
+import Queue
+
+
+def compute_bbox_regression_targets(rois, overlaps, labels):
+    """
+    given rois, overlaps, gt labels, compute bounding box regression targets
+    :param rois: roidb[i]['boxes'] k * 4
+    :param overlaps: roidb[i]['max_overlaps'] k * 1
+    :param labels: roidb[i]['max_classes'] k * 1
+    :return: targets[i][class, dx, dy, dw, dh] k * 5
+    """
+    # Ensure ROIs are floats
+    rois = rois.astype(np.float, copy=False)
+
+    # Sanity check
+    if len(rois) != len(overlaps):
+        logger.warning('bbox regression: len(rois) != len(overlaps)')
+
+    # Indices of ground-truth ROIs
+    gt_inds = np.where(overlaps == 1)[0]
+    if len(gt_inds) == 0:
+        logger.warning('bbox regression: len(gt_inds) == 0')
+
+    # Indices of examples for which we try to make predictions
+    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
+
+    # Get IoU overlap between each ex ROI and gt ROI
+    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
+
+    # Find which gt ROI each ex ROI has max overlap with:
+    # this will be the ex ROI's gt target
+    gt_assignment = ex_gt_overlaps.argmax(axis=1)
+    gt_rois = rois[gt_inds[gt_assignment], :]
+    ex_rois = rois[ex_inds, :]
+
+    targets = np.zeros((rois.shape[0], 5), dtype=np.float32)
+    targets[ex_inds, 0] = labels[ex_inds]
+    targets[ex_inds, 1:] = bbox_transform(ex_rois, gt_rois)
+    return targets
+
+
+def add_bbox_regression_targets(roidb):
+    """
+    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    :return: means, std variances of targets
+    """
+    logger.info('bbox regression: add bounding box regression targets')
+    assert len(roidb) > 0
+    assert 'max_classes' in roidb[0]
+
+    num_images = len(roidb)
+    num_classes = roidb[0]['gt_overlaps'].shape[1]
+    for im_i in range(num_images):
+        rois = roidb[im_i]['boxes']
+        max_overlaps = roidb[im_i]['max_overlaps']
+        max_classes = roidb[im_i]['max_classes']
+        roidb[im_i]['bbox_targets'] = compute_bbox_regression_targets(
+            rois, max_overlaps, max_classes)
+
+    if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+        # use fixed / precomputed means and stds instead of empirical values
+        means = np.tile(np.array(config.TRAIN.BBOX_MEANS), (num_classes, 1))
+        stds = np.tile(np.array(config.TRAIN.BBOX_STDS), (num_classes, 1))
+    else:
+        # compute mean, std values
+        class_counts = np.zeros((num_classes, 1)) + 1e-14
+        sums = np.zeros((num_classes, 4))
+        squared_sums = np.zeros((num_classes, 4))
+        for im_i in range(num_images):
+            targets = roidb[im_i]['bbox_targets']
+            for cls in range(1, num_classes):
+                cls_indexes = np.where(targets[:, 0] == cls)[0]
+                if cls_indexes.size > 0:
+                    class_counts[cls] += cls_indexes.size
+                    sums[cls, :] += targets[cls_indexes, 1:].sum(axis=0)
+                    squared_sums[cls, :] += (targets[cls_indexes,
+                                                     1:]**2).sum(axis=0)
+
+        means = sums / class_counts
+        # var(x) = E(x^2) - E(x)^2
+        stds = np.sqrt(squared_sums / class_counts - means**2)
+
+    # normalized targets
+    for im_i in range(num_images):
+        targets = roidb[im_i]['bbox_targets']
+        for cls in range(1, num_classes):
+            cls_indexes = np.where(targets[:, 0] == cls)[0]
+            roidb[im_i]['bbox_targets'][cls_indexes, 1:] -= means[cls, :]
+            roidb[im_i]['bbox_targets'][cls_indexes, 1:] /= stds[cls, :]
+
+    return means.ravel(), stds.ravel()
+
+
+def expand_bbox_regression_targets(bbox_targets_data, num_classes):
+    """
+    expand from 5 to 4 * num_classes; only the right class has non-zero bbox regression targets
+    :param bbox_targets_data: [k * 5]
+    :param num_classes: number of classes
+    :return: bbox target processed [k * 4 num_classes]
+    bbox_weights ! only foreground boxes have bbox regression computation!
+    """
+    classes = bbox_targets_data[:, 0]
+    bbox_targets = np.zeros((classes.size, 4 * num_classes), dtype=np.float32)
+    bbox_weights = np.zeros(bbox_targets.shape, dtype=np.float32)
+    indexes = np.where(classes > 0)[0]
+    for index in indexes:
+        cls = classes[index]
+        start = int(4 * cls)
+        end = start + 4
+        bbox_targets[index, start:end] = bbox_targets_data[index, 1:]
+        bbox_weights[index, start:end] = config.TRAIN.BBOX_WEIGHTS
+    return bbox_targets, bbox_weights
+
+
+def compute_mask_and_label(ex_rois, ex_labels, seg, flipped):
+    # assert os.path.exists(seg_gt), 'Path does not exist: {}'.format(seg_gt)
+    # im = Image.open(seg_gt)
+    # pixel = list(im.getdata())
+    # pixel = np.array(pixel).reshape([im.size[1], im.size[0]])
+    im = Image.open(seg)
+    pixel = list(im.getdata())
+    ins_seg = np.array(pixel).reshape([im.size[1], im.size[0]])
+    if flipped:
+        ins_seg = ins_seg[:, ::-1]
+    rois = ex_rois
+    n_rois = ex_rois.shape[0]
+    label = ex_labels
+    class_id = config.CLASS_ID
+    mask_target = np.zeros((n_rois, 28, 28), dtype=np.int8)
+    mask_label = np.zeros((n_rois), dtype=np.int8)
+    for n in range(n_rois):
+        target = ins_seg[int(rois[n, 1]):int(rois[n, 3]),
+                         int(rois[n, 0]):int(rois[n, 2])]
+        ids = np.unique(target)
+        ins_id = 0
+        max_count = 0
+        for id in ids:
+            if math.floor(id / 1000) == class_id[int(label[int(n)])]:
+                px = np.where(ins_seg == int(id))
+                x_min = np.min(px[1])
+                y_min = np.min(px[0])
+                x_max = np.max(px[1])
+                y_max = np.max(px[0])
+                x1 = max(rois[n, 0], x_min)
+                y1 = max(rois[n, 1], y_min)
+                x2 = min(rois[n, 2], x_max)
+                y2 = min(rois[n, 3], y_max)
+                iou = (x2 - x1) * (y2 - y1)
+                iou = iou / ((rois[n, 2] - rois[n, 0]) *
+                             (rois[n, 3] - rois[n, 1]) + (x_max - x_min) *
+                             (y_max - y_min) - iou)
+                if iou > max_count:
+                    ins_id = id
+                    max_count = iou
+
+        if max_count == 0:
+            continue
+        # print max_count
+        mask = np.zeros(target.shape)
+        idx = np.where(target == ins_id)
+        mask[idx] = 1
+        mask = cv2.resize(mask, (28, 28), interpolation=cv2.INTER_NEAREST)
+
+        mask_target[n] = mask
+        mask_label[n] = label[int(n)]
+    return mask_target, mask_label
+
+
+def compute_bbox_mask_targets_and_label(rois, overlaps, labels, seg, flipped):
+    """
+    given rois, overlaps, gt labels, seg, compute bounding box mask targets
+    :param rois: roidb[i]['boxes'] k * 4
+    :param overlaps: roidb[i]['max_overlaps'] k * 1
+    :param labels: roidb[i]['max_classes'] k * 1
+    :return: targets[i][class, dx, dy, dw, dh] k * 5
+    """
+    # Ensure ROIs are floats
+    rois = rois.astype(np.float, copy=False)
+
+    # Sanity check
+    if len(rois) != len(overlaps):
+        print 'bbox regression: this should not happen'
+
+    # Indices of ground-truth ROIs
+    gt_inds = np.where(overlaps == 1)[0]
+    if len(gt_inds) == 0:
+        print 'something wrong : zero ground truth rois'
+    # Indices of examples for which we try to make predictions
+    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
+
+    # Get IoU overlap between each ex ROI and gt ROI
+    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
+
+    # Find which gt ROI each ex ROI has max overlap with:
+    # this will be the ex ROI's gt target
+    gt_assignment = ex_gt_overlaps.argmax(axis=1)
+    gt_rois = rois[gt_inds[gt_assignment], :]
+    ex_rois = rois[ex_inds, :]
+
+    mask_targets, mask_label = compute_mask_and_label(ex_rois, labels[ex_inds],
+                                                      seg, flipped)
+    return mask_targets, mask_label, ex_inds
+
+
+def add_mask_targets(roidb):
+    """
+    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    :return: means, std variances of targets
+    """
+    print 'add bounding box mask targets'
+    assert len(roidb) > 0
+    assert 'max_classes' in roidb[0]
+
+    num_images = len(roidb)
+
+    # Multi threads processing
+    im_quene = Queue.Queue(maxsize=0)
+    for im_i in range(num_images):
+        im_quene.put(im_i)
+
+    def process():
+        while not im_quene.empty():
+            im_i = im_quene.get()
+            print "-----process img {}".format(im_i)
+            rois = roidb[im_i]['boxes']
+            max_overlaps = roidb[im_i]['max_overlaps']
+            max_classes = roidb[im_i]['max_classes']
+            ins_seg = roidb[im_i]['ins_seg']
+            flipped = roidb[im_i]['flipped']
+            roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
+                compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg, flipped)
+
+    threads = [threading.Thread(target=process, args=()) for i in range(10)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+    # Single thread
+    # for im_i in range(num_images):
+    #     print "-----processing img {}".format(im_i)
+    #     rois = roidb[im_i]['boxes']
+    #     max_overlaps = roidb[im_i]['max_overlaps']
+    #     max_classes = roidb[im_i]['max_classes']
+    #     ins_seg = roidb[im_i]['ins_seg']
+    #     # roidb[im_i]['mask_targets'] = compute_bbox_mask_targets(rois, max_overlaps, max_classes, ins_seg)
+    #     roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
+    #         compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg)
diff --git a/detection/retinaface/rcnn/processing/bbox_transform.py b/detection/retinaface/rcnn/processing/bbox_transform.py
new file mode 100644
index 000000000..5ee3646ab
--- /dev/null
+++ b/detection/retinaface/rcnn/processing/bbox_transform.py
@@ -0,0 +1,223 @@
+import numpy as np
+from ..cython.bbox import bbox_overlaps_cython
+#from rcnn.config import config
+
+
+def bbox_overlaps(boxes, query_boxes):
+    return bbox_overlaps_cython(boxes, query_boxes)
+
+
+def bbox_overlaps_py(boxes, query_boxes):
+    """
+    determine overlaps between boxes and query_boxes
+    :param boxes: n * 4 bounding boxes
+    :param query_boxes: k * 4 bounding boxes
+    :return: overlaps: n * k overlaps
+    """
+    n_ = boxes.shape[0]
+    k_ = query_boxes.shape[0]
+    overlaps = np.zeros((n_, k_), dtype=np.float)
+    for k in range(k_):
+        query_box_area = (query_boxes[k, 2] - query_boxes[k, 0] +
+                          1) * (query_boxes[k, 3] - query_boxes[k, 1] + 1)
+        for n in range(n_):
+            iw = min(boxes[n, 2], query_boxes[k, 2]) - max(
+                boxes[n, 0], query_boxes[k, 0]) + 1
+            if iw > 0:
+                ih = min(boxes[n, 3], query_boxes[k, 3]) - max(
+                    boxes[n, 1], query_boxes[k, 1]) + 1
+                if ih > 0:
+                    box_area = (boxes[n, 2] - boxes[n, 0] +
+                                1) * (boxes[n, 3] - boxes[n, 1] + 1)
+                    all_area = float(box_area + query_box_area - iw * ih)
+                    overlaps[n, k] = iw * ih / all_area
+    return overlaps
+
+
+def clip_boxes(boxes, im_shape):
+    """
+    Clip boxes to image boundaries.
+    :param boxes: [N, 4* num_classes]
+    :param im_shape: tuple of 2
+    :return: [N, 4* num_classes]
+    """
+    # x1 >= 0
+    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
+    # y1 >= 0
+    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
+    # x2 < im_shape[1]
+    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
+    # y2 < im_shape[0]
+    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
+    return boxes
+
+
+def nonlinear_transform(ex_rois, gt_rois):
+    """
+    compute bounding box regression targets from ex_rois to gt_rois
+    :param ex_rois: [N, 4]
+    :param gt_rois: [N, 4]
+    :return: [N, 4]
+    """
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+
+    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
+    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
+    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
+    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
+
+    gt_widths = gt_rois[:, 2] - gt_rois[:, 0] + 1.0
+    gt_heights = gt_rois[:, 3] - gt_rois[:, 1] + 1.0
+    gt_ctr_x = gt_rois[:, 0] + 0.5 * (gt_widths - 1.0)
+    gt_ctr_y = gt_rois[:, 1] + 0.5 * (gt_heights - 1.0)
+
+    targets_dx = (gt_ctr_x - ex_ctr_x) / (ex_widths + 1e-14)
+    targets_dy = (gt_ctr_y - ex_ctr_y) / (ex_heights + 1e-14)
+    targets_dw = np.log(gt_widths / ex_widths)
+    targets_dh = np.log(gt_heights / ex_heights)
+
+    if gt_rois.shape[1] <= 4:
+        targets = np.vstack(
+            (targets_dx, targets_dy, targets_dw, targets_dh)).transpose()
+        return targets
+    else:
+        targets = [targets_dx, targets_dy, targets_dw, targets_dh]
+        #if config.USE_BLUR:
+        #  for i in range(4, gt_rois.shape[1]):
+        #    t = gt_rois[:,i]
+        #    targets.append(t)
+        targets = np.vstack(targets).transpose()
+        return targets
+
+
+def landmark_transform(ex_rois, gt_rois):
+
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+
+    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
+    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
+    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
+    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
+
+    targets = []
+    for i in range(gt_rois.shape[1]):
+        for j in range(gt_rois.shape[2]):
+            #if not config.USE_OCCLUSION and j==2:
+            #  continue
+            if j == 2:
+                continue
+            if j == 0:  #w
+                target = (gt_rois[:, i, j] - ex_ctr_x) / (ex_widths + 1e-14)
+            elif j == 1:  #h
+                target = (gt_rois[:, i, j] - ex_ctr_y) / (ex_heights + 1e-14)
+            else:  #visibile
+                target = gt_rois[:, i, j]
+            targets.append(target)
+
+    targets = np.vstack(targets).transpose()
+    return targets
+
+
+def nonlinear_pred(boxes, box_deltas):
+    """
+    Transform the set of class-agnostic boxes into class-specific boxes
+    by applying the predicted offsets (box_deltas)
+    :param boxes: !important [N 4]
+    :param box_deltas: [N, 4 * num_classes]
+    :return: [N 4 * num_classes]
+    """
+    if boxes.shape[0] == 0:
+        return np.zeros((0, box_deltas.shape[1]))
+
+    boxes = boxes.astype(np.float, copy=False)
+    widths = boxes[:, 2] - boxes[:, 0] + 1.0
+    heights = boxes[:, 3] - boxes[:, 1] + 1.0
+    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+
+    dx = box_deltas[:, 0::4]
+    dy = box_deltas[:, 1::4]
+    dw = box_deltas[:, 2::4]
+    dh = box_deltas[:, 3::4]
+
+    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
+    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
+    pred_w = np.exp(dw) * widths[:, np.newaxis]
+    pred_h = np.exp(dh) * heights[:, np.newaxis]
+
+    pred_boxes = np.zeros(box_deltas.shape)
+    # x1
+    pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * (pred_w - 1.0)
+    # y1
+    pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * (pred_h - 1.0)
+    # x2
+    pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * (pred_w - 1.0)
+    # y2
+    pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
+
+    return pred_boxes
+
+
+def landmark_pred(boxes, landmark_deltas):
+    if boxes.shape[0] == 0:
+        return np.zeros((0, landmark_deltas.shape[1]))
+    boxes = boxes.astype(np.float, copy=False)
+    widths = boxes[:, 2] - boxes[:, 0] + 1.0
+    heights = boxes[:, 3] - boxes[:, 1] + 1.0
+    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+    preds = []
+    for i in range(landmark_deltas.shape[1]):
+        if i % 2 == 0:
+            pred = (landmark_deltas[:, i] * widths + ctr_x)
+        else:
+            pred = (landmark_deltas[:, i] * heights + ctr_y)
+        preds.append(pred)
+    preds = np.vstack(preds).transpose()
+    return preds
+
+
+def iou_transform(ex_rois, gt_rois):
+    """ return bbox targets, IoU loss uses gt_rois as gt """
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+    return gt_rois
+
+
+def iou_pred(boxes, box_deltas):
+    """
+    Transform the set of class-agnostic boxes into class-specific boxes
+    by applying the predicted offsets (box_deltas)
+    :param boxes: !important [N 4]
+    :param box_deltas: [N, 4 * num_classes]
+    :return: [N 4 * num_classes]
+    """
+    if boxes.shape[0] == 0:
+        return np.zeros((0, box_deltas.shape[1]))
+
+    boxes = boxes.astype(np.float, copy=False)
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+
+    dx1 = box_deltas[:, 0::4]
+    dy1 = box_deltas[:, 1::4]
+    dx2 = box_deltas[:, 2::4]
+    dy2 = box_deltas[:, 3::4]
+
+    pred_boxes = np.zeros(box_deltas.shape)
+    # x1
+    pred_boxes[:, 0::4] = dx1 + x1[:, np.newaxis]
+    # y1
+    pred_boxes[:, 1::4] = dy1 + y1[:, np.newaxis]
+    # x2
+    pred_boxes[:, 2::4] = dx2 + x2[:, np.newaxis]
+    # y2
+    pred_boxes[:, 3::4] = dy2 + y2[:, np.newaxis]
+
+    return pred_boxes
+
+
+# define bbox_transform and bbox_pred
+bbox_transform = nonlinear_transform
+bbox_pred = nonlinear_pred
diff --git a/detection/retinaface/rcnn/processing/generate_anchor.py b/detection/retinaface/rcnn/processing/generate_anchor.py
new file mode 100644
index 000000000..83c5ada2a
--- /dev/null
+++ b/detection/retinaface/rcnn/processing/generate_anchor.py
@@ -0,0 +1,135 @@
+"""
+Generate base anchors on index 0
+"""
+from __future__ import print_function
+import sys
+from builtins import range
+import numpy as np
+from ..cython.anchors import anchors_cython
+#from ..config import config
+
+
+def anchors_plane(feat_h, feat_w, stride, base_anchor):
+    return anchors_cython(feat_h, feat_w, stride, base_anchor)
+
+
+def generate_anchors(base_size=16,
+                     ratios=[0.5, 1, 2],
+                     scales=2**np.arange(3, 6),
+                     stride=16,
+                     dense_anchor=False):
+    """
+    Generate anchor (reference) windows by enumerating aspect ratios X
+    scales wrt a reference (0, 0, 15, 15) window.
+    """
+
+    base_anchor = np.array([1, 1, base_size, base_size]) - 1
+    ratio_anchors = _ratio_enum(base_anchor, ratios)
+    anchors = np.vstack([
+        _scale_enum(ratio_anchors[i, :], scales)
+        for i in range(ratio_anchors.shape[0])
+    ])
+    if dense_anchor:
+        assert stride % 2 == 0
+        anchors2 = anchors.copy()
+        anchors2[:, :] += int(stride / 2)
+        anchors = np.vstack((anchors, anchors2))
+    #print('GA',base_anchor.shape, ratio_anchors.shape, anchors.shape)
+    return anchors
+
+
+#def generate_anchors_fpn(base_size=[64,32,16,8,4], ratios=[0.5, 1, 2], scales=8):
+#    """
+#    Generate anchor (reference) windows by enumerating aspect ratios X
+#    scales wrt a reference (0, 0, 15, 15) window.
+#    """
+#    anchors = []
+#    _ratios = ratios.reshape( (len(base_size), -1) )
+#    _scales = scales.reshape( (len(base_size), -1) )
+#    for i,bs in enumerate(base_size):
+#      __ratios = _ratios[i]
+#      __scales = _scales[i]
+#      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
+#      r = generate_anchors(bs, __ratios, __scales)
+#      #print('anchors_fpn', r.shape, file=sys.stderr)
+#      anchors.append(r)
+#    return anchors
+
+
+def generate_anchors_fpn(dense_anchor=False, cfg=None):
+    #assert(False)
+    """
+    Generate anchor (reference) windows by enumerating aspect ratios X
+    scales wrt a reference (0, 0, 15, 15) window.
+    """
+    if cfg is None:
+        from ..config import config
+        cfg = config.RPN_ANCHOR_CFG
+    RPN_FEAT_STRIDE = []
+    for k in cfg:
+        RPN_FEAT_STRIDE.append(int(k))
+    RPN_FEAT_STRIDE = sorted(RPN_FEAT_STRIDE, reverse=True)
+    anchors = []
+    for k in RPN_FEAT_STRIDE:
+        v = cfg[str(k)]
+        bs = v['BASE_SIZE']
+        __ratios = np.array(v['RATIOS'])
+        __scales = np.array(v['SCALES'])
+        stride = int(k)
+        #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
+        r = generate_anchors(bs, __ratios, __scales, stride, dense_anchor)
+        #print('anchors_fpn', r.shape, file=sys.stderr)
+        anchors.append(r)
+
+    return anchors
+
+
+def _whctrs(anchor):
+    """
+    Return width, height, x center, and y center for an anchor (window).
+    """
+
+    w = anchor[2] - anchor[0] + 1
+    h = anchor[3] - anchor[1] + 1
+    x_ctr = anchor[0] + 0.5 * (w - 1)
+    y_ctr = anchor[1] + 0.5 * (h - 1)
+    return w, h, x_ctr, y_ctr
+
+
+def _mkanchors(ws, hs, x_ctr, y_ctr):
+    """
+    Given a vector of widths (ws) and heights (hs) around a center
+    (x_ctr, y_ctr), output a set of anchors (windows).
+    """
+
+    ws = ws[:, np.newaxis]
+    hs = hs[:, np.newaxis]
+    anchors = np.hstack((x_ctr - 0.5 * (ws - 1), y_ctr - 0.5 * (hs - 1),
+                         x_ctr + 0.5 * (ws - 1), y_ctr + 0.5 * (hs - 1)))
+    return anchors
+
+
+def _ratio_enum(anchor, ratios):
+    """
+    Enumerate a set of anchors for each aspect ratio wrt an anchor.
+    """
+
+    w, h, x_ctr, y_ctr = _whctrs(anchor)
+    size = w * h
+    size_ratios = size / ratios
+    ws = np.round(np.sqrt(size_ratios))
+    hs = np.round(ws * ratios)
+    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
+    return anchors
+
+
+def _scale_enum(anchor, scales):
+    """
+    Enumerate a set of anchors for each scale wrt an anchor.
+    """
+
+    w, h, x_ctr, y_ctr = _whctrs(anchor)
+    ws = w * scales
+    hs = h * scales
+    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
+    return anchors
diff --git a/detection/retinaface/rcnn/processing/nms.py b/detection/retinaface/rcnn/processing/nms.py
new file mode 100644
index 000000000..b32d92d0f
--- /dev/null
+++ b/detection/retinaface/rcnn/processing/nms.py
@@ -0,0 +1,67 @@
+import numpy as np
+from ..cython.cpu_nms import cpu_nms
+try:
+    from ..cython.gpu_nms import gpu_nms
+except ImportError:
+    gpu_nms = None
+
+
+def py_nms_wrapper(thresh):
+    def _nms(dets):
+        return nms(dets, thresh)
+
+    return _nms
+
+
+def cpu_nms_wrapper(thresh):
+    def _nms(dets):
+        return cpu_nms(dets, thresh)
+
+    return _nms
+
+
+def gpu_nms_wrapper(thresh, device_id):
+    def _nms(dets):
+        return gpu_nms(dets, thresh, device_id)
+
+    if gpu_nms is not None:
+        return _nms
+    else:
+        return cpu_nms_wrapper(thresh)
+
+
+def nms(dets, thresh):
+    """
+    greedily select boxes with high confidence and overlap with current maximum <= thresh
+    rule out overlap >= thresh
+    :param dets: [[x1, y1, x2, y2 score]]
+    :param thresh: retain overlap < thresh
+    :return: indexes to keep
+    """
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    scores = dets[:, 4]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= thresh)[0]
+        order = order[inds + 1]
+
+    return keep
diff --git a/RetinaFace/rcnn/pycocotools/UPSTREAM_REV b/detection/retinaface/rcnn/pycocotools/UPSTREAM_REV
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/UPSTREAM_REV
rename to detection/retinaface/rcnn/pycocotools/UPSTREAM_REV
diff --git a/RetinaFace/rcnn/pycocotools/__init__.py b/detection/retinaface/rcnn/pycocotools/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/__init__.py
rename to detection/retinaface/rcnn/pycocotools/__init__.py
diff --git a/RetinaFace/rcnn/pycocotools/_mask.c b/detection/retinaface/rcnn/pycocotools/_mask.c
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/_mask.c
rename to detection/retinaface/rcnn/pycocotools/_mask.c
diff --git a/RetinaFace/rcnn/pycocotools/_mask.pyx b/detection/retinaface/rcnn/pycocotools/_mask.pyx
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/_mask.pyx
rename to detection/retinaface/rcnn/pycocotools/_mask.pyx
diff --git a/detection/retinaface/rcnn/pycocotools/coco.py b/detection/retinaface/rcnn/pycocotools/coco.py
new file mode 100644
index 000000000..4f7923652
--- /dev/null
+++ b/detection/retinaface/rcnn/pycocotools/coco.py
@@ -0,0 +1,477 @@
+__author__ = 'tylin'
+__version__ = '2.0'
+# Interface for accessing the Microsoft COCO dataset.
+
+# Microsoft COCO is a large image dataset designed for object detection,
+# segmentation, and caption generation. pycocotools is a Python API that
+# assists in loading, parsing and visualizing the annotations in COCO.
+# Please visit http://mscoco.org/ for more information on COCO, including
+# for the data, paper, and tutorials. The exact format of the annotations
+# is also described on the COCO website. For example usage of the pycocotools
+# please see pycocotools_demo.ipynb. In addition to this API, please download both
+# the COCO images and annotations in order to run the demo.
+
+# An alternative to using the API is to load the annotations directly
+# into Python dictionary
+# Using the API provides additional utility functions. Note that this API
+# supports both *instance* and *caption* annotations. In the case of
+# captions not all functions are defined (e.g. categories are undefined).
+
+# The following API functions are defined:
+#  COCO       - COCO api class that loads COCO annotation file and prepare data structures.
+#  decodeMask - Decode binary mask M encoded via run-length encoding.
+#  encodeMask - Encode binary mask M using run-length encoding.
+#  getAnnIds  - Get ann ids that satisfy given filter conditions.
+#  getCatIds  - Get cat ids that satisfy given filter conditions.
+#  getImgIds  - Get img ids that satisfy given filter conditions.
+#  loadAnns   - Load anns with the specified ids.
+#  loadCats   - Load cats with the specified ids.
+#  loadImgs   - Load imgs with the specified ids.
+#  annToMask  - Convert segmentation in an annotation to binary mask.
+#  showAnns   - Display the specified annotations.
+#  loadRes    - Load algorithm results and create API for accessing them.
+#  download   - Download COCO images from mscoco.org server.
+# Throughout the API "ann"=annotation, "cat"=category, and "img"=image.
+# Help on each functions can be accessed by: "help COCO>function".
+
+# See also COCO>decodeMask,
+# COCO>encodeMask, COCO>getAnnIds, COCO>getCatIds,
+# COCO>getImgIds, COCO>loadAnns, COCO>loadCats,
+# COCO>loadImgs, COCO>annToMask, COCO>showAnns
+
+# Microsoft COCO Toolbox.      version 2.0
+# Data, paper, and tutorials available at:  http://mscoco.org/
+# Code written by Piotr Dollar and Tsung-Yi Lin, 2014.
+# Licensed under the Simplified BSD License [see bsd.txt]
+
+import json
+import time
+import matplotlib.pyplot as plt
+from matplotlib.collections import PatchCollection
+from matplotlib.patches import Polygon
+import numpy as np
+import copy
+import itertools
+from . import mask as maskUtils
+import os
+from collections import defaultdict
+import sys
+PYTHON_VERSION = sys.version_info[0]
+if PYTHON_VERSION == 2:
+    from urllib import urlretrieve
+elif PYTHON_VERSION == 3:
+    from urllib.request import urlretrieve
+
+
+class COCO:
+    def __init__(self, annotation_file=None):
+        """
+        Constructor of Microsoft COCO helper class for reading and visualizing annotations.
+        :param annotation_file (str): location of annotation file
+        :param image_folder (str): location to the folder that hosts images.
+        :return:
+        """
+        # load dataset
+        self.dataset, self.anns, self.cats, self.imgs = dict(), dict(), dict(
+        ), dict()
+        self.imgToAnns, self.catToImgs = defaultdict(list), defaultdict(list)
+        if not annotation_file == None:
+            print('loading annotations into memory...')
+            tic = time.time()
+            dataset = json.load(open(annotation_file, 'r'))
+            assert type(
+                dataset
+            ) == dict, 'annotation file format {} not supported'.format(
+                type(dataset))
+            print('Done (t={:0.2f}s)'.format(time.time() - tic))
+            self.dataset = dataset
+            self.createIndex()
+
+    def createIndex(self):
+        # create index
+        print('creating index...')
+        anns, cats, imgs = {}, {}, {}
+        imgToAnns, catToImgs = defaultdict(list), defaultdict(list)
+        if 'annotations' in self.dataset:
+            for ann in self.dataset['annotations']:
+                imgToAnns[ann['image_id']].append(ann)
+                anns[ann['id']] = ann
+
+        if 'images' in self.dataset:
+            for img in self.dataset['images']:
+                imgs[img['id']] = img
+
+        if 'categories' in self.dataset:
+            for cat in self.dataset['categories']:
+                cats[cat['id']] = cat
+
+        if 'annotations' in self.dataset and 'categories' in self.dataset:
+            for ann in self.dataset['annotations']:
+                catToImgs[ann['category_id']].append(ann['image_id'])
+
+        print('index created!')
+
+        # create class members
+        self.anns = anns
+        self.imgToAnns = imgToAnns
+        self.catToImgs = catToImgs
+        self.imgs = imgs
+        self.cats = cats
+
+    def info(self):
+        """
+        Print information about the annotation file.
+        :return:
+        """
+        for key, value in self.dataset['info'].items():
+            print('{}: {}'.format(key, value))
+
+    def getAnnIds(self, imgIds=[], catIds=[], areaRng=[], iscrowd=None):
+        """
+        Get ann ids that satisfy given filter conditions. default skips that filter
+        :param imgIds  (int array)     : get anns for given imgs
+               catIds  (int array)     : get anns for given cats
+               areaRng (float array)   : get anns for given area range (e.g. [0 inf])
+               iscrowd (boolean)       : get anns for given crowd label (False or True)
+        :return: ids (int array)       : integer array of ann ids
+        """
+        imgIds = imgIds if type(imgIds) == list else [imgIds]
+        catIds = catIds if type(catIds) == list else [catIds]
+
+        if len(imgIds) == len(catIds) == len(areaRng) == 0:
+            anns = self.dataset['annotations']
+        else:
+            if not len(imgIds) == 0:
+                lists = [
+                    self.imgToAnns[imgId] for imgId in imgIds
+                    if imgId in self.imgToAnns
+                ]
+                anns = list(itertools.chain.from_iterable(lists))
+            else:
+                anns = self.dataset['annotations']
+            anns = anns if len(catIds) == 0 else [
+                ann for ann in anns if ann['category_id'] in catIds
+            ]
+            anns = anns if len(areaRng) == 0 else [
+                ann for ann in anns
+                if ann['area'] > areaRng[0] and ann['area'] < areaRng[1]
+            ]
+        if not iscrowd == None:
+            ids = [ann['id'] for ann in anns if ann['iscrowd'] == iscrowd]
+        else:
+            ids = [ann['id'] for ann in anns]
+        return ids
+
+    def getCatIds(self, catNms=[], supNms=[], catIds=[]):
+        """
+        filtering parameters. default skips that filter.
+        :param catNms (str array)  : get cats for given cat names
+        :param supNms (str array)  : get cats for given supercategory names
+        :param catIds (int array)  : get cats for given cat ids
+        :return: ids (int array)   : integer array of cat ids
+        """
+        catNms = catNms if type(catNms) == list else [catNms]
+        supNms = supNms if type(supNms) == list else [supNms]
+        catIds = catIds if type(catIds) == list else [catIds]
+
+        if len(catNms) == len(supNms) == len(catIds) == 0:
+            cats = self.dataset['categories']
+        else:
+            cats = self.dataset['categories']
+            cats = cats if len(catNms) == 0 else [
+                cat for cat in cats if cat['name'] in catNms
+            ]
+            cats = cats if len(supNms) == 0 else [
+                cat for cat in cats if cat['supercategory'] in supNms
+            ]
+            cats = cats if len(catIds) == 0 else [
+                cat for cat in cats if cat['id'] in catIds
+            ]
+        ids = [cat['id'] for cat in cats]
+        return ids
+
+    def getImgIds(self, imgIds=[], catIds=[]):
+        '''
+        Get img ids that satisfy given filter conditions.
+        :param imgIds (int array) : get imgs for given ids
+        :param catIds (int array) : get imgs with all given cats
+        :return: ids (int array)  : integer array of img ids
+        '''
+        imgIds = imgIds if type(imgIds) == list else [imgIds]
+        catIds = catIds if type(catIds) == list else [catIds]
+
+        if len(imgIds) == len(catIds) == 0:
+            ids = self.imgs.keys()
+        else:
+            ids = set(imgIds)
+            for i, catId in enumerate(catIds):
+                if i == 0 and len(ids) == 0:
+                    ids = set(self.catToImgs[catId])
+                else:
+                    ids &= set(self.catToImgs[catId])
+        return list(ids)
+
+    def loadAnns(self, ids=[]):
+        """
+        Load anns with the specified ids.
+        :param ids (int array)       : integer ids specifying anns
+        :return: anns (object array) : loaded ann objects
+        """
+        if type(ids) == list:
+            return [self.anns[id] for id in ids]
+        elif type(ids) == int:
+            return [self.anns[ids]]
+
+    def loadCats(self, ids=[]):
+        """
+        Load cats with the specified ids.
+        :param ids (int array)       : integer ids specifying cats
+        :return: cats (object array) : loaded cat objects
+        """
+        if type(ids) == list:
+            return [self.cats[id] for id in ids]
+        elif type(ids) == int:
+            return [self.cats[ids]]
+
+    def loadImgs(self, ids=[]):
+        """
+        Load anns with the specified ids.
+        :param ids (int array)       : integer ids specifying img
+        :return: imgs (object array) : loaded img objects
+        """
+        if type(ids) == list:
+            return [self.imgs[id] for id in ids]
+        elif type(ids) == int:
+            return [self.imgs[ids]]
+
+    def showAnns(self, anns):
+        """
+        Display the specified annotations.
+        :param anns (array of object): annotations to display
+        :return: None
+        """
+        if len(anns) == 0:
+            return 0
+        if 'segmentation' in anns[0] or 'keypoints' in anns[0]:
+            datasetType = 'instances'
+        elif 'caption' in anns[0]:
+            datasetType = 'captions'
+        else:
+            raise Exception('datasetType not supported')
+        if datasetType == 'instances':
+            ax = plt.gca()
+            ax.set_autoscale_on(False)
+            polygons = []
+            color = []
+            for ann in anns:
+                c = (np.random.random((1, 3)) * 0.6 + 0.4).tolist()[0]
+                if 'segmentation' in ann:
+                    if type(ann['segmentation']) == list:
+                        # polygon
+                        for seg in ann['segmentation']:
+                            poly = np.array(seg).reshape(
+                                (int(len(seg) / 2), 2))
+                            polygons.append(Polygon(poly))
+                            color.append(c)
+                    else:
+                        # mask
+                        t = self.imgs[ann['image_id']]
+                        if type(ann['segmentation']['counts']) == list:
+                            rle = maskUtils.frPyObjects([ann['segmentation']],
+                                                        t['height'],
+                                                        t['width'])
+                        else:
+                            rle = [ann['segmentation']]
+                        m = maskUtils.decode(rle)
+                        img = np.ones((m.shape[0], m.shape[1], 3))
+                        if ann['iscrowd'] == 1:
+                            color_mask = np.array([2.0, 166.0, 101.0]) / 255
+                        if ann['iscrowd'] == 0:
+                            color_mask = np.random.random((1, 3)).tolist()[0]
+                        for i in range(3):
+                            img[:, :, i] = color_mask[i]
+                        ax.imshow(np.dstack((img, m * 0.5)))
+                if 'keypoints' in ann and type(ann['keypoints']) == list:
+                    # turn skeleton into zero-based index
+                    sks = np.array(
+                        self.loadCats(ann['category_id'])[0]['skeleton']) - 1
+                    kp = np.array(ann['keypoints'])
+                    x = kp[0::3]
+                    y = kp[1::3]
+                    v = kp[2::3]
+                    for sk in sks:
+                        if np.all(v[sk] > 0):
+                            plt.plot(x[sk], y[sk], linewidth=3, color=c)
+                    plt.plot(x[v > 0],
+                             y[v > 0],
+                             'o',
+                             markersize=8,
+                             markerfacecolor=c,
+                             markeredgecolor='k',
+                             markeredgewidth=2)
+                    plt.plot(x[v > 1],
+                             y[v > 1],
+                             'o',
+                             markersize=8,
+                             markerfacecolor=c,
+                             markeredgecolor=c,
+                             markeredgewidth=2)
+            p = PatchCollection(polygons,
+                                facecolor=color,
+                                linewidths=0,
+                                alpha=0.4)
+            ax.add_collection(p)
+            p = PatchCollection(polygons,
+                                facecolor='none',
+                                edgecolors=color,
+                                linewidths=2)
+            ax.add_collection(p)
+        elif datasetType == 'captions':
+            for ann in anns:
+                print(ann['caption'])
+
+    def loadRes(self, resFile):
+        """
+        Load result file and return a result api object.
+        :param   resFile (str)     : file name of result file
+        :return: res (obj)         : result api object
+        """
+        res = COCO()
+        res.dataset['images'] = [img for img in self.dataset['images']]
+
+        print('Loading and preparing results...')
+        tic = time.time()
+        if type(resFile) == str or type(resFile) == unicode:
+            anns = json.load(open(resFile))
+        elif type(resFile) == np.ndarray:
+            anns = self.loadNumpyAnnotations(resFile)
+        else:
+            anns = resFile
+        assert type(anns) == list, 'results in not an array of objects'
+        annsImgIds = [ann['image_id'] for ann in anns]
+        assert set(annsImgIds) == (set(annsImgIds) & set(self.getImgIds())), \
+               'Results do not correspond to current coco set'
+        if 'caption' in anns[0]:
+            imgIds = set([img['id'] for img in res.dataset['images']]) & set(
+                [ann['image_id'] for ann in anns])
+            res.dataset['images'] = [
+                img for img in res.dataset['images'] if img['id'] in imgIds
+            ]
+            for id, ann in enumerate(anns):
+                ann['id'] = id + 1
+        elif 'bbox' in anns[0] and not anns[0]['bbox'] == []:
+            res.dataset['categories'] = copy.deepcopy(
+                self.dataset['categories'])
+            for id, ann in enumerate(anns):
+                bb = ann['bbox']
+                x1, x2, y1, y2 = [bb[0], bb[0] + bb[2], bb[1], bb[1] + bb[3]]
+                if not 'segmentation' in ann:
+                    ann['segmentation'] = [[x1, y1, x1, y2, x2, y2, x2, y1]]
+                ann['area'] = bb[2] * bb[3]
+                ann['id'] = id + 1
+                ann['iscrowd'] = 0
+        elif 'segmentation' in anns[0]:
+            res.dataset['categories'] = copy.deepcopy(
+                self.dataset['categories'])
+            for id, ann in enumerate(anns):
+                # now only support compressed RLE format as segmentation results
+                ann['area'] = maskUtils.area(ann['segmentation'])
+                if not 'bbox' in ann:
+                    ann['bbox'] = maskUtils.toBbox(ann['segmentation'])
+                ann['id'] = id + 1
+                ann['iscrowd'] = 0
+        elif 'keypoints' in anns[0]:
+            res.dataset['categories'] = copy.deepcopy(
+                self.dataset['categories'])
+            for id, ann in enumerate(anns):
+                s = ann['keypoints']
+                x = s[0::3]
+                y = s[1::3]
+                x0, x1, y0, y1 = np.min(x), np.max(x), np.min(y), np.max(y)
+                ann['area'] = (x1 - x0) * (y1 - y0)
+                ann['id'] = id + 1
+                ann['bbox'] = [x0, y0, x1 - x0, y1 - y0]
+        print('DONE (t={:0.2f}s)'.format(time.time() - tic))
+
+        res.dataset['annotations'] = anns
+        res.createIndex()
+        return res
+
+    def download(self, tarDir=None, imgIds=[]):
+        '''
+        Download COCO images from mscoco.org server.
+        :param tarDir (str): COCO results directory name
+               imgIds (list): images to be downloaded
+        :return:
+        '''
+        if tarDir is None:
+            print('Please specify target directory')
+            return -1
+        if len(imgIds) == 0:
+            imgs = self.imgs.values()
+        else:
+            imgs = self.loadImgs(imgIds)
+        N = len(imgs)
+        if not os.path.exists(tarDir):
+            os.makedirs(tarDir)
+        for i, img in enumerate(imgs):
+            tic = time.time()
+            fname = os.path.join(tarDir, img['file_name'])
+            if not os.path.exists(fname):
+                urlretrieve(img['coco_url'], fname)
+            print('downloaded {}/{} images (t={:0.1f}s)'.format(
+                i, N,
+                time.time() - tic))
+
+    def loadNumpyAnnotations(self, data):
+        """
+        Convert result data from a numpy array [Nx7] where each row contains {imageID,x1,y1,w,h,score,class}
+        :param  data (numpy.ndarray)
+        :return: annotations (python nested list)
+        """
+        print('Converting ndarray to lists...')
+        assert (type(data) == np.ndarray)
+        print(data.shape)
+        assert (data.shape[1] == 7)
+        N = data.shape[0]
+        ann = []
+        for i in range(N):
+            if i % 1000000 == 0:
+                print('{}/{}'.format(i, N))
+            ann += [{
+                'image_id': int(data[i, 0]),
+                'bbox': [data[i, 1], data[i, 2], data[i, 3], data[i, 4]],
+                'score': data[i, 5],
+                'category_id': int(data[i, 6]),
+            }]
+        return ann
+
+    def annToRLE(self, ann):
+        """
+        Convert annotation which can be polygons, uncompressed RLE to RLE.
+        :return: binary mask (numpy 2D array)
+        """
+        t = self.imgs[ann['image_id']]
+        h, w = t['height'], t['width']
+        segm = ann['segmentation']
+        if type(segm) == list:
+            # polygon -- a single object might consist of multiple parts
+            # we merge all parts into one mask rle code
+            rles = maskUtils.frPyObjects(segm, h, w)
+            rle = maskUtils.merge(rles)
+        elif type(segm['counts']) == list:
+            # uncompressed RLE
+            rle = maskUtils.frPyObjects(segm, h, w)
+        else:
+            # rle
+            rle = ann['segmentation']
+        return rle
+
+    def annToMask(self, ann):
+        """
+        Convert annotation which can be polygons, uncompressed RLE, or RLE to binary mask.
+        :return: binary mask (numpy 2D array)
+        """
+        rle = self.annToRLE(ann)
+        m = maskUtils.decode(rle)
+        return m
diff --git a/detection/retinaface/rcnn/pycocotools/cocoeval.py b/detection/retinaface/rcnn/pycocotools/cocoeval.py
new file mode 100644
index 000000000..1d3531847
--- /dev/null
+++ b/detection/retinaface/rcnn/pycocotools/cocoeval.py
@@ -0,0 +1,592 @@
+__author__ = 'tsungyi'
+
+import numpy as np
+import datetime
+import time
+from collections import defaultdict
+from .mask import *
+import copy
+
+
+class COCOeval:
+    # Interface for evaluating detection on the Microsoft COCO dataset.
+    #
+    # The usage for CocoEval is as follows:
+    #  cocoGt=..., cocoDt=...       # load dataset and results
+    #  E = CocoEval(cocoGt,cocoDt); # initialize CocoEval object
+    #  E.params.recThrs = ...;      # set parameters as desired
+    #  E.evaluate();                # run per image evaluation
+    #  E.accumulate();              # accumulate per image results
+    #  E.summarize();               # display summary metrics of results
+    # For example usage see evalDemo.m and http://mscoco.org/.
+    #
+    # The evaluation parameters are as follows (defaults in brackets):
+    #  imgIds     - [all] N img ids to use for evaluation
+    #  catIds     - [all] K cat ids to use for evaluation
+    #  iouThrs    - [.5:.05:.95] T=10 IoU thresholds for evaluation
+    #  recThrs    - [0:.01:1] R=101 recall thresholds for evaluation
+    #  areaRng    - [...] A=4 object area ranges for evaluation
+    #  maxDets    - [1 10 100] M=3 thresholds on max detections per image
+    #  iouType    - ['segm'] set iouType to 'segm', 'bbox' or 'keypoints'
+    #  iouType replaced the now DEPRECATED useSegm parameter.
+    #  useCats    - [1] if true use category labels for evaluation
+    # Note: if useCats=0 category labels are ignored as in proposal scoring.
+    # Note: multiple areaRngs [Ax2] and maxDets [Mx1] can be specified.
+    #
+    # evaluate(): evaluates detections on every image and every category and
+    # concats the results into the "evalImgs" with fields:
+    #  dtIds      - [1xD] id for each of the D detections (dt)
+    #  gtIds      - [1xG] id for each of the G ground truths (gt)
+    #  dtMatches  - [TxD] matching gt id at each IoU or 0
+    #  gtMatches  - [TxG] matching dt id at each IoU or 0
+    #  dtScores   - [1xD] confidence of each dt
+    #  gtIgnore   - [1xG] ignore flag for each gt
+    #  dtIgnore   - [TxD] ignore flag for each dt at each IoU
+    #
+    # accumulate(): accumulates the per-image, per-category evaluation
+    # results in "evalImgs" into the dictionary "eval" with fields:
+    #  params     - parameters used for evaluation
+    #  date       - date evaluation was performed
+    #  counts     - [T,R,K,A,M] parameter dimensions (see above)
+    #  precision  - [TxRxKxAxM] precision for every evaluation setting
+    #  recall     - [TxKxAxM] max recall for every evaluation setting
+    # Note: precision and recall==-1 for settings with no gt objects.
+    #
+    # See also coco, mask, pycocoDemo, pycocoEvalDemo
+    #
+    # Microsoft COCO Toolbox.      version 2.0
+    # Data, paper, and tutorials available at:  http://mscoco.org/
+    # Code written by Piotr Dollar and Tsung-Yi Lin, 2015.
+    # Licensed under the Simplified BSD License [see coco/license.txt]
+    def __init__(self, cocoGt=None, cocoDt=None, iouType='segm'):
+        '''
+        Initialize CocoEval using coco APIs for gt and dt
+        :param cocoGt: coco object with ground truth annotations
+        :param cocoDt: coco object with detection results
+        :return: None
+        '''
+        if not iouType:
+            print('iouType not specified. use default iouType segm')
+        self.cocoGt = cocoGt  # ground truth COCO API
+        self.cocoDt = cocoDt  # detections COCO API
+        self.params = {}  # evaluation parameters
+        self.evalImgs = defaultdict(
+            list)  # per-image per-category evaluation results [KxAxI] elements
+        self.eval = {}  # accumulated evaluation results
+        self._gts = defaultdict(list)  # gt for evaluation
+        self._dts = defaultdict(list)  # dt for evaluation
+        self.params = Params(iouType=iouType)  # parameters
+        self._paramsEval = {}  # parameters for evaluation
+        self.stats = []  # result summarization
+        self.ious = {}  # ious between all gts and dts
+        if not cocoGt is None:
+            self.params.imgIds = sorted(cocoGt.getImgIds())
+            self.params.catIds = sorted(cocoGt.getCatIds())
+
+    def _prepare(self):
+        '''
+        Prepare ._gts and ._dts for evaluation based on params
+        :return: None
+        '''
+        def _toMask(anns, coco):
+            # modify ann['segmentation'] by reference
+            for ann in anns:
+                rle = coco.annToRLE(ann)
+                ann['segmentation'] = rle
+
+        p = self.params
+        if p.useCats:
+            gts = self.cocoGt.loadAnns(
+                self.cocoGt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
+            dts = self.cocoDt.loadAnns(
+                self.cocoDt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
+        else:
+            gts = self.cocoGt.loadAnns(self.cocoGt.getAnnIds(imgIds=p.imgIds))
+            dts = self.cocoDt.loadAnns(self.cocoDt.getAnnIds(imgIds=p.imgIds))
+
+        # convert ground truth to mask if iouType == 'segm'
+        if p.iouType == 'segm':
+            _toMask(gts, self.cocoGt)
+            _toMask(dts, self.cocoDt)
+        # set ignore flag
+        for gt in gts:
+            gt['ignore'] = gt['ignore'] if 'ignore' in gt else 0
+            gt['ignore'] = 'iscrowd' in gt and gt['iscrowd']
+            if p.iouType == 'keypoints':
+                gt['ignore'] = (gt['num_keypoints'] == 0) or gt['ignore']
+        self._gts = defaultdict(list)  # gt for evaluation
+        self._dts = defaultdict(list)  # dt for evaluation
+        for gt in gts:
+            self._gts[gt['image_id'], gt['category_id']].append(gt)
+        for dt in dts:
+            self._dts[dt['image_id'], dt['category_id']].append(dt)
+        self.evalImgs = defaultdict(
+            list)  # per-image per-category evaluation results
+        self.eval = {}  # accumulated evaluation results
+
+    def evaluate(self):
+        '''
+        Run per image evaluation on given images and store results (a list of dict) in self.evalImgs
+        :return: None
+        '''
+        tic = time.time()
+        print('Running per image evaluation...')
+        p = self.params
+        # add backward compatibility if useSegm is specified in params
+        if not p.useSegm is None:
+            p.iouType = 'segm' if p.useSegm == 1 else 'bbox'
+            print('useSegm (deprecated) is not None. Running {} evaluation'.
+                  format(p.iouType))
+        print('Evaluate annotation type *{}*'.format(p.iouType))
+        p.imgIds = list(np.unique(p.imgIds))
+        if p.useCats:
+            p.catIds = list(np.unique(p.catIds))
+        p.maxDets = sorted(p.maxDets)
+        self.params = p
+
+        self._prepare()
+        # loop through images, area range, max detection number
+        catIds = p.catIds if p.useCats else [-1]
+
+        if p.iouType == 'segm' or p.iouType == 'bbox':
+            computeIoU = self.computeIoU
+        elif p.iouType == 'keypoints':
+            computeIoU = self.computeOks
+        self.ious = {(imgId, catId): computeIoU(imgId, catId) \
+                        for imgId in p.imgIds
+                        for catId in catIds}
+
+        evaluateImg = self.evaluateImg
+        maxDet = p.maxDets[-1]
+        self.evalImgs = [
+            evaluateImg(imgId, catId, areaRng, maxDet) for catId in catIds
+            for areaRng in p.areaRng for imgId in p.imgIds
+        ]
+        self._paramsEval = copy.deepcopy(self.params)
+        toc = time.time()
+        print('DONE (t={:0.2f}s).'.format(toc - tic))
+
+    def computeIoU(self, imgId, catId):
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return []
+        inds = np.argsort([-d['score'] for d in dt], kind='mergesort')
+        dt = [dt[i] for i in inds]
+        if len(dt) > p.maxDets[-1]:
+            dt = dt[0:p.maxDets[-1]]
+
+        if p.iouType == 'segm':
+            g = [g['segmentation'] for g in gt]
+            d = [d['segmentation'] for d in dt]
+        elif p.iouType == 'bbox':
+            g = [g['bbox'] for g in gt]
+            d = [d['bbox'] for d in dt]
+        else:
+            raise Exception('unknown iouType for iou computation')
+
+        # compute iou between each dt and gt region
+        iscrowd = [int(o['iscrowd']) for o in gt]
+        ious = iou(d, g, iscrowd)
+        return ious
+
+    def computeOks(self, imgId, catId):
+        p = self.params
+        # dimention here should be Nxm
+        gts = self._gts[imgId, catId]
+        dts = self._dts[imgId, catId]
+        inds = np.argsort([-d['score'] for d in dts], kind='mergesort')
+        dts = [dts[i] for i in inds]
+        if len(dts) > p.maxDets[-1]:
+            dts = dts[0:p.maxDets[-1]]
+        # if len(gts) == 0 and len(dts) == 0:
+        if len(gts) == 0 or len(dts) == 0:
+            return []
+        ious = np.zeros((len(dts), len(gts)))
+        sigmas = np.array([
+            .26, .25, .25, .35, .35, .79, .79, .72, .72, .62, .62, 1.07, 1.07,
+            .87, .87, .89, .89
+        ]) / 10.0
+        vars = (sigmas * 2)**2
+        k = len(sigmas)
+        # compute oks between each detection and ground truth object
+        for j, gt in enumerate(gts):
+            # create bounds for ignore regions(double the gt bbox)
+            g = np.array(gt['keypoints'])
+            xg = g[0::3]
+            yg = g[1::3]
+            vg = g[2::3]
+            k1 = np.count_nonzero(vg > 0)
+            bb = gt['bbox']
+            x0 = bb[0] - bb[2]
+            x1 = bb[0] + bb[2] * 2
+            y0 = bb[1] - bb[3]
+            y1 = bb[1] + bb[3] * 2
+            for i, dt in enumerate(dts):
+                d = np.array(dt['keypoints'])
+                xd = d[0::3]
+                yd = d[1::3]
+                if k1 > 0:
+                    # measure the per-keypoint distance if keypoints visible
+                    dx = xd - xg
+                    dy = yd - yg
+                else:
+                    # measure minimum distance to keypoints in (x0,y0) & (x1,y1)
+                    z = np.zeros((k))
+                    dx = np.max((z, x0 - xd), axis=0) + np.max(
+                        (z, xd - x1), axis=0)
+                    dy = np.max((z, y0 - yd), axis=0) + np.max(
+                        (z, yd - y1), axis=0)
+                e = (dx**2 + dy**2) / vars / (gt['area'] + np.spacing(1)) / 2
+                if k1 > 0:
+                    e = e[vg > 0]
+                ious[i, j] = np.sum(np.exp(-e)) / e.shape[0]
+        return ious
+
+    def evaluateImg(self, imgId, catId, aRng, maxDet):
+        '''
+        perform evaluation for single category and image
+        :return: dict (single image results)
+        '''
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return None
+
+        for g in gt:
+            if g['ignore'] or (g['area'] < aRng[0] or g['area'] > aRng[1]):
+                g['_ignore'] = 1
+            else:
+                g['_ignore'] = 0
+
+        # sort dt highest score first, sort gt ignore last
+        gtind = np.argsort([g['_ignore'] for g in gt], kind='mergesort')
+        gt = [gt[i] for i in gtind]
+        dtind = np.argsort([-d['score'] for d in dt], kind='mergesort')
+        dt = [dt[i] for i in dtind[0:maxDet]]
+        iscrowd = [int(o['iscrowd']) for o in gt]
+        # load computed ious
+        ious = self.ious[imgId, catId][:, gtind] if len(
+            self.ious[imgId, catId]) > 0 else self.ious[imgId, catId]
+
+        T = len(p.iouThrs)
+        G = len(gt)
+        D = len(dt)
+        gtm = np.zeros((T, G))
+        dtm = np.zeros((T, D))
+        gtIg = np.array([g['_ignore'] for g in gt])
+        dtIg = np.zeros((T, D))
+        if not len(ious) == 0:
+            for tind, t in enumerate(p.iouThrs):
+                for dind, d in enumerate(dt):
+                    # information about best match so far (m=-1 -> unmatched)
+                    iou = min([t, 1 - 1e-10])
+                    m = -1
+                    for gind, g in enumerate(gt):
+                        # if this gt already matched, and not a crowd, continue
+                        if gtm[tind, gind] > 0 and not iscrowd[gind]:
+                            continue
+                        # if dt matched to reg gt, and on ignore gt, stop
+                        if m > -1 and gtIg[m] == 0 and gtIg[gind] == 1:
+                            break
+                        # continue to next gt unless better match made
+                        if ious[dind, gind] < iou:
+                            continue
+                        # if match successful and best so far, store appropriately
+                        iou = ious[dind, gind]
+                        m = gind
+                    # if match made store id of match for both dt and gt
+                    if m == -1:
+                        continue
+                    dtIg[tind, dind] = gtIg[m]
+                    dtm[tind, dind] = gt[m]['id']
+                    gtm[tind, m] = d['id']
+        # set unmatched detections outside of area range to ignore
+        a = np.array([d['area'] < aRng[0] or d['area'] > aRng[1]
+                      for d in dt]).reshape((1, len(dt)))
+        dtIg = np.logical_or(dtIg, np.logical_and(dtm == 0, np.repeat(a, T,
+                                                                      0)))
+        # store results for given image and category
+        return {
+            'image_id': imgId,
+            'category_id': catId,
+            'aRng': aRng,
+            'maxDet': maxDet,
+            'dtIds': [d['id'] for d in dt],
+            'gtIds': [g['id'] for g in gt],
+            'dtMatches': dtm,
+            'gtMatches': gtm,
+            'dtScores': [d['score'] for d in dt],
+            'gtIgnore': gtIg,
+            'dtIgnore': dtIg,
+        }
+
+    def accumulate(self, p=None):
+        '''
+        Accumulate per image evaluation results and store the result in self.eval
+        :param p: input params for evaluation
+        :return: None
+        '''
+        print('Accumulating evaluation results...')
+        tic = time.time()
+        if not self.evalImgs:
+            print('Please run evaluate() first')
+        # allows input customized parameters
+        if p is None:
+            p = self.params
+        p.catIds = p.catIds if p.useCats == 1 else [-1]
+        T = len(p.iouThrs)
+        R = len(p.recThrs)
+        K = len(p.catIds) if p.useCats else 1
+        A = len(p.areaRng)
+        M = len(p.maxDets)
+        precision = -np.ones(
+            (T, R, K, A, M))  # -1 for the precision of absent categories
+        recall = -np.ones((T, K, A, M))
+
+        # create dictionary for future indexing
+        _pe = self._paramsEval
+        catIds = _pe.catIds if _pe.useCats else [-1]
+        setK = set(catIds)
+        setA = set(map(tuple, _pe.areaRng))
+        setM = set(_pe.maxDets)
+        setI = set(_pe.imgIds)
+        # get inds to evaluate
+        k_list = [n for n, k in enumerate(p.catIds) if k in setK]
+        m_list = [m for n, m in enumerate(p.maxDets) if m in setM]
+        a_list = [
+            n for n, a in enumerate(map(lambda x: tuple(x), p.areaRng))
+            if a in setA
+        ]
+        i_list = [n for n, i in enumerate(p.imgIds) if i in setI]
+        I0 = len(_pe.imgIds)
+        A0 = len(_pe.areaRng)
+        # retrieve E at each category, area range, and max number of detections
+        for k, k0 in enumerate(k_list):
+            Nk = k0 * A0 * I0
+            for a, a0 in enumerate(a_list):
+                Na = a0 * I0
+                for m, maxDet in enumerate(m_list):
+                    E = [self.evalImgs[Nk + Na + i] for i in i_list]
+                    E = [e for e in E if not e is None]
+                    if len(E) == 0:
+                        continue
+                    dtScores = np.concatenate(
+                        [e['dtScores'][0:maxDet] for e in E])
+
+                    # different sorting method generates slightly different results.
+                    # mergesort is used to be consistent as Matlab implementation.
+                    inds = np.argsort(-dtScores, kind='mergesort')
+
+                    dtm = np.concatenate(
+                        [e['dtMatches'][:, 0:maxDet] for e in E], axis=1)[:,
+                                                                          inds]
+                    dtIg = np.concatenate(
+                        [e['dtIgnore'][:, 0:maxDet] for e in E], axis=1)[:,
+                                                                         inds]
+                    gtIg = np.concatenate([e['gtIgnore'] for e in E])
+                    npig = np.count_nonzero(gtIg == 0)
+                    if npig == 0:
+                        continue
+                    tps = np.logical_and(dtm, np.logical_not(dtIg))
+                    fps = np.logical_and(np.logical_not(dtm),
+                                         np.logical_not(dtIg))
+
+                    tp_sum = np.cumsum(tps, axis=1).astype(dtype=np.float)
+                    fp_sum = np.cumsum(fps, axis=1).astype(dtype=np.float)
+                    for t, (tp, fp) in enumerate(zip(tp_sum, fp_sum)):
+                        tp = np.array(tp)
+                        fp = np.array(fp)
+                        nd = len(tp)
+                        rc = tp / npig
+                        pr = tp / (fp + tp + np.spacing(1))
+                        q = np.zeros((R, ))
+
+                        if nd:
+                            recall[t, k, a, m] = rc[-1]
+                        else:
+                            recall[t, k, a, m] = 0
+
+                        # numpy is slow without cython optimization for accessing elements
+                        # use python array gets significant speed improvement
+                        pr = pr.tolist()
+                        q = q.tolist()
+
+                        for i in range(nd - 1, 0, -1):
+                            if pr[i] > pr[i - 1]:
+                                pr[i - 1] = pr[i]
+
+                        inds = np.searchsorted(rc, p.recThrs, side='left')
+                        try:
+                            for ri, pi in enumerate(inds):
+                                q[ri] = pr[pi]
+                        except:
+                            pass
+                        precision[t, :, k, a, m] = np.array(q)
+        self.eval = {
+            'params': p,
+            'counts': [T, R, K, A, M],
+            'date': datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
+            'precision': precision,
+            'recall': recall,
+        }
+        toc = time.time()
+        print('DONE (t={:0.2f}s).'.format(toc - tic))
+
+    def summarize(self):
+        '''
+        Compute and display summary metrics for evaluation results.
+        Note this functin can *only* be applied on the default parameter setting
+        '''
+        def _summarize(ap=1, iouThr=None, areaRng='all', maxDets=100):
+            p = self.params
+            iStr = ' {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}'
+            titleStr = 'Average Precision' if ap == 1 else 'Average Recall'
+            typeStr = '(AP)' if ap == 1 else '(AR)'
+            iouStr = '{:0.2f}:{:0.2f}'.format(p.iouThrs[0], p.iouThrs[-1]) \
+                if iouThr is None else '{:0.2f}'.format(iouThr)
+
+            aind = [
+                i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng
+            ]
+            mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets]
+            if ap == 1:
+                # dimension of precision: [TxRxKxAxM]
+                s = self.eval['precision']
+                # IoU
+                if iouThr is not None:
+                    t = np.where(iouThr == p.iouThrs)[0]
+                    s = s[t]
+                s = s[:, :, :, aind, mind]
+            else:
+                # dimension of recall: [TxKxAxM]
+                s = self.eval['recall']
+                if iouThr is not None:
+                    t = np.where(iouThr == p.iouThrs)[0]
+                    s = s[t]
+                s = s[:, :, aind, mind]
+            if len(s[s > -1]) == 0:
+                mean_s = -1
+            else:
+                mean_s = np.mean(s[s > -1])
+            print(
+                iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets,
+                            mean_s))
+            return mean_s
+
+        def _summarizeDets():
+            stats = np.zeros((12, ))
+            stats[0] = _summarize(1)
+            stats[1] = _summarize(1, iouThr=.5, maxDets=self.params.maxDets[2])
+            stats[2] = _summarize(1,
+                                  iouThr=.75,
+                                  maxDets=self.params.maxDets[2])
+            stats[3] = _summarize(1,
+                                  areaRng='small',
+                                  maxDets=self.params.maxDets[2])
+            stats[4] = _summarize(1,
+                                  areaRng='medium',
+                                  maxDets=self.params.maxDets[2])
+            stats[5] = _summarize(1,
+                                  areaRng='large',
+                                  maxDets=self.params.maxDets[2])
+            stats[6] = _summarize(0, maxDets=self.params.maxDets[0])
+            stats[7] = _summarize(0, maxDets=self.params.maxDets[1])
+            stats[8] = _summarize(0, maxDets=self.params.maxDets[2])
+            stats[9] = _summarize(0,
+                                  areaRng='small',
+                                  maxDets=self.params.maxDets[2])
+            stats[10] = _summarize(0,
+                                   areaRng='medium',
+                                   maxDets=self.params.maxDets[2])
+            stats[11] = _summarize(0,
+                                   areaRng='large',
+                                   maxDets=self.params.maxDets[2])
+            return stats
+
+        def _summarizeKps():
+            stats = np.zeros((10, ))
+            stats[0] = _summarize(1, maxDets=20)
+            stats[1] = _summarize(1, maxDets=20, iouThr=.5)
+            stats[2] = _summarize(1, maxDets=20, iouThr=.75)
+            stats[3] = _summarize(1, maxDets=20, areaRng='medium')
+            stats[4] = _summarize(1, maxDets=20, areaRng='large')
+            stats[5] = _summarize(0, maxDets=20)
+            stats[6] = _summarize(0, maxDets=20, iouThr=.5)
+            stats[7] = _summarize(0, maxDets=20, iouThr=.75)
+            stats[8] = _summarize(0, maxDets=20, areaRng='medium')
+            stats[9] = _summarize(0, maxDets=20, areaRng='large')
+            return stats
+
+        if not self.eval:
+            raise Exception('Please run accumulate() first')
+        iouType = self.params.iouType
+        if iouType == 'segm' or iouType == 'bbox':
+            summarize = _summarizeDets
+        elif iouType == 'keypoints':
+            summarize = _summarizeKps
+        self.stats = summarize()
+
+    def __str__(self):
+        self.summarize()
+
+
+class Params:
+    '''
+    Params for coco evaluation api
+    '''
+    def setDetParams(self):
+        self.imgIds = []
+        self.catIds = []
+        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
+        self.iouThrs = np.linspace(.5,
+                                   0.95,
+                                   np.round((0.95 - .5) / .05) + 1,
+                                   endpoint=True)
+        self.recThrs = np.linspace(.0,
+                                   1.00,
+                                   np.round((1.00 - .0) / .01) + 1,
+                                   endpoint=True)
+        self.maxDets = [1, 10, 100]
+        self.areaRng = [[0**2, 1e5**2], [0**2, 32**2], [32**2, 96**2],
+                        [96**2, 1e5**2]]
+        self.areaRngLbl = ['all', 'small', 'medium', 'large']
+        self.useCats = 1
+
+    def setKpParams(self):
+        self.imgIds = []
+        self.catIds = []
+        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
+        self.iouThrs = np.linspace(.5,
+                                   0.95,
+                                   np.round((0.95 - .5) / .05) + 1,
+                                   endpoint=True)
+        self.recThrs = np.linspace(.0,
+                                   1.00,
+                                   np.round((1.00 - .0) / .01) + 1,
+                                   endpoint=True)
+        self.maxDets = [20]
+        self.areaRng = [[0**2, 1e5**2], [32**2, 96**2], [96**2, 1e5**2]]
+        self.areaRngLbl = ['all', 'medium', 'large']
+        self.useCats = 1
+
+    def __init__(self, iouType='segm'):
+        if iouType == 'segm' or iouType == 'bbox':
+            self.setDetParams()
+        elif iouType == 'keypoints':
+            self.setKpParams()
+        else:
+            raise Exception('iouType not supported')
+        self.iouType = iouType
+        # useSegm is deprecated
+        self.useSegm = None
diff --git a/RetinaFace/rcnn/pycocotools/mask.py b/detection/retinaface/rcnn/pycocotools/mask.py
similarity index 97%
rename from RetinaFace/rcnn/pycocotools/mask.py
rename to detection/retinaface/rcnn/pycocotools/mask.py
index bba4ee0d5..3a9c14cf0 100644
--- a/RetinaFace/rcnn/pycocotools/mask.py
+++ b/detection/retinaface/rcnn/pycocotools/mask.py
@@ -73,10 +73,11 @@
 # Code written by Piotr Dollar and Tsung-Yi Lin, 2015.
 # Licensed under the Simplified BSD License [see coco/license.txt]
 
-iou         = _mask.iou
-merge       = _mask.merge
+iou = _mask.iou
+merge = _mask.merge
 frPyObjects = _mask.frPyObjects
 
+
 def encode(bimask):
     if len(bimask.shape) == 3:
         return _mask.encode(bimask)
@@ -84,11 +85,13 @@ def encode(bimask):
         h, w = bimask.shape
         return _mask.encode(bimask.reshape((h, w, 1), order='F'))[0]
 
+
 def decode(rleObjs):
     if type(rleObjs) == list:
         return _mask.decode(rleObjs)
     else:
-        return _mask.decode([rleObjs])[:,:,0]
+        return _mask.decode([rleObjs])[:, :, 0]
+
 
 def area(rleObjs):
     if type(rleObjs) == list:
@@ -96,6 +99,7 @@ def area(rleObjs):
     else:
         return _mask.area([rleObjs])[0]
 
+
 def toBbox(rleObjs):
     if type(rleObjs) == list:
         return _mask.toBbox(rleObjs)
diff --git a/RetinaFace/rcnn/pycocotools/maskApi.c b/detection/retinaface/rcnn/pycocotools/maskApi.c
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/maskApi.c
rename to detection/retinaface/rcnn/pycocotools/maskApi.c
diff --git a/RetinaFace/rcnn/pycocotools/maskApi.h b/detection/retinaface/rcnn/pycocotools/maskApi.h
similarity index 100%
rename from RetinaFace/rcnn/pycocotools/maskApi.h
rename to detection/retinaface/rcnn/pycocotools/maskApi.h
diff --git a/detection/retinaface/rcnn/pycocotools/setup.py b/detection/retinaface/rcnn/pycocotools/setup.py
new file mode 100644
index 000000000..d1e0d2e89
--- /dev/null
+++ b/detection/retinaface/rcnn/pycocotools/setup.py
@@ -0,0 +1,18 @@
+from distutils.core import setup
+from Cython.Build import cythonize
+from distutils.extension import Extension
+import numpy as np
+
+# To compile and install locally run "python setup.py build_ext --inplace"
+# To install library to Python site-packages run "python setup.py build_ext install"
+
+ext_modules = [
+    Extension(
+        '_mask',
+        sources=['maskApi.c', '_mask.pyx'],
+        include_dirs=[np.get_include()],
+        extra_compile_args=['-Wno-cpp', '-Wno-unused-function', '-std=c99'],
+    )
+]
+
+setup(name='pycocotools', ext_modules=cythonize(ext_modules))
diff --git a/detection/retinaface/rcnn/sample_config.py b/detection/retinaface/rcnn/sample_config.py
new file mode 100644
index 000000000..88fff9f04
--- /dev/null
+++ b/detection/retinaface/rcnn/sample_config.py
@@ -0,0 +1,351 @@
+import numpy as np
+from easydict import EasyDict as edict
+
+config = edict()
+
+# network related params
+config.PIXEL_MEANS = np.array([103.939, 116.779, 123.68])
+config.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
+config.PIXEL_SCALE = 1.0
+config.IMAGE_STRIDE = 0
+
+# dataset related params
+config.NUM_CLASSES = 2
+config.PRE_SCALES = [(1200, 1600)
+                     ]  # first is scale (the shorter side); second is max size
+config.SCALES = [(640, 640)
+                 ]  # first is scale (the shorter side); second is max size
+#config.SCALES = [(800, 800)]  # first is scale (the shorter side); second is max size
+config.ORIGIN_SCALE = False
+
+_ratio = (1., )
+
+RAC_SSH = {
+    '32': {
+        'SCALES': (32, 16),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '16': {
+        'SCALES': (8, 4),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '8': {
+        'SCALES': (2, 1),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+}
+
+_ratio = (1., 1.5)
+RAC_SSH2 = {
+    '32': {
+        'SCALES': (32, 16),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '16': {
+        'SCALES': (8, 4),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '8': {
+        'SCALES': (2, 1),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+}
+
+_ratio = (1., 1.5)
+RAC_SSH3 = {
+    '32': {
+        'SCALES': (32, 16),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '16': {
+        'SCALES': (8, 4),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '8': {
+        'SCALES': (2, 1),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+    '4': {
+        'SCALES': (2, 1),
+        'BASE_SIZE': 16,
+        'RATIOS': _ratio,
+        'ALLOWED_BORDER': 9999
+    },
+}
+
+RAC_RETINA = {}
+_ratios = (1.0, )
+_ass = 2.0**(1.0 / 3)
+_basescale = 1.0
+for _stride in [4, 8, 16, 32, 64]:
+    key = str(_stride)
+    value = {'BASE_SIZE': 16, 'RATIOS': _ratios, 'ALLOWED_BORDER': 9999}
+    scales = []
+    for _ in range(3):
+        scales.append(_basescale)
+        _basescale *= _ass
+    value['SCALES'] = tuple(scales)
+    RAC_RETINA[key] = value
+
+config.RPN_ANCHOR_CFG = RAC_SSH  #default
+
+config.NET_MODE = 2
+config.HEAD_MODULE = 'SSH'
+#config.HEAD_MODULE = 'RF'
+config.LR_MODE = 0
+config.LANDMARK_LR_MULT = 2.0
+config.HEAD_FILTER_NUM = 256
+config.CONTEXT_FILTER_RATIO = 1
+config.max_feat_channel = 9999
+
+config.USE_CROP = True
+config.USE_FPN = True
+config.USE_DCN = 0
+config.FACE_LANDMARK = True
+config.USE_OCCLUSION = False
+config.USE_BLUR = False
+config.MORE_SMALL_BOX = True
+
+config.LAYER_FIX = False
+
+config.CASCADE = 0
+config.CASCADE_MODE = 1
+#config.CASCADE_CLS_STRIDES = [16,8,4]
+#config.CASCADE_BBOX_STRIDES = [64,32]
+config.CASCADE_CLS_STRIDES = [64, 32, 16, 8, 4]
+config.CASCADE_BBOX_STRIDES = [64, 32, 16, 8, 4]
+#config.CASCADE_BBOX_STRIDES = [64,32,16,8]
+
+config.HEAD_BOX = False
+config.DENSE_ANCHOR = False
+config.USE_MAXOUT = 0
+config.SHARE_WEIGHT_BBOX = False
+config.SHARE_WEIGHT_LANDMARK = False
+
+config.RANDOM_FEAT_STRIDE = False
+config.NUM_CPU = 4
+config.MIXUP = 0.0
+config.USE_3D = False
+
+#config.BBOX_MASK_THRESH = 0
+config.COLOR_MODE = 2
+config.COLOR_JITTERING = 0.125
+#config.COLOR_JITTERING = 0
+#config.COLOR_JITTERING = 0.2
+
+config.TRAIN = edict()
+
+config.TRAIN.IMAGE_ALIGN = 0
+config.TRAIN.MIN_BOX_SIZE = 0
+config.BBOX_MASK_THRESH = config.TRAIN.MIN_BOX_SIZE
+# R-CNN and RPN
+# size of images for each device, 2 for rcnn, 1 for rpn and e2e
+config.TRAIN.BATCH_IMAGES = 8
+# e2e changes behavior of anchor loader and metric
+config.TRAIN.END2END = True
+# group images with similar aspect ratio
+config.TRAIN.ASPECT_GROUPING = False
+
+# RPN anchor loader
+# rpn anchors batch size
+config.TRAIN.RPN_ENABLE_OHEM = 2
+config.TRAIN.OHEM_MODE = 1
+config.TRAIN.RPN_BATCH_SIZE = 256
+# rpn anchors sampling params
+config.TRAIN.RPN_FG_FRACTION = 0.25
+config.TRAIN.RPN_POSITIVE_OVERLAP = 0.5
+config.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3
+if config.CASCADE > 0:
+    config.TRAIN.RPN_POSITIVE_OVERLAP = 0.7
+config.TRAIN.CASCADE_OVERLAP = [0.4, 0.5]
+config.TRAIN.RPN_CLOBBER_POSITIVES = False
+config.TRAIN.RPN_FORCE_POSITIVE = False
+# rpn bounding box regression params
+config.TRAIN.BBOX_STDS = (1.0, 1.0, 1.0, 1.0)
+config.TRAIN.LANDMARK_STD = 1.0
+
+config.TEST = edict()
+
+# R-CNN testing
+# use rpn to generate proposal
+config.TEST.HAS_RPN = False
+# size of images for each device
+config.TEST.BATCH_IMAGES = 1
+
+# RPN proposal
+config.TEST.CXX_PROPOSAL = True
+config.TEST.RPN_NMS_THRESH = 0.3
+config.TEST.RPN_PRE_NMS_TOP_N = 1000
+config.TEST.RPN_POST_NMS_TOP_N = 3000
+#config.TEST.RPN_MIN_SIZE = config.RPN_FEAT_STRIDE
+#config.TEST.RPN_MIN_SIZE = [0,0,0]
+
+# RCNN nms
+config.TEST.NMS = 0.3
+
+config.TEST.SCORE_THRESH = 0.05
+config.TEST.IOU_THRESH = 0.5
+
+# network settings
+network = edict()
+
+network.ssh = edict()
+
+network.mnet = edict()
+#network.mnet.pretrained = 'model/mnasnet'
+#network.mnet.pretrained = 'model/mobilenetv2_0_5'
+#network.mnet.pretrained = 'model/mobilenet_0_5'
+#network.mnet.MULTIPLIER = 0.5
+#network.mnet.pretrained = 'model/mobilenet_0_25'
+#network.mnet.pretrained_epoch = 0
+#network.mnet.PIXEL_MEANS = np.array([0.406, 0.456, 0.485])
+#network.mnet.PIXEL_STDS = np.array([0.225, 0.224, 0.229])
+#network.mnet.PIXEL_SCALE = 255.0
+network.mnet.FIXED_PARAMS = ['^stage1', '^.*upsampling']
+network.mnet.BATCH_IMAGES = 16
+network.mnet.HEAD_FILTER_NUM = 64
+network.mnet.CONTEXT_FILTER_RATIO = 1
+
+network.mnet.PIXEL_MEANS = np.array([0.0, 0.0, 0.0])
+network.mnet.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
+network.mnet.PIXEL_SCALE = 1.0
+#network.mnet.pretrained = 'model/mobilenetfd_0_25' #78
+#network.mnet.pretrained = 'model/mobilenetfd2' #75
+network.mnet.pretrained = 'model/mobilenet025fd0'  #78
+#network.mnet.pretrained = 'model/mobilenet025fd1' #75
+#network.mnet.pretrained = 'model/mobilenet025fd2' #
+network.mnet.pretrained_epoch = 0
+network.mnet.max_feat_channel = 8888
+network.mnet.COLOR_MODE = 1
+network.mnet.USE_CROP = True
+network.mnet.RPN_ANCHOR_CFG = RAC_SSH
+network.mnet.LAYER_FIX = True
+network.mnet.LANDMARK_LR_MULT = 2.5
+
+network.resnet = edict()
+#network.resnet.pretrained = 'model/ResNet50_v1d'
+#network.resnet.pretrained = 'model/resnet-50'
+network.resnet.pretrained = 'model/resnet-152'
+#network.resnet.pretrained = 'model/senet154'
+#network.resnet.pretrained = 'model/densenet161'
+network.resnet.pretrained_epoch = 0
+#network.mnet.PIXEL_MEANS = np.array([103.939, 116.779, 123.68])
+#network.mnet.PIXEL_STDS = np.array([57.375, 57.12, 58.393])
+#network.resnet.PIXEL_MEANS = np.array([0.406, 0.456, 0.485])
+#network.resnet.PIXEL_STDS = np.array([0.225, 0.224, 0.229])
+#network.resnet.PIXEL_SCALE = 255.0
+network.resnet.lr_step = '1,2,3,4,5,55,68,80'
+network.resnet.lr = 0.001
+network.resnet.PIXEL_MEANS = np.array([0.0, 0.0, 0.0])
+network.resnet.PIXEL_STDS = np.array([1.0, 1.0, 1.0])
+network.resnet.PIXEL_SCALE = 1.0
+network.resnet.FIXED_PARAMS = ['^stage1', '^.*upsampling']
+network.resnet.BATCH_IMAGES = 8
+network.resnet.HEAD_FILTER_NUM = 256
+network.resnet.CONTEXT_FILTER_RATIO = 1
+network.resnet.USE_DCN = 2
+network.resnet.RPN_BATCH_SIZE = 256
+network.resnet.RPN_ANCHOR_CFG = RAC_RETINA
+
+network.resnet.USE_DCN = 0
+network.resnet.pretrained = 'model/resnet-50'
+network.resnet.RPN_ANCHOR_CFG = RAC_SSH
+
+# dataset settings
+dataset = edict()
+
+dataset.widerface = edict()
+dataset.widerface.dataset = 'widerface'
+dataset.widerface.image_set = 'train'
+dataset.widerface.test_image_set = 'val'
+dataset.widerface.root_path = 'data'
+dataset.widerface.dataset_path = 'data/widerface'
+dataset.widerface.NUM_CLASSES = 2
+
+dataset.retinaface = edict()
+dataset.retinaface.dataset = 'retinaface'
+dataset.retinaface.image_set = 'train'
+dataset.retinaface.test_image_set = 'val'
+dataset.retinaface.root_path = 'data'
+dataset.retinaface.dataset_path = 'data/retinaface'
+dataset.retinaface.NUM_CLASSES = 2
+
+# default settings
+default = edict()
+
+config.FIXED_PARAMS = ['^conv1', '^conv2', '^conv3', '^.*upsampling']
+#config.FIXED_PARAMS = ['^.*upsampling']
+#config.FIXED_PARAMS = ['^conv1', '^conv2', '^conv3']
+#config.FIXED_PARAMS = ['^conv0', '^stage1', 'gamma', 'beta']  #for resnet
+
+# default network
+default.network = 'resnet'
+default.pretrained = 'model/resnet-152'
+#default.network = 'resnetssh'
+default.pretrained_epoch = 0
+# default dataset
+default.dataset = 'retinaface'
+default.image_set = 'train'
+default.test_image_set = 'val'
+default.root_path = 'data'
+default.dataset_path = 'data/retinaface'
+# default training
+default.frequent = 20
+default.kvstore = 'device'
+# default e2e
+default.prefix = 'model/retinaface'
+default.end_epoch = 10000
+default.lr_step = '55,68,80'
+default.lr = 0.01
+default.wd = 0.0005
+
+
+def generate_config(_network, _dataset):
+    for k, v in network[_network].items():
+        if k in config:
+            config[k] = v
+        elif k in default:
+            default[k] = v
+        if k in config.TRAIN:
+            config.TRAIN[k] = v
+    for k, v in dataset[_dataset].items():
+        if k in config:
+            config[k] = v
+        elif k in default:
+            default[k] = v
+        if k in config.TRAIN:
+            config.TRAIN[k] = v
+    config.network = _network
+    config.dataset = _dataset
+    config.RPN_FEAT_STRIDE = []
+    num_anchors = []
+    for k in config.RPN_ANCHOR_CFG:
+        config.RPN_FEAT_STRIDE.append(int(k))
+        _num_anchors = len(config.RPN_ANCHOR_CFG[k]['SCALES']) * len(
+            config.RPN_ANCHOR_CFG[k]['RATIOS'])
+        if config.DENSE_ANCHOR:
+            _num_anchors *= 2
+        config.RPN_ANCHOR_CFG[k]['NUM_ANCHORS'] = _num_anchors
+        num_anchors.append(_num_anchors)
+    config.RPN_FEAT_STRIDE = sorted(config.RPN_FEAT_STRIDE, reverse=True)
+    for j in range(1, len(num_anchors)):
+        assert num_anchors[0] == num_anchors[j]
+    config.NUM_ANCHORS = num_anchors[0]
diff --git a/RetinaFace/rcnn/symbol/__init__.py b/detection/retinaface/rcnn/symbol/__init__.py
similarity index 100%
rename from RetinaFace/rcnn/symbol/__init__.py
rename to detection/retinaface/rcnn/symbol/__init__.py
diff --git a/detection/retinaface/rcnn/symbol/pyramidbox.py b/detection/retinaface/rcnn/symbol/pyramidbox.py
new file mode 100644
index 000000000..64ae7ce94
--- /dev/null
+++ b/detection/retinaface/rcnn/symbol/pyramidbox.py
@@ -0,0 +1,489 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import six
+import paddle.fluid as fluid
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.initializer import Xavier
+from paddle.fluid.initializer import Constant
+from paddle.fluid.initializer import Bilinear
+from paddle.fluid.regularizer import L2Decay
+
+
+def conv_bn(input,
+            filter,
+            ksize,
+            stride,
+            padding,
+            act='relu',
+            bias_attr=False):
+    conv = fluid.layers.conv2d(input=input,
+                               filter_size=ksize,
+                               num_filters=filter,
+                               stride=stride,
+                               padding=padding,
+                               act=None,
+                               bias_attr=bias_attr)
+    return fluid.layers.batch_norm(input=conv, act=act)
+
+
+def conv_block(input, groups, filters, ksizes, strides=None, with_pool=True):
+    assert len(filters) == groups
+    assert len(ksizes) == groups
+    strides = [1] * groups if strides is None else strides
+    w_attr = ParamAttr(learning_rate=1., initializer=Xavier())
+    b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
+    conv = input
+    for i in six.moves.xrange(groups):
+        conv = fluid.layers.conv2d(input=conv,
+                                   num_filters=filters[i],
+                                   filter_size=ksizes[i],
+                                   stride=strides[i],
+                                   padding=(ksizes[i] - 1) // 2,
+                                   param_attr=w_attr,
+                                   bias_attr=b_attr,
+                                   act='relu')
+    if with_pool:
+        pool = fluid.layers.pool2d(input=conv,
+                                   pool_size=2,
+                                   pool_type='max',
+                                   pool_stride=2,
+                                   ceil_mode=True)
+        return conv, pool
+    else:
+        return conv
+
+
+class PyramidBox(object):
+    def __init__(self,
+                 data_shape,
+                 num_classes=None,
+                 use_transposed_conv2d=True,
+                 is_infer=False,
+                 sub_network=False):
+        """
+        TODO(qingqing): add comments.
+        """
+        self.data_shape = data_shape
+        self.min_sizes = [16., 32., 64., 128., 256., 512.]
+        self.steps = [4., 8., 16., 32., 64., 128.]
+        self.num_classes = num_classes
+        self.use_transposed_conv2d = use_transposed_conv2d
+        self.is_infer = is_infer
+        self.sub_network = sub_network
+
+        # the base network is VGG with atrous layers
+        self._input()
+        self._vgg()
+        if sub_network:
+            self._low_level_fpn()
+            self._cpm_module()
+            self._pyramidbox()
+        else:
+            self._vgg_ssd()
+
+    def feeds(self):
+        if self.is_infer:
+            return [self.image]
+        else:
+            return [self.image, self.face_box, self.head_box, self.gt_label]
+
+    def _input(self):
+        self.image = fluid.layers.data(name='image',
+                                       shape=self.data_shape,
+                                       dtype='float32')
+        if not self.is_infer:
+            self.face_box = fluid.layers.data(name='face_box',
+                                              shape=[4],
+                                              dtype='float32',
+                                              lod_level=1)
+            self.head_box = fluid.layers.data(name='head_box',
+                                              shape=[4],
+                                              dtype='float32',
+                                              lod_level=1)
+            self.gt_label = fluid.layers.data(name='gt_label',
+                                              shape=[1],
+                                              dtype='int32',
+                                              lod_level=1)
+
+    def _vgg(self):
+        self.conv1, self.pool1 = conv_block(self.image, 2, [64] * 2, [3] * 2)
+        self.conv2, self.pool2 = conv_block(self.pool1, 2, [128] * 2, [3] * 2)
+
+        #priorbox min_size is 16
+        self.conv3, self.pool3 = conv_block(self.pool2, 3, [256] * 3, [3] * 3)
+        #priorbox min_size is 32
+        self.conv4, self.pool4 = conv_block(self.pool3, 3, [512] * 3, [3] * 3)
+        #priorbox min_size is 64
+        self.conv5, self.pool5 = conv_block(self.pool4, 3, [512] * 3, [3] * 3)
+
+        # fc6 and fc7 in paper, priorbox min_size is 128
+        self.conv6 = conv_block(self.pool5,
+                                2, [1024, 1024], [3, 1],
+                                with_pool=False)
+        # conv6_1 and conv6_2 in paper, priorbox min_size is 256
+        self.conv7 = conv_block(self.conv6,
+                                2, [256, 512], [1, 3], [1, 2],
+                                with_pool=False)
+        # conv7_1 and conv7_2 in paper, priorbox mini_size is 512
+        self.conv8 = conv_block(self.conv7,
+                                2, [128, 256], [1, 3], [1, 2],
+                                with_pool=False)
+
+    def _low_level_fpn(self):
+        """
+        Low-level feature pyramid network.
+        """
+        def fpn(up_from, up_to):
+            ch = up_to.shape[1]
+            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
+            conv1 = fluid.layers.conv2d(up_from,
+                                        ch,
+                                        1,
+                                        act='relu',
+                                        bias_attr=b_attr)
+            if self.use_transposed_conv2d:
+                w_attr = ParamAttr(learning_rate=0.,
+                                   regularizer=L2Decay(0.),
+                                   initializer=Bilinear())
+                upsampling = fluid.layers.conv2d_transpose(conv1,
+                                                           ch,
+                                                           output_size=None,
+                                                           filter_size=4,
+                                                           padding=1,
+                                                           stride=2,
+                                                           groups=ch,
+                                                           param_attr=w_attr,
+                                                           bias_attr=False,
+                                                           use_cudnn=True)
+            else:
+                upsampling = fluid.layers.resize_bilinear(
+                    conv1, out_shape=up_to.shape[2:])
+
+            conv2 = fluid.layers.conv2d(up_to,
+                                        ch,
+                                        1,
+                                        act='relu',
+                                        bias_attr=b_attr)
+            if self.is_infer:
+                upsampling = fluid.layers.crop(upsampling, shape=conv2)
+            # eltwise mul
+            conv_fuse = upsampling * conv2
+            return conv_fuse
+
+        self.lfpn2_on_conv5 = fpn(self.conv6, self.conv5)
+        self.lfpn1_on_conv4 = fpn(self.lfpn2_on_conv5, self.conv4)
+        self.lfpn0_on_conv3 = fpn(self.lfpn1_on_conv4, self.conv3)
+
+    def _cpm_module(self):
+        """
+        Context-sensitive Prediction Module 
+        """
+        def cpm(input):
+            # residual
+            branch1 = conv_bn(input, 1024, 1, 1, 0, None)
+            branch2a = conv_bn(input, 256, 1, 1, 0, act='relu')
+            branch2b = conv_bn(branch2a, 256, 3, 1, 1, act='relu')
+            branch2c = conv_bn(branch2b, 1024, 1, 1, 0, None)
+            sum = branch1 + branch2c
+            rescomb = fluid.layers.relu(x=sum)
+
+            # ssh
+            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
+            ssh_1 = fluid.layers.conv2d(rescomb,
+                                        256,
+                                        3,
+                                        1,
+                                        1,
+                                        bias_attr=b_attr)
+            ssh_dimred = fluid.layers.conv2d(rescomb,
+                                             128,
+                                             3,
+                                             1,
+                                             1,
+                                             act='relu',
+                                             bias_attr=b_attr)
+            ssh_2 = fluid.layers.conv2d(ssh_dimred,
+                                        128,
+                                        3,
+                                        1,
+                                        1,
+                                        bias_attr=b_attr)
+            ssh_3a = fluid.layers.conv2d(ssh_dimred,
+                                         128,
+                                         3,
+                                         1,
+                                         1,
+                                         act='relu',
+                                         bias_attr=b_attr)
+            ssh_3b = fluid.layers.conv2d(ssh_3a,
+                                         128,
+                                         3,
+                                         1,
+                                         1,
+                                         bias_attr=b_attr)
+
+            ssh_concat = fluid.layers.concat([ssh_1, ssh_2, ssh_3b], axis=1)
+            ssh_out = fluid.layers.relu(x=ssh_concat)
+            return ssh_out
+
+        self.ssh_conv3 = cpm(self.lfpn0_on_conv3)
+        self.ssh_conv4 = cpm(self.lfpn1_on_conv4)
+        self.ssh_conv5 = cpm(self.lfpn2_on_conv5)
+        self.ssh_conv6 = cpm(self.conv6)
+        self.ssh_conv7 = cpm(self.conv7)
+        self.ssh_conv8 = cpm(self.conv8)
+
+    def _l2_norm_scale(self, input, init_scale=1.0, channel_shared=False):
+        from paddle.fluid.layer_helper import LayerHelper
+        helper = LayerHelper("Scale")
+        l2_norm = fluid.layers.l2_normalize(input,
+                                            axis=1)  # l2 norm along channel
+        shape = [1] if channel_shared else [input.shape[1]]
+        scale = helper.create_parameter(
+            attr=helper.param_attr,
+            shape=shape,
+            dtype=input.dtype,
+            default_initializer=Constant(init_scale))
+        out = fluid.layers.elementwise_mul(x=l2_norm,
+                                           y=scale,
+                                           axis=-1 if channel_shared else 1)
+        return out
+
+    def _pyramidbox(self):
+        """
+        Get prior-boxes and pyramid-box
+        """
+        self.ssh_conv3_norm = self._l2_norm_scale(self.ssh_conv3,
+                                                  init_scale=10.)
+        self.ssh_conv4_norm = self._l2_norm_scale(self.ssh_conv4,
+                                                  init_scale=8.)
+        self.ssh_conv5_norm = self._l2_norm_scale(self.ssh_conv5,
+                                                  init_scale=5.)
+
+        def permute_and_reshape(input, last_dim):
+            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
+            compile_shape = [
+                trans.shape[0],
+                np.prod(trans.shape[1:]) // last_dim, last_dim
+            ]
+            run_shape = fluid.layers.assign(
+                np.array([0, -1, last_dim]).astype("int32"))
+            return fluid.layers.reshape(trans,
+                                        shape=compile_shape,
+                                        actual_shape=run_shape)
+
+        face_locs, face_confs = [], []
+        head_locs, head_confs = [], []
+        boxes, vars = [], []
+        inputs = [
+            self.ssh_conv3_norm, self.ssh_conv4_norm, self.ssh_conv5_norm,
+            self.ssh_conv6, self.ssh_conv7, self.ssh_conv8
+        ]
+        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
+        for i, input in enumerate(inputs):
+            mbox_loc = fluid.layers.conv2d(input, 8, 3, 1, 1, bias_attr=b_attr)
+            face_loc, head_loc = fluid.layers.split(mbox_loc,
+                                                    num_or_sections=2,
+                                                    dim=1)
+            face_loc = permute_and_reshape(face_loc, 4)
+            head_loc = permute_and_reshape(head_loc, 4)
+
+            mbox_conf = fluid.layers.conv2d(input,
+                                            6,
+                                            3,
+                                            1,
+                                            1,
+                                            bias_attr=b_attr)
+            face_conf1, face_conf3, head_conf = fluid.layers.split(
+                mbox_conf, num_or_sections=[1, 3, 2], dim=1)
+            face_conf3_maxin = fluid.layers.reduce_max(face_conf3,
+                                                       dim=1,
+                                                       keep_dim=True)
+            face_conf = fluid.layers.concat([face_conf1, face_conf3_maxin],
+                                            axis=1)
+
+            face_conf = permute_and_reshape(face_conf, 2)
+            head_conf = permute_and_reshape(head_conf, 2)
+
+            face_locs.append(face_loc)
+            face_confs.append(face_conf)
+
+            head_locs.append(head_loc)
+            head_confs.append(head_conf)
+
+            box, var = fluid.layers.prior_box(input,
+                                              self.image,
+                                              min_sizes=[self.min_sizes[i]],
+                                              steps=[self.steps[i]] * 2,
+                                              aspect_ratios=[1.],
+                                              clip=False,
+                                              flip=True,
+                                              offset=0.5)
+            box = fluid.layers.reshape(box, shape=[-1, 4])
+            var = fluid.layers.reshape(var, shape=[-1, 4])
+
+            boxes.append(box)
+            vars.append(var)
+
+        self.face_mbox_loc = fluid.layers.concat(face_locs, axis=1)
+        self.face_mbox_conf = fluid.layers.concat(face_confs, axis=1)
+
+        self.head_mbox_loc = fluid.layers.concat(head_locs, axis=1)
+        self.head_mbox_conf = fluid.layers.concat(head_confs, axis=1)
+
+        self.prior_boxes = fluid.layers.concat(boxes)
+        self.box_vars = fluid.layers.concat(vars)
+
+    def _vgg_ssd(self):
+        self.conv3_norm = self._l2_norm_scale(self.conv3, init_scale=10.)
+        self.conv4_norm = self._l2_norm_scale(self.conv4, init_scale=8.)
+        self.conv5_norm = self._l2_norm_scale(self.conv5, init_scale=5.)
+
+        def permute_and_reshape(input, last_dim):
+            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
+            compile_shape = [
+                trans.shape[0],
+                np.prod(trans.shape[1:]) // last_dim, last_dim
+            ]
+            run_shape = fluid.layers.assign(
+                np.array([0, -1, last_dim]).astype("int32"))
+            return fluid.layers.reshape(trans,
+                                        shape=compile_shape,
+                                        actual_shape=run_shape)
+
+        locs, confs = [], []
+        boxes, vars = [], []
+        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
+
+        # conv3
+        mbox_loc = fluid.layers.conv2d(self.conv3_norm,
+                                       4,
+                                       3,
+                                       1,
+                                       1,
+                                       bias_attr=b_attr)
+        loc = permute_and_reshape(mbox_loc, 4)
+        mbox_conf = fluid.layers.conv2d(self.conv3_norm,
+                                        4,
+                                        3,
+                                        1,
+                                        1,
+                                        bias_attr=b_attr)
+        conf1, conf3 = fluid.layers.split(mbox_conf,
+                                          num_or_sections=[1, 3],
+                                          dim=1)
+        conf3_maxin = fluid.layers.reduce_max(conf3, dim=1, keep_dim=True)
+        conf = fluid.layers.concat([conf1, conf3_maxin], axis=1)
+        conf = permute_and_reshape(conf, 2)
+        box, var = fluid.layers.prior_box(self.conv3_norm,
+                                          self.image,
+                                          min_sizes=[16.],
+                                          steps=[4, 4],
+                                          aspect_ratios=[1.],
+                                          clip=False,
+                                          flip=True,
+                                          offset=0.5)
+        box = fluid.layers.reshape(box, shape=[-1, 4])
+        var = fluid.layers.reshape(var, shape=[-1, 4])
+
+        locs.append(loc)
+        confs.append(conf)
+        boxes.append(box)
+        vars.append(var)
+
+        min_sizes = [32., 64., 128., 256., 512.]
+        steps = [8., 16., 32., 64., 128.]
+        inputs = [
+            self.conv4_norm, self.conv5_norm, self.conv6, self.conv7,
+            self.conv8
+        ]
+        for i, input in enumerate(inputs):
+            mbox_loc = fluid.layers.conv2d(input, 4, 3, 1, 1, bias_attr=b_attr)
+            loc = permute_and_reshape(mbox_loc, 4)
+
+            mbox_conf = fluid.layers.conv2d(input,
+                                            2,
+                                            3,
+                                            1,
+                                            1,
+                                            bias_attr=b_attr)
+            conf = permute_and_reshape(mbox_conf, 2)
+            box, var = fluid.layers.prior_box(input,
+                                              self.image,
+                                              min_sizes=[min_sizes[i]],
+                                              steps=[steps[i]] * 2,
+                                              aspect_ratios=[1.],
+                                              clip=False,
+                                              flip=True,
+                                              offset=0.5)
+            box = fluid.layers.reshape(box, shape=[-1, 4])
+            var = fluid.layers.reshape(var, shape=[-1, 4])
+
+            locs.append(loc)
+            confs.append(conf)
+            boxes.append(box)
+            vars.append(var)
+
+        self.face_mbox_loc = fluid.layers.concat(locs, axis=1)
+        self.face_mbox_conf = fluid.layers.concat(confs, axis=1)
+        self.prior_boxes = fluid.layers.concat(boxes)
+        self.box_vars = fluid.layers.concat(vars)
+
+    def vgg_ssd_loss(self):
+        loss = fluid.layers.ssd_loss(self.face_mbox_loc,
+                                     self.face_mbox_conf,
+                                     self.face_box,
+                                     self.gt_label,
+                                     self.prior_boxes,
+                                     self.box_vars,
+                                     overlap_threshold=0.35,
+                                     neg_overlap=0.35)
+        loss = fluid.layers.reduce_sum(loss)
+        return loss
+
+    def train(self):
+        face_loss = fluid.layers.ssd_loss(self.face_mbox_loc,
+                                          self.face_mbox_conf,
+                                          self.face_box,
+                                          self.gt_label,
+                                          self.prior_boxes,
+                                          self.box_vars,
+                                          overlap_threshold=0.35,
+                                          neg_overlap=0.35)
+        face_loss.persistable = True
+        head_loss = fluid.layers.ssd_loss(self.head_mbox_loc,
+                                          self.head_mbox_conf,
+                                          self.head_box,
+                                          self.gt_label,
+                                          self.prior_boxes,
+                                          self.box_vars,
+                                          overlap_threshold=0.35,
+                                          neg_overlap=0.35)
+        head_loss.persistable = True
+        face_loss = fluid.layers.reduce_sum(face_loss)
+        face_loss.persistable = True
+        head_loss = fluid.layers.reduce_sum(head_loss)
+        head_loss.persistable = True
+        total_loss = face_loss + head_loss
+        total_loss.persistable = True
+        return face_loss, head_loss, total_loss
+
+    def infer(self, main_program=None):
+        if main_program is None:
+            test_program = fluid.default_main_program().clone(for_test=True)
+        else:
+            test_program = main_program.clone(for_test=True)
+        with fluid.program_guard(test_program):
+            face_nmsed_out = fluid.layers.detection_output(
+                self.face_mbox_loc,
+                self.face_mbox_conf,
+                self.prior_boxes,
+                self.box_vars,
+                nms_threshold=0.3,
+                nms_top_k=5000,
+                keep_top_k=750,
+                score_threshold=0.01)
+        return test_program, face_nmsed_out
diff --git a/detection/retinaface/rcnn/symbol/symbol_common.py b/detection/retinaface/rcnn/symbol/symbol_common.py
new file mode 100644
index 000000000..1343dd963
--- /dev/null
+++ b/detection/retinaface/rcnn/symbol/symbol_common.py
@@ -0,0 +1,757 @@
+import mxnet as mx
+import mxnet.ndarray as nd
+import numpy as np
+from rcnn.config import config
+from rcnn.PY_OP import rpn_fpn_ohem3, cascade_refine
+
+PREFIX = 'RF'
+F1 = 0
+F2 = 0
+_bwm = 1.0
+
+def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), bias_wd_mult=0.0, shared_weight=None, shared_bias = None):
+    if shared_weight is None:
+        weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                    init=mx.init.Normal(0.01),
+                                    attr={'__lr_mult__': '1.0'})
+        bias = mx.symbol.Variable(name="{}_bias".format(name),
+                                  init=mx.init.Constant(0.0),
+                                  attr={
+                                      '__lr_mult__': '2.0',
+                                      '__wd_mult__': str(bias_wd_mult)
+                                  })
+    else:
+        weight = shared_weight
+        bias = shared_bias
+        print('reuse shared var in', name)
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
+    return conv
+
+
+def conv_deformable(net, num_filter, num_group=1, act_type='relu', name=''):
+    if config.USE_DCN == 1:
+        f = num_group * 18
+        conv_offset = mx.symbol.Convolution(name=name + '_conv_offset',
+                                            data=net,
+                                            num_filter=f,
+                                            pad=(1, 1),
+                                            kernel=(3, 3),
+                                            stride=(1, 1))
+        net = mx.contrib.symbol.DeformableConvolution(
+            name=name + "_conv",
+            data=net,
+            offset=conv_offset,
+            num_filter=num_filter,
+            pad=(1, 1),
+            kernel=(3, 3),
+            num_deformable_group=num_group,
+            stride=(1, 1),
+            no_bias=False)
+    else:
+        print('use dcnv2 at', name)
+        lr_mult = 0.1
+        weight_var = mx.sym.Variable(name=name + '_conv2_offset_weight',
+                                     init=mx.init.Zero(),
+                                     lr_mult=lr_mult)
+        bias_var = mx.sym.Variable(name=name + '_conv2_offset_bias',
+                                   init=mx.init.Zero(),
+                                   lr_mult=lr_mult)
+        conv2_offset = mx.symbol.Convolution(name=name + '_conv2_offset',
+                                             data=net,
+                                             num_filter=27,
+                                             pad=(1, 1),
+                                             kernel=(3, 3),
+                                             stride=(1, 1),
+                                             weight=weight_var,
+                                             bias=bias_var,
+                                             lr_mult=lr_mult)
+        conv2_offset_t = mx.sym.slice_axis(conv2_offset,
+                                           axis=1,
+                                           begin=0,
+                                           end=18)
+        conv2_mask = mx.sym.slice_axis(conv2_offset,
+                                       axis=1,
+                                       begin=18,
+                                       end=None)
+        conv2_mask = 2 * mx.sym.Activation(conv2_mask, act_type='sigmoid')
+
+        conv2 = mx.contrib.symbol.ModulatedDeformableConvolution(
+            name=name + '_conv2',
+            data=net,
+            offset=conv2_offset_t,
+            mask=conv2_mask,
+            num_filter=num_filter,
+            pad=(1, 1),
+            kernel=(3, 3),
+            stride=(1, 1),
+            num_deformable_group=num_group,
+            no_bias=True)
+        net = conv2
+    net = mx.sym.BatchNorm(data=net,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=0.9,
+                           name=name + '_bn')
+    if len(act_type) > 0:
+        net = mx.symbol.Activation(data=net,
+                                   act_type=act_type,
+                                   name=name + '_act')
+    return net
+
+def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
+    assert kernel[0] == 3
+    weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                init=mx.init.Normal(0.01),
+                                attr={'__lr_mult__': '1.0'})
+    bias = mx.symbol.Variable(name="{}_bias".format(name),
+                              init=mx.init.Constant(0.0),
+                              attr={
+                                  '__lr_mult__': '2.0',
+                                  '__wd_mult__': str(bias_wd_mult)
+                              })
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
+    conv = mx.sym.BatchNorm(data=conv,
+                            fix_gamma=False,
+                            eps=2e-5,
+                            momentum=0.9,
+                            name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
+
+    if config.USE_DCN > 1 and kernel == (3, 3) and pad == (
+            1, 1) and stride == (1, 1) and not separable:
+        return conv_deformable(from_layer,
+                               num_filter,
+                               num_group=1,
+                               act_type=act_type,
+                               name=name)
+
+    if separable:
+        assert kernel[0] > 1
+        assert filter_in > 0
+    if not separable:
+        weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                    init=mx.init.Normal(0.01),
+                                    attr={'__lr_mult__': '1.0'})
+        bias = mx.symbol.Variable(name="{}_bias".format(name),
+                                  init=mx.init.Constant(0.0),
+                                  attr={
+                                      '__lr_mult__': '2.0',
+                                      '__wd_mult__': str(bias_wd_mult)
+                                  })
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    else:
+        if filter_in < 0:
+            filter_in = num_filter
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_sep_bn')
+        conv = mx.symbol.Activation(data=conv, act_type='relu', \
+            name="{}_sep_bn_relu".format(name))
+        conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
+            stride=(1,1), num_filter=num_filter, name="{}".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+
+def ssh_context_module(body, num_filter, filter_in, name):
+    conv_dimred = conv_act_layer(body,
+                                 name + '_conv1',
+                                 num_filter,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 separable=False,
+                                 filter_in=filter_in)
+    conv5x5 = conv_act_layer(conv_dimred,
+                             name + '_conv2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=False)
+    conv7x7_1 = conv_act_layer(conv_dimred,
+                               name + '_conv3_1',
+                               num_filter,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               separable=False)
+    conv7x7 = conv_act_layer(conv7x7_1,
+                             name + '_conv3_2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=False)
+    return (conv5x5, conv7x7)
+
+
+def ssh_detection_module(body, num_filter, filter_in, name):
+    assert num_filter % 4 == 0
+    conv3x3 = conv_act_layer(body,
+                             name + '_conv1',
+                             num_filter // 2,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=False,
+                             filter_in=filter_in)
+    #_filter = max(num_filter//4, 16)
+    _filter = num_filter // 4
+    conv5x5, conv7x7 = ssh_context_module(body, _filter, filter_in,
+                                          name + '_context')
+    ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7],
+                        dim=1,
+                        name=name + '_concat')
+    ret = mx.symbol.Activation(data=ret,
+                               act_type='relu',
+                               name=name + '_concat_relu')
+    out_filter = num_filter // 2 + _filter * 2
+    if config.USE_DCN > 0:
+        ret = conv_deformable(ret,
+                              num_filter=out_filter,
+                              name=name + '_concat_dcn')
+    return ret
+
+
+#def retina_context_module(body, kernel, num_filter, filter_in, name):
+#  conv_dimred = conv_act_layer(body, name+'_conv0',
+#      num_filter, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False, filter_in = filter_in)
+#  conv1 = conv_act_layer(conv_dimred, name+'_conv1',
+#      num_filter*6, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False, filter_in = filter_in)
+#  conv2 = conv_act_layer(conv1, name+'_conv2',
+#      num_filter*6, kernel=kernel, pad=((kernel[0]-1)//2, (kernel[1]-1)//2), stride=(1, 1), act_type='relu', separable=True, filter_in = num_filter*6)
+#  conv3 = conv_act_layer(conv2, name+'_conv3',
+#      num_filter, kernel=(1,1), pad=(0,0), stride=(1, 1), act_type='relu', separable=False)
+#  conv3 = conv3 + conv_dimred
+#  return conv3
+
+
+def retina_detection_module(body, num_filter, filter_in, name):
+    assert num_filter % 4 == 0
+    conv1 = conv_act_layer(body,
+                           name + '_conv1',
+                           num_filter // 2,
+                           kernel=(3, 3),
+                           pad=(1, 1),
+                           stride=(1, 1),
+                           act_type='relu',
+                           separable=False,
+                           filter_in=filter_in)
+    conv2 = conv_act_layer(conv1,
+                           name + '_conv2',
+                           num_filter // 2,
+                           kernel=(3, 3),
+                           pad=(1, 1),
+                           stride=(1, 1),
+                           act_type='relu',
+                           separable=False,
+                           filter_in=num_filter // 2)
+    conv3 = conv_act_layer(conv2,
+                           name + '_conv3',
+                           num_filter // 2,
+                           kernel=(3, 3),
+                           pad=(1, 1),
+                           stride=(1, 1),
+                           act_type='relu',
+                           separable=False,
+                           filter_in=num_filter // 2)
+    conv4 = conv2 + conv3
+    body = mx.sym.concat(*[conv1, conv4], dim=1, name=name + '_concat')
+    if config.USE_DCN > 0:
+        body = conv_deformable(body,
+                               num_filter=num_filter,
+                               name=name + '_concat_dcn')
+    return body
+
+
+def head_module(body, num_filter, filter_in, name):
+    if config.HEAD_MODULE == 'SSH':
+        return ssh_detection_module(body, num_filter, filter_in, name)
+    else:
+        return retina_detection_module(body, num_filter, filter_in, name)
+
+
+def upsampling(data, num_filter, name):
+    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(4,4),  stride=(2, 2), pad=(1,1),
+    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+    #    name=name)
+    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(2,2),  stride=(2, 2), pad=(0,0),
+    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+    #    name=name)
+    ret = mx.symbol.UpSampling(data,
+                               scale=2,
+                               sample_type='nearest',
+                               workspace=512,
+                               name=name,
+                               num_args=1)
+    return ret
+
+
+def get_sym_by_name(name, sym_buffer):
+    if name in sym_buffer:
+        return sym_buffer[name]
+    ret = None
+    name_key = name[0:1]
+    name_num = int(name[1:])
+    #print('getting', name, name_key, name_num)
+    if name_key == 'C':
+        assert name_num % 2 == 0
+        bottom = get_sym_by_name('C%d' % (name_num // 2), sym_buffer)
+        ret = conv_act_layer(bottom,
+                             '%s_C%d' (PREFIX, name_num),
+                             F1,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(2, 2),
+                             act_type='relu',
+                             bias_wd_mult=_bwm)
+    elif name_key == 'P':
+        assert name_num % 2 == 0
+        assert name_num <= max(config.RPN_FEAT_STRIDE)
+        lateral = get_sym_by_name('L%d' % (name_num), sym_buffer)
+        if name_num == max(config.RPN_FEAT_STRIDE) or name_num > 32:
+            ret = mx.sym.identity(lateral, name='%s_P%d' % (PREFIX, name_num))
+        else:
+            bottom = get_sym_by_name('L%d' % (name_num * 2), sym_buffer)
+            bottom_up = upsampling(bottom, F1, '%s_U%d' % (PREFIX, name_num))
+            if config.USE_CROP:
+                bottom_up = mx.symbol.Crop(*[bottom_up, lateral])
+            aggr = lateral + bottom_up
+            aggr = conv_act_layer(aggr,
+                                  '%s_A%d' % (PREFIX, name_num),
+                                  F1,
+                                  kernel=(3, 3),
+                                  pad=(1, 1),
+                                  stride=(1, 1),
+                                  act_type='relu',
+                                  bias_wd_mult=_bwm)
+            ret = mx.sym.identity(aggr, name='%s_P%d' % (PREFIX, name_num))
+    elif name_key == 'L':
+        c = get_sym_by_name('C%d' % (name_num), sym_buffer)
+        #print('L', name, F1)
+        ret = conv_act_layer(c,
+                             '%s_L%d' % (PREFIX, name_num),
+                             F1,
+                             kernel=(1, 1),
+                             pad=(0, 0),
+                             stride=(1, 1),
+                             act_type='relu',
+                             bias_wd_mult=_bwm)
+    else:
+        raise RuntimeError('%s is not a valid sym key name' % name)
+    sym_buffer[name] = ret
+    return ret
+
+
+def get_sym_conv(data, sym):
+    all_layers = sym.get_internals()
+
+    isize = 640
+    _, out_shape, _ = all_layers.infer_shape(data=(1, 3, isize, isize))
+    last_entry = None
+    c1 = None
+    c2 = None
+    c3 = None
+    c1_name = None
+    c2_name = None
+    c3_name = None
+    c1_filter = -1
+    c2_filter = -1
+    c3_filter = -1
+    #print(len(all_layers), len(out_shape))
+    #print(all_layers.__class__)
+    outputs = all_layers.list_outputs()
+    #print(outputs.__class__, len(outputs))
+    count = len(outputs)
+    stride2name = {}
+    stride2layer = {}
+    stride2shape = {}
+    for i in range(count):
+        name = outputs[i]
+        shape = out_shape[i]
+        print(i, name, count, shape)
+        if not name.endswith('_output'):
+            continue
+        if len(shape) != 4:
+            continue
+        assert isize % shape[2] == 0
+        if shape[1] > config.max_feat_channel:
+            break
+        stride = isize // shape[2]
+        stride2name[stride] = name
+        stride2layer[stride] = all_layers[name]
+        stride2shape[stride] = shape
+
+    strides = sorted(stride2name.keys())
+    for stride in strides:
+        print('stride', stride, stride2name[stride], stride2shape[stride])
+    print('F1_F2', F1, F2)
+    #print('cnames', c1_name, c2_name, c3_name, F1, F2)
+    _bwm = 1.0
+    ret = {}
+    sym_buffer = {}
+    for stride in [4, 8, 16, 32]:
+        sym_buffer['C%d' % stride] = stride2layer[stride]
+    if not config.USE_FPN:
+        for stride in config.RPN_FEAT_STRIDE:
+            name = 'L%d' % stride
+            ret[stride] = get_sym_by_name(name, sym_buffer)
+    else:
+        for stride in config.RPN_FEAT_STRIDE:
+            name = 'P%d' % stride
+            ret[stride] = get_sym_by_name(name, sym_buffer)
+
+    return ret
+
+
+def get_out(conv_fpn_feat,
+            prefix,
+            stride,
+            landmark=False,
+            lr_mult=1.0,
+            gt_boxes=None):
+    A = config.NUM_ANCHORS
+    bbox_pred_len = 4
+    landmark_pred_len = 10
+    if config.USE_BLUR:
+        bbox_pred_len = 5
+    if config.USE_OCCLUSION:
+        landmark_pred_len = 15
+    ret_group = []
+    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
+    cls_label = mx.symbol.Variable(name='%s_label_stride%d' % (prefix, stride))
+    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d' %
+                                     (prefix, stride))
+    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d' %
+                                     (prefix, stride))
+    if landmark:
+        landmark_target = mx.symbol.Variable(
+            name='%s_landmark_target_stride%d' % (prefix, stride))
+        landmark_weight = mx.symbol.Variable(
+            name='%s_landmark_weight_stride%d' % (prefix, stride))
+    conv_feat = conv_fpn_feat[stride]
+    rpn_relu = head_module(conv_feat, F2 * config.CONTEXT_FILTER_RATIO, F1,
+                           'rf_head_stride%d' % stride)
+
+    rpn_cls_score = conv_only(rpn_relu,
+                              '%s_rpn_cls_score_stride%d' % (prefix, stride),
+                              2 * num_anchors,
+                              kernel=(1, 1),
+                              pad=(0, 0),
+                              stride=(1, 1))
+
+    rpn_bbox_pred = conv_only(rpn_relu,
+                              '%s_rpn_bbox_pred_stride%d' % (prefix, stride),
+                              bbox_pred_len * num_anchors,
+                              kernel=(1, 1),
+                              pad=(0, 0),
+                              stride=(1, 1))
+
+    # prepare rpn data
+    rpn_cls_score_reshape = mx.symbol.Reshape(
+        data=rpn_cls_score,
+        shape=(0, 2, -1),
+        name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+
+    rpn_bbox_pred_reshape = mx.symbol.Reshape(
+        data=rpn_bbox_pred,
+        shape=(0, 0, -1),
+        name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix, stride))
+    if landmark:
+        rpn_landmark_pred = conv_only(rpn_relu,
+                                      '%s_rpn_landmark_pred_stride%d' %
+                                      (prefix, stride),
+                                      landmark_pred_len * num_anchors,
+                                      kernel=(1, 1),
+                                      pad=(0, 0),
+                                      stride=(1, 1))
+        rpn_landmark_pred_reshape = mx.symbol.Reshape(
+            data=rpn_landmark_pred,
+            shape=(0, 0, -1),
+            name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix, stride))
+
+    if config.TRAIN.RPN_ENABLE_OHEM >= 2:
+        label, anchor_weight, pos_count = mx.sym.Custom(
+            op_type='rpn_fpn_ohem3',
+            stride=int(stride),
+            network=config.network,
+            dataset=config.dataset,
+            prefix=prefix,
+            cls_score=rpn_cls_score_reshape,
+            labels=cls_label)
+
+        _bbox_weight = mx.sym.tile(anchor_weight, (1, 1, bbox_pred_len))
+        _bbox_weight = _bbox_weight.reshape(
+            (0, -1, A * bbox_pred_len)).transpose((0, 2, 1))
+        bbox_weight = mx.sym.elemwise_mul(bbox_weight,
+                                          _bbox_weight,
+                                          name='%s_bbox_weight_mul_stride%s' %
+                                          (prefix, stride))
+
+        if landmark:
+            _landmark_weight = mx.sym.tile(anchor_weight,
+                                           (1, 1, landmark_pred_len))
+            _landmark_weight = _landmark_weight.reshape(
+                (0, -1, A * landmark_pred_len)).transpose((0, 2, 1))
+            landmark_weight = mx.sym.elemwise_mul(
+                landmark_weight,
+                _landmark_weight,
+                name='%s_landmark_weight_mul_stride%s' % (prefix, stride))
+    else:
+        label = cls_label
+        #if not config.FACE_LANDMARK:
+        #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
+        #else:
+        #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
+    #cls loss
+    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
+                                           label=label,
+                                           multi_output=True,
+                                           normalization='valid',
+                                           use_ignore=True,
+                                           ignore_label=-1,
+                                           grad_scale=lr_mult,
+                                           name='%s_rpn_cls_prob_stride%d' %
+                                           (prefix, stride))
+    ret_group.append(rpn_cls_prob)
+    ret_group.append(mx.sym.BlockGrad(label))
+
+    pos_count = mx.symbol.sum(pos_count)
+    pos_count = pos_count + 0.001  #avoid zero
+
+    #bbox loss
+    bbox_diff = rpn_bbox_pred_reshape - bbox_target
+    bbox_diff = bbox_diff * bbox_weight
+    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_' %
+                                         (prefix, stride),
+                                         scalar=3.0,
+                                         data=bbox_diff)
+    bbox_lr_mode0 = 0.25 * lr_mult * config.TRAIN.BATCH_IMAGES / config.TRAIN.RPN_BATCH_SIZE
+    landmark_lr_mode0 = 0.4 * config.LANDMARK_LR_MULT * bbox_lr_mode0
+    if config.LR_MODE == 0:
+        rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d' %
+                                        (prefix, stride),
+                                        data=rpn_bbox_loss_,
+                                        grad_scale=bbox_lr_mode0)
+    else:
+        rpn_bbox_loss_ = mx.symbol.broadcast_div(rpn_bbox_loss_, pos_count)
+        rpn_bbox_loss = mx.sym.MakeLoss(name='%s_rpn_bbox_loss_stride%d' %
+                                        (prefix, stride),
+                                        data=rpn_bbox_loss_,
+                                        grad_scale=0.5 * lr_mult)
+    ret_group.append(rpn_bbox_loss)
+    ret_group.append(mx.sym.BlockGrad(bbox_weight))
+
+    #landmark loss
+    if landmark:
+        landmark_diff = rpn_landmark_pred_reshape - landmark_target
+        landmark_diff = landmark_diff * landmark_weight
+        rpn_landmark_loss_ = mx.symbol.smooth_l1(
+            name='%s_rpn_landmark_loss_stride%d_' % (prefix, stride),
+            scalar=3.0,
+            data=landmark_diff)
+        if config.LR_MODE == 0:
+            rpn_landmark_loss = mx.sym.MakeLoss(
+                name='%s_rpn_landmark_loss_stride%d' % (prefix, stride),
+                data=rpn_landmark_loss_,
+                grad_scale=landmark_lr_mode0)
+        else:
+            rpn_landmark_loss_ = mx.symbol.broadcast_div(
+                rpn_landmark_loss_, pos_count)
+            rpn_landmark_loss = mx.sym.MakeLoss(
+                name='%s_rpn_landmark_loss_stride%d' % (prefix, stride),
+                data=rpn_landmark_loss_,
+                grad_scale=0.2 * config.LANDMARK_LR_MULT * lr_mult)
+        ret_group.append(rpn_landmark_loss)
+        ret_group.append(mx.sym.BlockGrad(landmark_weight))
+    if config.USE_3D:
+        from rcnn.PY_OP import rpn_3d_mesh
+        pass
+    if config.CASCADE > 0:
+        if config.CASCADE_MODE == 0:
+            body = rpn_relu
+        elif config.CASCADE_MODE == 1:
+            body = head_module(conv_feat, F2 * config.CONTEXT_FILTER_RATIO, F1,
+                               '%s_head_stride%d_cas' % (PREFIX, stride))
+        elif config.CASCADE_MODE == 2:
+            body = conv_feat + rpn_relu
+            body = head_module(body, F2 * config.CONTEXT_FILTER_RATIO, F1,
+                               '%s_head_stride%d_cas' % (PREFIX, stride))
+        else:
+            body = head_module(conv_feat, F2 * config.CONTEXT_FILTER_RATIO, F1,
+                               '%s_head_stride%d_cas' % (PREFIX, stride))
+            body = mx.sym.concat(body,
+                                 rpn_cls_score,
+                                 rpn_bbox_pred,
+                                 rpn_landmark_pred,
+                                 dim=1)
+
+        #cls_pred = rpn_cls_prob
+        cls_pred_t0 = rpn_cls_score_reshape
+        cls_label_raw = cls_label
+        cls_label_t0 = label
+        bbox_pred_t0 = rpn_bbox_pred_reshape
+        #bbox_pred = rpn_bbox_pred
+        #bbox_pred = mx.sym.transpose(bbox_pred, (0, 2, 3, 1))
+        #bbox_pred_len = 4
+        #bbox_pred = mx.sym.reshape(bbox_pred, (0, -1, bbox_pred_len))
+        bbox_label_t0 = bbox_target
+        #prefix = prefix+'2'
+        for casid in range(config.CASCADE):
+            #pseudo-code
+            #anchor_label = GENANCHOR(bbox_label, bbox_pred, stride)
+            #bbox_label = F(anchor_label, bbox_pred)
+            #bbox_label = bbox_label - bbox_pred
+            cls_pred = conv_only(body,
+                                 '%s_rpn_cls_score_stride%d_cas%d' %
+                                 (prefix, stride, casid),
+                                 2 * num_anchors,
+                                 kernel=(1, 1),
+                                 pad=(0, 0),
+                                 stride=(1, 1))
+            rpn_cls_score_reshape = mx.symbol.Reshape(
+                data=cls_pred,
+                shape=(0, 2, -1),
+                name="%s_rpn_cls_score_reshape_stride%s_cas%d" %
+                (prefix, stride, casid))
+
+            #bbox_label equals to bbox_target
+            #cls_pred, cls_label, bbox_pred, bbox_label, bbox_weight, pos_count = mx.sym.Custom(op_type='cascade_refine', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_pred=cls_pred, cls_label = cls_label, bbox_pred = bbox_pred, bbox_label = bbox_label)
+            #cls_label, bbox_label, anchor_weight, pos_count = mx.sym.Custom(op_type='cascade_refine', stride=int(stride), network=config.network, dataset=config.dataset, prefix=prefix, cls_pred_t0=cls_pred_t0, cls_label_t0 = cls_label_t0, cls_pred = rpn_cls_score_reshape, bbox_pred_t0 = bbox_pred_t0, bbox_label_t0 = bbox_label_t0)
+            cls_label, bbox_label, anchor_weight, pos_count = mx.sym.Custom(
+                op_type='cascade_refine',
+                stride=int(stride),
+                network=config.network,
+                dataset=config.dataset,
+                prefix=prefix,
+                cls_label_t0=cls_label_t0,
+                cls_pred_t0=cls_pred_t0,
+                cls_pred=rpn_cls_score_reshape,
+                bbox_pred_t0=bbox_pred_t0,
+                bbox_label_t0=bbox_label_t0,
+                cls_label_raw=cls_label_raw,
+                cas_gt_boxes=gt_boxes)
+            if stride in config.CASCADE_CLS_STRIDES:
+                rpn_cls_prob = mx.symbol.SoftmaxOutput(
+                    data=rpn_cls_score_reshape,
+                    label=cls_label,
+                    multi_output=True,
+                    normalization='valid',
+                    use_ignore=True,
+                    ignore_label=-1,
+                    grad_scale=lr_mult,
+                    name='%s_rpn_cls_prob_stride%d_cas%d' %
+                    (prefix, stride, casid))
+                ret_group.append(rpn_cls_prob)
+                ret_group.append(mx.sym.BlockGrad(cls_label))
+            if stride in config.CASCADE_BBOX_STRIDES:
+                bbox_pred = conv_only(body,
+                                      '%s_rpn_bbox_pred_stride%d_cas%d' %
+                                      (prefix, stride, casid),
+                                      bbox_pred_len * num_anchors,
+                                      kernel=(1, 1),
+                                      pad=(0, 0),
+                                      stride=(1, 1))
+
+                rpn_bbox_pred_reshape = mx.symbol.Reshape(
+                    data=bbox_pred,
+                    shape=(0, 0, -1),
+                    name="%s_rpn_bbox_pred_reshape_stride%s_cas%d" %
+                    (prefix, stride, casid))
+                _bbox_weight = mx.sym.tile(anchor_weight,
+                                           (1, 1, bbox_pred_len))
+                _bbox_weight = _bbox_weight.reshape(
+                    (0, -1, A * bbox_pred_len)).transpose((0, 2, 1))
+                bbox_weight = _bbox_weight
+                pos_count = mx.symbol.sum(pos_count)
+                pos_count = pos_count + 0.01  #avoid zero
+                #bbox_weight = mx.sym.elemwise_mul(bbox_weight, _bbox_weight, name='%s_bbox_weight_mul_stride%s'%(prefix,stride))
+                #bbox loss
+                bbox_diff = rpn_bbox_pred_reshape - bbox_label
+                bbox_diff = bbox_diff * bbox_weight
+                rpn_bbox_loss_ = mx.symbol.smooth_l1(
+                    name='%s_rpn_bbox_loss_stride%d_cas%d' %
+                    (prefix, stride, casid),
+                    scalar=3.0,
+                    data=bbox_diff)
+                if config.LR_MODE == 0:
+                    rpn_bbox_loss = mx.sym.MakeLoss(
+                        name='%s_rpn_bbox_loss_stride%d_cas%d' %
+                        (prefix, stride, casid),
+                        data=rpn_bbox_loss_,
+                        grad_scale=bbox_lr_mode0)
+                else:
+                    rpn_bbox_loss_ = mx.symbol.broadcast_div(
+                        rpn_bbox_loss_, pos_count)
+                    rpn_bbox_loss = mx.sym.MakeLoss(
+                        name='%s_rpn_bbox_loss_stride%d_cas%d' %
+                        (prefix, stride, casid),
+                        data=rpn_bbox_loss_,
+                        grad_scale=0.5 * lr_mult)
+                ret_group.append(rpn_bbox_loss)
+                ret_group.append(mx.sym.BlockGrad(bbox_weight))
+                #bbox_pred = rpn_bbox_pred_reshape
+
+    return ret_group
+
+
+def get_sym_train(sym):
+    data = mx.symbol.Variable(name="data")
+    global F1, F2
+    F1 = config.HEAD_FILTER_NUM
+    F2 = F1
+
+    # shared convolutional layers
+    conv_fpn_feat = get_sym_conv(data, sym)
+    ret_group = []
+    gt_boxes = None
+    if config.CASCADE > 0:
+        gt_boxes = mx.sym.Variable('gt_boxes')
+
+    for stride in config.RPN_FEAT_STRIDE:
+        ret = get_out(conv_fpn_feat,
+                      'face',
+                      stride,
+                      config.FACE_LANDMARK,
+                      lr_mult=1.0,
+                      gt_boxes=gt_boxes)
+        ret_group += ret
+
+    return mx.sym.Group(ret_group)
diff --git a/RetinaFace/rcnn/symbol/symbol_common.py.bak b/detection/retinaface/rcnn/symbol/symbol_common.py.bak
similarity index 100%
rename from RetinaFace/rcnn/symbol/symbol_common.py.bak
rename to detection/retinaface/rcnn/symbol/symbol_common.py.bak
diff --git a/detection/retinaface/rcnn/symbol/symbol_mnet.py b/detection/retinaface/rcnn/symbol/symbol_mnet.py
new file mode 100644
index 000000000..86f65a3eb
--- /dev/null
+++ b/detection/retinaface/rcnn/symbol/symbol_mnet.py
@@ -0,0 +1,834 @@
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import numpy as np
+from rcnn.config import config
+from rcnn.PY_OP import rpn_fpn_ohem3
+from rcnn.symbol.symbol_common import get_sym_train
+
+
+def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), bias_wd_mult=0.0, shared_weight=None, shared_bias = None):
+    if shared_weight is None:
+        weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                    init=mx.init.Normal(0.01),
+                                    attr={'__lr_mult__': '1.0'})
+        bias = mx.symbol.Variable(name="{}_bias".format(name),
+                                  init=mx.init.Constant(0.0),
+                                  attr={
+                                      '__lr_mult__': '2.0',
+                                      '__wd_mult__': str(bias_wd_mult)
+                                  })
+    else:
+        weight = shared_weight
+        bias = shared_bias
+        print('reuse shared var in', name)
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
+    return conv
+
+def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
+    assert kernel[0] == 3
+    weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                init=mx.init.Normal(0.01),
+                                attr={'__lr_mult__': '1.0'})
+    bias = mx.symbol.Variable(name="{}_bias".format(name),
+                              init=mx.init.Constant(0.0),
+                              attr={
+                                  '__lr_mult__': '2.0',
+                                  '__wd_mult__': str(bias_wd_mult)
+                              })
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
+    conv = mx.sym.BatchNorm(data=conv,
+                            fix_gamma=False,
+                            eps=2e-5,
+                            momentum=0.9,
+                            name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
+
+    separable = False
+    if separable:
+        assert kernel[0] == 3
+    if not separable:
+        weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                    init=mx.init.Normal(0.01),
+                                    attr={'__lr_mult__': '1.0'})
+        bias = mx.symbol.Variable(name="{}_bias".format(name),
+                                  init=mx.init.Constant(0.0),
+                                  attr={
+                                      '__lr_mult__': '2.0',
+                                      '__wd_mult__': str(bias_wd_mult)
+                                  })
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    else:
+        if filter_in < 0:
+            filter_in = num_filter
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_sep_bn')
+        conv = mx.symbol.Activation(data=conv, act_type='relu', \
+            name="{}_sep_bn_relu".format(name))
+        conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
+            stride=(1,1), num_filter=num_filter, name="{}".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+
+def ssh_context_module(body, num_filter, filter_in, name):
+    conv_dimred = conv_act_layer(body,
+                                 name + '_conv1',
+                                 num_filter,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 separable=True,
+                                 filter_in=filter_in)
+    conv5x5 = conv_act_layer(conv_dimred,
+                             name + '_conv2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True)
+    conv7x7_1 = conv_act_layer(conv_dimred,
+                               name + '_conv3_1',
+                               num_filter,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               separable=True)
+    conv7x7 = conv_act_layer(conv7x7_1,
+                             name + '_conv3_2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True)
+    return (conv5x5, conv7x7)
+
+
+def ssh_detection_module(body, num_filter, filter_in, name):
+    conv3x3 = conv_act_layer(body,
+                             name + '_conv1',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True,
+                             filter_in=filter_in)
+    conv5x5, conv7x7 = ssh_context_module(body, num_filter // 2, filter_in,
+                                          name + '_context')
+    ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7],
+                        dim=1,
+                        name=name + '_concat')
+    ret = mx.symbol.Activation(data=ret,
+                               act_type='relu',
+                               name=name + '_concat_relu')
+    return ret
+
+
+def upsampling(data, num_filter, name):
+    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(4,4),  stride=(2, 2), pad=(1,1),
+    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+    #    name=name)
+    #ret = mx.symbol.Deconvolution(data=data, num_filter=num_filter, kernel=(2,2),  stride=(2, 2), pad=(0,0),
+    #    num_group = num_filter, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+    #    name=name)
+    ret = mx.symbol.UpSampling(data,
+                               scale=2,
+                               sample_type='nearest',
+                               workspace=512,
+                               name=name,
+                               num_args=1)
+    return ret
+
+
+def get_mnet_conv(data, sym):
+    mm = config.MULTIPLIER
+    all_layers = sym.get_internals()
+    #print(all_layers)
+    ##c1 = all_layers['mobilenetv20_features_linearbottleneck6_relu60_relu6_output'] #96
+    #c1 = all_layers['mobilenetv20_features_linearbottleneck5_elemwise_add0_output'] # 16
+    ##c2 = all_layers['mobilenetv20_features_linearbottleneck13_relu60_relu6_output']
+    #c2 = all_layers['mobilenetv20_features_linearbottleneck12_elemwise_add0_output'] # 48
+    ##c3 = all_layers['mobilenetv20_features_linearbottleneck16_batchnorm2_fwd_output'] # 160
+    #c3 = all_layers['mobilenetv20_features_linearbottleneck13_batchnorm2_fwd_output'] # 80
+    #c1_filter = int(32*mm)
+    #c2_filter = int(96*mm)
+    #c3_filter = int(160*mm)
+
+    #c1 = all_layers['mobilenet0_relu10_fwd_output']
+    #c2 = all_layers['mobilenet0_relu22_fwd_output']
+    #c3 = all_layers['mobilenet0_relu26_fwd_output']
+
+    #c1 = all_layers['conv_6_relu_output']
+    #c2 = all_layers['conv_12_relu_output']
+    #c3 = all_layers['conv_14_relu_output']
+    #c1_filter = int(256*mm)
+    #c2_filter = int(512*mm)
+    #c3_filter = int(1024*mm)
+
+    isize = 640
+    _, out_shape, _ = all_layers.infer_shape(data=(1, 3, isize, isize))
+    last_entry = None
+    c1 = None
+    c2 = None
+    c3 = None
+    c1_name = None
+    c2_name = None
+    c3_name = None
+    c1_filter = -1
+    c2_filter = -1
+    c3_filter = -1
+    #print(len(all_layers), len(out_shape))
+    #print(all_layers.__class__)
+    outputs = all_layers.list_outputs()
+    #print(outputs.__class__, len(outputs))
+    count = len(outputs)
+    for i in range(count):
+        name = outputs[i]
+        shape = out_shape[i]
+        if not name.endswith('_output'):
+            continue
+        if len(shape) != 4:
+            continue
+        #print(name, shape)
+        if c1 is None and shape[2] == isize // 16:
+            cname = last_entry[0]
+            #print('c1', last_entry)
+            c1 = all_layers[cname]
+            c1_name = cname
+        if c2 is None and shape[2] == isize // 32:
+            cname = last_entry[0]
+            #print('c2', last_entry)
+            c2 = all_layers[cname]
+            c2_name = cname
+        if shape[2] == isize // 32:
+            c3 = all_layers[name]
+            #print('c3', name, shape)
+            c3_name = name
+
+        last_entry = (name, shape)
+    print('cnames', c1_name, c2_name, c3_name)
+
+    F1 = int(256 * mm)
+    F2 = int(128 * mm)
+    if config.SHARE_WEIGHT_BBOX or config.SHARE_WEIGHT_LANDMARK:
+        F2 = F1
+    _bwm = 1.0
+    if config.NET_MODE == 0:
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_m2_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
+        #    num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+        #    name='ssh_m2_red_upsampling')
+        #c2_up = mx.symbol.UpSampling(c2_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2_lateral, F2, 'ssh_m2_red_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+
+        c1 = c1_lateral + c2_up
+
+        c1 = conv_act_layer(c1,
+                            'ssh_m1_conv',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 1:
+        c3_lateral = conv_act_layer(c3,
+                                    'ssh_c3_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #c3_up = mx.symbol.UpSampling(c3_lateral, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
+        c3_up = upsampling(c3_lateral, F2, 'ssh_c3_upsampling')
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+
+        c1 = c1_lateral + c2_up
+
+        c1 = conv_act_layer(c1,
+                            'ssh_m1_conv',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 2:
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
+        c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 3:
+        #c3 = conv_act_layer(c3, 'ssh_c3_lateral',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c3 = ssh_detection_module(c3, F2 // 2, c3_filter, 'ssh_c3_lateral')
+        #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
+        c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
+        #c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c2_lateral = ssh_detection_module(c2, F2 // 2, c2_filter,
+                                          'ssh_c2_lateral')
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        #c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c1_lateral = ssh_detection_module(c1, F2 // 2, c1_filter,
+                                          'ssh_c1_lateral')
+        #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 4:
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
+        c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2 // 2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1 // 2, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1 // 2, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 5:
+        c3 = conv_act_layer_dw(c3,
+                               'ssh_c3_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        #c3_up = mx.symbol.UpSampling(c3, scale=2, sample_type='nearest', workspace=512, name='ssh_c3_up', num_args=1)
+        c3_up = upsampling(c3, F2, 'ssh_c3_upsampling')
+        c2 = conv_act_layer_dw(c2,
+                               'ssh_c2_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1 = conv_act_layer_dw(c1,
+                               'ssh_c1_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #c2_up = mx.symbol.UpSampling(c2, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        c2_up = upsampling(c2, F2, 'ssh_c2_upsampling')
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+
+    return {8: m1, 16: m2, 32: m3}
+
+
+def get_out(conv_fpn_feat,
+            prefix,
+            stride,
+            landmark=False,
+            lr_mult=1.0,
+            shared_vars=None):
+    A = config.NUM_ANCHORS
+    bbox_pred_len = 4
+    landmark_pred_len = 10
+    if config.USE_BLUR:
+        bbox_pred_len = 5
+    if config.USE_OCCLUSION:
+        landmark_pred_len = 15
+    ret_group = []
+    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
+    label = mx.symbol.Variable(name='%s_label_stride%d' % (prefix, stride))
+    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d' %
+                                     (prefix, stride))
+    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d' %
+                                     (prefix, stride))
+    if landmark:
+        landmark_target = mx.symbol.Variable(
+            name='%s_landmark_target_stride%d' % (prefix, stride))
+        landmark_weight = mx.symbol.Variable(
+            name='%s_landmark_weight_stride%d' % (prefix, stride))
+    rpn_relu = conv_fpn_feat[stride]
+    maxout_stat = 0
+    if config.USE_MAXOUT >= 1 and stride == config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 1
+    if config.USE_MAXOUT >= 2 and stride != config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 2
+
+    if maxout_stat == 0:
+        rpn_cls_score = conv_only(rpn_relu,
+                                  '%s_rpn_cls_score_stride%d' %
+                                  (prefix, stride),
+                                  2 * num_anchors,
+                                  kernel=(1, 1),
+                                  pad=(0, 0),
+                                  stride=(1, 1),
+                                  shared_weight=shared_vars[0][0],
+                                  shared_bias=shared_vars[0][1])
+    elif maxout_stat == 1:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+    else:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+
+    rpn_bbox_pred = conv_only(rpn_relu,
+                              '%s_rpn_bbox_pred_stride%d' % (prefix, stride),
+                              bbox_pred_len * num_anchors,
+                              kernel=(1, 1),
+                              pad=(0, 0),
+                              stride=(1, 1),
+                              shared_weight=shared_vars[1][0],
+                              shared_bias=shared_vars[1][1])
+
+    # prepare rpn data
+    if not config.FBN:
+        rpn_cls_score_reshape = mx.symbol.Reshape(
+            data=rpn_cls_score,
+            shape=(0, 2, -1),
+            name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+    else:
+        rpn_cls_score_reshape = mx.symbol.Reshape(
+            data=rpn_cls_score,
+            shape=(0, 2, -1),
+            name="%s_rpn_cls_score_reshape_stride%s_pre" % (prefix, stride))
+        rpn_cls_score_reshape = mx.symbol.BatchNorm(
+            rpn_cls_score_reshape,
+            fix_gamma=True,
+            eps=2e-5,
+            name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+
+    rpn_bbox_pred_reshape = mx.symbol.Reshape(
+        data=rpn_bbox_pred,
+        shape=(0, 0, -1),
+        name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix, stride))
+    if landmark:
+        rpn_landmark_pred = conv_only(rpn_relu,
+                                      '%s_rpn_landmark_pred_stride%d' %
+                                      (prefix, stride),
+                                      landmark_pred_len * num_anchors,
+                                      kernel=(1, 1),
+                                      pad=(0, 0),
+                                      stride=(1, 1),
+                                      shared_weight=shared_vars[2][0],
+                                      shared_bias=shared_vars[2][1])
+        rpn_landmark_pred_reshape = mx.symbol.Reshape(
+            data=rpn_landmark_pred,
+            shape=(0, 0, -1),
+            name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix, stride))
+
+    if config.TRAIN.RPN_ENABLE_OHEM >= 2:
+        label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3',
+                                             stride=int(stride),
+                                             network=config.network,
+                                             dataset=config.dataset,
+                                             prefix=prefix,
+                                             cls_score=rpn_cls_score_reshape,
+                                             labels=label)
+
+        _bbox_weight = mx.sym.tile(anchor_weight, (1, 1, bbox_pred_len))
+        _bbox_weight = _bbox_weight.reshape(
+            (0, -1, A * bbox_pred_len)).transpose((0, 2, 1))
+        bbox_weight = mx.sym.elemwise_mul(bbox_weight,
+                                          _bbox_weight,
+                                          name='%s_bbox_weight_mul_stride%s' %
+                                          (prefix, stride))
+
+        if landmark:
+            _landmark_weight = mx.sym.tile(anchor_weight,
+                                           (1, 1, landmark_pred_len))
+            _landmark_weight = _landmark_weight.reshape(
+                (0, -1, A * landmark_pred_len)).transpose((0, 2, 1))
+            landmark_weight = mx.sym.elemwise_mul(
+                landmark_weight,
+                _landmark_weight,
+                name='%s_landmark_weight_mul_stride%s' % (prefix, stride))
+        #if not config.FACE_LANDMARK:
+        #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
+        #else:
+        #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
+    #cls loss
+    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
+                                           label=label,
+                                           multi_output=True,
+                                           normalization='valid',
+                                           use_ignore=True,
+                                           ignore_label=-1,
+                                           grad_scale=lr_mult,
+                                           name='%s_rpn_cls_prob_stride%d' %
+                                           (prefix, stride))
+    ret_group.append(rpn_cls_prob)
+    ret_group.append(mx.sym.BlockGrad(label))
+
+    #bbox loss
+    bbox_diff = rpn_bbox_pred_reshape - bbox_target
+    bbox_diff = bbox_diff * bbox_weight
+    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_' %
+                                         (prefix, stride),
+                                         scalar=3.0,
+                                         data=bbox_diff)
+    rpn_bbox_loss = mx.sym.MakeLoss(
+        name='%s_rpn_bbox_loss_stride%d' % (prefix, stride),
+        data=rpn_bbox_loss_,
+        grad_scale=1.0 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+    ret_group.append(rpn_bbox_loss)
+    ret_group.append(mx.sym.BlockGrad(bbox_weight))
+
+    #landmark loss
+    if landmark:
+        landmark_diff = rpn_landmark_pred_reshape - landmark_target
+        landmark_diff = landmark_diff * landmark_weight
+        rpn_landmark_loss_ = mx.symbol.smooth_l1(
+            name='%s_rpn_landmark_loss_stride%d_' % (prefix, stride),
+            scalar=3.0,
+            data=landmark_diff)
+        rpn_landmark_loss = mx.sym.MakeLoss(
+            name='%s_rpn_landmark_loss_stride%d' % (prefix, stride),
+            data=rpn_landmark_loss_,
+            grad_scale=0.5 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+        ret_group.append(rpn_landmark_loss)
+        ret_group.append(mx.sym.BlockGrad(landmark_weight))
+    return ret_group
+
+
+def get_mnet_train(sym):
+    return get_sym_train(sym)
+    #data = mx.symbol.Variable(name="data")
+    ## shared convolutional layers
+    #conv_fpn_feat = get_mnet_conv(data, sym)
+    #ret_group = []
+    #shared_vars = []
+    #if config.SHARE_WEIGHT_BBOX:
+    #  assert config.USE_MAXOUT==0
+    #  _name = 'face_rpn_cls_score_share'
+    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),
+    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
+    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),
+    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
+    #  shared_vars.append( [shared_weight, shared_bias] )
+    #  _name = 'face_rpn_bbox_pred_share'
+    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),
+    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
+    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),
+    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
+    #  shared_vars.append( [shared_weight, shared_bias] )
+    #else:
+    #  shared_vars.append( [None, None] )
+    #  shared_vars.append( [None, None] )
+    #if config.SHARE_WEIGHT_LANDMARK:
+    #  _name = 'face_rpn_landmark_pred_share'
+    #  shared_weight = mx.symbol.Variable(name="{}_weight".format(_name),
+    #      init=mx.init.Normal(0.01), attr={'__lr_mult__': '1.0'})
+    #  shared_bias = mx.symbol.Variable(name="{}_bias".format(_name),
+    #      init=mx.init.Constant(0.0), attr={'__lr_mult__': '2.0', '__wd_mult__': str(0.0)})
+    #  shared_vars.append( [shared_weight, shared_bias] )
+    #else:
+    #  shared_vars.append( [None, None] )
+
+    #for stride in config.RPN_FEAT_STRIDE:
+    #  ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0, shared_vars = shared_vars)
+    #  ret_group += ret
+    #  if config.HEAD_BOX:
+    #    ret = get_out(conv_fpn_feat, 'head', stride, False, lr_mult=0.5)
+    #    ret_group += ret
+
+    #return mx.sym.Group(ret_group)
diff --git a/RetinaFace/rcnn/symbol/symbol_mnet.py.bak b/detection/retinaface/rcnn/symbol/symbol_mnet.py.bak
similarity index 100%
rename from RetinaFace/rcnn/symbol/symbol_mnet.py.bak
rename to detection/retinaface/rcnn/symbol/symbol_mnet.py.bak
diff --git a/detection/retinaface/rcnn/symbol/symbol_resnet.py b/detection/retinaface/rcnn/symbol/symbol_resnet.py
new file mode 100644
index 000000000..7e6312b1b
--- /dev/null
+++ b/detection/retinaface/rcnn/symbol/symbol_resnet.py
@@ -0,0 +1,827 @@
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import numpy as np
+from rcnn.config import config
+from rcnn.PY_OP import rpn_fpn_ohem3
+from rcnn.symbol.symbol_common import get_sym_train
+
+def conv_only(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), bias_wd_mult=0.0):
+    weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                init=mx.init.Normal(0.01),
+                                attr={'__lr_mult__': '1.0'})
+    bias = mx.symbol.Variable(name="{}_bias".format(name),
+                              init=mx.init.Constant(0.0),
+                              attr={
+                                  '__lr_mult__': '2.0',
+                                  '__wd_mult__': str(bias_wd_mult)
+                              })
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
+    return conv
+
+def conv_act_layer_dw(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0):
+    assert kernel[0] == 3
+    weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                init=mx.init.Normal(0.01),
+                                attr={'__lr_mult__': '1.0'})
+    bias = mx.symbol.Variable(name="{}_bias".format(name),
+                              init=mx.init.Constant(0.0),
+                              attr={
+                                  '__lr_mult__': '2.0',
+                                  '__wd_mult__': str(bias_wd_mult)
+                              })
+    conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+        stride=stride, num_filter=num_filter, num_group=num_filter, name="{}".format(name), weight=weight, bias=bias)
+    conv = mx.sym.BatchNorm(data=conv,
+                            fix_gamma=False,
+                            eps=2e-5,
+                            momentum=0.9,
+                            name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0, separable=False, filter_in = -1):
+
+    separable = False
+    if separable:
+        assert kernel[0] == 3
+    if not separable:
+        weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                    init=mx.init.Normal(0.01),
+                                    attr={'__lr_mult__': '1.0'})
+        bias = mx.symbol.Variable(name="{}_bias".format(name),
+                                  init=mx.init.Constant(0.0),
+                                  attr={
+                                      '__lr_mult__': '2.0',
+                                      '__wd_mult__': str(bias_wd_mult)
+                                  })
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=num_filter, name="{}".format(name), weight=weight, bias=bias)
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    else:
+        if filter_in < 0:
+            filter_in = num_filter
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=filter_in, num_group=filter_in, name="{}_sep".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_sep_bn')
+        conv = mx.symbol.Activation(data=conv, act_type='relu', \
+            name="{}_sep_bn_relu".format(name))
+        conv = mx.symbol.Convolution(data=conv, kernel=(1,1), pad=(0,0), \
+            stride=(1,1), num_filter=num_filter, name="{}".format(name))
+        conv = mx.sym.BatchNorm(data=conv,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name=name + '_bn')
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+
+def ssh_context_module(body, num_filter, filter_in, name):
+    conv_dimred = conv_act_layer(body,
+                                 name + '_conv1',
+                                 num_filter,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 separable=True,
+                                 filter_in=filter_in)
+    conv5x5 = conv_act_layer(conv_dimred,
+                             name + '_conv2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True)
+    conv7x7_1 = conv_act_layer(conv_dimred,
+                               name + '_conv3_1',
+                               num_filter,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               separable=True)
+    conv7x7 = conv_act_layer(conv7x7_1,
+                             name + '_conv3_2',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True)
+    return (conv5x5, conv7x7)
+
+
+def conv_deformable(net, num_filter, num_group=1, act_type='relu', name=''):
+    f = num_group * 18
+    conv_offset = mx.symbol.Convolution(name=name + '_conv_offset',
+                                        data=net,
+                                        num_filter=f,
+                                        pad=(1, 1),
+                                        kernel=(3, 3),
+                                        stride=(1, 1))
+    net = mx.contrib.symbol.DeformableConvolution(
+        name=name + "_conv",
+        data=net,
+        offset=conv_offset,
+        num_filter=num_filter,
+        pad=(1, 1),
+        kernel=(3, 3),
+        num_deformable_group=num_group,
+        stride=(1, 1),
+        no_bias=False)
+    net = mx.sym.BatchNorm(data=net,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=0.9,
+                           name=name + '_bn')
+    if len(act_type) > 0:
+        net = mx.symbol.Activation(data=net,
+                                   act_type=act_type,
+                                   name=name + '_act')
+    return net
+
+
+def ssh_detection_module(body, num_filter, filter_in, name):
+    conv3x3 = conv_act_layer(body,
+                             name + '_conv1',
+                             num_filter,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             separable=True,
+                             filter_in=filter_in)
+    conv5x5, conv7x7 = ssh_context_module(body, num_filter // 2, filter_in,
+                                          name + '_context')
+    ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7],
+                        dim=1,
+                        name=name + '_concat')
+    ret = mx.symbol.Activation(data=ret,
+                               act_type='relu',
+                               name=name + '_concat_relu')
+    if config.USE_DCN == 1:
+        ret = conv_deformable(ret,
+                              num_filter=num_filter * 2,
+                              name=name + '_concat_dcn')
+    elif config.USE_DCN == 2:
+        ret = conv_deformable2(ret,
+                               num_filter=num_filter * 2,
+                               name=name + '_concat_dcn')
+    return ret
+
+
+def get_resnet_conv(data, sym):
+    all_layers = sym.get_internals()
+    isize = 640
+    _, out_shape, _ = all_layers.infer_shape(data=(1, 3, isize, isize))
+    last_entry = None
+    c1 = None
+    c2 = None
+    c3 = None
+    #print(len(all_layers), len(out_shape))
+    #print(all_layers.__class__)
+    outputs = all_layers.list_outputs()
+    #print(outputs.__class__, len(outputs))
+    count = len(outputs)
+    for i in range(count):
+        name = outputs[i]
+        shape = out_shape[i]
+        if not name.endswith('_output'):
+            continue
+        if len(shape) != 4:
+            continue
+        print(name, shape)
+        if c1 is None and shape[2] == isize // 16:
+            cname = last_entry[0]
+            print('c1', last_entry)
+            c1 = all_layers[cname]
+        if c2 is None and shape[2] == isize // 32:
+            cname = last_entry[0]
+            print('c2', last_entry)
+            c2 = all_layers[cname]
+        if shape[2] == isize // 32:
+            c3 = all_layers[name]
+            print('c3', name, shape)
+
+        last_entry = (name, shape)
+
+    c1_filter = -1
+    c2_filter = -1
+    c3_filter = -1
+
+    F1 = 256
+    F2 = 256
+    _bwm = 1.0
+    if config.NET_MODE == 0:
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_m2_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        #conv5_128_up = mx.symbol.Deconvolution(data=conv5_128, num_filter=F2, kernel=(4,4),  stride=(2, 2), pad=(1,1),
+        #    num_group = F2, no_bias = True, attr={'__lr_mult__': '0.0', '__wd_mult__': '0.0'},
+        #    name='ssh_m2_red_upsampling')
+        c2_up = mx.symbol.UpSampling(c2_lateral,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+
+        c1 = c1_lateral + c2_up
+
+        c1 = conv_act_layer(c1,
+                            'ssh_m1_conv',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 1:
+        c3_lateral = conv_act_layer(c3,
+                                    'ssh_c3_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.UpSampling(c3_lateral,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_c3_up',
+                                     num_args=1)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_up = mx.symbol.UpSampling(c2,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+
+        c1 = c1_lateral + c2_up
+
+        c1 = conv_act_layer(c1,
+                            'ssh_m1_conv',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 2:
+        n1 = ssh_detection_module(c1, F2, F2, 'ssh_n1_det')
+        n2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_n2_det')
+        n3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_n3_det')
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c3_up = mx.symbol.UpSampling(c3,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_c3_up',
+                                     num_args=1)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_up = mx.symbol.UpSampling(c2,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 3:
+        #c3 = conv_act_layer(c3, 'ssh_c3_lateral',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c3 = ssh_detection_module(c3, F2 // 2, c3_filter, 'ssh_c3_lateral')
+        c3_up = mx.symbol.UpSampling(c3,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_c3_up',
+                                     num_args=1)
+        #c2_lateral = conv_act_layer(c2, 'ssh_c2_lateral',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c2_lateral = ssh_detection_module(c2, F2 // 2, c2_filter,
+                                          'ssh_c2_lateral')
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        #c1_lateral = conv_act_layer(c1, 'ssh_m1_red_conv',
+        #    F2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), act_type='relu', bias_wd_mult=_bwm)
+        c1_lateral = ssh_detection_module(c1, F2 // 2, c1_filter,
+                                          'ssh_c1_lateral')
+        c2_up = mx.symbol.UpSampling(c2,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 4:
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c3_up = mx.symbol.UpSampling(c3,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_c3_up',
+                                     num_args=1)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_up = mx.symbol.UpSampling(c2,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2 // 2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1 // 2, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1 // 2, c3_filter, 'ssh_m3_det')
+    elif config.NET_MODE == 5:
+        c3 = conv_act_layer_dw(c3,
+                               'ssh_c3_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c3 = conv_act_layer(c3,
+                            'ssh_c3_lateral',
+                            F2,
+                            kernel=(1, 1),
+                            pad=(0, 0),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c3_up = mx.symbol.UpSampling(c3,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_c3_up',
+                                     num_args=1)
+        c2 = conv_act_layer_dw(c2,
+                               'ssh_c2_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c2_lateral = conv_act_layer(c2,
+                                    'ssh_c2_lateral',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c3_up = mx.symbol.Crop(*[c3_up, c2_lateral])
+        c2 = c2_lateral + c3_up
+        c2 = conv_act_layer(c2,
+                            'ssh_c2_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+        c1 = conv_act_layer_dw(c1,
+                               'ssh_c1_lateral_m',
+                               F2,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               bias_wd_mult=_bwm)
+        c1_lateral = conv_act_layer(c1,
+                                    'ssh_m1_red_conv',
+                                    F2,
+                                    kernel=(1, 1),
+                                    pad=(0, 0),
+                                    stride=(1, 1),
+                                    act_type='relu',
+                                    bias_wd_mult=_bwm)
+        c2_up = mx.symbol.UpSampling(c2,
+                                     scale=2,
+                                     sample_type='nearest',
+                                     workspace=512,
+                                     name='ssh_m2_red_up',
+                                     num_args=1)
+        #conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        c2_up = mx.symbol.Crop(*[c2_up, c1_lateral])
+        c1 = c1_lateral + c2_up
+        c1 = conv_act_layer(c1,
+                            'ssh_c1_aggr',
+                            F2,
+                            kernel=(3, 3),
+                            pad=(1, 1),
+                            stride=(1, 1),
+                            act_type='relu',
+                            bias_wd_mult=_bwm)
+
+        m1 = ssh_detection_module(c1, F2, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(c2, F1, c2_filter, 'ssh_m2_det')
+        m3 = ssh_detection_module(c3, F1, c3_filter, 'ssh_m3_det')
+
+    return {8: m1, 16: m2, 32: m3}, {8: n1, 16: n2, 32: n3}
+
+
+def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0):
+    A = config.NUM_ANCHORS
+    bbox_pred_len = 4
+    landmark_pred_len = 10
+    if config.USE_BLUR:
+        bbox_pred_len = 5
+    if config.USE_OCCLUSION:
+        landmark_pred_len = 15
+    ret_group = []
+    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
+    label = mx.symbol.Variable(name='%s_label_stride%d' % (prefix, stride))
+    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d' %
+                                     (prefix, stride))
+    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d' %
+                                     (prefix, stride))
+    if landmark:
+        landmark_target = mx.symbol.Variable(
+            name='%s_landmark_target_stride%d' % (prefix, stride))
+        landmark_weight = mx.symbol.Variable(
+            name='%s_landmark_weight_stride%d' % (prefix, stride))
+    rpn_relu = conv_fpn_feat[stride]
+    maxout_stat = 0
+    if config.USE_MAXOUT >= 1 and stride == config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 1
+    if config.USE_MAXOUT >= 2 and stride != config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 2
+
+    if maxout_stat == 0:
+        rpn_cls_score = conv_only(rpn_relu,
+                                  '%s_rpn_cls_score_stride%d' %
+                                  (prefix, stride),
+                                  2 * num_anchors,
+                                  kernel=(1, 1),
+                                  pad=(0, 0),
+                                  stride=(1, 1))
+    elif maxout_stat == 1:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+    else:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_only(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1))
+            rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+
+    rpn_bbox_pred = conv_only(rpn_relu,
+                              '%s_rpn_bbox_pred_stride%d' % (prefix, stride),
+                              bbox_pred_len * num_anchors,
+                              kernel=(1, 1),
+                              pad=(0, 0),
+                              stride=(1, 1))
+
+    # prepare rpn data
+    if not config.FBN:
+        rpn_cls_score_reshape = mx.symbol.Reshape(
+            data=rpn_cls_score,
+            shape=(0, 2, -1),
+            name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+    else:
+        rpn_cls_score_reshape = mx.symbol.Reshape(
+            data=rpn_cls_score,
+            shape=(0, 2, -1),
+            name="%s_rpn_cls_score_reshape_stride%s_pre" % (prefix, stride))
+        rpn_cls_score_reshape = mx.symbol.BatchNorm(
+            rpn_cls_score_reshape,
+            fix_gamma=True,
+            eps=2e-5,
+            name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+
+    rpn_bbox_pred_reshape = mx.symbol.Reshape(
+        data=rpn_bbox_pred,
+        shape=(0, 0, -1),
+        name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix, stride))
+    if landmark:
+        rpn_landmark_pred = conv_only(rpn_relu,
+                                      '%s_rpn_landmark_pred_stride%d' %
+                                      (prefix, stride),
+                                      landmark_pred_len * num_anchors,
+                                      kernel=(1, 1),
+                                      pad=(0, 0),
+                                      stride=(1, 1))
+        rpn_landmark_pred_reshape = mx.symbol.Reshape(
+            data=rpn_landmark_pred,
+            shape=(0, 0, -1),
+            name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix, stride))
+
+    if config.TRAIN.RPN_ENABLE_OHEM >= 2:
+        label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3',
+                                             stride=int(stride),
+                                             network=config.network,
+                                             dataset=config.dataset,
+                                             prefix=prefix,
+                                             cls_score=rpn_cls_score_reshape,
+                                             labels=label)
+
+        _bbox_weight = mx.sym.tile(anchor_weight, (1, 1, bbox_pred_len))
+        _bbox_weight = _bbox_weight.reshape(
+            (0, -1, A * bbox_pred_len)).transpose((0, 2, 1))
+        bbox_weight = mx.sym.elemwise_mul(bbox_weight,
+                                          _bbox_weight,
+                                          name='%s_bbox_weight_mul_stride%s' %
+                                          (prefix, stride))
+
+        if landmark:
+            _landmark_weight = mx.sym.tile(anchor_weight,
+                                           (1, 1, landmark_pred_len))
+            _landmark_weight = _landmark_weight.reshape(
+                (0, -1, A * landmark_pred_len)).transpose((0, 2, 1))
+            landmark_weight = mx.sym.elemwise_mul(
+                landmark_weight,
+                _landmark_weight,
+                name='%s_landmark_weight_mul_stride%s' % (prefix, stride))
+        #if not config.FACE_LANDMARK:
+        #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
+        #else:
+        #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
+    #cls loss
+    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
+                                           label=label,
+                                           multi_output=True,
+                                           normalization='valid',
+                                           use_ignore=True,
+                                           ignore_label=-1,
+                                           grad_scale=lr_mult,
+                                           name='%s_rpn_cls_prob_stride%d' %
+                                           (prefix, stride))
+    ret_group.append(rpn_cls_prob)
+    ret_group.append(mx.sym.BlockGrad(label))
+
+    #bbox loss
+    bbox_diff = rpn_bbox_pred_reshape - bbox_target
+    bbox_diff = bbox_diff * bbox_weight
+    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_' %
+                                         (prefix, stride),
+                                         scalar=3.0,
+                                         data=bbox_diff)
+    rpn_bbox_loss = mx.sym.MakeLoss(
+        name='%s_rpn_bbox_loss_stride%d' % (prefix, stride),
+        data=rpn_bbox_loss_,
+        grad_scale=1.0 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+    ret_group.append(rpn_bbox_loss)
+    ret_group.append(mx.sym.BlockGrad(bbox_weight))
+
+    #landmark loss
+    if landmark:
+        landmark_diff = rpn_landmark_pred_reshape - landmark_target
+        landmark_diff = landmark_diff * landmark_weight
+        rpn_landmark_loss_ = mx.symbol.smooth_l1(
+            name='%s_rpn_landmark_loss_stride%d_' % (prefix, stride),
+            scalar=3.0,
+            data=landmark_diff)
+        rpn_landmark_loss = mx.sym.MakeLoss(
+            name='%s_rpn_landmark_loss_stride%d' % (prefix, stride),
+            data=rpn_landmark_loss_,
+            grad_scale=0.5 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+        ret_group.append(rpn_landmark_loss)
+        ret_group.append(mx.sym.BlockGrad(landmark_weight))
+    return ret_group
+
+
+def get_resnet_train(sym):
+    return get_sym_train(sym)
+    #data = mx.symbol.Variable(name="data")
+    ## shared convolutional layers
+    #conv_fpn_feat, conv_fpn_feat2 = get_resnet_conv(data, sym)
+    #ret_group = []
+    #for stride in config.RPN_FEAT_STRIDE:
+    #  ret = get_out(conv_fpn_feat, 'face', stride, config.FACE_LANDMARK, lr_mult=1.0)
+    #  ret_group += ret
+    #  if config.HEAD_BOX:
+    #    ret = get_out(conv_fpn_feat2, 'head', stride, False, lr_mult=1.0)
+    #    ret_group += ret
+
+    #return mx.sym.Group(ret_group)
diff --git a/detection/retinaface/rcnn/symbol/symbol_ssh.py b/detection/retinaface/rcnn/symbol/symbol_ssh.py
new file mode 100644
index 000000000..f8be30d0a
--- /dev/null
+++ b/detection/retinaface/rcnn/symbol/symbol_ssh.py
@@ -0,0 +1,725 @@
+import mxnet as mx
+import numpy as np
+from rcnn.config import config
+from rcnn.PY_OP import rpn_fpn_ohem3
+FPN = False
+USE_DCN = False
+
+def conv_act_layer(from_layer, name, num_filter, kernel=(1,1), pad=(0,0), \
+    stride=(1,1), act_type="relu", bias_wd_mult=0.0, dcn=False):
+
+    weight = mx.symbol.Variable(name="{}_weight".format(name),
+                                init=mx.init.Normal(0.01),
+                                attr={'__lr_mult__': '1.0'})
+    bias = mx.symbol.Variable(name="{}_bias".format(name),
+                              init=mx.init.Constant(0.0),
+                              attr={
+                                  '__lr_mult__': '2.0',
+                                  '__wd_mult__': str(bias_wd_mult)
+                              })
+    if not dcn:
+        conv = mx.symbol.Convolution(data=from_layer, kernel=kernel, pad=pad, \
+            stride=stride, num_filter=num_filter, name="{}".format(name), weight = weight, bias=bias)
+    else:
+        assert kernel[0] == 3 and kernel[1] == 3
+        num_group = 1
+        f = num_group * 18
+        offset_weight = mx.symbol.Variable(
+            name="{}_offset_weight".format(name),
+            init=mx.init.Constant(0.0),
+            attr={'__lr_mult__': '1.0'})
+        offset_bias = mx.symbol.Variable(name="{}_offset_bias".format(name),
+                                         init=mx.init.Constant(0.0),
+                                         attr={
+                                             '__lr_mult__': '2.0',
+                                             '__wd_mult__': str(bias_wd_mult)
+                                         })
+        conv_offset = mx.symbol.Convolution(name=name + '_offset',
+                                            data=from_layer,
+                                            weight=offset_weight,
+                                            bias=offset_bias,
+                                            num_filter=f,
+                                            pad=(1, 1),
+                                            kernel=(3, 3),
+                                            stride=(1, 1))
+        conv = mx.contrib.symbol.DeformableConvolution(
+            name=name,
+            data=from_layer,
+            offset=conv_offset,
+            weight=weight,
+            bias=bias,
+            num_filter=num_filter,
+            pad=(1, 1),
+            kernel=(3, 3),
+            num_deformable_group=num_group,
+            stride=(1, 1),
+            no_bias=False)
+    if len(act_type) > 0:
+        relu = mx.symbol.Activation(data=conv, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    else:
+        relu = conv
+    return relu
+
+
+def ssh_context_module(body, num_filters, name):
+    conv_dimred = conv_act_layer(body,
+                                 name + '_conv1',
+                                 num_filters,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 dcn=False)
+    conv5x5 = conv_act_layer(conv_dimred,
+                             name + '_conv2',
+                             num_filters,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             dcn=USE_DCN)
+    conv7x7_1 = conv_act_layer(conv_dimred,
+                               name + '_conv3_1',
+                               num_filters,
+                               kernel=(3, 3),
+                               pad=(1, 1),
+                               stride=(1, 1),
+                               act_type='relu',
+                               dcn=False)
+    conv7x7 = conv_act_layer(conv7x7_1,
+                             name + '_conv3_2',
+                             num_filters,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             dcn=USE_DCN)
+    return (conv5x5, conv7x7)
+
+
+def ssh_detection_module(body, num_filters, name):
+    conv3x3 = conv_act_layer(body,
+                             name + '_conv1',
+                             num_filters,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='',
+                             dcn=USE_DCN)
+    conv5x5, conv7x7 = ssh_context_module(body, num_filters // 2,
+                                          name + '_context')
+    ret = mx.sym.concat(*[conv3x3, conv5x5, conv7x7],
+                        dim=1,
+                        name=name + '_concat')
+    ret = mx.symbol.Activation(data=ret,
+                               act_type='relu',
+                               name=name + '_concat_relu')
+    return ret
+
+
+def conv_bn(input, filter, ksize, stride, padding, act_type='relu', name=''):
+    conv = mx.symbol.Convolution(data=input, kernel=(ksize,ksize), pad=(padding,padding), \
+        stride=(stride,stride), num_filter=filter, name=name+"_conv")
+    ret = mx.sym.BatchNorm(data=conv,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=0.9,
+                           name=name + '_bn')
+    if act_type is not None:
+        ret = mx.symbol.Activation(data=ret, act_type=act_type, \
+            name="{}_{}".format(name, act_type))
+    return ret
+
+
+def cpm(input, name):
+    # residual
+    branch1 = conv_bn(input,
+                      1024,
+                      1,
+                      1,
+                      0,
+                      act_type=None,
+                      name=name + "_branch1")
+    branch2a = conv_bn(input,
+                       256,
+                       1,
+                       1,
+                       0,
+                       act_type='relu',
+                       name=name + "_branch2a")
+    branch2b = conv_bn(branch2a,
+                       256,
+                       3,
+                       1,
+                       1,
+                       act_type='relu',
+                       name=name + "_branch2b")
+    branch2c = conv_bn(branch2b,
+                       1024,
+                       1,
+                       1,
+                       0,
+                       act_type=None,
+                       name=name + "_branch2c")
+    sum = branch1 + branch2c
+    rescomb = mx.symbol.Activation(data=sum,
+                                   act_type='relu',
+                                   name="%s_relu2" % (name))
+
+    ssh_out = ssh_detection_module(rescomb, 256, name=name + "_ssh")
+    return ssh_out
+
+
+def get_feat_down(conv_feat):
+    #P5 = mx.symbol.Convolution(data=conv_feat[0], kernel=(1, 1), num_filter=256, name="P5_lateral")
+    P5 = conv_act_layer(conv_feat[0],
+                        'P5_lateral',
+                        256,
+                        kernel=(1, 1),
+                        pad=(0, 0),
+                        stride=(1, 1),
+                        act_type='relu')
+
+    # P5 2x upsampling + C4 = P4
+    P5_up = mx.symbol.UpSampling(P5,
+                                 scale=2,
+                                 sample_type='nearest',
+                                 workspace=512,
+                                 name='P5_upsampling',
+                                 num_args=1)
+    #P4_la   = mx.symbol.Convolution(data=conv_feat[1], kernel=(1, 1), num_filter=256, name="P4_lateral")
+    P4_la = conv_act_layer(conv_feat[1],
+                           'P4_lateral',
+                           256,
+                           kernel=(1, 1),
+                           pad=(0, 0),
+                           stride=(1, 1),
+                           act_type='relu')
+    P5_clip = mx.symbol.Crop(*[P5_up, P4_la], name="P4_clip")
+    P4 = mx.sym.ElementWiseSum(*[P5_clip, P4_la], name="P4_sum")
+    #P4      = mx.symbol.Convolution(data=P4, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P4_aggregate")
+    P4 = conv_act_layer(P4,
+                        'P4_aggregate',
+                        256,
+                        kernel=(3, 3),
+                        pad=(1, 1),
+                        stride=(1, 1),
+                        act_type='relu')
+
+    # P4 2x upsampling + C3 = P3
+    P4_up = mx.symbol.UpSampling(P4,
+                                 scale=2,
+                                 sample_type='nearest',
+                                 workspace=512,
+                                 name='P4_upsampling',
+                                 num_args=1)
+    #P3_la   = mx.symbol.Convolution(data=conv_feat[2], kernel=(1, 1), num_filter=256, name="P3_lateral")
+    P3_la = conv_act_layer(conv_feat[2],
+                           'P3_lateral',
+                           256,
+                           kernel=(1, 1),
+                           pad=(0, 0),
+                           stride=(1, 1),
+                           act_type='relu')
+    P4_clip = mx.symbol.Crop(*[P4_up, P3_la], name="P3_clip")
+    P3 = mx.sym.ElementWiseSum(*[P4_clip, P3_la], name="P3_sum")
+    #P3      = mx.symbol.Convolution(data=P3, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P3_aggregate")
+    P3 = conv_act_layer(P3,
+                        'P3_aggregate',
+                        256,
+                        kernel=(3, 3),
+                        pad=(1, 1),
+                        stride=(1, 1),
+                        act_type='relu')
+
+    return P3, P4, P5
+
+
+def get_ssh_conv(data):
+    """
+    shared convolutional layers
+    :param data: Symbol
+    :return: Symbol
+    """
+    # group 1
+    #conv1_1 = mx.symbol.Convolution(
+    #    data=data, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_1")
+    #relu1_1 = mx.symbol.Activation(data=conv1_1, act_type="relu", name="relu1_1")
+    relu1_1 = conv_act_layer(data,
+                             'conv1_1',
+                             64,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv1_2 = mx.symbol.Convolution(
+    #    data=relu1_1, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_2")
+    #relu1_2 = mx.symbol.Activation(data=conv1_2, act_type="relu", name="relu1_2")
+    relu1_2 = conv_act_layer(relu1_1,
+                             'conv1_2',
+                             64,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    pool1 = mx.symbol.Pooling(data=relu1_2,
+                              pool_type="max",
+                              kernel=(2, 2),
+                              stride=(2, 2),
+                              name="pool1")
+    # group 2
+    #conv2_1 = mx.symbol.Convolution(
+    #    data=pool1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_1")
+    #relu2_1 = mx.symbol.Activation(data=conv2_1, act_type="relu", name="relu2_1")
+    relu2_1 = conv_act_layer(pool1,
+                             'conv2_1',
+                             128,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv2_2 = mx.symbol.Convolution(
+    #    data=relu2_1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_2")
+    #relu2_2 = mx.symbol.Activation(data=conv2_2, act_type="relu", name="relu2_2")
+    relu2_2 = conv_act_layer(relu2_1,
+                             'conv2_2',
+                             128,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    pool2 = mx.symbol.Pooling(data=relu2_2,
+                              pool_type="max",
+                              kernel=(2, 2),
+                              stride=(2, 2),
+                              name="pool2")
+    # group 3
+    #conv3_1 = mx.symbol.Convolution(
+    #    data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_1")
+    #relu3_1 = mx.symbol.Activation(data=conv3_1, act_type="relu", name="relu3_1")
+    relu3_1 = conv_act_layer(pool2,
+                             'conv3_1',
+                             256,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv3_2 = mx.symbol.Convolution(
+    #    data=relu3_1, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_2")
+    #relu3_2 = mx.symbol.Activation(data=conv3_2, act_type="relu", name="relu3_2")
+    relu3_2 = conv_act_layer(relu3_1,
+                             'conv3_2',
+                             256,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv3_3 = mx.symbol.Convolution(
+    #    data=relu3_2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_3")
+    #relu3_3 = mx.symbol.Activation(data=conv3_3, act_type="relu", name="relu3_3")
+    relu3_3 = conv_act_layer(relu3_2,
+                             'conv3_3',
+                             256,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    pool3 = mx.symbol.Pooling(data=relu3_3,
+                              pool_type="max",
+                              kernel=(2, 2),
+                              stride=(2, 2),
+                              name="pool3")
+    # group 4
+    #conv4_1 = mx.symbol.Convolution(
+    #    data=pool3, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_1")
+    #relu4_1 = mx.symbol.Activation(data=conv4_1, act_type="relu", name="relu4_1")
+    relu4_1 = conv_act_layer(pool3,
+                             'conv4_1',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv4_2 = mx.symbol.Convolution(
+    #    data=relu4_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_2")
+    #relu4_2 = mx.symbol.Activation(data=conv4_2, act_type="relu", name="relu4_2")
+    relu4_2 = conv_act_layer(relu4_1,
+                             'conv4_2',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv4_3 = mx.symbol.Convolution(
+    #    data=relu4_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_3")
+    #relu4_3 = mx.symbol.Activation(data=conv4_3, act_type="relu", name="relu4_3")
+    relu4_3 = conv_act_layer(relu4_2,
+                             'conv4_3',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    pool4 = mx.symbol.Pooling(data=relu4_3,
+                              pool_type="max",
+                              kernel=(2, 2),
+                              stride=(2, 2),
+                              name="pool4")
+    # group 5
+    #conv5_1 = mx.symbol.Convolution(
+    #    data=pool4, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_1")
+    #relu5_1 = mx.symbol.Activation(data=conv5_1, act_type="relu", name="relu5_1")
+    relu5_1 = conv_act_layer(pool4,
+                             'conv5_1',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv5_2 = mx.symbol.Convolution(
+    #    data=relu5_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_2")
+    #relu5_2 = mx.symbol.Activation(data=conv5_2, act_type="relu", name="relu5_2")
+    relu5_2 = conv_act_layer(relu5_1,
+                             'conv5_2',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    #conv5_3 = mx.symbol.Convolution(
+    #    data=relu5_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_3")
+    #relu5_3 = mx.symbol.Activation(data=conv5_3, act_type="relu", name="relu5_3")
+    relu5_3 = conv_act_layer(relu5_2,
+                             'conv5_3',
+                             512,
+                             kernel=(3, 3),
+                             pad=(1, 1),
+                             stride=(1, 1),
+                             act_type='relu')
+    m3_pool = mx.sym.Pooling(data=relu5_3,
+                             kernel=(2, 2),
+                             stride=(2, 2),
+                             pad=(0, 0),
+                             pool_type='max')
+    if config.SSH_MODE <= 5:
+        #if FPN:
+        #  relu4_3, relu5_3, m3_pool = get_feat_down([m3_pool, relu5_3, relu4_3])
+
+        F1 = 256
+        F2 = 128
+        if config.SSH_MODE == 1:
+            F2 = 256
+        _bwm = 1.0
+        conv4_128 = conv_act_layer(relu4_3,
+                                   'ssh_m1_red_conv',
+                                   F2,
+                                   kernel=(1, 1),
+                                   pad=(0, 0),
+                                   stride=(1, 1),
+                                   act_type='relu',
+                                   bias_wd_mult=_bwm)
+        conv5_128 = conv_act_layer(relu5_3,
+                                   'ssh_m2_red_conv',
+                                   F2,
+                                   kernel=(1, 1),
+                                   pad=(0, 0),
+                                   stride=(1, 1),
+                                   act_type='relu',
+                                   bias_wd_mult=_bwm)
+        conv5_128_up = mx.symbol.Deconvolution(data=conv5_128,
+                                               num_filter=F2,
+                                               kernel=(4, 4),
+                                               stride=(2, 2),
+                                               pad=(1, 1),
+                                               num_group=F2,
+                                               no_bias=True,
+                                               attr={
+                                                   '__lr_mult__': '0.0',
+                                                   '__wd_mult__': '0.0'
+                                               },
+                                               name='ssh_m2_red_upsampling')
+        #conv5_128_up = mx.symbol.UpSampling(conv5_128, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        #conv5_128_up = mx.symbol.Crop(*[conv5_128_up, conv4_128])
+
+        conv_sum = conv4_128 + conv5_128_up
+        #conv_sum = conv_1x1
+
+        m1_conv = conv_act_layer(conv_sum,
+                                 'ssh_m1_conv',
+                                 F2,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 bias_wd_mult=_bwm)
+        m1 = ssh_detection_module(m1_conv, F2, 'ssh_m1_det')
+        m2 = ssh_detection_module(relu5_3, F1, 'ssh_m2_det')
+        m3 = ssh_detection_module(m3_pool, F1, 'ssh_m3_det')
+        return {8: m1, 16: m2, 32: m3}
+    else:
+        F1 = 256
+        F2 = 256
+        _bwm = 1.0
+        conv4_128 = conv_act_layer(relu4_3,
+                                   'ssh_m1_red_conv',
+                                   F2,
+                                   kernel=(1, 1),
+                                   pad=(0, 0),
+                                   stride=(1, 1),
+                                   act_type='relu',
+                                   bias_wd_mult=_bwm)
+        conv5_128 = conv_act_layer(relu5_3,
+                                   'ssh_m2_red_conv',
+                                   F2,
+                                   kernel=(1, 1),
+                                   pad=(0, 0),
+                                   stride=(1, 1),
+                                   act_type='relu',
+                                   bias_wd_mult=_bwm)
+        conv5_128_up = mx.symbol.Deconvolution(data=conv5_128,
+                                               num_filter=F2,
+                                               kernel=(4, 4),
+                                               stride=(2, 2),
+                                               pad=(1, 1),
+                                               num_group=F2,
+                                               no_bias=True,
+                                               attr={
+                                                   '__lr_mult__': '0.0',
+                                                   '__wd_mult__': '0.0'
+                                               },
+                                               name='ssh_m2_red_upsampling')
+        #conv5_128_up = mx.symbol.UpSampling(conv5_128, scale=2, sample_type='nearest', workspace=512, name='ssh_m2_red_up', num_args=1)
+        conv4_128 = mx.symbol.Crop(*[conv4_128, conv5_128_up])
+        #conv5_128_up = mx.symbol.Crop(*[conv5_128_up, conv4_128])
+
+        conv_sum = conv4_128 + conv5_128_up
+        m1_conv = conv_act_layer(conv_sum,
+                                 'ssh_m1_conv',
+                                 F2,
+                                 kernel=(3, 3),
+                                 pad=(1, 1),
+                                 stride=(1, 1),
+                                 act_type='relu',
+                                 bias_wd_mult=_bwm)
+        m1 = cpm(m1_conv, 'ssh_m1_det')
+        m2 = cpm(relu5_3, 'ssh_m2_det')
+        m3 = cpm(m3_pool, 'ssh_m3_det')
+        return {8: m1, 16: m2, 32: m3}
+
+
+def get_out(conv_fpn_feat, prefix, stride, landmark=False, lr_mult=1.0):
+    A = config.NUM_ANCHORS
+    ret_group = []
+    num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
+    label = mx.symbol.Variable(name='%s_label_stride%d' % (prefix, stride))
+    bbox_target = mx.symbol.Variable(name='%s_bbox_target_stride%d' %
+                                     (prefix, stride))
+    bbox_weight = mx.symbol.Variable(name='%s_bbox_weight_stride%d' %
+                                     (prefix, stride))
+    if landmark:
+        landmark_target = mx.symbol.Variable(
+            name='%s_landmark_target_stride%d' % (prefix, stride))
+        landmark_weight = mx.symbol.Variable(
+            name='%s_landmark_weight_stride%d' % (prefix, stride))
+    rpn_relu = conv_fpn_feat[stride]
+    maxout_stat = 0
+    if config.USE_MAXOUT >= 1 and stride == config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 1
+    if config.USE_MAXOUT >= 2 and stride != config.RPN_FEAT_STRIDE[-1]:
+        maxout_stat = 2
+
+    if maxout_stat == 0:
+        rpn_cls_score = conv_act_layer(rpn_relu,
+                                       '%s_rpn_cls_score_stride%d' %
+                                       (prefix, stride),
+                                       2 * num_anchors,
+                                       kernel=(1, 1),
+                                       pad=(0, 0),
+                                       stride=(1, 1),
+                                       act_type='')
+    elif maxout_stat == 1:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_act_layer(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                act_type='')
+            rpn_cls_score_bg = mx.sym.max(rpn_cls_score_bg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_act_layer(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                act_type='')
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+    else:
+        cls_list = []
+        for a in range(num_anchors):
+            rpn_cls_score_bg = conv_act_layer(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_bg' % (prefix, stride, a),
+                1,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                act_type='')
+            cls_list.append(rpn_cls_score_bg)
+            rpn_cls_score_fg = conv_act_layer(
+                rpn_relu,
+                '%s_rpn_cls_score_stride%d_anchor%d_fg' % (prefix, stride, a),
+                3,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                act_type='')
+            rpn_cls_score_fg = mx.sym.max(rpn_cls_score_fg,
+                                          axis=1,
+                                          keepdims=True)
+            cls_list.append(rpn_cls_score_fg)
+        rpn_cls_score = mx.sym.concat(*cls_list,
+                                      dim=1,
+                                      name='%s_rpn_cls_score_stride%d' %
+                                      (prefix, stride))
+
+    rpn_bbox_pred = conv_act_layer(rpn_relu,
+                                   '%s_rpn_bbox_pred_stride%d' %
+                                   (prefix, stride),
+                                   4 * num_anchors,
+                                   kernel=(1, 1),
+                                   pad=(0, 0),
+                                   stride=(1, 1),
+                                   act_type='')
+
+    # prepare rpn data
+    rpn_cls_score_reshape = mx.symbol.Reshape(
+        data=rpn_cls_score,
+        shape=(0, 2, -1),
+        name="%s_rpn_cls_score_reshape_stride%s" % (prefix, stride))
+    rpn_bbox_pred_reshape = mx.symbol.Reshape(
+        data=rpn_bbox_pred,
+        shape=(0, 0, -1),
+        name="%s_rpn_bbox_pred_reshape_stride%s" % (prefix, stride))
+    if landmark:
+        rpn_landmark_pred = conv_act_layer(rpn_relu,
+                                           '%s_rpn_landmark_pred_stride%d' %
+                                           (prefix, stride),
+                                           10 * num_anchors,
+                                           kernel=(1, 1),
+                                           pad=(0, 0),
+                                           stride=(1, 1),
+                                           act_type='')
+        rpn_landmark_pred_reshape = mx.symbol.Reshape(
+            data=rpn_landmark_pred,
+            shape=(0, 0, -1),
+            name="%s_rpn_landmark_pred_reshape_stride%s" % (prefix, stride))
+
+    if config.TRAIN.RPN_ENABLE_OHEM >= 2:
+        label, anchor_weight = mx.sym.Custom(op_type='rpn_fpn_ohem3',
+                                             stride=int(stride),
+                                             network=config.network,
+                                             dataset=config.dataset,
+                                             prefix=prefix,
+                                             cls_score=rpn_cls_score_reshape,
+                                             labels=label)
+
+        _bbox_weight = mx.sym.tile(anchor_weight, (1, 1, 4))
+        _bbox_weight = _bbox_weight.reshape((0, -1, A * 4)).transpose(
+            (0, 2, 1))
+        bbox_weight = mx.sym.elemwise_mul(bbox_weight,
+                                          _bbox_weight,
+                                          name='%s_bbox_weight_mul_stride%s' %
+                                          (prefix, stride))
+
+        if landmark:
+            _landmark_weight = mx.sym.tile(anchor_weight, (1, 1, 10))
+            _landmark_weight = _landmark_weight.reshape(
+                (0, -1, A * 10)).transpose((0, 2, 1))
+            landmark_weight = mx.sym.elemwise_mul(
+                landmark_weight,
+                _landmark_weight,
+                name='%s_landmark_weight_mul_stride%s' % (prefix, stride))
+        #if not config.FACE_LANDMARK:
+        #  label, bbox_weight = mx.sym.Custom(op_type='rpn_fpn_ohem', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight , labels = label)
+        #else:
+        #  label, bbox_weight, landmark_weight = mx.sym.Custom(op_type='rpn_fpn_ohem2', stride=int(stride), cls_score=rpn_cls_score_reshape, bbox_weight = bbox_weight, landmark_weight=landmark_weight, labels = label)
+    #cls loss
+    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape,
+                                           label=label,
+                                           multi_output=True,
+                                           normalization='valid',
+                                           use_ignore=True,
+                                           ignore_label=-1,
+                                           grad_scale=lr_mult,
+                                           name='%s_rpn_cls_prob_stride%d' %
+                                           (prefix, stride))
+    ret_group.append(rpn_cls_prob)
+    ret_group.append(mx.sym.BlockGrad(label))
+
+    #bbox loss
+    bbox_diff = rpn_bbox_pred_reshape - bbox_target
+    bbox_diff = bbox_diff * bbox_weight
+    rpn_bbox_loss_ = mx.symbol.smooth_l1(name='%s_rpn_bbox_loss_stride%d_' %
+                                         (prefix, stride),
+                                         scalar=3.0,
+                                         data=bbox_diff)
+    rpn_bbox_loss = mx.sym.MakeLoss(
+        name='%s_rpn_bbox_loss_stride%d' % (prefix, stride),
+        data=rpn_bbox_loss_,
+        grad_scale=1.0 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+    ret_group.append(rpn_bbox_loss)
+    ret_group.append(mx.sym.BlockGrad(bbox_weight))
+
+    #landmark loss
+    if landmark:
+        landmark_diff = rpn_landmark_pred_reshape - landmark_target
+        landmark_diff = landmark_diff * landmark_weight
+        rpn_landmark_loss_ = mx.symbol.smooth_l1(
+            name='%s_rpn_landmark_loss_stride%d_' % (prefix, stride),
+            scalar=3.0,
+            data=landmark_diff)
+        rpn_landmark_loss = mx.sym.MakeLoss(
+            name='%s_rpn_landmark_loss_stride%d' % (prefix, stride),
+            data=rpn_landmark_loss_,
+            grad_scale=0.5 * lr_mult / (config.TRAIN.RPN_BATCH_SIZE))
+        ret_group.append(rpn_landmark_loss)
+        ret_group.append(mx.sym.BlockGrad(landmark_weight))
+    return ret_group
+
+
+def get_ssh_train():
+    """
+    Region Proposal Network with VGG
+    :return: Symbol
+    """
+    data = mx.symbol.Variable(name="data")
+
+    # shared convolutional layers
+    conv_fpn_feat = get_ssh_conv(data)
+    ret_group = []
+    for stride in config.RPN_FEAT_STRIDE:
+        ret = get_out(conv_fpn_feat,
+                      'face',
+                      stride,
+                      config.FACE_LANDMARK,
+                      lr_mult=1.0)
+        ret_group += ret
+        if config.HEAD_BOX:
+            ret = get_out(conv_fpn_feat, 'head', stride, False, lr_mult=1.0)
+            ret_group += ret
+
+    return mx.sym.Group(ret_group)
diff --git a/detection/retinaface/rcnn/tools/__init__.py b/detection/retinaface/rcnn/tools/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/detection/retinaface/rcnn/tools/reeval.py b/detection/retinaface/rcnn/tools/reeval.py
new file mode 100644
index 000000000..0903a06b4
--- /dev/null
+++ b/detection/retinaface/rcnn/tools/reeval.py
@@ -0,0 +1,68 @@
+import argparse
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+import os
+import mxnet as mx
+
+from ..logger import logger
+from ..config import config, default, generate_config
+from ..dataset import *
+
+
+def reeval(args):
+    # load imdb
+    imdb = eval(args.dataset)(args.image_set, args.root_path,
+                              args.dataset_path)
+
+    # load detection results
+    cache_file = os.path.join(imdb.cache_path, imdb.name, 'detections.pkl')
+    with open(cache_file) as f:
+        detections = pickle.load(f)
+
+    # eval
+    imdb.evaluate_detections(detections)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='imdb test')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # other
+    parser.add_argument('--no_shuffle',
+                        help='disable random shuffle',
+                        action='store_true')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger.info('Called with argument: %s' % args)
+    reeval(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/rcnn/tools/test_rcnn.py b/detection/retinaface/rcnn/tools/test_rcnn.py
new file mode 100644
index 000000000..d571ed60f
--- /dev/null
+++ b/detection/retinaface/rcnn/tools/test_rcnn.py
@@ -0,0 +1,161 @@
+import argparse
+import pprint
+import mxnet as mx
+
+from ..logger import logger
+from ..config import config, default, generate_config
+from ..symbol import *
+from ..dataset import *
+from ..core.loader import TestLoader
+from ..core.tester import Predictor, pred_eval
+from ..utils.load_model import load_param
+
+
+def test_rcnn(network, dataset, image_set, root_path, dataset_path, ctx,
+              prefix, epoch, vis, shuffle, has_rpn, proposal, thresh):
+    # set config
+    if has_rpn:
+        config.TEST.HAS_RPN = True
+
+    # print config
+    logger.info(pprint.pformat(config))
+
+    # load symbol and testing data
+    if has_rpn:
+        sym = eval('get_' + network + '_test')(num_classes=config.NUM_CLASSES,
+                                               num_anchors=config.NUM_ANCHORS)
+        imdb = eval(dataset)(image_set, root_path, dataset_path)
+        roidb = imdb.gt_roidb()
+    else:
+        sym = eval('get_' + network +
+                   '_rcnn_test')(num_classes=config.NUM_CLASSES)
+        imdb = eval(dataset)(image_set, root_path, dataset_path)
+        gt_roidb = imdb.gt_roidb()
+        roidb = eval('imdb.' + proposal + '_roidb')(gt_roidb)
+
+    # get test data iter
+    test_data = TestLoader(roidb,
+                           batch_size=1,
+                           shuffle=shuffle,
+                           has_rpn=has_rpn)
+
+    # load model
+    arg_params, aux_params = load_param(prefix,
+                                        epoch,
+                                        convert=True,
+                                        ctx=ctx,
+                                        process=True)
+
+    # infer shape
+    data_shape_dict = dict(test_data.provide_data)
+    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
+    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
+
+    # check parameters
+    for k in sym.list_arguments():
+        if k in data_shape_dict or 'label' in k:
+            continue
+        assert k in arg_params, k + ' not initialized'
+        assert arg_params[k].shape == arg_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
+    for k in sym.list_auxiliary_states():
+        assert k in aux_params, k + ' not initialized'
+        assert aux_params[k].shape == aux_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
+
+    # decide maximum shape
+    data_names = [k[0] for k in test_data.provide_data]
+    label_names = None
+    max_data_shape = [('data', (1, 3, max([v[0] for v in config.SCALES]),
+                                max([v[1] for v in config.SCALES])))]
+    if not has_rpn:
+        max_data_shape.append(
+            ('rois', (1, config.TEST.PROPOSAL_POST_NMS_TOP_N + 30, 5)))
+
+    # create predictor
+    predictor = Predictor(sym,
+                          data_names,
+                          label_names,
+                          context=ctx,
+                          max_data_shapes=max_data_shape,
+                          provide_data=test_data.provide_data,
+                          provide_label=test_data.provide_label,
+                          arg_params=arg_params,
+                          aux_params=aux_params)
+
+    # start detection
+    pred_eval(predictor, test_data, imdb, vis=vis, thresh=thresh)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Test a Fast R-CNN network')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.test_image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # testing
+    parser.add_argument('--prefix',
+                        help='model to test with',
+                        default=default.rcnn_prefix,
+                        type=str)
+    parser.add_argument('--epoch',
+                        help='model to test with',
+                        default=default.rcnn_epoch,
+                        type=int)
+    parser.add_argument('--gpu',
+                        help='GPU device to test with',
+                        default=0,
+                        type=int)
+    # rcnn
+    parser.add_argument('--vis',
+                        help='turn on visualization',
+                        action='store_true')
+    parser.add_argument('--thresh',
+                        help='valid detection threshold',
+                        default=1e-3,
+                        type=float)
+    parser.add_argument('--shuffle',
+                        help='shuffle data on visualization',
+                        action='store_true')
+    parser.add_argument('--has_rpn',
+                        help='generate proposals on the fly',
+                        action='store_true')
+    parser.add_argument('--proposal',
+                        help='can be ss for selective search or rpn',
+                        default='rpn',
+                        type=str)
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger.info('Called with argument: %s' % args)
+    ctx = mx.gpu(args.gpu)
+    test_rcnn(args.network, args.dataset, args.image_set, args.root_path,
+              args.dataset_path, ctx, args.prefix, args.epoch, args.vis,
+              args.shuffle, args.has_rpn, args.proposal, args.thresh)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/rcnn/tools/test_rpn.py b/detection/retinaface/rcnn/tools/test_rpn.py
new file mode 100644
index 000000000..e77040683
--- /dev/null
+++ b/detection/retinaface/rcnn/tools/test_rpn.py
@@ -0,0 +1,160 @@
+import argparse
+import pprint
+import mxnet as mx
+
+from ..logger import logger
+from ..config import config, default, generate_config
+from ..symbol import *
+from ..dataset import *
+from ..core.loader import TestLoader
+from ..core.tester import Predictor, generate_proposals, test_proposals
+from ..utils.load_model import load_param
+
+
+def test_rpn(network,
+             dataset,
+             image_set,
+             root_path,
+             dataset_path,
+             ctx,
+             prefix,
+             epoch,
+             vis,
+             shuffle,
+             thresh,
+             test_output=False):
+    # rpn generate proposal config
+    config.TEST.HAS_RPN = True
+
+    # print config
+    logger.info(pprint.pformat(config))
+
+    # load symbol
+    sym = eval('get_' + network + '_rpn_test')()
+
+    # load dataset and prepare imdb for training
+    imdb = eval(dataset)(image_set, root_path, dataset_path)
+    roidb = imdb.gt_roidb()
+    test_data = TestLoader(roidb,
+                           batch_size=1,
+                           shuffle=shuffle,
+                           has_rpn=True,
+                           withlabel=True)
+
+    # load model
+    arg_params, aux_params = load_param(prefix, epoch, convert=True, ctx=ctx)
+
+    # infer shape
+    data_shape_dict = dict(test_data.provide_data)
+    arg_shape, _, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
+    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
+
+    # check parameters
+    for k in sym.list_arguments():
+        if k in data_shape_dict or 'label' in k:
+            continue
+        assert k in arg_params, k + ' not initialized'
+        assert arg_params[k].shape == arg_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
+    for k in sym.list_auxiliary_states():
+        assert k in aux_params, k + ' not initialized'
+        assert aux_params[k].shape == aux_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
+
+    # decide maximum shape
+    data_names = [k[0] for k in test_data.provide_data]
+    label_names = None if test_data.provide_label is None else [
+        k[0] for k in test_data.provide_label
+    ]
+    max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]),
+                                max([v[1] for v in config.SCALES])))]
+
+    # create predictor
+    predictor = Predictor(sym,
+                          data_names,
+                          label_names,
+                          context=ctx,
+                          max_data_shapes=max_data_shape,
+                          provide_data=test_data.provide_data,
+                          provide_label=test_data.provide_label,
+                          arg_params=arg_params,
+                          aux_params=aux_params)
+
+    # start testing
+    if not test_output:
+        imdb_boxes = generate_proposals(predictor,
+                                        test_data,
+                                        imdb,
+                                        vis=vis,
+                                        thresh=thresh)
+        imdb.evaluate_recall(roidb, candidate_boxes=imdb_boxes)
+    else:
+        test_proposals(predictor, test_data, imdb, roidb, vis=vis)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Test a Region Proposal Network')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.test_image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # testing
+    parser.add_argument('--prefix',
+                        help='model to test with',
+                        default=default.rpn_prefix,
+                        type=str)
+    parser.add_argument('--epoch',
+                        help='model to test with',
+                        default=default.rpn_epoch,
+                        type=int)
+    # rpn
+    parser.add_argument('--gpu',
+                        help='GPU device to test with',
+                        default=0,
+                        type=int)
+    parser.add_argument('--vis',
+                        help='turn on visualization',
+                        action='store_true')
+    parser.add_argument('--thresh',
+                        help='rpn proposal threshold',
+                        default=0,
+                        type=float)
+    parser.add_argument('--shuffle',
+                        help='shuffle data on visualization',
+                        action='store_true')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger.info('Called with argument: %s' % args)
+    ctx = mx.gpu(args.gpu)
+    test_rpn(args.network, args.dataset, args.image_set, args.root_path,
+             args.dataset_path, ctx, args.prefix, args.epoch, args.vis,
+             args.shuffle, args.thresh)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/rcnn/tools/train_rcnn.py b/detection/retinaface/rcnn/tools/train_rcnn.py
new file mode 100644
index 000000000..167a3b0b1
--- /dev/null
+++ b/detection/retinaface/rcnn/tools/train_rcnn.py
@@ -0,0 +1,286 @@
+import argparse
+import pprint
+import mxnet as mx
+
+from ..logger import logger
+from ..config import config, default, generate_config
+from ..symbol import *
+from ..core import callback, metric
+from ..core.loader import ROIIter
+from ..core.module import MutableModule
+from ..processing.bbox_regression import add_bbox_regression_targets
+from ..utils.load_data import load_proposal_roidb, merge_roidb, filter_roidb
+from ..utils.load_model import load_param
+
+
+def train_rcnn(network, dataset, image_set, root_path, dataset_path, frequent,
+               kvstore, work_load_list, no_flip, no_shuffle, resume, ctx,
+               pretrained, epoch, prefix, begin_epoch, end_epoch, train_shared,
+               lr, lr_step, proposal):
+    # set up config
+    config.TRAIN.BATCH_IMAGES = 2
+    config.TRAIN.BATCH_ROIS = 128
+    if proposal == 'ss':
+        config.TRAIN.BG_THRESH_LO = 0.1  # reproduce Fast R-CNN
+
+    # load symbol
+    sym = eval('get_' + network + '_rcnn')(num_classes=config.NUM_CLASSES)
+
+    # setup multi-gpu
+    batch_size = len(ctx)
+    input_batch_size = config.TRAIN.BATCH_IMAGES * batch_size
+
+    # print config
+    logger.info(pprint.pformat(config))
+
+    # load dataset and prepare imdb for training
+    image_sets = [iset for iset in image_set.split('+')]
+    roidbs = [
+        load_proposal_roidb(dataset,
+                            image_set,
+                            root_path,
+                            dataset_path,
+                            proposal=proposal,
+                            append_gt=True,
+                            flip=not no_flip) for image_set in image_sets
+    ]
+    roidb = merge_roidb(roidbs)
+    roidb = filter_roidb(roidb)
+    means, stds = add_bbox_regression_targets(roidb)
+
+    # load training data
+    train_data = ROIIter(roidb,
+                         batch_size=input_batch_size,
+                         shuffle=not no_shuffle,
+                         ctx=ctx,
+                         work_load_list=work_load_list,
+                         aspect_grouping=config.TRAIN.ASPECT_GROUPING)
+
+    # infer max shape
+    max_data_shape = [('data', (input_batch_size, 3,
+                                max([v[0] for v in config.SCALES]),
+                                max([v[1] for v in config.SCALES])))]
+    logger.info('providing maximum shape %s' % max_data_shape)
+
+    # infer shape
+    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
+    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
+    out_shape_dict = dict(zip(sym.list_outputs(), out_shape))
+    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
+    logger.info('output shape %s' % pprint.pformat(out_shape_dict))
+
+    # load and initialize params
+    if resume:
+        arg_params, aux_params = load_param(prefix, begin_epoch, convert=True)
+    else:
+        arg_params, aux_params = load_param(pretrained, epoch, convert=True)
+        arg_params['cls_score_weight'] = mx.random.normal(
+            0, 0.01, shape=arg_shape_dict['cls_score_weight'])
+        arg_params['cls_score_bias'] = mx.nd.zeros(
+            shape=arg_shape_dict['cls_score_bias'])
+        arg_params['bbox_pred_weight'] = mx.random.normal(
+            0, 0.001, shape=arg_shape_dict['bbox_pred_weight'])
+        arg_params['bbox_pred_bias'] = mx.nd.zeros(
+            shape=arg_shape_dict['bbox_pred_bias'])
+
+    # check parameter shapes
+    for k in sym.list_arguments():
+        if k in data_shape_dict:
+            continue
+        assert k in arg_params, k + ' not initialized'
+        assert arg_params[k].shape == arg_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
+    for k in sym.list_auxiliary_states():
+        assert k in aux_params, k + ' not initialized'
+        assert aux_params[k].shape == aux_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
+
+    # prepare training
+    # create solver
+    data_names = [k[0] for k in train_data.provide_data]
+    label_names = [k[0] for k in train_data.provide_label]
+    if train_shared:
+        fixed_param_prefix = config.FIXED_PARAMS_SHARED
+    else:
+        fixed_param_prefix = config.FIXED_PARAMS
+    mod = MutableModule(sym,
+                        data_names=data_names,
+                        label_names=label_names,
+                        logger=logger,
+                        context=ctx,
+                        work_load_list=work_load_list,
+                        max_data_shapes=max_data_shape,
+                        fixed_param_prefix=fixed_param_prefix)
+
+    # decide training params
+    # metric
+    eval_metric = metric.RCNNAccMetric()
+    cls_metric = metric.RCNNLogLossMetric()
+    bbox_metric = metric.RCNNL1LossMetric()
+    eval_metrics = mx.metric.CompositeEvalMetric()
+    for child_metric in [eval_metric, cls_metric, bbox_metric]:
+        eval_metrics.add(child_metric)
+    # callback
+    batch_end_callback = mx.callback.Speedometer(train_data.batch_size,
+                                                 frequent=frequent,
+                                                 auto_reset=False)
+    epoch_end_callback = callback.do_checkpoint(prefix, means, stds)
+    # decide learning rate
+    base_lr = lr
+    lr_factor = 0.1
+    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
+    lr_epoch_diff = [
+        epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch
+    ]
+    lr = base_lr * (lr_factor**(len(lr_epoch) - len(lr_epoch_diff)))
+    lr_iters = [
+        int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff
+    ]
+    logger.info('lr %f lr_epoch_diff %s lr_iters %s' %
+                (lr, lr_epoch_diff, lr_iters))
+    lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
+    # optimizer
+    optimizer_params = {
+        'momentum': 0.9,
+        'wd': 0.0005,
+        'learning_rate': lr,
+        'lr_scheduler': lr_scheduler,
+        'rescale_grad': (1.0 / batch_size),
+        'clip_gradient': 5
+    }
+
+    # train
+    mod.fit(train_data,
+            eval_metric=eval_metrics,
+            epoch_end_callback=epoch_end_callback,
+            batch_end_callback=batch_end_callback,
+            kvstore=kvstore,
+            optimizer='sgd',
+            optimizer_params=optimizer_params,
+            arg_params=arg_params,
+            aux_params=aux_params,
+            begin_epoch=begin_epoch,
+            num_epoch=end_epoch)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train a Fast R-CNN Network')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # training
+    parser.add_argument('--frequent',
+                        help='frequency of logging',
+                        default=default.frequent,
+                        type=int)
+    parser.add_argument('--kvstore',
+                        help='the kv-store type',
+                        default=default.kvstore,
+                        type=str)
+    parser.add_argument('--work_load_list',
+                        help='work load for different devices',
+                        default=None,
+                        type=list)
+    parser.add_argument('--no_flip',
+                        help='disable flip images',
+                        action='store_true')
+    parser.add_argument('--no_shuffle',
+                        help='disable random shuffle',
+                        action='store_true')
+    parser.add_argument('--resume',
+                        help='continue training',
+                        action='store_true')
+    # rcnn
+    parser.add_argument('--gpus',
+                        help='GPU device to train with',
+                        default='0',
+                        type=str)
+    parser.add_argument('--pretrained',
+                        help='pretrained model prefix',
+                        default=default.pretrained,
+                        type=str)
+    parser.add_argument('--pretrained_epoch',
+                        help='pretrained model epoch',
+                        default=default.pretrained_epoch,
+                        type=int)
+    parser.add_argument('--prefix',
+                        help='new model prefix',
+                        default=default.rcnn_prefix,
+                        type=str)
+    parser.add_argument('--begin_epoch',
+                        help='begin epoch of training',
+                        default=0,
+                        type=int)
+    parser.add_argument('--end_epoch',
+                        help='end epoch of training',
+                        default=default.rcnn_epoch,
+                        type=int)
+    parser.add_argument('--lr',
+                        help='base learning rate',
+                        default=default.rcnn_lr,
+                        type=float)
+    parser.add_argument('--lr_step',
+                        help='learning rate steps (in epoch)',
+                        default=default.rcnn_lr_step,
+                        type=str)
+    parser.add_argument('--train_shared',
+                        help='second round train shared params',
+                        action='store_true')
+    parser.add_argument('--proposal',
+                        help='can be ss for selective search or rpn',
+                        default='rpn',
+                        type=str)
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger.info('Called with argument: %s' % args)
+    ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
+    train_rcnn(args.network,
+               args.dataset,
+               args.image_set,
+               args.root_path,
+               args.dataset_path,
+               args.frequent,
+               args.kvstore,
+               args.work_load_list,
+               args.no_flip,
+               args.no_shuffle,
+               args.resume,
+               ctx,
+               args.pretrained,
+               args.pretrained_epoch,
+               args.prefix,
+               args.begin_epoch,
+               args.end_epoch,
+               train_shared=args.train_shared,
+               lr=args.lr,
+               lr_step=args.lr_step,
+               proposal=args.proposal)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/rcnn/tools/train_rpn.py b/detection/retinaface/rcnn/tools/train_rpn.py
new file mode 100644
index 000000000..9d6008093
--- /dev/null
+++ b/detection/retinaface/rcnn/tools/train_rpn.py
@@ -0,0 +1,300 @@
+import argparse
+import logging
+import pprint
+import mxnet as mx
+
+from ..config import config, default, generate_config
+from ..symbol import *
+from ..core import callback, metric
+from ..core.loader import AnchorLoaderFPN
+from ..core.module import MutableModule
+from ..utils.load_data import load_gt_roidb, merge_roidb, filter_roidb
+from ..utils.load_model import load_param
+
+
+def train_rpn(network, dataset, image_set, root_path, dataset_path, frequent,
+              kvstore, work_load_list, no_flip, no_shuffle, resume, ctx,
+              pretrained, epoch, prefix, begin_epoch, end_epoch, train_shared,
+              lr, lr_step):
+    # set up logger
+    logging.basicConfig()
+    logger = logging.getLogger()
+    logger.setLevel(logging.INFO)
+
+    # setup config
+    assert config.TRAIN.BATCH_IMAGES == 1
+
+    # load symbol
+    sym = eval('get_' + network + '_rpn')()
+    feat_sym = []
+    for stride in config.RPN_FEAT_STRIDE:
+        feat_sym.append(sym.get_internals()['rpn_cls_score_stride%s_output' %
+                                            stride])
+
+    # setup multi-gpu
+    batch_size = len(ctx)
+    input_batch_size = config.TRAIN.BATCH_IMAGES * batch_size
+
+    # print config
+    pprint.pprint(config)
+
+    # load dataset and prepare imdb for training
+    image_sets = [iset for iset in image_set.split('+')]
+    roidbs = [
+        load_gt_roidb(dataset,
+                      image_set,
+                      root_path,
+                      dataset_path,
+                      flip=not no_flip) for image_set in image_sets
+    ]
+    roidb = merge_roidb(roidbs)
+    roidb = filter_roidb(roidb)
+
+    # load training data
+    #train_data = AnchorLoaderFPN(feat_sym, roidb, batch_size=input_batch_size, shuffle=not no_shuffle,
+    #                              ctx=ctx, work_load_list=work_load_list,
+    #                              feat_stride=config.RPN_FEAT_STRIDE, anchor_scales=config.ANCHOR_SCALES,
+    #                              anchor_ratios=config.ANCHOR_RATIOS, aspect_grouping=config.TRAIN.ASPECT_GROUPING,
+    #                              allowed_border=9999)
+    train_data = AnchorLoaderFPN(feat_sym,
+                                 roidb,
+                                 batch_size=input_batch_size,
+                                 shuffle=not no_shuffle,
+                                 ctx=ctx,
+                                 work_load_list=work_load_list)
+
+    # infer max shape
+    max_data_shape = [('data', (input_batch_size, 3,
+                                max([v[0] for v in config.SCALES]),
+                                max([v[1] for v in config.SCALES])))]
+    max_data_shape, max_label_shape = train_data.infer_shape(max_data_shape)
+    print 'providing maximum shape', max_data_shape, max_label_shape
+
+    # infer shape
+    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
+    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
+    out_shape_dict = zip(sym.list_outputs(), out_shape)
+    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
+    print 'output shape'
+    pprint.pprint(out_shape_dict)
+
+    # load and initialize params
+    if resume:
+        arg_params, aux_params = load_param(prefix, begin_epoch, convert=True)
+    else:
+        arg_params, aux_params = load_param(pretrained, epoch, convert=True)
+        init = mx.init.Xavier(factor_type="in",
+                              rnd_type='gaussian',
+                              magnitude=2)
+        init_internal = mx.init.Normal(sigma=0.01)
+        for k in sym.list_arguments():
+            if k in data_shape_dict:
+                continue
+            if k not in arg_params:
+                print 'init', k
+                arg_params[k] = mx.nd.zeros(shape=arg_shape_dict[k])
+                if not k.endswith('bias'):
+                    init_internal(k, arg_params[k])
+
+        for k in sym.list_auxiliary_states():
+            if k not in aux_params:
+                print 'init', k
+                aux_params[k] = mx.nd.zeros(shape=aux_shape_dict[k])
+                init(k, aux_params[k])
+
+    # check parameter shapes
+    for k in sym.list_arguments():
+        if k in data_shape_dict:
+            continue
+        assert k in arg_params, k + ' not initialized'
+        assert arg_params[k].shape == arg_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
+    for k in sym.list_auxiliary_states():
+        assert k in aux_params, k + ' not initialized'
+        assert aux_params[k].shape == aux_shape_dict[k], \
+            'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
+
+    # create solver
+    data_names = [k[0] for k in train_data.provide_data]
+    label_names = [k[0] for k in train_data.provide_label]
+    if train_shared:
+        fixed_param_prefix = config.FIXED_PARAMS_SHARED
+    else:
+        fixed_param_prefix = config.FIXED_PARAMS
+    mod = MutableModule(sym,
+                        data_names=data_names,
+                        label_names=label_names,
+                        logger=logger,
+                        context=ctx,
+                        work_load_list=work_load_list,
+                        max_data_shapes=max_data_shape,
+                        max_label_shapes=max_label_shape,
+                        fixed_param_prefix=fixed_param_prefix)
+
+    # decide training params
+    # metric
+    eval_metric = metric.RPNAccMetric()
+    cls_metric = metric.RPNLogLossMetric()
+    bbox_metric = metric.RPNL1LossMetric()
+    eval_metrics = mx.metric.CompositeEvalMetric()
+    for child_metric in [eval_metric, cls_metric, bbox_metric]:
+        eval_metrics.add(child_metric)
+    # callback
+    batch_end_callback = []
+    batch_end_callback.append(
+        mx.callback.Speedometer(train_data.batch_size, frequent=frequent))
+    epoch_end_callback = mx.callback.do_checkpoint(prefix)
+    # decide learning rate
+    base_lr = lr
+    lr_factor = 0.1
+    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
+    lr_epoch_diff = [
+        epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch
+    ]
+    lr = base_lr * (lr_factor**(len(lr_epoch) - len(lr_epoch_diff)))
+    lr_iters = [
+        int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff
+    ]
+    print 'lr', lr, 'lr_epoch_diff', lr_epoch_diff, 'lr_iters', lr_iters
+    lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
+    # optimizer
+    optimizer_params = {
+        'momentum': 0.9,
+        'wd': 0.0001,
+        'learning_rate': lr,
+        'lr_scheduler': lr_scheduler,
+        'rescale_grad': (1.0 / batch_size),
+        'clip_gradient': 5
+    }
+
+    # train
+    mod.fit(train_data,
+            eval_metric=eval_metrics,
+            epoch_end_callback=epoch_end_callback,
+            batch_end_callback=batch_end_callback,
+            kvstore=kvstore,
+            optimizer='sgd',
+            optimizer_params=optimizer_params,
+            arg_params=arg_params,
+            aux_params=aux_params,
+            begin_epoch=begin_epoch,
+            num_epoch=end_epoch)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Train a Region Proposal Network')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # training
+    parser.add_argument('--frequent',
+                        help='frequency of logging',
+                        default=default.frequent,
+                        type=int)
+    parser.add_argument('--kvstore',
+                        help='the kv-store type',
+                        default=default.kvstore,
+                        type=str)
+    parser.add_argument('--work_load_list',
+                        help='work load for different devices',
+                        default=None,
+                        type=list)
+    parser.add_argument('--no_flip',
+                        help='disable flip images',
+                        action='store_true')
+    parser.add_argument('--no_shuffle',
+                        help='disable random shuffle',
+                        action='store_true')
+    parser.add_argument('--resume',
+                        help='continue training',
+                        action='store_true')
+    # rpn
+    parser.add_argument('--gpus',
+                        help='GPU device to train with',
+                        default='0',
+                        type=str)
+    parser.add_argument('--pretrained',
+                        help='pretrained model prefix',
+                        default=default.pretrained,
+                        type=str)
+    parser.add_argument('--pretrained_epoch',
+                        help='pretrained model epoch',
+                        default=default.pretrained_epoch,
+                        type=int)
+    parser.add_argument('--prefix',
+                        help='new model prefix',
+                        default=default.rpn_prefix,
+                        type=str)
+    parser.add_argument('--begin_epoch',
+                        help='begin epoch of training',
+                        default=0,
+                        type=int)
+    parser.add_argument('--end_epoch',
+                        help='end epoch of training',
+                        default=default.rpn_epoch,
+                        type=int)
+    parser.add_argument('--lr',
+                        help='base learning rate',
+                        default=default.rpn_lr,
+                        type=float)
+    parser.add_argument('--lr_step',
+                        help='learning rate steps (in epoch)',
+                        default=default.rpn_lr_step,
+                        type=str)
+    parser.add_argument('--train_shared',
+                        help='second round train shared params',
+                        action='store_true')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    print 'Called with argument:', args
+    ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
+    train_rpn(args.network,
+              args.dataset,
+              args.image_set,
+              args.root_path,
+              args.dataset_path,
+              args.frequent,
+              args.kvstore,
+              args.work_load_list,
+              args.no_flip,
+              args.no_shuffle,
+              args.resume,
+              ctx,
+              args.pretrained,
+              args.pretrained_epoch,
+              args.prefix,
+              args.begin_epoch,
+              args.end_epoch,
+              train_shared=args.train_shared,
+              lr=args.lr,
+              lr_step=args.lr_step)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/rcnn/utils/__init__.py b/detection/retinaface/rcnn/utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/RetinaFace/rcnn/utils/combine_model.py b/detection/retinaface/rcnn/utils/combine_model.py
similarity index 100%
rename from RetinaFace/rcnn/utils/combine_model.py
rename to detection/retinaface/rcnn/utils/combine_model.py
diff --git a/detection/retinaface/rcnn/utils/load_data.py b/detection/retinaface/rcnn/utils/load_data.py
new file mode 100644
index 000000000..5c9d3a45a
--- /dev/null
+++ b/detection/retinaface/rcnn/utils/load_data.py
@@ -0,0 +1,64 @@
+import numpy as np
+from ..logger import logger
+from ..config import config
+from ..dataset import *
+
+
+def load_gt_roidb(dataset_name,
+                  image_set_name,
+                  root_path,
+                  dataset_path,
+                  flip=False):
+    """ load ground truth roidb """
+    imdb = eval(dataset_name)(image_set_name, root_path, dataset_path)
+    roidb = imdb.gt_roidb()
+    print('roidb size', len(roidb))
+    if flip:
+        roidb = imdb.append_flipped_images(roidb)
+    print('flipped roidb size', len(roidb))
+    return roidb
+
+
+def load_proposal_roidb(dataset_name,
+                        image_set_name,
+                        root_path,
+                        dataset_path,
+                        proposal='rpn',
+                        append_gt=True,
+                        flip=False):
+    """ load proposal roidb (append_gt when training) """
+    imdb = eval(dataset_name)(image_set_name, root_path, dataset_path)
+    gt_roidb = imdb.gt_roidb()
+    roidb = eval('imdb.' + proposal + '_roidb')(gt_roidb, append_gt)
+    if flip:
+        roidb = imdb.append_flipped_images(roidb)
+    return roidb
+
+
+def merge_roidb(roidbs):
+    """ roidb are list, concat them together """
+    roidb = roidbs[0]
+    for r in roidbs[1:]:
+        roidb.extend(r)
+    return roidb
+
+
+def filter_roidb(roidb):
+    """ remove roidb entries without usable rois """
+    def is_valid(entry):
+        """ valid images have at least 1 fg or bg roi """
+        overlaps = entry['max_overlaps']
+        fg_inds = np.where(overlaps >= config.TRAIN.FG_THRESH)[0]
+        bg_inds = np.where((overlaps < config.TRAIN.BG_THRESH_HI)
+                           & (overlaps >= config.TRAIN.BG_THRESH_LO))[0]
+        valid = len(fg_inds) > 0 or len(bg_inds) > 0
+        #valid = len(fg_inds) > 0
+        return valid
+
+    num = len(roidb)
+    filtered_roidb = [entry for entry in roidb if is_valid(entry)]
+    num_after = len(filtered_roidb)
+    logger.info('load data: filtered %d roidb entries: %d -> %d' %
+                (num - num_after, num, num_after))
+
+    return filtered_roidb
diff --git a/RetinaFace/rcnn/utils/load_model.py b/detection/retinaface/rcnn/utils/load_model.py
similarity index 100%
rename from RetinaFace/rcnn/utils/load_model.py
rename to detection/retinaface/rcnn/utils/load_model.py
diff --git a/RetinaFace/rcnn/utils/save_model.py b/detection/retinaface/rcnn/utils/save_model.py
similarity index 81%
rename from RetinaFace/rcnn/utils/save_model.py
rename to detection/retinaface/rcnn/utils/save_model.py
index 1c9886973..9c595f4db 100644
--- a/RetinaFace/rcnn/utils/save_model.py
+++ b/detection/retinaface/rcnn/utils/save_model.py
@@ -12,7 +12,7 @@ def save_checkpoint(prefix, epoch, arg_params, aux_params):
     :return: None
     prefix-epoch.params will be saved for parameters.
     """
-    save_dict = {('arg:%s' % k) : v for k, v in arg_params.items()}
-    save_dict.update({('aux:%s' % k) : v for k, v in aux_params.items()})
+    save_dict = {('arg:%s' % k): v for k, v in arg_params.items()}
+    save_dict.update({('aux:%s' % k): v for k, v in aux_params.items()})
     param_name = '%s-%04d.params' % (prefix, epoch)
     mx.nd.save(param_name, save_dict)
diff --git a/detection/retinaface/retinaface.py b/detection/retinaface/retinaface.py
new file mode 100644
index 000000000..5b274e3da
--- /dev/null
+++ b/detection/retinaface/retinaface.py
@@ -0,0 +1,839 @@
+from __future__ import print_function
+import sys
+import os
+import datetime
+import time
+import numpy as np
+import mxnet as mx
+from mxnet import ndarray as nd
+import cv2
+#from rcnn import config
+from rcnn.logger import logger
+#from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
+from rcnn.processing.bbox_transform import clip_boxes
+from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
+from rcnn.processing.nms import gpu_nms_wrapper, cpu_nms_wrapper
+from rcnn.processing.bbox_transform import bbox_overlaps
+
+
+class RetinaFace:
+    def __init__(self,
+                 prefix,
+                 epoch,
+                 ctx_id=0,
+                 network='net3',
+                 nms=0.4,
+                 nocrop=False,
+                 decay4=0.5,
+                 vote=False):
+        self.ctx_id = ctx_id
+        self.network = network
+        self.decay4 = decay4
+        self.nms_threshold = nms
+        self.vote = vote
+        self.nocrop = nocrop
+        self.debug = False
+        self.fpn_keys = []
+        self.anchor_cfg = None
+        pixel_means = [0.0, 0.0, 0.0]
+        pixel_stds = [1.0, 1.0, 1.0]
+        pixel_scale = 1.0
+        self.preprocess = False
+        _ratio = (1., )
+        fmc = 3
+        if network == 'ssh' or network == 'vgg':
+            pixel_means = [103.939, 116.779, 123.68]
+            self.preprocess = True
+        elif network == 'net3':
+            _ratio = (1., )
+        elif network == 'net3a':
+            _ratio = (1., 1.5)
+        elif network == 'net6':  #like pyramidbox or s3fd
+            fmc = 6
+        elif network == 'net5':  #retinaface
+            fmc = 5
+        elif network == 'net5a':
+            fmc = 5
+            _ratio = (1., 1.5)
+        elif network == 'net4':
+            fmc = 4
+        elif network == 'net4a':
+            fmc = 4
+            _ratio = (1., 1.5)
+        elif network == 'x5':
+            fmc = 5
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        elif network == 'x3':
+            fmc = 3
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        elif network == 'x3a':
+            fmc = 3
+            _ratio = (1., 1.5)
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        else:
+            assert False, 'network setting error %s' % network
+
+        if fmc == 3:
+            self._feat_stride_fpn = [32, 16, 8]
+            self.anchor_cfg = {
+                '32': {
+                    'SCALES': (32, 16),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (8, 4),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 4:
+            self._feat_stride_fpn = [32, 16, 8, 4]
+            self.anchor_cfg = {
+                '32': {
+                    'SCALES': (32, 16),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (8, 4),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '4': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 6:
+            self._feat_stride_fpn = [128, 64, 32, 16, 8, 4]
+            self.anchor_cfg = {
+                '128': {
+                    'SCALES': (32, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '64': {
+                    'SCALES': (16, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '32': {
+                    'SCALES': (8, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (4, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '4': {
+                    'SCALES': (1, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 5:
+            self._feat_stride_fpn = [64, 32, 16, 8, 4]
+            self.anchor_cfg = {}
+            _ass = 2.0**(1.0 / 3)
+            _basescale = 1.0
+            for _stride in [4, 8, 16, 32, 64]:
+                key = str(_stride)
+                value = {
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                }
+                scales = []
+                for _ in range(3):
+                    scales.append(_basescale)
+                    _basescale *= _ass
+                value['SCALES'] = tuple(scales)
+                self.anchor_cfg[key] = value
+
+        print(self._feat_stride_fpn, self.anchor_cfg)
+
+        for s in self._feat_stride_fpn:
+            self.fpn_keys.append('stride%s' % s)
+
+        dense_anchor = False
+        #self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(base_size=fpn_base_size, scales=self._scales, ratios=self._ratios)))
+        self._anchors_fpn = dict(
+            zip(
+                self.fpn_keys,
+                generate_anchors_fpn(dense_anchor=dense_anchor,
+                                     cfg=self.anchor_cfg)))
+        for k in self._anchors_fpn:
+            v = self._anchors_fpn[k].astype(np.float32)
+            self._anchors_fpn[k] = v
+
+        self._num_anchors = dict(
+            zip(self.fpn_keys,
+                [anchors.shape[0] for anchors in self._anchors_fpn.values()]))
+        #self._bbox_pred = nonlinear_pred
+        #self._landmark_pred = landmark_pred
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        if self.ctx_id >= 0:
+            self.ctx = mx.gpu(self.ctx_id)
+            self.nms = gpu_nms_wrapper(self.nms_threshold, self.ctx_id)
+        else:
+            self.ctx = mx.cpu()
+            self.nms = cpu_nms_wrapper(self.nms_threshold)
+        self.pixel_means = np.array(pixel_means, dtype=np.float32)
+        self.pixel_stds = np.array(pixel_stds, dtype=np.float32)
+        self.pixel_scale = float(pixel_scale)
+        print('means', self.pixel_means)
+        self.use_landmarks = False
+        if len(sym) // len(self._feat_stride_fpn) >= 3:
+            self.use_landmarks = True
+        print('use_landmarks', self.use_landmarks)
+        self.cascade = 0
+        if float(len(sym)) // len(self._feat_stride_fpn) > 3.0:
+            self.cascade = 1
+        print('cascade', self.cascade)
+        #self.bbox_stds = [0.1, 0.1, 0.2, 0.2]
+        #self.landmark_std = 0.1
+        self.bbox_stds = [1.0, 1.0, 1.0, 1.0]
+        self.landmark_std = 1.0
+
+        if self.debug:
+            c = len(sym) // len(self._feat_stride_fpn)
+            sym = sym[(c * 0):]
+            self._feat_stride_fpn = [32, 16, 8]
+        print('sym size:', len(sym))
+
+        image_size = (640, 640)
+        self.model = mx.mod.Module(symbol=sym,
+                                   context=self.ctx,
+                                   label_names=None)
+        self.model.bind(data_shapes=[('data', (1, 3, image_size[0],
+                                               image_size[1]))],
+                        for_training=False)
+        self.model.set_params(arg_params, aux_params)
+
+    def get_input(self, img):
+        im = img.astype(np.float32)
+        im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
+        for i in range(3):
+            im_tensor[
+                0,
+                i, :, :] = (im[:, :, 2 - i] / self.pixel_scale -
+                            self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
+        #if self.debug:
+        #  timeb = datetime.datetime.now()
+        #  diff = timeb - timea
+        #  print('X2 uses', diff.total_seconds(), 'seconds')
+        data = nd.array(im_tensor)
+        return data
+
+    def detect(self, img, threshold=0.5, scales=[1.0], do_flip=False):
+        #print('in_detect', threshold, scales, do_flip, do_nms)
+        proposals_list = []
+        scores_list = []
+        landmarks_list = []
+        strides_list = []
+        timea = datetime.datetime.now()
+        flips = [0]
+        if do_flip:
+            flips = [0, 1]
+
+        imgs = [img]
+        if isinstance(img, list):
+            imgs = img
+        for img in imgs:
+            for im_scale in scales:
+                for flip in flips:
+                    if im_scale != 1.0:
+                        im = cv2.resize(img,
+                                        None,
+                                        None,
+                                        fx=im_scale,
+                                        fy=im_scale,
+                                        interpolation=cv2.INTER_LINEAR)
+                    else:
+                        im = img.copy()
+                    if flip:
+                        im = im[:, ::-1, :]
+                    if self.nocrop:
+                        if im.shape[0] % 32 == 0:
+                            h = im.shape[0]
+                        else:
+                            h = (im.shape[0] // 32 + 1) * 32
+                        if im.shape[1] % 32 == 0:
+                            w = im.shape[1]
+                        else:
+                            w = (im.shape[1] // 32 + 1) * 32
+                        _im = np.zeros((h, w, 3), dtype=np.float32)
+                        _im[0:im.shape[0], 0:im.shape[1], :] = im
+                        im = _im
+                    else:
+                        im = im.astype(np.float32)
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X1 uses', diff.total_seconds(), 'seconds')
+                    #self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
+                    #im_info = [im.shape[0], im.shape[1], im_scale]
+                    im_info = [im.shape[0], im.shape[1]]
+                    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
+                    for i in range(3):
+                        im_tensor[0, i, :, :] = (
+                            im[:, :, 2 - i] / self.pixel_scale -
+                            self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X2 uses', diff.total_seconds(), 'seconds')
+                    data = nd.array(im_tensor)
+                    db = mx.io.DataBatch(data=(data, ),
+                                         provide_data=[('data', data.shape)])
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X3 uses', diff.total_seconds(), 'seconds')
+                    self.model.forward(db, is_train=False)
+                    net_out = self.model.get_outputs()
+                    #post_nms_topN = self._rpn_post_nms_top_n
+                    #min_size_dict = self._rpn_min_size_fpn
+
+                    sym_idx = 0
+
+                    for _idx, s in enumerate(self._feat_stride_fpn):
+                        #if len(scales)>1 and s==32 and im_scale==scales[-1]:
+                        #  continue
+                        _key = 'stride%s' % s
+                        stride = int(s)
+                        is_cascade = False
+                        if self.cascade:
+                            is_cascade = True
+                        #if self.vote and stride==4 and len(scales)>2 and (im_scale==scales[0]):
+                        #  continue
+                        #print('getting', im_scale, stride, idx, len(net_out), data.shape, file=sys.stderr)
+                        scores = net_out[sym_idx].asnumpy()
+                        if self.debug:
+                            timeb = datetime.datetime.now()
+                            diff = timeb - timea
+                            print('A uses', diff.total_seconds(), 'seconds')
+                        #print(scores.shape)
+                        #print('scores',stride, scores.shape, file=sys.stderr)
+                        scores = scores[:, self._num_anchors['stride%s' %
+                                                             s]:, :, :]
+
+                        bbox_deltas = net_out[sym_idx + 1].asnumpy()
+
+                        #if DEBUG:
+                        #    print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
+                        #    print 'scale: {}'.format(im_info[2])
+
+                        #_height, _width = int(im_info[0] / stride), int(im_info[1] / stride)
+                        height, width = bbox_deltas.shape[
+                            2], bbox_deltas.shape[3]
+
+                        A = self._num_anchors['stride%s' % s]
+                        K = height * width
+                        anchors_fpn = self._anchors_fpn['stride%s' % s]
+                        anchors = anchors_plane(height, width, stride,
+                                                anchors_fpn)
+                        #print((height, width), (_height, _width), anchors.shape, bbox_deltas.shape, scores.shape, file=sys.stderr)
+                        anchors = anchors.reshape((K * A, 4))
+                        #print('num_anchors', self._num_anchors['stride%s'%s], file=sys.stderr)
+                        #print('HW', (height, width), file=sys.stderr)
+                        #print('anchors_fpn', anchors_fpn.shape, file=sys.stderr)
+                        #print('anchors', anchors.shape, file=sys.stderr)
+                        #print('bbox_deltas', bbox_deltas.shape, file=sys.stderr)
+                        #print('scores', scores.shape, file=sys.stderr)
+
+                        #scores = self._clip_pad(scores, (height, width))
+                        scores = scores.transpose((0, 2, 3, 1)).reshape(
+                            (-1, 1))
+
+                        #print('pre', bbox_deltas.shape, height, width)
+                        #bbox_deltas = self._clip_pad(bbox_deltas, (height, width))
+                        #print('after', bbox_deltas.shape, height, width)
+                        bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
+                        bbox_pred_len = bbox_deltas.shape[3] // A
+                        #print(bbox_deltas.shape)
+                        bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
+                        bbox_deltas[:,
+                                    0::4] = bbox_deltas[:, 0::
+                                                        4] * self.bbox_stds[0]
+                        bbox_deltas[:,
+                                    1::4] = bbox_deltas[:, 1::
+                                                        4] * self.bbox_stds[1]
+                        bbox_deltas[:,
+                                    2::4] = bbox_deltas[:, 2::
+                                                        4] * self.bbox_stds[2]
+                        bbox_deltas[:,
+                                    3::4] = bbox_deltas[:, 3::
+                                                        4] * self.bbox_stds[3]
+                        proposals = self.bbox_pred(anchors, bbox_deltas)
+
+                        #print(anchors.shape, bbox_deltas.shape, A, K, file=sys.stderr)
+                        if is_cascade:
+                            cascade_sym_num = 0
+                            cls_cascade = False
+                            bbox_cascade = False
+                            __idx = [3, 4]
+                            if not self.use_landmarks:
+                                __idx = [2, 3]
+                            for diff_idx in __idx:
+                                if sym_idx + diff_idx >= len(net_out):
+                                    break
+                                body = net_out[sym_idx + diff_idx].asnumpy()
+                                if body.shape[1] // A == 2:  #cls branch
+                                    if cls_cascade or bbox_cascade:
+                                        break
+                                    else:
+                                        cascade_scores = body[:, self.
+                                                              _num_anchors[
+                                                                  'stride%s' %
+                                                                  s]:, :, :]
+                                        cascade_scores = cascade_scores.transpose(
+                                            (0, 2, 3, 1)).reshape((-1, 1))
+                                        #scores = (scores+cascade_scores)/2.0
+                                        scores = cascade_scores  #TODO?
+                                        cascade_sym_num += 1
+                                        cls_cascade = True
+                                        #print('find cascade cls at stride', stride)
+                                elif body.shape[1] // A == 4:  #bbox branch
+                                    cascade_deltas = body.transpose(
+                                        (0, 2, 3, 1)).reshape(
+                                            (-1, bbox_pred_len))
+                                    cascade_deltas[:, 0::
+                                                   4] = cascade_deltas[:, 0::
+                                                                       4] * self.bbox_stds[
+                                                                           0]
+                                    cascade_deltas[:, 1::
+                                                   4] = cascade_deltas[:, 1::
+                                                                       4] * self.bbox_stds[
+                                                                           1]
+                                    cascade_deltas[:, 2::
+                                                   4] = cascade_deltas[:, 2::
+                                                                       4] * self.bbox_stds[
+                                                                           2]
+                                    cascade_deltas[:, 3::
+                                                   4] = cascade_deltas[:, 3::
+                                                                       4] * self.bbox_stds[
+                                                                           3]
+                                    proposals = self.bbox_pred(
+                                        proposals, cascade_deltas)
+                                    cascade_sym_num += 1
+                                    bbox_cascade = True
+                                    #print('find cascade bbox at stride', stride)
+
+                        proposals = clip_boxes(proposals, im_info[:2])
+
+                        #if self.vote:
+                        #  if im_scale>1.0:
+                        #    keep = self._filter_boxes2(proposals, 160*im_scale, -1)
+                        #  else:
+                        #    keep = self._filter_boxes2(proposals, -1, 100*im_scale)
+                        #  if stride==4:
+                        #    keep = self._filter_boxes2(proposals, 12*im_scale, -1)
+                        #    proposals = proposals[keep, :]
+                        #    scores = scores[keep]
+
+                        #keep = self._filter_boxes(proposals, min_size_dict['stride%s'%s] * im_info[2])
+                        #proposals = proposals[keep, :]
+                        #scores = scores[keep]
+                        #print('333', proposals.shape)
+                        if stride == 4 and self.decay4 < 1.0:
+                            scores *= self.decay4
+
+                        scores_ravel = scores.ravel()
+                        #print('__shapes', proposals.shape, scores_ravel.shape)
+                        #print('max score', np.max(scores_ravel))
+                        order = np.where(scores_ravel >= threshold)[0]
+                        #_scores = scores_ravel[order]
+                        #_order = _scores.argsort()[::-1]
+                        #order = order[_order]
+                        proposals = proposals[order, :]
+                        scores = scores[order]
+                        if flip:
+                            oldx1 = proposals[:, 0].copy()
+                            oldx2 = proposals[:, 2].copy()
+                            proposals[:, 0] = im.shape[1] - oldx2 - 1
+                            proposals[:, 2] = im.shape[1] - oldx1 - 1
+
+                        proposals[:, 0:4] /= im_scale
+
+                        proposals_list.append(proposals)
+                        scores_list.append(scores)
+                        if self.nms_threshold < 0.0:
+                            _strides = np.empty(shape=(scores.shape),
+                                                dtype=np.float32)
+                            _strides.fill(stride)
+                            strides_list.append(_strides)
+
+                        if not self.vote and self.use_landmarks:
+                            landmark_deltas = net_out[sym_idx + 2].asnumpy()
+                            #landmark_deltas = self._clip_pad(landmark_deltas, (height, width))
+                            landmark_pred_len = landmark_deltas.shape[1] // A
+                            landmark_deltas = landmark_deltas.transpose(
+                                (0, 2, 3, 1)).reshape(
+                                    (-1, 5, landmark_pred_len // 5))
+                            landmark_deltas *= self.landmark_std
+                            #print(landmark_deltas.shape, landmark_deltas)
+                            landmarks = self.landmark_pred(
+                                anchors, landmark_deltas)
+                            landmarks = landmarks[order, :]
+
+                            if flip:
+                                landmarks[:, :,
+                                          0] = im.shape[1] - landmarks[:, :,
+                                                                       0] - 1
+                                #for a in range(5):
+                                #  oldx1 = landmarks[:, a].copy()
+                                #  landmarks[:,a] = im.shape[1] - oldx1 - 1
+                                order = [1, 0, 2, 4, 3]
+                                flandmarks = landmarks.copy()
+                                for idx, a in enumerate(order):
+                                    flandmarks[:, idx, :] = landmarks[:, a, :]
+                                    #flandmarks[:, idx*2] = landmarks[:,a*2]
+                                    #flandmarks[:, idx*2+1] = landmarks[:,a*2+1]
+                                landmarks = flandmarks
+                            landmarks[:, :, 0:2] /= im_scale
+                            #landmarks /= im_scale
+                            #landmarks = landmarks.reshape( (-1, landmark_pred_len) )
+                            landmarks_list.append(landmarks)
+                            #proposals = np.hstack((proposals, landmarks))
+                        if self.use_landmarks:
+                            sym_idx += 3
+                        else:
+                            sym_idx += 2
+                        if is_cascade:
+                            sym_idx += cascade_sym_num
+
+        if self.debug:
+            timeb = datetime.datetime.now()
+            diff = timeb - timea
+            print('B uses', diff.total_seconds(), 'seconds')
+        proposals = np.vstack(proposals_list)
+        landmarks = None
+        if proposals.shape[0] == 0:
+            if self.use_landmarks:
+                landmarks = np.zeros((0, 5, 2))
+            if self.nms_threshold < 0.0:
+                return np.zeros((0, 6)), landmarks
+            else:
+                return np.zeros((0, 5)), landmarks
+        scores = np.vstack(scores_list)
+        #print('shapes', proposals.shape, scores.shape)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        #if config.TEST.SCORE_THRESH>0.0:
+        #  _count = np.sum(scores_ravel>config.TEST.SCORE_THRESH)
+        #  order = order[:_count]
+        proposals = proposals[order, :]
+        scores = scores[order]
+        if self.nms_threshold < 0.0:
+            strides = np.vstack(strides_list)
+            strides = strides[order]
+        if not self.vote and self.use_landmarks:
+            landmarks = np.vstack(landmarks_list)
+            landmarks = landmarks[order].astype(np.float32, copy=False)
+
+        if self.nms_threshold > 0.0:
+            pre_det = np.hstack((proposals[:, 0:4], scores)).astype(np.float32,
+                                                                    copy=False)
+            if not self.vote:
+                keep = self.nms(pre_det)
+                det = np.hstack((pre_det, proposals[:, 4:]))
+                det = det[keep, :]
+                if self.use_landmarks:
+                    landmarks = landmarks[keep]
+            else:
+                det = np.hstack((pre_det, proposals[:, 4:]))
+                det = self.bbox_vote(det)
+        elif self.nms_threshold < 0.0:
+            det = np.hstack(
+                (proposals[:, 0:4], scores, strides)).astype(np.float32,
+                                                             copy=False)
+        else:
+            det = np.hstack((proposals[:, 0:4], scores)).astype(np.float32,
+                                                                copy=False)
+
+        if self.debug:
+            timeb = datetime.datetime.now()
+            diff = timeb - timea
+            print('C uses', diff.total_seconds(), 'seconds')
+        return det, landmarks
+
+    def detect_center(self, img, threshold=0.5, scales=[1.0], do_flip=False):
+        det, landmarks = self.detect(img, threshold, scales, do_flip)
+        if det.shape[0] == 0:
+            return None, None
+        bindex = 0
+        if det.shape[0] > 1:
+            img_size = np.asarray(img.shape)[0:2]
+            bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                           det[:, 1])
+            img_center = img_size / 2
+            offsets = np.vstack([(det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                                 (det[:, 1] + det[:, 3]) / 2 - img_center[0]])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            bindex = np.argmax(bounding_box_size - offset_dist_squared *
+                               2.0)  # some extra weight on the centering
+        bbox = det[bindex, :]
+        landmark = landmarks[bindex, :, :]
+        return bbox, landmark
+
+    @staticmethod
+    def check_large_pose(landmark, bbox):
+        assert landmark.shape == (5, 2)
+        assert len(bbox) == 4
+
+        def get_theta(base, x, y):
+            vx = x - base
+            vy = y - base
+            vx[1] *= -1
+            vy[1] *= -1
+            tx = np.arctan2(vx[1], vx[0])
+            ty = np.arctan2(vy[1], vy[0])
+            d = ty - tx
+            d = np.degrees(d)
+            #print(vx, tx, vy, ty, d)
+            #if d<-1.*math.pi:
+            #  d+=2*math.pi
+            #elif d>math.pi:
+            #  d-=2*math.pi
+            if d < -180.0:
+                d += 360.
+            elif d > 180.0:
+                d -= 360.0
+            return d
+
+        landmark = landmark.astype(np.float32)
+
+        theta1 = get_theta(landmark[0], landmark[3], landmark[2])
+        theta2 = get_theta(landmark[1], landmark[2], landmark[4])
+        #print(va, vb, theta2)
+        theta3 = get_theta(landmark[0], landmark[2], landmark[1])
+        theta4 = get_theta(landmark[1], landmark[0], landmark[2])
+        theta5 = get_theta(landmark[3], landmark[4], landmark[2])
+        theta6 = get_theta(landmark[4], landmark[2], landmark[3])
+        theta7 = get_theta(landmark[3], landmark[2], landmark[0])
+        theta8 = get_theta(landmark[4], landmark[1], landmark[2])
+        #print(theta1, theta2, theta3, theta4, theta5, theta6, theta7, theta8)
+        left_score = 0.0
+        right_score = 0.0
+        up_score = 0.0
+        down_score = 0.0
+        if theta1 <= 0.0:
+            left_score = 10.0
+        elif theta2 <= 0.0:
+            right_score = 10.0
+        else:
+            left_score = theta2 / theta1
+            right_score = theta1 / theta2
+        if theta3 <= 10.0 or theta4 <= 10.0:
+            up_score = 10.0
+        else:
+            up_score = max(theta1 / theta3, theta2 / theta4)
+        if theta5 <= 10.0 or theta6 <= 10.0:
+            down_score = 10.0
+        else:
+            down_score = max(theta7 / theta5, theta8 / theta6)
+        mleft = (landmark[0][0] + landmark[3][0]) / 2
+        mright = (landmark[1][0] + landmark[4][0]) / 2
+        box_center = ((bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2)
+        ret = 0
+        if left_score >= 3.0:
+            ret = 1
+        if ret == 0 and left_score >= 2.0:
+            if mright <= box_center[0]:
+                ret = 1
+        if ret == 0 and right_score >= 3.0:
+            ret = 2
+        if ret == 0 and right_score >= 2.0:
+            if mleft >= box_center[0]:
+                ret = 2
+        if ret == 0 and up_score >= 2.0:
+            ret = 3
+        if ret == 0 and down_score >= 5.0:
+            ret = 4
+        return ret, left_score, right_score, up_score, down_score
+
+    @staticmethod
+    def _filter_boxes(boxes, min_size):
+        """ Remove all boxes with any side smaller than min_size """
+        ws = boxes[:, 2] - boxes[:, 0] + 1
+        hs = boxes[:, 3] - boxes[:, 1] + 1
+        keep = np.where((ws >= min_size) & (hs >= min_size))[0]
+        return keep
+
+    @staticmethod
+    def _filter_boxes2(boxes, max_size, min_size):
+        """ Remove all boxes with any side smaller than min_size """
+        ws = boxes[:, 2] - boxes[:, 0] + 1
+        hs = boxes[:, 3] - boxes[:, 1] + 1
+        if max_size > 0:
+            keep = np.where(np.minimum(ws, hs) < max_size)[0]
+        elif min_size > 0:
+            keep = np.where(np.maximum(ws, hs) > min_size)[0]
+        return keep
+
+    @staticmethod
+    def _clip_pad(tensor, pad_shape):
+        """
+      Clip boxes of the pad area.
+      :param tensor: [n, c, H, W]
+      :param pad_shape: [h, w]
+      :return: [n, c, h, w]
+      """
+        H, W = tensor.shape[2:]
+        h, w = pad_shape
+
+        if h < H or w < W:
+            tensor = tensor[:, :, :h, :w].copy()
+
+        return tensor
+
+    @staticmethod
+    def bbox_pred(boxes, box_deltas):
+        """
+      Transform the set of class-agnostic boxes into class-specific boxes
+      by applying the predicted offsets (box_deltas)
+      :param boxes: !important [N 4]
+      :param box_deltas: [N, 4 * num_classes]
+      :return: [N 4 * num_classes]
+      """
+        if boxes.shape[0] == 0:
+            return np.zeros((0, box_deltas.shape[1]))
+
+        boxes = boxes.astype(np.float, copy=False)
+        widths = boxes[:, 2] - boxes[:, 0] + 1.0
+        heights = boxes[:, 3] - boxes[:, 1] + 1.0
+        ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+        ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+
+        dx = box_deltas[:, 0:1]
+        dy = box_deltas[:, 1:2]
+        dw = box_deltas[:, 2:3]
+        dh = box_deltas[:, 3:4]
+
+        pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
+        pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
+        pred_w = np.exp(dw) * widths[:, np.newaxis]
+        pred_h = np.exp(dh) * heights[:, np.newaxis]
+
+        pred_boxes = np.zeros(box_deltas.shape)
+        # x1
+        pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
+        # y1
+        pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
+        # x2
+        pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
+        # y2
+        pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
+
+        if box_deltas.shape[1] > 4:
+            pred_boxes[:, 4:] = box_deltas[:, 4:]
+
+        return pred_boxes
+
+    @staticmethod
+    def landmark_pred(boxes, landmark_deltas):
+        if boxes.shape[0] == 0:
+            return np.zeros((0, landmark_deltas.shape[1]))
+        boxes = boxes.astype(np.float, copy=False)
+        widths = boxes[:, 2] - boxes[:, 0] + 1.0
+        heights = boxes[:, 3] - boxes[:, 1] + 1.0
+        ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+        ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+        pred = landmark_deltas.copy()
+        for i in range(5):
+            pred[:, i, 0] = landmark_deltas[:, i, 0] * widths + ctr_x
+            pred[:, i, 1] = landmark_deltas[:, i, 1] * heights + ctr_y
+        return pred
+        #preds = []
+        #for i in range(landmark_deltas.shape[1]):
+        #  if i%2==0:
+        #    pred = (landmark_deltas[:,i]*widths + ctr_x)
+        #  else:
+        #    pred = (landmark_deltas[:,i]*heights + ctr_y)
+        #  preds.append(pred)
+        #preds = np.vstack(preds).transpose()
+        #return preds
+
+    def bbox_vote(self, det):
+        #order = det[:, 4].ravel().argsort()[::-1]
+        #det = det[order, :]
+        if det.shape[0] == 0:
+            return np.zeros((0, 5))
+            #dets = np.array([[10, 10, 20, 20, 0.002]])
+            #det = np.empty(shape=[0, 5])
+        dets = None
+        while det.shape[0] > 0:
+            if dets is not None and dets.shape[0] >= 750:
+                break
+            # IOU
+            area = (det[:, 2] - det[:, 0] + 1) * (det[:, 3] - det[:, 1] + 1)
+            xx1 = np.maximum(det[0, 0], det[:, 0])
+            yy1 = np.maximum(det[0, 1], det[:, 1])
+            xx2 = np.minimum(det[0, 2], det[:, 2])
+            yy2 = np.minimum(det[0, 3], det[:, 3])
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            o = inter / (area[0] + area[:] - inter)
+
+            # nms
+            merge_index = np.where(o >= self.nms_threshold)[0]
+            det_accu = det[merge_index, :]
+            det = np.delete(det, merge_index, 0)
+            if merge_index.shape[0] <= 1:
+                if det.shape[0] == 0:
+                    try:
+                        dets = np.row_stack((dets, det_accu))
+                    except:
+                        dets = det_accu
+                continue
+            det_accu[:,
+                     0:4] = det_accu[:, 0:4] * np.tile(det_accu[:, -1:],
+                                                       (1, 4))
+            max_score = np.max(det_accu[:, 4])
+            det_accu_sum = np.zeros((1, 5))
+            det_accu_sum[:, 0:4] = np.sum(det_accu[:, 0:4], axis=0) / np.sum(
+                det_accu[:, -1:])
+            det_accu_sum[:, 4] = max_score
+            if dets is None:
+                dets = det_accu_sum
+            else:
+                dets = np.row_stack((dets, det_accu_sum))
+        dets = dets[0:750, :]
+        return dets
diff --git a/detection/retinaface/test.py b/detection/retinaface/test.py
new file mode 100644
index 000000000..b88c82ba6
--- /dev/null
+++ b/detection/retinaface/test.py
@@ -0,0 +1,63 @@
+import cv2
+import sys
+import numpy as np
+import datetime
+import os
+import glob
+from retinaface import RetinaFace
+
+thresh = 0.8
+scales = [1024, 1980]
+
+count = 1
+
+gpuid = 0
+detector = RetinaFace('./model/R50', 0, gpuid, 'net3')
+
+img = cv2.imread('t1.jpg')
+print(img.shape)
+im_shape = img.shape
+target_size = scales[0]
+max_size = scales[1]
+im_size_min = np.min(im_shape[0:2])
+im_size_max = np.max(im_shape[0:2])
+#im_scale = 1.0
+#if im_size_min>target_size or im_size_max>max_size:
+im_scale = float(target_size) / float(im_size_min)
+# prevent bigger axis from being more than max_size:
+if np.round(im_scale * im_size_max) > max_size:
+    im_scale = float(max_size) / float(im_size_max)
+
+print('im_scale', im_scale)
+
+scales = [im_scale]
+flip = False
+
+for c in range(count):
+    faces, landmarks = detector.detect(img,
+                                       thresh,
+                                       scales=scales,
+                                       do_flip=flip)
+    print(c, faces.shape, landmarks.shape)
+
+if faces is not None:
+    print('find', faces.shape[0], 'faces')
+    for i in range(faces.shape[0]):
+        #print('score', faces[i][4])
+        box = faces[i].astype(np.int)
+        #color = (255,0,0)
+        color = (0, 0, 255)
+        cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
+        if landmarks is not None:
+            landmark5 = landmarks[i].astype(np.int)
+            #print(landmark.shape)
+            for l in range(landmark5.shape[0]):
+                color = (0, 0, 255)
+                if l == 0 or l == 3:
+                    color = (0, 255, 0)
+                cv2.circle(img, (landmark5[l][0], landmark5[l][1]), 1, color,
+                           2)
+
+    filename = './detector_test.jpg'
+    print('writing', filename)
+    cv2.imwrite(filename, img)
diff --git a/detection/retinaface/test_widerface.py b/detection/retinaface/test_widerface.py
new file mode 100644
index 000000000..78c2f83d9
--- /dev/null
+++ b/detection/retinaface/test_widerface.py
@@ -0,0 +1,259 @@
+from __future__ import print_function
+
+import argparse
+import sys
+import os
+import time
+import numpy as np
+import mxnet as mx
+from mxnet import ndarray as nd
+import cv2
+from rcnn.logger import logger
+#from rcnn.config import config, default, generate_config
+#from rcnn.tools.test_rcnn import test_rcnn
+#from rcnn.tools.test_rpn import test_rpn
+from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
+from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
+from rcnn.processing.nms import gpu_nms_wrapper
+from rcnn.processing.bbox_transform import bbox_overlaps
+from rcnn.dataset import retinaface
+from retinaface import RetinaFace
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Test widerface by retinaface detector')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default='net3',
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default='retinaface',
+                        type=str)
+    parser.add_argument('--image-set',
+                        help='image_set name',
+                        default='val',
+                        type=str)
+    parser.add_argument('--root-path',
+                        help='output data folder',
+                        default='./data',
+                        type=str)
+    parser.add_argument('--dataset-path',
+                        help='dataset path',
+                        default='./data/retinaface',
+                        type=str)
+    parser.add_argument('--gpu',
+                        help='GPU device to test with',
+                        default=0,
+                        type=int)
+    # testing
+    parser.add_argument('--prefix',
+                        help='model to test with',
+                        default='',
+                        type=str)
+    parser.add_argument('--epoch',
+                        help='model to test with',
+                        default=0,
+                        type=int)
+    parser.add_argument('--output',
+                        help='output folder',
+                        default='./wout',
+                        type=str)
+    parser.add_argument('--nocrop', help='', action='store_true')
+    parser.add_argument('--thresh',
+                        help='valid detection threshold',
+                        default=0.02,
+                        type=float)
+    parser.add_argument('--mode',
+                        help='test mode, 0 for fast, 1 for accurate',
+                        default=1,
+                        type=int)
+    #parser.add_argument('--pyramid', help='enable pyramid test', action='store_true')
+    #parser.add_argument('--bbox-vote', help='', action='store_true')
+    parser.add_argument('--part', help='', default=0, type=int)
+    parser.add_argument('--parts', help='', default=1, type=int)
+    args = parser.parse_args()
+    return args
+
+
+detector = None
+args = None
+imgid = -1
+
+
+def get_boxes(roi, pyramid):
+    global imgid
+    im = cv2.imread(roi['image'])
+    do_flip = False
+    if not pyramid:
+        target_size = 1200
+        max_size = 1600
+        #do_flip = True
+        target_size = 1504
+        max_size = 2000
+        target_size = 1600
+        max_size = 2150
+        im_shape = im.shape
+        im_size_min = np.min(im_shape[0:2])
+        im_size_max = np.max(im_shape[0:2])
+        im_scale = float(target_size) / float(im_size_min)
+        # prevent bigger axis from being more than max_size:
+        if np.round(im_scale * im_size_max) > max_size:
+            im_scale = float(max_size) / float(im_size_max)
+        scales = [im_scale]
+    else:
+        do_flip = True
+        #TEST_SCALES = [500, 800, 1200, 1600]
+        TEST_SCALES = [500, 800, 1100, 1400, 1700]
+        target_size = 800
+        max_size = 1200
+        im_shape = im.shape
+        im_size_min = np.min(im_shape[0:2])
+        im_size_max = np.max(im_shape[0:2])
+        im_scale = float(target_size) / float(im_size_min)
+        # prevent bigger axis from being more than max_size:
+        if np.round(im_scale * im_size_max) > max_size:
+            im_scale = float(max_size) / float(im_size_max)
+        scales = [
+            float(scale) / target_size * im_scale for scale in TEST_SCALES
+        ]
+    boxes, landmarks = detector.detect(im,
+                                       threshold=args.thresh,
+                                       scales=scales,
+                                       do_flip=do_flip)
+    #print(boxes.shape, landmarks.shape)
+    if imgid >= 0 and imgid < 100:
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        for i in range(boxes.shape[0]):
+            box = boxes[i]
+            ibox = box[0:4].copy().astype(np.int)
+            cv2.rectangle(im, (ibox[0], ibox[1]), (ibox[2], ibox[3]),
+                          (255, 0, 0), 2)
+            #print('box', ibox)
+            #if len(ibox)>5:
+            #  for l in range(5):
+            #    pp = (ibox[5+l*2], ibox[6+l*2])
+            #    cv2.circle(im, (pp[0], pp[1]), 1, (0, 0, 255), 1)
+            blur = box[5]
+            k = "%.3f" % blur
+            cv2.putText(im, k, (ibox[0] + 2, ibox[1] + 14), font, 0.6,
+                        (0, 255, 0), 2)
+            #landmarks = box[6:21].reshape( (5,3) )
+            if landmarks is not None:
+                for l in range(5):
+                    color = (0, 255, 0)
+                    landmark = landmarks[i][l]
+                    pp = (int(landmark[0]), int(landmark[1]))
+                    if landmark[2] - 0.5 < 0.0:
+                        color = (0, 0, 255)
+                    cv2.circle(im, (pp[0], pp[1]), 1, color, 2)
+        filename = './testimages/%d.jpg' % imgid
+        cv2.imwrite(filename, im)
+        print(filename, 'wrote')
+        imgid += 1
+
+    return boxes
+
+
+def test(args):
+    print('test with', args)
+    global detector
+    output_folder = args.output
+    if not os.path.exists(output_folder):
+        os.mkdir(output_folder)
+    detector = RetinaFace(args.prefix,
+                          args.epoch,
+                          args.gpu,
+                          network=args.network,
+                          nocrop=args.nocrop,
+                          vote=args.bbox_vote)
+    imdb = eval(args.dataset)(args.image_set, args.root_path,
+                              args.dataset_path)
+    roidb = imdb.gt_roidb()
+    gt_overlaps = np.zeros(0)
+    overall = [0.0, 0.0]
+    gt_max = np.array((0.0, 0.0))
+    num_pos = 0
+    print('roidb size', len(roidb))
+
+    for i in range(len(roidb)):
+        if i % args.parts != args.part:
+            continue
+        #if i%10==0:
+        #  print('processing', i, file=sys.stderr)
+        roi = roidb[i]
+        boxes = get_boxes(roi, args.pyramid)
+        if 'boxes' in roi:
+            gt_boxes = roi['boxes'].copy()
+            gt_areas = (gt_boxes[:, 2] - gt_boxes[:, 0] +
+                        1) * (gt_boxes[:, 3] - gt_boxes[:, 1] + 1)
+            num_pos += gt_boxes.shape[0]
+
+            overlaps = bbox_overlaps(boxes.astype(np.float),
+                                     gt_boxes.astype(np.float))
+            #print(im_info, gt_boxes.shape, boxes.shape, overlaps.shape, file=sys.stderr)
+
+            _gt_overlaps = np.zeros((gt_boxes.shape[0]))
+
+            if boxes.shape[0] > 0:
+                _gt_overlaps = overlaps.max(axis=0)
+                #print('max_overlaps', _gt_overlaps, file=sys.stderr)
+                for j in range(len(_gt_overlaps)):
+                    if _gt_overlaps[j] > 0.5:
+                        continue
+                    #print(j, 'failed', gt_boxes[j],  'max_overlap:', _gt_overlaps[j], file=sys.stderr)
+
+                # append recorded IoU coverage level
+                found = (_gt_overlaps > 0.5).sum()
+                recall = found / float(gt_boxes.shape[0])
+                #print('recall', _recall, gt_boxes.shape[0], boxes.shape[0], gt_areas, 'num:', i, file=sys.stderr)
+                overall[0] += found
+                overall[1] += gt_boxes.shape[0]
+                #gt_overlaps = np.hstack((gt_overlaps, _gt_overlaps))
+                #_recall = (gt_overlaps >= threshold).sum() / float(num_pos)
+                recall_all = float(overall[0]) / overall[1]
+                #print('recall_all', _recall, file=sys.stderr)
+                print('[%d]' % i,
+                      'recall',
+                      recall, (gt_boxes.shape[0], boxes.shape[0]),
+                      'all:',
+                      recall_all,
+                      file=sys.stderr)
+        else:
+            print('[%d]' % i, 'detect %d faces' % boxes.shape[0])
+
+        _vec = roidb[i]['image'].split('/')
+        out_dir = os.path.join(output_folder, _vec[-2])
+        if not os.path.exists(out_dir):
+            os.mkdir(out_dir)
+        out_file = os.path.join(out_dir, _vec[-1].replace('jpg', 'txt'))
+        with open(out_file, 'w') as f:
+            name = '/'.join(roidb[i]['image'].split('/')[-2:])
+            f.write("%s\n" % (name))
+            f.write("%d\n" % (boxes.shape[0]))
+            for b in range(boxes.shape[0]):
+                box = boxes[b]
+                f.write(
+                    "%d %d %d %d %g \n" %
+                    (box[0], box[1], box[2] - box[0], box[3] - box[1], box[4]))
+
+
+def main():
+    global args
+    args = parse_args()
+    args.pyramid = False
+    args.bbox_vote = False
+    if args.mode == 1:
+        args.pyramid = True
+        args.bbox_vote = True
+    elif args.mode == 2:
+        args.pyramid = True
+        args.bbox_vote = False
+    logger.info('Called with argument: %s' % args)
+    test(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface/train.py b/detection/retinaface/train.py
new file mode 100644
index 000000000..5fbcc70b9
--- /dev/null
+++ b/detection/retinaface/train.py
@@ -0,0 +1,502 @@
+from __future__ import print_function
+import sys
+import argparse
+import os
+import pprint
+import re
+import mxnet as mx
+import numpy as np
+from mxnet.module import Module
+import mxnet.optimizer as optimizer
+
+from rcnn.logger import logger
+from rcnn.config import config, default, generate_config
+from rcnn.symbol import *
+from rcnn.core import callback, metric
+from rcnn.core.loader import CropLoader, CropLoader2
+from rcnn.core.module import MutableModule
+from rcnn.utils.load_data import load_gt_roidb, merge_roidb, filter_roidb
+from rcnn.utils.load_model import load_param
+
+
+def get_fixed_params(symbol, fixed_param):
+    if not config.LAYER_FIX:
+        return []
+    fixed_param_names = []
+    #for name in symbol.list_arguments():
+    #  for f in fixed_param:
+    #    if re.match(f, name):
+    #      fixed_param_names.append(name)
+    #pre = 'mobilenetv20_features_linearbottleneck'
+    idx = 0
+    for name in symbol.list_arguments():
+        #print(idx, name)
+        if idx < 7 and name != 'data':
+            fixed_param_names.append(name)
+        #elif name.startswith('stage1_'):
+        #  fixed_param_names.append(name)
+        if name.find('upsampling') >= 0:
+            fixed_param_names.append(name)
+
+        idx += 1
+    return fixed_param_names
+
+
+def train_net(args,
+              ctx,
+              pretrained,
+              epoch,
+              prefix,
+              begin_epoch,
+              end_epoch,
+              lr=0.001,
+              lr_step='5'):
+    # setup config
+    #init_config()
+    #print(config)
+    # setup multi-gpu
+
+    input_batch_size = config.TRAIN.BATCH_IMAGES * len(ctx)
+
+    # print config
+    logger.info(pprint.pformat(config))
+
+    # load dataset and prepare imdb for training
+    image_sets = [iset for iset in args.image_set.split('+')]
+    roidbs = [
+        load_gt_roidb(args.dataset,
+                      image_set,
+                      args.root_path,
+                      args.dataset_path,
+                      flip=not args.no_flip) for image_set in image_sets
+    ]
+    #roidb = merge_roidb(roidbs)
+    #roidb = filter_roidb(roidb)
+    roidb = roidbs[0]
+
+    # load symbol
+    #sym = eval('get_' + args.network + '_train')(num_classes=config.NUM_CLASSES, num_anchors=config.NUM_ANCHORS)
+    #feat_sym = sym.get_internals()['rpn_cls_score_output']
+    #train_data = AnchorLoader(feat_sym, roidb, batch_size=input_batch_size, shuffle=not args.no_shuffle,
+    #                          ctx=ctx, work_load_list=args.work_load_list,
+    #                          feat_stride=config.RPN_FEAT_STRIDE, anchor_scales=config.ANCHOR_SCALES,
+    #                          anchor_ratios=config.ANCHOR_RATIOS, aspect_grouping=config.TRAIN.ASPECT_GROUPING)
+
+    # load and initialize params
+    sym = None
+    if len(pretrained) == 0:
+        arg_params = {}
+        aux_params = {}
+    else:
+        logger.info('loading %s,%d' % (pretrained, epoch))
+        sym, arg_params, aux_params = mx.model.load_checkpoint(
+            pretrained, epoch)
+        #arg_params, aux_params = load_param(pretrained, epoch, convert=True)
+        #for k in ['rpn_conv_3x3', 'rpn_cls_score', 'rpn_bbox_pred', 'cls_score', 'bbox_pred']:
+        #  _k = k+"_weight"
+        #  if _k in arg_shape_dict:
+        #    v = 0.001 if _k.startswith('bbox_') else 0.01
+        #    arg_params[_k] = mx.random.normal(0, v, shape=arg_shape_dict[_k])
+        #    print('init %s with normal %.5f'%(_k,v))
+        #  _k = k+"_bias"
+        #  if _k in arg_shape_dict:
+        #    arg_params[_k] = mx.nd.zeros(shape=arg_shape_dict[_k])
+        #    print('init %s with zero'%(_k))
+
+    sym = eval('get_' + args.network + '_train')(sym)
+    #print(sym.get_internals())
+    feat_sym = []
+    for stride in config.RPN_FEAT_STRIDE:
+        feat_sym.append(
+            sym.get_internals()['face_rpn_cls_score_stride%s_output' % stride])
+
+    train_data = CropLoader(feat_sym,
+                            roidb,
+                            batch_size=input_batch_size,
+                            shuffle=not args.no_shuffle,
+                            ctx=ctx,
+                            work_load_list=args.work_load_list)
+
+    # infer max shape
+    max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]),
+                                max([v[1] for v in config.SCALES])))]
+    #max_data_shape = [('data', (1, 3, max([v[1] for v in config.SCALES]), max([v[1] for v in config.SCALES])))]
+    max_data_shape, max_label_shape = train_data.infer_shape(max_data_shape)
+    max_data_shape.append(('gt_boxes', (1, roidb[0]['max_num_boxes'], 5)))
+    logger.info('providing maximum shape %s %s' %
+                (max_data_shape, max_label_shape))
+
+    # infer shape
+    data_shape_dict = dict(train_data.provide_data + train_data.provide_label)
+    arg_shape, out_shape, aux_shape = sym.infer_shape(**data_shape_dict)
+    arg_shape_dict = dict(zip(sym.list_arguments(), arg_shape))
+    out_shape_dict = dict(zip(sym.list_outputs(), out_shape))
+    aux_shape_dict = dict(zip(sym.list_auxiliary_states(), aux_shape))
+    logger.info('output shape %s' % pprint.pformat(out_shape_dict))
+
+    for k in arg_shape_dict:
+        v = arg_shape_dict[k]
+        if k.find('upsampling') >= 0:
+            print('initializing upsampling_weight', k)
+            arg_params[k] = mx.nd.zeros(shape=v)
+            init = mx.init.Initializer()
+            init._init_bilinear(k, arg_params[k])
+            #print(args[k])
+
+    # check parameter shapes
+    #for k in sym.list_arguments():
+    #    if k in data_shape_dict:
+    #        continue
+    #    assert k in arg_params, k + ' not initialized'
+    #    assert arg_params[k].shape == arg_shape_dict[k], \
+    #        'shape inconsistent for ' + k + ' inferred ' + str(arg_shape_dict[k]) + ' provided ' + str(arg_params[k].shape)
+    #for k in sym.list_auxiliary_states():
+    #    assert k in aux_params, k + ' not initialized'
+    #    assert aux_params[k].shape == aux_shape_dict[k], \
+    #        'shape inconsistent for ' + k + ' inferred ' + str(aux_shape_dict[k]) + ' provided ' + str(aux_params[k].shape)
+
+    fixed_param_prefix = config.FIXED_PARAMS
+    # create solver
+    data_names = [k[0] for k in train_data.provide_data]
+    label_names = [k[0] for k in train_data.provide_label]
+    fixed_param_names = get_fixed_params(sym, fixed_param_prefix)
+    print('fixed', fixed_param_names, file=sys.stderr)
+    mod = Module(sym,
+                 data_names=data_names,
+                 label_names=label_names,
+                 logger=logger,
+                 context=ctx,
+                 work_load_list=args.work_load_list,
+                 fixed_param_names=fixed_param_names)
+
+    # metric
+    eval_metrics = mx.metric.CompositeEvalMetric()
+    mid = 0
+    for m in range(len(config.RPN_FEAT_STRIDE)):
+        stride = config.RPN_FEAT_STRIDE[m]
+        #mid = m*MSTEP
+        _metric = metric.RPNAccMetric(pred_idx=mid,
+                                      label_idx=mid + 1,
+                                      name='RPNAcc_s%s' % stride)
+        eval_metrics.add(_metric)
+        mid += 2
+        #_metric = metric.RPNLogLossMetric(pred_idx=mid, label_idx=mid+1)
+        #eval_metrics.add(_metric)
+
+        _metric = metric.RPNL1LossMetric(loss_idx=mid,
+                                         weight_idx=mid + 1,
+                                         name='RPNL1Loss_s%s' % stride)
+        eval_metrics.add(_metric)
+        mid += 2
+        if config.FACE_LANDMARK:
+            _metric = metric.RPNL1LossMetric(loss_idx=mid,
+                                             weight_idx=mid + 1,
+                                             name='RPNLandMarkL1Loss_s%s' %
+                                             stride)
+            eval_metrics.add(_metric)
+            mid += 2
+        if config.HEAD_BOX:
+            _metric = metric.RPNAccMetric(pred_idx=mid,
+                                          label_idx=mid + 1,
+                                          name='RPNAcc_head_s%s' % stride)
+            eval_metrics.add(_metric)
+            mid += 2
+            #_metric = metric.RPNLogLossMetric(pred_idx=mid, label_idx=mid+1)
+            #eval_metrics.add(_metric)
+
+            _metric = metric.RPNL1LossMetric(loss_idx=mid,
+                                             weight_idx=mid + 1,
+                                             name='RPNL1Loss_head_s%s' %
+                                             stride)
+            eval_metrics.add(_metric)
+            mid += 2
+        if config.CASCADE > 0:
+            for _idx in range(config.CASCADE):
+                if stride in config.CASCADE_CLS_STRIDES:
+                    _metric = metric.RPNAccMetric(pred_idx=mid,
+                                                  label_idx=mid + 1,
+                                                  name='RPNAccCAS%d_s%s' %
+                                                  (_idx, stride))
+                    eval_metrics.add(_metric)
+                    mid += 2
+                if stride in config.CASCADE_BBOX_STRIDES:
+                    _metric = metric.RPNL1LossMetric(
+                        loss_idx=mid,
+                        weight_idx=mid + 1,
+                        name='RPNL1LossCAS%d_s%s' % (_idx, stride))
+                    eval_metrics.add(_metric)
+                    mid += 2
+
+    # callback
+    #means = np.tile(np.array(config.TRAIN.BBOX_MEANS), config.NUM_CLASSES)
+    #stds = np.tile(np.array(config.TRAIN.BBOX_STDS), config.NUM_CLASSES)
+    #epoch_end_callback = callback.do_checkpoint(prefix, means, stds)
+    epoch_end_callback = None
+    # decide learning rate
+    #base_lr = lr
+    #lr_factor = 0.1
+    #lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
+
+    lr_epoch = [int(epoch) for epoch in lr_step.split(',')]
+    lr_epoch_diff = [
+        epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch
+    ]
+    lr_iters = [
+        int(epoch * len(roidb) / input_batch_size) for epoch in lr_epoch_diff
+    ]
+    iter_per_epoch = int(len(roidb) / input_batch_size)
+
+    lr_steps = []
+    if len(lr_iters) == 5:
+        factors = [0.5, 0.5, 0.4, 0.1, 0.1]
+        for i in range(5):
+            lr_steps.append((lr_iters[i], factors[i]))
+    elif len(lr_iters) == 8:  #warmup
+        for li in lr_iters[0:5]:
+            lr_steps.append((li, 1.5849))
+        for li in lr_iters[5:]:
+            lr_steps.append((li, 0.1))
+    else:
+        for li in lr_iters:
+            lr_steps.append((li, 0.1))
+    #lr_steps = [ (10,0.1), (20, 0.1) ] #XXX
+
+    end_epoch = 10000
+    logger.info('lr %f lr_epoch_diff %s lr_steps %s' %
+                (lr, lr_epoch_diff, lr_steps))
+    # optimizer
+    opt = optimizer.SGD(learning_rate=lr,
+                        momentum=0.9,
+                        wd=args.wd,
+                        rescale_grad=1.0 / len(ctx),
+                        clip_gradient=None)
+    initializer = mx.init.Xavier()
+    #initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
+
+    train_data = mx.io.PrefetchingIter(train_data)
+
+    _cb = mx.callback.Speedometer(train_data.batch_size,
+                                  frequent=args.frequent,
+                                  auto_reset=False)
+    global_step = [0]
+
+    def save_model(epoch):
+        arg, aux = mod.get_params()
+        all_layers = mod.symbol.get_internals()
+        outs = []
+        for stride in config.RPN_FEAT_STRIDE:
+            num_anchors = config.RPN_ANCHOR_CFG[str(stride)]['NUM_ANCHORS']
+            if config.CASCADE > 0:
+                _name = 'face_rpn_cls_score_stride%d_output' % (stride)
+                cls_pred = all_layers[_name]
+                cls_pred = mx.symbol.Reshape(data=cls_pred,
+                                             shape=(0, 2, -1, 0))
+
+                cls_pred = mx.symbol.SoftmaxActivation(data=cls_pred,
+                                                       mode="channel")
+                cls_pred = mx.symbol.Reshape(data=cls_pred,
+                                             shape=(0, 2 * num_anchors, -1, 0))
+                outs.append(cls_pred)
+                _name = 'face_rpn_bbox_pred_stride%d_output' % stride
+                rpn_bbox_pred = all_layers[_name]
+                outs.append(rpn_bbox_pred)
+                if config.FACE_LANDMARK:
+                    _name = 'face_rpn_landmark_pred_stride%d_output' % stride
+                    rpn_landmark_pred = all_layers[_name]
+                    outs.append(rpn_landmark_pred)
+                for casid in range(config.CASCADE):
+                    if stride in config.CASCADE_CLS_STRIDES:
+                        _name = 'face_rpn_cls_score_stride%d_cas%d_output' % (
+                            stride, casid)
+                        cls_pred = all_layers[_name]
+                        cls_pred = mx.symbol.Reshape(data=cls_pred,
+                                                     shape=(0, 2, -1, 0))
+                        cls_pred = mx.symbol.SoftmaxActivation(data=cls_pred,
+                                                               mode="channel")
+                        cls_pred = mx.symbol.Reshape(data=cls_pred,
+                                                     shape=(0, 2 * num_anchors,
+                                                            -1, 0))
+                        outs.append(cls_pred)
+                    if stride in config.CASCADE_BBOX_STRIDES:
+                        _name = 'face_rpn_bbox_pred_stride%d_cas%d_output' % (
+                            stride, casid)
+                        bbox_pred = all_layers[_name]
+                        outs.append(bbox_pred)
+            else:
+                _name = 'face_rpn_cls_score_stride%d_output' % stride
+                rpn_cls_score = all_layers[_name]
+
+                # prepare rpn data
+                rpn_cls_score_reshape = mx.symbol.Reshape(
+                    data=rpn_cls_score,
+                    shape=(0, 2, -1, 0),
+                    name="face_rpn_cls_score_reshape_stride%d" % stride)
+
+                rpn_cls_prob = mx.symbol.SoftmaxActivation(
+                    data=rpn_cls_score_reshape,
+                    mode="channel",
+                    name="face_rpn_cls_prob_stride%d" % stride)
+                rpn_cls_prob_reshape = mx.symbol.Reshape(
+                    data=rpn_cls_prob,
+                    shape=(0, 2 * num_anchors, -1, 0),
+                    name='face_rpn_cls_prob_reshape_stride%d' % stride)
+                _name = 'face_rpn_bbox_pred_stride%d_output' % stride
+                rpn_bbox_pred = all_layers[_name]
+                outs.append(rpn_cls_prob_reshape)
+                outs.append(rpn_bbox_pred)
+                if config.FACE_LANDMARK:
+                    _name = 'face_rpn_landmark_pred_stride%d_output' % stride
+                    rpn_landmark_pred = all_layers[_name]
+                    outs.append(rpn_landmark_pred)
+        _sym = mx.sym.Group(outs)
+        mx.model.save_checkpoint(prefix, epoch, _sym, arg, aux)
+
+    def _batch_callback(param):
+        #global global_step
+        _cb(param)
+        global_step[0] += 1
+        mbatch = global_step[0]
+        for step in lr_steps:
+            if mbatch == step[0]:
+                opt.lr *= step[1]
+                print('lr change to',
+                      opt.lr,
+                      ' in batch',
+                      mbatch,
+                      file=sys.stderr)
+                break
+
+        if mbatch % iter_per_epoch == 0:
+            print('saving checkpoint', mbatch, file=sys.stderr)
+            save_model(0)
+        if mbatch == lr_steps[-1][0]:
+            print('saving final checkpoint', mbatch, file=sys.stderr)
+            save_model(0)
+            #arg, aux = mod.get_params()
+            #mx.model.save_checkpoint(prefix, 99, mod.symbol, arg, aux)
+            sys.exit(0)
+
+    # train
+    mod.fit(train_data,
+            eval_metric=eval_metrics,
+            epoch_end_callback=epoch_end_callback,
+            batch_end_callback=_batch_callback,
+            kvstore=args.kvstore,
+            optimizer=opt,
+            initializer=initializer,
+            allow_missing=True,
+            arg_params=arg_params,
+            aux_params=aux_params,
+            begin_epoch=begin_epoch,
+            num_epoch=end_epoch)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train RetinaFace')
+    # general
+    parser.add_argument('--network',
+                        help='network name',
+                        default=default.network,
+                        type=str)
+    parser.add_argument('--dataset',
+                        help='dataset name',
+                        default=default.dataset,
+                        type=str)
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset)
+    parser.add_argument('--image_set',
+                        help='image_set name',
+                        default=default.image_set,
+                        type=str)
+    parser.add_argument('--root_path',
+                        help='output data folder',
+                        default=default.root_path,
+                        type=str)
+    parser.add_argument('--dataset_path',
+                        help='dataset path',
+                        default=default.dataset_path,
+                        type=str)
+    # training
+    parser.add_argument('--frequent',
+                        help='frequency of logging',
+                        default=default.frequent,
+                        type=int)
+    parser.add_argument('--kvstore',
+                        help='the kv-store type',
+                        default=default.kvstore,
+                        type=str)
+    parser.add_argument('--work_load_list',
+                        help='work load for different devices',
+                        default=None,
+                        type=list)
+    parser.add_argument('--no_flip',
+                        help='disable flip images',
+                        action='store_true')
+    parser.add_argument('--no_shuffle',
+                        help='disable random shuffle',
+                        action='store_true')
+    # e2e
+    #parser.add_argument('--gpus', help='GPU device to train with', default='0,1,2,3', type=str)
+    parser.add_argument('--pretrained',
+                        help='pretrained model prefix',
+                        default=default.pretrained,
+                        type=str)
+    parser.add_argument('--pretrained_epoch',
+                        help='pretrained model epoch',
+                        default=default.pretrained_epoch,
+                        type=int)
+    parser.add_argument('--prefix',
+                        help='new model prefix',
+                        default=default.prefix,
+                        type=str)
+    parser.add_argument('--begin_epoch',
+                        help='begin epoch of training, use with resume',
+                        default=0,
+                        type=int)
+    parser.add_argument('--end_epoch',
+                        help='end epoch of training',
+                        default=default.end_epoch,
+                        type=int)
+    parser.add_argument('--lr',
+                        help='base learning rate',
+                        default=default.lr,
+                        type=float)
+    parser.add_argument('--lr_step',
+                        help='learning rate steps (in epoch)',
+                        default=default.lr_step,
+                        type=str)
+    parser.add_argument('--wd',
+                        help='weight decay',
+                        default=default.wd,
+                        type=float)
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger.info('Called with argument: %s' % args)
+    #ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    train_net(args,
+              ctx,
+              args.pretrained,
+              args.pretrained_epoch,
+              args.prefix,
+              args.begin_epoch,
+              args.end_epoch,
+              lr=args.lr,
+              lr_step=args.lr_step)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/retinaface_anticov/README.md b/detection/retinaface_anticov/README.md
new file mode 100644
index 000000000..97b3b2449
--- /dev/null
+++ b/detection/retinaface_anticov/README.md
@@ -0,0 +1,36 @@
+# RetinaFace Anti Cov Face Detector
+
+## Introduction
+
+RetinaFace-Anti-Cov is a customized one stage face detector to help people protect themselves from CovID-19.
+
+![demoimg1](https://insightface.ai/assets/img/github/cov_test.jpg)
+
+
+## Testing
+
+Please check ``test.py`` for testing.
+
+Make sure that you set ``network='net3l'`` instead of ``'net3'`` for 'mnet_cov2' model, otherwise you will get incorrect landmarks.
+
+## Pretrained Models
+
+~~MobileNet0.25([baidu cloud](https://pan.baidu.com/s/1p8n4R2W-9WmmBWxYQEFcWg),code: fmfm)~~
+
+Better: MobileNet0.25 ([baidu cloud](https://pan.baidu.com/s/16ihzPxjTObdbv0D6P6LmEQ), code: j3b6, [dropbox](https://www.dropbox.com/s/6rhhxsbh2qik65k/cov2.zip?dl=0))
+
+
+
+## References
+
+```
+  
+@inproceedings{deng2019retinaface,
+title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
+author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
+booktitle={arxiv},
+year={2019}
+}
+```
+
+
diff --git a/detection/retinaface_anticov/rcnn/processing/__init__.py b/detection/retinaface_anticov/rcnn/processing/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/detection/retinaface_anticov/rcnn/processing/assign_levels.py b/detection/retinaface_anticov/rcnn/processing/assign_levels.py
new file mode 100755
index 000000000..012d73d21
--- /dev/null
+++ b/detection/retinaface_anticov/rcnn/processing/assign_levels.py
@@ -0,0 +1,36 @@
+from rcnn.config import config
+import numpy as np
+
+
+def compute_assign_targets(rois, threshold):
+    rois_area = np.sqrt(
+        (rois[:, 2] - rois[:, 0] + 1) * (rois[:, 3] - rois[:, 1] + 1))
+    num_rois = np.shape(rois)[0]
+    assign_levels = np.zeros(num_rois, dtype=np.uint8)
+    for i, stride in enumerate(config.RCNN_FEAT_STRIDE):
+        thd = threshold[i]
+        idx = np.logical_and(thd[1] <= rois_area, rois_area < thd[0])
+        assign_levels[idx] = stride
+
+    assert 0 not in assign_levels, "All rois should assign to specify levels."
+    return assign_levels
+
+
+def add_assign_targets(roidb):
+    """
+    given roidb, add ['assign_level']
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    """
+    print 'add assign targets'
+    assert len(roidb) > 0
+    assert 'boxes' in roidb[0]
+
+    area_threshold = [[np.inf, 448], [448, 224], [224, 112], [112, 0]]
+
+    assert len(config.RCNN_FEAT_STRIDE) == len(area_threshold)
+
+    num_images = len(roidb)
+    for im_i in range(num_images):
+        rois = roidb[im_i]['boxes']
+        roidb[im_i]['assign_levels'] = compute_assign_targets(
+            rois, area_threshold)
diff --git a/detection/retinaface_anticov/rcnn/processing/bbox_regression.py b/detection/retinaface_anticov/rcnn/processing/bbox_regression.py
new file mode 100644
index 000000000..0eaf917a6
--- /dev/null
+++ b/detection/retinaface_anticov/rcnn/processing/bbox_regression.py
@@ -0,0 +1,263 @@
+"""
+This file has functions about generating bounding box regression targets
+"""
+
+from ..pycocotools.mask import encode
+import numpy as np
+
+from ..logger import logger
+from .bbox_transform import bbox_overlaps, bbox_transform
+from rcnn.config import config
+import math
+import cv2
+import PIL.Image as Image
+import threading
+import Queue
+
+
+def compute_bbox_regression_targets(rois, overlaps, labels):
+    """
+    given rois, overlaps, gt labels, compute bounding box regression targets
+    :param rois: roidb[i]['boxes'] k * 4
+    :param overlaps: roidb[i]['max_overlaps'] k * 1
+    :param labels: roidb[i]['max_classes'] k * 1
+    :return: targets[i][class, dx, dy, dw, dh] k * 5
+    """
+    # Ensure ROIs are floats
+    rois = rois.astype(np.float, copy=False)
+
+    # Sanity check
+    if len(rois) != len(overlaps):
+        logger.warning('bbox regression: len(rois) != len(overlaps)')
+
+    # Indices of ground-truth ROIs
+    gt_inds = np.where(overlaps == 1)[0]
+    if len(gt_inds) == 0:
+        logger.warning('bbox regression: len(gt_inds) == 0')
+
+    # Indices of examples for which we try to make predictions
+    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
+
+    # Get IoU overlap between each ex ROI and gt ROI
+    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
+
+    # Find which gt ROI each ex ROI has max overlap with:
+    # this will be the ex ROI's gt target
+    gt_assignment = ex_gt_overlaps.argmax(axis=1)
+    gt_rois = rois[gt_inds[gt_assignment], :]
+    ex_rois = rois[ex_inds, :]
+
+    targets = np.zeros((rois.shape[0], 5), dtype=np.float32)
+    targets[ex_inds, 0] = labels[ex_inds]
+    targets[ex_inds, 1:] = bbox_transform(ex_rois, gt_rois)
+    return targets
+
+
+def add_bbox_regression_targets(roidb):
+    """
+    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    :return: means, std variances of targets
+    """
+    logger.info('bbox regression: add bounding box regression targets')
+    assert len(roidb) > 0
+    assert 'max_classes' in roidb[0]
+
+    num_images = len(roidb)
+    num_classes = roidb[0]['gt_overlaps'].shape[1]
+    for im_i in range(num_images):
+        rois = roidb[im_i]['boxes']
+        max_overlaps = roidb[im_i]['max_overlaps']
+        max_classes = roidb[im_i]['max_classes']
+        roidb[im_i]['bbox_targets'] = compute_bbox_regression_targets(
+            rois, max_overlaps, max_classes)
+
+    if config.TRAIN.BBOX_NORMALIZATION_PRECOMPUTED:
+        # use fixed / precomputed means and stds instead of empirical values
+        means = np.tile(np.array(config.TRAIN.BBOX_MEANS), (num_classes, 1))
+        stds = np.tile(np.array(config.TRAIN.BBOX_STDS), (num_classes, 1))
+    else:
+        # compute mean, std values
+        class_counts = np.zeros((num_classes, 1)) + 1e-14
+        sums = np.zeros((num_classes, 4))
+        squared_sums = np.zeros((num_classes, 4))
+        for im_i in range(num_images):
+            targets = roidb[im_i]['bbox_targets']
+            for cls in range(1, num_classes):
+                cls_indexes = np.where(targets[:, 0] == cls)[0]
+                if cls_indexes.size > 0:
+                    class_counts[cls] += cls_indexes.size
+                    sums[cls, :] += targets[cls_indexes, 1:].sum(axis=0)
+                    squared_sums[cls, :] += (targets[cls_indexes,
+                                                     1:]**2).sum(axis=0)
+
+        means = sums / class_counts
+        # var(x) = E(x^2) - E(x)^2
+        stds = np.sqrt(squared_sums / class_counts - means**2)
+
+    # normalized targets
+    for im_i in range(num_images):
+        targets = roidb[im_i]['bbox_targets']
+        for cls in range(1, num_classes):
+            cls_indexes = np.where(targets[:, 0] == cls)[0]
+            roidb[im_i]['bbox_targets'][cls_indexes, 1:] -= means[cls, :]
+            roidb[im_i]['bbox_targets'][cls_indexes, 1:] /= stds[cls, :]
+
+    return means.ravel(), stds.ravel()
+
+
+def expand_bbox_regression_targets(bbox_targets_data, num_classes):
+    """
+    expand from 5 to 4 * num_classes; only the right class has non-zero bbox regression targets
+    :param bbox_targets_data: [k * 5]
+    :param num_classes: number of classes
+    :return: bbox target processed [k * 4 num_classes]
+    bbox_weights ! only foreground boxes have bbox regression computation!
+    """
+    classes = bbox_targets_data[:, 0]
+    bbox_targets = np.zeros((classes.size, 4 * num_classes), dtype=np.float32)
+    bbox_weights = np.zeros(bbox_targets.shape, dtype=np.float32)
+    indexes = np.where(classes > 0)[0]
+    for index in indexes:
+        cls = classes[index]
+        start = int(4 * cls)
+        end = start + 4
+        bbox_targets[index, start:end] = bbox_targets_data[index, 1:]
+        bbox_weights[index, start:end] = config.TRAIN.BBOX_WEIGHTS
+    return bbox_targets, bbox_weights
+
+
+def compute_mask_and_label(ex_rois, ex_labels, seg, flipped):
+    # assert os.path.exists(seg_gt), 'Path does not exist: {}'.format(seg_gt)
+    # im = Image.open(seg_gt)
+    # pixel = list(im.getdata())
+    # pixel = np.array(pixel).reshape([im.size[1], im.size[0]])
+    im = Image.open(seg)
+    pixel = list(im.getdata())
+    ins_seg = np.array(pixel).reshape([im.size[1], im.size[0]])
+    if flipped:
+        ins_seg = ins_seg[:, ::-1]
+    rois = ex_rois
+    n_rois = ex_rois.shape[0]
+    label = ex_labels
+    class_id = config.CLASS_ID
+    mask_target = np.zeros((n_rois, 28, 28), dtype=np.int8)
+    mask_label = np.zeros((n_rois), dtype=np.int8)
+    for n in range(n_rois):
+        target = ins_seg[int(rois[n, 1]):int(rois[n, 3]),
+                         int(rois[n, 0]):int(rois[n, 2])]
+        ids = np.unique(target)
+        ins_id = 0
+        max_count = 0
+        for id in ids:
+            if math.floor(id / 1000) == class_id[int(label[int(n)])]:
+                px = np.where(ins_seg == int(id))
+                x_min = np.min(px[1])
+                y_min = np.min(px[0])
+                x_max = np.max(px[1])
+                y_max = np.max(px[0])
+                x1 = max(rois[n, 0], x_min)
+                y1 = max(rois[n, 1], y_min)
+                x2 = min(rois[n, 2], x_max)
+                y2 = min(rois[n, 3], y_max)
+                iou = (x2 - x1) * (y2 - y1)
+                iou = iou / ((rois[n, 2] - rois[n, 0]) *
+                             (rois[n, 3] - rois[n, 1]) + (x_max - x_min) *
+                             (y_max - y_min) - iou)
+                if iou > max_count:
+                    ins_id = id
+                    max_count = iou
+
+        if max_count == 0:
+            continue
+        # print max_count
+        mask = np.zeros(target.shape)
+        idx = np.where(target == ins_id)
+        mask[idx] = 1
+        mask = cv2.resize(mask, (28, 28), interpolation=cv2.INTER_NEAREST)
+
+        mask_target[n] = mask
+        mask_label[n] = label[int(n)]
+    return mask_target, mask_label
+
+
+def compute_bbox_mask_targets_and_label(rois, overlaps, labels, seg, flipped):
+    """
+    given rois, overlaps, gt labels, seg, compute bounding box mask targets
+    :param rois: roidb[i]['boxes'] k * 4
+    :param overlaps: roidb[i]['max_overlaps'] k * 1
+    :param labels: roidb[i]['max_classes'] k * 1
+    :return: targets[i][class, dx, dy, dw, dh] k * 5
+    """
+    # Ensure ROIs are floats
+    rois = rois.astype(np.float, copy=False)
+
+    # Sanity check
+    if len(rois) != len(overlaps):
+        print 'bbox regression: this should not happen'
+
+    # Indices of ground-truth ROIs
+    gt_inds = np.where(overlaps == 1)[0]
+    if len(gt_inds) == 0:
+        print 'something wrong : zero ground truth rois'
+    # Indices of examples for which we try to make predictions
+    ex_inds = np.where(overlaps >= config.TRAIN.BBOX_REGRESSION_THRESH)[0]
+
+    # Get IoU overlap between each ex ROI and gt ROI
+    ex_gt_overlaps = bbox_overlaps(rois[ex_inds, :], rois[gt_inds, :])
+
+    # Find which gt ROI each ex ROI has max overlap with:
+    # this will be the ex ROI's gt target
+    gt_assignment = ex_gt_overlaps.argmax(axis=1)
+    gt_rois = rois[gt_inds[gt_assignment], :]
+    ex_rois = rois[ex_inds, :]
+
+    mask_targets, mask_label = compute_mask_and_label(ex_rois, labels[ex_inds],
+                                                      seg, flipped)
+    return mask_targets, mask_label, ex_inds
+
+
+def add_mask_targets(roidb):
+    """
+    given roidb, add ['bbox_targets'] and normalize bounding box regression targets
+    :param roidb: roidb to be processed. must have gone through imdb.prepare_roidb
+    :return: means, std variances of targets
+    """
+    print 'add bounding box mask targets'
+    assert len(roidb) > 0
+    assert 'max_classes' in roidb[0]
+
+    num_images = len(roidb)
+
+    # Multi threads processing
+    im_quene = Queue.Queue(maxsize=0)
+    for im_i in range(num_images):
+        im_quene.put(im_i)
+
+    def process():
+        while not im_quene.empty():
+            im_i = im_quene.get()
+            print "-----process img {}".format(im_i)
+            rois = roidb[im_i]['boxes']
+            max_overlaps = roidb[im_i]['max_overlaps']
+            max_classes = roidb[im_i]['max_classes']
+            ins_seg = roidb[im_i]['ins_seg']
+            flipped = roidb[im_i]['flipped']
+            roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
+                compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg, flipped)
+
+    threads = [threading.Thread(target=process, args=()) for i in range(10)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+    # Single thread
+    # for im_i in range(num_images):
+    #     print "-----processing img {}".format(im_i)
+    #     rois = roidb[im_i]['boxes']
+    #     max_overlaps = roidb[im_i]['max_overlaps']
+    #     max_classes = roidb[im_i]['max_classes']
+    #     ins_seg = roidb[im_i]['ins_seg']
+    #     # roidb[im_i]['mask_targets'] = compute_bbox_mask_targets(rois, max_overlaps, max_classes, ins_seg)
+    #     roidb[im_i]['mask_targets'], roidb[im_i]['mask_labels'], roidb[im_i]['mask_inds'] = \
+    #         compute_bbox_mask_targets_and_label(rois, max_overlaps, max_classes, ins_seg)
diff --git a/detection/retinaface_anticov/rcnn/processing/bbox_transform.py b/detection/retinaface_anticov/rcnn/processing/bbox_transform.py
new file mode 100644
index 000000000..5ee3646ab
--- /dev/null
+++ b/detection/retinaface_anticov/rcnn/processing/bbox_transform.py
@@ -0,0 +1,223 @@
+import numpy as np
+from ..cython.bbox import bbox_overlaps_cython
+#from rcnn.config import config
+
+
+def bbox_overlaps(boxes, query_boxes):
+    return bbox_overlaps_cython(boxes, query_boxes)
+
+
+def bbox_overlaps_py(boxes, query_boxes):
+    """
+    determine overlaps between boxes and query_boxes
+    :param boxes: n * 4 bounding boxes
+    :param query_boxes: k * 4 bounding boxes
+    :return: overlaps: n * k overlaps
+    """
+    n_ = boxes.shape[0]
+    k_ = query_boxes.shape[0]
+    overlaps = np.zeros((n_, k_), dtype=np.float)
+    for k in range(k_):
+        query_box_area = (query_boxes[k, 2] - query_boxes[k, 0] +
+                          1) * (query_boxes[k, 3] - query_boxes[k, 1] + 1)
+        for n in range(n_):
+            iw = min(boxes[n, 2], query_boxes[k, 2]) - max(
+                boxes[n, 0], query_boxes[k, 0]) + 1
+            if iw > 0:
+                ih = min(boxes[n, 3], query_boxes[k, 3]) - max(
+                    boxes[n, 1], query_boxes[k, 1]) + 1
+                if ih > 0:
+                    box_area = (boxes[n, 2] - boxes[n, 0] +
+                                1) * (boxes[n, 3] - boxes[n, 1] + 1)
+                    all_area = float(box_area + query_box_area - iw * ih)
+                    overlaps[n, k] = iw * ih / all_area
+    return overlaps
+
+
+def clip_boxes(boxes, im_shape):
+    """
+    Clip boxes to image boundaries.
+    :param boxes: [N, 4* num_classes]
+    :param im_shape: tuple of 2
+    :return: [N, 4* num_classes]
+    """
+    # x1 >= 0
+    boxes[:, 0::4] = np.maximum(np.minimum(boxes[:, 0::4], im_shape[1] - 1), 0)
+    # y1 >= 0
+    boxes[:, 1::4] = np.maximum(np.minimum(boxes[:, 1::4], im_shape[0] - 1), 0)
+    # x2 < im_shape[1]
+    boxes[:, 2::4] = np.maximum(np.minimum(boxes[:, 2::4], im_shape[1] - 1), 0)
+    # y2 < im_shape[0]
+    boxes[:, 3::4] = np.maximum(np.minimum(boxes[:, 3::4], im_shape[0] - 1), 0)
+    return boxes
+
+
+def nonlinear_transform(ex_rois, gt_rois):
+    """
+    compute bounding box regression targets from ex_rois to gt_rois
+    :param ex_rois: [N, 4]
+    :param gt_rois: [N, 4]
+    :return: [N, 4]
+    """
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+
+    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
+    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
+    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
+    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
+
+    gt_widths = gt_rois[:, 2] - gt_rois[:, 0] + 1.0
+    gt_heights = gt_rois[:, 3] - gt_rois[:, 1] + 1.0
+    gt_ctr_x = gt_rois[:, 0] + 0.5 * (gt_widths - 1.0)
+    gt_ctr_y = gt_rois[:, 1] + 0.5 * (gt_heights - 1.0)
+
+    targets_dx = (gt_ctr_x - ex_ctr_x) / (ex_widths + 1e-14)
+    targets_dy = (gt_ctr_y - ex_ctr_y) / (ex_heights + 1e-14)
+    targets_dw = np.log(gt_widths / ex_widths)
+    targets_dh = np.log(gt_heights / ex_heights)
+
+    if gt_rois.shape[1] <= 4:
+        targets = np.vstack(
+            (targets_dx, targets_dy, targets_dw, targets_dh)).transpose()
+        return targets
+    else:
+        targets = [targets_dx, targets_dy, targets_dw, targets_dh]
+        #if config.USE_BLUR:
+        #  for i in range(4, gt_rois.shape[1]):
+        #    t = gt_rois[:,i]
+        #    targets.append(t)
+        targets = np.vstack(targets).transpose()
+        return targets
+
+
+def landmark_transform(ex_rois, gt_rois):
+
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+
+    ex_widths = ex_rois[:, 2] - ex_rois[:, 0] + 1.0
+    ex_heights = ex_rois[:, 3] - ex_rois[:, 1] + 1.0
+    ex_ctr_x = ex_rois[:, 0] + 0.5 * (ex_widths - 1.0)
+    ex_ctr_y = ex_rois[:, 1] + 0.5 * (ex_heights - 1.0)
+
+    targets = []
+    for i in range(gt_rois.shape[1]):
+        for j in range(gt_rois.shape[2]):
+            #if not config.USE_OCCLUSION and j==2:
+            #  continue
+            if j == 2:
+                continue
+            if j == 0:  #w
+                target = (gt_rois[:, i, j] - ex_ctr_x) / (ex_widths + 1e-14)
+            elif j == 1:  #h
+                target = (gt_rois[:, i, j] - ex_ctr_y) / (ex_heights + 1e-14)
+            else:  #visibile
+                target = gt_rois[:, i, j]
+            targets.append(target)
+
+    targets = np.vstack(targets).transpose()
+    return targets
+
+
+def nonlinear_pred(boxes, box_deltas):
+    """
+    Transform the set of class-agnostic boxes into class-specific boxes
+    by applying the predicted offsets (box_deltas)
+    :param boxes: !important [N 4]
+    :param box_deltas: [N, 4 * num_classes]
+    :return: [N 4 * num_classes]
+    """
+    if boxes.shape[0] == 0:
+        return np.zeros((0, box_deltas.shape[1]))
+
+    boxes = boxes.astype(np.float, copy=False)
+    widths = boxes[:, 2] - boxes[:, 0] + 1.0
+    heights = boxes[:, 3] - boxes[:, 1] + 1.0
+    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+
+    dx = box_deltas[:, 0::4]
+    dy = box_deltas[:, 1::4]
+    dw = box_deltas[:, 2::4]
+    dh = box_deltas[:, 3::4]
+
+    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
+    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
+    pred_w = np.exp(dw) * widths[:, np.newaxis]
+    pred_h = np.exp(dh) * heights[:, np.newaxis]
+
+    pred_boxes = np.zeros(box_deltas.shape)
+    # x1
+    pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * (pred_w - 1.0)
+    # y1
+    pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * (pred_h - 1.0)
+    # x2
+    pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * (pred_w - 1.0)
+    # y2
+    pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
+
+    return pred_boxes
+
+
+def landmark_pred(boxes, landmark_deltas):
+    if boxes.shape[0] == 0:
+        return np.zeros((0, landmark_deltas.shape[1]))
+    boxes = boxes.astype(np.float, copy=False)
+    widths = boxes[:, 2] - boxes[:, 0] + 1.0
+    heights = boxes[:, 3] - boxes[:, 1] + 1.0
+    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+    preds = []
+    for i in range(landmark_deltas.shape[1]):
+        if i % 2 == 0:
+            pred = (landmark_deltas[:, i] * widths + ctr_x)
+        else:
+            pred = (landmark_deltas[:, i] * heights + ctr_y)
+        preds.append(pred)
+    preds = np.vstack(preds).transpose()
+    return preds
+
+
+def iou_transform(ex_rois, gt_rois):
+    """ return bbox targets, IoU loss uses gt_rois as gt """
+    assert ex_rois.shape[0] == gt_rois.shape[0], 'inconsistent rois number'
+    return gt_rois
+
+
+def iou_pred(boxes, box_deltas):
+    """
+    Transform the set of class-agnostic boxes into class-specific boxes
+    by applying the predicted offsets (box_deltas)
+    :param boxes: !important [N 4]
+    :param box_deltas: [N, 4 * num_classes]
+    :return: [N 4 * num_classes]
+    """
+    if boxes.shape[0] == 0:
+        return np.zeros((0, box_deltas.shape[1]))
+
+    boxes = boxes.astype(np.float, copy=False)
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+
+    dx1 = box_deltas[:, 0::4]
+    dy1 = box_deltas[:, 1::4]
+    dx2 = box_deltas[:, 2::4]
+    dy2 = box_deltas[:, 3::4]
+
+    pred_boxes = np.zeros(box_deltas.shape)
+    # x1
+    pred_boxes[:, 0::4] = dx1 + x1[:, np.newaxis]
+    # y1
+    pred_boxes[:, 1::4] = dy1 + y1[:, np.newaxis]
+    # x2
+    pred_boxes[:, 2::4] = dx2 + x2[:, np.newaxis]
+    # y2
+    pred_boxes[:, 3::4] = dy2 + y2[:, np.newaxis]
+
+    return pred_boxes
+
+
+# define bbox_transform and bbox_pred
+bbox_transform = nonlinear_transform
+bbox_pred = nonlinear_pred
diff --git a/RetinaFaceAntiCov/rcnn/processing/bbox_transform.py.orig b/detection/retinaface_anticov/rcnn/processing/bbox_transform.py.orig
similarity index 100%
rename from RetinaFaceAntiCov/rcnn/processing/bbox_transform.py.orig
rename to detection/retinaface_anticov/rcnn/processing/bbox_transform.py.orig
diff --git a/detection/retinaface_anticov/rcnn/processing/generate_anchor.py b/detection/retinaface_anticov/rcnn/processing/generate_anchor.py
new file mode 100644
index 000000000..83c5ada2a
--- /dev/null
+++ b/detection/retinaface_anticov/rcnn/processing/generate_anchor.py
@@ -0,0 +1,135 @@
+"""
+Generate base anchors on index 0
+"""
+from __future__ import print_function
+import sys
+from builtins import range
+import numpy as np
+from ..cython.anchors import anchors_cython
+#from ..config import config
+
+
+def anchors_plane(feat_h, feat_w, stride, base_anchor):
+    return anchors_cython(feat_h, feat_w, stride, base_anchor)
+
+
+def generate_anchors(base_size=16,
+                     ratios=[0.5, 1, 2],
+                     scales=2**np.arange(3, 6),
+                     stride=16,
+                     dense_anchor=False):
+    """
+    Generate anchor (reference) windows by enumerating aspect ratios X
+    scales wrt a reference (0, 0, 15, 15) window.
+    """
+
+    base_anchor = np.array([1, 1, base_size, base_size]) - 1
+    ratio_anchors = _ratio_enum(base_anchor, ratios)
+    anchors = np.vstack([
+        _scale_enum(ratio_anchors[i, :], scales)
+        for i in range(ratio_anchors.shape[0])
+    ])
+    if dense_anchor:
+        assert stride % 2 == 0
+        anchors2 = anchors.copy()
+        anchors2[:, :] += int(stride / 2)
+        anchors = np.vstack((anchors, anchors2))
+    #print('GA',base_anchor.shape, ratio_anchors.shape, anchors.shape)
+    return anchors
+
+
+#def generate_anchors_fpn(base_size=[64,32,16,8,4], ratios=[0.5, 1, 2], scales=8):
+#    """
+#    Generate anchor (reference) windows by enumerating aspect ratios X
+#    scales wrt a reference (0, 0, 15, 15) window.
+#    """
+#    anchors = []
+#    _ratios = ratios.reshape( (len(base_size), -1) )
+#    _scales = scales.reshape( (len(base_size), -1) )
+#    for i,bs in enumerate(base_size):
+#      __ratios = _ratios[i]
+#      __scales = _scales[i]
+#      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
+#      r = generate_anchors(bs, __ratios, __scales)
+#      #print('anchors_fpn', r.shape, file=sys.stderr)
+#      anchors.append(r)
+#    return anchors
+
+
+def generate_anchors_fpn(dense_anchor=False, cfg=None):
+    #assert(False)
+    """
+    Generate anchor (reference) windows by enumerating aspect ratios X
+    scales wrt a reference (0, 0, 15, 15) window.
+    """
+    if cfg is None:
+        from ..config import config
+        cfg = config.RPN_ANCHOR_CFG
+    RPN_FEAT_STRIDE = []
+    for k in cfg:
+        RPN_FEAT_STRIDE.append(int(k))
+    RPN_FEAT_STRIDE = sorted(RPN_FEAT_STRIDE, reverse=True)
+    anchors = []
+    for k in RPN_FEAT_STRIDE:
+        v = cfg[str(k)]
+        bs = v['BASE_SIZE']
+        __ratios = np.array(v['RATIOS'])
+        __scales = np.array(v['SCALES'])
+        stride = int(k)
+        #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
+        r = generate_anchors(bs, __ratios, __scales, stride, dense_anchor)
+        #print('anchors_fpn', r.shape, file=sys.stderr)
+        anchors.append(r)
+
+    return anchors
+
+
+def _whctrs(anchor):
+    """
+    Return width, height, x center, and y center for an anchor (window).
+    """
+
+    w = anchor[2] - anchor[0] + 1
+    h = anchor[3] - anchor[1] + 1
+    x_ctr = anchor[0] + 0.5 * (w - 1)
+    y_ctr = anchor[1] + 0.5 * (h - 1)
+    return w, h, x_ctr, y_ctr
+
+
+def _mkanchors(ws, hs, x_ctr, y_ctr):
+    """
+    Given a vector of widths (ws) and heights (hs) around a center
+    (x_ctr, y_ctr), output a set of anchors (windows).
+    """
+
+    ws = ws[:, np.newaxis]
+    hs = hs[:, np.newaxis]
+    anchors = np.hstack((x_ctr - 0.5 * (ws - 1), y_ctr - 0.5 * (hs - 1),
+                         x_ctr + 0.5 * (ws - 1), y_ctr + 0.5 * (hs - 1)))
+    return anchors
+
+
+def _ratio_enum(anchor, ratios):
+    """
+    Enumerate a set of anchors for each aspect ratio wrt an anchor.
+    """
+
+    w, h, x_ctr, y_ctr = _whctrs(anchor)
+    size = w * h
+    size_ratios = size / ratios
+    ws = np.round(np.sqrt(size_ratios))
+    hs = np.round(ws * ratios)
+    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
+    return anchors
+
+
+def _scale_enum(anchor, scales):
+    """
+    Enumerate a set of anchors for each scale wrt an anchor.
+    """
+
+    w, h, x_ctr, y_ctr = _whctrs(anchor)
+    ws = w * scales
+    hs = h * scales
+    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
+    return anchors
diff --git a/detection/retinaface_anticov/rcnn/processing/nms.py b/detection/retinaface_anticov/rcnn/processing/nms.py
new file mode 100644
index 000000000..b32d92d0f
--- /dev/null
+++ b/detection/retinaface_anticov/rcnn/processing/nms.py
@@ -0,0 +1,67 @@
+import numpy as np
+from ..cython.cpu_nms import cpu_nms
+try:
+    from ..cython.gpu_nms import gpu_nms
+except ImportError:
+    gpu_nms = None
+
+
+def py_nms_wrapper(thresh):
+    def _nms(dets):
+        return nms(dets, thresh)
+
+    return _nms
+
+
+def cpu_nms_wrapper(thresh):
+    def _nms(dets):
+        return cpu_nms(dets, thresh)
+
+    return _nms
+
+
+def gpu_nms_wrapper(thresh, device_id):
+    def _nms(dets):
+        return gpu_nms(dets, thresh, device_id)
+
+    if gpu_nms is not None:
+        return _nms
+    else:
+        return cpu_nms_wrapper(thresh)
+
+
+def nms(dets, thresh):
+    """
+    greedily select boxes with high confidence and overlap with current maximum <= thresh
+    rule out overlap >= thresh
+    :param dets: [[x1, y1, x2, y2 score]]
+    :param thresh: retain overlap < thresh
+    :return: indexes to keep
+    """
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    scores = dets[:, 4]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= thresh)[0]
+        order = order[inds + 1]
+
+    return keep
diff --git a/detection/retinaface_anticov/retinaface_cov.py b/detection/retinaface_anticov/retinaface_cov.py
new file mode 100644
index 000000000..a4d1c1ac4
--- /dev/null
+++ b/detection/retinaface_anticov/retinaface_cov.py
@@ -0,0 +1,752 @@
+from __future__ import print_function
+import sys
+import os
+import datetime
+import time
+import numpy as np
+import mxnet as mx
+from mxnet import ndarray as nd
+import cv2
+#from rcnn import config
+#from rcnn.processing.bbox_transform import nonlinear_pred, clip_boxes, landmark_pred
+from rcnn.processing.bbox_transform import clip_boxes
+from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
+from rcnn.processing.nms import gpu_nms_wrapper, cpu_nms_wrapper
+from rcnn.processing.bbox_transform import bbox_overlaps
+
+
+class RetinaFaceCoV:
+    def __init__(self,
+                 prefix,
+                 epoch,
+                 ctx_id=0,
+                 network='net3',
+                 nms=0.4,
+                 nocrop=False):
+        self.ctx_id = ctx_id
+        self.network = network
+        self.nms_threshold = nms
+        self.nocrop = nocrop
+        self.debug = False
+        self.fpn_keys = []
+        self.anchor_cfg = None
+        pixel_means = [0.0, 0.0, 0.0]
+        pixel_stds = [1.0, 1.0, 1.0]
+        pixel_scale = 1.0
+        self.bbox_stds = [1.0, 1.0, 1.0, 1.0]
+        self.landmark_std = 1.0
+        self.preprocess = False
+        _ratio = (1., )
+        fmc = 3
+        if network == 'ssh' or network == 'vgg':
+            pixel_means = [103.939, 116.779, 123.68]
+            self.preprocess = True
+        elif network == 'net3':
+            _ratio = (1., )
+        elif network == 'net3l':
+            _ratio = (1., )
+            self.landmark_std = 0.2
+        elif network == 'net3a':
+            _ratio = (1., 1.5)
+        elif network == 'net6':  #like pyramidbox or s3fd
+            fmc = 6
+        elif network == 'net5':  #retinaface
+            fmc = 5
+        elif network == 'net5a':
+            fmc = 5
+            _ratio = (1., 1.5)
+        elif network == 'net4':
+            fmc = 4
+        elif network == 'net4a':
+            fmc = 4
+            _ratio = (1., 1.5)
+        elif network == 'x5':
+            fmc = 5
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        elif network == 'x3':
+            fmc = 3
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        elif network == 'x3a':
+            fmc = 3
+            _ratio = (1., 1.5)
+            pixel_means = [103.52, 116.28, 123.675]
+            pixel_stds = [57.375, 57.12, 58.395]
+        else:
+            assert False, 'network setting error %s' % network
+
+        if fmc == 3:
+            self._feat_stride_fpn = [32, 16, 8]
+            self.anchor_cfg = {
+                '32': {
+                    'SCALES': (32, 16),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (8, 4),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 4:
+            self._feat_stride_fpn = [32, 16, 8, 4]
+            self.anchor_cfg = {
+                '32': {
+                    'SCALES': (32, 16),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (8, 4),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '4': {
+                    'SCALES': (2, 1),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 6:
+            self._feat_stride_fpn = [128, 64, 32, 16, 8, 4]
+            self.anchor_cfg = {
+                '128': {
+                    'SCALES': (32, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '64': {
+                    'SCALES': (16, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '32': {
+                    'SCALES': (8, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '16': {
+                    'SCALES': (4, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '8': {
+                    'SCALES': (2, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+                '4': {
+                    'SCALES': (1, ),
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                },
+            }
+        elif fmc == 5:
+            self._feat_stride_fpn = [64, 32, 16, 8, 4]
+            self.anchor_cfg = {}
+            _ass = 2.0**(1.0 / 3)
+            _basescale = 1.0
+            for _stride in [4, 8, 16, 32, 64]:
+                key = str(_stride)
+                value = {
+                    'BASE_SIZE': 16,
+                    'RATIOS': _ratio,
+                    'ALLOWED_BORDER': 9999
+                }
+                scales = []
+                for _ in range(3):
+                    scales.append(_basescale)
+                    _basescale *= _ass
+                value['SCALES'] = tuple(scales)
+                self.anchor_cfg[key] = value
+
+        #print(self._feat_stride_fpn, self.anchor_cfg)
+
+        for s in self._feat_stride_fpn:
+            self.fpn_keys.append('stride%s' % s)
+
+        dense_anchor = False
+        #self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(base_size=fpn_base_size, scales=self._scales, ratios=self._ratios)))
+        self._anchors_fpn = dict(
+            zip(
+                self.fpn_keys,
+                generate_anchors_fpn(dense_anchor=dense_anchor,
+                                     cfg=self.anchor_cfg)))
+        for k in self._anchors_fpn:
+            v = self._anchors_fpn[k].astype(np.float32)
+            self._anchors_fpn[k] = v
+
+        self._num_anchors = dict(
+            zip(self.fpn_keys,
+                [anchors.shape[0] for anchors in self._anchors_fpn.values()]))
+        #self._bbox_pred = nonlinear_pred
+        #self._landmark_pred = landmark_pred
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        if self.ctx_id >= 0:
+            self.ctx = mx.gpu(self.ctx_id)
+            self.nms = gpu_nms_wrapper(self.nms_threshold, self.ctx_id)
+        else:
+            self.ctx = mx.cpu()
+            self.nms = cpu_nms_wrapper(self.nms_threshold)
+        self.pixel_means = np.array(pixel_means, dtype=np.float32)
+        self.pixel_stds = np.array(pixel_stds, dtype=np.float32)
+        self.pixel_scale = float(pixel_scale)
+        #print('means', self.pixel_means)
+        self.use_landmarks = True
+        #print('use_landmarks', self.use_landmarks)
+        self.cascade = 0
+
+        if self.debug:
+            c = len(sym) // len(self._feat_stride_fpn)
+            sym = sym[(c * 0):]
+            self._feat_stride_fpn = [32, 16, 8]
+        #print('sym size:', len(sym))
+
+        image_size = (640, 640)
+        self.model = mx.mod.Module(symbol=sym,
+                                   context=self.ctx,
+                                   label_names=None)
+        self.model.bind(data_shapes=[('data', (1, 3, image_size[0],
+                                               image_size[1]))],
+                        for_training=False)
+        self.model.set_params(arg_params, aux_params)
+
+    def get_input(self, img):
+        im = img.astype(np.float32)
+        im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
+        for i in range(3):
+            im_tensor[
+                0,
+                i, :, :] = (im[:, :, 2 - i] / self.pixel_scale -
+                            self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
+        #if self.debug:
+        #  timeb = datetime.datetime.now()
+        #  diff = timeb - timea
+        #  print('X2 uses', diff.total_seconds(), 'seconds')
+        data = nd.array(im_tensor)
+        return data
+
+    def detect(self, img, threshold=0.5, scales=[1.0], do_flip=False):
+        #print('in_detect', threshold, scales, do_flip, do_nms)
+        proposals_list = []
+        scores_list = []
+        mask_scores_list = []
+        landmarks_list = []
+        strides_list = []
+        timea = datetime.datetime.now()
+        flips = [0]
+        if do_flip:
+            flips = [0, 1]
+
+        imgs = [img]
+        if isinstance(img, list):
+            imgs = img
+        for img in imgs:
+            for im_scale in scales:
+                for flip in flips:
+                    if im_scale != 1.0:
+                        im = cv2.resize(img,
+                                        None,
+                                        None,
+                                        fx=im_scale,
+                                        fy=im_scale,
+                                        interpolation=cv2.INTER_LINEAR)
+                    else:
+                        im = img.copy()
+                    if flip:
+                        im = im[:, ::-1, :]
+                    if self.nocrop:
+                        if im.shape[0] % 32 == 0:
+                            h = im.shape[0]
+                        else:
+                            h = (im.shape[0] // 32 + 1) * 32
+                        if im.shape[1] % 32 == 0:
+                            w = im.shape[1]
+                        else:
+                            w = (im.shape[1] // 32 + 1) * 32
+                        _im = np.zeros((h, w, 3), dtype=np.float32)
+                        _im[0:im.shape[0], 0:im.shape[1], :] = im
+                        im = _im
+                    else:
+                        im = im.astype(np.float32)
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X1 uses', diff.total_seconds(), 'seconds')
+                    #self.model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))], for_training=False)
+                    #im_info = [im.shape[0], im.shape[1], im_scale]
+                    im_info = [im.shape[0], im.shape[1]]
+                    im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
+                    for i in range(3):
+                        im_tensor[0, i, :, :] = (
+                            im[:, :, 2 - i] / self.pixel_scale -
+                            self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X2 uses', diff.total_seconds(), 'seconds')
+                    data = nd.array(im_tensor)
+                    db = mx.io.DataBatch(data=(data, ),
+                                         provide_data=[('data', data.shape)])
+                    if self.debug:
+                        timeb = datetime.datetime.now()
+                        diff = timeb - timea
+                        print('X3 uses', diff.total_seconds(), 'seconds')
+                    self.model.forward(db, is_train=False)
+                    net_out = self.model.get_outputs()
+                    #post_nms_topN = self._rpn_post_nms_top_n
+                    #min_size_dict = self._rpn_min_size_fpn
+
+                    sym_idx = 0
+
+                    for _idx, s in enumerate(self._feat_stride_fpn):
+                        #if len(scales)>1 and s==32 and im_scale==scales[-1]:
+                        #  continue
+                        _key = 'stride%s' % s
+                        stride = int(s)
+                        is_cascade = False
+                        #if self.vote and stride==4 and len(scales)>2 and (im_scale==scales[0]):
+                        #  continue
+                        #print('getting', im_scale, stride, idx, len(net_out), data.shape, file=sys.stderr)
+                        scores = net_out[sym_idx].asnumpy()
+                        type_scores = net_out[sym_idx + 3].asnumpy()
+                        print(scores.shape, type_scores.shape)
+                        if self.debug:
+                            timeb = datetime.datetime.now()
+                            diff = timeb - timea
+                            print('A uses', diff.total_seconds(), 'seconds')
+                        A = self._num_anchors['stride%s' % s]
+                        #print(scores.shape)
+                        #print('scores',stride, scores.shape, file=sys.stderr)
+                        scores = scores[:, A:, :, :]
+                        mask_scores = type_scores[:, A * 2:, :, :]  #x, A, x, x
+
+                        bbox_deltas = net_out[sym_idx + 1].asnumpy()
+
+                        #if DEBUG:
+                        #    print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
+                        #    print 'scale: {}'.format(im_info[2])
+
+                        #_height, _width = int(im_info[0] / stride), int(im_info[1] / stride)
+                        height, width = bbox_deltas.shape[
+                            2], bbox_deltas.shape[3]
+
+                        K = height * width
+                        anchors_fpn = self._anchors_fpn['stride%s' % s]
+                        anchors = anchors_plane(height, width, stride,
+                                                anchors_fpn)
+                        #print((height, width), (_height, _width), anchors.shape, bbox_deltas.shape, scores.shape, file=sys.stderr)
+                        anchors = anchors.reshape((K * A, 4))
+                        #print('num_anchors', self._num_anchors['stride%s'%s], file=sys.stderr)
+                        #print('HW', (height, width), file=sys.stderr)
+                        #print('anchors_fpn', anchors_fpn.shape, file=sys.stderr)
+                        #print('anchors', anchors.shape, file=sys.stderr)
+                        #print('bbox_deltas', bbox_deltas.shape, file=sys.stderr)
+                        #print('scores', scores.shape, file=sys.stderr)
+
+                        #scores = self._clip_pad(scores, (height, width))
+                        scores = scores.transpose((0, 2, 3, 1)).reshape(
+                            (-1, 1))
+                        mask_scores = mask_scores.transpose(
+                            (0, 2, 3, 1)).reshape((-1, 1))
+
+                        #print('pre', bbox_deltas.shape, height, width)
+                        #bbox_deltas = self._clip_pad(bbox_deltas, (height, width))
+                        #print('after', bbox_deltas.shape, height, width)
+                        bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
+                        bbox_pred_len = bbox_deltas.shape[3] // A
+                        #print(bbox_deltas.shape)
+                        bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
+                        bbox_deltas[:,
+                                    0::4] = bbox_deltas[:, 0::
+                                                        4] * self.bbox_stds[0]
+                        bbox_deltas[:,
+                                    1::4] = bbox_deltas[:, 1::
+                                                        4] * self.bbox_stds[1]
+                        bbox_deltas[:,
+                                    2::4] = bbox_deltas[:, 2::
+                                                        4] * self.bbox_stds[2]
+                        bbox_deltas[:,
+                                    3::4] = bbox_deltas[:, 3::
+                                                        4] * self.bbox_stds[3]
+                        proposals = self.bbox_pred(anchors, bbox_deltas)
+
+                        proposals = clip_boxes(proposals, im_info[:2])
+
+                        #if self.vote:
+                        #  if im_scale>1.0:
+                        #    keep = self._filter_boxes2(proposals, 160*im_scale, -1)
+                        #  else:
+                        #    keep = self._filter_boxes2(proposals, -1, 100*im_scale)
+                        #  if stride==4:
+                        #    keep = self._filter_boxes2(proposals, 12*im_scale, -1)
+                        #    proposals = proposals[keep, :]
+                        #    scores = scores[keep]
+
+                        #keep = self._filter_boxes(proposals, min_size_dict['stride%s'%s] * im_info[2])
+                        #proposals = proposals[keep, :]
+                        #scores = scores[keep]
+                        #print('333', proposals.shape)
+                        if stride == 4 and self.decay4 < 1.0:
+                            scores *= self.decay4
+
+                        scores_ravel = scores.ravel()
+                        #mask_scores_ravel = mask_scores.ravel()
+                        #print('__shapes', proposals.shape, scores_ravel.shape)
+                        #print('max score', np.max(scores_ravel))
+                        order = np.where(scores_ravel >= threshold)[0]
+                        #_scores = scores_ravel[order]
+                        #_order = _scores.argsort()[::-1]
+                        #order = order[_order]
+                        proposals = proposals[order, :]
+                        scores = scores[order]
+                        mask_scores = mask_scores[order]
+                        if flip:
+                            oldx1 = proposals[:, 0].copy()
+                            oldx2 = proposals[:, 2].copy()
+                            proposals[:, 0] = im.shape[1] - oldx2 - 1
+                            proposals[:, 2] = im.shape[1] - oldx1 - 1
+
+                        proposals[:, 0:4] /= im_scale
+
+                        proposals_list.append(proposals)
+                        scores_list.append(scores)
+                        mask_scores_list.append(mask_scores)
+
+                        landmark_deltas = net_out[sym_idx + 2].asnumpy()
+                        #landmark_deltas = self._clip_pad(landmark_deltas, (height, width))
+                        landmark_pred_len = landmark_deltas.shape[1] // A
+                        landmark_deltas = landmark_deltas.transpose(
+                            (0, 2, 3, 1)).reshape(
+                                (-1, 5, landmark_pred_len // 5))
+                        landmark_deltas *= self.landmark_std
+                        #print(landmark_deltas.shape, landmark_deltas)
+                        landmarks = self.landmark_pred(anchors,
+                                                       landmark_deltas)
+                        landmarks = landmarks[order, :]
+
+                        if flip:
+                            landmarks[:, :,
+                                      0] = im.shape[1] - landmarks[:, :, 0] - 1
+                            #for a in range(5):
+                            #  oldx1 = landmarks[:, a].copy()
+                            #  landmarks[:,a] = im.shape[1] - oldx1 - 1
+                            order = [1, 0, 2, 4, 3]
+                            flandmarks = landmarks.copy()
+                            for idx, a in enumerate(order):
+                                flandmarks[:, idx, :] = landmarks[:, a, :]
+                                #flandmarks[:, idx*2] = landmarks[:,a*2]
+                                #flandmarks[:, idx*2+1] = landmarks[:,a*2+1]
+                            landmarks = flandmarks
+                        landmarks[:, :, 0:2] /= im_scale
+                        #landmarks /= im_scale
+                        #landmarks = landmarks.reshape( (-1, landmark_pred_len) )
+                        landmarks_list.append(landmarks)
+                        #proposals = np.hstack((proposals, landmarks))
+                        sym_idx += 4
+
+        if self.debug:
+            timeb = datetime.datetime.now()
+            diff = timeb - timea
+            print('B uses', diff.total_seconds(), 'seconds')
+        proposals = np.vstack(proposals_list)
+        landmarks = None
+        if proposals.shape[0] == 0:
+            landmarks = np.zeros((0, 5, 2))
+            return np.zeros((0, 6)), landmarks
+        scores = np.vstack(scores_list)
+        mask_scores = np.vstack(mask_scores_list)
+        #print('shapes', proposals.shape, scores.shape)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        #if config.TEST.SCORE_THRESH>0.0:
+        #  _count = np.sum(scores_ravel>config.TEST.SCORE_THRESH)
+        #  order = order[:_count]
+        proposals = proposals[order, :]
+        scores = scores[order]
+        mask_scores = mask_scores[order]
+        landmarks = np.vstack(landmarks_list)
+        landmarks = landmarks[order].astype(np.float32, copy=False)
+
+        pre_det = np.hstack((proposals[:, 0:4], scores)).astype(np.float32,
+                                                                copy=False)
+        keep = self.nms(pre_det)
+        det = np.hstack((pre_det, mask_scores))
+        det = det[keep, :]
+        landmarks = landmarks[keep]
+
+        if self.debug:
+            timeb = datetime.datetime.now()
+            diff = timeb - timea
+            print('C uses', diff.total_seconds(), 'seconds')
+        return det, landmarks
+
+    def detect_center(self, img, threshold=0.5, scales=[1.0], do_flip=False):
+        det, landmarks = self.detect(img, threshold, scales, do_flip)
+        if det.shape[0] == 0:
+            return None, None
+        bindex = 0
+        if det.shape[0] > 1:
+            img_size = np.asarray(img.shape)[0:2]
+            bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                           det[:, 1])
+            img_center = img_size / 2
+            offsets = np.vstack([(det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                                 (det[:, 1] + det[:, 3]) / 2 - img_center[0]])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            bindex = np.argmax(bounding_box_size - offset_dist_squared *
+                               2.0)  # some extra weight on the centering
+        bbox = det[bindex, :]
+        landmark = landmarks[bindex, :, :]
+        return bbox, landmark
+
+    @staticmethod
+    def check_large_pose(landmark, bbox):
+        assert landmark.shape == (5, 2)
+        assert len(bbox) == 4
+
+        def get_theta(base, x, y):
+            vx = x - base
+            vy = y - base
+            vx[1] *= -1
+            vy[1] *= -1
+            tx = np.arctan2(vx[1], vx[0])
+            ty = np.arctan2(vy[1], vy[0])
+            d = ty - tx
+            d = np.degrees(d)
+            #print(vx, tx, vy, ty, d)
+            #if d<-1.*math.pi:
+            #  d+=2*math.pi
+            #elif d>math.pi:
+            #  d-=2*math.pi
+            if d < -180.0:
+                d += 360.
+            elif d > 180.0:
+                d -= 360.0
+            return d
+
+        landmark = landmark.astype(np.float32)
+
+        theta1 = get_theta(landmark[0], landmark[3], landmark[2])
+        theta2 = get_theta(landmark[1], landmark[2], landmark[4])
+        #print(va, vb, theta2)
+        theta3 = get_theta(landmark[0], landmark[2], landmark[1])
+        theta4 = get_theta(landmark[1], landmark[0], landmark[2])
+        theta5 = get_theta(landmark[3], landmark[4], landmark[2])
+        theta6 = get_theta(landmark[4], landmark[2], landmark[3])
+        theta7 = get_theta(landmark[3], landmark[2], landmark[0])
+        theta8 = get_theta(landmark[4], landmark[1], landmark[2])
+        #print(theta1, theta2, theta3, theta4, theta5, theta6, theta7, theta8)
+        left_score = 0.0
+        right_score = 0.0
+        up_score = 0.0
+        down_score = 0.0
+        if theta1 <= 0.0:
+            left_score = 10.0
+        elif theta2 <= 0.0:
+            right_score = 10.0
+        else:
+            left_score = theta2 / theta1
+            right_score = theta1 / theta2
+        if theta3 <= 10.0 or theta4 <= 10.0:
+            up_score = 10.0
+        else:
+            up_score = max(theta1 / theta3, theta2 / theta4)
+        if theta5 <= 10.0 or theta6 <= 10.0:
+            down_score = 10.0
+        else:
+            down_score = max(theta7 / theta5, theta8 / theta6)
+        mleft = (landmark[0][0] + landmark[3][0]) / 2
+        mright = (landmark[1][0] + landmark[4][0]) / 2
+        box_center = ((bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2)
+        ret = 0
+        if left_score >= 3.0:
+            ret = 1
+        if ret == 0 and left_score >= 2.0:
+            if mright <= box_center[0]:
+                ret = 1
+        if ret == 0 and right_score >= 3.0:
+            ret = 2
+        if ret == 0 and right_score >= 2.0:
+            if mleft >= box_center[0]:
+                ret = 2
+        if ret == 0 and up_score >= 2.0:
+            ret = 3
+        if ret == 0 and down_score >= 5.0:
+            ret = 4
+        return ret, left_score, right_score, up_score, down_score
+
+    @staticmethod
+    def _filter_boxes(boxes, min_size):
+        """ Remove all boxes with any side smaller than min_size """
+        ws = boxes[:, 2] - boxes[:, 0] + 1
+        hs = boxes[:, 3] - boxes[:, 1] + 1
+        keep = np.where((ws >= min_size) & (hs >= min_size))[0]
+        return keep
+
+    @staticmethod
+    def _filter_boxes2(boxes, max_size, min_size):
+        """ Remove all boxes with any side smaller than min_size """
+        ws = boxes[:, 2] - boxes[:, 0] + 1
+        hs = boxes[:, 3] - boxes[:, 1] + 1
+        if max_size > 0:
+            keep = np.where(np.minimum(ws, hs) < max_size)[0]
+        elif min_size > 0:
+            keep = np.where(np.maximum(ws, hs) > min_size)[0]
+        return keep
+
+    @staticmethod
+    def _clip_pad(tensor, pad_shape):
+        """
+      Clip boxes of the pad area.
+      :param tensor: [n, c, H, W]
+      :param pad_shape: [h, w]
+      :return: [n, c, h, w]
+      """
+        H, W = tensor.shape[2:]
+        h, w = pad_shape
+
+        if h < H or w < W:
+            tensor = tensor[:, :, :h, :w].copy()
+
+        return tensor
+
+    @staticmethod
+    def bbox_pred(boxes, box_deltas):
+        """
+      Transform the set of class-agnostic boxes into class-specific boxes
+      by applying the predicted offsets (box_deltas)
+      :param boxes: !important [N 4]
+      :param box_deltas: [N, 4 * num_classes]
+      :return: [N 4 * num_classes]
+      """
+        if boxes.shape[0] == 0:
+            return np.zeros((0, box_deltas.shape[1]))
+
+        boxes = boxes.astype(np.float, copy=False)
+        widths = boxes[:, 2] - boxes[:, 0] + 1.0
+        heights = boxes[:, 3] - boxes[:, 1] + 1.0
+        ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+        ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+
+        dx = box_deltas[:, 0:1]
+        dy = box_deltas[:, 1:2]
+        dw = box_deltas[:, 2:3]
+        dh = box_deltas[:, 3:4]
+
+        pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
+        pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
+        pred_w = np.exp(dw) * widths[:, np.newaxis]
+        pred_h = np.exp(dh) * heights[:, np.newaxis]
+
+        pred_boxes = np.zeros(box_deltas.shape)
+        # x1
+        pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
+        # y1
+        pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
+        # x2
+        pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
+        # y2
+        pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
+
+        if box_deltas.shape[1] > 4:
+            pred_boxes[:, 4:] = box_deltas[:, 4:]
+
+        return pred_boxes
+
+    @staticmethod
+    def landmark_pred(boxes, landmark_deltas):
+        if boxes.shape[0] == 0:
+            return np.zeros((0, landmark_deltas.shape[1]))
+        boxes = boxes.astype(np.float, copy=False)
+        widths = boxes[:, 2] - boxes[:, 0] + 1.0
+        heights = boxes[:, 3] - boxes[:, 1] + 1.0
+        ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
+        ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
+        pred = landmark_deltas.copy()
+        for i in range(5):
+            pred[:, i, 0] = landmark_deltas[:, i, 0] * widths + ctr_x
+            pred[:, i, 1] = landmark_deltas[:, i, 1] * heights + ctr_y
+        return pred
+        #preds = []
+        #for i in range(landmark_deltas.shape[1]):
+        #  if i%2==0:
+        #    pred = (landmark_deltas[:,i]*widths + ctr_x)
+        #  else:
+        #    pred = (landmark_deltas[:,i]*heights + ctr_y)
+        #  preds.append(pred)
+        #preds = np.vstack(preds).transpose()
+        #return preds
+
+    def vote(self, det):
+        #order = det[:, 4].ravel().argsort()[::-1]
+        #det = det[order, :]
+        if det.shape[0] == 0:
+            return np.zeros((0, 5))
+            #dets = np.array([[10, 10, 20, 20, 0.002]])
+            #det = np.empty(shape=[0, 5])
+        dets = None
+        while det.shape[0] > 0:
+            if dets is not None and dets.shape[0] >= 750:
+                break
+            # IOU
+            area = (det[:, 2] - det[:, 0] + 1) * (det[:, 3] - det[:, 1] + 1)
+            xx1 = np.maximum(det[0, 0], det[:, 0])
+            yy1 = np.maximum(det[0, 1], det[:, 1])
+            xx2 = np.minimum(det[0, 2], det[:, 2])
+            yy2 = np.minimum(det[0, 3], det[:, 3])
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            o = inter / (area[0] + area[:] - inter)
+
+            # nms
+            merge_index = np.where(o >= self.nms_threshold)[0]
+            det_accu = det[merge_index, :]
+            det = np.delete(det, merge_index, 0)
+            if merge_index.shape[0] <= 1:
+                if det.shape[0] == 0:
+                    try:
+                        dets = np.row_stack((dets, det_accu))
+                    except:
+                        dets = det_accu
+                continue
+            det_accu[:,
+                     0:4] = det_accu[:, 0:4] * np.tile(det_accu[:, -1:],
+                                                       (1, 4))
+            max_score = np.max(det_accu[:, 4])
+            det_accu_sum = np.zeros((1, 5))
+            det_accu_sum[:, 0:4] = np.sum(det_accu[:, 0:4], axis=0) / np.sum(
+                det_accu[:, -1:])
+            det_accu_sum[:, 4] = max_score
+            if dets is None:
+                dets = det_accu_sum
+            else:
+                dets = np.row_stack((dets, det_accu_sum))
+        dets = dets[0:750, :]
+        return dets
diff --git a/detection/retinaface_anticov/test.py b/detection/retinaface_anticov/test.py
new file mode 100644
index 000000000..f7c282b6d
--- /dev/null
+++ b/detection/retinaface_anticov/test.py
@@ -0,0 +1,66 @@
+import cv2
+import sys
+import numpy as np
+import datetime
+import os
+import glob
+from retinaface_cov import RetinaFaceCoV
+
+thresh = 0.8
+mask_thresh = 0.2
+scales = [640, 1080]
+
+count = 1
+
+gpuid = 0
+#detector = RetinaFaceCoV('./model/mnet_cov1', 0, gpuid, 'net3')
+detector = RetinaFaceCoV('./model/mnet_cov2', 0, gpuid, 'net3l')
+
+img = cv2.imread('n1.jpg')
+print(img.shape)
+im_shape = img.shape
+target_size = scales[0]
+max_size = scales[1]
+im_size_min = np.min(im_shape[0:2])
+im_size_max = np.max(im_shape[0:2])
+#im_scale = 1.0
+#if im_size_min>target_size or im_size_max>max_size:
+im_scale = float(target_size) / float(im_size_min)
+# prevent bigger axis from being more than max_size:
+if np.round(im_scale * im_size_max) > max_size:
+    im_scale = float(max_size) / float(im_size_max)
+
+print('im_scale', im_scale)
+
+scales = [im_scale]
+flip = False
+
+for c in range(count):
+    faces, landmarks = detector.detect(img,
+                                       thresh,
+                                       scales=scales,
+                                       do_flip=flip)
+
+if faces is not None:
+    print('find', faces.shape[0], 'faces')
+    for i in range(faces.shape[0]):
+        #print('score', faces[i][4])
+        face = faces[i]
+        box = face[0:4].astype(np.int)
+        mask = face[5]
+        print(i, box, mask)
+        #color = (255,0,0)
+        if mask >= mask_thresh:
+            color = (0, 0, 255)
+        else:
+            color = (0, 255, 0)
+        cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), color, 2)
+        landmark5 = landmarks[i].astype(np.int)
+        #print(landmark.shape)
+        for l in range(landmark5.shape[0]):
+            color = (255, 0, 0)
+            cv2.circle(img, (landmark5[l][0], landmark5[l][1]), 1, color, 2)
+
+    filename = './cov_test.jpg'
+    print('writing', filename)
+    cv2.imwrite(filename, img)
diff --git a/detection/scrfd/LICENSE b/detection/scrfd/LICENSE
new file mode 100755
index 000000000..04adf5cbc
--- /dev/null
+++ b/detection/scrfd/LICENSE
@@ -0,0 +1,203 @@
+Copyright 2018-2019 Open-MMLab. All rights reserved.
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2018-2019 Open-MMLab.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/detection/scrfd/README.md b/detection/scrfd/README.md
new file mode 100755
index 000000000..43c26f8fe
--- /dev/null
+++ b/detection/scrfd/README.md
@@ -0,0 +1,183 @@
+## Introduction
+
+SCRFD is an efficient high accuracy face detection approach which initially described in [Arxiv](https://arxiv.org/abs/2105.04714), and accepted by ICLR-2022.
+
+Try out the Gradio Web Demo: [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/hysts/insightface-SCRFD)
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/scrfd_evelope.jpg" width="400" alt="prcurve"/>
+
+## Performance
+
+Precision, flops and infer time are all evaluated on **VGA resolution**.
+
+#### ResNet family
+
+| Method              | Backbone        | Easy  | Medium | Hard  | \#Params(M) | \#Flops(G) | Infer(ms) |
+| ------------------- | --------------- | ----- | ------ | ----- | ----------- | ---------- | --------- |
+| DSFD (CVPR19)       | ResNet152       | 94.29 | 91.47  | 71.39 | 120.06      | 259.55     | 55.6      |
+| RetinaFace (CVPR20) | ResNet50        | 94.92 | 91.90  | 64.17 | 29.50       | 37.59      | 21.7      |
+| HAMBox (CVPR20)     | ResNet50        | 95.27 | 93.76  | 76.75 | 30.24       | 43.28      | 25.9      |
+| TinaFace (Arxiv20)  | ResNet50        | 95.61 | 94.25  | 81.43 | 37.98       | 172.95     | 38.9      |
+| - | - | - | - | - | - | - | - |
+| ResNet-34GF         | ResNet50        | 95.64 | 94.22  | 84.02 | 24.81       | 34.16      | 11.8      |
+| **SCRFD-34GF**      | Bottleneck Res  | 96.06 | 94.92  | 85.29 | 9.80        | 34.13      | 11.7      |
+| ResNet-10GF         | ResNet34x0.5    | 94.69 | 92.90  | 80.42 | 6.85        | 10.18      | 6.3       |
+| **SCRFD-10GF**      | Basic Res       | 95.16 | 93.87  | 83.05 | 3.86        | 9.98       | 4.9       |
+| ResNet-2.5GF        | ResNet34x0.25   | 93.21 | 91.11  | 74.47 | 1.62        | 2.57       | 5.4       |
+| **SCRFD-2.5GF**     | Basic Res       | 93.78 | 92.16  | 77.87 | 0.67        | 2.53       | 4.2       |
+
+
+#### Mobile family
+
+| Method              | Backbone        | Easy  | Medium | Hard  | \#Params(M) | \#Flops(G) | Infer(ms) |
+| ------------------- | --------------- | ----- | ------ | ----- | ----------- | ---------- | --------- |
+| RetinaFace (CVPR20) | MobileNet0.25   | 87.78 | 81.16  | 47.32 | 0.44        | 0.802      | 7.9       |
+| FaceBoxes (IJCB17)  | -               | 76.17 | 57.17  | 24.18 | 1.01        | 0.275      | 2.5       |
+| - | - | - | - | - | - | - | - |
+| MobileNet-0.5GF     | MobileNetx0.25  | 90.38 | 87.05  | 66.68 | 0.37        | 0.507      | 3.7       |
+| **SCRFD-0.5GF**     | Depth-wise Conv | 90.57 | 88.12  | 68.51 | 0.57        | 0.508      | 3.6       |
+
+
+**X64 CPU Performance of SCRFD-0.5GF:**
+
+| Test-Input-Size         | CPU Single-Thread   | Easy  | Medium | Hard  |
+| ----------------------- | -----------------   | ----- | ------ | ----- |
+| Original-Size(scale1.0) | -                   | 90.91 | 89.49  | 82.03 |
+| 640x480                 | 28.3ms              | 90.57 | 88.12  | 68.51 |
+| 320x240                 | 11.4ms              | -     | -      | -     |
+
+*precision and infer time are evaluated on AMD Ryzen 9 3950X, using the simple PyTorch CPU inference by setting `OMP_NUM_THREADS=1` (no mkldnn).*
+
+## Installation
+
+Please refer to [mmdetection](https://github.com/open-mmlab/mmdetection/blob/master/docs/en/get_started.md#installation) for installation.
+ 
+  1. Install [mmcv](https://github.com/open-mmlab/mmcv). (mmcv-full==1.2.6 and 1.3.3 was tested)
+  2. Install build requirements and then install mmdet.
+       ```
+       pip install -r requirements/build.txt
+       pip install -v -e .  # or "python setup.py develop"
+       ```
+
+## Data preparation
+
+### WIDERFace:
+  1. Download WIDERFace datasets and put it under `data/retinaface`.
+  2. Download annotation files from [gdrive](https://drive.google.com/file/d/1UW3KoApOhusyqSHX96yEDRYiNkd3Iv3Z/view?usp=sharing) and put them under `data/retinaface/`
+ 
+   ```
+     data/retinaface/
+         train/
+             images/
+             labelv2.txt
+         val/
+             images/
+             labelv2.txt
+             gt/
+                 *.mat
+             
+   ```
+ 
+
+#### Annotation Format 
+
+*please refer to labelv2.txt for detail*
+
+For each image:
+  ```
+  # <image_path> image_width image_height
+  bbox_x1 bbox_y1 bbox_x2 bbox_y2 (<keypoint,3>*N)
+  ...
+  ...
+  # <image_path> image_width image_height
+  bbox_x1 bbox_y1 bbox_x2 bbox_y2 (<keypoint,3>*N)
+  ...
+  ...
+  ```
+Keypoints can be ignored if there is bbox annotation only.
+
+
+## Training
+
+Example training command, with 4 GPUs:
+```
+CUDA_VISIBLE_DEVICES="0,1,2,3" PORT=29701 bash ./tools/dist_train.sh ./configs/scrfd/scrfd_1g.py 4
+```
+
+## WIDERFace Evaluation
+
+We use a pure python evaluation script without Matlab.
+
+```
+GPU=0
+GROUP=scrfd
+TASK=scrfd_2.5g
+CUDA_VISIBLE_DEVICES="$GPU" python -u tools/test_widerface.py ./configs/"$GROUP"/"$TASK".py ./work_dirs/"$TASK"/model.pth --mode 0 --out wouts
+```
+
+
+## Pretrained-Models
+
+|      Name      | Easy  | Medium | Hard  | FLOPs | Params(M) | Infer(ms) | Link                                                         |
+| :------------: | ----- | ------ | ----- | ----- | --------- | --------- | ------------------------------------------------------------ |
+|   SCRFD_500M   | 90.57 | 88.12  | 68.51 | 500M  | 0.57      | 3.6       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyYWxScdiTITY4TQ?e=DjXof9) |
+|    SCRFD_1G    | 92.38 | 90.57  | 74.80 | 1G    | 0.64      | 4.1       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyPVLI44ahNBsOMR?e=esPrBL) |
+|   SCRFD_2.5G   | 93.78 | 92.16  | 77.87 | 2.5G  | 0.67      | 4.2       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyTIXnzB1ujPq4th?e=5t1VNv) |
+|   SCRFD_10G    | 95.16 | 93.87  | 83.05 | 10G   | 3.86      | 4.9       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyUKwTiwXv2kaa8o?e=umfepO) |
+|   SCRFD_34G    | 96.06 | 94.92  | 85.29 | 34G   | 9.80      | 11.7      | [download](https://1drv.ms/u/s!AswpsDO2toNKqyKZwFebVlmlOvzz?e=V2rqUy) |
+| SCRFD_500M_KPS | 90.97 | 88.44  | 69.49 | 500M  | 0.57      | 3.6      | [download](https://1drv.ms/u/s!AswpsDO2toNKri_NDM0GIkPpkE2f?e=JkebJo) |
+| SCRFD_2.5G_KPS | 93.80 | 92.02  | 77.13 | 2.5G  | 0.82      | 4.3       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyGlhxnCg3smyQqX?e=A6Hufm) |
+| SCRFD_10G_KPS  | 95.40 | 94.01  | 82.80 | 10G   | 4.23      | 5.0       | [download](https://1drv.ms/u/s!AswpsDO2toNKqycsF19UbaCWaLWx?e=F6i5Vm) |
+
+mAP, FLOPs and inference latency are all evaluated on VGA resolution.
+``_KPS`` means the model includes 5 keypoints prediction.
+
+## Convert to ONNX
+
+Please refer to `tools/scrfd2onnx.py`
+
+Generated onnx model can accept dynamic input as default.
+
+You can also set specific input shape by pass ``--shape 640 640``, then output onnx model can be optimized by onnx-simplifier.
+
+
+## Inference
+
+Please refer to `tools/scrfd.py` which uses onnxruntime to do inference.
+
+## Network Search
+
+For two-steps search as we described in paper, we target hard mAP on how we select best candidate models.
+
+We provide an example for searching SCRFD-2.5GF in this repo as below.
+
+1. For searching backbones: 
+
+    ```
+    python search_tools/generate_configs_2.5g.py --mode 1
+    ```
+   Where ``mode==1`` means searching backbone only. For other parameters, please check the code.
+2. After step-1 done, there will be ``configs/scrfdgen2.5g/scrfdgen2.5g_1.py`` to ``configs/scrfdgen2.5g/scrfdgen2.5g_64.py`` if ``num_configs`` is set to 64.
+3. Do training for every generated configs for 80 epochs, please check ``search_tools/search_train.sh``
+4. Test WIDERFace precision for every generated configs, using ``search_tools/search_test.sh``.
+5. Select the top accurate config as the base template(assume the 10-th config is the best), then do the overall network search. 
+    ```
+    python search_tools/generate_configs_2.5g.py --mode 2 --template 10
+    ```
+6. Test these new generated configs again and select the top accurate one(s).
+
+
+## Acknowledgments
+
+We thank [nihui](https://github.com/nihui) for the excellent [mobile-phone demo](https://github.com/nihui/ncnn-android-scrfd).
+
+## Demo
+
+1. [ncnn-android-scrfd](https://github.com/nihui/ncnn-android-scrfd)
+2. [scrfd-MNN C++](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/mnn/cv/mnn_scrfd.cpp)
+3. [scrfd-TNN C++](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/tnn/cv/tnn_scrfd.cpp)
+4. [scrfd-NCNN C++](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/ncnn/cv/ncnn_scrfd.cpp)
+5. [scrfd-ONNXRuntime C++](https://github.com/DefTruth/lite.ai.toolkit/blob/main/lite/ort/cv/scrfd.cpp)
+6. [TensorRT Python](https://github.com/SthPhoenix/InsightFace-REST/blob/master/src/api_trt/modules/model_zoo/detectors/scrfd.py)
+7. [Modelscope demo for rotated face](https://modelscope.cn/models/damo/cv_resnet_facedetection_scrfd10gkps/summary)
+8. [Modelscope demo for card detection](https://modelscope.cn/models/damo/cv_resnet_carddetection_scrfd34gkps/summary)
diff --git a/detection/scrfd/configs/_base_/datasets/cityscapes_detection.py b/detection/scrfd/configs/_base_/datasets/cityscapes_detection.py
new file mode 100755
index 000000000..156aca025
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/cityscapes_detection.py
@@ -0,0 +1,55 @@
+dataset_type = 'CityscapesDataset'
+data_root = 'data/cityscapes/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='Resize', img_scale=[(2048, 800), (2048, 1024)], keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 1024),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=1,
+    workers_per_gpu=2,
+    train=dict(
+        type='RepeatDataset',
+        times=8,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=data_root +
+            'annotations/instancesonly_filtered_gtFine_train.json',
+            img_prefix=data_root + 'leftImg8bit/train/',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root +
+        'annotations/instancesonly_filtered_gtFine_val.json',
+        img_prefix=data_root + 'leftImg8bit/val/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root +
+        'annotations/instancesonly_filtered_gtFine_test.json',
+        img_prefix=data_root + 'leftImg8bit/test/',
+        pipeline=test_pipeline))
+evaluation = dict(interval=1, metric='bbox')
diff --git a/detection/scrfd/configs/_base_/datasets/cityscapes_instance.py b/detection/scrfd/configs/_base_/datasets/cityscapes_instance.py
new file mode 100755
index 000000000..3c5472aab
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/cityscapes_instance.py
@@ -0,0 +1,55 @@
+dataset_type = 'CityscapesDataset'
+data_root = 'data/cityscapes/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
+    dict(
+        type='Resize', img_scale=[(2048, 800), (2048, 1024)], keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 1024),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=1,
+    workers_per_gpu=2,
+    train=dict(
+        type='RepeatDataset',
+        times=8,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=data_root +
+            'annotations/instancesonly_filtered_gtFine_train.json',
+            img_prefix=data_root + 'leftImg8bit/train/',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root +
+        'annotations/instancesonly_filtered_gtFine_val.json',
+        img_prefix=data_root + 'leftImg8bit/val/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root +
+        'annotations/instancesonly_filtered_gtFine_test.json',
+        img_prefix=data_root + 'leftImg8bit/test/',
+        pipeline=test_pipeline))
+evaluation = dict(metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/coco_detection.py b/detection/scrfd/configs/_base_/datasets/coco_detection.py
new file mode 100755
index 000000000..09a75c404
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/coco_detection.py
@@ -0,0 +1,48 @@
+dataset_type = 'CocoDataset'
+data_root = 'data/coco/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_train2017.json',
+        img_prefix=data_root + 'train2017/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline))
+evaluation = dict(interval=1, metric='bbox')
diff --git a/detection/scrfd/configs/_base_/datasets/coco_instance.py b/detection/scrfd/configs/_base_/datasets/coco_instance.py
new file mode 100755
index 000000000..f6ea4f456
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/coco_instance.py
@@ -0,0 +1,48 @@
+dataset_type = 'CocoDataset'
+data_root = 'data/coco/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
+    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_train2017.json',
+        img_prefix=data_root + 'train2017/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline))
+evaluation = dict(metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/coco_instance_semantic.py b/detection/scrfd/configs/_base_/datasets/coco_instance_semantic.py
new file mode 100755
index 000000000..f7c072ec9
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/coco_instance_semantic.py
@@ -0,0 +1,53 @@
+dataset_type = 'CocoDataset'
+data_root = 'data/coco/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='LoadAnnotations', with_bbox=True, with_mask=True, with_seg=True),
+    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='SegRescale', scale_factor=1 / 8),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_train2017.json',
+        img_prefix=data_root + 'train2017/',
+        seg_prefix=data_root + 'stuffthingmaps/train2017/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        pipeline=test_pipeline))
+evaluation = dict(metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/deepfashion.py b/detection/scrfd/configs/_base_/datasets/deepfashion.py
new file mode 100755
index 000000000..308b4b2ac
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/deepfashion.py
@@ -0,0 +1,53 @@
+# dataset settings
+dataset_type = 'DeepFashionDataset'
+data_root = 'data/DeepFashion/In-shop/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
+    dict(type='Resize', img_scale=(750, 1101), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(750, 1101),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    imgs_per_gpu=2,
+    workers_per_gpu=1,
+    train=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/DeepFashion_segmentation_query.json',
+        img_prefix=data_root + 'Img/',
+        pipeline=train_pipeline,
+        data_root=data_root),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/DeepFashion_segmentation_query.json',
+        img_prefix=data_root + 'Img/',
+        pipeline=test_pipeline,
+        data_root=data_root),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root +
+        'annotations/DeepFashion_segmentation_gallery.json',
+        img_prefix=data_root + 'Img/',
+        pipeline=test_pipeline,
+        data_root=data_root))
+evaluation = dict(interval=5, metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/lvis_v0.5_instance.py b/detection/scrfd/configs/_base_/datasets/lvis_v0.5_instance.py
new file mode 100755
index 000000000..f3da861d6
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/lvis_v0.5_instance.py
@@ -0,0 +1,23 @@
+_base_ = 'coco_instance.py'
+dataset_type = 'LVISV05Dataset'
+data_root = 'data/lvis_v0.5/'
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        _delete_=True,
+        type='ClassBalancedDataset',
+        oversample_thr=1e-3,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=data_root + 'annotations/lvis_v0.5_train.json',
+            img_prefix=data_root + 'train2017/')),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/lvis_v0.5_val.json',
+        img_prefix=data_root + 'val2017/'),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/lvis_v0.5_val.json',
+        img_prefix=data_root + 'val2017/'))
+evaluation = dict(metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/lvis_v1_instance.py b/detection/scrfd/configs/_base_/datasets/lvis_v1_instance.py
new file mode 100755
index 000000000..e8c5d1b14
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/lvis_v1_instance.py
@@ -0,0 +1,23 @@
+_base_ = 'coco_instance.py'
+dataset_type = 'LVISV1Dataset'
+data_root = 'data/lvis_v1/'
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        _delete_=True,
+        type='ClassBalancedDataset',
+        oversample_thr=1e-3,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=data_root + 'annotations/lvis_v1_train.json',
+            img_prefix=data_root)),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/lvis_v1_val.json',
+        img_prefix=data_root),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/lvis_v1_val.json',
+        img_prefix=data_root))
+evaluation = dict(metric=['bbox', 'segm'])
diff --git a/detection/scrfd/configs/_base_/datasets/retinaface.py b/detection/scrfd/configs/_base_/datasets/retinaface.py
new file mode 100755
index 000000000..4c5b36961
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/retinaface.py
@@ -0,0 +1,85 @@
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop',
+         crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0]),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1100, 1650),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+#evaluation = dict(interval=1, metric='bbox')
+evaluation = dict(interval=10, metric='mAP')
diff --git a/detection/scrfd/configs/_base_/datasets/voc0712.py b/detection/scrfd/configs/_base_/datasets/voc0712.py
new file mode 100755
index 000000000..ae09acdd5
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/voc0712.py
@@ -0,0 +1,55 @@
+# dataset settings
+dataset_type = 'VOCDataset'
+data_root = 'data/VOCdevkit/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='Resize', img_scale=(1000, 600), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1000, 600),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type='RepeatDataset',
+        times=3,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=[
+                data_root + 'VOC2007/ImageSets/Main/trainval.txt',
+                data_root + 'VOC2012/ImageSets/Main/trainval.txt'
+            ],
+            img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
+        img_prefix=data_root + 'VOC2007/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
+        img_prefix=data_root + 'VOC2007/',
+        pipeline=test_pipeline))
+evaluation = dict(interval=1, metric='mAP')
diff --git a/detection/scrfd/configs/_base_/datasets/wider_face.py b/detection/scrfd/configs/_base_/datasets/wider_face.py
new file mode 100755
index 000000000..d1d649be4
--- /dev/null
+++ b/detection/scrfd/configs/_base_/datasets/wider_face.py
@@ -0,0 +1,63 @@
+# dataset settings
+dataset_type = 'WIDERFaceDataset'
+data_root = 'data/WIDERFace/'
+img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[1, 1, 1], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Expand',
+        mean=img_norm_cfg['mean'],
+        to_rgb=img_norm_cfg['to_rgb'],
+        ratio_range=(1, 4)),
+    dict(
+        type='MinIoURandomCrop',
+        min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
+        min_crop_size=0.3),
+    dict(type='Resize', img_scale=(300, 300), keep_ratio=False),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(300, 300),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=False),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=60,
+    workers_per_gpu=2,
+    train=dict(
+        type='RepeatDataset',
+        times=2,
+        dataset=dict(
+            type=dataset_type,
+            ann_file=data_root + 'train.txt',
+            img_prefix=data_root + 'WIDER_train/',
+            min_size=17,
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'val.txt',
+        img_prefix=data_root + 'WIDER_val/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + 'val.txt',
+        img_prefix=data_root + 'WIDER_val/',
+        pipeline=test_pipeline))
diff --git a/detection/scrfd/configs/_base_/default_runtime.py b/detection/scrfd/configs/_base_/default_runtime.py
new file mode 100755
index 000000000..594de8dcc
--- /dev/null
+++ b/detection/scrfd/configs/_base_/default_runtime.py
@@ -0,0 +1,14 @@
+checkpoint_config = dict(interval=1)
+# yapf:disable
+log_config = dict(
+    interval=50,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+# yapf:enable
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
diff --git a/detection/scrfd/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py b/detection/scrfd/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py
new file mode 100755
index 000000000..f90b78cef
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py
@@ -0,0 +1,200 @@
+# model settings
+model = dict(
+    type='CascadeRCNN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=256,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[8],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+    roi_head=dict(
+        type='CascadeRoIHead',
+        num_stages=3,
+        stage_loss_weights=[1, 0.5, 0.25],
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        bbox_head=[
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.1, 0.1, 0.2, 0.2]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
+                               loss_weight=1.0)),
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.05, 0.05, 0.1, 0.1]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
+                               loss_weight=1.0)),
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.033, 0.033, 0.067, 0.067]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
+        ],
+        mask_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        mask_head=dict(
+            type='FCNMaskHead',
+            num_convs=4,
+            in_channels=256,
+            conv_out_channels=256,
+            num_classes=80,
+            loss_mask=dict(
+                type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=2000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=[
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.5,
+                neg_iou_thr=0.5,
+                min_pos_iou=0.5,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            mask_size=28,
+            pos_weight=-1,
+            debug=False),
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.6,
+                neg_iou_thr=0.6,
+                min_pos_iou=0.6,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            mask_size=28,
+            pos_weight=-1,
+            debug=False),
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.7,
+                neg_iou_thr=0.7,
+                min_pos_iou=0.7,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            mask_size=28,
+            pos_weight=-1,
+            debug=False)
+    ])
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=1000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100,
+        mask_thr_binary=0.5))
diff --git a/detection/scrfd/configs/_base_/models/cascade_rcnn_r50_fpn.py b/detection/scrfd/configs/_base_/models/cascade_rcnn_r50_fpn.py
new file mode 100755
index 000000000..303276b84
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/cascade_rcnn_r50_fpn.py
@@ -0,0 +1,183 @@
+# model settings
+model = dict(
+    type='CascadeRCNN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=256,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[8],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+    roi_head=dict(
+        type='CascadeRoIHead',
+        num_stages=3,
+        stage_loss_weights=[1, 0.5, 0.25],
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        bbox_head=[
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.1, 0.1, 0.2, 0.2]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
+                               loss_weight=1.0)),
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.05, 0.05, 0.1, 0.1]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
+                               loss_weight=1.0)),
+            dict(
+                type='Shared2FCBBoxHead',
+                in_channels=256,
+                fc_out_channels=1024,
+                roi_feat_size=7,
+                num_classes=80,
+                bbox_coder=dict(
+                    type='DeltaXYWHBBoxCoder',
+                    target_means=[0., 0., 0., 0.],
+                    target_stds=[0.033, 0.033, 0.067, 0.067]),
+                reg_class_agnostic=True,
+                loss_cls=dict(
+                    type='CrossEntropyLoss',
+                    use_sigmoid=False,
+                    loss_weight=1.0),
+                loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
+        ]))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=2000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=[
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.5,
+                neg_iou_thr=0.5,
+                min_pos_iou=0.5,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            pos_weight=-1,
+            debug=False),
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.6,
+                neg_iou_thr=0.6,
+                min_pos_iou=0.6,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            pos_weight=-1,
+            debug=False),
+        dict(
+            assigner=dict(
+                type='MaxIoUAssigner',
+                pos_iou_thr=0.7,
+                neg_iou_thr=0.7,
+                min_pos_iou=0.7,
+                match_low_quality=False,
+                ignore_iof_thr=-1),
+            sampler=dict(
+                type='RandomSampler',
+                num=512,
+                pos_fraction=0.25,
+                neg_pos_ub=-1,
+                add_gt_as_proposals=True),
+            pos_weight=-1,
+            debug=False)
+    ])
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=1000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100))
diff --git a/detection/scrfd/configs/_base_/models/fast_rcnn_r50_fpn.py b/detection/scrfd/configs/_base_/models/fast_rcnn_r50_fpn.py
new file mode 100755
index 000000000..b8d9570de
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/fast_rcnn_r50_fpn.py
@@ -0,0 +1,62 @@
+# model settings
+model = dict(
+    type='FastRCNN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    roi_head=dict(
+        type='StandardRoIHead',
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=256,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100))
diff --git a/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_c4.py b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_c4.py
new file mode 100755
index 000000000..5a3816363
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_c4.py
@@ -0,0 +1,116 @@
+# model settings
+norm_cfg = dict(type='BN', requires_grad=False)
+model = dict(
+    type='FasterRCNN',
+    pretrained='open-mmlab://detectron2/resnet50_caffe',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=3,
+        strides=(1, 2, 2),
+        dilations=(1, 1, 1),
+        out_indices=(2, ),
+        frozen_stages=1,
+        norm_cfg=norm_cfg,
+        norm_eval=True,
+        style='caffe'),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=1024,
+        feat_channels=1024,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[2, 4, 8, 16, 32],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[16]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        shared_head=dict(
+            type='ResLayer',
+            depth=50,
+            stage=3,
+            stride=2,
+            dilation=1,
+            style='caffe',
+            norm_cfg=norm_cfg,
+            norm_eval=True),
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
+            out_channels=1024,
+            featmap_strides=[16]),
+        bbox_head=dict(
+            type='BBoxHead',
+            with_avg_pool=True,
+            roi_feat_size=7,
+            in_channels=2048,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=12000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=6000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100))
diff --git a/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_dc5.py b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_dc5.py
new file mode 100755
index 000000000..5b4e4c3d6
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_caffe_dc5.py
@@ -0,0 +1,107 @@
+# model settings
+norm_cfg = dict(type='BN', requires_grad=False)
+model = dict(
+    type='FasterRCNN',
+    pretrained='open-mmlab://detectron2/resnet50_caffe',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        strides=(1, 2, 2, 1),
+        dilations=(1, 1, 1, 2),
+        out_indices=(3, ),
+        frozen_stages=1,
+        norm_cfg=norm_cfg,
+        norm_eval=True,
+        style='caffe'),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=2048,
+        feat_channels=2048,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[2, 4, 8, 16, 32],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[16]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=2048,
+            featmap_strides=[16]),
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=2048,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=12000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=6000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100))
diff --git a/detection/scrfd/configs/_base_/models/faster_rcnn_r50_fpn.py b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_fpn.py
new file mode 100755
index 000000000..338a5c6b6
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/faster_rcnn_r50_fpn.py
@@ -0,0 +1,111 @@
+model = dict(
+    type='FasterRCNN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=256,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[8],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=256,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=-1,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=2000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=1000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100)
+    # soft-nms is also supported for rcnn testing
+    # e.g., nms=dict(type='soft_nms', iou_threshold=0.5, min_score=0.05)
+)
diff --git a/detection/scrfd/configs/_base_/models/mask_rcnn_r50_caffe_c4.py b/detection/scrfd/configs/_base_/models/mask_rcnn_r50_caffe_c4.py
new file mode 100755
index 000000000..b9b29b0b9
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/mask_rcnn_r50_caffe_c4.py
@@ -0,0 +1,127 @@
+# model settings
+norm_cfg = dict(type='BN', requires_grad=False)
+model = dict(
+    type='MaskRCNN',
+    pretrained='open-mmlab://detectron2/resnet50_caffe',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=3,
+        strides=(1, 2, 2),
+        dilations=(1, 1, 1),
+        out_indices=(2, ),
+        frozen_stages=1,
+        norm_cfg=norm_cfg,
+        norm_eval=True,
+        style='caffe'),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=1024,
+        feat_channels=1024,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[2, 4, 8, 16, 32],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[16]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        shared_head=dict(
+            type='ResLayer',
+            depth=50,
+            stage=3,
+            stride=2,
+            dilation=1,
+            style='caffe',
+            norm_cfg=norm_cfg,
+            norm_eval=True),
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
+            out_channels=1024,
+            featmap_strides=[16]),
+        bbox_head=dict(
+            type='BBoxHead',
+            with_avg_pool=True,
+            roi_feat_size=7,
+            in_channels=2048,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+        mask_roi_extractor=None,
+        mask_head=dict(
+            type='FCNMaskHead',
+            num_convs=0,
+            in_channels=2048,
+            conv_out_channels=256,
+            num_classes=80,
+            loss_mask=dict(
+                type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=12000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        mask_size=14,
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=6000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100,
+        mask_thr_binary=0.5))
diff --git a/detection/scrfd/configs/_base_/models/mask_rcnn_r50_fpn.py b/detection/scrfd/configs/_base_/models/mask_rcnn_r50_fpn.py
new file mode 100755
index 000000000..4472bd0a8
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/mask_rcnn_r50_fpn.py
@@ -0,0 +1,124 @@
+# model settings
+model = dict(
+    type='MaskRCNN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=256,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[8],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=256,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+        mask_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
+            out_channels=256,
+            featmap_strides=[4, 8, 16, 32]),
+        mask_head=dict(
+            type='FCNMaskHead',
+            num_convs=4,
+            in_channels=256,
+            conv_out_channels=256,
+            num_classes=80,
+            loss_mask=dict(
+                type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=-1,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=2000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        mask_size=28,
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=1000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100,
+        mask_thr_binary=0.5))
diff --git a/detection/scrfd/configs/_base_/models/retinanet_r50_fpn.py b/detection/scrfd/configs/_base_/models/retinanet_r50_fpn.py
new file mode 100755
index 000000000..a08b14f60
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/retinanet_r50_fpn.py
@@ -0,0 +1,60 @@
+# model settings
+model = dict(
+    type='RetinaNet',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        start_level=1,
+        add_extra_convs='on_input',
+        num_outs=5),
+    bbox_head=dict(
+        type='RetinaHead',
+        num_classes=80,
+        in_channels=256,
+        stacked_convs=4,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            octave_base_scale=4,
+            scales_per_octave=3,
+            ratios=[0.5, 1.0, 2.0],
+            strides=[8, 16, 32, 64, 128]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)))
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(
+        type='MaxIoUAssigner',
+        pos_iou_thr=0.5,
+        neg_iou_thr=0.4,
+        min_pos_iou=0,
+        ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=1000,
+    min_bbox_size=0,
+    score_thr=0.05,
+    nms=dict(type='nms', iou_threshold=0.5),
+    max_per_img=100)
diff --git a/detection/scrfd/configs/_base_/models/rpn_r50_caffe_c4.py b/detection/scrfd/configs/_base_/models/rpn_r50_caffe_c4.py
new file mode 100755
index 000000000..bd5d665e0
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/rpn_r50_caffe_c4.py
@@ -0,0 +1,58 @@
+# model settings
+model = dict(
+    type='RPN',
+    pretrained='open-mmlab://detectron2/resnet50_caffe',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=3,
+        strides=(1, 2, 2),
+        dilations=(1, 1, 1),
+        out_indices=(2, ),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=False),
+        norm_eval=True,
+        style='caffe'),
+    neck=None,
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=1024,
+        feat_channels=1024,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[2, 4, 8, 16, 32],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[16]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=12000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0))
diff --git a/detection/scrfd/configs/_base_/models/rpn_r50_fpn.py b/detection/scrfd/configs/_base_/models/rpn_r50_fpn.py
new file mode 100755
index 000000000..13e96191d
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/rpn_r50_fpn.py
@@ -0,0 +1,60 @@
+# model settings
+model = dict(
+    type='RPN',
+    pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=True,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024, 2048],
+        out_channels=256,
+        num_outs=5),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=256,
+        feat_channels=256,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[8],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=2000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0))
diff --git a/detection/scrfd/configs/_base_/models/ssd300.py b/detection/scrfd/configs/_base_/models/ssd300.py
new file mode 100755
index 000000000..ee7cf3adc
--- /dev/null
+++ b/detection/scrfd/configs/_base_/models/ssd300.py
@@ -0,0 +1,49 @@
+# model settings
+input_size = 300
+model = dict(
+    type='SingleStageDetector',
+    pretrained='open-mmlab://vgg16_caffe',
+    backbone=dict(
+        type='SSDVGG',
+        input_size=input_size,
+        depth=16,
+        with_last_pool=False,
+        ceil_mode=True,
+        out_indices=(3, 4),
+        out_feature_indices=(22, 34),
+        l2_norm_scale=20),
+    neck=None,
+    bbox_head=dict(
+        type='SSDHead',
+        in_channels=(512, 1024, 512, 256, 256, 256),
+        num_classes=80,
+        anchor_generator=dict(
+            type='SSDAnchorGenerator',
+            scale_major=False,
+            input_size=input_size,
+            basesize_ratio_range=(0.15, 0.9),
+            strides=[8, 16, 32, 64, 100, 300],
+            ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[0.1, 0.1, 0.2, 0.2])))
+cudnn_benchmark = True
+train_cfg = dict(
+    assigner=dict(
+        type='MaxIoUAssigner',
+        pos_iou_thr=0.5,
+        neg_iou_thr=0.5,
+        min_pos_iou=0.,
+        ignore_iof_thr=-1,
+        gt_max_assign_all=False),
+    smoothl1_beta=1.,
+    allowed_border=-1,
+    pos_weight=-1,
+    neg_pos_ratio=3,
+    debug=False)
+test_cfg = dict(
+    nms=dict(type='nms', iou_threshold=0.45),
+    min_bbox_size=0,
+    score_thr=0.02,
+    max_per_img=200)
diff --git a/detection/scrfd/configs/_base_/schedules/schedule_1x.py b/detection/scrfd/configs/_base_/schedules/schedule_1x.py
new file mode 100755
index 000000000..12694c87a
--- /dev/null
+++ b/detection/scrfd/configs/_base_/schedules/schedule_1x.py
@@ -0,0 +1,11 @@
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[8, 11])
+total_epochs = 12
diff --git a/detection/scrfd/configs/_base_/schedules/schedule_20e.py b/detection/scrfd/configs/_base_/schedules/schedule_20e.py
new file mode 100755
index 000000000..0559030c2
--- /dev/null
+++ b/detection/scrfd/configs/_base_/schedules/schedule_20e.py
@@ -0,0 +1,11 @@
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[16, 19])
+total_epochs = 20
diff --git a/detection/scrfd/configs/_base_/schedules/schedule_2x.py b/detection/scrfd/configs/_base_/schedules/schedule_2x.py
new file mode 100755
index 000000000..e34095ff2
--- /dev/null
+++ b/detection/scrfd/configs/_base_/schedules/schedule_2x.py
@@ -0,0 +1,11 @@
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[16, 22])
+total_epochs = 24
diff --git a/detection/scrfd/configs/_base_/schedules/schedule_retinaface_sgd.py b/detection/scrfd/configs/_base_/schedules/schedule_retinaface_sgd.py
new file mode 100755
index 000000000..8512bc5d8
--- /dev/null
+++ b/detection/scrfd/configs/_base_/schedules/schedule_retinaface_sgd.py
@@ -0,0 +1,11 @@
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
+optimizer_config = dict(grad_clip=None)
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=500,
+    warmup_ratio=0.001,
+    step=[55, 68])
+total_epochs = 80
diff --git a/detection/scrfd/configs/scrfd/base_1.3g.py b/detection/scrfd/configs/scrfd/base_1.3g.py
new file mode 100755
index 000000000..9a28b937c
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_1.3g.py
@@ -0,0 +1,185 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    #dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        #img_scale=(1100, 1650),
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            #dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=18,
+        base_channels=16,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        #frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[64, 128, 256, 512],
+        in_channels=[16, 32, 64, 128],
+        out_channels=32,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=32,
+        stacked_convs=2,
+        feat_channels=64,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share = True,
+        strides_share = True,
+        scale_mode = 2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=80)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
+
diff --git a/detection/scrfd/configs/scrfd/base_10g.py b/detection/scrfd/configs/scrfd/base_10g.py
new file mode 100755
index 000000000..c857c98fc
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_10g.py
@@ -0,0 +1,191 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop',
+         crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        #img_scale=(1100, 1650),
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            #dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+
+
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNetV1e',
+        #depth=0,
+        #block_cfg=dict(
+        #    block='BasicBlock',
+        #    stage_blocks=(3, 4, 2, 3),
+        #    stage_planes=[24, 40, 40, 96]),
+        #base_channels=24,
+        depth=34,
+        base_channels=32,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        #frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[24, 40, 40, 96],
+        in_channels=[32, 64, 128, 256],
+        out_channels=128,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=128,
+        stacked_convs=2,
+        feat_channels=160,
+        norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
diff --git a/detection/scrfd/configs/scrfd/base_1g.py b/detection/scrfd/configs/scrfd/base_1g.py
new file mode 100755
index 000000000..55ec81770
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_1g.py
@@ -0,0 +1,185 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop',
+         crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(3, 3, 7, 3), stage_planes=[16, 16, 32, 64, 128,
+                                                     256]),
+        #frozen_stages=1,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        #norm_eval=False,
+        #style='pytorch',
+        ),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[64, 128, 256, 512],
+        in_channels=[32, 64, 128, 256],
+        out_channels=64,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=64,
+        stacked_convs=2,
+        feat_channels=128,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share = True,
+        strides_share = True,
+        dw_conv = True,
+        scale_mode = 2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            #octave_base_scale=3,
+            #scales_per_octave=1,
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=80)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
diff --git a/detection/scrfd/configs/scrfd/base_2.5g.py b/detection/scrfd/configs/scrfd/base_2.5g.py
new file mode 100755
index 000000000..d37d56acc
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_2.5g.py
@@ -0,0 +1,185 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    #dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        #img_scale=(1100, 1650),
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            #dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=34,
+        base_channels=16,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        #frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[64, 128, 256, 512],
+        in_channels=[16, 32, 64, 128],
+        out_channels=48,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=48,
+        stacked_convs=2,
+        feat_channels=96,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share = True,
+        strides_share = True,
+        scale_mode = 2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=80)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
+
diff --git a/detection/scrfd/configs/scrfd/base_34g.py b/detection/scrfd/configs/scrfd/base_34g.py
new file mode 100755
index 000000000..da807cd14
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_34g.py
@@ -0,0 +1,185 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop',
+         crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            #dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=56,
+        base_channels=64,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        #frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        norm_eval=False,
+        style='pytorch'),
+
+    #backbone=dict(type='Res2Net', depth=50, scales=4, base_width=26),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[64, 512, 1024, 2048],
+        #in_channels=[16, 32, 64, 128],
+        #in_channels=[32, 64, 128, 256],
+        out_channels=128,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=128,
+        stacked_convs=2,
+        feat_channels=256,
+        norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 5
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
diff --git a/detection/scrfd/configs/scrfd/base_500m.py b/detection/scrfd/configs/scrfd/base_500m.py
new file mode 100755
index 000000000..dd83b2876
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/base_500m.py
@@ -0,0 +1,185 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(type='LoadAnnotations', with_bbox=True),
+    #dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    #dict(type='RandomFlip', flip_ratio=0.5),
+    #dict(type='Normalize', **img_norm_cfg),
+    #dict(type='Pad', size_divisor=32),
+    #dict(type='DefaultFormatBundle'),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_keypoints', 'gt_labels']),
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop',
+         crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+    #dict(type='LoadImageFromFile'),
+    #dict(
+    #    type='MultiScaleFlipAug',
+    #    img_scale=(1333, 800),
+    #    flip=False,
+    #    transforms=[
+    #        dict(type='Resize', keep_ratio=True),
+    #        dict(type='RandomFlip'),
+    #        dict(type='Normalize', **img_norm_cfg),
+    #        dict(type='Pad', size_divisor=32),
+    #        dict(type='ImageToTensor', keys=['img']),
+    #        dict(type='Collect', keys=['img']),
+    #    ])
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    #pretrained='torchvision://resnet50',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(2, 2, 6, 3), stage_planes=[8, 16, 32, 64, 128,
+                                                     256]),
+        #frozen_stages=1,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        #norm_eval=False,
+        #style='pytorch',
+        ),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[64, 128, 256, 512],
+        in_channels=[32, 64, 128, 256],
+        out_channels=32,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=32,
+        stacked_convs=2,
+        feat_channels=80,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share = True,
+        strides_share = True,
+        dw_conv = True,
+        scale_mode = 2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            #octave_base_scale=3,
+            #scales_per_octave=1,
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=80)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
diff --git a/detection/scrfd/configs/scrfd/scrfd_1.3g.py b/detection/scrfd/configs/scrfd/scrfd_1.3g.py
new file mode 100755
index 000000000..41c7be239
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_1.3g.py
@@ -0,0 +1,235 @@
+lr_mult = 8
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0],
+        bbox_clip_border=False),
+    dict(
+        type='Resize',
+        img_scale=(640, 640),
+        keep_ratio=False,
+        bbox_clip_border=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 5, 3, 2),
+            stage_planes=[16, 32, 40, 72]),
+        base_channels=16,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[16, 32, 40, 72],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=48,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9, mode=0),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_10g.py b/detection/scrfd/configs/scrfd/scrfd_10g.py
new file mode 100755
index 000000000..ad8cb60c0
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_10g.py
@@ -0,0 +1,235 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0],
+        bbox_clip_border=False),
+    dict(
+        type='Resize',
+        img_scale=(640, 640),
+        keep_ratio=False,
+        bbox_clip_border=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=3,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 4, 2, 3),
+            stage_planes=[56, 88, 88, 224]),
+        base_channels=56,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[56, 88, 88, 224],
+        out_channels=56,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=56,
+        stacked_convs=3,
+        feat_channels=80,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_10g_bnkps.py b/detection/scrfd/configs/scrfd/scrfd_10g_bnkps.py
new file mode 100755
index 000000000..668da0d24
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_10g_bnkps.py
@@ -0,0 +1,235 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0],
+        bbox_clip_border=False),
+    dict(
+        type='Resize',
+        img_scale=(640, 640),
+        keep_ratio=False,
+        bbox_clip_border=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=3,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 4, 2, 3),
+            stage_planes=[56, 88, 88, 224]),
+        base_channels=56,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[56, 88, 88, 224],
+        out_channels=56,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=56,
+        stacked_convs=3,
+        feat_channels=80,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=False,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=True,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_1g.py b/detection/scrfd/configs/scrfd/scrfd_1g.py
new file mode 100755
index 000000000..8a7397449
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_1g.py
@@ -0,0 +1,217 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80*4)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ]),
+            dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(3, 2, 1, 5),
+            stage_planes=[32, 48, 48, 160, 216, 312])),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[48, 160, 216, 312],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=96,
+        norm_cfg=dict(type='GN', num_groups=8, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        dw_conv=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[1, 2],
+            base_sizes=[16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80*2, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_1gbn.py b/detection/scrfd/configs/scrfd/scrfd_1gbn.py
new file mode 100755
index 000000000..2082624ea
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_1gbn.py
@@ -0,0 +1,218 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80*4)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ]),
+            dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(3, 2, 1, 5),
+            stage_planes=[32, 48, 48, 160, 216, 312])),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[48, 160, 216, 312],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=96,
+        #norm_cfg=dict(type='GN', num_groups=8, requires_grad=True),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        cls_reg_share=True,
+        strides_share=False,
+        dw_conv=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[1, 2],
+            base_sizes=[16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80*2, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_2.5g.py b/detection/scrfd/configs/scrfd/scrfd_2.5g.py
new file mode 100755
index 000000000..a0c6a3b9f
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_2.5g.py
@@ -0,0 +1,235 @@
+lr_mult = 8
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0],
+        bbox_clip_border=False),
+    dict(
+        type='Resize',
+        img_scale=(640, 640),
+        keep_ratio=False,
+        bbox_clip_border=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 5, 3, 2),
+            stage_planes=[24, 48, 48, 80]),
+        base_channels=24,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[24, 48, 48, 80],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=64,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9, mode=0),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_2.5g_bnkps.py b/detection/scrfd/configs/scrfd/scrfd_2.5g_bnkps.py
new file mode 100755
index 000000000..7cbed91a7
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_2.5g_bnkps.py
@@ -0,0 +1,235 @@
+lr_mult = 8
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0],
+        bbox_clip_border=False),
+    dict(
+        type='Resize',
+        img_scale=(640, 640),
+        keep_ratio=False,
+        bbox_clip_border=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 5, 3, 2),
+            stage_planes=[24, 48, 48, 80]),
+        base_channels=24,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[24, 48, 48, 80],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=64,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=False,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=True,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9, mode=0),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_34g.py b/detection/scrfd/configs/scrfd/scrfd_34g.py
new file mode 100755
index 000000000..842010137
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_34g.py
@@ -0,0 +1,225 @@
+optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=3,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ]),
+            dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='Bottleneck',
+            stage_blocks=(17, 16, 2, 8),
+            stage_planes=[56, 56, 144, 184]),
+        base_channels=56,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[224, 224, 576, 736],
+        out_channels=128,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=128,
+        stacked_convs=2,
+        feat_channels=256,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[1, 2],
+            base_sizes=[16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_500m.py b/detection/scrfd/configs/scrfd/scrfd_500m.py
new file mode 100755
index 000000000..4c75a5377
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_500m.py
@@ -0,0 +1,218 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=3,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ]),
+            dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(2, 3, 2, 6), stage_planes=[16, 16, 40, 72, 152,
+                                                     288])),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[40, 72, 152, 288],
+        out_channels=16,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=16,
+        stacked_convs=2,
+        feat_channels=64,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=True,
+        dw_conv=True,
+        scale_mode=2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[1, 2],
+            base_sizes=[16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=False,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd/scrfd_500m_bnkps.py b/detection/scrfd/configs/scrfd/scrfd_500m_bnkps.py
new file mode 100755
index 000000000..6fdeab39e
--- /dev/null
+++ b/detection/scrfd/configs/scrfd/scrfd_500m_bnkps.py
@@ -0,0 +1,218 @@
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_mult = 8
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=80)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = 'data/retinaface/train/'
+val_root = 'data/retinaface/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(
+        type='RandomSquareCrop',
+        crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(
+        type='Normalize',
+        mean=[127.5, 127.5, 127.5],
+        std=[128.0, 128.0, 128.0],
+        to_rgb=True),
+    dict(type='DefaultFormatBundle'),
+    dict(
+        type='Collect',
+        keys=[
+            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+            'gt_keypointss'
+        ])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='Pad', size=(640, 640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=16,
+    workers_per_gpu=4,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/train/labelv2.txt',
+        img_prefix='data/retinaface/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ]),
+            dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/retinaface/val/labelv2.txt',
+        img_prefix='data/retinaface/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='MobileNetV1',
+        block_cfg=dict(
+            stage_blocks=(2, 3, 2, 6), stage_planes=[16, 16, 40, 72, 152,
+                                                     288])),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[40, 72, 152, 288],
+        out_channels=16,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=16,
+        stacked_convs=2,
+        feat_channels=64,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=False,
+        dw_conv=True,
+        scale_mode=0,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[1, 2],
+            base_sizes=[16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=True,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+epoch_multi = 1
+evaluation = dict(interval=80, metric='mAP')
diff --git a/detection/scrfd/configs/scrfd_crowdhuman/scrfd_crowdhuman_2.5g_bnkps.py b/detection/scrfd/configs/scrfd_crowdhuman/scrfd_crowdhuman_2.5g_bnkps.py
new file mode 100755
index 000000000..61f63c339
--- /dev/null
+++ b/detection/scrfd/configs/scrfd_crowdhuman/scrfd_crowdhuman_2.5g_bnkps.py
@@ -0,0 +1,189 @@
+lr_mult = 4
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+optimizer_config = dict(grad_clip=None)
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=2500,
+    warmup_ratio=0.001,
+    step=[55*lr_mult, 68*lr_mult])
+total_epochs = 80*lr_mult
+checkpoint_config = dict(interval=total_epochs)
+log_config = dict(interval=100, hooks=[dict(type='TextLoggerHook')])
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/crowdhuman/'
+train_root = 'data/crowdhuman/train/'
+val_root = 'data/crowdhuman/val/'
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/crowdhuman/train/label_fullbody.txt',
+        img_prefix='data/crowdhuman/train/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+            dict(
+                type='RandomSquareCrop',
+                crop_choice=[
+                    0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0
+                ],
+                bbox_clip_border=False),
+            dict(
+                type='Resize',
+                img_scale=(640, 640),
+                keep_ratio=False,
+                bbox_clip_border=False),
+            dict(type='RandomFlip', flip_ratio=0.5),
+            dict(
+                type='PhotoMetricDistortion',
+                brightness_delta=32,
+                contrast_range=(0.5, 1.5),
+                saturation_range=(0.5, 1.5),
+                hue_delta=18),
+            dict(
+                type='Normalize',
+                mean=[127.5, 127.5, 127.5],
+                std=[128.0, 128.0, 128.0],
+                to_rgb=True),
+            dict(type='DefaultFormatBundle'),
+            dict(
+                type='Collect',
+                keys=[
+                    'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore',
+                    'gt_keypointss'
+                ])
+        ]),
+    val=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/crowdhuman/val/label_fullbody.txt',
+        img_prefix='data/crowdhuman/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    #dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]),
+    test=dict(
+        type='RetinaFaceDataset',
+        ann_file='data/crowdhuman/val/label_fullbody.txt',
+        img_prefix='data/crowdhuman/val/images/',
+        pipeline=[
+            dict(type='LoadImageFromFile'),
+            dict(
+                type='MultiScaleFlipAug',
+                img_scale=(640, 640),
+                flip=False,
+                transforms=[
+                    dict(type='Resize', keep_ratio=True),
+                    #dict(type='RandomFlip', flip_ratio=0.0),
+                    dict(
+                        type='Normalize',
+                        mean=[127.5, 127.5, 127.5],
+                        std=[128.0, 128.0, 128.0],
+                        to_rgb=True),
+                    dict(type='Pad', size=(640, 640), pad_val=0),
+                    dict(type='ImageToTensor', keys=['img']),
+                    dict(type='Collect', keys=['img'])
+                ])
+        ]))
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0,
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 5, 3, 2),
+            stage_planes=[24, 48, 48, 80]),
+        base_channels=24,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        norm_cfg=dict(type='BN', requires_grad=True),
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        in_channels=[24, 48, 48, 80],
+        out_channels=24,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=5),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=24,
+        stacked_convs=2,
+        feat_channels=64,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share=True,
+        strides_share=False,
+        scale_mode=2,
+        #anchor_generator=dict(
+        #    type='AnchorGenerator',
+        #    ratios=[1.0],
+        #    scales = [1,2],
+        #    base_sizes = [16, 64, 256],
+        #    strides=[8, 16, 32]),
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[2.0],
+            scales = [3],
+            base_sizes = [8, 16, 32, 64, 128],
+            strides=[8, 16, 32, 64, 128]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        loss_dfl=False,
+        reg_max=8,
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps=True,
+        loss_kps=dict(
+            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=0.1),
+        train_cfg=dict(
+            assigner=dict(type='ATSSAssigner', topk=9),
+            allowed_border=-1,
+            pos_weight=-1,
+            debug=False),
+        test_cfg=dict(
+            nms_pre=-1,
+            min_bbox_size=0,
+            score_thr=0.02,
+            nms=dict(type='nms', iou_threshold=0.45),
+            max_per_img=-1)))
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9, mode=0),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+evaluation = dict(interval=40, metric='mAP')
diff --git a/detection/scrfd/configs/scrfdgen2.5g/scrfdgen2.5g_0.py b/detection/scrfd/configs/scrfdgen2.5g/scrfdgen2.5g_0.py
new file mode 100755
index 000000000..b1052eff5
--- /dev/null
+++ b/detection/scrfd/configs/scrfdgen2.5g/scrfdgen2.5g_0.py
@@ -0,0 +1,167 @@
+_base_ = [
+    #'../_base_/datasets/retinaface.py',
+    '../_base_/schedules/schedule_retinaface_sgd.py', '../_base_/default_runtime.py'
+]
+dataset_type = 'RetinaFaceDataset'
+data_root = 'data/retinaface/'
+train_root = data_root+'train/'
+val_root = data_root+'val/'
+#img_norm_cfg = dict(
+#    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_norm_cfg = dict(
+    mean=[127.5, 127.5, 127.5], std=[128.0, 128.0, 128.0], to_rgb=True)
+train_pipeline = [
+
+    dict(type='LoadImageFromFile', to_float32=True),
+    dict(type='LoadAnnotations', with_bbox=True, with_keypoints=True),
+    dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]),
+    #dict(type='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0]),
+    dict(type='Resize', img_scale=(640, 640), keep_ratio=False),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='PhotoMetricDistortion',
+        brightness_delta=32,
+        contrast_range=(0.5, 1.5),
+        saturation_range=(0.5, 1.5),
+        hue_delta=18),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore', 'gt_keypointss']),
+    #dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']),
+]
+test_pipeline = [
+
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        #img_scale=(1100, 1650),
+        img_scale=(640, 640),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip', flip_ratio=0.0),
+            dict(type='Normalize', **img_norm_cfg),
+            #dict(type='Pad', size_divisor=32, pad_val=0),
+            dict(type='Pad', size=(640,640), pad_val=0),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=8,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=train_root + 'labelv2.txt',
+        #ann_file=train_root + 'label_wo.txt',
+        img_prefix=train_root+ 'images/',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        ann_file=val_root + 'labelv2.txt',
+        img_prefix=val_root+ 'images/',
+        pipeline=test_pipeline),
+    )
+model = dict(
+    type='SCRFD',
+    backbone=dict(
+        type='ResNetV1e',
+        depth=0, #refer to depth-34
+        block_cfg=dict(
+            block='BasicBlock',
+            stage_blocks=(3, 4, 6, 3),
+            stage_planes=[16, 32, 64, 128]),
+        base_channels=16,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        #frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        #norm_eval=True,
+        norm_eval=False,
+        style='pytorch'),
+    neck=dict(
+        type='PAFPN',
+        #in_channels=[64, 128, 256, 512],
+        in_channels=[16, 32, 64, 128],
+        out_channels=48,
+        start_level=1,
+        add_extra_convs='on_output',
+        num_outs=3),
+    bbox_head=dict(
+        type='SCRFDHead',
+        num_classes=1,
+        in_channels=48,
+        stacked_convs=2,
+        feat_channels=96,
+        #norm_cfg=dict(type='BN', requires_grad=True),
+        norm_cfg=dict(type='GN', num_groups=16, requires_grad=True),
+        cls_reg_share = True,
+        strides_share = True,
+        scale_mode = 2,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales = [1,2],
+            base_sizes = [16, 64, 256],
+            strides=[8, 16, 32]),
+        loss_cls=dict(
+            type='QualityFocalLoss',
+            use_sigmoid=True,
+            beta=2.0,
+            loss_weight=1.0),
+        #loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+        loss_dfl=False,
+        reg_max=8,
+        #loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
+        loss_bbox=dict(type='DIoULoss', loss_weight=2.0),
+        use_kps = False,
+        loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+        )
+    )
+# training and testing settings
+train_cfg = dict(
+    assigner=dict(type='ATSSAssigner', topk=9),
+    #assigner=dict(
+    #    type='MaxIoUAssigner',
+    #    pos_iou_thr=0.5,
+    #    neg_iou_thr=0.3,
+    #    min_pos_iou=0.5,
+    #    ignore_iof_thr=-1),
+    allowed_border=-1,
+    pos_weight=-1,
+    debug=False)
+test_cfg = dict(
+    #nms_pre=1000,
+    nms_pre=-1,
+    min_bbox_size=0,
+    score_thr=0.02,
+    nms=dict(type='nms', iou_threshold=0.45),
+    max_per_img=-1)
+# optimizer
+optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
+epoch_multi = 1
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=0.001,
+    step=[55*epoch_multi, 68*epoch_multi])
+total_epochs = 80*epoch_multi
+#checkpoint_config = dict(interval=1)
+checkpoint_config = dict(interval=80)
+log_config = dict(
+    interval=100,
+    hooks=[
+        dict(type='TextLoggerHook'),
+        # dict(type='TensorboardLoggerHook')
+    ])
+evaluation = dict(interval=80, metric='mAP')
+
+
+
+
diff --git a/detection/scrfd/demo/image_demo.py b/detection/scrfd/demo/image_demo.py
new file mode 100755
index 000000000..5fbf93d3b
--- /dev/null
+++ b/detection/scrfd/demo/image_demo.py
@@ -0,0 +1,26 @@
+from argparse import ArgumentParser
+
+from mmdet.apis import inference_detector, init_detector, show_result_pyplot
+
+
+def main():
+    parser = ArgumentParser()
+    parser.add_argument('img', help='Image file')
+    parser.add_argument('config', help='Config file')
+    parser.add_argument('checkpoint', help='Checkpoint file')
+    parser.add_argument(
+        '--device', default='cuda:0', help='Device used for inference')
+    parser.add_argument(
+        '--score-thr', type=float, default=0.3, help='bbox score threshold')
+    args = parser.parse_args()
+
+    # build the model from a config file and a checkpoint file
+    model = init_detector(args.config, args.checkpoint, device=args.device)
+    # test a single image
+    result = inference_detector(model, args.img)
+    # show the results
+    show_result_pyplot(model, args.img, result, score_thr=args.score_thr)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/demo/webcam_demo.py b/detection/scrfd/demo/webcam_demo.py
new file mode 100755
index 000000000..5bded14ff
--- /dev/null
+++ b/detection/scrfd/demo/webcam_demo.py
@@ -0,0 +1,46 @@
+import argparse
+
+import cv2
+import torch
+
+from mmdet.apis import inference_detector, init_detector
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='MMDetection webcam demo')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument('checkpoint', help='checkpoint file')
+    parser.add_argument(
+        '--device', type=str, default='cuda:0', help='CPU/CUDA device option')
+    parser.add_argument(
+        '--camera-id', type=int, default=0, help='camera device id')
+    parser.add_argument(
+        '--score-thr', type=float, default=0.5, help='bbox score threshold')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    device = torch.device(args.device)
+
+    model = init_detector(args.config, args.checkpoint, device=device)
+
+    camera = cv2.VideoCapture(args.camera_id)
+
+    print('Press "Esc", "q" or "Q" to exit.')
+    while True:
+        ret_val, img = camera.read()
+        result = inference_detector(model, img)
+
+        ch = cv2.waitKey(1)
+        if ch == 27 or ch == ord('q') or ch == ord('Q'):
+            break
+
+        model.show_result(
+            img, result, score_thr=args.score_thr, wait_time=1, show=True)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/mmdet/__init__.py b/detection/scrfd/mmdet/__init__.py
new file mode 100755
index 000000000..5a8e4bd0e
--- /dev/null
+++ b/detection/scrfd/mmdet/__init__.py
@@ -0,0 +1,29 @@
+import mmcv
+
+from .version import __version__, short_version
+
+
+def digit_version(version_str):
+    digit_version = []
+    for x in version_str.split('.'):
+        if x.isdigit():
+            digit_version.append(int(x))
+        elif x.find('rc') != -1:
+            patch_version = x.split('rc')
+            digit_version.append(int(patch_version[0]) - 1)
+            digit_version.append(int(patch_version[1]))
+    return digit_version
+
+
+mmcv_minimum_version = '1.1.5'
+mmcv_maximum_version = '1.4'
+mmcv_version = digit_version(mmcv.__version__)
+
+
+assert (mmcv_version >= digit_version(mmcv_minimum_version)
+        and mmcv_version <= digit_version(mmcv_maximum_version)), \
+    f'MMCV=={mmcv.__version__} is used but incompatible. ' \
+    f'Please install mmcv>={mmcv_minimum_version}, <={mmcv_maximum_version}.'
+
+__all__ = ['__version__', 'short_version']
+
diff --git a/detection/scrfd/mmdet/apis/__init__.py b/detection/scrfd/mmdet/apis/__init__.py
new file mode 100755
index 000000000..1d8035b74
--- /dev/null
+++ b/detection/scrfd/mmdet/apis/__init__.py
@@ -0,0 +1,10 @@
+from .inference import (async_inference_detector, inference_detector,
+                        init_detector, show_result_pyplot)
+from .test import multi_gpu_test, single_gpu_test
+from .train import get_root_logger, set_random_seed, train_detector
+
+__all__ = [
+    'get_root_logger', 'set_random_seed', 'train_detector', 'init_detector',
+    'async_inference_detector', 'inference_detector', 'show_result_pyplot',
+    'multi_gpu_test', 'single_gpu_test'
+]
diff --git a/detection/scrfd/mmdet/apis/inference.py b/detection/scrfd/mmdet/apis/inference.py
new file mode 100755
index 000000000..6fa19cd58
--- /dev/null
+++ b/detection/scrfd/mmdet/apis/inference.py
@@ -0,0 +1,187 @@
+import warnings
+
+import matplotlib.pyplot as plt
+import mmcv
+import numpy as np
+import torch
+from mmcv.ops import RoIPool
+from mmcv.parallel import collate, scatter
+from mmcv.runner import load_checkpoint
+
+from mmdet.core import get_classes
+from mmdet.datasets.pipelines import Compose
+from mmdet.models import build_detector
+
+
+def init_detector(config, checkpoint=None, device='cuda:0', cfg_options=None):
+    """Initialize a detector from config file.
+
+    Args:
+        config (str or :obj:`mmcv.Config`): Config file path or the config
+            object.
+        checkpoint (str, optional): Checkpoint path. If left as None, the model
+            will not load any weights.
+        cfg_options (dict): Options to override some settings in the used
+            config.
+
+    Returns:
+        nn.Module: The constructed detector.
+    """
+    if isinstance(config, str):
+        config = mmcv.Config.fromfile(config)
+    elif not isinstance(config, mmcv.Config):
+        raise TypeError('config must be a filename or Config object, '
+                        f'but got {type(config)}')
+    if cfg_options is not None:
+        config.merge_from_dict(cfg_options)
+    config.model.pretrained = None
+    model = build_detector(config.model, test_cfg=config.test_cfg)
+    if checkpoint is not None:
+        map_loc = 'cpu' if device == 'cpu' else None
+        checkpoint = load_checkpoint(model, checkpoint, map_location=map_loc)
+        if 'CLASSES' in checkpoint['meta']:
+            model.CLASSES = checkpoint['meta']['CLASSES']
+        else:
+            warnings.simplefilter('once')
+            warnings.warn('Class names are not saved in the checkpoint\'s '
+                          'meta data, use COCO classes by default.')
+            model.CLASSES = get_classes('coco')
+    model.cfg = config  # save the config in the model for convenience
+    model.to(device)
+    model.eval()
+    return model
+
+
+class LoadImage(object):
+    """A simple pipeline to load image."""
+
+    def __call__(self, results):
+        """Call function to load images into results.
+
+        Args:
+            results (dict): A result dict contains the file name
+                of the image to be read.
+
+        Returns:
+            dict: ``results`` will be returned containing loaded image.
+        """
+        if isinstance(results['img'], str):
+            results['filename'] = results['img']
+            results['ori_filename'] = results['img']
+        else:
+            results['filename'] = None
+            results['ori_filename'] = None
+        img = mmcv.imread(results['img'])
+        results['img'] = img
+        results['img_fields'] = ['img']
+        results['img_shape'] = img.shape
+        results['ori_shape'] = img.shape
+        return results
+
+
+def inference_detector(model, img):
+    """Inference image(s) with the detector.
+
+    Args:
+        model (nn.Module): The loaded detector.
+        imgs (str/ndarray or list[str/ndarray]): Either image files or loaded
+            images.
+
+    Returns:
+        If imgs is a str, a generator will be returned, otherwise return the
+        detection results directly.
+    """
+    cfg = model.cfg
+    device = next(model.parameters()).device  # model device
+    # prepare data
+    if isinstance(img, np.ndarray):
+        # directly add img
+        data = dict(img=img)
+        cfg = cfg.copy()
+        # set loading pipeline type
+        cfg.data.test.pipeline[0].type = 'LoadImageFromWebcam'
+    else:
+        # add information into dict
+        data = dict(img_info=dict(filename=img), img_prefix=None)
+    # build the data pipeline
+    test_pipeline = Compose(cfg.data.test.pipeline)
+    data = test_pipeline(data)
+    data = collate([data], samples_per_gpu=1)
+    if next(model.parameters()).is_cuda:
+        # scatter to specified GPU
+        data = scatter(data, [device])[0]
+    else:
+        for m in model.modules():
+            assert not isinstance(
+                m, RoIPool
+            ), 'CPU inference with RoIPool is not supported currently.'
+        # just get the actual data from DataContainer
+        data['img_metas'] = data['img_metas'][0].data
+
+    # forward the model
+    with torch.no_grad():
+        result = model(return_loss=False, rescale=True, **data)[0]
+    return result
+
+
+async def async_inference_detector(model, img):
+    """Async inference image(s) with the detector.
+
+    Args:
+        model (nn.Module): The loaded detector.
+        img (str | ndarray): Either image files or loaded images.
+
+    Returns:
+        Awaitable detection results.
+    """
+    cfg = model.cfg
+    device = next(model.parameters()).device  # model device
+    # prepare data
+    if isinstance(img, np.ndarray):
+        # directly add img
+        data = dict(img=img)
+        cfg = cfg.copy()
+        # set loading pipeline type
+        cfg.data.test.pipeline[0].type = 'LoadImageFromWebcam'
+    else:
+        # add information into dict
+        data = dict(img_info=dict(filename=img), img_prefix=None)
+    # build the data pipeline
+    test_pipeline = Compose(cfg.data.test.pipeline)
+    data = test_pipeline(data)
+    data = scatter(collate([data], samples_per_gpu=1), [device])[0]
+
+    # We don't restore `torch.is_grad_enabled()` value during concurrent
+    # inference since execution can overlap
+    torch.set_grad_enabled(False)
+    result = await model.aforward_test(rescale=True, **data)
+    return result
+
+
+def show_result_pyplot(model,
+                       img,
+                       result,
+                       score_thr=0.3,
+                       fig_size=(15, 10),
+                       title='result',
+                       block=True):
+    """Visualize the detection results on the image.
+
+    Args:
+        model (nn.Module): The loaded detector.
+        img (str or np.ndarray): Image filename or loaded image.
+        result (tuple[list] or list): The detection result, can be either
+            (bbox, segm) or just bbox.
+        score_thr (float): The threshold to visualize the bboxes and masks.
+        fig_size (tuple): Figure size of the pyplot figure.
+        title (str): Title of the pyplot figure.
+        block (bool): Whether to block GUI.
+    """
+    if hasattr(model, 'module'):
+        model = model.module
+    img = model.show_result(img, result, score_thr=score_thr, show=False)
+    plt.figure(figsize=fig_size)
+    plt.imshow(mmcv.bgr2rgb(img))
+    plt.title(title)
+    plt.tight_layout()
+    plt.show(block=block)
diff --git a/detection/scrfd/mmdet/apis/test.py b/detection/scrfd/mmdet/apis/test.py
new file mode 100755
index 000000000..e54b1b8c2
--- /dev/null
+++ b/detection/scrfd/mmdet/apis/test.py
@@ -0,0 +1,190 @@
+import os.path as osp
+import pickle
+import shutil
+import tempfile
+import time
+
+import mmcv
+import torch
+import torch.distributed as dist
+from mmcv.image import tensor2imgs
+from mmcv.runner import get_dist_info
+
+from mmdet.core import encode_mask_results
+
+
+def single_gpu_test(model,
+                    data_loader,
+                    show=False,
+                    out_dir=None,
+                    show_score_thr=0.3):
+    model.eval()
+    results = []
+    dataset = data_loader.dataset
+    prog_bar = mmcv.ProgressBar(len(dataset))
+    for i, data in enumerate(data_loader):
+        with torch.no_grad():
+            result = model(return_loss=False, rescale=True, **data)
+
+        batch_size = len(result)
+        if show or out_dir:
+            if batch_size == 1 and isinstance(data['img'][0], torch.Tensor):
+                img_tensor = data['img'][0]
+            else:
+                img_tensor = data['img'][0].data[0]
+            img_metas = data['img_metas'][0].data[0]
+            imgs = tensor2imgs(img_tensor, **img_metas[0]['img_norm_cfg'])
+            assert len(imgs) == len(img_metas)
+
+            for i, (img, img_meta) in enumerate(zip(imgs, img_metas)):
+                h, w, _ = img_meta['img_shape']
+                img_show = img[:h, :w, :]
+
+                ori_h, ori_w = img_meta['ori_shape'][:-1]
+                img_show = mmcv.imresize(img_show, (ori_w, ori_h))
+
+                if out_dir:
+                    out_file = osp.join(out_dir, img_meta['ori_filename'])
+                else:
+                    out_file = None
+
+                model.module.show_result(
+                    img_show,
+                    result[i],
+                    show=show,
+                    out_file=out_file,
+                    score_thr=show_score_thr)
+
+        # encode mask results
+        if isinstance(result[0], tuple):
+            result = [(bbox_results, encode_mask_results(mask_results))
+                      for bbox_results, mask_results in result]
+        results.extend(result)
+
+        for _ in range(batch_size):
+            prog_bar.update()
+    return results
+
+
+def multi_gpu_test(model, data_loader, tmpdir=None, gpu_collect=False):
+    """Test model with multiple gpus.
+
+    This method tests model with multiple gpus and collects the results
+    under two different modes: gpu and cpu modes. By setting 'gpu_collect=True'
+    it encodes results to gpu tensors and use gpu communication for results
+    collection. On cpu mode it saves the results on different gpus to 'tmpdir'
+    and collects them by the rank 0 worker.
+
+    Args:
+        model (nn.Module): Model to be tested.
+        data_loader (nn.Dataloader): Pytorch data loader.
+        tmpdir (str): Path of directory to save the temporary results from
+            different gpus under cpu mode.
+        gpu_collect (bool): Option to use either gpu or cpu to collect results.
+
+    Returns:
+        list: The prediction results.
+    """
+    model.eval()
+    results = []
+    dataset = data_loader.dataset
+    rank, world_size = get_dist_info()
+    if rank == 0:
+        prog_bar = mmcv.ProgressBar(len(dataset))
+    time.sleep(2)  # This line can prevent deadlock problem in some cases.
+    for i, data in enumerate(data_loader):
+        with torch.no_grad():
+            result = model(return_loss=False, rescale=True, **data)
+            # encode mask results
+            if isinstance(result[0], tuple):
+                result = [(bbox_results, encode_mask_results(mask_results))
+                          for bbox_results, mask_results in result]
+        results.extend(result)
+
+        if rank == 0:
+            batch_size = len(result)
+            for _ in range(batch_size * world_size):
+                prog_bar.update()
+
+    # collect results from all ranks
+    if gpu_collect:
+        results = collect_results_gpu(results, len(dataset))
+    else:
+        results = collect_results_cpu(results, len(dataset), tmpdir)
+    return results
+
+
+def collect_results_cpu(result_part, size, tmpdir=None):
+    rank, world_size = get_dist_info()
+    # create a tmp dir if it is not specified
+    if tmpdir is None:
+        MAX_LEN = 512
+        # 32 is whitespace
+        dir_tensor = torch.full((MAX_LEN, ),
+                                32,
+                                dtype=torch.uint8,
+                                device='cuda')
+        if rank == 0:
+            mmcv.mkdir_or_exist('.dist_test')
+            tmpdir = tempfile.mkdtemp(dir='.dist_test')
+            tmpdir = torch.tensor(
+                bytearray(tmpdir.encode()), dtype=torch.uint8, device='cuda')
+            dir_tensor[:len(tmpdir)] = tmpdir
+        dist.broadcast(dir_tensor, 0)
+        tmpdir = dir_tensor.cpu().numpy().tobytes().decode().rstrip()
+    else:
+        mmcv.mkdir_or_exist(tmpdir)
+    # dump the part result to the dir
+    mmcv.dump(result_part, osp.join(tmpdir, f'part_{rank}.pkl'))
+    dist.barrier()
+    # collect all parts
+    if rank != 0:
+        return None
+    else:
+        # load results of all parts from tmp dir
+        part_list = []
+        for i in range(world_size):
+            part_file = osp.join(tmpdir, f'part_{i}.pkl')
+            part_list.append(mmcv.load(part_file))
+        # sort the results
+        ordered_results = []
+        for res in zip(*part_list):
+            ordered_results.extend(list(res))
+        # the dataloader may pad some samples
+        ordered_results = ordered_results[:size]
+        # remove tmp dir
+        shutil.rmtree(tmpdir)
+        return ordered_results
+
+
+def collect_results_gpu(result_part, size):
+    rank, world_size = get_dist_info()
+    # dump result part to tensor with pickle
+    part_tensor = torch.tensor(
+        bytearray(pickle.dumps(result_part)), dtype=torch.uint8, device='cuda')
+    # gather all result part tensor shape
+    shape_tensor = torch.tensor(part_tensor.shape, device='cuda')
+    shape_list = [shape_tensor.clone() for _ in range(world_size)]
+    dist.all_gather(shape_list, shape_tensor)
+    # padding result part tensor to max length
+    shape_max = torch.tensor(shape_list).max()
+    part_send = torch.zeros(shape_max, dtype=torch.uint8, device='cuda')
+    part_send[:shape_tensor[0]] = part_tensor
+    part_recv_list = [
+        part_tensor.new_zeros(shape_max) for _ in range(world_size)
+    ]
+    # gather all result part
+    dist.all_gather(part_recv_list, part_send)
+
+    if rank == 0:
+        part_list = []
+        for recv, shape in zip(part_recv_list, shape_list):
+            part_list.append(
+                pickle.loads(recv[:shape[0]].cpu().numpy().tobytes()))
+        # sort the results
+        ordered_results = []
+        for res in zip(*part_list):
+            ordered_results.extend(list(res))
+        # the dataloader may pad some samples
+        ordered_results = ordered_results[:size]
+        return ordered_results
diff --git a/detection/scrfd/mmdet/apis/train.py b/detection/scrfd/mmdet/apis/train.py
new file mode 100755
index 000000000..ad17a5379
--- /dev/null
+++ b/detection/scrfd/mmdet/apis/train.py
@@ -0,0 +1,150 @@
+import random
+
+import numpy as np
+import torch
+from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
+from mmcv.runner import (HOOKS, DistSamplerSeedHook, EpochBasedRunner,
+                         Fp16OptimizerHook, OptimizerHook, build_optimizer)
+from mmcv.utils import build_from_cfg
+
+from mmdet.core import DistEvalHook, EvalHook
+from mmdet.datasets import (build_dataloader, build_dataset,
+                            replace_ImageToTensor)
+from mmdet.utils import get_root_logger
+
+
+def set_random_seed(seed, deterministic=False):
+    """Set random seed.
+
+    Args:
+        seed (int): Seed to be used.
+        deterministic (bool): Whether to set the deterministic option for
+            CUDNN backend, i.e., set `torch.backends.cudnn.deterministic`
+            to True and `torch.backends.cudnn.benchmark` to False.
+            Default: False.
+    """
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    if deterministic:
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+
+
+def train_detector(model,
+                   dataset,
+                   cfg,
+                   distributed=False,
+                   validate=False,
+                   timestamp=None,
+                   meta=None):
+    logger = get_root_logger(cfg.log_level)
+
+    # prepare data loaders
+    dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
+    if 'imgs_per_gpu' in cfg.data:
+        logger.warning('"imgs_per_gpu" is deprecated in MMDet V2.0. '
+                       'Please use "samples_per_gpu" instead')
+        if 'samples_per_gpu' in cfg.data:
+            logger.warning(
+                f'Got "imgs_per_gpu"={cfg.data.imgs_per_gpu} and '
+                f'"samples_per_gpu"={cfg.data.samples_per_gpu}, "imgs_per_gpu"'
+                f'={cfg.data.imgs_per_gpu} is used in this experiments')
+        else:
+            logger.warning(
+                'Automatically set "samples_per_gpu"="imgs_per_gpu"='
+                f'{cfg.data.imgs_per_gpu} in this experiments')
+        cfg.data.samples_per_gpu = cfg.data.imgs_per_gpu
+
+    data_loaders = [
+        build_dataloader(
+            ds,
+            cfg.data.samples_per_gpu,
+            cfg.data.workers_per_gpu,
+            # cfg.gpus will be ignored if distributed
+            len(cfg.gpu_ids),
+            dist=distributed,
+            seed=cfg.seed) for ds in dataset
+    ]
+
+    # put model on gpus
+    if distributed:
+        find_unused_parameters = cfg.get('find_unused_parameters', False)
+        # Sets the `find_unused_parameters` parameter in
+        # torch.nn.parallel.DistributedDataParallel
+        model = MMDistributedDataParallel(
+            model.cuda(),
+            device_ids=[torch.cuda.current_device()],
+            broadcast_buffers=False,
+            find_unused_parameters=find_unused_parameters)
+    else:
+        model = MMDataParallel(
+            model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
+
+    # build runner
+    optimizer = build_optimizer(model, cfg.optimizer)
+    runner = EpochBasedRunner(
+        model,
+        optimizer=optimizer,
+        work_dir=cfg.work_dir,
+        logger=logger,
+        meta=meta)
+    # an ugly workaround to make .log and .log.json filenames the same
+    runner.timestamp = timestamp
+
+    # fp16 setting
+    fp16_cfg = cfg.get('fp16', None)
+    if fp16_cfg is not None:
+        optimizer_config = Fp16OptimizerHook(
+            **cfg.optimizer_config, **fp16_cfg, distributed=distributed)
+    elif distributed and 'type' not in cfg.optimizer_config:
+        optimizer_config = OptimizerHook(**cfg.optimizer_config)
+    else:
+        optimizer_config = cfg.optimizer_config
+
+    # register hooks
+    runner.register_training_hooks(cfg.lr_config, optimizer_config,
+                                   cfg.checkpoint_config, cfg.log_config,
+                                   cfg.get('momentum_config', None))
+    if distributed:
+        runner.register_hook(DistSamplerSeedHook())
+
+    # register eval hooks
+    if validate:
+        # Support batch_size > 1 in validation
+        val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
+        if val_samples_per_gpu > 1:
+            # Replace 'ImageToTensor' to 'DefaultFormatBundle'
+            cfg.data.val.pipeline = replace_ImageToTensor(
+                cfg.data.val.pipeline)
+        val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
+        val_dataloader = build_dataloader(
+            val_dataset,
+            samples_per_gpu=val_samples_per_gpu,
+            workers_per_gpu=cfg.data.workers_per_gpu,
+            dist=distributed,
+            shuffle=False)
+        eval_cfg = cfg.get('evaluation', {})
+        eval_hook = DistEvalHook if distributed else EvalHook
+        runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
+
+    # user-defined hooks
+    if cfg.get('custom_hooks', None):
+        custom_hooks = cfg.custom_hooks
+        assert isinstance(custom_hooks, list), \
+            f'custom_hooks expect list type, but got {type(custom_hooks)}'
+        for hook_cfg in cfg.custom_hooks:
+            assert isinstance(hook_cfg, dict), \
+                'Each item in custom_hooks expects dict type, but got ' \
+                f'{type(hook_cfg)}'
+            hook_cfg = hook_cfg.copy()
+            priority = hook_cfg.pop('priority', 'NORMAL')
+            hook = build_from_cfg(hook_cfg, HOOKS)
+            runner.register_hook(hook, priority=priority)
+
+    if cfg.resume_from:
+        runner.resume(cfg.resume_from)
+    elif cfg.load_from:
+        runner.load_checkpoint(cfg.load_from)
+    runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
diff --git a/detection/scrfd/mmdet/core/__init__.py b/detection/scrfd/mmdet/core/__init__.py
new file mode 100755
index 000000000..b07536909
--- /dev/null
+++ b/detection/scrfd/mmdet/core/__init__.py
@@ -0,0 +1,8 @@
+from .anchor import *  # noqa: F401, F403
+from .bbox import *  # noqa: F401, F403
+from .evaluation import *  # noqa: F401, F403
+from .export import *  # noqa: F401, F403
+from .fp16 import *  # noqa: F401, F403
+from .mask import *  # noqa: F401, F403
+from .post_processing import *  # noqa: F401, F403
+from .utils import *  # noqa: F401, F403
diff --git a/detection/scrfd/mmdet/core/anchor/__init__.py b/detection/scrfd/mmdet/core/anchor/__init__.py
new file mode 100755
index 000000000..5838ff3ee
--- /dev/null
+++ b/detection/scrfd/mmdet/core/anchor/__init__.py
@@ -0,0 +1,11 @@
+from .anchor_generator import (AnchorGenerator, LegacyAnchorGenerator,
+                               YOLOAnchorGenerator)
+from .builder import ANCHOR_GENERATORS, build_anchor_generator
+from .point_generator import PointGenerator
+from .utils import anchor_inside_flags, calc_region, images_to_levels
+
+__all__ = [
+    'AnchorGenerator', 'LegacyAnchorGenerator', 'anchor_inside_flags',
+    'PointGenerator', 'images_to_levels', 'calc_region',
+    'build_anchor_generator', 'ANCHOR_GENERATORS', 'YOLOAnchorGenerator'
+]
diff --git a/detection/scrfd/mmdet/core/anchor/anchor_generator.py b/detection/scrfd/mmdet/core/anchor/anchor_generator.py
new file mode 100755
index 000000000..29b5ed04b
--- /dev/null
+++ b/detection/scrfd/mmdet/core/anchor/anchor_generator.py
@@ -0,0 +1,728 @@
+import mmcv
+import numpy as np
+import torch
+from torch.nn.modules.utils import _pair
+
+from .builder import ANCHOR_GENERATORS
+
+
+@ANCHOR_GENERATORS.register_module()
+class AnchorGenerator(object):
+    """Standard anchor generator for 2D anchor-based detectors.
+
+    Args:
+        strides (list[int] | list[tuple[int, int]]): Strides of anchors
+            in multiple feature levels in order (w, h).
+        ratios (list[float]): The list of ratios between the height and width
+            of anchors in a single level.
+        scales (list[int] | None): Anchor scales for anchors in a single level.
+            It cannot be set at the same time if `octave_base_scale` and
+            `scales_per_octave` are set.
+        base_sizes (list[int] | None): The basic sizes
+            of anchors in multiple levels.
+            If None is given, strides will be used as base_sizes.
+            (If strides are non square, the shortest stride is taken.)
+        scale_major (bool): Whether to multiply scales first when generating
+            base anchors. If true, the anchors in the same row will have the
+            same scales. By default it is True in V2.0
+        octave_base_scale (int): The base scale of octave.
+        scales_per_octave (int): Number of scales for each octave.
+            `octave_base_scale` and `scales_per_octave` are usually used in
+            retinanet and the `scales` should be None when they are set.
+        centers (list[tuple[float, float]] | None): The centers of the anchor
+            relative to the feature grid center in multiple feature levels.
+            By default it is set to be None and not used. If a list of tuple of
+            float is given, they will be used to shift the centers of anchors.
+        center_offset (float): The offset of center in proportion to anchors'
+            width and height. By default it is 0 in V2.0.
+
+    Examples:
+        >>> from mmdet.core import AnchorGenerator
+        >>> self = AnchorGenerator([16], [1.], [1.], [9])
+        >>> all_anchors = self.grid_anchors([(2, 2)], device='cpu')
+        >>> print(all_anchors)
+        [tensor([[-4.5000, -4.5000,  4.5000,  4.5000],
+                [11.5000, -4.5000, 20.5000,  4.5000],
+                [-4.5000, 11.5000,  4.5000, 20.5000],
+                [11.5000, 11.5000, 20.5000, 20.5000]])]
+        >>> self = AnchorGenerator([16, 32], [1.], [1.], [9, 18])
+        >>> all_anchors = self.grid_anchors([(2, 2), (1, 1)], device='cpu')
+        >>> print(all_anchors)
+        [tensor([[-4.5000, -4.5000,  4.5000,  4.5000],
+                [11.5000, -4.5000, 20.5000,  4.5000],
+                [-4.5000, 11.5000,  4.5000, 20.5000],
+                [11.5000, 11.5000, 20.5000, 20.5000]]), \
+        tensor([[-9., -9., 9., 9.]])]
+    """
+
+    def __init__(self,
+                 strides,
+                 ratios,
+                 scales=None,
+                 base_sizes=None,
+                 scale_major=True,
+                 octave_base_scale=None,
+                 scales_per_octave=None,
+                 centers=None,
+                 center_offset=0.):
+        # check center and center_offset
+        if center_offset != 0:
+            assert centers is None, 'center cannot be set when center_offset' \
+                f'!=0, {centers} is given.'
+        if not (0 <= center_offset <= 1):
+            raise ValueError('center_offset should be in range [0, 1], '
+                             f'{center_offset} is given.')
+        if centers is not None:
+            assert len(centers) == len(strides), \
+                'The number of strides should be the same as centers, got ' \
+                f'{strides} and {centers}'
+
+        # calculate base sizes of anchors
+        self.strides = [_pair(stride) for stride in strides]
+        self.base_sizes = [min(stride) for stride in self.strides
+                           ] if base_sizes is None else base_sizes
+        assert len(self.base_sizes) == len(self.strides), \
+            'The number of strides should be the same as base sizes, got ' \
+            f'{self.strides} and {self.base_sizes}'
+
+        # calculate scales of anchors
+        assert ((octave_base_scale is not None
+                and scales_per_octave is not None) ^ (scales is not None)), \
+            'scales and octave_base_scale with scales_per_octave cannot' \
+            ' be set at the same time'
+        if scales is not None:
+            self.scales = torch.Tensor(scales)
+        elif octave_base_scale is not None and scales_per_octave is not None:
+            octave_scales = np.array(
+                [2**(i / scales_per_octave) for i in range(scales_per_octave)])
+            scales = octave_scales * octave_base_scale
+            self.scales = torch.Tensor(scales)
+        else:
+            raise ValueError('Either scales or octave_base_scale with '
+                             'scales_per_octave should be set')
+
+        self.octave_base_scale = octave_base_scale
+        self.scales_per_octave = scales_per_octave
+        self.ratios = torch.Tensor(ratios)
+        self.scale_major = scale_major
+        self.centers = centers
+        self.center_offset = center_offset
+        self.base_anchors = self.gen_base_anchors()
+
+    @property
+    def num_base_anchors(self):
+        """list[int]: total number of base anchors in a feature grid"""
+        return [base_anchors.size(0) for base_anchors in self.base_anchors]
+
+    @property
+    def num_levels(self):
+        """int: number of feature levels that the generator will be applied"""
+        return len(self.strides)
+
+    def gen_base_anchors(self):
+        """Generate base anchors.
+
+        Returns:
+            list(torch.Tensor): Base anchors of a feature grid in multiple \
+                feature levels.
+        """
+        multi_level_base_anchors = []
+        for i, base_size in enumerate(self.base_sizes):
+            center = None
+            if self.centers is not None:
+                center = self.centers[i]
+            multi_level_base_anchors.append(
+                self.gen_single_level_base_anchors(
+                    base_size,
+                    scales=self.scales,
+                    ratios=self.ratios,
+                    center=center))
+        return multi_level_base_anchors
+
+    def gen_single_level_base_anchors(self,
+                                      base_size,
+                                      scales,
+                                      ratios,
+                                      center=None):
+        """Generate base anchors of a single level.
+
+        Args:
+            base_size (int | float): Basic size of an anchor.
+            scales (torch.Tensor): Scales of the anchor.
+            ratios (torch.Tensor): The ratio between between the height
+                and width of anchors in a single level.
+            center (tuple[float], optional): The center of the base anchor
+                related to a single feature grid. Defaults to None.
+
+        Returns:
+            torch.Tensor: Anchors in a single-level feature maps.
+        """
+        w = base_size
+        h = base_size
+        if center is None:
+            x_center = self.center_offset * w
+            y_center = self.center_offset * h
+        else:
+            x_center, y_center = center
+
+        h_ratios = torch.sqrt(ratios)
+        w_ratios = 1 / h_ratios
+        if self.scale_major:
+            ws = (w * w_ratios[:, None] * scales[None, :]).view(-1)
+            hs = (h * h_ratios[:, None] * scales[None, :]).view(-1)
+        else:
+            ws = (w * scales[:, None] * w_ratios[None, :]).view(-1)
+            hs = (h * scales[:, None] * h_ratios[None, :]).view(-1)
+
+        # use float anchor and the anchor's center is aligned with the
+        # pixel center
+        base_anchors = [
+            x_center - 0.5 * ws, y_center - 0.5 * hs, x_center + 0.5 * ws,
+            y_center + 0.5 * hs
+        ]
+        base_anchors = torch.stack(base_anchors, dim=-1)
+
+        return base_anchors
+
+    def _meshgrid(self, x, y, row_major=True):
+        """Generate mesh grid of x and y.
+
+        Args:
+            x (torch.Tensor): Grids of x dimension.
+            y (torch.Tensor): Grids of y dimension.
+            row_major (bool, optional): Whether to return y grids first.
+                Defaults to True.
+
+        Returns:
+            tuple[torch.Tensor]: The mesh grids of x and y.
+        """
+        xx = x.repeat(len(y))
+        yy = y.view(-1, 1).repeat(1, len(x)).view(-1)
+        if row_major:
+            return xx, yy
+        else:
+            return yy, xx
+
+    def grid_anchors(self, featmap_sizes, device='cuda'):
+        """Generate grid anchors in multiple feature levels.
+
+        Args:
+            featmap_sizes (list[tuple]): List of feature map sizes in
+                multiple feature levels.
+            device (str): Device where the anchors will be put on.
+
+        Return:
+            list[torch.Tensor]: Anchors in multiple feature levels. \
+                The sizes of each tensor should be [N, 4], where \
+                N = width * height * num_base_anchors, width and height \
+                are the sizes of the corresponding feature level, \
+                num_base_anchors is the number of anchors for that level.
+        """
+        assert self.num_levels == len(featmap_sizes)
+        multi_level_anchors = []
+        for i in range(self.num_levels):
+            anchors = self.single_level_grid_anchors(
+                self.base_anchors[i].to(device),
+                featmap_sizes[i],
+                self.strides[i],
+                device=device)
+            multi_level_anchors.append(anchors)
+        return multi_level_anchors
+
+    def single_level_grid_anchors(self,
+                                  base_anchors,
+                                  featmap_size,
+                                  stride=(16, 16),
+                                  device='cuda'):
+        """Generate grid anchors of a single level.
+
+        Note:
+            This function is usually called by method ``self.grid_anchors``.
+
+        Args:
+            base_anchors (torch.Tensor): The base anchors of a feature grid.
+            featmap_size (tuple[int]): Size of the feature maps.
+            stride (tuple[int], optional): Stride of the feature map in order
+                (w, h). Defaults to (16, 16).
+            device (str, optional): Device the tensor will be put on.
+                Defaults to 'cuda'.
+
+        Returns:
+            torch.Tensor: Anchors in the overall feature maps.
+        """
+        feat_h, feat_w = featmap_size
+        # convert Tensor to int, so that we can covert to ONNX correctlly
+        feat_h = int(feat_h)
+        feat_w = int(feat_w)
+        shift_x = torch.arange(0, feat_w, device=device) * stride[0]
+        shift_y = torch.arange(0, feat_h, device=device) * stride[1]
+
+        shift_xx, shift_yy = self._meshgrid(shift_x, shift_y)
+        shifts = torch.stack([shift_xx, shift_yy, shift_xx, shift_yy], dim=-1)
+        shifts = shifts.type_as(base_anchors)
+        # first feat_w elements correspond to the first row of shifts
+        # add A anchors (1, A, 4) to K shifts (K, 1, 4) to get
+        # shifted anchors (K, A, 4), reshape to (K*A, 4)
+
+        all_anchors = base_anchors[None, :, :] + shifts[:, None, :]
+        all_anchors = all_anchors.view(-1, 4)
+        # first A rows correspond to A anchors of (0, 0) in feature map,
+        # then (0, 1), (0, 2), ...
+        return all_anchors
+
+    def valid_flags(self, featmap_sizes, pad_shape, device='cuda'):
+        """Generate valid flags of anchors in multiple feature levels.
+
+        Args:
+            featmap_sizes (list(tuple)): List of feature map sizes in
+                multiple feature levels.
+            pad_shape (tuple): The padded shape of the image.
+            device (str): Device where the anchors will be put on.
+
+        Return:
+            list(torch.Tensor): Valid flags of anchors in multiple levels.
+        """
+        assert self.num_levels == len(featmap_sizes)
+        multi_level_flags = []
+        for i in range(self.num_levels):
+            anchor_stride = self.strides[i]
+            feat_h, feat_w = featmap_sizes[i]
+            h, w = pad_shape[:2]
+            valid_feat_h = min(int(np.ceil(h / anchor_stride[1])), feat_h)
+            valid_feat_w = min(int(np.ceil(w / anchor_stride[0])), feat_w)
+            flags = self.single_level_valid_flags((feat_h, feat_w),
+                                                  (valid_feat_h, valid_feat_w),
+                                                  self.num_base_anchors[i],
+                                                  device=device)
+            multi_level_flags.append(flags)
+        return multi_level_flags
+
+    def single_level_valid_flags(self,
+                                 featmap_size,
+                                 valid_size,
+                                 num_base_anchors,
+                                 device='cuda'):
+        """Generate the valid flags of anchor in a single feature map.
+
+        Args:
+            featmap_size (tuple[int]): The size of feature maps.
+            valid_size (tuple[int]): The valid size of the feature maps.
+            num_base_anchors (int): The number of base anchors.
+            device (str, optional): Device where the flags will be put on.
+                Defaults to 'cuda'.
+
+        Returns:
+            torch.Tensor: The valid flags of each anchor in a single level \
+                feature map.
+        """
+        feat_h, feat_w = featmap_size
+        valid_h, valid_w = valid_size
+        assert valid_h <= feat_h and valid_w <= feat_w
+        valid_x = torch.zeros(feat_w, dtype=torch.bool, device=device)
+        valid_y = torch.zeros(feat_h, dtype=torch.bool, device=device)
+        valid_x[:valid_w] = 1
+        valid_y[:valid_h] = 1
+        valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)
+        valid = valid_xx & valid_yy
+        valid = valid[:, None].expand(valid.size(0),
+                                      num_base_anchors).contiguous().view(-1)
+        return valid
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        indent_str = '    '
+        repr_str = self.__class__.__name__ + '(\n'
+        repr_str += f'{indent_str}strides={self.strides},\n'
+        repr_str += f'{indent_str}ratios={self.ratios},\n'
+        repr_str += f'{indent_str}scales={self.scales},\n'
+        repr_str += f'{indent_str}base_sizes={self.base_sizes},\n'
+        repr_str += f'{indent_str}scale_major={self.scale_major},\n'
+        repr_str += f'{indent_str}octave_base_scale='
+        repr_str += f'{self.octave_base_scale},\n'
+        repr_str += f'{indent_str}scales_per_octave='
+        repr_str += f'{self.scales_per_octave},\n'
+        repr_str += f'{indent_str}num_levels={self.num_levels}\n'
+        repr_str += f'{indent_str}centers={self.centers},\n'
+        repr_str += f'{indent_str}center_offset={self.center_offset})'
+        return repr_str
+
+
+@ANCHOR_GENERATORS.register_module()
+class SSDAnchorGenerator(AnchorGenerator):
+    """Anchor generator for SSD.
+
+    Args:
+        strides (list[int]  | list[tuple[int, int]]): Strides of anchors
+            in multiple feature levels.
+        ratios (list[float]): The list of ratios between the height and width
+            of anchors in a single level.
+        basesize_ratio_range (tuple(float)): Ratio range of anchors.
+        input_size (int): Size of feature map, 300 for SSD300,
+            512 for SSD512.
+        scale_major (bool): Whether to multiply scales first when generating
+            base anchors. If true, the anchors in the same row will have the
+            same scales. It is always set to be False in SSD.
+    """
+
+    def __init__(self,
+                 strides,
+                 ratios,
+                 basesize_ratio_range,
+                 input_size=300,
+                 scale_major=True):
+        assert len(strides) == len(ratios)
+        assert mmcv.is_tuple_of(basesize_ratio_range, float)
+
+        self.strides = [_pair(stride) for stride in strides]
+        self.input_size = input_size
+        self.centers = [(stride[0] / 2., stride[1] / 2.)
+                        for stride in self.strides]
+        self.basesize_ratio_range = basesize_ratio_range
+
+        # calculate anchor ratios and sizes
+        min_ratio, max_ratio = basesize_ratio_range
+        min_ratio = int(min_ratio * 100)
+        max_ratio = int(max_ratio * 100)
+        step = int(np.floor(max_ratio - min_ratio) / (self.num_levels - 2))
+        min_sizes = []
+        max_sizes = []
+        for ratio in range(int(min_ratio), int(max_ratio) + 1, step):
+            min_sizes.append(int(self.input_size * ratio / 100))
+            max_sizes.append(int(self.input_size * (ratio + step) / 100))
+        if self.input_size == 300:
+            if basesize_ratio_range[0] == 0.15:  # SSD300 COCO
+                min_sizes.insert(0, int(self.input_size * 7 / 100))
+                max_sizes.insert(0, int(self.input_size * 15 / 100))
+            elif basesize_ratio_range[0] == 0.2:  # SSD300 VOC
+                min_sizes.insert(0, int(self.input_size * 10 / 100))
+                max_sizes.insert(0, int(self.input_size * 20 / 100))
+            else:
+                raise ValueError(
+                    'basesize_ratio_range[0] should be either 0.15'
+                    'or 0.2 when input_size is 300, got '
+                    f'{basesize_ratio_range[0]}.')
+        elif self.input_size == 512:
+            if basesize_ratio_range[0] == 0.1:  # SSD512 COCO
+                min_sizes.insert(0, int(self.input_size * 4 / 100))
+                max_sizes.insert(0, int(self.input_size * 10 / 100))
+            elif basesize_ratio_range[0] == 0.15:  # SSD512 VOC
+                min_sizes.insert(0, int(self.input_size * 7 / 100))
+                max_sizes.insert(0, int(self.input_size * 15 / 100))
+            else:
+                raise ValueError('basesize_ratio_range[0] should be either 0.1'
+                                 'or 0.15 when input_size is 512, got'
+                                 f' {basesize_ratio_range[0]}.')
+        else:
+            raise ValueError('Only support 300 or 512 in SSDAnchorGenerator'
+                             f', got {self.input_size}.')
+
+        anchor_ratios = []
+        anchor_scales = []
+        for k in range(len(self.strides)):
+            scales = [1., np.sqrt(max_sizes[k] / min_sizes[k])]
+            anchor_ratio = [1.]
+            for r in ratios[k]:
+                anchor_ratio += [1 / r, r]  # 4 or 6 ratio
+            anchor_ratios.append(torch.Tensor(anchor_ratio))
+            anchor_scales.append(torch.Tensor(scales))
+
+        self.base_sizes = min_sizes
+        self.scales = anchor_scales
+        self.ratios = anchor_ratios
+        self.scale_major = scale_major
+        self.center_offset = 0
+        self.base_anchors = self.gen_base_anchors()
+
+    def gen_base_anchors(self):
+        """Generate base anchors.
+
+        Returns:
+            list(torch.Tensor): Base anchors of a feature grid in multiple \
+                feature levels.
+        """
+        multi_level_base_anchors = []
+        for i, base_size in enumerate(self.base_sizes):
+            base_anchors = self.gen_single_level_base_anchors(
+                base_size,
+                scales=self.scales[i],
+                ratios=self.ratios[i],
+                center=self.centers[i])
+            indices = list(range(len(self.ratios[i])))
+            indices.insert(1, len(indices))
+            base_anchors = torch.index_select(base_anchors, 0,
+                                              torch.LongTensor(indices))
+            multi_level_base_anchors.append(base_anchors)
+        return multi_level_base_anchors
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        indent_str = '    '
+        repr_str = self.__class__.__name__ + '(\n'
+        repr_str += f'{indent_str}strides={self.strides},\n'
+        repr_str += f'{indent_str}scales={self.scales},\n'
+        repr_str += f'{indent_str}scale_major={self.scale_major},\n'
+        repr_str += f'{indent_str}input_size={self.input_size},\n'
+        repr_str += f'{indent_str}scales={self.scales},\n'
+        repr_str += f'{indent_str}ratios={self.ratios},\n'
+        repr_str += f'{indent_str}num_levels={self.num_levels},\n'
+        repr_str += f'{indent_str}base_sizes={self.base_sizes},\n'
+        repr_str += f'{indent_str}basesize_ratio_range='
+        repr_str += f'{self.basesize_ratio_range})'
+        return repr_str
+
+
+@ANCHOR_GENERATORS.register_module()
+class LegacyAnchorGenerator(AnchorGenerator):
+    """Legacy anchor generator used in MMDetection V1.x.
+
+    Note:
+        Difference to the V2.0 anchor generator:
+
+        1. The center offset of V1.x anchors are set to be 0.5 rather than 0.
+        2. The width/height are minused by 1 when calculating the anchors' \
+            centers and corners to meet the V1.x coordinate system.
+        3. The anchors' corners are quantized.
+
+    Args:
+        strides (list[int] | list[tuple[int]]): Strides of anchors
+            in multiple feature levels.
+        ratios (list[float]): The list of ratios between the height and width
+            of anchors in a single level.
+        scales (list[int] | None): Anchor scales for anchors in a single level.
+            It cannot be set at the same time if `octave_base_scale` and
+            `scales_per_octave` are set.
+        base_sizes (list[int]): The basic sizes of anchors in multiple levels.
+            If None is given, strides will be used to generate base_sizes.
+        scale_major (bool): Whether to multiply scales first when generating
+            base anchors. If true, the anchors in the same row will have the
+            same scales. By default it is True in V2.0
+        octave_base_scale (int): The base scale of octave.
+        scales_per_octave (int): Number of scales for each octave.
+            `octave_base_scale` and `scales_per_octave` are usually used in
+            retinanet and the `scales` should be None when they are set.
+        centers (list[tuple[float, float]] | None): The centers of the anchor
+            relative to the feature grid center in multiple feature levels.
+            By default it is set to be None and not used. It a list of float
+            is given, this list will be used to shift the centers of anchors.
+        center_offset (float): The offset of center in propotion to anchors'
+            width and height. By default it is 0.5 in V2.0 but it should be 0.5
+            in v1.x models.
+
+    Examples:
+        >>> from mmdet.core import LegacyAnchorGenerator
+        >>> self = LegacyAnchorGenerator(
+        >>>     [16], [1.], [1.], [9], center_offset=0.5)
+        >>> all_anchors = self.grid_anchors(((2, 2),), device='cpu')
+        >>> print(all_anchors)
+        [tensor([[ 0.,  0.,  8.,  8.],
+                [16.,  0., 24.,  8.],
+                [ 0., 16.,  8., 24.],
+                [16., 16., 24., 24.]])]
+    """
+
+    def gen_single_level_base_anchors(self,
+                                      base_size,
+                                      scales,
+                                      ratios,
+                                      center=None):
+        """Generate base anchors of a single level.
+
+        Note:
+            The width/height of anchors are minused by 1 when calculating \
+                the centers and corners to meet the V1.x coordinate system.
+
+        Args:
+            base_size (int | float): Basic size of an anchor.
+            scales (torch.Tensor): Scales of the anchor.
+            ratios (torch.Tensor): The ratio between between the height.
+                and width of anchors in a single level.
+            center (tuple[float], optional): The center of the base anchor
+                related to a single feature grid. Defaults to None.
+
+        Returns:
+            torch.Tensor: Anchors in a single-level feature map.
+        """
+        w = base_size
+        h = base_size
+        if center is None:
+            x_center = self.center_offset * (w - 1)
+            y_center = self.center_offset * (h - 1)
+        else:
+            x_center, y_center = center
+
+        h_ratios = torch.sqrt(ratios)
+        w_ratios = 1 / h_ratios
+        if self.scale_major:
+            ws = (w * w_ratios[:, None] * scales[None, :]).view(-1)
+            hs = (h * h_ratios[:, None] * scales[None, :]).view(-1)
+        else:
+            ws = (w * scales[:, None] * w_ratios[None, :]).view(-1)
+            hs = (h * scales[:, None] * h_ratios[None, :]).view(-1)
+
+        # use float anchor and the anchor's center is aligned with the
+        # pixel center
+        base_anchors = [
+            x_center - 0.5 * (ws - 1), y_center - 0.5 * (hs - 1),
+            x_center + 0.5 * (ws - 1), y_center + 0.5 * (hs - 1)
+        ]
+        base_anchors = torch.stack(base_anchors, dim=-1).round()
+
+        return base_anchors
+
+
+@ANCHOR_GENERATORS.register_module()
+class LegacySSDAnchorGenerator(SSDAnchorGenerator, LegacyAnchorGenerator):
+    """Legacy anchor generator used in MMDetection V1.x.
+
+    The difference between `LegacySSDAnchorGenerator` and `SSDAnchorGenerator`
+    can be found in `LegacyAnchorGenerator`.
+    """
+
+    def __init__(self,
+                 strides,
+                 ratios,
+                 basesize_ratio_range,
+                 input_size=300,
+                 scale_major=True):
+        super(LegacySSDAnchorGenerator,
+              self).__init__(strides, ratios, basesize_ratio_range, input_size,
+                             scale_major)
+        self.centers = [((stride - 1) / 2., (stride - 1) / 2.)
+                        for stride in strides]
+        self.base_anchors = self.gen_base_anchors()
+
+
+@ANCHOR_GENERATORS.register_module()
+class YOLOAnchorGenerator(AnchorGenerator):
+    """Anchor generator for YOLO.
+
+    Args:
+        strides (list[int] | list[tuple[int, int]]): Strides of anchors
+            in multiple feature levels.
+        base_sizes (list[list[tuple[int, int]]]): The basic sizes
+            of anchors in multiple levels.
+    """
+
+    def __init__(self, strides, base_sizes):
+        self.strides = [_pair(stride) for stride in strides]
+        self.centers = [(stride[0] / 2., stride[1] / 2.)
+                        for stride in self.strides]
+        self.base_sizes = []
+        num_anchor_per_level = len(base_sizes[0])
+        for base_sizes_per_level in base_sizes:
+            assert num_anchor_per_level == len(base_sizes_per_level)
+            self.base_sizes.append(
+                [_pair(base_size) for base_size in base_sizes_per_level])
+        self.base_anchors = self.gen_base_anchors()
+
+    @property
+    def num_levels(self):
+        """int: number of feature levels that the generator will be applied"""
+        return len(self.base_sizes)
+
+    def gen_base_anchors(self):
+        """Generate base anchors.
+
+        Returns:
+            list(torch.Tensor): Base anchors of a feature grid in multiple \
+                feature levels.
+        """
+        multi_level_base_anchors = []
+        for i, base_sizes_per_level in enumerate(self.base_sizes):
+            center = None
+            if self.centers is not None:
+                center = self.centers[i]
+            multi_level_base_anchors.append(
+                self.gen_single_level_base_anchors(base_sizes_per_level,
+                                                   center))
+        return multi_level_base_anchors
+
+    def gen_single_level_base_anchors(self, base_sizes_per_level, center=None):
+        """Generate base anchors of a single level.
+
+        Args:
+            base_sizes_per_level (list[tuple[int, int]]): Basic sizes of
+                anchors.
+            center (tuple[float], optional): The center of the base anchor
+                related to a single feature grid. Defaults to None.
+
+        Returns:
+            torch.Tensor: Anchors in a single-level feature maps.
+        """
+        x_center, y_center = center
+        base_anchors = []
+        for base_size in base_sizes_per_level:
+            w, h = base_size
+
+            # use float anchor and the anchor's center is aligned with the
+            # pixel center
+            base_anchor = torch.Tensor([
+                x_center - 0.5 * w, y_center - 0.5 * h, x_center + 0.5 * w,
+                y_center + 0.5 * h
+            ])
+            base_anchors.append(base_anchor)
+        base_anchors = torch.stack(base_anchors, dim=0)
+
+        return base_anchors
+
+    def responsible_flags(self, featmap_sizes, gt_bboxes, device='cuda'):
+        """Generate responsible anchor flags of grid cells in multiple scales.
+
+        Args:
+            featmap_sizes (list(tuple)): List of feature map sizes in multiple
+                feature levels.
+            gt_bboxes (Tensor): Ground truth boxes, shape (n, 4).
+            device (str): Device where the anchors will be put on.
+
+        Return:
+            list(torch.Tensor): responsible flags of anchors in multiple level
+        """
+        assert self.num_levels == len(featmap_sizes)
+        multi_level_responsible_flags = []
+        for i in range(self.num_levels):
+            anchor_stride = self.strides[i]
+            flags = self.single_level_responsible_flags(
+                featmap_sizes[i],
+                gt_bboxes,
+                anchor_stride,
+                self.num_base_anchors[i],
+                device=device)
+            multi_level_responsible_flags.append(flags)
+        return multi_level_responsible_flags
+
+    def single_level_responsible_flags(self,
+                                       featmap_size,
+                                       gt_bboxes,
+                                       stride,
+                                       num_base_anchors,
+                                       device='cuda'):
+        """Generate the responsible flags of anchor in a single feature map.
+
+        Args:
+            featmap_size (tuple[int]): The size of feature maps.
+            gt_bboxes (Tensor): Ground truth boxes, shape (n, 4).
+            stride (tuple(int)): stride of current level
+            num_base_anchors (int): The number of base anchors.
+            device (str, optional): Device where the flags will be put on.
+                Defaults to 'cuda'.
+
+        Returns:
+            torch.Tensor: The valid flags of each anchor in a single level \
+                feature map.
+        """
+        feat_h, feat_w = featmap_size
+        gt_bboxes_cx = ((gt_bboxes[:, 0] + gt_bboxes[:, 2]) * 0.5).to(device)
+        gt_bboxes_cy = ((gt_bboxes[:, 1] + gt_bboxes[:, 3]) * 0.5).to(device)
+        gt_bboxes_grid_x = torch.floor(gt_bboxes_cx / stride[0]).long()
+        gt_bboxes_grid_y = torch.floor(gt_bboxes_cy / stride[1]).long()
+
+        # row major indexing
+        gt_bboxes_grid_idx = gt_bboxes_grid_y * feat_w + gt_bboxes_grid_x
+
+        responsible_grid = torch.zeros(
+            feat_h * feat_w, dtype=torch.uint8, device=device)
+        responsible_grid[gt_bboxes_grid_idx] = 1
+
+        responsible_grid = responsible_grid[:, None].expand(
+            responsible_grid.size(0), num_base_anchors).contiguous().view(-1)
+        return responsible_grid
diff --git a/detection/scrfd/mmdet/core/anchor/builder.py b/detection/scrfd/mmdet/core/anchor/builder.py
new file mode 100755
index 000000000..d79b448eb
--- /dev/null
+++ b/detection/scrfd/mmdet/core/anchor/builder.py
@@ -0,0 +1,7 @@
+from mmcv.utils import Registry, build_from_cfg
+
+ANCHOR_GENERATORS = Registry('Anchor generator')
+
+
+def build_anchor_generator(cfg, default_args=None):
+    return build_from_cfg(cfg, ANCHOR_GENERATORS, default_args)
diff --git a/detection/scrfd/mmdet/core/anchor/point_generator.py b/detection/scrfd/mmdet/core/anchor/point_generator.py
new file mode 100755
index 000000000..e6fbd988c
--- /dev/null
+++ b/detection/scrfd/mmdet/core/anchor/point_generator.py
@@ -0,0 +1,37 @@
+import torch
+
+from .builder import ANCHOR_GENERATORS
+
+
+@ANCHOR_GENERATORS.register_module()
+class PointGenerator(object):
+
+    def _meshgrid(self, x, y, row_major=True):
+        xx = x.repeat(len(y))
+        yy = y.view(-1, 1).repeat(1, len(x)).view(-1)
+        if row_major:
+            return xx, yy
+        else:
+            return yy, xx
+
+    def grid_points(self, featmap_size, stride=16, device='cuda'):
+        feat_h, feat_w = featmap_size
+        shift_x = torch.arange(0., feat_w, device=device) * stride
+        shift_y = torch.arange(0., feat_h, device=device) * stride
+        shift_xx, shift_yy = self._meshgrid(shift_x, shift_y)
+        stride = shift_x.new_full((shift_xx.shape[0], ), stride)
+        shifts = torch.stack([shift_xx, shift_yy, stride], dim=-1)
+        all_points = shifts.to(device)
+        return all_points
+
+    def valid_flags(self, featmap_size, valid_size, device='cuda'):
+        feat_h, feat_w = featmap_size
+        valid_h, valid_w = valid_size
+        assert valid_h <= feat_h and valid_w <= feat_w
+        valid_x = torch.zeros(feat_w, dtype=torch.bool, device=device)
+        valid_y = torch.zeros(feat_h, dtype=torch.bool, device=device)
+        valid_x[:valid_w] = 1
+        valid_y[:valid_h] = 1
+        valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)
+        valid = valid_xx & valid_yy
+        return valid
diff --git a/detection/scrfd/mmdet/core/anchor/utils.py b/detection/scrfd/mmdet/core/anchor/utils.py
new file mode 100755
index 000000000..ab9b53f37
--- /dev/null
+++ b/detection/scrfd/mmdet/core/anchor/utils.py
@@ -0,0 +1,71 @@
+import torch
+
+
+def images_to_levels(target, num_levels):
+    """Convert targets by image to targets by feature level.
+
+    [target_img0, target_img1] -> [target_level0, target_level1, ...]
+    """
+    target = torch.stack(target, 0)
+    level_targets = []
+    start = 0
+    for n in num_levels:
+        end = start + n
+        # level_targets.append(target[:, start:end].squeeze(0))
+        level_targets.append(target[:, start:end])
+        start = end
+    return level_targets
+
+
+def anchor_inside_flags(flat_anchors,
+                        valid_flags,
+                        img_shape,
+                        allowed_border=0):
+    """Check whether the anchors are inside the border.
+
+    Args:
+        flat_anchors (torch.Tensor): Flatten anchors, shape (n, 4).
+        valid_flags (torch.Tensor): An existing valid flags of anchors.
+        img_shape (tuple(int)): Shape of current image.
+        allowed_border (int, optional): The border to allow the valid anchor.
+            Defaults to 0.
+
+    Returns:
+        torch.Tensor: Flags indicating whether the anchors are inside a \
+            valid range.
+    """
+    img_h, img_w = img_shape[:2]
+    if allowed_border >= 0:
+        inside_flags = valid_flags & \
+            (flat_anchors[:, 0] >= -allowed_border) & \
+            (flat_anchors[:, 1] >= -allowed_border) & \
+            (flat_anchors[:, 2] < img_w + allowed_border) & \
+            (flat_anchors[:, 3] < img_h + allowed_border)
+    else:
+        inside_flags = valid_flags
+    return inside_flags
+
+
+def calc_region(bbox, ratio, featmap_size=None):
+    """Calculate a proportional bbox region.
+
+    The bbox center are fixed and the new h' and w' is h * ratio and w * ratio.
+
+    Args:
+        bbox (Tensor): Bboxes to calculate regions, shape (n, 4).
+        ratio (float): Ratio of the output region.
+        featmap_size (tuple): Feature map size used for clipping the boundary.
+
+    Returns:
+        tuple: x1, y1, x2, y2
+    """
+    x1 = torch.round((1 - ratio) * bbox[0] + ratio * bbox[2]).long()
+    y1 = torch.round((1 - ratio) * bbox[1] + ratio * bbox[3]).long()
+    x2 = torch.round(ratio * bbox[0] + (1 - ratio) * bbox[2]).long()
+    y2 = torch.round(ratio * bbox[1] + (1 - ratio) * bbox[3]).long()
+    if featmap_size is not None:
+        x1 = x1.clamp(min=0, max=featmap_size[1])
+        y1 = y1.clamp(min=0, max=featmap_size[0])
+        x2 = x2.clamp(min=0, max=featmap_size[1])
+        y2 = y2.clamp(min=0, max=featmap_size[0])
+    return (x1, y1, x2, y2)
diff --git a/detection/scrfd/mmdet/core/bbox/__init__.py b/detection/scrfd/mmdet/core/bbox/__init__.py
new file mode 100755
index 000000000..28a1b9f95
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/__init__.py
@@ -0,0 +1,27 @@
+from .assigners import (AssignResult, BaseAssigner, CenterRegionAssigner,
+                        MaxIoUAssigner)
+from .builder import build_assigner, build_bbox_coder, build_sampler
+from .coder import (BaseBBoxCoder, DeltaXYWHBBoxCoder, PseudoBBoxCoder,
+                    TBLRBBoxCoder)
+from .iou_calculators import BboxOverlaps2D, bbox_overlaps
+from .samplers import (BaseSampler, CombinedSampler,
+                       InstanceBalancedPosSampler, IoUBalancedNegSampler,
+                       OHEMSampler, PseudoSampler, RandomSampler,
+                       SamplingResult, ScoreHLRSampler)
+from .transforms import (bbox2distance, bbox2result, bbox2roi, kps2distance,
+                         bbox_cxcywh_to_xyxy, bbox_flip, bbox_mapping,
+                         bbox_mapping_back, bbox_rescale, bbox_xyxy_to_cxcywh,
+                         distance2bbox, distance2kps, roi2bbox)
+
+__all__ = [
+    'bbox_overlaps', 'BboxOverlaps2D', 'BaseAssigner', 'MaxIoUAssigner',
+    'AssignResult', 'BaseSampler', 'PseudoSampler', 'RandomSampler',
+    'InstanceBalancedPosSampler', 'IoUBalancedNegSampler', 'CombinedSampler',
+    'OHEMSampler', 'SamplingResult', 'ScoreHLRSampler', 'build_assigner',
+    'build_sampler', 'bbox_flip', 'bbox_mapping', 'bbox_mapping_back',
+    'bbox2roi', 'roi2bbox', 'bbox2result', 
+    'distance2bbox', 'bbox2distance', 'distance2kps', 'kps2distance',
+    'build_bbox_coder', 'BaseBBoxCoder', 'PseudoBBoxCoder',
+    'DeltaXYWHBBoxCoder', 'TBLRBBoxCoder', 'CenterRegionAssigner',
+    'bbox_rescale', 'bbox_cxcywh_to_xyxy', 'bbox_xyxy_to_cxcywh'
+]
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/__init__.py b/detection/scrfd/mmdet/core/bbox/assigners/__init__.py
new file mode 100755
index 000000000..b8f0f48d8
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/__init__.py
@@ -0,0 +1,15 @@
+from .approx_max_iou_assigner import ApproxMaxIoUAssigner
+from .assign_result import AssignResult
+from .atss_assigner import ATSSAssigner
+from .base_assigner import BaseAssigner
+from .center_region_assigner import CenterRegionAssigner
+from .grid_assigner import GridAssigner
+from .hungarian_assigner import HungarianAssigner
+from .max_iou_assigner import MaxIoUAssigner
+from .point_assigner import PointAssigner
+
+__all__ = [
+    'BaseAssigner', 'MaxIoUAssigner', 'ApproxMaxIoUAssigner', 'AssignResult',
+    'PointAssigner', 'ATSSAssigner', 'CenterRegionAssigner', 'GridAssigner',
+    'HungarianAssigner'
+]
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/approx_max_iou_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/approx_max_iou_assigner.py
new file mode 100755
index 000000000..6d07656d1
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/approx_max_iou_assigner.py
@@ -0,0 +1,145 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from .max_iou_assigner import MaxIoUAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class ApproxMaxIoUAssigner(MaxIoUAssigner):
+    """Assign a corresponding gt bbox or background to each bbox.
+
+    Each proposals will be assigned with an integer indicating the ground truth
+     index. (semi-positive index: gt label (0-based), -1: background)
+
+    - -1: negative sample, no assigned gt
+    - semi-positive integer: positive sample, index (0-based) of assigned gt
+
+    Args:
+        pos_iou_thr (float): IoU threshold for positive bboxes.
+        neg_iou_thr (float or tuple): IoU threshold for negative bboxes.
+        min_pos_iou (float): Minimum iou for a bbox to be considered as a
+            positive bbox. Positive samples can have smaller IoU than
+            pos_iou_thr due to the 4th step (assign max IoU sample to each gt).
+        gt_max_assign_all (bool): Whether to assign all bboxes with the same
+            highest overlap with some gt to that gt.
+        ignore_iof_thr (float): IoF threshold for ignoring bboxes (if
+            `gt_bboxes_ignore` is specified). Negative values mean not
+            ignoring any bboxes.
+        ignore_wrt_candidates (bool): Whether to compute the iof between
+            `bboxes` and `gt_bboxes_ignore`, or the contrary.
+        match_low_quality (bool): Whether to allow quality matches. This is
+            usually allowed for RPN and single stage detectors, but not allowed
+            in the second stage.
+        gpu_assign_thr (int): The upper bound of the number of GT for GPU
+            assign. When the number of gt is above this threshold, will assign
+            on CPU device. Negative values mean not assign on CPU.
+    """
+
+    def __init__(self,
+                 pos_iou_thr,
+                 neg_iou_thr,
+                 min_pos_iou=.0,
+                 gt_max_assign_all=True,
+                 ignore_iof_thr=-1,
+                 ignore_wrt_candidates=True,
+                 match_low_quality=True,
+                 gpu_assign_thr=-1,
+                 iou_calculator=dict(type='BboxOverlaps2D')):
+        self.pos_iou_thr = pos_iou_thr
+        self.neg_iou_thr = neg_iou_thr
+        self.min_pos_iou = min_pos_iou
+        self.gt_max_assign_all = gt_max_assign_all
+        self.ignore_iof_thr = ignore_iof_thr
+        self.ignore_wrt_candidates = ignore_wrt_candidates
+        self.gpu_assign_thr = gpu_assign_thr
+        self.match_low_quality = match_low_quality
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+
+    def assign(self,
+               approxs,
+               squares,
+               approxs_per_octave,
+               gt_bboxes,
+               gt_bboxes_ignore=None,
+               gt_labels=None):
+        """Assign gt to approxs.
+
+        This method assign a gt bbox to each group of approxs (bboxes),
+        each group of approxs is represent by a base approx (bbox) and
+        will be assigned with -1, or a semi-positive number.
+        background_label (-1) means negative sample,
+        semi-positive number is the index (0-based) of assigned gt.
+        The assignment is done in following steps, the order matters.
+
+        1. assign every bbox to background_label (-1)
+        2. use the max IoU of each group of approxs to assign
+        2. assign proposals whose iou with all gts < neg_iou_thr to background
+        3. for each bbox, if the iou with its nearest gt >= pos_iou_thr,
+           assign it to that bbox
+        4. for each gt bbox, assign its nearest proposals (may be more than
+           one) to itself
+
+        Args:
+            approxs (Tensor): Bounding boxes to be assigned,
+                shape(approxs_per_octave*n, 4).
+            squares (Tensor): Base Bounding boxes to be assigned,
+                shape(n, 4).
+            approxs_per_octave (int): number of approxs per octave
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
+                labelled as `ignored`, e.g., crowd boxes in COCO.
+            gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+        """
+        num_squares = squares.size(0)
+        num_gts = gt_bboxes.size(0)
+
+        if num_squares == 0 or num_gts == 0:
+            # No predictions and/or truth, return empty assignment
+            overlaps = approxs.new(num_gts, num_squares)
+            assign_result = self.assign_wrt_overlaps(overlaps, gt_labels)
+            return assign_result
+
+        # re-organize anchors by approxs_per_octave x num_squares
+        approxs = torch.transpose(
+            approxs.view(num_squares, approxs_per_octave, 4), 0,
+            1).contiguous().view(-1, 4)
+        assign_on_cpu = True if (self.gpu_assign_thr > 0) and (
+            num_gts > self.gpu_assign_thr) else False
+        # compute overlap and assign gt on CPU when number of GT is large
+        if assign_on_cpu:
+            device = approxs.device
+            approxs = approxs.cpu()
+            gt_bboxes = gt_bboxes.cpu()
+            if gt_bboxes_ignore is not None:
+                gt_bboxes_ignore = gt_bboxes_ignore.cpu()
+            if gt_labels is not None:
+                gt_labels = gt_labels.cpu()
+        all_overlaps = self.iou_calculator(approxs, gt_bboxes)
+
+        overlaps, _ = all_overlaps.view(approxs_per_octave, num_squares,
+                                        num_gts).max(dim=0)
+        overlaps = torch.transpose(overlaps, 0, 1)
+
+        if (self.ignore_iof_thr > 0 and gt_bboxes_ignore is not None
+                and gt_bboxes_ignore.numel() > 0 and squares.numel() > 0):
+            if self.ignore_wrt_candidates:
+                ignore_overlaps = self.iou_calculator(
+                    squares, gt_bboxes_ignore, mode='iof')
+                ignore_max_overlaps, _ = ignore_overlaps.max(dim=1)
+            else:
+                ignore_overlaps = self.iou_calculator(
+                    gt_bboxes_ignore, squares, mode='iof')
+                ignore_max_overlaps, _ = ignore_overlaps.max(dim=0)
+            overlaps[:, ignore_max_overlaps > self.ignore_iof_thr] = -1
+
+        assign_result = self.assign_wrt_overlaps(overlaps, gt_labels)
+        if assign_on_cpu:
+            assign_result.gt_inds = assign_result.gt_inds.to(device)
+            assign_result.max_overlaps = assign_result.max_overlaps.to(device)
+            if assign_result.labels is not None:
+                assign_result.labels = assign_result.labels.to(device)
+        return assign_result
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/assign_result.py b/detection/scrfd/mmdet/core/bbox/assigners/assign_result.py
new file mode 100755
index 000000000..4639fbdba
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/assign_result.py
@@ -0,0 +1,204 @@
+import torch
+
+from mmdet.utils import util_mixins
+
+
+class AssignResult(util_mixins.NiceRepr):
+    """Stores assignments between predicted and truth boxes.
+
+    Attributes:
+        num_gts (int): the number of truth boxes considered when computing this
+            assignment
+
+        gt_inds (LongTensor): for each predicted box indicates the 1-based
+            index of the assigned truth box. 0 means unassigned and -1 means
+            ignore.
+
+        max_overlaps (FloatTensor): the iou between the predicted box and its
+            assigned truth box.
+
+        labels (None | LongTensor): If specified, for each predicted box
+            indicates the category label of the assigned truth box.
+
+    Example:
+        >>> # An assign result between 4 predicted boxes and 9 true boxes
+        >>> # where only two boxes were assigned.
+        >>> num_gts = 9
+        >>> max_overlaps = torch.LongTensor([0, .5, .9, 0])
+        >>> gt_inds = torch.LongTensor([-1, 1, 2, 0])
+        >>> labels = torch.LongTensor([0, 3, 4, 0])
+        >>> self = AssignResult(num_gts, gt_inds, max_overlaps, labels)
+        >>> print(str(self))  # xdoctest: +IGNORE_WANT
+        <AssignResult(num_gts=9, gt_inds.shape=(4,), max_overlaps.shape=(4,),
+                      labels.shape=(4,))>
+        >>> # Force addition of gt labels (when adding gt as proposals)
+        >>> new_labels = torch.LongTensor([3, 4, 5])
+        >>> self.add_gt_(new_labels)
+        >>> print(str(self))  # xdoctest: +IGNORE_WANT
+        <AssignResult(num_gts=9, gt_inds.shape=(7,), max_overlaps.shape=(7,),
+                      labels.shape=(7,))>
+    """
+
+    def __init__(self, num_gts, gt_inds, max_overlaps, labels=None):
+        self.num_gts = num_gts
+        self.gt_inds = gt_inds
+        self.max_overlaps = max_overlaps
+        self.labels = labels
+        # Interface for possible user-defined properties
+        self._extra_properties = {}
+
+    @property
+    def num_preds(self):
+        """int: the number of predictions in this assignment"""
+        return len(self.gt_inds)
+
+    def set_extra_property(self, key, value):
+        """Set user-defined new property."""
+        assert key not in self.info
+        self._extra_properties[key] = value
+
+    def get_extra_property(self, key):
+        """Get user-defined property."""
+        return self._extra_properties.get(key, None)
+
+    @property
+    def info(self):
+        """dict: a dictionary of info about the object"""
+        basic_info = {
+            'num_gts': self.num_gts,
+            'num_preds': self.num_preds,
+            'gt_inds': self.gt_inds,
+            'max_overlaps': self.max_overlaps,
+            'labels': self.labels,
+        }
+        basic_info.update(self._extra_properties)
+        return basic_info
+
+    def __nice__(self):
+        """str: a "nice" summary string describing this assign result"""
+        parts = []
+        parts.append(f'num_gts={self.num_gts!r}')
+        if self.gt_inds is None:
+            parts.append(f'gt_inds={self.gt_inds!r}')
+        else:
+            parts.append(f'gt_inds.shape={tuple(self.gt_inds.shape)!r}')
+        if self.max_overlaps is None:
+            parts.append(f'max_overlaps={self.max_overlaps!r}')
+        else:
+            parts.append('max_overlaps.shape='
+                         f'{tuple(self.max_overlaps.shape)!r}')
+        if self.labels is None:
+            parts.append(f'labels={self.labels!r}')
+        else:
+            parts.append(f'labels.shape={tuple(self.labels.shape)!r}')
+        return ', '.join(parts)
+
+    @classmethod
+    def random(cls, **kwargs):
+        """Create random AssignResult for tests or debugging.
+
+        Args:
+            num_preds: number of predicted boxes
+            num_gts: number of true boxes
+            p_ignore (float): probability of a predicted box assinged to an
+                ignored truth
+            p_assigned (float): probability of a predicted box not being
+                assigned
+            p_use_label (float | bool): with labels or not
+            rng (None | int | numpy.random.RandomState): seed or state
+
+        Returns:
+            :obj:`AssignResult`: Randomly generated assign results.
+
+        Example:
+            >>> from mmdet.core.bbox.assigners.assign_result import *  # NOQA
+            >>> self = AssignResult.random()
+            >>> print(self.info)
+        """
+        from mmdet.core.bbox import demodata
+        rng = demodata.ensure_rng(kwargs.get('rng', None))
+
+        num_gts = kwargs.get('num_gts', None)
+        num_preds = kwargs.get('num_preds', None)
+        p_ignore = kwargs.get('p_ignore', 0.3)
+        p_assigned = kwargs.get('p_assigned', 0.7)
+        p_use_label = kwargs.get('p_use_label', 0.5)
+        num_classes = kwargs.get('p_use_label', 3)
+
+        if num_gts is None:
+            num_gts = rng.randint(0, 8)
+        if num_preds is None:
+            num_preds = rng.randint(0, 16)
+
+        if num_gts == 0:
+            max_overlaps = torch.zeros(num_preds, dtype=torch.float32)
+            gt_inds = torch.zeros(num_preds, dtype=torch.int64)
+            if p_use_label is True or p_use_label < rng.rand():
+                labels = torch.zeros(num_preds, dtype=torch.int64)
+            else:
+                labels = None
+        else:
+            import numpy as np
+            # Create an overlap for each predicted box
+            max_overlaps = torch.from_numpy(rng.rand(num_preds))
+
+            # Construct gt_inds for each predicted box
+            is_assigned = torch.from_numpy(rng.rand(num_preds) < p_assigned)
+            # maximum number of assignments constraints
+            n_assigned = min(num_preds, min(num_gts, is_assigned.sum()))
+
+            assigned_idxs = np.where(is_assigned)[0]
+            rng.shuffle(assigned_idxs)
+            assigned_idxs = assigned_idxs[0:n_assigned]
+            assigned_idxs.sort()
+
+            is_assigned[:] = 0
+            is_assigned[assigned_idxs] = True
+
+            is_ignore = torch.from_numpy(
+                rng.rand(num_preds) < p_ignore) & is_assigned
+
+            gt_inds = torch.zeros(num_preds, dtype=torch.int64)
+
+            true_idxs = np.arange(num_gts)
+            rng.shuffle(true_idxs)
+            true_idxs = torch.from_numpy(true_idxs)
+            gt_inds[is_assigned] = true_idxs[:n_assigned]
+
+            gt_inds = torch.from_numpy(
+                rng.randint(1, num_gts + 1, size=num_preds))
+            gt_inds[is_ignore] = -1
+            gt_inds[~is_assigned] = 0
+            max_overlaps[~is_assigned] = 0
+
+            if p_use_label is True or p_use_label < rng.rand():
+                if num_classes == 0:
+                    labels = torch.zeros(num_preds, dtype=torch.int64)
+                else:
+                    labels = torch.from_numpy(
+                        # remind that we set FG labels to [0, num_class-1]
+                        # since mmdet v2.0
+                        # BG cat_id: num_class
+                        rng.randint(0, num_classes, size=num_preds))
+                    labels[~is_assigned] = 0
+            else:
+                labels = None
+
+        self = cls(num_gts, gt_inds, max_overlaps, labels)
+        return self
+
+    def add_gt_(self, gt_labels):
+        """Add ground truth as assigned results.
+
+        Args:
+            gt_labels (torch.Tensor): Labels of gt boxes
+        """
+        self_inds = torch.arange(
+            1, len(gt_labels) + 1, dtype=torch.long, device=gt_labels.device)
+        self.gt_inds = torch.cat([self_inds, self.gt_inds])
+
+        self.max_overlaps = torch.cat(
+            [self.max_overlaps.new_ones(len(gt_labels)), self.max_overlaps])
+
+        if self.labels is not None:
+            self.labels = torch.cat([gt_labels, self.labels])
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/atss_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/atss_assigner.py
new file mode 100755
index 000000000..d41a2f2cc
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/atss_assigner.py
@@ -0,0 +1,215 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class ATSSAssigner(BaseAssigner):
+    """Assign a corresponding gt bbox or background to each bbox.
+
+    Each proposals will be assigned with `0` or a positive integer
+    indicating the ground truth index.
+
+    - 0: negative sample, no assigned gt
+    - positive integer: positive sample, index (1-based) of assigned gt
+
+    Args:
+        topk (float): number of bbox selected in each level
+    """
+
+    def __init__(self,
+                 topk,
+                 mode=0,
+                 iou_calculator=dict(type='BboxOverlaps2D'),
+                 ignore_iof_thr=-1):
+        self.topk = topk
+        self.mode = mode
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+        self.ignore_iof_thr = ignore_iof_thr
+
+    # https://github.com/sfzhang15/ATSS/blob/master/atss_core/modeling/rpn/atss/loss.py
+
+    def assign(self,
+               bboxes,
+               num_level_bboxes,
+               gt_bboxes,
+               gt_bboxes_ignore=None,
+               gt_labels=None):
+        """Assign gt to bboxes.
+
+        The assignment is done in following steps
+
+        1. compute iou between all bbox (bbox of all pyramid levels) and gt
+        2. compute center distance between all bbox and gt
+        3. on each pyramid level, for each gt, select k bbox whose center
+           are closest to the gt center, so we total select k*l bbox as
+           candidates for each gt
+        4. get corresponding iou for the these candidates, and compute the
+           mean and std, set mean + std as the iou threshold
+        5. select these candidates whose iou are greater than or equal to
+           the threshold as postive
+        6. limit the positive sample's center in gt
+
+
+        Args:
+            bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).
+            num_level_bboxes (List): num of bboxes in each level
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
+                labelled as `ignored`, e.g., crowd boxes in COCO.
+            gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+        """
+        INF = 100000000
+        bboxes = bboxes[:, :4]
+        num_gt, num_bboxes = gt_bboxes.size(0), bboxes.size(0)
+        #print('AT1:', num_gt, num_bboxes)
+
+        # compute iou between all bbox and gt
+        overlaps = self.iou_calculator(bboxes, gt_bboxes)
+
+        # assign 0 by default
+        assigned_gt_inds = overlaps.new_full((num_bboxes, ),
+                                             0,
+                                             dtype=torch.long)
+
+        if num_gt == 0 or num_bboxes == 0:
+            # No ground truth or boxes, return empty assignment
+            max_overlaps = overlaps.new_zeros((num_bboxes, ))
+            if num_gt == 0:
+                # No truth, assign everything to background
+                assigned_gt_inds[:] = 0
+            if gt_labels is None:
+                assigned_labels = None
+            else:
+                assigned_labels = overlaps.new_full((num_bboxes, ),
+                                                    -1,
+                                                    dtype=torch.long)
+            return AssignResult(
+                num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)
+
+        # compute center distance between all bbox and gt
+        gt_cx = (gt_bboxes[:, 0] + gt_bboxes[:, 2]) / 2.0
+        gt_cy = (gt_bboxes[:, 1] + gt_bboxes[:, 3]) / 2.0
+        gt_points = torch.stack((gt_cx, gt_cy), dim=1)
+
+        gt_width = gt_bboxes[:,2] - gt_bboxes[:,0]
+        gt_height = gt_bboxes[:,3] - gt_bboxes[:,1]
+        gt_area = torch.sqrt( torch.clamp(gt_width*gt_height, min=1e-4) )
+
+        bboxes_cx = (bboxes[:, 0] + bboxes[:, 2]) / 2.0
+        bboxes_cy = (bboxes[:, 1] + bboxes[:, 3]) / 2.0
+        bboxes_points = torch.stack((bboxes_cx, bboxes_cy), dim=1)
+
+        distances = (bboxes_points[:, None, :] -
+                     gt_points[None, :, :]).pow(2).sum(-1).sqrt()
+
+        if (self.ignore_iof_thr > 0 and gt_bboxes_ignore is not None
+                and gt_bboxes_ignore.numel() > 0 and bboxes.numel() > 0):
+            ignore_overlaps = self.iou_calculator(
+                bboxes, gt_bboxes_ignore, mode='iof')
+            ignore_max_overlaps, _ = ignore_overlaps.max(dim=1)
+            ignore_idxs = ignore_max_overlaps > self.ignore_iof_thr
+            distances[ignore_idxs, :] = INF
+            assigned_gt_inds[ignore_idxs] = -1
+
+        # Selecting candidates based on the center distance
+        candidate_idxs = []
+        start_idx = 0
+        for level, bboxes_per_level in enumerate(num_level_bboxes):
+            # on each pyramid level, for each gt,
+            # select k bbox whose center are closest to the gt center
+            end_idx = start_idx + bboxes_per_level
+            distances_per_level = distances[start_idx:end_idx, :] #(A,G)
+            selectable_k = min(self.topk, bboxes_per_level)
+            _, topk_idxs_per_level = distances_per_level.topk(
+                selectable_k, dim=0, largest=False)
+            #print('AT-LEVEL:', start_idx, end_idx, bboxes_per_level, topk_idxs_per_level.shape)
+            candidate_idxs.append(topk_idxs_per_level + start_idx)
+            start_idx = end_idx
+        candidate_idxs = torch.cat(candidate_idxs, dim=0)# candidate anchors (topk*num_level_bboxes, G) = (AK, G)
+
+        # get corresponding iou for the these candidates, and compute the
+        # mean and std, set mean + std as the iou threshold
+        candidate_overlaps = overlaps[candidate_idxs, torch.arange(num_gt)] #(AK,G)
+        overlaps_mean_per_gt = candidate_overlaps.mean(0)
+        overlaps_std_per_gt = candidate_overlaps.std(0)
+        overlaps_thr_per_gt = overlaps_mean_per_gt + overlaps_std_per_gt
+
+        is_pos = candidate_overlaps >= overlaps_thr_per_gt[None, :]
+        #print('CAND:', candidate_idxs.shape, candidate_overlaps.shape, is_pos.shape)
+        #print('BOXES:', bboxes_cx.shape)
+
+        # limit the positive sample's center in gt
+        for gt_idx in range(num_gt):
+            candidate_idxs[:, gt_idx] += gt_idx * num_bboxes
+        ep_bboxes_cx = bboxes_cx.view(1, -1).expand(
+            num_gt, num_bboxes).contiguous().view(-1)
+        ep_bboxes_cy = bboxes_cy.view(1, -1).expand(
+            num_gt, num_bboxes).contiguous().view(-1)
+        candidate_idxs = candidate_idxs.view(-1)
+
+        # calculate the left, top, right, bottom distance between positive
+        # bbox center and gt side
+        l_ = ep_bboxes_cx[candidate_idxs].view(-1, num_gt) - gt_bboxes[:, 0]
+        t_ = ep_bboxes_cy[candidate_idxs].view(-1, num_gt) - gt_bboxes[:, 1]
+        r_ = gt_bboxes[:, 2] - ep_bboxes_cx[candidate_idxs].view(-1, num_gt)
+        b_ = gt_bboxes[:, 3] - ep_bboxes_cy[candidate_idxs].view(-1, num_gt)
+        #is_in_gts = torch.stack([l_, t_, r_, b_], dim=1).min(dim=1)[0] > 0.01
+        dist_min = torch.stack([l_, t_, r_, b_], dim=1).min(dim=1)[0] # (A,G)
+        dist_min.div_(gt_area)
+        #print('ATTT:', l_.shape, t_.shape, dist_min.shape, self.mode)
+        if self.mode==0:
+            is_in_gts = dist_min > 0.001
+        elif self.mode==1:
+            is_in_gts = dist_min > -0.25
+        elif self.mode==2:
+            is_in_gts = dist_min > -0.15
+            #dist_expand = torch.clamp(gt_area / 16.0, min=1.0, max=3.0)
+            #dist_min.mul_(dist_expand)
+            #is_in_gts = dist_min > -0.25
+        elif self.mode==3:
+            dist_expand = torch.clamp(gt_area / 16.0, min=1.0, max=6.0)
+            dist_min.mul_(dist_expand)
+            is_in_gts = dist_min > -0.2
+        elif self.mode==4:
+            dist_expand = torch.clamp(gt_area / 16.0, min=0.5, max=6.0)
+            dist_min.mul_(dist_expand)
+            is_in_gts = dist_min > -0.2
+        elif self.mode==5:
+            dist_div = torch.clamp(gt_area / 16.0, min=0.5, max=3.0)
+            dist_min.div_(dist_div)
+            is_in_gts = dist_min > -0.2
+        else:
+            raise ValueError
+        #print(gt_area.shape, is_in_gts.shape, is_pos.shape)
+        is_pos = is_pos & is_in_gts
+
+        # if an anchor box is assigned to multiple gts,
+        # the one with the highest IoU will be selected.
+        overlaps_inf = torch.full_like(overlaps,
+                                       -INF).t().contiguous().view(-1)
+        index = candidate_idxs.view(-1)[is_pos.view(-1)]
+        overlaps_inf[index] = overlaps.t().contiguous().view(-1)[index]
+        overlaps_inf = overlaps_inf.view(num_gt, -1).t()
+
+        max_overlaps, argmax_overlaps = overlaps_inf.max(dim=1)
+        assigned_gt_inds[
+            max_overlaps != -INF] = argmax_overlaps[max_overlaps != -INF] + 1
+
+        if gt_labels is not None:
+            assigned_labels = assigned_gt_inds.new_full((num_bboxes, ), -1)
+            pos_inds = torch.nonzero(
+                assigned_gt_inds > 0, as_tuple=False).squeeze()
+            if pos_inds.numel() > 0:
+                assigned_labels[pos_inds] = gt_labels[
+                    assigned_gt_inds[pos_inds] - 1]
+        else:
+            assigned_labels = None
+        return AssignResult(
+            num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/base_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/base_assigner.py
new file mode 100755
index 000000000..2da9e0f4a
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/base_assigner.py
@@ -0,0 +1,10 @@
+from abc import ABCMeta, abstractmethod
+
+
+class BaseAssigner(metaclass=ABCMeta):
+    """Base assigner that assigns boxes to ground truth boxes."""
+
+    @abstractmethod
+    def assign(self, bboxes, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None):
+        """Assign boxes to either a ground truth boxe or a negative boxes."""
+        pass
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/center_region_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/center_region_assigner.py
new file mode 100755
index 000000000..488e3b615
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/center_region_assigner.py
@@ -0,0 +1,335 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+def scale_boxes(bboxes, scale):
+    """Expand an array of boxes by a given scale.
+
+    Args:
+        bboxes (Tensor): Shape (m, 4)
+        scale (float): The scale factor of bboxes
+
+    Returns:
+        (Tensor): Shape (m, 4). Scaled bboxes
+    """
+    assert bboxes.size(1) == 4
+    w_half = (bboxes[:, 2] - bboxes[:, 0]) * .5
+    h_half = (bboxes[:, 3] - bboxes[:, 1]) * .5
+    x_c = (bboxes[:, 2] + bboxes[:, 0]) * .5
+    y_c = (bboxes[:, 3] + bboxes[:, 1]) * .5
+
+    w_half *= scale
+    h_half *= scale
+
+    boxes_scaled = torch.zeros_like(bboxes)
+    boxes_scaled[:, 0] = x_c - w_half
+    boxes_scaled[:, 2] = x_c + w_half
+    boxes_scaled[:, 1] = y_c - h_half
+    boxes_scaled[:, 3] = y_c + h_half
+    return boxes_scaled
+
+
+def is_located_in(points, bboxes):
+    """Are points located in bboxes.
+
+    Args:
+      points (Tensor): Points, shape: (m, 2).
+      bboxes (Tensor): Bounding boxes, shape: (n, 4).
+
+    Return:
+      Tensor: Flags indicating if points are located in bboxes, shape: (m, n).
+    """
+    assert points.size(1) == 2
+    assert bboxes.size(1) == 4
+    return (points[:, 0].unsqueeze(1) > bboxes[:, 0].unsqueeze(0)) & \
+           (points[:, 0].unsqueeze(1) < bboxes[:, 2].unsqueeze(0)) & \
+           (points[:, 1].unsqueeze(1) > bboxes[:, 1].unsqueeze(0)) & \
+           (points[:, 1].unsqueeze(1) < bboxes[:, 3].unsqueeze(0))
+
+
+def bboxes_area(bboxes):
+    """Compute the area of an array of bboxes.
+
+    Args:
+        bboxes (Tensor): The coordinates ox bboxes. Shape: (m, 4)
+
+    Returns:
+        Tensor: Area of the bboxes. Shape: (m, )
+    """
+    assert bboxes.size(1) == 4
+    w = (bboxes[:, 2] - bboxes[:, 0])
+    h = (bboxes[:, 3] - bboxes[:, 1])
+    areas = w * h
+    return areas
+
+
+@BBOX_ASSIGNERS.register_module()
+class CenterRegionAssigner(BaseAssigner):
+    """Assign pixels at the center region of a bbox as positive.
+
+    Each proposals will be assigned with `-1`, `0`, or a positive integer
+    indicating the ground truth index.
+    - -1: negative samples
+    - semi-positive numbers: positive sample, index (0-based) of assigned gt
+
+    Args:
+        pos_scale (float): Threshold within which pixels are
+          labelled as positive.
+        neg_scale (float): Threshold above which pixels are
+          labelled as positive.
+        min_pos_iof (float): Minimum iof of a pixel with a gt to be
+          labelled as positive. Default: 1e-2
+        ignore_gt_scale (float): Threshold within which the pixels
+          are ignored when the gt is labelled as shadowed. Default: 0.5
+        foreground_dominate (bool): If True, the bbox will be assigned as
+          positive when a gt's kernel region overlaps with another's shadowed
+          (ignored) region, otherwise it is set as ignored. Default to False.
+    """
+
+    def __init__(self,
+                 pos_scale,
+                 neg_scale,
+                 min_pos_iof=1e-2,
+                 ignore_gt_scale=0.5,
+                 foreground_dominate=False,
+                 iou_calculator=dict(type='BboxOverlaps2D')):
+        self.pos_scale = pos_scale
+        self.neg_scale = neg_scale
+        self.min_pos_iof = min_pos_iof
+        self.ignore_gt_scale = ignore_gt_scale
+        self.foreground_dominate = foreground_dominate
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+
+    def get_gt_priorities(self, gt_bboxes):
+        """Get gt priorities according to their areas.
+
+        Smaller gt has higher priority.
+
+        Args:
+            gt_bboxes (Tensor): Ground truth boxes, shape (k, 4).
+
+        Returns:
+            Tensor: The priority of gts so that gts with larger priority is \
+              more likely to be assigned. Shape (k, )
+        """
+        gt_areas = bboxes_area(gt_bboxes)
+        # Rank all gt bbox areas. Smaller objects has larger priority
+        _, sort_idx = gt_areas.sort(descending=True)
+        sort_idx = sort_idx.argsort()
+        return sort_idx
+
+    def assign(self, bboxes, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None):
+        """Assign gt to bboxes.
+
+        This method assigns gts to every bbox (proposal/anchor), each bbox \
+        will be assigned with -1, or a semi-positive number. -1 means \
+        negative sample, semi-positive number is the index (0-based) of \
+        assigned gt.
+
+        Args:
+            bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_bboxes_ignore (tensor, optional): Ground truth bboxes that are
+              labelled as `ignored`, e.g., crowd boxes in COCO.
+            gt_labels (tensor, optional): Label of gt_bboxes, shape (num_gts,).
+
+        Returns:
+            :obj:`AssignResult`: The assigned result. Note that \
+              shadowed_labels of shape (N, 2) is also added as an \
+              `assign_result` attribute. `shadowed_labels` is a tensor \
+              composed of N pairs of anchor_ind, class_label], where N \
+              is the number of anchors that lie in the outer region of a \
+              gt, anchor_ind is the shadowed anchor index and class_label \
+              is the shadowed class label.
+
+        Example:
+            >>> self = CenterRegionAssigner(0.2, 0.2)
+            >>> bboxes = torch.Tensor([[0, 0, 10, 10], [10, 10, 20, 20]])
+            >>> gt_bboxes = torch.Tensor([[0, 0, 10, 10]])
+            >>> assign_result = self.assign(bboxes, gt_bboxes)
+            >>> expected_gt_inds = torch.LongTensor([1, 0])
+            >>> assert torch.all(assign_result.gt_inds == expected_gt_inds)
+        """
+        # There are in total 5 steps in the pixel assignment
+        # 1. Find core (the center region, say inner 0.2)
+        #     and shadow (the relatively ourter part, say inner 0.2-0.5)
+        #     regions of every gt.
+        # 2. Find all prior bboxes that lie in gt_core and gt_shadow regions
+        # 3. Assign prior bboxes in gt_core with a one-hot id of the gt in
+        #      the image.
+        #    3.1. For overlapping objects, the prior bboxes in gt_core is
+        #           assigned with the object with smallest area
+        # 4. Assign prior bboxes with class label according to its gt id.
+        #    4.1. Assign -1 to prior bboxes lying in shadowed gts
+        #    4.2. Assign positive prior boxes with the corresponding label
+        # 5. Find pixels lying in the shadow of an object and assign them with
+        #      background label, but set the loss weight of its corresponding
+        #      gt to zero.
+        assert bboxes.size(1) == 4, 'bboxes must have size of 4'
+        # 1. Find core positive and shadow region of every gt
+        gt_core = scale_boxes(gt_bboxes, self.pos_scale)
+        gt_shadow = scale_boxes(gt_bboxes, self.neg_scale)
+
+        # 2. Find prior bboxes that lie in gt_core and gt_shadow regions
+        bbox_centers = (bboxes[:, 2:4] + bboxes[:, 0:2]) / 2
+        # The center points lie within the gt boxes
+        is_bbox_in_gt = is_located_in(bbox_centers, gt_bboxes)
+        # Only calculate bbox and gt_core IoF. This enables small prior bboxes
+        #   to match large gts
+        bbox_and_gt_core_overlaps = self.iou_calculator(
+            bboxes, gt_core, mode='iof')
+        # The center point of effective priors should be within the gt box
+        is_bbox_in_gt_core = is_bbox_in_gt & (
+            bbox_and_gt_core_overlaps > self.min_pos_iof)  # shape (n, k)
+
+        is_bbox_in_gt_shadow = (
+            self.iou_calculator(bboxes, gt_shadow, mode='iof') >
+            self.min_pos_iof)
+        # Rule out center effective positive pixels
+        is_bbox_in_gt_shadow &= (~is_bbox_in_gt_core)
+
+        num_gts, num_bboxes = gt_bboxes.size(0), bboxes.size(0)
+        if num_gts == 0 or num_bboxes == 0:
+            # If no gts exist, assign all pixels to negative
+            assigned_gt_ids = \
+                is_bbox_in_gt_core.new_zeros((num_bboxes,),
+                                             dtype=torch.long)
+            pixels_in_gt_shadow = assigned_gt_ids.new_empty((0, 2))
+        else:
+            # Step 3: assign a one-hot gt id to each pixel, and smaller objects
+            #    have high priority to assign the pixel.
+            sort_idx = self.get_gt_priorities(gt_bboxes)
+            assigned_gt_ids, pixels_in_gt_shadow = \
+                self.assign_one_hot_gt_indices(is_bbox_in_gt_core,
+                                               is_bbox_in_gt_shadow,
+                                               gt_priority=sort_idx)
+
+        if gt_bboxes_ignore is not None and gt_bboxes_ignore.numel() > 0:
+            # No ground truth or boxes, return empty assignment
+            gt_bboxes_ignore = scale_boxes(
+                gt_bboxes_ignore, scale=self.ignore_gt_scale)
+            is_bbox_in_ignored_gts = is_located_in(bbox_centers,
+                                                   gt_bboxes_ignore)
+            is_bbox_in_ignored_gts = is_bbox_in_ignored_gts.any(dim=1)
+            assigned_gt_ids[is_bbox_in_ignored_gts] = -1
+
+        # 4. Assign prior bboxes with class label according to its gt id.
+        assigned_labels = None
+        shadowed_pixel_labels = None
+        if gt_labels is not None:
+            # Default assigned label is the background (-1)
+            assigned_labels = assigned_gt_ids.new_full((num_bboxes, ), -1)
+            pos_inds = torch.nonzero(
+                assigned_gt_ids > 0, as_tuple=False).squeeze()
+            if pos_inds.numel() > 0:
+                assigned_labels[pos_inds] = gt_labels[assigned_gt_ids[pos_inds]
+                                                      - 1]
+            # 5. Find pixels lying in the shadow of an object
+            shadowed_pixel_labels = pixels_in_gt_shadow.clone()
+            if pixels_in_gt_shadow.numel() > 0:
+                pixel_idx, gt_idx =\
+                    pixels_in_gt_shadow[:, 0], pixels_in_gt_shadow[:, 1]
+                assert (assigned_gt_ids[pixel_idx] != gt_idx).all(), \
+                    'Some pixels are dually assigned to ignore and gt!'
+                shadowed_pixel_labels[:, 1] = gt_labels[gt_idx - 1]
+                override = (
+                    assigned_labels[pixel_idx] == shadowed_pixel_labels[:, 1])
+                if self.foreground_dominate:
+                    # When a pixel is both positive and shadowed, set it as pos
+                    shadowed_pixel_labels = shadowed_pixel_labels[~override]
+                else:
+                    # When a pixel is both pos and shadowed, set it as shadowed
+                    assigned_labels[pixel_idx[override]] = -1
+                    assigned_gt_ids[pixel_idx[override]] = 0
+
+        assign_result = AssignResult(
+            num_gts, assigned_gt_ids, None, labels=assigned_labels)
+        # Add shadowed_labels as assign_result property. Shape: (num_shadow, 2)
+        assign_result.set_extra_property('shadowed_labels',
+                                         shadowed_pixel_labels)
+        return assign_result
+
+    def assign_one_hot_gt_indices(self,
+                                  is_bbox_in_gt_core,
+                                  is_bbox_in_gt_shadow,
+                                  gt_priority=None):
+        """Assign only one gt index to each prior box.
+
+        Gts with large gt_priority are more likely to be assigned.
+
+        Args:
+            is_bbox_in_gt_core (Tensor): Bool tensor indicating the bbox center
+              is in the core area of a gt (e.g. 0-0.2).
+              Shape: (num_prior, num_gt).
+            is_bbox_in_gt_shadow (Tensor): Bool tensor indicating the bbox
+              center is in the shadowed area of a gt (e.g. 0.2-0.5).
+              Shape: (num_prior, num_gt).
+            gt_priority (Tensor): Priorities of gts. The gt with a higher
+              priority is more likely to be assigned to the bbox when the bbox
+              match with multiple gts. Shape: (num_gt, ).
+
+        Returns:
+            tuple: Returns (assigned_gt_inds, shadowed_gt_inds).
+
+                - assigned_gt_inds: The assigned gt index of each prior bbox \
+                    (i.e. index from 1 to num_gts). Shape: (num_prior, ).
+                - shadowed_gt_inds: shadowed gt indices. It is a tensor of \
+                    shape (num_ignore, 2) with first column being the \
+                    shadowed prior bbox indices and the second column the \
+                    shadowed gt indices (1-based).
+        """
+        num_bboxes, num_gts = is_bbox_in_gt_core.shape
+
+        if gt_priority is None:
+            gt_priority = torch.arange(
+                num_gts, device=is_bbox_in_gt_core.device)
+        assert gt_priority.size(0) == num_gts
+        # The bigger gt_priority, the more preferable to be assigned
+        # The assigned inds are by default 0 (background)
+        assigned_gt_inds = is_bbox_in_gt_core.new_zeros((num_bboxes, ),
+                                                        dtype=torch.long)
+        # Shadowed bboxes are assigned to be background. But the corresponding
+        #   label is ignored during loss calculation, which is done through
+        #   shadowed_gt_inds
+        shadowed_gt_inds = torch.nonzero(is_bbox_in_gt_shadow, as_tuple=False)
+        if is_bbox_in_gt_core.sum() == 0:  # No gt match
+            shadowed_gt_inds[:, 1] += 1  # 1-based. For consistency issue
+            return assigned_gt_inds, shadowed_gt_inds
+
+        # The priority of each prior box and gt pair. If one prior box is
+        #  matched bo multiple gts. Only the pair with the highest priority
+        #  is saved
+        pair_priority = is_bbox_in_gt_core.new_full((num_bboxes, num_gts),
+                                                    -1,
+                                                    dtype=torch.long)
+
+        # Each bbox could match with multiple gts.
+        # The following codes deal with this situation
+        # Matched  bboxes (to any gt). Shape: (num_pos_anchor, )
+        inds_of_match = torch.any(is_bbox_in_gt_core, dim=1)
+        # The matched gt index of each positive bbox. Length >= num_pos_anchor
+        #   , since one bbox could match multiple gts
+        matched_bbox_gt_inds = torch.nonzero(
+            is_bbox_in_gt_core, as_tuple=False)[:, 1]
+        # Assign priority to each bbox-gt pair.
+        pair_priority[is_bbox_in_gt_core] = gt_priority[matched_bbox_gt_inds]
+        _, argmax_priority = pair_priority[inds_of_match].max(dim=1)
+        assigned_gt_inds[inds_of_match] = argmax_priority + 1  # 1-based
+        # Zero-out the assigned anchor box to filter the shadowed gt indices
+        is_bbox_in_gt_core[inds_of_match, argmax_priority] = 0
+        # Concat the shadowed indices due to overlapping with that out side of
+        #   effective scale. shape: (total_num_ignore, 2)
+        shadowed_gt_inds = torch.cat(
+            (shadowed_gt_inds, torch.nonzero(
+                is_bbox_in_gt_core, as_tuple=False)),
+            dim=0)
+        # `is_bbox_in_gt_core` should be changed back to keep arguments intact.
+        is_bbox_in_gt_core[inds_of_match, argmax_priority] = 1
+        # 1-based shadowed gt indices, to be consistent with `assigned_gt_inds`
+        if shadowed_gt_inds.numel() > 0:
+            shadowed_gt_inds[:, 1] += 1
+        return assigned_gt_inds, shadowed_gt_inds
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/grid_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/grid_assigner.py
new file mode 100755
index 000000000..7390ea637
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/grid_assigner.py
@@ -0,0 +1,155 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class GridAssigner(BaseAssigner):
+    """Assign a corresponding gt bbox or background to each bbox.
+
+    Each proposals will be assigned with `-1`, `0`, or a positive integer
+    indicating the ground truth index.
+
+    - -1: don't care
+    - 0: negative sample, no assigned gt
+    - positive integer: positive sample, index (1-based) of assigned gt
+
+    Args:
+        pos_iou_thr (float): IoU threshold for positive bboxes.
+        neg_iou_thr (float or tuple): IoU threshold for negative bboxes.
+        min_pos_iou (float): Minimum iou for a bbox to be considered as a
+            positive bbox. Positive samples can have smaller IoU than
+            pos_iou_thr due to the 4th step (assign max IoU sample to each gt).
+        gt_max_assign_all (bool): Whether to assign all bboxes with the same
+            highest overlap with some gt to that gt.
+    """
+
+    def __init__(self,
+                 pos_iou_thr,
+                 neg_iou_thr,
+                 min_pos_iou=.0,
+                 gt_max_assign_all=True,
+                 iou_calculator=dict(type='BboxOverlaps2D')):
+        self.pos_iou_thr = pos_iou_thr
+        self.neg_iou_thr = neg_iou_thr
+        self.min_pos_iou = min_pos_iou
+        self.gt_max_assign_all = gt_max_assign_all
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+
+    def assign(self, bboxes, box_responsible_flags, gt_bboxes, gt_labels=None):
+        """Assign gt to bboxes. The process is very much like the max iou
+        assigner, except that positive samples are constrained within the cell
+        that the gt boxes fell in.
+
+        This method assign a gt bbox to every bbox (proposal/anchor), each bbox
+        will be assigned with -1, 0, or a positive number. -1 means don't care,
+        0 means negative sample, positive number is the index (1-based) of
+        assigned gt.
+        The assignment is done in following steps, the order matters.
+
+        1. assign every bbox to -1
+        2. assign proposals whose iou with all gts <= neg_iou_thr to 0
+        3. for each bbox within a cell, if the iou with its nearest gt >
+            pos_iou_thr and the center of that gt falls inside the cell,
+            assign it to that bbox
+        4. for each gt bbox, assign its nearest proposals within the cell the
+            gt bbox falls in to itself.
+
+        Args:
+            bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).
+            box_responsible_flags (Tensor): flag to indicate whether box is
+                responsible for prediction, shape(n, )
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+        """
+        num_gts, num_bboxes = gt_bboxes.size(0), bboxes.size(0)
+
+        # compute iou between all gt and bboxes
+        overlaps = self.iou_calculator(gt_bboxes, bboxes)
+
+        # 1. assign -1 by default
+        assigned_gt_inds = overlaps.new_full((num_bboxes, ),
+                                             -1,
+                                             dtype=torch.long)
+
+        if num_gts == 0 or num_bboxes == 0:
+            # No ground truth or boxes, return empty assignment
+            max_overlaps = overlaps.new_zeros((num_bboxes, ))
+            if num_gts == 0:
+                # No truth, assign everything to background
+                assigned_gt_inds[:] = 0
+            if gt_labels is None:
+                assigned_labels = None
+            else:
+                assigned_labels = overlaps.new_full((num_bboxes, ),
+                                                    -1,
+                                                    dtype=torch.long)
+            return AssignResult(
+                num_gts,
+                assigned_gt_inds,
+                max_overlaps,
+                labels=assigned_labels)
+
+        # 2. assign negative: below
+        # for each anchor, which gt best overlaps with it
+        # for each anchor, the max iou of all gts
+        # shape of max_overlaps == argmax_overlaps == num_bboxes
+        max_overlaps, argmax_overlaps = overlaps.max(dim=0)
+
+        if isinstance(self.neg_iou_thr, float):
+            assigned_gt_inds[(max_overlaps >= 0)
+                             & (max_overlaps <= self.neg_iou_thr)] = 0
+        elif isinstance(self.neg_iou_thr, (tuple, list)):
+            assert len(self.neg_iou_thr) == 2
+            assigned_gt_inds[(max_overlaps > self.neg_iou_thr[0])
+                             & (max_overlaps <= self.neg_iou_thr[1])] = 0
+
+        # 3. assign positive: falls into responsible cell and above
+        # positive IOU threshold, the order matters.
+        # the prior condition of comparision is to filter out all
+        # unrelated anchors, i.e. not box_responsible_flags
+        overlaps[:, ~box_responsible_flags.type(torch.bool)] = -1.
+
+        # calculate max_overlaps again, but this time we only consider IOUs
+        # for anchors responsible for prediction
+        max_overlaps, argmax_overlaps = overlaps.max(dim=0)
+
+        # for each gt, which anchor best overlaps with it
+        # for each gt, the max iou of all proposals
+        # shape of gt_max_overlaps == gt_argmax_overlaps == num_gts
+        gt_max_overlaps, gt_argmax_overlaps = overlaps.max(dim=1)
+
+        pos_inds = (max_overlaps >
+                    self.pos_iou_thr) & box_responsible_flags.type(torch.bool)
+        assigned_gt_inds[pos_inds] = argmax_overlaps[pos_inds] + 1
+
+        # 4. assign positive to max overlapped anchors within responsible cell
+        for i in range(num_gts):
+            if gt_max_overlaps[i] > self.min_pos_iou:
+                if self.gt_max_assign_all:
+                    max_iou_inds = (overlaps[i, :] == gt_max_overlaps[i]) & \
+                         box_responsible_flags.type(torch.bool)
+                    assigned_gt_inds[max_iou_inds] = i + 1
+                elif box_responsible_flags[gt_argmax_overlaps[i]]:
+                    assigned_gt_inds[gt_argmax_overlaps[i]] = i + 1
+
+        # assign labels of positive anchors
+        if gt_labels is not None:
+            assigned_labels = assigned_gt_inds.new_full((num_bboxes, ), -1)
+            pos_inds = torch.nonzero(
+                assigned_gt_inds > 0, as_tuple=False).squeeze()
+            if pos_inds.numel() > 0:
+                assigned_labels[pos_inds] = gt_labels[
+                    assigned_gt_inds[pos_inds] - 1]
+
+        else:
+            assigned_labels = None
+
+        return AssignResult(
+            num_gts, assigned_gt_inds, max_overlaps, labels=assigned_labels)
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/hungarian_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/hungarian_assigner.py
new file mode 100755
index 000000000..224609300
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/hungarian_assigner.py
@@ -0,0 +1,158 @@
+import torch
+from scipy.optimize import linear_sum_assignment
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from ..transforms import bbox_cxcywh_to_xyxy, bbox_xyxy_to_cxcywh
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class HungarianAssigner(BaseAssigner):
+    """Computes one-to-one matching between predictions and ground truth.
+
+    This class computes an assignment between the targets and the predictions
+    based on the costs. The costs are weighted sum of three components:
+    classfication cost, regression L1 cost and regression iou cost. The
+    targets don't include the no_object, so generally there are more
+    predictions than targets. After the one-to-one matching, the un-matched
+    are treated as backgrounds. Thus each query prediction will be assigned
+    with `0` or a positive integer indicating the ground truth index:
+
+    - 0: negative sample, no assigned gt
+    - positive integer: positive sample, index (1-based) of assigned gt
+
+    Args:
+        cls_weight (int | float, optional): The scale factor for classification
+            cost. Default 1.0.
+        bbox_weight (int | float, optional): The scale factor for regression
+            L1 cost. Default 1.0.
+        iou_weight (int | float, optional): The scale factor for regression
+            iou cost. Default 1.0.
+        iou_calculator (dict | optional): The config for the iou calculation.
+            Default type `BboxOverlaps2D`.
+        iou_mode (str | optional): "iou" (intersection over union), "iof"
+                (intersection over foreground), or "giou" (generalized
+                intersection over union). Default "giou".
+    """
+
+    def __init__(self,
+                 cls_weight=1.,
+                 bbox_weight=1.,
+                 iou_weight=1.,
+                 iou_calculator=dict(type='BboxOverlaps2D'),
+                 iou_mode='giou'):
+        # defaultly giou cost is used in the official DETR repo.
+        self.iou_mode = iou_mode
+        self.cls_weight = cls_weight
+        self.bbox_weight = bbox_weight
+        self.iou_weight = iou_weight
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+
+    def assign(self,
+               bbox_pred,
+               cls_pred,
+               gt_bboxes,
+               gt_labels,
+               img_meta,
+               gt_bboxes_ignore=None,
+               eps=1e-7):
+        """Computes one-to-one matching based on the weighted costs.
+
+        This method assign each query prediction to a ground truth or
+        background. The `assigned_gt_inds` with -1 means don't care,
+        0 means negative sample, and positive number is the index (1-based)
+        of assigned gt.
+        The assignment is done in the following steps, the order matters.
+
+        1. assign every prediction to -1
+        2. compute the weighted costs
+        3. do Hungarian matching on CPU based on the costs
+        4. assign all to 0 (background) first, then for each matched pair
+           between predictions and gts, treat this prediction as foreground
+           and assign the corresponding gt index (plus 1) to it.
+
+        Args:
+            bbox_pred (Tensor): Predicted boxes with normalized coordinates
+                (cx, cy, w, h), which are all in range [0, 1]. Shape
+                [num_query, 4].
+            cls_pred (Tensor): Predicted classification logits, shape
+                [num_query, num_class].
+            gt_bboxes (Tensor): Ground truth boxes with unnormalized
+                coordinates (x1, y1, x2, y2). Shape [num_gt, 4].
+            gt_labels (Tensor): Label of `gt_bboxes`, shape (num_gt,).
+            img_meta (dict): Meta information for current image.
+            gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
+                labelled as `ignored`. Default None.
+            eps (int | float, optional): A value added to the denominator for
+                numerical stability. Default 1e-7.
+
+        Returns:
+            :obj:`AssignResult`: The assigned result.
+        """
+        assert gt_bboxes_ignore is None, \
+            'Only case when gt_bboxes_ignore is None is supported.'
+        num_gts, num_bboxes = gt_bboxes.size(0), bbox_pred.size(0)
+
+        # 1. assign -1 by default
+        assigned_gt_inds = bbox_pred.new_full((num_bboxes, ),
+                                              -1,
+                                              dtype=torch.long)
+        assigned_labels = bbox_pred.new_full((num_bboxes, ),
+                                             -1,
+                                             dtype=torch.long)
+        if num_gts == 0 or num_bboxes == 0:
+            # No ground truth or boxes, return empty assignment
+            if num_gts == 0:
+                # No ground truth, assign all to background
+                assigned_gt_inds[:] = 0
+            return AssignResult(
+                num_gts, assigned_gt_inds, None, labels=assigned_labels)
+
+        # 2. compute the weighted costs
+        # classification cost.
+        # Following the official DETR repo, contrary to the loss that
+        # NLL is used, we approximate it in 1 - cls_score[gt_label].
+        # The 1 is a constant that doesn't change the matching,
+        # so it can be ommitted.
+        cls_score = cls_pred.softmax(-1)
+        cls_cost = -cls_score[:, gt_labels]  # [num_bboxes, num_gt]
+
+        # regression L1 cost
+        img_h, img_w, _ = img_meta['img_shape']
+        factor = torch.Tensor([img_w, img_h, img_w,
+                               img_h]).unsqueeze(0).to(gt_bboxes.device)
+        gt_bboxes_normalized = gt_bboxes / factor
+        bbox_cost = torch.cdist(
+            bbox_pred, bbox_xyxy_to_cxcywh(gt_bboxes_normalized),
+            p=1)  # [num_bboxes, num_gt]
+
+        # regression iou cost, defaultly giou is used in official DETR.
+        bboxes = bbox_cxcywh_to_xyxy(bbox_pred) * factor
+        # overlaps: [num_bboxes, num_gt]
+        overlaps = self.iou_calculator(
+            bboxes, gt_bboxes, mode=self.iou_mode, is_aligned=False)
+        # The 1 is a constant that doesn't change the matching, so ommitted.
+        iou_cost = -overlaps
+
+        # weighted sum of above three costs
+        cost = self.cls_weight * cls_cost + self.bbox_weight * bbox_cost
+        cost = cost + self.iou_weight * iou_cost
+
+        # 3. do Hungarian matching on CPU using linear_sum_assignment
+        cost = cost.detach().cpu()
+        matched_row_inds, matched_col_inds = linear_sum_assignment(cost)
+        matched_row_inds = torch.from_numpy(matched_row_inds).to(
+            bbox_pred.device)
+        matched_col_inds = torch.from_numpy(matched_col_inds).to(
+            bbox_pred.device)
+
+        # 4. assign backgrounds and foregrounds
+        # assign all indices to backgrounds first
+        assigned_gt_inds[:] = 0
+        # assign foregrounds based on matching results
+        assigned_gt_inds[matched_row_inds] = matched_col_inds + 1
+        assigned_labels[matched_row_inds] = gt_labels[matched_col_inds]
+        return AssignResult(
+            num_gts, assigned_gt_inds, None, labels=assigned_labels)
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/max_iou_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/max_iou_assigner.py
new file mode 100755
index 000000000..a99f77e10
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/max_iou_assigner.py
@@ -0,0 +1,212 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from ..iou_calculators import build_iou_calculator
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class MaxIoUAssigner(BaseAssigner):
+    """Assign a corresponding gt bbox or background to each bbox.
+
+    Each proposals will be assigned with `-1`, or a semi-positive integer
+    indicating the ground truth index.
+
+    - -1: negative sample, no assigned gt
+    - semi-positive integer: positive sample, index (0-based) of assigned gt
+
+    Args:
+        pos_iou_thr (float): IoU threshold for positive bboxes.
+        neg_iou_thr (float or tuple): IoU threshold for negative bboxes.
+        min_pos_iou (float): Minimum iou for a bbox to be considered as a
+            positive bbox. Positive samples can have smaller IoU than
+            pos_iou_thr due to the 4th step (assign max IoU sample to each gt).
+        gt_max_assign_all (bool): Whether to assign all bboxes with the same
+            highest overlap with some gt to that gt.
+        ignore_iof_thr (float): IoF threshold for ignoring bboxes (if
+            `gt_bboxes_ignore` is specified). Negative values mean not
+            ignoring any bboxes.
+        ignore_wrt_candidates (bool): Whether to compute the iof between
+            `bboxes` and `gt_bboxes_ignore`, or the contrary.
+        match_low_quality (bool): Whether to allow low quality matches. This is
+            usually allowed for RPN and single stage detectors, but not allowed
+            in the second stage. Details are demonstrated in Step 4.
+        gpu_assign_thr (int): The upper bound of the number of GT for GPU
+            assign. When the number of gt is above this threshold, will assign
+            on CPU device. Negative values mean not assign on CPU.
+    """
+
+    def __init__(self,
+                 pos_iou_thr,
+                 neg_iou_thr,
+                 min_pos_iou=.0,
+                 gt_max_assign_all=True,
+                 ignore_iof_thr=-1,
+                 ignore_wrt_candidates=True,
+                 match_low_quality=True,
+                 gpu_assign_thr=-1,
+                 iou_calculator=dict(type='BboxOverlaps2D')):
+        self.pos_iou_thr = pos_iou_thr
+        self.neg_iou_thr = neg_iou_thr
+        self.min_pos_iou = min_pos_iou
+        self.gt_max_assign_all = gt_max_assign_all
+        self.ignore_iof_thr = ignore_iof_thr
+        self.ignore_wrt_candidates = ignore_wrt_candidates
+        self.gpu_assign_thr = gpu_assign_thr
+        self.match_low_quality = match_low_quality
+        self.iou_calculator = build_iou_calculator(iou_calculator)
+
+    def assign(self, bboxes, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None):
+        """Assign gt to bboxes.
+
+        This method assign a gt bbox to every bbox (proposal/anchor), each bbox
+        will be assigned with -1, or a semi-positive number. -1 means negative
+        sample, semi-positive number is the index (0-based) of assigned gt.
+        The assignment is done in following steps, the order matters.
+
+        1. assign every bbox to the background
+        2. assign proposals whose iou with all gts < neg_iou_thr to 0
+        3. for each bbox, if the iou with its nearest gt >= pos_iou_thr,
+           assign it to that bbox
+        4. for each gt bbox, assign its nearest proposals (may be more than
+           one) to itself
+
+        Args:
+            bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
+                labelled as `ignored`, e.g., crowd boxes in COCO.
+            gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+
+        Example:
+            >>> self = MaxIoUAssigner(0.5, 0.5)
+            >>> bboxes = torch.Tensor([[0, 0, 10, 10], [10, 10, 20, 20]])
+            >>> gt_bboxes = torch.Tensor([[0, 0, 10, 9]])
+            >>> assign_result = self.assign(bboxes, gt_bboxes)
+            >>> expected_gt_inds = torch.LongTensor([1, 0])
+            >>> assert torch.all(assign_result.gt_inds == expected_gt_inds)
+        """
+        assign_on_cpu = True if (self.gpu_assign_thr > 0) and (
+            gt_bboxes.shape[0] > self.gpu_assign_thr) else False
+        # compute overlap and assign gt on CPU when number of GT is large
+        if assign_on_cpu:
+            device = bboxes.device
+            bboxes = bboxes.cpu()
+            gt_bboxes = gt_bboxes.cpu()
+            if gt_bboxes_ignore is not None:
+                gt_bboxes_ignore = gt_bboxes_ignore.cpu()
+            if gt_labels is not None:
+                gt_labels = gt_labels.cpu()
+
+        overlaps = self.iou_calculator(gt_bboxes, bboxes)
+
+        if (self.ignore_iof_thr > 0 and gt_bboxes_ignore is not None
+                and gt_bboxes_ignore.numel() > 0 and bboxes.numel() > 0):
+            if self.ignore_wrt_candidates:
+                ignore_overlaps = self.iou_calculator(
+                    bboxes, gt_bboxes_ignore, mode='iof')
+                ignore_max_overlaps, _ = ignore_overlaps.max(dim=1)
+            else:
+                ignore_overlaps = self.iou_calculator(
+                    gt_bboxes_ignore, bboxes, mode='iof')
+                ignore_max_overlaps, _ = ignore_overlaps.max(dim=0)
+            overlaps[:, ignore_max_overlaps > self.ignore_iof_thr] = -1
+
+        assign_result = self.assign_wrt_overlaps(overlaps, gt_labels)
+        if assign_on_cpu:
+            assign_result.gt_inds = assign_result.gt_inds.to(device)
+            assign_result.max_overlaps = assign_result.max_overlaps.to(device)
+            if assign_result.labels is not None:
+                assign_result.labels = assign_result.labels.to(device)
+        return assign_result
+
+    def assign_wrt_overlaps(self, overlaps, gt_labels=None):
+        """Assign w.r.t. the overlaps of bboxes with gts.
+
+        Args:
+            overlaps (Tensor): Overlaps between k gt_bboxes and n bboxes,
+                shape(k, n).
+            gt_labels (Tensor, optional): Labels of k gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+        """
+        num_gts, num_bboxes = overlaps.size(0), overlaps.size(1)
+
+        # 1. assign -1 by default
+        assigned_gt_inds = overlaps.new_full((num_bboxes, ),
+                                             -1,
+                                             dtype=torch.long)
+
+        if num_gts == 0 or num_bboxes == 0:
+            # No ground truth or boxes, return empty assignment
+            max_overlaps = overlaps.new_zeros((num_bboxes, ))
+            if num_gts == 0:
+                # No truth, assign everything to background
+                assigned_gt_inds[:] = 0
+            if gt_labels is None:
+                assigned_labels = None
+            else:
+                assigned_labels = overlaps.new_full((num_bboxes, ),
+                                                    -1,
+                                                    dtype=torch.long)
+            return AssignResult(
+                num_gts,
+                assigned_gt_inds,
+                max_overlaps,
+                labels=assigned_labels)
+
+        # for each anchor, which gt best overlaps with it
+        # for each anchor, the max iou of all gts
+        max_overlaps, argmax_overlaps = overlaps.max(dim=0)
+        # for each gt, which anchor best overlaps with it
+        # for each gt, the max iou of all proposals
+        gt_max_overlaps, gt_argmax_overlaps = overlaps.max(dim=1)
+
+        # 2. assign negative: below
+        # the negative inds are set to be 0
+        if isinstance(self.neg_iou_thr, float):
+            assigned_gt_inds[(max_overlaps >= 0)
+                             & (max_overlaps < self.neg_iou_thr)] = 0
+        elif isinstance(self.neg_iou_thr, tuple):
+            assert len(self.neg_iou_thr) == 2
+            assigned_gt_inds[(max_overlaps >= self.neg_iou_thr[0])
+                             & (max_overlaps < self.neg_iou_thr[1])] = 0
+
+        # 3. assign positive: above positive IoU threshold
+        pos_inds = max_overlaps >= self.pos_iou_thr
+        assigned_gt_inds[pos_inds] = argmax_overlaps[pos_inds] + 1
+
+        if self.match_low_quality:
+            # Low-quality matching will overwirte the assigned_gt_inds assigned
+            # in Step 3. Thus, the assigned gt might not be the best one for
+            # prediction.
+            # For example, if bbox A has 0.9 and 0.8 iou with GT bbox 1 & 2,
+            # bbox 1 will be assigned as the best target for bbox A in step 3.
+            # However, if GT bbox 2's gt_argmax_overlaps = A, bbox A's
+            # assigned_gt_inds will be overwritten to be bbox B.
+            # This might be the reason that it is not used in ROI Heads.
+            for i in range(num_gts):
+                if gt_max_overlaps[i] >= self.min_pos_iou:
+                    if self.gt_max_assign_all:
+                        max_iou_inds = overlaps[i, :] == gt_max_overlaps[i]
+                        assigned_gt_inds[max_iou_inds] = i + 1
+                    else:
+                        assigned_gt_inds[gt_argmax_overlaps[i]] = i + 1
+
+        if gt_labels is not None:
+            assigned_labels = assigned_gt_inds.new_full((num_bboxes, ), -1)
+            pos_inds = torch.nonzero(
+                assigned_gt_inds > 0, as_tuple=False).squeeze()
+            if pos_inds.numel() > 0:
+                assigned_labels[pos_inds] = gt_labels[
+                    assigned_gt_inds[pos_inds] - 1]
+        else:
+            assigned_labels = None
+
+        return AssignResult(
+            num_gts, assigned_gt_inds, max_overlaps, labels=assigned_labels)
diff --git a/detection/scrfd/mmdet/core/bbox/assigners/point_assigner.py b/detection/scrfd/mmdet/core/bbox/assigners/point_assigner.py
new file mode 100755
index 000000000..fb8f5e4ed
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/assigners/point_assigner.py
@@ -0,0 +1,133 @@
+import torch
+
+from ..builder import BBOX_ASSIGNERS
+from .assign_result import AssignResult
+from .base_assigner import BaseAssigner
+
+
+@BBOX_ASSIGNERS.register_module()
+class PointAssigner(BaseAssigner):
+    """Assign a corresponding gt bbox or background to each point.
+
+    Each proposals will be assigned with `0`, or a positive integer
+    indicating the ground truth index.
+
+    - 0: negative sample, no assigned gt
+    - positive integer: positive sample, index (1-based) of assigned gt
+    """
+
+    def __init__(self, scale=4, pos_num=3):
+        self.scale = scale
+        self.pos_num = pos_num
+
+    def assign(self, points, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None):
+        """Assign gt to points.
+
+        This method assign a gt bbox to every points set, each points set
+        will be assigned with  the background_label (-1), or a label number.
+        -1 is background, and semi-positive number is the index (0-based) of
+        assigned gt.
+        The assignment is done in following steps, the order matters.
+
+        1. assign every points to the background_label (-1)
+        2. A point is assigned to some gt bbox if
+            (i) the point is within the k closest points to the gt bbox
+            (ii) the distance between this point and the gt is smaller than
+                other gt bboxes
+
+        Args:
+            points (Tensor): points to be assigned, shape(n, 3) while last
+                dimension stands for (x, y, stride).
+            gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
+            gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
+                labelled as `ignored`, e.g., crowd boxes in COCO.
+                NOTE: currently unused.
+            gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
+
+        Returns:
+            :obj:`AssignResult`: The assign result.
+        """
+        num_points = points.shape[0]
+        num_gts = gt_bboxes.shape[0]
+
+        if num_gts == 0 or num_points == 0:
+            # If no truth assign everything to the background
+            assigned_gt_inds = points.new_full((num_points, ),
+                                               0,
+                                               dtype=torch.long)
+            if gt_labels is None:
+                assigned_labels = None
+            else:
+                assigned_labels = points.new_full((num_points, ),
+                                                  -1,
+                                                  dtype=torch.long)
+            return AssignResult(
+                num_gts, assigned_gt_inds, None, labels=assigned_labels)
+
+        points_xy = points[:, :2]
+        points_stride = points[:, 2]
+        points_lvl = torch.log2(
+            points_stride).int()  # [3...,4...,5...,6...,7...]
+        lvl_min, lvl_max = points_lvl.min(), points_lvl.max()
+
+        # assign gt box
+        gt_bboxes_xy = (gt_bboxes[:, :2] + gt_bboxes[:, 2:]) / 2
+        gt_bboxes_wh = (gt_bboxes[:, 2:] - gt_bboxes[:, :2]).clamp(min=1e-6)
+        scale = self.scale
+        gt_bboxes_lvl = ((torch.log2(gt_bboxes_wh[:, 0] / scale) +
+                          torch.log2(gt_bboxes_wh[:, 1] / scale)) / 2).int()
+        gt_bboxes_lvl = torch.clamp(gt_bboxes_lvl, min=lvl_min, max=lvl_max)
+
+        # stores the assigned gt index of each point
+        assigned_gt_inds = points.new_zeros((num_points, ), dtype=torch.long)
+        # stores the assigned gt dist (to this point) of each point
+        assigned_gt_dist = points.new_full((num_points, ), float('inf'))
+        points_range = torch.arange(points.shape[0])
+
+        for idx in range(num_gts):
+            gt_lvl = gt_bboxes_lvl[idx]
+            # get the index of points in this level
+            lvl_idx = gt_lvl == points_lvl
+            points_index = points_range[lvl_idx]
+            # get the points in this level
+            lvl_points = points_xy[lvl_idx, :]
+            # get the center point of gt
+            gt_point = gt_bboxes_xy[[idx], :]
+            # get width and height of gt
+            gt_wh = gt_bboxes_wh[[idx], :]
+            # compute the distance between gt center and
+            #   all points in this level
+            points_gt_dist = ((lvl_points - gt_point) / gt_wh).norm(dim=1)
+            # find the nearest k points to gt center in this level
+            min_dist, min_dist_index = torch.topk(
+                points_gt_dist, self.pos_num, largest=False)
+            # the index of nearest k points to gt center in this level
+            min_dist_points_index = points_index[min_dist_index]
+            # The less_than_recorded_index stores the index
+            #   of min_dist that is less then the assigned_gt_dist. Where
+            #   assigned_gt_dist stores the dist from previous assigned gt
+            #   (if exist) to each point.
+            less_than_recorded_index = min_dist < assigned_gt_dist[
+                min_dist_points_index]
+            # The min_dist_points_index stores the index of points satisfy:
+            #   (1) it is k nearest to current gt center in this level.
+            #   (2) it is closer to current gt center than other gt center.
+            min_dist_points_index = min_dist_points_index[
+                less_than_recorded_index]
+            # assign the result
+            assigned_gt_inds[min_dist_points_index] = idx + 1
+            assigned_gt_dist[min_dist_points_index] = min_dist[
+                less_than_recorded_index]
+
+        if gt_labels is not None:
+            assigned_labels = assigned_gt_inds.new_full((num_points, ), -1)
+            pos_inds = torch.nonzero(
+                assigned_gt_inds > 0, as_tuple=False).squeeze()
+            if pos_inds.numel() > 0:
+                assigned_labels[pos_inds] = gt_labels[
+                    assigned_gt_inds[pos_inds] - 1]
+        else:
+            assigned_labels = None
+
+        return AssignResult(
+            num_gts, assigned_gt_inds, None, labels=assigned_labels)
diff --git a/detection/scrfd/mmdet/core/bbox/builder.py b/detection/scrfd/mmdet/core/bbox/builder.py
new file mode 100755
index 000000000..682683b62
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/builder.py
@@ -0,0 +1,20 @@
+from mmcv.utils import Registry, build_from_cfg
+
+BBOX_ASSIGNERS = Registry('bbox_assigner')
+BBOX_SAMPLERS = Registry('bbox_sampler')
+BBOX_CODERS = Registry('bbox_coder')
+
+
+def build_assigner(cfg, **default_args):
+    """Builder of box assigner."""
+    return build_from_cfg(cfg, BBOX_ASSIGNERS, default_args)
+
+
+def build_sampler(cfg, **default_args):
+    """Builder of box sampler."""
+    return build_from_cfg(cfg, BBOX_SAMPLERS, default_args)
+
+
+def build_bbox_coder(cfg, **default_args):
+    """Builder of box coder."""
+    return build_from_cfg(cfg, BBOX_CODERS, default_args)
diff --git a/detection/scrfd/mmdet/core/bbox/coder/__init__.py b/detection/scrfd/mmdet/core/bbox/coder/__init__.py
new file mode 100755
index 000000000..ae455ba8f
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/__init__.py
@@ -0,0 +1,13 @@
+from .base_bbox_coder import BaseBBoxCoder
+from .bucketing_bbox_coder import BucketingBBoxCoder
+from .delta_xywh_bbox_coder import DeltaXYWHBBoxCoder
+from .legacy_delta_xywh_bbox_coder import LegacyDeltaXYWHBBoxCoder
+from .pseudo_bbox_coder import PseudoBBoxCoder
+from .tblr_bbox_coder import TBLRBBoxCoder
+from .yolo_bbox_coder import YOLOBBoxCoder
+
+__all__ = [
+    'BaseBBoxCoder', 'PseudoBBoxCoder', 'DeltaXYWHBBoxCoder',
+    'LegacyDeltaXYWHBBoxCoder', 'TBLRBBoxCoder', 'YOLOBBoxCoder',
+    'BucketingBBoxCoder'
+]
diff --git a/detection/scrfd/mmdet/core/bbox/coder/base_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/base_bbox_coder.py
new file mode 100755
index 000000000..6e4272721
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/base_bbox_coder.py
@@ -0,0 +1,19 @@
+from abc import ABCMeta, abstractmethod
+
+
+class BaseBBoxCoder(metaclass=ABCMeta):
+    """Base bounding box coder."""
+
+    def __init__(self, **kwargs):
+        pass
+
+    @abstractmethod
+    def encode(self, bboxes, gt_bboxes):
+        """Encode deltas between bboxes and ground truth boxes."""
+        pass
+
+    @abstractmethod
+    def decode(self, bboxes, bboxes_pred):
+        """Decode the predicted bboxes according to prediction and base
+        boxes."""
+        pass
diff --git a/detection/scrfd/mmdet/core/bbox/coder/bucketing_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/bucketing_bbox_coder.py
new file mode 100755
index 000000000..e8c450c5f
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/bucketing_bbox_coder.py
@@ -0,0 +1,346 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from ..builder import BBOX_CODERS
+from ..transforms import bbox_rescale
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class BucketingBBoxCoder(BaseBBoxCoder):
+    """Bucketing BBox Coder for Side-Aware Bounday Localization (SABL).
+
+    Boundary Localization with Bucketing and Bucketing Guided Rescoring
+    are implemented here.
+
+    Please refer to https://arxiv.org/abs/1912.04260 for more details.
+
+    Args:
+        num_buckets (int): Number of buckets.
+        scale_factor (int): Scale factor of proposals to generate buckets.
+        offset_topk (int): Topk buckets are used to generate
+             bucket fine regression targets. Defaults to 2.
+        offset_upperbound (float): Offset upperbound to generate
+             bucket fine regression targets.
+             To avoid too large offset displacements. Defaults to 1.0.
+        cls_ignore_neighbor (bool): Ignore second nearest bucket or Not.
+             Defaults to True.
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+    """
+
+    def __init__(self,
+                 num_buckets,
+                 scale_factor,
+                 offset_topk=2,
+                 offset_upperbound=1.0,
+                 cls_ignore_neighbor=True,
+                 clip_border=True):
+        super(BucketingBBoxCoder, self).__init__()
+        self.num_buckets = num_buckets
+        self.scale_factor = scale_factor
+        self.offset_topk = offset_topk
+        self.offset_upperbound = offset_upperbound
+        self.cls_ignore_neighbor = cls_ignore_neighbor
+        self.clip_border = clip_border
+
+    def encode(self, bboxes, gt_bboxes):
+        """Get bucketing estimation and fine regression targets during
+        training.
+
+        Args:
+            bboxes (torch.Tensor): source boxes, e.g., object proposals.
+            gt_bboxes (torch.Tensor): target of the transformation, e.g.,
+                ground truth boxes.
+
+        Returns:
+           encoded_bboxes(tuple[Tensor]): bucketing estimation
+            and fine regression targets and weights
+        """
+
+        assert bboxes.size(0) == gt_bboxes.size(0)
+        assert bboxes.size(-1) == gt_bboxes.size(-1) == 4
+        encoded_bboxes = bbox2bucket(bboxes, gt_bboxes, self.num_buckets,
+                                     self.scale_factor, self.offset_topk,
+                                     self.offset_upperbound,
+                                     self.cls_ignore_neighbor)
+        return encoded_bboxes
+
+    def decode(self, bboxes, pred_bboxes, max_shape=None):
+        """Apply transformation `pred_bboxes` to `boxes`.
+        Args:
+            boxes (torch.Tensor): Basic boxes.
+            pred_bboxes (torch.Tensor): Predictions for bucketing estimation
+                and fine regression
+            max_shape (tuple[int], optional): Maximum shape of boxes.
+                Defaults to None.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+        assert len(pred_bboxes) == 2
+        cls_preds, offset_preds = pred_bboxes
+        assert cls_preds.size(0) == bboxes.size(0) and offset_preds.size(
+            0) == bboxes.size(0)
+        decoded_bboxes = bucket2bbox(bboxes, cls_preds, offset_preds,
+                                     self.num_buckets, self.scale_factor,
+                                     max_shape, self.clip_border)
+
+        return decoded_bboxes
+
+
+def generat_buckets(proposals, num_buckets, scale_factor=1.0):
+    """Generate buckets w.r.t bucket number and scale factor of proposals.
+
+    Args:
+        proposals (Tensor): Shape (n, 4)
+        num_buckets (int): Number of buckets.
+        scale_factor (float): Scale factor to rescale proposals.
+
+    Returns:
+        tuple[Tensor]: (bucket_w, bucket_h, l_buckets, r_buckets,
+         t_buckets, d_buckets)
+
+            - bucket_w: Width of buckets on x-axis. Shape (n, ).
+            - bucket_h: Height of buckets on y-axis. Shape (n, ).
+            - l_buckets: Left buckets. Shape (n, ceil(side_num/2)).
+            - r_buckets: Right buckets. Shape (n, ceil(side_num/2)).
+            - t_buckets: Top buckets. Shape (n, ceil(side_num/2)).
+            - d_buckets: Down buckets. Shape (n, ceil(side_num/2)).
+    """
+    proposals = bbox_rescale(proposals, scale_factor)
+
+    # number of buckets in each side
+    side_num = int(np.ceil(num_buckets / 2.0))
+    pw = proposals[..., 2] - proposals[..., 0]
+    ph = proposals[..., 3] - proposals[..., 1]
+    px1 = proposals[..., 0]
+    py1 = proposals[..., 1]
+    px2 = proposals[..., 2]
+    py2 = proposals[..., 3]
+
+    bucket_w = pw / num_buckets
+    bucket_h = ph / num_buckets
+
+    # left buckets
+    l_buckets = px1[:, None] + (0.5 + torch.arange(
+        0, side_num).to(proposals).float())[None, :] * bucket_w[:, None]
+    # right buckets
+    r_buckets = px2[:, None] - (0.5 + torch.arange(
+        0, side_num).to(proposals).float())[None, :] * bucket_w[:, None]
+    # top buckets
+    t_buckets = py1[:, None] + (0.5 + torch.arange(
+        0, side_num).to(proposals).float())[None, :] * bucket_h[:, None]
+    # down buckets
+    d_buckets = py2[:, None] - (0.5 + torch.arange(
+        0, side_num).to(proposals).float())[None, :] * bucket_h[:, None]
+    return bucket_w, bucket_h, l_buckets, r_buckets, t_buckets, d_buckets
+
+
+def bbox2bucket(proposals,
+                gt,
+                num_buckets,
+                scale_factor,
+                offset_topk=2,
+                offset_upperbound=1.0,
+                cls_ignore_neighbor=True):
+    """Generate buckets estimation and fine regression targets.
+
+    Args:
+        proposals (Tensor): Shape (n, 4)
+        gt (Tensor): Shape (n, 4)
+        num_buckets (int): Number of buckets.
+        scale_factor (float): Scale factor to rescale proposals.
+        offset_topk (int): Topk buckets are used to generate
+             bucket fine regression targets. Defaults to 2.
+        offset_upperbound (float): Offset allowance to generate
+             bucket fine regression targets.
+             To avoid too large offset displacements. Defaults to 1.0.
+        cls_ignore_neighbor (bool): Ignore second nearest bucket or Not.
+             Defaults to True.
+
+    Returns:
+        tuple[Tensor]: (offsets, offsets_weights, bucket_labels, cls_weights).
+
+            - offsets: Fine regression targets. \
+                Shape (n, num_buckets*2).
+            - offsets_weights: Fine regression weights. \
+                Shape (n, num_buckets*2).
+            - bucket_labels: Bucketing estimation labels. \
+                Shape (n, num_buckets*2).
+            - cls_weights: Bucketing estimation weights. \
+                Shape (n, num_buckets*2).
+    """
+    assert proposals.size() == gt.size()
+
+    # generate buckets
+    proposals = proposals.float()
+    gt = gt.float()
+    (bucket_w, bucket_h, l_buckets, r_buckets, t_buckets,
+     d_buckets) = generat_buckets(proposals, num_buckets, scale_factor)
+
+    gx1 = gt[..., 0]
+    gy1 = gt[..., 1]
+    gx2 = gt[..., 2]
+    gy2 = gt[..., 3]
+
+    # generate offset targets and weights
+    # offsets from buckets to gts
+    l_offsets = (l_buckets - gx1[:, None]) / bucket_w[:, None]
+    r_offsets = (r_buckets - gx2[:, None]) / bucket_w[:, None]
+    t_offsets = (t_buckets - gy1[:, None]) / bucket_h[:, None]
+    d_offsets = (d_buckets - gy2[:, None]) / bucket_h[:, None]
+
+    # select top-k nearset buckets
+    l_topk, l_label = l_offsets.abs().topk(
+        offset_topk, dim=1, largest=False, sorted=True)
+    r_topk, r_label = r_offsets.abs().topk(
+        offset_topk, dim=1, largest=False, sorted=True)
+    t_topk, t_label = t_offsets.abs().topk(
+        offset_topk, dim=1, largest=False, sorted=True)
+    d_topk, d_label = d_offsets.abs().topk(
+        offset_topk, dim=1, largest=False, sorted=True)
+
+    offset_l_weights = l_offsets.new_zeros(l_offsets.size())
+    offset_r_weights = r_offsets.new_zeros(r_offsets.size())
+    offset_t_weights = t_offsets.new_zeros(t_offsets.size())
+    offset_d_weights = d_offsets.new_zeros(d_offsets.size())
+    inds = torch.arange(0, proposals.size(0)).to(proposals).long()
+
+    # generate offset weights of top-k nearset buckets
+    for k in range(offset_topk):
+        if k >= 1:
+            offset_l_weights[inds, l_label[:,
+                                           k]] = (l_topk[:, k] <
+                                                  offset_upperbound).float()
+            offset_r_weights[inds, r_label[:,
+                                           k]] = (r_topk[:, k] <
+                                                  offset_upperbound).float()
+            offset_t_weights[inds, t_label[:,
+                                           k]] = (t_topk[:, k] <
+                                                  offset_upperbound).float()
+            offset_d_weights[inds, d_label[:,
+                                           k]] = (d_topk[:, k] <
+                                                  offset_upperbound).float()
+        else:
+            offset_l_weights[inds, l_label[:, k]] = 1.0
+            offset_r_weights[inds, r_label[:, k]] = 1.0
+            offset_t_weights[inds, t_label[:, k]] = 1.0
+            offset_d_weights[inds, d_label[:, k]] = 1.0
+
+    offsets = torch.cat([l_offsets, r_offsets, t_offsets, d_offsets], dim=-1)
+    offsets_weights = torch.cat([
+        offset_l_weights, offset_r_weights, offset_t_weights, offset_d_weights
+    ],
+                                dim=-1)
+
+    # generate bucket labels and weight
+    side_num = int(np.ceil(num_buckets / 2.0))
+    labels = torch.stack(
+        [l_label[:, 0], r_label[:, 0], t_label[:, 0], d_label[:, 0]], dim=-1)
+
+    batch_size = labels.size(0)
+    bucket_labels = F.one_hot(labels.view(-1), side_num).view(batch_size,
+                                                              -1).float()
+    bucket_cls_l_weights = (l_offsets.abs() < 1).float()
+    bucket_cls_r_weights = (r_offsets.abs() < 1).float()
+    bucket_cls_t_weights = (t_offsets.abs() < 1).float()
+    bucket_cls_d_weights = (d_offsets.abs() < 1).float()
+    bucket_cls_weights = torch.cat([
+        bucket_cls_l_weights, bucket_cls_r_weights, bucket_cls_t_weights,
+        bucket_cls_d_weights
+    ],
+                                   dim=-1)
+    # ignore second nearest buckets for cls if necessay
+    if cls_ignore_neighbor:
+        bucket_cls_weights = (~((bucket_cls_weights == 1) &
+                                (bucket_labels == 0))).float()
+    else:
+        bucket_cls_weights[:] = 1.0
+    return offsets, offsets_weights, bucket_labels, bucket_cls_weights
+
+
+def bucket2bbox(proposals,
+                cls_preds,
+                offset_preds,
+                num_buckets,
+                scale_factor=1.0,
+                max_shape=None,
+                clip_border=True):
+    """Apply bucketing estimation (cls preds) and fine regression (offset
+    preds) to generate det bboxes.
+
+    Args:
+        proposals (Tensor): Boxes to be transformed. Shape (n, 4)
+        cls_preds (Tensor): bucketing estimation. Shape (n, num_buckets*2).
+        offset_preds (Tensor): fine regression. Shape (n, num_buckets*2).
+        num_buckets (int): Number of buckets.
+        scale_factor (float): Scale factor to rescale proposals.
+        max_shape (tuple[int, int]): Maximum bounds for boxes. specifies (H, W)
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+
+    Returns:
+        tuple[Tensor]: (bboxes, loc_confidence).
+
+            - bboxes: predicted bboxes. Shape (n, 4)
+            - loc_confidence: localization confidence of predicted bboxes.
+                Shape (n,).
+    """
+
+    side_num = int(np.ceil(num_buckets / 2.0))
+    cls_preds = cls_preds.view(-1, side_num)
+    offset_preds = offset_preds.view(-1, side_num)
+
+    scores = F.softmax(cls_preds, dim=1)
+    score_topk, score_label = scores.topk(2, dim=1, largest=True, sorted=True)
+
+    rescaled_proposals = bbox_rescale(proposals, scale_factor)
+
+    pw = rescaled_proposals[..., 2] - rescaled_proposals[..., 0]
+    ph = rescaled_proposals[..., 3] - rescaled_proposals[..., 1]
+    px1 = rescaled_proposals[..., 0]
+    py1 = rescaled_proposals[..., 1]
+    px2 = rescaled_proposals[..., 2]
+    py2 = rescaled_proposals[..., 3]
+
+    bucket_w = pw / num_buckets
+    bucket_h = ph / num_buckets
+
+    score_inds_l = score_label[0::4, 0]
+    score_inds_r = score_label[1::4, 0]
+    score_inds_t = score_label[2::4, 0]
+    score_inds_d = score_label[3::4, 0]
+    l_buckets = px1 + (0.5 + score_inds_l.float()) * bucket_w
+    r_buckets = px2 - (0.5 + score_inds_r.float()) * bucket_w
+    t_buckets = py1 + (0.5 + score_inds_t.float()) * bucket_h
+    d_buckets = py2 - (0.5 + score_inds_d.float()) * bucket_h
+
+    offsets = offset_preds.view(-1, 4, side_num)
+    inds = torch.arange(proposals.size(0)).to(proposals).long()
+    l_offsets = offsets[:, 0, :][inds, score_inds_l]
+    r_offsets = offsets[:, 1, :][inds, score_inds_r]
+    t_offsets = offsets[:, 2, :][inds, score_inds_t]
+    d_offsets = offsets[:, 3, :][inds, score_inds_d]
+
+    x1 = l_buckets - l_offsets * bucket_w
+    x2 = r_buckets - r_offsets * bucket_w
+    y1 = t_buckets - t_offsets * bucket_h
+    y2 = d_buckets - d_offsets * bucket_h
+
+    if clip_border and max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1] - 1)
+        y1 = y1.clamp(min=0, max=max_shape[0] - 1)
+        x2 = x2.clamp(min=0, max=max_shape[1] - 1)
+        y2 = y2.clamp(min=0, max=max_shape[0] - 1)
+    bboxes = torch.cat([x1[:, None], y1[:, None], x2[:, None], y2[:, None]],
+                       dim=-1)
+
+    # bucketing guided rescoring
+    loc_confidence = score_topk[:, 0]
+    top2_neighbor_inds = (score_label[:, 0] - score_label[:, 1]).abs() == 1
+    loc_confidence += score_topk[:, 1] * top2_neighbor_inds.float()
+    loc_confidence = loc_confidence.view(-1, 4).mean(dim=1)
+
+    return bboxes, loc_confidence
diff --git a/detection/scrfd/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py
new file mode 100755
index 000000000..e9eb35790
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py
@@ -0,0 +1,204 @@
+import numpy as np
+import torch
+
+from ..builder import BBOX_CODERS
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class DeltaXYWHBBoxCoder(BaseBBoxCoder):
+    """Delta XYWH BBox coder.
+
+    Following the practice in `R-CNN <https://arxiv.org/abs/1311.2524>`_,
+    this coder encodes bbox (x1, y1, x2, y2) into delta (dx, dy, dw, dh) and
+    decodes delta (dx, dy, dw, dh) back to original bbox (x1, y1, x2, y2).
+
+    Args:
+        target_means (Sequence[float]): Denormalizing means of target for
+            delta coordinates
+        target_stds (Sequence[float]): Denormalizing standard deviation of
+            target for delta coordinates
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+    """
+
+    def __init__(self,
+                 target_means=(0., 0., 0., 0.),
+                 target_stds=(1., 1., 1., 1.),
+                 clip_border=True):
+        super(BaseBBoxCoder, self).__init__()
+        self.means = target_means
+        self.stds = target_stds
+        self.clip_border = clip_border
+
+    def encode(self, bboxes, gt_bboxes):
+        """Get box regression transformation deltas that can be used to
+        transform the ``bboxes`` into the ``gt_bboxes``.
+
+        Args:
+            bboxes (torch.Tensor): Source boxes, e.g., object proposals.
+            gt_bboxes (torch.Tensor): Target of the transformation, e.g.,
+                ground-truth boxes.
+
+        Returns:
+            torch.Tensor: Box transformation deltas
+        """
+
+        assert bboxes.size(0) == gt_bboxes.size(0)
+        assert bboxes.size(-1) == gt_bboxes.size(-1) == 4
+        encoded_bboxes = bbox2delta(bboxes, gt_bboxes, self.means, self.stds)
+        return encoded_bboxes
+
+    def decode(self,
+               bboxes,
+               pred_bboxes,
+               max_shape=None,
+               wh_ratio_clip=16 / 1000):
+        """Apply transformation `pred_bboxes` to `boxes`.
+
+        Args:
+            boxes (torch.Tensor): Basic boxes.
+            pred_bboxes (torch.Tensor): Encoded boxes with shape
+            max_shape (tuple[int], optional): Maximum shape of boxes.
+                Defaults to None.
+            wh_ratio_clip (float, optional): The allowed ratio between
+                width and height.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+
+        assert pred_bboxes.size(0) == bboxes.size(0)
+        decoded_bboxes = delta2bbox(bboxes, pred_bboxes, self.means, self.stds,
+                                    max_shape, wh_ratio_clip, self.clip_border)
+
+        return decoded_bboxes
+
+
+def bbox2delta(proposals, gt, means=(0., 0., 0., 0.), stds=(1., 1., 1., 1.)):
+    """Compute deltas of proposals w.r.t. gt.
+
+    We usually compute the deltas of x, y, w, h of proposals w.r.t ground
+    truth bboxes to get regression target.
+    This is the inverse function of :func:`delta2bbox`.
+
+    Args:
+        proposals (Tensor): Boxes to be transformed, shape (N, ..., 4)
+        gt (Tensor): Gt bboxes to be used as base, shape (N, ..., 4)
+        means (Sequence[float]): Denormalizing means for delta coordinates
+        stds (Sequence[float]): Denormalizing standard deviation for delta
+            coordinates
+
+    Returns:
+        Tensor: deltas with shape (N, 4), where columns represent dx, dy,
+            dw, dh.
+    """
+    assert proposals.size() == gt.size()
+
+    proposals = proposals.float()
+    gt = gt.float()
+    px = (proposals[..., 0] + proposals[..., 2]) * 0.5
+    py = (proposals[..., 1] + proposals[..., 3]) * 0.5
+    pw = proposals[..., 2] - proposals[..., 0]
+    ph = proposals[..., 3] - proposals[..., 1]
+
+    gx = (gt[..., 0] + gt[..., 2]) * 0.5
+    gy = (gt[..., 1] + gt[..., 3]) * 0.5
+    gw = gt[..., 2] - gt[..., 0]
+    gh = gt[..., 3] - gt[..., 1]
+
+    dx = (gx - px) / pw
+    dy = (gy - py) / ph
+    dw = torch.log(gw / pw)
+    dh = torch.log(gh / ph)
+    deltas = torch.stack([dx, dy, dw, dh], dim=-1)
+
+    means = deltas.new_tensor(means).unsqueeze(0)
+    stds = deltas.new_tensor(stds).unsqueeze(0)
+    deltas = deltas.sub_(means).div_(stds)
+
+    return deltas
+
+
+def delta2bbox(rois,
+               deltas,
+               means=(0., 0., 0., 0.),
+               stds=(1., 1., 1., 1.),
+               max_shape=None,
+               wh_ratio_clip=16 / 1000,
+               clip_border=True):
+    """Apply deltas to shift/scale base boxes.
+
+    Typically the rois are anchor or proposed bounding boxes and the deltas are
+    network outputs used to shift/scale those boxes.
+    This is the inverse function of :func:`bbox2delta`.
+
+    Args:
+        rois (Tensor): Boxes to be transformed. Has shape (N, 4)
+        deltas (Tensor): Encoded offsets with respect to each roi.
+            Has shape (N, 4 * num_classes). Note N = num_anchors * W * H when
+            rois is a grid of anchors. Offset encoding follows [1]_.
+        means (Sequence[float]): Denormalizing means for delta coordinates
+        stds (Sequence[float]): Denormalizing standard deviation for delta
+            coordinates
+        max_shape (tuple[int, int]): Maximum bounds for boxes. specifies (H, W)
+        wh_ratio_clip (float): Maximum aspect ratio for boxes.
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+
+    Returns:
+        Tensor: Boxes with shape (N, 4), where columns represent
+            tl_x, tl_y, br_x, br_y.
+
+    References:
+        .. [1] https://arxiv.org/abs/1311.2524
+
+    Example:
+        >>> rois = torch.Tensor([[ 0.,  0.,  1.,  1.],
+        >>>                      [ 0.,  0.,  1.,  1.],
+        >>>                      [ 0.,  0.,  1.,  1.],
+        >>>                      [ 5.,  5.,  5.,  5.]])
+        >>> deltas = torch.Tensor([[  0.,   0.,   0.,   0.],
+        >>>                        [  1.,   1.,   1.,   1.],
+        >>>                        [  0.,   0.,   2.,  -1.],
+        >>>                        [ 0.7, -1.9, -0.5,  0.3]])
+        >>> delta2bbox(rois, deltas, max_shape=(32, 32))
+        tensor([[0.0000, 0.0000, 1.0000, 1.0000],
+                [0.1409, 0.1409, 2.8591, 2.8591],
+                [0.0000, 0.3161, 4.1945, 0.6839],
+                [5.0000, 5.0000, 5.0000, 5.0000]])
+    """
+    means = deltas.new_tensor(means).view(1, -1).repeat(1, deltas.size(1) // 4)
+    stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(1) // 4)
+    denorm_deltas = deltas * stds + means
+    dx = denorm_deltas[:, 0::4]
+    dy = denorm_deltas[:, 1::4]
+    dw = denorm_deltas[:, 2::4]
+    dh = denorm_deltas[:, 3::4]
+    max_ratio = np.abs(np.log(wh_ratio_clip))
+    dw = dw.clamp(min=-max_ratio, max=max_ratio)
+    dh = dh.clamp(min=-max_ratio, max=max_ratio)
+    # Compute center of each roi
+    px = ((rois[:, 0] + rois[:, 2]) * 0.5).unsqueeze(1).expand_as(dx)
+    py = ((rois[:, 1] + rois[:, 3]) * 0.5).unsqueeze(1).expand_as(dy)
+    # Compute width/height of each roi
+    pw = (rois[:, 2] - rois[:, 0]).unsqueeze(1).expand_as(dw)
+    ph = (rois[:, 3] - rois[:, 1]).unsqueeze(1).expand_as(dh)
+    # Use exp(network energy) to enlarge/shrink each roi
+    gw = pw * dw.exp()
+    gh = ph * dh.exp()
+    # Use network energy to shift the center of each roi
+    gx = px + pw * dx
+    gy = py + ph * dy
+    # Convert center-xy/width/height to top-left, bottom-right
+    x1 = gx - gw * 0.5
+    y1 = gy - gh * 0.5
+    x2 = gx + gw * 0.5
+    y2 = gy + gh * 0.5
+    if clip_border and max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    bboxes = torch.stack([x1, y1, x2, y2], dim=-1).view(deltas.size())
+    return bboxes
diff --git a/detection/scrfd/mmdet/core/bbox/coder/legacy_delta_xywh_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/legacy_delta_xywh_bbox_coder.py
new file mode 100755
index 000000000..74e801a15
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/legacy_delta_xywh_bbox_coder.py
@@ -0,0 +1,212 @@
+import numpy as np
+import torch
+
+from ..builder import BBOX_CODERS
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class LegacyDeltaXYWHBBoxCoder(BaseBBoxCoder):
+    """Legacy Delta XYWH BBox coder used in MMDet V1.x.
+
+    Following the practice in R-CNN [1]_, this coder encodes bbox (x1, y1, x2,
+    y2) into delta (dx, dy, dw, dh) and decodes delta (dx, dy, dw, dh)
+    back to original bbox (x1, y1, x2, y2).
+
+    Note:
+        The main difference between :class`LegacyDeltaXYWHBBoxCoder` and
+        :class:`DeltaXYWHBBoxCoder` is whether ``+ 1`` is used during width and
+        height calculation. We suggest to only use this coder when testing with
+        MMDet V1.x models.
+
+    References:
+        .. [1] https://arxiv.org/abs/1311.2524
+
+    Args:
+        target_means (Sequence[float]): denormalizing means of target for
+            delta coordinates
+        target_stds (Sequence[float]): denormalizing standard deviation of
+            target for delta coordinates
+    """
+
+    def __init__(self,
+                 target_means=(0., 0., 0., 0.),
+                 target_stds=(1., 1., 1., 1.)):
+        super(BaseBBoxCoder, self).__init__()
+        self.means = target_means
+        self.stds = target_stds
+
+    def encode(self, bboxes, gt_bboxes):
+        """Get box regression transformation deltas that can be used to
+        transform the ``bboxes`` into the ``gt_bboxes``.
+
+        Args:
+            bboxes (torch.Tensor): source boxes, e.g., object proposals.
+            gt_bboxes (torch.Tensor): target of the transformation, e.g.,
+                ground-truth boxes.
+
+        Returns:
+            torch.Tensor: Box transformation deltas
+        """
+        assert bboxes.size(0) == gt_bboxes.size(0)
+        assert bboxes.size(-1) == gt_bboxes.size(-1) == 4
+        encoded_bboxes = legacy_bbox2delta(bboxes, gt_bboxes, self.means,
+                                           self.stds)
+        return encoded_bboxes
+
+    def decode(self,
+               bboxes,
+               pred_bboxes,
+               max_shape=None,
+               wh_ratio_clip=16 / 1000):
+        """Apply transformation `pred_bboxes` to `boxes`.
+
+        Args:
+            boxes (torch.Tensor): Basic boxes.
+            pred_bboxes (torch.Tensor): Encoded boxes with shape
+            max_shape (tuple[int], optional): Maximum shape of boxes.
+                Defaults to None.
+            wh_ratio_clip (float, optional): The allowed ratio between
+                width and height.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+        assert pred_bboxes.size(0) == bboxes.size(0)
+        decoded_bboxes = legacy_delta2bbox(bboxes, pred_bboxes, self.means,
+                                           self.stds, max_shape, wh_ratio_clip)
+
+        return decoded_bboxes
+
+
+def legacy_bbox2delta(proposals,
+                      gt,
+                      means=(0., 0., 0., 0.),
+                      stds=(1., 1., 1., 1.)):
+    """Compute deltas of proposals w.r.t. gt in the MMDet V1.x manner.
+
+    We usually compute the deltas of x, y, w, h of proposals w.r.t ground
+    truth bboxes to get regression target.
+    This is the inverse function of `delta2bbox()`
+
+    Args:
+        proposals (Tensor): Boxes to be transformed, shape (N, ..., 4)
+        gt (Tensor): Gt bboxes to be used as base, shape (N, ..., 4)
+        means (Sequence[float]): Denormalizing means for delta coordinates
+        stds (Sequence[float]): Denormalizing standard deviation for delta
+            coordinates
+
+    Returns:
+        Tensor: deltas with shape (N, 4), where columns represent dx, dy,
+            dw, dh.
+    """
+    assert proposals.size() == gt.size()
+
+    proposals = proposals.float()
+    gt = gt.float()
+    px = (proposals[..., 0] + proposals[..., 2]) * 0.5
+    py = (proposals[..., 1] + proposals[..., 3]) * 0.5
+    pw = proposals[..., 2] - proposals[..., 0] + 1.0
+    ph = proposals[..., 3] - proposals[..., 1] + 1.0
+
+    gx = (gt[..., 0] + gt[..., 2]) * 0.5
+    gy = (gt[..., 1] + gt[..., 3]) * 0.5
+    gw = gt[..., 2] - gt[..., 0] + 1.0
+    gh = gt[..., 3] - gt[..., 1] + 1.0
+
+    dx = (gx - px) / pw
+    dy = (gy - py) / ph
+    dw = torch.log(gw / pw)
+    dh = torch.log(gh / ph)
+    deltas = torch.stack([dx, dy, dw, dh], dim=-1)
+
+    means = deltas.new_tensor(means).unsqueeze(0)
+    stds = deltas.new_tensor(stds).unsqueeze(0)
+    deltas = deltas.sub_(means).div_(stds)
+
+    return deltas
+
+
+def legacy_delta2bbox(rois,
+                      deltas,
+                      means=(0., 0., 0., 0.),
+                      stds=(1., 1., 1., 1.),
+                      max_shape=None,
+                      wh_ratio_clip=16 / 1000):
+    """Apply deltas to shift/scale base boxes in the MMDet V1.x manner.
+
+    Typically the rois are anchor or proposed bounding boxes and the deltas are
+    network outputs used to shift/scale those boxes.
+    This is the inverse function of `bbox2delta()`
+
+    Args:
+        rois (Tensor): Boxes to be transformed. Has shape (N, 4)
+        deltas (Tensor): Encoded offsets with respect to each roi.
+            Has shape (N, 4 * num_classes). Note N = num_anchors * W * H when
+            rois is a grid of anchors. Offset encoding follows [1]_.
+        means (Sequence[float]): Denormalizing means for delta coordinates
+        stds (Sequence[float]): Denormalizing standard deviation for delta
+            coordinates
+        max_shape (tuple[int, int]): Maximum bounds for boxes. specifies (H, W)
+        wh_ratio_clip (float): Maximum aspect ratio for boxes.
+
+    Returns:
+        Tensor: Boxes with shape (N, 4), where columns represent
+            tl_x, tl_y, br_x, br_y.
+
+    References:
+        .. [1] https://arxiv.org/abs/1311.2524
+
+    Example:
+        >>> rois = torch.Tensor([[ 0.,  0.,  1.,  1.],
+        >>>                      [ 0.,  0.,  1.,  1.],
+        >>>                      [ 0.,  0.,  1.,  1.],
+        >>>                      [ 5.,  5.,  5.,  5.]])
+        >>> deltas = torch.Tensor([[  0.,   0.,   0.,   0.],
+        >>>                        [  1.,   1.,   1.,   1.],
+        >>>                        [  0.,   0.,   2.,  -1.],
+        >>>                        [ 0.7, -1.9, -0.5,  0.3]])
+        >>> legacy_delta2bbox(rois, deltas, max_shape=(32, 32))
+        tensor([[0.0000, 0.0000, 1.5000, 1.5000],
+                [0.0000, 0.0000, 5.2183, 5.2183],
+                [0.0000, 0.1321, 7.8891, 0.8679],
+                [5.3967, 2.4251, 6.0033, 3.7749]])
+    """
+    means = deltas.new_tensor(means).repeat(1, deltas.size(1) // 4)
+    stds = deltas.new_tensor(stds).repeat(1, deltas.size(1) // 4)
+    denorm_deltas = deltas * stds + means
+    dx = denorm_deltas[:, 0::4]
+    dy = denorm_deltas[:, 1::4]
+    dw = denorm_deltas[:, 2::4]
+    dh = denorm_deltas[:, 3::4]
+    max_ratio = np.abs(np.log(wh_ratio_clip))
+    dw = dw.clamp(min=-max_ratio, max=max_ratio)
+    dh = dh.clamp(min=-max_ratio, max=max_ratio)
+    # Compute center of each roi
+    px = ((rois[:, 0] + rois[:, 2]) * 0.5).unsqueeze(1).expand_as(dx)
+    py = ((rois[:, 1] + rois[:, 3]) * 0.5).unsqueeze(1).expand_as(dy)
+    # Compute width/height of each roi
+    pw = (rois[:, 2] - rois[:, 0] + 1.0).unsqueeze(1).expand_as(dw)
+    ph = (rois[:, 3] - rois[:, 1] + 1.0).unsqueeze(1).expand_as(dh)
+    # Use exp(network energy) to enlarge/shrink each roi
+    gw = pw * dw.exp()
+    gh = ph * dh.exp()
+    # Use network energy to shift the center of each roi
+    gx = px + pw * dx
+    gy = py + ph * dy
+    # Convert center-xy/width/height to top-left, bottom-right
+
+    # The true legacy box coder should +- 0.5 here.
+    # However, current implementation improves the performance when testing
+    # the models trained in MMDetection 1.X (~0.5 bbox AP, 0.2 mask AP)
+    x1 = gx - gw * 0.5
+    y1 = gy - gh * 0.5
+    x2 = gx + gw * 0.5
+    y2 = gy + gh * 0.5
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1] - 1)
+        y1 = y1.clamp(min=0, max=max_shape[0] - 1)
+        x2 = x2.clamp(min=0, max=max_shape[1] - 1)
+        y2 = y2.clamp(min=0, max=max_shape[0] - 1)
+    bboxes = torch.stack([x1, y1, x2, y2], dim=-1).view_as(deltas)
+    return bboxes
diff --git a/detection/scrfd/mmdet/core/bbox/coder/pseudo_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/pseudo_bbox_coder.py
new file mode 100755
index 000000000..1c8346f4a
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/pseudo_bbox_coder.py
@@ -0,0 +1,18 @@
+from ..builder import BBOX_CODERS
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class PseudoBBoxCoder(BaseBBoxCoder):
+    """Pseudo bounding box coder."""
+
+    def __init__(self, **kwargs):
+        super(BaseBBoxCoder, self).__init__(**kwargs)
+
+    def encode(self, bboxes, gt_bboxes):
+        """torch.Tensor: return the given ``bboxes``"""
+        return gt_bboxes
+
+    def decode(self, bboxes, pred_bboxes):
+        """torch.Tensor: return the given ``pred_bboxes``"""
+        return pred_bboxes
diff --git a/detection/scrfd/mmdet/core/bbox/coder/tblr_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/tblr_bbox_coder.py
new file mode 100755
index 000000000..436670b4c
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/tblr_bbox_coder.py
@@ -0,0 +1,172 @@
+import torch
+
+from ..builder import BBOX_CODERS
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class TBLRBBoxCoder(BaseBBoxCoder):
+    """TBLR BBox coder.
+
+    Following the practice in `FSAF <https://arxiv.org/abs/1903.00621>`_,
+    this coder encodes gt bboxes (x1, y1, x2, y2) into (top, bottom, left,
+    right) and decode it back to the original.
+
+    Args:
+        normalizer (list | float): Normalization factor to be
+          divided with when coding the coordinates. If it is a list, it should
+          have length of 4 indicating normalization factor in tblr dims.
+          Otherwise it is a unified float factor for all dims. Default: 4.0
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+    """
+
+    def __init__(self, normalizer=4.0, clip_border=True):
+        super(BaseBBoxCoder, self).__init__()
+        self.normalizer = normalizer
+        self.clip_border = clip_border
+
+    def encode(self, bboxes, gt_bboxes):
+        """Get box regression transformation deltas that can be used to
+        transform the ``bboxes`` into the ``gt_bboxes`` in the (top, left,
+        bottom, right) order.
+
+        Args:
+            bboxes (torch.Tensor): source boxes, e.g., object proposals.
+            gt_bboxes (torch.Tensor): target of the transformation, e.g.,
+                ground truth boxes.
+
+        Returns:
+            torch.Tensor: Box transformation deltas
+        """
+        assert bboxes.size(0) == gt_bboxes.size(0)
+        assert bboxes.size(-1) == gt_bboxes.size(-1) == 4
+        encoded_bboxes = bboxes2tblr(
+            bboxes, gt_bboxes, normalizer=self.normalizer)
+        return encoded_bboxes
+
+    def decode(self, bboxes, pred_bboxes, max_shape=None):
+        """Apply transformation `pred_bboxes` to `boxes`.
+
+        Args:
+            boxes (torch.Tensor): Basic boxes.
+            pred_bboxes (torch.Tensor): Encoded boxes with shape
+            max_shape (tuple[int], optional): Maximum shape of boxes.
+                Defaults to None.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+        assert pred_bboxes.size(0) == bboxes.size(0)
+        decoded_bboxes = tblr2bboxes(
+            bboxes,
+            pred_bboxes,
+            normalizer=self.normalizer,
+            max_shape=max_shape,
+            clip_border=self.clip_border)
+
+        return decoded_bboxes
+
+
+def bboxes2tblr(priors, gts, normalizer=4.0, normalize_by_wh=True):
+    """Encode ground truth boxes to tblr coordinate.
+
+    It first convert the gt coordinate to tblr format,
+     (top, bottom, left, right), relative to prior box centers.
+     The tblr coordinate may be normalized by the side length of prior bboxes
+     if `normalize_by_wh` is specified as True, and it is then normalized by
+     the `normalizer` factor.
+
+    Args:
+        priors (Tensor): Prior boxes in point form
+            Shape: (num_proposals,4).
+        gts (Tensor): Coords of ground truth for each prior in point-form
+            Shape: (num_proposals, 4).
+        normalizer (Sequence[float] | float): normalization parameter of
+            encoded boxes. If it is a list, it has to have length = 4.
+            Default: 4.0
+        normalize_by_wh (bool): Whether to normalize tblr coordinate by the
+            side length (wh) of prior bboxes.
+
+    Return:
+        encoded boxes (Tensor), Shape: (num_proposals, 4)
+    """
+
+    # dist b/t match center and prior's center
+    if not isinstance(normalizer, float):
+        normalizer = torch.tensor(normalizer, device=priors.device)
+        assert len(normalizer) == 4, 'Normalizer must have length = 4'
+    assert priors.size(0) == gts.size(0)
+    prior_centers = (priors[:, 0:2] + priors[:, 2:4]) / 2
+    xmin, ymin, xmax, ymax = gts.split(1, dim=1)
+    top = prior_centers[:, 1].unsqueeze(1) - ymin
+    bottom = ymax - prior_centers[:, 1].unsqueeze(1)
+    left = prior_centers[:, 0].unsqueeze(1) - xmin
+    right = xmax - prior_centers[:, 0].unsqueeze(1)
+    loc = torch.cat((top, bottom, left, right), dim=1)
+    if normalize_by_wh:
+        # Normalize tblr by anchor width and height
+        wh = priors[:, 2:4] - priors[:, 0:2]
+        w, h = torch.split(wh, 1, dim=1)
+        loc[:, :2] /= h  # tb is normalized by h
+        loc[:, 2:] /= w  # lr is normalized by w
+    # Normalize tblr by the given normalization factor
+    return loc / normalizer
+
+
+def tblr2bboxes(priors,
+                tblr,
+                normalizer=4.0,
+                normalize_by_wh=True,
+                max_shape=None,
+                clip_border=True):
+    """Decode tblr outputs to prediction boxes.
+
+    The process includes 3 steps: 1) De-normalize tblr coordinates by
+    multiplying it with `normalizer`; 2) De-normalize tblr coordinates by the
+    prior bbox width and height if `normalize_by_wh` is `True`; 3) Convert
+    tblr (top, bottom, left, right) pair relative to the center of priors back
+    to (xmin, ymin, xmax, ymax) coordinate.
+
+    Args:
+        priors (Tensor): Prior boxes in point form (x0, y0, x1, y1)
+          Shape: (n,4).
+        tblr (Tensor): Coords of network output in tblr form
+          Shape: (n, 4).
+        normalizer (Sequence[float] | float): Normalization parameter of
+          encoded boxes. By list, it represents the normalization factors at
+          tblr dims. By float, it is the unified normalization factor at all
+          dims. Default: 4.0
+        normalize_by_wh (bool): Whether the tblr coordinates have been
+          normalized by the side length (wh) of prior bboxes.
+        max_shape (tuple, optional): Shape of the image. Decoded bboxes
+          exceeding which will be clamped.
+        clip_border (bool, optional): Whether clip the objects outside the
+            border of the image. Defaults to True.
+
+    Return:
+        encoded boxes (Tensor), Shape: (n, 4)
+    """
+    if not isinstance(normalizer, float):
+        normalizer = torch.tensor(normalizer, device=priors.device)
+        assert len(normalizer) == 4, 'Normalizer must have length = 4'
+    assert priors.size(0) == tblr.size(0)
+    loc_decode = tblr * normalizer
+    prior_centers = (priors[:, 0:2] + priors[:, 2:4]) / 2
+    if normalize_by_wh:
+        wh = priors[:, 2:4] - priors[:, 0:2]
+        w, h = torch.split(wh, 1, dim=1)
+        loc_decode[:, :2] *= h  # tb
+        loc_decode[:, 2:] *= w  # lr
+    top, bottom, left, right = loc_decode.split((1, 1, 1, 1), dim=1)
+    xmin = prior_centers[:, 0].unsqueeze(1) - left
+    xmax = prior_centers[:, 0].unsqueeze(1) + right
+    ymin = prior_centers[:, 1].unsqueeze(1) - top
+    ymax = prior_centers[:, 1].unsqueeze(1) + bottom
+    boxes = torch.cat((xmin, ymin, xmax, ymax), dim=1)
+    if clip_border and max_shape is not None:
+        boxes[:, 0].clamp_(min=0, max=max_shape[1])
+        boxes[:, 1].clamp_(min=0, max=max_shape[0])
+        boxes[:, 2].clamp_(min=0, max=max_shape[1])
+        boxes[:, 3].clamp_(min=0, max=max_shape[0])
+    return boxes
diff --git a/detection/scrfd/mmdet/core/bbox/coder/yolo_bbox_coder.py b/detection/scrfd/mmdet/core/bbox/coder/yolo_bbox_coder.py
new file mode 100755
index 000000000..2a1dc34fc
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/coder/yolo_bbox_coder.py
@@ -0,0 +1,86 @@
+import torch
+
+from ..builder import BBOX_CODERS
+from .base_bbox_coder import BaseBBoxCoder
+
+
+@BBOX_CODERS.register_module()
+class YOLOBBoxCoder(BaseBBoxCoder):
+    """YOLO BBox coder.
+
+    Following `YOLO <https://arxiv.org/abs/1506.02640>`_, this coder divide
+    image into grids, and encode bbox (x1, y1, x2, y2) into (cx, cy, dw, dh).
+    cx, cy in [0., 1.], denotes relative center position w.r.t the center of
+    bboxes. dw, dh are the same as :obj:`DeltaXYWHBBoxCoder`.
+
+    Args:
+        eps (float): Min value of cx, cy when encoding.
+    """
+
+    def __init__(self, eps=1e-6):
+        super(BaseBBoxCoder, self).__init__()
+        self.eps = eps
+
+    def encode(self, bboxes, gt_bboxes, stride):
+        """Get box regression transformation deltas that can be used to
+        transform the ``bboxes`` into the ``gt_bboxes``.
+
+        Args:
+            bboxes (torch.Tensor): Source boxes, e.g., anchors.
+            gt_bboxes (torch.Tensor): Target of the transformation, e.g.,
+                ground-truth boxes.
+            stride (torch.Tensor | int): Stride of bboxes.
+
+        Returns:
+            torch.Tensor: Box transformation deltas
+        """
+
+        assert bboxes.size(0) == gt_bboxes.size(0)
+        assert bboxes.size(-1) == gt_bboxes.size(-1) == 4
+        x_center_gt = (gt_bboxes[..., 0] + gt_bboxes[..., 2]) * 0.5
+        y_center_gt = (gt_bboxes[..., 1] + gt_bboxes[..., 3]) * 0.5
+        w_gt = gt_bboxes[..., 2] - gt_bboxes[..., 0]
+        h_gt = gt_bboxes[..., 3] - gt_bboxes[..., 1]
+        x_center = (bboxes[..., 0] + bboxes[..., 2]) * 0.5
+        y_center = (bboxes[..., 1] + bboxes[..., 3]) * 0.5
+        w = bboxes[..., 2] - bboxes[..., 0]
+        h = bboxes[..., 3] - bboxes[..., 1]
+        w_target = torch.log((w_gt / w).clamp(min=self.eps))
+        h_target = torch.log((h_gt / h).clamp(min=self.eps))
+        x_center_target = ((x_center_gt - x_center) / stride + 0.5).clamp(
+            self.eps, 1 - self.eps)
+        y_center_target = ((y_center_gt - y_center) / stride + 0.5).clamp(
+            self.eps, 1 - self.eps)
+        encoded_bboxes = torch.stack(
+            [x_center_target, y_center_target, w_target, h_target], dim=-1)
+        return encoded_bboxes
+
+    def decode(self, bboxes, pred_bboxes, stride):
+        """Apply transformation `pred_bboxes` to `boxes`.
+
+        Args:
+            boxes (torch.Tensor): Basic boxes, e.g. anchors.
+            pred_bboxes (torch.Tensor): Encoded boxes with shape
+            stride (torch.Tensor | int): Strides of bboxes.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+        assert pred_bboxes.size(0) == bboxes.size(0)
+        assert pred_bboxes.size(-1) == bboxes.size(-1) == 4
+        x_center = (bboxes[..., 0] + bboxes[..., 2]) * 0.5
+        y_center = (bboxes[..., 1] + bboxes[..., 3]) * 0.5
+        w = bboxes[..., 2] - bboxes[..., 0]
+        h = bboxes[..., 3] - bboxes[..., 1]
+        # Get outputs x, y
+        x_center_pred = (pred_bboxes[..., 0] - 0.5) * stride + x_center
+        y_center_pred = (pred_bboxes[..., 1] - 0.5) * stride + y_center
+        w_pred = torch.exp(pred_bboxes[..., 2]) * w
+        h_pred = torch.exp(pred_bboxes[..., 3]) * h
+
+        decoded_bboxes = torch.stack(
+            (x_center_pred - w_pred / 2, y_center_pred - h_pred / 2,
+             x_center_pred + w_pred / 2, y_center_pred + h_pred / 2),
+            dim=-1)
+
+        return decoded_bboxes
diff --git a/detection/scrfd/mmdet/core/bbox/demodata.py b/detection/scrfd/mmdet/core/bbox/demodata.py
new file mode 100755
index 000000000..943085859
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/demodata.py
@@ -0,0 +1,63 @@
+import numpy as np
+import torch
+
+
+def ensure_rng(rng=None):
+    """Simple version of the ``kwarray.ensure_rng``
+
+    Args:
+        rng (int | numpy.random.RandomState | None):
+            if None, then defaults to the global rng. Otherwise this can be an
+            integer or a RandomState class
+    Returns:
+        (numpy.random.RandomState) : rng -
+            a numpy random number generator
+
+    References:
+        https://gitlab.kitware.com/computer-vision/kwarray/blob/master/kwarray/util_random.py#L270
+    """
+
+    if rng is None:
+        rng = np.random.mtrand._rand
+    elif isinstance(rng, int):
+        rng = np.random.RandomState(rng)
+    else:
+        rng = rng
+    return rng
+
+
+def random_boxes(num=1, scale=1, rng=None):
+    """Simple version of ``kwimage.Boxes.random``
+
+    Returns:
+        Tensor: shape (n, 4) in x1, y1, x2, y2 format.
+
+    References:
+        https://gitlab.kitware.com/computer-vision/kwimage/blob/master/kwimage/structs/boxes.py#L1390
+
+    Example:
+        >>> num = 3
+        >>> scale = 512
+        >>> rng = 0
+        >>> boxes = random_boxes(num, scale, rng)
+        >>> print(boxes)
+        tensor([[280.9925, 278.9802, 308.6148, 366.1769],
+                [216.9113, 330.6978, 224.0446, 456.5878],
+                [405.3632, 196.3221, 493.3953, 270.7942]])
+    """
+    rng = ensure_rng(rng)
+
+    tlbr = rng.rand(num, 4).astype(np.float32)
+
+    tl_x = np.minimum(tlbr[:, 0], tlbr[:, 2])
+    tl_y = np.minimum(tlbr[:, 1], tlbr[:, 3])
+    br_x = np.maximum(tlbr[:, 0], tlbr[:, 2])
+    br_y = np.maximum(tlbr[:, 1], tlbr[:, 3])
+
+    tlbr[:, 0] = tl_x * scale
+    tlbr[:, 1] = tl_y * scale
+    tlbr[:, 2] = br_x * scale
+    tlbr[:, 3] = br_y * scale
+
+    boxes = torch.from_numpy(tlbr)
+    return boxes
diff --git a/detection/scrfd/mmdet/core/bbox/iou_calculators/__init__.py b/detection/scrfd/mmdet/core/bbox/iou_calculators/__init__.py
new file mode 100755
index 000000000..e71369a58
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/iou_calculators/__init__.py
@@ -0,0 +1,4 @@
+from .builder import build_iou_calculator
+from .iou2d_calculator import BboxOverlaps2D, bbox_overlaps
+
+__all__ = ['build_iou_calculator', 'BboxOverlaps2D', 'bbox_overlaps']
diff --git a/detection/scrfd/mmdet/core/bbox/iou_calculators/builder.py b/detection/scrfd/mmdet/core/bbox/iou_calculators/builder.py
new file mode 100755
index 000000000..09094d7ec
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/iou_calculators/builder.py
@@ -0,0 +1,8 @@
+from mmcv.utils import Registry, build_from_cfg
+
+IOU_CALCULATORS = Registry('IoU calculator')
+
+
+def build_iou_calculator(cfg, default_args=None):
+    """Builder of IoU calculator."""
+    return build_from_cfg(cfg, IOU_CALCULATORS, default_args)
diff --git a/detection/scrfd/mmdet/core/bbox/iou_calculators/iou2d_calculator.py b/detection/scrfd/mmdet/core/bbox/iou_calculators/iou2d_calculator.py
new file mode 100755
index 000000000..97adcca5e
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/iou_calculators/iou2d_calculator.py
@@ -0,0 +1,159 @@
+import torch
+
+from .builder import IOU_CALCULATORS
+
+
+@IOU_CALCULATORS.register_module()
+class BboxOverlaps2D(object):
+    """2D Overlaps (e.g. IoUs, GIoUs) Calculator."""
+
+    def __call__(self, bboxes1, bboxes2, mode='iou', is_aligned=False):
+        """Calculate IoU between 2D bboxes.
+
+        Args:
+            bboxes1 (Tensor): bboxes have shape (m, 4) in <x1, y1, x2, y2>
+                format, or shape (m, 5) in <x1, y1, x2, y2, score> format.
+            bboxes2 (Tensor): bboxes have shape (m, 4) in <x1, y1, x2, y2>
+                format, shape (m, 5) in <x1, y1, x2, y2, score> format, or be
+                empty. If ``is_aligned `` is ``True``, then m and n must be
+                equal.
+            mode (str): "iou" (intersection over union), "iof" (intersection
+                over foreground), or "giou" (generalized intersection over
+                union).
+            is_aligned (bool, optional): If True, then m and n must be equal.
+                Default False.
+
+        Returns:
+            Tensor: shape (m, n) if ``is_aligned `` is False else shape (m,)
+        """
+        assert bboxes1.size(-1) in [0, 4, 5]
+        assert bboxes2.size(-1) in [0, 4, 5]
+        if bboxes2.size(-1) == 5:
+            bboxes2 = bboxes2[..., :4]
+        if bboxes1.size(-1) == 5:
+            bboxes1 = bboxes1[..., :4]
+        return bbox_overlaps(bboxes1, bboxes2, mode, is_aligned)
+
+    def __repr__(self):
+        """str: a string describing the module"""
+        repr_str = self.__class__.__name__ + '()'
+        return repr_str
+
+
+def bbox_overlaps(bboxes1, bboxes2, mode='iou', is_aligned=False, eps=1e-6):
+    """Calculate overlap between two set of bboxes.
+
+    If ``is_aligned `` is ``False``, then calculate the overlaps between each
+    bbox of bboxes1 and bboxes2, otherwise the overlaps between each aligned
+    pair of bboxes1 and bboxes2.
+
+    Args:
+        bboxes1 (Tensor): shape (B, m, 4) in <x1, y1, x2, y2> format or empty.
+        bboxes2 (Tensor): shape (B, n, 4) in <x1, y1, x2, y2> format or empty.
+            B indicates the batch dim, in shape (B1, B2, ..., Bn).
+            If ``is_aligned `` is ``True``, then m and n must be equal.
+        mode (str): "iou" (intersection over union), "iof" (intersection over
+            foreground) or "giou" (generalized intersection over union).
+            Default "iou".
+        is_aligned (bool, optional): If True, then m and n must be equal.
+            Default False.
+        eps (float, optional): A value added to the denominator for numerical
+            stability. Default 1e-6.
+
+    Returns:
+        Tensor: shape (m, n) if ``is_aligned `` is False else shape (m,)
+
+    Example:
+        >>> bboxes1 = torch.FloatTensor([
+        >>>     [0, 0, 10, 10],
+        >>>     [10, 10, 20, 20],
+        >>>     [32, 32, 38, 42],
+        >>> ])
+        >>> bboxes2 = torch.FloatTensor([
+        >>>     [0, 0, 10, 20],
+        >>>     [0, 10, 10, 19],
+        >>>     [10, 10, 20, 20],
+        >>> ])
+        >>> overlaps = bbox_overlaps(bboxes1, bboxes2)
+        >>> assert overlaps.shape == (3, 3)
+        >>> overlaps = bbox_overlaps(bboxes1, bboxes2, is_aligned=True)
+        >>> assert overlaps.shape == (3, )
+
+    Example:
+        >>> empty = torch.empty(0, 4)
+        >>> nonempty = torch.FloatTensor([[0, 0, 10, 9]])
+        >>> assert tuple(bbox_overlaps(empty, nonempty).shape) == (0, 1)
+        >>> assert tuple(bbox_overlaps(nonempty, empty).shape) == (1, 0)
+        >>> assert tuple(bbox_overlaps(empty, empty).shape) == (0, 0)
+    """
+
+    assert mode in ['iou', 'iof', 'giou'], f'Unsupported mode {mode}'
+    # Either the boxes are empty or the length of boxes's last dimenstion is 4
+    assert (bboxes1.size(-1) == 4 or bboxes1.size(0) == 0)
+    assert (bboxes2.size(-1) == 4 or bboxes2.size(0) == 0)
+
+    # Batch dim must be the same
+    # Batch dim: (B1, B2, ... Bn)
+    assert bboxes1.shape[:-2] == bboxes2.shape[:-2]
+    batch_shape = bboxes1.shape[:-2]
+
+    rows = bboxes1.size(-2)
+    cols = bboxes2.size(-2)
+    if is_aligned:
+        assert rows == cols
+
+    if rows * cols == 0:
+        if is_aligned:
+            return bboxes1.new(batch_shape + (rows, ))
+        else:
+            return bboxes1.new(batch_shape + (rows, cols))
+
+    area1 = (bboxes1[..., 2] - bboxes1[..., 0]) * (
+        bboxes1[..., 3] - bboxes1[..., 1])
+    area2 = (bboxes2[..., 2] - bboxes2[..., 0]) * (
+        bboxes2[..., 3] - bboxes2[..., 1])
+
+    if is_aligned:
+        lt = torch.max(bboxes1[..., :2], bboxes2[..., :2])  # [B, rows, 2]
+        rb = torch.min(bboxes1[..., 2:], bboxes2[..., 2:])  # [B, rows, 2]
+
+        wh = (rb - lt).clamp(min=0)  # [B, rows, 2]
+        overlap = wh[..., 0] * wh[..., 1]
+
+        if mode in ['iou', 'giou']:
+            union = area1 + area2 - overlap
+        else:
+            union = area1
+        if mode == 'giou':
+            enclosed_lt = torch.min(bboxes1[..., :2], bboxes2[..., :2])
+            enclosed_rb = torch.max(bboxes1[..., 2:], bboxes2[..., 2:])
+    else:
+        lt = torch.max(bboxes1[..., :, None, :2],
+                       bboxes2[..., None, :, :2])  # [B, rows, cols, 2]
+        rb = torch.min(bboxes1[..., :, None, 2:],
+                       bboxes2[..., None, :, 2:])  # [B, rows, cols, 2]
+
+        wh = (rb - lt).clamp(min=0)  # [B, rows, cols, 2]
+        overlap = wh[..., 0] * wh[..., 1]
+
+        if mode in ['iou', 'giou']:
+            union = area1[..., None] + area2[..., None, :] - overlap
+        else:
+            union = area1[..., None]
+        if mode == 'giou':
+            enclosed_lt = torch.min(bboxes1[..., :, None, :2],
+                                    bboxes2[..., None, :, :2])
+            enclosed_rb = torch.max(bboxes1[..., :, None, 2:],
+                                    bboxes2[..., None, :, 2:])
+
+    eps = union.new_tensor([eps])
+    union = torch.max(union, eps)
+    ious = overlap / union
+    if mode in ['iou', 'iof']:
+        return ious
+    # calculate gious
+    enclose_wh = (enclosed_rb - enclosed_lt).clamp(min=0)
+    enclose_area = enclose_wh[..., 0] * enclose_wh[..., 1]
+    enclose_area = torch.max(enclose_area, eps)
+    gious = ious - (enclose_area - union) / enclose_area
+    return gious
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/__init__.py b/detection/scrfd/mmdet/core/bbox/samplers/__init__.py
new file mode 100755
index 000000000..0b06303fe
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/__init__.py
@@ -0,0 +1,15 @@
+from .base_sampler import BaseSampler
+from .combined_sampler import CombinedSampler
+from .instance_balanced_pos_sampler import InstanceBalancedPosSampler
+from .iou_balanced_neg_sampler import IoUBalancedNegSampler
+from .ohem_sampler import OHEMSampler
+from .pseudo_sampler import PseudoSampler
+from .random_sampler import RandomSampler
+from .sampling_result import SamplingResult
+from .score_hlr_sampler import ScoreHLRSampler
+
+__all__ = [
+    'BaseSampler', 'PseudoSampler', 'RandomSampler',
+    'InstanceBalancedPosSampler', 'IoUBalancedNegSampler', 'CombinedSampler',
+    'OHEMSampler', 'SamplingResult', 'ScoreHLRSampler'
+]
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/base_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/base_sampler.py
new file mode 100755
index 000000000..9ea35def1
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/base_sampler.py
@@ -0,0 +1,101 @@
+from abc import ABCMeta, abstractmethod
+
+import torch
+
+from .sampling_result import SamplingResult
+
+
+class BaseSampler(metaclass=ABCMeta):
+    """Base class of samplers."""
+
+    def __init__(self,
+                 num,
+                 pos_fraction,
+                 neg_pos_ub=-1,
+                 add_gt_as_proposals=True,
+                 **kwargs):
+        self.num = num
+        self.pos_fraction = pos_fraction
+        self.neg_pos_ub = neg_pos_ub
+        self.add_gt_as_proposals = add_gt_as_proposals
+        self.pos_sampler = self
+        self.neg_sampler = self
+
+    @abstractmethod
+    def _sample_pos(self, assign_result, num_expected, **kwargs):
+        """Sample positive samples."""
+        pass
+
+    @abstractmethod
+    def _sample_neg(self, assign_result, num_expected, **kwargs):
+        """Sample negative samples."""
+        pass
+
+    def sample(self,
+               assign_result,
+               bboxes,
+               gt_bboxes,
+               gt_labels=None,
+               **kwargs):
+        """Sample positive and negative bboxes.
+
+        This is a simple implementation of bbox sampling given candidates,
+        assigning results and ground truth bboxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): Bbox assigning results.
+            bboxes (Tensor): Boxes to be sampled from.
+            gt_bboxes (Tensor): Ground truth bboxes.
+            gt_labels (Tensor, optional): Class labels of ground truth bboxes.
+
+        Returns:
+            :obj:`SamplingResult`: Sampling result.
+
+        Example:
+            >>> from mmdet.core.bbox import RandomSampler
+            >>> from mmdet.core.bbox import AssignResult
+            >>> from mmdet.core.bbox.demodata import ensure_rng, random_boxes
+            >>> rng = ensure_rng(None)
+            >>> assign_result = AssignResult.random(rng=rng)
+            >>> bboxes = random_boxes(assign_result.num_preds, rng=rng)
+            >>> gt_bboxes = random_boxes(assign_result.num_gts, rng=rng)
+            >>> gt_labels = None
+            >>> self = RandomSampler(num=32, pos_fraction=0.5, neg_pos_ub=-1,
+            >>>                      add_gt_as_proposals=False)
+            >>> self = self.sample(assign_result, bboxes, gt_bboxes, gt_labels)
+        """
+        if len(bboxes.shape) < 2:
+            bboxes = bboxes[None, :]
+
+        bboxes = bboxes[:, :4]
+
+        gt_flags = bboxes.new_zeros((bboxes.shape[0], ), dtype=torch.uint8)
+        if self.add_gt_as_proposals and len(gt_bboxes) > 0:
+            if gt_labels is None:
+                raise ValueError(
+                    'gt_labels must be given when add_gt_as_proposals is True')
+            bboxes = torch.cat([gt_bboxes, bboxes], dim=0)
+            assign_result.add_gt_(gt_labels)
+            gt_ones = bboxes.new_ones(gt_bboxes.shape[0], dtype=torch.uint8)
+            gt_flags = torch.cat([gt_ones, gt_flags])
+
+        num_expected_pos = int(self.num * self.pos_fraction)
+        pos_inds = self.pos_sampler._sample_pos(
+            assign_result, num_expected_pos, bboxes=bboxes, **kwargs)
+        # We found that sampled indices have duplicated items occasionally.
+        # (may be a bug of PyTorch)
+        pos_inds = pos_inds.unique()
+        num_sampled_pos = pos_inds.numel()
+        num_expected_neg = self.num - num_sampled_pos
+        if self.neg_pos_ub >= 0:
+            _pos = max(1, num_sampled_pos)
+            neg_upper_bound = int(self.neg_pos_ub * _pos)
+            if num_expected_neg > neg_upper_bound:
+                num_expected_neg = neg_upper_bound
+        neg_inds = self.neg_sampler._sample_neg(
+            assign_result, num_expected_neg, bboxes=bboxes, **kwargs)
+        neg_inds = neg_inds.unique()
+
+        sampling_result = SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes,
+                                         assign_result, gt_flags)
+        return sampling_result
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/combined_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/combined_sampler.py
new file mode 100755
index 000000000..564729f08
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/combined_sampler.py
@@ -0,0 +1,20 @@
+from ..builder import BBOX_SAMPLERS, build_sampler
+from .base_sampler import BaseSampler
+
+
+@BBOX_SAMPLERS.register_module()
+class CombinedSampler(BaseSampler):
+    """A sampler that combines positive sampler and negative sampler."""
+
+    def __init__(self, pos_sampler, neg_sampler, **kwargs):
+        super(CombinedSampler, self).__init__(**kwargs)
+        self.pos_sampler = build_sampler(pos_sampler, **kwargs)
+        self.neg_sampler = build_sampler(neg_sampler, **kwargs)
+
+    def _sample_pos(self, **kwargs):
+        """Sample positive samples."""
+        raise NotImplementedError
+
+    def _sample_neg(self, **kwargs):
+        """Sample negative samples."""
+        raise NotImplementedError
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/instance_balanced_pos_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/instance_balanced_pos_sampler.py
new file mode 100755
index 000000000..c73529848
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/instance_balanced_pos_sampler.py
@@ -0,0 +1,55 @@
+import numpy as np
+import torch
+
+from ..builder import BBOX_SAMPLERS
+from .random_sampler import RandomSampler
+
+
+@BBOX_SAMPLERS.register_module()
+class InstanceBalancedPosSampler(RandomSampler):
+    """Instance balanced sampler that samples equal number of positive samples
+    for each instance."""
+
+    def _sample_pos(self, assign_result, num_expected, **kwargs):
+        """Sample positive boxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): The assigned results of boxes.
+            num_expected (int): The number of expected positive samples
+
+        Returns:
+            Tensor or ndarray: sampled indices.
+        """
+        pos_inds = torch.nonzero(assign_result.gt_inds > 0, as_tuple=False)
+        if pos_inds.numel() != 0:
+            pos_inds = pos_inds.squeeze(1)
+        if pos_inds.numel() <= num_expected:
+            return pos_inds
+        else:
+            unique_gt_inds = assign_result.gt_inds[pos_inds].unique()
+            num_gts = len(unique_gt_inds)
+            num_per_gt = int(round(num_expected / float(num_gts)) + 1)
+            sampled_inds = []
+            for i in unique_gt_inds:
+                inds = torch.nonzero(
+                    assign_result.gt_inds == i.item(), as_tuple=False)
+                if inds.numel() != 0:
+                    inds = inds.squeeze(1)
+                else:
+                    continue
+                if len(inds) > num_per_gt:
+                    inds = self.random_choice(inds, num_per_gt)
+                sampled_inds.append(inds)
+            sampled_inds = torch.cat(sampled_inds)
+            if len(sampled_inds) < num_expected:
+                num_extra = num_expected - len(sampled_inds)
+                extra_inds = np.array(
+                    list(set(pos_inds.cpu()) - set(sampled_inds.cpu())))
+                if len(extra_inds) > num_extra:
+                    extra_inds = self.random_choice(extra_inds, num_extra)
+                extra_inds = torch.from_numpy(extra_inds).to(
+                    assign_result.gt_inds.device).long()
+                sampled_inds = torch.cat([sampled_inds, extra_inds])
+            elif len(sampled_inds) > num_expected:
+                sampled_inds = self.random_choice(sampled_inds, num_expected)
+            return sampled_inds
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/iou_balanced_neg_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/iou_balanced_neg_sampler.py
new file mode 100755
index 000000000..f275e430d
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/iou_balanced_neg_sampler.py
@@ -0,0 +1,157 @@
+import numpy as np
+import torch
+
+from ..builder import BBOX_SAMPLERS
+from .random_sampler import RandomSampler
+
+
+@BBOX_SAMPLERS.register_module()
+class IoUBalancedNegSampler(RandomSampler):
+    """IoU Balanced Sampling.
+
+    arXiv: https://arxiv.org/pdf/1904.02701.pdf (CVPR 2019)
+
+    Sampling proposals according to their IoU. `floor_fraction` of needed RoIs
+    are sampled from proposals whose IoU are lower than `floor_thr` randomly.
+    The others are sampled from proposals whose IoU are higher than
+    `floor_thr`. These proposals are sampled from some bins evenly, which are
+    split by `num_bins` via IoU evenly.
+
+    Args:
+        num (int): number of proposals.
+        pos_fraction (float): fraction of positive proposals.
+        floor_thr (float): threshold (minimum) IoU for IoU balanced sampling,
+            set to -1 if all using IoU balanced sampling.
+        floor_fraction (float): sampling fraction of proposals under floor_thr.
+        num_bins (int): number of bins in IoU balanced sampling.
+    """
+
+    def __init__(self,
+                 num,
+                 pos_fraction,
+                 floor_thr=-1,
+                 floor_fraction=0,
+                 num_bins=3,
+                 **kwargs):
+        super(IoUBalancedNegSampler, self).__init__(num, pos_fraction,
+                                                    **kwargs)
+        assert floor_thr >= 0 or floor_thr == -1
+        assert 0 <= floor_fraction <= 1
+        assert num_bins >= 1
+
+        self.floor_thr = floor_thr
+        self.floor_fraction = floor_fraction
+        self.num_bins = num_bins
+
+    def sample_via_interval(self, max_overlaps, full_set, num_expected):
+        """Sample according to the iou interval.
+
+        Args:
+            max_overlaps (torch.Tensor): IoU between bounding boxes and ground
+                truth boxes.
+            full_set (set(int)): A full set of indices of boxes。
+            num_expected (int): Number of expected samples。
+
+        Returns:
+            np.ndarray: Indices  of samples
+        """
+        max_iou = max_overlaps.max()
+        iou_interval = (max_iou - self.floor_thr) / self.num_bins
+        per_num_expected = int(num_expected / self.num_bins)
+
+        sampled_inds = []
+        for i in range(self.num_bins):
+            start_iou = self.floor_thr + i * iou_interval
+            end_iou = self.floor_thr + (i + 1) * iou_interval
+            tmp_set = set(
+                np.where(
+                    np.logical_and(max_overlaps >= start_iou,
+                                   max_overlaps < end_iou))[0])
+            tmp_inds = list(tmp_set & full_set)
+            if len(tmp_inds) > per_num_expected:
+                tmp_sampled_set = self.random_choice(tmp_inds,
+                                                     per_num_expected)
+            else:
+                tmp_sampled_set = np.array(tmp_inds, dtype=np.int)
+            sampled_inds.append(tmp_sampled_set)
+
+        sampled_inds = np.concatenate(sampled_inds)
+        if len(sampled_inds) < num_expected:
+            num_extra = num_expected - len(sampled_inds)
+            extra_inds = np.array(list(full_set - set(sampled_inds)))
+            if len(extra_inds) > num_extra:
+                extra_inds = self.random_choice(extra_inds, num_extra)
+            sampled_inds = np.concatenate([sampled_inds, extra_inds])
+
+        return sampled_inds
+
+    def _sample_neg(self, assign_result, num_expected, **kwargs):
+        """Sample negative boxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): The assigned results of boxes.
+            num_expected (int): The number of expected negative samples
+
+        Returns:
+            Tensor or ndarray: sampled indices.
+        """
+        neg_inds = torch.nonzero(assign_result.gt_inds == 0, as_tuple=False)
+        if neg_inds.numel() != 0:
+            neg_inds = neg_inds.squeeze(1)
+        if len(neg_inds) <= num_expected:
+            return neg_inds
+        else:
+            max_overlaps = assign_result.max_overlaps.cpu().numpy()
+            # balance sampling for negative samples
+            neg_set = set(neg_inds.cpu().numpy())
+
+            if self.floor_thr > 0:
+                floor_set = set(
+                    np.where(
+                        np.logical_and(max_overlaps >= 0,
+                                       max_overlaps < self.floor_thr))[0])
+                iou_sampling_set = set(
+                    np.where(max_overlaps >= self.floor_thr)[0])
+            elif self.floor_thr == 0:
+                floor_set = set(np.where(max_overlaps == 0)[0])
+                iou_sampling_set = set(
+                    np.where(max_overlaps > self.floor_thr)[0])
+            else:
+                floor_set = set()
+                iou_sampling_set = set(
+                    np.where(max_overlaps > self.floor_thr)[0])
+                # for sampling interval calculation
+                self.floor_thr = 0
+
+            floor_neg_inds = list(floor_set & neg_set)
+            iou_sampling_neg_inds = list(iou_sampling_set & neg_set)
+            num_expected_iou_sampling = int(num_expected *
+                                            (1 - self.floor_fraction))
+            if len(iou_sampling_neg_inds) > num_expected_iou_sampling:
+                if self.num_bins >= 2:
+                    iou_sampled_inds = self.sample_via_interval(
+                        max_overlaps, set(iou_sampling_neg_inds),
+                        num_expected_iou_sampling)
+                else:
+                    iou_sampled_inds = self.random_choice(
+                        iou_sampling_neg_inds, num_expected_iou_sampling)
+            else:
+                iou_sampled_inds = np.array(
+                    iou_sampling_neg_inds, dtype=np.int)
+            num_expected_floor = num_expected - len(iou_sampled_inds)
+            if len(floor_neg_inds) > num_expected_floor:
+                sampled_floor_inds = self.random_choice(
+                    floor_neg_inds, num_expected_floor)
+            else:
+                sampled_floor_inds = np.array(floor_neg_inds, dtype=np.int)
+            sampled_inds = np.concatenate(
+                (sampled_floor_inds, iou_sampled_inds))
+            if len(sampled_inds) < num_expected:
+                num_extra = num_expected - len(sampled_inds)
+                extra_inds = np.array(list(neg_set - set(sampled_inds)))
+                if len(extra_inds) > num_extra:
+                    extra_inds = self.random_choice(extra_inds, num_extra)
+                sampled_inds = np.concatenate((sampled_inds, extra_inds))
+            sampled_inds = torch.from_numpy(sampled_inds).long().to(
+                assign_result.gt_inds.device)
+            return sampled_inds
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/ohem_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/ohem_sampler.py
new file mode 100755
index 000000000..8b99f60ef
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/ohem_sampler.py
@@ -0,0 +1,107 @@
+import torch
+
+from ..builder import BBOX_SAMPLERS
+from ..transforms import bbox2roi
+from .base_sampler import BaseSampler
+
+
+@BBOX_SAMPLERS.register_module()
+class OHEMSampler(BaseSampler):
+    r"""Online Hard Example Mining Sampler described in `Training Region-based
+    Object Detectors with Online Hard Example Mining
+    <https://arxiv.org/abs/1604.03540>`_.
+    """
+
+    def __init__(self,
+                 num,
+                 pos_fraction,
+                 context,
+                 neg_pos_ub=-1,
+                 add_gt_as_proposals=True,
+                 **kwargs):
+        super(OHEMSampler, self).__init__(num, pos_fraction, neg_pos_ub,
+                                          add_gt_as_proposals)
+        self.context = context
+        if not hasattr(self.context, 'num_stages'):
+            self.bbox_head = self.context.bbox_head
+        else:
+            self.bbox_head = self.context.bbox_head[self.context.current_stage]
+
+    def hard_mining(self, inds, num_expected, bboxes, labels, feats):
+        with torch.no_grad():
+            rois = bbox2roi([bboxes])
+            if not hasattr(self.context, 'num_stages'):
+                bbox_results = self.context._bbox_forward(feats, rois)
+            else:
+                bbox_results = self.context._bbox_forward(
+                    self.context.current_stage, feats, rois)
+            cls_score = bbox_results['cls_score']
+            loss = self.bbox_head.loss(
+                cls_score=cls_score,
+                bbox_pred=None,
+                rois=rois,
+                labels=labels,
+                label_weights=cls_score.new_ones(cls_score.size(0)),
+                bbox_targets=None,
+                bbox_weights=None,
+                reduction_override='none')['loss_cls']
+            _, topk_loss_inds = loss.topk(num_expected)
+        return inds[topk_loss_inds]
+
+    def _sample_pos(self,
+                    assign_result,
+                    num_expected,
+                    bboxes=None,
+                    feats=None,
+                    **kwargs):
+        """Sample positive boxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): Assigned results
+            num_expected (int): Number of expected positive samples
+            bboxes (torch.Tensor, optional): Boxes. Defaults to None.
+            feats (list[torch.Tensor], optional): Multi-level features.
+                Defaults to None.
+
+        Returns:
+            torch.Tensor: Indices  of positive samples
+        """
+        # Sample some hard positive samples
+        pos_inds = torch.nonzero(assign_result.gt_inds > 0, as_tuple=False)
+        if pos_inds.numel() != 0:
+            pos_inds = pos_inds.squeeze(1)
+        if pos_inds.numel() <= num_expected:
+            return pos_inds
+        else:
+            return self.hard_mining(pos_inds, num_expected, bboxes[pos_inds],
+                                    assign_result.labels[pos_inds], feats)
+
+    def _sample_neg(self,
+                    assign_result,
+                    num_expected,
+                    bboxes=None,
+                    feats=None,
+                    **kwargs):
+        """Sample negative boxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): Assigned results
+            num_expected (int): Number of expected negative samples
+            bboxes (torch.Tensor, optional): Boxes. Defaults to None.
+            feats (list[torch.Tensor], optional): Multi-level features.
+                Defaults to None.
+
+        Returns:
+            torch.Tensor: Indices  of negative samples
+        """
+        # Sample some hard negative samples
+        neg_inds = torch.nonzero(assign_result.gt_inds == 0, as_tuple=False)
+        if neg_inds.numel() != 0:
+            neg_inds = neg_inds.squeeze(1)
+        if len(neg_inds) <= num_expected:
+            return neg_inds
+        else:
+            neg_labels = assign_result.labels.new_empty(
+                neg_inds.size(0)).fill_(self.bbox_head.num_classes)
+            return self.hard_mining(neg_inds, num_expected, bboxes[neg_inds],
+                                    neg_labels, feats)
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/pseudo_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/pseudo_sampler.py
new file mode 100755
index 000000000..2bd81abcd
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/pseudo_sampler.py
@@ -0,0 +1,41 @@
+import torch
+
+from ..builder import BBOX_SAMPLERS
+from .base_sampler import BaseSampler
+from .sampling_result import SamplingResult
+
+
+@BBOX_SAMPLERS.register_module()
+class PseudoSampler(BaseSampler):
+    """A pseudo sampler that does not do sampling actually."""
+
+    def __init__(self, **kwargs):
+        pass
+
+    def _sample_pos(self, **kwargs):
+        """Sample positive samples."""
+        raise NotImplementedError
+
+    def _sample_neg(self, **kwargs):
+        """Sample negative samples."""
+        raise NotImplementedError
+
+    def sample(self, assign_result, bboxes, gt_bboxes, **kwargs):
+        """Directly returns the positive and negative indices  of samples.
+
+        Args:
+            assign_result (:obj:`AssignResult`): Assigned results
+            bboxes (torch.Tensor): Bounding boxes
+            gt_bboxes (torch.Tensor): Ground truth boxes
+
+        Returns:
+            :obj:`SamplingResult`: sampler results
+        """
+        pos_inds = torch.nonzero(
+            assign_result.gt_inds > 0, as_tuple=False).squeeze(-1).unique()
+        neg_inds = torch.nonzero(
+            assign_result.gt_inds == 0, as_tuple=False).squeeze(-1).unique()
+        gt_flags = bboxes.new_zeros(bboxes.shape[0], dtype=torch.uint8)
+        sampling_result = SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes,
+                                         assign_result, gt_flags)
+        return sampling_result
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/random_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/random_sampler.py
new file mode 100755
index 000000000..f34b006e8
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/random_sampler.py
@@ -0,0 +1,78 @@
+import torch
+
+from ..builder import BBOX_SAMPLERS
+from .base_sampler import BaseSampler
+
+
+@BBOX_SAMPLERS.register_module()
+class RandomSampler(BaseSampler):
+    """Random sampler.
+
+    Args:
+        num (int): Number of samples
+        pos_fraction (float): Fraction of positive samples
+        neg_pos_up (int, optional): Upper bound number of negative and
+            positive samples. Defaults to -1.
+        add_gt_as_proposals (bool, optional): Whether to add ground truth
+            boxes as proposals. Defaults to True.
+    """
+
+    def __init__(self,
+                 num,
+                 pos_fraction,
+                 neg_pos_ub=-1,
+                 add_gt_as_proposals=True,
+                 **kwargs):
+        from mmdet.core.bbox import demodata
+        super(RandomSampler, self).__init__(num, pos_fraction, neg_pos_ub,
+                                            add_gt_as_proposals)
+        self.rng = demodata.ensure_rng(kwargs.get('rng', None))
+
+    def random_choice(self, gallery, num):
+        """Random select some elements from the gallery.
+
+        If `gallery` is a Tensor, the returned indices will be a Tensor;
+        If `gallery` is a ndarray or list, the returned indices will be a
+        ndarray.
+
+        Args:
+            gallery (Tensor | ndarray | list): indices pool.
+            num (int): expected sample num.
+
+        Returns:
+            Tensor or ndarray: sampled indices.
+        """
+        assert len(gallery) >= num
+
+        is_tensor = isinstance(gallery, torch.Tensor)
+        if not is_tensor:
+            if torch.cuda.is_available():
+                device = torch.cuda.current_device()
+            else:
+                device = 'cpu'
+            gallery = torch.tensor(gallery, dtype=torch.long, device=device)
+        perm = torch.randperm(gallery.numel(), device=gallery.device)[:num]
+        rand_inds = gallery[perm]
+        if not is_tensor:
+            rand_inds = rand_inds.cpu().numpy()
+        return rand_inds
+
+    def _sample_pos(self, assign_result, num_expected, **kwargs):
+        """Randomly sample some positive samples."""
+        pos_inds = torch.nonzero(assign_result.gt_inds > 0, as_tuple=False)
+        if pos_inds.numel() != 0:
+            pos_inds = pos_inds.squeeze(1)
+        if pos_inds.numel() <= num_expected:
+            return pos_inds
+        else:
+            return self.random_choice(pos_inds, num_expected)
+
+    def _sample_neg(self, assign_result, num_expected, **kwargs):
+        """Randomly sample some negative samples."""
+        neg_inds = torch.nonzero(assign_result.gt_inds == 0, as_tuple=False)
+        if neg_inds.numel() != 0:
+            neg_inds = neg_inds.squeeze(1)
+        if len(neg_inds) <= num_expected:
+            return neg_inds
+        else:
+            return self.random_choice(neg_inds, num_expected)
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/sampling_result.py b/detection/scrfd/mmdet/core/bbox/samplers/sampling_result.py
new file mode 100755
index 000000000..419a8e39a
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/sampling_result.py
@@ -0,0 +1,152 @@
+import torch
+
+from mmdet.utils import util_mixins
+
+
+class SamplingResult(util_mixins.NiceRepr):
+    """Bbox sampling result.
+
+    Example:
+        >>> # xdoctest: +IGNORE_WANT
+        >>> from mmdet.core.bbox.samplers.sampling_result import *  # NOQA
+        >>> self = SamplingResult.random(rng=10)
+        >>> print(f'self = {self}')
+        self = <SamplingResult({
+            'neg_bboxes': torch.Size([12, 4]),
+            'neg_inds': tensor([ 0,  1,  2,  4,  5,  6,  7,  8,  9, 10, 11, 12]),
+            'num_gts': 4,
+            'pos_assigned_gt_inds': tensor([], dtype=torch.int64),
+            'pos_bboxes': torch.Size([0, 4]),
+            'pos_inds': tensor([], dtype=torch.int64),
+            'pos_is_gt': tensor([], dtype=torch.uint8)
+        })>
+    """
+
+    def __init__(self, pos_inds, neg_inds, bboxes, gt_bboxes, assign_result,
+                 gt_flags):
+        self.pos_inds = pos_inds
+        self.neg_inds = neg_inds
+        self.pos_bboxes = bboxes[pos_inds]
+        self.neg_bboxes = bboxes[neg_inds]
+        self.pos_is_gt = gt_flags[pos_inds]
+
+        self.num_gts = gt_bboxes.shape[0]
+        self.pos_assigned_gt_inds = assign_result.gt_inds[pos_inds] - 1
+
+        if gt_bboxes.numel() == 0:
+            # hack for index error case
+            assert self.pos_assigned_gt_inds.numel() == 0
+            self.pos_gt_bboxes = torch.empty_like(gt_bboxes).view(-1, 4)
+        else:
+            if len(gt_bboxes.shape) < 2:
+                gt_bboxes = gt_bboxes.view(-1, 4)
+
+            self.pos_gt_bboxes = gt_bboxes[self.pos_assigned_gt_inds, :]
+
+        if assign_result.labels is not None:
+            self.pos_gt_labels = assign_result.labels[pos_inds]
+        else:
+            self.pos_gt_labels = None
+
+    @property
+    def bboxes(self):
+        """torch.Tensor: concatenated positive and negative boxes"""
+        return torch.cat([self.pos_bboxes, self.neg_bboxes])
+
+    def to(self, device):
+        """Change the device of the data inplace.
+
+        Example:
+            >>> self = SamplingResult.random()
+            >>> print(f'self = {self.to(None)}')
+            >>> # xdoctest: +REQUIRES(--gpu)
+            >>> print(f'self = {self.to(0)}')
+        """
+        _dict = self.__dict__
+        for key, value in _dict.items():
+            if isinstance(value, torch.Tensor):
+                _dict[key] = value.to(device)
+        return self
+
+    def __nice__(self):
+        data = self.info.copy()
+        data['pos_bboxes'] = data.pop('pos_bboxes').shape
+        data['neg_bboxes'] = data.pop('neg_bboxes').shape
+        parts = [f"'{k}': {v!r}" for k, v in sorted(data.items())]
+        body = '    ' + ',\n    '.join(parts)
+        return '{\n' + body + '\n}'
+
+    @property
+    def info(self):
+        """Returns a dictionary of info about the object."""
+        return {
+            'pos_inds': self.pos_inds,
+            'neg_inds': self.neg_inds,
+            'pos_bboxes': self.pos_bboxes,
+            'neg_bboxes': self.neg_bboxes,
+            'pos_is_gt': self.pos_is_gt,
+            'num_gts': self.num_gts,
+            'pos_assigned_gt_inds': self.pos_assigned_gt_inds,
+        }
+
+    @classmethod
+    def random(cls, rng=None, **kwargs):
+        """
+        Args:
+            rng (None | int | numpy.random.RandomState): seed or state.
+            kwargs (keyword arguments):
+                - num_preds: number of predicted boxes
+                - num_gts: number of true boxes
+                - p_ignore (float): probability of a predicted box assinged to \
+                    an ignored truth.
+                - p_assigned (float): probability of a predicted box not being \
+                    assigned.
+                - p_use_label (float | bool): with labels or not.
+
+        Returns:
+            :obj:`SamplingResult`: Randomly generated sampling result.
+
+        Example:
+            >>> from mmdet.core.bbox.samplers.sampling_result import *  # NOQA
+            >>> self = SamplingResult.random()
+            >>> print(self.__dict__)
+        """
+        from mmdet.core.bbox.samplers.random_sampler import RandomSampler
+        from mmdet.core.bbox.assigners.assign_result import AssignResult
+        from mmdet.core.bbox import demodata
+        rng = demodata.ensure_rng(rng)
+
+        # make probabalistic?
+        num = 32
+        pos_fraction = 0.5
+        neg_pos_ub = -1
+
+        assign_result = AssignResult.random(rng=rng, **kwargs)
+
+        # Note we could just compute an assignment
+        bboxes = demodata.random_boxes(assign_result.num_preds, rng=rng)
+        gt_bboxes = demodata.random_boxes(assign_result.num_gts, rng=rng)
+
+        if rng.rand() > 0.2:
+            # sometimes algorithms squeeze their data, be robust to that
+            gt_bboxes = gt_bboxes.squeeze()
+            bboxes = bboxes.squeeze()
+
+        if assign_result.labels is None:
+            gt_labels = None
+        else:
+            gt_labels = None  # todo
+
+        if gt_labels is None:
+            add_gt_as_proposals = False
+        else:
+            add_gt_as_proposals = True  # make probabalistic?
+
+        sampler = RandomSampler(
+            num,
+            pos_fraction,
+            neg_pos_ub=neg_pos_ub,
+            add_gt_as_proposals=add_gt_as_proposals,
+            rng=rng)
+        self = sampler.sample(assign_result, bboxes, gt_bboxes, gt_labels)
+        return self
diff --git a/detection/scrfd/mmdet/core/bbox/samplers/score_hlr_sampler.py b/detection/scrfd/mmdet/core/bbox/samplers/score_hlr_sampler.py
new file mode 100755
index 000000000..11d46b977
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/samplers/score_hlr_sampler.py
@@ -0,0 +1,264 @@
+import torch
+from mmcv.ops import nms_match
+
+from ..builder import BBOX_SAMPLERS
+from ..transforms import bbox2roi
+from .base_sampler import BaseSampler
+from .sampling_result import SamplingResult
+
+
+@BBOX_SAMPLERS.register_module()
+class ScoreHLRSampler(BaseSampler):
+    r"""Importance-based Sample Reweighting (ISR_N), described in `Prime Sample
+    Attention in Object Detection <https://arxiv.org/abs/1904.04821>`_.
+
+    Score hierarchical local rank (HLR) differentiates with RandomSampler in
+    negative part. It firstly computes Score-HLR in a two-step way,
+    then linearly maps score hlr to the loss weights.
+
+    Args:
+        num (int): Total number of sampled RoIs.
+        pos_fraction (float): Fraction of positive samples.
+        context (:class:`BaseRoIHead`): RoI head that the sampler belongs to.
+        neg_pos_ub (int): Upper bound of the ratio of num negative to num
+            positive, -1 means no upper bound.
+        add_gt_as_proposals (bool): Whether to add ground truth as proposals.
+        k (float): Power of the non-linear mapping.
+        bias (float): Shift of the non-linear mapping.
+        score_thr (float): Minimum score that a negative sample is to be
+            considered as valid bbox.
+    """
+
+    def __init__(self,
+                 num,
+                 pos_fraction,
+                 context,
+                 neg_pos_ub=-1,
+                 add_gt_as_proposals=True,
+                 k=0.5,
+                 bias=0,
+                 score_thr=0.05,
+                 iou_thr=0.5,
+                 **kwargs):
+        super().__init__(num, pos_fraction, neg_pos_ub, add_gt_as_proposals)
+        self.k = k
+        self.bias = bias
+        self.score_thr = score_thr
+        self.iou_thr = iou_thr
+        self.context = context
+        # context of cascade detectors is a list, so distinguish them here.
+        if not hasattr(context, 'num_stages'):
+            self.bbox_roi_extractor = context.bbox_roi_extractor
+            self.bbox_head = context.bbox_head
+            self.with_shared_head = context.with_shared_head
+            if self.with_shared_head:
+                self.shared_head = context.shared_head
+        else:
+            self.bbox_roi_extractor = context.bbox_roi_extractor[
+                context.current_stage]
+            self.bbox_head = context.bbox_head[context.current_stage]
+
+    @staticmethod
+    def random_choice(gallery, num):
+        """Randomly select some elements from the gallery.
+
+        If `gallery` is a Tensor, the returned indices will be a Tensor;
+        If `gallery` is a ndarray or list, the returned indices will be a
+        ndarray.
+
+        Args:
+            gallery (Tensor | ndarray | list): indices pool.
+            num (int): expected sample num.
+
+        Returns:
+            Tensor or ndarray: sampled indices.
+        """
+        assert len(gallery) >= num
+
+        is_tensor = isinstance(gallery, torch.Tensor)
+        if not is_tensor:
+            if torch.cuda.is_available():
+                device = torch.cuda.current_device()
+            else:
+                device = 'cpu'
+            gallery = torch.tensor(gallery, dtype=torch.long, device=device)
+        perm = torch.randperm(gallery.numel(), device=gallery.device)[:num]
+        rand_inds = gallery[perm]
+        if not is_tensor:
+            rand_inds = rand_inds.cpu().numpy()
+        return rand_inds
+
+    def _sample_pos(self, assign_result, num_expected, **kwargs):
+        """Randomly sample some positive samples."""
+        pos_inds = torch.nonzero(assign_result.gt_inds > 0).flatten()
+        if pos_inds.numel() <= num_expected:
+            return pos_inds
+        else:
+            return self.random_choice(pos_inds, num_expected)
+
+    def _sample_neg(self,
+                    assign_result,
+                    num_expected,
+                    bboxes,
+                    feats=None,
+                    img_meta=None,
+                    **kwargs):
+        """Sample negative samples.
+
+        Score-HLR sampler is done in the following steps:
+        1. Take the maximum positive score prediction of each negative samples
+            as s_i.
+        2. Filter out negative samples whose s_i <= score_thr, the left samples
+            are called valid samples.
+        3. Use NMS-Match to divide valid samples into different groups,
+            samples in the same group will greatly overlap with each other
+        4. Rank the matched samples in two-steps to get Score-HLR.
+            (1) In the same group, rank samples with their scores.
+            (2) In the same score rank across different groups,
+                rank samples with their scores again.
+        5. Linearly map Score-HLR to the final label weights.
+
+        Args:
+            assign_result (:obj:`AssignResult`): result of assigner.
+            num_expected (int): Expected number of samples.
+            bboxes (Tensor): bbox to be sampled.
+            feats (Tensor): Features come from FPN.
+            img_meta (dict): Meta information dictionary.
+        """
+        neg_inds = torch.nonzero(assign_result.gt_inds == 0).flatten()
+        num_neg = neg_inds.size(0)
+        if num_neg == 0:
+            return neg_inds, None
+        with torch.no_grad():
+            neg_bboxes = bboxes[neg_inds]
+            neg_rois = bbox2roi([neg_bboxes])
+            bbox_result = self.context._bbox_forward(feats, neg_rois)
+            cls_score, bbox_pred = bbox_result['cls_score'], bbox_result[
+                'bbox_pred']
+
+            ori_loss = self.bbox_head.loss(
+                cls_score=cls_score,
+                bbox_pred=None,
+                rois=None,
+                labels=neg_inds.new_full((num_neg, ),
+                                         self.bbox_head.num_classes),
+                label_weights=cls_score.new_ones(num_neg),
+                bbox_targets=None,
+                bbox_weights=None,
+                reduction_override='none')['loss_cls']
+
+            # filter out samples with the max score lower than score_thr
+            max_score, argmax_score = cls_score.softmax(-1)[:, :-1].max(-1)
+            valid_inds = (max_score > self.score_thr).nonzero().view(-1)
+            invalid_inds = (max_score <= self.score_thr).nonzero().view(-1)
+            num_valid = valid_inds.size(0)
+            num_invalid = invalid_inds.size(0)
+
+            num_expected = min(num_neg, num_expected)
+            num_hlr = min(num_valid, num_expected)
+            num_rand = num_expected - num_hlr
+            if num_valid > 0:
+                valid_rois = neg_rois[valid_inds]
+                valid_max_score = max_score[valid_inds]
+                valid_argmax_score = argmax_score[valid_inds]
+                valid_bbox_pred = bbox_pred[valid_inds]
+
+                # valid_bbox_pred shape: [num_valid, #num_classes, 4]
+                valid_bbox_pred = valid_bbox_pred.view(
+                    valid_bbox_pred.size(0), -1, 4)
+                selected_bbox_pred = valid_bbox_pred[range(num_valid),
+                                                     valid_argmax_score]
+                pred_bboxes = self.bbox_head.bbox_coder.decode(
+                    valid_rois[:, 1:], selected_bbox_pred)
+                pred_bboxes_with_score = torch.cat(
+                    [pred_bboxes, valid_max_score[:, None]], -1)
+                group = nms_match(pred_bboxes_with_score, self.iou_thr)
+
+                # imp: importance
+                imp = cls_score.new_zeros(num_valid)
+                for g in group:
+                    g_score = valid_max_score[g]
+                    # g_score has already sorted
+                    rank = g_score.new_tensor(range(g_score.size(0)))
+                    imp[g] = num_valid - rank + g_score
+                _, imp_rank_inds = imp.sort(descending=True)
+                _, imp_rank = imp_rank_inds.sort()
+                hlr_inds = imp_rank_inds[:num_expected]
+
+                if num_rand > 0:
+                    rand_inds = torch.randperm(num_invalid)[:num_rand]
+                    select_inds = torch.cat(
+                        [valid_inds[hlr_inds], invalid_inds[rand_inds]])
+                else:
+                    select_inds = valid_inds[hlr_inds]
+
+                neg_label_weights = cls_score.new_ones(num_expected)
+
+                up_bound = max(num_expected, num_valid)
+                imp_weights = (up_bound -
+                               imp_rank[hlr_inds].float()) / up_bound
+                neg_label_weights[:num_hlr] = imp_weights
+                neg_label_weights[num_hlr:] = imp_weights.min()
+                neg_label_weights = (self.bias +
+                                     (1 - self.bias) * neg_label_weights).pow(
+                                         self.k)
+                ori_selected_loss = ori_loss[select_inds]
+                new_loss = ori_selected_loss * neg_label_weights
+                norm_ratio = ori_selected_loss.sum() / new_loss.sum()
+                neg_label_weights *= norm_ratio
+            else:
+                neg_label_weights = cls_score.new_ones(num_expected)
+                select_inds = torch.randperm(num_neg)[:num_expected]
+
+            return neg_inds[select_inds], neg_label_weights
+
+    def sample(self,
+               assign_result,
+               bboxes,
+               gt_bboxes,
+               gt_labels=None,
+               img_meta=None,
+               **kwargs):
+        """Sample positive and negative bboxes.
+
+        This is a simple implementation of bbox sampling given candidates,
+        assigning results and ground truth bboxes.
+
+        Args:
+            assign_result (:obj:`AssignResult`): Bbox assigning results.
+            bboxes (Tensor): Boxes to be sampled from.
+            gt_bboxes (Tensor): Ground truth bboxes.
+            gt_labels (Tensor, optional): Class labels of ground truth bboxes.
+
+        Returns:
+            tuple[:obj:`SamplingResult`, Tensor]: Sampling result and negetive
+                label weights.
+        """
+        bboxes = bboxes[:, :4]
+
+        gt_flags = bboxes.new_zeros((bboxes.shape[0], ), dtype=torch.uint8)
+        if self.add_gt_as_proposals:
+            bboxes = torch.cat([gt_bboxes, bboxes], dim=0)
+            assign_result.add_gt_(gt_labels)
+            gt_ones = bboxes.new_ones(gt_bboxes.shape[0], dtype=torch.uint8)
+            gt_flags = torch.cat([gt_ones, gt_flags])
+
+        num_expected_pos = int(self.num * self.pos_fraction)
+        pos_inds = self.pos_sampler._sample_pos(
+            assign_result, num_expected_pos, bboxes=bboxes, **kwargs)
+        num_sampled_pos = pos_inds.numel()
+        num_expected_neg = self.num - num_sampled_pos
+        if self.neg_pos_ub >= 0:
+            _pos = max(1, num_sampled_pos)
+            neg_upper_bound = int(self.neg_pos_ub * _pos)
+            if num_expected_neg > neg_upper_bound:
+                num_expected_neg = neg_upper_bound
+        neg_inds, neg_label_weights = self.neg_sampler._sample_neg(
+            assign_result,
+            num_expected_neg,
+            bboxes,
+            img_meta=img_meta,
+            **kwargs)
+
+        return SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes,
+                              assign_result, gt_flags), neg_label_weights
diff --git a/detection/scrfd/mmdet/core/bbox/transforms.py b/detection/scrfd/mmdet/core/bbox/transforms.py
new file mode 100755
index 000000000..8bd6d4fb1
--- /dev/null
+++ b/detection/scrfd/mmdet/core/bbox/transforms.py
@@ -0,0 +1,270 @@
+import numpy as np
+import torch
+
+
+def bbox_flip(bboxes, img_shape, direction='horizontal'):
+    """Flip bboxes horizontally or vertically.
+
+    Args:
+        bboxes (Tensor): Shape (..., 4*k)
+        img_shape (tuple): Image shape.
+        direction (str): Flip direction, options are "horizontal", "vertical",
+            "diagonal". Default: "horizontal"
+
+    Returns:
+        Tensor: Flipped bboxes.
+    """
+    assert bboxes.shape[-1] % 4 == 0
+    assert direction in ['horizontal', 'vertical', 'diagonal']
+    flipped = bboxes.clone()
+    if direction == 'horizontal':
+        flipped[..., 0::4] = img_shape[1] - bboxes[..., 2::4]
+        flipped[..., 2::4] = img_shape[1] - bboxes[..., 0::4]
+    elif direction == 'vertical':
+        flipped[..., 1::4] = img_shape[0] - bboxes[..., 3::4]
+        flipped[..., 3::4] = img_shape[0] - bboxes[..., 1::4]
+    else:
+        flipped[..., 0::4] = img_shape[1] - bboxes[..., 2::4]
+        flipped[..., 1::4] = img_shape[0] - bboxes[..., 3::4]
+        flipped[..., 2::4] = img_shape[1] - bboxes[..., 0::4]
+        flipped[..., 3::4] = img_shape[0] - bboxes[..., 1::4]
+    return flipped
+
+
+def bbox_mapping(bboxes,
+                 img_shape,
+                 scale_factor,
+                 flip,
+                 flip_direction='horizontal'):
+    """Map bboxes from the original image scale to testing scale."""
+    new_bboxes = bboxes * bboxes.new_tensor(scale_factor)
+    if flip:
+        new_bboxes = bbox_flip(new_bboxes, img_shape, flip_direction)
+    return new_bboxes
+
+
+def bbox_mapping_back(bboxes,
+                      img_shape,
+                      scale_factor,
+                      flip,
+                      flip_direction='horizontal'):
+    """Map bboxes from testing scale to original image scale."""
+    new_bboxes = bbox_flip(bboxes, img_shape,
+                           flip_direction) if flip else bboxes
+    new_bboxes = new_bboxes.view(-1, 4) / new_bboxes.new_tensor(scale_factor)
+    return new_bboxes.view(bboxes.shape)
+
+
+def bbox2roi(bbox_list):
+    """Convert a list of bboxes to roi format.
+
+    Args:
+        bbox_list (list[Tensor]): a list of bboxes corresponding to a batch
+            of images.
+
+    Returns:
+        Tensor: shape (n, 5), [batch_ind, x1, y1, x2, y2]
+    """
+    rois_list = []
+    for img_id, bboxes in enumerate(bbox_list):
+        if bboxes.size(0) > 0:
+            img_inds = bboxes.new_full((bboxes.size(0), 1), img_id)
+            rois = torch.cat([img_inds, bboxes[:, :4]], dim=-1)
+        else:
+            rois = bboxes.new_zeros((0, 5))
+        rois_list.append(rois)
+    rois = torch.cat(rois_list, 0)
+    return rois
+
+
+def roi2bbox(rois):
+    """Convert rois to bounding box format.
+
+    Args:
+        rois (torch.Tensor): RoIs with the shape (n, 5) where the first
+            column indicates batch id of each RoI.
+
+    Returns:
+        list[torch.Tensor]: Converted boxes of corresponding rois.
+    """
+    bbox_list = []
+    img_ids = torch.unique(rois[:, 0].cpu(), sorted=True)
+    for img_id in img_ids:
+        inds = (rois[:, 0] == img_id.item())
+        bbox = rois[inds, 1:]
+        bbox_list.append(bbox)
+    return bbox_list
+
+
+def bbox2result(bboxes, labels, num_classes):
+    """Convert detection results to a list of numpy arrays.
+
+    Args:
+        bboxes (torch.Tensor | np.ndarray): shape (n, 5)
+        labels (torch.Tensor | np.ndarray): shape (n, )
+        num_classes (int): class number, including background class
+
+    Returns:
+        list(ndarray): bbox results of each class
+    """
+    if bboxes.shape[0] == 0:
+        return [np.zeros((0, 5), dtype=np.float32) for i in range(num_classes)]
+    else:
+        if isinstance(bboxes, torch.Tensor):
+            bboxes = bboxes.detach().cpu().numpy()
+            labels = labels.detach().cpu().numpy()
+        return [bboxes[labels == i, :] for i in range(num_classes)]
+
+
+def distance2bbox(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    return torch.stack([x1, y1, x2, y2], -1)
+
+def distance2kps(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded kps.
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i%2] + distance[:, i]
+        py = points[:, i%2+1] + distance[:, i+1]
+        if max_shape is not None:
+            px = px.clamp(min=0, max=max_shape[1])
+            py = py.clamp(min=0, max=max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return torch.stack(preds, -1)
+
+
+def bbox2distance(points, bbox, max_dis=None, eps=0.1):
+    """Decode bounding box based on distances.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        bbox (Tensor): Shape (n, 4), "xyxy" format
+        max_dis (float): Upper bound of the distance.
+        eps (float): a small value to ensure target < max_dis, instead <=
+
+    Returns:
+        Tensor: Decoded distances.
+    """
+    left = points[:, 0] - bbox[:, 0]
+    top = points[:, 1] - bbox[:, 1]
+    right = bbox[:, 2] - points[:, 0]
+    bottom = bbox[:, 3] - points[:, 1]
+    if max_dis is not None:
+        left = left.clamp(min=0, max=max_dis - eps)
+        top = top.clamp(min=0, max=max_dis - eps)
+        right = right.clamp(min=0, max=max_dis - eps)
+        bottom = bottom.clamp(min=0, max=max_dis - eps)
+    return torch.stack([left, top, right, bottom], -1)
+
+def kps2distance(points, kps, max_dis=None, eps=0.1):
+    """Decode bounding box based on distances.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        kps (Tensor): Shape (n, K), "xyxy" format
+        max_dis (float): Upper bound of the distance.
+        eps (float): a small value to ensure target < max_dis, instead <=
+
+    Returns:
+        Tensor: Decoded distances.
+    """
+
+    preds = []
+    for i in range(0, kps.shape[1], 2):
+        px = kps[:, i] - points[:, i%2]
+        py = kps[:, i+1] - points[:, i%2+1]
+        if max_dis is not None:
+            px = px.clamp(min=0, max=max_dis - eps)
+            py = py.clamp(min=0, max=max_dis - eps)
+        preds.append(px)
+        preds.append(py)
+    return torch.stack(preds, -1)
+
+def bbox_rescale(bboxes, scale_factor=1.0):
+    """Rescale bounding box w.r.t. scale_factor.
+
+    Args:
+        bboxes (Tensor): Shape (n, 4) for bboxes or (n, 5) for rois
+        scale_factor (float): rescale factor
+
+    Returns:
+        Tensor: Rescaled bboxes.
+    """
+    if bboxes.size(1) == 5:
+        bboxes_ = bboxes[:, 1:]
+        inds_ = bboxes[:, 0]
+    else:
+        bboxes_ = bboxes
+    cx = (bboxes_[:, 0] + bboxes_[:, 2]) * 0.5
+    cy = (bboxes_[:, 1] + bboxes_[:, 3]) * 0.5
+    w = bboxes_[:, 2] - bboxes_[:, 0]
+    h = bboxes_[:, 3] - bboxes_[:, 1]
+    w = w * scale_factor
+    h = h * scale_factor
+    x1 = cx - 0.5 * w
+    x2 = cx + 0.5 * w
+    y1 = cy - 0.5 * h
+    y2 = cy + 0.5 * h
+    if bboxes.size(1) == 5:
+        rescaled_bboxes = torch.stack([inds_, x1, y1, x2, y2], dim=-1)
+    else:
+        rescaled_bboxes = torch.stack([x1, y1, x2, y2], dim=-1)
+    return rescaled_bboxes
+
+
+def bbox_cxcywh_to_xyxy(bbox):
+    """Convert bbox coordinates from (cx, cy, w, h) to (x1, y1, x2, y2).
+
+    Args:
+        bbox (Tensor): Shape (n, 4) for bboxes.
+
+    Returns:
+        Tensor: Converted bboxes.
+    """
+    cx, cy, w, h = bbox.split((1, 1, 1, 1), dim=-1)
+    bbox_new = [(cx - 0.5 * w), (cy - 0.5 * h), (cx + 0.5 * w), (cy + 0.5 * h)]
+    return torch.cat(bbox_new, dim=-1)
+
+
+def bbox_xyxy_to_cxcywh(bbox):
+    """Convert bbox coordinates from (x1, y1, x2, y2) to (cx, cy, w, h).
+
+    Args:
+        bbox (Tensor): Shape (n, 4) for bboxes.
+
+    Returns:
+        Tensor: Converted bboxes.
+    """
+    x1, y1, x2, y2 = bbox.split((1, 1, 1, 1), dim=-1)
+    bbox_new = [(x1 + x2) / 2, (y1 + y2) / 2, (x2 - x1), (y2 - y1)]
+    return torch.cat(bbox_new, dim=-1)
diff --git a/detection/scrfd/mmdet/core/evaluation/__init__.py b/detection/scrfd/mmdet/core/evaluation/__init__.py
new file mode 100755
index 000000000..5fef9b346
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/__init__.py
@@ -0,0 +1,17 @@
+from .class_names import (cityscapes_classes, coco_classes, dataset_aliases,
+                          get_classes, imagenet_det_classes,
+                          imagenet_vid_classes, voc_classes)
+from .eval_hooks import DistEvalHook, EvalHook
+from .mean_ap import average_precision, eval_map, print_map_summary
+from .recall import (eval_recalls, plot_iou_recall, plot_num_recall,
+                     print_recall_summary)
+from .widerface import wider_evaluation, get_widerface_gts
+
+__all__ = [
+    'voc_classes', 'imagenet_det_classes', 'imagenet_vid_classes',
+    'coco_classes', 'cityscapes_classes', 'dataset_aliases', 'get_classes',
+    'DistEvalHook', 'EvalHook', 'average_precision', 'eval_map',
+    'print_map_summary', 'eval_recalls', 'print_recall_summary',
+    'plot_num_recall', 'plot_iou_recall',
+    'wider_evaluation', 'get_widerface_gts',
+]
diff --git a/detection/scrfd/mmdet/core/evaluation/bbox_overlaps.py b/detection/scrfd/mmdet/core/evaluation/bbox_overlaps.py
new file mode 100755
index 000000000..93559ea0f
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/bbox_overlaps.py
@@ -0,0 +1,48 @@
+import numpy as np
+
+
+def bbox_overlaps(bboxes1, bboxes2, mode='iou', eps=1e-6):
+    """Calculate the ious between each bbox of bboxes1 and bboxes2.
+
+    Args:
+        bboxes1(ndarray): shape (n, 4)
+        bboxes2(ndarray): shape (k, 4)
+        mode(str): iou (intersection over union) or iof (intersection
+            over foreground)
+
+    Returns:
+        ious(ndarray): shape (n, k)
+    """
+
+    assert mode in ['iou', 'iof']
+
+    bboxes1 = bboxes1.astype(np.float32)
+    bboxes2 = bboxes2.astype(np.float32)
+    rows = bboxes1.shape[0]
+    cols = bboxes2.shape[0]
+    ious = np.zeros((rows, cols), dtype=np.float32)
+    if rows * cols == 0:
+        return ious
+    exchange = False
+    if bboxes1.shape[0] > bboxes2.shape[0]:
+        bboxes1, bboxes2 = bboxes2, bboxes1
+        ious = np.zeros((cols, rows), dtype=np.float32)
+        exchange = True
+    area1 = (bboxes1[:, 2] - bboxes1[:, 0]) * (bboxes1[:, 3] - bboxes1[:, 1])
+    area2 = (bboxes2[:, 2] - bboxes2[:, 0]) * (bboxes2[:, 3] - bboxes2[:, 1])
+    for i in range(bboxes1.shape[0]):
+        x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0])
+        y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1])
+        x_end = np.minimum(bboxes1[i, 2], bboxes2[:, 2])
+        y_end = np.minimum(bboxes1[i, 3], bboxes2[:, 3])
+        overlap = np.maximum(x_end - x_start, 0) * np.maximum(
+            y_end - y_start, 0)
+        if mode == 'iou':
+            union = area1[i] + area2 - overlap
+        else:
+            union = area1[i] if not exchange else area2
+        union = np.maximum(union, eps)
+        ious[i, :] = overlap / union
+    if exchange:
+        ious = ious.T
+    return ious
diff --git a/detection/scrfd/mmdet/core/evaluation/class_names.py b/detection/scrfd/mmdet/core/evaluation/class_names.py
new file mode 100755
index 000000000..4b8845f3f
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/class_names.py
@@ -0,0 +1,116 @@
+import mmcv
+
+
+def wider_face_classes():
+    return ['face']
+
+
+def voc_classes():
+    return [
+        'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat',
+        'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person',
+        'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
+    ]
+
+
+def imagenet_det_classes():
+    return [
+        'accordion', 'airplane', 'ant', 'antelope', 'apple', 'armadillo',
+        'artichoke', 'axe', 'baby_bed', 'backpack', 'bagel', 'balance_beam',
+        'banana', 'band_aid', 'banjo', 'baseball', 'basketball', 'bathing_cap',
+        'beaker', 'bear', 'bee', 'bell_pepper', 'bench', 'bicycle', 'binder',
+        'bird', 'bookshelf', 'bow_tie', 'bow', 'bowl', 'brassiere', 'burrito',
+        'bus', 'butterfly', 'camel', 'can_opener', 'car', 'cart', 'cattle',
+        'cello', 'centipede', 'chain_saw', 'chair', 'chime', 'cocktail_shaker',
+        'coffee_maker', 'computer_keyboard', 'computer_mouse', 'corkscrew',
+        'cream', 'croquet_ball', 'crutch', 'cucumber', 'cup_or_mug', 'diaper',
+        'digital_clock', 'dishwasher', 'dog', 'domestic_cat', 'dragonfly',
+        'drum', 'dumbbell', 'electric_fan', 'elephant', 'face_powder', 'fig',
+        'filing_cabinet', 'flower_pot', 'flute', 'fox', 'french_horn', 'frog',
+        'frying_pan', 'giant_panda', 'goldfish', 'golf_ball', 'golfcart',
+        'guacamole', 'guitar', 'hair_dryer', 'hair_spray', 'hamburger',
+        'hammer', 'hamster', 'harmonica', 'harp', 'hat_with_a_wide_brim',
+        'head_cabbage', 'helmet', 'hippopotamus', 'horizontal_bar', 'horse',
+        'hotdog', 'iPod', 'isopod', 'jellyfish', 'koala_bear', 'ladle',
+        'ladybug', 'lamp', 'laptop', 'lemon', 'lion', 'lipstick', 'lizard',
+        'lobster', 'maillot', 'maraca', 'microphone', 'microwave', 'milk_can',
+        'miniskirt', 'monkey', 'motorcycle', 'mushroom', 'nail', 'neck_brace',
+        'oboe', 'orange', 'otter', 'pencil_box', 'pencil_sharpener', 'perfume',
+        'person', 'piano', 'pineapple', 'ping-pong_ball', 'pitcher', 'pizza',
+        'plastic_bag', 'plate_rack', 'pomegranate', 'popsicle', 'porcupine',
+        'power_drill', 'pretzel', 'printer', 'puck', 'punching_bag', 'purse',
+        'rabbit', 'racket', 'ray', 'red_panda', 'refrigerator',
+        'remote_control', 'rubber_eraser', 'rugby_ball', 'ruler',
+        'salt_or_pepper_shaker', 'saxophone', 'scorpion', 'screwdriver',
+        'seal', 'sheep', 'ski', 'skunk', 'snail', 'snake', 'snowmobile',
+        'snowplow', 'soap_dispenser', 'soccer_ball', 'sofa', 'spatula',
+        'squirrel', 'starfish', 'stethoscope', 'stove', 'strainer',
+        'strawberry', 'stretcher', 'sunglasses', 'swimming_trunks', 'swine',
+        'syringe', 'table', 'tape_player', 'tennis_ball', 'tick', 'tie',
+        'tiger', 'toaster', 'traffic_light', 'train', 'trombone', 'trumpet',
+        'turtle', 'tv_or_monitor', 'unicycle', 'vacuum', 'violin',
+        'volleyball', 'waffle_iron', 'washer', 'water_bottle', 'watercraft',
+        'whale', 'wine_bottle', 'zebra'
+    ]
+
+
+def imagenet_vid_classes():
+    return [
+        'airplane', 'antelope', 'bear', 'bicycle', 'bird', 'bus', 'car',
+        'cattle', 'dog', 'domestic_cat', 'elephant', 'fox', 'giant_panda',
+        'hamster', 'horse', 'lion', 'lizard', 'monkey', 'motorcycle', 'rabbit',
+        'red_panda', 'sheep', 'snake', 'squirrel', 'tiger', 'train', 'turtle',
+        'watercraft', 'whale', 'zebra'
+    ]
+
+
+def coco_classes():
+    return [
+        'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
+        'truck', 'boat', 'traffic_light', 'fire_hydrant', 'stop_sign',
+        'parking_meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',
+        'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella',
+        'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard',
+        'sports_ball', 'kite', 'baseball_bat', 'baseball_glove', 'skateboard',
+        'surfboard', 'tennis_racket', 'bottle', 'wine_glass', 'cup', 'fork',
+        'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
+        'broccoli', 'carrot', 'hot_dog', 'pizza', 'donut', 'cake', 'chair',
+        'couch', 'potted_plant', 'bed', 'dining_table', 'toilet', 'tv',
+        'laptop', 'mouse', 'remote', 'keyboard', 'cell_phone', 'microwave',
+        'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase',
+        'scissors', 'teddy_bear', 'hair_drier', 'toothbrush'
+    ]
+
+
+def cityscapes_classes():
+    return [
+        'person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',
+        'bicycle'
+    ]
+
+
+dataset_aliases = {
+    'voc': ['voc', 'pascal_voc', 'voc07', 'voc12'],
+    'imagenet_det': ['det', 'imagenet_det', 'ilsvrc_det'],
+    'imagenet_vid': ['vid', 'imagenet_vid', 'ilsvrc_vid'],
+    'coco': ['coco', 'mscoco', 'ms_coco'],
+    'wider_face': ['WIDERFaceDataset', 'wider_face', 'WDIERFace'],
+    'cityscapes': ['cityscapes']
+}
+
+
+def get_classes(dataset):
+    """Get class names of a dataset."""
+    alias2name = {}
+    for name, aliases in dataset_aliases.items():
+        for alias in aliases:
+            alias2name[alias] = name
+
+    if mmcv.is_str(dataset):
+        if dataset in alias2name:
+            labels = eval(alias2name[dataset] + '_classes()')
+        else:
+            raise ValueError(f'Unrecognized dataset: {dataset}')
+    else:
+        raise TypeError(f'dataset must a str, but got {type(dataset)}')
+    return labels
diff --git a/detection/scrfd/mmdet/core/evaluation/eval_hooks.py b/detection/scrfd/mmdet/core/evaluation/eval_hooks.py
new file mode 100755
index 000000000..15d9abd92
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/eval_hooks.py
@@ -0,0 +1,255 @@
+import os.path as osp
+import warnings
+from math import inf
+
+import mmcv
+from mmcv.runner import Hook
+from torch.utils.data import DataLoader
+
+from mmdet.utils import get_root_logger
+
+
+class EvalHook(Hook):
+    """Evaluation hook.
+
+    Notes:
+        If new arguments are added for EvalHook, tools/test.py,
+        tools/eval_metric.py may be effected.
+
+    Attributes:
+        dataloader (DataLoader): A PyTorch dataloader.
+        start (int, optional): Evaluation starting epoch. It enables evaluation
+            before the training starts if ``start`` <= the resuming epoch.
+            If None, whether to evaluate is merely decided by ``interval``.
+            Default: None.
+        interval (int): Evaluation interval (by epochs). Default: 1.
+        save_best (str, optional): If a metric is specified, it would measure
+            the best checkpoint during evaluation. The information about best
+            checkpoint would be save in best.json.
+            Options are the evaluation metrics to the test dataset. e.g.,
+            ``bbox_mAP``, ``segm_mAP`` for bbox detection and instance
+            segmentation. ``AR@100`` for proposal recall. If ``save_best`` is
+            ``auto``, the first key will be used. The interval of
+            ``CheckpointHook`` should device EvalHook. Default: None.
+        rule (str, optional): Comparison rule for best score. If set to None,
+            it will infer a reasonable rule. Keys such as 'mAP' or 'AR' will
+            be inferred by 'greater' rule. Keys contain 'loss' will be inferred
+             by 'less' rule. Options are 'greater', 'less'. Default: None.
+        **eval_kwargs: Evaluation arguments fed into the evaluate function of
+            the dataset.
+    """
+
+    rule_map = {'greater': lambda x, y: x > y, 'less': lambda x, y: x < y}
+    init_value_map = {'greater': -inf, 'less': inf}
+    greater_keys = ['mAP', 'AR']
+    less_keys = ['loss']
+
+    def __init__(self,
+                 dataloader,
+                 start=None,
+                 interval=1,
+                 save_best=None,
+                 rule=None,
+                 **eval_kwargs):
+        if not isinstance(dataloader, DataLoader):
+            raise TypeError('dataloader must be a pytorch DataLoader, but got'
+                            f' {type(dataloader)}')
+        if not interval > 0:
+            raise ValueError(f'interval must be positive, but got {interval}')
+        if start is not None and start < 0:
+            warnings.warn(
+                f'The evaluation start epoch {start} is smaller than 0, '
+                f'use 0 instead', UserWarning)
+            start = 0
+        self.dataloader = dataloader
+        self.interval = interval
+        self.start = start
+        assert isinstance(save_best, str) or save_best is None
+        self.save_best = save_best
+        self.eval_kwargs = eval_kwargs
+        self.initial_epoch_flag = True
+
+        self.logger = get_root_logger()
+
+        if self.save_best is not None:
+            self._init_rule(rule, self.save_best)
+
+    def _init_rule(self, rule, key_indicator):
+        """Initialize rule, key_indicator, comparison_func, and best score.
+
+        Args:
+            rule (str | None): Comparison rule for best score.
+            key_indicator (str | None): Key indicator to determine the
+                comparison rule.
+        """
+        if rule not in self.rule_map and rule is not None:
+            raise KeyError(f'rule must be greater, less or None, '
+                           f'but got {rule}.')
+
+        if rule is None:
+            if key_indicator != 'auto':
+                if any(key in key_indicator for key in self.greater_keys):
+                    rule = 'greater'
+                elif any(key in key_indicator for key in self.less_keys):
+                    rule = 'less'
+                else:
+                    raise ValueError(f'Cannot infer the rule for key '
+                                     f'{key_indicator}, thus a specific rule '
+                                     f'must be specified.')
+        self.rule = rule
+        self.key_indicator = key_indicator
+        if self.rule is not None:
+            self.compare_func = self.rule_map[self.rule]
+
+    def before_run(self, runner):
+        if self.save_best is not None:
+            if runner.meta is None:
+                warnings.warn('runner.meta is None. Creating a empty one.')
+                runner.meta = dict()
+            runner.meta.setdefault('hook_msgs', dict())
+
+    def before_train_epoch(self, runner):
+        """Evaluate the model only at the start of training."""
+        if not self.initial_epoch_flag:
+            return
+        if self.start is not None and runner.epoch >= self.start:
+            self.after_train_epoch(runner)
+        self.initial_epoch_flag = False
+
+    def evaluation_flag(self, runner):
+        """Judge whether to perform_evaluation after this epoch.
+
+        Returns:
+            bool: The flag indicating whether to perform evaluation.
+        """
+        if self.start is None:
+            if not self.every_n_epochs(runner, self.interval):
+                # No evaluation during the interval epochs.
+                return False
+        elif (runner.epoch + 1) < self.start:
+            # No evaluation if start is larger than the current epoch.
+            return False
+        else:
+            # Evaluation only at epochs 3, 5, 7... if start==3 and interval==2
+            if (runner.epoch + 1 - self.start) % self.interval:
+                return False
+        return True
+
+    def after_train_epoch(self, runner):
+        if not self.evaluation_flag(runner):
+            return
+        from mmdet.apis import single_gpu_test
+        results = single_gpu_test(runner.model, self.dataloader, show=False)
+        key_score = self.evaluate(runner, results)
+        if self.save_best:
+            best_score = runner.meta['hook_msgs'].get(
+                'best_score', self.init_value_map[self.rule])
+            if self.compare_func(key_score, best_score):
+                best_score = key_score
+                runner.meta['hook_msgs']['best_score'] = best_score
+                last_ckpt = runner.meta['hook_msgs']['last_ckpt']
+                runner.meta['hook_msgs']['best_ckpt'] = last_ckpt
+                mmcv.symlink(
+                    last_ckpt,
+                    osp.join(runner.work_dir,
+                             f'best_{self.key_indicator}.pth'))
+                self.logger.info(
+                    f'Now best checkpoint is epoch_{runner.epoch + 1}.pth.'
+                    f'Best {self.key_indicator} is {best_score:0.4f}')
+
+    def evaluate(self, runner, results):
+        eval_res = self.dataloader.dataset.evaluate(
+            results, logger=runner.logger, **self.eval_kwargs)
+        for name, val in eval_res.items():
+            runner.log_buffer.output[name] = val
+        runner.log_buffer.ready = True
+        if self.save_best is not None:
+            if self.key_indicator == 'auto':
+                # infer from eval_results
+                self._init_rule(self.rule, list(eval_res.keys())[0])
+            return eval_res[self.key_indicator]
+        else:
+            return None
+
+
+class DistEvalHook(EvalHook):
+    """Distributed evaluation hook.
+
+    Notes:
+        If new arguments are added, tools/test.py may be effected.
+
+    Attributes:
+        dataloader (DataLoader): A PyTorch dataloader.
+        start (int, optional): Evaluation starting epoch. It enables evaluation
+            before the training starts if ``start`` <= the resuming epoch.
+            If None, whether to evaluate is merely decided by ``interval``.
+            Default: None.
+        interval (int): Evaluation interval (by epochs). Default: 1.
+        tmpdir (str | None): Temporary directory to save the results of all
+            processes. Default: None.
+        gpu_collect (bool): Whether to use gpu or cpu to collect results.
+            Default: False.
+        save_best (str, optional): If a metric is specified, it would measure
+            the best checkpoint during evaluation. The information about best
+            checkpoint would be save in best.json.
+            Options are the evaluation metrics to the test dataset. e.g.,
+            ``bbox_mAP``, ``segm_mAP`` for bbox detection and instance
+            segmentation. ``AR@100`` for proposal recall. If ``save_best`` is
+            ``auto``, the first key will be used. The interval of
+            ``CheckpointHook`` should device EvalHook. Default: None.
+        rule (str | None): Comparison rule for best score. If set to None,
+            it will infer a reasonable rule. Default: 'None'.
+        **eval_kwargs: Evaluation arguments fed into the evaluate function of
+            the dataset.
+    """
+
+    def __init__(self,
+                 dataloader,
+                 start=None,
+                 interval=1,
+                 tmpdir=None,
+                 gpu_collect=False,
+                 save_best=None,
+                 rule=None,
+                 **eval_kwargs):
+        super().__init__(
+            dataloader,
+            start=start,
+            interval=interval,
+            save_best=save_best,
+            rule=rule,
+            **eval_kwargs)
+        self.tmpdir = tmpdir
+        self.gpu_collect = gpu_collect
+
+    def after_train_epoch(self, runner):
+        if not self.evaluation_flag(runner):
+            return
+
+        from mmdet.apis import multi_gpu_test
+        tmpdir = self.tmpdir
+        if tmpdir is None:
+            tmpdir = osp.join(runner.work_dir, '.eval_hook')
+        results = multi_gpu_test(
+            runner.model,
+            self.dataloader,
+            tmpdir=tmpdir,
+            gpu_collect=self.gpu_collect)
+        if runner.rank == 0:
+            print('\n')
+            key_score = self.evaluate(runner, results)
+            if self.save_best:
+                best_score = runner.meta['hook_msgs'].get(
+                    'best_score', self.init_value_map[self.rule])
+                if self.compare_func(key_score, best_score):
+                    best_score = key_score
+                    runner.meta['hook_msgs']['best_score'] = best_score
+                    last_ckpt = runner.meta['hook_msgs']['last_ckpt']
+                    runner.meta['hook_msgs']['best_ckpt'] = last_ckpt
+                    mmcv.symlink(
+                        last_ckpt,
+                        osp.join(runner.work_dir,
+                                 f'best_{self.key_indicator}.pth'))
+                    self.logger.info(
+                        f'Now best checkpoint is {last_ckpt}.'
+                        f'Best {self.key_indicator} is {best_score:0.4f}')
diff --git a/detection/scrfd/mmdet/core/evaluation/mean_ap.py b/detection/scrfd/mmdet/core/evaluation/mean_ap.py
new file mode 100755
index 000000000..f44314b42
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/mean_ap.py
@@ -0,0 +1,469 @@
+from multiprocessing import Pool
+
+import mmcv
+import numpy as np
+from mmcv.utils import print_log
+from terminaltables import AsciiTable
+
+from .bbox_overlaps import bbox_overlaps
+from .class_names import get_classes
+
+
+def average_precision(recalls, precisions, mode='area'):
+    """Calculate average precision (for single or multiple scales).
+
+    Args:
+        recalls (ndarray): shape (num_scales, num_dets) or (num_dets, )
+        precisions (ndarray): shape (num_scales, num_dets) or (num_dets, )
+        mode (str): 'area' or '11points', 'area' means calculating the area
+            under precision-recall curve, '11points' means calculating
+            the average precision of recalls at [0, 0.1, ..., 1]
+
+    Returns:
+        float or ndarray: calculated average precision
+    """
+    no_scale = False
+    if recalls.ndim == 1:
+        no_scale = True
+        recalls = recalls[np.newaxis, :]
+        precisions = precisions[np.newaxis, :]
+    assert recalls.shape == precisions.shape and recalls.ndim == 2
+    num_scales = recalls.shape[0]
+    ap = np.zeros(num_scales, dtype=np.float32)
+    if mode == 'area':
+        zeros = np.zeros((num_scales, 1), dtype=recalls.dtype)
+        ones = np.ones((num_scales, 1), dtype=recalls.dtype)
+        mrec = np.hstack((zeros, recalls, ones))
+        mpre = np.hstack((zeros, precisions, zeros))
+        for i in range(mpre.shape[1] - 1, 0, -1):
+            mpre[:, i - 1] = np.maximum(mpre[:, i - 1], mpre[:, i])
+        for i in range(num_scales):
+            ind = np.where(mrec[i, 1:] != mrec[i, :-1])[0]
+            ap[i] = np.sum(
+                (mrec[i, ind + 1] - mrec[i, ind]) * mpre[i, ind + 1])
+    elif mode == '11points':
+        for i in range(num_scales):
+            for thr in np.arange(0, 1 + 1e-3, 0.1):
+                precs = precisions[i, recalls[i, :] >= thr]
+                prec = precs.max() if precs.size > 0 else 0
+                ap[i] += prec
+            ap /= 11
+    else:
+        raise ValueError(
+            'Unrecognized mode, only "area" and "11points" are supported')
+    if no_scale:
+        ap = ap[0]
+    return ap
+
+
+def tpfp_imagenet(det_bboxes,
+                  gt_bboxes,
+                  gt_bboxes_ignore=None,
+                  default_iou_thr=0.5,
+                  area_ranges=None):
+    """Check if detected bboxes are true positive or false positive.
+
+    Args:
+        det_bbox (ndarray): Detected bboxes of this image, of shape (m, 5).
+        gt_bboxes (ndarray): GT bboxes of this image, of shape (n, 4).
+        gt_bboxes_ignore (ndarray): Ignored gt bboxes of this image,
+            of shape (k, 4). Default: None
+        default_iou_thr (float): IoU threshold to be considered as matched for
+            medium and large bboxes (small ones have special rules).
+            Default: 0.5.
+        area_ranges (list[tuple] | None): Range of bbox areas to be evaluated,
+            in the format [(min1, max1), (min2, max2), ...]. Default: None.
+
+    Returns:
+        tuple[np.ndarray]: (tp, fp) whose elements are 0 and 1. The shape of
+            each array is (num_scales, m).
+    """
+    # an indicator of ignored gts
+    gt_ignore_inds = np.concatenate(
+        (np.zeros(gt_bboxes.shape[0], dtype=np.bool),
+         np.ones(gt_bboxes_ignore.shape[0], dtype=np.bool)))
+    # stack gt_bboxes and gt_bboxes_ignore for convenience
+    gt_bboxes = np.vstack((gt_bboxes, gt_bboxes_ignore))
+
+    num_dets = det_bboxes.shape[0]
+    num_gts = gt_bboxes.shape[0]
+    if area_ranges is None:
+        area_ranges = [(None, None)]
+    num_scales = len(area_ranges)
+    # tp and fp are of shape (num_scales, num_gts), each row is tp or fp
+    # of a certain scale.
+    tp = np.zeros((num_scales, num_dets), dtype=np.float32)
+    fp = np.zeros((num_scales, num_dets), dtype=np.float32)
+    if gt_bboxes.shape[0] == 0:
+        if area_ranges == [(None, None)]:
+            fp[...] = 1
+        else:
+            det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0]) * (
+                det_bboxes[:, 3] - det_bboxes[:, 1])
+            for i, (min_area, max_area) in enumerate(area_ranges):
+                fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
+        return tp, fp
+    ious = bbox_overlaps(det_bboxes, gt_bboxes - 1)
+    gt_w = gt_bboxes[:, 2] - gt_bboxes[:, 0]
+    gt_h = gt_bboxes[:, 3] - gt_bboxes[:, 1]
+    iou_thrs = np.minimum((gt_w * gt_h) / ((gt_w + 10.0) * (gt_h + 10.0)),
+                          default_iou_thr)
+    # sort all detections by scores in descending order
+    sort_inds = np.argsort(-det_bboxes[:, -1])
+    for k, (min_area, max_area) in enumerate(area_ranges):
+        gt_covered = np.zeros(num_gts, dtype=bool)
+        # if no area range is specified, gt_area_ignore is all False
+        if min_area is None:
+            gt_area_ignore = np.zeros_like(gt_ignore_inds, dtype=bool)
+        else:
+            gt_areas = gt_w * gt_h
+            gt_area_ignore = (gt_areas < min_area) | (gt_areas >= max_area)
+        for i in sort_inds:
+            max_iou = -1
+            matched_gt = -1
+            # find best overlapped available gt
+            for j in range(num_gts):
+                # different from PASCAL VOC: allow finding other gts if the
+                # best overlaped ones are already matched by other det bboxes
+                if gt_covered[j]:
+                    continue
+                elif ious[i, j] >= iou_thrs[j] and ious[i, j] > max_iou:
+                    max_iou = ious[i, j]
+                    matched_gt = j
+            # there are 4 cases for a det bbox:
+            # 1. it matches a gt, tp = 1, fp = 0
+            # 2. it matches an ignored gt, tp = 0, fp = 0
+            # 3. it matches no gt and within area range, tp = 0, fp = 1
+            # 4. it matches no gt but is beyond area range, tp = 0, fp = 0
+            if matched_gt >= 0:
+                gt_covered[matched_gt] = 1
+                if not (gt_ignore_inds[matched_gt]
+                        or gt_area_ignore[matched_gt]):
+                    tp[k, i] = 1
+            elif min_area is None:
+                fp[k, i] = 1
+            else:
+                bbox = det_bboxes[i, :4]
+                area = (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
+                if area >= min_area and area < max_area:
+                    fp[k, i] = 1
+    return tp, fp
+
+
+def tpfp_default(det_bboxes,
+                 gt_bboxes,
+                 gt_bboxes_ignore=None,
+                 iou_thr=0.5,
+                 area_ranges=None):
+    """Check if detected bboxes are true positive or false positive.
+
+    Args:
+        det_bbox (ndarray): Detected bboxes of this image, of shape (m, 5).
+        gt_bboxes (ndarray): GT bboxes of this image, of shape (n, 4).
+        gt_bboxes_ignore (ndarray): Ignored gt bboxes of this image,
+            of shape (k, 4). Default: None
+        iou_thr (float): IoU threshold to be considered as matched.
+            Default: 0.5.
+        area_ranges (list[tuple] | None): Range of bbox areas to be evaluated,
+            in the format [(min1, max1), (min2, max2), ...]. Default: None.
+
+    Returns:
+        tuple[np.ndarray]: (tp, fp) whose elements are 0 and 1. The shape of
+            each array is (num_scales, m).
+    """
+    # an indicator of ignored gts
+    gt_ignore_inds = np.concatenate(
+        (np.zeros(gt_bboxes.shape[0], dtype=np.bool),
+         np.ones(gt_bboxes_ignore.shape[0], dtype=np.bool)))
+    # stack gt_bboxes and gt_bboxes_ignore for convenience
+    gt_bboxes = np.vstack((gt_bboxes, gt_bboxes_ignore))
+
+    num_dets = det_bboxes.shape[0]
+    num_gts = gt_bboxes.shape[0]
+    if area_ranges is None:
+        area_ranges = [(None, None)]
+    num_scales = len(area_ranges)
+    # tp and fp are of shape (num_scales, num_gts), each row is tp or fp of
+    # a certain scale
+    tp = np.zeros((num_scales, num_dets), dtype=np.float32)
+    fp = np.zeros((num_scales, num_dets), dtype=np.float32)
+
+    # if there is no gt bboxes in this image, then all det bboxes
+    # within area range are false positives
+    if gt_bboxes.shape[0] == 0:
+        if area_ranges == [(None, None)]:
+            fp[...] = 1
+        else:
+            det_areas = (det_bboxes[:, 2] - det_bboxes[:, 0]) * (
+                det_bboxes[:, 3] - det_bboxes[:, 1])
+            for i, (min_area, max_area) in enumerate(area_ranges):
+                fp[i, (det_areas >= min_area) & (det_areas < max_area)] = 1
+        return tp, fp
+
+    ious = bbox_overlaps(det_bboxes, gt_bboxes)
+    # for each det, the max iou with all gts
+    ious_max = ious.max(axis=1)
+    # for each det, which gt overlaps most with it
+    ious_argmax = ious.argmax(axis=1)
+    # sort all dets in descending order by scores
+    sort_inds = np.argsort(-det_bboxes[:, -1])
+    for k, (min_area, max_area) in enumerate(area_ranges):
+        gt_covered = np.zeros(num_gts, dtype=bool)
+        # if no area range is specified, gt_area_ignore is all False
+        if min_area is None:
+            gt_area_ignore = np.zeros_like(gt_ignore_inds, dtype=bool)
+        else:
+            gt_areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0]) * (
+                gt_bboxes[:, 3] - gt_bboxes[:, 1])
+            gt_area_ignore = (gt_areas < min_area) | (gt_areas >= max_area)
+        for i in sort_inds:
+            if ious_max[i] >= iou_thr:
+                matched_gt = ious_argmax[i]
+                if not (gt_ignore_inds[matched_gt]
+                        or gt_area_ignore[matched_gt]):
+                    if not gt_covered[matched_gt]:
+                        gt_covered[matched_gt] = True
+                        tp[k, i] = 1
+                    else:
+                        fp[k, i] = 1
+                # otherwise ignore this detected bbox, tp = 0, fp = 0
+            elif min_area is None:
+                fp[k, i] = 1
+            else:
+                bbox = det_bboxes[i, :4]
+                area = (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
+                if area >= min_area and area < max_area:
+                    fp[k, i] = 1
+    return tp, fp
+
+
+def get_cls_results(det_results, annotations, class_id):
+    """Get det results and gt information of a certain class.
+
+    Args:
+        det_results (list[list]): Same as `eval_map()`.
+        annotations (list[dict]): Same as `eval_map()`.
+        class_id (int): ID of a specific class.
+
+    Returns:
+        tuple[list[np.ndarray]]: detected bboxes, gt bboxes, ignored gt bboxes
+    """
+    cls_dets = [img_res[class_id] for img_res in det_results]
+    cls_gts = []
+    cls_gts_ignore = []
+    for ann in annotations:
+        gt_inds = ann['labels'] == class_id
+        cls_gts.append(ann['bboxes'][gt_inds, :])
+
+        if ann.get('labels_ignore', None) is not None:
+            ignore_inds = ann['labels_ignore'] == class_id
+            cls_gts_ignore.append(ann['bboxes_ignore'][ignore_inds, :])
+        else:
+            cls_gts_ignore.append(np.empty((0, 4), dtype=np.float32))
+
+    return cls_dets, cls_gts, cls_gts_ignore
+
+
+def eval_map(det_results,
+             annotations,
+             scale_ranges=None,
+             iou_thr=0.5,
+             dataset=None,
+             logger=None,
+             tpfp_fn=None,
+             nproc=4):
+    """Evaluate mAP of a dataset.
+
+    Args:
+        det_results (list[list]): [[cls1_det, cls2_det, ...], ...].
+            The outer list indicates images, and the inner list indicates
+            per-class detected bboxes.
+        annotations (list[dict]): Ground truth annotations where each item of
+            the list indicates an image. Keys of annotations are:
+
+            - `bboxes`: numpy array of shape (n, 4)
+            - `labels`: numpy array of shape (n, )
+            - `bboxes_ignore` (optional): numpy array of shape (k, 4)
+            - `labels_ignore` (optional): numpy array of shape (k, )
+        scale_ranges (list[tuple] | None): Range of scales to be evaluated,
+            in the format [(min1, max1), (min2, max2), ...]. A range of
+            (32, 64) means the area range between (32**2, 64**2).
+            Default: None.
+        iou_thr (float): IoU threshold to be considered as matched.
+            Default: 0.5.
+        dataset (list[str] | str | None): Dataset name or dataset classes,
+            there are minor differences in metrics for different datsets, e.g.
+            "voc07", "imagenet_det", etc. Default: None.
+        logger (logging.Logger | str | None): The way to print the mAP
+            summary. See `mmdet.utils.print_log()` for details. Default: None.
+        tpfp_fn (callable | None): The function used to determine true/
+            false positives. If None, :func:`tpfp_default` is used as default
+            unless dataset is 'det' or 'vid' (:func:`tpfp_imagenet` in this
+            case). If it is given as a function, then this function is used
+            to evaluate tp & fp. Default None.
+        nproc (int): Processes used for computing TP and FP.
+            Default: 4.
+
+    Returns:
+        tuple: (mAP, [dict, dict, ...])
+    """
+    assert len(det_results) == len(annotations)
+
+    num_imgs = len(det_results)
+    num_scales = len(scale_ranges) if scale_ranges is not None else 1
+    num_classes = len(det_results[0])  # positive class num
+    area_ranges = ([(rg[0]**2, rg[1]**2) for rg in scale_ranges]
+                   if scale_ranges is not None else None)
+
+    pool = Pool(nproc)
+    eval_results = []
+    for i in range(num_classes):
+        # get gt and det bboxes of this class
+        cls_dets, cls_gts, cls_gts_ignore = get_cls_results(
+            det_results, annotations, i)
+        # choose proper function according to datasets to compute tp and fp
+        if tpfp_fn is None:
+            if dataset in ['det', 'vid']:
+                tpfp_fn = tpfp_imagenet
+            else:
+                tpfp_fn = tpfp_default
+        if not callable(tpfp_fn):
+            raise ValueError(
+                f'tpfp_fn has to be a function or None, but got {tpfp_fn}')
+
+        # compute tp and fp for each image with multiple processes
+        tpfp = pool.starmap(
+            tpfp_fn,
+            zip(cls_dets, cls_gts, cls_gts_ignore,
+                [iou_thr for _ in range(num_imgs)],
+                [area_ranges for _ in range(num_imgs)]))
+        tp, fp = tuple(zip(*tpfp))
+        # calculate gt number of each scale
+        # ignored gts or gts beyond the specific scale are not counted
+        num_gts = np.zeros(num_scales, dtype=int)
+        for j, bbox in enumerate(cls_gts):
+            if area_ranges is None:
+                num_gts[0] += bbox.shape[0]
+            else:
+                gt_areas = (bbox[:, 2] - bbox[:, 0]) * (
+                    bbox[:, 3] - bbox[:, 1])
+                for k, (min_area, max_area) in enumerate(area_ranges):
+                    num_gts[k] += np.sum((gt_areas >= min_area)
+                                         & (gt_areas < max_area))
+        # sort all det bboxes by score, also sort tp and fp
+        cls_dets = np.vstack(cls_dets)
+        num_dets = cls_dets.shape[0]
+        sort_inds = np.argsort(-cls_dets[:, -1])
+        tp = np.hstack(tp)[:, sort_inds]
+        fp = np.hstack(fp)[:, sort_inds]
+        # calculate recall and precision with tp and fp
+        tp = np.cumsum(tp, axis=1)
+        fp = np.cumsum(fp, axis=1)
+        eps = np.finfo(np.float32).eps
+        recalls = tp / np.maximum(num_gts[:, np.newaxis], eps)
+        precisions = tp / np.maximum((tp + fp), eps)
+        # calculate AP
+        if scale_ranges is None:
+            recalls = recalls[0, :]
+            precisions = precisions[0, :]
+            num_gts = num_gts.item()
+        mode = 'area' if dataset != 'voc07' else '11points'
+        ap = average_precision(recalls, precisions, mode)
+        eval_results.append({
+            'num_gts': num_gts,
+            'num_dets': num_dets,
+            'recall': recalls,
+            'precision': precisions,
+            'ap': ap
+        })
+    pool.close()
+    if scale_ranges is not None:
+        # shape (num_classes, num_scales)
+        all_ap = np.vstack([cls_result['ap'] for cls_result in eval_results])
+        all_num_gts = np.vstack(
+            [cls_result['num_gts'] for cls_result in eval_results])
+        mean_ap = []
+        for i in range(num_scales):
+            if np.any(all_num_gts[:, i] > 0):
+                mean_ap.append(all_ap[all_num_gts[:, i] > 0, i].mean())
+            else:
+                mean_ap.append(0.0)
+    else:
+        aps = []
+        for cls_result in eval_results:
+            if cls_result['num_gts'] > 0:
+                aps.append(cls_result['ap'])
+        mean_ap = np.array(aps).mean().item() if aps else 0.0
+
+    print_map_summary(
+        mean_ap, eval_results, dataset, area_ranges, logger=logger)
+
+    return mean_ap, eval_results
+
+
+def print_map_summary(mean_ap,
+                      results,
+                      dataset=None,
+                      scale_ranges=None,
+                      logger=None):
+    """Print mAP and results of each class.
+
+    A table will be printed to show the gts/dets/recall/AP of each class and
+    the mAP.
+
+    Args:
+        mean_ap (float): Calculated from `eval_map()`.
+        results (list[dict]): Calculated from `eval_map()`.
+        dataset (list[str] | str | None): Dataset name or dataset classes.
+        scale_ranges (list[tuple] | None): Range of scales to be evaluated.
+        logger (logging.Logger | str | None): The way to print the mAP
+            summary. See `mmdet.utils.print_log()` for details. Default: None.
+    """
+
+    if logger == 'silent':
+        return
+
+    if isinstance(results[0]['ap'], np.ndarray):
+        num_scales = len(results[0]['ap'])
+    else:
+        num_scales = 1
+
+    if scale_ranges is not None:
+        assert len(scale_ranges) == num_scales
+
+    num_classes = len(results)
+
+    recalls = np.zeros((num_scales, num_classes), dtype=np.float32)
+    aps = np.zeros((num_scales, num_classes), dtype=np.float32)
+    num_gts = np.zeros((num_scales, num_classes), dtype=int)
+    for i, cls_result in enumerate(results):
+        if cls_result['recall'].size > 0:
+            recalls[:, i] = np.array(cls_result['recall'], ndmin=2)[:, -1]
+        aps[:, i] = cls_result['ap']
+        num_gts[:, i] = cls_result['num_gts']
+
+    if dataset is None:
+        label_names = [str(i) for i in range(num_classes)]
+    elif mmcv.is_str(dataset):
+        label_names = get_classes(dataset)
+    else:
+        label_names = dataset
+
+    if not isinstance(mean_ap, list):
+        mean_ap = [mean_ap]
+
+    header = ['class', 'gts', 'dets', 'recall', 'ap']
+    for i in range(num_scales):
+        if scale_ranges is not None:
+            print_log(f'Scale range {scale_ranges[i]}', logger=logger)
+        table_data = [header]
+        for j in range(num_classes):
+            row_data = [
+                label_names[j], num_gts[i, j], results[j]['num_dets'],
+                f'{recalls[i, j]:.3f}', f'{aps[i, j]:.3f}'
+            ]
+            table_data.append(row_data)
+        table_data.append(['mAP', '', '', '', f'{mean_ap[i]:.3f}'])
+        table = AsciiTable(table_data)
+        table.inner_footing_row_border = True
+        print_log('\n' + table.table, logger=logger)
diff --git a/detection/scrfd/mmdet/core/evaluation/recall.py b/detection/scrfd/mmdet/core/evaluation/recall.py
new file mode 100755
index 000000000..ea6277abc
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/recall.py
@@ -0,0 +1,189 @@
+from collections.abc import Sequence
+
+import numpy as np
+from mmcv.utils import print_log
+from terminaltables import AsciiTable
+
+from .bbox_overlaps import bbox_overlaps
+
+
+def _recalls(all_ious, proposal_nums, thrs):
+
+    img_num = all_ious.shape[0]
+    total_gt_num = sum([ious.shape[0] for ious in all_ious])
+
+    _ious = np.zeros((proposal_nums.size, total_gt_num), dtype=np.float32)
+    for k, proposal_num in enumerate(proposal_nums):
+        tmp_ious = np.zeros(0)
+        for i in range(img_num):
+            ious = all_ious[i][:, :proposal_num].copy()
+            gt_ious = np.zeros((ious.shape[0]))
+            if ious.size == 0:
+                tmp_ious = np.hstack((tmp_ious, gt_ious))
+                continue
+            for j in range(ious.shape[0]):
+                gt_max_overlaps = ious.argmax(axis=1)
+                max_ious = ious[np.arange(0, ious.shape[0]), gt_max_overlaps]
+                gt_idx = max_ious.argmax()
+                gt_ious[j] = max_ious[gt_idx]
+                box_idx = gt_max_overlaps[gt_idx]
+                ious[gt_idx, :] = -1
+                ious[:, box_idx] = -1
+            tmp_ious = np.hstack((tmp_ious, gt_ious))
+        _ious[k, :] = tmp_ious
+
+    _ious = np.fliplr(np.sort(_ious, axis=1))
+    recalls = np.zeros((proposal_nums.size, thrs.size))
+    for i, thr in enumerate(thrs):
+        recalls[:, i] = (_ious >= thr).sum(axis=1) / float(total_gt_num)
+
+    return recalls
+
+
+def set_recall_param(proposal_nums, iou_thrs):
+    """Check proposal_nums and iou_thrs and set correct format."""
+    if isinstance(proposal_nums, Sequence):
+        _proposal_nums = np.array(proposal_nums)
+    elif isinstance(proposal_nums, int):
+        _proposal_nums = np.array([proposal_nums])
+    else:
+        _proposal_nums = proposal_nums
+
+    if iou_thrs is None:
+        _iou_thrs = np.array([0.5])
+    elif isinstance(iou_thrs, Sequence):
+        _iou_thrs = np.array(iou_thrs)
+    elif isinstance(iou_thrs, float):
+        _iou_thrs = np.array([iou_thrs])
+    else:
+        _iou_thrs = iou_thrs
+
+    return _proposal_nums, _iou_thrs
+
+
+def eval_recalls(gts,
+                 proposals,
+                 proposal_nums=None,
+                 iou_thrs=0.5,
+                 logger=None):
+    """Calculate recalls.
+
+    Args:
+        gts (list[ndarray]): a list of arrays of shape (n, 4)
+        proposals (list[ndarray]): a list of arrays of shape (k, 4) or (k, 5)
+        proposal_nums (int | Sequence[int]): Top N proposals to be evaluated.
+        iou_thrs (float | Sequence[float]): IoU thresholds. Default: 0.5.
+        logger (logging.Logger | str | None): The way to print the recall
+            summary. See `mmdet.utils.print_log()` for details. Default: None.
+
+    Returns:
+        ndarray: recalls of different ious and proposal nums
+    """
+
+    img_num = len(gts)
+    assert img_num == len(proposals)
+
+    proposal_nums, iou_thrs = set_recall_param(proposal_nums, iou_thrs)
+
+    all_ious = []
+    for i in range(img_num):
+        if proposals[i].ndim == 2 and proposals[i].shape[1] == 5:
+            scores = proposals[i][:, 4]
+            sort_idx = np.argsort(scores)[::-1]
+            img_proposal = proposals[i][sort_idx, :]
+        else:
+            img_proposal = proposals[i]
+        prop_num = min(img_proposal.shape[0], proposal_nums[-1])
+        if gts[i] is None or gts[i].shape[0] == 0:
+            ious = np.zeros((0, img_proposal.shape[0]), dtype=np.float32)
+        else:
+            ious = bbox_overlaps(gts[i], img_proposal[:prop_num, :4])
+        all_ious.append(ious)
+    all_ious = np.array(all_ious)
+    recalls = _recalls(all_ious, proposal_nums, iou_thrs)
+
+    print_recall_summary(recalls, proposal_nums, iou_thrs, logger=logger)
+    return recalls
+
+
+def print_recall_summary(recalls,
+                         proposal_nums,
+                         iou_thrs,
+                         row_idxs=None,
+                         col_idxs=None,
+                         logger=None):
+    """Print recalls in a table.
+
+    Args:
+        recalls (ndarray): calculated from `bbox_recalls`
+        proposal_nums (ndarray or list): top N proposals
+        iou_thrs (ndarray or list): iou thresholds
+        row_idxs (ndarray): which rows(proposal nums) to print
+        col_idxs (ndarray): which cols(iou thresholds) to print
+        logger (logging.Logger | str | None): The way to print the recall
+            summary. See `mmdet.utils.print_log()` for details. Default: None.
+    """
+    proposal_nums = np.array(proposal_nums, dtype=np.int32)
+    iou_thrs = np.array(iou_thrs)
+    if row_idxs is None:
+        row_idxs = np.arange(proposal_nums.size)
+    if col_idxs is None:
+        col_idxs = np.arange(iou_thrs.size)
+    row_header = [''] + iou_thrs[col_idxs].tolist()
+    table_data = [row_header]
+    for i, num in enumerate(proposal_nums[row_idxs]):
+        row = [f'{val:.3f}' for val in recalls[row_idxs[i], col_idxs].tolist()]
+        row.insert(0, num)
+        table_data.append(row)
+    table = AsciiTable(table_data)
+    print_log('\n' + table.table, logger=logger)
+
+
+def plot_num_recall(recalls, proposal_nums):
+    """Plot Proposal_num-Recalls curve.
+
+    Args:
+        recalls(ndarray or list): shape (k,)
+        proposal_nums(ndarray or list): same shape as `recalls`
+    """
+    if isinstance(proposal_nums, np.ndarray):
+        _proposal_nums = proposal_nums.tolist()
+    else:
+        _proposal_nums = proposal_nums
+    if isinstance(recalls, np.ndarray):
+        _recalls = recalls.tolist()
+    else:
+        _recalls = recalls
+
+    import matplotlib.pyplot as plt
+    f = plt.figure()
+    plt.plot([0] + _proposal_nums, [0] + _recalls)
+    plt.xlabel('Proposal num')
+    plt.ylabel('Recall')
+    plt.axis([0, proposal_nums.max(), 0, 1])
+    f.show()
+
+
+def plot_iou_recall(recalls, iou_thrs):
+    """Plot IoU-Recalls curve.
+
+    Args:
+        recalls(ndarray or list): shape (k,)
+        iou_thrs(ndarray or list): same shape as `recalls`
+    """
+    if isinstance(iou_thrs, np.ndarray):
+        _iou_thrs = iou_thrs.tolist()
+    else:
+        _iou_thrs = iou_thrs
+    if isinstance(recalls, np.ndarray):
+        _recalls = recalls.tolist()
+    else:
+        _recalls = recalls
+
+    import matplotlib.pyplot as plt
+    f = plt.figure()
+    plt.plot(_iou_thrs + [1.0], _recalls + [0.])
+    plt.xlabel('IoU')
+    plt.ylabel('Recall')
+    plt.axis([iou_thrs.min(), 1, 0, 1])
+    f.show()
diff --git a/detection/scrfd/mmdet/core/evaluation/widerface.py b/detection/scrfd/mmdet/core/evaluation/widerface.py
new file mode 100755
index 000000000..a1a3ad375
--- /dev/null
+++ b/detection/scrfd/mmdet/core/evaluation/widerface.py
@@ -0,0 +1,564 @@
+"""
+WiderFace evaluation code
+author: wondervictor
+mail: tianhengcheng@gmail.com
+copyright@wondervictor
+"""
+
+from __future__ import absolute_import
+import os
+import tqdm
+import pickle
+import datetime
+import argparse
+import numpy as np
+from scipy.io import loadmat
+#from facedet.evaluation.box_utils import jaccard
+#from facedet.evaluation.bbox import bbox_overlaps
+#import torch
+#from mmdet.core.bbox import bbox_overlaps
+
+#def intersect(box_a, box_b):
+#    A = box_a.size(0)
+#    B = box_b.size(0)
+#    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
+#                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
+#    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
+#                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
+#    inter = torch.clamp((max_xy - min_xy), min=0)
+#    return inter[:, :, 0] * inter[:, :, 1]
+#
+#def jaccard(box_a, box_b):
+#    inter = intersect(box_a, box_b)
+#    #torch.cuda.empty_cache()
+#    if not inter.is_cuda:
+#        box_a_cpu = box_a.cpu()
+#        box_b_cpu = box_b.cpu()
+#        area_a_cpu = ((box_a_cpu[:, 2]-box_a_cpu[:, 0]) *
+#              (box_a_cpu[:, 3]-box_a_cpu[:, 1])).unsqueeze(1).expand_as(inter)  # [A,B]
+#        area_b_cpu = ((box_b_cpu[:, 2]-box_b_cpu[:, 0]) *
+#              (box_b_cpu[:, 3]-box_b_cpu[:, 1])).unsqueeze(0).expand_as(inter)  # [A,B]
+#        union_cpu = area_a_cpu + area_b_cpu - inter.cpu()
+#        return inter / union_cpu
+#    else:
+#        area_a = ((box_a[:, 2]-box_a[:, 0]) *
+#              (box_a[:, 3]-box_a[:, 1])).unsqueeze(1).expand_as(inter)  # [A,B]
+#        area_b = ((box_b[:, 2]-box_b[:, 0]) *
+#              (box_b[:, 3]-box_b[:, 1])).unsqueeze(0).expand_as(inter)  # [A,B]
+#        union = area_a + area_b - inter
+#
+#        return inter / union  # [A,B]
+#
+def bbox_overlaps(boxes, query_boxes):
+    n_ = boxes.shape[0]
+    k_ = query_boxes.shape[0]
+    overlaps = np.zeros((n_, k_), dtype=np.float)
+    for k in range(k_):
+        query_box_area = (query_boxes[k, 2] - query_boxes[k, 0] +
+                          1) * (query_boxes[k, 3] - query_boxes[k, 1] + 1)
+        for n in range(n_):
+            iw = min(boxes[n, 2], query_boxes[k, 2]) - max(
+                boxes[n, 0], query_boxes[k, 0]) + 1
+            if iw > 0:
+                ih = min(boxes[n, 3], query_boxes[k, 3]) - max(
+                    boxes[n, 1], query_boxes[k, 1]) + 1
+                if ih > 0:
+                    box_area = (boxes[n, 2] - boxes[n, 0] +
+                                1) * (boxes[n, 3] - boxes[n, 1] + 1)
+                    all_area = float(box_area + query_box_area - iw * ih)
+                    overlaps[n, k] = iw * ih / all_area
+    return overlaps
+
+def bbox_overlap(a, b):
+    x1 = np.maximum(a[:,0], b[0])
+    y1 = np.maximum(a[:,1], b[1])
+    x2 = np.minimum(a[:,2], b[2])
+    y2 = np.minimum(a[:,3], b[3])
+    w = x2-x1+1
+    h = y2-y1+1
+    inter = w*h
+    aarea = (a[:,2]-a[:,0]+1) * (a[:,3]-a[:,1]+1)
+    barea = (b[2]-b[0]+1) * (b[3]-b[1]+1)
+    o = inter / (aarea+barea-inter)
+    o[w<=0] = 0
+    o[h<=0] = 0
+    return o
+
+def __bbox_overlap(a, b):
+    x1 = torch.max(a[:,0], b[0])
+    y1 = torch.max(a[:,1], b[1])
+    x2 = torch.min(a[:,2], b[2])
+    y2 = torch.min(a[:,3], b[3])
+    w = x2-x1+1
+    h = y2-y1+1
+    inter = w*h
+    aarea = (a[:,2]-a[:,0]+1) * (a[:,3]-a[:,1]+1)
+    barea = (b[2]-b[0]+1) * (b[3]-b[1]+1)
+    o = inter / (aarea+barea-inter)
+    o[w<=0] = 0
+    o[h<=0] = 0
+    return o
+
+def np_around(array, num_decimals=0):
+    #return array
+    return np.around(array, decimals=num_decimals)
+
+#def compute_iou(box_a, box_b):
+#    x0 = np.maximum(box_a[:,0], box_b[0])
+#    y0 = np.maximum(box_a[:,1], box_b[1])
+#    x1 = np.minimum(box_a[:,2], box_b[2])
+#    y1 = np.minimum(box_a[:,3], box_b[3])
+#    #print ('x0', x0[0], x1[0], y0[0], y1[0], box_a[0], box_b[:])
+#    #w = np.maximum(x1 - x0 + 1, 0) 
+#    w = np_around(x1 - x0 + 1) 
+#    #h = np.maximum(y1 - y0 + 1, 0)
+#    h = np_around(y1 - y0 + 1)
+#    inter = np_around(w * h)
+#    area_a = (box_a[:,2] - box_a[:,0] + 1) * (box_a[:,3] - box_a[:,1] + 1)
+#    area_a = np_around(area_a)
+#    area_b = (box_b[2] - box_b[0] + 1) * (box_b[3] - box_b[1] + 1)
+#    area_b = np_around(area_b)
+#    iou = inter / (area_a + area_b - inter)
+#    iou[w <= 0] = 0
+#    iou[h <=0] = 0
+#    return iou
+
+def np_round(val, decimals=4):
+    return val
+    #if isinstance(val, np.ndarray):
+    #    val = np.around(val, decimals=decimals)
+    #return val
+
+
+def get_gt_boxes(gt_dir):
+    """ gt dir: (wider_face_val.mat, wider_easy_val.mat, wider_medium_val.mat, wider_hard_val.mat)"""
+
+    gt_mat = loadmat(os.path.join(gt_dir, 'wider_face_val.mat'))
+    hard_mat = loadmat(os.path.join(gt_dir, 'wider_hard_val.mat'))
+    medium_mat = loadmat(os.path.join(gt_dir, 'wider_medium_val.mat'))
+    easy_mat = loadmat(os.path.join(gt_dir, 'wider_easy_val.mat'))
+
+    facebox_list = gt_mat['face_bbx_list']
+    event_list = gt_mat['event_list']
+    file_list = gt_mat['file_list']
+
+    hard_gt_list = hard_mat['gt_list']
+    medium_gt_list = medium_mat['gt_list']
+    easy_gt_list = easy_mat['gt_list']
+
+    return facebox_list, event_list, file_list, hard_gt_list, medium_gt_list, easy_gt_list
+
+
+def get_gt_boxes_from_txt(gt_path, cache_dir):
+
+    cache_file = os.path.join(cache_dir, 'gt_cache.pkl')
+    if os.path.exists(cache_file):
+        with open(cache_file, 'rb') as f:
+            boxes = pickle.load(f)
+        return boxes
+
+    f = open(gt_path, 'r')
+    state = 0
+    lines = f.readlines()
+    lines = list(map(lambda x: x.rstrip('\r\n'), lines))
+    boxes = {}
+    #print(len(lines))
+    f.close()
+    current_boxes = []
+    current_name = None
+    for line in lines:
+        if state == 0 and '--' in line:
+            state = 1
+            current_name = line
+            continue
+        if state == 1:
+            state = 2
+            continue
+
+        if state == 2 and '--' in line:
+            state = 1
+            boxes[current_name] = np.array(current_boxes).astype('float32')
+            current_name = line
+            current_boxes = []
+            continue
+
+        if state == 2:
+            box = [float(x) for x in line.split(' ')[:4]]
+            current_boxes.append(box)
+            continue
+
+    with open(cache_file, 'wb') as f:
+        pickle.dump(boxes, f)
+    return boxes
+
+
+def read_pred_file(filepath):
+
+    with open(filepath, 'r') as f:
+        lines = f.readlines()
+        img_file = lines[0].rstrip('\n\r')
+        lines = lines[2:]
+
+    boxes = np.array(list(map(lambda x: [float(a) for a in x.rstrip('\r\n').split(' ')], lines))).astype('float')
+    return img_file.split('/')[-1], boxes
+
+
+def get_preds(pred_dir):
+    events = os.listdir(pred_dir)
+    boxes = dict()
+    pbar = tqdm.tqdm(events)
+
+    for event in pbar:
+        pbar.set_description('Reading Predictions ')
+        event_dir = os.path.join(pred_dir, event)
+        event_images = os.listdir(event_dir)
+        current_event = dict()
+        for imgtxt in event_images:
+            imgname, _boxes = read_pred_file(os.path.join(event_dir, imgtxt))
+            current_event[imgname.rstrip('.jpg')] = _boxes
+        boxes[event] = current_event
+    return boxes
+
+
+def norm_score(pred):
+    """ norm score
+    pred {key: [[x1,y1,x2,y2,s]]}
+    """
+
+    max_score = -1 
+    min_score = 2
+
+    for _, k in pred.items():
+        for _, v in k.items():
+            if len(v) == 0:
+                continue
+            _min = np.min(v[:, -1])
+            _max = np.max(v[:, -1])
+            max_score = max(_max, max_score)
+            min_score = min(_min, min_score)
+
+    diff = max_score - min_score
+    for _, k in pred.items():
+        for _, v in k.items():
+            if len(v) == 0:
+                continue
+            v[:, -1] = (v[:, -1] - min_score).astype(np.float64)/diff
+    return pred
+
+
+def image_eval(pred, gt, ignore, iou_thresh, mpp):
+    """ single image evaluation
+    pred: Nx5
+    gt: Nx4
+    ignore:
+    """
+
+    
+    _pred = pred.copy()
+    _gt = gt.copy()
+    pred_recall = np.zeros(_pred.shape[0])
+    recall_list = np.zeros(_gt.shape[0])
+    proposal_list = np.ones(_pred.shape[0])
+
+    _pred[:, 2] = _pred[:, 2] + _pred[:, 0]
+    _pred[:, 3] = _pred[:, 3] + _pred[:, 1]
+    _gt[:, 2] = _gt[:, 2] + _gt[:, 0]
+    _gt[:, 3] = _gt[:, 3] + _gt[:, 1]
+
+    gt_overlap_list = mpp.starmap(bbox_overlap, zip([_gt]*_pred.shape[0],[_pred[h] for h in range(_pred.shape[0])]))
+
+    #use_cuda = True
+    #if use_cuda:
+    #    _pred = torch.cuda.FloatTensor(_pred[:,:4])
+    #    _gt = torch.cuda.FloatTensor(_gt)
+    #else:
+    #    _pred = torch.FloatTensor(_pred[:,:4])
+    #    _gt = torch.FloatTensor(_gt)
+
+    #overlaps = jaccard(_pred, _gt).cpu().numpy()
+    #overlaps = compute_iou((_pred[:, :4]), (_gt))
+
+    #overlaps = bbox_overlaps(_pred, _gt)
+
+    #if use_cuda:
+    #    overlaps = overlaps.cpu().numpy()
+    #else:
+    #    overlaps = overlaps.numpy()
+
+    for h in range(_pred.shape[0]):
+
+        #gt_overlap = overlaps[h]
+        #gt_overlap = bbox_overlap(_gt, _pred[h])
+        gt_overlap = gt_overlap_list[h]
+        #if use_cuda:
+        #    gt_overlap = gt_overlap.cpu().numpy()
+        #else:
+        #    gt_overlap = gt_overlap.numpy()
+
+        #max_overlap, max_idx = gt_overlap.max(), gt_overlap.argmax()
+        #gt_overlap = compute_iou(_gt, _pred[h, :4])
+        #exit()
+        #exit()
+        #print ('overlap', gt_overlap)
+        max_overlap, max_idx = gt_overlap.max(), gt_overlap.argmax()
+
+        if max_overlap >= iou_thresh:
+            if ignore[max_idx] == 0:
+                recall_list[max_idx] = -1
+                proposal_list[h] = -1
+            elif recall_list[max_idx] == 0:
+                recall_list[max_idx] = 1
+
+        r_keep_index = np.where(recall_list == 1)[0]
+        pred_recall[h] = len(r_keep_index)
+
+    return pred_recall, proposal_list
+
+
+def img_pr_info(thresh_num, pred_info, proposal_list, pred_recall):
+    pr_info = np.zeros((thresh_num, 2)).astype('float')
+    fp = np.zeros((pred_info.shape[0],), dtype=np.int)
+    last_info = [-1, -1]
+    for t in range(thresh_num):
+
+        thresh = 1 - (t+1)/thresh_num
+        r_index = np.where(pred_info[:, 4] >= thresh)[0]
+        if len(r_index) == 0:
+            pr_info[t, 0] = 0
+            pr_info[t, 1] = 0
+        else:
+            r_index = r_index[-1]
+            p_index = np.where(proposal_list[:r_index+1] == 1)[0]
+            pr_info[t, 0] = len(p_index) #valid pred number
+            pr_info[t, 1] = pred_recall[r_index] # valid gt number
+
+            if t>0 and pr_info[t, 0] > pr_info[t-1,0] and pr_info[t, 1]==pr_info[t-1,1]:
+                fp[r_index] = 1
+                #if thresh>=0.85:
+                #    print(thresh, t, pr_info[t])
+    #print(pr_info[:10,0])
+    #print(pr_info[:10,1])
+    return pr_info, fp
+
+
+def dataset_pr_info(thresh_num, pr_curve, count_face):
+    _pr_curve = np.zeros((thresh_num, 2))
+    for i in range(thresh_num):
+        #_pr_curve[i, 0] = round(pr_curve[i, 1] / pr_curve[i, 0], 4)
+        #_pr_curve[i, 1] = round(pr_curve[i, 1] / count_face, 4)
+        _pr_curve[i, 0] = pr_curve[i, 1] / pr_curve[i, 0]
+        _pr_curve[i, 1] = pr_curve[i, 1] / count_face
+    return _pr_curve
+
+
+def voc_ap(rec, prec):
+
+    # correct AP calculation
+    # first append sentinel values at the end
+    #print ('rec:', rec)
+    #print ('pre:', prec)
+    mrec = np.concatenate(([0.], rec, [1.]))
+    mpre = np.concatenate(([0.], prec, [0.]))
+
+    # compute the precision envelope
+    for i in range(mpre.size - 1, 0, -1):
+        mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
+
+    # to calculate area under PR curve, look for points
+    # where X axis (recall) changes value
+    i = np.where(mrec[1:] != mrec[:-1])[0]
+
+    # and sum (\Delta recall) * prec
+    ap = np_round(np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]))
+    return ap
+
+
+def wider_evaluation(pred, gt_path, iou_thresh=0.5, debug=False):
+    #pred = get_preds(pred)
+    pred = norm_score(pred)
+    thresh_num = 1000
+    #thresh_num = 2000
+    facebox_list, event_list, file_list, hard_gt_list, medium_gt_list, easy_gt_list = get_gt_boxes(gt_path)
+    event_num = len(event_list)
+    settings = ['easy', 'medium', 'hard']
+    setting_gts = [easy_gt_list, medium_gt_list, hard_gt_list]
+    from multiprocessing import Pool
+    #from multiprocessing.pool import ThreadPool
+    mpp = Pool(8)
+    aps = [-1.0, -1.0, -1.0]
+    meta = {}
+    #setting_id = 2
+    print('')
+    for setting_id in range(3):
+    #for setting_id in range(1):
+        ta = datetime.datetime.now()
+        # different setting
+        #iou_th = 0.5 #+ 0.05 * idx
+        iou_th = iou_thresh
+        # different setting
+        gt_list = setting_gts[setting_id]
+        count_face = 0
+        pr_curve = np.zeros((thresh_num, 2)).astype('float')
+        # [hard, medium, easy]
+        #pbar = tqdm.tqdm(range(event_num))
+        #for i in pbar:
+        high_score_count = 0
+        high_score_fp_count = 0
+        for i in range(event_num):
+            #pbar.set_description('Processing {}'.format(settings[setting_id]))
+            event_name = str(event_list[i][0][0])
+            img_list = file_list[i][0]
+            pred_list = pred[event_name]
+            sub_gt_list = gt_list[i][0]
+            # img_pr_info_list = np.zeros((len(img_list), thresh_num, 2))
+            gt_bbx_list = facebox_list[i][0]
+
+            for j in range(len(img_list)):
+                img_name = str(img_list[j][0][0])
+                pred_info = pred_list[img_name]
+
+                gt_boxes = gt_bbx_list[j][0].astype('float')
+                keep_index = sub_gt_list[j][0]
+                #print ('keep_index', keep_index)
+                count_face += len(keep_index)
+                
+
+                if len(gt_boxes) == 0 or len(pred_info) == 0:
+                    continue
+                #ignore = np.zeros(gt_boxes.shape[0])
+                #if len(keep_index) != 0:
+                #    ignore[keep_index-1] = 1
+                #assert len(keep_index)>0
+                ignore = np.zeros(gt_boxes.shape[0], dtype=np.int)
+                if len(keep_index) != 0:
+                    ignore[keep_index-1] = 1
+                pred_info = np_round(pred_info,1)
+                #print('ignore:', len(ignore), len(np.where(ignore==1)[0]))
+                #pred_sort_idx= np.argsort(pred_info[:,4])
+                #pred_info = pred_info[pred_sort_idx][::-1]
+                #print ('pred_info', pred_info[:20, 4])
+                #exit()
+
+
+                gt_boxes = np_round(gt_boxes)
+                #ignore = np_round(ignore)
+                pred_recall, proposal_list = image_eval(pred_info, gt_boxes, ignore, iou_th, mpp)
+                #print(pred_recall[:10], proposal_list[:10])
+                #print('1 stage', pred_recall, proposal_list)
+                #print(pred_info.shape, pred_recall.shape)
+
+                _img_pr_info, fp = img_pr_info(thresh_num, pred_info, proposal_list, pred_recall)
+                #for f in range(pred_info.shape[0]):
+                #    _score = pred_info[f,4]
+                #    if _score<0.929:
+                #        break
+                #    high_score_count+=1
+                #    if fp[f]==1:
+                #        w = pred_info[f, 2]
+                #        h = pred_info[f, 3]
+                #        print('fp:', event_name, img_name, _score, w, h)
+                #        high_score_fp_count+=1
+                pr_curve += _img_pr_info
+        #print ('pr_curve', pr_curve, count_face)
+        pr_curve = dataset_pr_info(thresh_num, pr_curve, count_face)
+        #print(pr_curve.shape)
+
+        propose = pr_curve[:, 0]
+        recall = pr_curve[:, 1]
+        #for f in range(thresh_num):
+        #    print('R-P:', recall[f], propose[f])
+        for srecall in np.arange(0.1, 1.0001, 0.1):
+            rindex = len(np.where(recall<=srecall)[0])-1
+            rthresh = 1.0 - float(rindex)/thresh_num
+            print('Recall-Precision-Thresh:', recall[rindex], propose[rindex], rthresh)
+
+        ap = voc_ap(recall, propose)
+        aps[setting_id] = ap
+        tb = datetime.datetime.now()
+        #print('high score count:', high_score_count)
+        #print('high score fp count:', high_score_fp_count)
+        print('%s cost %.4f seconds, ap: %.5f'%(settings[setting_id], (tb-ta).total_seconds(), ap))
+
+    return aps
+
+def get_widerface_gts(gt_path):
+    facebox_list, event_list, file_list, hard_gt_list, medium_gt_list, easy_gt_list = get_gt_boxes(gt_path)
+    event_num = len(event_list)
+
+    settings = ['easy', 'medium', 'hard']
+    setting_gts = [easy_gt_list, medium_gt_list, hard_gt_list]
+    all_results = []
+    for setting_id in range(3):
+        results = {}
+        gt_list = setting_gts[setting_id]
+        count_face = 0
+        # [hard, medium, easy]
+        #pbar = tqdm.tqdm(range(event_num))
+        #for i in pbar:
+        for i in range(event_num):
+            #pbar.set_description('Processing {}'.format(settings[setting_id]))
+            event_name = str(event_list[i][0][0])
+            img_list = file_list[i][0]
+            sub_gt_list = gt_list[i][0]
+            # img_pr_info_list = np.zeros((len(img_list), thresh_num, 2))
+            gt_bbx_list = facebox_list[i][0]
+            results[event_name] = {}
+
+            for j in range(len(img_list)):
+
+                gt_boxes = gt_bbx_list[j][0].astype('float').copy()
+                gt_boxes[:,2] += gt_boxes[:,0]
+                gt_boxes[:,3] += gt_boxes[:,1]
+                keep_index = sub_gt_list[j][0].copy()
+                #print ('keep_index', keep_index.shape)
+                count_face += len(keep_index)
+                
+
+                if len(gt_boxes) == 0:
+                    results[event_name][str(img_list[j][0][0])] = np.empty( (0,4) )
+                    continue
+                keep_index -= 1
+                keep_index = keep_index.flatten()
+                #ignore = np.zeros(gt_boxes.shape[0])
+                #if len(keep_index) != 0:
+                #    ignore[keep_index-1] = 1
+                #assert len(keep_index)>0
+                #ignore = np.zeros(gt_boxes.shape[0], dtype=np.int)
+                #if len(keep_index) != 0:
+                #    ignore[keep_index-1] = 1
+                #print('ignore:', len(ignore), len(np.where(ignore==1)[0]))
+                #pred_sort_idx= np.argsort(pred_info[:,4])
+                #pred_info = pred_info[pred_sort_idx][::-1]
+                #print ('pred_info', pred_info[:20, 4])
+                #exit()
+                #if setting_id==2 and len(keep_index)<gt_boxes.shape[0]:
+                #    print(gt_boxes.shape, keep_index.shape)
+                
+                gt_boxes = np_round(gt_boxes)[keep_index,:]
+
+                results[event_name][str(img_list[j][0][0])] = gt_boxes
+        all_results.append(results)
+    return all_results
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-p', '--pred', default='')
+    parser.add_argument('-g', '--gt', default='./ground_truth/')
+
+    args = parser.parse_args()
+    evaluation(args.pred, args.gt)
+
+
+
+
+
+
+
+
+
+
+
+
+
+
diff --git a/detection/scrfd/mmdet/core/export/__init__.py b/detection/scrfd/mmdet/core/export/__init__.py
new file mode 100755
index 000000000..76589b1f2
--- /dev/null
+++ b/detection/scrfd/mmdet/core/export/__init__.py
@@ -0,0 +1,8 @@
+from .pytorch2onnx import (build_model_from_cfg,
+                           generate_inputs_and_wrap_model,
+                           preprocess_example_input)
+
+__all__ = [
+    'build_model_from_cfg', 'generate_inputs_and_wrap_model',
+    'preprocess_example_input'
+]
diff --git a/detection/scrfd/mmdet/core/export/pytorch2onnx.py b/detection/scrfd/mmdet/core/export/pytorch2onnx.py
new file mode 100755
index 000000000..8f9309df8
--- /dev/null
+++ b/detection/scrfd/mmdet/core/export/pytorch2onnx.py
@@ -0,0 +1,143 @@
+from functools import partial
+
+import mmcv
+import numpy as np
+import torch
+from mmcv.runner import load_checkpoint
+
+
+def generate_inputs_and_wrap_model(config_path, checkpoint_path, input_config):
+    """Prepare sample input and wrap model for ONNX export.
+
+    The ONNX export API only accept args, and all inputs should be
+    torch.Tensor or corresponding types (such as tuple of tensor).
+    So we should call this function before exporting. This function will:
+
+    1. generate corresponding inputs which are used to execute the model.
+    2. Wrap the model's forward function.
+
+    For example, the MMDet models' forward function has a parameter
+    ``return_loss:bool``. As we want to set it as False while export API
+    supports neither bool type or kwargs. So we have to replace the forward
+    like: ``model.forward = partial(model.forward, return_loss=False)``
+
+    Args:
+        config_path (str): the OpenMMLab config for the model we want to
+            export to ONNX
+        checkpoint_path (str): Path to the corresponding checkpoint
+        input_config (dict): the exactly data in this dict depends on the
+            framework. For MMSeg, we can just declare the input shape,
+            and generate the dummy data accordingly. However, for MMDet,
+            we may pass the real img path, or the NMS will return None
+            as there is no legal bbox.
+
+    Returns:
+        tuple: (model, tensor_data) wrapped model which can be called by \
+        model(*tensor_data) and a list of inputs which are used to execute \
+            the model while exporting.
+    """
+
+    model = build_model_from_cfg(config_path, checkpoint_path)
+    one_img, one_meta = preprocess_example_input(input_config)
+    tensor_data = [one_img]
+    model.forward = partial(
+        model.forward, img_metas=[[one_meta]], return_loss=False)
+
+    # pytorch has some bug in pytorch1.3, we have to fix it
+    # by replacing these existing op
+    opset_version = 11
+    # put the import within the function thus it will not cause import error
+    # when not using this function
+    try:
+        from mmcv.onnx.symbolic import register_extra_symbolics
+    except ModuleNotFoundError:
+        raise NotImplementedError('please update mmcv to version>=v1.0.4')
+    register_extra_symbolics(opset_version)
+
+    return model, tensor_data
+
+
+def build_model_from_cfg(config_path, checkpoint_path):
+    """Build a model from config and load the given checkpoint.
+
+    Args:
+        config_path (str): the OpenMMLab config for the model we want to
+            export to ONNX
+        checkpoint_path (str): Path to the corresponding checkpoint
+
+    Returns:
+        torch.nn.Module: the built model
+    """
+    from mmdet.models import build_detector
+
+    cfg = mmcv.Config.fromfile(config_path)
+    # import modules from string list.
+    if cfg.get('custom_imports', None):
+        from mmcv.utils import import_modules_from_strings
+        import_modules_from_strings(**cfg['custom_imports'])
+    cfg.model.pretrained = None
+    cfg.data.test.test_mode = True
+
+    # build the model
+    model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
+    load_checkpoint(model, checkpoint_path, map_location='cpu')
+    model.cpu().eval()
+    return model
+
+
+def preprocess_example_input(input_config):
+    """Prepare an example input image for ``generate_inputs_and_wrap_model``.
+
+    Args:
+        input_config (dict): customized config describing the example input.
+
+    Returns:
+        tuple: (one_img, one_meta), tensor of the example input image and \
+            meta information for the example input image.
+
+    Examples:
+        >>> from mmdet.core.export import preprocess_example_input
+        >>> input_config = {
+        >>>         'input_shape': (1,3,224,224),
+        >>>         'input_path': 'demo/demo.jpg',
+        >>>         'normalize_cfg': {
+        >>>             'mean': (123.675, 116.28, 103.53),
+        >>>             'std': (58.395, 57.12, 57.375)
+        >>>             }
+        >>>         }
+        >>> one_img, one_meta = preprocess_example_input(input_config)
+        >>> print(one_img.shape)
+        torch.Size([1, 3, 224, 224])
+        >>> print(one_meta)
+        {'img_shape': (224, 224, 3),
+        'ori_shape': (224, 224, 3),
+        'pad_shape': (224, 224, 3),
+        'filename': '<demo>.png',
+        'scale_factor': 1.0,
+        'flip': False}
+    """
+    input_path = input_config['input_path']
+    input_shape = input_config['input_shape']
+    one_img = mmcv.imread(input_path)
+    one_img = mmcv.imresize(one_img, input_shape[2:][::-1])
+    show_img = one_img.copy()
+    if 'normalize_cfg' in input_config.keys():
+        normalize_cfg = input_config['normalize_cfg']
+        mean = np.array(normalize_cfg['mean'], dtype=np.float32)
+        std = np.array(normalize_cfg['std'], dtype=np.float32)
+        one_img = mmcv.imnormalize(one_img, mean, std)
+    one_img = one_img.transpose(2, 0, 1)
+    one_img = torch.from_numpy(one_img).unsqueeze(0).float().requires_grad_(
+        True)
+    (_, C, H, W) = input_shape
+    one_meta = {
+        'img_shape': (H, W, C),
+        'ori_shape': (H, W, C),
+        'pad_shape': (H, W, C),
+        'filename': '<demo>.png',
+        'scale_factor': 1.0,
+        'flip': False,
+        'show_img': show_img,
+    }
+
+    return one_img, one_meta
diff --git a/detection/scrfd/mmdet/core/fp16/__init__.py b/detection/scrfd/mmdet/core/fp16/__init__.py
new file mode 100755
index 000000000..0a68c28b9
--- /dev/null
+++ b/detection/scrfd/mmdet/core/fp16/__init__.py
@@ -0,0 +1,8 @@
+from .deprecated_fp16_utils import \
+    DeprecatedFp16OptimizerHook as Fp16OptimizerHook
+from .deprecated_fp16_utils import deprecated_auto_fp16 as auto_fp16
+from .deprecated_fp16_utils import deprecated_force_fp32 as force_fp32
+from .deprecated_fp16_utils import \
+    deprecated_wrap_fp16_model as wrap_fp16_model
+
+__all__ = ['auto_fp16', 'force_fp32', 'Fp16OptimizerHook', 'wrap_fp16_model']
diff --git a/detection/scrfd/mmdet/core/fp16/deprecated_fp16_utils.py b/detection/scrfd/mmdet/core/fp16/deprecated_fp16_utils.py
new file mode 100755
index 000000000..1b15b047e
--- /dev/null
+++ b/detection/scrfd/mmdet/core/fp16/deprecated_fp16_utils.py
@@ -0,0 +1,47 @@
+import warnings
+
+from mmcv.runner import (Fp16OptimizerHook, auto_fp16, force_fp32,
+                         wrap_fp16_model)
+
+
+class DeprecatedFp16OptimizerHook(Fp16OptimizerHook):
+    """A wrapper class for the FP16 optimizer hook. This class wraps
+    :class:`Fp16OptimizerHook` in `mmcv.runner` and shows a warning that the
+    :class:`Fp16OptimizerHook` from `mmdet.core` will be deprecated.
+
+    Refer to :class:`Fp16OptimizerHook` in `mmcv.runner` for more details.
+
+    Args:
+        loss_scale (float): Scale factor multiplied with loss.
+    """
+
+    def __init__(*args, **kwargs):
+        super().__init__(*args, **kwargs)
+        warnings.warn(
+            'Importing Fp16OptimizerHook from "mmdet.core" will be '
+            'deprecated in the future. Please import them from "mmcv.runner" '
+            'instead')
+
+
+def deprecated_auto_fp16(*args, **kwargs):
+    warnings.warn(
+        'Importing auto_fp16 from "mmdet.core" will be '
+        'deprecated in the future. Please import them from "mmcv.runner" '
+        'instead')
+    return auto_fp16(*args, **kwargs)
+
+
+def deprecated_force_fp32(*args, **kwargs):
+    warnings.warn(
+        'Importing force_fp32 from "mmdet.core" will be '
+        'deprecated in the future. Please import them from "mmcv.runner" '
+        'instead')
+    return force_fp32(*args, **kwargs)
+
+
+def deprecated_wrap_fp16_model(*args, **kwargs):
+    warnings.warn(
+        'Importing wrap_fp16_model from "mmdet.core" will be '
+        'deprecated in the future. Please import them from "mmcv.runner" '
+        'instead')
+    wrap_fp16_model(*args, **kwargs)
diff --git a/detection/scrfd/mmdet/core/mask/__init__.py b/detection/scrfd/mmdet/core/mask/__init__.py
new file mode 100755
index 000000000..ab1e88bc6
--- /dev/null
+++ b/detection/scrfd/mmdet/core/mask/__init__.py
@@ -0,0 +1,8 @@
+from .mask_target import mask_target
+from .structures import BaseInstanceMasks, BitmapMasks, PolygonMasks
+from .utils import encode_mask_results, split_combined_polys
+
+__all__ = [
+    'split_combined_polys', 'mask_target', 'BaseInstanceMasks', 'BitmapMasks',
+    'PolygonMasks', 'encode_mask_results'
+]
diff --git a/detection/scrfd/mmdet/core/mask/mask_target.py b/detection/scrfd/mmdet/core/mask/mask_target.py
new file mode 100755
index 000000000..18e423507
--- /dev/null
+++ b/detection/scrfd/mmdet/core/mask/mask_target.py
@@ -0,0 +1,62 @@
+import numpy as np
+import torch
+from torch.nn.modules.utils import _pair
+
+
+def mask_target(pos_proposals_list, pos_assigned_gt_inds_list, gt_masks_list,
+                cfg):
+    """Compute mask target for positive proposals in multiple images.
+
+    Args:
+        pos_proposals_list (list[Tensor]): Positive proposals in multiple
+            images.
+        pos_assigned_gt_inds_list (list[Tensor]): Assigned GT indices for each
+            positive proposals.
+        gt_masks_list (list[:obj:`BaseInstanceMasks`]): Ground truth masks of
+            each image.
+        cfg (dict): Config dict that specifies the mask size.
+
+    Returns:
+        list[Tensor]: Mask target of each image.
+    """
+    cfg_list = [cfg for _ in range(len(pos_proposals_list))]
+    mask_targets = map(mask_target_single, pos_proposals_list,
+                       pos_assigned_gt_inds_list, gt_masks_list, cfg_list)
+    mask_targets = list(mask_targets)
+    if len(mask_targets) > 0:
+        mask_targets = torch.cat(mask_targets)
+    return mask_targets
+
+
+def mask_target_single(pos_proposals, pos_assigned_gt_inds, gt_masks, cfg):
+    """Compute mask target for each positive proposal in the image.
+
+    Args:
+        pos_proposals (Tensor): Positive proposals.
+        pos_assigned_gt_inds (Tensor): Assigned GT inds of positive proposals.
+        gt_masks (:obj:`BaseInstanceMasks`): GT masks in the format of Bitmap
+            or Polygon.
+        cfg (dict): Config dict that indicate the mask size.
+
+    Returns:
+        Tensor: Mask target of each positive proposals in the image.
+    """
+    device = pos_proposals.device
+    mask_size = _pair(cfg.mask_size)
+    num_pos = pos_proposals.size(0)
+    if num_pos > 0:
+        proposals_np = pos_proposals.cpu().numpy()
+        maxh, maxw = gt_masks.height, gt_masks.width
+        proposals_np[:, [0, 2]] = np.clip(proposals_np[:, [0, 2]], 0, maxw)
+        proposals_np[:, [1, 3]] = np.clip(proposals_np[:, [1, 3]], 0, maxh)
+        pos_assigned_gt_inds = pos_assigned_gt_inds.cpu().numpy()
+
+        mask_targets = gt_masks.crop_and_resize(
+            proposals_np, mask_size, device=device,
+            inds=pos_assigned_gt_inds).to_ndarray()
+
+        mask_targets = torch.from_numpy(mask_targets).float().to(device)
+    else:
+        mask_targets = pos_proposals.new_zeros((0, ) + mask_size)
+
+    return mask_targets
diff --git a/detection/scrfd/mmdet/core/mask/structures.py b/detection/scrfd/mmdet/core/mask/structures.py
new file mode 100755
index 000000000..6f5da5547
--- /dev/null
+++ b/detection/scrfd/mmdet/core/mask/structures.py
@@ -0,0 +1,827 @@
+from abc import ABCMeta, abstractmethod
+
+import cv2
+import mmcv
+import numpy as np
+import pycocotools.mask as maskUtils
+import torch
+from mmcv.ops.roi_align import roi_align
+
+
+class BaseInstanceMasks(metaclass=ABCMeta):
+    """Base class for instance masks."""
+
+    @abstractmethod
+    def rescale(self, scale, interpolation='nearest'):
+        """Rescale masks as large as possible while keeping the aspect ratio.
+        For details can refer to `mmcv.imrescale`.
+
+        Args:
+            scale (tuple[int]): The maximum size (h, w) of rescaled mask.
+            interpolation (str): Same as :func:`mmcv.imrescale`.
+
+        Returns:
+            BaseInstanceMasks: The rescaled masks.
+        """
+        pass
+
+    @abstractmethod
+    def resize(self, out_shape, interpolation='nearest'):
+        """Resize masks to the given out_shape.
+
+        Args:
+            out_shape: Target (h, w) of resized mask.
+            interpolation (str): See :func:`mmcv.imresize`.
+
+        Returns:
+            BaseInstanceMasks: The resized masks.
+        """
+        pass
+
+    @abstractmethod
+    def flip(self, flip_direction='horizontal'):
+        """Flip masks alone the given direction.
+
+        Args:
+            flip_direction (str): Either 'horizontal' or 'vertical'.
+
+        Returns:
+            BaseInstanceMasks: The flipped masks.
+        """
+        pass
+
+    @abstractmethod
+    def pad(self, out_shape, pad_val):
+        """Pad masks to the given size of (h, w).
+
+        Args:
+            out_shape (tuple[int]): Target (h, w) of padded mask.
+            pad_val (int): The padded value.
+
+        Returns:
+            BaseInstanceMasks: The padded masks.
+        """
+        pass
+
+    @abstractmethod
+    def crop(self, bbox):
+        """Crop each mask by the given bbox.
+
+        Args:
+            bbox (ndarray): Bbox in format [x1, y1, x2, y2], shape (4, ).
+
+        Return:
+            BaseInstanceMasks: The cropped masks.
+        """
+        pass
+
+    @abstractmethod
+    def crop_and_resize(self,
+                        bboxes,
+                        out_shape,
+                        inds,
+                        device,
+                        interpolation='bilinear'):
+        """Crop and resize masks by the given bboxes.
+
+        This function is mainly used in mask targets computation.
+        It firstly align mask to bboxes by assigned_inds, then crop mask by the
+        assigned bbox and resize to the size of (mask_h, mask_w)
+
+        Args:
+            bboxes (Tensor): Bboxes in format [x1, y1, x2, y2], shape (N, 4)
+            out_shape (tuple[int]): Target (h, w) of resized mask
+            inds (ndarray): Indexes to assign masks to each bbox
+            device (str): Device of bboxes
+            interpolation (str): See `mmcv.imresize`
+
+        Return:
+            BaseInstanceMasks: the cropped and resized masks.
+        """
+        pass
+
+    @abstractmethod
+    def expand(self, expanded_h, expanded_w, top, left):
+        """see :class:`Expand`."""
+        pass
+
+    @property
+    @abstractmethod
+    def areas(self):
+        """ndarray: areas of each instance."""
+        pass
+
+    @abstractmethod
+    def to_ndarray(self):
+        """Convert masks to the format of ndarray.
+
+        Return:
+            ndarray: Converted masks in the format of ndarray.
+        """
+        pass
+
+    @abstractmethod
+    def to_tensor(self, dtype, device):
+        """Convert masks to the format of Tensor.
+
+        Args:
+            dtype (str): Dtype of converted mask.
+            device (torch.device): Device of converted masks.
+
+        Returns:
+            Tensor: Converted masks in the format of Tensor.
+        """
+        pass
+
+    @abstractmethod
+    def translate(self,
+                  out_shape,
+                  offset,
+                  direction='horizontal',
+                  fill_val=0,
+                  interpolation='bilinear'):
+        """Translate the masks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            offset (int | float): The offset for translate.
+            direction (str): The translate direction, either "horizontal"
+                or "vertical".
+            fill_val (int | float): Border value. Default 0.
+            interpolation (str): Same as :func:`mmcv.imtranslate`.
+
+        Returns:
+            Translated masks.
+        """
+        pass
+
+    def shear(self,
+              out_shape,
+              magnitude,
+              direction='horizontal',
+              border_value=0,
+              interpolation='bilinear'):
+        """Shear the masks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            magnitude (int | float): The magnitude used for shear.
+            direction (str): The shear direction, either "horizontal"
+                or "vertical".
+            border_value (int | tuple[int]): Value used in case of a
+                constant border. Default 0.
+            interpolation (str): Same as in :func:`mmcv.imshear`.
+
+        Returns:
+            ndarray: Sheared masks.
+        """
+        pass
+
+    @abstractmethod
+    def rotate(self, out_shape, angle, center=None, scale=1.0, fill_val=0):
+        """Rotate the masks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            angle (int | float): Rotation angle in degrees. Positive values
+                mean counter-clockwise rotation.
+            center (tuple[float], optional): Center point (w, h) of the
+                rotation in source image. If not specified, the center of
+                the image will be used.
+            scale (int | float): Isotropic scale factor.
+            fill_val (int | float): Border value. Default 0 for masks.
+
+        Returns:
+            Rotated masks.
+        """
+        pass
+
+
+class BitmapMasks(BaseInstanceMasks):
+    """This class represents masks in the form of bitmaps.
+
+    Args:
+        masks (ndarray): ndarray of masks in shape (N, H, W), where N is
+            the number of objects.
+        height (int): height of masks
+        width (int): width of masks
+    """
+
+    def __init__(self, masks, height, width):
+        self.height = height
+        self.width = width
+        if len(masks) == 0:
+            self.masks = np.empty((0, self.height, self.width), dtype=np.uint8)
+        else:
+            assert isinstance(masks, (list, np.ndarray))
+            if isinstance(masks, list):
+                assert isinstance(masks[0], np.ndarray)
+                assert masks[0].ndim == 2  # (H, W)
+            else:
+                assert masks.ndim == 3  # (N, H, W)
+
+            self.masks = np.stack(masks).reshape(-1, height, width)
+            assert self.masks.shape[1] == self.height
+            assert self.masks.shape[2] == self.width
+
+    def __getitem__(self, index):
+        """Index the BitmapMask.
+
+        Args:
+            index (int | ndarray): Indices in the format of integer or ndarray.
+
+        Returns:
+            :obj:`BitmapMasks`: Indexed bitmap masks.
+        """
+        masks = self.masks[index].reshape(-1, self.height, self.width)
+        return BitmapMasks(masks, self.height, self.width)
+
+    def __iter__(self):
+        return iter(self.masks)
+
+    def __repr__(self):
+        s = self.__class__.__name__ + '('
+        s += f'num_masks={len(self.masks)}, '
+        s += f'height={self.height}, '
+        s += f'width={self.width})'
+        return s
+
+    def __len__(self):
+        """Number of masks."""
+        return len(self.masks)
+
+    def rescale(self, scale, interpolation='nearest'):
+        """See :func:`BaseInstanceMasks.rescale`."""
+        if len(self.masks) == 0:
+            new_w, new_h = mmcv.rescale_size((self.width, self.height), scale)
+            rescaled_masks = np.empty((0, new_h, new_w), dtype=np.uint8)
+        else:
+            rescaled_masks = np.stack([
+                mmcv.imrescale(mask, scale, interpolation=interpolation)
+                for mask in self.masks
+            ])
+        height, width = rescaled_masks.shape[1:]
+        return BitmapMasks(rescaled_masks, height, width)
+
+    def resize(self, out_shape, interpolation='nearest'):
+        """See :func:`BaseInstanceMasks.resize`."""
+        if len(self.masks) == 0:
+            resized_masks = np.empty((0, *out_shape), dtype=np.uint8)
+        else:
+            resized_masks = np.stack([
+                mmcv.imresize(mask, out_shape, interpolation=interpolation)
+                for mask in self.masks
+            ])
+        return BitmapMasks(resized_masks, *out_shape)
+
+    def flip(self, flip_direction='horizontal'):
+        """See :func:`BaseInstanceMasks.flip`."""
+        assert flip_direction in ('horizontal', 'vertical', 'diagonal')
+
+        if len(self.masks) == 0:
+            flipped_masks = self.masks
+        else:
+            flipped_masks = np.stack([
+                mmcv.imflip(mask, direction=flip_direction)
+                for mask in self.masks
+            ])
+        return BitmapMasks(flipped_masks, self.height, self.width)
+
+    def pad(self, out_shape, pad_val=0):
+        """See :func:`BaseInstanceMasks.pad`."""
+        if len(self.masks) == 0:
+            padded_masks = np.empty((0, *out_shape), dtype=np.uint8)
+        else:
+            padded_masks = np.stack([
+                mmcv.impad(mask, shape=out_shape, pad_val=pad_val)
+                for mask in self.masks
+            ])
+        return BitmapMasks(padded_masks, *out_shape)
+
+    def crop(self, bbox):
+        """See :func:`BaseInstanceMasks.crop`."""
+        assert isinstance(bbox, np.ndarray)
+        assert bbox.ndim == 1
+
+        # clip the boundary
+        bbox = bbox.copy()
+        bbox[0::2] = np.clip(bbox[0::2], 0, self.width)
+        bbox[1::2] = np.clip(bbox[1::2], 0, self.height)
+        x1, y1, x2, y2 = bbox
+        w = np.maximum(x2 - x1, 1)
+        h = np.maximum(y2 - y1, 1)
+
+        if len(self.masks) == 0:
+            cropped_masks = np.empty((0, h, w), dtype=np.uint8)
+        else:
+            cropped_masks = self.masks[:, y1:y1 + h, x1:x1 + w]
+        return BitmapMasks(cropped_masks, h, w)
+
+    def crop_and_resize(self,
+                        bboxes,
+                        out_shape,
+                        inds,
+                        device='cpu',
+                        interpolation='bilinear'):
+        """See :func:`BaseInstanceMasks.crop_and_resize`."""
+        if len(self.masks) == 0:
+            empty_masks = np.empty((0, *out_shape), dtype=np.uint8)
+            return BitmapMasks(empty_masks, *out_shape)
+
+        # convert bboxes to tensor
+        if isinstance(bboxes, np.ndarray):
+            bboxes = torch.from_numpy(bboxes).to(device=device)
+        if isinstance(inds, np.ndarray):
+            inds = torch.from_numpy(inds).to(device=device)
+
+        num_bbox = bboxes.shape[0]
+        fake_inds = torch.arange(
+            num_bbox, device=device).to(dtype=bboxes.dtype)[:, None]
+        rois = torch.cat([fake_inds, bboxes], dim=1)  # Nx5
+        rois = rois.to(device=device)
+        if num_bbox > 0:
+            gt_masks_th = torch.from_numpy(self.masks).to(device).index_select(
+                0, inds).to(dtype=rois.dtype)
+            targets = roi_align(gt_masks_th[:, None, :, :], rois, out_shape,
+                                1.0, 0, 'avg', True).squeeze(1)
+            resized_masks = (targets >= 0.5).cpu().numpy()
+        else:
+            resized_masks = []
+        return BitmapMasks(resized_masks, *out_shape)
+
+    def expand(self, expanded_h, expanded_w, top, left):
+        """See :func:`BaseInstanceMasks.expand`."""
+        if len(self.masks) == 0:
+            expanded_mask = np.empty((0, expanded_h, expanded_w),
+                                     dtype=np.uint8)
+        else:
+            expanded_mask = np.zeros((len(self), expanded_h, expanded_w),
+                                     dtype=np.uint8)
+            expanded_mask[:, top:top + self.height,
+                          left:left + self.width] = self.masks
+        return BitmapMasks(expanded_mask, expanded_h, expanded_w)
+
+    def translate(self,
+                  out_shape,
+                  offset,
+                  direction='horizontal',
+                  fill_val=0,
+                  interpolation='bilinear'):
+        """Translate the BitmapMasks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            offset (int | float): The offset for translate.
+            direction (str): The translate direction, either "horizontal"
+                or "vertical".
+            fill_val (int | float): Border value. Default 0 for masks.
+            interpolation (str): Same as :func:`mmcv.imtranslate`.
+
+        Returns:
+            BitmapMasks: Translated BitmapMasks.
+        """
+        if len(self.masks) == 0:
+            translated_masks = np.empty((0, *out_shape), dtype=np.uint8)
+        else:
+            translated_masks = mmcv.imtranslate(
+                self.masks.transpose((1, 2, 0)),
+                offset,
+                direction,
+                border_value=fill_val,
+                interpolation=interpolation)
+            if translated_masks.ndim == 2:
+                translated_masks = translated_masks[:, :, None]
+            translated_masks = translated_masks.transpose(
+                (2, 0, 1)).astype(self.masks.dtype)
+        return BitmapMasks(translated_masks, *out_shape)
+
+    def shear(self,
+              out_shape,
+              magnitude,
+              direction='horizontal',
+              border_value=0,
+              interpolation='bilinear'):
+        """Shear the BitmapMasks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            magnitude (int | float): The magnitude used for shear.
+            direction (str): The shear direction, either "horizontal"
+                or "vertical".
+            border_value (int | tuple[int]): Value used in case of a
+                constant border.
+            interpolation (str): Same as in :func:`mmcv.imshear`.
+
+        Returns:
+            BitmapMasks: The sheared masks.
+        """
+        if len(self.masks) == 0:
+            sheared_masks = np.empty((0, *out_shape), dtype=np.uint8)
+        else:
+            sheared_masks = mmcv.imshear(
+                self.masks.transpose((1, 2, 0)),
+                magnitude,
+                direction,
+                border_value=border_value,
+                interpolation=interpolation)
+            if sheared_masks.ndim == 2:
+                sheared_masks = sheared_masks[:, :, None]
+            sheared_masks = sheared_masks.transpose(
+                (2, 0, 1)).astype(self.masks.dtype)
+        return BitmapMasks(sheared_masks, *out_shape)
+
+    def rotate(self, out_shape, angle, center=None, scale=1.0, fill_val=0):
+        """Rotate the BitmapMasks.
+
+        Args:
+            out_shape (tuple[int]): Shape for output mask, format (h, w).
+            angle (int | float): Rotation angle in degrees. Positive values
+                mean counter-clockwise rotation.
+            center (tuple[float], optional): Center point (w, h) of the
+                rotation in source image. If not specified, the center of
+                the image will be used.
+            scale (int | float): Isotropic scale factor.
+            fill_val (int | float): Border value. Default 0 for masks.
+
+        Returns:
+            BitmapMasks: Rotated BitmapMasks.
+        """
+        if len(self.masks) == 0:
+            rotated_masks = np.empty((0, *out_shape), dtype=self.masks.dtype)
+        else:
+            rotated_masks = mmcv.imrotate(
+                self.masks.transpose((1, 2, 0)),
+                angle,
+                center=center,
+                scale=scale,
+                border_value=fill_val)
+            if rotated_masks.ndim == 2:
+                # case when only one mask, (h, w)
+                rotated_masks = rotated_masks[:, :, None]  # (h, w, 1)
+            rotated_masks = rotated_masks.transpose(
+                (2, 0, 1)).astype(self.masks.dtype)
+        return BitmapMasks(rotated_masks, *out_shape)
+
+    @property
+    def areas(self):
+        """See :py:attr:`BaseInstanceMasks.areas`."""
+        return self.masks.sum((1, 2))
+
+    def to_ndarray(self):
+        """See :func:`BaseInstanceMasks.to_ndarray`."""
+        return self.masks
+
+    def to_tensor(self, dtype, device):
+        """See :func:`BaseInstanceMasks.to_tensor`."""
+        return torch.tensor(self.masks, dtype=dtype, device=device)
+
+
+class PolygonMasks(BaseInstanceMasks):
+    """This class represents masks in the form of polygons.
+
+    Polygons is a list of three levels. The first level of the list
+    corresponds to objects, the second level to the polys that compose the
+    object, the third level to the poly coordinates
+
+    Args:
+        masks (list[list[ndarray]]): The first level of the list
+            corresponds to objects, the second level to the polys that
+            compose the object, the third level to the poly coordinates
+        height (int): height of masks
+        width (int): width of masks
+    """
+
+    def __init__(self, masks, height, width):
+        assert isinstance(masks, list)
+        if len(masks) > 0:
+            assert isinstance(masks[0], list)
+            assert isinstance(masks[0][0], np.ndarray)
+
+        self.height = height
+        self.width = width
+        self.masks = masks
+
+    def __getitem__(self, index):
+        """Index the polygon masks.
+
+        Args:
+            index (ndarray | List): The indices.
+
+        Returns:
+            :obj:`PolygonMasks`: The indexed polygon masks.
+        """
+        if isinstance(index, np.ndarray):
+            index = index.tolist()
+        if isinstance(index, list):
+            masks = [self.masks[i] for i in index]
+        else:
+            try:
+                masks = self.masks[index]
+            except Exception:
+                raise ValueError(
+                    f'Unsupported input of type {type(index)} for indexing!')
+        if len(masks) and isinstance(masks[0], np.ndarray):
+            masks = [masks]  # ensure a list of three levels
+        return PolygonMasks(masks, self.height, self.width)
+
+    def __iter__(self):
+        return iter(self.masks)
+
+    def __repr__(self):
+        s = self.__class__.__name__ + '('
+        s += f'num_masks={len(self.masks)}, '
+        s += f'height={self.height}, '
+        s += f'width={self.width})'
+        return s
+
+    def __len__(self):
+        """Number of masks."""
+        return len(self.masks)
+
+    def rescale(self, scale, interpolation=None):
+        """see :func:`BaseInstanceMasks.rescale`"""
+        new_w, new_h = mmcv.rescale_size((self.width, self.height), scale)
+        if len(self.masks) == 0:
+            rescaled_masks = PolygonMasks([], new_h, new_w)
+        else:
+            rescaled_masks = self.resize((new_h, new_w))
+        return rescaled_masks
+
+    def resize(self, out_shape, interpolation=None):
+        """see :func:`BaseInstanceMasks.resize`"""
+        if len(self.masks) == 0:
+            resized_masks = PolygonMasks([], *out_shape)
+        else:
+            h_scale = out_shape[0] / self.height
+            w_scale = out_shape[1] / self.width
+            resized_masks = []
+            for poly_per_obj in self.masks:
+                resized_poly = []
+                for p in poly_per_obj:
+                    p = p.copy()
+                    p[0::2] *= w_scale
+                    p[1::2] *= h_scale
+                    resized_poly.append(p)
+                resized_masks.append(resized_poly)
+            resized_masks = PolygonMasks(resized_masks, *out_shape)
+        return resized_masks
+
+    def flip(self, flip_direction='horizontal'):
+        """see :func:`BaseInstanceMasks.flip`"""
+        assert flip_direction in ('horizontal', 'vertical', 'diagonal')
+        if len(self.masks) == 0:
+            flipped_masks = PolygonMasks([], self.height, self.width)
+        else:
+            flipped_masks = []
+            for poly_per_obj in self.masks:
+                flipped_poly_per_obj = []
+                for p in poly_per_obj:
+                    p = p.copy()
+                    if flip_direction == 'horizontal':
+                        p[0::2] = self.width - p[0::2]
+                    elif flip_direction == 'vertical':
+                        p[1::2] = self.height - p[1::2]
+                    else:
+                        p[0::2] = self.width - p[0::2]
+                        p[1::2] = self.height - p[1::2]
+                    flipped_poly_per_obj.append(p)
+                flipped_masks.append(flipped_poly_per_obj)
+            flipped_masks = PolygonMasks(flipped_masks, self.height,
+                                         self.width)
+        return flipped_masks
+
+    def crop(self, bbox):
+        """see :func:`BaseInstanceMasks.crop`"""
+        assert isinstance(bbox, np.ndarray)
+        assert bbox.ndim == 1
+
+        # clip the boundary
+        bbox = bbox.copy()
+        bbox[0::2] = np.clip(bbox[0::2], 0, self.width)
+        bbox[1::2] = np.clip(bbox[1::2], 0, self.height)
+        x1, y1, x2, y2 = bbox
+        w = np.maximum(x2 - x1, 1)
+        h = np.maximum(y2 - y1, 1)
+
+        if len(self.masks) == 0:
+            cropped_masks = PolygonMasks([], h, w)
+        else:
+            cropped_masks = []
+            for poly_per_obj in self.masks:
+                cropped_poly_per_obj = []
+                for p in poly_per_obj:
+                    # pycocotools will clip the boundary
+                    p = p.copy()
+                    p[0::2] -= bbox[0]
+                    p[1::2] -= bbox[1]
+                    cropped_poly_per_obj.append(p)
+                cropped_masks.append(cropped_poly_per_obj)
+            cropped_masks = PolygonMasks(cropped_masks, h, w)
+        return cropped_masks
+
+    def pad(self, out_shape, pad_val=0):
+        """padding has no effect on polygons`"""
+        return PolygonMasks(self.masks, *out_shape)
+
+    def expand(self, *args, **kwargs):
+        """TODO: Add expand for polygon"""
+        raise NotImplementedError
+
+    def crop_and_resize(self,
+                        bboxes,
+                        out_shape,
+                        inds,
+                        device='cpu',
+                        interpolation='bilinear'):
+        """see :func:`BaseInstanceMasks.crop_and_resize`"""
+        out_h, out_w = out_shape
+        if len(self.masks) == 0:
+            return PolygonMasks([], out_h, out_w)
+
+        resized_masks = []
+        for i in range(len(bboxes)):
+            mask = self.masks[inds[i]]
+            bbox = bboxes[i, :]
+            x1, y1, x2, y2 = bbox
+            w = np.maximum(x2 - x1, 1)
+            h = np.maximum(y2 - y1, 1)
+            h_scale = out_h / max(h, 0.1)  # avoid too large scale
+            w_scale = out_w / max(w, 0.1)
+
+            resized_mask = []
+            for p in mask:
+                p = p.copy()
+                # crop
+                # pycocotools will clip the boundary
+                p[0::2] -= bbox[0]
+                p[1::2] -= bbox[1]
+
+                # resize
+                p[0::2] *= w_scale
+                p[1::2] *= h_scale
+                resized_mask.append(p)
+            resized_masks.append(resized_mask)
+        return PolygonMasks(resized_masks, *out_shape)
+
+    def translate(self,
+                  out_shape,
+                  offset,
+                  direction='horizontal',
+                  fill_val=None,
+                  interpolation=None):
+        """Translate the PolygonMasks."""
+        assert fill_val is None or fill_val == 0, 'Here fill_val is not '\
+            f'used, and defaultly should be None or 0. got {fill_val}.'
+        if len(self.masks) == 0:
+            translated_masks = PolygonMasks([], *out_shape)
+        else:
+            translated_masks = []
+            for poly_per_obj in self.masks:
+                translated_poly_per_obj = []
+                for p in poly_per_obj:
+                    p = p.copy()
+                    if direction == 'horizontal':
+                        p[0::2] = np.clip(p[0::2] + offset, 0, out_shape[1])
+                    elif direction == 'vertical':
+                        p[1::2] = np.clip(p[1::2] + offset, 0, out_shape[0])
+                    translated_poly_per_obj.append(p)
+                translated_masks.append(translated_poly_per_obj)
+            translated_masks = PolygonMasks(translated_masks, *out_shape)
+        return translated_masks
+
+    def shear(self,
+              out_shape,
+              magnitude,
+              direction='horizontal',
+              border_value=0,
+              interpolation='bilinear'):
+        """See :func:`BaseInstanceMasks.shear`."""
+        if len(self.masks) == 0:
+            sheared_masks = PolygonMasks([], *out_shape)
+        else:
+            sheared_masks = []
+            if direction == 'horizontal':
+                shear_matrix = np.stack([[1, magnitude],
+                                         [0, 1]]).astype(np.float32)
+            elif direction == 'vertical':
+                shear_matrix = np.stack([[1, 0], [magnitude,
+                                                  1]]).astype(np.float32)
+            for poly_per_obj in self.masks:
+                sheared_poly = []
+                for p in poly_per_obj:
+                    p = np.stack([p[0::2], p[1::2]], axis=0)  # [2, n]
+                    new_coords = np.matmul(shear_matrix, p)  # [2, n]
+                    new_coords[0, :] = np.clip(new_coords[0, :], 0,
+                                               out_shape[1])
+                    new_coords[1, :] = np.clip(new_coords[1, :], 0,
+                                               out_shape[0])
+                    sheared_poly.append(
+                        new_coords.transpose((1, 0)).reshape(-1))
+                sheared_masks.append(sheared_poly)
+            sheared_masks = PolygonMasks(sheared_masks, *out_shape)
+        return sheared_masks
+
+    def rotate(self, out_shape, angle, center=None, scale=1.0, fill_val=0):
+        """See :func:`BaseInstanceMasks.rotate`."""
+        if len(self.masks) == 0:
+            rotated_masks = PolygonMasks([], *out_shape)
+        else:
+            rotated_masks = []
+            rotate_matrix = cv2.getRotationMatrix2D(center, -angle, scale)
+            for poly_per_obj in self.masks:
+                rotated_poly = []
+                for p in poly_per_obj:
+                    p = p.copy()
+                    coords = np.stack([p[0::2], p[1::2]], axis=1)  # [n, 2]
+                    # pad 1 to convert from format [x, y] to homogeneous
+                    # coordinates format [x, y, 1]
+                    coords = np.concatenate(
+                        (coords, np.ones((coords.shape[0], 1), coords.dtype)),
+                        axis=1)  # [n, 3]
+                    rotated_coords = np.matmul(
+                        rotate_matrix[None, :, :],
+                        coords[:, :, None])[..., 0]  # [n, 2, 1] -> [n, 2]
+                    rotated_coords[:, 0] = np.clip(rotated_coords[:, 0], 0,
+                                                   out_shape[1])
+                    rotated_coords[:, 1] = np.clip(rotated_coords[:, 1], 0,
+                                                   out_shape[0])
+                    rotated_poly.append(rotated_coords.reshape(-1))
+                rotated_masks.append(rotated_poly)
+            rotated_masks = PolygonMasks(rotated_masks, *out_shape)
+        return rotated_masks
+
+    def to_bitmap(self):
+        """convert polygon masks to bitmap masks."""
+        bitmap_masks = self.to_ndarray()
+        return BitmapMasks(bitmap_masks, self.height, self.width)
+
+    @property
+    def areas(self):
+        """Compute areas of masks.
+
+        This func is modified from `detectron2
+        <https://github.com/facebookresearch/detectron2/blob/ffff8acc35ea88ad1cb1806ab0f00b4c1c5dbfd9/detectron2/structures/masks.py#L387>`_.
+        The function only works with Polygons using the shoelace formula.
+
+        Return:
+            ndarray: areas of each instance
+        """  # noqa: W501
+        area = []
+        for polygons_per_obj in self.masks:
+            area_per_obj = 0
+            for p in polygons_per_obj:
+                area_per_obj += self._polygon_area(p[0::2], p[1::2])
+            area.append(area_per_obj)
+        return np.asarray(area)
+
+    def _polygon_area(self, x, y):
+        """Compute the area of a component of a polygon.
+
+        Using the shoelace formula:
+        https://stackoverflow.com/questions/24467972/calculate-area-of-polygon-given-x-y-coordinates
+
+        Args:
+            x (ndarray): x coordinates of the component
+            y (ndarray): y coordinates of the component
+
+        Return:
+            float: the are of the component
+        """  # noqa: 501
+        return 0.5 * np.abs(
+            np.dot(x, np.roll(y, 1)) - np.dot(y, np.roll(x, 1)))
+
+    def to_ndarray(self):
+        """Convert masks to the format of ndarray."""
+        if len(self.masks) == 0:
+            return np.empty((0, self.height, self.width), dtype=np.uint8)
+        bitmap_masks = []
+        for poly_per_obj in self.masks:
+            bitmap_masks.append(
+                polygon_to_bitmap(poly_per_obj, self.height, self.width))
+        return np.stack(bitmap_masks)
+
+    def to_tensor(self, dtype, device):
+        """See :func:`BaseInstanceMasks.to_tensor`."""
+        if len(self.masks) == 0:
+            return torch.empty((0, self.height, self.width),
+                               dtype=dtype,
+                               device=device)
+        ndarray_masks = self.to_ndarray()
+        return torch.tensor(ndarray_masks, dtype=dtype, device=device)
+
+
+def polygon_to_bitmap(polygons, height, width):
+    """Convert masks from the form of polygons to bitmaps.
+
+    Args:
+        polygons (list[ndarray]): masks in polygon representation
+        height (int): mask height
+        width (int): mask width
+
+    Return:
+        ndarray: the converted masks in bitmap representation
+    """
+    rles = maskUtils.frPyObjects(polygons, height, width)
+    rle = maskUtils.merge(rles)
+    bitmap_mask = maskUtils.decode(rle).astype(np.bool)
+    return bitmap_mask
diff --git a/detection/scrfd/mmdet/core/mask/utils.py b/detection/scrfd/mmdet/core/mask/utils.py
new file mode 100755
index 000000000..c88208291
--- /dev/null
+++ b/detection/scrfd/mmdet/core/mask/utils.py
@@ -0,0 +1,63 @@
+import mmcv
+import numpy as np
+import pycocotools.mask as mask_util
+
+
+def split_combined_polys(polys, poly_lens, polys_per_mask):
+    """Split the combined 1-D polys into masks.
+
+    A mask is represented as a list of polys, and a poly is represented as
+    a 1-D array. In dataset, all masks are concatenated into a single 1-D
+    tensor. Here we need to split the tensor into original representations.
+
+    Args:
+        polys (list): a list (length = image num) of 1-D tensors
+        poly_lens (list): a list (length = image num) of poly length
+        polys_per_mask (list): a list (length = image num) of poly number
+            of each mask
+
+    Returns:
+        list: a list (length = image num) of list (length = mask num) of \
+            list (length = poly num) of numpy array.
+    """
+    mask_polys_list = []
+    for img_id in range(len(polys)):
+        polys_single = polys[img_id]
+        polys_lens_single = poly_lens[img_id].tolist()
+        polys_per_mask_single = polys_per_mask[img_id].tolist()
+
+        split_polys = mmcv.slice_list(polys_single, polys_lens_single)
+        mask_polys = mmcv.slice_list(split_polys, polys_per_mask_single)
+        mask_polys_list.append(mask_polys)
+    return mask_polys_list
+
+
+# TODO: move this function to more proper place
+def encode_mask_results(mask_results):
+    """Encode bitmap mask to RLE code.
+
+    Args:
+        mask_results (list | tuple[list]): bitmap mask results.
+            In mask scoring rcnn, mask_results is a tuple of (segm_results,
+            segm_cls_score).
+
+    Returns:
+        list | tuple: RLE encoded mask.
+    """
+    if isinstance(mask_results, tuple):  # mask scoring
+        cls_segms, cls_mask_scores = mask_results
+    else:
+        cls_segms = mask_results
+    num_classes = len(cls_segms)
+    encoded_mask_results = [[] for _ in range(num_classes)]
+    for i in range(len(cls_segms)):
+        for cls_segm in cls_segms[i]:
+            encoded_mask_results[i].append(
+                mask_util.encode(
+                    np.array(
+                        cls_segm[:, :, np.newaxis], order='F',
+                        dtype='uint8'))[0])  # encoded with RLE
+    if isinstance(mask_results, tuple):
+        return encoded_mask_results, cls_mask_scores
+    else:
+        return encoded_mask_results
diff --git a/detection/scrfd/mmdet/core/post_processing/__init__.py b/detection/scrfd/mmdet/core/post_processing/__init__.py
new file mode 100755
index 000000000..880b3f066
--- /dev/null
+++ b/detection/scrfd/mmdet/core/post_processing/__init__.py
@@ -0,0 +1,8 @@
+from .bbox_nms import fast_nms, multiclass_nms
+from .merge_augs import (merge_aug_bboxes, merge_aug_masks,
+                         merge_aug_proposals, merge_aug_scores)
+
+__all__ = [
+    'multiclass_nms', 'merge_aug_proposals', 'merge_aug_bboxes',
+    'merge_aug_scores', 'merge_aug_masks', 'fast_nms'
+]
diff --git a/detection/scrfd/mmdet/core/post_processing/bbox_nms.py b/detection/scrfd/mmdet/core/post_processing/bbox_nms.py
new file mode 100755
index 000000000..7ab322a2a
--- /dev/null
+++ b/detection/scrfd/mmdet/core/post_processing/bbox_nms.py
@@ -0,0 +1,149 @@
+import torch
+from mmcv.ops.nms import batched_nms
+
+from mmdet.core.bbox.iou_calculators import bbox_overlaps
+
+
+def multiclass_nms(multi_bboxes,
+                   multi_scores,
+                   score_thr,
+                   nms_cfg,
+                   max_num=-1,
+                   score_factors=None,
+                   return_inds=False):
+    """NMS for multi-class bboxes.
+
+    Args:
+        multi_bboxes (Tensor): shape (n, #class*4) or (n, 4)
+        multi_scores (Tensor): shape (n, #class), where the last column
+            contains scores of the background class, but this will be ignored.
+        score_thr (float): bbox threshold, bboxes with scores lower than it
+            will not be considered.
+        nms_thr (float): NMS IoU threshold
+        max_num (int, optional): if there are more than max_num bboxes after
+            NMS, only top max_num will be kept. Default to -1.
+        score_factors (Tensor, optional): The factors multiplied to scores
+            before applying NMS. Default to None.
+        return_inds (bool, optional): Whether return the indices of kept
+            bboxes. Default to False.
+
+    Returns:
+        tuple: (bboxes, labels, indices (optional)), tensors of shape (k, 5),
+            (k), and (k). Labels are 0-based.
+    """
+    num_classes = multi_scores.size(1) - 1
+    #print('!!!!!', multi_bboxes.shape)
+    # exclude background category
+    if multi_bboxes.shape[1] > 4:
+        bboxes = multi_bboxes.view(multi_scores.size(0), -1, 4)
+    else:
+        bboxes = multi_bboxes[:, None].expand(
+            multi_scores.size(0), num_classes, 4)
+
+    scores = multi_scores[:, :-1]
+    if score_factors is not None:
+        scores = scores * score_factors[:, None]
+
+    labels = torch.arange(num_classes, dtype=torch.long)
+    labels = labels.view(1, -1).expand_as(scores)
+
+    bboxes = bboxes.reshape(-1, 4)
+    scores = scores.reshape(-1)
+    labels = labels.reshape(-1)
+
+    # remove low scoring boxes
+    valid_mask = scores > score_thr
+    inds = valid_mask.nonzero(as_tuple=False).squeeze(1)
+    bboxes, scores, labels = bboxes[inds], scores[inds], labels[inds]
+    if inds.numel() == 0:
+        if torch.onnx.is_in_onnx_export():
+            raise RuntimeError('[ONNX Error] Can not record NMS '
+                               'as it has not been executed this time')
+        return bboxes, labels
+
+    # TODO: add size check before feed into batched_nms
+    dets, keep = batched_nms(bboxes, scores, labels, nms_cfg)
+
+    if max_num > 0:
+        dets = dets[:max_num]
+        keep = keep[:max_num]
+
+    if return_inds:
+        return dets, labels[keep], keep
+    else:
+        return dets, labels[keep]
+
+
+def fast_nms(multi_bboxes,
+             multi_scores,
+             multi_coeffs,
+             score_thr,
+             iou_thr,
+             top_k,
+             max_num=-1):
+    """Fast NMS in `YOLACT <https://arxiv.org/abs/1904.02689>`_.
+
+    Fast NMS allows already-removed detections to suppress other detections so
+    that every instance can be decided to be kept or discarded in parallel,
+    which is not possible in traditional NMS. This relaxation allows us to
+    implement Fast NMS entirely in standard GPU-accelerated matrix operations.
+
+    Args:
+        multi_bboxes (Tensor): shape (n, #class*4) or (n, 4)
+        multi_scores (Tensor): shape (n, #class+1), where the last column
+            contains scores of the background class, but this will be ignored.
+        multi_coeffs (Tensor): shape (n, #class*coeffs_dim).
+        score_thr (float): bbox threshold, bboxes with scores lower than it
+            will not be considered.
+        iou_thr (float): IoU threshold to be considered as conflicted.
+        top_k (int): if there are more than top_k bboxes before NMS,
+            only top top_k will be kept.
+        max_num (int): if there are more than max_num bboxes after NMS,
+            only top max_num will be kept. If -1, keep all the bboxes.
+            Default: -1.
+
+    Returns:
+        tuple: (bboxes, labels, coefficients), tensors of shape (k, 5), (k, 1),
+            and (k, coeffs_dim). Labels are 0-based.
+    """
+
+    scores = multi_scores[:, :-1].t()  # [#class, n]
+    scores, idx = scores.sort(1, descending=True)
+
+    idx = idx[:, :top_k].contiguous()
+    scores = scores[:, :top_k]  # [#class, topk]
+    num_classes, num_dets = idx.size()
+    boxes = multi_bboxes[idx.view(-1), :].view(num_classes, num_dets, 4)
+    coeffs = multi_coeffs[idx.view(-1), :].view(num_classes, num_dets, -1)
+
+    iou = bbox_overlaps(boxes, boxes)  # [#class, topk, topk]
+    iou.triu_(diagonal=1)
+    iou_max, _ = iou.max(dim=1)
+
+    # Now just filter out the ones higher than the threshold
+    keep = iou_max <= iou_thr
+
+    # Second thresholding introduces 0.2 mAP gain at negligible time cost
+    keep *= scores > score_thr
+
+    # Assign each kept detection to its corresponding class
+    classes = torch.arange(
+        num_classes, device=boxes.device)[:, None].expand_as(keep)
+    classes = classes[keep]
+
+    boxes = boxes[keep]
+    coeffs = coeffs[keep]
+    scores = scores[keep]
+
+    # Only keep the top max_num highest scores across all classes
+    scores, idx = scores.sort(0, descending=True)
+    if max_num > 0:
+        idx = idx[:max_num]
+        scores = scores[:max_num]
+
+    classes = classes[idx]
+    boxes = boxes[idx]
+    coeffs = coeffs[idx]
+
+    cls_dets = torch.cat([boxes, scores[:, None]], dim=1)
+    return cls_dets, classes, coeffs
diff --git a/detection/scrfd/mmdet/core/post_processing/merge_augs.py b/detection/scrfd/mmdet/core/post_processing/merge_augs.py
new file mode 100755
index 000000000..167093ebf
--- /dev/null
+++ b/detection/scrfd/mmdet/core/post_processing/merge_augs.py
@@ -0,0 +1,117 @@
+import numpy as np
+import torch
+from mmcv.ops import nms
+
+from ..bbox import bbox_mapping_back
+
+
+def merge_aug_proposals(aug_proposals, img_metas, rpn_test_cfg):
+    """Merge augmented proposals (multiscale, flip, etc.)
+
+    Args:
+        aug_proposals (list[Tensor]): proposals from different testing
+            schemes, shape (n, 5). Note that they are not rescaled to the
+            original image size.
+
+        img_metas (list[dict]): list of image info dict where each dict has:
+            'img_shape', 'scale_factor', 'flip', and may also contain
+            'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+            For details on the values of these keys see
+            `mmdet/datasets/pipelines/formatting.py:Collect`.
+
+        rpn_test_cfg (dict): rpn test config.
+
+    Returns:
+        Tensor: shape (n, 4), proposals corresponding to original image scale.
+    """
+    recovered_proposals = []
+    for proposals, img_info in zip(aug_proposals, img_metas):
+        img_shape = img_info['img_shape']
+        scale_factor = img_info['scale_factor']
+        flip = img_info['flip']
+        flip_direction = img_info['flip_direction']
+        _proposals = proposals.clone()
+        _proposals[:, :4] = bbox_mapping_back(_proposals[:, :4], img_shape,
+                                              scale_factor, flip,
+                                              flip_direction)
+        recovered_proposals.append(_proposals)
+    aug_proposals = torch.cat(recovered_proposals, dim=0)
+    merged_proposals, _ = nms(aug_proposals[:, :4].contiguous(),
+                              aug_proposals[:, -1].contiguous(),
+                              rpn_test_cfg.nms_thr)
+    scores = merged_proposals[:, 4]
+    _, order = scores.sort(0, descending=True)
+    num = min(rpn_test_cfg.max_num, merged_proposals.shape[0])
+    order = order[:num]
+    merged_proposals = merged_proposals[order, :]
+    return merged_proposals
+
+
+def merge_aug_bboxes(aug_bboxes, aug_scores, img_metas, rcnn_test_cfg):
+    """Merge augmented detection bboxes and scores.
+
+    Args:
+        aug_bboxes (list[Tensor]): shape (n, 4*#class)
+        aug_scores (list[Tensor] or None): shape (n, #class)
+        img_shapes (list[Tensor]): shape (3, ).
+        rcnn_test_cfg (dict): rcnn test config.
+
+    Returns:
+        tuple: (bboxes, scores)
+    """
+    recovered_bboxes = []
+    for bboxes, img_info in zip(aug_bboxes, img_metas):
+        img_shape = img_info[0]['img_shape']
+        scale_factor = img_info[0]['scale_factor']
+        flip = img_info[0]['flip']
+        flip_direction = img_info[0]['flip_direction']
+        bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip,
+                                   flip_direction)
+        recovered_bboxes.append(bboxes)
+    bboxes = torch.stack(recovered_bboxes).mean(dim=0)
+    if aug_scores is None:
+        return bboxes
+    else:
+        scores = torch.stack(aug_scores).mean(dim=0)
+        return bboxes, scores
+
+
+def merge_aug_scores(aug_scores):
+    """Merge augmented bbox scores."""
+    if isinstance(aug_scores[0], torch.Tensor):
+        return torch.mean(torch.stack(aug_scores), dim=0)
+    else:
+        return np.mean(aug_scores, axis=0)
+
+
+def merge_aug_masks(aug_masks, img_metas, rcnn_test_cfg, weights=None):
+    """Merge augmented mask prediction.
+
+    Args:
+        aug_masks (list[ndarray]): shape (n, #class, h, w)
+        img_shapes (list[ndarray]): shape (3, ).
+        rcnn_test_cfg (dict): rcnn test config.
+
+    Returns:
+        tuple: (bboxes, scores)
+    """
+    recovered_masks = []
+    for mask, img_info in zip(aug_masks, img_metas):
+        flip = img_info[0]['flip']
+        flip_direction = img_info[0]['flip_direction']
+        if flip:
+            if flip_direction == 'horizontal':
+                mask = mask[:, :, :, ::-1]
+            elif flip_direction == 'vertical':
+                mask = mask[:, :, ::-1, :]
+            else:
+                raise ValueError(
+                    f"Invalid flipping direction '{flip_direction}'")
+        recovered_masks.append(mask)
+
+    if weights is None:
+        merged_masks = np.mean(recovered_masks, axis=0)
+    else:
+        merged_masks = np.average(
+            np.array(recovered_masks), axis=0, weights=np.array(weights))
+    return merged_masks
diff --git a/detection/scrfd/mmdet/core/utils/__init__.py b/detection/scrfd/mmdet/core/utils/__init__.py
new file mode 100755
index 000000000..b49b3fc46
--- /dev/null
+++ b/detection/scrfd/mmdet/core/utils/__init__.py
@@ -0,0 +1,7 @@
+from .dist_utils import DistOptimizerHook, allreduce_grads, reduce_mean
+from .misc import multi_apply, unmap
+
+__all__ = [
+    'allreduce_grads', 'DistOptimizerHook', 'reduce_mean', 'multi_apply',
+    'unmap'
+]
diff --git a/detection/scrfd/mmdet/core/utils/dist_utils.py b/detection/scrfd/mmdet/core/utils/dist_utils.py
new file mode 100755
index 000000000..5fe777533
--- /dev/null
+++ b/detection/scrfd/mmdet/core/utils/dist_utils.py
@@ -0,0 +1,69 @@
+import warnings
+from collections import OrderedDict
+
+import torch.distributed as dist
+from mmcv.runner import OptimizerHook
+from torch._utils import (_flatten_dense_tensors, _take_tensors,
+                          _unflatten_dense_tensors)
+
+
+def _allreduce_coalesced(tensors, world_size, bucket_size_mb=-1):
+    if bucket_size_mb > 0:
+        bucket_size_bytes = bucket_size_mb * 1024 * 1024
+        buckets = _take_tensors(tensors, bucket_size_bytes)
+    else:
+        buckets = OrderedDict()
+        for tensor in tensors:
+            tp = tensor.type()
+            if tp not in buckets:
+                buckets[tp] = []
+            buckets[tp].append(tensor)
+        buckets = buckets.values()
+
+    for bucket in buckets:
+        flat_tensors = _flatten_dense_tensors(bucket)
+        dist.all_reduce(flat_tensors)
+        flat_tensors.div_(world_size)
+        for tensor, synced in zip(
+                bucket, _unflatten_dense_tensors(flat_tensors, bucket)):
+            tensor.copy_(synced)
+
+
+def allreduce_grads(params, coalesce=True, bucket_size_mb=-1):
+    """Allreduce gradients.
+
+    Args:
+        params (list[torch.Parameters]): List of parameters of a model
+        coalesce (bool, optional): Whether allreduce parameters as a whole.
+            Defaults to True.
+        bucket_size_mb (int, optional): Size of bucket, the unit is MB.
+            Defaults to -1.
+    """
+    grads = [
+        param.grad.data for param in params
+        if param.requires_grad and param.grad is not None
+    ]
+    world_size = dist.get_world_size()
+    if coalesce:
+        _allreduce_coalesced(grads, world_size, bucket_size_mb)
+    else:
+        for tensor in grads:
+            dist.all_reduce(tensor.div_(world_size))
+
+
+class DistOptimizerHook(OptimizerHook):
+    """Deprecated optimizer hook for distributed training."""
+
+    def __init__(self, *args, **kwargs):
+        warnings.warn('"DistOptimizerHook" is deprecated, please switch to'
+                      '"mmcv.runner.OptimizerHook".')
+        super().__init__(*args, **kwargs)
+
+
+def reduce_mean(tensor):
+    """"Obtain the mean of tensor on different GPUs."""
+    if not (dist.is_available() and dist.is_initialized()):
+        return tensor
+    tensor = tensor.clone()
+    dist.all_reduce(tensor.div_(dist.get_world_size()), op=dist.ReduceOp.SUM)
+    return tensor
diff --git a/detection/scrfd/mmdet/core/utils/misc.py b/detection/scrfd/mmdet/core/utils/misc.py
new file mode 100755
index 000000000..407116e99
--- /dev/null
+++ b/detection/scrfd/mmdet/core/utils/misc.py
@@ -0,0 +1,41 @@
+from functools import partial
+
+import torch
+from six.moves import map, zip
+
+
+def multi_apply(func, *args, **kwargs):
+    """Apply function to a list of arguments.
+
+    Note:
+        This function applies the ``func`` to multiple inputs and
+        map the multiple outputs of the ``func`` into different
+        list. Each list contains the same type of outputs corresponding
+        to different inputs.
+
+    Args:
+        func (Function): A function that will be applied to a list of
+            arguments
+
+    Returns:
+        tuple(list): A tuple containing multiple list, each list contains \
+            a kind of returned results by the function
+    """
+    pfunc = partial(func, **kwargs) if kwargs else func
+    map_results = map(pfunc, *args)
+    return tuple(map(list, zip(*map_results)))
+
+
+def unmap(data, count, inds, fill=0):
+    """Unmap a subset of item (data) back to the original set of items (of size
+    count)"""
+    if data.dim() == 1:
+        ret = data.new_full((count, ), fill)
+        ret[inds.type(torch.bool)] = data
+    else:
+        new_size = (count, ) + data.size()[1:]
+        ret = data.new_full(new_size, fill)
+        #print(inds)
+        #print('CCC', ret.shape, inds.shape, data.shape)
+        ret[inds.type(torch.bool), :] = data
+    return ret
diff --git a/detection/scrfd/mmdet/datasets/__init__.py b/detection/scrfd/mmdet/datasets/__init__.py
new file mode 100755
index 000000000..2302e4e32
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/__init__.py
@@ -0,0 +1,24 @@
+from .builder import DATASETS, PIPELINES, build_dataloader, build_dataset
+from .cityscapes import CityscapesDataset
+from .coco import CocoDataset
+from .custom import CustomDataset
+from .retinaface import RetinaFaceDataset
+from .dataset_wrappers import (ClassBalancedDataset, ConcatDataset,
+                               RepeatDataset)
+from .deepfashion import DeepFashionDataset
+from .lvis import LVISDataset, LVISV1Dataset, LVISV05Dataset
+from .samplers import DistributedGroupSampler, DistributedSampler, GroupSampler
+from .utils import replace_ImageToTensor
+from .voc import VOCDataset
+from .wider_face import WIDERFaceDataset
+from .xml_style import XMLDataset
+
+__all__ = [
+    'CustomDataset', 'XMLDataset', 'CocoDataset', 'DeepFashionDataset',
+    'VOCDataset', 'CityscapesDataset', 'LVISDataset', 'LVISV05Dataset',
+    'LVISV1Dataset', 'GroupSampler', 'DistributedGroupSampler',
+    'DistributedSampler', 'build_dataloader', 'ConcatDataset', 'RepeatDataset',
+    'RetinaFaceDataset',
+    'ClassBalancedDataset', 'WIDERFaceDataset', 'DATASETS', 'PIPELINES',
+    'build_dataset', 'replace_ImageToTensor'
+]
diff --git a/detection/scrfd/mmdet/datasets/builder.py b/detection/scrfd/mmdet/datasets/builder.py
new file mode 100755
index 000000000..16d9ae34b
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/builder.py
@@ -0,0 +1,143 @@
+import copy
+import platform
+import random
+from functools import partial
+
+import numpy as np
+from mmcv.parallel import collate
+from mmcv.runner import get_dist_info
+from mmcv.utils import Registry, build_from_cfg
+from torch.utils.data import DataLoader
+
+from .samplers import DistributedGroupSampler, DistributedSampler, GroupSampler
+
+if platform.system() != 'Windows':
+    # https://github.com/pytorch/pytorch/issues/973
+    import resource
+    rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
+    hard_limit = rlimit[1]
+    soft_limit = min(4096, hard_limit)
+    resource.setrlimit(resource.RLIMIT_NOFILE, (soft_limit, hard_limit))
+
+DATASETS = Registry('dataset')
+PIPELINES = Registry('pipeline')
+
+
+def _concat_dataset(cfg, default_args=None):
+    from .dataset_wrappers import ConcatDataset
+    ann_files = cfg['ann_file']
+    img_prefixes = cfg.get('img_prefix', None)
+    seg_prefixes = cfg.get('seg_prefix', None)
+    proposal_files = cfg.get('proposal_file', None)
+    separate_eval = cfg.get('separate_eval', True)
+
+    datasets = []
+    num_dset = len(ann_files)
+    for i in range(num_dset):
+        data_cfg = copy.deepcopy(cfg)
+        # pop 'separate_eval' since it is not a valid key for common datasets.
+        if 'separate_eval' in data_cfg:
+            data_cfg.pop('separate_eval')
+        data_cfg['ann_file'] = ann_files[i]
+        if isinstance(img_prefixes, (list, tuple)):
+            data_cfg['img_prefix'] = img_prefixes[i]
+        if isinstance(seg_prefixes, (list, tuple)):
+            data_cfg['seg_prefix'] = seg_prefixes[i]
+        if isinstance(proposal_files, (list, tuple)):
+            data_cfg['proposal_file'] = proposal_files[i]
+        datasets.append(build_dataset(data_cfg, default_args))
+
+    return ConcatDataset(datasets, separate_eval)
+
+
+def build_dataset(cfg, default_args=None):
+    from .dataset_wrappers import (ConcatDataset, RepeatDataset,
+                                   ClassBalancedDataset)
+    if isinstance(cfg, (list, tuple)):
+        dataset = ConcatDataset([build_dataset(c, default_args) for c in cfg])
+    elif cfg['type'] == 'ConcatDataset':
+        dataset = ConcatDataset(
+            [build_dataset(c, default_args) for c in cfg['datasets']],
+            cfg.get('separate_eval', True))
+    elif cfg['type'] == 'RepeatDataset':
+        dataset = RepeatDataset(
+            build_dataset(cfg['dataset'], default_args), cfg['times'])
+    elif cfg['type'] == 'ClassBalancedDataset':
+        dataset = ClassBalancedDataset(
+            build_dataset(cfg['dataset'], default_args), cfg['oversample_thr'])
+    elif isinstance(cfg.get('ann_file'), (list, tuple)):
+        dataset = _concat_dataset(cfg, default_args)
+    else:
+        dataset = build_from_cfg(cfg, DATASETS, default_args)
+
+    return dataset
+
+
+def build_dataloader(dataset,
+                     samples_per_gpu,
+                     workers_per_gpu,
+                     num_gpus=1,
+                     dist=True,
+                     shuffle=True,
+                     seed=None,
+                     **kwargs):
+    """Build PyTorch DataLoader.
+
+    In distributed training, each GPU/process has a dataloader.
+    In non-distributed training, there is only one dataloader for all GPUs.
+
+    Args:
+        dataset (Dataset): A PyTorch dataset.
+        samples_per_gpu (int): Number of training samples on each GPU, i.e.,
+            batch size of each GPU.
+        workers_per_gpu (int): How many subprocesses to use for data loading
+            for each GPU.
+        num_gpus (int): Number of GPUs. Only used in non-distributed training.
+        dist (bool): Distributed training/test or not. Default: True.
+        shuffle (bool): Whether to shuffle the data at every epoch.
+            Default: True.
+        kwargs: any keyword argument to be used to initialize DataLoader
+
+    Returns:
+        DataLoader: A PyTorch dataloader.
+    """
+    rank, world_size = get_dist_info()
+    if dist:
+        # DistributedGroupSampler will definitely shuffle the data to satisfy
+        # that images on each GPU are in the same group
+        if shuffle:
+            sampler = DistributedGroupSampler(dataset, samples_per_gpu,
+                                              world_size, rank)
+        else:
+            sampler = DistributedSampler(
+                dataset, world_size, rank, shuffle=False)
+        batch_size = samples_per_gpu
+        num_workers = workers_per_gpu
+    else:
+        sampler = GroupSampler(dataset, samples_per_gpu) if shuffle else None
+        batch_size = num_gpus * samples_per_gpu
+        num_workers = num_gpus * workers_per_gpu
+
+    init_fn = partial(
+        worker_init_fn, num_workers=num_workers, rank=rank,
+        seed=seed) if seed is not None else None
+
+    data_loader = DataLoader(
+        dataset,
+        batch_size=batch_size,
+        sampler=sampler,
+        num_workers=num_workers,
+        collate_fn=partial(collate, samples_per_gpu=samples_per_gpu),
+        pin_memory=False,
+        worker_init_fn=init_fn,
+        **kwargs)
+
+    return data_loader
+
+
+def worker_init_fn(worker_id, num_workers, rank, seed):
+    # The seed of each worker equals to
+    # num_worker * rank + worker_id + user_seed
+    worker_seed = num_workers * rank + worker_id + seed
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
diff --git a/detection/scrfd/mmdet/datasets/cityscapes.py b/detection/scrfd/mmdet/datasets/cityscapes.py
new file mode 100755
index 000000000..71eead87e
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/cityscapes.py
@@ -0,0 +1,334 @@
+# Modified from https://github.com/facebookresearch/detectron2/blob/master/detectron2/data/datasets/cityscapes.py # noqa
+# and https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalInstanceLevelSemanticLabeling.py # noqa
+
+import glob
+import os
+import os.path as osp
+import tempfile
+from collections import OrderedDict
+
+import mmcv
+import numpy as np
+import pycocotools.mask as maskUtils
+from mmcv.utils import print_log
+
+from .builder import DATASETS
+from .coco import CocoDataset
+
+
+@DATASETS.register_module()
+class CityscapesDataset(CocoDataset):
+
+    CLASSES = ('person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',
+               'bicycle')
+
+    def _filter_imgs(self, min_size=32):
+        """Filter images too small or without ground truths."""
+        valid_inds = []
+        # obtain images that contain annotation
+        ids_with_ann = set(_['image_id'] for _ in self.coco.anns.values())
+        # obtain images that contain annotations of the required categories
+        ids_in_cat = set()
+        for i, class_id in enumerate(self.cat_ids):
+            ids_in_cat |= set(self.coco.cat_img_map[class_id])
+        # merge the image id sets of the two conditions and use the merged set
+        # to filter out images if self.filter_empty_gt=True
+        ids_in_cat &= ids_with_ann
+
+        valid_img_ids = []
+        for i, img_info in enumerate(self.data_infos):
+            img_id = img_info['id']
+            ann_ids = self.coco.getAnnIds(imgIds=[img_id])
+            ann_info = self.coco.loadAnns(ann_ids)
+            all_iscrowd = all([_['iscrowd'] for _ in ann_info])
+            if self.filter_empty_gt and (self.img_ids[i] not in ids_in_cat
+                                         or all_iscrowd):
+                continue
+            if min(img_info['width'], img_info['height']) >= min_size:
+                valid_inds.append(i)
+                valid_img_ids.append(img_id)
+        self.img_ids = valid_img_ids
+        return valid_inds
+
+    def _parse_ann_info(self, img_info, ann_info):
+        """Parse bbox and mask annotation.
+
+        Args:
+            img_info (dict): Image info of an image.
+            ann_info (list[dict]): Annotation info of an image.
+
+        Returns:
+            dict: A dict containing the following keys: bboxes, \
+                bboxes_ignore, labels, masks, seg_map. \
+                "masks" are already decoded into binary masks.
+        """
+        gt_bboxes = []
+        gt_labels = []
+        gt_bboxes_ignore = []
+        gt_masks_ann = []
+
+        for i, ann in enumerate(ann_info):
+            if ann.get('ignore', False):
+                continue
+            x1, y1, w, h = ann['bbox']
+            if ann['area'] <= 0 or w < 1 or h < 1:
+                continue
+            if ann['category_id'] not in self.cat_ids:
+                continue
+            bbox = [x1, y1, x1 + w, y1 + h]
+            if ann.get('iscrowd', False):
+                gt_bboxes_ignore.append(bbox)
+            else:
+                gt_bboxes.append(bbox)
+                gt_labels.append(self.cat2label[ann['category_id']])
+                gt_masks_ann.append(ann['segmentation'])
+
+        if gt_bboxes:
+            gt_bboxes = np.array(gt_bboxes, dtype=np.float32)
+            gt_labels = np.array(gt_labels, dtype=np.int64)
+        else:
+            gt_bboxes = np.zeros((0, 4), dtype=np.float32)
+            gt_labels = np.array([], dtype=np.int64)
+
+        if gt_bboxes_ignore:
+            gt_bboxes_ignore = np.array(gt_bboxes_ignore, dtype=np.float32)
+        else:
+            gt_bboxes_ignore = np.zeros((0, 4), dtype=np.float32)
+
+        ann = dict(
+            bboxes=gt_bboxes,
+            labels=gt_labels,
+            bboxes_ignore=gt_bboxes_ignore,
+            masks=gt_masks_ann,
+            seg_map=img_info['segm_file'])
+
+        return ann
+
+    def results2txt(self, results, outfile_prefix):
+        """Dump the detection results to a txt file.
+
+        Args:
+            results (list[list | tuple]): Testing results of the
+                dataset.
+            outfile_prefix (str): The filename prefix of the json files.
+                If the prefix is "somepath/xxx",
+                the txt files will be named "somepath/xxx.txt".
+
+        Returns:
+            list[str]: Result txt files which contains corresponding \
+                instance segmentation images.
+        """
+        try:
+            import cityscapesscripts.helpers.labels as CSLabels
+        except ImportError:
+            raise ImportError('Please run "pip install citscapesscripts" to '
+                              'install cityscapesscripts first.')
+        result_files = []
+        os.makedirs(outfile_prefix, exist_ok=True)
+        prog_bar = mmcv.ProgressBar(len(self))
+        for idx in range(len(self)):
+            result = results[idx]
+            filename = self.data_infos[idx]['filename']
+            basename = osp.splitext(osp.basename(filename))[0]
+            pred_txt = osp.join(outfile_prefix, basename + '_pred.txt')
+
+            bbox_result, segm_result = result
+            bboxes = np.vstack(bbox_result)
+            # segm results
+            if isinstance(segm_result, tuple):
+                # Some detectors use different scores for bbox and mask,
+                # like Mask Scoring R-CNN. Score of segm will be used instead
+                # of bbox score.
+                segms = mmcv.concat_list(segm_result[0])
+                mask_score = segm_result[1]
+            else:
+                # use bbox score for mask score
+                segms = mmcv.concat_list(segm_result)
+                mask_score = [bbox[-1] for bbox in bboxes]
+            labels = [
+                np.full(bbox.shape[0], i, dtype=np.int32)
+                for i, bbox in enumerate(bbox_result)
+            ]
+            labels = np.concatenate(labels)
+
+            assert len(bboxes) == len(segms) == len(labels)
+            num_instances = len(bboxes)
+            prog_bar.update()
+            with open(pred_txt, 'w') as fout:
+                for i in range(num_instances):
+                    pred_class = labels[i]
+                    classes = self.CLASSES[pred_class]
+                    class_id = CSLabels.name2label[classes].id
+                    score = mask_score[i]
+                    mask = maskUtils.decode(segms[i]).astype(np.uint8)
+                    png_filename = osp.join(outfile_prefix,
+                                            basename + f'_{i}_{classes}.png')
+                    mmcv.imwrite(mask, png_filename)
+                    fout.write(f'{osp.basename(png_filename)} {class_id} '
+                               f'{score}\n')
+            result_files.append(pred_txt)
+
+        return result_files
+
+    def format_results(self, results, txtfile_prefix=None):
+        """Format the results to txt (standard format for Cityscapes
+        evaluation).
+
+        Args:
+            results (list): Testing results of the dataset.
+            txtfile_prefix (str | None): The prefix of txt files. It includes
+                the file path and the prefix of filename, e.g., "a/b/prefix".
+                If not specified, a temp file will be created. Default: None.
+
+        Returns:
+            tuple: (result_files, tmp_dir), result_files is a dict containing \
+                the json filepaths, tmp_dir is the temporal directory created \
+                for saving txt/png files when txtfile_prefix is not specified.
+        """
+        assert isinstance(results, list), 'results must be a list'
+        assert len(results) == len(self), (
+            'The length of results is not equal to the dataset len: {} != {}'.
+            format(len(results), len(self)))
+
+        assert isinstance(results, list), 'results must be a list'
+        assert len(results) == len(self), (
+            'The length of results is not equal to the dataset len: {} != {}'.
+            format(len(results), len(self)))
+
+        if txtfile_prefix is None:
+            tmp_dir = tempfile.TemporaryDirectory()
+            txtfile_prefix = osp.join(tmp_dir.name, 'results')
+        else:
+            tmp_dir = None
+        result_files = self.results2txt(results, txtfile_prefix)
+
+        return result_files, tmp_dir
+
+    def evaluate(self,
+                 results,
+                 metric='bbox',
+                 logger=None,
+                 outfile_prefix=None,
+                 classwise=False,
+                 proposal_nums=(100, 300, 1000),
+                 iou_thrs=np.arange(0.5, 0.96, 0.05)):
+        """Evaluation in Cityscapes/COCO protocol.
+
+        Args:
+            results (list[list | tuple]): Testing results of the dataset.
+            metric (str | list[str]): Metrics to be evaluated. Options are
+                'bbox', 'segm', 'proposal', 'proposal_fast'.
+            logger (logging.Logger | str | None): Logger used for printing
+                related information during evaluation. Default: None.
+            outfile_prefix (str | None): The prefix of output file. It includes
+                the file path and the prefix of filename, e.g., "a/b/prefix".
+                If results are evaluated with COCO protocol, it would be the
+                prefix of output json file. For example, the metric is 'bbox'
+                and 'segm', then json files would be "a/b/prefix.bbox.json" and
+                "a/b/prefix.segm.json".
+                If results are evaluated with cityscapes protocol, it would be
+                the prefix of output txt/png files. The output files would be
+                png images under folder "a/b/prefix/xxx/" and the file name of
+                images would be written into a txt file
+                "a/b/prefix/xxx_pred.txt", where "xxx" is the video name of
+                cityscapes. If not specified, a temp file will be created.
+                Default: None.
+            classwise (bool): Whether to evaluating the AP for each class.
+            proposal_nums (Sequence[int]): Proposal number used for evaluating
+                recalls, such as recall@100, recall@1000.
+                Default: (100, 300, 1000).
+            iou_thrs (Sequence[float]): IoU threshold used for evaluating
+                recalls. If set to a list, the average recall of all IoUs will
+                also be computed. Default: 0.5.
+
+        Returns:
+            dict[str, float]: COCO style evaluation metric or cityscapes mAP \
+                and AP@50.
+        """
+        eval_results = dict()
+
+        metrics = metric.copy() if isinstance(metric, list) else [metric]
+
+        if 'cityscapes' in metrics:
+            eval_results.update(
+                self._evaluate_cityscapes(results, outfile_prefix, logger))
+            metrics.remove('cityscapes')
+
+        # left metrics are all coco metric
+        if len(metrics) > 0:
+            # create CocoDataset with CityscapesDataset annotation
+            self_coco = CocoDataset(self.ann_file, self.pipeline.transforms,
+                                    None, self.data_root, self.img_prefix,
+                                    self.seg_prefix, self.proposal_file,
+                                    self.test_mode, self.filter_empty_gt)
+            # TODO: remove this in the future
+            # reload annotations of correct class
+            self_coco.CLASSES = self.CLASSES
+            self_coco.data_infos = self_coco.load_annotations(self.ann_file)
+            eval_results.update(
+                self_coco.evaluate(results, metrics, logger, outfile_prefix,
+                                   classwise, proposal_nums, iou_thrs))
+
+        return eval_results
+
+    def _evaluate_cityscapes(self, results, txtfile_prefix, logger):
+        """Evaluation in Cityscapes protocol.
+
+        Args:
+            results (list): Testing results of the dataset.
+            txtfile_prefix (str | None): The prefix of output txt file
+            logger (logging.Logger | str | None): Logger used for printing
+                related information during evaluation. Default: None.
+
+        Returns:
+            dict[str: float]: Cityscapes evaluation results, contains 'mAP' \
+                and 'AP@50'.
+        """
+
+        try:
+            import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as CSEval  # noqa
+        except ImportError:
+            raise ImportError('Please run "pip install citscapesscripts" to '
+                              'install cityscapesscripts first.')
+        msg = 'Evaluating in Cityscapes style'
+        if logger is None:
+            msg = '\n' + msg
+        print_log(msg, logger=logger)
+
+        result_files, tmp_dir = self.format_results(results, txtfile_prefix)
+
+        if tmp_dir is None:
+            result_dir = osp.join(txtfile_prefix, 'results')
+        else:
+            result_dir = osp.join(tmp_dir.name, 'results')
+
+        eval_results = OrderedDict()
+        print_log(f'Evaluating results under {result_dir} ...', logger=logger)
+
+        # set global states in cityscapes evaluation API
+        CSEval.args.cityscapesPath = os.path.join(self.img_prefix, '../..')
+        CSEval.args.predictionPath = os.path.abspath(result_dir)
+        CSEval.args.predictionWalk = None
+        CSEval.args.JSONOutput = False
+        CSEval.args.colorized = False
+        CSEval.args.gtInstancesFile = os.path.join(result_dir,
+                                                   'gtInstances.json')
+        CSEval.args.groundTruthSearch = os.path.join(
+            self.img_prefix.replace('leftImg8bit', 'gtFine'),
+            '*/*_gtFine_instanceIds.png')
+
+        groundTruthImgList = glob.glob(CSEval.args.groundTruthSearch)
+        assert len(groundTruthImgList), 'Cannot find ground truth images' \
+            f' in {CSEval.args.groundTruthSearch}.'
+        predictionImgList = []
+        for gt in groundTruthImgList:
+            predictionImgList.append(CSEval.getPrediction(gt, CSEval.args))
+        CSEval_results = CSEval.evaluateImgLists(predictionImgList,
+                                                 groundTruthImgList,
+                                                 CSEval.args)['averages']
+
+        eval_results['mAP'] = CSEval_results['allAp']
+        eval_results['AP@50'] = CSEval_results['allAp50%']
+        if tmp_dir is not None:
+            tmp_dir.cleanup()
+        return eval_results
diff --git a/detection/scrfd/mmdet/datasets/coco.py b/detection/scrfd/mmdet/datasets/coco.py
new file mode 100755
index 000000000..9eea6a4fd
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/coco.py
@@ -0,0 +1,544 @@
+import itertools
+import logging
+import os.path as osp
+import tempfile
+from collections import OrderedDict
+
+import mmcv
+import numpy as np
+from mmcv.utils import print_log
+from pycocotools.coco import COCO
+from pycocotools.cocoeval import COCOeval
+from terminaltables import AsciiTable
+
+from mmdet.core import eval_recalls
+from .builder import DATASETS
+from .custom import CustomDataset
+
+try:
+    import pycocotools
+    if not hasattr(pycocotools, '__sphinx_mock__'):  # for doc generation
+        assert pycocotools.__version__ >= '12.0.2'
+except AssertionError:
+    raise AssertionError('Incompatible version of pycocotools is installed. '
+                         'Run pip uninstall pycocotools first. Then run pip '
+                         'install mmpycocotools to install open-mmlab forked '
+                         'pycocotools.')
+
+
+@DATASETS.register_module()
+class CocoDataset(CustomDataset):
+
+    CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
+               'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
+               'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
+               'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe',
+               'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
+               'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat',
+               'baseball glove', 'skateboard', 'surfboard', 'tennis racket',
+               'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
+               'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot',
+               'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
+               'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop',
+               'mouse', 'remote', 'keyboard', 'cell phone', 'microwave',
+               'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock',
+               'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush')
+
+    def load_annotations(self, ann_file):
+        """Load annotation from COCO style annotation file.
+
+        Args:
+            ann_file (str): Path of annotation file.
+
+        Returns:
+            list[dict]: Annotation info from COCO api.
+        """
+
+        self.coco = COCO(ann_file)
+        self.cat_ids = self.coco.get_cat_ids(cat_names=self.CLASSES)
+        self.cat2label = {cat_id: i for i, cat_id in enumerate(self.cat_ids)}
+        self.img_ids = self.coco.get_img_ids()
+        data_infos = []
+        for i in self.img_ids:
+            info = self.coco.load_imgs([i])[0]
+            info['filename'] = info['file_name']
+            data_infos.append(info)
+        return data_infos
+
+    def get_ann_info(self, idx):
+        """Get COCO annotation by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Annotation info of specified index.
+        """
+
+        img_id = self.data_infos[idx]['id']
+        ann_ids = self.coco.get_ann_ids(img_ids=[img_id])
+        ann_info = self.coco.load_anns(ann_ids)
+        return self._parse_ann_info(self.data_infos[idx], ann_info)
+
+    def get_cat_ids(self, idx):
+        """Get COCO category ids by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            list[int]: All categories in the image of specified index.
+        """
+
+        img_id = self.data_infos[idx]['id']
+        ann_ids = self.coco.get_ann_ids(img_ids=[img_id])
+        ann_info = self.coco.load_anns(ann_ids)
+        return [ann['category_id'] for ann in ann_info]
+
+    def _filter_imgs(self, min_size=32):
+        """Filter images too small or without ground truths."""
+        valid_inds = []
+        # obtain images that contain annotation
+        ids_with_ann = set(_['image_id'] for _ in self.coco.anns.values())
+        # obtain images that contain annotations of the required categories
+        ids_in_cat = set()
+        for i, class_id in enumerate(self.cat_ids):
+            ids_in_cat |= set(self.coco.cat_img_map[class_id])
+        # merge the image id sets of the two conditions and use the merged set
+        # to filter out images if self.filter_empty_gt=True
+        ids_in_cat &= ids_with_ann
+
+        valid_img_ids = []
+        for i, img_info in enumerate(self.data_infos):
+            img_id = self.img_ids[i]
+            if self.filter_empty_gt and img_id not in ids_in_cat:
+                continue
+            if min(img_info['width'], img_info['height']) >= min_size:
+                valid_inds.append(i)
+                valid_img_ids.append(img_id)
+        self.img_ids = valid_img_ids
+        return valid_inds
+
+    def _parse_ann_info(self, img_info, ann_info):
+        """Parse bbox and mask annotation.
+
+        Args:
+            ann_info (list[dict]): Annotation info of an image.
+            with_mask (bool): Whether to parse mask annotations.
+
+        Returns:
+            dict: A dict containing the following keys: bboxes, bboxes_ignore,\
+                labels, masks, seg_map. "masks" are raw annotations and not \
+                decoded into binary masks.
+        """
+        gt_bboxes = []
+        gt_labels = []
+        gt_bboxes_ignore = []
+        gt_masks_ann = []
+        for i, ann in enumerate(ann_info):
+            if ann.get('ignore', False):
+                continue
+            x1, y1, w, h = ann['bbox']
+            inter_w = max(0, min(x1 + w, img_info['width']) - max(x1, 0))
+            inter_h = max(0, min(y1 + h, img_info['height']) - max(y1, 0))
+            if inter_w * inter_h == 0:
+                continue
+            if ann['area'] <= 0 or w < 1 or h < 1:
+                continue
+            if ann['category_id'] not in self.cat_ids:
+                continue
+            bbox = [x1, y1, x1 + w, y1 + h]
+            if ann.get('iscrowd', False):
+                gt_bboxes_ignore.append(bbox)
+            else:
+                gt_bboxes.append(bbox)
+                gt_labels.append(self.cat2label[ann['category_id']])
+                gt_masks_ann.append(ann.get('segmentation', None))
+
+        if gt_bboxes:
+            gt_bboxes = np.array(gt_bboxes, dtype=np.float32)
+            gt_labels = np.array(gt_labels, dtype=np.int64)
+        else:
+            gt_bboxes = np.zeros((0, 4), dtype=np.float32)
+            gt_labels = np.array([], dtype=np.int64)
+
+        if gt_bboxes_ignore:
+            gt_bboxes_ignore = np.array(gt_bboxes_ignore, dtype=np.float32)
+        else:
+            gt_bboxes_ignore = np.zeros((0, 4), dtype=np.float32)
+
+        seg_map = img_info['filename'].replace('jpg', 'png')
+
+        ann = dict(
+            bboxes=gt_bboxes,
+            labels=gt_labels,
+            bboxes_ignore=gt_bboxes_ignore,
+            masks=gt_masks_ann,
+            seg_map=seg_map)
+
+        return ann
+
+    def xyxy2xywh(self, bbox):
+        """Convert ``xyxy`` style bounding boxes to ``xywh`` style for COCO
+        evaluation.
+
+        Args:
+            bbox (numpy.ndarray): The bounding boxes, shape (4, ), in
+                ``xyxy`` order.
+
+        Returns:
+            list[float]: The converted bounding boxes, in ``xywh`` order.
+        """
+
+        _bbox = bbox.tolist()
+        return [
+            _bbox[0],
+            _bbox[1],
+            _bbox[2] - _bbox[0],
+            _bbox[3] - _bbox[1],
+        ]
+
+    def _proposal2json(self, results):
+        """Convert proposal results to COCO json style."""
+        json_results = []
+        for idx in range(len(self)):
+            img_id = self.img_ids[idx]
+            bboxes = results[idx]
+            for i in range(bboxes.shape[0]):
+                data = dict()
+                data['image_id'] = img_id
+                data['bbox'] = self.xyxy2xywh(bboxes[i])
+                data['score'] = float(bboxes[i][4])
+                data['category_id'] = 1
+                json_results.append(data)
+        return json_results
+
+    def _det2json(self, results):
+        """Convert detection results to COCO json style."""
+        json_results = []
+        for idx in range(len(self)):
+            img_id = self.img_ids[idx]
+            result = results[idx]
+            for label in range(len(result)):
+                bboxes = result[label]
+                for i in range(bboxes.shape[0]):
+                    data = dict()
+                    data['image_id'] = img_id
+                    data['bbox'] = self.xyxy2xywh(bboxes[i])
+                    data['score'] = float(bboxes[i][4])
+                    data['category_id'] = self.cat_ids[label]
+                    json_results.append(data)
+        return json_results
+
+    def _segm2json(self, results):
+        """Convert instance segmentation results to COCO json style."""
+        bbox_json_results = []
+        segm_json_results = []
+        for idx in range(len(self)):
+            img_id = self.img_ids[idx]
+            det, seg = results[idx]
+            for label in range(len(det)):
+                # bbox results
+                bboxes = det[label]
+                for i in range(bboxes.shape[0]):
+                    data = dict()
+                    data['image_id'] = img_id
+                    data['bbox'] = self.xyxy2xywh(bboxes[i])
+                    data['score'] = float(bboxes[i][4])
+                    data['category_id'] = self.cat_ids[label]
+                    bbox_json_results.append(data)
+
+                # segm results
+                # some detectors use different scores for bbox and mask
+                if isinstance(seg, tuple):
+                    segms = seg[0][label]
+                    mask_score = seg[1][label]
+                else:
+                    segms = seg[label]
+                    mask_score = [bbox[4] for bbox in bboxes]
+                for i in range(bboxes.shape[0]):
+                    data = dict()
+                    data['image_id'] = img_id
+                    data['bbox'] = self.xyxy2xywh(bboxes[i])
+                    data['score'] = float(mask_score[i])
+                    data['category_id'] = self.cat_ids[label]
+                    if isinstance(segms[i]['counts'], bytes):
+                        segms[i]['counts'] = segms[i]['counts'].decode()
+                    data['segmentation'] = segms[i]
+                    segm_json_results.append(data)
+        return bbox_json_results, segm_json_results
+
+    def results2json(self, results, outfile_prefix):
+        """Dump the detection results to a COCO style json file.
+
+        There are 3 types of results: proposals, bbox predictions, mask
+        predictions, and they have different data types. This method will
+        automatically recognize the type, and dump them to json files.
+
+        Args:
+            results (list[list | tuple | ndarray]): Testing results of the
+                dataset.
+            outfile_prefix (str): The filename prefix of the json files. If the
+                prefix is "somepath/xxx", the json files will be named
+                "somepath/xxx.bbox.json", "somepath/xxx.segm.json",
+                "somepath/xxx.proposal.json".
+
+        Returns:
+            dict[str: str]: Possible keys are "bbox", "segm", "proposal", and \
+                values are corresponding filenames.
+        """
+        result_files = dict()
+        if isinstance(results[0], list):
+            json_results = self._det2json(results)
+            result_files['bbox'] = f'{outfile_prefix}.bbox.json'
+            result_files['proposal'] = f'{outfile_prefix}.bbox.json'
+            mmcv.dump(json_results, result_files['bbox'])
+        elif isinstance(results[0], tuple):
+            json_results = self._segm2json(results)
+            result_files['bbox'] = f'{outfile_prefix}.bbox.json'
+            result_files['proposal'] = f'{outfile_prefix}.bbox.json'
+            result_files['segm'] = f'{outfile_prefix}.segm.json'
+            mmcv.dump(json_results[0], result_files['bbox'])
+            mmcv.dump(json_results[1], result_files['segm'])
+        elif isinstance(results[0], np.ndarray):
+            json_results = self._proposal2json(results)
+            result_files['proposal'] = f'{outfile_prefix}.proposal.json'
+            mmcv.dump(json_results, result_files['proposal'])
+        else:
+            raise TypeError('invalid type of results')
+        return result_files
+
+    def fast_eval_recall(self, results, proposal_nums, iou_thrs, logger=None):
+        gt_bboxes = []
+        for i in range(len(self.img_ids)):
+            ann_ids = self.coco.get_ann_ids(img_ids=self.img_ids[i])
+            ann_info = self.coco.load_anns(ann_ids)
+            if len(ann_info) == 0:
+                gt_bboxes.append(np.zeros((0, 4)))
+                continue
+            bboxes = []
+            for ann in ann_info:
+                if ann.get('ignore', False) or ann['iscrowd']:
+                    continue
+                x1, y1, w, h = ann['bbox']
+                bboxes.append([x1, y1, x1 + w, y1 + h])
+            bboxes = np.array(bboxes, dtype=np.float32)
+            if bboxes.shape[0] == 0:
+                bboxes = np.zeros((0, 4))
+            gt_bboxes.append(bboxes)
+
+        recalls = eval_recalls(
+            gt_bboxes, results, proposal_nums, iou_thrs, logger=logger)
+        ar = recalls.mean(axis=1)
+        return ar
+
+    def format_results(self, results, jsonfile_prefix=None, **kwargs):
+        """Format the results to json (standard format for COCO evaluation).
+
+        Args:
+            results (list[tuple | numpy.ndarray]): Testing results of the
+                dataset.
+            jsonfile_prefix (str | None): The prefix of json files. It includes
+                the file path and the prefix of filename, e.g., "a/b/prefix".
+                If not specified, a temp file will be created. Default: None.
+
+        Returns:
+            tuple: (result_files, tmp_dir), result_files is a dict containing \
+                the json filepaths, tmp_dir is the temporal directory created \
+                for saving json files when jsonfile_prefix is not specified.
+        """
+        assert isinstance(results, list), 'results must be a list'
+        assert len(results) == len(self), (
+            'The length of results is not equal to the dataset len: {} != {}'.
+            format(len(results), len(self)))
+
+        if jsonfile_prefix is None:
+            tmp_dir = tempfile.TemporaryDirectory()
+            jsonfile_prefix = osp.join(tmp_dir.name, 'results')
+        else:
+            tmp_dir = None
+        result_files = self.results2json(results, jsonfile_prefix)
+        return result_files, tmp_dir
+
+    def evaluate(self,
+                 results,
+                 metric='bbox',
+                 logger=None,
+                 jsonfile_prefix=None,
+                 classwise=False,
+                 proposal_nums=(100, 300, 1000),
+                 iou_thrs=None,
+                 metric_items=None):
+        """Evaluation in COCO protocol.
+
+        Args:
+            results (list[list | tuple]): Testing results of the dataset.
+            metric (str | list[str]): Metrics to be evaluated. Options are
+                'bbox', 'segm', 'proposal', 'proposal_fast'.
+            logger (logging.Logger | str | None): Logger used for printing
+                related information during evaluation. Default: None.
+            jsonfile_prefix (str | None): The prefix of json files. It includes
+                the file path and the prefix of filename, e.g., "a/b/prefix".
+                If not specified, a temp file will be created. Default: None.
+            classwise (bool): Whether to evaluating the AP for each class.
+            proposal_nums (Sequence[int]): Proposal number used for evaluating
+                recalls, such as recall@100, recall@1000.
+                Default: (100, 300, 1000).
+            iou_thrs (Sequence[float], optional): IoU threshold used for
+                evaluating recalls/mAPs. If set to a list, the average of all
+                IoUs will also be computed. If not specified, [0.50, 0.55,
+                0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95] will be used.
+                Default: None.
+            metric_items (list[str] | str, optional): Metric items that will
+                be returned. If not specified, ``['AR@100', 'AR@300',
+                'AR@1000', 'AR_s@1000', 'AR_m@1000', 'AR_l@1000' ]`` will be
+                used when ``metric=='proposal'``, ``['mAP', 'mAP_50', 'mAP_75',
+                'mAP_s', 'mAP_m', 'mAP_l']`` will be used when
+                ``metric=='bbox' or metric=='segm'``.
+
+        Returns:
+            dict[str, float]: COCO style evaluation metric.
+        """
+
+        metrics = metric if isinstance(metric, list) else [metric]
+        allowed_metrics = ['bbox', 'segm', 'proposal', 'proposal_fast']
+        for metric in metrics:
+            if metric not in allowed_metrics:
+                raise KeyError(f'metric {metric} is not supported')
+        if iou_thrs is None:
+            iou_thrs = np.linspace(
+                .5, 0.95, int(np.round((0.95 - .5) / .05)) + 1, endpoint=True)
+        if metric_items is not None:
+            if not isinstance(metric_items, list):
+                metric_items = [metric_items]
+
+        result_files, tmp_dir = self.format_results(results, jsonfile_prefix)
+
+        eval_results = OrderedDict()
+        cocoGt = self.coco
+        for metric in metrics:
+            msg = f'Evaluating {metric}...'
+            if logger is None:
+                msg = '\n' + msg
+            print_log(msg, logger=logger)
+
+            if metric == 'proposal_fast':
+                ar = self.fast_eval_recall(
+                    results, proposal_nums, iou_thrs, logger='silent')
+                log_msg = []
+                for i, num in enumerate(proposal_nums):
+                    eval_results[f'AR@{num}'] = ar[i]
+                    log_msg.append(f'\nAR@{num}\t{ar[i]:.4f}')
+                log_msg = ''.join(log_msg)
+                print_log(log_msg, logger=logger)
+                continue
+
+            if metric not in result_files:
+                raise KeyError(f'{metric} is not in results')
+            try:
+                cocoDt = cocoGt.loadRes(result_files[metric])
+            except IndexError:
+                print_log(
+                    'The testing results of the whole dataset is empty.',
+                    logger=logger,
+                    level=logging.ERROR)
+                break
+
+            iou_type = 'bbox' if metric == 'proposal' else metric
+            cocoEval = COCOeval(cocoGt, cocoDt, iou_type)
+            cocoEval.params.catIds = self.cat_ids
+            cocoEval.params.imgIds = self.img_ids
+            cocoEval.params.maxDets = list(proposal_nums)
+            cocoEval.params.iouThrs = iou_thrs
+            # mapping of cocoEval.stats
+            coco_metric_names = {
+                'mAP': 0,
+                'mAP_50': 1,
+                'mAP_75': 2,
+                'mAP_s': 3,
+                'mAP_m': 4,
+                'mAP_l': 5,
+                'AR@100': 6,
+                'AR@300': 7,
+                'AR@1000': 8,
+                'AR_s@1000': 9,
+                'AR_m@1000': 10,
+                'AR_l@1000': 11
+            }
+            if metric_items is not None:
+                for metric_item in metric_items:
+                    if metric_item not in coco_metric_names:
+                        raise KeyError(
+                            f'metric item {metric_item} is not supported')
+
+            if metric == 'proposal':
+                cocoEval.params.useCats = 0
+                cocoEval.evaluate()
+                cocoEval.accumulate()
+                cocoEval.summarize()
+                if metric_items is None:
+                    metric_items = [
+                        'AR@100', 'AR@300', 'AR@1000', 'AR_s@1000',
+                        'AR_m@1000', 'AR_l@1000'
+                    ]
+
+                for item in metric_items:
+                    val = float(
+                        f'{cocoEval.stats[coco_metric_names[item]]:.3f}')
+                    eval_results[item] = val
+            else:
+                cocoEval.evaluate()
+                cocoEval.accumulate()
+                cocoEval.summarize()
+                if classwise:  # Compute per-category AP
+                    # Compute per-category AP
+                    # from https://github.com/facebookresearch/detectron2/
+                    precisions = cocoEval.eval['precision']
+                    # precision: (iou, recall, cls, area range, max dets)
+                    assert len(self.cat_ids) == precisions.shape[2]
+
+                    results_per_category = []
+                    for idx, catId in enumerate(self.cat_ids):
+                        # area range index 0: all area ranges
+                        # max dets index -1: typically 100 per image
+                        nm = self.coco.loadCats(catId)[0]
+                        precision = precisions[:, :, idx, 0, -1]
+                        precision = precision[precision > -1]
+                        if precision.size:
+                            ap = np.mean(precision)
+                        else:
+                            ap = float('nan')
+                        results_per_category.append(
+                            (f'{nm["name"]}', f'{float(ap):0.3f}'))
+
+                    num_columns = min(6, len(results_per_category) * 2)
+                    results_flatten = list(
+                        itertools.chain(*results_per_category))
+                    headers = ['category', 'AP'] * (num_columns // 2)
+                    results_2d = itertools.zip_longest(*[
+                        results_flatten[i::num_columns]
+                        for i in range(num_columns)
+                    ])
+                    table_data = [headers]
+                    table_data += [result for result in results_2d]
+                    table = AsciiTable(table_data)
+                    print_log('\n' + table.table, logger=logger)
+
+                if metric_items is None:
+                    metric_items = [
+                        'mAP', 'mAP_50', 'mAP_75', 'mAP_s', 'mAP_m', 'mAP_l'
+                    ]
+
+                for metric_item in metric_items:
+                    key = f'{metric}_{metric_item}'
+                    val = float(
+                        f'{cocoEval.stats[coco_metric_names[metric_item]]:.3f}'
+                    )
+                    eval_results[key] = val
+                ap = cocoEval.stats[:6]
+                eval_results[f'{metric}_mAP_copypaste'] = (
+                    f'{ap[0]:.3f} {ap[1]:.3f} {ap[2]:.3f} {ap[3]:.3f} '
+                    f'{ap[4]:.3f} {ap[5]:.3f}')
+        if tmp_dir is not None:
+            tmp_dir.cleanup()
+        return eval_results
diff --git a/detection/scrfd/mmdet/datasets/custom.py b/detection/scrfd/mmdet/datasets/custom.py
new file mode 100755
index 000000000..1f78a1f5a
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/custom.py
@@ -0,0 +1,363 @@
+import os.path as osp
+import warnings
+from collections import OrderedDict
+
+import mmcv
+import numpy as np
+from torch.utils.data import Dataset
+
+from mmdet.core import eval_map, eval_recalls
+from .builder import DATASETS
+from .pipelines import Compose
+
+
+@DATASETS.register_module()
+class CustomDataset(Dataset):
+    """Custom dataset for detection.
+
+    The annotation format is shown as follows. The `ann` field is optional for
+    testing.
+
+    .. code-block:: none
+
+        [
+            {
+                'filename': 'a.jpg',
+                'width': 1280,
+                'height': 720,
+                'ann': {
+                    'bboxes': <np.ndarray> (n, 4) in (x1, y1, x2, y2) order.
+                    'labels': <np.ndarray> (n, ),
+                    'bboxes_ignore': <np.ndarray> (k, 4), (optional field)
+                    'labels_ignore': <np.ndarray> (k, 4) (optional field)
+                }
+            },
+            ...
+        ]
+
+    Args:
+        ann_file (str): Annotation file path.
+        pipeline (list[dict]): Processing pipeline.
+        classes (str | Sequence[str], optional): Specify classes to load.
+            If is None, ``cls.CLASSES`` will be used. Default: None.
+        data_root (str, optional): Data root for ``ann_file``,
+            ``img_prefix``, ``seg_prefix``, ``proposal_file`` if specified.
+        test_mode (bool, optional): If set True, annotation will not be loaded.
+        filter_empty_gt (bool, optional): If set true, images without bounding
+            boxes of the dataset's classes will be filtered out. This option
+            only works when `test_mode=False`, i.e., we never filter images
+            during tests.
+    """
+
+    CLASSES = None
+
+    def __init__(self,
+                 ann_file,
+                 pipeline,
+                 classes=None,
+                 data_root=None,
+                 img_prefix='',
+                 seg_prefix=None,
+                 proposal_file=None,
+                 test_mode=False,
+                 filter_empty_gt=True,
+                 support_mosaic=False,
+                 ):
+        self.ann_file = ann_file
+        self.data_root = data_root
+        self.img_prefix = img_prefix
+        self.seg_prefix = seg_prefix
+        self.proposal_file = proposal_file
+        self.test_mode = test_mode
+        self.filter_empty_gt = filter_empty_gt
+        self.support_mosaic = support_mosaic
+        self.CLASSES = self.get_classes(classes)
+
+        # join paths if data_root is specified
+        if self.data_root is not None:
+            if not osp.isabs(self.ann_file):
+                self.ann_file = osp.join(self.data_root, self.ann_file)
+            if not (self.img_prefix is None or osp.isabs(self.img_prefix)):
+                self.img_prefix = osp.join(self.data_root, self.img_prefix)
+            if not (self.seg_prefix is None or osp.isabs(self.seg_prefix)):
+                self.seg_prefix = osp.join(self.data_root, self.seg_prefix)
+            if not (self.proposal_file is None
+                    or osp.isabs(self.proposal_file)):
+                self.proposal_file = osp.join(self.data_root,
+                                              self.proposal_file)
+        # load annotations (and proposals)
+        self.data_infos = self.load_annotations(self.ann_file)
+
+        if self.proposal_file is not None:
+            self.proposals = self.load_proposals(self.proposal_file)
+        else:
+            self.proposals = None
+
+        # filter images too small and containing no annotations
+        if not test_mode:
+            valid_inds = self._filter_imgs()
+            self.data_infos = [self.data_infos[i] for i in valid_inds]
+            if self.proposals is not None:
+                self.proposals = [self.proposals[i] for i in valid_inds]
+            # set group flag for the sampler
+            self._set_group_flag()
+
+        # processing pipeline
+        self.pipeline = Compose(pipeline)
+
+    def __len__(self):
+        """Total number of samples of data."""
+        return len(self.data_infos)
+
+    def load_annotations(self, ann_file):
+        """Load annotation from annotation file."""
+        return mmcv.load(ann_file)
+
+    def load_proposals(self, proposal_file):
+        """Load proposal from proposal file."""
+        return mmcv.load(proposal_file)
+
+    def get_ann_info(self, idx):
+        """Get annotation by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Annotation info of specified index.
+        """
+
+        return self.data_infos[idx]['ann']
+
+    def get_cat_ids(self, idx):
+        """Get category ids by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            list[int]: All categories in the image of specified index.
+        """
+
+        return self.data_infos[idx]['ann']['labels'].astype(np.int).tolist()
+
+    def pre_pipeline(self, results):
+        """Prepare results dict for pipeline."""
+        results['img_prefix'] = self.img_prefix
+        results['seg_prefix'] = self.seg_prefix
+        results['proposal_file'] = self.proposal_file
+        results['bbox_fields'] = []
+        results['mask_fields'] = []
+        results['seg_fields'] = []
+
+    def _filter_imgs(self, min_size=32):
+        """Filter images too small."""
+        if self.filter_empty_gt:
+            warnings.warn(
+                'CustomDataset does not support filtering empty gt images.')
+        valid_inds = []
+        for i, img_info in enumerate(self.data_infos):
+            if min(img_info['width'], img_info['height']) >= min_size:
+                valid_inds.append(i)
+        return valid_inds
+
+    def _set_group_flag(self):
+        """Set flag according to image aspect ratio.
+
+        Images with aspect ratio greater than 1 will be set as group 1,
+        otherwise group 0.
+        """
+        self.flag = np.zeros(len(self), dtype=np.uint8)
+        for i in range(len(self)):
+            img_info = self.data_infos[i]
+            if img_info['width'] / img_info['height'] > 1:
+                self.flag[i] = 1
+
+    def _rand_another(self, idx):
+        """Get another random index from the same group as the given index."""
+        pool = np.where(self.flag == self.flag[idx])[0]
+        return np.random.choice(pool)
+
+    def __getitem__(self, idx):
+        """Get training/test data after pipeline.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Training/test data (with annotation if `test_mode` is set \
+                True).
+        """
+
+        if self.test_mode:
+            return self.prepare_test_img(idx)
+        while True:
+            if self.support_mosaic:
+                num_samples = np.random.choice([1,4,9], p=[0.4,0.4,0.2])
+            else:
+                num_samples = 1
+            if num_samples==1:
+                data = self.prepare_train_img(idx)
+            else:
+                data = self.prepare_multiple_train_imgs(idx, num_samples)
+            if data is None:
+                idx = self._rand_another(idx)
+                continue
+            return data
+
+    def prepare_train_img(self, idx):
+        """Get training data and annotations after pipeline.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Training data and annotation after pipeline with new keys \
+                introduced by pipeline.
+        """
+
+        img_info = self.data_infos[idx]
+        ann_info = self.get_ann_info(idx)
+        results = dict(img_info=img_info, ann_info=ann_info)
+        if self.proposals is not None:
+            results['proposals'] = self.proposals[idx]
+        self.pre_pipeline(results)
+        return self.pipeline(results)
+
+    def prepare_multiple_train_imgs(self, idx, num_samples):
+        """Get training data and annotations (of multiple images) after
+        pipeline.
+        While the main sample is determined by idx, the other samples
+        are selected randomly.
+        Args:
+            idx (int): Index of the main sample.
+        Returns:
+            dict: Training data and annotation after pipeline with new keys \
+                introduced by pipeline.
+        """
+        num = num_samples - 1
+
+        n_samples = len(self)
+        total_indices = list(range(n_samples))
+        total_indices.remove(idx)
+
+        selected_indices = random.choice(total_indices, num, False)
+        selected_indices = np.insert(selected_indices, 0, idx)
+
+        results = []
+        for i in selected_indices:
+            img_info = self.data_infos[i]
+            ann_info = self.get_ann_info(i)
+            results_i = dict(img_info=img_info, ann_info=ann_info)
+            if self.proposals is not None:
+                results_i['proposals'] = self.proposals[i]
+            self.pre_pipeline(results_i)
+            results.append(results_i)
+        return self.pipeline(results)
+
+
+    def prepare_test_img(self, idx):
+        """Get testing data  after pipeline.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Testing data after pipeline with new keys intorduced by \
+                piepline.
+        """
+
+        img_info = self.data_infos[idx]
+        results = dict(img_info=img_info)
+        if self.proposals is not None:
+            results['proposals'] = self.proposals[idx]
+        self.pre_pipeline(results)
+        return self.pipeline(results)
+
+    @classmethod
+    def get_classes(cls, classes=None):
+        """Get class names of current dataset.
+
+        Args:
+            classes (Sequence[str] | str | None): If classes is None, use
+                default CLASSES defined by builtin dataset. If classes is a
+                string, take it as a file name. The file contains the name of
+                classes where each line contains one class name. If classes is
+                a tuple or list, override the CLASSES defined by the dataset.
+
+        Returns:
+            tuple[str] or list[str]: Names of categories of the dataset.
+        """
+        if classes is None:
+            return cls.CLASSES
+
+        if isinstance(classes, str):
+            # take it as a file path
+            class_names = mmcv.list_from_file(classes)
+        elif isinstance(classes, (tuple, list)):
+            class_names = classes
+        else:
+            raise ValueError(f'Unsupported type {type(classes)} of classes.')
+
+        return class_names
+
+    def format_results(self, results, **kwargs):
+        """Place holder to format result to dataset specific output."""
+        pass
+
+    def evaluate(self,
+                 results,
+                 metric='mAP',
+                 logger=None,
+                 proposal_nums=(100, 300, 1000),
+                 iou_thr=0.5,
+                 scale_ranges=None):
+        """Evaluate the dataset.
+
+        Args:
+            results (list): Testing results of the dataset.
+            metric (str | list[str]): Metrics to be evaluated.
+            logger (logging.Logger | None | str): Logger used for printing
+                related information during evaluation. Default: None.
+            proposal_nums (Sequence[int]): Proposal number used for evaluating
+                recalls, such as recall@100, recall@1000.
+                Default: (100, 300, 1000).
+            iou_thr (float | list[float]): IoU threshold. It must be a float
+                when evaluating mAP, and can be a list when evaluating recall.
+                Default: 0.5.
+            scale_ranges (list[tuple] | None): Scale ranges for evaluating mAP.
+                Default: None.
+        """
+
+        if not isinstance(metric, str):
+            assert len(metric) == 1
+            metric = metric[0]
+        allowed_metrics = ['mAP', 'recall']
+        if metric not in allowed_metrics:
+            raise KeyError(f'metric {metric} is not supported')
+        annotations = [self.get_ann_info(i) for i in range(len(self))]
+        eval_results = OrderedDict()
+        if metric == 'mAP':
+            assert isinstance(iou_thr, float)
+            mean_ap, _ = eval_map(
+                results,
+                annotations,
+                scale_ranges=scale_ranges,
+                iou_thr=iou_thr,
+                dataset=self.CLASSES,
+                logger=logger)
+            eval_results['mAP'] = mean_ap
+        elif metric == 'recall':
+            gt_bboxes = [ann['bboxes'] for ann in annotations]
+            if isinstance(iou_thr, float):
+                iou_thr = [iou_thr]
+            recalls = eval_recalls(
+                gt_bboxes, results, proposal_nums, iou_thr, logger=logger)
+            for i, num in enumerate(proposal_nums):
+                for j, iou in enumerate(iou_thr):
+                    eval_results[f'recall@{num}@{iou}'] = recalls[i, j]
+            if recalls.shape[1] > 1:
+                ar = recalls.mean(axis=1)
+                for i, num in enumerate(proposal_nums):
+                    eval_results[f'AR@{num}'] = ar[i]
+        return eval_results
diff --git a/detection/scrfd/mmdet/datasets/dataset_wrappers.py b/detection/scrfd/mmdet/datasets/dataset_wrappers.py
new file mode 100755
index 000000000..55ad5cb60
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/dataset_wrappers.py
@@ -0,0 +1,282 @@
+import bisect
+import math
+from collections import defaultdict
+
+import numpy as np
+from mmcv.utils import print_log
+from torch.utils.data.dataset import ConcatDataset as _ConcatDataset
+
+from .builder import DATASETS
+from .coco import CocoDataset
+
+
+@DATASETS.register_module()
+class ConcatDataset(_ConcatDataset):
+    """A wrapper of concatenated dataset.
+
+    Same as :obj:`torch.utils.data.dataset.ConcatDataset`, but
+    concat the group flag for image aspect ratio.
+
+    Args:
+        datasets (list[:obj:`Dataset`]): A list of datasets.
+        separate_eval (bool): Whether to evaluate the results
+            separately if it is used as validation dataset.
+            Defaults to True.
+    """
+
+    def __init__(self, datasets, separate_eval=True):
+        super(ConcatDataset, self).__init__(datasets)
+        self.CLASSES = datasets[0].CLASSES
+        self.separate_eval = separate_eval
+        if not separate_eval:
+            if any([isinstance(ds, CocoDataset) for ds in datasets]):
+                raise NotImplementedError(
+                    'Evaluating concatenated CocoDataset as a whole is not'
+                    ' supported! Please set "separate_eval=True"')
+            elif len(set([type(ds) for ds in datasets])) != 1:
+                raise NotImplementedError(
+                    'All the datasets should have same types')
+
+        if hasattr(datasets[0], 'flag'):
+            flags = []
+            for i in range(0, len(datasets)):
+                flags.append(datasets[i].flag)
+            self.flag = np.concatenate(flags)
+
+    def get_cat_ids(self, idx):
+        """Get category ids of concatenated dataset by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            list[int]: All categories in the image of specified index.
+        """
+
+        if idx < 0:
+            if -idx > len(self):
+                raise ValueError(
+                    'absolute value of index should not exceed dataset length')
+            idx = len(self) + idx
+        dataset_idx = bisect.bisect_right(self.cumulative_sizes, idx)
+        if dataset_idx == 0:
+            sample_idx = idx
+        else:
+            sample_idx = idx - self.cumulative_sizes[dataset_idx - 1]
+        return self.datasets[dataset_idx].get_cat_ids(sample_idx)
+
+    def evaluate(self, results, logger=None, **kwargs):
+        """Evaluate the results.
+
+        Args:
+            results (list[list | tuple]): Testing results of the dataset.
+            logger (logging.Logger | str | None): Logger used for printing
+                related information during evaluation. Default: None.
+
+        Returns:
+            dict[str: float]: AP results of the total dataset or each separate
+            dataset if `self.separate_eval=True`.
+        """
+        assert len(results) == self.cumulative_sizes[-1], \
+            ('Dataset and results have different sizes: '
+             f'{self.cumulative_sizes[-1]} v.s. {len(results)}')
+
+        # Check whether all the datasets support evaluation
+        for dataset in self.datasets:
+            assert hasattr(dataset, 'evaluate'), \
+                    f'{type(dataset)} does not implement evaluate function'
+
+        if self.separate_eval:
+            dataset_idx = -1
+            total_eval_results = dict()
+            for size, dataset in zip(self.cumulative_sizes, self.datasets):
+                start_idx = 0 if dataset_idx == -1 else \
+                    self.cumulative_sizes[dataset_idx]
+                end_idx = self.cumulative_sizes[dataset_idx + 1]
+
+                results_per_dataset = results[start_idx:end_idx]
+                print_log(
+                    f'\nEvaluateing {dataset.ann_file} with '
+                    f'{len(results_per_dataset)} images now',
+                    logger=logger)
+
+                eval_results_per_dataset = dataset.evaluate(
+                    results_per_dataset, logger=logger, **kwargs)
+                dataset_idx += 1
+                for k, v in eval_results_per_dataset.items():
+                    total_eval_results.update({f'{dataset_idx}_{k}': v})
+
+            return total_eval_results
+        elif any([isinstance(ds, CocoDataset) for ds in self.datasets]):
+            raise NotImplementedError(
+                'Evaluating concatenated CocoDataset as a whole is not'
+                ' supported! Please set "separate_eval=True"')
+        elif len(set([type(ds) for ds in self.datasets])) != 1:
+            raise NotImplementedError(
+                'All the datasets should have same types')
+        else:
+            original_data_infos = self.datasets[0].data_infos
+            self.datasets[0].data_infos = sum(
+                [dataset.data_infos for dataset in self.datasets], [])
+            eval_results = self.datasets[0].evaluate(
+                results, logger=logger, **kwargs)
+            self.datasets[0].data_infos = original_data_infos
+            return eval_results
+
+
+@DATASETS.register_module()
+class RepeatDataset(object):
+    """A wrapper of repeated dataset.
+
+    The length of repeated dataset will be `times` larger than the original
+    dataset. This is useful when the data loading time is long but the dataset
+    is small. Using RepeatDataset can reduce the data loading time between
+    epochs.
+
+    Args:
+        dataset (:obj:`Dataset`): The dataset to be repeated.
+        times (int): Repeat times.
+    """
+
+    def __init__(self, dataset, times):
+        self.dataset = dataset
+        self.times = times
+        self.CLASSES = dataset.CLASSES
+        if hasattr(self.dataset, 'flag'):
+            self.flag = np.tile(self.dataset.flag, times)
+
+        self._ori_len = len(self.dataset)
+
+    def __getitem__(self, idx):
+        return self.dataset[idx % self._ori_len]
+
+    def get_cat_ids(self, idx):
+        """Get category ids of repeat dataset by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            list[int]: All categories in the image of specified index.
+        """
+
+        return self.dataset.get_cat_ids(idx % self._ori_len)
+
+    def __len__(self):
+        """Length after repetition."""
+        return self.times * self._ori_len
+
+
+# Modified from https://github.com/facebookresearch/detectron2/blob/41d475b75a230221e21d9cac5d69655e3415e3a4/detectron2/data/samplers/distributed_sampler.py#L57 # noqa
+@DATASETS.register_module()
+class ClassBalancedDataset(object):
+    """A wrapper of repeated dataset with repeat factor.
+
+    Suitable for training on class imbalanced datasets like LVIS. Following
+    the sampling strategy in the `paper <https://arxiv.org/abs/1908.03195>`_,
+    in each epoch, an image may appear multiple times based on its
+    "repeat factor".
+    The repeat factor for an image is a function of the frequency the rarest
+    category labeled in that image. The "frequency of category c" in [0, 1]
+    is defined by the fraction of images in the training set (without repeats)
+    in which category c appears.
+    The dataset needs to instantiate :func:`self.get_cat_ids` to support
+    ClassBalancedDataset.
+
+    The repeat factor is computed as followed.
+
+    1. For each category c, compute the fraction # of images
+       that contain it: :math:`f(c)`
+    2. For each category c, compute the category-level repeat factor:
+       :math:`r(c) = max(1, sqrt(t/f(c)))`
+    3. For each image I, compute the image-level repeat factor:
+       :math:`r(I) = max_{c in I} r(c)`
+
+    Args:
+        dataset (:obj:`CustomDataset`): The dataset to be repeated.
+        oversample_thr (float): frequency threshold below which data is
+            repeated. For categories with ``f_c >= oversample_thr``, there is
+            no oversampling. For categories with ``f_c < oversample_thr``, the
+            degree of oversampling following the square-root inverse frequency
+            heuristic above.
+        filter_empty_gt (bool, optional): If set true, images without bounding
+            boxes will not be oversampled. Otherwise, they will be categorized
+            as the pure background class and involved into the oversampling.
+            Default: True.
+    """
+
+    def __init__(self, dataset, oversample_thr, filter_empty_gt=True):
+        self.dataset = dataset
+        self.oversample_thr = oversample_thr
+        self.filter_empty_gt = filter_empty_gt
+        self.CLASSES = dataset.CLASSES
+
+        repeat_factors = self._get_repeat_factors(dataset, oversample_thr)
+        repeat_indices = []
+        for dataset_idx, repeat_factor in enumerate(repeat_factors):
+            repeat_indices.extend([dataset_idx] * math.ceil(repeat_factor))
+        self.repeat_indices = repeat_indices
+
+        flags = []
+        if hasattr(self.dataset, 'flag'):
+            for flag, repeat_factor in zip(self.dataset.flag, repeat_factors):
+                flags.extend([flag] * int(math.ceil(repeat_factor)))
+            assert len(flags) == len(repeat_indices)
+        self.flag = np.asarray(flags, dtype=np.uint8)
+
+    def _get_repeat_factors(self, dataset, repeat_thr):
+        """Get repeat factor for each images in the dataset.
+
+        Args:
+            dataset (:obj:`CustomDataset`): The dataset
+            repeat_thr (float): The threshold of frequency. If an image
+                contains the categories whose frequency below the threshold,
+                it would be repeated.
+
+        Returns:
+            list[float]: The repeat factors for each images in the dataset.
+        """
+
+        # 1. For each category c, compute the fraction # of images
+        #   that contain it: f(c)
+        category_freq = defaultdict(int)
+        num_images = len(dataset)
+        for idx in range(num_images):
+            cat_ids = set(self.dataset.get_cat_ids(idx))
+            if len(cat_ids) == 0 and not self.filter_empty_gt:
+                cat_ids = set([len(self.CLASSES)])
+            for cat_id in cat_ids:
+                category_freq[cat_id] += 1
+        for k, v in category_freq.items():
+            category_freq[k] = v / num_images
+
+        # 2. For each category c, compute the category-level repeat factor:
+        #    r(c) = max(1, sqrt(t/f(c)))
+        category_repeat = {
+            cat_id: max(1.0, math.sqrt(repeat_thr / cat_freq))
+            for cat_id, cat_freq in category_freq.items()
+        }
+
+        # 3. For each image I, compute the image-level repeat factor:
+        #    r(I) = max_{c in I} r(c)
+        repeat_factors = []
+        for idx in range(num_images):
+            cat_ids = set(self.dataset.get_cat_ids(idx))
+            if len(cat_ids) == 0 and not self.filter_empty_gt:
+                cat_ids = set([len(self.CLASSES)])
+            repeat_factor = 1
+            if len(cat_ids) > 0:
+                repeat_factor = max(
+                    {category_repeat[cat_id]
+                     for cat_id in cat_ids})
+            repeat_factors.append(repeat_factor)
+
+        return repeat_factors
+
+    def __getitem__(self, idx):
+        ori_index = self.repeat_indices[idx]
+        return self.dataset[ori_index]
+
+    def __len__(self):
+        """Length after repetition."""
+        return len(self.repeat_indices)
diff --git a/detection/scrfd/mmdet/datasets/deepfashion.py b/detection/scrfd/mmdet/datasets/deepfashion.py
new file mode 100755
index 000000000..112537609
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/deepfashion.py
@@ -0,0 +1,10 @@
+from .builder import DATASETS
+from .coco import CocoDataset
+
+
+@DATASETS.register_module()
+class DeepFashionDataset(CocoDataset):
+
+    CLASSES = ('top', 'skirt', 'leggings', 'dress', 'outer', 'pants', 'bag',
+               'neckwear', 'headwear', 'eyeglass', 'belt', 'footwear', 'hair',
+               'skin', 'face')
diff --git a/detection/scrfd/mmdet/datasets/lvis.py b/detection/scrfd/mmdet/datasets/lvis.py
new file mode 100755
index 000000000..9f3eba066
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/lvis.py
@@ -0,0 +1,744 @@
+import itertools
+import logging
+import os.path as osp
+import tempfile
+from collections import OrderedDict
+
+import numpy as np
+from mmcv.utils import print_log
+from terminaltables import AsciiTable
+
+from .builder import DATASETS
+from .coco import CocoDataset
+
+
+@DATASETS.register_module()
+class LVISV05Dataset(CocoDataset):
+
+    CLASSES = (
+        'acorn', 'aerosol_can', 'air_conditioner', 'airplane', 'alarm_clock',
+        'alcohol', 'alligator', 'almond', 'ambulance', 'amplifier', 'anklet',
+        'antenna', 'apple', 'apple_juice', 'applesauce', 'apricot', 'apron',
+        'aquarium', 'armband', 'armchair', 'armoire', 'armor', 'artichoke',
+        'trash_can', 'ashtray', 'asparagus', 'atomizer', 'avocado', 'award',
+        'awning', 'ax', 'baby_buggy', 'basketball_backboard', 'backpack',
+        'handbag', 'suitcase', 'bagel', 'bagpipe', 'baguet', 'bait', 'ball',
+        'ballet_skirt', 'balloon', 'bamboo', 'banana', 'Band_Aid', 'bandage',
+        'bandanna', 'banjo', 'banner', 'barbell', 'barge', 'barrel',
+        'barrette', 'barrow', 'baseball_base', 'baseball', 'baseball_bat',
+        'baseball_cap', 'baseball_glove', 'basket', 'basketball_hoop',
+        'basketball', 'bass_horn', 'bat_(animal)', 'bath_mat', 'bath_towel',
+        'bathrobe', 'bathtub', 'batter_(food)', 'battery', 'beachball', 'bead',
+        'beaker', 'bean_curd', 'beanbag', 'beanie', 'bear', 'bed',
+        'bedspread', 'cow', 'beef_(food)', 'beeper', 'beer_bottle', 'beer_can',
+        'beetle', 'bell', 'bell_pepper', 'belt', 'belt_buckle', 'bench',
+        'beret', 'bib', 'Bible', 'bicycle', 'visor', 'binder', 'binoculars',
+        'bird', 'birdfeeder', 'birdbath', 'birdcage', 'birdhouse',
+        'birthday_cake', 'birthday_card', 'biscuit_(bread)', 'pirate_flag',
+        'black_sheep', 'blackboard', 'blanket', 'blazer', 'blender', 'blimp',
+        'blinker', 'blueberry', 'boar', 'gameboard', 'boat', 'bobbin',
+        'bobby_pin', 'boiled_egg', 'bolo_tie', 'deadbolt', 'bolt', 'bonnet',
+        'book', 'book_bag', 'bookcase', 'booklet', 'bookmark',
+        'boom_microphone', 'boot', 'bottle', 'bottle_opener', 'bouquet',
+        'bow_(weapon)', 'bow_(decorative_ribbons)', 'bow-tie', 'bowl',
+        'pipe_bowl', 'bowler_hat', 'bowling_ball', 'bowling_pin',
+        'boxing_glove', 'suspenders', 'bracelet', 'brass_plaque', 'brassiere',
+        'bread-bin', 'breechcloth', 'bridal_gown', 'briefcase',
+        'bristle_brush', 'broccoli', 'broach', 'broom', 'brownie',
+        'brussels_sprouts', 'bubble_gum', 'bucket', 'horse_buggy', 'bull',
+        'bulldog', 'bulldozer', 'bullet_train', 'bulletin_board',
+        'bulletproof_vest', 'bullhorn', 'corned_beef', 'bun', 'bunk_bed',
+        'buoy', 'burrito', 'bus_(vehicle)', 'business_card', 'butcher_knife',
+        'butter', 'butterfly', 'button', 'cab_(taxi)', 'cabana', 'cabin_car',
+        'cabinet', 'locker', 'cake', 'calculator', 'calendar', 'calf',
+        'camcorder', 'camel', 'camera', 'camera_lens', 'camper_(vehicle)',
+        'can', 'can_opener', 'candelabrum', 'candle', 'candle_holder',
+        'candy_bar', 'candy_cane', 'walking_cane', 'canister', 'cannon',
+        'canoe', 'cantaloup', 'canteen', 'cap_(headwear)', 'bottle_cap',
+        'cape', 'cappuccino', 'car_(automobile)', 'railcar_(part_of_a_train)',
+        'elevator_car', 'car_battery', 'identity_card', 'card', 'cardigan',
+        'cargo_ship', 'carnation', 'horse_carriage', 'carrot', 'tote_bag',
+        'cart', 'carton', 'cash_register', 'casserole', 'cassette', 'cast',
+        'cat', 'cauliflower', 'caviar', 'cayenne_(spice)', 'CD_player',
+        'celery', 'cellular_telephone', 'chain_mail', 'chair', 'chaise_longue',
+        'champagne', 'chandelier', 'chap', 'checkbook', 'checkerboard',
+        'cherry', 'chessboard', 'chest_of_drawers_(furniture)',
+        'chicken_(animal)', 'chicken_wire', 'chickpea', 'Chihuahua',
+        'chili_(vegetable)', 'chime', 'chinaware', 'crisp_(potato_chip)',
+        'poker_chip', 'chocolate_bar', 'chocolate_cake', 'chocolate_milk',
+        'chocolate_mousse', 'choker', 'chopping_board', 'chopstick',
+        'Christmas_tree', 'slide', 'cider', 'cigar_box', 'cigarette',
+        'cigarette_case', 'cistern', 'clarinet', 'clasp', 'cleansing_agent',
+        'clementine', 'clip', 'clipboard', 'clock', 'clock_tower',
+        'clothes_hamper', 'clothespin', 'clutch_bag', 'coaster', 'coat',
+        'coat_hanger', 'coatrack', 'cock', 'coconut', 'coffee_filter',
+        'coffee_maker', 'coffee_table', 'coffeepot', 'coil', 'coin',
+        'colander', 'coleslaw', 'coloring_material', 'combination_lock',
+        'pacifier', 'comic_book', 'computer_keyboard', 'concrete_mixer',
+        'cone', 'control', 'convertible_(automobile)', 'sofa_bed', 'cookie',
+        'cookie_jar', 'cooking_utensil', 'cooler_(for_food)',
+        'cork_(bottle_plug)', 'corkboard', 'corkscrew', 'edible_corn',
+        'cornbread', 'cornet', 'cornice', 'cornmeal', 'corset',
+        'romaine_lettuce', 'costume', 'cougar', 'coverall', 'cowbell',
+        'cowboy_hat', 'crab_(animal)', 'cracker', 'crape', 'crate', 'crayon',
+        'cream_pitcher', 'credit_card', 'crescent_roll', 'crib', 'crock_pot',
+        'crossbar', 'crouton', 'crow', 'crown', 'crucifix', 'cruise_ship',
+        'police_cruiser', 'crumb', 'crutch', 'cub_(animal)', 'cube',
+        'cucumber', 'cufflink', 'cup', 'trophy_cup', 'cupcake', 'hair_curler',
+        'curling_iron', 'curtain', 'cushion', 'custard', 'cutting_tool',
+        'cylinder', 'cymbal', 'dachshund', 'dagger', 'dartboard',
+        'date_(fruit)', 'deck_chair', 'deer', 'dental_floss', 'desk',
+        'detergent', 'diaper', 'diary', 'die', 'dinghy', 'dining_table', 'tux',
+        'dish', 'dish_antenna', 'dishrag', 'dishtowel', 'dishwasher',
+        'dishwasher_detergent', 'diskette', 'dispenser', 'Dixie_cup', 'dog',
+        'dog_collar', 'doll', 'dollar', 'dolphin', 'domestic_ass', 'eye_mask',
+        'doorbell', 'doorknob', 'doormat', 'doughnut', 'dove', 'dragonfly',
+        'drawer', 'underdrawers', 'dress', 'dress_hat', 'dress_suit',
+        'dresser', 'drill', 'drinking_fountain', 'drone', 'dropper',
+        'drum_(musical_instrument)', 'drumstick', 'duck', 'duckling',
+        'duct_tape', 'duffel_bag', 'dumbbell', 'dumpster', 'dustpan',
+        'Dutch_oven', 'eagle', 'earphone', 'earplug', 'earring', 'easel',
+        'eclair', 'eel', 'egg', 'egg_roll', 'egg_yolk', 'eggbeater',
+        'eggplant', 'electric_chair', 'refrigerator', 'elephant', 'elk',
+        'envelope', 'eraser', 'escargot', 'eyepatch', 'falcon', 'fan',
+        'faucet', 'fedora', 'ferret', 'Ferris_wheel', 'ferry', 'fig_(fruit)',
+        'fighter_jet', 'figurine', 'file_cabinet', 'file_(tool)', 'fire_alarm',
+        'fire_engine', 'fire_extinguisher', 'fire_hose', 'fireplace',
+        'fireplug', 'fish', 'fish_(food)', 'fishbowl', 'fishing_boat',
+        'fishing_rod', 'flag', 'flagpole', 'flamingo', 'flannel', 'flash',
+        'flashlight', 'fleece', 'flip-flop_(sandal)', 'flipper_(footwear)',
+        'flower_arrangement', 'flute_glass', 'foal', 'folding_chair',
+        'food_processor', 'football_(American)', 'football_helmet',
+        'footstool', 'fork', 'forklift', 'freight_car', 'French_toast',
+        'freshener', 'frisbee', 'frog', 'fruit_juice', 'fruit_salad',
+        'frying_pan', 'fudge', 'funnel', 'futon', 'gag', 'garbage',
+        'garbage_truck', 'garden_hose', 'gargle', 'gargoyle', 'garlic',
+        'gasmask', 'gazelle', 'gelatin', 'gemstone', 'giant_panda',
+        'gift_wrap', 'ginger', 'giraffe', 'cincture',
+        'glass_(drink_container)', 'globe', 'glove', 'goat', 'goggles',
+        'goldfish', 'golf_club', 'golfcart', 'gondola_(boat)', 'goose',
+        'gorilla', 'gourd', 'surgical_gown', 'grape', 'grasshopper', 'grater',
+        'gravestone', 'gravy_boat', 'green_bean', 'green_onion', 'griddle',
+        'grillroom', 'grinder_(tool)', 'grits', 'grizzly', 'grocery_bag',
+        'guacamole', 'guitar', 'gull', 'gun', 'hair_spray', 'hairbrush',
+        'hairnet', 'hairpin', 'ham', 'hamburger', 'hammer', 'hammock',
+        'hamper', 'hamster', 'hair_dryer', 'hand_glass', 'hand_towel',
+        'handcart', 'handcuff', 'handkerchief', 'handle', 'handsaw',
+        'hardback_book', 'harmonium', 'hat', 'hatbox', 'hatch', 'veil',
+        'headband', 'headboard', 'headlight', 'headscarf', 'headset',
+        'headstall_(for_horses)', 'hearing_aid', 'heart', 'heater',
+        'helicopter', 'helmet', 'heron', 'highchair', 'hinge', 'hippopotamus',
+        'hockey_stick', 'hog', 'home_plate_(baseball)', 'honey', 'fume_hood',
+        'hook', 'horse', 'hose', 'hot-air_balloon', 'hotplate', 'hot_sauce',
+        'hourglass', 'houseboat', 'hummingbird', 'hummus', 'polar_bear',
+        'icecream', 'popsicle', 'ice_maker', 'ice_pack', 'ice_skate',
+        'ice_tea', 'igniter', 'incense', 'inhaler', 'iPod',
+        'iron_(for_clothing)', 'ironing_board', 'jacket', 'jam', 'jean',
+        'jeep', 'jelly_bean', 'jersey', 'jet_plane', 'jewelry', 'joystick',
+        'jumpsuit', 'kayak', 'keg', 'kennel', 'kettle', 'key', 'keycard',
+        'kilt', 'kimono', 'kitchen_sink', 'kitchen_table', 'kite', 'kitten',
+        'kiwi_fruit', 'knee_pad', 'knife', 'knight_(chess_piece)',
+        'knitting_needle', 'knob', 'knocker_(on_a_door)', 'koala', 'lab_coat',
+        'ladder', 'ladle', 'ladybug', 'lamb_(animal)', 'lamb-chop', 'lamp',
+        'lamppost', 'lampshade', 'lantern', 'lanyard', 'laptop_computer',
+        'lasagna', 'latch', 'lawn_mower', 'leather', 'legging_(clothing)',
+        'Lego', 'lemon', 'lemonade', 'lettuce', 'license_plate', 'life_buoy',
+        'life_jacket', 'lightbulb', 'lightning_rod', 'lime', 'limousine',
+        'linen_paper', 'lion', 'lip_balm', 'lipstick', 'liquor', 'lizard',
+        'Loafer_(type_of_shoe)', 'log', 'lollipop', 'lotion',
+        'speaker_(stero_equipment)', 'loveseat', 'machine_gun', 'magazine',
+        'magnet', 'mail_slot', 'mailbox_(at_home)', 'mallet', 'mammoth',
+        'mandarin_orange', 'manger', 'manhole', 'map', 'marker', 'martini',
+        'mascot', 'mashed_potato', 'masher', 'mask', 'mast',
+        'mat_(gym_equipment)', 'matchbox', 'mattress', 'measuring_cup',
+        'measuring_stick', 'meatball', 'medicine', 'melon', 'microphone',
+        'microscope', 'microwave_oven', 'milestone', 'milk', 'minivan',
+        'mint_candy', 'mirror', 'mitten', 'mixer_(kitchen_tool)', 'money',
+        'monitor_(computer_equipment) computer_monitor', 'monkey', 'motor',
+        'motor_scooter', 'motor_vehicle', 'motorboat', 'motorcycle',
+        'mound_(baseball)', 'mouse_(animal_rodent)',
+        'mouse_(computer_equipment)', 'mousepad', 'muffin', 'mug', 'mushroom',
+        'music_stool', 'musical_instrument', 'nailfile', 'nameplate', 'napkin',
+        'neckerchief', 'necklace', 'necktie', 'needle', 'nest', 'newsstand',
+        'nightshirt', 'nosebag_(for_animals)', 'noseband_(for_animals)',
+        'notebook', 'notepad', 'nut', 'nutcracker', 'oar', 'octopus_(food)',
+        'octopus_(animal)', 'oil_lamp', 'olive_oil', 'omelet', 'onion',
+        'orange_(fruit)', 'orange_juice', 'oregano', 'ostrich', 'ottoman',
+        'overalls_(clothing)', 'owl', 'packet', 'inkpad', 'pad', 'paddle',
+        'padlock', 'paintbox', 'paintbrush', 'painting', 'pajamas', 'palette',
+        'pan_(for_cooking)', 'pan_(metal_container)', 'pancake', 'pantyhose',
+        'papaya', 'paperclip', 'paper_plate', 'paper_towel', 'paperback_book',
+        'paperweight', 'parachute', 'parakeet', 'parasail_(sports)',
+        'parchment', 'parka', 'parking_meter', 'parrot',
+        'passenger_car_(part_of_a_train)', 'passenger_ship', 'passport',
+        'pastry', 'patty_(food)', 'pea_(food)', 'peach', 'peanut_butter',
+        'pear', 'peeler_(tool_for_fruit_and_vegetables)', 'pegboard',
+        'pelican', 'pen', 'pencil', 'pencil_box', 'pencil_sharpener',
+        'pendulum', 'penguin', 'pennant', 'penny_(coin)', 'pepper',
+        'pepper_mill', 'perfume', 'persimmon', 'baby', 'pet', 'petfood',
+        'pew_(church_bench)', 'phonebook', 'phonograph_record', 'piano',
+        'pickle', 'pickup_truck', 'pie', 'pigeon', 'piggy_bank', 'pillow',
+        'pin_(non_jewelry)', 'pineapple', 'pinecone', 'ping-pong_ball',
+        'pinwheel', 'tobacco_pipe', 'pipe', 'pistol', 'pita_(bread)',
+        'pitcher_(vessel_for_liquid)', 'pitchfork', 'pizza', 'place_mat',
+        'plate', 'platter', 'playing_card', 'playpen', 'pliers',
+        'plow_(farm_equipment)', 'pocket_watch', 'pocketknife',
+        'poker_(fire_stirring_tool)', 'pole', 'police_van', 'polo_shirt',
+        'poncho', 'pony', 'pool_table', 'pop_(soda)', 'portrait',
+        'postbox_(public)', 'postcard', 'poster', 'pot', 'flowerpot', 'potato',
+        'potholder', 'pottery', 'pouch', 'power_shovel', 'prawn', 'printer',
+        'projectile_(weapon)', 'projector', 'propeller', 'prune', 'pudding',
+        'puffer_(fish)', 'puffin', 'pug-dog', 'pumpkin', 'puncher', 'puppet',
+        'puppy', 'quesadilla', 'quiche', 'quilt', 'rabbit', 'race_car',
+        'racket', 'radar', 'radiator', 'radio_receiver', 'radish', 'raft',
+        'rag_doll', 'raincoat', 'ram_(animal)', 'raspberry', 'rat',
+        'razorblade', 'reamer_(juicer)', 'rearview_mirror', 'receipt',
+        'recliner', 'record_player', 'red_cabbage', 'reflector',
+        'remote_control', 'rhinoceros', 'rib_(food)', 'rifle', 'ring',
+        'river_boat', 'road_map', 'robe', 'rocking_chair', 'roller_skate',
+        'Rollerblade', 'rolling_pin', 'root_beer',
+        'router_(computer_equipment)', 'rubber_band', 'runner_(carpet)',
+        'plastic_bag', 'saddle_(on_an_animal)', 'saddle_blanket', 'saddlebag',
+        'safety_pin', 'sail', 'salad', 'salad_plate', 'salami',
+        'salmon_(fish)', 'salmon_(food)', 'salsa', 'saltshaker',
+        'sandal_(type_of_shoe)', 'sandwich', 'satchel', 'saucepan', 'saucer',
+        'sausage', 'sawhorse', 'saxophone', 'scale_(measuring_instrument)',
+        'scarecrow', 'scarf', 'school_bus', 'scissors', 'scoreboard',
+        'scrambled_eggs', 'scraper', 'scratcher', 'screwdriver',
+        'scrubbing_brush', 'sculpture', 'seabird', 'seahorse', 'seaplane',
+        'seashell', 'seedling', 'serving_dish', 'sewing_machine', 'shaker',
+        'shampoo', 'shark', 'sharpener', 'Sharpie', 'shaver_(electric)',
+        'shaving_cream', 'shawl', 'shears', 'sheep', 'shepherd_dog',
+        'sherbert', 'shield', 'shirt', 'shoe', 'shopping_bag', 'shopping_cart',
+        'short_pants', 'shot_glass', 'shoulder_bag', 'shovel', 'shower_head',
+        'shower_curtain', 'shredder_(for_paper)', 'sieve', 'signboard', 'silo',
+        'sink', 'skateboard', 'skewer', 'ski', 'ski_boot', 'ski_parka',
+        'ski_pole', 'skirt', 'sled', 'sleeping_bag', 'sling_(bandage)',
+        'slipper_(footwear)', 'smoothie', 'snake', 'snowboard', 'snowman',
+        'snowmobile', 'soap', 'soccer_ball', 'sock', 'soda_fountain',
+        'carbonated_water', 'sofa', 'softball', 'solar_array', 'sombrero',
+        'soup', 'soup_bowl', 'soupspoon', 'sour_cream', 'soya_milk',
+        'space_shuttle', 'sparkler_(fireworks)', 'spatula', 'spear',
+        'spectacles', 'spice_rack', 'spider', 'sponge', 'spoon', 'sportswear',
+        'spotlight', 'squirrel', 'stapler_(stapling_machine)', 'starfish',
+        'statue_(sculpture)', 'steak_(food)', 'steak_knife',
+        'steamer_(kitchen_appliance)', 'steering_wheel', 'stencil',
+        'stepladder', 'step_stool', 'stereo_(sound_system)', 'stew', 'stirrer',
+        'stirrup', 'stockings_(leg_wear)', 'stool', 'stop_sign', 'brake_light',
+        'stove', 'strainer', 'strap', 'straw_(for_drinking)', 'strawberry',
+        'street_sign', 'streetlight', 'string_cheese', 'stylus', 'subwoofer',
+        'sugar_bowl', 'sugarcane_(plant)', 'suit_(clothing)', 'sunflower',
+        'sunglasses', 'sunhat', 'sunscreen', 'surfboard', 'sushi', 'mop',
+        'sweat_pants', 'sweatband', 'sweater', 'sweatshirt', 'sweet_potato',
+        'swimsuit', 'sword', 'syringe', 'Tabasco_sauce', 'table-tennis_table',
+        'table', 'table_lamp', 'tablecloth', 'tachometer', 'taco', 'tag',
+        'taillight', 'tambourine', 'army_tank', 'tank_(storage_vessel)',
+        'tank_top_(clothing)', 'tape_(sticky_cloth_or_paper)', 'tape_measure',
+        'tapestry', 'tarp', 'tartan', 'tassel', 'tea_bag', 'teacup',
+        'teakettle', 'teapot', 'teddy_bear', 'telephone', 'telephone_booth',
+        'telephone_pole', 'telephoto_lens', 'television_camera',
+        'television_set', 'tennis_ball', 'tennis_racket', 'tequila',
+        'thermometer', 'thermos_bottle', 'thermostat', 'thimble', 'thread',
+        'thumbtack', 'tiara', 'tiger', 'tights_(clothing)', 'timer', 'tinfoil',
+        'tinsel', 'tissue_paper', 'toast_(food)', 'toaster', 'toaster_oven',
+        'toilet', 'toilet_tissue', 'tomato', 'tongs', 'toolbox', 'toothbrush',
+        'toothpaste', 'toothpick', 'cover', 'tortilla', 'tow_truck', 'towel',
+        'towel_rack', 'toy', 'tractor_(farm_equipment)', 'traffic_light',
+        'dirt_bike', 'trailer_truck', 'train_(railroad_vehicle)', 'trampoline',
+        'tray', 'tree_house', 'trench_coat', 'triangle_(musical_instrument)',
+        'tricycle', 'tripod', 'trousers', 'truck', 'truffle_(chocolate)',
+        'trunk', 'vat', 'turban', 'turkey_(bird)', 'turkey_(food)', 'turnip',
+        'turtle', 'turtleneck_(clothing)', 'typewriter', 'umbrella',
+        'underwear', 'unicycle', 'urinal', 'urn', 'vacuum_cleaner', 'valve',
+        'vase', 'vending_machine', 'vent', 'videotape', 'vinegar', 'violin',
+        'vodka', 'volleyball', 'vulture', 'waffle', 'waffle_iron', 'wagon',
+        'wagon_wheel', 'walking_stick', 'wall_clock', 'wall_socket', 'wallet',
+        'walrus', 'wardrobe', 'wasabi', 'automatic_washer', 'watch',
+        'water_bottle', 'water_cooler', 'water_faucet', 'water_filter',
+        'water_heater', 'water_jug', 'water_gun', 'water_scooter', 'water_ski',
+        'water_tower', 'watering_can', 'watermelon', 'weathervane', 'webcam',
+        'wedding_cake', 'wedding_ring', 'wet_suit', 'wheel', 'wheelchair',
+        'whipped_cream', 'whiskey', 'whistle', 'wick', 'wig', 'wind_chime',
+        'windmill', 'window_box_(for_plants)', 'windshield_wiper', 'windsock',
+        'wine_bottle', 'wine_bucket', 'wineglass', 'wing_chair',
+        'blinder_(for_horses)', 'wok', 'wolf', 'wooden_spoon', 'wreath',
+        'wrench', 'wristband', 'wristlet', 'yacht', 'yak', 'yogurt',
+        'yoke_(animal_equipment)', 'zebra', 'zucchini')
+
+    def load_annotations(self, ann_file):
+        """Load annotation from lvis style annotation file.
+
+        Args:
+            ann_file (str): Path of annotation file.
+
+        Returns:
+            list[dict]: Annotation info from LVIS api.
+        """
+
+        try:
+            import lvis
+            assert lvis.__version__ >= '10.5.3'
+            from lvis import LVIS
+        except AssertionError:
+            raise AssertionError('Incompatible version of lvis is installed. '
+                                 'Run pip uninstall lvis first. Then run pip '
+                                 'install mmlvis to install open-mmlab forked '
+                                 'lvis. ')
+        except ImportError:
+            raise ImportError('Package lvis is not installed. Please run pip '
+                              'install mmlvis to install open-mmlab forked '
+                              'lvis.')
+        self.coco = LVIS(ann_file)
+        assert not self.custom_classes, 'LVIS custom classes is not supported'
+        self.cat_ids = self.coco.get_cat_ids()
+        self.cat2label = {cat_id: i for i, cat_id in enumerate(self.cat_ids)}
+        self.img_ids = self.coco.get_img_ids()
+        data_infos = []
+        for i in self.img_ids:
+            info = self.coco.load_imgs([i])[0]
+            if info['file_name'].startswith('COCO'):
+                # Convert form the COCO 2014 file naming convention of
+                # COCO_[train/val/test]2014_000000000000.jpg to the 2017
+                # naming convention of 000000000000.jpg
+                # (LVIS v1 will fix this naming issue)
+                info['filename'] = info['file_name'][-16:]
+            else:
+                info['filename'] = info['file_name']
+            data_infos.append(info)
+        return data_infos
+
+    def evaluate(self,
+                 results,
+                 metric='bbox',
+                 logger=None,
+                 jsonfile_prefix=None,
+                 classwise=False,
+                 proposal_nums=(100, 300, 1000),
+                 iou_thrs=np.arange(0.5, 0.96, 0.05)):
+        """Evaluation in LVIS protocol.
+
+        Args:
+            results (list[list | tuple]): Testing results of the dataset.
+            metric (str | list[str]): Metrics to be evaluated. Options are
+                'bbox', 'segm', 'proposal', 'proposal_fast'.
+            logger (logging.Logger | str | None): Logger used for printing
+                related information during evaluation. Default: None.
+            jsonfile_prefix (str | None):
+            classwise (bool): Whether to evaluating the AP for each class.
+            proposal_nums (Sequence[int]): Proposal number used for evaluating
+                recalls, such as recall@100, recall@1000.
+                Default: (100, 300, 1000).
+            iou_thrs (Sequence[float]): IoU threshold used for evaluating
+                recalls. If set to a list, the average recall of all IoUs will
+                also be computed. Default: 0.5.
+
+        Returns:
+            dict[str, float]: LVIS style metrics.
+        """
+
+        try:
+            import lvis
+            assert lvis.__version__ >= '10.5.3'
+            from lvis import LVISResults, LVISEval
+        except AssertionError:
+            raise AssertionError('Incompatible version of lvis is installed. '
+                                 'Run pip uninstall lvis first. Then run pip '
+                                 'install mmlvis to install open-mmlab forked '
+                                 'lvis. ')
+        except ImportError:
+            raise ImportError('Package lvis is not installed. Please run pip '
+                              'install mmlvis to install open-mmlab forked '
+                              'lvis.')
+        assert isinstance(results, list), 'results must be a list'
+        assert len(results) == len(self), (
+            'The length of results is not equal to the dataset len: {} != {}'.
+            format(len(results), len(self)))
+
+        metrics = metric if isinstance(metric, list) else [metric]
+        allowed_metrics = ['bbox', 'segm', 'proposal', 'proposal_fast']
+        for metric in metrics:
+            if metric not in allowed_metrics:
+                raise KeyError('metric {} is not supported'.format(metric))
+
+        if jsonfile_prefix is None:
+            tmp_dir = tempfile.TemporaryDirectory()
+            jsonfile_prefix = osp.join(tmp_dir.name, 'results')
+        else:
+            tmp_dir = None
+        result_files = self.results2json(results, jsonfile_prefix)
+
+        eval_results = OrderedDict()
+        # get original api
+        lvis_gt = self.coco
+        for metric in metrics:
+            msg = 'Evaluating {}...'.format(metric)
+            if logger is None:
+                msg = '\n' + msg
+            print_log(msg, logger=logger)
+
+            if metric == 'proposal_fast':
+                ar = self.fast_eval_recall(
+                    results, proposal_nums, iou_thrs, logger='silent')
+                log_msg = []
+                for i, num in enumerate(proposal_nums):
+                    eval_results['AR@{}'.format(num)] = ar[i]
+                    log_msg.append('\nAR@{}\t{:.4f}'.format(num, ar[i]))
+                log_msg = ''.join(log_msg)
+                print_log(log_msg, logger=logger)
+                continue
+
+            if metric not in result_files:
+                raise KeyError('{} is not in results'.format(metric))
+            try:
+                lvis_dt = LVISResults(lvis_gt, result_files[metric])
+            except IndexError:
+                print_log(
+                    'The testing results of the whole dataset is empty.',
+                    logger=logger,
+                    level=logging.ERROR)
+                break
+
+            iou_type = 'bbox' if metric == 'proposal' else metric
+            lvis_eval = LVISEval(lvis_gt, lvis_dt, iou_type)
+            lvis_eval.params.imgIds = self.img_ids
+            if metric == 'proposal':
+                lvis_eval.params.useCats = 0
+                lvis_eval.params.maxDets = list(proposal_nums)
+                lvis_eval.evaluate()
+                lvis_eval.accumulate()
+                lvis_eval.summarize()
+                for k, v in lvis_eval.get_results().items():
+                    if k.startswith('AR'):
+                        val = float('{:.3f}'.format(float(v)))
+                        eval_results[k] = val
+            else:
+                lvis_eval.evaluate()
+                lvis_eval.accumulate()
+                lvis_eval.summarize()
+                lvis_results = lvis_eval.get_results()
+                if classwise:  # Compute per-category AP
+                    # Compute per-category AP
+                    # from https://github.com/facebookresearch/detectron2/
+                    precisions = lvis_eval.eval['precision']
+                    # precision: (iou, recall, cls, area range, max dets)
+                    assert len(self.cat_ids) == precisions.shape[2]
+
+                    results_per_category = []
+                    for idx, catId in enumerate(self.cat_ids):
+                        # area range index 0: all area ranges
+                        # max dets index -1: typically 100 per image
+                        nm = self.coco.load_cats(catId)[0]
+                        precision = precisions[:, :, idx, 0, -1]
+                        precision = precision[precision > -1]
+                        if precision.size:
+                            ap = np.mean(precision)
+                        else:
+                            ap = float('nan')
+                        results_per_category.append(
+                            (f'{nm["name"]}', f'{float(ap):0.3f}'))
+
+                    num_columns = min(6, len(results_per_category) * 2)
+                    results_flatten = list(
+                        itertools.chain(*results_per_category))
+                    headers = ['category', 'AP'] * (num_columns // 2)
+                    results_2d = itertools.zip_longest(*[
+                        results_flatten[i::num_columns]
+                        for i in range(num_columns)
+                    ])
+                    table_data = [headers]
+                    table_data += [result for result in results_2d]
+                    table = AsciiTable(table_data)
+                    print_log('\n' + table.table, logger=logger)
+
+                for k, v in lvis_results.items():
+                    if k.startswith('AP'):
+                        key = '{}_{}'.format(metric, k)
+                        val = float('{:.3f}'.format(float(v)))
+                        eval_results[key] = val
+                ap_summary = ' '.join([
+                    '{}:{:.3f}'.format(k, float(v))
+                    for k, v in lvis_results.items() if k.startswith('AP')
+                ])
+                eval_results['{}_mAP_copypaste'.format(metric)] = ap_summary
+            lvis_eval.print_results()
+        if tmp_dir is not None:
+            tmp_dir.cleanup()
+        return eval_results
+
+
+LVISDataset = LVISV05Dataset
+DATASETS.register_module(name='LVISDataset', module=LVISDataset)
+
+
+@DATASETS.register_module()
+class LVISV1Dataset(LVISDataset):
+
+    CLASSES = (
+        'aerosol_can', 'air_conditioner', 'airplane', 'alarm_clock', 'alcohol',
+        'alligator', 'almond', 'ambulance', 'amplifier', 'anklet', 'antenna',
+        'apple', 'applesauce', 'apricot', 'apron', 'aquarium',
+        'arctic_(type_of_shoe)', 'armband', 'armchair', 'armoire', 'armor',
+        'artichoke', 'trash_can', 'ashtray', 'asparagus', 'atomizer',
+        'avocado', 'award', 'awning', 'ax', 'baboon', 'baby_buggy',
+        'basketball_backboard', 'backpack', 'handbag', 'suitcase', 'bagel',
+        'bagpipe', 'baguet', 'bait', 'ball', 'ballet_skirt', 'balloon',
+        'bamboo', 'banana', 'Band_Aid', 'bandage', 'bandanna', 'banjo',
+        'banner', 'barbell', 'barge', 'barrel', 'barrette', 'barrow',
+        'baseball_base', 'baseball', 'baseball_bat', 'baseball_cap',
+        'baseball_glove', 'basket', 'basketball', 'bass_horn', 'bat_(animal)',
+        'bath_mat', 'bath_towel', 'bathrobe', 'bathtub', 'batter_(food)',
+        'battery', 'beachball', 'bead', 'bean_curd', 'beanbag', 'beanie',
+        'bear', 'bed', 'bedpan', 'bedspread', 'cow', 'beef_(food)', 'beeper',
+        'beer_bottle', 'beer_can', 'beetle', 'bell', 'bell_pepper', 'belt',
+        'belt_buckle', 'bench', 'beret', 'bib', 'Bible', 'bicycle', 'visor',
+        'billboard', 'binder', 'binoculars', 'bird', 'birdfeeder', 'birdbath',
+        'birdcage', 'birdhouse', 'birthday_cake', 'birthday_card',
+        'pirate_flag', 'black_sheep', 'blackberry', 'blackboard', 'blanket',
+        'blazer', 'blender', 'blimp', 'blinker', 'blouse', 'blueberry',
+        'gameboard', 'boat', 'bob', 'bobbin', 'bobby_pin', 'boiled_egg',
+        'bolo_tie', 'deadbolt', 'bolt', 'bonnet', 'book', 'bookcase',
+        'booklet', 'bookmark', 'boom_microphone', 'boot', 'bottle',
+        'bottle_opener', 'bouquet', 'bow_(weapon)', 'bow_(decorative_ribbons)',
+        'bow-tie', 'bowl', 'pipe_bowl', 'bowler_hat', 'bowling_ball', 'box',
+        'boxing_glove', 'suspenders', 'bracelet', 'brass_plaque', 'brassiere',
+        'bread-bin', 'bread', 'breechcloth', 'bridal_gown', 'briefcase',
+        'broccoli', 'broach', 'broom', 'brownie', 'brussels_sprouts',
+        'bubble_gum', 'bucket', 'horse_buggy', 'bull', 'bulldog', 'bulldozer',
+        'bullet_train', 'bulletin_board', 'bulletproof_vest', 'bullhorn',
+        'bun', 'bunk_bed', 'buoy', 'burrito', 'bus_(vehicle)', 'business_card',
+        'butter', 'butterfly', 'button', 'cab_(taxi)', 'cabana', 'cabin_car',
+        'cabinet', 'locker', 'cake', 'calculator', 'calendar', 'calf',
+        'camcorder', 'camel', 'camera', 'camera_lens', 'camper_(vehicle)',
+        'can', 'can_opener', 'candle', 'candle_holder', 'candy_bar',
+        'candy_cane', 'walking_cane', 'canister', 'canoe', 'cantaloup',
+        'canteen', 'cap_(headwear)', 'bottle_cap', 'cape', 'cappuccino',
+        'car_(automobile)', 'railcar_(part_of_a_train)', 'elevator_car',
+        'car_battery', 'identity_card', 'card', 'cardigan', 'cargo_ship',
+        'carnation', 'horse_carriage', 'carrot', 'tote_bag', 'cart', 'carton',
+        'cash_register', 'casserole', 'cassette', 'cast', 'cat', 'cauliflower',
+        'cayenne_(spice)', 'CD_player', 'celery', 'cellular_telephone',
+        'chain_mail', 'chair', 'chaise_longue', 'chalice', 'chandelier',
+        'chap', 'checkbook', 'checkerboard', 'cherry', 'chessboard',
+        'chicken_(animal)', 'chickpea', 'chili_(vegetable)', 'chime',
+        'chinaware', 'crisp_(potato_chip)', 'poker_chip', 'chocolate_bar',
+        'chocolate_cake', 'chocolate_milk', 'chocolate_mousse', 'choker',
+        'chopping_board', 'chopstick', 'Christmas_tree', 'slide', 'cider',
+        'cigar_box', 'cigarette', 'cigarette_case', 'cistern', 'clarinet',
+        'clasp', 'cleansing_agent', 'cleat_(for_securing_rope)', 'clementine',
+        'clip', 'clipboard', 'clippers_(for_plants)', 'cloak', 'clock',
+        'clock_tower', 'clothes_hamper', 'clothespin', 'clutch_bag', 'coaster',
+        'coat', 'coat_hanger', 'coatrack', 'cock', 'cockroach',
+        'cocoa_(beverage)', 'coconut', 'coffee_maker', 'coffee_table',
+        'coffeepot', 'coil', 'coin', 'colander', 'coleslaw',
+        'coloring_material', 'combination_lock', 'pacifier', 'comic_book',
+        'compass', 'computer_keyboard', 'condiment', 'cone', 'control',
+        'convertible_(automobile)', 'sofa_bed', 'cooker', 'cookie',
+        'cooking_utensil', 'cooler_(for_food)', 'cork_(bottle_plug)',
+        'corkboard', 'corkscrew', 'edible_corn', 'cornbread', 'cornet',
+        'cornice', 'cornmeal', 'corset', 'costume', 'cougar', 'coverall',
+        'cowbell', 'cowboy_hat', 'crab_(animal)', 'crabmeat', 'cracker',
+        'crape', 'crate', 'crayon', 'cream_pitcher', 'crescent_roll', 'crib',
+        'crock_pot', 'crossbar', 'crouton', 'crow', 'crowbar', 'crown',
+        'crucifix', 'cruise_ship', 'police_cruiser', 'crumb', 'crutch',
+        'cub_(animal)', 'cube', 'cucumber', 'cufflink', 'cup', 'trophy_cup',
+        'cupboard', 'cupcake', 'hair_curler', 'curling_iron', 'curtain',
+        'cushion', 'cylinder', 'cymbal', 'dagger', 'dalmatian', 'dartboard',
+        'date_(fruit)', 'deck_chair', 'deer', 'dental_floss', 'desk',
+        'detergent', 'diaper', 'diary', 'die', 'dinghy', 'dining_table', 'tux',
+        'dish', 'dish_antenna', 'dishrag', 'dishtowel', 'dishwasher',
+        'dishwasher_detergent', 'dispenser', 'diving_board', 'Dixie_cup',
+        'dog', 'dog_collar', 'doll', 'dollar', 'dollhouse', 'dolphin',
+        'domestic_ass', 'doorknob', 'doormat', 'doughnut', 'dove', 'dragonfly',
+        'drawer', 'underdrawers', 'dress', 'dress_hat', 'dress_suit',
+        'dresser', 'drill', 'drone', 'dropper', 'drum_(musical_instrument)',
+        'drumstick', 'duck', 'duckling', 'duct_tape', 'duffel_bag', 'dumbbell',
+        'dumpster', 'dustpan', 'eagle', 'earphone', 'earplug', 'earring',
+        'easel', 'eclair', 'eel', 'egg', 'egg_roll', 'egg_yolk', 'eggbeater',
+        'eggplant', 'electric_chair', 'refrigerator', 'elephant', 'elk',
+        'envelope', 'eraser', 'escargot', 'eyepatch', 'falcon', 'fan',
+        'faucet', 'fedora', 'ferret', 'Ferris_wheel', 'ferry', 'fig_(fruit)',
+        'fighter_jet', 'figurine', 'file_cabinet', 'file_(tool)', 'fire_alarm',
+        'fire_engine', 'fire_extinguisher', 'fire_hose', 'fireplace',
+        'fireplug', 'first-aid_kit', 'fish', 'fish_(food)', 'fishbowl',
+        'fishing_rod', 'flag', 'flagpole', 'flamingo', 'flannel', 'flap',
+        'flash', 'flashlight', 'fleece', 'flip-flop_(sandal)',
+        'flipper_(footwear)', 'flower_arrangement', 'flute_glass', 'foal',
+        'folding_chair', 'food_processor', 'football_(American)',
+        'football_helmet', 'footstool', 'fork', 'forklift', 'freight_car',
+        'French_toast', 'freshener', 'frisbee', 'frog', 'fruit_juice',
+        'frying_pan', 'fudge', 'funnel', 'futon', 'gag', 'garbage',
+        'garbage_truck', 'garden_hose', 'gargle', 'gargoyle', 'garlic',
+        'gasmask', 'gazelle', 'gelatin', 'gemstone', 'generator',
+        'giant_panda', 'gift_wrap', 'ginger', 'giraffe', 'cincture',
+        'glass_(drink_container)', 'globe', 'glove', 'goat', 'goggles',
+        'goldfish', 'golf_club', 'golfcart', 'gondola_(boat)', 'goose',
+        'gorilla', 'gourd', 'grape', 'grater', 'gravestone', 'gravy_boat',
+        'green_bean', 'green_onion', 'griddle', 'grill', 'grits', 'grizzly',
+        'grocery_bag', 'guitar', 'gull', 'gun', 'hairbrush', 'hairnet',
+        'hairpin', 'halter_top', 'ham', 'hamburger', 'hammer', 'hammock',
+        'hamper', 'hamster', 'hair_dryer', 'hand_glass', 'hand_towel',
+        'handcart', 'handcuff', 'handkerchief', 'handle', 'handsaw',
+        'hardback_book', 'harmonium', 'hat', 'hatbox', 'veil', 'headband',
+        'headboard', 'headlight', 'headscarf', 'headset',
+        'headstall_(for_horses)', 'heart', 'heater', 'helicopter', 'helmet',
+        'heron', 'highchair', 'hinge', 'hippopotamus', 'hockey_stick', 'hog',
+        'home_plate_(baseball)', 'honey', 'fume_hood', 'hook', 'hookah',
+        'hornet', 'horse', 'hose', 'hot-air_balloon', 'hotplate', 'hot_sauce',
+        'hourglass', 'houseboat', 'hummingbird', 'hummus', 'polar_bear',
+        'icecream', 'popsicle', 'ice_maker', 'ice_pack', 'ice_skate',
+        'igniter', 'inhaler', 'iPod', 'iron_(for_clothing)', 'ironing_board',
+        'jacket', 'jam', 'jar', 'jean', 'jeep', 'jelly_bean', 'jersey',
+        'jet_plane', 'jewel', 'jewelry', 'joystick', 'jumpsuit', 'kayak',
+        'keg', 'kennel', 'kettle', 'key', 'keycard', 'kilt', 'kimono',
+        'kitchen_sink', 'kitchen_table', 'kite', 'kitten', 'kiwi_fruit',
+        'knee_pad', 'knife', 'knitting_needle', 'knob', 'knocker_(on_a_door)',
+        'koala', 'lab_coat', 'ladder', 'ladle', 'ladybug', 'lamb_(animal)',
+        'lamb-chop', 'lamp', 'lamppost', 'lampshade', 'lantern', 'lanyard',
+        'laptop_computer', 'lasagna', 'latch', 'lawn_mower', 'leather',
+        'legging_(clothing)', 'Lego', 'legume', 'lemon', 'lemonade', 'lettuce',
+        'license_plate', 'life_buoy', 'life_jacket', 'lightbulb',
+        'lightning_rod', 'lime', 'limousine', 'lion', 'lip_balm', 'liquor',
+        'lizard', 'log', 'lollipop', 'speaker_(stero_equipment)', 'loveseat',
+        'machine_gun', 'magazine', 'magnet', 'mail_slot', 'mailbox_(at_home)',
+        'mallard', 'mallet', 'mammoth', 'manatee', 'mandarin_orange', 'manger',
+        'manhole', 'map', 'marker', 'martini', 'mascot', 'mashed_potato',
+        'masher', 'mask', 'mast', 'mat_(gym_equipment)', 'matchbox',
+        'mattress', 'measuring_cup', 'measuring_stick', 'meatball', 'medicine',
+        'melon', 'microphone', 'microscope', 'microwave_oven', 'milestone',
+        'milk', 'milk_can', 'milkshake', 'minivan', 'mint_candy', 'mirror',
+        'mitten', 'mixer_(kitchen_tool)', 'money',
+        'monitor_(computer_equipment) computer_monitor', 'monkey', 'motor',
+        'motor_scooter', 'motor_vehicle', 'motorcycle', 'mound_(baseball)',
+        'mouse_(computer_equipment)', 'mousepad', 'muffin', 'mug', 'mushroom',
+        'music_stool', 'musical_instrument', 'nailfile', 'napkin',
+        'neckerchief', 'necklace', 'necktie', 'needle', 'nest', 'newspaper',
+        'newsstand', 'nightshirt', 'nosebag_(for_animals)',
+        'noseband_(for_animals)', 'notebook', 'notepad', 'nut', 'nutcracker',
+        'oar', 'octopus_(food)', 'octopus_(animal)', 'oil_lamp', 'olive_oil',
+        'omelet', 'onion', 'orange_(fruit)', 'orange_juice', 'ostrich',
+        'ottoman', 'oven', 'overalls_(clothing)', 'owl', 'packet', 'inkpad',
+        'pad', 'paddle', 'padlock', 'paintbrush', 'painting', 'pajamas',
+        'palette', 'pan_(for_cooking)', 'pan_(metal_container)', 'pancake',
+        'pantyhose', 'papaya', 'paper_plate', 'paper_towel', 'paperback_book',
+        'paperweight', 'parachute', 'parakeet', 'parasail_(sports)', 'parasol',
+        'parchment', 'parka', 'parking_meter', 'parrot',
+        'passenger_car_(part_of_a_train)', 'passenger_ship', 'passport',
+        'pastry', 'patty_(food)', 'pea_(food)', 'peach', 'peanut_butter',
+        'pear', 'peeler_(tool_for_fruit_and_vegetables)', 'wooden_leg',
+        'pegboard', 'pelican', 'pen', 'pencil', 'pencil_box',
+        'pencil_sharpener', 'pendulum', 'penguin', 'pennant', 'penny_(coin)',
+        'pepper', 'pepper_mill', 'perfume', 'persimmon', 'person', 'pet',
+        'pew_(church_bench)', 'phonebook', 'phonograph_record', 'piano',
+        'pickle', 'pickup_truck', 'pie', 'pigeon', 'piggy_bank', 'pillow',
+        'pin_(non_jewelry)', 'pineapple', 'pinecone', 'ping-pong_ball',
+        'pinwheel', 'tobacco_pipe', 'pipe', 'pistol', 'pita_(bread)',
+        'pitcher_(vessel_for_liquid)', 'pitchfork', 'pizza', 'place_mat',
+        'plate', 'platter', 'playpen', 'pliers', 'plow_(farm_equipment)',
+        'plume', 'pocket_watch', 'pocketknife', 'poker_(fire_stirring_tool)',
+        'pole', 'polo_shirt', 'poncho', 'pony', 'pool_table', 'pop_(soda)',
+        'postbox_(public)', 'postcard', 'poster', 'pot', 'flowerpot', 'potato',
+        'potholder', 'pottery', 'pouch', 'power_shovel', 'prawn', 'pretzel',
+        'printer', 'projectile_(weapon)', 'projector', 'propeller', 'prune',
+        'pudding', 'puffer_(fish)', 'puffin', 'pug-dog', 'pumpkin', 'puncher',
+        'puppet', 'puppy', 'quesadilla', 'quiche', 'quilt', 'rabbit',
+        'race_car', 'racket', 'radar', 'radiator', 'radio_receiver', 'radish',
+        'raft', 'rag_doll', 'raincoat', 'ram_(animal)', 'raspberry', 'rat',
+        'razorblade', 'reamer_(juicer)', 'rearview_mirror', 'receipt',
+        'recliner', 'record_player', 'reflector', 'remote_control',
+        'rhinoceros', 'rib_(food)', 'rifle', 'ring', 'river_boat', 'road_map',
+        'robe', 'rocking_chair', 'rodent', 'roller_skate', 'Rollerblade',
+        'rolling_pin', 'root_beer', 'router_(computer_equipment)',
+        'rubber_band', 'runner_(carpet)', 'plastic_bag',
+        'saddle_(on_an_animal)', 'saddle_blanket', 'saddlebag', 'safety_pin',
+        'sail', 'salad', 'salad_plate', 'salami', 'salmon_(fish)',
+        'salmon_(food)', 'salsa', 'saltshaker', 'sandal_(type_of_shoe)',
+        'sandwich', 'satchel', 'saucepan', 'saucer', 'sausage', 'sawhorse',
+        'saxophone', 'scale_(measuring_instrument)', 'scarecrow', 'scarf',
+        'school_bus', 'scissors', 'scoreboard', 'scraper', 'screwdriver',
+        'scrubbing_brush', 'sculpture', 'seabird', 'seahorse', 'seaplane',
+        'seashell', 'sewing_machine', 'shaker', 'shampoo', 'shark',
+        'sharpener', 'Sharpie', 'shaver_(electric)', 'shaving_cream', 'shawl',
+        'shears', 'sheep', 'shepherd_dog', 'sherbert', 'shield', 'shirt',
+        'shoe', 'shopping_bag', 'shopping_cart', 'short_pants', 'shot_glass',
+        'shoulder_bag', 'shovel', 'shower_head', 'shower_cap',
+        'shower_curtain', 'shredder_(for_paper)', 'signboard', 'silo', 'sink',
+        'skateboard', 'skewer', 'ski', 'ski_boot', 'ski_parka', 'ski_pole',
+        'skirt', 'skullcap', 'sled', 'sleeping_bag', 'sling_(bandage)',
+        'slipper_(footwear)', 'smoothie', 'snake', 'snowboard', 'snowman',
+        'snowmobile', 'soap', 'soccer_ball', 'sock', 'sofa', 'softball',
+        'solar_array', 'sombrero', 'soup', 'soup_bowl', 'soupspoon',
+        'sour_cream', 'soya_milk', 'space_shuttle', 'sparkler_(fireworks)',
+        'spatula', 'spear', 'spectacles', 'spice_rack', 'spider', 'crawfish',
+        'sponge', 'spoon', 'sportswear', 'spotlight', 'squid_(food)',
+        'squirrel', 'stagecoach', 'stapler_(stapling_machine)', 'starfish',
+        'statue_(sculpture)', 'steak_(food)', 'steak_knife', 'steering_wheel',
+        'stepladder', 'step_stool', 'stereo_(sound_system)', 'stew', 'stirrer',
+        'stirrup', 'stool', 'stop_sign', 'brake_light', 'stove', 'strainer',
+        'strap', 'straw_(for_drinking)', 'strawberry', 'street_sign',
+        'streetlight', 'string_cheese', 'stylus', 'subwoofer', 'sugar_bowl',
+        'sugarcane_(plant)', 'suit_(clothing)', 'sunflower', 'sunglasses',
+        'sunhat', 'surfboard', 'sushi', 'mop', 'sweat_pants', 'sweatband',
+        'sweater', 'sweatshirt', 'sweet_potato', 'swimsuit', 'sword',
+        'syringe', 'Tabasco_sauce', 'table-tennis_table', 'table',
+        'table_lamp', 'tablecloth', 'tachometer', 'taco', 'tag', 'taillight',
+        'tambourine', 'army_tank', 'tank_(storage_vessel)',
+        'tank_top_(clothing)', 'tape_(sticky_cloth_or_paper)', 'tape_measure',
+        'tapestry', 'tarp', 'tartan', 'tassel', 'tea_bag', 'teacup',
+        'teakettle', 'teapot', 'teddy_bear', 'telephone', 'telephone_booth',
+        'telephone_pole', 'telephoto_lens', 'television_camera',
+        'television_set', 'tennis_ball', 'tennis_racket', 'tequila',
+        'thermometer', 'thermos_bottle', 'thermostat', 'thimble', 'thread',
+        'thumbtack', 'tiara', 'tiger', 'tights_(clothing)', 'timer', 'tinfoil',
+        'tinsel', 'tissue_paper', 'toast_(food)', 'toaster', 'toaster_oven',
+        'toilet', 'toilet_tissue', 'tomato', 'tongs', 'toolbox', 'toothbrush',
+        'toothpaste', 'toothpick', 'cover', 'tortilla', 'tow_truck', 'towel',
+        'towel_rack', 'toy', 'tractor_(farm_equipment)', 'traffic_light',
+        'dirt_bike', 'trailer_truck', 'train_(railroad_vehicle)', 'trampoline',
+        'tray', 'trench_coat', 'triangle_(musical_instrument)', 'tricycle',
+        'tripod', 'trousers', 'truck', 'truffle_(chocolate)', 'trunk', 'vat',
+        'turban', 'turkey_(food)', 'turnip', 'turtle', 'turtleneck_(clothing)',
+        'typewriter', 'umbrella', 'underwear', 'unicycle', 'urinal', 'urn',
+        'vacuum_cleaner', 'vase', 'vending_machine', 'vent', 'vest',
+        'videotape', 'vinegar', 'violin', 'vodka', 'volleyball', 'vulture',
+        'waffle', 'waffle_iron', 'wagon', 'wagon_wheel', 'walking_stick',
+        'wall_clock', 'wall_socket', 'wallet', 'walrus', 'wardrobe',
+        'washbasin', 'automatic_washer', 'watch', 'water_bottle',
+        'water_cooler', 'water_faucet', 'water_heater', 'water_jug',
+        'water_gun', 'water_scooter', 'water_ski', 'water_tower',
+        'watering_can', 'watermelon', 'weathervane', 'webcam', 'wedding_cake',
+        'wedding_ring', 'wet_suit', 'wheel', 'wheelchair', 'whipped_cream',
+        'whistle', 'wig', 'wind_chime', 'windmill', 'window_box_(for_plants)',
+        'windshield_wiper', 'windsock', 'wine_bottle', 'wine_bucket',
+        'wineglass', 'blinder_(for_horses)', 'wok', 'wolf', 'wooden_spoon',
+        'wreath', 'wrench', 'wristband', 'wristlet', 'yacht', 'yogurt',
+        'yoke_(animal_equipment)', 'zebra', 'zucchini')
+
+    def load_annotations(self, ann_file):
+        try:
+            import lvis
+            assert lvis.__version__ >= '10.5.3'
+            from lvis import LVIS
+        except AssertionError:
+            raise AssertionError('Incompatible version of lvis is installed. '
+                                 'Run pip uninstall lvis first. Then run pip '
+                                 'install mmlvis to install open-mmlab forked '
+                                 'lvis. ')
+        except ImportError:
+            raise ImportError('Package lvis is not installed. Please run pip '
+                              'install mmlvis to install open-mmlab forked '
+                              'lvis.')
+        self.coco = LVIS(ann_file)
+        assert not self.custom_classes, 'LVIS custom classes is not supported'
+        self.cat_ids = self.coco.get_cat_ids()
+        self.cat2label = {cat_id: i for i, cat_id in enumerate(self.cat_ids)}
+        self.img_ids = self.coco.get_img_ids()
+        data_infos = []
+        for i in self.img_ids:
+            info = self.coco.load_imgs([i])[0]
+            # coco_url is used in LVISv1 instead of file_name
+            # e.g. http://images.cocodataset.org/train2017/000000391895.jpg
+            # train/val split in specified in url
+            info['filename'] = info['coco_url'].replace(
+                'http://images.cocodataset.org/', '')
+            data_infos.append(info)
+        return data_infos
diff --git a/detection/scrfd/mmdet/datasets/pipelines/__init__.py b/detection/scrfd/mmdet/datasets/pipelines/__init__.py
new file mode 100755
index 000000000..ccaabe4bc
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/__init__.py
@@ -0,0 +1,27 @@
+from .auto_augment import (AutoAugment, BrightnessTransform, ColorTransform,
+                           ContrastTransform, EqualizeTransform, Rotate, Shear,
+                           Translate)
+from .compose import Compose
+from .formating import (Collect, DefaultFormatBundle, ImageToTensor,
+                        ToDataContainer, ToTensor, Transpose, to_tensor)
+from .instaboost import InstaBoost
+from .loading import (LoadAnnotations, LoadImageFromFile, LoadImageFromWebcam,
+                      LoadMultiChannelImageFromFiles, LoadProposals)
+from .test_time_aug import MultiScaleFlipAug
+from .transforms import (Albu, CutOut, Expand, MinIoURandomCrop, Normalize,
+                         Pad, PhotoMetricDistortion, RandomCenterCropPad,
+                         RandomSquareCrop,
+                         RandomCrop, RandomFlip, Resize, SegRescale)
+
+__all__ = [
+    'Compose', 'to_tensor', 'ToTensor', 'ImageToTensor', 'ToDataContainer',
+    'Transpose', 'Collect', 'DefaultFormatBundle', 'LoadAnnotations',
+    'LoadImageFromFile', 'LoadImageFromWebcam',
+    'LoadMultiChannelImageFromFiles', 'LoadProposals', 'MultiScaleFlipAug',
+    'Resize', 'RandomFlip', 'Pad', 'RandomCrop', 'Normalize', 'SegRescale',
+    'RandomSquareCrop',
+    'MinIoURandomCrop', 'Expand', 'PhotoMetricDistortion', 'Albu',
+    'InstaBoost', 'RandomCenterCropPad', 'AutoAugment', 'CutOut', 'Shear',
+    'Rotate', 'ColorTransform', 'EqualizeTransform', 'BrightnessTransform',
+    'ContrastTransform', 'Translate'
+]
diff --git a/detection/scrfd/mmdet/datasets/pipelines/auto_augment.py b/detection/scrfd/mmdet/datasets/pipelines/auto_augment.py
new file mode 100755
index 000000000..0d225331a
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/auto_augment.py
@@ -0,0 +1,890 @@
+import copy
+
+import cv2
+import mmcv
+import numpy as np
+
+from ..builder import PIPELINES
+from .compose import Compose
+
+_MAX_LEVEL = 10
+
+
+def level_to_value(level, max_value):
+    """Map from level to values based on max_value."""
+    return (level / _MAX_LEVEL) * max_value
+
+
+def enhance_level_to_value(level, a=1.8, b=0.1):
+    """Map from level to values."""
+    return (level / _MAX_LEVEL) * a + b
+
+
+def random_negative(value, random_negative_prob):
+    """Randomly negate value based on random_negative_prob."""
+    return -value if np.random.rand() < random_negative_prob else value
+
+
+def bbox2fields():
+    """The key correspondence from bboxes to labels, masks and
+    segmentations."""
+    bbox2label = {
+        'gt_bboxes': 'gt_labels',
+        'gt_bboxes_ignore': 'gt_labels_ignore'
+    }
+    bbox2mask = {
+        'gt_bboxes': 'gt_masks',
+        'gt_bboxes_ignore': 'gt_masks_ignore'
+    }
+    bbox2seg = {
+        'gt_bboxes': 'gt_semantic_seg',
+    }
+    return bbox2label, bbox2mask, bbox2seg
+
+
+@PIPELINES.register_module()
+class AutoAugment(object):
+    """Auto augmentation.
+
+    This data augmentation is proposed in `Learning Data Augmentation
+    Strategies for Object Detection <https://arxiv.org/pdf/1906.11172>`_.
+
+    TODO: Implement 'Shear', 'Sharpness' and 'Rotate' transforms
+
+    Args:
+        policies (list[list[dict]]): The policies of auto augmentation. Each
+            policy in ``policies`` is a specific augmentation policy, and is
+            composed by several augmentations (dict). When AutoAugment is
+            called, a random policy in ``policies`` will be selected to
+            augment images.
+
+    Examples:
+        >>> replace = (104, 116, 124)
+        >>> policies = [
+        >>>     [
+        >>>         dict(type='Sharpness', prob=0.0, level=8),
+        >>>         dict(
+        >>>             type='Shear',
+        >>>             prob=0.4,
+        >>>             level=0,
+        >>>             replace=replace,
+        >>>             axis='x')
+        >>>     ],
+        >>>     [
+        >>>         dict(
+        >>>             type='Rotate',
+        >>>             prob=0.6,
+        >>>             level=10,
+        >>>             replace=replace),
+        >>>         dict(type='Color', prob=1.0, level=6)
+        >>>     ]
+        >>> ]
+        >>> augmentation = AutoAugment(policies)
+        >>> img = np.ones(100, 100, 3)
+        >>> gt_bboxes = np.ones(10, 4)
+        >>> results = dict(img=img, gt_bboxes=gt_bboxes)
+        >>> results = augmentation(results)
+    """
+
+    def __init__(self, policies):
+        assert isinstance(policies, list) and len(policies) > 0, \
+            'Policies must be a non-empty list.'
+        for policy in policies:
+            assert isinstance(policy, list) and len(policy) > 0, \
+                'Each policy in policies must be a non-empty list.'
+            for augment in policy:
+                assert isinstance(augment, dict) and 'type' in augment, \
+                    'Each specific augmentation must be a dict with key' \
+                    ' "type".'
+
+        self.policies = copy.deepcopy(policies)
+        self.transforms = [Compose(policy) for policy in self.policies]
+
+    def __call__(self, results):
+        transform = np.random.choice(self.transforms)
+        return transform(results)
+
+    def __repr__(self):
+        return f'{self.__class__.__name__}(policies={self.policies})'
+
+
+@PIPELINES.register_module()
+class Shear(object):
+    """Apply Shear Transformation to image (and its corresponding bbox, mask,
+    segmentation).
+
+    Args:
+        level (int | float): The level should be in range [0,_MAX_LEVEL].
+        img_fill_val (int | float | tuple): The filled values for image border.
+            If float, the same fill value will be used for all the three
+            channels of image. If tuple, the should be 3 elements.
+        seg_ignore_label (int): The fill value used for segmentation map.
+            Note this value must equals ``ignore_label`` in ``semantic_head``
+            of the corresponding config. Default 255.
+        prob (float): The probability for performing Shear and should be in
+            range [0, 1].
+        direction (str): The direction for shear, either "horizontal"
+            or "vertical".
+        max_shear_magnitude (float): The maximum magnitude for Shear
+            transformation.
+        random_negative_prob (float): The probability that turns the
+                offset negative. Should be in range [0,1]
+        interpolation (str): Same as in :func:`mmcv.imshear`.
+    """
+
+    def __init__(self,
+                 level,
+                 img_fill_val=128,
+                 seg_ignore_label=255,
+                 prob=0.5,
+                 direction='horizontal',
+                 max_shear_magnitude=0.3,
+                 random_negative_prob=0.5,
+                 interpolation='bilinear'):
+        assert isinstance(level, (int, float)), 'The level must be type ' \
+            f'int or float, got {type(level)}.'
+        assert 0 <= level <= _MAX_LEVEL, 'The level should be in range ' \
+            f'[0,{_MAX_LEVEL}], got {level}.'
+        if isinstance(img_fill_val, (float, int)):
+            img_fill_val = tuple([float(img_fill_val)] * 3)
+        elif isinstance(img_fill_val, tuple):
+            assert len(img_fill_val) == 3, 'img_fill_val as tuple must ' \
+                f'have 3 elements. got {len(img_fill_val)}.'
+            img_fill_val = tuple([float(val) for val in img_fill_val])
+        else:
+            raise ValueError(
+                'img_fill_val must be float or tuple with 3 elements.')
+        assert np.all([0 <= val <= 255 for val in img_fill_val]), 'all ' \
+            'elements of img_fill_val should between range [0,255].' \
+            f'got {img_fill_val}.'
+        assert 0 <= prob <= 1.0, 'The probability of shear should be in ' \
+            f'range [0,1]. got {prob}.'
+        assert direction in ('horizontal', 'vertical'), 'direction must ' \
+            f'in be either "horizontal" or "vertical". got {direction}.'
+        assert isinstance(max_shear_magnitude, float), 'max_shear_magnitude ' \
+            f'should be type float. got {type(max_shear_magnitude)}.'
+        assert 0. <= max_shear_magnitude <= 1., 'Defaultly ' \
+            'max_shear_magnitude should be in range [0,1]. ' \
+            f'got {max_shear_magnitude}.'
+        self.level = level
+        self.magnitude = level_to_value(level, max_shear_magnitude)
+        self.img_fill_val = img_fill_val
+        self.seg_ignore_label = seg_ignore_label
+        self.prob = prob
+        self.direction = direction
+        self.max_shear_magnitude = max_shear_magnitude
+        self.random_negative_prob = random_negative_prob
+        self.interpolation = interpolation
+
+    def _shear_img(self,
+                   results,
+                   magnitude,
+                   direction='horizontal',
+                   interpolation='bilinear'):
+        """Shear the image.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+            magnitude (int | float): The magnitude used for shear.
+            direction (str): The direction for shear, either "horizontal"
+                or "vertical".
+            interpolation (str): Same as in :func:`mmcv.imshear`.
+        """
+        for key in results.get('img_fields', ['img']):
+            img = results[key]
+            img_sheared = mmcv.imshear(
+                img,
+                magnitude,
+                direction,
+                border_value=self.img_fill_val,
+                interpolation=interpolation)
+            results[key] = img_sheared.astype(img.dtype)
+
+    def _shear_bboxes(self, results, magnitude):
+        """Shear the bboxes."""
+        h, w, c = results['img_shape']
+        if self.direction == 'horizontal':
+            shear_matrix = np.stack([[1, magnitude],
+                                     [0, 1]]).astype(np.float32)  # [2, 2]
+        else:
+            shear_matrix = np.stack([[1, 0], [magnitude,
+                                              1]]).astype(np.float32)
+        for key in results.get('bbox_fields', []):
+            min_x, min_y, max_x, max_y = np.split(
+                results[key], results[key].shape[-1], axis=-1)
+            coordinates = np.stack([[min_x, min_y], [max_x, min_y],
+                                    [min_x, max_y],
+                                    [max_x, max_y]])  # [4, 2, nb_box, 1]
+            coordinates = coordinates[..., 0].transpose(
+                (2, 1, 0)).astype(np.float32)  # [nb_box, 2, 4]
+            new_coords = np.matmul(shear_matrix[None, :, :],
+                                   coordinates)  # [nb_box, 2, 4]
+            min_x = np.min(new_coords[:, 0, :], axis=-1)
+            min_y = np.min(new_coords[:, 1, :], axis=-1)
+            max_x = np.max(new_coords[:, 0, :], axis=-1)
+            max_y = np.max(new_coords[:, 1, :], axis=-1)
+            min_x = np.clip(min_x, a_min=0, a_max=w)
+            min_y = np.clip(min_y, a_min=0, a_max=h)
+            max_x = np.clip(max_x, a_min=min_x, a_max=w)
+            max_y = np.clip(max_y, a_min=min_y, a_max=h)
+            results[key] = np.stack([min_x, min_y, max_x, max_y],
+                                    axis=-1).astype(results[key].dtype)
+
+    def _shear_masks(self,
+                     results,
+                     magnitude,
+                     direction='horizontal',
+                     fill_val=0,
+                     interpolation='bilinear'):
+        """Shear the masks."""
+        h, w, c = results['img_shape']
+        for key in results.get('mask_fields', []):
+            masks = results[key]
+            results[key] = masks.shear((h, w),
+                                       magnitude,
+                                       direction,
+                                       border_value=fill_val,
+                                       interpolation=interpolation)
+
+    def _shear_seg(self,
+                   results,
+                   magnitude,
+                   direction='horizontal',
+                   fill_val=255,
+                   interpolation='bilinear'):
+        """Shear the segmentation maps."""
+        for key in results.get('seg_fields', []):
+            seg = results[key]
+            results[key] = mmcv.imshear(
+                seg,
+                magnitude,
+                direction,
+                border_value=fill_val,
+                interpolation=interpolation).astype(seg.dtype)
+
+    def _filter_invalid(self, results, min_bbox_size=0):
+        """Filter bboxes and corresponding masks too small after shear
+        augmentation."""
+        bbox2label, bbox2mask, _ = bbox2fields()
+        for key in results.get('bbox_fields', []):
+            bbox_w = results[key][:, 2] - results[key][:, 0]
+            bbox_h = results[key][:, 3] - results[key][:, 1]
+            valid_inds = (bbox_w > min_bbox_size) & (bbox_h > min_bbox_size)
+            valid_inds = np.nonzero(valid_inds)[0]
+            results[key] = results[key][valid_inds]
+            # label fields. e.g. gt_labels and gt_labels_ignore
+            label_key = bbox2label.get(key)
+            if label_key in results:
+                results[label_key] = results[label_key][valid_inds]
+            # mask fields, e.g. gt_masks and gt_masks_ignore
+            mask_key = bbox2mask.get(key)
+            if mask_key in results:
+                results[mask_key] = results[mask_key][valid_inds]
+
+    def __call__(self, results):
+        """Call function to shear images, bounding boxes, masks and semantic
+        segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Sheared results.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        magnitude = random_negative(self.magnitude, self.random_negative_prob)
+        self._shear_img(results, magnitude, self.direction, self.interpolation)
+        self._shear_bboxes(results, magnitude)
+        # fill_val set to 0 for background of mask.
+        self._shear_masks(
+            results,
+            magnitude,
+            self.direction,
+            fill_val=0,
+            interpolation=self.interpolation)
+        self._shear_seg(
+            results,
+            magnitude,
+            self.direction,
+            fill_val=self.seg_ignore_label,
+            interpolation=self.interpolation)
+        self._filter_invalid(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(level={self.level}, '
+        repr_str += f'img_fill_val={self.img_fill_val}, '
+        repr_str += f'seg_ignore_label={self.seg_ignore_label}, '
+        repr_str += f'prob={self.prob}, '
+        repr_str += f'direction={self.direction}, '
+        repr_str += f'max_shear_magnitude={self.max_shear_magnitude}, '
+        repr_str += f'random_negative_prob={self.random_negative_prob}, '
+        repr_str += f'interpolation={self.interpolation})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Rotate(object):
+    """Apply Rotate Transformation to image (and its corresponding bbox, mask,
+    segmentation).
+
+    Args:
+        level (int | float): The level should be in range (0,_MAX_LEVEL].
+        scale (int | float): Isotropic scale factor. Same in
+            ``mmcv.imrotate``.
+        center (int | float | tuple[float]): Center point (w, h) of the
+            rotation in the source image. If None, the center of the
+            image will be used. Same in ``mmcv.imrotate``.
+        img_fill_val (int | float | tuple): The fill value for image border.
+            If float, the same value will be used for all the three
+            channels of image. If tuple, the should be 3 elements (e.g.
+            equals the number of channels for image).
+        seg_ignore_label (int): The fill value used for segmentation map.
+            Note this value must equals ``ignore_label`` in ``semantic_head``
+            of the corresponding config. Default 255.
+        prob (float): The probability for perform transformation and
+            should be in range 0 to 1.
+        max_rotate_angle (int | float): The maximum angles for rotate
+            transformation.
+        random_negative_prob (float): The probability that turns the
+             offset negative.
+    """
+
+    def __init__(self,
+                 level,
+                 scale=1,
+                 center=None,
+                 img_fill_val=128,
+                 seg_ignore_label=255,
+                 prob=0.5,
+                 max_rotate_angle=30,
+                 random_negative_prob=0.5):
+        assert isinstance(level, (int, float)), \
+            f'The level must be type int or float. got {type(level)}.'
+        assert 0 <= level <= _MAX_LEVEL, \
+            f'The level should be in range (0,{_MAX_LEVEL}]. got {level}.'
+        assert isinstance(scale, (int, float)), \
+            f'The scale must be type int or float. got type {type(scale)}.'
+        if isinstance(center, (int, float)):
+            center = (center, center)
+        elif isinstance(center, tuple):
+            assert len(center) == 2, 'center with type tuple must have '\
+                f'2 elements. got {len(center)} elements.'
+        else:
+            assert center is None, 'center must be None or type int, '\
+                f'float or tuple, got type {type(center)}.'
+        if isinstance(img_fill_val, (float, int)):
+            img_fill_val = tuple([float(img_fill_val)] * 3)
+        elif isinstance(img_fill_val, tuple):
+            assert len(img_fill_val) == 3, 'img_fill_val as tuple must '\
+                f'have 3 elements. got {len(img_fill_val)}.'
+            img_fill_val = tuple([float(val) for val in img_fill_val])
+        else:
+            raise ValueError(
+                'img_fill_val must be float or tuple with 3 elements.')
+        assert np.all([0 <= val <= 255 for val in img_fill_val]), \
+            'all elements of img_fill_val should between range [0,255]. '\
+            f'got {img_fill_val}.'
+        assert 0 <= prob <= 1.0, 'The probability should be in range [0,1]. '\
+            'got {prob}.'
+        assert isinstance(max_rotate_angle, (int, float)), 'max_rotate_angle '\
+            f'should be type int or float. got type {type(max_rotate_angle)}.'
+        self.level = level
+        self.scale = scale
+        # Rotation angle in degrees. Positive values mean
+        # clockwise rotation.
+        self.angle = level_to_value(level, max_rotate_angle)
+        self.center = center
+        self.img_fill_val = img_fill_val
+        self.seg_ignore_label = seg_ignore_label
+        self.prob = prob
+        self.max_rotate_angle = max_rotate_angle
+        self.random_negative_prob = random_negative_prob
+
+    def _rotate_img(self, results, angle, center=None, scale=1.0):
+        """Rotate the image.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+            angle (float): Rotation angle in degrees, positive values
+                mean clockwise rotation. Same in ``mmcv.imrotate``.
+            center (tuple[float], optional): Center point (w, h) of the
+                rotation. Same in ``mmcv.imrotate``.
+            scale (int | float): Isotropic scale factor. Same in
+                ``mmcv.imrotate``.
+        """
+        for key in results.get('img_fields', ['img']):
+            img = results[key].copy()
+            img_rotated = mmcv.imrotate(
+                img, angle, center, scale, border_value=self.img_fill_val)
+            results[key] = img_rotated.astype(img.dtype)
+
+    def _rotate_bboxes(self, results, rotate_matrix):
+        """Rotate the bboxes."""
+        h, w, c = results['img_shape']
+        for key in results.get('bbox_fields', []):
+            min_x, min_y, max_x, max_y = np.split(
+                results[key], results[key].shape[-1], axis=-1)
+            coordinates = np.stack([[min_x, min_y], [max_x, min_y],
+                                    [min_x, max_y],
+                                    [max_x, max_y]])  # [4, 2, nb_bbox, 1]
+            # pad 1 to convert from format [x, y] to homogeneous
+            # coordinates format [x, y, 1]
+            coordinates = np.concatenate(
+                (coordinates,
+                 np.ones((4, 1, coordinates.shape[2], 1), coordinates.dtype)),
+                axis=1)  # [4, 3, nb_bbox, 1]
+            coordinates = coordinates.transpose(
+                (2, 0, 1, 3))  # [nb_bbox, 4, 3, 1]
+            rotated_coords = np.matmul(rotate_matrix,
+                                       coordinates)  # [nb_bbox, 4, 2, 1]
+            rotated_coords = rotated_coords[..., 0]  # [nb_bbox, 4, 2]
+            min_x, min_y = np.min(
+                rotated_coords[:, :, 0], axis=1), np.min(
+                    rotated_coords[:, :, 1], axis=1)
+            max_x, max_y = np.max(
+                rotated_coords[:, :, 0], axis=1), np.max(
+                    rotated_coords[:, :, 1], axis=1)
+            min_x, min_y = np.clip(
+                min_x, a_min=0, a_max=w), np.clip(
+                    min_y, a_min=0, a_max=h)
+            max_x, max_y = np.clip(
+                max_x, a_min=min_x, a_max=w), np.clip(
+                    max_y, a_min=min_y, a_max=h)
+            results[key] = np.stack([min_x, min_y, max_x, max_y],
+                                    axis=-1).astype(results[key].dtype)
+
+    def _rotate_masks(self,
+                      results,
+                      angle,
+                      center=None,
+                      scale=1.0,
+                      fill_val=0):
+        """Rotate the masks."""
+        h, w, c = results['img_shape']
+        for key in results.get('mask_fields', []):
+            masks = results[key]
+            results[key] = masks.rotate((h, w), angle, center, scale, fill_val)
+
+    def _rotate_seg(self,
+                    results,
+                    angle,
+                    center=None,
+                    scale=1.0,
+                    fill_val=255):
+        """Rotate the segmentation map."""
+        for key in results.get('seg_fields', []):
+            seg = results[key].copy()
+            results[key] = mmcv.imrotate(
+                seg, angle, center, scale,
+                border_value=fill_val).astype(seg.dtype)
+
+    def _filter_invalid(self, results, min_bbox_size=0):
+        """Filter bboxes and corresponding masks too small after rotate
+        augmentation."""
+        bbox2label, bbox2mask, _ = bbox2fields()
+        for key in results.get('bbox_fields', []):
+            bbox_w = results[key][:, 2] - results[key][:, 0]
+            bbox_h = results[key][:, 3] - results[key][:, 1]
+            valid_inds = (bbox_w > min_bbox_size) & (bbox_h > min_bbox_size)
+            valid_inds = np.nonzero(valid_inds)[0]
+            results[key] = results[key][valid_inds]
+            # label fields. e.g. gt_labels and gt_labels_ignore
+            label_key = bbox2label.get(key)
+            if label_key in results:
+                results[label_key] = results[label_key][valid_inds]
+            # mask fields, e.g. gt_masks and gt_masks_ignore
+            mask_key = bbox2mask.get(key)
+            if mask_key in results:
+                results[mask_key] = results[mask_key][valid_inds]
+
+    def __call__(self, results):
+        """Call function to rotate images, bounding boxes, masks and semantic
+        segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Rotated results.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        h, w = results['img'].shape[:2]
+        center = self.center
+        if center is None:
+            center = ((w - 1) * 0.5, (h - 1) * 0.5)
+        angle = random_negative(self.angle, self.random_negative_prob)
+        self._rotate_img(results, angle, center, self.scale)
+        rotate_matrix = cv2.getRotationMatrix2D(center, -angle, self.scale)
+        self._rotate_bboxes(results, rotate_matrix)
+        self._rotate_masks(results, angle, center, self.scale, fill_val=0)
+        self._rotate_seg(
+            results, angle, center, self.scale, fill_val=self.seg_ignore_label)
+        self._filter_invalid(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(level={self.level}, '
+        repr_str += f'scale={self.scale}, '
+        repr_str += f'center={self.center}, '
+        repr_str += f'img_fill_val={self.img_fill_val}, '
+        repr_str += f'seg_ignore_label={self.seg_ignore_label}, '
+        repr_str += f'prob={self.prob}, '
+        repr_str += f'max_rotate_angle={self.max_rotate_angle}, '
+        repr_str += f'random_negative_prob={self.random_negative_prob})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Translate(object):
+    """Translate the images, bboxes, masks and segmentation maps horizontally
+    or vertically.
+
+    Args:
+        level (int | float): The level for Translate and should be in
+            range [0,_MAX_LEVEL].
+        prob (float): The probability for performing translation and
+            should be in range [0, 1].
+        img_fill_val (int | float | tuple): The filled value for image
+            border. If float, the same fill value will be used for all
+            the three channels of image. If tuple, the should be 3
+            elements (e.g. equals the number of channels for image).
+        seg_ignore_label (int): The fill value used for segmentation map.
+            Note this value must equals ``ignore_label`` in ``semantic_head``
+            of the corresponding config. Default 255.
+        direction (str): The translate direction, either "horizontal"
+            or "vertical".
+        max_translate_offset (int | float): The maximum pixel's offset for
+            Translate.
+        random_negative_prob (float): The probability that turns the
+            offset negative.
+        min_size (int | float): The minimum pixel for filtering
+            invalid bboxes after the translation.
+    """
+
+    def __init__(self,
+                 level,
+                 prob=0.5,
+                 img_fill_val=128,
+                 seg_ignore_label=255,
+                 direction='horizontal',
+                 max_translate_offset=250.,
+                 random_negative_prob=0.5,
+                 min_size=0):
+        assert isinstance(level, (int, float)), \
+            'The level must be type int or float.'
+        assert 0 <= level <= _MAX_LEVEL, \
+            'The level used for calculating Translate\'s offset should be ' \
+            'in range [0,_MAX_LEVEL]'
+        assert 0 <= prob <= 1.0, \
+            'The probability of translation should be in range [0, 1].'
+        if isinstance(img_fill_val, (float, int)):
+            img_fill_val = tuple([float(img_fill_val)] * 3)
+        elif isinstance(img_fill_val, tuple):
+            assert len(img_fill_val) == 3, \
+                'img_fill_val as tuple must have 3 elements.'
+            img_fill_val = tuple([float(val) for val in img_fill_val])
+        else:
+            raise ValueError('img_fill_val must be type float or tuple.')
+        assert np.all([0 <= val <= 255 for val in img_fill_val]), \
+            'all elements of img_fill_val should between range [0,255].'
+        assert direction in ('horizontal', 'vertical'), \
+            'direction should be "horizontal" or "vertical".'
+        assert isinstance(max_translate_offset, (int, float)), \
+            'The max_translate_offset must be type int or float.'
+        # the offset used for translation
+        self.offset = int(level_to_value(level, max_translate_offset))
+        self.level = level
+        self.prob = prob
+        self.img_fill_val = img_fill_val
+        self.seg_ignore_label = seg_ignore_label
+        self.direction = direction
+        self.max_translate_offset = max_translate_offset
+        self.random_negative_prob = random_negative_prob
+        self.min_size = min_size
+
+    def _translate_img(self, results, offset, direction='horizontal'):
+        """Translate the image.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+            offset (int | float): The offset for translate.
+            direction (str): The translate direction, either "horizontal"
+                or "vertical".
+        """
+        for key in results.get('img_fields', ['img']):
+            img = results[key].copy()
+            results[key] = mmcv.imtranslate(
+                img, offset, direction, self.img_fill_val).astype(img.dtype)
+
+    def _translate_bboxes(self, results, offset):
+        """Shift bboxes horizontally or vertically, according to offset."""
+        h, w, c = results['img_shape']
+        for key in results.get('bbox_fields', []):
+            min_x, min_y, max_x, max_y = np.split(
+                results[key], results[key].shape[-1], axis=-1)
+            if self.direction == 'horizontal':
+                min_x = np.maximum(0, min_x + offset)
+                max_x = np.minimum(w, max_x + offset)
+            elif self.direction == 'vertical':
+                min_y = np.maximum(0, min_y + offset)
+                max_y = np.minimum(h, max_y + offset)
+
+            # the boxs translated outside of image will be filtered along with
+            # the corresponding masks, by invoking ``_filter_invalid``.
+            results[key] = np.concatenate([min_x, min_y, max_x, max_y],
+                                          axis=-1)
+
+    def _translate_masks(self,
+                         results,
+                         offset,
+                         direction='horizontal',
+                         fill_val=0):
+        """Translate masks horizontally or vertically."""
+        h, w, c = results['img_shape']
+        for key in results.get('mask_fields', []):
+            masks = results[key]
+            results[key] = masks.translate((h, w), offset, direction, fill_val)
+
+    def _translate_seg(self,
+                       results,
+                       offset,
+                       direction='horizontal',
+                       fill_val=255):
+        """Translate segmentation maps horizontally or vertically."""
+        for key in results.get('seg_fields', []):
+            seg = results[key].copy()
+            results[key] = mmcv.imtranslate(seg, offset, direction,
+                                            fill_val).astype(seg.dtype)
+
+    def _filter_invalid(self, results, min_size=0):
+        """Filter bboxes and masks too small or translated out of image."""
+        bbox2label, bbox2mask, _ = bbox2fields()
+        for key in results.get('bbox_fields', []):
+            bbox_w = results[key][:, 2] - results[key][:, 0]
+            bbox_h = results[key][:, 3] - results[key][:, 1]
+            valid_inds = (bbox_w > min_size) & (bbox_h > min_size)
+            valid_inds = np.nonzero(valid_inds)[0]
+            results[key] = results[key][valid_inds]
+            # label fields. e.g. gt_labels and gt_labels_ignore
+            label_key = bbox2label.get(key)
+            if label_key in results:
+                results[label_key] = results[label_key][valid_inds]
+            # mask fields, e.g. gt_masks and gt_masks_ignore
+            mask_key = bbox2mask.get(key)
+            if mask_key in results:
+                results[mask_key] = results[mask_key][valid_inds]
+        return results
+
+    def __call__(self, results):
+        """Call function to translate images, bounding boxes, masks and
+        semantic segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Translated results.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        offset = random_negative(self.offset, self.random_negative_prob)
+        self._translate_img(results, offset, self.direction)
+        self._translate_bboxes(results, offset)
+        # fill_val defaultly 0 for BitmapMasks and None for PolygonMasks.
+        self._translate_masks(results, offset, self.direction)
+        # fill_val set to ``seg_ignore_label`` for the ignored value
+        # of segmentation map.
+        self._translate_seg(
+            results, offset, self.direction, fill_val=self.seg_ignore_label)
+        self._filter_invalid(results, min_size=self.min_size)
+        return results
+
+
+@PIPELINES.register_module()
+class ColorTransform(object):
+    """Apply Color transformation to image. The bboxes, masks, and
+    segmentations are not modified.
+
+    Args:
+        level (int | float): Should be in range [0,_MAX_LEVEL].
+        prob (float): The probability for performing Color transformation.
+    """
+
+    def __init__(self, level, prob=0.5):
+        assert isinstance(level, (int, float)), \
+            'The level must be type int or float.'
+        assert 0 <= level <= _MAX_LEVEL, \
+            'The level should be in range [0,_MAX_LEVEL].'
+        assert 0 <= prob <= 1.0, \
+            'The probability should be in range [0,1].'
+        self.level = level
+        self.prob = prob
+        self.factor = enhance_level_to_value(level)
+
+    def _adjust_color_img(self, results, factor=1.0):
+        """Apply Color transformation to image."""
+        for key in results.get('img_fields', ['img']):
+            # NOTE defaultly the image should be BGR format
+            img = results[key]
+            results[key] = mmcv.adjust_color(img, factor).astype(img.dtype)
+
+    def __call__(self, results):
+        """Call function for Color transformation.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Colored results.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        self._adjust_color_img(results, self.factor)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(level={self.level}, '
+        repr_str += f'prob={self.prob})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class EqualizeTransform(object):
+    """Apply Equalize transformation to image. The bboxes, masks and
+    segmentations are not modified.
+
+    Args:
+        prob (float): The probability for performing Equalize transformation.
+    """
+
+    def __init__(self, prob=0.5):
+        assert 0 <= prob <= 1.0, \
+            'The probability should be in range [0,1].'
+        self.prob = prob
+
+    def _imequalize(self, results):
+        """Equalizes the histogram of one image."""
+        for key in results.get('img_fields', ['img']):
+            img = results[key]
+            results[key] = mmcv.imequalize(img).astype(img.dtype)
+
+    def __call__(self, results):
+        """Call function for Equalize transformation.
+
+        Args:
+            results (dict): Results dict from loading pipeline.
+
+        Returns:
+            dict: Results after the transformation.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        self._imequalize(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(prob={self.prob})'
+
+
+@PIPELINES.register_module()
+class BrightnessTransform(object):
+    """Apply Brightness transformation to image. The bboxes, masks and
+    segmentations are not modified.
+
+    Args:
+        level (int | float): Should be in range [0,_MAX_LEVEL].
+        prob (float): The probability for performing Brightness transformation.
+    """
+
+    def __init__(self, level, prob=0.5):
+        assert isinstance(level, (int, float)), \
+            'The level must be type int or float.'
+        assert 0 <= level <= _MAX_LEVEL, \
+            'The level should be in range [0,_MAX_LEVEL].'
+        assert 0 <= prob <= 1.0, \
+            'The probability should be in range [0,1].'
+        self.level = level
+        self.prob = prob
+        self.factor = enhance_level_to_value(level)
+
+    def _adjust_brightness_img(self, results, factor=1.0):
+        """Adjust the brightness of image."""
+        for key in results.get('img_fields', ['img']):
+            img = results[key]
+            results[key] = mmcv.adjust_brightness(img,
+                                                  factor).astype(img.dtype)
+
+    def __call__(self, results):
+        """Call function for Brightness transformation.
+
+        Args:
+            results (dict): Results dict from loading pipeline.
+
+        Returns:
+            dict: Results after the transformation.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        self._adjust_brightness_img(results, self.factor)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(level={self.level}, '
+        repr_str += f'prob={self.prob})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class ContrastTransform(object):
+    """Apply Contrast transformation to image. The bboxes, masks and
+    segmentations are not modified.
+
+    Args:
+        level (int | float): Should be in range [0,_MAX_LEVEL].
+        prob (float): The probability for performing Contrast transformation.
+    """
+
+    def __init__(self, level, prob=0.5):
+        assert isinstance(level, (int, float)), \
+            'The level must be type int or float.'
+        assert 0 <= level <= _MAX_LEVEL, \
+            'The level should be in range [0,_MAX_LEVEL].'
+        assert 0 <= prob <= 1.0, \
+            'The probability should be in range [0,1].'
+        self.level = level
+        self.prob = prob
+        self.factor = enhance_level_to_value(level)
+
+    def _adjust_contrast_img(self, results, factor=1.0):
+        """Adjust the image contrast."""
+        for key in results.get('img_fields', ['img']):
+            img = results[key]
+            results[key] = mmcv.adjust_contrast(img, factor).astype(img.dtype)
+
+    def __call__(self, results):
+        """Call function for Contrast transformation.
+
+        Args:
+            results (dict): Results dict from loading pipeline.
+
+        Returns:
+            dict: Results after the transformation.
+        """
+        if np.random.rand() > self.prob:
+            return results
+        self._adjust_contrast_img(results, self.factor)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(level={self.level}, '
+        repr_str += f'prob={self.prob})'
+        return repr_str
diff --git a/detection/scrfd/mmdet/datasets/pipelines/compose.py b/detection/scrfd/mmdet/datasets/pipelines/compose.py
new file mode 100755
index 000000000..ca48f1c93
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/compose.py
@@ -0,0 +1,51 @@
+import collections
+
+from mmcv.utils import build_from_cfg
+
+from ..builder import PIPELINES
+
+
+@PIPELINES.register_module()
+class Compose(object):
+    """Compose multiple transforms sequentially.
+
+    Args:
+        transforms (Sequence[dict | callable]): Sequence of transform object or
+            config dict to be composed.
+    """
+
+    def __init__(self, transforms):
+        assert isinstance(transforms, collections.abc.Sequence)
+        self.transforms = []
+        for transform in transforms:
+            if isinstance(transform, dict):
+                transform = build_from_cfg(transform, PIPELINES)
+                self.transforms.append(transform)
+            elif callable(transform):
+                self.transforms.append(transform)
+            else:
+                raise TypeError('transform must be callable or a dict')
+
+    def __call__(self, data):
+        """Call function to apply transforms sequentially.
+
+        Args:
+            data (dict): A result dict contains the data to transform.
+
+        Returns:
+           dict: Transformed data.
+        """
+
+        for t in self.transforms:
+            data = t(data)
+            if data is None:
+                return None
+        return data
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + '('
+        for t in self.transforms:
+            format_string += '\n'
+            format_string += f'    {t}'
+        format_string += '\n)'
+        return format_string
diff --git a/detection/scrfd/mmdet/datasets/pipelines/formating.py b/detection/scrfd/mmdet/datasets/pipelines/formating.py
new file mode 100755
index 000000000..afb75edc2
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/formating.py
@@ -0,0 +1,364 @@
+from collections.abc import Sequence
+
+import mmcv
+import numpy as np
+import torch
+from mmcv.parallel import DataContainer as DC
+
+from ..builder import PIPELINES
+
+
+def to_tensor(data):
+    """Convert objects of various python types to :obj:`torch.Tensor`.
+
+    Supported types are: :class:`numpy.ndarray`, :class:`torch.Tensor`,
+    :class:`Sequence`, :class:`int` and :class:`float`.
+
+    Args:
+        data (torch.Tensor | numpy.ndarray | Sequence | int | float): Data to
+            be converted.
+    """
+
+    if isinstance(data, torch.Tensor):
+        return data
+    elif isinstance(data, np.ndarray):
+        return torch.from_numpy(data)
+    elif isinstance(data, Sequence) and not mmcv.is_str(data):
+        return torch.tensor(data)
+    elif isinstance(data, int):
+        return torch.LongTensor([data])
+    elif isinstance(data, float):
+        return torch.FloatTensor([data])
+    else:
+        raise TypeError(f'type {type(data)} cannot be converted to tensor.')
+
+
+@PIPELINES.register_module()
+class ToTensor(object):
+    """Convert some results to :obj:`torch.Tensor` by given keys.
+
+    Args:
+        keys (Sequence[str]): Keys that need to be converted to Tensor.
+    """
+
+    def __init__(self, keys):
+        self.keys = keys
+
+    def __call__(self, results):
+        """Call function to convert data in results to :obj:`torch.Tensor`.
+
+        Args:
+            results (dict): Result dict contains the data to convert.
+
+        Returns:
+            dict: The result dict contains the data converted
+                to :obj:`torch.Tensor`.
+        """
+        for key in self.keys:
+            results[key] = to_tensor(results[key])
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + f'(keys={self.keys})'
+
+
+@PIPELINES.register_module()
+class ImageToTensor(object):
+    """Convert image to :obj:`torch.Tensor` by given keys.
+
+    The dimension order of input image is (H, W, C). The pipeline will convert
+    it to (C, H, W). If only 2 dimension (H, W) is given, the output would be
+    (1, H, W).
+
+    Args:
+        keys (Sequence[str]): Key of images to be converted to Tensor.
+    """
+
+    def __init__(self, keys):
+        self.keys = keys
+
+    def __call__(self, results):
+        """Call function to convert image in results to :obj:`torch.Tensor` and
+        transpose the channel order.
+
+        Args:
+            results (dict): Result dict contains the image data to convert.
+
+        Returns:
+            dict: The result dict contains the image converted
+                to :obj:`torch.Tensor` and transposed to (C, H, W) order.
+        """
+        for key in self.keys:
+            img = results[key]
+            if len(img.shape) < 3:
+                img = np.expand_dims(img, -1)
+            results[key] = to_tensor(img.transpose(2, 0, 1))
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + f'(keys={self.keys})'
+
+
+@PIPELINES.register_module()
+class Transpose(object):
+    """Transpose some results by given keys.
+
+    Args:
+        keys (Sequence[str]): Keys of results to be transposed.
+        order (Sequence[int]): Order of transpose.
+    """
+
+    def __init__(self, keys, order):
+        self.keys = keys
+        self.order = order
+
+    def __call__(self, results):
+        """Call function to transpose the channel order of data in results.
+
+        Args:
+            results (dict): Result dict contains the data to transpose.
+
+        Returns:
+            dict: The result dict contains the data transposed to \
+                ``self.order``.
+        """
+        for key in self.keys:
+            results[key] = results[key].transpose(self.order)
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + \
+            f'(keys={self.keys}, order={self.order})'
+
+
+@PIPELINES.register_module()
+class ToDataContainer(object):
+    """Convert results to :obj:`mmcv.DataContainer` by given fields.
+
+    Args:
+        fields (Sequence[dict]): Each field is a dict like
+            ``dict(key='xxx', **kwargs)``. The ``key`` in result will
+            be converted to :obj:`mmcv.DataContainer` with ``**kwargs``.
+            Default: ``(dict(key='img', stack=True), dict(key='gt_bboxes'),
+            dict(key='gt_labels'))``.
+    """
+
+    def __init__(self,
+                 fields=(dict(key='img', stack=True), dict(key='gt_bboxes'),
+                         dict(key='gt_labels'))):
+        self.fields = fields
+
+    def __call__(self, results):
+        """Call function to convert data in results to
+        :obj:`mmcv.DataContainer`.
+
+        Args:
+            results (dict): Result dict contains the data to convert.
+
+        Returns:
+            dict: The result dict contains the data converted to \
+                :obj:`mmcv.DataContainer`.
+        """
+
+        for field in self.fields:
+            field = field.copy()
+            key = field.pop('key')
+            results[key] = DC(results[key], **field)
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + f'(fields={self.fields})'
+
+
+@PIPELINES.register_module()
+class DefaultFormatBundle(object):
+    """Default formatting bundle.
+
+    It simplifies the pipeline of formatting common fields, including "img",
+    "proposals", "gt_bboxes", "gt_labels", "gt_masks" and "gt_semantic_seg".
+    These fields are formatted as follows.
+
+    - img: (1)transpose, (2)to tensor, (3)to DataContainer (stack=True)
+    - proposals: (1)to tensor, (2)to DataContainer
+    - gt_bboxes: (1)to tensor, (2)to DataContainer
+    - gt_bboxes_ignore: (1)to tensor, (2)to DataContainer
+    - gt_labels: (1)to tensor, (2)to DataContainer
+    - gt_masks: (1)to tensor, (2)to DataContainer (cpu_only=True)
+    - gt_semantic_seg: (1)unsqueeze dim-0 (2)to tensor, \
+                       (3)to DataContainer (stack=True)
+    """
+
+    def __call__(self, results):
+        """Call function to transform and format common fields in results.
+
+        Args:
+            results (dict): Result dict contains the data to convert.
+
+        Returns:
+            dict: The result dict contains the data that is formatted with \
+                default bundle.
+        """
+
+        if 'img' in results:
+            img = results['img']
+            # add default meta keys
+            results = self._add_default_meta_keys(results)
+            if len(img.shape) < 3:
+                img = np.expand_dims(img, -1)
+            img = np.ascontiguousarray(img.transpose(2, 0, 1))
+            results['img'] = DC(to_tensor(img), stack=True)
+        for key in ['proposals', 'gt_bboxes', 'gt_bboxes_ignore', 'gt_keypointss', 'gt_labels']:
+            if key not in results:
+                continue
+            results[key] = DC(to_tensor(results[key]))
+        if 'gt_masks' in results:
+            results['gt_masks'] = DC(results['gt_masks'], cpu_only=True)
+        if 'gt_semantic_seg' in results:
+            results['gt_semantic_seg'] = DC(
+                to_tensor(results['gt_semantic_seg'][None, ...]), stack=True)
+        return results
+
+    def _add_default_meta_keys(self, results):
+        """Add default meta keys.
+
+        We set default meta keys including `pad_shape`, `scale_factor` and
+        `img_norm_cfg` to avoid the case where no `Resize`, `Normalize` and
+        `Pad` are implemented during the whole pipeline.
+
+        Args:
+            results (dict): Result dict contains the data to convert.
+
+        Returns:
+            results (dict): Updated result dict contains the data to convert.
+        """
+        img = results['img']
+        results.setdefault('pad_shape', img.shape)
+        results.setdefault('scale_factor', 1.0)
+        num_channels = 1 if len(img.shape) < 3 else img.shape[2]
+        results.setdefault(
+            'img_norm_cfg',
+            dict(
+                mean=np.zeros(num_channels, dtype=np.float32),
+                std=np.ones(num_channels, dtype=np.float32),
+                to_rgb=False))
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__
+
+
+@PIPELINES.register_module()
+class Collect(object):
+    """Collect data from the loader relevant to the specific task.
+
+    This is usually the last stage of the data loader pipeline. Typically keys
+    is set to some subset of "img", "proposals", "gt_bboxes",
+    "gt_bboxes_ignore", "gt_labels", and/or "gt_masks".
+
+    The "img_meta" item is always populated.  The contents of the "img_meta"
+    dictionary depends on "meta_keys". By default this includes:
+
+        - "img_shape": shape of the image input to the network as a tuple \
+            (h, w, c).  Note that images may be zero padded on the \
+            bottom/right if the batch tensor is larger than this shape.
+
+        - "scale_factor": a float indicating the preprocessing scale
+
+        - "flip": a boolean indicating if image flip transform was used
+
+        - "filename": path to the image file
+
+        - "ori_shape": original shape of the image as a tuple (h, w, c)
+
+        - "pad_shape": image shape after padding
+
+        - "img_norm_cfg": a dict of normalization information:
+
+            - mean - per channel mean subtraction
+            - std - per channel std divisor
+            - to_rgb - bool indicating if bgr was converted to rgb
+
+    Args:
+        keys (Sequence[str]): Keys of results to be collected in ``data``.
+        meta_keys (Sequence[str], optional): Meta keys to be converted to
+            ``mmcv.DataContainer`` and collected in ``data[img_metas]``.
+            Default: ``('filename', 'ori_filename', 'ori_shape', 'img_shape',
+            'pad_shape', 'scale_factor', 'flip', 'flip_direction',
+            'img_norm_cfg')``
+    """
+
+    def __init__(self,
+                 keys,
+                 meta_keys=('filename', 'ori_filename', 'ori_shape',
+                            'img_shape', 'pad_shape', 'scale_factor', 'flip',
+                            'flip_direction', 'img_norm_cfg')):
+        self.keys = keys
+        self.meta_keys = meta_keys
+
+    def __call__(self, results):
+        """Call function to collect keys in results. The keys in ``meta_keys``
+        will be converted to :obj:mmcv.DataContainer.
+
+        Args:
+            results (dict): Result dict contains the data to collect.
+
+        Returns:
+            dict: The result dict contains the following keys
+
+                - keys in``self.keys``
+                - ``img_metas``
+        """
+
+        data = {}
+        img_meta = {}
+        for key in self.meta_keys:
+            img_meta[key] = results[key]
+        data['img_metas'] = DC(img_meta, cpu_only=True)
+        for key in self.keys:
+            data[key] = results[key]
+        return data
+
+    def __repr__(self):
+        return self.__class__.__name__ + \
+            f'(keys={self.keys}, meta_keys={self.meta_keys})'
+
+
+@PIPELINES.register_module()
+class WrapFieldsToLists(object):
+    """Wrap fields of the data dictionary into lists for evaluation.
+
+    This class can be used as a last step of a test or validation
+    pipeline for single image evaluation or inference.
+
+    Example:
+        >>> test_pipeline = [
+        >>>    dict(type='LoadImageFromFile'),
+        >>>    dict(type='Normalize',
+                    mean=[123.675, 116.28, 103.53],
+                    std=[58.395, 57.12, 57.375],
+                    to_rgb=True),
+        >>>    dict(type='Pad', size_divisor=32),
+        >>>    dict(type='ImageToTensor', keys=['img']),
+        >>>    dict(type='Collect', keys=['img']),
+        >>>    dict(type='WrapFieldsToLists')
+        >>> ]
+    """
+
+    def __call__(self, results):
+        """Call function to wrap fields into lists.
+
+        Args:
+            results (dict): Result dict contains the data to wrap.
+
+        Returns:
+            dict: The result dict where value of ``self.keys`` are wrapped \
+                into list.
+        """
+
+        # Wrap dict fields into lists
+        for key, val in results.items():
+            results[key] = [val]
+        return results
+
+    def __repr__(self):
+        return f'{self.__class__.__name__}()'
diff --git a/detection/scrfd/mmdet/datasets/pipelines/instaboost.py b/detection/scrfd/mmdet/datasets/pipelines/instaboost.py
new file mode 100755
index 000000000..38b6819f6
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/instaboost.py
@@ -0,0 +1,98 @@
+import numpy as np
+
+from ..builder import PIPELINES
+
+
+@PIPELINES.register_module()
+class InstaBoost(object):
+    r"""Data augmentation method in `InstaBoost: Boosting Instance
+    Segmentation Via Probability Map Guided Copy-Pasting
+    <https://arxiv.org/abs/1908.07801>`_.
+
+    Refer to https://github.com/GothicAi/Instaboost for implementation details.
+    """
+
+    def __init__(self,
+                 action_candidate=('normal', 'horizontal', 'skip'),
+                 action_prob=(1, 0, 0),
+                 scale=(0.8, 1.2),
+                 dx=15,
+                 dy=15,
+                 theta=(-1, 1),
+                 color_prob=0.5,
+                 hflag=False,
+                 aug_ratio=0.5):
+        try:
+            import instaboostfast as instaboost
+        except ImportError:
+            raise ImportError(
+                'Please run "pip install instaboostfast" '
+                'to install instaboostfast first for instaboost augmentation.')
+        self.cfg = instaboost.InstaBoostConfig(action_candidate, action_prob,
+                                               scale, dx, dy, theta,
+                                               color_prob, hflag)
+        self.aug_ratio = aug_ratio
+
+    def _load_anns(self, results):
+        labels = results['ann_info']['labels']
+        masks = results['ann_info']['masks']
+        bboxes = results['ann_info']['bboxes']
+        n = len(labels)
+
+        anns = []
+        for i in range(n):
+            label = labels[i]
+            bbox = bboxes[i]
+            mask = masks[i]
+            x1, y1, x2, y2 = bbox
+            # assert (x2 - x1) >= 1 and (y2 - y1) >= 1
+            bbox = [x1, y1, x2 - x1, y2 - y1]
+            anns.append({
+                'category_id': label,
+                'segmentation': mask,
+                'bbox': bbox
+            })
+
+        return anns
+
+    def _parse_anns(self, results, anns, img):
+        gt_bboxes = []
+        gt_labels = []
+        gt_masks_ann = []
+        for ann in anns:
+            x1, y1, w, h = ann['bbox']
+            # TODO: more essential bug need to be fixed in instaboost
+            if w <= 0 or h <= 0:
+                continue
+            bbox = [x1, y1, x1 + w, y1 + h]
+            gt_bboxes.append(bbox)
+            gt_labels.append(ann['category_id'])
+            gt_masks_ann.append(ann['segmentation'])
+        gt_bboxes = np.array(gt_bboxes, dtype=np.float32)
+        gt_labels = np.array(gt_labels, dtype=np.int64)
+        results['ann_info']['labels'] = gt_labels
+        results['ann_info']['bboxes'] = gt_bboxes
+        results['ann_info']['masks'] = gt_masks_ann
+        results['img'] = img
+        return results
+
+    def __call__(self, results):
+        img = results['img']
+        orig_type = img.dtype
+        anns = self._load_anns(results)
+        if np.random.choice([0, 1], p=[1 - self.aug_ratio, self.aug_ratio]):
+            try:
+                import instaboostfast as instaboost
+            except ImportError:
+                raise ImportError('Please run "pip install instaboostfast" '
+                                  'to install instaboostfast first.')
+            anns, img = instaboost.get_new_data(
+                anns, img.astype(np.uint8), self.cfg, background=None)
+
+        results = self._parse_anns(results, anns, img.astype(orig_type))
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(cfg={self.cfg}, aug_ratio={self.aug_ratio})'
+        return repr_str
diff --git a/detection/scrfd/mmdet/datasets/pipelines/loading.py b/detection/scrfd/mmdet/datasets/pipelines/loading.py
new file mode 100755
index 000000000..7f8c1a5c1
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/loading.py
@@ -0,0 +1,480 @@
+import os.path as osp
+
+import mmcv
+import numpy as np
+import pycocotools.mask as maskUtils
+
+from mmdet.core import BitmapMasks, PolygonMasks
+from ..builder import PIPELINES
+
+
+@PIPELINES.register_module()
+class LoadImageFromFile(object):
+    """Load an image from file.
+
+    Required keys are "img_prefix" and "img_info" (a dict that must contain the
+    key "filename"). Added or updated keys are "filename", "img", "img_shape",
+    "ori_shape" (same as `img_shape`), "pad_shape" (same as `img_shape`),
+    "scale_factor" (1.0) and "img_norm_cfg" (means=0 and stds=1).
+
+    Args:
+        to_float32 (bool): Whether to convert the loaded image to a float32
+            numpy array. If set to False, the loaded image is an uint8 array.
+            Defaults to False.
+        color_type (str): The flag argument for :func:`mmcv.imfrombytes`.
+            Defaults to 'color'.
+        file_client_args (dict): Arguments to instantiate a FileClient.
+            See :class:`mmcv.fileio.FileClient` for details.
+            Defaults to ``dict(backend='disk')``.
+    """
+
+    def __init__(self,
+                 to_float32=False,
+                 color_type='color',
+                 file_client_args=dict(backend='disk')):
+        self.to_float32 = to_float32
+        self.color_type = color_type
+        self.file_client_args = file_client_args.copy()
+        self.file_client = None
+
+    def __call__(self, results):
+        """Call functions to load image and get image meta information.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded image and meta information.
+        """
+
+        if self.file_client is None:
+            self.file_client = mmcv.FileClient(**self.file_client_args)
+
+        if results['img_prefix'] is not None:
+            filename = osp.join(results['img_prefix'],
+                                results['img_info']['filename'])
+        else:
+            filename = results['img_info']['filename']
+
+        img_bytes = self.file_client.get(filename)
+        img = mmcv.imfrombytes(img_bytes, flag=self.color_type)
+        if self.to_float32:
+            img = img.astype(np.float32)
+
+        results['filename'] = filename
+        results['ori_filename'] = results['img_info']['filename']
+        results['img'] = img
+        results['img_shape'] = img.shape
+        results['ori_shape'] = img.shape
+        results['img_fields'] = ['img']
+        return results
+
+    def __repr__(self):
+        repr_str = (f'{self.__class__.__name__}('
+                    f'to_float32={self.to_float32}, '
+                    f"color_type='{self.color_type}', "
+                    f'file_client_args={self.file_client_args})')
+        return repr_str
+
+
+@PIPELINES.register_module()
+class LoadImageFromWebcam(LoadImageFromFile):
+    """Load an image from webcam.
+
+    Similar with :obj:`LoadImageFromFile`, but the image read from webcam is in
+    ``results['img']``.
+    """
+
+    def __call__(self, results):
+        """Call functions to add image meta information.
+
+        Args:
+            results (dict): Result dict with Webcam read image in
+                ``results['img']``.
+
+        Returns:
+            dict: The dict contains loaded image and meta information.
+        """
+
+        img = results['img']
+        if self.to_float32:
+            img = img.astype(np.float32)
+
+        results['filename'] = None
+        results['ori_filename'] = None
+        results['img'] = img
+        results['img_shape'] = img.shape
+        results['ori_shape'] = img.shape
+        results['img_fields'] = ['img']
+        return results
+
+
+@PIPELINES.register_module()
+class LoadMultiChannelImageFromFiles(object):
+    """Load multi-channel images from a list of separate channel files.
+
+    Required keys are "img_prefix" and "img_info" (a dict that must contain the
+    key "filename", which is expected to be a list of filenames).
+    Added or updated keys are "filename", "img", "img_shape",
+    "ori_shape" (same as `img_shape`), "pad_shape" (same as `img_shape`),
+    "scale_factor" (1.0) and "img_norm_cfg" (means=0 and stds=1).
+
+    Args:
+        to_float32 (bool): Whether to convert the loaded image to a float32
+            numpy array. If set to False, the loaded image is an uint8 array.
+            Defaults to False.
+        color_type (str): The flag argument for :func:`mmcv.imfrombytes`.
+            Defaults to 'color'.
+        file_client_args (dict): Arguments to instantiate a FileClient.
+            See :class:`mmcv.fileio.FileClient` for details.
+            Defaults to ``dict(backend='disk')``.
+    """
+
+    def __init__(self,
+                 to_float32=False,
+                 color_type='unchanged',
+                 file_client_args=dict(backend='disk')):
+        self.to_float32 = to_float32
+        self.color_type = color_type
+        self.file_client_args = file_client_args.copy()
+        self.file_client = None
+
+    def __call__(self, results):
+        """Call functions to load multiple images and get images meta
+        information.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded images and meta information.
+        """
+
+        if self.file_client is None:
+            self.file_client = mmcv.FileClient(**self.file_client_args)
+
+        if results['img_prefix'] is not None:
+            filename = [
+                osp.join(results['img_prefix'], fname)
+                for fname in results['img_info']['filename']
+            ]
+        else:
+            filename = results['img_info']['filename']
+
+        img = []
+        for name in filename:
+            img_bytes = self.file_client.get(name)
+            img.append(mmcv.imfrombytes(img_bytes, flag=self.color_type))
+        img = np.stack(img, axis=-1)
+        if self.to_float32:
+            img = img.astype(np.float32)
+
+        results['filename'] = filename
+        results['ori_filename'] = results['img_info']['filename']
+        results['img'] = img
+        results['img_shape'] = img.shape
+        results['ori_shape'] = img.shape
+        # Set initial values for default meta_keys
+        results['pad_shape'] = img.shape
+        results['scale_factor'] = 1.0
+        num_channels = 1 if len(img.shape) < 3 else img.shape[2]
+        results['img_norm_cfg'] = dict(
+            mean=np.zeros(num_channels, dtype=np.float32),
+            std=np.ones(num_channels, dtype=np.float32),
+            to_rgb=False)
+        return results
+
+    def __repr__(self):
+        repr_str = (f'{self.__class__.__name__}('
+                    f'to_float32={self.to_float32}, '
+                    f"color_type='{self.color_type}', "
+                    f'file_client_args={self.file_client_args})')
+        return repr_str
+
+
+@PIPELINES.register_module()
+class LoadAnnotations(object):
+    """Load mutiple types of annotations.
+
+    Args:
+        with_bbox (bool): Whether to parse and load the bbox annotation.
+             Default: True.
+        with_label (bool): Whether to parse and load the label annotation.
+            Default: True.
+        with_keypoints (bool): Whether to parse and load the keypoints annotation.
+            Default: False.
+        with_mask (bool): Whether to parse and load the mask annotation.
+             Default: False.
+        with_seg (bool): Whether to parse and load the semantic segmentation
+            annotation. Default: False.
+        poly2mask (bool): Whether to convert the instance masks from polygons
+            to bitmaps. Default: True.
+        file_client_args (dict): Arguments to instantiate a FileClient.
+            See :class:`mmcv.fileio.FileClient` for details.
+            Defaults to ``dict(backend='disk')``.
+    """
+
+    def __init__(self,
+                 with_bbox=True,
+                 with_label=True,
+                 with_keypoints=False,
+                 with_mask=False,
+                 with_seg=False,
+                 poly2mask=True,
+                 file_client_args=dict(backend='disk')):
+        self.with_bbox = with_bbox
+        self.with_label = with_label
+        self.with_keypoints = with_keypoints
+        self.with_mask = with_mask
+        self.with_seg = with_seg
+        self.poly2mask = poly2mask
+        self.file_client_args = file_client_args.copy()
+        self.file_client = None
+
+    def _load_bboxes(self, results):
+        """Private function to load bounding box annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded bounding box annotations.
+        """
+
+        ann_info = results['ann_info']
+        results['gt_bboxes'] = ann_info['bboxes'].copy()
+
+        gt_bboxes_ignore = ann_info.get('bboxes_ignore', None)
+        if gt_bboxes_ignore is not None:
+            results['gt_bboxes_ignore'] = gt_bboxes_ignore.copy()
+            results['bbox_fields'].append('gt_bboxes_ignore')
+        results['bbox_fields'].append('gt_bboxes')
+        return results
+    def _load_keypoints(self, results):
+        """Private function to load bounding box annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded bounding box annotations.
+        """
+
+        ann_info = results['ann_info']
+        results['gt_keypointss'] = ann_info['keypointss'].copy()
+
+        results['keypoints_fields'] = ['gt_keypointss']
+        return results
+
+    def _load_labels(self, results):
+        """Private function to load label annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded label annotations.
+        """
+
+        results['gt_labels'] = results['ann_info']['labels'].copy()
+        return results
+
+    def _poly2mask(self, mask_ann, img_h, img_w):
+        """Private function to convert masks represented with polygon to
+        bitmaps.
+
+        Args:
+            mask_ann (list | dict): Polygon mask annotation input.
+            img_h (int): The height of output mask.
+            img_w (int): The width of output mask.
+
+        Returns:
+            numpy.ndarray: The decode bitmap mask of shape (img_h, img_w).
+        """
+
+        if isinstance(mask_ann, list):
+            # polygon -- a single object might consist of multiple parts
+            # we merge all parts into one mask rle code
+            rles = maskUtils.frPyObjects(mask_ann, img_h, img_w)
+            rle = maskUtils.merge(rles)
+        elif isinstance(mask_ann['counts'], list):
+            # uncompressed RLE
+            rle = maskUtils.frPyObjects(mask_ann, img_h, img_w)
+        else:
+            # rle
+            rle = mask_ann
+        mask = maskUtils.decode(rle)
+        return mask
+
+    def process_polygons(self, polygons):
+        """Convert polygons to list of ndarray and filter invalid polygons.
+
+        Args:
+            polygons (list[list]): Polygons of one instance.
+
+        Returns:
+            list[numpy.ndarray]: Processed polygons.
+        """
+
+        polygons = [np.array(p) for p in polygons]
+        valid_polygons = []
+        for polygon in polygons:
+            if len(polygon) % 2 == 0 and len(polygon) >= 6:
+                valid_polygons.append(polygon)
+        return valid_polygons
+
+    def _load_masks(self, results):
+        """Private function to load mask annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded mask annotations.
+                If ``self.poly2mask`` is set ``True``, `gt_mask` will contain
+                :obj:`PolygonMasks`. Otherwise, :obj:`BitmapMasks` is used.
+        """
+
+        h, w = results['img_info']['height'], results['img_info']['width']
+        gt_masks = results['ann_info']['masks']
+        if self.poly2mask:
+            gt_masks = BitmapMasks(
+                [self._poly2mask(mask, h, w) for mask in gt_masks], h, w)
+        else:
+            gt_masks = PolygonMasks(
+                [self.process_polygons(polygons) for polygons in gt_masks], h,
+                w)
+        results['gt_masks'] = gt_masks
+        results['mask_fields'].append('gt_masks')
+        return results
+
+    def _load_semantic_seg(self, results):
+        """Private function to load semantic segmentation annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`dataset`.
+
+        Returns:
+            dict: The dict contains loaded semantic segmentation annotations.
+        """
+
+        if self.file_client is None:
+            self.file_client = mmcv.FileClient(**self.file_client_args)
+
+        filename = osp.join(results['seg_prefix'],
+                            results['ann_info']['seg_map'])
+        img_bytes = self.file_client.get(filename)
+        results['gt_semantic_seg'] = mmcv.imfrombytes(
+            img_bytes, flag='unchanged').squeeze()
+        results['seg_fields'].append('gt_semantic_seg')
+        return results
+
+    def __call__(self, results):
+        """Call function to load multiple types annotations.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded bounding box, label, mask and
+                semantic segmentation annotations.
+        """
+
+        if self.with_bbox:
+            results = self._load_bboxes(results)
+            if results is None:
+                return None
+        if self.with_label:
+            results = self._load_labels(results)
+        if self.with_keypoints:
+            results = self._load_keypoints(results)
+        if self.with_mask:
+            results = self._load_masks(results)
+        if self.with_seg:
+            results = self._load_semantic_seg(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(with_bbox={self.with_bbox}, '
+        repr_str += f'with_label={self.with_label}, '
+        repr_str += f'with_keypoints={self.with_keypoints}, '
+        repr_str += f'with_mask={self.with_mask}, '
+        repr_str += f'with_seg={self.with_seg})'
+        repr_str += f'poly2mask={self.poly2mask})'
+        repr_str += f'poly2mask={self.file_client_args})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class LoadProposals(object):
+    """Load proposal pipeline.
+
+    Required key is "proposals". Updated keys are "proposals", "bbox_fields".
+
+    Args:
+        num_max_proposals (int, optional): Maximum number of proposals to load.
+            If not specified, all proposals will be loaded.
+    """
+
+    def __init__(self, num_max_proposals=None):
+        self.num_max_proposals = num_max_proposals
+
+    def __call__(self, results):
+        """Call function to load proposals from file.
+
+        Args:
+            results (dict): Result dict from :obj:`mmdet.CustomDataset`.
+
+        Returns:
+            dict: The dict contains loaded proposal annotations.
+        """
+
+        proposals = results['proposals']
+        if proposals.shape[1] not in (4, 5):
+            raise AssertionError(
+                'proposals should have shapes (n, 4) or (n, 5), '
+                f'but found {proposals.shape}')
+        proposals = proposals[:, :4]
+
+        if self.num_max_proposals is not None:
+            proposals = proposals[:self.num_max_proposals]
+
+        if len(proposals) == 0:
+            proposals = np.array([[0, 0, 0, 0]], dtype=np.float32)
+        results['proposals'] = proposals
+        results['bbox_fields'].append('proposals')
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + \
+            f'(num_max_proposals={self.num_max_proposals})'
+
+
+@PIPELINES.register_module()
+class FilterAnnotations(object):
+    """Filter invalid annotations.
+
+    Args:
+        min_gt_bbox_wh (tuple[int]): Minimum width and height of ground truth
+            boxes.
+    """
+
+    def __init__(self, min_gt_bbox_wh):
+        # TODO: add more filter options
+        self.min_gt_bbox_wh = min_gt_bbox_wh
+
+    def __call__(self, results):
+        assert 'gt_bboxes' in results
+        gt_bboxes = results['gt_bboxes']
+        w = gt_bboxes[:, 2] - gt_bboxes[:, 0]
+        h = gt_bboxes[:, 3] - gt_bboxes[:, 1]
+        keep = (w > self.min_gt_bbox_wh[0]) & (h > self.min_gt_bbox_wh[1])
+        if not keep.any():
+            return None
+        else:
+            keys = ('gt_bboxes', 'gt_labels', 'gt_masks', 'gt_semantic_seg')
+            for key in keys:
+                if key in results:
+                    results[key] = results[key][keep]
+            return results
diff --git a/detection/scrfd/mmdet/datasets/pipelines/test_time_aug.py b/detection/scrfd/mmdet/datasets/pipelines/test_time_aug.py
new file mode 100755
index 000000000..2b0ad4db3
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/test_time_aug.py
@@ -0,0 +1,119 @@
+import warnings
+
+import mmcv
+
+from ..builder import PIPELINES
+from .compose import Compose
+
+
+@PIPELINES.register_module()
+class MultiScaleFlipAug(object):
+    """Test-time augmentation with multiple scales and flipping.
+
+    An example configuration is as followed:
+
+    .. code-block::
+
+        img_scale=[(1333, 400), (1333, 800)],
+        flip=True,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ]
+
+    After MultiScaleFLipAug with above configuration, the results are wrapped
+    into lists of the same length as followed:
+
+    .. code-block::
+
+        dict(
+            img=[...],
+            img_shape=[...],
+            scale=[(1333, 400), (1333, 400), (1333, 800), (1333, 800)]
+            flip=[False, True, False, True]
+            ...
+        )
+
+    Args:
+        transforms (list[dict]): Transforms to apply in each augmentation.
+        img_scale (tuple | list[tuple] | None): Images scales for resizing.
+        scale_factor (float | list[float] | None): Scale factors for resizing.
+        flip (bool): Whether apply flip augmentation. Default: False.
+        flip_direction (str | list[str]): Flip augmentation directions,
+            options are "horizontal" and "vertical". If flip_direction is list,
+            multiple flip augmentations will be applied.
+            It has no effect when flip == False. Default: "horizontal".
+    """
+
+    def __init__(self,
+                 transforms,
+                 img_scale=None,
+                 scale_factor=None,
+                 flip=False,
+                 flip_direction='horizontal'):
+        self.transforms = Compose(transforms)
+        assert (img_scale is None) ^ (scale_factor is None), (
+            'Must have but only one variable can be setted')
+        if img_scale is not None:
+            self.img_scale = img_scale if isinstance(img_scale,
+                                                     list) else [img_scale]
+            self.scale_key = 'scale'
+            assert mmcv.is_list_of(self.img_scale, tuple)
+        else:
+            self.img_scale = scale_factor if isinstance(
+                scale_factor, list) else [scale_factor]
+            self.scale_key = 'scale_factor'
+
+        self.flip = flip
+        self.flip_direction = flip_direction if isinstance(
+            flip_direction, list) else [flip_direction]
+        assert mmcv.is_list_of(self.flip_direction, str)
+        if not self.flip and self.flip_direction != ['horizontal']:
+            warnings.warn(
+                'flip_direction has no effect when flip is set to False')
+        if (self.flip
+                and not any([t['type'] == 'RandomFlip' for t in transforms])):
+            warnings.warn(
+                'flip has no effect when RandomFlip is not in transforms')
+
+    def __call__(self, results):
+        """Call function to apply test time augment transforms on results.
+
+        Args:
+            results (dict): Result dict contains the data to transform.
+
+        Returns:
+           dict[str: list]: The augmented data, where each value is wrapped
+               into a list.
+        """
+
+        aug_data = []
+        flip_args = [(False, None)]
+        if self.flip:
+            flip_args += [(True, direction)
+                          for direction in self.flip_direction]
+        for scale in self.img_scale:
+            for flip, direction in flip_args:
+                _results = results.copy()
+                _results[self.scale_key] = scale
+                _results['flip'] = flip
+                _results['flip_direction'] = direction
+                data = self.transforms(_results)
+                aug_data.append(data)
+        # list of dict to dict of list
+        aug_data_dict = {key: [] for key in aug_data[0]}
+        for data in aug_data:
+            for key, val in data.items():
+                aug_data_dict[key].append(val)
+        return aug_data_dict
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(transforms={self.transforms}, '
+        repr_str += f'img_scale={self.img_scale}, flip={self.flip})'
+        repr_str += f'flip_direction={self.flip_direction}'
+        return repr_str
diff --git a/detection/scrfd/mmdet/datasets/pipelines/transforms.py b/detection/scrfd/mmdet/datasets/pipelines/transforms.py
new file mode 100755
index 000000000..869acfb5b
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/pipelines/transforms.py
@@ -0,0 +1,2037 @@
+import inspect
+
+import mmcv
+import numpy as np
+from numpy import random
+import cv2
+
+from mmdet.core import PolygonMasks
+from mmdet.core.evaluation.bbox_overlaps import bbox_overlaps
+from ..builder import PIPELINES
+
+try:
+    from imagecorruptions import corrupt
+except ImportError:
+    corrupt = None
+
+try:
+    import albumentations
+    from albumentations import Compose
+except ImportError:
+    albumentations = None
+    Compose = None
+
+
+@PIPELINES.register_module()
+class Resize(object):
+    """Resize images & bbox & mask.
+
+    This transform resizes the input image to some scale. Bboxes and masks are
+    then resized with the same scale factor. If the input dict contains the key
+    "scale", then the scale in the input dict is used, otherwise the specified
+    scale in the init method is used. If the input dict contains the key
+    "scale_factor" (if MultiScaleFlipAug does not give img_scale but
+    scale_factor), the actual scale will be computed by image shape and
+    scale_factor.
+
+    `img_scale` can either be a tuple (single-scale) or a list of tuple
+    (multi-scale). There are 3 multiscale modes:
+
+    - ``ratio_range is not None``: randomly sample a ratio from the ratio \
+      range and multiply it with the image scale.
+    - ``ratio_range is None`` and ``multiscale_mode == "range"``: randomly \
+      sample a scale from the multiscale range.
+    - ``ratio_range is None`` and ``multiscale_mode == "value"``: randomly \
+      sample a scale from multiple scales.
+
+    Args:
+        img_scale (tuple or list[tuple]): Images scales for resizing.
+        multiscale_mode (str): Either "range" or "value".
+        ratio_range (tuple[float]): (min_ratio, max_ratio)
+        keep_ratio (bool): Whether to keep the aspect ratio when resizing the
+            image.
+        bbox_clip_border (bool, optional): Whether clip the objects outside
+            the border of the image. Defaults to True.
+        backend (str): Image resize backend, choices are 'cv2' and 'pillow'.
+            These two backends generates slightly different results. Defaults
+            to 'cv2'.
+        override (bool, optional): Whether to override `scale` and
+            `scale_factor` so as to call resize twice. Default False. If True,
+            after the first resizing, the existed `scale` and `scale_factor`
+            will be ignored so the second resizing can be allowed.
+            This option is a work-around for multiple times of resize in DETR.
+            Defaults to False.
+    """
+
+    def __init__(self,
+                 img_scale=None,
+                 multiscale_mode='range',
+                 ratio_range=None,
+                 keep_ratio=True,
+                 bbox_clip_border=True,
+                 backend='cv2',
+                 override=False):
+        if img_scale is None:
+            self.img_scale = None
+        else:
+            if isinstance(img_scale, list):
+                self.img_scale = img_scale
+            else:
+                self.img_scale = [img_scale]
+            assert mmcv.is_list_of(self.img_scale, tuple)
+
+        if ratio_range is not None:
+            # mode 1: given a scale and a range of image ratio
+            assert len(self.img_scale) == 1
+        else:
+            # mode 2: given multiple scales or a range of scales
+            assert multiscale_mode in ['value', 'range']
+
+        self.backend = backend
+        self.multiscale_mode = multiscale_mode
+        self.ratio_range = ratio_range
+        self.keep_ratio = keep_ratio
+        # TODO: refactor the override option in Resize
+        self.override = override
+        self.bbox_clip_border = bbox_clip_border
+
+    @staticmethod
+    def random_select(img_scales):
+        """Randomly select an img_scale from given candidates.
+
+        Args:
+            img_scales (list[tuple]): Images scales for selection.
+
+        Returns:
+            (tuple, int): Returns a tuple ``(img_scale, scale_dix)``, \
+                where ``img_scale`` is the selected image scale and \
+                ``scale_idx`` is the selected index in the given candidates.
+        """
+
+        assert mmcv.is_list_of(img_scales, tuple)
+        scale_idx = np.random.randint(len(img_scales))
+        img_scale = img_scales[scale_idx]
+        return img_scale, scale_idx
+
+    @staticmethod
+    def random_sample(img_scales):
+        """Randomly sample an img_scale when ``multiscale_mode=='range'``.
+
+        Args:
+            img_scales (list[tuple]): Images scale range for sampling.
+                There must be two tuples in img_scales, which specify the lower
+                and uper bound of image scales.
+
+        Returns:
+            (tuple, None): Returns a tuple ``(img_scale, None)``, where \
+                ``img_scale`` is sampled scale and None is just a placeholder \
+                to be consistent with :func:`random_select`.
+        """
+
+        assert mmcv.is_list_of(img_scales, tuple) and len(img_scales) == 2
+        img_scale_long = [max(s) for s in img_scales]
+        img_scale_short = [min(s) for s in img_scales]
+        long_edge = np.random.randint(
+            min(img_scale_long),
+            max(img_scale_long) + 1)
+        short_edge = np.random.randint(
+            min(img_scale_short),
+            max(img_scale_short) + 1)
+        img_scale = (long_edge, short_edge)
+        return img_scale, None
+
+    @staticmethod
+    def random_sample_ratio(img_scale, ratio_range):
+        """Randomly sample an img_scale when ``ratio_range`` is specified.
+
+        A ratio will be randomly sampled from the range specified by
+        ``ratio_range``. Then it would be multiplied with ``img_scale`` to
+        generate sampled scale.
+
+        Args:
+            img_scale (tuple): Images scale base to multiply with ratio.
+            ratio_range (tuple[float]): The minimum and maximum ratio to scale
+                the ``img_scale``.
+
+        Returns:
+            (tuple, None): Returns a tuple ``(scale, None)``, where \
+                ``scale`` is sampled ratio multiplied with ``img_scale`` and \
+                None is just a placeholder to be consistent with \
+                :func:`random_select`.
+        """
+
+        assert isinstance(img_scale, tuple) and len(img_scale) == 2
+        min_ratio, max_ratio = ratio_range
+        assert min_ratio <= max_ratio
+        ratio = np.random.random_sample() * (max_ratio - min_ratio) + min_ratio
+        scale = int(img_scale[0] * ratio), int(img_scale[1] * ratio)
+        return scale, None
+
+    def _random_scale(self, results):
+        """Randomly sample an img_scale according to ``ratio_range`` and
+        ``multiscale_mode``.
+
+        If ``ratio_range`` is specified, a ratio will be sampled and be
+        multiplied with ``img_scale``.
+        If multiple scales are specified by ``img_scale``, a scale will be
+        sampled according to ``multiscale_mode``.
+        Otherwise, single scale will be used.
+
+        Args:
+            results (dict): Result dict from :obj:`dataset`.
+
+        Returns:
+            dict: Two new keys 'scale` and 'scale_idx` are added into \
+                ``results``, which would be used by subsequent pipelines.
+        """
+
+        if self.ratio_range is not None:
+            scale, scale_idx = self.random_sample_ratio(
+                self.img_scale[0], self.ratio_range)
+        elif len(self.img_scale) == 1:
+            scale, scale_idx = self.img_scale[0], 0
+        elif self.multiscale_mode == 'range':
+            scale, scale_idx = self.random_sample(self.img_scale)
+        elif self.multiscale_mode == 'value':
+            scale, scale_idx = self.random_select(self.img_scale)
+        else:
+            raise NotImplementedError
+
+        results['scale'] = scale
+        results['scale_idx'] = scale_idx
+
+    def _resize_img(self, results):
+        """Resize images with ``results['scale']``."""
+        for key in results.get('img_fields', ['img']):
+            if self.keep_ratio:
+                img, scale_factor = mmcv.imrescale(
+                    results[key],
+                    results['scale'],
+                    return_scale=True,
+                    backend=self.backend)
+                # the w_scale and h_scale has minor difference
+                # a real fix should be done in the mmcv.imrescale in the future
+                new_h, new_w = img.shape[:2]
+                h, w = results[key].shape[:2]
+                w_scale = new_w / w
+                h_scale = new_h / h
+            else:
+                img, w_scale, h_scale = mmcv.imresize(
+                    results[key],
+                    results['scale'],
+                    return_scale=True,
+                    backend=self.backend)
+            results[key] = img
+
+            scale_factor = np.array([w_scale, h_scale, w_scale, h_scale],
+                                    dtype=np.float32)
+            results['img_shape'] = img.shape
+            # in case that there is no padding
+            results['pad_shape'] = img.shape
+            results['scale_factor'] = scale_factor
+            results['keep_ratio'] = self.keep_ratio
+
+    def _resize_bboxes(self, results):
+        """Resize bounding boxes with ``results['scale_factor']``."""
+        for key in results.get('bbox_fields', []):
+            bboxes = results[key] * results['scale_factor']
+            if self.bbox_clip_border:
+                img_shape = results['img_shape']
+                bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, img_shape[1])
+                bboxes[:, 1::2] = np.clip(bboxes[:, 1::2], 0, img_shape[0])
+            results[key] = bboxes
+
+    def _resize_keypoints(self, results):
+        for key in results.get('keypoints_fields', []):
+            keypointss = results[key].copy()
+            factors = results['scale_factor']
+            assert factors[0]==factors[2]
+            assert factors[1]==factors[3]
+            #print('AAA', results['scale_factor'])
+            keypointss[:,:,0] *= factors[0]
+            keypointss[:,:,1] *= factors[1]
+            if self.bbox_clip_border:
+                img_shape = results['img_shape']
+                keypointss[:,:, 0] = np.clip(keypointss[:,:, 0], 0, img_shape[1])
+                keypointss[:,:, 1] = np.clip(keypointss[:,:, 1], 0, img_shape[0])
+            results[key] = keypointss
+
+    def _resize_masks(self, results):
+        """Resize masks with ``results['scale']``"""
+        for key in results.get('mask_fields', []):
+            if results[key] is None:
+                continue
+            if self.keep_ratio:
+                results[key] = results[key].rescale(results['scale'])
+            else:
+                results[key] = results[key].resize(results['img_shape'][:2])
+
+    def _resize_seg(self, results):
+        """Resize semantic segmentation map with ``results['scale']``."""
+        for key in results.get('seg_fields', []):
+            if self.keep_ratio:
+                gt_seg = mmcv.imrescale(
+                    results[key],
+                    results['scale'],
+                    interpolation='nearest',
+                    backend=self.backend)
+            else:
+                gt_seg = mmcv.imresize(
+                    results[key],
+                    results['scale'],
+                    interpolation='nearest',
+                    backend=self.backend)
+            results['gt_semantic_seg'] = gt_seg
+
+    def __call__(self, results):
+        """Call function to resize images, bounding boxes, masks, semantic
+        segmentation map.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Resized results, 'img_shape', 'pad_shape', 'scale_factor', \
+                'keep_ratio' keys are added into result dict.
+        """
+
+        if 'scale' not in results:
+            if 'scale_factor' in results:
+                img_shape = results['img'].shape[:2]
+                scale_factor = results['scale_factor']
+                assert isinstance(scale_factor, float)
+                results['scale'] = tuple(
+                    [int(x * scale_factor) for x in img_shape][::-1])
+            else:
+                self._random_scale(results)
+        else:
+            if not self.override:
+                assert 'scale_factor' not in results, (
+                    'scale and scale_factor cannot be both set.')
+            else:
+                results.pop('scale')
+                if 'scale_factor' in results:
+                    results.pop('scale_factor')
+                self._random_scale(results)
+
+        self._resize_img(results)
+        self._resize_bboxes(results)
+        self._resize_keypoints(results)
+        self._resize_masks(results)
+        self._resize_seg(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(img_scale={self.img_scale}, '
+        repr_str += f'multiscale_mode={self.multiscale_mode}, '
+        repr_str += f'ratio_range={self.ratio_range}, '
+        repr_str += f'keep_ratio={self.keep_ratio})'
+        repr_str += f'bbox_clip_border={self.bbox_clip_border})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class RandomFlip(object):
+    """Flip the image & bbox & mask.
+
+    If the input dict contains the key "flip", then the flag will be used,
+    otherwise it will be randomly decided by a ratio specified in the init
+    method.
+
+    When random flip is enabled, ``flip_ratio``/``direction`` can either be a
+    float/string or tuple of float/string. There are 3 flip modes:
+
+    - ``flip_ratio`` is float, ``direction`` is string: the image will be
+        ``direction``ly flipped with probability of ``flip_ratio`` .
+        E.g., ``flip_ratio=0.5``, ``direction='horizontal'``,
+        then image will be horizontally flipped with probability of 0.5.
+    - ``flip_ratio`` is float, ``direction`` is list of string: the image wil
+        be ``direction[i]``ly flipped with probability of
+        ``flip_ratio/len(direction)``.
+        E.g., ``flip_ratio=0.5``, ``direction=['horizontal', 'vertical']``,
+        then image will be horizontally flipped with probability of 0.25,
+        vertically with probability of 0.25.
+    - ``flip_ratio`` is list of float, ``direction`` is list of string:
+        given ``len(flip_ratio) == len(direction)``, the image wil
+        be ``direction[i]``ly flipped with probability of ``flip_ratio[i]``.
+        E.g., ``flip_ratio=[0.3, 0.5]``, ``direction=['horizontal',
+        'vertical']``, then image will be horizontally flipped with probability
+         of 0.3, vertically with probability of 0.5
+
+    Args:
+        flip_ratio (float | list[float], optional): The flipping probability.
+            Default: None.
+        direction(str | list[str], optional): The flipping direction. Options
+            are 'horizontal', 'vertical', 'diagonal'. Default: 'horizontal'.
+            If input is a list, the length must equal ``flip_ratio``. Each
+            element in ``flip_ratio`` indicates the flip probability of
+            corresponding direction.
+    """
+
+    def __init__(self, flip_ratio=None, direction='horizontal'):
+        if isinstance(flip_ratio, list):
+            assert mmcv.is_list_of(flip_ratio, float)
+            assert 0 <= sum(flip_ratio) <= 1
+        elif isinstance(flip_ratio, float):
+            assert 0 <= flip_ratio <= 1
+        elif flip_ratio is None:
+            pass
+        else:
+            raise ValueError('flip_ratios must be None, float, '
+                             'or list of float')
+        self.flip_ratio = flip_ratio
+
+        valid_directions = ['horizontal', 'vertical', 'diagonal']
+        if isinstance(direction, str):
+            assert direction in valid_directions
+        elif isinstance(direction, list):
+            assert mmcv.is_list_of(direction, str)
+            assert set(direction).issubset(set(valid_directions))
+        else:
+            raise ValueError('direction must be either str or list of str')
+        self.direction = direction
+
+        if isinstance(flip_ratio, list):
+            assert len(self.flip_ratio) == len(self.direction)
+
+    def bbox_flip(self, bboxes, img_shape, direction):
+        """Flip bboxes horizontally.
+
+        Args:
+            bboxes (numpy.ndarray): Bounding boxes, shape (..., 4*k)
+            img_shape (tuple[int]): Image shape (height, width)
+            direction (str): Flip direction. Options are 'horizontal',
+                'vertical'.
+
+        Returns:
+            numpy.ndarray: Flipped bounding boxes.
+        """
+
+        assert bboxes.shape[-1] % 4 == 0
+        flipped = bboxes.copy()
+        if direction == 'horizontal':
+            w = img_shape[1]
+            flipped[..., 0::4] = w - bboxes[..., 2::4]
+            flipped[..., 2::4] = w - bboxes[..., 0::4]
+        elif direction == 'vertical':
+            h = img_shape[0]
+            flipped[..., 1::4] = h - bboxes[..., 3::4]
+            flipped[..., 3::4] = h - bboxes[..., 1::4]
+        elif direction == 'diagonal':
+            w = img_shape[1]
+            h = img_shape[0]
+            flipped[..., 0::4] = w - bboxes[..., 2::4]
+            flipped[..., 1::4] = h - bboxes[..., 3::4]
+            flipped[..., 2::4] = w - bboxes[..., 0::4]
+            flipped[..., 3::4] = h - bboxes[..., 1::4]
+        else:
+            raise ValueError(f"Invalid flipping direction '{direction}'")
+        return flipped
+
+    def keypoints_flip(self, keypointss, img_shape, direction):
+
+        assert direction == 'horizontal'
+        assert keypointss.shape[-1] == 3
+        assert keypointss.shape[1]==5
+
+        assert keypointss.ndim==3
+        flipped = keypointss.copy()
+        flip_order = [1,0,2,4,3]
+        for idx, a in enumerate(flip_order):
+            flipped[:,idx,:] = keypointss[:,a,:]
+        w = img_shape[1]
+        flipped[..., 0] = w - flipped[..., 0]
+        return flipped
+
+    def __call__(self, results):
+        """Call function to flip bounding boxes, masks, semantic segmentation
+        maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Flipped results, 'flip', 'flip_direction' keys are added \
+                into result dict.
+        """
+
+        if 'flip' not in results:
+            if isinstance(self.direction, list):
+                # None means non-flip
+                direction_list = self.direction + [None]
+            else:
+                # None means non-flip
+                direction_list = [self.direction, None]
+
+            if isinstance(self.flip_ratio, list):
+                non_flip_ratio = 1 - sum(self.flip_ratio)
+                flip_ratio_list = self.flip_ratio + [non_flip_ratio]
+            else:
+                non_flip_ratio = 1 - self.flip_ratio
+                # exclude non-flip
+                single_ratio = self.flip_ratio / (len(direction_list) - 1)
+                flip_ratio_list = [single_ratio] * (len(direction_list) -
+                                                    1) + [non_flip_ratio]
+
+            cur_dir = np.random.choice(direction_list, p=flip_ratio_list)
+
+            results['flip'] = cur_dir is not None
+        if 'flip_direction' not in results:
+            results['flip_direction'] = cur_dir
+        if results['flip']:
+            # flip image
+            for key in results.get('img_fields', ['img']):
+                results[key] = mmcv.imflip(
+                    results[key], direction=results['flip_direction'])
+            # flip bboxes
+            for key in results.get('bbox_fields', []):
+                results[key] = self.bbox_flip(results[key],
+                                              results['img_shape'],
+                                              results['flip_direction'])
+            for key in results.get('keypoints_fields', []):
+                results[key] = self.keypoints_flip(results[key],
+                                              results['img_shape'],
+                                              results['flip_direction'])
+            # flip masks
+            for key in results.get('mask_fields', []):
+                results[key] = results[key].flip(results['flip_direction'])
+
+            # flip segs
+            for key in results.get('seg_fields', []):
+                results[key] = mmcv.imflip(
+                    results[key], direction=results['flip_direction'])
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + f'(flip_ratio={self.flip_ratio})'
+
+
+@PIPELINES.register_module()
+class Pad(object):
+    """Pad the image & mask.
+
+    There are two padding modes: (1) pad to a fixed size and (2) pad to the
+    minimum size that is divisible by some number.
+    Added keys are "pad_shape", "pad_fixed_size", "pad_size_divisor",
+
+    Args:
+        size (tuple, optional): Fixed padding size.
+        size_divisor (int, optional): The divisor of padded size.
+        pad_val (float, optional): Padding value, 0 by default.
+    """
+
+    def __init__(self, size=None, size_divisor=None, pad_val=0):
+        self.size = size
+        self.size_divisor = size_divisor
+        self.pad_val = pad_val
+        # only one of size and size_divisor should be valid
+        assert size is not None or size_divisor is not None
+        assert size is None or size_divisor is None
+
+    def _pad_img(self, results):
+        """Pad images according to ``self.size``."""
+        for key in results.get('img_fields', ['img']):
+            if self.size is not None:
+                padded_img = mmcv.impad(
+                    results[key], shape=self.size, pad_val=self.pad_val)
+            elif self.size_divisor is not None:
+                padded_img = mmcv.impad_to_multiple(
+                    results[key], self.size_divisor, pad_val=self.pad_val)
+            results[key] = padded_img
+        results['pad_shape'] = padded_img.shape
+        results['pad_fixed_size'] = self.size
+        results['pad_size_divisor'] = self.size_divisor
+
+    def _pad_masks(self, results):
+        """Pad masks according to ``results['pad_shape']``."""
+        pad_shape = results['pad_shape'][:2]
+        for key in results.get('mask_fields', []):
+            results[key] = results[key].pad(pad_shape, pad_val=self.pad_val)
+
+    def _pad_seg(self, results):
+        """Pad semantic segmentation map according to
+        ``results['pad_shape']``."""
+        for key in results.get('seg_fields', []):
+            results[key] = mmcv.impad(
+                results[key], shape=results['pad_shape'][:2])
+
+    def __call__(self, results):
+        """Call function to pad images, masks, semantic segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Updated result dict.
+        """
+        self._pad_img(results)
+        self._pad_masks(results)
+        self._pad_seg(results)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(size={self.size}, '
+        repr_str += f'size_divisor={self.size_divisor}, '
+        repr_str += f'pad_val={self.pad_val})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Normalize(object):
+    """Normalize the image.
+
+    Added key is "img_norm_cfg".
+
+    Args:
+        mean (sequence): Mean values of 3 channels.
+        std (sequence): Std values of 3 channels.
+        to_rgb (bool): Whether to convert the image from BGR to RGB,
+            default is true.
+    """
+
+    def __init__(self, mean, std, to_rgb=True):
+        self.mean = np.array(mean, dtype=np.float32)
+        self.std = np.array(std, dtype=np.float32)
+        self.to_rgb = to_rgb
+
+    def __call__(self, results):
+        """Call function to normalize images.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Normalized results, 'img_norm_cfg' key is added into
+                result dict.
+        """
+        for key in results.get('img_fields', ['img']):
+            results[key] = mmcv.imnormalize(results[key], self.mean, self.std,
+                                            self.to_rgb)
+        results['img_norm_cfg'] = dict(
+            mean=self.mean, std=self.std, to_rgb=self.to_rgb)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(mean={self.mean}, std={self.std}, to_rgb={self.to_rgb})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class RandomCrop(object):
+    """Random crop the image & bboxes & masks.
+
+    The absolute `crop_size` is sampled based on `crop_type` and `image_size`,
+    then the cropped results are generated.
+
+    Args:
+        crop_size (tuple): The relative ratio or absolute pixels of
+            height and width.
+        crop_type (str, optional): one of "relative_range", "relative",
+            "absolute", "absolute_range". "relative" randomly crops
+            (h * crop_size[0], w * crop_size[1]) part from an input of size
+            (h, w). "relative_range" uniformly samples relative crop size from
+            range [crop_size[0], 1] and [crop_size[1], 1] for height and width
+            respectively. "absolute" crops from an input with absolute size
+            (crop_size[0], crop_size[1]). "absolute_range" uniformly samples
+            crop_h in range [crop_size[0], min(h, crop_size[1])] and crop_w
+            in range [crop_size[0], min(w, crop_size[1])]. Default "absolute".
+        allow_negative_crop (bool, optional): Whether to allow a crop that does
+            not contain any bbox area. Default False.
+        bbox_clip_border (bool, optional): Whether clip the objects outside
+            the border of the image. Defaults to True.
+
+    Note:
+        - If the image is smaller than the absolute crop size, return the
+            original image.
+        - The keys for bboxes, labels and masks must be aligned. That is,
+          `gt_bboxes` corresponds to `gt_labels` and `gt_masks`, and
+          `gt_bboxes_ignore` corresponds to `gt_labels_ignore` and
+          `gt_masks_ignore`.
+        - If the crop does not contain any gt-bbox region and
+          `allow_negative_crop` is set to False, skip this image.
+    """
+
+    def __init__(self,
+                 crop_size,
+                 crop_type='absolute',
+                 allow_negative_crop=False,
+                 bbox_clip_border=True):
+        if crop_type not in [
+                'relative_range', 'relative', 'absolute', 'absolute_range'
+        ]:
+            raise ValueError(f'Invalid crop_type {crop_type}.')
+        if crop_type in ['absolute', 'absolute_range']:
+            assert crop_size[0] > 0 and crop_size[1] > 0
+            assert isinstance(crop_size[0], int) and isinstance(
+                crop_size[1], int)
+        else:
+            assert 0 < crop_size[0] <= 1 and 0 < crop_size[1] <= 1
+        self.crop_size = crop_size
+        self.crop_type = crop_type
+        self.allow_negative_crop = allow_negative_crop
+        self.bbox_clip_border = bbox_clip_border
+        # The key correspondence from bboxes to labels and masks.
+        self.bbox2label = {
+            'gt_bboxes': 'gt_labels',
+            'gt_bboxes_ignore': 'gt_labels_ignore'
+        }
+        self.bbox2mask = {
+            'gt_bboxes': 'gt_masks',
+            'gt_bboxes_ignore': 'gt_masks_ignore'
+        }
+
+    def _crop_data(self, results, crop_size, allow_negative_crop):
+        """Function to randomly crop images, bounding boxes, masks, semantic
+        segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+            crop_size (tuple): Expected absolute size after cropping, (h, w).
+            allow_negative_crop (bool): Whether to allow a crop that does not
+                contain any bbox area. Default to False.
+
+        Returns:
+            dict: Randomly cropped results, 'img_shape' key in result dict is
+                updated according to crop size.
+        """
+        assert crop_size[0] > 0 and crop_size[1] > 0
+        for key in results.get('img_fields', ['img']):
+            img = results[key]
+            margin_h = max(img.shape[0] - crop_size[0], 0)
+            margin_w = max(img.shape[1] - crop_size[1], 0)
+            offset_h = np.random.randint(0, margin_h + 1)
+            offset_w = np.random.randint(0, margin_w + 1)
+            crop_y1, crop_y2 = offset_h, offset_h + crop_size[0]
+            crop_x1, crop_x2 = offset_w, offset_w + crop_size[1]
+
+            # crop the image
+            img = img[crop_y1:crop_y2, crop_x1:crop_x2, ...]
+            img_shape = img.shape
+            results[key] = img
+        results['img_shape'] = img_shape
+
+        # crop bboxes accordingly and clip to the image boundary
+        for key in results.get('bbox_fields', []):
+            # e.g. gt_bboxes and gt_bboxes_ignore
+            bbox_offset = np.array([offset_w, offset_h, offset_w, offset_h],
+                                   dtype=np.float32)
+            bboxes = results[key] - bbox_offset
+            if self.bbox_clip_border:
+                bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, img_shape[1])
+                bboxes[:, 1::2] = np.clip(bboxes[:, 1::2], 0, img_shape[0])
+            valid_inds = (bboxes[:, 2] > bboxes[:, 0]) & (
+                bboxes[:, 3] > bboxes[:, 1])
+            # If the crop does not contain any gt-bbox area and
+            # allow_negative_crop is False, skip this image.
+            if (key == 'gt_bboxes' and not valid_inds.any()
+                    and not allow_negative_crop):
+                return None
+            results[key] = bboxes[valid_inds, :]
+            # label fields. e.g. gt_labels and gt_labels_ignore
+            label_key = self.bbox2label.get(key)
+            if label_key in results:
+                results[label_key] = results[label_key][valid_inds]
+
+            # mask fields, e.g. gt_masks and gt_masks_ignore
+            mask_key = self.bbox2mask.get(key)
+            if mask_key in results:
+                results[mask_key] = results[mask_key][
+                    valid_inds.nonzero()[0]].crop(
+                        np.asarray([crop_x1, crop_y1, crop_x2, crop_y2]))
+
+        # crop semantic seg
+        for key in results.get('seg_fields', []):
+            results[key] = results[key][crop_y1:crop_y2, crop_x1:crop_x2]
+
+        return results
+
+    def _get_crop_size(self, image_size):
+        """Randomly generates the absolute crop size based on `crop_type` and
+        `image_size`.
+
+        Args:
+            image_size (tuple): (h, w).
+
+        Returns:
+            crop_size (tuple): (crop_h, crop_w) in absolute pixels.
+        """
+        h, w = image_size
+        if self.crop_type == 'absolute':
+            return (min(self.crop_size[0], h), min(self.crop_size[1], w))
+        elif self.crop_type == 'absolute_range':
+            assert self.crop_size[0] <= self.crop_size[1]
+            crop_h = np.random.randint(
+                min(h, self.crop_size[0]),
+                min(h, self.crop_size[1]) + 1)
+            crop_w = np.random.randint(
+                min(w, self.crop_size[0]),
+                min(w, self.crop_size[1]) + 1)
+            return crop_h, crop_w
+        elif self.crop_type == 'relative':
+            crop_h, crop_w = self.crop_size
+            return int(h * crop_h + 0.5), int(w * crop_w + 0.5)
+        elif self.crop_type == 'relative_range':
+            crop_size = np.asarray(self.crop_size, dtype=np.float32)
+            crop_h, crop_w = crop_size + np.random.rand(2) * (1 - crop_size)
+            return int(h * crop_h + 0.5), int(w * crop_w + 0.5)
+
+    def __call__(self, results):
+        """Call function to randomly crop images, bounding boxes, masks,
+        semantic segmentation maps.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Randomly cropped results, 'img_shape' key in result dict is
+                updated according to crop size.
+        """
+        image_size = results['img'].shape[:2]
+        crop_size = self._get_crop_size(image_size)
+        results = self._crop_data(results, crop_size, self.allow_negative_crop)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(crop_size={self.crop_size}, '
+        repr_str += f'crop_type={self.crop_type}, '
+        repr_str += f'allow_negative_crop={self.allow_negative_crop}, '
+        repr_str += f'bbox_clip_border={self.bbox_clip_border})'
+        return repr_str
+
+@PIPELINES.register_module()
+class RandomSquareCrop(object):
+    """Random crop the image & bboxes, the cropped patches have minimum IoU
+    requirement with original image & bboxes, the IoU threshold is randomly
+    selected from min_ious.
+
+    Args:
+        min_ious (tuple): minimum IoU threshold for all intersections with
+        bounding boxes
+        min_crop_size (float): minimum crop's size (i.e. h,w := a*h, a*w,
+        where a >= min_crop_size).
+
+    Note:
+        The keys for bboxes, labels and masks should be paired. That is, \
+        `gt_bboxes` corresponds to `gt_labels` and `gt_masks`, and \
+        `gt_bboxes_ignore` to `gt_labels_ignore` and `gt_masks_ignore`.
+    """
+
+    def __init__(self, crop_ratio_range=None, crop_choice=None, bbox_clip_border=True):
+
+        self.crop_ratio_range = crop_ratio_range
+        self.crop_choice = crop_choice
+        self.bbox_clip_border = bbox_clip_border
+
+        assert (self.crop_ratio_range is None) ^ (self.crop_choice is None)
+        if self.crop_ratio_range is not None:
+            self.crop_ratio_min, self.crop_ratio_max = self.crop_ratio_range
+
+        self.bbox2label = {
+            'gt_bboxes': 'gt_labels',
+            'gt_bboxes_ignore': 'gt_labels_ignore'
+        }
+        self.bbox2mask = {
+            'gt_bboxes': 'gt_masks',
+            'gt_bboxes_ignore': 'gt_masks_ignore'
+        }
+
+    def __call__(self, results):
+        """Call function to crop images and bounding boxes with minimum IoU
+        constraint.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with images and bounding boxes cropped, \
+                'img_shape' key is updated.
+        """
+
+        if 'img_fields' in results:
+            assert results['img_fields'] == ['img'], \
+                'Only single img_fields is allowed'
+        img = results['img']
+        assert 'bbox_fields' in results
+        assert 'gt_bboxes' in results
+        boxes = results['gt_bboxes']
+        #boxes = [results[key] for key in results['bbox_fields']]
+        #boxes = np.concatenate(boxes, 0)
+        h, w, c = img.shape
+        scale_retry = 0
+        if self.crop_ratio_range is not None:
+            max_scale = self.crop_ratio_max
+        else:
+            max_scale = np.amax(self.crop_choice)
+        #max_scale = max(max_scale, float(max(w,h))/min(w,h))
+
+        while True:
+            scale_retry += 1
+
+            if scale_retry==1 or max_scale>1.0:
+                if self.crop_ratio_range is not None:
+                    scale = np.random.uniform(self.crop_ratio_min,
+                                              self.crop_ratio_max)
+                elif self.crop_choice is not None:
+                    scale = np.random.choice(self.crop_choice)
+            else:
+                #scale = min(scale*1.2, max_scale)
+                scale = scale*1.2
+
+            # print(scale, img.shape[:2], boxes)
+            # import cv2
+            # cv2.imwrite('aaa.png', img)
+
+            for i in range(250):
+                short_side = min(w, h)
+                cw = int(scale * short_side)
+                ch = cw
+
+                # TODO +1
+                if w==cw:
+                    left = 0
+                elif w>cw:
+                    #left = random.uniform(w - cw)
+                    left = random.randint(0, w - cw)
+                else:
+                    left = random.randint(w - cw, 0)
+                if h==ch:
+                    top = 0
+                elif h>ch:
+                    #top = random.uniform(h - ch)
+                    top = random.randint(0, h - ch)
+                else:
+                    top = random.randint(h - ch, 0)
+
+                patch = np.array(
+                    (int(left), int(top), int(left + cw), int(top + ch)), dtype=np.int)
+
+                # center of boxes should inside the crop img
+                # only adjust boxes and instance masks when the gt is not empty
+                # adjust boxes
+                def is_center_of_bboxes_in_patch(boxes, patch):
+                    # TODO >=
+                    center = (boxes[:, :2] + boxes[:, 2:]) / 2
+                    mask = ((center[:, 0] > patch[0]) *
+                            (center[:, 1] > patch[1]) *
+                            (center[:, 0] < patch[2]) *
+                            (center[:, 1] < patch[3]))
+                    return mask
+
+                mask = is_center_of_bboxes_in_patch(boxes, patch)
+                if not mask.any():
+                    continue
+                for key in results.get('bbox_fields', []):
+                    boxes = results[key].copy()
+                    #print('BBB', key, boxes.shape)
+                    mask = is_center_of_bboxes_in_patch(boxes, patch)
+                    boxes = boxes[mask]
+                    if self.bbox_clip_border:
+                        boxes[:, 2:] = boxes[:, 2:].clip(max=patch[2:])
+                        boxes[:, :2] = boxes[:, :2].clip(min=patch[:2])
+                    boxes -= np.tile(patch[:2], 2)
+
+                    results[key] = boxes
+                    # labels
+                    label_key = self.bbox2label.get(key)
+                    if label_key in results:
+                        results[label_key] = results[label_key][mask]
+
+                    # keypoints field
+                    if key=='gt_bboxes':
+                        for kps_key in results.get('keypoints_fields', []):
+                            keypointss = results[kps_key].copy()
+                            #print('AAAA', kps_key, keypointss.shape, mask.shape)
+                            keypointss = keypointss[mask,:,:]
+                            if self.bbox_clip_border:
+                                keypointss[:,:,:2] = keypointss[:,:,:2].clip(max=patch[2:])
+                                keypointss[:,:,:2] = keypointss[:,:,:2].clip(min=patch[:2])
+                            #keypointss[:,:,:2] -= np.tile(patch[:2], 2)
+                            keypointss[:,:,0] -= patch[0]
+                            keypointss[:,:,1] -= patch[1]
+                            results[kps_key] = keypointss
+
+                    # mask fields
+                    mask_key = self.bbox2mask.get(key)
+                    if mask_key in results:
+                        results[mask_key] = results[mask_key][mask.nonzero()
+                                                              [0]].crop(patch)
+
+                # adjust the img no matter whether the gt is empty before crop
+                #img = img[patch[1]:patch[3], patch[0]:patch[2]]
+                rimg = np.ones( (ch, cw, 3), dtype=img.dtype) * 128
+                patch_from = patch.copy()
+                patch_from[0] = max(0, patch_from[0])
+                patch_from[1] = max(0, patch_from[1])
+                patch_from[2] = min(img.shape[1], patch_from[2])
+                patch_from[3] = min(img.shape[0], patch_from[3])
+                patch_to = patch.copy()
+                patch_to[0] = max(0, patch_to[0]*-1)
+                patch_to[1] = max(0, patch_to[1]*-1)
+                patch_to[2] = patch_to[0] + (patch_from[2] - patch_from[0])
+                patch_to[3] = patch_to[1] + (patch_from[3] - patch_from[1])
+                rimg[patch_to[1]:patch_to[3], patch_to[0]:patch_to[2],:] = img[patch_from[1]:patch_from[3], patch_from[0]:patch_from[2], :]
+                #print(img.shape, scale, patch, patch_from, patch_to, rimg.shape)
+                img = rimg
+                results['img'] = img
+                results['img_shape'] = img.shape
+
+                # seg fields
+                #for key in results.get('seg_fields', []):
+                #    results[key] = results[key][patch[1]:patch[3],
+                #                                patch[0]:patch[2]]
+                return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(min_ious={self.min_iou}, '
+        repr_str += f'crop_size={self.crop_size})'
+        return repr_str
+
+@PIPELINES.register_module()
+class SegRescale(object):
+    """Rescale semantic segmentation maps.
+
+    Args:
+        scale_factor (float): The scale factor of the final output.
+        backend (str): Image rescale backend, choices are 'cv2' and 'pillow'.
+            These two backends generates slightly different results. Defaults
+            to 'cv2'.
+    """
+
+    def __init__(self, scale_factor=1, backend='cv2'):
+        self.scale_factor = scale_factor
+        self.backend = backend
+
+    def __call__(self, results):
+        """Call function to scale the semantic segmentation map.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with semantic segmentation map scaled.
+        """
+
+        for key in results.get('seg_fields', []):
+            if self.scale_factor != 1:
+                results[key] = mmcv.imrescale(
+                    results[key],
+                    self.scale_factor,
+                    interpolation='nearest',
+                    backend=self.backend)
+        return results
+
+    def __repr__(self):
+        return self.__class__.__name__ + f'(scale_factor={self.scale_factor})'
+
+
+@PIPELINES.register_module()
+class PhotoMetricDistortion(object):
+    """Apply photometric distortion to image sequentially, every transformation
+    is applied with a probability of 0.5. The position of random contrast is in
+    second or second to last.
+
+    1. random brightness
+    2. random contrast (mode 0)
+    3. convert color from BGR to HSV
+    4. random saturation
+    5. random hue
+    6. convert color from HSV to BGR
+    7. random contrast (mode 1)
+    8. randomly swap channels
+    9. random grayscale
+
+    Args:
+        brightness_delta (int): delta of brightness.
+        contrast_range (tuple): range of contrast.
+        saturation_range (tuple): range of saturation.
+        hue_delta (int): delta of hue.
+        gray_prob (float): prob of grayscale.
+    """
+
+    def __init__(self,
+                 brightness_delta=32,
+                 contrast_range=(0.5, 1.5),
+                 saturation_range=(0.5, 1.5),
+                 hue_delta=18,
+                 gray_prob=0.0):
+        self.brightness_delta = brightness_delta
+        self.contrast_lower, self.contrast_upper = contrast_range
+        self.saturation_lower, self.saturation_upper = saturation_range
+        self.hue_delta = hue_delta
+        self.gray_prob = gray_prob
+
+    def __call__(self, results):
+        """Call function to perform photometric distortion on images.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with images distorted.
+        """
+
+        if 'img_fields' in results:
+            assert results['img_fields'] == ['img'], \
+                'Only single img_fields is allowed'
+        img = results['img']
+        assert img.dtype == np.float32, \
+            'PhotoMetricDistortion needs the input image of dtype np.float32,'\
+            ' please set "to_float32=True" in "LoadImageFromFile" pipeline'
+        # random brightness
+        if random.randint(2):
+            delta = random.uniform(-self.brightness_delta,
+                                   self.brightness_delta)
+            img += delta
+
+        # mode == 0 --> do random contrast first
+        # mode == 1 --> do random contrast last
+        mode = random.randint(2)
+        if mode == 1:
+            if random.randint(2):
+                alpha = random.uniform(self.contrast_lower,
+                                       self.contrast_upper)
+                img *= alpha
+
+        # convert color from BGR to HSV
+        img = mmcv.bgr2hsv(img)
+
+        # random saturation
+        if random.randint(2):
+            img[..., 1] *= random.uniform(self.saturation_lower,
+                                          self.saturation_upper)
+
+        # random hue
+        if random.randint(2):
+            img[..., 0] += random.uniform(-self.hue_delta, self.hue_delta)
+            img[..., 0][img[..., 0] > 360] -= 360
+            img[..., 0][img[..., 0] < 0] += 360
+
+        # convert color from HSV to BGR
+        img = mmcv.hsv2bgr(img)
+
+        # random contrast
+        if mode == 0:
+            if random.randint(2):
+                alpha = random.uniform(self.contrast_lower,
+                                       self.contrast_upper)
+                img *= alpha
+
+        # randomly swap channels
+        if random.randint(2):
+            img = img[..., random.permutation(3)]
+
+        if self.gray_prob>0.0:
+            if random.random()<self.gray_prob:
+                gray = mmcv.bgr2gray(img)
+                img = cv2.merge([gray, gray, gray])
+        results['img'] = img
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(\nbrightness_delta={self.brightness_delta},\n'
+        repr_str += 'contrast_range='
+        repr_str += f'{(self.contrast_lower, self.contrast_upper)},\n'
+        repr_str += 'saturation_range='
+        repr_str += f'{(self.saturation_lower, self.saturation_upper)},\n'
+        repr_str += f'hue_delta={self.hue_delta},\n'
+        repr_str += f'gray_prob={self.gray_prob})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Expand(object):
+    """Random expand the image & bboxes.
+
+    Randomly place the original image on a canvas of 'ratio' x original image
+    size filled with mean values. The ratio is in the range of ratio_range.
+
+    Args:
+        mean (tuple): mean value of dataset.
+        to_rgb (bool): if need to convert the order of mean to align with RGB.
+        ratio_range (tuple): range of expand ratio.
+        prob (float): probability of applying this transformation
+    """
+
+    def __init__(self,
+                 mean=(0, 0, 0),
+                 to_rgb=True,
+                 ratio_range=(1, 4),
+                 seg_ignore_label=None,
+                 prob=0.5):
+        self.to_rgb = to_rgb
+        self.ratio_range = ratio_range
+        if to_rgb:
+            self.mean = mean[::-1]
+        else:
+            self.mean = mean
+        self.min_ratio, self.max_ratio = ratio_range
+        self.seg_ignore_label = seg_ignore_label
+        self.prob = prob
+
+    def __call__(self, results):
+        """Call function to expand images, bounding boxes.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with images, bounding boxes expanded
+        """
+
+        if random.uniform(0, 1) > self.prob:
+            return results
+
+        if 'img_fields' in results:
+            assert results['img_fields'] == ['img'], \
+                'Only single img_fields is allowed'
+        img = results['img']
+
+        h, w, c = img.shape
+        ratio = random.uniform(self.min_ratio, self.max_ratio)
+        # speedup expand when meets large image
+        if np.all(self.mean == self.mean[0]):
+            expand_img = np.empty((int(h * ratio), int(w * ratio), c),
+                                  img.dtype)
+            expand_img.fill(self.mean[0])
+        else:
+            expand_img = np.full((int(h * ratio), int(w * ratio), c),
+                                 self.mean,
+                                 dtype=img.dtype)
+        left = int(random.uniform(0, w * ratio - w))
+        top = int(random.uniform(0, h * ratio - h))
+        expand_img[top:top + h, left:left + w] = img
+
+        results['img'] = expand_img
+        # expand bboxes
+        for key in results.get('bbox_fields', []):
+            results[key] = results[key] + np.tile(
+                (left, top), 2).astype(results[key].dtype)
+
+        # expand masks
+        for key in results.get('mask_fields', []):
+            results[key] = results[key].expand(
+                int(h * ratio), int(w * ratio), top, left)
+
+        # expand segs
+        for key in results.get('seg_fields', []):
+            gt_seg = results[key]
+            expand_gt_seg = np.full((int(h * ratio), int(w * ratio)),
+                                    self.seg_ignore_label,
+                                    dtype=gt_seg.dtype)
+            expand_gt_seg[top:top + h, left:left + w] = gt_seg
+            results[key] = expand_gt_seg
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(mean={self.mean}, to_rgb={self.to_rgb}, '
+        repr_str += f'ratio_range={self.ratio_range}, '
+        repr_str += f'seg_ignore_label={self.seg_ignore_label})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class MinIoURandomCrop(object):
+    """Random crop the image & bboxes, the cropped patches have minimum IoU
+    requirement with original image & bboxes, the IoU threshold is randomly
+    selected from min_ious.
+
+    Args:
+        min_ious (tuple): minimum IoU threshold for all intersections with
+        bounding boxes
+        min_crop_size (float): minimum crop's size (i.e. h,w := a*h, a*w,
+        where a >= min_crop_size).
+        bbox_clip_border (bool, optional): Whether clip the objects outside
+            the border of the image. Defaults to True.
+
+    Note:
+        The keys for bboxes, labels and masks should be paired. That is, \
+        `gt_bboxes` corresponds to `gt_labels` and `gt_masks`, and \
+        `gt_bboxes_ignore` to `gt_labels_ignore` and `gt_masks_ignore`.
+    """
+
+    def __init__(self,
+                 min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
+                 min_crop_size=0.3,
+                 bbox_clip_border=True):
+        # 1: return ori img
+        self.min_ious = min_ious
+        self.sample_mode = (1, *min_ious, 0)
+        self.min_crop_size = min_crop_size
+        self.bbox_clip_border = bbox_clip_border
+        self.bbox2label = {
+            'gt_bboxes': 'gt_labels',
+            'gt_bboxes_ignore': 'gt_labels_ignore'
+        }
+        self.bbox2mask = {
+            'gt_bboxes': 'gt_masks',
+            'gt_bboxes_ignore': 'gt_masks_ignore'
+        }
+
+    def __call__(self, results):
+        """Call function to crop images and bounding boxes with minimum IoU
+        constraint.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with images and bounding boxes cropped, \
+                'img_shape' key is updated.
+        """
+
+        if 'img_fields' in results:
+            assert results['img_fields'] == ['img'], \
+                'Only single img_fields is allowed'
+        img = results['img']
+        assert 'bbox_fields' in results
+        boxes = [results[key] for key in results['bbox_fields']]
+        boxes = np.concatenate(boxes, 0)
+        h, w, c = img.shape
+        while True:
+            mode = random.choice(self.sample_mode)
+            self.mode = mode
+            if mode == 1:
+                return results
+
+            min_iou = mode
+            for i in range(50):
+                new_w = random.uniform(self.min_crop_size * w, w)
+                new_h = random.uniform(self.min_crop_size * h, h)
+
+                # h / w in [0.5, 2]
+                if new_h / new_w < 0.5 or new_h / new_w > 2:
+                    continue
+
+                left = random.uniform(w - new_w)
+                top = random.uniform(h - new_h)
+
+                patch = np.array(
+                    (int(left), int(top), int(left + new_w), int(top + new_h)))
+                # Line or point crop is not allowed
+                if patch[2] == patch[0] or patch[3] == patch[1]:
+                    continue
+                overlaps = bbox_overlaps(
+                    patch.reshape(-1, 4), boxes.reshape(-1, 4)).reshape(-1)
+                if len(overlaps) > 0 and overlaps.min() < min_iou:
+                    continue
+
+                # center of boxes should inside the crop img
+                # only adjust boxes and instance masks when the gt is not empty
+                if len(overlaps) > 0:
+                    # adjust boxes
+                    def is_center_of_bboxes_in_patch(boxes, patch):
+                        center = (boxes[:, :2] + boxes[:, 2:]) / 2
+                        mask = ((center[:, 0] > patch[0]) *
+                                (center[:, 1] > patch[1]) *
+                                (center[:, 0] < patch[2]) *
+                                (center[:, 1] < patch[3]))
+                        return mask
+
+                    mask = is_center_of_bboxes_in_patch(boxes, patch)
+                    if not mask.any():
+                        continue
+                    for key in results.get('bbox_fields', []):
+                        boxes = results[key].copy()
+                        mask = is_center_of_bboxes_in_patch(boxes, patch)
+                        boxes = boxes[mask]
+                        if self.bbox_clip_border:
+                            boxes[:, 2:] = boxes[:, 2:].clip(max=patch[2:])
+                            boxes[:, :2] = boxes[:, :2].clip(min=patch[:2])
+                        boxes -= np.tile(patch[:2], 2)
+
+                        results[key] = boxes
+                        # labels
+                        label_key = self.bbox2label.get(key)
+                        if label_key in results:
+                            results[label_key] = results[label_key][mask]
+
+                        # mask fields
+                        mask_key = self.bbox2mask.get(key)
+                        if mask_key in results:
+                            results[mask_key] = results[mask_key][
+                                mask.nonzero()[0]].crop(patch)
+                # adjust the img no matter whether the gt is empty before crop
+                img = img[patch[1]:patch[3], patch[0]:patch[2]]
+                results['img'] = img
+                results['img_shape'] = img.shape
+
+                # seg fields
+                for key in results.get('seg_fields', []):
+                    results[key] = results[key][patch[1]:patch[3],
+                                                patch[0]:patch[2]]
+                return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(min_ious={self.min_ious}, '
+        repr_str += f'min_crop_size={self.min_crop_size}), '
+        repr_str += f'bbox_clip_border={self.bbox_clip_border})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Corrupt(object):
+    """Corruption augmentation.
+
+    Corruption transforms implemented based on
+    `imagecorruptions <https://github.com/bethgelab/imagecorruptions>`_.
+
+    Args:
+        corruption (str): Corruption name.
+        severity (int, optional): The severity of corruption. Default: 1.
+    """
+
+    def __init__(self, corruption, severity=1):
+        self.corruption = corruption
+        self.severity = severity
+
+    def __call__(self, results):
+        """Call function to corrupt image.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Result dict with images corrupted.
+        """
+
+        if corrupt is None:
+            raise RuntimeError('imagecorruptions is not installed')
+        if 'img_fields' in results:
+            assert results['img_fields'] == ['img'], \
+                'Only single img_fields is allowed'
+        results['img'] = corrupt(
+            results['img'].astype(np.uint8),
+            corruption_name=self.corruption,
+            severity=self.severity)
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(corruption={self.corruption}, '
+        repr_str += f'severity={self.severity})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class Albu(object):
+    """Albumentation augmentation.
+
+    Adds custom transformations from Albumentations library.
+    Please, visit `https://albumentations.readthedocs.io`
+    to get more information.
+
+    An example of ``transforms`` is as followed:
+
+    .. code-block::
+
+        [
+            dict(
+                type='ShiftScaleRotate',
+                shift_limit=0.0625,
+                scale_limit=0.0,
+                rotate_limit=0,
+                interpolation=1,
+                p=0.5),
+            dict(
+                type='RandomBrightnessContrast',
+                brightness_limit=[0.1, 0.3],
+                contrast_limit=[0.1, 0.3],
+                p=0.2),
+            dict(type='ChannelShuffle', p=0.1),
+            dict(
+                type='OneOf',
+                transforms=[
+                    dict(type='Blur', blur_limit=3, p=1.0),
+                    dict(type='MedianBlur', blur_limit=3, p=1.0)
+                ],
+                p=0.1),
+        ]
+
+    Args:
+        transforms (list[dict]): A list of albu transformations
+        bbox_params (dict): Bbox_params for albumentation `Compose`
+        keymap (dict): Contains {'input key':'albumentation-style key'}
+        skip_img_without_anno (bool): Whether to skip the image if no ann left
+            after aug
+    """
+
+    def __init__(self,
+                 transforms,
+                 bbox_params=None,
+                 keymap=None,
+                 update_pad_shape=False,
+                 skip_img_without_anno=False):
+        if Compose is None:
+            raise RuntimeError('albumentations is not installed')
+
+        self.transforms = transforms
+        self.filter_lost_elements = False
+        self.update_pad_shape = update_pad_shape
+        self.skip_img_without_anno = skip_img_without_anno
+
+        # A simple workaround to remove masks without boxes
+        if (isinstance(bbox_params, dict) and 'label_fields' in bbox_params
+                and 'filter_lost_elements' in bbox_params):
+            self.filter_lost_elements = True
+            self.origin_label_fields = bbox_params['label_fields']
+            bbox_params['label_fields'] = ['idx_mapper']
+            del bbox_params['filter_lost_elements']
+
+        self.bbox_params = (
+            self.albu_builder(bbox_params) if bbox_params else None)
+        self.aug = Compose([self.albu_builder(t) for t in self.transforms],
+                           bbox_params=self.bbox_params)
+
+        if not keymap:
+            self.keymap_to_albu = {
+                'img': 'image',
+                'gt_masks': 'masks',
+                'gt_bboxes': 'bboxes'
+            }
+        else:
+            self.keymap_to_albu = keymap
+        self.keymap_back = {v: k for k, v in self.keymap_to_albu.items()}
+
+    def albu_builder(self, cfg):
+        """Import a module from albumentations.
+
+        It inherits some of :func:`build_from_cfg` logic.
+
+        Args:
+            cfg (dict): Config dict. It should at least contain the key "type".
+
+        Returns:
+            obj: The constructed object.
+        """
+
+        assert isinstance(cfg, dict) and 'type' in cfg
+        args = cfg.copy()
+
+        obj_type = args.pop('type')
+        if mmcv.is_str(obj_type):
+            if albumentations is None:
+                raise RuntimeError('albumentations is not installed')
+            obj_cls = getattr(albumentations, obj_type)
+        elif inspect.isclass(obj_type):
+            obj_cls = obj_type
+        else:
+            raise TypeError(
+                f'type must be a str or valid type, but got {type(obj_type)}')
+
+        if 'transforms' in args:
+            args['transforms'] = [
+                self.albu_builder(transform)
+                for transform in args['transforms']
+            ]
+
+        return obj_cls(**args)
+
+    @staticmethod
+    def mapper(d, keymap):
+        """Dictionary mapper. Renames keys according to keymap provided.
+
+        Args:
+            d (dict): old dict
+            keymap (dict): {'old_key':'new_key'}
+        Returns:
+            dict: new dict.
+        """
+
+        updated_dict = {}
+        for k, v in zip(d.keys(), d.values()):
+            new_k = keymap.get(k, k)
+            updated_dict[new_k] = d[k]
+        return updated_dict
+
+    def __call__(self, results):
+        # dict to albumentations format
+        results = self.mapper(results, self.keymap_to_albu)
+        # TODO: add bbox_fields
+        if 'bboxes' in results:
+            # to list of boxes
+            if isinstance(results['bboxes'], np.ndarray):
+                results['bboxes'] = [x for x in results['bboxes']]
+            # add pseudo-field for filtration
+            if self.filter_lost_elements:
+                results['idx_mapper'] = np.arange(len(results['bboxes']))
+
+        # TODO: Support mask structure in albu
+        if 'masks' in results:
+            if isinstance(results['masks'], PolygonMasks):
+                raise NotImplementedError(
+                    'Albu only supports BitMap masks now')
+            ori_masks = results['masks']
+            if albumentations.__version__ < '0.5':
+                results['masks'] = results['masks'].masks
+            else:
+                results['masks'] = [mask for mask in results['masks'].masks]
+
+        results = self.aug(**results)
+
+        if 'bboxes' in results:
+            if isinstance(results['bboxes'], list):
+                results['bboxes'] = np.array(
+                    results['bboxes'], dtype=np.float32)
+            results['bboxes'] = results['bboxes'].reshape(-1, 4)
+
+            # filter label_fields
+            if self.filter_lost_elements:
+
+                for label in self.origin_label_fields:
+                    results[label] = np.array(
+                        [results[label][i] for i in results['idx_mapper']])
+                if 'masks' in results:
+                    results['masks'] = np.array(
+                        [results['masks'][i] for i in results['idx_mapper']])
+                    results['masks'] = ori_masks.__class__(
+                        results['masks'], results['image'].shape[0],
+                        results['image'].shape[1])
+
+                if (not len(results['idx_mapper'])
+                        and self.skip_img_without_anno):
+                    return None
+
+        if 'gt_labels' in results:
+            if isinstance(results['gt_labels'], list):
+                results['gt_labels'] = np.array(results['gt_labels'])
+            results['gt_labels'] = results['gt_labels'].astype(np.int64)
+
+        # back to the original format
+        results = self.mapper(results, self.keymap_back)
+
+        # update final shape
+        if self.update_pad_shape:
+            results['pad_shape'] = results['img'].shape
+
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__ + f'(transforms={self.transforms})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class RandomCenterCropPad(object):
+    """Random center crop and random around padding for CornerNet.
+
+    This operation generates randomly cropped image from the original image and
+    pads it simultaneously. Different from :class:`RandomCrop`, the output
+    shape may not equal to ``crop_size`` strictly. We choose a random value
+    from ``ratios`` and the output shape could be larger or smaller than
+    ``crop_size``. The padding operation is also different from :class:`Pad`,
+    here we use around padding instead of right-bottom padding.
+
+    The relation between output image (padding image) and original image:
+
+    .. code:: text
+
+                        output image
+
+               +----------------------------+
+               |          padded area       |
+        +------|----------------------------|----------+
+        |      |         cropped area       |          |
+        |      |         +---------------+  |          |
+        |      |         |    .   center |  |          | original image
+        |      |         |        range  |  |          |
+        |      |         +---------------+  |          |
+        +------|----------------------------|----------+
+               |          padded area       |
+               +----------------------------+
+
+    There are 5 main areas in the figure:
+
+    - output image: output image of this operation, also called padding
+      image in following instruction.
+    - original image: input image of this operation.
+    - padded area: non-intersect area of output image and original image.
+    - cropped area: the overlap of output image and original image.
+    - center range: a smaller area where random center chosen from.
+      center range is computed by ``border`` and original image's shape
+      to avoid our random center is too close to original image's border.
+
+    Also this operation act differently in train and test mode, the summary
+    pipeline is listed below.
+
+    Train pipeline:
+
+    1. Choose a ``random_ratio`` from ``ratios``, the shape of padding image
+       will be ``random_ratio * crop_size``.
+    2. Choose a ``random_center`` in center range.
+    3. Generate padding image with center matches the ``random_center``.
+    4. Initialize the padding image with pixel value equals to ``mean``.
+    5. Copy the cropped area to padding image.
+    6. Refine annotations.
+
+    Test pipeline:
+
+    1. Compute output shape according to ``test_pad_mode``.
+    2. Generate padding image with center matches the original image
+       center.
+    3. Initialize the padding image with pixel value equals to ``mean``.
+    4. Copy the ``cropped area`` to padding image.
+
+    Args:
+        crop_size (tuple | None): expected size after crop, final size will
+            computed according to ratio. Requires (h, w) in train mode, and
+            None in test mode.
+        ratios (tuple): random select a ratio from tuple and crop image to
+            (crop_size[0] * ratio) * (crop_size[1] * ratio).
+            Only available in train mode.
+        border (int): max distance from center select area to image border.
+            Only available in train mode.
+        mean (sequence): Mean values of 3 channels.
+        std (sequence): Std values of 3 channels.
+        to_rgb (bool): Whether to convert the image from BGR to RGB.
+        test_mode (bool): whether involve random variables in transform.
+            In train mode, crop_size is fixed, center coords and ratio is
+            random selected from predefined lists. In test mode, crop_size
+            is image's original shape, center coords and ratio is fixed.
+        test_pad_mode (tuple): padding method and padding shape value, only
+            available in test mode. Default is using 'logical_or' with
+            127 as padding shape value.
+
+            - 'logical_or': final_shape = input_shape | padding_shape_value
+            - 'size_divisor': final_shape = int(
+              ceil(input_shape / padding_shape_value) * padding_shape_value)
+        bbox_clip_border (bool, optional): Whether clip the objects outside
+            the border of the image. Defaults to True.
+    """
+
+    def __init__(self,
+                 crop_size=None,
+                 ratios=(0.9, 1.0, 1.1),
+                 border=128,
+                 mean=None,
+                 std=None,
+                 to_rgb=None,
+                 test_mode=False,
+                 test_pad_mode=('logical_or', 127),
+                 bbox_clip_border=True):
+        if test_mode:
+            assert crop_size is None, 'crop_size must be None in test mode'
+            assert ratios is None, 'ratios must be None in test mode'
+            assert border is None, 'border must be None in test mode'
+            assert isinstance(test_pad_mode, (list, tuple))
+            assert test_pad_mode[0] in ['logical_or', 'size_divisor']
+        else:
+            assert isinstance(crop_size, (list, tuple))
+            assert crop_size[0] > 0 and crop_size[1] > 0, (
+                'crop_size must > 0 in train mode')
+            assert isinstance(ratios, (list, tuple))
+            assert test_pad_mode is None, (
+                'test_pad_mode must be None in train mode')
+
+        self.crop_size = crop_size
+        self.ratios = ratios
+        self.border = border
+        # We do not set default value to mean, std and to_rgb because these
+        # hyper-parameters are easy to forget but could affect the performance.
+        # Please use the same setting as Normalize for performance assurance.
+        assert mean is not None and std is not None and to_rgb is not None
+        self.to_rgb = to_rgb
+        self.input_mean = mean
+        self.input_std = std
+        if to_rgb:
+            self.mean = mean[::-1]
+            self.std = std[::-1]
+        else:
+            self.mean = mean
+            self.std = std
+        self.test_mode = test_mode
+        self.test_pad_mode = test_pad_mode
+        self.bbox_clip_border = bbox_clip_border
+
+    def _get_border(self, border, size):
+        """Get final border for the target size.
+
+        This function generates a ``final_border`` according to image's shape.
+        The area between ``final_border`` and ``size - final_border`` is the
+        ``center range``. We randomly choose center from the ``center range``
+        to avoid our random center is too close to original image's border.
+        Also ``center range`` should be larger than 0.
+
+        Args:
+            border (int): The initial border, default is 128.
+            size (int): The width or height of original image.
+        Returns:
+            int: The final border.
+        """
+        k = 2 * border / size
+        i = pow(2, np.ceil(np.log2(np.ceil(k))) + (k == int(k)))
+        return border // i
+
+    def _filter_boxes(self, patch, boxes):
+        """Check whether the center of each box is in the patch.
+
+        Args:
+            patch (list[int]): The cropped area, [left, top, right, bottom].
+            boxes (numpy array, (N x 4)): Ground truth boxes.
+
+        Returns:
+            mask (numpy array, (N,)): Each box is inside or outside the patch.
+        """
+        center = (boxes[:, :2] + boxes[:, 2:]) / 2
+        mask = (center[:, 0] > patch[0]) * (center[:, 1] > patch[1]) * (
+            center[:, 0] < patch[2]) * (
+                center[:, 1] < patch[3])
+        return mask
+
+    def _crop_image_and_paste(self, image, center, size):
+        """Crop image with a given center and size, then paste the cropped
+        image to a blank image with two centers align.
+
+        This function is equivalent to generating a blank image with ``size``
+        as its shape. Then cover it on the original image with two centers (
+        the center of blank image and the random center of original image)
+        aligned. The overlap area is paste from the original image and the
+        outside area is filled with ``mean pixel``.
+
+        Args:
+            image (np array, H x W x C): Original image.
+            center (list[int]): Target crop center coord.
+            size (list[int]): Target crop size. [target_h, target_w]
+
+        Returns:
+            cropped_img (np array, target_h x target_w x C): Cropped image.
+            border (np array, 4): The distance of four border of
+                ``cropped_img`` to the original image area, [top, bottom,
+                left, right]
+            patch (list[int]): The cropped area, [left, top, right, bottom].
+        """
+        center_y, center_x = center
+        target_h, target_w = size
+        img_h, img_w, img_c = image.shape
+
+        x0 = max(0, center_x - target_w // 2)
+        x1 = min(center_x + target_w // 2, img_w)
+        y0 = max(0, center_y - target_h // 2)
+        y1 = min(center_y + target_h // 2, img_h)
+        patch = np.array((int(x0), int(y0), int(x1), int(y1)))
+
+        left, right = center_x - x0, x1 - center_x
+        top, bottom = center_y - y0, y1 - center_y
+
+        cropped_center_y, cropped_center_x = target_h // 2, target_w // 2
+        cropped_img = np.zeros((target_h, target_w, img_c), dtype=image.dtype)
+        for i in range(img_c):
+            cropped_img[:, :, i] += self.mean[i]
+        y_slice = slice(cropped_center_y - top, cropped_center_y + bottom)
+        x_slice = slice(cropped_center_x - left, cropped_center_x + right)
+        cropped_img[y_slice, x_slice, :] = image[y0:y1, x0:x1, :]
+
+        border = np.array([
+            cropped_center_y - top, cropped_center_y + bottom,
+            cropped_center_x - left, cropped_center_x + right
+        ],
+                          dtype=np.float32)
+
+        return cropped_img, border, patch
+
+    def _train_aug(self, results):
+        """Random crop and around padding the original image.
+
+        Args:
+            results (dict): Image infomations in the augment pipeline.
+
+        Returns:
+            results (dict): The updated dict.
+        """
+        img = results['img']
+        h, w, c = img.shape
+        boxes = results['gt_bboxes']
+        while True:
+            scale = random.choice(self.ratios)
+            new_h = int(self.crop_size[0] * scale)
+            new_w = int(self.crop_size[1] * scale)
+            h_border = self._get_border(self.border, h)
+            w_border = self._get_border(self.border, w)
+
+            for i in range(50):
+                center_x = random.randint(low=w_border, high=w - w_border)
+                center_y = random.randint(low=h_border, high=h - h_border)
+
+                cropped_img, border, patch = self._crop_image_and_paste(
+                    img, [center_y, center_x], [new_h, new_w])
+
+                mask = self._filter_boxes(patch, boxes)
+                # if image do not have valid bbox, any crop patch is valid.
+                if not mask.any() and len(boxes) > 0:
+                    continue
+
+                results['img'] = cropped_img
+                results['img_shape'] = cropped_img.shape
+                results['pad_shape'] = cropped_img.shape
+
+                x0, y0, x1, y1 = patch
+
+                left_w, top_h = center_x - x0, center_y - y0
+                cropped_center_x, cropped_center_y = new_w // 2, new_h // 2
+
+                # crop bboxes accordingly and clip to the image boundary
+                for key in results.get('bbox_fields', []):
+                    mask = self._filter_boxes(patch, results[key])
+                    bboxes = results[key][mask]
+                    bboxes[:, 0:4:2] += cropped_center_x - left_w - x0
+                    bboxes[:, 1:4:2] += cropped_center_y - top_h - y0
+                    if self.bbox_clip_border:
+                        bboxes[:, 0:4:2] = np.clip(bboxes[:, 0:4:2], 0, new_w)
+                        bboxes[:, 1:4:2] = np.clip(bboxes[:, 1:4:2], 0, new_h)
+                    keep = (bboxes[:, 2] > bboxes[:, 0]) & (
+                        bboxes[:, 3] > bboxes[:, 1])
+                    bboxes = bboxes[keep]
+                    results[key] = bboxes
+                    if key in ['gt_bboxes']:
+                        if 'gt_labels' in results:
+                            labels = results['gt_labels'][mask]
+                            labels = labels[keep]
+                            results['gt_labels'] = labels
+                        if 'gt_masks' in results:
+                            raise NotImplementedError(
+                                'RandomCenterCropPad only supports bbox.')
+
+                # crop semantic seg
+                for key in results.get('seg_fields', []):
+                    raise NotImplementedError(
+                        'RandomCenterCropPad only supports bbox.')
+                return results
+
+    def _test_aug(self, results):
+        """Around padding the original image without cropping.
+
+        The padding mode and value are from ``test_pad_mode``.
+
+        Args:
+            results (dict): Image infomations in the augment pipeline.
+
+        Returns:
+            results (dict): The updated dict.
+        """
+        img = results['img']
+        h, w, c = img.shape
+        results['img_shape'] = img.shape
+        if self.test_pad_mode[0] in ['logical_or']:
+            target_h = h | self.test_pad_mode[1]
+            target_w = w | self.test_pad_mode[1]
+        elif self.test_pad_mode[0] in ['size_divisor']:
+            divisor = self.test_pad_mode[1]
+            target_h = int(np.ceil(h / divisor)) * divisor
+            target_w = int(np.ceil(w / divisor)) * divisor
+        else:
+            raise NotImplementedError(
+                'RandomCenterCropPad only support two testing pad mode:'
+                'logical-or and size_divisor.')
+
+        cropped_img, border, _ = self._crop_image_and_paste(
+            img, [h // 2, w // 2], [target_h, target_w])
+        results['img'] = cropped_img
+        results['pad_shape'] = cropped_img.shape
+        results['border'] = border
+        return results
+
+    def __call__(self, results):
+        img = results['img']
+        assert img.dtype == np.float32, (
+            'RandomCenterCropPad needs the input image of dtype np.float32,'
+            ' please set "to_float32=True" in "LoadImageFromFile" pipeline')
+        h, w, c = img.shape
+        assert c == len(self.mean)
+        if self.test_mode:
+            return self._test_aug(results)
+        else:
+            return self._train_aug(results)
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(crop_size={self.crop_size}, '
+        repr_str += f'ratios={self.ratios}, '
+        repr_str += f'border={self.border}, '
+        repr_str += f'mean={self.input_mean}, '
+        repr_str += f'std={self.input_std}, '
+        repr_str += f'to_rgb={self.to_rgb}, '
+        repr_str += f'test_mode={self.test_mode}, '
+        repr_str += f'test_pad_mode={self.test_pad_mode}), '
+        repr_str += f'bbox_clip_border={self.bbox_clip_border})'
+        return repr_str
+
+
+@PIPELINES.register_module()
+class CutOut(object):
+    """CutOut operation.
+
+    Randomly drop some regions of image used in
+    `Cutout <https://arxiv.org/abs/1708.04552>`_.
+
+    Args:
+        n_holes (int | tuple[int, int]): Number of regions to be dropped.
+            If it is given as a list, number of holes will be randomly
+            selected from the closed interval [`n_holes[0]`, `n_holes[1]`].
+        cutout_shape (tuple[int, int] | list[tuple[int, int]]): The candidate
+            shape of dropped regions. It can be `tuple[int, int]` to use a
+            fixed cutout shape, or `list[tuple[int, int]]` to randomly choose
+            shape from the list.
+        cutout_ratio (tuple[float, float] | list[tuple[float, float]]): The
+            candidate ratio of dropped regions. It can be `tuple[float, float]`
+            to use a fixed ratio or `list[tuple[float, float]]` to randomly
+            choose ratio from the list. Please note that `cutout_shape`
+            and `cutout_ratio` cannot be both given at the same time.
+        fill_in (tuple[float, float, float] | tuple[int, int, int]): The value
+            of pixel to fill in the dropped regions. Default: (0, 0, 0).
+    """
+
+    def __init__(self,
+                 n_holes,
+                 cutout_shape=None,
+                 cutout_ratio=None,
+                 fill_in=(0, 0, 0)):
+
+        assert (cutout_shape is None) ^ (cutout_ratio is None), \
+            'Either cutout_shape or cutout_ratio should be specified.'
+        assert (isinstance(cutout_shape, (list, tuple))
+                or isinstance(cutout_ratio, (list, tuple)))
+        if isinstance(n_holes, tuple):
+            assert len(n_holes) == 2 and 0 <= n_holes[0] < n_holes[1]
+        else:
+            n_holes = (n_holes, n_holes)
+        self.n_holes = n_holes
+        self.fill_in = fill_in
+        self.with_ratio = cutout_ratio is not None
+        self.candidates = cutout_ratio if self.with_ratio else cutout_shape
+        if not isinstance(self.candidates, list):
+            self.candidates = [self.candidates]
+
+    def __call__(self, results):
+        """Call function to drop some regions of image."""
+        h, w, c = results['img'].shape
+        n_holes = np.random.randint(self.n_holes[0], self.n_holes[1] + 1)
+        for _ in range(n_holes):
+            x1 = np.random.randint(0, w)
+            y1 = np.random.randint(0, h)
+            index = np.random.randint(0, len(self.candidates))
+            if not self.with_ratio:
+                cutout_w, cutout_h = self.candidates[index]
+            else:
+                cutout_w = int(self.candidates[index][0] * w)
+                cutout_h = int(self.candidates[index][1] * h)
+
+            x2 = np.clip(x1 + cutout_w, 0, w)
+            y2 = np.clip(y1 + cutout_h, 0, h)
+            results['img'][y1:y2, x1:x2, :] = self.fill_in
+
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(n_holes={self.n_holes}, '
+        repr_str += (f'cutout_ratio={self.candidates}, ' if self.with_ratio
+                     else f'cutout_shape={self.candidates}, ')
+        repr_str += f'fill_in={self.fill_in})'
+        return repr_str
diff --git a/detection/scrfd/mmdet/datasets/retinaface.py b/detection/scrfd/mmdet/datasets/retinaface.py
new file mode 100755
index 000000000..ed9c8b4f0
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/retinaface.py
@@ -0,0 +1,169 @@
+import itertools
+import logging
+import os.path as osp
+import tempfile
+from collections import OrderedDict
+
+import mmcv
+import numpy as np
+from mmcv.utils import print_log
+from terminaltables import AsciiTable
+
+from mmdet.core import eval_recalls
+from .builder import DATASETS
+from .custom import CustomDataset
+
+try:
+    import pycocotools
+    if not hasattr(pycocotools, '__sphinx_mock__'):  # for doc generation
+        assert pycocotools.__version__ >= '12.0.2'
+except AssertionError:
+    raise AssertionError('Incompatible version of pycocotools is installed. '
+                         'Run pip uninstall pycocotools first. Then run pip '
+                         'install mmpycocotools to install open-mmlab forked '
+                         'pycocotools.')
+
+@DATASETS.register_module()
+class RetinaFaceDataset(CustomDataset):
+
+    CLASSES = ('FG', )
+    def __init__(self, min_size=None, **kwargs):
+        self.NK = 5
+        self.cat2label = {cat: i for i, cat in enumerate(self.CLASSES)}
+        self.min_size = min_size
+        self.gt_path = kwargs.get('gt_path')
+        super(RetinaFaceDataset, self).__init__(**kwargs)
+        #print(self.cat2label)
+
+    def _parse_ann_line(self, line):
+        values = [float(x) for x in line.strip().split()]
+        bbox = np.array(values[0:4], dtype=np.float32 )
+        kps = np.zeros( (self.NK,3), dtype=np.float32 )
+        ignore = False
+        if self.min_size is not None:
+            assert not self.test_mode
+            w = bbox[2] - bbox[0]
+            h = bbox[3] - bbox[1]
+            if w < self.min_size or h < self.min_size:
+                ignore = True
+        if len(values)>4:
+            if len(values)>5:
+                #print(values)
+                kps = np.array( values[4:19], dtype=np.float32 ).reshape((self.NK,3))
+                for li in range(kps.shape[0]):
+                    if (kps[li,:]==-1).all():
+                        #assert kps[li][2]==-1
+                        kps[li][2] = 0.0 #weight = 0, ignore
+                    else:
+                        assert kps[li][2]>=0
+                        kps[li][2] = 1.0 #weight
+                        #if li==0:
+                        #  landmark_num+=1
+                        #if kps[li][2]==0.0:#visible
+                        #  kps[li][2] = 1.0
+                        #else:
+                        #  kps[li][2] = 0.0
+            else: #len(values)==5
+                if not ignore:
+                    ignore = (values[4]==1)
+        else:
+            assert self.test_mode
+
+        return dict(bbox=bbox, kps=kps, ignore=ignore, cat='FG')
+
+
+    def load_annotations(self, ann_file):
+        """Load annotation from COCO style annotation file.
+
+        Args:
+            ann_file (str): Path of annotation file.
+
+        Returns:
+            list[dict]: Annotation info from COCO api.
+        """
+        name = None
+        bbox_map = {}
+        for line in open(ann_file, 'r'):
+            line = line.strip()
+            if line.startswith('#'):
+                value = line[1:].strip().split()
+                name = value[0]
+                width = int(value[1])
+                height = int(value[2])
+
+                bbox_map[name] = dict(width=width, height=height, objs=[])
+                continue
+            assert name is not None
+            assert name in bbox_map
+            bbox_map[name]['objs'].append(line)
+        print('origin image size', len(bbox_map))
+        data_infos = []
+        for name in bbox_map:
+            item = bbox_map[name]
+            width = item['width']
+            height = item['height']
+            vals = item['objs']
+            objs = []
+            for line in vals:
+                data = self._parse_ann_line(line)
+                if data is None:
+                    continue
+                objs.append( data ) #data is (bbox, kps, cat)
+            if len(objs)==0 and not self.test_mode:
+                continue
+            data_infos.append(dict(filename=name, width = width, height=height, objs = objs))
+        return data_infos
+
+
+    def get_ann_info(self, idx):
+        """Get COCO annotation by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Annotation info of specified index.
+        """
+        data_info = self.data_infos[idx]
+
+        bboxes = []
+        keypointss = []
+        labels = []
+        bboxes_ignore = []
+        labels_ignore = []
+        for obj in data_info['objs']:
+            label = self.cat2label[obj['cat']]
+            bbox = obj['bbox']
+            keypoints = obj['kps']
+            ignore = obj['ignore']
+            if ignore:
+                bboxes_ignore.append(bbox)
+                labels_ignore.append(label)
+            else:
+                bboxes.append(bbox)
+                labels.append(label)
+                keypointss.append(keypoints)
+        if not bboxes:
+            bboxes = np.zeros((0, 4))
+            labels = np.zeros((0, ))
+            keypointss = np.zeros((0, self.NK, 3))
+        else:
+            #bboxes = np.array(bboxes, ndmin=2) - 1
+            bboxes = np.array(bboxes, ndmin=2)
+            labels = np.array(labels)
+            keypointss = np.array(keypointss, ndmin=3)
+        if not bboxes_ignore:
+            bboxes_ignore = np.zeros((0, 4))
+            labels_ignore = np.zeros((0, ))
+        else:
+            #bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1
+            bboxes_ignore = np.array(bboxes_ignore, ndmin=2)
+            labels_ignore = np.array(labels_ignore)
+        ann = dict(
+            bboxes=bboxes.astype(np.float32),
+            labels=labels.astype(np.int64),
+            keypointss = keypointss.astype(np.float32),
+            bboxes_ignore=bboxes_ignore.astype(np.float32),
+            labels_ignore=labels_ignore.astype(np.int64))
+        return ann
+
diff --git a/detection/scrfd/mmdet/datasets/samplers/__init__.py b/detection/scrfd/mmdet/datasets/samplers/__init__.py
new file mode 100755
index 000000000..2596aeb2c
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/samplers/__init__.py
@@ -0,0 +1,4 @@
+from .distributed_sampler import DistributedSampler
+from .group_sampler import DistributedGroupSampler, GroupSampler
+
+__all__ = ['DistributedSampler', 'DistributedGroupSampler', 'GroupSampler']
diff --git a/detection/scrfd/mmdet/datasets/samplers/distributed_sampler.py b/detection/scrfd/mmdet/datasets/samplers/distributed_sampler.py
new file mode 100755
index 000000000..a9a1fc0b1
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/samplers/distributed_sampler.py
@@ -0,0 +1,32 @@
+import math
+
+import torch
+from torch.utils.data import DistributedSampler as _DistributedSampler
+
+
+class DistributedSampler(_DistributedSampler):
+
+    def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank)
+        self.shuffle = shuffle
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = torch.arange(len(self.dataset)).tolist()
+
+        # add extra samples to make it evenly divisible
+        # in case that indices is shorter than half of total_size
+        indices = (indices *
+                   math.ceil(self.total_size / len(indices)))[:self.total_size]
+        assert len(indices) == self.total_size
+
+        # subsample
+        indices = indices[self.rank:self.total_size:self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        return iter(indices)
diff --git a/detection/scrfd/mmdet/datasets/samplers/group_sampler.py b/detection/scrfd/mmdet/datasets/samplers/group_sampler.py
new file mode 100755
index 000000000..a691b949d
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/samplers/group_sampler.py
@@ -0,0 +1,143 @@
+from __future__ import division
+import math
+
+import numpy as np
+import torch
+from mmcv.runner import get_dist_info
+from torch.utils.data import Sampler
+
+
+class GroupSampler(Sampler):
+
+    def __init__(self, dataset, samples_per_gpu=1):
+        assert hasattr(dataset, 'flag')
+        self.dataset = dataset
+        self.samples_per_gpu = samples_per_gpu
+        self.flag = dataset.flag.astype(np.int64)
+        self.group_sizes = np.bincount(self.flag)
+        self.num_samples = 0
+        for i, size in enumerate(self.group_sizes):
+            self.num_samples += int(np.ceil(
+                size / self.samples_per_gpu)) * self.samples_per_gpu
+
+    def __iter__(self):
+        indices = []
+        for i, size in enumerate(self.group_sizes):
+            if size == 0:
+                continue
+            indice = np.where(self.flag == i)[0]
+            assert len(indice) == size
+            np.random.shuffle(indice)
+            num_extra = int(np.ceil(size / self.samples_per_gpu)
+                            ) * self.samples_per_gpu - len(indice)
+            indice = np.concatenate(
+                [indice, np.random.choice(indice, num_extra)])
+            indices.append(indice)
+        indices = np.concatenate(indices)
+        indices = [
+            indices[i * self.samples_per_gpu:(i + 1) * self.samples_per_gpu]
+            for i in np.random.permutation(
+                range(len(indices) // self.samples_per_gpu))
+        ]
+        indices = np.concatenate(indices)
+        indices = indices.astype(np.int64).tolist()
+        assert len(indices) == self.num_samples
+        return iter(indices)
+
+    def __len__(self):
+        return self.num_samples
+
+
+class DistributedGroupSampler(Sampler):
+    """Sampler that restricts data loading to a subset of the dataset.
+
+    It is especially useful in conjunction with
+    :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each
+    process can pass a DistributedSampler instance as a DataLoader sampler,
+    and load a subset of the original dataset that is exclusive to it.
+
+    .. note::
+        Dataset is assumed to be of constant size.
+
+    Arguments:
+        dataset: Dataset used for sampling.
+        num_replicas (optional): Number of processes participating in
+            distributed training.
+        rank (optional): Rank of the current process within num_replicas.
+    """
+
+    def __init__(self,
+                 dataset,
+                 samples_per_gpu=1,
+                 num_replicas=None,
+                 rank=None):
+        _rank, _num_replicas = get_dist_info()
+        if num_replicas is None:
+            num_replicas = _num_replicas
+        if rank is None:
+            rank = _rank
+        self.dataset = dataset
+        self.samples_per_gpu = samples_per_gpu
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.epoch = 0
+
+        assert hasattr(self.dataset, 'flag')
+        self.flag = self.dataset.flag
+        self.group_sizes = np.bincount(self.flag)
+
+        self.num_samples = 0
+        for i, j in enumerate(self.group_sizes):
+            self.num_samples += int(
+                math.ceil(self.group_sizes[i] * 1.0 / self.samples_per_gpu /
+                          self.num_replicas)) * self.samples_per_gpu
+        self.total_size = self.num_samples * self.num_replicas
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+
+        indices = []
+        for i, size in enumerate(self.group_sizes):
+            if size > 0:
+                indice = np.where(self.flag == i)[0]
+                assert len(indice) == size
+                # add .numpy() to avoid bug when selecting indice in parrots.
+                # TODO: check whether torch.randperm() can be replaced by
+                # numpy.random.permutation().
+                indice = indice[list(
+                    torch.randperm(int(size), generator=g).numpy())].tolist()
+                extra = int(
+                    math.ceil(
+                        size * 1.0 / self.samples_per_gpu / self.num_replicas)
+                ) * self.samples_per_gpu * self.num_replicas - len(indice)
+                # pad indice
+                tmp = indice.copy()
+                for _ in range(extra // size):
+                    indice.extend(tmp)
+                indice.extend(tmp[:extra % size])
+                indices.extend(indice)
+
+        assert len(indices) == self.total_size
+
+        indices = [
+            indices[j] for i in list(
+                torch.randperm(
+                    len(indices) // self.samples_per_gpu, generator=g))
+            for j in range(i * self.samples_per_gpu, (i + 1) *
+                           self.samples_per_gpu)
+        ]
+
+        # subsample
+        offset = self.num_samples * self.rank
+        indices = indices[offset:offset + self.num_samples]
+        assert len(indices) == self.num_samples
+
+        return iter(indices)
+
+    def __len__(self):
+        return self.num_samples
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
diff --git a/detection/scrfd/mmdet/datasets/utils.py b/detection/scrfd/mmdet/datasets/utils.py
new file mode 100755
index 000000000..04a4c4c92
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/utils.py
@@ -0,0 +1,62 @@
+import copy
+import warnings
+
+
+def replace_ImageToTensor(pipelines):
+    """Replace the ImageToTensor transform in a data pipeline to
+    DefaultFormatBundle, which is normally useful in batch inference.
+
+    Args:
+        pipelines (list[dict]): Data pipeline configs.
+
+    Returns:
+        list: The new pipeline list with all ImageToTensor replaced by
+            DefaultFormatBundle.
+
+    Examples:
+        >>> pipelines = [
+        ...    dict(type='LoadImageFromFile'),
+        ...    dict(
+        ...        type='MultiScaleFlipAug',
+        ...        img_scale=(1333, 800),
+        ...        flip=False,
+        ...        transforms=[
+        ...            dict(type='Resize', keep_ratio=True),
+        ...            dict(type='RandomFlip'),
+        ...            dict(type='Normalize', mean=[0, 0, 0], std=[1, 1, 1]),
+        ...            dict(type='Pad', size_divisor=32),
+        ...            dict(type='ImageToTensor', keys=['img']),
+        ...            dict(type='Collect', keys=['img']),
+        ...        ])
+        ...    ]
+        >>> expected_pipelines = [
+        ...    dict(type='LoadImageFromFile'),
+        ...    dict(
+        ...        type='MultiScaleFlipAug',
+        ...        img_scale=(1333, 800),
+        ...        flip=False,
+        ...        transforms=[
+        ...            dict(type='Resize', keep_ratio=True),
+        ...            dict(type='RandomFlip'),
+        ...            dict(type='Normalize', mean=[0, 0, 0], std=[1, 1, 1]),
+        ...            dict(type='Pad', size_divisor=32),
+        ...            dict(type='DefaultFormatBundle'),
+        ...            dict(type='Collect', keys=['img']),
+        ...        ])
+        ...    ]
+        >>> assert expected_pipelines == replace_ImageToTensor(pipelines)
+    """
+    pipelines = copy.deepcopy(pipelines)
+    for i, pipeline in enumerate(pipelines):
+        if pipeline['type'] == 'MultiScaleFlipAug':
+            assert 'transforms' in pipeline
+            pipeline['transforms'] = replace_ImageToTensor(
+                pipeline['transforms'])
+        elif pipeline['type'] == 'ImageToTensor':
+            warnings.warn(
+                '"ImageToTensor" pipeline is replaced by '
+                '"DefaultFormatBundle" for batch inference. It is '
+                'recommended to manually replace it in the test '
+                'data pipeline in your config file.', UserWarning)
+            pipelines[i] = {'type': 'DefaultFormatBundle'}
+    return pipelines
diff --git a/detection/scrfd/mmdet/datasets/voc.py b/detection/scrfd/mmdet/datasets/voc.py
new file mode 100755
index 000000000..4d1e861ea
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/voc.py
@@ -0,0 +1,89 @@
+from collections import OrderedDict
+
+from mmdet.core import eval_map, eval_recalls
+from .builder import DATASETS
+from .xml_style import XMLDataset
+
+
+@DATASETS.register_module()
+class VOCDataset(XMLDataset):
+
+    CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',
+               'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
+               'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
+               'tvmonitor')
+
+    def __init__(self, **kwargs):
+        super(VOCDataset, self).__init__(**kwargs)
+        if 'VOC2007' in self.img_prefix:
+            self.year = 2007
+        elif 'VOC2012' in self.img_prefix:
+            self.year = 2012
+        else:
+            raise ValueError('Cannot infer dataset year from img_prefix')
+
+    def evaluate(self,
+                 results,
+                 metric='mAP',
+                 logger=None,
+                 proposal_nums=(100, 300, 1000),
+                 iou_thr=0.5,
+                 scale_ranges=None):
+        """Evaluate in VOC protocol.
+
+        Args:
+            results (list[list | tuple]): Testing results of the dataset.
+            metric (str | list[str]): Metrics to be evaluated. Options are
+                'mAP', 'recall'.
+            logger (logging.Logger | str, optional): Logger used for printing
+                related information during evaluation. Default: None.
+            proposal_nums (Sequence[int]): Proposal number used for evaluating
+                recalls, such as recall@100, recall@1000.
+                Default: (100, 300, 1000).
+            iou_thr (float | list[float]): IoU threshold. It must be a float
+                when evaluating mAP, and can be a list when evaluating recall.
+                Default: 0.5.
+            scale_ranges (list[tuple], optional): Scale ranges for evaluating
+                mAP. If not specified, all bounding boxes would be included in
+                evaluation. Default: None.
+
+        Returns:
+            dict[str, float]: AP/recall metrics.
+        """
+
+        if not isinstance(metric, str):
+            assert len(metric) == 1
+            metric = metric[0]
+        allowed_metrics = ['mAP', 'recall']
+        if metric not in allowed_metrics:
+            raise KeyError(f'metric {metric} is not supported')
+        annotations = [self.get_ann_info(i) for i in range(len(self))]
+        eval_results = OrderedDict()
+        if metric == 'mAP':
+            assert isinstance(iou_thr, float)
+            if self.year == 2007:
+                ds_name = 'voc07'
+            else:
+                ds_name = self.CLASSES
+            mean_ap, _ = eval_map(
+                results,
+                annotations,
+                scale_ranges=None,
+                iou_thr=iou_thr,
+                dataset=ds_name,
+                logger=logger)
+            eval_results['mAP'] = mean_ap
+        elif metric == 'recall':
+            gt_bboxes = [ann['bboxes'] for ann in annotations]
+            if isinstance(iou_thr, float):
+                iou_thr = [iou_thr]
+            recalls = eval_recalls(
+                gt_bboxes, results, proposal_nums, iou_thr, logger=logger)
+            for i, num in enumerate(proposal_nums):
+                for j, iou in enumerate(iou_thr):
+                    eval_results[f'recall@{num}@{iou}'] = recalls[i, j]
+            if recalls.shape[1] > 1:
+                ar = recalls.mean(axis=1)
+                for i, num in enumerate(proposal_nums):
+                    eval_results[f'AR@{num}'] = ar[i]
+        return eval_results
diff --git a/detection/scrfd/mmdet/datasets/wider_face.py b/detection/scrfd/mmdet/datasets/wider_face.py
new file mode 100755
index 000000000..3a13907db
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/wider_face.py
@@ -0,0 +1,51 @@
+import os.path as osp
+import xml.etree.ElementTree as ET
+
+import mmcv
+
+from .builder import DATASETS
+from .xml_style import XMLDataset
+
+
+@DATASETS.register_module()
+class WIDERFaceDataset(XMLDataset):
+    """Reader for the WIDER Face dataset in PASCAL VOC format.
+
+    Conversion scripts can be found in
+    https://github.com/sovrasov/wider-face-pascal-voc-annotations
+    """
+    CLASSES = ('face', )
+
+    def __init__(self, **kwargs):
+        super(WIDERFaceDataset, self).__init__(**kwargs)
+
+    def load_annotations(self, ann_file):
+        """Load annotation from WIDERFace XML style annotation file.
+
+        Args:
+            ann_file (str): Path of XML file.
+
+        Returns:
+            list[dict]: Annotation info from XML file.
+        """
+
+        data_infos = []
+        img_ids = mmcv.list_from_file(ann_file)
+        for img_id in img_ids:
+            filename = f'{img_id}.jpg'
+            xml_path = osp.join(self.img_prefix, 'Annotations',
+                                f'{img_id}.xml')
+            tree = ET.parse(xml_path)
+            root = tree.getroot()
+            size = root.find('size')
+            width = int(size.find('width').text)
+            height = int(size.find('height').text)
+            folder = root.find('folder').text
+            data_infos.append(
+                dict(
+                    id=img_id,
+                    filename=osp.join(folder, filename),
+                    width=width,
+                    height=height))
+
+        return data_infos
diff --git a/detection/scrfd/mmdet/datasets/xml_style.py b/detection/scrfd/mmdet/datasets/xml_style.py
new file mode 100755
index 000000000..b912de38d
--- /dev/null
+++ b/detection/scrfd/mmdet/datasets/xml_style.py
@@ -0,0 +1,169 @@
+import os.path as osp
+import xml.etree.ElementTree as ET
+
+import mmcv
+import numpy as np
+from PIL import Image
+
+from .builder import DATASETS
+from .custom import CustomDataset
+
+
+@DATASETS.register_module()
+class XMLDataset(CustomDataset):
+    """XML dataset for detection.
+
+    Args:
+        min_size (int | float, optional): The minimum size of bounding
+            boxes in the images. If the size of a bounding box is less than
+            ``min_size``, it would be add to ignored field.
+    """
+
+    def __init__(self, min_size=None, **kwargs):
+        super(XMLDataset, self).__init__(**kwargs)
+        self.cat2label = {cat: i for i, cat in enumerate(self.CLASSES)}
+        self.min_size = min_size
+
+    def load_annotations(self, ann_file):
+        """Load annotation from XML style ann_file.
+
+        Args:
+            ann_file (str): Path of XML file.
+
+        Returns:
+            list[dict]: Annotation info from XML file.
+        """
+
+        data_infos = []
+        img_ids = mmcv.list_from_file(ann_file)
+        for img_id in img_ids:
+            filename = f'JPEGImages/{img_id}.jpg'
+            xml_path = osp.join(self.img_prefix, 'Annotations',
+                                f'{img_id}.xml')
+            tree = ET.parse(xml_path)
+            root = tree.getroot()
+            size = root.find('size')
+            width = 0
+            height = 0
+            if size is not None:
+                width = int(size.find('width').text)
+                height = int(size.find('height').text)
+            else:
+                img_path = osp.join(self.img_prefix, 'JPEGImages',
+                                    '{}.jpg'.format(img_id))
+                img = Image.open(img_path)
+                width, height = img.size
+            data_infos.append(
+                dict(id=img_id, filename=filename, width=width, height=height))
+
+        return data_infos
+
+    def _filter_imgs(self, min_size=32):
+        """Filter images too small or without annotation."""
+        valid_inds = []
+        for i, img_info in enumerate(self.data_infos):
+            if min(img_info['width'], img_info['height']) < min_size:
+                continue
+            if self.filter_empty_gt:
+                img_id = img_info['id']
+                xml_path = osp.join(self.img_prefix, 'Annotations',
+                                    f'{img_id}.xml')
+                tree = ET.parse(xml_path)
+                root = tree.getroot()
+                for obj in root.findall('object'):
+                    name = obj.find('name').text
+                    if name in self.CLASSES:
+                        valid_inds.append(i)
+                        break
+            else:
+                valid_inds.append(i)
+        return valid_inds
+
+    def get_ann_info(self, idx):
+        """Get annotation from XML file by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            dict: Annotation info of specified index.
+        """
+
+        img_id = self.data_infos[idx]['id']
+        xml_path = osp.join(self.img_prefix, 'Annotations', f'{img_id}.xml')
+        tree = ET.parse(xml_path)
+        root = tree.getroot()
+        bboxes = []
+        labels = []
+        bboxes_ignore = []
+        labels_ignore = []
+        for obj in root.findall('object'):
+            name = obj.find('name').text
+            if name not in self.CLASSES:
+                continue
+            label = self.cat2label[name]
+            difficult = int(obj.find('difficult').text)
+            bnd_box = obj.find('bndbox')
+            # TODO: check whether it is necessary to use int
+            # Coordinates may be float type
+            bbox = [
+                int(float(bnd_box.find('xmin').text)),
+                int(float(bnd_box.find('ymin').text)),
+                int(float(bnd_box.find('xmax').text)),
+                int(float(bnd_box.find('ymax').text))
+            ]
+            ignore = False
+            if self.min_size:
+                assert not self.test_mode
+                w = bbox[2] - bbox[0]
+                h = bbox[3] - bbox[1]
+                if w < self.min_size or h < self.min_size:
+                    ignore = True
+            if difficult or ignore:
+                bboxes_ignore.append(bbox)
+                labels_ignore.append(label)
+            else:
+                bboxes.append(bbox)
+                labels.append(label)
+        if not bboxes:
+            bboxes = np.zeros((0, 4))
+            labels = np.zeros((0, ))
+        else:
+            bboxes = np.array(bboxes, ndmin=2) - 1
+            labels = np.array(labels)
+        if not bboxes_ignore:
+            bboxes_ignore = np.zeros((0, 4))
+            labels_ignore = np.zeros((0, ))
+        else:
+            bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1
+            labels_ignore = np.array(labels_ignore)
+        ann = dict(
+            bboxes=bboxes.astype(np.float32),
+            labels=labels.astype(np.int64),
+            bboxes_ignore=bboxes_ignore.astype(np.float32),
+            labels_ignore=labels_ignore.astype(np.int64))
+        return ann
+
+    def get_cat_ids(self, idx):
+        """Get category ids in XML file by index.
+
+        Args:
+            idx (int): Index of data.
+
+        Returns:
+            list[int]: All categories in the image of specified index.
+        """
+
+        cat_ids = []
+        img_id = self.data_infos[idx]['id']
+        xml_path = osp.join(self.img_prefix, 'Annotations', f'{img_id}.xml')
+        tree = ET.parse(xml_path)
+        root = tree.getroot()
+        for obj in root.findall('object'):
+            name = obj.find('name').text
+            if name not in self.CLASSES:
+                continue
+            label = self.cat2label[name]
+            cat_ids.append(label)
+
+        return cat_ids
diff --git a/detection/scrfd/mmdet/models/__init__.py b/detection/scrfd/mmdet/models/__init__.py
new file mode 100755
index 000000000..44ac99855
--- /dev/null
+++ b/detection/scrfd/mmdet/models/__init__.py
@@ -0,0 +1,16 @@
+from .backbones import *  # noqa: F401,F403
+from .builder import (BACKBONES, DETECTORS, HEADS, LOSSES, NECKS,
+                      ROI_EXTRACTORS, SHARED_HEADS, build_backbone,
+                      build_detector, build_head, build_loss, build_neck,
+                      build_roi_extractor, build_shared_head)
+from .dense_heads import *  # noqa: F401,F403
+from .detectors import *  # noqa: F401,F403
+from .losses import *  # noqa: F401,F403
+from .necks import *  # noqa: F401,F403
+from .roi_heads import *  # noqa: F401,F403
+
+__all__ = [
+    'BACKBONES', 'NECKS', 'ROI_EXTRACTORS', 'SHARED_HEADS', 'HEADS', 'LOSSES',
+    'DETECTORS', 'build_backbone', 'build_neck', 'build_roi_extractor',
+    'build_shared_head', 'build_head', 'build_loss', 'build_detector'
+]
diff --git a/detection/scrfd/mmdet/models/backbones/__init__.py b/detection/scrfd/mmdet/models/backbones/__init__.py
new file mode 100755
index 000000000..2dc4665de
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/__init__.py
@@ -0,0 +1,20 @@
+from .darknet import Darknet
+from .detectors_resnet import DetectoRS_ResNet
+from .detectors_resnext import DetectoRS_ResNeXt
+from .hourglass import HourglassNet
+from .hrnet import HRNet
+from .regnet import RegNet
+from .res2net import Res2Net
+from .resnest import ResNeSt
+from .resnet import ResNet, ResNetV1d, ResNetV1e
+from .resnext import ResNeXt
+from .ssd_vgg import SSDVGG
+from .trident_resnet import TridentResNet
+
+from .mobilenet import MobileNetV1
+
+__all__ = [
+    'RegNet', 'ResNet', 'ResNetV1d', 'ResNetV1e', 'ResNeXt', 'SSDVGG', 'HRNet', 'Res2Net',
+    'HourglassNet', 'DetectoRS_ResNet', 'DetectoRS_ResNeXt', 'Darknet',
+    'ResNeSt', 'TridentResNet', 'MobileNetV1'
+]
diff --git a/detection/scrfd/mmdet/models/backbones/darknet.py b/detection/scrfd/mmdet/models/backbones/darknet.py
new file mode 100755
index 000000000..517fe2625
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/darknet.py
@@ -0,0 +1,199 @@
+# Copyright (c) 2019 Western Digital Corporation or its affiliates.
+
+import logging
+
+import torch.nn as nn
+from mmcv.cnn import ConvModule, constant_init, kaiming_init
+from mmcv.runner import load_checkpoint
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from ..builder import BACKBONES
+
+
+class ResBlock(nn.Module):
+    """The basic residual block used in Darknet. Each ResBlock consists of two
+    ConvModules and the input is added to the final output. Each ConvModule is
+    composed of Conv, BN, and LeakyReLU. In YoloV3 paper, the first convLayer
+    has half of the number of the filters as much as the second convLayer. The
+    first convLayer has filter size of 1x1 and the second one has the filter
+    size of 3x3.
+
+    Args:
+        in_channels (int): The input channels. Must be even.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+            Default: dict(type='BN', requires_grad=True)
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='LeakyReLU', negative_slope=0.1).
+    """
+
+    def __init__(self,
+                 in_channels,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 act_cfg=dict(type='LeakyReLU', negative_slope=0.1)):
+        super(ResBlock, self).__init__()
+        assert in_channels % 2 == 0  # ensure the in_channels is even
+        half_in_channels = in_channels // 2
+
+        # shortcut
+        cfg = dict(conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg)
+
+        self.conv1 = ConvModule(in_channels, half_in_channels, 1, **cfg)
+        self.conv2 = ConvModule(
+            half_in_channels, in_channels, 3, padding=1, **cfg)
+
+    def forward(self, x):
+        residual = x
+        out = self.conv1(x)
+        out = self.conv2(out)
+        out = out + residual
+
+        return out
+
+
+@BACKBONES.register_module()
+class Darknet(nn.Module):
+    """Darknet backbone.
+
+    Args:
+        depth (int): Depth of Darknet. Currently only support 53.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters. Default: -1.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+            Default: dict(type='BN', requires_grad=True)
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='LeakyReLU', negative_slope=0.1).
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+
+    Example:
+        >>> from mmdet.models import Darknet
+        >>> import torch
+        >>> self = Darknet(depth=53)
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 416, 416)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        ...
+        (1, 256, 52, 52)
+        (1, 512, 26, 26)
+        (1, 1024, 13, 13)
+    """
+
+    # Dict(depth: (layers, channels))
+    arch_settings = {
+        53: ((1, 2, 8, 8, 4), ((32, 64), (64, 128), (128, 256), (256, 512),
+                               (512, 1024)))
+    }
+
+    def __init__(self,
+                 depth=53,
+                 out_indices=(3, 4, 5),
+                 frozen_stages=-1,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 act_cfg=dict(type='LeakyReLU', negative_slope=0.1),
+                 norm_eval=True):
+        super(Darknet, self).__init__()
+        if depth not in self.arch_settings:
+            raise KeyError(f'invalid depth {depth} for darknet')
+        self.depth = depth
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+        self.layers, self.channels = self.arch_settings[depth]
+
+        cfg = dict(conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg)
+
+        self.conv1 = ConvModule(3, 32, 3, padding=1, **cfg)
+
+        self.cr_blocks = ['conv1']
+        for i, n_layers in enumerate(self.layers):
+            layer_name = f'conv_res_block{i + 1}'
+            in_c, out_c = self.channels[i]
+            self.add_module(
+                layer_name,
+                self.make_conv_res_block(in_c, out_c, n_layers, **cfg))
+            self.cr_blocks.append(layer_name)
+
+        self.norm_eval = norm_eval
+
+    def forward(self, x):
+        outs = []
+        for i, layer_name in enumerate(self.cr_blocks):
+            cr_block = getattr(self, layer_name)
+            x = cr_block(x)
+            if i in self.out_indices:
+                outs.append(x)
+
+        return tuple(outs)
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = logging.getLogger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
+                    constant_init(m, 1)
+
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            for i in range(self.frozen_stages):
+                m = getattr(self, self.cr_blocks[i])
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+    def train(self, mode=True):
+        super(Darknet, self).train(mode)
+        self._freeze_stages()
+        if mode and self.norm_eval:
+            for m in self.modules():
+                if isinstance(m, _BatchNorm):
+                    m.eval()
+
+    @staticmethod
+    def make_conv_res_block(in_channels,
+                            out_channels,
+                            res_repeat,
+                            conv_cfg=None,
+                            norm_cfg=dict(type='BN', requires_grad=True),
+                            act_cfg=dict(type='LeakyReLU',
+                                         negative_slope=0.1)):
+        """In Darknet backbone, ConvLayer is usually followed by ResBlock. This
+        function will make that. The Conv layers always have 3x3 filters with
+        stride=2. The number of the filters in Conv layer is the same as the
+        out channels of the ResBlock.
+
+        Args:
+            in_channels (int): The number of input channels.
+            out_channels (int): The number of output channels.
+            res_repeat (int): The number of ResBlocks.
+            conv_cfg (dict): Config dict for convolution layer. Default: None.
+            norm_cfg (dict): Dictionary to construct and config norm layer.
+                Default: dict(type='BN', requires_grad=True)
+            act_cfg (dict): Config dict for activation layer.
+                Default: dict(type='LeakyReLU', negative_slope=0.1).
+        """
+
+        cfg = dict(conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg)
+
+        model = nn.Sequential()
+        model.add_module(
+            'conv',
+            ConvModule(
+                in_channels, out_channels, 3, stride=2, padding=1, **cfg))
+        for idx in range(res_repeat):
+            model.add_module('res{}'.format(idx),
+                             ResBlock(out_channels, **cfg))
+        return model
diff --git a/detection/scrfd/mmdet/models/backbones/detectors_resnet.py b/detection/scrfd/mmdet/models/backbones/detectors_resnet.py
new file mode 100755
index 000000000..324e737de
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/detectors_resnet.py
@@ -0,0 +1,305 @@
+import torch.nn as nn
+import torch.utils.checkpoint as cp
+from mmcv.cnn import build_conv_layer, build_norm_layer, constant_init
+
+from ..builder import BACKBONES
+from .resnet import Bottleneck as _Bottleneck
+from .resnet import ResNet
+
+
+class Bottleneck(_Bottleneck):
+    r"""Bottleneck for the ResNet backbone in `DetectoRS
+    <https://arxiv.org/pdf/2006.02334.pdf>`_.
+
+    This bottleneck allows the users to specify whether to use
+    SAC (Switchable Atrous Convolution) and RFP (Recursive Feature Pyramid).
+
+    Args:
+         inplanes (int): The number of input channels.
+         planes (int): The number of output channels before expansion.
+         rfp_inplanes (int, optional): The number of channels from RFP.
+             Default: None. If specified, an additional conv layer will be
+             added for ``rfp_feat``. Otherwise, the structure is the same as
+             base class.
+         sac (dict, optional): Dictionary to construct SAC. Default: None.
+    """
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 rfp_inplanes=None,
+                 sac=None,
+                 **kwargs):
+        super(Bottleneck, self).__init__(inplanes, planes, **kwargs)
+
+        assert sac is None or isinstance(sac, dict)
+        self.sac = sac
+        self.with_sac = sac is not None
+        if self.with_sac:
+            self.conv2 = build_conv_layer(
+                self.sac,
+                planes,
+                planes,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                bias=False)
+
+        self.rfp_inplanes = rfp_inplanes
+        if self.rfp_inplanes:
+            self.rfp_conv = build_conv_layer(
+                None,
+                self.rfp_inplanes,
+                planes * self.expansion,
+                1,
+                stride=1,
+                bias=True)
+        self.init_weights()
+
+    def init_weights(self):
+        """Initialize the weights."""
+        if self.rfp_inplanes:
+            constant_init(self.rfp_conv, 0)
+
+    def rfp_forward(self, x, rfp_feat):
+        """The forward function that also takes the RFP features as input."""
+
+        def _inner_forward(x):
+            identity = x
+
+            out = self.conv1(x)
+            out = self.norm1(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv1_plugin_names)
+
+            out = self.conv2(out)
+            out = self.norm2(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv2_plugin_names)
+
+            out = self.conv3(out)
+            out = self.norm3(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv3_plugin_names)
+
+            if self.downsample is not None:
+                identity = self.downsample(x)
+
+            out += identity
+
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        if self.rfp_inplanes:
+            rfp_feat = self.rfp_conv(rfp_feat)
+            out = out + rfp_feat
+
+        out = self.relu(out)
+
+        return out
+
+
+class ResLayer(nn.Sequential):
+    """ResLayer to build ResNet style backbone for RPF in detectoRS.
+
+    The difference between this module and base class is that we pass
+    ``rfp_inplanes`` to the first block.
+
+    Args:
+        block (nn.Module): block used to build ResLayer.
+        inplanes (int): inplanes of block.
+        planes (int): planes of block.
+        num_blocks (int): number of blocks.
+        stride (int): stride of the first block. Default: 1
+        avg_down (bool): Use AvgPool instead of stride conv when
+            downsampling in the bottleneck. Default: False
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: dict(type='BN')
+        downsample_first (bool): Downsample at the first block or last block.
+            False for Hourglass, True for ResNet. Default: True
+        rfp_inplanes (int, optional): The number of channels from RFP.
+            Default: None. If specified, an additional conv layer will be
+            added for ``rfp_feat``. Otherwise, the structure is the same as
+            base class.
+    """
+
+    def __init__(self,
+                 block,
+                 inplanes,
+                 planes,
+                 num_blocks,
+                 stride=1,
+                 avg_down=False,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 downsample_first=True,
+                 rfp_inplanes=None,
+                 **kwargs):
+        self.block = block
+        assert downsample_first, f'downsampel_first={downsample_first} is ' \
+                                 'not supported in DetectoRS'
+
+        downsample = None
+        if stride != 1 or inplanes != planes * block.expansion:
+            downsample = []
+            conv_stride = stride
+            if avg_down and stride != 1:
+                conv_stride = 1
+                downsample.append(
+                    nn.AvgPool2d(
+                        kernel_size=stride,
+                        stride=stride,
+                        ceil_mode=True,
+                        count_include_pad=False))
+            downsample.extend([
+                build_conv_layer(
+                    conv_cfg,
+                    inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=conv_stride,
+                    bias=False),
+                build_norm_layer(norm_cfg, planes * block.expansion)[1]
+            ])
+            downsample = nn.Sequential(*downsample)
+
+        layers = []
+        layers.append(
+            block(
+                inplanes=inplanes,
+                planes=planes,
+                stride=stride,
+                downsample=downsample,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                rfp_inplanes=rfp_inplanes,
+                **kwargs))
+        inplanes = planes * block.expansion
+        for _ in range(1, num_blocks):
+            layers.append(
+                block(
+                    inplanes=inplanes,
+                    planes=planes,
+                    stride=1,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    **kwargs))
+
+        super(ResLayer, self).__init__(*layers)
+
+
+@BACKBONES.register_module()
+class DetectoRS_ResNet(ResNet):
+    """ResNet backbone for DetectoRS.
+
+    Args:
+        sac (dict, optional): Dictionary to construct SAC (Switchable Atrous
+            Convolution). Default: None.
+        stage_with_sac (list): Which stage to use sac. Default: (False, False,
+            False, False).
+        rfp_inplanes (int, optional): The number of channels from RFP.
+            Default: None. If specified, an additional conv layer will be
+            added for ``rfp_feat``. Otherwise, the structure is the same as
+            base class.
+        output_img (bool): If ``True``, the input image will be inserted into
+            the starting position of output. Default: False.
+        pretrained (str, optional): The pretrained model to load.
+    """
+
+    arch_settings = {
+        50: (Bottleneck, (3, 4, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3))
+    }
+
+    def __init__(self,
+                 sac=None,
+                 stage_with_sac=(False, False, False, False),
+                 rfp_inplanes=None,
+                 output_img=False,
+                 pretrained=None,
+                 **kwargs):
+        self.sac = sac
+        self.stage_with_sac = stage_with_sac
+        self.rfp_inplanes = rfp_inplanes
+        self.output_img = output_img
+        self.pretrained = pretrained
+        super(DetectoRS_ResNet, self).__init__(**kwargs)
+
+        self.inplanes = self.stem_channels
+        self.res_layers = []
+        for i, num_blocks in enumerate(self.stage_blocks):
+            stride = self.strides[i]
+            dilation = self.dilations[i]
+            dcn = self.dcn if self.stage_with_dcn[i] else None
+            sac = self.sac if self.stage_with_sac[i] else None
+            if self.plugins is not None:
+                stage_plugins = self.make_stage_plugins(self.plugins, i)
+            else:
+                stage_plugins = None
+            planes = self.base_channels * 2**i
+            res_layer = self.make_res_layer(
+                block=self.block,
+                inplanes=self.inplanes,
+                planes=planes,
+                num_blocks=num_blocks,
+                stride=stride,
+                dilation=dilation,
+                style=self.style,
+                avg_down=self.avg_down,
+                with_cp=self.with_cp,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg,
+                dcn=dcn,
+                sac=sac,
+                rfp_inplanes=rfp_inplanes if i > 0 else None,
+                plugins=stage_plugins)
+            self.inplanes = planes * self.block.expansion
+            layer_name = f'layer{i + 1}'
+            self.add_module(layer_name, res_layer)
+            self.res_layers.append(layer_name)
+
+        self._freeze_stages()
+
+    def make_res_layer(self, **kwargs):
+        """Pack all blocks in a stage into a ``ResLayer`` for DetectoRS."""
+        return ResLayer(**kwargs)
+
+    def forward(self, x):
+        """Forward function."""
+        outs = list(super(DetectoRS_ResNet, self).forward(x))
+        if self.output_img:
+            outs.insert(0, x)
+        return tuple(outs)
+
+    def rfp_forward(self, x, rfp_feats):
+        """Forward function for RFP."""
+        if self.deep_stem:
+            x = self.stem(x)
+        else:
+            x = self.conv1(x)
+            x = self.norm1(x)
+            x = self.relu(x)
+        x = self.maxpool(x)
+        outs = []
+        for i, layer_name in enumerate(self.res_layers):
+            res_layer = getattr(self, layer_name)
+            rfp_feat = rfp_feats[i] if i > 0 else None
+            for layer in res_layer:
+                x = layer.rfp_forward(x, rfp_feat)
+            if i in self.out_indices:
+                outs.append(x)
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/backbones/detectors_resnext.py b/detection/scrfd/mmdet/models/backbones/detectors_resnext.py
new file mode 100755
index 000000000..57d032fe3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/detectors_resnext.py
@@ -0,0 +1,122 @@
+import math
+
+from mmcv.cnn import build_conv_layer, build_norm_layer
+
+from ..builder import BACKBONES
+from .detectors_resnet import Bottleneck as _Bottleneck
+from .detectors_resnet import DetectoRS_ResNet
+
+
+class Bottleneck(_Bottleneck):
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 groups=1,
+                 base_width=4,
+                 base_channels=64,
+                 **kwargs):
+        """Bottleneck block for ResNeXt.
+
+        If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
+        it is "caffe", the stride-two layer is the first 1x1 conv layer.
+        """
+        super(Bottleneck, self).__init__(inplanes, planes, **kwargs)
+
+        if groups == 1:
+            width = self.planes
+        else:
+            width = math.floor(self.planes *
+                               (base_width / base_channels)) * groups
+
+        self.norm1_name, norm1 = build_norm_layer(
+            self.norm_cfg, width, postfix=1)
+        self.norm2_name, norm2 = build_norm_layer(
+            self.norm_cfg, width, postfix=2)
+        self.norm3_name, norm3 = build_norm_layer(
+            self.norm_cfg, self.planes * self.expansion, postfix=3)
+
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            self.inplanes,
+            width,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+        fallback_on_stride = False
+        self.with_modulated_dcn = False
+        if self.with_dcn:
+            fallback_on_stride = self.dcn.pop('fallback_on_stride', False)
+        if self.with_sac:
+            self.conv2 = build_conv_layer(
+                self.sac,
+                width,
+                width,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                groups=groups,
+                bias=False)
+        elif not self.with_dcn or fallback_on_stride:
+            self.conv2 = build_conv_layer(
+                self.conv_cfg,
+                width,
+                width,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                groups=groups,
+                bias=False)
+        else:
+            assert self.conv_cfg is None, 'conv_cfg must be None for DCN'
+            self.conv2 = build_conv_layer(
+                self.dcn,
+                width,
+                width,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                groups=groups,
+                bias=False)
+
+        self.add_module(self.norm2_name, norm2)
+        self.conv3 = build_conv_layer(
+            self.conv_cfg,
+            width,
+            self.planes * self.expansion,
+            kernel_size=1,
+            bias=False)
+        self.add_module(self.norm3_name, norm3)
+
+
+@BACKBONES.register_module()
+class DetectoRS_ResNeXt(DetectoRS_ResNet):
+    """ResNeXt backbone for DetectoRS.
+
+    Args:
+        groups (int): The number of groups in ResNeXt.
+        base_width (int): The base width of ResNeXt.
+    """
+
+    arch_settings = {
+        50: (Bottleneck, (3, 4, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3))
+    }
+
+    def __init__(self, groups=1, base_width=4, **kwargs):
+        self.groups = groups
+        self.base_width = base_width
+        super(DetectoRS_ResNeXt, self).__init__(**kwargs)
+
+    def make_res_layer(self, **kwargs):
+        return super().make_res_layer(
+            groups=self.groups,
+            base_width=self.base_width,
+            base_channels=self.base_channels,
+            **kwargs)
diff --git a/detection/scrfd/mmdet/models/backbones/hourglass.py b/detection/scrfd/mmdet/models/backbones/hourglass.py
new file mode 100755
index 000000000..3422acee3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/hourglass.py
@@ -0,0 +1,198 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule
+
+from ..builder import BACKBONES
+from ..utils import ResLayer
+from .resnet import BasicBlock
+
+
+class HourglassModule(nn.Module):
+    """Hourglass Module for HourglassNet backbone.
+
+    Generate module recursively and use BasicBlock as the base unit.
+
+    Args:
+        depth (int): Depth of current HourglassModule.
+        stage_channels (list[int]): Feature channels of sub-modules in current
+            and follow-up HourglassModule.
+        stage_blocks (list[int]): Number of sub-modules stacked in current and
+            follow-up HourglassModule.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+    """
+
+    def __init__(self,
+                 depth,
+                 stage_channels,
+                 stage_blocks,
+                 norm_cfg=dict(type='BN', requires_grad=True)):
+        super(HourglassModule, self).__init__()
+
+        self.depth = depth
+
+        cur_block = stage_blocks[0]
+        next_block = stage_blocks[1]
+
+        cur_channel = stage_channels[0]
+        next_channel = stage_channels[1]
+
+        self.up1 = ResLayer(
+            BasicBlock, cur_channel, cur_channel, cur_block, norm_cfg=norm_cfg)
+
+        self.low1 = ResLayer(
+            BasicBlock,
+            cur_channel,
+            next_channel,
+            cur_block,
+            stride=2,
+            norm_cfg=norm_cfg)
+
+        if self.depth > 1:
+            self.low2 = HourglassModule(depth - 1, stage_channels[1:],
+                                        stage_blocks[1:])
+        else:
+            self.low2 = ResLayer(
+                BasicBlock,
+                next_channel,
+                next_channel,
+                next_block,
+                norm_cfg=norm_cfg)
+
+        self.low3 = ResLayer(
+            BasicBlock,
+            next_channel,
+            cur_channel,
+            cur_block,
+            norm_cfg=norm_cfg,
+            downsample_first=False)
+
+        self.up2 = nn.Upsample(scale_factor=2)
+
+    def forward(self, x):
+        """Forward function."""
+        up1 = self.up1(x)
+        low1 = self.low1(x)
+        low2 = self.low2(low1)
+        low3 = self.low3(low2)
+        up2 = self.up2(low3)
+        return up1 + up2
+
+
+@BACKBONES.register_module()
+class HourglassNet(nn.Module):
+    """HourglassNet backbone.
+
+    Stacked Hourglass Networks for Human Pose Estimation.
+    More details can be found in the `paper
+    <https://arxiv.org/abs/1603.06937>`_ .
+
+    Args:
+        downsample_times (int): Downsample times in a HourglassModule.
+        num_stacks (int): Number of HourglassModule modules stacked,
+            1 for Hourglass-52, 2 for Hourglass-104.
+        stage_channels (list[int]): Feature channel of each sub-module in a
+            HourglassModule.
+        stage_blocks (list[int]): Number of sub-modules stacked in a
+            HourglassModule.
+        feat_channel (int): Feature channel of conv after a HourglassModule.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+
+    Example:
+        >>> from mmdet.models import HourglassNet
+        >>> import torch
+        >>> self = HourglassNet()
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 511, 511)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_output in level_outputs:
+        ...     print(tuple(level_output.shape))
+        (1, 256, 128, 128)
+        (1, 256, 128, 128)
+    """
+
+    def __init__(self,
+                 downsample_times=5,
+                 num_stacks=2,
+                 stage_channels=(256, 256, 384, 384, 384, 512),
+                 stage_blocks=(2, 2, 2, 2, 2, 4),
+                 feat_channel=256,
+                 norm_cfg=dict(type='BN', requires_grad=True)):
+        super(HourglassNet, self).__init__()
+
+        self.num_stacks = num_stacks
+        assert self.num_stacks >= 1
+        assert len(stage_channels) == len(stage_blocks)
+        assert len(stage_channels) > downsample_times
+
+        cur_channel = stage_channels[0]
+
+        self.stem = nn.Sequential(
+            ConvModule(3, 128, 7, padding=3, stride=2, norm_cfg=norm_cfg),
+            ResLayer(BasicBlock, 128, 256, 1, stride=2, norm_cfg=norm_cfg))
+
+        self.hourglass_modules = nn.ModuleList([
+            HourglassModule(downsample_times, stage_channels, stage_blocks)
+            for _ in range(num_stacks)
+        ])
+
+        self.inters = ResLayer(
+            BasicBlock,
+            cur_channel,
+            cur_channel,
+            num_stacks - 1,
+            norm_cfg=norm_cfg)
+
+        self.conv1x1s = nn.ModuleList([
+            ConvModule(
+                cur_channel, cur_channel, 1, norm_cfg=norm_cfg, act_cfg=None)
+            for _ in range(num_stacks - 1)
+        ])
+
+        self.out_convs = nn.ModuleList([
+            ConvModule(
+                cur_channel, feat_channel, 3, padding=1, norm_cfg=norm_cfg)
+            for _ in range(num_stacks)
+        ])
+
+        self.remap_convs = nn.ModuleList([
+            ConvModule(
+                feat_channel, cur_channel, 1, norm_cfg=norm_cfg, act_cfg=None)
+            for _ in range(num_stacks - 1)
+        ])
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def init_weights(self, pretrained=None):
+        """Init module weights.
+
+        We do nothing in this function because all modules we used
+        (ConvModule, BasicBlock and etc.) have default initialization, and
+        currently we don't provide pretrained model of HourglassNet.
+
+        Detector's __init__() will call backbone's init_weights() with
+        pretrained as input, so we keep this function.
+        """
+        # Training Centripetal Model needs to reset parameters for Conv2d
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                m.reset_parameters()
+
+    def forward(self, x):
+        """Forward function."""
+        inter_feat = self.stem(x)
+        out_feats = []
+
+        for ind in range(self.num_stacks):
+            single_hourglass = self.hourglass_modules[ind]
+            out_conv = self.out_convs[ind]
+
+            hourglass_feat = single_hourglass(inter_feat)
+            out_feat = out_conv(hourglass_feat)
+            out_feats.append(out_feat)
+
+            if ind < self.num_stacks - 1:
+                inter_feat = self.conv1x1s[ind](
+                    inter_feat) + self.remap_convs[ind](
+                        out_feat)
+                inter_feat = self.inters[ind](self.relu(inter_feat))
+
+        return out_feats
diff --git a/detection/scrfd/mmdet/models/backbones/hrnet.py b/detection/scrfd/mmdet/models/backbones/hrnet.py
new file mode 100755
index 000000000..1ecc79f12
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/hrnet.py
@@ -0,0 +1,537 @@
+import torch.nn as nn
+from mmcv.cnn import (build_conv_layer, build_norm_layer, constant_init,
+                      kaiming_init)
+from mmcv.runner import load_checkpoint
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from mmdet.utils import get_root_logger
+from ..builder import BACKBONES
+from .resnet import BasicBlock, Bottleneck
+
+
+class HRModule(nn.Module):
+    """High-Resolution Module for HRNet.
+
+    In this module, every branch has 4 BasicBlocks/Bottlenecks. Fusion/Exchange
+    is in this module.
+    """
+
+    def __init__(self,
+                 num_branches,
+                 blocks,
+                 num_blocks,
+                 in_channels,
+                 num_channels,
+                 multiscale_output=True,
+                 with_cp=False,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN')):
+        super(HRModule, self).__init__()
+        self._check_branches(num_branches, num_blocks, in_channels,
+                             num_channels)
+
+        self.in_channels = in_channels
+        self.num_branches = num_branches
+
+        self.multiscale_output = multiscale_output
+        self.norm_cfg = norm_cfg
+        self.conv_cfg = conv_cfg
+        self.with_cp = with_cp
+        self.branches = self._make_branches(num_branches, blocks, num_blocks,
+                                            num_channels)
+        self.fuse_layers = self._make_fuse_layers()
+        self.relu = nn.ReLU(inplace=False)
+
+    def _check_branches(self, num_branches, num_blocks, in_channels,
+                        num_channels):
+        if num_branches != len(num_blocks):
+            error_msg = f'NUM_BRANCHES({num_branches}) ' \
+                f'!= NUM_BLOCKS({len(num_blocks)})'
+            raise ValueError(error_msg)
+
+        if num_branches != len(num_channels):
+            error_msg = f'NUM_BRANCHES({num_branches}) ' \
+                f'!= NUM_CHANNELS({len(num_channels)})'
+            raise ValueError(error_msg)
+
+        if num_branches != len(in_channels):
+            error_msg = f'NUM_BRANCHES({num_branches}) ' \
+                f'!= NUM_INCHANNELS({len(in_channels)})'
+            raise ValueError(error_msg)
+
+    def _make_one_branch(self,
+                         branch_index,
+                         block,
+                         num_blocks,
+                         num_channels,
+                         stride=1):
+        downsample = None
+        if stride != 1 or \
+                self.in_channels[branch_index] != \
+                num_channels[branch_index] * block.expansion:
+            downsample = nn.Sequential(
+                build_conv_layer(
+                    self.conv_cfg,
+                    self.in_channels[branch_index],
+                    num_channels[branch_index] * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    bias=False),
+                build_norm_layer(self.norm_cfg, num_channels[branch_index] *
+                                 block.expansion)[1])
+
+        layers = []
+        layers.append(
+            block(
+                self.in_channels[branch_index],
+                num_channels[branch_index],
+                stride,
+                downsample=downsample,
+                with_cp=self.with_cp,
+                norm_cfg=self.norm_cfg,
+                conv_cfg=self.conv_cfg))
+        self.in_channels[branch_index] = \
+            num_channels[branch_index] * block.expansion
+        for i in range(1, num_blocks[branch_index]):
+            layers.append(
+                block(
+                    self.in_channels[branch_index],
+                    num_channels[branch_index],
+                    with_cp=self.with_cp,
+                    norm_cfg=self.norm_cfg,
+                    conv_cfg=self.conv_cfg))
+
+        return nn.Sequential(*layers)
+
+    def _make_branches(self, num_branches, block, num_blocks, num_channels):
+        branches = []
+
+        for i in range(num_branches):
+            branches.append(
+                self._make_one_branch(i, block, num_blocks, num_channels))
+
+        return nn.ModuleList(branches)
+
+    def _make_fuse_layers(self):
+        if self.num_branches == 1:
+            return None
+
+        num_branches = self.num_branches
+        in_channels = self.in_channels
+        fuse_layers = []
+        num_out_branches = num_branches if self.multiscale_output else 1
+        for i in range(num_out_branches):
+            fuse_layer = []
+            for j in range(num_branches):
+                if j > i:
+                    fuse_layer.append(
+                        nn.Sequential(
+                            build_conv_layer(
+                                self.conv_cfg,
+                                in_channels[j],
+                                in_channels[i],
+                                kernel_size=1,
+                                stride=1,
+                                padding=0,
+                                bias=False),
+                            build_norm_layer(self.norm_cfg, in_channels[i])[1],
+                            nn.Upsample(
+                                scale_factor=2**(j - i), mode='nearest')))
+                elif j == i:
+                    fuse_layer.append(None)
+                else:
+                    conv_downsamples = []
+                    for k in range(i - j):
+                        if k == i - j - 1:
+                            conv_downsamples.append(
+                                nn.Sequential(
+                                    build_conv_layer(
+                                        self.conv_cfg,
+                                        in_channels[j],
+                                        in_channels[i],
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=1,
+                                        bias=False),
+                                    build_norm_layer(self.norm_cfg,
+                                                     in_channels[i])[1]))
+                        else:
+                            conv_downsamples.append(
+                                nn.Sequential(
+                                    build_conv_layer(
+                                        self.conv_cfg,
+                                        in_channels[j],
+                                        in_channels[j],
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=1,
+                                        bias=False),
+                                    build_norm_layer(self.norm_cfg,
+                                                     in_channels[j])[1],
+                                    nn.ReLU(inplace=False)))
+                    fuse_layer.append(nn.Sequential(*conv_downsamples))
+            fuse_layers.append(nn.ModuleList(fuse_layer))
+
+        return nn.ModuleList(fuse_layers)
+
+    def forward(self, x):
+        """Forward function."""
+        if self.num_branches == 1:
+            return [self.branches[0](x[0])]
+
+        for i in range(self.num_branches):
+            x[i] = self.branches[i](x[i])
+
+        x_fuse = []
+        for i in range(len(self.fuse_layers)):
+            y = 0
+            for j in range(self.num_branches):
+                if i == j:
+                    y += x[j]
+                else:
+                    y += self.fuse_layers[i][j](x[j])
+            x_fuse.append(self.relu(y))
+        return x_fuse
+
+
+@BACKBONES.register_module()
+class HRNet(nn.Module):
+    """HRNet backbone.
+
+    High-Resolution Representations for Labeling Pixels and Regions
+    arXiv: https://arxiv.org/abs/1904.04514
+
+    Args:
+        extra (dict): detailed configuration for each stage of HRNet.
+        in_channels (int): Number of input image channels. Default: 3.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        zero_init_residual (bool): whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
+
+    Example:
+        >>> from mmdet.models import HRNet
+        >>> import torch
+        >>> extra = dict(
+        >>>     stage1=dict(
+        >>>         num_modules=1,
+        >>>         num_branches=1,
+        >>>         block='BOTTLENECK',
+        >>>         num_blocks=(4, ),
+        >>>         num_channels=(64, )),
+        >>>     stage2=dict(
+        >>>         num_modules=1,
+        >>>         num_branches=2,
+        >>>         block='BASIC',
+        >>>         num_blocks=(4, 4),
+        >>>         num_channels=(32, 64)),
+        >>>     stage3=dict(
+        >>>         num_modules=4,
+        >>>         num_branches=3,
+        >>>         block='BASIC',
+        >>>         num_blocks=(4, 4, 4),
+        >>>         num_channels=(32, 64, 128)),
+        >>>     stage4=dict(
+        >>>         num_modules=3,
+        >>>         num_branches=4,
+        >>>         block='BASIC',
+        >>>         num_blocks=(4, 4, 4, 4),
+        >>>         num_channels=(32, 64, 128, 256)))
+        >>> self = HRNet(extra, in_channels=1)
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 1, 32, 32)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        (1, 32, 8, 8)
+        (1, 64, 4, 4)
+        (1, 128, 2, 2)
+        (1, 256, 1, 1)
+    """
+
+    blocks_dict = {'BASIC': BasicBlock, 'BOTTLENECK': Bottleneck}
+
+    def __init__(self,
+                 extra,
+                 in_channels=3,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 norm_eval=True,
+                 with_cp=False,
+                 zero_init_residual=False):
+        super(HRNet, self).__init__()
+        self.extra = extra
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.norm_eval = norm_eval
+        self.with_cp = with_cp
+        self.zero_init_residual = zero_init_residual
+
+        # stem net
+        self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
+        self.norm2_name, norm2 = build_norm_layer(self.norm_cfg, 64, postfix=2)
+
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            in_channels,
+            64,
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            bias=False)
+
+        self.add_module(self.norm1_name, norm1)
+        self.conv2 = build_conv_layer(
+            self.conv_cfg,
+            64,
+            64,
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            bias=False)
+
+        self.add_module(self.norm2_name, norm2)
+        self.relu = nn.ReLU(inplace=True)
+
+        # stage 1
+        self.stage1_cfg = self.extra['stage1']
+        num_channels = self.stage1_cfg['num_channels'][0]
+        block_type = self.stage1_cfg['block']
+        num_blocks = self.stage1_cfg['num_blocks'][0]
+
+        block = self.blocks_dict[block_type]
+        stage1_out_channels = num_channels * block.expansion
+        self.layer1 = self._make_layer(block, 64, num_channels, num_blocks)
+
+        # stage 2
+        self.stage2_cfg = self.extra['stage2']
+        num_channels = self.stage2_cfg['num_channels']
+        block_type = self.stage2_cfg['block']
+
+        block = self.blocks_dict[block_type]
+        num_channels = [channel * block.expansion for channel in num_channels]
+        self.transition1 = self._make_transition_layer([stage1_out_channels],
+                                                       num_channels)
+        self.stage2, pre_stage_channels = self._make_stage(
+            self.stage2_cfg, num_channels)
+
+        # stage 3
+        self.stage3_cfg = self.extra['stage3']
+        num_channels = self.stage3_cfg['num_channels']
+        block_type = self.stage3_cfg['block']
+
+        block = self.blocks_dict[block_type]
+        num_channels = [channel * block.expansion for channel in num_channels]
+        self.transition2 = self._make_transition_layer(pre_stage_channels,
+                                                       num_channels)
+        self.stage3, pre_stage_channels = self._make_stage(
+            self.stage3_cfg, num_channels)
+
+        # stage 4
+        self.stage4_cfg = self.extra['stage4']
+        num_channels = self.stage4_cfg['num_channels']
+        block_type = self.stage4_cfg['block']
+
+        block = self.blocks_dict[block_type]
+        num_channels = [channel * block.expansion for channel in num_channels]
+        self.transition3 = self._make_transition_layer(pre_stage_channels,
+                                                       num_channels)
+        self.stage4, pre_stage_channels = self._make_stage(
+            self.stage4_cfg, num_channels)
+
+    @property
+    def norm1(self):
+        """nn.Module: the normalization layer named "norm1" """
+        return getattr(self, self.norm1_name)
+
+    @property
+    def norm2(self):
+        """nn.Module: the normalization layer named "norm2" """
+        return getattr(self, self.norm2_name)
+
+    def _make_transition_layer(self, num_channels_pre_layer,
+                               num_channels_cur_layer):
+        num_branches_cur = len(num_channels_cur_layer)
+        num_branches_pre = len(num_channels_pre_layer)
+
+        transition_layers = []
+        for i in range(num_branches_cur):
+            if i < num_branches_pre:
+                if num_channels_cur_layer[i] != num_channels_pre_layer[i]:
+                    transition_layers.append(
+                        nn.Sequential(
+                            build_conv_layer(
+                                self.conv_cfg,
+                                num_channels_pre_layer[i],
+                                num_channels_cur_layer[i],
+                                kernel_size=3,
+                                stride=1,
+                                padding=1,
+                                bias=False),
+                            build_norm_layer(self.norm_cfg,
+                                             num_channels_cur_layer[i])[1],
+                            nn.ReLU(inplace=True)))
+                else:
+                    transition_layers.append(None)
+            else:
+                conv_downsamples = []
+                for j in range(i + 1 - num_branches_pre):
+                    in_channels = num_channels_pre_layer[-1]
+                    out_channels = num_channels_cur_layer[i] \
+                        if j == i - num_branches_pre else in_channels
+                    conv_downsamples.append(
+                        nn.Sequential(
+                            build_conv_layer(
+                                self.conv_cfg,
+                                in_channels,
+                                out_channels,
+                                kernel_size=3,
+                                stride=2,
+                                padding=1,
+                                bias=False),
+                            build_norm_layer(self.norm_cfg, out_channels)[1],
+                            nn.ReLU(inplace=True)))
+                transition_layers.append(nn.Sequential(*conv_downsamples))
+
+        return nn.ModuleList(transition_layers)
+
+    def _make_layer(self, block, inplanes, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                build_conv_layer(
+                    self.conv_cfg,
+                    inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    bias=False),
+                build_norm_layer(self.norm_cfg, planes * block.expansion)[1])
+
+        layers = []
+        layers.append(
+            block(
+                inplanes,
+                planes,
+                stride,
+                downsample=downsample,
+                with_cp=self.with_cp,
+                norm_cfg=self.norm_cfg,
+                conv_cfg=self.conv_cfg))
+        inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(
+                block(
+                    inplanes,
+                    planes,
+                    with_cp=self.with_cp,
+                    norm_cfg=self.norm_cfg,
+                    conv_cfg=self.conv_cfg))
+
+        return nn.Sequential(*layers)
+
+    def _make_stage(self, layer_config, in_channels, multiscale_output=True):
+        num_modules = layer_config['num_modules']
+        num_branches = layer_config['num_branches']
+        num_blocks = layer_config['num_blocks']
+        num_channels = layer_config['num_channels']
+        block = self.blocks_dict[layer_config['block']]
+
+        hr_modules = []
+        for i in range(num_modules):
+            # multi_scale_output is only used for the last module
+            if not multiscale_output and i == num_modules - 1:
+                reset_multiscale_output = False
+            else:
+                reset_multiscale_output = True
+
+            hr_modules.append(
+                HRModule(
+                    num_branches,
+                    block,
+                    num_blocks,
+                    in_channels,
+                    num_channels,
+                    reset_multiscale_output,
+                    with_cp=self.with_cp,
+                    norm_cfg=self.norm_cfg,
+                    conv_cfg=self.conv_cfg))
+
+        return nn.Sequential(*hr_modules), in_channels
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in backbone.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
+                    constant_init(m, 1)
+
+            if self.zero_init_residual:
+                for m in self.modules():
+                    if isinstance(m, Bottleneck):
+                        constant_init(m.norm3, 0)
+                    elif isinstance(m, BasicBlock):
+                        constant_init(m.norm2, 0)
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.relu(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        x = self.relu(x)
+        x = self.layer1(x)
+
+        x_list = []
+        for i in range(self.stage2_cfg['num_branches']):
+            if self.transition1[i] is not None:
+                x_list.append(self.transition1[i](x))
+            else:
+                x_list.append(x)
+        y_list = self.stage2(x_list)
+
+        x_list = []
+        for i in range(self.stage3_cfg['num_branches']):
+            if self.transition2[i] is not None:
+                x_list.append(self.transition2[i](y_list[-1]))
+            else:
+                x_list.append(y_list[i])
+        y_list = self.stage3(x_list)
+
+        x_list = []
+        for i in range(self.stage4_cfg['num_branches']):
+            if self.transition3[i] is not None:
+                x_list.append(self.transition3[i](y_list[-1]))
+            else:
+                x_list.append(y_list[i])
+        y_list = self.stage4(x_list)
+
+        return y_list
+
+    def train(self, mode=True):
+        """Convert the model into training mode whill keeping the normalization
+        layer freezed."""
+        super(HRNet, self).train(mode)
+        if mode and self.norm_eval:
+            for m in self.modules():
+                # trick: eval have effect on BatchNorm only
+                if isinstance(m, _BatchNorm):
+                    m.eval()
diff --git a/detection/scrfd/mmdet/models/backbones/mobilenet.py b/detection/scrfd/mmdet/models/backbones/mobilenet.py
new file mode 100755
index 000000000..8dccab8fd
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/mobilenet.py
@@ -0,0 +1,129 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn as nn
+from mmdet.utils import get_root_logger
+from mmcv.cnn import (build_conv_layer, build_norm_layer, build_plugin_layer,
+                      constant_init, kaiming_init)
+from mmcv.runner import load_checkpoint
+from torch.nn.modules.batchnorm import _BatchNorm
+from ..builder import BACKBONES
+
+@BACKBONES.register_module()
+class MobileNetV1(nn.Module):
+    def __init__(self,
+                 in_channels=3,
+                 #base_channels=32,
+                 block_cfg = None,
+                 num_stages=4,
+                 out_indices=(0, 1, 2, 3)):
+        super(MobileNetV1, self).__init__()
+        self.out_indices = out_indices
+
+        def conv_bn(inp, oup, stride):
+            return nn.Sequential(
+                nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+                nn.BatchNorm2d(oup),
+                nn.ReLU(inplace=True)
+            )
+
+        def conv_dw(inp, oup, stride):
+            return nn.Sequential(
+                nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
+                nn.BatchNorm2d(inp),
+                nn.ReLU(inplace=True),
+
+                nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
+                nn.BatchNorm2d(oup),
+                nn.ReLU(inplace=True),
+            )
+
+        if block_cfg is None:
+            stage_planes = [8, 16, 32, 64, 128, 256]  #0.25 default
+            stage_blocks = [2,4,4,2]
+        else:
+            stage_planes = block_cfg['stage_planes']
+            stage_blocks = block_cfg['stage_blocks']
+        assert len(stage_planes)==6
+        assert len(stage_blocks)==4
+        self.stem = nn.Sequential(
+                conv_bn(3, stage_planes[0], 2),
+                conv_dw(stage_planes[0], stage_planes[1], 1),
+                )
+        self.stage_layers = []
+        for i, num_blocks in enumerate(stage_blocks):
+            _layers = []
+            for n in range(num_blocks):
+                if n==0:
+                    _layer = conv_dw(stage_planes[i+1], stage_planes[i+2], 2)
+                else:
+                    _layer = conv_dw(stage_planes[i+2], stage_planes[i+2], 1)
+                _layers.append(_layer)
+
+            _block = nn.Sequential(*_layers)
+            layer_name = f'layer{i + 1}'
+            self.add_module(layer_name, _block)
+            self.stage_layers.append(layer_name)
+
+        #bc = base_channels
+        #self.stages = nn.ModuleDict()
+        #self.stage0 = nn.Sequential(
+        #    conv_bn(3, bc, 2),
+        #    conv_dw(bc, bc*2, 1),
+        #    conv_dw(bc*2, bc*4, 2),
+        #    conv_dw(bc*4, bc*4, 1),
+        #)
+        #self.stage1 = nn.Sequential(
+        #    conv_dw(bc*4, bc*8, 2),
+        #    conv_dw(bc*8, bc*8, 1),
+
+        #    conv_dw(bc*8, bc*8, 1),
+        #    conv_dw(bc*8, bc*8, 1),
+        #)
+        #self.stage2 = nn.Sequential(
+        #    conv_dw(bc*8, bc*16, 2),
+        #    conv_dw(bc*16, bc*16, 1),
+        #    conv_dw(bc*16, bc*16, 1),
+        #    conv_dw(bc*16, bc*16, 1),
+        #    #conv_dw(bc*16, bc*16, 1),
+        #    #conv_dw(bc*16, bc*16, 1),
+        #)
+        #self.stage3 = nn.Sequential(
+        #    conv_dw(bc*16, bc*32, 2),
+        #    conv_dw(bc*32, bc*32, 1),
+        #)
+        #self.stages = [self.stage0, self.stage1, self.stage2, self.stage3]
+
+
+
+    def forward(self, x):
+        output = []
+        x = self.stem(x)
+        for i, layer_name in enumerate(self.stage_layers):
+            stage_layer = getattr(self, layer_name)
+            x = stage_layer(x)
+            if i in self.out_indices:
+                output.append(x)
+
+        return tuple(output)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in backbone.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
+                    constant_init(m, 1)
+        else:
+            raise TypeError('pretrained must be a str or None')
diff --git a/detection/scrfd/mmdet/models/backbones/regnet.py b/detection/scrfd/mmdet/models/backbones/regnet.py
new file mode 100755
index 000000000..b786a3f8a
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/regnet.py
@@ -0,0 +1,325 @@
+import numpy as np
+import torch.nn as nn
+from mmcv.cnn import build_conv_layer, build_norm_layer
+
+from ..builder import BACKBONES
+from .resnet import ResNet
+from .resnext import Bottleneck
+
+
+@BACKBONES.register_module()
+class RegNet(ResNet):
+    """RegNet backbone.
+
+    More details can be found in `paper <https://arxiv.org/abs/2003.13678>`_ .
+
+    Args:
+        arch (dict): The parameter of RegNets.
+
+            - w0 (int): initial width
+            - wa (float): slope of width
+            - wm (float): quantization parameter to quantize the width
+            - depth (int): depth of the backbone
+            - group_w (int): width of group
+            - bot_mul (float): bottleneck ratio, i.e. expansion of bottlneck.
+        strides (Sequence[int]): Strides of the first block of each stage.
+        base_channels (int): Base channels after stem layer.
+        in_channels (int): Number of input image channels. Default: 3.
+        dilations (Sequence[int]): Dilation of each stage.
+        out_indices (Sequence[int]): Output from which stages.
+        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
+            layer is the 3x3 conv layer, otherwise the stride-two layer is
+            the first 1x1 conv layer.
+        frozen_stages (int): Stages to be frozen (all param fixed). -1 means
+            not freezing any parameters.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        zero_init_residual (bool): whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
+
+    Example:
+        >>> from mmdet.models import RegNet
+        >>> import torch
+        >>> self = RegNet(
+                arch=dict(
+                    w0=88,
+                    wa=26.31,
+                    wm=2.25,
+                    group_w=48,
+                    depth=25,
+                    bot_mul=1.0))
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 32, 32)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        (1, 96, 8, 8)
+        (1, 192, 4, 4)
+        (1, 432, 2, 2)
+        (1, 1008, 1, 1)
+    """
+    arch_settings = {
+        'regnetx_400mf':
+        dict(w0=24, wa=24.48, wm=2.54, group_w=16, depth=22, bot_mul=1.0),
+        'regnetx_800mf':
+        dict(w0=56, wa=35.73, wm=2.28, group_w=16, depth=16, bot_mul=1.0),
+        'regnetx_1.6gf':
+        dict(w0=80, wa=34.01, wm=2.25, group_w=24, depth=18, bot_mul=1.0),
+        'regnetx_3.2gf':
+        dict(w0=88, wa=26.31, wm=2.25, group_w=48, depth=25, bot_mul=1.0),
+        'regnetx_4.0gf':
+        dict(w0=96, wa=38.65, wm=2.43, group_w=40, depth=23, bot_mul=1.0),
+        'regnetx_6.4gf':
+        dict(w0=184, wa=60.83, wm=2.07, group_w=56, depth=17, bot_mul=1.0),
+        'regnetx_8.0gf':
+        dict(w0=80, wa=49.56, wm=2.88, group_w=120, depth=23, bot_mul=1.0),
+        'regnetx_12gf':
+        dict(w0=168, wa=73.36, wm=2.37, group_w=112, depth=19, bot_mul=1.0),
+    }
+
+    def __init__(self,
+                 arch,
+                 in_channels=3,
+                 stem_channels=32,
+                 base_channels=32,
+                 strides=(2, 2, 2, 2),
+                 dilations=(1, 1, 1, 1),
+                 out_indices=(0, 1, 2, 3),
+                 style='pytorch',
+                 deep_stem=False,
+                 avg_down=False,
+                 frozen_stages=-1,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 norm_eval=True,
+                 dcn=None,
+                 stage_with_dcn=(False, False, False, False),
+                 plugins=None,
+                 with_cp=False,
+                 zero_init_residual=True):
+        super(ResNet, self).__init__()
+
+        # Generate RegNet parameters first
+        if isinstance(arch, str):
+            assert arch in self.arch_settings, \
+                f'"arch": "{arch}" is not one of the' \
+                ' arch_settings'
+            arch = self.arch_settings[arch]
+        elif not isinstance(arch, dict):
+            raise ValueError('Expect "arch" to be either a string '
+                             f'or a dict, got {type(arch)}')
+
+        widths, num_stages = self.generate_regnet(
+            arch['w0'],
+            arch['wa'],
+            arch['wm'],
+            arch['depth'],
+        )
+        # Convert to per stage format
+        stage_widths, stage_blocks = self.get_stages_from_blocks(widths)
+        # Generate group widths and bot muls
+        group_widths = [arch['group_w'] for _ in range(num_stages)]
+        self.bottleneck_ratio = [arch['bot_mul'] for _ in range(num_stages)]
+        # Adjust the compatibility of stage_widths and group_widths
+        stage_widths, group_widths = self.adjust_width_group(
+            stage_widths, self.bottleneck_ratio, group_widths)
+
+        # Group params by stage
+        self.stage_widths = stage_widths
+        self.group_widths = group_widths
+        self.depth = sum(stage_blocks)
+        self.stem_channels = stem_channels
+        self.base_channels = base_channels
+        self.num_stages = num_stages
+        assert num_stages >= 1 and num_stages <= 4
+        self.strides = strides
+        self.dilations = dilations
+        assert len(strides) == len(dilations) == num_stages
+        self.out_indices = out_indices
+        assert max(out_indices) < num_stages
+        self.style = style
+        self.deep_stem = deep_stem
+        self.avg_down = avg_down
+        self.frozen_stages = frozen_stages
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.with_cp = with_cp
+        self.norm_eval = norm_eval
+        self.dcn = dcn
+        self.stage_with_dcn = stage_with_dcn
+        if dcn is not None:
+            assert len(stage_with_dcn) == num_stages
+        self.plugins = plugins
+        self.zero_init_residual = zero_init_residual
+        self.block = Bottleneck
+        expansion_bak = self.block.expansion
+        self.block.expansion = 1
+        self.stage_blocks = stage_blocks[:num_stages]
+
+        self._make_stem_layer(in_channels, stem_channels)
+
+        self.inplanes = stem_channels
+        self.res_layers = []
+        for i, num_blocks in enumerate(self.stage_blocks):
+            stride = self.strides[i]
+            dilation = self.dilations[i]
+            group_width = self.group_widths[i]
+            width = int(round(self.stage_widths[i] * self.bottleneck_ratio[i]))
+            stage_groups = width // group_width
+
+            dcn = self.dcn if self.stage_with_dcn[i] else None
+            if self.plugins is not None:
+                stage_plugins = self.make_stage_plugins(self.plugins, i)
+            else:
+                stage_plugins = None
+
+            res_layer = self.make_res_layer(
+                block=self.block,
+                inplanes=self.inplanes,
+                planes=self.stage_widths[i],
+                num_blocks=num_blocks,
+                stride=stride,
+                dilation=dilation,
+                style=self.style,
+                avg_down=self.avg_down,
+                with_cp=self.with_cp,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg,
+                dcn=dcn,
+                plugins=stage_plugins,
+                groups=stage_groups,
+                base_width=group_width,
+                base_channels=self.stage_widths[i])
+            self.inplanes = self.stage_widths[i]
+            layer_name = f'layer{i + 1}'
+            self.add_module(layer_name, res_layer)
+            self.res_layers.append(layer_name)
+
+        self._freeze_stages()
+
+        self.feat_dim = stage_widths[-1]
+        self.block.expansion = expansion_bak
+
+    def _make_stem_layer(self, in_channels, base_channels):
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            in_channels,
+            base_channels,
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            bias=False)
+        self.norm1_name, norm1 = build_norm_layer(
+            self.norm_cfg, base_channels, postfix=1)
+        self.add_module(self.norm1_name, norm1)
+        self.relu = nn.ReLU(inplace=True)
+
+    def generate_regnet(self,
+                        initial_width,
+                        width_slope,
+                        width_parameter,
+                        depth,
+                        divisor=8):
+        """Generates per block width from RegNet parameters.
+
+        Args:
+            initial_width ([int]): Initial width of the backbone
+            width_slope ([float]): Slope of the quantized linear function
+            width_parameter ([int]): Parameter used to quantize the width.
+            depth ([int]): Depth of the backbone.
+            divisor (int, optional): The divisor of channels. Defaults to 8.
+
+        Returns:
+            list, int: return a list of widths of each stage and the number \
+                of stages
+        """
+        assert width_slope >= 0
+        assert initial_width > 0
+        assert width_parameter > 1
+        assert initial_width % divisor == 0
+        widths_cont = np.arange(depth) * width_slope + initial_width
+        ks = np.round(
+            np.log(widths_cont / initial_width) / np.log(width_parameter))
+        widths = initial_width * np.power(width_parameter, ks)
+        widths = np.round(np.divide(widths, divisor)) * divisor
+        num_stages = len(np.unique(widths))
+        widths, widths_cont = widths.astype(int).tolist(), widths_cont.tolist()
+        return widths, num_stages
+
+    @staticmethod
+    def quantize_float(number, divisor):
+        """Converts a float to closest non-zero int divisible by divior.
+
+        Args:
+            number (int): Original number to be quantized.
+            divisor (int): Divisor used to quantize the number.
+
+        Returns:
+            int: quantized number that is divisible by devisor.
+        """
+        return int(round(number / divisor) * divisor)
+
+    def adjust_width_group(self, widths, bottleneck_ratio, groups):
+        """Adjusts the compatibility of widths and groups.
+
+        Args:
+            widths (list[int]): Width of each stage.
+            bottleneck_ratio (float): Bottleneck ratio.
+            groups (int): number of groups in each stage
+
+        Returns:
+            tuple(list): The adjusted widths and groups of each stage.
+        """
+        bottleneck_width = [
+            int(w * b) for w, b in zip(widths, bottleneck_ratio)
+        ]
+        groups = [min(g, w_bot) for g, w_bot in zip(groups, bottleneck_width)]
+        bottleneck_width = [
+            self.quantize_float(w_bot, g)
+            for w_bot, g in zip(bottleneck_width, groups)
+        ]
+        widths = [
+            int(w_bot / b)
+            for w_bot, b in zip(bottleneck_width, bottleneck_ratio)
+        ]
+        return widths, groups
+
+    def get_stages_from_blocks(self, widths):
+        """Gets widths/stage_blocks of network at each stage.
+
+        Args:
+            widths (list[int]): Width in each stage.
+
+        Returns:
+            tuple(list): width and depth of each stage
+        """
+        width_diff = [
+            width != width_prev
+            for width, width_prev in zip(widths + [0], [0] + widths)
+        ]
+        stage_widths = [
+            width for width, diff in zip(widths, width_diff[:-1]) if diff
+        ]
+        stage_blocks = np.diff([
+            depth for depth, diff in zip(range(len(width_diff)), width_diff)
+            if diff
+        ]).tolist()
+        return stage_widths, stage_blocks
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.relu(x)
+
+        outs = []
+        for i, layer_name in enumerate(self.res_layers):
+            res_layer = getattr(self, layer_name)
+            x = res_layer(x)
+            if i in self.out_indices:
+                outs.append(x)
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/backbones/res2net.py b/detection/scrfd/mmdet/models/backbones/res2net.py
new file mode 100755
index 000000000..7901b7f2f
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/res2net.py
@@ -0,0 +1,351 @@
+import math
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as cp
+from mmcv.cnn import (build_conv_layer, build_norm_layer, constant_init,
+                      kaiming_init)
+from mmcv.runner import load_checkpoint
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from mmdet.utils import get_root_logger
+from ..builder import BACKBONES
+from .resnet import Bottleneck as _Bottleneck
+from .resnet import ResNet
+
+
+class Bottle2neck(_Bottleneck):
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 scales=4,
+                 base_width=26,
+                 base_channels=64,
+                 stage_type='normal',
+                 **kwargs):
+        """Bottle2neck block for Res2Net.
+
+        If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
+        it is "caffe", the stride-two layer is the first 1x1 conv layer.
+        """
+        super(Bottle2neck, self).__init__(inplanes, planes, **kwargs)
+        assert scales > 1, 'Res2Net degenerates to ResNet when scales = 1.'
+        width = int(math.floor(self.planes * (base_width / base_channels)))
+
+        self.norm1_name, norm1 = build_norm_layer(
+            self.norm_cfg, width * scales, postfix=1)
+        self.norm3_name, norm3 = build_norm_layer(
+            self.norm_cfg, self.planes * self.expansion, postfix=3)
+
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            self.inplanes,
+            width * scales,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+
+        if stage_type == 'stage' and self.conv2_stride != 1:
+            self.pool = nn.AvgPool2d(
+                kernel_size=3, stride=self.conv2_stride, padding=1)
+        convs = []
+        bns = []
+
+        fallback_on_stride = False
+        if self.with_dcn:
+            fallback_on_stride = self.dcn.pop('fallback_on_stride', False)
+        if not self.with_dcn or fallback_on_stride:
+            for i in range(scales - 1):
+                convs.append(
+                    build_conv_layer(
+                        self.conv_cfg,
+                        width,
+                        width,
+                        kernel_size=3,
+                        stride=self.conv2_stride,
+                        padding=self.dilation,
+                        dilation=self.dilation,
+                        bias=False))
+                bns.append(
+                    build_norm_layer(self.norm_cfg, width, postfix=i + 1)[1])
+            self.convs = nn.ModuleList(convs)
+            self.bns = nn.ModuleList(bns)
+        else:
+            assert self.conv_cfg is None, 'conv_cfg must be None for DCN'
+            for i in range(scales - 1):
+                convs.append(
+                    build_conv_layer(
+                        self.dcn,
+                        width,
+                        width,
+                        kernel_size=3,
+                        stride=self.conv2_stride,
+                        padding=self.dilation,
+                        dilation=self.dilation,
+                        bias=False))
+                bns.append(
+                    build_norm_layer(self.norm_cfg, width, postfix=i + 1)[1])
+            self.convs = nn.ModuleList(convs)
+            self.bns = nn.ModuleList(bns)
+
+        self.conv3 = build_conv_layer(
+            self.conv_cfg,
+            width * scales,
+            self.planes * self.expansion,
+            kernel_size=1,
+            bias=False)
+        self.add_module(self.norm3_name, norm3)
+
+        self.stage_type = stage_type
+        self.scales = scales
+        self.width = width
+        delattr(self, 'conv2')
+        delattr(self, self.norm2_name)
+
+    def forward(self, x):
+        """Forward function."""
+
+        def _inner_forward(x):
+            identity = x
+
+            out = self.conv1(x)
+            out = self.norm1(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv1_plugin_names)
+
+            spx = torch.split(out, self.width, 1)
+            sp = self.convs[0](spx[0].contiguous())
+            sp = self.relu(self.bns[0](sp))
+            out = sp
+            for i in range(1, self.scales - 1):
+                if self.stage_type == 'stage':
+                    sp = spx[i]
+                else:
+                    sp = sp + spx[i]
+                sp = self.convs[i](sp.contiguous())
+                sp = self.relu(self.bns[i](sp))
+                out = torch.cat((out, sp), 1)
+
+            if self.stage_type == 'normal' or self.conv2_stride == 1:
+                out = torch.cat((out, spx[self.scales - 1]), 1)
+            elif self.stage_type == 'stage':
+                out = torch.cat((out, self.pool(spx[self.scales - 1])), 1)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv2_plugin_names)
+
+            out = self.conv3(out)
+            out = self.norm3(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv3_plugin_names)
+
+            if self.downsample is not None:
+                identity = self.downsample(x)
+
+            out += identity
+
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        out = self.relu(out)
+
+        return out
+
+
+class Res2Layer(nn.Sequential):
+    """Res2Layer to build Res2Net style backbone.
+
+    Args:
+        block (nn.Module): block used to build ResLayer.
+        inplanes (int): inplanes of block.
+        planes (int): planes of block.
+        num_blocks (int): number of blocks.
+        stride (int): stride of the first block. Default: 1
+        avg_down (bool): Use AvgPool instead of stride conv when
+            downsampling in the bottle2neck. Default: False
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: dict(type='BN')
+        scales (int): Scales used in Res2Net. Default: 4
+        base_width (int): Basic width of each scale. Default: 26
+    """
+
+    def __init__(self,
+                 block,
+                 inplanes,
+                 planes,
+                 num_blocks,
+                 stride=1,
+                 avg_down=True,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 scales=4,
+                 base_width=26,
+                 **kwargs):
+        self.block = block
+
+        downsample = None
+        if stride != 1 or inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.AvgPool2d(
+                    kernel_size=stride,
+                    stride=stride,
+                    ceil_mode=True,
+                    count_include_pad=False),
+                build_conv_layer(
+                    conv_cfg,
+                    inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=1,
+                    bias=False),
+                build_norm_layer(norm_cfg, planes * block.expansion)[1],
+            )
+
+        layers = []
+        layers.append(
+            block(
+                inplanes=inplanes,
+                planes=planes,
+                stride=stride,
+                downsample=downsample,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                scales=scales,
+                base_width=base_width,
+                stage_type='stage',
+                **kwargs))
+        inplanes = planes * block.expansion
+        for i in range(1, num_blocks):
+            layers.append(
+                block(
+                    inplanes=inplanes,
+                    planes=planes,
+                    stride=1,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    scales=scales,
+                    base_width=base_width,
+                    **kwargs))
+        super(Res2Layer, self).__init__(*layers)
+
+
+@BACKBONES.register_module()
+class Res2Net(ResNet):
+    """Res2Net backbone.
+
+    Args:
+        scales (int): Scales used in Res2Net. Default: 4
+        base_width (int): Basic width of each scale. Default: 26
+        depth (int): Depth of res2net, from {50, 101, 152}.
+        in_channels (int): Number of input image channels. Default: 3.
+        num_stages (int): Res2net stages. Default: 4.
+        strides (Sequence[int]): Strides of the first block of each stage.
+        dilations (Sequence[int]): Dilation of each stage.
+        out_indices (Sequence[int]): Output from which stages.
+        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
+            layer is the 3x3 conv layer, otherwise the stride-two layer is
+            the first 1x1 conv layer.
+        deep_stem (bool): Replace 7x7 conv in input stem with 3 3x3 conv
+        avg_down (bool): Use AvgPool instead of stride conv when
+            downsampling in the bottle2neck.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+        plugins (list[dict]): List of plugins for stages, each dict contains:
+
+            - cfg (dict, required): Cfg dict to build plugin.
+            - position (str, required): Position inside block to insert
+              plugin, options are 'after_conv1', 'after_conv2', 'after_conv3'.
+            - stages (tuple[bool], optional): Stages to apply plugin, length
+              should be same as 'num_stages'.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        zero_init_residual (bool): Whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
+
+    Example:
+        >>> from mmdet.models import Res2Net
+        >>> import torch
+        >>> self = Res2Net(depth=50, scales=4, base_width=26)
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 32, 32)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        (1, 256, 8, 8)
+        (1, 512, 4, 4)
+        (1, 1024, 2, 2)
+        (1, 2048, 1, 1)
+    """
+
+    arch_settings = {
+        50: (Bottle2neck, (3, 4, 6, 3)),
+        101: (Bottle2neck, (3, 4, 23, 3)),
+        152: (Bottle2neck, (3, 8, 36, 3))
+    }
+
+    def __init__(self,
+                 scales=4,
+                 base_width=26,
+                 style='pytorch',
+                 deep_stem=True,
+                 avg_down=True,
+                 **kwargs):
+        self.scales = scales
+        self.base_width = base_width
+        super(Res2Net, self).__init__(
+            style='pytorch', deep_stem=True, avg_down=True, **kwargs)
+
+    def make_res_layer(self, **kwargs):
+        return Res2Layer(
+            scales=self.scales,
+            base_width=self.base_width,
+            base_channels=self.base_channels,
+            **kwargs)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in backbone.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
+                    constant_init(m, 1)
+
+            if self.dcn is not None:
+                for m in self.modules():
+                    if isinstance(m, Bottle2neck):
+                        # dcn in Res2Net bottle2neck is in ModuleList
+                        for n in m.convs:
+                            if hasattr(n, 'conv_offset'):
+                                constant_init(n.conv_offset, 0)
+
+            if self.zero_init_residual:
+                for m in self.modules():
+                    if isinstance(m, Bottle2neck):
+                        constant_init(m.norm3, 0)
+        else:
+            raise TypeError('pretrained must be a str or None')
diff --git a/detection/scrfd/mmdet/models/backbones/resnest.py b/detection/scrfd/mmdet/models/backbones/resnest.py
new file mode 100755
index 000000000..48e1d8bfa
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/resnest.py
@@ -0,0 +1,317 @@
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as cp
+from mmcv.cnn import build_conv_layer, build_norm_layer
+
+from ..builder import BACKBONES
+from ..utils import ResLayer
+from .resnet import Bottleneck as _Bottleneck
+from .resnet import ResNetV1d
+
+
+class RSoftmax(nn.Module):
+    """Radix Softmax module in ``SplitAttentionConv2d``.
+
+    Args:
+        radix (int): Radix of input.
+        groups (int): Groups of input.
+    """
+
+    def __init__(self, radix, groups):
+        super().__init__()
+        self.radix = radix
+        self.groups = groups
+
+    def forward(self, x):
+        batch = x.size(0)
+        if self.radix > 1:
+            x = x.view(batch, self.groups, self.radix, -1).transpose(1, 2)
+            x = F.softmax(x, dim=1)
+            x = x.reshape(batch, -1)
+        else:
+            x = torch.sigmoid(x)
+        return x
+
+
+class SplitAttentionConv2d(nn.Module):
+    """Split-Attention Conv2d in ResNeSt.
+
+    Args:
+        in_channels (int): Number of channels in the input feature map.
+        channels (int): Number of intermediate channels.
+        kernel_size (int | tuple[int]): Size of the convolution kernel.
+        stride (int | tuple[int]): Stride of the convolution.
+        padding (int | tuple[int]): Zero-padding added to both sides of
+        dilation (int | tuple[int]): Spacing between kernel elements.
+        groups (int): Number of blocked connections from input channels to
+            output channels.
+        groups (int): Same as nn.Conv2d.
+        radix (int): Radix of SpltAtConv2d. Default: 2
+        reduction_factor (int): Reduction factor of inter_channels. Default: 4.
+        conv_cfg (dict): Config dict for convolution layer. Default: None,
+            which means using conv2d.
+        norm_cfg (dict): Config dict for normalization layer. Default: None.
+        dcn (dict): Config dict for DCN. Default: None.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1,
+                 radix=2,
+                 reduction_factor=4,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 dcn=None):
+        super(SplitAttentionConv2d, self).__init__()
+        inter_channels = max(in_channels * radix // reduction_factor, 32)
+        self.radix = radix
+        self.groups = groups
+        self.channels = channels
+        self.with_dcn = dcn is not None
+        self.dcn = dcn
+        fallback_on_stride = False
+        if self.with_dcn:
+            fallback_on_stride = self.dcn.pop('fallback_on_stride', False)
+        if self.with_dcn and not fallback_on_stride:
+            assert conv_cfg is None, 'conv_cfg must be None for DCN'
+            conv_cfg = dcn
+        self.conv = build_conv_layer(
+            conv_cfg,
+            in_channels,
+            channels * radix,
+            kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups * radix,
+            bias=False)
+        # To be consistent with original implementation, starting from 0
+        self.norm0_name, norm0 = build_norm_layer(
+            norm_cfg, channels * radix, postfix=0)
+        self.add_module(self.norm0_name, norm0)
+        self.relu = nn.ReLU(inplace=True)
+        self.fc1 = build_conv_layer(
+            None, channels, inter_channels, 1, groups=self.groups)
+        self.norm1_name, norm1 = build_norm_layer(
+            norm_cfg, inter_channels, postfix=1)
+        self.add_module(self.norm1_name, norm1)
+        self.fc2 = build_conv_layer(
+            None, inter_channels, channels * radix, 1, groups=self.groups)
+        self.rsoftmax = RSoftmax(radix, groups)
+
+    @property
+    def norm0(self):
+        """nn.Module: the normalization layer named "norm0" """
+        return getattr(self, self.norm0_name)
+
+    @property
+    def norm1(self):
+        """nn.Module: the normalization layer named "norm1" """
+        return getattr(self, self.norm1_name)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.norm0(x)
+        x = self.relu(x)
+
+        batch, rchannel = x.shape[:2]
+        batch = x.size(0)
+        if self.radix > 1:
+            splits = x.view(batch, self.radix, -1, *x.shape[2:])
+            gap = splits.sum(dim=1)
+        else:
+            gap = x
+        gap = F.adaptive_avg_pool2d(gap, 1)
+        gap = self.fc1(gap)
+
+        gap = self.norm1(gap)
+        gap = self.relu(gap)
+
+        atten = self.fc2(gap)
+        atten = self.rsoftmax(atten).view(batch, -1, 1, 1)
+
+        if self.radix > 1:
+            attens = atten.view(batch, self.radix, -1, *atten.shape[2:])
+            out = torch.sum(attens * splits, dim=1)
+        else:
+            out = atten * x
+        return out.contiguous()
+
+
+class Bottleneck(_Bottleneck):
+    """Bottleneck block for ResNeSt.
+
+    Args:
+        inplane (int): Input planes of this block.
+        planes (int): Middle planes of this block.
+        groups (int): Groups of conv2.
+        base_width (int): Base of width in terms of base channels. Default: 4.
+        base_channels (int): Base of channels for calculating width.
+            Default: 64.
+        radix (int): Radix of SpltAtConv2d. Default: 2
+        reduction_factor (int): Reduction factor of inter_channels in
+            SplitAttentionConv2d. Default: 4.
+        avg_down_stride (bool): Whether to use average pool for stride in
+            Bottleneck. Default: True.
+        kwargs (dict): Key word arguments for base class.
+    """
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 groups=1,
+                 base_width=4,
+                 base_channels=64,
+                 radix=2,
+                 reduction_factor=4,
+                 avg_down_stride=True,
+                 **kwargs):
+        """Bottleneck block for ResNeSt."""
+        super(Bottleneck, self).__init__(inplanes, planes, **kwargs)
+
+        if groups == 1:
+            width = self.planes
+        else:
+            width = math.floor(self.planes *
+                               (base_width / base_channels)) * groups
+
+        self.avg_down_stride = avg_down_stride and self.conv2_stride > 1
+
+        self.norm1_name, norm1 = build_norm_layer(
+            self.norm_cfg, width, postfix=1)
+        self.norm3_name, norm3 = build_norm_layer(
+            self.norm_cfg, self.planes * self.expansion, postfix=3)
+
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            self.inplanes,
+            width,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+        self.with_modulated_dcn = False
+        self.conv2 = SplitAttentionConv2d(
+            width,
+            width,
+            kernel_size=3,
+            stride=1 if self.avg_down_stride else self.conv2_stride,
+            padding=self.dilation,
+            dilation=self.dilation,
+            groups=groups,
+            radix=radix,
+            reduction_factor=reduction_factor,
+            conv_cfg=self.conv_cfg,
+            norm_cfg=self.norm_cfg,
+            dcn=self.dcn)
+        delattr(self, self.norm2_name)
+
+        if self.avg_down_stride:
+            self.avd_layer = nn.AvgPool2d(3, self.conv2_stride, padding=1)
+
+        self.conv3 = build_conv_layer(
+            self.conv_cfg,
+            width,
+            self.planes * self.expansion,
+            kernel_size=1,
+            bias=False)
+        self.add_module(self.norm3_name, norm3)
+
+    def forward(self, x):
+
+        def _inner_forward(x):
+            identity = x
+
+            out = self.conv1(x)
+            out = self.norm1(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv1_plugin_names)
+
+            out = self.conv2(out)
+
+            if self.avg_down_stride:
+                out = self.avd_layer(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv2_plugin_names)
+
+            out = self.conv3(out)
+            out = self.norm3(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv3_plugin_names)
+
+            if self.downsample is not None:
+                identity = self.downsample(x)
+
+            out += identity
+
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        out = self.relu(out)
+
+        return out
+
+
+@BACKBONES.register_module()
+class ResNeSt(ResNetV1d):
+    """ResNeSt backbone.
+
+    Args:
+        groups (int): Number of groups of Bottleneck. Default: 1
+        base_width (int): Base width of Bottleneck. Default: 4
+        radix (int): Radix of SplitAttentionConv2d. Default: 2
+        reduction_factor (int): Reduction factor of inter_channels in
+            SplitAttentionConv2d. Default: 4.
+        avg_down_stride (bool): Whether to use average pool for stride in
+            Bottleneck. Default: True.
+        kwargs (dict): Keyword arguments for ResNet.
+    """
+
+    arch_settings = {
+        50: (Bottleneck, (3, 4, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3)),
+        200: (Bottleneck, (3, 24, 36, 3))
+    }
+
+    def __init__(self,
+                 groups=1,
+                 base_width=4,
+                 radix=2,
+                 reduction_factor=4,
+                 avg_down_stride=True,
+                 **kwargs):
+        self.groups = groups
+        self.base_width = base_width
+        self.radix = radix
+        self.reduction_factor = reduction_factor
+        self.avg_down_stride = avg_down_stride
+        super(ResNeSt, self).__init__(**kwargs)
+
+    def make_res_layer(self, **kwargs):
+        """Pack all blocks in a stage into a ``ResLayer``."""
+        return ResLayer(
+            groups=self.groups,
+            base_width=self.base_width,
+            base_channels=self.base_channels,
+            radix=self.radix,
+            reduction_factor=self.reduction_factor,
+            avg_down_stride=self.avg_down_stride,
+            **kwargs)
diff --git a/detection/scrfd/mmdet/models/backbones/resnet.py b/detection/scrfd/mmdet/models/backbones/resnet.py
new file mode 100755
index 000000000..e39e402f0
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/resnet.py
@@ -0,0 +1,713 @@
+import torch.nn as nn
+import torch.utils.checkpoint as cp
+from mmcv.cnn import (build_conv_layer, build_norm_layer, build_plugin_layer,
+                      constant_init, kaiming_init)
+from mmcv.runner import load_checkpoint
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from mmdet.utils import get_root_logger
+from ..builder import BACKBONES
+from ..utils import ResLayer
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 stride=1,
+                 dilation=1,
+                 downsample=None,
+                 style='pytorch',
+                 with_cp=False,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 dcn=None,
+                 plugins=None):
+        super(BasicBlock, self).__init__()
+        assert dcn is None, 'Not implemented yet.'
+        assert plugins is None, 'Not implemented yet.'
+
+        self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
+        self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
+        #print('init basic block:', inplanes, planes)
+
+        self.conv1 = build_conv_layer(
+            conv_cfg,
+            inplanes,
+            planes,
+            3,
+            stride=stride,
+            padding=dilation,
+            dilation=dilation,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+        self.conv2 = build_conv_layer(
+            conv_cfg, planes, planes, 3, padding=1, bias=False)
+        self.add_module(self.norm2_name, norm2)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+        self.dilation = dilation
+        self.with_cp = with_cp
+
+    @property
+    def norm1(self):
+        """nn.Module: normalization layer after the first convolution layer"""
+        return getattr(self, self.norm1_name)
+
+    @property
+    def norm2(self):
+        """nn.Module: normalization layer after the second convolution layer"""
+        return getattr(self, self.norm2_name)
+
+    def forward(self, x):
+        """Forward function."""
+
+        def _inner_forward(x):
+            identity = x
+
+            out = self.conv1(x)
+            out = self.norm1(out)
+            out = self.relu(out)
+
+            out = self.conv2(out)
+            out = self.norm2(out)
+
+            if self.downsample is not None:
+                identity = self.downsample(x)
+
+            out += identity
+
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 stride=1,
+                 dilation=1,
+                 downsample=None,
+                 style='pytorch',
+                 with_cp=False,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 dcn=None,
+                 plugins=None):
+        """Bottleneck block for ResNet.
+
+        If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
+        it is "caffe", the stride-two layer is the first 1x1 conv layer.
+        """
+        super(Bottleneck, self).__init__()
+        assert style in ['pytorch', 'caffe']
+        assert dcn is None or isinstance(dcn, dict)
+        assert plugins is None or isinstance(plugins, list)
+        if plugins is not None:
+            allowed_position = ['after_conv1', 'after_conv2', 'after_conv3']
+            assert all(p['position'] in allowed_position for p in plugins)
+
+        self.inplanes = inplanes
+        self.planes = planes
+        self.stride = stride
+        self.dilation = dilation
+        self.style = style
+        self.with_cp = with_cp
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.dcn = dcn
+        self.with_dcn = dcn is not None
+        self.plugins = plugins
+        self.with_plugins = plugins is not None
+
+        if self.with_plugins:
+            # collect plugins for conv1/conv2/conv3
+            self.after_conv1_plugins = [
+                plugin['cfg'] for plugin in plugins
+                if plugin['position'] == 'after_conv1'
+            ]
+            self.after_conv2_plugins = [
+                plugin['cfg'] for plugin in plugins
+                if plugin['position'] == 'after_conv2'
+            ]
+            self.after_conv3_plugins = [
+                plugin['cfg'] for plugin in plugins
+                if plugin['position'] == 'after_conv3'
+            ]
+
+        if self.style == 'pytorch':
+            self.conv1_stride = 1
+            self.conv2_stride = stride
+        else:
+            self.conv1_stride = stride
+            self.conv2_stride = 1
+
+        self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
+        self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
+        self.norm3_name, norm3 = build_norm_layer(
+            norm_cfg, planes * self.expansion, postfix=3)
+
+        self.conv1 = build_conv_layer(
+            conv_cfg,
+            inplanes,
+            planes,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+        fallback_on_stride = False
+        if self.with_dcn:
+            fallback_on_stride = dcn.pop('fallback_on_stride', False)
+        if not self.with_dcn or fallback_on_stride:
+            self.conv2 = build_conv_layer(
+                conv_cfg,
+                planes,
+                planes,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=dilation,
+                dilation=dilation,
+                bias=False)
+        else:
+            assert self.conv_cfg is None, 'conv_cfg must be None for DCN'
+            self.conv2 = build_conv_layer(
+                dcn,
+                planes,
+                planes,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=dilation,
+                dilation=dilation,
+                bias=False)
+
+        self.add_module(self.norm2_name, norm2)
+        self.conv3 = build_conv_layer(
+            conv_cfg,
+            planes,
+            planes * self.expansion,
+            kernel_size=1,
+            bias=False)
+        self.add_module(self.norm3_name, norm3)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+
+        if self.with_plugins:
+            self.after_conv1_plugin_names = self.make_block_plugins(
+                planes, self.after_conv1_plugins)
+            self.after_conv2_plugin_names = self.make_block_plugins(
+                planes, self.after_conv2_plugins)
+            self.after_conv3_plugin_names = self.make_block_plugins(
+                planes * self.expansion, self.after_conv3_plugins)
+
+    def make_block_plugins(self, in_channels, plugins):
+        """make plugins for block.
+
+        Args:
+            in_channels (int): Input channels of plugin.
+            plugins (list[dict]): List of plugins cfg to build.
+
+        Returns:
+            list[str]: List of the names of plugin.
+        """
+        assert isinstance(plugins, list)
+        plugin_names = []
+        for plugin in plugins:
+            plugin = plugin.copy()
+            name, layer = build_plugin_layer(
+                plugin,
+                in_channels=in_channels,
+                postfix=plugin.pop('postfix', ''))
+            assert not hasattr(self, name), f'duplicate plugin {name}'
+            self.add_module(name, layer)
+            plugin_names.append(name)
+        return plugin_names
+
+    def forward_plugin(self, x, plugin_names):
+        out = x
+        for name in plugin_names:
+            out = getattr(self, name)(x)
+        return out
+
+    @property
+    def norm1(self):
+        """nn.Module: normalization layer after the first convolution layer"""
+        return getattr(self, self.norm1_name)
+
+    @property
+    def norm2(self):
+        """nn.Module: normalization layer after the second convolution layer"""
+        return getattr(self, self.norm2_name)
+
+    @property
+    def norm3(self):
+        """nn.Module: normalization layer after the third convolution layer"""
+        return getattr(self, self.norm3_name)
+
+    def forward(self, x):
+        """Forward function."""
+
+        def _inner_forward(x):
+            identity = x
+            out = self.conv1(x)
+            out = self.norm1(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv1_plugin_names)
+
+            out = self.conv2(out)
+            out = self.norm2(out)
+            out = self.relu(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv2_plugin_names)
+
+            out = self.conv3(out)
+            out = self.norm3(out)
+
+            if self.with_plugins:
+                out = self.forward_plugin(out, self.after_conv3_plugin_names)
+
+            if self.downsample is not None:
+                identity = self.downsample(x)
+
+            out += identity
+
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        out = self.relu(out)
+
+        return out
+
+
+@BACKBONES.register_module()
+class ResNet(nn.Module):
+    """ResNet backbone.
+
+    Args:
+        depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
+        stem_channels (int | None): Number of stem channels. If not specified,
+            it will be the same as `base_channels`. Default: None.
+        base_channels (int): Number of base channels of res layer. Default: 64.
+        in_channels (int): Number of input image channels. Default: 3.
+        num_stages (int): Resnet stages. Default: 4.
+        strides (Sequence[int]): Strides of the first block of each stage.
+        dilations (Sequence[int]): Dilation of each stage.
+        out_indices (Sequence[int]): Output from which stages.
+        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
+            layer is the 3x3 conv layer, otherwise the stride-two layer is
+            the first 1x1 conv layer.
+        deep_stem (bool): Replace 7x7 conv in input stem with 3 3x3 conv
+        avg_down (bool): Use AvgPool instead of stride conv when
+            downsampling in the bottleneck.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+        plugins (list[dict]): List of plugins for stages, each dict contains:
+
+            - cfg (dict, required): Cfg dict to build plugin.
+            - position (str, required): Position inside block to insert
+              plugin, options are 'after_conv1', 'after_conv2', 'after_conv3'.
+            - stages (tuple[bool], optional): Stages to apply plugin, length
+              should be same as 'num_stages'.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        zero_init_residual (bool): Whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
+
+    Example:
+        >>> from mmdet.models import ResNet
+        >>> import torch
+        >>> self = ResNet(depth=18)
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 32, 32)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        (1, 64, 8, 8)
+        (1, 128, 4, 4)
+        (1, 256, 2, 2)
+        (1, 512, 1, 1)
+    """
+
+    arch_settings = {
+        0: (BasicBlock, (2, 2, 2, 2)),
+        18: (BasicBlock, (2, 2, 2, 2)),
+        19: (BasicBlock, (2, 4, 4, 1)),
+        20: (BasicBlock, (2, 3, 2, 2)),
+        22: (BasicBlock, (2, 4, 3, 1)),
+        24: (BasicBlock, (2, 4, 4, 1)),
+        26: (BasicBlock, (2, 4, 4, 2)),
+        28: (BasicBlock, (2, 5, 4, 2)),
+        29: (BasicBlock, (2, 6, 3, 2)),
+        30: (BasicBlock, (2, 5, 5, 2)),
+        32: (BasicBlock, (2, 6, 5, 2)),
+        34: (BasicBlock, (3, 4, 6, 3)),
+        35: (BasicBlock, (3, 6, 4, 3)),
+        38: (BasicBlock, (3, 8, 4, 3)),
+        40: (BasicBlock, (3, 8, 5, 3)),
+        50: (Bottleneck, (3, 4, 6, 3)),
+        56: (Bottleneck, (3, 8, 4, 3)),
+        68: (Bottleneck, (3, 10, 6, 3)),
+        74: (Bottleneck, (3, 12, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3))
+    }
+
+    def __init__(self,
+                 depth,
+                 in_channels=3,
+                 stem_channels=None,
+                 base_channels=64,
+                 num_stages=4,
+                 block_cfg = None,
+                 strides=(1, 2, 2, 2),
+                 dilations=(1, 1, 1, 1),
+                 out_indices=(0, 1, 2, 3),
+                 style='pytorch',
+                 deep_stem=False,
+                 avg_down=False,
+                 no_pool33=False,
+                 frozen_stages=-1,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 norm_eval=True,
+                 dcn=None,
+                 stage_with_dcn=(False, False, False, False),
+                 plugins=None,
+                 with_cp=False,
+                 zero_init_residual=True):
+        super(ResNet, self).__init__()
+        if depth not in self.arch_settings:
+            raise KeyError(f'invalid depth {depth} for resnet')
+        self.depth = depth
+        if stem_channels is None:
+            stem_channels = base_channels
+        self.stem_channels = stem_channels
+        self.base_channels = base_channels
+        self.num_stages = num_stages
+        assert num_stages >= 1 and num_stages <= 4
+        self.strides = strides
+        self.dilations = dilations
+        assert len(strides) == len(dilations) == num_stages
+        self.out_indices = out_indices
+        assert max(out_indices) < num_stages
+        self.style = style
+        self.deep_stem = deep_stem
+        self.avg_down = avg_down
+        self.no_pool33 = no_pool33
+        self.frozen_stages = frozen_stages
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.with_cp = with_cp
+        self.norm_eval = norm_eval
+        self.dcn = dcn
+        self.stage_with_dcn = stage_with_dcn
+        if dcn is not None:
+            assert len(stage_with_dcn) == num_stages
+        self.plugins = plugins
+        self.zero_init_residual = zero_init_residual
+        if block_cfg is None:
+            self.block, stage_blocks = self.arch_settings[depth]
+        else:
+            self.block = BasicBlock if block_cfg['block']=='BasicBlock' else Bottleneck
+            stage_blocks = block_cfg['stage_blocks']
+            assert len(stage_blocks)>=num_stages
+        self.stage_blocks = stage_blocks[:num_stages]
+        self.inplanes = stem_channels
+
+        self._make_stem_layer(in_channels, stem_channels)
+        if block_cfg is not None and 'stage_planes' in block_cfg:
+            stage_planes = block_cfg['stage_planes']
+        else:
+            stage_planes = [base_channels * 2**i for i in range(num_stages)]
+
+        #print('resnet cfg:', stage_blocks, stage_planes)
+        self.res_layers = []
+        for i, num_blocks in enumerate(self.stage_blocks):
+            stride = strides[i]
+            dilation = dilations[i]
+            dcn = self.dcn if self.stage_with_dcn[i] else None
+            if plugins is not None:
+                stage_plugins = self.make_stage_plugins(plugins, i)
+            else:
+                stage_plugins = None
+            #planes = base_channels * 2**i
+            planes = stage_planes[i]
+            #print('block detail:', i, self.inplanes, planes, stride)
+            res_layer = self.make_res_layer(
+                block=self.block,
+                inplanes=self.inplanes,
+                planes=planes,
+                num_blocks=num_blocks,
+                stride=stride,
+                dilation=dilation,
+                style=self.style,
+                avg_down=self.avg_down,
+                with_cp=with_cp,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                dcn=dcn,
+                plugins=stage_plugins)
+            self.inplanes = planes * self.block.expansion
+            layer_name = f'layer{i + 1}'
+            self.add_module(layer_name, res_layer)
+            self.res_layers.append(layer_name)
+
+        self._freeze_stages()
+
+        self.feat_dim = self.block.expansion * base_channels * 2**(
+            len(self.stage_blocks) - 1)
+
+    def make_stage_plugins(self, plugins, stage_idx):
+        """Make plugins for ResNet ``stage_idx`` th stage.
+
+        Currently we support to insert ``context_block``,
+        ``empirical_attention_block``, ``nonlocal_block`` into the backbone
+        like ResNet/ResNeXt. They could be inserted after conv1/conv2/conv3 of
+        Bottleneck.
+
+        An example of plugins format could be:
+
+        Examples:
+            >>> plugins=[
+            ...     dict(cfg=dict(type='xxx', arg1='xxx'),
+            ...          stages=(False, True, True, True),
+            ...          position='after_conv2'),
+            ...     dict(cfg=dict(type='yyy'),
+            ...          stages=(True, True, True, True),
+            ...          position='after_conv3'),
+            ...     dict(cfg=dict(type='zzz', postfix='1'),
+            ...          stages=(True, True, True, True),
+            ...          position='after_conv3'),
+            ...     dict(cfg=dict(type='zzz', postfix='2'),
+            ...          stages=(True, True, True, True),
+            ...          position='after_conv3')
+            ... ]
+            >>> self = ResNet(depth=18)
+            >>> stage_plugins = self.make_stage_plugins(plugins, 0)
+            >>> assert len(stage_plugins) == 3
+
+        Suppose ``stage_idx=0``, the structure of blocks in the stage would be:
+
+        .. code-block:: none
+
+            conv1-> conv2->conv3->yyy->zzz1->zzz2
+
+        Suppose 'stage_idx=1', the structure of blocks in the stage would be:
+
+        .. code-block:: none
+
+            conv1-> conv2->xxx->conv3->yyy->zzz1->zzz2
+
+        If stages is missing, the plugin would be applied to all stages.
+
+        Args:
+            plugins (list[dict]): List of plugins cfg to build. The postfix is
+                required if multiple same type plugins are inserted.
+            stage_idx (int): Index of stage to build
+
+        Returns:
+            list[dict]: Plugins for current stage
+        """
+        stage_plugins = []
+        for plugin in plugins:
+            plugin = plugin.copy()
+            stages = plugin.pop('stages', None)
+            assert stages is None or len(stages) == self.num_stages
+            # whether to insert plugin into current stage
+            if stages is None or stages[stage_idx]:
+                stage_plugins.append(plugin)
+
+        return stage_plugins
+
+    def make_res_layer(self, **kwargs):
+        """Pack all blocks in a stage into a ``ResLayer``."""
+        return ResLayer(**kwargs)
+
+    @property
+    def norm1(self):
+        """nn.Module: the normalization layer named "norm1" """
+        return getattr(self, self.norm1_name)
+
+    def _make_stem_layer(self, in_channels, stem_channels):
+        if self.deep_stem:
+            self.stem = nn.Sequential(
+                build_conv_layer(
+                    self.conv_cfg,
+                    in_channels,
+                    stem_channels // 2,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    bias=False),
+                build_norm_layer(self.norm_cfg, stem_channels // 2)[1],
+                nn.ReLU(inplace=True),
+                build_conv_layer(
+                    self.conv_cfg,
+                    stem_channels // 2,
+                    stem_channels // 2,
+                    kernel_size=3,
+                    stride=1,
+                    padding=1,
+                    bias=False),
+                build_norm_layer(self.norm_cfg, stem_channels // 2)[1],
+                nn.ReLU(inplace=True),
+                build_conv_layer(
+                    self.conv_cfg,
+                    stem_channels // 2,
+                    stem_channels,
+                    kernel_size=3,
+                    stride=1,
+                    padding=1,
+                    bias=False),
+                build_norm_layer(self.norm_cfg, stem_channels)[1],
+                nn.ReLU(inplace=True))
+        else:
+            self.conv1 = build_conv_layer(
+                self.conv_cfg,
+                in_channels,
+                stem_channels,
+                kernel_size=7,
+                stride=2,
+                padding=3,
+                bias=False)
+            self.norm1_name, norm1 = build_norm_layer(
+                self.norm_cfg, stem_channels, postfix=1)
+            self.add_module(self.norm1_name, norm1)
+            self.relu = nn.ReLU(inplace=True)
+        if self.no_pool33:
+            assert self.deep_stem
+            self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
+        else:
+            self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            if self.deep_stem:
+                self.stem.eval()
+                for param in self.stem.parameters():
+                    param.requires_grad = False
+            else:
+                self.norm1.eval()
+                for m in [self.conv1, self.norm1]:
+                    for param in m.parameters():
+                        param.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = getattr(self, f'layer{i}')
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in backbone.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
+                    constant_init(m, 1)
+
+            if self.dcn is not None:
+                for m in self.modules():
+                    if isinstance(m, Bottleneck) and hasattr(
+                            m.conv2, 'conv_offset'):
+                        constant_init(m.conv2.conv_offset, 0)
+
+            if self.zero_init_residual:
+                for m in self.modules():
+                    if isinstance(m, Bottleneck):
+                        constant_init(m.norm3, 0)
+                    elif isinstance(m, BasicBlock):
+                        constant_init(m.norm2, 0)
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+    def forward(self, x):
+        """Forward function."""
+        if self.deep_stem:
+            x = self.stem(x)
+        else:
+            x = self.conv1(x)
+            x = self.norm1(x)
+            x = self.relu(x)
+        x = self.maxpool(x)
+        outs = []
+        for i, layer_name in enumerate(self.res_layers):
+            res_layer = getattr(self, layer_name)
+            x = res_layer(x)
+            if i in self.out_indices:
+                outs.append(x)
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep normalization layer
+        freezed."""
+        super(ResNet, self).train(mode)
+        self._freeze_stages()
+        if mode and self.norm_eval:
+            for m in self.modules():
+                # trick: eval have effect on BatchNorm only
+                if isinstance(m, _BatchNorm):
+                    m.eval()
+
+
+@BACKBONES.register_module()
+class ResNetV1d(ResNet):
+    r"""ResNetV1d variant described in `Bag of Tricks
+    <https://arxiv.org/pdf/1812.01187.pdf>`_.
+
+    Compared with default ResNet(ResNetV1b), ResNetV1d replaces the 7x7 conv in
+    the input stem with three 3x3 convs. And in the downsampling block, a 2x2
+    avg_pool with stride 2 is added before conv, whose stride is changed to 1.
+    """
+
+    def __init__(self, **kwargs):
+        super(ResNetV1d, self).__init__(
+            deep_stem=True, avg_down=True, **kwargs)
+
+@BACKBONES.register_module()
+class ResNetV1e(ResNet):
+    r"""ResNetV1d variant described in `Bag of Tricks
+    <https://arxiv.org/pdf/1812.01187.pdf>`_.
+
+    Compared with default ResNet(ResNetV1b), ResNetV1d replaces the 7x7 conv in
+    the input stem with three 3x3 convs. And in the downsampling block, a 2x2
+    avg_pool with stride 2 is added before conv, whose stride is changed to 1.
+    """
+
+    def __init__(self, **kwargs):
+        super(ResNetV1e, self).__init__(
+            deep_stem=True, avg_down=True, no_pool33=True, **kwargs)
diff --git a/detection/scrfd/mmdet/models/backbones/resnext.py b/detection/scrfd/mmdet/models/backbones/resnext.py
new file mode 100755
index 000000000..bf0360ea7
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/resnext.py
@@ -0,0 +1,132 @@
+import math
+
+from mmcv.cnn import build_conv_layer, build_norm_layer
+
+from ..builder import BACKBONES
+from ..utils import ResLayer
+from .resnet import Bottleneck as _Bottleneck
+from .resnet import ResNet
+
+
+class Bottleneck(_Bottleneck):
+    expansion = 4
+
+    def __init__(self,
+                 inplanes,
+                 planes,
+                 groups=1,
+                 base_width=4,
+                 base_channels=64,
+                 **kwargs):
+        """Bottleneck block for ResNeXt.
+
+        If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
+        it is "caffe", the stride-two layer is the first 1x1 conv layer.
+        """
+        super(Bottleneck, self).__init__(inplanes, planes, **kwargs)
+
+        if groups == 1:
+            width = self.planes
+        else:
+            width = math.floor(self.planes *
+                               (base_width / base_channels)) * groups
+
+        self.norm1_name, norm1 = build_norm_layer(
+            self.norm_cfg, width, postfix=1)
+        self.norm2_name, norm2 = build_norm_layer(
+            self.norm_cfg, width, postfix=2)
+        self.norm3_name, norm3 = build_norm_layer(
+            self.norm_cfg, self.planes * self.expansion, postfix=3)
+
+        self.conv1 = build_conv_layer(
+            self.conv_cfg,
+            self.inplanes,
+            width,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.add_module(self.norm1_name, norm1)
+        fallback_on_stride = False
+        self.with_modulated_dcn = False
+        if self.with_dcn:
+            fallback_on_stride = self.dcn.pop('fallback_on_stride', False)
+        if not self.with_dcn or fallback_on_stride:
+            self.conv2 = build_conv_layer(
+                self.conv_cfg,
+                width,
+                width,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                groups=groups,
+                bias=False)
+        else:
+            assert self.conv_cfg is None, 'conv_cfg must be None for DCN'
+            self.conv2 = build_conv_layer(
+                self.dcn,
+                width,
+                width,
+                kernel_size=3,
+                stride=self.conv2_stride,
+                padding=self.dilation,
+                dilation=self.dilation,
+                groups=groups,
+                bias=False)
+
+        self.add_module(self.norm2_name, norm2)
+        self.conv3 = build_conv_layer(
+            self.conv_cfg,
+            width,
+            self.planes * self.expansion,
+            kernel_size=1,
+            bias=False)
+        self.add_module(self.norm3_name, norm3)
+
+
+@BACKBONES.register_module()
+class ResNeXt(ResNet):
+    """ResNeXt backbone.
+
+    Args:
+        depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
+        in_channels (int): Number of input image channels. Default: 3.
+        num_stages (int): Resnet stages. Default: 4.
+        groups (int): Group of resnext.
+        base_width (int): Base width of resnext.
+        strides (Sequence[int]): Strides of the first block of each stage.
+        dilations (Sequence[int]): Dilation of each stage.
+        out_indices (Sequence[int]): Output from which stages.
+        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
+            layer is the 3x3 conv layer, otherwise the stride-two layer is
+            the first 1x1 conv layer.
+        frozen_stages (int): Stages to be frozen (all param fixed). -1 means
+            not freezing any parameters.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        zero_init_residual (bool): whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
+    """
+
+    arch_settings = {
+        50: (Bottleneck, (3, 4, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3))
+    }
+
+    def __init__(self, groups=1, base_width=4, **kwargs):
+        self.groups = groups
+        self.base_width = base_width
+        super(ResNeXt, self).__init__(**kwargs)
+
+    def make_res_layer(self, **kwargs):
+        """Pack all blocks in a stage into a ``ResLayer``"""
+        return ResLayer(
+            groups=self.groups,
+            base_width=self.base_width,
+            base_channels=self.base_channels,
+            **kwargs)
diff --git a/detection/scrfd/mmdet/models/backbones/ssd_vgg.py b/detection/scrfd/mmdet/models/backbones/ssd_vgg.py
new file mode 100755
index 000000000..cbc4fbb23
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/ssd_vgg.py
@@ -0,0 +1,169 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import VGG, constant_init, kaiming_init, normal_init, xavier_init
+from mmcv.runner import load_checkpoint
+
+from mmdet.utils import get_root_logger
+from ..builder import BACKBONES
+
+
+@BACKBONES.register_module()
+class SSDVGG(VGG):
+    """VGG Backbone network for single-shot-detection.
+
+    Args:
+        input_size (int): width and height of input, from {300, 512}.
+        depth (int): Depth of vgg, from {11, 13, 16, 19}.
+        out_indices (Sequence[int]): Output from which stages.
+
+    Example:
+        >>> self = SSDVGG(input_size=300, depth=11)
+        >>> self.eval()
+        >>> inputs = torch.rand(1, 3, 300, 300)
+        >>> level_outputs = self.forward(inputs)
+        >>> for level_out in level_outputs:
+        ...     print(tuple(level_out.shape))
+        (1, 1024, 19, 19)
+        (1, 512, 10, 10)
+        (1, 256, 5, 5)
+        (1, 256, 3, 3)
+        (1, 256, 1, 1)
+    """
+    extra_setting = {
+        300: (256, 'S', 512, 128, 'S', 256, 128, 256, 128, 256),
+        512: (256, 'S', 512, 128, 'S', 256, 128, 'S', 256, 128, 'S', 256, 128),
+    }
+
+    def __init__(self,
+                 input_size,
+                 depth,
+                 with_last_pool=False,
+                 ceil_mode=True,
+                 out_indices=(3, 4),
+                 out_feature_indices=(22, 34),
+                 l2_norm_scale=20.):
+        # TODO: in_channels for mmcv.VGG
+        super(SSDVGG, self).__init__(
+            depth,
+            with_last_pool=with_last_pool,
+            ceil_mode=ceil_mode,
+            out_indices=out_indices)
+        assert input_size in (300, 512)
+        self.input_size = input_size
+
+        self.features.add_module(
+            str(len(self.features)),
+            nn.MaxPool2d(kernel_size=3, stride=1, padding=1))
+        self.features.add_module(
+            str(len(self.features)),
+            nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6))
+        self.features.add_module(
+            str(len(self.features)), nn.ReLU(inplace=True))
+        self.features.add_module(
+            str(len(self.features)), nn.Conv2d(1024, 1024, kernel_size=1))
+        self.features.add_module(
+            str(len(self.features)), nn.ReLU(inplace=True))
+        self.out_feature_indices = out_feature_indices
+
+        self.inplanes = 1024
+        self.extra = self._make_extra_layers(self.extra_setting[input_size])
+        self.l2_norm = L2Norm(
+            self.features[out_feature_indices[0] - 1].out_channels,
+            l2_norm_scale)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in backbone.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.features.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, nn.BatchNorm2d):
+                    constant_init(m, 1)
+                elif isinstance(m, nn.Linear):
+                    normal_init(m, std=0.01)
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+        for m in self.extra.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+        constant_init(self.l2_norm, self.l2_norm.scale)
+
+    def forward(self, x):
+        """Forward function."""
+        outs = []
+        for i, layer in enumerate(self.features):
+            x = layer(x)
+            if i in self.out_feature_indices:
+                outs.append(x)
+        for i, layer in enumerate(self.extra):
+            x = F.relu(layer(x), inplace=True)
+            if i % 2 == 1:
+                outs.append(x)
+        outs[0] = self.l2_norm(outs[0])
+        if len(outs) == 1:
+            return outs[0]
+        else:
+            return tuple(outs)
+
+    def _make_extra_layers(self, outplanes):
+        layers = []
+        kernel_sizes = (1, 3)
+        num_layers = 0
+        outplane = None
+        for i in range(len(outplanes)):
+            if self.inplanes == 'S':
+                self.inplanes = outplane
+                continue
+            k = kernel_sizes[num_layers % 2]
+            if outplanes[i] == 'S':
+                outplane = outplanes[i + 1]
+                conv = nn.Conv2d(
+                    self.inplanes, outplane, k, stride=2, padding=1)
+            else:
+                outplane = outplanes[i]
+                conv = nn.Conv2d(
+                    self.inplanes, outplane, k, stride=1, padding=0)
+            layers.append(conv)
+            self.inplanes = outplanes[i]
+            num_layers += 1
+        if self.input_size == 512:
+            layers.append(nn.Conv2d(self.inplanes, 256, 4, padding=1))
+
+        return nn.Sequential(*layers)
+
+
+class L2Norm(nn.Module):
+
+    def __init__(self, n_dims, scale=20., eps=1e-10):
+        """L2 normalization layer.
+
+        Args:
+            n_dims (int): Number of dimensions to be normalized
+            scale (float, optional): Defaults to 20..
+            eps (float, optional): Used to avoid division by zero.
+                Defaults to 1e-10.
+        """
+        super(L2Norm, self).__init__()
+        self.n_dims = n_dims
+        self.weight = nn.Parameter(torch.Tensor(self.n_dims))
+        self.eps = eps
+        self.scale = scale
+
+    def forward(self, x):
+        """Forward function."""
+        # normalization layer convert to FP32 in FP16 training
+        x_float = x.float()
+        norm = x_float.pow(2).sum(1, keepdim=True).sqrt() + self.eps
+        return (self.weight[None, :, None, None].float().expand_as(x_float) *
+                x_float / norm).type_as(x)
diff --git a/detection/scrfd/mmdet/models/backbones/trident_resnet.py b/detection/scrfd/mmdet/models/backbones/trident_resnet.py
new file mode 100755
index 000000000..e6100132b
--- /dev/null
+++ b/detection/scrfd/mmdet/models/backbones/trident_resnet.py
@@ -0,0 +1,292 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as cp
+from mmcv.cnn import build_conv_layer, build_norm_layer, kaiming_init
+from torch.nn.modules.utils import _pair
+
+from mmdet.models.backbones.resnet import Bottleneck, ResNet
+from mmdet.models.builder import BACKBONES
+
+
+class TridentConv(nn.Module):
+    """Trident Convolution Module.
+
+    Args:
+        in_channels (int): Number of channels in input.
+        out_channels (int): Number of channels in output.
+        kernel_size (int): Size of convolution kernel.
+        stride (int, optional): Convolution stride. Default: 1.
+        trident_dilations (tuple[int, int, int], optional): Dilations of
+            different trident branch. Default: (1, 2, 3).
+        test_branch_idx (int, optional): In inference, all 3 branches will
+            be used if `test_branch_idx==-1`, otherwise only branch with
+            index `test_branch_idx` will be used. Default: 1.
+        bias (bool, optional): Whether to use bias in convolution or not.
+            Default: False.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 trident_dilations=(1, 2, 3),
+                 test_branch_idx=1,
+                 bias=False):
+        super(TridentConv, self).__init__()
+        self.num_branch = len(trident_dilations)
+        self.with_bias = bias
+        self.test_branch_idx = test_branch_idx
+        self.stride = _pair(stride)
+        self.kernel_size = _pair(kernel_size)
+        self.paddings = _pair(trident_dilations)
+        self.dilations = trident_dilations
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.bias = bias
+
+        self.weight = nn.Parameter(
+            torch.Tensor(out_channels, in_channels, *self.kernel_size))
+        if bias:
+            self.bias = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.bias = None
+        self.init_weights()
+
+    def init_weights(self):
+        kaiming_init(self, distribution='uniform', mode='fan_in')
+
+    def extra_repr(self):
+        tmpstr = f'in_channels={self.in_channels}'
+        tmpstr += f', out_channels={self.out_channels}'
+        tmpstr += f', kernel_size={self.kernel_size}'
+        tmpstr += f', num_branch={self.num_branch}'
+        tmpstr += f', test_branch_idx={self.test_branch_idx}'
+        tmpstr += f', stride={self.stride}'
+        tmpstr += f', paddings={self.paddings}'
+        tmpstr += f', dilations={self.dilations}'
+        tmpstr += f', bias={self.bias}'
+        return tmpstr
+
+    def forward(self, inputs):
+        if self.training or self.test_branch_idx == -1:
+            outputs = [
+                F.conv2d(input, self.weight, self.bias, self.stride, padding,
+                         dilation) for input, dilation, padding in zip(
+                             inputs, self.dilations, self.paddings)
+            ]
+        else:
+            assert len(inputs) == 1
+            outputs = [
+                F.conv2d(inputs[0], self.weight, self.bias, self.stride,
+                         self.paddings[self.test_branch_idx],
+                         self.dilations[self.test_branch_idx])
+            ]
+
+        return outputs
+
+
+# Since TridentNet is defined over ResNet50 and ResNet101, here we
+# only support TridentBottleneckBlock.
+class TridentBottleneck(Bottleneck):
+    """BottleBlock for TridentResNet.
+
+    Args:
+        trident_dilations (tuple[int, int, int]): Dilations of different
+            trident branch.
+        test_branch_idx (int): In inference, all 3 branches will be used
+            if `test_branch_idx==-1`, otherwise only branch with index
+            `test_branch_idx` will be used.
+        concat_output (bool): Whether to concat the output list to a Tensor.
+            `True` only in the last Block.
+    """
+
+    def __init__(self, trident_dilations, test_branch_idx, concat_output,
+                 **kwargs):
+
+        super(TridentBottleneck, self).__init__(**kwargs)
+        self.trident_dilations = trident_dilations
+        self.num_branch = len(trident_dilations)
+        self.concat_output = concat_output
+        self.test_branch_idx = test_branch_idx
+        self.conv2 = TridentConv(
+            self.planes,
+            self.planes,
+            kernel_size=3,
+            stride=self.conv2_stride,
+            bias=False,
+            trident_dilations=self.trident_dilations,
+            test_branch_idx=test_branch_idx)
+
+    def forward(self, x):
+
+        def _inner_forward(x):
+            num_branch = (
+                self.num_branch
+                if self.training or self.test_branch_idx == -1 else 1)
+            identity = x
+            if not isinstance(x, list):
+                x = (x, ) * num_branch
+                identity = x
+                if self.downsample is not None:
+                    identity = [self.downsample(b) for b in x]
+
+            out = [self.conv1(b) for b in x]
+            out = [self.norm1(b) for b in out]
+            out = [self.relu(b) for b in out]
+
+            if self.with_plugins:
+                for k in range(len(out)):
+                    out[k] = self.forward_plugin(out[k],
+                                                 self.after_conv1_plugin_names)
+
+            out = self.conv2(out)
+            out = [self.norm2(b) for b in out]
+            out = [self.relu(b) for b in out]
+            if self.with_plugins:
+                for k in range(len(out)):
+                    out[k] = self.forward_plugin(out[k],
+                                                 self.after_conv2_plugin_names)
+
+            out = [self.conv3(b) for b in out]
+            out = [self.norm3(b) for b in out]
+
+            if self.with_plugins:
+                for k in range(len(out)):
+                    out[k] = self.forward_plugin(out[k],
+                                                 self.after_conv3_plugin_names)
+
+            out = [
+                out_b + identity_b for out_b, identity_b in zip(out, identity)
+            ]
+            return out
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
+        else:
+            out = _inner_forward(x)
+
+        out = [self.relu(b) for b in out]
+        if self.concat_output:
+            out = torch.cat(out, dim=0)
+        return out
+
+
+def make_trident_res_layer(block,
+                           inplanes,
+                           planes,
+                           num_blocks,
+                           stride=1,
+                           trident_dilations=(1, 2, 3),
+                           style='pytorch',
+                           with_cp=False,
+                           conv_cfg=None,
+                           norm_cfg=dict(type='BN'),
+                           dcn=None,
+                           plugins=None,
+                           test_branch_idx=-1):
+    """Build Trident Res Layers."""
+
+    downsample = None
+    if stride != 1 or inplanes != planes * block.expansion:
+        downsample = []
+        conv_stride = stride
+        downsample.extend([
+            build_conv_layer(
+                conv_cfg,
+                inplanes,
+                planes * block.expansion,
+                kernel_size=1,
+                stride=conv_stride,
+                bias=False),
+            build_norm_layer(norm_cfg, planes * block.expansion)[1]
+        ])
+        downsample = nn.Sequential(*downsample)
+
+    layers = []
+    for i in range(num_blocks):
+        layers.append(
+            block(
+                inplanes=inplanes,
+                planes=planes,
+                stride=stride if i == 0 else 1,
+                trident_dilations=trident_dilations,
+                downsample=downsample if i == 0 else None,
+                style=style,
+                with_cp=with_cp,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                dcn=dcn,
+                plugins=plugins,
+                test_branch_idx=test_branch_idx,
+                concat_output=True if i == num_blocks - 1 else False))
+        inplanes = planes * block.expansion
+    return nn.Sequential(*layers)
+
+
+@BACKBONES.register_module()
+class TridentResNet(ResNet):
+    """The stem layer, stage 1 and stage 2 in Trident ResNet are identical to
+    ResNet, while in stage 3, Trident BottleBlock is utilized to replace the
+    normal BottleBlock to yield trident output. Different branch shares the
+    convolution weight but uses different dilations to achieve multi-scale
+    output.
+
+                               / stage3(b0) \
+    x - stem - stage1 - stage2 - stage3(b1) - output
+                               \ stage3(b2) /
+
+    Args:
+        depth (int): Depth of resnet, from {50, 101, 152}.
+        num_branch (int): Number of branches in TridentNet.
+        test_branch_idx (int): In inference, all 3 branches will be used
+            if `test_branch_idx==-1`, otherwise only branch with index
+            `test_branch_idx` will be used.
+        trident_dilations (tuple[int]): Dilations of different trident branch.
+            len(trident_dilations) should be equal to num_branch.
+    """  # noqa
+
+    def __init__(self, depth, num_branch, test_branch_idx, trident_dilations,
+                 **kwargs):
+
+        assert num_branch == len(trident_dilations)
+        assert depth in (50, 101, 152)
+        super(TridentResNet, self).__init__(depth, **kwargs)
+        assert self.num_stages == 3
+        self.test_branch_idx = test_branch_idx
+        self.num_branch = num_branch
+
+        last_stage_idx = self.num_stages - 1
+        stride = self.strides[last_stage_idx]
+        dilation = trident_dilations
+        dcn = self.dcn if self.stage_with_dcn[last_stage_idx] else None
+        if self.plugins is not None:
+            stage_plugins = self.make_stage_plugins(self.plugins,
+                                                    last_stage_idx)
+        else:
+            stage_plugins = None
+        planes = self.base_channels * 2**last_stage_idx
+        res_layer = make_trident_res_layer(
+            TridentBottleneck,
+            inplanes=(self.block.expansion * self.base_channels *
+                      2**(last_stage_idx - 1)),
+            planes=planes,
+            num_blocks=self.stage_blocks[last_stage_idx],
+            stride=stride,
+            trident_dilations=dilation,
+            style=self.style,
+            with_cp=self.with_cp,
+            conv_cfg=self.conv_cfg,
+            norm_cfg=self.norm_cfg,
+            dcn=dcn,
+            plugins=stage_plugins,
+            test_branch_idx=self.test_branch_idx)
+
+        layer_name = f'layer{last_stage_idx + 1}'
+
+        self.__setattr__(layer_name, res_layer)
+        self.res_layers.pop(last_stage_idx)
+        self.res_layers.insert(last_stage_idx, layer_name)
+
+        self._freeze_stages()
diff --git a/detection/scrfd/mmdet/models/builder.py b/detection/scrfd/mmdet/models/builder.py
new file mode 100755
index 000000000..899e78744
--- /dev/null
+++ b/detection/scrfd/mmdet/models/builder.py
@@ -0,0 +1,67 @@
+from mmcv.utils import Registry, build_from_cfg
+from torch import nn
+
+BACKBONES = Registry('backbone')
+NECKS = Registry('neck')
+ROI_EXTRACTORS = Registry('roi_extractor')
+SHARED_HEADS = Registry('shared_head')
+HEADS = Registry('head')
+LOSSES = Registry('loss')
+DETECTORS = Registry('detector')
+
+
+def build(cfg, registry, default_args=None):
+    """Build a module.
+
+    Args:
+        cfg (dict, list[dict]): The config of modules, is is either a dict
+            or a list of configs.
+        registry (:obj:`Registry`): A registry the module belongs to.
+        default_args (dict, optional): Default arguments to build the module.
+            Defaults to None.
+
+    Returns:
+        nn.Module: A built nn module.
+    """
+    if isinstance(cfg, list):
+        modules = [
+            build_from_cfg(cfg_, registry, default_args) for cfg_ in cfg
+        ]
+        return nn.Sequential(*modules)
+    else:
+        return build_from_cfg(cfg, registry, default_args)
+
+
+def build_backbone(cfg):
+    """Build backbone."""
+    return build(cfg, BACKBONES)
+
+
+def build_neck(cfg):
+    """Build neck."""
+    return build(cfg, NECKS)
+
+
+def build_roi_extractor(cfg):
+    """Build roi extractor."""
+    return build(cfg, ROI_EXTRACTORS)
+
+
+def build_shared_head(cfg):
+    """Build shared head."""
+    return build(cfg, SHARED_HEADS)
+
+
+def build_head(cfg):
+    """Build head."""
+    return build(cfg, HEADS)
+
+
+def build_loss(cfg):
+    """Build loss."""
+    return build(cfg, LOSSES)
+
+
+def build_detector(cfg, train_cfg=None, test_cfg=None):
+    """Build detector."""
+    return build(cfg, DETECTORS, dict(train_cfg=train_cfg, test_cfg=test_cfg))
diff --git a/detection/scrfd/mmdet/models/dense_heads/__init__.py b/detection/scrfd/mmdet/models/dense_heads/__init__.py
new file mode 100755
index 000000000..a762e1fd2
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/__init__.py
@@ -0,0 +1,39 @@
+from .anchor_free_head import AnchorFreeHead
+from .anchor_head import AnchorHead
+from .atss_head import ATSSHead
+from .centripetal_head import CentripetalHead
+from .corner_head import CornerHead
+from .fcos_head import FCOSHead
+from .fovea_head import FoveaHead
+from .free_anchor_retina_head import FreeAnchorRetinaHead
+from .fsaf_head import FSAFHead
+from .ga_retina_head import GARetinaHead
+from .ga_rpn_head import GARPNHead
+from .gfl_head import GFLHead
+from .guided_anchor_head import FeatureAdaption, GuidedAnchorHead
+from .nasfcos_head import NASFCOSHead
+from .paa_head import PAAHead
+from .pisa_retinanet_head import PISARetinaHead
+from .pisa_ssd_head import PISASSDHead
+from .reppoints_head import RepPointsHead
+from .retina_head import RetinaHead
+from .retina_sepbn_head import RetinaSepBNHead
+from .rpn_head import RPNHead
+from .sabl_retina_head import SABLRetinaHead
+from .ssd_head import SSDHead
+from .transformer_head import TransformerHead
+from .vfnet_head import VFNetHead
+from .yolact_head import YOLACTHead, YOLACTProtonet, YOLACTSegmHead
+from .yolo_head import YOLOV3Head
+from .scrfd_head import SCRFDHead
+
+__all__ = [
+    'AnchorFreeHead', 'AnchorHead', 'GuidedAnchorHead', 'FeatureAdaption',
+    'RPNHead', 'GARPNHead', 'RetinaHead', 'RetinaSepBNHead', 'GARetinaHead',
+    'SSDHead', 'FCOSHead', 'RepPointsHead', 'FoveaHead',
+    'FreeAnchorRetinaHead', 'ATSSHead', 'FSAFHead', 'NASFCOSHead',
+    'PISARetinaHead', 'PISASSDHead', 'GFLHead', 'CornerHead', 'YOLACTHead',
+    'YOLACTSegmHead', 'YOLACTProtonet', 'YOLOV3Head', 'PAAHead',
+    'SABLRetinaHead', 'CentripetalHead', 'VFNetHead', 'TransformerHead',
+    'SCRFDHead'
+]
diff --git a/detection/scrfd/mmdet/models/dense_heads/anchor_free_head.py b/detection/scrfd/mmdet/models/dense_heads/anchor_free_head.py
new file mode 100755
index 000000000..917acde63
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/anchor_free_head.py
@@ -0,0 +1,340 @@
+from abc import abstractmethod
+
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import multi_apply
+from ..builder import HEADS, build_loss
+from .base_dense_head import BaseDenseHead
+from .dense_test_mixins import BBoxTestMixin
+
+
+@HEADS.register_module()
+class AnchorFreeHead(BaseDenseHead, BBoxTestMixin):
+    """Anchor-free head (FCOS, Fovea, RepPoints, etc.).
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        feat_channels (int): Number of hidden channels. Used in child classes.
+        stacked_convs (int): Number of stacking convs of the head.
+        strides (tuple): Downsample factor of each feature map.
+        dcn_on_last_conv (bool): If true, use dcn in the last layer of
+            towers. Default: False.
+        conv_bias (bool | str): If specified as `auto`, it will be decided by
+            the norm_cfg. Bias of conv will be set as True if `norm_cfg` is
+            None, otherwise False. Default: "auto".
+        loss_cls (dict): Config of classification loss.
+        loss_bbox (dict): Config of localization loss.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Config dict for normalization layer. Default: None.
+        train_cfg (dict): Training config of anchor head.
+        test_cfg (dict): Testing config of anchor head.
+    """  # noqa: W605
+
+    _version = 1
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 feat_channels=256,
+                 stacked_convs=4,
+                 strides=(4, 8, 16, 32, 64),
+                 dcn_on_last_conv=False,
+                 conv_bias='auto',
+                 loss_cls=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 loss_bbox=dict(type='IoULoss', loss_weight=1.0),
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 train_cfg=None,
+                 test_cfg=None):
+        super(AnchorFreeHead, self).__init__()
+        self.num_classes = num_classes
+        self.cls_out_channels = num_classes
+        self.in_channels = in_channels
+        self.feat_channels = feat_channels
+        self.stacked_convs = stacked_convs
+        self.strides = strides
+        self.dcn_on_last_conv = dcn_on_last_conv
+        assert conv_bias == 'auto' or isinstance(conv_bias, bool)
+        self.conv_bias = conv_bias
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox = build_loss(loss_bbox)
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.fp16_enabled = False
+
+        self._init_layers()
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self._init_cls_convs()
+        self._init_reg_convs()
+        self._init_predictor()
+
+    def _init_cls_convs(self):
+        """Initialize classification conv layers of the head."""
+        self.cls_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            if self.dcn_on_last_conv and i == self.stacked_convs - 1:
+                conv_cfg = dict(type='DCNv2')
+            else:
+                conv_cfg = self.conv_cfg
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=self.norm_cfg,
+                    bias=self.conv_bias))
+
+    def _init_reg_convs(self):
+        """Initialize bbox regression conv layers of the head."""
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            if self.dcn_on_last_conv and i == self.stacked_convs - 1:
+                conv_cfg = dict(type='DCNv2')
+            else:
+                conv_cfg = self.conv_cfg
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=self.norm_cfg,
+                    bias=self.conv_bias))
+
+    def _init_predictor(self):
+        """Initialize predictor layers of the head."""
+        self.conv_cls = nn.Conv2d(
+            self.feat_channels, self.cls_out_channels, 3, padding=1)
+        self.conv_reg = nn.Conv2d(self.feat_channels, 4, 3, padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            if isinstance(m.conv, nn.Conv2d):
+                normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            if isinstance(m.conv, nn.Conv2d):
+                normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.conv_cls, std=0.01, bias=bias_cls)
+        normal_init(self.conv_reg, std=0.01)
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        """Hack some keys of the model state dict so that can load checkpoints
+        of previous version."""
+        version = local_metadata.get('version', None)
+        if version is None:
+            # the key is different in early versions
+            # for example, 'fcos_cls' become 'conv_cls' now
+            bbox_head_keys = [
+                k for k in state_dict.keys() if k.startswith(prefix)
+            ]
+            ori_predictor_keys = []
+            new_predictor_keys = []
+            # e.g. 'fcos_cls' or 'fcos_reg'
+            for key in bbox_head_keys:
+                ori_predictor_keys.append(key)
+                key = key.split('.')
+                conv_name = None
+                if key[1].endswith('cls'):
+                    conv_name = 'conv_cls'
+                elif key[1].endswith('reg'):
+                    conv_name = 'conv_reg'
+                elif key[1].endswith('centerness'):
+                    conv_name = 'conv_centerness'
+                else:
+                    assert NotImplementedError
+                if conv_name is not None:
+                    key[1] = conv_name
+                    new_predictor_keys.append('.'.join(key))
+                else:
+                    ori_predictor_keys.pop(-1)
+            for i in range(len(new_predictor_keys)):
+                state_dict[new_predictor_keys[i]] = state_dict.pop(
+                    ori_predictor_keys[i])
+        super()._load_from_state_dict(state_dict, prefix, local_metadata,
+                                      strict, missing_keys, unexpected_keys,
+                                      error_msgs)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually contain classification scores and bbox predictions.
+                cls_scores (list[Tensor]): Box scores for each scale level,
+                    each is a 4D-tensor, the channel number is
+                    num_points * num_classes.
+                bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                    level, each is a 4D-tensor, the channel number is
+                    num_points * 4.
+        """
+        return multi_apply(self.forward_single, feats)[:2]
+
+    def forward_single(self, x):
+        """Forward features of a single scale levle.
+
+        Args:
+            x (Tensor): FPN feature maps of the specified stride.
+
+        Returns:
+            tuple: Scores for each class, bbox predictions, features
+                after classification and regression conv layers, some
+                models needs these features like FCOS.
+        """
+        cls_feat = x
+        reg_feat = x
+
+        for cls_layer in self.cls_convs:
+            cls_feat = cls_layer(cls_feat)
+        cls_score = self.conv_cls(cls_feat)
+
+        for reg_layer in self.reg_convs:
+            reg_feat = reg_layer(reg_feat)
+        bbox_pred = self.conv_reg(reg_feat)
+        return cls_score, bbox_pred, cls_feat, reg_feat
+
+    @abstractmethod
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute loss of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level,
+                each is a 4D-tensor, the channel number is
+                num_points * num_classes.
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level, each is a 4D-tensor, the channel number is
+                num_points * 4.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+        """
+
+        raise NotImplementedError
+
+    @abstractmethod
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=None):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_points * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_points * 4, H, W)
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used
+            rescale (bool): If True, return boxes in original image space
+        """
+
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_targets(self, points, gt_bboxes_list, gt_labels_list):
+        """Compute regression, classification and centerss targets for points
+        in multiple images.
+
+        Args:
+            points (list[Tensor]): Points of each fpn level, each has shape
+                (num_points, 2).
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image,
+                each has shape (num_gt, 4).
+            gt_labels_list (list[Tensor]): Ground truth labels of each box,
+                each has shape (num_gt,).
+        """
+        raise NotImplementedError
+
+    def _get_points_single(self,
+                           featmap_size,
+                           stride,
+                           dtype,
+                           device,
+                           flatten=False):
+        """Get points of a single scale level."""
+        h, w = featmap_size
+        x_range = torch.arange(w, dtype=dtype, device=device)
+        y_range = torch.arange(h, dtype=dtype, device=device)
+        y, x = torch.meshgrid(y_range, x_range)
+        if flatten:
+            y = y.flatten()
+            x = x.flatten()
+        return y, x
+
+    def get_points(self, featmap_sizes, dtype, device, flatten=False):
+        """Get points according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            dtype (torch.dtype): Type of points.
+            device (torch.device): Device of points.
+
+        Returns:
+            tuple: points of each image.
+        """
+        mlvl_points = []
+        for i in range(len(featmap_sizes)):
+            mlvl_points.append(
+                self._get_points_single(featmap_sizes[i], self.strides[i],
+                                        dtype, device, flatten))
+        return mlvl_points
+
+    def aug_test(self, feats, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            feats (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains features for all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[ndarray]: bbox results of each class
+        """
+        return self.aug_test_bboxes(feats, img_metas, rescale=rescale)
diff --git a/detection/scrfd/mmdet/models/dense_heads/anchor_head.py b/detection/scrfd/mmdet/models/dense_heads/anchor_head.py
new file mode 100755
index 000000000..a5bb41373
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/anchor_head.py
@@ -0,0 +1,682 @@
+import torch
+import torch.nn as nn
+from mmcv.cnn import normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, build_anchor_generator,
+                        build_assigner, build_bbox_coder, build_sampler,
+                        images_to_levels, multi_apply, multiclass_nms, unmap)
+from ..builder import HEADS, build_loss
+from .base_dense_head import BaseDenseHead
+from .dense_test_mixins import BBoxTestMixin
+
+
+@HEADS.register_module()
+class AnchorHead(BaseDenseHead, BBoxTestMixin):
+    """Anchor-based head (RPN, RetinaNet, SSD, etc.).
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        feat_channels (int): Number of hidden channels. Used in child classes.
+        anchor_generator (dict): Config dict for anchor generator
+        bbox_coder (dict): Config of bounding box coder.
+        reg_decoded_bbox (bool): If true, the regression loss would be
+            applied on decoded bounding boxes. Default: False
+        loss_cls (dict): Config of classification loss.
+        loss_bbox (dict): Config of localization loss.
+        train_cfg (dict): Training config of anchor head.
+        test_cfg (dict): Testing config of anchor head.
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 feat_channels=256,
+                 anchor_generator=dict(
+                     type='AnchorGenerator',
+                     scales=[8, 16, 32],
+                     ratios=[0.5, 1.0, 2.0],
+                     strides=[4, 8, 16, 32, 64]),
+                 bbox_coder=dict(
+                     type='DeltaXYWHBBoxCoder',
+                     clip_border=True,
+                     target_means=(.0, .0, .0, .0),
+                     target_stds=(1.0, 1.0, 1.0, 1.0)),
+                 reg_decoded_bbox=False,
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 loss_bbox=dict(
+                     type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0),
+                 train_cfg=None,
+                 test_cfg=None):
+        super(AnchorHead, self).__init__()
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.feat_channels = feat_channels
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        # TODO better way to determine whether sample or not
+        self.sampling = loss_cls['type'] not in [
+            'FocalLoss', 'GHMC', 'QualityFocalLoss'
+        ]
+        if self.use_sigmoid_cls:
+            self.cls_out_channels = num_classes
+        else:
+            self.cls_out_channels = num_classes + 1
+
+        if self.cls_out_channels <= 0:
+            raise ValueError(f'num_classes={num_classes} is too small')
+        self.reg_decoded_bbox = reg_decoded_bbox
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox = build_loss(loss_bbox)
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # use PseudoSampler when sampling is False
+            if self.sampling and hasattr(self.train_cfg, 'sampler'):
+                sampler_cfg = self.train_cfg.sampler
+            else:
+                sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+        self.fp16_enabled = False
+
+        self.anchor_generator = build_anchor_generator(anchor_generator)
+        # usually the numbers of anchors for each level are the same
+        # except SSD detectors
+        self.num_anchors = self.anchor_generator.num_base_anchors[0]
+        self._init_layers()
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.conv_cls = nn.Conv2d(self.in_channels,
+                                  self.num_anchors * self.cls_out_channels, 1)
+        self.conv_reg = nn.Conv2d(self.in_channels, self.num_anchors * 4, 1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        normal_init(self.conv_cls, std=0.01)
+        normal_init(self.conv_reg, std=0.01)
+
+    def forward_single(self, x):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls scores for a single scale level \
+                    the channels number is num_anchors * num_classes.
+                bbox_pred (Tensor): Box energies / deltas for a single scale \
+                    level, the channels number is num_anchors * 4.
+        """
+        cls_score = self.conv_cls(x)
+        bbox_pred = self.conv_reg(x)
+        return cls_score, bbox_pred
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: A tuple of classification scores and bbox prediction.
+
+                - cls_scores (list[Tensor]): Classification scores for all \
+                    scale levels, each is a 4D-tensor, the channels number \
+                    is num_anchors * num_classes.
+                - bbox_preds (list[Tensor]): Box energies / deltas for all \
+                    scale levels, each is a 4D-tensor, the channels number \
+                    is num_anchors * 4.
+        """
+        return multi_apply(self.forward_single, feats)
+
+    def get_anchors(self, featmap_sizes, img_metas, device='cuda'):
+        """Get anchors according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+            device (torch.device | str): Device for returned tensors
+
+        Returns:
+            tuple:
+                anchor_list (list[Tensor]): Anchors of each image.
+                valid_flag_list (list[Tensor]): Valid flags of each image.
+        """
+        num_imgs = len(img_metas)
+
+        # since feature map sizes of all images are the same, we only compute
+        # anchors for one time
+        multi_level_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device)
+        anchor_list = [multi_level_anchors for _ in range(num_imgs)]
+
+        # for each image, we compute valid flags of multi level anchors
+        valid_flag_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_flags = self.anchor_generator.valid_flags(
+                featmap_sizes, img_meta['pad_shape'], device)
+            valid_flag_list.append(multi_level_flags)
+
+        return anchor_list, valid_flag_list
+
+    def _get_targets_single(self,
+                            flat_anchors,
+                            valid_flags,
+                            gt_bboxes,
+                            gt_bboxes_ignore,
+                            gt_labels,
+                            img_meta,
+                            label_channels=1,
+                            unmap_outputs=True):
+        """Compute regression and classification targets for anchors in a
+        single image.
+
+        Args:
+            flat_anchors (Tensor): Multi-level anchors of the image, which are
+                concatenated into a single tensor of shape (num_anchors ,4)
+            valid_flags (Tensor): Multi level valid flags of the image,
+                which are concatenated into a single tensor of
+                    shape (num_anchors,).
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            img_meta (dict): Meta info of the image.
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            img_meta (dict): Meta info of the image.
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple:
+                labels_list (list[Tensor]): Labels of each level
+                label_weights_list (list[Tensor]): Label weights of each level
+                bbox_targets_list (list[Tensor]): BBox targets of each level
+                bbox_weights_list (list[Tensor]): BBox weights of each level
+                num_total_pos (int): Number of positive samples in all images
+                num_total_neg (int): Number of negative samples in all images
+        """
+        inside_flags = anchor_inside_flags(flat_anchors, valid_flags,
+                                           img_meta['img_shape'][:2],
+                                           self.train_cfg.allowed_border)
+        if not inside_flags.any():
+            return (None, ) * 7
+        # assign gt and sample anchors
+        anchors = flat_anchors[inside_flags, :]
+
+        assign_result = self.assigner.assign(
+            anchors, gt_bboxes, gt_bboxes_ignore,
+            None if self.sampling else gt_labels)
+        sampling_result = self.sampler.sample(assign_result, anchors,
+                                              gt_bboxes)
+
+        num_valid_anchors = anchors.shape[0]
+        bbox_targets = torch.zeros_like(anchors)
+        bbox_weights = torch.zeros_like(anchors)
+        labels = anchors.new_full((num_valid_anchors, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = anchors.new_zeros(num_valid_anchors, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            if not self.reg_decoded_bbox:
+                pos_bbox_targets = self.bbox_coder.encode(
+                    sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes)
+            else:
+                pos_bbox_targets = sampling_result.pos_gt_bboxes
+            bbox_targets[pos_inds, :] = pos_bbox_targets
+            bbox_weights[pos_inds, :] = 1.0
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class since v2.5.0
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_anchors.size(0)
+            labels = unmap(
+                labels, num_total_anchors, inside_flags,
+                fill=self.num_classes)  # fill bg label
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_targets = unmap(bbox_targets, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+
+        return (labels, label_weights, bbox_targets, bbox_weights, pos_inds,
+                neg_inds, sampling_result)
+
+    def get_targets(self,
+                    anchor_list,
+                    valid_flag_list,
+                    gt_bboxes_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None,
+                    gt_labels_list=None,
+                    label_channels=1,
+                    unmap_outputs=True,
+                    return_sampling_results=False):
+        """Compute regression and classification targets for anchors in
+        multiple images.
+
+        Args:
+            anchor_list (list[list[Tensor]]): Multi level anchors of each
+                image. The outer list indicates images, and the inner list
+                corresponds to feature levels of the image. Each element of
+                the inner list is a tensor of shape (num_anchors, 4).
+            valid_flag_list (list[list[Tensor]]): Multi level valid flags of
+                each image. The outer list indicates images, and the inner list
+                corresponds to feature levels of the image. Each element of
+                the inner list is a tensor of shape (num_anchors, )
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            img_metas (list[dict]): Meta info of each image.
+            gt_bboxes_ignore_list (list[Tensor]): Ground truth bboxes to be
+                ignored.
+            gt_labels_list (list[Tensor]): Ground truth labels of each box.
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple: Usually returns a tuple containing learning targets.
+
+                - labels_list (list[Tensor]): Labels of each level.
+                - label_weights_list (list[Tensor]): Label weights of each \
+                    level.
+                - bbox_targets_list (list[Tensor]): BBox targets of each level.
+                - bbox_weights_list (list[Tensor]): BBox weights of each level.
+                - num_total_pos (int): Number of positive samples in all \
+                    images.
+                - num_total_neg (int): Number of negative samples in all \
+                    images.
+            additional_returns: This function enables user-defined returns from
+                `self._get_targets_single`. These returns are currently refined
+                to properties at each feature map (i.e. having HxW dimension).
+                The results will be concatenated after the end
+        """
+        num_imgs = len(img_metas)
+        assert len(anchor_list) == len(valid_flag_list) == num_imgs
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        # concat all level anchors to a single tensor
+        concat_anchor_list = []
+        concat_valid_flag_list = []
+        for i in range(num_imgs):
+            assert len(anchor_list[i]) == len(valid_flag_list[i])
+            concat_anchor_list.append(torch.cat(anchor_list[i]))
+            concat_valid_flag_list.append(torch.cat(valid_flag_list[i]))
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        results = multi_apply(
+            self._get_targets_single,
+            concat_anchor_list,
+            concat_valid_flag_list,
+            gt_bboxes_list,
+            gt_bboxes_ignore_list,
+            gt_labels_list,
+            img_metas,
+            label_channels=label_channels,
+            unmap_outputs=unmap_outputs)
+        (all_labels, all_label_weights, all_bbox_targets, all_bbox_weights,
+         pos_inds_list, neg_inds_list, sampling_results_list) = results[:7]
+        rest_results = list(results[7:])  # user-added return values
+        # no valid anchors
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        labels_list = images_to_levels(all_labels, num_level_anchors)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_anchors)
+        bbox_targets_list = images_to_levels(all_bbox_targets,
+                                             num_level_anchors)
+        bbox_weights_list = images_to_levels(all_bbox_weights,
+                                             num_level_anchors)
+        res = (labels_list, label_weights_list, bbox_targets_list,
+               bbox_weights_list, num_total_pos, num_total_neg)
+        if return_sampling_results:
+            res = res + (sampling_results_list, )
+        for i, r in enumerate(rest_results):  # user-added return values
+            rest_results[i] = images_to_levels(r, num_level_anchors)
+
+        return res + tuple(rest_results)
+
+    def loss_single(self, cls_score, bbox_pred, anchors, labels, label_weights,
+                    bbox_targets, bbox_weights, num_total_samples):
+        """Compute loss of a single scale level.
+
+        Args:
+            cls_score (Tensor): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W).
+            bbox_pred (Tensor): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W).
+            anchors (Tensor): Box reference for each scale level with shape
+                (N, num_total_anchors, 4).
+            labels (Tensor): Labels of each anchors with shape
+                (N, num_total_anchors).
+            label_weights (Tensor): Label weights of each anchor with shape
+                (N, num_total_anchors)
+            bbox_targets (Tensor): BBox regression targets of each anchor wight
+                shape (N, num_total_anchors, 4).
+            bbox_weights (Tensor): BBox regression loss weights of each anchor
+                with shape (N, num_total_anchors, 4).
+            num_total_samples (int): If sampling, num total samples equal to
+                the number of total anchors; Otherwise, it is the number of
+                positive anchors.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        # classification loss
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+        cls_score = cls_score.permute(0, 2, 3,
+                                      1).reshape(-1, self.cls_out_channels)
+        loss_cls = self.loss_cls(
+            cls_score, labels, label_weights, avg_factor=num_total_samples)
+        # regression loss
+        bbox_targets = bbox_targets.reshape(-1, 4)
+        bbox_weights = bbox_weights.reshape(-1, 4)
+        bbox_pred = bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)
+        if self.reg_decoded_bbox:
+            anchors = anchors.reshape(-1, 4)
+            bbox_pred = self.bbox_coder.decode(anchors, bbox_pred)
+        loss_bbox = self.loss_bbox(
+            bbox_pred,
+            bbox_targets,
+            bbox_weights,
+            avg_factor=num_total_samples)
+        return loss_cls, loss_bbox
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss. Default: None
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+        num_total_samples = (
+            num_total_pos + num_total_neg if self.sampling else num_total_pos)
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        # concat all level anchors and flags to a single tensor
+        concat_anchor_list = []
+        for i in range(len(anchor_list)):
+            concat_anchor_list.append(torch.cat(anchor_list[i]))
+        all_anchor_list = images_to_levels(concat_anchor_list,
+                                           num_level_anchors)
+
+        losses_cls, losses_bbox = multi_apply(
+            self.loss_single,
+            cls_scores,
+            bbox_preds,
+            all_anchor_list,
+            labels_list,
+            label_weights_list,
+            bbox_targets_list,
+            bbox_weights_list,
+            num_total_samples=num_total_samples)
+        return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class labelof the
+                corresponding box.
+
+        Example:
+            >>> import mmcv
+            >>> self = AnchorHead(
+            >>>     num_classes=9,
+            >>>     in_channels=1,
+            >>>     anchor_generator=dict(
+            >>>         type='AnchorGenerator',
+            >>>         scales=[8],
+            >>>         ratios=[0.5, 1.0, 2.0],
+            >>>         strides=[4,]))
+            >>> img_metas = [{'img_shape': (32, 32, 3), 'scale_factor': 1}]
+            >>> cfg = mmcv.Config(dict(
+            >>>     score_thr=0.00,
+            >>>     nms=dict(type='nms', iou_thr=1.0),
+            >>>     max_per_img=10))
+            >>> feat = torch.rand(1, 1, 3, 3)
+            >>> cls_score, bbox_pred = self.forward_single(feat)
+            >>> # note the input lists are over different levels, not images
+            >>> cls_scores, bbox_preds = [cls_score], [bbox_pred]
+            >>> result_list = self.get_bboxes(cls_scores, bbox_preds,
+            >>>                               img_metas, cfg)
+            >>> det_bboxes, det_labels = result_list[0]
+            >>> assert len(result_list) == 1
+            >>> assert det_bboxes.shape[1] == 5
+            >>> assert len(det_bboxes) == len(det_labels) == cfg.max_per_img
+        """
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+
+        device = cls_scores[0].device
+        featmap_sizes = [cls_scores[i].shape[-2:] for i in range(num_levels)]
+        mlvl_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            if with_nms:
+                # some heads don't support with_nms argument
+                proposals = self._get_bboxes_single(cls_score_list,
+                                                    bbox_pred_list,
+                                                    mlvl_anchors, img_shape,
+                                                    scale_factor, cfg, rescale)
+            else:
+                proposals = self._get_bboxes_single(cls_score_list,
+                                                    bbox_pred_list,
+                                                    mlvl_anchors, img_shape,
+                                                    scale_factor, cfg, rescale,
+                                                    with_nms)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_score_list,
+                           bbox_pred_list,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            cls_score_list (list[Tensor]): Box scores for a single scale level
+                Has shape (num_anchors * num_classes, H, W).
+            bbox_pred_list (list[Tensor]): Box energies / deltas for a single
+                scale level with shape (num_anchors * 4, H, W).
+            mlvl_anchors (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            Tensor: Labeled boxes in shape (n, 5), where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_score_list) == len(bbox_pred_list) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, anchors in zip(cls_score_list,
+                                                 bbox_pred_list, mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            cls_score = cls_score.permute(1, 2,
+                                          0).reshape(-1, self.cls_out_channels)
+            if self.use_sigmoid_cls:
+                scores = cls_score.sigmoid()
+            else:
+                scores = cls_score.softmax(-1)
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                # Get maximum scores for foreground classes.
+                if self.use_sigmoid_cls:
+                    max_scores, _ = scores.max(dim=1)
+                else:
+                    # remind that we set FG labels to [0, num_class-1]
+                    # since mmdet v2.0
+                    # BG cat_id: num_class
+                    max_scores, _ = scores[:, :-1].max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+            bboxes = self.bbox_coder.decode(
+                anchors, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        if self.use_sigmoid_cls:
+            # Add a dummy background class to the backend when using sigmoid
+            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+            # BG cat_id: num_class
+            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores
+
+    def aug_test(self, feats, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            feats (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains features for all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[ndarray]: bbox results of each class
+        """
+        return self.aug_test_bboxes(feats, img_metas, rescale=rescale)
diff --git a/detection/scrfd/mmdet/models/dense_heads/atss_head.py b/detection/scrfd/mmdet/models/dense_heads/atss_head.py
new file mode 100755
index 000000000..764938629
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/atss_head.py
@@ -0,0 +1,650 @@
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, Scale, bias_init_with_prob, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, build_assigner, build_sampler,
+                        images_to_levels, multi_apply, multiclass_nms,
+                        reduce_mean, unmap)
+from ..builder import HEADS, build_loss
+from .anchor_head import AnchorHead
+
+EPS = 1e-12
+
+
+@HEADS.register_module()
+class ATSSHead(AnchorHead):
+    """Bridging the Gap Between Anchor-based and Anchor-free Detection via
+    Adaptive Training Sample Selection.
+
+    ATSS head structure is similar with FCOS, however ATSS use anchor boxes
+    and assign label by Adaptive Training Sample Selection instead max-iou.
+
+    https://arxiv.org/abs/1912.02424
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+                 loss_centerness=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        super(ATSSHead, self).__init__(num_classes, in_channels, **kwargs)
+
+        self.sampling = False
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # SSD sampling=False so use PseudoSampler
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+        self.loss_centerness = build_loss(loss_centerness)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        self.atss_cls = nn.Conv2d(
+            self.feat_channels,
+            self.num_anchors * self.cls_out_channels,
+            3,
+            padding=1)
+        self.atss_reg = nn.Conv2d(
+            self.feat_channels, self.num_anchors * 4, 3, padding=1)
+        self.atss_centerness = nn.Conv2d(
+            self.feat_channels, self.num_anchors * 1, 3, padding=1)
+        self.scales = nn.ModuleList(
+            [Scale(1.0) for _ in self.anchor_generator.strides])
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.atss_cls, std=0.01, bias=bias_cls)
+        normal_init(self.atss_reg, std=0.01)
+        normal_init(self.atss_centerness, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually a tuple of classification scores and bbox prediction
+                cls_scores (list[Tensor]): Classification scores for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * num_classes.
+                bbox_preds (list[Tensor]): Box energies / deltas for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * 4.
+        """
+        return multi_apply(self.forward_single, feats, self.scales)
+
+    def forward_single(self, x, scale):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls scores for a single scale level
+                    the channels number is num_anchors * num_classes.
+                bbox_pred (Tensor): Box energies / deltas for a single scale
+                    level, the channels number is num_anchors * 4.
+                centerness (Tensor): Centerness for a single scale level, the
+                    channel number is (N, num_anchors * 1, H, W).
+        """
+        cls_feat = x
+        reg_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            reg_feat = reg_conv(reg_feat)
+        cls_score = self.atss_cls(cls_feat)
+        # we just follow atss, not apply exp in bbox_pred
+        bbox_pred = scale(self.atss_reg(reg_feat)).float()
+        centerness = self.atss_centerness(reg_feat)
+        return cls_score, bbox_pred, centerness
+
+    def loss_single(self, anchors, cls_score, bbox_pred, centerness, labels,
+                    label_weights, bbox_targets, num_total_samples):
+        """Compute loss of a single scale level.
+
+        Args:
+            cls_score (Tensor): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W).
+            bbox_pred (Tensor): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W).
+            anchors (Tensor): Box reference for each scale level with shape
+                (N, num_total_anchors, 4).
+            labels (Tensor): Labels of each anchors with shape
+                (N, num_total_anchors).
+            label_weights (Tensor): Label weights of each anchor with shape
+                (N, num_total_anchors)
+            bbox_targets (Tensor): BBox regression targets of each anchor wight
+                shape (N, num_total_anchors, 4).
+            num_total_samples (int): Number os positive samples that is
+                reduced over all GPUs.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+
+        anchors = anchors.reshape(-1, 4)
+        cls_score = cls_score.permute(0, 2, 3, 1).reshape(
+            -1, self.cls_out_channels).contiguous()
+        bbox_pred = bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)
+        centerness = centerness.permute(0, 2, 3, 1).reshape(-1)
+        bbox_targets = bbox_targets.reshape(-1, 4)
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+
+        # classification loss
+        loss_cls = self.loss_cls(
+            cls_score, labels, label_weights, avg_factor=num_total_samples)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((labels >= 0)
+                    & (labels < bg_class_ind)).nonzero().squeeze(1)
+
+        if len(pos_inds) > 0:
+            pos_bbox_targets = bbox_targets[pos_inds]
+            pos_bbox_pred = bbox_pred[pos_inds]
+            pos_anchors = anchors[pos_inds]
+            pos_centerness = centerness[pos_inds]
+
+            centerness_targets = self.centerness_target(
+                pos_anchors, pos_bbox_targets)
+            pos_decode_bbox_pred = self.bbox_coder.decode(
+                pos_anchors, pos_bbox_pred)
+            pos_decode_bbox_targets = self.bbox_coder.decode(
+                pos_anchors, pos_bbox_targets)
+
+            # regression loss
+            loss_bbox = self.loss_bbox(
+                pos_decode_bbox_pred,
+                pos_decode_bbox_targets,
+                weight=centerness_targets,
+                avg_factor=1.0)
+
+            # centerness loss
+            loss_centerness = self.loss_centerness(
+                pos_centerness,
+                centerness_targets,
+                avg_factor=num_total_samples)
+
+        else:
+            loss_bbox = bbox_pred.sum() * 0
+            loss_centerness = centerness.sum() * 0
+            centerness_targets = bbox_targets.new_tensor(0.)
+
+        return loss_cls, loss_bbox, loss_centerness, centerness_targets.sum()
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'centernesses'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             centernesses,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            centernesses (list[Tensor]): Centerness for each scale
+                level with shape (N, num_anchors * 1, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+
+        (anchor_list, labels_list, label_weights_list, bbox_targets_list,
+         bbox_weights_list, num_total_pos, num_total_neg) = cls_reg_targets
+
+        num_total_samples = reduce_mean(
+            torch.tensor(num_total_pos).cuda()).item()
+        num_total_samples = max(num_total_samples, 1.0)
+
+        losses_cls, losses_bbox, loss_centerness,\
+            bbox_avg_factor = multi_apply(
+                self.loss_single,
+                anchor_list,
+                cls_scores,
+                bbox_preds,
+                centernesses,
+                labels_list,
+                label_weights_list,
+                bbox_targets_list,
+                num_total_samples=num_total_samples)
+
+        bbox_avg_factor = sum(bbox_avg_factor)
+        bbox_avg_factor = reduce_mean(bbox_avg_factor).item()
+        if bbox_avg_factor < EPS:
+            bbox_avg_factor = 1
+        losses_bbox = list(map(lambda x: x / bbox_avg_factor, losses_bbox))
+        return dict(
+            loss_cls=losses_cls,
+            loss_bbox=losses_bbox,
+            loss_centerness=loss_centerness)
+
+    def centerness_target(self, anchors, bbox_targets):
+        # only calculate pos centerness targets, otherwise there may be nan
+        gts = self.bbox_coder.decode(anchors, bbox_targets)
+        anchors_cx = (anchors[:, 2] + anchors[:, 0]) / 2
+        anchors_cy = (anchors[:, 3] + anchors[:, 1]) / 2
+        l_ = anchors_cx - gts[:, 0]
+        t_ = anchors_cy - gts[:, 1]
+        r_ = gts[:, 2] - anchors_cx
+        b_ = gts[:, 3] - anchors_cy
+
+        left_right = torch.stack([l_, r_], dim=1)
+        top_bottom = torch.stack([t_, b_], dim=1)
+        centerness = torch.sqrt(
+            (left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0]) *
+            (top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0]))
+        assert not torch.isnan(centerness).any()
+        return centerness
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'centernesses'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   centernesses,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                with shape (N, num_anchors * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W).
+            centernesses (list[Tensor]): Centerness for each scale level with
+                shape (N, num_anchors * 1, H, W).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used. Default: None.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class label of the
+                corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+        device = cls_scores[0].device
+        featmap_sizes = [cls_scores[i].shape[-2:] for i in range(num_levels)]
+        mlvl_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            centerness_pred_list = [
+                centernesses[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self._get_bboxes_single(cls_score_list, bbox_pred_list,
+                                                centerness_pred_list,
+                                                mlvl_anchors, img_shape,
+                                                scale_factor, cfg, rescale,
+                                                with_nms)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           centernesses,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into labeled boxes.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for a single scale level
+                with shape (num_anchors * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for a single
+                scale level with shape (num_anchors * 4, H, W).
+            centernesses (list[Tensor]): Centerness for a single scale level
+                with shape (num_anchors * 1, H, W).
+            mlvl_anchors (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arrange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): BBox predictions in shape (n, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (n,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        mlvl_centerness = []
+        for cls_score, bbox_pred, centerness, anchors in zip(
+                cls_scores, bbox_preds, centernesses, mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            centerness = centerness.permute(1, 2, 0).reshape(-1).sigmoid()
+
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                max_scores, _ = (scores * centerness[:, None]).max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+                centerness = centerness[topk_inds]
+
+            bboxes = self.bbox_coder.decode(
+                anchors, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+            mlvl_centerness.append(centerness)
+
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        # Add a dummy background class to the backend when using sigmoid
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        mlvl_centerness = torch.cat(mlvl_centerness)
+
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(
+                mlvl_bboxes,
+                mlvl_scores,
+                cfg.score_thr,
+                cfg.nms,
+                cfg.max_per_img,
+                score_factors=mlvl_centerness)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores, mlvl_centerness
+
+    def get_targets(self,
+                    anchor_list,
+                    valid_flag_list,
+                    gt_bboxes_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None,
+                    gt_labels_list=None,
+                    label_channels=1,
+                    unmap_outputs=True):
+        """Get targets for ATSS head.
+
+        This method is almost the same as `AnchorHead.get_targets()`. Besides
+        returning the targets as the parent method does, it also returns the
+        anchors as the first element of the returned tuple.
+        """
+        num_imgs = len(img_metas)
+        assert len(anchor_list) == len(valid_flag_list) == num_imgs
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        num_level_anchors_list = [num_level_anchors] * num_imgs
+
+        # concat all level anchors and flags to a single tensor
+        for i in range(num_imgs):
+            assert len(anchor_list[i]) == len(valid_flag_list[i])
+            anchor_list[i] = torch.cat(anchor_list[i])
+            valid_flag_list[i] = torch.cat(valid_flag_list[i])
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        (all_anchors, all_labels, all_label_weights, all_bbox_targets,
+         all_bbox_weights, pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single,
+             anchor_list,
+             valid_flag_list,
+             num_level_anchors_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             gt_labels_list,
+             img_metas,
+             label_channels=label_channels,
+             unmap_outputs=unmap_outputs)
+        # no valid anchors
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        anchors_list = images_to_levels(all_anchors, num_level_anchors)
+        labels_list = images_to_levels(all_labels, num_level_anchors)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_anchors)
+        bbox_targets_list = images_to_levels(all_bbox_targets,
+                                             num_level_anchors)
+        bbox_weights_list = images_to_levels(all_bbox_weights,
+                                             num_level_anchors)
+        return (anchors_list, labels_list, label_weights_list,
+                bbox_targets_list, bbox_weights_list, num_total_pos,
+                num_total_neg)
+
+    def _get_target_single(self,
+                           flat_anchors,
+                           valid_flags,
+                           num_level_anchors,
+                           gt_bboxes,
+                           gt_bboxes_ignore,
+                           gt_labels,
+                           img_meta,
+                           label_channels=1,
+                           unmap_outputs=True):
+        """Compute regression, classification targets for anchors in a single
+        image.
+
+        Args:
+            flat_anchors (Tensor): Multi-level anchors of the image, which are
+                concatenated into a single tensor of shape (num_anchors ,4)
+            valid_flags (Tensor): Multi level valid flags of the image,
+                which are concatenated into a single tensor of
+                    shape (num_anchors,).
+            num_level_anchors Tensor): Number of anchors of each scale level.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            img_meta (dict): Meta info of the image.
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple: N is the number of total anchors in the image.
+                labels (Tensor): Labels of all anchors in the image with shape
+                    (N,).
+                label_weights (Tensor): Label weights of all anchor in the
+                    image with shape (N,).
+                bbox_targets (Tensor): BBox targets of all anchors in the
+                    image with shape (N, 4).
+                bbox_weights (Tensor): BBox weights of all anchors in the
+                    image with shape (N, 4)
+                pos_inds (Tensor): Indices of postive anchor with shape
+                    (num_pos,).
+                neg_inds (Tensor): Indices of negative anchor with shape
+                    (num_neg,).
+        """
+        inside_flags = anchor_inside_flags(flat_anchors, valid_flags,
+                                           img_meta['img_shape'][:2],
+                                           self.train_cfg.allowed_border)
+        if not inside_flags.any():
+            return (None, ) * 7
+        # assign gt and sample anchors
+        anchors = flat_anchors[inside_flags, :]
+
+        num_level_anchors_inside = self.get_num_level_anchors_inside(
+            num_level_anchors, inside_flags)
+        assign_result = self.assigner.assign(anchors, num_level_anchors_inside,
+                                             gt_bboxes, gt_bboxes_ignore,
+                                             gt_labels)
+
+        sampling_result = self.sampler.sample(assign_result, anchors,
+                                              gt_bboxes)
+
+        num_valid_anchors = anchors.shape[0]
+        bbox_targets = torch.zeros_like(anchors)
+        bbox_weights = torch.zeros_like(anchors)
+        labels = anchors.new_full((num_valid_anchors, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = anchors.new_zeros(num_valid_anchors, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            if hasattr(self, 'bbox_coder'):
+                pos_bbox_targets = self.bbox_coder.encode(
+                    sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes)
+            else:
+                # used in VFNetHead
+                pos_bbox_targets = sampling_result.pos_gt_bboxes
+            bbox_targets[pos_inds, :] = pos_bbox_targets
+            bbox_weights[pos_inds, :] = 1.0
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class since v2.5.0
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_anchors.size(0)
+            anchors = unmap(anchors, num_total_anchors, inside_flags)
+            labels = unmap(
+                labels, num_total_anchors, inside_flags, fill=self.num_classes)
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_targets = unmap(bbox_targets, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+
+        return (anchors, labels, label_weights, bbox_targets, bbox_weights,
+                pos_inds, neg_inds)
+
+    def get_num_level_anchors_inside(self, num_level_anchors, inside_flags):
+        split_inside_flags = torch.split(inside_flags, num_level_anchors)
+        num_level_anchors_inside = [
+            int(flags.sum()) for flags in split_inside_flags
+        ]
+        return num_level_anchors_inside
diff --git a/detection/scrfd/mmdet/models/dense_heads/base_dense_head.py b/detection/scrfd/mmdet/models/dense_heads/base_dense_head.py
new file mode 100755
index 000000000..de11e4a21
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/base_dense_head.py
@@ -0,0 +1,59 @@
+from abc import ABCMeta, abstractmethod
+
+import torch.nn as nn
+
+
+class BaseDenseHead(nn.Module, metaclass=ABCMeta):
+    """Base class for DenseHeads."""
+
+    def __init__(self):
+        super(BaseDenseHead, self).__init__()
+
+    @abstractmethod
+    def loss(self, **kwargs):
+        """Compute losses of the head."""
+        pass
+
+    @abstractmethod
+    def get_bboxes(self, **kwargs):
+        """Transform network output for a batch into bbox predictions."""
+        pass
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels=None,
+                      gt_bboxes_ignore=None,
+                      proposal_cfg=None,
+                      **kwargs):
+        """
+        Args:
+            x (list[Tensor]): Features from FPN.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            proposal_cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used
+
+        Returns:
+            tuple:
+                losses: (dict[str, Tensor]): A dictionary of loss components.
+                proposal_list (list[Tensor]): Proposals of each image.
+        """
+        outs = self(x)
+        if gt_labels is None:
+            loss_inputs = outs + (gt_bboxes, img_metas)
+        else:
+            loss_inputs = outs + (gt_bboxes, gt_labels, img_metas)
+        losses = self.loss(*loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
+        if proposal_cfg is None:
+            return losses
+        else:
+            proposal_list = self.get_bboxes(*outs, img_metas, cfg=proposal_cfg)
+            return losses, proposal_list
diff --git a/detection/scrfd/mmdet/models/dense_heads/centripetal_head.py b/detection/scrfd/mmdet/models/dense_heads/centripetal_head.py
new file mode 100755
index 000000000..6728218b6
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/centripetal_head.py
@@ -0,0 +1,421 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, normal_init
+from mmcv.ops import DeformConv2d
+
+from mmdet.core import multi_apply
+from ..builder import HEADS, build_loss
+from .corner_head import CornerHead
+
+
+@HEADS.register_module()
+class CentripetalHead(CornerHead):
+    """Head of CentripetalNet: Pursuing High-quality Keypoint Pairs for Object
+    Detection.
+
+    CentripetalHead inherits from :class:`CornerHead`. It removes the
+    embedding branch and adds guiding shift and centripetal shift branches.
+    More details can be found in the `paper
+    <https://arxiv.org/abs/2003.09119>`_ .
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        num_feat_levels (int): Levels of feature from the previous module. 2
+            for HourglassNet-104 and 1 for HourglassNet-52. HourglassNet-104
+            outputs the final feature and intermediate supervision feature and
+            HourglassNet-52 only outputs the final feature. Default: 2.
+        corner_emb_channels (int): Channel of embedding vector. Default: 1.
+        train_cfg (dict | None): Training config. Useless in CornerHead,
+            but we keep this variable for SingleStageDetector. Default: None.
+        test_cfg (dict | None): Testing config of CornerHead. Default: None.
+        loss_heatmap (dict | None): Config of corner heatmap loss. Default:
+            GaussianFocalLoss.
+        loss_embedding (dict | None): Config of corner embedding loss. Default:
+            AssociativeEmbeddingLoss.
+        loss_offset (dict | None): Config of corner offset loss. Default:
+            SmoothL1Loss.
+        loss_guiding_shift (dict): Config of guiding shift loss. Default:
+            SmoothL1Loss.
+        loss_centripetal_shift (dict): Config of centripetal shift loss.
+            Default: SmoothL1Loss.
+    """
+
+    def __init__(self,
+                 *args,
+                 centripetal_shift_channels=2,
+                 guiding_shift_channels=2,
+                 feat_adaption_conv_kernel=3,
+                 loss_guiding_shift=dict(
+                     type='SmoothL1Loss', beta=1.0, loss_weight=0.05),
+                 loss_centripetal_shift=dict(
+                     type='SmoothL1Loss', beta=1.0, loss_weight=1),
+                 **kwargs):
+        assert centripetal_shift_channels == 2, (
+            'CentripetalHead only support centripetal_shift_channels == 2')
+        self.centripetal_shift_channels = centripetal_shift_channels
+        assert guiding_shift_channels == 2, (
+            'CentripetalHead only support guiding_shift_channels == 2')
+        self.guiding_shift_channels = guiding_shift_channels
+        self.feat_adaption_conv_kernel = feat_adaption_conv_kernel
+        super(CentripetalHead, self).__init__(*args, **kwargs)
+        self.loss_guiding_shift = build_loss(loss_guiding_shift)
+        self.loss_centripetal_shift = build_loss(loss_centripetal_shift)
+
+    def _init_centripetal_layers(self):
+        """Initialize centripetal layers.
+
+        Including feature adaption deform convs (feat_adaption), deform offset
+        prediction convs (dcn_off), guiding shift (guiding_shift) and
+        centripetal shift ( centripetal_shift). Each branch has two parts:
+        prefix `tl_` for top-left and `br_` for bottom-right.
+        """
+        self.tl_feat_adaption = nn.ModuleList()
+        self.br_feat_adaption = nn.ModuleList()
+        self.tl_dcn_offset = nn.ModuleList()
+        self.br_dcn_offset = nn.ModuleList()
+        self.tl_guiding_shift = nn.ModuleList()
+        self.br_guiding_shift = nn.ModuleList()
+        self.tl_centripetal_shift = nn.ModuleList()
+        self.br_centripetal_shift = nn.ModuleList()
+
+        for _ in range(self.num_feat_levels):
+            self.tl_feat_adaption.append(
+                DeformConv2d(self.in_channels, self.in_channels,
+                             self.feat_adaption_conv_kernel, 1, 1))
+            self.br_feat_adaption.append(
+                DeformConv2d(self.in_channels, self.in_channels,
+                             self.feat_adaption_conv_kernel, 1, 1))
+
+            self.tl_guiding_shift.append(
+                self._make_layers(
+                    out_channels=self.guiding_shift_channels,
+                    in_channels=self.in_channels))
+            self.br_guiding_shift.append(
+                self._make_layers(
+                    out_channels=self.guiding_shift_channels,
+                    in_channels=self.in_channels))
+
+            self.tl_dcn_offset.append(
+                ConvModule(
+                    self.guiding_shift_channels,
+                    self.feat_adaption_conv_kernel**2 *
+                    self.guiding_shift_channels,
+                    1,
+                    bias=False,
+                    act_cfg=None))
+            self.br_dcn_offset.append(
+                ConvModule(
+                    self.guiding_shift_channels,
+                    self.feat_adaption_conv_kernel**2 *
+                    self.guiding_shift_channels,
+                    1,
+                    bias=False,
+                    act_cfg=None))
+
+            self.tl_centripetal_shift.append(
+                self._make_layers(
+                    out_channels=self.centripetal_shift_channels,
+                    in_channels=self.in_channels))
+            self.br_centripetal_shift.append(
+                self._make_layers(
+                    out_channels=self.centripetal_shift_channels,
+                    in_channels=self.in_channels))
+
+    def _init_layers(self):
+        """Initialize layers for CentripetalHead.
+
+        Including two parts: CornerHead layers and CentripetalHead layers
+        """
+        super()._init_layers()  # using _init_layers in CornerHead
+        self._init_centripetal_layers()
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        super().init_weights()
+        for i in range(self.num_feat_levels):
+            normal_init(self.tl_feat_adaption[i], std=0.01)
+            normal_init(self.br_feat_adaption[i], std=0.01)
+            normal_init(self.tl_dcn_offset[i].conv, std=0.1)
+            normal_init(self.br_dcn_offset[i].conv, std=0.1)
+            _ = [x.conv.reset_parameters() for x in self.tl_guiding_shift[i]]
+            _ = [x.conv.reset_parameters() for x in self.br_guiding_shift[i]]
+            _ = [
+                x.conv.reset_parameters() for x in self.tl_centripetal_shift[i]
+            ]
+            _ = [
+                x.conv.reset_parameters() for x in self.br_centripetal_shift[i]
+            ]
+
+    def forward_single(self, x, lvl_ind):
+        """Forward feature of a single level.
+
+        Args:
+            x (Tensor): Feature of a single level.
+            lvl_ind (int): Level index of current feature.
+
+        Returns:
+            tuple[Tensor]: A tuple of CentripetalHead's output for current
+            feature level. Containing the following Tensors:
+
+                - tl_heat (Tensor): Predicted top-left corner heatmap.
+                - br_heat (Tensor): Predicted bottom-right corner heatmap.
+                - tl_off (Tensor): Predicted top-left offset heatmap.
+                - br_off (Tensor): Predicted bottom-right offset heatmap.
+                - tl_guiding_shift (Tensor): Predicted top-left guiding shift
+                  heatmap.
+                - br_guiding_shift (Tensor): Predicted bottom-right guiding
+                  shift heatmap.
+                - tl_centripetal_shift (Tensor): Predicted top-left centripetal
+                  shift heatmap.
+                - br_centripetal_shift (Tensor): Predicted bottom-right
+                  centripetal shift heatmap.
+        """
+        tl_heat, br_heat, _, _, tl_off, br_off, tl_pool, br_pool = super(
+        ).forward_single(
+            x, lvl_ind, return_pool=True)
+
+        tl_guiding_shift = self.tl_guiding_shift[lvl_ind](tl_pool)
+        br_guiding_shift = self.br_guiding_shift[lvl_ind](br_pool)
+
+        tl_dcn_offset = self.tl_dcn_offset[lvl_ind](tl_guiding_shift.detach())
+        br_dcn_offset = self.br_dcn_offset[lvl_ind](br_guiding_shift.detach())
+
+        tl_feat_adaption = self.tl_feat_adaption[lvl_ind](tl_pool,
+                                                          tl_dcn_offset)
+        br_feat_adaption = self.br_feat_adaption[lvl_ind](br_pool,
+                                                          br_dcn_offset)
+
+        tl_centripetal_shift = self.tl_centripetal_shift[lvl_ind](
+            tl_feat_adaption)
+        br_centripetal_shift = self.br_centripetal_shift[lvl_ind](
+            br_feat_adaption)
+
+        result_list = [
+            tl_heat, br_heat, tl_off, br_off, tl_guiding_shift,
+            br_guiding_shift, tl_centripetal_shift, br_centripetal_shift
+        ]
+        return result_list
+
+    def loss(self,
+             tl_heats,
+             br_heats,
+             tl_offs,
+             br_offs,
+             tl_guiding_shifts,
+             br_guiding_shifts,
+             tl_centripetal_shifts,
+             br_centripetal_shifts,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
+                with shape (N, num_classes, H, W).
+            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
+                level with shape (N, num_classes, H, W).
+            tl_offs (list[Tensor]): Top-left corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            br_offs (list[Tensor]): Bottom-right corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            tl_guiding_shifts (list[Tensor]): Top-left guiding shifts for each
+                level with shape (N, guiding_shift_channels, H, W).
+            br_guiding_shifts (list[Tensor]): Bottom-right guiding shifts for
+                each level with shape (N, guiding_shift_channels, H, W).
+            tl_centripetal_shifts (list[Tensor]): Top-left centripetal shifts
+                for each level with shape (N, centripetal_shift_channels, H,
+                W).
+            br_centripetal_shifts (list[Tensor]): Bottom-right centripetal
+                shifts for each level with shape (N,
+                centripetal_shift_channels, H, W).
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [left, top, right, bottom] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): Specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components. Containing the
+            following losses:
+
+                - det_loss (list[Tensor]): Corner keypoint losses of all
+                  feature levels.
+                - off_loss (list[Tensor]): Corner offset losses of all feature
+                  levels.
+                - guiding_loss (list[Tensor]): Guiding shift losses of all
+                  feature levels.
+                - centripetal_loss (list[Tensor]): Centripetal shift losses of
+                  all feature levels.
+        """
+        targets = self.get_targets(
+            gt_bboxes,
+            gt_labels,
+            tl_heats[-1].shape,
+            img_metas[0]['pad_shape'],
+            with_corner_emb=self.with_corner_emb,
+            with_guiding_shift=True,
+            with_centripetal_shift=True)
+        mlvl_targets = [targets for _ in range(self.num_feat_levels)]
+        [det_losses, off_losses, guiding_losses, centripetal_losses
+         ] = multi_apply(self.loss_single, tl_heats, br_heats, tl_offs,
+                         br_offs, tl_guiding_shifts, br_guiding_shifts,
+                         tl_centripetal_shifts, br_centripetal_shifts,
+                         mlvl_targets)
+        loss_dict = dict(
+            det_loss=det_losses,
+            off_loss=off_losses,
+            guiding_loss=guiding_losses,
+            centripetal_loss=centripetal_losses)
+        return loss_dict
+
+    def loss_single(self, tl_hmp, br_hmp, tl_off, br_off, tl_guiding_shift,
+                    br_guiding_shift, tl_centripetal_shift,
+                    br_centripetal_shift, targets):
+        """Compute losses for single level.
+
+        Args:
+            tl_hmp (Tensor): Top-left corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            br_hmp (Tensor): Bottom-right corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            tl_off (Tensor): Top-left corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            br_off (Tensor): Bottom-right corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            tl_guiding_shift (Tensor): Top-left guiding shift for current level
+                with shape (N, guiding_shift_channels, H, W).
+            br_guiding_shift (Tensor): Bottom-right guiding shift for current
+                level with shape (N, guiding_shift_channels, H, W).
+            tl_centripetal_shift (Tensor): Top-left centripetal shift for
+                current level with shape (N, centripetal_shift_channels, H, W).
+            br_centripetal_shift (Tensor): Bottom-right centripetal shift for
+                current level with shape (N, centripetal_shift_channels, H, W).
+            targets (dict): Corner target generated by `get_targets`.
+
+        Returns:
+            tuple[torch.Tensor]: Losses of the head's differnet branches
+            containing the following losses:
+
+                - det_loss (Tensor): Corner keypoint loss.
+                - off_loss (Tensor): Corner offset loss.
+                - guiding_loss (Tensor): Guiding shift loss.
+                - centripetal_loss (Tensor): Centripetal shift loss.
+        """
+        targets['corner_embedding'] = None
+
+        det_loss, _, _, off_loss = super().loss_single(tl_hmp, br_hmp, None,
+                                                       None, tl_off, br_off,
+                                                       targets)
+
+        gt_tl_guiding_shift = targets['topleft_guiding_shift']
+        gt_br_guiding_shift = targets['bottomright_guiding_shift']
+        gt_tl_centripetal_shift = targets['topleft_centripetal_shift']
+        gt_br_centripetal_shift = targets['bottomright_centripetal_shift']
+
+        gt_tl_heatmap = targets['topleft_heatmap']
+        gt_br_heatmap = targets['bottomright_heatmap']
+        # We only compute the offset loss at the real corner position.
+        # The value of real corner would be 1 in heatmap ground truth.
+        # The mask is computed in class agnostic mode and its shape is
+        # batch * 1 * width * height.
+        tl_mask = gt_tl_heatmap.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
+            gt_tl_heatmap)
+        br_mask = gt_br_heatmap.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
+            gt_br_heatmap)
+
+        # Guiding shift loss
+        tl_guiding_loss = self.loss_guiding_shift(
+            tl_guiding_shift,
+            gt_tl_guiding_shift,
+            tl_mask,
+            avg_factor=tl_mask.sum())
+        br_guiding_loss = self.loss_guiding_shift(
+            br_guiding_shift,
+            gt_br_guiding_shift,
+            br_mask,
+            avg_factor=br_mask.sum())
+        guiding_loss = (tl_guiding_loss + br_guiding_loss) / 2.0
+        # Centripetal shift loss
+        tl_centripetal_loss = self.loss_centripetal_shift(
+            tl_centripetal_shift,
+            gt_tl_centripetal_shift,
+            tl_mask,
+            avg_factor=tl_mask.sum())
+        br_centripetal_loss = self.loss_centripetal_shift(
+            br_centripetal_shift,
+            gt_br_centripetal_shift,
+            br_mask,
+            avg_factor=br_mask.sum())
+        centripetal_loss = (tl_centripetal_loss + br_centripetal_loss) / 2.0
+
+        return det_loss, off_loss, guiding_loss, centripetal_loss
+
+    def get_bboxes(self,
+                   tl_heats,
+                   br_heats,
+                   tl_offs,
+                   br_offs,
+                   tl_guiding_shifts,
+                   br_guiding_shifts,
+                   tl_centripetal_shifts,
+                   br_centripetal_shifts,
+                   img_metas,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
+                with shape (N, num_classes, H, W).
+            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
+                level with shape (N, num_classes, H, W).
+            tl_offs (list[Tensor]): Top-left corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            br_offs (list[Tensor]): Bottom-right corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            tl_guiding_shifts (list[Tensor]): Top-left guiding shifts for each
+                level with shape (N, guiding_shift_channels, H, W). Useless in
+                this function, we keep this arg because it's the raw output
+                from CentripetalHead.
+            br_guiding_shifts (list[Tensor]): Bottom-right guiding shifts for
+                each level with shape (N, guiding_shift_channels, H, W).
+                Useless in this function, we keep this arg because it's the
+                raw output from CentripetalHead.
+            tl_centripetal_shifts (list[Tensor]): Top-left centripetal shifts
+                for each level with shape (N, centripetal_shift_channels, H,
+                W).
+            br_centripetal_shifts (list[Tensor]): Bottom-right centripetal
+                shifts for each level with shape (N,
+                centripetal_shift_channels, H, W).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+        """
+        assert tl_heats[-1].shape[0] == br_heats[-1].shape[0] == len(img_metas)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            result_list.append(
+                self._get_bboxes_single(
+                    tl_heats[-1][img_id:img_id + 1, :],
+                    br_heats[-1][img_id:img_id + 1, :],
+                    tl_offs[-1][img_id:img_id + 1, :],
+                    br_offs[-1][img_id:img_id + 1, :],
+                    img_metas[img_id],
+                    tl_emb=None,
+                    br_emb=None,
+                    tl_centripetal_shift=tl_centripetal_shifts[-1][
+                        img_id:img_id + 1, :],
+                    br_centripetal_shift=br_centripetal_shifts[-1][
+                        img_id:img_id + 1, :],
+                    rescale=rescale,
+                    with_nms=with_nms))
+
+        return result_list
diff --git a/detection/scrfd/mmdet/models/dense_heads/corner_head.py b/detection/scrfd/mmdet/models/dense_heads/corner_head.py
new file mode 100755
index 000000000..cdaeca431
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/corner_head.py
@@ -0,0 +1,1064 @@
+from math import ceil, log
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, bias_init_with_prob
+from mmcv.ops import CornerPool, batched_nms
+
+from mmdet.core import multi_apply
+from ..builder import HEADS, build_loss
+from ..utils import gaussian_radius, gen_gaussian_target
+from .base_dense_head import BaseDenseHead
+
+
+class BiCornerPool(nn.Module):
+    """Bidirectional Corner Pooling Module (TopLeft, BottomRight, etc.)
+
+    Args:
+        in_channels (int): Input channels of module.
+        out_channels (int): Output channels of module.
+        feat_channels (int): Feature channels of module.
+        directions (list[str]): Directions of two CornerPools.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 directions,
+                 feat_channels=128,
+                 out_channels=128,
+                 norm_cfg=dict(type='BN', requires_grad=True)):
+        super(BiCornerPool, self).__init__()
+        self.direction1_conv = ConvModule(
+            in_channels, feat_channels, 3, padding=1, norm_cfg=norm_cfg)
+        self.direction2_conv = ConvModule(
+            in_channels, feat_channels, 3, padding=1, norm_cfg=norm_cfg)
+
+        self.aftpool_conv = ConvModule(
+            feat_channels,
+            out_channels,
+            3,
+            padding=1,
+            norm_cfg=norm_cfg,
+            act_cfg=None)
+
+        self.conv1 = ConvModule(
+            in_channels, out_channels, 1, norm_cfg=norm_cfg, act_cfg=None)
+        self.conv2 = ConvModule(
+            in_channels, out_channels, 3, padding=1, norm_cfg=norm_cfg)
+
+        self.direction1_pool = CornerPool(directions[0])
+        self.direction2_pool = CornerPool(directions[1])
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        """Forward features from the upstream network.
+
+        Args:
+            x (tensor): Input feature of BiCornerPool.
+
+        Returns:
+            conv2 (tensor): Output feature of BiCornerPool.
+        """
+        direction1_conv = self.direction1_conv(x)
+        direction2_conv = self.direction2_conv(x)
+        direction1_feat = self.direction1_pool(direction1_conv)
+        direction2_feat = self.direction2_pool(direction2_conv)
+        aftpool_conv = self.aftpool_conv(direction1_feat + direction2_feat)
+        conv1 = self.conv1(x)
+        relu = self.relu(aftpool_conv + conv1)
+        conv2 = self.conv2(relu)
+        return conv2
+
+
+@HEADS.register_module()
+class CornerHead(BaseDenseHead):
+    """Head of CornerNet: Detecting Objects as Paired Keypoints.
+
+    Code is modified from the `official github repo
+    <https://github.com/princeton-vl/CornerNet/blob/master/models/py_utils/
+    kp.py#L73>`_ .
+
+    More details can be found in the `paper
+    <https://arxiv.org/abs/1808.01244>`_ .
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        num_feat_levels (int): Levels of feature from the previous module. 2
+            for HourglassNet-104 and 1 for HourglassNet-52. Because
+            HourglassNet-104 outputs the final feature and intermediate
+            supervision feature and HourglassNet-52 only outputs the final
+            feature. Default: 2.
+        corner_emb_channels (int): Channel of embedding vector. Default: 1.
+        train_cfg (dict | None): Training config. Useless in CornerHead,
+            but we keep this variable for SingleStageDetector. Default: None.
+        test_cfg (dict | None): Testing config of CornerHead. Default: None.
+        loss_heatmap (dict | None): Config of corner heatmap loss. Default:
+            GaussianFocalLoss.
+        loss_embedding (dict | None): Config of corner embedding loss. Default:
+            AssociativeEmbeddingLoss.
+        loss_offset (dict | None): Config of corner offset loss. Default:
+            SmoothL1Loss.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 num_feat_levels=2,
+                 corner_emb_channels=1,
+                 train_cfg=None,
+                 test_cfg=None,
+                 loss_heatmap=dict(
+                     type='GaussianFocalLoss',
+                     alpha=2.0,
+                     gamma=4.0,
+                     loss_weight=1),
+                 loss_embedding=dict(
+                     type='AssociativeEmbeddingLoss',
+                     pull_weight=0.25,
+                     push_weight=0.25),
+                 loss_offset=dict(
+                     type='SmoothL1Loss', beta=1.0, loss_weight=1)):
+        super(CornerHead, self).__init__()
+        self.num_classes = num_classes
+        self.in_channels = in_channels
+        self.corner_emb_channels = corner_emb_channels
+        self.with_corner_emb = self.corner_emb_channels > 0
+        self.corner_offset_channels = 2
+        self.num_feat_levels = num_feat_levels
+        self.loss_heatmap = build_loss(
+            loss_heatmap) if loss_heatmap is not None else None
+        self.loss_embedding = build_loss(
+            loss_embedding) if loss_embedding is not None else None
+        self.loss_offset = build_loss(
+            loss_offset) if loss_offset is not None else None
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+
+        self._init_layers()
+
+    def _make_layers(self, out_channels, in_channels=256, feat_channels=256):
+        """Initialize conv sequential for CornerHead."""
+        return nn.Sequential(
+            ConvModule(in_channels, feat_channels, 3, padding=1),
+            ConvModule(
+                feat_channels, out_channels, 1, norm_cfg=None, act_cfg=None))
+
+    def _init_corner_kpt_layers(self):
+        """Initialize corner keypoint layers.
+
+        Including corner heatmap branch and corner offset branch. Each branch
+        has two parts: prefix `tl_` for top-left and `br_` for bottom-right.
+        """
+        self.tl_pool, self.br_pool = nn.ModuleList(), nn.ModuleList()
+        self.tl_heat, self.br_heat = nn.ModuleList(), nn.ModuleList()
+        self.tl_off, self.br_off = nn.ModuleList(), nn.ModuleList()
+
+        for _ in range(self.num_feat_levels):
+            self.tl_pool.append(
+                BiCornerPool(
+                    self.in_channels, ['top', 'left'],
+                    out_channels=self.in_channels))
+            self.br_pool.append(
+                BiCornerPool(
+                    self.in_channels, ['bottom', 'right'],
+                    out_channels=self.in_channels))
+
+            self.tl_heat.append(
+                self._make_layers(
+                    out_channels=self.num_classes,
+                    in_channels=self.in_channels))
+            self.br_heat.append(
+                self._make_layers(
+                    out_channels=self.num_classes,
+                    in_channels=self.in_channels))
+
+            self.tl_off.append(
+                self._make_layers(
+                    out_channels=self.corner_offset_channels,
+                    in_channels=self.in_channels))
+            self.br_off.append(
+                self._make_layers(
+                    out_channels=self.corner_offset_channels,
+                    in_channels=self.in_channels))
+
+    def _init_corner_emb_layers(self):
+        """Initialize corner embedding layers.
+
+        Only include corner embedding branch with two parts: prefix `tl_` for
+        top-left and `br_` for bottom-right.
+        """
+        self.tl_emb, self.br_emb = nn.ModuleList(), nn.ModuleList()
+
+        for _ in range(self.num_feat_levels):
+            self.tl_emb.append(
+                self._make_layers(
+                    out_channels=self.corner_emb_channels,
+                    in_channels=self.in_channels))
+            self.br_emb.append(
+                self._make_layers(
+                    out_channels=self.corner_emb_channels,
+                    in_channels=self.in_channels))
+
+    def _init_layers(self):
+        """Initialize layers for CornerHead.
+
+        Including two parts: corner keypoint layers and corner embedding layers
+        """
+        self._init_corner_kpt_layers()
+        if self.with_corner_emb:
+            self._init_corner_emb_layers()
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        bias_init = bias_init_with_prob(0.1)
+        for i in range(self.num_feat_levels):
+            # The initialization of parameters are different between nn.Conv2d
+            # and ConvModule. Our experiments show that using the original
+            # initialization of nn.Conv2d increases the final mAP by about 0.2%
+            self.tl_heat[i][-1].conv.reset_parameters()
+            self.tl_heat[i][-1].conv.bias.data.fill_(bias_init)
+            self.br_heat[i][-1].conv.reset_parameters()
+            self.br_heat[i][-1].conv.bias.data.fill_(bias_init)
+            self.tl_off[i][-1].conv.reset_parameters()
+            self.br_off[i][-1].conv.reset_parameters()
+            if self.with_corner_emb:
+                self.tl_emb[i][-1].conv.reset_parameters()
+                self.br_emb[i][-1].conv.reset_parameters()
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually a tuple of corner heatmaps, offset heatmaps and
+            embedding heatmaps.
+                - tl_heats (list[Tensor]): Top-left corner heatmaps for all
+                  levels, each is a 4D-tensor, the channels number is
+                  num_classes.
+                - br_heats (list[Tensor]): Bottom-right corner heatmaps for all
+                  levels, each is a 4D-tensor, the channels number is
+                  num_classes.
+                - tl_embs (list[Tensor] | list[None]): Top-left embedding
+                  heatmaps for all levels, each is a 4D-tensor or None.
+                  If not None, the channels number is corner_emb_channels.
+                - br_embs (list[Tensor] | list[None]): Bottom-right embedding
+                  heatmaps for all levels, each is a 4D-tensor or None.
+                  If not None, the channels number is corner_emb_channels.
+                - tl_offs (list[Tensor]): Top-left offset heatmaps for all
+                  levels, each is a 4D-tensor. The channels number is
+                  corner_offset_channels.
+                - br_offs (list[Tensor]): Bottom-right offset heatmaps for all
+                  levels, each is a 4D-tensor. The channels number is
+                  corner_offset_channels.
+        """
+        lvl_ind = list(range(self.num_feat_levels))
+        return multi_apply(self.forward_single, feats, lvl_ind)
+
+    def forward_single(self, x, lvl_ind, return_pool=False):
+        """Forward feature of a single level.
+
+        Args:
+            x (Tensor): Feature of a single level.
+            lvl_ind (int): Level index of current feature.
+            return_pool (bool): Return corner pool feature or not.
+
+        Returns:
+            tuple[Tensor]: A tuple of CornerHead's output for current feature
+            level. Containing the following Tensors:
+
+                - tl_heat (Tensor): Predicted top-left corner heatmap.
+                - br_heat (Tensor): Predicted bottom-right corner heatmap.
+                - tl_emb (Tensor | None): Predicted top-left embedding heatmap.
+                  None for `self.with_corner_emb == False`.
+                - br_emb (Tensor | None): Predicted bottom-right embedding
+                  heatmap. None for `self.with_corner_emb == False`.
+                - tl_off (Tensor): Predicted top-left offset heatmap.
+                - br_off (Tensor): Predicted bottom-right offset heatmap.
+                - tl_pool (Tensor): Top-left corner pool feature. Not must
+                  have.
+                - br_pool (Tensor): Bottom-right corner pool feature. Not must
+                  have.
+        """
+        tl_pool = self.tl_pool[lvl_ind](x)
+        tl_heat = self.tl_heat[lvl_ind](tl_pool)
+        br_pool = self.br_pool[lvl_ind](x)
+        br_heat = self.br_heat[lvl_ind](br_pool)
+
+        tl_emb, br_emb = None, None
+        if self.with_corner_emb:
+            tl_emb = self.tl_emb[lvl_ind](tl_pool)
+            br_emb = self.br_emb[lvl_ind](br_pool)
+
+        tl_off = self.tl_off[lvl_ind](tl_pool)
+        br_off = self.br_off[lvl_ind](br_pool)
+
+        result_list = [tl_heat, br_heat, tl_emb, br_emb, tl_off, br_off]
+        if return_pool:
+            result_list.append(tl_pool)
+            result_list.append(br_pool)
+
+        return result_list
+
+    def get_targets(self,
+                    gt_bboxes,
+                    gt_labels,
+                    feat_shape,
+                    img_shape,
+                    with_corner_emb=False,
+                    with_guiding_shift=False,
+                    with_centripetal_shift=False):
+        """Generate corner targets.
+
+        Including corner heatmap, corner offset.
+
+        Optional: corner embedding, corner guiding shift, centripetal shift.
+
+        For CornerNet, we generate corner heatmap, corner offset and corner
+        embedding from this function.
+
+        For CentripetalNet, we generate corner heatmap, corner offset, guiding
+        shift and centripetal shift from this function.
+
+        Args:
+            gt_bboxes (list[Tensor]): Ground truth bboxes of each image, each
+                has shape (num_gt, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each box, each has
+                shape (num_gt,).
+            feat_shape (list[int]): Shape of output feature,
+                [batch, channel, height, width].
+            img_shape (list[int]): Shape of input image,
+                [height, width, channel].
+            with_corner_emb (bool): Generate corner embedding target or not.
+                Default: False.
+            with_guiding_shift (bool): Generate guiding shift target or not.
+                Default: False.
+            with_centripetal_shift (bool): Generate centripetal shift target or
+                not. Default: False.
+
+        Returns:
+            dict: Ground truth of corner heatmap, corner offset, corner
+            embedding, guiding shift and centripetal shift. Containing the
+            following keys:
+
+                - topleft_heatmap (Tensor): Ground truth top-left corner
+                  heatmap.
+                - bottomright_heatmap (Tensor): Ground truth bottom-right
+                  corner heatmap.
+                - topleft_offset (Tensor): Ground truth top-left corner offset.
+                - bottomright_offset (Tensor): Ground truth bottom-right corner
+                  offset.
+                - corner_embedding (list[list[list[int]]]): Ground truth corner
+                  embedding. Not must have.
+                - topleft_guiding_shift (Tensor): Ground truth top-left corner
+                  guiding shift. Not must have.
+                - bottomright_guiding_shift (Tensor): Ground truth bottom-right
+                  corner guiding shift. Not must have.
+                - topleft_centripetal_shift (Tensor): Ground truth top-left
+                  corner centripetal shift. Not must have.
+                - bottomright_centripetal_shift (Tensor): Ground truth
+                  bottom-right corner centripetal shift. Not must have.
+        """
+        batch_size, _, height, width = feat_shape
+        img_h, img_w = img_shape[:2]
+
+        width_ratio = float(width / img_w)
+        height_ratio = float(height / img_h)
+
+        gt_tl_heatmap = gt_bboxes[-1].new_zeros(
+            [batch_size, self.num_classes, height, width])
+        gt_br_heatmap = gt_bboxes[-1].new_zeros(
+            [batch_size, self.num_classes, height, width])
+        gt_tl_offset = gt_bboxes[-1].new_zeros([batch_size, 2, height, width])
+        gt_br_offset = gt_bboxes[-1].new_zeros([batch_size, 2, height, width])
+
+        if with_corner_emb:
+            match = []
+
+        # Guiding shift is a kind of offset, from center to corner
+        if with_guiding_shift:
+            gt_tl_guiding_shift = gt_bboxes[-1].new_zeros(
+                [batch_size, 2, height, width])
+            gt_br_guiding_shift = gt_bboxes[-1].new_zeros(
+                [batch_size, 2, height, width])
+        # Centripetal shift is also a kind of offset, from center to corner
+        # and normalized by log.
+        if with_centripetal_shift:
+            gt_tl_centripetal_shift = gt_bboxes[-1].new_zeros(
+                [batch_size, 2, height, width])
+            gt_br_centripetal_shift = gt_bboxes[-1].new_zeros(
+                [batch_size, 2, height, width])
+
+        for batch_id in range(batch_size):
+            # Ground truth of corner embedding per image is a list of coord set
+            corner_match = []
+            for box_id in range(len(gt_labels[batch_id])):
+                left, top, right, bottom = gt_bboxes[batch_id][box_id]
+                center_x = (left + right) / 2.0
+                center_y = (top + bottom) / 2.0
+                label = gt_labels[batch_id][box_id]
+
+                # Use coords in the feature level to generate ground truth
+                scale_left = left * width_ratio
+                scale_right = right * width_ratio
+                scale_top = top * height_ratio
+                scale_bottom = bottom * height_ratio
+                scale_center_x = center_x * width_ratio
+                scale_center_y = center_y * height_ratio
+
+                # Int coords on feature map/ground truth tensor
+                left_idx = int(min(scale_left, width - 1))
+                right_idx = int(min(scale_right, width - 1))
+                top_idx = int(min(scale_top, height - 1))
+                bottom_idx = int(min(scale_bottom, height - 1))
+
+                # Generate gaussian heatmap
+                scale_box_width = ceil(scale_right - scale_left)
+                scale_box_height = ceil(scale_bottom - scale_top)
+                radius = gaussian_radius((scale_box_height, scale_box_width),
+                                         min_overlap=0.3)
+                radius = max(0, int(radius))
+                gt_tl_heatmap[batch_id, label] = gen_gaussian_target(
+                    gt_tl_heatmap[batch_id, label], [left_idx, top_idx],
+                    radius)
+                gt_br_heatmap[batch_id, label] = gen_gaussian_target(
+                    gt_br_heatmap[batch_id, label], [right_idx, bottom_idx],
+                    radius)
+
+                # Generate corner offset
+                left_offset = scale_left - left_idx
+                top_offset = scale_top - top_idx
+                right_offset = scale_right - right_idx
+                bottom_offset = scale_bottom - bottom_idx
+                gt_tl_offset[batch_id, 0, top_idx, left_idx] = left_offset
+                gt_tl_offset[batch_id, 1, top_idx, left_idx] = top_offset
+                gt_br_offset[batch_id, 0, bottom_idx, right_idx] = right_offset
+                gt_br_offset[batch_id, 1, bottom_idx,
+                             right_idx] = bottom_offset
+
+                # Generate corner embedding
+                if with_corner_emb:
+                    corner_match.append([[top_idx, left_idx],
+                                         [bottom_idx, right_idx]])
+                # Generate guiding shift
+                if with_guiding_shift:
+                    gt_tl_guiding_shift[batch_id, 0, top_idx,
+                                        left_idx] = scale_center_x - left_idx
+                    gt_tl_guiding_shift[batch_id, 1, top_idx,
+                                        left_idx] = scale_center_y - top_idx
+                    gt_br_guiding_shift[batch_id, 0, bottom_idx,
+                                        right_idx] = right_idx - scale_center_x
+                    gt_br_guiding_shift[
+                        batch_id, 1, bottom_idx,
+                        right_idx] = bottom_idx - scale_center_y
+                # Generate centripetal shift
+                if with_centripetal_shift:
+                    gt_tl_centripetal_shift[batch_id, 0, top_idx,
+                                            left_idx] = log(scale_center_x -
+                                                            scale_left)
+                    gt_tl_centripetal_shift[batch_id, 1, top_idx,
+                                            left_idx] = log(scale_center_y -
+                                                            scale_top)
+                    gt_br_centripetal_shift[batch_id, 0, bottom_idx,
+                                            right_idx] = log(scale_right -
+                                                             scale_center_x)
+                    gt_br_centripetal_shift[batch_id, 1, bottom_idx,
+                                            right_idx] = log(scale_bottom -
+                                                             scale_center_y)
+
+            if with_corner_emb:
+                match.append(corner_match)
+
+        target_result = dict(
+            topleft_heatmap=gt_tl_heatmap,
+            topleft_offset=gt_tl_offset,
+            bottomright_heatmap=gt_br_heatmap,
+            bottomright_offset=gt_br_offset)
+
+        if with_corner_emb:
+            target_result.update(corner_embedding=match)
+        if with_guiding_shift:
+            target_result.update(
+                topleft_guiding_shift=gt_tl_guiding_shift,
+                bottomright_guiding_shift=gt_br_guiding_shift)
+        if with_centripetal_shift:
+            target_result.update(
+                topleft_centripetal_shift=gt_tl_centripetal_shift,
+                bottomright_centripetal_shift=gt_br_centripetal_shift)
+
+        return target_result
+
+    def loss(self,
+             tl_heats,
+             br_heats,
+             tl_embs,
+             br_embs,
+             tl_offs,
+             br_offs,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
+                with shape (N, num_classes, H, W).
+            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
+                level with shape (N, num_classes, H, W).
+            tl_embs (list[Tensor]): Top-left corner embeddings for each level
+                with shape (N, corner_emb_channels, H, W).
+            br_embs (list[Tensor]): Bottom-right corner embeddings for each
+                level with shape (N, corner_emb_channels, H, W).
+            tl_offs (list[Tensor]): Top-left corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            br_offs (list[Tensor]): Bottom-right corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [left, top, right, bottom] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): Specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components. Containing the
+            following losses:
+
+                - det_loss (list[Tensor]): Corner keypoint losses of all
+                  feature levels.
+                - pull_loss (list[Tensor]): Part one of AssociativeEmbedding
+                  losses of all feature levels.
+                - push_loss (list[Tensor]): Part two of AssociativeEmbedding
+                  losses of all feature levels.
+                - off_loss (list[Tensor]): Corner offset losses of all feature
+                  levels.
+        """
+        targets = self.get_targets(
+            gt_bboxes,
+            gt_labels,
+            tl_heats[-1].shape,
+            img_metas[0]['pad_shape'],
+            with_corner_emb=self.with_corner_emb)
+        mlvl_targets = [targets for _ in range(self.num_feat_levels)]
+        det_losses, pull_losses, push_losses, off_losses = multi_apply(
+            self.loss_single, tl_heats, br_heats, tl_embs, br_embs, tl_offs,
+            br_offs, mlvl_targets)
+        loss_dict = dict(det_loss=det_losses, off_loss=off_losses)
+        if self.with_corner_emb:
+            loss_dict.update(pull_loss=pull_losses, push_loss=push_losses)
+        return loss_dict
+
+    def loss_single(self, tl_hmp, br_hmp, tl_emb, br_emb, tl_off, br_off,
+                    targets):
+        """Compute losses for single level.
+
+        Args:
+            tl_hmp (Tensor): Top-left corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            br_hmp (Tensor): Bottom-right corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            tl_emb (Tensor): Top-left corner embedding for current level with
+                shape (N, corner_emb_channels, H, W).
+            br_emb (Tensor): Bottom-right corner embedding for current level
+                with shape (N, corner_emb_channels, H, W).
+            tl_off (Tensor): Top-left corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            br_off (Tensor): Bottom-right corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            targets (dict): Corner target generated by `get_targets`.
+
+        Returns:
+            tuple[torch.Tensor]: Losses of the head's differnet branches
+            containing the following losses:
+
+                - det_loss (Tensor): Corner keypoint loss.
+                - pull_loss (Tensor): Part one of AssociativeEmbedding loss.
+                - push_loss (Tensor): Part two of AssociativeEmbedding loss.
+                - off_loss (Tensor): Corner offset loss.
+        """
+        gt_tl_hmp = targets['topleft_heatmap']
+        gt_br_hmp = targets['bottomright_heatmap']
+        gt_tl_off = targets['topleft_offset']
+        gt_br_off = targets['bottomright_offset']
+        gt_embedding = targets['corner_embedding']
+
+        # Detection loss
+        tl_det_loss = self.loss_heatmap(
+            tl_hmp.sigmoid(),
+            gt_tl_hmp,
+            avg_factor=max(1,
+                           gt_tl_hmp.eq(1).sum()))
+        br_det_loss = self.loss_heatmap(
+            br_hmp.sigmoid(),
+            gt_br_hmp,
+            avg_factor=max(1,
+                           gt_br_hmp.eq(1).sum()))
+        det_loss = (tl_det_loss + br_det_loss) / 2.0
+
+        # AssociativeEmbedding loss
+        if self.with_corner_emb and self.loss_embedding is not None:
+            pull_loss, push_loss = self.loss_embedding(tl_emb, br_emb,
+                                                       gt_embedding)
+        else:
+            pull_loss, push_loss = None, None
+
+        # Offset loss
+        # We only compute the offset loss at the real corner position.
+        # The value of real corner would be 1 in heatmap ground truth.
+        # The mask is computed in class agnostic mode and its shape is
+        # batch * 1 * width * height.
+        tl_off_mask = gt_tl_hmp.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
+            gt_tl_hmp)
+        br_off_mask = gt_br_hmp.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
+            gt_br_hmp)
+        tl_off_loss = self.loss_offset(
+            tl_off,
+            gt_tl_off,
+            tl_off_mask,
+            avg_factor=max(1, tl_off_mask.sum()))
+        br_off_loss = self.loss_offset(
+            br_off,
+            gt_br_off,
+            br_off_mask,
+            avg_factor=max(1, br_off_mask.sum()))
+
+        off_loss = (tl_off_loss + br_off_loss) / 2.0
+
+        return det_loss, pull_loss, push_loss, off_loss
+
+    def get_bboxes(self,
+                   tl_heats,
+                   br_heats,
+                   tl_embs,
+                   br_embs,
+                   tl_offs,
+                   br_offs,
+                   img_metas,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
+                with shape (N, num_classes, H, W).
+            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
+                level with shape (N, num_classes, H, W).
+            tl_embs (list[Tensor]): Top-left corner embeddings for each level
+                with shape (N, corner_emb_channels, H, W).
+            br_embs (list[Tensor]): Bottom-right corner embeddings for each
+                level with shape (N, corner_emb_channels, H, W).
+            tl_offs (list[Tensor]): Top-left corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            br_offs (list[Tensor]): Bottom-right corner offsets for each level
+                with shape (N, corner_offset_channels, H, W).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+        """
+        assert tl_heats[-1].shape[0] == br_heats[-1].shape[0] == len(img_metas)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            result_list.append(
+                self._get_bboxes_single(
+                    tl_heats[-1][img_id:img_id + 1, :],
+                    br_heats[-1][img_id:img_id + 1, :],
+                    tl_offs[-1][img_id:img_id + 1, :],
+                    br_offs[-1][img_id:img_id + 1, :],
+                    img_metas[img_id],
+                    tl_emb=tl_embs[-1][img_id:img_id + 1, :],
+                    br_emb=br_embs[-1][img_id:img_id + 1, :],
+                    rescale=rescale,
+                    with_nms=with_nms))
+
+        return result_list
+
+    def _get_bboxes_single(self,
+                           tl_heat,
+                           br_heat,
+                           tl_off,
+                           br_off,
+                           img_meta,
+                           tl_emb=None,
+                           br_emb=None,
+                           tl_centripetal_shift=None,
+                           br_centripetal_shift=None,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            tl_heat (Tensor): Top-left corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            br_heat (Tensor): Bottom-right corner heatmap for current level
+                with shape (N, num_classes, H, W).
+            tl_off (Tensor): Top-left corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            br_off (Tensor): Bottom-right corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            img_meta (dict): Meta information of current image, e.g.,
+                image size, scaling factor, etc.
+            tl_emb (Tensor): Top-left corner embedding for current level with
+                shape (N, corner_emb_channels, H, W).
+            br_emb (Tensor): Bottom-right corner embedding for current level
+                with shape (N, corner_emb_channels, H, W).
+            tl_centripetal_shift: Top-left corner's centripetal shift for
+                current level with shape (N, 2, H, W).
+            br_centripetal_shift: Bottom-right corner's centripetal shift for
+                current level with shape (N, 2, H, W).
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+        """
+        if isinstance(img_meta, (list, tuple)):
+            img_meta = img_meta[0]
+
+        batch_bboxes, batch_scores, batch_clses = self.decode_heatmap(
+            tl_heat=tl_heat.sigmoid(),
+            br_heat=br_heat.sigmoid(),
+            tl_off=tl_off,
+            br_off=br_off,
+            tl_emb=tl_emb,
+            br_emb=br_emb,
+            tl_centripetal_shift=tl_centripetal_shift,
+            br_centripetal_shift=br_centripetal_shift,
+            img_meta=img_meta,
+            k=self.test_cfg.corner_topk,
+            kernel=self.test_cfg.local_maximum_kernel,
+            distance_threshold=self.test_cfg.distance_threshold)
+
+        if rescale:
+            batch_bboxes /= img_meta['scale_factor']
+
+        bboxes = batch_bboxes.view([-1, 4])
+        scores = batch_scores.view([-1, 1])
+        clses = batch_clses.view([-1, 1])
+
+        idx = scores.argsort(dim=0, descending=True)
+        bboxes = bboxes[idx].view([-1, 4])
+        scores = scores[idx].view(-1)
+        clses = clses[idx].view(-1)
+
+        detections = torch.cat([bboxes, scores.unsqueeze(-1)], -1)
+        keepinds = (detections[:, -1] > -0.1)
+        detections = detections[keepinds]
+        labels = clses[keepinds]
+
+        if with_nms:
+            detections, labels = self._bboxes_nms(detections, labels,
+                                                  self.test_cfg)
+
+        return detections, labels
+
+    def _bboxes_nms(self, bboxes, labels, cfg):
+        out_bboxes, keep = batched_nms(bboxes[:, :4], bboxes[:, -1], labels,
+                                       cfg.nms_cfg)
+        out_labels = labels[keep]
+
+        if len(out_bboxes) > 0:
+            idx = torch.argsort(out_bboxes[:, -1], descending=True)
+            idx = idx[:cfg.max_per_img]
+            out_bboxes = out_bboxes[idx]
+            out_labels = out_labels[idx]
+
+        return out_bboxes, out_labels
+
+    def _gather_feat(self, feat, ind, mask=None):
+        """Gather feature according to index.
+
+        Args:
+            feat (Tensor): Target feature map.
+            ind (Tensor): Target coord index.
+            mask (Tensor | None): Mask of featuremap. Default: None.
+
+        Returns:
+            feat (Tensor): Gathered feature.
+        """
+        dim = feat.size(2)
+        ind = ind.unsqueeze(2).repeat(1, 1, dim)
+        feat = feat.gather(1, ind)
+        if mask is not None:
+            mask = mask.unsqueeze(2).expand_as(feat)
+            feat = feat[mask]
+            feat = feat.view(-1, dim)
+        return feat
+
+    def _local_maximum(self, heat, kernel=3):
+        """Extract local maximum pixel with given kernal.
+
+        Args:
+            heat (Tensor): Target heatmap.
+            kernel (int): Kernel size of max pooling. Default: 3.
+
+        Returns:
+            heat (Tensor): A heatmap where local maximum pixels maintain its
+                own value and other positions are 0.
+        """
+        pad = (kernel - 1) // 2
+        hmax = F.max_pool2d(heat, kernel, stride=1, padding=pad)
+        keep = (hmax == heat).float()
+        return heat * keep
+
+    def _transpose_and_gather_feat(self, feat, ind):
+        """Transpose and gather feature according to index.
+
+        Args:
+            feat (Tensor): Target feature map.
+            ind (Tensor): Target coord index.
+
+        Returns:
+            feat (Tensor): Transposed and gathered feature.
+        """
+        feat = feat.permute(0, 2, 3, 1).contiguous()
+        feat = feat.view(feat.size(0), -1, feat.size(3))
+        feat = self._gather_feat(feat, ind)
+        return feat
+
+    def _topk(self, scores, k=20):
+        """Get top k positions from heatmap.
+
+        Args:
+            scores (Tensor): Target heatmap with shape
+                [batch, num_classes, height, width].
+            k (int): Target number. Default: 20.
+
+        Returns:
+            tuple[torch.Tensor]: Scores, indexes, categories and coords of
+                topk keypoint. Containing following Tensors:
+
+            - topk_scores (Tensor): Max scores of each topk keypoint.
+            - topk_inds (Tensor): Indexes of each topk keypoint.
+            - topk_clses (Tensor): Categories of each topk keypoint.
+            - topk_ys (Tensor): Y-coord of each topk keypoint.
+            - topk_xs (Tensor): X-coord of each topk keypoint.
+        """
+        batch, _, height, width = scores.size()
+        topk_scores, topk_inds = torch.topk(scores.view(batch, -1), k)
+        topk_clses = topk_inds // (height * width)
+        topk_inds = topk_inds % (height * width)
+        topk_ys = topk_inds // width
+        topk_xs = (topk_inds % width).int().float()
+        return topk_scores, topk_inds, topk_clses, topk_ys, topk_xs
+
+    def decode_heatmap(self,
+                       tl_heat,
+                       br_heat,
+                       tl_off,
+                       br_off,
+                       tl_emb=None,
+                       br_emb=None,
+                       tl_centripetal_shift=None,
+                       br_centripetal_shift=None,
+                       img_meta=None,
+                       k=100,
+                       kernel=3,
+                       distance_threshold=0.5,
+                       num_dets=1000):
+        """Transform outputs for a single batch item into raw bbox predictions.
+
+        Args:
+            tl_heat (Tensor): Top-left corner heatmap for current level with
+                shape (N, num_classes, H, W).
+            br_heat (Tensor): Bottom-right corner heatmap for current level
+                with shape (N, num_classes, H, W).
+            tl_off (Tensor): Top-left corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            br_off (Tensor): Bottom-right corner offset for current level with
+                shape (N, corner_offset_channels, H, W).
+            tl_emb (Tensor | None): Top-left corner embedding for current
+                level with shape (N, corner_emb_channels, H, W).
+            br_emb (Tensor | None): Bottom-right corner embedding for current
+                level with shape (N, corner_emb_channels, H, W).
+            tl_centripetal_shift (Tensor | None): Top-left centripetal shift
+                for current level with shape (N, 2, H, W).
+            br_centripetal_shift (Tensor | None): Bottom-right centripetal
+                shift for current level with shape (N, 2, H, W).
+            img_meta (dict): Meta information of current image, e.g.,
+                image size, scaling factor, etc.
+            k (int): Get top k corner keypoints from heatmap.
+            kernel (int): Max pooling kernel for extract local maximum pixels.
+            distance_threshold (float): Distance threshold. Top-left and
+                bottom-right corner keypoints with feature distance less than
+                the threshold will be regarded as keypoints from same object.
+            num_dets (int): Num of raw boxes before doing nms.
+
+        Returns:
+            tuple[torch.Tensor]: Decoded output of CornerHead, containing the
+            following Tensors:
+
+            - bboxes (Tensor): Coords of each box.
+            - scores (Tensor): Scores of each box.
+            - clses (Tensor): Categories of each box.
+        """
+        with_embedding = tl_emb is not None and br_emb is not None
+        with_centripetal_shift = (
+            tl_centripetal_shift is not None
+            and br_centripetal_shift is not None)
+        assert with_embedding + with_centripetal_shift == 1
+        batch, _, height, width = tl_heat.size()
+        inp_h, inp_w, _ = img_meta['pad_shape']
+
+        # perform nms on heatmaps
+        tl_heat = self._local_maximum(tl_heat, kernel=kernel)
+        br_heat = self._local_maximum(br_heat, kernel=kernel)
+
+        tl_scores, tl_inds, tl_clses, tl_ys, tl_xs = self._topk(tl_heat, k=k)
+        br_scores, br_inds, br_clses, br_ys, br_xs = self._topk(br_heat, k=k)
+
+        # We use repeat instead of expand here because expand is a
+        # shallow-copy function. Thus it could cause unexpected testing result
+        # sometimes. Using expand will decrease about 10% mAP during testing
+        # compared to repeat.
+        tl_ys = tl_ys.view(batch, k, 1).repeat(1, 1, k)
+        tl_xs = tl_xs.view(batch, k, 1).repeat(1, 1, k)
+        br_ys = br_ys.view(batch, 1, k).repeat(1, k, 1)
+        br_xs = br_xs.view(batch, 1, k).repeat(1, k, 1)
+
+        tl_off = self._transpose_and_gather_feat(tl_off, tl_inds)
+        tl_off = tl_off.view(batch, k, 1, 2)
+        br_off = self._transpose_and_gather_feat(br_off, br_inds)
+        br_off = br_off.view(batch, 1, k, 2)
+
+        tl_xs = tl_xs + tl_off[..., 0]
+        tl_ys = tl_ys + tl_off[..., 1]
+        br_xs = br_xs + br_off[..., 0]
+        br_ys = br_ys + br_off[..., 1]
+
+        if with_centripetal_shift:
+            tl_centripetal_shift = self._transpose_and_gather_feat(
+                tl_centripetal_shift, tl_inds).view(batch, k, 1, 2).exp()
+            br_centripetal_shift = self._transpose_and_gather_feat(
+                br_centripetal_shift, br_inds).view(batch, 1, k, 2).exp()
+
+            tl_ctxs = tl_xs + tl_centripetal_shift[..., 0]
+            tl_ctys = tl_ys + tl_centripetal_shift[..., 1]
+            br_ctxs = br_xs - br_centripetal_shift[..., 0]
+            br_ctys = br_ys - br_centripetal_shift[..., 1]
+
+        # all possible boxes based on top k corners (ignoring class)
+        tl_xs *= (inp_w / width)
+        tl_ys *= (inp_h / height)
+        br_xs *= (inp_w / width)
+        br_ys *= (inp_h / height)
+
+        if with_centripetal_shift:
+            tl_ctxs *= (inp_w / width)
+            tl_ctys *= (inp_h / height)
+            br_ctxs *= (inp_w / width)
+            br_ctys *= (inp_h / height)
+
+        x_off = img_meta['border'][2]
+        y_off = img_meta['border'][0]
+
+        tl_xs -= x_off
+        tl_ys -= y_off
+        br_xs -= x_off
+        br_ys -= y_off
+
+        tl_xs *= tl_xs.gt(0.0).type_as(tl_xs)
+        tl_ys *= tl_ys.gt(0.0).type_as(tl_ys)
+        br_xs *= br_xs.gt(0.0).type_as(br_xs)
+        br_ys *= br_ys.gt(0.0).type_as(br_ys)
+
+        bboxes = torch.stack((tl_xs, tl_ys, br_xs, br_ys), dim=3)
+        area_bboxes = ((br_xs - tl_xs) * (br_ys - tl_ys)).abs()
+
+        if with_centripetal_shift:
+            tl_ctxs -= x_off
+            tl_ctys -= y_off
+            br_ctxs -= x_off
+            br_ctys -= y_off
+
+            tl_ctxs *= tl_ctxs.gt(0.0).type_as(tl_ctxs)
+            tl_ctys *= tl_ctys.gt(0.0).type_as(tl_ctys)
+            br_ctxs *= br_ctxs.gt(0.0).type_as(br_ctxs)
+            br_ctys *= br_ctys.gt(0.0).type_as(br_ctys)
+
+            ct_bboxes = torch.stack((tl_ctxs, tl_ctys, br_ctxs, br_ctys),
+                                    dim=3)
+            area_ct_bboxes = ((br_ctxs - tl_ctxs) * (br_ctys - tl_ctys)).abs()
+
+            rcentral = torch.zeros_like(ct_bboxes)
+            # magic nums from paper section 4.1
+            mu = torch.ones_like(area_bboxes) / 2.4
+            mu[area_bboxes > 3500] = 1 / 2.1  # large bbox have smaller mu
+
+            bboxes_center_x = (bboxes[..., 0] + bboxes[..., 2]) / 2
+            bboxes_center_y = (bboxes[..., 1] + bboxes[..., 3]) / 2
+            rcentral[..., 0] = bboxes_center_x - mu * (bboxes[..., 2] -
+                                                       bboxes[..., 0]) / 2
+            rcentral[..., 1] = bboxes_center_y - mu * (bboxes[..., 3] -
+                                                       bboxes[..., 1]) / 2
+            rcentral[..., 2] = bboxes_center_x + mu * (bboxes[..., 2] -
+                                                       bboxes[..., 0]) / 2
+            rcentral[..., 3] = bboxes_center_y + mu * (bboxes[..., 3] -
+                                                       bboxes[..., 1]) / 2
+            area_rcentral = ((rcentral[..., 2] - rcentral[..., 0]) *
+                             (rcentral[..., 3] - rcentral[..., 1])).abs()
+            dists = area_ct_bboxes / area_rcentral
+
+            tl_ctx_inds = (ct_bboxes[..., 0] <= rcentral[..., 0]) | (
+                ct_bboxes[..., 0] >= rcentral[..., 2])
+            tl_cty_inds = (ct_bboxes[..., 1] <= rcentral[..., 1]) | (
+                ct_bboxes[..., 1] >= rcentral[..., 3])
+            br_ctx_inds = (ct_bboxes[..., 2] <= rcentral[..., 0]) | (
+                ct_bboxes[..., 2] >= rcentral[..., 2])
+            br_cty_inds = (ct_bboxes[..., 3] <= rcentral[..., 1]) | (
+                ct_bboxes[..., 3] >= rcentral[..., 3])
+
+        if with_embedding:
+            tl_emb = self._transpose_and_gather_feat(tl_emb, tl_inds)
+            tl_emb = tl_emb.view(batch, k, 1)
+            br_emb = self._transpose_and_gather_feat(br_emb, br_inds)
+            br_emb = br_emb.view(batch, 1, k)
+            dists = torch.abs(tl_emb - br_emb)
+
+        tl_scores = tl_scores.view(batch, k, 1).repeat(1, 1, k)
+        br_scores = br_scores.view(batch, 1, k).repeat(1, k, 1)
+
+        scores = (tl_scores + br_scores) / 2  # scores for all possible boxes
+
+        # tl and br should have same class
+        tl_clses = tl_clses.view(batch, k, 1).repeat(1, 1, k)
+        br_clses = br_clses.view(batch, 1, k).repeat(1, k, 1)
+        cls_inds = (tl_clses != br_clses)
+
+        # reject boxes based on distances
+        dist_inds = dists > distance_threshold
+
+        # reject boxes based on widths and heights
+        width_inds = (br_xs <= tl_xs)
+        height_inds = (br_ys <= tl_ys)
+
+        scores[cls_inds] = -1
+        scores[width_inds] = -1
+        scores[height_inds] = -1
+        scores[dist_inds] = -1
+        if with_centripetal_shift:
+            scores[tl_ctx_inds] = -1
+            scores[tl_cty_inds] = -1
+            scores[br_ctx_inds] = -1
+            scores[br_cty_inds] = -1
+
+        scores = scores.view(batch, -1)
+        scores, inds = torch.topk(scores, num_dets)
+        scores = scores.unsqueeze(2)
+
+        bboxes = bboxes.view(batch, -1, 4)
+        bboxes = self._gather_feat(bboxes, inds)
+
+        clses = tl_clses.contiguous().view(batch, -1, 1)
+        clses = self._gather_feat(clses, inds).float()
+
+        return bboxes, scores, clses
diff --git a/detection/scrfd/mmdet/models/dense_heads/dense_test_mixins.py b/detection/scrfd/mmdet/models/dense_heads/dense_test_mixins.py
new file mode 100755
index 000000000..a07c9d423
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/dense_test_mixins.py
@@ -0,0 +1,97 @@
+from inspect import signature
+
+import torch
+
+from mmdet.core import bbox2result, bbox_mapping_back, multiclass_nms
+
+
+class BBoxTestMixin(object):
+    """Mixin class for test time augmentation of bboxes."""
+
+    def merge_aug_bboxes(self, aug_bboxes, aug_scores, img_metas):
+        """Merge augmented detection bboxes and scores.
+
+        Args:
+            aug_bboxes (list[Tensor]): shape (n, 4*#class)
+            aug_scores (list[Tensor] or None): shape (n, #class)
+            img_shapes (list[Tensor]): shape (3, ).
+
+        Returns:
+            tuple: (bboxes, scores)
+        """
+        recovered_bboxes = []
+        for bboxes, img_info in zip(aug_bboxes, img_metas):
+            img_shape = img_info[0]['img_shape']
+            scale_factor = img_info[0]['scale_factor']
+            flip = img_info[0]['flip']
+            flip_direction = img_info[0]['flip_direction']
+            bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip,
+                                       flip_direction)
+            recovered_bboxes.append(bboxes)
+        bboxes = torch.cat(recovered_bboxes, dim=0)
+        if aug_scores is None:
+            return bboxes
+        else:
+            scores = torch.cat(aug_scores, dim=0)
+            return bboxes, scores
+
+    def aug_test_bboxes(self, feats, img_metas, rescale=False):
+        """Test det bboxes with test time augmentation.
+
+        Args:
+            feats (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains features for all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[ndarray]: bbox results of each class
+        """
+        # check with_nms argument
+        gb_sig = signature(self.get_bboxes)
+        gb_args = [p.name for p in gb_sig.parameters.values()]
+        gbs_sig = signature(self._get_bboxes_single)
+        gbs_args = [p.name for p in gbs_sig.parameters.values()]
+        assert ('with_nms' in gb_args) and ('with_nms' in gbs_args), \
+            f'{self.__class__.__name__}' \
+            ' does not support test-time augmentation'
+
+        aug_bboxes = []
+        aug_scores = []
+        aug_factors = []  # score_factors for NMS
+        for x, img_meta in zip(feats, img_metas):
+            # only one image in the batch
+            outs = self.forward(x)
+            bbox_inputs = outs + (img_meta, self.test_cfg, False, False)
+            bbox_outputs = self.get_bboxes(*bbox_inputs)[0]
+            aug_bboxes.append(bbox_outputs[0])
+            aug_scores.append(bbox_outputs[1])
+            # bbox_outputs of some detectors (e.g., ATSS, FCOS, YOLOv3)
+            # contains additional element to adjust scores before NMS
+            if len(bbox_outputs) >= 3:
+                aug_factors.append(bbox_outputs[2])
+
+        # after merging, bboxes will be rescaled to the original image size
+        merged_bboxes, merged_scores = self.merge_aug_bboxes(
+            aug_bboxes, aug_scores, img_metas)
+        merged_factors = torch.cat(aug_factors, dim=0) if aug_factors else None
+        det_bboxes, det_labels = multiclass_nms(
+            merged_bboxes,
+            merged_scores,
+            self.test_cfg.score_thr,
+            self.test_cfg.nms,
+            self.test_cfg.max_per_img,
+            score_factors=merged_factors)
+
+        if rescale:
+            _det_bboxes = det_bboxes
+        else:
+            _det_bboxes = det_bboxes.clone()
+            _det_bboxes[:, :4] *= det_bboxes.new_tensor(
+                img_metas[0][0]['scale_factor'])
+        bbox_results = bbox2result(_det_bboxes, det_labels, self.num_classes)
+        return bbox_results
diff --git a/detection/scrfd/mmdet/models/dense_heads/fcos_head.py b/detection/scrfd/mmdet/models/dense_heads/fcos_head.py
new file mode 100755
index 000000000..51639c984
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/fcos_head.py
@@ -0,0 +1,574 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import Scale, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import distance2bbox, multi_apply, multiclass_nms
+from ..builder import HEADS, build_loss
+from .anchor_free_head import AnchorFreeHead
+
+INF = 1e8
+
+
+@HEADS.register_module()
+class FCOSHead(AnchorFreeHead):
+    """Anchor-free head used in `FCOS <https://arxiv.org/abs/1904.01355>`_.
+
+    The FCOS head does not use anchor boxes. Instead bounding boxes are
+    predicted at each pixel and a centerness measure is used to supress
+    low-quality predictions.
+    Here norm_on_bbox, centerness_on_reg, dcn_on_last_conv are training
+    tricks used in official repo, which will bring remarkable mAP gains
+    of up to 4.9. Please see https://github.com/tianzhi0549/FCOS for
+    more detail.
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        strides (list[int] | list[tuple[int, int]]): Strides of points
+            in multiple feature levels. Default: (4, 8, 16, 32, 64).
+        regress_ranges (tuple[tuple[int, int]]): Regress range of multiple
+            level points.
+        center_sampling (bool): If true, use center sampling. Default: False.
+        center_sample_radius (float): Radius of center sampling. Default: 1.5.
+        norm_on_bbox (bool): If true, normalize the regression targets
+            with FPN strides. Default: False.
+        centerness_on_reg (bool): If true, position centerness on the
+            regress branch. Please refer to https://github.com/tianzhi0549/FCOS/issues/89#issuecomment-516877042.
+            Default: False.
+        conv_bias (bool | str): If specified as `auto`, it will be decided by the
+            norm_cfg. Bias of conv will be set as True if `norm_cfg` is None, otherwise
+            False. Default: "auto".
+        loss_cls (dict): Config of classification loss.
+        loss_bbox (dict): Config of localization loss.
+        loss_centerness (dict): Config of centerness loss.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: norm_cfg=dict(type='GN', num_groups=32, requires_grad=True).
+
+    Example:
+        >>> self = FCOSHead(11, 7)
+        >>> feats = [torch.rand(1, 7, s, s) for s in [4, 8, 16, 32, 64]]
+        >>> cls_score, bbox_pred, centerness = self.forward(feats)
+        >>> assert len(cls_score) == len(self.scales)
+    """  # noqa: E501
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 regress_ranges=((-1, 64), (64, 128), (128, 256), (256, 512),
+                                 (512, INF)),
+                 center_sampling=False,
+                 center_sample_radius=1.5,
+                 norm_on_bbox=False,
+                 centerness_on_reg=False,
+                 loss_cls=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 loss_bbox=dict(type='IoULoss', loss_weight=1.0),
+                 loss_centerness=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+                 **kwargs):
+        self.regress_ranges = regress_ranges
+        self.center_sampling = center_sampling
+        self.center_sample_radius = center_sample_radius
+        self.norm_on_bbox = norm_on_bbox
+        self.centerness_on_reg = centerness_on_reg
+        super().__init__(
+            num_classes,
+            in_channels,
+            loss_cls=loss_cls,
+            loss_bbox=loss_bbox,
+            norm_cfg=norm_cfg,
+            **kwargs)
+        self.loss_centerness = build_loss(loss_centerness)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        super()._init_layers()
+        self.conv_centerness = nn.Conv2d(self.feat_channels, 1, 3, padding=1)
+        self.scales = nn.ModuleList([Scale(1.0) for _ in self.strides])
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        super().init_weights()
+        normal_init(self.conv_centerness, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple:
+                cls_scores (list[Tensor]): Box scores for each scale level, \
+                    each is a 4D-tensor, the channel number is \
+                    num_points * num_classes.
+                bbox_preds (list[Tensor]): Box energies / deltas for each \
+                    scale level, each is a 4D-tensor, the channel number is \
+                    num_points * 4.
+                centernesses (list[Tensor]): Centerss for each scale level, \
+                    each is a 4D-tensor, the channel number is num_points * 1.
+        """
+        return multi_apply(self.forward_single, feats, self.scales,
+                           self.strides)
+
+    def forward_single(self, x, scale, stride):
+        """Forward features of a single scale levle.
+
+        Args:
+            x (Tensor): FPN feature maps of the specified stride.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+            stride (int): The corresponding stride for feature maps, only
+                used to normalize the bbox prediction when self.norm_on_bbox
+                is True.
+
+        Returns:
+            tuple: scores for each class, bbox predictions and centerness \
+                predictions of input feature maps.
+        """
+        cls_score, bbox_pred, cls_feat, reg_feat = super().forward_single(x)
+        if self.centerness_on_reg:
+            centerness = self.conv_centerness(reg_feat)
+        else:
+            centerness = self.conv_centerness(cls_feat)
+        # scale the bbox_pred of different level
+        # float to avoid overflow when enabling FP16
+        bbox_pred = scale(bbox_pred).float()
+        if self.norm_on_bbox:
+            bbox_pred = F.relu(bbox_pred)
+            if not self.training:
+                bbox_pred *= stride
+        else:
+            bbox_pred = bbox_pred.exp()
+        return cls_score, bbox_pred, centerness
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'centernesses'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             centernesses,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute loss of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level,
+                each is a 4D-tensor, the channel number is
+                num_points * num_classes.
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level, each is a 4D-tensor, the channel number is
+                num_points * 4.
+            centernesses (list[Tensor]): Centerss for each scale level, each
+                is a 4D-tensor, the channel number is num_points * 1.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert len(cls_scores) == len(bbox_preds) == len(centernesses)
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        all_level_points = self.get_points(featmap_sizes, bbox_preds[0].dtype,
+                                           bbox_preds[0].device)
+        labels, bbox_targets = self.get_targets(all_level_points, gt_bboxes,
+                                                gt_labels)
+
+        num_imgs = cls_scores[0].size(0)
+        # flatten cls_scores, bbox_preds and centerness
+        flatten_cls_scores = [
+            cls_score.permute(0, 2, 3, 1).reshape(-1, self.cls_out_channels)
+            for cls_score in cls_scores
+        ]
+        flatten_bbox_preds = [
+            bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)
+            for bbox_pred in bbox_preds
+        ]
+        flatten_centerness = [
+            centerness.permute(0, 2, 3, 1).reshape(-1)
+            for centerness in centernesses
+        ]
+        flatten_cls_scores = torch.cat(flatten_cls_scores)
+        flatten_bbox_preds = torch.cat(flatten_bbox_preds)
+        flatten_centerness = torch.cat(flatten_centerness)
+        flatten_labels = torch.cat(labels)
+        flatten_bbox_targets = torch.cat(bbox_targets)
+        # repeat points to align with bbox_preds
+        flatten_points = torch.cat(
+            [points.repeat(num_imgs, 1) for points in all_level_points])
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((flatten_labels >= 0)
+                    & (flatten_labels < bg_class_ind)).nonzero().reshape(-1)
+        num_pos = len(pos_inds)
+        loss_cls = self.loss_cls(
+            flatten_cls_scores, flatten_labels,
+            avg_factor=num_pos + num_imgs)  # avoid num_pos is 0
+
+        pos_bbox_preds = flatten_bbox_preds[pos_inds]
+        pos_centerness = flatten_centerness[pos_inds]
+
+        if num_pos > 0:
+            pos_bbox_targets = flatten_bbox_targets[pos_inds]
+            pos_centerness_targets = self.centerness_target(pos_bbox_targets)
+            pos_points = flatten_points[pos_inds]
+            pos_decoded_bbox_preds = distance2bbox(pos_points, pos_bbox_preds)
+            pos_decoded_target_preds = distance2bbox(pos_points,
+                                                     pos_bbox_targets)
+            # centerness weighted iou loss
+            loss_bbox = self.loss_bbox(
+                pos_decoded_bbox_preds,
+                pos_decoded_target_preds,
+                weight=pos_centerness_targets,
+                avg_factor=pos_centerness_targets.sum())
+            loss_centerness = self.loss_centerness(pos_centerness,
+                                                   pos_centerness_targets)
+        else:
+            loss_bbox = pos_bbox_preds.sum()
+            loss_centerness = pos_centerness.sum()
+
+        return dict(
+            loss_cls=loss_cls,
+            loss_bbox=loss_bbox,
+            loss_centerness=loss_centerness)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'centernesses'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   centernesses,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                with shape (N, num_points * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_points * 4, H, W).
+            centernesses (list[Tensor]): Centerness for each scale level with
+                shape (N, num_points * 1, H, W).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used. Default: None.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class label of the
+                corresponding box.
+        """
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        mlvl_points = self.get_points(featmap_sizes, bbox_preds[0].dtype,
+                                      bbox_preds[0].device)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            centerness_pred_list = [
+                centernesses[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            det_bboxes = self._get_bboxes_single(
+                cls_score_list, bbox_pred_list, centerness_pred_list,
+                mlvl_points, img_shape, scale_factor, cfg, rescale, with_nms)
+            result_list.append(det_bboxes)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           centernesses,
+                           mlvl_points,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for a single scale level
+                with shape (num_points * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for a single scale
+                level with shape (num_points * 4, H, W).
+            centernesses (list[Tensor]): Centerness for a single scale level
+                with shape (num_points * 4, H, W).
+            mlvl_points (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_points, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arrange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): BBox predictions in shape (n, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (n,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_points)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        mlvl_centerness = []
+        for cls_score, bbox_pred, centerness, points in zip(
+                cls_scores, bbox_preds, centernesses, mlvl_points):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            centerness = centerness.permute(1, 2, 0).reshape(-1).sigmoid()
+
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                max_scores, _ = (scores * centerness[:, None]).max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                points = points[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+                centerness = centerness[topk_inds]
+            bboxes = distance2bbox(points, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+            mlvl_centerness.append(centerness)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        mlvl_centerness = torch.cat(mlvl_centerness)
+
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(
+                mlvl_bboxes,
+                mlvl_scores,
+                cfg.score_thr,
+                cfg.nms,
+                cfg.max_per_img,
+                score_factors=mlvl_centerness)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores, mlvl_centerness
+
+    def _get_points_single(self,
+                           featmap_size,
+                           stride,
+                           dtype,
+                           device,
+                           flatten=False):
+        """Get points according to feature map sizes."""
+        y, x = super()._get_points_single(featmap_size, stride, dtype, device)
+        points = torch.stack((x.reshape(-1) * stride, y.reshape(-1) * stride),
+                             dim=-1) + stride // 2
+        return points
+
+    def get_targets(self, points, gt_bboxes_list, gt_labels_list):
+        """Compute regression, classification and centerss targets for points
+        in multiple images.
+
+        Args:
+            points (list[Tensor]): Points of each fpn level, each has shape
+                (num_points, 2).
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image,
+                each has shape (num_gt, 4).
+            gt_labels_list (list[Tensor]): Ground truth labels of each box,
+                each has shape (num_gt,).
+
+        Returns:
+            tuple:
+                concat_lvl_labels (list[Tensor]): Labels of each level. \
+                concat_lvl_bbox_targets (list[Tensor]): BBox targets of each \
+                    level.
+        """
+        assert len(points) == len(self.regress_ranges)
+        num_levels = len(points)
+        # expand regress ranges to align with points
+        expanded_regress_ranges = [
+            points[i].new_tensor(self.regress_ranges[i])[None].expand_as(
+                points[i]) for i in range(num_levels)
+        ]
+        # concat all levels points and regress ranges
+        concat_regress_ranges = torch.cat(expanded_regress_ranges, dim=0)
+        concat_points = torch.cat(points, dim=0)
+
+        # the number of points per img, per lvl
+        num_points = [center.size(0) for center in points]
+
+        # get labels and bbox_targets of each image
+        labels_list, bbox_targets_list = multi_apply(
+            self._get_target_single,
+            gt_bboxes_list,
+            gt_labels_list,
+            points=concat_points,
+            regress_ranges=concat_regress_ranges,
+            num_points_per_lvl=num_points)
+
+        # split to per img, per level
+        labels_list = [labels.split(num_points, 0) for labels in labels_list]
+        bbox_targets_list = [
+            bbox_targets.split(num_points, 0)
+            for bbox_targets in bbox_targets_list
+        ]
+
+        # concat per level image
+        concat_lvl_labels = []
+        concat_lvl_bbox_targets = []
+        for i in range(num_levels):
+            concat_lvl_labels.append(
+                torch.cat([labels[i] for labels in labels_list]))
+            bbox_targets = torch.cat(
+                [bbox_targets[i] for bbox_targets in bbox_targets_list])
+            if self.norm_on_bbox:
+                bbox_targets = bbox_targets / self.strides[i]
+            concat_lvl_bbox_targets.append(bbox_targets)
+        return concat_lvl_labels, concat_lvl_bbox_targets
+
+    def _get_target_single(self, gt_bboxes, gt_labels, points, regress_ranges,
+                           num_points_per_lvl):
+        """Compute regression and classification targets for a single image."""
+        num_points = points.size(0)
+        num_gts = gt_labels.size(0)
+        if num_gts == 0:
+            return gt_labels.new_full((num_points,), self.num_classes), \
+                   gt_bboxes.new_zeros((num_points, 4))
+
+        areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0]) * (
+            gt_bboxes[:, 3] - gt_bboxes[:, 1])
+        # TODO: figure out why these two are different
+        # areas = areas[None].expand(num_points, num_gts)
+        areas = areas[None].repeat(num_points, 1)
+        regress_ranges = regress_ranges[:, None, :].expand(
+            num_points, num_gts, 2)
+        gt_bboxes = gt_bboxes[None].expand(num_points, num_gts, 4)
+        xs, ys = points[:, 0], points[:, 1]
+        xs = xs[:, None].expand(num_points, num_gts)
+        ys = ys[:, None].expand(num_points, num_gts)
+
+        left = xs - gt_bboxes[..., 0]
+        right = gt_bboxes[..., 2] - xs
+        top = ys - gt_bboxes[..., 1]
+        bottom = gt_bboxes[..., 3] - ys
+        bbox_targets = torch.stack((left, top, right, bottom), -1)
+
+        if self.center_sampling:
+            # condition1: inside a `center bbox`
+            radius = self.center_sample_radius
+            center_xs = (gt_bboxes[..., 0] + gt_bboxes[..., 2]) / 2
+            center_ys = (gt_bboxes[..., 1] + gt_bboxes[..., 3]) / 2
+            center_gts = torch.zeros_like(gt_bboxes)
+            stride = center_xs.new_zeros(center_xs.shape)
+
+            # project the points on current lvl back to the `original` sizes
+            lvl_begin = 0
+            for lvl_idx, num_points_lvl in enumerate(num_points_per_lvl):
+                lvl_end = lvl_begin + num_points_lvl
+                stride[lvl_begin:lvl_end] = self.strides[lvl_idx] * radius
+                lvl_begin = lvl_end
+
+            x_mins = center_xs - stride
+            y_mins = center_ys - stride
+            x_maxs = center_xs + stride
+            y_maxs = center_ys + stride
+            center_gts[..., 0] = torch.where(x_mins > gt_bboxes[..., 0],
+                                             x_mins, gt_bboxes[..., 0])
+            center_gts[..., 1] = torch.where(y_mins > gt_bboxes[..., 1],
+                                             y_mins, gt_bboxes[..., 1])
+            center_gts[..., 2] = torch.where(x_maxs > gt_bboxes[..., 2],
+                                             gt_bboxes[..., 2], x_maxs)
+            center_gts[..., 3] = torch.where(y_maxs > gt_bboxes[..., 3],
+                                             gt_bboxes[..., 3], y_maxs)
+
+            cb_dist_left = xs - center_gts[..., 0]
+            cb_dist_right = center_gts[..., 2] - xs
+            cb_dist_top = ys - center_gts[..., 1]
+            cb_dist_bottom = center_gts[..., 3] - ys
+            center_bbox = torch.stack(
+                (cb_dist_left, cb_dist_top, cb_dist_right, cb_dist_bottom), -1)
+            inside_gt_bbox_mask = center_bbox.min(-1)[0] > 0
+        else:
+            # condition1: inside a gt bbox
+            inside_gt_bbox_mask = bbox_targets.min(-1)[0] > 0
+
+        # condition2: limit the regression range for each location
+        max_regress_distance = bbox_targets.max(-1)[0]
+        inside_regress_range = (
+            (max_regress_distance >= regress_ranges[..., 0])
+            & (max_regress_distance <= regress_ranges[..., 1]))
+
+        # if there are still more than one objects for a location,
+        # we choose the one with minimal area
+        areas[inside_gt_bbox_mask == 0] = INF
+        areas[inside_regress_range == 0] = INF
+        min_area, min_area_inds = areas.min(dim=1)
+
+        labels = gt_labels[min_area_inds]
+        labels[min_area == INF] = self.num_classes  # set as BG
+        bbox_targets = bbox_targets[range(num_points), min_area_inds]
+
+        return labels, bbox_targets
+
+    def centerness_target(self, pos_bbox_targets):
+        """Compute centerness targets.
+
+        Args:
+            pos_bbox_targets (Tensor): BBox targets of positive bboxes in shape
+                (num_pos, 4)
+
+        Returns:
+            Tensor: Centerness target.
+        """
+        # only calculate pos centerness targets, otherwise there may be nan
+        left_right = pos_bbox_targets[:, [0, 2]]
+        top_bottom = pos_bbox_targets[:, [1, 3]]
+        centerness_targets = (
+            left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0]) * (
+                top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0])
+        return torch.sqrt(centerness_targets)
diff --git a/detection/scrfd/mmdet/models/dense_heads/fovea_head.py b/detection/scrfd/mmdet/models/dense_heads/fovea_head.py
new file mode 100755
index 000000000..c8ccea787
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/fovea_head.py
@@ -0,0 +1,341 @@
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, normal_init
+from mmcv.ops import DeformConv2d
+
+from mmdet.core import multi_apply, multiclass_nms
+from ..builder import HEADS
+from .anchor_free_head import AnchorFreeHead
+
+INF = 1e8
+
+
+class FeatureAlign(nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 deform_groups=4):
+        super(FeatureAlign, self).__init__()
+        offset_channels = kernel_size * kernel_size * 2
+        self.conv_offset = nn.Conv2d(
+            4, deform_groups * offset_channels, 1, bias=False)
+        self.conv_adaption = DeformConv2d(
+            in_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            padding=(kernel_size - 1) // 2,
+            deform_groups=deform_groups)
+        self.relu = nn.ReLU(inplace=True)
+
+    def init_weights(self):
+        normal_init(self.conv_offset, std=0.1)
+        normal_init(self.conv_adaption, std=0.01)
+
+    def forward(self, x, shape):
+        offset = self.conv_offset(shape)
+        x = self.relu(self.conv_adaption(x, offset))
+        return x
+
+
+@HEADS.register_module()
+class FoveaHead(AnchorFreeHead):
+    """FoveaBox: Beyond Anchor-based Object Detector
+    https://arxiv.org/abs/1904.03797
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 base_edge_list=(16, 32, 64, 128, 256),
+                 scale_ranges=((8, 32), (16, 64), (32, 128), (64, 256), (128,
+                                                                         512)),
+                 sigma=0.4,
+                 with_deform=False,
+                 deform_groups=4,
+                 **kwargs):
+        self.base_edge_list = base_edge_list
+        self.scale_ranges = scale_ranges
+        self.sigma = sigma
+        self.with_deform = with_deform
+        self.deform_groups = deform_groups
+        super().__init__(num_classes, in_channels, **kwargs)
+
+    def _init_layers(self):
+        # box branch
+        super()._init_reg_convs()
+        self.conv_reg = nn.Conv2d(self.feat_channels, 4, 3, padding=1)
+
+        # cls branch
+        if not self.with_deform:
+            super()._init_cls_convs()
+            self.conv_cls = nn.Conv2d(
+                self.feat_channels, self.cls_out_channels, 3, padding=1)
+        else:
+            self.cls_convs = nn.ModuleList()
+            self.cls_convs.append(
+                ConvModule(
+                    self.feat_channels, (self.feat_channels * 4),
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg,
+                    bias=self.norm_cfg is None))
+            self.cls_convs.append(
+                ConvModule((self.feat_channels * 4), (self.feat_channels * 4),
+                           1,
+                           stride=1,
+                           padding=0,
+                           conv_cfg=self.conv_cfg,
+                           norm_cfg=self.norm_cfg,
+                           bias=self.norm_cfg is None))
+            self.feature_adaption = FeatureAlign(
+                self.feat_channels,
+                self.feat_channels,
+                kernel_size=3,
+                deform_groups=self.deform_groups)
+            self.conv_cls = nn.Conv2d(
+                int(self.feat_channels * 4),
+                self.cls_out_channels,
+                3,
+                padding=1)
+
+    def init_weights(self):
+        super().init_weights()
+        if self.with_deform:
+            self.feature_adaption.init_weights()
+
+    def forward_single(self, x):
+        cls_feat = x
+        reg_feat = x
+        for reg_layer in self.reg_convs:
+            reg_feat = reg_layer(reg_feat)
+        bbox_pred = self.conv_reg(reg_feat)
+        if self.with_deform:
+            cls_feat = self.feature_adaption(cls_feat, bbox_pred.exp())
+        for cls_layer in self.cls_convs:
+            cls_feat = cls_layer(cls_feat)
+        cls_score = self.conv_cls(cls_feat)
+        return cls_score, bbox_pred
+
+    def _get_points_single(self, *args, **kwargs):
+        y, x = super()._get_points_single(*args, **kwargs)
+        return y + 0.5, x + 0.5
+
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bbox_list,
+             gt_label_list,
+             img_metas,
+             gt_bboxes_ignore=None):
+        assert len(cls_scores) == len(bbox_preds)
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        points = self.get_points(featmap_sizes, bbox_preds[0].dtype,
+                                 bbox_preds[0].device)
+        num_imgs = cls_scores[0].size(0)
+        flatten_cls_scores = [
+            cls_score.permute(0, 2, 3, 1).reshape(-1, self.cls_out_channels)
+            for cls_score in cls_scores
+        ]
+        flatten_bbox_preds = [
+            bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)
+            for bbox_pred in bbox_preds
+        ]
+        flatten_cls_scores = torch.cat(flatten_cls_scores)
+        flatten_bbox_preds = torch.cat(flatten_bbox_preds)
+        flatten_labels, flatten_bbox_targets = self.get_targets(
+            gt_bbox_list, gt_label_list, featmap_sizes, points)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        pos_inds = ((flatten_labels >= 0)
+                    & (flatten_labels < self.num_classes)).nonzero().view(-1)
+        num_pos = len(pos_inds)
+
+        loss_cls = self.loss_cls(
+            flatten_cls_scores, flatten_labels, avg_factor=num_pos + num_imgs)
+        if num_pos > 0:
+            pos_bbox_preds = flatten_bbox_preds[pos_inds]
+            pos_bbox_targets = flatten_bbox_targets[pos_inds]
+            pos_weights = pos_bbox_targets.new_zeros(
+                pos_bbox_targets.size()) + 1.0
+            loss_bbox = self.loss_bbox(
+                pos_bbox_preds,
+                pos_bbox_targets,
+                pos_weights,
+                avg_factor=num_pos)
+        else:
+            loss_bbox = torch.tensor(
+                0,
+                dtype=flatten_bbox_preds.dtype,
+                device=flatten_bbox_preds.device)
+        return dict(loss_cls=loss_cls, loss_bbox=loss_bbox)
+
+    def get_targets(self, gt_bbox_list, gt_label_list, featmap_sizes, points):
+        label_list, bbox_target_list = multi_apply(
+            self._get_target_single,
+            gt_bbox_list,
+            gt_label_list,
+            featmap_size_list=featmap_sizes,
+            point_list=points)
+        flatten_labels = [
+            torch.cat([
+                labels_level_img.flatten() for labels_level_img in labels_level
+            ]) for labels_level in zip(*label_list)
+        ]
+        flatten_bbox_targets = [
+            torch.cat([
+                bbox_targets_level_img.reshape(-1, 4)
+                for bbox_targets_level_img in bbox_targets_level
+            ]) for bbox_targets_level in zip(*bbox_target_list)
+        ]
+        flatten_labels = torch.cat(flatten_labels)
+        flatten_bbox_targets = torch.cat(flatten_bbox_targets)
+        return flatten_labels, flatten_bbox_targets
+
+    def _get_target_single(self,
+                           gt_bboxes_raw,
+                           gt_labels_raw,
+                           featmap_size_list=None,
+                           point_list=None):
+
+        gt_areas = torch.sqrt((gt_bboxes_raw[:, 2] - gt_bboxes_raw[:, 0]) *
+                              (gt_bboxes_raw[:, 3] - gt_bboxes_raw[:, 1]))
+        label_list = []
+        bbox_target_list = []
+        # for each pyramid, find the cls and box target
+        for base_len, (lower_bound, upper_bound), stride, featmap_size, \
+            (y, x) in zip(self.base_edge_list, self.scale_ranges,
+                          self.strides, featmap_size_list, point_list):
+            # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+            labels = gt_labels_raw.new_zeros(featmap_size) + self.num_classes
+            bbox_targets = gt_bboxes_raw.new(featmap_size[0], featmap_size[1],
+                                             4) + 1
+            # scale assignment
+            hit_indices = ((gt_areas >= lower_bound) &
+                           (gt_areas <= upper_bound)).nonzero().flatten()
+            if len(hit_indices) == 0:
+                label_list.append(labels)
+                bbox_target_list.append(torch.log(bbox_targets))
+                continue
+            _, hit_index_order = torch.sort(-gt_areas[hit_indices])
+            hit_indices = hit_indices[hit_index_order]
+            gt_bboxes = gt_bboxes_raw[hit_indices, :] / stride
+            gt_labels = gt_labels_raw[hit_indices]
+            half_w = 0.5 * (gt_bboxes[:, 2] - gt_bboxes[:, 0])
+            half_h = 0.5 * (gt_bboxes[:, 3] - gt_bboxes[:, 1])
+            # valid fovea area: left, right, top, down
+            pos_left = torch.ceil(
+                gt_bboxes[:, 0] + (1 - self.sigma) * half_w - 0.5).long().\
+                clamp(0, featmap_size[1] - 1)
+            pos_right = torch.floor(
+                gt_bboxes[:, 0] + (1 + self.sigma) * half_w - 0.5).long().\
+                clamp(0, featmap_size[1] - 1)
+            pos_top = torch.ceil(
+                gt_bboxes[:, 1] + (1 - self.sigma) * half_h - 0.5).long().\
+                clamp(0, featmap_size[0] - 1)
+            pos_down = torch.floor(
+                gt_bboxes[:, 1] + (1 + self.sigma) * half_h - 0.5).long().\
+                clamp(0, featmap_size[0] - 1)
+            for px1, py1, px2, py2, label, (gt_x1, gt_y1, gt_x2, gt_y2) in \
+                    zip(pos_left, pos_top, pos_right, pos_down, gt_labels,
+                        gt_bboxes_raw[hit_indices, :]):
+                labels[py1:py2 + 1, px1:px2 + 1] = label
+                bbox_targets[py1:py2 + 1, px1:px2 + 1, 0] = \
+                    (stride * x[py1:py2 + 1, px1:px2 + 1] - gt_x1) / base_len
+                bbox_targets[py1:py2 + 1, px1:px2 + 1, 1] = \
+                    (stride * y[py1:py2 + 1, px1:px2 + 1] - gt_y1) / base_len
+                bbox_targets[py1:py2 + 1, px1:px2 + 1, 2] = \
+                    (gt_x2 - stride * x[py1:py2 + 1, px1:px2 + 1]) / base_len
+                bbox_targets[py1:py2 + 1, px1:px2 + 1, 3] = \
+                    (gt_y2 - stride * y[py1:py2 + 1, px1:px2 + 1]) / base_len
+            bbox_targets = bbox_targets.clamp(min=1. / 16, max=16.)
+            label_list.append(labels)
+            bbox_target_list.append(torch.log(bbox_targets))
+        return label_list, bbox_target_list
+
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=None):
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        points = self.get_points(
+            featmap_sizes,
+            bbox_preds[0].dtype,
+            bbox_preds[0].device,
+            flatten=True)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            det_bboxes = self._get_bboxes_single(cls_score_list,
+                                                 bbox_pred_list, featmap_sizes,
+                                                 points, img_shape,
+                                                 scale_factor, cfg, rescale)
+            result_list.append(det_bboxes)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           featmap_sizes,
+                           point_list,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False):
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(point_list)
+        det_bboxes = []
+        det_scores = []
+        for cls_score, bbox_pred, featmap_size, stride, base_len, (y, x) \
+                in zip(cls_scores, bbox_preds, featmap_sizes, self.strides,
+                       self.base_edge_list, point_list):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4).exp()
+            nms_pre = cfg.get('nms_pre', -1)
+            if (nms_pre > 0) and (scores.shape[0] > nms_pre):
+                max_scores, _ = scores.max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+                y = y[topk_inds]
+                x = x[topk_inds]
+            x1 = (stride * x - base_len * bbox_pred[:, 0]).\
+                clamp(min=0, max=img_shape[1] - 1)
+            y1 = (stride * y - base_len * bbox_pred[:, 1]).\
+                clamp(min=0, max=img_shape[0] - 1)
+            x2 = (stride * x + base_len * bbox_pred[:, 2]).\
+                clamp(min=0, max=img_shape[1] - 1)
+            y2 = (stride * y + base_len * bbox_pred[:, 3]).\
+                clamp(min=0, max=img_shape[0] - 1)
+            bboxes = torch.stack([x1, y1, x2, y2], -1)
+            det_bboxes.append(bboxes)
+            det_scores.append(scores)
+        det_bboxes = torch.cat(det_bboxes)
+        if rescale:
+            det_bboxes /= det_bboxes.new_tensor(scale_factor)
+        det_scores = torch.cat(det_scores)
+        padding = det_scores.new_zeros(det_scores.shape[0], 1)
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        det_scores = torch.cat([det_scores, padding], dim=1)
+        det_bboxes, det_labels = multiclass_nms(det_bboxes, det_scores,
+                                                cfg.score_thr, cfg.nms,
+                                                cfg.max_per_img)
+        return det_bboxes, det_labels
diff --git a/detection/scrfd/mmdet/models/dense_heads/free_anchor_retina_head.py b/detection/scrfd/mmdet/models/dense_heads/free_anchor_retina_head.py
new file mode 100755
index 000000000..79879fdc3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/free_anchor_retina_head.py
@@ -0,0 +1,270 @@
+import torch
+import torch.nn.functional as F
+
+from mmdet.core import bbox_overlaps
+from ..builder import HEADS
+from .retina_head import RetinaHead
+
+EPS = 1e-12
+
+
+@HEADS.register_module()
+class FreeAnchorRetinaHead(RetinaHead):
+    """FreeAnchor RetinaHead used in https://arxiv.org/abs/1909.02466.
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        stacked_convs (int): Number of conv layers in cls and reg tower.
+            Default: 4.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: norm_cfg=dict(type='GN', num_groups=32,
+            requires_grad=True).
+        pre_anchor_topk (int): Number of boxes that be token in each bag.
+        bbox_thr (float): The threshold of the saturated linear function. It is
+            usually the same with the IoU threshold used in NMS.
+        gamma (float): Gamma parameter in focal loss.
+        alpha (float): Alpha parameter in focal loss.
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 pre_anchor_topk=50,
+                 bbox_thr=0.6,
+                 gamma=2.0,
+                 alpha=0.5,
+                 **kwargs):
+        super(FreeAnchorRetinaHead,
+              self).__init__(num_classes, in_channels, stacked_convs, conv_cfg,
+                             norm_cfg, **kwargs)
+
+        self.pre_anchor_topk = pre_anchor_topk
+        self.bbox_thr = bbox_thr
+        self.gamma = gamma
+        self.alpha = alpha
+
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == len(self.anchor_generator.base_anchors)
+
+        anchor_list, _ = self.get_anchors(featmap_sizes, img_metas)
+        anchors = [torch.cat(anchor) for anchor in anchor_list]
+
+        # concatenate each level
+        cls_scores = [
+            cls.permute(0, 2, 3,
+                        1).reshape(cls.size(0), -1, self.cls_out_channels)
+            for cls in cls_scores
+        ]
+        bbox_preds = [
+            bbox_pred.permute(0, 2, 3, 1).reshape(bbox_pred.size(0), -1, 4)
+            for bbox_pred in bbox_preds
+        ]
+        cls_scores = torch.cat(cls_scores, dim=1)
+        bbox_preds = torch.cat(bbox_preds, dim=1)
+
+        cls_prob = torch.sigmoid(cls_scores)
+        box_prob = []
+        num_pos = 0
+        positive_losses = []
+        for _, (anchors_, gt_labels_, gt_bboxes_, cls_prob_,
+                bbox_preds_) in enumerate(
+                    zip(anchors, gt_labels, gt_bboxes, cls_prob, bbox_preds)):
+
+            with torch.no_grad():
+                if len(gt_bboxes_) == 0:
+                    image_box_prob = torch.zeros(
+                        anchors_.size(0),
+                        self.cls_out_channels).type_as(bbox_preds_)
+                else:
+                    # box_localization: a_{j}^{loc}, shape: [j, 4]
+                    pred_boxes = self.bbox_coder.decode(anchors_, bbox_preds_)
+
+                    # object_box_iou: IoU_{ij}^{loc}, shape: [i, j]
+                    object_box_iou = bbox_overlaps(gt_bboxes_, pred_boxes)
+
+                    # object_box_prob: P{a_{j} -> b_{i}}, shape: [i, j]
+                    t1 = self.bbox_thr
+                    t2 = object_box_iou.max(
+                        dim=1, keepdim=True).values.clamp(min=t1 + 1e-12)
+                    object_box_prob = ((object_box_iou - t1) /
+                                       (t2 - t1)).clamp(
+                                           min=0, max=1)
+
+                    # object_cls_box_prob: P{a_{j} -> b_{i}}, shape: [i, c, j]
+                    num_obj = gt_labels_.size(0)
+                    indices = torch.stack([
+                        torch.arange(num_obj).type_as(gt_labels_), gt_labels_
+                    ],
+                                          dim=0)
+                    object_cls_box_prob = torch.sparse_coo_tensor(
+                        indices, object_box_prob)
+
+                    # image_box_iou: P{a_{j} \in A_{+}}, shape: [c, j]
+                    """
+                    from "start" to "end" implement:
+                    image_box_iou = torch.sparse.max(object_cls_box_prob,
+                                                     dim=0).t()
+
+                    """
+                    # start
+                    box_cls_prob = torch.sparse.sum(
+                        object_cls_box_prob, dim=0).to_dense()
+
+                    indices = torch.nonzero(box_cls_prob, as_tuple=False).t_()
+                    if indices.numel() == 0:
+                        image_box_prob = torch.zeros(
+                            anchors_.size(0),
+                            self.cls_out_channels).type_as(object_box_prob)
+                    else:
+                        nonzero_box_prob = torch.where(
+                            (gt_labels_.unsqueeze(dim=-1) == indices[0]),
+                            object_box_prob[:, indices[1]],
+                            torch.tensor([
+                                0
+                            ]).type_as(object_box_prob)).max(dim=0).values
+
+                        # upmap to shape [j, c]
+                        image_box_prob = torch.sparse_coo_tensor(
+                            indices.flip([0]),
+                            nonzero_box_prob,
+                            size=(anchors_.size(0),
+                                  self.cls_out_channels)).to_dense()
+                    # end
+
+                box_prob.append(image_box_prob)
+
+            # construct bags for objects
+            match_quality_matrix = bbox_overlaps(gt_bboxes_, anchors_)
+            _, matched = torch.topk(
+                match_quality_matrix,
+                self.pre_anchor_topk,
+                dim=1,
+                sorted=False)
+            del match_quality_matrix
+
+            # matched_cls_prob: P_{ij}^{cls}
+            matched_cls_prob = torch.gather(
+                cls_prob_[matched], 2,
+                gt_labels_.view(-1, 1, 1).repeat(1, self.pre_anchor_topk,
+                                                 1)).squeeze(2)
+
+            # matched_box_prob: P_{ij}^{loc}
+            matched_anchors = anchors_[matched]
+            matched_object_targets = self.bbox_coder.encode(
+                matched_anchors,
+                gt_bboxes_.unsqueeze(dim=1).expand_as(matched_anchors))
+            loss_bbox = self.loss_bbox(
+                bbox_preds_[matched],
+                matched_object_targets,
+                reduction_override='none').sum(-1)
+            matched_box_prob = torch.exp(-loss_bbox)
+
+            # positive_losses: {-log( Mean-max(P_{ij}^{cls} * P_{ij}^{loc}) )}
+            num_pos += len(gt_bboxes_)
+            positive_losses.append(
+                self.positive_bag_loss(matched_cls_prob, matched_box_prob))
+        positive_loss = torch.cat(positive_losses).sum() / max(1, num_pos)
+
+        # box_prob: P{a_{j} \in A_{+}}
+        box_prob = torch.stack(box_prob, dim=0)
+
+        # negative_loss:
+        # \sum_{j}{ FL((1 - P{a_{j} \in A_{+}}) * (1 - P_{j}^{bg})) } / n||B||
+        negative_loss = self.negative_bag_loss(cls_prob, box_prob).sum() / max(
+            1, num_pos * self.pre_anchor_topk)
+
+        # avoid the absence of gradients in regression subnet
+        # when no ground-truth in a batch
+        if num_pos == 0:
+            positive_loss = bbox_preds.sum() * 0
+
+        losses = {
+            'positive_bag_loss': positive_loss,
+            'negative_bag_loss': negative_loss
+        }
+        return losses
+
+    def positive_bag_loss(self, matched_cls_prob, matched_box_prob):
+        """Compute positive bag loss.
+
+        :math:`-log( Mean-max(P_{ij}^{cls} * P_{ij}^{loc}) )`.
+
+        :math:`P_{ij}^{cls}`: matched_cls_prob, classification probability of matched samples.
+
+        :math:`P_{ij}^{loc}`: matched_box_prob, box probability of matched samples.
+
+        Args:
+            matched_cls_prob (Tensor): Classification probabilty of matched
+                samples in shape (num_gt, pre_anchor_topk).
+            matched_box_prob (Tensor): BBox probability of matched samples,
+                in shape (num_gt, pre_anchor_topk).
+
+        Returns:
+            Tensor: Positive bag loss in shape (num_gt,).
+        """  # noqa: E501, W605
+        # bag_prob = Mean-max(matched_prob)
+        matched_prob = matched_cls_prob * matched_box_prob
+        weight = 1 / torch.clamp(1 - matched_prob, 1e-12, None)
+        weight /= weight.sum(dim=1).unsqueeze(dim=-1)
+        bag_prob = (weight * matched_prob).sum(dim=1)
+        # positive_bag_loss = -self.alpha * log(bag_prob)
+        return self.alpha * F.binary_cross_entropy(
+            bag_prob, torch.ones_like(bag_prob), reduction='none')
+
+    def negative_bag_loss(self, cls_prob, box_prob):
+        """Compute negative bag loss.
+
+        :math:`FL((1 - P_{a_{j} \in A_{+}}) * (1 - P_{j}^{bg}))`.
+
+        :math:`P_{a_{j} \in A_{+}}`: Box_probability of matched samples.
+
+        :math:`P_{j}^{bg}`: Classification probability of negative samples.
+
+        Args:
+            cls_prob (Tensor): Classification probability, in shape
+                (num_img, num_anchors, num_classes).
+            box_prob (Tensor): Box probability, in shape
+                (num_img, num_anchors, num_classes).
+
+        Returns:
+            Tensor: Negative bag loss in shape (num_img, num_anchors, num_classes).
+        """  # noqa: E501, W605
+        prob = cls_prob * (1 - box_prob)
+        # There are some cases when neg_prob = 0.
+        # This will cause the neg_prob.log() to be inf without clamp.
+        prob = prob.clamp(min=EPS, max=1 - EPS)
+        negative_bag_loss = prob**self.gamma * F.binary_cross_entropy(
+            prob, torch.zeros_like(prob), reduction='none')
+        return (1 - self.alpha) * negative_bag_loss
diff --git a/detection/scrfd/mmdet/models/dense_heads/fsaf_head.py b/detection/scrfd/mmdet/models/dense_heads/fsaf_head.py
new file mode 100755
index 000000000..c23e3699f
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/fsaf_head.py
@@ -0,0 +1,418 @@
+import numpy as np
+import torch
+from mmcv.cnn import normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, images_to_levels, multi_apply,
+                        unmap)
+from ..builder import HEADS
+from ..losses.accuracy import accuracy
+from ..losses.utils import weight_reduce_loss
+from .retina_head import RetinaHead
+
+
+@HEADS.register_module()
+class FSAFHead(RetinaHead):
+    """Anchor-free head used in `FSAF <https://arxiv.org/abs/1903.00621>`_.
+
+    The head contains two subnetworks. The first classifies anchor boxes and
+    the second regresses deltas for the anchors (num_anchors is 1 for anchor-
+    free methods)
+
+    Args:
+        *args: Same as its base class in :class:`RetinaHead`
+        score_threshold (float, optional): The score_threshold to calculate
+            positive recall. If given, prediction scores lower than this value
+            is counted as incorrect prediction. Default to None.
+        **kwargs: Same as its base class in :class:`RetinaHead`
+
+    Example:
+        >>> import torch
+        >>> self = FSAFHead(11, 7)
+        >>> x = torch.rand(1, 7, 32, 32)
+        >>> cls_score, bbox_pred = self.forward_single(x)
+        >>> # Each anchor predicts a score for each class except background
+        >>> cls_per_anchor = cls_score.shape[1] / self.num_anchors
+        >>> box_per_anchor = bbox_pred.shape[1] / self.num_anchors
+        >>> assert cls_per_anchor == self.num_classes
+        >>> assert box_per_anchor == 4
+    """
+
+    def __init__(self, *args, score_threshold=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.score_threshold = score_threshold
+
+    def forward_single(self, x):
+        """Forward feature map of a single scale level.
+
+        Args:
+            x (Tensor): Feature map of a single scale level.
+
+        Returns:
+            tuple (Tensor):
+                cls_score (Tensor): Box scores for each scale level
+                    Has shape (N, num_points * num_classes, H, W).
+                bbox_pred (Tensor): Box energies / deltas for each scale
+                    level with shape (N, num_points * 4, H, W).
+        """
+        cls_score, bbox_pred = super().forward_single(x)
+        # relu: TBLR encoder only accepts positive bbox_pred
+        return cls_score, self.relu(bbox_pred)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        super(FSAFHead, self).init_weights()
+        # The positive bias in self.retina_reg conv is to prevent predicted \
+        #  bbox with 0 area
+        normal_init(self.retina_reg, std=0.01, bias=0.25)
+
+    def _get_targets_single(self,
+                            flat_anchors,
+                            valid_flags,
+                            gt_bboxes,
+                            gt_bboxes_ignore,
+                            gt_labels,
+                            img_meta,
+                            label_channels=1,
+                            unmap_outputs=True):
+        """Compute regression and classification targets for anchors in a
+        single image.
+
+        Most of the codes are the same with the base class
+          :obj: `AnchorHead`, except that it also collects and returns
+          the matched gt index in the image (from 0 to num_gt-1). If the
+          anchor bbox is not matched to any gt, the corresponding value in
+          pos_gt_inds is -1.
+        """
+        inside_flags = anchor_inside_flags(flat_anchors, valid_flags,
+                                           img_meta['img_shape'][:2],
+                                           self.train_cfg.allowed_border)
+        if not inside_flags.any():
+            return (None, ) * 7
+        # Assign gt and sample anchors
+        anchors = flat_anchors[inside_flags.type(torch.bool), :]
+        assign_result = self.assigner.assign(
+            anchors, gt_bboxes, gt_bboxes_ignore,
+            None if self.sampling else gt_labels)
+
+        sampling_result = self.sampler.sample(assign_result, anchors,
+                                              gt_bboxes)
+
+        num_valid_anchors = anchors.shape[0]
+        bbox_targets = torch.zeros_like(anchors)
+        bbox_weights = torch.zeros_like(anchors)
+        labels = anchors.new_full((num_valid_anchors, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = anchors.new_zeros((num_valid_anchors, label_channels),
+                                          dtype=torch.float)
+        pos_gt_inds = anchors.new_full((num_valid_anchors, ),
+                                       -1,
+                                       dtype=torch.long)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+
+        if len(pos_inds) > 0:
+            if not self.reg_decoded_bbox:
+                pos_bbox_targets = self.bbox_coder.encode(
+                    sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes)
+            else:
+                pos_bbox_targets = sampling_result.pos_gt_bboxes
+            bbox_targets[pos_inds, :] = pos_bbox_targets
+            bbox_weights[pos_inds, :] = 1.0
+            # The assigned gt_index for each anchor. (0-based)
+            pos_gt_inds[pos_inds] = sampling_result.pos_assigned_gt_inds
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # shadowed_labels is a tensor composed of tuples
+        #  (anchor_inds, class_label) that indicate those anchors lying in the
+        #  outer region of a gt or overlapped by another gt with a smaller
+        #  area.
+        #
+        # Therefore, only the shadowed labels are ignored for loss calculation.
+        # the key `shadowed_labels` is defined in :obj:`CenterRegionAssigner`
+        shadowed_labels = assign_result.get_extra_property('shadowed_labels')
+        if shadowed_labels is not None and shadowed_labels.numel():
+            if len(shadowed_labels.shape) == 2:
+                idx_, label_ = shadowed_labels[:, 0], shadowed_labels[:, 1]
+                assert (labels[idx_] != label_).all(), \
+                    'One label cannot be both positive and ignored'
+                label_weights[idx_, label_] = 0
+            else:
+                label_weights[shadowed_labels] = 0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_anchors.size(0)
+            labels = unmap(labels, num_total_anchors, inside_flags)
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_targets = unmap(bbox_targets, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+            pos_gt_inds = unmap(
+                pos_gt_inds, num_total_anchors, inside_flags, fill=-1)
+
+        return (labels, label_weights, bbox_targets, bbox_weights, pos_inds,
+                neg_inds, sampling_result, pos_gt_inds)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute loss of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_points * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_points * 4, H, W).
+            gt_bboxes (list[Tensor]): each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        for i in range(len(bbox_preds)):  # loop over fpn level
+            # avoid 0 area of the predicted bbox
+            bbox_preds[i] = bbox_preds[i].clamp(min=1e-4)
+        # TODO: It may directly use the base-class loss function.
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+        batch_size = len(gt_bboxes)
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg,
+         pos_assigned_gt_inds_list) = cls_reg_targets
+
+        num_gts = np.array(list(map(len, gt_labels)))
+        num_total_samples = (
+            num_total_pos + num_total_neg if self.sampling else num_total_pos)
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        # concat all level anchors and flags to a single tensor
+        concat_anchor_list = []
+        for i in range(len(anchor_list)):
+            concat_anchor_list.append(torch.cat(anchor_list[i]))
+        all_anchor_list = images_to_levels(concat_anchor_list,
+                                           num_level_anchors)
+        losses_cls, losses_bbox = multi_apply(
+            self.loss_single,
+            cls_scores,
+            bbox_preds,
+            all_anchor_list,
+            labels_list,
+            label_weights_list,
+            bbox_targets_list,
+            bbox_weights_list,
+            num_total_samples=num_total_samples)
+
+        # `pos_assigned_gt_inds_list` (length: fpn_levels) stores the assigned
+        # gt index of each anchor bbox in each fpn level.
+        cum_num_gts = list(np.cumsum(num_gts))  # length of batch_size
+        for i, assign in enumerate(pos_assigned_gt_inds_list):
+            # loop over fpn levels
+            for j in range(1, batch_size):
+                # loop over batch size
+                # Convert gt indices in each img to those in the batch
+                assign[j][assign[j] >= 0] += int(cum_num_gts[j - 1])
+            pos_assigned_gt_inds_list[i] = assign.flatten()
+            labels_list[i] = labels_list[i].flatten()
+        num_gts = sum(map(len, gt_labels))  # total number of gt in the batch
+        # The unique label index of each gt in the batch
+        label_sequence = torch.arange(num_gts, device=device)
+        # Collect the average loss of each gt in each level
+        with torch.no_grad():
+            loss_levels, = multi_apply(
+                self.collect_loss_level_single,
+                losses_cls,
+                losses_bbox,
+                pos_assigned_gt_inds_list,
+                labels_seq=label_sequence)
+            # Shape: (fpn_levels, num_gts). Loss of each gt at each fpn level
+            loss_levels = torch.stack(loss_levels, dim=0)
+            # Locate the best fpn level for loss back-propagation
+            if loss_levels.numel() == 0:  # zero gt
+                argmin = loss_levels.new_empty((num_gts, ), dtype=torch.long)
+            else:
+                _, argmin = loss_levels.min(dim=0)
+
+        # Reweight the loss of each (anchor, label) pair, so that only those
+        #  at the best gt level are back-propagated.
+        losses_cls, losses_bbox, pos_inds = multi_apply(
+            self.reweight_loss_single,
+            losses_cls,
+            losses_bbox,
+            pos_assigned_gt_inds_list,
+            labels_list,
+            list(range(len(losses_cls))),
+            min_levels=argmin)
+        num_pos = torch.cat(pos_inds, 0).sum().float()
+        pos_recall = self.calculate_pos_recall(cls_scores, labels_list,
+                                               pos_inds)
+
+        if num_pos == 0:  # No gt
+            avg_factor = num_pos + float(num_total_neg)
+        else:
+            avg_factor = num_pos
+        for i in range(len(losses_cls)):
+            losses_cls[i] /= avg_factor
+            losses_bbox[i] /= avg_factor
+        return dict(
+            loss_cls=losses_cls,
+            loss_bbox=losses_bbox,
+            num_pos=num_pos / batch_size,
+            pos_recall=pos_recall)
+
+    def calculate_pos_recall(self, cls_scores, labels_list, pos_inds):
+        """Calculate positive recall with score threshold.
+
+        Args:
+            cls_scores (list[Tensor]): Classification scores at all fpn levels.
+                Each tensor is in shape (N, num_classes * num_anchors, H, W)
+            labels_list (list[Tensor]): The label that each anchor is assigned
+                to. Shape (N * H * W * num_anchors, )
+            pos_inds (list[Tensor]): List of bool tensors indicating whether
+                the anchor is assigned to a positive label.
+                Shape (N * H * W * num_anchors, )
+
+        Returns:
+            Tensor: A single float number indicating the positive recall.
+        """
+        with torch.no_grad():
+            num_class = self.num_classes
+            scores = [
+                cls.permute(0, 2, 3, 1).reshape(-1, num_class)[pos]
+                for cls, pos in zip(cls_scores, pos_inds)
+            ]
+            labels = [
+                label.reshape(-1)[pos]
+                for label, pos in zip(labels_list, pos_inds)
+            ]
+            scores = torch.cat(scores, dim=0)
+            labels = torch.cat(labels, dim=0)
+            if self.use_sigmoid_cls:
+                scores = scores.sigmoid()
+            else:
+                scores = scores.softmax(dim=1)
+
+            return accuracy(scores, labels, thresh=self.score_threshold)
+
+    def collect_loss_level_single(self, cls_loss, reg_loss, assigned_gt_inds,
+                                  labels_seq):
+        """Get the average loss in each FPN level w.r.t. each gt label.
+
+        Args:
+            cls_loss (Tensor): Classification loss of each feature map pixel,
+              shape (num_anchor, num_class)
+            reg_loss (Tensor): Regression loss of each feature map pixel,
+              shape (num_anchor, 4)
+            assigned_gt_inds (Tensor): It indicates which gt the prior is
+              assigned to (0-based, -1: no assignment). shape (num_anchor),
+            labels_seq: The rank of labels. shape (num_gt)
+
+        Returns:
+            shape: (num_gt), average loss of each gt in this level
+        """
+        if len(reg_loss.shape) == 2:  # iou loss has shape (num_prior, 4)
+            reg_loss = reg_loss.sum(dim=-1)  # sum loss in tblr dims
+        if len(cls_loss.shape) == 2:
+            cls_loss = cls_loss.sum(dim=-1)  # sum loss in class dims
+        loss = cls_loss + reg_loss
+        assert loss.size(0) == assigned_gt_inds.size(0)
+        # Default loss value is 1e6 for a layer where no anchor is positive
+        #  to ensure it will not be chosen to back-propagate gradient
+        losses_ = loss.new_full(labels_seq.shape, 1e6)
+        for i, l in enumerate(labels_seq):
+            match = assigned_gt_inds == l
+            if match.any():
+                losses_[i] = loss[match].mean()
+        return losses_,
+
+    def reweight_loss_single(self, cls_loss, reg_loss, assigned_gt_inds,
+                             labels, level, min_levels):
+        """Reweight loss values at each level.
+
+        Reassign loss values at each level by masking those where the
+        pre-calculated loss is too large. Then return the reduced losses.
+
+        Args:
+            cls_loss (Tensor): Element-wise classification loss.
+              Shape: (num_anchors, num_classes)
+            reg_loss (Tensor): Element-wise regression loss.
+              Shape: (num_anchors, 4)
+            assigned_gt_inds (Tensor): The gt indices that each anchor bbox
+              is assigned to. -1 denotes a negative anchor, otherwise it is the
+              gt index (0-based). Shape: (num_anchors, ),
+            labels (Tensor): Label assigned to anchors. Shape: (num_anchors, ).
+            level (int): The current level index in the pyramid
+              (0-4 for RetinaNet)
+            min_levels (Tensor): The best-matching level for each gt.
+              Shape: (num_gts, ),
+
+        Returns:
+            tuple:
+                - cls_loss: Reduced corrected classification loss. Scalar.
+                - reg_loss: Reduced corrected regression loss. Scalar.
+                - pos_flags (Tensor): Corrected bool tensor indicating the
+                  final postive anchors. Shape: (num_anchors, ).
+        """
+        loc_weight = torch.ones_like(reg_loss)
+        cls_weight = torch.ones_like(cls_loss)
+        pos_flags = assigned_gt_inds >= 0  # positive pixel flag
+        pos_indices = torch.nonzero(pos_flags, as_tuple=False).flatten()
+
+        if pos_flags.any():  # pos pixels exist
+            pos_assigned_gt_inds = assigned_gt_inds[pos_flags]
+            zeroing_indices = (min_levels[pos_assigned_gt_inds] != level)
+            neg_indices = pos_indices[zeroing_indices]
+
+            if neg_indices.numel():
+                pos_flags[neg_indices] = 0
+                loc_weight[neg_indices] = 0
+                # Only the weight corresponding to the label is
+                #  zeroed out if not selected
+                zeroing_labels = labels[neg_indices]
+                assert (zeroing_labels >= 0).all()
+                cls_weight[neg_indices, zeroing_labels] = 0
+
+        # Weighted loss for both cls and reg loss
+        cls_loss = weight_reduce_loss(cls_loss, cls_weight, reduction='sum')
+        reg_loss = weight_reduce_loss(reg_loss, loc_weight, reduction='sum')
+
+        return cls_loss, reg_loss, pos_flags
diff --git a/detection/scrfd/mmdet/models/dense_heads/ga_retina_head.py b/detection/scrfd/mmdet/models/dense_heads/ga_retina_head.py
new file mode 100755
index 000000000..8822d1ca7
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/ga_retina_head.py
@@ -0,0 +1,109 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.ops import MaskedConv2d
+
+from ..builder import HEADS
+from .guided_anchor_head import FeatureAdaption, GuidedAnchorHead
+
+
+@HEADS.register_module()
+class GARetinaHead(GuidedAnchorHead):
+    """Guided-Anchor-based RetinaNet head."""
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        super(GARetinaHead, self).__init__(num_classes, in_channels, **kwargs)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+
+        self.conv_loc = nn.Conv2d(self.feat_channels, 1, 1)
+        self.conv_shape = nn.Conv2d(self.feat_channels, self.num_anchors * 2,
+                                    1)
+        self.feature_adaption_cls = FeatureAdaption(
+            self.feat_channels,
+            self.feat_channels,
+            kernel_size=3,
+            deform_groups=self.deform_groups)
+        self.feature_adaption_reg = FeatureAdaption(
+            self.feat_channels,
+            self.feat_channels,
+            kernel_size=3,
+            deform_groups=self.deform_groups)
+        self.retina_cls = MaskedConv2d(
+            self.feat_channels,
+            self.num_anchors * self.cls_out_channels,
+            3,
+            padding=1)
+        self.retina_reg = MaskedConv2d(
+            self.feat_channels, self.num_anchors * 4, 3, padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the layer."""
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+
+        self.feature_adaption_cls.init_weights()
+        self.feature_adaption_reg.init_weights()
+
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.conv_loc, std=0.01, bias=bias_cls)
+        normal_init(self.conv_shape, std=0.01)
+        normal_init(self.retina_cls, std=0.01, bias=bias_cls)
+        normal_init(self.retina_reg, std=0.01)
+
+    def forward_single(self, x):
+        """Forward feature map of a single scale level."""
+        cls_feat = x
+        reg_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            reg_feat = reg_conv(reg_feat)
+
+        loc_pred = self.conv_loc(cls_feat)
+        shape_pred = self.conv_shape(reg_feat)
+
+        cls_feat = self.feature_adaption_cls(cls_feat, shape_pred)
+        reg_feat = self.feature_adaption_reg(reg_feat, shape_pred)
+
+        if not self.training:
+            mask = loc_pred.sigmoid()[0] >= self.loc_filter_thr
+        else:
+            mask = None
+        cls_score = self.retina_cls(cls_feat, mask)
+        bbox_pred = self.retina_reg(reg_feat, mask)
+        return cls_score, bbox_pred, shape_pred, loc_pred
diff --git a/detection/scrfd/mmdet/models/dense_heads/ga_rpn_head.py b/detection/scrfd/mmdet/models/dense_heads/ga_rpn_head.py
new file mode 100755
index 000000000..d3c3a84b2
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/ga_rpn_head.py
@@ -0,0 +1,133 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import normal_init
+from mmcv.ops import nms
+
+from ..builder import HEADS
+from .guided_anchor_head import GuidedAnchorHead
+from .rpn_test_mixin import RPNTestMixin
+
+
+@HEADS.register_module()
+class GARPNHead(RPNTestMixin, GuidedAnchorHead):
+    """Guided-Anchor-based RPN head."""
+
+    def __init__(self, in_channels, **kwargs):
+        super(GARPNHead, self).__init__(1, in_channels, **kwargs)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.rpn_conv = nn.Conv2d(
+            self.in_channels, self.feat_channels, 3, padding=1)
+        super(GARPNHead, self)._init_layers()
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        normal_init(self.rpn_conv, std=0.01)
+        super(GARPNHead, self).init_weights()
+
+    def forward_single(self, x):
+        """Forward feature of a single scale level."""
+
+        x = self.rpn_conv(x)
+        x = F.relu(x, inplace=True)
+        (cls_score, bbox_pred, shape_pred,
+         loc_pred) = super(GARPNHead, self).forward_single(x)
+        return cls_score, bbox_pred, shape_pred, loc_pred
+
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             shape_preds,
+             loc_preds,
+             gt_bboxes,
+             img_metas,
+             gt_bboxes_ignore=None):
+        losses = super(GARPNHead, self).loss(
+            cls_scores,
+            bbox_preds,
+            shape_preds,
+            loc_preds,
+            gt_bboxes,
+            None,
+            img_metas,
+            gt_bboxes_ignore=gt_bboxes_ignore)
+        return dict(
+            loss_rpn_cls=losses['loss_cls'],
+            loss_rpn_bbox=losses['loss_bbox'],
+            loss_anchor_shape=losses['loss_shape'],
+            loss_anchor_loc=losses['loss_loc'])
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_anchors,
+                           mlvl_masks,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False):
+        cfg = self.test_cfg if cfg is None else cfg
+        mlvl_proposals = []
+        for idx in range(len(cls_scores)):
+            rpn_cls_score = cls_scores[idx]
+            rpn_bbox_pred = bbox_preds[idx]
+            anchors = mlvl_anchors[idx]
+            mask = mlvl_masks[idx]
+            assert rpn_cls_score.size()[-2:] == rpn_bbox_pred.size()[-2:]
+            # if no location is kept, end.
+            if mask.sum() == 0:
+                continue
+            rpn_cls_score = rpn_cls_score.permute(1, 2, 0)
+            if self.use_sigmoid_cls:
+                rpn_cls_score = rpn_cls_score.reshape(-1)
+                scores = rpn_cls_score.sigmoid()
+            else:
+                rpn_cls_score = rpn_cls_score.reshape(-1, 2)
+                # remind that we set FG labels to [0, num_class-1]
+                # since mmdet v2.0
+                # BG cat_id: num_class
+                scores = rpn_cls_score.softmax(dim=1)[:, :-1]
+            # filter scores, bbox_pred w.r.t. mask.
+            # anchors are filtered in get_anchors() beforehand.
+            scores = scores[mask]
+            rpn_bbox_pred = rpn_bbox_pred.permute(1, 2, 0).reshape(-1,
+                                                                   4)[mask, :]
+            if scores.dim() == 0:
+                rpn_bbox_pred = rpn_bbox_pred.unsqueeze(0)
+                anchors = anchors.unsqueeze(0)
+                scores = scores.unsqueeze(0)
+            # filter anchors, bbox_pred, scores w.r.t. scores
+            if cfg.nms_pre > 0 and scores.shape[0] > cfg.nms_pre:
+                _, topk_inds = scores.topk(cfg.nms_pre)
+                rpn_bbox_pred = rpn_bbox_pred[topk_inds, :]
+                anchors = anchors[topk_inds, :]
+                scores = scores[topk_inds]
+            # get proposals w.r.t. anchors and rpn_bbox_pred
+            proposals = self.bbox_coder.decode(
+                anchors, rpn_bbox_pred, max_shape=img_shape)
+            # filter out too small bboxes
+            if cfg.min_bbox_size > 0:
+                w = proposals[:, 2] - proposals[:, 0]
+                h = proposals[:, 3] - proposals[:, 1]
+                valid_inds = torch.nonzero(
+                    (w >= cfg.min_bbox_size) & (h >= cfg.min_bbox_size),
+                    as_tuple=False).squeeze()
+                proposals = proposals[valid_inds, :]
+                scores = scores[valid_inds]
+            # NMS in current level
+            proposals, _ = nms(proposals, scores, cfg.nms_thr)
+            proposals = proposals[:cfg.nms_post, :]
+            mlvl_proposals.append(proposals)
+        proposals = torch.cat(mlvl_proposals, 0)
+        if cfg.nms_across_levels:
+            # NMS across multi levels
+            proposals, _ = nms(proposals[:, :4], proposals[:, -1], cfg.nms_thr)
+            proposals = proposals[:cfg.max_num, :]
+        else:
+            scores = proposals[:, 4]
+            num = min(cfg.max_num, proposals.shape[0])
+            _, topk_inds = scores.topk(num)
+            proposals = proposals[topk_inds, :]
+        return proposals
diff --git a/detection/scrfd/mmdet/models/dense_heads/gfl_head.py b/detection/scrfd/mmdet/models/dense_heads/gfl_head.py
new file mode 100755
index 000000000..ab7a69abc
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/gfl_head.py
@@ -0,0 +1,638 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, Scale, bias_init_with_prob, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, bbox2distance, bbox_overlaps,
+                        build_assigner, build_sampler, distance2bbox,
+                        images_to_levels, multi_apply, multiclass_nms,
+                        reduce_mean, unmap)
+from ..builder import HEADS, build_loss
+from .anchor_head import AnchorHead
+
+
+class Integral(nn.Module):
+    """A fixed layer for calculating integral result from distribution.
+
+    This layer calculates the target location by :math: `sum{P(y_i) * y_i}`,
+    P(y_i) denotes the softmax vector that represents the discrete distribution
+    y_i denotes the discrete set, usually {0, 1, 2, ..., reg_max}
+
+    Args:
+        reg_max (int): The maximal value of the discrete set. Default: 16. You
+            may want to reset it according to your new dataset or related
+            settings.
+    """
+
+    def __init__(self, reg_max=16):
+        super(Integral, self).__init__()
+        self.reg_max = reg_max
+        self.register_buffer('project',
+                             torch.linspace(0, self.reg_max, self.reg_max + 1))
+
+    def forward(self, x):
+        """Forward feature from the regression head to get integral result of
+        bounding box location.
+
+        Args:
+            x (Tensor): Features of the regression head, shape (N, 4*(n+1)),
+                n is self.reg_max.
+
+        Returns:
+            x (Tensor): Integral result of box locations, i.e., distance
+                offsets from the box center in four directions, shape (N, 4).
+        """
+        x = F.softmax(x.reshape(-1, self.reg_max + 1), dim=1)
+        x = F.linear(x, self.project.type_as(x)).reshape(-1, 4)
+        return x
+
+
+@HEADS.register_module()
+class GFLHead(AnchorHead):
+    """Generalized Focal Loss: Learning Qualified and Distributed Bounding
+    Boxes for Dense Object Detection.
+
+    GFL head structure is similar with ATSS, however GFL uses
+    1) joint representation for classification and localization quality, and
+    2) flexible General distribution for bounding box locations,
+    which are supervised by
+    Quality Focal Loss (QFL) and Distribution Focal Loss (DFL), respectively
+
+    https://arxiv.org/abs/2006.04388
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        stacked_convs (int): Number of conv layers in cls and reg tower.
+            Default: 4.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: dict(type='GN', num_groups=32, requires_grad=True).
+        loss_qfl (dict): Config of Quality Focal Loss (QFL).
+        reg_max (int): Max value of integral set :math: `{0, ..., reg_max}`
+            in QFL setting. Default: 16.
+    Example:
+        >>> self = GFLHead(11, 7)
+        >>> feats = [torch.rand(1, 7, s, s) for s in [4, 8, 16, 32, 64]]
+        >>> cls_quality_score, bbox_pred = self.forward(feats)
+        >>> assert len(cls_quality_score) == len(self.scales)
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+                 loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
+                 reg_max=16,
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.reg_max = reg_max
+        super(GFLHead, self).__init__(num_classes, in_channels, **kwargs)
+
+        self.sampling = False
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # SSD sampling=False so use PseudoSampler
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+        self.integral = Integral(self.reg_max)
+        self.loss_dfl = build_loss(loss_dfl)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        assert self.num_anchors == 1, 'anchor free version'
+        self.gfl_cls = nn.Conv2d(
+            self.feat_channels, self.cls_out_channels, 3, padding=1)
+        self.gfl_reg = nn.Conv2d(
+            self.feat_channels, 4 * (self.reg_max + 1), 3, padding=1)
+        self.scales = nn.ModuleList(
+            [Scale(1.0) for _ in self.anchor_generator.strides])
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.gfl_cls, std=0.01, bias=bias_cls)
+        normal_init(self.gfl_reg, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually a tuple of classification scores and bbox prediction
+                cls_scores (list[Tensor]): Classification and quality (IoU)
+                    joint scores for all scale levels, each is a 4D-tensor,
+                    the channel number is num_classes.
+                bbox_preds (list[Tensor]): Box distribution logits for all
+                    scale levels, each is a 4D-tensor, the channel number is
+                    4*(n+1), n is max value of integral set.
+        """
+        return multi_apply(self.forward_single, feats, self.scales)
+
+    def forward_single(self, x, scale):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls and quality joint scores for a single
+                    scale level the channel number is num_classes.
+                bbox_pred (Tensor): Box distribution logits for a single scale
+                    level, the channel number is 4*(n+1), n is max value of
+                    integral set.
+        """
+        cls_feat = x
+        reg_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            reg_feat = reg_conv(reg_feat)
+        cls_score = self.gfl_cls(cls_feat)
+        bbox_pred = scale(self.gfl_reg(reg_feat)).float()
+        return cls_score, bbox_pred
+
+    def anchor_center(self, anchors):
+        """Get anchor centers from anchors.
+
+        Args:
+            anchors (Tensor): Anchor list with shape (N, 4), "xyxy" format.
+
+        Returns:
+            Tensor: Anchor centers with shape (N, 2), "xy" format.
+        """
+        anchors_cx = (anchors[:, 2] + anchors[:, 0]) / 2
+        anchors_cy = (anchors[:, 3] + anchors[:, 1]) / 2
+        return torch.stack([anchors_cx, anchors_cy], dim=-1)
+
+    def loss_single(self, anchors, cls_score, bbox_pred, labels, label_weights,
+                    bbox_targets, stride, num_total_samples):
+        """Compute loss of a single scale level.
+
+        Args:
+            anchors (Tensor): Box reference for each scale level with shape
+                (N, num_total_anchors, 4).
+            cls_score (Tensor): Cls and quality joint scores for each scale
+                level has shape (N, num_classes, H, W).
+            bbox_pred (Tensor): Box distribution logits for each scale
+                level with shape (N, 4*(n+1), H, W), n is max value of integral
+                set.
+            labels (Tensor): Labels of each anchors with shape
+                (N, num_total_anchors).
+            label_weights (Tensor): Label weights of each anchor with shape
+                (N, num_total_anchors)
+            bbox_targets (Tensor): BBox regression targets of each anchor wight
+                shape (N, num_total_anchors, 4).
+            stride (tuple): Stride in this scale level.
+            num_total_samples (int): Number of positive samples that is
+                reduced over all GPUs.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert stride[0] == stride[1], 'h stride is not equal to w stride!'
+        anchors = anchors.reshape(-1, 4)
+        cls_score = cls_score.permute(0, 2, 3,
+                                      1).reshape(-1, self.cls_out_channels)
+        bbox_pred = bbox_pred.permute(0, 2, 3,
+                                      1).reshape(-1, 4 * (self.reg_max + 1))
+        bbox_targets = bbox_targets.reshape(-1, 4)
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((labels >= 0)
+                    & (labels < bg_class_ind)).nonzero().squeeze(1)
+        score = label_weights.new_zeros(labels.shape)
+
+        if len(pos_inds) > 0:
+            pos_bbox_targets = bbox_targets[pos_inds]
+            pos_bbox_pred = bbox_pred[pos_inds]
+            pos_anchors = anchors[pos_inds]
+            pos_anchor_centers = self.anchor_center(pos_anchors) / stride[0]
+
+            weight_targets = cls_score.detach().sigmoid()
+            weight_targets = weight_targets.max(dim=1)[0][pos_inds]
+            pos_bbox_pred_corners = self.integral(pos_bbox_pred)
+            pos_decode_bbox_pred = distance2bbox(pos_anchor_centers,
+                                                 pos_bbox_pred_corners)
+            pos_decode_bbox_targets = pos_bbox_targets / stride[0]
+            score[pos_inds] = bbox_overlaps(
+                pos_decode_bbox_pred.detach(),
+                pos_decode_bbox_targets,
+                is_aligned=True)
+            pred_corners = pos_bbox_pred.reshape(-1, self.reg_max + 1)
+            target_corners = bbox2distance(pos_anchor_centers,
+                                           pos_decode_bbox_targets,
+                                           self.reg_max).reshape(-1)
+
+            # regression loss
+            loss_bbox = self.loss_bbox(
+                pos_decode_bbox_pred,
+                pos_decode_bbox_targets,
+                weight=weight_targets,
+                avg_factor=1.0)
+
+            # dfl loss
+            loss_dfl = self.loss_dfl(
+                pred_corners,
+                target_corners,
+                weight=weight_targets[:, None].expand(-1, 4).reshape(-1),
+                avg_factor=4.0)
+        else:
+            loss_bbox = bbox_pred.sum() * 0
+            loss_dfl = bbox_pred.sum() * 0
+            weight_targets = torch.tensor(0).cuda()
+
+        # cls (qfl) loss
+        loss_cls = self.loss_cls(
+            cls_score, (labels, score),
+            weight=label_weights,
+            avg_factor=num_total_samples)
+
+        return loss_cls, loss_bbox, loss_dfl, weight_targets.sum()
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Cls and quality scores for each scale
+                level has shape (N, num_classes, H, W).
+            bbox_preds (list[Tensor]): Box distribution logits for each scale
+                level with shape (N, 4*(n+1), H, W), n is max value of integral
+                set.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+
+        (anchor_list, labels_list, label_weights_list, bbox_targets_list,
+         bbox_weights_list, num_total_pos, num_total_neg) = cls_reg_targets
+
+        num_total_samples = reduce_mean(
+            torch.tensor(num_total_pos, dtype=torch.float,
+                         device=device)).item()
+        num_total_samples = max(num_total_samples, 1.0)
+
+        losses_cls, losses_bbox, losses_dfl,\
+            avg_factor = multi_apply(
+                self.loss_single,
+                anchor_list,
+                cls_scores,
+                bbox_preds,
+                labels_list,
+                label_weights_list,
+                bbox_targets_list,
+                self.anchor_generator.strides,
+                num_total_samples=num_total_samples)
+
+        avg_factor = sum(avg_factor)
+        avg_factor = reduce_mean(avg_factor).item()
+        losses_bbox = list(map(lambda x: x / avg_factor, losses_bbox))
+        losses_dfl = list(map(lambda x: x / avg_factor, losses_dfl))
+        return dict(
+            loss_cls=losses_cls, loss_bbox=losses_bbox, loss_dfl=losses_dfl)
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into labeled boxes.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for a single scale level
+                has shape (num_classes, H, W).
+            bbox_preds (list[Tensor]): Box distribution logits for a single
+                scale level with shape (4*(n+1), H, W), n is max value of
+                integral set.
+            mlvl_anchors (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): Bbox predictions in shape (N, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (N,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, stride, anchors in zip(
+                cls_scores, bbox_preds, self.anchor_generator.strides,
+                mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            assert stride[0] == stride[1]
+
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0)
+            bbox_pred = self.integral(bbox_pred) * stride[0]
+
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                max_scores, _ = scores.max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+
+            bboxes = distance2bbox(
+                self.anchor_center(anchors), bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+
+        mlvl_scores = torch.cat(mlvl_scores)
+        # Add a dummy background class to the backend when using sigmoid
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores
+
+    def get_targets(self,
+                    anchor_list,
+                    valid_flag_list,
+                    gt_bboxes_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None,
+                    gt_labels_list=None,
+                    label_channels=1,
+                    unmap_outputs=True):
+        """Get targets for GFL head.
+
+        This method is almost the same as `AnchorHead.get_targets()`. Besides
+        returning the targets as the parent method does, it also returns the
+        anchors as the first element of the returned tuple.
+        """
+        num_imgs = len(img_metas)
+        assert len(anchor_list) == len(valid_flag_list) == num_imgs
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        num_level_anchors_list = [num_level_anchors] * num_imgs
+
+        # concat all level anchors and flags to a single tensor
+        for i in range(num_imgs):
+            assert len(anchor_list[i]) == len(valid_flag_list[i])
+            anchor_list[i] = torch.cat(anchor_list[i])
+            valid_flag_list[i] = torch.cat(valid_flag_list[i])
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        (all_anchors, all_labels, all_label_weights, all_bbox_targets,
+         all_bbox_weights, pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single,
+             anchor_list,
+             valid_flag_list,
+             num_level_anchors_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             gt_labels_list,
+             img_metas,
+             label_channels=label_channels,
+             unmap_outputs=unmap_outputs)
+        # no valid anchors
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        anchors_list = images_to_levels(all_anchors, num_level_anchors)
+        labels_list = images_to_levels(all_labels, num_level_anchors)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_anchors)
+        bbox_targets_list = images_to_levels(all_bbox_targets,
+                                             num_level_anchors)
+        bbox_weights_list = images_to_levels(all_bbox_weights,
+                                             num_level_anchors)
+        return (anchors_list, labels_list, label_weights_list,
+                bbox_targets_list, bbox_weights_list, num_total_pos,
+                num_total_neg)
+
+    def _get_target_single(self,
+                           flat_anchors,
+                           valid_flags,
+                           num_level_anchors,
+                           gt_bboxes,
+                           gt_bboxes_ignore,
+                           gt_labels,
+                           img_meta,
+                           label_channels=1,
+                           unmap_outputs=True):
+        """Compute regression, classification targets for anchors in a single
+        image.
+
+        Args:
+            flat_anchors (Tensor): Multi-level anchors of the image, which are
+                concatenated into a single tensor of shape (num_anchors, 4)
+            valid_flags (Tensor): Multi level valid flags of the image,
+                which are concatenated into a single tensor of
+                    shape (num_anchors,).
+            num_level_anchors Tensor): Number of anchors of each scale level.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            img_meta (dict): Meta info of the image.
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple: N is the number of total anchors in the image.
+                anchors (Tensor): All anchors in the image with shape (N, 4).
+                labels (Tensor): Labels of all anchors in the image with shape
+                    (N,).
+                label_weights (Tensor): Label weights of all anchor in the
+                    image with shape (N,).
+                bbox_targets (Tensor): BBox targets of all anchors in the
+                    image with shape (N, 4).
+                bbox_weights (Tensor): BBox weights of all anchors in the
+                    image with shape (N, 4).
+                pos_inds (Tensor): Indices of postive anchor with shape
+                    (num_pos,).
+                neg_inds (Tensor): Indices of negative anchor with shape
+                    (num_neg,).
+        """
+        inside_flags = anchor_inside_flags(flat_anchors, valid_flags,
+                                           img_meta['img_shape'][:2],
+                                           self.train_cfg.allowed_border)
+        if not inside_flags.any():
+            return (None, ) * 7
+        # assign gt and sample anchors
+        anchors = flat_anchors[inside_flags, :]
+
+        num_level_anchors_inside = self.get_num_level_anchors_inside(
+            num_level_anchors, inside_flags)
+        #print('NNN:', self.assigner.__class__.__name__)
+        if self.assigner.__class__.__name__=='ATSSAssigner':
+            assign_result = self.assigner.assign(anchors, num_level_anchors_inside,
+                                                 gt_bboxes, gt_bboxes_ignore,
+                                                 gt_labels)
+        else:
+            assign_result = self.assigner.assign(anchors, 
+                                                 gt_bboxes, gt_bboxes_ignore,
+                                                 gt_labels)
+
+        sampling_result = self.sampler.sample(assign_result, anchors,
+                                              gt_bboxes)
+
+        num_valid_anchors = anchors.shape[0]
+        bbox_targets = torch.zeros_like(anchors)
+        bbox_weights = torch.zeros_like(anchors)
+        labels = anchors.new_full((num_valid_anchors, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = anchors.new_zeros(num_valid_anchors, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            pos_bbox_targets = sampling_result.pos_gt_bboxes
+            bbox_targets[pos_inds, :] = pos_bbox_targets
+            bbox_weights[pos_inds, :] = 1.0
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_anchors.size(0)
+            anchors = unmap(anchors, num_total_anchors, inside_flags)
+            labels = unmap(
+                labels, num_total_anchors, inside_flags, fill=self.num_classes)
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_targets = unmap(bbox_targets, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+
+        return (anchors, labels, label_weights, bbox_targets, bbox_weights,
+                pos_inds, neg_inds)
+
+    def get_num_level_anchors_inside(self, num_level_anchors, inside_flags):
+        split_inside_flags = torch.split(inside_flags, num_level_anchors)
+        num_level_anchors_inside = [
+            int(flags.sum()) for flags in split_inside_flags
+        ]
+        return num_level_anchors_inside
diff --git a/detection/scrfd/mmdet/models/dense_heads/guided_anchor_head.py b/detection/scrfd/mmdet/models/dense_heads/guided_anchor_head.py
new file mode 100755
index 000000000..062df4864
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/guided_anchor_head.py
@@ -0,0 +1,855 @@
+import torch
+import torch.nn as nn
+from mmcv.cnn import bias_init_with_prob, normal_init
+from mmcv.ops import DeformConv2d, MaskedConv2d
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, build_anchor_generator,
+                        build_assigner, build_bbox_coder, build_sampler,
+                        calc_region, images_to_levels, multi_apply,
+                        multiclass_nms, unmap)
+from ..builder import HEADS, build_loss
+from .anchor_head import AnchorHead
+
+
+class FeatureAdaption(nn.Module):
+    """Feature Adaption Module.
+
+    Feature Adaption Module is implemented based on DCN v1.
+    It uses anchor shape prediction rather than feature map to
+    predict offsets of deform conv layer.
+
+    Args:
+        in_channels (int): Number of channels in the input feature map.
+        out_channels (int): Number of channels in the output feature map.
+        kernel_size (int): Deformable conv kernel size.
+        deform_groups (int): Deformable conv group size.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 deform_groups=4):
+        super(FeatureAdaption, self).__init__()
+        offset_channels = kernel_size * kernel_size * 2
+        self.conv_offset = nn.Conv2d(
+            2, deform_groups * offset_channels, 1, bias=False)
+        self.conv_adaption = DeformConv2d(
+            in_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            padding=(kernel_size - 1) // 2,
+            deform_groups=deform_groups)
+        self.relu = nn.ReLU(inplace=True)
+
+    def init_weights(self):
+        normal_init(self.conv_offset, std=0.1)
+        normal_init(self.conv_adaption, std=0.01)
+
+    def forward(self, x, shape):
+        offset = self.conv_offset(shape.detach())
+        x = self.relu(self.conv_adaption(x, offset))
+        return x
+
+
+@HEADS.register_module()
+class GuidedAnchorHead(AnchorHead):
+    """Guided-Anchor-based head (GA-RPN, GA-RetinaNet, etc.).
+
+    This GuidedAnchorHead will predict high-quality feature guided
+    anchors and locations where anchors will be kept in inference.
+    There are mainly 3 categories of bounding-boxes.
+
+    - Sampled 9 pairs for target assignment. (approxes)
+    - The square boxes where the predicted anchors are based on. (squares)
+    - Guided anchors.
+
+    Please refer to https://arxiv.org/abs/1901.03278 for more details.
+
+    Args:
+        num_classes (int): Number of classes.
+        in_channels (int): Number of channels in the input feature map.
+        feat_channels (int): Number of hidden channels.
+        approx_anchor_generator (dict): Config dict for approx generator
+        square_anchor_generator (dict): Config dict for square generator
+        anchor_coder (dict): Config dict for anchor coder
+        bbox_coder (dict): Config dict for bbox coder
+        deform_groups: (int): Group number of DCN in
+            FeatureAdaption module.
+        loc_filter_thr (float): Threshold to filter out unconcerned regions.
+        loss_loc (dict): Config of location loss.
+        loss_shape (dict): Config of anchor shape loss.
+        loss_cls (dict): Config of classification loss.
+        loss_bbox (dict): Config of bbox regression loss.
+    """
+
+    def __init__(
+        self,
+        num_classes,
+        in_channels,
+        feat_channels=256,
+        approx_anchor_generator=dict(
+            type='AnchorGenerator',
+            octave_base_scale=8,
+            scales_per_octave=3,
+            ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64]),
+        square_anchor_generator=dict(
+            type='AnchorGenerator',
+            ratios=[1.0],
+            scales=[8],
+            strides=[4, 8, 16, 32, 64]),
+        anchor_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]
+        ),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]
+        ),
+        reg_decoded_bbox=False,
+        deform_groups=4,
+        loc_filter_thr=0.01,
+        train_cfg=None,
+        test_cfg=None,
+        loss_loc=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_shape=dict(type='BoundedIoULoss', beta=0.2, loss_weight=1.0),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
+                       loss_weight=1.0)):  # yapf: disable
+        super(AnchorHead, self).__init__()
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.feat_channels = feat_channels
+        self.deform_groups = deform_groups
+        self.loc_filter_thr = loc_filter_thr
+
+        # build approx_anchor_generator and square_anchor_generator
+        assert (approx_anchor_generator['octave_base_scale'] ==
+                square_anchor_generator['scales'][0])
+        assert (approx_anchor_generator['strides'] ==
+                square_anchor_generator['strides'])
+        self.approx_anchor_generator = build_anchor_generator(
+            approx_anchor_generator)
+        self.square_anchor_generator = build_anchor_generator(
+            square_anchor_generator)
+        self.approxs_per_octave = self.approx_anchor_generator \
+            .num_base_anchors[0]
+
+        self.reg_decoded_bbox = reg_decoded_bbox
+
+        # one anchor per location
+        self.num_anchors = 1
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        self.loc_focal_loss = loss_loc['type'] in ['FocalLoss']
+        self.sampling = loss_cls['type'] not in ['FocalLoss']
+        self.ga_sampling = train_cfg is not None and hasattr(
+            train_cfg, 'ga_sampler')
+        if self.use_sigmoid_cls:
+            self.cls_out_channels = self.num_classes
+        else:
+            self.cls_out_channels = self.num_classes + 1
+
+        # build bbox_coder
+        self.anchor_coder = build_bbox_coder(anchor_coder)
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+
+        # build losses
+        self.loss_loc = build_loss(loss_loc)
+        self.loss_shape = build_loss(loss_shape)
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox = build_loss(loss_bbox)
+
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # use PseudoSampler when sampling is False
+            if self.sampling and hasattr(self.train_cfg, 'sampler'):
+                sampler_cfg = self.train_cfg.sampler
+            else:
+                sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+            self.ga_assigner = build_assigner(self.train_cfg.ga_assigner)
+            if self.ga_sampling:
+                ga_sampler_cfg = self.train_cfg.ga_sampler
+            else:
+                ga_sampler_cfg = dict(type='PseudoSampler')
+            self.ga_sampler = build_sampler(ga_sampler_cfg, context=self)
+
+        self.fp16_enabled = False
+
+        self._init_layers()
+
+    def _init_layers(self):
+        self.relu = nn.ReLU(inplace=True)
+        self.conv_loc = nn.Conv2d(self.in_channels, 1, 1)
+        self.conv_shape = nn.Conv2d(self.in_channels, self.num_anchors * 2, 1)
+        self.feature_adaption = FeatureAdaption(
+            self.in_channels,
+            self.feat_channels,
+            kernel_size=3,
+            deform_groups=self.deform_groups)
+        self.conv_cls = MaskedConv2d(self.feat_channels,
+                                     self.num_anchors * self.cls_out_channels,
+                                     1)
+        self.conv_reg = MaskedConv2d(self.feat_channels, self.num_anchors * 4,
+                                     1)
+
+    def init_weights(self):
+        normal_init(self.conv_cls, std=0.01)
+        normal_init(self.conv_reg, std=0.01)
+
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.conv_loc, std=0.01, bias=bias_cls)
+        normal_init(self.conv_shape, std=0.01)
+
+        self.feature_adaption.init_weights()
+
+    def forward_single(self, x):
+        loc_pred = self.conv_loc(x)
+        shape_pred = self.conv_shape(x)
+        x = self.feature_adaption(x, shape_pred)
+        # masked conv is only used during inference for speed-up
+        if not self.training:
+            mask = loc_pred.sigmoid()[0] >= self.loc_filter_thr
+        else:
+            mask = None
+        cls_score = self.conv_cls(x, mask)
+        bbox_pred = self.conv_reg(x, mask)
+        return cls_score, bbox_pred, shape_pred, loc_pred
+
+    def forward(self, feats):
+        return multi_apply(self.forward_single, feats)
+
+    def get_sampled_approxs(self, featmap_sizes, img_metas, device='cuda'):
+        """Get sampled approxs and inside flags according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+            device (torch.device | str): device for returned tensors
+
+        Returns:
+            tuple: approxes of each image, inside flags of each image
+        """
+        num_imgs = len(img_metas)
+
+        # since feature map sizes of all images are the same, we only compute
+        # approxes for one time
+        multi_level_approxs = self.approx_anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+        approxs_list = [multi_level_approxs for _ in range(num_imgs)]
+
+        # for each image, we compute inside flags of multi level approxes
+        inside_flag_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_flags = []
+            multi_level_approxs = approxs_list[img_id]
+
+            # obtain valid flags for each approx first
+            multi_level_approx_flags = self.approx_anchor_generator \
+                .valid_flags(featmap_sizes,
+                             img_meta['pad_shape'],
+                             device=device)
+
+            for i, flags in enumerate(multi_level_approx_flags):
+                approxs = multi_level_approxs[i]
+                inside_flags_list = []
+                for i in range(self.approxs_per_octave):
+                    split_valid_flags = flags[i::self.approxs_per_octave]
+                    split_approxs = approxs[i::self.approxs_per_octave, :]
+                    inside_flags = anchor_inside_flags(
+                        split_approxs, split_valid_flags,
+                        img_meta['img_shape'][:2],
+                        self.train_cfg.allowed_border)
+                    inside_flags_list.append(inside_flags)
+                # inside_flag for a position is true if any anchor in this
+                # position is true
+                inside_flags = (
+                    torch.stack(inside_flags_list, 0).sum(dim=0) > 0)
+                multi_level_flags.append(inside_flags)
+            inside_flag_list.append(multi_level_flags)
+        return approxs_list, inside_flag_list
+
+    def get_anchors(self,
+                    featmap_sizes,
+                    shape_preds,
+                    loc_preds,
+                    img_metas,
+                    use_loc_filter=False,
+                    device='cuda'):
+        """Get squares according to feature map sizes and guided anchors.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            shape_preds (list[tensor]): Multi-level shape predictions.
+            loc_preds (list[tensor]): Multi-level location predictions.
+            img_metas (list[dict]): Image meta info.
+            use_loc_filter (bool): Use loc filter or not.
+            device (torch.device | str): device for returned tensors
+
+        Returns:
+            tuple: square approxs of each image, guided anchors of each image,
+                loc masks of each image
+        """
+        num_imgs = len(img_metas)
+        num_levels = len(featmap_sizes)
+
+        # since feature map sizes of all images are the same, we only compute
+        # squares for one time
+        multi_level_squares = self.square_anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+        squares_list = [multi_level_squares for _ in range(num_imgs)]
+
+        # for each image, we compute multi level guided anchors
+        guided_anchors_list = []
+        loc_mask_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_guided_anchors = []
+            multi_level_loc_mask = []
+            for i in range(num_levels):
+                squares = squares_list[img_id][i]
+                shape_pred = shape_preds[i][img_id]
+                loc_pred = loc_preds[i][img_id]
+                guided_anchors, loc_mask = self._get_guided_anchors_single(
+                    squares,
+                    shape_pred,
+                    loc_pred,
+                    use_loc_filter=use_loc_filter)
+                multi_level_guided_anchors.append(guided_anchors)
+                multi_level_loc_mask.append(loc_mask)
+            guided_anchors_list.append(multi_level_guided_anchors)
+            loc_mask_list.append(multi_level_loc_mask)
+        return squares_list, guided_anchors_list, loc_mask_list
+
+    def _get_guided_anchors_single(self,
+                                   squares,
+                                   shape_pred,
+                                   loc_pred,
+                                   use_loc_filter=False):
+        """Get guided anchors and loc masks for a single level.
+
+        Args:
+            square (tensor): Squares of a single level.
+            shape_pred (tensor): Shape predections of a single level.
+            loc_pred (tensor): Loc predections of a single level.
+            use_loc_filter (list[tensor]): Use loc filter or not.
+
+        Returns:
+            tuple: guided anchors, location masks
+        """
+        # calculate location filtering mask
+        loc_pred = loc_pred.sigmoid().detach()
+        if use_loc_filter:
+            loc_mask = loc_pred >= self.loc_filter_thr
+        else:
+            loc_mask = loc_pred >= 0.0
+        mask = loc_mask.permute(1, 2, 0).expand(-1, -1, self.num_anchors)
+        mask = mask.contiguous().view(-1)
+        # calculate guided anchors
+        squares = squares[mask]
+        anchor_deltas = shape_pred.permute(1, 2, 0).contiguous().view(
+            -1, 2).detach()[mask]
+        bbox_deltas = anchor_deltas.new_full(squares.size(), 0)
+        bbox_deltas[:, 2:] = anchor_deltas
+        guided_anchors = self.anchor_coder.decode(
+            squares, bbox_deltas, wh_ratio_clip=1e-6)
+        return guided_anchors, mask
+
+    def ga_loc_targets(self, gt_bboxes_list, featmap_sizes):
+        """Compute location targets for guided anchoring.
+
+        Each feature map is divided into positive, negative and ignore regions.
+        - positive regions: target 1, weight 1
+        - ignore regions: target 0, weight 0
+        - negative regions: target 0, weight 0.1
+
+        Args:
+            gt_bboxes_list (list[Tensor]): Gt bboxes of each image.
+            featmap_sizes (list[tuple]): Multi level sizes of each feature
+                maps.
+
+        Returns:
+            tuple
+        """
+        anchor_scale = self.approx_anchor_generator.octave_base_scale
+        anchor_strides = self.approx_anchor_generator.strides
+        # Currently only supports same stride in x and y direction.
+        for stride in anchor_strides:
+            assert (stride[0] == stride[1])
+        anchor_strides = [stride[0] for stride in anchor_strides]
+
+        center_ratio = self.train_cfg.center_ratio
+        ignore_ratio = self.train_cfg.ignore_ratio
+        img_per_gpu = len(gt_bboxes_list)
+        num_lvls = len(featmap_sizes)
+        r1 = (1 - center_ratio) / 2
+        r2 = (1 - ignore_ratio) / 2
+        all_loc_targets = []
+        all_loc_weights = []
+        all_ignore_map = []
+        for lvl_id in range(num_lvls):
+            h, w = featmap_sizes[lvl_id]
+            loc_targets = torch.zeros(
+                img_per_gpu,
+                1,
+                h,
+                w,
+                device=gt_bboxes_list[0].device,
+                dtype=torch.float32)
+            loc_weights = torch.full_like(loc_targets, -1)
+            ignore_map = torch.zeros_like(loc_targets)
+            all_loc_targets.append(loc_targets)
+            all_loc_weights.append(loc_weights)
+            all_ignore_map.append(ignore_map)
+        for img_id in range(img_per_gpu):
+            gt_bboxes = gt_bboxes_list[img_id]
+            scale = torch.sqrt((gt_bboxes[:, 2] - gt_bboxes[:, 0]) *
+                               (gt_bboxes[:, 3] - gt_bboxes[:, 1]))
+            min_anchor_size = scale.new_full(
+                (1, ), float(anchor_scale * anchor_strides[0]))
+            # assign gt bboxes to different feature levels w.r.t. their scales
+            target_lvls = torch.floor(
+                torch.log2(scale) - torch.log2(min_anchor_size) + 0.5)
+            target_lvls = target_lvls.clamp(min=0, max=num_lvls - 1).long()
+            for gt_id in range(gt_bboxes.size(0)):
+                lvl = target_lvls[gt_id].item()
+                # rescaled to corresponding feature map
+                gt_ = gt_bboxes[gt_id, :4] / anchor_strides[lvl]
+                # calculate ignore regions
+                ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
+                    gt_, r2, featmap_sizes[lvl])
+                # calculate positive (center) regions
+                ctr_x1, ctr_y1, ctr_x2, ctr_y2 = calc_region(
+                    gt_, r1, featmap_sizes[lvl])
+                all_loc_targets[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
+                                     ctr_x1:ctr_x2 + 1] = 1
+                all_loc_weights[lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
+                                     ignore_x1:ignore_x2 + 1] = 0
+                all_loc_weights[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
+                                     ctr_x1:ctr_x2 + 1] = 1
+                # calculate ignore map on nearby low level feature
+                if lvl > 0:
+                    d_lvl = lvl - 1
+                    # rescaled to corresponding feature map
+                    gt_ = gt_bboxes[gt_id, :4] / anchor_strides[d_lvl]
+                    ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
+                        gt_, r2, featmap_sizes[d_lvl])
+                    all_ignore_map[d_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
+                                          ignore_x1:ignore_x2 + 1] = 1
+                # calculate ignore map on nearby high level feature
+                if lvl < num_lvls - 1:
+                    u_lvl = lvl + 1
+                    # rescaled to corresponding feature map
+                    gt_ = gt_bboxes[gt_id, :4] / anchor_strides[u_lvl]
+                    ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
+                        gt_, r2, featmap_sizes[u_lvl])
+                    all_ignore_map[u_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
+                                          ignore_x1:ignore_x2 + 1] = 1
+        for lvl_id in range(num_lvls):
+            # ignore negative regions w.r.t. ignore map
+            all_loc_weights[lvl_id][(all_loc_weights[lvl_id] < 0)
+                                    & (all_ignore_map[lvl_id] > 0)] = 0
+            # set negative regions with weight 0.1
+            all_loc_weights[lvl_id][all_loc_weights[lvl_id] < 0] = 0.1
+        # loc average factor to balance loss
+        loc_avg_factor = sum(
+            [t.size(0) * t.size(-1) * t.size(-2)
+             for t in all_loc_targets]) / 200
+        return all_loc_targets, all_loc_weights, loc_avg_factor
+
+    def _ga_shape_target_single(self,
+                                flat_approxs,
+                                inside_flags,
+                                flat_squares,
+                                gt_bboxes,
+                                gt_bboxes_ignore,
+                                img_meta,
+                                unmap_outputs=True):
+        """Compute guided anchoring targets.
+
+        This function returns sampled anchors and gt bboxes directly
+        rather than calculates regression targets.
+
+        Args:
+            flat_approxs (Tensor): flat approxs of a single image,
+                shape (n, 4)
+            inside_flags (Tensor): inside flags of a single image,
+                shape (n, ).
+            flat_squares (Tensor): flat squares of a single image,
+                shape (approxs_per_octave * n, 4)
+            gt_bboxes (Tensor): Ground truth bboxes of a single image.
+            img_meta (dict): Meta info of a single image.
+            approxs_per_octave (int): number of approxs per octave
+            cfg (dict): RPN train configs.
+            unmap_outputs (bool): unmap outputs or not.
+
+        Returns:
+            tuple
+        """
+        if not inside_flags.any():
+            return (None, ) * 5
+        # assign gt and sample anchors
+        expand_inside_flags = inside_flags[:, None].expand(
+            -1, self.approxs_per_octave).reshape(-1)
+        approxs = flat_approxs[expand_inside_flags, :]
+        squares = flat_squares[inside_flags, :]
+
+        assign_result = self.ga_assigner.assign(approxs, squares,
+                                                self.approxs_per_octave,
+                                                gt_bboxes, gt_bboxes_ignore)
+        sampling_result = self.ga_sampler.sample(assign_result, squares,
+                                                 gt_bboxes)
+
+        bbox_anchors = torch.zeros_like(squares)
+        bbox_gts = torch.zeros_like(squares)
+        bbox_weights = torch.zeros_like(squares)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            bbox_anchors[pos_inds, :] = sampling_result.pos_bboxes
+            bbox_gts[pos_inds, :] = sampling_result.pos_gt_bboxes
+            bbox_weights[pos_inds, :] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_squares.size(0)
+            bbox_anchors = unmap(bbox_anchors, num_total_anchors, inside_flags)
+            bbox_gts = unmap(bbox_gts, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+
+        return (bbox_anchors, bbox_gts, bbox_weights, pos_inds, neg_inds)
+
+    def ga_shape_targets(self,
+                         approx_list,
+                         inside_flag_list,
+                         square_list,
+                         gt_bboxes_list,
+                         img_metas,
+                         gt_bboxes_ignore_list=None,
+                         unmap_outputs=True):
+        """Compute guided anchoring targets.
+
+        Args:
+            approx_list (list[list]): Multi level approxs of each image.
+            inside_flag_list (list[list]): Multi level inside flags of each
+                image.
+            square_list (list[list]): Multi level squares of each image.
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            img_metas (list[dict]): Meta info of each image.
+            gt_bboxes_ignore_list (list[Tensor]): ignore list of gt bboxes.
+            unmap_outputs (bool): unmap outputs or not.
+
+        Returns:
+            tuple
+        """
+        num_imgs = len(img_metas)
+        assert len(approx_list) == len(inside_flag_list) == len(
+            square_list) == num_imgs
+        # anchor number of multi levels
+        num_level_squares = [squares.size(0) for squares in square_list[0]]
+        # concat all level anchors and flags to a single tensor
+        inside_flag_flat_list = []
+        approx_flat_list = []
+        square_flat_list = []
+        for i in range(num_imgs):
+            assert len(square_list[i]) == len(inside_flag_list[i])
+            inside_flag_flat_list.append(torch.cat(inside_flag_list[i]))
+            approx_flat_list.append(torch.cat(approx_list[i]))
+            square_flat_list.append(torch.cat(square_list[i]))
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        (all_bbox_anchors, all_bbox_gts, all_bbox_weights, pos_inds_list,
+         neg_inds_list) = multi_apply(
+             self._ga_shape_target_single,
+             approx_flat_list,
+             inside_flag_flat_list,
+             square_flat_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             img_metas,
+             unmap_outputs=unmap_outputs)
+        # no valid anchors
+        if any([bbox_anchors is None for bbox_anchors in all_bbox_anchors]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        bbox_anchors_list = images_to_levels(all_bbox_anchors,
+                                             num_level_squares)
+        bbox_gts_list = images_to_levels(all_bbox_gts, num_level_squares)
+        bbox_weights_list = images_to_levels(all_bbox_weights,
+                                             num_level_squares)
+        return (bbox_anchors_list, bbox_gts_list, bbox_weights_list,
+                num_total_pos, num_total_neg)
+
+    def loss_shape_single(self, shape_pred, bbox_anchors, bbox_gts,
+                          anchor_weights, anchor_total_num):
+        shape_pred = shape_pred.permute(0, 2, 3, 1).contiguous().view(-1, 2)
+        bbox_anchors = bbox_anchors.contiguous().view(-1, 4)
+        bbox_gts = bbox_gts.contiguous().view(-1, 4)
+        anchor_weights = anchor_weights.contiguous().view(-1, 4)
+        bbox_deltas = bbox_anchors.new_full(bbox_anchors.size(), 0)
+        bbox_deltas[:, 2:] += shape_pred
+        # filter out negative samples to speed-up weighted_bounded_iou_loss
+        inds = torch.nonzero(
+            anchor_weights[:, 0] > 0, as_tuple=False).squeeze(1)
+        bbox_deltas_ = bbox_deltas[inds]
+        bbox_anchors_ = bbox_anchors[inds]
+        bbox_gts_ = bbox_gts[inds]
+        anchor_weights_ = anchor_weights[inds]
+        pred_anchors_ = self.anchor_coder.decode(
+            bbox_anchors_, bbox_deltas_, wh_ratio_clip=1e-6)
+        loss_shape = self.loss_shape(
+            pred_anchors_,
+            bbox_gts_,
+            anchor_weights_,
+            avg_factor=anchor_total_num)
+        return loss_shape
+
+    def loss_loc_single(self, loc_pred, loc_target, loc_weight,
+                        loc_avg_factor):
+        loss_loc = self.loss_loc(
+            loc_pred.reshape(-1, 1),
+            loc_target.reshape(-1).long(),
+            loc_weight.reshape(-1),
+            avg_factor=loc_avg_factor)
+        return loss_loc
+
+    @force_fp32(
+        apply_to=('cls_scores', 'bbox_preds', 'shape_preds', 'loc_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             shape_preds,
+             loc_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.approx_anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        # get loc targets
+        loc_targets, loc_weights, loc_avg_factor = self.ga_loc_targets(
+            gt_bboxes, featmap_sizes)
+
+        # get sampled approxes
+        approxs_list, inside_flag_list = self.get_sampled_approxs(
+            featmap_sizes, img_metas, device=device)
+        # get squares and guided anchors
+        squares_list, guided_anchors_list, _ = self.get_anchors(
+            featmap_sizes, shape_preds, loc_preds, img_metas, device=device)
+
+        # get shape targets
+        shape_targets = self.ga_shape_targets(approxs_list, inside_flag_list,
+                                              squares_list, gt_bboxes,
+                                              img_metas)
+        if shape_targets is None:
+            return None
+        (bbox_anchors_list, bbox_gts_list, anchor_weights_list, anchor_fg_num,
+         anchor_bg_num) = shape_targets
+        anchor_total_num = (
+            anchor_fg_num if not self.ga_sampling else anchor_fg_num +
+            anchor_bg_num)
+
+        # get anchor targets
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            guided_anchors_list,
+            inside_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+        num_total_samples = (
+            num_total_pos + num_total_neg if self.sampling else num_total_pos)
+
+        # anchor number of multi levels
+        num_level_anchors = [
+            anchors.size(0) for anchors in guided_anchors_list[0]
+        ]
+        # concat all level anchors to a single tensor
+        concat_anchor_list = []
+        for i in range(len(guided_anchors_list)):
+            concat_anchor_list.append(torch.cat(guided_anchors_list[i]))
+        all_anchor_list = images_to_levels(concat_anchor_list,
+                                           num_level_anchors)
+
+        # get classification and bbox regression losses
+        losses_cls, losses_bbox = multi_apply(
+            self.loss_single,
+            cls_scores,
+            bbox_preds,
+            all_anchor_list,
+            labels_list,
+            label_weights_list,
+            bbox_targets_list,
+            bbox_weights_list,
+            num_total_samples=num_total_samples)
+
+        # get anchor location loss
+        losses_loc = []
+        for i in range(len(loc_preds)):
+            loss_loc = self.loss_loc_single(
+                loc_preds[i],
+                loc_targets[i],
+                loc_weights[i],
+                loc_avg_factor=loc_avg_factor)
+            losses_loc.append(loss_loc)
+
+        # get anchor shape loss
+        losses_shape = []
+        for i in range(len(shape_preds)):
+            loss_shape = self.loss_shape_single(
+                shape_preds[i],
+                bbox_anchors_list[i],
+                bbox_gts_list[i],
+                anchor_weights_list[i],
+                anchor_total_num=anchor_total_num)
+            losses_shape.append(loss_shape)
+
+        return dict(
+            loss_cls=losses_cls,
+            loss_bbox=losses_bbox,
+            loss_shape=losses_shape,
+            loss_loc=losses_loc)
+
+    @force_fp32(
+        apply_to=('cls_scores', 'bbox_preds', 'shape_preds', 'loc_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   shape_preds,
+                   loc_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=False):
+        assert len(cls_scores) == len(bbox_preds) == len(shape_preds) == len(
+            loc_preds)
+        num_levels = len(cls_scores)
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        device = cls_scores[0].device
+        # get guided anchors
+        _, guided_anchors, loc_masks = self.get_anchors(
+            featmap_sizes,
+            shape_preds,
+            loc_preds,
+            img_metas,
+            use_loc_filter=not self.training,
+            device=device)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            guided_anchor_list = [
+                guided_anchors[img_id][i].detach() for i in range(num_levels)
+            ]
+            loc_mask_list = [
+                loc_masks[img_id][i].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self._get_bboxes_single(cls_score_list, bbox_pred_list,
+                                                guided_anchor_list,
+                                                loc_mask_list, img_shape,
+                                                scale_factor, cfg, rescale)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_anchors,
+                           mlvl_masks,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False):
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, anchors, mask in zip(cls_scores, bbox_preds,
+                                                       mlvl_anchors,
+                                                       mlvl_masks):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            # if no location is kept, end.
+            if mask.sum() == 0:
+                continue
+            # reshape scores and bbox_pred
+            cls_score = cls_score.permute(1, 2,
+                                          0).reshape(-1, self.cls_out_channels)
+            if self.use_sigmoid_cls:
+                scores = cls_score.sigmoid()
+            else:
+                scores = cls_score.softmax(-1)
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            # filter scores, bbox_pred w.r.t. mask.
+            # anchors are filtered in get_anchors() beforehand.
+            scores = scores[mask, :]
+            bbox_pred = bbox_pred[mask, :]
+            if scores.dim() == 0:
+                anchors = anchors.unsqueeze(0)
+                scores = scores.unsqueeze(0)
+                bbox_pred = bbox_pred.unsqueeze(0)
+            # filter anchors, bbox_pred, scores w.r.t. scores
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                if self.use_sigmoid_cls:
+                    max_scores, _ = scores.max(dim=1)
+                else:
+                    # remind that we set FG labels to [0, num_class-1]
+                    # since mmdet v2.0
+                    # BG cat_id: num_class
+                    max_scores, _ = scores[:, :-1].max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+            bboxes = self.bbox_coder.decode(
+                anchors, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        if self.use_sigmoid_cls:
+            # Add a dummy background class to the backend when using sigmoid
+            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+            # BG cat_id: num_class
+            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        # multi class NMS
+        det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                cfg.score_thr, cfg.nms,
+                                                cfg.max_per_img)
+        return det_bboxes, det_labels
diff --git a/detection/scrfd/mmdet/models/dense_heads/nasfcos_head.py b/detection/scrfd/mmdet/models/dense_heads/nasfcos_head.py
new file mode 100755
index 000000000..994ce0455
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/nasfcos_head.py
@@ -0,0 +1,75 @@
+import copy
+
+import torch.nn as nn
+from mmcv.cnn import (ConvModule, Scale, bias_init_with_prob,
+                      caffe2_xavier_init, normal_init)
+
+from mmdet.models.dense_heads.fcos_head import FCOSHead
+from ..builder import HEADS
+
+
+@HEADS.register_module()
+class NASFCOSHead(FCOSHead):
+    """Anchor-free head used in `NASFCOS <https://arxiv.org/abs/1906.04423>`_.
+
+    It is quite similar with FCOS head, except for the searched structure of
+    classification branch and bbox regression branch, where a structure of
+    "dconv3x3, conv3x3, dconv3x3, conv1x1" is utilized instead.
+    """
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        dconv3x3_config = dict(
+            type='DCNv2',
+            kernel_size=3,
+            use_bias=True,
+            deform_groups=2,
+            padding=1)
+        conv3x3_config = dict(type='Conv', kernel_size=3, padding=1)
+        conv1x1_config = dict(type='Conv', kernel_size=1)
+
+        self.arch_config = [
+            dconv3x3_config, conv3x3_config, dconv3x3_config, conv1x1_config
+        ]
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i, op_ in enumerate(self.arch_config):
+            op = copy.deepcopy(op_)
+            chn = self.in_channels if i == 0 else self.feat_channels
+            assert isinstance(op, dict)
+            use_bias = op.pop('use_bias', False)
+            padding = op.pop('padding', 0)
+            kernel_size = op.pop('kernel_size')
+            module = ConvModule(
+                chn,
+                self.feat_channels,
+                kernel_size,
+                stride=1,
+                padding=padding,
+                norm_cfg=self.norm_cfg,
+                bias=use_bias,
+                conv_cfg=op)
+
+            self.cls_convs.append(copy.deepcopy(module))
+            self.reg_convs.append(copy.deepcopy(module))
+
+        self.conv_cls = nn.Conv2d(
+            self.feat_channels, self.cls_out_channels, 3, padding=1)
+        self.conv_reg = nn.Conv2d(self.feat_channels, 4, 3, padding=1)
+        self.conv_centerness = nn.Conv2d(self.feat_channels, 1, 3, padding=1)
+
+        self.scales = nn.ModuleList([Scale(1.0) for _ in self.strides])
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        # retinanet_bias_init
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.conv_reg, std=0.01)
+        normal_init(self.conv_centerness, std=0.01)
+        normal_init(self.conv_cls, std=0.01, bias=bias_cls)
+
+        for branch in [self.cls_convs, self.reg_convs]:
+            for module in branch.modules():
+                if isinstance(module, ConvModule) \
+                        and isinstance(module.conv, nn.Conv2d):
+                    caffe2_xavier_init(module.conv)
diff --git a/detection/scrfd/mmdet/models/dense_heads/paa_head.py b/detection/scrfd/mmdet/models/dense_heads/paa_head.py
new file mode 100755
index 000000000..db28abc7e
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/paa_head.py
@@ -0,0 +1,653 @@
+import numpy as np
+import torch
+from mmcv.runner import force_fp32
+
+from mmdet.core import multi_apply, multiclass_nms
+from mmdet.core.bbox.iou_calculators import bbox_overlaps
+from mmdet.models import HEADS
+from mmdet.models.dense_heads import ATSSHead
+
+EPS = 1e-12
+try:
+    import sklearn.mixture as skm
+except ImportError:
+    skm = None
+
+
+def levels_to_images(mlvl_tensor):
+    """Concat multi-level feature maps by image.
+
+    [feature_level0, feature_level1...] -> [feature_image0, feature_image1...]
+    Convert the shape of each element in mlvl_tensor from (N, C, H, W) to
+    (N, H*W , C), then split the element to N elements with shape (H*W, C), and
+    concat elements in same image of all level along first dimension.
+
+    Args:
+        mlvl_tensor (list[torch.Tensor]): list of Tensor which collect from
+            corresponding level. Each element is of shape (N, C, H, W)
+
+    Returns:
+        list[torch.Tensor]: A list that contains N tensors and each tensor is
+            of shape (num_elements, C)
+    """
+    batch_size = mlvl_tensor[0].size(0)
+    batch_list = [[] for _ in range(batch_size)]
+    channels = mlvl_tensor[0].size(1)
+    for t in mlvl_tensor:
+        t = t.permute(0, 2, 3, 1)
+        t = t.view(batch_size, -1, channels).contiguous()
+        for img in range(batch_size):
+            batch_list[img].append(t[img])
+    return [torch.cat(item, 0) for item in batch_list]
+
+
+@HEADS.register_module()
+class PAAHead(ATSSHead):
+    """Head of PAAAssignment: Probabilistic Anchor Assignment with IoU
+    Prediction for Object Detection.
+
+    Code is modified from the `official github repo
+    <https://github.com/kkhoot/PAA/blob/master/paa_core
+    /modeling/rpn/paa/loss.py>`_.
+
+    More details can be found in the `paper
+    <https://arxiv.org/abs/2007.08103>`_ .
+
+    Args:
+        topk (int): Select topk samples with smallest loss in
+            each level.
+        score_voting (bool): Whether to use score voting in post-process.
+        covariance_type : String describing the type of covariance parameters
+            to be used in :class:`sklearn.mixture.GaussianMixture`.
+            It must be one of:
+
+            - 'full': each component has its own general covariance matrix
+            - 'tied': all components share the same general covariance matrix
+            - 'diag': each component has its own diagonal covariance matrix
+            - 'spherical': each component has its own single variance
+            Default: 'diag'. From 'full' to 'spherical', the gmm fitting
+            process is faster yet the performance could be influenced. For most
+            cases, 'diag' should be a good choice.
+    """
+
+    def __init__(self,
+                 *args,
+                 topk=9,
+                 score_voting=True,
+                 covariance_type='diag',
+                 **kwargs):
+        # topk used in paa reassign process
+        self.topk = topk
+        self.with_score_voting = score_voting
+        self.covariance_type = covariance_type
+        super(PAAHead, self).__init__(*args, **kwargs)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'iou_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             iou_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            iou_preds (list[Tensor]): iou_preds for each scale
+                level with shape (N, num_anchors * 1, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): Specify which bounding
+                boxes can be ignored when are computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss gmm_assignment.
+        """
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels,
+        )
+        (labels, labels_weight, bboxes_target, bboxes_weight, pos_inds,
+         pos_gt_index) = cls_reg_targets
+        cls_scores = levels_to_images(cls_scores)
+        cls_scores = [
+            item.reshape(-1, self.cls_out_channels) for item in cls_scores
+        ]
+        bbox_preds = levels_to_images(bbox_preds)
+        bbox_preds = [item.reshape(-1, 4) for item in bbox_preds]
+        iou_preds = levels_to_images(iou_preds)
+        iou_preds = [item.reshape(-1, 1) for item in iou_preds]
+        pos_losses_list, = multi_apply(self.get_pos_loss, anchor_list,
+                                       cls_scores, bbox_preds, labels,
+                                       labels_weight, bboxes_target,
+                                       bboxes_weight, pos_inds)
+
+        with torch.no_grad():
+            labels, label_weights, bbox_weights, num_pos = multi_apply(
+                self.paa_reassign,
+                pos_losses_list,
+                labels,
+                labels_weight,
+                bboxes_weight,
+                pos_inds,
+                pos_gt_index,
+                anchor_list,
+            )
+            num_pos = sum(num_pos)
+        # convert all tensor list to a flatten tensor
+        cls_scores = torch.cat(cls_scores, 0).view(-1, cls_scores[0].size(-1))
+        bbox_preds = torch.cat(bbox_preds, 0).view(-1, bbox_preds[0].size(-1))
+        iou_preds = torch.cat(iou_preds, 0).view(-1, iou_preds[0].size(-1))
+        labels = torch.cat(labels, 0).view(-1)
+        flatten_anchors = torch.cat(
+            [torch.cat(item, 0) for item in anchor_list])
+        labels_weight = torch.cat(labels_weight, 0).view(-1)
+        bboxes_target = torch.cat(bboxes_target,
+                                  0).view(-1, bboxes_target[0].size(-1))
+
+        pos_inds_flatten = ((labels >= 0)
+                            &
+                            (labels < self.num_classes)).nonzero().reshape(-1)
+
+        losses_cls = self.loss_cls(
+            cls_scores,
+            labels,
+            labels_weight,
+            avg_factor=max(num_pos, len(img_metas)))  # avoid num_pos=0
+        if num_pos:
+            pos_bbox_pred = self.bbox_coder.decode(
+                flatten_anchors[pos_inds_flatten],
+                bbox_preds[pos_inds_flatten])
+            pos_bbox_target = bboxes_target[pos_inds_flatten]
+            iou_target = bbox_overlaps(
+                pos_bbox_pred.detach(), pos_bbox_target, is_aligned=True)
+            losses_iou = self.loss_centerness(
+                iou_preds[pos_inds_flatten],
+                iou_target.unsqueeze(-1),
+                avg_factor=num_pos)
+            losses_bbox = self.loss_bbox(
+                pos_bbox_pred,
+                pos_bbox_target,
+                iou_target.clamp(min=EPS),
+                avg_factor=iou_target.sum())
+        else:
+            losses_iou = iou_preds.sum() * 0
+            losses_bbox = bbox_preds.sum() * 0
+
+        return dict(
+            loss_cls=losses_cls, loss_bbox=losses_bbox, loss_iou=losses_iou)
+
+    def get_pos_loss(self, anchors, cls_score, bbox_pred, label, label_weight,
+                     bbox_target, bbox_weight, pos_inds):
+        """Calculate loss of all potential positive samples obtained from first
+        match process.
+
+        Args:
+            anchors (list[Tensor]): Anchors of each scale.
+            cls_score (Tensor): Box scores of single image with shape
+                (num_anchors, num_classes)
+            bbox_pred (Tensor): Box energies / deltas of single image
+                with shape (num_anchors, 4)
+            label (Tensor): classification target of each anchor with
+                shape (num_anchors,)
+            label_weight (Tensor): Classification loss weight of each
+                anchor with shape (num_anchors).
+            bbox_target (dict): Regression target of each anchor with
+                shape (num_anchors, 4).
+            bbox_weight (Tensor): Bbox weight of each anchor with shape
+                (num_anchors, 4).
+            pos_inds (Tensor): Index of all positive samples got from
+                first assign process.
+
+        Returns:
+            Tensor: Losses of all positive samples in single image.
+        """
+        if not len(pos_inds):
+            return cls_score.new([]),
+        anchors_all_level = torch.cat(anchors, 0)
+        pos_scores = cls_score[pos_inds]
+        pos_bbox_pred = bbox_pred[pos_inds]
+        pos_label = label[pos_inds]
+        pos_label_weight = label_weight[pos_inds]
+        pos_bbox_target = bbox_target[pos_inds]
+        pos_bbox_weight = bbox_weight[pos_inds]
+        pos_anchors = anchors_all_level[pos_inds]
+        pos_bbox_pred = self.bbox_coder.decode(pos_anchors, pos_bbox_pred)
+
+        # to keep loss dimension
+        loss_cls = self.loss_cls(
+            pos_scores,
+            pos_label,
+            pos_label_weight,
+            avg_factor=self.loss_cls.loss_weight,
+            reduction_override='none')
+
+        loss_bbox = self.loss_bbox(
+            pos_bbox_pred,
+            pos_bbox_target,
+            pos_bbox_weight,
+            avg_factor=self.loss_cls.loss_weight,
+            reduction_override='none')
+
+        loss_cls = loss_cls.sum(-1)
+        pos_loss = loss_bbox + loss_cls
+        return pos_loss,
+
+    def paa_reassign(self, pos_losses, label, label_weight, bbox_weight,
+                     pos_inds, pos_gt_inds, anchors):
+        """Fit loss to GMM distribution and separate positive, ignore, negative
+        samples again with GMM model.
+
+        Args:
+            pos_losses (Tensor): Losses of all positive samples in
+                single image.
+            label (Tensor): classification target of each anchor with
+                shape (num_anchors,)
+            label_weight (Tensor): Classification loss weight of each
+                anchor with shape (num_anchors).
+            bbox_weight (Tensor): Bbox weight of each anchor with shape
+                (num_anchors, 4).
+            pos_inds (Tensor): Index of all positive samples got from
+                first assign process.
+            pos_gt_inds (Tensor): Gt_index of all positive samples got
+                from first assign process.
+            anchors (list[Tensor]): Anchors of each scale.
+
+        Returns:
+            tuple: Usually returns a tuple containing learning targets.
+
+                - label (Tensor): classification target of each anchor after
+                  paa assign, with shape (num_anchors,)
+                - label_weight (Tensor): Classification loss weight of each
+                  anchor after paa assign, with shape (num_anchors).
+                - bbox_weight (Tensor): Bbox weight of each anchor with shape
+                  (num_anchors, 4).
+                - num_pos (int): The number of positive samples after paa
+                  assign.
+        """
+        if not len(pos_inds):
+            return label, label_weight, bbox_weight, 0
+
+        num_gt = pos_gt_inds.max() + 1
+        num_level = len(anchors)
+        num_anchors_each_level = [item.size(0) for item in anchors]
+        num_anchors_each_level.insert(0, 0)
+        inds_level_interval = np.cumsum(num_anchors_each_level)
+        pos_level_mask = []
+        for i in range(num_level):
+            mask = (pos_inds >= inds_level_interval[i]) & (
+                pos_inds < inds_level_interval[i + 1])
+            pos_level_mask.append(mask)
+        pos_inds_after_paa = [label.new_tensor([])]
+        ignore_inds_after_paa = [label.new_tensor([])]
+        for gt_ind in range(num_gt):
+            pos_inds_gmm = []
+            pos_loss_gmm = []
+            gt_mask = pos_gt_inds == gt_ind
+            for level in range(num_level):
+                level_mask = pos_level_mask[level]
+                level_gt_mask = level_mask & gt_mask
+                value, topk_inds = pos_losses[level_gt_mask].topk(
+                    min(level_gt_mask.sum(), self.topk), largest=False)
+                pos_inds_gmm.append(pos_inds[level_gt_mask][topk_inds])
+                pos_loss_gmm.append(value)
+            pos_inds_gmm = torch.cat(pos_inds_gmm)
+            pos_loss_gmm = torch.cat(pos_loss_gmm)
+            # fix gmm need at least two sample
+            if len(pos_inds_gmm) < 2:
+                continue
+            device = pos_inds_gmm.device
+            pos_loss_gmm, sort_inds = pos_loss_gmm.sort()
+            pos_inds_gmm = pos_inds_gmm[sort_inds]
+            pos_loss_gmm = pos_loss_gmm.view(-1, 1).cpu().numpy()
+            min_loss, max_loss = pos_loss_gmm.min(), pos_loss_gmm.max()
+            means_init = np.array([min_loss, max_loss]).reshape(2, 1)
+            weights_init = np.array([0.5, 0.5])
+            precisions_init = np.array([1.0, 1.0]).reshape(2, 1, 1)  # full
+            if self.covariance_type == 'spherical':
+                precisions_init = precisions_init.reshape(2)
+            elif self.covariance_type == 'diag':
+                precisions_init = precisions_init.reshape(2, 1)
+            elif self.covariance_type == 'tied':
+                precisions_init = np.array([[1.0]])
+            if skm is None:
+                raise ImportError('Please run "pip install sklearn" '
+                                  'to install sklearn first.')
+            gmm = skm.GaussianMixture(
+                2,
+                weights_init=weights_init,
+                means_init=means_init,
+                precisions_init=precisions_init,
+                covariance_type=self.covariance_type)
+            gmm.fit(pos_loss_gmm)
+            gmm_assignment = gmm.predict(pos_loss_gmm)
+            scores = gmm.score_samples(pos_loss_gmm)
+            gmm_assignment = torch.from_numpy(gmm_assignment).to(device)
+            scores = torch.from_numpy(scores).to(device)
+
+            pos_inds_temp, ignore_inds_temp = self.gmm_separation_scheme(
+                gmm_assignment, scores, pos_inds_gmm)
+            pos_inds_after_paa.append(pos_inds_temp)
+            ignore_inds_after_paa.append(ignore_inds_temp)
+
+        pos_inds_after_paa = torch.cat(pos_inds_after_paa)
+        ignore_inds_after_paa = torch.cat(ignore_inds_after_paa)
+        reassign_mask = (pos_inds.unsqueeze(1) != pos_inds_after_paa).all(1)
+        reassign_ids = pos_inds[reassign_mask]
+        label[reassign_ids] = self.num_classes
+        label_weight[ignore_inds_after_paa] = 0
+        bbox_weight[reassign_ids] = 0
+        num_pos = len(pos_inds_after_paa)
+        return label, label_weight, bbox_weight, num_pos
+
+    def gmm_separation_scheme(self, gmm_assignment, scores, pos_inds_gmm):
+        """A general separation scheme for gmm model.
+
+        It separates a GMM distribution of candidate samples into three
+        parts, 0 1 and uncertain areas, and you can implement other
+        separation schemes by rewriting this function.
+
+        Args:
+            gmm_assignment (Tensor): The prediction of GMM which is of shape
+                (num_samples,). The 0/1 value indicates the distribution
+                that each sample comes from.
+            scores (Tensor): The probability of sample coming from the
+                fit GMM distribution. The tensor is of shape (num_samples,).
+            pos_inds_gmm (Tensor): All the indexes of samples which are used
+                to fit GMM model. The tensor is of shape (num_samples,)
+
+        Returns:
+            tuple[Tensor]: The indices of positive and ignored samples.
+
+                - pos_inds_temp (Tensor): Indices of positive samples.
+                - ignore_inds_temp (Tensor): Indices of ignore samples.
+        """
+        # The implementation is (c) in Fig.3 in origin paper intead of (b).
+        # You can refer to issues such as
+        # https://github.com/kkhoot/PAA/issues/8 and
+        # https://github.com/kkhoot/PAA/issues/9.
+        fgs = gmm_assignment == 0
+        pos_inds_temp = fgs.new_tensor([], dtype=torch.long)
+        ignore_inds_temp = fgs.new_tensor([], dtype=torch.long)
+        if fgs.nonzero().numel():
+            _, pos_thr_ind = scores[fgs].topk(1)
+            pos_inds_temp = pos_inds_gmm[fgs][:pos_thr_ind + 1]
+            ignore_inds_temp = pos_inds_gmm.new_tensor([])
+        return pos_inds_temp, ignore_inds_temp
+
+    def get_targets(
+        self,
+        anchor_list,
+        valid_flag_list,
+        gt_bboxes_list,
+        img_metas,
+        gt_bboxes_ignore_list=None,
+        gt_labels_list=None,
+        label_channels=1,
+        unmap_outputs=True,
+    ):
+        """Get targets for PAA head.
+
+        This method is almost the same as `AnchorHead.get_targets()`. We direct
+        return the results from _get_targets_single instead map it to levels
+        by images_to_levels function.
+
+        Args:
+            anchor_list (list[list[Tensor]]): Multi level anchors of each
+                image. The outer list indicates images, and the inner list
+                corresponds to feature levels of the image. Each element of
+                the inner list is a tensor of shape (num_anchors, 4).
+            valid_flag_list (list[list[Tensor]]): Multi level valid flags of
+                each image. The outer list indicates images, and the inner list
+                corresponds to feature levels of the image. Each element of
+                the inner list is a tensor of shape (num_anchors, )
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            img_metas (list[dict]): Meta info of each image.
+            gt_bboxes_ignore_list (list[Tensor]): Ground truth bboxes to be
+                ignored.
+            gt_labels_list (list[Tensor]): Ground truth labels of each box.
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple: Usually returns a tuple containing learning targets.
+
+                - labels (list[Tensor]): Labels of all anchors, each with
+                    shape (num_anchors,).
+                - label_weights (list[Tensor]): Label weights of all anchor.
+                    each with shape (num_anchors,).
+                - bbox_targets (list[Tensor]): BBox targets of all anchors.
+                    each with shape (num_anchors, 4).
+                - bbox_weights (list[Tensor]): BBox weights of all anchors.
+                    each with shape (num_anchors, 4).
+                - pos_inds (list[Tensor]): Contains all index of positive
+                    sample in all anchor.
+                - gt_inds (list[Tensor]): Contains all gt_index of positive
+                    sample in all anchor.
+        """
+
+        num_imgs = len(img_metas)
+        assert len(anchor_list) == len(valid_flag_list) == num_imgs
+        concat_anchor_list = []
+        concat_valid_flag_list = []
+        for i in range(num_imgs):
+            assert len(anchor_list[i]) == len(valid_flag_list[i])
+            concat_anchor_list.append(torch.cat(anchor_list[i]))
+            concat_valid_flag_list.append(torch.cat(valid_flag_list[i]))
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        results = multi_apply(
+            self._get_targets_single,
+            concat_anchor_list,
+            concat_valid_flag_list,
+            gt_bboxes_list,
+            gt_bboxes_ignore_list,
+            gt_labels_list,
+            img_metas,
+            label_channels=label_channels,
+            unmap_outputs=unmap_outputs)
+
+        (labels, label_weights, bbox_targets, bbox_weights, valid_pos_inds,
+         valid_neg_inds, sampling_result) = results
+
+        # Due to valid flag of anchors, we have to calculate the real pos_inds
+        # in origin anchor set.
+        pos_inds = []
+        for i, single_labels in enumerate(labels):
+            pos_mask = (0 <= single_labels) & (
+                single_labels < self.num_classes)
+            pos_inds.append(pos_mask.nonzero().view(-1))
+
+        gt_inds = [item.pos_assigned_gt_inds for item in sampling_result]
+        return (labels, label_weights, bbox_targets, bbox_weights, pos_inds,
+                gt_inds)
+
+    def _get_targets_single(self,
+                            flat_anchors,
+                            valid_flags,
+                            gt_bboxes,
+                            gt_bboxes_ignore,
+                            gt_labels,
+                            img_meta,
+                            label_channels=1,
+                            unmap_outputs=True):
+        """Compute regression and classification targets for anchors in a
+        single image.
+
+        This method is same as `AnchorHead._get_targets_single()`.
+        """
+        assert unmap_outputs, 'We must map outputs back to the original' \
+            'set of anchors in PAAhead'
+        return super(ATSSHead, self)._get_targets_single(
+            flat_anchors,
+            valid_flags,
+            gt_bboxes,
+            gt_bboxes_ignore,
+            gt_labels,
+            img_meta,
+            label_channels=1,
+            unmap_outputs=True)
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           iou_preds,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into labeled boxes.
+
+        This method is almost same as `ATSSHead._get_bboxes_single()`.
+        We use sqrt(iou_preds * cls_scores) in NMS process instead of just
+        cls_scores. Besides, score voting is used when `` score_voting``
+        is set to True.
+        """
+        assert with_nms, 'PAA only supports "with_nms=True" now'
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        mlvl_iou_preds = []
+        for cls_score, bbox_pred, iou_preds, anchors in zip(
+                cls_scores, bbox_preds, iou_preds, mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            iou_preds = iou_preds.permute(1, 2, 0).reshape(-1).sigmoid()
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                max_scores, _ = (scores * iou_preds[:, None]).sqrt().max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+                iou_preds = iou_preds[topk_inds]
+
+            bboxes = self.bbox_coder.decode(
+                anchors, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+            mlvl_iou_preds.append(iou_preds)
+
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        # Add a dummy background class to the backend when using sigmoid
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        mlvl_iou_preds = torch.cat(mlvl_iou_preds)
+        mlvl_nms_scores = (mlvl_scores * mlvl_iou_preds[:, None]).sqrt()
+        det_bboxes, det_labels = multiclass_nms(
+            mlvl_bboxes,
+            mlvl_nms_scores,
+            cfg.score_thr,
+            cfg.nms,
+            cfg.max_per_img,
+            score_factors=None)
+        if self.with_score_voting:
+            det_bboxes, det_labels = self.score_voting(det_bboxes, det_labels,
+                                                       mlvl_bboxes,
+                                                       mlvl_nms_scores,
+                                                       cfg.score_thr)
+
+        return det_bboxes, det_labels
+
+    def score_voting(self, det_bboxes, det_labels, mlvl_bboxes,
+                     mlvl_nms_scores, score_thr):
+        """Implementation of score voting method works on each remaining boxes
+        after NMS procedure.
+
+        Args:
+            det_bboxes (Tensor): Remaining boxes after NMS procedure,
+                with shape (k, 5), each dimension means
+                (x1, y1, x2, y2, score).
+            det_labels (Tensor): The label of remaining boxes, with shape
+                (k, 1),Labels are 0-based.
+            mlvl_bboxes (Tensor): All boxes before the NMS procedure,
+                with shape (num_anchors,4).
+            mlvl_nms_scores (Tensor): The scores of all boxes which is used
+                in the NMS procedure, with shape (num_anchors, num_class)
+            mlvl_iou_preds (Tensot): The predictions of IOU of all boxes
+                before the NMS procedure, with shape (num_anchors, 1)
+            score_thr (float): The score threshold of bboxes.
+
+        Returns:
+            tuple: Usually returns a tuple containing voting results.
+
+                - det_bboxes_voted (Tensor): Remaining boxes after
+                    score voting procedure, with shape (k, 5), each
+                    dimension means (x1, y1, x2, y2, score).
+                - det_labels_voted (Tensor): Label of remaining bboxes
+                    after voting, with shape (num_anchors,).
+        """
+        candidate_mask = mlvl_nms_scores > score_thr
+        candidate_mask_nozeros = candidate_mask.nonzero()
+        candidate_inds = candidate_mask_nozeros[:, 0]
+        candidate_labels = candidate_mask_nozeros[:, 1]
+        candidate_bboxes = mlvl_bboxes[candidate_inds]
+        candidate_scores = mlvl_nms_scores[candidate_mask]
+        det_bboxes_voted = []
+        det_labels_voted = []
+        for cls in range(self.cls_out_channels):
+            candidate_cls_mask = candidate_labels == cls
+            if not candidate_cls_mask.any():
+                continue
+            candidate_cls_scores = candidate_scores[candidate_cls_mask]
+            candidate_cls_bboxes = candidate_bboxes[candidate_cls_mask]
+            det_cls_mask = det_labels == cls
+            det_cls_bboxes = det_bboxes[det_cls_mask].view(
+                -1, det_bboxes.size(-1))
+            det_candidate_ious = bbox_overlaps(det_cls_bboxes[:, :4],
+                                               candidate_cls_bboxes)
+            for det_ind in range(len(det_cls_bboxes)):
+                single_det_ious = det_candidate_ious[det_ind]
+                pos_ious_mask = single_det_ious > 0.01
+                pos_ious = single_det_ious[pos_ious_mask]
+                pos_bboxes = candidate_cls_bboxes[pos_ious_mask]
+                pos_scores = candidate_cls_scores[pos_ious_mask]
+                pis = (torch.exp(-(1 - pos_ious)**2 / 0.025) *
+                       pos_scores)[:, None]
+                voted_box = torch.sum(
+                    pis * pos_bboxes, dim=0) / torch.sum(
+                        pis, dim=0)
+                voted_score = det_cls_bboxes[det_ind][-1:][None, :]
+                det_bboxes_voted.append(
+                    torch.cat((voted_box[None, :], voted_score), dim=1))
+                det_labels_voted.append(cls)
+
+        det_bboxes_voted = torch.cat(det_bboxes_voted, dim=0)
+        det_labels_voted = det_labels.new_tensor(det_labels_voted)
+        return det_bboxes_voted, det_labels_voted
diff --git a/detection/scrfd/mmdet/models/dense_heads/pisa_retinanet_head.py b/detection/scrfd/mmdet/models/dense_heads/pisa_retinanet_head.py
new file mode 100755
index 000000000..bd87b9aeb
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/pisa_retinanet_head.py
@@ -0,0 +1,154 @@
+import torch
+from mmcv.runner import force_fp32
+
+from mmdet.core import images_to_levels
+from ..builder import HEADS
+from ..losses import carl_loss, isr_p
+from .retina_head import RetinaHead
+
+
+@HEADS.register_module()
+class PISARetinaHead(RetinaHead):
+    """PISA Retinanet Head.
+
+    The head owns the same structure with Retinanet Head, but differs in two
+        aspects:
+        1. Importance-based Sample Reweighting Positive (ISR-P) is applied to
+            change the positive loss weights.
+        2. Classification-aware regression loss is adopted as a third loss.
+    """
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes of each image
+                with shape (num_obj, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each image
+                with shape (num_obj, 4).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor]): Ignored gt bboxes of each image.
+                Default: None.
+
+        Returns:
+            dict: Loss dict, comprise classification loss, regression loss and
+                carl loss.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels,
+            return_sampling_results=True)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg, sampling_results_list) = cls_reg_targets
+        num_total_samples = (
+            num_total_pos + num_total_neg if self.sampling else num_total_pos)
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        # concat all level anchors and flags to a single tensor
+        concat_anchor_list = []
+        for i in range(len(anchor_list)):
+            concat_anchor_list.append(torch.cat(anchor_list[i]))
+        all_anchor_list = images_to_levels(concat_anchor_list,
+                                           num_level_anchors)
+
+        num_imgs = len(img_metas)
+        flatten_cls_scores = [
+            cls_score.permute(0, 2, 3, 1).reshape(num_imgs, -1, label_channels)
+            for cls_score in cls_scores
+        ]
+        flatten_cls_scores = torch.cat(
+            flatten_cls_scores, dim=1).reshape(-1,
+                                               flatten_cls_scores[0].size(-1))
+        flatten_bbox_preds = [
+            bbox_pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, 4)
+            for bbox_pred in bbox_preds
+        ]
+        flatten_bbox_preds = torch.cat(
+            flatten_bbox_preds, dim=1).view(-1, flatten_bbox_preds[0].size(-1))
+        flatten_labels = torch.cat(labels_list, dim=1).reshape(-1)
+        flatten_label_weights = torch.cat(
+            label_weights_list, dim=1).reshape(-1)
+        flatten_anchors = torch.cat(all_anchor_list, dim=1).reshape(-1, 4)
+        flatten_bbox_targets = torch.cat(
+            bbox_targets_list, dim=1).reshape(-1, 4)
+        flatten_bbox_weights = torch.cat(
+            bbox_weights_list, dim=1).reshape(-1, 4)
+
+        # Apply ISR-P
+        isr_cfg = self.train_cfg.get('isr', None)
+        if isr_cfg is not None:
+            all_targets = (flatten_labels, flatten_label_weights,
+                           flatten_bbox_targets, flatten_bbox_weights)
+            with torch.no_grad():
+                all_targets = isr_p(
+                    flatten_cls_scores,
+                    flatten_bbox_preds,
+                    all_targets,
+                    flatten_anchors,
+                    sampling_results_list,
+                    bbox_coder=self.bbox_coder,
+                    loss_cls=self.loss_cls,
+                    num_class=self.num_classes,
+                    **self.train_cfg.isr)
+            (flatten_labels, flatten_label_weights, flatten_bbox_targets,
+             flatten_bbox_weights) = all_targets
+
+        # For convenience we compute loss once instead separating by fpn level,
+        # so that we don't need to separate the weights by level again.
+        # The result should be the same
+        losses_cls = self.loss_cls(
+            flatten_cls_scores,
+            flatten_labels,
+            flatten_label_weights,
+            avg_factor=num_total_samples)
+        losses_bbox = self.loss_bbox(
+            flatten_bbox_preds,
+            flatten_bbox_targets,
+            flatten_bbox_weights,
+            avg_factor=num_total_samples)
+        loss_dict = dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
+
+        # CARL Loss
+        carl_cfg = self.train_cfg.get('carl', None)
+        if carl_cfg is not None:
+            loss_carl = carl_loss(
+                flatten_cls_scores,
+                flatten_labels,
+                flatten_bbox_preds,
+                flatten_bbox_targets,
+                self.loss_bbox,
+                **self.train_cfg.carl,
+                avg_factor=num_total_pos,
+                sigmoid=True,
+                num_class=self.num_classes)
+            loss_dict.update(loss_carl)
+
+        return loss_dict
diff --git a/detection/scrfd/mmdet/models/dense_heads/pisa_ssd_head.py b/detection/scrfd/mmdet/models/dense_heads/pisa_ssd_head.py
new file mode 100755
index 000000000..90ef3c83e
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/pisa_ssd_head.py
@@ -0,0 +1,139 @@
+import torch
+
+from mmdet.core import multi_apply
+from ..builder import HEADS
+from ..losses import CrossEntropyLoss, SmoothL1Loss, carl_loss, isr_p
+from .ssd_head import SSDHead
+
+
+# TODO: add loss evaluator for SSD
+@HEADS.register_module()
+class PISASSDHead(SSDHead):
+
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes of each image
+                with shape (num_obj, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each image
+                with shape (num_obj, 4).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor]): Ignored gt bboxes of each image.
+                Default: None.
+
+        Returns:
+            dict: Loss dict, comprise classification loss regression loss and
+                carl loss.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=1,
+            unmap_outputs=False,
+            return_sampling_results=True)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg, sampling_results_list) = cls_reg_targets
+
+        num_images = len(img_metas)
+        all_cls_scores = torch.cat([
+            s.permute(0, 2, 3, 1).reshape(
+                num_images, -1, self.cls_out_channels) for s in cls_scores
+        ], 1)
+        all_labels = torch.cat(labels_list, -1).view(num_images, -1)
+        all_label_weights = torch.cat(label_weights_list,
+                                      -1).view(num_images, -1)
+        all_bbox_preds = torch.cat([
+            b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
+            for b in bbox_preds
+        ], -2)
+        all_bbox_targets = torch.cat(bbox_targets_list,
+                                     -2).view(num_images, -1, 4)
+        all_bbox_weights = torch.cat(bbox_weights_list,
+                                     -2).view(num_images, -1, 4)
+
+        # concat all level anchors to a single tensor
+        all_anchors = []
+        for i in range(num_images):
+            all_anchors.append(torch.cat(anchor_list[i]))
+
+        isr_cfg = self.train_cfg.get('isr', None)
+        all_targets = (all_labels.view(-1), all_label_weights.view(-1),
+                       all_bbox_targets.view(-1,
+                                             4), all_bbox_weights.view(-1, 4))
+        # apply ISR-P
+        if isr_cfg is not None:
+            all_targets = isr_p(
+                all_cls_scores.view(-1, all_cls_scores.size(-1)),
+                all_bbox_preds.view(-1, 4),
+                all_targets,
+                torch.cat(all_anchors),
+                sampling_results_list,
+                loss_cls=CrossEntropyLoss(),
+                bbox_coder=self.bbox_coder,
+                **self.train_cfg.isr,
+                num_class=self.num_classes)
+            (new_labels, new_label_weights, new_bbox_targets,
+             new_bbox_weights) = all_targets
+            all_labels = new_labels.view(all_labels.shape)
+            all_label_weights = new_label_weights.view(all_label_weights.shape)
+            all_bbox_targets = new_bbox_targets.view(all_bbox_targets.shape)
+            all_bbox_weights = new_bbox_weights.view(all_bbox_weights.shape)
+
+        # add CARL loss
+        carl_loss_cfg = self.train_cfg.get('carl', None)
+        if carl_loss_cfg is not None:
+            loss_carl = carl_loss(
+                all_cls_scores.view(-1, all_cls_scores.size(-1)),
+                all_targets[0],
+                all_bbox_preds.view(-1, 4),
+                all_targets[2],
+                SmoothL1Loss(beta=1.),
+                **self.train_cfg.carl,
+                avg_factor=num_total_pos,
+                num_class=self.num_classes)
+
+        # check NaN and Inf
+        assert torch.isfinite(all_cls_scores).all().item(), \
+            'classification scores become infinite or NaN!'
+        assert torch.isfinite(all_bbox_preds).all().item(), \
+            'bbox predications become infinite or NaN!'
+
+        losses_cls, losses_bbox = multi_apply(
+            self.loss_single,
+            all_cls_scores,
+            all_bbox_preds,
+            all_anchors,
+            all_labels,
+            all_label_weights,
+            all_bbox_targets,
+            all_bbox_weights,
+            num_total_samples=num_total_pos)
+        loss_dict = dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
+        if carl_loss_cfg is not None:
+            loss_dict.update(loss_carl)
+        return loss_dict
diff --git a/detection/scrfd/mmdet/models/dense_heads/reppoints_head.py b/detection/scrfd/mmdet/models/dense_heads/reppoints_head.py
new file mode 100755
index 000000000..03e3fa0f1
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/reppoints_head.py
@@ -0,0 +1,763 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.ops import DeformConv2d
+
+from mmdet.core import (PointGenerator, build_assigner, build_sampler,
+                        images_to_levels, multi_apply, multiclass_nms, unmap)
+from ..builder import HEADS, build_loss
+from .anchor_free_head import AnchorFreeHead
+
+
+@HEADS.register_module()
+class RepPointsHead(AnchorFreeHead):
+    """RepPoint head.
+
+    Args:
+        point_feat_channels (int): Number of channels of points features.
+        gradient_mul (float): The multiplier to gradients from
+            points refinement and recognition.
+        point_strides (Iterable): points strides.
+        point_base_scale (int): bbox scale for assigning labels.
+        loss_cls (dict): Config of classification loss.
+        loss_bbox_init (dict): Config of initial points loss.
+        loss_bbox_refine (dict): Config of points loss in refinement.
+        use_grid_points (bool): If we use bounding box representation, the
+        reppoints is represented as grid points on the bounding box.
+        center_init (bool): Whether to use center point assignment.
+        transform_method (str): The methods to transform RepPoints to bbox.
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 point_feat_channels=256,
+                 num_points=9,
+                 gradient_mul=0.1,
+                 point_strides=[8, 16, 32, 64, 128],
+                 point_base_scale=4,
+                 loss_cls=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 loss_bbox_init=dict(
+                     type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.5),
+                 loss_bbox_refine=dict(
+                     type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0),
+                 use_grid_points=False,
+                 center_init=True,
+                 transform_method='moment',
+                 moment_mul=0.01,
+                 **kwargs):
+        self.num_points = num_points
+        self.point_feat_channels = point_feat_channels
+        self.use_grid_points = use_grid_points
+        self.center_init = center_init
+
+        # we use deform conv to extract points features
+        self.dcn_kernel = int(np.sqrt(num_points))
+        self.dcn_pad = int((self.dcn_kernel - 1) / 2)
+        assert self.dcn_kernel * self.dcn_kernel == num_points, \
+            'The points number should be a square number.'
+        assert self.dcn_kernel % 2 == 1, \
+            'The points number should be an odd square number.'
+        dcn_base = np.arange(-self.dcn_pad,
+                             self.dcn_pad + 1).astype(np.float64)
+        dcn_base_y = np.repeat(dcn_base, self.dcn_kernel)
+        dcn_base_x = np.tile(dcn_base, self.dcn_kernel)
+        dcn_base_offset = np.stack([dcn_base_y, dcn_base_x], axis=1).reshape(
+            (-1))
+        self.dcn_base_offset = torch.tensor(dcn_base_offset).view(1, -1, 1, 1)
+
+        super().__init__(num_classes, in_channels, loss_cls=loss_cls, **kwargs)
+
+        self.gradient_mul = gradient_mul
+        self.point_base_scale = point_base_scale
+        self.point_strides = point_strides
+        self.point_generators = [PointGenerator() for _ in self.point_strides]
+
+        self.sampling = loss_cls['type'] not in ['FocalLoss']
+        if self.train_cfg:
+            self.init_assigner = build_assigner(self.train_cfg.init.assigner)
+            self.refine_assigner = build_assigner(
+                self.train_cfg.refine.assigner)
+            # use PseudoSampler when sampling is False
+            if self.sampling and hasattr(self.train_cfg, 'sampler'):
+                sampler_cfg = self.train_cfg.sampler
+            else:
+                sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+        self.transform_method = transform_method
+        if self.transform_method == 'moment':
+            self.moment_transfer = nn.Parameter(
+                data=torch.zeros(2), requires_grad=True)
+            self.moment_mul = moment_mul
+
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        if self.use_sigmoid_cls:
+            self.cls_out_channels = self.num_classes
+        else:
+            self.cls_out_channels = self.num_classes + 1
+        self.loss_bbox_init = build_loss(loss_bbox_init)
+        self.loss_bbox_refine = build_loss(loss_bbox_refine)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        pts_out_dim = 4 if self.use_grid_points else 2 * self.num_points
+        self.reppoints_cls_conv = DeformConv2d(self.feat_channels,
+                                               self.point_feat_channels,
+                                               self.dcn_kernel, 1,
+                                               self.dcn_pad)
+        self.reppoints_cls_out = nn.Conv2d(self.point_feat_channels,
+                                           self.cls_out_channels, 1, 1, 0)
+        self.reppoints_pts_init_conv = nn.Conv2d(self.feat_channels,
+                                                 self.point_feat_channels, 3,
+                                                 1, 1)
+        self.reppoints_pts_init_out = nn.Conv2d(self.point_feat_channels,
+                                                pts_out_dim, 1, 1, 0)
+        self.reppoints_pts_refine_conv = DeformConv2d(self.feat_channels,
+                                                      self.point_feat_channels,
+                                                      self.dcn_kernel, 1,
+                                                      self.dcn_pad)
+        self.reppoints_pts_refine_out = nn.Conv2d(self.point_feat_channels,
+                                                  pts_out_dim, 1, 1, 0)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.reppoints_cls_conv, std=0.01)
+        normal_init(self.reppoints_cls_out, std=0.01, bias=bias_cls)
+        normal_init(self.reppoints_pts_init_conv, std=0.01)
+        normal_init(self.reppoints_pts_init_out, std=0.01)
+        normal_init(self.reppoints_pts_refine_conv, std=0.01)
+        normal_init(self.reppoints_pts_refine_out, std=0.01)
+
+    def points2bbox(self, pts, y_first=True):
+        """Converting the points set into bounding box.
+
+        :param pts: the input points sets (fields), each points
+            set (fields) is represented as 2n scalar.
+        :param y_first: if y_fisrt=True, the point set is represented as
+            [y1, x1, y2, x2 ... yn, xn], otherwise the point set is
+            represented as [x1, y1, x2, y2 ... xn, yn].
+        :return: each points set is converting to a bbox [x1, y1, x2, y2].
+        """
+        pts_reshape = pts.view(pts.shape[0], -1, 2, *pts.shape[2:])
+        pts_y = pts_reshape[:, :, 0, ...] if y_first else pts_reshape[:, :, 1,
+                                                                      ...]
+        pts_x = pts_reshape[:, :, 1, ...] if y_first else pts_reshape[:, :, 0,
+                                                                      ...]
+        if self.transform_method == 'minmax':
+            bbox_left = pts_x.min(dim=1, keepdim=True)[0]
+            bbox_right = pts_x.max(dim=1, keepdim=True)[0]
+            bbox_up = pts_y.min(dim=1, keepdim=True)[0]
+            bbox_bottom = pts_y.max(dim=1, keepdim=True)[0]
+            bbox = torch.cat([bbox_left, bbox_up, bbox_right, bbox_bottom],
+                             dim=1)
+        elif self.transform_method == 'partial_minmax':
+            pts_y = pts_y[:, :4, ...]
+            pts_x = pts_x[:, :4, ...]
+            bbox_left = pts_x.min(dim=1, keepdim=True)[0]
+            bbox_right = pts_x.max(dim=1, keepdim=True)[0]
+            bbox_up = pts_y.min(dim=1, keepdim=True)[0]
+            bbox_bottom = pts_y.max(dim=1, keepdim=True)[0]
+            bbox = torch.cat([bbox_left, bbox_up, bbox_right, bbox_bottom],
+                             dim=1)
+        elif self.transform_method == 'moment':
+            pts_y_mean = pts_y.mean(dim=1, keepdim=True)
+            pts_x_mean = pts_x.mean(dim=1, keepdim=True)
+            pts_y_std = torch.std(pts_y - pts_y_mean, dim=1, keepdim=True)
+            pts_x_std = torch.std(pts_x - pts_x_mean, dim=1, keepdim=True)
+            moment_transfer = (self.moment_transfer * self.moment_mul) + (
+                self.moment_transfer.detach() * (1 - self.moment_mul))
+            moment_width_transfer = moment_transfer[0]
+            moment_height_transfer = moment_transfer[1]
+            half_width = pts_x_std * torch.exp(moment_width_transfer)
+            half_height = pts_y_std * torch.exp(moment_height_transfer)
+            bbox = torch.cat([
+                pts_x_mean - half_width, pts_y_mean - half_height,
+                pts_x_mean + half_width, pts_y_mean + half_height
+            ],
+                             dim=1)
+        else:
+            raise NotImplementedError
+        return bbox
+
+    def gen_grid_from_reg(self, reg, previous_boxes):
+        """Base on the previous bboxes and regression values, we compute the
+        regressed bboxes and generate the grids on the bboxes.
+
+        :param reg: the regression value to previous bboxes.
+        :param previous_boxes: previous bboxes.
+        :return: generate grids on the regressed bboxes.
+        """
+        b, _, h, w = reg.shape
+        bxy = (previous_boxes[:, :2, ...] + previous_boxes[:, 2:, ...]) / 2.
+        bwh = (previous_boxes[:, 2:, ...] -
+               previous_boxes[:, :2, ...]).clamp(min=1e-6)
+        grid_topleft = bxy + bwh * reg[:, :2, ...] - 0.5 * bwh * torch.exp(
+            reg[:, 2:, ...])
+        grid_wh = bwh * torch.exp(reg[:, 2:, ...])
+        grid_left = grid_topleft[:, [0], ...]
+        grid_top = grid_topleft[:, [1], ...]
+        grid_width = grid_wh[:, [0], ...]
+        grid_height = grid_wh[:, [1], ...]
+        intervel = torch.linspace(0., 1., self.dcn_kernel).view(
+            1, self.dcn_kernel, 1, 1).type_as(reg)
+        grid_x = grid_left + grid_width * intervel
+        grid_x = grid_x.unsqueeze(1).repeat(1, self.dcn_kernel, 1, 1, 1)
+        grid_x = grid_x.view(b, -1, h, w)
+        grid_y = grid_top + grid_height * intervel
+        grid_y = grid_y.unsqueeze(2).repeat(1, 1, self.dcn_kernel, 1, 1)
+        grid_y = grid_y.view(b, -1, h, w)
+        grid_yx = torch.stack([grid_y, grid_x], dim=2)
+        grid_yx = grid_yx.view(b, -1, h, w)
+        regressed_bbox = torch.cat([
+            grid_left, grid_top, grid_left + grid_width, grid_top + grid_height
+        ], 1)
+        return grid_yx, regressed_bbox
+
+    def forward(self, feats):
+        return multi_apply(self.forward_single, feats)
+
+    def forward_single(self, x):
+        """Forward feature map of a single FPN level."""
+        dcn_base_offset = self.dcn_base_offset.type_as(x)
+        # If we use center_init, the initial reppoints is from center points.
+        # If we use bounding bbox representation, the initial reppoints is
+        #   from regular grid placed on a pre-defined bbox.
+        if self.use_grid_points or not self.center_init:
+            scale = self.point_base_scale / 2
+            points_init = dcn_base_offset / dcn_base_offset.max() * scale
+            bbox_init = x.new_tensor([-scale, -scale, scale,
+                                      scale]).view(1, 4, 1, 1)
+        else:
+            points_init = 0
+        cls_feat = x
+        pts_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            pts_feat = reg_conv(pts_feat)
+        # initialize reppoints
+        pts_out_init = self.reppoints_pts_init_out(
+            self.relu(self.reppoints_pts_init_conv(pts_feat)))
+        if self.use_grid_points:
+            pts_out_init, bbox_out_init = self.gen_grid_from_reg(
+                pts_out_init, bbox_init.detach())
+        else:
+            pts_out_init = pts_out_init + points_init
+        # refine and classify reppoints
+        pts_out_init_grad_mul = (1 - self.gradient_mul) * pts_out_init.detach(
+        ) + self.gradient_mul * pts_out_init
+        dcn_offset = pts_out_init_grad_mul - dcn_base_offset
+        cls_out = self.reppoints_cls_out(
+            self.relu(self.reppoints_cls_conv(cls_feat, dcn_offset)))
+        pts_out_refine = self.reppoints_pts_refine_out(
+            self.relu(self.reppoints_pts_refine_conv(pts_feat, dcn_offset)))
+        if self.use_grid_points:
+            pts_out_refine, bbox_out_refine = self.gen_grid_from_reg(
+                pts_out_refine, bbox_out_init.detach())
+        else:
+            pts_out_refine = pts_out_refine + pts_out_init.detach()
+        return cls_out, pts_out_init, pts_out_refine
+
+    def get_points(self, featmap_sizes, img_metas, device):
+        """Get points according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+
+        Returns:
+            tuple: points of each image, valid flags of each image
+        """
+        num_imgs = len(img_metas)
+        num_levels = len(featmap_sizes)
+
+        # since feature map sizes of all images are the same, we only compute
+        # points center for one time
+        multi_level_points = []
+        for i in range(num_levels):
+            points = self.point_generators[i].grid_points(
+                featmap_sizes[i], self.point_strides[i], device)
+            multi_level_points.append(points)
+        points_list = [[point.clone() for point in multi_level_points]
+                       for _ in range(num_imgs)]
+
+        # for each image, we compute valid flags of multi level grids
+        valid_flag_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_flags = []
+            for i in range(num_levels):
+                point_stride = self.point_strides[i]
+                feat_h, feat_w = featmap_sizes[i]
+                h, w = img_meta['pad_shape'][:2]
+                valid_feat_h = min(int(np.ceil(h / point_stride)), feat_h)
+                valid_feat_w = min(int(np.ceil(w / point_stride)), feat_w)
+                flags = self.point_generators[i].valid_flags(
+                    (feat_h, feat_w), (valid_feat_h, valid_feat_w), device)
+                multi_level_flags.append(flags)
+            valid_flag_list.append(multi_level_flags)
+
+        return points_list, valid_flag_list
+
+    def centers_to_bboxes(self, point_list):
+        """Get bboxes according to center points.
+
+        Only used in :class:`MaxIoUAssigner`.
+        """
+        bbox_list = []
+        for i_img, point in enumerate(point_list):
+            bbox = []
+            for i_lvl in range(len(self.point_strides)):
+                scale = self.point_base_scale * self.point_strides[i_lvl] * 0.5
+                bbox_shift = torch.Tensor([-scale, -scale, scale,
+                                           scale]).view(1, 4).type_as(point[0])
+                bbox_center = torch.cat(
+                    [point[i_lvl][:, :2], point[i_lvl][:, :2]], dim=1)
+                bbox.append(bbox_center + bbox_shift)
+            bbox_list.append(bbox)
+        return bbox_list
+
+    def offset_to_pts(self, center_list, pred_list):
+        """Change from point offset to point coordinate."""
+        pts_list = []
+        for i_lvl in range(len(self.point_strides)):
+            pts_lvl = []
+            for i_img in range(len(center_list)):
+                pts_center = center_list[i_img][i_lvl][:, :2].repeat(
+                    1, self.num_points)
+                pts_shift = pred_list[i_lvl][i_img]
+                yx_pts_shift = pts_shift.permute(1, 2, 0).view(
+                    -1, 2 * self.num_points)
+                y_pts_shift = yx_pts_shift[..., 0::2]
+                x_pts_shift = yx_pts_shift[..., 1::2]
+                xy_pts_shift = torch.stack([x_pts_shift, y_pts_shift], -1)
+                xy_pts_shift = xy_pts_shift.view(*yx_pts_shift.shape[:-1], -1)
+                pts = xy_pts_shift * self.point_strides[i_lvl] + pts_center
+                pts_lvl.append(pts)
+            pts_lvl = torch.stack(pts_lvl, 0)
+            pts_list.append(pts_lvl)
+        return pts_list
+
+    def _point_target_single(self,
+                             flat_proposals,
+                             valid_flags,
+                             gt_bboxes,
+                             gt_bboxes_ignore,
+                             gt_labels,
+                             label_channels=1,
+                             stage='init',
+                             unmap_outputs=True):
+        inside_flags = valid_flags
+        if not inside_flags.any():
+            return (None, ) * 7
+        # assign gt and sample proposals
+        proposals = flat_proposals[inside_flags, :]
+
+        if stage == 'init':
+            assigner = self.init_assigner
+            pos_weight = self.train_cfg.init.pos_weight
+        else:
+            assigner = self.refine_assigner
+            pos_weight = self.train_cfg.refine.pos_weight
+        assign_result = assigner.assign(proposals, gt_bboxes, gt_bboxes_ignore,
+                                        None if self.sampling else gt_labels)
+        sampling_result = self.sampler.sample(assign_result, proposals,
+                                              gt_bboxes)
+
+        num_valid_proposals = proposals.shape[0]
+        bbox_gt = proposals.new_zeros([num_valid_proposals, 4])
+        pos_proposals = torch.zeros_like(proposals)
+        proposals_weights = proposals.new_zeros([num_valid_proposals, 4])
+        labels = proposals.new_full((num_valid_proposals, ),
+                                    self.num_classes,
+                                    dtype=torch.long)
+        label_weights = proposals.new_zeros(
+            num_valid_proposals, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            pos_gt_bboxes = sampling_result.pos_gt_bboxes
+            bbox_gt[pos_inds, :] = pos_gt_bboxes
+            pos_proposals[pos_inds, :] = proposals[pos_inds, :]
+            proposals_weights[pos_inds, :] = 1.0
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of proposals
+        if unmap_outputs:
+            num_total_proposals = flat_proposals.size(0)
+            labels = unmap(labels, num_total_proposals, inside_flags)
+            label_weights = unmap(label_weights, num_total_proposals,
+                                  inside_flags)
+            bbox_gt = unmap(bbox_gt, num_total_proposals, inside_flags)
+            pos_proposals = unmap(pos_proposals, num_total_proposals,
+                                  inside_flags)
+            proposals_weights = unmap(proposals_weights, num_total_proposals,
+                                      inside_flags)
+
+        return (labels, label_weights, bbox_gt, pos_proposals,
+                proposals_weights, pos_inds, neg_inds)
+
+    def get_targets(self,
+                    proposals_list,
+                    valid_flag_list,
+                    gt_bboxes_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None,
+                    gt_labels_list=None,
+                    stage='init',
+                    label_channels=1,
+                    unmap_outputs=True):
+        """Compute corresponding GT box and classification targets for
+        proposals.
+
+        Args:
+            proposals_list (list[list]): Multi level points/bboxes of each
+                image.
+            valid_flag_list (list[list]): Multi level valid flags of each
+                image.
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            img_metas (list[dict]): Meta info of each image.
+            gt_bboxes_ignore_list (list[Tensor]): Ground truth bboxes to be
+                ignored.
+            gt_bboxes_list (list[Tensor]): Ground truth labels of each box.
+            stage (str): `init` or `refine`. Generate target for init stage or
+                refine stage
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple:
+                - labels_list (list[Tensor]): Labels of each level.
+                - label_weights_list (list[Tensor]): Label weights of each level.  # noqa: E501
+                - bbox_gt_list (list[Tensor]): Ground truth bbox of each level.
+                - proposal_list (list[Tensor]): Proposals(points/bboxes) of each level.  # noqa: E501
+                - proposal_weights_list (list[Tensor]): Proposal weights of each level.  # noqa: E501
+                - num_total_pos (int): Number of positive samples in all images.  # noqa: E501
+                - num_total_neg (int): Number of negative samples in all images.  # noqa: E501
+        """
+        assert stage in ['init', 'refine']
+        num_imgs = len(img_metas)
+        assert len(proposals_list) == len(valid_flag_list) == num_imgs
+
+        # points number of multi levels
+        num_level_proposals = [points.size(0) for points in proposals_list[0]]
+
+        # concat all level points and flags to a single tensor
+        for i in range(num_imgs):
+            assert len(proposals_list[i]) == len(valid_flag_list[i])
+            proposals_list[i] = torch.cat(proposals_list[i])
+            valid_flag_list[i] = torch.cat(valid_flag_list[i])
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        (all_labels, all_label_weights, all_bbox_gt, all_proposals,
+         all_proposal_weights, pos_inds_list, neg_inds_list) = multi_apply(
+             self._point_target_single,
+             proposals_list,
+             valid_flag_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             gt_labels_list,
+             stage=stage,
+             label_channels=label_channels,
+             unmap_outputs=unmap_outputs)
+        # no valid points
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled points of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        labels_list = images_to_levels(all_labels, num_level_proposals)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_proposals)
+        bbox_gt_list = images_to_levels(all_bbox_gt, num_level_proposals)
+        proposals_list = images_to_levels(all_proposals, num_level_proposals)
+        proposal_weights_list = images_to_levels(all_proposal_weights,
+                                                 num_level_proposals)
+        return (labels_list, label_weights_list, bbox_gt_list, proposals_list,
+                proposal_weights_list, num_total_pos, num_total_neg)
+
+    def loss_single(self, cls_score, pts_pred_init, pts_pred_refine, labels,
+                    label_weights, bbox_gt_init, bbox_weights_init,
+                    bbox_gt_refine, bbox_weights_refine, stride,
+                    num_total_samples_init, num_total_samples_refine):
+        # classification loss
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+        cls_score = cls_score.permute(0, 2, 3,
+                                      1).reshape(-1, self.cls_out_channels)
+        cls_score = cls_score.contiguous()
+        loss_cls = self.loss_cls(
+            cls_score,
+            labels,
+            label_weights,
+            avg_factor=num_total_samples_refine)
+
+        # points loss
+        bbox_gt_init = bbox_gt_init.reshape(-1, 4)
+        bbox_weights_init = bbox_weights_init.reshape(-1, 4)
+        bbox_pred_init = self.points2bbox(
+            pts_pred_init.reshape(-1, 2 * self.num_points), y_first=False)
+        bbox_gt_refine = bbox_gt_refine.reshape(-1, 4)
+        bbox_weights_refine = bbox_weights_refine.reshape(-1, 4)
+        bbox_pred_refine = self.points2bbox(
+            pts_pred_refine.reshape(-1, 2 * self.num_points), y_first=False)
+        normalize_term = self.point_base_scale * stride
+        loss_pts_init = self.loss_bbox_init(
+            bbox_pred_init / normalize_term,
+            bbox_gt_init / normalize_term,
+            bbox_weights_init,
+            avg_factor=num_total_samples_init)
+        loss_pts_refine = self.loss_bbox_refine(
+            bbox_pred_refine / normalize_term,
+            bbox_gt_refine / normalize_term,
+            bbox_weights_refine,
+            avg_factor=num_total_samples_refine)
+        return loss_cls, loss_pts_init, loss_pts_refine
+
+    def loss(self,
+             cls_scores,
+             pts_preds_init,
+             pts_preds_refine,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == len(self.point_generators)
+        device = cls_scores[0].device
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        # target for initial stage
+        center_list, valid_flag_list = self.get_points(featmap_sizes,
+                                                       img_metas, device)
+        pts_coordinate_preds_init = self.offset_to_pts(center_list,
+                                                       pts_preds_init)
+        if self.train_cfg.init.assigner['type'] == 'PointAssigner':
+            # Assign target for center list
+            candidate_list = center_list
+        else:
+            # transform center list to bbox list and
+            #   assign target for bbox list
+            bbox_list = self.centers_to_bboxes(center_list)
+            candidate_list = bbox_list
+        cls_reg_targets_init = self.get_targets(
+            candidate_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            stage='init',
+            label_channels=label_channels)
+        (*_, bbox_gt_list_init, candidate_list_init, bbox_weights_list_init,
+         num_total_pos_init, num_total_neg_init) = cls_reg_targets_init
+        num_total_samples_init = (
+            num_total_pos_init +
+            num_total_neg_init if self.sampling else num_total_pos_init)
+
+        # target for refinement stage
+        center_list, valid_flag_list = self.get_points(featmap_sizes,
+                                                       img_metas, device)
+        pts_coordinate_preds_refine = self.offset_to_pts(
+            center_list, pts_preds_refine)
+        bbox_list = []
+        for i_img, center in enumerate(center_list):
+            bbox = []
+            for i_lvl in range(len(pts_preds_refine)):
+                bbox_preds_init = self.points2bbox(
+                    pts_preds_init[i_lvl].detach())
+                bbox_shift = bbox_preds_init * self.point_strides[i_lvl]
+                bbox_center = torch.cat(
+                    [center[i_lvl][:, :2], center[i_lvl][:, :2]], dim=1)
+                bbox.append(bbox_center +
+                            bbox_shift[i_img].permute(1, 2, 0).reshape(-1, 4))
+            bbox_list.append(bbox)
+        cls_reg_targets_refine = self.get_targets(
+            bbox_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            stage='refine',
+            label_channels=label_channels)
+        (labels_list, label_weights_list, bbox_gt_list_refine,
+         candidate_list_refine, bbox_weights_list_refine, num_total_pos_refine,
+         num_total_neg_refine) = cls_reg_targets_refine
+        num_total_samples_refine = (
+            num_total_pos_refine +
+            num_total_neg_refine if self.sampling else num_total_pos_refine)
+
+        # compute loss
+        losses_cls, losses_pts_init, losses_pts_refine = multi_apply(
+            self.loss_single,
+            cls_scores,
+            pts_coordinate_preds_init,
+            pts_coordinate_preds_refine,
+            labels_list,
+            label_weights_list,
+            bbox_gt_list_init,
+            bbox_weights_list_init,
+            bbox_gt_list_refine,
+            bbox_weights_list_refine,
+            self.point_strides,
+            num_total_samples_init=num_total_samples_init,
+            num_total_samples_refine=num_total_samples_refine)
+        loss_dict_all = {
+            'loss_cls': losses_cls,
+            'loss_pts_init': losses_pts_init,
+            'loss_pts_refine': losses_pts_refine
+        }
+        return loss_dict_all
+
+    def get_bboxes(self,
+                   cls_scores,
+                   pts_preds_init,
+                   pts_preds_refine,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        assert len(cls_scores) == len(pts_preds_refine)
+        device = cls_scores[0].device
+        bbox_preds_refine = [
+            self.points2bbox(pts_pred_refine)
+            for pts_pred_refine in pts_preds_refine
+        ]
+        num_levels = len(cls_scores)
+        mlvl_points = [
+            self.point_generators[i].grid_points(cls_scores[i].size()[-2:],
+                                                 self.point_strides[i], device)
+            for i in range(num_levels)
+        ]
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds_refine[i][img_id].detach()
+                for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self._get_bboxes_single(cls_score_list, bbox_pred_list,
+                                                mlvl_points, img_shape,
+                                                scale_factor, cfg, rescale,
+                                                with_nms)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_points,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_points)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for i_lvl, (cls_score, bbox_pred, points) in enumerate(
+                zip(cls_scores, bbox_preds, mlvl_points)):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            cls_score = cls_score.permute(1, 2,
+                                          0).reshape(-1, self.cls_out_channels)
+            if self.use_sigmoid_cls:
+                scores = cls_score.sigmoid()
+            else:
+                scores = cls_score.softmax(-1)
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                if self.use_sigmoid_cls:
+                    max_scores, _ = scores.max(dim=1)
+                else:
+                    # remind that we set FG labels to [0, num_class-1]
+                    # since mmdet v2.0
+                    # BG cat_id: num_class
+                    max_scores, _ = scores[:, :-1].max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                points = points[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+            bbox_pos_center = torch.cat([points[:, :2], points[:, :2]], dim=1)
+            bboxes = bbox_pred * self.point_strides[i_lvl] + bbox_pos_center
+            x1 = bboxes[:, 0].clamp(min=0, max=img_shape[1])
+            y1 = bboxes[:, 1].clamp(min=0, max=img_shape[0])
+            x2 = bboxes[:, 2].clamp(min=0, max=img_shape[1])
+            y2 = bboxes[:, 3].clamp(min=0, max=img_shape[0])
+            bboxes = torch.stack([x1, y1, x2, y2], dim=-1)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        if self.use_sigmoid_cls:
+            # Add a dummy background class to the backend when using sigmoid
+            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+            # BG cat_id: num_class
+            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores
diff --git a/detection/scrfd/mmdet/models/dense_heads/retina_head.py b/detection/scrfd/mmdet/models/dense_heads/retina_head.py
new file mode 100755
index 000000000..b12416fa8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/retina_head.py
@@ -0,0 +1,114 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+
+from ..builder import HEADS
+from .anchor_head import AnchorHead
+
+
+@HEADS.register_module()
+class RetinaHead(AnchorHead):
+    r"""An anchor-based head used in `RetinaNet
+    <https://arxiv.org/pdf/1708.02002.pdf>`_.
+
+    The head contains two subnetworks. The first classifies anchor boxes and
+    the second regresses deltas for the anchors.
+
+    Example:
+        >>> import torch
+        >>> self = RetinaHead(11, 7)
+        >>> x = torch.rand(1, 7, 32, 32)
+        >>> cls_score, bbox_pred = self.forward_single(x)
+        >>> # Each anchor predicts a score for each class except background
+        >>> cls_per_anchor = cls_score.shape[1] / self.num_anchors
+        >>> box_per_anchor = bbox_pred.shape[1] / self.num_anchors
+        >>> assert cls_per_anchor == (self.num_classes)
+        >>> assert box_per_anchor == 4
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 anchor_generator=dict(
+                     type='AnchorGenerator',
+                     octave_base_scale=4,
+                     scales_per_octave=3,
+                     ratios=[0.5, 1.0, 2.0],
+                     strides=[8, 16, 32, 64, 128]),
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        super(RetinaHead, self).__init__(
+            num_classes,
+            in_channels,
+            anchor_generator=anchor_generator,
+            **kwargs)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        self.retina_cls = nn.Conv2d(
+            self.feat_channels,
+            self.num_anchors * self.cls_out_channels,
+            3,
+            padding=1)
+        self.retina_reg = nn.Conv2d(
+            self.feat_channels, self.num_anchors * 4, 3, padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.retina_cls, std=0.01, bias=bias_cls)
+        normal_init(self.retina_reg, std=0.01)
+
+    def forward_single(self, x):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls scores for a single scale level
+                    the channels number is num_anchors * num_classes.
+                bbox_pred (Tensor): Box energies / deltas for a single scale
+                    level, the channels number is num_anchors * 4.
+        """
+        cls_feat = x
+        reg_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            reg_feat = reg_conv(reg_feat)
+        cls_score = self.retina_cls(cls_feat)
+        bbox_pred = self.retina_reg(reg_feat)
+        return cls_score, bbox_pred
diff --git a/detection/scrfd/mmdet/models/dense_heads/retina_sepbn_head.py b/detection/scrfd/mmdet/models/dense_heads/retina_sepbn_head.py
new file mode 100755
index 000000000..6b8ce7f01
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/retina_sepbn_head.py
@@ -0,0 +1,113 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+
+from ..builder import HEADS
+from .anchor_head import AnchorHead
+
+
+@HEADS.register_module()
+class RetinaSepBNHead(AnchorHead):
+    """"RetinaHead with separate BN.
+
+    In RetinaHead, conv/norm layers are shared across different FPN levels,
+    while in RetinaSepBNHead, conv layers are shared across different FPN
+    levels, but BN layers are separated.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 num_ins,
+                 in_channels,
+                 stacked_convs=4,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.num_ins = num_ins
+        super(RetinaSepBNHead, self).__init__(num_classes, in_channels,
+                                              **kwargs)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.num_ins):
+            cls_convs = nn.ModuleList()
+            reg_convs = nn.ModuleList()
+            for i in range(self.stacked_convs):
+                chn = self.in_channels if i == 0 else self.feat_channels
+                cls_convs.append(
+                    ConvModule(
+                        chn,
+                        self.feat_channels,
+                        3,
+                        stride=1,
+                        padding=1,
+                        conv_cfg=self.conv_cfg,
+                        norm_cfg=self.norm_cfg))
+                reg_convs.append(
+                    ConvModule(
+                        chn,
+                        self.feat_channels,
+                        3,
+                        stride=1,
+                        padding=1,
+                        conv_cfg=self.conv_cfg,
+                        norm_cfg=self.norm_cfg))
+            self.cls_convs.append(cls_convs)
+            self.reg_convs.append(reg_convs)
+        for i in range(self.stacked_convs):
+            for j in range(1, self.num_ins):
+                self.cls_convs[j][i].conv = self.cls_convs[0][i].conv
+                self.reg_convs[j][i].conv = self.reg_convs[0][i].conv
+        self.retina_cls = nn.Conv2d(
+            self.feat_channels,
+            self.num_anchors * self.cls_out_channels,
+            3,
+            padding=1)
+        self.retina_reg = nn.Conv2d(
+            self.feat_channels, self.num_anchors * 4, 3, padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs[0]:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs[0]:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.retina_cls, std=0.01, bias=bias_cls)
+        normal_init(self.retina_reg, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually a tuple of classification scores and bbox prediction
+                cls_scores (list[Tensor]): Classification scores for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * num_classes.
+                bbox_preds (list[Tensor]): Box energies / deltas for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * 4.
+        """
+        cls_scores = []
+        bbox_preds = []
+        for i, x in enumerate(feats):
+            cls_feat = feats[i]
+            reg_feat = feats[i]
+            for cls_conv in self.cls_convs[i]:
+                cls_feat = cls_conv(cls_feat)
+            for reg_conv in self.reg_convs[i]:
+                reg_feat = reg_conv(reg_feat)
+            cls_score = self.retina_cls(cls_feat)
+            bbox_pred = self.retina_reg(reg_feat)
+            cls_scores.append(cls_score)
+            bbox_preds.append(bbox_pred)
+        return cls_scores, bbox_preds
diff --git a/detection/scrfd/mmdet/models/dense_heads/rpn_head.py b/detection/scrfd/mmdet/models/dense_heads/rpn_head.py
new file mode 100755
index 000000000..f565d1a41
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/rpn_head.py
@@ -0,0 +1,168 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import normal_init
+from mmcv.ops import batched_nms
+
+from ..builder import HEADS
+from .anchor_head import AnchorHead
+from .rpn_test_mixin import RPNTestMixin
+
+
+@HEADS.register_module()
+class RPNHead(RPNTestMixin, AnchorHead):
+    """RPN head.
+
+    Args:
+        in_channels (int): Number of channels in the input feature map.
+    """  # noqa: W605
+
+    def __init__(self, in_channels, **kwargs):
+        super(RPNHead, self).__init__(1, in_channels, **kwargs)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.rpn_conv = nn.Conv2d(
+            self.in_channels, self.feat_channels, 3, padding=1)
+        self.rpn_cls = nn.Conv2d(self.feat_channels,
+                                 self.num_anchors * self.cls_out_channels, 1)
+        self.rpn_reg = nn.Conv2d(self.feat_channels, self.num_anchors * 4, 1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        normal_init(self.rpn_conv, std=0.01)
+        normal_init(self.rpn_cls, std=0.01)
+        normal_init(self.rpn_reg, std=0.01)
+
+    def forward_single(self, x):
+        """Forward feature map of a single scale level."""
+        x = self.rpn_conv(x)
+        x = F.relu(x, inplace=True)
+        rpn_cls_score = self.rpn_cls(x)
+        rpn_bbox_pred = self.rpn_reg(x)
+        return rpn_cls_score, rpn_bbox_pred
+
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        losses = super(RPNHead, self).loss(
+            cls_scores,
+            bbox_preds,
+            gt_bboxes,
+            None,
+            img_metas,
+            gt_bboxes_ignore=gt_bboxes_ignore)
+        return dict(
+            loss_rpn_cls=losses['loss_cls'], loss_rpn_bbox=losses['loss_bbox'])
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (num_anchors * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (num_anchors * 4, H, W).
+            mlvl_anchors (list[Tensor]): Box reference for each scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+
+        Returns:
+            Tensor: Labeled boxes in shape (n, 5), where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        # bboxes from different level should be independent during NMS,
+        # level_ids are used as labels for batched NMS to separate them
+        level_ids = []
+        mlvl_scores = []
+        mlvl_bbox_preds = []
+        mlvl_valid_anchors = []
+        for idx in range(len(cls_scores)):
+            rpn_cls_score = cls_scores[idx]
+            rpn_bbox_pred = bbox_preds[idx]
+            assert rpn_cls_score.size()[-2:] == rpn_bbox_pred.size()[-2:]
+            rpn_cls_score = rpn_cls_score.permute(1, 2, 0)
+            if self.use_sigmoid_cls:
+                rpn_cls_score = rpn_cls_score.reshape(-1)
+                scores = rpn_cls_score.sigmoid()
+            else:
+                rpn_cls_score = rpn_cls_score.reshape(-1, 2)
+                # We set FG labels to [0, num_class-1] and BG label to
+                # num_class in RPN head since mmdet v2.5, which is unified to
+                # be consistent with other head since mmdet v2.0. In mmdet v2.0
+                # to v2.4 we keep BG label as 0 and FG label as 1 in rpn head.
+                scores = rpn_cls_score.softmax(dim=1)[:, 0]
+            rpn_bbox_pred = rpn_bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            anchors = mlvl_anchors[idx]
+            if cfg.nms_pre > 0 and scores.shape[0] > cfg.nms_pre:
+                # sort is faster than topk
+                # _, topk_inds = scores.topk(cfg.nms_pre)
+                ranked_scores, rank_inds = scores.sort(descending=True)
+                topk_inds = rank_inds[:cfg.nms_pre]
+                scores = ranked_scores[:cfg.nms_pre]
+                rpn_bbox_pred = rpn_bbox_pred[topk_inds, :]
+                anchors = anchors[topk_inds, :]
+            mlvl_scores.append(scores)
+            mlvl_bbox_preds.append(rpn_bbox_pred)
+            mlvl_valid_anchors.append(anchors)
+            level_ids.append(
+                scores.new_full((scores.size(0), ), idx, dtype=torch.long))
+
+        scores = torch.cat(mlvl_scores)
+        anchors = torch.cat(mlvl_valid_anchors)
+        rpn_bbox_pred = torch.cat(mlvl_bbox_preds)
+        proposals = self.bbox_coder.decode(
+            anchors, rpn_bbox_pred, max_shape=img_shape)
+        ids = torch.cat(level_ids)
+
+        if cfg.min_bbox_size > 0:
+            w = proposals[:, 2] - proposals[:, 0]
+            h = proposals[:, 3] - proposals[:, 1]
+            valid_inds = torch.nonzero(
+                (w >= cfg.min_bbox_size)
+                & (h >= cfg.min_bbox_size),
+                as_tuple=False).squeeze()
+            if valid_inds.sum().item() != len(proposals):
+                proposals = proposals[valid_inds, :]
+                scores = scores[valid_inds]
+                ids = ids[valid_inds]
+
+        # TODO: remove the hard coded nms type
+        nms_cfg = dict(type='nms', iou_threshold=cfg.nms_thr)
+        dets, keep = batched_nms(proposals, scores, ids, nms_cfg)
+        return dets[:cfg.nms_post]
diff --git a/detection/scrfd/mmdet/models/dense_heads/rpn_test_mixin.py b/detection/scrfd/mmdet/models/dense_heads/rpn_test_mixin.py
new file mode 100755
index 000000000..4ce5c66f8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/rpn_test_mixin.py
@@ -0,0 +1,59 @@
+import sys
+
+from mmdet.core import merge_aug_proposals
+
+if sys.version_info >= (3, 7):
+    from mmdet.utils.contextmanagers import completed
+
+
+class RPNTestMixin(object):
+    """Test methods of RPN."""
+
+    if sys.version_info >= (3, 7):
+
+        async def async_simple_test_rpn(self, x, img_metas):
+            sleep_interval = self.test_cfg.pop('async_sleep_interval', 0.025)
+            async with completed(
+                    __name__, 'rpn_head_forward',
+                    sleep_interval=sleep_interval):
+                rpn_outs = self(x)
+
+            proposal_list = self.get_bboxes(*rpn_outs, img_metas)
+            return proposal_list
+
+    def simple_test_rpn(self, x, img_metas):
+        """Test without augmentation.
+
+        Args:
+            x (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+            img_metas (list[dict]): Meta info of each image.
+
+        Returns:
+            list[Tensor]: Proposals of each image.
+        """
+        rpn_outs = self(x)
+        proposal_list = self.get_bboxes(*rpn_outs, img_metas)
+        return proposal_list
+
+    def aug_test_rpn(self, feats, img_metas):
+        samples_per_gpu = len(img_metas[0])
+        aug_proposals = [[] for _ in range(samples_per_gpu)]
+        for x, img_meta in zip(feats, img_metas):
+            proposal_list = self.simple_test_rpn(x, img_meta)
+            for i, proposals in enumerate(proposal_list):
+                aug_proposals[i].append(proposals)
+        # reorganize the order of 'img_metas' to match the dimensions
+        # of 'aug_proposals'
+        aug_img_metas = []
+        for i in range(samples_per_gpu):
+            aug_img_meta = []
+            for j in range(len(img_metas)):
+                aug_img_meta.append(img_metas[j][i])
+            aug_img_metas.append(aug_img_meta)
+        # after merging, proposals will be rescaled to the original image size
+        merged_proposals = [
+            merge_aug_proposals(proposals, aug_img_meta, self.test_cfg)
+            for proposals, aug_img_meta in zip(aug_proposals, aug_img_metas)
+        ]
+        return merged_proposals
diff --git a/detection/scrfd/mmdet/models/dense_heads/sabl_retina_head.py b/detection/scrfd/mmdet/models/dense_heads/sabl_retina_head.py
new file mode 100755
index 000000000..73143da4a
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/sabl_retina_head.py
@@ -0,0 +1,618 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (build_anchor_generator, build_assigner,
+                        build_bbox_coder, build_sampler, images_to_levels,
+                        multi_apply, multiclass_nms, unmap)
+from ..builder import HEADS, build_loss
+from .base_dense_head import BaseDenseHead
+from .guided_anchor_head import GuidedAnchorHead
+
+
+@HEADS.register_module()
+class SABLRetinaHead(BaseDenseHead):
+    """Side-Aware Boundary Localization (SABL) for RetinaNet.
+
+    The anchor generation, assigning and sampling in SABLRetinaHead
+    are the same as GuidedAnchorHead for guided anchoring.
+
+    Please refer to https://arxiv.org/abs/1912.04260 for more details.
+
+    Args:
+        num_classes (int): Number of classes.
+        in_channels (int): Number of channels in the input feature map.
+        stacked_convs (int): Number of Convs for classification \
+            and regression branches. Defaults to 4.
+        feat_channels (int): Number of hidden channels. \
+            Defaults to 256.
+        approx_anchor_generator (dict): Config dict for approx generator.
+        square_anchor_generator (dict): Config dict for square generator.
+        conv_cfg (dict): Config dict for ConvModule. Defaults to None.
+        norm_cfg (dict): Config dict for Norm Layer. Defaults to None.
+        bbox_coder (dict): Config dict for bbox coder.
+        reg_decoded_bbox (bool): Whether to regress decoded bbox. \
+            Defaults to False.
+        train_cfg (dict): Training config of SABLRetinaHead.
+        test_cfg (dict): Testing config of SABLRetinaHead.
+        loss_cls (dict): Config of classification loss.
+        loss_bbox_cls (dict): Config of classification loss for bbox branch.
+        loss_bbox_reg (dict): Config of regression loss for bbox branch.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 feat_channels=256,
+                 approx_anchor_generator=dict(
+                     type='AnchorGenerator',
+                     octave_base_scale=4,
+                     scales_per_octave=3,
+                     ratios=[0.5, 1.0, 2.0],
+                     strides=[8, 16, 32, 64, 128]),
+                 square_anchor_generator=dict(
+                     type='AnchorGenerator',
+                     ratios=[1.0],
+                     scales=[4],
+                     strides=[8, 16, 32, 64, 128]),
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 bbox_coder=dict(
+                     type='BucketingBBoxCoder',
+                     num_buckets=14,
+                     scale_factor=3.0),
+                 reg_decoded_bbox=False,
+                 train_cfg=None,
+                 test_cfg=None,
+                 loss_cls=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 loss_bbox_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.5),
+                 loss_bbox_reg=dict(
+                     type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.5)):
+        super(SABLRetinaHead, self).__init__()
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.feat_channels = feat_channels
+        self.num_buckets = bbox_coder['num_buckets']
+        self.side_num = int(np.ceil(self.num_buckets / 2))
+
+        assert (approx_anchor_generator['octave_base_scale'] ==
+                square_anchor_generator['scales'][0])
+        assert (approx_anchor_generator['strides'] ==
+                square_anchor_generator['strides'])
+
+        self.approx_anchor_generator = build_anchor_generator(
+            approx_anchor_generator)
+        self.square_anchor_generator = build_anchor_generator(
+            square_anchor_generator)
+        self.approxs_per_octave = (
+            self.approx_anchor_generator.num_base_anchors[0])
+
+        # one anchor per location
+        self.num_anchors = 1
+        self.stacked_convs = stacked_convs
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+
+        self.reg_decoded_bbox = reg_decoded_bbox
+
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        self.sampling = loss_cls['type'] not in [
+            'FocalLoss', 'GHMC', 'QualityFocalLoss'
+        ]
+        if self.use_sigmoid_cls:
+            self.cls_out_channels = num_classes
+        else:
+            self.cls_out_channels = num_classes + 1
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox_cls = build_loss(loss_bbox_cls)
+        self.loss_bbox_reg = build_loss(loss_bbox_reg)
+
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # use PseudoSampler when sampling is False
+            if self.sampling and hasattr(self.train_cfg, 'sampler'):
+                sampler_cfg = self.train_cfg.sampler
+            else:
+                sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+        self.fp16_enabled = False
+        self._init_layers()
+
+    def _init_layers(self):
+        self.relu = nn.ReLU(inplace=True)
+        self.cls_convs = nn.ModuleList()
+        self.reg_convs = nn.ModuleList()
+        for i in range(self.stacked_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.cls_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+            self.reg_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        self.retina_cls = nn.Conv2d(
+            self.feat_channels, self.cls_out_channels, 3, padding=1)
+        self.retina_bbox_reg = nn.Conv2d(
+            self.feat_channels, self.side_num * 4, 3, padding=1)
+        self.retina_bbox_cls = nn.Conv2d(
+            self.feat_channels, self.side_num * 4, 3, padding=1)
+
+    def init_weights(self):
+        for m in self.cls_convs:
+            normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            normal_init(m.conv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.retina_cls, std=0.01, bias=bias_cls)
+        normal_init(self.retina_bbox_reg, std=0.01)
+        normal_init(self.retina_bbox_cls, std=0.01)
+
+    def forward_single(self, x):
+        cls_feat = x
+        reg_feat = x
+        for cls_conv in self.cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        for reg_conv in self.reg_convs:
+            reg_feat = reg_conv(reg_feat)
+        cls_score = self.retina_cls(cls_feat)
+        bbox_cls_pred = self.retina_bbox_cls(reg_feat)
+        bbox_reg_pred = self.retina_bbox_reg(reg_feat)
+        bbox_pred = (bbox_cls_pred, bbox_reg_pred)
+        return cls_score, bbox_pred
+
+    def forward(self, feats):
+        return multi_apply(self.forward_single, feats)
+
+    def get_anchors(self, featmap_sizes, img_metas, device='cuda'):
+        """Get squares according to feature map sizes and guided anchors.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+            device (torch.device | str): device for returned tensors
+
+        Returns:
+            tuple: square approxs of each image
+        """
+        num_imgs = len(img_metas)
+
+        # since feature map sizes of all images are the same, we only compute
+        # squares for one time
+        multi_level_squares = self.square_anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+        squares_list = [multi_level_squares for _ in range(num_imgs)]
+
+        return squares_list
+
+    def get_target(self,
+                   approx_list,
+                   inside_flag_list,
+                   square_list,
+                   gt_bboxes_list,
+                   img_metas,
+                   gt_bboxes_ignore_list=None,
+                   gt_labels_list=None,
+                   label_channels=None,
+                   sampling=True,
+                   unmap_outputs=True):
+        """Compute bucketing targets.
+        Args:
+            approx_list (list[list]): Multi level approxs of each image.
+            inside_flag_list (list[list]): Multi level inside flags of each
+                image.
+            square_list (list[list]): Multi level squares of each image.
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            img_metas (list[dict]): Meta info of each image.
+            gt_bboxes_ignore_list (list[Tensor]): ignore list of gt bboxes.
+            gt_bboxes_list (list[Tensor]): Gt bboxes of each image.
+            label_channels (int): Channel of label.
+            sampling (bool): Sample Anchors or not.
+            unmap_outputs (bool): unmap outputs or not.
+
+        Returns:
+            tuple: Returns a tuple containing learning targets.
+
+                - labels_list (list[Tensor]): Labels of each level.
+                - label_weights_list (list[Tensor]): Label weights of each \
+                    level.
+                - bbox_cls_targets_list (list[Tensor]): BBox cls targets of \
+                    each level.
+                - bbox_cls_weights_list (list[Tensor]): BBox cls weights of \
+                    each level.
+                - bbox_reg_targets_list (list[Tensor]): BBox reg targets of \
+                    each level.
+                - bbox_reg_weights_list (list[Tensor]): BBox reg weights of \
+                    each level.
+                - num_total_pos (int): Number of positive samples in all \
+                    images.
+                - num_total_neg (int): Number of negative samples in all \
+                    images.
+        """
+        num_imgs = len(img_metas)
+        assert len(approx_list) == len(inside_flag_list) == len(
+            square_list) == num_imgs
+        # anchor number of multi levels
+        num_level_squares = [squares.size(0) for squares in square_list[0]]
+        # concat all level anchors and flags to a single tensor
+        inside_flag_flat_list = []
+        approx_flat_list = []
+        square_flat_list = []
+        for i in range(num_imgs):
+            assert len(square_list[i]) == len(inside_flag_list[i])
+            inside_flag_flat_list.append(torch.cat(inside_flag_list[i]))
+            approx_flat_list.append(torch.cat(approx_list[i]))
+            square_flat_list.append(torch.cat(square_list[i]))
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        (all_labels, all_label_weights, all_bbox_cls_targets,
+         all_bbox_cls_weights, all_bbox_reg_targets, all_bbox_reg_weights,
+         pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single,
+             approx_flat_list,
+             inside_flag_flat_list,
+             square_flat_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             gt_labels_list,
+             img_metas,
+             label_channels=label_channels,
+             sampling=sampling,
+             unmap_outputs=unmap_outputs)
+        # no valid anchors
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        labels_list = images_to_levels(all_labels, num_level_squares)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_squares)
+        bbox_cls_targets_list = images_to_levels(all_bbox_cls_targets,
+                                                 num_level_squares)
+        bbox_cls_weights_list = images_to_levels(all_bbox_cls_weights,
+                                                 num_level_squares)
+        bbox_reg_targets_list = images_to_levels(all_bbox_reg_targets,
+                                                 num_level_squares)
+        bbox_reg_weights_list = images_to_levels(all_bbox_reg_weights,
+                                                 num_level_squares)
+        return (labels_list, label_weights_list, bbox_cls_targets_list,
+                bbox_cls_weights_list, bbox_reg_targets_list,
+                bbox_reg_weights_list, num_total_pos, num_total_neg)
+
+    def _get_target_single(self,
+                           flat_approxs,
+                           inside_flags,
+                           flat_squares,
+                           gt_bboxes,
+                           gt_bboxes_ignore,
+                           gt_labels,
+                           img_meta,
+                           label_channels=None,
+                           sampling=True,
+                           unmap_outputs=True):
+        """Compute regression and classification targets for anchors in a
+        single image.
+
+        Args:
+            flat_approxs (Tensor): flat approxs of a single image,
+                shape (n, 4)
+            inside_flags (Tensor): inside flags of a single image,
+                shape (n, ).
+            flat_squares (Tensor): flat squares of a single image,
+                shape (approxs_per_octave * n, 4)
+            gt_bboxes (Tensor): Ground truth bboxes of a single image, \
+                shape (num_gts, 4).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            img_meta (dict): Meta info of the image.
+            label_channels (int): Channel of label.
+            sampling (bool): Sample Anchors or not.
+            unmap_outputs (bool): unmap outputs or not.
+
+        Returns:
+            tuple:
+
+                - labels_list (Tensor): Labels in a single image
+                - label_weights (Tensor): Label weights in a single image
+                - bbox_cls_targets (Tensor): BBox cls targets in a single image
+                - bbox_cls_weights (Tensor): BBox cls weights in a single image
+                - bbox_reg_targets (Tensor): BBox reg targets in a single image
+                - bbox_reg_weights (Tensor): BBox reg weights in a single image
+                - num_total_pos (int): Number of positive samples \
+                    in a single image
+                - num_total_neg (int): Number of negative samples \
+                    in a single image
+        """
+        if not inside_flags.any():
+            return (None, ) * 8
+        # assign gt and sample anchors
+        expand_inside_flags = inside_flags[:, None].expand(
+            -1, self.approxs_per_octave).reshape(-1)
+        approxs = flat_approxs[expand_inside_flags, :]
+        squares = flat_squares[inside_flags, :]
+
+        assign_result = self.assigner.assign(approxs, squares,
+                                             self.approxs_per_octave,
+                                             gt_bboxes, gt_bboxes_ignore)
+        sampling_result = self.sampler.sample(assign_result, squares,
+                                              gt_bboxes)
+
+        num_valid_squares = squares.shape[0]
+        bbox_cls_targets = squares.new_zeros(
+            (num_valid_squares, self.side_num * 4))
+        bbox_cls_weights = squares.new_zeros(
+            (num_valid_squares, self.side_num * 4))
+        bbox_reg_targets = squares.new_zeros(
+            (num_valid_squares, self.side_num * 4))
+        bbox_reg_weights = squares.new_zeros(
+            (num_valid_squares, self.side_num * 4))
+        labels = squares.new_full((num_valid_squares, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = squares.new_zeros(num_valid_squares, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            (pos_bbox_reg_targets, pos_bbox_reg_weights, pos_bbox_cls_targets,
+             pos_bbox_cls_weights) = self.bbox_coder.encode(
+                 sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes)
+
+            bbox_cls_targets[pos_inds, :] = pos_bbox_cls_targets
+            bbox_reg_targets[pos_inds, :] = pos_bbox_reg_targets
+            bbox_cls_weights[pos_inds, :] = pos_bbox_cls_weights
+            bbox_reg_weights[pos_inds, :] = pos_bbox_reg_weights
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_squares.size(0)
+            labels = unmap(
+                labels, num_total_anchors, inside_flags, fill=self.num_classes)
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_cls_targets = unmap(bbox_cls_targets, num_total_anchors,
+                                     inside_flags)
+            bbox_cls_weights = unmap(bbox_cls_weights, num_total_anchors,
+                                     inside_flags)
+            bbox_reg_targets = unmap(bbox_reg_targets, num_total_anchors,
+                                     inside_flags)
+            bbox_reg_weights = unmap(bbox_reg_weights, num_total_anchors,
+                                     inside_flags)
+        return (labels, label_weights, bbox_cls_targets, bbox_cls_weights,
+                bbox_reg_targets, bbox_reg_weights, pos_inds, neg_inds)
+
+    def loss_single(self, cls_score, bbox_pred, labels, label_weights,
+                    bbox_cls_targets, bbox_cls_weights, bbox_reg_targets,
+                    bbox_reg_weights, num_total_samples):
+        # classification loss
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+        cls_score = cls_score.permute(0, 2, 3,
+                                      1).reshape(-1, self.cls_out_channels)
+        loss_cls = self.loss_cls(
+            cls_score, labels, label_weights, avg_factor=num_total_samples)
+        # regression loss
+        bbox_cls_targets = bbox_cls_targets.reshape(-1, self.side_num * 4)
+        bbox_cls_weights = bbox_cls_weights.reshape(-1, self.side_num * 4)
+        bbox_reg_targets = bbox_reg_targets.reshape(-1, self.side_num * 4)
+        bbox_reg_weights = bbox_reg_weights.reshape(-1, self.side_num * 4)
+        (bbox_cls_pred, bbox_reg_pred) = bbox_pred
+        bbox_cls_pred = bbox_cls_pred.permute(0, 2, 3, 1).reshape(
+            -1, self.side_num * 4)
+        bbox_reg_pred = bbox_reg_pred.permute(0, 2, 3, 1).reshape(
+            -1, self.side_num * 4)
+        loss_bbox_cls = self.loss_bbox_cls(
+            bbox_cls_pred,
+            bbox_cls_targets.long(),
+            bbox_cls_weights,
+            avg_factor=num_total_samples * 4 * self.side_num)
+        loss_bbox_reg = self.loss_bbox_reg(
+            bbox_reg_pred,
+            bbox_reg_targets,
+            bbox_reg_weights,
+            avg_factor=num_total_samples * 4 * self.bbox_coder.offset_topk)
+        return loss_cls, loss_bbox_cls, loss_bbox_reg
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.approx_anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        # get sampled approxes
+        approxs_list, inside_flag_list = GuidedAnchorHead.get_sampled_approxs(
+            self, featmap_sizes, img_metas, device=device)
+
+        square_list = self.get_anchors(featmap_sizes, img_metas, device=device)
+
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        cls_reg_targets = self.get_target(
+            approxs_list,
+            inside_flag_list,
+            square_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels,
+            sampling=self.sampling)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_cls_targets_list,
+         bbox_cls_weights_list, bbox_reg_targets_list, bbox_reg_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+        num_total_samples = (
+            num_total_pos + num_total_neg if self.sampling else num_total_pos)
+        losses_cls, losses_bbox_cls, losses_bbox_reg = multi_apply(
+            self.loss_single,
+            cls_scores,
+            bbox_preds,
+            labels_list,
+            label_weights_list,
+            bbox_cls_targets_list,
+            bbox_cls_weights_list,
+            bbox_reg_targets_list,
+            bbox_reg_weights_list,
+            num_total_samples=num_total_samples)
+        return dict(
+            loss_cls=losses_cls,
+            loss_bbox_cls=losses_bbox_cls,
+            loss_bbox_reg=losses_bbox_reg)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=False):
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+
+        device = cls_scores[0].device
+        mlvl_anchors = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_cls_pred_list = [
+                bbox_preds[i][0][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_reg_pred_list = [
+                bbox_preds[i][1][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self.get_bboxes_single(cls_score_list,
+                                               bbox_cls_pred_list,
+                                               bbox_reg_pred_list,
+                                               mlvl_anchors[img_id], img_shape,
+                                               scale_factor, cfg, rescale)
+            result_list.append(proposals)
+        return result_list
+
+    def get_bboxes_single(self,
+                          cls_scores,
+                          bbox_cls_preds,
+                          bbox_reg_preds,
+                          mlvl_anchors,
+                          img_shape,
+                          scale_factor,
+                          cfg,
+                          rescale=False):
+        cfg = self.test_cfg if cfg is None else cfg
+        mlvl_bboxes = []
+        mlvl_scores = []
+        mlvl_confids = []
+        assert len(cls_scores) == len(bbox_cls_preds) == len(
+            bbox_reg_preds) == len(mlvl_anchors)
+        for cls_score, bbox_cls_pred, bbox_reg_pred, anchors in zip(
+                cls_scores, bbox_cls_preds, bbox_reg_preds, mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_cls_pred.size(
+            )[-2:] == bbox_reg_pred.size()[-2::]
+            cls_score = cls_score.permute(1, 2,
+                                          0).reshape(-1, self.cls_out_channels)
+            if self.use_sigmoid_cls:
+                scores = cls_score.sigmoid()
+            else:
+                scores = cls_score.softmax(-1)
+            bbox_cls_pred = bbox_cls_pred.permute(1, 2, 0).reshape(
+                -1, self.side_num * 4)
+            bbox_reg_pred = bbox_reg_pred.permute(1, 2, 0).reshape(
+                -1, self.side_num * 4)
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                if self.use_sigmoid_cls:
+                    max_scores, _ = scores.max(dim=1)
+                else:
+                    max_scores, _ = scores[:, :-1].max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_cls_pred = bbox_cls_pred[topk_inds, :]
+                bbox_reg_pred = bbox_reg_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+            bbox_preds = [
+                bbox_cls_pred.contiguous(),
+                bbox_reg_pred.contiguous()
+            ]
+            bboxes, confids = self.bbox_coder.decode(
+                anchors.contiguous(), bbox_preds, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+            mlvl_confids.append(confids)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        mlvl_confids = torch.cat(mlvl_confids)
+        if self.use_sigmoid_cls:
+            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        det_bboxes, det_labels = multiclass_nms(
+            mlvl_bboxes,
+            mlvl_scores,
+            cfg.score_thr,
+            cfg.nms,
+            cfg.max_per_img,
+            score_factors=mlvl_confids)
+        return det_bboxes, det_labels
diff --git a/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py b/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py
new file mode 100755
index 000000000..44d64730a
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py
@@ -0,0 +1,1063 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, Scale, bias_init_with_prob, normal_init, DepthwiseSeparableConvModule
+from mmcv.cnn import constant_init, kaiming_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (anchor_inside_flags, bbox2distance, bbox_overlaps,
+                        build_assigner, build_sampler, distance2bbox, distance2kps, kps2distance,
+                        images_to_levels, multi_apply, multiclass_nms,
+                        reduce_mean, unmap)
+from ..builder import HEADS, build_loss
+from .anchor_head import AnchorHead
+from .base_dense_head import BaseDenseHead
+from .dense_test_mixins import BBoxTestMixin
+
+
+def last_zero_init(m):
+    if isinstance(m, nn.Sequential):
+        constant_init(m[-1], val=0)
+    else:
+        constant_init(m, val=0)
+
+
+class Integral(nn.Module):
+    """A fixed layer for calculating integral result from distribution.
+
+    This layer calculates the target location by :math: `sum{P(y_i) * y_i}`,
+    P(y_i) denotes the softmax vector that represents the discrete distribution
+    y_i denotes the discrete set, usually {0, 1, 2, ..., reg_max}
+
+    Args:
+        reg_max (int): The maximal value of the discrete set. Default: 16. You
+            may want to reset it according to your new dataset or related
+            settings.
+    """
+
+    def __init__(self, reg_max=16):
+        super(Integral, self).__init__()
+        self.reg_max = reg_max
+        self.register_buffer('project',
+                             torch.linspace(0, self.reg_max, self.reg_max + 1))
+
+    def forward(self, x):
+        """Forward feature from the regression head to get integral result of
+        bounding box location.
+
+        Args:
+            x (Tensor): Features of the regression head, shape (N, 4*(n+1)),
+                n is self.reg_max.
+
+        Returns:
+            x (Tensor): Integral result of box locations, i.e., distance
+                offsets from the box center in four directions, shape (N, 4).
+        """
+        x = F.softmax(x.reshape(-1, self.reg_max + 1), dim=1)
+        x = F.linear(x, self.project.type_as(x)).reshape(-1, 4)
+        return x
+
+
+@HEADS.register_module()
+class SCRFDHead(AnchorHead):
+    """Generalized Focal Loss: Learning Qualified and Distributed Bounding
+    Boxes for Dense Object Detection.
+
+    GFL head structure is similar with ATSS, however GFL uses
+    1) joint representation for classification and localization quality, and
+    2) flexible General distribution for bounding box locations,
+    which are supervised by
+    Quality Focal Loss (QFL) and Distribution Focal Loss (DFL), respectively
+
+    https://arxiv.org/abs/2006.04388
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        stacked_convs (int): Number of conv layers in cls and reg tower.
+            Default: 4.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: dict(type='GN', num_groups=32, requires_grad=True).
+        loss_qfl (dict): Config of Quality Focal Loss (QFL).
+        reg_max (int): Max value of integral set :math: `{0, ..., reg_max}`
+            in QFL setting. Default: 16.
+    Example:
+        >>> self = GFLHead(11, 7)
+        >>> feats = [torch.rand(1, 7, s, s) for s in [4, 8, 16, 32, 64]]
+        >>> cls_quality_score, bbox_pred = self.forward(feats)
+        >>> assert len(cls_quality_score) == len(self.scales)
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 stacked_convs=4,
+                 feat_mults=None,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+                 loss_dfl=None,
+                 reg_max=8,
+                 cls_reg_share=False,
+                 strides_share=True,
+                 scale_mode = 1,
+                 dw_conv = False,
+                 use_kps = False,
+                 loss_kps=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=0.1),
+                 #loss_kps=dict(type='SmoothL1Loss', beta=1.0, loss_weight=0.3),
+                 **kwargs):
+        self.stacked_convs = stacked_convs
+        self.feat_mults = feat_mults
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.reg_max = reg_max
+        self.cls_reg_share = cls_reg_share
+        self.strides_share = strides_share
+        self.scale_mode = scale_mode
+        self.use_dfl = True
+        self.dw_conv = dw_conv
+        self.NK = 5
+        self.extra_flops = 0.0
+        if loss_dfl is None or not loss_dfl:
+            self.use_dfl = False
+        self.use_scale = False
+        self.use_kps = use_kps
+        if self.scale_mode>0 and (self.strides_share or self.scale_mode==2):
+            self.use_scale = True
+        #print('USE-SCALE:', self.use_scale)
+        super(SCRFDHead, self).__init__(num_classes, in_channels, **kwargs)
+
+        self.sampling = False
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # SSD sampling=False so use PseudoSampler
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+        self.integral = Integral(self.reg_max)
+        if self.use_dfl:
+            self.loss_dfl = build_loss(loss_dfl)
+        #print('USE_DFL:', self.use_dfl)
+        self.loss_kps = build_loss(loss_kps)
+        self.loss_kps_std = 1.0
+        #print(self.bbox_coder.__class__)
+        self.train_step = 0
+        self.pos_count = {}
+        self.gtgroup_count = {}
+        for stride in self.anchor_generator.strides:
+            self.pos_count[stride[0]] = 0
+
+    def _get_conv_module(self, in_channel, out_channel):
+        if not self.dw_conv:
+            conv = ConvModule(
+                    in_channel,
+                    out_channel,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg)
+        else:
+            conv = DepthwiseSeparableConvModule(
+                    in_channel,
+                    out_channel,
+                    3,
+                    stride=1,
+                    padding=1,
+                    pw_norm_cfg=self.norm_cfg,
+                    dw_norm_cfg=self.norm_cfg)
+        return conv
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        conv_strides = [0] if self.strides_share else self.anchor_generator.strides
+        self.cls_stride_convs = nn.ModuleDict()
+        self.reg_stride_convs = nn.ModuleDict()
+        self.stride_cls = nn.ModuleDict()
+        self.stride_reg = nn.ModuleDict()
+        if self.use_kps:
+            self.stride_kps = nn.ModuleDict()
+        for stride_idx, conv_stride in enumerate(conv_strides):
+            #print('create convs for stride:', conv_stride)
+            key = str(conv_stride)
+            cls_convs = nn.ModuleList()
+            reg_convs = nn.ModuleList()
+            stacked_convs = self.stacked_convs[stride_idx] if isinstance(self.stacked_convs, (list, tuple)) else self.stacked_convs
+            feat_mult = self.feat_mults[stride_idx] if self.feat_mults is not None else 1
+            feat_ch = int(self.feat_channels*feat_mult)
+            for i in range(stacked_convs):
+                chn = self.in_channels if i == 0 else last_feat_ch
+                cls_convs.append( self._get_conv_module(chn, feat_ch) )
+                if not self.cls_reg_share:
+                    reg_convs.append( self._get_conv_module(chn, feat_ch) )
+                last_feat_ch = feat_ch
+            self.cls_stride_convs[key] = cls_convs
+            self.reg_stride_convs[key] = reg_convs
+            self.stride_cls[key] = nn.Conv2d(
+                feat_ch, self.cls_out_channels * self.num_anchors, 3, padding=1)
+            if not self.use_dfl:
+                self.stride_reg[key] = nn.Conv2d(
+                    feat_ch, 4 * self.num_anchors, 3, padding=1)
+            else:
+                self.stride_reg[key] = nn.Conv2d(
+                    feat_ch, 4 * (self.reg_max + 1) * self.num_anchors, 3, padding=1)
+            if self.use_kps:
+                self.stride_kps[key] = nn.Conv2d(
+                    feat_ch, self.NK*2*self.num_anchors, 3, padding=1)
+        #assert self.num_anchors == 1, 'anchor free version'
+        #extra_gflops /= 1e9
+        #print('extra_gflops: %.6fG'%extra_gflops)
+        if self.use_scale:
+            self.scales = nn.ModuleList(
+                [Scale(1.0) for _ in self.anchor_generator.strides])
+        else:
+            self.scales = [None for _ in self.anchor_generator.strides]
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for stride, cls_convs in self.cls_stride_convs.items():
+            #print('init cls for stride:', stride)
+            for m in cls_convs:
+                if not self.dw_conv:
+                    try:
+                        normal_init(m.conv, std=0.01)
+                    except:
+                        pass
+                else:
+                    normal_init(m.depthwise_conv.conv, std=0.01)
+                    normal_init(m.pointwise_conv.conv, std=0.01)
+        for stride, reg_convs in self.reg_stride_convs.items():
+            for m in reg_convs:
+                if not self.dw_conv:
+                    normal_init(m.conv, std=0.01)
+                else:
+                    normal_init(m.depthwise_conv.conv, std=0.01)
+                    normal_init(m.pointwise_conv.conv, std=0.01)
+        #bias_cls = bias_init_with_prob(0.01)
+        bias_cls = -4.595
+        #bias_cls = -1.595
+        for stride, conv in self.stride_cls.items():
+            normal_init(conv, std=0.01, bias=bias_cls)
+        for stride, conv in self.stride_reg.items():
+            normal_init(conv, std=0.01)
+        if self.use_kps:
+            for stride, conv in self.stride_kps.items():
+                normal_init(conv, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple: Usually a tuple of classification scores and bbox prediction
+                cls_scores (list[Tensor]): Classification and quality (IoU)
+                    joint scores for all scale levels, each is a 4D-tensor,
+                    the channel number is num_classes.
+                bbox_preds (list[Tensor]): Box distribution logits for all
+                    scale levels, each is a 4D-tensor, the channel number is
+                    4*(n+1), n is max value of integral set.
+        """
+        return multi_apply(self.forward_single, feats, self.scales, self.anchor_generator.strides)
+
+    def forward_single(self, x, scale, stride):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls and quality joint scores for a single
+                    scale level the channel number is num_classes.
+                bbox_pred (Tensor): Box distribution logits for a single scale
+                    level, the channel number is 4*(n+1), n is max value of
+                    integral set.
+        """
+        cls_feat = x
+        reg_feat = x
+        #print('forward_single in stride:', stride)
+        cls_convs = self.cls_stride_convs['0'] if self.strides_share else self.cls_stride_convs[str(stride)]
+        for cls_conv in cls_convs:
+            cls_feat = cls_conv(cls_feat)
+        if not self.cls_reg_share:
+            reg_convs = self.reg_stride_convs['0'] if self.strides_share else self.reg_stride_convs[str(stride)]
+            for reg_conv in reg_convs:
+                reg_feat = reg_conv(reg_feat)
+        else:
+            reg_feat = cls_feat
+        cls_pred_module = self.stride_cls['0'] if self.strides_share else self.stride_cls[str(stride)]
+        cls_score = cls_pred_module(cls_feat)
+        reg_pred_module = self.stride_reg['0'] if self.strides_share else self.stride_reg[str(stride)]
+        _bbox_pred = reg_pred_module(reg_feat)
+        if self.use_scale:
+            bbox_pred = scale(_bbox_pred)
+        else:
+            bbox_pred = _bbox_pred
+        if self.use_kps:
+            kps_pred_module = self.stride_kps['0'] if self.strides_share else self.stride_kps[str(stride)]
+            kps_pred = kps_pred_module(reg_feat)
+        else:
+            kps_pred = bbox_pred.new_zeros( (bbox_pred.shape[0], self.NK*2, bbox_pred.shape[2], bbox_pred.shape[3]) )
+        if torch.onnx.is_in_onnx_export():
+            assert not self.use_dfl
+            print('in-onnx-export', cls_score.shape, bbox_pred.shape)
+            #print(scale.parameters())
+            #for p in scale.parameters():
+                #print(p.name, p.data)
+                #scale_val = p.data.item()
+                #print(scale_val)
+            #print('EEE1', cls_score.shape)
+            #cls_score = torch.sigmoid(cls_score).reshape(1, self.cls_out_channels, -1).permute(0, 2, 1)
+            #print('EEE2', cls_score.shape)
+            #if self.use_dfl:
+            #    bbox_pred = self.integral(bbox_pred) * stride[0]
+            #else:
+            #    bbox_pred = bbox_pred.reshape( (-1,4) ) * stride[0]
+            #if self.use_dfl:
+            #    bbox_pred = bbox_pred.reshape(1, (self.reg_max+1)*4, -1).permute(0, 2, 1)
+            #    bbox_pred = bbox_pred.reshape( (1, -1, 4, self.reg_max+1) )
+            #    bbox_pred = F.softmax(bbox_pred, dim=3)
+            #else:
+            #    bbox_pred = bbox_pred.reshape(1, 4, -1).permute(0, 2, 1)
+            #kps_pred = kps_pred.reshape(1, 10, -1).permute(0, 2, 1)
+
+            # Add output batch dim, based on pull request #1593
+            batch_size = cls_score.shape[0]
+            cls_score = cls_score.permute(0, 2, 3, 1).reshape(batch_size, -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(0, 2, 3, 1).reshape(batch_size, -1, 4)
+            kps_pred = kps_pred.permute(0, 2, 3, 1).reshape(batch_size, -1, 10)
+
+        return cls_score, bbox_pred, kps_pred
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels=None,
+                      gt_keypointss=None,
+                      gt_bboxes_ignore=None,
+                      proposal_cfg=None,
+                      **kwargs):
+        """
+        Args:
+            x (list[Tensor]): Features from FPN.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            proposal_cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used
+
+        Returns:
+            tuple:
+                losses: (dict[str, Tensor]): A dictionary of loss components.
+                proposal_list (list[Tensor]): Proposals of each image.
+        """
+        outs = self(x)
+        if gt_labels is None:
+            loss_inputs = outs + (gt_bboxes, img_metas)
+        else:
+            loss_inputs = outs + (gt_bboxes, gt_labels, gt_keypointss, img_metas)
+        #print('AAA', gt_bboxes[0].shape, gt_keypointss[0].shape)
+        losses = self.loss(*loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
+        if proposal_cfg is None:
+            return losses
+        else:
+            proposal_list = self.get_bboxes(*outs, img_metas, cfg=proposal_cfg)
+            return losses, proposal_list
+
+    def get_anchors(self, featmap_sizes, img_metas, device='cuda'):
+        """Get anchors according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+            device (torch.device | str): Device for returned tensors
+
+        Returns:
+            tuple:
+                anchor_list (list[Tensor]): Anchors of each image.
+                valid_flag_list (list[Tensor]): Valid flags of each image.
+        """
+        num_imgs = len(img_metas)
+
+        # since feature map sizes of all images are the same, we only compute
+        # anchors for one time
+        multi_level_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device)
+        anchor_list = [multi_level_anchors for _ in range(num_imgs)]
+
+        # for each image, we compute valid flags of multi level anchors
+        valid_flag_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_flags = self.anchor_generator.valid_flags(
+                featmap_sizes, img_meta['pad_shape'], device)
+            valid_flag_list.append(multi_level_flags)
+
+        return anchor_list, valid_flag_list
+
+    def anchor_center(self, anchors):
+        """Get anchor centers from anchors.
+
+        Args:
+            anchors (Tensor): Anchor list with shape (N, 4), "xyxy" format.
+
+        Returns:
+            Tensor: Anchor centers with shape (N, 2), "xy" format.
+        """
+        anchors_cx = (anchors[:, 2] + anchors[:, 0]) / 2
+        anchors_cy = (anchors[:, 3] + anchors[:, 1]) / 2
+        return torch.stack([anchors_cx, anchors_cy], dim=-1)
+
+    def loss_single(self, anchors, cls_score, bbox_pred, kps_pred, labels, label_weights,
+                    bbox_targets, kps_targets, kps_weights, stride, num_total_samples):
+        """Compute loss of a single scale level.
+
+        Args:
+            anchors (Tensor): Box reference for each scale level with shape
+                (N, num_total_anchors, 4).
+            cls_score (Tensor): Cls and quality joint scores for each scale
+                level has shape (N, num_classes, H, W).
+            bbox_pred (Tensor): Box distribution logits for each scale
+                level with shape (N, 4*(n+1), H, W), n is max value of integral
+                set.
+            labels (Tensor): Labels of each anchors with shape
+                (N, num_total_anchors).
+            label_weights (Tensor): Label weights of each anchor with shape
+                (N, num_total_anchors)
+            bbox_targets (Tensor): BBox regression targets of each anchor wight
+                shape (N, num_total_anchors, 4).
+            stride (tuple): Stride in this scale level.
+            num_total_samples (int): Number of positive samples that is
+                reduced over all GPUs.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert stride[0] == stride[1], 'h stride is not equal to w stride!'
+        use_qscore = True
+        anchors = anchors.reshape(-1, 4)
+        cls_score = cls_score.permute(0, 2, 3,
+                                      1).reshape(-1, self.cls_out_channels)
+        if not self.use_dfl:
+            bbox_pred = bbox_pred.permute(0, 2, 3,
+                                          1).reshape(-1, 4)
+        else:
+            bbox_pred = bbox_pred.permute(0, 2, 3,
+                                          1).reshape(-1, 4 * (self.reg_max + 1))
+        bbox_targets = bbox_targets.reshape(-1, 4)
+        labels = labels.reshape(-1)
+        label_weights = label_weights.reshape(-1)
+
+        if self.use_kps:
+            kps_pred = kps_pred.permute(0, 2, 3,
+                                          1).reshape(-1, self.NK*2)
+            kps_targets = kps_targets.reshape( (-1, self.NK*2) )
+            kps_weights = kps_weights.reshape( (-1, self.NK*2) )
+            #print('AAA000', kps_targets.shape, kps_weights.shape)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((labels >= 0)
+                    & (labels < bg_class_ind)).nonzero().squeeze(1)
+        score = label_weights.new_zeros(labels.shape)
+
+        if len(pos_inds) > 0:
+            pos_bbox_targets = bbox_targets[pos_inds]
+            pos_bbox_pred = bbox_pred[pos_inds]
+            pos_anchors = anchors[pos_inds]
+            pos_anchor_centers = self.anchor_center(pos_anchors) / stride[0]
+
+            weight_targets = cls_score.detach().sigmoid()
+            weight_targets = weight_targets.max(dim=1)[0][pos_inds]
+            pos_decode_bbox_targets = pos_bbox_targets / stride[0]
+
+
+            if self.use_dfl:
+                pos_bbox_pred_corners = self.integral(pos_bbox_pred)
+                pos_decode_bbox_pred = distance2bbox(pos_anchor_centers,
+                                                     pos_bbox_pred_corners)
+            else:
+                pos_decode_bbox_pred = distance2bbox(pos_anchor_centers,
+                                                     pos_bbox_pred)
+            if self.use_kps:
+                pos_kps_targets = kps_targets[pos_inds]
+                pos_kps_pred = kps_pred[pos_inds]
+                #print('CCC000', kps_weights.shape)
+                pos_kps_weights = kps_weights.max(dim=1)[0][pos_inds] * weight_targets
+                #pos_kps_weights = kps_weights.max(dim=1)[0][pos_inds]
+                pos_kps_weights = pos_kps_weights.reshape( (-1, 1) )
+                #pos_kps_weights = kps_weights.max(dim=1, keepdims=True)[0][pos_inds]
+                #print('SSS', pos_kps_weights.sum())
+
+                #pos_decode_kps_targets = pos_kps_targets / stride[0]
+                #pos_decode_kps_pred = distance2kps(pos_anchor_centers, pos_kps_pred)
+
+                pos_decode_kps_targets = kps2distance(pos_anchor_centers, pos_kps_targets / stride[0])
+                pos_decode_kps_pred = pos_kps_pred
+                #print('ZZZ', pos_decode_kps_targets.shape, pos_decode_kps_pred.shape)
+                #print(pos_kps_weights[0,:].detach().cpu().numpy())
+                #print(pos_decode_kps_targets[0,:].detach().cpu().numpy())
+                #print(pos_decode_kps_pred[0,:].detach().cpu().numpy())
+
+
+                #print('CCC111', weight_targets.shape, pos_bbox_pred.shape, pos_decode_bbox_pred.shape, pos_kps_pred.shape, pos_decode_kps_pred.shape, pos_kps_weights.shape)
+
+            if use_qscore:
+                score[pos_inds] = bbox_overlaps(
+                    pos_decode_bbox_pred.detach(),
+                    pos_decode_bbox_targets,
+                    is_aligned=True)
+            else:
+                score[pos_inds] = 1.0
+
+            # regression loss
+            loss_bbox = self.loss_bbox(
+                pos_decode_bbox_pred,
+                pos_decode_bbox_targets,
+                weight=weight_targets,
+                avg_factor=1.0)
+
+            if self.use_kps:
+                loss_kps = self.loss_kps(
+                    pos_decode_kps_pred * self.loss_kps_std,
+                    pos_decode_kps_targets * self.loss_kps_std,
+                    weight=pos_kps_weights,
+                    avg_factor=1.0)
+            else:
+                loss_kps = kps_pred.sum() * 0
+
+            # dfl loss
+            if self.use_dfl:
+                pred_corners = pos_bbox_pred.reshape(-1, self.reg_max + 1)
+                target_corners = bbox2distance(pos_anchor_centers,
+                                               pos_decode_bbox_targets,
+                                               self.reg_max).reshape(-1)
+                loss_dfl = self.loss_dfl(
+                    pred_corners,
+                    target_corners,
+                    weight=weight_targets[:, None].expand(-1, 4).reshape(-1),
+                    avg_factor=4.0)
+            else:
+                loss_dfl = bbox_pred.sum() * 0
+        else:
+            loss_bbox = bbox_pred.sum() * 0
+            loss_dfl = bbox_pred.sum() * 0
+            loss_kps = kps_pred.sum() * 0
+            weight_targets = torch.tensor(0).cuda()
+
+        loss_cls = self.loss_cls(
+            cls_score, (labels, score),
+            weight=label_weights,
+            avg_factor=num_total_samples)
+
+
+        return loss_cls, loss_bbox, loss_dfl, loss_kps, weight_targets.sum()
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             kps_preds,
+             gt_bboxes,
+             gt_labels,
+             gt_keypointss,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Cls and quality scores for each scale
+                level has shape (N, num_classes, H, W).
+            bbox_preds (list[Tensor]): Box distribution logits for each scale
+                level with shape (N, 4*(n+1), H, W), n is max value of integral
+                set.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (list[Tensor] | None): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            gt_keypointss,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels)
+        if cls_reg_targets is None:
+            return None
+
+        (anchor_list, labels_list, label_weights_list, bbox_targets_list,
+         bbox_weights_list, keypoints_targets_list, keypoints_weights_list, num_total_pos, num_total_neg) = cls_reg_targets
+
+        num_total_samples = reduce_mean(
+            torch.tensor(num_total_pos, dtype=torch.float,
+                         device=device)).item()
+        num_total_samples = max(num_total_samples, 1.0)
+
+        losses_cls, losses_bbox, losses_dfl, losses_kps,\
+            avg_factor = multi_apply(
+                self.loss_single,
+                anchor_list,
+                cls_scores,
+                bbox_preds,
+                kps_preds,
+                labels_list,
+                label_weights_list,
+                bbox_targets_list,
+                keypoints_targets_list,
+                keypoints_weights_list,
+                self.anchor_generator.strides,
+                num_total_samples=num_total_samples)
+
+        #if self.train_step%100==0:
+        #    print('loss_cls:', losses_cls)
+        #    print('avg_factor:', avg_factor)
+
+
+        avg_factor = sum(avg_factor)
+        avg_factor = reduce_mean(avg_factor).item()
+        losses_bbox = list(map(lambda x: x / avg_factor, losses_bbox))
+        losses = dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
+        if self.use_kps:
+            losses_kps = list(map(lambda x: x / avg_factor, losses_kps))
+            losses['loss_kps'] = losses_kps
+        if self.use_dfl:
+            losses_dfl = list(map(lambda x: x / avg_factor, losses_dfl))
+            losses['loss_dfl'] = losses_dfl
+        return losses
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'kps_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   kps_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class labelof the
+                corresponding box.
+
+        Example:
+            >>> import mmcv
+            >>> self = AnchorHead(
+            >>>     num_classes=9,
+            >>>     in_channels=1,
+            >>>     anchor_generator=dict(
+            >>>         type='AnchorGenerator',
+            >>>         scales=[8],
+            >>>         ratios=[0.5, 1.0, 2.0],
+            >>>         strides=[4,]))
+            >>> img_metas = [{'img_shape': (32, 32, 3), 'scale_factor': 1}]
+            >>> cfg = mmcv.Config(dict(
+            >>>     score_thr=0.00,
+            >>>     nms=dict(type='nms', iou_thr=1.0),
+            >>>     max_per_img=10))
+            >>> feat = torch.rand(1, 1, 3, 3)
+            >>> cls_score, bbox_pred = self.forward_single(feat)
+            >>> # note the input lists are over different levels, not images
+            >>> cls_scores, bbox_preds = [cls_score], [bbox_pred]
+            >>> result_list = self.get_bboxes(cls_scores, bbox_preds,
+            >>>                               img_metas, cfg)
+            >>> det_bboxes, det_labels = result_list[0]
+            >>> assert len(result_list) == 1
+            >>> assert det_bboxes.shape[1] == 5
+            >>> assert len(det_bboxes) == len(det_labels) == cfg.max_per_img
+        """
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+
+        device = cls_scores[0].device
+        featmap_sizes = [cls_scores[i].shape[-2:] for i in range(num_levels)]
+        mlvl_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            if with_nms:
+                # some heads don't support with_nms argument
+                proposals = self._get_bboxes_single(cls_score_list,
+                                                    bbox_pred_list,
+                                                    mlvl_anchors, img_shape,
+                                                    scale_factor, cfg, rescale)
+            else:
+                proposals = self._get_bboxes_single(cls_score_list,
+                                                    bbox_pred_list,
+                                                    mlvl_anchors, img_shape,
+                                                    scale_factor, cfg, rescale,
+                                                    with_nms)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into labeled boxes.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for a single scale level
+                has shape (num_classes, H, W).
+            bbox_preds (list[Tensor]): Box distribution logits for a single
+                scale level with shape (4*(n+1), H, W), n is max value of
+                integral set.
+            mlvl_anchors (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): Bbox predictions in shape (N, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (N,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, stride, anchors in zip(
+                cls_scores, bbox_preds, self.anchor_generator.strides,
+                mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            assert stride[0] == stride[1]
+
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0)
+            if self.use_dfl:
+                bbox_pred = self.integral(bbox_pred) * stride[0]
+            else:
+                bbox_pred = bbox_pred.reshape( (-1,4) ) * stride[0]
+
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                max_scores, _ = scores.max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+
+            bboxes = distance2bbox(
+                self.anchor_center(anchors), bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+
+        mlvl_scores = torch.cat(mlvl_scores)
+        # Add a dummy background class to the backend when using sigmoid
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores
+
+    def get_targets(self,
+                    anchor_list,
+                    valid_flag_list,
+                    gt_bboxes_list,
+                    gt_keypointss_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None,
+                    gt_labels_list=None,
+                    label_channels=1,
+                    unmap_outputs=True):
+        """Get targets for GFL head.
+
+        This method is almost the same as `AnchorHead.get_targets()`. Besides
+        returning the targets as the parent method does, it also returns the
+        anchors as the first element of the returned tuple.
+        """
+        num_imgs = len(img_metas)
+        assert len(anchor_list) == len(valid_flag_list) == num_imgs
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+        num_level_anchors_list = [num_level_anchors] * num_imgs
+
+        # concat all level anchors and flags to a single tensor
+        for i in range(num_imgs):
+            assert len(anchor_list[i]) == len(valid_flag_list[i])
+            anchor_list[i] = torch.cat(anchor_list[i])
+            valid_flag_list[i] = torch.cat(valid_flag_list[i])
+
+        # compute targets for each image
+        if gt_bboxes_ignore_list is None:
+            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
+        if gt_labels_list is None:
+            gt_labels_list = [None for _ in range(num_imgs)]
+        if gt_keypointss_list is None:
+            gt_keypointss_list = [None for _ in range(num_imgs)]
+        #print('QQQ:', num_imgs, gt_bboxes_list[0].shape)
+        (all_anchors, all_labels, all_label_weights, all_bbox_targets,
+         all_bbox_weights, all_keypoints_targets, all_keypoints_weights, 
+         pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single,
+             anchor_list,
+             valid_flag_list,
+             num_level_anchors_list,
+             gt_bboxes_list,
+             gt_bboxes_ignore_list,
+             gt_labels_list,
+             gt_keypointss_list,
+             img_metas,
+             label_channels=label_channels,
+             unmap_outputs=unmap_outputs)
+        # no valid anchors
+        if any([labels is None for labels in all_labels]):
+            return None
+        # sampled anchors of all images
+        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
+        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
+        # split targets to a list w.r.t. multiple levels
+        anchors_list = images_to_levels(all_anchors, num_level_anchors)
+        labels_list = images_to_levels(all_labels, num_level_anchors)
+        label_weights_list = images_to_levels(all_label_weights,
+                                              num_level_anchors)
+        bbox_targets_list = images_to_levels(all_bbox_targets,
+                                             num_level_anchors)
+        bbox_weights_list = images_to_levels(all_bbox_weights,
+                                             num_level_anchors)
+        keypoints_targets_list = images_to_levels(all_keypoints_targets,
+                                             num_level_anchors)
+        keypoints_weights_list = images_to_levels(all_keypoints_weights,
+                                             num_level_anchors)
+        return (anchors_list, labels_list, label_weights_list,
+                bbox_targets_list, bbox_weights_list, keypoints_targets_list, keypoints_weights_list,
+                num_total_pos,
+                num_total_neg)
+
+    def _get_target_single(self,
+                           flat_anchors,
+                           valid_flags,
+                           num_level_anchors,
+                           gt_bboxes,
+                           gt_bboxes_ignore,
+                           gt_labels,
+                           gt_keypointss,
+                           img_meta,
+                           label_channels=1,
+                           unmap_outputs=True):
+        """Compute regression, classification targets for anchors in a single
+        image.
+
+        Args:
+            flat_anchors (Tensor): Multi-level anchors of the image, which are
+                concatenated into a single tensor of shape (num_anchors, 4)
+            valid_flags (Tensor): Multi level valid flags of the image,
+                which are concatenated into a single tensor of
+                    shape (num_anchors,).
+            num_level_anchors Tensor): Number of anchors of each scale level.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            img_meta (dict): Meta info of the image.
+            label_channels (int): Channel of label.
+            unmap_outputs (bool): Whether to map outputs back to the original
+                set of anchors.
+
+        Returns:
+            tuple: N is the number of total anchors in the image.
+                anchors (Tensor): All anchors in the image with shape (N, 4).
+                labels (Tensor): Labels of all anchors in the image with shape
+                    (N,).
+                label_weights (Tensor): Label weights of all anchor in the
+                    image with shape (N,).
+                bbox_targets (Tensor): BBox targets of all anchors in the
+                    image with shape (N, 4).
+                bbox_weights (Tensor): BBox weights of all anchors in the
+                    image with shape (N, 4).
+                pos_inds (Tensor): Indices of postive anchor with shape
+                    (num_pos,).
+                neg_inds (Tensor): Indices of negative anchor with shape
+                    (num_neg,).
+        """
+        inside_flags = anchor_inside_flags(flat_anchors, valid_flags,
+                                           img_meta['img_shape'][:2],
+                                           self.train_cfg.allowed_border)
+        if not inside_flags.any():
+            return (None, ) * 7
+        # assign gt and sample anchors
+        anchors = flat_anchors[inside_flags, :]
+
+        num_level_anchors_inside = self.get_num_level_anchors_inside(
+            num_level_anchors, inside_flags)
+        if self.assigner.__class__.__name__=='ATSSAssigner':
+            assign_result = self.assigner.assign(anchors, num_level_anchors_inside,
+                                                 gt_bboxes, gt_bboxes_ignore,
+                                                 gt_labels)
+        else:
+            assign_result = self.assigner.assign(anchors, 
+                                                 gt_bboxes, gt_bboxes_ignore,
+                                                 gt_labels)
+
+        sampling_result = self.sampler.sample(assign_result, anchors,
+                                              gt_bboxes)
+
+        num_valid_anchors = anchors.shape[0]
+        bbox_targets = torch.zeros_like(anchors)
+        bbox_weights = torch.zeros_like(anchors)
+        kps_targets = anchors.new_zeros(size=(anchors.shape[0], self.NK*2))
+        kps_weights = anchors.new_zeros(size=(anchors.shape[0], self.NK*2))
+        labels = anchors.new_full((num_valid_anchors, ),
+                                  self.num_classes,
+                                  dtype=torch.long)
+        label_weights = anchors.new_zeros(num_valid_anchors, dtype=torch.float)
+
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        if len(pos_inds) > 0:
+            pos_bbox_targets = sampling_result.pos_gt_bboxes
+            bbox_targets[pos_inds, :] = pos_bbox_targets
+            bbox_weights[pos_inds, :] = 1.0
+            if self.use_kps:
+                pos_assigned_gt_inds = sampling_result.pos_assigned_gt_inds
+                #print('BBB', anchors.shape, gt_bboxes.shape, gt_keypointss.shape, pos_inds.shape, bbox_targets.shape, pos_bbox_targets.shape)
+                kps_targets[pos_inds, :] = gt_keypointss[pos_assigned_gt_inds,:,:2].reshape( (-1, self.NK*2) )
+                kps_weights[pos_inds, :] = torch.mean(gt_keypointss[pos_assigned_gt_inds,:,2], dim=1, keepdims=True)
+            #kps_weights[pos_inds, :] = 1.0
+            if gt_labels is None:
+                # Only rpn gives gt_labels as None
+                # Foreground is the first class
+                labels[pos_inds] = 0
+            else:
+                labels[pos_inds] = gt_labels[
+                    sampling_result.pos_assigned_gt_inds]
+            if self.train_cfg.pos_weight <= 0:
+                label_weights[pos_inds] = 1.0
+            else:
+                label_weights[pos_inds] = self.train_cfg.pos_weight
+        if len(neg_inds) > 0:
+            label_weights[neg_inds] = 1.0
+
+        # map up to original set of anchors
+        if unmap_outputs:
+            num_total_anchors = flat_anchors.size(0)
+            anchors = unmap(anchors, num_total_anchors, inside_flags)
+            labels = unmap(
+                labels, num_total_anchors, inside_flags, fill=self.num_classes)
+            label_weights = unmap(label_weights, num_total_anchors,
+                                  inside_flags)
+            bbox_targets = unmap(bbox_targets, num_total_anchors, inside_flags)
+            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)
+            if self.use_kps:
+                kps_targets = unmap(kps_targets, num_total_anchors, inside_flags)
+                kps_weights = unmap(kps_weights, num_total_anchors, inside_flags)
+
+        return (anchors, labels, label_weights, bbox_targets, bbox_weights,
+                kps_targets, kps_weights,
+                pos_inds, neg_inds)
+
+    def get_num_level_anchors_inside(self, num_level_anchors, inside_flags):
+        split_inside_flags = torch.split(inside_flags, num_level_anchors)
+        num_level_anchors_inside = [
+            int(flags.sum()) for flags in split_inside_flags
+        ]
+        return num_level_anchors_inside
+
+    def aug_test(self, feats, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            feats (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains features for all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[ndarray]: bbox results of each class
+        """
+        return self.aug_test_bboxes(feats, img_metas, rescale=rescale)
diff --git a/detection/scrfd/mmdet/models/dense_heads/ssd_head.py b/detection/scrfd/mmdet/models/dense_heads/ssd_head.py
new file mode 100755
index 000000000..42554c12c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/ssd_head.py
@@ -0,0 +1,259 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import xavier_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (build_anchor_generator, build_assigner,
+                        build_bbox_coder, build_sampler, multi_apply)
+from ..builder import HEADS
+from ..losses import smooth_l1_loss
+from .anchor_head import AnchorHead
+
+
+# TODO: add loss evaluator for SSD
+@HEADS.register_module()
+class SSDHead(AnchorHead):
+    """SSD head used in https://arxiv.org/abs/1512.02325.
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        anchor_generator (dict): Config dict for anchor generator
+        bbox_coder (dict): Config of bounding box coder.
+        reg_decoded_bbox (bool): If true, the regression loss would be
+            applied on decoded bounding boxes. Default: False
+        train_cfg (dict): Training config of anchor head.
+        test_cfg (dict): Testing config of anchor head.
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_classes=80,
+                 in_channels=(512, 1024, 512, 256, 256, 256),
+                 anchor_generator=dict(
+                     type='SSDAnchorGenerator',
+                     scale_major=False,
+                     input_size=300,
+                     strides=[8, 16, 32, 64, 100, 300],
+                     ratios=([2], [2, 3], [2, 3], [2, 3], [2], [2]),
+                     basesize_ratio_range=(0.1, 0.9)),
+                 bbox_coder=dict(
+                     type='DeltaXYWHBBoxCoder',
+                     clip_border=True,
+                     target_means=[.0, .0, .0, .0],
+                     target_stds=[1.0, 1.0, 1.0, 1.0],
+                 ),
+                 reg_decoded_bbox=False,
+                 train_cfg=None,
+                 test_cfg=None):
+        super(AnchorHead, self).__init__()
+        self.num_classes = num_classes
+        self.in_channels = in_channels
+        self.cls_out_channels = num_classes + 1  # add background class
+        self.anchor_generator = build_anchor_generator(anchor_generator)
+        num_anchors = self.anchor_generator.num_base_anchors
+
+        reg_convs = []
+        cls_convs = []
+        for i in range(len(in_channels)):
+            reg_convs.append(
+                nn.Conv2d(
+                    in_channels[i],
+                    num_anchors[i] * 4,
+                    kernel_size=3,
+                    padding=1))
+            cls_convs.append(
+                nn.Conv2d(
+                    in_channels[i],
+                    num_anchors[i] * (num_classes + 1),
+                    kernel_size=3,
+                    padding=1))
+        self.reg_convs = nn.ModuleList(reg_convs)
+        self.cls_convs = nn.ModuleList(cls_convs)
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.reg_decoded_bbox = reg_decoded_bbox
+        self.use_sigmoid_cls = False
+        self.cls_focal_loss = False
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        # set sampling=False for archor_target
+        self.sampling = False
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            # SSD sampling=False so use PseudoSampler
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+        self.fp16_enabled = False
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform', bias=0)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple:
+                cls_scores (list[Tensor]): Classification scores for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * num_classes.
+                bbox_preds (list[Tensor]): Box energies / deltas for all scale
+                    levels, each is a 4D-tensor, the channels number is
+                    num_anchors * 4.
+        """
+        cls_scores = []
+        bbox_preds = []
+        for feat, reg_conv, cls_conv in zip(feats, self.reg_convs,
+                                            self.cls_convs):
+            cls_scores.append(cls_conv(feat))
+            bbox_preds.append(reg_conv(feat))
+        return cls_scores, bbox_preds
+
+    def loss_single(self, cls_score, bbox_pred, anchor, labels, label_weights,
+                    bbox_targets, bbox_weights, num_total_samples):
+        """Compute loss of a single image.
+
+        Args:
+            cls_score (Tensor): Box scores for eachimage
+                Has shape (num_total_anchors, num_classes).
+            bbox_pred (Tensor): Box energies / deltas for each image
+                level with shape (num_total_anchors, 4).
+            anchors (Tensor): Box reference for each scale level with shape
+                (num_total_anchors, 4).
+            labels (Tensor): Labels of each anchors with shape
+                (num_total_anchors,).
+            label_weights (Tensor): Label weights of each anchor with shape
+                (num_total_anchors,)
+            bbox_targets (Tensor): BBox regression targets of each anchor wight
+                shape (num_total_anchors, 4).
+            bbox_weights (Tensor): BBox regression loss weights of each anchor
+                with shape (num_total_anchors, 4).
+            num_total_samples (int): If sampling, num total samples equal to
+                the number of total anchors; Otherwise, it is the number of
+                positive anchors.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+
+        loss_cls_all = F.cross_entropy(
+            cls_score, labels, reduction='none') * label_weights
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        pos_inds = ((labels >= 0) &
+                    (labels < self.num_classes)).nonzero().reshape(-1)
+        neg_inds = (labels == self.num_classes).nonzero().view(-1)
+
+        num_pos_samples = pos_inds.size(0)
+        num_neg_samples = self.train_cfg.neg_pos_ratio * num_pos_samples
+        if num_neg_samples > neg_inds.size(0):
+            num_neg_samples = neg_inds.size(0)
+        topk_loss_cls_neg, _ = loss_cls_all[neg_inds].topk(num_neg_samples)
+        loss_cls_pos = loss_cls_all[pos_inds].sum()
+        loss_cls_neg = topk_loss_cls_neg.sum()
+        loss_cls = (loss_cls_pos + loss_cls_neg) / num_total_samples
+
+        if self.reg_decoded_bbox:
+            bbox_pred = self.bbox_coder.decode(anchor, bbox_pred)
+
+        loss_bbox = smooth_l1_loss(
+            bbox_pred,
+            bbox_targets,
+            bbox_weights,
+            beta=self.train_cfg.smoothl1_beta,
+            avg_factor=num_total_samples)
+        return loss_cls[None], loss_bbox
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute losses of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=1,
+            unmap_outputs=False)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+
+        num_images = len(img_metas)
+        all_cls_scores = torch.cat([
+            s.permute(0, 2, 3, 1).reshape(
+                num_images, -1, self.cls_out_channels) for s in cls_scores
+        ], 1)
+        all_labels = torch.cat(labels_list, -1).view(num_images, -1)
+        all_label_weights = torch.cat(label_weights_list,
+                                      -1).view(num_images, -1)
+        all_bbox_preds = torch.cat([
+            b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
+            for b in bbox_preds
+        ], -2)
+        all_bbox_targets = torch.cat(bbox_targets_list,
+                                     -2).view(num_images, -1, 4)
+        all_bbox_weights = torch.cat(bbox_weights_list,
+                                     -2).view(num_images, -1, 4)
+
+        # concat all level anchors to a single tensor
+        all_anchors = []
+        for i in range(num_images):
+            all_anchors.append(torch.cat(anchor_list[i]))
+
+        # check NaN and Inf
+        assert torch.isfinite(all_cls_scores).all().item(), \
+            'classification scores become infinite or NaN!'
+        assert torch.isfinite(all_bbox_preds).all().item(), \
+            'bbox predications become infinite or NaN!'
+
+        losses_cls, losses_bbox = multi_apply(
+            self.loss_single,
+            all_cls_scores,
+            all_bbox_preds,
+            all_anchors,
+            all_labels,
+            all_label_weights,
+            all_bbox_targets,
+            all_bbox_weights,
+            num_total_samples=num_total_pos)
+        return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
diff --git a/detection/scrfd/mmdet/models/dense_heads/transformer_head.py b/detection/scrfd/mmdet/models/dense_heads/transformer_head.py
new file mode 100755
index 000000000..df3fc9411
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/transformer_head.py
@@ -0,0 +1,655 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import Conv2d, Linear, build_activation_layer
+from mmcv.runner import force_fp32
+
+from mmdet.core import (bbox_cxcywh_to_xyxy, bbox_xyxy_to_cxcywh,
+                        build_assigner, build_sampler, multi_apply,
+                        reduce_mean)
+from mmdet.models.utils import (FFN, build_positional_encoding,
+                                build_transformer)
+from ..builder import HEADS, build_loss
+from .anchor_free_head import AnchorFreeHead
+
+
+@HEADS.register_module()
+class TransformerHead(AnchorFreeHead):
+    """Implements the DETR transformer head.
+
+    See `paper: End-to-End Object Detection with Transformers
+    <https://arxiv.org/pdf/2005.12872>`_ for details.
+
+    Args:
+        num_classes (int): Number of categories excluding the background.
+        in_channels (int): Number of channels in the input feature map.
+        num_fcs (int, optional): Number of fully-connected layers used in
+            `FFN`, which is then used for the regression head. Default 2.
+        transformer (dict, optional): Config for transformer.
+        positional_encoding (dict, optional): Config for position encoding.
+        loss_cls (dict, optional): Config of the classification loss.
+            Default `CrossEntropyLoss`.
+        loss_bbox (dict, optional): Config of the regression loss.
+            Default `L1Loss`.
+        loss_iou (dict, optional): Config of the regression iou loss.
+            Default `GIoULoss`.
+        tran_cfg (dict, optional): Training config of transformer head.
+        test_cfg (dict, optional): Testing config of transformer head.
+
+    Example:
+        >>> import torch
+        >>> self = TransformerHead(80, 2048)
+        >>> x = torch.rand(1, 2048, 32, 32)
+        >>> mask = torch.ones(1, 32, 32).to(x.dtype)
+        >>> mask[:, :16, :15] = 0
+        >>> all_cls_scores, all_bbox_preds = self(x, mask)
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 num_fcs=2,
+                 transformer=dict(
+                     type='Transformer',
+                     embed_dims=256,
+                     num_heads=8,
+                     num_encoder_layers=6,
+                     num_decoder_layers=6,
+                     feedforward_channels=2048,
+                     dropout=0.1,
+                     act_cfg=dict(type='ReLU', inplace=True),
+                     norm_cfg=dict(type='LN'),
+                     num_fcs=2,
+                     pre_norm=False,
+                     return_intermediate_dec=True),
+                 positional_encoding=dict(
+                     type='SinePositionalEncoding',
+                     num_feats=128,
+                     normalize=True),
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     bg_cls_weight=0.1,
+                     use_sigmoid=False,
+                     loss_weight=1.0,
+                     class_weight=1.0),
+                 loss_bbox=dict(type='L1Loss', loss_weight=5.0),
+                 loss_iou=dict(type='GIoULoss', loss_weight=2.0),
+                 train_cfg=dict(
+                     assigner=dict(
+                         type='HungarianAssigner',
+                         cls_weight=1.,
+                         bbox_weight=5.,
+                         iou_weight=2.,
+                         iou_calculator=dict(type='BboxOverlaps2D'),
+                         iou_mode='giou')),
+                 test_cfg=dict(max_per_img=100),
+                 **kwargs):
+        # NOTE here use `AnchorFreeHead` instead of `TransformerHead`,
+        # since it brings inconvenience when the initialization of
+        # `AnchorFreeHead` is called.
+        super(AnchorFreeHead, self).__init__()
+        use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        assert not use_sigmoid_cls, 'setting use_sigmoid_cls as True is ' \
+            'not supported in DETR, since background is needed for the ' \
+            'matching process.'
+        assert 'embed_dims' in transformer \
+            and 'num_feats' in positional_encoding
+        num_feats = positional_encoding['num_feats']
+        embed_dims = transformer['embed_dims']
+        assert num_feats * 2 == embed_dims, 'embed_dims should' \
+            f' be exactly 2 times of num_feats. Found {embed_dims}' \
+            f' and {num_feats}.'
+        assert test_cfg is not None and 'max_per_img' in test_cfg
+
+        class_weight = loss_cls.get('class_weight', None)
+        if class_weight is not None:
+            assert isinstance(class_weight, float), 'Expected ' \
+                'class_weight to have type float. Found ' \
+                f'{type(class_weight)}.'
+            # NOTE following the official DETR rep0, bg_cls_weight means
+            # relative classification weight of the no-object class.
+            bg_cls_weight = loss_cls.get('bg_cls_weight', class_weight)
+            assert isinstance(bg_cls_weight, float), 'Expected ' \
+                'bg_cls_weight to have type float. Found ' \
+                f'{type(bg_cls_weight)}.'
+            class_weight = torch.ones(num_classes + 1) * class_weight
+            # set background class as the last indice
+            class_weight[num_classes] = bg_cls_weight
+            loss_cls.update({'class_weight': class_weight})
+            if 'bg_cls_weight' in loss_cls:
+                loss_cls.pop('bg_cls_weight')
+            self.bg_cls_weight = bg_cls_weight
+
+        if train_cfg:
+            assert 'assigner' in train_cfg, 'assigner should be provided '\
+                'when train_cfg is set.'
+            assigner = train_cfg['assigner']
+            assert loss_cls['loss_weight'] == assigner['cls_weight'], \
+                'The classification weight for loss and matcher should be' \
+                'exactly the same.'
+            assert loss_bbox['loss_weight'] == assigner['bbox_weight'], \
+                'The regression L1 weight for loss and matcher should be' \
+                'exactly the same.'
+            assert loss_iou['loss_weight'] == assigner['iou_weight'], \
+                'The regression iou weight for loss and matcher should be' \
+                'exactly the same.'
+            self.assigner = build_assigner(assigner)
+            # DETR sampling=False, so use PseudoSampler
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+        self.num_classes = num_classes
+        self.cls_out_channels = num_classes + 1
+        self.in_channels = in_channels
+        self.num_fcs = num_fcs
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        self.use_sigmoid_cls = use_sigmoid_cls
+        self.embed_dims = embed_dims
+        self.num_query = test_cfg['max_per_img']
+        self.fp16_enabled = False
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox = build_loss(loss_bbox)
+        self.loss_iou = build_loss(loss_iou)
+        self.act_cfg = transformer.get('act_cfg',
+                                       dict(type='ReLU', inplace=True))
+        self.activate = build_activation_layer(self.act_cfg)
+        self.positional_encoding = build_positional_encoding(
+            positional_encoding)
+        self.transformer = build_transformer(transformer)
+        self._init_layers()
+
+    def _init_layers(self):
+        """Initialize layers of the transformer head."""
+        self.input_proj = Conv2d(
+            self.in_channels, self.embed_dims, kernel_size=1)
+        self.fc_cls = Linear(self.embed_dims, self.cls_out_channels)
+        self.reg_ffn = FFN(
+            self.embed_dims,
+            self.embed_dims,
+            self.num_fcs,
+            self.act_cfg,
+            dropout=0.0,
+            add_residual=False)
+        self.fc_reg = Linear(self.embed_dims, 4)
+        self.query_embedding = nn.Embedding(self.num_query, self.embed_dims)
+
+    def init_weights(self, distribution='uniform'):
+        """Initialize weights of the transformer head."""
+        # The initialization for transformer is important
+        self.transformer.init_weights()
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        """load checkpoints."""
+        # NOTE here use `AnchorFreeHead` instead of `TransformerHead`,
+        # since `AnchorFreeHead._load_from_state_dict` should not be
+        # called here. Invoking the default `Module._load_from_state_dict`
+        # is enough.
+        super(AnchorFreeHead,
+              self)._load_from_state_dict(state_dict, prefix, local_metadata,
+                                          strict, missing_keys,
+                                          unexpected_keys, error_msgs)
+
+    def forward(self, feats, img_metas):
+        """Forward function.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+            img_metas (list[dict]): List of image information.
+
+        Returns:
+            tuple[list[Tensor], list[Tensor]]: Outputs for all scale levels.
+
+                - all_cls_scores_list (list[Tensor]): Classification scores \
+                    for each scale level. Each is a 4D-tensor with shape \
+                    [nb_dec, bs, num_query, cls_out_channels]. Note \
+                    `cls_out_channels` should includes background.
+                - all_bbox_preds_list (list[Tensor]): Sigmoid regression \
+                    outputs for each scale level. Each is a 4D-tensor with \
+                    normalized coordinate format (cx, cy, w, h) and shape \
+                    [nb_dec, bs, num_query, 4].
+        """
+        num_levels = len(feats)
+        img_metas_list = [img_metas for _ in range(num_levels)]
+        return multi_apply(self.forward_single, feats, img_metas_list)
+
+    def forward_single(self, x, img_metas):
+        """"Forward function for a single feature level.
+
+        Args:
+            x (Tensor): Input feature from backbone's single stage, shape
+                [bs, c, h, w].
+            img_metas (list[dict]): List of image information.
+
+        Returns:
+            all_cls_scores (Tensor): Outputs from the classification head,
+                shape [nb_dec, bs, num_query, cls_out_channels]. Note
+                cls_out_channels should includes background.
+            all_bbox_preds (Tensor): Sigmoid outputs from the regression
+                head with normalized coordinate format (cx, cy, w, h).
+                Shape [nb_dec, bs, num_query, 4].
+        """
+        # construct binary masks which used for the transformer.
+        # NOTE following the official DETR repo, non-zero values representing
+        # ignored positions, while zero values means valid positions.
+        batch_size = x.size(0)
+        input_img_h, input_img_w = img_metas[0]['batch_input_shape']
+        masks = x.new_ones((batch_size, input_img_h, input_img_w))
+        for img_id in range(batch_size):
+            img_h, img_w, _ = img_metas[img_id]['img_shape']
+            masks[img_id, :img_h, :img_w] = 0
+
+        x = self.input_proj(x)
+        # interpolate masks to have the same spatial shape with x
+        masks = F.interpolate(
+            masks.unsqueeze(1), size=x.shape[-2:]).to(torch.bool).squeeze(1)
+        # position encoding
+        pos_embed = self.positional_encoding(masks)  # [bs, embed_dim, h, w]
+        # outs_dec: [nb_dec, bs, num_query, embed_dim]
+        outs_dec, _ = self.transformer(x, masks, self.query_embedding.weight,
+                                       pos_embed)
+
+        all_cls_scores = self.fc_cls(outs_dec)
+        all_bbox_preds = self.fc_reg(self.activate(
+            self.reg_ffn(outs_dec))).sigmoid()
+        return all_cls_scores, all_bbox_preds
+
+    @force_fp32(apply_to=('all_cls_scores_list', 'all_bbox_preds_list'))
+    def loss(self,
+             all_cls_scores_list,
+             all_bbox_preds_list,
+             gt_bboxes_list,
+             gt_labels_list,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """"Loss function.
+
+        Only outputs from the last feature level are used for computing
+        losses by default.
+
+        Args:
+            all_cls_scores_list (list[Tensor]): Classification outputs
+                for each feature level. Each is a 4D-tensor with shape
+                [nb_dec, bs, num_query, cls_out_channels].
+            all_bbox_preds_list (list[Tensor]): Sigmoid regression
+                outputs for each feature level. Each is a 4D-tensor with
+                normalized coordinate format (cx, cy, w, h) and shape
+                [nb_dec, bs, num_query, 4].
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes for each image
+                with shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+            gt_bboxes_ignore (list[Tensor], optional): Bounding boxes
+                which can be ignored for each image. Default None.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        # NOTE defaultly only the outputs from the last feature scale is used.
+        all_cls_scores = all_cls_scores_list[-1]
+        all_bbox_preds = all_bbox_preds_list[-1]
+        assert gt_bboxes_ignore is None, \
+            'Only supports for gt_bboxes_ignore setting to None.'
+
+        num_dec_layers = len(all_cls_scores)
+        all_gt_bboxes_list = [gt_bboxes_list for _ in range(num_dec_layers)]
+        all_gt_labels_list = [gt_labels_list for _ in range(num_dec_layers)]
+        all_gt_bboxes_ignore_list = [
+            gt_bboxes_ignore for _ in range(num_dec_layers)
+        ]
+        img_metas_list = [img_metas for _ in range(num_dec_layers)]
+
+        losses_cls, losses_bbox, losses_iou = multi_apply(
+            self.loss_single, all_cls_scores, all_bbox_preds,
+            all_gt_bboxes_list, all_gt_labels_list, img_metas_list,
+            all_gt_bboxes_ignore_list)
+
+        loss_dict = dict()
+        # loss from the last decoder layer
+        loss_dict['loss_cls'] = losses_cls[-1]
+        loss_dict['loss_bbox'] = losses_bbox[-1]
+        loss_dict['loss_iou'] = losses_iou[-1]
+        # loss from other decoder layers
+        num_dec_layer = 0
+        for loss_cls_i, loss_bbox_i, loss_iou_i in zip(losses_cls[:-1],
+                                                       losses_bbox[:-1],
+                                                       losses_iou[:-1]):
+            loss_dict[f'd{num_dec_layer}.loss_cls'] = loss_cls_i
+            loss_dict[f'd{num_dec_layer}.loss_bbox'] = loss_bbox_i
+            loss_dict[f'd{num_dec_layer}.loss_iou'] = loss_iou_i
+            num_dec_layer += 1
+        return loss_dict
+
+    def loss_single(self,
+                    cls_scores,
+                    bbox_preds,
+                    gt_bboxes_list,
+                    gt_labels_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None):
+        """"Loss function for outputs from a single decoder layer of a single
+        feature level.
+
+        Args:
+            cls_scores (Tensor): Box score logits from a single decoder layer
+                for all images. Shape [bs, num_query, cls_out_channels].
+            bbox_preds (Tensor): Sigmoid outputs from a single decoder layer
+                for all images, with normalized coordinate (cx, cy, w, h) and
+                shape [bs, num_query, 4].
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes for each image
+                with shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+            gt_bboxes_ignore_list (list[Tensor], optional): Bounding
+                boxes which can be ignored for each image. Default None.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components for outputs from
+                a single decoder layer.
+        """
+        num_imgs = cls_scores.size(0)
+        cls_scores_list = [cls_scores[i] for i in range(num_imgs)]
+        bbox_preds_list = [bbox_preds[i] for i in range(num_imgs)]
+        cls_reg_targets = self.get_targets(cls_scores_list, bbox_preds_list,
+                                           gt_bboxes_list, gt_labels_list,
+                                           img_metas, gt_bboxes_ignore_list)
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+        labels = torch.cat(labels_list, 0)
+        label_weights = torch.cat(label_weights_list, 0)
+        bbox_targets = torch.cat(bbox_targets_list, 0)
+        bbox_weights = torch.cat(bbox_weights_list, 0)
+
+        # classification loss
+        cls_scores = cls_scores.reshape(-1, self.cls_out_channels)
+        # construct weighted avg_factor to match with the official DETR repo
+        cls_avg_factor = num_total_pos * 1.0 + \
+            num_total_neg * self.bg_cls_weight
+        loss_cls = self.loss_cls(
+            cls_scores, labels, label_weights, avg_factor=cls_avg_factor)
+
+        # Compute the average number of gt boxes accross all gpus, for
+        # normalization purposes
+        num_total_pos = loss_cls.new_tensor([num_total_pos])
+        num_total_pos = torch.clamp(reduce_mean(num_total_pos), min=1).item()
+
+        # construct factors used for rescale bboxes
+        factors = []
+        for img_meta, bbox_pred in zip(img_metas, bbox_preds):
+            img_h, img_w, _ = img_meta['img_shape']
+            factor = bbox_pred.new_tensor([img_w, img_h, img_w,
+                                           img_h]).unsqueeze(0).repeat(
+                                               bbox_pred.size(0), 1)
+            factors.append(factor)
+        factors = torch.cat(factors, 0)
+
+        # DETR regress the relative position of boxes (cxcywh) in the image,
+        # thus the learning target is normalized by the image size. So here
+        # we need to re-scale them for calculating IoU loss
+        bbox_preds = bbox_preds.reshape(-1, 4)
+        bboxes = bbox_cxcywh_to_xyxy(bbox_preds) * factors
+        bboxes_gt = bbox_cxcywh_to_xyxy(bbox_targets) * factors
+
+        # regression IoU loss, defaultly GIoU loss
+        loss_iou = self.loss_iou(
+            bboxes, bboxes_gt, bbox_weights, avg_factor=num_total_pos)
+
+        # regression L1 loss
+        loss_bbox = self.loss_bbox(
+            bbox_preds, bbox_targets, bbox_weights, avg_factor=num_total_pos)
+        return loss_cls, loss_bbox, loss_iou
+
+    def get_targets(self,
+                    cls_scores_list,
+                    bbox_preds_list,
+                    gt_bboxes_list,
+                    gt_labels_list,
+                    img_metas,
+                    gt_bboxes_ignore_list=None):
+        """"Compute regression and classification targets for a batch image.
+
+        Outputs from a single decoder layer of a single feature level are used.
+
+        Args:
+            cls_scores_list (list[Tensor]): Box score logits from a single
+                decoder layer for each image with shape [num_query,
+                cls_out_channels].
+            bbox_preds_list (list[Tensor]): Sigmoid outputs from a single
+                decoder layer for each image, with normalized coordinate
+                (cx, cy, w, h) and shape [num_query, 4].
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes for each image
+                with shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+            gt_bboxes_ignore_list (list[Tensor], optional): Bounding
+                boxes which can be ignored for each image. Default None.
+
+        Returns:
+            tuple: a tuple containing the following targets.
+
+                - labels_list (list[Tensor]): Labels for all images.
+                - label_weights_list (list[Tensor]): Label weights for all \
+                    images.
+                - bbox_targets_list (list[Tensor]): BBox targets for all \
+                    images.
+                - bbox_weights_list (list[Tensor]): BBox weights for all \
+                    images.
+                - num_total_pos (int): Number of positive samples in all \
+                    images.
+                - num_total_neg (int): Number of negative samples in all \
+                    images.
+        """
+        assert gt_bboxes_ignore_list is None, \
+            'Only supports for gt_bboxes_ignore setting to None.'
+        num_imgs = len(cls_scores_list)
+        gt_bboxes_ignore_list = [
+            gt_bboxes_ignore_list for _ in range(num_imgs)
+        ]
+
+        (labels_list, label_weights_list, bbox_targets_list,
+         bbox_weights_list, pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single, cls_scores_list, bbox_preds_list,
+             gt_bboxes_list, gt_labels_list, img_metas, gt_bboxes_ignore_list)
+        num_total_pos = sum((inds.numel() for inds in pos_inds_list))
+        num_total_neg = sum((inds.numel() for inds in neg_inds_list))
+        return (labels_list, label_weights_list, bbox_targets_list,
+                bbox_weights_list, num_total_pos, num_total_neg)
+
+    def _get_target_single(self,
+                           cls_score,
+                           bbox_pred,
+                           gt_bboxes,
+                           gt_labels,
+                           img_meta,
+                           gt_bboxes_ignore=None):
+        """"Compute regression and classification targets for one image.
+
+        Outputs from a single decoder layer of a single feature level are used.
+
+        Args:
+            cls_score (Tensor): Box score logits from a single decoder layer
+                for one image. Shape [num_query, cls_out_channels].
+            bbox_pred (Tensor): Sigmoid outputs from a single decoder layer
+                for one image, with normalized coordinate (cx, cy, w, h) and
+                shape [num_query, 4].
+            gt_bboxes (Tensor): Ground truth bboxes for one image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (Tensor): Ground truth class indices for one image
+                with shape (num_gts, ).
+            img_meta (dict): Meta information for one image.
+            gt_bboxes_ignore (Tensor, optional): Bounding boxes
+                which can be ignored. Default None.
+
+        Returns:
+            tuple[Tensor]: a tuple containing the following for one image.
+
+                - labels (Tensor): Labels of each image.
+                - label_weights (Tensor]): Label weights of each image.
+                - bbox_targets (Tensor): BBox targets of each image.
+                - bbox_weights (Tensor): BBox weights of each image.
+                - pos_inds (Tensor): Sampled positive indices for each image.
+                - neg_inds (Tensor): Sampled negative indices for each image.
+        """
+
+        num_bboxes = bbox_pred.size(0)
+        # assigner and sampler
+        assign_result = self.assigner.assign(bbox_pred, cls_score, gt_bboxes,
+                                             gt_labels, img_meta,
+                                             gt_bboxes_ignore)
+        sampling_result = self.sampler.sample(assign_result, bbox_pred,
+                                              gt_bboxes)
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+
+        # label targets
+        labels = gt_bboxes.new_full((num_bboxes, ),
+                                    self.num_classes,
+                                    dtype=torch.long)
+        labels[pos_inds] = gt_labels[sampling_result.pos_assigned_gt_inds]
+        label_weights = gt_bboxes.new_ones(num_bboxes)
+
+        # bbox targets
+        bbox_targets = torch.zeros_like(bbox_pred)
+        bbox_weights = torch.zeros_like(bbox_pred)
+        bbox_weights[pos_inds] = 1.0
+        img_h, img_w, _ = img_meta['img_shape']
+
+        # DETR regress the relative position of boxes (cxcywh) in the image.
+        # Thus the learning target should be normalized by the image size, also
+        # the box format should be converted from defaultly x1y1x2y2 to cxcywh.
+        factor = bbox_pred.new_tensor([img_w, img_h, img_w,
+                                       img_h]).unsqueeze(0)
+        pos_gt_bboxes_normalized = sampling_result.pos_gt_bboxes / factor
+        pos_gt_bboxes_targets = bbox_xyxy_to_cxcywh(pos_gt_bboxes_normalized)
+        bbox_targets[pos_inds] = pos_gt_bboxes_targets
+        return (labels, label_weights, bbox_targets, bbox_weights, pos_inds,
+                neg_inds)
+
+    # over-write because img_metas are needed as inputs for bbox_head.
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels=None,
+                      gt_bboxes_ignore=None,
+                      proposal_cfg=None,
+                      **kwargs):
+        """Forward function for training mode.
+
+        Args:
+            x (list[Tensor]): Features from backbone.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes (Tensor): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_labels (Tensor): Ground truth labels of each box,
+                shape (num_gts,).
+            gt_bboxes_ignore (Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            proposal_cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert proposal_cfg is None, '"proposal_cfg" must be None'
+        outs = self(x, img_metas)
+        if gt_labels is None:
+            loss_inputs = outs + (gt_bboxes, img_metas)
+        else:
+            loss_inputs = outs + (gt_bboxes, gt_labels, img_metas)
+        losses = self.loss(*loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
+        return losses
+
+    @force_fp32(apply_to=('all_cls_scores_list', 'all_bbox_preds_list'))
+    def get_bboxes(self,
+                   all_cls_scores_list,
+                   all_bbox_preds_list,
+                   img_metas,
+                   rescale=False):
+        """Transform network outputs for a batch into bbox predictions.
+
+        Args:
+            all_cls_scores_list (list[Tensor]): Classification outputs
+                for each feature level. Each is a 4D-tensor with shape
+                [nb_dec, bs, num_query, cls_out_channels].
+            all_bbox_preds_list (list[Tensor]): Sigmoid regression
+                outputs for each feature level. Each is a 4D-tensor with
+                normalized coordinate format (cx, cy, w, h) and shape
+                [nb_dec, bs, num_query, 4].
+            img_metas (list[dict]): Meta information of each image.
+            rescale (bool, optional): If True, return boxes in original
+                image space. Defalut False.
+
+        Returns:
+            list[list[Tensor, Tensor]]: Each item in result_list is 2-tuple. \
+                The first item is an (n, 5) tensor, where the first 4 columns \
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the \
+                5-th column is a score between 0 and 1. The second item is a \
+                (n,) tensor where each item is the predicted class label of \
+                the corresponding box.
+        """
+        # NOTE defaultly only using outputs from the last feature level,
+        # and only the ouputs from the last decoder layer is used.
+        cls_scores = all_cls_scores_list[-1][-1]
+        bbox_preds = all_bbox_preds_list[-1][-1]
+
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score = cls_scores[img_id]
+            bbox_pred = bbox_preds[img_id]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self._get_bboxes_single(cls_score, bbox_pred,
+                                                img_shape, scale_factor,
+                                                rescale)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_score,
+                           bbox_pred,
+                           img_shape,
+                           scale_factor,
+                           rescale=False):
+        """Transform outputs from the last decoder layer into bbox predictions
+        for each image.
+
+        Args:
+            cls_score (Tensor): Box score logits from the last decoder layer
+                for each image. Shape [num_query, cls_out_channels].
+            bbox_pred (Tensor): Sigmoid outputs from the last decoder layer
+                for each image, with coordinate format (cx, cy, w, h) and
+                shape [num_query, 4].
+            img_shape (tuple[int]): Shape of input image, (height, width, 3).
+            scale_factor (ndarray, optional): Scale factor of the image arange
+                as (w_scale, h_scale, w_scale, h_scale).
+            rescale (bool, optional): If True, return boxes in original image
+                space. Default False.
+
+        Returns:
+            tuple[Tensor]: Results of detected bboxes and labels.
+
+                - det_bboxes: Predicted bboxes with shape [num_query, 5], \
+                    where the first 4 columns are bounding box positions \
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column are scores \
+                    between 0 and 1.
+                - det_labels: Predicted labels of the corresponding box with \
+                    shape [num_query].
+        """
+        assert len(cls_score) == len(bbox_pred)
+        # exclude background
+        scores, det_labels = F.softmax(cls_score, dim=-1)[..., :-1].max(-1)
+        det_bboxes = bbox_cxcywh_to_xyxy(bbox_pred)
+        det_bboxes[:, 0::2] = det_bboxes[:, 0::2] * img_shape[1]
+        det_bboxes[:, 1::2] = det_bboxes[:, 1::2] * img_shape[0]
+        det_bboxes[:, 0::2].clamp_(min=0, max=img_shape[1])
+        det_bboxes[:, 1::2].clamp_(min=0, max=img_shape[0])
+        if rescale:
+            det_bboxes /= det_bboxes.new_tensor(scale_factor)
+        det_bboxes = torch.cat((det_bboxes, scores.unsqueeze(1)), -1)
+        return det_bboxes, det_labels
diff --git a/detection/scrfd/mmdet/models/dense_heads/vfnet_head.py b/detection/scrfd/mmdet/models/dense_heads/vfnet_head.py
new file mode 100755
index 000000000..7243bb628
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/vfnet_head.py
@@ -0,0 +1,794 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, Scale, bias_init_with_prob, normal_init
+from mmcv.ops import DeformConv2d
+from mmcv.runner import force_fp32
+
+from mmdet.core import (bbox2distance, bbox_overlaps, build_anchor_generator,
+                        build_assigner, build_sampler, distance2bbox,
+                        multi_apply, multiclass_nms, reduce_mean)
+from ..builder import HEADS, build_loss
+from .atss_head import ATSSHead
+from .fcos_head import FCOSHead
+
+INF = 1e8
+
+
+@HEADS.register_module()
+class VFNetHead(ATSSHead, FCOSHead):
+    """Head of `VarifocalNet (VFNet): An IoU-aware Dense Object
+    Detector.<https://arxiv.org/abs/2008.13367>`_.
+
+    The VFNet predicts IoU-aware classification scores which mix the
+    object presence confidence and object localization accuracy as the
+    detection score. It is built on the FCOS architecture and uses ATSS
+    for defining positive/negative training examples. The VFNet is trained
+    with Varifocal Loss and empolys star-shaped deformable convolution to
+    extract features for a bbox.
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        regress_ranges (tuple[tuple[int, int]]): Regress range of multiple
+            level points.
+        center_sampling (bool): If true, use center sampling. Default: False.
+        center_sample_radius (float): Radius of center sampling. Default: 1.5.
+        sync_num_pos (bool): If true, synchronize the number of positive
+            examples across GPUs. Default: True
+        gradient_mul (float): The multiplier to gradients from bbox refinement
+            and recognition. Default: 0.1.
+        bbox_norm_type (str): The bbox normalization type, 'reg_denom' or
+            'stride'. Default: reg_denom
+        loss_cls_fl (dict): Config of focal loss.
+        use_vfl (bool): If true, use varifocal loss for training.
+            Default: True.
+        loss_cls (dict): Config of varifocal loss.
+        loss_bbox (dict): Config of localization loss, GIoU Loss.
+        loss_bbox (dict): Config of localization refinement loss, GIoU Loss.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: norm_cfg=dict(type='GN', num_groups=32,
+            requires_grad=True).
+        use_atss (bool): If true, use ATSS to define positive/negative
+            examples. Default: True.
+        anchor_generator (dict): Config of anchor generator for ATSS.
+
+    Example:
+        >>> self = VFNetHead(11, 7)
+        >>> feats = [torch.rand(1, 7, s, s) for s in [4, 8, 16, 32, 64]]
+        >>> cls_score, bbox_pred, bbox_pred_refine= self.forward(feats)
+        >>> assert len(cls_score) == len(self.scales)
+    """  # noqa: E501
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 regress_ranges=((-1, 64), (64, 128), (128, 256), (256, 512),
+                                 (512, INF)),
+                 center_sampling=False,
+                 center_sample_radius=1.5,
+                 sync_num_pos=True,
+                 gradient_mul=0.1,
+                 bbox_norm_type='reg_denom',
+                 loss_cls_fl=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 use_vfl=True,
+                 loss_cls=dict(
+                     type='VarifocalLoss',
+                     use_sigmoid=True,
+                     alpha=0.75,
+                     gamma=2.0,
+                     iou_weighted=True,
+                     loss_weight=1.0),
+                 loss_bbox=dict(type='GIoULoss', loss_weight=1.5),
+                 loss_bbox_refine=dict(type='GIoULoss', loss_weight=2.0),
+                 norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
+                 use_atss=True,
+                 anchor_generator=dict(
+                     type='AnchorGenerator',
+                     ratios=[1.0],
+                     octave_base_scale=8,
+                     scales_per_octave=1,
+                     center_offset=0.0,
+                     strides=[8, 16, 32, 64, 128]),
+                 **kwargs):
+        # dcn base offsets, adapted from reppoints_head.py
+        self.num_dconv_points = 9
+        self.dcn_kernel = int(np.sqrt(self.num_dconv_points))
+        self.dcn_pad = int((self.dcn_kernel - 1) / 2)
+        dcn_base = np.arange(-self.dcn_pad,
+                             self.dcn_pad + 1).astype(np.float64)
+        dcn_base_y = np.repeat(dcn_base, self.dcn_kernel)
+        dcn_base_x = np.tile(dcn_base, self.dcn_kernel)
+        dcn_base_offset = np.stack([dcn_base_y, dcn_base_x], axis=1).reshape(
+            (-1))
+        self.dcn_base_offset = torch.tensor(dcn_base_offset).view(1, -1, 1, 1)
+
+        super(FCOSHead, self).__init__(
+            num_classes, in_channels, norm_cfg=norm_cfg, **kwargs)
+        self.regress_ranges = regress_ranges
+        self.reg_denoms = [
+            regress_range[-1] for regress_range in regress_ranges
+        ]
+        self.reg_denoms[-1] = self.reg_denoms[-2] * 2
+        self.center_sampling = center_sampling
+        self.center_sample_radius = center_sample_radius
+        self.sync_num_pos = sync_num_pos
+        self.bbox_norm_type = bbox_norm_type
+        self.gradient_mul = gradient_mul
+        self.use_vfl = use_vfl
+        if self.use_vfl:
+            self.loss_cls = build_loss(loss_cls)
+        else:
+            self.loss_cls = build_loss(loss_cls_fl)
+        self.loss_bbox = build_loss(loss_bbox)
+        self.loss_bbox_refine = build_loss(loss_bbox_refine)
+
+        # for getting ATSS targets
+        self.use_atss = use_atss
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+        self.anchor_generator = build_anchor_generator(anchor_generator)
+        self.anchor_center_offset = anchor_generator['center_offset']
+        self.num_anchors = self.anchor_generator.num_base_anchors[0]
+        self.sampling = False
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        super(FCOSHead, self)._init_cls_convs()
+        super(FCOSHead, self)._init_reg_convs()
+        self.relu = nn.ReLU(inplace=True)
+        self.vfnet_reg_conv = ConvModule(
+            self.feat_channels,
+            self.feat_channels,
+            3,
+            stride=1,
+            padding=1,
+            conv_cfg=self.conv_cfg,
+            norm_cfg=self.norm_cfg,
+            bias=self.conv_bias)
+        self.vfnet_reg = nn.Conv2d(self.feat_channels, 4, 3, padding=1)
+        self.scales = nn.ModuleList([Scale(1.0) for _ in self.strides])
+
+        self.vfnet_reg_refine_dconv = DeformConv2d(
+            self.feat_channels,
+            self.feat_channels,
+            self.dcn_kernel,
+            1,
+            padding=self.dcn_pad)
+        self.vfnet_reg_refine = nn.Conv2d(self.feat_channels, 4, 3, padding=1)
+        self.scales_refine = nn.ModuleList([Scale(1.0) for _ in self.strides])
+
+        self.vfnet_cls_dconv = DeformConv2d(
+            self.feat_channels,
+            self.feat_channels,
+            self.dcn_kernel,
+            1,
+            padding=self.dcn_pad)
+        self.vfnet_cls = nn.Conv2d(
+            self.feat_channels, self.cls_out_channels, 3, padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.cls_convs:
+            if isinstance(m.conv, nn.Conv2d):
+                normal_init(m.conv, std=0.01)
+        for m in self.reg_convs:
+            if isinstance(m.conv, nn.Conv2d):
+                normal_init(m.conv, std=0.01)
+        normal_init(self.vfnet_reg_conv.conv, std=0.01)
+        normal_init(self.vfnet_reg, std=0.01)
+        normal_init(self.vfnet_reg_refine_dconv, std=0.01)
+        normal_init(self.vfnet_reg_refine, std=0.01)
+        normal_init(self.vfnet_cls_dconv, std=0.01)
+        bias_cls = bias_init_with_prob(0.01)
+        normal_init(self.vfnet_cls, std=0.01, bias=bias_cls)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple:
+                cls_scores (list[Tensor]): Box iou-aware scores for each scale
+                    level, each is a 4D-tensor, the channel number is
+                    num_points * num_classes.
+                bbox_preds (list[Tensor]): Box offsets for each
+                    scale level, each is a 4D-tensor, the channel number is
+                    num_points * 4.
+                bbox_preds_refine (list[Tensor]): Refined Box offsets for
+                    each scale level, each is a 4D-tensor, the channel
+                    number is num_points * 4.
+        """
+        return multi_apply(self.forward_single, feats, self.scales,
+                           self.scales_refine, self.strides, self.reg_denoms)
+
+    def forward_single(self, x, scale, scale_refine, stride, reg_denom):
+        """Forward features of a single scale level.
+
+        Args:
+            x (Tensor): FPN feature maps of the specified stride.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+            scale_refine (:obj: `mmcv.cnn.Scale`): Learnable scale module to
+                resize the refined bbox prediction.
+            stride (int): The corresponding stride for feature maps,
+                used to normalize the bbox prediction when
+                bbox_norm_type = 'stride'.
+            reg_denom (int): The corresponding regression range for feature
+                maps, only used to normalize the bbox prediction when
+                bbox_norm_type = 'reg_denom'.
+
+        Returns:
+            tuple: iou-aware cls scores for each box, bbox predictions and
+                refined bbox predictions of input feature maps.
+        """
+        cls_feat = x
+        reg_feat = x
+
+        for cls_layer in self.cls_convs:
+            cls_feat = cls_layer(cls_feat)
+
+        for reg_layer in self.reg_convs:
+            reg_feat = reg_layer(reg_feat)
+
+        # predict the bbox_pred of different level
+        reg_feat_init = self.vfnet_reg_conv(reg_feat)
+        if self.bbox_norm_type == 'reg_denom':
+            bbox_pred = scale(
+                self.vfnet_reg(reg_feat_init)).float().exp() * reg_denom
+        elif self.bbox_norm_type == 'stride':
+            bbox_pred = scale(
+                self.vfnet_reg(reg_feat_init)).float().exp() * stride
+        else:
+            raise NotImplementedError
+
+        # compute star deformable convolution offsets
+        # converting dcn_offset to reg_feat.dtype thus VFNet can be
+        # trained with FP16
+        dcn_offset = self.star_dcn_offset(bbox_pred, self.gradient_mul,
+                                          stride).to(reg_feat.dtype)
+
+        # refine the bbox_pred
+        reg_feat = self.relu(self.vfnet_reg_refine_dconv(reg_feat, dcn_offset))
+        bbox_pred_refine = scale_refine(
+            self.vfnet_reg_refine(reg_feat)).float().exp()
+        bbox_pred_refine = bbox_pred_refine * bbox_pred.detach()
+
+        # predict the iou-aware cls score
+        cls_feat = self.relu(self.vfnet_cls_dconv(cls_feat, dcn_offset))
+        cls_score = self.vfnet_cls(cls_feat)
+
+        return cls_score, bbox_pred, bbox_pred_refine
+
+    def star_dcn_offset(self, bbox_pred, gradient_mul, stride):
+        """Compute the star deformable conv offsets.
+
+        Args:
+            bbox_pred (Tensor): Predicted bbox distance offsets (l, r, t, b).
+            gradient_mul (float): Gradient multiplier.
+            stride (int): The corresponding stride for feature maps,
+                used to project the bbox onto the feature map.
+
+        Returns:
+            dcn_offsets (Tensor): The offsets for deformable convolution.
+        """
+        dcn_base_offset = self.dcn_base_offset.type_as(bbox_pred)
+        bbox_pred_grad_mul = (1 - gradient_mul) * bbox_pred.detach() + \
+            gradient_mul * bbox_pred
+        # map to the feature map scale
+        bbox_pred_grad_mul = bbox_pred_grad_mul / stride
+        N, C, H, W = bbox_pred.size()
+
+        x1 = bbox_pred_grad_mul[:, 0, :, :]
+        y1 = bbox_pred_grad_mul[:, 1, :, :]
+        x2 = bbox_pred_grad_mul[:, 2, :, :]
+        y2 = bbox_pred_grad_mul[:, 3, :, :]
+        bbox_pred_grad_mul_offset = bbox_pred.new_zeros(
+            N, 2 * self.num_dconv_points, H, W)
+        bbox_pred_grad_mul_offset[:, 0, :, :] = -1.0 * y1  # -y1
+        bbox_pred_grad_mul_offset[:, 1, :, :] = -1.0 * x1  # -x1
+        bbox_pred_grad_mul_offset[:, 2, :, :] = -1.0 * y1  # -y1
+        bbox_pred_grad_mul_offset[:, 4, :, :] = -1.0 * y1  # -y1
+        bbox_pred_grad_mul_offset[:, 5, :, :] = x2  # x2
+        bbox_pred_grad_mul_offset[:, 7, :, :] = -1.0 * x1  # -x1
+        bbox_pred_grad_mul_offset[:, 11, :, :] = x2  # x2
+        bbox_pred_grad_mul_offset[:, 12, :, :] = y2  # y2
+        bbox_pred_grad_mul_offset[:, 13, :, :] = -1.0 * x1  # -x1
+        bbox_pred_grad_mul_offset[:, 14, :, :] = y2  # y2
+        bbox_pred_grad_mul_offset[:, 16, :, :] = y2  # y2
+        bbox_pred_grad_mul_offset[:, 17, :, :] = x2  # x2
+        dcn_offset = bbox_pred_grad_mul_offset - dcn_base_offset
+
+        return dcn_offset
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'bbox_preds_refine'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             bbox_preds_refine,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute loss of the head.
+
+        Args:
+            cls_scores (list[Tensor]): Box iou-aware scores for each scale
+                level, each is a 4D-tensor, the channel number is
+                num_points * num_classes.
+            bbox_preds (list[Tensor]): Box offsets for each
+                scale level, each is a 4D-tensor, the channel number is
+                num_points * 4.
+            bbox_preds_refine (list[Tensor]): Refined Box offsets for
+                each scale level, each is a 4D-tensor, the channel
+                number is num_points * 4.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+                Default: None.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert len(cls_scores) == len(bbox_preds) == len(bbox_preds_refine)
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        all_level_points = self.get_points(featmap_sizes, bbox_preds[0].dtype,
+                                           bbox_preds[0].device)
+        labels, label_weights, bbox_targets, bbox_weights = self.get_targets(
+            cls_scores, all_level_points, gt_bboxes, gt_labels, img_metas,
+            gt_bboxes_ignore)
+
+        num_imgs = cls_scores[0].size(0)
+        # flatten cls_scores, bbox_preds and bbox_preds_refine
+        flatten_cls_scores = [
+            cls_score.permute(0, 2, 3,
+                              1).reshape(-1,
+                                         self.cls_out_channels).contiguous()
+            for cls_score in cls_scores
+        ]
+        flatten_bbox_preds = [
+            bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4).contiguous()
+            for bbox_pred in bbox_preds
+        ]
+        flatten_bbox_preds_refine = [
+            bbox_pred_refine.permute(0, 2, 3, 1).reshape(-1, 4).contiguous()
+            for bbox_pred_refine in bbox_preds_refine
+        ]
+        flatten_cls_scores = torch.cat(flatten_cls_scores)
+        flatten_bbox_preds = torch.cat(flatten_bbox_preds)
+        flatten_bbox_preds_refine = torch.cat(flatten_bbox_preds_refine)
+        flatten_labels = torch.cat(labels)
+        flatten_bbox_targets = torch.cat(bbox_targets)
+        # repeat points to align with bbox_preds
+        flatten_points = torch.cat(
+            [points.repeat(num_imgs, 1) for points in all_level_points])
+
+        # FG cat_id: [0, num_classes - 1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = torch.where(
+            ((flatten_labels >= 0) & (flatten_labels < bg_class_ind)) > 0)[0]
+        num_pos = len(pos_inds)
+
+        pos_bbox_preds = flatten_bbox_preds[pos_inds]
+        pos_bbox_preds_refine = flatten_bbox_preds_refine[pos_inds]
+        pos_labels = flatten_labels[pos_inds]
+
+        # sync num_pos across all gpus
+        if self.sync_num_pos:
+            num_pos_avg_per_gpu = reduce_mean(
+                pos_inds.new_tensor(num_pos).float()).item()
+            num_pos_avg_per_gpu = max(num_pos_avg_per_gpu, 1.0)
+        else:
+            num_pos_avg_per_gpu = num_pos
+
+        if num_pos > 0:
+            pos_bbox_targets = flatten_bbox_targets[pos_inds]
+            pos_points = flatten_points[pos_inds]
+
+            pos_decoded_bbox_preds = distance2bbox(pos_points, pos_bbox_preds)
+            pos_decoded_target_preds = distance2bbox(pos_points,
+                                                     pos_bbox_targets)
+            iou_targets_ini = bbox_overlaps(
+                pos_decoded_bbox_preds,
+                pos_decoded_target_preds.detach(),
+                is_aligned=True).clamp(min=1e-6)
+            bbox_weights_ini = iou_targets_ini.clone().detach()
+            iou_targets_ini_avg_per_gpu = reduce_mean(
+                bbox_weights_ini.sum()).item()
+            bbox_avg_factor_ini = max(iou_targets_ini_avg_per_gpu, 1.0)
+            loss_bbox = self.loss_bbox(
+                pos_decoded_bbox_preds,
+                pos_decoded_target_preds.detach(),
+                weight=bbox_weights_ini,
+                avg_factor=bbox_avg_factor_ini)
+
+            pos_decoded_bbox_preds_refine = \
+                distance2bbox(pos_points, pos_bbox_preds_refine)
+            iou_targets_rf = bbox_overlaps(
+                pos_decoded_bbox_preds_refine,
+                pos_decoded_target_preds.detach(),
+                is_aligned=True).clamp(min=1e-6)
+            bbox_weights_rf = iou_targets_rf.clone().detach()
+            iou_targets_rf_avg_per_gpu = reduce_mean(
+                bbox_weights_rf.sum()).item()
+            bbox_avg_factor_rf = max(iou_targets_rf_avg_per_gpu, 1.0)
+            loss_bbox_refine = self.loss_bbox_refine(
+                pos_decoded_bbox_preds_refine,
+                pos_decoded_target_preds.detach(),
+                weight=bbox_weights_rf,
+                avg_factor=bbox_avg_factor_rf)
+
+            # build IoU-aware cls_score targets
+            if self.use_vfl:
+                pos_ious = iou_targets_rf.clone().detach()
+                cls_iou_targets = torch.zeros_like(flatten_cls_scores)
+                cls_iou_targets[pos_inds, pos_labels] = pos_ious
+        else:
+            loss_bbox = pos_bbox_preds.sum() * 0
+            loss_bbox_refine = pos_bbox_preds_refine.sum() * 0
+            if self.use_vfl:
+                cls_iou_targets = torch.zeros_like(flatten_cls_scores)
+
+        if self.use_vfl:
+            loss_cls = self.loss_cls(
+                flatten_cls_scores,
+                cls_iou_targets,
+                avg_factor=num_pos_avg_per_gpu)
+        else:
+            loss_cls = self.loss_cls(
+                flatten_cls_scores,
+                flatten_labels,
+                weight=label_weights,
+                avg_factor=num_pos_avg_per_gpu)
+
+        return dict(
+            loss_cls=loss_cls,
+            loss_bbox=loss_bbox,
+            loss_bbox_rf=loss_bbox_refine)
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'bbox_preds_refine'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   bbox_preds_refine,
+                   img_metas,
+                   cfg=None,
+                   rescale=None,
+                   with_nms=True):
+        """Transform network outputs for a batch into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box iou-aware scores for each scale
+                level with shape (N, num_points * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box offsets for each scale
+                level with shape (N, num_points * 4, H, W).
+            bbox_preds_refine (list[Tensor]): Refined Box offsets for
+                each scale level with shape (N, num_points * 4, H, W).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used. Default: None.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before returning boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class label of
+                the corresponding box.
+        """
+        assert len(cls_scores) == len(bbox_preds) == len(bbox_preds_refine)
+        num_levels = len(cls_scores)
+
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        mlvl_points = self.get_points(featmap_sizes, bbox_preds[0].dtype,
+                                      bbox_preds[0].device)
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds_refine[i][img_id].detach()
+                for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            det_bboxes = self._get_bboxes_single(cls_score_list,
+                                                 bbox_pred_list, mlvl_points,
+                                                 img_shape, scale_factor, cfg,
+                                                 rescale, with_nms)
+            result_list.append(det_bboxes)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           cls_scores,
+                           bbox_preds,
+                           mlvl_points,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            cls_scores (list[Tensor]): Box iou-aware scores for a single scale
+                level with shape (num_points * num_classes, H, W).
+            bbox_preds (list[Tensor]): Box offsets for a single scale
+                level with shape (num_points * 4, H, W).
+            mlvl_points (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_points, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arrange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before returning boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): BBox predictions in shape (n, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (n,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_points)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, points in zip(cls_scores, bbox_preds,
+                                                mlvl_points):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            scores = cls_score.permute(1, 2, 0).reshape(
+                -1, self.cls_out_channels).contiguous().sigmoid()
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4).contiguous()
+
+            nms_pre = cfg.get('nms_pre', -1)
+            if 0 < nms_pre < scores.shape[0]:
+                max_scores, _ = scores.max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                points = points[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+            bboxes = distance2bbox(points, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+        # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+        # BG cat_id: num_class
+        mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        if with_nms:
+            det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+            return det_bboxes, det_labels
+        else:
+            return mlvl_bboxes, mlvl_scores
+
+    def _get_points_single(self,
+                           featmap_size,
+                           stride,
+                           dtype,
+                           device,
+                           flatten=False):
+        """Get points according to feature map sizes."""
+        h, w = featmap_size
+        x_range = torch.arange(
+            0, w * stride, stride, dtype=dtype, device=device)
+        y_range = torch.arange(
+            0, h * stride, stride, dtype=dtype, device=device)
+        y, x = torch.meshgrid(y_range, x_range)
+        # to be compatible with anchor points in ATSS
+        if self.use_atss:
+            points = torch.stack(
+                (x.reshape(-1), y.reshape(-1)), dim=-1) + \
+                     stride * self.anchor_center_offset
+        else:
+            points = torch.stack(
+                (x.reshape(-1), y.reshape(-1)), dim=-1) + stride // 2
+        return points
+
+    def get_targets(self, cls_scores, mlvl_points, gt_bboxes, gt_labels,
+                    img_metas, gt_bboxes_ignore):
+        """A wrapper for computing ATSS and FCOS targets for points in multiple
+        images.
+
+        Args:
+            cls_scores (list[Tensor]): Box iou-aware scores for each scale
+                level with shape (N, num_points * num_classes, H, W).
+            mlvl_points (list[Tensor]): Points of each fpn level, each has
+                shape (num_points, 2).
+            gt_bboxes (list[Tensor]): Ground truth bboxes of each image,
+                each has shape (num_gt, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each box,
+                each has shape (num_gt,).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+
+        Returns:
+            tuple:
+                labels_list (list[Tensor]): Labels of each level.
+                label_weights (Tensor/None): Label weights of all levels.
+                bbox_targets_list (list[Tensor]): Regression targets of each
+                    level, (l, t, r, b).
+                bbox_weights (Tensor/None): Bbox weights of all levels.
+        """
+        if self.use_atss:
+            return self.get_atss_targets(cls_scores, mlvl_points, gt_bboxes,
+                                         gt_labels, img_metas,
+                                         gt_bboxes_ignore)
+        else:
+            self.norm_on_bbox = False
+            return self.get_fcos_targets(mlvl_points, gt_bboxes, gt_labels)
+
+    def _get_target_single(self, *args, **kwargs):
+        """Avoid ambiguity in multiple inheritance."""
+        if self.use_atss:
+            return ATSSHead._get_target_single(self, *args, **kwargs)
+        else:
+            return FCOSHead._get_target_single(self, *args, **kwargs)
+
+    def get_fcos_targets(self, points, gt_bboxes_list, gt_labels_list):
+        """Compute FCOS regression and classification targets for points in
+        multiple images.
+
+        Args:
+            points (list[Tensor]): Points of each fpn level, each has shape
+                (num_points, 2).
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image,
+                each has shape (num_gt, 4).
+            gt_labels_list (list[Tensor]): Ground truth labels of each box,
+                each has shape (num_gt,).
+
+        Returns:
+            tuple:
+                labels (list[Tensor]): Labels of each level.
+                label_weights: None, to be compatible with ATSS targets.
+                bbox_targets (list[Tensor]): BBox targets of each level.
+                bbox_weights: None, to be compatible with ATSS targets.
+        """
+        labels, bbox_targets = FCOSHead.get_targets(self, points,
+                                                    gt_bboxes_list,
+                                                    gt_labels_list)
+        label_weights = None
+        bbox_weights = None
+        return labels, label_weights, bbox_targets, bbox_weights
+
+    def get_atss_targets(self,
+                         cls_scores,
+                         mlvl_points,
+                         gt_bboxes,
+                         gt_labels,
+                         img_metas,
+                         gt_bboxes_ignore=None):
+        """A wrapper for computing ATSS targets for points in multiple images.
+
+        Args:
+            cls_scores (list[Tensor]): Box iou-aware scores for each scale
+                level with shape (N, num_points * num_classes, H, W).
+            mlvl_points (list[Tensor]): Points of each fpn level, each has
+                shape (num_points, 2).
+            gt_bboxes (list[Tensor]): Ground truth bboxes of each image,
+                each has shape (num_gt, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each box,
+                each has shape (num_gt,).
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | Tensor): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4). Default: None.
+
+        Returns:
+            tuple:
+                labels_list (list[Tensor]): Labels of each level.
+                label_weights (Tensor): Label weights of all levels.
+                bbox_targets_list (list[Tensor]): Regression targets of each
+                    level, (l, t, r, b).
+                bbox_weights (Tensor): Bbox weights of all levels.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+
+        cls_reg_targets = ATSSHead.get_targets(
+            self,
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels,
+            unmap_outputs=True)
+        if cls_reg_targets is None:
+            return None
+
+        (anchor_list, labels_list, label_weights_list, bbox_targets_list,
+         bbox_weights_list, num_total_pos, num_total_neg) = cls_reg_targets
+
+        bbox_targets_list = [
+            bbox_targets.reshape(-1, 4) for bbox_targets in bbox_targets_list
+        ]
+
+        num_imgs = len(img_metas)
+        # transform bbox_targets (x1, y1, x2, y2) into (l, t, r, b) format
+        bbox_targets_list = self.transform_bbox_targets(
+            bbox_targets_list, mlvl_points, num_imgs)
+
+        labels_list = [labels.reshape(-1) for labels in labels_list]
+        label_weights_list = [
+            label_weights.reshape(-1) for label_weights in label_weights_list
+        ]
+        bbox_weights_list = [
+            bbox_weights.reshape(-1) for bbox_weights in bbox_weights_list
+        ]
+        label_weights = torch.cat(label_weights_list)
+        bbox_weights = torch.cat(bbox_weights_list)
+        return labels_list, label_weights, bbox_targets_list, bbox_weights
+
+    def transform_bbox_targets(self, decoded_bboxes, mlvl_points, num_imgs):
+        """Transform bbox_targets (x1, y1, x2, y2) into (l, t, r, b) format.
+
+        Args:
+            decoded_bboxes (list[Tensor]): Regression targets of each level,
+                in the form of (x1, y1, x2, y2).
+            mlvl_points (list[Tensor]): Points of each fpn level, each has
+                shape (num_points, 2).
+            num_imgs (int): the number of images in a batch.
+
+        Returns:
+            bbox_targets (list[Tensor]): Regression targets of each level in
+                the form of (l, t, r, b).
+        """
+        # TODO: Re-implemented in Class PointCoder
+        assert len(decoded_bboxes) == len(mlvl_points)
+        num_levels = len(decoded_bboxes)
+        mlvl_points = [points.repeat(num_imgs, 1) for points in mlvl_points]
+        bbox_targets = []
+        for i in range(num_levels):
+            bbox_target = bbox2distance(mlvl_points[i], decoded_bboxes[i])
+            bbox_targets.append(bbox_target)
+
+        return bbox_targets
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        """Override the method in the parent class to avoid changing para's
+        name."""
+        pass
diff --git a/detection/scrfd/mmdet/models/dense_heads/yolact_head.py b/detection/scrfd/mmdet/models/dense_heads/yolact_head.py
new file mode 100755
index 000000000..824246ce7
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/yolact_head.py
@@ -0,0 +1,939 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, xavier_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import build_sampler, fast_nms, images_to_levels, multi_apply
+from ..builder import HEADS, build_loss
+from .anchor_head import AnchorHead
+
+
+@HEADS.register_module()
+class YOLACTHead(AnchorHead):
+    """YOLACT box head used in https://arxiv.org/abs/1904.02689.
+
+    Note that YOLACT head is a light version of RetinaNet head.
+    Four differences are described as follows:
+
+    1. YOLACT box head has three-times fewer anchors.
+    2. YOLACT box head shares the convs for box and cls branches.
+    3. YOLACT box head uses OHEM instead of Focal loss.
+    4. YOLACT box head predicts a set of mask coefficients for each box.
+
+    Args:
+        num_classes (int): Number of categories excluding the background
+            category.
+        in_channels (int): Number of channels in the input feature map.
+        anchor_generator (dict): Config dict for anchor generator
+        loss_cls (dict): Config of classification loss.
+        loss_bbox (dict): Config of localization loss.
+        num_head_convs (int): Number of the conv layers shared by
+            box and cls branches.
+        num_protos (int): Number of the mask coefficients.
+        use_ohem (bool): If true, ``loss_single_OHEM`` will be used for
+            cls loss calculation. If false, ``loss_single`` will be used.
+        conv_cfg (dict): Dictionary to construct and config conv layer.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 anchor_generator=dict(
+                     type='AnchorGenerator',
+                     octave_base_scale=3,
+                     scales_per_octave=1,
+                     ratios=[0.5, 1.0, 2.0],
+                     strides=[8, 16, 32, 64, 128]),
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=False,
+                     reduction='none',
+                     loss_weight=1.0),
+                 loss_bbox=dict(
+                     type='SmoothL1Loss', beta=1.0, loss_weight=1.5),
+                 num_head_convs=1,
+                 num_protos=32,
+                 use_ohem=True,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 **kwargs):
+        self.num_head_convs = num_head_convs
+        self.num_protos = num_protos
+        self.use_ohem = use_ohem
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        super(YOLACTHead, self).__init__(
+            num_classes,
+            in_channels,
+            loss_cls=loss_cls,
+            loss_bbox=loss_bbox,
+            anchor_generator=anchor_generator,
+            **kwargs)
+        if self.use_ohem:
+            sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+            self.sampling = False
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.relu = nn.ReLU(inplace=True)
+        self.head_convs = nn.ModuleList()
+        for i in range(self.num_head_convs):
+            chn = self.in_channels if i == 0 else self.feat_channels
+            self.head_convs.append(
+                ConvModule(
+                    chn,
+                    self.feat_channels,
+                    3,
+                    stride=1,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        self.conv_cls = nn.Conv2d(
+            self.feat_channels,
+            self.num_anchors * self.cls_out_channels,
+            3,
+            padding=1)
+        self.conv_reg = nn.Conv2d(
+            self.feat_channels, self.num_anchors * 4, 3, padding=1)
+        self.conv_coeff = nn.Conv2d(
+            self.feat_channels,
+            self.num_anchors * self.num_protos,
+            3,
+            padding=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.head_convs:
+            xavier_init(m.conv, distribution='uniform', bias=0)
+        xavier_init(self.conv_cls, distribution='uniform', bias=0)
+        xavier_init(self.conv_reg, distribution='uniform', bias=0)
+        xavier_init(self.conv_coeff, distribution='uniform', bias=0)
+
+    def forward_single(self, x):
+        """Forward feature of a single scale level.
+
+        Args:
+            x (Tensor): Features of a single scale level.
+
+        Returns:
+            tuple:
+                cls_score (Tensor): Cls scores for a single scale level \
+                    the channels number is num_anchors * num_classes.
+                bbox_pred (Tensor): Box energies / deltas for a single scale \
+                    level, the channels number is num_anchors * 4.
+                coeff_pred (Tensor): Mask coefficients for a single scale \
+                    level, the channels number is num_anchors * num_protos.
+        """
+        for head_conv in self.head_convs:
+            x = head_conv(x)
+        cls_score = self.conv_cls(x)
+        bbox_pred = self.conv_reg(x)
+        coeff_pred = self.conv_coeff(x).tanh()
+        return cls_score, bbox_pred, coeff_pred
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
+    def loss(self,
+             cls_scores,
+             bbox_preds,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """A combination of the func:``AnchorHead.loss`` and
+        func:``SSDHead.loss``.
+
+        When ``self.use_ohem == True``, it functions like ``SSDHead.loss``,
+        otherwise, it follows ``AnchorHead.loss``. Besides, it additionally
+        returns ``sampling_results``.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                Has shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
+                boxes can be ignored when computing the loss. Default: None
+
+        Returns:
+            tuple:
+                dict[str, Tensor]: A dictionary of loss components.
+                List[:obj:``SamplingResult``]: Sampler results for each image.
+        """
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == self.anchor_generator.num_levels
+
+        device = cls_scores[0].device
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas, device=device)
+        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
+        cls_reg_targets = self.get_targets(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            gt_bboxes_ignore_list=gt_bboxes_ignore,
+            gt_labels_list=gt_labels,
+            label_channels=label_channels,
+            unmap_outputs=not self.use_ohem,
+            return_sampling_results=True)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg, sampling_results) = cls_reg_targets
+
+        if self.use_ohem:
+            num_images = len(img_metas)
+            all_cls_scores = torch.cat([
+                s.permute(0, 2, 3, 1).reshape(
+                    num_images, -1, self.cls_out_channels) for s in cls_scores
+            ], 1)
+            all_labels = torch.cat(labels_list, -1).view(num_images, -1)
+            all_label_weights = torch.cat(label_weights_list,
+                                          -1).view(num_images, -1)
+            all_bbox_preds = torch.cat([
+                b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
+                for b in bbox_preds
+            ], -2)
+            all_bbox_targets = torch.cat(bbox_targets_list,
+                                         -2).view(num_images, -1, 4)
+            all_bbox_weights = torch.cat(bbox_weights_list,
+                                         -2).view(num_images, -1, 4)
+
+            # concat all level anchors to a single tensor
+            all_anchors = []
+            for i in range(num_images):
+                all_anchors.append(torch.cat(anchor_list[i]))
+
+            # check NaN and Inf
+            assert torch.isfinite(all_cls_scores).all().item(), \
+                'classification scores become infinite or NaN!'
+            assert torch.isfinite(all_bbox_preds).all().item(), \
+                'bbox predications become infinite or NaN!'
+
+            losses_cls, losses_bbox = multi_apply(
+                self.loss_single_OHEM,
+                all_cls_scores,
+                all_bbox_preds,
+                all_anchors,
+                all_labels,
+                all_label_weights,
+                all_bbox_targets,
+                all_bbox_weights,
+                num_total_samples=num_total_pos)
+        else:
+            num_total_samples = (
+                num_total_pos +
+                num_total_neg if self.sampling else num_total_pos)
+
+            # anchor number of multi levels
+            num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+            # concat all level anchors and flags to a single tensor
+            concat_anchor_list = []
+            for i in range(len(anchor_list)):
+                concat_anchor_list.append(torch.cat(anchor_list[i]))
+            all_anchor_list = images_to_levels(concat_anchor_list,
+                                               num_level_anchors)
+            losses_cls, losses_bbox = multi_apply(
+                self.loss_single,
+                cls_scores,
+                bbox_preds,
+                all_anchor_list,
+                labels_list,
+                label_weights_list,
+                bbox_targets_list,
+                bbox_weights_list,
+                num_total_samples=num_total_samples)
+
+        return dict(
+            loss_cls=losses_cls, loss_bbox=losses_bbox), sampling_results
+
+    def loss_single_OHEM(self, cls_score, bbox_pred, anchors, labels,
+                         label_weights, bbox_targets, bbox_weights,
+                         num_total_samples):
+        """"See func:``SSDHead.loss``."""
+        loss_cls_all = self.loss_cls(cls_score, labels, label_weights)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        pos_inds = ((labels >= 0) &
+                    (labels < self.num_classes)).nonzero().reshape(-1)
+        neg_inds = (labels == self.num_classes).nonzero().view(-1)
+
+        num_pos_samples = pos_inds.size(0)
+        if num_pos_samples == 0:
+            num_neg_samples = neg_inds.size(0)
+        else:
+            num_neg_samples = self.train_cfg.neg_pos_ratio * num_pos_samples
+            if num_neg_samples > neg_inds.size(0):
+                num_neg_samples = neg_inds.size(0)
+        topk_loss_cls_neg, _ = loss_cls_all[neg_inds].topk(num_neg_samples)
+        loss_cls_pos = loss_cls_all[pos_inds].sum()
+        loss_cls_neg = topk_loss_cls_neg.sum()
+        loss_cls = (loss_cls_pos + loss_cls_neg) / num_total_samples
+        if self.reg_decoded_bbox:
+            bbox_pred = self.bbox_coder.decode(anchors, bbox_pred)
+        loss_bbox = self.loss_bbox(
+            bbox_pred,
+            bbox_targets,
+            bbox_weights,
+            avg_factor=num_total_samples)
+        return loss_cls[None], loss_bbox
+
+    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'coeff_preds'))
+    def get_bboxes(self,
+                   cls_scores,
+                   bbox_preds,
+                   coeff_preds,
+                   img_metas,
+                   cfg=None,
+                   rescale=False):
+        """"Similiar to func:``AnchorHead.get_bboxes``, but additionally
+        processes coeff_preds.
+
+        Args:
+            cls_scores (list[Tensor]): Box scores for each scale level
+                with shape (N, num_anchors * num_classes, H, W)
+            bbox_preds (list[Tensor]): Box energies / deltas for each scale
+                level with shape (N, num_anchors * 4, H, W)
+            coeff_preds (list[Tensor]): Mask coefficients for each scale
+                level with shape (N, num_anchors * num_protos, H, W)
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+
+        Returns:
+            list[tuple[Tensor, Tensor, Tensor]]: Each item in result_list is
+                a 3-tuple. The first item is an (n, 5) tensor, where the
+                first 4 columns are bounding box positions
+                (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                between 0 and 1. The second item is an (n,) tensor where each
+                item is the predicted class label of the corresponding box.
+                The third item is an (n, num_protos) tensor where each item
+                is the predicted mask coefficients of instance inside the
+                corresponding box.
+        """
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+
+        device = cls_scores[0].device
+        featmap_sizes = [cls_scores[i].shape[-2:] for i in range(num_levels)]
+        mlvl_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device=device)
+
+        det_bboxes = []
+        det_labels = []
+        det_coeffs = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            coeff_pred_list = [
+                coeff_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            bbox_res = self._get_bboxes_single(cls_score_list, bbox_pred_list,
+                                               coeff_pred_list, mlvl_anchors,
+                                               img_shape, scale_factor, cfg,
+                                               rescale)
+            det_bboxes.append(bbox_res[0])
+            det_labels.append(bbox_res[1])
+            det_coeffs.append(bbox_res[2])
+        return det_bboxes, det_labels, det_coeffs
+
+    def _get_bboxes_single(self,
+                           cls_score_list,
+                           bbox_pred_list,
+                           coeff_preds_list,
+                           mlvl_anchors,
+                           img_shape,
+                           scale_factor,
+                           cfg,
+                           rescale=False):
+        """"Similiar to func:``AnchorHead._get_bboxes_single``, but
+        additionally processes coeff_preds_list and uses fast NMS instead of
+        traditional NMS.
+
+        Args:
+            cls_score_list (list[Tensor]): Box scores for a single scale level
+                Has shape (num_anchors * num_classes, H, W).
+            bbox_pred_list (list[Tensor]): Box energies / deltas for a single
+                scale level with shape (num_anchors * 4, H, W).
+            coeff_preds_list (list[Tensor]): Mask coefficients for a single
+                scale level with shape (num_anchors * num_protos, H, W).
+            mlvl_anchors (list[Tensor]): Box reference for a single scale level
+                with shape (num_total_anchors, 4).
+            img_shape (tuple[int]): Shape of the input image,
+                (height, width, 3).
+            scale_factor (ndarray): Scale factor of the image arange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+
+        Returns:
+            tuple[Tensor, Tensor, Tensor]: The first item is an (n, 5) tensor,
+                where the first 4 columns are bounding box positions
+                (tl_x, tl_y, br_x, br_y) and the 5-th column is a score between
+                0 and 1. The second item is an (n,) tensor where each item is
+                the predicted class label of the corresponding box. The third
+                item is an (n, num_protos) tensor where each item is the
+                predicted mask coefficients of instance inside the
+                corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(cls_score_list) == len(bbox_pred_list) == len(mlvl_anchors)
+        mlvl_bboxes = []
+        mlvl_scores = []
+        mlvl_coeffs = []
+        for cls_score, bbox_pred, coeff_pred, anchors in \
+                zip(cls_score_list, bbox_pred_list,
+                    coeff_preds_list, mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            cls_score = cls_score.permute(1, 2,
+                                          0).reshape(-1, self.cls_out_channels)
+            if self.use_sigmoid_cls:
+                scores = cls_score.sigmoid()
+            else:
+                scores = cls_score.softmax(-1)
+            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
+            coeff_pred = coeff_pred.permute(1, 2,
+                                            0).reshape(-1, self.num_protos)
+            nms_pre = cfg.get('nms_pre', -1)
+            if nms_pre > 0 and scores.shape[0] > nms_pre:
+                # Get maximum scores for foreground classes.
+                if self.use_sigmoid_cls:
+                    max_scores, _ = scores.max(dim=1)
+                else:
+                    # remind that we set FG labels to [0, num_class-1]
+                    # since mmdet v2.0
+                    # BG cat_id: num_class
+                    max_scores, _ = scores[:, :-1].max(dim=1)
+                _, topk_inds = max_scores.topk(nms_pre)
+                anchors = anchors[topk_inds, :]
+                bbox_pred = bbox_pred[topk_inds, :]
+                scores = scores[topk_inds, :]
+                coeff_pred = coeff_pred[topk_inds, :]
+            bboxes = self.bbox_coder.decode(
+                anchors, bbox_pred, max_shape=img_shape)
+            mlvl_bboxes.append(bboxes)
+            mlvl_scores.append(scores)
+            mlvl_coeffs.append(coeff_pred)
+        mlvl_bboxes = torch.cat(mlvl_bboxes)
+        if rescale:
+            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        mlvl_coeffs = torch.cat(mlvl_coeffs)
+        if self.use_sigmoid_cls:
+            # Add a dummy background class to the backend when using sigmoid
+            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
+            # BG cat_id: num_class
+            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
+            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
+        det_bboxes, det_labels, det_coeffs = fast_nms(mlvl_bboxes, mlvl_scores,
+                                                      mlvl_coeffs,
+                                                      cfg.score_thr,
+                                                      cfg.iou_thr, cfg.top_k,
+                                                      cfg.max_per_img)
+        return det_bboxes, det_labels, det_coeffs
+
+
+@HEADS.register_module()
+class YOLACTSegmHead(nn.Module):
+    """YOLACT segmentation head used in https://arxiv.org/abs/1904.02689.
+
+    Apply a semantic segmentation loss on feature space using layers that are
+    only evaluated during training to increase performance with no speed
+    penalty.
+
+    Args:
+        in_channels (int): Number of channels in the input feature map.
+        num_classes (int): Number of categories excluding the background
+            category.
+        loss_segm (dict): Config of semantic segmentation loss.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels=256,
+                 loss_segm=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0)):
+        super(YOLACTSegmHead, self).__init__()
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.loss_segm = build_loss(loss_segm)
+        self._init_layers()
+        self.fp16_enabled = False
+
+    def _init_layers(self):
+        """Initialize layers of the head."""
+        self.segm_conv = nn.Conv2d(
+            self.in_channels, self.num_classes, kernel_size=1)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        xavier_init(self.segm_conv, distribution='uniform')
+
+    def forward(self, x):
+        """Forward feature from the upstream network.
+
+        Args:
+            x (Tensor): Feature from the upstream network, which is
+                a 4D-tensor.
+
+        Returns:
+            Tensor: Predicted semantic segmentation map with shape
+                (N, num_classes, H, W).
+        """
+        return self.segm_conv(x)
+
+    @force_fp32(apply_to=('segm_pred', ))
+    def loss(self, segm_pred, gt_masks, gt_labels):
+        """Compute loss of the head.
+
+        Args:
+            segm_pred (list[Tensor]): Predicted semantic segmentation map
+                with shape (N, num_classes, H, W).
+            gt_masks (list[Tensor]): Ground truth masks for each image with
+                the same shape of the input image.
+            gt_labels (list[Tensor]): Class indices corresponding to each box.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        loss_segm = []
+        num_imgs, num_classes, mask_h, mask_w = segm_pred.size()
+        for idx in range(num_imgs):
+            cur_segm_pred = segm_pred[idx]
+            cur_gt_masks = gt_masks[idx].float()
+            cur_gt_labels = gt_labels[idx]
+            segm_targets = self.get_targets(cur_segm_pred, cur_gt_masks,
+                                            cur_gt_labels)
+            if segm_targets is None:
+                loss = self.loss_segm(cur_segm_pred,
+                                      torch.zeros_like(cur_segm_pred),
+                                      torch.zeros_like(cur_segm_pred))
+            else:
+                loss = self.loss_segm(
+                    cur_segm_pred,
+                    segm_targets,
+                    avg_factor=num_imgs * mask_h * mask_w)
+            loss_segm.append(loss)
+        return dict(loss_segm=loss_segm)
+
+    def get_targets(self, segm_pred, gt_masks, gt_labels):
+        """Compute semantic segmentation targets for each image.
+
+        Args:
+            segm_pred (Tensor): Predicted semantic segmentation map
+                with shape (num_classes, H, W).
+            gt_masks (Tensor): Ground truth masks for each image with
+                the same shape of the input image.
+            gt_labels (Tensor): Class indices corresponding to each box.
+
+        Returns:
+            Tensor: Semantic segmentation targets with shape
+                (num_classes, H, W).
+        """
+        if gt_masks.size(0) == 0:
+            return None
+        num_classes, mask_h, mask_w = segm_pred.size()
+        with torch.no_grad():
+            downsampled_masks = F.interpolate(
+                gt_masks.unsqueeze(0), (mask_h, mask_w),
+                mode='bilinear',
+                align_corners=False).squeeze(0)
+            downsampled_masks = downsampled_masks.gt(0.5).float()
+            segm_targets = torch.zeros_like(segm_pred, requires_grad=False)
+            for obj_idx in range(downsampled_masks.size(0)):
+                segm_targets[gt_labels[obj_idx] - 1] = torch.max(
+                    segm_targets[gt_labels[obj_idx] - 1],
+                    downsampled_masks[obj_idx])
+            return segm_targets
+
+
+@HEADS.register_module()
+class YOLACTProtonet(nn.Module):
+    """YOLACT mask head used in https://arxiv.org/abs/1904.02689.
+
+    This head outputs the mask prototypes for YOLACT.
+
+    Args:
+        in_channels (int): Number of channels in the input feature map.
+        proto_channels (tuple[int]): Output channels of protonet convs.
+        proto_kernel_sizes (tuple[int]): Kernel sizes of protonet convs.
+        include_last_relu (Bool): If keep the last relu of protonet.
+        num_protos (int): Number of prototypes.
+        num_classes (int): Number of categories excluding the background
+            category.
+        loss_mask_weight (float): Reweight the mask loss by this factor.
+        max_masks_to_train (int): Maximum number of masks to train for
+            each image.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels=256,
+                 proto_channels=(256, 256, 256, None, 256, 32),
+                 proto_kernel_sizes=(3, 3, 3, -2, 3, 1),
+                 include_last_relu=True,
+                 num_protos=32,
+                 loss_mask_weight=1.0,
+                 max_masks_to_train=100):
+        super(YOLACTProtonet, self).__init__()
+        self.in_channels = in_channels
+        self.proto_channels = proto_channels
+        self.proto_kernel_sizes = proto_kernel_sizes
+        self.include_last_relu = include_last_relu
+        self.protonet = self._init_layers()
+
+        self.loss_mask_weight = loss_mask_weight
+        self.num_protos = num_protos
+        self.num_classes = num_classes
+        self.max_masks_to_train = max_masks_to_train
+        self.fp16_enabled = False
+
+    def _init_layers(self):
+        """A helper function to take a config setting and turn it into a
+        network."""
+        # Possible patterns:
+        # ( 256, 3) -> conv
+        # ( 256,-2) -> deconv
+        # (None,-2) -> bilinear interpolate
+        in_channels = self.in_channels
+        protonets = nn.ModuleList()
+        for num_channels, kernel_size in zip(self.proto_channels,
+                                             self.proto_kernel_sizes):
+            if kernel_size > 0:
+                layer = nn.Conv2d(
+                    in_channels,
+                    num_channels,
+                    kernel_size,
+                    padding=kernel_size // 2)
+            else:
+                if num_channels is None:
+                    layer = InterpolateModule(
+                        scale_factor=-kernel_size,
+                        mode='bilinear',
+                        align_corners=False)
+                else:
+                    layer = nn.ConvTranspose2d(
+                        in_channels,
+                        num_channels,
+                        -kernel_size,
+                        padding=kernel_size // 2)
+            protonets.append(layer)
+            protonets.append(nn.ReLU(inplace=True))
+            in_channels = num_channels if num_channels is not None \
+                else in_channels
+        if not self.include_last_relu:
+            protonets = protonets[:-1]
+        return nn.Sequential(*protonets)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.protonet:
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    def forward(self, x, coeff_pred, bboxes, img_meta, sampling_results=None):
+        """Forward feature from the upstream network to get prototypes and
+        linearly combine the prototypes, using masks coefficients, into
+        instance masks. Finally, crop the instance masks with given bboxes.
+
+        Args:
+            x (Tensor): Feature from the upstream network, which is
+                a 4D-tensor.
+            coeff_pred (list[Tensor]): Mask coefficients for each scale
+                level with shape (N, num_anchors * num_protos, H, W).
+            bboxes (list[Tensor]): Box used for cropping with shape
+                (N, num_anchors * 4, H, W). During training, they are
+                ground truth boxes. During testing, they are predicted
+                boxes.
+            img_meta (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            sampling_results (List[:obj:``SamplingResult``]): Sampler results
+                for each image.
+
+        Returns:
+            list[Tensor]: Predicted instance segmentation masks.
+        """
+        prototypes = self.protonet(x)
+        prototypes = prototypes.permute(0, 2, 3, 1).contiguous()
+
+        num_imgs = x.size(0)
+        # Training state
+        if self.training:
+            coeff_pred_list = []
+            for coeff_pred_per_level in coeff_pred:
+                coeff_pred_per_level = \
+                    coeff_pred_per_level.permute(0, 2, 3, 1)\
+                    .reshape(num_imgs, -1, self.num_protos)
+                coeff_pred_list.append(coeff_pred_per_level)
+            coeff_pred = torch.cat(coeff_pred_list, dim=1)
+
+        mask_pred_list = []
+        for idx in range(num_imgs):
+            cur_prototypes = prototypes[idx]
+            cur_coeff_pred = coeff_pred[idx]
+            cur_bboxes = bboxes[idx]
+            cur_img_meta = img_meta[idx]
+
+            # Testing state
+            if not self.training:
+                bboxes_for_cropping = cur_bboxes
+            else:
+                cur_sampling_results = sampling_results[idx]
+                pos_assigned_gt_inds = \
+                    cur_sampling_results.pos_assigned_gt_inds
+                bboxes_for_cropping = cur_bboxes[pos_assigned_gt_inds].clone()
+                pos_inds = cur_sampling_results.pos_inds
+                cur_coeff_pred = cur_coeff_pred[pos_inds]
+
+            # Linearly combine the prototypes with the mask coefficients
+            mask_pred = cur_prototypes @ cur_coeff_pred.t()
+            mask_pred = torch.sigmoid(mask_pred)
+
+            h, w = cur_img_meta['img_shape'][:2]
+            bboxes_for_cropping[:, 0] /= w
+            bboxes_for_cropping[:, 1] /= h
+            bboxes_for_cropping[:, 2] /= w
+            bboxes_for_cropping[:, 3] /= h
+
+            mask_pred = self.crop(mask_pred, bboxes_for_cropping)
+            mask_pred = mask_pred.permute(2, 0, 1).contiguous()
+            mask_pred_list.append(mask_pred)
+        return mask_pred_list
+
+    @force_fp32(apply_to=('mask_pred', ))
+    def loss(self, mask_pred, gt_masks, gt_bboxes, img_meta, sampling_results):
+        """Compute loss of the head.
+
+        Args:
+            mask_pred (list[Tensor]): Predicted prototypes with shape
+                (num_classes, H, W).
+            gt_masks (list[Tensor]): Ground truth masks for each image with
+                the same shape of the input image.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            img_meta (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            sampling_results (List[:obj:``SamplingResult``]): Sampler results
+                for each image.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        loss_mask = []
+        num_imgs = len(mask_pred)
+        total_pos = 0
+        for idx in range(num_imgs):
+            cur_mask_pred = mask_pred[idx]
+            cur_gt_masks = gt_masks[idx].float()
+            cur_gt_bboxes = gt_bboxes[idx]
+            cur_img_meta = img_meta[idx]
+            cur_sampling_results = sampling_results[idx]
+
+            pos_assigned_gt_inds = cur_sampling_results.pos_assigned_gt_inds
+            num_pos = pos_assigned_gt_inds.size(0)
+            # Since we're producing (near) full image masks,
+            # it'd take too much vram to backprop on every single mask.
+            # Thus we select only a subset.
+            if num_pos > self.max_masks_to_train:
+                perm = torch.randperm(num_pos)
+                select = perm[:self.max_masks_to_train]
+                cur_mask_pred = cur_mask_pred[select]
+                pos_assigned_gt_inds = pos_assigned_gt_inds[select]
+                num_pos = self.max_masks_to_train
+            total_pos += num_pos
+
+            gt_bboxes_for_reweight = cur_gt_bboxes[pos_assigned_gt_inds]
+
+            mask_targets = self.get_targets(cur_mask_pred, cur_gt_masks,
+                                            pos_assigned_gt_inds)
+            if num_pos == 0:
+                loss = cur_mask_pred.sum() * 0.
+            elif mask_targets is None:
+                loss = F.binary_cross_entropy(cur_mask_pred,
+                                              torch.zeros_like(cur_mask_pred),
+                                              torch.zeros_like(cur_mask_pred))
+            else:
+                cur_mask_pred = torch.clamp(cur_mask_pred, 0, 1)
+                loss = F.binary_cross_entropy(
+                    cur_mask_pred, mask_targets,
+                    reduction='none') * self.loss_mask_weight
+
+                h, w = cur_img_meta['img_shape'][:2]
+                gt_bboxes_width = (gt_bboxes_for_reweight[:, 2] -
+                                   gt_bboxes_for_reweight[:, 0]) / w
+                gt_bboxes_height = (gt_bboxes_for_reweight[:, 3] -
+                                    gt_bboxes_for_reweight[:, 1]) / h
+                loss = loss.mean(dim=(1,
+                                      2)) / gt_bboxes_width / gt_bboxes_height
+                loss = torch.sum(loss)
+            loss_mask.append(loss)
+
+        if total_pos == 0:
+            total_pos += 1  # avoid nan
+        loss_mask = [x / total_pos for x in loss_mask]
+
+        return dict(loss_mask=loss_mask)
+
+    def get_targets(self, mask_pred, gt_masks, pos_assigned_gt_inds):
+        """Compute instance segmentation targets for each image.
+
+        Args:
+            mask_pred (Tensor): Predicted prototypes with shape
+                (num_classes, H, W).
+            gt_masks (Tensor): Ground truth masks for each image with
+                the same shape of the input image.
+            pos_assigned_gt_inds (Tensor): GT indices of the corresponding
+                positive samples.
+        Returns:
+            Tensor: Instance segmentation targets with shape
+                (num_instances, H, W).
+        """
+        if gt_masks.size(0) == 0:
+            return None
+        mask_h, mask_w = mask_pred.shape[-2:]
+        gt_masks = F.interpolate(
+            gt_masks.unsqueeze(0), (mask_h, mask_w),
+            mode='bilinear',
+            align_corners=False).squeeze(0)
+        gt_masks = gt_masks.gt(0.5).float()
+        mask_targets = gt_masks[pos_assigned_gt_inds]
+        return mask_targets
+
+    def get_seg_masks(self, mask_pred, label_pred, img_meta, rescale):
+        """Resize, binarize, and format the instance mask predictions.
+
+        Args:
+            mask_pred (Tensor): shape (N, H, W).
+            label_pred (Tensor): shape (N, ).
+            img_meta (dict): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            rescale (bool): If rescale is False, then returned masks will
+                fit the scale of imgs[0].
+        Returns:
+            list[ndarray]: Mask predictions grouped by their predicted classes.
+        """
+        ori_shape = img_meta['ori_shape']
+        scale_factor = img_meta['scale_factor']
+        if rescale:
+            img_h, img_w = ori_shape[:2]
+        else:
+            img_h = np.round(ori_shape[0] * scale_factor[1]).astype(np.int32)
+            img_w = np.round(ori_shape[1] * scale_factor[0]).astype(np.int32)
+
+        cls_segms = [[] for _ in range(self.num_classes)]
+        if mask_pred.size(0) == 0:
+            return cls_segms
+
+        mask_pred = F.interpolate(
+            mask_pred.unsqueeze(0), (img_h, img_w),
+            mode='bilinear',
+            align_corners=False).squeeze(0) > 0.5
+        mask_pred = mask_pred.cpu().numpy().astype(np.uint8)
+
+        for m, l in zip(mask_pred, label_pred):
+            cls_segms[l].append(m)
+        return cls_segms
+
+    def crop(self, masks, boxes, padding=1):
+        """Crop predicted masks by zeroing out everything not in the predicted
+        bbox.
+
+        Args:
+            masks (Tensor): shape [H, W, N].
+            boxes (Tensor): bbox coords in relative point form with
+                shape [N, 4].
+
+        Return:
+            Tensor: The cropped masks.
+        """
+        h, w, n = masks.size()
+        x1, x2 = self.sanitize_coordinates(
+            boxes[:, 0], boxes[:, 2], w, padding, cast=False)
+        y1, y2 = self.sanitize_coordinates(
+            boxes[:, 1], boxes[:, 3], h, padding, cast=False)
+
+        rows = torch.arange(
+            w, device=masks.device, dtype=x1.dtype).view(1, -1,
+                                                         1).expand(h, w, n)
+        cols = torch.arange(
+            h, device=masks.device, dtype=x1.dtype).view(-1, 1,
+                                                         1).expand(h, w, n)
+
+        masks_left = rows >= x1.view(1, 1, -1)
+        masks_right = rows < x2.view(1, 1, -1)
+        masks_up = cols >= y1.view(1, 1, -1)
+        masks_down = cols < y2.view(1, 1, -1)
+
+        crop_mask = masks_left * masks_right * masks_up * masks_down
+
+        return masks * crop_mask.float()
+
+    def sanitize_coordinates(self, x1, x2, img_size, padding=0, cast=True):
+        """Sanitizes the input coordinates so that x1 < x2, x1 != x2, x1 >= 0,
+        and x2 <= image_size. Also converts from relative to absolute
+        coordinates and casts the results to long tensors.
+
+        Warning: this does things in-place behind the scenes so
+        copy if necessary.
+
+        Args:
+            _x1 (Tensor): shape (N, ).
+            _x2 (Tensor): shape (N, ).
+            img_size (int): Size of the input image.
+            padding (int): x1 >= padding, x2 <= image_size-padding.
+            cast (bool): If cast is false, the result won't be cast to longs.
+
+        Returns:
+            tuple:
+                x1 (Tensor): Sanitized _x1.
+                x2 (Tensor): Sanitized _x2.
+        """
+        x1 = x1 * img_size
+        x2 = x2 * img_size
+        if cast:
+            x1 = x1.long()
+            x2 = x2.long()
+        x1 = torch.min(x1, x2)
+        x2 = torch.max(x1, x2)
+        x1 = torch.clamp(x1 - padding, min=0)
+        x2 = torch.clamp(x2 + padding, max=img_size)
+        return x1, x2
+
+
+class InterpolateModule(nn.Module):
+    """This is a module version of F.interpolate.
+
+    Any arguments you give it just get passed along for the ride.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+        self.args = args
+        self.kwargs = kwargs
+
+    def forward(self, x):
+        """Forward features from the upstream network."""
+        return F.interpolate(x, *self.args, **self.kwargs)
diff --git a/detection/scrfd/mmdet/models/dense_heads/yolo_head.py b/detection/scrfd/mmdet/models/dense_heads/yolo_head.py
new file mode 100755
index 000000000..93d051e7a
--- /dev/null
+++ b/detection/scrfd/mmdet/models/dense_heads/yolo_head.py
@@ -0,0 +1,536 @@
+# Copyright (c) 2019 Western Digital Corporation or its affiliates.
+
+import warnings
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, normal_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import (build_anchor_generator, build_assigner,
+                        build_bbox_coder, build_sampler, images_to_levels,
+                        multi_apply, multiclass_nms)
+from ..builder import HEADS, build_loss
+from .base_dense_head import BaseDenseHead
+from .dense_test_mixins import BBoxTestMixin
+
+
+@HEADS.register_module()
+class YOLOV3Head(BaseDenseHead, BBoxTestMixin):
+    """YOLOV3Head Paper link: https://arxiv.org/abs/1804.02767.
+
+    Args:
+        num_classes (int): The number of object classes (w/o background)
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (List[int]): The number of output channels per scale
+            before the final 1x1 layer. Default: (1024, 512, 256).
+        anchor_generator (dict): Config dict for anchor generator
+        bbox_coder (dict): Config of bounding box coder.
+        featmap_strides (List[int]): The stride of each scale.
+            Should be in descending order. Default: (32, 16, 8).
+        one_hot_smoother (float): Set a non-zero value to enable label-smooth
+            Default: 0.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+            Default: dict(type='BN', requires_grad=True)
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='LeakyReLU', negative_slope=0.1).
+        loss_cls (dict): Config of classification loss.
+        loss_conf (dict): Config of confidence loss.
+        loss_xy (dict): Config of xy coordinate loss.
+        loss_wh (dict): Config of wh coordinate loss.
+        train_cfg (dict): Training config of YOLOV3 head. Default: None.
+        test_cfg (dict): Testing config of YOLOV3 head. Default: None.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 out_channels=(1024, 512, 256),
+                 anchor_generator=dict(
+                     type='YOLOAnchorGenerator',
+                     base_sizes=[[(116, 90), (156, 198), (373, 326)],
+                                 [(30, 61), (62, 45), (59, 119)],
+                                 [(10, 13), (16, 30), (33, 23)]],
+                     strides=[32, 16, 8]),
+                 bbox_coder=dict(type='YOLOBBoxCoder'),
+                 featmap_strides=[32, 16, 8],
+                 one_hot_smoother=0.,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 act_cfg=dict(type='LeakyReLU', negative_slope=0.1),
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 loss_conf=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 loss_xy=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 loss_wh=dict(type='MSELoss', loss_weight=1.0),
+                 train_cfg=None,
+                 test_cfg=None):
+        super(YOLOV3Head, self).__init__()
+        # Check params
+        assert (len(in_channels) == len(out_channels) == len(featmap_strides))
+
+        self.num_classes = num_classes
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.featmap_strides = featmap_strides
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        if self.train_cfg:
+            self.assigner = build_assigner(self.train_cfg.assigner)
+            if hasattr(self.train_cfg, 'sampler'):
+                sampler_cfg = self.train_cfg.sampler
+            else:
+                sampler_cfg = dict(type='PseudoSampler')
+            self.sampler = build_sampler(sampler_cfg, context=self)
+
+        self.one_hot_smoother = one_hot_smoother
+
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.act_cfg = act_cfg
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.anchor_generator = build_anchor_generator(anchor_generator)
+
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_conf = build_loss(loss_conf)
+        self.loss_xy = build_loss(loss_xy)
+        self.loss_wh = build_loss(loss_wh)
+        # usually the numbers of anchors for each level are the same
+        # except SSD detectors
+        self.num_anchors = self.anchor_generator.num_base_anchors[0]
+        assert len(
+            self.anchor_generator.num_base_anchors) == len(featmap_strides)
+        self._init_layers()
+
+    @property
+    def num_levels(self):
+        return len(self.featmap_strides)
+
+    @property
+    def num_attrib(self):
+        """int: number of attributes in pred_map, bboxes (4) +
+        objectness (1) + num_classes"""
+
+        return 5 + self.num_classes
+
+    def _init_layers(self):
+        self.convs_bridge = nn.ModuleList()
+        self.convs_pred = nn.ModuleList()
+        for i in range(self.num_levels):
+            conv_bridge = ConvModule(
+                self.in_channels[i],
+                self.out_channels[i],
+                3,
+                padding=1,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg,
+                act_cfg=self.act_cfg)
+            conv_pred = nn.Conv2d(self.out_channels[i],
+                                  self.num_anchors * self.num_attrib, 1)
+
+            self.convs_bridge.append(conv_bridge)
+            self.convs_pred.append(conv_pred)
+
+    def init_weights(self):
+        """Initialize weights of the head."""
+        for m in self.convs_pred:
+            normal_init(m, std=0.01)
+
+    def forward(self, feats):
+        """Forward features from the upstream network.
+
+        Args:
+            feats (tuple[Tensor]): Features from the upstream network, each is
+                a 4D-tensor.
+
+        Returns:
+            tuple[Tensor]: A tuple of multi-level predication map, each is a
+                4D-tensor of shape (batch_size, 5+num_classes, height, width).
+        """
+
+        assert len(feats) == self.num_levels
+        pred_maps = []
+        for i in range(self.num_levels):
+            x = feats[i]
+            x = self.convs_bridge[i](x)
+            pred_map = self.convs_pred[i](x)
+            pred_maps.append(pred_map)
+
+        return tuple(pred_maps),
+
+    @force_fp32(apply_to=('pred_maps', ))
+    def get_bboxes(self,
+                   pred_maps,
+                   img_metas,
+                   cfg=None,
+                   rescale=False,
+                   with_nms=True):
+        """Transform network output for a batch into bbox predictions.
+
+        Args:
+            pred_maps (list[Tensor]): Raw predictions for a batch of images.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used. Default: None.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class label of the
+                corresponding box.
+        """
+        result_list = []
+        num_levels = len(pred_maps)
+        for img_id in range(len(img_metas)):
+            pred_maps_list = [
+                pred_maps[i][img_id].detach() for i in range(num_levels)
+            ]
+            scale_factor = img_metas[img_id]['scale_factor']
+            proposals = self._get_bboxes_single(pred_maps_list, scale_factor,
+                                                cfg, rescale, with_nms)
+            result_list.append(proposals)
+        return result_list
+
+    def _get_bboxes_single(self,
+                           pred_maps_list,
+                           scale_factor,
+                           cfg,
+                           rescale=False,
+                           with_nms=True):
+        """Transform outputs for a single batch item into bbox predictions.
+
+        Args:
+            pred_maps_list (list[Tensor]): Prediction maps for different scales
+                of each single image in the batch.
+            scale_factor (ndarray): Scale factor of the image arrange as
+                (w_scale, h_scale, w_scale, h_scale).
+            cfg (mmcv.Config | None): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+            with_nms (bool): If True, do nms before return boxes.
+                Default: True.
+
+        Returns:
+            tuple(Tensor):
+                det_bboxes (Tensor): BBox predictions in shape (n, 5), where
+                    the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
+                    between 0 and 1.
+                det_labels (Tensor): A (n,) tensor where each item is the
+                    predicted class label of the corresponding box.
+        """
+        cfg = self.test_cfg if cfg is None else cfg
+        assert len(pred_maps_list) == self.num_levels
+        multi_lvl_bboxes = []
+        multi_lvl_cls_scores = []
+        multi_lvl_conf_scores = []
+        num_levels = len(pred_maps_list)
+        featmap_sizes = [
+            pred_maps_list[i].shape[-2:] for i in range(num_levels)
+        ]
+        multi_lvl_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, pred_maps_list[0][0].device)
+        for i in range(self.num_levels):
+            # get some key info for current scale
+            pred_map = pred_maps_list[i]
+            stride = self.featmap_strides[i]
+
+            # (h, w, num_anchors*num_attrib) -> (h*w*num_anchors, num_attrib)
+            pred_map = pred_map.permute(1, 2, 0).reshape(-1, self.num_attrib)
+
+            pred_map[..., :2] = torch.sigmoid(pred_map[..., :2])
+            bbox_pred = self.bbox_coder.decode(multi_lvl_anchors[i],
+                                               pred_map[..., :4], stride)
+            # conf and cls
+            conf_pred = torch.sigmoid(pred_map[..., 4]).view(-1)
+            cls_pred = torch.sigmoid(pred_map[..., 5:]).view(
+                -1, self.num_classes)  # Cls pred one-hot.
+
+            # Filtering out all predictions with conf < conf_thr
+            conf_thr = cfg.get('conf_thr', -1)
+            if conf_thr > 0:
+                # add as_tuple=False for compatibility in Pytorch 1.6
+                # flatten would create a Reshape op with constant values,
+                # and raise RuntimeError when doing inference in ONNX Runtime
+                # with a different input image (#4221).
+                conf_inds = conf_pred.ge(conf_thr).nonzero(
+                    as_tuple=False).squeeze(1)
+                bbox_pred = bbox_pred[conf_inds, :]
+                cls_pred = cls_pred[conf_inds, :]
+                conf_pred = conf_pred[conf_inds]
+
+            # Get top-k prediction
+            nms_pre = cfg.get('nms_pre', -1)
+            if 0 < nms_pre < conf_pred.size(0) and (
+                    not torch.onnx.is_in_onnx_export()):
+                _, topk_inds = conf_pred.topk(nms_pre)
+                bbox_pred = bbox_pred[topk_inds, :]
+                cls_pred = cls_pred[topk_inds, :]
+                conf_pred = conf_pred[topk_inds]
+
+            # Save the result of current scale
+            multi_lvl_bboxes.append(bbox_pred)
+            multi_lvl_cls_scores.append(cls_pred)
+            multi_lvl_conf_scores.append(conf_pred)
+
+        # Merge the results of different scales together
+        multi_lvl_bboxes = torch.cat(multi_lvl_bboxes)
+        multi_lvl_cls_scores = torch.cat(multi_lvl_cls_scores)
+        multi_lvl_conf_scores = torch.cat(multi_lvl_conf_scores)
+
+        if with_nms and (multi_lvl_conf_scores.size(0) == 0):
+            return torch.zeros((0, 5)), torch.zeros((0, ))
+
+        if rescale:
+            multi_lvl_bboxes /= multi_lvl_bboxes.new_tensor(scale_factor)
+
+        # In mmdet 2.x, the class_id for background is num_classes.
+        # i.e., the last column.
+        padding = multi_lvl_cls_scores.new_zeros(multi_lvl_cls_scores.shape[0],
+                                                 1)
+        multi_lvl_cls_scores = torch.cat([multi_lvl_cls_scores, padding],
+                                         dim=1)
+
+        # Support exporting to onnx without nms
+        if with_nms and cfg.get('nms', None) is not None:
+            det_bboxes, det_labels = multiclass_nms(
+                multi_lvl_bboxes,
+                multi_lvl_cls_scores,
+                cfg.score_thr,
+                cfg.nms,
+                cfg.max_per_img,
+                score_factors=multi_lvl_conf_scores)
+            return det_bboxes, det_labels
+        else:
+            return (multi_lvl_bboxes, multi_lvl_cls_scores,
+                    multi_lvl_conf_scores)
+
+    @force_fp32(apply_to=('pred_maps', ))
+    def loss(self,
+             pred_maps,
+             gt_bboxes,
+             gt_labels,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Compute loss of the head.
+
+        Args:
+            pred_maps (list[Tensor]): Prediction map for each scale level,
+                shape (N, num_anchors * num_attrib, H, W)
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        num_imgs = len(img_metas)
+        device = pred_maps[0][0].device
+
+        featmap_sizes = [
+            pred_maps[i].shape[-2:] for i in range(self.num_levels)
+        ]
+        multi_level_anchors = self.anchor_generator.grid_anchors(
+            featmap_sizes, device)
+        anchor_list = [multi_level_anchors for _ in range(num_imgs)]
+
+        responsible_flag_list = []
+        for img_id in range(len(img_metas)):
+            responsible_flag_list.append(
+                self.anchor_generator.responsible_flags(
+                    featmap_sizes, gt_bboxes[img_id], device))
+
+        target_maps_list, neg_maps_list = self.get_targets(
+            anchor_list, responsible_flag_list, gt_bboxes, gt_labels)
+
+        losses_cls, losses_conf, losses_xy, losses_wh = multi_apply(
+            self.loss_single, pred_maps, target_maps_list, neg_maps_list)
+
+        return dict(
+            loss_cls=losses_cls,
+            loss_conf=losses_conf,
+            loss_xy=losses_xy,
+            loss_wh=losses_wh)
+
+    def loss_single(self, pred_map, target_map, neg_map):
+        """Compute loss of a single image from a batch.
+
+        Args:
+            pred_map (Tensor): Raw predictions for a single level.
+            target_map (Tensor): The Ground-Truth target for a single level.
+            neg_map (Tensor): The negative masks for a single level.
+
+        Returns:
+            tuple:
+                loss_cls (Tensor): Classification loss.
+                loss_conf (Tensor): Confidence loss.
+                loss_xy (Tensor): Regression loss of x, y coordinate.
+                loss_wh (Tensor): Regression loss of w, h coordinate.
+        """
+
+        num_imgs = len(pred_map)
+        pred_map = pred_map.permute(0, 2, 3,
+                                    1).reshape(num_imgs, -1, self.num_attrib)
+        neg_mask = neg_map.float()
+        pos_mask = target_map[..., 4]
+        pos_and_neg_mask = neg_mask + pos_mask
+        pos_mask = pos_mask.unsqueeze(dim=-1)
+        if torch.max(pos_and_neg_mask) > 1.:
+            warnings.warn('There is overlap between pos and neg sample.')
+            pos_and_neg_mask = pos_and_neg_mask.clamp(min=0., max=1.)
+
+        pred_xy = pred_map[..., :2]
+        pred_wh = pred_map[..., 2:4]
+        pred_conf = pred_map[..., 4]
+        pred_label = pred_map[..., 5:]
+
+        target_xy = target_map[..., :2]
+        target_wh = target_map[..., 2:4]
+        target_conf = target_map[..., 4]
+        target_label = target_map[..., 5:]
+
+        loss_cls = self.loss_cls(pred_label, target_label, weight=pos_mask)
+        loss_conf = self.loss_conf(
+            pred_conf, target_conf, weight=pos_and_neg_mask)
+        loss_xy = self.loss_xy(pred_xy, target_xy, weight=pos_mask)
+        loss_wh = self.loss_wh(pred_wh, target_wh, weight=pos_mask)
+
+        return loss_cls, loss_conf, loss_xy, loss_wh
+
+    def get_targets(self, anchor_list, responsible_flag_list, gt_bboxes_list,
+                    gt_labels_list):
+        """Compute target maps for anchors in multiple images.
+
+        Args:
+            anchor_list (list[list[Tensor]]): Multi level anchors of each
+                image. The outer list indicates images, and the inner list
+                corresponds to feature levels of the image. Each element of
+                the inner list is a tensor of shape (num_total_anchors, 4).
+            responsible_flag_list (list[list[Tensor]]): Multi level responsible
+                flags of each image. Each element is a tensor of shape
+                (num_total_anchors, )
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
+            gt_labels_list (list[Tensor]): Ground truth labels of each box.
+
+        Returns:
+            tuple: Usually returns a tuple containing learning targets.
+                - target_map_list (list[Tensor]): Target map of each level.
+                - neg_map_list (list[Tensor]): Negative map of each level.
+        """
+        num_imgs = len(anchor_list)
+
+        # anchor number of multi levels
+        num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
+
+        results = multi_apply(self._get_targets_single, anchor_list,
+                              responsible_flag_list, gt_bboxes_list,
+                              gt_labels_list)
+
+        all_target_maps, all_neg_maps = results
+        assert num_imgs == len(all_target_maps) == len(all_neg_maps)
+        target_maps_list = images_to_levels(all_target_maps, num_level_anchors)
+        neg_maps_list = images_to_levels(all_neg_maps, num_level_anchors)
+
+        return target_maps_list, neg_maps_list
+
+    def _get_targets_single(self, anchors, responsible_flags, gt_bboxes,
+                            gt_labels):
+        """Generate matching bounding box prior and converted GT.
+
+        Args:
+            anchors (list[Tensor]): Multi-level anchors of the image.
+            responsible_flags (list[Tensor]): Multi-level responsible flags of
+                anchors
+            gt_bboxes (Tensor): Ground truth bboxes of single image.
+            gt_labels (Tensor): Ground truth labels of single image.
+
+        Returns:
+            tuple:
+                target_map (Tensor): Predication target map of each
+                    scale level, shape (num_total_anchors,
+                    5+num_classes)
+                neg_map (Tensor): Negative map of each scale level,
+                    shape (num_total_anchors,)
+        """
+
+        anchor_strides = []
+        for i in range(len(anchors)):
+            anchor_strides.append(
+                torch.tensor(self.featmap_strides[i],
+                             device=gt_bboxes.device).repeat(len(anchors[i])))
+        concat_anchors = torch.cat(anchors)
+        concat_responsible_flags = torch.cat(responsible_flags)
+
+        anchor_strides = torch.cat(anchor_strides)
+        assert len(anchor_strides) == len(concat_anchors) == \
+               len(concat_responsible_flags)
+        assign_result = self.assigner.assign(concat_anchors,
+                                             concat_responsible_flags,
+                                             gt_bboxes)
+        sampling_result = self.sampler.sample(assign_result, concat_anchors,
+                                              gt_bboxes)
+
+        target_map = concat_anchors.new_zeros(
+            concat_anchors.size(0), self.num_attrib)
+
+        target_map[sampling_result.pos_inds, :4] = self.bbox_coder.encode(
+            sampling_result.pos_bboxes, sampling_result.pos_gt_bboxes,
+            anchor_strides[sampling_result.pos_inds])
+
+        target_map[sampling_result.pos_inds, 4] = 1
+
+        gt_labels_one_hot = F.one_hot(
+            gt_labels, num_classes=self.num_classes).float()
+        if self.one_hot_smoother != 0:  # label smooth
+            gt_labels_one_hot = gt_labels_one_hot * (
+                1 - self.one_hot_smoother
+            ) + self.one_hot_smoother / self.num_classes
+        target_map[sampling_result.pos_inds, 5:] = gt_labels_one_hot[
+            sampling_result.pos_assigned_gt_inds]
+
+        neg_map = concat_anchors.new_zeros(
+            concat_anchors.size(0), dtype=torch.uint8)
+        neg_map[sampling_result.neg_inds] = 1
+
+        return target_map, neg_map
+
+    def aug_test(self, feats, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            feats (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains features for all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[ndarray]: bbox results of each class
+        """
+        return self.aug_test_bboxes(feats, img_metas, rescale=rescale)
diff --git a/detection/scrfd/mmdet/models/detectors/__init__.py b/detection/scrfd/mmdet/models/detectors/__init__.py
new file mode 100755
index 000000000..e54a219c2
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/__init__.py
@@ -0,0 +1,36 @@
+from .atss import ATSS
+from .base import BaseDetector
+from .cascade_rcnn import CascadeRCNN
+from .cornernet import CornerNet
+from .detr import DETR
+from .fast_rcnn import FastRCNN
+from .faster_rcnn import FasterRCNN
+from .fcos import FCOS
+from .fovea import FOVEA
+from .fsaf import FSAF
+from .gfl import GFL
+from .grid_rcnn import GridRCNN
+from .htc import HybridTaskCascade
+from .mask_rcnn import MaskRCNN
+from .mask_scoring_rcnn import MaskScoringRCNN
+from .nasfcos import NASFCOS
+from .paa import PAA
+from .point_rend import PointRend
+from .reppoints_detector import RepPointsDetector
+from .retinanet import RetinaNet
+from .rpn import RPN
+from .single_stage import SingleStageDetector
+from .trident_faster_rcnn import TridentFasterRCNN
+from .two_stage import TwoStageDetector
+from .vfnet import VFNet
+from .yolact import YOLACT
+from .yolo import YOLOV3
+from .scrfd import SCRFD
+
+__all__ = [
+    'ATSS', 'BaseDetector', 'SingleStageDetector', 'TwoStageDetector', 'RPN',
+    'FastRCNN', 'FasterRCNN', 'MaskRCNN', 'CascadeRCNN', 'HybridTaskCascade',
+    'RetinaNet', 'FCOS', 'GridRCNN', 'MaskScoringRCNN', 'RepPointsDetector',
+    'FOVEA', 'FSAF', 'NASFCOS', 'PointRend', 'GFL', 'CornerNet', 'PAA',
+    'YOLOV3', 'YOLACT', 'VFNet', 'DETR', 'TridentFasterRCNN', 'SCRFD'
+]
diff --git a/detection/scrfd/mmdet/models/detectors/atss.py b/detection/scrfd/mmdet/models/detectors/atss.py
new file mode 100755
index 000000000..db7139c6b
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/atss.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class ATSS(SingleStageDetector):
+    """Implementation of `ATSS <https://arxiv.org/abs/1912.02424>`_."""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(ATSS, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                   test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/base.py b/detection/scrfd/mmdet/models/detectors/base.py
new file mode 100755
index 000000000..7c6d5e961
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/base.py
@@ -0,0 +1,355 @@
+from abc import ABCMeta, abstractmethod
+from collections import OrderedDict
+
+import mmcv
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from mmcv.runner import auto_fp16
+from mmcv.utils import print_log
+
+from mmdet.utils import get_root_logger
+
+
+class BaseDetector(nn.Module, metaclass=ABCMeta):
+    """Base class for detectors."""
+
+    def __init__(self):
+        super(BaseDetector, self).__init__()
+        self.fp16_enabled = False
+
+    @property
+    def with_neck(self):
+        """bool: whether the detector has a neck"""
+        return hasattr(self, 'neck') and self.neck is not None
+
+    # TODO: these properties need to be carefully handled
+    # for both single stage & two stage detectors
+    @property
+    def with_shared_head(self):
+        """bool: whether the detector has a shared head in the RoI Head"""
+        return hasattr(self, 'roi_head') and self.roi_head.with_shared_head
+
+    @property
+    def with_bbox(self):
+        """bool: whether the detector has a bbox head"""
+        return ((hasattr(self, 'roi_head') and self.roi_head.with_bbox)
+                or (hasattr(self, 'bbox_head') and self.bbox_head is not None))
+
+    @property
+    def with_mask(self):
+        """bool: whether the detector has a mask head"""
+        return ((hasattr(self, 'roi_head') and self.roi_head.with_mask)
+                or (hasattr(self, 'mask_head') and self.mask_head is not None))
+
+    @abstractmethod
+    def extract_feat(self, imgs):
+        """Extract features from images."""
+        pass
+
+    def extract_feats(self, imgs):
+        """Extract features from multiple images.
+
+        Args:
+            imgs (list[torch.Tensor]): A list of images. The images are
+                augmented from the same image but in different ways.
+
+        Returns:
+            list[torch.Tensor]: Features of different images
+        """
+        assert isinstance(imgs, list)
+        return [self.extract_feat(img) for img in imgs]
+
+    def forward_train(self, imgs, img_metas, **kwargs):
+        """
+        Args:
+            img (list[Tensor]): List of tensors of shape (1, C, H, W).
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys, see
+                :class:`mmdet.datasets.pipelines.Collect`.
+            kwargs (keyword arguments): Specific to concrete implementation.
+        """
+        # NOTE the batched image size information may be useful, e.g.
+        # in DETR, this is needed for the construction of masks, which is
+        # then used for the transformer_head.
+        batch_input_shape = tuple(imgs[0].size()[-2:])
+        for img_meta in img_metas:
+            img_meta['batch_input_shape'] = batch_input_shape
+
+    async def async_simple_test(self, img, img_metas, **kwargs):
+        raise NotImplementedError
+
+    @abstractmethod
+    def simple_test(self, img, img_metas, **kwargs):
+        pass
+
+    @abstractmethod
+    def aug_test(self, imgs, img_metas, **kwargs):
+        """Test function with test time augmentation."""
+        pass
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in detector.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if pretrained is not None:
+            logger = get_root_logger()
+            print_log(f'load model from: {pretrained}', logger=logger)
+
+    async def aforward_test(self, *, img, img_metas, **kwargs):
+        for var, name in [(img, 'img'), (img_metas, 'img_metas')]:
+            if not isinstance(var, list):
+                raise TypeError(f'{name} must be a list, but got {type(var)}')
+
+        num_augs = len(img)
+        if num_augs != len(img_metas):
+            raise ValueError(f'num of augmentations ({len(img)}) '
+                             f'!= num of image metas ({len(img_metas)})')
+        # TODO: remove the restriction of samples_per_gpu == 1 when prepared
+        samples_per_gpu = img[0].size(0)
+        assert samples_per_gpu == 1
+
+        if num_augs == 1:
+            return await self.async_simple_test(img[0], img_metas[0], **kwargs)
+        else:
+            raise NotImplementedError
+
+    def forward_test(self, imgs, img_metas, **kwargs):
+        """
+        Args:
+            imgs (List[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains all images in the batch.
+            img_metas (List[List[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch.
+        """
+        for var, name in [(imgs, 'imgs'), (img_metas, 'img_metas')]:
+            if not isinstance(var, list):
+                raise TypeError(f'{name} must be a list, but got {type(var)}')
+
+        num_augs = len(imgs)
+        if num_augs != len(img_metas):
+            raise ValueError(f'num of augmentations ({len(imgs)}) '
+                             f'!= num of image meta ({len(img_metas)})')
+
+        # NOTE the batched image size information may be useful, e.g.
+        # in DETR, this is needed for the construction of masks, which is
+        # then used for the transformer_head.
+        for img, img_meta in zip(imgs, img_metas):
+            batch_size = len(img_meta)
+            for img_id in range(batch_size):
+                img_meta[img_id]['batch_input_shape'] = tuple(img.size()[-2:])
+
+        if num_augs == 1:
+            # proposals (List[List[Tensor]]): the outer list indicates
+            # test-time augs (multiscale, flip, etc.) and the inner list
+            # indicates images in a batch.
+            # The Tensor should have a shape Px4, where P is the number of
+            # proposals.
+            if 'proposals' in kwargs:
+                kwargs['proposals'] = kwargs['proposals'][0]
+            return self.simple_test(imgs[0], img_metas[0], **kwargs)
+        else:
+            assert imgs[0].size(0) == 1, 'aug test does not support ' \
+                                         'inference with batch size ' \
+                                         f'{imgs[0].size(0)}'
+            # TODO: support test augmentation for predefined proposals
+            assert 'proposals' not in kwargs
+            return self.aug_test(imgs, img_metas, **kwargs)
+
+    @auto_fp16(apply_to=('img', ))
+    def forward(self, img, img_metas, return_loss=True, **kwargs):
+        """Calls either :func:`forward_train` or :func:`forward_test` depending
+        on whether ``return_loss`` is ``True``.
+
+        Note this setting will change the expected inputs. When
+        ``return_loss=True``, img and img_meta are single-nested (i.e. Tensor
+        and List[dict]), and when ``resturn_loss=False``, img and img_meta
+        should be double nested (i.e.  List[Tensor], List[List[dict]]), with
+        the outer list indicating test time augmentations.
+        """
+        if return_loss:
+            return self.forward_train(img, img_metas, **kwargs)
+        else:
+            return self.forward_test(img, img_metas, **kwargs)
+
+    def _parse_losses(self, losses):
+        """Parse the raw outputs (losses) of the network.
+
+        Args:
+            losses (dict): Raw output of the network, which usually contain
+                losses and other necessary infomation.
+
+        Returns:
+            tuple[Tensor, dict]: (loss, log_vars), loss is the loss tensor \
+                which may be a weighted sum of all losses, log_vars contains \
+                all the variables to be sent to the logger.
+        """
+        log_vars = OrderedDict()
+        for loss_name, loss_value in losses.items():
+            if isinstance(loss_value, torch.Tensor):
+                log_vars[loss_name] = loss_value.mean()
+            elif isinstance(loss_value, list):
+                log_vars[loss_name] = sum(_loss.mean() for _loss in loss_value)
+            else:
+                raise TypeError(
+                    f'{loss_name} is not a tensor or list of tensors')
+
+        loss = sum(_value for _key, _value in log_vars.items()
+                   if 'loss' in _key)
+
+        log_vars['loss'] = loss
+        for loss_name, loss_value in log_vars.items():
+            # reduce loss when distributed training
+            if dist.is_available() and dist.is_initialized():
+                loss_value = loss_value.data.clone()
+                dist.all_reduce(loss_value.div_(dist.get_world_size()))
+            log_vars[loss_name] = loss_value.item()
+
+        return loss, log_vars
+
+    def train_step(self, data, optimizer):
+        """The iteration step during training.
+
+        This method defines an iteration step during training, except for the
+        back propagation and optimizer updating, which are done in an optimizer
+        hook. Note that in some complicated cases or models, the whole process
+        including back propagation and optimizer updating is also defined in
+        this method, such as GAN.
+
+        Args:
+            data (dict): The output of dataloader.
+            optimizer (:obj:`torch.optim.Optimizer` | dict): The optimizer of
+                runner is passed to ``train_step()``. This argument is unused
+                and reserved.
+
+        Returns:
+            dict: It should contain at least 3 keys: ``loss``, ``log_vars``, \
+                ``num_samples``.
+
+                - ``loss`` is a tensor for back propagation, which can be a \
+                weighted sum of multiple losses.
+                - ``log_vars`` contains all the variables to be sent to the
+                logger.
+                - ``num_samples`` indicates the batch size (when the model is \
+                DDP, it means the batch size on each GPU), which is used for \
+                averaging the logs.
+        """
+        losses = self(**data)
+        loss, log_vars = self._parse_losses(losses)
+
+        outputs = dict(
+            loss=loss, log_vars=log_vars, num_samples=len(data['img_metas']))
+
+        return outputs
+
+    def val_step(self, data, optimizer):
+        """The iteration step during validation.
+
+        This method shares the same signature as :func:`train_step`, but used
+        during val epochs. Note that the evaluation after training epochs is
+        not implemented with this method, but an evaluation hook.
+        """
+        losses = self(**data)
+        loss, log_vars = self._parse_losses(losses)
+
+        outputs = dict(
+            loss=loss, log_vars=log_vars, num_samples=len(data['img_metas']))
+
+        return outputs
+
+    def show_result(self,
+                    img,
+                    result,
+                    score_thr=0.3,
+                    bbox_color='green',
+                    text_color='green',
+                    thickness=1,
+                    font_scale=0.5,
+                    win_name='',
+                    show=False,
+                    wait_time=0,
+                    out_file=None):
+        """Draw `result` over `img`.
+
+        Args:
+            img (str or Tensor): The image to be displayed.
+            result (Tensor or tuple): The results to draw over `img`
+                bbox_result or (bbox_result, segm_result).
+            score_thr (float, optional): Minimum score of bboxes to be shown.
+                Default: 0.3.
+            bbox_color (str or tuple or :obj:`Color`): Color of bbox lines.
+            text_color (str or tuple or :obj:`Color`): Color of texts.
+            thickness (int): Thickness of lines.
+            font_scale (float): Font scales of texts.
+            win_name (str): The window name.
+            wait_time (int): Value of waitKey param.
+                Default: 0.
+            show (bool): Whether to show the image.
+                Default: False.
+            out_file (str or None): The filename to write the image.
+                Default: None.
+
+        Returns:
+            img (Tensor): Only if not `show` or `out_file`
+        """
+        img = mmcv.imread(img)
+        img = img.copy()
+        if isinstance(result, tuple):
+            bbox_result, segm_result = result
+            if isinstance(segm_result, tuple):
+                segm_result = segm_result[0]  # ms rcnn
+        else:
+            bbox_result, segm_result = result, None
+        bboxes = np.vstack(bbox_result)
+        labels = [
+            np.full(bbox.shape[0], i, dtype=np.int32)
+            for i, bbox in enumerate(bbox_result)
+        ]
+        labels = np.concatenate(labels)
+        # draw segmentation masks
+        if segm_result is not None and len(labels) > 0:  # non empty
+            segms = mmcv.concat_list(segm_result)
+            inds = np.where(bboxes[:, -1] > score_thr)[0]
+            np.random.seed(42)
+            color_masks = [
+                np.random.randint(0, 256, (1, 3), dtype=np.uint8)
+                for _ in range(max(labels) + 1)
+            ]
+            for i in inds:
+                i = int(i)
+                color_mask = color_masks[labels[i]]
+                sg = segms[i]
+                if isinstance(sg, torch.Tensor):
+                    sg = sg.detach().cpu().numpy()
+                mask = sg.astype(bool)
+                img[mask] = img[mask] * 0.5 + color_mask * 0.5
+        # if out_file specified, do not show image in window
+        if out_file is not None:
+            show = False
+        # draw bounding boxes
+        mmcv.imshow_det_bboxes(
+            img,
+            bboxes,
+            labels,
+            class_names=self.CLASSES,
+            score_thr=score_thr,
+            bbox_color=bbox_color,
+            text_color=text_color,
+            thickness=thickness,
+            font_scale=font_scale,
+            win_name=win_name,
+            show=show,
+            wait_time=wait_time,
+            out_file=out_file)
+
+        if not (show or out_file):
+            return img
diff --git a/detection/scrfd/mmdet/models/detectors/cascade_rcnn.py b/detection/scrfd/mmdet/models/detectors/cascade_rcnn.py
new file mode 100755
index 000000000..47cc7cef9
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/cascade_rcnn.py
@@ -0,0 +1,37 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class CascadeRCNN(TwoStageDetector):
+    r"""Implementation of `Cascade R-CNN: Delving into High Quality Object
+    Detection <https://arxiv.org/abs/1906.09756>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck=None,
+                 rpn_head=None,
+                 roi_head=None,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(CascadeRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
+
+    def show_result(self, data, result, **kwargs):
+        """Show prediction results of the detector."""
+        if self.with_mask:
+            ms_bbox_result, ms_segm_result = result
+            if isinstance(ms_bbox_result, dict):
+                result = (ms_bbox_result['ensemble'],
+                          ms_segm_result['ensemble'])
+        else:
+            if isinstance(result, dict):
+                result = result['ensemble']
+        return super(CascadeRCNN, self).show_result(data, result, **kwargs)
diff --git a/detection/scrfd/mmdet/models/detectors/cornernet.py b/detection/scrfd/mmdet/models/detectors/cornernet.py
new file mode 100755
index 000000000..bb8ccc146
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/cornernet.py
@@ -0,0 +1,95 @@
+import torch
+
+from mmdet.core import bbox2result, bbox_mapping_back
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class CornerNet(SingleStageDetector):
+    """CornerNet.
+
+    This detector is the implementation of the paper `CornerNet: Detecting
+    Objects as Paired Keypoints <https://arxiv.org/abs/1808.01244>`_ .
+    """
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(CornerNet, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                        test_cfg, pretrained)
+
+    def merge_aug_results(self, aug_results, img_metas):
+        """Merge augmented detection bboxes and score.
+
+        Args:
+            aug_results (list[list[Tensor]]): Det_bboxes and det_labels of each
+                image.
+            img_metas (list[list[dict]]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+
+        Returns:
+            tuple: (bboxes, labels)
+        """
+        recovered_bboxes, aug_labels = [], []
+        for bboxes_labels, img_info in zip(aug_results, img_metas):
+            img_shape = img_info[0]['img_shape']  # using shape before padding
+            scale_factor = img_info[0]['scale_factor']
+            flip = img_info[0]['flip']
+            bboxes, labels = bboxes_labels
+            bboxes, scores = bboxes[:, :4], bboxes[:, -1:]
+            bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
+            recovered_bboxes.append(torch.cat([bboxes, scores], dim=-1))
+            aug_labels.append(labels)
+
+        bboxes = torch.cat(recovered_bboxes, dim=0)
+        labels = torch.cat(aug_labels)
+
+        if bboxes.shape[0] > 0:
+            out_bboxes, out_labels = self.bbox_head._bboxes_nms(
+                bboxes, labels, self.bbox_head.test_cfg)
+        else:
+            out_bboxes, out_labels = bboxes, labels
+
+        return out_bboxes, out_labels
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Augment testing of CornerNet.
+
+        Args:
+            imgs (list[Tensor]): Augmented images.
+            img_metas (list[list[dict]]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            rescale (bool): If True, return boxes in original image space.
+                Default: False.
+
+        Note:
+            ``imgs`` must including flipped image pairs.
+
+        Returns:
+            list[list[np.ndarray]]: BBox results of each image and classes.
+                The outer list corresponds to each image. The inner list
+                corresponds to each class.
+        """
+        img_inds = list(range(len(imgs)))
+
+        assert img_metas[0][0]['flip'] + img_metas[1][0]['flip'], (
+            'aug test must have flipped image pair')
+        aug_results = []
+        for ind, flip_ind in zip(img_inds[0::2], img_inds[1::2]):
+            img_pair = torch.cat([imgs[ind], imgs[flip_ind]])
+            x = self.extract_feat(img_pair)
+            outs = self.bbox_head(x)
+            bbox_list = self.bbox_head.get_bboxes(
+                *outs, [img_metas[ind], img_metas[flip_ind]], False, False)
+            aug_results.append(bbox_list[0])
+            aug_results.append(bbox_list[1])
+
+        bboxes, labels = self.merge_aug_results(aug_results, img_metas)
+        bbox_results = bbox2result(bboxes, labels, self.bbox_head.num_classes)
+
+        return [bbox_results]
diff --git a/detection/scrfd/mmdet/models/detectors/detr.py b/detection/scrfd/mmdet/models/detectors/detr.py
new file mode 100755
index 000000000..5ff82a280
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/detr.py
@@ -0,0 +1,46 @@
+from mmdet.core import bbox2result
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class DETR(SingleStageDetector):
+    r"""Implementation of `DETR: End-to-End Object Detection with
+    Transformers <https://arxiv.org/pdf/2005.12872>`_"""
+
+    def __init__(self,
+                 backbone,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(DETR, self).__init__(backbone, None, bbox_head, train_cfg,
+                                   test_cfg, pretrained)
+
+    def simple_test(self, img, img_metas, rescale=False):
+        """Test function without test time augmentation.
+
+        Args:
+            imgs (list[torch.Tensor]): List of multiple images
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[list[np.ndarray]]: BBox results of each image and classes.
+                The outer list corresponds to each image. The inner list
+                corresponds to each class.
+        """
+        batch_size = len(img_metas)
+        assert batch_size == 1, 'Currently only batch_size 1 for inference ' \
+            f'mode is supported. Found batch_size {batch_size}.'
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x, img_metas)
+        bbox_list = self.bbox_head.get_bboxes(
+            *outs, img_metas, rescale=rescale)
+
+        bbox_results = [
+            bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
+            for det_bboxes, det_labels in bbox_list
+        ]
+        return bbox_results
diff --git a/detection/scrfd/mmdet/models/detectors/fast_rcnn.py b/detection/scrfd/mmdet/models/detectors/fast_rcnn.py
new file mode 100755
index 000000000..3d6e24276
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/fast_rcnn.py
@@ -0,0 +1,52 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class FastRCNN(TwoStageDetector):
+    """Implementation of `Fast R-CNN <https://arxiv.org/abs/1504.08083>`_"""
+
+    def __init__(self,
+                 backbone,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(FastRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
+
+    def forward_test(self, imgs, img_metas, proposals, **kwargs):
+        """
+        Args:
+            imgs (List[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains all images in the batch.
+            img_metas (List[List[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch.
+            proposals (List[List[Tensor]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. The Tensor should have a shape Px4, where
+                P is the number of proposals.
+        """
+        for var, name in [(imgs, 'imgs'), (img_metas, 'img_metas')]:
+            if not isinstance(var, list):
+                raise TypeError(f'{name} must be a list, but got {type(var)}')
+
+        num_augs = len(imgs)
+        if num_augs != len(img_metas):
+            raise ValueError(f'num of augmentations ({len(imgs)}) '
+                             f'!= num of image meta ({len(img_metas)})')
+
+        if num_augs == 1:
+            return self.simple_test(imgs[0], img_metas[0], proposals[0],
+                                    **kwargs)
+        else:
+            # TODO: support test-time augmentation
+            assert NotImplementedError
diff --git a/detection/scrfd/mmdet/models/detectors/faster_rcnn.py b/detection/scrfd/mmdet/models/detectors/faster_rcnn.py
new file mode 100755
index 000000000..81bad0f43
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/faster_rcnn.py
@@ -0,0 +1,24 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class FasterRCNN(TwoStageDetector):
+    """Implementation of `Faster R-CNN <https://arxiv.org/abs/1506.01497>`_"""
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(FasterRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/fcos.py b/detection/scrfd/mmdet/models/detectors/fcos.py
new file mode 100755
index 000000000..58485c186
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/fcos.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class FCOS(SingleStageDetector):
+    """Implementation of `FCOS <https://arxiv.org/abs/1904.01355>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(FCOS, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                   test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/fovea.py b/detection/scrfd/mmdet/models/detectors/fovea.py
new file mode 100755
index 000000000..22a578eff
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/fovea.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class FOVEA(SingleStageDetector):
+    """Implementation of `FoveaBox <https://arxiv.org/abs/1904.03797>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(FOVEA, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                    test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/fsaf.py b/detection/scrfd/mmdet/models/detectors/fsaf.py
new file mode 100755
index 000000000..9f10fa1ae
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/fsaf.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class FSAF(SingleStageDetector):
+    """Implementation of `FSAF <https://arxiv.org/abs/1903.00621>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(FSAF, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                   test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/gfl.py b/detection/scrfd/mmdet/models/detectors/gfl.py
new file mode 100755
index 000000000..64d65cb2d
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/gfl.py
@@ -0,0 +1,16 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class GFL(SingleStageDetector):
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(GFL, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                  test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/grid_rcnn.py b/detection/scrfd/mmdet/models/detectors/grid_rcnn.py
new file mode 100755
index 000000000..b6145a146
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/grid_rcnn.py
@@ -0,0 +1,29 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class GridRCNN(TwoStageDetector):
+    """Grid R-CNN.
+
+    This detector is the implementation of:
+    - Grid R-CNN (https://arxiv.org/abs/1811.12030)
+    - Grid R-CNN Plus: Faster and Better (https://arxiv.org/abs/1906.05688)
+    """
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(GridRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/htc.py b/detection/scrfd/mmdet/models/detectors/htc.py
new file mode 100755
index 000000000..d9efdf420
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/htc.py
@@ -0,0 +1,15 @@
+from ..builder import DETECTORS
+from .cascade_rcnn import CascadeRCNN
+
+
+@DETECTORS.register_module()
+class HybridTaskCascade(CascadeRCNN):
+    """Implementation of `HTC <https://arxiv.org/abs/1901.07518>`_"""
+
+    def __init__(self, **kwargs):
+        super(HybridTaskCascade, self).__init__(**kwargs)
+
+    @property
+    def with_semantic(self):
+        """bool: whether the detector has a semantic head"""
+        return self.roi_head.with_semantic
diff --git a/detection/scrfd/mmdet/models/detectors/mask_rcnn.py b/detection/scrfd/mmdet/models/detectors/mask_rcnn.py
new file mode 100755
index 000000000..c15a77331
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/mask_rcnn.py
@@ -0,0 +1,24 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class MaskRCNN(TwoStageDetector):
+    """Implementation of `Mask R-CNN <https://arxiv.org/abs/1703.06870>`_"""
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(MaskRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/mask_scoring_rcnn.py b/detection/scrfd/mmdet/models/detectors/mask_scoring_rcnn.py
new file mode 100755
index 000000000..b6252b6e1
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/mask_scoring_rcnn.py
@@ -0,0 +1,27 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class MaskScoringRCNN(TwoStageDetector):
+    """Mask Scoring RCNN.
+
+    https://arxiv.org/abs/1903.00241
+    """
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(MaskScoringRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/nasfcos.py b/detection/scrfd/mmdet/models/detectors/nasfcos.py
new file mode 100755
index 000000000..fb0148351
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/nasfcos.py
@@ -0,0 +1,20 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class NASFCOS(SingleStageDetector):
+    """NAS-FCOS: Fast Neural Architecture Search for Object Detection.
+
+    https://arxiv.org/abs/1906.0442
+    """
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(NASFCOS, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                      test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/paa.py b/detection/scrfd/mmdet/models/detectors/paa.py
new file mode 100755
index 000000000..9b4bb5e09
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/paa.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class PAA(SingleStageDetector):
+    """Implementation of `PAA <https://arxiv.org/pdf/2007.08103.pdf>`_."""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(PAA, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                  test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/point_rend.py b/detection/scrfd/mmdet/models/detectors/point_rend.py
new file mode 100755
index 000000000..808ef2258
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/point_rend.py
@@ -0,0 +1,29 @@
+from ..builder import DETECTORS
+from .two_stage import TwoStageDetector
+
+
+@DETECTORS.register_module()
+class PointRend(TwoStageDetector):
+    """PointRend: Image Segmentation as Rendering
+
+    This detector is the implementation of
+    `PointRend <https://arxiv.org/abs/1912.08193>`_.
+
+    """
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+        super(PointRend, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/reppoints_detector.py b/detection/scrfd/mmdet/models/detectors/reppoints_detector.py
new file mode 100755
index 000000000..a5f6be31e
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/reppoints_detector.py
@@ -0,0 +1,22 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class RepPointsDetector(SingleStageDetector):
+    """RepPoints: Point Set Representation for Object Detection.
+
+        This detector is the implementation of:
+        - RepPoints detector (https://arxiv.org/pdf/1904.11490)
+    """
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(RepPointsDetector,
+              self).__init__(backbone, neck, bbox_head, train_cfg, test_cfg,
+                             pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/retinanet.py b/detection/scrfd/mmdet/models/detectors/retinanet.py
new file mode 100755
index 000000000..41378e8bc
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/retinanet.py
@@ -0,0 +1,17 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class RetinaNet(SingleStageDetector):
+    """Implementation of `RetinaNet <https://arxiv.org/abs/1708.02002>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(RetinaNet, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                        test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/rpn.py b/detection/scrfd/mmdet/models/detectors/rpn.py
new file mode 100755
index 000000000..36c38afc5
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/rpn.py
@@ -0,0 +1,153 @@
+import mmcv
+from mmcv.image import tensor2imgs
+
+from mmdet.core import bbox_mapping
+from ..builder import DETECTORS, build_backbone, build_head, build_neck
+from .base import BaseDetector
+
+
+@DETECTORS.register_module()
+class RPN(BaseDetector):
+    """Implementation of Region Proposal Network."""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 rpn_head,
+                 train_cfg,
+                 test_cfg,
+                 pretrained=None):
+        super(RPN, self).__init__()
+        self.backbone = build_backbone(backbone)
+        self.neck = build_neck(neck) if neck is not None else None
+        rpn_train_cfg = train_cfg.rpn if train_cfg is not None else None
+        rpn_head.update(train_cfg=rpn_train_cfg)
+        rpn_head.update(test_cfg=test_cfg.rpn)
+        self.rpn_head = build_head(rpn_head)
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        self.init_weights(pretrained=pretrained)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in detector.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(RPN, self).init_weights(pretrained)
+        self.backbone.init_weights(pretrained=pretrained)
+        if self.with_neck:
+            self.neck.init_weights()
+        self.rpn_head.init_weights()
+
+    def extract_feat(self, img):
+        """Extract features.
+
+        Args:
+            img (torch.Tensor): Image tensor with shape (n, c, h ,w).
+
+        Returns:
+            list[torch.Tensor]: Multi-level features that may have
+                different resolutions.
+        """
+        x = self.backbone(img)
+        if self.with_neck:
+            x = self.neck(x)
+        return x
+
+    def forward_dummy(self, img):
+        """Dummy forward function."""
+        x = self.extract_feat(img)
+        rpn_outs = self.rpn_head(x)
+        return rpn_outs
+
+    def forward_train(self,
+                      img,
+                      img_metas,
+                      gt_bboxes=None,
+                      gt_bboxes_ignore=None):
+        """
+        Args:
+            img (Tensor): Input images of shape (N, C, H, W).
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): A List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                :class:`mmdet.datasets.pipelines.Collect`.
+            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        if self.train_cfg.rpn.get('debug', False):
+            self.rpn_head.debug_imgs = tensor2imgs(img)
+
+        x = self.extract_feat(img)
+        losses = self.rpn_head.forward_train(x, img_metas, gt_bboxes, None,
+                                             gt_bboxes_ignore)
+        return losses
+
+    def simple_test(self, img, img_metas, rescale=False):
+        """Test function without test time augmentation.
+
+        Args:
+            imgs (list[torch.Tensor]): List of multiple images
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[np.ndarray]: proposals
+        """
+        x = self.extract_feat(img)
+        proposal_list = self.rpn_head.simple_test_rpn(x, img_metas)
+        if rescale:
+            for proposals, meta in zip(proposal_list, img_metas):
+                proposals[:, :4] /= proposals.new_tensor(meta['scale_factor'])
+
+        return [proposal.cpu().numpy() for proposal in proposal_list]
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            imgs (list[torch.Tensor]): List of multiple images
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[np.ndarray]: proposals
+        """
+        proposal_list = self.rpn_head.aug_test_rpn(
+            self.extract_feats(imgs), img_metas)
+        if not rescale:
+            for proposals, img_meta in zip(proposal_list, img_metas[0]):
+                img_shape = img_meta['img_shape']
+                scale_factor = img_meta['scale_factor']
+                flip = img_meta['flip']
+                flip_direction = img_meta['flip_direction']
+                proposals[:, :4] = bbox_mapping(proposals[:, :4], img_shape,
+                                                scale_factor, flip,
+                                                flip_direction)
+        return [proposal.cpu().numpy() for proposal in proposal_list]
+
+    def show_result(self, data, result, dataset=None, top_k=20):
+        """Show RPN proposals on the image.
+
+        Although we assume batch size is 1, this method supports arbitrary
+        batch size.
+        """
+        img_tensor = data['img'][0]
+        img_metas = data['img_metas'][0].data[0]
+        imgs = tensor2imgs(img_tensor, **img_metas[0]['img_norm_cfg'])
+        assert len(imgs) == len(img_metas)
+        for img, img_meta in zip(imgs, img_metas):
+            h, w, _ = img_meta['img_shape']
+            img_show = img[:h, :w, :]
+            mmcv.imshow_bboxes(img_show, result, top_k=top_k)
diff --git a/detection/scrfd/mmdet/models/detectors/scrfd.py b/detection/scrfd/mmdet/models/detectors/scrfd.py
new file mode 100755
index 000000000..332e0a30c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/scrfd.py
@@ -0,0 +1,100 @@
+from mmdet.core import bbox2result
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+import torch
+
+
+@DETECTORS.register_module()
+class SCRFD(SingleStageDetector):
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(SCRFD, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                  test_cfg, pretrained)
+
+    def forward_train(self,
+                      img,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_keypointss=None,
+                      gt_bboxes_ignore=None):
+        """
+        Args:
+            img (Tensor): Input images of shape (N, C, H, W).
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): A List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                :class:`mmdet.datasets.pipelines.Collect`.
+            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box
+            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        super(SingleStageDetector, self).forward_train(img, img_metas)
+        x = self.extract_feat(img)
+        losses = self.bbox_head.forward_train(x, img_metas, gt_bboxes,
+                                              gt_labels, gt_keypointss, gt_bboxes_ignore)
+        return losses
+
+    def simple_test(self, img, img_metas, rescale=False):
+        """Test function without test time augmentation.
+
+        Args:
+            imgs (list[torch.Tensor]): List of multiple images
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[list[np.ndarray]]: BBox results of each image and classes.
+                The outer list corresponds to each image. The inner list
+                corresponds to each class.
+        """
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x)
+        #print(len(outs))
+        if torch.onnx.is_in_onnx_export():
+            print('single_stage.py in-onnx-export')
+            print(outs.__class__)
+            cls_score, bbox_pred, kps_pred = outs
+            for c in cls_score:
+                print(c.shape)
+            for c in bbox_pred:
+                print(c.shape)
+            #print(outs[0].shape, outs[1].shape)
+            if self.bbox_head.use_kps:
+                for c in kps_pred:
+                    print(c.shape)
+                return (cls_score, bbox_pred, kps_pred)
+            else:
+                return (cls_score, bbox_pred)
+            #return outs
+        bbox_list = self.bbox_head.get_bboxes(
+            *outs, img_metas, rescale=rescale)
+        # skip post-processing when exporting to ONNX
+        #if torch.onnx.is_in_onnx_export():
+        #    return bbox_list
+
+        bbox_results = [
+            bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
+            for det_bboxes, det_labels in bbox_list
+        ]
+        return bbox_results
+
+    def feature_test(self, img):
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x)
+        return outs
+
diff --git a/detection/scrfd/mmdet/models/detectors/single_stage.py b/detection/scrfd/mmdet/models/detectors/single_stage.py
new file mode 100755
index 000000000..6ed6edece
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/single_stage.py
@@ -0,0 +1,160 @@
+import torch
+import torch.nn as nn
+
+from mmdet.core import bbox2result
+from ..builder import DETECTORS, build_backbone, build_head, build_neck
+from .base import BaseDetector
+
+
+@DETECTORS.register_module()
+class SingleStageDetector(BaseDetector):
+    """Base class for single-stage detectors.
+
+    Single-stage detectors directly and densely predict bounding boxes on the
+    output features of the backbone+neck.
+    """
+
+    def __init__(self,
+                 backbone,
+                 neck=None,
+                 bbox_head=None,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(SingleStageDetector, self).__init__()
+        self.backbone = build_backbone(backbone)
+        if neck is not None:
+            self.neck = build_neck(neck)
+        bbox_head.update(train_cfg=train_cfg)
+        bbox_head.update(test_cfg=test_cfg)
+        self.bbox_head = build_head(bbox_head)
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        self.init_weights(pretrained=pretrained)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in detector.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(SingleStageDetector, self).init_weights(pretrained)
+        self.backbone.init_weights(pretrained=pretrained)
+        if self.with_neck:
+            if isinstance(self.neck, nn.Sequential):
+                for m in self.neck:
+                    m.init_weights()
+            else:
+                self.neck.init_weights()
+        self.bbox_head.init_weights()
+
+    def extract_feat(self, img):
+        """Directly extract features from the backbone+neck."""
+        x = self.backbone(img)
+        if self.with_neck:
+            x = self.neck(x)
+        return x
+
+    def forward_dummy(self, img):
+        """Used for computing network flops.
+
+        See `mmdetection/tools/get_flops.py`
+        """
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x)
+        return outs
+
+    def forward_train(self,
+                      img,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None):
+        """
+        Args:
+            img (Tensor): Input images of shape (N, C, H, W).
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): A List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                :class:`mmdet.datasets.pipelines.Collect`.
+            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box
+            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
+                boxes can be ignored when computing the loss.
+
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        super(SingleStageDetector, self).forward_train(img, img_metas)
+        x = self.extract_feat(img)
+        losses = self.bbox_head.forward_train(x, img_metas, gt_bboxes,
+                                              gt_labels, gt_bboxes_ignore)
+        return losses
+
+    def simple_test(self, img, img_metas, rescale=False):
+        """Test function without test time augmentation.
+
+        Args:
+            imgs (list[torch.Tensor]): List of multiple images
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[list[np.ndarray]]: BBox results of each image and classes.
+                The outer list corresponds to each image. The inner list
+                corresponds to each class.
+        """
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x)
+        #print(len(outs))
+        if torch.onnx.is_in_onnx_export():
+            print('single_stage.py in-onnx-export')
+            print(outs.__class__)
+            cls_score, bbox_pred = outs
+            for c in cls_score:
+                print(c.shape)
+            for c in bbox_pred:
+                print(c.shape)
+            #print(outs[0].shape, outs[1].shape)
+            return outs
+        bbox_list = self.bbox_head.get_bboxes(
+            *outs, img_metas, rescale=rescale)
+        # skip post-processing when exporting to ONNX
+        if torch.onnx.is_in_onnx_export():
+            return bbox_list
+
+        bbox_results = [
+            bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
+            for det_bboxes, det_labels in bbox_list
+        ]
+        return bbox_results
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Test function with test time augmentation.
+
+        Args:
+            imgs (list[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains all images in the batch.
+            img_metas (list[list[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch. each dict has image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+
+        Returns:
+            list[list[np.ndarray]]: BBox results of each image and classes.
+                The outer list corresponds to each image. The inner list
+                corresponds to each class.
+        """
+        assert hasattr(self.bbox_head, 'aug_test'), \
+            f'{self.bbox_head.__class__.__name__}' \
+            ' does not support test-time augmentation'
+        print('aug-test:', len(imgs))
+        feats = self.extract_feats(imgs)
+        return [self.bbox_head.aug_test(feats, img_metas, rescale=rescale)]
diff --git a/detection/scrfd/mmdet/models/detectors/trident_faster_rcnn.py b/detection/scrfd/mmdet/models/detectors/trident_faster_rcnn.py
new file mode 100755
index 000000000..f0fd80d41
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/trident_faster_rcnn.py
@@ -0,0 +1,66 @@
+from ..builder import DETECTORS
+from .faster_rcnn import FasterRCNN
+
+
+@DETECTORS.register_module()
+class TridentFasterRCNN(FasterRCNN):
+    """Implementation of `TridentNet <https://arxiv.org/abs/1901.01892>`_"""
+
+    def __init__(self,
+                 backbone,
+                 rpn_head,
+                 roi_head,
+                 train_cfg,
+                 test_cfg,
+                 neck=None,
+                 pretrained=None):
+
+        super(TridentFasterRCNN, self).__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            pretrained=pretrained)
+        assert self.backbone.num_branch == self.roi_head.num_branch
+        assert self.backbone.test_branch_idx == self.roi_head.test_branch_idx
+        self.num_branch = self.backbone.num_branch
+        self.test_branch_idx = self.backbone.test_branch_idx
+
+    def simple_test(self, img, img_metas, proposals=None, rescale=False):
+        """Test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+        x = self.extract_feat(img)
+        if proposals is None:
+            num_branch = (self.num_branch if self.test_branch_idx == -1 else 1)
+            trident_img_metas = img_metas * num_branch
+            proposal_list = self.rpn_head.simple_test_rpn(x, trident_img_metas)
+        else:
+            proposal_list = proposals
+
+        return self.roi_head.simple_test(
+            x, proposal_list, trident_img_metas, rescale=rescale)
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        x = self.extract_feats(imgs)
+        num_branch = (self.num_branch if self.test_branch_idx == -1 else 1)
+        trident_img_metas = [img_metas * num_branch for img_metas in img_metas]
+        proposal_list = self.rpn_head.aug_test_rpn(x, trident_img_metas)
+        return self.roi_head.aug_test(
+            x, proposal_list, img_metas, rescale=rescale)
+
+    def forward_train(self, img, img_metas, gt_bboxes, gt_labels, **kwargs):
+        """make copies of img and gts to fit multi-branch."""
+        trident_gt_bboxes = tuple(gt_bboxes * self.num_branch)
+        trident_gt_labels = tuple(gt_labels * self.num_branch)
+        trident_img_metas = tuple(img_metas * self.num_branch)
+
+        return super(TridentFasterRCNN,
+                     self).forward_train(img, trident_img_metas,
+                                         trident_gt_bboxes, trident_gt_labels)
diff --git a/detection/scrfd/mmdet/models/detectors/two_stage.py b/detection/scrfd/mmdet/models/detectors/two_stage.py
new file mode 100755
index 000000000..d66923f32
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/two_stage.py
@@ -0,0 +1,210 @@
+import torch
+import torch.nn as nn
+
+# from mmdet.core import bbox2result, bbox2roi, build_assigner, build_sampler
+from ..builder import DETECTORS, build_backbone, build_head, build_neck
+from .base import BaseDetector
+
+
+@DETECTORS.register_module()
+class TwoStageDetector(BaseDetector):
+    """Base class for two-stage detectors.
+
+    Two-stage detectors typically consisting of a region proposal network and a
+    task-specific regression head.
+    """
+
+    def __init__(self,
+                 backbone,
+                 neck=None,
+                 rpn_head=None,
+                 roi_head=None,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(TwoStageDetector, self).__init__()
+        self.backbone = build_backbone(backbone)
+
+        if neck is not None:
+            self.neck = build_neck(neck)
+
+        if rpn_head is not None:
+            rpn_train_cfg = train_cfg.rpn if train_cfg is not None else None
+            rpn_head_ = rpn_head.copy()
+            rpn_head_.update(train_cfg=rpn_train_cfg, test_cfg=test_cfg.rpn)
+            self.rpn_head = build_head(rpn_head_)
+
+        if roi_head is not None:
+            # update train and test cfg here for now
+            # TODO: refactor assigner & sampler
+            rcnn_train_cfg = train_cfg.rcnn if train_cfg is not None else None
+            roi_head.update(train_cfg=rcnn_train_cfg)
+            roi_head.update(test_cfg=test_cfg.rcnn)
+            self.roi_head = build_head(roi_head)
+
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+
+        self.init_weights(pretrained=pretrained)
+
+    @property
+    def with_rpn(self):
+        """bool: whether the detector has RPN"""
+        return hasattr(self, 'rpn_head') and self.rpn_head is not None
+
+    @property
+    def with_roi_head(self):
+        """bool: whether the detector has a RoI head"""
+        return hasattr(self, 'roi_head') and self.roi_head is not None
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in detector.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(TwoStageDetector, self).init_weights(pretrained)
+        self.backbone.init_weights(pretrained=pretrained)
+        if self.with_neck:
+            if isinstance(self.neck, nn.Sequential):
+                for m in self.neck:
+                    m.init_weights()
+            else:
+                self.neck.init_weights()
+        if self.with_rpn:
+            self.rpn_head.init_weights()
+        if self.with_roi_head:
+            self.roi_head.init_weights(pretrained)
+
+    def extract_feat(self, img):
+        """Directly extract features from the backbone+neck."""
+        x = self.backbone(img)
+        if self.with_neck:
+            x = self.neck(x)
+        return x
+
+    def forward_dummy(self, img):
+        """Used for computing network flops.
+
+        See `mmdetection/tools/get_flops.py`
+        """
+        outs = ()
+        # backbone
+        x = self.extract_feat(img)
+        # rpn
+        if self.with_rpn:
+            rpn_outs = self.rpn_head(x)
+            outs = outs + (rpn_outs, )
+        proposals = torch.randn(1000, 4).to(img.device)
+        # roi_head
+        roi_outs = self.roi_head.forward_dummy(x, proposals)
+        outs = outs + (roi_outs, )
+        return outs
+
+    def forward_train(self,
+                      img,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None,
+                      proposals=None,
+                      **kwargs):
+        """
+        Args:
+            img (Tensor): of shape (N, C, H, W) encoding input images.
+                Typically these should be mean centered and std scaled.
+
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+
+            gt_labels (list[Tensor]): class indices corresponding to each box
+
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+            gt_masks (None | Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+            proposals : override rpn proposals with custom proposals. Use when
+                `with_rpn` is False.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        x = self.extract_feat(img)
+
+        losses = dict()
+
+        # RPN forward and loss
+        if self.with_rpn:
+            proposal_cfg = self.train_cfg.get('rpn_proposal',
+                                              self.test_cfg.rpn)
+            rpn_losses, proposal_list = self.rpn_head.forward_train(
+                x,
+                img_metas,
+                gt_bboxes,
+                gt_labels=None,
+                gt_bboxes_ignore=gt_bboxes_ignore,
+                proposal_cfg=proposal_cfg)
+            losses.update(rpn_losses)
+        else:
+            proposal_list = proposals
+
+        roi_losses = self.roi_head.forward_train(x, img_metas, proposal_list,
+                                                 gt_bboxes, gt_labels,
+                                                 gt_bboxes_ignore, gt_masks,
+                                                 **kwargs)
+        losses.update(roi_losses)
+
+        return losses
+
+    async def async_simple_test(self,
+                                img,
+                                img_meta,
+                                proposals=None,
+                                rescale=False):
+        """Async test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+        x = self.extract_feat(img)
+
+        if proposals is None:
+            proposal_list = await self.rpn_head.async_simple_test_rpn(
+                x, img_meta)
+        else:
+            proposal_list = proposals
+
+        return await self.roi_head.async_simple_test(
+            x, proposal_list, img_meta, rescale=rescale)
+
+    def simple_test(self, img, img_metas, proposals=None, rescale=False):
+        """Test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+
+        x = self.extract_feat(img)
+
+        if proposals is None:
+            proposal_list = self.rpn_head.simple_test_rpn(x, img_metas)
+        else:
+            proposal_list = proposals
+
+        return self.roi_head.simple_test(
+            x, proposal_list, img_metas, rescale=rescale)
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        x = self.extract_feats(imgs)
+        proposal_list = self.rpn_head.aug_test_rpn(x, img_metas)
+        return self.roi_head.aug_test(
+            x, proposal_list, img_metas, rescale=rescale)
diff --git a/detection/scrfd/mmdet/models/detectors/vfnet.py b/detection/scrfd/mmdet/models/detectors/vfnet.py
new file mode 100755
index 000000000..e23f89674
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/vfnet.py
@@ -0,0 +1,18 @@
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class VFNet(SingleStageDetector):
+    """Implementation of `VarifocalNet
+    (VFNet).<https://arxiv.org/abs/2008.13367>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(VFNet, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                    test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/detectors/yolact.py b/detection/scrfd/mmdet/models/detectors/yolact.py
new file mode 100755
index 000000000..5f6c7ff09
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/yolact.py
@@ -0,0 +1,146 @@
+import torch
+
+from mmdet.core import bbox2result
+from ..builder import DETECTORS, build_head
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class YOLACT(SingleStageDetector):
+    """Implementation of `YOLACT <https://arxiv.org/abs/1904.02689>`_"""
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 segm_head,
+                 mask_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(YOLACT, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                     test_cfg, pretrained)
+        self.segm_head = build_head(segm_head)
+        self.mask_head = build_head(mask_head)
+        self.init_segm_mask_weights()
+
+    def init_segm_mask_weights(self):
+        """Initialize weights of the YOLACT semg head and YOLACT mask head."""
+        self.segm_head.init_weights()
+        self.mask_head.init_weights()
+
+    def forward_dummy(self, img):
+        """Used for computing network flops.
+
+        See `mmdetection/tools/get_flops.py`
+        """
+        raise NotImplementedError
+
+    def forward_train(self,
+                      img,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None):
+        """
+        Args:
+            img (Tensor): of shape (N, C, H, W) encoding input images.
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+            gt_masks (None | Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        # convert Bitmap mask or Polygon Mask to Tensor here
+        gt_masks = [
+            gt_mask.to_tensor(dtype=torch.uint8, device=img.device)
+            for gt_mask in gt_masks
+        ]
+
+        x = self.extract_feat(img)
+
+        cls_score, bbox_pred, coeff_pred = self.bbox_head(x)
+        bbox_head_loss_inputs = (cls_score, bbox_pred) + (gt_bboxes, gt_labels,
+                                                          img_metas)
+        losses, sampling_results = self.bbox_head.loss(
+            *bbox_head_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
+
+        segm_head_outs = self.segm_head(x[0])
+        loss_segm = self.segm_head.loss(segm_head_outs, gt_masks, gt_labels)
+        losses.update(loss_segm)
+
+        mask_pred = self.mask_head(x[0], coeff_pred, gt_bboxes, img_metas,
+                                   sampling_results)
+        loss_mask = self.mask_head.loss(mask_pred, gt_masks, gt_bboxes,
+                                        img_metas, sampling_results)
+        losses.update(loss_mask)
+
+        # check NaN and Inf
+        for loss_name in losses.keys():
+            assert torch.isfinite(torch.stack(losses[loss_name]))\
+                .all().item(), '{} becomes infinite or NaN!'\
+                .format(loss_name)
+
+        return losses
+
+    def simple_test(self, img, img_metas, rescale=False):
+        """Test function without test time augmentation."""
+        x = self.extract_feat(img)
+
+        cls_score, bbox_pred, coeff_pred = self.bbox_head(x)
+
+        bbox_inputs = (cls_score, bbox_pred,
+                       coeff_pred) + (img_metas, self.test_cfg, rescale)
+        det_bboxes, det_labels, det_coeffs = self.bbox_head.get_bboxes(
+            *bbox_inputs)
+        bbox_results = [
+            bbox2result(det_bbox, det_label, self.bbox_head.num_classes)
+            for det_bbox, det_label in zip(det_bboxes, det_labels)
+        ]
+
+        num_imgs = len(img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+        if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+            segm_results = [[[] for _ in range(self.mask_head.num_classes)]
+                            for _ in range(num_imgs)]
+        else:
+            # if det_bboxes is rescaled to the original image size, we need to
+            # rescale it back to the testing scale to obtain RoIs.
+            if rescale and not isinstance(scale_factors[0], float):
+                scale_factors = [
+                    torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                    for scale_factor in scale_factors
+                ]
+            _bboxes = [
+                det_bboxes[i][:, :4] *
+                scale_factors[i] if rescale else det_bboxes[i][:, :4]
+                for i in range(len(det_bboxes))
+            ]
+            mask_preds = self.mask_head(x[0], det_coeffs, _bboxes, img_metas)
+            # apply mask post-processing to each image individually
+            segm_results = []
+            for i in range(num_imgs):
+                if det_bboxes[i].shape[0] == 0:
+                    segm_results.append(
+                        [[] for _ in range(self.mask_head.num_classes)])
+                else:
+                    segm_result = self.mask_head.get_seg_masks(
+                        mask_preds[i], det_labels[i], img_metas[i], rescale)
+                    segm_results.append(segm_result)
+        return list(zip(bbox_results, segm_results))
+
+    def aug_test(self, imgs, img_metas, rescale=False):
+        """Test with augmentations."""
+        raise NotImplementedError
diff --git a/detection/scrfd/mmdet/models/detectors/yolo.py b/detection/scrfd/mmdet/models/detectors/yolo.py
new file mode 100755
index 000000000..240aab20f
--- /dev/null
+++ b/detection/scrfd/mmdet/models/detectors/yolo.py
@@ -0,0 +1,18 @@
+# Copyright (c) 2019 Western Digital Corporation or its affiliates.
+
+from ..builder import DETECTORS
+from .single_stage import SingleStageDetector
+
+
+@DETECTORS.register_module()
+class YOLOV3(SingleStageDetector):
+
+    def __init__(self,
+                 backbone,
+                 neck,
+                 bbox_head,
+                 train_cfg=None,
+                 test_cfg=None,
+                 pretrained=None):
+        super(YOLOV3, self).__init__(backbone, neck, bbox_head, train_cfg,
+                                     test_cfg, pretrained)
diff --git a/detection/scrfd/mmdet/models/losses/__init__.py b/detection/scrfd/mmdet/models/losses/__init__.py
new file mode 100755
index 000000000..bb887d373
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/__init__.py
@@ -0,0 +1,28 @@
+from .accuracy import Accuracy, accuracy
+from .ae_loss import AssociativeEmbeddingLoss
+from .balanced_l1_loss import BalancedL1Loss, balanced_l1_loss
+from .cross_entropy_loss import (CrossEntropyLoss, binary_cross_entropy,
+                                 cross_entropy, mask_cross_entropy)
+from .focal_loss import FocalLoss, sigmoid_focal_loss
+from .gaussian_focal_loss import GaussianFocalLoss
+from .gfocal_loss import DistributionFocalLoss, QualityFocalLoss
+from .ghm_loss import GHMC, GHMR
+from .iou_loss import (BoundedIoULoss, CIoULoss, DIoULoss, GIoULoss, IoULoss,
+                       bounded_iou_loss, iou_loss)
+from .mse_loss import MSELoss, mse_loss
+from .pisa_loss import carl_loss, isr_p
+from .smooth_l1_loss import L1Loss, SmoothL1Loss, l1_loss, smooth_l1_loss
+from .utils import reduce_loss, weight_reduce_loss, weighted_loss
+from .varifocal_loss import VarifocalLoss
+
+__all__ = [
+    'accuracy', 'Accuracy', 'cross_entropy', 'binary_cross_entropy',
+    'mask_cross_entropy', 'CrossEntropyLoss', 'sigmoid_focal_loss',
+    'FocalLoss', 'smooth_l1_loss', 'SmoothL1Loss', 'balanced_l1_loss',
+    'BalancedL1Loss', 'mse_loss', 'MSELoss', 'iou_loss', 'bounded_iou_loss',
+    'IoULoss', 'BoundedIoULoss', 'GIoULoss', 'DIoULoss', 'CIoULoss', 'GHMC',
+    'GHMR', 'reduce_loss', 'weight_reduce_loss', 'weighted_loss', 'L1Loss',
+    'l1_loss', 'isr_p', 'carl_loss', 'AssociativeEmbeddingLoss',
+    'GaussianFocalLoss', 'QualityFocalLoss', 'DistributionFocalLoss',
+    'VarifocalLoss'
+]
diff --git a/detection/scrfd/mmdet/models/losses/accuracy.py b/detection/scrfd/mmdet/models/losses/accuracy.py
new file mode 100755
index 000000000..924ebbed7
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/accuracy.py
@@ -0,0 +1,76 @@
+import torch.nn as nn
+
+
+def accuracy(pred, target, topk=1, thresh=None):
+    """Calculate accuracy according to the prediction and target.
+
+    Args:
+        pred (torch.Tensor): The model prediction, shape (N, num_class)
+        target (torch.Tensor): The target of each prediction, shape (N, )
+        topk (int | tuple[int], optional): If the predictions in ``topk``
+            matches the target, the predictions will be regarded as
+            correct ones. Defaults to 1.
+        thresh (float, optional): If not None, predictions with scores under
+            this threshold are considered incorrect. Default to None.
+
+    Returns:
+        float | tuple[float]: If the input ``topk`` is a single integer,
+            the function will return a single float as accuracy. If
+            ``topk`` is a tuple containing multiple integers, the
+            function will return a tuple containing accuracies of
+            each ``topk`` number.
+    """
+    assert isinstance(topk, (int, tuple))
+    if isinstance(topk, int):
+        topk = (topk, )
+        return_single = True
+    else:
+        return_single = False
+
+    maxk = max(topk)
+    if pred.size(0) == 0:
+        accu = [pred.new_tensor(0.) for i in range(len(topk))]
+        return accu[0] if return_single else accu
+    assert pred.ndim == 2 and target.ndim == 1
+    assert pred.size(0) == target.size(0)
+    assert maxk <= pred.size(1), \
+        f'maxk {maxk} exceeds pred dimension {pred.size(1)}'
+    pred_value, pred_label = pred.topk(maxk, dim=1)
+    pred_label = pred_label.t()  # transpose to shape (maxk, N)
+    correct = pred_label.eq(target.view(1, -1).expand_as(pred_label))
+    if thresh is not None:
+        # Only prediction values larger than thresh are counted as correct
+        correct = correct & (pred_value > thresh).t()
+    res = []
+    for k in topk:
+        correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
+        res.append(correct_k.mul_(100.0 / pred.size(0)))
+    return res[0] if return_single else res
+
+
+class Accuracy(nn.Module):
+
+    def __init__(self, topk=(1, ), thresh=None):
+        """Module to calculate the accuracy.
+
+        Args:
+            topk (tuple, optional): The criterion used to calculate the
+                accuracy. Defaults to (1,).
+            thresh (float, optional): If not None, predictions with scores
+                under this threshold are considered incorrect. Default to None.
+        """
+        super().__init__()
+        self.topk = topk
+        self.thresh = thresh
+
+    def forward(self, pred, target):
+        """Forward function to calculate accuracy.
+
+        Args:
+            pred (torch.Tensor): Prediction of models.
+            target (torch.Tensor): Target for each prediction.
+
+        Returns:
+            tuple[float]: The accuracies under different topk criterions.
+        """
+        return accuracy(pred, target, self.topk, self.thresh)
diff --git a/detection/scrfd/mmdet/models/losses/ae_loss.py b/detection/scrfd/mmdet/models/losses/ae_loss.py
new file mode 100755
index 000000000..6077652cf
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/ae_loss.py
@@ -0,0 +1,100 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+
+
+def ae_loss_per_image(tl_preds, br_preds, match):
+    """Associative Embedding Loss in one image.
+
+    Associative Embedding Loss including two parts: pull loss and push loss.
+    Pull loss makes embedding vectors from same object closer to each other.
+    Push loss distinguish embedding vector from different objects, and makes
+        the gap between them is large enough.
+
+    During computing, usually there are 3 cases:
+        - no object in image: both pull loss and push loss will be 0.
+        - one object in image: push loss will be 0 and pull loss is computed
+            by the two corner of the only object.
+        - more than one objects in image: pull loss is computed by corner pairs
+            from each object, push loss is computed by each object with all
+            other objects. We use confusion matrix with 0 in diagonal to
+            compute the push loss.
+
+    Args:
+        tl_preds (tensor): Embedding feature map of left-top corner.
+        br_preds (tensor): Embedding feature map of bottim-right corner.
+        match (list): Downsampled coordinates pair of each ground truth box.
+    """
+
+    tl_list, br_list, me_list = [], [], []
+    if len(match) == 0:  # no object in image
+        pull_loss = tl_preds.sum() * 0.
+        push_loss = tl_preds.sum() * 0.
+    else:
+        for m in match:
+            [tl_y, tl_x], [br_y, br_x] = m
+            tl_e = tl_preds[:, tl_y, tl_x].view(-1, 1)
+            br_e = br_preds[:, br_y, br_x].view(-1, 1)
+            tl_list.append(tl_e)
+            br_list.append(br_e)
+            me_list.append((tl_e + br_e) / 2.0)
+
+        tl_list = torch.cat(tl_list)
+        br_list = torch.cat(br_list)
+        me_list = torch.cat(me_list)
+
+        assert tl_list.size() == br_list.size()
+
+        # N is object number in image, M is dimension of embedding vector
+        N, M = tl_list.size()
+
+        pull_loss = (tl_list - me_list).pow(2) + (br_list - me_list).pow(2)
+        pull_loss = pull_loss.sum() / N
+
+        margin = 1  # exp setting of CornerNet, details in section 3.3 of paper
+
+        # confusion matrix of push loss
+        conf_mat = me_list.expand((N, N, M)).permute(1, 0, 2) - me_list
+        conf_weight = 1 - torch.eye(N).type_as(me_list)
+        conf_mat = conf_weight * (margin - conf_mat.sum(-1).abs())
+
+        if N > 1:  # more than one object in current image
+            push_loss = F.relu(conf_mat).sum() / (N * (N - 1))
+        else:
+            push_loss = tl_preds.sum() * 0.
+
+    return pull_loss, push_loss
+
+
+@LOSSES.register_module()
+class AssociativeEmbeddingLoss(nn.Module):
+    """Associative Embedding Loss.
+
+    More details can be found in
+    `Associative Embedding <https://arxiv.org/abs/1611.05424>`_ and
+    `CornerNet <https://arxiv.org/abs/1808.01244>`_ .
+    Code is modified from `kp_utils.py <https://github.com/princeton-vl/CornerNet/blob/master/models/py_utils/kp_utils.py#L180>`_  # noqa: E501
+
+    Args:
+        pull_weight (float): Loss weight for corners from same object.
+        push_weight (float): Loss weight for corners from different object.
+    """
+
+    def __init__(self, pull_weight=0.25, push_weight=0.25):
+        super(AssociativeEmbeddingLoss, self).__init__()
+        self.pull_weight = pull_weight
+        self.push_weight = push_weight
+
+    def forward(self, pred, target, match):
+        """Forward function."""
+        batch = pred.size(0)
+        pull_all, push_all = 0.0, 0.0
+        for i in range(batch):
+            pull, push = ae_loss_per_image(pred[i], target[i], match[i])
+
+            pull_all += self.pull_weight * pull
+            push_all += self.push_weight * push
+
+        return pull_all, push_all
diff --git a/detection/scrfd/mmdet/models/losses/balanced_l1_loss.py b/detection/scrfd/mmdet/models/losses/balanced_l1_loss.py
new file mode 100755
index 000000000..3790a80b8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/balanced_l1_loss.py
@@ -0,0 +1,118 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def balanced_l1_loss(pred,
+                     target,
+                     beta=1.0,
+                     alpha=0.5,
+                     gamma=1.5,
+                     reduction='mean'):
+    """Calculate balanced L1 loss.
+
+    Please see the `Libra R-CNN <https://arxiv.org/pdf/1904.02701.pdf>`_
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, 4).
+        target (torch.Tensor): The learning target of the prediction with
+            shape (N, 4).
+        beta (float): The loss is a piecewise function of prediction and target
+            and ``beta`` serves as a threshold for the difference between the
+            prediction and target. Defaults to 1.0.
+        alpha (float): The denominator ``alpha`` in the balanced L1 loss.
+            Defaults to 0.5.
+        gamma (float): The ``gamma`` in the balanced L1 loss.
+            Defaults to 1.5.
+        reduction (str, optional): The method that reduces the loss to a
+            scalar. Options are "none", "mean" and "sum".
+
+    Returns:
+        torch.Tensor: The calculated loss
+    """
+    assert beta > 0
+    assert pred.size() == target.size() and target.numel() > 0
+
+    diff = torch.abs(pred - target)
+    b = np.e**(gamma / alpha) - 1
+    loss = torch.where(
+        diff < beta, alpha / b *
+        (b * diff + 1) * torch.log(b * diff / beta + 1) - alpha * diff,
+        gamma * diff + gamma / b - alpha * beta)
+
+    return loss
+
+
+@LOSSES.register_module()
+class BalancedL1Loss(nn.Module):
+    """Balanced L1 Loss.
+
+    arXiv: https://arxiv.org/pdf/1904.02701.pdf (CVPR 2019)
+
+    Args:
+        alpha (float): The denominator ``alpha`` in the balanced L1 loss.
+            Defaults to 0.5.
+        gamma (float): The ``gamma`` in the balanced L1 loss. Defaults to 1.5.
+        beta (float, optional): The loss is a piecewise function of prediction
+            and target. ``beta`` serves as a threshold for the difference
+            between the prediction and target. Defaults to 1.0.
+        reduction (str, optional): The method that reduces the loss to a
+            scalar. Options are "none", "mean" and "sum".
+        loss_weight (float, optional): The weight of the loss. Defaults to 1.0
+    """
+
+    def __init__(self,
+                 alpha=0.5,
+                 gamma=1.5,
+                 beta=1.0,
+                 reduction='mean',
+                 loss_weight=1.0):
+        super(BalancedL1Loss, self).__init__()
+        self.alpha = alpha
+        self.gamma = gamma
+        self.beta = beta
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        """Forward function of loss.
+
+        Args:
+            pred (torch.Tensor): The prediction with shape (N, 4).
+            target (torch.Tensor): The learning target of the prediction with
+                shape (N, 4).
+            weight (torch.Tensor, optional): Sample-wise loss weight with
+                shape (N, ).
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Options are "none", "mean" and "sum".
+
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_bbox = self.loss_weight * balanced_l1_loss(
+            pred,
+            target,
+            weight,
+            alpha=self.alpha,
+            gamma=self.gamma,
+            beta=self.beta,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss_bbox
diff --git a/detection/scrfd/mmdet/models/losses/cross_entropy_loss.py b/detection/scrfd/mmdet/models/losses/cross_entropy_loss.py
new file mode 100755
index 000000000..1f283dd8d
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/cross_entropy_loss.py
@@ -0,0 +1,201 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+from .utils import weight_reduce_loss
+
+
+def cross_entropy(pred,
+                  label,
+                  weight=None,
+                  reduction='mean',
+                  avg_factor=None,
+                  class_weight=None):
+    """Calculate the CrossEntropy loss.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the number
+            of classes.
+        label (torch.Tensor): The learning label of the prediction.
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        reduction (str, optional): The method used to reduce the loss.
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+
+    Returns:
+        torch.Tensor: The calculated loss
+    """
+    # element-wise losses
+    loss = F.cross_entropy(pred, label, weight=class_weight, reduction='none')
+
+    # apply weights and do the reduction
+    if weight is not None:
+        weight = weight.float()
+    loss = weight_reduce_loss(
+        loss, weight=weight, reduction=reduction, avg_factor=avg_factor)
+
+    return loss
+
+
+def _expand_onehot_labels(labels, label_weights, label_channels):
+    bin_labels = labels.new_full((labels.size(0), label_channels), 0)
+    inds = torch.nonzero(
+        (labels >= 0) & (labels < label_channels), as_tuple=False).squeeze()
+    if inds.numel() > 0:
+        bin_labels[inds, labels[inds]] = 1
+
+    if label_weights is None:
+        bin_label_weights = None
+    else:
+        bin_label_weights = label_weights.view(-1, 1).expand(
+            label_weights.size(0), label_channels)
+
+    return bin_labels, bin_label_weights
+
+
+def binary_cross_entropy(pred,
+                         label,
+                         weight=None,
+                         reduction='mean',
+                         avg_factor=None,
+                         class_weight=None):
+    """Calculate the binary CrossEntropy loss.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, 1).
+        label (torch.Tensor): The learning label of the prediction.
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        reduction (str, optional): The method used to reduce the loss.
+            Options are "none", "mean" and "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+
+    Returns:
+        torch.Tensor: The calculated loss
+    """
+    if pred.dim() != label.dim():
+        label, weight = _expand_onehot_labels(label, weight, pred.size(-1))
+
+    # weighted element-wise losses
+    if weight is not None:
+        weight = weight.float()
+    loss = F.binary_cross_entropy_with_logits(
+        pred, label.float(), pos_weight=class_weight, reduction='none')
+    # do the reduction for the weighted loss
+    loss = weight_reduce_loss(
+        loss, weight, reduction=reduction, avg_factor=avg_factor)
+
+    return loss
+
+
+def mask_cross_entropy(pred,
+                       target,
+                       label,
+                       reduction='mean',
+                       avg_factor=None,
+                       class_weight=None):
+    """Calculate the CrossEntropy loss for masks.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the number
+            of classes.
+        target (torch.Tensor): The learning label of the prediction.
+        label (torch.Tensor): ``label`` indicates the class label of the mask'
+            corresponding object. This will be used to select the mask in the
+            of the class which the object belongs to when the mask prediction
+            if not class-agnostic.
+        reduction (str, optional): The method used to reduce the loss.
+            Options are "none", "mean" and "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+
+    Returns:
+        torch.Tensor: The calculated loss
+    """
+    # TODO: handle these two reserved arguments
+    assert reduction == 'mean' and avg_factor is None
+    num_rois = pred.size()[0]
+    inds = torch.arange(0, num_rois, dtype=torch.long, device=pred.device)
+    pred_slice = pred[inds, label].squeeze(1)
+    return F.binary_cross_entropy_with_logits(
+        pred_slice, target, weight=class_weight, reduction='mean')[None]
+
+
+@LOSSES.register_module()
+class CrossEntropyLoss(nn.Module):
+
+    def __init__(self,
+                 use_sigmoid=False,
+                 use_mask=False,
+                 reduction='mean',
+                 class_weight=None,
+                 loss_weight=1.0):
+        """CrossEntropyLoss.
+
+        Args:
+            use_sigmoid (bool, optional): Whether the prediction uses sigmoid
+                of softmax. Defaults to False.
+            use_mask (bool, optional): Whether to use mask cross entropy loss.
+                Defaults to False.
+            reduction (str, optional): . Defaults to 'mean'.
+                Options are "none", "mean" and "sum".
+            class_weight (list[float], optional): Weight of each class.
+                Defaults to None.
+            loss_weight (float, optional): Weight of the loss. Defaults to 1.0.
+        """
+        super(CrossEntropyLoss, self).__init__()
+        assert (use_sigmoid is False) or (use_mask is False)
+        self.use_sigmoid = use_sigmoid
+        self.use_mask = use_mask
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+        self.class_weight = class_weight
+
+        if self.use_sigmoid:
+            self.cls_criterion = binary_cross_entropy
+        elif self.use_mask:
+            self.cls_criterion = mask_cross_entropy
+        else:
+            self.cls_criterion = cross_entropy
+
+    def forward(self,
+                cls_score,
+                label,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        """Forward function.
+
+        Args:
+            cls_score (torch.Tensor): The prediction.
+            label (torch.Tensor): The learning label of the prediction.
+            weight (torch.Tensor, optional): Sample-wise loss weight.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction (str, optional): The method used to reduce the loss.
+                Options are "none", "mean" and "sum".
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if self.class_weight is not None:
+            class_weight = cls_score.new_tensor(
+                self.class_weight, device=cls_score.device)
+        else:
+            class_weight = None
+        loss_cls = self.loss_weight * self.cls_criterion(
+            cls_score,
+            label,
+            weight,
+            class_weight=class_weight,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss_cls
diff --git a/detection/scrfd/mmdet/models/losses/focal_loss.py b/detection/scrfd/mmdet/models/losses/focal_loss.py
new file mode 100755
index 000000000..be252e579
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/focal_loss.py
@@ -0,0 +1,157 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.ops import sigmoid_focal_loss as _sigmoid_focal_loss
+
+from ..builder import LOSSES
+from .utils import weight_reduce_loss
+
+
+# This method is only for debugging
+def py_sigmoid_focal_loss(pred,
+                          target,
+                          weight=None,
+                          gamma=2.0,
+                          alpha=0.25,
+                          reduction='mean',
+                          avg_factor=None):
+    """PyTorch version of `Focal Loss <https://arxiv.org/abs/1708.02002>`_.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the
+            number of classes
+        target (torch.Tensor): The learning label of the prediction.
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        gamma (float, optional): The gamma for calculating the modulating
+            factor. Defaults to 2.0.
+        alpha (float, optional): A balanced form for Focal Loss.
+            Defaults to 0.25.
+        reduction (str, optional): The method used to reduce the loss into
+            a scalar. Defaults to 'mean'.
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+    """
+    pred_sigmoid = pred.sigmoid()
+    target = target.type_as(pred)
+    pt = (1 - pred_sigmoid) * target + pred_sigmoid * (1 - target)
+    focal_weight = (alpha * target + (1 - alpha) *
+                    (1 - target)) * pt.pow(gamma)
+    loss = F.binary_cross_entropy_with_logits(
+        pred, target, reduction='none') * focal_weight
+    loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
+    return loss
+
+
+def sigmoid_focal_loss(pred,
+                       target,
+                       weight=None,
+                       gamma=2.0,
+                       alpha=0.25,
+                       reduction='mean',
+                       avg_factor=None):
+    r"""A warpper of cuda version `Focal Loss
+    <https://arxiv.org/abs/1708.02002>`_.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the number
+            of classes.
+        target (torch.Tensor): The learning label of the prediction.
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        gamma (float, optional): The gamma for calculating the modulating
+            factor. Defaults to 2.0.
+        alpha (float, optional): A balanced form for Focal Loss.
+            Defaults to 0.25.
+        reduction (str, optional): The method used to reduce the loss into
+            a scalar. Defaults to 'mean'. Options are "none", "mean" and "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+    """
+    # Function.apply does not accept keyword arguments, so the decorator
+    # "weighted_loss" is not applicable
+    loss = _sigmoid_focal_loss(pred.contiguous(), target, gamma, alpha, None,
+                               'none')
+    if weight is not None:
+        if weight.shape != loss.shape:
+            if weight.size(0) == loss.size(0):
+                # For most cases, weight is of shape (num_priors, ),
+                #  which means it does not have the second axis num_class
+                weight = weight.view(-1, 1)
+            else:
+                # Sometimes, weight per anchor per class is also needed. e.g.
+                #  in FSAF. But it may be flattened of shape
+                #  (num_priors x num_class, ), while loss is still of shape
+                #  (num_priors, num_class).
+                assert weight.numel() == loss.numel()
+                weight = weight.view(loss.size(0), -1)
+        assert weight.ndim == loss.ndim
+    loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
+    return loss
+
+
+@LOSSES.register_module()
+class FocalLoss(nn.Module):
+
+    def __init__(self,
+                 use_sigmoid=True,
+                 gamma=2.0,
+                 alpha=0.25,
+                 reduction='mean',
+                 loss_weight=1.0):
+        """`Focal Loss <https://arxiv.org/abs/1708.02002>`_
+
+        Args:
+            use_sigmoid (bool, optional): Whether to the prediction is
+                used for sigmoid or softmax. Defaults to True.
+            gamma (float, optional): The gamma for calculating the modulating
+                factor. Defaults to 2.0.
+            alpha (float, optional): A balanced form for Focal Loss.
+                Defaults to 0.25.
+            reduction (str, optional): The method used to reduce the loss into
+                a scalar. Defaults to 'mean'. Options are "none", "mean" and
+                "sum".
+            loss_weight (float, optional): Weight of loss. Defaults to 1.0.
+        """
+        super(FocalLoss, self).__init__()
+        assert use_sigmoid is True, 'Only sigmoid focal loss supported now.'
+        self.use_sigmoid = use_sigmoid
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning label of the prediction.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Options are "none", "mean" and "sum".
+
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if self.use_sigmoid:
+            loss_cls = self.loss_weight * sigmoid_focal_loss(
+                pred,
+                target,
+                weight,
+                gamma=self.gamma,
+                alpha=self.alpha,
+                reduction=reduction,
+                avg_factor=avg_factor)
+        else:
+            raise NotImplementedError
+        return loss_cls
diff --git a/detection/scrfd/mmdet/models/losses/gaussian_focal_loss.py b/detection/scrfd/mmdet/models/losses/gaussian_focal_loss.py
new file mode 100755
index 000000000..54307a1e0
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/gaussian_focal_loss.py
@@ -0,0 +1,89 @@
+import torch.nn as nn
+
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def gaussian_focal_loss(pred, gaussian_target, alpha=2.0, gamma=4.0):
+    """`Focal Loss <https://arxiv.org/abs/1708.02002>`_ for targets in gaussian
+    distribution.
+
+    Args:
+        pred (torch.Tensor): The prediction.
+        gaussian_target (torch.Tensor): The learning target of the prediction
+            in gaussian distribution.
+        alpha (float, optional): A balanced form for Focal Loss.
+            Defaults to 2.0.
+        gamma (float, optional): The gamma for calculating the modulating
+            factor. Defaults to 4.0.
+    """
+    eps = 1e-12
+    pos_weights = gaussian_target.eq(1)
+    neg_weights = (1 - gaussian_target).pow(gamma)
+    pos_loss = -(pred + eps).log() * (1 - pred).pow(alpha) * pos_weights
+    neg_loss = -(1 - pred + eps).log() * pred.pow(alpha) * neg_weights
+    return pos_loss + neg_loss
+
+
+@LOSSES.register_module()
+class GaussianFocalLoss(nn.Module):
+    """GaussianFocalLoss is a variant of focal loss.
+
+    More details can be found in the `paper
+    <https://arxiv.org/abs/1808.01244>`_
+    Code is modified from `kp_utils.py
+    <https://github.com/princeton-vl/CornerNet/blob/master/models/py_utils/kp_utils.py#L152>`_  # noqa: E501
+    Please notice that the target in GaussianFocalLoss is a gaussian heatmap,
+    not 0/1 binary target.
+
+    Args:
+        alpha (float): Power of prediction.
+        gamma (float): Power of target for negtive samples.
+        reduction (str): Options are "none", "mean" and "sum".
+        loss_weight (float): Loss weight of current loss.
+    """
+
+    def __init__(self,
+                 alpha=2.0,
+                 gamma=4.0,
+                 reduction='mean',
+                 loss_weight=1.0):
+        super(GaussianFocalLoss, self).__init__()
+        self.alpha = alpha
+        self.gamma = gamma
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction
+                in gaussian distribution.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_reg = self.loss_weight * gaussian_focal_loss(
+            pred,
+            target,
+            weight,
+            alpha=self.alpha,
+            gamma=self.gamma,
+            reduction=reduction,
+            avg_factor=avg_factor)
+        return loss_reg
diff --git a/detection/scrfd/mmdet/models/losses/gfocal_loss.py b/detection/scrfd/mmdet/models/losses/gfocal_loss.py
new file mode 100755
index 000000000..73102da82
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/gfocal_loss.py
@@ -0,0 +1,185 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def quality_focal_loss(pred, target, beta=2.0):
+    r"""Quality Focal Loss (QFL) is from `Generalized Focal Loss: Learning
+    Qualified and Distributed Bounding Boxes for Dense Object Detection
+    <https://arxiv.org/abs/2006.04388>`_.
+
+    Args:
+        pred (torch.Tensor): Predicted joint representation of classification
+            and quality (IoU) estimation with shape (N, C), C is the number of
+            classes.
+        target (tuple([torch.Tensor])): Target category label with shape (N,)
+            and target quality label with shape (N,).
+        beta (float): The beta parameter for calculating the modulating factor.
+            Defaults to 2.0.
+
+    Returns:
+        torch.Tensor: Loss tensor with shape (N,).
+    """
+    assert len(target) == 2, """target for QFL must be a tuple of two elements,
+        including category label and quality label, respectively"""
+    # label denotes the category id, score denotes the quality score
+    label, score = target
+
+    # negatives are supervised by 0 quality score
+    pred_sigmoid = pred.sigmoid()
+    scale_factor = pred_sigmoid
+    zerolabel = scale_factor.new_zeros(pred.shape)
+    loss = F.binary_cross_entropy_with_logits(
+        pred, zerolabel, reduction='none') * scale_factor.pow(beta)
+
+    # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+    bg_class_ind = pred.size(1)
+    pos = ((label >= 0) & (label < bg_class_ind)).nonzero().squeeze(1)
+    pos_label = label[pos].long()
+    # positives are supervised by bbox quality (IoU) score
+    scale_factor = score[pos] - pred_sigmoid[pos, pos_label]
+    loss[pos, pos_label] = F.binary_cross_entropy_with_logits(
+        pred[pos, pos_label], score[pos],
+        reduction='none') * scale_factor.abs().pow(beta)
+
+    loss = loss.sum(dim=1, keepdim=False)
+    return loss
+
+
+@weighted_loss
+def distribution_focal_loss(pred, label):
+    r"""Distribution Focal Loss (DFL) is from `Generalized Focal Loss: Learning
+    Qualified and Distributed Bounding Boxes for Dense Object Detection
+    <https://arxiv.org/abs/2006.04388>`_.
+
+    Args:
+        pred (torch.Tensor): Predicted general distribution of bounding boxes
+            (before softmax) with shape (N, n+1), n is the max value of the
+            integral set `{0, ..., n}` in paper.
+        label (torch.Tensor): Target distance label for bounding boxes with
+            shape (N,).
+
+    Returns:
+        torch.Tensor: Loss tensor with shape (N,).
+    """
+    dis_left = label.long()
+    dis_right = dis_left + 1
+    weight_left = dis_right.float() - label
+    weight_right = label - dis_left.float()
+    loss = F.cross_entropy(pred, dis_left, reduction='none') * weight_left \
+        + F.cross_entropy(pred, dis_right, reduction='none') * weight_right
+    return loss
+
+
+@LOSSES.register_module()
+class QualityFocalLoss(nn.Module):
+    r"""Quality Focal Loss (QFL) is a variant of `Generalized Focal Loss:
+    Learning Qualified and Distributed Bounding Boxes for Dense Object
+    Detection <https://arxiv.org/abs/2006.04388>`_.
+
+    Args:
+        use_sigmoid (bool): Whether sigmoid operation is conducted in QFL.
+            Defaults to True.
+        beta (float): The beta parameter for calculating the modulating factor.
+            Defaults to 2.0.
+        reduction (str): Options are "none", "mean" and "sum".
+        loss_weight (float): Loss weight of current loss.
+    """
+
+    def __init__(self,
+                 use_sigmoid=True,
+                 beta=2.0,
+                 reduction='mean',
+                 loss_weight=1.0):
+        super(QualityFocalLoss, self).__init__()
+        assert use_sigmoid is True, 'Only sigmoid in QFL supported now.'
+        self.use_sigmoid = use_sigmoid
+        self.beta = beta
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): Predicted joint representation of
+                classification and quality (IoU) estimation with shape (N, C),
+                C is the number of classes.
+            target (tuple([torch.Tensor])): Target category label with shape
+                (N,) and target quality label with shape (N,).
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if self.use_sigmoid:
+            loss_cls = self.loss_weight * quality_focal_loss(
+                pred,
+                target,
+                weight,
+                beta=self.beta,
+                reduction=reduction,
+                avg_factor=avg_factor)
+        else:
+            raise NotImplementedError
+        return loss_cls
+
+
+@LOSSES.register_module()
+class DistributionFocalLoss(nn.Module):
+    r"""Distribution Focal Loss (DFL) is a variant of `Generalized Focal Loss:
+    Learning Qualified and Distributed Bounding Boxes for Dense Object
+    Detection <https://arxiv.org/abs/2006.04388>`_.
+
+    Args:
+        reduction (str): Options are `'none'`, `'mean'` and `'sum'`.
+        loss_weight (float): Loss weight of current loss.
+    """
+
+    def __init__(self, reduction='mean', loss_weight=1.0):
+        super(DistributionFocalLoss, self).__init__()
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): Predicted general distribution of bounding
+                boxes (before softmax) with shape (N, n+1), n is the max value
+                of the integral set `{0, ..., n}` in paper.
+            target (torch.Tensor): Target distance label for bounding boxes
+                with shape (N,).
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_cls = self.loss_weight * distribution_focal_loss(
+            pred, target, weight, reduction=reduction, avg_factor=avg_factor)
+        return loss_cls
diff --git a/detection/scrfd/mmdet/models/losses/ghm_loss.py b/detection/scrfd/mmdet/models/losses/ghm_loss.py
new file mode 100755
index 000000000..8969a23fd
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/ghm_loss.py
@@ -0,0 +1,172 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+
+
+def _expand_onehot_labels(labels, label_weights, label_channels):
+    bin_labels = labels.new_full((labels.size(0), label_channels), 0)
+    inds = torch.nonzero(
+        (labels >= 0) & (labels < label_channels), as_tuple=False).squeeze()
+    if inds.numel() > 0:
+        bin_labels[inds, labels[inds]] = 1
+    bin_label_weights = label_weights.view(-1, 1).expand(
+        label_weights.size(0), label_channels)
+    return bin_labels, bin_label_weights
+
+
+# TODO: code refactoring to make it consistent with other losses
+@LOSSES.register_module()
+class GHMC(nn.Module):
+    """GHM Classification Loss.
+
+    Details of the theorem can be viewed in the paper
+    `Gradient Harmonized Single-stage Detector
+    <https://arxiv.org/abs/1811.05181>`_.
+
+    Args:
+        bins (int): Number of the unit regions for distribution calculation.
+        momentum (float): The parameter for moving average.
+        use_sigmoid (bool): Can only be true for BCE based loss now.
+        loss_weight (float): The weight of the total GHM-C loss.
+    """
+
+    def __init__(self, bins=10, momentum=0, use_sigmoid=True, loss_weight=1.0):
+        super(GHMC, self).__init__()
+        self.bins = bins
+        self.momentum = momentum
+        edges = torch.arange(bins + 1).float() / bins
+        self.register_buffer('edges', edges)
+        self.edges[-1] += 1e-6
+        if momentum > 0:
+            acc_sum = torch.zeros(bins)
+            self.register_buffer('acc_sum', acc_sum)
+        self.use_sigmoid = use_sigmoid
+        if not self.use_sigmoid:
+            raise NotImplementedError
+        self.loss_weight = loss_weight
+
+    def forward(self, pred, target, label_weight, *args, **kwargs):
+        """Calculate the GHM-C loss.
+
+        Args:
+            pred (float tensor of size [batch_num, class_num]):
+                The direct prediction of classification fc layer.
+            target (float tensor of size [batch_num, class_num]):
+                Binary class target for each sample.
+            label_weight (float tensor of size [batch_num, class_num]):
+                the value is 1 if the sample is valid and 0 if ignored.
+        Returns:
+            The gradient harmonized loss.
+        """
+        # the target should be binary class label
+        if pred.dim() != target.dim():
+            target, label_weight = _expand_onehot_labels(
+                target, label_weight, pred.size(-1))
+        target, label_weight = target.float(), label_weight.float()
+        edges = self.edges
+        mmt = self.momentum
+        weights = torch.zeros_like(pred)
+
+        # gradient length
+        g = torch.abs(pred.sigmoid().detach() - target)
+
+        valid = label_weight > 0
+        tot = max(valid.float().sum().item(), 1.0)
+        n = 0  # n valid bins
+        for i in range(self.bins):
+            inds = (g >= edges[i]) & (g < edges[i + 1]) & valid
+            num_in_bin = inds.sum().item()
+            if num_in_bin > 0:
+                if mmt > 0:
+                    self.acc_sum[i] = mmt * self.acc_sum[i] \
+                        + (1 - mmt) * num_in_bin
+                    weights[inds] = tot / self.acc_sum[i]
+                else:
+                    weights[inds] = tot / num_in_bin
+                n += 1
+        if n > 0:
+            weights = weights / n
+
+        loss = F.binary_cross_entropy_with_logits(
+            pred, target, weights, reduction='sum') / tot
+        return loss * self.loss_weight
+
+
+# TODO: code refactoring to make it consistent with other losses
+@LOSSES.register_module()
+class GHMR(nn.Module):
+    """GHM Regression Loss.
+
+    Details of the theorem can be viewed in the paper
+    `Gradient Harmonized Single-stage Detector
+    <https://arxiv.org/abs/1811.05181>`_.
+
+    Args:
+        mu (float): The parameter for the Authentic Smooth L1 loss.
+        bins (int): Number of the unit regions for distribution calculation.
+        momentum (float): The parameter for moving average.
+        loss_weight (float): The weight of the total GHM-R loss.
+    """
+
+    def __init__(self, mu=0.02, bins=10, momentum=0, loss_weight=1.0):
+        super(GHMR, self).__init__()
+        self.mu = mu
+        self.bins = bins
+        edges = torch.arange(bins + 1).float() / bins
+        self.register_buffer('edges', edges)
+        self.edges[-1] = 1e3
+        self.momentum = momentum
+        if momentum > 0:
+            acc_sum = torch.zeros(bins)
+            self.register_buffer('acc_sum', acc_sum)
+        self.loss_weight = loss_weight
+
+    # TODO: support reduction parameter
+    def forward(self, pred, target, label_weight, avg_factor=None):
+        """Calculate the GHM-R loss.
+
+        Args:
+            pred (float tensor of size [batch_num, 4 (* class_num)]):
+                The prediction of box regression layer. Channel number can be 4
+                or 4 * class_num depending on whether it is class-agnostic.
+            target (float tensor of size [batch_num, 4 (* class_num)]):
+                The target regression values with the same size of pred.
+            label_weight (float tensor of size [batch_num, 4 (* class_num)]):
+                The weight of each sample, 0 if ignored.
+        Returns:
+            The gradient harmonized loss.
+        """
+        mu = self.mu
+        edges = self.edges
+        mmt = self.momentum
+
+        # ASL1 loss
+        diff = pred - target
+        loss = torch.sqrt(diff * diff + mu * mu) - mu
+
+        # gradient length
+        g = torch.abs(diff / torch.sqrt(mu * mu + diff * diff)).detach()
+        weights = torch.zeros_like(g)
+
+        valid = label_weight > 0
+        tot = max(label_weight.float().sum().item(), 1.0)
+        n = 0  # n: valid bins
+        for i in range(self.bins):
+            inds = (g >= edges[i]) & (g < edges[i + 1]) & valid
+            num_in_bin = inds.sum().item()
+            if num_in_bin > 0:
+                n += 1
+                if mmt > 0:
+                    self.acc_sum[i] = mmt * self.acc_sum[i] \
+                        + (1 - mmt) * num_in_bin
+                    weights[inds] = tot / self.acc_sum[i]
+                else:
+                    weights[inds] = tot / num_in_bin
+        if n > 0:
+            weights /= n
+
+        loss = loss * weights
+        loss = loss.sum() / tot
+        return loss * self.loss_weight
diff --git a/detection/scrfd/mmdet/models/losses/iou_loss.py b/detection/scrfd/mmdet/models/losses/iou_loss.py
new file mode 100755
index 000000000..7fd06d463
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/iou_loss.py
@@ -0,0 +1,420 @@
+import math
+
+import torch
+import torch.nn as nn
+
+from mmdet.core import bbox_overlaps
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def iou_loss(pred, target, eps=1e-6):
+    """IoU loss.
+
+    Computing the IoU loss between a set of predicted bboxes and target bboxes.
+    The loss is calculated as negative log of IoU.
+
+    Args:
+        pred (torch.Tensor): Predicted bboxes of format (x1, y1, x2, y2),
+            shape (n, 4).
+        target (torch.Tensor): Corresponding gt bboxes, shape (n, 4).
+        eps (float): Eps to avoid log(0).
+
+    Return:
+        torch.Tensor: Loss tensor.
+    """
+    ious = bbox_overlaps(pred, target, is_aligned=True).clamp(min=eps)
+    loss = -ious.log()
+    return loss
+
+
+@weighted_loss
+def bounded_iou_loss(pred, target, beta=0.2, eps=1e-3):
+    """BIoULoss.
+
+    This is an implementation of paper
+    `Improving Object Localization with Fitness NMS and Bounded IoU Loss.
+    <https://arxiv.org/abs/1711.00164>`_.
+
+    Args:
+        pred (torch.Tensor): Predicted bboxes.
+        target (torch.Tensor): Target bboxes.
+        beta (float): beta parameter in smoothl1.
+        eps (float): eps to avoid NaN.
+    """
+    pred_ctrx = (pred[:, 0] + pred[:, 2]) * 0.5
+    pred_ctry = (pred[:, 1] + pred[:, 3]) * 0.5
+    pred_w = pred[:, 2] - pred[:, 0]
+    pred_h = pred[:, 3] - pred[:, 1]
+    with torch.no_grad():
+        target_ctrx = (target[:, 0] + target[:, 2]) * 0.5
+        target_ctry = (target[:, 1] + target[:, 3]) * 0.5
+        target_w = target[:, 2] - target[:, 0]
+        target_h = target[:, 3] - target[:, 1]
+
+    dx = target_ctrx - pred_ctrx
+    dy = target_ctry - pred_ctry
+
+    loss_dx = 1 - torch.max(
+        (target_w - 2 * dx.abs()) /
+        (target_w + 2 * dx.abs() + eps), torch.zeros_like(dx))
+    loss_dy = 1 - torch.max(
+        (target_h - 2 * dy.abs()) /
+        (target_h + 2 * dy.abs() + eps), torch.zeros_like(dy))
+    loss_dw = 1 - torch.min(target_w / (pred_w + eps), pred_w /
+                            (target_w + eps))
+    loss_dh = 1 - torch.min(target_h / (pred_h + eps), pred_h /
+                            (target_h + eps))
+    loss_comb = torch.stack([loss_dx, loss_dy, loss_dw, loss_dh],
+                            dim=-1).view(loss_dx.size(0), -1)
+
+    loss = torch.where(loss_comb < beta, 0.5 * loss_comb * loss_comb / beta,
+                       loss_comb - 0.5 * beta)
+    return loss
+
+
+@weighted_loss
+def giou_loss(pred, target, eps=1e-7):
+    r"""`Generalized Intersection over Union: A Metric and A Loss for Bounding
+    Box Regression <https://arxiv.org/abs/1902.09630>`_.
+
+    Args:
+        pred (torch.Tensor): Predicted bboxes of format (x1, y1, x2, y2),
+            shape (n, 4).
+        target (torch.Tensor): Corresponding gt bboxes, shape (n, 4).
+        eps (float): Eps to avoid log(0).
+
+    Return:
+        Tensor: Loss tensor.
+    """
+    gious = bbox_overlaps(pred, target, mode='giou', is_aligned=True, eps=eps)
+    loss = 1 - gious
+    return loss
+
+
+@weighted_loss
+def diou_loss(pred, target, eps=1e-7):
+    r"""`Implementation of Distance-IoU Loss: Faster and Better
+    Learning for Bounding Box Regression, https://arxiv.org/abs/1911.08287`_.
+
+    Code is modified from https://github.com/Zzh-tju/DIoU.
+
+    Args:
+        pred (Tensor): Predicted bboxes of format (x1, y1, x2, y2),
+            shape (n, 4).
+        target (Tensor): Corresponding gt bboxes, shape (n, 4).
+        eps (float): Eps to avoid log(0).
+    Return:
+        Tensor: Loss tensor.
+    """
+    # overlap
+    lt = torch.max(pred[:, :2], target[:, :2])
+    rb = torch.min(pred[:, 2:], target[:, 2:])
+    wh = (rb - lt).clamp(min=0)
+    overlap = wh[:, 0] * wh[:, 1]
+
+    # union
+    ap = (pred[:, 2] - pred[:, 0]) * (pred[:, 3] - pred[:, 1])
+    ag = (target[:, 2] - target[:, 0]) * (target[:, 3] - target[:, 1])
+    union = ap + ag - overlap + eps
+
+    # IoU
+    ious = overlap / union
+
+    # enclose area
+    enclose_x1y1 = torch.min(pred[:, :2], target[:, :2])
+    enclose_x2y2 = torch.max(pred[:, 2:], target[:, 2:])
+    enclose_wh = (enclose_x2y2 - enclose_x1y1).clamp(min=0)
+
+    cw = enclose_wh[:, 0]
+    ch = enclose_wh[:, 1]
+
+    c2 = cw**2 + ch**2 + eps
+
+    b1_x1, b1_y1 = pred[:, 0], pred[:, 1]
+    b1_x2, b1_y2 = pred[:, 2], pred[:, 3]
+    b2_x1, b2_y1 = target[:, 0], target[:, 1]
+    b2_x2, b2_y2 = target[:, 2], target[:, 3]
+
+    left = ((b2_x1 + b2_x2) - (b1_x1 + b1_x2))**2 / 4
+    right = ((b2_y1 + b2_y2) - (b1_y1 + b1_y2))**2 / 4
+    rho2 = left + right
+
+    # DIoU
+    dious = ious - rho2 / c2
+    loss = 1 - dious
+    return loss
+
+
+@weighted_loss
+def ciou_loss(pred, target, eps=1e-7):
+    r"""`Implementation of paper `Enhancing Geometric Factors into
+    Model Learning and Inference for Object Detection and Instance
+    Segmentation <https://arxiv.org/abs/2005.03572>`_.
+
+    Code is modified from https://github.com/Zzh-tju/CIoU.
+
+    Args:
+        pred (Tensor): Predicted bboxes of format (x1, y1, x2, y2),
+            shape (n, 4).
+        target (Tensor): Corresponding gt bboxes, shape (n, 4).
+        eps (float): Eps to avoid log(0).
+    Return:
+        Tensor: Loss tensor.
+    """
+    # overlap
+    lt = torch.max(pred[:, :2], target[:, :2])
+    rb = torch.min(pred[:, 2:], target[:, 2:])
+    wh = (rb - lt).clamp(min=0)
+    overlap = wh[:, 0] * wh[:, 1]
+
+    # union
+    ap = (pred[:, 2] - pred[:, 0]) * (pred[:, 3] - pred[:, 1])
+    ag = (target[:, 2] - target[:, 0]) * (target[:, 3] - target[:, 1])
+    union = ap + ag - overlap + eps
+
+    # IoU
+    ious = overlap / union
+
+    # enclose area
+    enclose_x1y1 = torch.min(pred[:, :2], target[:, :2])
+    enclose_x2y2 = torch.max(pred[:, 2:], target[:, 2:])
+    enclose_wh = (enclose_x2y2 - enclose_x1y1).clamp(min=0)
+
+    cw = enclose_wh[:, 0]
+    ch = enclose_wh[:, 1]
+
+    c2 = cw**2 + ch**2 + eps
+
+    b1_x1, b1_y1 = pred[:, 0], pred[:, 1]
+    b1_x2, b1_y2 = pred[:, 2], pred[:, 3]
+    b2_x1, b2_y1 = target[:, 0], target[:, 1]
+    b2_x2, b2_y2 = target[:, 2], target[:, 3]
+
+    w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
+    w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
+
+    left = ((b2_x1 + b2_x2) - (b1_x1 + b1_x2))**2 / 4
+    right = ((b2_y1 + b2_y2) - (b1_y1 + b1_y2))**2 / 4
+    rho2 = left + right
+
+    factor = 4 / math.pi**2
+    v = factor * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
+
+    # CIoU
+    cious = ious - (rho2 / c2 + v**2 / (1 - ious + v))
+    loss = 1 - cious
+    return loss
+
+
+@LOSSES.register_module()
+class IoULoss(nn.Module):
+    """IoULoss.
+
+    Computing the IoU loss between a set of predicted bboxes and target bboxes.
+
+    Args:
+        eps (float): Eps to avoid log(0).
+        reduction (str): Options are "none", "mean" and "sum".
+        loss_weight (float): Weight of loss.
+    """
+
+    def __init__(self, eps=1e-6, reduction='mean', loss_weight=1.0):
+        super(IoULoss, self).__init__()
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None. Options are "none", "mean" and "sum".
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if (weight is not None) and (not torch.any(weight > 0)) and (
+                reduction != 'none'):
+            return (pred * weight).sum()  # 0
+        if weight is not None and weight.dim() > 1:
+            # TODO: remove this in the future
+            # reduce the weight of shape (n, 4) to (n,) to match the
+            # iou_loss of shape (n,)
+            assert weight.shape == pred.shape
+            weight = weight.mean(-1)
+        loss = self.loss_weight * iou_loss(
+            pred,
+            target,
+            weight,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
+
+
+@LOSSES.register_module()
+class BoundedIoULoss(nn.Module):
+
+    def __init__(self, beta=0.2, eps=1e-3, reduction='mean', loss_weight=1.0):
+        super(BoundedIoULoss, self).__init__()
+        self.beta = beta
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        if weight is not None and not torch.any(weight > 0):
+            return (pred * weight).sum()  # 0
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss = self.loss_weight * bounded_iou_loss(
+            pred,
+            target,
+            weight,
+            beta=self.beta,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
+
+
+@LOSSES.register_module()
+class GIoULoss(nn.Module):
+
+    def __init__(self, eps=1e-6, reduction='mean', loss_weight=1.0):
+        super(GIoULoss, self).__init__()
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        if weight is not None and not torch.any(weight > 0):
+            print('XXXXX', pred.shape, weight.shape)
+            return (weight.unsqueeze(1) * pred).sum()
+            #return (pred * weight).sum()  # 0
+        
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if weight is not None and weight.dim() > 1:
+            # TODO: remove this in the future
+            # reduce the weight of shape (n, 4) to (n,) to match the
+            # giou_loss of shape (n,)
+            assert weight.shape == pred.shape
+            weight = weight.mean(-1)
+        loss = self.loss_weight * giou_loss(
+            pred,
+            target,
+            weight,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
+
+
+@LOSSES.register_module()
+class DIoULoss(nn.Module):
+
+    def __init__(self, eps=1e-6, reduction='mean', loss_weight=1.0):
+        super(DIoULoss, self).__init__()
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        if weight is not None and not torch.any(weight > 0):
+            return (pred * weight).sum()  # 0
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if weight is not None and weight.dim() > 1:
+            # TODO: remove this in the future
+            # reduce the weight of shape (n, 4) to (n,) to match the
+            # giou_loss of shape (n,)
+            assert weight.shape == pred.shape
+            weight = weight.mean(-1)
+        loss = self.loss_weight * diou_loss(
+            pred,
+            target,
+            weight,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
+
+
+@LOSSES.register_module()
+class CIoULoss(nn.Module):
+
+    def __init__(self, eps=1e-6, reduction='mean', loss_weight=1.0):
+        super(CIoULoss, self).__init__()
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        if weight is not None and not torch.any(weight > 0):
+            return (pred * weight).sum()  # 0
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if weight is not None and weight.dim() > 1:
+            # TODO: remove this in the future
+            # reduce the weight of shape (n, 4) to (n,) to match the
+            # giou_loss of shape (n,)
+            assert weight.shape == pred.shape
+            weight = weight.mean(-1)
+        loss = self.loss_weight * ciou_loss(
+            pred,
+            target,
+            weight,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
diff --git a/detection/scrfd/mmdet/models/losses/mse_loss.py b/detection/scrfd/mmdet/models/losses/mse_loss.py
new file mode 100755
index 000000000..68d05752a
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/mse_loss.py
@@ -0,0 +1,49 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def mse_loss(pred, target):
+    """Warpper of mse loss."""
+    return F.mse_loss(pred, target, reduction='none')
+
+
+@LOSSES.register_module()
+class MSELoss(nn.Module):
+    """MSELoss.
+
+    Args:
+        reduction (str, optional): The method that reduces the loss to a
+            scalar. Options are "none", "mean" and "sum".
+        loss_weight (float, optional): The weight of the loss. Defaults to 1.0
+    """
+
+    def __init__(self, reduction='mean', loss_weight=1.0):
+        super().__init__()
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self, pred, target, weight=None, avg_factor=None):
+        """Forward function of loss.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction.
+            weight (torch.Tensor, optional): Weight of the loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+        loss = self.loss_weight * mse_loss(
+            pred,
+            target,
+            weight,
+            reduction=self.reduction,
+            avg_factor=avg_factor)
+        return loss
diff --git a/detection/scrfd/mmdet/models/losses/pisa_loss.py b/detection/scrfd/mmdet/models/losses/pisa_loss.py
new file mode 100755
index 000000000..fb907a2ea
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/pisa_loss.py
@@ -0,0 +1,180 @@
+import torch
+
+from mmdet.core import bbox_overlaps
+
+
+def isr_p(cls_score,
+          bbox_pred,
+          bbox_targets,
+          rois,
+          sampling_results,
+          loss_cls,
+          bbox_coder,
+          k=2,
+          bias=0,
+          num_class=80):
+    """Importance-based Sample Reweighting (ISR_P), positive part.
+
+    Args:
+        cls_score (Tensor): Predicted classification scores.
+        bbox_pred (Tensor): Predicted bbox deltas.
+        bbox_targets (tuple[Tensor]): A tuple of bbox targets, the are
+            labels, label_weights, bbox_targets, bbox_weights, respectively.
+        rois (Tensor): Anchors (single_stage) in shape (n, 4) or RoIs
+            (two_stage) in shape (n, 5).
+        sampling_results (obj): Sampling results.
+        loss_cls (func): Classification loss func of the head.
+        bbox_coder (obj): BBox coder of the head.
+        k (float): Power of the non-linear mapping.
+        bias (float): Shift of the non-linear mapping.
+        num_class (int): Number of classes, default: 80.
+
+    Return:
+        tuple([Tensor]): labels, imp_based_label_weights, bbox_targets,
+            bbox_target_weights
+    """
+
+    labels, label_weights, bbox_targets, bbox_weights = bbox_targets
+    pos_label_inds = ((labels >= 0) &
+                      (labels < num_class)).nonzero().reshape(-1)
+    pos_labels = labels[pos_label_inds]
+
+    # if no positive samples, return the original targets
+    num_pos = float(pos_label_inds.size(0))
+    if num_pos == 0:
+        return labels, label_weights, bbox_targets, bbox_weights
+
+    # merge pos_assigned_gt_inds of per image to a single tensor
+    gts = list()
+    last_max_gt = 0
+    for i in range(len(sampling_results)):
+        gt_i = sampling_results[i].pos_assigned_gt_inds
+        gts.append(gt_i + last_max_gt)
+        if len(gt_i) != 0:
+            last_max_gt = gt_i.max() + 1
+    gts = torch.cat(gts)
+    assert len(gts) == num_pos
+
+    cls_score = cls_score.detach()
+    bbox_pred = bbox_pred.detach()
+
+    # For single stage detectors, rois here indicate anchors, in shape (N, 4)
+    # For two stage detectors, rois are in shape (N, 5)
+    if rois.size(-1) == 5:
+        pos_rois = rois[pos_label_inds][:, 1:]
+    else:
+        pos_rois = rois[pos_label_inds]
+
+    if bbox_pred.size(-1) > 4:
+        bbox_pred = bbox_pred.view(bbox_pred.size(0), -1, 4)
+        pos_delta_pred = bbox_pred[pos_label_inds, pos_labels].view(-1, 4)
+    else:
+        pos_delta_pred = bbox_pred[pos_label_inds].view(-1, 4)
+
+    # compute iou of the predicted bbox and the corresponding GT
+    pos_delta_target = bbox_targets[pos_label_inds].view(-1, 4)
+    pos_bbox_pred = bbox_coder.decode(pos_rois, pos_delta_pred)
+    target_bbox_pred = bbox_coder.decode(pos_rois, pos_delta_target)
+    ious = bbox_overlaps(pos_bbox_pred, target_bbox_pred, is_aligned=True)
+
+    pos_imp_weights = label_weights[pos_label_inds]
+    # Two steps to compute IoU-HLR. Samples are first sorted by IoU locally,
+    # then sorted again within the same-rank group
+    max_l_num = pos_labels.bincount().max()
+    for label in pos_labels.unique():
+        l_inds = (pos_labels == label).nonzero().view(-1)
+        l_gts = gts[l_inds]
+        for t in l_gts.unique():
+            t_inds = l_inds[l_gts == t]
+            t_ious = ious[t_inds]
+            _, t_iou_rank_idx = t_ious.sort(descending=True)
+            _, t_iou_rank = t_iou_rank_idx.sort()
+            ious[t_inds] += max_l_num - t_iou_rank.float()
+        l_ious = ious[l_inds]
+        _, l_iou_rank_idx = l_ious.sort(descending=True)
+        _, l_iou_rank = l_iou_rank_idx.sort()  # IoU-HLR
+        # linearly map HLR to label weights
+        pos_imp_weights[l_inds] *= (max_l_num - l_iou_rank.float()) / max_l_num
+
+    pos_imp_weights = (bias + pos_imp_weights * (1 - bias)).pow(k)
+
+    # normalize to make the new weighted loss value equal to the original loss
+    pos_loss_cls = loss_cls(
+        cls_score[pos_label_inds], pos_labels, reduction_override='none')
+    if pos_loss_cls.dim() > 1:
+        ori_pos_loss_cls = pos_loss_cls * label_weights[pos_label_inds][:,
+                                                                        None]
+        new_pos_loss_cls = pos_loss_cls * pos_imp_weights[:, None]
+    else:
+        ori_pos_loss_cls = pos_loss_cls * label_weights[pos_label_inds]
+        new_pos_loss_cls = pos_loss_cls * pos_imp_weights
+    pos_loss_cls_ratio = ori_pos_loss_cls.sum() / new_pos_loss_cls.sum()
+    pos_imp_weights = pos_imp_weights * pos_loss_cls_ratio
+    label_weights[pos_label_inds] = pos_imp_weights
+
+    bbox_targets = labels, label_weights, bbox_targets, bbox_weights
+    return bbox_targets
+
+
+def carl_loss(cls_score,
+              labels,
+              bbox_pred,
+              bbox_targets,
+              loss_bbox,
+              k=1,
+              bias=0.2,
+              avg_factor=None,
+              sigmoid=False,
+              num_class=80):
+    """Classification-Aware Regression Loss (CARL).
+
+    Args:
+        cls_score (Tensor): Predicted classification scores.
+        labels (Tensor): Targets of classification.
+        bbox_pred (Tensor): Predicted bbox deltas.
+        bbox_targets (Tensor): Target of bbox regression.
+        loss_bbox (func): Regression loss func of the head.
+        bbox_coder (obj): BBox coder of the head.
+        k (float): Power of the non-linear mapping.
+        bias (float): Shift of the non-linear mapping.
+        avg_factor (int): Average factor used in regression loss.
+        sigmoid (bool): Activation of the classification score.
+        num_class (int): Number of classes, default: 80.
+
+    Return:
+        dict: CARL loss dict.
+    """
+    pos_label_inds = ((labels >= 0) &
+                      (labels < num_class)).nonzero().reshape(-1)
+    if pos_label_inds.numel() == 0:
+        return dict(loss_carl=cls_score.sum()[None] * 0.)
+    pos_labels = labels[pos_label_inds]
+
+    # multiply pos_cls_score with the corresponding bbox weight
+    # and remain gradient
+    if sigmoid:
+        pos_cls_score = cls_score.sigmoid()[pos_label_inds, pos_labels]
+    else:
+        pos_cls_score = cls_score.softmax(-1)[pos_label_inds, pos_labels]
+    carl_loss_weights = (bias + (1 - bias) * pos_cls_score).pow(k)
+
+    # normalize carl_loss_weight to make its sum equal to num positive
+    num_pos = float(pos_cls_score.size(0))
+    weight_ratio = num_pos / carl_loss_weights.sum()
+    carl_loss_weights *= weight_ratio
+
+    if avg_factor is None:
+        avg_factor = bbox_targets.size(0)
+    # if is class agnostic, bbox pred is in shape (N, 4)
+    # otherwise, bbox pred is in shape (N, #classes, 4)
+    if bbox_pred.size(-1) > 4:
+        bbox_pred = bbox_pred.view(bbox_pred.size(0), -1, 4)
+        pos_bbox_preds = bbox_pred[pos_label_inds, pos_labels]
+    else:
+        pos_bbox_preds = bbox_pred[pos_label_inds]
+    ori_loss_reg = loss_bbox(
+        pos_bbox_preds,
+        bbox_targets[pos_label_inds],
+        reduction_override='none') / avg_factor
+    loss_carl = (ori_loss_reg * carl_loss_weights[:, None]).sum()
+    return dict(loss_carl=loss_carl[None])
diff --git a/detection/scrfd/mmdet/models/losses/smooth_l1_loss.py b/detection/scrfd/mmdet/models/losses/smooth_l1_loss.py
new file mode 100755
index 000000000..ad5e8a4df
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/smooth_l1_loss.py
@@ -0,0 +1,136 @@
+import torch
+import torch.nn as nn
+
+from ..builder import LOSSES
+from .utils import weighted_loss
+
+
+@weighted_loss
+def smooth_l1_loss(pred, target, beta=1.0):
+    """Smooth L1 loss.
+
+    Args:
+        pred (torch.Tensor): The prediction.
+        target (torch.Tensor): The learning target of the prediction.
+        beta (float, optional): The threshold in the piecewise function.
+            Defaults to 1.0.
+
+    Returns:
+        torch.Tensor: Calculated loss
+    """
+    assert beta > 0
+    assert pred.size() == target.size() and target.numel() > 0
+    diff = torch.abs(pred - target)
+    loss = torch.where(diff < beta, 0.5 * diff * diff / beta,
+                       diff - 0.5 * beta)
+    return loss
+
+
+@weighted_loss
+def l1_loss(pred, target):
+    """L1 loss.
+
+    Args:
+        pred (torch.Tensor): The prediction.
+        target (torch.Tensor): The learning target of the prediction.
+
+    Returns:
+        torch.Tensor: Calculated loss
+    """
+    assert pred.size() == target.size() and target.numel() > 0
+    loss = torch.abs(pred - target)
+    return loss
+
+
+@LOSSES.register_module()
+class SmoothL1Loss(nn.Module):
+    """Smooth L1 loss.
+
+    Args:
+        beta (float, optional): The threshold in the piecewise function.
+            Defaults to 1.0.
+        reduction (str, optional): The method to reduce the loss.
+            Options are "none", "mean" and "sum". Defaults to "mean".
+        loss_weight (float, optional): The weight of loss.
+    """
+
+    def __init__(self, beta=1.0, reduction='mean', loss_weight=1.0):
+        super(SmoothL1Loss, self).__init__()
+        self.beta = beta
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_bbox = self.loss_weight * smooth_l1_loss(
+            pred,
+            target,
+            weight,
+            beta=self.beta,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss_bbox
+
+
+@LOSSES.register_module()
+class L1Loss(nn.Module):
+    """L1 loss.
+
+    Args:
+        reduction (str, optional): The method to reduce the loss.
+            Options are "none", "mean" and "sum".
+        loss_weight (float, optional): The weight of loss.
+    """
+
+    def __init__(self, reduction='mean', loss_weight=1.0):
+        super(L1Loss, self).__init__()
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_bbox = self.loss_weight * l1_loss(
+            pred, target, weight, reduction=reduction, avg_factor=avg_factor)
+        return loss_bbox
diff --git a/detection/scrfd/mmdet/models/losses/utils.py b/detection/scrfd/mmdet/models/losses/utils.py
new file mode 100755
index 000000000..ee90f3ac0
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/utils.py
@@ -0,0 +1,99 @@
+import functools
+
+import torch.nn.functional as F
+
+
+def reduce_loss(loss, reduction):
+    """Reduce loss as specified.
+
+    Args:
+        loss (Tensor): Elementwise loss tensor.
+        reduction (str): Options are "none", "mean" and "sum".
+
+    Return:
+        Tensor: Reduced loss tensor.
+    """
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, elementwise_mean:1, sum: 2
+    if reduction_enum == 0:
+        return loss
+    elif reduction_enum == 1:
+        return loss.mean()
+    elif reduction_enum == 2:
+        return loss.sum()
+
+
+def weight_reduce_loss(loss, weight=None, reduction='mean', avg_factor=None):
+    """Apply element-wise weight and reduce loss.
+
+    Args:
+        loss (Tensor): Element-wise loss.
+        weight (Tensor): Element-wise weights.
+        reduction (str): Same as built-in losses of PyTorch.
+        avg_factor (float): Avarage factor when computing the mean of losses.
+
+    Returns:
+        Tensor: Processed loss values.
+    """
+    # if weight is specified, apply element-wise weight
+    if weight is not None:
+        loss = loss * weight
+
+    # if avg_factor is not specified, just reduce the loss
+    if avg_factor is None:
+        loss = reduce_loss(loss, reduction)
+    else:
+        # if reduction is mean, then average the loss by avg_factor
+        if reduction == 'mean':
+            loss = loss.sum() / avg_factor
+        # if reduction is 'none', then do nothing, otherwise raise an error
+        elif reduction != 'none':
+            raise ValueError('avg_factor can not be used with reduction="sum"')
+    return loss
+
+
+def weighted_loss(loss_func):
+    """Create a weighted version of a given loss function.
+
+    To use this decorator, the loss function must have the signature like
+    `loss_func(pred, target, **kwargs)`. The function only needs to compute
+    element-wise loss without any reduction. This decorator will add weight
+    and reduction arguments to the function. The decorated function will have
+    the signature like `loss_func(pred, target, weight=None, reduction='mean',
+    avg_factor=None, **kwargs)`.
+
+    :Example:
+
+    >>> import torch
+    >>> @weighted_loss
+    >>> def l1_loss(pred, target):
+    >>>     return (pred - target).abs()
+
+    >>> pred = torch.Tensor([0, 2, 3])
+    >>> target = torch.Tensor([1, 1, 1])
+    >>> weight = torch.Tensor([1, 0, 1])
+
+    >>> l1_loss(pred, target)
+    tensor(1.3333)
+    >>> l1_loss(pred, target, weight)
+    tensor(1.)
+    >>> l1_loss(pred, target, reduction='none')
+    tensor([1., 1., 2.])
+    >>> l1_loss(pred, target, weight, avg_factor=2)
+    tensor(1.5000)
+    """
+
+    @functools.wraps(loss_func)
+    def wrapper(pred,
+                target,
+                weight=None,
+                reduction='mean',
+                avg_factor=None,
+                **kwargs):
+        # get element-wise loss
+        loss = loss_func(pred, target, **kwargs)
+        #print('LLL', pred.shape, target.shape, loss.shape, weight.shape)
+        loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
+        return loss
+
+    return wrapper
diff --git a/detection/scrfd/mmdet/models/losses/varifocal_loss.py b/detection/scrfd/mmdet/models/losses/varifocal_loss.py
new file mode 100755
index 000000000..6a84307fe
--- /dev/null
+++ b/detection/scrfd/mmdet/models/losses/varifocal_loss.py
@@ -0,0 +1,131 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..builder import LOSSES
+from .utils import weight_reduce_loss
+
+
+def varifocal_loss(pred,
+                   target,
+                   weight=None,
+                   alpha=0.75,
+                   gamma=2.0,
+                   iou_weighted=True,
+                   reduction='mean',
+                   avg_factor=None):
+    """`Varifocal Loss <https://arxiv.org/abs/2008.13367>`_
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the
+            number of classes
+        target (torch.Tensor): The learning target of the iou-aware
+            classification score with shape (N, C), C is the number of classes.
+        weight (torch.Tensor, optional): The weight of loss for each
+            prediction. Defaults to None.
+        alpha (float, optional): A balance factor for the negative part of
+            Varifocal Loss, which is different from the alpha of Focal Loss.
+            Defaults to 0.75.
+        gamma (float, optional): The gamma for calculating the modulating
+            factor. Defaults to 2.0.
+        iou_weighted (bool, optional): Whether to weight the loss of the
+            positive example with the iou target. Defaults to True.
+        reduction (str, optional): The method used to reduce the loss into
+            a scalar. Defaults to 'mean'. Options are "none", "mean" and
+            "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+    """
+    # pred and target should be of the same size
+    assert pred.size() == target.size()
+    pred_sigmoid = pred.sigmoid()
+    target = target.type_as(pred)
+    if iou_weighted:
+        focal_weight = target * (target > 0.0).float() + \
+            alpha * (pred_sigmoid - target).abs().pow(gamma) * \
+            (target <= 0.0).float()
+    else:
+        focal_weight = (target > 0.0).float() + \
+            alpha * (pred_sigmoid - target).abs().pow(gamma) * \
+            (target <= 0.0).float()
+    loss = F.binary_cross_entropy_with_logits(
+        pred, target, reduction='none') * focal_weight
+    loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
+    return loss
+
+
+@LOSSES.register_module()
+class VarifocalLoss(nn.Module):
+
+    def __init__(self,
+                 use_sigmoid=True,
+                 alpha=0.75,
+                 gamma=2.0,
+                 iou_weighted=True,
+                 reduction='mean',
+                 loss_weight=1.0):
+        """`Varifocal Loss <https://arxiv.org/abs/2008.13367>`_
+
+        Args:
+            use_sigmoid (bool, optional): Whether the prediction is
+                used for sigmoid or softmax. Defaults to True.
+            alpha (float, optional): A balance factor for the negative part of
+                Varifocal Loss, which is different from the alpha of Focal
+                Loss. Defaults to 0.75.
+            gamma (float, optional): The gamma for calculating the modulating
+                factor. Defaults to 2.0.
+            iou_weighted (bool, optional): Whether to weight the loss of the
+                positive examples with the iou target. Defaults to True.
+            reduction (str, optional): The method used to reduce the loss into
+                a scalar. Defaults to 'mean'. Options are "none", "mean" and
+                "sum".
+            loss_weight (float, optional): Weight of loss. Defaults to 1.0.
+        """
+        super(VarifocalLoss, self).__init__()
+        assert use_sigmoid is True, \
+            'Only sigmoid varifocal loss supported now.'
+        assert alpha >= 0.0
+        self.use_sigmoid = use_sigmoid
+        self.alpha = alpha
+        self.gamma = gamma
+        self.iou_weighted = iou_weighted
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction.
+            target (torch.Tensor): The learning target of the prediction.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Options are "none", "mean" and "sum".
+
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if self.use_sigmoid:
+            loss_cls = self.loss_weight * varifocal_loss(
+                pred,
+                target,
+                weight,
+                alpha=self.alpha,
+                gamma=self.gamma,
+                iou_weighted=self.iou_weighted,
+                reduction=reduction,
+                avg_factor=avg_factor)
+        else:
+            raise NotImplementedError
+        return loss_cls
diff --git a/detection/scrfd/mmdet/models/necks/__init__.py b/detection/scrfd/mmdet/models/necks/__init__.py
new file mode 100755
index 000000000..6a45250fe
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/__init__.py
@@ -0,0 +1,17 @@
+from .bfp import BFP
+from .channel_mapper import ChannelMapper
+from .fpn import FPN
+from .fpn_carafe import FPN_CARAFE
+from .hrfpn import HRFPN
+from .nas_fpn import NASFPN
+from .nasfcos_fpn import NASFCOS_FPN
+from .pafpn import PAFPN
+from .rfp import RFP
+from .yolo_neck import YOLOV3Neck
+from .lfpn import LFPN
+
+__all__ = [
+    'FPN', 'BFP', 'ChannelMapper', 'HRFPN', 'NASFPN', 'FPN_CARAFE', 'PAFPN',
+    'NASFCOS_FPN', 'RFP', 'YOLOV3Neck',
+    'LFPN'
+]
diff --git a/detection/scrfd/mmdet/models/necks/bfp.py b/detection/scrfd/mmdet/models/necks/bfp.py
new file mode 100755
index 000000000..65f00fb30
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/bfp.py
@@ -0,0 +1,104 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, xavier_init
+from mmcv.cnn.bricks import NonLocal2d
+import numpy as np
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class BFP(nn.Module):
+    """BFP (Balanced Feature Pyrmamids)
+
+    BFP takes multi-level features as inputs and gather them into a single one,
+    then refine the gathered feature and scatter the refined results to
+    multi-level features. This module is used in Libra R-CNN (CVPR 2019), see
+    the paper `Libra R-CNN: Towards Balanced Learning for Object Detection
+    <https://arxiv.org/abs/1904.02701>`_ for details.
+
+    Args:
+        in_channels (int): Number of input channels (feature maps of all levels
+            should have the same channels).
+        num_levels (int): Number of input feature levels.
+        conv_cfg (dict): The config dict for convolution layers.
+        norm_cfg (dict): The config dict for normalization layers.
+        refine_level (int): Index of integration and refine level of BSF in
+            multi-level features from bottom to top.
+        refine_type (str): Type of the refine op, currently support
+            [None, 'conv', 'non_local'].
+    """
+
+    def __init__(self,
+                 in_channels,
+                 num_levels,
+                 refine_level=2,
+                 refine_type=None,
+                 conv_cfg=None,
+                 norm_cfg=None):
+        super(BFP, self).__init__()
+        assert refine_type in [None, 'conv', 'non_local']
+
+        self.in_channels = in_channels
+        self.num_levels = num_levels
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+
+        self.refine_level = refine_level
+        self.refine_type = refine_type
+        assert 0 <= self.refine_level < self.num_levels
+
+        if self.refine_type == 'conv':
+            self.refine = ConvModule(
+                self.in_channels,
+                self.in_channels,
+                3,
+                padding=1,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg)
+        elif self.refine_type == 'non_local':
+            self.refine = NonLocal2d(
+                self.in_channels,
+                reduction=1,
+                use_scale=False,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg)
+
+    def init_weights(self):
+        """Initialize the weights of FPN module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == self.num_levels
+
+        # step 1: gather multi-level features by resize and average
+        feats = []
+        gather_size = inputs[self.refine_level].size()[2:]
+        for i in range(self.num_levels):
+            if i < self.refine_level:
+                gathered = F.adaptive_max_pool2d(
+                    inputs[i], output_size=gather_size)
+            else:
+                gathered = F.interpolate(
+                    inputs[i], size=gather_size, mode='nearest')
+            feats.append(gathered)
+
+        bsf = np.mean(feats)
+
+        # step 2: refine gathered features
+        if self.refine_type is not None:
+            bsf = self.refine(bsf)
+
+        # step 3: scatter refined features to multi-levels by a residual path
+        outs = []
+        for i in range(self.num_levels):
+            out_size = inputs[i].size()[2:]
+            if i < self.refine_level:
+                residual = F.interpolate(bsf, size=out_size, mode='nearest')
+            else:
+                residual = F.adaptive_max_pool2d(bsf, output_size=out_size)
+            outs.append(residual + inputs[i])
+
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/channel_mapper.py b/detection/scrfd/mmdet/models/necks/channel_mapper.py
new file mode 100755
index 000000000..a4f5ed44c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/channel_mapper.py
@@ -0,0 +1,74 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, xavier_init
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class ChannelMapper(nn.Module):
+    r"""Channel Mapper to reduce/increase channels of backbone features.
+
+    This is used to reduce/increase channels of backbone features.
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale).
+        kernel_size (int, optional): kernel_size for reducing channels (used
+            at each scale). Default: 3.
+        conv_cfg (dict, optional): Config dict for convolution layer.
+            Default: None.
+        norm_cfg (dict, optional): Config dict for normalization layer.
+            Default: None.
+        act_cfg (dict, optional): Config dict for activation layer in
+            ConvModule. Default: dict(type='ReLU').
+
+    Example:
+        >>> import torch
+        >>> in_channels = [2, 3, 5, 7]
+        >>> scales = [340, 170, 84, 43]
+        >>> inputs = [torch.rand(1, c, s, s)
+        ...           for c, s in zip(in_channels, scales)]
+        >>> self = ChannelMapper(in_channels, 11, 3).eval()
+        >>> outputs = self.forward(inputs)
+        >>> for i in range(len(outputs)):
+        ...     print(f'outputs[{i}].shape = {outputs[i].shape}')
+        outputs[0].shape = torch.Size([1, 11, 340, 340])
+        outputs[1].shape = torch.Size([1, 11, 170, 170])
+        outputs[2].shape = torch.Size([1, 11, 84, 84])
+        outputs[3].shape = torch.Size([1, 11, 43, 43])
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 act_cfg=dict(type='ReLU')):
+        super(ChannelMapper, self).__init__()
+        assert isinstance(in_channels, list)
+
+        self.convs = nn.ModuleList()
+        for in_channel in in_channels:
+            self.convs.append(
+                ConvModule(
+                    in_channel,
+                    out_channels,
+                    kernel_size,
+                    padding=(kernel_size - 1) // 2,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    act_cfg=act_cfg))
+
+    # default init_weights for conv(msra) and norm in ConvModule
+    def init_weights(self):
+        """Initialize the weights of ChannelMapper module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.convs)
+        outs = [self.convs[i](inputs[i]) for i in range(len(inputs))]
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/fpn.py b/detection/scrfd/mmdet/models/necks/fpn.py
new file mode 100755
index 000000000..5e5dfe685
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/fpn.py
@@ -0,0 +1,221 @@
+import warnings
+
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, xavier_init
+from mmcv.runner import auto_fp16
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class FPN(nn.Module):
+    r"""Feature Pyramid Network.
+
+    This is an implementation of paper `Feature Pyramid Networks for Object
+    Detection <https://arxiv.org/abs/1612.03144>`_.
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale)
+        num_outs (int): Number of output scales.
+        start_level (int): Index of the start input backbone level used to
+            build the feature pyramid. Default: 0.
+        end_level (int): Index of the end input backbone level (exclusive) to
+            build the feature pyramid. Default: -1, which means the last level.
+        add_extra_convs (bool | str): If bool, it decides whether to add conv
+            layers on top of the original feature maps. Default to False.
+            If True, its actual mode is specified by `extra_convs_on_inputs`.
+            If str, it specifies the source feature map of the extra convs.
+            Only the following options are allowed
+
+            - 'on_input': Last feat map of neck inputs (i.e. backbone feature).
+            - 'on_lateral':  Last feature map after lateral convs.
+            - 'on_output': The last output feature map after fpn convs.
+        extra_convs_on_inputs (bool, deprecated): Whether to apply extra convs
+            on the original feature from the backbone. If True,
+            it is equivalent to `add_extra_convs='on_input'`. If False, it is
+            equivalent to set `add_extra_convs='on_output'`. Default to True.
+        relu_before_extra_convs (bool): Whether to apply relu before the extra
+            conv. Default: False.
+        no_norm_on_lateral (bool): Whether to apply norm on lateral.
+            Default: False.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Config dict for normalization layer. Default: None.
+        act_cfg (str): Config dict for activation layer in ConvModule.
+            Default: None.
+        upsample_cfg (dict): Config dict for interpolate layer.
+            Default: `dict(mode='nearest')`
+
+    Example:
+        >>> import torch
+        >>> in_channels = [2, 3, 5, 7]
+        >>> scales = [340, 170, 84, 43]
+        >>> inputs = [torch.rand(1, c, s, s)
+        ...           for c, s in zip(in_channels, scales)]
+        >>> self = FPN(in_channels, 11, len(in_channels)).eval()
+        >>> outputs = self.forward(inputs)
+        >>> for i in range(len(outputs)):
+        ...     print(f'outputs[{i}].shape = {outputs[i].shape}')
+        outputs[0].shape = torch.Size([1, 11, 340, 340])
+        outputs[1].shape = torch.Size([1, 11, 170, 170])
+        outputs[2].shape = torch.Size([1, 11, 84, 84])
+        outputs[3].shape = torch.Size([1, 11, 43, 43])
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 start_level=0,
+                 end_level=-1,
+                 add_extra_convs=False,
+                 extra_convs_on_inputs=True,
+                 relu_before_extra_convs=False,
+                 no_norm_on_lateral=False,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 act_cfg=None,
+                 upsample_cfg=dict(mode='nearest')):
+        super(FPN, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)
+        self.num_outs = num_outs
+        self.relu_before_extra_convs = relu_before_extra_convs
+        self.no_norm_on_lateral = no_norm_on_lateral
+        self.fp16_enabled = False
+        self.upsample_cfg = upsample_cfg.copy()
+
+        if end_level == -1:
+            self.backbone_end_level = self.num_ins
+            assert num_outs >= self.num_ins - start_level
+        else:
+            # if end_level < inputs, no extra level is allowed
+            self.backbone_end_level = end_level
+            assert end_level <= len(in_channels)
+            assert num_outs == end_level - start_level
+        self.start_level = start_level
+        self.end_level = end_level
+        self.add_extra_convs = add_extra_convs
+        assert isinstance(add_extra_convs, (str, bool))
+        if isinstance(add_extra_convs, str):
+            # Extra_convs_source choices: 'on_input', 'on_lateral', 'on_output'
+            assert add_extra_convs in ('on_input', 'on_lateral', 'on_output')
+        elif add_extra_convs:  # True
+            if extra_convs_on_inputs:
+                # TODO: deprecate `extra_convs_on_inputs`
+                warnings.simplefilter('once')
+                warnings.warn(
+                    '"extra_convs_on_inputs" will be deprecated in v2.9.0,'
+                    'Please use "add_extra_convs"', DeprecationWarning)
+                self.add_extra_convs = 'on_input'
+            else:
+                self.add_extra_convs = 'on_output'
+
+        self.lateral_convs = nn.ModuleList()
+        self.fpn_convs = nn.ModuleList()
+
+        for i in range(self.start_level, self.backbone_end_level):
+            l_conv = ConvModule(
+                in_channels[i],
+                out_channels,
+                1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
+                act_cfg=act_cfg,
+                inplace=False)
+            fpn_conv = ConvModule(
+                out_channels,
+                out_channels,
+                3,
+                padding=1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                act_cfg=act_cfg,
+                inplace=False)
+
+            self.lateral_convs.append(l_conv)
+            self.fpn_convs.append(fpn_conv)
+
+        # add extra conv layers (e.g., RetinaNet)
+        extra_levels = num_outs - self.backbone_end_level + self.start_level
+        if self.add_extra_convs and extra_levels >= 1:
+            for i in range(extra_levels):
+                if i == 0 and self.add_extra_convs == 'on_input':
+                    in_channels = self.in_channels[self.backbone_end_level - 1]
+                else:
+                    in_channels = out_channels
+                extra_fpn_conv = ConvModule(
+                    in_channels,
+                    out_channels,
+                    3,
+                    stride=2,
+                    padding=1,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    act_cfg=act_cfg,
+                    inplace=False)
+                self.fpn_convs.append(extra_fpn_conv)
+
+    # default init_weights for conv(msra) and norm in ConvModule
+    def init_weights(self):
+        """Initialize the weights of FPN module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    @auto_fp16()
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.in_channels)
+
+        # build laterals
+        laterals = [
+            lateral_conv(inputs[i + self.start_level])
+            for i, lateral_conv in enumerate(self.lateral_convs)
+        ]
+
+        # build top-down path
+        used_backbone_levels = len(laterals)
+        for i in range(used_backbone_levels - 1, 0, -1):
+            # In some cases, fixing `scale factor` (e.g. 2) is preferred, but
+            #  it cannot co-exist with `size` in `F.interpolate`.
+            if 'scale_factor' in self.upsample_cfg:
+                laterals[i - 1] += F.interpolate(laterals[i],
+                                                 **self.upsample_cfg)
+            else:
+                prev_shape = laterals[i - 1].shape[2:]
+                laterals[i - 1] += F.interpolate(
+                    laterals[i], size=prev_shape, **self.upsample_cfg)
+
+        # build outputs
+        # part 1: from original levels
+        outs = [
+            self.fpn_convs[i](laterals[i]) for i in range(used_backbone_levels)
+        ]
+        # part 2: add extra levels
+        if self.num_outs > len(outs):
+            # use max pool to get more levels on top of outputs
+            # (e.g., Faster R-CNN, Mask R-CNN)
+            if not self.add_extra_convs:
+                for i in range(self.num_outs - used_backbone_levels):
+                    outs.append(F.max_pool2d(outs[-1], 1, stride=2))
+            # add conv layers on top of original feature maps (RetinaNet)
+            else:
+                if self.add_extra_convs == 'on_input':
+                    extra_source = inputs[self.backbone_end_level - 1]
+                elif self.add_extra_convs == 'on_lateral':
+                    extra_source = laterals[-1]
+                elif self.add_extra_convs == 'on_output':
+                    extra_source = outs[-1]
+                else:
+                    raise NotImplementedError
+                outs.append(self.fpn_convs[used_backbone_levels](extra_source))
+                for i in range(used_backbone_levels + 1, self.num_outs):
+                    if self.relu_before_extra_convs:
+                        outs.append(self.fpn_convs[i](F.relu(outs[-1])))
+                    else:
+                        outs.append(self.fpn_convs[i](outs[-1]))
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/fpn_carafe.py b/detection/scrfd/mmdet/models/necks/fpn_carafe.py
new file mode 100755
index 000000000..b97a6aa73
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/fpn_carafe.py
@@ -0,0 +1,267 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, build_upsample_layer, xavier_init
+from mmcv.ops.carafe import CARAFEPack
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class FPN_CARAFE(nn.Module):
+    """FPN_CARAFE is a more flexible implementation of FPN. It allows more
+    choice for upsample methods during the top-down pathway.
+
+    It can reproduce the preformance of ICCV 2019 paper
+    CARAFE: Content-Aware ReAssembly of FEatures
+    Please refer to https://arxiv.org/abs/1905.02188 for more details.
+
+    Args:
+        in_channels (list[int]): Number of channels for each input feature map.
+        out_channels (int): Output channels of feature pyramids.
+        num_outs (int): Number of output stages.
+        start_level (int): Start level of feature pyramids.
+            (Default: 0)
+        end_level (int): End level of feature pyramids.
+            (Default: -1 indicates the last level).
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+        activate (str): Type of activation function in ConvModule
+            (Default: None indicates w/o activation).
+        order (dict): Order of components in ConvModule.
+        upsample (str): Type of upsample layer.
+        upsample_cfg (dict): Dictionary to construct and config upsample layer.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 start_level=0,
+                 end_level=-1,
+                 norm_cfg=None,
+                 act_cfg=None,
+                 order=('conv', 'norm', 'act'),
+                 upsample_cfg=dict(
+                     type='carafe',
+                     up_kernel=5,
+                     up_group=1,
+                     encoder_kernel=3,
+                     encoder_dilation=1)):
+        super(FPN_CARAFE, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)
+        self.num_outs = num_outs
+        self.norm_cfg = norm_cfg
+        self.act_cfg = act_cfg
+        self.with_bias = norm_cfg is None
+        self.upsample_cfg = upsample_cfg.copy()
+        self.upsample = self.upsample_cfg.get('type')
+        self.relu = nn.ReLU(inplace=False)
+
+        self.order = order
+        assert order in [('conv', 'norm', 'act'), ('act', 'conv', 'norm')]
+
+        assert self.upsample in [
+            'nearest', 'bilinear', 'deconv', 'pixel_shuffle', 'carafe', None
+        ]
+        if self.upsample in ['deconv', 'pixel_shuffle']:
+            assert hasattr(
+                self.upsample_cfg,
+                'upsample_kernel') and self.upsample_cfg.upsample_kernel > 0
+            self.upsample_kernel = self.upsample_cfg.pop('upsample_kernel')
+
+        if end_level == -1:
+            self.backbone_end_level = self.num_ins
+            assert num_outs >= self.num_ins - start_level
+        else:
+            # if end_level < inputs, no extra level is allowed
+            self.backbone_end_level = end_level
+            assert end_level <= len(in_channels)
+            assert num_outs == end_level - start_level
+        self.start_level = start_level
+        self.end_level = end_level
+
+        self.lateral_convs = nn.ModuleList()
+        self.fpn_convs = nn.ModuleList()
+        self.upsample_modules = nn.ModuleList()
+
+        for i in range(self.start_level, self.backbone_end_level):
+            l_conv = ConvModule(
+                in_channels[i],
+                out_channels,
+                1,
+                norm_cfg=norm_cfg,
+                bias=self.with_bias,
+                act_cfg=act_cfg,
+                inplace=False,
+                order=self.order)
+            fpn_conv = ConvModule(
+                out_channels,
+                out_channels,
+                3,
+                padding=1,
+                norm_cfg=self.norm_cfg,
+                bias=self.with_bias,
+                act_cfg=act_cfg,
+                inplace=False,
+                order=self.order)
+            if i != self.backbone_end_level - 1:
+                upsample_cfg_ = self.upsample_cfg.copy()
+                if self.upsample == 'deconv':
+                    upsample_cfg_.update(
+                        in_channels=out_channels,
+                        out_channels=out_channels,
+                        kernel_size=self.upsample_kernel,
+                        stride=2,
+                        padding=(self.upsample_kernel - 1) // 2,
+                        output_padding=(self.upsample_kernel - 1) // 2)
+                elif self.upsample == 'pixel_shuffle':
+                    upsample_cfg_.update(
+                        in_channels=out_channels,
+                        out_channels=out_channels,
+                        scale_factor=2,
+                        upsample_kernel=self.upsample_kernel)
+                elif self.upsample == 'carafe':
+                    upsample_cfg_.update(channels=out_channels, scale_factor=2)
+                else:
+                    # suppress warnings
+                    align_corners = (None
+                                     if self.upsample == 'nearest' else False)
+                    upsample_cfg_.update(
+                        scale_factor=2,
+                        mode=self.upsample,
+                        align_corners=align_corners)
+                upsample_module = build_upsample_layer(upsample_cfg_)
+                self.upsample_modules.append(upsample_module)
+            self.lateral_convs.append(l_conv)
+            self.fpn_convs.append(fpn_conv)
+
+        # add extra conv layers (e.g., RetinaNet)
+        extra_out_levels = (
+            num_outs - self.backbone_end_level + self.start_level)
+        if extra_out_levels >= 1:
+            for i in range(extra_out_levels):
+                in_channels = (
+                    self.in_channels[self.backbone_end_level -
+                                     1] if i == 0 else out_channels)
+                extra_l_conv = ConvModule(
+                    in_channels,
+                    out_channels,
+                    3,
+                    stride=2,
+                    padding=1,
+                    norm_cfg=norm_cfg,
+                    bias=self.with_bias,
+                    act_cfg=act_cfg,
+                    inplace=False,
+                    order=self.order)
+                if self.upsample == 'deconv':
+                    upsampler_cfg_ = dict(
+                        in_channels=out_channels,
+                        out_channels=out_channels,
+                        kernel_size=self.upsample_kernel,
+                        stride=2,
+                        padding=(self.upsample_kernel - 1) // 2,
+                        output_padding=(self.upsample_kernel - 1) // 2)
+                elif self.upsample == 'pixel_shuffle':
+                    upsampler_cfg_ = dict(
+                        in_channels=out_channels,
+                        out_channels=out_channels,
+                        scale_factor=2,
+                        upsample_kernel=self.upsample_kernel)
+                elif self.upsample == 'carafe':
+                    upsampler_cfg_ = dict(
+                        channels=out_channels,
+                        scale_factor=2,
+                        **self.upsample_cfg)
+                else:
+                    # suppress warnings
+                    align_corners = (None
+                                     if self.upsample == 'nearest' else False)
+                    upsampler_cfg_ = dict(
+                        scale_factor=2,
+                        mode=self.upsample,
+                        align_corners=align_corners)
+                upsampler_cfg_['type'] = self.upsample
+                upsample_module = build_upsample_layer(upsampler_cfg_)
+                extra_fpn_conv = ConvModule(
+                    out_channels,
+                    out_channels,
+                    3,
+                    padding=1,
+                    norm_cfg=self.norm_cfg,
+                    bias=self.with_bias,
+                    act_cfg=act_cfg,
+                    inplace=False,
+                    order=self.order)
+                self.upsample_modules.append(upsample_module)
+                self.fpn_convs.append(extra_fpn_conv)
+                self.lateral_convs.append(extra_l_conv)
+
+    # default init_weights for conv(msra) and norm in ConvModule
+    def init_weights(self):
+        """Initialize the weights of module."""
+        for m in self.modules():
+            if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
+                xavier_init(m, distribution='uniform')
+        for m in self.modules():
+            if isinstance(m, CARAFEPack):
+                m.init_weights()
+
+    def slice_as(self, src, dst):
+        """Slice ``src`` as ``dst``
+
+        Note:
+            ``src`` should have the same or larger size than ``dst``.
+
+        Args:
+            src (torch.Tensor): Tensors to be sliced.
+            dst (torch.Tensor): ``src`` will be sliced to have the same
+                size as ``dst``.
+
+        Returns:
+            torch.Tensor: Sliced tensor.
+        """
+        assert (src.size(2) >= dst.size(2)) and (src.size(3) >= dst.size(3))
+        if src.size(2) == dst.size(2) and src.size(3) == dst.size(3):
+            return src
+        else:
+            return src[:, :, :dst.size(2), :dst.size(3)]
+
+    def tensor_add(self, a, b):
+        """Add tensors ``a`` and ``b`` that might have different sizes."""
+        if a.size() == b.size():
+            c = a + b
+        else:
+            c = a + self.slice_as(b, a)
+        return c
+
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.in_channels)
+
+        # build laterals
+        laterals = []
+        for i, lateral_conv in enumerate(self.lateral_convs):
+            if i <= self.backbone_end_level - self.start_level:
+                input = inputs[min(i + self.start_level, len(inputs) - 1)]
+            else:
+                input = laterals[-1]
+            lateral = lateral_conv(input)
+            laterals.append(lateral)
+
+        # build top-down path
+        for i in range(len(laterals) - 1, 0, -1):
+            if self.upsample is not None:
+                upsample_feat = self.upsample_modules[i - 1](laterals[i])
+            else:
+                upsample_feat = laterals[i]
+            laterals[i - 1] = self.tensor_add(laterals[i - 1], upsample_feat)
+
+        # build outputs
+        num_conv_outs = len(self.fpn_convs)
+        outs = []
+        for i in range(num_conv_outs):
+            out = self.fpn_convs[i](laterals[i])
+            outs.append(out)
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/hrfpn.py b/detection/scrfd/mmdet/models/necks/hrfpn.py
new file mode 100755
index 000000000..cf87cfa79
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/hrfpn.py
@@ -0,0 +1,102 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, caffe2_xavier_init
+from torch.utils.checkpoint import checkpoint
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class HRFPN(nn.Module):
+    """HRFPN (High Resolution Feature Pyrmamids)
+
+    paper: `High-Resolution Representations for Labeling Pixels and Regions
+    <https://arxiv.org/abs/1904.04514>`_.
+
+    Args:
+        in_channels (list): number of channels for each branch.
+        out_channels (int): output channels of feature pyramids.
+        num_outs (int): number of output stages.
+        pooling_type (str): pooling for generating feature pyramids
+            from {MAX, AVG}.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+        with_cp  (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+        stride (int): stride of 3x3 convolutional layers
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs=5,
+                 pooling_type='AVG',
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 with_cp=False,
+                 stride=1):
+        super(HRFPN, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)
+        self.num_outs = num_outs
+        self.with_cp = with_cp
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+
+        self.reduction_conv = ConvModule(
+            sum(in_channels),
+            out_channels,
+            kernel_size=1,
+            conv_cfg=self.conv_cfg,
+            act_cfg=None)
+
+        self.fpn_convs = nn.ModuleList()
+        for i in range(self.num_outs):
+            self.fpn_convs.append(
+                ConvModule(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    padding=1,
+                    stride=stride,
+                    conv_cfg=self.conv_cfg,
+                    act_cfg=None))
+
+        if pooling_type == 'MAX':
+            self.pooling = F.max_pool2d
+        else:
+            self.pooling = F.avg_pool2d
+
+    def init_weights(self):
+        """Initialize the weights of module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                caffe2_xavier_init(m)
+
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == self.num_ins
+        outs = [inputs[0]]
+        for i in range(1, self.num_ins):
+            outs.append(
+                F.interpolate(inputs[i], scale_factor=2**i, mode='bilinear'))
+        out = torch.cat(outs, dim=1)
+        if out.requires_grad and self.with_cp:
+            out = checkpoint(self.reduction_conv, out)
+        else:
+            out = self.reduction_conv(out)
+        outs = [out]
+        for i in range(1, self.num_outs):
+            outs.append(self.pooling(out, kernel_size=2**i, stride=2**i))
+        outputs = []
+
+        for i in range(self.num_outs):
+            if outs[i].requires_grad and self.with_cp:
+                tmp_out = checkpoint(self.fpn_convs[i], outs[i])
+            else:
+                tmp_out = self.fpn_convs[i](outs[i])
+            outputs.append(tmp_out)
+        return tuple(outputs)
diff --git a/detection/scrfd/mmdet/models/necks/lfpn.py b/detection/scrfd/mmdet/models/necks/lfpn.py
new file mode 100755
index 000000000..a82cf0658
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/lfpn.py
@@ -0,0 +1,176 @@
+import warnings
+
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, xavier_init
+from mmcv.runner import auto_fp16
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class LFPN(nn.Module):
+    r"""Feature Pyramid Network.
+
+    This is an implementation of paper `Feature Pyramid Networks for Object
+    Detection <https://arxiv.org/abs/1612.03144>`_.
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale)
+        num_outs (int): Number of output scales.
+        start_level (int): Index of the start input backbone level used to
+            build the feature pyramid. Default: 0.
+        end_level (int): Index of the end input backbone level (exclusive) to
+            build the feature pyramid. Default: -1, which means the last level.
+        add_extra_convs (bool | str): If bool, it decides whether to add conv
+            layers on top of the original feature maps. Default to False.
+            If True, its actual mode is specified by `extra_convs_on_inputs`.
+            If str, it specifies the source feature map of the extra convs.
+            Only the following options are allowed
+
+            - 'on_input': Last feat map of neck inputs (i.e. backbone feature).
+            - 'on_lateral':  Last feature map after lateral convs.
+            - 'on_output': The last output feature map after fpn convs.
+        extra_convs_on_inputs (bool, deprecated): Whether to apply extra convs
+            on the original feature from the backbone. If True,
+            it is equivalent to `add_extra_convs='on_input'`. If False, it is
+            equivalent to set `add_extra_convs='on_output'`. Default to True.
+        relu_before_extra_convs (bool): Whether to apply relu before the extra
+            conv. Default: False.
+        no_norm_on_lateral (bool): Whether to apply norm on lateral.
+            Default: False.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Config dict for normalization layer. Default: None.
+        act_cfg (str): Config dict for activation layer in ConvModule.
+            Default: None.
+        upsample_cfg (dict): Config dict for interpolate layer.
+            Default: `dict(mode='nearest')`
+
+    Example:
+        >>> import torch
+        >>> in_channels = [2, 3, 5, 7]
+        >>> scales = [340, 170, 84, 43]
+        >>> inputs = [torch.rand(1, c, s, s)
+        ...           for c, s in zip(in_channels, scales)]
+        >>> self = FPN(in_channels, 11, len(in_channels)).eval()
+        >>> outputs = self.forward(inputs)
+        >>> for i in range(len(outputs)):
+        ...     print(f'outputs[{i}].shape = {outputs[i].shape}')
+        outputs[0].shape = torch.Size([1, 11, 340, 340])
+        outputs[1].shape = torch.Size([1, 11, 170, 170])
+        outputs[2].shape = torch.Size([1, 11, 84, 84])
+        outputs[3].shape = torch.Size([1, 11, 43, 43])
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 start_level=0,
+                 end_level=-1,
+                 add_extra_convs=False,
+                 extra_convs_on_inputs=True,
+                 relu_before_extra_convs=False,
+                 no_norm_on_lateral=False,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 act_cfg=None,
+                 upsample_cfg=dict(mode='nearest')):
+        super(LFPN, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)
+        self.num_outs = num_outs
+        self.relu_before_extra_convs = relu_before_extra_convs
+        self.no_norm_on_lateral = no_norm_on_lateral
+        self.fp16_enabled = False
+        self.upsample_cfg = upsample_cfg.copy()
+
+        if end_level == -1:
+            self.backbone_end_level = self.num_ins
+            assert num_outs >= self.num_ins - start_level
+        else:
+            # if end_level < inputs, no extra level is allowed
+            self.backbone_end_level = end_level
+            assert end_level <= len(in_channels)
+            assert num_outs == end_level - start_level
+        self.start_level = start_level
+        self.end_level = end_level
+        self.add_extra_convs = add_extra_convs
+        assert isinstance(add_extra_convs, (str, bool))
+        if isinstance(add_extra_convs, str):
+            # Extra_convs_source choices: 'on_input', 'on_lateral', 'on_output'
+            assert add_extra_convs in ('on_input', 'on_lateral', 'on_output')
+        elif add_extra_convs:  # True
+            if extra_convs_on_inputs:
+                # TODO: deprecate `extra_convs_on_inputs`
+                warnings.simplefilter('once')
+                warnings.warn(
+                    '"extra_convs_on_inputs" will be deprecated in v2.9.0,'
+                    'Please use "add_extra_convs"', DeprecationWarning)
+                self.add_extra_convs = 'on_input'
+            else:
+                self.add_extra_convs = 'on_output'
+
+        self.lateral_convs = nn.ModuleList()
+
+        for i in range(self.start_level, self.backbone_end_level):
+            l_conv = ConvModule(
+                in_channels[i],
+                out_channels,
+                1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
+                act_cfg=act_cfg,
+                inplace=False)
+
+            self.lateral_convs.append(l_conv)
+            #self.fpn_convs.append(fpn_conv)
+
+        # add extra conv layers (e.g., RetinaNet)
+
+    # default init_weights for conv(msra) and norm in ConvModule
+    def init_weights(self):
+        """Initialize the weights of FPN module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    @auto_fp16()
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.in_channels)
+
+        # build laterals
+        laterals = [
+            lateral_conv(inputs[i + self.start_level])
+            for i, lateral_conv in enumerate(self.lateral_convs)
+        ]
+
+        # build top-down path
+        used_backbone_levels = len(laterals)
+        for i in range(used_backbone_levels - 1, 0, -1):
+            # In some cases, fixing `scale factor` (e.g. 2) is preferred, but
+            #  it cannot co-exist with `size` in `F.interpolate`.
+            if 'scale_factor' in self.upsample_cfg:
+                laterals[i - 1] += F.interpolate(laterals[i],
+                                                 **self.upsample_cfg)
+            else:
+                prev_shape = laterals[i - 1].shape[2:]
+                laterals[i - 1] += F.interpolate(
+                    laterals[i], size=prev_shape, **self.upsample_cfg)
+
+        # build outputs
+        # part 1: from original levels
+        outs = [
+            laterals[i] for i in range(used_backbone_levels)
+        ]
+        # part 2: add extra levels
+        if self.num_outs > len(outs):
+            # use max pool to get more levels on top of outputs
+            # (e.g., Faster R-CNN, Mask R-CNN)
+            for i in range(self.num_outs - used_backbone_levels):
+                outs.append(F.max_pool2d(outs[-1], 2, stride=2))
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/nas_fpn.py b/detection/scrfd/mmdet/models/necks/nas_fpn.py
new file mode 100755
index 000000000..8e333ce65
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/nas_fpn.py
@@ -0,0 +1,160 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, caffe2_xavier_init
+from mmcv.ops.merge_cells import GlobalPoolingCell, SumCell
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class NASFPN(nn.Module):
+    """NAS-FPN.
+
+    Implementation of `NAS-FPN: Learning Scalable Feature Pyramid Architecture
+    for Object Detection <https://arxiv.org/abs/1904.07392>`_
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale)
+        num_outs (int): Number of output scales.
+        stack_times (int): The number of times the pyramid architecture will
+            be stacked.
+        start_level (int): Index of the start input backbone level used to
+            build the feature pyramid. Default: 0.
+        end_level (int): Index of the end input backbone level (exclusive) to
+            build the feature pyramid. Default: -1, which means the last level.
+        add_extra_convs (bool): It decides whether to add conv
+            layers on top of the original feature maps. Default to False.
+            If True, its actual mode is specified by `extra_convs_on_inputs`.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 stack_times,
+                 start_level=0,
+                 end_level=-1,
+                 add_extra_convs=False,
+                 norm_cfg=None):
+        super(NASFPN, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)  # num of input feature levels
+        self.num_outs = num_outs  # num of output feature levels
+        self.stack_times = stack_times
+        self.norm_cfg = norm_cfg
+
+        if end_level == -1:
+            self.backbone_end_level = self.num_ins
+            assert num_outs >= self.num_ins - start_level
+        else:
+            # if end_level < inputs, no extra level is allowed
+            self.backbone_end_level = end_level
+            assert end_level <= len(in_channels)
+            assert num_outs == end_level - start_level
+        self.start_level = start_level
+        self.end_level = end_level
+        self.add_extra_convs = add_extra_convs
+
+        # add lateral connections
+        self.lateral_convs = nn.ModuleList()
+        for i in range(self.start_level, self.backbone_end_level):
+            l_conv = ConvModule(
+                in_channels[i],
+                out_channels,
+                1,
+                norm_cfg=norm_cfg,
+                act_cfg=None)
+            self.lateral_convs.append(l_conv)
+
+        # add extra downsample layers (stride-2 pooling or conv)
+        extra_levels = num_outs - self.backbone_end_level + self.start_level
+        self.extra_downsamples = nn.ModuleList()
+        for i in range(extra_levels):
+            extra_conv = ConvModule(
+                out_channels, out_channels, 1, norm_cfg=norm_cfg, act_cfg=None)
+            self.extra_downsamples.append(
+                nn.Sequential(extra_conv, nn.MaxPool2d(2, 2)))
+
+        # add NAS FPN connections
+        self.fpn_stages = nn.ModuleList()
+        for _ in range(self.stack_times):
+            stage = nn.ModuleDict()
+            # gp(p6, p4) -> p4_1
+            stage['gp_64_4'] = GlobalPoolingCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # sum(p4_1, p4) -> p4_2
+            stage['sum_44_4'] = SumCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # sum(p4_2, p3) -> p3_out
+            stage['sum_43_3'] = SumCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # sum(p3_out, p4_2) -> p4_out
+            stage['sum_34_4'] = SumCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # sum(p5, gp(p4_out, p3_out)) -> p5_out
+            stage['gp_43_5'] = GlobalPoolingCell(with_out_conv=False)
+            stage['sum_55_5'] = SumCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # sum(p7, gp(p5_out, p4_2)) -> p7_out
+            stage['gp_54_7'] = GlobalPoolingCell(with_out_conv=False)
+            stage['sum_77_7'] = SumCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            # gp(p7_out, p5_out) -> p6_out
+            stage['gp_75_6'] = GlobalPoolingCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                out_norm_cfg=norm_cfg)
+            self.fpn_stages.append(stage)
+
+    def init_weights(self):
+        """Initialize the weights of module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                caffe2_xavier_init(m)
+
+    def forward(self, inputs):
+        """Forward function."""
+        # build P3-P5
+        feats = [
+            lateral_conv(inputs[i + self.start_level])
+            for i, lateral_conv in enumerate(self.lateral_convs)
+        ]
+        # build P6-P7 on top of P5
+        for downsample in self.extra_downsamples:
+            feats.append(downsample(feats[-1]))
+
+        p3, p4, p5, p6, p7 = feats
+
+        for stage in self.fpn_stages:
+            # gp(p6, p4) -> p4_1
+            p4_1 = stage['gp_64_4'](p6, p4, out_size=p4.shape[-2:])
+            # sum(p4_1, p4) -> p4_2
+            p4_2 = stage['sum_44_4'](p4_1, p4, out_size=p4.shape[-2:])
+            # sum(p4_2, p3) -> p3_out
+            p3 = stage['sum_43_3'](p4_2, p3, out_size=p3.shape[-2:])
+            # sum(p3_out, p4_2) -> p4_out
+            p4 = stage['sum_34_4'](p3, p4_2, out_size=p4.shape[-2:])
+            # sum(p5, gp(p4_out, p3_out)) -> p5_out
+            p5_tmp = stage['gp_43_5'](p4, p3, out_size=p5.shape[-2:])
+            p5 = stage['sum_55_5'](p5, p5_tmp, out_size=p5.shape[-2:])
+            # sum(p7, gp(p5_out, p4_2)) -> p7_out
+            p7_tmp = stage['gp_54_7'](p5, p4_2, out_size=p7.shape[-2:])
+            p7 = stage['sum_77_7'](p7, p7_tmp, out_size=p7.shape[-2:])
+            # gp(p7_out, p5_out) -> p6_out
+            p6 = stage['gp_75_6'](p7, p5, out_size=p6.shape[-2:])
+
+        return p3, p4, p5, p6, p7
diff --git a/detection/scrfd/mmdet/models/necks/nasfcos_fpn.py b/detection/scrfd/mmdet/models/necks/nasfcos_fpn.py
new file mode 100755
index 000000000..2daf79ef5
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/nasfcos_fpn.py
@@ -0,0 +1,161 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, caffe2_xavier_init
+from mmcv.ops.merge_cells import ConcatCell
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class NASFCOS_FPN(nn.Module):
+    """FPN structure in NASFPN.
+
+    Implementation of paper `NAS-FCOS: Fast Neural Architecture Search for
+    Object Detection <https://arxiv.org/abs/1906.04423>`_
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale)
+        num_outs (int): Number of output scales.
+        start_level (int): Index of the start input backbone level used to
+            build the feature pyramid. Default: 0.
+        end_level (int): Index of the end input backbone level (exclusive) to
+            build the feature pyramid. Default: -1, which means the last level.
+        add_extra_convs (bool): It decides whether to add conv
+            layers on top of the original feature maps. Default to False.
+            If True, its actual mode is specified by `extra_convs_on_inputs`.
+        conv_cfg (dict): dictionary to construct and config conv layer.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 start_level=1,
+                 end_level=-1,
+                 add_extra_convs=False,
+                 conv_cfg=None,
+                 norm_cfg=None):
+        super(NASFCOS_FPN, self).__init__()
+        assert isinstance(in_channels, list)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_ins = len(in_channels)
+        self.num_outs = num_outs
+        self.norm_cfg = norm_cfg
+        self.conv_cfg = conv_cfg
+
+        if end_level == -1:
+            self.backbone_end_level = self.num_ins
+            assert num_outs >= self.num_ins - start_level
+        else:
+            self.backbone_end_level = end_level
+            assert end_level <= len(in_channels)
+            assert num_outs == end_level - start_level
+        self.start_level = start_level
+        self.end_level = end_level
+        self.add_extra_convs = add_extra_convs
+
+        self.adapt_convs = nn.ModuleList()
+        for i in range(self.start_level, self.backbone_end_level):
+            adapt_conv = ConvModule(
+                in_channels[i],
+                out_channels,
+                1,
+                stride=1,
+                padding=0,
+                bias=False,
+                norm_cfg=dict(type='BN'),
+                act_cfg=dict(type='ReLU', inplace=False))
+            self.adapt_convs.append(adapt_conv)
+
+        # C2 is omitted according to the paper
+        extra_levels = num_outs - self.backbone_end_level + self.start_level
+
+        def build_concat_cell(with_input1_conv, with_input2_conv):
+            cell_conv_cfg = dict(
+                kernel_size=1, padding=0, bias=False, groups=out_channels)
+            return ConcatCell(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                with_out_conv=True,
+                out_conv_cfg=cell_conv_cfg,
+                out_norm_cfg=dict(type='BN'),
+                out_conv_order=('norm', 'act', 'conv'),
+                with_input1_conv=with_input1_conv,
+                with_input2_conv=with_input2_conv,
+                input_conv_cfg=conv_cfg,
+                input_norm_cfg=norm_cfg,
+                upsample_mode='nearest')
+
+        # Denote c3=f0, c4=f1, c5=f2 for convince
+        self.fpn = nn.ModuleDict()
+        self.fpn['c22_1'] = build_concat_cell(True, True)
+        self.fpn['c22_2'] = build_concat_cell(True, True)
+        self.fpn['c32'] = build_concat_cell(True, False)
+        self.fpn['c02'] = build_concat_cell(True, False)
+        self.fpn['c42'] = build_concat_cell(True, True)
+        self.fpn['c36'] = build_concat_cell(True, True)
+        self.fpn['c61'] = build_concat_cell(True, True)  # f9
+        self.extra_downsamples = nn.ModuleList()
+        for i in range(extra_levels):
+            extra_act_cfg = None if i == 0 \
+                else dict(type='ReLU', inplace=False)
+            self.extra_downsamples.append(
+                ConvModule(
+                    out_channels,
+                    out_channels,
+                    3,
+                    stride=2,
+                    padding=1,
+                    act_cfg=extra_act_cfg,
+                    order=('act', 'norm', 'conv')))
+
+    def forward(self, inputs):
+        """Forward function."""
+        feats = [
+            adapt_conv(inputs[i + self.start_level])
+            for i, adapt_conv in enumerate(self.adapt_convs)
+        ]
+
+        for (i, module_name) in enumerate(self.fpn):
+            idx_1, idx_2 = int(module_name[1]), int(module_name[2])
+            res = self.fpn[module_name](feats[idx_1], feats[idx_2])
+            feats.append(res)
+
+        ret = []
+        for (idx, input_idx) in zip([9, 8, 7], [1, 2, 3]):  # add P3, P4, P5
+            feats1, feats2 = feats[idx], feats[5]
+            feats2_resize = F.interpolate(
+                feats2,
+                size=feats1.size()[2:],
+                mode='bilinear',
+                align_corners=False)
+
+            feats_sum = feats1 + feats2_resize
+            ret.append(
+                F.interpolate(
+                    feats_sum,
+                    size=inputs[input_idx].size()[2:],
+                    mode='bilinear',
+                    align_corners=False))
+
+        for submodule in self.extra_downsamples:
+            ret.append(submodule(ret[-1]))
+
+        return tuple(ret)
+
+    def init_weights(self):
+        """Initialize the weights of module."""
+        for module in self.fpn.values():
+            if hasattr(module, 'conv_out'):
+                caffe2_xavier_init(module.out_conv.conv)
+
+        for modules in [
+                self.adapt_convs.modules(),
+                self.extra_downsamples.modules()
+        ]:
+            for module in modules:
+                if isinstance(module, nn.Conv2d):
+                    caffe2_xavier_init(module)
diff --git a/detection/scrfd/mmdet/models/necks/pafpn.py b/detection/scrfd/mmdet/models/necks/pafpn.py
new file mode 100755
index 000000000..d7c0b50f2
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/pafpn.py
@@ -0,0 +1,142 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule
+from mmcv.runner import auto_fp16
+
+from ..builder import NECKS
+from .fpn import FPN
+
+
+@NECKS.register_module()
+class PAFPN(FPN):
+    """Path Aggregation Network for Instance Segmentation.
+
+    This is an implementation of the `PAFPN in Path Aggregation Network
+    <https://arxiv.org/abs/1803.01534>`_.
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale)
+        num_outs (int): Number of output scales.
+        start_level (int): Index of the start input backbone level used to
+            build the feature pyramid. Default: 0.
+        end_level (int): Index of the end input backbone level (exclusive) to
+            build the feature pyramid. Default: -1, which means the last level.
+        add_extra_convs (bool): Whether to add conv layers on top of the
+            original feature maps. Default: False.
+        extra_convs_on_inputs (bool): Whether to apply extra conv on
+            the original feature from the backbone. Default: False.
+        relu_before_extra_convs (bool): Whether to apply relu before the extra
+            conv. Default: False.
+        no_norm_on_lateral (bool): Whether to apply norm on lateral.
+            Default: False.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Config dict for normalization layer. Default: None.
+        act_cfg (str): Config dict for activation layer in ConvModule.
+            Default: None.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 num_outs,
+                 start_level=0,
+                 end_level=-1,
+                 add_extra_convs=False,
+                 extra_convs_on_inputs=True,
+                 relu_before_extra_convs=False,
+                 no_norm_on_lateral=False,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 act_cfg=None):
+        super(PAFPN,
+              self).__init__(in_channels, out_channels, num_outs, start_level,
+                             end_level, add_extra_convs, extra_convs_on_inputs,
+                             relu_before_extra_convs, no_norm_on_lateral,
+                             conv_cfg, norm_cfg, act_cfg)
+        # add extra bottom up pathway
+        self.downsample_convs = nn.ModuleList()
+        self.pafpn_convs = nn.ModuleList()
+        for i in range(self.start_level + 1, self.backbone_end_level):
+            d_conv = ConvModule(
+                out_channels,
+                out_channels,
+                3,
+                stride=2,
+                padding=1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                act_cfg=act_cfg,
+                inplace=False)
+            pafpn_conv = ConvModule(
+                out_channels,
+                out_channels,
+                3,
+                padding=1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                act_cfg=act_cfg,
+                inplace=False)
+            self.downsample_convs.append(d_conv)
+            self.pafpn_convs.append(pafpn_conv)
+
+    @auto_fp16()
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.in_channels)
+
+        # build laterals
+        laterals = [
+            lateral_conv(inputs[i + self.start_level])
+            for i, lateral_conv in enumerate(self.lateral_convs)
+        ]
+
+        # build top-down path
+        used_backbone_levels = len(laterals)
+        for i in range(used_backbone_levels - 1, 0, -1):
+            prev_shape = laterals[i - 1].shape[2:]
+            laterals[i - 1] += F.interpolate(
+                laterals[i], size=prev_shape, mode='nearest')
+
+        # build outputs
+        # part 1: from original levels
+        inter_outs = [
+            self.fpn_convs[i](laterals[i]) for i in range(used_backbone_levels)
+        ]
+
+        # part 2: add bottom-up path
+        for i in range(0, used_backbone_levels - 1):
+            inter_outs[i + 1] += self.downsample_convs[i](inter_outs[i])
+
+        outs = []
+        outs.append(inter_outs[0])
+        outs.extend([
+            self.pafpn_convs[i - 1](inter_outs[i])
+            for i in range(1, used_backbone_levels)
+        ])
+
+        # part 3: add extra levels
+        if self.num_outs > len(outs):
+            # use max pool to get more levels on top of outputs
+            # (e.g., Faster R-CNN, Mask R-CNN)
+            if not self.add_extra_convs:
+                for i in range(self.num_outs - used_backbone_levels):
+                    outs.append(F.max_pool2d(outs[-1], 1, stride=2))
+            # add conv layers on top of original feature maps (RetinaNet)
+            else:
+                if self.add_extra_convs == 'on_input':
+                    orig = inputs[self.backbone_end_level - 1]
+                    outs.append(self.fpn_convs[used_backbone_levels](orig))
+                elif self.add_extra_convs == 'on_lateral':
+                    outs.append(self.fpn_convs[used_backbone_levels](
+                        laterals[-1]))
+                elif self.add_extra_convs == 'on_output':
+                    outs.append(self.fpn_convs[used_backbone_levels](outs[-1]))
+                else:
+                    raise NotImplementedError
+                for i in range(used_backbone_levels + 1, self.num_outs):
+                    if self.relu_before_extra_convs:
+                        outs.append(self.fpn_convs[i](F.relu(outs[-1])))
+                    else:
+                        outs.append(self.fpn_convs[i](outs[-1]))
+        return tuple(outs)
diff --git a/detection/scrfd/mmdet/models/necks/rfp.py b/detection/scrfd/mmdet/models/necks/rfp.py
new file mode 100755
index 000000000..8a63e63bd
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/rfp.py
@@ -0,0 +1,128 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import constant_init, kaiming_init, xavier_init
+
+from ..builder import NECKS, build_backbone
+from .fpn import FPN
+
+
+class ASPP(nn.Module):
+    """ASPP (Atrous Spatial Pyramid Pooling)
+
+    This is an implementation of the ASPP module used in DetectoRS
+    (https://arxiv.org/pdf/2006.02334.pdf)
+
+    Args:
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of channels produced by this module
+        dilations (tuple[int]): Dilations of the four branches.
+            Default: (1, 3, 6, 1)
+    """
+
+    def __init__(self, in_channels, out_channels, dilations=(1, 3, 6, 1)):
+        super().__init__()
+        assert dilations[-1] == 1
+        self.aspp = nn.ModuleList()
+        for dilation in dilations:
+            kernel_size = 3 if dilation > 1 else 1
+            padding = dilation if dilation > 1 else 0
+            conv = nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=kernel_size,
+                stride=1,
+                dilation=dilation,
+                padding=padding,
+                bias=True)
+            self.aspp.append(conv)
+        self.gap = nn.AdaptiveAvgPool2d(1)
+        self.init_weights()
+
+    def init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                kaiming_init(m)
+
+    def forward(self, x):
+        avg_x = self.gap(x)
+        out = []
+        for aspp_idx in range(len(self.aspp)):
+            inp = avg_x if (aspp_idx == len(self.aspp) - 1) else x
+            out.append(F.relu_(self.aspp[aspp_idx](inp)))
+        out[-1] = out[-1].expand_as(out[-2])
+        out = torch.cat(out, dim=1)
+        return out
+
+
+@NECKS.register_module()
+class RFP(FPN):
+    """RFP (Recursive Feature Pyramid)
+
+    This is an implementation of RFP in `DetectoRS
+    <https://arxiv.org/pdf/2006.02334.pdf>`_. Different from standard FPN, the
+    input of RFP should be multi level features along with origin input image
+    of backbone.
+
+    Args:
+        rfp_steps (int): Number of unrolled steps of RFP.
+        rfp_backbone (dict): Configuration of the backbone for RFP.
+        aspp_out_channels (int): Number of output channels of ASPP module.
+        aspp_dilations (tuple[int]): Dilation rates of four branches.
+            Default: (1, 3, 6, 1)
+    """
+
+    def __init__(self,
+                 rfp_steps,
+                 rfp_backbone,
+                 aspp_out_channels,
+                 aspp_dilations=(1, 3, 6, 1),
+                 **kwargs):
+        super().__init__(**kwargs)
+        self.rfp_steps = rfp_steps
+        self.rfp_modules = nn.ModuleList()
+        for rfp_idx in range(1, rfp_steps):
+            rfp_module = build_backbone(rfp_backbone)
+            self.rfp_modules.append(rfp_module)
+        self.rfp_aspp = ASPP(self.out_channels, aspp_out_channels,
+                             aspp_dilations)
+        self.rfp_weight = nn.Conv2d(
+            self.out_channels,
+            1,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True)
+
+    def init_weights(self):
+        # Avoid using super().init_weights(), which may alter the default
+        # initialization of the modules in self.rfp_modules that have missing
+        # keys in the pretrained checkpoint.
+        for convs in [self.lateral_convs, self.fpn_convs]:
+            for m in convs.modules():
+                if isinstance(m, nn.Conv2d):
+                    xavier_init(m, distribution='uniform')
+        for rfp_idx in range(self.rfp_steps - 1):
+            self.rfp_modules[rfp_idx].init_weights(
+                self.rfp_modules[rfp_idx].pretrained)
+        constant_init(self.rfp_weight, 0)
+
+    def forward(self, inputs):
+        inputs = list(inputs)
+        assert len(inputs) == len(self.in_channels) + 1  # +1 for input image
+        img = inputs.pop(0)
+        # FPN forward
+        x = super().forward(tuple(inputs))
+        for rfp_idx in range(self.rfp_steps - 1):
+            rfp_feats = [x[0]] + list(
+                self.rfp_aspp(x[i]) for i in range(1, len(x)))
+            x_idx = self.rfp_modules[rfp_idx].rfp_forward(img, rfp_feats)
+            # FPN forward
+            x_idx = super().forward(x_idx)
+            x_new = []
+            for ft_idx in range(len(x_idx)):
+                add_weight = torch.sigmoid(self.rfp_weight(x_idx[ft_idx]))
+                x_new.append(add_weight * x_idx[ft_idx] +
+                             (1 - add_weight) * x[ft_idx])
+            x = x_new
+        return x
diff --git a/detection/scrfd/mmdet/models/necks/yolo_neck.py b/detection/scrfd/mmdet/models/necks/yolo_neck.py
new file mode 100755
index 000000000..c2f9b9ef3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/necks/yolo_neck.py
@@ -0,0 +1,136 @@
+# Copyright (c) 2019 Western Digital Corporation or its affiliates.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule
+
+from ..builder import NECKS
+
+
+class DetectionBlock(nn.Module):
+    """Detection block in YOLO neck.
+
+    Let out_channels = n, the DetectionBlock contains:
+    Six ConvLayers, 1 Conv2D Layer and 1 YoloLayer.
+    The first 6 ConvLayers are formed the following way:
+        1x1xn, 3x3x2n, 1x1xn, 3x3x2n, 1x1xn, 3x3x2n.
+    The Conv2D layer is 1x1x255.
+    Some block will have branch after the fifth ConvLayer.
+    The input channel is arbitrary (in_channels)
+
+    Args:
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+            Default: dict(type='BN', requires_grad=True)
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='LeakyReLU', negative_slope=0.1).
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 act_cfg=dict(type='LeakyReLU', negative_slope=0.1)):
+        super(DetectionBlock, self).__init__()
+        double_out_channels = out_channels * 2
+
+        # shortcut
+        cfg = dict(conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg)
+        self.conv1 = ConvModule(in_channels, out_channels, 1, **cfg)
+        self.conv2 = ConvModule(
+            out_channels, double_out_channels, 3, padding=1, **cfg)
+        self.conv3 = ConvModule(double_out_channels, out_channels, 1, **cfg)
+        self.conv4 = ConvModule(
+            out_channels, double_out_channels, 3, padding=1, **cfg)
+        self.conv5 = ConvModule(double_out_channels, out_channels, 1, **cfg)
+
+    def forward(self, x):
+        tmp = self.conv1(x)
+        tmp = self.conv2(tmp)
+        tmp = self.conv3(tmp)
+        tmp = self.conv4(tmp)
+        out = self.conv5(tmp)
+        return out
+
+
+@NECKS.register_module()
+class YOLOV3Neck(nn.Module):
+    """The neck of YOLOV3.
+
+    It can be treated as a simplified version of FPN. It
+    will take the result from Darknet backbone and do some upsampling and
+    concatenation. It will finally output the detection result.
+
+    Note:
+        The input feats should be from top to bottom.
+            i.e., from high-lvl to low-lvl
+        But YOLOV3Neck will process them in reversed order.
+            i.e., from bottom (high-lvl) to top (low-lvl)
+
+    Args:
+        num_scales (int): The number of scales / stages.
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        conv_cfg (dict): Config dict for convolution layer. Default: None.
+        norm_cfg (dict): Dictionary to construct and config norm layer.
+            Default: dict(type='BN', requires_grad=True)
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='LeakyReLU', negative_slope=0.1).
+    """
+
+    def __init__(self,
+                 num_scales,
+                 in_channels,
+                 out_channels,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 act_cfg=dict(type='LeakyReLU', negative_slope=0.1)):
+        super(YOLOV3Neck, self).__init__()
+        assert (num_scales == len(in_channels) == len(out_channels))
+        self.num_scales = num_scales
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        # shortcut
+        cfg = dict(conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg)
+
+        # To support arbitrary scales, the code looks awful, but it works.
+        # Better solution is welcomed.
+        self.detect1 = DetectionBlock(in_channels[0], out_channels[0], **cfg)
+        for i in range(1, self.num_scales):
+            in_c, out_c = self.in_channels[i], self.out_channels[i]
+            self.add_module(f'conv{i}', ConvModule(in_c, out_c, 1, **cfg))
+            # in_c + out_c : High-lvl feats will be cat with low-lvl feats
+            self.add_module(f'detect{i+1}',
+                            DetectionBlock(in_c + out_c, out_c, **cfg))
+
+    def forward(self, feats):
+        assert len(feats) == self.num_scales
+
+        # processed from bottom (high-lvl) to top (low-lvl)
+        outs = []
+        out = self.detect1(feats[-1])
+        outs.append(out)
+
+        for i, x in enumerate(reversed(feats[:-1])):
+            conv = getattr(self, f'conv{i+1}')
+            tmp = conv(out)
+
+            # Cat with low-lvl feats
+            tmp = F.interpolate(tmp, scale_factor=2)
+            tmp = torch.cat((tmp, x), 1)
+
+            detect = getattr(self, f'detect{i+2}')
+            out = detect(tmp)
+            outs.append(out)
+
+        return tuple(outs)
+
+    def init_weights(self):
+        """Initialize the weights of module."""
+        # init is done in ConvModule
+        pass
diff --git a/detection/scrfd/mmdet/models/roi_heads/__init__.py b/detection/scrfd/mmdet/models/roi_heads/__init__.py
new file mode 100755
index 000000000..e08c3b928
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/__init__.py
@@ -0,0 +1,27 @@
+from .base_roi_head import BaseRoIHead
+from .bbox_heads import (BBoxHead, ConvFCBBoxHead, DoubleConvFCBBoxHead,
+                         Shared2FCBBoxHead, Shared4Conv1FCBBoxHead)
+from .cascade_roi_head import CascadeRoIHead
+from .double_roi_head import DoubleHeadRoIHead
+from .dynamic_roi_head import DynamicRoIHead
+from .grid_roi_head import GridRoIHead
+from .htc_roi_head import HybridTaskCascadeRoIHead
+from .mask_heads import (CoarseMaskHead, FCNMaskHead, FusedSemanticHead,
+                         GridHead, HTCMaskHead, MaskIoUHead, MaskPointHead)
+from .mask_scoring_roi_head import MaskScoringRoIHead
+from .pisa_roi_head import PISARoIHead
+from .point_rend_roi_head import PointRendRoIHead
+from .roi_extractors import SingleRoIExtractor
+from .shared_heads import ResLayer
+from .standard_roi_head import StandardRoIHead
+from .trident_roi_head import TridentRoIHead
+
+__all__ = [
+    'BaseRoIHead', 'CascadeRoIHead', 'DoubleHeadRoIHead', 'MaskScoringRoIHead',
+    'HybridTaskCascadeRoIHead', 'GridRoIHead', 'ResLayer', 'BBoxHead',
+    'ConvFCBBoxHead', 'Shared2FCBBoxHead', 'StandardRoIHead',
+    'Shared4Conv1FCBBoxHead', 'DoubleConvFCBBoxHead', 'FCNMaskHead',
+    'HTCMaskHead', 'FusedSemanticHead', 'GridHead', 'MaskIoUHead',
+    'SingleRoIExtractor', 'PISARoIHead', 'PointRendRoIHead', 'MaskPointHead',
+    'CoarseMaskHead', 'DynamicRoIHead', 'TridentRoIHead'
+]
diff --git a/detection/scrfd/mmdet/models/roi_heads/base_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/base_roi_head.py
new file mode 100755
index 000000000..ec027dda8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/base_roi_head.py
@@ -0,0 +1,106 @@
+from abc import ABCMeta, abstractmethod
+
+import torch.nn as nn
+
+from ..builder import build_shared_head
+
+
+class BaseRoIHead(nn.Module, metaclass=ABCMeta):
+    """Base class for RoIHeads."""
+
+    def __init__(self,
+                 bbox_roi_extractor=None,
+                 bbox_head=None,
+                 mask_roi_extractor=None,
+                 mask_head=None,
+                 shared_head=None,
+                 train_cfg=None,
+                 test_cfg=None):
+        super(BaseRoIHead, self).__init__()
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+        if shared_head is not None:
+            self.shared_head = build_shared_head(shared_head)
+
+        if bbox_head is not None:
+            self.init_bbox_head(bbox_roi_extractor, bbox_head)
+
+        if mask_head is not None:
+            self.init_mask_head(mask_roi_extractor, mask_head)
+
+        self.init_assigner_sampler()
+
+    @property
+    def with_bbox(self):
+        """bool: whether the RoI head contains a `bbox_head`"""
+        return hasattr(self, 'bbox_head') and self.bbox_head is not None
+
+    @property
+    def with_mask(self):
+        """bool: whether the RoI head contains a `mask_head`"""
+        return hasattr(self, 'mask_head') and self.mask_head is not None
+
+    @property
+    def with_shared_head(self):
+        """bool: whether the RoI head contains a `shared_head`"""
+        return hasattr(self, 'shared_head') and self.shared_head is not None
+
+    @abstractmethod
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        pass
+
+    @abstractmethod
+    def init_bbox_head(self):
+        """Initialize ``bbox_head``"""
+        pass
+
+    @abstractmethod
+    def init_mask_head(self):
+        """Initialize ``mask_head``"""
+        pass
+
+    @abstractmethod
+    def init_assigner_sampler(self):
+        """Initialize assigner and sampler."""
+        pass
+
+    @abstractmethod
+    def forward_train(self,
+                      x,
+                      img_meta,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None,
+                      **kwargs):
+        """Forward function during training."""
+        pass
+
+    async def async_simple_test(self, x, img_meta, **kwargs):
+        """Asynchronized test function."""
+        raise NotImplementedError
+
+    def simple_test(self,
+                    x,
+                    proposal_list,
+                    img_meta,
+                    proposals=None,
+                    rescale=False,
+                    **kwargs):
+        """Test without augmentation."""
+        pass
+
+    def aug_test(self, x, proposal_list, img_metas, rescale=False, **kwargs):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        pass
diff --git a/detection/scrfd/mmdet/models/roi_heads/bbox_heads/__init__.py b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/__init__.py
new file mode 100755
index 000000000..acfb71ebe
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/__init__.py
@@ -0,0 +1,10 @@
+from .bbox_head import BBoxHead
+from .convfc_bbox_head import (ConvFCBBoxHead, Shared2FCBBoxHead,
+                               Shared4Conv1FCBBoxHead)
+from .double_bbox_head import DoubleConvFCBBoxHead
+from .sabl_head import SABLHead
+
+__all__ = [
+    'BBoxHead', 'ConvFCBBoxHead', 'Shared2FCBBoxHead',
+    'Shared4Conv1FCBBoxHead', 'DoubleConvFCBBoxHead', 'SABLHead'
+]
diff --git a/detection/scrfd/mmdet/models/roi_heads/bbox_heads/bbox_head.py b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/bbox_head.py
new file mode 100755
index 000000000..e0931e176
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/bbox_head.py
@@ -0,0 +1,335 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.runner import auto_fp16, force_fp32
+from torch.nn.modules.utils import _pair
+
+from mmdet.core import build_bbox_coder, multi_apply, multiclass_nms
+from mmdet.models.builder import HEADS, build_loss
+from mmdet.models.losses import accuracy
+
+
+@HEADS.register_module()
+class BBoxHead(nn.Module):
+    """Simplest RoI head, with only two fc layers for classification and
+    regression respectively."""
+
+    def __init__(self,
+                 with_avg_pool=False,
+                 with_cls=True,
+                 with_reg=True,
+                 roi_feat_size=7,
+                 in_channels=256,
+                 num_classes=80,
+                 bbox_coder=dict(
+                     type='DeltaXYWHBBoxCoder',
+                     clip_border=True,
+                     target_means=[0., 0., 0., 0.],
+                     target_stds=[0.1, 0.1, 0.2, 0.2]),
+                 reg_class_agnostic=False,
+                 reg_decoded_bbox=False,
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=False,
+                     loss_weight=1.0),
+                 loss_bbox=dict(
+                     type='SmoothL1Loss', beta=1.0, loss_weight=1.0)):
+        super(BBoxHead, self).__init__()
+        assert with_cls or with_reg
+        self.with_avg_pool = with_avg_pool
+        self.with_cls = with_cls
+        self.with_reg = with_reg
+        self.roi_feat_size = _pair(roi_feat_size)
+        self.roi_feat_area = self.roi_feat_size[0] * self.roi_feat_size[1]
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.reg_class_agnostic = reg_class_agnostic
+        self.reg_decoded_bbox = reg_decoded_bbox
+        self.fp16_enabled = False
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox = build_loss(loss_bbox)
+
+        in_channels = self.in_channels
+        if self.with_avg_pool:
+            self.avg_pool = nn.AvgPool2d(self.roi_feat_size)
+        else:
+            in_channels *= self.roi_feat_area
+        if self.with_cls:
+            # need to add background class
+            self.fc_cls = nn.Linear(in_channels, num_classes + 1)
+        if self.with_reg:
+            out_dim_reg = 4 if reg_class_agnostic else 4 * num_classes
+            self.fc_reg = nn.Linear(in_channels, out_dim_reg)
+        self.debug_imgs = None
+
+    def init_weights(self):
+        # conv layers are already initialized by ConvModule
+        if self.with_cls:
+            nn.init.normal_(self.fc_cls.weight, 0, 0.01)
+            nn.init.constant_(self.fc_cls.bias, 0)
+        if self.with_reg:
+            nn.init.normal_(self.fc_reg.weight, 0, 0.001)
+            nn.init.constant_(self.fc_reg.bias, 0)
+
+    @auto_fp16()
+    def forward(self, x):
+        if self.with_avg_pool:
+            x = self.avg_pool(x)
+        x = x.view(x.size(0), -1)
+        cls_score = self.fc_cls(x) if self.with_cls else None
+        bbox_pred = self.fc_reg(x) if self.with_reg else None
+        return cls_score, bbox_pred
+
+    def _get_target_single(self, pos_bboxes, neg_bboxes, pos_gt_bboxes,
+                           pos_gt_labels, cfg):
+        num_pos = pos_bboxes.size(0)
+        num_neg = neg_bboxes.size(0)
+        num_samples = num_pos + num_neg
+
+        # original implementation uses new_zeros since BG are set to be 0
+        # now use empty & fill because BG cat_id = num_classes,
+        # FG cat_id = [0, num_classes-1]
+        labels = pos_bboxes.new_full((num_samples, ),
+                                     self.num_classes,
+                                     dtype=torch.long)
+        label_weights = pos_bboxes.new_zeros(num_samples)
+        bbox_targets = pos_bboxes.new_zeros(num_samples, 4)
+        bbox_weights = pos_bboxes.new_zeros(num_samples, 4)
+        if num_pos > 0:
+            labels[:num_pos] = pos_gt_labels
+            pos_weight = 1.0 if cfg.pos_weight <= 0 else cfg.pos_weight
+            label_weights[:num_pos] = pos_weight
+            if not self.reg_decoded_bbox:
+                pos_bbox_targets = self.bbox_coder.encode(
+                    pos_bboxes, pos_gt_bboxes)
+            else:
+                pos_bbox_targets = pos_gt_bboxes
+            bbox_targets[:num_pos, :] = pos_bbox_targets
+            bbox_weights[:num_pos, :] = 1
+        if num_neg > 0:
+            label_weights[-num_neg:] = 1.0
+
+        return labels, label_weights, bbox_targets, bbox_weights
+
+    def get_targets(self,
+                    sampling_results,
+                    gt_bboxes,
+                    gt_labels,
+                    rcnn_train_cfg,
+                    concat=True):
+        pos_bboxes_list = [res.pos_bboxes for res in sampling_results]
+        neg_bboxes_list = [res.neg_bboxes for res in sampling_results]
+        pos_gt_bboxes_list = [res.pos_gt_bboxes for res in sampling_results]
+        pos_gt_labels_list = [res.pos_gt_labels for res in sampling_results]
+        labels, label_weights, bbox_targets, bbox_weights = multi_apply(
+            self._get_target_single,
+            pos_bboxes_list,
+            neg_bboxes_list,
+            pos_gt_bboxes_list,
+            pos_gt_labels_list,
+            cfg=rcnn_train_cfg)
+
+        if concat:
+            labels = torch.cat(labels, 0)
+            label_weights = torch.cat(label_weights, 0)
+            bbox_targets = torch.cat(bbox_targets, 0)
+            bbox_weights = torch.cat(bbox_weights, 0)
+        return labels, label_weights, bbox_targets, bbox_weights
+
+    @force_fp32(apply_to=('cls_score', 'bbox_pred'))
+    def loss(self,
+             cls_score,
+             bbox_pred,
+             rois,
+             labels,
+             label_weights,
+             bbox_targets,
+             bbox_weights,
+             reduction_override=None):
+        losses = dict()
+        if cls_score is not None:
+            avg_factor = max(torch.sum(label_weights > 0).float().item(), 1.)
+            if cls_score.numel() > 0:
+                losses['loss_cls'] = self.loss_cls(
+                    cls_score,
+                    labels,
+                    label_weights,
+                    avg_factor=avg_factor,
+                    reduction_override=reduction_override)
+                losses['acc'] = accuracy(cls_score, labels)
+        if bbox_pred is not None:
+            bg_class_ind = self.num_classes
+            # 0~self.num_classes-1 are FG, self.num_classes is BG
+            pos_inds = (labels >= 0) & (labels < bg_class_ind)
+            # do not perform bounding box regression for BG anymore.
+            if pos_inds.any():
+                if self.reg_decoded_bbox:
+                    bbox_pred = self.bbox_coder.decode(rois[:, 1:], bbox_pred)
+                if self.reg_class_agnostic:
+                    pos_bbox_pred = bbox_pred.view(
+                        bbox_pred.size(0), 4)[pos_inds.type(torch.bool)]
+                else:
+                    pos_bbox_pred = bbox_pred.view(
+                        bbox_pred.size(0), -1,
+                        4)[pos_inds.type(torch.bool),
+                           labels[pos_inds.type(torch.bool)]]
+                losses['loss_bbox'] = self.loss_bbox(
+                    pos_bbox_pred,
+                    bbox_targets[pos_inds.type(torch.bool)],
+                    bbox_weights[pos_inds.type(torch.bool)],
+                    avg_factor=bbox_targets.size(0),
+                    reduction_override=reduction_override)
+            else:
+                losses['loss_bbox'] = bbox_pred[pos_inds].sum()
+        return losses
+
+    @force_fp32(apply_to=('cls_score', 'bbox_pred'))
+    def get_bboxes(self,
+                   rois,
+                   cls_score,
+                   bbox_pred,
+                   img_shape,
+                   scale_factor,
+                   rescale=False,
+                   cfg=None):
+        if isinstance(cls_score, list):
+            cls_score = sum(cls_score) / float(len(cls_score))
+        scores = F.softmax(cls_score, dim=1) if cls_score is not None else None
+
+        if bbox_pred is not None:
+            bboxes = self.bbox_coder.decode(
+                rois[:, 1:], bbox_pred, max_shape=img_shape)
+        else:
+            bboxes = rois[:, 1:].clone()
+            if img_shape is not None:
+                bboxes[:, [0, 2]].clamp_(min=0, max=img_shape[1])
+                bboxes[:, [1, 3]].clamp_(min=0, max=img_shape[0])
+
+        if rescale and bboxes.size(0) > 0:
+            if isinstance(scale_factor, float):
+                bboxes /= scale_factor
+            else:
+                scale_factor = bboxes.new_tensor(scale_factor)
+                bboxes = (bboxes.view(bboxes.size(0), -1, 4) /
+                          scale_factor).view(bboxes.size()[0], -1)
+
+        if cfg is None:
+            return bboxes, scores
+        else:
+            det_bboxes, det_labels = multiclass_nms(bboxes, scores,
+                                                    cfg.score_thr, cfg.nms,
+                                                    cfg.max_per_img)
+
+            return det_bboxes, det_labels
+
+    @force_fp32(apply_to=('bbox_preds', ))
+    def refine_bboxes(self, rois, labels, bbox_preds, pos_is_gts, img_metas):
+        """Refine bboxes during training.
+
+        Args:
+            rois (Tensor): Shape (n*bs, 5), where n is image number per GPU,
+                and bs is the sampled RoIs per image. The first column is
+                the image id and the next 4 columns are x1, y1, x2, y2.
+            labels (Tensor): Shape (n*bs, ).
+            bbox_preds (Tensor): Shape (n*bs, 4) or (n*bs, 4*#class).
+            pos_is_gts (list[Tensor]): Flags indicating if each positive bbox
+                is a gt bbox.
+            img_metas (list[dict]): Meta info of each image.
+
+        Returns:
+            list[Tensor]: Refined bboxes of each image in a mini-batch.
+
+        Example:
+            >>> # xdoctest: +REQUIRES(module:kwarray)
+            >>> import kwarray
+            >>> import numpy as np
+            >>> from mmdet.core.bbox.demodata import random_boxes
+            >>> self = BBoxHead(reg_class_agnostic=True)
+            >>> n_roi = 2
+            >>> n_img = 4
+            >>> scale = 512
+            >>> rng = np.random.RandomState(0)
+            >>> img_metas = [{'img_shape': (scale, scale)}
+            ...              for _ in range(n_img)]
+            >>> # Create rois in the expected format
+            >>> roi_boxes = random_boxes(n_roi, scale=scale, rng=rng)
+            >>> img_ids = torch.randint(0, n_img, (n_roi,))
+            >>> img_ids = img_ids.float()
+            >>> rois = torch.cat([img_ids[:, None], roi_boxes], dim=1)
+            >>> # Create other args
+            >>> labels = torch.randint(0, 2, (n_roi,)).long()
+            >>> bbox_preds = random_boxes(n_roi, scale=scale, rng=rng)
+            >>> # For each image, pretend random positive boxes are gts
+            >>> is_label_pos = (labels.numpy() > 0).astype(np.int)
+            >>> lbl_per_img = kwarray.group_items(is_label_pos,
+            ...                                   img_ids.numpy())
+            >>> pos_per_img = [sum(lbl_per_img.get(gid, []))
+            ...                for gid in range(n_img)]
+            >>> pos_is_gts = [
+            >>>     torch.randint(0, 2, (npos,)).byte().sort(
+            >>>         descending=True)[0]
+            >>>     for npos in pos_per_img
+            >>> ]
+            >>> bboxes_list = self.refine_bboxes(rois, labels, bbox_preds,
+            >>>                    pos_is_gts, img_metas)
+            >>> print(bboxes_list)
+        """
+        img_ids = rois[:, 0].long().unique(sorted=True)
+        assert img_ids.numel() <= len(img_metas)
+
+        bboxes_list = []
+        for i in range(len(img_metas)):
+            inds = torch.nonzero(
+                rois[:, 0] == i, as_tuple=False).squeeze(dim=1)
+            num_rois = inds.numel()
+
+            bboxes_ = rois[inds, 1:]
+            label_ = labels[inds]
+            bbox_pred_ = bbox_preds[inds]
+            img_meta_ = img_metas[i]
+            pos_is_gts_ = pos_is_gts[i]
+
+            bboxes = self.regress_by_class(bboxes_, label_, bbox_pred_,
+                                           img_meta_)
+
+            # filter gt bboxes
+            pos_keep = 1 - pos_is_gts_
+            keep_inds = pos_is_gts_.new_ones(num_rois)
+            keep_inds[:len(pos_is_gts_)] = pos_keep
+
+            bboxes_list.append(bboxes[keep_inds.type(torch.bool)])
+
+        return bboxes_list
+
+    @force_fp32(apply_to=('bbox_pred', ))
+    def regress_by_class(self, rois, label, bbox_pred, img_meta):
+        """Regress the bbox for the predicted class. Used in Cascade R-CNN.
+
+        Args:
+            rois (Tensor): shape (n, 4) or (n, 5)
+            label (Tensor): shape (n, )
+            bbox_pred (Tensor): shape (n, 4*(#class)) or (n, 4)
+            img_meta (dict): Image meta info.
+
+        Returns:
+            Tensor: Regressed bboxes, the same shape as input rois.
+        """
+        assert rois.size(1) == 4 or rois.size(1) == 5, repr(rois.shape)
+
+        if not self.reg_class_agnostic:
+            label = label * 4
+            inds = torch.stack((label, label + 1, label + 2, label + 3), 1)
+            bbox_pred = torch.gather(bbox_pred, 1, inds)
+        assert bbox_pred.size(1) == 4
+
+        if rois.size(1) == 4:
+            new_rois = self.bbox_coder.decode(
+                rois, bbox_pred, max_shape=img_meta['img_shape'])
+        else:
+            bboxes = self.bbox_coder.decode(
+                rois[:, 1:], bbox_pred, max_shape=img_meta['img_shape'])
+            new_rois = torch.cat((rois[:, [0]], bboxes), dim=1)
+
+        return new_rois
diff --git a/detection/scrfd/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py
new file mode 100755
index 000000000..0e86d2ea6
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py
@@ -0,0 +1,205 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule
+
+from mmdet.models.builder import HEADS
+from .bbox_head import BBoxHead
+
+
+@HEADS.register_module()
+class ConvFCBBoxHead(BBoxHead):
+    r"""More general bbox head, with shared conv and fc layers and two optional
+    separated branches.
+
+    .. code-block:: none
+
+                                    /-> cls convs -> cls fcs -> cls
+        shared convs -> shared fcs
+                                    \-> reg convs -> reg fcs -> reg
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_shared_convs=0,
+                 num_shared_fcs=0,
+                 num_cls_convs=0,
+                 num_cls_fcs=0,
+                 num_reg_convs=0,
+                 num_reg_fcs=0,
+                 conv_out_channels=256,
+                 fc_out_channels=1024,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 *args,
+                 **kwargs):
+        super(ConvFCBBoxHead, self).__init__(*args, **kwargs)
+        assert (num_shared_convs + num_shared_fcs + num_cls_convs +
+                num_cls_fcs + num_reg_convs + num_reg_fcs > 0)
+        if num_cls_convs > 0 or num_reg_convs > 0:
+            assert num_shared_fcs == 0
+        if not self.with_cls:
+            assert num_cls_convs == 0 and num_cls_fcs == 0
+        if not self.with_reg:
+            assert num_reg_convs == 0 and num_reg_fcs == 0
+        self.num_shared_convs = num_shared_convs
+        self.num_shared_fcs = num_shared_fcs
+        self.num_cls_convs = num_cls_convs
+        self.num_cls_fcs = num_cls_fcs
+        self.num_reg_convs = num_reg_convs
+        self.num_reg_fcs = num_reg_fcs
+        self.conv_out_channels = conv_out_channels
+        self.fc_out_channels = fc_out_channels
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+
+        # add shared convs and fcs
+        self.shared_convs, self.shared_fcs, last_layer_dim = \
+            self._add_conv_fc_branch(
+                self.num_shared_convs, self.num_shared_fcs, self.in_channels,
+                True)
+        self.shared_out_channels = last_layer_dim
+
+        # add cls specific branch
+        self.cls_convs, self.cls_fcs, self.cls_last_dim = \
+            self._add_conv_fc_branch(
+                self.num_cls_convs, self.num_cls_fcs, self.shared_out_channels)
+
+        # add reg specific branch
+        self.reg_convs, self.reg_fcs, self.reg_last_dim = \
+            self._add_conv_fc_branch(
+                self.num_reg_convs, self.num_reg_fcs, self.shared_out_channels)
+
+        if self.num_shared_fcs == 0 and not self.with_avg_pool:
+            if self.num_cls_fcs == 0:
+                self.cls_last_dim *= self.roi_feat_area
+            if self.num_reg_fcs == 0:
+                self.reg_last_dim *= self.roi_feat_area
+
+        self.relu = nn.ReLU(inplace=True)
+        # reconstruct fc_cls and fc_reg since input channels are changed
+        if self.with_cls:
+            self.fc_cls = nn.Linear(self.cls_last_dim, self.num_classes + 1)
+        if self.with_reg:
+            out_dim_reg = (4 if self.reg_class_agnostic else 4 *
+                           self.num_classes)
+            self.fc_reg = nn.Linear(self.reg_last_dim, out_dim_reg)
+
+    def _add_conv_fc_branch(self,
+                            num_branch_convs,
+                            num_branch_fcs,
+                            in_channels,
+                            is_shared=False):
+        """Add shared or separable branch.
+
+        convs -> avg pool (optional) -> fcs
+        """
+        last_layer_dim = in_channels
+        # add branch specific conv layers
+        branch_convs = nn.ModuleList()
+        if num_branch_convs > 0:
+            for i in range(num_branch_convs):
+                conv_in_channels = (
+                    last_layer_dim if i == 0 else self.conv_out_channels)
+                branch_convs.append(
+                    ConvModule(
+                        conv_in_channels,
+                        self.conv_out_channels,
+                        3,
+                        padding=1,
+                        conv_cfg=self.conv_cfg,
+                        norm_cfg=self.norm_cfg))
+            last_layer_dim = self.conv_out_channels
+        # add branch specific fc layers
+        branch_fcs = nn.ModuleList()
+        if num_branch_fcs > 0:
+            # for shared branch, only consider self.with_avg_pool
+            # for separated branches, also consider self.num_shared_fcs
+            if (is_shared
+                    or self.num_shared_fcs == 0) and not self.with_avg_pool:
+                last_layer_dim *= self.roi_feat_area
+            for i in range(num_branch_fcs):
+                fc_in_channels = (
+                    last_layer_dim if i == 0 else self.fc_out_channels)
+                branch_fcs.append(
+                    nn.Linear(fc_in_channels, self.fc_out_channels))
+            last_layer_dim = self.fc_out_channels
+        return branch_convs, branch_fcs, last_layer_dim
+
+    def init_weights(self):
+        super(ConvFCBBoxHead, self).init_weights()
+        # conv layers are already initialized by ConvModule
+        for module_list in [self.shared_fcs, self.cls_fcs, self.reg_fcs]:
+            for m in module_list.modules():
+                if isinstance(m, nn.Linear):
+                    nn.init.xavier_uniform_(m.weight)
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        # shared part
+        if self.num_shared_convs > 0:
+            for conv in self.shared_convs:
+                x = conv(x)
+
+        if self.num_shared_fcs > 0:
+            if self.with_avg_pool:
+                x = self.avg_pool(x)
+
+            x = x.flatten(1)
+
+            for fc in self.shared_fcs:
+                x = self.relu(fc(x))
+        # separate branches
+        x_cls = x
+        x_reg = x
+
+        for conv in self.cls_convs:
+            x_cls = conv(x_cls)
+        if x_cls.dim() > 2:
+            if self.with_avg_pool:
+                x_cls = self.avg_pool(x_cls)
+            x_cls = x_cls.flatten(1)
+        for fc in self.cls_fcs:
+            x_cls = self.relu(fc(x_cls))
+
+        for conv in self.reg_convs:
+            x_reg = conv(x_reg)
+        if x_reg.dim() > 2:
+            if self.with_avg_pool:
+                x_reg = self.avg_pool(x_reg)
+            x_reg = x_reg.flatten(1)
+        for fc in self.reg_fcs:
+            x_reg = self.relu(fc(x_reg))
+
+        cls_score = self.fc_cls(x_cls) if self.with_cls else None
+        bbox_pred = self.fc_reg(x_reg) if self.with_reg else None
+        return cls_score, bbox_pred
+
+
+@HEADS.register_module()
+class Shared2FCBBoxHead(ConvFCBBoxHead):
+
+    def __init__(self, fc_out_channels=1024, *args, **kwargs):
+        super(Shared2FCBBoxHead, self).__init__(
+            num_shared_convs=0,
+            num_shared_fcs=2,
+            num_cls_convs=0,
+            num_cls_fcs=0,
+            num_reg_convs=0,
+            num_reg_fcs=0,
+            fc_out_channels=fc_out_channels,
+            *args,
+            **kwargs)
+
+
+@HEADS.register_module()
+class Shared4Conv1FCBBoxHead(ConvFCBBoxHead):
+
+    def __init__(self, fc_out_channels=1024, *args, **kwargs):
+        super(Shared4Conv1FCBBoxHead, self).__init__(
+            num_shared_convs=4,
+            num_shared_fcs=1,
+            num_cls_convs=0,
+            num_cls_fcs=0,
+            num_reg_convs=0,
+            num_reg_fcs=0,
+            fc_out_channels=fc_out_channels,
+            *args,
+            **kwargs)
diff --git a/detection/scrfd/mmdet/models/roi_heads/bbox_heads/double_bbox_head.py b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/double_bbox_head.py
new file mode 100755
index 000000000..6c154cb3c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/double_bbox_head.py
@@ -0,0 +1,172 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, normal_init, xavier_init
+
+from mmdet.models.backbones.resnet import Bottleneck
+from mmdet.models.builder import HEADS
+from .bbox_head import BBoxHead
+
+
+class BasicResBlock(nn.Module):
+    """Basic residual block.
+
+    This block is a little different from the block in the ResNet backbone.
+    The kernel size of conv1 is 1 in this block while 3 in ResNet BasicBlock.
+
+    Args:
+        in_channels (int): Channels of the input feature map.
+        out_channels (int): Channels of the output feature map.
+        conv_cfg (dict): The config dict for convolution layers.
+        norm_cfg (dict): The config dict for normalization layers.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN')):
+        super(BasicResBlock, self).__init__()
+
+        # main path
+        self.conv1 = ConvModule(
+            in_channels,
+            in_channels,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+            conv_cfg=conv_cfg,
+            norm_cfg=norm_cfg)
+        self.conv2 = ConvModule(
+            in_channels,
+            out_channels,
+            kernel_size=1,
+            bias=False,
+            conv_cfg=conv_cfg,
+            norm_cfg=norm_cfg,
+            act_cfg=None)
+
+        # identity path
+        self.conv_identity = ConvModule(
+            in_channels,
+            out_channels,
+            kernel_size=1,
+            conv_cfg=conv_cfg,
+            norm_cfg=norm_cfg,
+            act_cfg=None)
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        identity = x
+
+        x = self.conv1(x)
+        x = self.conv2(x)
+
+        identity = self.conv_identity(identity)
+        out = x + identity
+
+        out = self.relu(out)
+        return out
+
+
+@HEADS.register_module()
+class DoubleConvFCBBoxHead(BBoxHead):
+    r"""Bbox head used in Double-Head R-CNN
+
+    .. code-block:: none
+
+                                          /-> cls
+                      /-> shared convs ->
+                                          \-> reg
+        roi features
+                                          /-> cls
+                      \-> shared fc    ->
+                                          \-> reg
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_convs=0,
+                 num_fcs=0,
+                 conv_out_channels=1024,
+                 fc_out_channels=1024,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 **kwargs):
+        kwargs.setdefault('with_avg_pool', True)
+        super(DoubleConvFCBBoxHead, self).__init__(**kwargs)
+        assert self.with_avg_pool
+        assert num_convs > 0
+        assert num_fcs > 0
+        self.num_convs = num_convs
+        self.num_fcs = num_fcs
+        self.conv_out_channels = conv_out_channels
+        self.fc_out_channels = fc_out_channels
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+
+        # increase the channel of input features
+        self.res_block = BasicResBlock(self.in_channels,
+                                       self.conv_out_channels)
+
+        # add conv heads
+        self.conv_branch = self._add_conv_branch()
+        # add fc heads
+        self.fc_branch = self._add_fc_branch()
+
+        out_dim_reg = 4 if self.reg_class_agnostic else 4 * self.num_classes
+        self.fc_reg = nn.Linear(self.conv_out_channels, out_dim_reg)
+
+        self.fc_cls = nn.Linear(self.fc_out_channels, self.num_classes + 1)
+        self.relu = nn.ReLU(inplace=True)
+
+    def _add_conv_branch(self):
+        """Add the fc branch which consists of a sequential of conv layers."""
+        branch_convs = nn.ModuleList()
+        for i in range(self.num_convs):
+            branch_convs.append(
+                Bottleneck(
+                    inplanes=self.conv_out_channels,
+                    planes=self.conv_out_channels // 4,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        return branch_convs
+
+    def _add_fc_branch(self):
+        """Add the fc branch which consists of a sequential of fc layers."""
+        branch_fcs = nn.ModuleList()
+        for i in range(self.num_fcs):
+            fc_in_channels = (
+                self.in_channels *
+                self.roi_feat_area if i == 0 else self.fc_out_channels)
+            branch_fcs.append(nn.Linear(fc_in_channels, self.fc_out_channels))
+        return branch_fcs
+
+    def init_weights(self):
+        # conv layers are already initialized by ConvModule
+        normal_init(self.fc_cls, std=0.01)
+        normal_init(self.fc_reg, std=0.001)
+
+        for m in self.fc_branch.modules():
+            if isinstance(m, nn.Linear):
+                xavier_init(m, distribution='uniform')
+
+    def forward(self, x_cls, x_reg):
+        # conv head
+        x_conv = self.res_block(x_reg)
+
+        for conv in self.conv_branch:
+            x_conv = conv(x_conv)
+
+        if self.with_avg_pool:
+            x_conv = self.avg_pool(x_conv)
+
+        x_conv = x_conv.view(x_conv.size(0), -1)
+        bbox_pred = self.fc_reg(x_conv)
+
+        # fc head
+        x_fc = x_cls.view(x_cls.size(0), -1)
+        for fc in self.fc_branch:
+            x_fc = self.relu(fc(x_fc))
+
+        cls_score = self.fc_cls(x_fc)
+
+        return cls_score, bbox_pred
diff --git a/detection/scrfd/mmdet/models/roi_heads/bbox_heads/sabl_head.py b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/sabl_head.py
new file mode 100755
index 000000000..370a933bf
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/bbox_heads/sabl_head.py
@@ -0,0 +1,572 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, kaiming_init, normal_init, xavier_init
+from mmcv.runner import force_fp32
+
+from mmdet.core import build_bbox_coder, multi_apply, multiclass_nms
+from mmdet.models.builder import HEADS, build_loss
+from mmdet.models.losses import accuracy
+
+
+@HEADS.register_module()
+class SABLHead(nn.Module):
+    """Side-Aware Boundary Localization (SABL) for RoI-Head.
+
+    Side-Aware features are extracted by conv layers
+    with an attention mechanism.
+    Boundary Localization with Bucketing and Bucketing Guided Rescoring
+    are implemented in BucketingBBoxCoder.
+
+    Please refer to https://arxiv.org/abs/1912.04260 for more details.
+
+    Args:
+        cls_in_channels (int): Input channels of cls RoI feature. \
+            Defaults to 256.
+        reg_in_channels (int): Input channels of reg RoI feature. \
+            Defaults to 256.
+        roi_feat_size (int): Size of RoI features. Defaults to 7.
+        reg_feat_up_ratio (int): Upsample ratio of reg features. \
+            Defaults to 2.
+        reg_pre_kernel (int): Kernel of 2D conv layers before \
+            attention pooling. Defaults to 3.
+        reg_post_kernel (int): Kernel of 1D conv layers after \
+            attention pooling. Defaults to 3.
+        reg_pre_num (int): Number of pre convs. Defaults to 2.
+        reg_post_num (int): Number of post convs. Defaults to 1.
+        num_classes (int): Number of classes in dataset. Defaults to 80.
+        cls_out_channels (int): Hidden channels in cls fcs. Defaults to 1024.
+        reg_offset_out_channels (int): Hidden and output channel \
+            of reg offset branch. Defaults to 256.
+        reg_cls_out_channels (int): Hidden and output channel \
+            of reg cls branch. Defaults to 256.
+        num_cls_fcs (int): Number of fcs for cls branch. Defaults to 1.
+        num_reg_fcs (int): Number of fcs for reg branch.. Defaults to 0.
+        reg_class_agnostic (bool): Class agnostic regresion or not. \
+            Defaults to True.
+        norm_cfg (dict): Config of norm layers. Defaults to None.
+        bbox_coder (dict): Config of bbox coder. Defaults 'BucketingBBoxCoder'.
+        loss_cls (dict): Config of classification loss.
+        loss_bbox_cls (dict): Config of classification loss for bbox branch.
+        loss_bbox_reg (dict): Config of regression loss for bbox branch.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 cls_in_channels=256,
+                 reg_in_channels=256,
+                 roi_feat_size=7,
+                 reg_feat_up_ratio=2,
+                 reg_pre_kernel=3,
+                 reg_post_kernel=3,
+                 reg_pre_num=2,
+                 reg_post_num=1,
+                 cls_out_channels=1024,
+                 reg_offset_out_channels=256,
+                 reg_cls_out_channels=256,
+                 num_cls_fcs=1,
+                 num_reg_fcs=0,
+                 reg_class_agnostic=True,
+                 norm_cfg=None,
+                 bbox_coder=dict(
+                     type='BucketingBBoxCoder',
+                     num_buckets=14,
+                     scale_factor=1.7),
+                 loss_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=False,
+                     loss_weight=1.0),
+                 loss_bbox_cls=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0),
+                 loss_bbox_reg=dict(
+                     type='SmoothL1Loss', beta=0.1, loss_weight=1.0)):
+        super(SABLHead, self).__init__()
+        self.cls_in_channels = cls_in_channels
+        self.reg_in_channels = reg_in_channels
+        self.roi_feat_size = roi_feat_size
+        self.reg_feat_up_ratio = int(reg_feat_up_ratio)
+        self.num_buckets = bbox_coder['num_buckets']
+        assert self.reg_feat_up_ratio // 2 >= 1
+        self.up_reg_feat_size = roi_feat_size * self.reg_feat_up_ratio
+        assert self.up_reg_feat_size == bbox_coder['num_buckets']
+        self.reg_pre_kernel = reg_pre_kernel
+        self.reg_post_kernel = reg_post_kernel
+        self.reg_pre_num = reg_pre_num
+        self.reg_post_num = reg_post_num
+        self.num_classes = num_classes
+        self.cls_out_channels = cls_out_channels
+        self.reg_offset_out_channels = reg_offset_out_channels
+        self.reg_cls_out_channels = reg_cls_out_channels
+        self.num_cls_fcs = num_cls_fcs
+        self.num_reg_fcs = num_reg_fcs
+        self.reg_class_agnostic = reg_class_agnostic
+        assert self.reg_class_agnostic
+        self.norm_cfg = norm_cfg
+
+        self.bbox_coder = build_bbox_coder(bbox_coder)
+        self.loss_cls = build_loss(loss_cls)
+        self.loss_bbox_cls = build_loss(loss_bbox_cls)
+        self.loss_bbox_reg = build_loss(loss_bbox_reg)
+
+        self.cls_fcs = self._add_fc_branch(self.num_cls_fcs,
+                                           self.cls_in_channels,
+                                           self.roi_feat_size,
+                                           self.cls_out_channels)
+
+        self.side_num = int(np.ceil(self.num_buckets / 2))
+
+        if self.reg_feat_up_ratio > 1:
+            self.upsample_x = nn.ConvTranspose1d(
+                reg_in_channels,
+                reg_in_channels,
+                self.reg_feat_up_ratio,
+                stride=self.reg_feat_up_ratio)
+            self.upsample_y = nn.ConvTranspose1d(
+                reg_in_channels,
+                reg_in_channels,
+                self.reg_feat_up_ratio,
+                stride=self.reg_feat_up_ratio)
+
+        self.reg_pre_convs = nn.ModuleList()
+        for i in range(self.reg_pre_num):
+            reg_pre_conv = ConvModule(
+                reg_in_channels,
+                reg_in_channels,
+                kernel_size=reg_pre_kernel,
+                padding=reg_pre_kernel // 2,
+                norm_cfg=norm_cfg,
+                act_cfg=dict(type='ReLU'))
+            self.reg_pre_convs.append(reg_pre_conv)
+
+        self.reg_post_conv_xs = nn.ModuleList()
+        for i in range(self.reg_post_num):
+            reg_post_conv_x = ConvModule(
+                reg_in_channels,
+                reg_in_channels,
+                kernel_size=(1, reg_post_kernel),
+                padding=(0, reg_post_kernel // 2),
+                norm_cfg=norm_cfg,
+                act_cfg=dict(type='ReLU'))
+            self.reg_post_conv_xs.append(reg_post_conv_x)
+        self.reg_post_conv_ys = nn.ModuleList()
+        for i in range(self.reg_post_num):
+            reg_post_conv_y = ConvModule(
+                reg_in_channels,
+                reg_in_channels,
+                kernel_size=(reg_post_kernel, 1),
+                padding=(reg_post_kernel // 2, 0),
+                norm_cfg=norm_cfg,
+                act_cfg=dict(type='ReLU'))
+            self.reg_post_conv_ys.append(reg_post_conv_y)
+
+        self.reg_conv_att_x = nn.Conv2d(reg_in_channels, 1, 1)
+        self.reg_conv_att_y = nn.Conv2d(reg_in_channels, 1, 1)
+
+        self.fc_cls = nn.Linear(self.cls_out_channels, self.num_classes + 1)
+        self.relu = nn.ReLU(inplace=True)
+
+        self.reg_cls_fcs = self._add_fc_branch(self.num_reg_fcs,
+                                               self.reg_in_channels, 1,
+                                               self.reg_cls_out_channels)
+        self.reg_offset_fcs = self._add_fc_branch(self.num_reg_fcs,
+                                                  self.reg_in_channels, 1,
+                                                  self.reg_offset_out_channels)
+        self.fc_reg_cls = nn.Linear(self.reg_cls_out_channels, 1)
+        self.fc_reg_offset = nn.Linear(self.reg_offset_out_channels, 1)
+
+    def _add_fc_branch(self, num_branch_fcs, in_channels, roi_feat_size,
+                       fc_out_channels):
+        in_channels = in_channels * roi_feat_size * roi_feat_size
+        branch_fcs = nn.ModuleList()
+        for i in range(num_branch_fcs):
+            fc_in_channels = (in_channels if i == 0 else fc_out_channels)
+            branch_fcs.append(nn.Linear(fc_in_channels, fc_out_channels))
+        return branch_fcs
+
+    def init_weights(self):
+        for module_list in [
+                self.reg_cls_fcs, self.reg_offset_fcs, self.cls_fcs
+        ]:
+            for m in module_list.modules():
+                if isinstance(m, nn.Linear):
+                    xavier_init(m, distribution='uniform')
+        if self.reg_feat_up_ratio > 1:
+            kaiming_init(self.upsample_x, distribution='normal')
+            kaiming_init(self.upsample_y, distribution='normal')
+
+        normal_init(self.reg_conv_att_x, 0, 0.01)
+        normal_init(self.reg_conv_att_y, 0, 0.01)
+        normal_init(self.fc_reg_offset, 0, 0.001)
+        normal_init(self.fc_reg_cls, 0, 0.01)
+        normal_init(self.fc_cls, 0, 0.01)
+
+    def cls_forward(self, cls_x):
+        cls_x = cls_x.view(cls_x.size(0), -1)
+        for fc in self.cls_fcs:
+            cls_x = self.relu(fc(cls_x))
+        cls_score = self.fc_cls(cls_x)
+        return cls_score
+
+    def attention_pool(self, reg_x):
+        """Extract direction-specific features fx and fy with attention
+        methanism."""
+        reg_fx = reg_x
+        reg_fy = reg_x
+        reg_fx_att = self.reg_conv_att_x(reg_fx).sigmoid()
+        reg_fy_att = self.reg_conv_att_y(reg_fy).sigmoid()
+        reg_fx_att = reg_fx_att / reg_fx_att.sum(dim=2).unsqueeze(2)
+        reg_fy_att = reg_fy_att / reg_fy_att.sum(dim=3).unsqueeze(3)
+        reg_fx = (reg_fx * reg_fx_att).sum(dim=2)
+        reg_fy = (reg_fy * reg_fy_att).sum(dim=3)
+        return reg_fx, reg_fy
+
+    def side_aware_feature_extractor(self, reg_x):
+        """Refine and extract side-aware features without split them."""
+        for reg_pre_conv in self.reg_pre_convs:
+            reg_x = reg_pre_conv(reg_x)
+        reg_fx, reg_fy = self.attention_pool(reg_x)
+
+        if self.reg_post_num > 0:
+            reg_fx = reg_fx.unsqueeze(2)
+            reg_fy = reg_fy.unsqueeze(3)
+            for i in range(self.reg_post_num):
+                reg_fx = self.reg_post_conv_xs[i](reg_fx)
+                reg_fy = self.reg_post_conv_ys[i](reg_fy)
+            reg_fx = reg_fx.squeeze(2)
+            reg_fy = reg_fy.squeeze(3)
+        if self.reg_feat_up_ratio > 1:
+            reg_fx = self.relu(self.upsample_x(reg_fx))
+            reg_fy = self.relu(self.upsample_y(reg_fy))
+        reg_fx = torch.transpose(reg_fx, 1, 2)
+        reg_fy = torch.transpose(reg_fy, 1, 2)
+        return reg_fx.contiguous(), reg_fy.contiguous()
+
+    def reg_pred(self, x, offfset_fcs, cls_fcs):
+        """Predict bucketing esimation (cls_pred) and fine regression (offset
+        pred) with side-aware features."""
+        x_offset = x.view(-1, self.reg_in_channels)
+        x_cls = x.view(-1, self.reg_in_channels)
+
+        for fc in offfset_fcs:
+            x_offset = self.relu(fc(x_offset))
+        for fc in cls_fcs:
+            x_cls = self.relu(fc(x_cls))
+        offset_pred = self.fc_reg_offset(x_offset)
+        cls_pred = self.fc_reg_cls(x_cls)
+
+        offset_pred = offset_pred.view(x.size(0), -1)
+        cls_pred = cls_pred.view(x.size(0), -1)
+
+        return offset_pred, cls_pred
+
+    def side_aware_split(self, feat):
+        """Split side-aware features aligned with orders of bucketing
+        targets."""
+        l_end = int(np.ceil(self.up_reg_feat_size / 2))
+        r_start = int(np.floor(self.up_reg_feat_size / 2))
+        feat_fl = feat[:, :l_end]
+        feat_fr = feat[:, r_start:].flip(dims=(1, ))
+        feat_fl = feat_fl.contiguous()
+        feat_fr = feat_fr.contiguous()
+        feat = torch.cat([feat_fl, feat_fr], dim=-1)
+        return feat
+
+    def bbox_pred_split(self, bbox_pred, num_proposals_per_img):
+        """Split batch bbox prediction back to each image."""
+        bucket_cls_preds, bucket_offset_preds = bbox_pred
+        bucket_cls_preds = bucket_cls_preds.split(num_proposals_per_img, 0)
+        bucket_offset_preds = bucket_offset_preds.split(
+            num_proposals_per_img, 0)
+        bbox_pred = tuple(zip(bucket_cls_preds, bucket_offset_preds))
+        return bbox_pred
+
+    def reg_forward(self, reg_x):
+        outs = self.side_aware_feature_extractor(reg_x)
+        edge_offset_preds = []
+        edge_cls_preds = []
+        reg_fx = outs[0]
+        reg_fy = outs[1]
+        offset_pred_x, cls_pred_x = self.reg_pred(reg_fx, self.reg_offset_fcs,
+                                                  self.reg_cls_fcs)
+        offset_pred_y, cls_pred_y = self.reg_pred(reg_fy, self.reg_offset_fcs,
+                                                  self.reg_cls_fcs)
+        offset_pred_x = self.side_aware_split(offset_pred_x)
+        offset_pred_y = self.side_aware_split(offset_pred_y)
+        cls_pred_x = self.side_aware_split(cls_pred_x)
+        cls_pred_y = self.side_aware_split(cls_pred_y)
+        edge_offset_preds = torch.cat([offset_pred_x, offset_pred_y], dim=-1)
+        edge_cls_preds = torch.cat([cls_pred_x, cls_pred_y], dim=-1)
+
+        return (edge_cls_preds, edge_offset_preds)
+
+    def forward(self, x):
+
+        bbox_pred = self.reg_forward(x)
+        cls_score = self.cls_forward(x)
+
+        return cls_score, bbox_pred
+
+    def get_targets(self, sampling_results, gt_bboxes, gt_labels,
+                    rcnn_train_cfg):
+        pos_proposals = [res.pos_bboxes for res in sampling_results]
+        neg_proposals = [res.neg_bboxes for res in sampling_results]
+        pos_gt_bboxes = [res.pos_gt_bboxes for res in sampling_results]
+        pos_gt_labels = [res.pos_gt_labels for res in sampling_results]
+        cls_reg_targets = self.bucket_target(pos_proposals, neg_proposals,
+                                             pos_gt_bboxes, pos_gt_labels,
+                                             rcnn_train_cfg)
+        (labels, label_weights, bucket_cls_targets, bucket_cls_weights,
+         bucket_offset_targets, bucket_offset_weights) = cls_reg_targets
+        return (labels, label_weights, (bucket_cls_targets,
+                                        bucket_offset_targets),
+                (bucket_cls_weights, bucket_offset_weights))
+
+    def bucket_target(self,
+                      pos_proposals_list,
+                      neg_proposals_list,
+                      pos_gt_bboxes_list,
+                      pos_gt_labels_list,
+                      rcnn_train_cfg,
+                      concat=True):
+        (labels, label_weights, bucket_cls_targets, bucket_cls_weights,
+         bucket_offset_targets, bucket_offset_weights) = multi_apply(
+             self._bucket_target_single,
+             pos_proposals_list,
+             neg_proposals_list,
+             pos_gt_bboxes_list,
+             pos_gt_labels_list,
+             cfg=rcnn_train_cfg)
+
+        if concat:
+            labels = torch.cat(labels, 0)
+            label_weights = torch.cat(label_weights, 0)
+            bucket_cls_targets = torch.cat(bucket_cls_targets, 0)
+            bucket_cls_weights = torch.cat(bucket_cls_weights, 0)
+            bucket_offset_targets = torch.cat(bucket_offset_targets, 0)
+            bucket_offset_weights = torch.cat(bucket_offset_weights, 0)
+        return (labels, label_weights, bucket_cls_targets, bucket_cls_weights,
+                bucket_offset_targets, bucket_offset_weights)
+
+    def _bucket_target_single(self, pos_proposals, neg_proposals,
+                              pos_gt_bboxes, pos_gt_labels, cfg):
+        """Compute bucketing estimation targets and fine regression targets for
+        a single image.
+
+        Args:
+            pos_proposals (Tensor): positive proposals of a single image,
+                 Shape (n_pos, 4)
+            neg_proposals (Tensor): negative proposals of a single image,
+                 Shape (n_neg, 4).
+            pos_gt_bboxes (Tensor): gt bboxes assigned to positive proposals
+                 of a single image, Shape (n_pos, 4).
+            pos_gt_labels (Tensor): gt labels assigned to positive proposals
+                 of a single image, Shape (n_pos, ).
+            cfg (dict): Config of calculating targets
+
+        Returns:
+            tuple:
+
+                - labels (Tensor): Labels in a single image. \
+                    Shape (n,).
+                - label_weights (Tensor): Label weights in a single image.\
+                    Shape (n,)
+                - bucket_cls_targets (Tensor): Bucket cls targets in \
+                    a single image. Shape (n, num_buckets*2).
+                - bucket_cls_weights (Tensor): Bucket cls weights in \
+                    a single image. Shape (n, num_buckets*2).
+                - bucket_offset_targets (Tensor): Bucket offset targets \
+                    in a single image. Shape (n, num_buckets*2).
+                - bucket_offset_targets (Tensor): Bucket offset weights \
+                    in a single image. Shape (n, num_buckets*2).
+        """
+        num_pos = pos_proposals.size(0)
+        num_neg = neg_proposals.size(0)
+        num_samples = num_pos + num_neg
+        labels = pos_gt_bboxes.new_full((num_samples, ),
+                                        self.num_classes,
+                                        dtype=torch.long)
+        label_weights = pos_proposals.new_zeros(num_samples)
+        bucket_cls_targets = pos_proposals.new_zeros(num_samples,
+                                                     4 * self.side_num)
+        bucket_cls_weights = pos_proposals.new_zeros(num_samples,
+                                                     4 * self.side_num)
+        bucket_offset_targets = pos_proposals.new_zeros(
+            num_samples, 4 * self.side_num)
+        bucket_offset_weights = pos_proposals.new_zeros(
+            num_samples, 4 * self.side_num)
+        if num_pos > 0:
+            labels[:num_pos] = pos_gt_labels
+            label_weights[:num_pos] = 1.0
+            (pos_bucket_offset_targets, pos_bucket_offset_weights,
+             pos_bucket_cls_targets,
+             pos_bucket_cls_weights) = self.bbox_coder.encode(
+                 pos_proposals, pos_gt_bboxes)
+            bucket_cls_targets[:num_pos, :] = pos_bucket_cls_targets
+            bucket_cls_weights[:num_pos, :] = pos_bucket_cls_weights
+            bucket_offset_targets[:num_pos, :] = pos_bucket_offset_targets
+            bucket_offset_weights[:num_pos, :] = pos_bucket_offset_weights
+        if num_neg > 0:
+            label_weights[-num_neg:] = 1.0
+        return (labels, label_weights, bucket_cls_targets, bucket_cls_weights,
+                bucket_offset_targets, bucket_offset_weights)
+
+    def loss(self,
+             cls_score,
+             bbox_pred,
+             rois,
+             labels,
+             label_weights,
+             bbox_targets,
+             bbox_weights,
+             reduction_override=None):
+        losses = dict()
+        if cls_score is not None:
+            avg_factor = max(torch.sum(label_weights > 0).float().item(), 1.)
+            losses['loss_cls'] = self.loss_cls(
+                cls_score,
+                labels,
+                label_weights,
+                avg_factor=avg_factor,
+                reduction_override=reduction_override)
+            losses['acc'] = accuracy(cls_score, labels)
+
+        if bbox_pred is not None:
+            bucket_cls_preds, bucket_offset_preds = bbox_pred
+            bucket_cls_targets, bucket_offset_targets = bbox_targets
+            bucket_cls_weights, bucket_offset_weights = bbox_weights
+            # edge cls
+            bucket_cls_preds = bucket_cls_preds.view(-1, self.side_num)
+            bucket_cls_targets = bucket_cls_targets.view(-1, self.side_num)
+            bucket_cls_weights = bucket_cls_weights.view(-1, self.side_num)
+            losses['loss_bbox_cls'] = self.loss_bbox_cls(
+                bucket_cls_preds,
+                bucket_cls_targets,
+                bucket_cls_weights,
+                avg_factor=bucket_cls_targets.size(0),
+                reduction_override=reduction_override)
+
+            losses['loss_bbox_reg'] = self.loss_bbox_reg(
+                bucket_offset_preds,
+                bucket_offset_targets,
+                bucket_offset_weights,
+                avg_factor=bucket_offset_targets.size(0),
+                reduction_override=reduction_override)
+
+        return losses
+
+    @force_fp32(apply_to=('cls_score', 'bbox_pred'))
+    def get_bboxes(self,
+                   rois,
+                   cls_score,
+                   bbox_pred,
+                   img_shape,
+                   scale_factor,
+                   rescale=False,
+                   cfg=None):
+        if isinstance(cls_score, list):
+            cls_score = sum(cls_score) / float(len(cls_score))
+        scores = F.softmax(cls_score, dim=1) if cls_score is not None else None
+
+        if bbox_pred is not None:
+            bboxes, confids = self.bbox_coder.decode(rois[:, 1:], bbox_pred,
+                                                     img_shape)
+        else:
+            bboxes = rois[:, 1:].clone()
+            confids = None
+            if img_shape is not None:
+                bboxes[:, [0, 2]].clamp_(min=0, max=img_shape[1] - 1)
+                bboxes[:, [1, 3]].clamp_(min=0, max=img_shape[0] - 1)
+
+        if rescale and bboxes.size(0) > 0:
+            if isinstance(scale_factor, float):
+                bboxes /= scale_factor
+            else:
+                bboxes /= torch.from_numpy(scale_factor).to(bboxes.device)
+
+        if cfg is None:
+            return bboxes, scores
+        else:
+            det_bboxes, det_labels = multiclass_nms(
+                bboxes,
+                scores,
+                cfg.score_thr,
+                cfg.nms,
+                cfg.max_per_img,
+                score_factors=confids)
+
+            return det_bboxes, det_labels
+
+    @force_fp32(apply_to=('bbox_preds', ))
+    def refine_bboxes(self, rois, labels, bbox_preds, pos_is_gts, img_metas):
+        """Refine bboxes during training.
+
+        Args:
+            rois (Tensor): Shape (n*bs, 5), where n is image number per GPU,
+                and bs is the sampled RoIs per image.
+            labels (Tensor): Shape (n*bs, ).
+            bbox_preds (list[Tensor]): Shape [(n*bs, num_buckets*2), \
+                (n*bs, num_buckets*2)].
+            pos_is_gts (list[Tensor]): Flags indicating if each positive bbox
+                is a gt bbox.
+            img_metas (list[dict]): Meta info of each image.
+
+        Returns:
+            list[Tensor]: Refined bboxes of each image in a mini-batch.
+        """
+        img_ids = rois[:, 0].long().unique(sorted=True)
+        assert img_ids.numel() == len(img_metas)
+
+        bboxes_list = []
+        for i in range(len(img_metas)):
+            inds = torch.nonzero(
+                rois[:, 0] == i, as_tuple=False).squeeze(dim=1)
+            num_rois = inds.numel()
+
+            bboxes_ = rois[inds, 1:]
+            label_ = labels[inds]
+            edge_cls_preds, edge_offset_preds = bbox_preds
+            edge_cls_preds_ = edge_cls_preds[inds]
+            edge_offset_preds_ = edge_offset_preds[inds]
+            bbox_pred_ = [edge_cls_preds_, edge_offset_preds_]
+            img_meta_ = img_metas[i]
+            pos_is_gts_ = pos_is_gts[i]
+
+            bboxes = self.regress_by_class(bboxes_, label_, bbox_pred_,
+                                           img_meta_)
+            # filter gt bboxes
+            pos_keep = 1 - pos_is_gts_
+            keep_inds = pos_is_gts_.new_ones(num_rois)
+            keep_inds[:len(pos_is_gts_)] = pos_keep
+
+            bboxes_list.append(bboxes[keep_inds.type(torch.bool)])
+
+        return bboxes_list
+
+    @force_fp32(apply_to=('bbox_pred', ))
+    def regress_by_class(self, rois, label, bbox_pred, img_meta):
+        """Regress the bbox for the predicted class. Used in Cascade R-CNN.
+
+        Args:
+            rois (Tensor): shape (n, 4) or (n, 5)
+            label (Tensor): shape (n, )
+            bbox_pred (list[Tensor]): shape [(n, num_buckets *2), \
+                (n, num_buckets *2)]
+            img_meta (dict): Image meta info.
+
+        Returns:
+            Tensor: Regressed bboxes, the same shape as input rois.
+        """
+        assert rois.size(1) == 4 or rois.size(1) == 5
+
+        if rois.size(1) == 4:
+            new_rois, _ = self.bbox_coder.decode(rois, bbox_pred,
+                                                 img_meta['img_shape'])
+        else:
+            bboxes, _ = self.bbox_coder.decode(rois[:, 1:], bbox_pred,
+                                               img_meta['img_shape'])
+            new_rois = torch.cat((rois[:, [0]], bboxes), dim=1)
+
+        return new_rois
diff --git a/detection/scrfd/mmdet/models/roi_heads/cascade_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/cascade_roi_head.py
new file mode 100755
index 000000000..45b6f36a3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/cascade_roi_head.py
@@ -0,0 +1,507 @@
+import torch
+import torch.nn as nn
+
+from mmdet.core import (bbox2result, bbox2roi, bbox_mapping, build_assigner,
+                        build_sampler, merge_aug_bboxes, merge_aug_masks,
+                        multiclass_nms)
+from ..builder import HEADS, build_head, build_roi_extractor
+from .base_roi_head import BaseRoIHead
+from .test_mixins import BBoxTestMixin, MaskTestMixin
+
+
+@HEADS.register_module()
+class CascadeRoIHead(BaseRoIHead, BBoxTestMixin, MaskTestMixin):
+    """Cascade roi head including one bbox head and one mask head.
+
+    https://arxiv.org/abs/1712.00726
+    """
+
+    def __init__(self,
+                 num_stages,
+                 stage_loss_weights,
+                 bbox_roi_extractor=None,
+                 bbox_head=None,
+                 mask_roi_extractor=None,
+                 mask_head=None,
+                 shared_head=None,
+                 train_cfg=None,
+                 test_cfg=None):
+        assert bbox_roi_extractor is not None
+        assert bbox_head is not None
+        assert shared_head is None, \
+            'Shared head is not supported in Cascade RCNN anymore'
+        self.num_stages = num_stages
+        self.stage_loss_weights = stage_loss_weights
+        super(CascadeRoIHead, self).__init__(
+            bbox_roi_extractor=bbox_roi_extractor,
+            bbox_head=bbox_head,
+            mask_roi_extractor=mask_roi_extractor,
+            mask_head=mask_head,
+            shared_head=shared_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg)
+
+    def init_bbox_head(self, bbox_roi_extractor, bbox_head):
+        """Initialize box head and box roi extractor.
+
+        Args:
+            bbox_roi_extractor (dict): Config of box roi extractor.
+            bbox_head (dict): Config of box in box head.
+        """
+        self.bbox_roi_extractor = nn.ModuleList()
+        self.bbox_head = nn.ModuleList()
+        if not isinstance(bbox_roi_extractor, list):
+            bbox_roi_extractor = [
+                bbox_roi_extractor for _ in range(self.num_stages)
+            ]
+        if not isinstance(bbox_head, list):
+            bbox_head = [bbox_head for _ in range(self.num_stages)]
+        assert len(bbox_roi_extractor) == len(bbox_head) == self.num_stages
+        for roi_extractor, head in zip(bbox_roi_extractor, bbox_head):
+            self.bbox_roi_extractor.append(build_roi_extractor(roi_extractor))
+            self.bbox_head.append(build_head(head))
+
+    def init_mask_head(self, mask_roi_extractor, mask_head):
+        """Initialize mask head and mask roi extractor.
+
+        Args:
+            mask_roi_extractor (dict): Config of mask roi extractor.
+            mask_head (dict): Config of mask in mask head.
+        """
+        self.mask_head = nn.ModuleList()
+        if not isinstance(mask_head, list):
+            mask_head = [mask_head for _ in range(self.num_stages)]
+        assert len(mask_head) == self.num_stages
+        for head in mask_head:
+            self.mask_head.append(build_head(head))
+        if mask_roi_extractor is not None:
+            self.share_roi_extractor = False
+            self.mask_roi_extractor = nn.ModuleList()
+            if not isinstance(mask_roi_extractor, list):
+                mask_roi_extractor = [
+                    mask_roi_extractor for _ in range(self.num_stages)
+                ]
+            assert len(mask_roi_extractor) == self.num_stages
+            for roi_extractor in mask_roi_extractor:
+                self.mask_roi_extractor.append(
+                    build_roi_extractor(roi_extractor))
+        else:
+            self.share_roi_extractor = True
+            self.mask_roi_extractor = self.bbox_roi_extractor
+
+    def init_assigner_sampler(self):
+        """Initialize assigner and sampler for each stage."""
+        self.bbox_assigner = []
+        self.bbox_sampler = []
+        if self.train_cfg is not None:
+            for idx, rcnn_train_cfg in enumerate(self.train_cfg):
+                self.bbox_assigner.append(
+                    build_assigner(rcnn_train_cfg.assigner))
+                self.current_stage = idx
+                self.bbox_sampler.append(
+                    build_sampler(rcnn_train_cfg.sampler, context=self))
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if self.with_shared_head:
+            self.shared_head.init_weights(pretrained=pretrained)
+        for i in range(self.num_stages):
+            if self.with_bbox:
+                self.bbox_roi_extractor[i].init_weights()
+                self.bbox_head[i].init_weights()
+            if self.with_mask:
+                if not self.share_roi_extractor:
+                    self.mask_roi_extractor[i].init_weights()
+                self.mask_head[i].init_weights()
+
+    def forward_dummy(self, x, proposals):
+        """Dummy forward function."""
+        # bbox head
+        outs = ()
+        rois = bbox2roi([proposals])
+        if self.with_bbox:
+            for i in range(self.num_stages):
+                bbox_results = self._bbox_forward(i, x, rois)
+                outs = outs + (bbox_results['cls_score'],
+                               bbox_results['bbox_pred'])
+        # mask heads
+        if self.with_mask:
+            mask_rois = rois[:100]
+            for i in range(self.num_stages):
+                mask_results = self._mask_forward(i, x, mask_rois)
+                outs = outs + (mask_results['mask_pred'], )
+        return outs
+
+    def _bbox_forward(self, stage, x, rois):
+        """Box head forward function used in both training and testing."""
+        bbox_roi_extractor = self.bbox_roi_extractor[stage]
+        bbox_head = self.bbox_head[stage]
+        bbox_feats = bbox_roi_extractor(x[:bbox_roi_extractor.num_inputs],
+                                        rois)
+        # do not support caffe_c4 model anymore
+        cls_score, bbox_pred = bbox_head(bbox_feats)
+
+        bbox_results = dict(
+            cls_score=cls_score, bbox_pred=bbox_pred, bbox_feats=bbox_feats)
+        return bbox_results
+
+    def _bbox_forward_train(self, stage, x, sampling_results, gt_bboxes,
+                            gt_labels, rcnn_train_cfg):
+        """Run forward function and calculate loss for box head in training."""
+        rois = bbox2roi([res.bboxes for res in sampling_results])
+        bbox_results = self._bbox_forward(stage, x, rois)
+        bbox_targets = self.bbox_head[stage].get_targets(
+            sampling_results, gt_bboxes, gt_labels, rcnn_train_cfg)
+        loss_bbox = self.bbox_head[stage].loss(bbox_results['cls_score'],
+                                               bbox_results['bbox_pred'], rois,
+                                               *bbox_targets)
+
+        bbox_results.update(
+            loss_bbox=loss_bbox, rois=rois, bbox_targets=bbox_targets)
+        return bbox_results
+
+    def _mask_forward(self, stage, x, rois):
+        """Mask head forward function used in both training and testing."""
+        mask_roi_extractor = self.mask_roi_extractor[stage]
+        mask_head = self.mask_head[stage]
+        mask_feats = mask_roi_extractor(x[:mask_roi_extractor.num_inputs],
+                                        rois)
+        # do not support caffe_c4 model anymore
+        mask_pred = mask_head(mask_feats)
+
+        mask_results = dict(mask_pred=mask_pred)
+        return mask_results
+
+    def _mask_forward_train(self,
+                            stage,
+                            x,
+                            sampling_results,
+                            gt_masks,
+                            rcnn_train_cfg,
+                            bbox_feats=None):
+        """Run forward function and calculate loss for mask head in
+        training."""
+        pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
+        mask_results = self._mask_forward(stage, x, pos_rois)
+
+        mask_targets = self.mask_head[stage].get_targets(
+            sampling_results, gt_masks, rcnn_train_cfg)
+        pos_labels = torch.cat([res.pos_gt_labels for res in sampling_results])
+        loss_mask = self.mask_head[stage].loss(mask_results['mask_pred'],
+                                               mask_targets, pos_labels)
+
+        mask_results.update(loss_mask=loss_mask)
+        return mask_results
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None):
+        """
+        Args:
+            x (list[Tensor]): list of multi-level img features.
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+            proposals (list[Tensors]): list of region proposals.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+            gt_masks (None | Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        losses = dict()
+        for i in range(self.num_stages):
+            self.current_stage = i
+            rcnn_train_cfg = self.train_cfg[i]
+            lw = self.stage_loss_weights[i]
+
+            # assign gts and sample proposals
+            sampling_results = []
+            if self.with_bbox or self.with_mask:
+                bbox_assigner = self.bbox_assigner[i]
+                bbox_sampler = self.bbox_sampler[i]
+                num_imgs = len(img_metas)
+                if gt_bboxes_ignore is None:
+                    gt_bboxes_ignore = [None for _ in range(num_imgs)]
+
+                for j in range(num_imgs):
+                    assign_result = bbox_assigner.assign(
+                        proposal_list[j], gt_bboxes[j], gt_bboxes_ignore[j],
+                        gt_labels[j])
+                    sampling_result = bbox_sampler.sample(
+                        assign_result,
+                        proposal_list[j],
+                        gt_bboxes[j],
+                        gt_labels[j],
+                        feats=[lvl_feat[j][None] for lvl_feat in x])
+                    sampling_results.append(sampling_result)
+
+            # bbox head forward and loss
+            bbox_results = self._bbox_forward_train(i, x, sampling_results,
+                                                    gt_bboxes, gt_labels,
+                                                    rcnn_train_cfg)
+
+            for name, value in bbox_results['loss_bbox'].items():
+                losses[f's{i}.{name}'] = (
+                    value * lw if 'loss' in name else value)
+
+            # mask head forward and loss
+            if self.with_mask:
+                mask_results = self._mask_forward_train(
+                    i, x, sampling_results, gt_masks, rcnn_train_cfg,
+                    bbox_results['bbox_feats'])
+                for name, value in mask_results['loss_mask'].items():
+                    losses[f's{i}.{name}'] = (
+                        value * lw if 'loss' in name else value)
+
+            # refine bboxes
+            if i < self.num_stages - 1:
+                pos_is_gts = [res.pos_is_gt for res in sampling_results]
+                # bbox_targets is a tuple
+                roi_labels = bbox_results['bbox_targets'][0]
+                with torch.no_grad():
+                    roi_labels = torch.where(
+                        roi_labels == self.bbox_head[i].num_classes,
+                        bbox_results['cls_score'][:, :-1].argmax(1),
+                        roi_labels)
+                    proposal_list = self.bbox_head[i].refine_bboxes(
+                        bbox_results['rois'], roi_labels,
+                        bbox_results['bbox_pred'], pos_is_gts, img_metas)
+
+        return losses
+
+    def simple_test(self, x, proposal_list, img_metas, rescale=False):
+        """Test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+        num_imgs = len(proposal_list)
+        img_shapes = tuple(meta['img_shape'] for meta in img_metas)
+        ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+
+        # "ms" in variable names means multi-stage
+        ms_bbox_result = {}
+        ms_segm_result = {}
+        ms_scores = []
+        rcnn_test_cfg = self.test_cfg
+
+        rois = bbox2roi(proposal_list)
+        for i in range(self.num_stages):
+            bbox_results = self._bbox_forward(i, x, rois)
+
+            # split batch bbox prediction back to each image
+            cls_score = bbox_results['cls_score']
+            bbox_pred = bbox_results['bbox_pred']
+            num_proposals_per_img = tuple(
+                len(proposals) for proposals in proposal_list)
+            rois = rois.split(num_proposals_per_img, 0)
+            cls_score = cls_score.split(num_proposals_per_img, 0)
+            if isinstance(bbox_pred, torch.Tensor):
+                bbox_pred = bbox_pred.split(num_proposals_per_img, 0)
+            else:
+                bbox_pred = self.bbox_head[i].bbox_pred_split(
+                    bbox_pred, num_proposals_per_img)
+            ms_scores.append(cls_score)
+
+            if i < self.num_stages - 1:
+                bbox_label = [s[:, :-1].argmax(dim=1) for s in cls_score]
+                rois = torch.cat([
+                    self.bbox_head[i].regress_by_class(rois[j], bbox_label[j],
+                                                       bbox_pred[j],
+                                                       img_metas[j])
+                    for j in range(num_imgs)
+                ])
+
+        # average scores of each image by stages
+        cls_score = [
+            sum([score[i] for score in ms_scores]) / float(len(ms_scores))
+            for i in range(num_imgs)
+        ]
+
+        # apply bbox post-processing to each image individually
+        det_bboxes = []
+        det_labels = []
+        for i in range(num_imgs):
+            det_bbox, det_label = self.bbox_head[-1].get_bboxes(
+                rois[i],
+                cls_score[i],
+                bbox_pred[i],
+                img_shapes[i],
+                scale_factors[i],
+                rescale=rescale,
+                cfg=rcnn_test_cfg)
+            det_bboxes.append(det_bbox)
+            det_labels.append(det_label)
+
+        if torch.onnx.is_in_onnx_export():
+            return det_bboxes, det_labels
+        bbox_results = [
+            bbox2result(det_bboxes[i], det_labels[i],
+                        self.bbox_head[-1].num_classes)
+            for i in range(num_imgs)
+        ]
+        ms_bbox_result['ensemble'] = bbox_results
+
+        if self.with_mask:
+            if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+                mask_classes = self.mask_head[-1].num_classes
+                segm_results = [[[] for _ in range(mask_classes)]
+                                for _ in range(num_imgs)]
+            else:
+                if rescale and not isinstance(scale_factors[0], float):
+                    scale_factors = [
+                        torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                        for scale_factor in scale_factors
+                    ]
+                _bboxes = [
+                    det_bboxes[i][:, :4] *
+                    scale_factors[i] if rescale else det_bboxes[i][:, :4]
+                    for i in range(len(det_bboxes))
+                ]
+                mask_rois = bbox2roi(_bboxes)
+                num_mask_rois_per_img = tuple(
+                    _bbox.size(0) for _bbox in _bboxes)
+                aug_masks = []
+                for i in range(self.num_stages):
+                    mask_results = self._mask_forward(i, x, mask_rois)
+                    mask_pred = mask_results['mask_pred']
+                    # split batch mask prediction back to each image
+                    mask_pred = mask_pred.split(num_mask_rois_per_img, 0)
+                    aug_masks.append(
+                        [m.sigmoid().cpu().numpy() for m in mask_pred])
+
+                # apply mask post-processing to each image individually
+                segm_results = []
+                for i in range(num_imgs):
+                    if det_bboxes[i].shape[0] == 0:
+                        segm_results.append(
+                            [[]
+                             for _ in range(self.mask_head[-1].num_classes)])
+                    else:
+                        aug_mask = [mask[i] for mask in aug_masks]
+                        merged_masks = merge_aug_masks(
+                            aug_mask, [[img_metas[i]]] * self.num_stages,
+                            rcnn_test_cfg)
+                        segm_result = self.mask_head[-1].get_seg_masks(
+                            merged_masks, _bboxes[i], det_labels[i],
+                            rcnn_test_cfg, ori_shapes[i], scale_factors[i],
+                            rescale)
+                        segm_results.append(segm_result)
+            ms_segm_result['ensemble'] = segm_results
+
+        if self.with_mask:
+            results = list(
+                zip(ms_bbox_result['ensemble'], ms_segm_result['ensemble']))
+        else:
+            results = ms_bbox_result['ensemble']
+
+        return results
+
+    def aug_test(self, features, proposal_list, img_metas, rescale=False):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        rcnn_test_cfg = self.test_cfg
+        aug_bboxes = []
+        aug_scores = []
+        for x, img_meta in zip(features, img_metas):
+            # only one image in the batch
+            img_shape = img_meta[0]['img_shape']
+            scale_factor = img_meta[0]['scale_factor']
+            flip = img_meta[0]['flip']
+            flip_direction = img_meta[0]['flip_direction']
+
+            proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
+                                     scale_factor, flip, flip_direction)
+            # "ms" in variable names means multi-stage
+            ms_scores = []
+
+            rois = bbox2roi([proposals])
+            for i in range(self.num_stages):
+                bbox_results = self._bbox_forward(i, x, rois)
+                ms_scores.append(bbox_results['cls_score'])
+
+                if i < self.num_stages - 1:
+                    bbox_label = bbox_results['cls_score'][:, :-1].argmax(
+                        dim=1)
+                    rois = self.bbox_head[i].regress_by_class(
+                        rois, bbox_label, bbox_results['bbox_pred'],
+                        img_meta[0])
+
+            cls_score = sum(ms_scores) / float(len(ms_scores))
+            bboxes, scores = self.bbox_head[-1].get_bboxes(
+                rois,
+                cls_score,
+                bbox_results['bbox_pred'],
+                img_shape,
+                scale_factor,
+                rescale=False,
+                cfg=None)
+            aug_bboxes.append(bboxes)
+            aug_scores.append(scores)
+
+        # after merging, bboxes will be rescaled to the original image size
+        merged_bboxes, merged_scores = merge_aug_bboxes(
+            aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
+        det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
+                                                rcnn_test_cfg.score_thr,
+                                                rcnn_test_cfg.nms,
+                                                rcnn_test_cfg.max_per_img)
+
+        bbox_result = bbox2result(det_bboxes, det_labels,
+                                  self.bbox_head[-1].num_classes)
+
+        if self.with_mask:
+            if det_bboxes.shape[0] == 0:
+                segm_result = [[[]
+                                for _ in range(self.mask_head[-1].num_classes)]
+                               ]
+            else:
+                aug_masks = []
+                aug_img_metas = []
+                for x, img_meta in zip(features, img_metas):
+                    img_shape = img_meta[0]['img_shape']
+                    scale_factor = img_meta[0]['scale_factor']
+                    flip = img_meta[0]['flip']
+                    flip_direction = img_meta[0]['flip_direction']
+                    _bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
+                                           scale_factor, flip, flip_direction)
+                    mask_rois = bbox2roi([_bboxes])
+                    for i in range(self.num_stages):
+                        mask_results = self._mask_forward(i, x, mask_rois)
+                        aug_masks.append(
+                            mask_results['mask_pred'].sigmoid().cpu().numpy())
+                        aug_img_metas.append(img_meta)
+                merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
+                                               self.test_cfg)
+
+                ori_shape = img_metas[0][0]['ori_shape']
+                segm_result = self.mask_head[-1].get_seg_masks(
+                    merged_masks,
+                    det_bboxes,
+                    det_labels,
+                    rcnn_test_cfg,
+                    ori_shape,
+                    scale_factor=1.0,
+                    rescale=False)
+            return [(bbox_result, segm_result)]
+        else:
+            return [bbox_result]
diff --git a/detection/scrfd/mmdet/models/roi_heads/double_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/double_roi_head.py
new file mode 100755
index 000000000..a1aa6c824
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/double_roi_head.py
@@ -0,0 +1,33 @@
+from ..builder import HEADS
+from .standard_roi_head import StandardRoIHead
+
+
+@HEADS.register_module()
+class DoubleHeadRoIHead(StandardRoIHead):
+    """RoI head for Double Head RCNN.
+
+    https://arxiv.org/abs/1904.06493
+    """
+
+    def __init__(self, reg_roi_scale_factor, **kwargs):
+        super(DoubleHeadRoIHead, self).__init__(**kwargs)
+        self.reg_roi_scale_factor = reg_roi_scale_factor
+
+    def _bbox_forward(self, x, rois):
+        """Box head forward function used in both training and testing time."""
+        bbox_cls_feats = self.bbox_roi_extractor(
+            x[:self.bbox_roi_extractor.num_inputs], rois)
+        bbox_reg_feats = self.bbox_roi_extractor(
+            x[:self.bbox_roi_extractor.num_inputs],
+            rois,
+            roi_scale_factor=self.reg_roi_scale_factor)
+        if self.with_shared_head:
+            bbox_cls_feats = self.shared_head(bbox_cls_feats)
+            bbox_reg_feats = self.shared_head(bbox_reg_feats)
+        cls_score, bbox_pred = self.bbox_head(bbox_cls_feats, bbox_reg_feats)
+
+        bbox_results = dict(
+            cls_score=cls_score,
+            bbox_pred=bbox_pred,
+            bbox_feats=bbox_cls_feats)
+        return bbox_results
diff --git a/detection/scrfd/mmdet/models/roi_heads/dynamic_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/dynamic_roi_head.py
new file mode 100755
index 000000000..89427a931
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/dynamic_roi_head.py
@@ -0,0 +1,154 @@
+import numpy as np
+import torch
+
+from mmdet.core import bbox2roi
+from mmdet.models.losses import SmoothL1Loss
+from ..builder import HEADS
+from .standard_roi_head import StandardRoIHead
+
+EPS = 1e-15
+
+
+@HEADS.register_module()
+class DynamicRoIHead(StandardRoIHead):
+    """RoI head for `Dynamic R-CNN <https://arxiv.org/abs/2004.06002>`_."""
+
+    def __init__(self, **kwargs):
+        super(DynamicRoIHead, self).__init__(**kwargs)
+        assert isinstance(self.bbox_head.loss_bbox, SmoothL1Loss)
+        # the IoU history of the past `update_iter_interval` iterations
+        self.iou_history = []
+        # the beta history of the past `update_iter_interval` iterations
+        self.beta_history = []
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None):
+        """Forward function for training.
+
+        Args:
+            x (list[Tensor]): list of multi-level img features.
+
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+
+            proposals (list[Tensors]): list of region proposals.
+
+            gt_bboxes (list[Tensor]): each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+
+            gt_labels (list[Tensor]): class indices corresponding to each box
+
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+            gt_masks (None | Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        # assign gts and sample proposals
+        if self.with_bbox or self.with_mask:
+            num_imgs = len(img_metas)
+            if gt_bboxes_ignore is None:
+                gt_bboxes_ignore = [None for _ in range(num_imgs)]
+            sampling_results = []
+            cur_iou = []
+            for i in range(num_imgs):
+                assign_result = self.bbox_assigner.assign(
+                    proposal_list[i], gt_bboxes[i], gt_bboxes_ignore[i],
+                    gt_labels[i])
+                sampling_result = self.bbox_sampler.sample(
+                    assign_result,
+                    proposal_list[i],
+                    gt_bboxes[i],
+                    gt_labels[i],
+                    feats=[lvl_feat[i][None] for lvl_feat in x])
+                # record the `iou_topk`-th largest IoU in an image
+                iou_topk = min(self.train_cfg.dynamic_rcnn.iou_topk,
+                               len(assign_result.max_overlaps))
+                ious, _ = torch.topk(assign_result.max_overlaps, iou_topk)
+                cur_iou.append(ious[-1].item())
+                sampling_results.append(sampling_result)
+            # average the current IoUs over images
+            cur_iou = np.mean(cur_iou)
+            self.iou_history.append(cur_iou)
+
+        losses = dict()
+        # bbox head forward and loss
+        if self.with_bbox:
+            bbox_results = self._bbox_forward_train(x, sampling_results,
+                                                    gt_bboxes, gt_labels,
+                                                    img_metas)
+            losses.update(bbox_results['loss_bbox'])
+
+        # mask head forward and loss
+        if self.with_mask:
+            mask_results = self._mask_forward_train(x, sampling_results,
+                                                    bbox_results['bbox_feats'],
+                                                    gt_masks, img_metas)
+            losses.update(mask_results['loss_mask'])
+
+        # update IoU threshold and SmoothL1 beta
+        update_iter_interval = self.train_cfg.dynamic_rcnn.update_iter_interval
+        if len(self.iou_history) % update_iter_interval == 0:
+            new_iou_thr, new_beta = self.update_hyperparameters()
+
+        return losses
+
+    def _bbox_forward_train(self, x, sampling_results, gt_bboxes, gt_labels,
+                            img_metas):
+        num_imgs = len(img_metas)
+        rois = bbox2roi([res.bboxes for res in sampling_results])
+        bbox_results = self._bbox_forward(x, rois)
+
+        bbox_targets = self.bbox_head.get_targets(sampling_results, gt_bboxes,
+                                                  gt_labels, self.train_cfg)
+        # record the `beta_topk`-th smallest target
+        # `bbox_targets[2]` and `bbox_targets[3]` stand for bbox_targets
+        # and bbox_weights, respectively
+        pos_inds = bbox_targets[3][:, 0].nonzero().squeeze(1)
+        num_pos = len(pos_inds)
+        cur_target = bbox_targets[2][pos_inds, :2].abs().mean(dim=1)
+        beta_topk = min(self.train_cfg.dynamic_rcnn.beta_topk * num_imgs,
+                        num_pos)
+        cur_target = torch.kthvalue(cur_target, beta_topk)[0].item()
+        self.beta_history.append(cur_target)
+        loss_bbox = self.bbox_head.loss(bbox_results['cls_score'],
+                                        bbox_results['bbox_pred'], rois,
+                                        *bbox_targets)
+
+        bbox_results.update(loss_bbox=loss_bbox)
+        return bbox_results
+
+    def update_hyperparameters(self):
+        """Update hyperparameters like IoU thresholds for assigner and beta for
+        SmoothL1 loss based on the training statistics.
+
+        Returns:
+            tuple[float]: the updated ``iou_thr`` and ``beta``.
+        """
+        new_iou_thr = max(self.train_cfg.dynamic_rcnn.initial_iou,
+                          np.mean(self.iou_history))
+        self.iou_history = []
+        self.bbox_assigner.pos_iou_thr = new_iou_thr
+        self.bbox_assigner.neg_iou_thr = new_iou_thr
+        self.bbox_assigner.min_pos_iou = new_iou_thr
+        if (np.median(self.beta_history) < EPS):
+            # avoid 0 or too small value for new_beta
+            new_beta = self.bbox_head.loss_bbox.beta
+        else:
+            new_beta = min(self.train_cfg.dynamic_rcnn.initial_beta,
+                           np.median(self.beta_history))
+        self.beta_history = []
+        self.bbox_head.loss_bbox.beta = new_beta
+        return new_iou_thr, new_beta
diff --git a/detection/scrfd/mmdet/models/roi_heads/grid_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/grid_roi_head.py
new file mode 100755
index 000000000..4c52c7986
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/grid_roi_head.py
@@ -0,0 +1,176 @@
+import torch
+
+from mmdet.core import bbox2result, bbox2roi
+from ..builder import HEADS, build_head, build_roi_extractor
+from .standard_roi_head import StandardRoIHead
+
+
+@HEADS.register_module()
+class GridRoIHead(StandardRoIHead):
+    """Grid roi head for Grid R-CNN.
+
+    https://arxiv.org/abs/1811.12030
+    """
+
+    def __init__(self, grid_roi_extractor, grid_head, **kwargs):
+        assert grid_head is not None
+        super(GridRoIHead, self).__init__(**kwargs)
+        if grid_roi_extractor is not None:
+            self.grid_roi_extractor = build_roi_extractor(grid_roi_extractor)
+            self.share_roi_extractor = False
+        else:
+            self.share_roi_extractor = True
+            self.grid_roi_extractor = self.bbox_roi_extractor
+        self.grid_head = build_head(grid_head)
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(GridRoIHead, self).init_weights(pretrained)
+        self.grid_head.init_weights()
+        if not self.share_roi_extractor:
+            self.grid_roi_extractor.init_weights()
+
+    def _random_jitter(self, sampling_results, img_metas, amplitude=0.15):
+        """Ramdom jitter positive proposals for training."""
+        for sampling_result, img_meta in zip(sampling_results, img_metas):
+            bboxes = sampling_result.pos_bboxes
+            random_offsets = bboxes.new_empty(bboxes.shape[0], 4).uniform_(
+                -amplitude, amplitude)
+            # before jittering
+            cxcy = (bboxes[:, 2:4] + bboxes[:, :2]) / 2
+            wh = (bboxes[:, 2:4] - bboxes[:, :2]).abs()
+            # after jittering
+            new_cxcy = cxcy + wh * random_offsets[:, :2]
+            new_wh = wh * (1 + random_offsets[:, 2:])
+            # xywh to xyxy
+            new_x1y1 = (new_cxcy - new_wh / 2)
+            new_x2y2 = (new_cxcy + new_wh / 2)
+            new_bboxes = torch.cat([new_x1y1, new_x2y2], dim=1)
+            # clip bboxes
+            max_shape = img_meta['img_shape']
+            if max_shape is not None:
+                new_bboxes[:, 0::2].clamp_(min=0, max=max_shape[1] - 1)
+                new_bboxes[:, 1::2].clamp_(min=0, max=max_shape[0] - 1)
+
+            sampling_result.pos_bboxes = new_bboxes
+        return sampling_results
+
+    def forward_dummy(self, x, proposals):
+        """Dummy forward function."""
+        # bbox head
+        outs = ()
+        rois = bbox2roi([proposals])
+        if self.with_bbox:
+            bbox_results = self._bbox_forward(x, rois)
+            outs = outs + (bbox_results['cls_score'],
+                           bbox_results['bbox_pred'])
+
+        # grid head
+        grid_rois = rois[:100]
+        grid_feats = self.grid_roi_extractor(
+            x[:self.grid_roi_extractor.num_inputs], grid_rois)
+        if self.with_shared_head:
+            grid_feats = self.shared_head(grid_feats)
+        grid_pred = self.grid_head(grid_feats)
+        outs = outs + (grid_pred, )
+
+        # mask head
+        if self.with_mask:
+            mask_rois = rois[:100]
+            mask_results = self._mask_forward(x, mask_rois)
+            outs = outs + (mask_results['mask_pred'], )
+        return outs
+
+    def _bbox_forward_train(self, x, sampling_results, gt_bboxes, gt_labels,
+                            img_metas):
+        """Run forward function and calculate loss for box head in training."""
+        bbox_results = super(GridRoIHead,
+                             self)._bbox_forward_train(x, sampling_results,
+                                                       gt_bboxes, gt_labels,
+                                                       img_metas)
+
+        # Grid head forward and loss
+        sampling_results = self._random_jitter(sampling_results, img_metas)
+        pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
+
+        # GN in head does not support zero shape input
+        if pos_rois.shape[0] == 0:
+            return bbox_results
+
+        grid_feats = self.grid_roi_extractor(
+            x[:self.grid_roi_extractor.num_inputs], pos_rois)
+        if self.with_shared_head:
+            grid_feats = self.shared_head(grid_feats)
+        # Accelerate training
+        max_sample_num_grid = self.train_cfg.get('max_num_grid', 192)
+        sample_idx = torch.randperm(
+            grid_feats.shape[0])[:min(grid_feats.shape[0], max_sample_num_grid
+                                      )]
+        grid_feats = grid_feats[sample_idx]
+
+        grid_pred = self.grid_head(grid_feats)
+
+        grid_targets = self.grid_head.get_targets(sampling_results,
+                                                  self.train_cfg)
+        grid_targets = grid_targets[sample_idx]
+
+        loss_grid = self.grid_head.loss(grid_pred, grid_targets)
+
+        bbox_results['loss_bbox'].update(loss_grid)
+        return bbox_results
+
+    def simple_test(self,
+                    x,
+                    proposal_list,
+                    img_metas,
+                    proposals=None,
+                    rescale=False):
+        """Test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+
+        det_bboxes, det_labels = self.simple_test_bboxes(
+            x, img_metas, proposal_list, self.test_cfg, rescale=False)
+        # pack rois into bboxes
+        grid_rois = bbox2roi([det_bbox[:, :4] for det_bbox in det_bboxes])
+        if grid_rois.shape[0] != 0:
+            grid_feats = self.grid_roi_extractor(
+                x[:len(self.grid_roi_extractor.featmap_strides)], grid_rois)
+            self.grid_head.test_mode = True
+            grid_pred = self.grid_head(grid_feats)
+            # split batch grid head prediction back to each image
+            num_roi_per_img = tuple(len(det_bbox) for det_bbox in det_bboxes)
+            grid_pred = {
+                k: v.split(num_roi_per_img, 0)
+                for k, v in grid_pred.items()
+            }
+
+            # apply bbox post-processing to each image individually
+            bbox_results = []
+            num_imgs = len(det_bboxes)
+            for i in range(num_imgs):
+                if det_bboxes[i].shape[0] == 0:
+                    bbox_results.append(grid_rois.new_tensor([]))
+                else:
+                    det_bbox = self.grid_head.get_bboxes(
+                        det_bboxes[i], grid_pred['fused'][i], [img_metas[i]])
+                    if rescale:
+                        det_bbox[:, :4] /= img_metas[i]['scale_factor']
+                    bbox_results.append(
+                        bbox2result(det_bbox, det_labels[i],
+                                    self.bbox_head.num_classes))
+        else:
+            bbox_results = [
+                grid_rois.new_tensor([]) for _ in range(len(det_bboxes))
+            ]
+
+        if not self.with_mask:
+            return bbox_results
+        else:
+            segm_results = self.simple_test_mask(
+                x, img_metas, det_bboxes, det_labels, rescale=rescale)
+            return list(zip(bbox_results, segm_results))
diff --git a/detection/scrfd/mmdet/models/roi_heads/htc_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/htc_roi_head.py
new file mode 100755
index 000000000..e2d3d7384
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/htc_roi_head.py
@@ -0,0 +1,589 @@
+import torch
+import torch.nn.functional as F
+
+from mmdet.core import (bbox2result, bbox2roi, bbox_mapping, merge_aug_bboxes,
+                        merge_aug_masks, multiclass_nms)
+from ..builder import HEADS, build_head, build_roi_extractor
+from .cascade_roi_head import CascadeRoIHead
+
+
+@HEADS.register_module()
+class HybridTaskCascadeRoIHead(CascadeRoIHead):
+    """Hybrid task cascade roi head including one bbox head and one mask head.
+
+    https://arxiv.org/abs/1901.07518
+    """
+
+    def __init__(self,
+                 num_stages,
+                 stage_loss_weights,
+                 semantic_roi_extractor=None,
+                 semantic_head=None,
+                 semantic_fusion=('bbox', 'mask'),
+                 interleaved=True,
+                 mask_info_flow=True,
+                 **kwargs):
+        super(HybridTaskCascadeRoIHead,
+              self).__init__(num_stages, stage_loss_weights, **kwargs)
+        assert self.with_bbox and self.with_mask
+        assert not self.with_shared_head  # shared head is not supported
+
+        if semantic_head is not None:
+            self.semantic_roi_extractor = build_roi_extractor(
+                semantic_roi_extractor)
+            self.semantic_head = build_head(semantic_head)
+
+        self.semantic_fusion = semantic_fusion
+        self.interleaved = interleaved
+        self.mask_info_flow = mask_info_flow
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(HybridTaskCascadeRoIHead, self).init_weights(pretrained)
+        if self.with_semantic:
+            self.semantic_head.init_weights()
+
+    @property
+    def with_semantic(self):
+        """bool: whether the head has semantic head"""
+        if hasattr(self, 'semantic_head') and self.semantic_head is not None:
+            return True
+        else:
+            return False
+
+    def forward_dummy(self, x, proposals):
+        """Dummy forward function."""
+        outs = ()
+        # semantic head
+        if self.with_semantic:
+            _, semantic_feat = self.semantic_head(x)
+        else:
+            semantic_feat = None
+        # bbox heads
+        rois = bbox2roi([proposals])
+        for i in range(self.num_stages):
+            bbox_results = self._bbox_forward(
+                i, x, rois, semantic_feat=semantic_feat)
+            outs = outs + (bbox_results['cls_score'],
+                           bbox_results['bbox_pred'])
+        # mask heads
+        if self.with_mask:
+            mask_rois = rois[:100]
+            mask_roi_extractor = self.mask_roi_extractor[-1]
+            mask_feats = mask_roi_extractor(
+                x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
+            if self.with_semantic and 'mask' in self.semantic_fusion:
+                mask_semantic_feat = self.semantic_roi_extractor(
+                    [semantic_feat], mask_rois)
+                mask_feats += mask_semantic_feat
+            last_feat = None
+            for i in range(self.num_stages):
+                mask_head = self.mask_head[i]
+                if self.mask_info_flow:
+                    mask_pred, last_feat = mask_head(mask_feats, last_feat)
+                else:
+                    mask_pred = mask_head(mask_feats)
+                outs = outs + (mask_pred, )
+        return outs
+
+    def _bbox_forward_train(self,
+                            stage,
+                            x,
+                            sampling_results,
+                            gt_bboxes,
+                            gt_labels,
+                            rcnn_train_cfg,
+                            semantic_feat=None):
+        """Run forward function and calculate loss for box head in training."""
+        bbox_head = self.bbox_head[stage]
+        rois = bbox2roi([res.bboxes for res in sampling_results])
+        bbox_results = self._bbox_forward(
+            stage, x, rois, semantic_feat=semantic_feat)
+
+        bbox_targets = bbox_head.get_targets(sampling_results, gt_bboxes,
+                                             gt_labels, rcnn_train_cfg)
+        loss_bbox = bbox_head.loss(bbox_results['cls_score'],
+                                   bbox_results['bbox_pred'], rois,
+                                   *bbox_targets)
+
+        bbox_results.update(
+            loss_bbox=loss_bbox,
+            rois=rois,
+            bbox_targets=bbox_targets,
+        )
+        return bbox_results
+
+    def _mask_forward_train(self,
+                            stage,
+                            x,
+                            sampling_results,
+                            gt_masks,
+                            rcnn_train_cfg,
+                            semantic_feat=None):
+        """Run forward function and calculate loss for mask head in
+        training."""
+        mask_roi_extractor = self.mask_roi_extractor[stage]
+        mask_head = self.mask_head[stage]
+        pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
+        mask_feats = mask_roi_extractor(x[:mask_roi_extractor.num_inputs],
+                                        pos_rois)
+
+        # semantic feature fusion
+        # element-wise sum for original features and pooled semantic features
+        if self.with_semantic and 'mask' in self.semantic_fusion:
+            mask_semantic_feat = self.semantic_roi_extractor([semantic_feat],
+                                                             pos_rois)
+            if mask_semantic_feat.shape[-2:] != mask_feats.shape[-2:]:
+                mask_semantic_feat = F.adaptive_avg_pool2d(
+                    mask_semantic_feat, mask_feats.shape[-2:])
+            mask_feats += mask_semantic_feat
+
+        # mask information flow
+        # forward all previous mask heads to obtain last_feat, and fuse it
+        # with the normal mask feature
+        if self.mask_info_flow:
+            last_feat = None
+            for i in range(stage):
+                last_feat = self.mask_head[i](
+                    mask_feats, last_feat, return_logits=False)
+            mask_pred = mask_head(mask_feats, last_feat, return_feat=False)
+        else:
+            mask_pred = mask_head(mask_feats, return_feat=False)
+
+        mask_targets = mask_head.get_targets(sampling_results, gt_masks,
+                                             rcnn_train_cfg)
+        pos_labels = torch.cat([res.pos_gt_labels for res in sampling_results])
+        loss_mask = mask_head.loss(mask_pred, mask_targets, pos_labels)
+
+        mask_results = dict(loss_mask=loss_mask)
+        return mask_results
+
+    def _bbox_forward(self, stage, x, rois, semantic_feat=None):
+        """Box head forward function used in both training and testing."""
+        bbox_roi_extractor = self.bbox_roi_extractor[stage]
+        bbox_head = self.bbox_head[stage]
+        bbox_feats = bbox_roi_extractor(
+            x[:len(bbox_roi_extractor.featmap_strides)], rois)
+        if self.with_semantic and 'bbox' in self.semantic_fusion:
+            bbox_semantic_feat = self.semantic_roi_extractor([semantic_feat],
+                                                             rois)
+            if bbox_semantic_feat.shape[-2:] != bbox_feats.shape[-2:]:
+                bbox_semantic_feat = F.adaptive_avg_pool2d(
+                    bbox_semantic_feat, bbox_feats.shape[-2:])
+            bbox_feats += bbox_semantic_feat
+        cls_score, bbox_pred = bbox_head(bbox_feats)
+
+        bbox_results = dict(cls_score=cls_score, bbox_pred=bbox_pred)
+        return bbox_results
+
+    def _mask_forward_test(self, stage, x, bboxes, semantic_feat=None):
+        """Mask head forward function for testing."""
+        mask_roi_extractor = self.mask_roi_extractor[stage]
+        mask_head = self.mask_head[stage]
+        mask_rois = bbox2roi([bboxes])
+        mask_feats = mask_roi_extractor(
+            x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
+        if self.with_semantic and 'mask' in self.semantic_fusion:
+            mask_semantic_feat = self.semantic_roi_extractor([semantic_feat],
+                                                             mask_rois)
+            if mask_semantic_feat.shape[-2:] != mask_feats.shape[-2:]:
+                mask_semantic_feat = F.adaptive_avg_pool2d(
+                    mask_semantic_feat, mask_feats.shape[-2:])
+            mask_feats += mask_semantic_feat
+        if self.mask_info_flow:
+            last_feat = None
+            last_pred = None
+            for i in range(stage):
+                mask_pred, last_feat = self.mask_head[i](mask_feats, last_feat)
+                if last_pred is not None:
+                    mask_pred = mask_pred + last_pred
+                last_pred = mask_pred
+            mask_pred = mask_head(mask_feats, last_feat, return_feat=False)
+            if last_pred is not None:
+                mask_pred = mask_pred + last_pred
+        else:
+            mask_pred = mask_head(mask_feats)
+        return mask_pred
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None,
+                      gt_semantic_seg=None):
+        """
+        Args:
+            x (list[Tensor]): list of multi-level img features.
+
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+
+            proposal_list (list[Tensors]): list of region proposals.
+
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+
+            gt_labels (list[Tensor]): class indices corresponding to each box
+
+            gt_bboxes_ignore (None, list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+
+            gt_masks (None, Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+            gt_semantic_seg (None, list[Tensor]): semantic segmentation masks
+                used if the architecture supports semantic segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        # semantic segmentation part
+        # 2 outputs: segmentation prediction and embedded features
+        losses = dict()
+        if self.with_semantic:
+            semantic_pred, semantic_feat = self.semantic_head(x)
+            loss_seg = self.semantic_head.loss(semantic_pred, gt_semantic_seg)
+            losses['loss_semantic_seg'] = loss_seg
+        else:
+            semantic_feat = None
+
+        for i in range(self.num_stages):
+            self.current_stage = i
+            rcnn_train_cfg = self.train_cfg[i]
+            lw = self.stage_loss_weights[i]
+
+            # assign gts and sample proposals
+            sampling_results = []
+            bbox_assigner = self.bbox_assigner[i]
+            bbox_sampler = self.bbox_sampler[i]
+            num_imgs = len(img_metas)
+            if gt_bboxes_ignore is None:
+                gt_bboxes_ignore = [None for _ in range(num_imgs)]
+
+            for j in range(num_imgs):
+                assign_result = bbox_assigner.assign(proposal_list[j],
+                                                     gt_bboxes[j],
+                                                     gt_bboxes_ignore[j],
+                                                     gt_labels[j])
+                sampling_result = bbox_sampler.sample(
+                    assign_result,
+                    proposal_list[j],
+                    gt_bboxes[j],
+                    gt_labels[j],
+                    feats=[lvl_feat[j][None] for lvl_feat in x])
+                sampling_results.append(sampling_result)
+
+            # bbox head forward and loss
+            bbox_results = \
+                self._bbox_forward_train(
+                    i, x, sampling_results, gt_bboxes, gt_labels,
+                    rcnn_train_cfg, semantic_feat)
+            roi_labels = bbox_results['bbox_targets'][0]
+
+            for name, value in bbox_results['loss_bbox'].items():
+                losses[f's{i}.{name}'] = (
+                    value * lw if 'loss' in name else value)
+
+            # mask head forward and loss
+            if self.with_mask:
+                # interleaved execution: use regressed bboxes by the box branch
+                # to train the mask branch
+                if self.interleaved:
+                    pos_is_gts = [res.pos_is_gt for res in sampling_results]
+                    with torch.no_grad():
+                        proposal_list = self.bbox_head[i].refine_bboxes(
+                            bbox_results['rois'], roi_labels,
+                            bbox_results['bbox_pred'], pos_is_gts, img_metas)
+                        # re-assign and sample 512 RoIs from 512 RoIs
+                        sampling_results = []
+                        for j in range(num_imgs):
+                            assign_result = bbox_assigner.assign(
+                                proposal_list[j], gt_bboxes[j],
+                                gt_bboxes_ignore[j], gt_labels[j])
+                            sampling_result = bbox_sampler.sample(
+                                assign_result,
+                                proposal_list[j],
+                                gt_bboxes[j],
+                                gt_labels[j],
+                                feats=[lvl_feat[j][None] for lvl_feat in x])
+                            sampling_results.append(sampling_result)
+                mask_results = self._mask_forward_train(
+                    i, x, sampling_results, gt_masks, rcnn_train_cfg,
+                    semantic_feat)
+                for name, value in mask_results['loss_mask'].items():
+                    losses[f's{i}.{name}'] = (
+                        value * lw if 'loss' in name else value)
+
+            # refine bboxes (same as Cascade R-CNN)
+            if i < self.num_stages - 1 and not self.interleaved:
+                pos_is_gts = [res.pos_is_gt for res in sampling_results]
+                with torch.no_grad():
+                    proposal_list = self.bbox_head[i].refine_bboxes(
+                        bbox_results['rois'], roi_labels,
+                        bbox_results['bbox_pred'], pos_is_gts, img_metas)
+
+        return losses
+
+    def simple_test(self, x, proposal_list, img_metas, rescale=False):
+        """Test without augmentation."""
+        if self.with_semantic:
+            _, semantic_feat = self.semantic_head(x)
+        else:
+            semantic_feat = None
+
+        num_imgs = len(proposal_list)
+        img_shapes = tuple(meta['img_shape'] for meta in img_metas)
+        ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+
+        # "ms" in variable names means multi-stage
+        ms_bbox_result = {}
+        ms_segm_result = {}
+        ms_scores = []
+        rcnn_test_cfg = self.test_cfg
+
+        rois = bbox2roi(proposal_list)
+        for i in range(self.num_stages):
+            bbox_head = self.bbox_head[i]
+            bbox_results = self._bbox_forward(
+                i, x, rois, semantic_feat=semantic_feat)
+            # split batch bbox prediction back to each image
+            cls_score = bbox_results['cls_score']
+            bbox_pred = bbox_results['bbox_pred']
+            num_proposals_per_img = tuple(len(p) for p in proposal_list)
+            rois = rois.split(num_proposals_per_img, 0)
+            cls_score = cls_score.split(num_proposals_per_img, 0)
+            bbox_pred = bbox_pred.split(num_proposals_per_img, 0)
+            ms_scores.append(cls_score)
+
+            if i < self.num_stages - 1:
+                bbox_label = [s[:, :-1].argmax(dim=1) for s in cls_score]
+                rois = torch.cat([
+                    bbox_head.regress_by_class(rois[i], bbox_label[i],
+                                               bbox_pred[i], img_metas[i])
+                    for i in range(num_imgs)
+                ])
+
+        # average scores of each image by stages
+        cls_score = [
+            sum([score[i] for score in ms_scores]) / float(len(ms_scores))
+            for i in range(num_imgs)
+        ]
+
+        # apply bbox post-processing to each image individually
+        det_bboxes = []
+        det_labels = []
+        for i in range(num_imgs):
+            det_bbox, det_label = self.bbox_head[-1].get_bboxes(
+                rois[i],
+                cls_score[i],
+                bbox_pred[i],
+                img_shapes[i],
+                scale_factors[i],
+                rescale=rescale,
+                cfg=rcnn_test_cfg)
+            det_bboxes.append(det_bbox)
+            det_labels.append(det_label)
+        bbox_result = [
+            bbox2result(det_bboxes[i], det_labels[i],
+                        self.bbox_head[-1].num_classes)
+            for i in range(num_imgs)
+        ]
+        ms_bbox_result['ensemble'] = bbox_result
+
+        if self.with_mask:
+            if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+                mask_classes = self.mask_head[-1].num_classes
+                segm_results = [[[] for _ in range(mask_classes)]
+                                for _ in range(num_imgs)]
+            else:
+                if rescale and not isinstance(scale_factors[0], float):
+                    scale_factors = [
+                        torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                        for scale_factor in scale_factors
+                    ]
+                _bboxes = [
+                    det_bboxes[i][:, :4] *
+                    scale_factors[i] if rescale else det_bboxes[i]
+                    for i in range(num_imgs)
+                ]
+                mask_rois = bbox2roi(_bboxes)
+                aug_masks = []
+                mask_roi_extractor = self.mask_roi_extractor[-1]
+                mask_feats = mask_roi_extractor(
+                    x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
+                if self.with_semantic and 'mask' in self.semantic_fusion:
+                    mask_semantic_feat = self.semantic_roi_extractor(
+                        [semantic_feat], mask_rois)
+                    mask_feats += mask_semantic_feat
+                last_feat = None
+
+                num_bbox_per_img = tuple(len(_bbox) for _bbox in _bboxes)
+                for i in range(self.num_stages):
+                    mask_head = self.mask_head[i]
+                    if self.mask_info_flow:
+                        mask_pred, last_feat = mask_head(mask_feats, last_feat)
+                    else:
+                        mask_pred = mask_head(mask_feats)
+
+                    # split batch mask prediction back to each image
+                    mask_pred = mask_pred.split(num_bbox_per_img, 0)
+                    aug_masks.append(
+                        [mask.sigmoid().cpu().numpy() for mask in mask_pred])
+
+                # apply mask post-processing to each image individually
+                segm_results = []
+                for i in range(num_imgs):
+                    if det_bboxes[i].shape[0] == 0:
+                        segm_results.append(
+                            [[]
+                             for _ in range(self.mask_head[-1].num_classes)])
+                    else:
+                        aug_mask = [mask[i] for mask in aug_masks]
+                        merged_mask = merge_aug_masks(
+                            aug_mask, [[img_metas[i]]] * self.num_stages,
+                            rcnn_test_cfg)
+                        segm_result = self.mask_head[-1].get_seg_masks(
+                            merged_mask, _bboxes[i], det_labels[i],
+                            rcnn_test_cfg, ori_shapes[i], scale_factors[i],
+                            rescale)
+                        segm_results.append(segm_result)
+            ms_segm_result['ensemble'] = segm_results
+
+        if self.with_mask:
+            results = list(
+                zip(ms_bbox_result['ensemble'], ms_segm_result['ensemble']))
+        else:
+            results = ms_bbox_result['ensemble']
+
+        return results
+
+    def aug_test(self, img_feats, proposal_list, img_metas, rescale=False):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        if self.with_semantic:
+            semantic_feats = [
+                self.semantic_head(feat)[1] for feat in img_feats
+            ]
+        else:
+            semantic_feats = [None] * len(img_metas)
+
+        rcnn_test_cfg = self.test_cfg
+        aug_bboxes = []
+        aug_scores = []
+        for x, img_meta, semantic in zip(img_feats, img_metas, semantic_feats):
+            # only one image in the batch
+            img_shape = img_meta[0]['img_shape']
+            scale_factor = img_meta[0]['scale_factor']
+            flip = img_meta[0]['flip']
+            flip_direction = img_meta[0]['flip_direction']
+
+            proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
+                                     scale_factor, flip, flip_direction)
+            # "ms" in variable names means multi-stage
+            ms_scores = []
+
+            rois = bbox2roi([proposals])
+            for i in range(self.num_stages):
+                bbox_head = self.bbox_head[i]
+                bbox_results = self._bbox_forward(
+                    i, x, rois, semantic_feat=semantic)
+                ms_scores.append(bbox_results['cls_score'])
+
+                if i < self.num_stages - 1:
+                    bbox_label = bbox_results['cls_score'].argmax(dim=1)
+                    rois = bbox_head.regress_by_class(
+                        rois, bbox_label, bbox_results['bbox_pred'],
+                        img_meta[0])
+
+            cls_score = sum(ms_scores) / float(len(ms_scores))
+            bboxes, scores = self.bbox_head[-1].get_bboxes(
+                rois,
+                cls_score,
+                bbox_results['bbox_pred'],
+                img_shape,
+                scale_factor,
+                rescale=False,
+                cfg=None)
+            aug_bboxes.append(bboxes)
+            aug_scores.append(scores)
+
+        # after merging, bboxes will be rescaled to the original image size
+        merged_bboxes, merged_scores = merge_aug_bboxes(
+            aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
+        det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
+                                                rcnn_test_cfg.score_thr,
+                                                rcnn_test_cfg.nms,
+                                                rcnn_test_cfg.max_per_img)
+
+        bbox_result = bbox2result(det_bboxes, det_labels,
+                                  self.bbox_head[-1].num_classes)
+
+        if self.with_mask:
+            if det_bboxes.shape[0] == 0:
+                segm_result = [[[]
+                                for _ in range(self.mask_head[-1].num_classes -
+                                               1)]]
+            else:
+                aug_masks = []
+                aug_img_metas = []
+                for x, img_meta, semantic in zip(img_feats, img_metas,
+                                                 semantic_feats):
+                    img_shape = img_meta[0]['img_shape']
+                    scale_factor = img_meta[0]['scale_factor']
+                    flip = img_meta[0]['flip']
+                    flip_direction = img_meta[0]['flip_direction']
+                    _bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
+                                           scale_factor, flip, flip_direction)
+                    mask_rois = bbox2roi([_bboxes])
+                    mask_feats = self.mask_roi_extractor[-1](
+                        x[:len(self.mask_roi_extractor[-1].featmap_strides)],
+                        mask_rois)
+                    if self.with_semantic:
+                        semantic_feat = semantic
+                        mask_semantic_feat = self.semantic_roi_extractor(
+                            [semantic_feat], mask_rois)
+                        if mask_semantic_feat.shape[-2:] != mask_feats.shape[
+                                -2:]:
+                            mask_semantic_feat = F.adaptive_avg_pool2d(
+                                mask_semantic_feat, mask_feats.shape[-2:])
+                        mask_feats += mask_semantic_feat
+                    last_feat = None
+                    for i in range(self.num_stages):
+                        mask_head = self.mask_head[i]
+                        if self.mask_info_flow:
+                            mask_pred, last_feat = mask_head(
+                                mask_feats, last_feat)
+                        else:
+                            mask_pred = mask_head(mask_feats)
+                        aug_masks.append(mask_pred.sigmoid().cpu().numpy())
+                        aug_img_metas.append(img_meta)
+                merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
+                                               self.test_cfg)
+
+                ori_shape = img_metas[0][0]['ori_shape']
+                segm_result = self.mask_head[-1].get_seg_masks(
+                    merged_masks,
+                    det_bboxes,
+                    det_labels,
+                    rcnn_test_cfg,
+                    ori_shape,
+                    scale_factor=1.0,
+                    rescale=False)
+            return [(bbox_result, segm_result)]
+        else:
+            return [bbox_result]
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/__init__.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/__init__.py
new file mode 100755
index 000000000..26c3e95a6
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/__init__.py
@@ -0,0 +1,12 @@
+from .coarse_mask_head import CoarseMaskHead
+from .fcn_mask_head import FCNMaskHead
+from .fused_semantic_head import FusedSemanticHead
+from .grid_head import GridHead
+from .htc_mask_head import HTCMaskHead
+from .mask_point_head import MaskPointHead
+from .maskiou_head import MaskIoUHead
+
+__all__ = [
+    'FCNMaskHead', 'HTCMaskHead', 'FusedSemanticHead', 'GridHead',
+    'MaskIoUHead', 'CoarseMaskHead', 'MaskPointHead'
+]
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/coarse_mask_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/coarse_mask_head.py
new file mode 100755
index 000000000..d665dfff8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/coarse_mask_head.py
@@ -0,0 +1,91 @@
+import torch.nn as nn
+from mmcv.cnn import ConvModule, Linear, constant_init, xavier_init
+from mmcv.runner import auto_fp16
+
+from mmdet.models.builder import HEADS
+from .fcn_mask_head import FCNMaskHead
+
+
+@HEADS.register_module()
+class CoarseMaskHead(FCNMaskHead):
+    """Coarse mask head used in PointRend.
+
+    Compared with standard ``FCNMaskHead``, ``CoarseMaskHead`` will downsample
+    the input feature map instead of upsample it.
+
+    Args:
+        num_convs (int): Number of conv layers in the head. Default: 0.
+        num_fcs (int): Number of fc layers in the head. Default: 2.
+        fc_out_channels (int): Number of output channels of fc layer.
+            Default: 1024.
+        downsample_factor (int): The factor that feature map is downsampled by.
+            Default: 2.
+    """
+
+    def __init__(self,
+                 num_convs=0,
+                 num_fcs=2,
+                 fc_out_channels=1024,
+                 downsample_factor=2,
+                 *arg,
+                 **kwarg):
+        super(CoarseMaskHead, self).__init__(
+            *arg, num_convs=num_convs, upsample_cfg=dict(type=None), **kwarg)
+        self.num_fcs = num_fcs
+        assert self.num_fcs > 0
+        self.fc_out_channels = fc_out_channels
+        self.downsample_factor = downsample_factor
+        assert self.downsample_factor >= 1
+        # remove conv_logit
+        delattr(self, 'conv_logits')
+
+        if downsample_factor > 1:
+            downsample_in_channels = (
+                self.conv_out_channels
+                if self.num_convs > 0 else self.in_channels)
+            self.downsample_conv = ConvModule(
+                downsample_in_channels,
+                self.conv_out_channels,
+                kernel_size=downsample_factor,
+                stride=downsample_factor,
+                padding=0,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg)
+        else:
+            self.downsample_conv = None
+
+        self.output_size = (self.roi_feat_size[0] // downsample_factor,
+                            self.roi_feat_size[1] // downsample_factor)
+        self.output_area = self.output_size[0] * self.output_size[1]
+
+        last_layer_dim = self.conv_out_channels * self.output_area
+
+        self.fcs = nn.ModuleList()
+        for i in range(num_fcs):
+            fc_in_channels = (
+                last_layer_dim if i == 0 else self.fc_out_channels)
+            self.fcs.append(Linear(fc_in_channels, self.fc_out_channels))
+        last_layer_dim = self.fc_out_channels
+        output_channels = self.num_classes * self.output_area
+        self.fc_logits = Linear(last_layer_dim, output_channels)
+
+    def init_weights(self):
+        for m in self.fcs.modules():
+            if isinstance(m, nn.Linear):
+                xavier_init(m)
+        constant_init(self.fc_logits, 0.001)
+
+    @auto_fp16()
+    def forward(self, x):
+        for conv in self.convs:
+            x = conv(x)
+
+        if self.downsample_conv is not None:
+            x = self.downsample_conv(x)
+
+        x = x.flatten(1)
+        for fc in self.fcs:
+            x = self.relu(fc(x))
+        mask_pred = self.fc_logits(x).view(
+            x.size(0), self.num_classes, *self.output_size)
+        return mask_pred
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/fcn_mask_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/fcn_mask_head.py
new file mode 100755
index 000000000..0cba3cda0
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/fcn_mask_head.py
@@ -0,0 +1,328 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import Conv2d, ConvModule, build_upsample_layer
+from mmcv.ops.carafe import CARAFEPack
+from mmcv.runner import auto_fp16, force_fp32
+from torch.nn.modules.utils import _pair
+
+from mmdet.core import mask_target
+from mmdet.models.builder import HEADS, build_loss
+
+BYTES_PER_FLOAT = 4
+# TODO: This memory limit may be too much or too little. It would be better to
+# determine it based on available resources.
+GPU_MEM_LIMIT = 1024**3  # 1 GB memory limit
+
+
+@HEADS.register_module()
+class FCNMaskHead(nn.Module):
+
+    def __init__(self,
+                 num_convs=4,
+                 roi_feat_size=14,
+                 in_channels=256,
+                 conv_kernel_size=3,
+                 conv_out_channels=256,
+                 num_classes=80,
+                 class_agnostic=False,
+                 upsample_cfg=dict(type='deconv', scale_factor=2),
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 loss_mask=dict(
+                     type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)):
+        super(FCNMaskHead, self).__init__()
+        self.upsample_cfg = upsample_cfg.copy()
+        if self.upsample_cfg['type'] not in [
+                None, 'deconv', 'nearest', 'bilinear', 'carafe'
+        ]:
+            raise ValueError(
+                f'Invalid upsample method {self.upsample_cfg["type"]}, '
+                'accepted methods are "deconv", "nearest", "bilinear", '
+                '"carafe"')
+        self.num_convs = num_convs
+        # WARN: roi_feat_size is reserved and not used
+        self.roi_feat_size = _pair(roi_feat_size)
+        self.in_channels = in_channels
+        self.conv_kernel_size = conv_kernel_size
+        self.conv_out_channels = conv_out_channels
+        self.upsample_method = self.upsample_cfg.get('type')
+        self.scale_factor = self.upsample_cfg.pop('scale_factor', None)
+        self.num_classes = num_classes
+        self.class_agnostic = class_agnostic
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.fp16_enabled = False
+        self.loss_mask = build_loss(loss_mask)
+
+        self.convs = nn.ModuleList()
+        for i in range(self.num_convs):
+            in_channels = (
+                self.in_channels if i == 0 else self.conv_out_channels)
+            padding = (self.conv_kernel_size - 1) // 2
+            self.convs.append(
+                ConvModule(
+                    in_channels,
+                    self.conv_out_channels,
+                    self.conv_kernel_size,
+                    padding=padding,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg))
+        upsample_in_channels = (
+            self.conv_out_channels if self.num_convs > 0 else in_channels)
+        upsample_cfg_ = self.upsample_cfg.copy()
+        if self.upsample_method is None:
+            self.upsample = None
+        elif self.upsample_method == 'deconv':
+            upsample_cfg_.update(
+                in_channels=upsample_in_channels,
+                out_channels=self.conv_out_channels,
+                kernel_size=self.scale_factor,
+                stride=self.scale_factor)
+            self.upsample = build_upsample_layer(upsample_cfg_)
+        elif self.upsample_method == 'carafe':
+            upsample_cfg_.update(
+                channels=upsample_in_channels, scale_factor=self.scale_factor)
+            self.upsample = build_upsample_layer(upsample_cfg_)
+        else:
+            # suppress warnings
+            align_corners = (None
+                             if self.upsample_method == 'nearest' else False)
+            upsample_cfg_.update(
+                scale_factor=self.scale_factor,
+                mode=self.upsample_method,
+                align_corners=align_corners)
+            self.upsample = build_upsample_layer(upsample_cfg_)
+
+        out_channels = 1 if self.class_agnostic else self.num_classes
+        logits_in_channel = (
+            self.conv_out_channels
+            if self.upsample_method == 'deconv' else upsample_in_channels)
+        self.conv_logits = Conv2d(logits_in_channel, out_channels, 1)
+        self.relu = nn.ReLU(inplace=True)
+        self.debug_imgs = None
+
+    def init_weights(self):
+        for m in [self.upsample, self.conv_logits]:
+            if m is None:
+                continue
+            elif isinstance(m, CARAFEPack):
+                m.init_weights()
+            else:
+                nn.init.kaiming_normal_(
+                    m.weight, mode='fan_out', nonlinearity='relu')
+                nn.init.constant_(m.bias, 0)
+
+    @auto_fp16()
+    def forward(self, x):
+        for conv in self.convs:
+            x = conv(x)
+        if self.upsample is not None:
+            x = self.upsample(x)
+            if self.upsample_method == 'deconv':
+                x = self.relu(x)
+        mask_pred = self.conv_logits(x)
+        return mask_pred
+
+    def get_targets(self, sampling_results, gt_masks, rcnn_train_cfg):
+        pos_proposals = [res.pos_bboxes for res in sampling_results]
+        pos_assigned_gt_inds = [
+            res.pos_assigned_gt_inds for res in sampling_results
+        ]
+        mask_targets = mask_target(pos_proposals, pos_assigned_gt_inds,
+                                   gt_masks, rcnn_train_cfg)
+        return mask_targets
+
+    @force_fp32(apply_to=('mask_pred', ))
+    def loss(self, mask_pred, mask_targets, labels):
+        loss = dict()
+        if mask_pred.size(0) == 0:
+            loss_mask = mask_pred.sum()
+        else:
+            if self.class_agnostic:
+                loss_mask = self.loss_mask(mask_pred, mask_targets,
+                                           torch.zeros_like(labels))
+            else:
+                loss_mask = self.loss_mask(mask_pred, mask_targets, labels)
+        loss['loss_mask'] = loss_mask
+        return loss
+
+    def get_seg_masks(self, mask_pred, det_bboxes, det_labels, rcnn_test_cfg,
+                      ori_shape, scale_factor, rescale):
+        """Get segmentation masks from mask_pred and bboxes.
+
+        Args:
+            mask_pred (Tensor or ndarray): shape (n, #class, h, w).
+                For single-scale testing, mask_pred is the direct output of
+                model, whose type is Tensor, while for multi-scale testing,
+                it will be converted to numpy array outside of this method.
+            det_bboxes (Tensor): shape (n, 4/5)
+            det_labels (Tensor): shape (n, )
+            img_shape (Tensor): shape (3, )
+            rcnn_test_cfg (dict): rcnn testing config
+            ori_shape: original image size
+
+        Returns:
+            list[list]: encoded masks
+        """
+        if isinstance(mask_pred, torch.Tensor):
+            mask_pred = mask_pred.sigmoid()
+        else:
+            mask_pred = det_bboxes.new_tensor(mask_pred)
+
+        device = mask_pred.device
+        cls_segms = [[] for _ in range(self.num_classes)
+                     ]  # BG is not included in num_classes
+        bboxes = det_bboxes[:, :4]
+        labels = det_labels
+
+        if rescale:
+            img_h, img_w = ori_shape[:2]
+        else:
+            if isinstance(scale_factor, float):
+                img_h = np.round(ori_shape[0] * scale_factor).astype(np.int32)
+                img_w = np.round(ori_shape[1] * scale_factor).astype(np.int32)
+            else:
+                w_scale, h_scale = scale_factor[0], scale_factor[1]
+                img_h = np.round(ori_shape[0] * h_scale.item()).astype(
+                    np.int32)
+                img_w = np.round(ori_shape[1] * w_scale.item()).astype(
+                    np.int32)
+            scale_factor = 1.0
+
+        if not isinstance(scale_factor, (float, torch.Tensor)):
+            scale_factor = bboxes.new_tensor(scale_factor)
+        bboxes = bboxes / scale_factor
+
+        if torch.onnx.is_in_onnx_export():
+            # TODO: Remove after F.grid_sample is supported.
+            from torchvision.models.detection.roi_heads \
+                import paste_masks_in_image
+            masks = paste_masks_in_image(mask_pred, bboxes, ori_shape[:2])
+            thr = rcnn_test_cfg.get('mask_thr_binary', 0)
+            if thr > 0:
+                masks = masks >= thr
+            return masks
+
+        N = len(mask_pred)
+        # The actual implementation split the input into chunks,
+        # and paste them chunk by chunk.
+        if device.type == 'cpu':
+            # CPU is most efficient when they are pasted one by one with
+            # skip_empty=True, so that it performs minimal number of
+            # operations.
+            num_chunks = N
+        else:
+            # GPU benefits from parallelism for larger chunks,
+            # but may have memory issue
+            num_chunks = int(
+                np.ceil(N * img_h * img_w * BYTES_PER_FLOAT / GPU_MEM_LIMIT))
+            assert (num_chunks <=
+                    N), 'Default GPU_MEM_LIMIT is too small; try increasing it'
+        chunks = torch.chunk(torch.arange(N, device=device), num_chunks)
+
+        threshold = rcnn_test_cfg.mask_thr_binary
+        im_mask = torch.zeros(
+            N,
+            img_h,
+            img_w,
+            device=device,
+            dtype=torch.bool if threshold >= 0 else torch.uint8)
+
+        if not self.class_agnostic:
+            mask_pred = mask_pred[range(N), labels][:, None]
+
+        for inds in chunks:
+            masks_chunk, spatial_inds = _do_paste_mask(
+                mask_pred[inds],
+                bboxes[inds],
+                img_h,
+                img_w,
+                skip_empty=device.type == 'cpu')
+
+            if threshold >= 0:
+                masks_chunk = (masks_chunk >= threshold).to(dtype=torch.bool)
+            else:
+                # for visualization and debugging
+                masks_chunk = (masks_chunk * 255).to(dtype=torch.uint8)
+
+            im_mask[(inds, ) + spatial_inds] = masks_chunk
+
+        for i in range(N):
+            cls_segms[labels[i]].append(im_mask[i].detach().cpu().numpy())
+        return cls_segms
+
+
+def _do_paste_mask(masks, boxes, img_h, img_w, skip_empty=True):
+    """Paste instance masks acoording to boxes.
+
+    This implementation is modified from
+    https://github.com/facebookresearch/detectron2/
+
+    Args:
+        masks (Tensor): N, 1, H, W
+        boxes (Tensor): N, 4
+        img_h (int): Height of the image to be pasted.
+        img_w (int): Width of the image to be pasted.
+        skip_empty (bool): Only paste masks within the region that
+            tightly bound all boxes, and returns the results this region only.
+            An important optimization for CPU.
+
+    Returns:
+        tuple: (Tensor, tuple). The first item is mask tensor, the second one
+            is the slice object.
+        If skip_empty == False, the whole image will be pasted. It will
+            return a mask of shape (N, img_h, img_w) and an empty tuple.
+        If skip_empty == True, only area around the mask will be pasted.
+            A mask of shape (N, h', w') and its start and end coordinates
+            in the original image will be returned.
+    """
+    # On GPU, paste all masks together (up to chunk size)
+    # by using the entire image to sample the masks
+    # Compared to pasting them one by one,
+    # this has more operations but is faster on COCO-scale dataset.
+    device = masks.device
+    if skip_empty:
+        x0_int, y0_int = torch.clamp(
+            boxes.min(dim=0).values.floor()[:2] - 1,
+            min=0).to(dtype=torch.int32)
+        x1_int = torch.clamp(
+            boxes[:, 2].max().ceil() + 1, max=img_w).to(dtype=torch.int32)
+        y1_int = torch.clamp(
+            boxes[:, 3].max().ceil() + 1, max=img_h).to(dtype=torch.int32)
+    else:
+        x0_int, y0_int = 0, 0
+        x1_int, y1_int = img_w, img_h
+    x0, y0, x1, y1 = torch.split(boxes, 1, dim=1)  # each is Nx1
+
+    N = masks.shape[0]
+
+    img_y = torch.arange(
+        y0_int, y1_int, device=device, dtype=torch.float32) + 0.5
+    img_x = torch.arange(
+        x0_int, x1_int, device=device, dtype=torch.float32) + 0.5
+    img_y = (img_y - y0) / (y1 - y0) * 2 - 1
+    img_x = (img_x - x0) / (x1 - x0) * 2 - 1
+    # img_x, img_y have shapes (N, w), (N, h)
+    if torch.isinf(img_x).any():
+        inds = torch.where(torch.isinf(img_x))
+        img_x[inds] = 0
+    if torch.isinf(img_y).any():
+        inds = torch.where(torch.isinf(img_y))
+        img_y[inds] = 0
+
+    gx = img_x[:, None, :].expand(N, img_y.size(1), img_x.size(1))
+    gy = img_y[:, :, None].expand(N, img_y.size(1), img_x.size(1))
+    grid = torch.stack([gx, gy], dim=3)
+
+    if torch.onnx.is_in_onnx_export():
+        raise RuntimeError(
+            'Exporting F.grid_sample from Pytorch to ONNX is not supported.')
+    img_masks = F.grid_sample(
+        masks.to(dtype=torch.float32), grid, align_corners=False)
+
+    if skip_empty:
+        return img_masks[:, 0], (slice(y0_int, y1_int), slice(x0_int, x1_int))
+    else:
+        return img_masks[:, 0], ()
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/fused_semantic_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/fused_semantic_head.py
new file mode 100755
index 000000000..2aa6033ee
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/fused_semantic_head.py
@@ -0,0 +1,107 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, kaiming_init
+from mmcv.runner import auto_fp16, force_fp32
+
+from mmdet.models.builder import HEADS
+
+
+@HEADS.register_module()
+class FusedSemanticHead(nn.Module):
+    r"""Multi-level fused semantic segmentation head.
+
+    .. code-block:: none
+
+        in_1 -> 1x1 conv ---
+                            |
+        in_2 -> 1x1 conv -- |
+                           ||
+        in_3 -> 1x1 conv - ||
+                          |||                  /-> 1x1 conv (mask prediction)
+        in_4 -> 1x1 conv -----> 3x3 convs (*4)
+                            |                  \-> 1x1 conv (feature)
+        in_5 -> 1x1 conv ---
+    """  # noqa: W605
+
+    def __init__(self,
+                 num_ins,
+                 fusion_level,
+                 num_convs=4,
+                 in_channels=256,
+                 conv_out_channels=256,
+                 num_classes=183,
+                 ignore_label=255,
+                 loss_weight=0.2,
+                 conv_cfg=None,
+                 norm_cfg=None):
+        super(FusedSemanticHead, self).__init__()
+        self.num_ins = num_ins
+        self.fusion_level = fusion_level
+        self.num_convs = num_convs
+        self.in_channels = in_channels
+        self.conv_out_channels = conv_out_channels
+        self.num_classes = num_classes
+        self.ignore_label = ignore_label
+        self.loss_weight = loss_weight
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.fp16_enabled = False
+
+        self.lateral_convs = nn.ModuleList()
+        for i in range(self.num_ins):
+            self.lateral_convs.append(
+                ConvModule(
+                    self.in_channels,
+                    self.in_channels,
+                    1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg,
+                    inplace=False))
+
+        self.convs = nn.ModuleList()
+        for i in range(self.num_convs):
+            in_channels = self.in_channels if i == 0 else conv_out_channels
+            self.convs.append(
+                ConvModule(
+                    in_channels,
+                    conv_out_channels,
+                    3,
+                    padding=1,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg))
+        self.conv_embedding = ConvModule(
+            conv_out_channels,
+            conv_out_channels,
+            1,
+            conv_cfg=self.conv_cfg,
+            norm_cfg=self.norm_cfg)
+        self.conv_logits = nn.Conv2d(conv_out_channels, self.num_classes, 1)
+
+        self.criterion = nn.CrossEntropyLoss(ignore_index=ignore_label)
+
+    def init_weights(self):
+        kaiming_init(self.conv_logits)
+
+    @auto_fp16()
+    def forward(self, feats):
+        x = self.lateral_convs[self.fusion_level](feats[self.fusion_level])
+        fused_size = tuple(x.shape[-2:])
+        for i, feat in enumerate(feats):
+            if i != self.fusion_level:
+                feat = F.interpolate(
+                    feat, size=fused_size, mode='bilinear', align_corners=True)
+                x += self.lateral_convs[i](feat)
+
+        for i in range(self.num_convs):
+            x = self.convs[i](x)
+
+        mask_pred = self.conv_logits(x)
+        x = self.conv_embedding(x)
+        return mask_pred, x
+
+    @force_fp32(apply_to=('mask_pred', ))
+    def loss(self, mask_pred, labels):
+        labels = labels.squeeze(1).long()
+        loss_semantic_seg = self.criterion(mask_pred, labels)
+        loss_semantic_seg *= self.loss_weight
+        return loss_semantic_seg
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/grid_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/grid_head.py
new file mode 100755
index 000000000..83058cbdd
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/grid_head.py
@@ -0,0 +1,359 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import ConvModule, kaiming_init, normal_init
+
+from mmdet.models.builder import HEADS, build_loss
+
+
+@HEADS.register_module()
+class GridHead(nn.Module):
+
+    def __init__(self,
+                 grid_points=9,
+                 num_convs=8,
+                 roi_feat_size=14,
+                 in_channels=256,
+                 conv_kernel_size=3,
+                 point_feat_channels=64,
+                 deconv_kernel_size=4,
+                 class_agnostic=False,
+                 loss_grid=dict(
+                     type='CrossEntropyLoss', use_sigmoid=True,
+                     loss_weight=15),
+                 conv_cfg=None,
+                 norm_cfg=dict(type='GN', num_groups=36)):
+        super(GridHead, self).__init__()
+        self.grid_points = grid_points
+        self.num_convs = num_convs
+        self.roi_feat_size = roi_feat_size
+        self.in_channels = in_channels
+        self.conv_kernel_size = conv_kernel_size
+        self.point_feat_channels = point_feat_channels
+        self.conv_out_channels = self.point_feat_channels * self.grid_points
+        self.class_agnostic = class_agnostic
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        if isinstance(norm_cfg, dict) and norm_cfg['type'] == 'GN':
+            assert self.conv_out_channels % norm_cfg['num_groups'] == 0
+
+        assert self.grid_points >= 4
+        self.grid_size = int(np.sqrt(self.grid_points))
+        if self.grid_size * self.grid_size != self.grid_points:
+            raise ValueError('grid_points must be a square number')
+
+        # the predicted heatmap is half of whole_map_size
+        if not isinstance(self.roi_feat_size, int):
+            raise ValueError('Only square RoIs are supporeted in Grid R-CNN')
+        self.whole_map_size = self.roi_feat_size * 4
+
+        # compute point-wise sub-regions
+        self.sub_regions = self.calc_sub_regions()
+
+        self.convs = []
+        for i in range(self.num_convs):
+            in_channels = (
+                self.in_channels if i == 0 else self.conv_out_channels)
+            stride = 2 if i == 0 else 1
+            padding = (self.conv_kernel_size - 1) // 2
+            self.convs.append(
+                ConvModule(
+                    in_channels,
+                    self.conv_out_channels,
+                    self.conv_kernel_size,
+                    stride=stride,
+                    padding=padding,
+                    conv_cfg=self.conv_cfg,
+                    norm_cfg=self.norm_cfg,
+                    bias=True))
+        self.convs = nn.Sequential(*self.convs)
+
+        self.deconv1 = nn.ConvTranspose2d(
+            self.conv_out_channels,
+            self.conv_out_channels,
+            kernel_size=deconv_kernel_size,
+            stride=2,
+            padding=(deconv_kernel_size - 2) // 2,
+            groups=grid_points)
+        self.norm1 = nn.GroupNorm(grid_points, self.conv_out_channels)
+        self.deconv2 = nn.ConvTranspose2d(
+            self.conv_out_channels,
+            grid_points,
+            kernel_size=deconv_kernel_size,
+            stride=2,
+            padding=(deconv_kernel_size - 2) // 2,
+            groups=grid_points)
+
+        # find the 4-neighbor of each grid point
+        self.neighbor_points = []
+        grid_size = self.grid_size
+        for i in range(grid_size):  # i-th column
+            for j in range(grid_size):  # j-th row
+                neighbors = []
+                if i > 0:  # left: (i - 1, j)
+                    neighbors.append((i - 1) * grid_size + j)
+                if j > 0:  # up: (i, j - 1)
+                    neighbors.append(i * grid_size + j - 1)
+                if j < grid_size - 1:  # down: (i, j + 1)
+                    neighbors.append(i * grid_size + j + 1)
+                if i < grid_size - 1:  # right: (i + 1, j)
+                    neighbors.append((i + 1) * grid_size + j)
+                self.neighbor_points.append(tuple(neighbors))
+        # total edges in the grid
+        self.num_edges = sum([len(p) for p in self.neighbor_points])
+
+        self.forder_trans = nn.ModuleList()  # first-order feature transition
+        self.sorder_trans = nn.ModuleList()  # second-order feature transition
+        for neighbors in self.neighbor_points:
+            fo_trans = nn.ModuleList()
+            so_trans = nn.ModuleList()
+            for _ in range(len(neighbors)):
+                # each transition module consists of a 5x5 depth-wise conv and
+                # 1x1 conv.
+                fo_trans.append(
+                    nn.Sequential(
+                        nn.Conv2d(
+                            self.point_feat_channels,
+                            self.point_feat_channels,
+                            5,
+                            stride=1,
+                            padding=2,
+                            groups=self.point_feat_channels),
+                        nn.Conv2d(self.point_feat_channels,
+                                  self.point_feat_channels, 1)))
+                so_trans.append(
+                    nn.Sequential(
+                        nn.Conv2d(
+                            self.point_feat_channels,
+                            self.point_feat_channels,
+                            5,
+                            1,
+                            2,
+                            groups=self.point_feat_channels),
+                        nn.Conv2d(self.point_feat_channels,
+                                  self.point_feat_channels, 1)))
+            self.forder_trans.append(fo_trans)
+            self.sorder_trans.append(so_trans)
+
+        self.loss_grid = build_loss(loss_grid)
+
+    def init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
+                # TODO: compare mode = "fan_in" or "fan_out"
+                kaiming_init(m)
+        for m in self.modules():
+            if isinstance(m, nn.ConvTranspose2d):
+                normal_init(m, std=0.001)
+        nn.init.constant_(self.deconv2.bias, -np.log(0.99 / 0.01))
+
+    def forward(self, x):
+        assert x.shape[-1] == x.shape[-2] == self.roi_feat_size
+        # RoI feature transformation, downsample 2x
+        x = self.convs(x)
+
+        c = self.point_feat_channels
+        # first-order fusion
+        x_fo = [None for _ in range(self.grid_points)]
+        for i, points in enumerate(self.neighbor_points):
+            x_fo[i] = x[:, i * c:(i + 1) * c]
+            for j, point_idx in enumerate(points):
+                x_fo[i] = x_fo[i] + self.forder_trans[i][j](
+                    x[:, point_idx * c:(point_idx + 1) * c])
+
+        # second-order fusion
+        x_so = [None for _ in range(self.grid_points)]
+        for i, points in enumerate(self.neighbor_points):
+            x_so[i] = x[:, i * c:(i + 1) * c]
+            for j, point_idx in enumerate(points):
+                x_so[i] = x_so[i] + self.sorder_trans[i][j](x_fo[point_idx])
+
+        # predicted heatmap with fused features
+        x2 = torch.cat(x_so, dim=1)
+        x2 = self.deconv1(x2)
+        x2 = F.relu(self.norm1(x2), inplace=True)
+        heatmap = self.deconv2(x2)
+
+        # predicted heatmap with original features (applicable during training)
+        if self.training:
+            x1 = x
+            x1 = self.deconv1(x1)
+            x1 = F.relu(self.norm1(x1), inplace=True)
+            heatmap_unfused = self.deconv2(x1)
+        else:
+            heatmap_unfused = heatmap
+
+        return dict(fused=heatmap, unfused=heatmap_unfused)
+
+    def calc_sub_regions(self):
+        """Compute point specific representation regions.
+
+        See Grid R-CNN Plus (https://arxiv.org/abs/1906.05688) for details.
+        """
+        # to make it consistent with the original implementation, half_size
+        # is computed as 2 * quarter_size, which is smaller
+        half_size = self.whole_map_size // 4 * 2
+        sub_regions = []
+        for i in range(self.grid_points):
+            x_idx = i // self.grid_size
+            y_idx = i % self.grid_size
+            if x_idx == 0:
+                sub_x1 = 0
+            elif x_idx == self.grid_size - 1:
+                sub_x1 = half_size
+            else:
+                ratio = x_idx / (self.grid_size - 1) - 0.25
+                sub_x1 = max(int(ratio * self.whole_map_size), 0)
+
+            if y_idx == 0:
+                sub_y1 = 0
+            elif y_idx == self.grid_size - 1:
+                sub_y1 = half_size
+            else:
+                ratio = y_idx / (self.grid_size - 1) - 0.25
+                sub_y1 = max(int(ratio * self.whole_map_size), 0)
+            sub_regions.append(
+                (sub_x1, sub_y1, sub_x1 + half_size, sub_y1 + half_size))
+        return sub_regions
+
+    def get_targets(self, sampling_results, rcnn_train_cfg):
+        # mix all samples (across images) together.
+        pos_bboxes = torch.cat([res.pos_bboxes for res in sampling_results],
+                               dim=0).cpu()
+        pos_gt_bboxes = torch.cat(
+            [res.pos_gt_bboxes for res in sampling_results], dim=0).cpu()
+        assert pos_bboxes.shape == pos_gt_bboxes.shape
+
+        # expand pos_bboxes to 2x of original size
+        x1 = pos_bboxes[:, 0] - (pos_bboxes[:, 2] - pos_bboxes[:, 0]) / 2
+        y1 = pos_bboxes[:, 1] - (pos_bboxes[:, 3] - pos_bboxes[:, 1]) / 2
+        x2 = pos_bboxes[:, 2] + (pos_bboxes[:, 2] - pos_bboxes[:, 0]) / 2
+        y2 = pos_bboxes[:, 3] + (pos_bboxes[:, 3] - pos_bboxes[:, 1]) / 2
+        pos_bboxes = torch.stack([x1, y1, x2, y2], dim=-1)
+        pos_bbox_ws = (pos_bboxes[:, 2] - pos_bboxes[:, 0]).unsqueeze(-1)
+        pos_bbox_hs = (pos_bboxes[:, 3] - pos_bboxes[:, 1]).unsqueeze(-1)
+
+        num_rois = pos_bboxes.shape[0]
+        map_size = self.whole_map_size
+        # this is not the final target shape
+        targets = torch.zeros((num_rois, self.grid_points, map_size, map_size),
+                              dtype=torch.float)
+
+        # pre-compute interpolation factors for all grid points.
+        # the first item is the factor of x-dim, and the second is y-dim.
+        # for a 9-point grid, factors are like (1, 0), (0.5, 0.5), (0, 1)
+        factors = []
+        for j in range(self.grid_points):
+            x_idx = j // self.grid_size
+            y_idx = j % self.grid_size
+            factors.append((1 - x_idx / (self.grid_size - 1),
+                            1 - y_idx / (self.grid_size - 1)))
+
+        radius = rcnn_train_cfg.pos_radius
+        radius2 = radius**2
+        for i in range(num_rois):
+            # ignore small bboxes
+            if (pos_bbox_ws[i] <= self.grid_size
+                    or pos_bbox_hs[i] <= self.grid_size):
+                continue
+            # for each grid point, mark a small circle as positive
+            for j in range(self.grid_points):
+                factor_x, factor_y = factors[j]
+                gridpoint_x = factor_x * pos_gt_bboxes[i, 0] + (
+                    1 - factor_x) * pos_gt_bboxes[i, 2]
+                gridpoint_y = factor_y * pos_gt_bboxes[i, 1] + (
+                    1 - factor_y) * pos_gt_bboxes[i, 3]
+
+                cx = int((gridpoint_x - pos_bboxes[i, 0]) / pos_bbox_ws[i] *
+                         map_size)
+                cy = int((gridpoint_y - pos_bboxes[i, 1]) / pos_bbox_hs[i] *
+                         map_size)
+
+                for x in range(cx - radius, cx + radius + 1):
+                    for y in range(cy - radius, cy + radius + 1):
+                        if x >= 0 and x < map_size and y >= 0 and y < map_size:
+                            if (x - cx)**2 + (y - cy)**2 <= radius2:
+                                targets[i, j, y, x] = 1
+        # reduce the target heatmap size by a half
+        # proposed in Grid R-CNN Plus (https://arxiv.org/abs/1906.05688).
+        sub_targets = []
+        for i in range(self.grid_points):
+            sub_x1, sub_y1, sub_x2, sub_y2 = self.sub_regions[i]
+            sub_targets.append(targets[:, [i], sub_y1:sub_y2, sub_x1:sub_x2])
+        sub_targets = torch.cat(sub_targets, dim=1)
+        sub_targets = sub_targets.to(sampling_results[0].pos_bboxes.device)
+        return sub_targets
+
+    def loss(self, grid_pred, grid_targets):
+        loss_fused = self.loss_grid(grid_pred['fused'], grid_targets)
+        loss_unfused = self.loss_grid(grid_pred['unfused'], grid_targets)
+        loss_grid = loss_fused + loss_unfused
+        return dict(loss_grid=loss_grid)
+
+    def get_bboxes(self, det_bboxes, grid_pred, img_metas):
+        # TODO: refactoring
+        assert det_bboxes.shape[0] == grid_pred.shape[0]
+        det_bboxes = det_bboxes.cpu()
+        cls_scores = det_bboxes[:, [4]]
+        det_bboxes = det_bboxes[:, :4]
+        grid_pred = grid_pred.sigmoid().cpu()
+
+        R, c, h, w = grid_pred.shape
+        half_size = self.whole_map_size // 4 * 2
+        assert h == w == half_size
+        assert c == self.grid_points
+
+        # find the point with max scores in the half-sized heatmap
+        grid_pred = grid_pred.view(R * c, h * w)
+        pred_scores, pred_position = grid_pred.max(dim=1)
+        xs = pred_position % w
+        ys = pred_position // w
+
+        # get the position in the whole heatmap instead of half-sized heatmap
+        for i in range(self.grid_points):
+            xs[i::self.grid_points] += self.sub_regions[i][0]
+            ys[i::self.grid_points] += self.sub_regions[i][1]
+
+        # reshape to (num_rois, grid_points)
+        pred_scores, xs, ys = tuple(
+            map(lambda x: x.view(R, c), [pred_scores, xs, ys]))
+
+        # get expanded pos_bboxes
+        widths = (det_bboxes[:, 2] - det_bboxes[:, 0]).unsqueeze(-1)
+        heights = (det_bboxes[:, 3] - det_bboxes[:, 1]).unsqueeze(-1)
+        x1 = (det_bboxes[:, 0, None] - widths / 2)
+        y1 = (det_bboxes[:, 1, None] - heights / 2)
+        # map the grid point to the absolute coordinates
+        abs_xs = (xs.float() + 0.5) / w * widths + x1
+        abs_ys = (ys.float() + 0.5) / h * heights + y1
+
+        # get the grid points indices that fall on the bbox boundaries
+        x1_inds = [i for i in range(self.grid_size)]
+        y1_inds = [i * self.grid_size for i in range(self.grid_size)]
+        x2_inds = [
+            self.grid_points - self.grid_size + i
+            for i in range(self.grid_size)
+        ]
+        y2_inds = [(i + 1) * self.grid_size - 1 for i in range(self.grid_size)]
+
+        # voting of all grid points on some boundary
+        bboxes_x1 = (abs_xs[:, x1_inds] * pred_scores[:, x1_inds]).sum(
+            dim=1, keepdim=True) / (
+                pred_scores[:, x1_inds].sum(dim=1, keepdim=True))
+        bboxes_y1 = (abs_ys[:, y1_inds] * pred_scores[:, y1_inds]).sum(
+            dim=1, keepdim=True) / (
+                pred_scores[:, y1_inds].sum(dim=1, keepdim=True))
+        bboxes_x2 = (abs_xs[:, x2_inds] * pred_scores[:, x2_inds]).sum(
+            dim=1, keepdim=True) / (
+                pred_scores[:, x2_inds].sum(dim=1, keepdim=True))
+        bboxes_y2 = (abs_ys[:, y2_inds] * pred_scores[:, y2_inds]).sum(
+            dim=1, keepdim=True) / (
+                pred_scores[:, y2_inds].sum(dim=1, keepdim=True))
+
+        bbox_res = torch.cat(
+            [bboxes_x1, bboxes_y1, bboxes_x2, bboxes_y2, cls_scores], dim=1)
+        bbox_res[:, [0, 2]].clamp_(min=0, max=img_metas[0]['img_shape'][1])
+        bbox_res[:, [1, 3]].clamp_(min=0, max=img_metas[0]['img_shape'][0])
+
+        return bbox_res
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/htc_mask_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/htc_mask_head.py
new file mode 100755
index 000000000..330b778eb
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/htc_mask_head.py
@@ -0,0 +1,43 @@
+from mmcv.cnn import ConvModule
+
+from mmdet.models.builder import HEADS
+from .fcn_mask_head import FCNMaskHead
+
+
+@HEADS.register_module()
+class HTCMaskHead(FCNMaskHead):
+
+    def __init__(self, with_conv_res=True, *args, **kwargs):
+        super(HTCMaskHead, self).__init__(*args, **kwargs)
+        self.with_conv_res = with_conv_res
+        if self.with_conv_res:
+            self.conv_res = ConvModule(
+                self.conv_out_channels,
+                self.conv_out_channels,
+                1,
+                conv_cfg=self.conv_cfg,
+                norm_cfg=self.norm_cfg)
+
+    def init_weights(self):
+        super(HTCMaskHead, self).init_weights()
+        if self.with_conv_res:
+            self.conv_res.init_weights()
+
+    def forward(self, x, res_feat=None, return_logits=True, return_feat=True):
+        if res_feat is not None:
+            assert self.with_conv_res
+            res_feat = self.conv_res(res_feat)
+            x = x + res_feat
+        for conv in self.convs:
+            x = conv(x)
+        res_feat = x
+        outs = []
+        if return_logits:
+            x = self.upsample(x)
+            if self.upsample_method == 'deconv':
+                x = self.relu(x)
+            mask_pred = self.conv_logits(x)
+            outs.append(mask_pred)
+        if return_feat:
+            outs.append(res_feat)
+        return outs if len(outs) > 1 else outs[0]
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/mask_point_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/mask_point_head.py
new file mode 100755
index 000000000..f38a5c9d7
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/mask_point_head.py
@@ -0,0 +1,300 @@
+# Modified from https://github.com/facebookresearch/detectron2/tree/master/projects/PointRend/point_head/point_head.py  # noqa
+
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, normal_init
+from mmcv.ops import point_sample, rel_roi_point_to_rel_img_point
+
+from mmdet.models.builder import HEADS, build_loss
+
+
+@HEADS.register_module()
+class MaskPointHead(nn.Module):
+    """A mask point head use in PointRend.
+
+    ``MaskPointHead`` use shared multi-layer perceptron (equivalent to
+    nn.Conv1d) to predict the logit of input points. The fine-grained feature
+    and coarse feature will be concatenate together for predication.
+
+    Args:
+        num_fcs (int): Number of fc layers in the head. Default: 3.
+        in_channels (int): Number of input channels. Default: 256.
+        fc_channels (int): Number of fc channels. Default: 256.
+        num_classes (int): Number of classes for logits. Default: 80.
+        class_agnostic (bool): Whether use class agnostic classification.
+            If so, the output channels of logits will be 1. Default: False.
+        coarse_pred_each_layer (bool): Whether concatenate coarse feature with
+            the output of each fc layer. Default: True.
+        conv_cfg (dict | None): Dictionary to construct and config conv layer.
+            Default: dict(type='Conv1d'))
+        norm_cfg (dict | None): Dictionary to construct and config norm layer.
+            Default: None.
+        loss_point (dict): Dictionary to construct and config loss layer of
+            point head. Default: dict(type='CrossEntropyLoss', use_mask=True,
+            loss_weight=1.0).
+    """
+
+    def __init__(self,
+                 num_classes,
+                 num_fcs=3,
+                 in_channels=256,
+                 fc_channels=256,
+                 class_agnostic=False,
+                 coarse_pred_each_layer=True,
+                 conv_cfg=dict(type='Conv1d'),
+                 norm_cfg=None,
+                 act_cfg=dict(type='ReLU'),
+                 loss_point=dict(
+                     type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)):
+        super().__init__()
+        self.num_fcs = num_fcs
+        self.in_channels = in_channels
+        self.fc_channles = fc_channels
+        self.num_classes = num_classes
+        self.class_agnostic = class_agnostic
+        self.coarse_pred_each_layer = coarse_pred_each_layer
+        self.conv_cfg = conv_cfg
+        self.norm_cfg = norm_cfg
+        self.loss_point = build_loss(loss_point)
+
+        fc_in_channels = in_channels + num_classes
+        self.fcs = nn.ModuleList()
+        for _ in range(num_fcs):
+            fc = ConvModule(
+                fc_in_channels,
+                fc_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                act_cfg=act_cfg)
+            self.fcs.append(fc)
+            fc_in_channels = fc_channels
+            fc_in_channels += num_classes if self.coarse_pred_each_layer else 0
+
+        out_channels = 1 if self.class_agnostic else self.num_classes
+        self.fc_logits = nn.Conv1d(
+            fc_in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def init_weights(self):
+        """Initialize last classification layer of MaskPointHead, conv layers
+        are already initialized by ConvModule."""
+        normal_init(self.fc_logits, std=0.001)
+
+    def forward(self, fine_grained_feats, coarse_feats):
+        """Classify each point base on fine grained and coarse feats.
+
+        Args:
+            fine_grained_feats (Tensor): Fine grained feature sampled from FPN,
+                shape (num_rois, in_channels, num_points).
+            coarse_feats (Tensor): Coarse feature sampled from CoarseMaskHead,
+                shape (num_rois, num_classes, num_points).
+
+        Returns:
+            Tensor: Point classification results,
+                shape (num_rois, num_class, num_points).
+        """
+
+        x = torch.cat([fine_grained_feats, coarse_feats], dim=1)
+        for fc in self.fcs:
+            x = fc(x)
+            if self.coarse_pred_each_layer:
+                x = torch.cat((x, coarse_feats), dim=1)
+        return self.fc_logits(x)
+
+    def get_targets(self, rois, rel_roi_points, sampling_results, gt_masks,
+                    cfg):
+        """Get training targets of MaskPointHead for all images.
+
+        Args:
+            rois (Tensor): Region of Interest, shape (num_rois, 5).
+            rel_roi_points: Points coordinates relative to RoI, shape
+                (num_rois, num_points, 2).
+            sampling_results (:obj:`SamplingResult`): Sampling result after
+                sampling and assignment.
+            gt_masks (Tensor) : Ground truth segmentation masks of
+                corresponding boxes, shape (num_rois, height, width).
+            cfg (dict): Training cfg.
+
+        Returns:
+            Tensor: Point target, shape (num_rois, num_points).
+        """
+
+        num_imgs = len(sampling_results)
+        rois_list = []
+        rel_roi_points_list = []
+        for batch_ind in range(num_imgs):
+            inds = (rois[:, 0] == batch_ind)
+            rois_list.append(rois[inds])
+            rel_roi_points_list.append(rel_roi_points[inds])
+        pos_assigned_gt_inds_list = [
+            res.pos_assigned_gt_inds for res in sampling_results
+        ]
+        cfg_list = [cfg for _ in range(num_imgs)]
+
+        point_targets = map(self._get_target_single, rois_list,
+                            rel_roi_points_list, pos_assigned_gt_inds_list,
+                            gt_masks, cfg_list)
+        point_targets = list(point_targets)
+
+        if len(point_targets) > 0:
+            point_targets = torch.cat(point_targets)
+
+        return point_targets
+
+    def _get_target_single(self, rois, rel_roi_points, pos_assigned_gt_inds,
+                           gt_masks, cfg):
+        """Get training target of MaskPointHead for each image."""
+        num_pos = rois.size(0)
+        num_points = cfg.num_points
+        if num_pos > 0:
+            gt_masks_th = (
+                gt_masks.to_tensor(rois.dtype, rois.device).index_select(
+                    0, pos_assigned_gt_inds))
+            gt_masks_th = gt_masks_th.unsqueeze(1)
+            rel_img_points = rel_roi_point_to_rel_img_point(
+                rois, rel_roi_points, gt_masks_th.shape[2:])
+            point_targets = point_sample(gt_masks_th,
+                                         rel_img_points).squeeze(1)
+        else:
+            point_targets = rois.new_zeros((0, num_points))
+        return point_targets
+
+    def loss(self, point_pred, point_targets, labels):
+        """Calculate loss for MaskPointHead.
+
+        Args:
+            point_pred (Tensor): Point predication result, shape
+                (num_rois, num_classes, num_points).
+            point_targets (Tensor): Point targets, shape (num_roi, num_points).
+            labels (Tensor): Class label of corresponding boxes,
+                shape (num_rois, )
+
+        Returns:
+            dict[str, Tensor]: a dictionary of point loss components
+        """
+
+        loss = dict()
+        if self.class_agnostic:
+            loss_point = self.loss_point(point_pred, point_targets,
+                                         torch.zeros_like(labels))
+        else:
+            loss_point = self.loss_point(point_pred, point_targets, labels)
+        loss['loss_point'] = loss_point
+        return loss
+
+    def _get_uncertainty(self, mask_pred, labels):
+        """Estimate uncertainty based on pred logits.
+
+        We estimate uncertainty as L1 distance between 0.0 and the logits
+        prediction in 'mask_pred' for the foreground class in `classes`.
+
+        Args:
+            mask_pred (Tensor): mask predication logits, shape (num_rois,
+                num_classes, mask_height, mask_width).
+
+            labels (list[Tensor]): Either predicted or ground truth label for
+                each predicted mask, of length num_rois.
+
+        Returns:
+            scores (Tensor): Uncertainty scores with the most uncertain
+                locations having the highest uncertainty score,
+                shape (num_rois, 1, mask_height, mask_width)
+        """
+        if mask_pred.shape[1] == 1:
+            gt_class_logits = mask_pred.clone()
+        else:
+            inds = torch.arange(mask_pred.shape[0], device=mask_pred.device)
+            gt_class_logits = mask_pred[inds, labels].unsqueeze(1)
+        return -torch.abs(gt_class_logits)
+
+    def get_roi_rel_points_train(self, mask_pred, labels, cfg):
+        """Get ``num_points`` most uncertain points with random points during
+        train.
+
+        Sample points in [0, 1] x [0, 1] coordinate space based on their
+        uncertainty. The uncertainties are calculated for each point using
+        '_get_uncertainty()' function that takes point's logit prediction as
+        input.
+
+        Args:
+            mask_pred (Tensor): A tensor of shape (num_rois, num_classes,
+                mask_height, mask_width) for class-specific or class-agnostic
+                prediction.
+            labels (list): The ground truth class for each instance.
+            cfg (dict): Training config of point head.
+
+        Returns:
+            point_coords (Tensor): A tensor of shape (num_rois, num_points, 2)
+                that contains the coordinates sampled points.
+        """
+        num_points = cfg.num_points
+        oversample_ratio = cfg.oversample_ratio
+        importance_sample_ratio = cfg.importance_sample_ratio
+        assert oversample_ratio >= 1
+        assert 0 <= importance_sample_ratio <= 1
+        batch_size = mask_pred.shape[0]
+        num_sampled = int(num_points * oversample_ratio)
+        point_coords = torch.rand(
+            batch_size, num_sampled, 2, device=mask_pred.device)
+        point_logits = point_sample(mask_pred, point_coords)
+        # It is crucial to calculate uncertainty based on the sampled
+        # prediction value for the points. Calculating uncertainties of the
+        # coarse predictions first and sampling them for points leads to
+        # incorrect results.  To illustrate this: assume uncertainty func(
+        # logits)=-abs(logits), a sampled point between two coarse
+        # predictions with -1 and 1 logits has 0 logits, and therefore 0
+        # uncertainty value. However, if we calculate uncertainties for the
+        # coarse predictions first, both will have -1 uncertainty,
+        # and sampled point will get -1 uncertainty.
+        point_uncertainties = self._get_uncertainty(point_logits, labels)
+        num_uncertain_points = int(importance_sample_ratio * num_points)
+        num_random_points = num_points - num_uncertain_points
+        idx = torch.topk(
+            point_uncertainties[:, 0, :], k=num_uncertain_points, dim=1)[1]
+        shift = num_sampled * torch.arange(
+            batch_size, dtype=torch.long, device=mask_pred.device)
+        idx += shift[:, None]
+        point_coords = point_coords.view(-1, 2)[idx.view(-1), :].view(
+            batch_size, num_uncertain_points, 2)
+        if num_random_points > 0:
+            rand_roi_coords = torch.rand(
+                batch_size, num_random_points, 2, device=mask_pred.device)
+            point_coords = torch.cat((point_coords, rand_roi_coords), dim=1)
+        return point_coords
+
+    def get_roi_rel_points_test(self, mask_pred, pred_label, cfg):
+        """Get ``num_points`` most uncertain points during test.
+
+        Args:
+            mask_pred (Tensor): A tensor of shape (num_rois, num_classes,
+                mask_height, mask_width) for class-specific or class-agnostic
+                prediction.
+            pred_label (list): The predication class for each instance.
+            cfg (dict): Testing config of point head.
+
+        Returns:
+            point_indices (Tensor): A tensor of shape (num_rois, num_points)
+                that contains indices from [0, mask_height x mask_width) of the
+                most uncertain points.
+            point_coords (Tensor): A tensor of shape (num_rois, num_points, 2)
+                that contains [0, 1] x [0, 1] normalized coordinates of the
+                most uncertain points from the [mask_height, mask_width] grid .
+        """
+        num_points = cfg.subdivision_num_points
+        uncertainty_map = self._get_uncertainty(mask_pred, pred_label)
+        num_rois, _, mask_height, mask_width = uncertainty_map.shape
+        h_step = 1.0 / mask_height
+        w_step = 1.0 / mask_width
+
+        uncertainty_map = uncertainty_map.view(num_rois,
+                                               mask_height * mask_width)
+        num_points = min(mask_height * mask_width, num_points)
+        point_indices = uncertainty_map.topk(num_points, dim=1)[1]
+        point_coords = uncertainty_map.new_zeros(num_rois, num_points, 2)
+        point_coords[:, :, 0] = w_step / 2.0 + (point_indices %
+                                                mask_width).float() * w_step
+        point_coords[:, :, 1] = h_step / 2.0 + (point_indices //
+                                                mask_width).float() * h_step
+        return point_indices, point_coords
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_heads/maskiou_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_heads/maskiou_head.py
new file mode 100755
index 000000000..39bcd6a7d
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_heads/maskiou_head.py
@@ -0,0 +1,186 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.cnn import Conv2d, Linear, MaxPool2d, kaiming_init, normal_init
+from mmcv.runner import force_fp32
+from torch.nn.modules.utils import _pair
+
+from mmdet.models.builder import HEADS, build_loss
+
+
+@HEADS.register_module()
+class MaskIoUHead(nn.Module):
+    """Mask IoU Head.
+
+    This head predicts the IoU of predicted masks and corresponding gt masks.
+    """
+
+    def __init__(self,
+                 num_convs=4,
+                 num_fcs=2,
+                 roi_feat_size=14,
+                 in_channels=256,
+                 conv_out_channels=256,
+                 fc_out_channels=1024,
+                 num_classes=80,
+                 loss_iou=dict(type='MSELoss', loss_weight=0.5)):
+        super(MaskIoUHead, self).__init__()
+        self.in_channels = in_channels
+        self.conv_out_channels = conv_out_channels
+        self.fc_out_channels = fc_out_channels
+        self.num_classes = num_classes
+        self.fp16_enabled = False
+
+        self.convs = nn.ModuleList()
+        for i in range(num_convs):
+            if i == 0:
+                # concatenation of mask feature and mask prediction
+                in_channels = self.in_channels + 1
+            else:
+                in_channels = self.conv_out_channels
+            stride = 2 if i == num_convs - 1 else 1
+            self.convs.append(
+                Conv2d(
+                    in_channels,
+                    self.conv_out_channels,
+                    3,
+                    stride=stride,
+                    padding=1))
+
+        roi_feat_size = _pair(roi_feat_size)
+        pooled_area = (roi_feat_size[0] // 2) * (roi_feat_size[1] // 2)
+        self.fcs = nn.ModuleList()
+        for i in range(num_fcs):
+            in_channels = (
+                self.conv_out_channels *
+                pooled_area if i == 0 else self.fc_out_channels)
+            self.fcs.append(Linear(in_channels, self.fc_out_channels))
+
+        self.fc_mask_iou = Linear(self.fc_out_channels, self.num_classes)
+        self.relu = nn.ReLU()
+        self.max_pool = MaxPool2d(2, 2)
+        self.loss_iou = build_loss(loss_iou)
+
+    def init_weights(self):
+        for conv in self.convs:
+            kaiming_init(conv)
+        for fc in self.fcs:
+            kaiming_init(
+                fc,
+                a=1,
+                mode='fan_in',
+                nonlinearity='leaky_relu',
+                distribution='uniform')
+        normal_init(self.fc_mask_iou, std=0.01)
+
+    def forward(self, mask_feat, mask_pred):
+        mask_pred = mask_pred.sigmoid()
+        mask_pred_pooled = self.max_pool(mask_pred.unsqueeze(1))
+
+        x = torch.cat((mask_feat, mask_pred_pooled), 1)
+
+        for conv in self.convs:
+            x = self.relu(conv(x))
+        x = x.flatten(1)
+        for fc in self.fcs:
+            x = self.relu(fc(x))
+        mask_iou = self.fc_mask_iou(x)
+        return mask_iou
+
+    @force_fp32(apply_to=('mask_iou_pred', ))
+    def loss(self, mask_iou_pred, mask_iou_targets):
+        pos_inds = mask_iou_targets > 0
+        if pos_inds.sum() > 0:
+            loss_mask_iou = self.loss_iou(mask_iou_pred[pos_inds],
+                                          mask_iou_targets[pos_inds])
+        else:
+            loss_mask_iou = mask_iou_pred.sum() * 0
+        return dict(loss_mask_iou=loss_mask_iou)
+
+    @force_fp32(apply_to=('mask_pred', ))
+    def get_targets(self, sampling_results, gt_masks, mask_pred, mask_targets,
+                    rcnn_train_cfg):
+        """Compute target of mask IoU.
+
+        Mask IoU target is the IoU of the predicted mask (inside a bbox) and
+        the gt mask of corresponding gt mask (the whole instance).
+        The intersection area is computed inside the bbox, and the gt mask area
+        is computed with two steps, firstly we compute the gt area inside the
+        bbox, then divide it by the area ratio of gt area inside the bbox and
+        the gt area of the whole instance.
+
+        Args:
+            sampling_results (list[:obj:`SamplingResult`]): sampling results.
+            gt_masks (BitmapMask | PolygonMask): Gt masks (the whole instance)
+                of each image, with the same shape of the input image.
+            mask_pred (Tensor): Predicted masks of each positive proposal,
+                shape (num_pos, h, w).
+            mask_targets (Tensor): Gt mask of each positive proposal,
+                binary map of the shape (num_pos, h, w).
+            rcnn_train_cfg (dict): Training config for R-CNN part.
+
+        Returns:
+            Tensor: mask iou target (length == num positive).
+        """
+        pos_proposals = [res.pos_bboxes for res in sampling_results]
+        pos_assigned_gt_inds = [
+            res.pos_assigned_gt_inds for res in sampling_results
+        ]
+
+        # compute the area ratio of gt areas inside the proposals and
+        # the whole instance
+        area_ratios = map(self._get_area_ratio, pos_proposals,
+                          pos_assigned_gt_inds, gt_masks)
+        area_ratios = torch.cat(list(area_ratios))
+        assert mask_targets.size(0) == area_ratios.size(0)
+
+        mask_pred = (mask_pred > rcnn_train_cfg.mask_thr_binary).float()
+        mask_pred_areas = mask_pred.sum((-1, -2))
+
+        # mask_pred and mask_targets are binary maps
+        overlap_areas = (mask_pred * mask_targets).sum((-1, -2))
+
+        # compute the mask area of the whole instance
+        gt_full_areas = mask_targets.sum((-1, -2)) / (area_ratios + 1e-7)
+
+        mask_iou_targets = overlap_areas / (
+            mask_pred_areas + gt_full_areas - overlap_areas)
+        return mask_iou_targets
+
+    def _get_area_ratio(self, pos_proposals, pos_assigned_gt_inds, gt_masks):
+        """Compute area ratio of the gt mask inside the proposal and the gt
+        mask of the corresponding instance."""
+        num_pos = pos_proposals.size(0)
+        if num_pos > 0:
+            area_ratios = []
+            proposals_np = pos_proposals.cpu().numpy()
+            pos_assigned_gt_inds = pos_assigned_gt_inds.cpu().numpy()
+            # compute mask areas of gt instances (batch processing for speedup)
+            gt_instance_mask_area = gt_masks.areas
+            for i in range(num_pos):
+                gt_mask = gt_masks[pos_assigned_gt_inds[i]]
+
+                # crop the gt mask inside the proposal
+                bbox = proposals_np[i, :].astype(np.int32)
+                gt_mask_in_proposal = gt_mask.crop(bbox)
+
+                ratio = gt_mask_in_proposal.areas[0] / (
+                    gt_instance_mask_area[pos_assigned_gt_inds[i]] + 1e-7)
+                area_ratios.append(ratio)
+            area_ratios = torch.from_numpy(np.stack(area_ratios)).float().to(
+                pos_proposals.device)
+        else:
+            area_ratios = pos_proposals.new_zeros((0, ))
+        return area_ratios
+
+    @force_fp32(apply_to=('mask_iou_pred', ))
+    def get_mask_scores(self, mask_iou_pred, det_bboxes, det_labels):
+        """Get the mask scores.
+
+        mask_score = bbox_score * mask_iou
+        """
+        inds = range(det_labels.size(0))
+        mask_scores = mask_iou_pred[inds, det_labels] * det_bboxes[inds, -1]
+        mask_scores = mask_scores.cpu().numpy()
+        det_labels = det_labels.cpu().numpy()
+        return [mask_scores[det_labels == i] for i in range(self.num_classes)]
diff --git a/detection/scrfd/mmdet/models/roi_heads/mask_scoring_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/mask_scoring_roi_head.py
new file mode 100755
index 000000000..c6e55c775
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/mask_scoring_roi_head.py
@@ -0,0 +1,122 @@
+import torch
+
+from mmdet.core import bbox2roi
+from ..builder import HEADS, build_head
+from .standard_roi_head import StandardRoIHead
+
+
+@HEADS.register_module()
+class MaskScoringRoIHead(StandardRoIHead):
+    """Mask Scoring RoIHead for Mask Scoring RCNN.
+
+    https://arxiv.org/abs/1903.00241
+    """
+
+    def __init__(self, mask_iou_head, **kwargs):
+        assert mask_iou_head is not None
+        super(MaskScoringRoIHead, self).__init__(**kwargs)
+        self.mask_iou_head = build_head(mask_iou_head)
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        super(MaskScoringRoIHead, self).init_weights(pretrained)
+        self.mask_iou_head.init_weights()
+
+    def _mask_forward_train(self, x, sampling_results, bbox_feats, gt_masks,
+                            img_metas):
+        """Run forward function and calculate loss for Mask head in
+        training."""
+        pos_labels = torch.cat([res.pos_gt_labels for res in sampling_results])
+        mask_results = super(MaskScoringRoIHead,
+                             self)._mask_forward_train(x, sampling_results,
+                                                       bbox_feats, gt_masks,
+                                                       img_metas)
+        if mask_results['loss_mask'] is None:
+            return mask_results
+
+        # mask iou head forward and loss
+        pos_mask_pred = mask_results['mask_pred'][
+            range(mask_results['mask_pred'].size(0)), pos_labels]
+        mask_iou_pred = self.mask_iou_head(mask_results['mask_feats'],
+                                           pos_mask_pred)
+        pos_mask_iou_pred = mask_iou_pred[range(mask_iou_pred.size(0)),
+                                          pos_labels]
+
+        mask_iou_targets = self.mask_iou_head.get_targets(
+            sampling_results, gt_masks, pos_mask_pred,
+            mask_results['mask_targets'], self.train_cfg)
+        loss_mask_iou = self.mask_iou_head.loss(pos_mask_iou_pred,
+                                                mask_iou_targets)
+        mask_results['loss_mask'].update(loss_mask_iou)
+        return mask_results
+
+    def simple_test_mask(self,
+                         x,
+                         img_metas,
+                         det_bboxes,
+                         det_labels,
+                         rescale=False):
+        """Obtain mask prediction without augmentation."""
+        # image shapes of images in the batch
+        ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+
+        num_imgs = len(det_bboxes)
+        if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+            num_classes = self.mask_head.num_classes
+            segm_results = [[[] for _ in range(num_classes)]
+                            for _ in range(num_imgs)]
+            mask_scores = [[[] for _ in range(num_classes)]
+                           for _ in range(num_imgs)]
+        else:
+            # if det_bboxes is rescaled to the original image size, we need to
+            # rescale it back to the testing scale to obtain RoIs.
+            if rescale and not isinstance(scale_factors[0], float):
+                scale_factors = [
+                    torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                    for scale_factor in scale_factors
+                ]
+            _bboxes = [
+                det_bboxes[i][:, :4] *
+                scale_factors[i] if rescale else det_bboxes[i]
+                for i in range(num_imgs)
+            ]
+            mask_rois = bbox2roi(_bboxes)
+            mask_results = self._mask_forward(x, mask_rois)
+            concat_det_labels = torch.cat(det_labels)
+            # get mask scores with mask iou head
+            mask_feats = mask_results['mask_feats']
+            mask_pred = mask_results['mask_pred']
+            mask_iou_pred = self.mask_iou_head(
+                mask_feats, mask_pred[range(concat_det_labels.size(0)),
+                                      concat_det_labels])
+            # split batch mask prediction back to each image
+            num_bboxes_per_img = tuple(len(_bbox) for _bbox in _bboxes)
+            mask_preds = mask_pred.split(num_bboxes_per_img, 0)
+            mask_iou_preds = mask_iou_pred.split(num_bboxes_per_img, 0)
+
+            # apply mask post-processing to each image individually
+            segm_results = []
+            mask_scores = []
+            for i in range(num_imgs):
+                if det_bboxes[i].shape[0] == 0:
+                    segm_results.append(
+                        [[] for _ in range(self.mask_head.num_classes)])
+                    mask_scores.append(
+                        [[] for _ in range(self.mask_head.num_classes)])
+                else:
+                    segm_result = self.mask_head.get_seg_masks(
+                        mask_preds[i], _bboxes[i], det_labels[i],
+                        self.test_cfg, ori_shapes[i], scale_factors[i],
+                        rescale)
+                    # get mask scores with mask iou head
+                    mask_score = self.mask_iou_head.get_mask_scores(
+                        mask_iou_preds[i], det_bboxes[i], det_labels[i])
+                    segm_results.append(segm_result)
+                    mask_scores.append(mask_score)
+        return list(zip(segm_results, mask_scores))
diff --git a/detection/scrfd/mmdet/models/roi_heads/pisa_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/pisa_roi_head.py
new file mode 100755
index 000000000..e01113629
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/pisa_roi_head.py
@@ -0,0 +1,159 @@
+from mmdet.core import bbox2roi
+from ..builder import HEADS
+from ..losses.pisa_loss import carl_loss, isr_p
+from .standard_roi_head import StandardRoIHead
+
+
+@HEADS.register_module()
+class PISARoIHead(StandardRoIHead):
+    r"""The RoI head for `Prime Sample Attention in Object Detection
+    <https://arxiv.org/abs/1904.04821>`_."""
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None):
+        """Forward function for training.
+
+        Args:
+            x (list[Tensor]): List of multi-level img features.
+            img_metas (list[dict]): List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+            proposals (list[Tensors]): List of region proposals.
+            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box
+            gt_bboxes_ignore (list[Tensor], optional): Specify which bounding
+                boxes can be ignored when computing the loss.
+            gt_masks (None | Tensor) : True segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        # assign gts and sample proposals
+        if self.with_bbox or self.with_mask:
+            num_imgs = len(img_metas)
+            if gt_bboxes_ignore is None:
+                gt_bboxes_ignore = [None for _ in range(num_imgs)]
+            sampling_results = []
+            neg_label_weights = []
+            for i in range(num_imgs):
+                assign_result = self.bbox_assigner.assign(
+                    proposal_list[i], gt_bboxes[i], gt_bboxes_ignore[i],
+                    gt_labels[i])
+                sampling_result = self.bbox_sampler.sample(
+                    assign_result,
+                    proposal_list[i],
+                    gt_bboxes[i],
+                    gt_labels[i],
+                    feats=[lvl_feat[i][None] for lvl_feat in x])
+                # neg label weight is obtained by sampling when using ISR-N
+                neg_label_weight = None
+                if isinstance(sampling_result, tuple):
+                    sampling_result, neg_label_weight = sampling_result
+                sampling_results.append(sampling_result)
+                neg_label_weights.append(neg_label_weight)
+
+        losses = dict()
+        # bbox head forward and loss
+        if self.with_bbox:
+            bbox_results = self._bbox_forward_train(
+                x,
+                sampling_results,
+                gt_bboxes,
+                gt_labels,
+                img_metas,
+                neg_label_weights=neg_label_weights)
+            losses.update(bbox_results['loss_bbox'])
+
+        # mask head forward and loss
+        if self.with_mask:
+            mask_results = self._mask_forward_train(x, sampling_results,
+                                                    bbox_results['bbox_feats'],
+                                                    gt_masks, img_metas)
+            losses.update(mask_results['loss_mask'])
+
+        return losses
+
+    def _bbox_forward(self, x, rois):
+        """Box forward function used in both training and testing."""
+        # TODO: a more flexible way to decide which feature maps to use
+        bbox_feats = self.bbox_roi_extractor(
+            x[:self.bbox_roi_extractor.num_inputs], rois)
+        if self.with_shared_head:
+            bbox_feats = self.shared_head(bbox_feats)
+        cls_score, bbox_pred = self.bbox_head(bbox_feats)
+
+        bbox_results = dict(
+            cls_score=cls_score, bbox_pred=bbox_pred, bbox_feats=bbox_feats)
+        return bbox_results
+
+    def _bbox_forward_train(self,
+                            x,
+                            sampling_results,
+                            gt_bboxes,
+                            gt_labels,
+                            img_metas,
+                            neg_label_weights=None):
+        """Run forward function and calculate loss for box head in training."""
+        rois = bbox2roi([res.bboxes for res in sampling_results])
+
+        bbox_results = self._bbox_forward(x, rois)
+
+        bbox_targets = self.bbox_head.get_targets(sampling_results, gt_bboxes,
+                                                  gt_labels, self.train_cfg)
+
+        # neg_label_weights obtained by sampler is image-wise, mapping back to
+        # the corresponding location in label weights
+        if neg_label_weights[0] is not None:
+            label_weights = bbox_targets[1]
+            cur_num_rois = 0
+            for i in range(len(sampling_results)):
+                num_pos = sampling_results[i].pos_inds.size(0)
+                num_neg = sampling_results[i].neg_inds.size(0)
+                label_weights[cur_num_rois + num_pos:cur_num_rois + num_pos +
+                              num_neg] = neg_label_weights[i]
+                cur_num_rois += num_pos + num_neg
+
+        cls_score = bbox_results['cls_score']
+        bbox_pred = bbox_results['bbox_pred']
+
+        # Apply ISR-P
+        isr_cfg = self.train_cfg.get('isr', None)
+        if isr_cfg is not None:
+            bbox_targets = isr_p(
+                cls_score,
+                bbox_pred,
+                bbox_targets,
+                rois,
+                sampling_results,
+                self.bbox_head.loss_cls,
+                self.bbox_head.bbox_coder,
+                **isr_cfg,
+                num_class=self.bbox_head.num_classes)
+        loss_bbox = self.bbox_head.loss(cls_score, bbox_pred, rois,
+                                        *bbox_targets)
+
+        # Add CARL Loss
+        carl_cfg = self.train_cfg.get('carl', None)
+        if carl_cfg is not None:
+            loss_carl = carl_loss(
+                cls_score,
+                bbox_targets[0],
+                bbox_pred,
+                bbox_targets[2],
+                self.bbox_head.loss_bbox,
+                **carl_cfg,
+                num_class=self.bbox_head.num_classes)
+            loss_bbox.update(loss_carl)
+
+        bbox_results.update(loss_bbox=loss_bbox)
+        return bbox_results
diff --git a/detection/scrfd/mmdet/models/roi_heads/point_rend_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/point_rend_roi_head.py
new file mode 100755
index 000000000..478cdf5bf
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/point_rend_roi_head.py
@@ -0,0 +1,218 @@
+# Modified from https://github.com/facebookresearch/detectron2/tree/master/projects/PointRend  # noqa
+
+import torch
+import torch.nn.functional as F
+from mmcv.ops import point_sample, rel_roi_point_to_rel_img_point
+
+from mmdet.core import bbox2roi, bbox_mapping, merge_aug_masks
+from .. import builder
+from ..builder import HEADS
+from .standard_roi_head import StandardRoIHead
+
+
+@HEADS.register_module()
+class PointRendRoIHead(StandardRoIHead):
+    """`PointRend <https://arxiv.org/abs/1912.08193>`_."""
+
+    def __init__(self, point_head, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.with_bbox and self.with_mask
+        self.init_point_head(point_head)
+
+    def init_point_head(self, point_head):
+        """Initialize ``point_head``"""
+        self.point_head = builder.build_head(point_head)
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+        """
+        super().init_weights(pretrained)
+        self.point_head.init_weights()
+
+    def _mask_forward_train(self, x, sampling_results, bbox_feats, gt_masks,
+                            img_metas):
+        """Run forward function and calculate loss for mask head and point head
+        in training."""
+        mask_results = super()._mask_forward_train(x, sampling_results,
+                                                   bbox_feats, gt_masks,
+                                                   img_metas)
+        if mask_results['loss_mask'] is not None:
+            loss_point = self._mask_point_forward_train(
+                x, sampling_results, mask_results['mask_pred'], gt_masks,
+                img_metas)
+            mask_results['loss_mask'].update(loss_point)
+
+        return mask_results
+
+    def _mask_point_forward_train(self, x, sampling_results, mask_pred,
+                                  gt_masks, img_metas):
+        """Run forward function and calculate loss for point head in
+        training."""
+        pos_labels = torch.cat([res.pos_gt_labels for res in sampling_results])
+        rel_roi_points = self.point_head.get_roi_rel_points_train(
+            mask_pred, pos_labels, cfg=self.train_cfg)
+        rois = bbox2roi([res.pos_bboxes for res in sampling_results])
+
+        fine_grained_point_feats = self._get_fine_grained_point_feats(
+            x, rois, rel_roi_points, img_metas)
+        coarse_point_feats = point_sample(mask_pred, rel_roi_points)
+        mask_point_pred = self.point_head(fine_grained_point_feats,
+                                          coarse_point_feats)
+        mask_point_target = self.point_head.get_targets(
+            rois, rel_roi_points, sampling_results, gt_masks, self.train_cfg)
+        loss_mask_point = self.point_head.loss(mask_point_pred,
+                                               mask_point_target, pos_labels)
+
+        return loss_mask_point
+
+    def _get_fine_grained_point_feats(self, x, rois, rel_roi_points,
+                                      img_metas):
+        """Sample fine grained feats from each level feature map and
+        concatenate them together."""
+        num_imgs = len(img_metas)
+        fine_grained_feats = []
+        for idx in range(self.mask_roi_extractor.num_inputs):
+            feats = x[idx]
+            spatial_scale = 1. / float(
+                self.mask_roi_extractor.featmap_strides[idx])
+            point_feats = []
+            for batch_ind in range(num_imgs):
+                # unravel batch dim
+                feat = feats[batch_ind].unsqueeze(0)
+                inds = (rois[:, 0].long() == batch_ind)
+                if inds.any():
+                    rel_img_points = rel_roi_point_to_rel_img_point(
+                        rois[inds], rel_roi_points[inds], feat.shape[2:],
+                        spatial_scale).unsqueeze(0)
+                    point_feat = point_sample(feat, rel_img_points)
+                    point_feat = point_feat.squeeze(0).transpose(0, 1)
+                    point_feats.append(point_feat)
+            fine_grained_feats.append(torch.cat(point_feats, dim=0))
+        return torch.cat(fine_grained_feats, dim=1)
+
+    def _mask_point_forward_test(self, x, rois, label_pred, mask_pred,
+                                 img_metas):
+        """Mask refining process with point head in testing."""
+        refined_mask_pred = mask_pred.clone()
+        for subdivision_step in range(self.test_cfg.subdivision_steps):
+            refined_mask_pred = F.interpolate(
+                refined_mask_pred,
+                scale_factor=self.test_cfg.scale_factor,
+                mode='bilinear',
+                align_corners=False)
+            # If `subdivision_num_points` is larger or equal to the
+            # resolution of the next step, then we can skip this step
+            num_rois, channels, mask_height, mask_width = \
+                refined_mask_pred.shape
+            if (self.test_cfg.subdivision_num_points >=
+                    self.test_cfg.scale_factor**2 * mask_height * mask_width
+                    and
+                    subdivision_step < self.test_cfg.subdivision_steps - 1):
+                continue
+            point_indices, rel_roi_points = \
+                self.point_head.get_roi_rel_points_test(
+                    refined_mask_pred, label_pred, cfg=self.test_cfg)
+            fine_grained_point_feats = self._get_fine_grained_point_feats(
+                x, rois, rel_roi_points, img_metas)
+            coarse_point_feats = point_sample(mask_pred, rel_roi_points)
+            mask_point_pred = self.point_head(fine_grained_point_feats,
+                                              coarse_point_feats)
+
+            point_indices = point_indices.unsqueeze(1).expand(-1, channels, -1)
+            refined_mask_pred = refined_mask_pred.reshape(
+                num_rois, channels, mask_height * mask_width)
+            refined_mask_pred = refined_mask_pred.scatter_(
+                2, point_indices, mask_point_pred)
+            refined_mask_pred = refined_mask_pred.view(num_rois, channels,
+                                                       mask_height, mask_width)
+
+        return refined_mask_pred
+
+    def simple_test_mask(self,
+                         x,
+                         img_metas,
+                         det_bboxes,
+                         det_labels,
+                         rescale=False):
+        """Obtain mask prediction without augmentation."""
+        ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+        num_imgs = len(det_bboxes)
+        if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+            segm_results = [[[] for _ in range(self.mask_head.num_classes)]
+                            for _ in range(num_imgs)]
+        else:
+            # if det_bboxes is rescaled to the original image size, we need to
+            # rescale it back to the testing scale to obtain RoIs.
+            if rescale and not isinstance(scale_factors[0], float):
+                scale_factors = [
+                    torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                    for scale_factor in scale_factors
+                ]
+            _bboxes = [
+                det_bboxes[i][:, :4] *
+                scale_factors[i] if rescale else det_bboxes[i][:, :4]
+                for i in range(len(det_bboxes))
+            ]
+            mask_rois = bbox2roi(_bboxes)
+            mask_results = self._mask_forward(x, mask_rois)
+            # split batch mask prediction back to each image
+            mask_pred = mask_results['mask_pred']
+            num_mask_roi_per_img = [len(det_bbox) for det_bbox in det_bboxes]
+            mask_preds = mask_pred.split(num_mask_roi_per_img, 0)
+            mask_rois = mask_rois.split(num_mask_roi_per_img, 0)
+
+            # apply mask post-processing to each image individually
+            segm_results = []
+            for i in range(num_imgs):
+                if det_bboxes[i].shape[0] == 0:
+                    segm_results.append(
+                        [[] for _ in range(self.mask_head.num_classes)])
+                else:
+                    x_i = [xx[[i]] for xx in x]
+                    mask_rois_i = mask_rois[i]
+                    mask_rois_i[:, 0] = 0  # TODO: remove this hack
+                    mask_pred_i = self._mask_point_forward_test(
+                        x_i, mask_rois_i, det_labels[i], mask_preds[i],
+                        [img_metas])
+                    segm_result = self.mask_head.get_seg_masks(
+                        mask_pred_i, _bboxes[i], det_labels[i], self.test_cfg,
+                        ori_shapes[i], scale_factors[i], rescale)
+                    segm_results.append(segm_result)
+        return segm_results
+
+    def aug_test_mask(self, feats, img_metas, det_bboxes, det_labels):
+        """Test for mask head with test time augmentation."""
+        if det_bboxes.shape[0] == 0:
+            segm_result = [[] for _ in range(self.mask_head.num_classes)]
+        else:
+            aug_masks = []
+            for x, img_meta in zip(feats, img_metas):
+                img_shape = img_meta[0]['img_shape']
+                scale_factor = img_meta[0]['scale_factor']
+                flip = img_meta[0]['flip']
+                _bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
+                                       scale_factor, flip)
+                mask_rois = bbox2roi([_bboxes])
+                mask_results = self._mask_forward(x, mask_rois)
+                mask_results['mask_pred'] = self._mask_point_forward_test(
+                    x, mask_rois, det_labels, mask_results['mask_pred'],
+                    img_metas)
+                # convert to numpy array to save memory
+                aug_masks.append(
+                    mask_results['mask_pred'].sigmoid().cpu().numpy())
+            merged_masks = merge_aug_masks(aug_masks, img_metas, self.test_cfg)
+
+            ori_shape = img_metas[0][0]['ori_shape']
+            segm_result = self.mask_head.get_seg_masks(
+                merged_masks,
+                det_bboxes,
+                det_labels,
+                self.test_cfg,
+                ori_shape,
+                scale_factor=1.0,
+                rescale=False)
+        return segm_result
diff --git a/detection/scrfd/mmdet/models/roi_heads/roi_extractors/__init__.py b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/__init__.py
new file mode 100755
index 000000000..a6ec0ecc3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/__init__.py
@@ -0,0 +1,7 @@
+from .generic_roi_extractor import GenericRoIExtractor
+from .single_level_roi_extractor import SingleRoIExtractor
+
+__all__ = [
+    'SingleRoIExtractor',
+    'GenericRoIExtractor',
+]
diff --git a/detection/scrfd/mmdet/models/roi_heads/roi_extractors/base_roi_extractor.py b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/base_roi_extractor.py
new file mode 100755
index 000000000..0e42b52f3
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/base_roi_extractor.py
@@ -0,0 +1,83 @@
+from abc import ABCMeta, abstractmethod
+
+import torch
+import torch.nn as nn
+from mmcv import ops
+
+
+class BaseRoIExtractor(nn.Module, metaclass=ABCMeta):
+    """Base class for RoI extractor.
+
+    Args:
+        roi_layer (dict): Specify RoI layer type and arguments.
+        out_channels (int): Output channels of RoI layers.
+        featmap_strides (int): Strides of input feature maps.
+    """
+
+    def __init__(self, roi_layer, out_channels, featmap_strides):
+        super(BaseRoIExtractor, self).__init__()
+        self.roi_layers = self.build_roi_layers(roi_layer, featmap_strides)
+        self.out_channels = out_channels
+        self.featmap_strides = featmap_strides
+        self.fp16_enabled = False
+
+    @property
+    def num_inputs(self):
+        """int: Number of input feature maps."""
+        return len(self.featmap_strides)
+
+    def init_weights(self):
+        pass
+
+    def build_roi_layers(self, layer_cfg, featmap_strides):
+        """Build RoI operator to extract feature from each level feature map.
+
+        Args:
+            layer_cfg (dict): Dictionary to construct and config RoI layer
+                operation. Options are modules under ``mmcv/ops`` such as
+                ``RoIAlign``.
+            featmap_strides (int): The stride of input feature map w.r.t to the
+                original image size, which would be used to scale RoI
+                coordinate (original image coordinate system) to feature
+                coordinate system.
+
+        Returns:
+            nn.ModuleList: The RoI extractor modules for each level feature
+                map.
+        """
+
+        cfg = layer_cfg.copy()
+        layer_type = cfg.pop('type')
+        assert hasattr(ops, layer_type)
+        layer_cls = getattr(ops, layer_type)
+        roi_layers = nn.ModuleList(
+            [layer_cls(spatial_scale=1 / s, **cfg) for s in featmap_strides])
+        return roi_layers
+
+    def roi_rescale(self, rois, scale_factor):
+        """Scale RoI coordinates by scale factor.
+
+        Args:
+            rois (torch.Tensor): RoI (Region of Interest), shape (n, 5)
+            scale_factor (float): Scale factor that RoI will be multiplied by.
+
+        Returns:
+            torch.Tensor: Scaled RoI.
+        """
+
+        cx = (rois[:, 1] + rois[:, 3]) * 0.5
+        cy = (rois[:, 2] + rois[:, 4]) * 0.5
+        w = rois[:, 3] - rois[:, 1]
+        h = rois[:, 4] - rois[:, 2]
+        new_w = w * scale_factor
+        new_h = h * scale_factor
+        x1 = cx - new_w * 0.5
+        x2 = cx + new_w * 0.5
+        y1 = cy - new_h * 0.5
+        y2 = cy + new_h * 0.5
+        new_rois = torch.stack((rois[:, 0], x1, y1, x2, y2), dim=-1)
+        return new_rois
+
+    @abstractmethod
+    def forward(self, feats, rois, roi_scale_factor=None):
+        pass
diff --git a/detection/scrfd/mmdet/models/roi_heads/roi_extractors/generic_roi_extractor.py b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/generic_roi_extractor.py
new file mode 100755
index 000000000..80c25bb8f
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/generic_roi_extractor.py
@@ -0,0 +1,83 @@
+from mmcv.cnn.bricks import build_plugin_layer
+from mmcv.runner import force_fp32
+
+from mmdet.models.builder import ROI_EXTRACTORS
+from .base_roi_extractor import BaseRoIExtractor
+
+
+@ROI_EXTRACTORS.register_module()
+class GenericRoIExtractor(BaseRoIExtractor):
+    """Extract RoI features from all level feature maps levels.
+
+    This is the implementation of `A novel Region of Interest Extraction Layer
+    for Instance Segmentation <https://arxiv.org/abs/2004.13665>`_.
+
+    Args:
+        aggregation (str): The method to aggregate multiple feature maps.
+            Options are 'sum', 'concat'. Default: 'sum'.
+        pre_cfg (dict | None): Specify pre-processing modules. Default: None.
+        post_cfg (dict | None): Specify post-processing modules. Default: None.
+        kwargs (keyword arguments): Arguments that are the same
+            as :class:`BaseRoIExtractor`.
+    """
+
+    def __init__(self,
+                 aggregation='sum',
+                 pre_cfg=None,
+                 post_cfg=None,
+                 **kwargs):
+        super(GenericRoIExtractor, self).__init__(**kwargs)
+
+        assert aggregation in ['sum', 'concat']
+
+        self.aggregation = aggregation
+        self.with_post = post_cfg is not None
+        self.with_pre = pre_cfg is not None
+        # build pre/post processing modules
+        if self.with_post:
+            self.post_module = build_plugin_layer(post_cfg, '_post_module')[1]
+        if self.with_pre:
+            self.pre_module = build_plugin_layer(pre_cfg, '_pre_module')[1]
+
+    @force_fp32(apply_to=('feats', ), out_fp16=True)
+    def forward(self, feats, rois, roi_scale_factor=None):
+        """Forward function."""
+        if len(feats) == 1:
+            return self.roi_layers[0](feats[0], rois)
+
+        out_size = self.roi_layers[0].output_size
+        num_levels = len(feats)
+        roi_feats = feats[0].new_zeros(
+            rois.size(0), self.out_channels, *out_size)
+
+        # some times rois is an empty tensor
+        if roi_feats.shape[0] == 0:
+            return roi_feats
+
+        if roi_scale_factor is not None:
+            rois = self.roi_rescale(rois, roi_scale_factor)
+
+        # mark the starting channels for concat mode
+        start_channels = 0
+        for i in range(num_levels):
+            roi_feats_t = self.roi_layers[i](feats[i], rois)
+            end_channels = start_channels + roi_feats_t.size(1)
+            if self.with_pre:
+                # apply pre-processing to a RoI extracted from each layer
+                roi_feats_t = self.pre_module(roi_feats_t)
+            if self.aggregation == 'sum':
+                # and sum them all
+                roi_feats += roi_feats_t
+            else:
+                # and concat them along channel dimension
+                roi_feats[:, start_channels:end_channels] = roi_feats_t
+            # update channels starting position
+            start_channels = end_channels
+        # check if concat channels match at the end
+        if self.aggregation == 'concat':
+            assert start_channels == self.out_channels
+
+        if self.with_post:
+            # apply post-processing before return the result
+            roi_feats = self.post_module(roi_feats)
+        return roi_feats
diff --git a/detection/scrfd/mmdet/models/roi_heads/roi_extractors/single_level_roi_extractor.py b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/single_level_roi_extractor.py
new file mode 100755
index 000000000..c0eebc4af
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/roi_extractors/single_level_roi_extractor.py
@@ -0,0 +1,99 @@
+import torch
+from mmcv.runner import force_fp32
+
+from mmdet.models.builder import ROI_EXTRACTORS
+from .base_roi_extractor import BaseRoIExtractor
+
+
+@ROI_EXTRACTORS.register_module()
+class SingleRoIExtractor(BaseRoIExtractor):
+    """Extract RoI features from a single level feature map.
+
+    If there are multiple input feature levels, each RoI is mapped to a level
+    according to its scale. The mapping rule is proposed in
+    `FPN <https://arxiv.org/abs/1612.03144>`_.
+
+    Args:
+        roi_layer (dict): Specify RoI layer type and arguments.
+        out_channels (int): Output channels of RoI layers.
+        featmap_strides (int): Strides of input feature maps.
+        finest_scale (int): Scale threshold of mapping to level 0. Default: 56.
+    """
+
+    def __init__(self,
+                 roi_layer,
+                 out_channels,
+                 featmap_strides,
+                 finest_scale=56):
+        super(SingleRoIExtractor, self).__init__(roi_layer, out_channels,
+                                                 featmap_strides)
+        self.finest_scale = finest_scale
+
+    def map_roi_levels(self, rois, num_levels):
+        """Map rois to corresponding feature levels by scales.
+
+        - scale < finest_scale * 2: level 0
+        - finest_scale * 2 <= scale < finest_scale * 4: level 1
+        - finest_scale * 4 <= scale < finest_scale * 8: level 2
+        - scale >= finest_scale * 8: level 3
+
+        Args:
+            rois (Tensor): Input RoIs, shape (k, 5).
+            num_levels (int): Total level number.
+
+        Returns:
+            Tensor: Level index (0-based) of each RoI, shape (k, )
+        """
+        scale = torch.sqrt(
+            (rois[:, 3] - rois[:, 1]) * (rois[:, 4] - rois[:, 2]))
+        target_lvls = torch.floor(torch.log2(scale / self.finest_scale + 1e-6))
+        target_lvls = target_lvls.clamp(min=0, max=num_levels - 1).long()
+        return target_lvls
+
+    @force_fp32(apply_to=('feats', ), out_fp16=True)
+    def forward(self, feats, rois, roi_scale_factor=None):
+        """Forward function."""
+        out_size = self.roi_layers[0].output_size
+        num_levels = len(feats)
+        if torch.onnx.is_in_onnx_export():
+            # Work around to export mask-rcnn to onnx
+            roi_feats = rois[:, :1].clone().detach()
+            roi_feats = roi_feats.expand(
+                -1, self.out_channels * out_size[0] * out_size[1])
+            roi_feats = roi_feats.reshape(-1, self.out_channels, *out_size)
+            roi_feats = roi_feats * 0
+        else:
+            roi_feats = feats[0].new_zeros(
+                rois.size(0), self.out_channels, *out_size)
+        # TODO: remove this when parrots supports
+        if torch.__version__ == 'parrots':
+            roi_feats.requires_grad = True
+
+        if num_levels == 1:
+            if len(rois) == 0:
+                return roi_feats
+            return self.roi_layers[0](feats[0], rois)
+
+        target_lvls = self.map_roi_levels(rois, num_levels)
+        if roi_scale_factor is not None:
+            rois = self.roi_rescale(rois, roi_scale_factor)
+
+        for i in range(num_levels):
+            mask = target_lvls == i
+            inds = mask.nonzero(as_tuple=False).squeeze(1)
+            # TODO: make it nicer when exporting to onnx
+            if torch.onnx.is_in_onnx_export():
+                # To keep all roi_align nodes exported to onnx
+                rois_ = rois[inds]
+                roi_feats_t = self.roi_layers[i](feats[i], rois_)
+                roi_feats[inds] = roi_feats_t
+                continue
+            if inds.numel() > 0:
+                rois_ = rois[inds]
+                roi_feats_t = self.roi_layers[i](feats[i], rois_)
+                roi_feats[inds] = roi_feats_t
+            else:
+                roi_feats += sum(
+                    x.view(-1)[0]
+                    for x in self.parameters()) * 0. + feats[i].sum() * 0.
+        return roi_feats
diff --git a/detection/scrfd/mmdet/models/roi_heads/shared_heads/__init__.py b/detection/scrfd/mmdet/models/roi_heads/shared_heads/__init__.py
new file mode 100755
index 000000000..bbe70145b
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/shared_heads/__init__.py
@@ -0,0 +1,3 @@
+from .res_layer import ResLayer
+
+__all__ = ['ResLayer']
diff --git a/detection/scrfd/mmdet/models/roi_heads/shared_heads/res_layer.py b/detection/scrfd/mmdet/models/roi_heads/shared_heads/res_layer.py
new file mode 100755
index 000000000..b5c343258
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/shared_heads/res_layer.py
@@ -0,0 +1,77 @@
+import torch.nn as nn
+from mmcv.cnn import constant_init, kaiming_init
+from mmcv.runner import auto_fp16, load_checkpoint
+
+from mmdet.models.backbones import ResNet
+from mmdet.models.builder import SHARED_HEADS
+from mmdet.models.utils import ResLayer as _ResLayer
+from mmdet.utils import get_root_logger
+
+
+@SHARED_HEADS.register_module()
+class ResLayer(nn.Module):
+
+    def __init__(self,
+                 depth,
+                 stage=3,
+                 stride=2,
+                 dilation=1,
+                 style='pytorch',
+                 norm_cfg=dict(type='BN', requires_grad=True),
+                 norm_eval=True,
+                 with_cp=False,
+                 dcn=None):
+        super(ResLayer, self).__init__()
+        self.norm_eval = norm_eval
+        self.norm_cfg = norm_cfg
+        self.stage = stage
+        self.fp16_enabled = False
+        block, stage_blocks = ResNet.arch_settings[depth]
+        stage_block = stage_blocks[stage]
+        planes = 64 * 2**stage
+        inplanes = 64 * 2**(stage - 1) * block.expansion
+
+        res_layer = _ResLayer(
+            block,
+            inplanes,
+            planes,
+            stage_block,
+            stride=stride,
+            dilation=dilation,
+            style=style,
+            with_cp=with_cp,
+            norm_cfg=self.norm_cfg,
+            dcn=dcn)
+        self.add_module(f'layer{stage + 1}', res_layer)
+
+    def init_weights(self, pretrained=None):
+        """Initialize the weights in the module.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if isinstance(pretrained, str):
+            logger = get_root_logger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, nn.BatchNorm2d):
+                    constant_init(m, 1)
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+    @auto_fp16()
+    def forward(self, x):
+        res_layer = getattr(self, f'layer{self.stage + 1}')
+        out = res_layer(x)
+        return out
+
+    def train(self, mode=True):
+        super(ResLayer, self).train(mode)
+        if self.norm_eval:
+            for m in self.modules():
+                if isinstance(m, nn.BatchNorm2d):
+                    m.eval()
diff --git a/detection/scrfd/mmdet/models/roi_heads/standard_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/standard_roi_head.py
new file mode 100755
index 000000000..c530f2a5c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/standard_roi_head.py
@@ -0,0 +1,295 @@
+import torch
+
+from mmdet.core import bbox2result, bbox2roi, build_assigner, build_sampler
+from ..builder import HEADS, build_head, build_roi_extractor
+from .base_roi_head import BaseRoIHead
+from .test_mixins import BBoxTestMixin, MaskTestMixin
+
+
+@HEADS.register_module()
+class StandardRoIHead(BaseRoIHead, BBoxTestMixin, MaskTestMixin):
+    """Simplest base roi head including one bbox head and one mask head."""
+
+    def init_assigner_sampler(self):
+        """Initialize assigner and sampler."""
+        self.bbox_assigner = None
+        self.bbox_sampler = None
+        if self.train_cfg:
+            self.bbox_assigner = build_assigner(self.train_cfg.assigner)
+            self.bbox_sampler = build_sampler(
+                self.train_cfg.sampler, context=self)
+
+    def init_bbox_head(self, bbox_roi_extractor, bbox_head):
+        """Initialize ``bbox_head``"""
+        self.bbox_roi_extractor = build_roi_extractor(bbox_roi_extractor)
+        self.bbox_head = build_head(bbox_head)
+
+    def init_mask_head(self, mask_roi_extractor, mask_head):
+        """Initialize ``mask_head``"""
+        if mask_roi_extractor is not None:
+            self.mask_roi_extractor = build_roi_extractor(mask_roi_extractor)
+            self.share_roi_extractor = False
+        else:
+            self.share_roi_extractor = True
+            self.mask_roi_extractor = self.bbox_roi_extractor
+        self.mask_head = build_head(mask_head)
+
+    def init_weights(self, pretrained):
+        """Initialize the weights in head.
+
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        if self.with_shared_head:
+            self.shared_head.init_weights(pretrained=pretrained)
+        if self.with_bbox:
+            self.bbox_roi_extractor.init_weights()
+            self.bbox_head.init_weights()
+        if self.with_mask:
+            self.mask_head.init_weights()
+            if not self.share_roi_extractor:
+                self.mask_roi_extractor.init_weights()
+
+    def forward_dummy(self, x, proposals):
+        """Dummy forward function."""
+        # bbox head
+        outs = ()
+        rois = bbox2roi([proposals])
+        if self.with_bbox:
+            bbox_results = self._bbox_forward(x, rois)
+            outs = outs + (bbox_results['cls_score'],
+                           bbox_results['bbox_pred'])
+        # mask head
+        if self.with_mask:
+            mask_rois = rois[:100]
+            mask_results = self._mask_forward(x, mask_rois)
+            outs = outs + (mask_results['mask_pred'], )
+        return outs
+
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      proposal_list,
+                      gt_bboxes,
+                      gt_labels,
+                      gt_bboxes_ignore=None,
+                      gt_masks=None):
+        """
+        Args:
+            x (list[Tensor]): list of multi-level img features.
+            img_metas (list[dict]): list of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `mmdet/datasets/pipelines/formatting.py:Collect`.
+            proposals (list[Tensors]): list of region proposals.
+            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
+                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): class indices corresponding to each box
+            gt_bboxes_ignore (None | list[Tensor]): specify which bounding
+                boxes can be ignored when computing the loss.
+            gt_masks (None | Tensor) : true segmentation masks for each box
+                used if the architecture supports a segmentation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        # assign gts and sample proposals
+        if self.with_bbox or self.with_mask:
+            num_imgs = len(img_metas)
+            if gt_bboxes_ignore is None:
+                gt_bboxes_ignore = [None for _ in range(num_imgs)]
+            sampling_results = []
+            for i in range(num_imgs):
+                assign_result = self.bbox_assigner.assign(
+                    proposal_list[i], gt_bboxes[i], gt_bboxes_ignore[i],
+                    gt_labels[i])
+                sampling_result = self.bbox_sampler.sample(
+                    assign_result,
+                    proposal_list[i],
+                    gt_bboxes[i],
+                    gt_labels[i],
+                    feats=[lvl_feat[i][None] for lvl_feat in x])
+                sampling_results.append(sampling_result)
+
+        losses = dict()
+        # bbox head forward and loss
+        if self.with_bbox:
+            bbox_results = self._bbox_forward_train(x, sampling_results,
+                                                    gt_bboxes, gt_labels,
+                                                    img_metas)
+            losses.update(bbox_results['loss_bbox'])
+
+        # mask head forward and loss
+        if self.with_mask:
+            mask_results = self._mask_forward_train(x, sampling_results,
+                                                    bbox_results['bbox_feats'],
+                                                    gt_masks, img_metas)
+            losses.update(mask_results['loss_mask'])
+
+        return losses
+
+    def _bbox_forward(self, x, rois):
+        """Box head forward function used in both training and testing."""
+        # TODO: a more flexible way to decide which feature maps to use
+        bbox_feats = self.bbox_roi_extractor(
+            x[:self.bbox_roi_extractor.num_inputs], rois)
+        if self.with_shared_head:
+            bbox_feats = self.shared_head(bbox_feats)
+        cls_score, bbox_pred = self.bbox_head(bbox_feats)
+
+        bbox_results = dict(
+            cls_score=cls_score, bbox_pred=bbox_pred, bbox_feats=bbox_feats)
+        return bbox_results
+
+    def _bbox_forward_train(self, x, sampling_results, gt_bboxes, gt_labels,
+                            img_metas):
+        """Run forward function and calculate loss for box head in training."""
+        rois = bbox2roi([res.bboxes for res in sampling_results])
+        bbox_results = self._bbox_forward(x, rois)
+
+        bbox_targets = self.bbox_head.get_targets(sampling_results, gt_bboxes,
+                                                  gt_labels, self.train_cfg)
+        loss_bbox = self.bbox_head.loss(bbox_results['cls_score'],
+                                        bbox_results['bbox_pred'], rois,
+                                        *bbox_targets)
+
+        bbox_results.update(loss_bbox=loss_bbox)
+        return bbox_results
+
+    def _mask_forward_train(self, x, sampling_results, bbox_feats, gt_masks,
+                            img_metas):
+        """Run forward function and calculate loss for mask head in
+        training."""
+        if not self.share_roi_extractor:
+            pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
+            mask_results = self._mask_forward(x, pos_rois)
+        else:
+            pos_inds = []
+            device = bbox_feats.device
+            for res in sampling_results:
+                pos_inds.append(
+                    torch.ones(
+                        res.pos_bboxes.shape[0],
+                        device=device,
+                        dtype=torch.uint8))
+                pos_inds.append(
+                    torch.zeros(
+                        res.neg_bboxes.shape[0],
+                        device=device,
+                        dtype=torch.uint8))
+            pos_inds = torch.cat(pos_inds)
+
+            mask_results = self._mask_forward(
+                x, pos_inds=pos_inds, bbox_feats=bbox_feats)
+
+        mask_targets = self.mask_head.get_targets(sampling_results, gt_masks,
+                                                  self.train_cfg)
+        pos_labels = torch.cat([res.pos_gt_labels for res in sampling_results])
+        loss_mask = self.mask_head.loss(mask_results['mask_pred'],
+                                        mask_targets, pos_labels)
+
+        mask_results.update(loss_mask=loss_mask, mask_targets=mask_targets)
+        return mask_results
+
+    def _mask_forward(self, x, rois=None, pos_inds=None, bbox_feats=None):
+        """Mask head forward function used in both training and testing."""
+        assert ((rois is not None) ^
+                (pos_inds is not None and bbox_feats is not None))
+        if rois is not None:
+            mask_feats = self.mask_roi_extractor(
+                x[:self.mask_roi_extractor.num_inputs], rois)
+            if self.with_shared_head:
+                mask_feats = self.shared_head(mask_feats)
+        else:
+            assert bbox_feats is not None
+            mask_feats = bbox_feats[pos_inds]
+
+        mask_pred = self.mask_head(mask_feats)
+        mask_results = dict(mask_pred=mask_pred, mask_feats=mask_feats)
+        return mask_results
+
+    async def async_simple_test(self,
+                                x,
+                                proposal_list,
+                                img_metas,
+                                proposals=None,
+                                rescale=False):
+        """Async test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+
+        det_bboxes, det_labels = await self.async_test_bboxes(
+            x, img_metas, proposal_list, self.test_cfg, rescale=rescale)
+        bbox_results = bbox2result(det_bboxes, det_labels,
+                                   self.bbox_head.num_classes)
+        if not self.with_mask:
+            return bbox_results
+        else:
+            segm_results = await self.async_test_mask(
+                x,
+                img_metas,
+                det_bboxes,
+                det_labels,
+                rescale=rescale,
+                mask_test_cfg=self.test_cfg.get('mask'))
+            return bbox_results, segm_results
+
+    def simple_test(self,
+                    x,
+                    proposal_list,
+                    img_metas,
+                    proposals=None,
+                    rescale=False):
+        """Test without augmentation."""
+        assert self.with_bbox, 'Bbox head must be implemented.'
+
+        det_bboxes, det_labels = self.simple_test_bboxes(
+            x, img_metas, proposal_list, self.test_cfg, rescale=rescale)
+        if torch.onnx.is_in_onnx_export():
+            if self.with_mask:
+                segm_results = self.simple_test_mask(
+                    x, img_metas, det_bboxes, det_labels, rescale=rescale)
+                return det_bboxes, det_labels, segm_results
+            else:
+                return det_bboxes, det_labels
+
+        bbox_results = [
+            bbox2result(det_bboxes[i], det_labels[i],
+                        self.bbox_head.num_classes)
+            for i in range(len(det_bboxes))
+        ]
+
+        if not self.with_mask:
+            return bbox_results
+        else:
+            segm_results = self.simple_test_mask(
+                x, img_metas, det_bboxes, det_labels, rescale=rescale)
+            return list(zip(bbox_results, segm_results))
+
+    def aug_test(self, x, proposal_list, img_metas, rescale=False):
+        """Test with augmentations.
+
+        If rescale is False, then returned bboxes and masks will fit the scale
+        of imgs[0].
+        """
+        det_bboxes, det_labels = self.aug_test_bboxes(x, img_metas,
+                                                      proposal_list,
+                                                      self.test_cfg)
+
+        if rescale:
+            _det_bboxes = det_bboxes
+        else:
+            _det_bboxes = det_bboxes.clone()
+            _det_bboxes[:, :4] *= det_bboxes.new_tensor(
+                img_metas[0][0]['scale_factor'])
+        bbox_results = bbox2result(_det_bboxes, det_labels,
+                                   self.bbox_head.num_classes)
+
+        # det_bboxes always keep the original scale
+        if self.with_mask:
+            segm_results = self.aug_test_mask(x, img_metas, det_bboxes,
+                                              det_labels)
+            return [(bbox_results, segm_results)]
+        else:
+            return [bbox_results]
diff --git a/detection/scrfd/mmdet/models/roi_heads/test_mixins.py b/detection/scrfd/mmdet/models/roi_heads/test_mixins.py
new file mode 100755
index 000000000..12684c52c
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/test_mixins.py
@@ -0,0 +1,268 @@
+import logging
+import sys
+
+import torch
+
+from mmdet.core import (bbox2roi, bbox_mapping, merge_aug_bboxes,
+                        merge_aug_masks, multiclass_nms)
+
+logger = logging.getLogger(__name__)
+
+if sys.version_info >= (3, 7):
+    from mmdet.utils.contextmanagers import completed
+
+
+class BBoxTestMixin(object):
+
+    if sys.version_info >= (3, 7):
+
+        async def async_test_bboxes(self,
+                                    x,
+                                    img_metas,
+                                    proposals,
+                                    rcnn_test_cfg,
+                                    rescale=False,
+                                    bbox_semaphore=None,
+                                    global_lock=None):
+            """Asynchronized test for box head without augmentation."""
+            rois = bbox2roi(proposals)
+            roi_feats = self.bbox_roi_extractor(
+                x[:len(self.bbox_roi_extractor.featmap_strides)], rois)
+            if self.with_shared_head:
+                roi_feats = self.shared_head(roi_feats)
+            sleep_interval = rcnn_test_cfg.get('async_sleep_interval', 0.017)
+
+            async with completed(
+                    __name__, 'bbox_head_forward',
+                    sleep_interval=sleep_interval):
+                cls_score, bbox_pred = self.bbox_head(roi_feats)
+
+            img_shape = img_metas[0]['img_shape']
+            scale_factor = img_metas[0]['scale_factor']
+            det_bboxes, det_labels = self.bbox_head.get_bboxes(
+                rois,
+                cls_score,
+                bbox_pred,
+                img_shape,
+                scale_factor,
+                rescale=rescale,
+                cfg=rcnn_test_cfg)
+            return det_bboxes, det_labels
+
+    def simple_test_bboxes(self,
+                           x,
+                           img_metas,
+                           proposals,
+                           rcnn_test_cfg,
+                           rescale=False):
+        """Test only det bboxes without augmentation."""
+        rois = bbox2roi(proposals)
+        bbox_results = self._bbox_forward(x, rois)
+        img_shapes = tuple(meta['img_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+
+        # split batch bbox prediction back to each image
+        cls_score = bbox_results['cls_score']
+        bbox_pred = bbox_results['bbox_pred']
+        num_proposals_per_img = tuple(len(p) for p in proposals)
+        rois = rois.split(num_proposals_per_img, 0)
+        cls_score = cls_score.split(num_proposals_per_img, 0)
+
+        # some detector with_reg is False, bbox_pred will be None
+        if bbox_pred is not None:
+            # the bbox prediction of some detectors like SABL is not Tensor
+            if isinstance(bbox_pred, torch.Tensor):
+                bbox_pred = bbox_pred.split(num_proposals_per_img, 0)
+            else:
+                bbox_pred = self.bbox_head.bbox_pred_split(
+                    bbox_pred, num_proposals_per_img)
+        else:
+            bbox_pred = (None, ) * len(proposals)
+
+        # apply bbox post-processing to each image individually
+        det_bboxes = []
+        det_labels = []
+        for i in range(len(proposals)):
+            det_bbox, det_label = self.bbox_head.get_bboxes(
+                rois[i],
+                cls_score[i],
+                bbox_pred[i],
+                img_shapes[i],
+                scale_factors[i],
+                rescale=rescale,
+                cfg=rcnn_test_cfg)
+            det_bboxes.append(det_bbox)
+            det_labels.append(det_label)
+        return det_bboxes, det_labels
+
+    def aug_test_bboxes(self, feats, img_metas, proposal_list, rcnn_test_cfg):
+        """Test det bboxes with test time augmentation."""
+        aug_bboxes = []
+        aug_scores = []
+        for x, img_meta in zip(feats, img_metas):
+            # only one image in the batch
+            img_shape = img_meta[0]['img_shape']
+            scale_factor = img_meta[0]['scale_factor']
+            flip = img_meta[0]['flip']
+            flip_direction = img_meta[0]['flip_direction']
+            # TODO more flexible
+            proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
+                                     scale_factor, flip, flip_direction)
+            rois = bbox2roi([proposals])
+            bbox_results = self._bbox_forward(x, rois)
+            bboxes, scores = self.bbox_head.get_bboxes(
+                rois,
+                bbox_results['cls_score'],
+                bbox_results['bbox_pred'],
+                img_shape,
+                scale_factor,
+                rescale=False,
+                cfg=None)
+            aug_bboxes.append(bboxes)
+            aug_scores.append(scores)
+        # after merging, bboxes will be rescaled to the original image size
+        merged_bboxes, merged_scores = merge_aug_bboxes(
+            aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
+        det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
+                                                rcnn_test_cfg.score_thr,
+                                                rcnn_test_cfg.nms,
+                                                rcnn_test_cfg.max_per_img)
+        return det_bboxes, det_labels
+
+
+class MaskTestMixin(object):
+
+    if sys.version_info >= (3, 7):
+
+        async def async_test_mask(self,
+                                  x,
+                                  img_metas,
+                                  det_bboxes,
+                                  det_labels,
+                                  rescale=False,
+                                  mask_test_cfg=None):
+            """Asynchronized test for mask head without augmentation."""
+            # image shape of the first image in the batch (only one)
+            ori_shape = img_metas[0]['ori_shape']
+            scale_factor = img_metas[0]['scale_factor']
+            if det_bboxes.shape[0] == 0:
+                segm_result = [[] for _ in range(self.mask_head.num_classes)]
+            else:
+                _bboxes = (
+                    det_bboxes[:, :4] *
+                    scale_factor if rescale else det_bboxes)
+                mask_rois = bbox2roi([_bboxes])
+                mask_feats = self.mask_roi_extractor(
+                    x[:len(self.mask_roi_extractor.featmap_strides)],
+                    mask_rois)
+
+                if self.with_shared_head:
+                    mask_feats = self.shared_head(mask_feats)
+                if mask_test_cfg and mask_test_cfg.get('async_sleep_interval'):
+                    sleep_interval = mask_test_cfg['async_sleep_interval']
+                else:
+                    sleep_interval = 0.035
+                async with completed(
+                        __name__,
+                        'mask_head_forward',
+                        sleep_interval=sleep_interval):
+                    mask_pred = self.mask_head(mask_feats)
+                segm_result = self.mask_head.get_seg_masks(
+                    mask_pred, _bboxes, det_labels, self.test_cfg, ori_shape,
+                    scale_factor, rescale)
+            return segm_result
+
+    def simple_test_mask(self,
+                         x,
+                         img_metas,
+                         det_bboxes,
+                         det_labels,
+                         rescale=False):
+        """Simple test for mask head without augmentation."""
+        # image shapes of images in the batch
+        ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
+        scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
+        num_imgs = len(det_bboxes)
+        if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
+            segm_results = [[[] for _ in range(self.mask_head.num_classes)]
+                            for _ in range(num_imgs)]
+        else:
+            # if det_bboxes is rescaled to the original image size, we need to
+            # rescale it back to the testing scale to obtain RoIs.
+            if rescale and not isinstance(scale_factors[0], float):
+                scale_factors = [
+                    torch.from_numpy(scale_factor).to(det_bboxes[0].device)
+                    for scale_factor in scale_factors
+                ]
+            if torch.onnx.is_in_onnx_export():
+                # avoid mask_pred.split with static number of prediction
+                mask_preds = []
+                _bboxes = []
+                for i, boxes in enumerate(det_bboxes):
+                    boxes = boxes[:, :4]
+                    if rescale:
+                        boxes *= scale_factors[i]
+                    _bboxes.append(boxes)
+                    img_inds = boxes[:, :1].clone() * 0 + i
+                    mask_rois = torch.cat([img_inds, boxes], dim=-1)
+                    mask_result = self._mask_forward(x, mask_rois)
+                    mask_preds.append(mask_result['mask_pred'])
+            else:
+                _bboxes = [
+                    det_bboxes[i][:, :4] *
+                    scale_factors[i] if rescale else det_bboxes[i][:, :4]
+                    for i in range(len(det_bboxes))
+                ]
+                mask_rois = bbox2roi(_bboxes)
+                mask_results = self._mask_forward(x, mask_rois)
+                mask_pred = mask_results['mask_pred']
+                # split batch mask prediction back to each image
+                num_mask_roi_per_img = [
+                    det_bbox.shape[0] for det_bbox in det_bboxes
+                ]
+                mask_preds = mask_pred.split(num_mask_roi_per_img, 0)
+
+            # apply mask post-processing to each image individually
+            segm_results = []
+            for i in range(num_imgs):
+                if det_bboxes[i].shape[0] == 0:
+                    segm_results.append(
+                        [[] for _ in range(self.mask_head.num_classes)])
+                else:
+                    segm_result = self.mask_head.get_seg_masks(
+                        mask_preds[i], _bboxes[i], det_labels[i],
+                        self.test_cfg, ori_shapes[i], scale_factors[i],
+                        rescale)
+                    segm_results.append(segm_result)
+        return segm_results
+
+    def aug_test_mask(self, feats, img_metas, det_bboxes, det_labels):
+        """Test for mask head with test time augmentation."""
+        if det_bboxes.shape[0] == 0:
+            segm_result = [[] for _ in range(self.mask_head.num_classes)]
+        else:
+            aug_masks = []
+            for x, img_meta in zip(feats, img_metas):
+                img_shape = img_meta[0]['img_shape']
+                scale_factor = img_meta[0]['scale_factor']
+                flip = img_meta[0]['flip']
+                flip_direction = img_meta[0]['flip_direction']
+                _bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
+                                       scale_factor, flip, flip_direction)
+                mask_rois = bbox2roi([_bboxes])
+                mask_results = self._mask_forward(x, mask_rois)
+                # convert to numpy array to save memory
+                aug_masks.append(
+                    mask_results['mask_pred'].sigmoid().cpu().numpy())
+            merged_masks = merge_aug_masks(aug_masks, img_metas, self.test_cfg)
+
+            ori_shape = img_metas[0][0]['ori_shape']
+            segm_result = self.mask_head.get_seg_masks(
+                merged_masks,
+                det_bboxes,
+                det_labels,
+                self.test_cfg,
+                ori_shape,
+                scale_factor=1.0,
+                rescale=False)
+        return segm_result
diff --git a/detection/scrfd/mmdet/models/roi_heads/trident_roi_head.py b/detection/scrfd/mmdet/models/roi_heads/trident_roi_head.py
new file mode 100755
index 000000000..da3eed8fa
--- /dev/null
+++ b/detection/scrfd/mmdet/models/roi_heads/trident_roi_head.py
@@ -0,0 +1,111 @@
+import torch
+from mmcv.ops import batched_nms
+
+from mmdet.core import (bbox2result, bbox2roi, bbox_mapping, merge_aug_bboxes,
+                        multiclass_nms)
+from mmdet.models.roi_heads.standard_roi_head import StandardRoIHead
+from ..builder import HEADS
+
+
+@HEADS.register_module()
+class TridentRoIHead(StandardRoIHead):
+    """Trident roi head.
+
+    Args:
+        num_branch (int): Number of branches in TridentNet.
+        test_branch_idx (int): In inference, all 3 branches will be used
+            if `test_branch_idx==-1`, otherwise only branch with index
+            `test_branch_idx` will be used.
+    """
+
+    def __init__(self, num_branch, test_branch_idx, **kwargs):
+        self.num_branch = num_branch
+        self.test_branch_idx = test_branch_idx
+        super(TridentRoIHead, self).__init__(**kwargs)
+
+    def simple_test(self,
+                    x,
+                    proposal_list,
+                    img_metas,
+                    proposals=None,
+                    rescale=False):
+        """Test without augmentation as follows:
+
+        1. Compute prediction bbox and label per branch.
+        2. Merge predictions of each branch according to scores of
+           bboxes, i.e., bboxes with higher score are kept to give
+           top-k prediction.
+        """
+        assert self.with_bbox, 'Bbox head must be implemented.'
+        det_bboxes_list, det_labels_list = self.simple_test_bboxes(
+            x, img_metas, proposal_list, self.test_cfg, rescale=rescale)
+
+        for _ in range(len(det_bboxes_list)):
+            if det_bboxes_list[_].shape[0] == 0:
+                det_bboxes_list[_] = det_bboxes_list[_].new_empty((0, 5))
+        trident_det_bboxes = torch.cat(det_bboxes_list, 0)
+        trident_det_labels = torch.cat(det_labels_list, 0)
+
+        if trident_det_bboxes.numel() == 0:
+            det_bboxes = trident_det_bboxes.new_zeros((0, 5))
+            det_labels = trident_det_bboxes.new_zeros((0, ), dtype=torch.long)
+        else:
+            nms_bboxes = trident_det_bboxes[:, :4]
+            nms_scores = trident_det_bboxes[:, 4].contiguous()
+            nms_inds = trident_det_labels
+            nms_cfg = self.test_cfg['nms']
+            det_bboxes, keep = batched_nms(nms_bboxes, nms_scores, nms_inds,
+                                           nms_cfg)
+            det_labels = trident_det_labels[keep]
+            if self.test_cfg['max_per_img'] > 0:
+                det_labels = det_labels[:self.test_cfg['max_per_img']]
+                det_bboxes = det_bboxes[:self.test_cfg['max_per_img']]
+
+        det_bboxes, det_labels = [det_bboxes], [det_labels]
+
+        bbox_results = [
+            bbox2result(det_bboxes[i], det_labels[i],
+                        self.bbox_head.num_classes)
+            for i in range(len(det_bboxes))
+        ]
+
+        return bbox_results
+
+    def aug_test_bboxes(self, feats, img_metas, proposal_list, rcnn_test_cfg):
+        """Test det bboxes with test time augmentation."""
+        aug_bboxes = []
+        aug_scores = []
+        for x, img_meta in zip(feats, img_metas):
+            # only one image in the batch
+            img_shape = img_meta[0]['img_shape']
+            scale_factor = img_meta[0]['scale_factor']
+            flip = img_meta[0]['flip']
+            flip_direction = img_meta[0]['flip_direction']
+
+            trident_bboxes, trident_scores = [], []
+            for branch_idx in range(len(proposal_list)):
+                proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
+                                         scale_factor, flip, flip_direction)
+                rois = bbox2roi([proposals])
+                bbox_results = self._bbox_forward(x, rois)
+                bboxes, scores = self.bbox_head.get_bboxes(
+                    rois,
+                    bbox_results['cls_score'],
+                    bbox_results['bbox_pred'],
+                    img_shape,
+                    scale_factor,
+                    rescale=False,
+                    cfg=None)
+                trident_bboxes.append(bboxes)
+                trident_scores.append(scores)
+
+            aug_bboxes.append(torch.cat(trident_bboxes, 0))
+            aug_scores.append(torch.cat(trident_scores, 0))
+        # after merging, bboxes will be rescaled to the original image size
+        merged_bboxes, merged_scores = merge_aug_bboxes(
+            aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
+        det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
+                                                rcnn_test_cfg.score_thr,
+                                                rcnn_test_cfg.nms,
+                                                rcnn_test_cfg.max_per_img)
+        return det_bboxes, det_labels
diff --git a/detection/scrfd/mmdet/models/utils/__init__.py b/detection/scrfd/mmdet/models/utils/__init__.py
new file mode 100755
index 000000000..1fc09b5e8
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/__init__.py
@@ -0,0 +1,16 @@
+from .builder import build_positional_encoding, build_transformer
+from .gaussian_target import gaussian_radius, gen_gaussian_target
+from .positional_encoding import (LearnedPositionalEncoding,
+                                  SinePositionalEncoding)
+from .res_layer import ResLayer
+from .transformer import (FFN, MultiheadAttention, Transformer,
+                          TransformerDecoder, TransformerDecoderLayer,
+                          TransformerEncoder, TransformerEncoderLayer)
+
+__all__ = [
+    'ResLayer', 'gaussian_radius', 'gen_gaussian_target', 'MultiheadAttention',
+    'FFN', 'TransformerEncoderLayer', 'TransformerEncoder',
+    'TransformerDecoderLayer', 'TransformerDecoder', 'Transformer',
+    'build_transformer', 'build_positional_encoding', 'SinePositionalEncoding',
+    'LearnedPositionalEncoding'
+]
diff --git a/detection/scrfd/mmdet/models/utils/builder.py b/detection/scrfd/mmdet/models/utils/builder.py
new file mode 100755
index 000000000..f362d1c92
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/builder.py
@@ -0,0 +1,14 @@
+from mmcv.utils import Registry, build_from_cfg
+
+TRANSFORMER = Registry('Transformer')
+POSITIONAL_ENCODING = Registry('Position encoding')
+
+
+def build_transformer(cfg, default_args=None):
+    """Builder for Transformer."""
+    return build_from_cfg(cfg, TRANSFORMER, default_args)
+
+
+def build_positional_encoding(cfg, default_args=None):
+    """Builder for Position Encoding."""
+    return build_from_cfg(cfg, POSITIONAL_ENCODING, default_args)
diff --git a/detection/scrfd/mmdet/models/utils/gaussian_target.py b/detection/scrfd/mmdet/models/utils/gaussian_target.py
new file mode 100755
index 000000000..7bb7160cb
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/gaussian_target.py
@@ -0,0 +1,185 @@
+from math import sqrt
+
+import torch
+
+
+def gaussian2D(radius, sigma=1, dtype=torch.float32, device='cpu'):
+    """Generate 2D gaussian kernel.
+
+    Args:
+        radius (int): Radius of gaussian kernel.
+        sigma (int): Sigma of gaussian function. Default: 1.
+        dtype (torch.dtype): Dtype of gaussian tensor. Default: torch.float32.
+        device (str): Device of gaussian tensor. Default: 'cpu'.
+
+    Returns:
+        h (Tensor): Gaussian kernel with a
+            ``(2 * radius + 1) * (2 * radius + 1)`` shape.
+    """
+    x = torch.arange(
+        -radius, radius + 1, dtype=dtype, device=device).view(1, -1)
+    y = torch.arange(
+        -radius, radius + 1, dtype=dtype, device=device).view(-1, 1)
+
+    h = (-(x * x + y * y) / (2 * sigma * sigma)).exp()
+
+    h[h < torch.finfo(h.dtype).eps * h.max()] = 0
+    return h
+
+
+def gen_gaussian_target(heatmap, center, radius, k=1):
+    """Generate 2D gaussian heatmap.
+
+    Args:
+        heatmap (Tensor): Input heatmap, the gaussian kernel will cover on
+            it and maintain the max value.
+        center (list[int]): Coord of gaussian kernel's center.
+        radius (int): Radius of gaussian kernel.
+        k (int): Coefficient of gaussian kernel. Default: 1.
+
+    Returns:
+        out_heatmap (Tensor): Updated heatmap covered by gaussian kernel.
+    """
+    diameter = 2 * radius + 1
+    gaussian_kernel = gaussian2D(
+        radius, sigma=diameter / 6, dtype=heatmap.dtype, device=heatmap.device)
+
+    x, y = center
+
+    height, width = heatmap.shape[:2]
+
+    left, right = min(x, radius), min(width - x, radius + 1)
+    top, bottom = min(y, radius), min(height - y, radius + 1)
+
+    masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right]
+    masked_gaussian = gaussian_kernel[radius - top:radius + bottom,
+                                      radius - left:radius + right]
+    out_heatmap = heatmap
+    torch.max(
+        masked_heatmap,
+        masked_gaussian * k,
+        out=out_heatmap[y - top:y + bottom, x - left:x + right])
+
+    return out_heatmap
+
+
+def gaussian_radius(det_size, min_overlap):
+    r"""Generate 2D gaussian radius.
+
+    This function is modified from the `official github repo
+    <https://github.com/princeton-vl/CornerNet-Lite/blob/master/core/sample/
+    utils.py#L65>`_.
+
+    Given ``min_overlap``, radius could computed by a quadratic equation
+    according to Vieta's formulas.
+
+    There are 3 cases for computing gaussian radius, details are following:
+
+    - Explanation of figure: ``lt`` and ``br`` indicates the left-top and
+      bottom-right corner of ground truth box. ``x`` indicates the
+      generated corner at the limited position when ``radius=r``.
+
+    - Case1: one corner is inside the gt box and the other is outside.
+
+    .. code:: text
+
+        |<   width   >|
+
+        lt-+----------+         -
+        |  |          |         ^
+        +--x----------+--+
+        |  |          |  |
+        |  |          |  |    height
+        |  | overlap  |  |
+        |  |          |  |
+        |  |          |  |      v
+        +--+---------br--+      -
+           |          |  |
+           +----------+--x
+
+    To ensure IoU of generated box and gt box is larger than ``min_overlap``:
+
+    .. math::
+        \cfrac{(w-r)*(h-r)}{w*h+(w+h)r-r^2} \ge {iou} \quad\Rightarrow\quad
+        {r^2-(w+h)r+\cfrac{1-iou}{1+iou}*w*h} \ge 0 \\
+        {a} = 1,\quad{b} = {-(w+h)},\quad{c} = {\cfrac{1-iou}{1+iou}*w*h}
+        {r} \le \cfrac{-b-\sqrt{b^2-4*a*c}}{2*a}
+
+    - Case2: both two corners are inside the gt box.
+
+    .. code:: text
+
+        |<   width   >|
+
+        lt-+----------+         -
+        |  |          |         ^
+        +--x-------+  |
+        |  |       |  |
+        |  |overlap|  |       height
+        |  |       |  |
+        |  +-------x--+
+        |          |  |         v
+        +----------+-br         -
+
+    To ensure IoU of generated box and gt box is larger than ``min_overlap``:
+
+    .. math::
+        \cfrac{(w-2*r)*(h-2*r)}{w*h} \ge {iou} \quad\Rightarrow\quad
+        {4r^2-2(w+h)r+(1-iou)*w*h} \ge 0 \\
+        {a} = 4,\quad {b} = {-2(w+h)},\quad {c} = {(1-iou)*w*h}
+        {r} \le \cfrac{-b-\sqrt{b^2-4*a*c}}{2*a}
+
+    - Case3: both two corners are outside the gt box.
+
+    .. code:: text
+
+           |<   width   >|
+
+        x--+----------------+
+        |  |                |
+        +-lt-------------+  |   -
+        |  |             |  |   ^
+        |  |             |  |
+        |  |   overlap   |  | height
+        |  |             |  |
+        |  |             |  |   v
+        |  +------------br--+   -
+        |                |  |
+        +----------------+--x
+
+    To ensure IoU of generated box and gt box is larger than ``min_overlap``:
+
+    .. math::
+        \cfrac{w*h}{(w+2*r)*(h+2*r)} \ge {iou} \quad\Rightarrow\quad
+        {4*iou*r^2+2*iou*(w+h)r+(iou-1)*w*h} \le 0 \\
+        {a} = {4*iou},\quad {b} = {2*iou*(w+h)},\quad {c} = {(iou-1)*w*h} \\
+        {r} \le \cfrac{-b+\sqrt{b^2-4*a*c}}{2*a}
+
+    Args:
+        det_size (list[int]): Shape of object.
+        min_overlap (float): Min IoU with ground truth for boxes generated by
+            keypoints inside the gaussian kernel.
+
+    Returns:
+        radius (int): Radius of gaussian kernel.
+    """
+    height, width = det_size
+
+    a1 = 1
+    b1 = (height + width)
+    c1 = width * height * (1 - min_overlap) / (1 + min_overlap)
+    sq1 = sqrt(b1**2 - 4 * a1 * c1)
+    r1 = (b1 - sq1) / (2 * a1)
+
+    a2 = 4
+    b2 = 2 * (height + width)
+    c2 = (1 - min_overlap) * width * height
+    sq2 = sqrt(b2**2 - 4 * a2 * c2)
+    r2 = (b2 - sq2) / (2 * a2)
+
+    a3 = 4 * min_overlap
+    b3 = -2 * min_overlap * (height + width)
+    c3 = (min_overlap - 1) * width * height
+    sq3 = sqrt(b3**2 - 4 * a3 * c3)
+    r3 = (b3 + sq3) / (2 * a3)
+    return min(r1, r2, r3)
diff --git a/detection/scrfd/mmdet/models/utils/positional_encoding.py b/detection/scrfd/mmdet/models/utils/positional_encoding.py
new file mode 100755
index 000000000..9bda2bbdb
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/positional_encoding.py
@@ -0,0 +1,150 @@
+import math
+
+import torch
+import torch.nn as nn
+from mmcv.cnn import uniform_init
+
+from .builder import POSITIONAL_ENCODING
+
+
+@POSITIONAL_ENCODING.register_module()
+class SinePositionalEncoding(nn.Module):
+    """Position encoding with sine and cosine functions.
+
+    See `End-to-End Object Detection with Transformers
+    <https://arxiv.org/pdf/2005.12872>`_ for details.
+
+    Args:
+        num_feats (int): The feature dimension for each position
+            along x-axis or y-axis. Note the final returned dimension
+            for each position is 2 times of this value.
+        temperature (int, optional): The temperature used for scaling
+            the position embedding. Default 10000.
+        normalize (bool, optional): Whether to normalize the position
+            embedding. Default False.
+        scale (float, optional): A scale factor that scales the position
+            embedding. The scale will be used only when `normalize` is True.
+            Default 2*pi.
+        eps (float, optional): A value added to the denominator for
+            numerical stability. Default 1e-6.
+    """
+
+    def __init__(self,
+                 num_feats,
+                 temperature=10000,
+                 normalize=False,
+                 scale=2 * math.pi,
+                 eps=1e-6):
+        super(SinePositionalEncoding, self).__init__()
+        if normalize:
+            assert isinstance(scale, (float, int)), 'when normalize is set,' \
+                'scale should be provided and in float or int type, ' \
+                f'found {type(scale)}'
+        self.num_feats = num_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        self.scale = scale
+        self.eps = eps
+
+    def forward(self, mask):
+        """Forward function for `SinePositionalEncoding`.
+
+        Args:
+            mask (Tensor): ByteTensor mask. Non-zero values representing
+                ignored positions, while zero values means valid positions
+                for this image. Shape [bs, h, w].
+
+        Returns:
+            pos (Tensor): Returned position embedding with shape
+                [bs, num_feats*2, h, w].
+        """
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            y_embed = y_embed / (y_embed[:, -1:, :] + self.eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + self.eps) * self.scale
+        dim_t = torch.arange(
+            self.num_feats, dtype=torch.float32, device=mask.device)
+        dim_t = self.temperature**(2 * (dim_t // 2) / self.num_feats)
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()),
+            dim=4).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()),
+            dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(num_feats={self.num_feats}, '
+        repr_str += f'temperature={self.temperature}, '
+        repr_str += f'normalize={self.normalize}, '
+        repr_str += f'scale={self.scale}, '
+        repr_str += f'eps={self.eps})'
+        return repr_str
+
+
+@POSITIONAL_ENCODING.register_module()
+class LearnedPositionalEncoding(nn.Module):
+    """Position embedding with learnable embedding weights.
+
+    Args:
+        num_feats (int): The feature dimension for each position
+            along x-axis or y-axis. The final returned dimension for
+            each position is 2 times of this value.
+        row_num_embed (int, optional): The dictionary size of row embeddings.
+            Default 50.
+        col_num_embed (int, optional): The dictionary size of col embeddings.
+            Default 50.
+    """
+
+    def __init__(self, num_feats, row_num_embed=50, col_num_embed=50):
+        super(LearnedPositionalEncoding, self).__init__()
+        self.row_embed = nn.Embedding(row_num_embed, num_feats)
+        self.col_embed = nn.Embedding(col_num_embed, num_feats)
+        self.num_feats = num_feats
+        self.row_num_embed = row_num_embed
+        self.col_num_embed = col_num_embed
+        self.init_weights()
+
+    def init_weights(self):
+        """Initialize the learnable weights."""
+        uniform_init(self.row_embed)
+        uniform_init(self.col_embed)
+
+    def forward(self, mask):
+        """Forward function for `LearnedPositionalEncoding`.
+
+        Args:
+            mask (Tensor): ByteTensor mask. Non-zero values representing
+                ignored positions, while zero values means valid positions
+                for this image. Shape [bs, h, w].
+
+        Returns:
+            pos (Tensor): Returned position embedding with shape
+                [bs, num_feats*2, h, w].
+        """
+        h, w = mask.shape[-2:]
+        x = torch.arange(w, device=mask.device)
+        y = torch.arange(h, device=mask.device)
+        x_embed = self.col_embed(x)
+        y_embed = self.row_embed(y)
+        pos = torch.cat(
+            (x_embed.unsqueeze(0).repeat(h, 1, 1), y_embed.unsqueeze(1).repeat(
+                1, w, 1)),
+            dim=-1).permute(2, 0,
+                            1).unsqueeze(0).repeat(mask.shape[0], 1, 1, 1)
+        return pos
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(num_feats={self.num_feats}, '
+        repr_str += f'row_num_embed={self.row_num_embed}, '
+        repr_str += f'col_num_embed={self.col_num_embed})'
+        return repr_str
diff --git a/detection/scrfd/mmdet/models/utils/res_layer.py b/detection/scrfd/mmdet/models/utils/res_layer.py
new file mode 100755
index 000000000..279024264
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/res_layer.py
@@ -0,0 +1,102 @@
+from mmcv.cnn import build_conv_layer, build_norm_layer
+from torch import nn as nn
+
+
+class ResLayer(nn.Sequential):
+    """ResLayer to build ResNet style backbone.
+
+    Args:
+        block (nn.Module): block used to build ResLayer.
+        inplanes (int): inplanes of block.
+        planes (int): planes of block.
+        num_blocks (int): number of blocks.
+        stride (int): stride of the first block. Default: 1
+        avg_down (bool): Use AvgPool instead of stride conv when
+            downsampling in the bottleneck. Default: False
+        conv_cfg (dict): dictionary to construct and config conv layer.
+            Default: None
+        norm_cfg (dict): dictionary to construct and config norm layer.
+            Default: dict(type='BN')
+        downsample_first (bool): Downsample at the first block or last block.
+            False for Hourglass, True for ResNet. Default: True
+    """
+
+    def __init__(self,
+                 block,
+                 inplanes,
+                 planes,
+                 num_blocks,
+                 stride=1,
+                 avg_down=False,
+                 conv_cfg=None,
+                 norm_cfg=dict(type='BN'),
+                 downsample_first=True,
+                 **kwargs):
+        self.block = block
+
+        downsample = None
+        if stride != 1 or inplanes != planes * block.expansion:
+            downsample = []
+            conv_stride = stride
+            if avg_down:
+                conv_stride = 1
+                downsample.append(
+                    nn.AvgPool2d(
+                        kernel_size=stride,
+                        stride=stride,
+                        ceil_mode=True,
+                        count_include_pad=False))
+            downsample.extend([
+                build_conv_layer(
+                    conv_cfg,
+                    inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=conv_stride,
+                    bias=False),
+                build_norm_layer(norm_cfg, planes * block.expansion)[1]
+            ])
+            downsample = nn.Sequential(*downsample)
+
+        layers = []
+        if downsample_first:
+            layers.append(
+                block(
+                    inplanes=inplanes,
+                    planes=planes,
+                    stride=stride,
+                    downsample=downsample,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    **kwargs))
+            inplanes = planes * block.expansion
+            for _ in range(1, num_blocks):
+                layers.append(
+                    block(
+                        inplanes=inplanes,
+                        planes=planes,
+                        stride=1,
+                        conv_cfg=conv_cfg,
+                        norm_cfg=norm_cfg,
+                        **kwargs))
+
+        else:  # downsample_first=False is for HourglassModule
+            for _ in range(num_blocks - 1):
+                layers.append(
+                    block(
+                        inplanes=inplanes,
+                        planes=inplanes,
+                        stride=1,
+                        conv_cfg=conv_cfg,
+                        norm_cfg=norm_cfg,
+                        **kwargs))
+            layers.append(
+                block(
+                    inplanes=inplanes,
+                    planes=planes,
+                    stride=stride,
+                    downsample=downsample,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    **kwargs))
+        super(ResLayer, self).__init__(*layers)
diff --git a/detection/scrfd/mmdet/models/utils/transformer.py b/detection/scrfd/mmdet/models/utils/transformer.py
new file mode 100755
index 000000000..f94b183b5
--- /dev/null
+++ b/detection/scrfd/mmdet/models/utils/transformer.py
@@ -0,0 +1,744 @@
+import torch
+import torch.nn as nn
+from mmcv.cnn import (Linear, build_activation_layer, build_norm_layer,
+                      xavier_init)
+
+from .builder import TRANSFORMER
+
+
+class MultiheadAttention(nn.Module):
+    """A warpper for torch.nn.MultiheadAttention.
+
+    This module implements MultiheadAttention with residual connection,
+    and positional encoding used in DETR is also passed as input.
+
+    Args:
+        embed_dims (int): The embedding dimension.
+        num_heads (int): Parallel attention heads. Same as
+            `nn.MultiheadAttention`.
+        dropout (float): A Dropout layer on attn_output_weights. Default 0.0.
+    """
+
+    def __init__(self, embed_dims, num_heads, dropout=0.0):
+        super(MultiheadAttention, self).__init__()
+        assert embed_dims % num_heads == 0, 'embed_dims must be ' \
+            f'divisible by num_heads. got {embed_dims} and {num_heads}.'
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.attn = nn.MultiheadAttention(embed_dims, num_heads, dropout)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self,
+                x,
+                key=None,
+                value=None,
+                residual=None,
+                query_pos=None,
+                key_pos=None,
+                attn_mask=None,
+                key_padding_mask=None):
+        """Forward function for `MultiheadAttention`.
+
+        Args:
+            x (Tensor): The input query with shape [num_query, bs,
+                embed_dims]. Same in `nn.MultiheadAttention.forward`.
+            key (Tensor): The key tensor with shape [num_key, bs,
+                embed_dims]. Same in `nn.MultiheadAttention.forward`.
+                Default None. If None, the `query` will be used.
+            value (Tensor): The value tensor with same shape as `key`.
+                Same in `nn.MultiheadAttention.forward`. Default None.
+                If None, the `key` will be used.
+            residual (Tensor): The tensor used for addition, with the
+                same shape as `x`. Default None. If None, `x` will be used.
+            query_pos (Tensor): The positional encoding for query, with
+                the same shape as `x`. Default None. If not None, it will
+                be added to `x` before forward function.
+            key_pos (Tensor): The positional encoding for `key`, with the
+                same shape as `key`. Default None. If not None, it will
+                be added to `key` before forward function. If None, and
+                `query_pos` has the same shape as `key`, then `query_pos`
+                will be used for `key_pos`.
+            attn_mask (Tensor): ByteTensor mask with shape [num_query,
+                num_key]. Same in `nn.MultiheadAttention.forward`.
+                Default None.
+            key_padding_mask (Tensor): ByteTensor with shape [bs, num_key].
+                Same in `nn.MultiheadAttention.forward`. Default None.
+
+        Returns:
+            Tensor: forwarded results with shape [num_query, bs, embed_dims].
+        """
+        query = x
+        if key is None:
+            key = query
+        if value is None:
+            value = key
+        if residual is None:
+            residual = x
+        if key_pos is None:
+            if query_pos is not None and key is not None:
+                if query_pos.shape == key.shape:
+                    key_pos = query_pos
+        if query_pos is not None:
+            query = query + query_pos
+        if key_pos is not None:
+            key = key + key_pos
+        out = self.attn(
+            query,
+            key,
+            value=value,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask)[0]
+
+        return residual + self.dropout(out)
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'dropout={self.dropout})'
+        return repr_str
+
+
+class FFN(nn.Module):
+    """Implements feed-forward networks (FFNs) with residual connection.
+
+    Args:
+        embed_dims (int): The feature dimension. Same as
+            `MultiheadAttention`.
+        feedforward_channels (int): The hidden dimension of FFNs.
+        num_fcs (int): The number of fully-connected layers in FFNs.
+        act_cfg (dict): The activation config for FFNs.
+        dropout (float): Probability of an element to be zeroed. Default 0.0.
+    """
+
+    def __init__(self,
+                 embed_dims,
+                 feedforward_channels,
+                 num_fcs=2,
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 dropout=0.0,
+                 add_residual=True):
+        super(FFN, self).__init__()
+        assert num_fcs >= 2, 'num_fcs should be no less ' \
+            f'than 2. got {num_fcs}.'
+        self.embed_dims = embed_dims
+        self.feedforward_channels = feedforward_channels
+        self.num_fcs = num_fcs
+        self.act_cfg = act_cfg
+        self.dropout = dropout
+        self.activate = build_activation_layer(act_cfg)
+
+        layers = nn.ModuleList()
+        in_channels = embed_dims
+        for _ in range(num_fcs - 1):
+            layers.append(
+                nn.Sequential(
+                    Linear(in_channels, feedforward_channels), self.activate,
+                    nn.Dropout(dropout)))
+            in_channels = feedforward_channels
+        layers.append(Linear(feedforward_channels, embed_dims))
+        self.layers = nn.Sequential(*layers)
+        self.dropout = nn.Dropout(dropout)
+        self.add_residual = add_residual
+
+    def forward(self, x, residual=None):
+        """Forward function for `FFN`."""
+        out = self.layers(x)
+        if not self.add_residual:
+            return out
+        if residual is None:
+            residual = x
+        return residual + self.dropout(out)
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(embed_dims={self.embed_dims}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'num_fcs={self.num_fcs}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'add_residual={self.add_residual})'
+        return repr_str
+
+
+class TransformerEncoderLayer(nn.Module):
+    """Implements one encoder layer in DETR transformer.
+
+    Args:
+        embed_dims (int): The feature dimension. Same as `FFN`.
+        num_heads (int): Parallel attention heads.
+        feedforward_channels (int): The hidden dimension for FFNs.
+        dropout (float): Probability of an element to be zeroed. Default 0.0.
+        order (tuple[str]): The order for encoder layer. Valid examples are
+            ('selfattn', 'norm', 'ffn', 'norm') and ('norm', 'selfattn',
+            'norm', 'ffn'). Default ('selfattn', 'norm', 'ffn', 'norm').
+        act_cfg (dict): The activation config for FFNs. Defalut ReLU.
+        norm_cfg (dict): Config dict for normalization layer. Default
+            layer normalization.
+        num_fcs (int): The number of fully-connected layers for FFNs.
+            Default 2.
+    """
+
+    def __init__(self,
+                 embed_dims,
+                 num_heads,
+                 feedforward_channels,
+                 dropout=0.0,
+                 order=('selfattn', 'norm', 'ffn', 'norm'),
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 norm_cfg=dict(type='LN'),
+                 num_fcs=2):
+        super(TransformerEncoderLayer, self).__init__()
+        assert isinstance(order, tuple) and len(order) == 4
+        assert set(order) == set(['selfattn', 'norm', 'ffn'])
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.feedforward_channels = feedforward_channels
+        self.dropout = dropout
+        self.order = order
+        self.act_cfg = act_cfg
+        self.norm_cfg = norm_cfg
+        self.num_fcs = num_fcs
+        self.pre_norm = order[0] == 'norm'
+        self.self_attn = MultiheadAttention(embed_dims, num_heads, dropout)
+        self.ffn = FFN(embed_dims, feedforward_channels, num_fcs, act_cfg,
+                       dropout)
+        self.norms = nn.ModuleList()
+        self.norms.append(build_norm_layer(norm_cfg, embed_dims)[1])
+        self.norms.append(build_norm_layer(norm_cfg, embed_dims)[1])
+
+    def forward(self, x, pos=None, attn_mask=None, key_padding_mask=None):
+        """Forward function for `TransformerEncoderLayer`.
+
+        Args:
+            x (Tensor): The input query with shape [num_key, bs,
+                embed_dims]. Same in `MultiheadAttention.forward`.
+            pos (Tensor): The positional encoding for query. Default None.
+                Same as `query_pos` in `MultiheadAttention.forward`.
+            attn_mask (Tensor): ByteTensor mask with shape [num_key,
+                num_key]. Same in `MultiheadAttention.forward`. Default None.
+            key_padding_mask (Tensor): ByteTensor with shape [bs, num_key].
+                Same in `MultiheadAttention.forward`. Default None.
+
+        Returns:
+            Tensor: forwarded results with shape [num_key, bs, embed_dims].
+        """
+        norm_cnt = 0
+        inp_residual = x
+        for layer in self.order:
+            if layer == 'selfattn':
+                # self attention
+                query = key = value = x
+                x = self.self_attn(
+                    query,
+                    key,
+                    value,
+                    inp_residual if self.pre_norm else None,
+                    query_pos=pos,
+                    key_pos=pos,
+                    attn_mask=attn_mask,
+                    key_padding_mask=key_padding_mask)
+                inp_residual = x
+            elif layer == 'norm':
+                x = self.norms[norm_cnt](x)
+                norm_cnt += 1
+            elif layer == 'ffn':
+                x = self.ffn(x, inp_residual if self.pre_norm else None)
+        return x
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'order={self.order}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'norm_cfg={self.norm_cfg}, '
+        repr_str += f'num_fcs={self.num_fcs})'
+        return repr_str
+
+
+class TransformerDecoderLayer(nn.Module):
+    """Implements one decoder layer in DETR transformer.
+
+    Args:
+        embed_dims (int): The feature dimension. Same as
+            `TransformerEncoderLayer`.
+        num_heads (int): Parallel attention heads.
+        feedforward_channels (int): Same as `TransformerEncoderLayer`.
+        dropout (float): Same as `TransformerEncoderLayer`. Default 0.0.
+        order (tuple[str]): The order for decoder layer. Valid examples are
+            ('selfattn', 'norm', 'multiheadattn', 'norm', 'ffn', 'norm') and
+            ('norm', 'selfattn', 'norm', 'multiheadattn', 'norm', 'ffn').
+            Default the former.
+        act_cfg (dict): Same as `TransformerEncoderLayer`. Defalut ReLU.
+        norm_cfg (dict): Config dict for normalization layer. Default
+            layer normalization.
+        num_fcs (int): The number of fully-connected layers in FFNs.
+    """
+
+    def __init__(self,
+                 embed_dims,
+                 num_heads,
+                 feedforward_channels,
+                 dropout=0.0,
+                 order=('selfattn', 'norm', 'multiheadattn', 'norm', 'ffn',
+                        'norm'),
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 norm_cfg=dict(type='LN'),
+                 num_fcs=2):
+        super(TransformerDecoderLayer, self).__init__()
+        assert isinstance(order, tuple) and len(order) == 6
+        assert set(order) == set(['selfattn', 'norm', 'multiheadattn', 'ffn'])
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.feedforward_channels = feedforward_channels
+        self.dropout = dropout
+        self.order = order
+        self.act_cfg = act_cfg
+        self.norm_cfg = norm_cfg
+        self.num_fcs = num_fcs
+        self.pre_norm = order[0] == 'norm'
+        self.self_attn = MultiheadAttention(embed_dims, num_heads, dropout)
+        self.multihead_attn = MultiheadAttention(embed_dims, num_heads,
+                                                 dropout)
+        self.ffn = FFN(embed_dims, feedforward_channels, num_fcs, act_cfg,
+                       dropout)
+        self.norms = nn.ModuleList()
+        # 3 norm layers in official DETR's TransformerDecoderLayer
+        for _ in range(3):
+            self.norms.append(build_norm_layer(norm_cfg, embed_dims)[1])
+
+    def forward(self,
+                x,
+                memory,
+                memory_pos=None,
+                query_pos=None,
+                memory_attn_mask=None,
+                target_attn_mask=None,
+                memory_key_padding_mask=None,
+                target_key_padding_mask=None):
+        """Forward function for `TransformerDecoderLayer`.
+
+        Args:
+            x (Tensor): Input query with shape [num_query, bs, embed_dims].
+            memory (Tensor): Tensor got from `TransformerEncoder`, with shape
+                [num_key, bs, embed_dims].
+            memory_pos (Tensor): The positional encoding for `memory`. Default
+                None. Same as `key_pos` in `MultiheadAttention.forward`.
+            query_pos (Tensor): The positional encoding for `query`. Default
+                None. Same as `query_pos` in `MultiheadAttention.forward`.
+            memory_attn_mask (Tensor): ByteTensor mask for `memory`, with
+                shape [num_key, num_key]. Same as `attn_mask` in
+                `MultiheadAttention.forward`. Default None.
+            target_attn_mask (Tensor): ByteTensor mask for `x`, with shape
+                [num_query, num_query]. Same as `attn_mask` in
+                `MultiheadAttention.forward`. Default None.
+            memory_key_padding_mask (Tensor): ByteTensor for `memory`, with
+                shape [bs, num_key]. Same as `key_padding_mask` in
+                `MultiheadAttention.forward`. Default None.
+            target_key_padding_mask (Tensor): ByteTensor for `x`, with shape
+                [bs, num_query]. Same as `key_padding_mask` in
+                `MultiheadAttention.forward`. Default None.
+
+        Returns:
+            Tensor: forwarded results with shape [num_query, bs, embed_dims].
+        """
+        norm_cnt = 0
+        inp_residual = x
+        for layer in self.order:
+            if layer == 'selfattn':
+                query = key = value = x
+                x = self.self_attn(
+                    query,
+                    key,
+                    value,
+                    inp_residual if self.pre_norm else None,
+                    query_pos,
+                    key_pos=query_pos,
+                    attn_mask=target_attn_mask,
+                    key_padding_mask=target_key_padding_mask)
+                inp_residual = x
+            elif layer == 'norm':
+                x = self.norms[norm_cnt](x)
+                norm_cnt += 1
+            elif layer == 'multiheadattn':
+                query = x
+                key = value = memory
+                x = self.multihead_attn(
+                    query,
+                    key,
+                    value,
+                    inp_residual if self.pre_norm else None,
+                    query_pos,
+                    key_pos=memory_pos,
+                    attn_mask=memory_attn_mask,
+                    key_padding_mask=memory_key_padding_mask)
+                inp_residual = x
+            elif layer == 'ffn':
+                x = self.ffn(x, inp_residual if self.pre_norm else None)
+        return x
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'order={self.order}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'norm_cfg={self.norm_cfg}, '
+        repr_str += f'num_fcs={self.num_fcs})'
+        return repr_str
+
+
+class TransformerEncoder(nn.Module):
+    """Implements the encoder in DETR transformer.
+
+    Args:
+        num_layers (int): The number of `TransformerEncoderLayer`.
+        embed_dims (int): Same as `TransformerEncoderLayer`.
+        num_heads (int): Same as `TransformerEncoderLayer`.
+        feedforward_channels (int): Same as `TransformerEncoderLayer`.
+        dropout (float): Same as `TransformerEncoderLayer`. Default 0.0.
+        order (tuple[str]): Same as `TransformerEncoderLayer`.
+        act_cfg (dict): Same as `TransformerEncoderLayer`. Defalut ReLU.
+        norm_cfg (dict): Same as `TransformerEncoderLayer`. Default
+            layer normalization.
+        num_fcs (int): Same as `TransformerEncoderLayer`. Default 2.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 embed_dims,
+                 num_heads,
+                 feedforward_channels,
+                 dropout=0.0,
+                 order=('selfattn', 'norm', 'ffn', 'norm'),
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 norm_cfg=dict(type='LN'),
+                 num_fcs=2):
+        super(TransformerEncoder, self).__init__()
+        assert isinstance(order, tuple) and len(order) == 4
+        assert set(order) == set(['selfattn', 'norm', 'ffn'])
+        self.num_layers = num_layers
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.feedforward_channels = feedforward_channels
+        self.dropout = dropout
+        self.order = order
+        self.act_cfg = act_cfg
+        self.norm_cfg = norm_cfg
+        self.num_fcs = num_fcs
+        self.pre_norm = order[0] == 'norm'
+        self.layers = nn.ModuleList()
+        for _ in range(num_layers):
+            self.layers.append(
+                TransformerEncoderLayer(embed_dims, num_heads,
+                                        feedforward_channels, dropout, order,
+                                        act_cfg, norm_cfg, num_fcs))
+        self.norm = build_norm_layer(norm_cfg,
+                                     embed_dims)[1] if self.pre_norm else None
+
+    def forward(self, x, pos=None, attn_mask=None, key_padding_mask=None):
+        """Forward function for `TransformerEncoder`.
+
+        Args:
+            x (Tensor): Input query. Same in `TransformerEncoderLayer.forward`.
+            pos (Tensor): Positional encoding for query. Default None.
+                Same in `TransformerEncoderLayer.forward`.
+            attn_mask (Tensor): ByteTensor attention mask. Default None.
+                Same in `TransformerEncoderLayer.forward`.
+            key_padding_mask (Tensor): Same in
+                `TransformerEncoderLayer.forward`. Default None.
+
+        Returns:
+            Tensor: Results with shape [num_key, bs, embed_dims].
+        """
+        for layer in self.layers:
+            x = layer(x, pos, attn_mask, key_padding_mask)
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(num_layers={self.num_layers}, '
+        repr_str += f'embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'order={self.order}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'norm_cfg={self.norm_cfg}, '
+        repr_str += f'num_fcs={self.num_fcs})'
+        return repr_str
+
+
+class TransformerDecoder(nn.Module):
+    """Implements the decoder in DETR transformer.
+
+    Args:
+        num_layers (int): The number of `TransformerDecoderLayer`.
+        embed_dims (int): Same as `TransformerDecoderLayer`.
+        num_heads (int): Same as `TransformerDecoderLayer`.
+        feedforward_channels (int): Same as `TransformerDecoderLayer`.
+        dropout (float): Same as `TransformerDecoderLayer`. Default 0.0.
+        order (tuple[str]): Same as `TransformerDecoderLayer`.
+        act_cfg (dict): Same as `TransformerDecoderLayer`. Defalut ReLU.
+        norm_cfg (dict): Same as `TransformerDecoderLayer`. Default
+            layer normalization.
+        num_fcs (int): Same as `TransformerDecoderLayer`. Default 2.
+    """
+
+    def __init__(self,
+                 num_layers,
+                 embed_dims,
+                 num_heads,
+                 feedforward_channels,
+                 dropout=0.0,
+                 order=('selfattn', 'norm', 'multiheadattn', 'norm', 'ffn',
+                        'norm'),
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 norm_cfg=dict(type='LN'),
+                 num_fcs=2,
+                 return_intermediate=False):
+        super(TransformerDecoder, self).__init__()
+        assert isinstance(order, tuple) and len(order) == 6
+        assert set(order) == set(['selfattn', 'norm', 'multiheadattn', 'ffn'])
+        self.num_layers = num_layers
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.feedforward_channels = feedforward_channels
+        self.dropout = dropout
+        self.order = order
+        self.act_cfg = act_cfg
+        self.norm_cfg = norm_cfg
+        self.num_fcs = num_fcs
+        self.return_intermediate = return_intermediate
+        self.layers = nn.ModuleList()
+        for _ in range(num_layers):
+            self.layers.append(
+                TransformerDecoderLayer(embed_dims, num_heads,
+                                        feedforward_channels, dropout, order,
+                                        act_cfg, norm_cfg, num_fcs))
+        self.norm = build_norm_layer(norm_cfg, embed_dims)[1]
+
+    def forward(self,
+                x,
+                memory,
+                memory_pos=None,
+                query_pos=None,
+                memory_attn_mask=None,
+                target_attn_mask=None,
+                memory_key_padding_mask=None,
+                target_key_padding_mask=None):
+        """Forward function for `TransformerDecoder`.
+
+        Args:
+            x (Tensor): Input query. Same in `TransformerDecoderLayer.forward`.
+            memory (Tensor): Same in `TransformerDecoderLayer.forward`.
+            memory_pos (Tensor): Same in `TransformerDecoderLayer.forward`.
+                Default None.
+            query_pos (Tensor): Same in `TransformerDecoderLayer.forward`.
+                Default None.
+            memory_attn_mask (Tensor): Same in
+                `TransformerDecoderLayer.forward`. Default None.
+            target_attn_mask (Tensor): Same in
+                `TransformerDecoderLayer.forward`. Default None.
+            memory_key_padding_mask (Tensor): Same in
+                `TransformerDecoderLayer.forward`. Default None.
+            target_key_padding_mask (Tensor): Same in
+                `TransformerDecoderLayer.forward`. Default None.
+
+        Returns:
+            Tensor: Results with shape [num_query, bs, embed_dims].
+        """
+        intermediate = []
+        for layer in self.layers:
+            x = layer(x, memory, memory_pos, query_pos, memory_attn_mask,
+                      target_attn_mask, memory_key_padding_mask,
+                      target_key_padding_mask)
+            if self.return_intermediate:
+                intermediate.append(self.norm(x))
+        if self.norm is not None:
+            x = self.norm(x)
+            if self.return_intermediate:
+                intermediate.pop()
+                intermediate.append(x)
+        if self.return_intermediate:
+            return torch.stack(intermediate)
+        return x.unsqueeze(0)
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(num_layers={self.num_layers}, '
+        repr_str += f'embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'order={self.order}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'norm_cfg={self.norm_cfg}, '
+        repr_str += f'num_fcs={self.num_fcs}, '
+        repr_str += f'return_intermediate={self.return_intermediate})'
+        return repr_str
+
+
+@TRANSFORMER.register_module()
+class Transformer(nn.Module):
+    """Implements the DETR transformer.
+
+    Following the official DETR implementation, this module copy-paste
+    from torch.nn.Transformer with modifications:
+
+        * positional encodings are passed in MultiheadAttention
+        * extra LN at the end of encoder is removed
+        * decoder returns a stack of activations from all decoding layers
+
+    See `paper: End-to-End Object Detection with Transformers
+    <https://arxiv.org/pdf/2005.12872>`_ for details.
+
+    Args:
+        embed_dims (int): The feature dimension.
+        num_heads (int): Parallel attention heads. Same as
+            `nn.MultiheadAttention`.
+        num_encoder_layers (int): Number of `TransformerEncoderLayer`.
+        num_decoder_layers (int): Number of `TransformerDecoderLayer`.
+        feedforward_channels (int): The hidden dimension for FFNs used in both
+            encoder and decoder.
+        dropout (float): Probability of an element to be zeroed. Default 0.0.
+        act_cfg (dict): Activation config for FFNs used in both encoder
+            and decoder. Defalut ReLU.
+        norm_cfg (dict): Config dict for normalization used in both encoder
+            and decoder. Default layer normalization.
+        num_fcs (int): The number of fully-connected layers in FFNs, which is
+            used for both encoder and decoder.
+        pre_norm (bool): Whether the normalization layer is ordered
+            first in the encoder and decoder. Default False.
+        return_intermediate_dec (bool): Whether to return the intermediate
+            output from each TransformerDecoderLayer or only the last
+            TransformerDecoderLayer. Default False. If False, the returned
+            `hs` has shape [num_decoder_layers, bs, num_query, embed_dims].
+            If True, the returned `hs` will have shape [1, bs, num_query,
+            embed_dims].
+    """
+
+    def __init__(self,
+                 embed_dims=512,
+                 num_heads=8,
+                 num_encoder_layers=6,
+                 num_decoder_layers=6,
+                 feedforward_channels=2048,
+                 dropout=0.0,
+                 act_cfg=dict(type='ReLU', inplace=True),
+                 norm_cfg=dict(type='LN'),
+                 num_fcs=2,
+                 pre_norm=False,
+                 return_intermediate_dec=False):
+        super(Transformer, self).__init__()
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.num_encoder_layers = num_encoder_layers
+        self.num_decoder_layers = num_decoder_layers
+        self.feedforward_channels = feedforward_channels
+        self.dropout = dropout
+        self.act_cfg = act_cfg
+        self.norm_cfg = norm_cfg
+        self.num_fcs = num_fcs
+        self.pre_norm = pre_norm
+        self.return_intermediate_dec = return_intermediate_dec
+        if self.pre_norm:
+            encoder_order = ('norm', 'selfattn', 'norm', 'ffn')
+            decoder_order = ('norm', 'selfattn', 'norm', 'multiheadattn',
+                             'norm', 'ffn')
+        else:
+            encoder_order = ('selfattn', 'norm', 'ffn', 'norm')
+            decoder_order = ('selfattn', 'norm', 'multiheadattn', 'norm',
+                             'ffn', 'norm')
+        self.encoder = TransformerEncoder(num_encoder_layers, embed_dims,
+                                          num_heads, feedforward_channels,
+                                          dropout, encoder_order, act_cfg,
+                                          norm_cfg, num_fcs)
+        self.decoder = TransformerDecoder(num_decoder_layers, embed_dims,
+                                          num_heads, feedforward_channels,
+                                          dropout, decoder_order, act_cfg,
+                                          norm_cfg, num_fcs,
+                                          return_intermediate_dec)
+
+    def init_weights(self, distribution='uniform'):
+        """Initialize the transformer weights."""
+        # follow the official DETR to init parameters
+        for m in self.modules():
+            if hasattr(m, 'weight') and m.weight.dim() > 1:
+                xavier_init(m, distribution=distribution)
+
+    def forward(self, x, mask, query_embed, pos_embed):
+        """Forward function for `Transformer`.
+
+        Args:
+            x (Tensor): Input query with shape [bs, c, h, w] where
+                c = embed_dims.
+            mask (Tensor): The key_padding_mask used for encoder and decoder,
+                with shape [bs, h, w].
+            query_embed (Tensor): The query embedding for decoder, with shape
+                [num_query, c].
+            pos_embed (Tensor): The positional encoding for encoder and
+                decoder, with the same shape as `x`.
+
+        Returns:
+            tuple[Tensor]: results of decoder containing the following tensor.
+
+                - out_dec: Output from decoder. If return_intermediate_dec \
+                      is True output has shape [num_dec_layers, bs,
+                      num_query, embed_dims], else has shape [1, bs, \
+                      num_query, embed_dims].
+                - memory: Output results from encoder, with shape \
+                      [bs, embed_dims, h, w].
+        """
+        bs, c, h, w = x.shape
+        x = x.flatten(2).permute(2, 0, 1)  # [bs, c, h, w] -> [h*w, bs, c]
+        pos_embed = pos_embed.flatten(2).permute(2, 0, 1)
+        query_embed = query_embed.unsqueeze(1).repeat(
+            1, bs, 1)  # [num_query, dim] -> [num_query, bs, dim]
+        mask = mask.flatten(1)  # [bs, h, w] -> [bs, h*w]
+        memory = self.encoder(
+            x, pos=pos_embed, attn_mask=None, key_padding_mask=mask)
+        target = torch.zeros_like(query_embed)
+        # out_dec: [num_layers, num_query, bs, dim]
+        out_dec = self.decoder(
+            target,
+            memory,
+            memory_pos=pos_embed,
+            query_pos=query_embed,
+            memory_attn_mask=None,
+            target_attn_mask=None,
+            memory_key_padding_mask=mask,
+            target_key_padding_mask=None)
+        out_dec = out_dec.transpose(1, 2)
+        memory = memory.permute(1, 2, 0).reshape(bs, c, h, w)
+        return out_dec, memory
+
+    def __repr__(self):
+        """str: a string that describes the module"""
+        repr_str = self.__class__.__name__
+        repr_str += f'(embed_dims={self.embed_dims}, '
+        repr_str += f'num_heads={self.num_heads}, '
+        repr_str += f'num_encoder_layers={self.num_encoder_layers}, '
+        repr_str += f'num_decoder_layers={self.num_decoder_layers}, '
+        repr_str += f'feedforward_channels={self.feedforward_channels}, '
+        repr_str += f'dropout={self.dropout}, '
+        repr_str += f'act_cfg={self.act_cfg}, '
+        repr_str += f'norm_cfg={self.norm_cfg}, '
+        repr_str += f'num_fcs={self.num_fcs}, '
+        repr_str += f'pre_norm={self.pre_norm}, '
+        repr_str += f'return_intermediate_dec={self.return_intermediate_dec})'
+        return repr_str
diff --git a/detection/scrfd/mmdet/ops/__init__.py b/detection/scrfd/mmdet/ops/__init__.py
new file mode 100755
index 000000000..ae6a83687
--- /dev/null
+++ b/detection/scrfd/mmdet/ops/__init__.py
@@ -0,0 +1,32 @@
+# This file is added for back-compatibility. Thus, downstream codebase
+# could still use and import mmdet.ops.
+
+# yapf: disable
+from mmcv.ops import (ContextBlock, Conv2d, ConvTranspose2d, ConvWS2d,
+                      CornerPool, DeformConv, DeformConvPack, DeformRoIPooling,
+                      DeformRoIPoolingPack, GeneralizedAttention, Linear,
+                      MaskedConv2d, MaxPool2d, ModulatedDeformConv,
+                      ModulatedDeformConvPack, ModulatedDeformRoIPoolingPack,
+                      NonLocal2D, RoIAlign, RoIPool, SAConv2d,
+                      SigmoidFocalLoss, SimpleRoIAlign, batched_nms,
+                      build_plugin_layer, conv_ws_2d, deform_conv,
+                      deform_roi_pooling, get_compiler_version,
+                      get_compiling_cuda_version, modulated_deform_conv, nms,
+                      nms_match, point_sample, rel_roi_point_to_rel_img_point,
+                      roi_align, roi_pool, sigmoid_focal_loss, soft_nms)
+
+# yapf: enable
+
+__all__ = [
+    'nms', 'soft_nms', 'RoIAlign', 'roi_align', 'RoIPool', 'roi_pool',
+    'DeformConv', 'DeformConvPack', 'DeformRoIPooling', 'DeformRoIPoolingPack',
+    'ModulatedDeformRoIPoolingPack', 'ModulatedDeformConv',
+    'ModulatedDeformConvPack', 'deform_conv', 'modulated_deform_conv',
+    'deform_roi_pooling', 'SigmoidFocalLoss', 'sigmoid_focal_loss',
+    'MaskedConv2d', 'ContextBlock', 'GeneralizedAttention', 'NonLocal2D',
+    'get_compiler_version', 'get_compiling_cuda_version', 'ConvWS2d',
+    'conv_ws_2d', 'build_plugin_layer', 'batched_nms', 'Conv2d',
+    'ConvTranspose2d', 'MaxPool2d', 'Linear', 'nms_match', 'CornerPool',
+    'point_sample', 'rel_roi_point_to_rel_img_point', 'SimpleRoIAlign',
+    'SAConv2d'
+]
diff --git a/detection/scrfd/mmdet/utils/__init__.py b/detection/scrfd/mmdet/utils/__init__.py
new file mode 100755
index 000000000..ac489e2db
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/__init__.py
@@ -0,0 +1,4 @@
+from .collect_env import collect_env
+from .logger import get_root_logger
+
+__all__ = ['get_root_logger', 'collect_env']
diff --git a/detection/scrfd/mmdet/utils/collect_env.py b/detection/scrfd/mmdet/utils/collect_env.py
new file mode 100755
index 000000000..89c064acc
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/collect_env.py
@@ -0,0 +1,16 @@
+from mmcv.utils import collect_env as collect_base_env
+from mmcv.utils import get_git_hash
+
+import mmdet
+
+
+def collect_env():
+    """Collect the information of the running environments."""
+    env_info = collect_base_env()
+    env_info['MMDetection'] = mmdet.__version__ + '+' + get_git_hash()[:7]
+    return env_info
+
+
+if __name__ == '__main__':
+    for name, val in collect_env().items():
+        print(f'{name}: {val}')
diff --git a/detection/scrfd/mmdet/utils/contextmanagers.py b/detection/scrfd/mmdet/utils/contextmanagers.py
new file mode 100755
index 000000000..38a639262
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/contextmanagers.py
@@ -0,0 +1,121 @@
+import asyncio
+import contextlib
+import logging
+import os
+import time
+from typing import List
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+DEBUG_COMPLETED_TIME = bool(os.environ.get('DEBUG_COMPLETED_TIME', False))
+
+
+@contextlib.asynccontextmanager
+async def completed(trace_name='',
+                    name='',
+                    sleep_interval=0.05,
+                    streams: List[torch.cuda.Stream] = None):
+    """Async context manager that waits for work to complete on given CUDA
+    streams."""
+    if not torch.cuda.is_available():
+        yield
+        return
+
+    stream_before_context_switch = torch.cuda.current_stream()
+    if not streams:
+        streams = [stream_before_context_switch]
+    else:
+        streams = [s if s else stream_before_context_switch for s in streams]
+
+    end_events = [
+        torch.cuda.Event(enable_timing=DEBUG_COMPLETED_TIME) for _ in streams
+    ]
+
+    if DEBUG_COMPLETED_TIME:
+        start = torch.cuda.Event(enable_timing=True)
+        stream_before_context_switch.record_event(start)
+
+        cpu_start = time.monotonic()
+    logger.debug('%s %s starting, streams: %s', trace_name, name, streams)
+    grad_enabled_before = torch.is_grad_enabled()
+    try:
+        yield
+    finally:
+        current_stream = torch.cuda.current_stream()
+        assert current_stream == stream_before_context_switch
+
+        if DEBUG_COMPLETED_TIME:
+            cpu_end = time.monotonic()
+        for i, stream in enumerate(streams):
+            event = end_events[i]
+            stream.record_event(event)
+
+        grad_enabled_after = torch.is_grad_enabled()
+
+        # observed change of torch.is_grad_enabled() during concurrent run of
+        # async_test_bboxes code
+        assert (grad_enabled_before == grad_enabled_after
+                ), 'Unexpected is_grad_enabled() value change'
+
+        are_done = [e.query() for e in end_events]
+        logger.debug('%s %s completed: %s streams: %s', trace_name, name,
+                     are_done, streams)
+        with torch.cuda.stream(stream_before_context_switch):
+            while not all(are_done):
+                await asyncio.sleep(sleep_interval)
+                are_done = [e.query() for e in end_events]
+                logger.debug(
+                    '%s %s completed: %s streams: %s',
+                    trace_name,
+                    name,
+                    are_done,
+                    streams,
+                )
+
+        current_stream = torch.cuda.current_stream()
+        assert current_stream == stream_before_context_switch
+
+        if DEBUG_COMPLETED_TIME:
+            cpu_time = (cpu_end - cpu_start) * 1000
+            stream_times_ms = ''
+            for i, stream in enumerate(streams):
+                elapsed_time = start.elapsed_time(end_events[i])
+                stream_times_ms += f' {stream} {elapsed_time:.2f} ms'
+            logger.info('%s %s %.2f ms %s', trace_name, name, cpu_time,
+                        stream_times_ms)
+
+
+@contextlib.asynccontextmanager
+async def concurrent(streamqueue: asyncio.Queue,
+                     trace_name='concurrent',
+                     name='stream'):
+    """Run code concurrently in different streams.
+
+    :param streamqueue: asyncio.Queue instance.
+
+    Queue tasks define the pool of streams used for concurrent execution.
+    """
+    if not torch.cuda.is_available():
+        yield
+        return
+
+    initial_stream = torch.cuda.current_stream()
+
+    with torch.cuda.stream(initial_stream):
+        stream = await streamqueue.get()
+        assert isinstance(stream, torch.cuda.Stream)
+
+        try:
+            with torch.cuda.stream(stream):
+                logger.debug('%s %s is starting, stream: %s', trace_name, name,
+                             stream)
+                yield
+                current = torch.cuda.current_stream()
+                assert current == stream
+                logger.debug('%s %s has finished, stream: %s', trace_name,
+                             name, stream)
+        finally:
+            streamqueue.task_done()
+            streamqueue.put_nowait(stream)
diff --git a/detection/scrfd/mmdet/utils/logger.py b/detection/scrfd/mmdet/utils/logger.py
new file mode 100755
index 000000000..6fc6e6b43
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/logger.py
@@ -0,0 +1,19 @@
+import logging
+
+from mmcv.utils import get_logger
+
+
+def get_root_logger(log_file=None, log_level=logging.INFO):
+    """Get root logger.
+
+    Args:
+        log_file (str, optional): File path of log. Defaults to None.
+        log_level (int, optional): The level of logger.
+            Defaults to logging.INFO.
+
+    Returns:
+        :obj:`logging.Logger`: The obtained logger
+    """
+    logger = get_logger(name='mmdet', log_file=log_file, log_level=log_level)
+
+    return logger
diff --git a/detection/scrfd/mmdet/utils/profiling.py b/detection/scrfd/mmdet/utils/profiling.py
new file mode 100755
index 000000000..4be9222c3
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/profiling.py
@@ -0,0 +1,39 @@
+import contextlib
+import sys
+import time
+
+import torch
+
+if sys.version_info >= (3, 7):
+
+    @contextlib.contextmanager
+    def profile_time(trace_name,
+                     name,
+                     enabled=True,
+                     stream=None,
+                     end_stream=None):
+        """Print time spent by CPU and GPU.
+
+        Useful as a temporary context manager to find sweet spots of code
+        suitable for async implementation.
+        """
+        if (not enabled) or not torch.cuda.is_available():
+            yield
+            return
+        stream = stream if stream else torch.cuda.current_stream()
+        end_stream = end_stream if end_stream else stream
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        stream.record_event(start)
+        try:
+            cpu_start = time.monotonic()
+            yield
+        finally:
+            cpu_end = time.monotonic()
+            end_stream.record_event(end)
+            end.synchronize()
+            cpu_time = (cpu_end - cpu_start) * 1000
+            gpu_time = start.elapsed_time(end)
+            msg = f'{trace_name} {name} cpu_time {cpu_time:.2f} ms '
+            msg += f'gpu_time {gpu_time:.2f} ms stream {stream}'
+            print(msg, end_stream)
diff --git a/detection/scrfd/mmdet/utils/util_mixins.py b/detection/scrfd/mmdet/utils/util_mixins.py
new file mode 100755
index 000000000..69669a3ca
--- /dev/null
+++ b/detection/scrfd/mmdet/utils/util_mixins.py
@@ -0,0 +1,104 @@
+"""This module defines the :class:`NiceRepr` mixin class, which defines a
+``__repr__`` and ``__str__`` method that only depend on a custom ``__nice__``
+method, which you must define. This means you only have to overload one
+function instead of two.  Furthermore, if the object defines a ``__len__``
+method, then the ``__nice__`` method defaults to something sensible, otherwise
+it is treated as abstract and raises ``NotImplementedError``.
+
+To use simply have your object inherit from :class:`NiceRepr`
+(multi-inheritance should be ok).
+
+This code was copied from the ubelt library: https://github.com/Erotemic/ubelt
+
+Example:
+    >>> # Objects that define __nice__ have a default __str__ and __repr__
+    >>> class Student(NiceRepr):
+    ...    def __init__(self, name):
+    ...        self.name = name
+    ...    def __nice__(self):
+    ...        return self.name
+    >>> s1 = Student('Alice')
+    >>> s2 = Student('Bob')
+    >>> print(f's1 = {s1}')
+    >>> print(f's2 = {s2}')
+    s1 = <Student(Alice)>
+    s2 = <Student(Bob)>
+
+Example:
+    >>> # Objects that define __len__ have a default __nice__
+    >>> class Group(NiceRepr):
+    ...    def __init__(self, data):
+    ...        self.data = data
+    ...    def __len__(self):
+    ...        return len(self.data)
+    >>> g = Group([1, 2, 3])
+    >>> print(f'g = {g}')
+    g = <Group(3)>
+"""
+import warnings
+
+
+class NiceRepr(object):
+    """Inherit from this class and define ``__nice__`` to "nicely" print your
+    objects.
+
+    Defines ``__str__`` and ``__repr__`` in terms of ``__nice__`` function
+    Classes that inherit from :class:`NiceRepr` should redefine ``__nice__``.
+    If the inheriting class has a ``__len__``, method then the default
+    ``__nice__`` method will return its length.
+
+    Example:
+        >>> class Foo(NiceRepr):
+        ...    def __nice__(self):
+        ...        return 'info'
+        >>> foo = Foo()
+        >>> assert str(foo) == '<Foo(info)>'
+        >>> assert repr(foo).startswith('<Foo(info) at ')
+
+    Example:
+        >>> class Bar(NiceRepr):
+        ...    pass
+        >>> bar = Bar()
+        >>> import pytest
+        >>> with pytest.warns(None) as record:
+        >>>     assert 'object at' in str(bar)
+        >>>     assert 'object at' in repr(bar)
+
+    Example:
+        >>> class Baz(NiceRepr):
+        ...    def __len__(self):
+        ...        return 5
+        >>> baz = Baz()
+        >>> assert str(baz) == '<Baz(5)>'
+    """
+
+    def __nice__(self):
+        """str: a "nice" summary string describing this module"""
+        if hasattr(self, '__len__'):
+            # It is a common pattern for objects to use __len__ in __nice__
+            # As a convenience we define a default __nice__ for these objects
+            return str(len(self))
+        else:
+            # In all other cases force the subclass to overload __nice__
+            raise NotImplementedError(
+                f'Define the __nice__ method for {self.__class__!r}')
+
+    def __repr__(self):
+        """str: the string of the module"""
+        try:
+            nice = self.__nice__()
+            classname = self.__class__.__name__
+            return f'<{classname}({nice}) at {hex(id(self))}>'
+        except NotImplementedError as ex:
+            warnings.warn(str(ex), category=RuntimeWarning)
+            return object.__repr__(self)
+
+    def __str__(self):
+        """str: the string of the module"""
+        try:
+            classname = self.__class__.__name__
+            nice = self.__nice__()
+            return f'<{classname}({nice})>'
+        except NotImplementedError as ex:
+            warnings.warn(str(ex), category=RuntimeWarning)
+            return object.__repr__(self)
diff --git a/detection/scrfd/mmdet/version.py b/detection/scrfd/mmdet/version.py
new file mode 100755
index 000000000..753ab3c29
--- /dev/null
+++ b/detection/scrfd/mmdet/version.py
@@ -0,0 +1,19 @@
+# Copyright (c) Open-MMLab. All rights reserved.
+
+__version__ = '2.7.0'
+short_version = __version__
+
+
+def parse_version_info(version_str):
+    version_info = []
+    for x in version_str.split('.'):
+        if x.isdigit():
+            version_info.append(int(x))
+        elif x.find('rc') != -1:
+            patch_version = x.split('rc')
+            version_info.append(int(patch_version[0]))
+            version_info.append(f'rc{patch_version[1]}')
+    return tuple(version_info)
+
+
+version_info = parse_version_info(__version__)
diff --git a/detection/scrfd/pytest.ini b/detection/scrfd/pytest.ini
new file mode 100755
index 000000000..9796e871e
--- /dev/null
+++ b/detection/scrfd/pytest.ini
@@ -0,0 +1,7 @@
+[pytest]
+addopts = --xdoctest --xdoctest-style=auto
+norecursedirs = .git ignore build __pycache__ data docker docs .eggs
+
+filterwarnings= default
+                ignore:.*No cfgstr given in Cacher constructor or call.*:Warning
+                ignore:.*Define the __nice__ method for.*:Warning
diff --git a/detection/scrfd/requirements.txt b/detection/scrfd/requirements.txt
new file mode 100755
index 000000000..6981bd723
--- /dev/null
+++ b/detection/scrfd/requirements.txt
@@ -0,0 +1,4 @@
+-r requirements/build.txt
+-r requirements/optional.txt
+-r requirements/runtime.txt
+-r requirements/tests.txt
diff --git a/detection/scrfd/requirements/build.txt b/detection/scrfd/requirements/build.txt
new file mode 100755
index 000000000..815582985
--- /dev/null
+++ b/detection/scrfd/requirements/build.txt
@@ -0,0 +1,3 @@
+# These must be installed before building mmdetection
+cython
+numpy
diff --git a/detection/scrfd/requirements/docs.txt b/detection/scrfd/requirements/docs.txt
new file mode 100755
index 000000000..89fbf86c0
--- /dev/null
+++ b/detection/scrfd/requirements/docs.txt
@@ -0,0 +1,4 @@
+recommonmark
+sphinx
+sphinx_markdown_tables
+sphinx_rtd_theme
diff --git a/detection/scrfd/requirements/optional.txt b/detection/scrfd/requirements/optional.txt
new file mode 100755
index 000000000..6654b5b96
--- /dev/null
+++ b/detection/scrfd/requirements/optional.txt
@@ -0,0 +1,4 @@
+albumentations>=0.3.2
+cityscapesscripts
+imagecorruptions
+mmlvis
diff --git a/detection/scrfd/requirements/readthedocs.txt b/detection/scrfd/requirements/readthedocs.txt
new file mode 100755
index 000000000..0542bfce6
--- /dev/null
+++ b/detection/scrfd/requirements/readthedocs.txt
@@ -0,0 +1,3 @@
+mmcv
+torch
+torchvision
diff --git a/detection/scrfd/requirements/runtime.txt b/detection/scrfd/requirements/runtime.txt
new file mode 100755
index 000000000..f559b9612
--- /dev/null
+++ b/detection/scrfd/requirements/runtime.txt
@@ -0,0 +1,7 @@
+matplotlib
+mmpycocotools
+numpy
+six
+terminaltables
+tqdm
+autotorch
diff --git a/detection/scrfd/requirements/tests.txt b/detection/scrfd/requirements/tests.txt
new file mode 100755
index 000000000..acd94d38b
--- /dev/null
+++ b/detection/scrfd/requirements/tests.txt
@@ -0,0 +1,12 @@
+asynctest
+codecov
+flake8
+interrogate
+isort==4.3.21
+# Note: used for kwarray.group_items, this may be ported to mmcv in the future.
+kwarray
+pytest
+ubelt
+xdoctest>=0.10.0
+yapf
+scipy
diff --git a/detection/scrfd/search_tools/generate_configs_2.5g.py b/detection/scrfd/search_tools/generate_configs_2.5g.py
new file mode 100755
index 000000000..e118a7f9a
--- /dev/null
+++ b/detection/scrfd/search_tools/generate_configs_2.5g.py
@@ -0,0 +1,258 @@
+
+import os
+import os.path as osp
+import io
+import numpy as np
+import argparse
+import datetime
+import importlib
+import configparser
+from tqdm import tqdm
+from mmdet.models import build_detector
+
+import torch
+import autotorch as at
+from mmcv import Config
+
+from mmdet.models import build_detector
+
+try:
+    from mmcv.cnn import get_model_complexity_info
+except ImportError:
+    raise ImportError('Please upgrade mmcv to >0.6.2')
+
+
+@at.obj(
+    block=at.Choice('BasicBlock', 'Bottleneck'),
+    base_channels=at.Int(8, 64),
+    stage_blocks=at.List(
+        at.Int(1,10),
+        at.Int(1,10),
+        at.Int(1,10),
+        at.Int(1,10),
+        ),
+    stage_planes_ratio=at.List(
+        at.Real(1.0,4.0),
+        at.Real(1.0,4.0),
+        at.Real(1.0,4.0),
+        ),
+)
+class GenConfigBackbone:
+    def __init__(self, **kwargs):
+        d = {}
+        d.update(**kwargs)
+        for k, v in d.items():
+            setattr(self, k, v)
+        self.m = 1.0
+
+    def stage_blocks_multi(self, m):
+        self.m = m
+
+    def merge_cfg(self, det_cfg):
+
+        base_channels = max(8, int(self.base_channels*self.m)//8 * 8)
+        stage_planes = [base_channels]
+        for ratio in self.stage_planes_ratio:
+            planes = int(stage_planes[-1] * ratio) //8 * 8
+            stage_planes.append(planes)
+        stage_blocks = [max(1, int(x*self.m)) for x in self.stage_blocks]
+        #print('Blocks:', stage_blocks)
+        #print('Planes:', stage_planes)
+        block_cfg=dict(block=self.block, stage_blocks=tuple(stage_blocks), stage_planes=stage_planes)
+        det_cfg['model']['backbone']['block_cfg'] = block_cfg
+        det_cfg['model']['backbone']['base_channels'] = base_channels
+        neck_in_planes = stage_planes if self.block=='BasicBlock' else [4*x for x in stage_planes]
+        det_cfg['model']['neck']['in_channels'] = neck_in_planes
+        return det_cfg
+
+@at.obj(
+    stage_blocks_ratio=at.Real(0.5, 3.0),
+    base_channels_ratio=at.Real(0.5, 3.0),
+    fpn_channel=at.Int(8,128),
+    head_channel=at.Int(8,256),
+    head_stack=at.Int(1,4),
+)
+class GenConfigAll:
+    def __init__(self, **kwargs):
+        d = {}
+        d.update(**kwargs)
+        for k, v in d.items():
+            setattr(self, k, v)
+        self.m = 1
+
+    def merge_cfg(self, det_cfg):
+        block_cfg = det_cfg['model']['backbone']['block_cfg']
+        stage_blocks = tuple([int(np.round(x*self.stage_blocks_ratio)) for x in block_cfg['stage_blocks']])
+        block_cfg['stage_blocks'] = stage_blocks
+        stage_planes = [int(np.round(x*self.base_channels_ratio))//8*8 for x in block_cfg['stage_planes']]
+        block_cfg['stage_planes'] = stage_planes
+        det_cfg['model']['backbone']['block_cfg'] = block_cfg
+        det_cfg['model']['backbone']['base_channels'] = stage_planes[0]
+        neck_in_planes = stage_planes if block_cfg['block']=='BasicBlock' else [4*x for x in stage_planes]
+        det_cfg['model']['neck']['in_channels'] = neck_in_planes
+
+        fpn_channel = self.fpn_channel//8*8
+        head_channel = self.head_channel//8*8
+        head_stack = self.head_stack
+        det_cfg['model']['neck']['out_channels'] = fpn_channel
+        det_cfg['model']['bbox_head']['in_channels'] = fpn_channel
+        det_cfg['model']['bbox_head']['feat_channels'] = head_channel
+        det_cfg['model']['bbox_head']['stacked_convs'] = head_stack
+        gn_num_groups = 8
+        for _gn_num_groups in [8, 16, 32, 64]:
+            if head_channel%_gn_num_groups!=0:
+                break
+            gn_num_groups = _gn_num_groups
+        det_cfg['model']['bbox_head']['norm_cfg']['num_groups'] = gn_num_groups
+        return det_cfg
+
+def get_args():
+    parser = argparse.ArgumentParser(description='Auto-SCRFD')
+    # config files
+    parser.add_argument('--group', type=str, default='configs/scrfdgen2.5g', help='configs work dir')
+    parser.add_argument('--template', type=int, default=0, help='template config index')
+    parser.add_argument('--gflops', type=float, default=2.5, help='expected flops')
+    parser.add_argument('--mode', type=int, default=1, help='generation mode, 1 for searching backbone, 2 for search all')
+    # target flops
+    parser.add_argument('--eps', type=float, default=2e-2,
+                         help='eps for expected flops')
+    # num configs
+    parser.add_argument('--num-configs', type=int, default=64, help='num of expected configs')
+    parser = parser
+
+    args = parser.parse_args()
+    return args
+
+
+def is_config_valid(cfg, target_flops, input_shape, eps):
+    model = build_detector(
+        cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
+    if torch.cuda.is_available():
+        model.cuda()
+    model.eval()
+    if hasattr(model, 'forward_dummy'):
+        model.forward = model.forward_dummy
+    else:
+        raise NotImplementedError(
+            'FLOPs counter is currently not currently supported with {}'.
+            format(model.__class__.__name__))
+    flops, params = get_model_complexity_info(model, input_shape, print_per_layer_stat=False, as_strings=False)
+    print('FLOPs:', flops/1e9)
+    return flops <= (1. + eps) * target_flops and \
+        flops >= (1. - eps) * target_flops
+
+def get_flops(cfg, input_shape):
+    model = build_detector(
+        cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
+    #if torch.cuda.is_available():
+    #    model.cuda()
+    model.eval()
+    if hasattr(model, 'forward_dummy'):
+        model.forward = model.forward_dummy
+    else:
+        raise NotImplementedError(
+            'FLOPs counter is currently not currently supported with {}'.
+            format(model.__class__.__name__))
+    buf = io.StringIO()
+    all_flops, params = get_model_complexity_info(model, input_shape, print_per_layer_stat=True, as_strings=False, ost=buf)
+    buf = buf.getvalue()
+    #print(buf)
+    lines = buf.split("\n")
+    names = ['(stem)', '(layer1)', '(layer2)', '(layer3)', '(layer4)', '(neck)', '(bbox_head)']
+    name_ptr = 0
+    line_num = 0
+    _flops = []
+    while name_ptr<len(names):
+        line = lines[line_num].strip()
+        name = names[name_ptr]
+        if line.startswith(name):
+            flops = float(lines[line_num+1].split(',')[2].strip().split(' ')[0])
+            _flops.append(flops)
+            name_ptr+=1
+        line_num+=1
+
+    backbone_flops = np.array(_flops[:-2], dtype=np.float32)
+    neck_flops = _flops[-2]
+    head_flops = _flops[-1]
+
+    return all_flops/1e9, backbone_flops, neck_flops, head_flops
+
+def is_flops_valid(flops, target_flops, eps):
+    return flops <= (1. + eps) * target_flops and \
+        flops >= (1. - eps) * target_flops
+
+def main():
+    args = get_args()
+    print(datetime.datetime.now())
+
+    input_shape = (3,480,640)
+    runtime_input_shape = input_shape
+    flops_mult = 1.0
+
+    assert osp.exists(args.group)
+    group_name = args.group.split('/')[-1]
+    assert len(group_name)>0
+    input_template = osp.join(args.group, "%s_%d.py"%(group_name, args.template))
+    assert osp.exists(input_template)
+    write_index = args.template+1
+    while True:
+        output_cfg = osp.join(args.group, "%s_%d.py"%(group_name, write_index))
+        if not osp.exists(output_cfg):
+            break
+        write_index+=1
+    print('write-index from:', write_index)
+
+    if args.mode==1:
+        gen = GenConfigBackbone()
+    elif args.mode==2:
+        gen = GenConfigAll()
+        det_cfg = Config.fromfile(input_template)
+        _, template_backbone_flops, _, _= get_flops(det_cfg, runtime_input_shape)
+        template_backbone_ratios = list(map(lambda x:x/template_backbone_flops[0], template_backbone_flops))
+        print('template_backbone_ratios:', template_backbone_ratios)
+
+
+
+    pp = 0
+    write_count = 0
+    while write_count < args.num_configs:
+        pp+=1
+        det_cfg = Config.fromfile(input_template)
+        config = gen.rand
+        det_cfg = config.merge_cfg(det_cfg)
+        all_flops, backbone_flops, neck_flops, head_flops = get_flops(det_cfg, runtime_input_shape)
+        assert len(backbone_flops)==5
+        all_flops *= flops_mult
+        backbone_flops *= flops_mult
+        neck_flops *= flops_mult
+        head_flops *= flops_mult
+        is_valid = True
+        if pp%10==0:
+            print(pp, all_flops, backbone_flops, neck_flops, head_flops, datetime.datetime.now())
+        if args.mode==2:
+            backbone_ratios = list(map(lambda x:x/backbone_flops[0], backbone_flops))
+            #if head_flops*0.8<neck_flops:
+            #    continue
+            for i in range(1,5):
+                if not is_flops_valid(template_backbone_ratios[i], backbone_ratios[i], args.eps*5):
+                    is_valid = False
+                    break
+        if not is_valid:
+            continue
+        #if args.mode==1:
+        #    if np.argmax(backbone_flops)!=1:
+        #        continue
+        #    if np.mean(backbone_flops[1:3])*0.8<np.mean(backbone_flops[-2:]):
+        #        continue
+        if not is_flops_valid(all_flops, args.gflops, args.eps):
+            continue
+
+        output_cfg_file = osp.join(args.group, "%s_%d.py"%(group_name, write_index))
+        det_cfg.dump(output_cfg_file)
+        print('SUCC', write_index, all_flops, backbone_flops, neck_flops, head_flops, datetime.datetime.now())
+        write_index += 1
+        write_count += 1
+
+if __name__ == '__main__':
+    main()
+
diff --git a/detection/scrfd/search_tools/search_stat.py b/detection/scrfd/search_tools/search_stat.py
new file mode 100755
index 000000000..f8b6013d8
--- /dev/null
+++ b/detection/scrfd/search_tools/search_stat.py
@@ -0,0 +1,78 @@
+import os
+import json
+import os.path as osp
+import io
+import torch
+import numpy as np
+from mmcv import Config
+
+from mmdet.models import build_detector
+from mmcv.cnn import get_model_complexity_info
+
+def get_flops(cfg, input_shape):
+    model = build_detector(
+        cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
+    if torch.cuda.is_available():
+        model.cuda()
+    model.eval()
+    if hasattr(model, 'forward_dummy'):
+        model.forward = model.forward_dummy
+    else:
+        raise NotImplementedError(
+            'FLOPs counter is currently not currently supported with {}'.
+            format(model.__class__.__name__))
+    buf = io.StringIO()
+    all_flops, params = get_model_complexity_info(model, input_shape, print_per_layer_stat=True, as_strings=False, ost=buf)
+    buf = buf.getvalue()
+    lines = buf.split("\n")
+    names = ['(stem)', '(layer1)', '(layer2)', '(layer3)', '(layer4)', '(neck)', '(bbox_head)']
+    name_ptr = 0
+    line_num = 0
+    _flops = []
+    while name_ptr<len(names):
+        line = lines[line_num].strip()
+        name = names[name_ptr]
+        if line.startswith(name):
+            flops = float(lines[line_num+1].split(',')[2].strip().split(' ')[0])
+            _flops.append(flops)
+            name_ptr+=1
+        line_num+=1
+
+    backbone_flops = _flops[:-2]
+    neck_flops = _flops[-2]
+    head_flops = _flops[-1]
+    return all_flops/1e9, backbone_flops, neck_flops, head_flops
+def get_stat(result_dir, group, prefix, idx):
+    curr_dir = osp.join(result_dir, group, "%s_%d"%(prefix, idx))
+    aps_file = osp.join(curr_dir, 'aps')
+    aps = []
+    if osp.exists(aps_file):
+        with open(aps_file, 'r') as f:
+            aps = [float(x) for x in f.readline().strip().split(',')]
+    cfg_file = osp.join('configs', group, '%s_%d.py'%(prefix, idx))
+    cfg = Config.fromfile(cfg_file)
+    all_flops, backbone_flops, neck_flops, head_flops = get_flops(cfg, (3,480,640))
+    
+    return aps, all_flops, backbone_flops, neck_flops, head_flops
+
+result_dir = './wouts'
+group = 'scrfdgen2.5g'
+prefix = group
+idx_from = 0
+idx_to = 320
+
+outf = open(osp.join(result_dir, "%s.txt"%prefix), 'w')
+
+for idx in range(idx_from, idx_to):
+    aps, all_flops, backbone_flops, neck_flops, head_flops = get_stat(result_dir, group, prefix, idx)
+    backbone_ratio = np.sum(backbone_flops) / all_flops
+    neck_ratio = neck_flops / all_flops
+    head_ratio = head_flops / all_flops
+    print(idx, aps, all_flops, backbone_flops, backbone_ratio, neck_ratio, head_ratio)
+    name = "%s_%d"%(prefix, idx)
+    data = dict(name=name, backbone_flops=backbone_flops, neck_flops=neck_flops, head_flops=head_flops, all_flops=all_flops, aps=aps)
+    data = json.dumps(data)
+    outf.write(data)
+    outf.write("\n")
+outf.close()
+
diff --git a/detection/scrfd/search_tools/search_test.sh b/detection/scrfd/search_tools/search_test.sh
new file mode 100755
index 000000000..cc4ec842e
--- /dev/null
+++ b/detection/scrfd/search_tools/search_test.sh
@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+
+GPU=0
+OUTPUT_DIR=wouts
+THR=0.02
+GROUP=scrfdgen2p5g
+PREFIX=$GROUP
+
+for i in {1..320}
+do
+    TASK="$PREFIX"_"$i"
+    echo $TASK
+    CUDA_VISIBLE_DEVICES="$GPU" python -u tools/test_widerface.py ./configs/"$GROUP"/"$TASK".py ./work_dirs/"$TASK"/latest.pth --mode 0 --thr "$THR" --out "$OUTPUT_DIR"/"$GROUP"/"$TASK"
+done
+
diff --git a/detection/scrfd/search_tools/search_train.py b/detection/scrfd/search_tools/search_train.py
new file mode 100755
index 000000000..2322e6567
--- /dev/null
+++ b/detection/scrfd/search_tools/search_train.py
@@ -0,0 +1,21 @@
+import os, sys
+
+
+
+def train(group, prefix, idx, gpuid):
+    assert idx>=0
+    cmd = "CUDA_VISIBLE_DEVICES='%d' PORT=%d bash ./tools/dist_train.sh ./configs/%s/%s_%d.py 1 --no-validate"%(gpuid,29100+idx, group, prefix, idx)
+    print(cmd)
+    os.system(cmd)
+
+
+gpuid = int(sys.argv[1])
+idx_from = int(sys.argv[2])
+idx_to = int(sys.argv[3])
+group = 'scrfdgen'
+if len(sys.argv)>4:
+    group = sys.argv[4]
+
+for idx in range(idx_from, idx_to):
+    train(group, group, idx, gpuid)
+
diff --git a/detection/scrfd/search_tools/search_train.sh b/detection/scrfd/search_tools/search_train.sh
new file mode 100755
index 000000000..53bcd46f2
--- /dev/null
+++ b/detection/scrfd/search_tools/search_train.sh
@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+
+GROUP=scrfdgen2.5g
+TASKS_PER_GPU=8
+OFFSET=1
+for i in {0..7}
+do
+    let a=TASKS_PER_GPU*i+OFFSET
+    let i2=i+1
+    let b=TASKS_PER_GPU*i2+OFFSET
+    echo $i,$a,$b,$GROUP
+    python -u search_tools/search_train.py $i $a $b $GROUP > "gpu$i".log 2>&1 &
+done
+
diff --git a/detection/scrfd/setup.cfg b/detection/scrfd/setup.cfg
new file mode 100755
index 000000000..873406e8f
--- /dev/null
+++ b/detection/scrfd/setup.cfg
@@ -0,0 +1,13 @@
+[isort]
+line_length = 79
+multi_line_output = 0
+known_standard_library = setuptools
+known_first_party = mmdet
+known_third_party = PIL,asynctest,cityscapesscripts,cv2,matplotlib,mmcv,numpy,onnx,onnxruntime,pycocotools,pytest,robustness_eval,scipy,seaborn,six,terminaltables,torch
+no_lines_before = STDLIB,LOCALFOLDER
+default_section = THIRDPARTY
+
+[yapf]
+BASED_ON_STYLE = pep8
+BLANK_LINE_BEFORE_NESTED_CLASS_OR_DEF = true
+SPLIT_BEFORE_EXPRESSION_AFTER_OPENING_PAREN = true
diff --git a/detection/scrfd/setup.py b/detection/scrfd/setup.py
new file mode 100755
index 000000000..55eea6ba6
--- /dev/null
+++ b/detection/scrfd/setup.py
@@ -0,0 +1,161 @@
+#!/usr/bin/env python
+import os
+from setuptools import find_packages, setup
+
+import torch
+from torch.utils.cpp_extension import (BuildExtension, CppExtension,
+                                       CUDAExtension)
+
+
+def readme():
+    with open('README.md', encoding='utf-8') as f:
+        content = f.read()
+    return content
+
+
+version_file = 'mmdet/version.py'
+
+
+def get_version():
+    with open(version_file, 'r') as f:
+        exec(compile(f.read(), version_file, 'exec'))
+    return locals()['__version__']
+
+
+def make_cuda_ext(name, module, sources, sources_cuda=[]):
+
+    define_macros = []
+    extra_compile_args = {'cxx': []}
+
+    if torch.cuda.is_available() or os.getenv('FORCE_CUDA', '0') == '1':
+        define_macros += [('WITH_CUDA', None)]
+        extension = CUDAExtension
+        extra_compile_args['nvcc'] = [
+            '-D__CUDA_NO_HALF_OPERATORS__',
+            '-D__CUDA_NO_HALF_CONVERSIONS__',
+            '-D__CUDA_NO_HALF2_OPERATORS__',
+        ]
+        sources += sources_cuda
+    else:
+        print(f'Compiling {name} without CUDA')
+        extension = CppExtension
+
+    return extension(
+        name=f'{module}.{name}',
+        sources=[os.path.join(*module.split('.'), p) for p in sources],
+        define_macros=define_macros,
+        extra_compile_args=extra_compile_args)
+
+
+def parse_requirements(fname='requirements.txt', with_version=True):
+    """Parse the package dependencies listed in a requirements file but strips
+    specific versioning information.
+
+    Args:
+        fname (str): path to requirements file
+        with_version (bool, default=False): if True include version specs
+
+    Returns:
+        List[str]: list of requirements items
+
+    CommandLine:
+        python -c "import setup; print(setup.parse_requirements())"
+    """
+    import sys
+    from os.path import exists
+    import re
+    require_fpath = fname
+
+    def parse_line(line):
+        """Parse information from a line in a requirements text file."""
+        if line.startswith('-r '):
+            # Allow specifying requirements in other files
+            target = line.split(' ')[1]
+            for info in parse_require_file(target):
+                yield info
+        else:
+            info = {'line': line}
+            if line.startswith('-e '):
+                info['package'] = line.split('#egg=')[1]
+            elif '@git+' in line:
+                info['package'] = line
+            else:
+                # Remove versioning from the package
+                pat = '(' + '|'.join(['>=', '==', '>']) + ')'
+                parts = re.split(pat, line, maxsplit=1)
+                parts = [p.strip() for p in parts]
+
+                info['package'] = parts[0]
+                if len(parts) > 1:
+                    op, rest = parts[1:]
+                    if ';' in rest:
+                        # Handle platform specific dependencies
+                        # http://setuptools.readthedocs.io/en/latest/setuptools.html#declaring-platform-specific-dependencies
+                        version, platform_deps = map(str.strip,
+                                                     rest.split(';'))
+                        info['platform_deps'] = platform_deps
+                    else:
+                        version = rest  # NOQA
+                    info['version'] = (op, version)
+            yield info
+
+    def parse_require_file(fpath):
+        with open(fpath, 'r') as f:
+            for line in f.readlines():
+                line = line.strip()
+                if line and not line.startswith('#'):
+                    for info in parse_line(line):
+                        yield info
+
+    def gen_packages_items():
+        if exists(require_fpath):
+            for info in parse_require_file(require_fpath):
+                parts = [info['package']]
+                if with_version and 'version' in info:
+                    parts.extend(info['version'])
+                if not sys.version.startswith('3.4'):
+                    # apparently package_deps are broken in 3.4
+                    platform_deps = info.get('platform_deps')
+                    if platform_deps is not None:
+                        parts.append(';' + platform_deps)
+                item = ''.join(parts)
+                yield item
+
+    packages = list(gen_packages_items())
+    return packages
+
+
+if __name__ == '__main__':
+    setup(
+        name='mmdet',
+        version=get_version(),
+        description='OpenMMLab Detection Toolbox and Benchmark',
+        long_description=readme(),
+        long_description_content_type='text/markdown',
+        author='OpenMMLab',
+        author_email='openmmlab@gmail.com',
+        keywords='computer vision, object detection',
+        url='https://github.com/open-mmlab/mmdetection',
+        packages=find_packages(exclude=('configs', 'tools', 'demo')),
+        classifiers=[
+            'Development Status :: 5 - Production/Stable',
+            'License :: OSI Approved :: Apache Software License',
+            'Operating System :: OS Independent',
+            'Programming Language :: Python :: 3',
+            'Programming Language :: Python :: 3.6',
+            'Programming Language :: Python :: 3.7',
+            'Programming Language :: Python :: 3.8',
+        ],
+        license='Apache License 2.0',
+        setup_requires=parse_requirements('requirements/build.txt'),
+        tests_require=parse_requirements('requirements/tests.txt'),
+        install_requires=parse_requirements('requirements/runtime.txt'),
+        extras_require={
+            'all': parse_requirements('requirements.txt'),
+            'tests': parse_requirements('requirements/tests.txt'),
+            'build': parse_requirements('requirements/build.txt'),
+            'optional': parse_requirements('requirements/optional.txt'),
+        },
+        ext_modules=[],
+        cmdclass={'build_ext': BuildExtension},
+        zip_safe=False)
diff --git a/detection/scrfd/tools/benchmark.py b/detection/scrfd/tools/benchmark.py
new file mode 100755
index 000000000..9c53a4291
--- /dev/null
+++ b/detection/scrfd/tools/benchmark.py
@@ -0,0 +1,116 @@
+import argparse
+import os
+import os.path as osp
+import pickle
+import numpy as np
+import datetime
+import warnings
+
+import mmcv
+import torch
+from mmcv import Config, DictAction
+from mmcv.cnn import fuse_conv_bn
+from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
+from mmcv.runner import (get_dist_info, init_dist, load_checkpoint,
+                         wrap_fp16_model)
+
+from mmdet.apis import multi_gpu_test, single_gpu_test
+from mmdet.datasets import (build_dataloader, build_dataset,
+                            replace_ImageToTensor)
+from mmdet.models import build_detector
+from mmdet.core.evaluation import wider_evaluation, get_widerface_gts
+#from torch.utils import mkldnn as mkldnn_utils
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='MMDet test (and eval) a model')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument('checkpoint', help='checkpoint file')
+    parser.add_argument('--cpu', action="store_true", default=False, help='Use cpu inference')
+    parser.add_argument(
+        '--shape',
+        type=int,
+        nargs='+',
+        default=[480, 640],
+        help='input image size')
+    args = parser.parse_args()
+
+    return args
+
+
+def main():
+    args = parse_args()
+
+
+    cfg = Config.fromfile(args.config)
+    cfg.model.pretrained = None
+
+    # in case the test dataset is concatenated
+    if isinstance(cfg.data.test, dict):
+        cfg.data.test.test_mode = True
+    elif isinstance(cfg.data.test, list):
+        for ds_cfg in cfg.data.test:
+            ds_cfg.test_mode = True
+
+    pipelines = cfg.data.test.pipeline
+    for pipeline in pipelines:
+        if pipeline.type=='MultiScaleFlipAug':
+            #pipeline.img_scale = (640, 640)
+            pipeline.img_scale = None
+            pipeline.scale_factor = 1.0
+            transforms = pipeline.transforms
+            for transform in transforms:
+                if transform.type=='Pad':
+                    #transform.size = pipeline.img_scale
+                    transform.size = None
+                    transform.size_divisor = 1
+    #print(cfg.data.test.pipeline)
+    distributed = False
+
+    # build the dataloader
+    samples_per_gpu = cfg.data.test.pop('samples_per_gpu', 1)
+    if samples_per_gpu > 1:
+        # Replace 'ImageToTensor' to 'DefaultFormatBundle'
+        cfg.data.test.pipeline = replace_ImageToTensor(cfg.data.test.pipeline)
+    dataset = build_dataset(cfg.data.test)
+    data_loader = build_dataloader(
+        dataset,
+        samples_per_gpu=samples_per_gpu,
+        workers_per_gpu=cfg.data.workers_per_gpu,
+        dist=distributed,
+        shuffle=False)
+
+    device = torch.device("cpu" if args.cpu else "cuda")
+
+    # build the model and load checkpoint
+    model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
+    fp16_cfg = cfg.get('fp16', None)
+    checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu')
+    if 'CLASSES' in checkpoint['meta']:
+        model.CLASSES = checkpoint['meta']['CLASSES']
+    else:
+        model.CLASSES = dataset.CLASSES
+
+    model = model.to(device)
+
+    model.eval()
+    dataset = data_loader.dataset
+    for i, data in enumerate(data_loader):
+        img = data['img'][0]
+        #print(img.shape)
+        img = img[:,:,:args.shape[0],:args.shape[1]]
+        img = img.to(device)
+        with torch.no_grad():
+            ta = datetime.datetime.now()
+            result = model.feature_test(img)
+            tb = datetime.datetime.now()
+            print('cost:', (tb-ta).total_seconds())
+            
+
+
+
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/browse_dataset.py b/detection/scrfd/tools/browse_dataset.py
new file mode 100755
index 000000000..55110e894
--- /dev/null
+++ b/detection/scrfd/tools/browse_dataset.py
@@ -0,0 +1,68 @@
+import argparse
+import os
+from pathlib import Path
+
+import mmcv
+from mmcv import Config
+
+from mmdet.datasets.builder import build_dataset
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Browse a dataset')
+    parser.add_argument('config', help='train config file path')
+    parser.add_argument(
+        '--skip-type',
+        type=str,
+        nargs='+',
+        default=['DefaultFormatBundle', 'Normalize', 'Collect'],
+        help='skip some useless pipeline')
+    parser.add_argument(
+        '--output-dir',
+        default=None,
+        type=str,
+        help='If there is no display interface, you can save it')
+    parser.add_argument('--not-show', default=False, action='store_true')
+    parser.add_argument(
+        '--show-interval',
+        type=int,
+        default=999,
+        help='the interval of show (ms)')
+    args = parser.parse_args()
+    return args
+
+
+def retrieve_data_cfg(config_path, skip_type):
+    cfg = Config.fromfile(config_path)
+    train_data_cfg = cfg.data.train
+    train_data_cfg['pipeline'] = [
+        x for x in train_data_cfg.pipeline if x['type'] not in skip_type
+    ]
+
+    return cfg
+
+
+def main():
+    args = parse_args()
+    cfg = retrieve_data_cfg(args.config, args.skip_type)
+
+    dataset = build_dataset(cfg.data.train)
+
+    progress_bar = mmcv.ProgressBar(len(dataset))
+    for item in dataset:
+        filename = os.path.join(args.output_dir,
+                                Path(item['filename']).name
+                                ) if args.output_dir is not None else None
+        mmcv.imshow_det_bboxes(
+            item['img'],
+            item['gt_bboxes'],
+            item['gt_labels'],
+            class_names=dataset.CLASSES,
+            show=not args.not_show,
+            out_file=filename,
+            wait_time=args.show_interval)
+        progress_bar.update()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/convert_crowdhuman.py b/detection/scrfd/tools/convert_crowdhuman.py
new file mode 100644
index 000000000..e116c5126
--- /dev/null
+++ b/detection/scrfd/tools/convert_crowdhuman.py
@@ -0,0 +1,92 @@
+from __future__ import print_function
+import cv2
+import argparse
+import os
+import os.path as osp
+import shutil
+import numpy as np
+import json
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='convert crowdhuman dataset to scrfd format')
+    parser.add_argument('--raw', help='raw dataset dir')
+    parser.add_argument('--save', default='data/crowdhuman', help='save path')
+    args = parser.parse_args()
+
+    return args
+
+def main():
+    args = parse_args()
+    raw_image_dir = osp.join(args.raw, 'Images')
+    for subset in ['train', 'val']:
+        save_image_dir = osp.join(args.save, subset, 'images')
+        if not osp.exists(save_image_dir):
+            os.makedirs(save_image_dir)
+        anno_file = osp.join(args.raw, 'annotation_%s.odgt'%subset)
+        fullbody_anno_file = osp.join(osp.join(args.save, subset, "label_fullbody.txt"))
+        head_anno_file = osp.join(osp.join(args.save, subset, "label_head.txt"))
+        fullbody_f = open(fullbody_anno_file, 'w')
+        head_f = open(head_anno_file, 'w')
+        for line in open(anno_file, 'r'):
+            data = json.loads(line)
+            img_id = data['ID']
+            img_name = "%s.jpg"%img_id
+            raw_image_file = osp.join(raw_image_dir, img_name)
+            target_image_file = osp.join(save_image_dir, img_name)
+            img = cv2.imread(raw_image_file)
+            print(raw_image_file, img.shape)
+            fullbody_f.write("# %s %d %d\n"%(img_name,img.shape[1],img.shape[0]))
+            head_f.write("# %s %d %d\n"%(img_name,img.shape[1],img.shape[0]))
+            shutil.copyfile(raw_image_file, target_image_file)
+            items = data['gtboxes']
+            for item in items:
+                fbox = item['fbox']
+                is_ignore = False
+                extra = item['extra']
+                if 'ignore' in extra:
+                    is_ignore = extra['ignore']==1
+                bbox = np.array(fbox, dtype=np.float32)
+                bbox[2] += bbox[0]
+                bbox[3] += bbox[1]
+                if is_ignore:
+                    fullbody_f.write("%.5f %.5f %.5f %.5f %d\n"%(bbox[0], bbox[1], bbox[2], bbox[3], is_ignore))
+                else:
+                    vbox = item['vbox']
+                    vbox = np.array(vbox, dtype=np.float32)
+                    vbox[2] += vbox[0]
+                    vbox[3] += vbox[1]
+                    x1, y1, x2, y2 = vbox[0], vbox[1], vbox[2], vbox[3]
+                    cx = (x1+x2)/2
+                    cy = (y1+y2)/2
+                    kps = np.ones( (5,3), dtype=np.float32)
+                    kps[0,0] = x1
+                    kps[0,1] = y1
+                    kps[1,0] = x2
+                    kps[1,1] = y1
+                    kps[2,0] = cx
+                    kps[2,1] = cy
+                    kps[3,0] = x1
+                    kps[3,1] = y2
+                    kps[4,0] = x2
+                    kps[4,1] = y2
+                    kps_str = " ".join(["%.5f"%x for x in kps.flatten()])
+                    fullbody_f.write("%.5f %.5f %.5f %.5f %s\n"%(bbox[0], bbox[1], bbox[2], bbox[3], kps_str))
+
+
+                hbox = item['hbox']
+                is_ignore = False
+                extra = item['head_attr']
+                if 'ignore' in extra:
+                    is_ignore = extra['ignore']==1
+                bbox = np.array(hbox, dtype=np.float32)
+                bbox[2] += bbox[0]
+                bbox[3] += bbox[1]
+                head_f.write("%.5f %.5f %.5f %.5f %d\n"%(bbox[0], bbox[1], bbox[2], bbox[3], is_ignore))
+        fullbody_f.close()
+        head_f.close()
+
+
+if __name__ == '__main__':
+    main()
+
diff --git a/detection/scrfd/tools/dist_test.sh b/detection/scrfd/tools/dist_test.sh
new file mode 100755
index 000000000..5488d938c
--- /dev/null
+++ b/detection/scrfd/tools/dist_test.sh
@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+CONFIG=$1
+CHECKPOINT=$2
+GPUS=$3
+PORT=${PORT:-29500}
+
+PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
+python -u -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
+    $(dirname "$0")/test.py $CONFIG $CHECKPOINT --launcher pytorch ${@:4}
diff --git a/detection/scrfd/tools/dist_train.sh b/detection/scrfd/tools/dist_train.sh
new file mode 100755
index 000000000..5b43fffbf
--- /dev/null
+++ b/detection/scrfd/tools/dist_train.sh
@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+
+CONFIG=$1
+GPUS=$2
+PORT=${PORT:-29500}
+
+PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
+python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
+    $(dirname "$0")/train.py $CONFIG --launcher pytorch ${@:3}
diff --git a/detection/scrfd/tools/get_flops.py b/detection/scrfd/tools/get_flops.py
new file mode 100755
index 000000000..3f42b7064
--- /dev/null
+++ b/detection/scrfd/tools/get_flops.py
@@ -0,0 +1,75 @@
+import argparse
+
+import torch
+from mmcv import Config
+
+from mmdet.models import build_detector
+
+try:
+    from mmcv.cnn import get_model_complexity_info
+except ImportError:
+    raise ImportError('Please upgrade mmcv to >0.6.2')
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train a detector')
+    parser.add_argument('config', help='train config file path')
+    parser.add_argument(
+        '--shape',
+        type=int,
+        nargs='+',
+        #default=[1280, 800],
+        default=[640, 640],
+        help='input image size')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+
+    args = parse_args()
+
+    if len(args.shape) == 1:
+        input_shape = (3, args.shape[0], args.shape[0])
+    elif len(args.shape) == 2:
+        input_shape = (3, ) + tuple(args.shape)
+    else:
+        raise ValueError('invalid input shape')
+
+    cfg = Config.fromfile(args.config)
+    # import modules from string list.
+    if cfg.get('custom_imports', None):
+        from mmcv.utils import import_modules_from_strings
+        import_modules_from_strings(**cfg['custom_imports'])
+
+    model = build_detector(
+        cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
+    if torch.cuda.is_available():
+        model.cuda()
+    model.eval()
+    #print(model.bbox_head)
+
+    if hasattr(model, 'forward_dummy'):
+        model.forward = model.forward_dummy
+    else:
+        raise NotImplementedError(
+            'FLOPs counter is currently not currently supported with {}'.
+            format(model.__class__.__name__))
+
+    #flops, params = get_model_complexity_info(model, input_shape)
+    flops, params = get_model_complexity_info(model, input_shape, print_per_layer_stat=True, as_strings=False)
+    flops += model.bbox_head.extra_flops
+    flops *= 0.75
+    split_line = '=' * 30
+    print('ON VGA:')
+    print(flops/1e9, 'G')
+    print(params/1e6, 'M')
+    #print(f'{split_line}\nInput shape: {input_shape}\n'
+    #      f'Flops: {flops}\nParams: {params}\n{split_line}')
+    #print('!!!Please be cautious if you use the results in papers. '
+    #      'You may need to check if all ops are supported and verify that the '
+    #      'flops computation is correct.')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/print_config.py b/detection/scrfd/tools/print_config.py
new file mode 100755
index 000000000..2ba994fb3
--- /dev/null
+++ b/detection/scrfd/tools/print_config.py
@@ -0,0 +1,26 @@
+import argparse
+
+from mmcv import Config, DictAction
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Print the whole config')
+    parser.add_argument('config', help='config file path')
+    parser.add_argument(
+        '--options', nargs='+', action=DictAction, help='arguments in dict')
+    args = parser.parse_args()
+
+    return args
+
+
+def main():
+    args = parse_args()
+
+    cfg = Config.fromfile(args.config)
+    if args.options is not None:
+        cfg.merge_from_dict(args.options)
+    print(f'Config:\n{cfg.pretty_text}')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/publish_model.py b/detection/scrfd/tools/publish_model.py
new file mode 100755
index 000000000..c20e7e38b
--- /dev/null
+++ b/detection/scrfd/tools/publish_model.py
@@ -0,0 +1,39 @@
+import argparse
+import subprocess
+
+import torch
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Process a checkpoint to be published')
+    parser.add_argument('in_file', help='input checkpoint filename')
+    parser.add_argument('out_file', help='output checkpoint filename')
+    args = parser.parse_args()
+    return args
+
+
+def process_checkpoint(in_file, out_file):
+    checkpoint = torch.load(in_file, map_location='cpu')
+    # remove optimizer for smaller file size
+    if 'optimizer' in checkpoint:
+        del checkpoint['optimizer']
+    # if it is necessary to remove some sensitive data in checkpoint['meta'],
+    # add the code here.
+    torch.save(checkpoint, out_file)
+    sha = subprocess.check_output(['sha256sum', out_file]).decode()
+    if out_file.endswith('.pth'):
+        out_file_name = out_file[:-4]
+    else:
+        out_file_name = out_file
+    final_file = out_file_name + f'-{sha[:8]}.pth'
+    subprocess.Popen(['mv', out_file, final_file])
+
+
+def main():
+    args = parse_args()
+    process_checkpoint(args.in_file, args.out_file)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/scrfd.py b/detection/scrfd/tools/scrfd.py
new file mode 100644
index 000000000..176d90e9d
--- /dev/null
+++ b/detection/scrfd/tools/scrfd.py
@@ -0,0 +1,337 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+from __future__ import division
+import datetime
+import numpy as np
+import onnx
+import onnxruntime
+import os
+import os.path as osp
+import cv2
+import sys
+
+def softmax(z):
+    assert len(z.shape) == 2
+    s = np.max(z, axis=1)
+    s = s[:, np.newaxis] # necessary step to do broadcasting
+    e_x = np.exp(z - s)
+    div = np.sum(e_x, axis=1)
+    div = div[:, np.newaxis] # dito
+    return e_x / div
+
+def distance2bbox(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def distance2kps(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i%2] + distance[:, i]
+        py = points[:, i%2+1] + distance[:, i+1]
+        if max_shape is not None:
+            px = px.clamp(min=0, max=max_shape[1])
+            py = py.clamp(min=0, max=max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return np.stack(preds, axis=-1)
+
+class SCRFD:
+    def __init__(self, model_file=None, session=None):
+        import onnxruntime
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'detection'
+        self.batched = False
+        if self.session is None:
+            assert self.model_file is not None
+            assert osp.exists(self.model_file)
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        self.center_cache = {}
+        self.nms_thresh = 0.4
+        self._init_vars()
+
+    def _init_vars(self):
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        if isinstance(input_shape[2], str):
+            self.input_size = None
+        else:
+            self.input_size = tuple(input_shape[2:4][::-1])
+        input_name = input_cfg.name
+        outputs = self.session.get_outputs()
+        if len(outputs[0].shape) == 3:
+            self.batched = True
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        self.use_kps = False
+        self._num_anchors = 1
+        if len(outputs)==6:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+        elif len(outputs)==9:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+            self.use_kps = True
+        elif len(outputs)==10:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+        elif len(outputs)==15:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+            self.use_kps = True
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+        nms_thresh = kwargs.get('nms_thresh', None)
+        if nms_thresh is not None:
+            self.nms_thresh = nms_thresh
+        input_size = kwargs.get('input_size', None)
+        if input_size is not None:
+            if self.input_size is not None:
+                print('warning: det_size is already set in scrfd model, ignore')
+            else:
+                self.input_size = input_size
+
+    def forward(self, img, thresh):
+        scores_list = []
+        bboxes_list = []
+        kpss_list = []
+        input_size = tuple(img.shape[0:2][::-1])
+        blob = cv2.dnn.blobFromImage(img, 1.0/128, input_size, (127.5, 127.5, 127.5), swapRB=True)
+        net_outs = self.session.run(self.output_names, {self.input_name : blob})
+
+        input_height = blob.shape[2]
+        input_width = blob.shape[3]
+        fmc = self.fmc
+        for idx, stride in enumerate(self._feat_stride_fpn):
+            # If model support batch dim, take first output
+            if self.batched:
+                scores = net_outs[idx][0]
+                bbox_preds = net_outs[idx + fmc][0]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2][0] * stride
+            # If model doesn't support batching take output as is
+            else:
+                scores = net_outs[idx]
+                bbox_preds = net_outs[idx + fmc]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2] * stride
+
+            height = input_height // stride
+            width = input_width // stride
+            K = height * width
+            key = (height, width, stride)
+            if key in self.center_cache:
+                anchor_centers = self.center_cache[key]
+            else:
+                #solution-1, c style:
+                #anchor_centers = np.zeros( (height, width, 2), dtype=np.float32 )
+                #for i in range(height):
+                #    anchor_centers[i, :, 1] = i
+                #for i in range(width):
+                #    anchor_centers[:, i, 0] = i
+
+                #solution-2:
+                #ax = np.arange(width, dtype=np.float32)
+                #ay = np.arange(height, dtype=np.float32)
+                #xv, yv = np.meshgrid(np.arange(width), np.arange(height))
+                #anchor_centers = np.stack([xv, yv], axis=-1).astype(np.float32)
+
+                #solution-3:
+                anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+                #print(anchor_centers.shape)
+
+                anchor_centers = (anchor_centers * stride).reshape( (-1, 2) )
+                if self._num_anchors>1:
+                    anchor_centers = np.stack([anchor_centers]*self._num_anchors, axis=1).reshape( (-1,2) )
+                if len(self.center_cache)<100:
+                    self.center_cache[key] = anchor_centers
+
+            pos_inds = np.where(scores>=thresh)[0]
+            bboxes = distance2bbox(anchor_centers, bbox_preds)
+            pos_scores = scores[pos_inds]
+            pos_bboxes = bboxes[pos_inds]
+            scores_list.append(pos_scores)
+            bboxes_list.append(pos_bboxes)
+            if self.use_kps:
+                kpss = distance2kps(anchor_centers, kps_preds)
+                #kpss = kps_preds
+                kpss = kpss.reshape( (kpss.shape[0], -1, 2) )
+                pos_kpss = kpss[pos_inds]
+                kpss_list.append(pos_kpss)
+        return scores_list, bboxes_list, kpss_list
+
+    def detect(self, img, thresh=0.5, input_size = None, max_num=0, metric='default'):
+        assert input_size is not None or self.input_size is not None
+        input_size = self.input_size if input_size is None else input_size
+            
+        im_ratio = float(img.shape[0]) / img.shape[1]
+        model_ratio = float(input_size[1]) / input_size[0]
+        if im_ratio>model_ratio:
+            new_height = input_size[1]
+            new_width = int(new_height / im_ratio)
+        else:
+            new_width = input_size[0]
+            new_height = int(new_width * im_ratio)
+        det_scale = float(new_height) / img.shape[0]
+        resized_img = cv2.resize(img, (new_width, new_height))
+        det_img = np.zeros( (input_size[1], input_size[0], 3), dtype=np.uint8 )
+        det_img[:new_height, :new_width, :] = resized_img
+
+        scores_list, bboxes_list, kpss_list = self.forward(det_img, thresh)
+
+        scores = np.vstack(scores_list)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        bboxes = np.vstack(bboxes_list) / det_scale
+        if self.use_kps:
+            kpss = np.vstack(kpss_list) / det_scale
+        pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
+        pre_det = pre_det[order, :]
+        keep = self.nms(pre_det)
+        det = pre_det[keep, :]
+        if self.use_kps:
+            kpss = kpss[order,:,:]
+            kpss = kpss[keep,:,:]
+        else:
+            kpss = None
+        if max_num > 0 and det.shape[0] > max_num:
+            area = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                    det[:, 1])
+            img_center = img.shape[0] // 2, img.shape[1] // 2
+            offsets = np.vstack([
+                (det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2 - img_center[0]
+            ])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            if metric=='max':
+                values = area
+            else:
+                values = area - offset_dist_squared * 2.0  # some extra weight on the centering
+            bindex = np.argsort(
+                values)[::-1]  # some extra weight on the centering
+            bindex = bindex[0:max_num]
+            det = det[bindex, :]
+            if kpss is not None:
+                kpss = kpss[bindex, :]
+        return det, kpss
+
+    def nms(self, dets):
+        thresh = self.nms_thresh
+        x1 = dets[:, 0]
+        y1 = dets[:, 1]
+        x2 = dets[:, 2]
+        y2 = dets[:, 3]
+        scores = dets[:, 4]
+
+        areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+        order = scores.argsort()[::-1]
+
+        keep = []
+        while order.size > 0:
+            i = order[0]
+            keep.append(i)
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+            inds = np.where(ovr <= thresh)[0]
+            order = order[inds + 1]
+
+        return keep
+
+def get_scrfd(name, download=False, root='~/.insightface/models', **kwargs):
+    if not download:
+        assert os.path.exists(name)
+        return SCRFD(name)
+    else:
+        from .model_store import get_model_file
+        _file = get_model_file("scrfd_%s" % name, root=root)
+        return SCRFD(_file)
+
+
+def scrfd_2p5gkps(**kwargs):
+    return get_scrfd("2p5gkps", download=True, **kwargs)
+
+
+if __name__ == '__main__':
+    import glob
+    #detector = SCRFD(model_file='./det.onnx')
+    detector = SCRFD(model_file='./det.onnx')
+    detector.prepare(-1)
+    img_paths = ['tests/data/t3.jpg']
+    for img_path in img_paths:
+        img = cv2.imread(img_path)
+
+        for _ in range(1):
+            ta = datetime.datetime.now()
+            bboxes, kpss = detector.detect(img, 0.5, input_size = (640, 640))
+            #bboxes, kpss = detector.detect(img, 0.5)
+            tb = datetime.datetime.now()
+            print('all cost:', (tb-ta).total_seconds()*1000)
+        print(img_path, bboxes.shape)
+        if kpss is not None:
+            print(kpss.shape)
+        for i in range(bboxes.shape[0]):
+            bbox = bboxes[i]
+            x1,y1,x2,y2,score = bbox.astype(np.int)
+            cv2.rectangle(img, (x1,y1)  , (x2,y2) , (255,0,0) , 2)
+            if kpss is not None:
+                kps = kpss[i]
+                for kp in kps:
+                    kp = kp.astype(np.int)
+                    cv2.circle(img, tuple(kp) , 1, (0,0,255) , 2)
+        filename = img_path.split('/')[-1]
+        print('output:', filename)
+        cv2.imwrite('./outputs/%s'%filename, img)
+
diff --git a/detection/scrfd/tools/scrfd2onnx.py b/detection/scrfd/tools/scrfd2onnx.py
new file mode 100755
index 000000000..218651574
--- /dev/null
+++ b/detection/scrfd/tools/scrfd2onnx.py
@@ -0,0 +1,201 @@
+import argparse
+import os.path as osp
+
+import numpy as np
+import onnx
+import os
+#import onnxruntime as rt
+import torch
+
+from mmdet.core import (build_model_from_cfg, generate_inputs_and_wrap_model,
+                        preprocess_example_input)
+
+#from mmdet.models import build
+
+def pytorch2onnx(config_path,
+                 checkpoint_path,
+                 input_img,
+                 input_shape,
+                 opset_version=11,
+                 show=False,
+                 output_file='tmp.onnx',
+                 verify=False,
+                 simplify = True,
+                 dynamic = True,
+                 normalize_cfg=None,
+                 dataset='coco',
+                 test_img=None):
+
+    input_config = {
+        'input_shape': input_shape,
+        'input_path': input_img,
+        'normalize_cfg': normalize_cfg
+    }
+    checkpoint = torch.load(checkpoint_path, map_location='cpu')
+    tmp_ckpt_file = None
+    # remove optimizer for smaller file size
+    if 'optimizer' in checkpoint:
+        del checkpoint['optimizer']
+        tmp_ckpt_file = checkpoint_path+"_slim.pth"
+        torch.save(checkpoint, tmp_ckpt_file)
+        print('remove optimizer params and save to', tmp_ckpt_file)
+        checkpoint_path = tmp_ckpt_file
+
+    model, tensor_data = generate_inputs_and_wrap_model(
+        config_path, checkpoint_path, input_config)
+
+    if tmp_ckpt_file is not None:
+        os.remove(tmp_ckpt_file)
+
+    if simplify or dynamic:
+        ori_output_file = output_file.split('.')[0]+"_ori.onnx"
+    else:
+        ori_output_file = output_file
+
+    # Define input and outputs names, which are required to properly define
+    # dynamic axes
+    input_names = ['input.1']
+    output_names = ['score_8', 'score_16', 'score_32',
+                    'bbox_8', 'bbox_16', 'bbox_32',
+                    ]
+
+    # If model graph contains keypoints strides add keypoints to outputs
+    if 'stride_kps' in str(model):
+        output_names += ['kps_8', 'kps_16', 'kps_32']
+
+    # Define dynamic axes for export
+    dynamic_axes = None
+    if dynamic:
+        dynamic_axes = {out: {0: '?', 1: '?'} for out in output_names}
+        dynamic_axes[input_names[0]] = {
+            0: '?',
+            2: '?',
+            3: '?'
+        }
+
+    torch.onnx.export(
+        model,
+        tensor_data,
+        ori_output_file,
+        keep_initializers_as_inputs=False,
+        verbose=False,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes=dynamic_axes,
+        opset_version=opset_version)
+
+    if simplify or dynamic:
+        model = onnx.load(ori_output_file)
+        if simplify:
+            from onnxsim import simplify
+            #print(model.graph.input[0])
+            if dynamic:
+                input_shapes = {model.graph.input[0].name : list(input_shape)}
+                model, check = simplify(model, input_shapes=input_shapes, dynamic_input_shape=True)
+            else:
+                model, check = simplify(model)
+            assert check, "Simplified ONNX model could not be validated"
+        onnx.save(model, output_file)
+        os.remove(ori_output_file)
+
+
+    print(f'Successfully exported ONNX model: {output_file}')
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Convert MMDetection models to ONNX')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument('checkpoint', help='checkpoint file')
+    parser.add_argument('--input-img', type=str, help='Images for input')
+    parser.add_argument('--show', action='store_true', help='show onnx graph')
+    parser.add_argument('--output-file', type=str, default='')
+    parser.add_argument('--opset-version', type=int, default=11)
+    parser.add_argument(
+        '--test-img', type=str, default=None, help='Images for test')
+    parser.add_argument(
+        '--dataset', type=str, default='coco', help='Dataset name')
+    parser.add_argument(
+        '--verify',
+        action='store_true',
+        help='verify the onnx model output against pytorch output')
+    parser.add_argument(
+        '--shape',
+        type=int,
+        nargs='+',
+        #default=[640, 640],
+        #default=[384, 384],
+        default=[-1, -1],
+        help='input image size')
+    parser.add_argument(
+        '--mean',
+        type=float,
+        nargs='+',
+        default=[127.5, 127.5, 127.5],
+        help='mean value used for preprocess input data')
+    parser.add_argument(
+        '--std',
+        type=float,
+        nargs='+',
+        default=[128.0, 128.0, 128.0],
+        help='variance value used for preprocess input data')
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    assert args.opset_version == 11, 'MMDet only support opset 11 now'
+
+    if not args.input_img:
+        args.input_img = osp.join(
+            osp.dirname(__file__), '../tests/data/t1.jpg')
+
+    if len(args.shape) == 1:
+        input_shape = (1, 3, args.shape[0], args.shape[0])
+    elif len(args.shape) == 2:
+        input_shape = (1, 3) + tuple(args.shape)
+    else:
+        raise ValueError('invalid input shape')
+
+    assert len(args.mean) == 3
+    assert len(args.std) == 3
+
+    simplify = True
+    dynamic = False
+    if input_shape[2]<=0 or input_shape[3]<=0:
+        input_shape = (1,3,640,640)
+        dynamic = True
+        #simplify = False
+        print('set to dynamic input with dummy shape:', input_shape)
+
+    normalize_cfg = {'mean': args.mean, 'std': args.std}
+
+    if len(args.output_file)==0:
+        output_dir = osp.join(osp.dirname(__file__), '../onnx')
+        if not osp.exists(output_dir):
+            os.makedirs(output_dir)
+        cfg_name = args.config.split('/')[-1]
+        pos = cfg_name.rfind('.')
+        cfg_name = cfg_name[:pos]
+        if dynamic:
+            args.output_file = osp.join(output_dir, "%s.onnx"%cfg_name)
+        else:
+            args.output_file = osp.join(output_dir, "%s_shape%dx%d.onnx"%(cfg_name,input_shape[2],input_shape[3]))
+
+    # convert model to onnx file
+    pytorch2onnx(
+        args.config,
+        args.checkpoint,
+        args.input_img,
+        input_shape,
+        opset_version=args.opset_version,
+        show=args.show,
+        output_file=args.output_file,
+        verify=args.verify,
+        simplify = simplify,
+        dynamic = dynamic,
+        normalize_cfg=normalize_cfg,
+        dataset=args.dataset,
+        test_img=args.test_img)
diff --git a/detection/scrfd/tools/test.py b/detection/scrfd/tools/test.py
new file mode 100755
index 000000000..8dcd305e1
--- /dev/null
+++ b/detection/scrfd/tools/test.py
@@ -0,0 +1,208 @@
+import argparse
+import os
+import warnings
+
+import mmcv
+import torch
+from mmcv import Config, DictAction
+from mmcv.cnn import fuse_conv_bn
+from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
+from mmcv.runner import (get_dist_info, init_dist, load_checkpoint,
+                         wrap_fp16_model)
+
+from mmdet.apis import multi_gpu_test, single_gpu_test
+from mmdet.datasets import (build_dataloader, build_dataset,
+                            replace_ImageToTensor)
+from mmdet.models import build_detector
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='MMDet test (and eval) a model')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument('checkpoint', help='checkpoint file')
+    parser.add_argument('--out', help='output result file in pickle format')
+    parser.add_argument(
+        '--fuse-conv-bn',
+        action='store_true',
+        help='Whether to fuse conv and bn, this will slightly increase'
+        'the inference speed')
+    parser.add_argument(
+        '--format-only',
+        action='store_true',
+        help='Format the output results without perform evaluation. It is'
+        'useful when you want to format the result to a specific format and '
+        'submit it to the test server')
+    parser.add_argument(
+        '--eval',
+        type=str,
+        nargs='+',
+        help='evaluation metrics, which depends on the dataset, e.g., "bbox",'
+        ' "segm", "proposal" for COCO, and "mAP", "recall" for PASCAL VOC')
+    parser.add_argument('--show', action='store_true', help='show results')
+    parser.add_argument(
+        '--show-dir', help='directory where painted images will be saved')
+    parser.add_argument(
+        '--show-score-thr',
+        type=float,
+        default=0.3,
+        help='score threshold (default: 0.3)')
+    parser.add_argument(
+        '--gpu-collect',
+        action='store_true',
+        help='whether to use gpu to collect results.')
+    parser.add_argument(
+        '--tmpdir',
+        help='tmp directory used for collecting results from multiple '
+        'workers, available when gpu-collect is not specified')
+    parser.add_argument(
+        '--cfg-options',
+        nargs='+',
+        action=DictAction,
+        help='override some settings in the used config, the key-value pair '
+        'in xxx=yyy format will be merged into config file.')
+    parser.add_argument(
+        '--options',
+        nargs='+',
+        action=DictAction,
+        help='custom options for evaluation, the key-value pair in xxx=yyy '
+        'format will be kwargs for dataset.evaluate() function (deprecate), '
+        'change to --eval-options instead.')
+    parser.add_argument(
+        '--eval-options',
+        nargs='+',
+        action=DictAction,
+        help='custom options for evaluation, the key-value pair in xxx=yyy '
+        'format will be kwargs for dataset.evaluate() function')
+    parser.add_argument(
+        '--launcher',
+        choices=['none', 'pytorch', 'slurm', 'mpi'],
+        default='none',
+        help='job launcher')
+    parser.add_argument('--local_rank', type=int, default=0)
+    args = parser.parse_args()
+    if 'LOCAL_RANK' not in os.environ:
+        os.environ['LOCAL_RANK'] = str(args.local_rank)
+
+    if args.options and args.eval_options:
+        raise ValueError(
+            '--options and --eval-options cannot be both '
+            'specified, --options is deprecated in favor of --eval-options')
+    if args.options:
+        warnings.warn('--options is deprecated in favor of --eval-options')
+        args.eval_options = args.options
+    return args
+
+
+def main():
+    args = parse_args()
+
+    assert args.out or args.eval or args.format_only or args.show \
+        or args.show_dir, \
+        ('Please specify at least one operation (save/eval/format/show the '
+         'results / save the results) with the argument "--out", "--eval"'
+         ', "--format-only", "--show" or "--show-dir"')
+
+    if args.eval and args.format_only:
+        raise ValueError('--eval and --format_only cannot be both specified')
+
+    if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
+        raise ValueError('The output file must be a pkl file.')
+
+    cfg = Config.fromfile(args.config)
+    if args.cfg_options is not None:
+        cfg.merge_from_dict(args.cfg_options)
+    # import modules from string list.
+    if cfg.get('custom_imports', None):
+        from mmcv.utils import import_modules_from_strings
+        import_modules_from_strings(**cfg['custom_imports'])
+    # set cudnn_benchmark
+    if cfg.get('cudnn_benchmark', False):
+        torch.backends.cudnn.benchmark = True
+    cfg.model.pretrained = None
+    if cfg.model.get('neck'):
+        if isinstance(cfg.model.neck, list):
+            for neck_cfg in cfg.model.neck:
+                if neck_cfg.get('rfp_backbone'):
+                    if neck_cfg.rfp_backbone.get('pretrained'):
+                        neck_cfg.rfp_backbone.pretrained = None
+        elif cfg.model.neck.get('rfp_backbone'):
+            if cfg.model.neck.rfp_backbone.get('pretrained'):
+                cfg.model.neck.rfp_backbone.pretrained = None
+
+    # in case the test dataset is concatenated
+    if isinstance(cfg.data.test, dict):
+        cfg.data.test.test_mode = True
+    elif isinstance(cfg.data.test, list):
+        for ds_cfg in cfg.data.test:
+            ds_cfg.test_mode = True
+
+    # init distributed env first, since logger depends on the dist info.
+    if args.launcher == 'none':
+        distributed = False
+    else:
+        distributed = True
+        init_dist(args.launcher, **cfg.dist_params)
+
+    # build the dataloader
+    samples_per_gpu = cfg.data.test.pop('samples_per_gpu', 1)
+    if samples_per_gpu > 1:
+        # Replace 'ImageToTensor' to 'DefaultFormatBundle'
+        cfg.data.test.pipeline = replace_ImageToTensor(cfg.data.test.pipeline)
+    dataset = build_dataset(cfg.data.test)
+    data_loader = build_dataloader(
+        dataset,
+        samples_per_gpu=samples_per_gpu,
+        workers_per_gpu=cfg.data.workers_per_gpu,
+        dist=distributed,
+        shuffle=False)
+
+    # build the model and load checkpoint
+    model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
+    fp16_cfg = cfg.get('fp16', None)
+    if fp16_cfg is not None:
+        wrap_fp16_model(model)
+    checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu')
+    if args.fuse_conv_bn:
+        model = fuse_conv_bn(model)
+    # old versions did not save class info in checkpoints, this walkaround is
+    # for backward compatibility
+    if 'CLASSES' in checkpoint['meta']:
+        model.CLASSES = checkpoint['meta']['CLASSES']
+    else:
+        model.CLASSES = dataset.CLASSES
+
+    if not distributed:
+        model = MMDataParallel(model, device_ids=[0])
+        outputs = single_gpu_test(model, data_loader, args.show, args.show_dir,
+                                  args.show_score_thr)
+    else:
+        model = MMDistributedDataParallel(
+            model.cuda(),
+            device_ids=[torch.cuda.current_device()],
+            broadcast_buffers=False)
+        outputs = multi_gpu_test(model, data_loader, args.tmpdir,
+                                 args.gpu_collect)
+
+    rank, _ = get_dist_info()
+    if rank == 0:
+        if args.out:
+            print(f'\nwriting results to {args.out}')
+            mmcv.dump(outputs, args.out)
+        kwargs = {} if args.eval_options is None else args.eval_options
+        if args.format_only:
+            dataset.format_results(outputs, **kwargs)
+        if args.eval:
+            eval_kwargs = cfg.get('evaluation', {}).copy()
+            # hard-code way to remove EvalHook args
+            for key in [
+                    'interval', 'tmpdir', 'start', 'gpu_collect', 'save_best',
+                    'rule'
+            ]:
+                eval_kwargs.pop(key, None)
+            eval_kwargs.update(dict(metric=args.eval, **kwargs))
+            print(dataset.evaluate(outputs, **eval_kwargs))
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/test_example.sh b/detection/scrfd/tools/test_example.sh
new file mode 100755
index 000000000..a3a1935fa
--- /dev/null
+++ b/detection/scrfd/tools/test_example.sh
@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+
+GPU=1
+GROUP=scrfd
+TASK=scrfd_2.5g_bnkps
+
+#CUDA_VISIBLE_DEVICES="$GPU" python -u tools/benchmark_vga.py ./configs/"$GROUP"/"$TASK".py ./work_dirs/"$TASK"/latest.pth #--cpu
+CUDA_VISIBLE_DEVICES="$GPU" python -u tools/test_widerface.py ./configs/"$GROUP"/"$TASK".py ./work_dirs/"$TASK"/model.pth --mode 0 --out wouts --save-preds
diff --git a/detection/scrfd/tools/test_widerface.py b/detection/scrfd/tools/test_widerface.py
new file mode 100755
index 000000000..6a388b8c7
--- /dev/null
+++ b/detection/scrfd/tools/test_widerface.py
@@ -0,0 +1,256 @@
+import argparse
+import os
+import os.path as osp
+import pickle
+import numpy as np
+import warnings
+
+import mmcv
+import torch
+from mmcv import Config, DictAction
+from mmcv.cnn import fuse_conv_bn
+from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
+from mmcv.runner import (get_dist_info, init_dist, load_checkpoint,
+                         wrap_fp16_model)
+
+from mmdet.apis import multi_gpu_test, single_gpu_test
+from mmdet.datasets import (build_dataloader, build_dataset,
+                            replace_ImageToTensor)
+from mmdet.models import build_detector
+from mmdet.core.evaluation import wider_evaluation, get_widerface_gts
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='MMDet test (and eval) a model')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument('checkpoint', help='checkpoint file')
+    parser.add_argument('--out', default='wout', help='output folder')
+    parser.add_argument(
+        '--eval',
+        type=str,
+        nargs='+',
+        help='evaluation metrics, which depends on the dataset, e.g., "bbox",'
+        ' "segm", "proposal" for COCO, and "mAP", "recall" for PASCAL VOC')
+    parser.add_argument('--show', action='store_true', help='show results')
+    parser.add_argument('--save-preds', action='store_true', help='save results')
+    parser.add_argument('--show-assign', action='store_true', help='show bbox assign')
+    parser.add_argument('--debug', action='store_true', help='debug flag')
+    parser.add_argument(
+        '--show-dir', help='directory where painted images will be saved')
+    parser.add_argument(
+        '--show-score-thr',
+        type=float,
+        default=0.3,
+        help='score threshold (default: 0.3)')
+    parser.add_argument(
+        '--thr',
+        type=float,
+        default=0.02,
+        help='score threshold')
+    parser.add_argument('--local_rank', type=int, default=0)
+    parser.add_argument('--mode', type=int, default=0)
+    args = parser.parse_args()
+    if 'LOCAL_RANK' not in os.environ:
+        os.environ['LOCAL_RANK'] = str(args.local_rank)
+
+    return args
+
+
+def main():
+    args = parse_args()
+
+
+    cfg = Config.fromfile(args.config)
+    if cfg.get('custom_imports', None):
+        from mmcv.utils import import_modules_from_strings
+        import_modules_from_strings(**cfg['custom_imports'])
+    # set cudnn_benchmark
+    if cfg.get('cudnn_benchmark', False):
+        torch.backends.cudnn.benchmark = True
+    cfg.model.pretrained = None
+    if cfg.model.get('neck'):
+        if isinstance(cfg.model.neck, list):
+            for neck_cfg in cfg.model.neck:
+                if neck_cfg.get('rfp_backbone'):
+                    if neck_cfg.rfp_backbone.get('pretrained'):
+                        neck_cfg.rfp_backbone.pretrained = None
+        elif cfg.model.neck.get('rfp_backbone'):
+            if cfg.model.neck.rfp_backbone.get('pretrained'):
+                cfg.model.neck.rfp_backbone.pretrained = None
+
+    # in case the test dataset is concatenated
+    if isinstance(cfg.data.test, dict):
+        cfg.data.test.test_mode = True
+    elif isinstance(cfg.data.test, list):
+        for ds_cfg in cfg.data.test:
+            ds_cfg.test_mode = True
+
+    gt_path = os.path.join(os.path.dirname(cfg.data.test.ann_file), 'gt')
+    pipelines = cfg.data.test.pipeline
+    for pipeline in pipelines:
+        if pipeline.type=='MultiScaleFlipAug':
+            if args.mode==0: #640 scale
+                pipeline.img_scale = (640, 640)
+            elif args.mode==1: #for single scale in other pages
+                pipeline.img_scale = (1100, 1650)
+            elif args.mode==2: #original scale
+                pipeline.img_scale = None
+                pipeline.scale_factor = 1.0
+            elif args.mode>30:
+                pipeline.img_scale = (args.mode, args.mode)
+            transforms = pipeline.transforms
+            for transform in transforms:
+                if transform.type=='Pad':
+                    if args.mode!=2:
+                        transform.size = pipeline.img_scale
+                    else:
+                        transform.size = None
+                        transform.size_divisor = 32
+    print(cfg.data.test.pipeline)
+    distributed = False
+
+    # build the dataloader
+    samples_per_gpu = cfg.data.test.pop('samples_per_gpu', 1)
+    if samples_per_gpu > 1:
+        # Replace 'ImageToTensor' to 'DefaultFormatBundle'
+        cfg.data.test.pipeline = replace_ImageToTensor(cfg.data.test.pipeline)
+    dataset = build_dataset(cfg.data.test)
+    data_loader = build_dataloader(
+        dataset,
+        samples_per_gpu=samples_per_gpu,
+        workers_per_gpu=cfg.data.workers_per_gpu,
+        dist=distributed,
+        shuffle=False)
+
+    cfg.test_cfg.score_thr = args.thr
+
+    # build the model and load checkpoint
+    model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
+    fp16_cfg = cfg.get('fp16', None)
+    checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu')
+    if 'CLASSES' in checkpoint['meta']:
+        model.CLASSES = checkpoint['meta']['CLASSES']
+    else:
+        model.CLASSES = dataset.CLASSES
+
+    if args.show_assign:
+        gts_easy, gts_medium, gts_hard = get_widerface_gts(gt_path)
+        assign_stat = [0, 0]
+        gts_size = []
+    model = MMDataParallel(model, device_ids=[0])
+    model.eval()
+    results = {}
+    output_folder = args.out
+    if not os.path.exists(output_folder):
+        os.makedirs(output_folder)
+    dataset = data_loader.dataset
+    prog_bar = mmcv.ProgressBar(len(dataset))
+    for i, data in enumerate(data_loader):
+        with torch.no_grad():
+            result = model(return_loss=False, rescale=True, **data)
+        assert len(result)==1
+        batch_size = 1
+        result = result[0][0]
+        img_metas = data['img_metas'][0].data[0][0]
+        filepath = img_metas['ori_filename']
+        det_scale = img_metas['scale_factor'][0]
+        #print(img_metas)
+        ori_shape = img_metas['ori_shape']
+        img_width = ori_shape[1]
+        img_height = ori_shape[0]
+        _vec = filepath.split('/')
+        pa, pb = _vec[-2], _vec[1]
+        if pa not in results:
+            results[pa] = {}
+        xywh = result.copy()
+        w = xywh[:,2] - xywh[:,0]
+        h = xywh[:,3] - xywh[:,1]
+        xywh[:,2] = w
+        xywh[:,3] = h
+
+        event_name = pa
+        img_name = pb.rstrip('.jpg')
+        results[event_name][img_name] = xywh
+        if args.save_preds:
+            out_dir = os.path.join(output_folder, pa)
+            if not os.path.exists(out_dir):
+                os.makedirs(out_dir)
+            out_file = os.path.join(out_dir, pb.replace('jpg', 'txt'))
+            boxes = result
+            with open(out_file, 'w') as f:
+                name = '/'.join([pa, pb])
+                f.write("%s\n"%(name))
+                f.write("%d\n"%(boxes.shape[0]))
+                for b in range(boxes.shape[0]):
+                    box = boxes[b]
+                    f.write("%.5f %.5f %.5f %.5f %g\n"%(box[0], box[1], box[2]-box[0], box[3]-box[1], box[4]))
+
+        if args.show_assign:
+            assert args.mode==0
+            input_height, input_width = 640, 640
+            gt_hard = gts_hard[event_name][img_name]
+            #print(event_name, img_name, gt_hard.shape)
+            gt_bboxes = gt_hard * det_scale
+            bbox_width = gt_bboxes[:,2] - gt_bboxes[:,0]
+            bbox_height = gt_bboxes[:,3] - gt_bboxes[:,1]
+            bbox_area = bbox_width * bbox_height
+            gt_size = np.sqrt(bbox_area+0.0001)
+            gts_size += list(gt_size)
+            anchor_cxs = []
+            anchor_cys = []
+            for idx, stride in enumerate([8,16,32,64,128]):
+                height = input_height // stride
+                width = input_width // stride
+                anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+                anchor_centers = (anchor_centers * stride).reshape( (-1, 2) )
+                anchor_cx = anchor_centers[:,0]
+                anchor_cy = anchor_centers[:,1]
+                anchor_cxs += list(anchor_cx)
+                anchor_cys += list(anchor_cy)
+            anchor_cx = np.array(anchor_cxs, dtype=np.float32)
+            anchor_cy = np.array(anchor_cys, dtype=np.float32)
+
+            num_gts = gt_bboxes.shape[0]
+            num_anchors = anchor_cx.shape[0]
+            anchor_cx = np.broadcast_to(anchor_cx.reshape((1,-1)), (num_gts, num_anchors)).reshape(num_anchors, num_gts)
+            anchor_cy = np.broadcast_to(anchor_cy.reshape((1,-1)), (num_gts, num_anchors)).reshape(num_anchors, num_gts)
+            gt_x1 = gt_bboxes[:,0]
+            gt_y1 = gt_bboxes[:,1]
+            gt_x2 = gt_bboxes[:,2]
+            gt_y2 = gt_bboxes[:,3]
+            gt_cover = np.zeros( (gt_bboxes.shape[0], ), dtype=np.float32)
+            l_ = anchor_cx - gt_x1
+            t_ = anchor_cy - gt_y1
+            r_ = gt_x2 - anchor_cx
+            b_ = gt_y2 - anchor_cy
+            dist = np.stack([l_, t_, r_, b_], axis=1).min(axis=1)
+            gt_dist = dist.max(axis=0)
+            gt_dist  = gt_dist / gt_size
+            center_thres = 0.01
+            #center_thres = -0.25
+            gt_cover_inds = np.where(gt_dist>center_thres)[0]
+            num_assigned = len(gt_cover_inds)
+            assign_stat[0] += num_gts
+            assign_stat[1] += num_assigned
+            
+
+
+
+
+        for _ in range(batch_size):
+            prog_bar.update()
+    aps = wider_evaluation(results, gt_path, 0.5, args.debug)
+    with open(os.path.join(output_folder, 'aps'), 'w') as f:
+        f.write("%f,%f,%f\n"%(aps[0],aps[1],aps[2]))
+    print('APS:', aps)
+    if args.show_assign:
+        print('ASSIGN:', assign_stat)
+        gts_size = np.array(gts_size, dtype=np.float32)
+        gts_size = np.sort(gts_size)
+        assert len(gts_size)==assign_stat[0]
+        print(gts_size[assign_stat[0]//2])
+
+
+if __name__ == '__main__':
+    main()
diff --git a/detection/scrfd/tools/train.py b/detection/scrfd/tools/train.py
new file mode 100755
index 000000000..e7c215481
--- /dev/null
+++ b/detection/scrfd/tools/train.py
@@ -0,0 +1,182 @@
+import argparse
+import copy
+import os
+import os.path as osp
+import time
+import warnings
+
+import mmcv
+import torch
+from mmcv import Config, DictAction
+from mmcv.runner import get_dist_info, init_dist
+from mmcv.utils import get_git_hash
+
+from mmdet import __version__
+from mmdet.apis import set_random_seed, train_detector
+from mmdet.datasets import build_dataset
+from mmdet.models import build_detector
+from mmdet.utils import collect_env, get_root_logger
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train a detector')
+    parser.add_argument('config', help='train config file path')
+    parser.add_argument('--work-dir', help='the dir to save logs and models')
+    parser.add_argument(
+        '--resume-from', help='the checkpoint file to resume from')
+    parser.add_argument(
+        '--no-validate',
+        action='store_true',
+        help='whether not to evaluate the checkpoint during training')
+    group_gpus = parser.add_mutually_exclusive_group()
+    group_gpus.add_argument(
+        '--gpus',
+        type=int,
+        help='number of gpus to use '
+        '(only applicable to non-distributed training)')
+    group_gpus.add_argument(
+        '--gpu-ids',
+        type=int,
+        nargs='+',
+        help='ids of gpus to use '
+        '(only applicable to non-distributed training)')
+    parser.add_argument('--seed', type=int, default=None, help='random seed')
+    parser.add_argument(
+        '--deterministic',
+        action='store_true',
+        help='whether to set deterministic options for CUDNN backend.')
+    parser.add_argument(
+        '--options',
+        nargs='+',
+        action=DictAction,
+        help='override some settings in the used config, the key-value pair '
+        'in xxx=yyy format will be merged into config file (deprecate), '
+        'change to --cfg-options instead.')
+    parser.add_argument(
+        '--cfg-options',
+        nargs='+',
+        action=DictAction,
+        help='override some settings in the used config, the key-value pair '
+        'in xxx=yyy format will be merged into config file.')
+    parser.add_argument(
+        '--launcher',
+        choices=['none', 'pytorch', 'slurm', 'mpi'],
+        default='none',
+        help='job launcher')
+    parser.add_argument('--local_rank', type=int, default=0)
+    args = parser.parse_args()
+    if 'LOCAL_RANK' not in os.environ:
+        os.environ['LOCAL_RANK'] = str(args.local_rank)
+
+    if args.options and args.cfg_options:
+        raise ValueError(
+            '--options and --cfg-options cannot be both '
+            'specified, --options is deprecated in favor of --cfg-options')
+    if args.options:
+        warnings.warn('--options is deprecated in favor of --cfg-options')
+        args.cfg_options = args.options
+
+    return args
+
+
+def main():
+    args = parse_args()
+
+    cfg = Config.fromfile(args.config)
+    if args.cfg_options is not None:
+        cfg.merge_from_dict(args.cfg_options)
+    # import modules from string list.
+    if cfg.get('custom_imports', None):
+        from mmcv.utils import import_modules_from_strings
+        import_modules_from_strings(**cfg['custom_imports'])
+    # set cudnn_benchmark
+    if cfg.get('cudnn_benchmark', False):
+        torch.backends.cudnn.benchmark = True
+
+    # work_dir is determined in this priority: CLI > segment in file > filename
+    if args.work_dir is not None:
+        # update configs according to CLI args if args.work_dir is not None
+        cfg.work_dir = args.work_dir
+    elif cfg.get('work_dir', None) is None:
+        # use config filename as default work_dir if cfg.work_dir is None
+        cfg.work_dir = osp.join('./work_dirs',
+                                osp.splitext(osp.basename(args.config))[0])
+    if args.resume_from is not None:
+        cfg.resume_from = args.resume_from
+    if args.gpu_ids is not None:
+        cfg.gpu_ids = args.gpu_ids
+    else:
+        cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus)
+
+    # init distributed env first, since logger depends on the dist info.
+    if args.launcher == 'none':
+        distributed = False
+    else:
+        distributed = True
+        init_dist(args.launcher, **cfg.dist_params)
+        # re-set gpu_ids with distributed training mode
+        _, world_size = get_dist_info()
+        cfg.gpu_ids = range(world_size)
+
+    # create work_dir
+    mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
+    # dump config
+    cfg.dump(osp.join(cfg.work_dir, osp.basename(args.config)))
+    # init the logger before other steps
+    timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
+    log_file = osp.join(cfg.work_dir, f'{timestamp}.log')
+    logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
+
+    # init the meta dict to record some important information such as
+    # environment info and seed, which will be logged
+    meta = dict()
+    # log env info
+    env_info_dict = collect_env()
+    env_info = '\n'.join([(f'{k}: {v}') for k, v in env_info_dict.items()])
+    dash_line = '-' * 60 + '\n'
+    logger.info('Environment info:\n' + dash_line + env_info + '\n' +
+                dash_line)
+    meta['env_info'] = env_info
+    meta['config'] = cfg.pretty_text
+    # log some basic info
+    logger.info(f'Distributed training: {distributed}')
+    logger.info(f'Config:\n{cfg.pretty_text}')
+
+    # set random seeds
+    if args.seed is not None:
+        logger.info(f'Set random seed to {args.seed}, '
+                    f'deterministic: {args.deterministic}')
+        set_random_seed(args.seed, deterministic=args.deterministic)
+    cfg.seed = args.seed
+    meta['seed'] = args.seed
+    meta['exp_name'] = osp.basename(args.config)
+
+    model = build_detector(
+        cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
+
+    datasets = [build_dataset(cfg.data.train)]
+    #print('ds count', len(datasets))
+    if len(cfg.workflow) == 2:
+        val_dataset = copy.deepcopy(cfg.data.val)
+        val_dataset.pipeline = cfg.data.train.pipeline
+        datasets.append(build_dataset(val_dataset))
+    if cfg.checkpoint_config is not None:
+        # save mmdet version, config file content and class names in
+        # checkpoints as meta data
+        cfg.checkpoint_config.meta = dict(
+            mmdet_version=__version__ + get_git_hash()[:7],
+            CLASSES=datasets[0].CLASSES)
+    # add an attribute for visualization convenience
+    model.CLASSES = datasets[0].CLASSES
+    train_detector(
+        model,
+        datasets,
+        cfg,
+        distributed=distributed,
+        validate=(not args.no_validate),
+        timestamp=timestamp,
+        meta=meta)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/examples/README.md b/examples/README.md
new file mode 100644
index 000000000..a414efb79
--- /dev/null
+++ b/examples/README.md
@@ -0,0 +1,4 @@
+InsightFace Example
+---
+
+Before running the examples, please install insightface package via `pip install -U insightface`
diff --git a/examples/demo_analysis.py b/examples/demo_analysis.py
new file mode 100644
index 000000000..fd7bf52f5
--- /dev/null
+++ b/examples/demo_analysis.py
@@ -0,0 +1,34 @@
+import argparse
+import cv2
+import sys
+import numpy as np
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+assert insightface.__version__>='0.3'
+
+parser = argparse.ArgumentParser(description='insightface app test')
+# general
+parser.add_argument('--ctx', default=0, type=int, help='ctx id, <0 means using cpu')
+parser.add_argument('--det-size', default=0, type=int, help='detection size; 0 means Auto (128 + 640)')
+args = parser.parse_args()
+
+app = FaceAnalysis()
+det_size = None if args.det_size <= 0 else (args.det_size, args.det_size)
+app.prepare(ctx_id=args.ctx, det_size=det_size)
+
+img = ins_get_image('t1')
+faces = app.get(img)
+assert len(faces)==6
+rimg = app.draw_on(img, faces)
+cv2.imwrite("./t1_output.jpg", rimg)
+
+# then print all-to-all face similarity
+feats = []
+for face in faces:
+    feats.append(face.normed_embedding)
+feats = np.array(feats, dtype=np.float32)
+sims = np.dot(feats, feats.T)
+print(sims)
+
diff --git a/examples/edge_inference/README.md b/examples/edge_inference/README.md
new file mode 100644
index 000000000..68fdd16aa
--- /dev/null
+++ b/examples/edge_inference/README.md
@@ -0,0 +1,33 @@
+# InsightFace Edge Inference and Deployment
+
+In this tutorial, we give examples and benchmarks of running insightface models on edge devices, mainly using 8-bits quantization technologies to make acceleration.
+
+## Recognition
+
+In recognition tutorial, we use an open-source model: *IR50@Glint360K*, and use a hard private 1:N testset(N=50000). The metric contains Rank1 and TAR@FAR<=e-3.
+
+
+
+Granularity and symmetry both stand for quantization setting, and mostly defined by hardware providers. Symmetric uses INT8 to save quantization results while Asymmetric uses UINT8 type.
+
+| Hardware    | Provider | Type | Backend     | Time | Granularity | Symmetry   | Rank1-Acc | TAR@FAR<=e-3 |
+| ----------- | -------- | ---- | ----------- | ---- | ----------- | ---------- | --------- | ------------ |
+| V100        | NVIDIA   | GPU  | onnxruntime | 4ms  | -           | -          | 80.94     | 30.77        |
+| Jetson NX   | NVIDIA   | GPU  | TensorRT    | 16ms | Per-channel | Symmetric  | 79.26     | 31.07        |
+| A311D       | Khadas   | ASIC | Tengine     | 26ms | Per-tensor  | Asymmetric | 77.83     | 26.58        |
+| A311D*      | Khadas   | ASIC | Tengine     | 26ms | Per-tensor  | Asymmetric | 79.38     | 28.59        |
+| NXP-IMX8P   | NXP      | ASIC | Tengine     | 24ms | Per-tensor  | Asymmetric | 77.87     | 26.80        |
+| NXP-IMX8P*  | NXP      | ASIC | Tengine     | 24ms | Per-tensor  | Asymmetric | 79.42     | 28.39        |
+| RV1126      | Rockchip | ASIC | RKNN        | 38ms | Per-tensor  | Asymmetric | 75.60     | 24.23        |
+| RV1126*     | Rockchip | ASIC | RKNN        | 38ms | Per-tensor  | Asymmetric | 77.82     | 26.30        |
+
+Suffix-* means mixed mode: using float32 model for gallery while using quantized model for probe images. Result features are all in float32 type.
+
+The example code of running quantized networks can be now found at [Tengine](https://github.com/OAID/Tengine/tree/tengine-lite/demos). Later, we will put a copy here and give full tutorial on how to quantize recognition models from 0 to 1.
+
+
+
+## Detection
+
+TODO
+
diff --git a/examples/face_detection/README.md b/examples/face_detection/README.md
new file mode 100644
index 000000000..d0942a86b
--- /dev/null
+++ b/examples/face_detection/README.md
@@ -0,0 +1,2 @@
+
+#
diff --git a/examples/face_recognition/README.md b/examples/face_recognition/README.md
new file mode 100644
index 000000000..d0942a86b
--- /dev/null
+++ b/examples/face_recognition/README.md
@@ -0,0 +1,2 @@
+
+#
diff --git a/examples/face_recognition/insightface_app.py b/examples/face_recognition/insightface_app.py
new file mode 100644
index 000000000..d5f515ea9
--- /dev/null
+++ b/examples/face_recognition/insightface_app.py
@@ -0,0 +1,46 @@
+import cv2
+import insightface
+import numpy as np
+from insightface.app import FaceAnalysis
+
+# Initialize face analysis model
+app = FaceAnalysis(name='buffalo_l', providers=['CPUExecutionProvider'])  # Use 'CUDAExecutionProvider' for GPU
+app.prepare(ctx_id=-1)  # ctx_id=-1 for CPU, 0 for GPU
+
+def get_face_embedding(image_path):
+    """Extract face embedding from an image"""
+    img = cv2.imread(image_path)
+    if img is None:
+        raise ValueError(f"Could not read image: {image_path}")
+    
+    faces = app.get(img)
+    
+    if len(faces) < 1:
+        raise ValueError("No faces detected in the image")
+    if len(faces) > 1:
+        print("Warning: Multiple faces detected. Using first detected face")
+    
+    return faces[0].embedding
+
+def compare_faces(emb1, emb2, threshold=0.65): # Adjust this threshold according to your usecase.
+    """Compare two embeddings using cosine similarity"""
+    similarity = np.dot(emb1, emb2) / (np.linalg.norm(emb1) * np.linalg.norm(emb2))
+    return similarity, similarity > threshold
+
+# Paths to your Indian face images
+image1_path = "path/to/face1.jpg"
+image2_path = "path/to/face2.jpg"
+
+try:
+    # Get embeddings
+    emb1 = get_face_embedding(image1_path)
+    emb2 = get_face_embedding(image2_path)
+    
+    # Compare faces
+    similarity_score, is_same_person = compare_faces(emb1, emb2)
+    
+    print(f"Similarity Score: {similarity_score:.4f}")
+    print(f"Same person? {'YES' if is_same_person else 'NO'}")
+    
+except Exception as e:
+    print(f"Error: {str(e)}")
diff --git a/examples/in_swapper/README.md b/examples/in_swapper/README.md
new file mode 100644
index 000000000..ce8f1965c
--- /dev/null
+++ b/examples/in_swapper/README.md
@@ -0,0 +1,44 @@
+# InsightFace Swapper
+
+## Update
+
+Please use our [Picsi.Ai face swapping product](https://www.picsi.ai) instead for higher resolution face swapping. This model and demo code is no longer maintained.
+
+
+In this example, we provide one-line simple code for subject agnostic identity transfer from source face to the target face.
+
+The input and output resolution of this tool is 128x128.
+
+
+## Usage
+
+Firstly install insightface python library, with version>=0.7:
+
+```
+pip install -U insightface
+```
+
+Second, download the `inswapper_128.onnx` swapping model from [googledrive]() and put it under `~/.insightface/models/`.
+
+Then use the recognition model from our `buffalo_l` pack and initialize the INSwapper class. 
+
+Note that now we can only accept latent embedding from the `buffalo_l` arcface model, otherwise the result will be not normal.
+
+For detail code, please check the [example](inswapper_main.py).
+
+## Result:
+
+Input: 
+
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/t1.jpg" width="640" />
+
+---Then we change the identity to Ross for all faces in this image.---
+
+Direct Outputs:
+
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/t1_swapped2.jpg" width="640" />
+
+Paste Back:
+
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/t1_swapped.jpg" width="640" />
+
diff --git a/examples/in_swapper/inswapper_main.py b/examples/in_swapper/inswapper_main.py
new file mode 100644
index 000000000..55d64666a
--- /dev/null
+++ b/examples/in_swapper/inswapper_main.py
@@ -0,0 +1,35 @@
+import datetime
+import numpy as np
+import os
+import os.path as osp
+import glob
+import cv2
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+
+assert insightface.__version__>='0.7'
+
+if __name__ == '__main__':
+    app = FaceAnalysis(name='buffalo_l')
+    app.prepare(ctx_id=0)
+    swapper = insightface.model_zoo.get_model('inswapper_128.onnx', download=True, download_zip=True)
+
+
+    img = ins_get_image('t1')
+    faces = app.get(img)
+    faces = sorted(faces, key = lambda x : x.bbox[0])
+    assert len(faces)==6
+    source_face = faces[2]
+    res = img.copy()
+    for face in faces:
+        res = swapper.get(res, face, source_face, paste_back=True)
+    cv2.imwrite("./t1_swapped.jpg", res)
+    res = []
+    for face in faces:
+        _img, _ = swapper.get(img, face, source_face, paste_back=False)
+        res.append(_img)
+    res = np.concatenate(res, axis=1)
+    cv2.imwrite("./t1_swapped2.jpg", res)
+
diff --git a/examples/mask_renderer.py b/examples/mask_renderer.py
new file mode 100644
index 000000000..5b047423b
--- /dev/null
+++ b/examples/mask_renderer.py
@@ -0,0 +1,22 @@
+import os, sys, datetime
+import numpy as np
+import os.path as osp
+import cv2
+import insightface
+from insightface.app import MaskRenderer
+from insightface.data import get_image as ins_get_image
+
+
+if __name__ == "__main__":
+    #make sure that you have download correct insightface model pack.
+    #make sure that BFM.mat and BFM_UV.mat have been generated
+    tool = MaskRenderer()
+    tool.prepare(ctx_id=0, det_size=(128,128))
+    image = ins_get_image('Tom_Hanks_54745')
+    mask_image  = "mask_blue"
+    params = tool.build_params(image)
+    mask_out = tool.render_mask(image, mask_image, params)
+
+    cv2.imwrite('output_mask.jpg', mask_out)
+
+
diff --git a/examples/mxnet_to_onnx.py b/examples/mxnet_to_onnx.py
new file mode 100644
index 000000000..e1fd04b15
--- /dev/null
+++ b/examples/mxnet_to_onnx.py
@@ -0,0 +1,178 @@
+import sys
+import os
+import argparse
+import onnx
+import json
+import mxnet as mx
+from onnx import helper
+from onnx import TensorProto
+from onnx import numpy_helper
+import onnxruntime
+import cv2
+
+print('mxnet version:', mx.__version__)
+print('onnx version:', onnx.__version__)
+
+assert mx.__version__ >= '1.8', 'mxnet version should >= 1.8'
+assert onnx.__version__ >= '1.2.1', 'onnx version should >= 1.2.1'
+
+import numpy as np
+from mxnet.contrib import onnx as onnx_mxnet
+
+def create_map(graph_member_list):
+    member_map={}
+    for n in graph_member_list:
+        member_map[n.name]=n
+    return member_map
+
+
+parser = argparse.ArgumentParser(description='convert mxnet model to onnx')
+# general
+parser.add_argument('params', default='./r100a/model-0000.params', help='mxnet params to load.')
+parser.add_argument('output', default='./r100a.onnx', help='path to write onnx model.')
+parser.add_argument('--eps', default=1.0e-8, type=float, help='eps for weights.')
+parser.add_argument('--input-shape', default='3,112,112', help='input shape.')
+parser.add_argument('--check', action='store_true')
+parser.add_argument('--batch', action='store_true')
+parser.add_argument('--input-mean', default=0.0, type=float, help='input mean for checking.')
+parser.add_argument('--input-std', default=1.0, type=float, help='input std for checking.')
+args = parser.parse_args()
+input_shape = (1,) + tuple( [int(x) for x in args.input_shape.split(',')] )
+
+params_file = args.params
+pos = params_file.rfind('-')
+prefix = params_file[:pos]
+epoch = int(params_file[pos+1:pos+5])
+sym_file = prefix + "-symbol.json"
+assert os.path.exists(sym_file)
+assert os.path.exists(params_file)
+
+sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+
+nodes = json.loads(sym.tojson())['nodes']
+bn_fixgamma_list = []
+for nodeid, node in enumerate(nodes):
+    if node['op'] == 'BatchNorm':
+        attr = node['attrs']
+        fix_gamma = False
+        if attr is not None and 'fix_gamma' in attr:
+            if str(attr['fix_gamma']).lower()=='true':
+                fix_gamma = True
+        if fix_gamma:
+            bn_fixgamma_list.append(node['name'])
+        #print(node, fix_gamma)
+
+print('fixgamma list:', bn_fixgamma_list)
+layer = None
+#layer = 'conv_2_dw_relu' #for debug
+
+if layer is not None:
+    all_layers = sym.get_internals()
+    sym = all_layers[layer + '_output']
+
+
+eps = args.eps
+
+arg = {}
+aux = {}
+invalid = 0
+ac = 0
+for k in arg_params:
+    v = arg_params[k]
+    nv = v.asnumpy()
+    nv = nv.astype(np.float32)
+    #print(k, nv.shape)
+    if k.endswith('_gamma'):
+        bnname = k[:-6]
+        if bnname in bn_fixgamma_list:
+            nv[:] = 1.0
+    ac += nv.size
+    invalid += np.count_nonzero(np.abs(nv)<eps)
+    nv[np.abs(nv) < eps] = 0.0
+    arg[k] = mx.nd.array(nv, dtype='float32')
+arg_params = arg
+invalid = 0
+ac = 0
+for k in aux_params:
+    v = aux_params[k]
+    nv = v.asnumpy().astype(np.float32)
+
+    ac += nv.size
+    invalid += np.count_nonzero(np.abs(nv)<eps)
+    nv[np.abs(nv) < eps] = 0.0
+    aux[k] = mx.nd.array(nv, dtype='float32')
+aux_params = aux
+
+all_args = {}
+all_args.update(arg_params)
+all_args.update(aux_params)
+converted_model_path = onnx_mxnet.export_model(sym, all_args, [input_shape], np.float32, args.output, opset_version=11)
+
+
+model = onnx.load(args.output)
+graph = model.graph
+input_map = create_map(graph.input)
+node_map = create_map(graph.node)
+init_map = create_map(graph.initializer)
+
+#fix PRelu issue
+for input_name in input_map.keys():
+    if input_name.endswith('_gamma'):
+        node_name = input_name[:-6]
+        if not node_name in node_map:
+            continue
+        node = node_map[node_name]
+        if node.op_type!='PRelu':
+            continue
+        _input_shape = input_map[input_name].type.tensor_type.shape.dim
+        input_dim_val=_input_shape[0].dim_value
+        
+        graph.initializer.remove(init_map[input_name])
+        weight_array = numpy_helper.to_array(init_map[input_name])
+        
+        b=[]
+        for w in weight_array:
+            b.append(w)
+        new_nv = helper.make_tensor(input_name, TensorProto.FLOAT, [input_dim_val,1,1], b)
+        graph.initializer.extend([new_nv])
+
+for init_name in init_map.keys():
+    weight_array = numpy_helper.to_array(init_map[init_name])
+    assert weight_array.dtype==np.float32
+    if init_name in input_map:
+        graph.input.remove(input_map[init_name])
+
+#support batch-inference
+if args.batch:
+    graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+
+onnx.save(model, args.output)
+
+#start to check correctness
+if args.check:
+    im_size = tuple(input_shape[2:])+(3,)
+    img = np.random.randint(0, 256, size=im_size, dtype=np.uint8)
+    input_size = tuple(input_shape[2:4][::-1])
+    input_std = args.input_std
+    input_mean = args.input_mean
+    #print(img.shape, input_size)
+    img = cv2.dnn.blobFromImage(img, 1.0/input_std, input_size, (input_mean, input_mean, input_mean), swapRB=True)
+    ctx = mx.cpu()
+    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
+    model.bind(for_training=False, data_shapes=[('data', input_shape)])
+    _, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch) #reload original params
+    model.set_params(arg_params, aux_params)
+
+    data = mx.nd.array(img)
+    db = mx.io.DataBatch(data=(data,))
+    model.forward(db, is_train=False)
+    x1 = model.get_outputs()[-1].asnumpy()
+
+    session = onnxruntime.InferenceSession(args.output, None)
+    input_name = session.get_inputs()[0].name
+    output_name = session.get_outputs()[0].name
+    x2 = session.run([output_name], {input_name : img})[0]
+    print(x1.shape, x2.shape)
+    print(x1.flatten()[:20])
+    print(x2.flatten()[:20])
+
diff --git a/examples/person_detection/README.md b/examples/person_detection/README.md
new file mode 100644
index 000000000..bd699b234
--- /dev/null
+++ b/examples/person_detection/README.md
@@ -0,0 +1,38 @@
+# Person Detection
+
+This person detection example is built by [SCRFD](../../detection/scrfd) approch.
+
+## Usage
+
+Firstly install insightface python library:
+
+```
+pip install -U insightface
+```
+
+and then load our person detection model by:
+
+```
+detector = insightface.model_zoo.get_model('scrfd_person_2.5g.onnx', download=True)
+detector.prepare(0, nms_thresh=0.5)
+```
+
+the model will be auto-downloaded from our storage server.
+
+## Detection Result:
+
+In this example, we support full-body detection and recognize the corresponding visible region in a single forward pass.
+
+Please see [scrfd_person.py](scrfd_person.py) for detail on how to visualize the results.
+
+The green bounding box shows the full-body while the blue mask indicates the visible region.
+
+(We make tests on the input size of 640)
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/283554,c2d0000d40862ba.jpg" width="640" />
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/283647,18e170005675c161.jpg" width="640" />
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/283554,2290700005b7d575.jpg" width="640" />
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/283554,175820000e7255da.jpg" width="640" />
diff --git a/examples/person_detection/scrfd_person.py b/examples/person_detection/scrfd_person.py
new file mode 100644
index 000000000..f50040bb5
--- /dev/null
+++ b/examples/person_detection/scrfd_person.py
@@ -0,0 +1,48 @@
+import datetime
+import numpy as np
+import os
+import os.path as osp
+import glob
+import cv2
+import insightface
+
+assert insightface.__version__>='0.4'
+
+
+def detect_person(img, detector):
+    bboxes, kpss = detector.detect(img)
+    bboxes = np.round(bboxes[:,:4]).astype(np.int)
+    kpss = np.round(kpss).astype(np.int)
+    kpss[:,:,0] = np.clip(kpss[:,:,0], 0, img.shape[1])
+    kpss[:,:,1] = np.clip(kpss[:,:,1], 0, img.shape[0])
+    vbboxes = bboxes.copy()
+    vbboxes[:,0] = kpss[:, 0, 0]
+    vbboxes[:,1] = kpss[:, 0, 1]
+    vbboxes[:,2] = kpss[:, 4, 0]
+    vbboxes[:,3] = kpss[:, 4, 1]
+    return bboxes, vbboxes
+
+if __name__ == '__main__':
+    import glob
+    detector = insightface.model_zoo.get_model('scrfd_person_2.5g.onnx', download=True)
+    detector.prepare(0, nms_thresh=0.5)
+    img_paths = glob.glob('data/images/*.jpg')
+    for img_path in img_paths:
+        img = cv2.imread(img_path)
+        bboxes, vbboxes = detect_person(img, detector)
+        for i in range(bboxes.shape[0]):
+            bbox = bboxes[i]
+            vbbox = vbboxes[i]
+            x1,y1,x2,y2 = bbox
+            vx1,vy1,vx2,vy2 = vbbox
+            cv2.rectangle(img, (x1,y1)  , (x2,y2) , (0,255,0) , 1)
+            alpha = 0.8
+            color = (255, 0, 0)
+            for c in range(3):
+                img[vy1:vy2,vx1:vx2,c] = img[vy1:vy2, vx1:vx2, c]*alpha + color[c]*(1.0-alpha)
+            cv2.circle(img, (vx1,vy1) , 1, color , 2)
+            cv2.circle(img, (vx1,vy2) , 1, color , 2)
+            cv2.circle(img, (vx2,vy1) , 1, color , 2)
+            cv2.circle(img, (vx2,vy2) , 1, color , 2)
+        filename = img_path.split('/')[-1]
+        cv2.imwrite('./outputs/%s'%filename, img)
diff --git a/gender-age/Tom_Hanks_54745.png b/gender-age/Tom_Hanks_54745.png
deleted file mode 100644
index 906315d13..000000000
Binary files a/gender-age/Tom_Hanks_54745.png and /dev/null differ
diff --git a/gender-age/data.py b/gender-age/data.py
deleted file mode 100644
index 21441f45c..000000000
--- a/gender-age/data.py
+++ /dev/null
@@ -1,276 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-from PIL import Image
-from io import BytesIO
-
-import mxnet as mx
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-
-logger = logging.getLogger()
-
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0, color_jittering = 0,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        logging.info('loading recordio %s...',
-                     path_imgrec)
-        path_imgidx = path_imgrec[0:-4]+".idx"
-        self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-        s = self.imgrec.read_idx(0)
-        header, _ = recordio.unpack(s)
-        self.imgidx = list(self.imgrec.keys)
-        self.seq = self.imgidx
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        self.color_jittering = color_jittering
-        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
-        self.provide_label = [(label_name, (batch_size,101))]
-        #print(self.provide_label[0][1])
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.shuffle:
-          random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        if self.cur >= len(self.seq):
-            raise StopIteration
-        idx = self.seq[self.cur]
-        self.cur += 1
-        s = self.imgrec.read_idx(idx)
-        header, img = recordio.unpack(s)
-        label = header.label
-        return label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = nd.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      #random.shuffle(augs)
-      #for aug in augs:
-      #  #print(img.shape)
-      #  img = aug(img, x)
-      #  #print(img.shape)
-      #return img
-      return self.CJA(img)
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in range(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-    def compress_aug(self, img):
-      buf = BytesIO()
-      img = Image.fromarray(img.asnumpy(), 'RGB')
-      q = random.randint(2, 20)
-      img.save(buf, format='JPEG', quality=q)
-      buf = buf.getvalue()
-      img = Image.open(BytesIO(buf))
-      return nd.array(np.asarray(img, 'float32'))
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                #print('XXXX', i)
-                label, s, bbox, landmark = self.next_sample()
-                gender = int(label[0])
-                age = int(label[1])
-                assert age>=0
-                #assert gender==0 or gender==1
-                plabel = np.zeros(shape=(101,), dtype=np.float32)
-                plabel[0] = gender
-                if age==0:
-                  age = 1
-                if age>100:
-                  age = 100
-                plabel[1:age+1] = 1
-                label = plabel
-                _data = self.imdecode(s)
-                if _data.shape[0]!=self.data_shape[1]:
-                  _data = mx.image.resize_short(_data, self.data_shape[1])
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.color_jittering>0:
-                  if self.color_jittering>1:
-                    _rd = random.randint(0,1)
-                    if _rd==1:
-                      _data = self.compress_aug(_data)
-                  #print('do color aug')
-                  _data = _data.astype('float32', copy=False)
-                  #print(_data.__class__)
-                  _data = self.color_aug(_data, 0.125)
-                if self.nd_mean is not None:
-                  _data = _data.astype('float32', copy=False)
-                  _data -= self.nd_mean
-                  _data *= 0.0078125
-                if self.cutoff>0:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    #print('do cutoff aug', self.cutoff)
-                    centerh = random.randint(0, _data.shape[0]-1)
-                    centerw = random.randint(0, _data.shape[1]-1)
-                    half = self.cutoff//2
-                    starth = max(0, centerh-half)
-                    endh = min(_data.shape[0], centerh+half)
-                    startw = max(0, centerw-half)
-                    endw = min(_data.shape[1], centerw+half)
-                    #print(starth, endh, startw, endw, _data.shape)
-                    _data[starth:endh, startw:endw, :] = 128
-                data = [_data]
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        return io.DataBatch([batch_data], [batch_label], batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/gender-age/face_model.py b/gender-age/face_model.py
deleted file mode 100644
index 1212a25e6..000000000
--- a/gender-age/face_model.py
+++ /dev/null
@@ -1,99 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-#import tensorflow as tf
-import numpy as np
-import mxnet as mx
-import random
-import cv2
-import sklearn
-from sklearn.decomposition import PCA
-from time import sleep
-from easydict import EasyDict as edict
-from mtcnn_detector import MtcnnDetector
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'src', 'common'))
-import face_image
-import face_preprocess
-
-
-def do_flip(data):
-  for idx in range(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_model(ctx, image_size, model_str, layer):
-  _vec = model_str.split(',')
-  assert len(_vec)==2
-  prefix = _vec[0]
-  epoch = int(_vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers[layer+'_output']
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-  model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  return model
-
-class FaceModel:
-  def __init__(self, args):
-    self.args = args
-    if args.gpu>=0:
-      ctx = mx.gpu(args.gpu)
-    else:
-      ctx = mx.cpu()
-    _vec = args.image_size.split(',')
-    assert len(_vec)==2
-    image_size = (int(_vec[0]), int(_vec[1]))
-    self.model = None
-    if len(args.model)>0:
-      self.model = get_model(ctx, image_size, args.model, 'fc1')
-
-    self.det_minsize = 50
-    self.det_threshold = [0.6,0.7,0.8]
-    #self.det_factor = 0.9
-    self.image_size = image_size
-    mtcnn_path = os.path.join(os.path.dirname(__file__), 'mtcnn-model')
-    if args.det==0:
-      detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=self.det_threshold)
-    else:
-      detector = MtcnnDetector(model_folder=mtcnn_path, ctx=ctx, num_worker=1, accurate_landmark = True, threshold=[0.0,0.0,0.2])
-    self.detector = detector
-
-
-  def get_input(self, face_img):
-    ret = self.detector.detect_face(face_img, det_type = self.args.det)
-    if ret is None:
-      return None
-    bbox, points = ret
-    if bbox.shape[0]==0:
-      return None
-    bbox = bbox[0,0:4]
-    points = points[0,:].reshape((2,5)).T
-    #print(bbox)
-    #print(points)
-    nimg = face_preprocess.preprocess(face_img, bbox, points, image_size='112,112')
-    nimg = cv2.cvtColor(nimg, cv2.COLOR_BGR2RGB)
-    aligned = np.transpose(nimg, (2,0,1))
-    input_blob = np.expand_dims(aligned, axis=0)
-    data = mx.nd.array(input_blob)
-    db = mx.io.DataBatch(data=(data,))
-    return db
-
-
-  def get_ga(self, data):
-    self.model.forward(data, is_train=False)
-    ret = self.model.get_outputs()[0].asnumpy()
-    g = ret[:,0:2].flatten()
-    gender = np.argmax(g)
-    a = ret[:,2:202].reshape( (100,2) )
-    a = np.argmax(a, axis=1)
-    age = int(sum(a))
-
-    return gender, age
-
diff --git a/gender-age/fmobilenet.py b/gender-age/fmobilenet.py
deleted file mode 100644
index d2bdeb8aa..000000000
--- a/gender-age/fmobilenet.py
+++ /dev/null
@@ -1,85 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-import mxnet as mx
-import symbol_utils
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    #body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    body = mx.sym.Activation(data=data, act_type='relu', name=name)
-    return body
-
-def Conv(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=True)
-    act = Act(data=bn, act_type='relu', name='%s%s_relu' %(name, suffix))
-    return act
-
-def ConvOnly(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    return conv
-
-def get_symbol(num_classes, **kwargs):
-    data = mx.symbol.Variable(name="data") # 224
-    data = data-127.5
-    data = data*0.0078125
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    multiplier = kwargs.get('multiplier', 1.0)
-    fc_type = version_output
-    base_filter = int(32*multiplier)
-    bf = base_filter
-    print(version_input, version_output, base_filter)
-
-    if version_input==0:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_1") # 224/112
-    else:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_1") # 224/112
-    conv_2_dw = Conv(conv_1, num_group=bf, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_2_dw") # 112/112
-    conv_2 = Conv(conv_2_dw, num_filter=bf*2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_2") # 112/112
-    conv_3_dw = Conv(conv_2, num_group=bf*2, num_filter=bf*2, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_3_dw") # 112/56
-    conv_3 = Conv(conv_3_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_3") # 56/56
-    conv_4_dw = Conv(conv_3, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_4_dw") # 56/56
-    conv_4 = Conv(conv_4_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_4") # 56/56
-    conv_5_dw = Conv(conv_4, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_5_dw") # 56/28
-    conv_5 = Conv(conv_5_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_5") # 28/28
-    conv_6_dw = Conv(conv_5, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_6_dw") # 28/28
-    conv_6 = Conv(conv_6_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_6") # 28/28
-    conv_7_dw = Conv(conv_6, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_7_dw") # 28/14
-    conv_7 = Conv(conv_7_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_7") # 14/14
-
-    conv_8_dw = Conv(conv_7, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_8_dw") # 14/14
-    conv_8 = Conv(conv_8_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_8") # 14/14
-    conv_9_dw = Conv(conv_8, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_9_dw") # 14/14
-    conv_9 = Conv(conv_9_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_9") # 14/14
-    conv_10_dw = Conv(conv_9, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_10_dw") # 14/14
-    conv_10 = Conv(conv_10_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_10") # 14/14
-    conv_11_dw = Conv(conv_10, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_11_dw") # 14/14
-    conv_11 = Conv(conv_11_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_11") # 14/14
-    conv_12_dw = Conv(conv_11, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_12_dw") # 14/14
-    conv_12 = Conv(conv_12_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_12") # 14/14
-
-    conv_13_dw = Conv(conv_12, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_13_dw") # 14/7
-    conv_13 = Conv(conv_13_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_13") # 7/7
-    conv_14_dw = Conv(conv_13, num_group=bf*32, num_filter=bf*32, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_14_dw") # 7/7
-    conv_14 = Conv(conv_14_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_14") # 7/7
-    body = conv_14
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
diff --git a/gender-age/fresnet.py b/gender-age/fresnet.py
deleted file mode 100644
index d4844b7dd..000000000
--- a/gender-age/fresnet.py
+++ /dev/null
@@ -1,601 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-'''
-Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
-Original author Wei Wu
-
-Implemented the following paper:
-
-Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
-'''
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import mxnet as mx
-import numpy as np
-import symbol_utils
-import sklearn
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-
-def Act(data, act_type, name):
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=stride, pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit2')
-    if bottle_neck:
-        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=stride, pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
-        conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv3 = mx.symbol.broadcast_mul(conv3, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv3 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv2 = mx.symbol.broadcast_mul(conv2, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv2 + shortcut
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit3')
-    if bottle_neck:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn4 = mx.symbol.broadcast_mul(bn4, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn4 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn3 + shortcut
-
-def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNeXt Unit symbol for building ResNeXt
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    assert(bottle_neck)
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    num_group = 32
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-    conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                               workspace=workspace, name=name + '_conv3')
-    bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-    if use_se:
-      #se begin
-      body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-      body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv1", workspace=workspace)
-      body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-      body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv2", workspace=workspace)
-      body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-      bn4 = mx.symbol.broadcast_mul(bn4, body)
-      #se end
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn4 + shortcut
-
-
-def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-  uv = kwargs.get('version_unit', 3)
-  version_input = kwargs.get('version_input', 1)
-  if uv==1:
-    if version_input==0:
-      return residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-    else:
-      return residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==2:
-    return residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==4:
-    return residual_unit_v4(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  else:
-    return residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-
-def resnet(units, num_stages, filter_list, num_classes, bottle_neck, **kwargs):
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    """Return ResNet symbol of
-    Parameters
-    ----------
-    units : list
-        Number of units in each stage
-    num_stages : int
-        Number of stage
-    filter_list : list
-        Channel size of each stage
-    num_classes : int
-        Ouput size of symbol
-    dataset : str
-        Dataset type, only cifar10 and imagenet supports
-    workspace : int
-        Workspace used in convolution operator
-    """
-    version_se = kwargs.get('version_se', 0)
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'GAP')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    act_type = kwargs.get('version_act', 'prelu')
-    print(version_se, version_input, version_output, version_unit, act_type)
-    num_unit = len(units)
-    assert(num_unit == num_stages)
-    data = mx.sym.Variable(name='data')
-    if version_input==0:
-      #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-      #body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    elif version_input==2:
-      data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1,1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-    else:
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = data
-      body = Conv(data=body, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-
-    for i in range(num_stages):
-      #if version_input==0:
-      #  body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
-      #                       name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      #else:
-      #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
-      #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      body = residual_unit(body, filter_list[i+1], (2, 2), False,
-        name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      for j in range(units[i]-1):
-        body = residual_unit(body, filter_list[i+1], (1,1), True, name='stage%d_unit%d' % (i+1, j+2),
-          bottle_neck=bottle_neck, **kwargs)
-
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
-def get_symbol(num_classes, num_layers, **kwargs):
-    """
-    Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
-    Original author Wei Wu
-    """
-    if num_layers >= 101:
-        filter_list = [64, 256, 512, 1024, 2048]
-        bottle_neck = True
-    else:
-        filter_list = [64, 64, 128, 256, 512]
-        bottle_neck = False
-    num_stages = 4
-    if num_layers == 18:
-        units = [2, 2, 2, 2]
-    elif num_layers == 34:
-        units = [3, 4, 6, 3]
-    elif num_layers == 49:
-        units = [3, 4, 14, 3]
-    elif num_layers == 50:
-        units = [3, 4, 14, 3]
-    elif num_layers == 74:
-        units = [3, 6, 24, 3]
-    elif num_layers == 90:
-        units = [3, 8, 30, 3]
-    elif num_layers == 100:
-        units = [3, 13, 30, 3]
-    elif num_layers == 124:
-        units = [3, 13, 40, 5]
-    elif num_layers == 101:
-        units = [3, 4, 23, 3]
-    elif num_layers == 152:
-        units = [3, 8, 36, 3]
-    elif num_layers == 200:
-        units = [3, 24, 36, 3]
-    elif num_layers == 269:
-        units = [3, 30, 48, 8]
-    else:
-        raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
-
-    return resnet(units       = units,
-                  num_stages  = num_stages,
-                  filter_list = filter_list,
-                  num_classes = num_classes,
-                  bottle_neck = bottle_neck,
-                  **kwargs)
-
diff --git a/gender-age/helper.py b/gender-age/helper.py
deleted file mode 100644
index b82c4b77a..000000000
--- a/gender-age/helper.py
+++ /dev/null
@@ -1,168 +0,0 @@
-# coding: utf-8
-# YuanYang
-import math
-import cv2
-import numpy as np
-
-
-def nms(boxes, overlap_threshold, mode='Union'):
-    """
-        non max suppression
-
-    Parameters:
-    ----------
-        box: numpy array n x 5
-            input bbox array
-        overlap_threshold: float number
-            threshold of overlap
-        mode: float number
-            how to compute overlap ratio, 'Union' or 'Min'
-    Returns:
-    -------
-        index array of the selected bbox
-    """
-    # if there are no boxes, return an empty list
-    if len(boxes) == 0:
-        return []
-
-    # if the bounding boxes integers, convert them to floats
-    if boxes.dtype.kind == "i":
-        boxes = boxes.astype("float")
-
-    # initialize the list of picked indexes
-    pick = []
-
-    # grab the coordinates of the bounding boxes
-    x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)]
-
-    area = (x2 - x1 + 1) * (y2 - y1 + 1)
-    idxs = np.argsort(score)
-
-    # keep looping while some indexes still remain in the indexes list
-    while len(idxs) > 0:
-        # grab the last index in the indexes list and add the index value to the list of picked indexes
-        last = len(idxs) - 1
-        i = idxs[last]
-        pick.append(i)
-
-        xx1 = np.maximum(x1[i], x1[idxs[:last]])
-        yy1 = np.maximum(y1[i], y1[idxs[:last]])
-        xx2 = np.minimum(x2[i], x2[idxs[:last]])
-        yy2 = np.minimum(y2[i], y2[idxs[:last]])
-
-        # compute the width and height of the bounding box
-        w = np.maximum(0, xx2 - xx1 + 1)
-        h = np.maximum(0, yy2 - yy1 + 1)
-
-        inter = w * h
-        if mode == 'Min':
-            overlap = inter / np.minimum(area[i], area[idxs[:last]])
-        else:
-            overlap = inter / (area[i] + area[idxs[:last]] - inter)
-
-        # delete all indexes from the index list that have
-        idxs = np.delete(idxs, np.concatenate(([last],
-                                               np.where(overlap > overlap_threshold)[0])))
-
-    return pick
-
-def adjust_input(in_data):
-    """
-        adjust the input from (h, w, c) to ( 1, c, h, w) for network input
-
-    Parameters:
-    ----------
-        in_data: numpy array of shape (h, w, c)
-            input data
-    Returns:
-    -------
-        out_data: numpy array of shape (1, c, h, w)
-            reshaped array
-    """
-    if in_data.dtype is not np.dtype('float32'):
-        out_data = in_data.astype(np.float32)
-    else:
-        out_data = in_data
-
-    out_data = out_data.transpose((2,0,1))
-    out_data = np.expand_dims(out_data, 0)
-    out_data = (out_data - 127.5)*0.0078125
-    return out_data
-
-def generate_bbox(map, reg, scale, threshold):
-     """
-         generate bbox from feature map
-     Parameters:
-     ----------
-         map: numpy array , n x m x 1
-             detect score for each position
-         reg: numpy array , n x m x 4
-             bbox
-         scale: float number
-             scale of this detection
-         threshold: float number
-             detect threshold
-     Returns:
-     -------
-         bbox array
-     """
-     stride = 2
-     cellsize = 12
-
-     t_index = np.where(map>threshold)
-
-     # find nothing
-     if t_index[0].size == 0:
-         return np.array([])
-
-     dx1, dy1, dx2, dy2 = [reg[0, i, t_index[0], t_index[1]] for i in range(4)]
-
-     reg = np.array([dx1, dy1, dx2, dy2])
-     score = map[t_index[0], t_index[1]]
-     boundingbox = np.vstack([np.round((stride*t_index[1]+1)/scale),
-                              np.round((stride*t_index[0]+1)/scale),
-                              np.round((stride*t_index[1]+1+cellsize)/scale),
-                              np.round((stride*t_index[0]+1+cellsize)/scale),
-                              score,
-                              reg])
-
-     return boundingbox.T
-
-
-def detect_first_stage(img, net, scale, threshold):
-    """
-        run PNet for first stage
-    
-    Parameters:
-    ----------
-        img: numpy array, bgr order
-            input image
-        scale: float number
-            how much should the input image scale
-        net: PNet
-            worker
-    Returns:
-    -------
-        total_boxes : bboxes
-    """
-    height, width, _ = img.shape
-    hs = int(math.ceil(height * scale))
-    ws = int(math.ceil(width * scale))
-    
-    im_data = cv2.resize(img, (ws,hs))
-    
-    # adjust for the network input
-    input_buf = adjust_input(im_data)
-    output = net.predict(input_buf)
-    boxes = generate_bbox(output[1][0,1,:,:], output[0], scale, threshold)
-
-    if boxes.size == 0:
-        return None
-
-    # nms
-    pick = nms(boxes[:,0:5], 0.5, mode='Union')
-    boxes = boxes[pick]
-    return boxes
-
-def detect_first_stage_warpper( args ):
-    return detect_first_stage(*args)
diff --git a/gender-age/model/model-0000.params b/gender-age/model/model-0000.params
deleted file mode 100644
index 35118adad..000000000
Binary files a/gender-age/model/model-0000.params and /dev/null differ
diff --git a/gender-age/model/model-symbol.json b/gender-age/model/model-symbol.json
deleted file mode 100644
index cea9abc0f..000000000
--- a/gender-age/model/model-symbol.json
+++ /dev/null
@@ -1,2399 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "name": "data", 
-      "inputs": []
-    }, 
-    {
-      "op": "_minus_scalar", 
-      "name": "_minusscalar0", 
-      "attrs": {"scalar": "127.5"}, 
-      "inputs": [[0, 0, 0]]
-    }, 
-    {
-      "op": "_mul_scalar", 
-      "name": "_mulscalar0", 
-      "attrs": {"scalar": "0.0078125"}, 
-      "inputs": [[1, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_1_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "8", 
-        "num_group": "1", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_1_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "8", 
-        "num_group": "1", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[2, 0, 0], [3, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_1_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_1_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_1_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_1_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_1_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[4, 0, 0], [5, 0, 0], [6, 0, 0], [7, 0, 1], [8, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_1_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[9, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "8", 
-        "num_group": "8", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_2_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "8", 
-        "num_group": "8", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[10, 0, 0], [11, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_2_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[12, 0, 0], [13, 0, 0], [14, 0, 0], [15, 0, 1], [16, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_2_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[17, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "16", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_2_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "16", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[18, 0, 0], [19, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_2_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_2_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[20, 0, 0], [21, 0, 0], [22, 0, 0], [23, 0, 1], [24, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_2_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[25, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "16", 
-        "num_group": "16", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_3_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "16", 
-        "num_group": "16", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": [[26, 0, 0], [27, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_3_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[28, 0, 0], [29, 0, 0], [30, 0, 0], [31, 0, 1], [32, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_3_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[33, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_3_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[34, 0, 0], [35, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_3_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_3_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[36, 0, 0], [37, 0, 0], [38, 0, 0], [39, 0, 1], [40, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_3_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[41, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "32", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_4_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "32", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[42, 0, 0], [43, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_4_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[44, 0, 0], [45, 0, 0], [46, 0, 0], [47, 0, 1], [48, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_4_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[49, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_4_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[50, 0, 0], [51, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_4_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_4_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[52, 0, 0], [53, 0, 0], [54, 0, 0], [55, 0, 1], [56, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_4_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[57, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "32", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_5_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "32", 
-        "num_group": "32", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": [[58, 0, 0], [59, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_5_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[60, 0, 0], [61, 0, 0], [62, 0, 0], [63, 0, 1], [64, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_5_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[65, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_5_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[66, 0, 0], [67, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_5_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_5_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[68, 0, 0], [69, 0, 0], [70, 0, 0], [71, 0, 1], [72, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_5_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[73, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "64", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_6_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "64", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[74, 0, 0], [75, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_6_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[76, 0, 0], [77, 0, 0], [78, 0, 0], [79, 0, 1], [80, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_6_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[81, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_6_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[82, 0, 0], [83, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_6_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_6_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[84, 0, 0], [85, 0, 0], [86, 0, 0], [87, 0, 1], [88, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_6_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[89, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "64", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_7_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "64", 
-        "num_group": "64", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": [[90, 0, 0], [91, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_7_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[92, 0, 0], [93, 0, 0], [94, 0, 0], [95, 0, 1], [96, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_7_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[97, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_7_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[98, 0, 0], [99, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_7_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_7_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[100, 0, 0], [101, 0, 0], [102, 0, 0], [103, 0, 1], [104, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_7_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[105, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_8_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[106, 0, 0], [107, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_8_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[108, 0, 0], [109, 0, 0], [110, 0, 0], [111, 0, 1], [112, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_8_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[113, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_8_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[114, 0, 0], [115, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_8_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_8_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[116, 0, 0], [117, 0, 0], [118, 0, 0], [119, 0, 1], [120, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_8_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[121, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_9_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[122, 0, 0], [123, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_9_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[124, 0, 0], [125, 0, 0], [126, 0, 0], [127, 0, 1], [128, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_9_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[129, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_9_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[130, 0, 0], [131, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_9_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_9_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[132, 0, 0], [133, 0, 0], [134, 0, 0], [135, 0, 1], [136, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_9_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[137, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_10_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[138, 0, 0], [139, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_10_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[140, 0, 0], [141, 0, 0], [142, 0, 0], [143, 0, 1], [144, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_10_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[145, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_10_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[146, 0, 0], [147, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_10_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_10_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[148, 0, 0], [149, 0, 0], [150, 0, 0], [151, 0, 1], [152, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_10_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[153, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_11_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[154, 0, 0], [155, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_11_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[156, 0, 0], [157, 0, 0], [158, 0, 0], [159, 0, 1], [160, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_11_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[161, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_11_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[162, 0, 0], [163, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_11_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_11_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[164, 0, 0], [165, 0, 0], [166, 0, 0], [167, 0, 1], [168, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_11_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[169, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_12_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[170, 0, 0], [171, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_12_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[172, 0, 0], [173, 0, 0], [174, 0, 0], [175, 0, 1], [176, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_12_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[177, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_12_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[178, 0, 0], [179, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_12_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_12_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[180, 0, 0], [181, 0, 0], [182, 0, 0], [183, 0, 1], [184, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_12_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[185, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_13_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "128", 
-        "num_group": "128", 
-        "pad": "(1, 1)", 
-        "stride": "(2, 2)"
-      }, 
-      "inputs": [[186, 0, 0], [187, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_13_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[188, 0, 0], [189, 0, 0], [190, 0, 0], [191, 0, 1], [192, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_13_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[193, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_13_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[194, 0, 0], [195, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_13_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_13_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[196, 0, 0], [197, 0, 0], [198, 0, 0], [199, 0, 1], [200, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_13_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[201, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_dw_conv2d_weight", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "256", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_14_dw_conv2d", 
-      "attrs": {
-        "kernel": "(3, 3)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "256", 
-        "pad": "(1, 1)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[202, 0, 0], [203, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_dw_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_dw_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_dw_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_dw_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_14_dw_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[204, 0, 0], [205, 0, 0], [206, 0, 0], [207, 0, 1], [208, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_14_dw_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[209, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_conv2d_weight", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "Convolution", 
-      "name": "conv_14_conv2d", 
-      "attrs": {
-        "kernel": "(1, 1)", 
-        "no_bias": "True", 
-        "num_filter": "256", 
-        "num_group": "1", 
-        "pad": "(0, 0)", 
-        "stride": "(1, 1)"
-      }, 
-      "inputs": [[210, 0, 0], [211, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_batchnorm_gamma", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_batchnorm_beta", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_batchnorm_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "conv_14_batchnorm_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "fix_gamma": "True"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "conv_14_batchnorm", 
-      "attrs": {"fix_gamma": "True"}, 
-      "inputs": [[212, 0, 0], [213, 0, 0], [214, 0, 0], [215, 0, 1], [216, 0, 1]]
-    }, 
-    {
-      "op": "Activation", 
-      "name": "conv_14_relu", 
-      "attrs": {"act_type": "relu"}, 
-      "inputs": [[217, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "bn1_gamma", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "False", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "bn1_beta", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "False", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "bn1_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "eps": "2e-05", 
-        "fix_gamma": "False", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "bn1_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "eps": "2e-05", 
-        "fix_gamma": "False", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "bn1", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "False", 
-        "momentum": "0.9"
-      }, 
-      "inputs": [[218, 0, 0], [219, 0, 0], [220, 0, 0], [221, 0, 1], [222, 0, 1]]
-    }, 
-    {
-      "op": "null", 
-      "name": "relu1_gamma", 
-      "attrs": {
-        "__init__": "[\"Constant\", {\"value\": 0.25}]", 
-        "act_type": "prelu"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "name": "relu1", 
-      "attrs": {"act_type": "prelu"}, 
-      "inputs": [[223, 0, 0], [224, 0, 0]]
-    }, 
-    {
-      "op": "Pooling", 
-      "name": "pool1", 
-      "attrs": {
-        "global_pool": "True", 
-        "kernel": "(7, 7)", 
-        "pool_type": "avg"
-      }, 
-      "inputs": [[225, 0, 0]]
-    }, 
-    {
-      "op": "Flatten", 
-      "name": "flatten0", 
-      "inputs": [[226, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "pre_fc1_weight", 
-      "attrs": {"num_hidden": "202"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "pre_fc1_bias", 
-      "attrs": {"num_hidden": "202"}, 
-      "inputs": []
-    }, 
-    {
-      "op": "FullyConnected", 
-      "name": "pre_fc1", 
-      "attrs": {"num_hidden": "202"}, 
-      "inputs": [[227, 0, 0], [228, 0, 0], [229, 0, 0]]
-    }, 
-    {
-      "op": "null", 
-      "name": "fc1_gamma", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "True", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "fc1_beta", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "True", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "fc1_moving_mean", 
-      "attrs": {
-        "__init__": "[\"zero\", {}]", 
-        "eps": "2e-05", 
-        "fix_gamma": "True", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "null", 
-      "name": "fc1_moving_var", 
-      "attrs": {
-        "__init__": "[\"one\", {}]", 
-        "eps": "2e-05", 
-        "fix_gamma": "True", 
-        "momentum": "0.9"
-      }, 
-      "inputs": []
-    }, 
-    {
-      "op": "BatchNorm", 
-      "name": "fc1", 
-      "attrs": {
-        "eps": "2e-05", 
-        "fix_gamma": "True", 
-        "momentum": "0.9"
-      }, 
-      "inputs": [[230, 0, 0], [231, 0, 0], [232, 0, 0], [233, 0, 1], [234, 0, 1]]
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    3, 
-    5, 
-    6, 
-    7, 
-    8, 
-    11, 
-    13, 
-    14, 
-    15, 
-    16, 
-    19, 
-    21, 
-    22, 
-    23, 
-    24, 
-    27, 
-    29, 
-    30, 
-    31, 
-    32, 
-    35, 
-    37, 
-    38, 
-    39, 
-    40, 
-    43, 
-    45, 
-    46, 
-    47, 
-    48, 
-    51, 
-    53, 
-    54, 
-    55, 
-    56, 
-    59, 
-    61, 
-    62, 
-    63, 
-    64, 
-    67, 
-    69, 
-    70, 
-    71, 
-    72, 
-    75, 
-    77, 
-    78, 
-    79, 
-    80, 
-    83, 
-    85, 
-    86, 
-    87, 
-    88, 
-    91, 
-    93, 
-    94, 
-    95, 
-    96, 
-    99, 
-    101, 
-    102, 
-    103, 
-    104, 
-    107, 
-    109, 
-    110, 
-    111, 
-    112, 
-    115, 
-    117, 
-    118, 
-    119, 
-    120, 
-    123, 
-    125, 
-    126, 
-    127, 
-    128, 
-    131, 
-    133, 
-    134, 
-    135, 
-    136, 
-    139, 
-    141, 
-    142, 
-    143, 
-    144, 
-    147, 
-    149, 
-    150, 
-    151, 
-    152, 
-    155, 
-    157, 
-    158, 
-    159, 
-    160, 
-    163, 
-    165, 
-    166, 
-    167, 
-    168, 
-    171, 
-    173, 
-    174, 
-    175, 
-    176, 
-    179, 
-    181, 
-    182, 
-    183, 
-    184, 
-    187, 
-    189, 
-    190, 
-    191, 
-    192, 
-    195, 
-    197, 
-    198, 
-    199, 
-    200, 
-    203, 
-    205, 
-    206, 
-    207, 
-    208, 
-    211, 
-    213, 
-    214, 
-    215, 
-    216, 
-    219, 
-    220, 
-    221, 
-    222, 
-    224, 
-    228, 
-    229, 
-    231, 
-    232, 
-    233, 
-    234
-  ], 
-  "node_row_ptr": [
-    0, 
-    1, 
-    2, 
-    3, 
-    4, 
-    5, 
-    6, 
-    7, 
-    8, 
-    9, 
-    12, 
-    13, 
-    14, 
-    15, 
-    16, 
-    17, 
-    18, 
-    19, 
-    22, 
-    23, 
-    24, 
-    25, 
-    26, 
-    27, 
-    28, 
-    29, 
-    32, 
-    33, 
-    34, 
-    35, 
-    36, 
-    37, 
-    38, 
-    39, 
-    42, 
-    43, 
-    44, 
-    45, 
-    46, 
-    47, 
-    48, 
-    49, 
-    52, 
-    53, 
-    54, 
-    55, 
-    56, 
-    57, 
-    58, 
-    59, 
-    62, 
-    63, 
-    64, 
-    65, 
-    66, 
-    67, 
-    68, 
-    69, 
-    72, 
-    73, 
-    74, 
-    75, 
-    76, 
-    77, 
-    78, 
-    79, 
-    82, 
-    83, 
-    84, 
-    85, 
-    86, 
-    87, 
-    88, 
-    89, 
-    92, 
-    93, 
-    94, 
-    95, 
-    96, 
-    97, 
-    98, 
-    99, 
-    102, 
-    103, 
-    104, 
-    105, 
-    106, 
-    107, 
-    108, 
-    109, 
-    112, 
-    113, 
-    114, 
-    115, 
-    116, 
-    117, 
-    118, 
-    119, 
-    122, 
-    123, 
-    124, 
-    125, 
-    126, 
-    127, 
-    128, 
-    129, 
-    132, 
-    133, 
-    134, 
-    135, 
-    136, 
-    137, 
-    138, 
-    139, 
-    142, 
-    143, 
-    144, 
-    145, 
-    146, 
-    147, 
-    148, 
-    149, 
-    152, 
-    153, 
-    154, 
-    155, 
-    156, 
-    157, 
-    158, 
-    159, 
-    162, 
-    163, 
-    164, 
-    165, 
-    166, 
-    167, 
-    168, 
-    169, 
-    172, 
-    173, 
-    174, 
-    175, 
-    176, 
-    177, 
-    178, 
-    179, 
-    182, 
-    183, 
-    184, 
-    185, 
-    186, 
-    187, 
-    188, 
-    189, 
-    192, 
-    193, 
-    194, 
-    195, 
-    196, 
-    197, 
-    198, 
-    199, 
-    202, 
-    203, 
-    204, 
-    205, 
-    206, 
-    207, 
-    208, 
-    209, 
-    212, 
-    213, 
-    214, 
-    215, 
-    216, 
-    217, 
-    218, 
-    219, 
-    222, 
-    223, 
-    224, 
-    225, 
-    226, 
-    227, 
-    228, 
-    229, 
-    232, 
-    233, 
-    234, 
-    235, 
-    236, 
-    237, 
-    238, 
-    239, 
-    242, 
-    243, 
-    244, 
-    245, 
-    246, 
-    247, 
-    248, 
-    249, 
-    252, 
-    253, 
-    254, 
-    255, 
-    256, 
-    257, 
-    258, 
-    259, 
-    262, 
-    263, 
-    264, 
-    265, 
-    266, 
-    267, 
-    268, 
-    269, 
-    272, 
-    273, 
-    274, 
-    275, 
-    276, 
-    277, 
-    280, 
-    281, 
-    282, 
-    283, 
-    284, 
-    285, 
-    286, 
-    287, 
-    288, 
-    289, 
-    290, 
-    291, 
-    294
-  ], 
-  "heads": [[235, 0, 0]], 
-  "attrs": {"mxnet_version": ["int", 10300]}
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det1-0001.params b/gender-age/mtcnn-model/det1-0001.params
deleted file mode 100644
index e4b04aa90..000000000
Binary files a/gender-age/mtcnn-model/det1-0001.params and /dev/null differ
diff --git a/gender-age/mtcnn-model/det1-symbol.json b/gender-age/mtcnn-model/det1-symbol.json
deleted file mode 100644
index bd9b7720b..000000000
--- a/gender-age/mtcnn-model/det1-symbol.json
+++ /dev/null
@@ -1,266 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "10", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(2,2)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "16", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[11, 0], [12, 0], [13, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[14, 0], [15, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(1,1)", 
-        "no_bias": "False", 
-        "num_filter": "4", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4_2", 
-      "inputs": [[16, 0], [17, 0], [18, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(1,1)", 
-        "no_bias": "False", 
-        "num_filter": "2", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4_1", 
-      "inputs": [[16, 0], [20, 0], [21, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxActivation", 
-      "param": {"mode": "channel"}, 
-      "name": "prob1", 
-      "inputs": [[22, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    12, 
-    13, 
-    15, 
-    17, 
-    18, 
-    20, 
-    21
-  ], 
-  "heads": [[19, 0], [23, 0]]
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det1.caffemodel b/gender-age/mtcnn-model/det1.caffemodel
deleted file mode 100644
index 79e93b488..000000000
Binary files a/gender-age/mtcnn-model/det1.caffemodel and /dev/null differ
diff --git a/gender-age/mtcnn-model/det1.prototxt b/gender-age/mtcnn-model/det1.prototxt
deleted file mode 100644
index c5c1657c1..000000000
--- a/gender-age/mtcnn-model/det1.prototxt
+++ /dev/null
@@ -1,177 +0,0 @@
-name: "PNet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 12
-input_dim: 12
-
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 10
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 2
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 16
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-}
-
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "conv2"
-  top: "conv3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 32
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-	  type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "PReLU3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-}
-
-
-layer {
-  name: "conv4-1"
-  type: "Convolution"
-  bottom: "conv3"
-  top: "conv4-1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 2
-    kernel_size: 1
-    stride: 1
-     weight_filler {
-      type: "xavier"
-    }
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "conv4-2"
-  type: "Convolution"
-  bottom: "conv3"
-  top: "conv4-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 4
-    kernel_size: 1
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv4-1"
-  top: "prob1"
-}
diff --git a/gender-age/mtcnn-model/det2-0001.params b/gender-age/mtcnn-model/det2-0001.params
deleted file mode 100644
index a14a478e1..000000000
Binary files a/gender-age/mtcnn-model/det2-0001.params and /dev/null differ
diff --git a/gender-age/mtcnn-model/det2-symbol.json b/gender-age/mtcnn-model/det2-symbol.json
deleted file mode 100644
index a13246a7a..000000000
--- a/gender-age/mtcnn-model/det2-symbol.json
+++ /dev/null
@@ -1,324 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2", 
-      "inputs": [[11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[12, 0], [13, 0], [14, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[15, 0], [16, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "128"
-      }, 
-      "name": "conv4", 
-      "inputs": [[17, 0], [18, 0], [19, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[20, 0], [21, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "4"
-      }, 
-      "name": "conv5_2", 
-      "inputs": [[22, 0], [23, 0], [24, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "conv5_1", 
-      "inputs": [[22, 0], [26, 0], [27, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prob1_label", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxOutput", 
-      "param": {
-        "grad_scale": "1", 
-        "ignore_label": "-1", 
-        "multi_output": "False", 
-        "normalization": "null", 
-        "use_ignore": "False"
-      }, 
-      "name": "prob1", 
-      "inputs": [[28, 0], [29, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    13, 
-    14, 
-    16, 
-    18, 
-    19, 
-    21, 
-    23, 
-    24, 
-    26, 
-    27, 
-    29
-  ], 
-  "heads": [[25, 0], [30, 0]]
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det2.caffemodel b/gender-age/mtcnn-model/det2.caffemodel
deleted file mode 100644
index a5a540c00..000000000
Binary files a/gender-age/mtcnn-model/det2.caffemodel and /dev/null differ
diff --git a/gender-age/mtcnn-model/det2.prototxt b/gender-age/mtcnn-model/det2.prototxt
deleted file mode 100644
index 51093e6d9..000000000
--- a/gender-age/mtcnn-model/det2.prototxt
+++ /dev/null
@@ -1,228 +0,0 @@
-name: "RNet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 24
-input_dim: 24
-
-
-##########################
-######################
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-  propagate_down: true
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-  propagate_down: true
-}
-layer {
-  name: "pool2"
-  type: "Pooling"
-  bottom: "conv2"
-  top: "pool2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-####################################
-
-##################################
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "pool2"
-  top: "conv3"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-  propagate_down: true
-}
-###############################
-
-###############################
-
-layer {
-  name: "conv4"
-  type: "InnerProduct"
-  bottom: "conv3"
-  top: "conv4"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  inner_product_param {
-    num_output: 128
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "conv4"
-  top: "conv4"
-}
-
-layer {
-  name: "conv5-1"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5-1"
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  param {
-    lr_mult: 0
-    decay_mult: 0
-  }
-  inner_product_param {
-    num_output: 2
-    #kernel_size: 1
-    #stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv5-2"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 4
-    #kernel_size: 1
-    #stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv5-1"
-  top: "prob1"
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det3-0001.params b/gender-age/mtcnn-model/det3-0001.params
deleted file mode 100644
index cae898b8a..000000000
Binary files a/gender-age/mtcnn-model/det3-0001.params and /dev/null differ
diff --git a/gender-age/mtcnn-model/det3-symbol.json b/gender-age/mtcnn-model/det3-symbol.json
deleted file mode 100644
index 00061edfd..000000000
--- a/gender-age/mtcnn-model/det3-symbol.json
+++ /dev/null
@@ -1,418 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "32", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1", 
-      "inputs": [[0, 0], [1, 0], [2, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1", 
-      "inputs": [[3, 0], [4, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1", 
-      "inputs": [[5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2", 
-      "inputs": [[6, 0], [7, 0], [8, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2", 
-      "inputs": [[9, 0], [10, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2", 
-      "inputs": [[11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3", 
-      "inputs": [[12, 0], [13, 0], [14, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3", 
-      "inputs": [[15, 0], [16, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(2,2)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool3", 
-      "inputs": [[17, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "128", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv4", 
-      "inputs": [[18, 0], [19, 0], [20, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[21, 0], [22, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "256"
-      }, 
-      "name": "conv5", 
-      "inputs": [[23, 0], [24, 0], [25, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu5", 
-      "inputs": [[26, 0], [27, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "10"
-      }, 
-      "name": "conv6_3", 
-      "inputs": [[28, 0], [29, 0], [30, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "4"
-      }, 
-      "name": "conv6_2", 
-      "inputs": [[28, 0], [32, 0], [33, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv6_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "conv6_1", 
-      "inputs": [[28, 0], [35, 0], [36, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prob1_label", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SoftmaxOutput", 
-      "param": {
-        "grad_scale": "1", 
-        "ignore_label": "-1", 
-        "multi_output": "False", 
-        "normalization": "null", 
-        "use_ignore": "False"
-      }, 
-      "name": "prob1", 
-      "inputs": [[37, 0], [38, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    1, 
-    2, 
-    4, 
-    7, 
-    8, 
-    10, 
-    13, 
-    14, 
-    16, 
-    19, 
-    20, 
-    22, 
-    24, 
-    25, 
-    27, 
-    29, 
-    30, 
-    32, 
-    33, 
-    35, 
-    36, 
-    38
-  ], 
-  "heads": [[31, 0], [34, 0], [39, 0]]
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det3.caffemodel b/gender-age/mtcnn-model/det3.caffemodel
deleted file mode 100644
index 7b4b8a4af..000000000
Binary files a/gender-age/mtcnn-model/det3.caffemodel and /dev/null differ
diff --git a/gender-age/mtcnn-model/det3.prototxt b/gender-age/mtcnn-model/det3.prototxt
deleted file mode 100644
index a19230786..000000000
--- a/gender-age/mtcnn-model/det3.prototxt
+++ /dev/null
@@ -1,294 +0,0 @@
-name: "ONet"
-input: "data"
-input_dim: 1
-input_dim: 3
-input_dim: 48
-input_dim: 48
-##################################
-layer {
-  name: "conv1"
-  type: "Convolution"
-  bottom: "data"
-  top: "conv1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 32
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu1"
-  type: "PReLU"
-  bottom: "conv1"
-  top: "conv1"
-}
-layer {
-  name: "pool1"
-  type: "Pooling"
-  bottom: "conv1"
-  top: "pool1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-layer {
-  name: "conv2"
-  type: "Convolution"
-  bottom: "pool1"
-  top: "conv2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "prelu2"
-  type: "PReLU"
-  bottom: "conv2"
-  top: "conv2"
-}
-layer {
-  name: "pool2"
-  type: "Pooling"
-  bottom: "conv2"
-  top: "pool2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv3"
-  type: "Convolution"
-  bottom: "pool2"
-  top: "conv3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-	num_output: 64
-	kernel_size: 3
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu3"
-  type: "PReLU"
-  bottom: "conv3"
-  top: "conv3"
-}
-layer {
-  name: "pool3"
-  type: "Pooling"
-  bottom: "conv3"
-  top: "pool3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 2
-    stride: 2
-  }
-}
-layer {
-  name: "conv4"
-  type: "Convolution"
-  bottom: "pool3"
-  top: "conv4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-	num_output: 128
-	kernel_size: 2
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "conv4"
-  top: "conv4"
-}
-
-
-layer {
-  name: "conv5"
-  type: "InnerProduct"
-  bottom: "conv4"
-  top: "conv5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-	#kernel_size: 3
-	num_output: 256
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-layer {
-  name: "drop5"
-  type: "Dropout"
-  bottom: "conv5"
-  top: "conv5"
-  dropout_param {
-    dropout_ratio: 0.25
-  }
-}
-layer {
-  name: "prelu5"
-  type: "PReLU"
-  bottom: "conv5"
-  top: "conv5"
-}
-
-
-layer {
-  name: "conv6-1"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    #kernel_size: 1
-	num_output: 2
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv6-2"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-  	#kernel_size: 1
-	num_output: 4
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "conv6-3"
-  type: "InnerProduct"
-  bottom: "conv5"
-  top: "conv6-3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-  	#kernel_size: 1
-	num_output: 10
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-layer {
-  name: "prob1"
-  type: "Softmax"
-  bottom: "conv6-1"
-  top: "prob1"
-}
diff --git a/gender-age/mtcnn-model/det4-0001.params b/gender-age/mtcnn-model/det4-0001.params
deleted file mode 100644
index efca9a959..000000000
Binary files a/gender-age/mtcnn-model/det4-0001.params and /dev/null differ
diff --git a/gender-age/mtcnn-model/det4-symbol.json b/gender-age/mtcnn-model/det4-symbol.json
deleted file mode 100644
index aa90e2a09..000000000
--- a/gender-age/mtcnn-model/det4-symbol.json
+++ /dev/null
@@ -1,1392 +0,0 @@
-{
-  "nodes": [
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "data", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "SliceChannel", 
-      "param": {
-        "axis": "1", 
-        "num_outputs": "5", 
-        "squeeze_axis": "False"
-      }, 
-      "name": "slice", 
-      "inputs": [[0, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_1", 
-      "inputs": [[1, 0], [2, 0], [3, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_1", 
-      "inputs": [[4, 0], [5, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_1", 
-      "inputs": [[6, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_1", 
-      "inputs": [[7, 0], [8, 0], [9, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_1", 
-      "inputs": [[10, 0], [11, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_1", 
-      "inputs": [[12, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_1", 
-      "inputs": [[13, 0], [14, 0], [15, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_1", 
-      "inputs": [[16, 0], [17, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_2", 
-      "inputs": [[1, 1], [19, 0], [20, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_2", 
-      "inputs": [[21, 0], [22, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_2", 
-      "inputs": [[23, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_2", 
-      "inputs": [[24, 0], [25, 0], [26, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_2", 
-      "inputs": [[27, 0], [28, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_2", 
-      "inputs": [[29, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_2", 
-      "inputs": [[30, 0], [31, 0], [32, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_2", 
-      "inputs": [[33, 0], [34, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_3", 
-      "inputs": [[1, 2], [36, 0], [37, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_3", 
-      "inputs": [[38, 0], [39, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_3", 
-      "inputs": [[40, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_3", 
-      "inputs": [[41, 0], [42, 0], [43, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_3", 
-      "inputs": [[44, 0], [45, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_3", 
-      "inputs": [[46, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_3", 
-      "inputs": [[47, 0], [48, 0], [49, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_3", 
-      "inputs": [[50, 0], [51, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_4", 
-      "inputs": [[1, 3], [53, 0], [54, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_4", 
-      "inputs": [[55, 0], [56, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_4", 
-      "inputs": [[57, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_4", 
-      "inputs": [[58, 0], [59, 0], [60, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_4", 
-      "inputs": [[61, 0], [62, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_4", 
-      "inputs": [[63, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_4", 
-      "inputs": [[64, 0], [65, 0], [66, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_4", 
-      "inputs": [[67, 0], [68, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv1_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "28", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv1_5", 
-      "inputs": [[1, 4], [70, 0], [71, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu1_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu1_5", 
-      "inputs": [[72, 0], [73, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool1_5", 
-      "inputs": [[74, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv2_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(3,3)", 
-        "no_bias": "False", 
-        "num_filter": "48", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv2_5", 
-      "inputs": [[75, 0], [76, 0], [77, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu2_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu2_5", 
-      "inputs": [[78, 0], [79, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Pooling", 
-      "param": {
-        "global_pool": "False", 
-        "kernel": "(3,3)", 
-        "pad": "(0,0)", 
-        "pool_type": "max", 
-        "pooling_convention": "full", 
-        "stride": "(2,2)"
-      }, 
-      "name": "pool2_5", 
-      "inputs": [[80, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "conv3_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Convolution", 
-      "param": {
-        "cudnn_off": "False", 
-        "cudnn_tune": "off", 
-        "dilate": "(1,1)", 
-        "kernel": "(2,2)", 
-        "no_bias": "False", 
-        "num_filter": "64", 
-        "num_group": "1", 
-        "pad": "(0,0)", 
-        "stride": "(1,1)", 
-        "workspace": "1024"
-      }, 
-      "name": "conv3_5", 
-      "inputs": [[81, 0], [82, 0], [83, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu3_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu3_5", 
-      "inputs": [[84, 0], [85, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "Concat", 
-      "param": {
-        "dim": "1", 
-        "num_args": "5"
-      }, 
-      "name": "concat", 
-      "inputs": [[18, 0], [35, 0], [52, 0], [69, 0], [86, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "256"
-      }, 
-      "name": "fc4", 
-      "inputs": [[87, 0], [88, 0], [89, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4", 
-      "inputs": [[90, 0], [91, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_1", 
-      "inputs": [[92, 0], [93, 0], [94, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_1_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_1", 
-      "inputs": [[95, 0], [96, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_1_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_1_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_1", 
-      "inputs": [[97, 0], [98, 0], [99, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_2", 
-      "inputs": [[92, 0], [101, 0], [102, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_2_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_2", 
-      "inputs": [[103, 0], [104, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_2_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_2_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_2", 
-      "inputs": [[105, 0], [106, 0], [107, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_3", 
-      "inputs": [[92, 0], [109, 0], [110, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_3_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_3", 
-      "inputs": [[111, 0], [112, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_3_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_3_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_3", 
-      "inputs": [[113, 0], [114, 0], [115, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_4", 
-      "inputs": [[92, 0], [117, 0], [118, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_4_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_4", 
-      "inputs": [[119, 0], [120, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_4_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_4_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_4", 
-      "inputs": [[121, 0], [122, 0], [123, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc4_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "64"
-      }, 
-      "name": "fc4_5", 
-      "inputs": [[92, 0], [125, 0], [126, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "prelu4_5_gamma", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "LeakyReLU", 
-      "param": {
-        "act_type": "prelu", 
-        "lower_bound": "0.125", 
-        "slope": "0.25", 
-        "upper_bound": "0.334"
-      }, 
-      "name": "prelu4_5", 
-      "inputs": [[127, 0], [128, 0]], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_5_weight", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "null", 
-      "param": {}, 
-      "name": "fc5_5_bias", 
-      "inputs": [], 
-      "backward_source_id": -1
-    }, 
-    {
-      "op": "FullyConnected", 
-      "param": {
-        "no_bias": "False", 
-        "num_hidden": "2"
-      }, 
-      "name": "fc5_5", 
-      "inputs": [[129, 0], [130, 0], [131, 0]], 
-      "backward_source_id": -1
-    }
-  ], 
-  "arg_nodes": [
-    0, 
-    2, 
-    3, 
-    5, 
-    8, 
-    9, 
-    11, 
-    14, 
-    15, 
-    17, 
-    19, 
-    20, 
-    22, 
-    25, 
-    26, 
-    28, 
-    31, 
-    32, 
-    34, 
-    36, 
-    37, 
-    39, 
-    42, 
-    43, 
-    45, 
-    48, 
-    49, 
-    51, 
-    53, 
-    54, 
-    56, 
-    59, 
-    60, 
-    62, 
-    65, 
-    66, 
-    68, 
-    70, 
-    71, 
-    73, 
-    76, 
-    77, 
-    79, 
-    82, 
-    83, 
-    85, 
-    88, 
-    89, 
-    91, 
-    93, 
-    94, 
-    96, 
-    98, 
-    99, 
-    101, 
-    102, 
-    104, 
-    106, 
-    107, 
-    109, 
-    110, 
-    112, 
-    114, 
-    115, 
-    117, 
-    118, 
-    120, 
-    122, 
-    123, 
-    125, 
-    126, 
-    128, 
-    130, 
-    131
-  ], 
-  "heads": [[100, 0], [108, 0], [116, 0], [124, 0], [132, 0]]
-}
\ No newline at end of file
diff --git a/gender-age/mtcnn-model/det4.caffemodel b/gender-age/mtcnn-model/det4.caffemodel
deleted file mode 100644
index 38353c4e3..000000000
Binary files a/gender-age/mtcnn-model/det4.caffemodel and /dev/null differ
diff --git a/gender-age/mtcnn-model/det4.prototxt b/gender-age/mtcnn-model/det4.prototxt
deleted file mode 100644
index 4cdc329d0..000000000
--- a/gender-age/mtcnn-model/det4.prototxt
+++ /dev/null
@@ -1,995 +0,0 @@
-name: "LNet"
-input: "data"
-input_dim: 1
-input_dim: 15
-input_dim: 24
-input_dim: 24
-
-layer {
-  name: "slicer_data"
-  type: "Slice"
-  bottom: "data"
-  top: "data241"
-  top: "data242"
-  top: "data243"
-  top: "data244"
-  top: "data245"
-  slice_param {
-    axis: 1
-    slice_point: 3
-    slice_point: 6
-    slice_point: 9
-    slice_point: 12
-  }
-}
-layer {
-  name: "conv1_1"
-  type: "Convolution"
-  bottom: "data241"
-  top: "conv1_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_1"
-  type: "PReLU"
-  bottom: "conv1_1"
-  top: "conv1_1"
-
-}
-layer {
-  name: "pool1_1"
-  type: "Pooling"
-  bottom: "conv1_1"
-  top: "pool1_1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_1"
-  type: "Convolution"
-  bottom: "pool1_1"
-  top: "conv2_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_1"
-  type: "PReLU"
-  bottom: "conv2_1"
-  top: "conv2_1"
-}
-layer {
-  name: "pool2_1"
-  type: "Pooling"
-  bottom: "conv2_1"
-  top: "pool2_1"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_1"
-  type: "Convolution"
-  bottom: "pool2_1"
-  top: "conv3_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_1"
-  type: "PReLU"
-  bottom: "conv3_1"
-  top: "conv3_1"
-}
-##########################
-layer {
-  name: "conv1_2"
-  type: "Convolution"
-  bottom: "data242"
-  top: "conv1_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_2"
-  type: "PReLU"
-  bottom: "conv1_2"
-  top: "conv1_2"
-
-}
-layer {
-  name: "pool1_2"
-  type: "Pooling"
-  bottom: "conv1_2"
-  top: "pool1_2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_2"
-  type: "Convolution"
-  bottom: "pool1_2"
-  top: "conv2_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_2"
-  type: "PReLU"
-  bottom: "conv2_2"
-  top: "conv2_2"
-}
-layer {
-  name: "pool2_2"
-  type: "Pooling"
-  bottom: "conv2_2"
-  top: "pool2_2"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_2"
-  type: "Convolution"
-  bottom: "pool2_2"
-  top: "conv3_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_2"
-  type: "PReLU"
-  bottom: "conv3_2"
-  top: "conv3_2"
-}
-##########################
-##########################
-layer {
-  name: "conv1_3"
-  type: "Convolution"
-  bottom: "data243"
-  top: "conv1_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_3"
-  type: "PReLU"
-  bottom: "conv1_3"
-  top: "conv1_3"
-
-}
-layer {
-  name: "pool1_3"
-  type: "Pooling"
-  bottom: "conv1_3"
-  top: "pool1_3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_3"
-  type: "Convolution"
-  bottom: "pool1_3"
-  top: "conv2_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_3"
-  type: "PReLU"
-  bottom: "conv2_3"
-  top: "conv2_3"
-}
-layer {
-  name: "pool2_3"
-  type: "Pooling"
-  bottom: "conv2_3"
-  top: "pool2_3"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_3"
-  type: "Convolution"
-  bottom: "pool2_3"
-  top: "conv3_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_3"
-  type: "PReLU"
-  bottom: "conv3_3"
-  top: "conv3_3"
-}
-##########################
-##########################
-layer {
-  name: "conv1_4"
-  type: "Convolution"
-  bottom: "data244"
-  top: "conv1_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_4"
-  type: "PReLU"
-  bottom: "conv1_4"
-  top: "conv1_4"
-
-}
-layer {
-  name: "pool1_4"
-  type: "Pooling"
-  bottom: "conv1_4"
-  top: "pool1_4"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_4"
-  type: "Convolution"
-  bottom: "pool1_4"
-  top: "conv2_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_4"
-  type: "PReLU"
-  bottom: "conv2_4"
-  top: "conv2_4"
-}
-layer {
-  name: "pool2_4"
-  type: "Pooling"
-  bottom: "conv2_4"
-  top: "pool2_4"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_4"
-  type: "Convolution"
-  bottom: "pool2_4"
-  top: "conv3_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_4"
-  type: "PReLU"
-  bottom: "conv3_4"
-  top: "conv3_4"
-}
-##########################
-##########################
-layer {
-  name: "conv1_5"
-  type: "Convolution"
-  bottom: "data245"
-  top: "conv1_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 28
-    kernel_size: 3
-    stride: 1
-     weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu1_5"
-  type: "PReLU"
-  bottom: "conv1_5"
-  top: "conv1_5"
-
-}
-layer {
-  name: "pool1_5"
-  type: "Pooling"
-  bottom: "conv1_5"
-  top: "pool1_5"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-}
-
-layer {
-  name: "conv2_5"
-  type: "Convolution"
-  bottom: "pool1_5"
-  top: "conv2_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 48
-    kernel_size: 3
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu2_5"
-  type: "PReLU"
-  bottom: "conv2_5"
-  top: "conv2_5"
-}
-layer {
-  name: "pool2_5"
-  type: "Pooling"
-  bottom: "conv2_5"
-  top: "pool2_5"
-  pooling_param {
-    pool: MAX
-    kernel_size: 3
-    stride: 2
-  }
-  
-}
-layer {
-  name: "conv3_5"
-  type: "Convolution"
-  bottom: "pool2_5"
-  top: "conv3_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  convolution_param {
-    num_output: 64
-    kernel_size: 2
-    stride: 1
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu3_5"
-  type: "PReLU"
-  bottom: "conv3_5"
-  top: "conv3_5"
-}
-##########################
-layer { 
-	name: "concat" 
-	bottom: "conv3_1" 
-	bottom: "conv3_2" 
-	bottom: "conv3_3" 
-	bottom: "conv3_4" 
-	bottom: "conv3_5" 
-	top: "conv3" 
-	type: "Concat" 
-	concat_param { 
-		axis: 1 
-	} 
-}
-##########################
-layer {
-  name: "fc4"
-  type: "InnerProduct"
-  bottom: "conv3"
-  top: "fc4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 256
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4"
-  type: "PReLU"
-  bottom: "fc4"
-  top: "fc4"
-}
-############################
-layer {
-  name: "fc4_1"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_1"
-  type: "PReLU"
-  bottom: "fc4_1"
-  top: "fc4_1"
-}
-layer {
-  name: "fc5_1"
-  type: "InnerProduct"
-  bottom: "fc4_1"
-  top: "fc5_1"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-
-#########################
-layer {
-  name: "fc4_2"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_2"
-  type: "PReLU"
-  bottom: "fc4_2"
-  top: "fc4_2"
-}
-layer {
-  name: "fc5_2"
-  type: "InnerProduct"
-  bottom: "fc4_2"
-  top: "fc5_2"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_3"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_3"
-  type: "PReLU"
-  bottom: "fc4_3"
-  top: "fc4_3"
-}
-layer {
-  name: "fc5_3"
-  type: "InnerProduct"
-  bottom: "fc4_3"
-  top: "fc5_3"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_4"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_4"
-  type: "PReLU"
-  bottom: "fc4_4"
-  top: "fc4_4"
-}
-layer {
-  name: "fc5_4"
-  type: "InnerProduct"
-  bottom: "fc4_4"
-  top: "fc5_4"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-layer {
-  name: "fc4_5"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc4_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 64
-    weight_filler {
-      type: "xavier"
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-  
-}
-layer {
-  name: "prelu4_5"
-  type: "PReLU"
-  bottom: "fc4_5"
-  top: "fc4_5"
-}
-layer {
-  name: "fc5_5"
-  type: "InnerProduct"
-  bottom: "fc4_5"
-  top: "fc5_5"
-  param {
-    lr_mult: 1
-    decay_mult: 1
-  }
-  param {
-    lr_mult: 2
-    decay_mult: 1
-  }
-  inner_product_param {
-    num_output: 2
-    weight_filler {
-      type: "xavier"
-	  #type: "constant"
-      #value: 0
-	}
-    bias_filler {
-      type: "constant"
-      value: 0
-    }
-  }
-}
-
-#########################
-
diff --git a/gender-age/mtcnn_detector.py b/gender-age/mtcnn_detector.py
deleted file mode 100644
index 2044f1db0..000000000
--- a/gender-age/mtcnn_detector.py
+++ /dev/null
@@ -1,655 +0,0 @@
-# coding: utf-8
-import os
-import mxnet as mx
-import numpy as np
-import math
-import cv2
-from multiprocessing import Pool
-from itertools import repeat
-from itertools import izip
-from helper import nms, adjust_input, generate_bbox, detect_first_stage_warpper
-
-class MtcnnDetector(object):
-    """
-        Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks
-        see https://github.com/kpzhang93/MTCNN_face_detection_alignment
-        this is a mxnet version
-    """
-    def __init__(self,
-                 model_folder='.',
-                 minsize = 20,
-                 threshold = [0.6, 0.7, 0.8],
-                 factor = 0.709,
-                 num_worker = 1,
-                 accurate_landmark = False,
-                 ctx=mx.cpu()):
-        """
-            Initialize the detector
-
-            Parameters:
-            ----------
-                model_folder : string
-                    path for the models
-                minsize : float number
-                    minimal face to detect
-                threshold : float number
-                    detect threshold for 3 stages
-                factor: float number
-                    scale factor for image pyramid
-                num_worker: int number
-                    number of processes we use for first stage
-                accurate_landmark: bool
-                    use accurate landmark localization or not
-
-        """
-        self.num_worker = num_worker
-        self.accurate_landmark = accurate_landmark
-
-        # load 4 models from folder
-        models = ['det1', 'det2', 'det3','det4']
-        models = [ os.path.join(model_folder, f) for f in models]
-        
-        self.PNets = []
-        for i in range(num_worker):
-            workner_net = mx.model.FeedForward.load(models[0], 1, ctx=ctx)
-            self.PNets.append(workner_net)
-
-        #self.Pool = Pool(num_worker)
-
-        self.RNet = mx.model.FeedForward.load(models[1], 1, ctx=ctx)
-        self.ONet = mx.model.FeedForward.load(models[2], 1, ctx=ctx)
-        self.LNet = mx.model.FeedForward.load(models[3], 1, ctx=ctx)
-
-        self.minsize   = float(minsize)
-        self.factor    = float(factor)
-        self.threshold = threshold
-
-
-    def convert_to_square(self, bbox):
-        """
-            convert bbox to square
-
-        Parameters:
-        ----------
-            bbox: numpy array , shape n x 5
-                input bbox
-
-        Returns:
-        -------
-            square bbox
-        """
-        square_bbox = bbox.copy()
-
-        h = bbox[:, 3] - bbox[:, 1] + 1
-        w = bbox[:, 2] - bbox[:, 0] + 1
-        max_side = np.maximum(h,w)
-        square_bbox[:, 0] = bbox[:, 0] + w*0.5 - max_side*0.5
-        square_bbox[:, 1] = bbox[:, 1] + h*0.5 - max_side*0.5
-        square_bbox[:, 2] = square_bbox[:, 0] + max_side - 1
-        square_bbox[:, 3] = square_bbox[:, 1] + max_side - 1
-        return square_bbox
-
-    def calibrate_box(self, bbox, reg):
-        """
-            calibrate bboxes
-
-        Parameters:
-        ----------
-            bbox: numpy array, shape n x 5
-                input bboxes
-            reg:  numpy array, shape n x 4
-                bboxex adjustment
-
-        Returns:
-        -------
-            bboxes after refinement
-
-        """
-        w = bbox[:, 2] - bbox[:, 0] + 1
-        w = np.expand_dims(w, 1)
-        h = bbox[:, 3] - bbox[:, 1] + 1
-        h = np.expand_dims(h, 1)
-        reg_m = np.hstack([w, h, w, h])
-        aug = reg_m * reg
-        bbox[:, 0:4] = bbox[:, 0:4] + aug
-        return bbox
-
- 
-    def pad(self, bboxes, w, h):
-        """
-            pad the the bboxes, alse restrict the size of it
-
-        Parameters:
-        ----------
-            bboxes: numpy array, n x 5
-                input bboxes
-            w: float number
-                width of the input image
-            h: float number
-                height of the input image
-        Returns :
-        ------s
-            dy, dx : numpy array, n x 1
-                start point of the bbox in target image
-            edy, edx : numpy array, n x 1
-                end point of the bbox in target image
-            y, x : numpy array, n x 1
-                start point of the bbox in original image
-            ex, ex : numpy array, n x 1
-                end point of the bbox in original image
-            tmph, tmpw: numpy array, n x 1
-                height and width of the bbox
-
-        """
-        tmpw, tmph = bboxes[:, 2] - bboxes[:, 0] + 1,  bboxes[:, 3] - bboxes[:, 1] + 1
-        num_box = bboxes.shape[0]
-
-        dx , dy= np.zeros((num_box, )), np.zeros((num_box, ))
-        edx, edy  = tmpw.copy()-1, tmph.copy()-1
-
-        x, y, ex, ey = bboxes[:, 0], bboxes[:, 1], bboxes[:, 2], bboxes[:, 3]
-
-        tmp_index = np.where(ex > w-1)
-        edx[tmp_index] = tmpw[tmp_index] + w - 2 - ex[tmp_index]
-        ex[tmp_index] = w - 1
-
-        tmp_index = np.where(ey > h-1)
-        edy[tmp_index] = tmph[tmp_index] + h - 2 - ey[tmp_index]
-        ey[tmp_index] = h - 1
-
-        tmp_index = np.where(x < 0)
-        dx[tmp_index] = 0 - x[tmp_index]
-        x[tmp_index] = 0
-
-        tmp_index = np.where(y < 0)
-        dy[tmp_index] = 0 - y[tmp_index]
-        y[tmp_index] = 0
-
-        return_list = [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph]
-        return_list = [item.astype(np.int32) for item in return_list]
-
-        return  return_list
-
-    def slice_index(self, number):
-        """
-            slice the index into (n,n,m), m < n
-        Parameters:
-        ----------
-            number: int number
-                number
-        """
-        def chunks(l, n):
-            """Yield successive n-sized chunks from l."""
-            for i in range(0, len(l), n):
-                yield l[i:i + n]
-        num_list = range(number)
-        return list(chunks(num_list, self.num_worker))
-        
-    def detect_face_limited(self, img, det_type=2):
-        height, width, _ = img.shape
-        if det_type>=2:
-          total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-          num_box = total_boxes.shape[0]
-
-          # pad the bbox
-          [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-          # (3, 24, 24) is the input shape for RNet
-          input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)
-
-          for i in range(num_box):
-              tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)
-              tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-              input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))
-
-          output = self.RNet.predict(input_buf)
-
-          # filter the total_boxes with threshold
-          passed = np.where(output[1][:, 1] > self.threshold[1])
-          total_boxes = total_boxes[passed]
-
-          if total_boxes.size == 0:
-              return None
-
-          total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))
-          reg = output[0][passed]
-
-          # nms
-          pick = nms(total_boxes, 0.7, 'Union')
-          total_boxes = total_boxes[pick]
-          total_boxes = self.calibrate_box(total_boxes, reg[pick])
-          total_boxes = self.convert_to_square(total_boxes)
-          total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-        else:
-          total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-        num_box = total_boxes.shape[0]
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 48, 48) is the input shape for ONet
-        input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))
-
-        output = self.ONet.predict(input_buf)
-        #print(output[2])
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[2][:, 1] > self.threshold[2])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))
-        reg = output[1][passed]
-        points = output[0][passed]
-
-        # compute landmark points
-        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-        points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]
-        points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]
-
-        # nms
-        total_boxes = self.calibrate_box(total_boxes, reg)
-        pick = nms(total_boxes, 0.7, 'Min')
-        total_boxes = total_boxes[pick]
-        points = points[pick]
-        
-        if not self.accurate_landmark:
-            return total_boxes, points
-
-        #############################################
-        # extended stage
-        #############################################
-        num_box = total_boxes.shape[0]
-        patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)
-        patchw = np.round(patchw*0.25)
-
-        # make it even
-        patchw[np.where(np.mod(patchw,2) == 1)] += 1
-
-        input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)
-        for i in range(5):
-            x, y = points[:, i], points[:, i+5]
-            x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)
-            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,
-                                                                    width,
-                                                                    height)
-            for j in range(num_box):
-                tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)
-                tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]
-                input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))
-
-        output = self.LNet.predict(input_buf)
-
-        pointx = np.zeros((num_box, 5))
-        pointy = np.zeros((num_box, 5))
-
-        for k in range(5):
-            # do not make a large movement
-            tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)
-            output[k][tmp_index[0]] = 0.5
-
-            pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw
-            pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw
-
-        points = np.hstack([pointx, pointy])
-        points = points.astype(np.int32)
-
-        return total_boxes, points
-
-    def detect_face(self, img, det_type=0):
-        """
-            detect face over img
-        Parameters:
-        ----------
-            img: numpy array, bgr order of shape (1, 3, n, m)
-                input image
-        Retures:
-        -------
-            bboxes: numpy array, n x 5 (x1,y2,x2,y2,score)
-                bboxes
-            points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
-                landmarks
-        """
-
-        # check input
-        height, width, _ = img.shape
-        if det_type==0:
-            MIN_DET_SIZE = 12
-
-            if img is None:
-                return None
-
-            # only works for color image
-            if len(img.shape) != 3:
-                return None
-
-            # detected boxes
-            total_boxes = []
-
-            minl = min( height, width)
-
-            # get all the valid scales
-            scales = []
-            m = MIN_DET_SIZE/self.minsize
-            minl *= m
-            factor_count = 0
-            while minl > MIN_DET_SIZE:
-                scales.append(m*self.factor**factor_count)
-                minl *= self.factor
-                factor_count += 1
-
-            #############################################
-            # first stage
-            #############################################
-            #for scale in scales:
-            #    return_boxes = self.detect_first_stage(img, scale, 0)
-            #    if return_boxes is not None:
-            #        total_boxes.append(return_boxes)
-            
-            sliced_index = self.slice_index(len(scales))
-            total_boxes = []
-            for batch in sliced_index:
-                #local_boxes = self.Pool.map( detect_first_stage_warpper, \
-                #        izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )
-                local_boxes = map( detect_first_stage_warpper, \
-                        izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )
-                total_boxes.extend(local_boxes)
-            
-            # remove the Nones 
-            total_boxes = [ i for i in total_boxes if i is not None]
-
-            if len(total_boxes) == 0:
-                return None
-            
-            total_boxes = np.vstack(total_boxes)
-
-            if total_boxes.size == 0:
-                return None
-
-            # merge the detection from first stage
-            pick = nms(total_boxes[:, 0:5], 0.7, 'Union')
-            total_boxes = total_boxes[pick]
-
-            bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-            bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-
-            # refine the bboxes
-            total_boxes = np.vstack([total_boxes[:, 0]+total_boxes[:, 5] * bbw,
-                                     total_boxes[:, 1]+total_boxes[:, 6] * bbh,
-                                     total_boxes[:, 2]+total_boxes[:, 7] * bbw,
-                                     total_boxes[:, 3]+total_boxes[:, 8] * bbh,
-                                     total_boxes[:, 4]
-                                     ])
-
-            total_boxes = total_boxes.T
-            total_boxes = self.convert_to_square(total_boxes)
-            total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-        else:
-            total_boxes = np.array( [ [0.0, 0.0, img.shape[1], img.shape[0], 0.9] ] ,dtype=np.float32)
-
-        #############################################
-        # second stage
-        #############################################
-        num_box = total_boxes.shape[0]
-
-        # pad the bbox
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 24, 24) is the input shape for RNet
-        input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))
-
-        output = self.RNet.predict(input_buf)
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[1][:, 1] > self.threshold[1])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))
-        reg = output[0][passed]
-
-        # nms
-        pick = nms(total_boxes, 0.7, 'Union')
-        total_boxes = total_boxes[pick]
-        total_boxes = self.calibrate_box(total_boxes, reg[pick])
-        total_boxes = self.convert_to_square(total_boxes)
-        total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])
-
-        #############################################
-        # third stage
-        #############################################
-        num_box = total_boxes.shape[0]
-
-        # pad the bbox
-        [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)
-        # (3, 48, 48) is the input shape for ONet
-        input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)
-
-        for i in range(num_box):
-            tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)
-            tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]
-            input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))
-
-        output = self.ONet.predict(input_buf)
-
-        # filter the total_boxes with threshold
-        passed = np.where(output[2][:, 1] > self.threshold[2])
-        total_boxes = total_boxes[passed]
-
-        if total_boxes.size == 0:
-            return None
-
-        total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))
-        reg = output[1][passed]
-        points = output[0][passed]
-
-        # compute landmark points
-        bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1
-        bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1
-        points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]
-        points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]
-
-        # nms
-        total_boxes = self.calibrate_box(total_boxes, reg)
-        pick = nms(total_boxes, 0.7, 'Min')
-        total_boxes = total_boxes[pick]
-        points = points[pick]
-        
-        if not self.accurate_landmark:
-            return total_boxes, points
-
-        #############################################
-        # extended stage
-        #############################################
-        num_box = total_boxes.shape[0]
-        patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)
-        patchw = np.round(patchw*0.25)
-
-        # make it even
-        patchw[np.where(np.mod(patchw,2) == 1)] += 1
-
-        input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)
-        for i in range(5):
-            x, y = points[:, i], points[:, i+5]
-            x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)
-            [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,
-                                                                    width,
-                                                                    height)
-            for j in range(num_box):
-                tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)
-                tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]
-                input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))
-
-        output = self.LNet.predict(input_buf)
-
-        pointx = np.zeros((num_box, 5))
-        pointy = np.zeros((num_box, 5))
-
-        for k in range(5):
-            # do not make a large movement
-            tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)
-            output[k][tmp_index[0]] = 0.5
-
-            pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw
-            pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw
-
-        points = np.hstack([pointx, pointy])
-        points = points.astype(np.int32)
-
-        return total_boxes, points
-
-
-
-    def list2colmatrix(self, pts_list):
-        """
-            convert list to column matrix
-        Parameters:
-        ----------
-            pts_list:
-                input list
-        Retures:
-        -------
-            colMat: 
-
-        """
-        assert len(pts_list) > 0
-        colMat = []
-        for i in range(len(pts_list)):
-            colMat.append(pts_list[i][0])
-            colMat.append(pts_list[i][1])
-        colMat = np.matrix(colMat).transpose()
-        return colMat
-
-    def find_tfrom_between_shapes(self, from_shape, to_shape):
-        """
-            find transform between shapes
-        Parameters:
-        ----------
-            from_shape: 
-            to_shape: 
-        Retures:
-        -------
-            tran_m:
-            tran_b:
-        """
-        assert from_shape.shape[0] == to_shape.shape[0] and from_shape.shape[0] % 2 == 0
-
-        sigma_from = 0.0
-        sigma_to = 0.0
-        cov = np.matrix([[0.0, 0.0], [0.0, 0.0]])
-
-        # compute the mean and cov
-        from_shape_points = from_shape.reshape(from_shape.shape[0]/2, 2)
-        to_shape_points = to_shape.reshape(to_shape.shape[0]/2, 2)
-        mean_from = from_shape_points.mean(axis=0)
-        mean_to = to_shape_points.mean(axis=0)
-
-        for i in range(from_shape_points.shape[0]):
-            temp_dis = np.linalg.norm(from_shape_points[i] - mean_from)
-            sigma_from += temp_dis * temp_dis
-            temp_dis = np.linalg.norm(to_shape_points[i] - mean_to)
-            sigma_to += temp_dis * temp_dis
-            cov += (to_shape_points[i].transpose() - mean_to.transpose()) * (from_shape_points[i] - mean_from)
-
-        sigma_from = sigma_from / to_shape_points.shape[0]
-        sigma_to = sigma_to / to_shape_points.shape[0]
-        cov = cov / to_shape_points.shape[0]
-
-        # compute the affine matrix
-        s = np.matrix([[1.0, 0.0], [0.0, 1.0]])
-        u, d, vt = np.linalg.svd(cov)
-
-        if np.linalg.det(cov) < 0:
-            if d[1] < d[0]:
-                s[1, 1] = -1
-            else:
-                s[0, 0] = -1
-        r = u * s * vt
-        c = 1.0
-        if sigma_from != 0:
-            c = 1.0 / sigma_from * np.trace(np.diag(d) * s)
-
-        tran_b = mean_to.transpose() - c * r * mean_from.transpose()
-        tran_m = c * r
-
-        return tran_m, tran_b
-
-    def extract_image_chips(self, img, points, desired_size=256, padding=0):
-        """
-            crop and align face
-        Parameters:
-        ----------
-            img: numpy array, bgr order of shape (1, 3, n, m)
-                input image
-            points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
-            desired_size: default 256
-            padding: default 0
-        Retures:
-        -------
-            crop_imgs: list, n
-                cropped and aligned faces 
-        """
-        crop_imgs = []
-        for p in points:
-            shape  =[]
-            for k in range(len(p)/2):
-                shape.append(p[k])
-                shape.append(p[k+5])
-
-            if padding > 0:
-                padding = padding
-            else:
-                padding = 0
-            # average positions of face points
-            mean_face_shape_x = [0.224152, 0.75610125, 0.490127, 0.254149, 0.726104]
-            mean_face_shape_y = [0.2119465, 0.2119465, 0.628106, 0.780233, 0.780233]
-
-            from_points = []
-            to_points = []
-
-            for i in range(len(shape)/2):
-                x = (padding + mean_face_shape_x[i]) / (2 * padding + 1) * desired_size
-                y = (padding + mean_face_shape_y[i]) / (2 * padding + 1) * desired_size
-                to_points.append([x, y])
-                from_points.append([shape[2*i], shape[2*i+1]])
-
-            # convert the points to Mat
-            from_mat = self.list2colmatrix(from_points)
-            to_mat = self.list2colmatrix(to_points)
-
-            # compute the similar transfrom
-            tran_m, tran_b = self.find_tfrom_between_shapes(from_mat, to_mat)
-
-            probe_vec = np.matrix([1.0, 0.0]).transpose()
-            probe_vec = tran_m * probe_vec
-
-            scale = np.linalg.norm(probe_vec)
-            angle = 180.0 / math.pi * math.atan2(probe_vec[1, 0], probe_vec[0, 0])
-
-            from_center = [(shape[0]+shape[2])/2.0, (shape[1]+shape[3])/2.0]
-            to_center = [0, 0]
-            to_center[1] = desired_size * 0.4
-            to_center[0] = desired_size * 0.5
-
-            ex = to_center[0] - from_center[0]
-            ey = to_center[1] - from_center[1]
-
-            rot_mat = cv2.getRotationMatrix2D((from_center[0], from_center[1]), -1*angle, scale)
-            rot_mat[0][2] += ex
-            rot_mat[1][2] += ey
-
-            chips = cv2.warpAffine(img, rot_mat, (desired_size, desired_size))
-            crop_imgs.append(chips)
-
-        return crop_imgs
-
diff --git a/gender-age/symbol_utils.py b/gender-age/symbol_utils.py
deleted file mode 100644
index e23b01bbc..000000000
--- a/gender-age/symbol_utils.py
+++ /dev/null
@@ -1,182 +0,0 @@
-import mxnet as mx
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    return body
-
-bn_mom = 0.9
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=bn_mom)    
-    return bn
-
-def get_fc1(last_conv, num_classes, fc_type):
-  bn_mom = 0.9
-  body = last_conv
-  if fc_type=='Z':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = body
-  elif fc_type=='E':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GAP':
-    bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    relu1 = Act(data=bn1, act_type='relu', name='relu1')
-    # Although kernel is not used here when global_pool=True, we should put one
-    pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
-    flat = mx.sym.Flatten(data=pool1)
-    fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GNAP': #mobilefacenet++
-    filters_in = 512 # param in mobilefacenet
-    if num_classes>filters_in:
-      body = mx.sym.Convolution(data=last_conv, num_filter=num_classes, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True, name='convx')
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=0.9, name='convx_bn')
-      body = Act(data=body, act_type='relu', name='convx_relu')
-      filters_in = num_classes
-    else:
-      body = last_conv
-    body = mx.sym.BatchNorm(data=body, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6f')  
-
-    spatial_norm=body*body
-    spatial_norm=mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
-    spatial_sqrt=mx.sym.sqrt(spatial_norm)
-    #spatial_mean=mx.sym.mean(spatial_sqrt, axis=(1,2,3), keepdims=True)
-    spatial_mean=mx.sym.mean(spatial_sqrt)
-    spatial_div_inverse=mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
-
-    spatial_attention_inverse=mx.symbol.tile(spatial_div_inverse, reps=(1,filters_in,1,1))   
-    body=body*spatial_attention_inverse
-    #body = mx.sym.broadcast_mul(body, spatial_div_inverse)
-
-    fc1 = mx.sym.Pooling(body, kernel=(7, 7), global_pool=True, pool_type='avg')
-    if num_classes<filters_in:
-      fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6w')
-      fc1 = mx.sym.FullyConnected(data=fc1, num_hidden=num_classes, name='pre_fc1')
-    else:
-      fc1 = mx.sym.Flatten(data=fc1)
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='fc1')
-  elif fc_type=="GDC": #mobilefacenet_v1
-    conv_6_dw = Linear(last_conv, num_filter=512, num_group=512, kernel=(7,7), pad=(0, 0), stride=(1, 1), name="conv_6dw7_7")  
-    conv_6_f = mx.sym.FullyConnected(data=conv_6_dw, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=conv_6_f, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='F':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='G':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='H':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='I':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='J':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  else:
-    bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    relu1 = Act(data=bn1, act_type='relu', name='relu1')
-    # Although kernel is not used here when global_pool=True, we should put one
-    pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
-    flat = mx.sym.Flatten(data=pool1)
-    if len(fc_type)>1:
-      if fc_type[1]=='X':
-        print('dropout mode')
-        flat = mx.symbol.Dropout(data=flat, p=0.2)
-      fc_type = fc_type[0]
-    if fc_type=='A':
-      fc1 = flat
-    else:
-      #B-D
-      #B
-      fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-      if fc_type=='C':
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-      elif fc_type=='D':
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-        fc1 = Act(data=fc1, act_type='relu', name='fc1_relu')
-  return fc1
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type='relu', name=name + '_relu1')
-    conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn3 + shortcut
-
-
-def get_head(data, version_input, num_filter):
-    bn_mom = 0.9
-    workspace = 256
-    kwargs = {'bn_mom': bn_mom, 'workspace' : workspace}
-    data = data-127.5
-    data = data*0.0078125
-    #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-    if version_input==0:
-      body = Conv(data=data, num_filter=num_filter, kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type='relu', name='relu0')
-      body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    else:
-      body = data
-      _num_filter = min(num_filter, 64)
-      body = Conv(data=body, num_filter=_num_filter, kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type='relu', name='relu0')
-      body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
-    return body
-
-
diff --git a/gender-age/test.py b/gender-age/test.py
deleted file mode 100644
index 663a8abbd..000000000
--- a/gender-age/test.py
+++ /dev/null
@@ -1,34 +0,0 @@
-import face_model
-import argparse
-import cv2
-import sys
-import numpy as np
-import datetime
-
-parser = argparse.ArgumentParser(description='face model test')
-# general
-parser.add_argument('--image-size', default='112,112', help='')
-parser.add_argument('--image', default='Tom_Hanks_54745.png', help='')
-parser.add_argument('--model', default='model/model,0', help='path to load model.')
-parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-parser.add_argument('--det', default=0, type=int, help='mtcnn option, 1 means using R+O, 0 means detect from begining')
-args = parser.parse_args()
-
-model = face_model.FaceModel(args)
-#img = cv2.imread('Tom_Hanks_54745.png')
-img = cv2.imread(args.image)
-img = model.get_input(img)
-#f1 = model.get_feature(img)
-#print(f1[0:10])
-for _ in range(5):
-  gender, age = model.get_ga(img)
-time_now = datetime.datetime.now()
-count = 200
-for _ in range(count):
-  gender, age = model.get_ga(img)
-time_now2 = datetime.datetime.now()
-diff = time_now2 - time_now
-print('time cost', diff.total_seconds()/count)
-print('gender is',gender)
-print('age is', age)
-
diff --git a/gender-age/train.py b/gender-age/train.py
deleted file mode 100644
index effeadb48..000000000
--- a/gender-age/train.py
+++ /dev/null
@@ -1,329 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import numpy as np
-import sklearn
-from data import FaceImageIter
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-#import face_image
-import fresnet
-import fmobilenet 
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-AGE=100
-
-
-args = None
-
-
-class AccMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(AccMetric, self).__init__(
-        'acc', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0].asnumpy()[:,0:1]
-    pred_label = preds[-1].asnumpy()[:,0:2]
-    pred_label = np.argmax(pred_label, axis=self.axis)
-    pred_label = pred_label.astype('int32').flatten()
-    label = label.astype('int32').flatten()
-    assert label.shape==pred_label.shape
-    self.sum_metric += (pred_label.flat == label.flat).sum()
-    self.num_inst += len(pred_label.flat)
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[-1].asnumpy()[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-    gt_label = preds[-2].asnumpy()
-    #print(gt_label)
-
-class MAEMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(MAEMetric, self).__init__(
-        'MAE', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0].asnumpy()
-    label_age = np.count_nonzero(label[:,1:], axis=1)
-    pred_age = np.zeros( label_age.shape, dtype=np.int)
-    #pred_age = np.zeros( label_age.shape, dtype=np.float32)
-    pred = preds[-1].asnumpy()
-    for i in range(AGE):
-      _pred = pred[:,2+i*2:4+i*2]
-      _pred = np.argmax(_pred, axis=1)
-      #pred = pred[:,1]
-      pred_age += _pred
-    #pred_age = pred_age.astype(np.int)
-    mae = np.mean(np.abs(label_age - pred_age))
-    self.sum_metric += mae
-    self.num_inst += 1.0
-
-class CUMMetric(mx.metric.EvalMetric):
-  def __init__(self, n=5):
-    self.axis = 1
-    self.n = n
-    super(CUMMetric, self).__init__(
-        'CUM_%d'%n, axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0].asnumpy()
-    label_age = np.count_nonzero(label[:,1:], axis=1)
-    pred_age = np.zeros( label_age.shape, dtype=np.int)
-    pred = preds[-1].asnumpy()
-    for i in range(AGE):
-      _pred = pred[:,2+i*2:4+i*2]
-      _pred = np.argmax(_pred, axis=1)
-      #pred = pred[:,1]
-      pred_age += _pred
-    diff = np.abs(label_age - pred_age)
-    cum = np.sum( (diff<self.n) )
-    self.sum_metric += cum
-    self.num_inst += len(label_age)
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--data-dir', default='', help='training set directory')
-  parser.add_argument('--prefix', default='../model/model', help='directory to save model.')
-  parser.add_argument('--pretrained', default='', help='pretrained model to load')
-  parser.add_argument('--ckpt', type=int, default=1, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--loss-type', type=int, default=4, help='loss type')
-  parser.add_argument('--verbose', type=int, default=2000, help='do verification testing and model saving every verbose batches')
-  parser.add_argument('--max-steps', type=int, default=0, help='max training batches')
-  parser.add_argument('--end-epoch', type=int, default=100000, help='training epoch size.')
-  parser.add_argument('--network', default='r50', help='specify network')
-  parser.add_argument('--image-size', default='112,112', help='specify input image height and width')
-  parser.add_argument('--version-input', type=int, default=1, help='network input config')
-  parser.add_argument('--version-output', type=str, default='GAP', help='network embedding output config')
-  parser.add_argument('--version-act', type=str, default='prelu', help='network activation config')
-  parser.add_argument('--multiplier', type=float, default=1.0, help='')
-  parser.add_argument('--lr', type=float, default=0.1, help='start learning rate')
-  parser.add_argument('--lr-steps', type=str, default='', help='steps of lr changing')
-  parser.add_argument('--wd', type=float, default=0.0005, help='weight decay')
-  parser.add_argument('--bn-mom', type=float, default=0.9, help='bn mom')
-  parser.add_argument('--mom', type=float, default=0.9, help='momentum')
-  parser.add_argument('--per-batch-size', type=int, default=128, help='batch size in each context')
-  parser.add_argument('--rand-mirror', type=int, default=1, help='if do random mirror in training')
-  parser.add_argument('--cutoff', type=int, default=0, help='cut off aug')
-  parser.add_argument('--color', type=int, default=0, help='color jittering aug')
-  parser.add_argument('--ce-loss', default=False, action='store_true', help='if output ce loss')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol(args, arg_params, aux_params):
-  data_shape = (args.image_channel,args.image_h,args.image_w)
-  image_shape = ",".join([str(x) for x in data_shape])
-  margin_symbols = []
-  if args.network[0]=='m':
-    fc1 = fmobilenet.get_symbol(AGE*2+2, 
-        multiplier = args.multiplier,
-        version_input=args.version_input, 
-        version_output=args.version_output)
-  else:
-    fc1 = fresnet.get_symbol(AGE*2+2, args.num_layers,
-        version_input=args.version_input, 
-        version_output=args.version_output)
-  label = mx.symbol.Variable('softmax_label')
-  gender_label = mx.symbol.slice_axis(data = label, axis=1, begin=0, end=1)
-  gender_label = mx.symbol.reshape(gender_label, shape=(args.per_batch_size,))
-  gender_fc1 = mx.symbol.slice_axis(data = fc1, axis=1, begin=0, end=2)
-  #gender_fc7 = mx.sym.FullyConnected(data=gender_fc1, num_hidden=2, name='gender_fc7')
-  gender_softmax = mx.symbol.SoftmaxOutput(data=gender_fc1, label = gender_label, name='gender_softmax', normalization='valid', use_ignore=True, ignore_label = 9999)
-  outs = [gender_softmax]
-  for i in range(AGE):
-    age_label = mx.symbol.slice_axis(data = label, axis=1, begin=i+1, end=i+2)
-    age_label = mx.symbol.reshape(age_label, shape=(args.per_batch_size,))
-    age_fc1 = mx.symbol.slice_axis(data = fc1, axis=1, begin=2+i*2, end=4+i*2)
-    #age_fc7 = mx.sym.FullyConnected(data=age_fc1, num_hidden=2, name='age_fc7_%i'%i)
-    age_softmax = mx.symbol.SoftmaxOutput(data=age_fc1, label = age_label, name='age_softmax_%d'%i, normalization='valid', grad_scale=1)
-    outs.append(age_softmax)
-  outs.append(mx.sym.BlockGrad(fc1))
-
-  out = mx.symbol.Group(outs)
-  return (out, arg_params, aux_params)
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in range(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = args.prefix
-    prefix_dir = os.path.dirname(prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    end_epoch = args.end_epoch
-    args.ctx_num = len(ctx)
-    args.num_layers = int(args.network[1:])
-    print('num_layers', args.num_layers)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = 3
-
-    data_dir_list = args.data_dir.split(',')
-    assert len(data_dir_list)==1
-    data_dir = data_dir_list[0]
-    path_imgrec = None
-    path_imglist = None
-    image_size = [int(x) for x in args.image_size.split(',')]
-    assert len(image_size)==2
-    assert image_size[0]==image_size[1]
-    args.image_h = image_size[0]
-    args.image_w = image_size[1]
-    print('image_size', image_size)
-    path_imgrec = os.path.join(data_dir, "train.rec")
-    path_imgrec_val = os.path.join(data_dir, "val.rec")
-
-
-    print('Called with argument:', args)
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-    mean = None
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    if len(args.pretrained)==0:
-      arg_params = None
-      aux_params = None
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-    else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      _, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-
-    #label_name = 'softmax_label'
-    #label_shape = (args.batch_size,)
-    model = mx.mod.Module(
-        context       = ctx,
-        symbol        = sym,
-    )
-    val_dataiter = None
-
-    train_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec,
-        shuffle              = True,
-        rand_mirror          = args.rand_mirror,
-        mean                 = mean,
-        cutoff               = args.cutoff,
-        color_jittering      = args.color,
-    )
-    val_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec_val,
-        shuffle              = False,
-        rand_mirror          = False,
-        mean                 = mean,
-    )
-
-    metric = mx.metric.CompositeEvalMetric([AccMetric(), MAEMetric(), CUMMetric()])
-
-    if args.network[0]=='r' or args.network[0]=='y':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    elif args.network[0]=='i' or args.network[0]=='x':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2) #inception
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-    _rescale = 1.0/args.ctx_num
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    #opt = optimizer.Nadam(learning_rate=base_lr, wd=base_wd, rescale_grad=_rescale)
-    som = 20
-    _cb = mx.callback.Speedometer(args.batch_size, som)
-    lr_steps = [int(x) for x in args.lr_steps.split(',')]
-
-    global_step = [0]
-    def _batch_callback(param):
-      _cb(param)
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for _lr in lr_steps:
-        if mbatch==_lr:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-      if mbatch==lr_steps[-1]:
-        arg, aux = model.get_params()
-        all_layers = model.symbol.get_internals()
-        _sym = all_layers['fc1_output']
-        mx.model.save_checkpoint(args.prefix, 0, _sym, arg, aux)
-        sys.exit(0)
-
-    epoch_cb = None
-    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
-    print('start fitting')
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = end_epoch,
-        eval_data          = val_dataiter,
-        eval_metric        = metric,
-        kvstore            = 'device',
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    #time.sleep(3600*6.5)
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/generation/README.md b/generation/README.md
new file mode 100644
index 000000000..9f8b7ecaf
--- /dev/null
+++ b/generation/README.md
@@ -0,0 +1 @@
+# InsightFace Generation Projects
diff --git a/iccv19-challenge/README.md b/iccv19-challenge/README.md
deleted file mode 100644
index 3052ecf18..000000000
--- a/iccv19-challenge/README.md
+++ /dev/null
@@ -1,111 +0,0 @@
-[The Lightweight Face Recognition Challenge & Workshop](https://ibug.doc.ic.ac.uk/resources/lightweight-face-recognition-challenge-workshop/) will be held in conjunction with the International Conference on Computer Vision (ICCV) 2019, Seoul Korea. 
-
-Please strictly follow the rules. For example, please use the same [method](https://github.com/deepinsight/insightface/blob/master/common/flops_counter.py) for the FLOPs calculation regardless of your training framework is insightface or not.
-
-[Test Server](http://www.insightface-challenge.com/overview) 
-
-**Sponsors:**
-
-The Lightweight Face Recognition Challenge has been supported by 
-
-EPSRC project FACER2VM (EP/N007743/1)
-
-Huawei (5000$)
-
-DeepGlint (3000$)
-
-iQIYI (3000$)
-
-Kingsoft Cloud (3000$)
-
-Pensees (3000$)
-
-Dynamic funding pool: (17000$)
-
-Cash sponsors and gift donations are welcome.
-
-Contact:
-insightface.challenge@gmail.com
-
-**Discussion Group**
-
-*For Chinese:*
-
-![wechat](https://github.com/deepinsight/insightface/blob/master/resources/lfr19_wechat1.jpg)
-
-*For English:*
-
-(in #lfr2019 channel)
-https://join.slack.com/t/insightface/shared_invite/enQtNjU0NDk2MjYyMTMzLTIzNDEwNmIxMjU5OGYzYzFhMjlkNjlhMTBkNWFiNjU4MTVhNTgzYjQ5ZTZiMGM3MzUyNzQ3OTBhZTg3MzM5M2I
-
-
-**NEWS**
-
-``2019.06.21`` We updated the groundtruth of Glint test dataset.
-
-``2019.06.04`` We will clean the groundtruth on deepglint testset.
-
-``2019.05.21`` Baseline models and training logs available.
-
-``2019.05.16`` The four tracks (deepglint-light, deepglint-large, iQIYI-light, iQIYI-large) will equally share the dynamic funding pool (14000$). From each track, the top 3 players will share the funding pool for 50%, 30% and 20% respectively.
-
-==================
-
-**How To Start:**
-
-**Training:**
-
-1. Download ms1m-retinaface from [baiducloud](https://pan.baidu.com/s/1rQxJ3drqm_071vpxBtp98A) or [dropbox](https://www.dropbox.com/s/ev5ezzcz79p2hge/ms1m-retinaface-t1.zip?dl=0) and unzip it to `$INSIGHTFACE_ROOT/datasets/`
-2. Go into `$INSIGHTFACE_ROOT/recognition/`
-3. Refer to the `retina` dataset configuration section in `sample_config.py` and copy it as your own configuration file `config.py`.
-4. Start training with `CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --dataset retina --network [your-network] --loss arcface`. It will output the accuracy of lfw, cfp_fp and agedb_30 every 2000 batches by default.
-5. Putting the training dataset on SSD hard disk will achieve better training efficiency.
-
-------------------
-
-**Testing:**
-
-1. Download testdata-image from [baiducloud](https://pan.baidu.com/s/1UKUYsRfVTSzj1tfU3BVFrw) or [dropbox](https://www.dropbox.com/s/r5y6xt754m36rh8/iccv19-challenge-data-v1.zip?dl=0). These face images are all pre-processed and aligned.
-2. To download testdata-video from iQIYI, please visit <http://challenge.ai.iqiyi.com/data-cluster>. You need to download iQIYI-VID-FACE.z01, iQIYI-VID-FACE.z02 and iQIYI-VID-FACE.zip after registration. These face frames are also pre-processed and aligned.
-   1. Unzip: ``zip iQIYI_VID_FACE.zip -s=0 --out iQIYI_VID_FACE_ALL.zip; unzip iQIYI_VID_FACE_ALL.zip``
-   2. We can get a directory named ``iQIYI_VID_FACE`` after decompression. Then, we have to move ``video_filelist.txt`` in testdata-image package to ``iQIYI_VID_FACE/filelist.txt``, to indicate the order of videos in our submission feature file.
-3. To generate image feature submission file: check ``gen_image_feature.py``
-4. To generate video feature submission file: check ``gen_video_feature.py``
-5. Submit binary feature to the right track of the test server.
-
-You can also check the verification performance during training time on LFW,CFP_FP,AgeDB_30 datasets.
-
-------------------
-
-**Evaluation:**
-
-Final ranking is determined by the TAR under 1:1 protocal only, for all valid submissions. 
-
-For image testset, we evaluate the TAR under FAR@e-8 while we choose the TAR under FAR@e-4 for video testset.
-
-------------------
-
-**Baseline:**
-
-1. Network y2(a deeper mobilefacenet): 933M FLOPs. TAR_image: 0.64691, TAR_video: 0.47191
-2. Network r100fc(ResNet100FC-IR): 24G FLOPs. TAR_image: 0.80312, TAR_video: 0.64894
-
-Baseline models download link: [baidu cloud](https://pan.baidu.com/s/1Em0ZFnefSoTsZoTd-9m8Nw)    [dropbox](https://www.dropbox.com/s/yqaziktiv38ehrv/iccv19-baseline-models.zip?dl=0)
-
-Training logs: [baidu cloud](https://pan.baidu.com/s/12rsp-oMzsjTeU6nugEvA9g)   [dropbox](https://www.dropbox.com/s/4ufb9g7n76rfav5/iccv-baseline-log.zip?dl=0)
-
-------------------
-
-**Discussion:**
-
-[https://github.com/deepinsight/insightface/issues/632](https://github.com/deepinsight/insightface/issues/632)
-
-------------------
-
-**Candidate solutions:**
-
-1. Manually design or automatically search different networks/losses.
-2. Use slightly deeper or wider mobile-level networks.
-3. [OctConv](https://arxiv.org/abs/1904.05049), to reduce FLOPs.
-4. [HRNet](https://arxiv.org/abs/1904.04514), for large FLOPs track.
-and so on
diff --git a/iccv19-challenge/gen_image_feature.py b/iccv19-challenge/gen_image_feature.py
deleted file mode 100644
index 55453ac07..000000000
--- a/iccv19-challenge/gen_image_feature.py
+++ /dev/null
@@ -1,153 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-
-from datetime import datetime
-import os.path
-from easydict import EasyDict as edict
-import time
-import json
-import sys
-import numpy as np
-import importlib
-import itertools
-import argparse
-import struct
-import cv2
-import sklearn
-from sklearn.preprocessing import normalize
-import mxnet as mx
-from mxnet import ndarray as nd
-
-image_shape = None
-net = None
-data_size = 1862120
-emb_size = 0
-use_flip = True
-
-
-
-def do_flip(data):
-  for idx in range(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_feature(buffer):
-  global emb_size
-  if use_flip:
-    input_blob = np.zeros( (len(buffer)*2, 3, image_shape[1], image_shape[2] ) )
-  else:
-    input_blob = np.zeros( (len(buffer), 3, image_shape[1], image_shape[2] ) )
-  idx = 0
-  for item in buffer:
-    img = cv2.imread(item)[:,:,::-1] #to rgb
-    img = np.transpose( img, (2,0,1) )
-    attempts = [0,1] if use_flip else [0]
-    for flipid in attempts:
-      _img = np.copy(img)
-      if flipid==1:
-        do_flip(_img)
-      input_blob[idx] = _img
-      idx+=1
-  data = mx.nd.array(input_blob)
-  db = mx.io.DataBatch(data=(data,))
-  net.model.forward(db, is_train=False)
-  _embedding = net.model.get_outputs()[0].asnumpy()
-  if emb_size==0:
-    emb_size = _embedding.shape[1]
-    print('set emb_size to ', emb_size)
-  embedding = np.zeros( (len(buffer), emb_size), dtype=np.float32 )
-  if use_flip:
-    embedding1 = _embedding[0::2]
-    embedding2 = _embedding[1::2]
-    embedding = embedding1+embedding2
-  else:
-    embedding = _embedding
-  embedding = sklearn.preprocessing.normalize(embedding)
-  return embedding
-
-
-
-def write_bin(path, m):
-  rows, cols = m.shape
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', rows,cols,cols*4,5))
-    f.write(m.data)
-
-def main(args):
-  global image_shape
-  global net
-
-  print(args)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in range(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  image_shape = [int(x) for x in args.image_size.split(',')]
-  vec = args.model.split(',')
-  assert len(vec)>1
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  net = edict()
-  net.ctx = ctx
-  net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(prefix, epoch)
-  #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
-  all_layers = net.sym.get_internals()
-  net.sym = all_layers['fc1_output']
-  net.model = mx.mod.Module(symbol=net.sym, context=net.ctx, label_names = None)
-  net.model.bind(data_shapes=[('data', (args.batch_size, 3, image_shape[1], image_shape[2]))])
-  net.model.set_params(net.arg_params, net.aux_params)
-
-  features_all = None
-
-  i = 0
-  fstart = 0
-  buffer = []
-  for line in open(os.path.join(args.input, 'filelist.txt'), 'r'):
-    if i%1000==0:
-      print("processing ",i)
-    i+=1
-    line = line.strip()
-    image_path = os.path.join(args.input, line)
-    buffer.append(image_path)
-    if len(buffer)==args.batch_size:
-      embedding = get_feature(buffer)
-      buffer = []
-      fend = fstart+embedding.shape[0]
-      if features_all is None:
-        features_all = np.zeros( (data_size, emb_size), dtype=np.float32 )
-      #print('writing', fstart, fend)
-      features_all[fstart:fend,:] = embedding
-      fstart = fend
-  if len(buffer)>0:
-    embedding = get_feature(buffer)
-    fend = fstart+embedding.shape[0]
-    print('writing', fstart, fend)
-    features_all[fstart:fend,:] = embedding
-  write_bin(args.output, features_all)
-  #os.system("bypy upload %s"%args.output)
-
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--batch_size', type=int, help='', default=32)
-  parser.add_argument('--image_size', type=str, help='', default='3,112,112')
-  parser.add_argument('--input', type=str, help='', default='')
-  parser.add_argument('--output', type=str, help='', default='')
-  parser.add_argument('--model', type=str, help='', default='')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/iccv19-challenge/gen_video_feature.py b/iccv19-challenge/gen_video_feature.py
deleted file mode 100644
index 14f4a1486..000000000
--- a/iccv19-challenge/gen_video_feature.py
+++ /dev/null
@@ -1,204 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-
-from datetime import datetime
-import os.path
-from easydict import EasyDict as edict
-import time
-import json
-import glob
-import sys
-import numpy as np
-import importlib
-import itertools
-import argparse
-import struct
-import cv2
-import sklearn
-from sklearn.preprocessing import normalize
-import mxnet as mx
-from mxnet import ndarray as nd
-
-image_shape = None
-net = None
-data_size = 203848
-emb_size = 0
-use_flip = False
-ctx_num = 0
-
-
-
-def do_flip(data):
-  for idx in range(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_feature(buffer):
-  global emb_size
-  input_count = len(buffer)
-  if use_flip:
-    input_count *= 2
-  network_count = input_count
-  if input_count%ctx_num!=0:
-    network_count = (input_count//ctx_num+1)*ctx_num
-
-  input_blob = np.zeros( (network_count, 3, image_shape[1], image_shape[2]), dtype=np.float32)
-  idx = 0
-  for item in buffer:
-    img = cv2.imread(item)[:,:,::-1] #to rgb
-    img = np.transpose( img, (2,0,1) )
-    attempts = [0,1] if use_flip else [0]
-    for flipid in attempts:
-      _img = np.copy(img)
-      if flipid==1:
-        do_flip(_img)
-      input_blob[idx] = _img
-      idx+=1
-  data = mx.nd.array(input_blob)
-  db = mx.io.DataBatch(data=(data,))
-  net.model.forward(db, is_train=False)
-  _embedding = net.model.get_outputs()[0].asnumpy()
-  _embedding = _embedding[0:input_count]
-  if emb_size==0:
-    emb_size = _embedding.shape[1]
-    print('set emb_size to ', emb_size)
-  embedding = np.zeros( (len(buffer), emb_size), dtype=np.float32 )
-  if use_flip:
-    embedding1 = _embedding[0::2]
-    embedding2 = _embedding[1::2]
-    embedding = embedding1+embedding2
-  else:
-    embedding = _embedding
-  embedding = sklearn.preprocessing.normalize(embedding)
-  return embedding
-
-def write_bin(path, m):
-  rows, cols = m.shape
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', rows,cols,cols*4,5))
-    f.write(m.data)
-
-def main(args):
-  global image_shape
-  global net
-  global ctx_num
-
-  print(args)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in range(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  ctx_num = len(ctx)
-  image_shape = [int(x) for x in args.image_size.split(',')]
-  vec = args.model.split(',')
-  assert len(vec)>1
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  net = edict()
-  net.ctx = ctx
-  net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(prefix, epoch)
-  #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
-  all_layers = net.sym.get_internals()
-  net.sym = all_layers['fc1_output']
-  net.model = mx.mod.Module(symbol=net.sym, context=net.ctx, label_names = None)
-  net.model.bind(data_shapes=[('data', (args.batch_size, 3, image_shape[1], image_shape[2]))])
-  net.model.set_params(net.arg_params, net.aux_params)
-
-  features_all = None
-
-  i = 0
-  filelist = os.path.join(args.input, 'filelist.txt')
-  #print(filelist)
-  buffer_images = []
-  buffer_embedding = np.zeros( (0,0), dtype=np.float32)
-  aggr_nums = []
-  row_idx = 0
-  for line in open(filelist, 'r'):
-    if i%1000==0:
-      print("processing ",i)
-    i+=1
-    #print('stat', i, len(buffer_images), buffer_embedding.shape, aggr_nums, row_idx)
-    videoname = line.strip().split()[0]
-    images = glob.glob("%s/%s/*.jpg"%(args.input, videoname))
-    assert len(images)>0
-    image_features = []
-    for image_path in images:
-      buffer_images.append(image_path)
-    aggr_nums.append(len(images))
-    while len(buffer_images)>=args.batch_size:
-      embedding = get_feature(buffer_images[0:args.batch_size])
-      buffer_images = buffer_images[args.batch_size:]
-      if buffer_embedding.shape[1]==0:
-        buffer_embedding = embedding.copy()
-      else:
-        buffer_embedding = np.concatenate( (buffer_embedding, embedding), axis=0)
-    buffer_idx = 0
-    acount = 0
-    for anum in aggr_nums:
-      if buffer_embedding.shape[0]>=anum+buffer_idx:
-        image_features = buffer_embedding[buffer_idx:buffer_idx+anum]
-        video_feature = np.sum(image_features, axis=0, keepdims=True)
-        video_feature = sklearn.preprocessing.normalize(video_feature)
-        if features_all is None:
-          features_all = np.zeros( (data_size, video_feature.shape[1]), dtype=np.float32 )
-        #print('write to', row_idx, anum, buffer_embedding.shape)
-        features_all[row_idx] = video_feature.flatten()
-        row_idx+=1
-        buffer_idx += anum
-        acount+=1
-      else:
-        break
-    aggr_nums = aggr_nums[acount:]
-    buffer_embedding = buffer_embedding[buffer_idx:]
-
-  if len(buffer_images)>0:
-    embedding = get_feature(buffer_images)
-    buffer_images = buffer_images[args.batch_size:]
-    buffer_embedding = np.concatenate( (buffer_embedding, embedding), axis=0)
-  buffer_idx = 0
-  acount = 0
-  for anum in aggr_nums:
-    assert buffer_embedding.shape[0]>=anum+buffer_idx
-    image_features = buffer_embedding[buffer_idx:buffer_idx+anum]
-    video_feature = np.sum(image_features, axis=0, keepdims=True)
-    video_feature = sklearn.preprocessing.normalize(video_feature)
-    #print('last write to', row_idx, anum, buffer_embedding.shape)
-    features_all[row_idx] = video_feature.flatten()
-    row_idx+=1
-    buffer_idx += anum
-    acount+=1
-
-  aggr_nums = aggr_nums[acount:]
-  buffer_embedding = buffer_embedding[buffer_idx:]
-  assert len(aggr_nums)==0
-  assert buffer_embedding.shape[0]==0
-
-  write_bin(args.output, features_all)
-  print(row_idx, features_all.shape)
-  #os.system("bypy upload %s"%args.output)
-
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--batch_size', type=int, help='', default=32)
-  parser.add_argument('--image_size', type=str, help='', default='3,112,112')
-  parser.add_argument('--input', type=str, help='', default='./testdata-video')
-  parser.add_argument('--output', type=str, help='', default='')
-  parser.add_argument('--model', type=str, help='', default='')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/model_zoo/README.md b/model_zoo/README.md
new file mode 100644
index 000000000..1991b989b
--- /dev/null
+++ b/model_zoo/README.md
@@ -0,0 +1,187 @@
+# InsightFace Model Zoo
+
+:bell:   **ALL models are available for non-commercial research purposes only.**
+
+## 0. Python Package models
+
+To check the detail of insightface python package, please see [here](../python-package).
+
+To install: ``pip install -U insightface``
+
+To use the specific model pack:
+
+```
+model_pack_name = 'buffalo_l'
+app = FaceAnalysis(name=model_pack_name)
+```
+
+Name in **bold** is the default model pack in latest version.
+
+
+| Name           | Detection Model | Recognition Model   | Alignment    | Attributes | Model-Size |
+| -------------- | --------------- | ------------------- | ------------ | ---------- | ---------- |
+| antelopev2 | RetinaFace-10GF      | ResNet100@Glint360K | 2d106 & 3d68 | Gender&Age | 407MB |
+| **buffalo_l**      | RetinaFace-10GF      | ResNet50@WebFace600K | 2d106 & 3d68 | Gender&Age | 326MB |
+| buffalo_m      | RetinaFace-2.5GF     | ResNet50@WebFace600K | 2d106 & 3d68 | Gender&Age | 313MB |
+| buffalo_s      | RetinaFace-500MF     | MBF@WebFace600K | 2d106 & 3d68 | Gender&Age | 159MB |
+| buffalo_sc      | RetinaFace-500MF     | MBF@WebFace600K | - | - | 16MB |
+
+### Recognition accuracy of python library model packs:
+
+| Name      | MR-ALL | African | Caucasian | South Asian | East Asian | LFW    | CFP-FP | AgeDB-30 | IJB-C(E4) |
+| :-------- | ------ | ------- | --------- | ----------- | ---------- | ------ | ------ | -------- | --------- |
+| buffalo_l | 91.25  | 90.29   | 94.70     | 93.16       | 74.96      | 99.83  | 99.33  | 98.23    | 97.25     |
+| buffalo_s	      | 71.87 | 69.45  | 80.45    | 73.39      | 51.03     | 99.70 | 98.00  | 96.58    | 95.02 |
+
+*buffalo_m has the same accuracy with buffalo_l.*
+
+*buffalo_sc has the same accuracy with buffalo_s.*
+
+(Note that almost all ONNX models in our model_zoo can be called by python library.)
+
+##  1. Face Recognition models.
+
+### Definition:
+
+The default training loss is margin based softmax if not specified.
+
+``MFN``: MobileFaceNet
+
+``MS1MV2``: MS1M-ArcFace
+
+``MS1MV3``: MS1M-RetinaFace
+
+``MS1M_MegaFace``: MS1MV2+MegaFace_train
+
+``_pfc``: using Partial FC, with sample-ratio=0.1
+
+``MegaFace``: MegaFace identification test, with gallery=1e6.
+
+``IJBC``: IJBC 1:1 test, under FAR<=1e-4.
+
+``BDrive``: BaiduDrive
+
+``GDrive``: GoogleDrive
+
+### List of models by MXNet and PaddlePaddle:
+
+| Backbone | Dataset | Method  | LFW   | CFP-FP | AgeDB-30 | MegaFace | Link.                                                        |
+| -------- | ------- | ------- | ----- | ------ | -------- | -------- | ------------------------------------------------------------ |
+| R100 (mxnet)     | MS1MV2  | ArcFace | 99.77 | 98.27  | 98.28    | 98.47    | [BDrive](https://pan.baidu.com/s/1wuRTf2YIsKt76TxFufsRNA), [GDrive](https://drive.google.com/file/d/1Hc5zUfBATaXUgcU2haUNa7dcaZSw95h2/view?usp=sharing) |
+| MFN (mxnet)     | MS1MV1  | ArcFace | 99.50 | 88.94  | 95.91    | -        | [BDrive](https://pan.baidu.com/s/1If28BkHde4fiuweJrbicVA), [GDrive](https://drive.google.com/file/d/1RHyJIeYuHduVDDBTn3ffpYEZoXWRamWI/view?usp=sharing) |
+| MFN (paddle)     | MS1MV2  | ArcFace | 99.45 | 93.43  | 96.13    |  -   | [pretrained model](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/MobileFaceNet_128_v1.0_pretrained.tar), [inference model](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar) |
+| iResNet50 (paddle)     | MS1MV2  | ArcFace | 99.73 | 97.43  | 97.88    |  -   | [pretrained model](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/arcface_iresnet50_v1.0_pretrained.tar), [inference model](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/arcface_iresnet50_v1.0_infer.tar) |
+
+
+
+### List of models by various depth IResNet and training datasets:
+
+| Backbone | Dataset   | MR-ALL | African | Caucasian | South Asian | East Asian | Link(onnx)                                                            |
+|----------|-----------|--------|---------|-----------|-------------|------------|-----------------------------------------------------------------------|
+| R100     | Casia     | 42.735 | 39.666  | 53.933    | 47.807      | 21.572     | [GDrive](https://drive.google.com/file/d/1WOrOK-qZO5FcagscCI3td6nnABUPPepD/view?usp=sharing) |
+| R100     | MS1MV2    | 80.725 | 79.117  | 87.176    | 85.501      | 55.807     | [GDrive](https://drive.google.com/file/d/1772DTho9EG047KNUIv2lop2e7EobiCFn/view?usp=sharing) |
+| R18      | MS1MV3    | 68.326 | 62.613  | 75.125    | 70.213      | 43.859     | [GDrive](https://drive.google.com/file/d/1dWZb0SLcdzr-toUzsVZ1zogn9dEIW1Dk/view?usp=sharing) |
+| R34      | MS1MV3    | 77.365 | 71.644  | 83.291    | 80.084      | 53.712     | [GDrive](https://drive.google.com/file/d/1ON6ImX-AigDKAi4pelFPf12vkJVyGFKl/view?usp=sharing) |
+| R50      | MS1MV3    | 80.533 | 75.488  | 86.115    | 84.305      | 57.352     | [GDrive](https://drive.google.com/file/d/1FPldzmZ6jHfaC-R-jLkxvQRP-cLgxjCT/view?usp=sharing) |
+| R100     | MS1MV3    | 84.312 | 81.083  | 89.040    | 88.082      | 62.193     | [GDrive](https://drive.google.com/file/d/1fZOfvfnavFYjzfFoKTh5j1YDcS8KCnio/view?usp=sharing) |
+| R18      | Glint360K | 72.074 | 68.230  | 80.575    | 75.852      | 47.831     | [GDrive](https://drive.google.com/file/d/1Z0eoO1Wqv32K8TdFHKqrlrxv46_W4390/view?usp=sharing) |
+| R34      | Glint360K | 83.015 | 79.907  | 88.620    | 86.815      | 60.604     | [GDrive](https://drive.google.com/file/d/1G1oeLkp_b3JA_z4wGs62RdLpg-u_Ov2Y/view?usp=sharing) |
+| R50      | Glint360K | 87.077 | 85.272  | 91.617    | 90.541      | 66.813     | [GDrive](https://drive.google.com/file/d/1MpRhM76OQ6cTzpr2ZSpHp2_CP19Er4PI/view?usp=sharing) |
+| R100     | Glint360K | 90.659 | 89.488  | 94.285    | 93.434      | 72.528     | [GDrive](https://drive.google.com/file/d/1Gh8C-bwl2B90RDrvKJkXafvZC3q4_H_z/view?usp=sharing) |
+
+
+### List of models by IResNet-50 and different training datasets:
+
+| Dataset           | MR-ALL | African | Caucasian | South Asian | East Asian | LFW   | CFP-FP | AgeDB-30 | IJB-C(E4) | Link(onnx) |
+| :--------         | ------ | ------- | ----      | ------      | --------   | ----- | ------ | -------- | --------- | --- |
+| CISIA	            | 36.794 | 42.550  | 55.825    | 49.618      | 19.611     | 99.450| 95.214 | 94.900   | 87.220    | [GDrive](https://drive.google.com/file/d/1km-cVFvUAPU1UumLLi1fIRasdg6VA-vM/view?usp=sharing) |
+| CISIA_pfc	        | 37.107 | 38.934  | 53.823    | 48.674      | 19.927     | 99.367| 95.429 | 94.600   | 84.970    | [GDrive](https://drive.google.com/file/d/1z8linstTZopL5Yy7NOUgVVtgzGtsu1LM/view?usp=sharing) |
+| VGG2	            | 38.578 | 35.259  | 54.304    | 44.081      | 24.095     | 99.550| 97.410 | 95.080   | 91.220    | [GDrive](https://drive.google.com/file/d/1UwyVIDSNDkHKClBANrWi8qpMU4nXizT6/view?usp=sharing) |
+| VGG2_pfc	        | 40.673 | 36.767  | 60.180    | 49.039      | 24.255     | 99.683| 98.529 | 95.400   | 92.490    | [GDrive](https://drive.google.com/file/d/1uW0EsctVyPklSyXMXF39AniIhSRXCRtp/view?usp=sharing) |
+| GlintAsia	        | 62.663 | 49.531  | 64.829    | 57.984      | 61.743     | 99.583| 93.186 | 95.400   | 91.500    | [GDrive](https://drive.google.com/file/d/1IyXh7m1HMwTZw4B5N1WMPIsN-S9kdS95/view?usp=sharing) |
+| GlintAsia_pfc	    | 63.149 | 50.366  | 65.227    | 57.936      | 61.820     | 99.650| 93.029 | 95.233   | 91.140    | [GDrive](https://drive.google.com/file/d/1CTjalggNucgPkmpFi5ij-NGG1Fy9sL5r/view?usp=sharing) |
+| MS1MV2	        | 77.696 | 74.596  | 84.126    | 82.041      | 51.105     | 99.833| 98.083 | 98.083   | 96.140    | [GDrive](https://drive.google.com/file/d/1rd4kbiXtXBTWE8nP7p4OTv_CAp2FUa1i/view?usp=sharing) |
+| MS1MV2_pfc	    | 77.738 | 74.728  | 84.883    | 82.798      | 52.507     | 99.783| 98.071 | 98.017   | 96.080    | [GDrive](https://drive.google.com/file/d/1ryrXenGQa-EGyk64mVaG136ihNUBmNMW/view?usp=sharing) |
+| MS1M_MegaFace	    | 78.372 | 74.138  | 82.251    | 77.223      | 60.203     | 99.750| 97.557 | 97.400   | 95.350    | [GDrive](https://drive.google.com/file/d/1c2JG0StcTMDrL4ywz3qWTN_9io3lo_ER/view?usp=sharing) |
+| MS1M_MegaFace_pfc | 78.773 | 73.690  | 82.947    | 78.793      | 57.566     | 99.800| 97.870 | 97.733   | 95.400    | [GDrive](https://drive.google.com/file/d/1BnG48LS_HIvYlSbSnP_LzpO3xjx0_rpu/view?usp=sharing) |
+| MS1MV3	        | 82.522 | 77.172  | 87.028    | 86.006      | 60.625     | 99.800| 98.529 | 98.267   | 96.580    | [GDrive](https://drive.google.com/file/d/1Tqorubgcl0qfjbjEM_Y9EDmjG5tCWzbr/view?usp=sharing) |
+| MS1MV3_pfc	    | 81.683 | 78.126  | 87.286    | 85.542      | 58.925     | 99.800| 98.443 | 98.167   | 96.430    | [GDrive](https://drive.google.com/file/d/15jrHCqhEmoSZ93kKL9orVMhbKfNWAhp-/view?usp=sharing) |
+| Glint360k	        | 86.789 | 84.749  | 91.414    | 90.088      | 66.168     | 99.817| 99.143 | 98.450   | 97.130    | [GDrive](https://drive.google.com/file/d/1gnt6P3jaiwfevV4hreWHPu0Mive5VRyP/view?usp=sharing) |
+| Glint360k_pfc	    | 87.077 | 85.272  | 91.616    | 90.541      | 66.813     | 99.817| 99.143 | 98.450   | 97.020    | [GDrive](https://drive.google.com/file/d/164o2Ct42tyJdQjckeMJH2-7KTXolu-EP/view?usp=sharing) |
+| WebFace600K	    | 90.566 | 89.355  | 94.177    | 92.358      | 73.852     | 99.800| 99.200 | 98.100   | 97.120    | [GDrive](https://drive.google.com/file/d/1N0GL-8ehw_bz2eZQWz2b0A5XBdXdxZhg/view?usp=sharing) |
+| WebFace600K_pfc    | 89.951 | 89.301  | 94.016    | 92.381      | 73.007     | 99.817| 99.143 | 98.117   | 97.010    | [GDrive](https://drive.google.com/file/d/11TASXssTnwLY1ZqKlRjsJiV-1nWu9pDY/view?usp=sharing) |
+| Average	        | 69.247 | 65.908  | 77.121    | 72.819      | 52.014     | 99.706| 97.374 | 96.962   | 93.925    |  |
+| Average_pfc	    | 69.519 | 65.898  | 77.497    | 73.213      | 51.853     | 99.715| 97.457 | 96.965   | 93.818    |  |
+
+### List of models by MobileFaceNet and different training datasets:
+
+**``FLOPS``:** 450M FLOPs
+
+**``Model-Size``:** 13MB
+
+| Dataset           | MR-ALL | African | Caucasian | South Asian | East Asian | LFW   | CFP-FP | AgeDB-30 | IJB-C(E4) | Link(onnx) |
+| :--------         | ------ | ------- | ----      | ------      | --------   | ----- | ------ | -------- | --------- | --- |
+| WebFace600K	      | 71.865 | 69.449  | 80.454    | 73.394      | 51.026     | 99.70 | 98.00  | 96.58    | 95.02     | - |
+
+
+## 2. Face Detection models.
+
+### 2.1 RetinaFace
+
+In RetinaFace, mAP was evaluated with multi-scale testing.
+
+``m025``: means MobileNet-0.25
+
+| Impelmentation           | Easy-Set | Medium-Set | Hard-Set | Link                                                         |
+| ------------------------ | -------- | ---------- | -------- | ------------------------------------------------------------ |
+| RetinaFace-R50           | 96.5     | 95.6       | 90.4     | [BDrive](https://pan.baidu.com/s/1C6nKq122gJxRhb37vK0_LQ), [GDrive](https://drive.google.com/file/d/1wm-6K688HQEx_H90UdAIuKv-NAsKBu85/view?usp=sharing) |
+| RetinaFace-m025(yangfly) | -        | -          | 82.5     | [BDrive](https://pan.baidu.com/s/1P1ypO7VYUbNAezdvLm2m9w)(nzof), [GDrive](https://drive.google.com/drive/folders/1OTXuAUdkLVaf78iz63D1uqGLZi4LbPeL?usp=sharing) |
+| BlazeFace-FPN-SSH	(paddle)           | 91.9     | 89.8       | 81.7%     | [pretrained model](https://paddledet.bj.bcebos.com/models/blazeface_fpn_ssh_1000e.pdparams), [inference model](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/blazeface_fpn_ssh_1000e_v1.0_infer.tar) |
+
+### 2.2 SCRFD
+
+In SCRFD, mAP was evaluated with single scale testing, VGA resolution.
+
+``2.5G``: means the model cost ``2.5G`` FLOPs while the input image is in VGA(640x480) resolution.
+
+``_KPS``: means this model can detect five facial keypoints.
+
+|      Name      | Easy  | Medium | Hard  | FLOPs | Params(M) | Infer(ms) | Link(pth)                                                    |
+| :------------: | ----- | ------ | ----- | ----- | --------- | --------- | ------------------------------------------------------------ |
+|   SCRFD_500M   | 90.57 | 88.12  | 68.51 | 500M  | 0.57      | 3.6       | [GDrive](https://drive.google.com/file/d/1OX0i_vWDp1Fp-ZynOUMZo-q1vB5g1pTN/view?usp=sharing) |
+|    SCRFD_1G    | 92.38 | 90.57  | 74.80 | 1G    | 0.64      | 4.1       | [GDrive](https://drive.google.com/file/d/1acd5wKjWnl1zMgS5YJBtCh13aWtw9dej/view?usp=sharing) |
+|   SCRFD_2.5G   | 93.78 | 92.16  | 77.87 | 2.5G  | 0.67      | 4.2       | [GDrive](https://drive.google.com/file/d/1wgg8GY2vyP3uUTaAKT0_MSpAPIhmDsCQ/view?usp=sharing) |
+|   SCRFD_10G    | 95.16 | 93.87  | 83.05 | 10G   | 3.86      | 4.9       | [GDrive](https://drive.google.com/file/d/1kUYa0s1XxLW37ZFRGeIfKNr9L_4ScpOg/view?usp=sharing) |
+|   SCRFD_34G    | 96.06 | 94.92  | 85.29 | 34G   | 9.80      | 11.7      | [GDrive](https://drive.google.com/file/d/1w9QOPilC9EhU0JgiVJoX0PLvfNSlm1XE/view?usp=sharing) |
+| SCRFD_500M_KPS | 90.97 | 88.44  | 69.49 | 500M  | 0.57      | 3.6       | [GDrive](https://drive.google.com/file/d/1TXvKmfLTTxtk7tMd2fEf-iWtAljlWDud/view?usp=sharing) |
+| SCRFD_2.5G_KPS | 93.80 | 92.02  | 77.13 | 2.5G  | 0.82      | 4.3       | [GDrive](https://drive.google.com/file/d/1KtOB9TocdPG9sk_S_-1QVG21y7OoLIIf/view?usp=sharing) |
+| SCRFD_10G_KPS  | 95.40 | 94.01  | 82.80 | 10G   | 4.23      | 5.0       | [GDrive](https://drive.google.com/file/d/1-2uy0tgkenw6ZLxfKV1qVhmkb5Ep_5yx/view?usp=sharing) |
+
+
+
+## 3. Face Alignment models.
+
+### 2.1 2D Face Alignment
+
+| Impelmentation        | Points | Backbone      | Params(M) | Link(onnx)                                                   |
+| --------------------- | ------ | ------------- | --------- | ------------------------------------------------------------ |
+| Coordinate-regression | 106    | MobileNet-0.5 | 1.2       | [GDrive](https://drive.google.com/file/d/1M5685m-bKnMCt0u2myJoEK5gUY3TDt_1/view?usp=sharing) |
+
+### 2.2 3D Face Alignment
+
+| Impelmentation | Points | Backbone  | Params(M) | Link(onnx)                                                   |
+| -------------- | ------ | --------- | --------- | ------------------------------------------------------------ |
+| -              | 68     | ResNet-50 | 34.2      | [GDrive](https://drive.google.com/file/d/1aJe5Rzoqrtf_a9U84E-V1b0rUi8-QbCI/view?usp=sharing) |
+
+### 2.3 Dense Face Alignment
+
+## 4. Face Attribute models.
+
+### 4.1 Gender&Age 
+
+| Training-Set | Backbone       | Params(M) | Link(onnx)                                                   |
+| ------------ | -------------- | --------- | ------------------------------------------------------------ |
+| CelebA       | MobileNet-0.25 | 0.3       | [GDrive](https://drive.google.com/file/d/1Mm3TeUuaZOwmEMp0nGOddvgXCjpRodPU/view?usp=sharing) |
+
+
+### 4.2 Expression
diff --git a/models/README.md b/models/README.md
deleted file mode 100644
index db5644a87..000000000
--- a/models/README.md
+++ /dev/null
@@ -1 +0,0 @@
-Put pretrained models here
diff --git a/parsing/dml_csr/README.md b/parsing/dml_csr/README.md
new file mode 100644
index 000000000..4b2ffceef
--- /dev/null
+++ b/parsing/dml_csr/README.md
@@ -0,0 +1,61 @@
+# Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing.
+
+The official repository of *[Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing. (CVPR 2022)](https://arxiv.org/abs/2203.14448)*. 
+
+## Installation
+
+Our model is based on Pytorch 1.7.1 with Python 3.6.2.
+
+```sh
+pip install -r requirements.txt
+```
+
+## Data
+You can download original datasets:
+- **Helen** : [https://www.sifeiliu.net/face-parsing](https://www.sifeiliu.net/face-parsing)
+- **LaPa** : [https://github.com/JDAI-CV/lapa-dataset](https://github.com/JDAI-CV/lapa-dataset)
+- **CelebAMask-HQ** : [https://github.com/switchablenorms/CelebAMask-HQ](https://github.com/switchablenorms/CelebAMask-HQ)
+
+and put them in ./dataset folder as below
+```
+dataset/
+    images/
+    labels/
+    edges/
+    train_list.txt
+    test_list.txt
+        each line: 'images/100032540_1.jpg labels/100032540_1.png'
+```
+Besides, we provide the edge genearation code in the *generate_edge.py*.
+
+## Usage
+
+If you need imagenet pretrained resent-101, please download from [baidu drive]() or [Google drive](https://drive.google.com/open?id=1rzLU-wK6rEorCNJfwrmIu5hY2wRMyKTK), and put it into snapshot folder.
+
+For dstributed(multi-gpu) training. Inplace-abn requires pytorch distributed data parallel.
+```
+GPU=4,5,6,7
+Node=4
+dataset=./datasets/CelebAMask-HQ/
+snapshot=./work_dirs/
+CUDA_VISIBLE_DEVICES="$GPU" python -m torch.distributed.launch --nproc_per_node="$Node"  --master_port=295002 train.py --data-dir "$dataset"  --random-mirror --random-scale \
+--gpu "$GPU" --batch-size 7 --input-size 473,473 --snapshot-dir "$snapshot" --num-classes 19 --epochs 200 --schp-start 150
+```
+
+For testing [pretrained models](https://drive.google.com/file/d/1-PjUts1AMzXNyvw3VaJQmg43GJbfEpEQ/view?usp=sharing)
+```
+python test.py --data-dir "$dataset" --out-dir "$out_dir" --restore-from "$snapshot" --gpu "$GPU" --batch-size 7 --input-size 473,473 --dataset test --num-classes 19
+```
+
+## Reference
+
+If you consider use our code, please cite our paper:
+
+```
+@inproceedings{Zheng2022DecoupledML,
+  title={Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing},
+  author={Qi Zheng and Jiankang Deng and Zheng Zhu and Ying Li and Stefanos Zafeiriou},
+  booktitle={Computer Vision and Pattern Recognition},
+  year={2022}
+}
+```
diff --git a/parsing/dml_csr/dataset/datasets.py b/parsing/dml_csr/dataset/datasets.py
new file mode 100644
index 000000000..9f8e211a9
--- /dev/null
+++ b/parsing/dml_csr/dataset/datasets.py
@@ -0,0 +1,180 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   datasets.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import cv2
+import numpy as np
+import os.path as osp
+import pickle
+import random
+
+import torch
+import torchvision.transforms.functional as TF
+
+from glob import glob
+from typing import Tuple
+from utils import transforms
+
+
+class FaceDataSet(torch.utils.data.Dataset):
+    """Face data set for model training and validating
+
+    Examples:
+
+        ./CelebAMask
+            |---test
+            |---train
+                |---images
+                    |---0.jpg
+                    |---1.jpg
+                |---labels
+                    |---0.png
+                    |---1.png
+                |---edges
+                    |---0.png
+                    |---1.png
+            |---valid
+            |---label_names.txt
+            |---test_list.txt
+            |---train_list.txt
+                |---images/0.jpg labels/0.png
+                |---images/1.jpg labels/1.png
+            |---valid_list.txt
+
+    Args:
+      root: A string, training/validating dataset path, e.g. "./CelebAMask"
+      dataset: A string, one of `"train"`, `"test"`, `"valid"`.
+      crop_size: A list of two intergers.
+      scale_factor: A float number.
+      rotation_factor: An integer number.
+      ignore_label: An integer number, default is 255.
+      transformer: A function of torchvision.transforms.Compose([])
+    """
+    def __init__(self, 
+                 root: str, 
+                 dataset: str, 
+                 crop_size: list=[473, 473], 
+                 scale_factor: float=0.25,
+                 rotation_factor: int=30, 
+                 ignore_label: int =255, 
+                 transform=None) -> None:
+
+        self.root = root
+        self.dataset = dataset
+        self.crop_size = np.asarray(crop_size)
+        self.scale_factor = scale_factor
+        self.rotation_factor = rotation_factor
+        self.ignore_label = ignore_label
+        self.transform = transform
+
+        self.flip_prob = 0.5
+        self.flip_pairs = [[4, 5], [6, 7]]
+        self.aspect_ratio = crop_size[1] * 1.0 / crop_size[0]
+
+        self.file_list_name = osp.join(root, dataset + '_list.txt')
+        self.im_list = [line.split()[0][7:-4] for line in open(self.file_list_name).readlines()]
+        self.number_samples = len(self.im_list)
+
+
+    def __len__(self) -> int:
+        return self.number_samples 
+
+    def _box2cs(self, box: list) -> tuple:
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x: float, y: float, w: float, h: float) -> tuple:
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w * 1.0, h * 1.0], dtype=np.float32)
+
+        return center, scale
+
+    def __getitem__(self, index: int) -> tuple:
+        # Load training image
+        im_name = self.im_list[index]
+        im_path = osp.join(self.root, self.dataset, 'images', im_name + '.jpg')
+        im = cv2.imread(im_path, cv2.IMREAD_COLOR)
+        h, w, _ = im.shape
+        parsing_anno = np.zeros((h, w), dtype=np.long)
+        # Get center and scale
+        center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+
+        if self.dataset not in ['test', 'valid']:
+            edge_path = osp.join(self.root, self.dataset, 'edges', im_name + '.png')
+            edge = cv2.imread(edge_path, cv2.IMREAD_GRAYSCALE)
+            parsing_anno_path = osp.join(self.root, self.dataset, 'labels', im_name + '.png')
+            parsing_anno = cv2.imread(parsing_anno_path, cv2.IMREAD_GRAYSCALE)
+
+            if self.dataset in 'train':
+
+                sf = self.scale_factor
+                rf = self.rotation_factor
+                s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
+                r = np.clip(np.random.randn() * rf, -rf * 2, rf * 2) \
+                    if random.random() <= 0.6 else 0
+
+        trans = transforms.get_affine_transform(center, s, r, self.crop_size)
+        image = cv2.warpAffine(
+            im,
+            trans,
+            (int(self.crop_size[1]), int(self.crop_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+        
+        if self.dataset not in ['test', 'valid']:
+            edge = cv2.warpAffine(
+                edge,
+                trans,
+                (int(self.crop_size[1]), int(self.crop_size[0])),
+                flags=cv2.INTER_LINEAR,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=(0, 0, 0))
+
+        if self.transform:
+            image = self.transform(image)
+
+        meta = {
+            'name': im_name,
+            'center': center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r,
+            'origin': image
+        }
+
+        if self.dataset not in 'train':
+            return image, meta
+        else:
+
+            label_parsing = cv2.warpAffine(
+                parsing_anno,
+                trans,
+                (int(self.crop_size[1]), int(self.crop_size[0])),
+                flags=cv2.INTER_NEAREST,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=(255))
+
+            label_parsing = torch.from_numpy(label_parsing)
+
+            return image, label_parsing, edge, meta
+
diff --git a/parsing/dml_csr/loss/consistency_loss.py b/parsing/dml_csr/loss/consistency_loss.py
new file mode 100644
index 000000000..2b6ac5992
--- /dev/null
+++ b/parsing/dml_csr/loss/consistency_loss.py
@@ -0,0 +1,75 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   consistency_loss.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+def generate_edge_tensor(label, edge_width=3):
+    label = label.type(torch.cuda.FloatTensor)
+    if len(label.shape) == 2:
+        label = label.unsqueeze(0)
+    n, h, w = label.shape
+    edge = torch.zeros(label.shape, dtype=torch.float).cuda()
+    # right
+    edge_right = edge[:, 1:h, :]
+    edge_right[(label[:, 1:h, :] != label[:, :h - 1, :]) & (label[:, 1:h, :] != 255)
+               & (label[:, :h - 1, :] != 255)] = 1
+
+    # up
+    edge_up = edge[:, :, :w - 1]
+    edge_up[(label[:, :, :w - 1] != label[:, :, 1:w])
+            & (label[:, :, :w - 1] != 255)
+            & (label[:, :, 1:w] != 255)] = 1
+
+    # upright
+    edge_upright = edge[:, :h - 1, :w - 1]
+    edge_upright[(label[:, :h - 1, :w - 1] != label[:, 1:h, 1:w])
+                 & (label[:, :h - 1, :w - 1] != 255)
+                 & (label[:, 1:h, 1:w] != 255)] = 1
+
+    # bottomright
+    edge_bottomright = edge[:, :h - 1, 1:w]
+    edge_bottomright[(label[:, :h - 1, 1:w] != label[:, 1:h, :w - 1])
+                     & (label[:, :h - 1, 1:w] != 255)
+                     & (label[:, 1:h, :w - 1] != 255)] = 1
+
+    kernel = torch.ones((1, 1, edge_width, edge_width), dtype=torch.float).cuda()
+    with torch.no_grad():
+        edge = edge.unsqueeze(1)
+        edge = F.conv2d(edge, kernel, stride=1, padding=1)
+    edge[edge!=0] = 1
+    edge = edge.squeeze()
+    return edge
+
+
+class ConsistencyLoss(nn.Module):
+    def __init__(self, ignore_index=255):
+        super(ConsistencyLoss, self).__init__()
+        self.ignore_index=ignore_index
+
+    def forward(self, parsing, edge, label):
+        parsing_pre = torch.argmax(parsing, dim=1)
+        parsing_pre[label==self.ignore_index]=self.ignore_index
+        generated_edge = generate_edge_tensor(parsing_pre)
+        edge_pre = torch.argmax(edge, dim=1)
+        v_generate_edge = generated_edge[label!=255]
+        v_edge_pre = edge_pre[label!=255]
+        one = torch.ones_like(v_edge_pre)
+        v_edge_pre = torch.where(v_edge_pre > 0, one, v_edge_pre)
+        v_edge_pre = v_edge_pre.type(torch.cuda.FloatTensor)
+        positive_union = (v_generate_edge==1)&(v_edge_pre==1) # only the positive values count
+        return F.smooth_l1_loss(v_generate_edge[positive_union].squeeze(0), v_edge_pre[positive_union].squeeze(0))
diff --git a/parsing/dml_csr/loss/criterion.py b/parsing/dml_csr/loss/criterion.py
new file mode 100644
index 000000000..9fa1bdc65
--- /dev/null
+++ b/parsing/dml_csr/loss/criterion.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   criterion.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch.nn as nn
+import torch
+import numpy as np
+
+from torch.nn import functional as F
+from torch.nn.modules.loss import _Loss
+from .lovasz_softmax import LovaszSoftmax
+from .kl_loss import KLDivergenceLoss
+from .consistency_loss import ConsistencyLoss
+
+
+class Criterion(nn.Module):
+    """DML_CSR loss for face parsing.
+    
+    Put more focus on facial components like eyes, eyebrow, nose and mouth
+    """
+    def __init__(self, loss_weight=[1.0, 1.0, 1.0, 1.0, 1.0], ignore_index=255, lambda_1=1, lambda_2=1, lambda_3=1, num_classes=11):
+        super(Criterion, self).__init__()
+        self.ignore_index = ignore_index   
+        self.loss_weight = loss_weight
+        self.criterion = torch.nn.CrossEntropyLoss(ignore_index=ignore_index) 
+        self.criterion_weight = torch.nn.CrossEntropyLoss(reduction='none', ignore_index=ignore_index)
+        # self.sed = WeightedCrossEntropyWithLogits()
+        self.lovasz = LovaszSoftmax(ignore_index=ignore_index)
+        self.kldiv = KLDivergenceLoss(ignore_index=ignore_index)
+        self.reg = ConsistencyLoss(ignore_index=ignore_index)
+        self.lamda_1 = lambda_1
+        self.lamda_2 = lambda_2
+        self.lamda_3 = lambda_3
+        self.num_classes = num_classes
+          
+    def forward(self, preds, target, cycle_n=None):
+        h, w = target[0].size(1), target[0].size(2)
+        
+        # binary edge
+        input_binary_labels = target[1].data.cpu().numpy().astype(np.int64)
+        binary_pos_num = np.sum(input_binary_labels==1).astype(np.float)
+        binary_neg_num = np.sum(input_binary_labels==0).astype(np.float)
+
+        binary_weight_pos = binary_neg_num/(binary_pos_num + binary_neg_num)
+        binary_weight_neg = binary_pos_num/(binary_pos_num + binary_neg_num)
+        binary_weights = (binary_weight_neg, binary_weight_pos)  
+        binary_weights = torch.from_numpy(np.array(binary_weights)).float().cuda()
+
+        # print('target', target[0].size(), target[1].size())
+        binary_edge_p_num = target[1].cpu().numpy().reshape(target[1].size(0),-1).sum(axis=1)
+        # print('edge_p_num_1', edge_p_num.shape)
+        binary_edge_p_num = np.tile(binary_edge_p_num, [h, w, 1]).transpose(2,1,0)
+        # print('edge_p_num_2', edge_p_num.shape)
+        binary_edge_p_num = torch.from_numpy(binary_edge_p_num).cuda().float()
+
+        # semantic edge
+        input_semantic_labels = target[2].data.cpu().numpy().astype(np.int64)
+        semantic_weights = []
+        semantic_pos_num = np.sum(input_semantic_labels>0).astype(np.float)
+        semantic_neg_num = np.sum(input_semantic_labels==0).astype(np.float)
+
+        for lbl in range(self.num_classes):
+            lbl_num = np.sum(input_semantic_labels==lbl).astype(np.float)
+            weight_lbl = lbl_num/(semantic_pos_num + semantic_neg_num)
+            semantic_weights.append(weight_lbl)
+        semantic_weights = torch.from_numpy(np.array(semantic_weights)).float().cuda()
+
+        # print('target', target[0].size(), target[1].size())
+        semantic_edge_p_num = np.count_nonzero(target[2].cpu().numpy().reshape(target[2].size(0),-1), axis=1)
+        # print('edge_p_num_1', edge_p_num.shape)
+        semantic_edge_p_num = np.tile(semantic_edge_p_num, [h, w, 1]).transpose(2,1,0)
+        # print('edge_p_num_2', edge_p_num.shape)
+        semantic_edge_p_num = torch.from_numpy(semantic_edge_p_num).cuda().float()
+        
+        loss_binary_edge = 0; loss_semantic_edge = 0; loss_parse = 0; loss_att_parse = 0; loss_att_binary_edge = 0; loss_att_semantic_edge = 0; loss_consistency = 0
+        # print(preds[1].size(), target[1].size(), weights, len(preds))
+
+        # loss for parsing
+        scale_parse = F.interpolate(input=preds[0], size=(h, w), mode='bilinear', align_corners=True) # parsing
+        loss_parse += 0.5 * self.lamda_1 * self.lovasz(scale_parse, target[0])
+        
+        if target[3] is None:
+            loss_parse += 0.5 * self.lamda_1 * self.criterion(scale_parse, target[0])
+        else:
+            soft_scale_parse = F.interpolate(input=target[2], size=(h, w), mode='bilinear', align_corners=True)
+            soft_scale_parse = moving_average(soft_scale_parse, to_one_hot(target[0], num_cls=self.num_classes),
+                                                1.0 / (cycle_n + 1.0))
+            loss_parse += 0.5 * self.lamda_1 * self.kldiv(scale_parse, soft_scale_parse, target[0])
+
+        # loss for binary edge
+        scale_binary_edge = F.interpolate(input=preds[1], size=(h, w), mode='bilinear', align_corners=True)  # edge 
+        
+        if target[4] is None:
+            loss_binary_edge = self.lamda_2 * F.cross_entropy(scale_binary_edge, target[1], binary_weights)
+        else:
+            soft_scale_binary_edge = F.interpolate(input=target[4], size=(h, w), mode='bilinear', align_corners=True)
+            soft_scale_binary_edge = moving_average(soft_scale_binary_edge, to_one_hot(target[1], num_cls=2),
+                                                1.0 / (cycle_n + 1.0))
+            loss_binary_edge += self.lamda_2 * self.kldiv(scale_binary_edge, soft_scale_binary_edge, target[0])
+
+        # loss for semantic edge
+        scale_semantic_edge = F.interpolate(input=preds[2], size=(h, w), mode='bilinear', align_corners=True)  # edge 
+        
+        if target[5] is None:
+            loss_semantic_edge = self.lamda_3 * F.cross_entropy(scale_semantic_edge, target[2], semantic_weights)
+            # loss_edge = self.lamda_2 * self.sed(scale_edge, target[1])
+        else:
+            soft_scale_semantic_edge = F.interpolate(input=target[5], size=(h, w), mode='bilinear', align_corners=True)
+            soft_scale_semantic_edge = moving_average(soft_scale_semantic_edge, to_one_hot(target[2], num_cls=self.num_classes),
+                                                1.0 / (cycle_n + 1.0))
+            loss_semantic_edge += self.lamda_3 * self.kldiv(scale_semantic_edge, soft_scale_semantic_edge, target[0])
+    
+        # binary edge attention loss
+        loss_att_binary_edge_ = self.criterion_weight(scale_parse, target[0]) * target[1].float()
+        loss_att_binary_edge_ = loss_att_binary_edge_ / binary_edge_p_num  # only compute the edge pixels
+        loss_att_binary_edge_ = torch.sum(loss_att_binary_edge_) / target[1].size(0)  # mean for batchsize      
+                
+        loss_parse += loss_parse
+        loss_att_binary_edge += loss_att_binary_edge
+        loss_att_binary_edge += loss_att_binary_edge_
+
+        # semantic edge attention loss
+        loss_att_semantic_edge_ = self.criterion_weight(scale_parse, target[0]) * target[2].float()
+        loss_att_semantic_edge_ = loss_att_semantic_edge_ / semantic_edge_p_num  # only compute the edge pixels
+        loss_att_semantic_edge_ = torch.sum(loss_att_semantic_edge_) / target[2].size(0)  # mean for batchsize      
+                
+        loss_parse += loss_parse
+        loss_semantic_edge += loss_semantic_edge
+        loss_att_semantic_edge += loss_att_semantic_edge_
+        # loss_consistency += loss_consistency
+        
+        # print('loss_parse: {}\t loss_edge: {}\t loss_att_edge: {}\t loss_semantic_edge: {}'.format(loss_parse,loss_edge,loss_att_edge, loss_consistency))
+        return self.loss_weight[0]*loss_parse + self.loss_weight[1]*loss_binary_edge + self.loss_weight[2]*loss_semantic_edge \
+            + self.loss_weight[3]*loss_att_binary_edge + self.loss_weight[4]*loss_att_semantic_edge
+
+
+def moving_average(target1, target2, alpha=1.0):
+    target = 0
+    target += (1.0 - alpha) * target1
+    target += target2 * alpha
+    return target
+
+
+def to_one_hot(tensor, num_cls, dim=1, ignore_index=255):
+    b, h, w = tensor.shape
+    tensor[tensor == ignore_index] = 0
+    onehot_tensor = torch.zeros(b, num_cls, h, w).cuda()
+    onehot_tensor.scatter_(dim, tensor.unsqueeze(dim), 1)
+    return onehot_tensor
diff --git a/parsing/dml_csr/loss/kl_loss.py b/parsing/dml_csr/loss/kl_loss.py
new file mode 100644
index 000000000..eb63cb1bb
--- /dev/null
+++ b/parsing/dml_csr/loss/kl_loss.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   kl_loss.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+def flatten_probas(input, target, labels, ignore=255):
+    """
+    Flattens predictions in the batch.
+    """
+    B, C, H, W = input.size()
+    input = input.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
+    target = target.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
+    labels = labels.view(-1)
+    if ignore is None:
+        return input, target
+    valid = (labels != ignore)
+    vinput = input[torch.nonzero(valid, as_tuple=False).squeeze()]
+    vtarget = target[torch.nonzero(valid, as_tuple=False).squeeze()]
+    return vinput, vtarget
+
+
+class KLDivergenceLoss(nn.Module):
+    def __init__(self, ignore_index=255, T=1):
+        super(KLDivergenceLoss, self).__init__()
+        self.ignore_index=ignore_index
+        self.T = T
+
+    def forward(self, input, target, label):
+        log_input_prob = F.log_softmax(input / self.T, dim=1)
+        target_porb = F.softmax(target / self.T, dim=1)
+        loss = F.kl_div(*flatten_probas(log_input_prob, target_porb, label, ignore=self.ignore_index), reduction='batchmean')
+        return self.T*self.T*loss # balanced
diff --git a/parsing/dml_csr/loss/lovasz_softmax.py b/parsing/dml_csr/loss/lovasz_softmax.py
new file mode 100644
index 000000000..739165072
--- /dev/null
+++ b/parsing/dml_csr/loss/lovasz_softmax.py
@@ -0,0 +1,280 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   lovaz_softmax.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from torch import nn
+from torch.autograd import Variable
+
+try:
+    from itertools import ifilterfalse
+except ImportError:  # py3k
+    from itertools import filterfalse as ifilterfalse
+
+
+def lovasz_grad(gt_sorted):
+    """
+    Computes gradient of the Lovasz extension w.r.t sorted errors
+    See Alg. 1 in paper
+    """
+    p = len(gt_sorted)
+    gts = gt_sorted.sum()
+    intersection = gts - gt_sorted.float().cumsum(0)
+    union = gts + (1 - gt_sorted).float().cumsum(0)
+    jaccard = 1. - intersection / union
+    if p > 1:  # cover 1-pixel case
+        jaccard[1:p] = jaccard[1:p] - jaccard[0:-1]
+    return jaccard
+
+
+def iou_binary(preds, labels, EMPTY=1., ignore=None, per_image=True):
+    """
+    IoU for foreground class
+    binary: 1 foreground, 0 background
+    """
+    if not per_image:
+        preds, labels = (preds,), (labels,)
+    ious = []
+    for pred, label in zip(preds, labels):
+        intersection = ((label == 1) & (pred == 1)).sum()
+        union = ((label == 1) | ((pred == 1) & (label != ignore))).sum()
+        if not union:
+            iou = EMPTY
+        else:
+            iou = float(intersection) / float(union)
+        ious.append(iou)
+    iou = mean(ious)  # mean accross images if per_image
+    return 100 * iou
+
+
+def iou(preds, labels, C, EMPTY=1., ignore=None, per_image=False):
+    """
+    Array of IoU for each (non ignored) class
+    """
+    if not per_image:
+        preds, labels = (preds,), (labels,)
+    ious = []
+    for pred, label in zip(preds, labels):
+        iou = []
+        for i in range(C):
+            if i != ignore:  # The ignored label is sometimes among predicted classes (ENet - CityScapes)
+                intersection = ((label == i) & (pred == i)).sum()
+                union = ((label == i) | ((pred == i) & (label != ignore))).sum()
+                if not union:
+                    iou.append(EMPTY)
+                else:
+                    iou.append(float(intersection) / float(union))
+        ious.append(iou)
+    ious = [mean(iou) for iou in zip(*ious)]  # mean accross images if per_image
+    return 100 * np.array(ious)
+
+
+# --------------------------- BINARY LOSSES ---------------------------
+
+
+def lovasz_hinge(logits, labels, per_image=True, ignore=None):
+    """
+    Binary Lovasz hinge loss
+      logits: [B, H, W] Variable, logits at each pixel (between -\infty and +\infty)
+      labels: [B, H, W] Tensor, binary ground truth masks (0 or 1)
+      per_image: compute the loss per image instead of per batch
+      ignore: void class id
+    """
+    if per_image:
+        loss = mean(lovasz_hinge_flat(*flatten_binary_scores(log.unsqueeze(0), lab.unsqueeze(0), ignore))
+                    for log, lab in zip(logits, labels))
+    else:
+        loss = lovasz_hinge_flat(*flatten_binary_scores(logits, labels, ignore))
+    return loss
+
+
+def lovasz_hinge_flat(logits, labels):
+    """
+    Binary Lovasz hinge loss
+      logits: [P] Variable, logits at each prediction (between -\infty and +\infty)
+      labels: [P] Tensor, binary ground truth labels (0 or 1)
+      ignore: label to ignore
+    """
+    if len(labels) == 0:
+        # only void pixels, the gradients should be 0
+        return logits.sum() * 0.
+    signs = 2. * labels.float() - 1.
+    errors = (1. - logits * Variable(signs))
+    errors_sorted, perm = torch.sort(errors, dim=0, descending=True)
+    perm = perm.data
+    gt_sorted = labels[perm]
+    grad = lovasz_grad(gt_sorted)
+    loss = torch.dot(F.relu(errors_sorted), Variable(grad))
+    return loss
+
+
+def flatten_binary_scores(scores, labels, ignore=None):
+    """
+    Flattens predictions in the batch (binary case)
+    Remove labels equal to 'ignore'
+    """
+    scores = scores.view(-1)
+    labels = labels.view(-1)
+    if ignore is None:
+        return scores, labels
+    valid = (labels != ignore)
+    vscores = scores[valid]
+    vlabels = labels[valid]
+    return vscores, vlabels
+
+
+class StableBCELoss(torch.nn.modules.Module):
+    def __init__(self):
+        super(StableBCELoss, self).__init__()
+
+    def forward(self, input, target):
+        neg_abs = - input.abs()
+        loss = input.clamp(min=0) - input * target + (1 + neg_abs.exp()).log()
+        return loss.mean()
+
+
+def binary_xloss(logits, labels, ignore=None):
+    """
+    Binary Cross entropy loss
+      logits: [B, H, W] Variable, logits at each pixel (between -\infty and +\infty)
+      labels: [B, H, W] Tensor, binary ground truth masks (0 or 1)
+      ignore: void class id
+    """
+    logits, labels = flatten_binary_scores(logits, labels, ignore)
+    loss = StableBCELoss()(logits, Variable(labels.float()))
+    return loss
+
+
+# --------------------------- MULTICLASS LOSSES ---------------------------
+
+
+def lovasz_softmax(probas, labels, classes='present', per_image=False, ignore=255, weighted=None):
+    """
+    Multi-class Lovasz-Softmax loss
+      probas: [B, C, H, W] Variable, class probabilities at each prediction (between 0 and 1).
+              Interpreted as binary (sigmoid) output with outputs of size [B, H, W].
+      labels: [B, H, W] Tensor, ground truth labels (between 0 and C - 1)
+      classes: 'all' for all, 'present' for classes present in labels, or a list of classes to average.
+      per_image: compute the loss per image instead of per batch
+      ignore: void class labels
+    """
+    if per_image:
+        loss = mean(lovasz_softmax_flat(*flatten_probas(prob.unsqueeze(0), lab.unsqueeze(0), ignore), classes=classes, weighted=weighted)
+                    for prob, lab in zip(probas, labels))
+    else:
+        loss = lovasz_softmax_flat(*flatten_probas(probas, labels, ignore), classes=classes, weighted=weighted )
+    return loss
+
+
+def lovasz_softmax_flat(probas, labels, classes='present', weighted=None):
+    """
+    Multi-class Lovasz-Softmax loss
+      probas: [P, C] Variable, class probabilities at each prediction (between 0 and 1)
+      labels: [P] Tensor, ground truth labels (between 0 and C - 1)
+      classes: 'all' for all, 'present' for classes present in labels, or a list of classes to average.
+    """
+    if probas.numel() == 0:
+        # only void pixels, the gradients should be 0
+        return probas * 0.
+    C = probas.size(1)
+    losses = []
+    class_to_sum = list(range(C)) if classes in ['all', 'present'] else classes
+    for c in class_to_sum:
+        fg = (labels == c).float()  # foreground for class c
+        if (classes is 'present' and fg.sum() == 0):
+            continue
+        if C == 1:
+            if len(classes) > 1:
+                raise ValueError('Sigmoid output possible only with 1 class')
+            class_pred = probas[:, 0]
+        else:
+            class_pred = probas[:, c]
+        errors = (Variable(fg) - class_pred).abs()
+        errors_sorted, perm = torch.sort(errors, 0, descending=True)
+        perm = perm.data
+        fg_sorted = fg[perm]
+        if weighted is not None:
+            losses.append(weighted[c]*torch.dot(errors_sorted, Variable(lovasz_grad(fg_sorted))))
+        else:
+            losses.append(torch.dot(errors_sorted, Variable(lovasz_grad(fg_sorted))))
+    return mean(losses)
+
+
+def flatten_probas(probas, labels, ignore=None):
+    """
+    Flattens predictions in the batch
+    """
+    if probas.dim() == 3:
+        # assumes output of a sigmoid layer
+        B, H, W = probas.size()
+        probas = probas.view(B, 1, H, W)
+    B, C, H, W = probas.size()
+    probas = probas.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
+    labels = labels.view(-1)
+    if ignore is None:
+        return probas, labels
+    valid = (labels != ignore)
+    # vprobas = probas[valid.nonzero().squeeze()]
+    vprobas = probas[torch.nonzero(valid, as_tuple =False).squeeze()]
+    vlabels = labels[valid]
+    return vprobas, vlabels
+
+
+def xloss(logits, labels, ignore=None):
+    """
+    Cross entropy loss
+    """
+    return F.cross_entropy(logits, Variable(labels), ignore_index=255)
+
+
+# --------------------------- HELPER FUNCTIONS ---------------------------
+def isnan(x):
+    return x != x
+
+
+def mean(l, ignore_nan=False, empty=0):
+    """
+    nanmean compatible with generators.
+    """
+    l = iter(l)
+    if ignore_nan:
+        l = ifilterfalse(isnan, l)
+    try:
+        n = 1
+        acc = next(l)
+    except StopIteration:
+        if empty == 'raise':
+            raise ValueError('Empty mean')
+        return empty
+    for n, v in enumerate(l, 2):
+        acc += v
+    if n == 1:
+        return acc
+    return acc / n
+
+# --------------------------- Class ---------------------------
+class LovaszSoftmax(nn.Module):
+    def __init__(self, per_image=False, ignore_index=255, weighted=None):
+        super(LovaszSoftmax, self).__init__()
+        self.lovasz_softmax = lovasz_softmax
+        self.per_image = per_image
+        self.ignore_index=ignore_index
+        self.weighted = weighted
+
+    def forward(self, pred, label):
+        pred = F.softmax(pred, dim=1)
+        return self.lovasz_softmax(pred, label, per_image=self.per_image, ignore=self.ignore_index, weighted=self.weighted)
diff --git a/parsing/dml_csr/networks/dml_csr.py b/parsing/dml_csr/networks/dml_csr.py
new file mode 100644
index 000000000..89b9408ec
--- /dev/null
+++ b/parsing/dml_csr/networks/dml_csr.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   dml_csr.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+import torch.nn as nn
+
+from torch.nn import functional as F
+from inplace_abn import InPlaceABNSync
+from .modules.ddgcn import DDualGCNHead
+from .modules.parsing import Parsing
+from .modules.edges import Edges
+from .modules.util import Bottleneck
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    "3x3 convolution with padding"
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class DML_CSR(nn.Module):
+    def __init__(self, 
+                 num_classes, 
+                 abn=InPlaceABNSync,
+                 trained=True):
+        super().__init__()
+        self.inplanes = 128
+        self.is_trained = trained
+
+        self.conv1 = conv3x3(3, 64, stride=2)
+        self.bn1 = abn(64)
+        self.relu1 = nn.ReLU(inplace=False)
+        self.conv2 = conv3x3(64, 64)
+        self.bn2 = abn(64)
+        self.relu2 = nn.ReLU(inplace=False)
+        self.conv3 = conv3x3(64, 128)
+        self.bn3 = abn(128)
+        self.relu3 = nn.ReLU(inplace=False)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layers = [3, 4, 23, 3]
+        self.abn = abn
+        strides = [1, 2, 1, 1]
+        dilations = [1, 1, 1, 2]
+
+        self.layer1 = self._make_layer(Bottleneck, 64, self.layers[0], stride=strides[0], dilation=dilations[0])
+        self.layer2 = self._make_layer(Bottleneck, 128, self.layers[1], stride=strides[1], dilation=dilations[1])
+        self.layer3 = self._make_layer(Bottleneck, 256, self.layers[2], stride=strides[2], dilation=dilations[2])
+        self.layer4 = self._make_layer(Bottleneck, 512, self.layers[3], stride=strides[3], dilation=dilations[3], multi_grid=(1,1,1))
+        # Context Aware
+        self.context = DDualGCNHead(2048, 512, abn)
+        self.layer6 = Parsing(512, 256, num_classes, abn)
+        # edge
+        if self.is_trained:
+            self.edge_layer = Edges(abn, out_fea=num_classes)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilation=1, multi_grid=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion,
+                          kernel_size=1, stride=stride, bias=False),
+                self.abn(planes * block.expansion, affine=True))
+
+        layers = []
+        generate_multi_grid = lambda index, grids: grids[index%len(grids)] if isinstance(grids, tuple) else 1
+        layers.append(block(self.inplanes, planes, stride, abn=self.abn, dilation=dilation, downsample=downsample, multi_grid=generate_multi_grid(0, multi_grid)))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, abn=self.abn, dilation=dilation, multi_grid=generate_multi_grid(i, multi_grid)))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        input = x
+        x = self.relu1(self.bn1(self.conv1(x)))
+        x = self.relu2(self.bn2(self.conv2(x)))
+        x1 = self.relu3(self.bn3(self.conv3(x)))
+        x = self.maxpool(x1)
+        x2 = self.layer1(x) # 119 x 119
+        x3 = self.layer2(x2) # 60 x 60
+        x4 = self.layer3(x3) # 60 x 60
+        x5 = self.layer4(x4) # 60 x 60
+        x = self.context(x5)
+        seg, x = self.layer6(x, x2)
+
+        if self.is_trained:
+            binary_edge, semantic_edge, edge_fea = self.edge_layer(x2,x3,x4)
+            return seg, binary_edge, semantic_edge
+        
+        return seg
+        
diff --git a/parsing/dml_csr/networks/modules/ddgcn.py b/parsing/dml_csr/networks/modules/ddgcn.py
new file mode 100644
index 000000000..368d60ee0
--- /dev/null
+++ b/parsing/dml_csr/networks/modules/ddgcn.py
@@ -0,0 +1,182 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   ddgcn.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+
+from inplace_abn import InPlaceABNSync
+
+
+class SpatialGCN(nn.Module):
+    def __init__(self, plane, abn=InPlaceABNSync):
+        super(SpatialGCN, self).__init__()
+        inter_plane = plane // 2
+        self.node_k = nn.Conv2d(plane, inter_plane, kernel_size=1)
+        self.node_v = nn.Conv2d(plane, inter_plane, kernel_size=1)
+        self.node_q = nn.Conv2d(plane, inter_plane, kernel_size=1)
+
+        self.conv_wg = nn.Conv1d(inter_plane, inter_plane, kernel_size=1, bias=False)
+        self.bn_wg   = nn.BatchNorm1d(inter_plane)
+        self.softmax = nn.Softmax(dim=2)
+
+        self.out = nn.Sequential(nn.Conv2d(inter_plane, plane, kernel_size=1),
+                                 abn(plane))
+
+        self.gamma = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x):
+        # b, c, h, w = x.size()
+        node_k = self.node_k(x)
+        node_v = self.node_v(x)
+        node_q = self.node_q(x)
+        b,c,h,w = node_k.size()
+        node_k = node_k.view(b, c, -1).permute(0, 2, 1)
+        node_q = node_q.view(b, c, -1)
+        node_v = node_v.view(b, c, -1).permute(0, 2, 1)
+        # A = k * q
+        # AV = k * q * v
+        # AVW = k *(q *v) * w
+        AV = torch.bmm(node_q,node_v)
+        AV = self.softmax(AV)
+        AV = torch.bmm(node_k, AV)
+        AV = AV.transpose(1, 2).contiguous()
+        AVW = self.conv_wg(AV)
+        AVW = self.bn_wg(AVW)
+        AVW = AVW.view(b, c, h, -1)
+        # out = F.relu_(self.out(AVW) + x)
+        out = self.gamma * self.out(AVW) + x
+        return out
+
+
+class DDualGCN(nn.Module):
+    """
+        Feature GCN with coordinate GCN
+    """
+    def __init__(self, planes, abn=InPlaceABNSync, ratio=4):
+        super(DDualGCN, self).__init__()
+
+        self.phi      = nn.Conv2d(planes, planes // ratio * 2, kernel_size=1, bias=False)
+        self.bn_phi   = abn(planes // ratio * 2)
+        self.theta    = nn.Conv2d(planes, planes // ratio, kernel_size=1, bias=False)
+        self.bn_theta = abn(planes // ratio)
+
+        #  Interaction Space
+        #  Adjacency Matrix: (-)A_g
+        self.conv_adj = nn.Conv1d(planes // ratio, planes // ratio, kernel_size=1, bias=False)
+        self.bn_adj   = nn.BatchNorm1d(planes // ratio)
+
+        #  State Update Function: W_g
+        self.conv_wg = nn.Conv1d(planes // ratio * 2, planes // ratio * 2, kernel_size=1, bias=False)
+        self.bn_wg   = nn.BatchNorm1d(planes // ratio * 2)
+
+        #  last fc
+        self.conv3 = nn.Conv2d(planes // ratio * 2, planes, kernel_size=1, bias=False)
+        self.bn3   = abn(planes)
+
+        self.local = nn.Sequential(
+            nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
+            abn(planes),
+            nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
+            abn(planes),
+            nn.Conv2d(planes, planes, 3, groups=planes, stride=2, padding=1, bias=False),
+            abn(planes))
+        self.gcn_local_attention = SpatialGCN(planes, abn)
+
+        self.final = nn.Sequential(nn.Conv2d(planes * 2, planes, kernel_size=1, bias=False),
+                                   abn(planes))
+
+        self.gamma1 = nn.Parameter(torch.zeros(1))
+
+    def to_matrix(self, x):
+        n, c, h, w = x.size()
+        x = x.view(n, c, -1)
+        return x
+
+    def forward(self, feat):
+        # # # # Local # # # #
+        x = feat
+        local = self.local(feat)
+        local = self.gcn_local_attention(local)
+        local = F.interpolate(local, size=x.size()[2:], mode='bilinear', align_corners=True)
+        spatial_local_feat = x * local + x
+
+        # # # # Projection Space # # # #
+        x_sqz, b = x, x
+
+        x_sqz = self.phi(x_sqz)
+        x_sqz = self.bn_phi(x_sqz)
+        x_sqz = self.to_matrix(x_sqz)
+
+        b = self.theta(b)
+        b = self.bn_theta(b)
+        b = self.to_matrix(b)
+
+        # Project
+        z_idt = torch.matmul(x_sqz, b.transpose(1, 2))  # channel
+
+        # # # # Interaction Space # # # #
+        z = z_idt.transpose(1, 2).contiguous()
+
+        z = self.conv_adj(z)
+        z = self.bn_adj(z)
+
+        z = z.transpose(1, 2).contiguous()
+        # Laplacian smoothing: (I - A_g)Z => Z - A_gZ
+        z += z_idt
+
+        z = self.conv_wg(z)
+        z = self.bn_wg(z)
+
+        # # # # Re-projection Space # # # #
+        # Re-project
+        y = torch.matmul(z, b)
+
+        n, _, h, w = x.size()
+        y = y.view(n, -1, h, w)
+
+        y = self.conv3(y)
+        y = self.bn3(y)
+
+        # g_out = x + y
+        # g_out = F.relu_(x+y)
+        g_out = self.gamma1*y + x  
+
+        # cat or sum, nearly the same results
+        out = self.final(torch.cat((spatial_local_feat, g_out), 1))
+
+        return out
+
+
+class DDualGCNHead(nn.Module):
+    def __init__(self, inplanes, interplanes, abn=InPlaceABNSync):
+        super(DDualGCNHead, self).__init__()
+        self.conva = nn.Sequential(nn.Conv2d(inplanes, interplanes, 3, padding=1, bias=False),
+                                   abn(interplanes))
+        self.dualgcn = DDualGCN(interplanes, abn)
+        self.convb = nn.Sequential(nn.Conv2d(interplanes, interplanes, 3, padding=1, bias=False),
+                                   abn(interplanes))
+
+        self.bottleneck = nn.Sequential(
+            nn.Conv2d(inplanes + interplanes, interplanes, kernel_size=3, padding=1, dilation=1, bias=False),
+            abn(interplanes)
+        )
+
+    def forward(self, x):
+        output = self.conva(x)
+        output = self.dualgcn(output)
+        output = self.convb(output)
+        output = self.bottleneck(torch.cat([x, output], 1))
+        return output
diff --git a/parsing/dml_csr/networks/modules/edges.py b/parsing/dml_csr/networks/modules/edges.py
new file mode 100644
index 000000000..5059f235d
--- /dev/null
+++ b/parsing/dml_csr/networks/modules/edges.py
@@ -0,0 +1,66 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   edges.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+
+from inplace_abn import InPlaceABNSync
+
+
+class Edges(nn.Module):
+
+    def __init__(self, abn=InPlaceABNSync, in_fea=[256,512,1024], mid_fea=256, out_fea=2):
+        super(Edges, self).__init__()
+        
+        self.conv1 =  nn.Sequential(
+            nn.Conv2d(in_fea[0], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(mid_fea)
+            ) 
+        self.conv2 =  nn.Sequential(
+            nn.Conv2d(in_fea[1], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(mid_fea)
+            )  
+        self.conv3 =  nn.Sequential(
+            nn.Conv2d(in_fea[2], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(mid_fea)
+        )
+        self.conv4 = nn.Conv2d(mid_fea,out_fea, kernel_size=3, padding=1, dilation=1, bias=True)
+        self.conv5_b = nn.Conv2d(out_fea*3,2, kernel_size=1, padding=0, dilation=1, bias=True)
+        self.conv5 = nn.Conv2d(out_fea*3,out_fea, kernel_size=1, padding=0, dilation=1, bias=True)
+            
+
+    def forward(self, x1, x2, x3):
+        _, _, h, w = x1.size()
+        
+        edge1_fea = self.conv1(x1)
+        edge1 = self.conv4(edge1_fea)
+        edge2_fea = self.conv2(x2)
+        edge2 = self.conv4(edge2_fea)
+        edge3_fea = self.conv3(x3)
+        edge3 = self.conv4(edge3_fea)        
+        
+        edge2_fea =  F.interpolate(edge2_fea, size=(h, w), mode='bilinear',align_corners=True) 
+        edge3_fea =  F.interpolate(edge3_fea, size=(h, w), mode='bilinear',align_corners=True) 
+        edge2 =  F.interpolate(edge2, size=(h, w), mode='bilinear',align_corners=True)
+        edge3 =  F.interpolate(edge3, size=(h, w), mode='bilinear',align_corners=True) 
+ 
+        edge = torch.cat([edge1, edge2, edge3], dim=1)
+        edge_fea = torch.cat([edge1_fea, edge2_fea, edge3_fea], dim=1)
+        semantic_edge = self.conv5(edge)
+        binary_edge = self.conv5_b(edge)
+         
+        return binary_edge, semantic_edge, edge_fea
+
diff --git a/parsing/dml_csr/networks/modules/parsing.py b/parsing/dml_csr/networks/modules/parsing.py
new file mode 100644
index 000000000..ed1addb31
--- /dev/null
+++ b/parsing/dml_csr/networks/modules/parsing.py
@@ -0,0 +1,51 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   parsing.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+
+from inplace_abn import InPlaceABNSync
+
+
+class Parsing(nn.Module):
+    def __init__(self, in_plane1, in_plane2, num_classes, abn=InPlaceABNSync):
+        super(Parsing, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_plane1, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(256)
+            )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_plane2, 48, kernel_size=1, stride=1, padding=0, dilation=1, bias=False),
+            abn(48)
+            )
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(304, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(256),
+            nn.Conv2d(256, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            abn(256)
+            )
+        self.conv4 = nn.Conv2d(256, num_classes, kernel_size=1, padding=0, dilation=1, bias=True)
+
+    def forward(self, xt, xl):
+        _, _, h, w = xl.size()
+
+        xt = F.interpolate(self.conv1(xt), size=(h, w), mode='bilinear', align_corners=True)
+        xl = self.conv2(xl)
+        x = torch.cat([xt, xl], dim=1)
+        x = self.conv3(x)
+        seg = self.conv4(x)
+        return seg, x 
+      
diff --git a/parsing/dml_csr/networks/modules/util.py b/parsing/dml_csr/networks/modules/util.py
new file mode 100644
index 000000000..fbe23dd8e
--- /dev/null
+++ b/parsing/dml_csr/networks/modules/util.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   util.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import torch.nn as nn
+
+from inplace_abn import InPlaceABNSync
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+    def __init__(self, inplanes, planes, stride=1, abn=InPlaceABNSync, dilation=1, downsample=None, fist_dilation=1, multi_grid=1):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = abn(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
+                               padding=dilation*multi_grid, dilation=dilation*multi_grid, bias=False)
+        self.bn2 = abn(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
+        self.bn3 = abn(planes * 4)
+        self.relu = nn.ReLU(inplace=False)
+        self.relu_inplace = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.dilation = dilation
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out = out + residual      
+        out = self.relu_inplace(out)
+
+        return out
diff --git a/parsing/dml_csr/test.py b/parsing/dml_csr/test.py
new file mode 100644
index 000000000..3aad2527a
--- /dev/null
+++ b/parsing/dml_csr/test.py
@@ -0,0 +1,182 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   datasets.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import cv2
+import numpy as np
+import matplotlib.pyplot as plt
+import os
+import torch
+import torch.backends.cudnn as cudnn
+import torchvision.transforms as transforms
+
+from copy import deepcopy
+from inplace_abn import InPlaceABN
+from dataset import datasets
+from networks import dml_csr
+from utils import miou
+
+torch.multiprocessing.set_start_method("spawn", force=True)
+
+DATA_DIRECTORY = './datasets/Helen'
+IGNORE_LABEL = 255
+NUM_CLASSES = 20
+SNAPSHOT_DIR = './snapshots/'
+INPUT_SIZE = (473,473)
+
+
+def get_arguments():
+    """Parse all the arguments provided from the CLI.
+    
+    Returns:
+      A list of parsed arguments.
+    """
+    parser = argparse.ArgumentParser(description="DML_CSR Network")
+    parser.add_argument("--batch-size", type=int, default=1,
+                        help="Number of images sent to the network in one step.")
+    parser.add_argument("--data-dir", type=str, default=DATA_DIRECTORY,
+                        help="Path to the directory containing the PASCAL VOC dataset.")
+    parser.add_argument("--out-dir", type=str, default=DATA_DIRECTORY,
+                        help="Path to the directory containing the PASCAL VOC dataset.")
+    parser.add_argument("--dataset", type=str, default='val',
+                        help="Path to the file listing the images in the dataset.")
+    parser.add_argument("--ignore-label", type=int, default=IGNORE_LABEL,
+                        help="The index of the label to ignore during the training.")
+    parser.add_argument("--num-classes", type=int, default=NUM_CLASSES,
+                        help="Number of classes to predict (including background).")
+    parser.add_argument("--restore-from", type=str,
+                        help="Where restore model parameters from.")
+    parser.add_argument("--gpu", type=str, default='7',
+                        help="choose gpu device.")
+    parser.add_argument("--input-size", type=str, default=INPUT_SIZE,
+                        help="Comma-separated string with height and width of images.")
+    parser.add_argument("--local_rank", type=int, default=0,
+                        help="choose gpu numbers") 
+    parser.add_argument('--dist-backend', default='nccl', type=str,
+                        help='distributed backend')
+    parser.add_argument("--model_type", type=int, default=0,
+                        help="choose model type") 
+    return parser.parse_args()
+
+
+def valid(model, valloader, input_size, num_samples, dir=None, dir_edge=None, dir_img=None):
+
+    height = input_size[0]
+    width  = input_size[1]
+    with torch.autograd.profiler.profile(enabled=True, use_cuda=True, \
+        record_shapes=False, profile_memory=False) as prof:
+        model.eval()
+        parsing_preds = np.zeros((num_samples, height, width), dtype=np.uint8)
+        scales = np.zeros((num_samples, 2), dtype=np.float32)
+        centers = np.zeros((num_samples, 2), dtype=np.int32)
+
+        idx = 0
+        interp = torch.nn.Upsample(size=(height, width), mode='bilinear', align_corners=True)
+
+        with torch.no_grad():
+            for index, batch in enumerate(valloader):
+                image, meta = batch
+                num_images = image.size(0)
+                if index % 10 == 0:
+                    print('%d  processd' % (index * num_images))
+
+                c = meta['center'].numpy()
+                s = meta['scale'].numpy()
+                scales[idx:idx + num_images, :] = s[:, :]
+                centers[idx:idx + num_images, :] = c[:, :]
+
+                results = model(image.cuda())
+                outputs = results
+
+                if isinstance(results, list):
+                    outputs = results[0]
+
+                if isinstance(outputs, list):
+                    for k, output in enumerate(outputs):
+                        parsing = output
+                        nums = len(parsing)
+                        parsing = interp(parsing).data.cpu().numpy()
+                        parsing = parsing.transpose(0, 2, 3, 1)  # NCHW NHWC
+                        parsing_preds[idx:idx + nums, :, :] = np.asarray(np.argmax(parsing, axis=3), dtype=np.uint8)
+                        idx += nums
+                else:
+                    parsing = outputs
+                    parsing = interp(parsing).data.cpu().numpy()
+                    parsing = parsing.transpose(0, 2, 3, 1)  # NCHW NHWC
+                    parsing_preds[idx:idx + num_images, :, :] = np.asarray(np.argmax(parsing, axis=3), dtype=np.uint8)
+
+                    if dir is not None:
+                        for i in range(len(meta['name'])):
+                            cv2.imwrite(os.path.join(dir, meta['name'][i] + '.png'), np.asarray(np.argmax(parsing, axis=3))[i])
+                    idx += num_images
+        parsing_preds = parsing_preds[:num_samples, :, :]
+
+    return parsing_preds, scales, centers
+
+
+def main():
+    """Create the model and start the evaluation process."""
+
+    args = get_arguments()
+
+    os.environ["CUDA_VISIBLE_DEVICES"]=args.gpu
+    gpus = [int(i) for i in args.gpu.split(',')]
+
+    print(args.gpu)
+
+    h, w = map(int, args.input_size.split(','))
+    
+    input_size = (h, w)
+
+    cudnn.benchmark = True
+    cudnn.enabled = True
+
+    model = dml_csr.DML_CSR(args.num_classes, InPlaceABN, False)
+    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
+                                     std=[0.229, 0.224, 0.225])
+
+    transform = transforms.Compose([
+        transforms.ToTensor(),
+        normalize,
+    ])
+
+    dataset = datasets.FaceDataSet(args.data_dir, args.dataset, \
+        crop_size=input_size, transform=transform)
+    num_samples = len(dataset)
+
+    valloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, \
+        shuffle=False, pin_memory=True)
+
+    restore_from = args.restore_from
+    print(restore_from)
+    state_dict = torch.load(restore_from,map_location='cuda:0')
+    model.load_state_dict(state_dict)
+        
+    model.cuda()
+    model.eval()
+
+    save_path =  os.path.join(args.out_dir, args.dataset, 'parsing')
+    if not os.path.exists(save_path):
+        os.makedirs(save_path)
+
+    parsing_preds, scales, centers = valid(model, valloader, input_size, num_samples, save_path)
+    mIoU, f1 = miou.compute_mean_ioU(parsing_preds, scales, centers, args.num_classes, args.data_dir, \
+        input_size, args.dataset, reverse=True)
+
+    print(mIoU)
+    print(f1)
+
+if __name__ == '__main__':
+    main()
diff --git a/parsing/dml_csr/train.py b/parsing/dml_csr/train.py
new file mode 100644
index 000000000..e6591bda6
--- /dev/null
+++ b/parsing/dml_csr/train.py
@@ -0,0 +1,418 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   train.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2015 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import os
+import os.path as osp
+import timeit
+import torch
+import torch.nn as nn
+import torch.backends.cudnn as cudnn
+import torch.distributed as dist
+import torch.optim as optim
+import torchvision.transforms as transforms
+import torchvision.utils as vutils
+import utils.schp as schp
+
+from datetime import datetime
+from tensorboardX import SummaryWriter
+from torch.utils.data.distributed import DistributedSampler
+from inplace_abn import InPlaceABNSync
+from inplace_abn import InPlaceABN
+
+from dataset import datasets
+from networks import dml_csr
+from utils.logging import get_root_logger
+from utils.utils import decode_parsing, inv_preprocess, SingleGPU
+from utils.encoding import DataParallelModel, DataParallelCriterion 
+from utils.miou import compute_mean_ioU
+from utils.warmup_scheduler import SGDRScheduler
+from loss.criterion import Criterion
+from test import valid
+
+torch.multiprocessing.set_start_method("spawn", force=True)
+
+RESTORE_FROM = 'resnet101-imagenet.pth'
+
+
+def str2bool(v):
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+
+def get_arguments():
+    """Parse all the arguments provided from the CLI.
+
+    Returns:
+      A list of parsed arguments.
+    """
+    parser = argparse.ArgumentParser(description="Training Network")
+    # Data Preference
+    parser.add_argument("--data-dir", type=str, default='./datasets',
+                        help="Path to the directory containing the dataset.")
+    parser.add_argument("--train-dataset", type=str, default='train', choices=['train', 'valid', 'test'],
+                        help="Path to the file listing the images in the dataset.")
+    parser.add_argument("--valid-dataset", type=str, default='test', choices=['valid', 'test_resize', 'test'],
+                        help="Path to the file listing the images in the dataset.")
+    parser.add_argument("--batch-size", type=int, default=8,
+                        help="Number of images sent to the network in one step.")
+    parser.add_argument("--input-size", type=str, default='473,473',
+                        help="Comma-separated string with height and width of images.")
+    parser.add_argument("--num-classes", type=int, default=11,
+                        help="Number of classes to predict (including background).")
+    parser.add_argument("--edge-classes", type=int, default=2,
+                        help="Number of classes to predict (including background).") 
+    parser.add_argument("--ignore-label", type=int, default=255,
+                        help="The index of the label to ignore during the training.")
+    parser.add_argument("--random-mirror", action="store_true",
+                        help="Whether to randomly mirror the inputs during the training.")
+    parser.add_argument("--random-scale", action="store_true",
+                        help="Whether to randomly scale the inputs during the training.")
+    parser.add_argument("--random-seed", type=int, default=1234,
+                        help="Random seed to have reproducible results.")
+    # Model
+    parser.add_argument("--restore-from", type=str, default=None,
+                        help="Where restore model parameters from.")
+    parser.add_argument("--snapshot-dir", type=str, default='./snapshots/',
+                        help="Where to save snapshots of the model.")
+    # Training Strategy
+    parser.add_argument("--save-num-images", type=int, default=2,
+                        help="How many images to save.")
+    parser.add_argument("--learning-rate", type=float, default=1e-3,
+                        help="Base learning rate for training with polynomial decay.")
+    parser.add_argument("--momentum", type=float, default=0.9,
+                        help="Momentum component of the optimiser.")
+    parser.add_argument("--weight-decay", type=float, default=0.0005,
+                        help="Regularisation parameter for L2-loss.")
+    parser.add_argument("--power", type=float, default=0.9,
+                        help="Decay parameter to compute the learning rate.")
+    parser.add_argument("--eval_epochs", type=int, default=1,
+                        help="Number of classes to predict (including background).")                                            
+    parser.add_argument("--start-epoch", type=int, default=0,
+                        help="choose the number of recurrence.")
+    parser.add_argument("--epochs", type=int, default=150,
+                        help="choose the number of recurrence.")
+    # Distributed Training
+    parser.add_argument("--gpu", type=str, default='None',
+                        help="choose gpu numbers")
+    parser.add_argument("--local_rank", type=int, default=0,
+                        help="choose gpu numbers")
+    parser.add_argument('--dist-backend', default='nccl', type=str,
+                        help='distributed backend')
+    # self correlation training
+    parser.add_argument("--schp-start", type=int, default=100, 
+                        help='schp start epoch')
+    parser.add_argument("--cycle-epochs", type=int, default=10, 
+                        help='schp cyclical epoch')
+    parser.add_argument("--schp-restore", type=str, default=None,
+                        help="Where restore schp model parameters from.")
+    parser.add_argument("--lambda-s", type=float, default=1,   
+                        help='segmentation loss weight')
+    parser.add_argument("--lambda-e", type=float, default=1, 
+                        help='edge loss weight')
+    parser.add_argument("--lambda-c", type=float, default=0.1, 
+                        help='segmentation-edge consistency loss weight')
+    return parser.parse_args()
+
+
+args = get_arguments()
+TIMESTAMP = "{0:%Y_%m_%dT%H_%M_%S/}".format(datetime.now())
+global writer
+
+def lr_poly(base_lr, iter, max_iter, power):
+    return base_lr * ((1 - float(iter) / max_iter) ** (power))
+
+
+def adjust_learning_rate(optimizer, i_iter, total_iters):
+    """Sets the learning rate to the initial LR divided by 5 at 60th, 120th and 160th epochs"""
+    lr = lr_poly(args.learning_rate, i_iter, total_iters, args.power)
+    optimizer.param_groups[0]['lr'] = lr
+    return lr
+
+
+def adjust_learning_rate_pose(optimizer, epoch):
+    decay = 0
+    if epoch + 1 >= 230:
+        decay = 0.05
+    elif epoch + 1 >= 200:
+        decay = 0.1
+    elif epoch + 1 >= 120:
+        decay = 0.25
+    elif epoch + 1 >= 90:
+        decay = 0.5
+    else:
+        decay = 1
+
+    lr = args.learning_rate * decay
+    for param_group in optimizer.param_groups:
+        param_group['lr'] = lr
+    return lr
+
+
+def set_bn_eval(m):
+    classname = m.__class__.__name__
+    if classname.find('BatchNorm') != -1:
+        m.eval()
+
+
+def set_bn_momentum(m):
+    classname = m.__class__.__name__
+    if classname.find('BatchNorm') != -1 or classname.find('InPlaceABN') != -1:
+        m.momentum = 0.0003
+
+
+def main():
+    """Create the model and start the training."""
+    cycle_n = 0
+    start_epoch = args.start_epoch
+    writer = SummaryWriter(osp.join(args.snapshot_dir, TIMESTAMP)) 
+    if not os.path.exists(args.snapshot_dir):
+        os.makedirs(args.snapshot_dir)
+
+    h, w = map(int, args.input_size.split(','))
+    input_size = [h, w]
+    best_f1 = 0
+
+    torch.cuda.set_device(args.local_rank)
+
+    try:
+        world_size = int(os.environ['WORLD_SIZE'])
+        distributed = world_size > 1
+    except:
+        distributed = False
+        world_size = 1
+    if distributed:
+        dist.init_process_group(backend=args.dist_backend, init_method='env://')
+    rank = 0 if not distributed else dist.get_rank()
+
+    log_file = args.snapshot_dir + '/' + TIMESTAMP + 'output.log'
+    logger = get_root_logger(log_file=log_file, log_level='INFO')
+    logger.info(f'Distributed training: {distributed}')
+
+    cudnn.enabled = True
+    cudnn.benchmark = True
+    torch.backends.cudnn.deterministic = False
+    torch.backends.cudnn.enabled = True
+ 
+    if distributed:
+        model = dml_csr.DML_CSR(args.num_classes)
+        schp_model = dml_csr.DML_CSR(args.num_classes)
+    else:
+        model = dml_csr.DML_CSR(args.num_classes, InPlaceABN)
+        schp_model = dml_csr.DML_CSR(args.num_classes, InPlaceABN)
+    
+    if args.restore_from is not None:
+        print('Resume training from {}'.format(args.restore_from))
+        model.load_state_dict(torch.load(args.restore_from), True)
+        start_epoch = int(float(args.restore_from.split('.')[0].split('_')[-1])) + 1
+    else:
+        resnet_params = torch.load(RESTORE_FROM)
+        new_params = model.state_dict().copy()
+        for i in resnet_params:
+            i_parts = i.split('.')
+            if not i_parts[0] == 'fc':
+                new_params['.'.join(i_parts[0:])] = resnet_params[i]
+        model.load_state_dict(new_params)
+    model.cuda()
+
+    args.schp_restore = osp.join(args.snapshot_dir, TIMESTAMP, 'best.pth')
+    if os.path.exists(args.schp_restore):
+        print('Resume schp checkpoint from {}'.format(args.schp_restore))
+        schp_model.load_state_dict(torch.load(args.schp_restore), True)
+    else:
+        schp_resnet_params = torch.load(RESTORE_FROM)
+        schp_new_params = schp_model.state_dict().copy()
+        for i in schp_resnet_params:
+            i_parts = i.split('.')
+            if not i_parts[0] == 'fc':
+                schp_new_params['.'.join(i_parts[0:])] = schp_resnet_params[i]
+        schp_model.load_state_dict(schp_new_params)
+    schp_model.cuda()
+
+    if distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+        schp_model = torch.nn.parallel.DistributedDataParallel(schp_model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+    else:
+        model = SingleGPU(model)
+        schp_model = SingleGPU(schp_model)
+
+    criterion = Criterion(loss_weight=[1, 1, 1, 4, 1], 
+                lambda_1=args.lambda_s, lambda_2=args.lambda_e, lambda_3=args.lambda_c, num_classes=args.num_classes)
+    criterion.cuda()
+
+    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+    transform = transforms.Compose([transforms.ToTensor(), normalize])
+    
+    train_dataset = FaceDataSet(args.data_dir, args.train_dataset, crop_size=input_size, transform=transform)
+    if distributed:
+        train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
+    else:
+        train_sampler = None
+    trainloader = data.DataLoader(train_dataset, batch_size=args.batch_size , shuffle=False, num_workers=2,
+                                  pin_memory=True, drop_last=True, sampler=train_sampler)
+    
+    val_dataset = datasets[str(args.model_type)](args.data_dir, args.valid_dataset, crop_size=input_size, transform=transform)
+    num_samples = len(val_dataset)
+    valloader = data.DataLoader(val_dataset, batch_size=args.batch_size , shuffle=False, pin_memory=True, drop_last=False)
+
+    # Optimizer Initialization
+    optimizer = optim.SGD(model.parameters(), lr=args.learning_rate, momentum=args.momentum,
+                        weight_decay=args.weight_decay)
+    lr_scheduler = SGDRScheduler(optimizer, total_epoch=args.epochs,
+                                eta_min=args.learning_rate / 100, warmup_epoch=10,
+                                start_cyclical=args.schp_start, cyclical_base_lr=args.learning_rate / 2,
+                                cyclical_epoch=args.cycle_epochs)
+
+    optimizer.zero_grad()
+
+    total_iters = args.epochs * len(trainloader)
+    start = timeit.default_timer()
+    for epoch in range(start_epoch, args.epochs):
+        model.train()
+        if distributed:
+            train_sampler.set_epoch(epoch)
+        for i_iter, batch in enumerate(trainloader):
+            i_iter += len(trainloader) * epoch
+
+            if epoch < args.schp_start:
+                lr = adjust_learning_rate(optimizer, i_iter, total_iters)
+            else:
+                lr = lr_scheduler.get_lr()[0]
+
+            images, labels, edges, semantic_edges, _ = batch
+            labels = labels.long().cuda(non_blocking=True)
+            edges = edges.long().cuda(non_blocking=True)
+            semantic_edges = semantic_edges.long().cuda(non_blocking=True)
+
+            preds = model(images)
+            
+            if cycle_n >= 1:
+                with torch.no_grad():
+                    soft_preds, soft_edges, soft_semantic_edges = schp_model(images)
+            else:
+                soft_preds = None
+                soft_edges = None
+                soft_semantic_edges = None
+
+            loss = criterion(preds, [labels, edges, semantic_edges, soft_preds, soft_edges, soft_semantic_edges], cycle_n)
+            
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            lr_scheduler.step()
+
+            with torch.no_grad():
+                loss = loss.detach() * labels.shape[0]
+                count = labels.new_tensor([labels.shape[0]], dtype=torch.long)
+                if dist.is_initialized():
+                    dist.all_reduce(count, dist.ReduceOp.SUM)
+                    dist.all_reduce(loss, dist.ReduceOp.SUM)
+                loss /= count.item()
+
+            if not dist.is_initialized() or dist.get_rank() == 0:
+                if i_iter % 50 == 0:
+                    writer.add_scalar('learning_rate', lr, i_iter)
+                    writer.add_scalar('loss', loss.data.cpu().numpy(), i_iter)
+
+                if i_iter % 500 == 0:
+                    images_inv = inv_preprocess(images, args.save_num_images)
+                    labels_colors = decode_parsing(labels, args.save_num_images, args.num_classes, is_pred=False)
+                    edges_colors = decode_parsing(edges, args.save_num_images, 2, is_pred=False)
+                    semantic_edges_colors = decode_parsing(semantic_edges, args.save_num_images, args.num_classes, is_pred=False)
+
+                    if isinstance(preds, list):
+                        preds = preds[0]
+                    preds_colors = decode_parsing(preds[0], args.save_num_images, args.num_classes, is_pred=True)
+                    pred_edges = decode_parsing(preds[1], args.save_num_images, 2, is_pred=True)
+                    pred_semantic_edges_colors = decode_parsing(preds[2], args.save_num_images, args.num_classes, is_pred=True)
+
+                    img = vutils.make_grid(images_inv, normalize=False, scale_each=True)
+                    lab = vutils.make_grid(labels_colors, normalize=False, scale_each=True)
+                    pred = vutils.make_grid(preds_colors, normalize=False, scale_each=True)
+                    edge = vutils.make_grid(edges_colors, normalize=False, scale_each=True)
+                    pred_edge = vutils.make_grid(pred_edges, normalize=False, scale_each=True)
+                    pred_semantic_edges = vutils.make_grid(pred_semantic_edges_colors, normalize=False, scale_each=True)
+
+
+                    writer.add_image('Images/', img, i_iter)
+                    writer.add_image('Labels/', lab, i_iter)
+                    writer.add_image('Preds/', pred, i_iter)
+                    writer.add_image('Edge/', edge, i_iter)
+                    writer.add_image('Pred_edge/', pred_edge, i_iter)
+    
+                cur_loss = loss.data.cpu().numpy()
+                logger.info(f'iter = {i_iter} of {total_iters} completed, loss = {cur_loss}, lr = {lr}')
+
+        if (epoch + 1) % (args.eval_epochs) == 0:
+            parsing_preds, scales, centers = valid(model, valloader, input_size, num_samples)
+            mIoU, f1 = compute_mean_ioU(parsing_preds, scales, centers, args.num_classes, args.data_dir, input_size, args.valid_dataset, True)
+
+            if not dist.is_initialized() or dist.get_rank() == 0:  
+                torch.save(model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'checkpoint_{}.pth'.format(epoch + 1)))
+                if 'Helen' in args.data_dir:
+                    if f1['overall'] > best_f1:
+                        torch.save(model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'best.pth'))
+                        best_f1 = f1['overall']
+                else:
+                    if f1['Mean_F1'] > best_f1:
+                        torch.save(model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'best.pth'))
+                        best_f1 = f1['Mean_F1']
+
+            writer.add_scalars('mIoU', mIoU, epoch)
+            writer.add_scalars('f1', f1, epoch)
+            logger.info(f'mIoU = {mIoU}, and f1 = {f1} of epoch = {epoch}, util now, best_f1 = {best_f1}')
+
+            if (epoch + 1) >= args.schp_start and (epoch + 1 - args.schp_start) % args.cycle_epochs == 0:
+                logger.info(f'Self-correction cycle number {cycle_n}')
+                schp.moving_average(schp_model, model, 1.0 / (cycle_n + 1))
+                cycle_n += 1
+                schp.bn_re_estimate(trainloader, schp_model)
+                parsing_preds, scales, centers = valid(schp_model, valloader, input_size, num_samples)
+                mIoU, f1 = compute_mean_ioU(parsing_preds, scales, centers, args.num_classes, args.data_dir, input_size, args.valid_dataset, True)
+
+                if not dist.is_initialized() or dist.get_rank() == 0: 
+                    torch.save(schp_model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'schp_{}_checkpoint.pth'.format(cycle_n)))
+
+                    if 'Helen' in args.data_dir:
+                        if f1['overall'] > best_f1:
+                            torch.save(schp_model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'best.pth'))
+                            best_f1 = f1['overall']
+                    else:
+                        if f1['Mean_F1'] > best_f1:
+                            torch.save(schp_model.module.state_dict(), osp.join(args.snapshot_dir, TIMESTAMP, 'best.pth'))
+                            best_f1 = f1['Mean_F1']
+                writer.add_scalars('mIoU', mIoU, epoch)
+                writer.add_scalars('f1', f1, epoch)
+                logger.info(f'mIoU = {mIoU}, and f1 = {f1} of epoch = {epoch}, util now, best_f1 = {best_f1}')
+
+            torch.cuda.empty_cache()
+            end = timeit.default_timer()
+            print('epoch = {} of {} completed using {} s'.format(epoch, args.epochs,
+                                                                (end - start) / (epoch - start_epoch + 1)))
+
+    end = timeit.default_timer()
+    print(end - start, 'seconds')
+ 
+
+if __name__ == '__main__':
+    main()
diff --git a/parsing/dml_csr/utils/encoding.py b/parsing/dml_csr/utils/encoding.py
new file mode 100644
index 000000000..098ca8185
--- /dev/null
+++ b/parsing/dml_csr/utils/encoding.py
@@ -0,0 +1,257 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   encoding.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import functools
+import threading
+import torch
+import torch.cuda.comm as comm
+
+from torch.autograd import Variable, Function
+from torch.nn.parallel.data_parallel import DataParallel
+from torch.nn.parallel.parallel_apply import get_a_var
+from torch.nn.parallel._functions import ReduceAddCoalesced, Broadcast
+
+torch_ver = torch.__version__[:3]
+
+__all__ = ['allreduce', 'DataParallelModel', 'DataParallelCriterion',
+           'patch_replication_callback']
+
+def allreduce(*inputs):
+    """Cross GPU all reduce autograd operation for calculate mean and
+    variance in SyncBN.
+    """
+    return AllReduce.apply(*inputs)
+
+class AllReduce(Function):
+    @staticmethod
+    def forward(ctx, num_inputs, *inputs):
+        ctx.num_inputs = num_inputs
+        ctx.target_gpus = [inputs[i].get_device() for i in range(0, len(inputs), num_inputs)]
+        inputs = [inputs[i:i + num_inputs]
+                 for i in range(0, len(inputs), num_inputs)]
+        # sort before reduce sum
+        inputs = sorted(inputs, key=lambda i: i[0].get_device())
+        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
+        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
+        return tuple([t for tensors in outputs for t in tensors])
+
+    @staticmethod
+    def backward(ctx, *inputs):
+        inputs = [i.data for i in inputs]
+        inputs = [inputs[i:i + ctx.num_inputs]
+                 for i in range(0, len(inputs), ctx.num_inputs)]
+        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
+        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
+        return (None,) + tuple([Variable(t) for tensors in outputs for t in tensors])
+
+
+class Reduce(Function):
+    @staticmethod
+    def forward(ctx, *inputs):
+        ctx.target_gpus = [inputs[i].get_device() for i in range(len(inputs))]
+        inputs = sorted(inputs, key=lambda i: i.get_device())
+        return comm.reduce_add(inputs)
+
+    @staticmethod
+    def backward(ctx, gradOutput):
+        return Broadcast.apply(ctx.target_gpus, gradOutput)
+
+
+class DataParallelModel(DataParallel):
+    """Implements data parallelism at the module level.
+
+    This container parallelizes the application of the given module by
+    splitting the input across the specified devices by chunking in the
+    batch dimension.
+    In the forward pass, the module is replicated on each device,
+    and each replica handles a portion of the input. During the backwards pass, gradients from each replica are summed into the original module.
+    Note that the outputs are not gathered, please use compatible
+    :class:`encoding.parallel.DataParallelCriterion`.
+
+    The batch size should be larger than the number of GPUs used. It should
+    also be an integer multiple of the number of GPUs so that each chunk is
+    the same size (so that each GPU processes the same number of samples).
+
+    Args:
+        module: module to be parallelized
+        device_ids: CUDA devices (default: all devices)
+
+    Reference:
+        Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi,
+        Amit Agrawal. “Context Encoding for Semantic Segmentation.
+        *The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018*
+
+    Example::
+
+        >>> net = encoding.nn.DataParallelModel(model, device_ids=[0, 1, 2])
+        >>> y = net(x)
+    """
+    def gather(self, outputs, output_device):
+        return outputs
+
+    def replicate(self, module, device_ids):
+        modules = super(DataParallelModel, self).replicate(module, device_ids)
+        execute_replication_callbacks(modules)
+        return modules
+
+
+class DataParallelCriterion(DataParallel):
+    """
+    Calculate loss in multiple-GPUs, which balance the memory usage for
+    Semantic Segmentation.
+
+    The targets are splitted across the specified devices by chunking in
+    the batch dimension. Please use together with :class:`encoding.parallel.DataParallelModel`.
+
+    Reference:
+        Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi,
+        Amit Agrawal. “Context Encoding for Semantic Segmentation.
+        *The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018*
+
+    Example::
+
+        >>> net = encoding.nn.DataParallelModel(model, device_ids=[0, 1, 2])
+        >>> criterion = encoding.nn.DataParallelCriterion(criterion, device_ids=[0, 1, 2])
+        >>> y = net(x)
+        >>> loss = criterion(y, target)
+    """
+    def forward(self, inputs, *targets, **kwargs):
+        # input should be already scatterd
+        # scattering the targets instead
+        if not self.device_ids:
+            return self.module(inputs, *targets, **kwargs)
+        targets, kwargs = self.scatter(targets, kwargs, self.device_ids)
+        if len(self.device_ids) == 1:
+            return self.module(inputs, *targets[0], **kwargs[0])
+        replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
+        outputs = _criterion_parallel_apply(replicas, inputs, targets, kwargs)
+        return Reduce.apply(*outputs) / len(outputs)
+        #return self.gather(outputs, self.output_device).mean()
+
+
+def _criterion_parallel_apply(modules, inputs, targets, kwargs_tup=None, devices=None):
+    assert len(modules) == len(inputs)
+    assert len(targets) == len(inputs)
+    if kwargs_tup:
+        assert len(modules) == len(kwargs_tup)
+    else:
+        kwargs_tup = ({},) * len(modules)
+    if devices is not None:
+        assert len(modules) == len(devices)
+    else:
+        devices = [None] * len(modules)
+
+    lock = threading.Lock()
+    results = {}
+    if torch_ver != "0.3":
+        grad_enabled = torch.is_grad_enabled()
+
+    def _worker(i, module, input, target, kwargs, device=None):
+        if torch_ver != "0.3":
+            torch.set_grad_enabled(grad_enabled)
+        if device is None:
+            device = get_a_var(input).get_device()
+        try:
+            if not isinstance(input, tuple):
+                input = (input,)
+            with torch.cuda.device(device):
+                output = module(*(input + target), **kwargs)
+            with lock:
+                results[i] = output
+        except Exception as e:
+            with lock:
+                results[i] = e
+
+    if len(modules) > 1:
+        threads = [threading.Thread(target=_worker,
+                                    args=(i, module, input, target,
+                                          kwargs, device),)
+                   for i, (module, input, target, kwargs, device) in
+                   enumerate(zip(modules, inputs, targets, kwargs_tup, devices))]
+
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+    else:
+        _worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
+
+    outputs = []
+    for i in range(len(inputs)):
+        output = results[i]
+        if isinstance(output, Exception):
+            raise output
+        outputs.append(output)
+    return outputs
+
+
+###########################################################################
+# Adapted from Synchronized-BatchNorm-PyTorch.
+# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
+#
+class CallbackContext(object):
+    pass
+
+
+def execute_replication_callbacks(modules):
+    """
+    Execute an replication callback `__data_parallel_replicate__` on each module created
+    by original replication.
+
+    The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)`
+
+    Note that, as all modules are isomorphism, we assign each sub-module with a context
+    (shared among multiple copies of this module on different devices).
+    Through this context, different copies can share some information.
+
+    We guarantee that the callback on the master copy (the first copy) will be called ahead
+    of calling the callback of any slave copies.
+    """
+    master_copy = modules[0]
+    nr_modules = len(list(master_copy.modules()))
+    ctxs = [CallbackContext() for _ in range(nr_modules)]
+
+    for i, module in enumerate(modules):
+        for j, m in enumerate(module.modules()):
+            if hasattr(m, '__data_parallel_replicate__'):
+                m.__data_parallel_replicate__(ctxs[j], i)
+
+
+def patch_replication_callback(data_parallel):
+    """
+    Monkey-patch an existing `DataParallel` object. Add the replication callback.
+    Useful when you have customized `DataParallel` implementation.
+
+    Examples:
+        > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False)
+        > sync_bn = DataParallel(sync_bn, device_ids=[0, 1])
+        > patch_replication_callback(sync_bn)
+        # this is equivalent to
+        > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False)
+        > sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1])
+    """
+
+    assert isinstance(data_parallel, DataParallel)
+
+    old_replicate = data_parallel.replicate
+
+    @functools.wraps(old_replicate)
+    def new_replicate(module, device_ids):
+        modules = old_replicate(module, device_ids)
+        execute_replication_callbacks(modules)
+        return modules
+
+    data_parallel.replicate = new_replicate
+    
diff --git a/parsing/dml_csr/utils/logging.py b/parsing/dml_csr/utils/logging.py
new file mode 100644
index 000000000..c89b6f0cc
--- /dev/null
+++ b/parsing/dml_csr/utils/logging.py
@@ -0,0 +1,135 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   logging.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import logging
+import torch.distributed as dist
+
+logger_initialized = {}
+
+
+def get_root_logger(log_file=None, log_level=logging.INFO):
+    """Get root logger.
+
+    Args:
+        log_file (str, optional): File path of log. Defaults to None.
+        log_level (int, optional): The level of logger.
+            Defaults to logging.INFO.
+
+    Returns:
+        :obj:`logging.Logger`: The obtained logger
+    """
+    logger = get_logger(name='face_parsing', log_file=log_file, log_level=log_level)
+
+    return logger
+
+
+def get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):
+    """Initialize and get a logger by name.
+    If the logger has not been initialized, this method will initialize the
+    logger by adding one or two handlers, otherwise the initialized logger will
+    be directly returned. During initialization, a StreamHandler will always be
+    added. If `log_file` is specified and the process rank is 0, a FileHandler
+    will also be added.
+    Args:
+        name (str): Logger name.
+        log_file (str | None): The log filename. If specified, a FileHandler
+            will be added to the logger.
+        log_level (int): The logger level. Note that only the process of
+            rank 0 is affected, and other processes will set the level to
+            "Error" thus be silent most of the time.
+        file_mode (str): The file mode used in opening log file.
+            Defaults to 'w'.
+    Returns:
+        logging.Logger: The expected logger.
+    """
+    logger = logging.getLogger(name)
+    if name in logger_initialized:
+        return logger
+    # handle hierarchical names
+    # e.g., logger "a" is initialized, then logger "a.b" will skip the
+    # initialization since it is a child of "a".
+    for logger_name in logger_initialized:
+        if name.startswith(logger_name):
+            return logger
+
+    # handle duplicate logs to the console
+    # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler <stderr> (NOTSET)
+    # to the root logger. As logger.propagate is True by default, this root
+    # level handler causes logging messages from rank>0 processes to
+    # unexpectedly show up on the console, creating much unwanted clutter.
+    # To fix this issue, we set the root logger's StreamHandler, if any, to log
+    # at the ERROR level.
+    for handler in logger.root.handlers:
+        if type(handler) is logging.StreamHandler:
+            handler.setLevel(logging.ERROR)
+
+    stream_handler = logging.StreamHandler()
+    handlers = [stream_handler]
+
+    if dist.is_available() and dist.is_initialized():
+        rank = dist.get_rank()
+    else:
+        rank = 0
+
+    # only rank 0 will add a FileHandler
+    if rank == 0 and log_file is not None:
+        # Here, the default behaviour of the official logger is 'a'. Thus, we
+        # provide an interface to change the file mode to the default
+        # behaviour.
+        file_handler = logging.FileHandler(log_file, file_mode)
+        handlers.append(file_handler)
+
+    formatter = logging.Formatter(
+        '%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+    for handler in handlers:
+        handler.setFormatter(formatter)
+        handler.setLevel(log_level)
+        logger.addHandler(handler)
+
+    if rank == 0:
+        logger.setLevel(log_level)
+    else:
+        logger.setLevel(logging.ERROR)
+
+    logger_initialized[name] = True
+
+    return logger
+
+
+def print_log(msg, logger=None, level=logging.INFO):
+    """Print a log message.
+    Args:
+        msg (str): The message to be logged.
+        logger (logging.Logger | str | None): The logger to be used.
+            Some special loggers are:
+            - "silent": no message will be printed.
+            - other str: the logger obtained with `get_root_logger(logger)`.
+            - None: The `print()` method will be used to print log messages.
+        level (int): Logging level. Only available when `logger` is a Logger
+            object or "root".
+    """
+    if logger is None:
+        print(msg)
+    elif isinstance(logger, logging.Logger):
+        logger.log(level, msg)
+    elif logger == 'silent':
+        pass
+    elif isinstance(logger, str):
+        _logger = get_logger(logger)
+        _logger.log(level, msg)
+    else:
+        raise TypeError(
+            'logger should be either a logging.Logger object, str, '
+            f'"silent" or None, but got {type(logger)}')
diff --git a/parsing/dml_csr/utils/miou.py b/parsing/dml_csr/utils/miou.py
new file mode 100644
index 000000000..373936095
--- /dev/null
+++ b/parsing/dml_csr/utils/miou.py
@@ -0,0 +1,242 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   miou.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import cv2
+import json
+import numpy as np
+import os
+
+from collections import OrderedDict
+from PIL import Image as PILImage
+from utils.transforms import transform_parsing
+
+
+LABELS = ['background', 'skin', 'nose', 'eye_g', 'l_eye', 'r_eye', \
+        'l_brow', 'r_brow', 'l_ear', 'r_ear', 'mouth', 'u_lip', \
+        'l_lip', 'hair', 'hat', 'ear_r', 'neck_l', 'neck', 'cloth']
+
+
+def get_palette(num_cls):
+    """ Returns the color map for visualizing the segmentation mask.
+    Args:
+        num_cls: Number of classes
+    Returns:
+        The color map
+    """
+
+    n = num_cls
+    palette = [0] * (n * 3)
+    for j in range(0, n):
+        lab = j
+        palette[j * 3 + 0] = 0
+        palette[j * 3 + 1] = 0
+        palette[j * 3 + 2] = 0
+        i = 0
+        while lab:
+            palette[j * 3 + 0] |= (((lab >> 0) & 1) << (7 - i))
+            palette[j * 3 + 1] |= (((lab >> 1) & 1) << (7 - i))
+            palette[j * 3 + 2] |= (((lab >> 2) & 1) << (7 - i))
+            i += 1
+            lab >>= 3
+    return palette
+
+
+def get_confusion_matrix(gt_label, pred_label, num_classes):
+    """
+    Calcute the confusion matrix by given label and pred
+    :param gt_label: the ground truth label
+    :param pred_label: the pred label
+    :param num_classes: the nunber of class
+    :return: the confusion matrix
+    """
+    index = (gt_label * num_classes + pred_label).astype('int32')
+    label_count = np.bincount(index)
+    confusion_matrix = np.zeros((num_classes, num_classes))
+
+    for i_label in range(num_classes):
+        for i_pred_label in range(num_classes):
+            cur_index = i_label * num_classes + i_pred_label
+            if cur_index < len(label_count):
+                confusion_matrix[i_label, i_pred_label] = label_count[cur_index]
+
+    return confusion_matrix
+
+
+def fast_histogram(a, b, na, nb):
+    '''
+    fast histogram calculation
+    ---
+    * a, b: non negative label ids, a.shape == b.shape, a in [0, ... na-1], b in [0, ..., nb-1]
+    '''
+    assert a.shape == b.shape
+    assert np.all((a >= 0) & (a < na) & (b >= 0) & (b < nb))
+    # k = (a >= 0) & (a < na) & (b >= 0) & (b < nb)
+    hist = np.bincount(
+        nb * a.reshape([-1]).astype(int) + b.reshape([-1]).astype(int),
+        minlength=na * nb).reshape(na, nb)
+    assert np.sum(hist) == a.size
+    return hist
+
+
+def _read_names(file_name):
+    label_names = []
+    for name in open(file_name, 'r'):
+        name = name.strip()
+        if len(name) > 0:
+            label_names.append(name)
+    return label_names
+
+
+def _merge(*list_pairs):
+    a = []
+    b = []
+    for al, bl in list_pairs:
+        a += al
+        b += bl
+    return a, b
+
+
+def compute_mean_ioU(preds, scales, centers, num_classes, datadir, input_size=[473, 473], dataset='val', reverse=False):
+    file_list_name = os.path.join(datadir, dataset + '_list.txt')
+    val_id = [line.split()[0][7:-4] for line in open(file_list_name).readlines()]
+
+    confusion_matrix = np.zeros((num_classes, num_classes))
+
+    label_names_file = os.path.join(datadir, 'label_names.txt')
+    gt_label_names = pred_label_names = _read_names(label_names_file)
+
+    assert gt_label_names[0] == pred_label_names[0] == 'bg'
+
+    hists = []
+    for i, im_name in enumerate(val_id):
+        gt_path = os.path.join(datadir, dataset, 'labels', im_name + '.png')
+        gt = cv2.imread(gt_path, cv2.IMREAD_GRAYSCALE)
+        h, w = gt.shape
+        pred_out = preds[i]
+        if scales is not None:
+            s = scales[i]
+            c = centers[i]
+        else:
+            s = None
+            c = None
+        pred_old = transform_parsing(pred_out, c, s, w, h, input_size)
+        gt = np.asarray(gt, dtype=np.int32)
+        pred = np.asarray(pred_old, dtype=np.int32)
+        ignore_index = gt != 255
+
+        gt = gt[ignore_index]
+        pred = pred[ignore_index]
+
+        hist = fast_histogram(gt, pred, len(gt_label_names), len(pred_label_names))
+        hists.append(hist)
+
+        confusion_matrix += get_confusion_matrix(gt, pred, num_classes)
+
+    hist_sum = np.sum(np.stack(hists, axis=0), axis=0)
+
+    eval_names = dict()
+    for label_name in gt_label_names:
+        gt_ind = gt_label_names.index(label_name)
+        pred_ind = pred_label_names.index(label_name)
+        eval_names[label_name] = ([gt_ind], [pred_ind])
+        
+    if 'le' in eval_names and 're' in eval_names:
+        eval_names['eyes'] = _merge(eval_names['le'], eval_names['re'])
+    if 'lb' in eval_names and 'rb' in eval_names:
+        eval_names['brows'] = _merge(eval_names['lb'], eval_names['rb'])
+    if 'ulip' in eval_names and 'imouth' in eval_names and 'llip' in eval_names:
+        eval_names['mouth'] = _merge(
+            eval_names['ulip'], eval_names['imouth'], eval_names['llip'])
+    
+    # Helen 
+    if 'eyes' in eval_names and 'brows' in eval_names and 'nose' in eval_names and 'mouth' in eval_names:
+        eval_names['overall'] = _merge(
+            eval_names['eyes'], eval_names['brows'], eval_names['nose'], eval_names['mouth'])
+
+    pos = confusion_matrix.sum(1)
+    res = confusion_matrix.sum(0)
+    tp = np.diag(confusion_matrix)
+
+    pixel_accuracy = (tp.sum() / pos.sum()) * 100
+    mean_accuracy = ((tp / np.maximum(1.0, pos)).mean()) * 100
+    IoU_array = (tp / np.maximum(1.0, pos + res - tp))
+    IoU_array = IoU_array * 100
+    mean_IoU = IoU_array.mean()
+    print('Pixel accuracy: %f \n' % pixel_accuracy)
+    print('Mean accuracy: %f \n' % mean_accuracy)
+    print('Mean IU: %f \n' % mean_IoU)
+    mIoU_value = []
+    f1_value = []
+    mf1_value = []
+
+    for i, (label, iou) in enumerate(zip(LABELS, IoU_array)):
+        mIoU_value.append((label, iou))
+
+    mIoU_value.append(('Pixel accuracy', pixel_accuracy))
+    mIoU_value.append(('Mean accuracy', mean_accuracy))
+    mIoU_value.append(('Mean IU', mean_IoU))
+    mIoU_value = OrderedDict(mIoU_value)
+
+    for eval_name, (gt_inds, pred_inds) in eval_names.items():
+        A = hist_sum[gt_inds, :].sum()
+        B = hist_sum[:, pred_inds].sum()
+        intersected = hist_sum[gt_inds, :][:, pred_inds].sum()
+        f1 = 2 * intersected / (A + B)
+
+        if eval_name in gt_label_names[1:]:
+            mf1_value.append(f1)
+        f1_value.append((eval_name, f1))
+
+    f1_value.append(('Mean_F1', np.array(mf1_value).mean()))
+    f1_value = OrderedDict(f1_value)
+
+    return mIoU_value, f1_value
+
+
+def write_results(preds, scales, centers, datadir, dataset, result_dir, input_size=[473, 473]):
+    palette = get_palette(20)
+    if not os.path.exists(result_dir):
+        os.makedirs(result_dir)
+
+    json_file = os.path.join(datadir, 'annotations', dataset + '.json')
+    with open(json_file) as data_file:
+        data_list = json.load(data_file)
+        data_list = data_list['root']
+    for item, pred_out, s, c in zip(data_list, preds, scales, centers):
+        im_name = item['im_name']
+        w = item['img_width']
+        h = item['img_height']
+        pred = transform_parsing(pred_out, c, s, w, h, input_size)
+
+        save_path = os.path.join(result_dir, im_name[:-4]+'.png')
+        output_im = PILImage.fromarray(np.asarray(pred, dtype=np.uint8))
+        output_im.putpalette(palette)
+        output_im.save(save_path)
+
+
+def get_arguments():
+    """Parse all the arguments provided from the CLI.
+
+    Returns:
+      A list of parsed arguments.
+    """
+    parser = argparse.ArgumentParser(description="DeepLabLFOV NetworkEv")
+    parser.add_argument("--pred-path", type=str, default='',
+                        help="Path to predicted segmentation.")
+    parser.add_argument("--gt-path", type=str, default='',
+                        help="Path to the groundtruth dir.")
+
+    return parser.parse_args()
diff --git a/parsing/dml_csr/utils/schp.py b/parsing/dml_csr/utils/schp.py
new file mode 100644
index 000000000..1f6f58f19
--- /dev/null
+++ b/parsing/dml_csr/utils/schp.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   schp.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import torch
+
+
+def moving_average(net1, net2, alpha=1):
+    for param1, param2 in zip(net1.parameters(), net2.parameters()):
+        param1.data *= (1.0 - alpha)
+        param1.data += param2.data * alpha
+
+
+def _check_bn(module, flag):
+    classname = module.__class__.__name__
+    if classname.find('BatchNorm') != -1 or classname.find('InPlaceABNSync') != -1:
+        flag[0] = True
+
+
+def check_bn(model):
+    flag = [False]
+    model.apply(lambda module: _check_bn(module, flag))
+    return flag[0]
+
+
+def reset_bn(module):
+    classname = module.__class__.__name__
+    if classname.find('BatchNorm') != -1 or classname.find('InPlaceABNSync') != -1:
+        module.running_mean = torch.zeros_like(module.running_mean)
+        module.running_var = torch.ones_like(module.running_var)
+
+
+def _get_momenta(module, momenta):
+    classname = module.__class__.__name__
+    if classname.find('BatchNorm') != -1 or classname.find('InPlaceABNSync') != -1:
+        momenta[module] = module.momentum
+
+
+def _set_momenta(module, momenta):
+    classname = module.__class__.__name__
+    if classname.find('BatchNorm') != -1 or classname.find('InPlaceABNSync') != -1:
+        module.momentum = momenta[module]
+
+
+def bn_re_estimate(loader, model):
+    if not check_bn(model):
+        print('No batch norm layer detected')
+        return
+    model.train()
+    momenta = {}
+    model.apply(reset_bn)
+    model.apply(lambda module: _get_momenta(module, momenta))
+    n = 0
+    for i_iter, batch in enumerate(loader):
+        # images, labels, edges, _ = batch
+        images = batch[0]
+        b = images.data.size(0)
+        momentum = b / (n + b)
+        for module in momenta.keys():
+            module.momentum = momentum
+        model(images)
+        n += b
+    model.apply(lambda module: _set_momenta(module, momenta))
+
+
+def save_schp_checkpoint(states, is_best_parsing, output_dir, filename='schp_checkpoint.pth.tar'):
+    save_path = os.path.join(output_dir, filename)
+    # if os.path.exists(save_path):
+    #     os.remove(save_path)
+    torch.save(states, save_path)
+    if is_best_parsing and 'state_dict' in states:
+        best_save_path = os.path.join(output_dir, 'model_parsing_best.pth.tar')
+        if os.path.exists(best_save_path):
+            os.remove(best_save_path)
+        torch.save(states, best_save_path)
diff --git a/parsing/dml_csr/utils/transforms.py b/parsing/dml_csr/utils/transforms.py
new file mode 100644
index 000000000..73ac2e99d
--- /dev/null
+++ b/parsing/dml_csr/utils/transforms.py
@@ -0,0 +1,122 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   transforms.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+import numpy as np
+import cv2
+
+
+def flip_back(output_flipped, matched_parts):
+    '''
+    ouput_flipped: numpy.ndarray(batch_size, num_joints, height, width)
+    '''
+    assert output_flipped.ndim == 4,\
+        'output_flipped should be [batch_size, num_joints, height, width]'
+
+    output_flipped = output_flipped[:, :, :, ::-1]
+
+    for pair in matched_parts:
+        tmp = output_flipped[:, pair[0], :, :].copy()
+        output_flipped[:, pair[0], :, :] = output_flipped[:, pair[1], :, :]
+        output_flipped[:, pair[1], :, :] = tmp
+
+    return output_flipped
+  
+
+def transform_parsing(pred, center, scale, width, height, input_size):
+
+    if center is not None:
+        trans = get_affine_transform(center, scale, 0, input_size, inv=1)
+        target_pred = cv2.warpAffine(
+                pred,
+                trans,
+                (int(width), int(height)), #(int(width), int(height)),
+                flags=cv2.INTER_NEAREST,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=(0))
+    else:
+        target_pred = cv2.resize(pred, (int(width), int(height)), interpolation=cv2.INTER_NEAREST)
+
+    return target_pred
+
+
+def get_affine_transform(center,
+                         scale,
+                         rot,
+                         output_size,
+                         shift=np.array([0, 0], dtype=np.float32),
+                         inv=0):
+    if not isinstance(scale, np.ndarray) and not isinstance(scale, list):
+        print(scale)
+        scale = np.array([scale, scale])
+
+    scale_tmp = scale
+
+    src_w = scale_tmp[0]
+    dst_w = output_size[1]
+    dst_h = output_size[0]
+
+    rot_rad = np.pi * rot / 180
+    src_dir = get_dir([0, src_w * -0.5], rot_rad)
+    dst_dir = np.array([0, dst_w * -0.5], np.float32)
+
+    src = np.zeros((3, 2), dtype=np.float32)
+    dst = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = center + scale_tmp * shift
+    src[1, :] = center + src_dir + scale_tmp * shift
+    dst[0, :] = [dst_w * 0.5, dst_h * 0.5]
+    dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir
+
+    src[2:, :] = get_3rd_point(src[0, :], src[1, :])
+    dst[2:, :] = get_3rd_point(dst[0, :], dst[1, :])
+
+    if inv:
+        trans = cv2.getAffineTransform(np.float32(dst), np.float32(src))
+    else:
+        trans = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return trans
+
+
+def affine_transform(pt, t):
+    new_pt = np.array([pt[0], pt[1], 1.]).T
+    new_pt = np.dot(t, new_pt)
+    return new_pt[:2]
+
+
+def get_3rd_point(a, b):
+    direct = a - b
+    return b + np.array([-direct[1], direct[0]], dtype=np.float32)
+
+
+def get_dir(src_point, rot_rad):
+    sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+
+    src_result = [0, 0]
+    src_result[0] = src_point[0] * cs - src_point[1] * sn
+    src_result[1] = src_point[0] * sn + src_point[1] * cs
+
+    return src_result
+
+
+def crop(img, center, scale, output_size, rot=0):
+    trans = get_affine_transform(center, scale, rot, output_size)
+
+    dst_img = cv2.warpAffine(img,
+                             trans,
+                             (int(output_size[1]), int(output_size[0])),
+                             flags=cv2.INTER_LINEAR)
+
+    return dst_img
diff --git a/parsing/dml_csr/utils/utils.py b/parsing/dml_csr/utils/utils.py
new file mode 100644
index 000000000..322c53d2b
--- /dev/null
+++ b/parsing/dml_csr/utils/utils.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   utils.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import torchvision
+import torch
+
+from PIL import Image
+from torch import nn
+
+# colour map
+COLORS = [(0,0,0)
+                # 0=background
+                ,(128,0,0),(0,128,0),(128,128,0),(0,0,128),(128,0,128)
+                # 1=aeroplane, 2=bicycle, 3=bird, 4=boat, 5=bottle
+                ,(0,128,128),(128,128,128),(64,0,0),(192,0,0),(64,128,0)
+                # 6=bus, 7=car, 8=cat, 9=chair, 10=cow
+                ,(192,128,0),(64,0,128),(192,0,128),(64,128,128),(192,128,128)
+                # 11=diningtable, 12=dog, 13=horse, 14=motorbike, 15=person
+                ,(0,64,0),(128,64,0),(0,192,0),(128,192,0),(0,64,128)]
+                # 16=potted plant, 17=sheep, 18=sofa, 19=train, 20=tv/monitor
+
+
+def decode_parsing(labels, num_images=1, num_classes=21, is_pred=False):
+    """Decode batch of segmentation masks.
+    
+    Args:
+      mask: result of inference after taking argmax.
+      num_images: number of images to decode from the batch.
+      num_classes: number of classes to predict (including background).
+    
+    Returns:
+      A batch with num_images RGB images of the same size as the input. 
+    """
+    pred_labels = labels[:num_images].clone().cpu().data
+    if is_pred:
+        pred_labels = torch.argmax(pred_labels, dim=1)
+    n, h, w = pred_labels.size()
+
+    labels_color = torch.zeros([n, 3, h, w], dtype=torch.uint8)
+    for i, c in enumerate(COLORS):
+        c0 = labels_color[:, 0, :, :]
+        c1 = labels_color[:, 1, :, :]
+        c2 = labels_color[:, 2, :, :]
+
+        c0[pred_labels == i] = c[0]
+        c1[pred_labels == i] = c[1]
+        c2[pred_labels == i] = c[2]
+
+    return labels_color
+
+
+def inv_preprocess(imgs, num_images):
+    """Inverse preprocessing of the batch of images.
+       Add the mean vector and convert from BGR to RGB.
+       
+    Args:
+      imgs: batch of input images.
+      num_images: number of images to apply the inverse transformations on.
+      img_mean: vector of mean colour values.
+  
+    Returns:
+      The batch of the size num_images with the same spatial dimensions as the input.
+    """
+    rev_imgs = imgs[:num_images].clone().cpu().data
+    rev_normalize = NormalizeInverse(mean=[0.485, 0.456, 0.406],
+                                     std=[0.229, 0.224, 0.225])
+    for i in range(num_images):
+        rev_imgs[i] = rev_normalize(rev_imgs[i])
+
+    return rev_imgs
+
+
+class NormalizeInverse(torchvision.transforms.Normalize):
+    """
+    Undoes the normalization and returns the reconstructed images in the input domain.
+    """
+
+    def __init__(self, mean, std):
+        mean = torch.as_tensor(mean)
+        std = torch.as_tensor(std)
+        std_inv = 1 / (std + 1e-7)
+        mean_inv = -mean * std_inv
+        super().__init__(mean=mean_inv, std=std_inv)
+
+
+class SingleGPU(nn.Module):
+    def __init__(self, module):
+        super(SingleGPU, self).__init__()
+        self.module = module
+
+    def forward(self, x):
+        return self.module(x.cuda(non_blocking=True))
+      
diff --git a/parsing/dml_csr/utils/warmup_scheduler.py b/parsing/dml_csr/utils/warmup_scheduler.py
new file mode 100644
index 000000000..a4925bac8
--- /dev/null
+++ b/parsing/dml_csr/utils/warmup_scheduler.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+"""
+@Author     :   Qingping Zheng
+@Contact    :   qingpingzheng2014@gmail.com
+@File       :   warmup_scheduler.py
+@Time       :   10/01/21 00:00 PM
+@Desc       :   
+@License    :   Licensed under the Apache License, Version 2.0 (the "License"); 
+@Copyright  :   Copyright 2022 The Authors. All Rights Reserved.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+from torch.optim.lr_scheduler import _LRScheduler
+
+
+class GradualWarmupScheduler(_LRScheduler):
+    """ Gradually warm-up learning rate with cosine annealing in optimizer.
+    Proposed in 'Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour'.
+    """
+
+    def __init__(self, optimizer, total_epoch, eta_min=0, warmup_epoch=10, last_epoch=-1):
+        self.total_epoch = total_epoch
+        self.eta_min = eta_min
+        self.warmup_epoch = warmup_epoch
+        super(GradualWarmupScheduler, self).__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if self.last_epoch <= self.warmup_epoch:
+            return [self.eta_min + self.last_epoch*(base_lr - self.eta_min)/self.warmup_epoch for base_lr in self.base_lrs]
+        else:
+            return [self.eta_min + (base_lr-self.eta_min)*(1+math.cos(math.pi*(self.last_epoch-self.warmup_epoch)/(self.total_epoch-self.warmup_epoch))) / 2 for base_lr in self.base_lrs]
+
+
+class SGDRScheduler(_LRScheduler):
+    """ Consine annealing with warm up and restarts.
+    Proposed in `SGDR: Stochastic Gradient Descent with Warm Restarts`.
+    """
+    def __init__(self, optimizer, total_epoch=150, start_cyclical=100, cyclical_base_lr=7e-4, cyclical_epoch=10, eta_min=0, warmup_epoch=10, last_epoch=-1):
+        self.total_epoch = total_epoch
+        self.start_cyclical = start_cyclical
+        self.cyclical_epoch = cyclical_epoch
+        self.cyclical_base_lr = cyclical_base_lr
+        self.eta_min = eta_min
+        self.warmup_epoch = warmup_epoch
+        super(SGDRScheduler, self).__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if self.last_epoch < self.warmup_epoch:
+            return [self.eta_min + self.last_epoch*(base_lr - self.eta_min)/self.warmup_epoch for base_lr in self.base_lrs]
+        elif self.last_epoch < self.start_cyclical:
+            return [self.eta_min + (base_lr-self.eta_min)*(1+math.cos(math.pi*(self.last_epoch-self.warmup_epoch)/(self.start_cyclical-self.warmup_epoch))) / 2 for base_lr in self.base_lrs]
+        else:
+            return [self.eta_min + (self.cyclical_base_lr-self.eta_min)*(1+math.cos(math.pi* ((self.last_epoch-self.start_cyclical)% self.cyclical_epoch)/self.cyclical_epoch)) / 2 for base_lr in self.base_lrs]
+
+
+if __name__ == '__main__':
+    import matplotlib.pyplot as plt
+    import torch
+    model = torch.nn.Linear(10, 2)
+    optimizer = torch.optim.SGD(params=model.parameters(), lr=7e-3, momentum=0.9, weight_decay=5e-4)
+    scheduler_warmup = SGDRScheduler(optimizer, total_epoch=150, eta_min=7e-5, warmup_epoch=10, start_cyclical=100, cyclical_base_lr=3.5e-3, cyclical_epoch=10)
+    lr = []
+    for epoch in range(0,150):
+        scheduler_warmup.step(epoch)
+        lr.append(scheduler_warmup.get_lr())
+    plt.style.use('ggplot')
+    plt.plot(list(range(0,150)), lr)
+    plt.show()
diff --git a/python-package/README.md b/python-package/README.md
index 08ce60754..4ac08c329 100644
--- a/python-package/README.md
+++ b/python-package/README.md
@@ -1,3 +1,267 @@
-InsightFace.ai README
+# InsightFace Python Library
 
+## License
 
+The code of InsightFace Python Library is released under the MIT License. There is no limitation for both academic and commercial usage.
+
+**The pretrained models we provided with this library are available for non-commercial research purposes only, including both auto-downloading models and manual-downloading models.**
+
+## Install
+
+### Install Inference Backend
+
+For ``insightface<=0.1.5``, we use MXNet as inference backend.
+
+Starting from insightface>=0.2, we use onnxruntime as inference backend.
+
+You have to install ``onnxruntime-gpu`` manually to enable GPU inference, or install ``onnxruntime`` to use CPU only inference.
+
+### Install InsightFace Evaluation Studio GUI
+
+InsightFace 1.0.1 includes a local desktop GUI:
+
+```
+pip install "insightface[gui]"
+insightface-gui
+```
+
+Development install:
+
+```
+cd python-package
+pip install -e ".[gui]"
+insightface-gui
+```
+
+Equivalent launch commands:
+
+```
+insightface-eval-studio
+insightface-desktop
+python -m insightface.gui
+```
+
+The GUI is called **InsightFace Evaluation Studio**. It provides local 1:1 face
+compare, People Library management, 1:N face search, multi-face photo
+recognition, batch folder processing, album people clustering, enterprise
+evaluation reports, and a face swap entry point. User images, videos,
+embeddings, databases, and reports are stored locally by default under
+``~/.insightface/gui`` and are not uploaded automatically.
+Image and video previews are clickable upload targets: click
+``Click to upload or drag a file here`` or drop a file onto the preview. The
+preview changes color on hover and during drag-over. Loaded previews show a
+small delete button and can be replaced by dragging in another file.
+
+The desktop app uses mode-based navigation. Choose **Face Recognition**,
+**Album Management**, **Face Swap**, or **Enterprise Evaluation** from the
+persistent **Workflows** rail on the left side of the window. Face Recognition
+is a single **Query & Gallery** workspace: upload
+one query image and one gallery image for 1:1 compare, or upload multiple
+gallery images / a folder for 1:N gallery search. Album Management uses a
+single **Album** workspace for adding one or more folders, refreshing new
+images, DBSCAN clustering with a default cosine similarity threshold of `0.48`,
+and reviewing original photo thumbnails. Album directories and clustering results are saved
+locally for the next launch. Enterprise Evaluation is a single workspace for
+local 1:1 and 1:N identity-folder evaluation, Auto Split, metrics, and PDF
+report export. Enterprise datasets must pass validation before evaluation; the
+validator checks folder layout, gallery/probe rules, and the selected
+multi-face handling policy.
+Global utilities are available from the top bar and **Tools** menu:
+**Settings**, **Models**, and **License**. **Settings** controls the UI theme
+and language. Language defaults to the operating system when it is supported,
+otherwise English. Supported GUI languages are English, Chinese, Japanese,
+Korean, Spanish, French, German, Portuguese, and Russian. Available themes
+include System, Precision Light, Studio Dark, Graphite Pro, Azure Lab, Emerald
+Focus, and Crimson Audit. Workspace paths are chosen on first launch and are
+not changed from the settings dialog.
+
+Models are not downloaded automatically by the GUI. Open **Models > Downloads**,
+click **Refresh Download URLs** to read the latest GitHub Releases
+asset URLs, then explicitly download the selected package. Downloaded zip files
+are cached under ``~/.insightface/gui/cache/models`` and extracted under
+``~/.insightface/models/<model_name>/``.
+The Downloads tab also lists GFPGANv1.4 as a third-party face restoration
+model. After it is downloaded, enable **GFPGAN post-processing** in
+**Models > Runtime** to run 512x512 GFPGAN restoration after face swap.
+Detection size defaults to **Auto**, which runs joint 128x128 and 640x640
+detection. Face swap models are selected in **Models > Runtime** from already
+downloaded swap models only; the Face Swap workspace loads the configured swap
+model only when a swap is run.
+
+### Optional face3d Build
+
+InsightFace 1.0.1 does not build the optional ``face3d`` Cython/C++ extension by
+default. This keeps the default install lighter and avoids local compiler
+requirements. Users who need the legacy mask renderer / face3d path can opt in:
+
+```
+pip install -e ".[face3d]" --no-build-isolation --config-settings editable_mode=compat
+python setup.py build_ext --inplace --with-face3d
+```
+
+The same build can also be enabled with:
+
+```
+INSIGHTFACE_WITH_FACE3D=1 python setup.py build_ext --inplace
+```
+
+More details:
+
+- ``docs/gui.md``
+- ``docs/commercial_evaluation.md``
+- ``docs/gui_packaging.md``
+
+## Change Log
+
+### [1.0.1] - 2026-05-23
+
+#### Changed
+
+- Remove the PyPI package metadata license classifier field while keeping the README license guidance.
+- Move direct `Pillow` and `scikit-learn` requirements to the GUI extra, and `matplotlib` to the optional `face3d` extra.
+- Remove unused base dependencies on `easydict` and `prettytable`.
+
+### [1.0] - 2026-05-23
+
+#### Added
+
+- Add **InsightFace Evaluation Studio**, a cross-platform PySide6 desktop GUI for local face recognition, album grouping, enterprise evaluation/report export, and face swap trials.
+- Add GUI launch commands: ``insightface-gui``, ``insightface-eval-studio``, ``insightface-desktop``, and ``python -m insightface.gui``.
+
+#### Changed
+
+- Default ``FaceAnalysis.prepare()`` detection size is now Auto, running SCRFD at both 128x128 and 640x640 before unified NMS.
+- Route detection models loaded by ``model_zoo.get_model()`` through ``SCRFD`` by default.
+- The optional ``face3d`` Cython/C++ extension is no longer built by default; use ``--with-face3d`` or ``INSIGHTFACE_WITH_FACE3D=1`` to opt in.
+
+### [0.7.1] - 2022-12-14
+  
+#### Changed
+  
+- Change model downloading provider to cloudfront.
+
+### [0.7] - 2022-11-28
+  
+#### Added
+
+- Add face swapping model and example.
+ 
+#### Changed
+  
+- Set default ORT provider to CUDA and CPU.
+ 
+### [0.6] - 2022-01-29
+  
+#### Added
+
+- Add pose estimation in face-analysis app.
+ 
+#### Changed
+  
+- Change model automated downloading url, to ucloud.
+ 
+
+## Quick Example
+
+```
+import cv2
+import numpy as np
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+app = FaceAnalysis(providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+app.prepare(ctx_id=0)  # Auto detection size: 128x128 + 640x640
+img = ins_get_image('t1')
+faces = app.get(img)
+rimg = app.draw_on(img, faces)
+cv2.imwrite("./t1_output.jpg", rimg)
+```
+
+This quick example will detect faces from the ``t1.jpg`` image and draw detection results on it.
+
+
+
+## Model Zoo
+
+In the latest version of insightface library, we provide following model packs:
+
+Name in **bold** is the default model pack. **Auto** means we can download the model pack through the python library directly.
+
+Once you manually downloaded the zip model pack, unzip it under `~/.insightface/models/` first before you call the program.
+
+| Name          | Detection Model | Recognition Model    | Alignment    | Attributes | Model-Size | Link                                                         | Auto |
+| ------------- | --------------- | -------------------- | ------------ | ---------- | ---------- | ------------------------------------------------------------ | ------------- |
+| antelopev2    | SCRFD-10GF      | ResNet100@Glint360K  | 2d106 & 3d68 | Gender&Age | 407MB      | [link](https://drive.google.com/file/d/18wEUfMNohBJ4K3Ly5wpTejPfDzp-8fI8/view?usp=sharing) | N             |
+| **buffalo_l** | SCRFD-10GF      | ResNet50@WebFace600K | 2d106 & 3d68 | Gender&Age | 326MB      | [link](https://drive.google.com/file/d/1qXsQJ8ZT42_xSmWIYy85IcidpiZudOCB/view?usp=sharing) | Y             |
+| buffalo_m     | SCRFD-2.5GF     | ResNet50@WebFace600K | 2d106 & 3d68 | Gender&Age | 313MB      | [link](https://drive.google.com/file/d/1net68yNxF33NNV6WP7k56FS6V53tq-64/view?usp=sharing) | N             |
+| buffalo_s     | SCRFD-500MF     | MBF@WebFace600K      | 2d106 & 3d68 | Gender&Age | 159MB      | [link](https://drive.google.com/file/d/1pKIusApEfoHKDjeBTXYB3yOQ0EtTonNE/view?usp=sharing) | N             |
+| buffalo_sc    | SCRFD-500MF     | MBF@WebFace600K      | -            | -          | 16MB       | [link](https://drive.google.com/file/d/19I-MZdctYKmVf3nu5Da3HS6KH5LBfdzG/view?usp=sharing) | N             |
+
+
+
+Recognition Accuracy:
+
+| Name      | MR-ALL | African | Caucasian | South Asian | East Asian | LFW   | CFP-FP | AgeDB-30 | IJB-C(E4) |
+| :-------- | ------ | ------- | --------- | ----------- | ---------- | ----- | ------ | -------- | --------- |
+| buffalo_l | 91.25  | 90.29   | 94.70     | 93.16       | 74.96      | 99.83 | 99.33  | 98.23    | 97.25     |
+| buffalo_s | 71.87  | 69.45   | 80.45     | 73.39       | 51.03      | 99.70 | 98.00  | 96.58    | 95.02     |
+
+*buffalo_m has the same accuracy with buffalo_l.*
+
+*buffalo_sc has the same accuracy with buffalo_s.*
+
+
+
+**Note that these models are available for non-commercial research purposes only.**
+
+
+
+For insightface>=0.3.3, models will be downloaded automatically once we init ``app = FaceAnalysis()`` instance.
+
+For insightface==0.3.2, you must first download the model package by command:
+
+```
+insightface-cli model.download buffalo_l
+```
+
+## Use Your Own Licensed Model
+
+You can simply create a new model directory under ``~/.insightface/models/`` and replace the pretrained models we provide with your own models. And then call ``app = FaceAnalysis(name='your_model_zoo')`` to load these models.
+
+## Call Models
+
+The latest insightface libary only supports onnx models. Once you have trained detection or recognition models by PyTorch, MXNet or any other frameworks, you can convert it to the onnx format and then they can be called with insightface library.
+
+### Call Detection Models
+
+```
+import cv2
+import numpy as np
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+# Method-1, use FaceAnalysis
+app = FaceAnalysis(allowed_modules=['detection']) # enable detection model only
+app.prepare(ctx_id=0) # Auto detection size: 128x128 + 640x640
+
+# Method-2, load model directly
+detector = insightface.model_zoo.get_model('your_detection_model.onnx')
+detector.prepare(ctx_id=0) # SCRFD defaults to Auto: 128x128 + 640x640
+
+```
+
+### Call Recognition Models
+
+```
+import cv2
+import numpy as np
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+handler = insightface.model_zoo.get_model('your_recognition_model.onnx')
+handler.prepare(ctx_id=0)
+
+```
diff --git a/python-package/docs/commercial_evaluation.md b/python-package/docs/commercial_evaluation.md
new file mode 100644
index 000000000..42a9ebde9
--- /dev/null
+++ b/python-package/docs/commercial_evaluation.md
@@ -0,0 +1,132 @@
+# Commercial Evaluation Mode
+
+Enterprise Evaluation Mode helps teams evaluate InsightFace locally with their
+own data before choosing a commercial model license, private model evaluation,
+SDK/API access, SLA, or custom training path.
+
+All processing is local by default. No images, embeddings, videos, or reports
+are uploaded automatically.
+
+## 1:1 Verification
+
+Select **1:1 Verification**, choose an identity-folder root, and decide whether
+to enable **Auto Split**.
+
+Identity folders use one subfolder per person:
+
+```text
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+```
+
+With **Auto Split**, a file containing `gallery` is used as that identity's
+gallery image. If no such file exists, the first sorted image is used. All
+other images become probes, and every comparison is probe vs gallery across
+identities. Without Auto Split, every image is treated as a probe and the page
+runs full pairwise probe-vs-probe comparisons.
+
+The evaluation reports the best cosine threshold accuracy, the selected
+threshold, positive and negative pair counts, and TAR@FAR for `1e-6`, `1e-5`,
+`1e-4`, and `1e-3` with corresponding thresholds.
+
+## Dataset Validation And Multi-face Policy
+
+Run **Validate Dataset** before every evaluation. The GUI checks folder layout,
+Auto Split rules, gallery/probe availability, generated 1:1 positive and
+negative pairs, 1:N gallery coverage, image readability, and detected face
+counts.
+
+The **Multi-face handling** option controls images with more than one detected
+face:
+
+- **Require exactly one face** is the default and blocks the run when any image
+  contains multiple faces.
+- **Use largest face** keeps the image and uses the largest detected face.
+- **Use largest centered face** keeps the image and selects the face with the
+  best area-minus-center-distance score, favoring large faces near the image
+  center.
+- **Mark as skip** skips multi-face images. If a required gallery image is
+  skipped, validation fails because the gallery sample is unavailable.
+
+## 1:N Identification
+
+Without **Auto Split**, prepare:
+
+```text
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+```
+
+With **Auto Split**, prepare:
+
+```text
+dataset/
+  identities/
+    0001__Alice/
+      img001.jpg
+      img002.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+```
+
+1:N evaluation always requires gallery images. The report includes Top1 and
+TAR@FAR for `1e-5`, `1e-4`, `1e-3`, and `1e-2`, with corresponding thresholds.
+
+## Interpreting Metrics
+
+- FAR: false accept rate. Lower means fewer different-person pairs accepted.
+- FRR: false reject rate. Lower means fewer same-person pairs rejected.
+- TAR: true accept rate at a target FAR.
+- Top1: correct person is the first search result.
+
+Thresholds are business decisions. Higher thresholds usually reduce false
+accepts but increase false rejects.
+
+## Exporting Reports
+
+Reports are exported automatically after a run. Click **Export PDF** to choose
+a PDF destination. Reports include:
+
+1. Executive Summary
+2. Evaluation Scenario
+3. Dataset Summary
+4. Model and Runtime
+5. License Status
+6. Metrics
+7. Threshold Recommendation
+8. Error Analysis
+9. Latency and Hardware
+10. Deployment Considerations
+11. Responsible Use and Compliance Notice
+12. Commercial Licensing Next Steps
+13. Appendix: Raw Results
+
+Markdown, HTML, and PDF are supported when the GUI extra is installed.
+
+## Commercial License Notice
+
+This evaluation may use research or non-commercial model files. Production or
+commercial deployment requires an appropriate commercial model license. Please
+contact InsightFace for commercial model licensing, private model evaluation,
+SDK/API access, SLA, or custom training.
+
+This report does not provide legal advice. Users are responsible for consent,
+privacy, retention, and compliance with applicable biometric regulations.
diff --git a/python-package/docs/gui.md b/python-package/docs/gui.md
new file mode 100644
index 000000000..ebc5201b7
--- /dev/null
+++ b/python-package/docs/gui.md
@@ -0,0 +1,316 @@
+# InsightFace Evaluation Studio GUI
+
+InsightFace Evaluation Studio is a local desktop GUI for InsightFace 1.0.1. It is
+designed for no-code face recognition testing, local People Library management,
+album organization, enterprise model evaluation, report export, and basic face
+swap trials.
+
+## Install
+
+```bash
+cd python-package
+pip install -e ".[gui]"
+insightface-gui
+```
+
+PyPI users can install:
+
+```bash
+pip install "insightface[gui]"
+insightface-gui
+```
+
+Aliases:
+
+```bash
+insightface-eval-studio
+insightface-desktop
+python -m insightface.gui
+```
+
+## Workspace
+
+By default, user data is stored in:
+
+```text
+~/.insightface/gui
+```
+
+The workspace contains the SQLite database, crops, exports, reports, and logs.
+The GUI does not upload images, videos, embeddings, or reports automatically.
+
+You can override the workspace:
+
+```bash
+insightface-gui --workspace /path/to/workspace
+```
+
+## File, Folder, and Video Inputs
+
+Image and video pages use the preview frame as the upload target. Click the
+empty preview frame labeled `Click to upload or drag a file here`, or drag a
+local file onto it. The preview changes color on hover to indicate it is
+clickable, and changes again while a valid file is dragged over it. After a file
+is loaded, the small `×` button removes it, and dragging another file onto the
+preview replaces it.
+
+Multi-file imports, folder imports, CSV files, and local model inputs use
+clickable drag-and-drop selectors with hover and drag-over color feedback.
+
+## Mode-based navigation
+
+InsightFace Evaluation Studio uses four workflow modes. The mode selector is a
+persistent **Workflows** rail on the left side of the window so the current
+workspace is always visible. Face Recognition, Album Management, and Face Swap
+use a single full-width workspace; modes with several workflows show a compact
+secondary sidebar next to the workflow rail:
+
+The Workflows rail also keeps the local-processing notice visible in every
+mode: **All processing is local. No images, embeddings, or reports are uploaded
+automatically.**
+
+1. **Face Recognition**: one Query & Gallery workspace. One gallery image runs
+   1:1 compare; multiple gallery images or a folder run 1:N gallery search.
+2. **Album Management**: one **Album** workspace for local folder import,
+   refresh, DBSCAN face clustering, and photo review.
+3. **Face Swap**: one Source + Target = Result workspace. Target can be an
+   image or video.
+4. **Enterprise Evaluation**: a single local 1:1 / 1:N evaluation workspace
+   with identity-folder import, Auto Split, metrics, and report export.
+
+Global utilities are always available from the top app bar:
+
+- **Settings** opens the application settings dialog for the UI theme and
+  language. Language defaults to the operating system when supported, otherwise
+  English. Supported GUI languages are English, Chinese, Japanese, Korean,
+  Spanish, French, German, Portuguese, and Russian. The included themes are
+  System, Precision Light, Studio Dark, Graphite Pro, Azure Lab, Emerald Focus,
+  and Crimson Audit. Workspace paths are chosen on first launch and are not
+  changed from this dialog.
+- **Models** opens runtime settings, model downloads, and custom model directory
+  tools.
+- **License** opens the License Center dialog.
+
+Settings, Model Settings, Model Downloads, and License Center are intentionally
+not shown in the left sidebar.
+
+## Models
+
+Open **Models** from the top app bar or **Tools > Models** to choose:
+
+- model pack: `buffalo_l`, `buffalo_s`, `antelopev2`, or a custom model folder
+- provider: Auto, CPU, CUDA when `CUDAExecutionProvider` is available
+- detection size: Auto, 128x128, 320x320, 640x640, 1024x1024
+- face swap model
+- GFPGAN post-processing model, if `GFPGANv1.4.onnx` has been downloaded
+
+**Auto** detection size is the default. It runs joint 128x128 and 640x640
+detection and merges duplicate boxes.
+
+The Runtime tab also lets you choose a face swap model. Only downloaded local
+swap models are listed. Download `inswapper_128.onnx` or another compatible
+swap model from **Models > Downloads** first.
+
+The Runtime tab can also enable GFPGAN post-processing for face swap. GFPGAN is
+listed in **Models > Downloads** as a third-party restore model and is not
+downloaded automatically. When enabled, face swap output is passed through the
+configured GFPGAN ONNX model using a 512x512 restore pass.
+
+The GUI opens even when a model is missing. In that case, pages show
+`Model is not loaded. Please open Models.`
+
+The GUI does not download models automatically. Open **Models > Downloads** and
+click **Refresh Download URLs** to fetch the latest release assets from:
+
+```text
+https://github.com/deepinsight/insightface/releases
+```
+
+The refreshed URLs are cached in:
+
+```text
+~/.insightface/gui/cache/model_download_urls.json
+```
+
+The GFPGANv1.4 third-party entry is added by the GUI alongside refreshed
+InsightFace release assets. It points to the Hugging Face `resolve/main` file
+URL for `Gourieff/ReActor` and is saved under the same local model root after
+manual download.
+
+Downloaded archives are cached in:
+
+```text
+~/.insightface/gui/cache/models
+```
+
+Zip model packages are extracted to:
+
+```text
+~/.insightface/models/<model_name>/
+```
+
+For example:
+
+```text
+~/.insightface/models/buffalo_l/
+~/.insightface/models/antelopev2/
+```
+
+Users can also manually place model directories under `~/.insightface/models/`
+or choose **custom model directory** from **Models > Runtime** to reveal the
+custom directory field.
+
+## Face Recognition
+
+Open **Face Recognition** and use the full-width **Query & Gallery** workspace.
+The Query preview accepts one image by click or drag. The Gallery panel accepts
+one image, multiple images, or a folder by click or drag. There is no separate
+Choose button in Gallery; click the Gallery panel or drag files/folders onto it.
+
+If Gallery contains one image, **Run Recognition** automatically runs 1:1
+compare and reports similarity, threshold, decision, and detection score. If
+Gallery contains multiple images or a folder, the same button runs 1:N gallery
+search and ranks the gallery images by similarity.
+
+## Multi-face Photo Recognition
+
+Open **Multi-face Photo Recognition** to detect all faces in a group photo,
+identify them against People Library, save results to the local database, and
+export annotated images or CSV/JSON.
+
+## Batch Processing
+
+Open **Batch Folder Processing**, select an image folder, and choose recursive
+scan, crop saving, and identification options. Batch results can be exported to
+CSV and JSON.
+
+## Album People Clustering
+
+Open **Album Management > Album** to add one or more album directories. Click
+**Import / Refresh** to scan new image files, detect faces, save local crops,
+and cluster all indexed faces from the selected directories. The page exposes a
+cosine similarity threshold with a default value of `0.48`; higher values make
+clusters stricter. DBSCAN internally receives cosine distance as
+`1 - cosine threshold`. The GUI extra installs scikit-learn for DBSCAN
+clustering.
+
+Album cluster IDs avoid duplicating existing People Library IDs. When a cluster
+matches an existing person within the configured duplicate distance threshold
+(`0.28` by default), the existing person ID is reused; otherwise the page assigns
+the next available album person ID. The cluster thumbnail is chosen from the
+face nearest the cluster centroid. Selecting a cluster shows all original photo
+thumbnails for that cluster, and double-clicking a thumbnail opens the original
+image.
+
+Album directories and clustering results are saved in the local SQLite
+database so the page can restore them on the next launch. **Clear** only clears
+the selected album directories and leaves the current clustering results
+visible. **Rebuild All** asks for confirmation, then reprocesses all selected
+album directories from scratch and replaces the saved clustering results. If no
+album directories are selected, **Rebuild All** clears the saved clustering
+results.
+
+## Enterprise Evaluation
+
+Open **Enterprise Evaluation** to run no-code local evaluations:
+
+- **1:1 Verification** from identity folders. Each subfolder is one identity.
+  With **Auto Split**, the file containing `gallery` or the first sorted image
+  becomes that identity's gallery image; other images are probes. Matching is
+  probe-vs-gallery across identities. Without Auto Split, all images are probes
+  and the page runs full pairwise probe-vs-probe comparisons.
+- **1:N Identification** from identity folders. With **Auto Split**, select a
+  dataset containing `identities/<identity folders>` or identity folders
+  directly; gallery images are selected with the same rule as 1:1. Without Auto
+  Split, use `gallery/<identity>`, `probe/<identity>`, and optional `unknown/`
+  folders.
+
+Choose a **Multi-face handling** policy before validation:
+
+- **Require exactly one face**: default. Validation fails if any evaluation
+  image contains more than one detected face.
+- **Use largest face**: validation warns on multi-face images and evaluation
+  uses the largest detected face.
+- **Use largest centered face**: validation warns on multi-face images and
+  evaluation uses the face with the best area-minus-center-distance score.
+- **Mark as skip**: multi-face images are skipped. If a required gallery image
+  would be skipped, validation fails because that identity no longer has a
+  usable gallery sample.
+
+Click **Validate Dataset** before running. Validation checks the folder layout,
+Auto Split rules, gallery/probe availability, generated positive and negative
+pairs for 1:1, required gallery coverage for 1:N, and the selected multi-face
+policy. **Run Evaluation** is enabled only after validation passes.
+
+1:1 reports include best cosine threshold accuracy, the threshold at that
+operating point, and TAR@FAR for `1e-6`, `1e-5`, `1e-4`, and `1e-3`. 1:N
+reports include Top1 plus TAR@FAR for `1e-5`, `1e-4`, `1e-3`, and `1e-2`,
+including the corresponding thresholds.
+
+## Report Export
+
+Reports are written to:
+
+```text
+~/.insightface/gui/reports
+```
+
+Markdown, HTML, and PDF are supported when the GUI extra is installed.
+
+## Face Swap
+
+Open **Face Swap** to use the full-width Source + Target = Result workspace.
+Choose the swap model in **Models > Runtime**; the page itself does not expose a
+model picker. The swap model is loaded only when you click **Run Face Swap**.
+If GFPGAN post-processing is enabled in **Models > Runtime**, the swapped image
+or swapped video frames are restored with GFPGAN before being saved.
+
+Source is always an image. Target can be an image or a video, and the workflow
+automatically chooses image swap or video swap from the target file type. Video
+swap writes an `.mp4` result to the exports folder and shows a preview frame
+when one is available.
+
+If the model is missing, it shows:
+
+```text
+Face swap model not found. Please download and choose a swap model in Models.
+```
+
+Face swap may require separate commercial authorization depending on usage and
+model license. Use only with appropriate rights and consent.
+
+## License Notice
+
+Code and model files may have different licenses. Research or publicly
+distributed pretrained models may be restricted to non-commercial or research
+usage. Commercial deployment requires appropriate model authorization.
+
+This tool does not provide legal advice. Users are responsible for consent,
+privacy, retention, and compliance with applicable biometric regulations.
+
+## Troubleshooting
+
+Run without automatic model loading:
+
+```bash
+insightface-gui --safe-mode
+```
+
+Safe mode is only applied to the current launch. It is intended for
+troubleshooting model/provider issues and is not saved to `config.json`.
+
+Force CPU:
+
+```bash
+insightface-gui --provider cpu
+```
+
+`CUDA` is selectable only when ONNX Runtime reports `CUDAExecutionProvider`.
+If CUDA is requested on a machine without a usable CUDA provider, the GUI
+falls back to Auto/CPU instead of exposing a broken GPU option.
+
+Logs are stored in:
+
+```text
+~/.insightface/gui/logs/app.log
+```
diff --git a/python-package/docs/gui_packaging.md b/python-package/docs/gui_packaging.md
new file mode 100644
index 000000000..95ee963e7
--- /dev/null
+++ b/python-package/docs/gui_packaging.md
@@ -0,0 +1,148 @@
+# GUI Packaging and Distribution
+
+InsightFace Evaluation Studio supports development installs, PyPI package
+installs, and desktop application builds.
+
+## Development Install
+
+```bash
+cd python-package
+pip install -e ".[gui]"
+insightface-gui
+```
+
+## Build Python Package
+
+```bash
+cd python-package
+python -m pip install build twine
+python -m build
+python -m twine check dist/*
+```
+
+## Optional face3d Extension
+
+The default 1.0.1 package does not compile the optional `face3d` Cython/C++
+extension. This avoids requiring a C++ compiler for normal inference and GUI
+users.
+
+To manually enable it, install the optional dependencies and pass the explicit
+build flag:
+
+```bash
+cd python-package
+pip install -e ".[face3d]" --no-build-isolation --config-settings editable_mode=compat
+python setup.py build_ext --inplace --with-face3d
+```
+
+Equivalent environment-variable control:
+
+```bash
+INSIGHTFACE_WITH_FACE3D=1 python setup.py build_ext --inplace
+```
+
+The chosen parameter name is `--with-face3d`.
+
+## Upload PyPI
+
+The repository includes a guarded release helper for official PyPI. It builds
+the source distribution and wheel, runs the GUI release smoke tests, checks that
+the version is not already published, runs `twine check`, and asks for explicit
+confirmation before upload.
+
+Dry run:
+
+```bash
+cd python-package
+python -m pip install build twine pytest
+bash packaging/pypi/build_upload_pypi.sh --dry-run
+```
+
+Official upload:
+
+```bash
+cd python-package
+export TWINE_USERNAME=__token__
+export TWINE_PASSWORD='pypi-your-official-token'
+bash packaging/pypi/build_upload_pypi.sh
+```
+
+The default release does not build `face3d`. To publish a package with the
+optional extension enabled, run:
+
+```bash
+bash packaging/pypi/build_upload_pypi.sh --with-face3d
+```
+
+Manual equivalent:
+
+```bash
+cd python-package
+python -m pip install build twine
+python -m build
+python -m twine check dist/*
+python -m twine upload dist/*
+```
+
+Only project maintainers or CI configured with PyPI Trusted Publisher should
+upload official PyPI releases. Codex should not attempt to upload PyPI.
+
+Before releasing 1.0.1, confirm model licenses, README content, version numbers,
+wheel contents, third-party notices, and that the package name is `insightface`.
+PyPI versions are immutable, so a released version cannot be overwritten.
+
+## Build Desktop Application
+
+The first desktop packaging path uses PyInstaller one-folder mode.
+
+Windows:
+
+```powershell
+cd python-package
+pip install -e ".[gui]"
+pip install pyinstaller
+powershell -ExecutionPolicy Bypass -File packaging/desktop/build_windows.ps1
+```
+
+macOS:
+
+```bash
+cd python-package
+pip install -e ".[gui]"
+pip install pyinstaller
+bash packaging/desktop/build_macos.sh
+```
+
+Linux:
+
+```bash
+cd python-package
+pip install -e ".[gui]"
+pip install pyinstaller
+bash packaging/desktop/build_linux.sh
+```
+
+Output:
+
+- Windows: `.exe` or `dist/InsightFace Evaluation Studio/`
+- macOS: `.app`, with `.dmg` creation left as a release step
+- Linux: executable directory, with AppImage/deb creation left as a release step
+
+## Notes
+
+- The GUI extra uses `PySide6-Essentials` for Qt Widgets and includes
+  `reportlab` for PDF reports without installing the much larger
+  `PySide6_Addons` wheel.
+- onnxruntime and onnxruntime-gpu may require additional dynamic library work.
+- CUDA builds are not recommended for default community installers.
+- A CPU provider build is the safest default.
+- GPU/CUDA builds can be distributed as separate enterprise builds.
+- Do not package user workspaces, SQLite databases, reports, images, videos, or
+  embeddings.
+- Do not package commercial model files by default.
+- The GUI workspace/cache defaults to `~/.insightface/gui/`; model packages are
+  manually downloaded into `~/.insightface/gui/cache/models` and extracted to
+  `~/.insightface/models/<model_name>/`.
+- Community builds must use PyInstaller one-folder mode, include third-party
+  license notices, include LGPL license text, include Qt/PySide6 source offer,
+  and must not restrict replacement of Qt/PySide6 shared libraries.
diff --git a/python-package/insightface/__init__.py b/python-package/insightface/__init__.py
index 571cc5f5b..83a71904b 100644
--- a/python-package/insightface/__init__.py
+++ b/python-package/insightface/__init__.py
@@ -3,26 +3,18 @@
 """InsightFace: A Face Analysis Toolkit."""
 from __future__ import absolute_import
 
-# mxnet version check
-#mx_version = '1.4.0'
 try:
-    import mxnet as mx
-    #from distutils.version import LooseVersion
-    #if LooseVersion(mx.__version__) < LooseVersion(mx_version):
-    #    msg = (
-    #        "Legacy mxnet-mkl=={} detected, some new modules may not work properly. "
-    #        "mxnet-mkl>={} is required. You can use pip to upgrade mxnet "
-    #        "`pip install mxnet-mkl --pre --upgrade` "
-    #        "or `pip install mxnet-cu90mkl --pre --upgrade`").format(mx.__version__, mx_version)
-    #    raise ImportError(msg)
+    #import mxnet as mx
+    import onnxruntime
 except ImportError:
     raise ImportError(
-        "Unable to import dependency mxnet. "
-        "A quick tip is to install via `pip install mxnet-mkl/mxnet-cu90mkl --pre`. ")
+        "Unable to import dependency onnxruntime. "
+    )
 
-__version__ = '0.1.5'
+__version__ = '1.0.1'
 
 from . import model_zoo
 from . import utils
 from . import app
-
+from . import data
+from . import thirdparty
diff --git a/python-package/insightface/app/__init__.py b/python-package/insightface/app/__init__.py
index cc5746168..16f441336 100644
--- a/python-package/insightface/app/__init__.py
+++ b/python-package/insightface/app/__init__.py
@@ -1 +1,10 @@
 from .face_analysis import *
+try:
+    import os
+
+    os.environ.setdefault("NO_ALBUMENTATIONS_UPDATE", "1")
+    from .mask_renderer import *
+except ImportError:
+    # The mask renderer depends on optional compiled face3d extensions. Keep the
+    # main InsightFace API importable in source-tree and GUI-safe environments.
+    pass
diff --git a/python-package/insightface/app/common.py b/python-package/insightface/app/common.py
new file mode 100644
index 000000000..5d5645b19
--- /dev/null
+++ b/python-package/insightface/app/common.py
@@ -0,0 +1,48 @@
+import numpy as np
+from numpy.linalg import norm as l2norm
+
+class Face(dict):
+
+    def __init__(self, d=None, **kwargs):
+        if d is None:
+            d = {}
+        if kwargs:
+            d.update(**kwargs)
+        for k, v in d.items():
+            setattr(self, k, v)
+        # Class attributes
+        #for k in self.__class__.__dict__.keys():
+        #    if not (k.startswith('__') and k.endswith('__')) and not k in ('update', 'pop'):
+        #        setattr(self, k, getattr(self, k))
+
+    def __setattr__(self, name, value):
+        if isinstance(value, (list, tuple)):
+            value = [self.__class__(x)
+                    if isinstance(x, dict) else x for x in value]
+        elif isinstance(value, dict) and not isinstance(value, self.__class__):
+            value = self.__class__(value)
+        super(Face, self).__setattr__(name, value)
+        super(Face, self).__setitem__(name, value)
+
+    __setitem__ = __setattr__
+
+    def __getattr__(self, name):
+        return None
+
+    @property
+    def embedding_norm(self):
+        if self.embedding is None:
+            return None
+        return l2norm(self.embedding)
+
+    @property 
+    def normed_embedding(self):
+        if self.embedding is None:
+            return None
+        return self.embedding / self.embedding_norm
+
+    @property 
+    def sex(self):
+        if self.gender is None:
+            return None
+        return 'M' if self.gender==1 else 'F'
diff --git a/python-package/insightface/app/face_analysis.py b/python-package/insightface/app/face_analysis.py
index 938b6b31d..ea9f9719c 100644
--- a/python-package/insightface/app/face_analysis.py
+++ b/python-package/insightface/app/face_analysis.py
@@ -1,73 +1,127 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+
 from __future__ import division
-import collections
-import mxnet as mx
+
+import glob
+import os.path as osp
+
 import numpy as np
+import onnxruntime
 from numpy.linalg import norm
-import mxnet.ndarray as nd
+
 from ..model_zoo import model_zoo
-from ..utils import face_align
+from ..utils import DEFAULT_MP_NAME, ensure_available
+from .common import Face
+
+__all__ = ['FaceAnalysis']
 
-__all__ = ['FaceAnalysis',
-           'Face']
+DEFAULT_DET_SIZES = [(128, 128), (640, 640)]
 
-Face = collections.namedtuple('Face', [
-        'bbox', 'landmark', 'det_score', 'embedding', 'gender', 'age', 'embedding_norm', 'normed_embedding'])
 
-Face.__new__.__defaults__ = (None,) * len(Face._fields)
+def _is_auto_det_size(det_size):
+    if det_size is None:
+        return True
+    if isinstance(det_size, np.ndarray):
+        det_size = det_size.tolist()
+    if isinstance(det_size, (list, tuple)) and len(det_size) == 2:
+        first, second = det_size
+        if not isinstance(first, (list, tuple, np.ndarray)) and not isinstance(second, (list, tuple, np.ndarray)):
+            try:
+                return int(first) == 0 and int(second) == 0
+            except (TypeError, ValueError):
+                return False
+    return False
+
 
 class FaceAnalysis:
-    def __init__(self, det_name='retinaface_r50_v1', rec_name='arcface_r100_v1', ga_name='genderage_v1'):
-        assert det_name is not None
-        self.det_model = model_zoo.get_model(det_name)
-        if rec_name is not None:
-            self.rec_model = model_zoo.get_model(rec_name)
-        else:
-            self.rec_model = None
-        if ga_name is not None:
-            self.ga_model = model_zoo.get_model(ga_name)
-        else:
-            self.ga_model = None
-
-    def prepare(self, ctx_id, nms=0.4):
-        self.det_model.prepare(ctx_id, nms)
-        if self.rec_model is not None:
-            self.rec_model.prepare(ctx_id)
-        if self.ga_model is not None:
-            self.ga_model.prepare(ctx_id)
-
-    def get(self, img, det_thresh = 0.8, det_scale = 1.0, max_num = 0):
-        bboxes, landmarks = self.det_model.detect(img, threshold=det_thresh, scale = det_scale)
-        if bboxes.shape[0]==0:
+    def __init__(self, name=DEFAULT_MP_NAME, root='~/.insightface', allowed_modules=None, **kwargs):
+        onnxruntime.set_default_logger_severity(3)
+        self.models = {}
+        self.model_dir = ensure_available('models', name, root=root)
+        onnx_files = glob.glob(osp.join(self.model_dir, '*.onnx'))
+        onnx_files = sorted(onnx_files)
+        for onnx_file in onnx_files:
+            model = model_zoo.get_model(onnx_file, **kwargs)
+            if model is None:
+                print('model not recognized:', onnx_file)
+            elif allowed_modules is not None and model.taskname not in allowed_modules:
+                print('model ignore:', onnx_file, model.taskname)
+                del model
+            elif model.taskname not in self.models and (allowed_modules is None or model.taskname in allowed_modules):
+                print('find model:', onnx_file, model.taskname, model.input_shape, model.input_mean, model.input_std)
+                self.models[model.taskname] = model
+            else:
+                print('duplicated model task type, ignore:', onnx_file, model.taskname)
+                del model
+        assert 'detection' in self.models
+        self.det_model = self.models['detection']
+
+
+    def prepare(self, ctx_id, det_thresh=0.5, det_size=None):
+        self.det_thresh = det_thresh
+        if _is_auto_det_size(det_size):
+            det_size = list(DEFAULT_DET_SIZES)
+        print('set det-size:', det_size)
+        self.det_size = det_size
+        for taskname, model in self.models.items():
+            if taskname=='detection':
+                model.prepare(ctx_id, input_size=det_size, det_thresh=det_thresh)
+            else:
+                model.prepare(ctx_id)
+
+    def get(self, img, max_num=0, det_metric='default'):
+        bboxes, kpss = self.det_model.detect(img,
+                                             max_num=max_num,
+                                             metric=det_metric)
+        if bboxes.shape[0] == 0:
             return []
-        if max_num>0 and bboxes.shape[0]>max_num:
-            area = (bboxes[:,2]-bboxes[:,0])*(bboxes[:,3]-bboxes[:,1])
-            img_center = img.shape[0]//2, img.shape[1]//2
-            offsets = np.vstack([ (bboxes[:,0]+bboxes[:,2])/2-img_center[1], (bboxes[:,1]+bboxes[:,3])/2-img_center[0] ])
-            offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-            values = area-offset_dist_squared*2.0 # some extra weight on the centering
-            bindex = np.argsort(values)[::-1] # some extra weight on the centering
-            bindex = bindex[0:max_num]
-            bboxes = bboxes[bindex, :]
-            landmarks = landmarks[bindex, :]
         ret = []
         for i in range(bboxes.shape[0]):
             bbox = bboxes[i, 0:4]
-            det_score = bboxes[i,4]
-            landmark = landmarks[i]
-            _img = face_align.norm_crop(img, landmark = landmark)
-            embedding = None
-            embedding_norm = None
-            normed_embedding = None
-            gender = None
-            age = None
-            if self.rec_model is not None:
-                embedding = self.rec_model.get_embedding(_img).flatten()
-                embedding_norm = norm(embedding)
-                normed_embedding = embedding / embedding_norm
-            if self.ga_model is not None:
-                gender, age = self.ga_model.get(_img)
-            face = Face(bbox = bbox, landmark = landmark, det_score = det_score, embedding = embedding, gender = gender, age = age
-                    , normed_embedding=normed_embedding, embedding_norm = embedding_norm)
+            det_score = bboxes[i, 4]
+            kps = None
+            if kpss is not None:
+                kps = kpss[i]
+            face = Face(bbox=bbox, kps=kps, det_score=det_score)
+            for taskname, model in self.models.items():
+                if taskname=='detection':
+                    continue
+                model.get(img, face)
             ret.append(face)
         return ret
 
+    def draw_on(self, img, faces):
+        import cv2
+        dimg = img.copy()
+        for i in range(len(faces)):
+            face = faces[i]
+            box = face.bbox.astype(int)
+            color = (0, 0, 255)
+            cv2.rectangle(dimg, (box[0], box[1]), (box[2], box[3]), color, 2)
+            if face.kps is not None:
+                kps = face.kps.astype(int)
+                #print(landmark.shape)
+                for l in range(kps.shape[0]):
+                    color = (0, 0, 255)
+                    if l == 0 or l == 3:
+                        color = (0, 255, 0)
+                    cv2.circle(dimg, (kps[l][0], kps[l][1]), 1, color,
+                               2)
+            if face.gender is not None and face.age is not None:
+                cv2.putText(dimg,'%s,%d'%(face.sex,face.age), (box[0]-1, box[1]-4),cv2.FONT_HERSHEY_COMPLEX,0.7,(0,255,0),1)
+
+            #for key, value in face.items():
+            #    if key.startswith('landmark_3d'):
+            #        print(key, value.shape)
+            #        print(value[0:10,:])
+            #        lmk = np.round(value).astype(int)
+            #        for l in range(lmk.shape[0]):
+            #            color = (255, 0, 0)
+            #            cv2.circle(dimg, (lmk[l][0], lmk[l][1]), 1, color,
+            #                       2)
+        return dimg
diff --git a/python-package/insightface/app/mask_renderer.py b/python-package/insightface/app/mask_renderer.py
new file mode 100644
index 000000000..2815f5018
--- /dev/null
+++ b/python-package/insightface/app/mask_renderer.py
@@ -0,0 +1,232 @@
+import os, sys, datetime
+import numpy as np
+import os.path as osp
+import albumentations as A
+from albumentations.core.transforms_interface import ImageOnlyTransform
+from .face_analysis import FaceAnalysis
+from ..utils import get_model_dir
+from ..thirdparty import face3d
+from ..data import get_image as ins_get_image
+from ..utils import DEFAULT_MP_NAME
+import cv2
+
+class MaskRenderer:
+    def __init__(self, name=DEFAULT_MP_NAME, root='~/.insightface', insfa=None):
+        #if insfa is None, enter render_only mode
+        self.mp_name = name
+        self.root = root
+        self.insfa = insfa
+        model_dir = get_model_dir(name, root)
+        bfm_file = osp.join(model_dir, 'BFM.mat')
+        assert osp.exists(bfm_file), 'should contains BFM.mat in your model directory'
+        self.bfm = face3d.morphable_model.MorphabelModel(bfm_file)
+        self.index_ind = self.bfm.kpt_ind
+        bfm_uv_file = osp.join(model_dir, 'BFM_UV.mat')
+        assert osp.exists(bfm_uv_file), 'should contains BFM_UV.mat in your model directory'
+        uv_coords = face3d.morphable_model.load.load_uv_coords(bfm_uv_file)
+        self.uv_size = (224,224)
+        self.mask_stxr =  0.1
+        self.mask_styr = 0.33
+        self.mask_etxr = 0.9
+        self.mask_etyr =  0.7
+        self.tex_h , self.tex_w, self.tex_c = self.uv_size[1] , self.uv_size[0],3
+        texcoord = np.zeros_like(uv_coords)
+        texcoord[:, 0] = uv_coords[:, 0] * (self.tex_h - 1)
+        texcoord[:, 1] = uv_coords[:, 1] * (self.tex_w - 1)
+        texcoord[:, 1] = self.tex_w - texcoord[:, 1] - 1
+        self.texcoord = np.hstack((texcoord, np.zeros((texcoord.shape[0], 1))))
+        self.X_ind = self.bfm.kpt_ind
+        self.mask_image_names = ['mask_white', 'mask_blue', 'mask_black', 'mask_green']
+        self.mask_aug_probs = [0.4, 0.4, 0.1, 0.1]
+        #self.mask_images = []
+        #self.mask_images_rgb = []
+        #for image_name in mask_image_names:
+        #    mask_image = ins_get_image(image_name)
+        #    self.mask_images.append(mask_image)
+        #    mask_image_rgb = mask_image[:,:,::-1]
+        #    self.mask_images_rgb.append(mask_image_rgb)
+
+
+    def prepare(self, ctx_id=0, det_thresh=0.5, det_size=(128, 128)):
+        self.pre_ctx_id = ctx_id
+        self.pre_det_thresh = det_thresh
+        self.pre_det_size = det_size
+
+    def transform(self, shape3D, R):
+        s = 1.0
+        shape3D[:2, :] = shape3D[:2, :]
+        shape3D = s * np.dot(R, shape3D)
+        return shape3D
+
+    def preprocess(self, vertices, w, h):
+        R1 = face3d.mesh.transform.angle2matrix([0, 180, 180])
+        t = np.array([-w // 2, -h // 2, 0])
+        vertices = vertices.T
+        vertices += t
+        vertices = self.transform(vertices.T, R1).T
+        return vertices
+
+    def project_to_2d(self,vertices,s,angles,t):
+        transformed_vertices = self.bfm.transform(vertices, s, angles, t)
+        projected_vertices = transformed_vertices.copy() # using stantard camera & orth projection
+        return projected_vertices[self.bfm.kpt_ind, :2]
+
+    def params_to_vertices(self,params  , H , W):
+        fitted_sp, fitted_ep, fitted_s, fitted_angles, fitted_t  = params
+        fitted_vertices = self.bfm.generate_vertices(fitted_sp, fitted_ep)
+        transformed_vertices = self.bfm.transform(fitted_vertices, fitted_s, fitted_angles,
+                                                  fitted_t)
+        transformed_vertices = self.preprocess(transformed_vertices.T, W, H)
+        image_vertices = face3d.mesh.transform.to_image(transformed_vertices, H, W)
+        return image_vertices
+
+    def draw_lmk(self, face_image):
+        faces = self.insfa.get(face_image, max_num=1)
+        if len(faces)==0:
+            return face_image
+        return self.insfa.draw_on(face_image, faces)
+
+    def build_params(self, face_image):
+        #landmark = self.if3d68_handler.get(face_image)
+        #if landmark is None:
+        #    return None #face not found
+        if self.insfa is None:
+            self.insfa = FaceAnalysis(name=self.mp_name, root=self.root, allowed_modules=['detection', 'landmark_3d_68'])
+            self.insfa.prepare(ctx_id=self.pre_ctx_id,  det_thresh=self.pre_det_thresh, det_size=self.pre_det_size)
+
+        faces = self.insfa.get(face_image, max_num=1)
+        if len(faces)==0:
+            return None
+        landmark = faces[0].landmark_3d_68[:,:2]
+        fitted_sp, fitted_ep, fitted_s, fitted_angles, fitted_t = self.bfm.fit(landmark, self.X_ind, max_iter = 3)
+        return [fitted_sp, fitted_ep, fitted_s, fitted_angles, fitted_t]
+
+    def generate_mask_uv(self,mask, positions):
+        uv_size = (self.uv_size[1], self.uv_size[0], 3)
+        h, w, c = uv_size
+        uv = np.zeros(shape=(self.uv_size[1],self.uv_size[0], 3), dtype=np.uint8)
+        stxr, styr  = positions[0], positions[1]
+        etxr, etyr = positions[2], positions[3]
+        stx, sty = int(w * stxr), int(h * styr)
+        etx, ety = int(w * etxr), int(h * etyr)
+        height = ety - sty
+        width = etx - stx
+        mask = cv2.resize(mask, (width, height))
+        uv[sty:ety, stx:etx] = mask
+        return uv
+
+    def render_mask(self,face_image, mask_image, params, input_is_rgb=False, auto_blend = True, positions=[0.1, 0.33, 0.9, 0.7]):
+        if isinstance(mask_image, str):
+            to_rgb = True if input_is_rgb else False
+            mask_image = ins_get_image(mask_image, to_rgb=to_rgb)
+        uv_mask_image = self.generate_mask_uv(mask_image, positions)
+        h,w,c = face_image.shape
+        image_vertices = self.params_to_vertices(params ,h,w)
+        output = (1-face3d.mesh.render.render_texture(image_vertices, self.bfm.full_triangles , uv_mask_image, self.texcoord, self.bfm.full_triangles, h , w ))*255
+        output = output.astype(np.uint8)
+        if auto_blend:
+            mask_bd = (output==255).astype(np.uint8)
+            final = face_image*mask_bd + (1-mask_bd)*output
+            return final
+        return output
+
+    #def mask_augmentation(self, face_image, label, input_is_rgb=False, p=0.1):
+    #    if np.random.random()<p:
+    #        assert isinstance(label, (list, np.ndarray)), 'make sure the rec dataset includes mask params'
+    #        assert len(label)==237 or len(lable)==235, 'make sure the rec dataset includes mask params'
+    #        if len(label)==237:
+    #            if label[1]<0.0: #invalid label for mask aug
+    #                return face_image
+    #            label = label[2:]
+    #        params = self.decode_params(label)
+    #        mask_image_name = np.random.choice(self.mask_image_names, p=self.mask_aug_probs)
+    #        pos = np.random.uniform(0.33, 0.5)
+    #        face_image = self.render_mask(face_image, mask_image_name, params, input_is_rgb=input_is_rgb, positions=[0.1, pos, 0.9, 0.7])
+    #    return face_image
+
+    @staticmethod
+    def encode_params(params):
+        p0 = list(params[0])
+        p1 = list(params[1])
+        p2 = [float(params[2])]
+        p3 = list(params[3])
+        p4 = list(params[4])
+        return p0+p1+p2+p3+p4
+
+    @staticmethod
+    def decode_params(params):
+        p0 = params[0:199]
+        p0 = np.array(p0, dtype=np.float32).reshape( (-1, 1))
+        p1 = params[199:228]
+        p1 = np.array(p1, dtype=np.float32).reshape( (-1, 1))
+        p2 = params[228]
+        p3 = tuple(params[229:232])
+        p4 = params[232:235]
+        p4 = np.array(p4, dtype=np.float32).reshape( (-1, 1))
+        return p0, p1, p2, p3, p4
+    
+class MaskAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'],
+            mask_probs=[0.4,0.4,0.1,0.1],
+            h_low = 0.33,
+            h_high = 0.35,
+            always_apply=False,
+            p=1.0,
+            ):
+        super(MaskAugmentation, self).__init__(always_apply, p)
+        self.renderer = MaskRenderer()
+        assert len(mask_names)>0
+        assert len(mask_names)==len(mask_probs)
+        self.mask_names = mask_names
+        self.mask_probs = mask_probs
+        self.h_low = h_low
+        self.h_high = h_high
+        #self.hlabel = None
+
+
+    def apply(self, image, hlabel, mask_name, h_pos, **params):
+        #print(params.keys())
+        #hlabel = params.get('hlabel')
+        assert len(hlabel)==237 or len(hlabel)==235, 'make sure the rec dataset includes mask params'
+        if len(hlabel)==237:
+            if hlabel[1]<0.0:
+                return image
+            hlabel = hlabel[2:]
+        #print(len(hlabel))
+        mask_params = self.renderer.decode_params(hlabel)
+        image = self.renderer.render_mask(image, mask_name, mask_params, input_is_rgb=True, positions=[0.1, h_pos, 0.9, 0.7])
+        return image
+
+    @property
+    def targets_as_params(self):
+        return ["image", "hlabel"]
+
+    def get_params_dependent_on_targets(self, params):
+        hlabel = params['hlabel']
+        mask_name = np.random.choice(self.mask_names, p=self.mask_probs)
+        h_pos = np.random.uniform(self.h_low, self.h_high)
+        return {'hlabel': hlabel, 'mask_name': mask_name, 'h_pos': h_pos}
+
+    def get_transform_init_args_names(self):
+        #return ("hlabel", 'mask_names', 'mask_probs', 'h_low', 'h_high')
+        return ('mask_names', 'mask_probs', 'h_low', 'h_high')
+
+
+if __name__ == "__main__":
+    tool = MaskRenderer('antelope')
+    tool.prepare(det_size=(128,128))
+    image = cv2.imread("Tom_Hanks_54745.png")
+    params = tool.build_params(image)
+    #out = tool.draw_lmk(image)
+    #cv2.imwrite('output_lmk.jpg', out)
+    #mask_image  = cv2.imread("masks/mask1.jpg")
+    #mask_image  = cv2.imread("masks/black-mask.png")
+    #mask_image  = cv2.imread("masks/mask2.jpg")
+    mask_out = tool.render_mask(image, 'mask_blue', params)# use single thread to test the time cost
+
+    cv2.imwrite('output_mask.jpg', mask_out)
+
+
diff --git a/python-package/insightface/commands/__init__.py b/python-package/insightface/commands/__init__.py
new file mode 100644
index 000000000..7fedaa0db
--- /dev/null
+++ b/python-package/insightface/commands/__init__.py
@@ -0,0 +1,13 @@
+from abc import ABC, abstractmethod
+from argparse import ArgumentParser
+
+
+class BaseInsightFaceCLICommand(ABC):
+    @staticmethod
+    @abstractmethod
+    def register_subcommand(parser: ArgumentParser):
+        raise NotImplementedError()
+
+    @abstractmethod
+    def run(self):
+        raise NotImplementedError()
diff --git a/python-package/insightface/commands/insightface_cli.py b/python-package/insightface/commands/insightface_cli.py
new file mode 100644
index 000000000..b8adbefcf
--- /dev/null
+++ b/python-package/insightface/commands/insightface_cli.py
@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+
+from argparse import ArgumentParser
+
+from .model_download import ModelDownloadCommand
+from .rec_add_mask_param import RecAddMaskParamCommand
+
+def main():
+    parser = ArgumentParser("InsightFace CLI tool", usage="insightface-cli <command> [<args>]")
+    commands_parser = parser.add_subparsers(help="insightface-cli command-line helpers")
+
+    # Register commands
+    ModelDownloadCommand.register_subcommand(commands_parser)
+    RecAddMaskParamCommand.register_subcommand(commands_parser)
+
+    args = parser.parse_args()
+
+    if not hasattr(args, "func"):
+        parser.print_help()
+        exit(1)
+
+    # Run
+    service = args.func(args)
+    service.run()
+
+
+if __name__ == "__main__":
+    main()
+
diff --git a/python-package/insightface/commands/model_download.py b/python-package/insightface/commands/model_download.py
new file mode 100644
index 000000000..c5a14a33d
--- /dev/null
+++ b/python-package/insightface/commands/model_download.py
@@ -0,0 +1,36 @@
+from argparse import ArgumentParser
+
+from . import BaseInsightFaceCLICommand
+import os
+import os.path as osp
+import zipfile
+import glob
+from ..utils import download
+
+
+def model_download_command_factory(args):
+    return ModelDownloadCommand(args.model, args.root, args.force)
+
+
+class ModelDownloadCommand(BaseInsightFaceCLICommand):
+    #_url_format = '{repo_url}models/{file_name}.zip'
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        download_parser = parser.add_parser("model.download")
+        download_parser.add_argument(
+            "--root", type=str, default='~/.insightface', help="Path to location to store the models"
+        )
+        download_parser.add_argument(
+            "--force", action="store_true", help="Force the model to be download even if already in root-dir"
+        )
+        download_parser.add_argument("model", type=str, help="Name of the model to download")
+        download_parser.set_defaults(func=model_download_command_factory)
+
+    def __init__(self, model: str, root: str, force: bool):
+        self._model = model
+        self._root = root
+        self._force = force
+
+    def run(self):
+        download('models', self._model, force=self._force, root=self._root)
+
diff --git a/python-package/insightface/commands/rec_add_mask_param.py b/python-package/insightface/commands/rec_add_mask_param.py
new file mode 100644
index 000000000..f73173d5f
--- /dev/null
+++ b/python-package/insightface/commands/rec_add_mask_param.py
@@ -0,0 +1,107 @@
+
+import numbers
+import os
+from argparse import ArgumentParser, Namespace
+
+import numpy as np
+
+from . import BaseInsightFaceCLICommand
+
+
+def rec_add_mask_param_command_factory(args: Namespace):
+
+    return RecAddMaskParamCommand(
+        args.input, args.output
+    )
+
+
+class RecAddMaskParamCommand(BaseInsightFaceCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        _parser = parser.add_parser("rec.addmaskparam")
+        _parser.add_argument("input", type=str, help="input rec")
+        _parser.add_argument("output", type=str, help="output rec, with mask param")
+        _parser.set_defaults(func=rec_add_mask_param_command_factory)
+
+    def __init__(
+        self,
+        input: str,
+        output: str,
+    ):
+        self._input = input
+        self._output = output
+
+
+    def run(self):
+        try:
+            import mxnet as mx
+        except ImportError as exc:
+            raise ImportError(
+                "The rec.addmaskparam command requires mxnet. "
+                "Please install mxnet before running this command."
+            ) from exc
+        try:
+            from ..app.mask_renderer import MaskRenderer
+            from ..data.rec_builder import RecBuilder
+        except ImportError as exc:
+            raise ImportError(
+                "The rec.addmaskparam command requires optional mask rendering "
+                "and MXNet record dependencies. Please install the required "
+                "optional dependencies before running this command."
+            ) from exc
+
+        tool = MaskRenderer()
+        tool.prepare(ctx_id=0, det_size=(128,128))
+        root_dir = self._input
+        path_imgrec = os.path.join(root_dir, 'train.rec')
+        path_imgidx = os.path.join(root_dir, 'train.idx')
+        imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+        save_path = self._output
+        wrec=RecBuilder(path=save_path)
+        s = imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        if header.flag > 0:
+            if len(header.label)==2:
+                imgidx = np.array(range(1, int(header.label[0])))
+            else:
+                imgidx = np.array(list(self.imgrec.keys))
+        else:
+            imgidx = np.array(list(self.imgrec.keys))
+        stat = [0, 0]
+        print('total:', len(imgidx))
+        for iid, idx in enumerate(imgidx):
+            #if iid==500000:
+            #    break
+            if iid%1000==0:
+                print('processing:', iid)
+            s = imgrec.read_idx(idx)
+            header, img = mx.recordio.unpack(s)
+            label = header.label
+            if not isinstance(label, numbers.Number):
+                label = label[0]
+            sample = mx.image.imdecode(img).asnumpy()
+            bgr = sample[:,:,::-1]
+            params = tool.build_params(bgr)
+            #if iid<10:
+            #    mask_out = tool.render_mask(bgr, 'mask_blue', params)
+            #    cv2.imwrite('maskout_%d.jpg'%iid, mask_out)
+            stat[1] += 1
+            if params is None:
+                wlabel = [label] + [-1.0]*236
+                stat[0] += 1
+            else:
+                #print(0, params[0].shape, params[0].dtype)
+                #print(1, params[1].shape, params[1].dtype)
+                #print(2, params[2])
+                #print(3, len(params[3]), params[3][0].__class__)
+                #print(4, params[4].shape, params[4].dtype)
+                mask_label = tool.encode_params(params)
+                wlabel = [label, 0.0]+mask_label # 237 including idlabel, total mask params size is 235
+                if iid==0:
+                    print('param size:', len(mask_label), len(wlabel), label)
+            assert len(wlabel)==237
+            wrec.add_image(img, wlabel)
+            #print(len(params))
+
+        wrec.close()
+        print('finished on', self._output, ', failed:', stat[0])
diff --git a/python-package/insightface/data/__init__.py b/python-package/insightface/data/__init__.py
new file mode 100644
index 000000000..665c59ec9
--- /dev/null
+++ b/python-package/insightface/data/__init__.py
@@ -0,0 +1,2 @@
+from .image import get_image
+from .pickle_object import get_object
diff --git a/python-package/insightface/data/image.py b/python-package/insightface/data/image.py
new file mode 100644
index 000000000..343ccc580
--- /dev/null
+++ b/python-package/insightface/data/image.py
@@ -0,0 +1,28 @@
+import cv2
+import os
+import os.path as osp
+from pathlib import Path
+
+class ImageCache:
+    data = {}
+
+def get_image(name, to_rgb=False, use_cache=True):
+    key = (name, to_rgb)
+    if key in ImageCache.data:
+        return ImageCache.data[key]
+    images_dir = osp.join(Path(__file__).parent.absolute(), 'images')
+    ext_names = ['.jpg', '.png', '.jpeg']
+    image_file = None
+    for ext_name in ext_names:
+        _image_file = osp.join(images_dir, "%s%s"%(name, ext_name))
+        if osp.exists(_image_file):
+            image_file = _image_file
+            break
+    assert image_file is not None, '%s not found'%name
+    img = cv2.imread(image_file)
+    if to_rgb:
+        img = img[:,:,::-1]
+    if use_cache:
+        ImageCache.data[key] = img
+    return img
+
diff --git a/deploy/Tom_Hanks_54745.png b/python-package/insightface/data/images/Tom_Hanks_54745.png
similarity index 100%
rename from deploy/Tom_Hanks_54745.png
rename to python-package/insightface/data/images/Tom_Hanks_54745.png
diff --git a/python-package/insightface/data/images/mask_black.jpg b/python-package/insightface/data/images/mask_black.jpg
new file mode 100644
index 000000000..0eab0df55
Binary files /dev/null and b/python-package/insightface/data/images/mask_black.jpg differ
diff --git a/python-package/insightface/data/images/mask_blue.jpg b/python-package/insightface/data/images/mask_blue.jpg
new file mode 100755
index 000000000..f71336b9a
Binary files /dev/null and b/python-package/insightface/data/images/mask_blue.jpg differ
diff --git a/python-package/insightface/data/images/mask_green.jpg b/python-package/insightface/data/images/mask_green.jpg
new file mode 100755
index 000000000..ac2ad55f4
Binary files /dev/null and b/python-package/insightface/data/images/mask_green.jpg differ
diff --git a/python-package/insightface/data/images/mask_white.jpg b/python-package/insightface/data/images/mask_white.jpg
new file mode 100644
index 000000000..2148ab2d0
Binary files /dev/null and b/python-package/insightface/data/images/mask_white.jpg differ
diff --git a/sample-images/t1.jpg b/python-package/insightface/data/images/t1.jpg
similarity index 100%
rename from sample-images/t1.jpg
rename to python-package/insightface/data/images/t1.jpg
diff --git a/python-package/insightface/data/objects/meanshape_68.pkl b/python-package/insightface/data/objects/meanshape_68.pkl
new file mode 100644
index 000000000..2af4d7602
Binary files /dev/null and b/python-package/insightface/data/objects/meanshape_68.pkl differ
diff --git a/python-package/insightface/data/pickle_object.py b/python-package/insightface/data/pickle_object.py
new file mode 100644
index 000000000..f028063e5
--- /dev/null
+++ b/python-package/insightface/data/pickle_object.py
@@ -0,0 +1,27 @@
+import sys
+import os
+import os.path as osp
+from pathlib import Path
+import pickle
+
+def get_object(name):
+    if getattr(sys, 'frozen', False):
+        base_dir = sys._MEIPASS
+    else:
+        base_dir = Path(__file__).parent.absolute()
+
+    objects_dir = osp.join(base_dir, 'objects')
+
+    if not name.endswith('.pkl'):
+        name = name + ".pkl"
+
+    filepath = osp.join(objects_dir, name)
+    
+    if not osp.exists(filepath):
+        print(f"[Error] File not found: {filepath}")
+        return None
+
+    with open(filepath, 'rb') as f:
+        obj = pickle.load(f)
+
+    return obj
diff --git a/python-package/insightface/data/rec_builder.py b/python-package/insightface/data/rec_builder.py
new file mode 100644
index 000000000..e02abc969
--- /dev/null
+++ b/python-package/insightface/data/rec_builder.py
@@ -0,0 +1,71 @@
+import pickle
+import numpy as np
+import os
+import os.path as osp
+import sys
+import mxnet as mx
+
+
+class RecBuilder():
+    def __init__(self, path, image_size=(112, 112)):
+        self.path = path
+        self.image_size = image_size
+        self.widx = 0
+        self.wlabel = 0
+        self.max_label = -1
+        assert not osp.exists(path), '%s exists' % path
+        os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(os.path.join(path, 'train.idx'), 
+                                                    os.path.join(path, 'train.rec'),
+                                                    'w')
+        self.meta = []
+
+    def add(self, imgs):
+        #!!! img should be BGR!!!!
+        #assert label >= 0
+        #assert label > self.last_label
+        assert len(imgs) > 0
+        label = self.wlabel
+        for img in imgs:
+            idx = self.widx
+            image_meta = {'image_index': idx, 'image_classes': [label]}
+            header = mx.recordio.IRHeader(0, label, idx, 0)
+            if isinstance(img, np.ndarray):
+                s = mx.recordio.pack_img(header,img,quality=95,img_fmt='.jpg')
+            else:
+                s = mx.recordio.pack(header, img)
+            self.writer.write_idx(idx, s)
+            self.meta.append(image_meta)
+            self.widx += 1
+        self.max_label = label
+        self.wlabel += 1
+
+
+    def add_image(self, img, label):
+        #!!! img should be BGR!!!!
+        #assert label >= 0
+        #assert label > self.last_label
+        idx = self.widx
+        header = mx.recordio.IRHeader(0, label, idx, 0)
+        if isinstance(label, list):
+            idlabel = label[0]
+        else:
+            idlabel = label
+        image_meta = {'image_index': idx, 'image_classes': [idlabel]}
+        if isinstance(img, np.ndarray):
+            s = mx.recordio.pack_img(header,img,quality=95,img_fmt='.jpg')
+        else:
+            s = mx.recordio.pack(header, img)
+        self.writer.write_idx(idx, s)
+        self.meta.append(image_meta)
+        self.widx += 1
+        self.max_label = max(self.max_label, idlabel)
+
+    def close(self):
+        with open(osp.join(self.path, 'train.meta'), 'wb') as pfile:
+            pickle.dump(self.meta, pfile, protocol=pickle.HIGHEST_PROTOCOL)
+        print('stat:', self.widx, self.wlabel)
+        with open(os.path.join(self.path, 'property'), 'w') as f:
+            f.write("%d,%d,%d\n" % (self.max_label+1, self.image_size[0], self.image_size[1]))
+            f.write("%d\n" % (self.widx))
+
diff --git a/python-package/insightface/gui/__init__.py b/python-package/insightface/gui/__init__.py
new file mode 100644
index 000000000..c326e5eff
--- /dev/null
+++ b/python-package/insightface/gui/__init__.py
@@ -0,0 +1,12 @@
+"""InsightFace Evaluation Studio.
+
+The GUI package is intentionally import-light. Importing ``insightface.gui``
+does not require PySide6; GUI dependencies are loaded by the entry point.
+"""
+
+__version__ = "1.0.1"
+
+APP_NAME = "InsightFace Evaluation Studio"
+APP_DISPLAY_NAME = "InsightFace Evaluation Studio v1.0.1"
+
+__all__ = ["__version__", "APP_NAME", "APP_DISPLAY_NAME"]
diff --git a/python-package/insightface/gui/__main__.py b/python-package/insightface/gui/__main__.py
new file mode 100644
index 000000000..3a97595e1
--- /dev/null
+++ b/python-package/insightface/gui/__main__.py
@@ -0,0 +1,46 @@
+"""Command line entry point for InsightFace Evaluation Studio."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+
+from . import __version__
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description="InsightFace Evaluation Studio")
+    parser.add_argument("--workspace", help="Use a specific local workspace directory")
+    parser.add_argument("--model", help="Model pack name or local model directory")
+    parser.add_argument("--provider", choices=["auto", "cpu", "cuda", "Auto", "CPU", "CUDA"], help="Execution provider")
+    parser.add_argument("--safe-mode", action="store_true", help="Open the GUI without loading a model")
+    parser.add_argument("--version", action="store_true", help="Print version information and exit")
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = build_parser()
+    args = parser.parse_args(argv)
+    if args.version:
+        try:
+            import insightface
+
+            insightface_version = getattr(insightface, "__version__", "1.0.1")
+        except Exception:
+            insightface_version = "1.0.1"
+        print(f"InsightFace Evaluation Studio {__version__}")
+        print(f"insightface {insightface_version}")
+        return 0
+    try:
+        from .app import run_app
+    except ImportError as exc:
+        if "PySide6" in str(exc):
+            print("InsightFace GUI requires PySide6.")
+            print("Please install with: pip install insightface[gui]")
+            return 1
+        raise
+    return run_app(args)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main(sys.argv[1:]))
diff --git a/python-package/insightface/gui/app.py b/python-package/insightface/gui/app.py
new file mode 100644
index 000000000..7373542df
--- /dev/null
+++ b/python-package/insightface/gui/app.py
@@ -0,0 +1,162 @@
+"""Application bootstrap for InsightFace Evaluation Studio."""
+
+from __future__ import annotations
+
+import os
+import sys
+from dataclasses import dataclass
+from typing import Iterable, Optional
+
+from pathlib import Path
+
+from .core.config import AppConfig, load_config, save_config
+from .core.face_engine import FaceEngine, is_cuda_provider_available, providers_from_choice
+from .core.logging import setup_logging
+from .core.storage import Storage
+
+
+@dataclass
+class StudioContext:
+    config: AppConfig
+    config_exists: bool
+    storage: Storage
+    engine: FaceEngine
+    log_file: str
+    runtime_safe_mode: bool = False
+
+
+def _unique_existing_paths(paths: Iterable[Path]) -> list[Path]:
+    seen = set()
+    unique_paths = []
+    for path in paths:
+        resolved = str(path)
+        if resolved in seen:
+            continue
+        seen.add(resolved)
+        if path.exists():
+            unique_paths.append(path)
+    return unique_paths
+
+
+def qt_plugin_root_candidates() -> list[Path]:
+    """Return plausible Qt plugin roots for PySide6 wheel and conda layouts."""
+
+    try:
+        import PySide6
+    except ImportError:
+        return []
+
+    candidates = []
+    for package_dir_raw in getattr(PySide6, "__path__", []):
+        package_dir = Path(package_dir_raw)
+        candidates.extend(
+            [
+                package_dir / "Qt" / "plugins",
+                package_dir / "plugins",
+            ]
+        )
+
+    try:
+        from PySide6.QtCore import QLibraryInfo
+
+        candidates.append(Path(QLibraryInfo.path(QLibraryInfo.LibraryPath.PluginsPath)))
+    except Exception:
+        pass
+
+    candidates.extend(
+        [
+            Path(sys.prefix) / "plugins",
+            Path(sys.prefix) / "Library" / "plugins",
+            Path(sys.prefix) / "lib" / "qt6" / "plugins",
+        ]
+    )
+    return _unique_existing_paths(candidates)
+
+
+def configure_qt_plugin_paths() -> None:
+    """Point Qt at PySide6-Essentials plugins when the PySide6 meta package is absent."""
+
+    for plugins in qt_plugin_root_candidates():
+        platforms = plugins / "platforms"
+        if platforms.exists():
+            if not os.environ.get("QT_PLUGIN_PATH"):
+                os.environ["QT_PLUGIN_PATH"] = str(plugins)
+            if not os.environ.get("QT_QPA_PLATFORM_PLUGIN_PATH"):
+                os.environ["QT_QPA_PLATFORM_PLUGIN_PATH"] = str(platforms)
+            return
+
+
+def create_context(args=None) -> StudioContext:
+    config_path = None
+    runtime_safe_mode = bool(args is not None and getattr(args, "safe_mode", False))
+    if args is not None and getattr(args, "workspace", None):
+        config_path = Path(args.workspace).expanduser() / "config.json"
+    config, exists = load_config(config_path)
+    if args is not None:
+        if getattr(args, "workspace", None):
+            config.workspace_path = args.workspace
+            config.database_path = ""
+            config.crop_dir = ""
+            config.export_dir = ""
+            config.report_dir = ""
+            config.log_dir = ""
+            config.cache_dir = ""
+            config.apply_workspace_defaults()
+        if getattr(args, "model", None):
+            config.model_name = args.model
+        if getattr(args, "provider", None):
+            value = str(args.provider).upper()
+            if value == "CPU":
+                config.provider = "CPU"
+            elif value == "CUDA":
+                config.provider = "CUDA" if is_cuda_provider_available() else "Auto"
+            else:
+                config.provider = "Auto"
+    if config.safe_mode:
+        # safe-mode is a startup troubleshooting flag. Older builds persisted it
+        # into config.json, which made normal launches silently skip model load.
+        config.safe_mode = False
+        if not config.auto_load_model:
+            config.auto_load_model = True
+    if str(config.provider).strip().lower() == "cuda" and not is_cuda_provider_available():
+        config.provider = "Auto"
+    config.apply_workspace_defaults()
+    save_config(config)
+    log_file = setup_logging(config.log_dir)
+    storage = Storage(config.database_path)
+    engine = FaceEngine(
+        model_name=config.model_name,
+        providers=providers_from_choice(config.provider),
+        det_size=config.det_size_tuple,
+        root=config.model_root,
+        custom_model_dir=config.custom_model_dir,
+    )
+    return StudioContext(
+        config=config,
+        config_exists=exists,
+        storage=storage,
+        engine=engine,
+        log_file=str(log_file),
+        runtime_safe_mode=runtime_safe_mode,
+    )
+
+
+def run_app(args=None) -> int:
+    configure_qt_plugin_paths()
+    try:
+        from PySide6.QtWidgets import QApplication
+    except ImportError:
+        print("InsightFace GUI requires PySide6.")
+        print("Please install with: pip install insightface[gui]")
+        return 1
+
+    from .main_window import MainWindow
+    from .resources import configure_application_metadata
+
+    app = QApplication.instance() or QApplication(sys.argv[:1])
+    configure_application_metadata(app)
+    context = create_context(args)
+    window = MainWindow(context)
+    window.resize(1320, 860)
+    window.show()
+    return int(app.exec())
diff --git a/python-package/insightface/gui/assets/app_icon.icns b/python-package/insightface/gui/assets/app_icon.icns
new file mode 100644
index 000000000..a555bc79b
Binary files /dev/null and b/python-package/insightface/gui/assets/app_icon.icns differ
diff --git a/python-package/insightface/gui/assets/app_icon.ico b/python-package/insightface/gui/assets/app_icon.ico
new file mode 100644
index 000000000..464538b5c
Binary files /dev/null and b/python-package/insightface/gui/assets/app_icon.ico differ
diff --git a/python-package/insightface/gui/assets/app_icon.png b/python-package/insightface/gui/assets/app_icon.png
new file mode 100644
index 000000000..c64d1e677
Binary files /dev/null and b/python-package/insightface/gui/assets/app_icon.png differ
diff --git a/python-package/insightface/gui/core/__init__.py b/python-package/insightface/gui/core/__init__.py
new file mode 100644
index 000000000..411634f51
--- /dev/null
+++ b/python-package/insightface/gui/core/__init__.py
@@ -0,0 +1 @@
+"""Core services for InsightFace Evaluation Studio."""
diff --git a/python-package/insightface/gui/core/camera.py b/python-package/insightface/gui/core/camera.py
new file mode 100644
index 000000000..4fe6088a2
--- /dev/null
+++ b/python-package/insightface/gui/core/camera.py
@@ -0,0 +1,19 @@
+"""Camera availability helpers."""
+
+from __future__ import annotations
+
+from typing import List
+
+
+def list_camera_indices(max_indices: int = 4) -> List[int]:
+    try:
+        import cv2
+    except Exception:
+        return []
+    indices = []
+    for idx in range(max_indices):
+        cap = cv2.VideoCapture(idx)
+        if cap.isOpened():
+            indices.append(idx)
+        cap.release()
+    return indices
diff --git a/python-package/insightface/gui/core/clustering.py b/python-package/insightface/gui/core/clustering.py
new file mode 100644
index 000000000..36eefb710
--- /dev/null
+++ b/python-package/insightface/gui/core/clustering.py
@@ -0,0 +1,65 @@
+"""Face embedding clustering helpers."""
+
+from __future__ import annotations
+
+from typing import Dict, Iterable, List
+
+import numpy as np
+
+from .recognition import cosine_similarity, normalize_embedding
+
+
+def cluster_embeddings(
+    embeddings: Iterable[np.ndarray],
+    threshold: float = 0.72,
+    min_samples: int = 2,
+) -> List[int]:
+    labels, _ = cluster_embeddings_dbscan(
+        embeddings,
+        distance_threshold=max(0.01, 1.0 - float(threshold)),
+        min_samples=min_samples,
+    )
+    return labels
+
+
+def cluster_embeddings_dbscan(
+    embeddings: Iterable[np.ndarray],
+    distance_threshold: float = 0.28,
+    min_samples: int = 2,
+) -> tuple[List[int], str]:
+    normalized = [normalize_embedding(embedding) for embedding in embeddings]
+    vectors = [embedding for embedding in normalized if embedding is not None]
+    if not vectors:
+        return [], "none"
+    matrix = np.vstack(vectors)
+    eps = max(0.01, float(distance_threshold))
+    try:
+        from sklearn.cluster import DBSCAN
+
+        labels = DBSCAN(eps=eps, min_samples=max(1, int(min_samples)), metric="cosine").fit_predict(matrix)
+        return [int(label) for label in labels], "DBSCAN"
+    except Exception:
+        labels: List[int] = []
+        centroids: Dict[int, np.ndarray] = {}
+        counts: Dict[int, int] = {}
+        next_label = 0
+        for vector in matrix:
+            best_label = None
+            best_score = -1.0
+            for label, centroid in centroids.items():
+                score = cosine_similarity(vector, centroid)
+                if score > best_score:
+                    best_label = label
+                    best_score = score
+            if best_label is not None and (1.0 - best_score) <= eps:
+                labels.append(best_label)
+                counts[best_label] += 1
+                centroids[best_label] = normalize_embedding(
+                    centroids[best_label] * (counts[best_label] - 1) + vector
+                )
+            else:
+                labels.append(next_label)
+                centroids[next_label] = vector
+                counts[next_label] = 1
+                next_label += 1
+        return labels, "centroid fallback"
diff --git a/python-package/insightface/gui/core/config.py b/python-package/insightface/gui/core/config.py
new file mode 100644
index 000000000..48a6248d6
--- /dev/null
+++ b/python-package/insightface/gui/core/config.py
@@ -0,0 +1,123 @@
+"""Configuration loading and saving."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from .constants import (
+    DEFAULT_DET_SIZE,
+    DEFAULT_LICENSE_STATUS,
+    DEFAULT_MODEL_NAME,
+    DEFAULT_PROVIDER,
+    DEFAULT_THRESHOLD,
+    DEFAULT_TOP_K,
+)
+from .paths import default_config_path, default_workspace, ensure_workspace, expand_path
+
+
+@dataclass
+class AppConfig:
+    workspace_path: str = str(default_workspace())
+    database_path: str = ""
+    crop_dir: str = ""
+    export_dir: str = ""
+    report_dir: str = ""
+    log_dir: str = ""
+    cache_dir: str = ""
+    mode: str = "Personal / Research"
+    model_name: str = DEFAULT_MODEL_NAME
+    model_root: str = "~/.insightface"
+    custom_model_dir: str = ""
+    provider: str = DEFAULT_PROVIDER
+    det_size: list[int] = None  # type: ignore[assignment]
+    recognition_threshold: float = DEFAULT_THRESHOLD
+    default_top_k: int = DEFAULT_TOP_K
+    min_detection_score: float = 0.45
+    min_face_size: int = 32
+    save_crops: bool = True
+    save_recognition_logs: bool = True
+    anonymize_report_paths: bool = False
+    batch_worker_count: int = 2
+    video_frame_interval: int = 10
+    camera_frame_skip: int = 3
+    ui_theme: str = "azure_lab"
+    ui_language: str = "system"
+    ui_default_mode: str = "face_verification"
+    ui_last_mode: str = "face_verification"
+    ui_last_page_face_verification: str = "verification"
+    ui_last_page_album_management: str = "album"
+    ui_last_page_face_swap: str = "image_face_swap"
+    ui_last_page_enterprise_evaluation: str = "enterprise_dashboard"
+    ui_window_width: int = 1320
+    ui_window_height: int = 860
+    ui_sidebar_width: int = 240
+    ui_settings_dialog_last_tab: int = 0
+    ui_sidebar_compact: bool = False
+    ui_show_status_chips: bool = True
+    license_status: str = DEFAULT_LICENSE_STATUS
+    safe_mode: bool = False
+    auto_load_model: bool = True
+    swap_model_path: str = ""
+    enable_gfpgan: bool = False
+    gfpgan_model_path: str = ""
+
+    def __post_init__(self) -> None:
+        if self.det_size is None:
+            self.det_size = [DEFAULT_DET_SIZE[0], DEFAULT_DET_SIZE[1]]
+        self.apply_workspace_defaults()
+
+    def apply_workspace_defaults(self) -> None:
+        paths = ensure_workspace(self.workspace_path)
+        self.workspace_path = str(paths["workspace"])
+        self.database_path = self.database_path or str(paths["database"])
+        self.crop_dir = self.crop_dir or str(paths["crops"])
+        self.export_dir = self.export_dir or str(paths["exports"])
+        self.report_dir = self.report_dir or str(paths["reports"])
+        self.log_dir = self.log_dir or str(paths["logs"])
+        self.cache_dir = self.cache_dir or str(paths["cache"])
+
+    @property
+    def det_size_tuple(self) -> tuple[int, int]:
+        return int(self.det_size[0]), int(self.det_size[1])
+
+    @property
+    def det_size_label(self) -> str:
+        width, height = self.det_size_tuple
+        if width <= 0 or height <= 0:
+            return "Auto"
+        return f"{width}x{height}"
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "AppConfig":
+        allowed = {field.name for field in cls.__dataclass_fields__.values()}  # type: ignore[attr-defined]
+        clean = {key: value for key, value in data.items() if key in allowed}
+        return cls(**clean)
+
+
+def load_config(path: Optional[str | Path] = None) -> tuple[AppConfig, bool]:
+    config_path = expand_path(path) if path else default_config_path()
+    if not config_path.exists():
+        if path is not None:
+            return AppConfig(workspace_path=str(config_path.parent)), False
+        return AppConfig(), False
+    try:
+        data = json.loads(config_path.read_text(encoding="utf-8"))
+        return AppConfig.from_dict(data), True
+    except Exception:
+        return AppConfig(), False
+
+
+def save_config(config: AppConfig, path: Optional[str | Path] = None) -> Path:
+    config.apply_workspace_defaults()
+    config_path = expand_path(path) if path else Path(config.workspace_path) / "config.json"
+    config_path.parent.mkdir(parents=True, exist_ok=True)
+    config_path.write_text(
+        json.dumps(config.to_dict(), indent=2, sort_keys=True), encoding="utf-8"
+    )
+    return config_path
diff --git a/python-package/insightface/gui/core/constants.py b/python-package/insightface/gui/core/constants.py
new file mode 100644
index 000000000..2488544f3
--- /dev/null
+++ b/python-package/insightface/gui/core/constants.py
@@ -0,0 +1,48 @@
+"""Shared constants and product copy."""
+
+from __future__ import annotations
+
+APP_NAME = "InsightFace Evaluation Studio"
+APP_VERSION = "1.0.1"
+APP_ORGANIZATION = "InsightFace"
+APP_DOMAIN = "insightface.ai"
+APP_ID = "ai.insightface.evaluationstudio"
+APP_PROCESS_NAME = APP_NAME
+WINDOW_TITLE = "InsightFace Evaluation Studio v1.0.1"
+SUBTITLE = (
+    "Local face recognition, face search, photo organization, model evaluation, "
+    "and face swap studio."
+)
+
+LOCAL_PROCESSING_NOTICE = (
+    "All processing is local by default. No images, videos, embeddings, or reports "
+    "are uploaded automatically."
+)
+
+LICENSE_NOTICE = (
+    "Code and model files may have different licenses. Commercial deployment "
+    "requires appropriate model authorization."
+)
+
+RESPONSIBLE_USE_NOTICE = (
+    "This tool does not provide legal advice. Users are responsible for consent, "
+    "privacy, retention, and compliance with applicable biometric regulations."
+)
+
+COMMERCIAL_NOTICE = (
+    "This evaluation may use research or non-commercial model files. Production "
+    "or commercial deployment requires an appropriate commercial model license. "
+    "Please contact InsightFace for commercial model licensing, private model "
+    "evaluation, SDK/API access, SLA, or custom training."
+)
+
+DEFAULT_MODEL_NAME = "buffalo_l"
+DEFAULT_PROVIDER = "Auto"
+DEFAULT_DET_SIZE = (0, 0)
+AUTO_DET_SIZES = ((128, 128), (640, 640))
+DEFAULT_THRESHOLD = 0.4
+DEFAULT_TOP_K = 5
+DEFAULT_LICENSE_STATUS = "Research / Non-commercial"
+
+IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".webp"}
+VIDEO_EXTENSIONS = {".mp4", ".mov", ".avi", ".mkv", ".webm", ".m4v"}
diff --git a/python-package/insightface/gui/core/evaluation.py b/python-package/insightface/gui/core/evaluation.py
new file mode 100644
index 000000000..fbcfa20bc
--- /dev/null
+++ b/python-package/insightface/gui/core/evaluation.py
@@ -0,0 +1,1165 @@
+"""No-code enterprise evaluation routines."""
+
+from __future__ import annotations
+
+import csv
+import math
+import platform
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+
+from .constants import DEFAULT_LICENSE_STATUS, DEFAULT_THRESHOLD, IMAGE_EXTENSIONS
+from .face_engine import FaceEngine
+from .models import EvaluationResult
+from .recognition import cosine_similarity
+from .utils import list_images, read_image, utc_now_iso
+
+MULTI_FACE_REQUIRE_ONE = "require_exactly_one"
+MULTI_FACE_USE_LARGEST = "use_largest_face"
+MULTI_FACE_USE_CENTERED_LARGEST = "use_centered_largest_face"
+MULTI_FACE_SKIP = "mark_as_skip"
+MULTI_FACE_POLICIES = {
+    MULTI_FACE_REQUIRE_ONE,
+    MULTI_FACE_USE_LARGEST,
+    MULTI_FACE_USE_CENTERED_LARGEST,
+    MULTI_FACE_SKIP,
+}
+
+
+@dataclass
+class DatasetValidationResult:
+    ok: bool
+    mode: str
+    auto_split: bool
+    multi_face_policy: str
+    root: str
+    summary: Dict[str, Any] = field(default_factory=dict)
+    errors: List[Dict[str, Any]] = field(default_factory=list)
+    warnings: List[Dict[str, Any]] = field(default_factory=list)
+
+
+def hardware_info() -> Dict[str, Any]:
+    return {
+        "platform": platform.platform(),
+        "machine": platform.machine(),
+        "processor": platform.processor(),
+        "python": platform.python_version(),
+    }
+
+
+def _metrics_at_threshold(rows: List[Dict[str, Any]], threshold: float) -> Dict[str, Any]:
+    positives = [row for row in rows if row.get("label") == 1 and row.get("similarity") is not None]
+    negatives = [row for row in rows if row.get("label") == 0 and row.get("similarity") is not None]
+    tp = sum(1 for row in positives if row["similarity"] >= threshold)
+    fn = sum(1 for row in positives if row["similarity"] < threshold)
+    fp = sum(1 for row in negatives if row["similarity"] >= threshold)
+    tn = sum(1 for row in negatives if row["similarity"] < threshold)
+    total = tp + tn + fp + fn
+    return {
+        "accuracy": (tp + tn) / total if total else 0.0,
+        "FAR": fp / max(1, len(negatives)),
+        "FRR": fn / max(1, len(positives)),
+        "TP": tp,
+        "TN": tn,
+        "FP": fp,
+        "FN": fn,
+    }
+
+
+def recommend_threshold(rows: List[Dict[str, Any]]) -> Optional[float]:
+    similarities = sorted({float(row["similarity"]) for row in rows if row.get("similarity") is not None})
+    if not similarities:
+        return None
+    best_threshold = similarities[0]
+    best_accuracy = -1.0
+    for threshold in similarities:
+        metrics = _metrics_at_threshold(rows, threshold)
+        if metrics["accuracy"] > best_accuracy:
+            best_accuracy = metrics["accuracy"]
+            best_threshold = threshold
+    return float(best_threshold)
+
+
+def tar_at_far(rows: List[Dict[str, Any]], target_far: float) -> Any:
+    positives = [row for row in rows if row.get("label") == 1 and row.get("similarity") is not None]
+    negatives = [row for row in rows if row.get("label") == 0 and row.get("similarity") is not None]
+    if len(negatives) < max(20, int(1 / max(target_far, 1e-9))):
+        return "insufficient data"
+    thresholds = sorted({float(row["similarity"]) for row in rows if row.get("similarity") is not None}, reverse=True)
+    best_tar = 0.0
+    for threshold in thresholds:
+        far = sum(1 for row in negatives if row["similarity"] >= threshold) / max(1, len(negatives))
+        if far <= target_far:
+            tar = sum(1 for row in positives if row["similarity"] >= threshold) / max(1, len(positives))
+            best_tar = max(best_tar, tar)
+    return best_tar
+
+
+def _identity_name(folder: Path) -> str:
+    return folder.name
+
+
+def _identity_dirs(root: Path) -> List[Path]:
+    if not root.exists() or not root.is_dir():
+        raise ValueError(f"Identity folder root not found: {root}")
+    return sorted(path for path in root.iterdir() if path.is_dir())
+
+
+def _auto_split_identity_images(identity_dir: Path) -> tuple[Path | None, List[Path]]:
+    images = sorted(list_images(identity_dir, recursive=True))
+    if not images:
+        return None, []
+    gallery_candidates = [path for path in images if "gallery" in path.stem.lower()]
+    gallery = sorted(gallery_candidates or images)[0]
+    probes = [path for path in images if path != gallery]
+    return gallery, probes
+
+
+def _normalize_multi_face_policy(policy: str) -> str:
+    policy = (policy or MULTI_FACE_REQUIRE_ONE).strip()
+    return policy if policy in MULTI_FACE_POLICIES else MULTI_FACE_REQUIRE_ONE
+
+
+def _face_bbox(face) -> np.ndarray:
+    bbox = getattr(face, "bbox", None)
+    if bbox is None:
+        bbox = face.bbox if hasattr(face, "bbox") else [0, 0, 0, 0]
+    return np.asarray(bbox, dtype=np.float32).reshape(-1)[:4]
+
+
+def _face_area(face) -> float:
+    try:
+        bbox = _face_bbox(face)
+        return max(0.0, float(bbox[2]) - float(bbox[0])) * max(0.0, float(bbox[3]) - float(bbox[1]))
+    except Exception:
+        return 0.0
+
+
+def _largest_centered_face(faces, image_shape):
+    if len(faces) == 1:
+        return faces[0]
+    try:
+        det = np.vstack([_face_bbox(face) for face in faces])
+        img_size = np.asarray(image_shape, dtype=np.float32)[0:2]
+        bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] - det[:, 1])
+        img_center = img_size / 2.0
+        offsets = np.vstack(
+            [
+                (det[:, 0] + det[:, 2]) / 2.0 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2.0 - img_center[0],
+            ]
+        )
+        offset_dist_squared = np.sum(np.power(offsets, 2.0), axis=0)
+        return faces[int(np.argmax(bounding_box_size - offset_dist_squared * 2.0))]
+    except Exception:
+        return max(faces, key=_face_area)
+
+
+def _select_face_from_faces(faces, image_shape, path: Path, policy: str, stage: str):
+    del path, stage
+    face_count = len(faces)
+    if face_count == 0:
+        raise ValueError("no face detected")
+    if face_count > 1 and policy == MULTI_FACE_REQUIRE_ONE:
+        raise ValueError(f"multiple faces detected ({face_count}); expected exactly one face")
+    if face_count > 1 and policy == MULTI_FACE_SKIP:
+        raise ValueError(f"skipped multi-face image ({face_count} faces)")
+    if face_count > 1 and policy == MULTI_FACE_USE_CENTERED_LARGEST:
+        return _largest_centered_face(faces, image_shape)
+    if face_count > 1:
+        return max(faces, key=_face_area)
+    return faces[0]
+
+
+def select_face_by_policy(faces, image_shape, policy: str = MULTI_FACE_REQUIRE_ONE, path: str | Path = "", stage: str = ""):
+    return _select_face_from_faces(
+        faces,
+        image_shape,
+        Path(path) if path else Path("image"),
+        _normalize_multi_face_policy(policy),
+        stage,
+    )
+
+
+def multi_face_policy_help(policy: str) -> str:
+    policy = _normalize_multi_face_policy(policy)
+    if policy == MULTI_FACE_REQUIRE_ONE:
+        return "Multi-face policy: require exactly one face. Images with multiple detected faces fail with a clear error."
+    if policy == MULTI_FACE_USE_LARGEST:
+        return "Multi-face policy: use largest face. If an image has multiple faces, the largest detected face is used."
+    if policy == MULTI_FACE_USE_CENTERED_LARGEST:
+        return (
+            "Multi-face policy: use largest centered face. If an image has multiple faces, the face with the best "
+            "area-minus-center-distance score is used."
+        )
+    return (
+        "Multi-face policy: mark as skip. Gallery images with multiple detected faces are skipped; a multi-face query "
+        "stops the current run."
+    )
+
+
+def _detect_faces_for_image(path: Path, engine: FaceEngine):
+    img = read_image(path)
+    if img is None:
+        raise ValueError("image read failure")
+    return img, engine.detect_faces(img, source_path=str(path))
+
+
+def _embedding_for_image(
+    path: Path,
+    engine: FaceEngine,
+    cache: Dict[str, np.ndarray],
+    errors: List[Dict[str, Any]],
+    stage: str,
+    multi_face_policy: str = MULTI_FACE_REQUIRE_ONE,
+) -> Optional[np.ndarray]:
+    key = str(path)
+    if key in cache:
+        return cache[key]
+    policy = _normalize_multi_face_policy(multi_face_policy)
+    try:
+        image, faces = _detect_faces_for_image(path, engine)
+        face = _select_face_from_faces(faces, image.shape, path, policy, stage)
+        if face is None or face.normed_embedding is None:
+            raise ValueError("no face or embedding")
+        embedding = np.asarray(face.normed_embedding, dtype=np.float32).reshape(-1)
+        cache[key] = embedding
+        return embedding
+    except Exception as exc:
+        errors.append({"path": key, "error": str(exc), "stage": stage})
+        return None
+
+
+def _collect_verification_specs(root: Path, auto_split: bool) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    identities = _identity_dirs(root)
+    specs: List[Dict[str, Any]] = []
+    structure_errors: List[Dict[str, Any]] = []
+    if auto_split:
+        for identity_dir in identities:
+            identity = _identity_name(identity_dir)
+            gallery, probes = _auto_split_identity_images(identity_dir)
+            if gallery is None:
+                structure_errors.append({"identity": identity, "error": "identity folder has no images"})
+                continue
+            specs.append({"identity": identity, "path": gallery, "role": "gallery", "selected_gallery": True})
+            specs.extend({"identity": identity, "path": path, "role": "probe"} for path in probes)
+            if not probes:
+                structure_errors.append(
+                    {"identity": identity, "error": "Auto Split requires at least one probe image after the gallery image"}
+                )
+    else:
+        for identity_dir in identities:
+            identity = _identity_name(identity_dir)
+            images = sorted(list_images(identity_dir, recursive=True))
+            if not images:
+                structure_errors.append({"identity": identity, "error": "identity folder has no images"})
+            specs.extend({"identity": identity, "path": path, "role": "probe"} for path in images)
+    return specs, structure_errors
+
+
+def _collect_identification_specs(root: Path, auto_split: bool) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    specs: List[Dict[str, Any]] = []
+    structure_errors: List[Dict[str, Any]] = []
+    if auto_split:
+        identities_root = _identity_root_for_auto_split(root)
+        identities = _identity_dirs(identities_root)
+        for identity_dir in identities:
+            identity = _identity_name(identity_dir)
+            gallery, probes = _auto_split_identity_images(identity_dir)
+            if gallery is None:
+                structure_errors.append({"identity": identity, "error": "identity folder has no images"})
+                continue
+            specs.append({"identity": identity, "path": gallery, "role": "gallery", "selected_gallery": True})
+            specs.extend({"identity": identity, "path": path, "role": "probe"} for path in probes)
+            if not probes:
+                structure_errors.append(
+                    {"identity": identity, "error": "Auto Split requires at least one probe image after the gallery image"}
+                )
+    else:
+        gallery_root = root / "gallery"
+        probe_root = root / "probe"
+        unknown_root = root / "unknown"
+        if not gallery_root.is_dir():
+            structure_errors.append({"path": str(gallery_root), "error": "1:N without Auto Split requires gallery/"})
+        if not probe_root.is_dir():
+            structure_errors.append({"path": str(probe_root), "error": "1:N without Auto Split requires probe/"})
+        if gallery_root.is_dir():
+            for identity_dir in _identity_dirs(gallery_root):
+                identity = _identity_name(identity_dir)
+                images = sorted(list_images(identity_dir, recursive=True))
+                if not images:
+                    structure_errors.append({"identity": identity, "role": "gallery", "error": "gallery identity has no images"})
+                specs.extend({"identity": identity, "path": path, "role": "gallery"} for path in images)
+        if probe_root.is_dir():
+            for identity_dir in _identity_dirs(probe_root):
+                identity = _identity_name(identity_dir)
+                images = sorted(list_images(identity_dir, recursive=True))
+                if not images:
+                    structure_errors.append({"identity": identity, "role": "probe", "error": "probe identity has no images"})
+                specs.extend({"identity": identity, "path": path, "role": "probe"} for path in images)
+        if unknown_root.is_dir():
+            specs.extend({"identity": "Unknown", "path": path, "role": "unknown"} for path in sorted(list_images(unknown_root, recursive=True)))
+    return specs, structure_errors
+
+
+def _validate_image_spec(
+    spec: Dict[str, Any],
+    engine: FaceEngine,
+    policy: str,
+    mode_key: str,
+) -> Dict[str, Any]:
+    path = Path(spec["path"])
+    row = {
+        "identity": spec.get("identity", ""),
+        "path": str(path),
+        "role": spec.get("role", ""),
+        "valid": False,
+        "skipped": False,
+        "face_count": 0,
+    }
+    try:
+        image, faces = _detect_faces_for_image(path, engine)
+        del image
+        row["face_count"] = len(faces)
+        if not faces:
+            row["error"] = "no face detected"
+            return row
+        if len(faces) > 1:
+            row["multi_face"] = True
+            if policy == MULTI_FACE_REQUIRE_ONE:
+                row["error"] = f"multiple faces detected ({len(faces)}); expected exactly one face"
+                return row
+            if policy == MULTI_FACE_SKIP:
+                row["skipped"] = True
+                row["warning"] = f"multi-face image skipped ({len(faces)} faces)"
+                if spec.get("role") == "gallery" or spec.get("selected_gallery"):
+                    row["error"] = (
+                        "multi-face gallery image would be skipped, leaving this gallery sample unavailable"
+                    )
+                return row
+            if policy == MULTI_FACE_USE_CENTERED_LARGEST:
+                row["warning"] = f"multiple faces detected ({len(faces)}); largest centered face will be used"
+            else:
+                row["warning"] = f"multiple faces detected ({len(faces)}); largest face will be used"
+        row["valid"] = True
+        return row
+    except Exception as exc:
+        row["error"] = str(exc)
+        return row
+
+
+def _quick_validation_specs(specs: List[Dict[str, Any]], policy: str) -> List[Dict[str, Any]]:
+    if policy == MULTI_FACE_REQUIRE_ONE:
+        return specs
+    gallery_specs = [spec for spec in specs if spec.get("role") == "gallery"]
+    if gallery_specs:
+        return gallery_specs
+    representatives: List[Dict[str, Any]] = []
+    seen = set()
+    for spec in specs:
+        identity = spec.get("identity", "")
+        if identity in seen:
+            continue
+        seen.add(identity)
+        representatives.append(spec)
+    return representatives
+
+
+def _validation_scope_label(policy: str, specs_to_check: List[Dict[str, Any]], specs: List[Dict[str, Any]]) -> str:
+    if policy == MULTI_FACE_REQUIRE_ONE:
+        return "all images"
+    if len(specs_to_check) == len(specs):
+        return "all gallery images"
+    return "gallery or representative images only"
+
+
+def validate_enterprise_dataset(
+    dataset_root: str | Path,
+    mode: str,
+    auto_split: bool,
+    engine: FaceEngine,
+    multi_face_policy: str = MULTI_FACE_REQUIRE_ONE,
+    progress_callback=None,
+    cancel_callback=None,
+) -> DatasetValidationResult:
+    root = Path(dataset_root).expanduser()
+    policy = _normalize_multi_face_policy(multi_face_policy)
+    result = DatasetValidationResult(
+        ok=False,
+        mode=mode,
+        auto_split=bool(auto_split),
+        multi_face_policy=policy,
+        root=str(root),
+    )
+    if not root.is_dir():
+        result.errors.append({"path": str(root), "error": "dataset root not found"})
+        return result
+
+    mode_key = "1n" if mode.strip().startswith("1:N") else "1to1"
+    try:
+        specs, structure_errors = (
+            _collect_identification_specs(root, auto_split)
+            if mode_key == "1n"
+            else _collect_verification_specs(root, auto_split)
+        )
+    except Exception as exc:
+        result.errors.append({"path": str(root), "error": str(exc)})
+        return result
+    result.errors.extend(structure_errors)
+    if not specs:
+        result.errors.append({"path": str(root), "error": "no image files found for this evaluation mode"})
+    if result.errors:
+        result.summary.update(
+            {
+                "total_images": len(specs),
+                "images_checked": 0,
+                "valid_images": 0,
+                "skipped_images": 0,
+                "multi_face_images": 0,
+                "validation_scope": "structure only",
+                "errors": len(result.errors),
+                "warnings": len(result.warnings),
+            }
+        )
+        return result
+
+    rows: List[Dict[str, Any]] = []
+    specs_to_check = _quick_validation_specs(specs, policy)
+    for index, spec in enumerate(specs_to_check):
+        if cancel_callback and cancel_callback():
+            result.errors.append({"error": "validation cancelled"})
+            break
+        row = _validate_image_spec(spec, engine, policy, mode_key)
+        rows.append(row)
+        if row.get("error"):
+            result.errors.append({key: row[key] for key in ("identity", "path", "role", "face_count", "error") if key in row})
+        elif row.get("warning"):
+            result.warnings.append({key: row[key] for key in ("identity", "path", "role", "face_count", "warning") if key in row})
+        if progress_callback:
+            progress_callback(index + 1, max(1, len(specs_to_check)), f"Validated image {index + 1}/{len(specs_to_check)}")
+        if row.get("error"):
+            break
+
+    if result.errors:
+        result.summary.update(
+            {
+                "total_images": len(specs),
+                "images_checked": len(rows),
+                "valid_images": sum(1 for row in rows if row.get("valid")),
+                "skipped_images": sum(1 for row in rows if row.get("skipped")),
+                "multi_face_images": sum(1 for row in rows if row.get("multi_face")),
+                "validation_scope": _validation_scope_label(policy, specs_to_check, specs),
+                "errors": len(result.errors),
+                "warnings": len(result.warnings),
+            }
+        )
+        return result
+
+    if policy == MULTI_FACE_REQUIRE_ONE:
+        valid_specs = [
+            dict(spec, **{"path": str(spec["path"])})
+            for spec, row in zip(specs_to_check, rows)
+            if row.get("valid")
+        ]
+    else:
+        valid_specs = [dict(spec, **{"path": str(spec["path"])}) for spec in specs]
+    valid_galleries = [spec for spec in valid_specs if spec.get("role") == "gallery"]
+    valid_probes = [spec for spec in valid_specs if spec.get("role") == "probe"]
+
+    if mode_key == "1to1":
+        if auto_split:
+            gallery_by_identity = {spec["identity"] for spec in valid_galleries}
+            probe_by_identity = {spec["identity"] for spec in valid_probes}
+            expected_identities = {spec["identity"] for spec in specs}
+            for identity in sorted(expected_identities - gallery_by_identity):
+                result.errors.append({"identity": identity, "role": "gallery", "error": "identity has no valid gallery image"})
+            for identity in sorted(expected_identities - probe_by_identity):
+                result.errors.append({"identity": identity, "role": "probe", "error": "identity has no valid probe image"})
+            positive_pairs = sum(1 for probe in valid_probes for gallery in valid_galleries if probe["identity"] == gallery["identity"])
+            negative_pairs = sum(1 for probe in valid_probes for gallery in valid_galleries if probe["identity"] != gallery["identity"])
+            pair_count = positive_pairs + negative_pairs
+        else:
+            pair_count = positive_pairs = negative_pairs = 0
+            for left_index, left in enumerate(valid_probes):
+                for right in valid_probes[left_index + 1 :]:
+                    pair_count += 1
+                    if left["identity"] == right["identity"]:
+                        positive_pairs += 1
+                    else:
+                        negative_pairs += 1
+        if pair_count <= 0:
+            result.errors.append({"error": "no valid 1:1 comparison pairs can be generated"})
+        if positive_pairs <= 0:
+            result.errors.append({"error": "no positive 1:1 pairs can be generated"})
+        if negative_pairs <= 0:
+            result.errors.append({"error": "no negative 1:1 pairs can be generated"})
+        result.summary.update(
+            {
+                "valid_gallery_images": len(valid_galleries),
+                "valid_probe_images": len(valid_probes),
+                "comparison_pairs": pair_count,
+                "positive_pairs": positive_pairs,
+                "negative_pairs": negative_pairs,
+            }
+        )
+    else:
+        gallery_identities = {spec["identity"] for spec in valid_galleries}
+        probe_identities = {spec["identity"] for spec in valid_probes}
+        if not valid_galleries:
+            result.errors.append({"role": "gallery", "error": "1:N evaluation requires at least one valid gallery image"})
+        if not valid_probes:
+            result.errors.append({"role": "probe", "error": "1:N evaluation requires at least one valid known probe image"})
+        missing_gallery = sorted(probe_identities - gallery_identities)
+        for identity in missing_gallery:
+            result.errors.append({"identity": identity, "role": "gallery", "error": "probe identity has no valid gallery image"})
+        result.summary.update(
+            {
+                "valid_gallery_images": len(valid_galleries),
+                "gallery_identities": len(gallery_identities),
+                "valid_known_probe_images": len(valid_probes),
+                "probe_identities": len(probe_identities),
+                "unknown_images": sum(1 for spec in valid_specs if spec.get("role") == "unknown"),
+            }
+        )
+
+    result.summary.update(
+        {
+            "total_images": len(specs),
+            "images_checked": len(rows),
+            "valid_images": len(valid_specs),
+            "skipped_images": sum(1 for row in rows if row.get("skipped")),
+            "multi_face_images": sum(1 for row in rows if row.get("multi_face")),
+            "validation_scope": _validation_scope_label(policy, specs_to_check, specs),
+            "errors": len(result.errors),
+            "warnings": len(result.warnings),
+        }
+    )
+    result.ok = not result.errors
+    return result
+
+
+def _metrics_at_best_threshold(rows: List[Dict[str, Any]]) -> tuple[float, Dict[str, Any]]:
+    completed = [row for row in rows if row.get("similarity") is not None]
+    if not completed:
+        return 0.0, {"best_accuracy": 0.0, "best_cosine_threshold": 0.0}
+    positives = np.sort(
+        np.asarray([float(row["similarity"]) for row in completed if row.get("label") == 1], dtype=np.float64)
+    )
+    negatives = np.sort(
+        np.asarray([float(row["similarity"]) for row in completed if row.get("label") == 0], dtype=np.float64)
+    )
+    candidates = np.unique(np.asarray([float(row["similarity"]) for row in completed], dtype=np.float64))
+    positive_total = int(positives.size)
+    negative_total = int(negatives.size)
+    tp = positive_total - np.searchsorted(positives, candidates, side="left")
+    fn = positive_total - tp
+    tn = np.searchsorted(negatives, candidates, side="left")
+    fp = negative_total - tn
+    total = positive_total + negative_total
+    accuracy = (tp + tn) / max(1, total)
+    best_index = int(np.argmax(accuracy))
+    best_threshold = float(candidates[best_index])
+    best_accuracy = float(accuracy[best_index])
+    best_metrics = {
+        "accuracy": best_accuracy,
+        "FAR": float(fp[best_index] / max(1, negative_total)),
+        "FRR": float(fn[best_index] / max(1, positive_total)),
+        "TP": int(tp[best_index]),
+        "TN": int(tn[best_index]),
+        "FP": int(fp[best_index]),
+        "FN": int(fn[best_index]),
+    }
+    best_metrics["best_accuracy"] = best_accuracy
+    best_metrics["best_cosine_threshold"] = best_threshold
+    return best_threshold, best_metrics
+
+
+def _tar_far_from_scores(
+    positive_scores: List[float],
+    positive_total: int,
+    negative_scores: List[float],
+    far_targets: List[float],
+) -> Dict[str, Any]:
+    metrics: Dict[str, Any] = {
+        "positive_trials_for_tar": int(positive_total),
+        "negative_trials_for_far": len(negative_scores),
+    }
+    far_targets_with_budget = [
+        far for far in far_targets if int(math.floor(far * len(negative_scores))) > 0
+    ]
+    if positive_total <= 0 or not negative_scores:
+        for far in far_targets_with_budget:
+            label = _far_label(far)
+            metrics[f"TAR@FAR={label}"] = "insufficient data"
+            metrics[f"Threshold@FAR={label}"] = "insufficient data"
+        return metrics
+
+    positives = np.sort(np.asarray(positive_scores, dtype=np.float64))
+    negatives = np.sort(np.asarray(negative_scores, dtype=np.float64))
+    all_scores = np.asarray(positive_scores + negative_scores, dtype=np.float64)
+    max_negative = float(negatives[-1])
+    candidates = np.unique(np.concatenate(([np.nextafter(max_negative, np.inf)], all_scores)))
+    positive_accepts = positive_total - np.searchsorted(positives, candidates, side="left")
+    false_accepts_by_threshold = len(negative_scores) - np.searchsorted(negatives, candidates, side="left")
+    for far in far_targets_with_budget:
+        label = _far_label(far)
+        allowed_fp = int(math.floor(far * len(negative_scores)))
+        valid = false_accepts_by_threshold <= allowed_fp
+        if not np.any(valid):
+            metrics[f"TAR@FAR={label}"] = 0.0
+            metrics[f"Threshold@FAR={label}"] = float(np.nextafter(max_negative, np.inf))
+            metrics[f"Achieved FAR@{label}"] = 0.0
+            metrics[f"False accept budget@{label}"] = allowed_fp
+            continue
+        valid_positive_accepts = positive_accepts[valid]
+        best_accepts = int(np.max(valid_positive_accepts))
+        best_mask = valid & (positive_accepts == best_accepts)
+        best_threshold = float(np.min(candidates[best_mask]))
+        best_false_accepts = int(
+            false_accepts_by_threshold[np.where(candidates == best_threshold)[0][0]]
+        )
+        best_tar = best_accepts / max(1, positive_total)
+        best_far = best_false_accepts / max(1, len(negative_scores))
+        metrics[f"TAR@FAR={label}"] = max(0.0, best_tar)
+        metrics[f"Threshold@FAR={label}"] = best_threshold
+        metrics[f"Achieved FAR@{label}"] = best_far
+        metrics[f"False accept budget@{label}"] = allowed_fp
+    return metrics
+
+
+def _far_label(value: float) -> str:
+    mantissa, exponent = f"{value:.0e}".split("e")
+    return f"{mantissa}e{int(exponent)}"
+
+
+def run_identity_verification_evaluation(
+    identity_root: str | Path,
+    engine: FaceEngine,
+    auto_split: bool = True,
+    multi_face_policy: str = MULTI_FACE_REQUIRE_ONE,
+    license_status: str = DEFAULT_LICENSE_STATUS,
+    progress_callback=None,
+    cancel_callback=None,
+) -> EvaluationResult:
+    root = Path(identity_root).expanduser()
+    identities = _identity_dirs(root)
+    errors: List[Dict[str, Any]] = []
+    cache: Dict[str, np.ndarray] = {}
+    pair_specs: List[Dict[str, Any]] = []
+
+    if auto_split:
+        galleries: List[Dict[str, Any]] = []
+        probes: List[Dict[str, Any]] = []
+        for identity_dir in identities:
+            gallery, identity_probes = _auto_split_identity_images(identity_dir)
+            if gallery is None:
+                continue
+            identity = _identity_name(identity_dir)
+            galleries.append({"identity": identity, "path": gallery})
+            probes.extend({"identity": identity, "path": path} for path in identity_probes)
+        for probe in probes:
+            for gallery in galleries:
+                pair_specs.append(
+                    {
+                        "image1_path": str(probe["path"]),
+                        "image2_path": str(gallery["path"]),
+                        "probe_identity": probe["identity"],
+                        "gallery_identity": gallery["identity"],
+                        "label": 1 if probe["identity"] == gallery["identity"] else 0,
+                    }
+                )
+    else:
+        images: List[Dict[str, Any]] = []
+        for identity_dir in identities:
+            identity = _identity_name(identity_dir)
+            images.extend({"identity": identity, "path": path} for path in sorted(list_images(identity_dir, recursive=True)))
+        for left_index, left in enumerate(images):
+            for right in images[left_index + 1 :]:
+                pair_specs.append(
+                    {
+                        "image1_path": str(left["path"]),
+                        "image2_path": str(right["path"]),
+                        "probe_identity": left["identity"],
+                        "gallery_identity": right["identity"],
+                        "label": 1 if left["identity"] == right["identity"] else 0,
+                    }
+                )
+
+    if not pair_specs:
+        raise ValueError("No verification pairs could be generated from the selected identity folders.")
+
+    start_all = time.perf_counter()
+    rows: List[Dict[str, Any]] = []
+    for index, spec in enumerate(pair_specs):
+        if cancel_callback and cancel_callback():
+            break
+        start = time.perf_counter()
+        row = dict(spec)
+        emb1 = _embedding_for_image(
+            Path(spec["image1_path"]),
+            engine,
+            cache,
+            errors,
+            "verification",
+            multi_face_policy=multi_face_policy,
+        )
+        emb2 = _embedding_for_image(
+            Path(spec["image2_path"]),
+            engine,
+            cache,
+            errors,
+            "verification",
+            multi_face_policy=multi_face_policy,
+        )
+        if emb1 is None or emb2 is None:
+            row.update({"similarity": None, "error": "embedding unavailable"})
+        else:
+            row.update(
+                {
+                    "similarity": cosine_similarity(emb1, emb2),
+                    "latency_ms": (time.perf_counter() - start) * 1000.0,
+                }
+            )
+        rows.append(row)
+        if progress_callback:
+            progress_callback(index + 1, len(pair_specs), f"Processed 1:1 pair {index + 1}/{len(pair_specs)}")
+
+    completed = [row for row in rows if row.get("similarity") is not None]
+    best_threshold, best_metrics = _metrics_at_best_threshold(completed)
+    positive_scores = [float(row["similarity"]) for row in completed if row.get("label") == 1]
+    negative_scores = [float(row["similarity"]) for row in completed if row.get("label") == 0]
+    metrics = {
+        "total_pairs": len(pair_specs),
+        "completed_pairs": len(completed),
+        "positive_pairs": sum(1 for row in pair_specs if row["label"] == 1),
+        "negative_pairs": sum(1 for row in pair_specs if row["label"] == 0),
+        "failed_pairs": len(pair_specs) - len(completed),
+        **best_metrics,
+        **_tar_far_from_scores(positive_scores, len(positive_scores), negative_scores, [1e-6, 1e-5, 1e-4, 1e-3]),
+        "average_latency_ms_per_pair": (
+            float(np.mean([row["latency_ms"] for row in completed if row.get("latency_ms") is not None]))
+            if completed
+            else 0.0
+        ),
+    }
+    return EvaluationResult(
+        scenario="Enterprise 1:1 Verification / Identity Folders",
+        model_name=engine.model_name,
+        provider=", ".join(engine.requested_providers),
+        threshold=best_threshold,
+        dataset_summary={
+            "identity_root": str(root),
+            "auto_split": bool(auto_split),
+            "multi_face_policy": _normalize_multi_face_policy(multi_face_policy),
+            "identities": len(identities),
+            "pairing_rule": (
+                "probe vs selected gallery for every identity"
+                if auto_split
+                else "all image pairs across identity folders"
+            ),
+        },
+        metrics=metrics,
+        errors=errors,
+        latency={"total_elapsed_ms": (time.perf_counter() - start_all) * 1000.0, "hardware": hardware_info()},
+        license_status=license_status,
+        created_at=utc_now_iso(),
+        raw_results=rows,
+        threshold_recommendation=best_threshold,
+    )
+
+
+def _identity_root_for_auto_split(root: Path) -> Path:
+    identities_root = root / "identities"
+    return identities_root if identities_root.is_dir() else root
+
+
+def _collect_gallery_probe_auto_split(root: Path) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]], List[Path]]:
+    gallery_items: List[Dict[str, Any]] = []
+    probe_items: List[Dict[str, Any]] = []
+    identities_root = _identity_root_for_auto_split(root)
+    for identity_dir in _identity_dirs(identities_root):
+        gallery, probes = _auto_split_identity_images(identity_dir)
+        if gallery is None:
+            continue
+        identity = _identity_name(identity_dir)
+        gallery_items.append({"identity": identity, "path": gallery})
+        probe_items.extend({"identity": identity, "path": path} for path in probes)
+    return gallery_items, probe_items, []
+
+
+def _collect_gallery_probe_structured(root: Path) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]], List[Path]]:
+    gallery_root = root / "gallery"
+    probe_root = root / "probe"
+    unknown_root = root / "unknown"
+    if not gallery_root.is_dir() or not probe_root.is_dir():
+        raise ValueError("1:N without Auto Split requires gallery/ and probe/ folders.")
+    gallery_items: List[Dict[str, Any]] = []
+    probe_items: List[Dict[str, Any]] = []
+    for identity_dir in _identity_dirs(gallery_root):
+        identity = _identity_name(identity_dir)
+        gallery_items.extend({"identity": identity, "path": path} for path in sorted(list_images(identity_dir, recursive=True)))
+    for identity_dir in _identity_dirs(probe_root):
+        identity = _identity_name(identity_dir)
+        probe_items.extend({"identity": identity, "path": path} for path in sorted(list_images(identity_dir, recursive=True)))
+    unknown_items = sorted(list_images(unknown_root, recursive=True)) if unknown_root.is_dir() else []
+    return gallery_items, probe_items, unknown_items
+
+
+def _rank_identities(query_embedding: np.ndarray, gallery: List[Dict[str, Any]]) -> List[tuple[str, float]]:
+    best_by_identity: Dict[str, float] = {}
+    for sample in gallery:
+        score = cosine_similarity(query_embedding, sample["embedding"])
+        identity = str(sample["identity"])
+        if identity not in best_by_identity or score > best_by_identity[identity]:
+            best_by_identity[identity] = score
+    return sorted(best_by_identity.items(), key=lambda item: item[1], reverse=True)
+
+
+def run_identity_identification_evaluation(
+    dataset_root: str | Path,
+    engine: FaceEngine,
+    auto_split: bool = False,
+    multi_face_policy: str = MULTI_FACE_REQUIRE_ONE,
+    license_status: str = DEFAULT_LICENSE_STATUS,
+    progress_callback=None,
+    cancel_callback=None,
+) -> EvaluationResult:
+    root = Path(dataset_root).expanduser()
+    if not root.is_dir():
+        raise ValueError(f"Dataset root not found: {root}")
+    gallery_items, probe_items, unknown_items = (
+        _collect_gallery_probe_auto_split(root) if auto_split else _collect_gallery_probe_structured(root)
+    )
+    if not gallery_items:
+        raise ValueError("1:N evaluation requires at least one gallery image.")
+    if not probe_items:
+        raise ValueError("1:N evaluation requires at least one known probe image.")
+
+    errors: List[Dict[str, Any]] = []
+    cache: Dict[str, np.ndarray] = {}
+    gallery: List[Dict[str, Any]] = []
+    start_all = time.perf_counter()
+    for index, item in enumerate(gallery_items):
+        embedding = _embedding_for_image(
+            Path(item["path"]),
+            engine,
+            cache,
+            errors,
+            "gallery",
+            multi_face_policy=multi_face_policy,
+        )
+        if embedding is not None:
+            gallery.append({"identity": item["identity"], "path": str(item["path"]), "embedding": embedding})
+        if progress_callback:
+            progress_callback(index + 1, len(gallery_items), f"Indexed gallery {index + 1}/{len(gallery_items)}")
+    if not gallery:
+        raise ValueError("No gallery embeddings could be extracted.")
+
+    rows: List[Dict[str, Any]] = []
+    known_total = len(probe_items)
+    for index, item in enumerate(probe_items):
+        if cancel_callback and cancel_callback():
+            break
+        embedding = _embedding_for_image(
+            Path(item["path"]),
+            engine,
+            cache,
+            errors,
+            "probe",
+            multi_face_policy=multi_face_policy,
+        )
+        if embedding is None:
+            rows.append({"probe_path": str(item["path"]), "ground_truth": item["identity"], "error": "embedding unavailable"})
+        else:
+            ranked = _rank_identities(embedding, gallery)
+            top1_identity, top1_score = ranked[0] if ranked else ("Unknown", 0.0)
+            top5 = [identity for identity, _score in ranked[:5]]
+            best_wrong = max([score for identity, score in ranked if identity != item["identity"]], default=None)
+            rows.append(
+                {
+                    "probe_path": str(item["path"]),
+                    "ground_truth": item["identity"],
+                    "top1": top1_identity,
+                    "top1_similarity": top1_score,
+                    "top5": top5,
+                    "correct_top1": top1_identity == item["identity"],
+                    "best_wrong_similarity": best_wrong,
+                    "label": "known",
+                }
+            )
+        if progress_callback:
+            progress_callback(index + 1, known_total, f"Processed known probe {index + 1}/{known_total}")
+
+    for index, path in enumerate(unknown_items):
+        if cancel_callback and cancel_callback():
+            break
+        embedding = _embedding_for_image(
+            path,
+            engine,
+            cache,
+            errors,
+            "unknown",
+            multi_face_policy=multi_face_policy,
+        )
+        if embedding is None:
+            rows.append({"probe_path": str(path), "ground_truth": "Unknown", "error": "embedding unavailable"})
+        else:
+            ranked = _rank_identities(embedding, gallery)
+            top1_identity, top1_score = ranked[0] if ranked else ("Unknown", 0.0)
+            rows.append(
+                {
+                    "probe_path": str(path),
+                    "ground_truth": "Unknown",
+                    "top1": top1_identity,
+                    "top1_similarity": top1_score,
+                    "top5": [identity for identity, _score in ranked[:5]],
+                    "correct_top1": False,
+                    "label": "unknown",
+                }
+            )
+        if progress_callback:
+            progress_callback(index + 1, max(1, len(unknown_items)), f"Processed unknown probe {index + 1}/{len(unknown_items)}")
+
+    known_rows = [row for row in rows if row.get("label") == "known" and row.get("top1_similarity") is not None]
+    correct_rows = [row for row in known_rows if row.get("correct_top1")]
+    positive_scores = [float(row["top1_similarity"]) for row in correct_rows]
+    negative_scores = [
+        float(row["top1_similarity"])
+        for row in rows
+        if row.get("label") == "unknown" and row.get("top1_similarity") is not None
+    ]
+    negative_scores.extend(
+        float(row["best_wrong_similarity"])
+        for row in known_rows
+        if row.get("best_wrong_similarity") is not None
+    )
+    metrics = {
+        "gallery_identities": len({item["identity"] for item in gallery}),
+        "gallery_images": len(gallery_items),
+        "gallery_embeddings": len(gallery),
+        "known_probe_images": len(probe_items),
+        "unknown_probe_images": len(unknown_items),
+        "completed_known_probes": len(known_rows),
+        "Top1": len(correct_rows) / max(1, len(known_rows)),
+        **_tar_far_from_scores(positive_scores, len(known_rows), negative_scores, [1e-5, 1e-4, 1e-3, 1e-2]),
+    }
+    threshold_recommendation = None
+    for key in ("Threshold@FAR=1e-3", "Threshold@FAR=1e-2", "Threshold@FAR=1e-4"):
+        value = metrics.get(key)
+        if isinstance(value, float):
+            threshold_recommendation = value
+            break
+    return EvaluationResult(
+        scenario="Enterprise 1:N Identification / Identity Folders",
+        model_name=engine.model_name,
+        provider=", ".join(engine.requested_providers),
+        threshold=threshold_recommendation or DEFAULT_THRESHOLD,
+        dataset_summary={
+            "dataset_root": str(root),
+            "auto_split": bool(auto_split),
+            "multi_face_policy": _normalize_multi_face_policy(multi_face_policy),
+            "structured_rule": "gallery/probe/unknown" if not auto_split else "identities auto split",
+            "gallery_required": True,
+        },
+        metrics=metrics,
+        errors=errors,
+        latency={"total_elapsed_ms": (time.perf_counter() - start_all) * 1000.0, "hardware": hardware_info()},
+        license_status=license_status,
+        created_at=utc_now_iso(),
+        raw_results=rows,
+        threshold_recommendation=threshold_recommendation,
+    )
+
+
+def run_kyc_pairs_evaluation(
+    pairs_csv: str | Path,
+    engine: FaceEngine,
+    threshold: float = DEFAULT_THRESHOLD,
+    license_status: str = DEFAULT_LICENSE_STATUS,
+    progress_callback=None,
+    cancel_callback=None,
+) -> EvaluationResult:
+    rows: List[Dict[str, Any]] = []
+    errors: List[Dict[str, Any]] = []
+    with Path(pairs_csv).open("r", newline="", encoding="utf-8-sig") as handle:
+        reader = csv.DictReader(handle)
+        pairs = list(reader)
+    start_all = time.perf_counter()
+    for index, row in enumerate(pairs):
+        if cancel_callback and cancel_callback():
+            break
+        image1_path = row.get("image1_path") or row.get("image1") or ""
+        image2_path = row.get("image2_path") or row.get("image2") or ""
+        label = int(row.get("label", "0"))
+        item: Dict[str, Any] = {"image1_path": image1_path, "image2_path": image2_path, "label": label}
+        start = time.perf_counter()
+        try:
+            img1 = read_image(image1_path)
+            img2 = read_image(image2_path)
+            if img1 is None or img2 is None:
+                raise ValueError("image read failure")
+            face1 = engine.detect_best_face(img1, source_path=image1_path)
+            face2 = engine.detect_best_face(img2, source_path=image2_path)
+            if face1 is None or face2 is None:
+                raise ValueError("failed detection")
+            if face1.normed_embedding is None or face2.normed_embedding is None:
+                raise ValueError("embedding unavailable")
+            similarity = cosine_similarity(face1.normed_embedding, face2.normed_embedding)
+            item.update(
+                {
+                    "similarity": similarity,
+                    "predicted": 1 if similarity >= threshold else 0,
+                    "latency_ms": (time.perf_counter() - start) * 1000.0,
+                }
+            )
+        except Exception as exc:
+            item.update({"similarity": None, "predicted": None, "error": str(exc)})
+            errors.append({"index": index, "error": str(exc), "row": dict(row)})
+        rows.append(item)
+        if progress_callback:
+            progress_callback(index + 1, len(pairs), f"Processed pair {index + 1}/{len(pairs)}")
+
+    completed = [row for row in rows if row.get("similarity") is not None]
+    metrics = _metrics_at_threshold(completed, threshold)
+    metrics.update(
+        {
+            "total_pairs": len(pairs),
+            "positive_pairs": sum(1 for row in rows if row.get("label") == 1),
+            "negative_pairs": sum(1 for row in rows if row.get("label") == 0),
+            "failed_detections": len(errors),
+            "TAR@FAR=1e-2": tar_at_far(completed, 1e-2),
+            "TAR@FAR=1e-3": tar_at_far(completed, 1e-3),
+            "TAR@FAR=1e-4": tar_at_far(completed, 1e-4),
+            "average_latency_ms_per_pair": (
+                float(np.mean([row["latency_ms"] for row in completed])) if completed else 0.0
+            ),
+        }
+    )
+    return EvaluationResult(
+        scenario="KYC / 1:1 Verification",
+        model_name=engine.model_name,
+        provider=", ".join(engine.requested_providers),
+        threshold=threshold,
+        dataset_summary={
+            "pairs_csv": str(pairs_csv),
+            "total_pairs": len(pairs),
+            "completed_pairs": len(completed),
+        },
+        metrics=metrics,
+        errors=errors,
+        latency={
+            "total_elapsed_ms": (time.perf_counter() - start_all) * 1000.0,
+            "hardware": hardware_info(),
+        },
+        license_status=license_status,
+        created_at=utc_now_iso(),
+        raw_results=rows,
+        threshold_recommendation=recommend_threshold(completed),
+    )
+
+
+def run_identification_evaluation(
+    gallery_folder: str | Path,
+    probe_folder: str | Path,
+    engine: FaceEngine,
+    threshold: float = DEFAULT_THRESHOLD,
+    ground_truth_csv: Optional[str | Path] = None,
+    license_status: str = DEFAULT_LICENSE_STATUS,
+    progress_callback=None,
+    cancel_callback=None,
+) -> EvaluationResult:
+    gallery: List[Dict[str, Any]] = []
+    errors: List[Dict[str, Any]] = []
+    gallery_root = Path(gallery_folder)
+    for person_dir in sorted(path for path in gallery_root.iterdir() if path.is_dir()):
+        for image_path in sorted(path for path in person_dir.rglob("*") if path.suffix.lower() in IMAGE_EXTENSIONS):
+            try:
+                img = read_image(image_path)
+                if img is None:
+                    raise ValueError("image read failure")
+                face = engine.detect_best_face(img, source_path=str(image_path))
+                if face is None or face.normed_embedding is None:
+                    raise ValueError("no face or embedding")
+                gallery.append(
+                    {
+                        "person_id": person_dir.name,
+                        "person_name": person_dir.name,
+                        "sample_id": len(gallery) + 1,
+                        "crop_path": "",
+                        "embedding": face.normed_embedding,
+                    }
+                )
+            except Exception as exc:
+                errors.append({"path": str(image_path), "error": str(exc), "stage": "gallery"})
+
+    ground_truth: Dict[str, str] = {}
+    if ground_truth_csv:
+        with Path(ground_truth_csv).open("r", newline="", encoding="utf-8-sig") as handle:
+            for row in csv.DictReader(handle):
+                ground_truth[row.get("image_path", "")] = row.get("person_name", "")
+
+    probes = list_images(probe_folder, recursive=True)
+    raw: List[Dict[str, Any]] = []
+    start_all = time.perf_counter()
+    for index, probe in enumerate(probes):
+        if cancel_callback and cancel_callback():
+            break
+        try:
+            img = read_image(probe)
+            if img is None:
+                raise ValueError("image read failure")
+            face = engine.detect_best_face(img, source_path=str(probe))
+            if face is None or face.normed_embedding is None:
+                raise ValueError("no face or embedding")
+            from .recognition import search_gallery
+
+            results = search_gallery(face.normed_embedding, gallery, top_k=5, threshold=threshold)
+            truth = ground_truth.get(str(probe), ground_truth.get(probe.name, ""))
+            row = {
+                "probe_path": str(probe),
+                "ground_truth": truth,
+                "top1": results[0].person_name if results else "Unknown",
+                "top1_similarity": results[0].similarity if results else 0.0,
+                "top5": [result.person_name for result in results],
+                "accepted": bool(results and results[0].similarity >= threshold),
+            }
+            raw.append(row)
+        except Exception as exc:
+            errors.append({"path": str(probe), "error": str(exc), "stage": "probe"})
+        if progress_callback:
+            progress_callback(index + 1, len(probes), f"Processed probe {index + 1}/{len(probes)}")
+
+    truth_rows = [row for row in raw if row.get("ground_truth")]
+    top1 = sum(1 for row in truth_rows if row["top1"] == row["ground_truth"]) / max(1, len(truth_rows))
+    top5 = sum(1 for row in truth_rows if row["ground_truth"] in row["top5"]) / max(1, len(truth_rows))
+    metrics = {
+        "gallery_persons": len({row["person_name"] for row in gallery}),
+        "gallery_face_samples": len(gallery),
+        "probe_images": len(probes),
+        "detected_probe_faces": len(raw),
+        "Top-1 accuracy": top1 if truth_rows else "ground truth not provided",
+        "Top-5 accuracy": top5 if truth_rows else "ground truth not provided",
+        "unknown_rejection_rate": sum(1 for row in raw if not row["accepted"]) / max(1, len(raw)),
+        "average_search_latency_ms": 0.0,
+    }
+    return EvaluationResult(
+        scenario="Access Control / 1:N Identification",
+        model_name=engine.model_name,
+        provider=", ".join(engine.requested_providers),
+        threshold=threshold,
+        dataset_summary={
+            "gallery_folder": str(gallery_folder),
+            "probe_folder": str(probe_folder),
+            "ground_truth_csv": str(ground_truth_csv or ""),
+        },
+        metrics=metrics,
+        errors=errors,
+        latency={"total_elapsed_ms": (time.perf_counter() - start_all) * 1000.0, "hardware": hardware_info()},
+        license_status=license_status,
+        created_at=utc_now_iso(),
+        raw_results=raw,
+    )
diff --git a/python-package/insightface/gui/core/exporters.py b/python-package/insightface/gui/core/exporters.py
new file mode 100644
index 000000000..ccd76f512
--- /dev/null
+++ b/python-package/insightface/gui/core/exporters.py
@@ -0,0 +1,51 @@
+"""Export helpers for JSON, CSV, Markdown, HTML, and annotated images."""
+
+from __future__ import annotations
+
+import csv
+import json
+from pathlib import Path
+from typing import Any, Iterable, Mapping
+
+from .utils import safe_json_dumps, save_image
+
+
+def export_json(path: str | Path, data: Any) -> Path:
+    out = Path(path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(data, indent=2, ensure_ascii=False, default=str), encoding="utf-8")
+    return out
+
+
+def export_csv(path: str | Path, rows: Iterable[Mapping[str, Any]]) -> Path:
+    out = Path(path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    rows = list(rows)
+    fieldnames = sorted({key for row in rows for key in row.keys()})
+    with out.open("w", newline="", encoding="utf-8") as handle:
+        writer = csv.DictWriter(handle, fieldnames=fieldnames)
+        writer.writeheader()
+        for row in rows:
+            writer.writerow({key: safe_json_dumps(value) if isinstance(value, (dict, list)) else value for key, value in row.items()})
+    return out
+
+
+def export_markdown(path: str | Path, text: str) -> Path:
+    out = Path(path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(text, encoding="utf-8")
+    return out
+
+
+def export_html(path: str | Path, html: str) -> Path:
+    out = Path(path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(html, encoding="utf-8")
+    return out
+
+
+def export_annotated_image(path: str | Path, image) -> Path:
+    out = Path(path)
+    if not save_image(out, image):
+        raise IOError(f"Unable to write image: {out}")
+    return out
diff --git a/python-package/insightface/gui/core/face_engine.py b/python-package/insightface/gui/core/face_engine.py
new file mode 100644
index 000000000..00f199203
--- /dev/null
+++ b/python-package/insightface/gui/core/face_engine.py
@@ -0,0 +1,420 @@
+"""InsightFace runtime wrapper used by the GUI."""
+
+from __future__ import annotations
+
+import glob
+import os
+import threading
+import time
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional
+
+import numpy as np
+
+from .constants import AUTO_DET_SIZES, DEFAULT_DET_SIZE, DEFAULT_MODEL_NAME, DEFAULT_THRESHOLD
+from .logging import get_logger
+from .models import CompareResult, FaceRecord
+from .quality import score_face
+from .recognition import compare_embeddings, cosine_similarity, normalize_embedding
+from .utils import crop_bbox
+
+LOGGER = get_logger("face_engine")
+
+
+class ModelNotLoadedError(RuntimeError):
+    pass
+
+
+class FaceEngine:
+    def __init__(
+        self,
+        model_name: str = DEFAULT_MODEL_NAME,
+        providers: Optional[Iterable[str]] = None,
+        det_size: tuple[int, int] = DEFAULT_DET_SIZE,
+        root: str | os.PathLike[str] | None = None,
+        custom_model_dir: str | os.PathLike[str] | None = None,
+    ):
+        self.model_name = model_name
+        self.root = Path(os.path.expanduser(str(root or "~/.insightface")))
+        self.custom_model_dir = Path(os.path.expanduser(str(custom_model_dir))) if custom_model_dir else None
+        self.requested_providers = list(providers or ["CPUExecutionProvider"])
+        self.det_size = det_size
+        self.det_sizes = self._resolve_det_sizes(det_size)
+        self.auto_det_size = len(self.det_sizes) > 1
+        self.models: Dict[str, Any] = {}
+        self.det_model = None
+        self.loaded = False
+        self.last_error = ""
+        self.embedding_dim: Optional[int] = None
+        self.last_latency_ms: Dict[str, float] = {}
+        self.active_providers: List[str] = []
+        self.ctx_id = -1
+        self._prepared_det_size = None
+        self._lock = threading.RLock()
+
+    def is_loaded(self) -> bool:
+        return self.loaded and self.det_model is not None
+
+    def resolve_model_dir(self) -> Path:
+        if self.custom_model_dir and self.custom_model_dir.exists():
+            return self.custom_model_dir
+        model_path = Path(os.path.expanduser(self.model_name))
+        if model_path.exists() and model_path.is_dir():
+            return model_path
+        return self.root / "models" / self.model_name
+
+    def load(self) -> None:
+        with self._lock:
+            self.loaded = False
+            self.last_error = ""
+            self.models.clear()
+            start = time.perf_counter()
+            model_dir = self.resolve_model_dir()
+            if not model_dir.exists():
+                self.last_error = (
+                    f"Model directory not found: {model_dir}. Configure a local model directory "
+                    "or install the model pack first."
+                )
+                LOGGER.warning(self.last_error)
+                return
+            onnx_files = sorted(glob.glob(str(model_dir / "*.onnx")))
+            if not onnx_files:
+                self.last_error = f"No .onnx model files found in {model_dir}."
+                LOGGER.warning(self.last_error)
+                return
+            try:
+                from insightface.app.common import Face
+            except Exception as exc:
+                self.last_error = f"InsightFace runtime import failed: {exc}"
+                LOGGER.exception(self.last_error)
+                return
+
+            del Face
+            try:
+                for onnx_file in onnx_files:
+                    model = self._load_model_from_onnx(onnx_file)
+                    if model is None:
+                        continue
+                    if model.taskname in self.models:
+                        continue
+                    self.models[model.taskname] = model
+                if "detection" not in self.models:
+                    self.last_error = f"No detection model found in {model_dir}."
+                    return
+                self.det_model = self.models["detection"]
+                ctx_id = 0 if any("CUDA" in provider for provider in self.requested_providers) else -1
+                self.ctx_id = ctx_id
+                for taskname, model in self.models.items():
+                    if taskname == "detection":
+                        self._prepare_detection()
+                    else:
+                        model.prepare(ctx_id)
+                self.loaded = True
+                self.embedding_dim = self._infer_embedding_dim()
+                self.active_providers = self.requested_providers
+                self.last_latency_ms["load"] = (time.perf_counter() - start) * 1000.0
+            except Exception as exc:
+                self.last_error = f"Model load failed: {exc}"
+                LOGGER.exception(self.last_error)
+                self.loaded = False
+
+    def _load_model_from_onnx(self, onnx_file: str):
+        """Load a model pack member, forcing GUI detection through SCRFD."""
+        try:
+            from insightface.model_zoo.arcface_onnx import ArcFaceONNX
+            from insightface.model_zoo.attribute import Attribute
+            from insightface.model_zoo.inswapper import INSwapper
+            from insightface.model_zoo.landmark import Landmark
+            from insightface.model_zoo.model_zoo import PickableInferenceSession
+            from insightface.model_zoo.scrfd import SCRFD
+        except Exception as exc:
+            raise RuntimeError(f"InsightFace model imports failed: {exc}") from exc
+
+        session_kwargs = {
+            "providers": self.requested_providers,
+            "provider_options": None,
+        }
+        session_options = self._quiet_onnxruntime_session_options()
+        if session_options is not None:
+            session_kwargs["sess_options"] = session_options
+        session = PickableInferenceSession(onnx_file, **session_kwargs)
+        LOGGER.info("Applied providers for %s: %s", onnx_file, getattr(session, "_providers", []))
+        inputs = session.get_inputs()
+        if not inputs:
+            return None
+        input_shape = inputs[0].shape
+        outputs = session.get_outputs()
+        input_height = self._shape_dim(input_shape, 2)
+        input_width = self._shape_dim(input_shape, 3)
+
+        if self._is_scrfd_detection_outputs(outputs):
+            LOGGER.info("GUI detection model routed to SCRFD: %s", onnx_file)
+            return SCRFD(model_file=onnx_file, session=session)
+        if input_height == 192 and input_width == 192:
+            return Landmark(model_file=onnx_file, session=session)
+        if input_height == 96 and input_width == 96:
+            return Attribute(model_file=onnx_file, session=session)
+        if len(inputs) == 2 and input_height == 128 and input_width == 128:
+            return INSwapper(model_file=onnx_file, session=session)
+        if (
+            input_height is not None
+            and input_width is not None
+            and input_height == input_width
+            and input_height >= 112
+            and input_height % 16 == 0
+        ):
+            return ArcFaceONNX(model_file=onnx_file, session=session)
+        return None
+
+    @staticmethod
+    def _shape_dim(shape, index: int) -> Optional[int]:
+        try:
+            value = shape[index]
+        except Exception:
+            return None
+        try:
+            return int(value)
+        except (TypeError, ValueError):
+            return None
+
+    @staticmethod
+    def _is_scrfd_detection_outputs(outputs) -> bool:
+        return len(outputs) in {6, 9, 10, 15}
+
+    @staticmethod
+    def _quiet_onnxruntime_session_options():
+        try:
+            import onnxruntime
+
+            onnxruntime.set_default_logger_severity(3)
+            session_options = onnxruntime.SessionOptions()
+            session_options.log_severity_level = 3
+            return session_options
+        except Exception:
+            return None
+
+    def _infer_embedding_dim(self) -> Optional[int]:
+        for model in self.models.values():
+            if getattr(model, "taskname", "") == "recognition":
+                output_shape = getattr(model, "output_shape", None)
+                if output_shape:
+                    try:
+                        return int(output_shape[-1])
+                    except Exception:
+                        return None
+        return None
+
+    def warmup(self) -> Dict[str, Any]:
+        if not self.is_loaded():
+            raise ModelNotLoadedError("Model is not loaded. Please open Models.")
+        warmup_width = max(size[0] for size in self.det_sizes)
+        warmup_height = max(size[1] for size in self.det_sizes)
+        image = np.zeros((warmup_height, warmup_width, 3), dtype=np.uint8)
+        start = time.perf_counter()
+        self.detect_faces(image)
+        elapsed = (time.perf_counter() - start) * 1000.0
+        self.last_latency_ms["warmup"] = elapsed
+        return {"warmup_ms": elapsed}
+
+    def detect_faces(self, image: np.ndarray, source_path: Optional[str] = None) -> List[FaceRecord]:
+        with self._lock:
+            if not self.is_loaded():
+                raise ModelNotLoadedError("Model is not loaded. Please open Models.")
+            if image is None:
+                return []
+            image = np.ascontiguousarray(image)
+            try:
+                from insightface.app.common import Face
+            except Exception as exc:
+                raise ModelNotLoadedError(f"InsightFace Face class import failed: {exc}") from exc
+
+            start = time.perf_counter()
+            bboxes, kpss = self._detect_raw(image)
+            records: List[FaceRecord] = []
+            if bboxes is None or bboxes.shape[0] == 0:
+                self.last_latency_ms["detect"] = (time.perf_counter() - start) * 1000.0
+                return records
+            for idx in range(bboxes.shape[0]):
+                bbox = bboxes[idx, 0:4]
+                det_score = float(bboxes[idx, 4])
+                kps = kpss[idx] if kpss is not None else None
+                face = Face(bbox=bbox, kps=kps, det_score=det_score)
+                for taskname, model in self.models.items():
+                    if taskname == "detection":
+                        continue
+                    model.get(image, face)
+                embedding = getattr(face, "embedding", None)
+                normed_embedding = getattr(face, "normed_embedding", None)
+                if normed_embedding is None and embedding is not None:
+                    normed_embedding = normalize_embedding(embedding)
+                crop = crop_bbox(image, bbox)
+                quality_score, quality_flags = score_face(image, bbox, kps, det_score)
+                records.append(
+                    FaceRecord(
+                        face_id=f"{source_path or 'image'}:{idx}",
+                        bbox=[float(v) for v in bbox],
+                        kps=kps.tolist() if kps is not None else None,
+                        det_score=det_score,
+                        embedding=np.asarray(embedding, dtype=np.float32).reshape(-1) if embedding is not None else None,
+                        normed_embedding=(
+                            np.asarray(normed_embedding, dtype=np.float32).reshape(-1)
+                            if normed_embedding is not None
+                            else None
+                        ),
+                        gender=getattr(face, "gender", None),
+                        age=getattr(face, "age", None),
+                        quality_score=quality_score,
+                        quality_flags=quality_flags,
+                        crop=crop,
+                        source_path=source_path,
+                    )
+                )
+            self.last_latency_ms["detect"] = (time.perf_counter() - start) * 1000.0
+            return records
+
+    def detect_best_face(self, image: np.ndarray, source_path: Optional[str] = None) -> Optional[FaceRecord]:
+        faces = self.detect_faces(image, source_path=source_path)
+        if not faces:
+            return None
+        return max(
+            faces,
+            key=lambda face: (
+                (face.bbox[2] - face.bbox[0]) * (face.bbox[3] - face.bbox[1]),
+                face.det_score,
+            ),
+        )
+
+    def compute_similarity(self, emb1: np.ndarray, emb2: np.ndarray) -> float:
+        return cosine_similarity(emb1, emb2)
+
+    def compare_images(
+        self,
+        image1: np.ndarray,
+        image2: np.ndarray,
+        threshold: float = DEFAULT_THRESHOLD,
+        path1: Optional[str] = None,
+        path2: Optional[str] = None,
+    ) -> CompareResult:
+        with self._lock:
+            face_a = self.detect_best_face(image1, source_path=path1)
+            face_b = self.detect_best_face(image2, source_path=path2)
+            if face_a is None:
+                raise ValueError("No face detected in Image A.")
+            if face_b is None:
+                raise ValueError("No face detected in Image B.")
+            if face_a.normed_embedding is None or face_b.normed_embedding is None:
+                raise ValueError("Recognition embedding is unavailable for one or both faces.")
+            comparison = compare_embeddings(face_a.normed_embedding, face_b.normed_embedding, threshold)
+            notes = []
+            for label, face in (("Image A", face_a), ("Image B", face_b)):
+                if face.quality_score is not None and face.quality_score < 0.45:
+                    notes.append(f"{label} has low quality. Recognition may be unreliable.")
+                if face.quality_flags:
+                    notes.append(f"{label}: {', '.join(face.quality_flags)}")
+            return CompareResult(
+                similarity=float(comparison["similarity"]),
+                threshold=threshold,
+                decision=str(comparison["decision"]),
+                face_a=face_a,
+                face_b=face_b,
+                notes=notes,
+            )
+
+    def get_runtime_info(self) -> Dict[str, Any]:
+        return {
+            "model_name": self.model_name,
+            "model_dir": str(self.resolve_model_dir()),
+            "providers": self.active_providers or self.requested_providers,
+            "detector": type(self.det_model).__name__ if self.det_model is not None else None,
+            "det_size": self._det_size_label(),
+            "embedding_dim": self.embedding_dim,
+            "loaded": self.is_loaded(),
+            "last_error": self.last_error,
+            "last_latency_ms": self.last_latency_ms,
+        }
+
+    def _resolve_det_sizes(self, det_size: tuple[int, int]) -> list[tuple[int, int]]:
+        width, height = int(det_size[0]), int(det_size[1])
+        if width <= 0 or height <= 0:
+            return [(int(width), int(height)) for width, height in AUTO_DET_SIZES]
+        return [(width, height)]
+
+    def _det_size_label(self) -> str:
+        if self.auto_det_size:
+            return "Auto (" + " + ".join(f"{width}x{height}" for width, height in self.det_sizes) + ")"
+        width, height = self.det_sizes[0]
+        return f"{width}x{height}"
+
+    def _prepare_detection(self) -> None:
+        if self.det_model is None:
+            return
+        input_size = self._detector_input_size()
+        if self._prepared_det_size == input_size:
+            return
+        LOGGER.debug("Preparing detection model with det_size=%s", input_size)
+        self.det_model.prepare(self.ctx_id, input_size=input_size, det_thresh=0.5)
+        self._prepared_det_size = input_size
+
+    def _detect_raw(self, image: np.ndarray):
+        self._prepare_detection()
+        return self.det_model.detect(image, max_num=0, metric="default")
+
+    def _detector_input_size(self):
+        if self.auto_det_size:
+            return list(self.det_sizes)
+        return self.det_sizes[0]
+
+    @staticmethod
+    def _nms(boxes: np.ndarray, scores: np.ndarray, iou_threshold: float = 0.4) -> list[int]:
+        if boxes.size == 0:
+            return []
+        x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+        areas = np.maximum(0.0, x2 - x1 + 1.0) * np.maximum(0.0, y2 - y1 + 1.0)
+        order = scores.argsort()[::-1]
+        keep: list[int] = []
+        while order.size > 0:
+            current = int(order[0])
+            keep.append(current)
+            if order.size == 1:
+                break
+            xx1 = np.maximum(x1[current], x1[order[1:]])
+            yy1 = np.maximum(y1[current], y1[order[1:]])
+            xx2 = np.minimum(x2[current], x2[order[1:]])
+            yy2 = np.minimum(y2[current], y2[order[1:]])
+            width = np.maximum(0.0, xx2 - xx1 + 1.0)
+            height = np.maximum(0.0, yy2 - yy1 + 1.0)
+            inter = width * height
+            union = areas[current] + areas[order[1:]] - inter
+            iou = np.where(union > 0, inter / union, 0.0)
+            order = order[np.where(iou <= iou_threshold)[0] + 1]
+        return keep
+
+
+def available_execution_providers() -> List[str]:
+    try:
+        import onnxruntime
+
+        providers = onnxruntime.get_available_providers()
+        return [str(provider) for provider in providers]
+    except Exception as exc:
+        LOGGER.debug("Unable to inspect ONNX Runtime providers: %s", exc)
+        return []
+
+
+def is_cuda_provider_available() -> bool:
+    return "CUDAExecutionProvider" in available_execution_providers()
+
+
+def providers_from_choice(choice: str) -> List[str]:
+    normalized = (choice or "CPU").strip().lower()
+    if normalized == "cpu":
+        return ["CPUExecutionProvider"]
+    if normalized == "cuda":
+        if is_cuda_provider_available():
+            return ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        LOGGER.warning("CUDA provider was requested but CUDAExecutionProvider is not available; using CPU.")
+        return ["CPUExecutionProvider"]
+    available = available_execution_providers()
+    if "CUDAExecutionProvider" in available:
+        return ["CUDAExecutionProvider", "CPUExecutionProvider"]
+    return ["CPUExecutionProvider"]
diff --git a/python-package/insightface/gui/core/i18n.py b/python-package/insightface/gui/core/i18n.py
new file mode 100644
index 000000000..93d53d0d3
--- /dev/null
+++ b/python-package/insightface/gui/core/i18n.py
@@ -0,0 +1,3078 @@
+"""Lightweight runtime localization for the desktop GUI."""
+
+from __future__ import annotations
+
+import locale
+from dataclasses import dataclass
+from typing import Any
+
+from PySide6.QtCore import QLocale, QObject, Qt
+from PySide6.QtGui import QAction
+from PySide6.QtWidgets import (
+    QAbstractButton,
+    QComboBox,
+    QDialog,
+    QGroupBox,
+    QLabel,
+    QLineEdit,
+    QMainWindow,
+    QMenu,
+    QTabWidget,
+    QTableWidget,
+    QWidget,
+)
+
+
+@dataclass(frozen=True)
+class LanguageOption:
+    value: str
+    label: str
+    native_label: str
+
+
+LANGUAGE_OPTIONS = [
+    LanguageOption("system", "System default", "System default"),
+    LanguageOption("en", "English", "English"),
+    LanguageOption("zh", "Chinese", "中文"),
+    LanguageOption("ja", "Japanese", "日本語"),
+    LanguageOption("ko", "Korean", "한국어"),
+    LanguageOption("es", "Spanish", "Español"),
+    LanguageOption("fr", "French", "Français"),
+    LanguageOption("de", "German", "Deutsch"),
+    LanguageOption("pt", "Portuguese", "Português"),
+    LanguageOption("ru", "Russian", "Русский"),
+]
+
+SUPPORTED_LANGUAGE_CODES = {option.value for option in LANGUAGE_OPTIONS if option.value != "system"}
+_SOURCE_PROPERTY = "_insightface_i18n_source"
+_PLACEHOLDER_PROPERTY = "_insightface_i18n_placeholder_source"
+_GENERIC_BUTTON_TOOLTIP = {
+    "zh": "点击执行：{action}。",
+    "ja": "クリックして「{action}」を実行します。",
+    "ko": "클릭하여 {action}을 실행합니다.",
+    "es": "Haga clic para ejecutar: {action}.",
+    "fr": "Cliquez pour exécuter : {action}.",
+    "de": "Klicken, um auszuführen: {action}.",
+    "pt": "Clique para executar: {action}.",
+    "ru": "Нажмите, чтобы выполнить: {action}.",
+}
+
+
+def normalize_language(language: str | None) -> str:
+    value = (language or "system").strip().replace("-", "_").lower()
+    if value in {"auto", "default", "system_default"}:
+        return "system"
+    code = value.split("_", 1)[0]
+    if code in SUPPORTED_LANGUAGE_CODES:
+        return code
+    return "system" if not value or value == "system" else "en"
+
+
+def detect_system_language() -> str:
+    candidates = []
+    try:
+        candidates.append(QLocale.system().name())
+    except Exception:
+        pass
+    try:
+        candidates.append(locale.getlocale()[0] or "")
+    except Exception:
+        pass
+    for candidate in candidates:
+        code = normalize_language(candidate)
+        if code in SUPPORTED_LANGUAGE_CODES:
+            return code
+    return "en"
+
+
+def effective_language(language: str | None) -> str:
+    value = normalize_language(language)
+    return detect_system_language() if value == "system" else value
+
+
+def language_label(language: str | None) -> str:
+    value = normalize_language(language)
+    for option in LANGUAGE_OPTIONS:
+        if option.value == value:
+            return option.native_label
+    return "English"
+
+
+def tr(text: str, language: str | None = None) -> str:
+    if not text:
+        return text
+    lang = effective_language(language)
+    if lang == "en":
+        return text
+    return _TRANSLATIONS.get(lang, {}).get(text, text)
+
+
+def apply_translations(root: Any, language: str | None) -> None:
+    """Translate common Qt widget texts in-place with English fallback."""
+    lang = effective_language(language)
+    _translate_object(root, lang)
+    if hasattr(root, "findChildren"):
+        for obj in root.findChildren(QObject):
+            _translate_object(obj, lang)
+
+
+def _source_text(obj: Any, text: str, prop: str = _SOURCE_PROPERTY) -> str:
+    source = obj.property(prop) if hasattr(obj, "property") else None
+    if source is None and hasattr(obj, "setProperty"):
+        obj.setProperty(prop, text)
+        return text
+    if source is not None and not _matches_source_or_translation(str(source), text) and hasattr(obj, "setProperty"):
+        obj.setProperty(prop, text)
+        return text
+    return str(source)
+
+
+def _matches_source_or_translation(source: str, text: str) -> bool:
+    if text == source:
+        return True
+    return any(translations.get(source) == text for translations in _TRANSLATIONS.values())
+
+
+def _help_source_text(obj: Any, text: str, prop: str) -> str:
+    source = obj.property(prop) if hasattr(obj, "property") else None
+    if source is None and hasattr(obj, "setProperty"):
+        obj.setProperty(prop, text)
+        return text
+    if source is not None:
+        if _matches_source_or_translation(str(source), text):
+            return str(source)
+        if isinstance(obj, QAbstractButton) and _looks_like_generic_button_tooltip(text):
+            return str(source)
+        if hasattr(obj, "setProperty"):
+            obj.setProperty(prop, text)
+            return text
+    return str(source)
+
+
+def _looks_like_generic_button_tooltip(text: str) -> bool:
+    for template in _GENERIC_BUTTON_TOOLTIP.values():
+        prefix, suffix = template.split("{action}", 1)
+        if text.startswith(prefix) and text.endswith(suffix):
+            return True
+    return False
+
+
+def _translate_object(obj: Any, lang: str) -> None:
+    _translate_object_help(obj, lang)
+    if isinstance(obj, (QDialog, QMainWindow)):
+        title = obj.windowTitle()
+        if title:
+            obj.setWindowTitle(tr(_source_text(obj, title), lang))
+    if isinstance(obj, QMenu):
+        title = obj.title()
+        if title:
+            obj.setTitle(tr(_source_text(obj, title), lang))
+    if isinstance(obj, QAction):
+        text = obj.text()
+        if text:
+            obj.setText(tr(_source_text(obj, text), lang))
+        _translate_action_help(obj, lang)
+    if isinstance(obj, QLabel):
+        text = obj.text()
+        if text:
+            obj.setText(tr(_source_text(obj, text), lang))
+    if isinstance(obj, QAbstractButton):
+        text = obj.text()
+        if text:
+            obj.setText(tr(_source_text(obj, text), lang))
+    if isinstance(obj, QGroupBox):
+        title = obj.title()
+        if title:
+            obj.setTitle(tr(_source_text(obj, title), lang))
+    if isinstance(obj, QLineEdit):
+        placeholder = obj.placeholderText()
+        if placeholder:
+            obj.setPlaceholderText(tr(_source_text(obj, placeholder, _PLACEHOLDER_PROPERTY), lang))
+    if isinstance(obj, QTabWidget):
+        for index in range(obj.count()):
+            text = obj.tabText(index)
+            key = f"{_SOURCE_PROPERTY}_tab_{index}"
+            source = obj.property(key)
+            if source is None:
+                obj.setProperty(key, text)
+                source = text
+            obj.setTabText(index, tr(str(source), lang))
+    if isinstance(obj, QComboBox) and obj.property("i18nItems"):
+        for index in range(obj.count()):
+            text = obj.itemText(index)
+            source = obj.itemData(index, Qt.UserRole + 3000)
+            if source is None:
+                obj.setItemData(index, text, Qt.UserRole + 3000)
+                source = text
+            obj.setItemText(index, tr(str(source), lang))
+    if isinstance(obj, QTableWidget):
+        for index in range(obj.columnCount()):
+            item = obj.horizontalHeaderItem(index)
+            if item is not None:
+                source = item.data(Qt.UserRole + 3000)
+                if source is None:
+                    source = item.text()
+                    item.setData(Qt.UserRole + 3000, source)
+                item.setText(tr(str(source), lang))
+
+
+def _translate_object_help(obj: Any, lang: str) -> None:
+    if not isinstance(obj, QWidget):
+        return
+    tooltip = obj.toolTip()
+    if tooltip:
+        tooltip_source = _help_source_text(obj, tooltip, f"{_SOURCE_PROPERTY}_tooltip")
+        translated = tr(tooltip_source, lang)
+        if translated == tooltip_source and lang != "en" and isinstance(obj, QAbstractButton):
+            translated = _generic_button_tooltip(obj, lang)
+        obj.setToolTip(translated)
+    status_tip = obj.statusTip()
+    if status_tip:
+        obj.setStatusTip(tr(_source_text(obj, status_tip, f"{_SOURCE_PROPERTY}_status_tip"), lang))
+    whats_this = obj.whatsThis()
+    if whats_this:
+        obj.setWhatsThis(tr(_source_text(obj, whats_this, f"{_SOURCE_PROPERTY}_whats_this"), lang))
+
+
+def _translate_action_help(action: QAction, lang: str) -> None:
+    tooltip = action.toolTip()
+    if tooltip:
+        action.setToolTip(tr(_source_text(action, tooltip, f"{_SOURCE_PROPERTY}_tooltip"), lang))
+    status_tip = action.statusTip()
+    if status_tip:
+        action.setStatusTip(tr(_source_text(action, status_tip, f"{_SOURCE_PROPERTY}_status_tip"), lang))
+    whats_this = action.whatsThis()
+    if whats_this:
+        action.setWhatsThis(tr(_source_text(action, whats_this, f"{_SOURCE_PROPERTY}_whats_this"), lang))
+
+
+def _generic_button_tooltip(button: QAbstractButton, lang: str) -> str:
+    source = button.property(_SOURCE_PROPERTY)
+    if source is None:
+        text = button.text().replace("&", "").strip()
+        source = "Remove" if text == "×" else text
+    action = tr(str(source), lang)
+    if not action or action == "×":
+        action = tr("Remove", lang)
+    return _GENERIC_BUTTON_TOOLTIP.get(lang, "{action}").format(action=action)
+
+
+_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "System default": "跟随系统",
+        "Settings": "设置",
+        "Application Settings": "应用设置",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "此处仅配置界面外观。工作目录在首次启动后保持固定。",
+        "UI theme": "界面主题",
+        "Language": "语言",
+        "Save": "保存",
+        "Apply": "应用",
+        "Cancel": "取消",
+        "Models": "模型",
+        "License": "授权",
+        "Research / Non-commercial": "研究 / 非商业",
+        "License Center": "授权中心",
+        "Workflows": "工作流",
+        "Choose the workspace for the task.": "选择当前任务的工作区。",
+        "Enterprise Help": "企业支持",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "所有处理默认在本地完成。图片、特征和报告不会自动上传。",
+        "Face Recognition": "人脸识别",
+        "Album Management": "相册管理",
+        "Face Swap": "人脸替换",
+        "Enterprise Evaluation": "企业评测",
+        "Query and gallery recognition": "查询图与底库识别",
+        "Photo clustering and review": "照片聚类与复核",
+        "Source + Target = Result": "源图 + 目标 = 结果",
+        "Datasets, metrics, and reports": "数据集、指标与报告",
+        "Upload a query face and recognize it against one gallery image or a local gallery.": "上传查询人脸，并与单张底库图或本地底库进行识别。",
+        "Organize local photos by people, clusters, and searchable face groups.": "按人物、聚类和可搜索人脸组整理本地照片。",
+        "Use a configured downloaded swap model for local image or video face swap.": "使用已配置并下载的换脸模型在本地处理图片或视频。",
+        "Run local 1:1 and 1:N model evaluations and export reports.": "运行本地 1:1 和 1:N 模型评测并导出报告。",
+        "Visit Homepage": "访问主页",
+        "Contact Enterprise Support": "联系企业支持",
+        "Usage": "用途",
+        "Status": "状态",
+        "Personal research": "个人研究",
+        "Internal evaluation": "内部评估",
+        "Commercial production": "商业生产",
+        "Redistribution": "再分发",
+        "SaaS / API usage": "SaaS / API 使用",
+        "Face swap commercial usage": "换脸商业使用",
+        "yes": "可以",
+        "depends on model license": "取决于模型授权",
+        "requires commercial model license": "需要商业模型授权",
+        "requires explicit permission": "需要明确许可",
+        "requires commercial agreement": "需要商业协议",
+        "A concise local-use and commercial-deployment summary.": "本地使用与商业部署的简明说明。",
+        "InsightFace Evaluation Studio runs locally and does not upload images, embeddings, videos, or reports automatically.": "InsightFace Evaluation Studio 在本地运行，不会自动上传图片、特征、视频或报告。",
+        "Code and model files may have different licenses, and commercial deployment requires appropriate model authorization.": "代码和模型文件可能适用不同授权；商业部署需要相应的模型授权。",
+        "Current package": "当前包",
+        "Evaluation Setup": "评测设置",
+        "Evaluation mode": "评测模式",
+        "Data split": "数据切分",
+        "Multi-face handling": "多人脸处理",
+        "Dataset": "数据集",
+        "Run": "运行",
+        "Dataset Rules": "数据集规则",
+        "Validate Dataset": "核验数据集",
+        "Run Evaluation": "运行评测",
+        "Export PDF": "导出 PDF",
+        "Open Report Folder": "打开报告目录",
+        "Validate the selected dataset and configuration before running evaluation.": "运行评测前请先核验所选数据集和配置。",
+        "Validate the dataset before running an evaluation.": "运行评测前请先核验数据集。",
+        "Dataset validation is required before running an evaluation.": "运行评测前需要完成数据集核验。",
+        "Evaluation Dataset Rules": "评测数据集规则",
+        "Prepare local identity folders using one of the supported layouts below.": "请按以下任一支持结构准备本地身份文件夹。",
+        "Recognition threshold": "识别阈值",
+        "Run Recognition": "运行识别",
+        "Clear": "清空",
+        "Query": "查询图",
+        "Gallery": "底库",
+        "Album": "相册",
+        "Album directories": "相册目录",
+        "Drag album folders into the directory list, or click Add Folder.": "将相册文件夹拖入目录列表，或点击“添加文件夹”。",
+        "Import / Refresh": "导入 / 刷新",
+        "Rebuild All": "全部重建",
+        "Add Folder": "添加文件夹",
+        "Remove Selected": "移除所选",
+        "Cosine threshold": "余弦阈值",
+        "Min samples": "最小样本数",
+        "Minimum face size": "最小人脸尺寸",
+        "Algorithm: DBSCAN": "算法：DBSCAN",
+        "Thumbnail": "缩略图",
+        "Photos": "照片",
+        "File": "文件",
+        "Faces in cluster": "聚类内人脸数",
+        "Run Face Swap": "运行换脸",
+        "Open Result Directory": "打开结果目录",
+        "Source": "源图",
+        "Target": "目标",
+        "Result": "结果",
+        "Result preview appears here. Video results are saved to the exports folder.": "结果预览会显示在这里。视频结果会保存到导出目录。",
+        "Select": "选择",
+        "Remove": "移除",
+        "Click to upload or drag a folder here.": "点击上传，或将文件夹拖到这里。",
+        "Click to upload or drag files here.": "点击上传，或将文件拖到这里。",
+        "Click to upload or drag a file here.": "点击上传，或将文件拖到这里。",
+        "File": "文件",
+        "Open Workspace Folder": "打开工作目录",
+        "Exit": "退出",
+        "Mode": "模式",
+        "Tools": "工具",
+        "Help": "帮助",
+        "About": "关于",
+        "Responsible Use Notice": "负责任使用提示",
+        "Ready": "就绪",
+        "Model": "模型",
+        "Provider": "运行后端",
+        "DB": "数据库",
+        "Coming soon": "即将推出",
+    },
+    "ja": {
+        "System default": "システム既定",
+        "Settings": "設定",
+        "Application Settings": "アプリケーション設定",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "ここでは表示設定のみ変更できます。ワークスペースのパスは初回起動後に固定されます。",
+        "UI theme": "UI テーマ",
+        "Language": "言語",
+        "Save": "保存",
+        "Apply": "適用",
+        "Cancel": "キャンセル",
+        "Models": "モデル",
+        "License": "ライセンス",
+        "Research / Non-commercial": "研究 / 非商用",
+        "License Center": "ライセンスセンター",
+        "Workflows": "ワークフロー",
+        "Choose the workspace for the task.": "作業内容に応じたワークスペースを選択します。",
+        "Enterprise Help": "エンタープライズサポート",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "すべての処理はローカルで実行されます。画像、特徴量、レポートは自動的にアップロードされません。",
+        "Face Recognition": "顔認識",
+        "Album Management": "アルバム管理",
+        "Face Swap": "顔交換",
+        "Enterprise Evaluation": "エンタープライズ評価",
+        "Query and gallery recognition": "クエリとギャラリー認識",
+        "Photo clustering and review": "写真クラスタリングと確認",
+        "Source + Target = Result": "ソース + ターゲット = 結果",
+        "Datasets, metrics, and reports": "データセット、指標、レポート",
+        "Upload a query face and recognize it against one gallery image or a local gallery.": "クエリ顔画像をアップロードし、1枚のギャラリー画像またはローカルギャラリーと照合します。",
+        "Organize local photos by people, clusters, and searchable face groups.": "人物、クラスタ、検索可能な顔グループでローカル写真を整理します。",
+        "Use a configured downloaded swap model for local image or video face swap.": "設定済みのダウンロード済み顔交換モデルで、画像または動画をローカル処理します。",
+        "Run local 1:1 and 1:N model evaluations and export reports.": "ローカルの 1:1 / 1:N モデル評価を実行し、レポートを出力します。",
+        "Visit Homepage": "ホームページを開く",
+        "Contact Enterprise Support": "企業向けサポートに連絡",
+        "Usage": "用途",
+        "Status": "状態",
+        "Personal research": "個人研究",
+        "Internal evaluation": "社内評価",
+        "Commercial production": "商用本番利用",
+        "Redistribution": "再配布",
+        "SaaS / API usage": "SaaS / API 利用",
+        "Face swap commercial usage": "顔交換の商用利用",
+        "yes": "可",
+        "depends on model license": "モデルライセンスによります",
+        "requires commercial model license": "商用モデルライセンスが必要",
+        "requires explicit permission": "明示的な許可が必要",
+        "requires commercial agreement": "商用契約が必要",
+        "A concise local-use and commercial-deployment summary.": "ローカル利用と商用展開に関する簡潔な要約です。",
+        "InsightFace Evaluation Studio runs locally and does not upload images, embeddings, videos, or reports automatically.": "InsightFace Evaluation Studio はローカルで動作し、画像、特徴量、動画、レポートを自動アップロードしません。",
+        "Code and model files may have different licenses, and commercial deployment requires appropriate model authorization.": "コードとモデルファイルには異なるライセンスが適用される場合があり、商用展開には適切なモデル認可が必要です。",
+        "Current package": "現在のパッケージ",
+        "Evaluation Setup": "評価設定",
+        "Evaluation mode": "評価モード",
+        "Data split": "データ分割",
+        "Multi-face handling": "複数顔の処理",
+        "Dataset": "データセット",
+        "Run": "実行",
+        "Dataset Rules": "データセット規則",
+        "Validate Dataset": "データセットを検証",
+        "Run Evaluation": "評価を実行",
+        "Export PDF": "PDF を出力",
+        "Open Report Folder": "レポートフォルダを開く",
+        "Validate the selected dataset and configuration before running evaluation.": "評価を実行する前に、選択したデータセットと設定を検証してください。",
+        "Validate the dataset before running an evaluation.": "評価を実行する前にデータセットを検証してください。",
+        "Dataset validation is required before running an evaluation.": "評価を実行する前にデータセット検証が必要です。",
+        "Evaluation Dataset Rules": "評価データセット規則",
+        "Prepare local identity folders using one of the supported layouts below.": "以下のいずれかの対応レイアウトでローカルの ID フォルダを準備してください。",
+        "Recognition threshold": "認識しきい値",
+        "Run Recognition": "認識を実行",
+        "Clear": "クリア",
+        "Query": "クエリ",
+        "Gallery": "ギャラリー",
+        "Album": "アルバム",
+        "Album directories": "アルバムディレクトリ",
+        "Drag album folders into the directory list, or click Add Folder.": "アルバムフォルダをリストにドラッグするか、「フォルダを追加」をクリックします。",
+        "Import / Refresh": "インポート / 更新",
+        "Rebuild All": "すべて再構築",
+        "Add Folder": "フォルダを追加",
+        "Remove Selected": "選択項目を削除",
+        "Cosine threshold": "コサインしきい値",
+        "Min samples": "最小サンプル数",
+        "Minimum face size": "最小顔サイズ",
+        "Algorithm: DBSCAN": "アルゴリズム: DBSCAN",
+        "Thumbnail": "サムネイル",
+        "Photos": "写真",
+        "File": "ファイル",
+        "Faces in cluster": "クラスタ内の顔数",
+        "Run Face Swap": "顔交換を実行",
+        "Open Result Directory": "結果ディレクトリを開く",
+        "Source": "ソース",
+        "Target": "ターゲット",
+        "Result": "結果",
+        "Result preview appears here. Video results are saved to the exports folder.": "結果プレビューがここに表示されます。動画結果はエクスポートフォルダに保存されます。",
+        "Select": "選択",
+        "Remove": "削除",
+        "Click to upload or drag a folder here.": "クリックしてアップロード、またはフォルダをここにドラッグします。",
+        "Click to upload or drag files here.": "クリックしてアップロード、またはファイルをここにドラッグします。",
+        "Click to upload or drag a file here.": "クリックしてアップロード、またはファイルをここにドラッグします。",
+        "Open Workspace Folder": "ワークスペースフォルダを開く",
+        "Exit": "終了",
+        "Mode": "モード",
+        "Tools": "ツール",
+        "Help": "ヘルプ",
+        "About": "このアプリについて",
+        "Responsible Use Notice": "責任ある利用に関する注意",
+        "Ready": "準備完了",
+        "Model": "モデル",
+        "Provider": "プロバイダー",
+        "DB": "DB",
+        "Coming soon": "近日公開",
+    },
+    "ko": {
+        "System default": "시스템 기본값",
+        "Settings": "설정",
+        "Application Settings": "애플리케이션 설정",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "여기서는 화면 표시 설정만 변경할 수 있습니다. 워크스페이스 경로는 최초 실행 후 고정됩니다.",
+        "UI theme": "UI 테마",
+        "Language": "언어",
+        "Save": "저장",
+        "Apply": "적용",
+        "Cancel": "취소",
+        "Models": "모델",
+        "License": "라이선스",
+        "Research / Non-commercial": "연구 / 비상업",
+        "License Center": "라이선스 센터",
+        "Workflows": "워크플로",
+        "Choose the workspace for the task.": "작업에 맞는 워크스페이스를 선택합니다.",
+        "Enterprise Help": "기업 지원",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "모든 처리는 로컬에서 수행됩니다. 이미지, 임베딩, 보고서는 자동으로 업로드되지 않습니다.",
+        "Face Recognition": "얼굴 인식",
+        "Album Management": "앨범 관리",
+        "Face Swap": "얼굴 교체",
+        "Enterprise Evaluation": "기업 평가",
+        "Query and gallery recognition": "쿼리 및 갤러리 인식",
+        "Photo clustering and review": "사진 클러스터링 및 검토",
+        "Source + Target = Result": "소스 + 타깃 = 결과",
+        "Datasets, metrics, and reports": "데이터셋, 지표, 보고서",
+        "Upload a query face and recognize it against one gallery image or a local gallery.": "쿼리 얼굴을 업로드하고 단일 갤러리 이미지 또는 로컬 갤러리와 비교합니다.",
+        "Organize local photos by people, clusters, and searchable face groups.": "로컬 사진을 인물, 클러스터, 검색 가능한 얼굴 그룹으로 정리합니다.",
+        "Use a configured downloaded swap model for local image or video face swap.": "설정된 다운로드 얼굴 교체 모델로 이미지 또는 비디오를 로컬 처리합니다.",
+        "Run local 1:1 and 1:N model evaluations and export reports.": "로컬 1:1 및 1:N 모델 평가를 실행하고 보고서를 내보냅니다.",
+        "Visit Homepage": "홈페이지 방문",
+        "Contact Enterprise Support": "기업 지원 문의",
+        "Usage": "용도",
+        "Status": "상태",
+        "Personal research": "개인 연구",
+        "Internal evaluation": "내부 평가",
+        "Commercial production": "상용 운영",
+        "Redistribution": "재배포",
+        "SaaS / API usage": "SaaS / API 사용",
+        "Face swap commercial usage": "얼굴 교체 상용 사용",
+        "yes": "가능",
+        "depends on model license": "모델 라이선스에 따라 다름",
+        "requires commercial model license": "상용 모델 라이선스 필요",
+        "requires explicit permission": "명시적 허가 필요",
+        "requires commercial agreement": "상용 계약 필요",
+        "A concise local-use and commercial-deployment summary.": "로컬 사용 및 상용 배포에 대한 간단한 요약입니다.",
+        "InsightFace Evaluation Studio runs locally and does not upload images, embeddings, videos, or reports automatically.": "InsightFace Evaluation Studio는 로컬에서 실행되며 이미지, 임베딩, 비디오, 보고서를 자동 업로드하지 않습니다.",
+        "Code and model files may have different licenses, and commercial deployment requires appropriate model authorization.": "코드와 모델 파일에는 서로 다른 라이선스가 적용될 수 있으며, 상용 배포에는 적절한 모델 승인이 필요합니다.",
+        "Current package": "현재 패키지",
+        "Evaluation Setup": "평가 설정",
+        "Evaluation mode": "평가 모드",
+        "Data split": "데이터 분할",
+        "Multi-face handling": "다중 얼굴 처리",
+        "Dataset": "데이터셋",
+        "Run": "실행",
+        "Dataset Rules": "데이터셋 규칙",
+        "Validate Dataset": "데이터셋 검증",
+        "Run Evaluation": "평가 실행",
+        "Export PDF": "PDF 내보내기",
+        "Open Report Folder": "보고서 폴더 열기",
+        "Recognition threshold": "인식 임계값",
+        "Run Recognition": "인식 실행",
+        "Clear": "지우기",
+        "Query": "쿼리",
+        "Gallery": "갤러리",
+        "Album": "앨범",
+        "Album directories": "앨범 디렉터리",
+        "Import / Refresh": "가져오기 / 새로고침",
+        "Rebuild All": "전체 재구축",
+        "Add Folder": "폴더 추가",
+        "Remove Selected": "선택 항목 제거",
+        "Cosine threshold": "코사인 임계값",
+        "Min samples": "최소 샘플 수",
+        "Minimum face size": "최소 얼굴 크기",
+        "Thumbnail": "썸네일",
+        "Photos": "사진",
+        "File": "파일",
+        "Faces in cluster": "클러스터 내 얼굴 수",
+        "Run Face Swap": "얼굴 교체 실행",
+        "Open Result Directory": "결과 디렉터리 열기",
+        "Source": "소스",
+        "Target": "타깃",
+        "Result": "결과",
+        "Select": "선택",
+        "Remove": "제거",
+        "Open Workspace Folder": "워크스페이스 폴더 열기",
+        "Exit": "종료",
+        "Mode": "모드",
+        "Tools": "도구",
+        "Help": "도움말",
+        "About": "정보",
+        "Responsible Use Notice": "책임 있는 사용 안내",
+        "Ready": "준비됨",
+        "Model": "모델",
+        "Provider": "프로바이더",
+        "DB": "DB",
+        "Coming soon": "출시 예정",
+    },
+    "es": {
+        "System default": "Predeterminado del sistema",
+        "Settings": "Configuración",
+        "Application Settings": "Configuración de la aplicación",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "Aquí solo se configuran opciones de apariencia. Las rutas del espacio de trabajo quedan fijadas después del primer inicio.",
+        "UI theme": "Tema de interfaz",
+        "Language": "Idioma",
+        "Save": "Guardar",
+        "Apply": "Aplicar",
+        "Cancel": "Cancelar",
+        "Models": "Modelos",
+        "License": "Licencia",
+        "Research / Non-commercial": "Investigación / no comercial",
+        "License Center": "Centro de licencias",
+        "Workflows": "Flujos de trabajo",
+        "Choose the workspace for the task.": "Elija el espacio de trabajo para la tarea.",
+        "Enterprise Help": "Soporte empresarial",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "Todo el procesamiento es local. No se suben automáticamente imágenes, embeddings ni informes.",
+        "Face Recognition": "Reconocimiento facial",
+        "Album Management": "Gestión de álbumes",
+        "Face Swap": "Intercambio facial",
+        "Enterprise Evaluation": "Evaluación empresarial",
+        "Query and gallery recognition": "Reconocimiento de consulta y galería",
+        "Photo clustering and review": "Agrupación y revisión de fotos",
+        "Source + Target = Result": "Origen + destino = resultado",
+        "Datasets, metrics, and reports": "Conjuntos de datos, métricas e informes",
+        "Visit Homepage": "Visitar sitio principal",
+        "Contact Enterprise Support": "Contactar soporte empresarial",
+        "Usage": "Uso",
+        "Status": "Estado",
+        "Personal research": "Investigación personal",
+        "Internal evaluation": "Evaluación interna",
+        "Commercial production": "Producción comercial",
+        "Redistribution": "Redistribución",
+        "SaaS / API usage": "Uso SaaS / API",
+        "Face swap commercial usage": "Uso comercial de intercambio facial",
+        "yes": "sí",
+        "depends on model license": "depende de la licencia del modelo",
+        "requires commercial model license": "requiere licencia comercial del modelo",
+        "requires explicit permission": "requiere permiso explícito",
+        "requires commercial agreement": "requiere acuerdo comercial",
+        "Current package": "Paquete actual",
+        "Evaluation Setup": "Configuración de evaluación",
+        "Evaluation mode": "Modo de evaluación",
+        "Data split": "División de datos",
+        "Multi-face handling": "Tratamiento de varias caras",
+        "Dataset": "Datos",
+        "Run": "Ejecutar",
+        "Dataset Rules": "Reglas del conjunto de datos",
+        "Validate Dataset": "Validar datos",
+        "Run Evaluation": "Ejecutar evaluación",
+        "Export PDF": "Exportar PDF",
+        "Open Report Folder": "Abrir carpeta de informes",
+        "Recognition threshold": "Umbral de reconocimiento",
+        "Run Recognition": "Ejecutar reconocimiento",
+        "Clear": "Limpiar",
+        "Query": "Consulta",
+        "Gallery": "Galería",
+        "Album": "Álbum",
+        "Album directories": "Directorios de álbum",
+        "Import / Refresh": "Importar / actualizar",
+        "Rebuild All": "Reconstruir todo",
+        "Add Folder": "Añadir carpeta",
+        "Remove Selected": "Quitar selección",
+        "Cosine threshold": "Umbral coseno",
+        "Min samples": "Muestras mínimas",
+        "Minimum face size": "Tamaño mínimo de cara",
+        "Thumbnail": "Miniatura",
+        "Photos": "Fotos",
+        "File": "Archivo",
+        "Faces in cluster": "Caras en el grupo",
+        "Run Face Swap": "Ejecutar intercambio facial",
+        "Open Result Directory": "Abrir directorio de resultados",
+        "Source": "Origen",
+        "Target": "Destino",
+        "Result": "Resultado",
+        "Select": "Seleccionar",
+        "Remove": "Quitar",
+        "Open Workspace Folder": "Abrir carpeta de trabajo",
+        "Exit": "Salir",
+        "Mode": "Modo",
+        "Tools": "Herramientas",
+        "Help": "Ayuda",
+        "About": "Acerca de",
+        "Responsible Use Notice": "Aviso de uso responsable",
+        "Ready": "Listo",
+        "Model": "Modelo",
+        "Provider": "Proveedor",
+        "DB": "BD",
+        "Coming soon": "Próximamente",
+    },
+    "fr": {
+        "System default": "Valeur système",
+        "Settings": "Paramètres",
+        "Application Settings": "Paramètres de l’application",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "Seuls les paramètres d’apparence sont configurables ici. Les chemins de l’espace de travail sont figés après le premier lancement.",
+        "UI theme": "Thème d’interface",
+        "Language": "Langue",
+        "Save": "Enregistrer",
+        "Apply": "Appliquer",
+        "Cancel": "Annuler",
+        "Models": "Modèles",
+        "License": "Licence",
+        "Research / Non-commercial": "Recherche / non commercial",
+        "License Center": "Centre de licences",
+        "Workflows": "Flux de travail",
+        "Choose the workspace for the task.": "Choisissez l’espace de travail adapté à la tâche.",
+        "Enterprise Help": "Support entreprise",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "Tous les traitements sont locaux. Aucune image, embedding ni rapport n’est téléversé automatiquement.",
+        "Face Recognition": "Reconnaissance faciale",
+        "Album Management": "Gestion d’albums",
+        "Face Swap": "Échange de visages",
+        "Enterprise Evaluation": "Évaluation entreprise",
+        "Query and gallery recognition": "Reconnaissance requête et galerie",
+        "Photo clustering and review": "Clustering et revue des photos",
+        "Source + Target = Result": "Source + cible = résultat",
+        "Datasets, metrics, and reports": "Jeux de données, métriques et rapports",
+        "Visit Homepage": "Visiter le site",
+        "Contact Enterprise Support": "Contacter le support entreprise",
+        "Usage": "Usage",
+        "Status": "Statut",
+        "Personal research": "Recherche personnelle",
+        "Internal evaluation": "Évaluation interne",
+        "Commercial production": "Production commerciale",
+        "Redistribution": "Redistribution",
+        "SaaS / API usage": "Usage SaaS / API",
+        "Face swap commercial usage": "Usage commercial du face swap",
+        "yes": "oui",
+        "depends on model license": "dépend de la licence du modèle",
+        "requires commercial model license": "requiert une licence commerciale du modèle",
+        "requires explicit permission": "requiert une autorisation explicite",
+        "requires commercial agreement": "requiert un accord commercial",
+        "Current package": "Package actuel",
+        "Evaluation Setup": "Configuration de l’évaluation",
+        "Evaluation mode": "Mode d’évaluation",
+        "Data split": "Répartition des données",
+        "Multi-face handling": "Gestion des visages multiples",
+        "Dataset": "Jeu de données",
+        "Run": "Exécuter",
+        "Dataset Rules": "Règles du jeu de données",
+        "Validate Dataset": "Valider le jeu de données",
+        "Run Evaluation": "Lancer l’évaluation",
+        "Export PDF": "Exporter en PDF",
+        "Open Report Folder": "Ouvrir le dossier des rapports",
+        "Recognition threshold": "Seuil de reconnaissance",
+        "Run Recognition": "Lancer la reconnaissance",
+        "Clear": "Effacer",
+        "Query": "Requête",
+        "Gallery": "Galerie",
+        "Album": "Album",
+        "Album directories": "Dossiers d’album",
+        "Import / Refresh": "Importer / actualiser",
+        "Rebuild All": "Tout reconstruire",
+        "Add Folder": "Ajouter un dossier",
+        "Remove Selected": "Retirer la sélection",
+        "Cosine threshold": "Seuil cosinus",
+        "Min samples": "Échantillons min.",
+        "Minimum face size": "Taille minimale du visage",
+        "Thumbnail": "Vignette",
+        "Photos": "Photos",
+        "File": "Fichier",
+        "Faces in cluster": "Visages dans le cluster",
+        "Run Face Swap": "Lancer le face swap",
+        "Open Result Directory": "Ouvrir le dossier des résultats",
+        "Source": "Source",
+        "Target": "Cible",
+        "Result": "Résultat",
+        "Select": "Sélectionner",
+        "Remove": "Retirer",
+        "Open Workspace Folder": "Ouvrir l’espace de travail",
+        "Exit": "Quitter",
+        "Mode": "Mode",
+        "Tools": "Outils",
+        "Help": "Aide",
+        "About": "À propos",
+        "Responsible Use Notice": "Avis d’utilisation responsable",
+        "Ready": "Prêt",
+        "Model": "Modèle",
+        "Provider": "Fournisseur",
+        "DB": "BD",
+        "Coming soon": "Bientôt disponible",
+    },
+    "de": {
+        "System default": "Systemstandard",
+        "Settings": "Einstellungen",
+        "Application Settings": "Anwendungseinstellungen",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "Hier können nur Darstellungseinstellungen geändert werden. Workspace-Pfade sind nach dem ersten Start festgelegt.",
+        "UI theme": "UI-Design",
+        "Language": "Sprache",
+        "Save": "Speichern",
+        "Apply": "Anwenden",
+        "Cancel": "Abbrechen",
+        "Models": "Modelle",
+        "License": "Lizenz",
+        "Research / Non-commercial": "Forschung / nicht kommerziell",
+        "License Center": "Lizenzzentrum",
+        "Workflows": "Workflows",
+        "Choose the workspace for the task.": "Wählen Sie den Arbeitsbereich für die Aufgabe.",
+        "Enterprise Help": "Enterprise-Support",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "Alle Verarbeitung erfolgt lokal. Bilder, Embeddings und Berichte werden nicht automatisch hochgeladen.",
+        "Face Recognition": "Gesichtserkennung",
+        "Album Management": "Albumverwaltung",
+        "Face Swap": "Face Swap",
+        "Enterprise Evaluation": "Enterprise-Evaluierung",
+        "Query and gallery recognition": "Query- und Galerieerkennung",
+        "Photo clustering and review": "Fotoclustering und Prüfung",
+        "Source + Target = Result": "Quelle + Ziel = Ergebnis",
+        "Datasets, metrics, and reports": "Datensätze, Metriken und Berichte",
+        "Visit Homepage": "Homepage besuchen",
+        "Contact Enterprise Support": "Enterprise-Support kontaktieren",
+        "Usage": "Nutzung",
+        "Status": "Status",
+        "Personal research": "Persönliche Forschung",
+        "Internal evaluation": "Interne Evaluierung",
+        "Commercial production": "Kommerzieller Produktivbetrieb",
+        "Redistribution": "Weiterverteilung",
+        "SaaS / API usage": "SaaS- / API-Nutzung",
+        "Face swap commercial usage": "Kommerzielle Face-Swap-Nutzung",
+        "yes": "ja",
+        "depends on model license": "abhängig von der Modelllizenz",
+        "requires commercial model license": "erfordert eine kommerzielle Modelllizenz",
+        "requires explicit permission": "erfordert ausdrückliche Genehmigung",
+        "requires commercial agreement": "erfordert eine kommerzielle Vereinbarung",
+        "Current package": "Aktuelles Paket",
+        "Evaluation Setup": "Evaluierung einrichten",
+        "Evaluation mode": "Evaluierungsmodus",
+        "Data split": "Datenaufteilung",
+        "Multi-face handling": "Umgang mit mehreren Gesichtern",
+        "Dataset": "Datensatz",
+        "Run": "Ausführen",
+        "Dataset Rules": "Datensatzregeln",
+        "Validate Dataset": "Datensatz validieren",
+        "Run Evaluation": "Evaluierung starten",
+        "Export PDF": "PDF exportieren",
+        "Open Report Folder": "Berichtsordner öffnen",
+        "Recognition threshold": "Erkennungsschwelle",
+        "Run Recognition": "Erkennung starten",
+        "Clear": "Leeren",
+        "Query": "Abfrage",
+        "Gallery": "Galerie",
+        "Album": "Album",
+        "Album directories": "Albumordner",
+        "Import / Refresh": "Importieren / aktualisieren",
+        "Rebuild All": "Alles neu erstellen",
+        "Add Folder": "Ordner hinzufügen",
+        "Remove Selected": "Auswahl entfernen",
+        "Cosine threshold": "Kosinus-Schwelle",
+        "Min samples": "Min. Stichproben",
+        "Minimum face size": "Minimale Gesichtsgröße",
+        "Thumbnail": "Miniatur",
+        "Photos": "Fotos",
+        "File": "Datei",
+        "Faces in cluster": "Gesichter im Cluster",
+        "Run Face Swap": "Face Swap ausführen",
+        "Open Result Directory": "Ergebnisordner öffnen",
+        "Source": "Quelle",
+        "Target": "Ziel",
+        "Result": "Ergebnis",
+        "Select": "Auswählen",
+        "Remove": "Entfernen",
+        "Open Workspace Folder": "Workspace-Ordner öffnen",
+        "Exit": "Beenden",
+        "Mode": "Modus",
+        "Tools": "Werkzeuge",
+        "Help": "Hilfe",
+        "About": "Info",
+        "Responsible Use Notice": "Hinweis zur verantwortungsvollen Nutzung",
+        "Ready": "Bereit",
+        "Model": "Modell",
+        "Provider": "Provider",
+        "DB": "DB",
+        "Coming soon": "Demnächst",
+    },
+    "pt": {
+        "System default": "Padrão do sistema",
+        "Settings": "Definições",
+        "Application Settings": "Definições da aplicação",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "Aqui só pode configurar a aparência. Os caminhos do workspace ficam fixos após o primeiro arranque.",
+        "UI theme": "Tema da interface",
+        "Language": "Idioma",
+        "Save": "Guardar",
+        "Apply": "Aplicar",
+        "Cancel": "Cancelar",
+        "Models": "Modelos",
+        "License": "Licença",
+        "Research / Non-commercial": "Investigação / não comercial",
+        "License Center": "Centro de licenças",
+        "Workflows": "Fluxos de trabalho",
+        "Choose the workspace for the task.": "Escolha o workspace para a tarefa.",
+        "Enterprise Help": "Suporte empresarial",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "Todo o processamento é local. Imagens, embeddings e relatórios não são carregados automaticamente.",
+        "Face Recognition": "Reconhecimento facial",
+        "Album Management": "Gestão de álbuns",
+        "Face Swap": "Troca de rostos",
+        "Enterprise Evaluation": "Avaliação empresarial",
+        "Query and gallery recognition": "Reconhecimento de consulta e galeria",
+        "Photo clustering and review": "Agrupamento e revisão de fotos",
+        "Source + Target = Result": "Origem + destino = resultado",
+        "Datasets, metrics, and reports": "Conjuntos de dados, métricas e relatórios",
+        "Visit Homepage": "Visitar página inicial",
+        "Contact Enterprise Support": "Contactar suporte empresarial",
+        "Usage": "Utilização",
+        "Status": "Estado",
+        "Personal research": "Investigação pessoal",
+        "Internal evaluation": "Avaliação interna",
+        "Commercial production": "Produção comercial",
+        "Redistribution": "Redistribuição",
+        "SaaS / API usage": "Utilização SaaS / API",
+        "Face swap commercial usage": "Uso comercial de troca de rostos",
+        "yes": "sim",
+        "depends on model license": "depende da licença do modelo",
+        "requires commercial model license": "requer licença comercial do modelo",
+        "requires explicit permission": "requer autorização explícita",
+        "requires commercial agreement": "requer acordo comercial",
+        "Current package": "Pacote atual",
+        "Evaluation Setup": "Configuração da avaliação",
+        "Evaluation mode": "Modo de avaliação",
+        "Data split": "Divisão de dados",
+        "Multi-face handling": "Tratamento de múltiplos rostos",
+        "Dataset": "Dados",
+        "Run": "Executar",
+        "Dataset Rules": "Regras do conjunto de dados",
+        "Validate Dataset": "Validar dados",
+        "Run Evaluation": "Executar avaliação",
+        "Export PDF": "Exportar PDF",
+        "Open Report Folder": "Abrir pasta de relatórios",
+        "Recognition threshold": "Limiar de reconhecimento",
+        "Run Recognition": "Executar reconhecimento",
+        "Clear": "Limpar",
+        "Query": "Consulta",
+        "Gallery": "Galeria",
+        "Album": "Álbum",
+        "Album directories": "Diretórios de álbum",
+        "Import / Refresh": "Importar / atualizar",
+        "Rebuild All": "Reconstruir tudo",
+        "Add Folder": "Adicionar pasta",
+        "Remove Selected": "Remover seleção",
+        "Cosine threshold": "Limiar de cosseno",
+        "Min samples": "Amostras mínimas",
+        "Minimum face size": "Tamanho mínimo do rosto",
+        "Thumbnail": "Miniatura",
+        "Photos": "Fotos",
+        "File": "Ficheiro",
+        "Faces in cluster": "Rostos no cluster",
+        "Run Face Swap": "Executar troca de rostos",
+        "Open Result Directory": "Abrir diretório de resultados",
+        "Source": "Origem",
+        "Target": "Destino",
+        "Result": "Resultado",
+        "Select": "Selecionar",
+        "Remove": "Remover",
+        "Open Workspace Folder": "Abrir pasta do workspace",
+        "Exit": "Sair",
+        "Mode": "Modo",
+        "Tools": "Ferramentas",
+        "Help": "Ajuda",
+        "About": "Sobre",
+        "Responsible Use Notice": "Aviso de utilização responsável",
+        "Ready": "Pronto",
+        "Model": "Modelo",
+        "Provider": "Fornecedor",
+        "DB": "BD",
+        "Coming soon": "Em breve",
+    },
+    "ru": {
+        "System default": "Как в системе",
+        "Settings": "Настройки",
+        "Application Settings": "Настройки приложения",
+        "Only appearance settings are configurable here. Workspace paths are fixed after first launch.": "Здесь настраивается только внешний вид. Пути рабочей области фиксируются после первого запуска.",
+        "UI theme": "Тема интерфейса",
+        "Language": "Язык",
+        "Save": "Сохранить",
+        "Apply": "Применить",
+        "Cancel": "Отмена",
+        "Models": "Модели",
+        "License": "Лицензия",
+        "Research / Non-commercial": "Исследование / некоммерческое использование",
+        "License Center": "Центр лицензий",
+        "Workflows": "Рабочие процессы",
+        "Choose the workspace for the task.": "Выберите рабочую область для задачи.",
+        "Enterprise Help": "Поддержка для бизнеса",
+        "All processing is local. No images, embeddings, or reports are uploaded automatically.": "Вся обработка выполняется локально. Изображения, эмбеддинги и отчёты не загружаются автоматически.",
+        "Face Recognition": "Распознавание лиц",
+        "Album Management": "Управление альбомами",
+        "Face Swap": "Замена лиц",
+        "Enterprise Evaluation": "Корпоративная оценка",
+        "Query and gallery recognition": "Распознавание запроса и галереи",
+        "Photo clustering and review": "Кластеризация и проверка фото",
+        "Source + Target = Result": "Источник + цель = результат",
+        "Datasets, metrics, and reports": "Датасеты, метрики и отчёты",
+        "Visit Homepage": "Открыть сайт",
+        "Contact Enterprise Support": "Связаться с поддержкой",
+        "Usage": "Использование",
+        "Status": "Статус",
+        "Personal research": "Личное исследование",
+        "Internal evaluation": "Внутренняя оценка",
+        "Commercial production": "Коммерческий продакшен",
+        "Redistribution": "Повторное распространение",
+        "SaaS / API usage": "Использование SaaS / API",
+        "Face swap commercial usage": "Коммерческое использование замены лиц",
+        "yes": "да",
+        "depends on model license": "зависит от лицензии модели",
+        "requires commercial model license": "требуется коммерческая лицензия модели",
+        "requires explicit permission": "требуется явное разрешение",
+        "requires commercial agreement": "требуется коммерческое соглашение",
+        "Current package": "Текущий пакет",
+        "Evaluation Setup": "Настройка оценки",
+        "Evaluation mode": "Режим оценки",
+        "Data split": "Разделение данных",
+        "Multi-face handling": "Обработка нескольких лиц",
+        "Dataset": "Датасет",
+        "Run": "Запуск",
+        "Dataset Rules": "Правила датасета",
+        "Validate Dataset": "Проверить датасет",
+        "Run Evaluation": "Запустить оценку",
+        "Export PDF": "Экспорт PDF",
+        "Open Report Folder": "Открыть папку отчётов",
+        "Recognition threshold": "Порог распознавания",
+        "Run Recognition": "Запустить распознавание",
+        "Clear": "Очистить",
+        "Query": "Запрос",
+        "Gallery": "Галерея",
+        "Album": "Альбом",
+        "Album directories": "Папки альбомов",
+        "Import / Refresh": "Импорт / обновление",
+        "Rebuild All": "Перестроить всё",
+        "Add Folder": "Добавить папку",
+        "Remove Selected": "Удалить выбранное",
+        "Cosine threshold": "Косинусный порог",
+        "Min samples": "Мин. образцов",
+        "Minimum face size": "Мин. размер лица",
+        "Thumbnail": "Миниатюра",
+        "Photos": "Фото",
+        "File": "Файл",
+        "Faces in cluster": "Лиц в кластере",
+        "Run Face Swap": "Запустить замену лиц",
+        "Open Result Directory": "Открыть папку результатов",
+        "Source": "Источник",
+        "Target": "Цель",
+        "Result": "Результат",
+        "Select": "Выбрать",
+        "Remove": "Удалить",
+        "Open Workspace Folder": "Открыть рабочую папку",
+        "Exit": "Выход",
+        "Mode": "Режим",
+        "Tools": "Инструменты",
+        "Help": "Справка",
+        "About": "О программе",
+        "Responsible Use Notice": "Уведомление об ответственном использовании",
+        "Ready": "Готово",
+        "Model": "Модель",
+        "Provider": "Провайдер",
+        "DB": "БД",
+        "Coming soon": "Скоро",
+    },
+}
+
+
+_EXTRA_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "About": "关于",
+        "Add Person": "添加人员",
+        "Add as New Person": "添加为新人员",
+        "Anonymize report paths": "匿名化报告路径",
+        "Assign to Existing Person": "分配给已有人员",
+        "Auto": "自动",
+        "Browse": "浏览",
+        "Camera index": "摄像头索引",
+        "Check Cameras": "检查摄像头",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "点击导入图片或文件夹，或拖拽到这里。文件夹会递归导入。",
+        "Click to upload images or a folder, or drag them here": "点击上传图片或文件夹，或拖拽到这里",
+        "Clustering is approximate. Please review and correct groups manually.": "聚类结果是近似的，请人工复核并修正分组。",
+        "Compare": "比对",
+        "Custom model directory": "自定义模型目录",
+        "DBSCAN distance threshold": "DBSCAN 距离阈值",
+        "Delete": "删除",
+        "Delete Selected": "删除所选",
+        "Download Selected": "下载所选",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "请先在 Models > Downloads 中下载 GFPGANv1.4.onnx。",
+        "Download inswapper_128.onnx from Models > Downloads first.": "请先在 Models > Downloads 中下载 inswapper_128.onnx。",
+        "Downloads are manual only. The GUI does not auto-download models. Model files may have different licenses from code; review usage before deployment.": "模型下载仅由用户手动触发，GUI 不会自动下载模型。模型文件和代码可能适用不同授权，部署前请核查使用范围。",
+        "Drag one or more album folders here, or use Add Folder.": "将一个或多个相册文件夹拖到这里，或使用“添加文件夹”。",
+        "Export": "导出",
+        "Export All Summary": "导出全部摘要",
+        "Export Annotated Image": "导出标注图片",
+        "Export CSV": "导出 CSV",
+        "Export CSV / JSON": "导出 CSV / JSON",
+        "Export Result": "导出结果",
+        "Export Results": "导出结果",
+        "Export Settings": "导出设置",
+        "Export Summary": "导出摘要",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "根据使用场景和模型授权，换脸功能可能需要单独的商业授权。请仅在拥有适当权利和同意的情况下使用。",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "边界框宽度或高度小于该值的人脸会被跳过。",
+        "Frame interval": "帧间隔",
+        "Frame skip": "跳帧",
+        "ID": "ID",
+        "Identify people": "识别人员",
+        "Import Image or Folder": "导入图片或文件夹",
+        "Import Settings": "导入设置",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "DBSCAN 核心点所需的最少邻近人脸数。",
+        "Name": "名称",
+        "No downloaded GFPGAN models": "没有已下载的 GFPGAN 模型",
+        "No downloaded swap models": "没有已下载的换脸模型",
+        "No images": "没有图片",
+        "Only downloaded GFPGAN restore models are shown.": "仅显示已下载的 GFPGAN 修复模型。",
+        "Only downloaded swap models are shown.": "仅显示已下载的换脸模型。",
+        "Open": "打开",
+        "Open GitHub Releases": "打开 GitHub Releases",
+        "Open License Center": "打开授权中心",
+        "Open Model Folder": "打开模型目录",
+        "Open Reports": "打开报告",
+        "Open Selected Report": "打开所选报告",
+        "People": "人员",
+        "Privacy": "隐私",
+        "Production or commercial deployment requires an appropriate commercial model license. This page summarizes common commercial paths; it does not provide legal advice.": "生产或商业部署需要适当的商业模型授权。本页总结常见商业路径，但不构成法律建议。",
+        "Re-export Selected": "重新导出所选",
+        "Recognize Faces": "识别人脸",
+        "Recursive": "递归",
+        "Refresh": "刷新",
+        "Refresh Download URLs": "刷新下载地址",
+        "Refresh People": "刷新人员",
+        "Register Current Face": "注册当前人脸",
+        "Reset": "重置",
+        "Reset Settings": "重置设置",
+        "Save Details": "保存详情",
+        "Save crops": "保存裁剪图",
+        "Save face crops": "保存人脸裁剪图",
+        "Save recognition logs": "保存识别日志",
+        "Scan and Cluster": "扫描并聚类",
+        "Search": "搜索",
+        "Select Folder": "选择文件夹",
+        "Select Image": "选择图片",
+        "Select gallery images or folders": "选择 Gallery 图片或文件夹",
+        "Select one or more image files or folders. Folders are scanned recursively.": "选择一个或多个图片文件或文件夹。文件夹会递归扫描。",
+        "Start": "开始",
+        "Stop": "停止",
+        "Target person": "目标人员",
+        "Test Model Load": "测试模型加载",
+        "Threshold": "阈值",
+        "Use Selected Model": "使用所选模型",
+        "Warmup": "预热",
+        "quality_score": "质量分数",
+        "rank": "排名",
+        "thumbnail": "缩略图",
+        "gallery_file": "Gallery 文件",
+        "similarity": "相似度",
+        "threshold": "阈值",
+        "decision": "判定",
+        "det_score": "检测分数",
+        "person_id": "人员 ID",
+        "person_name": "人员名称",
+        "status": "状态",
+        "sample_id": "样本 ID",
+        "crop_path": "裁剪路径",
+        "image_path": "图片路径",
+        "face_index": "人脸序号",
+        "bbox": "检测框",
+        "video_path": "视频路径",
+        "timestamp_ms": "时间戳 ms",
+        "timestamp_hhmmss": "时间",
+        "frame_index": "帧序号",
+        "id": "ID",
+        "date": "日期",
+        "scenario": "场景",
+        "model": "模型",
+        "provider": "运行后端",
+        "report_path": "报告路径",
+        "asset": "资源",
+        "source": "来源",
+        "kind": "类型",
+        "release": "版本",
+        "size": "大小",
+        "updated_at": "更新时间",
+        "local status": "本地状态",
+        "download url": "下载地址",
+        "Close": "关闭",
+        "First Launch Wizard": "首次启动向导",
+        "Workspace": "工作目录",
+        "Model package": "模型包",
+        "Provider": "运行后端",
+        "License notice": "授权提示",
+        "Personal / Research": "个人 / 研究",
+        "Enterprise Evaluation": "企业评测",
+        "custom model directory": "自定义模型目录",
+        "Runtime": "运行时",
+        "Downloads": "下载",
+        "Model Settings": "模型设置",
+        "Model Downloads": "模型下载",
+        "Runtime information": "运行时信息",
+        "Detection size": "检测尺寸",
+        "Face swap model": "换脸模型",
+        "GFPGAN post-processing": "GFPGAN 后处理",
+        "GFPGAN model": "GFPGAN 模型",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "下载 GFPGAN 模型后才能启用 GFPGAN。",
+        "Run GFPGAN face restoration after face swap.": "换脸后运行 GFPGAN 人脸修复。",
+        "Enable GFPGAN restore after face swap": "换脸后启用 GFPGAN 修复",
+        "Save Settings": "保存设置",
+        "Open Model Downloads": "打开模型下载",
+        "Test Model Load": "测试模型加载",
+        "Warmup": "预热",
+        "1:1 Verification": "1:1 验证",
+        "1:N Identification": "1:N 识别",
+        "Auto Split": "自动切分",
+        "Require exactly one face": "要求恰好一张脸",
+        "Use largest face": "使用最大人脸",
+        "Use largest centered face": "使用最大且居中的人脸",
+        "Mark as skip": "标记为跳过",
+        "Identity Folders / Dataset Root": "身份文件夹 / 数据集根目录",
+        "Open Dataset Rules for detailed folder examples.": "打开“数据集规则”查看详细目录示例。",
+        "1:1 verification expects one identity per subfolder. Auto Split selects one gallery image per identity and compares every probe against every identity gallery.": "1:1 验证要求每个子目录对应一个身份。自动切分会为每个身份选择一张 gallery 图，并将每张 probe 与所有身份的 gallery 逐一比对。",
+        "1:1 verification expects one identity per subfolder. Auto Split is off, so every image is treated as a probe and full pairwise probe-vs-probe comparisons are generated.": "1:1 验证要求每个子目录对应一个身份。关闭自动切分后，每张图片都作为 probe，并生成完整的 probe-vs-probe 两两比对。",
+        "1:N Auto Split can use identities/<identity folders> or identity folders directly. Each identity must provide a gallery image and at least one probe.": "1:N 自动切分可使用 identities/<身份文件夹>，也可以直接使用身份文件夹。每个身份必须能生成一张 gallery 图和至少一张 probe 图。",
+        "1:N structured evaluation requires gallery/<identity> and probe/<identity>. unknown/ is optional and contributes impostor scores for FAR thresholds.": "1:N 结构化评测要求包含 gallery/<身份> 和 probe/<身份>。unknown/ 为可选目录，会为 FAR 阈值提供冒名负样本分数。",
+        "Dataset validation passed. Evaluation is ready to run.": "数据集核验通过，可以运行评测。",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "数据集核验失败。请先修复列出的问题再运行评测。",
+        "Evaluation Dataset Rules": "评测数据集规则",
+        "Prepare local identity folders using one of the supported layouts below.": "请按以下任一支持结构准备本地身份文件夹。",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "打开 1:1、1:N 和自动切分的身份文件夹详细规则。",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "自动从文件名包含 gallery 的图片中选择每个身份的 gallery 图；如果不存在，则选择排序后的第一张图片。",
+        "Choose how evaluation handles images where more than one face is detected.": "选择评测中检测到多张人脸时的处理方式。",
+        "Similarity threshold for matching query and gallery faces.": "用于匹配查询人脸和底库人脸的相似度阈值。",
+        "Choose how query and gallery images handle multiple detected faces.": "选择查询图和底库图中检测到多张人脸时的处理方式。",
+        "Gallery face embeddings are cached in memory after the first run and reused while the gallery image list and multi-face policy are unchanged. Changing only the query image reuses the cached gallery; adding, removing, clearing gallery images, or changing multi-face handling clears and recomputes it.": "Gallery 人脸特征在首次运行后会常驻内存；只要底库图片列表和多人脸策略不变就会复用。仅更换 Query 会复用缓存；添加、移除、清空 Gallery 图片或修改多人脸处理策略会清空并重新计算。",
+        "Multi-face policy: require exactly one face. Images with multiple detected faces fail with a clear error.": "多人脸策略：要求恰好一张脸。检测到多张脸的图片会明确报错。",
+        "Multi-face policy: use largest face. If an image has multiple faces, the largest detected face is used.": "多人脸策略：使用最大人脸。如果图片中有多张脸，将使用检测框最大的那张脸。",
+        "Multi-face policy: use largest centered face. If an image has multiple faces, the face with the best area-minus-center-distance score is used.": "多人脸策略：使用最大且居中的人脸。如果图片中有多张脸，将使用面积与中心位置综合评分最高的人脸。",
+        "Multi-face policy: mark as skip. Gallery images with multiple detected faces are skipped; a multi-face query stops the current run.": "多人脸策略：标记为跳过。Gallery 中多脸图片会跳过；Query 中多脸会停止当前运行。",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "打开 InsightFace 联系页面，获取企业评测、授权、SDK/API、SLA 或定制训练支持。",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider 可用。",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider 不可用。请先安装匹配的 onnxruntime-gpu、CUDA 运行时和 GPU 驱动。",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Auto 会在 CUDAExecutionProvider 可用时使用 CUDA，否则使用 CPU。",
+        "Click to open the saved result file.": "点击打开已保存的结果文件。",
+    },
+    "ja": {
+        "Close": "閉じる",
+        "First Launch Wizard": "初回起動ウィザード",
+        "Workspace": "ワークスペース",
+        "Model package": "モデルパッケージ",
+        "License notice": "ライセンス通知",
+        "Personal / Research": "個人 / 研究",
+        "Runtime": "ランタイム",
+        "Downloads": "ダウンロード",
+        "Model Settings": "モデル設定",
+        "Model Downloads": "モデルダウンロード",
+        "Runtime information": "ランタイム情報",
+        "Detection size": "検出サイズ",
+        "Face swap model": "顔交換モデル",
+        "GFPGAN post-processing": "GFPGAN 後処理",
+        "GFPGAN model": "GFPGAN モデル",
+        "1:1 Verification": "1:1 検証",
+        "1:N Identification": "1:N 識別",
+        "Auto Split": "自動分割",
+        "Require exactly one face": "顔を1つだけ要求",
+        "Use largest face": "最大の顔を使用",
+        "Use largest centered face": "中央寄りの最大顔を使用",
+        "Mark as skip": "スキップとして扱う",
+        "Identity Folders / Dataset Root": "IDフォルダ / データセットルート",
+        "Open Dataset Rules for detailed folder examples.": "詳細なフォルダ例は「データセット規則」を開いて確認してください。",
+        "Dataset validation passed. Evaluation is ready to run.": "データセット検証に合格しました。評価を実行できます。",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "データセット検証に失敗しました。評価前に表示された問題を修正してください。",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "1:1、1:N、自動分割の詳細なIDフォルダ規則を開きます。",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "各IDの gallery 画像を、ファイル名に gallery を含む画像、またはソート順の最初の画像から自動選択します。",
+        "Choose how evaluation handles images where more than one face is detected.": "評価時に複数の顔が検出された画像の処理方法を選択します。",
+        "Similarity threshold for matching query and gallery faces.": "クエリ顔とギャラリー顔の照合に使う類似度しきい値です。",
+        "Choose how query and gallery images handle multiple detected faces.": "クエリ画像とギャラリー画像で複数顔が検出された場合の処理方法を選択します。",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "エンタープライズ評価、ライセンス、SDK/API、SLA、カスタム学習の相談ページを開きます。",
+        "Click to open the saved result file.": "クリックして保存済み結果ファイルを開きます。",
+    },
+    "ko": {
+        "Close": "닫기",
+        "First Launch Wizard": "첫 실행 마법사",
+        "Workspace": "워크스페이스",
+        "Model package": "모델 패키지",
+        "License notice": "라이선스 안내",
+        "Personal / Research": "개인 / 연구",
+        "Runtime": "런타임",
+        "Downloads": "다운로드",
+        "Model Settings": "모델 설정",
+        "Model Downloads": "모델 다운로드",
+        "Runtime information": "런타임 정보",
+        "Detection size": "검출 크기",
+        "Face swap model": "얼굴 교체 모델",
+        "GFPGAN post-processing": "GFPGAN 후처리",
+        "GFPGAN model": "GFPGAN 모델",
+        "1:1 Verification": "1:1 검증",
+        "1:N Identification": "1:N 식별",
+        "Auto Split": "자동 분할",
+        "Require exactly one face": "정확히 한 얼굴만 허용",
+        "Use largest face": "가장 큰 얼굴 사용",
+        "Use largest centered face": "중앙에 가까운 가장 큰 얼굴 사용",
+        "Mark as skip": "건너뛰기로 처리",
+        "Identity Folders / Dataset Root": "ID 폴더 / 데이터셋 루트",
+        "Open Dataset Rules for detailed folder examples.": "자세한 폴더 예시는 데이터셋 규칙을 여십시오.",
+        "Dataset validation passed. Evaluation is ready to run.": "데이터셋 검증을 통과했습니다. 평가를 실행할 수 있습니다.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "데이터셋 검증에 실패했습니다. 평가 전에 표시된 문제를 수정하십시오.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "1:1, 1:N, 자동 분할의 상세 ID 폴더 규칙을 엽니다.",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "각 ID의 gallery 이미지를 파일명에 gallery가 포함된 이미지 또는 정렬된 첫 이미지에서 자동 선택합니다.",
+        "Choose how evaluation handles images where more than one face is detected.": "평가 중 여러 얼굴이 검출된 이미지의 처리 방식을 선택합니다.",
+        "Similarity threshold for matching query and gallery faces.": "쿼리 얼굴과 갤러리 얼굴 매칭에 사용할 유사도 임계값입니다.",
+        "Choose how query and gallery images handle multiple detected faces.": "쿼리 및 갤러리 이미지에서 여러 얼굴이 검출될 때의 처리 방식을 선택합니다.",
+        "Click to open the saved result file.": "저장된 결과 파일을 열려면 클릭하십시오.",
+    },
+    "es": {
+        "Close": "Cerrar",
+        "First Launch Wizard": "Asistente de primer inicio",
+        "Workspace": "Espacio de trabajo",
+        "Model package": "Paquete de modelo",
+        "License notice": "Aviso de licencia",
+        "Personal / Research": "Personal / investigación",
+        "Runtime": "Ejecución",
+        "Downloads": "Descargas",
+        "Model Settings": "Configuración de modelos",
+        "Model Downloads": "Descargas de modelos",
+        "Runtime information": "Información de ejecución",
+        "Detection size": "Tamaño de detección",
+        "Face swap model": "Modelo de intercambio facial",
+        "GFPGAN post-processing": "Posprocesamiento GFPGAN",
+        "GFPGAN model": "Modelo GFPGAN",
+        "1:1 Verification": "Verificación 1:1",
+        "1:N Identification": "Identificación 1:N",
+        "Auto Split": "División automática",
+        "Require exactly one face": "Exigir exactamente una cara",
+        "Use largest face": "Usar la cara más grande",
+        "Use largest centered face": "Usar la cara más grande y centrada",
+        "Mark as skip": "Marcar como omitida",
+        "Identity Folders / Dataset Root": "Carpetas de identidad / raíz del dataset",
+        "Open Dataset Rules for detailed folder examples.": "Abra Reglas del dataset para ver ejemplos detallados de carpetas.",
+        "Dataset validation passed. Evaluation is ready to run.": "La validación del dataset se completó correctamente. La evaluación está lista.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "La validación del dataset falló. Corrija los problemas listados antes de ejecutar la evaluación.",
+        "Choose how evaluation handles images where more than one face is detected.": "Elija cómo tratar imágenes con más de una cara detectada durante la evaluación.",
+        "Similarity threshold for matching query and gallery faces.": "Umbral de similitud para comparar caras de consulta y galería.",
+        "Choose how query and gallery images handle multiple detected faces.": "Elija cómo tratar varias caras detectadas en imágenes de consulta y galería.",
+        "Click to open the saved result file.": "Haga clic para abrir el archivo de resultado guardado.",
+    },
+    "fr": {
+        "Close": "Fermer",
+        "First Launch Wizard": "Assistant de premier lancement",
+        "Workspace": "Espace de travail",
+        "Model package": "Package de modèle",
+        "License notice": "Avis de licence",
+        "Personal / Research": "Personnel / recherche",
+        "Runtime": "Exécution",
+        "Downloads": "Téléchargements",
+        "Model Settings": "Paramètres des modèles",
+        "Model Downloads": "Téléchargements de modèles",
+        "Runtime information": "Informations d’exécution",
+        "Detection size": "Taille de détection",
+        "Face swap model": "Modèle de face swap",
+        "GFPGAN post-processing": "Post-traitement GFPGAN",
+        "GFPGAN model": "Modèle GFPGAN",
+        "1:1 Verification": "Vérification 1:1",
+        "1:N Identification": "Identification 1:N",
+        "Auto Split": "Découpage automatique",
+        "Require exactly one face": "Exiger exactement un visage",
+        "Use largest face": "Utiliser le plus grand visage",
+        "Use largest centered face": "Utiliser le plus grand visage centré",
+        "Mark as skip": "Marquer comme ignoré",
+        "Identity Folders / Dataset Root": "Dossiers d’identité / racine du jeu de données",
+        "Open Dataset Rules for detailed folder examples.": "Ouvrez les règles du jeu de données pour des exemples de dossiers détaillés.",
+        "Dataset validation passed. Evaluation is ready to run.": "La validation du jeu de données a réussi. L’évaluation peut être lancée.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "La validation du jeu de données a échoué. Corrigez les problèmes listés avant l’évaluation.",
+        "Choose how evaluation handles images where more than one face is detected.": "Choisissez comment traiter les images contenant plusieurs visages détectés.",
+        "Similarity threshold for matching query and gallery faces.": "Seuil de similarité pour associer les visages requête et galerie.",
+        "Choose how query and gallery images handle multiple detected faces.": "Choisissez comment traiter plusieurs visages détectés dans la requête et la galerie.",
+        "Click to open the saved result file.": "Cliquez pour ouvrir le fichier de résultat enregistré.",
+    },
+    "de": {
+        "Close": "Schließen",
+        "First Launch Wizard": "Assistent für den ersten Start",
+        "Workspace": "Arbeitsbereich",
+        "Model package": "Modellpaket",
+        "License notice": "Lizenzhinweis",
+        "Personal / Research": "Privat / Forschung",
+        "Runtime": "Laufzeit",
+        "Downloads": "Downloads",
+        "Model Settings": "Modelleinstellungen",
+        "Model Downloads": "Modelldownloads",
+        "Runtime information": "Laufzeitinformationen",
+        "Detection size": "Erkennungsgröße",
+        "Face swap model": "Face-Swap-Modell",
+        "GFPGAN post-processing": "GFPGAN-Nachbearbeitung",
+        "GFPGAN model": "GFPGAN-Modell",
+        "1:1 Verification": "1:1-Verifikation",
+        "1:N Identification": "1:N-Identifikation",
+        "Auto Split": "Automatische Aufteilung",
+        "Require exactly one face": "Genau ein Gesicht verlangen",
+        "Use largest face": "Größtes Gesicht verwenden",
+        "Use largest centered face": "Größtes zentriertes Gesicht verwenden",
+        "Mark as skip": "Als überspringen markieren",
+        "Identity Folders / Dataset Root": "Identitätsordner / Datensatzwurzel",
+        "Open Dataset Rules for detailed folder examples.": "Öffnen Sie die Datensatzregeln für detaillierte Ordnerbeispiele.",
+        "Dataset validation passed. Evaluation is ready to run.": "Die Datensatzvalidierung war erfolgreich. Die Evaluierung kann gestartet werden.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "Die Datensatzvalidierung ist fehlgeschlagen. Beheben Sie die aufgeführten Probleme vor der Evaluierung.",
+        "Choose how evaluation handles images where more than one face is detected.": "Wählen Sie, wie Bilder mit mehr als einem erkannten Gesicht behandelt werden.",
+        "Similarity threshold for matching query and gallery faces.": "Ähnlichkeitsschwelle für den Abgleich von Abfrage- und Galeriegesichtern.",
+        "Choose how query and gallery images handle multiple detected faces.": "Wählen Sie, wie mehrere erkannte Gesichter in Abfrage- und Galeriebildern behandelt werden.",
+        "Click to open the saved result file.": "Klicken, um die gespeicherte Ergebnisdatei zu öffnen.",
+    },
+    "pt": {
+        "Close": "Fechar",
+        "First Launch Wizard": "Assistente do primeiro arranque",
+        "Workspace": "Workspace",
+        "Model package": "Pacote de modelo",
+        "License notice": "Aviso de licença",
+        "Personal / Research": "Pessoal / investigação",
+        "Runtime": "Execução",
+        "Downloads": "Transferências",
+        "Model Settings": "Definições de modelos",
+        "Model Downloads": "Transferências de modelos",
+        "Runtime information": "Informação de execução",
+        "Detection size": "Tamanho de deteção",
+        "Face swap model": "Modelo de troca de rostos",
+        "GFPGAN post-processing": "Pós-processamento GFPGAN",
+        "GFPGAN model": "Modelo GFPGAN",
+        "1:1 Verification": "Verificação 1:1",
+        "1:N Identification": "Identificação 1:N",
+        "Auto Split": "Divisão automática",
+        "Require exactly one face": "Exigir exatamente um rosto",
+        "Use largest face": "Usar o maior rosto",
+        "Use largest centered face": "Usar o maior rosto centrado",
+        "Mark as skip": "Marcar como ignorado",
+        "Identity Folders / Dataset Root": "Pastas de identidade / raiz do dataset",
+        "Open Dataset Rules for detailed folder examples.": "Abra as regras do dataset para exemplos detalhados de pastas.",
+        "Dataset validation passed. Evaluation is ready to run.": "A validação do dataset passou. A avaliação está pronta.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "A validação do dataset falhou. Corrija os problemas listados antes da avaliação.",
+        "Choose how evaluation handles images where more than one face is detected.": "Escolha como a avaliação trata imagens com mais de um rosto detetado.",
+        "Similarity threshold for matching query and gallery faces.": "Limiar de similaridade para comparar rostos de consulta e galeria.",
+        "Choose how query and gallery images handle multiple detected faces.": "Escolha como tratar múltiplos rostos em imagens de consulta e galeria.",
+        "Click to open the saved result file.": "Clique para abrir o ficheiro de resultado guardado.",
+    },
+    "ru": {
+        "Close": "Закрыть",
+        "First Launch Wizard": "Мастер первого запуска",
+        "Workspace": "Рабочая область",
+        "Model package": "Пакет модели",
+        "License notice": "Уведомление о лицензии",
+        "Personal / Research": "Личное / исследовательское",
+        "Runtime": "Среда выполнения",
+        "Downloads": "Загрузки",
+        "Model Settings": "Настройки моделей",
+        "Model Downloads": "Загрузки моделей",
+        "Runtime information": "Информация о среде выполнения",
+        "Detection size": "Размер детекции",
+        "Face swap model": "Модель замены лиц",
+        "GFPGAN post-processing": "Постобработка GFPGAN",
+        "GFPGAN model": "Модель GFPGAN",
+        "1:1 Verification": "Верификация 1:1",
+        "1:N Identification": "Идентификация 1:N",
+        "Auto Split": "Авторазделение",
+        "Require exactly one face": "Требовать ровно одно лицо",
+        "Use largest face": "Использовать самое крупное лицо",
+        "Use largest centered face": "Использовать самое крупное центральное лицо",
+        "Mark as skip": "Помечать как пропуск",
+        "Identity Folders / Dataset Root": "Папки идентичностей / корень датасета",
+        "Open Dataset Rules for detailed folder examples.": "Откройте правила датасета для подробных примеров структуры папок.",
+        "Dataset validation passed. Evaluation is ready to run.": "Проверка датасета пройдена. Оценку можно запускать.",
+        "Dataset validation failed. Fix the listed issues before running evaluation.": "Проверка датасета не пройдена. Исправьте перечисленные проблемы перед запуском оценки.",
+        "Choose how evaluation handles images where more than one face is detected.": "Выберите, как оценка будет обрабатывать изображения с несколькими найденными лицами.",
+        "Similarity threshold for matching query and gallery faces.": "Порог сходства для сопоставления лиц запроса и галереи.",
+        "Choose how query and gallery images handle multiple detected faces.": "Выберите обработку нескольких лиц в изображениях запроса и галереи.",
+        "Click to open the saved result file.": "Нажмите, чтобы открыть сохранённый файл результата.",
+    },
+}
+
+for _lang, _values in _EXTRA_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
+
+
+_ADDITIONAL_UI_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "<a href='#'>Browse</a>": "<a href='#'>浏览</a>",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "本应用在本地运行。代码与模型文件可能适用不同授权。研究模型可能仅限非商业使用。商业部署需要相应的模型授权。本工具不提供法律建议。",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "余弦阈值是人脸相似度截断值：数值越高，聚类越严格。DBSCAN 内部使用的余弦距离为 1 - 余弦阈值。",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "余弦相似度阈值。数值越高，聚类越严格；默认值为 0.48。",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "当前机器不可用 CUDA，因此 Auto 将使用 CPU。",
+        "Name, notes, and tags are shown here.": "姓名、备注和标签会显示在这里。",
+        "cluster_id": "聚类 ID",
+        "face_count": "人脸数",
+        "photo_count": "照片数",
+        "average_quality": "平均质量",
+        "sample_paths": "样本路径",
+        "model_name": "模型名称",
+        "source_image_path": "源图片路径",
+        "Min det score %": "最低检测分数 %",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+    },
+    "ja": {
+        "<a href='#'>Browse</a>": "<a href='#'>参照</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "カスタムモデルディレクトリ",
+        "Custom model directory": "カスタムモデルディレクトリ",
+        "Auto": "自動",
+        "Browse": "参照",
+        "Add Person": "人物を追加",
+        "Add as New Person": "新しい人物として追加",
+        "Assign to Existing Person": "既存の人物に割り当て",
+        "Compare": "比較",
+        "Search": "検索",
+        "Export": "エクスポート",
+        "Export Result": "結果をエクスポート",
+        "Export Results": "結果をエクスポート",
+        "Export Summary": "概要をエクスポート",
+        "Export Settings": "設定をエクスポート",
+        "Import Settings": "設定をインポート",
+        "Delete": "削除",
+        "Delete Selected": "選択項目を削除",
+        "Reset": "リセット",
+        "Reset Settings": "設定をリセット",
+        "Refresh": "更新",
+        "Start": "開始",
+        "Stop": "停止",
+        "Open": "開く",
+        "Save Settings": "設定を保存",
+        "Save Details": "詳細を保存",
+        "Save crops": "クロップを保存",
+        "Save face crops": "顔クロップを保存",
+        "Save recognition logs": "認識ログを保存",
+        "Anonymize report paths": "レポートパスを匿名化",
+        "Privacy": "プライバシー",
+        "Recursive": "再帰的に処理",
+        "Identify people": "人物を識別",
+        "Camera index": "カメラインデックス",
+        "Frame skip": "フレームスキップ",
+        "Frame interval": "フレーム間隔",
+        "Target person": "対象人物",
+        "Check Cameras": "カメラを確認",
+        "Register Current Face": "現在の顔を登録",
+        "Refresh People": "人物リストを更新",
+        "Recognize Faces": "顔を認識",
+        "Test Model Load": "モデル読み込みをテスト",
+        "Warmup": "ウォームアップ",
+        "Export Annotated Image": "注釈付き画像をエクスポート",
+        "Export CSV": "CSV をエクスポート",
+        "Export CSV / JSON": "CSV / JSON をエクスポート",
+        "Refresh Download URLs": "ダウンロード URL を更新",
+        "Download Selected": "選択項目をダウンロード",
+        "Use Selected Model": "選択したモデルを使用",
+        "Open Model Folder": "モデルフォルダを開く",
+        "Open GitHub Releases": "GitHub Releases を開く",
+        "Open Model Downloads": "モデルダウンロードを開く",
+        "Open License Center": "ライセンスセンターを開く",
+        "Open Reports": "レポートを開く",
+        "Open Selected Report": "選択したレポートを開く",
+        "Re-export Selected": "選択項目を再エクスポート",
+        "Export All Summary": "全概要をエクスポート",
+        "Import Image or Folder": "画像またはフォルダをインポート",
+        "Select Image": "画像を選択",
+        "Select Folder": "フォルダを選択",
+        "Select gallery images or folders": "ギャラリー画像またはフォルダを選択",
+        "Select one or more image files or folders. Folders are scanned recursively.": "1つ以上の画像ファイルまたはフォルダを選択してください。フォルダは再帰的にスキャンされます。",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "画像またはフォルダをクリックしてインポートするか、ここにドラッグしてください。フォルダは再帰的にインポートされます。",
+        "Click to upload images or a folder, or drag them here": "画像またはフォルダをクリックしてアップロードするか、ここにドラッグしてください",
+        "Drag one or more album folders here, or use Add Folder.": "1つ以上のアルバムフォルダをここにドラッグするか、「フォルダを追加」を使用してください。",
+        "Drag album folders into the directory list, or click Add Folder.": "アルバムフォルダをディレクトリ一覧にドラッグするか、「フォルダを追加」をクリックしてください。",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "コサインしきい値は顔類似度の判定基準です。値が高いほどクラスタリングは厳しくなります。DBSCAN では内部的にコサイン距離 = 1 - コサインしきい値を使用します。",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "コサイン類似度のしきい値です。値が高いほどクラスタリングは厳しくなります。既定値は 0.48 です。",
+        "DBSCAN distance threshold": "DBSCAN 距離しきい値",
+        "Algorithm: DBSCAN": "アルゴリズム: DBSCAN",
+        "Scan and Cluster": "スキャンしてクラスタリング",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "DBSCAN のコア点に必要な近傍顔数です。",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "バウンディングボックスの幅または高さがこの値より小さい顔はスキップされます。",
+        "No images": "画像なし",
+        "People": "人物",
+        "ID": "ID",
+        "Name": "名前",
+        "Name, notes, and tags are shown here.": "名前、メモ、タグがここに表示されます。",
+        "Threshold": "しきい値",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "このアプリケーションはローカルで動作します。コードとモデルのライセンスは異なる場合があります。研究用モデルは非商用利用に制限されることがあります。商用展開には適切なモデルライセンスが必要です。本ツールは法的助言を提供しません。",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "このマシンでは CUDA を利用できないため、自動設定では CPU を使用します。",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "CUDAExecutionProvider が利用可能な場合は自動で CUDA を使用し、それ以外は CPU を使用します。",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider を利用できます。",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider を利用できません。対応する onnxruntime-gpu、CUDA ランタイム、GPU ドライバを先にインストールしてください。",
+        "No downloaded swap models": "ダウンロード済みの顔交換モデルがありません",
+        "Only downloaded swap models are shown.": "ダウンロード済みの顔交換モデルのみ表示されます。",
+        "Download inswapper_128.onnx from Models > Downloads first.": "先に Models > Downloads から inswapper_128.onnx をダウンロードしてください。",
+        "No downloaded GFPGAN models": "ダウンロード済みの GFPGAN モデルがありません",
+        "Only downloaded GFPGAN restore models are shown.": "ダウンロード済みの GFPGAN 復元モデルのみ表示されます。",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "先に Models > Downloads から GFPGANv1.4.onnx をダウンロードしてください。",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN モデルをダウンロードするまで GFPGAN は利用できません。",
+        "Run GFPGAN face restoration after face swap.": "顔交換後に GFPGAN 顔復元を実行します。",
+        "Enable GFPGAN restore after face swap": "顔交換後に GFPGAN 復元を有効化",
+        "Validate the dataset before running an evaluation.": "評価を実行する前にデータセットを検証してください。",
+        "Validate the selected dataset and configuration before running evaluation.": "評価を実行する前に、選択したデータセットと設定を検証してください。",
+        "Result preview appears here. Video results are saved to the exports folder.": "結果プレビューがここに表示されます。動画結果はエクスポートフォルダに保存されます。",
+        "Min det score %": "最小検出スコア %",
+        "asset": "アセット",
+        "source": "ソース",
+        "kind": "種類",
+        "release": "リリース",
+        "size": "サイズ",
+        "updated_at": "更新日時",
+        "local status": "ローカル状態",
+        "download url": "ダウンロード URL",
+        "rank": "順位",
+        "thumbnail": "サムネイル",
+        "gallery_file": "ギャラリーファイル",
+        "similarity": "類似度",
+        "threshold": "しきい値",
+        "decision": "判定",
+        "det_score": "検出スコア",
+        "person_id": "人物 ID",
+        "person_name": "人物名",
+        "status": "状態",
+        "sample_id": "サンプル ID",
+        "crop_path": "クロップパス",
+        "image_path": "画像パス",
+        "face_index": "顔インデックス",
+        "bbox": "バウンディングボックス",
+        "video_path": "動画パス",
+        "timestamp_ms": "タイムスタンプ ms",
+        "timestamp_hhmmss": "時刻",
+        "frame_index": "フレーム番号",
+        "id": "ID",
+        "date": "日付",
+        "scenario": "シナリオ",
+        "model": "モデル",
+        "provider": "プロバイダ",
+        "report_path": "レポートパス",
+        "quality_score": "品質スコア",
+        "cluster_id": "クラスタ ID",
+        "face_count": "顔数",
+        "photo_count": "写真数",
+        "average_quality": "平均品質",
+        "sample_paths": "サンプルパス",
+        "model_name": "モデル名",
+        "source_image_path": "元画像パス",
+    },
+    "ko": {
+        "<a href='#'>Browse</a>": "<a href='#'>찾아보기</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "사용자 지정 모델 디렉터리",
+        "Custom model directory": "사용자 지정 모델 디렉터리",
+        "Auto": "자동",
+        "Browse": "찾아보기",
+        "Add Person": "사람 추가",
+        "Add as New Person": "새 사람으로 추가",
+        "Assign to Existing Person": "기존 사람에게 지정",
+        "Compare": "비교",
+        "Search": "검색",
+        "Export": "내보내기",
+        "Export Result": "결과 내보내기",
+        "Export Results": "결과 내보내기",
+        "Export Summary": "요약 내보내기",
+        "Export Settings": "설정 내보내기",
+        "Import Settings": "설정 가져오기",
+        "Delete": "삭제",
+        "Delete Selected": "선택 항목 삭제",
+        "Reset": "초기화",
+        "Reset Settings": "설정 초기화",
+        "Refresh": "새로고침",
+        "Start": "시작",
+        "Stop": "중지",
+        "Open": "열기",
+        "Save Settings": "설정 저장",
+        "Save Details": "세부 정보 저장",
+        "Save crops": "크롭 저장",
+        "Save face crops": "얼굴 크롭 저장",
+        "Save recognition logs": "인식 로그 저장",
+        "Anonymize report paths": "보고서 경로 익명화",
+        "Privacy": "개인정보",
+        "Recursive": "재귀적으로 처리",
+        "Identify people": "사람 식별",
+        "Camera index": "카메라 인덱스",
+        "Frame skip": "프레임 건너뛰기",
+        "Frame interval": "프레임 간격",
+        "Target person": "대상 사람",
+        "Check Cameras": "카메라 확인",
+        "Register Current Face": "현재 얼굴 등록",
+        "Refresh People": "사람 목록 새로고침",
+        "Recognize Faces": "얼굴 인식",
+        "Test Model Load": "모델 로드 테스트",
+        "Warmup": "워밍업",
+        "Export Annotated Image": "주석 이미지 내보내기",
+        "Export CSV": "CSV 내보내기",
+        "Export CSV / JSON": "CSV / JSON 내보내기",
+        "Refresh Download URLs": "다운로드 URL 새로고침",
+        "Download Selected": "선택 항목 다운로드",
+        "Use Selected Model": "선택한 모델 사용",
+        "Open Model Folder": "모델 폴더 열기",
+        "Open GitHub Releases": "GitHub Releases 열기",
+        "Open Model Downloads": "모델 다운로드 열기",
+        "Open License Center": "라이선스 센터 열기",
+        "Open Reports": "보고서 열기",
+        "Open Selected Report": "선택한 보고서 열기",
+        "Re-export Selected": "선택 항목 다시 내보내기",
+        "Export All Summary": "전체 요약 내보내기",
+        "Import Image or Folder": "이미지 또는 폴더 가져오기",
+        "Select Image": "이미지 선택",
+        "Select Folder": "폴더 선택",
+        "Select gallery images or folders": "갤러리 이미지 또는 폴더 선택",
+        "Select one or more image files or folders. Folders are scanned recursively.": "하나 이상의 이미지 파일 또는 폴더를 선택하십시오. 폴더는 재귀적으로 스캔됩니다.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "이미지 또는 폴더를 클릭해 가져오거나 여기로 끌어오십시오. 폴더는 재귀적으로 가져옵니다.",
+        "Click to upload images or a folder, or drag them here": "이미지 또는 폴더를 클릭해 업로드하거나 여기로 끌어오십시오",
+        "Drag one or more album folders here, or use Add Folder.": "하나 이상의 앨범 폴더를 여기로 끌어오거나 폴더 추가를 사용하십시오.",
+        "Drag album folders into the directory list, or click Add Folder.": "앨범 폴더를 디렉터리 목록으로 끌어오거나 폴더 추가를 클릭하십시오.",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "코사인 임계값은 얼굴 유사도의 기준값입니다. 값이 높을수록 클러스터가 더 엄격해집니다. DBSCAN은 내부적으로 코사인 거리 = 1 - 코사인 임계값을 사용합니다.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "코사인 유사도 임계값입니다. 값이 높을수록 클러스터가 더 엄격해지며 기본값은 0.48입니다.",
+        "DBSCAN distance threshold": "DBSCAN 거리 임계값",
+        "Algorithm: DBSCAN": "알고리즘: DBSCAN",
+        "Scan and Cluster": "스캔 및 클러스터링",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "DBSCAN 핵심점에 필요한 최소 이웃 얼굴 수입니다.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "바운딩 박스의 너비 또는 높이가 이 값보다 작은 얼굴은 건너뜁니다.",
+        "No images": "이미지 없음",
+        "People": "사람",
+        "ID": "ID",
+        "Name": "이름",
+        "Name, notes, and tags are shown here.": "이름, 메모, 태그가 여기에 표시됩니다.",
+        "Threshold": "임계값",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "이 애플리케이션은 로컬에서 실행됩니다. 코드와 모델 라이선스는 다를 수 있습니다. 연구용 모델은 비상업적 사용으로 제한될 수 있습니다. 상용 배포에는 적절한 모델 라이선스가 필요합니다. 이 도구는 법률 자문을 제공하지 않습니다.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "이 컴퓨터에서 CUDA를 사용할 수 없어 자동 설정은 CPU를 사용합니다.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "CUDAExecutionProvider를 사용할 수 있으면 자동으로 CUDA를 사용하고, 그렇지 않으면 CPU를 사용합니다.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider를 사용할 수 있습니다.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider를 사용할 수 없습니다. 먼저 호환되는 onnxruntime-gpu, CUDA 런타임, GPU 드라이버를 설치하십시오.",
+        "No downloaded swap models": "다운로드된 얼굴 교체 모델 없음",
+        "Only downloaded swap models are shown.": "다운로드된 얼굴 교체 모델만 표시됩니다.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "먼저 Models > Downloads에서 inswapper_128.onnx를 다운로드하십시오.",
+        "No downloaded GFPGAN models": "다운로드된 GFPGAN 모델 없음",
+        "Only downloaded GFPGAN restore models are shown.": "다운로드된 GFPGAN 복원 모델만 표시됩니다.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "먼저 Models > Downloads에서 GFPGANv1.4.onnx를 다운로드하십시오.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN 모델을 다운로드하기 전에는 GFPGAN을 사용할 수 없습니다.",
+        "Run GFPGAN face restoration after face swap.": "얼굴 교체 후 GFPGAN 얼굴 복원을 실행합니다.",
+        "Enable GFPGAN restore after face swap": "얼굴 교체 후 GFPGAN 복원 활성화",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "기업 평가, 라이선스, SDK/API, SLA 또는 맞춤 학습 지원을 위해 InsightFace 문의 페이지를 엽니다.",
+        "Evaluation Dataset Rules": "평가 데이터셋 규칙",
+        "Prepare local identity folders using one of the supported layouts below.": "아래 지원 레이아웃 중 하나로 로컬 ID 폴더를 준비하십시오.",
+        "Validate the dataset before running an evaluation.": "평가를 실행하기 전에 데이터셋을 검증하십시오.",
+        "Validate the selected dataset and configuration before running evaluation.": "평가를 실행하기 전에 선택한 데이터셋과 구성을 검증하십시오.",
+        "Result preview appears here. Video results are saved to the exports folder.": "결과 미리보기가 여기에 표시됩니다. 비디오 결과는 내보내기 폴더에 저장됩니다.",
+        "Min det score %": "최소 검출 점수 %",
+        "asset": "자산",
+        "source": "소스",
+        "kind": "종류",
+        "release": "릴리스",
+        "size": "크기",
+        "updated_at": "업데이트 시간",
+        "local status": "로컬 상태",
+        "download url": "다운로드 URL",
+        "rank": "순위",
+        "thumbnail": "썸네일",
+        "gallery_file": "갤러리 파일",
+        "similarity": "유사도",
+        "threshold": "임계값",
+        "decision": "판정",
+        "det_score": "검출 점수",
+        "person_id": "사람 ID",
+        "person_name": "사람 이름",
+        "status": "상태",
+        "sample_id": "샘플 ID",
+        "crop_path": "크롭 경로",
+        "image_path": "이미지 경로",
+        "face_index": "얼굴 인덱스",
+        "bbox": "바운딩 박스",
+        "video_path": "비디오 경로",
+        "timestamp_ms": "타임스탬프 ms",
+        "timestamp_hhmmss": "시간",
+        "frame_index": "프레임 인덱스",
+        "id": "ID",
+        "date": "날짜",
+        "scenario": "시나리오",
+        "model": "모델",
+        "provider": "제공자",
+        "report_path": "보고서 경로",
+        "quality_score": "품질 점수",
+        "cluster_id": "클러스터 ID",
+        "face_count": "얼굴 수",
+        "photo_count": "사진 수",
+        "average_quality": "평균 품질",
+        "sample_paths": "샘플 경로",
+        "model_name": "모델 이름",
+        "source_image_path": "원본 이미지 경로",
+    },
+    "es": {
+        "<a href='#'>Browse</a>": "<a href='#'>Explorar</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "directorio de modelo personalizado",
+        "Custom model directory": "Directorio de modelo personalizado",
+        "Auto": "Automático",
+        "Browse": "Explorar",
+        "Add Person": "Añadir persona",
+        "Add as New Person": "Añadir como nueva persona",
+        "Assign to Existing Person": "Asignar a una persona existente",
+        "Compare": "Comparar",
+        "Search": "Buscar",
+        "Export": "Exportar",
+        "Export Result": "Exportar resultado",
+        "Export Results": "Exportar resultados",
+        "Export Summary": "Exportar resumen",
+        "Export Settings": "Exportar configuración",
+        "Import Settings": "Importar configuración",
+        "Delete": "Eliminar",
+        "Delete Selected": "Eliminar selección",
+        "Reset": "Restablecer",
+        "Reset Settings": "Restablecer configuración",
+        "Refresh": "Actualizar",
+        "Start": "Iniciar",
+        "Stop": "Detener",
+        "Open": "Abrir",
+        "Save Settings": "Guardar configuración",
+        "Save Details": "Guardar detalles",
+        "Save crops": "Guardar recortes",
+        "Save face crops": "Guardar recortes de rostro",
+        "Save recognition logs": "Guardar registros de reconocimiento",
+        "Anonymize report paths": "Anonimizar rutas de informes",
+        "Privacy": "Privacidad",
+        "Recursive": "Recursivo",
+        "Identify people": "Identificar personas",
+        "Camera index": "Índice de cámara",
+        "Frame skip": "Salto de fotogramas",
+        "Frame interval": "Intervalo de fotogramas",
+        "Target person": "Persona objetivo",
+        "Check Cameras": "Comprobar cámaras",
+        "Register Current Face": "Registrar rostro actual",
+        "Refresh People": "Actualizar personas",
+        "Recognize Faces": "Reconocer rostros",
+        "Test Model Load": "Probar carga del modelo",
+        "Warmup": "Calentamiento",
+        "Export Annotated Image": "Exportar imagen anotada",
+        "Export CSV": "Exportar CSV",
+        "Export CSV / JSON": "Exportar CSV / JSON",
+        "Refresh Download URLs": "Actualizar URL de descarga",
+        "Download Selected": "Descargar selección",
+        "Use Selected Model": "Usar modelo seleccionado",
+        "Open Model Folder": "Abrir carpeta de modelos",
+        "Open GitHub Releases": "Abrir GitHub Releases",
+        "Open Model Downloads": "Abrir descargas de modelos",
+        "Open License Center": "Abrir centro de licencias",
+        "Open Reports": "Abrir informes",
+        "Open Selected Report": "Abrir informe seleccionado",
+        "Re-export Selected": "Reexportar selección",
+        "Export All Summary": "Exportar resumen completo",
+        "Import Image or Folder": "Importar imagen o carpeta",
+        "Select Image": "Seleccionar imagen",
+        "Select Folder": "Seleccionar carpeta",
+        "Select gallery images or folders": "Seleccionar imágenes o carpetas de galería",
+        "Select one or more image files or folders. Folders are scanned recursively.": "Seleccione uno o más archivos de imagen o carpetas. Las carpetas se escanean de forma recursiva.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "Haga clic para importar una imagen o carpeta, o arrástrela aquí. Las carpetas se importan de forma recursiva.",
+        "Click to upload images or a folder, or drag them here": "Haga clic para subir imágenes o una carpeta, o arrástrelas aquí",
+        "Drag one or more album folders here, or use Add Folder.": "Arrastre aquí una o más carpetas de álbum, o use Añadir carpeta.",
+        "Drag album folders into the directory list, or click Add Folder.": "Arrastre carpetas de álbum a la lista de directorios, o haga clic en Añadir carpeta.",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "El umbral de coseno es el corte de similitud facial: valores más altos hacen que los clústeres sean más estrictos. DBSCAN usa internamente distancia de coseno = 1 - umbral de coseno.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "Umbral de similitud de coseno. Valores más altos hacen que los clústeres sean más estrictos; el valor predeterminado es 0,48.",
+        "DBSCAN distance threshold": "Umbral de distancia DBSCAN",
+        "Algorithm: DBSCAN": "Algoritmo: DBSCAN",
+        "Scan and Cluster": "Escanear y agrupar",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "Número mínimo de rostros vecinos necesarios para un punto núcleo de DBSCAN.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "Se omiten los rostros cuyo cuadro delimitador tenga anchura o altura menor que este valor.",
+        "No images": "Sin imágenes",
+        "People": "Personas",
+        "ID": "ID",
+        "Name": "Nombre",
+        "Name, notes, and tags are shown here.": "Aquí se muestran el nombre, las notas y las etiquetas.",
+        "Threshold": "Umbral",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "Esta aplicación se ejecuta localmente. Las licencias del código y de los modelos pueden ser diferentes. Los modelos de investigación pueden estar limitados a uso no comercial. El despliegue comercial requiere una licencia de modelo adecuada. Esta herramienta no ofrece asesoramiento legal.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "CUDA no está disponible en esta máquina, por lo que Automático usará CPU.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Automático usa CUDA cuando CUDAExecutionProvider está disponible; en caso contrario usa CPU.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider está disponible.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider no está disponible. Instale primero una versión compatible de onnxruntime-gpu, el runtime CUDA y el controlador GPU.",
+        "No downloaded swap models": "No hay modelos de intercambio facial descargados",
+        "Only downloaded swap models are shown.": "Solo se muestran modelos de intercambio facial descargados.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "Descargue primero inswapper_128.onnx desde Models > Downloads.",
+        "No downloaded GFPGAN models": "No hay modelos GFPGAN descargados",
+        "Only downloaded GFPGAN restore models are shown.": "Solo se muestran modelos de restauración GFPGAN descargados.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "Descargue primero GFPGANv1.4.onnx desde Models > Downloads.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN no estará disponible hasta descargar un modelo GFPGAN.",
+        "Run GFPGAN face restoration after face swap.": "Ejecutar restauración facial GFPGAN después del intercambio.",
+        "Enable GFPGAN restore after face swap": "Activar restauración GFPGAN después del intercambio facial",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "Seleccione automáticamente la imagen de galería de cada identidad desde un archivo que contenga “gallery”, o la primera imagen ordenada.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "Abra las reglas detalladas de carpetas de identidad para 1:1, 1:N y División automática.",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "Abrir la página de contacto de InsightFace para evaluación empresarial, licencias, SDK/API, SLA o formación personalizada.",
+        "Evaluation Dataset Rules": "Reglas del dataset de evaluación",
+        "Prepare local identity folders using one of the supported layouts below.": "Prepare carpetas locales de identidad usando uno de los formatos admitidos a continuación.",
+        "Validate the dataset before running an evaluation.": "Valide el dataset antes de ejecutar una evaluación.",
+        "Validate the selected dataset and configuration before running evaluation.": "Valide el dataset y la configuración seleccionados antes de ejecutar la evaluación.",
+        "Result preview appears here. Video results are saved to the exports folder.": "La vista previa del resultado aparece aquí. Los resultados de vídeo se guardan en la carpeta de exportaciones.",
+        "Min det score %": "Puntuación mínima de detección %",
+        "asset": "recurso",
+        "source": "origen",
+        "kind": "tipo",
+        "release": "versión",
+        "size": "tamaño",
+        "updated_at": "actualizado",
+        "local status": "estado local",
+        "download url": "URL de descarga",
+        "rank": "posición",
+        "thumbnail": "miniatura",
+        "gallery_file": "archivo de galería",
+        "similarity": "similitud",
+        "threshold": "umbral",
+        "decision": "decisión",
+        "det_score": "puntuación de detección",
+        "person_id": "ID de persona",
+        "person_name": "nombre de persona",
+        "status": "estado",
+        "sample_id": "ID de muestra",
+        "crop_path": "ruta del recorte",
+        "image_path": "ruta de imagen",
+        "face_index": "índice de rostro",
+        "bbox": "cuadro",
+        "video_path": "ruta de vídeo",
+        "timestamp_ms": "marca temporal ms",
+        "timestamp_hhmmss": "hora",
+        "frame_index": "índice de fotograma",
+        "id": "ID",
+        "date": "fecha",
+        "scenario": "escenario",
+        "model": "modelo",
+        "provider": "proveedor",
+        "report_path": "ruta del informe",
+        "quality_score": "puntuación de calidad",
+        "cluster_id": "ID de clúster",
+        "face_count": "rostros",
+        "photo_count": "fotos",
+        "average_quality": "calidad media",
+        "sample_paths": "rutas de muestra",
+        "model_name": "nombre del modelo",
+        "source_image_path": "ruta de imagen origen",
+    },
+    "fr": {
+        "<a href='#'>Browse</a>": "<a href='#'>Parcourir</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "répertoire de modèle personnalisé",
+        "Custom model directory": "Répertoire de modèle personnalisé",
+        "Auto": "Auto",
+        "Browse": "Parcourir",
+        "Add Person": "Ajouter une personne",
+        "Add as New Person": "Ajouter comme nouvelle personne",
+        "Assign to Existing Person": "Attribuer à une personne existante",
+        "Compare": "Comparer",
+        "Search": "Rechercher",
+        "Export": "Exporter",
+        "Export Result": "Exporter le résultat",
+        "Export Results": "Exporter les résultats",
+        "Export Summary": "Exporter le résumé",
+        "Export Settings": "Exporter les paramètres",
+        "Import Settings": "Importer les paramètres",
+        "Delete": "Supprimer",
+        "Delete Selected": "Supprimer la sélection",
+        "Reset": "Réinitialiser",
+        "Reset Settings": "Réinitialiser les paramètres",
+        "Refresh": "Actualiser",
+        "Start": "Démarrer",
+        "Stop": "Arrêter",
+        "Open": "Ouvrir",
+        "Save Settings": "Enregistrer les paramètres",
+        "Save Details": "Enregistrer les détails",
+        "Save crops": "Enregistrer les recadrages",
+        "Save face crops": "Enregistrer les recadrages de visage",
+        "Save recognition logs": "Enregistrer les journaux de reconnaissance",
+        "Anonymize report paths": "Anonymiser les chemins des rapports",
+        "Privacy": "Confidentialité",
+        "Recursive": "Récursif",
+        "Identify people": "Identifier les personnes",
+        "Camera index": "Index de caméra",
+        "Frame skip": "Saut d’images",
+        "Frame interval": "Intervalle d’images",
+        "Target person": "Personne cible",
+        "Check Cameras": "Vérifier les caméras",
+        "Register Current Face": "Enregistrer le visage actuel",
+        "Refresh People": "Actualiser les personnes",
+        "Recognize Faces": "Reconnaître les visages",
+        "Test Model Load": "Tester le chargement du modèle",
+        "Warmup": "Préchauffage",
+        "Export Annotated Image": "Exporter l’image annotée",
+        "Export CSV": "Exporter CSV",
+        "Export CSV / JSON": "Exporter CSV / JSON",
+        "Refresh Download URLs": "Actualiser les URL de téléchargement",
+        "Download Selected": "Télécharger la sélection",
+        "Use Selected Model": "Utiliser le modèle sélectionné",
+        "Open Model Folder": "Ouvrir le dossier des modèles",
+        "Open GitHub Releases": "Ouvrir GitHub Releases",
+        "Open Model Downloads": "Ouvrir les téléchargements de modèles",
+        "Open License Center": "Ouvrir le centre de licences",
+        "Open Reports": "Ouvrir les rapports",
+        "Open Selected Report": "Ouvrir le rapport sélectionné",
+        "Re-export Selected": "Réexporter la sélection",
+        "Export All Summary": "Exporter le résumé complet",
+        "Import Image or Folder": "Importer une image ou un dossier",
+        "Select Image": "Sélectionner une image",
+        "Select Folder": "Sélectionner un dossier",
+        "Select gallery images or folders": "Sélectionner des images ou dossiers de galerie",
+        "Select one or more image files or folders. Folders are scanned recursively.": "Sélectionnez un ou plusieurs fichiers image ou dossiers. Les dossiers sont analysés récursivement.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "Cliquez pour importer une image ou un dossier, ou faites-le glisser ici. Les dossiers sont importés récursivement.",
+        "Click to upload images or a folder, or drag them here": "Cliquez pour charger des images ou un dossier, ou faites-les glisser ici",
+        "Drag one or more album folders here, or use Add Folder.": "Faites glisser un ou plusieurs dossiers d’album ici, ou utilisez Ajouter un dossier.",
+        "Drag album folders into the directory list, or click Add Folder.": "Faites glisser des dossiers d’album dans la liste, ou cliquez sur Ajouter un dossier.",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "Le seuil cosinus est le seuil de similarité faciale : plus il est élevé, plus les clusters sont stricts. DBSCAN utilise en interne la distance cosinus = 1 - seuil cosinus.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "Seuil de similarité cosinus. Plus la valeur est élevée, plus les clusters sont stricts ; la valeur par défaut est 0,48.",
+        "DBSCAN distance threshold": "Seuil de distance DBSCAN",
+        "Algorithm: DBSCAN": "Algorithme : DBSCAN",
+        "Scan and Cluster": "Analyser et regrouper",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "Nombre minimal de visages voisins nécessaire pour un point cœur DBSCAN.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "Les visages dont la boîte englobante est plus petite que cette valeur en largeur ou hauteur sont ignorés.",
+        "No images": "Aucune image",
+        "People": "Personnes",
+        "ID": "ID",
+        "Name": "Nom",
+        "Name, notes, and tags are shown here.": "Le nom, les notes et les tags apparaissent ici.",
+        "Threshold": "Seuil",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "Cette application s’exécute localement. Les licences du code et des modèles peuvent différer. Les modèles de recherche peuvent être limités à un usage non commercial. Un déploiement commercial requiert une licence de modèle appropriée. Cet outil ne fournit pas de conseil juridique.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "CUDA n’est pas disponible sur cette machine ; le mode Auto utilisera donc le CPU.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Le mode Auto utilise CUDA lorsque CUDAExecutionProvider est disponible, sinon le CPU.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider est disponible.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider n’est pas disponible. Installez d’abord une version compatible de onnxruntime-gpu, le runtime CUDA et le pilote GPU.",
+        "No downloaded swap models": "Aucun modèle de face swap téléchargé",
+        "Only downloaded swap models are shown.": "Seuls les modèles de face swap téléchargés sont affichés.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "Téléchargez d’abord inswapper_128.onnx depuis Models > Downloads.",
+        "No downloaded GFPGAN models": "Aucun modèle GFPGAN téléchargé",
+        "Only downloaded GFPGAN restore models are shown.": "Seuls les modèles de restauration GFPGAN téléchargés sont affichés.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "Téléchargez d’abord GFPGANv1.4.onnx depuis Models > Downloads.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN est indisponible tant qu’un modèle GFPGAN n’a pas été téléchargé.",
+        "Run GFPGAN face restoration after face swap.": "Exécuter la restauration de visage GFPGAN après le face swap.",
+        "Enable GFPGAN restore after face swap": "Activer la restauration GFPGAN après le face swap",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "Sélectionne automatiquement l’image de galerie de chaque identité à partir d’un fichier contenant « gallery », ou la première image triée.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "Ouvre les règles détaillées des dossiers d’identité pour 1:1, 1:N et le découpage automatique.",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "Ouvrir la page de contact InsightFace pour l’évaluation entreprise, les licences, SDK/API, SLA ou l’entraînement personnalisé.",
+        "Evaluation Dataset Rules": "Règles du jeu de données d’évaluation",
+        "Prepare local identity folders using one of the supported layouts below.": "Préparez des dossiers d’identité locaux selon l’une des structures prises en charge ci-dessous.",
+        "Validate the dataset before running an evaluation.": "Validez le jeu de données avant de lancer une évaluation.",
+        "Validate the selected dataset and configuration before running evaluation.": "Validez le jeu de données et la configuration sélectionnés avant de lancer l’évaluation.",
+        "Result preview appears here. Video results are saved to the exports folder.": "L’aperçu du résultat apparaît ici. Les résultats vidéo sont enregistrés dans le dossier d’export.",
+        "Min det score %": "Score de détection min. %",
+        "asset": "ressource",
+        "source": "source",
+        "kind": "type",
+        "release": "version",
+        "size": "taille",
+        "updated_at": "mis à jour",
+        "local status": "état local",
+        "download url": "URL de téléchargement",
+        "rank": "rang",
+        "thumbnail": "vignette",
+        "gallery_file": "fichier galerie",
+        "similarity": "similarité",
+        "threshold": "seuil",
+        "decision": "décision",
+        "det_score": "score de détection",
+        "person_id": "ID personne",
+        "person_name": "nom de personne",
+        "status": "état",
+        "sample_id": "ID échantillon",
+        "crop_path": "chemin du recadrage",
+        "image_path": "chemin image",
+        "face_index": "index visage",
+        "bbox": "boîte",
+        "video_path": "chemin vidéo",
+        "timestamp_ms": "horodatage ms",
+        "timestamp_hhmmss": "temps",
+        "frame_index": "index image",
+        "id": "ID",
+        "date": "date",
+        "scenario": "scénario",
+        "model": "modèle",
+        "provider": "fournisseur",
+        "report_path": "chemin du rapport",
+        "quality_score": "score qualité",
+        "cluster_id": "ID cluster",
+        "face_count": "visages",
+        "photo_count": "photos",
+        "average_quality": "qualité moyenne",
+        "sample_paths": "chemins des échantillons",
+        "model_name": "nom du modèle",
+        "source_image_path": "chemin de l’image source",
+    },
+    "de": {
+        "<a href='#'>Browse</a>": "<a href='#'>Durchsuchen</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "benutzerdefiniertes Modellverzeichnis",
+        "Custom model directory": "Benutzerdefiniertes Modellverzeichnis",
+        "Auto": "Auto",
+        "Browse": "Durchsuchen",
+        "Add Person": "Person hinzufügen",
+        "Add as New Person": "Als neue Person hinzufügen",
+        "Assign to Existing Person": "Bestehender Person zuweisen",
+        "Compare": "Vergleichen",
+        "Search": "Suchen",
+        "Export": "Exportieren",
+        "Export Result": "Ergebnis exportieren",
+        "Export Results": "Ergebnisse exportieren",
+        "Export Summary": "Zusammenfassung exportieren",
+        "Export Settings": "Einstellungen exportieren",
+        "Import Settings": "Einstellungen importieren",
+        "Delete": "Löschen",
+        "Delete Selected": "Auswahl löschen",
+        "Reset": "Zurücksetzen",
+        "Reset Settings": "Einstellungen zurücksetzen",
+        "Refresh": "Aktualisieren",
+        "Start": "Starten",
+        "Stop": "Stoppen",
+        "Open": "Öffnen",
+        "Save Settings": "Einstellungen speichern",
+        "Save Details": "Details speichern",
+        "Save crops": "Ausschnitte speichern",
+        "Save face crops": "Gesichtsausschnitte speichern",
+        "Save recognition logs": "Erkennungsprotokolle speichern",
+        "Anonymize report paths": "Berichtspfade anonymisieren",
+        "Privacy": "Datenschutz",
+        "Recursive": "Rekursiv",
+        "Identify people": "Personen identifizieren",
+        "Camera index": "Kameraindex",
+        "Frame skip": "Frames überspringen",
+        "Frame interval": "Frameintervall",
+        "Target person": "Zielperson",
+        "Check Cameras": "Kameras prüfen",
+        "Register Current Face": "Aktuelles Gesicht registrieren",
+        "Refresh People": "Personen aktualisieren",
+        "Recognize Faces": "Gesichter erkennen",
+        "Test Model Load": "Modellladen testen",
+        "Warmup": "Aufwärmen",
+        "Export Annotated Image": "Annotiertes Bild exportieren",
+        "Export CSV": "CSV exportieren",
+        "Export CSV / JSON": "CSV / JSON exportieren",
+        "Refresh Download URLs": "Download-URLs aktualisieren",
+        "Download Selected": "Auswahl herunterladen",
+        "Use Selected Model": "Ausgewähltes Modell verwenden",
+        "Open Model Folder": "Modellordner öffnen",
+        "Open GitHub Releases": "GitHub Releases öffnen",
+        "Open Model Downloads": "Modelldownloads öffnen",
+        "Open License Center": "Lizenzcenter öffnen",
+        "Open Reports": "Berichte öffnen",
+        "Open Selected Report": "Ausgewählten Bericht öffnen",
+        "Re-export Selected": "Auswahl erneut exportieren",
+        "Export All Summary": "Gesamtübersicht exportieren",
+        "Import Image or Folder": "Bild oder Ordner importieren",
+        "Select Image": "Bild auswählen",
+        "Select Folder": "Ordner auswählen",
+        "Select gallery images or folders": "Galeriebilder oder Ordner auswählen",
+        "Select one or more image files or folders. Folders are scanned recursively.": "Wählen Sie eine oder mehrere Bilddateien oder Ordner aus. Ordner werden rekursiv gescannt.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "Klicken Sie, um ein Bild oder einen Ordner zu importieren, oder ziehen Sie ihn hierher. Ordner werden rekursiv importiert.",
+        "Click to upload images or a folder, or drag them here": "Klicken Sie zum Hochladen von Bildern oder eines Ordners, oder ziehen Sie sie hierher",
+        "Drag one or more album folders here, or use Add Folder.": "Ziehen Sie einen oder mehrere Albumordner hierher oder verwenden Sie Ordner hinzufügen.",
+        "Drag album folders into the directory list, or click Add Folder.": "Ziehen Sie Albumordner in die Verzeichnisliste oder klicken Sie auf Ordner hinzufügen.",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "Der Kosinusschwellenwert ist die Grenze für Gesichtähnlichkeit: höhere Werte machen Cluster strenger. DBSCAN verwendet intern Kosinusdistanz = 1 - Kosinusschwellenwert.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "Kosinussimilaritätsschwelle. Höhere Werte machen Cluster strenger; Standard ist 0,48.",
+        "DBSCAN distance threshold": "DBSCAN-Distanzschwelle",
+        "Algorithm: DBSCAN": "Algorithmus: DBSCAN",
+        "Scan and Cluster": "Scannen und clustern",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "Mindestanzahl benachbarter Gesichter für einen DBSCAN-Kernpunkt.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "Gesichter, deren Boxbreite oder -höhe kleiner als dieser Wert ist, werden übersprungen.",
+        "No images": "Keine Bilder",
+        "People": "Personen",
+        "ID": "ID",
+        "Name": "Name",
+        "Name, notes, and tags are shown here.": "Name, Notizen und Tags werden hier angezeigt.",
+        "Threshold": "Schwelle",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "Diese Anwendung läuft lokal. Code- und Modelllizenzen können sich unterscheiden. Forschungsmodelle können auf nichtkommerzielle Nutzung beschränkt sein. Kommerzielle Bereitstellung erfordert eine geeignete Modelllizenz. Dieses Tool bietet keine Rechtsberatung.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "CUDA ist auf diesem Rechner nicht verfügbar; Auto verwendet daher die CPU.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Auto verwendet CUDA, wenn CUDAExecutionProvider verfügbar ist, andernfalls CPU.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider ist verfügbar.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider ist nicht verfügbar. Installieren Sie zuerst passendes onnxruntime-gpu, CUDA Runtime und GPU-Treiber.",
+        "No downloaded swap models": "Keine heruntergeladenen Face-Swap-Modelle",
+        "Only downloaded swap models are shown.": "Es werden nur heruntergeladene Face-Swap-Modelle angezeigt.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "Laden Sie zuerst inswapper_128.onnx unter Models > Downloads herunter.",
+        "No downloaded GFPGAN models": "Keine heruntergeladenen GFPGAN-Modelle",
+        "Only downloaded GFPGAN restore models are shown.": "Es werden nur heruntergeladene GFPGAN-Wiederherstellungsmodelle angezeigt.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "Laden Sie zuerst GFPGANv1.4.onnx unter Models > Downloads herunter.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN ist erst nach dem Herunterladen eines GFPGAN-Modells verfügbar.",
+        "Run GFPGAN face restoration after face swap.": "GFPGAN-Gesichtswiederherstellung nach dem Face Swap ausführen.",
+        "Enable GFPGAN restore after face swap": "GFPGAN-Wiederherstellung nach Face Swap aktivieren",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "Wählt das Galeriebild jeder Identität automatisch aus einer Datei mit „gallery“ im Namen oder dem ersten sortierten Bild.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "Öffnet die detaillierten Identitätsordnerregeln für 1:1, 1:N und automatische Aufteilung.",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "Öffnet die InsightFace-Kontaktseite für Unternehmensevaluierung, Lizenzierung, SDK/API, SLA oder individuelles Training.",
+        "Evaluation Dataset Rules": "Regeln für Evaluierungsdatensätze",
+        "Prepare local identity folders using one of the supported layouts below.": "Bereiten Sie lokale Identitätsordner mit einer der unten unterstützten Strukturen vor.",
+        "Validate the dataset before running an evaluation.": "Validieren Sie den Datensatz, bevor Sie eine Evaluierung starten.",
+        "Validate the selected dataset and configuration before running evaluation.": "Validieren Sie den ausgewählten Datensatz und die Konfiguration vor der Evaluierung.",
+        "Result preview appears here. Video results are saved to the exports folder.": "Die Ergebnisvorschau erscheint hier. Videoergebnisse werden im Exportordner gespeichert.",
+        "Min det score %": "Min. Detektionsscore %",
+        "asset": "Asset",
+        "source": "Quelle",
+        "kind": "Art",
+        "release": "Release",
+        "size": "Größe",
+        "updated_at": "aktualisiert",
+        "local status": "lokaler Status",
+        "download url": "Download-URL",
+        "rank": "Rang",
+        "thumbnail": "Miniatur",
+        "gallery_file": "Galeriedatei",
+        "similarity": "Ähnlichkeit",
+        "threshold": "Schwelle",
+        "decision": "Entscheidung",
+        "det_score": "Detektionsscore",
+        "person_id": "Personen-ID",
+        "person_name": "Personenname",
+        "status": "Status",
+        "sample_id": "Sample-ID",
+        "crop_path": "Ausschnittspfad",
+        "image_path": "Bildpfad",
+        "face_index": "Gesichtsindex",
+        "bbox": "Box",
+        "video_path": "Videopfad",
+        "timestamp_ms": "Zeitstempel ms",
+        "timestamp_hhmmss": "Zeit",
+        "frame_index": "Frameindex",
+        "id": "ID",
+        "date": "Datum",
+        "scenario": "Szenario",
+        "model": "Modell",
+        "provider": "Provider",
+        "report_path": "Berichtspfad",
+        "quality_score": "Qualitätsscore",
+        "cluster_id": "Cluster-ID",
+        "face_count": "Gesichter",
+        "photo_count": "Fotos",
+        "average_quality": "Durchschnittsqualität",
+        "sample_paths": "Sample-Pfade",
+        "model_name": "Modellname",
+        "source_image_path": "Quellbildpfad",
+    },
+    "pt": {
+        "<a href='#'>Browse</a>": "<a href='#'>Procurar</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "diretório de modelo personalizado",
+        "Custom model directory": "Diretório de modelo personalizado",
+        "Auto": "Automático",
+        "Browse": "Procurar",
+        "Add Person": "Adicionar pessoa",
+        "Add as New Person": "Adicionar como nova pessoa",
+        "Assign to Existing Person": "Atribuir a pessoa existente",
+        "Compare": "Comparar",
+        "Search": "Pesquisar",
+        "Export": "Exportar",
+        "Export Result": "Exportar resultado",
+        "Export Results": "Exportar resultados",
+        "Export Summary": "Exportar resumo",
+        "Export Settings": "Exportar definições",
+        "Import Settings": "Importar definições",
+        "Delete": "Eliminar",
+        "Delete Selected": "Eliminar seleção",
+        "Reset": "Repor",
+        "Reset Settings": "Repor definições",
+        "Refresh": "Atualizar",
+        "Start": "Iniciar",
+        "Stop": "Parar",
+        "Open": "Abrir",
+        "Save Settings": "Guardar definições",
+        "Save Details": "Guardar detalhes",
+        "Save crops": "Guardar recortes",
+        "Save face crops": "Guardar recortes de rosto",
+        "Save recognition logs": "Guardar registos de reconhecimento",
+        "Anonymize report paths": "Anonimizar caminhos dos relatórios",
+        "Privacy": "Privacidade",
+        "Recursive": "Recursivo",
+        "Identify people": "Identificar pessoas",
+        "Camera index": "Índice da câmara",
+        "Frame skip": "Saltar fotogramas",
+        "Frame interval": "Intervalo de fotogramas",
+        "Target person": "Pessoa alvo",
+        "Check Cameras": "Verificar câmaras",
+        "Register Current Face": "Registar rosto atual",
+        "Refresh People": "Atualizar pessoas",
+        "Recognize Faces": "Reconhecer rostos",
+        "Test Model Load": "Testar carregamento do modelo",
+        "Warmup": "Aquecimento",
+        "Export Annotated Image": "Exportar imagem anotada",
+        "Export CSV": "Exportar CSV",
+        "Export CSV / JSON": "Exportar CSV / JSON",
+        "Refresh Download URLs": "Atualizar URLs de transferência",
+        "Download Selected": "Transferir seleção",
+        "Use Selected Model": "Usar modelo selecionado",
+        "Open Model Folder": "Abrir pasta de modelos",
+        "Open GitHub Releases": "Abrir GitHub Releases",
+        "Open Model Downloads": "Abrir transferências de modelos",
+        "Open License Center": "Abrir centro de licenças",
+        "Open Reports": "Abrir relatórios",
+        "Open Selected Report": "Abrir relatório selecionado",
+        "Re-export Selected": "Reexportar seleção",
+        "Export All Summary": "Exportar resumo completo",
+        "Import Image or Folder": "Importar imagem ou pasta",
+        "Select Image": "Selecionar imagem",
+        "Select Folder": "Selecionar pasta",
+        "Select gallery images or folders": "Selecionar imagens ou pastas da galeria",
+        "Select one or more image files or folders. Folders are scanned recursively.": "Selecione um ou mais ficheiros de imagem ou pastas. As pastas são analisadas recursivamente.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "Clique para importar uma imagem ou pasta, ou arraste-a para aqui. As pastas são importadas recursivamente.",
+        "Click to upload images or a folder, or drag them here": "Clique para carregar imagens ou uma pasta, ou arraste-as para aqui",
+        "Drag one or more album folders here, or use Add Folder.": "Arraste uma ou mais pastas de álbum para aqui, ou use Adicionar pasta.",
+        "Drag album folders into the directory list, or click Add Folder.": "Arraste pastas de álbum para a lista de diretórios, ou clique em Adicionar pasta.",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "O limiar de cosseno é o corte de similaridade facial: valores mais altos tornam os clusters mais rigorosos. O DBSCAN usa internamente distância de cosseno = 1 - limiar de cosseno.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "Limiar de similaridade de cosseno. Valores mais altos tornam os clusters mais rigorosos; predefinição 0,48.",
+        "DBSCAN distance threshold": "Limiar de distância DBSCAN",
+        "Algorithm: DBSCAN": "Algoritmo: DBSCAN",
+        "Scan and Cluster": "Analisar e agrupar",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "Número mínimo de rostos vizinhos necessário para um ponto central do DBSCAN.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "Rostos cuja caixa delimitadora tenha largura ou altura inferior a este valor são ignorados.",
+        "No images": "Sem imagens",
+        "People": "Pessoas",
+        "ID": "ID",
+        "Name": "Nome",
+        "Name, notes, and tags are shown here.": "O nome, notas e etiquetas aparecem aqui.",
+        "Threshold": "Limiar",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "Esta aplicação é executada localmente. As licenças do código e dos modelos podem diferir. Modelos de investigação podem estar limitados a uso não comercial. A implantação comercial requer uma licença de modelo adequada. Esta ferramenta não fornece aconselhamento jurídico.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "CUDA não está disponível nesta máquina, por isso Automático usará CPU.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Automático usa CUDA quando CUDAExecutionProvider está disponível; caso contrário usa CPU.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider está disponível.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider não está disponível. Instale primeiro onnxruntime-gpu compatível, runtime CUDA e driver GPU.",
+        "No downloaded swap models": "Não há modelos de troca de rostos transferidos",
+        "Only downloaded swap models are shown.": "Só são mostrados modelos de troca de rostos transferidos.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "Transfira primeiro inswapper_128.onnx em Models > Downloads.",
+        "No downloaded GFPGAN models": "Não há modelos GFPGAN transferidos",
+        "Only downloaded GFPGAN restore models are shown.": "Só são mostrados modelos GFPGAN de restauro transferidos.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "Transfira primeiro GFPGANv1.4.onnx em Models > Downloads.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN fica indisponível até transferir um modelo GFPGAN.",
+        "Run GFPGAN face restoration after face swap.": "Executar restauro facial GFPGAN após a troca de rostos.",
+        "Enable GFPGAN restore after face swap": "Ativar restauro GFPGAN após troca de rostos",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "Seleciona automaticamente a imagem de galeria de cada identidade a partir de um ficheiro que contenha “gallery”, ou da primeira imagem ordenada.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "Abre as regras detalhadas de pastas de identidade para 1:1, 1:N e Divisão automática.",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "Abrir a página de contacto da InsightFace para avaliação empresarial, licenciamento, SDK/API, SLA ou treino personalizado.",
+        "Evaluation Dataset Rules": "Regras do dataset de avaliação",
+        "Prepare local identity folders using one of the supported layouts below.": "Prepare pastas locais de identidade usando uma das estruturas suportadas abaixo.",
+        "Validate the dataset before running an evaluation.": "Valide o dataset antes de executar uma avaliação.",
+        "Validate the selected dataset and configuration before running evaluation.": "Valide o dataset e a configuração selecionados antes de executar a avaliação.",
+        "Result preview appears here. Video results are saved to the exports folder.": "A pré-visualização do resultado aparece aqui. Os resultados de vídeo são guardados na pasta de exportações.",
+        "Min det score %": "Pontuação mínima de deteção %",
+        "asset": "recurso",
+        "source": "origem",
+        "kind": "tipo",
+        "release": "versão",
+        "size": "tamanho",
+        "updated_at": "atualizado",
+        "local status": "estado local",
+        "download url": "URL de transferência",
+        "rank": "classificação",
+        "thumbnail": "miniatura",
+        "gallery_file": "ficheiro da galeria",
+        "similarity": "similaridade",
+        "threshold": "limiar",
+        "decision": "decisão",
+        "det_score": "pontuação de deteção",
+        "person_id": "ID da pessoa",
+        "person_name": "nome da pessoa",
+        "status": "estado",
+        "sample_id": "ID da amostra",
+        "crop_path": "caminho do recorte",
+        "image_path": "caminho da imagem",
+        "face_index": "índice do rosto",
+        "bbox": "caixa",
+        "video_path": "caminho do vídeo",
+        "timestamp_ms": "timestamp ms",
+        "timestamp_hhmmss": "hora",
+        "frame_index": "índice do fotograma",
+        "id": "ID",
+        "date": "data",
+        "scenario": "cenário",
+        "model": "modelo",
+        "provider": "fornecedor",
+        "report_path": "caminho do relatório",
+        "quality_score": "pontuação de qualidade",
+        "cluster_id": "ID do cluster",
+        "face_count": "rostos",
+        "photo_count": "fotos",
+        "average_quality": "qualidade média",
+        "sample_paths": "caminhos das amostras",
+        "model_name": "nome do modelo",
+        "source_image_path": "caminho da imagem de origem",
+    },
+    "ru": {
+        "<a href='#'>Browse</a>": "<a href='#'>Обзор</a>",
+        "InsightFace Evaluation Studio": "InsightFace Evaluation Studio",
+        "custom model directory": "пользовательский каталог модели",
+        "Custom model directory": "Пользовательский каталог модели",
+        "Auto": "Авто",
+        "Browse": "Обзор",
+        "Add Person": "Добавить человека",
+        "Add as New Person": "Добавить как нового человека",
+        "Assign to Existing Person": "Назначить существующему человеку",
+        "Compare": "Сравнить",
+        "Search": "Поиск",
+        "Export": "Экспорт",
+        "Export Result": "Экспорт результата",
+        "Export Results": "Экспорт результатов",
+        "Export Summary": "Экспорт сводки",
+        "Export Settings": "Экспорт настроек",
+        "Import Settings": "Импорт настроек",
+        "Delete": "Удалить",
+        "Delete Selected": "Удалить выбранное",
+        "Reset": "Сброс",
+        "Reset Settings": "Сбросить настройки",
+        "Refresh": "Обновить",
+        "Start": "Старт",
+        "Stop": "Стоп",
+        "Open": "Открыть",
+        "Save Settings": "Сохранить настройки",
+        "Save Details": "Сохранить детали",
+        "Save crops": "Сохранять кадры",
+        "Save face crops": "Сохранять кадры лиц",
+        "Save recognition logs": "Сохранять журналы распознавания",
+        "Anonymize report paths": "Анонимизировать пути в отчётах",
+        "Privacy": "Конфиденциальность",
+        "Recursive": "Рекурсивно",
+        "Identify people": "Идентифицировать людей",
+        "Camera index": "Индекс камеры",
+        "Frame skip": "Пропуск кадров",
+        "Frame interval": "Интервал кадров",
+        "Target person": "Целевая персона",
+        "Check Cameras": "Проверить камеры",
+        "Register Current Face": "Зарегистрировать текущее лицо",
+        "Refresh People": "Обновить людей",
+        "Recognize Faces": "Распознать лица",
+        "Test Model Load": "Проверить загрузку модели",
+        "Warmup": "Прогрев",
+        "Export Annotated Image": "Экспортировать размеченное изображение",
+        "Export CSV": "Экспорт CSV",
+        "Export CSV / JSON": "Экспорт CSV / JSON",
+        "Refresh Download URLs": "Обновить URL загрузки",
+        "Download Selected": "Загрузить выбранное",
+        "Use Selected Model": "Использовать выбранную модель",
+        "Open Model Folder": "Открыть папку моделей",
+        "Open GitHub Releases": "Открыть GitHub Releases",
+        "Open Model Downloads": "Открыть загрузки моделей",
+        "Open License Center": "Открыть центр лицензий",
+        "Open Reports": "Открыть отчёты",
+        "Open Selected Report": "Открыть выбранный отчёт",
+        "Re-export Selected": "Переэкспортировать выбранное",
+        "Export All Summary": "Экспорт общей сводки",
+        "Import Image or Folder": "Импортировать изображение или папку",
+        "Select Image": "Выбрать изображение",
+        "Select Folder": "Выбрать папку",
+        "Select gallery images or folders": "Выбрать изображения или папки галереи",
+        "Select one or more image files or folders. Folders are scanned recursively.": "Выберите один или несколько файлов изображений или папок. Папки сканируются рекурсивно.",
+        "Click to import an image or folder, or drag one here. Folders are imported recursively.": "Нажмите для импорта изображения или папки либо перетащите сюда. Папки импортируются рекурсивно.",
+        "Click to upload images or a folder, or drag them here": "Нажмите для загрузки изображений или папки либо перетащите их сюда",
+        "Drag one or more album folders here, or use Add Folder.": "Перетащите сюда одну или несколько папок альбома либо используйте «Добавить папку».",
+        "Drag album folders into the directory list, or click Add Folder.": "Перетащите папки альбома в список каталогов или нажмите «Добавить папку».",
+        "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. DBSCAN internally uses cosine distance = 1 - cosine threshold.": "Косинусный порог задаёт отсечку сходства лиц: чем выше значение, тем строже кластеры. DBSCAN внутренне использует косинусную дистанцию = 1 - косинусный порог.",
+        "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48.": "Порог косинусного сходства. Более высокие значения делают кластеры строже; значение по умолчанию 0,48.",
+        "DBSCAN distance threshold": "Порог дистанции DBSCAN",
+        "Algorithm: DBSCAN": "Алгоритм: DBSCAN",
+        "Scan and Cluster": "Сканировать и кластеризовать",
+        "Minimum neighboring faces needed for a DBSCAN core point.": "Минимальное число соседних лиц для ядровой точки DBSCAN.",
+        "Faces whose bounding box width or height is smaller than this value are skipped.": "Лица, ширина или высота рамки которых меньше этого значения, пропускаются.",
+        "No images": "Нет изображений",
+        "People": "Люди",
+        "ID": "ID",
+        "Name": "Имя",
+        "Name, notes, and tags are shown here.": "Здесь отображаются имя, заметки и теги.",
+        "Threshold": "Порог",
+        "This application runs locally. Code and model licenses may differ. Research models may be limited to non-commercial use. Commercial deployment requires appropriate model license. This tool does not provide legal advice.": "Это приложение работает локально. Лицензии кода и моделей могут различаться. Исследовательские модели могут быть ограничены некоммерческим использованием. Коммерческое развертывание требует соответствующей лицензии на модель. Этот инструмент не предоставляет юридических консультаций.",
+        "CUDA is unavailable on this machine, so Auto will use CPU.": "CUDA недоступна на этом компьютере, поэтому Авто будет использовать CPU.",
+        "Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.": "Авто использует CUDA, когда доступен CUDAExecutionProvider, иначе CPU.",
+        "CUDAExecutionProvider is available.": "CUDAExecutionProvider доступен.",
+        "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first.": "CUDAExecutionProvider недоступен. Сначала установите совместимые onnxruntime-gpu, CUDA runtime и драйвер GPU.",
+        "No downloaded swap models": "Нет загруженных моделей замены лиц",
+        "Only downloaded swap models are shown.": "Показаны только загруженные модели замены лиц.",
+        "Download inswapper_128.onnx from Models > Downloads first.": "Сначала загрузите inswapper_128.onnx в Models > Downloads.",
+        "No downloaded GFPGAN models": "Нет загруженных моделей GFPGAN",
+        "Only downloaded GFPGAN restore models are shown.": "Показаны только загруженные модели восстановления GFPGAN.",
+        "Download GFPGANv1.4.onnx from Models > Downloads first.": "Сначала загрузите GFPGANv1.4.onnx в Models > Downloads.",
+        "GFPGAN is unavailable until a GFPGAN model is downloaded.": "GFPGAN недоступен, пока модель GFPGAN не загружена.",
+        "Run GFPGAN face restoration after face swap.": "Запускать восстановление лиц GFPGAN после замены лиц.",
+        "Enable GFPGAN restore after face swap": "Включить восстановление GFPGAN после замены лиц",
+        "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image.": "Автоматически выбирает галерейное изображение каждой идентичности из файла с «gallery» в имени или первое изображение по сортировке.",
+        "Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.": "Открыть подробные правила папок идентичностей для 1:1, 1:N и авторазделения.",
+        "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help.": "Открыть страницу контактов InsightFace для корпоративной оценки, лицензирования, SDK/API, SLA или индивидуального обучения.",
+        "Evaluation Dataset Rules": "Правила датасета оценки",
+        "Prepare local identity folders using one of the supported layouts below.": "Подготовьте локальные папки идентичностей в одной из поддерживаемых структур ниже.",
+        "Validate the dataset before running an evaluation.": "Проверьте датасет перед запуском оценки.",
+        "Validate the selected dataset and configuration before running evaluation.": "Проверьте выбранный датасет и конфигурацию перед запуском оценки.",
+        "Result preview appears here. Video results are saved to the exports folder.": "Предпросмотр результата появится здесь. Видео сохраняются в папку экспорта.",
+        "Min det score %": "Мин. score детекции %",
+        "asset": "ресурс",
+        "source": "источник",
+        "kind": "тип",
+        "release": "релиз",
+        "size": "размер",
+        "updated_at": "обновлено",
+        "local status": "локальный статус",
+        "download url": "URL загрузки",
+        "rank": "ранг",
+        "thumbnail": "миниатюра",
+        "gallery_file": "файл галереи",
+        "similarity": "сходство",
+        "threshold": "порог",
+        "decision": "решение",
+        "det_score": "score детекции",
+        "person_id": "ID человека",
+        "person_name": "имя человека",
+        "status": "статус",
+        "sample_id": "ID образца",
+        "crop_path": "путь кадра",
+        "image_path": "путь изображения",
+        "face_index": "индекс лица",
+        "bbox": "рамка",
+        "video_path": "путь видео",
+        "timestamp_ms": "временная метка ms",
+        "timestamp_hhmmss": "время",
+        "frame_index": "индекс кадра",
+        "id": "ID",
+        "date": "дата",
+        "scenario": "сценарий",
+        "model": "модель",
+        "provider": "провайдер",
+        "report_path": "путь отчёта",
+        "quality_score": "score качества",
+        "cluster_id": "ID кластера",
+        "face_count": "лиц",
+        "photo_count": "фото",
+        "average_quality": "среднее качество",
+        "sample_paths": "пути образцов",
+        "model_name": "имя модели",
+        "source_image_path": "путь исходного изображения",
+    },
+}
+
+for _lang, _values in _ADDITIONAL_UI_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
+
+
+_DYNAMIC_UI_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "Click to upload or drag a file here": "点击上传，或将文件拖到这里",
+        "Drop another file to replace it": "拖入另一个文件可替换当前文件",
+        "Click or drag more images or folders to add to Gallery.": "点击或继续拖入图片/文件夹以添加到 Gallery。",
+        "Image swap saved. Click Result to open it.": "图片换脸结果已保存。点击结果区域可打开文件。",
+        "Video swap saved. Click Result to open it.": "视频换脸结果已保存。点击结果区域可打开文件。",
+        "image": "image",
+        "images": "images",
+    },
+    "ja": {
+        "Click to upload or drag a file here": "クリックしてアップロードするか、ファイルをここにドラッグしてください",
+        "Drop another file to replace it": "別のファイルをドロップすると置き換えられます",
+        "Click or drag more images or folders to add to Gallery.": "クリックするか画像/フォルダをさらにドラッグしてギャラリーに追加します。",
+        "Image swap saved. Click Result to open it.": "画像の顔交換結果を保存しました。結果エリアをクリックして開けます。",
+        "Video swap saved. Click Result to open it.": "動画の顔交換結果を保存しました。結果エリアをクリックして開けます。",
+        "image": "画像",
+        "images": "画像",
+    },
+    "ko": {
+        "Click to upload or drag a file here": "클릭하여 업로드하거나 파일을 여기로 끌어오십시오",
+        "Drop another file to replace it": "다른 파일을 놓으면 현재 파일을 교체합니다",
+        "Click or drag more images or folders to add to Gallery.": "클릭하거나 이미지/폴더를 더 끌어와 Gallery에 추가하십시오.",
+        "Image swap saved. Click Result to open it.": "이미지 얼굴 교체 결과가 저장되었습니다. 결과 영역을 클릭하면 열 수 있습니다.",
+        "Video swap saved. Click Result to open it.": "비디오 얼굴 교체 결과가 저장되었습니다. 결과 영역을 클릭하면 열 수 있습니다.",
+        "image": "이미지",
+        "images": "이미지",
+    },
+    "es": {
+        "Click to upload or drag a file here": "Haga clic para subir un archivo o arrástrelo aquí",
+        "Drop another file to replace it": "Suelte otro archivo para reemplazarlo",
+        "Click or drag more images or folders to add to Gallery.": "Haga clic o arrastre más imágenes o carpetas para añadirlas a la galería.",
+        "Image swap saved. Click Result to open it.": "Intercambio de imagen guardado. Haga clic en el área de resultado para abrirlo.",
+        "Video swap saved. Click Result to open it.": "Intercambio de vídeo guardado. Haga clic en el área de resultado para abrirlo.",
+        "image": "imagen",
+        "images": "imágenes",
+    },
+    "fr": {
+        "Click to upload or drag a file here": "Cliquez pour charger un fichier ou faites-le glisser ici",
+        "Drop another file to replace it": "Déposez un autre fichier pour le remplacer",
+        "Click or drag more images or folders to add to Gallery.": "Cliquez ou faites glisser d’autres images ou dossiers pour les ajouter à la galerie.",
+        "Image swap saved. Click Result to open it.": "Face swap d’image enregistré. Cliquez sur la zone de résultat pour l’ouvrir.",
+        "Video swap saved. Click Result to open it.": "Face swap vidéo enregistré. Cliquez sur la zone de résultat pour l’ouvrir.",
+        "image": "image",
+        "images": "images",
+    },
+    "de": {
+        "Click to upload or drag a file here": "Klicken Sie zum Hochladen oder ziehen Sie eine Datei hierher",
+        "Drop another file to replace it": "Legen Sie eine andere Datei ab, um sie zu ersetzen",
+        "Click or drag more images or folders to add to Gallery.": "Klicken oder ziehen Sie weitere Bilder oder Ordner hierher, um sie zur Galerie hinzuzufügen.",
+        "Image swap saved. Click Result to open it.": "Bild-Face-Swap gespeichert. Klicken Sie auf den Ergebnisbereich, um ihn zu öffnen.",
+        "Video swap saved. Click Result to open it.": "Video-Face-Swap gespeichert. Klicken Sie auf den Ergebnisbereich, um ihn zu öffnen.",
+        "image": "Bild",
+        "images": "Bilder",
+    },
+    "pt": {
+        "Click to upload or drag a file here": "Clique para carregar um ficheiro ou arraste-o para aqui",
+        "Drop another file to replace it": "Largue outro ficheiro para substituir",
+        "Click or drag more images or folders to add to Gallery.": "Clique ou arraste mais imagens ou pastas para adicionar à galeria.",
+        "Image swap saved. Click Result to open it.": "Troca de rosto em imagem guardada. Clique na área de resultado para abrir.",
+        "Video swap saved. Click Result to open it.": "Troca de rosto em vídeo guardada. Clique na área de resultado para abrir.",
+        "image": "imagem",
+        "images": "imagens",
+    },
+    "ru": {
+        "Click to upload or drag a file here": "Нажмите для загрузки файла или перетащите его сюда",
+        "Drop another file to replace it": "Перетащите другой файл, чтобы заменить текущий",
+        "Click or drag more images or folders to add to Gallery.": "Нажмите или перетащите дополнительные изображения/папки, чтобы добавить их в галерею.",
+        "Image swap saved. Click Result to open it.": "Результат замены лиц на изображении сохранён. Нажмите область результата, чтобы открыть его.",
+        "Video swap saved. Click Result to open it.": "Результат замены лиц в видео сохранён. Нажмите область результата, чтобы открыть его.",
+        "image": "изображение",
+        "images": "изображений",
+    },
+}
+
+for _lang, _values in _DYNAMIC_UI_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
+
+
+_STATUS_ERROR_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "Select album folder": "选择相册文件夹",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "当前没有选择相册目录。全部重建将清空已保存的相册聚类结果。是否继续？",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "全部重建会重新处理所选相册目录中的每张图片，并替换已保存的聚类结果。是否继续？",
+        "Model is not loaded. Please open Models.": "模型尚未加载。请打开 Models。",
+        "Image read failure.": "图片读取失败。",
+        "Please upload a query image.": "请上传查询图片。",
+        "Please upload one or more gallery images or folders.": "请上传一个或多个 Gallery 图片或文件夹。",
+        "Please select identity folders or a dataset root.": "请选择身份文件夹或数据集根目录。",
+        "Please validate this dataset and configuration before running evaluation.": "请先核验该数据集和配置，再运行评测。",
+        "Run an evaluation first.": "请先运行一次评测。",
+        "PDF export is unavailable. Please install reportlab.": "PDF 导出不可用。请安装 reportlab。",
+        "Select a source image first.": "请先选择源图。",
+        "Select a target image or video first.": "请先选择目标图片或视频。",
+        "Face swap model not found. Please download and choose a swap model in Models.": "未找到换脸模型。请在 Models 中下载并选择换脸模型。",
+        "Unsupported target file type.": "不支持的目标文件类型。",
+        "No saved result file to open.": "没有可打开的已保存结果文件。",
+        "Add one or more album directories first.": "请先添加一个或多个相册目录。",
+        "No supported images found in the selected directories.": "所选目录中没有找到支持的图片。",
+        "Select a model asset first.": "请先选择一个模型资源。",
+        "Selected asset cannot be used as a model package.": "所选资源不能作为模型包使用。",
+        "Model settings saved.": "模型设置已保存。",
+        "Model loaded successfully.": "模型加载成功。",
+        "Model loaded.": "模型已加载。",
+        "Cleared.": "已清空。",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "多人脸处理策略已更改，Gallery 特征将重新计算。",
+        "Dataset validation passed.": "数据集核验通过。",
+        "Dataset validation failed.": "数据集核验失败。",
+        "Album directories cleared. Existing clustering results are still available.": "相册目录已清空。已有聚类结果仍可查看。",
+        "Saved album clustering results were cleared.": "已保存的相册聚类结果已清空。",
+        "Opened InsightFace homepage.": "已打开 InsightFace 主页。",
+        "Opened InsightFace enterprise contact page.": "已打开 InsightFace 企业联系页面。",
+    },
+    "ja": {
+        "Select album folder": "アルバムフォルダを選択",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "アルバムディレクトリが選択されていません。すべて再構築すると保存済みのクラスタリング結果が消去されます。続行しますか？",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "すべて再構築すると、選択したアルバムディレクトリ内の全画像を再処理し、保存済みのクラスタリング結果を置き換えます。続行しますか？",
+        "Model is not loaded. Please open Models.": "モデルが読み込まれていません。Models を開いてください。",
+        "Image read failure.": "画像の読み込みに失敗しました。",
+        "Please upload a query image.": "クエリ画像をアップロードしてください。",
+        "Please upload one or more gallery images or folders.": "1つ以上のギャラリー画像またはフォルダをアップロードしてください。",
+        "Please select identity folders or a dataset root.": "IDフォルダまたはデータセットルートを選択してください。",
+        "Please validate this dataset and configuration before running evaluation.": "評価を実行する前に、このデータセットと設定を検証してください。",
+        "Run an evaluation first.": "先に評価を実行してください。",
+        "PDF export is unavailable. Please install reportlab.": "PDF 出力は利用できません。reportlab をインストールしてください。",
+        "Select a source image first.": "先にソース画像を選択してください。",
+        "Select a target image or video first.": "先にターゲット画像または動画を選択してください。",
+        "Face swap model not found. Please download and choose a swap model in Models.": "顔交換モデルが見つかりません。Models でモデルをダウンロードして選択してください。",
+        "Unsupported target file type.": "サポートされていないターゲットファイル形式です。",
+        "No saved result file to open.": "開ける保存済み結果ファイルがありません。",
+        "Add one or more album directories first.": "先に1つ以上のアルバムディレクトリを追加してください。",
+        "No supported images found in the selected directories.": "選択したディレクトリに対応画像が見つかりません。",
+        "Select a model asset first.": "先にモデルアセットを選択してください。",
+        "Selected asset cannot be used as a model package.": "選択したアセットはモデルパッケージとして使用できません。",
+        "Model settings saved.": "モデル設定を保存しました。",
+        "Model loaded successfully.": "モデルを正常に読み込みました。",
+        "Model loaded.": "モデルを読み込みました。",
+        "Cleared.": "クリアしました。",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "複数顔の処理方法が変更されました。ギャラリー特徴量を再計算します。",
+        "Dataset validation passed.": "データセット検証に合格しました。",
+        "Dataset validation failed.": "データセット検証に失敗しました。",
+        "Album directories cleared. Existing clustering results are still available.": "アルバムディレクトリをクリアしました。既存のクラスタリング結果は引き続き利用できます。",
+        "Saved album clustering results were cleared.": "保存済みのアルバムクラスタリング結果を消去しました。",
+        "Opened InsightFace homepage.": "InsightFace ホームページを開きました。",
+        "Opened InsightFace enterprise contact page.": "InsightFace 企業向けお問い合わせページを開きました。",
+    },
+    "ko": {
+        "Select album folder": "앨범 폴더 선택",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "선택된 앨범 디렉터리가 없습니다. 전체 재빌드는 저장된 앨범 클러스터링 결과를 지웁니다. 계속하시겠습니까?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "전체 재빌드는 선택한 앨범 디렉터리의 모든 이미지를 다시 처리하고 저장된 클러스터링 결과를 교체합니다. 계속하시겠습니까?",
+        "Model is not loaded. Please open Models.": "모델이 로드되지 않았습니다. Models를 여십시오.",
+        "Image read failure.": "이미지를 읽지 못했습니다.",
+        "Please upload a query image.": "쿼리 이미지를 업로드하십시오.",
+        "Please upload one or more gallery images or folders.": "하나 이상의 갤러리 이미지 또는 폴더를 업로드하십시오.",
+        "Please select identity folders or a dataset root.": "ID 폴더 또는 데이터셋 루트를 선택하십시오.",
+        "Please validate this dataset and configuration before running evaluation.": "평가 실행 전에 이 데이터셋과 구성을 검증하십시오.",
+        "Run an evaluation first.": "먼저 평가를 실행하십시오.",
+        "PDF export is unavailable. Please install reportlab.": "PDF 내보내기를 사용할 수 없습니다. reportlab을 설치하십시오.",
+        "Select a source image first.": "먼저 소스 이미지를 선택하십시오.",
+        "Select a target image or video first.": "먼저 대상 이미지 또는 비디오를 선택하십시오.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "얼굴 교체 모델을 찾을 수 없습니다. Models에서 모델을 다운로드하고 선택하십시오.",
+        "Unsupported target file type.": "지원되지 않는 대상 파일 형식입니다.",
+        "No saved result file to open.": "열 수 있는 저장된 결과 파일이 없습니다.",
+        "Add one or more album directories first.": "먼저 하나 이상의 앨범 디렉터리를 추가하십시오.",
+        "No supported images found in the selected directories.": "선택한 디렉터리에서 지원되는 이미지를 찾지 못했습니다.",
+        "Select a model asset first.": "먼저 모델 자산을 선택하십시오.",
+        "Selected asset cannot be used as a model package.": "선택한 자산은 모델 패키지로 사용할 수 없습니다.",
+        "Model settings saved.": "모델 설정이 저장되었습니다.",
+        "Model loaded successfully.": "모델을 성공적으로 로드했습니다.",
+        "Model loaded.": "모델이 로드되었습니다.",
+        "Cleared.": "지웠습니다.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "다중 얼굴 처리 방식이 변경되었습니다. Gallery 임베딩을 다시 계산합니다.",
+        "Dataset validation passed.": "데이터셋 검증을 통과했습니다.",
+        "Dataset validation failed.": "데이터셋 검증에 실패했습니다.",
+        "Album directories cleared. Existing clustering results are still available.": "앨범 디렉터리가 지워졌습니다. 기존 클러스터링 결과는 계속 사용할 수 있습니다.",
+        "Saved album clustering results were cleared.": "저장된 앨범 클러스터링 결과가 지워졌습니다.",
+        "Opened InsightFace homepage.": "InsightFace 홈페이지를 열었습니다.",
+        "Opened InsightFace enterprise contact page.": "InsightFace 기업 문의 페이지를 열었습니다.",
+    },
+    "es": {
+        "Select album folder": "Seleccionar carpeta de álbum",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "No hay directorios de álbum seleccionados. Rebuild All borrará los resultados de agrupación guardados. ¿Continuar?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "Rebuild All reprocesará todas las imágenes de los directorios seleccionados y reemplazará los resultados guardados. ¿Continuar?",
+        "Model is not loaded. Please open Models.": "El modelo no está cargado. Abra Models.",
+        "Image read failure.": "Error al leer la imagen.",
+        "Please upload a query image.": "Suba una imagen de consulta.",
+        "Please upload one or more gallery images or folders.": "Suba una o más imágenes o carpetas de galería.",
+        "Please select identity folders or a dataset root.": "Seleccione carpetas de identidad o la raíz de un dataset.",
+        "Please validate this dataset and configuration before running evaluation.": "Valide este dataset y configuración antes de ejecutar la evaluación.",
+        "Run an evaluation first.": "Ejecute primero una evaluación.",
+        "PDF export is unavailable. Please install reportlab.": "La exportación PDF no está disponible. Instale reportlab.",
+        "Select a source image first.": "Seleccione primero una imagen fuente.",
+        "Select a target image or video first.": "Seleccione primero una imagen o vídeo de destino.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "No se encontró el modelo de intercambio facial. Descárguelo y selecciónelo en Models.",
+        "Unsupported target file type.": "Tipo de archivo de destino no compatible.",
+        "No saved result file to open.": "No hay archivo de resultado guardado para abrir.",
+        "Add one or more album directories first.": "Añada primero uno o más directorios de álbum.",
+        "No supported images found in the selected directories.": "No se encontraron imágenes compatibles en los directorios seleccionados.",
+        "Select a model asset first.": "Seleccione primero un recurso de modelo.",
+        "Selected asset cannot be used as a model package.": "El recurso seleccionado no puede usarse como paquete de modelo.",
+        "Model settings saved.": "Configuración del modelo guardada.",
+        "Model loaded successfully.": "Modelo cargado correctamente.",
+        "Model loaded.": "Modelo cargado.",
+        "Cleared.": "Borrado.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "Cambió el manejo de múltiples rostros. Se recalcularán los embeddings de la galería.",
+        "Dataset validation passed.": "Validación del dataset superada.",
+        "Dataset validation failed.": "Validación del dataset fallida.",
+        "Album directories cleared. Existing clustering results are still available.": "Directorios de álbum borrados. Los resultados de agrupación existentes siguen disponibles.",
+        "Saved album clustering results were cleared.": "Se borraron los resultados de agrupación guardados.",
+        "Opened InsightFace homepage.": "Página principal de InsightFace abierta.",
+        "Opened InsightFace enterprise contact page.": "Página de contacto empresarial de InsightFace abierta.",
+    },
+    "fr": {
+        "Select album folder": "Sélectionner un dossier d’album",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "Aucun dossier d’album n’est sélectionné. Rebuild All effacera les résultats de clustering enregistrés. Continuer ?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "Rebuild All retraitera toutes les images des dossiers sélectionnés et remplacera les résultats enregistrés. Continuer ?",
+        "Model is not loaded. Please open Models.": "Le modèle n’est pas chargé. Ouvrez Models.",
+        "Image read failure.": "Échec de lecture de l’image.",
+        "Please upload a query image.": "Chargez une image de requête.",
+        "Please upload one or more gallery images or folders.": "Chargez une ou plusieurs images ou dossiers de galerie.",
+        "Please select identity folders or a dataset root.": "Sélectionnez des dossiers d’identité ou la racine d’un jeu de données.",
+        "Please validate this dataset and configuration before running evaluation.": "Validez ce jeu de données et cette configuration avant l’évaluation.",
+        "Run an evaluation first.": "Lancez d’abord une évaluation.",
+        "PDF export is unavailable. Please install reportlab.": "L’export PDF est indisponible. Installez reportlab.",
+        "Select a source image first.": "Sélectionnez d’abord une image source.",
+        "Select a target image or video first.": "Sélectionnez d’abord une image ou vidéo cible.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "Modèle de face swap introuvable. Téléchargez et choisissez un modèle dans Models.",
+        "Unsupported target file type.": "Type de fichier cible non pris en charge.",
+        "No saved result file to open.": "Aucun fichier de résultat enregistré à ouvrir.",
+        "Add one or more album directories first.": "Ajoutez d’abord un ou plusieurs dossiers d’album.",
+        "No supported images found in the selected directories.": "Aucune image prise en charge trouvée dans les dossiers sélectionnés.",
+        "Select a model asset first.": "Sélectionnez d’abord une ressource de modèle.",
+        "Selected asset cannot be used as a model package.": "La ressource sélectionnée ne peut pas être utilisée comme package de modèle.",
+        "Model settings saved.": "Paramètres du modèle enregistrés.",
+        "Model loaded successfully.": "Modèle chargé avec succès.",
+        "Model loaded.": "Modèle chargé.",
+        "Cleared.": "Effacé.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "La gestion multi-visages a changé. Les embeddings de galerie seront recalculés.",
+        "Dataset validation passed.": "Validation du jeu de données réussie.",
+        "Dataset validation failed.": "Validation du jeu de données échouée.",
+        "Album directories cleared. Existing clustering results are still available.": "Dossiers d’album effacés. Les résultats de clustering existants restent disponibles.",
+        "Saved album clustering results were cleared.": "Les résultats de clustering enregistrés ont été effacés.",
+        "Opened InsightFace homepage.": "Page d’accueil InsightFace ouverte.",
+        "Opened InsightFace enterprise contact page.": "Page de contact entreprise InsightFace ouverte.",
+    },
+    "de": {
+        "Select album folder": "Albumordner auswählen",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "Es sind keine Albumordner ausgewählt. Rebuild All löscht gespeicherte Clustering-Ergebnisse. Fortfahren?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "Rebuild All verarbeitet jedes Bild in den ausgewählten Albumordnern erneut und ersetzt gespeicherte Clustering-Ergebnisse. Fortfahren?",
+        "Model is not loaded. Please open Models.": "Das Modell ist nicht geladen. Öffnen Sie Models.",
+        "Image read failure.": "Bild konnte nicht gelesen werden.",
+        "Please upload a query image.": "Laden Sie ein Abfragebild hoch.",
+        "Please upload one or more gallery images or folders.": "Laden Sie ein oder mehrere Galeriebilder oder Ordner hoch.",
+        "Please select identity folders or a dataset root.": "Wählen Sie Identitätsordner oder eine Datensatzwurzel aus.",
+        "Please validate this dataset and configuration before running evaluation.": "Validieren Sie diesen Datensatz und die Konfiguration vor der Evaluierung.",
+        "Run an evaluation first.": "Führen Sie zuerst eine Evaluierung aus.",
+        "PDF export is unavailable. Please install reportlab.": "PDF-Export ist nicht verfügbar. Installieren Sie reportlab.",
+        "Select a source image first.": "Wählen Sie zuerst ein Quellbild aus.",
+        "Select a target image or video first.": "Wählen Sie zuerst ein Zielbild oder Video aus.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "Face-Swap-Modell nicht gefunden. Laden Sie in Models ein Modell herunter und wählen Sie es aus.",
+        "Unsupported target file type.": "Nicht unterstützter Zieldateityp.",
+        "No saved result file to open.": "Keine gespeicherte Ergebnisdatei zum Öffnen.",
+        "Add one or more album directories first.": "Fügen Sie zuerst einen oder mehrere Albumordner hinzu.",
+        "No supported images found in the selected directories.": "In den ausgewählten Ordnern wurden keine unterstützten Bilder gefunden.",
+        "Select a model asset first.": "Wählen Sie zuerst ein Modell-Asset aus.",
+        "Selected asset cannot be used as a model package.": "Das ausgewählte Asset kann nicht als Modellpaket verwendet werden.",
+        "Model settings saved.": "Modelleinstellungen gespeichert.",
+        "Model loaded successfully.": "Modell erfolgreich geladen.",
+        "Model loaded.": "Modell geladen.",
+        "Cleared.": "Geleert.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "Die Mehrgesicht-Verarbeitung wurde geändert. Galerie-Embeddings werden neu berechnet.",
+        "Dataset validation passed.": "Datensatzvalidierung bestanden.",
+        "Dataset validation failed.": "Datensatzvalidierung fehlgeschlagen.",
+        "Album directories cleared. Existing clustering results are still available.": "Albumordner geleert. Bestehende Clustering-Ergebnisse bleiben verfügbar.",
+        "Saved album clustering results were cleared.": "Gespeicherte Album-Clustering-Ergebnisse wurden gelöscht.",
+        "Opened InsightFace homepage.": "InsightFace-Homepage geöffnet.",
+        "Opened InsightFace enterprise contact page.": "InsightFace-Kontaktseite für Unternehmen geöffnet.",
+    },
+    "pt": {
+        "Select album folder": "Selecionar pasta de álbum",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "Não há diretórios de álbum selecionados. Rebuild All irá limpar os resultados de clustering guardados. Continuar?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "Rebuild All irá reprocessar todas as imagens nos diretórios selecionados e substituir os resultados guardados. Continuar?",
+        "Model is not loaded. Please open Models.": "O modelo não está carregado. Abra Models.",
+        "Image read failure.": "Falha ao ler a imagem.",
+        "Please upload a query image.": "Carregue uma imagem de consulta.",
+        "Please upload one or more gallery images or folders.": "Carregue uma ou mais imagens ou pastas da galeria.",
+        "Please select identity folders or a dataset root.": "Selecione pastas de identidade ou a raiz de um dataset.",
+        "Please validate this dataset and configuration before running evaluation.": "Valide este dataset e configuração antes de executar a avaliação.",
+        "Run an evaluation first.": "Execute primeiro uma avaliação.",
+        "PDF export is unavailable. Please install reportlab.": "A exportação PDF não está disponível. Instale reportlab.",
+        "Select a source image first.": "Selecione primeiro uma imagem de origem.",
+        "Select a target image or video first.": "Selecione primeiro uma imagem ou vídeo de destino.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "Modelo de troca de rostos não encontrado. Transfira e escolha um modelo em Models.",
+        "Unsupported target file type.": "Tipo de ficheiro de destino não suportado.",
+        "No saved result file to open.": "Não há ficheiro de resultado guardado para abrir.",
+        "Add one or more album directories first.": "Adicione primeiro um ou mais diretórios de álbum.",
+        "No supported images found in the selected directories.": "Não foram encontradas imagens suportadas nos diretórios selecionados.",
+        "Select a model asset first.": "Selecione primeiro um recurso de modelo.",
+        "Selected asset cannot be used as a model package.": "O recurso selecionado não pode ser usado como pacote de modelo.",
+        "Model settings saved.": "Definições do modelo guardadas.",
+        "Model loaded successfully.": "Modelo carregado com sucesso.",
+        "Model loaded.": "Modelo carregado.",
+        "Cleared.": "Limpo.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "O tratamento de múltiplos rostos mudou. Os embeddings da galeria serão recalculados.",
+        "Dataset validation passed.": "Validação do dataset concluída.",
+        "Dataset validation failed.": "Validação do dataset falhou.",
+        "Album directories cleared. Existing clustering results are still available.": "Diretórios de álbum limpos. Os resultados de clustering existentes continuam disponíveis.",
+        "Saved album clustering results were cleared.": "Os resultados de clustering guardados foram limpos.",
+        "Opened InsightFace homepage.": "Página inicial InsightFace aberta.",
+        "Opened InsightFace enterprise contact page.": "Página de contacto empresarial InsightFace aberta.",
+    },
+    "ru": {
+        "Select album folder": "Выбрать папку альбома",
+        "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?": "Папки альбомов не выбраны. Rebuild All очистит сохранённые результаты кластеризации. Продолжить?",
+        "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?": "Rebuild All заново обработает все изображения в выбранных папках и заменит сохранённые результаты кластеризации. Продолжить?",
+        "Model is not loaded. Please open Models.": "Модель не загружена. Откройте Models.",
+        "Image read failure.": "Не удалось прочитать изображение.",
+        "Please upload a query image.": "Загрузите изображение запроса.",
+        "Please upload one or more gallery images or folders.": "Загрузите одно или несколько изображений/папок галереи.",
+        "Please select identity folders or a dataset root.": "Выберите папки идентичностей или корень датасета.",
+        "Please validate this dataset and configuration before running evaluation.": "Проверьте этот датасет и конфигурацию перед запуском оценки.",
+        "Run an evaluation first.": "Сначала запустите оценку.",
+        "PDF export is unavailable. Please install reportlab.": "Экспорт PDF недоступен. Установите reportlab.",
+        "Select a source image first.": "Сначала выберите исходное изображение.",
+        "Select a target image or video first.": "Сначала выберите целевое изображение или видео.",
+        "Face swap model not found. Please download and choose a swap model in Models.": "Модель замены лиц не найдена. Загрузите и выберите модель в Models.",
+        "Unsupported target file type.": "Неподдерживаемый тип целевого файла.",
+        "No saved result file to open.": "Нет сохранённого файла результата для открытия.",
+        "Add one or more album directories first.": "Сначала добавьте одну или несколько папок альбомов.",
+        "No supported images found in the selected directories.": "В выбранных папках не найдены поддерживаемые изображения.",
+        "Select a model asset first.": "Сначала выберите ресурс модели.",
+        "Selected asset cannot be used as a model package.": "Выбранный ресурс нельзя использовать как пакет модели.",
+        "Model settings saved.": "Настройки модели сохранены.",
+        "Model loaded successfully.": "Модель успешно загружена.",
+        "Model loaded.": "Модель загружена.",
+        "Cleared.": "Очищено.",
+        "Multi-face handling changed. Gallery embeddings will be recomputed.": "Обработка нескольких лиц изменена. Эмбеддинги галереи будут пересчитаны.",
+        "Dataset validation passed.": "Проверка датасета пройдена.",
+        "Dataset validation failed.": "Проверка датасета не пройдена.",
+        "Album directories cleared. Existing clustering results are still available.": "Папки альбомов очищены. Существующие результаты кластеризации всё ещё доступны.",
+        "Saved album clustering results were cleared.": "Сохранённые результаты кластеризации альбома очищены.",
+        "Opened InsightFace homepage.": "Домашняя страница InsightFace открыта.",
+        "Opened InsightFace enterprise contact page.": "Страница контактов InsightFace для предприятий открыта.",
+    },
+}
+
+for _lang, _values in _STATUS_ERROR_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
+
+
+_PAGE_COPY_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "导入或刷新本地相册文件夹，对检测到的人脸进行聚类，并查看每个人物组中的照片。",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "所有相册处理均在本地完成。导入 / 刷新会扫描每个已选择的文件夹，仅为新增图片提取特征，然后使用所选余弦阈值对全部已索引人脸运行 DBSCAN 聚类。",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "上传一张查询图片，以及一张 Gallery 图片、一组图片或一个文件夹。只有一张 Gallery 图片时执行 1:1 比对；多张 Gallery 图片时执行 1:N Gallery 搜索。",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "配置模型包、运行后端、换脸模型和运行时检查。",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "手动刷新 GitHub Release 模型下载地址，并将所选模型包下载到本地。",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "代码和模型文件可能适用不同授权。商业部署需要相应的模型授权。",
+        "Refresh source": "刷新来源",
+        "Local model root": "本地模型根目录",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "源图 + 目标 = 结果。目标可以是图片或视频，工作流会自动选择图片换脸或视频换脸。",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "根据使用场景和模型授权，换脸功能可能需要单独的商业授权。请仅在拥有适当权利和同意的情况下使用。",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "基于身份文件夹运行本地 1:1 验证或 1:N 识别评测，并导出可用于采购决策的报告。",
+    },
+    "ja": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "ローカルのアルバムフォルダをインポートまたは更新し、検出された顔をクラスタリングして、各人物グループの写真を確認します。",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "すべてのアルバム処理はローカルで実行されます。インポート / 更新では選択された各フォルダをスキャンし、新しい画像のみ特徴量を抽出してから、選択したコサインしきい値で全インデックス済み顔に DBSCAN クラスタリングを実行します。",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "1枚のクエリ画像と、ギャラリー画像、画像セット、またはフォルダをアップロードします。ギャラリーが1枚なら 1:1 比較、複数なら 1:N ギャラリー検索を実行します。",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "モデルパック、実行プロバイダ、顔交換モデル、ランタイム確認を設定します。",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "GitHub Release のモデル URL を手動で更新し、選択したモデルパッケージをローカルにダウンロードします。",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "コードとモデルファイルには異なるライセンスが適用される場合があります。商用展開には適切なモデル認可が必要です。",
+        "Refresh source": "更新元",
+        "Local model root": "ローカルモデルルート",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "ソース + ターゲット = 結果。ターゲットは画像または動画に対応し、ワークフローが画像または動画の顔交換を自動的に選択します。",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "利用目的やモデルライセンスによっては、顔交換に別途商用許諾が必要になる場合があります。適切な権利と同意がある場合にのみ使用してください。",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "IDフォルダからローカルの 1:1 検証または 1:N 識別評価を実行し、調達判断に使えるレポートを出力します。",
+    },
+    "ko": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "로컬 앨범 폴더를 가져오거나 새로고침하고, 검출된 얼굴을 클러스터링한 뒤 각 사람 그룹의 사진을 검토합니다.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "모든 앨범 처리는 로컬에서 수행됩니다. 가져오기 / 새로고침은 선택한 모든 폴더를 스캔하고 새 이미지에 대해서만 특징을 추출한 다음, 선택한 코사인 임계값으로 모든 인덱싱된 얼굴에 DBSCAN 클러스터링을 실행합니다.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "쿼리 이미지 하나와 갤러리 이미지, 이미지 세트 또는 폴더를 업로드합니다. 갤러리 이미지가 하나이면 1:1 비교를 실행하고, 여러 장이면 1:N 갤러리 검색을 실행합니다.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "모델 패키지, 실행 제공자, 얼굴 교체 모델, 런타임 점검을 구성합니다.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "GitHub Release 모델 URL을 수동으로 새로고침하고 선택한 모델 패키지를 로컬에 다운로드합니다.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "코드와 모델 파일에는 서로 다른 라이선스가 적용될 수 있습니다. 상용 배포에는 적절한 모델 승인이 필요합니다.",
+        "Refresh source": "새로고침 소스",
+        "Local model root": "로컬 모델 루트",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "소스 + 타깃 = 결과. 타깃은 이미지 또는 비디오일 수 있으며, 워크플로가 이미지 또는 비디오 얼굴 교체를 자동으로 선택합니다.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "사용 목적과 모델 라이선스에 따라 얼굴 교체에는 별도의 상업적 허가가 필요할 수 있습니다. 적절한 권리와 동의가 있는 경우에만 사용하십시오.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "ID 폴더에서 로컬 1:1 검증 또는 1:N 식별 평가를 실행하고 구매 검토용 보고서를 내보냅니다.",
+    },
+    "es": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "Importe o actualice carpetas de álbum locales, agrupe los rostros detectados y revise las fotos de cada grupo de personas.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "Todo el procesamiento del álbum es local. Importar / Actualizar escanea cada carpeta seleccionada, extrae características solo de imágenes nuevas y luego ejecuta DBSCAN sobre todos los rostros indexados con el umbral de coseno seleccionado.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "Suba una imagen de consulta y una imagen de galería, un conjunto de imágenes o una carpeta. Una sola imagen de galería ejecuta comparación 1:1; varias imágenes ejecutan búsqueda 1:N en la galería.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "Configure paquetes de modelos, proveedor de ejecución, modelos de intercambio facial y comprobaciones de runtime.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "Actualice manualmente las URL de modelos de GitHub Releases y descargue localmente los paquetes seleccionados.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "El código y los archivos de modelo pueden tener licencias diferentes. El despliegue comercial requiere autorización de modelo adecuada.",
+        "Refresh source": "Fuente de actualización",
+        "Local model root": "Raíz local de modelos",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "Origen + destino = resultado. El destino puede ser una imagen o un vídeo; el flujo elige automáticamente intercambio en imagen o vídeo.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "El intercambio facial puede requerir autorización comercial separada según el uso y la licencia del modelo. Úselo solo con los derechos y consentimientos adecuados.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "Ejecute evaluación local de verificación 1:1 o identificación 1:N desde carpetas de identidad y exporte informes listos para compras.",
+    },
+    "fr": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "Importez ou actualisez des dossiers d’album locaux, regroupez les visages détectés et examinez les photos de chaque groupe de personnes.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "Tout le traitement de l’album est local. Importer / Actualiser analyse chaque dossier sélectionné, extrait des caractéristiques uniquement pour les nouvelles images, puis exécute DBSCAN sur tous les visages indexés avec le seuil cosinus choisi.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "Chargez une image de requête et une image de galerie, un ensemble d’images ou un dossier. Une seule image de galerie lance une comparaison 1:1 ; plusieurs images lancent une recherche 1:N dans la galerie.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "Configurez les packages de modèles, le fournisseur d’exécution, les modèles de face swap et les vérifications runtime.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "Actualisez manuellement les URL des modèles GitHub Releases et téléchargez localement les packages sélectionnés.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "Le code et les fichiers de modèle peuvent avoir des licences différentes. Un déploiement commercial requiert une autorisation de modèle appropriée.",
+        "Refresh source": "Source d’actualisation",
+        "Local model root": "Racine locale des modèles",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "Source + cible = résultat. La cible peut être une image ou une vidéo ; le workflow choisit automatiquement le mode de face swap adapté.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "Le face swap peut nécessiter une autorisation commerciale distincte selon l’usage et la licence du modèle. Utilisez-le uniquement avec les droits et consentements appropriés.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "Exécutez une évaluation locale de vérification 1:1 ou d’identification 1:N depuis des dossiers d’identité et exportez des rapports prêts pour les achats.",
+    },
+    "de": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "Importieren oder aktualisieren Sie lokale Albumordner, clustern Sie erkannte Gesichter und prüfen Sie die Fotos jeder Personengruppe.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "Die gesamte Albumverarbeitung erfolgt lokal. Importieren / Aktualisieren scannt jeden ausgewählten Ordner, extrahiert Merkmale nur für neue Bilder und führt anschließend DBSCAN-Clustering über alle indizierten Gesichter mit dem gewählten Kosinusschwellenwert aus.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "Laden Sie ein Abfragebild und ein Galeriebild, einen Bildsatz oder einen Ordner hoch. Ein Galeriebild startet einen 1:1-Vergleich; mehrere Galeriebilder starten eine 1:N-Galeriesuche.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "Konfigurieren Sie Modellpakete, Ausführungsprovider, Face-Swap-Modelle und Laufzeitprüfungen.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "Aktualisieren Sie GitHub-Release-Modell-URLs manuell und laden Sie ausgewählte Modellpakete lokal herunter.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "Code- und Modelldateien können unterschiedliche Lizenzen haben. Kommerzielle Bereitstellung erfordert eine geeignete Modellautorisierung.",
+        "Refresh source": "Aktualisierungsquelle",
+        "Local model root": "Lokaler Modellstamm",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "Quelle + Ziel = Ergebnis. Das Ziel kann ein Bild oder Video sein; der Workflow wählt automatisch Bild- oder Video-Face-Swap.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "Face Swap kann je nach Nutzung und Modelllizenz eine separate kommerzielle Genehmigung erfordern. Verwenden Sie es nur mit geeigneten Rechten und Einwilligungen.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "Führen Sie lokale 1:1-Verifikation oder 1:N-Identifikation aus Identitätsordnern aus und exportieren Sie beschaffungsreife Berichte.",
+    },
+    "pt": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "Importe ou atualize pastas de álbum locais, agrupe os rostos detetados e reveja as fotos de cada grupo de pessoas.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "Todo o processamento do álbum é local. Importar / Atualizar analisa cada pasta selecionada, extrai características apenas de imagens novas e depois executa DBSCAN sobre todos os rostos indexados usando o limiar de cosseno selecionado.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "Carregue uma imagem de consulta e uma imagem de galeria, conjunto de imagens ou pasta. Uma imagem de galeria executa comparação 1:1; várias imagens executam pesquisa 1:N na galeria.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "Configure pacotes de modelos, fornecedor de execução, modelos de troca de rostos e verificações de runtime.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "Atualize manualmente os URLs de modelos do GitHub Releases e transfira localmente os pacotes selecionados.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "O código e os ficheiros de modelo podem ter licenças diferentes. A implantação comercial requer autorização de modelo adequada.",
+        "Refresh source": "Fonte de atualização",
+        "Local model root": "Raiz local dos modelos",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "Origem + destino = resultado. O destino pode ser imagem ou vídeo; o fluxo escolhe automaticamente troca em imagem ou vídeo.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "A troca de rostos pode exigir autorização comercial separada dependendo do uso e da licença do modelo. Use apenas com direitos e consentimento adequados.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "Execute avaliação local de verificação 1:1 ou identificação 1:N a partir de pastas de identidade e exporte relatórios prontos para aquisição.",
+    },
+    "ru": {
+        "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.": "Импортируйте или обновляйте локальные папки альбомов, кластеризуйте обнаруженные лица и просматривайте фотографии в каждой группе людей.",
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.": "Вся обработка альбомов выполняется локально. Import / Refresh сканирует каждую выбранную папку, извлекает признаки только для новых изображений, затем запускает DBSCAN по всем индексированным лицам с выбранным косинусным порогом.",
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.": "Загрузите одно изображение запроса и изображение галереи, набор изображений или папку. Одно изображение галереи запускает сравнение 1:1; несколько изображений запускают поиск 1:N по галерее.",
+        "Configure model packs, execution provider, face swap models, and runtime checks.": "Настройте пакеты моделей, провайдер выполнения, модели замены лиц и проверки runtime.",
+        "Manually refresh GitHub release model URLs and download selected model packages locally.": "Вручную обновляйте URL моделей из GitHub Releases и загружайте выбранные пакеты моделей локально.",
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.": "Код и файлы моделей могут иметь разные лицензии. Коммерческое развертывание требует соответствующей авторизации модели.",
+        "Refresh source": "Источник обновления",
+        "Local model root": "Локальный корень моделей",
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.": "Источник + цель = результат. Целью может быть изображение или видео; рабочий процесс автоматически выбирает замену лиц для изображения или видео.",
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.": "Замена лиц может требовать отдельного коммерческого разрешения в зависимости от использования и лицензии модели. Используйте только при наличии соответствующих прав и согласия.",
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.": "Запускайте локальную оценку верификации 1:1 или идентификации 1:N из папок идентичностей и экспортируйте отчёты для закупочных решений.",
+    },
+}
+
+for _lang, _values in _PAGE_COPY_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
+
+
+_REPORT_UI_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "zh": {
+        "Evaluation complete. Report saved to:": "评测完成。报告已保存至：",
+        "Evaluation report saved": "评测报告已保存",
+        "Evaluation report saved to:": "评测报告已保存至：",
+        "Scenario": "场景",
+        "Dataset": "数据集",
+        "Report PDF": "PDF 报告",
+        "Report Markdown": "Markdown 报告",
+        "Report HTML": "HTML 报告",
+        "not exported": "未导出",
+        "Metrics": "指标",
+        "Errors": "错误",
+        "First errors": "前几个错误",
+        "Selected dataset:": "已选择数据集：",
+        "No dataset selected.": "尚未选择数据集。",
+    },
+    "ja": {
+        "Evaluation complete. Report saved to:": "評価が完了しました。レポートの保存先：",
+        "Evaluation report saved": "評価レポートを保存しました",
+        "Evaluation report saved to:": "評価レポートの保存先：",
+        "Scenario": "シナリオ",
+        "Dataset": "データセット",
+        "Report PDF": "PDF レポート",
+        "Report Markdown": "Markdown レポート",
+        "Report HTML": "HTML レポート",
+        "not exported": "未出力",
+        "Metrics": "指標",
+        "Errors": "エラー",
+        "First errors": "最初のエラー",
+        "Selected dataset:": "選択中のデータセット：",
+        "No dataset selected.": "データセットが選択されていません。",
+    },
+    "ko": {
+        "Evaluation complete. Report saved to:": "평가가 완료되었습니다. 보고서 저장 위치:",
+        "Evaluation report saved": "평가 보고서 저장됨",
+        "Evaluation report saved to:": "평가 보고서 저장 위치:",
+        "Scenario": "시나리오",
+        "Dataset": "데이터셋",
+        "Report PDF": "PDF 보고서",
+        "Report Markdown": "Markdown 보고서",
+        "Report HTML": "HTML 보고서",
+        "not exported": "내보내지 않음",
+        "Metrics": "지표",
+        "Errors": "오류",
+        "First errors": "첫 번째 오류",
+        "Selected dataset:": "선택한 데이터셋:",
+        "No dataset selected.": "선택한 데이터셋이 없습니다.",
+    },
+    "es": {
+        "Evaluation complete. Report saved to:": "Evaluación completa. Informe guardado en:",
+        "Evaluation report saved": "Informe de evaluación guardado",
+        "Evaluation report saved to:": "Informe de evaluación guardado en:",
+        "Scenario": "Escenario",
+        "Dataset": "Dataset",
+        "Report PDF": "Informe PDF",
+        "Report Markdown": "Informe Markdown",
+        "Report HTML": "Informe HTML",
+        "not exported": "no exportado",
+        "Metrics": "Métricas",
+        "Errors": "Errores",
+        "First errors": "Primeros errores",
+        "Selected dataset:": "Dataset seleccionado:",
+        "No dataset selected.": "No se ha seleccionado ningún dataset.",
+    },
+    "fr": {
+        "Evaluation complete. Report saved to:": "Évaluation terminée. Rapport enregistré dans :",
+        "Evaluation report saved": "Rapport d’évaluation enregistré",
+        "Evaluation report saved to:": "Rapport d’évaluation enregistré dans :",
+        "Scenario": "Scénario",
+        "Dataset": "Jeu de données",
+        "Report PDF": "Rapport PDF",
+        "Report Markdown": "Rapport Markdown",
+        "Report HTML": "Rapport HTML",
+        "not exported": "non exporté",
+        "Metrics": "Métriques",
+        "Errors": "Erreurs",
+        "First errors": "Premières erreurs",
+        "Selected dataset:": "Jeu de données sélectionné :",
+        "No dataset selected.": "Aucun jeu de données sélectionné.",
+    },
+    "de": {
+        "Evaluation complete. Report saved to:": "Evaluierung abgeschlossen. Bericht gespeichert unter:",
+        "Evaluation report saved": "Evaluierungsbericht gespeichert",
+        "Evaluation report saved to:": "Evaluierungsbericht gespeichert unter:",
+        "Scenario": "Szenario",
+        "Dataset": "Datensatz",
+        "Report PDF": "PDF-Bericht",
+        "Report Markdown": "Markdown-Bericht",
+        "Report HTML": "HTML-Bericht",
+        "not exported": "nicht exportiert",
+        "Metrics": "Metriken",
+        "Errors": "Fehler",
+        "First errors": "Erste Fehler",
+        "Selected dataset:": "Ausgewählter Datensatz:",
+        "No dataset selected.": "Kein Datensatz ausgewählt.",
+    },
+    "pt": {
+        "Evaluation complete. Report saved to:": "Avaliação concluída. Relatório guardado em:",
+        "Evaluation report saved": "Relatório de avaliação guardado",
+        "Evaluation report saved to:": "Relatório de avaliação guardado em:",
+        "Scenario": "Cenário",
+        "Dataset": "Dataset",
+        "Report PDF": "Relatório PDF",
+        "Report Markdown": "Relatório Markdown",
+        "Report HTML": "Relatório HTML",
+        "not exported": "não exportado",
+        "Metrics": "Métricas",
+        "Errors": "Erros",
+        "First errors": "Primeiros erros",
+        "Selected dataset:": "Dataset selecionado:",
+        "No dataset selected.": "Nenhum dataset selecionado.",
+    },
+    "ru": {
+        "Evaluation complete. Report saved to:": "Оценка завершена. Отчёт сохранён в:",
+        "Evaluation report saved": "Отчёт оценки сохранён",
+        "Evaluation report saved to:": "Отчёт оценки сохранён в:",
+        "Scenario": "Сценарий",
+        "Dataset": "Датасет",
+        "Report PDF": "PDF-отчёт",
+        "Report Markdown": "Markdown-отчёт",
+        "Report HTML": "HTML-отчёт",
+        "not exported": "не экспортировано",
+        "Metrics": "Метрики",
+        "Errors": "Ошибки",
+        "First errors": "Первые ошибки",
+        "Selected dataset:": "Выбранный датасет:",
+        "No dataset selected.": "Датасет не выбран.",
+    },
+}
+
+for _lang, _values in _REPORT_UI_TRANSLATIONS.items():
+    _TRANSLATIONS.setdefault(_lang, {}).update(_values)
diff --git a/python-package/insightface/gui/core/licensing.py b/python-package/insightface/gui/core/licensing.py
new file mode 100644
index 000000000..baff5a34f
--- /dev/null
+++ b/python-package/insightface/gui/core/licensing.py
@@ -0,0 +1,46 @@
+"""License center helpers."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Dict
+
+from .constants import COMMERCIAL_NOTICE, LICENSE_NOTICE, RESPONSIBLE_USE_NOTICE
+
+
+def find_license_text(start: str | Path) -> str:
+    root = Path(start).resolve()
+    candidates = [root / "LICENSE", root / "LICENSE.md", root.parent / "LICENSE", root.parent / "LICENSE.md"]
+    for path in candidates:
+        if path.exists():
+            text = path.read_text(encoding="utf-8", errors="ignore").strip()
+            return text[:4000]
+    return "Please refer to the LICENSE file in this repository for the code license."
+
+
+def allowed_usage_summary() -> Dict[str, str]:
+    return {
+        "Personal research": "yes",
+        "Internal evaluation": "depends on model license",
+        "Commercial production": "requires commercial model license",
+        "Redistribution": "requires explicit permission",
+        "SaaS / API usage": "requires commercial agreement",
+        "Face swap commercial usage": "requires commercial agreement",
+    }
+
+
+def license_summary_text(status: str, model_name: str, provider: str, workspace: str) -> str:
+    lines = [
+        "# InsightFace License Summary",
+        "",
+        f"- Current status: {status}",
+        f"- Model: {model_name}",
+        f"- Provider: {provider}",
+        f"- Workspace: {workspace}",
+        "",
+        LICENSE_NOTICE,
+        COMMERCIAL_NOTICE,
+        RESPONSIBLE_USE_NOTICE,
+        "This tool does not provide legal advice.",
+    ]
+    return "\n".join(lines)
diff --git a/python-package/insightface/gui/core/links.py b/python-package/insightface/gui/core/links.py
new file mode 100644
index 000000000..eb0e8e82b
--- /dev/null
+++ b/python-package/insightface/gui/core/links.py
@@ -0,0 +1,40 @@
+"""External link helpers for GUI-origin attribution."""
+
+from __future__ import annotations
+
+from urllib.parse import parse_qsl, urlencode, urlsplit, urlunsplit
+
+from PySide6.QtCore import QUrl
+from PySide6.QtGui import QDesktopServices
+
+INSIGHTFACE_LINK_HOSTS = {"insightface.ai", "www.insightface.ai"}
+GUI_UTM_SOURCE = "insightface_gui"
+
+
+def add_gui_referrer(url: str, *, content: str | None = None) -> str:
+    """Add GUI source attribution to InsightFace website URLs.
+
+    Native desktop apps cannot reliably set the browser's HTTP Referer header
+    when delegating a URL to the user's default browser. Query attribution is
+    therefore the stable, privacy-visible path for website-side analytics.
+
+    The `content` argument is accepted for API compatibility but intentionally
+    not emitted; Vercel UTM reporting only needs `utm_source` for this app.
+    """
+
+    parts = urlsplit(url)
+    if parts.scheme not in {"http", "https"} or parts.hostname not in INSIGHTFACE_LINK_HOSTS:
+        return url
+
+    params = parse_qsl(parts.query, keep_blank_values=True)
+    existing = {key for key, _ in params}
+    if "utm_source" not in existing:
+        params.append(("utm_source", GUI_UTM_SOURCE))
+
+    return urlunsplit((parts.scheme, parts.netloc, parts.path, urlencode(params), parts.fragment))
+
+
+def open_insightface_url(url: str, *, content: str | None = None) -> bool:
+    """Open an InsightFace website URL with GUI attribution parameters."""
+
+    return QDesktopServices.openUrl(QUrl(add_gui_referrer(url, content=content)))
diff --git a/python-package/insightface/gui/core/logging.py b/python-package/insightface/gui/core/logging.py
new file mode 100644
index 000000000..45ae94a1e
--- /dev/null
+++ b/python-package/insightface/gui/core/logging.py
@@ -0,0 +1,25 @@
+"""Logging setup."""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+
+
+def setup_logging(log_dir: str | Path) -> Path:
+    path = Path(log_dir).expanduser()
+    path.mkdir(parents=True, exist_ok=True)
+    log_file = path / "app.log"
+    logger = logging.getLogger("insightface.gui")
+    logger.setLevel(logging.INFO)
+    if not any(isinstance(handler, logging.FileHandler) and handler.baseFilename == str(log_file) for handler in logger.handlers):
+        handler = logging.FileHandler(log_file, encoding="utf-8")
+        handler.setFormatter(
+            logging.Formatter("%(asctime)s %(levelname)s %(name)s: %(message)s")
+        )
+        logger.addHandler(handler)
+    return log_file
+
+
+def get_logger(name: str) -> logging.Logger:
+    return logging.getLogger(f"insightface.gui.{name}")
diff --git a/python-package/insightface/gui/core/model_downloads.py b/python-package/insightface/gui/core/model_downloads.py
new file mode 100644
index 000000000..3604b1207
--- /dev/null
+++ b/python-package/insightface/gui/core/model_downloads.py
@@ -0,0 +1,356 @@
+"""Manual model download catalog and helpers.
+
+The GUI never downloads models automatically. Users must open Model Downloads,
+refresh URLs, and explicitly start a download.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import time
+import urllib.error
+import urllib.request
+import zipfile
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Callable, Iterable, List, Optional
+
+from .utils import safe_json_dumps, utc_now_iso
+
+GITHUB_RELEASES_URL = "https://github.com/deepinsight/insightface/releases"
+GITHUB_LATEST_RELEASE_API = "https://api.github.com/repos/deepinsight/insightface/releases/latest"
+FALLBACK_RELEASE_TAG = "v0.7"
+FALLBACK_RELEASE_NAME = "insightface v0.7 model packages"
+GFPGAN_MODEL_NAME = "GFPGANv1.4.onnx"
+GFPGAN_DOWNLOAD_URL = (
+    "https://github.com/harisreedhar/Face-Upscalers-ONNX/releases/download/"
+    "Models/GFPGANv1.4.onnx"
+)
+
+
+@dataclass
+class ModelAsset:
+    name: str
+    browser_download_url: str
+    tag_name: str = FALLBACK_RELEASE_TAG
+    release_name: str = FALLBACK_RELEASE_NAME
+    size: int = 0
+    content_type: str = ""
+    updated_at: str = ""
+    source: str = "InsightFace"
+
+    @property
+    def stem(self) -> str:
+        if self.name.endswith(".zip"):
+            return self.name[:-4]
+        if self.name.endswith(".onnx"):
+            return self.name[:-5]
+        return Path(self.name).stem
+
+    @property
+    def kind(self) -> str:
+        if self.source.lower() != "insightface":
+            return "third-party restore model"
+        if self.name.endswith(".zip"):
+            return "model package"
+        if self.name.endswith(".onnx"):
+            return "onnx model"
+        return "asset"
+
+
+def fallback_model_assets() -> List[ModelAsset]:
+    # Latest GitHub release is v0.7. GitHub's HTML page reports nine assets,
+    # and the SourceForge mirror lists the uploaded model files. The GitHub
+    # download URLs follow the release asset URL pattern.
+    names = [
+        "antelopev2.zip",
+        "buffalo_l.zip",
+        "buffalo_m.zip",
+        "buffalo_s.zip",
+        "buffalo_sc.zip",
+        "inswapper_128.onnx",
+        "scrfd_person_2.5g.onnx",
+    ]
+    assets = [
+        ModelAsset(
+            name=name,
+            browser_download_url=f"https://github.com/deepinsight/insightface/releases/download/{FALLBACK_RELEASE_TAG}/{name}",
+            tag_name=FALLBACK_RELEASE_TAG,
+            release_name=FALLBACK_RELEASE_NAME,
+        )
+        for name in names
+    ]
+    return merge_required_assets(assets)
+
+
+def third_party_model_assets() -> List[ModelAsset]:
+    return [
+        ModelAsset(
+            name=GFPGAN_MODEL_NAME,
+            browser_download_url=GFPGAN_DOWNLOAD_URL,
+            tag_name="third-party",
+            release_name="harisreedhar/Face-Upscalers-ONNX GFPGAN v1.4",
+            size=340 * 1024 * 1024,
+            content_type="application/octet-stream",
+            source="third party",
+        )
+    ]
+
+
+def merge_required_assets(assets: Iterable[ModelAsset]) -> List[ModelAsset]:
+    merged = list(assets)
+    seen = {asset.name for asset in merged}
+    for asset in third_party_model_assets():
+        if asset.name not in seen:
+            merged.append(asset)
+            seen.add(asset.name)
+    return merged
+
+
+def cache_file(gui_cache_dir: str | os.PathLike[str]) -> Path:
+    path = Path(gui_cache_dir).expanduser() / "model_download_urls.json"
+    path.parent.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def asset_from_dict(data: dict) -> ModelAsset:
+    return ModelAsset(
+        name=str(data.get("name", "")),
+        browser_download_url=str(data.get("browser_download_url", "")),
+        tag_name=str(data.get("tag_name", data.get("release_tag", FALLBACK_RELEASE_TAG))),
+        release_name=str(data.get("release_name", FALLBACK_RELEASE_NAME)),
+        size=int(data.get("size") or 0),
+        content_type=str(data.get("content_type", "")),
+        updated_at=str(data.get("updated_at", "")),
+        source=str(data.get("source", "InsightFace")),
+    )
+
+
+def load_cached_assets(gui_cache_dir: str | os.PathLike[str]) -> List[ModelAsset]:
+    path = cache_file(gui_cache_dir)
+    if not path.exists():
+        return fallback_model_assets()
+    try:
+        payload = json.loads(path.read_text(encoding="utf-8"))
+        assets = [asset_from_dict(item) for item in payload.get("assets", []) if item.get("browser_download_url")]
+        return merge_required_assets(assets)
+    except Exception:
+        return fallback_model_assets()
+
+
+def save_cached_assets(gui_cache_dir: str | os.PathLike[str], assets: Iterable[ModelAsset], source: str = "") -> Path:
+    path = cache_file(gui_cache_dir)
+    payload = {
+        "source": source or GITHUB_LATEST_RELEASE_API,
+        "refreshed_at": utc_now_iso(),
+        "assets": [asdict(asset) for asset in assets],
+    }
+    path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")
+    return path
+
+
+def refresh_model_assets(gui_cache_dir: str | os.PathLike[str], timeout: int = 20) -> tuple[List[ModelAsset], str]:
+    request = urllib.request.Request(
+        GITHUB_LATEST_RELEASE_API,
+        headers={
+            "Accept": "application/vnd.github+json",
+            "User-Agent": "InsightFace-Evaluation-Studio/1.0.1",
+        },
+    )
+    try:
+        with urllib.request.urlopen(request, timeout=timeout) as response:
+            payload = json.loads(response.read().decode("utf-8"))
+    except (urllib.error.URLError, TimeoutError, json.JSONDecodeError) as exc:
+        assets = fallback_model_assets()
+        save_cached_assets(gui_cache_dir, assets, source=f"fallback after refresh error: {exc}")
+        return assets, f"Refresh failed; using bundled {FALLBACK_RELEASE_TAG} URLs. Error: {exc}"
+
+    tag = payload.get("tag_name") or FALLBACK_RELEASE_TAG
+    release_name = payload.get("name") or tag
+    assets = []
+    for item in payload.get("assets", []):
+        name = item.get("name") or ""
+        url = item.get("browser_download_url") or ""
+        if not url or not (name.endswith(".zip") or name.endswith(".onnx")):
+            continue
+        assets.append(
+            ModelAsset(
+                name=name,
+                browser_download_url=url,
+                tag_name=tag,
+                release_name=release_name,
+                size=int(item.get("size") or 0),
+                content_type=item.get("content_type") or "",
+                updated_at=item.get("updated_at") or "",
+            )
+        )
+    if not assets:
+        assets = fallback_model_assets()
+        message = f"Latest release returned no model assets; using bundled {FALLBACK_RELEASE_TAG} URLs."
+    else:
+        assets = merge_required_assets(assets)
+        message = f"Refreshed {len(assets)} asset(s), including third-party restore models."
+    save_cached_assets(gui_cache_dir, assets, source=GITHUB_LATEST_RELEASE_API)
+    return assets, message
+
+
+def local_model_status(asset: ModelAsset, model_root: str | os.PathLike[str]) -> str:
+    root = Path(model_root).expanduser()
+    if asset.name.endswith(".zip"):
+        target = root / "models" / asset.stem
+        if target.exists() and any(target.glob("*.onnx")):
+            return f"installed: {target}"
+        if target.exists():
+            return f"folder exists: {target}"
+        return "not installed"
+    target_file = root / "models" / asset.stem / asset.name
+    if target_file.exists():
+        return f"installed: {target_file}"
+    legacy_file = root / "models" / asset.name
+    if legacy_file.exists():
+        return f"installed: {legacy_file}"
+    return "not installed"
+
+
+def is_model_package_installed(model_name: str, model_root: str | os.PathLike[str]) -> bool:
+    target = Path(model_root).expanduser() / "models" / model_name
+    return target.exists() and any(target.glob("*.onnx"))
+
+
+def list_installed_swap_models(model_root: str | os.PathLike[str]) -> list[Path]:
+    root = Path(model_root).expanduser() / "models"
+    if not root.exists():
+        return []
+    paths = []
+    for path in root.rglob("*.onnx"):
+        name = path.name.lower()
+        parent = path.parent.name.lower()
+        if "swap" in name or "inswapper" in name or "swap" in parent or "inswapper" in parent:
+            paths.append(path)
+    return sorted(paths)
+
+
+def list_installed_gfpgan_models(model_root: str | os.PathLike[str]) -> list[Path]:
+    root = Path(model_root).expanduser() / "models"
+    if not root.exists():
+        return []
+    return sorted(path for path in root.rglob("*.onnx") if "gfpgan" in path.name.lower())
+
+
+def _content_range_total(value: str | None) -> int:
+    if not value or "/" not in value:
+        return 0
+    total = value.rsplit("/", 1)[-1].strip()
+    if not total or total == "*":
+        return 0
+    try:
+        return int(total)
+    except ValueError:
+        return 0
+
+
+def _download_with_retries(
+    url: str,
+    destination: Path,
+    asset_name: str,
+    expected_size: int = 0,
+    progress: Optional[Callable[[int, int, str], None]] = None,
+    retries: int = 4,
+) -> None:
+    partial = destination.with_suffix(destination.suffix + ".part")
+    last_error: Exception | None = None
+    for attempt in range(1, retries + 1):
+        existing = partial.stat().st_size if partial.exists() else 0
+        headers = {
+            "Accept": "application/octet-stream,*/*",
+            "User-Agent": "Mozilla/5.0 InsightFace-Evaluation-Studio/1.0.1",
+        }
+        if existing:
+            headers["Range"] = f"bytes={existing}-"
+        request = urllib.request.Request(url, headers=headers)
+        try:
+            with urllib.request.urlopen(request, timeout=120) as response:
+                status = getattr(response, "status", response.getcode())
+                content_length = int(response.headers.get("Content-Length") or 0)
+                content_range_total = _content_range_total(response.headers.get("Content-Range"))
+                if existing and status != 206:
+                    existing = 0
+                    mode = "wb"
+                elif existing:
+                    mode = "ab"
+                else:
+                    mode = "wb"
+                if content_range_total:
+                    total = content_range_total
+                elif status == 206:
+                    total = existing + content_length
+                else:
+                    total = content_length or expected_size
+                downloaded = existing if mode == "ab" else 0
+                with partial.open(mode) as handle:
+                    while True:
+                        chunk = response.read(1024 * 1024)
+                        if not chunk:
+                            break
+                        handle.write(chunk)
+                        downloaded += len(chunk)
+                        if progress:
+                            progress(downloaded, total or downloaded, f"Downloading {asset_name}")
+            partial.replace(destination)
+            return
+        except (urllib.error.URLError, TimeoutError, ConnectionResetError, OSError) as exc:
+            if isinstance(exc, urllib.error.HTTPError) and exc.code == 416 and partial.exists():
+                partial.replace(destination)
+                return
+            last_error = exc
+            if progress:
+                progress(
+                    existing,
+                    expected_size or existing or 1,
+                    f"Download interrupted; retrying {asset_name} ({attempt}/{retries})",
+                )
+            if attempt < retries:
+                time.sleep(min(2 * attempt, 8))
+    raise RuntimeError(f"Download failed for {asset_name} after {retries} attempt(s): {last_error}")
+
+
+def download_model_asset(
+    asset: ModelAsset,
+    model_root: str | os.PathLike[str],
+    gui_cache_dir: str | os.PathLike[str],
+    progress: Optional[Callable[[int, int, str], None]] = None,
+) -> Path:
+    cache_dir = Path(gui_cache_dir).expanduser() / "models"
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    model_root_path = Path(model_root).expanduser() / "models"
+    model_root_path.mkdir(parents=True, exist_ok=True)
+
+    archive_path = cache_dir / asset.name
+    _download_with_retries(
+        asset.browser_download_url,
+        archive_path,
+        asset.name,
+        expected_size=asset.size,
+        progress=progress,
+    )
+
+    if asset.name.endswith(".zip"):
+        target_dir = model_root_path / asset.stem
+        if target_dir.exists():
+            shutil.rmtree(target_dir)
+        target_dir.mkdir(parents=True, exist_ok=True)
+        with zipfile.ZipFile(archive_path) as zf:
+            zf.extractall(target_dir)
+        if progress:
+            progress(1, 1, f"Extracted to {target_dir}")
+        return target_dir
+
+    target_dir = model_root_path / asset.stem
+    target_dir.mkdir(parents=True, exist_ok=True)
+    target_file = target_dir / asset.name
+    shutil.copy2(archive_path, target_file)
+    if progress:
+        progress(1, 1, f"Saved to {target_file}")
+    return target_file
diff --git a/python-package/insightface/gui/core/models.py b/python-package/insightface/gui/core/models.py
new file mode 100644
index 000000000..cbe3f8cf7
--- /dev/null
+++ b/python-package/insightface/gui/core/models.py
@@ -0,0 +1,119 @@
+"""Dataclasses used by the GUI core."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+
+
+@dataclass
+class FaceRecord:
+    bbox: List[float]
+    kps: Optional[List[List[float]]]
+    det_score: float
+    embedding: Optional[np.ndarray]
+    normed_embedding: Optional[np.ndarray]
+    face_id: Optional[str] = None
+    gender: Optional[int] = None
+    age: Optional[int] = None
+    quality_score: Optional[float] = None
+    quality_flags: List[str] = field(default_factory=list)
+    crop: Optional[np.ndarray] = None
+    source_path: Optional[str] = None
+    frame_index: Optional[int] = None
+    timestamp_ms: Optional[int] = None
+
+    def to_json_dict(self, include_embedding: bool = False) -> Dict[str, Any]:
+        data: Dict[str, Any] = {
+            "face_id": self.face_id,
+            "bbox": self.bbox,
+            "kps": self.kps,
+            "det_score": self.det_score,
+            "gender": self.gender,
+            "age": self.age,
+            "source_path": self.source_path,
+            "frame_index": self.frame_index,
+            "timestamp_ms": self.timestamp_ms,
+        }
+        if include_embedding:
+            data["embedding"] = self.embedding.tolist() if self.embedding is not None else None
+            data["normed_embedding"] = (
+                self.normed_embedding.tolist() if self.normed_embedding is not None else None
+            )
+        return data
+
+
+@dataclass
+class CompareResult:
+    similarity: float
+    threshold: float
+    decision: str
+    face_a: FaceRecord
+    face_b: FaceRecord
+    notes: List[str] = field(default_factory=list)
+
+    def to_json_dict(self) -> Dict[str, Any]:
+        return {
+            "similarity": self.similarity,
+            "threshold": self.threshold,
+            "decision": self.decision,
+            "face_a": self.face_a.to_json_dict(),
+            "face_b": self.face_b.to_json_dict(),
+            "notes": self.notes,
+        }
+
+
+@dataclass
+class SearchResult:
+    person_id: Optional[int]
+    person_name: str
+    sample_id: Optional[int]
+    similarity: float
+    status: str
+    crop_path: Optional[str] = None
+
+    def to_json_dict(self) -> Dict[str, Any]:
+        return {
+            "person_id": self.person_id,
+            "person_name": self.person_name,
+            "sample_id": self.sample_id,
+            "similarity": self.similarity,
+            "status": self.status,
+            "crop_path": self.crop_path,
+        }
+
+
+@dataclass
+class EvaluationResult:
+    scenario: str
+    model_name: str
+    provider: str
+    threshold: float
+    dataset_summary: Dict[str, Any]
+    metrics: Dict[str, Any]
+    errors: List[Dict[str, Any]]
+    latency: Dict[str, Any]
+    license_status: str
+    created_at: str
+    raw_results: List[Dict[str, Any]] = field(default_factory=list)
+    threshold_recommendation: Optional[float] = None
+    report_path: Optional[str] = None
+
+    def to_json_dict(self) -> Dict[str, Any]:
+        return {
+            "scenario": self.scenario,
+            "model_name": self.model_name,
+            "provider": self.provider,
+            "threshold": self.threshold,
+            "dataset_summary": self.dataset_summary,
+            "metrics": self.metrics,
+            "errors": self.errors,
+            "latency": self.latency,
+            "license_status": self.license_status,
+            "created_at": self.created_at,
+            "raw_results": self.raw_results,
+            "threshold_recommendation": self.threshold_recommendation,
+            "report_path": self.report_path,
+        }
diff --git a/python-package/insightface/gui/core/navigation.py b/python-package/insightface/gui/core/navigation.py
new file mode 100644
index 000000000..abc69beb8
--- /dev/null
+++ b/python-package/insightface/gui/core/navigation.py
@@ -0,0 +1,91 @@
+"""Mode-based navigation specification for the desktop GUI."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+
+
+class AppMode(str, Enum):
+    FACE_VERIFICATION = "face_verification"
+    ALBUM_MANAGEMENT = "album_management"
+    FACE_SWAP = "face_swap"
+    ENTERPRISE_EVALUATION = "enterprise_evaluation"
+
+
+@dataclass(frozen=True)
+class NavigationItem:
+    id: str
+    title: str
+    page_key: str
+    description: str = ""
+    icon: str | None = None
+    enabled: bool = True
+    coming_soon: bool = False
+
+
+@dataclass(frozen=True)
+class NavigationMode:
+    id: AppMode
+    title: str
+    description: str
+    items: tuple[NavigationItem, ...]
+
+
+NAVIGATION_MODES: dict[AppMode, NavigationMode] = {
+    AppMode.FACE_VERIFICATION: NavigationMode(
+        id=AppMode.FACE_VERIFICATION,
+        title="Face Recognition",
+        description="Upload a query face and recognize it against one gallery image or a local gallery.",
+        items=(
+            NavigationItem("recognition", "Face Recognition", "verification", "Run 1:1 compare or 1:N gallery search from Query and Gallery uploads."),
+        ),
+    ),
+    AppMode.ALBUM_MANAGEMENT: NavigationMode(
+        id=AppMode.ALBUM_MANAGEMENT,
+        title="Album Management",
+        description="Organize local photos by people, clusters, and searchable face groups.",
+        items=(
+            NavigationItem("album", "Album", "album", "Import, refresh, cluster, and review local album folders."),
+        ),
+    ),
+    AppMode.FACE_SWAP: NavigationMode(
+        id=AppMode.FACE_SWAP,
+        title="Face Swap",
+        description="Use a configured downloaded swap model for local image or video face swap.",
+        items=(
+            NavigationItem("face_swap", "Face Swap", "image_face_swap", "Source + Target = Result for image or video swap."),
+        ),
+    ),
+    AppMode.ENTERPRISE_EVALUATION: NavigationMode(
+        id=AppMode.ENTERPRISE_EVALUATION,
+        title="Enterprise Evaluation",
+        description="Run local 1:1 and 1:N model evaluations and export reports.",
+        items=(
+            NavigationItem("enterprise_evaluation", "Enterprise Evaluation", "enterprise_evaluation", "Run local 1:1 or 1:N identity-folder evaluation and export reports."),
+        ),
+    ),
+}
+
+
+MODE_LABELS: dict[str, AppMode] = {mode.title: mode.id for mode in NAVIGATION_MODES.values()}
+
+
+GLOBAL_PAGE_TITLES = {"Settings", "Model Settings", "Model Downloads", "License Center"}
+
+
+def mode_from_value(value: str | AppMode | None) -> AppMode:
+    if isinstance(value, AppMode):
+        return value
+    if value:
+        for mode in AppMode:
+            if value == mode.value:
+                return mode
+        for mode in NAVIGATION_MODES.values():
+            if value == mode.title:
+                return mode.id
+    return AppMode.FACE_VERIFICATION
+
+
+def last_page_attr(mode: AppMode) -> str:
+    return f"ui_last_page_{mode.value}"
diff --git a/python-package/insightface/gui/core/paths.py b/python-package/insightface/gui/core/paths.py
new file mode 100644
index 000000000..7435470fc
--- /dev/null
+++ b/python-package/insightface/gui/core/paths.py
@@ -0,0 +1,44 @@
+"""Workspace and filesystem helpers."""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from typing import Dict
+
+
+def default_workspace() -> Path:
+    return Path.home() / ".insightface" / "gui"
+
+
+def default_config_path() -> Path:
+    return default_workspace() / "config.json"
+
+
+def expand_path(value: str | os.PathLike[str] | None) -> Path:
+    if value is None or str(value).strip() == "":
+        return default_workspace()
+    return Path(os.path.expandvars(os.path.expanduser(str(value)))).resolve()
+
+
+def workspace_paths(workspace: str | os.PathLike[str] | None = None) -> Dict[str, Path]:
+    root = expand_path(workspace)
+    return {
+        "workspace": root,
+        "database": root / "insightface_gui.db",
+        "crops": root / "crops",
+        "exports": root / "exports",
+        "reports": root / "reports",
+        "logs": root / "logs",
+        "cache": root / "cache",
+    }
+
+
+def ensure_workspace(workspace: str | os.PathLike[str] | None = None) -> Dict[str, Path]:
+    paths = workspace_paths(workspace)
+    for path in paths.values():
+        if path.suffix:
+            path.parent.mkdir(parents=True, exist_ok=True)
+        else:
+            path.mkdir(parents=True, exist_ok=True)
+    return paths
diff --git a/python-package/insightface/gui/core/quality.py b/python-package/insightface/gui/core/quality.py
new file mode 100644
index 000000000..71966c411
--- /dev/null
+++ b/python-package/insightface/gui/core/quality.py
@@ -0,0 +1,103 @@
+"""Heuristic face quality scoring."""
+
+from __future__ import annotations
+
+from typing import Iterable, List, Optional, Tuple
+
+import numpy as np
+
+
+def _to_gray(image: np.ndarray) -> np.ndarray:
+    arr = np.asarray(image)
+    if arr.ndim == 2:
+        return arr.astype(np.float32)
+    if arr.shape[2] >= 3:
+        return (0.114 * arr[..., 0] + 0.587 * arr[..., 1] + 0.299 * arr[..., 2]).astype(np.float32)
+    return arr[..., 0].astype(np.float32)
+
+
+def blur_score(image: np.ndarray) -> float:
+    gray = _to_gray(image)
+    try:
+        import cv2
+
+        value = float(cv2.Laplacian(gray, cv2.CV_64F).var())
+    except Exception:
+        gx = np.diff(gray, axis=1)
+        gy = np.diff(gray, axis=0)
+        value = float(np.var(gx) + np.var(gy))
+    return max(0.0, value)
+
+
+def _bbox_area_ratio(bbox: Iterable[float], image_shape: Tuple[int, ...]) -> float:
+    x1, y1, x2, y2 = [float(v) for v in bbox]
+    face_area = max(0.0, x2 - x1) * max(0.0, y2 - y1)
+    image_area = float(max(1, image_shape[0] * image_shape[1]))
+    return face_area / image_area
+
+
+def _frontal_score(kps: Optional[Iterable[Iterable[float]]]) -> tuple[float, List[str]]:
+    if kps is None:
+        return 0.55, ["landmarks_missing"]
+    pts = np.asarray(list(kps), dtype=np.float32)
+    if pts.shape[0] < 5:
+        return 0.55, ["landmarks_incomplete"]
+    left_eye, right_eye, nose, left_mouth, right_mouth = pts[:5]
+    eye_mid = (left_eye + right_eye) / 2.0
+    mouth_mid = (left_mouth + right_mouth) / 2.0
+    eye_distance = max(float(np.linalg.norm(left_eye - right_eye)), 1.0)
+    nose_offset = abs(float(nose[0] - eye_mid[0])) / eye_distance
+    vertical_span = max(abs(float(mouth_mid[1] - eye_mid[1])), 1.0)
+    nose_vertical = abs(float(nose[1] - (eye_mid[1] + mouth_mid[1]) / 2.0)) / vertical_span
+    score = 1.0 - min(0.7, nose_offset) - min(0.3, nose_vertical * 0.3)
+    flags: List[str] = []
+    if nose_offset > 0.35:
+        flags.append("side_face")
+    return max(0.0, min(1.0, score)), flags
+
+
+def score_face(
+    image: np.ndarray,
+    bbox: Iterable[float],
+    kps: Optional[Iterable[Iterable[float]]] = None,
+    det_score: float = 1.0,
+) -> tuple[float, List[str]]:
+    flags: List[str] = []
+    det_component = max(0.0, min(1.0, float(det_score)))
+    if det_component < 0.55:
+        flags.append("low_detection_score")
+
+    area_ratio = _bbox_area_ratio(bbox, image.shape)
+    if area_ratio < 0.01:
+        flags.append("face_too_small")
+    size_component = max(0.0, min(1.0, area_ratio / 0.08))
+
+    x1, y1, x2, y2 = [int(round(float(v))) for v in bbox]
+    x1, y1 = max(0, x1), max(0, y1)
+    x2, y2 = min(image.shape[1], x2), min(image.shape[0], y2)
+    crop = image[y1:y2, x1:x2] if x2 > x1 and y2 > y1 else image
+
+    blur = blur_score(crop)
+    if blur < 45:
+        flags.append("blurry")
+    blur_component = max(0.0, min(1.0, blur / 250.0))
+
+    gray = _to_gray(crop)
+    brightness = float(np.mean(gray)) if gray.size else 128.0
+    exposure_component = 1.0 - min(1.0, abs(brightness - 128.0) / 128.0)
+    if brightness < 45:
+        flags.append("too_dark")
+    elif brightness > 220:
+        flags.append("too_bright")
+
+    frontal_component, frontal_flags = _frontal_score(kps)
+    flags.extend(frontal_flags)
+
+    score = (
+        0.28 * det_component
+        + 0.22 * size_component
+        + 0.22 * blur_component
+        + 0.18 * frontal_component
+        + 0.10 * exposure_component
+    )
+    return max(0.0, min(1.0, float(score))), sorted(set(flags))
diff --git a/python-package/insightface/gui/core/recognition.py b/python-package/insightface/gui/core/recognition.py
new file mode 100644
index 000000000..626df1433
--- /dev/null
+++ b/python-package/insightface/gui/core/recognition.py
@@ -0,0 +1,118 @@
+"""Embedding normalization, comparison, and gallery search."""
+
+from __future__ import annotations
+
+from collections import defaultdict
+from typing import Any, Dict, Iterable, List, Optional
+
+import numpy as np
+
+from .constants import DEFAULT_THRESHOLD
+from .models import SearchResult
+
+
+def normalize_embedding(embedding: np.ndarray | Iterable[float] | None) -> Optional[np.ndarray]:
+    if embedding is None:
+        return None
+    arr = np.asarray(embedding, dtype=np.float32).reshape(-1)
+    norm = float(np.linalg.norm(arr))
+    if norm <= 1e-12:
+        return arr
+    return arr / norm
+
+
+def cosine_similarity(a: np.ndarray | Iterable[float] | None, b: np.ndarray | Iterable[float] | None) -> float:
+    na = normalize_embedding(a)
+    nb = normalize_embedding(b)
+    if na is None or nb is None or na.shape != nb.shape:
+        return 0.0
+    return float(np.dot(na, nb))
+
+
+def compare_embeddings(a: np.ndarray, b: np.ndarray, threshold: float = DEFAULT_THRESHOLD) -> Dict[str, Any]:
+    similarity = cosine_similarity(a, b)
+    if similarity >= threshold:
+        decision = "Same Person"
+    elif similarity >= threshold - 0.05:
+        decision = "Uncertain"
+    else:
+        decision = "Different Person"
+    return {"similarity": similarity, "threshold": threshold, "decision": decision}
+
+
+def search_gallery(
+    query_embedding: np.ndarray,
+    gallery_embeddings: Iterable[Dict[str, Any]],
+    top_k: int = 5,
+    threshold: float = DEFAULT_THRESHOLD,
+) -> List[SearchResult]:
+    query = normalize_embedding(query_embedding)
+    if query is None:
+        return []
+    best_by_person: Dict[Any, SearchResult] = {}
+    for item in gallery_embeddings:
+        emb = item.get("embedding")
+        if emb is None:
+            continue
+        emb_norm = normalize_embedding(emb)
+        if emb_norm is None or emb_norm.shape != query.shape:
+            continue
+        similarity = float(np.dot(query, emb_norm))
+        person_id = item.get("person_id")
+        status = "matched" if similarity >= threshold else "below_threshold"
+        result = SearchResult(
+            person_id=person_id,
+            person_name=item.get("person_name") or "Unknown",
+            sample_id=item.get("sample_id"),
+            similarity=similarity,
+            status=status,
+            crop_path=item.get("crop_path"),
+        )
+        existing = best_by_person.get(person_id)
+        if existing is None or result.similarity > existing.similarity:
+            best_by_person[person_id] = result
+    results = sorted(best_by_person.values(), key=lambda row: row.similarity, reverse=True)
+    return results[: max(1, int(top_k))]
+
+
+def aggregate_person_embeddings(face_samples: Iterable[Dict[str, Any]], method: str = "mean") -> Dict[int, np.ndarray]:
+    grouped: Dict[int, List[np.ndarray]] = defaultdict(list)
+    for sample in face_samples:
+        person_id = sample.get("person_id")
+        emb = normalize_embedding(sample.get("embedding"))
+        if person_id is None or emb is None:
+            continue
+        grouped[int(person_id)].append(emb)
+    aggregated: Dict[int, np.ndarray] = {}
+    for person_id, embeddings in grouped.items():
+        stack = np.vstack(embeddings)
+        if method == "max":
+            aggregated[person_id] = stack[0]
+        else:
+            aggregated[person_id] = normalize_embedding(np.mean(stack, axis=0))  # type: ignore[assignment]
+    return aggregated
+
+
+def identify_face(
+    query_embedding: np.ndarray,
+    people_gallery: Iterable[Dict[str, Any]],
+    threshold: float = DEFAULT_THRESHOLD,
+    top_k: int = 5,
+) -> List[SearchResult]:
+    results = search_gallery(query_embedding, people_gallery, top_k=top_k, threshold=threshold)
+    if not results:
+        return [
+            SearchResult(
+                person_id=None,
+                person_name="Unknown",
+                sample_id=None,
+                similarity=0.0,
+                status="unknown",
+            )
+        ]
+    for result in results:
+        if result.similarity < threshold:
+            result.status = "unknown"
+            if result.person_name == "":
+                result.person_name = "Unknown"
+    return results
diff --git a/python-package/insightface/gui/core/reporting.py b/python-package/insightface/gui/core/reporting.py
new file mode 100644
index 000000000..d1b83fa08
--- /dev/null
+++ b/python-package/insightface/gui/core/reporting.py
@@ -0,0 +1,760 @@
+"""Enterprise evaluation report generation."""
+
+from __future__ import annotations
+
+import html
+import sys
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Sequence, Tuple
+
+from .i18n import effective_language
+from .models import EvaluationResult
+from .utils import safe_json_dumps, timestamp_for_filename
+
+
+_REPORT_TRANSLATIONS: dict[str, dict[str, str]] = {
+    "en": {
+        "title": "InsightFace Enterprise Evaluation Report",
+        "local_notice": "All processing is local. No images, embeddings, or reports are uploaded automatically.",
+        "executive_summary": "Executive Summary",
+        "evaluation_scenario": "Evaluation Scenario",
+        "dataset_summary": "Dataset Summary",
+        "model_runtime": "Model and Runtime",
+        "license_status": "License Status",
+        "metrics": "Metrics",
+        "threshold_recommendation": "Threshold Recommendation",
+        "error_analysis": "Error Analysis",
+        "latency_hardware": "Latency and Hardware",
+        "deployment_considerations": "Deployment Considerations",
+        "responsible_use": "Responsible Use and Compliance Notice",
+        "commercial_next_steps": "Commercial Licensing Next Steps",
+        "appendix_raw": "Appendix: Raw Results",
+        "scenario": "Scenario",
+        "created_at": "Created at",
+        "license": "License",
+        "model": "Model",
+        "provider": "Provider",
+        "threshold": "Threshold",
+        "field": "Field",
+        "value": "Value",
+        "metric": "Metric",
+        "not_available": "Not available",
+        "no_errors": "No failed cases were recorded.",
+        "raw_note": "Showing the first {count} raw result rows.",
+        "commercial_notice": "This evaluation may use research or non-commercial model files. Production or commercial deployment requires an appropriate commercial model license.",
+        "responsible_notice": "Users are responsible for consent, privacy, retention, and compliance with applicable biometric regulations.",
+        "legal_notice": "This report does not provide legal advice.",
+        "contact_notice": "For report interpretation, commercial model licensing, private model evaluation, SDK/API access, SLA, or custom training, contact InsightFace: https://www.insightface.ai/contact",
+        "generated_by": "Generated by InsightFace Evaluation Studio",
+    },
+    "zh": {
+        "title": "InsightFace 企业评测报告",
+        "local_notice": "所有处理均在本地完成。图片、特征和报告不会被自动上传。",
+        "executive_summary": "执行摘要",
+        "evaluation_scenario": "评测场景",
+        "dataset_summary": "数据集摘要",
+        "model_runtime": "模型与运行环境",
+        "license_status": "授权状态",
+        "metrics": "指标",
+        "threshold_recommendation": "阈值建议",
+        "error_analysis": "错误分析",
+        "latency_hardware": "延迟与硬件",
+        "deployment_considerations": "部署考虑",
+        "responsible_use": "负责任使用与合规提示",
+        "commercial_next_steps": "商业授权下一步",
+        "appendix_raw": "附录：原始结果",
+        "scenario": "场景",
+        "created_at": "生成时间",
+        "license": "授权",
+        "model": "模型",
+        "provider": "运行后端",
+        "threshold": "阈值",
+        "field": "字段",
+        "value": "值",
+        "metric": "指标",
+        "not_available": "不可用",
+        "no_errors": "未记录失败样例。",
+        "raw_note": "仅显示前 {count} 条原始结果。",
+        "commercial_notice": "本次评测可能使用研究或非商业模型文件。生产或商业部署需要相应的商业模型授权。",
+        "responsible_notice": "用户需自行负责同意、隐私、数据留存以及适用生物识别法规的合规要求。",
+        "legal_notice": "本报告不构成法律建议。",
+        "contact_notice": "如需报告解读、商业模型授权、私有模型评测、SDK/API、SLA 或定制训练，请联系 InsightFace：https://www.insightface.ai/contact",
+        "generated_by": "由 InsightFace Evaluation Studio 生成",
+    },
+    "ja": {
+        "title": "InsightFace エンタープライズ評価レポート",
+        "local_notice": "すべての処理はローカルで実行されます。画像、特徴量、レポートは自動アップロードされません。",
+        "executive_summary": "エグゼクティブサマリー",
+        "evaluation_scenario": "評価シナリオ",
+        "dataset_summary": "データセット概要",
+        "model_runtime": "モデルとランタイム",
+        "license_status": "ライセンス状況",
+        "metrics": "指標",
+        "threshold_recommendation": "しきい値推奨",
+        "error_analysis": "エラー分析",
+        "latency_hardware": "レイテンシとハードウェア",
+        "deployment_considerations": "導入時の考慮事項",
+        "responsible_use": "責任ある利用とコンプライアンス注意",
+        "commercial_next_steps": "商用ライセンスの次のステップ",
+        "appendix_raw": "付録：生データ結果",
+        "scenario": "シナリオ",
+        "created_at": "作成日時",
+        "license": "ライセンス",
+        "model": "モデル",
+        "provider": "プロバイダ",
+        "threshold": "しきい値",
+        "field": "項目",
+        "value": "値",
+        "metric": "指標",
+        "not_available": "利用不可",
+        "no_errors": "失敗ケースは記録されていません。",
+        "raw_note": "最初の {count} 件の生データ結果を表示しています。",
+        "commercial_notice": "この評価では研究用または非商用モデルファイルを使用している可能性があります。本番または商用展開には適切な商用モデルライセンスが必要です。",
+        "responsible_notice": "同意、プライバシー、保存期間、適用される生体情報規制への準拠は利用者の責任です。",
+        "legal_notice": "本レポートは法的助言ではありません。",
+        "contact_notice": "レポート内容の解釈、商用モデルライセンス、非公開モデル評価、SDK/API、SLA、カスタム学習については InsightFace までお問い合わせください: https://www.insightface.ai/contact",
+        "generated_by": "InsightFace Evaluation Studio により生成",
+    },
+    "ko": {
+        "title": "InsightFace 엔터프라이즈 평가 보고서",
+        "local_notice": "모든 처리는 로컬에서 수행됩니다. 이미지, 임베딩, 보고서는 자동으로 업로드되지 않습니다.",
+        "executive_summary": "요약",
+        "evaluation_scenario": "평가 시나리오",
+        "dataset_summary": "데이터셋 요약",
+        "model_runtime": "모델 및 런타임",
+        "license_status": "라이선스 상태",
+        "metrics": "지표",
+        "threshold_recommendation": "임계값 권장",
+        "error_analysis": "오류 분석",
+        "latency_hardware": "지연 시간 및 하드웨어",
+        "deployment_considerations": "배포 고려 사항",
+        "responsible_use": "책임 있는 사용 및 컴플라이언스 고지",
+        "commercial_next_steps": "상용 라이선스 다음 단계",
+        "appendix_raw": "부록: 원시 결과",
+        "scenario": "시나리오",
+        "created_at": "생성 시간",
+        "license": "라이선스",
+        "model": "모델",
+        "provider": "제공자",
+        "threshold": "임계값",
+        "field": "항목",
+        "value": "값",
+        "metric": "지표",
+        "not_available": "사용 불가",
+        "no_errors": "기록된 실패 사례가 없습니다.",
+        "raw_note": "처음 {count}개의 원시 결과 행을 표시합니다.",
+        "commercial_notice": "이 평가는 연구용 또는 비상업용 모델 파일을 사용할 수 있습니다. 프로덕션 또는 상업적 배포에는 적절한 상용 모델 라이선스가 필요합니다.",
+        "responsible_notice": "동의, 개인정보, 보존, 적용 가능한 생체정보 규정 준수는 사용자 책임입니다.",
+        "legal_notice": "이 보고서는 법률 자문을 제공하지 않습니다.",
+        "contact_notice": "보고서 해석, 상용 모델 라이선스, 비공개 모델 평가, SDK/API, SLA 또는 맞춤 학습은 InsightFace에 문의하십시오: https://www.insightface.ai/contact",
+        "generated_by": "InsightFace Evaluation Studio에서 생성",
+    },
+    "es": {
+        "title": "Informe de evaluación empresarial de InsightFace",
+        "local_notice": "Todo el procesamiento es local. Las imágenes, embeddings e informes no se suben automáticamente.",
+        "executive_summary": "Resumen ejecutivo",
+        "evaluation_scenario": "Escenario de evaluación",
+        "dataset_summary": "Resumen del dataset",
+        "model_runtime": "Modelo y runtime",
+        "license_status": "Estado de licencia",
+        "metrics": "Métricas",
+        "threshold_recommendation": "Recomendación de umbral",
+        "error_analysis": "Análisis de errores",
+        "latency_hardware": "Latencia y hardware",
+        "deployment_considerations": "Consideraciones de despliegue",
+        "responsible_use": "Uso responsable y aviso de cumplimiento",
+        "commercial_next_steps": "Próximos pasos de licencia comercial",
+        "appendix_raw": "Apéndice: resultados sin procesar",
+        "scenario": "Escenario",
+        "created_at": "Creado el",
+        "license": "Licencia",
+        "model": "Modelo",
+        "provider": "Proveedor",
+        "threshold": "Umbral",
+        "field": "Campo",
+        "value": "Valor",
+        "metric": "Métrica",
+        "not_available": "No disponible",
+        "no_errors": "No se registraron casos fallidos.",
+        "raw_note": "Se muestran las primeras {count} filas de resultados sin procesar.",
+        "commercial_notice": "Esta evaluación puede usar archivos de modelo de investigación o no comerciales. La producción o despliegue comercial requiere una licencia comercial de modelo adecuada.",
+        "responsible_notice": "Los usuarios son responsables del consentimiento, privacidad, retención y cumplimiento de la normativa biométrica aplicable.",
+        "legal_notice": "Este informe no proporciona asesoramiento legal.",
+        "contact_notice": "Para interpretar el informe, obtener licencias comerciales de modelos, evaluación privada, SDK/API, SLA o entrenamiento personalizado, contacte con InsightFace: https://www.insightface.ai/contact",
+        "generated_by": "Generado por InsightFace Evaluation Studio",
+    },
+    "fr": {
+        "title": "Rapport d’évaluation entreprise InsightFace",
+        "local_notice": "Tout le traitement est local. Les images, embeddings et rapports ne sont pas téléversés automatiquement.",
+        "executive_summary": "Synthèse",
+        "evaluation_scenario": "Scénario d’évaluation",
+        "dataset_summary": "Résumé du jeu de données",
+        "model_runtime": "Modèle et runtime",
+        "license_status": "Statut de licence",
+        "metrics": "Métriques",
+        "threshold_recommendation": "Recommandation de seuil",
+        "error_analysis": "Analyse des erreurs",
+        "latency_hardware": "Latence et matériel",
+        "deployment_considerations": "Considérations de déploiement",
+        "responsible_use": "Usage responsable et conformité",
+        "commercial_next_steps": "Prochaines étapes de licence commerciale",
+        "appendix_raw": "Annexe : résultats bruts",
+        "scenario": "Scénario",
+        "created_at": "Créé le",
+        "license": "Licence",
+        "model": "Modèle",
+        "provider": "Fournisseur",
+        "threshold": "Seuil",
+        "field": "Champ",
+        "value": "Valeur",
+        "metric": "Métrique",
+        "not_available": "Non disponible",
+        "no_errors": "Aucun cas d’échec n’a été enregistré.",
+        "raw_note": "Affichage des {count} premières lignes de résultats bruts.",
+        "commercial_notice": "Cette évaluation peut utiliser des fichiers de modèle de recherche ou non commerciaux. Un déploiement en production ou commercial nécessite une licence commerciale appropriée.",
+        "responsible_notice": "Les utilisateurs sont responsables du consentement, de la confidentialité, de la conservation et du respect des réglementations biométriques applicables.",
+        "legal_notice": "Ce rapport ne fournit pas de conseil juridique.",
+        "contact_notice": "Pour l’interprétation du rapport, les licences commerciales, l’évaluation privée, l’accès SDK/API, les SLA ou l’entraînement personnalisé, contactez InsightFace : https://www.insightface.ai/contact",
+        "generated_by": "Généré par InsightFace Evaluation Studio",
+    },
+    "de": {
+        "title": "InsightFace Enterprise-Evaluierungsbericht",
+        "local_notice": "Die gesamte Verarbeitung erfolgt lokal. Bilder, Embeddings und Berichte werden nicht automatisch hochgeladen.",
+        "executive_summary": "Management Summary",
+        "evaluation_scenario": "Evaluierungsszenario",
+        "dataset_summary": "Datensatzübersicht",
+        "model_runtime": "Modell und Laufzeit",
+        "license_status": "Lizenzstatus",
+        "metrics": "Metriken",
+        "threshold_recommendation": "Schwellenwertempfehlung",
+        "error_analysis": "Fehleranalyse",
+        "latency_hardware": "Latenz und Hardware",
+        "deployment_considerations": "Bereitstellungsaspekte",
+        "responsible_use": "Verantwortungsvolle Nutzung und Compliance-Hinweis",
+        "commercial_next_steps": "Nächste Schritte zur kommerziellen Lizenzierung",
+        "appendix_raw": "Anhang: Rohresultate",
+        "scenario": "Szenario",
+        "created_at": "Erstellt am",
+        "license": "Lizenz",
+        "model": "Modell",
+        "provider": "Provider",
+        "threshold": "Schwellenwert",
+        "field": "Feld",
+        "value": "Wert",
+        "metric": "Metrik",
+        "not_available": "Nicht verfügbar",
+        "no_errors": "Es wurden keine Fehlerfälle erfasst.",
+        "raw_note": "Angezeigt werden die ersten {count} Rohresultate.",
+        "commercial_notice": "Diese Evaluierung kann Forschungs- oder nicht-kommerzielle Modelldateien verwenden. Produktion oder kommerzielle Bereitstellung erfordert eine geeignete kommerzielle Modelllizenz.",
+        "responsible_notice": "Nutzer sind für Einwilligung, Datenschutz, Aufbewahrung und die Einhaltung geltender biometrischer Vorschriften verantwortlich.",
+        "legal_notice": "Dieser Bericht stellt keine Rechtsberatung dar.",
+        "contact_notice": "Für die Interpretation des Berichts, kommerzielle Modelllizenzen, private Modellevaluierung, SDK/API-Zugang, SLA oder individuelles Training kontaktieren Sie InsightFace: https://www.insightface.ai/contact",
+        "generated_by": "Erstellt mit InsightFace Evaluation Studio",
+    },
+    "pt": {
+        "title": "Relatório de avaliação empresarial InsightFace",
+        "local_notice": "Todo o processamento é local. Imagens, embeddings e relatórios não são enviados automaticamente.",
+        "executive_summary": "Resumo executivo",
+        "evaluation_scenario": "Cenário de avaliação",
+        "dataset_summary": "Resumo do dataset",
+        "model_runtime": "Modelo e runtime",
+        "license_status": "Estado da licença",
+        "metrics": "Métricas",
+        "threshold_recommendation": "Recomendação de limiar",
+        "error_analysis": "Análise de erros",
+        "latency_hardware": "Latência e hardware",
+        "deployment_considerations": "Considerações de implantação",
+        "responsible_use": "Uso responsável e aviso de conformidade",
+        "commercial_next_steps": "Próximos passos de licenciamento comercial",
+        "appendix_raw": "Apêndice: resultados brutos",
+        "scenario": "Cenário",
+        "created_at": "Criado em",
+        "license": "Licença",
+        "model": "Modelo",
+        "provider": "Fornecedor",
+        "threshold": "Limiar",
+        "field": "Campo",
+        "value": "Valor",
+        "metric": "Métrica",
+        "not_available": "Não disponível",
+        "no_errors": "Nenhum caso de falha foi registado.",
+        "raw_note": "A mostrar as primeiras {count} linhas de resultados brutos.",
+        "commercial_notice": "Esta avaliação pode usar ficheiros de modelo de investigação ou não comerciais. Produção ou implantação comercial requer uma licença comercial de modelo adequada.",
+        "responsible_notice": "Os utilizadores são responsáveis por consentimento, privacidade, retenção e conformidade com regulamentos biométricos aplicáveis.",
+        "legal_notice": "Este relatório não fornece aconselhamento jurídico.",
+        "contact_notice": "Para interpretação do relatório, licenciamento comercial, avaliação privada, SDK/API, SLA ou treino personalizado, contacte a InsightFace: https://www.insightface.ai/contact",
+        "generated_by": "Gerado por InsightFace Evaluation Studio",
+    },
+    "ru": {
+        "title": "Корпоративный отчёт оценки InsightFace",
+        "local_notice": "Вся обработка выполняется локально. Изображения, эмбеддинги и отчёты не загружаются автоматически.",
+        "executive_summary": "Резюме",
+        "evaluation_scenario": "Сценарий оценки",
+        "dataset_summary": "Сводка датасета",
+        "model_runtime": "Модель и среда выполнения",
+        "license_status": "Статус лицензии",
+        "metrics": "Метрики",
+        "threshold_recommendation": "Рекомендация порога",
+        "error_analysis": "Анализ ошибок",
+        "latency_hardware": "Задержка и оборудование",
+        "deployment_considerations": "Рекомендации по внедрению",
+        "responsible_use": "Ответственное использование и соответствие требованиям",
+        "commercial_next_steps": "Следующие шаги коммерческого лицензирования",
+        "appendix_raw": "Приложение: сырые результаты",
+        "scenario": "Сценарий",
+        "created_at": "Создано",
+        "license": "Лицензия",
+        "model": "Модель",
+        "provider": "Провайдер",
+        "threshold": "Порог",
+        "field": "Поле",
+        "value": "Значение",
+        "metric": "Метрика",
+        "not_available": "Недоступно",
+        "no_errors": "Случаи ошибок не зафиксированы.",
+        "raw_note": "Показаны первые {count} строк сырых результатов.",
+        "commercial_notice": "Эта оценка может использовать исследовательские или некоммерческие файлы моделей. Производственное или коммерческое внедрение требует соответствующей коммерческой лицензии модели.",
+        "responsible_notice": "Пользователи отвечают за согласие, конфиденциальность, хранение и соблюдение применимых биометрических норм.",
+        "legal_notice": "Этот отчёт не является юридической консультацией.",
+        "contact_notice": "Для разбора отчета, коммерческого лицензирования модели, приватной оценки, доступа к SDK/API, SLA или индивидуального обучения свяжитесь с InsightFace: https://www.insightface.ai/contact",
+        "generated_by": "Создано InsightFace Evaluation Studio",
+    },
+}
+
+
+def _rt(key: str, language: str | None = None) -> str:
+    lang = effective_language(language)
+    return _REPORT_TRANSLATIONS.get(lang, _REPORT_TRANSLATIONS["en"]).get(
+        key, _REPORT_TRANSLATIONS["en"].get(key, key)
+    )
+
+
+def _value_to_text(value: Any) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, float):
+        return f"{value:.6f}"
+    if isinstance(value, (dict, list, tuple)):
+        return safe_json_dumps(value)
+    return str(value)
+
+
+def _metric_lines(metrics: Dict[str, Any]) -> Iterable[str]:
+    for key, value in metrics.items():
+        yield f"- **{key}**: {_value_to_text(value)}"
+
+
+def _markdown_escape(value: Any) -> str:
+    return _value_to_text(value).replace("|", "\\|").replace("\n", "<br>")
+
+
+def _markdown_table(rows: Sequence[Tuple[str, Any]], language: str | None = None) -> List[str]:
+    lines = [
+        f"| {_rt('field', language)} | {_rt('value', language)} |",
+        "| --- | --- |",
+    ]
+    for key, value in rows:
+        lines.append(f"| {_markdown_escape(key)} | {_markdown_escape(value)} |")
+    return lines
+
+
+def _summary_rows(result: EvaluationResult, language: str | None) -> List[Tuple[str, Any]]:
+    threshold_recommendation = (
+        f"{result.threshold_recommendation:.6f}"
+        if result.threshold_recommendation is not None
+        else _rt("not_available", language)
+    )
+    return [
+        (_rt("scenario", language), result.scenario),
+        (_rt("created_at", language), result.created_at),
+        (_rt("model", language), result.model_name),
+        (_rt("provider", language), result.provider),
+        (_rt("license", language), result.license_status),
+        (_rt("threshold_recommendation", language), threshold_recommendation),
+    ]
+
+
+def generate_markdown_report(result: EvaluationResult, language: str | None = None) -> str:
+    raw_rows = result.raw_results[:50]
+    raw_block = safe_json_dumps(raw_rows)
+    lines = [
+        f"# {_rt('title', language)}",
+        "",
+        f"> {_rt('local_notice', language)}",
+        "",
+        f"## 1. {_rt('executive_summary', language)}",
+        *_markdown_table(_summary_rows(result, language), language),
+        "",
+        f"## 2. {_rt('evaluation_scenario', language)}",
+        result.scenario,
+        "",
+        f"## 3. {_rt('dataset_summary', language)}",
+        *_markdown_table(list(result.dataset_summary.items()), language),
+        "",
+        f"## 4. {_rt('model_runtime', language)}",
+        *_markdown_table(
+            [
+                (_rt("model", language), result.model_name),
+                (_rt("provider", language), result.provider),
+                (_rt("threshold", language), f"{result.threshold:.6f}"),
+            ],
+            language,
+        ),
+        "",
+        f"## 5. {_rt('license_status', language)}",
+        result.license_status,
+        "",
+        f"## 6. {_rt('metrics', language)}",
+        *list(_metric_lines(result.metrics)),
+        "",
+        f"## 7. {_rt('threshold_recommendation', language)}",
+        _summary_rows(result, language)[-1][1],
+        "",
+        f"## 8. {_rt('error_analysis', language)}",
+        safe_json_dumps(result.errors[:50]) if result.errors else _rt("no_errors", language),
+        "",
+        f"## 9. {_rt('latency_hardware', language)}",
+        *_markdown_table(list(result.latency.items()), language),
+        "",
+        f"## 10. {_rt('deployment_considerations', language)}",
+        _rt("commercial_notice", language),
+        "",
+        f"## 11. {_rt('responsible_use', language)}",
+        _rt("responsible_notice", language),
+        "",
+        _rt("legal_notice", language),
+        "",
+        f"## 12. {_rt('commercial_next_steps', language)}",
+        _rt("contact_notice", language),
+        "",
+        f"## 13. {_rt('appendix_raw', language)}",
+        _rt("raw_note", language).format(count=len(raw_rows)),
+        "",
+        "```json",
+        raw_block,
+        "```",
+        "",
+    ]
+    return "\n".join(str(line) for line in lines)
+
+
+def generate_html_report(result: EvaluationResult, language: str | None = None) -> str:
+    markdown = generate_markdown_report(result, language)
+    lines = []
+    in_code = False
+    for line in markdown.splitlines():
+        if line.startswith("```"):
+            lines.append("</pre>" if in_code else "<pre>")
+            in_code = not in_code
+        elif in_code:
+            lines.append(html.escape(line))
+        elif line.startswith("# "):
+            lines.append(f"<h1>{html.escape(line[2:])}</h1>")
+        elif line.startswith("## "):
+            lines.append(f"<h2>{html.escape(line[3:])}</h2>")
+        elif line.startswith("> "):
+            lines.append(f"<blockquote>{html.escape(line[2:])}</blockquote>")
+        elif line.startswith("| "):
+            lines.append(f"<p class='table-line'>{html.escape(line)}</p>")
+        elif line.startswith("- "):
+            lines.append(f"<p>{html.escape(line)}</p>")
+        elif line.strip():
+            lines.append(f"<p>{html.escape(line)}</p>")
+        else:
+            lines.append("")
+    return (
+        "<!doctype html><html><head><meta charset='utf-8'>"
+        f"<title>{html.escape(_rt('title', language))}</title>"
+        "<style>body{font-family:-apple-system,BlinkMacSystemFont,Segoe UI,sans-serif;"
+        "max-width:1040px;margin:32px auto;line-height:1.55;color:#1f2937}"
+        "h1{font-size:30px;color:#0f172a}h2{margin-top:28px;color:#1e3a8a}"
+        "blockquote{background:#eaf3ff;border-left:4px solid #2563eb;padding:12px 16px}"
+        "pre{background:#f3f4f6;padding:16px;overflow:auto;border-radius:8px}"
+        ".table-line{font-family:ui-monospace,SFMono-Regular,Menlo,monospace}</style></head><body>"
+        + "\n".join(lines)
+        + "</body></html>"
+    )
+
+
+def _candidate_font_paths() -> List[Path]:
+    py_site = Path(sys.prefix) / "lib" / f"python{sys.version_info.major}.{sys.version_info.minor}" / "site-packages"
+    win_site = Path(sys.prefix) / "Lib" / "site-packages"
+    candidates = [
+        py_site / "matplotlib" / "mpl-data" / "fonts" / "ttf" / "DejaVuSans.ttf",
+        py_site / "matplotlib" / "mpl-data" / "fonts" / "ttf" / "DejaVuSans-Bold.ttf",
+        win_site / "matplotlib" / "mpl-data" / "fonts" / "ttf" / "DejaVuSans.ttf",
+        win_site / "matplotlib" / "mpl-data" / "fonts" / "ttf" / "DejaVuSans-Bold.ttf",
+        Path("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf"),
+        Path("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf"),
+        Path("/usr/share/fonts/truetype/liberation2/LiberationSans-Regular.ttf"),
+        Path("/usr/share/fonts/truetype/liberation2/LiberationSans-Bold.ttf"),
+        Path("/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc"),
+        Path("/usr/share/fonts/opentype/noto/NotoSansCJK-Bold.ttc"),
+        Path("/Library/Fonts/Arial Unicode.ttf"),
+        Path("/Library/Fonts/Arial.ttf"),
+        Path("/System/Library/Fonts/Supplemental/Arial.ttf"),
+        Path("/System/Library/Fonts/Supplemental/Arial Bold.ttf"),
+        Path("C:/Windows/Fonts/arial.ttf"),
+        Path("C:/Windows/Fonts/arialbd.ttf"),
+        Path("C:/Windows/Fonts/segoeui.ttf"),
+        Path("C:/Windows/Fonts/seguisb.ttf"),
+    ]
+    return [path for path in candidates if path.exists()]
+
+
+def _register_pdf_fonts(language: str | None = None) -> Tuple[str, str]:
+    from reportlab.pdfbase import pdfmetrics
+
+    lang = effective_language(language)
+    if lang in {"zh", "ja", "ko"}:
+        from reportlab.pdfbase.cidfonts import UnicodeCIDFont
+
+        cid_name = {"zh": "STSong-Light", "ja": "HeiseiMin-W3", "ko": "HYSMyeongJo-Medium"}[lang]
+        try:
+            pdfmetrics.registerFont(UnicodeCIDFont(cid_name))
+            return cid_name, cid_name
+        except Exception:
+            pass
+
+    regular_path = None
+    bold_path = None
+    for path in _candidate_font_paths():
+        lower = path.name.lower()
+        if regular_path is None and lower in {"dejavusans.ttf", "arial.ttf", "segoeui.ttf", "notosans-regular.ttf"}:
+            regular_path = path
+        if bold_path is None and lower in {"dejavusans-bold.ttf", "arialbd.ttf", "arial bold.ttf", "seguisb.ttf", "notosans-bold.ttf"}:
+            bold_path = path
+        if regular_path and bold_path:
+            break
+
+    if regular_path:
+        try:
+            from reportlab.pdfbase.ttfonts import TTFont
+
+            pdfmetrics.registerFont(TTFont("InsightFaceReport", str(regular_path)))
+            if bold_path:
+                pdfmetrics.registerFont(TTFont("InsightFaceReportBold", str(bold_path)))
+                return "InsightFaceReport", "InsightFaceReportBold"
+            return "InsightFaceReport", "InsightFaceReport"
+        except Exception:
+            pass
+    return "Helvetica", "Helvetica-Bold"
+
+
+def _pdf_paragraph(text: Any, style):
+    from reportlab.platypus import Paragraph
+
+    escaped = html.escape(_value_to_text(text)).replace("\n", "<br/>")
+    return Paragraph(escaped, style)
+
+
+def _pdf_table(rows: Sequence[Tuple[Any, Any]], styles, language: str | None = None, widths=None):
+    from reportlab.lib import colors
+    from reportlab.platypus import Table, TableStyle
+
+    data = [
+        [
+            _pdf_paragraph(_rt("field", language), styles["TableHeader"]),
+            _pdf_paragraph(_rt("value", language), styles["TableHeader"]),
+        ]
+    ]
+    for key, value in rows:
+        data.append([_pdf_paragraph(key, styles["TableCellStrong"]), _pdf_paragraph(value, styles["TableCell"])])
+    table = Table(data, colWidths=widths or [145, 340], repeatRows=1, hAlign="LEFT")
+    table.setStyle(
+        TableStyle(
+            [
+                ("BACKGROUND", (0, 0), (-1, 0), colors.HexColor("#dbeafe")),
+                ("TEXTCOLOR", (0, 0), (-1, 0), colors.HexColor("#0f172a")),
+                ("GRID", (0, 0), (-1, -1), 0.35, colors.HexColor("#cbd5e1")),
+                ("BACKGROUND", (0, 1), (-1, -1), colors.HexColor("#f8fafc")),
+                ("VALIGN", (0, 0), (-1, -1), "TOP"),
+                ("LEFTPADDING", (0, 0), (-1, -1), 8),
+                ("RIGHTPADDING", (0, 0), (-1, -1), 8),
+                ("TOPPADDING", (0, 0), (-1, -1), 6),
+                ("BOTTOMPADDING", (0, 0), (-1, -1), 6),
+            ]
+        )
+    )
+    return table
+
+
+def _section(title: str, styles):
+    from reportlab.platypus import Paragraph, Spacer
+
+    return [Spacer(1, 14), Paragraph(html.escape(title), styles["SectionHeading"]), Spacer(1, 7)]
+
+
+def write_pdf_report(result: EvaluationResult, path: str | Path, language: str | None = None) -> Path:
+    from reportlab.lib import colors
+    from reportlab.lib.enums import TA_CENTER
+    from reportlab.lib.pagesizes import A4
+    from reportlab.lib.styles import ParagraphStyle, getSampleStyleSheet
+    from reportlab.lib.units import mm
+    from reportlab.platypus import SimpleDocTemplate, Spacer
+
+    regular_font, bold_font = _register_pdf_fonts(language)
+    root = Path(path)
+    root.parent.mkdir(parents=True, exist_ok=True)
+
+    base = getSampleStyleSheet()
+    styles = {
+        "Title": ParagraphStyle(
+            "InsightFaceTitle",
+            parent=base["Title"],
+            fontName=bold_font,
+            fontSize=24,
+            leading=30,
+            textColor=colors.HexColor("#0f172a"),
+            alignment=TA_CENTER,
+            spaceAfter=8,
+        ),
+        "Subtitle": ParagraphStyle(
+            "InsightFaceSubtitle",
+            parent=base["BodyText"],
+            fontName=regular_font,
+            fontSize=10.5,
+            leading=14,
+            textColor=colors.HexColor("#475569"),
+            alignment=TA_CENTER,
+        ),
+        "Notice": ParagraphStyle(
+            "InsightFaceNotice",
+            parent=base["BodyText"],
+            fontName=regular_font,
+            fontSize=9.5,
+            leading=13,
+            textColor=colors.HexColor("#1e3a8a"),
+            backColor=colors.HexColor("#eaf3ff"),
+            borderColor=colors.HexColor("#93c5fd"),
+            borderWidth=0.8,
+            borderPadding=8,
+            spaceAfter=8,
+        ),
+        "SectionHeading": ParagraphStyle(
+            "InsightFaceSectionHeading",
+            parent=base["Heading2"],
+            fontName=bold_font,
+            fontSize=14,
+            leading=18,
+            textColor=colors.HexColor("#1e3a8a"),
+            spaceBefore=6,
+            spaceAfter=4,
+        ),
+        "Body": ParagraphStyle(
+            "InsightFaceBody",
+            parent=base["BodyText"],
+            fontName=regular_font,
+            fontSize=9.5,
+            leading=13.5,
+            textColor=colors.HexColor("#1f2937"),
+        ),
+        "TableHeader": ParagraphStyle(
+            "InsightFaceTableHeader",
+            parent=base["BodyText"],
+            fontName=bold_font,
+            fontSize=8.8,
+            leading=11,
+            textColor=colors.HexColor("#0f172a"),
+        ),
+        "TableCellStrong": ParagraphStyle(
+            "InsightFaceTableCellStrong",
+            parent=base["BodyText"],
+            fontName=bold_font,
+            fontSize=8.4,
+            leading=10.5,
+            textColor=colors.HexColor("#0f172a"),
+        ),
+        "TableCell": ParagraphStyle(
+            "InsightFaceTableCell",
+            parent=base["BodyText"],
+            fontName=regular_font,
+            fontSize=8.2,
+            leading=10.5,
+            textColor=colors.HexColor("#334155"),
+        ),
+    }
+
+    story: List[Any] = [
+        _pdf_paragraph(_rt("title", language), styles["Title"]),
+        _pdf_paragraph(f"{result.scenario} · {result.created_at}", styles["Subtitle"]),
+        Spacer(1, 8),
+        _pdf_paragraph(_rt("local_notice", language), styles["Notice"]),
+    ]
+    story.extend(_section(f"1. {_rt('executive_summary', language)}", styles))
+    story.append(_pdf_table(_summary_rows(result, language), styles, language))
+    story.extend(_section(f"2. {_rt('dataset_summary', language)}", styles))
+    story.append(_pdf_table(list(result.dataset_summary.items()), styles, language))
+    story.extend(_section(f"3. {_rt('model_runtime', language)}", styles))
+    story.append(
+        _pdf_table(
+            [
+                (_rt("model", language), result.model_name),
+                (_rt("provider", language), result.provider),
+                (_rt("threshold", language), f"{result.threshold:.6f}"),
+            ],
+            styles,
+            language,
+        )
+    )
+    story.extend(_section(f"4. {_rt('metrics', language)}", styles))
+    story.append(_pdf_table(list(result.metrics.items()), styles, language))
+    story.extend(_section(f"5. {_rt('error_analysis', language)}", styles))
+    story.append(
+        _pdf_paragraph(
+            safe_json_dumps(result.errors[:50]) if result.errors else _rt("no_errors", language),
+            styles["Body"],
+        )
+    )
+    story.extend(_section(f"6. {_rt('latency_hardware', language)}", styles))
+    story.append(_pdf_table(list(result.latency.items()), styles, language))
+    story.extend(_section(f"7. {_rt('deployment_considerations', language)}", styles))
+    story.append(_pdf_paragraph(_rt("commercial_notice", language), styles["Body"]))
+    story.extend(_section(f"8. {_rt('responsible_use', language)}", styles))
+    story.append(_pdf_paragraph(_rt("responsible_notice", language), styles["Body"]))
+    story.append(_pdf_paragraph(_rt("legal_notice", language), styles["Body"]))
+    story.extend(_section(f"9. {_rt('commercial_next_steps', language)}", styles))
+    story.append(_pdf_paragraph(_rt("contact_notice", language), styles["Body"]))
+    raw_rows = result.raw_results[:30]
+    story.extend(_section(f"10. {_rt('appendix_raw', language)}", styles))
+    story.append(_pdf_paragraph(_rt("raw_note", language).format(count=len(raw_rows)), styles["Body"]))
+    story.append(_pdf_paragraph(safe_json_dumps(raw_rows), styles["Body"]))
+
+    def footer(canvas, doc):
+        canvas.saveState()
+        canvas.setFillColor(colors.HexColor("#64748b"))
+        canvas.setFont(regular_font, 8)
+        canvas.drawString(18 * mm, 12 * mm, _rt("generated_by", language))
+        canvas.drawRightString(192 * mm, 12 * mm, str(doc.page))
+        canvas.restoreState()
+
+    doc = SimpleDocTemplate(
+        str(root),
+        pagesize=A4,
+        rightMargin=18 * mm,
+        leftMargin=18 * mm,
+        topMargin=16 * mm,
+        bottomMargin=18 * mm,
+        title=_rt("title", language),
+    )
+    doc.build(story, onFirstPage=footer, onLaterPages=footer)
+    return root
+
+
+def write_reports(result: EvaluationResult, report_dir: str | Path, language: str | None = None) -> Dict[str, str]:
+    root = Path(report_dir)
+    root.mkdir(parents=True, exist_ok=True)
+    stem = f"insightface_evaluation_{timestamp_for_filename()}"
+    markdown_path = root / f"{stem}.md"
+    html_path = root / f"{stem}.html"
+    pdf_path = root / f"{stem}.pdf"
+    markdown_path.write_text(generate_markdown_report(result, language), encoding="utf-8")
+    html_path.write_text(generate_html_report(result, language), encoding="utf-8")
+    paths = {"markdown": str(markdown_path), "html": str(html_path)}
+    try:
+        write_pdf_report(result, pdf_path, language)
+        paths["pdf"] = str(pdf_path)
+    except Exception:
+        pass
+    result.report_path = paths.get("pdf") or paths["markdown"]
+    return paths
diff --git a/python-package/insightface/gui/core/storage.py b/python-package/insightface/gui/core/storage.py
new file mode 100644
index 000000000..c65da732d
--- /dev/null
+++ b/python-package/insightface/gui/core/storage.py
@@ -0,0 +1,737 @@
+"""SQLite storage for local people, media, embeddings, and reports."""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Any, Dict, Iterable, Iterator, List, Optional
+
+import numpy as np
+
+from .logging import get_logger
+from .constants import DEFAULT_THRESHOLD
+from .recognition import search_gallery
+from .utils import safe_json_dumps, utc_now_iso
+
+LOGGER = get_logger("storage")
+ALBUM_DIRECTORIES_SETTING = "album.directories"
+ALBUM_RESULTS_SETTING = "album.results"
+
+
+def embedding_to_blob(embedding: np.ndarray | Iterable[float] | None) -> tuple[Optional[bytes], int]:
+    if embedding is None:
+        return None, 0
+    arr = np.asarray(embedding, dtype=np.float32).reshape(-1)
+    return arr.tobytes(), int(arr.shape[0])
+
+
+def blob_to_embedding(blob: bytes | memoryview | None, dim: int | None = None) -> Optional[np.ndarray]:
+    if blob is None:
+        return None
+    arr = np.frombuffer(blob, dtype=np.float32).copy()
+    if dim and arr.shape[0] != dim:
+        return None
+    return arr
+
+
+class Storage:
+    def __init__(self, database_path: str | Path):
+        self.database_path = Path(database_path).expanduser()
+        self.database_path.parent.mkdir(parents=True, exist_ok=True)
+        self.init_db()
+
+    @contextmanager
+    def connect(self) -> Iterator[sqlite3.Connection]:
+        conn = sqlite3.connect(str(self.database_path))
+        conn.row_factory = sqlite3.Row
+        try:
+            yield conn
+            conn.commit()
+        except Exception:
+            conn.rollback()
+            LOGGER.exception("Database operation failed")
+            raise
+        finally:
+            conn.close()
+
+    def init_db(self) -> None:
+        with self.connect() as conn:
+            conn.executescript(
+                """
+                CREATE TABLE IF NOT EXISTS people (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    name TEXT NOT NULL,
+                    display_name TEXT,
+                    notes TEXT,
+                    tags TEXT,
+                    cover_face_sample_id INTEGER,
+                    created_at TEXT,
+                    updated_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS face_samples (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    person_id INTEGER,
+                    source_image_path TEXT,
+                    crop_path TEXT,
+                    embedding BLOB,
+                    embedding_dim INTEGER,
+                    bbox_json TEXT,
+                    kps_json TEXT,
+                    det_score REAL,
+                    quality_score REAL,
+                    model_name TEXT,
+                    provider TEXT,
+                    created_at TEXT,
+                    updated_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS media_items (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    path TEXT UNIQUE,
+                    media_type TEXT,
+                    width INTEGER,
+                    height INTEGER,
+                    duration_ms INTEGER,
+                    file_size INTEGER,
+                    mtime REAL,
+                    sha256 TEXT,
+                    thumbnail BLOB,
+                    thumbnail_mime TEXT,
+                    processed_at TEXT,
+                    created_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS media_faces (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    media_id INTEGER,
+                    frame_index INTEGER,
+                    timestamp_ms INTEGER,
+                    crop_path TEXT,
+                    thumbnail BLOB,
+                    thumbnail_mime TEXT,
+                    embedding BLOB,
+                    embedding_dim INTEGER,
+                    bbox_json TEXT,
+                    kps_json TEXT,
+                    det_score REAL,
+                    quality_score REAL,
+                    assigned_person_id INTEGER,
+                    predicted_person_id INTEGER,
+                    similarity REAL,
+                    cluster_id INTEGER,
+                    status TEXT,
+                    created_at TEXT,
+                    updated_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS clusters (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    name TEXT,
+                    cover_media_face_id INTEGER,
+                    notes TEXT,
+                    created_at TEXT,
+                    updated_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS recognition_logs (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    source_type TEXT,
+                    source_path TEXT,
+                    query_face_sample_id INTEGER,
+                    predicted_person_id INTEGER,
+                    similarity REAL,
+                    threshold REAL,
+                    status TEXT,
+                    payload_json TEXT,
+                    created_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS evaluation_runs (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    scenario TEXT,
+                    model_name TEXT,
+                    provider TEXT,
+                    threshold REAL,
+                    dataset_summary_json TEXT,
+                    metrics_json TEXT,
+                    hardware_json TEXT,
+                    report_path TEXT,
+                    created_at TEXT
+                );
+
+                CREATE TABLE IF NOT EXISTS app_settings (
+                    key TEXT PRIMARY KEY,
+                    value TEXT,
+                    updated_at TEXT
+                );
+                """
+            )
+            self._ensure_column(conn, "media_items", "thumbnail", "BLOB")
+            self._ensure_column(conn, "media_items", "thumbnail_mime", "TEXT")
+            self._ensure_column(conn, "media_faces", "thumbnail", "BLOB")
+            self._ensure_column(conn, "media_faces", "thumbnail_mime", "TEXT")
+
+    @staticmethod
+    def _ensure_column(conn: sqlite3.Connection, table: str, column: str, definition: str) -> None:
+        rows = conn.execute(f"PRAGMA table_info({table})").fetchall()
+        existing = {str(row["name"]) for row in rows}
+        if column not in existing:
+            conn.execute(f"ALTER TABLE {table} ADD COLUMN {column} {definition}")
+
+    def add_person(
+        self,
+        name: str,
+        display_name: Optional[str] = None,
+        notes: str = "",
+        tags: str = "",
+    ) -> int:
+        now = utc_now_iso()
+        with self.connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO people (name, display_name, notes, tags, created_at, updated_at)
+                VALUES (?, ?, ?, ?, ?, ?)
+                """,
+                (name, display_name or name, notes, tags, now, now),
+            )
+            return int(cur.lastrowid)
+
+    def update_person(self, person_id: int, **fields: Any) -> None:
+        allowed = {"name", "display_name", "notes", "tags", "cover_face_sample_id"}
+        updates = {key: value for key, value in fields.items() if key in allowed}
+        if not updates:
+            return
+        updates["updated_at"] = utc_now_iso()
+        clause = ", ".join(f"{key}=?" for key in updates)
+        values = list(updates.values()) + [person_id]
+        with self.connect() as conn:
+            conn.execute(f"UPDATE people SET {clause} WHERE id=?", values)
+
+    def delete_person(self, person_id: int, delete_samples: bool = False) -> None:
+        with self.connect() as conn:
+            if delete_samples:
+                conn.execute("DELETE FROM face_samples WHERE person_id=?", (person_id,))
+            else:
+                conn.execute("UPDATE face_samples SET person_id=NULL WHERE person_id=?", (person_id,))
+            conn.execute("UPDATE media_faces SET assigned_person_id=NULL WHERE assigned_person_id=?", (person_id,))
+            conn.execute("DELETE FROM people WHERE id=?", (person_id,))
+
+    def merge_people(self, source_person_id: int, target_person_id: int) -> None:
+        now = utc_now_iso()
+        with self.connect() as conn:
+            conn.execute(
+                "UPDATE face_samples SET person_id=?, updated_at=? WHERE person_id=?",
+                (target_person_id, now, source_person_id),
+            )
+            conn.execute(
+                "UPDATE media_faces SET assigned_person_id=? WHERE assigned_person_id=?",
+                (target_person_id, source_person_id),
+            )
+            conn.execute("DELETE FROM people WHERE id=?", (source_person_id,))
+
+    def add_face_sample(
+        self,
+        person_id: Optional[int],
+        embedding: np.ndarray | Iterable[float] | None,
+        source_image_path: str = "",
+        crop_path: str = "",
+        bbox: Optional[Iterable[float]] = None,
+        kps: Optional[Iterable[Iterable[float]]] = None,
+        det_score: float = 0.0,
+        quality_score: Optional[float] = None,
+        model_name: str = "",
+        provider: str = "",
+    ) -> int:
+        blob, dim = embedding_to_blob(embedding)
+        now = utc_now_iso()
+        with self.connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO face_samples (
+                    person_id, source_image_path, crop_path, embedding, embedding_dim,
+                    bbox_json, kps_json, det_score, quality_score, model_name, provider,
+                    created_at, updated_at
+                )
+                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    person_id,
+                    source_image_path,
+                    crop_path,
+                    blob,
+                    dim,
+                    safe_json_dumps(list(bbox) if bbox is not None else None),
+                    safe_json_dumps(list(kps) if kps is not None else None),
+                    det_score,
+                    quality_score,
+                    model_name,
+                    provider,
+                    now,
+                    now,
+                ),
+            )
+            sample_id = int(cur.lastrowid)
+            if person_id is not None:
+                conn.execute(
+                    """
+                    UPDATE people
+                    SET cover_face_sample_id=COALESCE(cover_face_sample_id, ?), updated_at=?
+                    WHERE id=?
+                    """,
+                    (sample_id, now, person_id),
+                )
+            return sample_id
+
+    def delete_face_sample(self, sample_id: int) -> None:
+        with self.connect() as conn:
+            conn.execute("DELETE FROM face_samples WHERE id=?", (sample_id,))
+
+    def list_people(self) -> List[Dict[str, Any]]:
+        with self.connect() as conn:
+            rows = conn.execute(
+                """
+                SELECT p.*,
+                       COUNT(fs.id) AS sample_count,
+                       AVG(fs.quality_score) AS avg_quality,
+                       COALESCE(
+                           (
+                               SELECT fs_cover.crop_path
+                               FROM face_samples fs_cover
+                               WHERE fs_cover.id = p.cover_face_sample_id
+                                 AND COALESCE(fs_cover.crop_path, '') <> ''
+                               LIMIT 1
+                           ),
+                           (
+                               SELECT fs2.crop_path
+                               FROM face_samples fs2
+                               WHERE fs2.person_id = p.id AND COALESCE(fs2.crop_path, '') <> ''
+                               ORDER BY fs2.id ASC
+                               LIMIT 1
+                           )
+                       ) AS cover_crop_path
+                FROM people p
+                LEFT JOIN face_samples fs ON fs.person_id = p.id
+                GROUP BY p.id
+                ORDER BY LOWER(COALESCE(p.display_name, p.name))
+                """
+            ).fetchall()
+            return [dict(row) for row in rows]
+
+    def list_face_samples(self, person_id: Optional[int] = None) -> List[Dict[str, Any]]:
+        query = "SELECT * FROM face_samples"
+        params: tuple[Any, ...] = ()
+        if person_id is not None:
+            query += " WHERE person_id=?"
+            params = (person_id,)
+        query += " ORDER BY created_at DESC, id DESC"
+        with self.connect() as conn:
+            rows = conn.execute(query, params).fetchall()
+            return [self._decode_sample(row) for row in rows]
+
+    def _decode_sample(self, row: sqlite3.Row) -> Dict[str, Any]:
+        data = dict(row)
+        data["embedding"] = blob_to_embedding(data.get("embedding"), data.get("embedding_dim"))
+        for key in ("bbox_json", "kps_json"):
+            try:
+                data[key.replace("_json", "")] = json.loads(data[key]) if data.get(key) else None
+            except Exception:
+                data[key.replace("_json", "")] = None
+        return data
+
+    def add_media_item(
+        self,
+        path: str,
+        media_type: str = "image",
+        width: Optional[int] = None,
+        height: Optional[int] = None,
+        duration_ms: Optional[int] = None,
+        file_size: Optional[int] = None,
+        mtime: Optional[float] = None,
+        sha256: Optional[str] = None,
+        thumbnail: Optional[bytes] = None,
+        thumbnail_mime: str = "image/webp",
+        processed_at: Optional[str] = None,
+    ) -> int:
+        now = utc_now_iso()
+        with self.connect() as conn:
+            conn.execute(
+                """
+                INSERT OR IGNORE INTO media_items (
+                    path, media_type, width, height, duration_ms, file_size,
+                    mtime, sha256, thumbnail, thumbnail_mime, processed_at, created_at
+                )
+                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    path,
+                    media_type,
+                    width,
+                    height,
+                    duration_ms,
+                    file_size,
+                    mtime,
+                    sha256,
+                    thumbnail,
+                    thumbnail_mime if thumbnail else "",
+                    processed_at,
+                    now,
+                ),
+            )
+            row = conn.execute("SELECT id FROM media_items WHERE path=?", (path,)).fetchone()
+            return int(row["id"])
+
+    def add_media_face(
+        self,
+        media_id: int,
+        embedding: np.ndarray | Iterable[float] | None,
+        crop_path: str = "",
+        thumbnail: Optional[bytes] = None,
+        thumbnail_mime: str = "image/webp",
+        bbox: Optional[Iterable[float]] = None,
+        kps: Optional[Iterable[Iterable[float]]] = None,
+        det_score: float = 0.0,
+        quality_score: Optional[float] = None,
+        frame_index: Optional[int] = None,
+        timestamp_ms: Optional[int] = None,
+        assigned_person_id: Optional[int] = None,
+        predicted_person_id: Optional[int] = None,
+        similarity: Optional[float] = None,
+        cluster_id: Optional[int] = None,
+        status: str = "unknown",
+    ) -> int:
+        blob, dim = embedding_to_blob(embedding)
+        now = utc_now_iso()
+        with self.connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO media_faces (
+                    media_id, frame_index, timestamp_ms, crop_path, thumbnail, thumbnail_mime, embedding, embedding_dim,
+                    bbox_json, kps_json, det_score, quality_score, assigned_person_id,
+                    predicted_person_id, similarity, cluster_id, status, created_at, updated_at
+                )
+                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    media_id,
+                    frame_index,
+                    timestamp_ms,
+                    crop_path,
+                    thumbnail,
+                    thumbnail_mime if thumbnail else "",
+                    blob,
+                    dim,
+                    safe_json_dumps(list(bbox) if bbox is not None else None),
+                    safe_json_dumps(list(kps) if kps is not None else None),
+                    det_score,
+                    quality_score,
+                    assigned_person_id,
+                    predicted_person_id,
+                    similarity,
+                    cluster_id,
+                    status,
+                    now,
+                    now,
+                ),
+            )
+            return int(cur.lastrowid)
+
+    def load_all_gallery_embeddings(self) -> List[Dict[str, Any]]:
+        with self.connect() as conn:
+            rows = conn.execute(
+                """
+                SELECT fs.id AS sample_id, fs.person_id, fs.crop_path, fs.embedding,
+                       fs.embedding_dim,
+                       COALESCE(p.display_name, p.name, 'Unknown') AS person_name
+                FROM face_samples fs
+                LEFT JOIN people p ON p.id = fs.person_id
+                WHERE fs.embedding IS NOT NULL AND fs.person_id IS NOT NULL
+                """
+            ).fetchall()
+            gallery = []
+            for row in rows:
+                item = dict(row)
+                item["embedding"] = blob_to_embedding(item.get("embedding"), item.get("embedding_dim"))
+                gallery.append(item)
+            return gallery
+
+    def search_embeddings(
+        self,
+        query_embedding: np.ndarray,
+        top_k: int = 5,
+        threshold: float = DEFAULT_THRESHOLD,
+    ):
+        return search_gallery(query_embedding, self.load_all_gallery_embeddings(), top_k=top_k, threshold=threshold)
+
+    def log_recognition(
+        self,
+        source_type: str,
+        source_path: str,
+        predicted_person_id: Optional[int],
+        similarity: float,
+        threshold: float,
+        status: str,
+        payload: Optional[Dict[str, Any]] = None,
+        query_face_sample_id: Optional[int] = None,
+    ) -> int:
+        with self.connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO recognition_logs (
+                    source_type, source_path, query_face_sample_id, predicted_person_id,
+                    similarity, threshold, status, payload_json, created_at
+                )
+                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    source_type,
+                    source_path,
+                    query_face_sample_id,
+                    predicted_person_id,
+                    similarity,
+                    threshold,
+                    status,
+                    safe_json_dumps(payload or {}),
+                    utc_now_iso(),
+                ),
+            )
+            return int(cur.lastrowid)
+
+    def save_evaluation_run(
+        self,
+        scenario: str,
+        model_name: str,
+        provider: str,
+        threshold: float,
+        dataset_summary: Dict[str, Any],
+        metrics: Dict[str, Any],
+        hardware: Dict[str, Any],
+        report_path: str = "",
+        created_at: Optional[str] = None,
+    ) -> int:
+        with self.connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO evaluation_runs (
+                    scenario, model_name, provider, threshold, dataset_summary_json,
+                    metrics_json, hardware_json, report_path, created_at
+                )
+                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    scenario,
+                    model_name,
+                    provider,
+                    threshold,
+                    safe_json_dumps(dataset_summary),
+                    safe_json_dumps(metrics),
+                    safe_json_dumps(hardware),
+                    report_path,
+                    created_at or utc_now_iso(),
+                ),
+            )
+            return int(cur.lastrowid)
+
+    def list_evaluation_runs(self) -> List[Dict[str, Any]]:
+        with self.connect() as conn:
+            rows = conn.execute("SELECT * FROM evaluation_runs ORDER BY created_at DESC, id DESC").fetchall()
+            result = []
+            for row in rows:
+                item = dict(row)
+                for key in ("dataset_summary_json", "metrics_json", "hardware_json"):
+                    try:
+                        item[key.replace("_json", "")] = json.loads(item.get(key) or "{}")
+                    except Exception:
+                        item[key.replace("_json", "")] = {}
+                result.append(item)
+            return result
+
+    def delete_evaluation_run(self, run_id: int) -> None:
+        with self.connect() as conn:
+            conn.execute("DELETE FROM evaluation_runs WHERE id=?", (run_id,))
+
+    def get_setting(self, key: str, default: Optional[str] = None) -> Optional[str]:
+        with self.connect() as conn:
+            row = conn.execute("SELECT value FROM app_settings WHERE key=?", (key,)).fetchone()
+            return row["value"] if row else default
+
+    def set_setting(self, key: str, value: Any) -> None:
+        with self.connect() as conn:
+            conn.execute(
+                """
+                INSERT INTO app_settings (key, value, updated_at)
+                VALUES (?, ?, ?)
+                ON CONFLICT(key) DO UPDATE SET value=excluded.value, updated_at=excluded.updated_at
+                """,
+                (key, str(value), utc_now_iso()),
+            )
+
+    def delete_setting(self, key: str) -> None:
+        with self.connect() as conn:
+            conn.execute("DELETE FROM app_settings WHERE key=?", (key,))
+
+    def list_album_directories(self) -> List[str]:
+        raw = self.get_setting(ALBUM_DIRECTORIES_SETTING, "[]")
+        try:
+            values = json.loads(raw or "[]")
+        except Exception:
+            return []
+        if not isinstance(values, list):
+            return []
+        directories = []
+        seen = set()
+        for value in values:
+            path = str(value)
+            if path and path not in seen:
+                directories.append(path)
+                seen.add(path)
+        return directories
+
+    def save_album_directories(self, directories: Iterable[str]) -> None:
+        values = []
+        seen = set()
+        for directory in directories:
+            path = str(directory)
+            if path and path not in seen:
+                values.append(path)
+                seen.add(path)
+        self.set_setting(ALBUM_DIRECTORIES_SETTING, json.dumps(values, ensure_ascii=False))
+
+    def save_album_results(
+        self,
+        clusters: Iterable[Dict[str, Any]],
+        cluster_items: Dict[int, List[Dict[str, Any]]],
+        algorithm: str,
+        cluster_threshold: Optional[float] = None,
+        min_samples: int = 2,
+        min_face_size: Optional[int] = None,
+    ) -> None:
+        now = utc_now_iso()
+        serializable_clusters: List[Dict[str, Any]] = []
+        assignments: List[tuple[int, int]] = []
+        for cluster in clusters:
+            cluster_id = int(cluster["id"])
+            items = cluster_items.get(cluster_id, [])
+            face_ids = [int(item["id"]) for item in items if item.get("id") is not None]
+            for face_id in face_ids:
+                assignments.append((cluster_id, face_id))
+            data = {
+                "id": cluster_id,
+                "label": cluster.get("label"),
+                "name": cluster.get("name", f"Album Person {cluster_id}"),
+                "source": cluster.get("source", "album"),
+                "face_count": int(cluster.get("face_count") or len(face_ids)),
+                "photo_count": int(cluster.get("photo_count") or 0),
+                "avg_quality": float(cluster.get("avg_quality") or 0.0),
+                "thumbnail_path": cluster.get("thumbnail_path") or "",
+                "thumbnail_face_id": cluster.get("thumbnail_face_id"),
+                "photos": list(cluster.get("photos") or []),
+                "face_ids": face_ids,
+            }
+            serializable_clusters.append(data)
+        payload = {
+            "version": 1,
+            "algorithm": algorithm,
+            "cluster_threshold": float(cluster_threshold) if cluster_threshold is not None else None,
+            "min_samples": int(min_samples),
+            "min_face_size": int(min_face_size) if min_face_size is not None else None,
+            "clusters": serializable_clusters,
+            "updated_at": now,
+        }
+        with self.connect() as conn:
+            conn.execute(
+                """
+                INSERT INTO app_settings (key, value, updated_at)
+                VALUES (?, ?, ?)
+                ON CONFLICT(key) DO UPDATE SET value=excluded.value, updated_at=excluded.updated_at
+                """,
+                (ALBUM_RESULTS_SETTING, json.dumps(payload, ensure_ascii=False), now),
+            )
+            conn.execute("UPDATE media_faces SET cluster_id=NULL, updated_at=?", (now,))
+            conn.executemany(
+                "UPDATE media_faces SET cluster_id=?, status='clustered', updated_at=? WHERE id=?",
+                [(cluster_id, now, face_id) for cluster_id, face_id in assignments],
+            )
+
+    def load_album_results(self) -> Dict[str, Any]:
+        raw = self.get_setting(ALBUM_RESULTS_SETTING, "{}")
+        try:
+            data = json.loads(raw or "{}")
+        except Exception:
+            return {}
+        return data if isinstance(data, dict) else {}
+
+    def clear_album_results(self) -> None:
+        now = utc_now_iso()
+        with self.connect() as conn:
+            conn.execute("DELETE FROM app_settings WHERE key=?", (ALBUM_RESULTS_SETTING,))
+            conn.execute("UPDATE media_faces SET cluster_id=NULL, updated_at=?", (now,))
+
+    def delete_media_items_by_paths(self, paths: Iterable[str]) -> int:
+        values = [str(path) for path in paths]
+        if not values:
+            return 0
+        deleted = 0
+        with self.connect() as conn:
+            for index in range(0, len(values), 900):
+                chunk = values[index : index + 900]
+                placeholders = ",".join("?" for _ in chunk)
+                rows = conn.execute(f"SELECT id FROM media_items WHERE path IN ({placeholders})", chunk).fetchall()
+                media_ids = [int(row["id"]) for row in rows]
+                if not media_ids:
+                    continue
+                id_placeholders = ",".join("?" for _ in media_ids)
+                conn.execute(f"DELETE FROM media_faces WHERE media_id IN ({id_placeholders})", media_ids)
+                conn.execute(f"DELETE FROM media_items WHERE id IN ({id_placeholders})", media_ids)
+                deleted += len(media_ids)
+        return deleted
+
+    def existing_media_paths(self, paths: Iterable[str]) -> set[str]:
+        values = [str(path) for path in paths]
+        if not values:
+            return set()
+        existing: set[str] = set()
+        with self.connect() as conn:
+            for index in range(0, len(values), 900):
+                chunk = values[index : index + 900]
+                placeholders = ",".join("?" for _ in chunk)
+                rows = conn.execute(f"SELECT path FROM media_items WHERE path IN ({placeholders})", chunk).fetchall()
+                existing.update(str(row["path"]) for row in rows)
+        return existing
+
+    def list_media_faces(self) -> List[Dict[str, Any]]:
+        with self.connect() as conn:
+            rows = conn.execute(
+                """
+                SELECT mf.*, mi.path AS media_path, mi.media_type, mi.width, mi.height,
+                       mi.thumbnail AS media_thumbnail,
+                       mi.thumbnail_mime AS media_thumbnail_mime
+                FROM media_faces mf
+                JOIN media_items mi ON mi.id = mf.media_id
+                WHERE mf.embedding IS NOT NULL
+                ORDER BY mf.created_at DESC, mf.id DESC
+                """
+            ).fetchall()
+            result = []
+            for row in rows:
+                item = dict(row)
+                item["embedding"] = blob_to_embedding(item.get("embedding"), item.get("embedding_dim"))
+                for key in ("bbox_json", "kps_json"):
+                    try:
+                        item[key.replace("_json", "")] = json.loads(item[key]) if item.get(key) else None
+                    except Exception:
+                        item[key.replace("_json", "")] = None
+                result.append(item)
+            return result
+
+    def counts(self) -> Dict[str, int]:
+        with self.connect() as conn:
+            return {
+                "people": int(conn.execute("SELECT COUNT(*) FROM people").fetchone()[0]),
+                "face_samples": int(conn.execute("SELECT COUNT(*) FROM face_samples").fetchone()[0]),
+                "media_items": int(conn.execute("SELECT COUNT(*) FROM media_items").fetchone()[0]),
+                "media_faces": int(conn.execute("SELECT COUNT(*) FROM media_faces").fetchone()[0]),
+                "evaluation_runs": int(conn.execute("SELECT COUNT(*) FROM evaluation_runs").fetchone()[0]),
+            }
diff --git a/python-package/insightface/gui/core/swap.py b/python-package/insightface/gui/core/swap.py
new file mode 100644
index 000000000..4b06738d4
--- /dev/null
+++ b/python-package/insightface/gui/core/swap.py
@@ -0,0 +1,116 @@
+"""Face swap model wrapper."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+
+
+class GFPGANRestorer:
+    """Small ONNXRuntime wrapper for GFPGAN 512x512 face restoration."""
+
+    def __init__(self, model_path: str = "", providers: Optional[list[str]] = None):
+        self.model_path = model_path
+        self.providers = providers or ["CPUExecutionProvider"]
+        self.session = None
+        self.input_name = ""
+        self.output_name = ""
+        self.last_error = ""
+
+    def load(self) -> bool:
+        path = Path(self.model_path).expanduser()
+        if not path.exists():
+            self.last_error = "GFPGAN model not found. Download GFPGANv1.4.onnx in Models > Downloads first."
+            return False
+        try:
+            import onnxruntime
+
+            onnxruntime.set_default_logger_severity(3)
+            self.session = onnxruntime.InferenceSession(str(path), providers=self.providers)
+            self.input_name = self.session.get_inputs()[0].name
+            self.output_name = self.session.get_outputs()[0].name
+            return True
+        except Exception as exc:
+            self.last_error = f"GFPGAN model load failed: {exc}"
+            return False
+
+    def restore(self, image: np.ndarray) -> np.ndarray:
+        if self.session is None:
+            raise RuntimeError("GFPGAN model is not loaded.")
+        try:
+            import cv2
+        except Exception as exc:
+            raise RuntimeError(f"OpenCV is required for GFPGAN post-processing: {exc}") from exc
+
+        original_height, original_width = image.shape[:2]
+        resized = cv2.resize(image, (512, 512), interpolation=cv2.INTER_LINEAR)
+        rgb = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
+        tensor = ((rgb - 0.5) / 0.5).transpose(2, 0, 1)[None, :, :, :]
+        outputs = self.session.run([self.output_name], {self.input_name: tensor.astype(np.float32)})
+        restored = np.asarray(outputs[0])
+        if restored.ndim == 4:
+            restored = restored[0]
+        if restored.shape[0] == 3:
+            restored = restored.transpose(1, 2, 0)
+        restored = restored.astype(np.float32)
+        if float(restored.min()) < 0.0:
+            restored = (restored + 1.0) * 0.5
+        restored = np.clip(restored, 0.0, 1.0)
+        restored = (restored * 255.0).round().astype(np.uint8)
+        bgr = cv2.cvtColor(restored, cv2.COLOR_RGB2BGR)
+        if (original_width, original_height) != (512, 512):
+            bgr = cv2.resize(bgr, (original_width, original_height), interpolation=cv2.INTER_LINEAR)
+        return bgr
+
+
+class FaceSwapEngine:
+    def __init__(
+        self,
+        model_path: str = "",
+        providers: Optional[list[str]] = None,
+        gfpgan_model_path: str = "",
+        enable_gfpgan: bool = False,
+    ):
+        self.model_path = model_path
+        self.providers = providers or ["CPUExecutionProvider"]
+        self.model = None
+        self.enable_gfpgan = enable_gfpgan
+        self.restorer = GFPGANRestorer(gfpgan_model_path, self.providers) if enable_gfpgan else None
+        self.last_error = ""
+
+    def is_available(self) -> bool:
+        return self.model is not None
+
+    def load(self) -> bool:
+        path = Path(self.model_path).expanduser()
+        if not path.exists():
+            self.last_error = (
+                "Face swap model not found. Please configure a valid swap model in Models."
+            )
+            return False
+        try:
+            import onnxruntime
+            from insightface.model_zoo import get_model
+
+            onnxruntime.set_default_logger_severity(3)
+            self.model = get_model(str(path), providers=self.providers)
+            if self.model is None:
+                self.last_error = "Configured face swap model could not be recognized."
+                return False
+            if self.restorer is not None and not self.restorer.load():
+                self.last_error = self.restorer.last_error
+                return False
+            return True
+        except Exception as exc:
+            self.last_error = f"Face swap model load failed: {exc}"
+            return False
+
+    def swap(self, image: np.ndarray, target_face, source_face) -> np.ndarray:
+        if self.model is None:
+            raise RuntimeError("Face swap model not found. Please configure a valid swap model in Models.")
+        output = self.model.get(image, target_face, source_face, paste_back=True)
+        if self.restorer is not None:
+            output = self.restorer.restore(output)
+        return output
diff --git a/python-package/insightface/gui/core/theme.py b/python-package/insightface/gui/core/theme.py
new file mode 100644
index 000000000..fbf2b99f5
--- /dev/null
+++ b/python-package/insightface/gui/core/theme.py
@@ -0,0 +1,622 @@
+"""Qt stylesheet helpers for the desktop GUI."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from PySide6.QtCore import Qt
+from PySide6.QtWidgets import QApplication
+
+
+@dataclass(frozen=True)
+class ThemeOption:
+    value: str
+    label: str
+    description: str
+
+
+THEME_OPTIONS = [
+    ThemeOption("system", "System", "Follow the operating system color scheme."),
+    ThemeOption("precision_light", "Precision Light", "Neutral white and graphite for long review sessions."),
+    ThemeOption("studio_dark", "Studio Dark", "Deep navy studio with violet and cyan model-workbench accents."),
+    ThemeOption("graphite_pro", "Graphite Pro", "Charcoal enterprise console with amber command accents."),
+    ThemeOption("azure_lab", "Azure Lab", "Bright blue laboratory workspace with crisp cyan interaction states."),
+    ThemeOption("emerald_focus", "Emerald Focus", "Green privacy-first workspace for local biometric review."),
+    ThemeOption("crimson_audit", "Crimson Audit", "Warm red audit theme for risk review and compliance workflows."),
+]
+
+_THEME_ALIASES = {
+    "light": "precision_light",
+    "dark": "studio_dark",
+}
+
+_LIGHT_THEME = "precision_light"
+_DARK_THEME = "studio_dark"
+
+
+def normalize_theme(theme: str | None) -> str:
+    value = (theme or "system").strip().lower()
+    value = _THEME_ALIASES.get(value, value)
+    valid = {option.value for option in THEME_OPTIONS}
+    return value if value in valid else "system"
+
+
+def theme_label(theme: str | None) -> str:
+    value = normalize_theme(theme)
+    for option in THEME_OPTIONS:
+        if option.value == value:
+            return option.label
+    return "System"
+
+
+def theme_description(theme: str | None) -> str:
+    value = normalize_theme(theme)
+    for option in THEME_OPTIONS:
+        if option.value == value:
+            return option.description
+    return THEME_OPTIONS[0].description
+
+
+def effective_theme(theme: str | None) -> str:
+    value = normalize_theme(theme)
+    if value != "system":
+        return value
+    app = QApplication.instance()
+    if app is not None:
+        try:
+            if app.styleHints().colorScheme() == Qt.ColorScheme.Dark:
+                return _DARK_THEME
+        except Exception:
+            pass
+    return _LIGHT_THEME
+
+
+def _palettes() -> dict[str, dict[str, str]]:
+    return {
+        "precision_light": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#f5f5f2",
+            "surface": "#ffffff",
+            "surface_alt": "#eceee9",
+            "panel": "#e2e4dd",
+            "panel_hover": "#d9dcd4",
+            "text": "#1f2428",
+            "muted": "#636a70",
+            "subtle": "#8a9096",
+            "border": "#d0d3ca",
+            "border_strong": "#a8ada4",
+            "accent": "#2f3a45",
+            "accent_2": "#5a6672",
+            "accent_soft": "#dfe3e7",
+            "accent_faint": "#f0f2f3",
+            "field": "#ffffff",
+            "button": "#ffffff",
+            "button_hover": "#ebece8",
+            "button_pressed": "#dfe3e7",
+            "selection": "#d2d8dd",
+            "success": "#2f855a",
+            "success_soft": "#dcefe5",
+            "warning": "#b7791f",
+            "danger": "#c53030",
+            "danger_hover": "#9b2c2c",
+            "upload_bg": "#f7f7f4",
+            "upload_hover": "#eceee9",
+            "upload_drag": "#e2f2e8",
+            "upload_file": "#ffffff",
+            "notice_bg": "#eceee9",
+            "notice_border": "#c8cdc4",
+            "table_alt": "#f7f7f4",
+            "header": "#e4e7e0",
+            "chip": "#f4f5f2",
+        },
+        "studio_dark": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#070b1d",
+            "surface": "#0d1533",
+            "surface_alt": "#121d42",
+            "panel": "#18295b",
+            "panel_hover": "#213575",
+            "text": "#edf3ff",
+            "muted": "#a9b8d8",
+            "subtle": "#7284ae",
+            "border": "#27386f",
+            "border_strong": "#4a63b3",
+            "accent": "#7c5cff",
+            "accent_2": "#24d6ff",
+            "accent_soft": "#241d63",
+            "accent_faint": "#111949",
+            "field": "#080f28",
+            "button": "#14204a",
+            "button_hover": "#1b2d65",
+            "button_pressed": "#251f6d",
+            "selection": "#332b82",
+            "success": "#2ee6a6",
+            "success_soft": "#0e3a33",
+            "warning": "#ffca4f",
+            "danger": "#ff5c8a",
+            "danger_hover": "#d93668",
+            "upload_bg": "#0b1431",
+            "upload_hover": "#13265a",
+            "upload_drag": "#083840",
+            "upload_file": "#0d1533",
+            "notice_bg": "#121d54",
+            "notice_border": "#3750a4",
+            "table_alt": "#101a3d",
+            "header": "#17265a",
+            "chip": "#121d42",
+        },
+        "graphite_pro": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#13110d",
+            "surface": "#1b1914",
+            "surface_alt": "#252119",
+            "panel": "#30291d",
+            "panel_hover": "#3a3122",
+            "text": "#f6f1e8",
+            "muted": "#c2b8a8",
+            "subtle": "#8d8375",
+            "border": "#3e3527",
+            "border_strong": "#6a5840",
+            "accent": "#f59e0b",
+            "accent_2": "#facc15",
+            "accent_soft": "#4a320d",
+            "accent_faint": "#2a2114",
+            "field": "#12100d",
+            "button": "#272218",
+            "button_hover": "#3a3122",
+            "button_pressed": "#4a320d",
+            "selection": "#62440f",
+            "success": "#84cc16",
+            "success_soft": "#26350f",
+            "warning": "#f97316",
+            "danger": "#f43f5e",
+            "danger_hover": "#be123c",
+            "upload_bg": "#19160f",
+            "upload_hover": "#2f2515",
+            "upload_drag": "#27340f",
+            "upload_file": "#1b1914",
+            "notice_bg": "#2a2114",
+            "notice_border": "#76541c",
+            "table_alt": "#211d16",
+            "header": "#30291d",
+            "chip": "#272218",
+        },
+        "azure_lab": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#e5f2ff",
+            "surface": "#fafdff",
+            "surface_alt": "#d8ebff",
+            "panel": "#c3ddff",
+            "panel_hover": "#aed0ff",
+            "text": "#09264a",
+            "muted": "#31577d",
+            "subtle": "#6685a4",
+            "border": "#a9cbf3",
+            "border_strong": "#5b9bed",
+            "accent": "#075eea",
+            "accent_2": "#00a4ff",
+            "accent_soft": "#bfe0ff",
+            "accent_faint": "#eef8ff",
+            "field": "#ffffff",
+            "button": "#fafdff",
+            "button_hover": "#d8ebff",
+            "button_pressed": "#bfe0ff",
+            "selection": "#93c5fd",
+            "success": "#009b72",
+            "success_soft": "#c9f5e7",
+            "warning": "#c27200",
+            "danger": "#d92d5b",
+            "danger_hover": "#a61d42",
+            "upload_bg": "#fafdff",
+            "upload_hover": "#d8ebff",
+            "upload_drag": "#c9f5ff",
+            "upload_file": "#ffffff",
+            "notice_bg": "#d8ebff",
+            "notice_border": "#77b5f7",
+            "table_alt": "#eef7ff",
+            "header": "#cfe4ff",
+            "chip": "#e8f4ff",
+        },
+        "emerald_focus": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#e8f7ef",
+            "surface": "#f8fff9",
+            "surface_alt": "#d7f2e2",
+            "panel": "#bfe8d1",
+            "panel_hover": "#a8ddc2",
+            "text": "#0d2f22",
+            "muted": "#3b6b58",
+            "subtle": "#6f927f",
+            "border": "#9bd2b5",
+            "border_strong": "#4fa878",
+            "accent": "#07844f",
+            "accent_2": "#00b894",
+            "accent_soft": "#b9ead1",
+            "accent_faint": "#eefbf3",
+            "field": "#ffffff",
+            "button": "#f8fff9",
+            "button_hover": "#d7f2e2",
+            "button_pressed": "#b9ead1",
+            "selection": "#8bdcb2",
+            "success": "#079455",
+            "success_soft": "#c6f6d5",
+            "warning": "#b7791f",
+            "danger": "#c2415b",
+            "danger_hover": "#9f1239",
+            "upload_bg": "#f8fff9",
+            "upload_hover": "#d7f2e2",
+            "upload_drag": "#c6f6d5",
+            "upload_file": "#ffffff",
+            "notice_bg": "#d7f2e2",
+            "notice_border": "#63bd8e",
+            "table_alt": "#eefaf3",
+            "header": "#cdebdc",
+            "chip": "#e9f8ef",
+        },
+        "crimson_audit": {
+            "font": '"Segoe UI", "Inter", "Noto Sans", "Helvetica Neue", Arial, sans-serif',
+            "bg": "#190b10",
+            "surface": "#241018",
+            "surface_alt": "#321522",
+            "panel": "#451b2c",
+            "panel_hover": "#5a2438",
+            "text": "#fff1f4",
+            "muted": "#e5a9b5",
+            "subtle": "#ae7280",
+            "border": "#653046",
+            "border_strong": "#9f4965",
+            "accent": "#ff3b6b",
+            "accent_2": "#ff8a3d",
+            "accent_soft": "#5a1830",
+            "accent_faint": "#32101f",
+            "field": "#1b0c12",
+            "button": "#351724",
+            "button_hover": "#512236",
+            "button_pressed": "#641d36",
+            "selection": "#7f2946",
+            "success": "#22c55e",
+            "success_soft": "#123422",
+            "warning": "#f97316",
+            "danger": "#ff3b6b",
+            "danger_hover": "#be123c",
+            "upload_bg": "#1f0d15",
+            "upload_hover": "#42182b",
+            "upload_drag": "#2d2a10",
+            "upload_file": "#241018",
+            "notice_bg": "#3a1427",
+            "notice_border": "#8c3a55",
+            "table_alt": "#2b1320",
+            "header": "#3b1828",
+            "chip": "#321522",
+        },
+    }
+
+
+def application_stylesheet(theme: str | None) -> str:
+    palette = _palettes()[effective_theme(theme)]
+
+    return f"""
+    * {{
+        font-family: {palette["font"]};
+        font-size: 13px;
+    }}
+    QMainWindow, QDialog {{
+        background: {palette["bg"]};
+        color: {palette["text"]};
+    }}
+    QWidget {{
+        color: {palette["text"]};
+        selection-background-color: {palette["selection"]};
+        selection-color: {palette["text"]};
+    }}
+    QWidget#topAppBar {{
+        background: {palette["surface"]};
+        border-bottom: 1px solid {palette["border"]};
+    }}
+    QWidget#modeRail {{
+        background: {palette["surface"]};
+        border-right: 1px solid {palette["border"]};
+    }}
+    QWidget#modeSidebar {{
+        background: {palette["surface_alt"]};
+        border-right: 1px solid {palette["border"]};
+    }}
+    QLabel[role="muted"], QLabel[role="status"], QLabel[role="secondary"] {{
+        color: {palette["muted"]};
+    }}
+    QLabel[role="statusChip"] {{
+        padding: 4px 9px;
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        background: {palette["chip"]};
+        color: {palette["text"]};
+        font-weight: 600;
+    }}
+    QLabel#noticeLabel {{
+        background: {palette["notice_bg"]};
+        border: 1px solid {palette["notice_border"]};
+        border-radius: 6px;
+        color: {palette["text"]};
+        padding: 9px 10px;
+    }}
+    QLabel#localProcessingNotice {{
+        background: {palette["notice_bg"]};
+        border: 1px solid {palette["notice_border"]};
+        border-radius: 8px;
+        color: {palette["muted"]};
+        padding: 9px 10px;
+        line-height: 1.25;
+    }}
+    QFrame#downloadSourceFooter {{
+        background: {palette["surface_alt"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+    }}
+    QPushButton {{
+        background: {palette["button"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        padding: 7px 12px;
+        font-weight: 600;
+    }}
+    QPushButton:hover {{
+        background: {palette["button_hover"]};
+        border-color: {palette["border_strong"]};
+    }}
+    QPushButton:pressed {{
+        background: {palette["button_pressed"]};
+        border-color: {palette["accent"]};
+    }}
+    QPushButton:disabled {{
+        color: {palette["subtle"]};
+        background: {palette["surface_alt"]};
+        border-color: {palette["border"]};
+    }}
+    QPushButton#primaryButton {{
+        background: {palette["accent"]};
+        border-color: {palette["accent"]};
+        color: #ffffff;
+    }}
+    QPushButton#primaryButton:hover {{
+        background: {palette["accent_2"]};
+        border-color: {palette["accent_2"]};
+    }}
+    QPushButton#removeUpload {{
+        background: {palette["button_pressed"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["border_strong"]};
+        border-radius: 11px;
+        font-weight: 700;
+        padding: 0;
+    }}
+    QPushButton#removeUpload:hover {{
+        background: {palette["danger"]};
+        border-color: {palette["danger"]};
+        color: #ffffff;
+    }}
+    QLineEdit, QComboBox, QSpinBox, QDoubleSpinBox, QTextEdit, QPlainTextEdit {{
+        background: {palette["field"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        padding: 6px 8px;
+    }}
+    QLineEdit:focus, QComboBox:focus, QSpinBox:focus, QDoubleSpinBox:focus,
+    QTextEdit:focus, QPlainTextEdit:focus {{
+        border-color: {palette["accent"]};
+        background: {palette["surface"]};
+    }}
+    QComboBox::drop-down {{
+        border: 0;
+        width: 24px;
+    }}
+    QAbstractItemView {{
+        background: {palette["surface"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["border"]};
+        selection-background-color: {palette["selection"]};
+    }}
+    QComboBox QAbstractItemView::item:disabled {{
+        color: {palette["subtle"]};
+        background: {palette["surface_alt"]};
+    }}
+    QListWidget {{
+        background: transparent;
+        border: 0;
+        outline: 0;
+    }}
+    QListWidget::item {{
+        border-radius: 6px;
+        padding: 8px 10px;
+        margin: 2px 0;
+    }}
+    QListWidget::item:selected {{
+        background: {palette["accent_soft"]};
+        color: {palette["text"]};
+    }}
+    QListWidget::item:hover {{
+        background: {palette["button_hover"]};
+    }}
+    QListWidget#modeList::item {{
+        border-radius: 8px;
+        padding: 10px 10px;
+        margin: 3px 0;
+        color: {palette["muted"]};
+        border: 1px solid transparent;
+    }}
+    QListWidget#modeList::item:selected {{
+        background: {palette["accent_soft"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["accent"]};
+    }}
+    QListWidget#modeList::item:hover {{
+        background: {palette["button_hover"]};
+        color: {palette["text"]};
+    }}
+    QTableWidget, QTableView {{
+        background: {palette["surface"]};
+        alternate-background-color: {palette["table_alt"]};
+        color: {palette["text"]};
+        gridline-color: {palette["border"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        selection-background-color: {palette["selection"]};
+    }}
+    QTableWidget::item, QTableView::item {{
+        padding: 4px;
+    }}
+    QTableWidget::item:selected, QTableView::item:selected {{
+        background: {palette["selection"]};
+        color: {palette["text"]};
+    }}
+    QHeaderView::section {{
+        background: {palette["header"]};
+        color: {palette["text"]};
+        border: 0;
+        border-bottom: 1px solid {palette["border"]};
+        padding: 7px;
+        font-weight: 650;
+    }}
+    QFrame, QWidget#dashboardCard {{
+        border-color: {palette["border"]};
+    }}
+    QWidget#dashboardCard {{
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        background: {palette["surface"]};
+    }}
+    QWidget#dashboardCard:hover {{
+        border-color: {palette["accent"]};
+        background: {palette["surface_alt"]};
+    }}
+    QFrame#enterpriseCard {{
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        background: {palette["surface"]};
+        padding: 10px;
+    }}
+    QLabel#enterpriseCardTitle {{
+        color: {palette["text"]};
+    }}
+    QFrame#uploadPreview, QFrame#dropInput, QFrame#galleryUpload,
+    QFrame#imageOrFolderImport, QListWidget#albumDirectoryList {{
+        border: 1px dashed {palette["border_strong"]};
+        border-radius: 6px;
+        background: {palette["upload_bg"]};
+        padding: 8px;
+    }}
+    QFrame#uploadPreview[hoverActive="true"], QFrame#dropInput[hoverActive="true"],
+    QFrame#galleryUpload[hoverActive="true"], QFrame#imageOrFolderImport[hoverActive="true"],
+    QListWidget#albumDirectoryList[hoverActive="true"] {{
+        border-color: {palette["accent"]};
+        background: {palette["upload_hover"]};
+    }}
+    QFrame#uploadPreview[dragActive="true"], QFrame#dropInput[dragActive="true"],
+    QFrame#galleryUpload[dragActive="true"], QFrame#imageOrFolderImport[dragActive="true"],
+    QListWidget#albumDirectoryList[dragActive="true"] {{
+        border-color: {palette["success"]};
+        background: {palette["upload_drag"]};
+    }}
+    QFrame#uploadPreview[hasFile="true"], QFrame#dropInput[hasFiles="true"],
+    QFrame#galleryUpload[hasFiles="true"] {{
+        border: 1px solid {palette["border"]};
+        background: {palette["upload_file"]};
+    }}
+    QFrame#uploadPreview QLabel, QFrame#dropInput QLabel, QFrame#galleryUpload QLabel,
+    QFrame#imageOrFolderImport QLabel {{
+        background: transparent;
+    }}
+    QFrame#uploadPreview QGraphicsView, QFrame#galleryUpload QListWidget {{
+        background: transparent;
+        border: 0;
+    }}
+    QGraphicsView#imageViewer {{
+        background: {palette["upload_bg"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+    }}
+    QWidget#imageViewerViewport {{
+        background: {palette["upload_bg"]};
+    }}
+    QLabel#uploadPrompt, QLabel#dropPrompt {{
+        color: {palette["muted"]};
+        font-size: 15px;
+        font-weight: 650;
+        padding: 18px;
+    }}
+    QLabel#pathLabel {{
+        color: {palette["muted"]};
+        padding: 0 8px 8px 8px;
+    }}
+    QMenuBar, QMenu {{
+        background: {palette["surface"]};
+        color: {palette["text"]};
+        border-color: {palette["border"]};
+    }}
+    QMenuBar::item:selected, QMenu::item:selected {{
+        background: {palette["accent_soft"]};
+    }}
+    QTabWidget::pane {{
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        background: {palette["surface"]};
+    }}
+    QTabBar::tab {{
+        background: {palette["surface_alt"]};
+        border: 1px solid {palette["border"]};
+        border-bottom: 0;
+        padding: 7px 12px;
+        margin-right: 2px;
+        border-top-left-radius: 6px;
+        border-top-right-radius: 6px;
+    }}
+    QTabBar::tab:selected {{
+        background: {palette["surface"]};
+        color: {palette["accent"]};
+    }}
+    QProgressBar {{
+        background: {palette["surface_alt"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 6px;
+        color: {palette["text"]};
+        text-align: center;
+    }}
+    QProgressBar::chunk {{
+        background: {palette["accent"]};
+        border-radius: 5px;
+    }}
+    QSlider::groove:horizontal {{
+        height: 6px;
+        background: {palette["surface_alt"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 3px;
+    }}
+    QSlider::handle:horizontal {{
+        background: {palette["accent"]};
+        border: 1px solid {palette["accent"]};
+        width: 16px;
+        height: 16px;
+        margin: -6px 0;
+        border-radius: 8px;
+    }}
+    QSplitter::handle {{
+        background: {palette["bg"]};
+    }}
+    QScrollArea {{
+        background: transparent;
+        border: 0;
+    }}
+    QStatusBar {{
+        background: {palette["surface"]};
+        color: {palette["muted"]};
+        border-top: 1px solid {palette["border"]};
+    }}
+    QToolTip {{
+        background: {palette["surface"]};
+        color: {palette["text"]};
+        border: 1px solid {palette["border"]};
+        border-radius: 4px;
+        padding: 4px;
+    }}
+    """
diff --git a/python-package/insightface/gui/core/tooltips.py b/python-package/insightface/gui/core/tooltips.py
new file mode 100644
index 000000000..ea958ce81
--- /dev/null
+++ b/python-package/insightface/gui/core/tooltips.py
@@ -0,0 +1,108 @@
+"""Shared button tooltip helpers."""
+
+from __future__ import annotations
+
+import re
+
+from PySide6.QtWidgets import QAbstractButton, QWidget
+
+
+BUTTON_TOOLTIPS = {
+    "Add Folder": "Add an album folder to the import list.",
+    "Add Person": "Create a new person entry in the local People Library.",
+    "Add as New Person": "Register the current face as a new person.",
+    "Apply": "Apply changes without closing this dialog.",
+    "Assign to Existing Person": "Assign the current face to an existing person.",
+    "Browse": "Choose a local file or folder.",
+    "Cancel": "Close without applying changes.",
+    "Check Cameras": "Check available local camera devices.",
+    "Clear": "Clear the current selection or form.",
+    "Compare": "Run 1:1 face comparison for the selected images.",
+    "Contact Enterprise Support": "Open the InsightFace enterprise support contact page.",
+    "Delete": "Delete the selected item.",
+    "Delete Selected": "Delete the selected record.",
+    "Download Selected": "Download the selected model package to the local model cache.",
+    "Exit": "Close the application.",
+    "Export": "Export the current results.",
+    "Export All Summary": "Export a summary of all available records.",
+    "Export Annotated Image": "Export the image with face labels drawn on it.",
+    "Export CSV": "Export the current table as CSV.",
+    "Export CSV / JSON": "Export current results as CSV and JSON.",
+    "Export License Summary": "Save a local license summary report.",
+    "Export Report": "Export the current evaluation report.",
+    "Export Result": "Export the current result.",
+    "Export Results": "Export the current results.",
+    "Export Settings": "Export application settings to a JSON file.",
+    "Import / Refresh": "Scan selected folders for new images and rebuild album clusters from indexed faces.",
+    "Import Image or Folder": "Import one image or a folder of images by clicking or dragging files here.",
+    "Import Settings": "Import application settings from a JSON file.",
+    "Open": "Open the selected item.",
+    "Open GitHub Releases": "Open the InsightFace GitHub Releases page.",
+    "Open License Center": "Open license and commercial usage information.",
+    "Open Model Downloads": "Open the model downloads tab.",
+    "Open Model Folder": "Open the local InsightFace model folder.",
+    "Open Report Folder": "Open the local folder where reports are saved.",
+    "Open Result Directory": "Open the local folder where face swap results are saved.",
+    "Open Selected Report": "Open the selected report file.",
+    "Open Workspace Folder": "Open the local workspace folder.",
+    "Re-export Selected": "Regenerate and export the selected report.",
+    "Rebuild All": "Reprocess selected album folders from scratch after confirmation.",
+    "Recognize Faces": "Detect and recognize all faces in the selected image.",
+    "Refresh": "Reload the current list from local storage.",
+    "Refresh Download URLs": "Refresh available model download URLs from GitHub Releases.",
+    "Refresh People": "Reload people from the local People Library.",
+    "Register Current Face": "Register the current camera face as a person.",
+    "Remove": "Remove the selected item.",
+    "Remove Selected": "Remove selected items from this list.",
+    "Request Commercial Model License": "Show contact options for commercial model licensing.",
+    "Request Face Swap Commercial License": "Show contact options for face swap commercial licensing.",
+    "Request Private Model Evaluation": "Show contact options for private model evaluation.",
+    "Request SDK / API": "Show contact options for SDK or API access.",
+    "Reset": "Reset this page to its initial state.",
+    "Reset Settings": "Reset application settings to their defaults.",
+    "Run Evaluation": "Run the selected local enterprise evaluation.",
+    "Run Face Swap": "Run face swap with the configured local swap model.",
+    "Run Recognition": "Run 1:1 comparison or 1:N gallery recognition.",
+    "Save": "Save changes and close this dialog.",
+    "Save Details": "Save the current person details.",
+    "Save Settings": "Save the current settings.",
+    "Scan": "Scan selected local data.",
+    "Scan and Cluster": "Scan images and cluster detected faces.",
+    "Search": "Search the local People Library for matching faces.",
+    "Select": "Select a local file or folder.",
+    "Start": "Start the current processing task.",
+    "Stop": "Stop the current processing task.",
+    "Test Model Load": "Test loading the selected model and runtime provider.",
+    "Use Selected Model": "Use the selected downloaded model package.",
+    "Visit Homepage": "Open the InsightFace homepage.",
+    "Warmup": "Run a short model warmup to measure startup latency.",
+    "×": "Remove the current file.",
+}
+
+
+def normalize_button_text(text: str) -> str:
+    value = str(text or "").replace("&", "").strip()
+    value = re.sub(r"\s+", " ", value)
+    return value
+
+
+def tooltip_for_button(text: str) -> str:
+    label = normalize_button_text(text)
+    if not label:
+        return "Click this button to run the related action."
+    if label in BUTTON_TOOLTIPS:
+        return BUTTON_TOOLTIPS[label]
+    if label.lower() == "ok":
+        return "Confirm and continue."
+    return f"Click to {label[0].lower() + label[1:]}."
+
+
+def set_button_tooltip(button: QAbstractButton, tooltip: str | None = None) -> None:
+    if button.toolTip():
+        return
+    button.setToolTip(tooltip or tooltip_for_button(button.text()))
+
+
+def apply_button_tooltips(widget: QWidget) -> None:
+    for button in widget.findChildren(QAbstractButton):
+        set_button_tooltip(button)
diff --git a/python-package/insightface/gui/core/utils.py b/python-package/insightface/gui/core/utils.py
new file mode 100644
index 000000000..bd245da3d
--- /dev/null
+++ b/python-package/insightface/gui/core/utils.py
@@ -0,0 +1,149 @@
+"""General utilities for image and file handling."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+from io import BytesIO
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Iterable, Optional
+
+import numpy as np
+
+
+def utc_now_iso() -> str:
+    return datetime.utcnow().replace(microsecond=0).isoformat() + "Z"
+
+
+def read_image(path: str | os.PathLike[str]) -> Optional[np.ndarray]:
+    file_path = Path(path)
+    try:
+        import cv2
+
+        data = np.fromfile(str(file_path), dtype=np.uint8)
+        img = cv2.imdecode(data, cv2.IMREAD_COLOR)
+        return img
+    except Exception:
+        try:
+            from PIL import Image
+
+            with Image.open(file_path) as image:
+                rgb = image.convert("RGB")
+                arr = np.asarray(rgb)
+                return arr[..., ::-1].copy()
+        except Exception:
+            return None
+
+
+def save_image(path: str | os.PathLike[str], image: np.ndarray) -> bool:
+    file_path = Path(path)
+    file_path.parent.mkdir(parents=True, exist_ok=True)
+    try:
+        import cv2
+
+        ext = file_path.suffix or ".png"
+        ok, data = cv2.imencode(ext, image)
+        if not ok:
+            return False
+        data.tofile(str(file_path))
+        return True
+    except Exception:
+        try:
+            from PIL import Image
+
+            rgb = image[..., ::-1] if image.ndim == 3 and image.shape[2] >= 3 else image
+            Image.fromarray(rgb).save(file_path)
+            return True
+        except Exception:
+            return False
+
+
+def encode_webp_thumbnail(
+    image: np.ndarray | None,
+    max_side: int,
+    quality: int,
+) -> bytes | None:
+    """Encode an image array as a small WebP thumbnail.
+
+    GUI image arrays are BGR when loaded through OpenCV. The helper converts to
+    RGB for Pillow, keeps aspect ratio, and returns compressed bytes for SQLite.
+    """
+
+    if image is None or image.size == 0:
+        return None
+    try:
+        from PIL import Image
+
+        arr = np.asarray(image)
+        if arr.ndim == 3 and arr.shape[2] >= 3:
+            arr = arr[..., :3][..., ::-1]
+        pil_image = Image.fromarray(arr).convert("RGB")
+        width, height = pil_image.size
+        longest = max(width, height)
+        if longest <= 0:
+            return None
+        scale = min(1.0, float(max_side) / float(longest))
+        target = (max(1, int(round(width * scale))), max(1, int(round(height * scale))))
+        if target != pil_image.size:
+            pil_image = pil_image.resize(target, Image.Resampling.LANCZOS)
+        buffer = BytesIO()
+        pil_image.save(buffer, format="WEBP", quality=int(quality), method=6)
+        return buffer.getvalue()
+    except Exception:
+        return None
+
+
+def crop_bbox(image: np.ndarray, bbox: Iterable[float], padding: float = 0.10) -> np.ndarray:
+    x1, y1, x2, y2 = [float(v) for v in bbox]
+    width = x2 - x1
+    height = y2 - y1
+    x1 -= width * padding
+    x2 += width * padding
+    y1 -= height * padding
+    y2 += height * padding
+    ix1, iy1 = max(0, int(round(x1))), max(0, int(round(y1)))
+    ix2, iy2 = min(image.shape[1], int(round(x2))), min(image.shape[0], int(round(y2)))
+    if ix2 <= ix1 or iy2 <= iy1:
+        return image.copy()
+    return image[iy1:iy2, ix1:ix2].copy()
+
+
+def file_sha256(path: str | os.PathLike[str], chunk_size: int = 1024 * 1024) -> str:
+    digest = hashlib.sha256()
+    with open(path, "rb") as handle:
+        for chunk in iter(lambda: handle.read(chunk_size), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def safe_json_dumps(data: Any) -> str:
+    def default(value: Any) -> Any:
+        if isinstance(value, np.ndarray):
+            return value.tolist()
+        if isinstance(value, (np.floating, np.integer)):
+            return value.item()
+        if isinstance(value, Path):
+            return str(value)
+        return str(value)
+
+    return json.dumps(data, ensure_ascii=False, default=default)
+
+
+def timestamp_for_filename() -> str:
+    return datetime.now().strftime("%Y%m%d_%H%M%S")
+
+
+def list_images(folder: str | os.PathLike[str], recursive: bool = True) -> list[Path]:
+    from .constants import IMAGE_EXTENSIONS
+
+    root = Path(folder)
+    iterator = root.rglob("*") if recursive else root.glob("*")
+    return sorted(path for path in iterator if path.is_file() and path.suffix.lower() in IMAGE_EXTENSIONS)
+
+
+def human_ms(ms: float) -> str:
+    if ms < 1000:
+        return f"{ms:.1f} ms"
+    return f"{ms / 1000:.2f} s"
diff --git a/python-package/insightface/gui/core/video.py b/python-package/insightface/gui/core/video.py
new file mode 100644
index 000000000..3ad5992ff
--- /dev/null
+++ b/python-package/insightface/gui/core/video.py
@@ -0,0 +1,51 @@
+"""Simple video helpers for local processing."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Iterator, Tuple
+
+import numpy as np
+
+
+def timestamp_hhmmss(timestamp_ms: int) -> str:
+    seconds = timestamp_ms // 1000
+    h = seconds // 3600
+    m = (seconds % 3600) // 60
+    s = seconds % 60
+    return f"{h:02d}:{m:02d}:{s:02d}"
+
+
+def iter_video_frames(path: str | Path, frame_interval: int = 10) -> Iterator[Tuple[int, int, np.ndarray]]:
+    import cv2
+
+    cap = cv2.VideoCapture(str(path))
+    if not cap.isOpened():
+        raise IOError(f"Unable to open video: {path}")
+    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+    index = 0
+    try:
+        while True:
+            ok, frame = cap.read()
+            if not ok:
+                break
+            if index % max(1, int(frame_interval)) == 0:
+                timestamp_ms = int(index / fps * 1000)
+                yield index, timestamp_ms, frame
+            index += 1
+    finally:
+        cap.release()
+
+
+def read_video_thumbnail(path: str | Path) -> np.ndarray | None:
+    """Read the first decodable frame from a video for GUI preview."""
+    import cv2
+
+    cap = cv2.VideoCapture(str(path))
+    if not cap.isOpened():
+        return None
+    try:
+        ok, frame = cap.read()
+        return frame if ok else None
+    finally:
+        cap.release()
diff --git a/python-package/insightface/gui/dialogs/__init__.py b/python-package/insightface/gui/dialogs/__init__.py
new file mode 100644
index 000000000..32f8e01bc
--- /dev/null
+++ b/python-package/insightface/gui/dialogs/__init__.py
@@ -0,0 +1 @@
+"""Global dialogs for InsightFace Evaluation Studio."""
diff --git a/python-package/insightface/gui/dialogs/license_dialog.py b/python-package/insightface/gui/dialogs/license_dialog.py
new file mode 100644
index 000000000..5bc6f58ec
--- /dev/null
+++ b/python-package/insightface/gui/dialogs/license_dialog.py
@@ -0,0 +1,27 @@
+"""License dialog."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QDialog, QLabel, QVBoxLayout
+
+from ..core.i18n import tr
+from ..pages.license_center_page import LicenseCenterPage
+
+
+class LicenseDialog(QDialog):
+    def __init__(self, context, parent=None):
+        super().__init__(parent)
+        self.context = context
+        self.setWindowTitle("License Center")
+        self.resize(900, 660)
+        layout = QVBoxLayout(self)
+        self.page = LicenseCenterPage(context)
+        layout.addWidget(self.page, 1)
+        self.status_label = QLabel("")
+        self.status_label.setWordWrap(True)
+        layout.addWidget(self.status_label)
+
+    def set_status(self, message: str) -> None:
+        self.status_label.setText(tr(message, self.context.config.ui_language))
+        if self.parent() is not None and hasattr(self.parent(), "set_status"):
+            self.parent().set_status(message)
diff --git a/python-package/insightface/gui/dialogs/model_manager_dialog.py b/python-package/insightface/gui/dialogs/model_manager_dialog.py
new file mode 100644
index 000000000..24c24d238
--- /dev/null
+++ b/python-package/insightface/gui/dialogs/model_manager_dialog.py
@@ -0,0 +1,74 @@
+"""Model runtime and download manager dialog."""
+
+from __future__ import annotations
+
+from PySide6.QtCore import Signal
+from PySide6.QtWidgets import QDialog, QLabel, QTabWidget, QVBoxLayout
+
+from ..core.i18n import tr
+from ..pages.model_download_page import ModelDownloadPage
+from ..pages.model_settings_page import ModelSettingsPage
+
+
+class ModelManagerDialog(QDialog):
+    modelChanged = Signal()
+
+    def __init__(self, context, parent=None):
+        super().__init__(parent)
+        self.context = context
+        self.setWindowTitle("Models")
+        self.setMinimumSize(1120, 700)
+        self.resize(1240, 740)
+        layout = QVBoxLayout(self)
+        self.tabs = QTabWidget()
+        self.runtime_page = ModelSettingsPage(context)
+        self.downloads_page = ModelDownloadPage(context)
+        self.tabs.addTab(self.runtime_page, "Runtime")
+        self.tabs.addTab(self.downloads_page, "Downloads")
+        self.tabs.currentChanged.connect(self._refresh_current_tab)
+        layout.addWidget(self.tabs, 1)
+        self.status_label = QLabel("")
+        self.status_label.setWordWrap(True)
+        layout.addWidget(self.status_label)
+
+    def open_page(self, label: str) -> None:
+        if label == "Model Downloads":
+            self.tabs.setCurrentWidget(self.downloads_page)
+            self.downloads_page.refresh()
+        elif label == "Model Settings":
+            self.tabs.setCurrentWidget(self.runtime_page)
+            self.runtime_page.refresh()
+
+    def set_status(self, message: str) -> None:
+        self.status_label.setText(tr(message, self.context.config.ui_language))
+        self.modelChanged.emit()
+        if self.parent() is not None and hasattr(self.parent(), "set_status"):
+            self.parent().set_status(message)
+
+    def run_task(self, title: str, fn, on_result=None, show_dialog: bool = True) -> None:
+        parent = self.parent()
+        if parent is not None and hasattr(parent, "run_task"):
+            parent.run_task(title, fn, on_result, show_dialog=show_dialog)
+            return
+        try:
+            result = fn()
+            if on_result:
+                on_result(result)
+        except Exception as exc:
+            self.set_status(str(exc))
+
+    def refresh_model_pages(self) -> None:
+        self.runtime_page.refresh()
+        self.downloads_page.populate()
+        self.modelChanged.emit()
+
+    def _refresh_current_tab(self, index: int | None = None) -> None:
+        del index
+        widget = self.tabs.currentWidget()
+        if hasattr(widget, "refresh"):
+            widget.refresh()
+
+    def refresh_statusbar(self) -> None:
+        self.modelChanged.emit()
+        if self.parent() is not None and hasattr(self.parent(), "refresh_statusbar"):
+            self.parent().refresh_statusbar()
diff --git a/python-package/insightface/gui/dialogs/settings_dialog.py b/python-package/insightface/gui/dialogs/settings_dialog.py
new file mode 100644
index 000000000..175c66d3a
--- /dev/null
+++ b/python-package/insightface/gui/dialogs/settings_dialog.py
@@ -0,0 +1,75 @@
+"""Application settings dialog."""
+
+from __future__ import annotations
+
+from PySide6.QtCore import Signal
+from PySide6.QtWidgets import QComboBox, QDialog, QDialogButtonBox, QFormLayout, QLabel, QVBoxLayout
+
+from ..core.config import save_config
+from ..core.i18n import LANGUAGE_OPTIONS, apply_translations, normalize_language, tr
+from ..core.theme import THEME_OPTIONS, normalize_theme, theme_description
+
+
+class SettingsDialog(QDialog):
+    settingsSaved = Signal()
+
+    def __init__(self, context, parent=None):
+        super().__init__(parent)
+        self.context = context
+        self.setWindowTitle("Settings")
+        self.resize(620, 300)
+
+        layout = QVBoxLayout(self)
+        title = QLabel("Application Settings")
+        title.setStyleSheet("font-size:18px; font-weight:700;")
+        note = QLabel("Only appearance settings are configurable here. Workspace paths are fixed after first launch.")
+        note.setWordWrap(True)
+        note.setProperty("role", "muted")
+        layout.addWidget(title)
+        layout.addWidget(note)
+
+        form = QFormLayout()
+        form.setFieldGrowthPolicy(QFormLayout.AllNonFixedFieldsGrow)
+        self.theme = QComboBox()
+        for option in THEME_OPTIONS:
+            self.theme.addItem(option.label, option.value)
+        current = normalize_theme(self.context.config.ui_theme)
+        current_index = self.theme.findData(current)
+        self.theme.setCurrentIndex(max(0, current_index))
+        self.theme.currentIndexChanged.connect(self._update_theme_description)
+        form.addRow("UI theme", self.theme)
+        self.language = QComboBox()
+        for option in LANGUAGE_OPTIONS:
+            self.language.addItem(option.native_label, option.value)
+        current_language = normalize_language(self.context.config.ui_language)
+        language_index = self.language.findData(current_language)
+        self.language.setCurrentIndex(max(0, language_index))
+        self.language.currentIndexChanged.connect(self._update_theme_description)
+        form.addRow("Language", self.language)
+        layout.addLayout(form, 1)
+        self.theme_help = QLabel("")
+        self.theme_help.setWordWrap(True)
+        self.theme_help.setProperty("role", "muted")
+        layout.addWidget(self.theme_help)
+        self._update_theme_description()
+
+        buttons = QDialogButtonBox(QDialogButtonBox.Save | QDialogButtonBox.Apply | QDialogButtonBox.Cancel)
+        buttons.button(QDialogButtonBox.Save).clicked.connect(self.save_and_close)
+        buttons.button(QDialogButtonBox.Apply).clicked.connect(self.apply)
+        buttons.button(QDialogButtonBox.Cancel).clicked.connect(self.reject)
+        layout.addWidget(buttons)
+        apply_translations(self, self.context.config.ui_language)
+
+    def apply(self) -> None:
+        self.context.config.ui_theme = self.theme.currentData()
+        self.context.config.ui_language = self.language.currentData()
+        save_config(self.context.config)
+        self.settingsSaved.emit()
+        apply_translations(self, self.context.config.ui_language)
+
+    def save_and_close(self) -> None:
+        self.apply()
+        self.accept()
+
+    def _update_theme_description(self) -> None:
+        self.theme_help.setText(tr(theme_description(self.theme.currentData()), self.language.currentData()))
diff --git a/python-package/insightface/gui/main_window.py b/python-package/insightface/gui/main_window.py
new file mode 100644
index 000000000..455271f6e
--- /dev/null
+++ b/python-package/insightface/gui/main_window.py
@@ -0,0 +1,616 @@
+"""Main window, mode navigation, and task orchestration."""
+
+from __future__ import annotations
+
+import inspect
+from pathlib import Path
+from typing import Callable
+
+from PySide6.QtCore import QObject, QRunnable, QSize, Qt, QThreadPool, QTimer, QUrl, Signal, Slot
+from PySide6.QtGui import QAction, QDesktopServices, QFontMetrics, QKeySequence
+from PySide6.QtWidgets import (
+    QApplication,
+    QComboBox,
+    QDialog,
+    QDialogButtonBox,
+    QFileDialog,
+    QFormLayout,
+    QHBoxLayout,
+    QLabel,
+    QLineEdit,
+    QListWidget,
+    QListWidgetItem,
+    QMainWindow,
+    QMessageBox,
+    QPushButton,
+    QStackedWidget,
+    QStatusBar,
+    QVBoxLayout,
+    QWidget,
+)
+
+from .core.config import save_config
+from .core.constants import APP_VERSION, LOCAL_PROCESSING_NOTICE, RESPONSIBLE_USE_NOTICE, WINDOW_TITLE
+from .core.face_engine import is_cuda_provider_available
+from .core.i18n import apply_translations, tr
+from .core.links import open_insightface_url
+from .core.navigation import (
+    AppMode,
+    NAVIGATION_MODES,
+    last_page_attr,
+    mode_from_value,
+)
+from .core.theme import application_stylesheet
+from .core.tooltips import apply_button_tooltips, set_button_tooltip
+from .dialogs.license_dialog import LicenseDialog
+from .dialogs.model_manager_dialog import ModelManagerDialog
+from .dialogs.settings_dialog import SettingsDialog
+from .page_registry import PageRegistry
+from .pages.license_center_page import ENTERPRISE_HELP_URL
+from .resources import app_icon
+from .widgets.progress_dialog import StudioProgressDialog
+
+
+class WorkerSignals(QObject):
+    result = Signal(object)
+    error = Signal(str)
+    progress = Signal(int, int, str)
+    finished = Signal()
+
+
+class Worker(QRunnable):
+    def __init__(self, fn: Callable):
+        super().__init__()
+        self.fn = fn
+        self.signals = WorkerSignals()
+        self.cancelled = False
+
+    def cancel(self) -> None:
+        self.cancelled = True
+
+    @Slot()
+    def run(self) -> None:
+        try:
+            signature = inspect.signature(self.fn)
+            if "progress" in signature.parameters or "is_cancelled" in signature.parameters:
+                result = self.fn(
+                    progress=lambda current, total, message="": self.signals.progress.emit(int(current), int(total), str(message)),
+                    is_cancelled=lambda: self.cancelled,
+                )
+            else:
+                result = self.fn()
+            self.signals.result.emit(result)
+        except Exception as exc:
+            self.signals.error.emit(str(exc))
+        finally:
+            self.signals.finished.emit()
+
+
+class FirstLaunchWizard(QDialog):
+    def __init__(self, config, parent=None):
+        super().__init__(parent)
+        self.setWindowTitle("First Launch Wizard")
+        self.config = config
+        layout = QVBoxLayout(self)
+        intro = QLabel(
+            "\n".join(
+                [
+                    "Welcome to InsightFace Evaluation Studio.",
+                    LOCAL_PROCESSING_NOTICE,
+                    RESPONSIBLE_USE_NOTICE,
+                ]
+            )
+        )
+        intro.setWordWrap(True)
+        layout.addWidget(intro)
+        form = QFormLayout()
+        self.workspace = QLineEdit(config.workspace_path)
+        browse = QLabel("<a href='#'>Browse</a>")
+        browse.linkActivated.connect(self.browse_workspace)
+        workspace_row = QWidget()
+        workspace_layout = QVBoxLayout(workspace_row)
+        workspace_layout.setContentsMargins(0, 0, 0, 0)
+        workspace_layout.addWidget(self.workspace)
+        workspace_layout.addWidget(browse)
+        self.mode = QComboBox()
+        self.mode.setProperty("i18nItems", True)
+        self.mode.addItem("Personal / Research", "Personal / Research")
+        self.mode.addItem("Enterprise Evaluation", "Enterprise Evaluation")
+        self.model = QComboBox()
+        self.model.addItems(["buffalo_l", "buffalo_s", "antelopev2", "custom model directory"])
+        self.provider = QComboBox()
+        self.provider.addItems(["Auto", "CPU", "CUDA"])
+        self._update_provider_availability()
+        self.license_notice = QLabel(
+            "This application runs locally. Code and model licenses may differ. "
+            "Research models may be limited to non-commercial use. Commercial deployment "
+            "requires appropriate model license. This tool does not provide legal advice."
+        )
+        self.license_notice.setWordWrap(True)
+        form.addRow("Workspace", workspace_row)
+        form.addRow("Mode", self.mode)
+        form.addRow("Model package", self.model)
+        form.addRow("Provider", self.provider)
+        form.addRow("License notice", self.license_notice)
+        layout.addLayout(form)
+        buttons = QDialogButtonBox(QDialogButtonBox.Ok | QDialogButtonBox.Cancel)
+        buttons.accepted.connect(self.accept)
+        buttons.rejected.connect(self.reject)
+        layout.addWidget(buttons)
+        apply_button_tooltips(self)
+
+    def browse_workspace(self) -> None:
+        folder = QFileDialog.getExistingDirectory(self, "Select workspace", str(Path.home()))
+        if folder:
+            self.workspace.setText(folder)
+
+    def accept(self) -> None:
+        self.config.workspace_path = self.workspace.text().strip()
+        self.config.mode = str(self.mode.currentData() or self.mode.currentText())
+        self.config.ui_last_mode = (
+            AppMode.ENTERPRISE_EVALUATION.value
+            if self.config.mode == "Enterprise Evaluation"
+            else AppMode.FACE_VERIFICATION.value
+        )
+        selected_model = self.model.currentText()
+        self.config.model_name = selected_model if selected_model != "custom model directory" else self.config.model_name
+        provider = self.provider.currentText()
+        self.config.provider = "Auto" if provider == "CUDA" and not is_cuda_provider_available() else provider
+        self.config.apply_workspace_defaults()
+        save_config(self.config)
+        super().accept()
+
+    def _update_provider_availability(self) -> None:
+        cuda_available = is_cuda_provider_available()
+        cuda_index = self.provider.findText("CUDA")
+        if cuda_index >= 0:
+            item = self.provider.model().item(cuda_index)
+            if item is not None:
+                item.setEnabled(cuda_available)
+                item.setToolTip(
+                    "CUDAExecutionProvider is available."
+                    if cuda_available
+                    else "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first."
+                )
+        self.provider.setToolTip("Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.")
+
+
+class MainWindow(QMainWindow):
+    LEGACY_PAGE_MAP = {
+        "Dashboard": "face_dashboard",
+        "Face Recognition": "verification",
+        "Verification": "verification",
+        "Verification Dashboard": "face_dashboard",
+        "Album Dashboard": "album_dashboard",
+        "Swap Dashboard": "swap_dashboard",
+        "Evaluation Dashboard": "enterprise_dashboard",
+        "1:1 Compare": "compare",
+        "1:N Face Search": "face_search",
+        "Multi-face Photo Recognition": "multiface_photo",
+        "Batch Folder Processing": "batch_processing",
+        "Camera Recognition": "camera_recognition",
+        "Video Person Search": "video_search",
+        "People Library": "people_library",
+        "Album People Clustering": "album_people_clustering",
+        "Enterprise Evaluation": "enterprise_evaluation",
+        "Reports": "reports",
+        "Image Face Swap": "image_face_swap",
+    }
+
+    GLOBAL_PAGES = {"Settings", "Model Settings", "Model Downloads", "License Center"}
+
+    def __init__(self, context, parent=None):
+        super().__init__(parent)
+        self.context = context
+        self.apply_theme()
+        self.setWindowTitle(WINDOW_TITLE)
+        icon = app_icon()
+        if icon is not None and not icon.isNull():
+            self.setWindowIcon(icon)
+        self.resize(int(context.config.ui_window_width), int(context.config.ui_window_height))
+        self.thread_pool = QThreadPool.globalInstance()
+        self.active_workers: set[Worker] = set()
+        self.page_registry = PageRegistry(context)
+        self.pages = self.page_registry.pages
+        self.current_mode = mode_from_value(context.config.ui_last_mode)
+        self.current_page_key = ""
+
+        self.stack = QStackedWidget()
+        self.mode_rail = QWidget()
+        self.mode_rail.setObjectName("modeRail")
+        self.mode_rail.setFixedWidth(230)
+        self.mode_list = QListWidget()
+        self.mode_list.setObjectName("modeList")
+        self.mode_list.setAlternatingRowColors(False)
+        self.mode_list.setWordWrap(True)
+        self.mode_list.setSpacing(2)
+        self.mode_list.itemClicked.connect(self._mode_list_clicked)
+        self.sidebar = QWidget()
+        self.sidebar.setObjectName("modeSidebar")
+        self.sidebar.setFixedWidth(int(context.config.ui_sidebar_width or 240))
+        self.sidebar_title = QLabel()
+        self.sidebar_title.setStyleSheet("font-size:18px; font-weight:700;")
+        self.sidebar_description = QLabel()
+        self.sidebar_description.setWordWrap(True)
+        self.sidebar_description.setProperty("role", "muted")
+        self.sidebar_list = QListWidget()
+        self.sidebar_list.setAlternatingRowColors(False)
+        self.sidebar_list.itemClicked.connect(self._sidebar_clicked)
+
+        central = QWidget()
+        root = QVBoxLayout(central)
+        root.setContentsMargins(0, 0, 0, 0)
+        root.setSpacing(0)
+        root.addWidget(self._build_top_bar())
+        body = QWidget()
+        body_layout = QHBoxLayout(body)
+        body_layout.setContentsMargins(0, 0, 0, 0)
+        body_layout.addWidget(self._build_mode_rail())
+        sidebar_layout = QVBoxLayout(self.sidebar)
+        sidebar_layout.setContentsMargins(14, 14, 10, 14)
+        sidebar_layout.addWidget(self.sidebar_title)
+        sidebar_layout.addWidget(self.sidebar_description)
+        sidebar_layout.addWidget(self.sidebar_list, 1)
+        body_layout.addWidget(self.sidebar)
+        body_layout.addWidget(self.stack, 1)
+        root.addWidget(body, 1)
+        self.setCentralWidget(central)
+
+        self._build_menu()
+        self._build_statusbar()
+        apply_button_tooltips(self)
+        self.change_mode(self.current_mode, restore_last=True, save=False)
+        self.apply_language()
+        self.refresh_statusbar()
+        if not self.context.config_exists:
+            QTimer.singleShot(100, self.show_first_launch)
+        if self.context.config.auto_load_model and not self.context.runtime_safe_mode:
+            QTimer.singleShot(250, self.auto_load_model)
+
+    def _build_top_bar(self) -> QWidget:
+        bar = QWidget()
+        bar.setObjectName("topAppBar")
+        layout = QHBoxLayout(bar)
+        layout.setContentsMargins(14, 8, 14, 8)
+        title = QLabel("InsightFace Evaluation Studio")
+        title.setStyleSheet("font-size:18px; font-weight:700; border:0;")
+        version = QLabel(f"v{APP_VERSION}")
+        version.setProperty("role", "muted")
+        self.model_chip = QLabel()
+        self.provider_chip = QLabel()
+        self.license_chip = QLabel()
+        for chip in (self.model_chip, self.provider_chip, self.license_chip):
+            chip.setProperty("role", "statusChip")
+        layout.addWidget(title)
+        layout.addWidget(version)
+        layout.addStretch(1)
+        layout.addWidget(self.model_chip)
+        layout.addWidget(self.provider_chip)
+        layout.addWidget(self.license_chip)
+        layout.addWidget(self._top_button("Models", self.open_model_manager))
+        layout.addWidget(self._top_button("License", self.open_license_dialog))
+        layout.addWidget(self._top_button("Settings", self.open_settings_dialog))
+        return bar
+
+    def _top_button(self, text: str, callback: Callable) -> QPushButton:
+        button = QPushButton(text)
+        button.clicked.connect(callback)
+        set_button_tooltip(button)
+        return button
+
+    def _build_mode_rail(self) -> QWidget:
+        layout = QVBoxLayout(self.mode_rail)
+        layout.setContentsMargins(14, 14, 12, 14)
+        layout.setSpacing(8)
+        title = QLabel("Workflows")
+        title.setObjectName("modeRailTitle")
+        title.setStyleSheet("font-size:14px; font-weight:700; border:0;")
+        description = QLabel("Choose the workspace for the task.")
+        description.setProperty("role", "muted")
+        description.setWordWrap(True)
+        layout.addWidget(title)
+        layout.addWidget(description)
+        for mode in NAVIGATION_MODES.values():
+            item = QListWidgetItem(f"{mode.title}\n{self._mode_subtitle(mode.id)}")
+            item.setData(Qt.UserRole, mode.id.value)
+            item.setToolTip(mode.description)
+            item.setSizeHint(QSize(190, 58))
+            self.mode_list.addItem(item)
+        layout.addWidget(self.mode_list, 1)
+        local_notice = QLabel("All processing is local. No images, embeddings, or reports are uploaded automatically.")
+        local_notice.setObjectName("localProcessingNotice")
+        local_notice.setWordWrap(True)
+        layout.addWidget(local_notice)
+        layout.addSpacing(6)
+        enterprise_help = QPushButton("Enterprise Help")
+        enterprise_help.setObjectName("enterpriseHelpButton")
+        enterprise_help.setToolTip(
+            "Open the InsightFace contact page for enterprise evaluation, licensing, SDK/API, SLA, or custom training help."
+        )
+        enterprise_help.clicked.connect(self.open_enterprise_help)
+        layout.addWidget(enterprise_help)
+        return self.mode_rail
+
+    def _mode_subtitle(self, mode: AppMode) -> str:
+        subtitles = {
+            AppMode.FACE_VERIFICATION: "Query and gallery recognition",
+            AppMode.ALBUM_MANAGEMENT: "Photo clustering and review",
+            AppMode.FACE_SWAP: "Source + Target = Result",
+            AppMode.ENTERPRISE_EVALUATION: "Datasets, metrics, and reports",
+        }
+        return tr(subtitles[mode], self.context.config.ui_language)
+
+    def _build_menu(self) -> None:
+        file_menu = self.menuBar().addMenu("File")
+        open_workspace = file_menu.addAction("Open Workspace Folder")
+        open_workspace.triggered.connect(lambda: QDesktopServices.openUrl(QUrl.fromLocalFile(self.context.config.workspace_path)))
+        quit_action = file_menu.addAction("Exit")
+        quit_action.setShortcut(QKeySequence("Ctrl+Q"))
+        quit_action.triggered.connect(self.close)
+
+        mode_menu = self.menuBar().addMenu("Mode")
+        for mode in NAVIGATION_MODES:
+            action = mode_menu.addAction(NAVIGATION_MODES[mode].title)
+            action.triggered.connect(lambda checked=False, target=mode: self.change_mode(target))
+
+        tools_menu = self.menuBar().addMenu("Tools")
+        models_action = QAction("Models", self)
+        models_action.setShortcut(QKeySequence("Ctrl+M"))
+        models_action.triggered.connect(self.open_model_manager)
+        settings_action = QAction("Settings", self)
+        settings_action.setShortcut(QKeySequence("Ctrl+,"))
+        settings_action.triggered.connect(self.open_settings_dialog)
+        license_action = QAction("License Center", self)
+        license_action.setShortcut(QKeySequence("Ctrl+L"))
+        license_action.triggered.connect(self.open_license_dialog)
+        tools_menu.addAction(models_action)
+        tools_menu.addAction(settings_action)
+        tools_menu.addAction(license_action)
+
+        help_menu = self.menuBar().addMenu("Help")
+        about_action = help_menu.addAction("About")
+        about_action.triggered.connect(
+            lambda: QMessageBox.information(
+                self,
+                tr("About", self.context.config.ui_language),
+                f"InsightFace Evaluation Studio v{APP_VERSION}",
+            )
+        )
+        responsible_action = help_menu.addAction("Responsible Use Notice")
+        responsible_action.triggered.connect(
+            lambda: QMessageBox.information(
+                self,
+                tr("Responsible Use Notice", self.context.config.ui_language),
+                tr(RESPONSIBLE_USE_NOTICE, self.context.config.ui_language),
+            )
+        )
+
+    def _build_statusbar(self) -> None:
+        self.status = QStatusBar()
+        self.setStatusBar(self.status)
+        self.status_labels = {
+            "model": QLabel(),
+            "provider": QLabel(),
+            "database": QLabel(),
+            "license": QLabel(),
+            "running": QLabel("Ready"),
+        }
+        for label in self.status_labels.values():
+            self.status.addPermanentWidget(label)
+
+    def _mode_list_clicked(self, item: QListWidgetItem) -> None:
+        mode = mode_from_value(item.data(Qt.UserRole))
+        if mode != self.current_mode:
+            self.change_mode(mode)
+
+    def change_mode(self, mode: AppMode | str, restore_last: bool = True, save: bool = True) -> None:
+        self.current_mode = mode_from_value(mode)
+        self._sync_mode_list_selection()
+        self._rebuild_sidebar()
+        page_key = getattr(self.context.config, last_page_attr(self.current_mode), "") if restore_last else ""
+        valid_keys = {item.page_key for item in NAVIGATION_MODES[self.current_mode].items}
+        if page_key not in valid_keys:
+            page_key = NAVIGATION_MODES[self.current_mode].items[0].page_key
+        if save:
+            self.context.config.ui_last_mode = self.current_mode.value
+            save_config(self.context.config)
+        self.open_page(page_key, from_mode_change=True)
+
+    def _sync_mode_list_selection(self) -> None:
+        for row in range(self.mode_list.count()):
+            item = self.mode_list.item(row)
+            if item.data(Qt.UserRole) == self.current_mode.value:
+                self.mode_list.blockSignals(True)
+                self.mode_list.setCurrentItem(item)
+                self.mode_list.blockSignals(False)
+                break
+
+    def _rebuild_sidebar(self) -> None:
+        mode = NAVIGATION_MODES[self.current_mode]
+        self.sidebar_title.setText(tr(mode.title, self.context.config.ui_language))
+        self.sidebar_description.setText(tr(mode.description, self.context.config.ui_language))
+        self.sidebar.setVisible(False)
+        self.sidebar_list.clear()
+        for nav_item in mode.items:
+            title = tr(nav_item.title, self.context.config.ui_language) + (
+                "  " + tr("Coming soon", self.context.config.ui_language) if nav_item.coming_soon else ""
+            )
+            item = QListWidgetItem(title)
+            item.setData(Qt.UserRole, nav_item.page_key)
+            item.setToolTip(tr(nav_item.description, self.context.config.ui_language))
+            if not nav_item.enabled:
+                item.setFlags(item.flags() & ~Qt.ItemIsEnabled)
+                item.setForeground(Qt.gray)
+            self.sidebar_list.addItem(item)
+
+    def _sidebar_clicked(self, item: QListWidgetItem) -> None:
+        if item.flags() & Qt.ItemIsEnabled:
+            self.open_page(str(item.data(Qt.UserRole)))
+
+    def open_page(self, page_key: str, from_mode_change: bool = False) -> None:
+        if page_key in self.GLOBAL_PAGES:
+            if page_key == "Settings":
+                self.open_settings_dialog()
+            elif page_key in {"Model Settings", "Model Downloads"}:
+                self.open_model_manager(initial=page_key)
+            elif page_key == "License Center":
+                self.open_license_dialog()
+            return
+        page_key = self.LEGACY_PAGE_MAP.get(page_key, page_key)
+        page = self.page_registry.get(page_key)
+        if self.stack.indexOf(page) < 0:
+            self.stack.addWidget(page)
+        apply_button_tooltips(page)
+        if hasattr(page, "refresh"):
+            page.refresh()
+        apply_translations(page, self.context.config.ui_language)
+        self.stack.setCurrentWidget(page)
+        self.current_page_key = page_key
+        for row in range(self.sidebar_list.count()):
+            item = self.sidebar_list.item(row)
+            if item.data(Qt.UserRole) == page_key:
+                self.sidebar_list.setCurrentItem(item)
+                break
+        if not from_mode_change:
+            setattr(self.context.config, last_page_attr(self.current_mode), page_key)
+            save_config(self.context.config)
+        self.set_status(self._page_title(page_key))
+
+    def _page_title(self, page_key: str) -> str:
+        for mode in NAVIGATION_MODES.values():
+            for item in mode.items:
+                if item.page_key == page_key:
+                    return tr(item.title, self.context.config.ui_language)
+        return page_key
+
+    def show_first_launch(self) -> None:
+        wizard = FirstLaunchWizard(self.context.config, self)
+        apply_translations(wizard, self.context.config.ui_language)
+        if wizard.exec() == QDialog.Accepted:
+            self.context.storage = self.context.storage.__class__(self.context.config.database_path)
+            self.change_mode(mode_from_value(self.context.config.ui_last_mode), restore_last=False)
+            self.refresh_statusbar()
+
+    def auto_load_model(self) -> None:
+        def task():
+            self.context.engine.load()
+            return self.context.engine
+
+        def done(engine):
+            self.refresh_statusbar()
+            if engine.is_loaded():
+                self.set_status("Model loaded.")
+            elif engine.last_error:
+                self.set_status(engine.last_error)
+
+        self.run_task("Loading model", task, done, show_dialog=False)
+
+    def open_settings_dialog(self) -> None:
+        dialog = SettingsDialog(self.context, self)
+        apply_button_tooltips(dialog)
+        apply_translations(dialog, self.context.config.ui_language)
+        dialog.settingsSaved.connect(self._settings_saved)
+        dialog.exec()
+        self.refresh_statusbar()
+
+    def _settings_saved(self) -> None:
+        self.apply_theme()
+        self.apply_language()
+        self.refresh_statusbar()
+
+    def apply_theme(self) -> None:
+        qt_app = QApplication.instance()
+        if qt_app is not None:
+            qt_app.setStyleSheet(application_stylesheet(self.context.config.ui_theme))
+
+    def apply_language(self) -> None:
+        self._refresh_mode_list_texts()
+        self._rebuild_sidebar()
+        apply_translations(self, self.context.config.ui_language)
+        self.refresh_statusbar()
+
+    def _refresh_mode_list_texts(self) -> None:
+        for row, mode in enumerate(NAVIGATION_MODES.values()):
+            item = self.mode_list.item(row)
+            if item is None:
+                continue
+            item.setText(f"{tr(mode.title, self.context.config.ui_language)}\n{self._mode_subtitle(mode.id)}")
+            item.setToolTip(tr(mode.description, self.context.config.ui_language))
+
+    def open_model_manager(self, initial: str | None = None) -> None:
+        dialog = ModelManagerDialog(self.context, self)
+        apply_button_tooltips(dialog)
+        apply_translations(dialog, self.context.config.ui_language)
+        dialog.modelChanged.connect(self.refresh_statusbar)
+        if initial:
+            dialog.open_page(initial)
+            apply_translations(dialog, self.context.config.ui_language)
+        dialog.exec()
+        self.refresh_statusbar()
+
+    def open_license_dialog(self) -> None:
+        dialog = LicenseDialog(self.context, self)
+        apply_button_tooltips(dialog)
+        apply_translations(dialog, self.context.config.ui_language)
+        dialog.exec()
+        self.refresh_statusbar()
+
+    def open_enterprise_help(self) -> None:
+        open_insightface_url(ENTERPRISE_HELP_URL, content="sidebar_enterprise_help")
+        self.set_status("Opened InsightFace enterprise contact page.")
+
+    def set_status(self, message: str) -> None:
+        self.status_labels["running"].setText(self._elide(tr(message, self.context.config.ui_language), 180))
+
+    def refresh_statusbar(self) -> None:
+        cfg = self.context.config
+        self.model_chip.setVisible(cfg.ui_show_status_chips)
+        self.provider_chip.setVisible(cfg.ui_show_status_chips)
+        self.license_chip.setVisible(cfg.ui_show_status_chips)
+        self.model_chip.setText(self._elide(f"{tr('Model', cfg.ui_language)}: {cfg.model_name}", 150))
+        self.provider_chip.setText(self._elide(f"{tr('Provider', cfg.ui_language)}: {cfg.provider}", 130))
+        self.license_chip.setText(self._elide(tr(cfg.license_status, cfg.ui_language), 180))
+        self.model_chip.setToolTip(cfg.model_name)
+        self.provider_chip.setToolTip(cfg.provider)
+        self.license_chip.setToolTip(cfg.license_status)
+        self.status_labels["model"].setText(self._elide(f"{tr('Model', cfg.ui_language)}: {cfg.model_name}", 160))
+        self.status_labels["provider"].setText(f"{tr('Provider', cfg.ui_language)}: {cfg.provider}")
+        self.status_labels["database"].setText(self._elide(f"{tr('DB', cfg.ui_language)}: {cfg.database_path}", 280))
+        self.status_labels["database"].setToolTip(cfg.database_path)
+        self.status_labels["license"].setText(self._elide(f"{tr('License', cfg.ui_language)}: {tr(cfg.license_status, cfg.ui_language)}", 220))
+
+    def run_task(self, title: str, fn: Callable, on_result: Callable | None = None, show_dialog: bool = True) -> None:
+        worker = Worker(fn)
+        worker.setAutoDelete(False)
+        self.active_workers.add(worker)
+        dialog = StudioProgressDialog(title, self) if show_dialog else None
+        if dialog:
+            dialog.canceled.connect(worker.cancel)
+            worker.signals.progress.connect(dialog.update_progress)
+            dialog.show()
+        worker.signals.result.connect(lambda result: on_result(result) if on_result else None)
+        worker.signals.error.connect(
+            lambda message: QMessageBox.warning(
+                self,
+                tr(title, self.context.config.ui_language),
+                tr(message, self.context.config.ui_language),
+            )
+        )
+        worker.signals.error.connect(self.set_status)
+        worker.signals.finished.connect(lambda: dialog.close() if dialog else None)
+        worker.signals.finished.connect(self.refresh_statusbar)
+        worker.signals.finished.connect(lambda: self.active_workers.discard(worker))
+        worker.signals.finished.connect(worker.signals.deleteLater)
+        self.set_status(title)
+        self.thread_pool.start(worker)
+
+    def closeEvent(self, event) -> None:  # noqa: N802
+        self.context.config.ui_window_width = max(800, self.width())
+        self.context.config.ui_window_height = max(600, self.height())
+        self.context.config.ui_sidebar_width = max(200, self.sidebar.width())
+        save_config(self.context.config)
+        super().closeEvent(event)
+
+    def _elide(self, text: str, width: int) -> str:
+        metrics = QFontMetrics(self.font())
+        return metrics.elidedText(str(text), Qt.ElideMiddle, width)
diff --git a/python-package/insightface/gui/page_registry.py b/python-package/insightface/gui/page_registry.py
new file mode 100644
index 000000000..a94c47343
--- /dev/null
+++ b/python-package/insightface/gui/page_registry.py
@@ -0,0 +1,109 @@
+"""Lazy page registry used by mode-based navigation."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QWidget
+
+from .pages.album_page import AlbumPage
+from .pages.album_people_page import AlbumPeoplePage
+from .pages.batch_processing_page import BatchProcessingPage
+from .pages.camera_recognition_page import CameraRecognitionPage
+from .pages.compare_page import ComparePage
+from .pages.enterprise_eval_page import EnterpriseEvalPage
+from .pages.face_search_page import FaceSearchPage
+from .pages.face_swap_page import FaceSwapPage
+from .pages.mode_dashboards import (
+    AlbumDashboardPage,
+    EnterpriseDashboardPage,
+    FaceDashboardPage,
+    SwapDashboardPage,
+)
+from .pages.multiface_photo_page import MultiFacePhotoPage
+from .pages.people_library_page import PeopleLibraryPage
+from .pages.placeholder_page import CommercialNextStepsPage, PlaceholderPage
+from .pages.reports_page import ReportsPage
+from .pages.verification_page import VerificationPage
+from .pages.video_search_page import VideoSearchPage
+
+
+class PageRegistry:
+    def __init__(self, context):
+        self.context = context
+        self._pages: dict[str, QWidget] = {}
+
+    def get(self, page_key: str) -> QWidget:
+        if page_key not in self._pages:
+            self._pages[page_key] = self._create_page(page_key)
+        return self._pages[page_key]
+
+    @property
+    def pages(self) -> dict[str, QWidget]:
+        return self._pages
+
+    def _create_page(self, page_key: str) -> QWidget:
+        factories = {
+            "verification": VerificationPage,
+            "face_dashboard": FaceDashboardPage,
+            "album_dashboard": AlbumDashboardPage,
+            "album": AlbumPage,
+            "swap_dashboard": SwapDashboardPage,
+            "enterprise_dashboard": EnterpriseDashboardPage,
+            "compare": ComparePage,
+            "face_search": FaceSearchPage,
+            "multiface_photo": MultiFacePhotoPage,
+            "people_library": PeopleLibraryPage,
+            "batch_processing": BatchProcessingPage,
+            "camera_recognition": CameraRecognitionPage,
+            "video_search": VideoSearchPage,
+            "album_import_scan": BatchProcessingPage,
+            "album_people_clustering": AlbumPeoplePage,
+            "album_person_search": FaceSearchPage,
+            "image_face_swap": FaceSwapPage,
+            "enterprise_evaluation": EnterpriseEvalPage,
+            "reports": ReportsPage,
+            "commercial_next_steps": CommercialNextStepsPage,
+        }
+        if page_key in factories:
+            return factories[page_key](self.context)
+        placeholders = {
+            "album_multi_person_search": (
+                "Multi-person Photo Search",
+                "Find photos that contain multiple selected people.",
+                "This page will search indexed album faces for photos containing combinations such as Alice and Bob.",
+            ),
+            "album_export": (
+                "Export Album Results",
+                "Export local album organization results.",
+                "This page will export per-person folders, face crops, CSV summaries, and review queues.",
+            ),
+            "batch_image_swap": (
+                "Batch Image Swap",
+                "Swap faces across a folder of images using the configured swap model.",
+                "Batch image swap is coming soon. Use Image Face Swap for single-image trials in v1.0.1.",
+            ),
+            "video_face_swap": (
+                "Video Face Swap",
+                "Swap faces in local videos.",
+                "Video Face Swap is coming soon and may require separate commercial authorization.",
+            ),
+            "source_face_library": (
+                "Source Face Library",
+                "Prepare high-quality source faces for future swap workflows.",
+                "This page will manage reusable source faces, consent notes, and quality review.",
+            ),
+            "enterprise_dataset_setup": (
+                "Dataset Setup",
+                "Prepare local evaluation datasets and templates.",
+                "Dataset setup is currently built into Run Evaluation. A guided setup workspace is coming soon.",
+            ),
+            "threshold_calibration": (
+                "Threshold Calibration",
+                "Review thresholds, FAR, FRR, and operating point recommendations.",
+                "Threshold calibration tables are currently generated in evaluation reports. A dedicated calibration page is coming soon.",
+            ),
+        }
+        title, description, body = placeholders.get(
+            page_key,
+            ("Coming Soon", "This workflow is planned for a future release.", "Coming soon."),
+        )
+        return PlaceholderPage(self.context, title, description, body)
diff --git a/python-package/insightface/gui/pages/__init__.py b/python-package/insightface/gui/pages/__init__.py
new file mode 100644
index 000000000..02a11ea97
--- /dev/null
+++ b/python-package/insightface/gui/pages/__init__.py
@@ -0,0 +1 @@
+"""Pages used by the main stacked GUI."""
diff --git a/python-package/insightface/gui/pages/album_page.py b/python-package/insightface/gui/pages/album_page.py
new file mode 100644
index 000000000..82231931a
--- /dev/null
+++ b/python-package/insightface/gui/pages/album_page.py
@@ -0,0 +1,627 @@
+"""Single-page album import, clustering, and review workflow."""
+
+from __future__ import annotations
+
+from collections import defaultdict
+from io import BytesIO
+from pathlib import Path
+
+import numpy as np
+from PySide6.QtCore import QEvent, QSize, Qt, QUrl, Signal
+from PySide6.QtGui import QCursor, QDesktopServices, QIcon, QImage, QPixmap
+from PySide6.QtWidgets import (
+    QAbstractItemView,
+    QFileDialog,
+    QDoubleSpinBox,
+    QHBoxLayout,
+    QLabel,
+    QListWidget,
+    QMessageBox,
+    QPushButton,
+    QSpinBox,
+    QSplitter,
+    QTableWidget,
+    QTableWidgetItem,
+    QVBoxLayout,
+    QWidget,
+)
+
+from ..core.clustering import cluster_embeddings_dbscan
+from ..core.i18n import tr
+from ..core.recognition import cosine_similarity, normalize_embedding
+from ..core.tooltips import set_button_tooltip
+from ..core.utils import crop_bbox, encode_webp_thumbnail, list_images, read_image, timestamp_for_filename
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+DEFAULT_ALBUM_MIN_FACE_SIZE = 80
+DEFAULT_ALBUM_COSINE_THRESHOLD = 0.48
+
+
+class AlbumDirectoryList(QListWidget):
+    foldersChanged = Signal()
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self.setObjectName("albumDirectoryList")
+        self.setAcceptDrops(True)
+        self.setMouseTracking(True)
+        self.setSelectionMode(QListWidget.ExtendedSelection)
+        self.setToolTip("Drag one or more album folders here, or use Add Folder.")
+        self.setProperty("hoverActive", False)
+        self.setProperty("dragActive", False)
+        self.installEventFilter(self)
+        self.viewport().installEventFilter(self)
+
+    def add_folder(self, folder: str) -> None:
+        folder = str(Path(folder).expanduser())
+        if folder and Path(folder).is_dir() and folder not in self.folders():
+            self.addItem(folder)
+            self.foldersChanged.emit()
+
+    def folders(self) -> list[str]:
+        return [self.item(index).text() for index in range(self.count())]
+
+    def dragEnterEvent(self, event) -> None:  # noqa: N802
+        if self._has_folder_urls(event):
+            self._drag(True)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def dragMoveEvent(self, event) -> None:  # noqa: N802
+        if self._has_folder_urls(event):
+            self._drag(True)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def dragLeaveEvent(self, event) -> None:  # noqa: N802
+        self._drag(False)
+        super().dragLeaveEvent(event)
+
+    def dropEvent(self, event) -> None:  # noqa: N802
+        self._drag(False)
+        added = False
+        for url in event.mimeData().urls():
+            if url.isLocalFile() and Path(url.toLocalFile()).is_dir():
+                self.add_folder(url.toLocalFile())
+                added = True
+        if added:
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def _drag(self, active: bool) -> None:
+        self.setProperty("dragActive", active)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    @staticmethod
+    def _has_folder_urls(event) -> bool:
+        return event.mimeData().hasUrls() and any(
+            Path(url.toLocalFile()).is_dir()
+            for url in event.mimeData().urls()
+            if url.isLocalFile()
+        )
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() == QEvent.Enter:
+            self._set_property("hoverActive", True)
+            return False
+        if event.type() == QEvent.Leave:
+            self._update_hover_from_cursor()
+            return False
+        return super().eventFilter(watched, event)
+
+    def enterEvent(self, event) -> None:  # noqa: N802
+        self._set_property("hoverActive", True)
+        super().enterEvent(event)
+
+    def leaveEvent(self, event) -> None:  # noqa: N802
+        self._update_hover_from_cursor()
+        super().leaveEvent(event)
+
+    def _set_property(self, name: str, value) -> None:
+        self.setProperty(name, value)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    def _update_hover_from_cursor(self) -> None:
+        inside = self.rect().contains(self.mapFromGlobal(QCursor.pos()))
+        self._set_property("hoverActive", inside)
+
+
+class AlbumPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Album",
+            "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.",
+            parent,
+        )
+        self.clusters: list[dict] = []
+        self.cluster_items: dict[int, list[dict]] = {}
+        self._loaded_saved_state = False
+
+        controls = QWidget()
+        controls_layout = QVBoxLayout(controls)
+        controls_layout.setContentsMargins(0, 0, 0, 0)
+        self.content.addWidget(
+            self.notice(
+                "All album processing is local. Import / Refresh scans every selected folder, extracts features "
+                "only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold."
+            )
+        )
+        self.folder_list = AlbumDirectoryList()
+        self.folder_list.setMinimumHeight(90)
+        self.folder_list.foldersChanged.connect(self._save_directories)
+        controls_layout.addWidget(QLabel("Album directories"))
+        controls_layout.addWidget(self.folder_list)
+        hint = QLabel("Drag album folders into the directory list, or click Add Folder.")
+        hint.setProperty("role", "muted")
+        controls_layout.addWidget(hint)
+        button_row = QHBoxLayout()
+        self.import_button = self._button("Import / Refresh", self.import_refresh)
+        self.rebuild_button = self._button("Rebuild All", self.rebuild_all)
+        for button in [
+            self._button("Add Folder", self.add_folder),
+            self._button("Remove Selected", self.remove_selected),
+            self._button("Clear", self.clear_directories),
+            self.import_button,
+            self.rebuild_button,
+        ]:
+            button_row.addWidget(button)
+        button_row.addStretch(1)
+        controls_layout.addLayout(button_row)
+
+        threshold_help = QLabel(
+            "Cosine threshold is a face similarity cutoff: higher values make clusters stricter. "
+            "DBSCAN internally uses cosine distance = 1 - cosine threshold."
+        )
+        threshold_help.setObjectName("albumThresholdHelp")
+        threshold_help.setProperty("role", "muted")
+        threshold_help.setWordWrap(True)
+        controls_layout.addWidget(threshold_help)
+
+        settings_row = QHBoxLayout()
+        self.cluster_threshold = QDoubleSpinBox()
+        self.cluster_threshold.setRange(0.01, 0.99)
+        self.cluster_threshold.setSingleStep(0.01)
+        self.cluster_threshold.setDecimals(2)
+        self.cluster_threshold.setValue(DEFAULT_ALBUM_COSINE_THRESHOLD)
+        self.cluster_threshold.setToolTip(
+            "Cosine similarity threshold. Higher values make clusters stricter; default is 0.48."
+        )
+        self.min_samples = QSpinBox()
+        self.min_samples.setRange(1, 50)
+        self.min_samples.setValue(2)
+        self.min_samples.setToolTip("Minimum neighboring faces needed for a DBSCAN core point.")
+        self.min_face_size = QSpinBox()
+        self.min_face_size.setRange(1, 4096)
+        self.min_face_size.setValue(DEFAULT_ALBUM_MIN_FACE_SIZE)
+        self.min_face_size.setToolTip(
+            "Faces whose bounding box width or height is smaller than this value are skipped."
+        )
+        self.algorithm_label = QLabel("Algorithm: DBSCAN")
+        settings_row.addWidget(QLabel("Cosine threshold"))
+        settings_row.addWidget(self.cluster_threshold)
+        settings_row.addWidget(QLabel("Min samples"))
+        settings_row.addWidget(self.min_samples)
+        settings_row.addWidget(QLabel("Minimum face size"))
+        settings_row.addWidget(self.min_face_size)
+        settings_row.addWidget(self.algorithm_label)
+        settings_row.addStretch(1)
+        controls_layout.addLayout(settings_row)
+        self.content.addWidget(controls)
+
+        splitter = QSplitter(Qt.Horizontal)
+        self.cluster_table = QTableWidget(0, 2)
+        self.cluster_table.setIconSize(QSize(56, 56))
+        self.cluster_table.setSelectionBehavior(QAbstractItemView.SelectRows)
+        self.cluster_table.setSelectionMode(QAbstractItemView.SingleSelection)
+        self.cluster_table.setEditTriggers(QAbstractItemView.NoEditTriggers)
+        self.cluster_table.setHorizontalHeaderLabels(["Thumbnail", "Photos"])
+        configure_table_columns(self.cluster_table, [120, 80])
+        self.cluster_table.currentCellChanged.connect(self.cluster_selected)
+        splitter.addWidget(self.cluster_table)
+
+        self.photo_table = QTableWidget(0, 3)
+        self.photo_table.setIconSize(QSize(96, 72))
+        self.photo_table.setSelectionBehavior(QAbstractItemView.SelectRows)
+        self.photo_table.setSelectionMode(QAbstractItemView.SingleSelection)
+        self.photo_table.setEditTriggers(QAbstractItemView.NoEditTriggers)
+        self.photo_table.setHorizontalHeaderLabels(["Thumbnail", "File", "Faces in cluster"])
+        configure_table_columns(self.photo_table, [120, 360, 130])
+        self.photo_table.cellDoubleClicked.connect(self.open_photo)
+        splitter.addWidget(self.photo_table)
+        splitter.setStretchFactor(0, 2)
+        splitter.setStretchFactor(1, 3)
+        self.content.addWidget(splitter, 1)
+
+    def _button(self, text: str, callback) -> QPushButton:
+        button = QPushButton(text)
+        button.clicked.connect(callback)
+        set_button_tooltip(button)
+        return button
+
+    def add_folder(self) -> None:
+        folder = QFileDialog.getExistingDirectory(
+            self,
+            tr("Select album folder", self.context.config.ui_language),
+            str(Path.home()),
+        )
+        if folder:
+            self.folder_list.add_folder(folder)
+
+    def remove_selected(self) -> None:
+        for item in self.folder_list.selectedItems():
+            self.folder_list.takeItem(self.folder_list.row(item))
+        self._save_directories()
+
+    def clear_directories(self) -> None:
+        self.folder_list.clear()
+        self._save_directories()
+        self.set_status("Album directories cleared. Existing clustering results are still available.")
+
+    def import_refresh(self) -> None:
+        self._run_import_refresh(rebuild=False)
+
+    def rebuild_all(self) -> None:
+        folders = [Path(folder) for folder in self.folder_list.folders()]
+        if not folders:
+            reply = QMessageBox.question(
+                self,
+                tr("Rebuild All", self.context.config.ui_language),
+                tr(
+                    "No album directories are selected. Rebuild All will clear saved album clustering results. Continue?",
+                    self.context.config.ui_language,
+                ),
+                QMessageBox.Yes | QMessageBox.No,
+                QMessageBox.No,
+            )
+            if reply == QMessageBox.Yes:
+                self.context.storage.clear_album_results()
+                self.clusters = []
+                self.cluster_items = {}
+                self._populate_clusters()
+                self.set_status("Saved album clustering results were cleared.")
+            return
+        reply = QMessageBox.question(
+            self,
+            tr("Rebuild All", self.context.config.ui_language),
+            tr(
+                "Rebuild All will reprocess every image in the selected album directories and replace saved clustering results. Continue?",
+                self.context.config.ui_language,
+            ),
+            QMessageBox.Yes | QMessageBox.No,
+            QMessageBox.No,
+        )
+        if reply == QMessageBox.Yes:
+            self._run_import_refresh(rebuild=True)
+
+    def _run_import_refresh(self, rebuild: bool = False) -> None:
+        folders = [Path(folder) for folder in self.folder_list.folders()]
+        if not folders:
+            self.show_error("Add one or more album directories first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        self._save_directories()
+        all_paths = []
+        for folder in folders:
+            all_paths.extend(str(path) for path in list_images(folder, recursive=True))
+        if not all_paths:
+            self.show_error("No supported images found in the selected directories.")
+            return
+        cosine_threshold = float(self.cluster_threshold.value())
+        min_samples = int(self.min_samples.value())
+        min_face_size = int(self.min_face_size.value())
+        existing = set() if rebuild else self.context.storage.existing_media_paths(all_paths)
+        new_paths = list(all_paths) if rebuild else [path for path in all_paths if path not in existing]
+
+        def task(progress=None, is_cancelled=None):
+            deleted = 0
+            if rebuild:
+                deleted = self.context.storage.delete_media_items_by_paths(all_paths)
+            imported = 0
+            faces_saved = 0
+            for index, path in enumerate(new_paths):
+                if is_cancelled and is_cancelled():
+                    break
+                image = read_image(path)
+                if image is None:
+                    if progress:
+                        progress(index + 1, len(new_paths), f"Skipped unreadable image: {Path(path).name}")
+                    continue
+                media_id = self.context.storage.add_media_item(
+                    path,
+                    "image",
+                    width=image.shape[1],
+                    height=image.shape[0],
+                    file_size=Path(path).stat().st_size if Path(path).exists() else None,
+                    mtime=Path(path).stat().st_mtime if Path(path).exists() else None,
+                    thumbnail=encode_webp_thumbnail(image, max_side=120, quality=35),
+                    processed_at=timestamp_for_filename(),
+                )
+                for face in self.context.engine.detect_faces(image, source_path=path):
+                    if face.normed_embedding is None:
+                        continue
+                    if self._face_box_size(face.bbox) < min_face_size:
+                        continue
+                    face_image = face.crop if face.crop is not None else crop_bbox(image, face.bbox)
+                    self.context.storage.add_media_face(
+                        media_id,
+                        face.normed_embedding,
+                        crop_path="",
+                        thumbnail=encode_webp_thumbnail(face_image, max_side=80, quality=70),
+                        bbox=face.bbox,
+                        kps=face.kps,
+                        det_score=face.det_score,
+                        quality_score=face.quality_score,
+                        status="unknown",
+                    )
+                    faces_saved += 1
+                imported += 1
+                if progress:
+                    progress(
+                        index + 1,
+                        max(1, len(new_paths)),
+                        f"Imported {imported} new images, saved {faces_saved} faces",
+                    )
+            faces = self._faces_for_folders(folders, min_face_size)
+            clusters, algorithm = self._cluster_faces(faces, cosine_threshold, min_samples)
+            self.context.storage.save_album_results(
+                clusters,
+                self.cluster_items,
+                algorithm,
+                cluster_threshold=cosine_threshold,
+                min_samples=min_samples,
+                min_face_size=min_face_size,
+            )
+            return {
+                "deleted": deleted,
+                "imported": imported,
+                "faces_saved": faces_saved,
+                "clusters": clusters,
+                "algorithm": algorithm,
+            }
+
+        def done(result):
+            self.clusters = result["clusters"]
+            self.algorithm_label.setText(f"Algorithm: {result['algorithm']}")
+            self._populate_clusters()
+            if rebuild:
+                self.set_status(
+                    f"Rebuilt album from scratch: removed {result['deleted']} indexed image(s), "
+                    f"processed {result['imported']} image(s), saved {result['faces_saved']} face(s), "
+                    f"built {len(self.clusters)} cluster(s)."
+                )
+            else:
+                self.set_status(
+                    f"Imported {result['imported']} new image(s), saved {result['faces_saved']} face(s), "
+                    f"built {len(self.clusters)} cluster(s)."
+                )
+
+        self.run_task("Rebuilding album" if rebuild else "Importing and clustering album", task, done)
+
+    def _faces_for_folders(self, folders: list[Path], min_face_size: int) -> list[dict]:
+        roots = [folder.resolve() for folder in folders]
+        faces = []
+        for face in self.context.storage.list_media_faces():
+            try:
+                media_path = Path(face["media_path"]).resolve()
+            except Exception:
+                continue
+            if any(media_path == root or root in media_path.parents for root in roots):
+                if (
+                    face.get("embedding") is not None
+                    and self._face_box_size(face.get("bbox")) >= min_face_size
+                ):
+                    faces.append(face)
+        return faces
+
+    def _cluster_faces(
+        self,
+        faces: list[dict],
+        cosine_threshold: float,
+        min_samples: int,
+    ) -> tuple[list[dict], str]:
+        embeddings = [face["embedding"] for face in faces]
+        distance_threshold = max(0.01, 1.0 - float(cosine_threshold))
+        labels, algorithm = cluster_embeddings_dbscan(
+            embeddings,
+            distance_threshold=distance_threshold,
+            min_samples=min_samples,
+        )
+        groups: dict[int, list[dict]] = defaultdict(list)
+        next_noise = max(labels, default=-1) + 1
+        for face, label in zip(faces, labels):
+            if label < 0:
+                label = next_noise
+                next_noise += 1
+            groups[int(label)].append(face)
+        next_album_id = 1
+        clusters = []
+        self.cluster_items = {}
+        for label, items in groups.items():
+            vectors = [normalize_embedding(item["embedding"]) for item in items if item.get("embedding") is not None]
+            vectors = [vector for vector in vectors if vector is not None]
+            if not vectors:
+                continue
+            centroid = normalize_embedding(np.mean(np.vstack(vectors), axis=0))
+            cluster_id = next_album_id
+            next_album_id += 1
+            name = f"Album Person {cluster_id}"
+            representative = max(items, key=lambda item: cosine_similarity(centroid, item.get("embedding")))
+            photos = sorted({item["media_path"] for item in items})
+            cluster = {
+                "id": cluster_id,
+                "label": label,
+                "name": name,
+                "source": "album",
+                "face_count": len(items),
+                "photo_count": len(photos),
+                "avg_quality": sum(float(item.get("quality_score") or 0.0) for item in items) / max(1, len(items)),
+                "thumbnail_face_id": representative.get("id"),
+                "thumbnail_path": representative.get("media_path"),
+                "photos": photos,
+            }
+            clusters.append(cluster)
+            self.cluster_items[cluster_id] = items
+        return sorted(clusters, key=lambda row: (-row["face_count"], row["id"])), algorithm
+
+    def _save_directories(self) -> None:
+        self.context.storage.save_album_directories(self.folder_list.folders())
+
+    def refresh(self) -> None:
+        if self._loaded_saved_state:
+            return
+        self._loaded_saved_state = True
+        self.folder_list.blockSignals(True)
+        self.folder_list.clear()
+        for folder in self.context.storage.list_album_directories():
+            if Path(folder).expanduser().is_dir():
+                self.folder_list.addItem(folder)
+        self.folder_list.blockSignals(False)
+        self._load_saved_results()
+
+    def _load_saved_results(self) -> None:
+        data = self.context.storage.load_album_results()
+        clusters = data.get("clusters") if isinstance(data, dict) else None
+        if not isinstance(clusters, list):
+            return
+        faces_by_id = {int(face["id"]): face for face in self.context.storage.list_media_faces() if face.get("id") is not None}
+        self.clusters = []
+        self.cluster_items = {}
+        for cluster in clusters:
+            try:
+                cluster_id = int(cluster["id"])
+            except Exception:
+                continue
+            face_ids = [int(face_id) for face_id in cluster.get("face_ids", []) if str(face_id).isdigit()]
+            self.cluster_items[cluster_id] = [faces_by_id[face_id] for face_id in face_ids if face_id in faces_by_id]
+            self.clusters.append(cluster)
+        algorithm = data.get("algorithm")
+        if algorithm not in {"DBSCAN", "centroid fallback", "none"}:
+            algorithm = "DBSCAN"
+        self.algorithm_label.setText(f"Algorithm: {algorithm}")
+        if data.get("cluster_threshold") is not None:
+            self.cluster_threshold.setValue(float(data["cluster_threshold"]))
+        if data.get("min_samples") is not None:
+            self.min_samples.setValue(int(data["min_samples"]))
+        if data.get("min_face_size") is not None:
+            self.min_face_size.setValue(int(data["min_face_size"]))
+        self._populate_clusters()
+
+    def _populate_clusters(self) -> None:
+        self.cluster_table.setRowCount(len(self.clusters))
+        for row, cluster in enumerate(self.clusters):
+            values = [
+                "",
+                cluster.get("photo_count", 0),
+            ]
+            for col, value in enumerate(values):
+                item = QTableWidgetItem(str(value))
+                if col == 0:
+                    icon = self._cluster_thumbnail_icon(cluster, QSize(56, 56))
+                    if icon:
+                        item.setIcon(icon)
+                item.setData(Qt.UserRole, cluster.get("id"))
+                item.setTextAlignment(Qt.AlignCenter)
+                item.setToolTip(
+                    f"ID: {cluster.get('id', '')}\n"
+                    f"Name: {cluster.get('name', '')}\n"
+                    f"Faces: {cluster.get('face_count', 0)}\n"
+                    f"Photos: {cluster.get('photo_count', 0)}"
+                )
+                self.cluster_table.setItem(row, col, item)
+            self.cluster_table.setRowHeight(row, 64)
+        refresh_table_columns(self.cluster_table)
+        if self.clusters:
+            self.cluster_table.selectRow(0)
+            self._populate_photos(int(self.clusters[0].get("id", 0)))
+        else:
+            self.photo_table.setRowCount(0)
+
+    def cluster_selected(self, current_row: int, current_column: int, previous_row: int, previous_column: int) -> None:
+        del current_column, previous_row, previous_column
+        if current_row < 0 or current_row >= len(self.clusters):
+            self.photo_table.setRowCount(0)
+            return
+        self._populate_photos(int(self.clusters[current_row].get("id", 0)))
+
+    def _populate_photos(self, cluster_id: int) -> None:
+        items = self.cluster_items.get(cluster_id, [])
+        grouped: dict[str, dict] = {}
+        for item in items:
+            path = item["media_path"]
+            if path not in grouped:
+                grouped[path] = {"count": 0, "thumbnail": item.get("media_thumbnail")}
+            grouped[path]["count"] += 1
+        if not grouped:
+            for cluster in self.clusters:
+                if int(cluster.get("id", -1)) == cluster_id:
+                    for path in cluster.get("photos", []):
+                        grouped[str(path)] = {"count": 1, "thumbnail": None}
+                    break
+        rows = sorted(grouped.items())
+        self.photo_table.setRowCount(len(rows))
+        for row, (path, data) in enumerate(rows):
+            thumb = QTableWidgetItem("")
+            icon = self._icon_from_bytes(data.get("thumbnail"), QSize(96, 72))
+            if icon:
+                thumb.setIcon(icon)
+            thumb.setData(Qt.UserRole, path)
+            file_item = QTableWidgetItem(path)
+            file_item.setData(Qt.UserRole, path)
+            self.photo_table.setItem(row, 0, thumb)
+            self.photo_table.setItem(row, 1, file_item)
+            self.photo_table.setItem(row, 2, QTableWidgetItem(str(data["count"])))
+            self.photo_table.setRowHeight(row, 80)
+        refresh_table_columns(self.photo_table)
+
+    def open_photo(self, row: int, column: int) -> None:
+        item = self.photo_table.item(row, 0) or self.photo_table.item(row, 1)
+        if item is None:
+            return
+        path = item.data(Qt.UserRole)
+        if path:
+            QDesktopServices.openUrl(QUrl.fromLocalFile(str(path)))
+
+    @staticmethod
+    def _face_box_size(bbox) -> float:
+        try:
+            x1, y1, x2, y2 = [float(value) for value in bbox[:4]]
+        except Exception:
+            return 0.0
+        return min(abs(x2 - x1), abs(y2 - y1))
+
+    def _cluster_thumbnail_icon(self, cluster: dict, size: QSize) -> QIcon | None:
+        thumbnail_face_id = cluster.get("thumbnail_face_id")
+        if thumbnail_face_id is not None:
+            for item in self.cluster_items.get(int(cluster.get("id", -1)), []):
+                if item.get("id") == thumbnail_face_id:
+                    return self._icon_from_bytes(item.get("thumbnail"), size)
+        for item in self.cluster_items.get(int(cluster.get("id", -1)), []):
+            icon = self._icon_from_bytes(item.get("thumbnail"), size)
+            if icon is not None:
+                return icon
+        return None
+
+    @staticmethod
+    def _icon_from_bytes(data, size: QSize) -> QIcon | None:
+        if not data:
+            return None
+        try:
+            from PIL import Image
+
+            with Image.open(BytesIO(bytes(data))) as image:
+                rgba = image.convert("RGBA")
+                raw = rgba.tobytes("raw", "RGBA")
+                qimage = QImage(raw, rgba.width, rgba.height, QImage.Format_RGBA8888).copy()
+            pixmap = QPixmap.fromImage(qimage)
+        except Exception:
+            return None
+        return QIcon(pixmap.scaled(size, Qt.KeepAspectRatio, Qt.SmoothTransformation))
diff --git a/python-package/insightface/gui/pages/album_people_page.py b/python-package/insightface/gui/pages/album_people_page.py
new file mode 100644
index 000000000..f12a4ee07
--- /dev/null
+++ b/python-package/insightface/gui/pages/album_people_page.py
@@ -0,0 +1,98 @@
+"""Album people clustering page."""
+
+from __future__ import annotations
+
+from collections import defaultdict
+from pathlib import Path
+
+from PySide6.QtWidgets import QLabel, QDoubleSpinBox, QTableWidget, QTableWidgetItem
+
+from ..core.clustering import cluster_embeddings_dbscan
+from ..core.exporters import export_csv
+from ..core.utils import list_images, read_image, timestamp_for_filename
+from ..widgets.drop_input import DropInput
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+
+class AlbumPeoplePage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Album People Clustering", "Approximate local person grouping for album review and correction.", parent)
+        self.folder = ""
+        self.rows = []
+        self.threshold = QDoubleSpinBox()
+        self.threshold.setRange(0.01, 0.99)
+        self.threshold.setSingleStep(0.01)
+        self.threshold.setValue(0.28)
+        self.content.addWidget(self.notice("Clustering is approximate. Please review and correct groups manually."))
+        self.folder_input = DropInput("Album Folder", mode="folder")
+        self.folder_input.pathsChanged.connect(lambda paths: self.set_folder(paths[0]) if paths else self.clear_folder())
+        self.content.addWidget(self.folder_input)
+        self.content.addWidget(self.row(QLabel("DBSCAN distance threshold"), self.threshold, self.button("Scan and Cluster", self.scan), self.button("Export CSV", self.export)))
+        self.table = QTableWidget(0, 5)
+        self.table.setHorizontalHeaderLabels(["cluster_id", "face_count", "photo_count", "average_quality", "sample_paths"])
+        configure_table_columns(self.table, [100, 100, 100, 120, 360])
+        self.content.addWidget(self.table, 1)
+
+    def set_folder(self, folder: str) -> None:
+        self.folder = folder
+        self.set_status(f"Selected album: {folder}")
+
+    def clear_folder(self) -> None:
+        self.folder = ""
+        self.rows = []
+        self.table.setRowCount(0)
+
+    def scan(self) -> None:
+        if not self.folder:
+            self.show_error("Select an album folder first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        paths = list_images(self.folder, recursive=True)
+
+        def task(progress=None, is_cancelled=None):
+            faces = []
+            embeddings = []
+            for index, path in enumerate(paths):
+                if is_cancelled and is_cancelled():
+                    break
+                image = read_image(path)
+                if image is None:
+                    continue
+                for face in self.context.engine.detect_faces(image, source_path=str(path)):
+                    if face.normed_embedding is not None:
+                        embeddings.append(face.normed_embedding)
+                        faces.append({"path": str(path), "quality": face.quality_score or 0.0, "bbox": face.bbox})
+                if progress:
+                    progress(index + 1, len(paths), f"Scanned {index + 1}/{len(paths)} photos")
+            labels, _algorithm = cluster_embeddings_dbscan(embeddings, distance_threshold=self.threshold.value()) if embeddings else ([], "none")
+            clusters = defaultdict(list)
+            for face, label in zip(faces, labels):
+                clusters[int(label)].append(face)
+            rows = []
+            for label, items in clusters.items():
+                rows.append({"cluster_id": label, "face_count": len(items), "photo_count": len({item["path"] for item in items}), "average_quality": sum(item["quality"] for item in items) / max(1, len(items)), "sample_paths": "; ".join(item["path"] for item in items[:5])})
+            return rows
+
+        self.run_task("Album clustering", task, lambda rows: (setattr(self, "rows", rows), self._populate(), self.set_status(f"Created {len(rows)} cluster(s).")))
+
+    def _populate(self) -> None:
+        columns = ["cluster_id", "face_count", "photo_count", "average_quality", "sample_paths"]
+        self.table.setRowCount(len(self.rows))
+        for row_index, row in enumerate(self.rows):
+            for col, key in enumerate(columns):
+                value = row[key]
+                if isinstance(value, float):
+                    value = f"{value:.4f}"
+                self.table.setItem(row_index, col, QTableWidgetItem(str(value)))
+        refresh_table_columns(self.table)
+
+    def export(self) -> None:
+        if not self.rows:
+            self.show_error("No clusters to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"album_clusters_{timestamp_for_filename()}.csv"
+        export_csv(path, self.rows)
+        self.set_status(f"Clusters exported to {path}")
diff --git a/python-package/insightface/gui/pages/base.py b/python-package/insightface/gui/pages/base.py
new file mode 100644
index 000000000..c1579217a
--- /dev/null
+++ b/python-package/insightface/gui/pages/base.py
@@ -0,0 +1,114 @@
+"""Base page helpers."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Callable, Optional
+
+from PySide6.QtCore import Qt
+from PySide6.QtWidgets import (
+    QFileDialog,
+    QFrame,
+    QHBoxLayout,
+    QLabel,
+    QMessageBox,
+    QPushButton,
+    QVBoxLayout,
+    QWidget,
+)
+
+from ..core.constants import LOCAL_PROCESSING_NOTICE
+from ..core.i18n import tr
+from ..core.tooltips import set_button_tooltip
+
+
+class BasePage(QWidget):
+    def __init__(self, context, title: str, description: str = "", parent=None):
+        super().__init__(parent)
+        self.context = context
+        self.title = title
+        self.root_layout = QVBoxLayout(self)
+        self.root_layout.setSpacing(12)
+        header = QLabel(title)
+        header.setStyleSheet("font-size: 24px; font-weight: 700;")
+        self.root_layout.addWidget(header)
+        if description:
+            desc = QLabel(description)
+            desc.setWordWrap(True)
+            desc.setProperty("role", "muted")
+            self.root_layout.addWidget(desc)
+        self.status_label = QLabel("")
+        self.status_label.setWordWrap(True)
+        self.status_label.setProperty("role", "status")
+        self.content = QVBoxLayout()
+        self.root_layout.addLayout(self.content, 1)
+        self.root_layout.addWidget(self.status_label)
+
+    def notice(self, text: str = LOCAL_PROCESSING_NOTICE) -> QLabel:
+        label = QLabel(text)
+        label.setObjectName("noticeLabel")
+        label.setWordWrap(True)
+        return label
+
+    def card(self) -> tuple[QFrame, QVBoxLayout]:
+        frame = QFrame()
+        frame.setFrameShape(QFrame.StyledPanel)
+        layout = QVBoxLayout(frame)
+        return frame, layout
+
+    def row(self, *widgets) -> QWidget:
+        wrapper = QWidget()
+        layout = QHBoxLayout(wrapper)
+        layout.setContentsMargins(0, 0, 0, 0)
+        for widget in widgets:
+            layout.addWidget(widget)
+        layout.addStretch(1)
+        return wrapper
+
+    def button(self, text: str, callback: Callable, enabled: bool = True) -> QPushButton:
+        btn = QPushButton(text)
+        btn.clicked.connect(callback)
+        btn.setEnabled(enabled)
+        set_button_tooltip(btn)
+        return btn
+
+    def choose_file(self, caption: str, filters: str = "Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)") -> Optional[str]:
+        path, _ = QFileDialog.getOpenFileName(self, caption, str(Path.home()), filters)
+        return path or None
+
+    def choose_files(self, caption: str, filters: str = "Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)") -> list[str]:
+        paths, _ = QFileDialog.getOpenFileNames(self, caption, str(Path.home()), filters)
+        return paths
+
+    def choose_folder(self, caption: str) -> Optional[str]:
+        path = QFileDialog.getExistingDirectory(self, caption, str(Path.home()))
+        return path or None
+
+    def save_file(self, caption: str, default_name: str, filters: str) -> Optional[str]:
+        path, _ = QFileDialog.getSaveFileName(self, caption, default_name, filters)
+        return path or None
+
+    def set_status(self, message: str) -> None:
+        self.status_label.setText(tr(message, self.context.config.ui_language))
+        main = self.window()
+        if main is not self and hasattr(main, "set_status"):
+            main.set_status(message)
+
+    def show_error(self, message: str) -> None:
+        self.set_status(message)
+        QMessageBox.warning(self, tr(self.title, self.context.config.ui_language), tr(message, self.context.config.ui_language))
+
+    def run_task(self, title: str, fn: Callable, on_result: Callable | None = None, show_dialog: bool = True) -> None:
+        main = self.window()
+        if hasattr(main, "run_task"):
+            main.run_task(title, fn, on_result, show_dialog=show_dialog)
+        else:
+            try:
+                result = fn()
+                if on_result:
+                    on_result(result)
+            except Exception as exc:
+                self.show_error(str(exc))
+
+    def refresh(self) -> None:
+        pass
diff --git a/python-package/insightface/gui/pages/batch_processing_page.py b/python-package/insightface/gui/pages/batch_processing_page.py
new file mode 100644
index 000000000..1b2679cf2
--- /dev/null
+++ b/python-package/insightface/gui/pages/batch_processing_page.py
@@ -0,0 +1,117 @@
+"""Batch folder processing page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtWidgets import QCheckBox, QLabel, QProgressBar, QSpinBox, QTableWidget, QTableWidgetItem
+
+from ..core.exporters import export_csv, export_json
+from ..core.utils import list_images, read_image, save_image, timestamp_for_filename
+from ..widgets.drop_input import DropInput
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+
+class BatchProcessingPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Batch Folder Processing", "Scan a folder, detect faces, optionally identify people, and export batch results.", parent)
+        self.folder = ""
+        self.rows = []
+        self.recursive = QCheckBox("Recursive")
+        self.recursive.setChecked(True)
+        self.identify = QCheckBox("Identify people")
+        self.identify.setChecked(True)
+        self.save_crops = QCheckBox("Save face crops")
+        self.save_crops.setChecked(context.config.save_crops)
+        self.min_score = QSpinBox()
+        self.min_score.setRange(0, 100)
+        self.min_score.setValue(int(context.config.min_detection_score * 100))
+        self.progress = QProgressBar()
+        self.table = QTableWidget(0, 10)
+        self.table.setHorizontalHeaderLabels(["image_path", "face_index", "person_id", "person_name", "similarity", "threshold", "status", "bbox", "det_score", "crop_path"])
+        configure_table_columns(self.table, [300, 80, 80, 160, 90, 90, 110, 220, 90, 260])
+        self.folder_input = DropInput("Image Folder", mode="folder")
+        self.folder_input.pathsChanged.connect(lambda paths: self.set_folder(paths[0]) if paths else self.clear_folder())
+        self.content.addWidget(self.folder_input)
+        self.content.addWidget(self.row(self.recursive, self.identify, self.save_crops, QLabel("Min det score %"), self.min_score, self.button("Start", self.start), self.button("Export", self.export)))
+        self.content.addWidget(self.progress)
+        self.content.addWidget(self.table, 1)
+
+    def set_folder(self, folder: str) -> None:
+        self.folder = folder
+        self.set_status(f"Selected folder: {folder}")
+
+    def clear_folder(self) -> None:
+        self.folder = ""
+        self.rows = []
+        self.progress.setValue(0)
+        self.table.setRowCount(0)
+
+    def start(self) -> None:
+        if not self.folder:
+            self.show_error("Select a folder first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        paths = [str(path) for path in list_images(self.folder, recursive=self.recursive.isChecked())]
+        min_score = self.min_score.value() / 100.0
+
+        def task(progress=None, is_cancelled=None):
+            rows = []
+            gallery = self.context.storage.load_all_gallery_embeddings() if self.identify.isChecked() else []
+            for index, path in enumerate(paths):
+                if is_cancelled and is_cancelled():
+                    break
+                image = read_image(path)
+                if image is None:
+                    rows.append({"image_path": path, "status": "error", "error": "image read failure"})
+                    continue
+                media_id = self.context.storage.add_media_item(path, "image", width=image.shape[1], height=image.shape[0], processed_at=timestamp_for_filename())
+                faces = [face for face in self.context.engine.detect_faces(image, source_path=path) if face.det_score >= min_score]
+                for face_index, face in enumerate(faces):
+                    result = None
+                    if face.normed_embedding is not None and gallery:
+                        matches = self.context.storage.search_embeddings(face.normed_embedding, self.context.config.default_top_k, self.context.config.recognition_threshold)
+                        result = matches[0] if matches else None
+                    person_name = result.person_name if result and result.similarity >= self.context.config.recognition_threshold else "Unknown"
+                    similarity = result.similarity if result else 0.0
+                    status = "matched" if result and result.similarity >= self.context.config.recognition_threshold else "unknown"
+                    crop_path = ""
+                    if self.save_crops.isChecked() and face.crop is not None:
+                        crop_path = str(Path(self.context.config.crop_dir) / f"batch_{media_id}_{face_index}_{timestamp_for_filename()}.png")
+                        save_image(crop_path, face.crop)
+                    self.context.storage.add_media_face(media_id, face.normed_embedding, crop_path=crop_path, bbox=face.bbox, kps=face.kps, det_score=face.det_score, predicted_person_id=result.person_id if result else None, similarity=similarity, status=status)
+                    rows.append({"image_path": path, "face_index": face_index, "person_id": result.person_id if result else "", "person_name": person_name, "similarity": similarity, "threshold": self.context.config.recognition_threshold, "status": status, "bbox": face.bbox, "det_score": face.det_score, "crop_path": crop_path})
+                if progress:
+                    progress(index + 1, len(paths), f"Processed {index + 1}/{len(paths)} files")
+            return rows
+
+        def done(rows):
+            self.rows = rows
+            self.progress.setValue(100)
+            self._populate()
+            self.set_status(f"Batch complete. {len(rows)} face row(s).")
+
+        self.run_task("Batch folder processing", task, done)
+
+    def _populate(self) -> None:
+        columns = ["image_path", "face_index", "person_id", "person_name", "similarity", "threshold", "status", "bbox", "det_score", "crop_path"]
+        self.table.setRowCount(len(self.rows))
+        for row_index, row in enumerate(self.rows):
+            for col, key in enumerate(columns):
+                value = row.get(key, "")
+                if isinstance(value, float):
+                    value = f"{value:.4f}"
+                self.table.setItem(row_index, col, QTableWidgetItem(str(value)))
+        refresh_table_columns(self.table)
+
+    def export(self) -> None:
+        if not self.rows:
+            self.show_error("No batch results to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"batch_results_{timestamp_for_filename()}.csv"
+        export_csv(path, self.rows)
+        export_json(path.with_suffix(".json"), self.rows)
+        self.set_status(f"Batch results exported to {path}")
diff --git a/python-package/insightface/gui/pages/camera_recognition_page.py b/python-package/insightface/gui/pages/camera_recognition_page.py
new file mode 100644
index 000000000..d74991837
--- /dev/null
+++ b/python-package/insightface/gui/pages/camera_recognition_page.py
@@ -0,0 +1,32 @@
+"""Camera recognition page."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QLabel, QSpinBox
+
+from ..core.camera import list_camera_indices
+from ..core.constants import RESPONSIBLE_USE_NOTICE
+from ..widgets.threshold_slider import ThresholdSlider
+from .base import BasePage
+
+
+class CameraRecognitionPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Camera Recognition", "Live recognition preview for lawful, consent-based controlled environments.", parent)
+        self.camera_index = QSpinBox()
+        self.camera_index.setRange(0, 16)
+        self.frame_skip = QSpinBox()
+        self.frame_skip.setRange(1, 60)
+        self.frame_skip.setValue(context.config.camera_frame_skip)
+        self.threshold = ThresholdSlider(context.config.recognition_threshold)
+        self.content.addWidget(self.row(QLabel("Camera index"), self.camera_index, QLabel("Frame skip"), self.frame_skip, self.button("Check Cameras", self.check), self.button("Start", self.unavailable), self.button("Stop", self.unavailable), self.button("Register Current Face", self.unavailable)))
+        self.content.addWidget(self.threshold)
+        self.content.addWidget(self.notice(RESPONSIBLE_USE_NOTICE + " Live recognition should only be used with consent and in lawful, controlled environments."))
+        self.content.addStretch(1)
+
+    def check(self) -> None:
+        indices = list_camera_indices()
+        self.set_status("Available camera indices: " + (", ".join(map(str, indices)) if indices else "none detected"))
+
+    def unavailable(self) -> None:
+        self.show_error("Live camera loop is not enabled in this first desktop build. Camera settings and responsible use notice are available.")
diff --git a/python-package/insightface/gui/pages/compare_page.py b/python-package/insightface/gui/pages/compare_page.py
new file mode 100644
index 000000000..9db0c3d86
--- /dev/null
+++ b/python-package/insightface/gui/pages/compare_page.py
@@ -0,0 +1,195 @@
+"""1:1 face compare page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtWidgets import QSplitter, QTextEdit, QWidget, QVBoxLayout
+from PySide6.QtCore import Qt
+
+from ..core.exporters import export_json, export_markdown
+from ..core.utils import read_image, timestamp_for_filename
+from ..widgets.threshold_slider import ThresholdSlider
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class ComparePage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "1:1 Face Compare", "Compare two local images and export a structured result.", parent)
+        self.path_a = ""
+        self.path_b = ""
+        self.image_a = None
+        self.image_b = None
+        self._compare_generation = 0
+        self._compare_running = False
+        image_filter = "Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)"
+        image_exts = [".jpg", ".jpeg", ".png", ".bmp", ".webp"]
+        self.input_a = UploadPreview("Image A", extensions=image_exts, dialog_filter=image_filter)
+        self.input_b = UploadPreview("Image B", extensions=image_exts, dialog_filter=image_filter)
+        self.input_a.pathChanged.connect(lambda path: self.load_a(path) if path else self.clear_a())
+        self.input_b.pathChanged.connect(lambda path: self.load_b(path) if path else self.clear_b())
+        self.compare_button = self.button("Compare", self.compare)
+        controls = self.row(
+            self.compare_button,
+            self.button("Reset", self.reset),
+            self.button("Export Result", self.export_result),
+        )
+        self.content.addWidget(controls)
+        self.threshold = ThresholdSlider(context.config.recognition_threshold)
+        self.content.addWidget(self.threshold)
+        splitter = QSplitter(Qt.Horizontal)
+        left = QWidget()
+        left_layout = QVBoxLayout(left)
+        self.viewer_a = self.input_a
+        left_layout.addWidget(self.viewer_a)
+        right = QWidget()
+        right_layout = QVBoxLayout(right)
+        self.viewer_b = self.input_b
+        right_layout.addWidget(self.viewer_b)
+        splitter.addWidget(left)
+        splitter.addWidget(right)
+        self.content.addWidget(splitter, 1)
+        self.result_text = QTextEdit()
+        self.result_text.setReadOnly(True)
+        self.content.addWidget(self.result_text)
+        self.result = None
+
+    def load_a(self, path: str) -> None:
+        self.path_a = path
+        self.image_a = read_image(path)
+        if self.image_a is None:
+            self.viewer_a.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self.viewer_a.set_image(self.image_a, path)
+        self._invalidate_result("Image A replaced.")
+
+    def load_b(self, path: str) -> None:
+        self.path_b = path
+        self.image_b = read_image(path)
+        if self.image_b is None:
+            self.viewer_b.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self.viewer_b.set_image(self.image_b, path)
+        self._invalidate_result("Image B replaced.")
+
+    def clear_a(self) -> None:
+        self.path_a = ""
+        self.image_a = None
+        self.viewer_a.set_image(None)
+        self._invalidate_result()
+
+    def clear_b(self) -> None:
+        self.path_b = ""
+        self.image_b = None
+        self.viewer_b.set_image(None)
+        self._invalidate_result()
+
+    def compare(self) -> None:
+        if self._compare_running:
+            self.set_status("Compare is already running.")
+            return
+        if self.image_a is None or self.image_b is None:
+            self.show_error("Please select both images first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        image_a = self.image_a.copy()
+        image_b = self.image_b.copy()
+        path_a = self.path_a
+        path_b = self.path_b
+        threshold = self.threshold.value()
+        self._compare_generation += 1
+        generation = self._compare_generation
+        self._compare_running = True
+        self.compare_button.setEnabled(False)
+
+        def task():
+            try:
+                result = self.context.engine.compare_images(
+                    image_a,
+                    image_b,
+                    threshold=threshold,
+                    path1=path_a,
+                    path2=path_b,
+                )
+                return {"generation": generation, "result": result, "error": ""}
+            except Exception as exc:
+                return {"generation": generation, "result": None, "error": str(exc)}
+
+        def done(payload):
+            self._compare_running = False
+            self.compare_button.setEnabled(True)
+            if payload["generation"] != self._compare_generation:
+                self.set_status("Ignored stale compare result.")
+                return
+            if payload["error"]:
+                self.show_error(payload["error"])
+                return
+            result = payload["result"]
+            self.result = result
+            self.viewer_a.set_faces([{"bbox": result.face_a.bbox, "label": "Face A"}])
+            self.viewer_b.set_faces([{"bbox": result.face_b.bbox, "label": "Face B"}])
+            self.result_text.setPlainText(
+                "\n".join(
+                    [
+                        f"Similarity: {result.similarity:.4f}",
+                        f"Threshold: {result.threshold:.4f}",
+                        f"Decision: {result.decision}",
+                        f"Det score A: {result.face_a.det_score:.4f}",
+                        f"Det score B: {result.face_b.det_score:.4f}",
+                        "Notes: " + ("; ".join(result.notes) if result.notes else "Both faces detected successfully."),
+                    ]
+                )
+            )
+            self.set_status(f"Compare complete: {result.decision}")
+
+        self.run_task("Comparing faces", task, done, show_dialog=False)
+
+    def reset(self) -> None:
+        self.path_a = self.path_b = ""
+        self.image_a = self.image_b = None
+        self.result = None
+        self.input_a.clear(emit=False)
+        self.input_b.clear(emit=False)
+        self._invalidate_result("Compare page reset.")
+        self._compare_running = False
+        self.compare_button.setEnabled(True)
+        self.set_status("Compare page reset.")
+
+    def _invalidate_result(self, status: str = "") -> None:
+        self._compare_generation += 1
+        self.result = None
+        self.result_text.clear()
+        self.viewer_a.set_faces([])
+        self.viewer_b.set_faces([])
+        if status:
+            self.set_status(status)
+
+    def export_result(self) -> None:
+        if self.result is None:
+            self.show_error("No compare result to export.")
+            return
+        default = str(Path(self.context.config.export_dir) / f"compare_{timestamp_for_filename()}.json")
+        path = self.save_file("Export Result", default, "JSON (*.json);;Markdown (*.md)")
+        if not path:
+            return
+        if path.lower().endswith(".md"):
+            text = "\n".join(
+                [
+                    "# InsightFace 1:1 Compare Result",
+                    f"- Image A: {self.path_a}",
+                    f"- Image B: {self.path_b}",
+                    f"- Similarity: {self.result.similarity:.4f}",
+                    f"- Threshold: {self.result.threshold:.4f}",
+                    f"- Decision: {self.result.decision}",
+                    f"- Notes: {'; '.join(self.result.notes)}",
+                ]
+            )
+            export_markdown(path, text)
+        else:
+            export_json(path, self.result.to_json_dict())
+        self.set_status(f"Result exported to {path}")
diff --git a/python-package/insightface/gui/pages/dashboard_page.py b/python-package/insightface/gui/pages/dashboard_page.py
new file mode 100644
index 000000000..a100da6e8
--- /dev/null
+++ b/python-package/insightface/gui/pages/dashboard_page.py
@@ -0,0 +1,59 @@
+"""Dashboard page."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QGridLayout, QPushButton, QWidget
+
+from ..core.constants import LOCAL_PROCESSING_NOTICE, SUBTITLE
+from ..core.tooltips import set_button_tooltip
+from ..widgets.metric_card import MetricCard
+from .base import BasePage
+
+
+class DashboardPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Dashboard", "Welcome to InsightFace Evaluation Studio.\n" + SUBTITLE, parent)
+        self.content.addWidget(self.notice(LOCAL_PROCESSING_NOTICE))
+        grid_widget = QWidget()
+        self.grid = QGridLayout(grid_widget)
+        self.cards = {
+            "workspace": MetricCard("Workspace", context.config.workspace_path),
+            "model": MetricCard("Model", context.config.model_name),
+            "provider": MetricCard("Provider", context.config.provider),
+            "license": MetricCard("License", context.config.license_status),
+            "people": MetricCard("People", "0"),
+            "samples": MetricCard("Face samples", "0"),
+            "media": MetricCard("Indexed photos", "0"),
+            "faces": MetricCard("Detected faces", "0"),
+        }
+        for index, card in enumerate(self.cards.values()):
+            self.grid.addWidget(card, index // 2, index % 2)
+        self.content.addWidget(grid_widget)
+        shortcuts = QWidget()
+        shortcut_layout = QGridLayout(shortcuts)
+        for index, (label, page) in enumerate(
+            [
+                ("Start 1:1 Compare", "1:1 Compare"),
+                ("Add People", "People Library"),
+                ("Scan Folder", "Batch Folder Processing"),
+                ("Run Enterprise Evaluation", "Enterprise Evaluation"),
+                ("Open License Center", "License Center"),
+            ]
+        ):
+            button = QPushButton(label)
+            button.clicked.connect(lambda checked=False, target=page: self.window().open_page(target))
+            set_button_tooltip(button)
+            shortcut_layout.addWidget(button, index // 3, index % 3)
+        self.content.addWidget(shortcuts)
+        self.content.addStretch(1)
+
+    def refresh(self) -> None:
+        counts = self.context.storage.counts()
+        self.cards["workspace"].set_value(self.context.config.workspace_path)
+        self.cards["model"].set_value(self.context.config.model_name)
+        self.cards["provider"].set_value(self.context.config.provider)
+        self.cards["license"].set_value(self.context.config.license_status)
+        self.cards["people"].set_value(counts["people"])
+        self.cards["samples"].set_value(counts["face_samples"])
+        self.cards["media"].set_value(counts["media_items"])
+        self.cards["faces"].set_value(counts["media_faces"])
diff --git a/python-package/insightface/gui/pages/enterprise_eval_page.py b/python-package/insightface/gui/pages/enterprise_eval_page.py
new file mode 100644
index 000000000..811400eb5
--- /dev/null
+++ b/python-package/insightface/gui/pages/enterprise_eval_page.py
@@ -0,0 +1,1255 @@
+"""Single-page enterprise 1:1 and 1:N evaluation workflow."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import QEvent, Qt
+from PySide6.QtWidgets import (
+    QCheckBox,
+    QComboBox,
+    QDialog,
+    QDialogButtonBox,
+    QFrame,
+    QGridLayout,
+    QHBoxLayout,
+    QLabel,
+    QMessageBox,
+    QTextEdit,
+    QVBoxLayout,
+    QWidget,
+)
+
+from ..core.evaluation import (
+    MULTI_FACE_REQUIRE_ONE,
+    MULTI_FACE_SKIP,
+    MULTI_FACE_USE_CENTERED_LARGEST,
+    MULTI_FACE_USE_LARGEST,
+    run_identity_identification_evaluation,
+    run_identity_verification_evaluation,
+    validate_enterprise_dataset,
+)
+from ..core.i18n import apply_translations, effective_language, tr
+from ..core.reporting import write_reports
+from ..widgets.drop_input import DropInput
+from .base import BasePage
+
+
+DATASET_RULES_TEXT = """Enterprise Evaluation Dataset Rules
+
+The Enterprise Evaluation page supports local 1:1 verification and 1:N identification evaluation from identity folders. No images, embeddings, or reports are uploaded automatically.
+
+1:1 Verification
+
+Auto Split enabled:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Each subfolder is one identity. A file whose name contains "gallery" is selected as that identity's gallery image. If no such file exists, the first sorted image is selected as gallery. All other images are probes. Evaluation compares every probe against every identity gallery, so each trial is probe vs gallery.
+
+Auto Split disabled:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Every image is treated as a probe. Evaluation runs full pairwise probe-vs-probe comparisons across all identity folders.
+
+1:N Identification
+
+Auto Split enabled:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+The identities/ folder is preferred when present. If it is not present, the selected dataset root may contain identity folders directly. Gallery selection follows the same Auto Split rule as 1:1.
+
+Auto Split disabled:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ and probe/ are required. unknown/ is optional. 1:N evaluation always requires gallery images.
+
+Validation Checks
+
+- The selected root exists and matches the selected mode.
+- Required gallery and probe folders/images exist.
+- Auto Split can create valid gallery/probe sets.
+- 1:1 can generate both positive and negative pairs.
+- 1:N probe identities have valid gallery coverage.
+- Each image can be read.
+- Detected face count follows the selected multi-face handling policy.
+
+Multi-face Handling
+
+- Require exactly one face: validation fails on multi-face images.
+- Use largest face: evaluation uses the largest detected face.
+- Use largest centered face: evaluation uses area - center_distance^2 * 2.0.
+- Mark as skip: multi-face images are skipped; skipped gallery images can make validation fail.
+
+Report Outputs
+
+1:1 reports include best cosine threshold accuracy, FAR/FRR, TAR@FAR, threshold recommendations, latency, and raw result examples.
+
+1:N reports include Top1, TAR@FAR, threshold recommendations, latency, and raw result examples.
+"""
+
+DATASET_RULES_LOCAL_TEXT = {
+    "zh": """Enterprise Evaluation 页面支持从身份文件夹进行本地 1:1 验证和 1:N 识别评测。图片、特征和报告不会被自动上传。
+
+1:1 验证
+
+启用 Auto Split:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+每个子文件夹代表一个身份。文件名中包含 "gallery" 的图片会被选为该身份的 gallery image。如果不存在这样的文件，则选择排序后的第一张图片作为 gallery。其余图片作为 probe。评测会将每个 probe 与每个身份的 gallery 做比对，因此每个 trial 都是 probe vs gallery。
+
+关闭 Auto Split:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+每张图片都会被视为 probe。评测会在所有身份文件夹中的图片之间执行完整的 probe-vs-probe 两两比对。
+
+1:N 识别
+
+启用 Auto Split:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+如果存在 identities/ 文件夹，会优先使用它。如果不存在，所选数据集根目录也可以直接包含身份文件夹。Gallery 选择规则与 1:1 的 Auto Split 相同。
+
+关闭 Auto Split:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ 和 probe/ 是必需的。unknown/ 是可选的。1:N 评测始终需要 gallery 图片。
+
+核验内容
+
+- 所选根目录存在，并且与当前评测模式匹配。
+- 必需的 gallery 和 probe 文件夹/图片存在。
+- Auto Split 能生成有效的 gallery/probe 集合。
+- 1:1 能生成正样本对和负样本对。
+- 1:N 的 probe 身份有对应的有效 gallery。
+- 每张需要核验的图片都可以读取。
+- 检测到的人脸数量符合所选多人脸处理策略。
+
+多人脸处理
+
+- Require exactly one face: 多人脸图片会在核验阶段失败。
+- Use largest face: 评测使用检测到的最大人脸。
+- Use largest centered face: 评测使用 area - center_distance^2 * 2.0 得分最高的人脸。
+- Mark as skip: 多人脸图片会被跳过；如果被跳过的是 gallery 图片，则可能导致核验失败。
+
+报告输出
+
+1:1 报告包含最佳 cosine 阈值下的 accuracy、FAR/FRR、TAR@FAR、阈值建议、延迟和原始结果样例。
+
+1:N 报告包含 Top1、TAR@FAR、阈值建议、延迟和原始结果样例。""",
+    "ja": """Enterprise Evaluation ページでは、ID フォルダからローカルの 1:1 検証と 1:N 識別評価を実行できます。画像、特徴量、レポートは自動アップロードされません。
+
+1:1 検証
+
+Auto Split 有効:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+各サブフォルダは 1 つの ID を表します。ファイル名に "gallery" を含む画像がその ID の gallery image として選択されます。該当ファイルがない場合は、ソート順で最初の画像が gallery になります。それ以外の画像は probe です。評価では各 probe をすべての ID の gallery と比較するため、各 trial は probe vs gallery です。
+
+Auto Split 無効:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+すべての画像が probe として扱われます。すべての ID フォルダ内の画像に対して、probe-vs-probe の総当たり比較を行います。
+
+1:N 識別
+
+Auto Split 有効:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+identities/ フォルダが存在する場合は優先されます。存在しない場合、選択したデータセットルートに ID フォルダを直接置くこともできます。Gallery の選択規則は 1:1 の Auto Split と同じです。
+
+Auto Split 無効:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ と probe/ は必須です。unknown/ は任意です。1:N 評価には常に gallery 画像が必要です。
+
+検証項目
+
+- 選択したルートが存在し、選択中の評価モードと一致していること。
+- 必須の gallery/probe フォルダまたは画像が存在すること。
+- Auto Split が有効な gallery/probe セットを作成できること。
+- 1:1 で正例ペアと負例ペアを生成できること。
+- 1:N の probe ID に有効な gallery があること。
+- 検証対象の画像を読み込めること。
+- 検出された顔数が選択した複数顔処理ポリシーに従うこと。
+
+複数顔の処理
+
+- Require exactly one face: 複数顔画像は検証で失敗します。
+- Use largest face: 検出された最大の顔を使用します。
+- Use largest centered face: area - center_distance^2 * 2.0 のスコアが最も高い顔を使用します。
+- Mark as skip: 複数顔画像をスキップします。gallery 画像がスキップされると検証失敗になる場合があります。
+
+レポート出力
+
+1:1 レポートには、最適な cosine しきい値での accuracy、FAR/FRR、TAR@FAR、しきい値推奨、レイテンシ、原始結果例が含まれます。
+
+1:N レポートには、Top1、TAR@FAR、しきい値推奨、レイテンシ、原始結果例が含まれます。""",
+    "ko": """Enterprise Evaluation 페이지는 ID 폴더에서 로컬 1:1 검증 및 1:N 식별 평가를 수행합니다. 이미지, 임베딩, 보고서는 자동 업로드되지 않습니다.
+
+1:1 검증
+
+Auto Split 켜짐:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+각 하위 폴더는 하나의 ID입니다. 파일명에 "gallery"가 포함된 이미지가 해당 ID의 gallery image로 선택됩니다. 그런 파일이 없으면 정렬된 첫 번째 이미지가 gallery가 됩니다. 나머지 이미지는 probe입니다. 평가는 각 probe를 모든 ID의 gallery와 비교하므로 각 trial은 probe vs gallery입니다.
+
+Auto Split 꺼짐:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+모든 이미지는 probe로 처리됩니다. 모든 ID 폴더의 이미지에 대해 전체 probe-vs-probe 쌍 비교를 수행합니다.
+
+1:N 식별
+
+Auto Split 켜짐:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+identities/ 폴더가 있으면 우선 사용됩니다. 없으면 선택한 데이터셋 루트에 ID 폴더를 직접 둘 수 있습니다. Gallery 선택 규칙은 1:1 Auto Split과 동일합니다.
+
+Auto Split 꺼짐:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/와 probe/는 필수입니다. unknown/은 선택 사항입니다. 1:N 평가는 항상 gallery 이미지가 필요합니다.
+
+검증 항목
+
+- 선택한 루트가 존재하고 선택한 평가 모드와 일치해야 합니다.
+- 필요한 gallery/probe 폴더 또는 이미지가 있어야 합니다.
+- Auto Split이 유효한 gallery/probe 세트를 만들 수 있어야 합니다.
+- 1:1은 양성 쌍과 음성 쌍을 모두 생성할 수 있어야 합니다.
+- 1:N probe ID에는 유효한 gallery가 있어야 합니다.
+- 검증 대상 이미지를 읽을 수 있어야 합니다.
+- 감지된 얼굴 수가 선택한 다중 얼굴 처리 정책을 따라야 합니다.
+
+다중 얼굴 처리
+
+- Require exactly one face: 다중 얼굴 이미지는 검증에서 실패합니다.
+- Use largest face: 감지된 가장 큰 얼굴을 사용합니다.
+- Use largest centered face: area - center_distance^2 * 2.0 점수가 가장 높은 얼굴을 사용합니다.
+- Mark as skip: 다중 얼굴 이미지는 건너뜁니다. gallery 이미지가 건너뛰어지면 검증이 실패할 수 있습니다.
+
+보고서 출력
+
+1:1 보고서에는 최적 cosine 임계값의 accuracy, FAR/FRR, TAR@FAR, 임계값 권장, 지연 시간, 원시 결과 예시가 포함됩니다.
+
+1:N 보고서에는 Top1, TAR@FAR, 임계값 권장, 지연 시간, 원시 결과 예시가 포함됩니다。""",
+    "es": """La página Enterprise Evaluation admite evaluación local 1:1 y 1:N desde carpetas de identidad. Las imágenes, embeddings e informes no se suben automáticamente.
+
+Verificación 1:1
+
+Auto Split activado:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Cada subcarpeta representa una identidad. Una imagen cuyo nombre contiene "gallery" se selecciona como gallery image de esa identidad. Si no existe, se usa la primera imagen ordenada como gallery. Las demás imágenes son probes. La evaluación compara cada probe contra la gallery de cada identidad, por lo que cada trial es probe vs gallery.
+
+Auto Split desactivado:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Cada imagen se trata como probe. La evaluación ejecuta comparaciones completas probe-vs-probe entre todas las carpetas de identidad.
+
+Identificación 1:N
+
+Auto Split activado:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+La carpeta identities/ se prefiere cuando existe. Si no existe, la raíz seleccionada puede contener carpetas de identidad directamente. La selección de gallery sigue la misma regla de Auto Split que 1:1.
+
+Auto Split desactivado:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ y probe/ son obligatorias. unknown/ es opcional. La evaluación 1:N siempre requiere imágenes gallery.
+
+Comprobaciones de validación
+
+- La raíz seleccionada existe y coincide con el modo elegido.
+- Las carpetas o imágenes gallery/probe requeridas existen.
+- Auto Split puede crear conjuntos gallery/probe válidos.
+- 1:1 puede generar pares positivos y negativos.
+- Las identidades probe de 1:N tienen cobertura gallery válida.
+- Cada imagen que debe validarse puede leerse.
+- El número de caras detectadas sigue la política de manejo de múltiples caras.
+
+Manejo de múltiples caras
+
+- Require exactly one face: las imágenes con varias caras fallan en validación.
+- Use largest face: se usa la cara detectada más grande.
+- Use largest centered face: se usa la cara con mayor puntuación area - center_distance^2 * 2.0.
+- Mark as skip: las imágenes con varias caras se omiten; omitir imágenes gallery puede hacer fallar la validación.
+
+Salidas del informe
+
+Los informes 1:1 incluyen accuracy con el mejor umbral cosine, FAR/FRR, TAR@FAR, recomendaciones de umbral, latencia y ejemplos de resultados sin procesar.
+
+Los informes 1:N incluyen Top1, TAR@FAR, recomendaciones de umbral, latencia y ejemplos de resultados sin procesar.""",
+    "fr": """La page Enterprise Evaluation prend en charge l’évaluation locale 1:1 et 1:N à partir de dossiers d’identité. Les images, embeddings et rapports ne sont pas téléversés automatiquement.
+
+Vérification 1:1
+
+Auto Split activé :
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Chaque sous-dossier représente une identité. Une image dont le nom contient "gallery" est sélectionnée comme gallery image de cette identité. Si aucun fichier ne correspond, la première image triée est utilisée comme gallery. Les autres images sont des probes. L’évaluation compare chaque probe à la gallery de chaque identité ; chaque trial est donc probe vs gallery.
+
+Auto Split désactivé :
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Chaque image est traitée comme probe. L’évaluation exécute des comparaisons probe-vs-probe complètes entre tous les dossiers d’identité.
+
+Identification 1:N
+
+Auto Split activé :
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+Le dossier identities/ est préféré lorsqu’il existe. Sinon, la racine sélectionnée peut contenir directement les dossiers d’identité. La sélection de gallery suit la même règle Auto Split que 1:1.
+
+Auto Split désactivé :
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ et probe/ sont obligatoires. unknown/ est optionnel. L’évaluation 1:N nécessite toujours des images gallery.
+
+Contrôles de validation
+
+- La racine sélectionnée existe et correspond au mode choisi.
+- Les dossiers ou images gallery/probe requis existent.
+- Auto Split peut créer des ensembles gallery/probe valides.
+- 1:1 peut générer des paires positives et négatives.
+- Les identités probe de 1:N disposent d’une gallery valide.
+- Chaque image à valider peut être lue.
+- Le nombre de visages détectés respecte la politique multi-visage sélectionnée.
+
+Gestion multi-visage
+
+- Require exactly one face: les images multi-visages échouent à la validation.
+- Use largest face: l’évaluation utilise le plus grand visage détecté.
+- Use largest centered face: l’évaluation utilise le visage au meilleur score area - center_distance^2 * 2.0.
+- Mark as skip: les images multi-visages sont ignorées ; ignorer une image gallery peut faire échouer la validation.
+
+Sorties de rapport
+
+Les rapports 1:1 incluent l’accuracy au meilleur seuil cosine, FAR/FRR, TAR@FAR, recommandations de seuil, latence et exemples de résultats bruts.
+
+Les rapports 1:N incluent Top1, TAR@FAR, recommandations de seuil, latence et exemples de résultats bruts.""",
+    "de": """Die Seite Enterprise Evaluation unterstützt lokale 1:1-Verifikation und 1:N-Identifikation aus Identitätsordnern. Bilder, Embeddings und Berichte werden nicht automatisch hochgeladen.
+
+1:1-Verifikation
+
+Auto Split aktiviert:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Jeder Unterordner steht für eine Identität. Eine Datei, deren Name "gallery" enthält, wird als gallery image dieser Identität gewählt. Wenn keine solche Datei vorhanden ist, wird das erste sortierte Bild als gallery verwendet. Alle anderen Bilder sind probes. Die Evaluierung vergleicht jede probe mit der gallery jeder Identität; jeder trial ist also probe vs gallery.
+
+Auto Split deaktiviert:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Jedes Bild wird als probe behandelt. Die Evaluierung führt vollständige probe-vs-probe-Paarvergleiche über alle Identitätsordner aus.
+
+1:N-Identifikation
+
+Auto Split aktiviert:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+Der Ordner identities/ wird bevorzugt, wenn er vorhanden ist. Falls nicht, kann der ausgewählte Dataset-Root Identitätsordner direkt enthalten. Die Gallery-Auswahl folgt derselben Auto-Split-Regel wie bei 1:1.
+
+Auto Split deaktiviert:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ und probe/ sind erforderlich. unknown/ ist optional. Eine 1:N-Evaluierung benötigt immer gallery-Bilder.
+
+Validierungsprüfungen
+
+- Der ausgewählte Root existiert und passt zum gewählten Modus.
+- Erforderliche gallery/probe-Ordner oder Bilder existieren.
+- Auto Split kann gültige gallery/probe-Sets erzeugen.
+- 1:1 kann positive und negative Paare erzeugen.
+- 1:N-probe-Identitäten haben gültige gallery-Abdeckung.
+- Jedes zu validierende Bild kann gelesen werden.
+- Die erkannte Gesichtszahl folgt der ausgewählten Mehrgesichter-Strategie.
+
+Mehrgesichter-Strategie
+
+- Require exactly one face: Bilder mit mehreren Gesichtern schlagen in der Validierung fehl.
+- Use largest face: Das größte erkannte Gesicht wird verwendet.
+- Use largest centered face: Das Gesicht mit dem besten Score area - center_distance^2 * 2.0 wird verwendet.
+- Mark as skip: Bilder mit mehreren Gesichtern werden übersprungen; übersprungene gallery-Bilder können die Validierung fehlschlagen lassen.
+
+Berichtsausgaben
+
+1:1-Berichte enthalten accuracy beim besten cosine-Schwellenwert, FAR/FRR, TAR@FAR, Schwellenwertempfehlungen, Latenz und Beispiele roher Ergebnisse.
+
+1:N-Berichte enthalten Top1, TAR@FAR, Schwellenwertempfehlungen, Latenz und Beispiele roher Ergebnisse.""",
+    "pt": """A página Enterprise Evaluation suporta avaliação local 1:1 e identificação 1:N a partir de pastas de identidade. Imagens, embeddings e relatórios não são carregados automaticamente.
+
+Verificação 1:1
+
+Auto Split ativado:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Cada subpasta representa uma identidade. Uma imagem cujo nome contém "gallery" é escolhida como gallery image dessa identidade. Se não existir, a primeira imagem ordenada é usada como gallery. As restantes imagens são probes. A avaliação compara cada probe com a gallery de cada identidade; cada trial é probe vs gallery.
+
+Auto Split desativado:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Cada imagem é tratada como probe. A avaliação executa comparações completas probe-vs-probe entre todas as pastas de identidade.
+
+Identificação 1:N
+
+Auto Split ativado:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+A pasta identities/ é preferida quando existe. Se não existir, a raiz selecionada pode conter diretamente pastas de identidade. A seleção de gallery segue a mesma regra de Auto Split do 1:1.
+
+Auto Split desativado:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ e probe/ são obrigatórios. unknown/ é opcional. A avaliação 1:N sempre requer imagens gallery.
+
+Verificações de validação
+
+- A raiz selecionada existe e corresponde ao modo selecionado.
+- As pastas ou imagens gallery/probe necessárias existem.
+- Auto Split consegue criar conjuntos gallery/probe válidos.
+- 1:1 consegue gerar pares positivos e negativos.
+- Identidades probe de 1:N têm cobertura gallery válida.
+- Cada imagem a validar pode ser lida.
+- A contagem de faces detectadas segue a política de múltiplas faces selecionada.
+
+Tratamento de múltiplas faces
+
+- Require exactly one face: imagens com múltiplas faces falham na validação.
+- Use largest face: a avaliação usa a maior face detectada.
+- Use largest centered face: a avaliação usa a face com maior pontuação area - center_distance^2 * 2.0.
+- Mark as skip: imagens com múltiplas faces são ignoradas; imagens gallery ignoradas podem fazer a validação falhar.
+
+Saídas do relatório
+
+Relatórios 1:1 incluem accuracy no melhor limiar cosine, FAR/FRR, TAR@FAR, recomendações de limiar, latência e exemplos de resultados brutos.
+
+Relatórios 1:N incluem Top1, TAR@FAR, recomendações de limiar, latência e exemplos de resultados brutos.""",
+    "ru": """Страница Enterprise Evaluation поддерживает локальную оценку 1:1 и 1:N из папок идентичностей. Изображения, эмбеддинги и отчеты не загружаются автоматически.
+
+Верификация 1:1
+
+Auto Split включен:
+
+dataset_1v1/
+  0001__Alice/
+    gallery.jpg
+    img002.jpg
+    img003.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Каждая подпапка представляет одну идентичность. Файл, имя которого содержит "gallery", выбирается как gallery image для этой идентичности. Если такого файла нет, первым gallery становится первое изображение в отсортированном списке. Все остальные изображения являются probes. Оценка сравнивает каждый probe с gallery каждой идентичности, поэтому каждый trial — это probe vs gallery.
+
+Auto Split отключен:
+
+dataset_1v1/
+  0001__Alice/
+    img001.jpg
+    img002.jpg
+  0002__Bob/
+    img001.jpg
+    img002.jpg
+
+Каждое изображение считается probe. Оценка выполняет полные попарные сравнения probe-vs-probe по всем папкам идентичностей.
+
+Идентификация 1:N
+
+Auto Split включен:
+
+dataset/
+  identities/
+    0001__Alice/
+      gallery.jpg
+      img002.jpg
+      img003.jpg
+    0002__Bob/
+      img001.jpg
+      img002.jpg
+
+Папка identities/ используется с приоритетом, если она существует. Если ее нет, выбранный корень датасета может напрямую содержать папки идентичностей. Выбор gallery следует тому же правилу Auto Split, что и в 1:1.
+
+Auto Split отключен:
+
+dataset_1n/
+  gallery/
+    0001__Alice/
+      enroll_001.jpg
+      enroll_002.jpg
+    0002__Bob/
+      enroll_001.jpg
+  probe/
+    0001__Alice/
+      test_001.jpg
+      test_002.jpg
+    0002__Bob/
+      test_001.jpg
+  unknown/
+    unknown_001.jpg
+    unknown_002.jpg
+
+gallery/ и probe/ обязательны. unknown/ является опциональной. Оценка 1:N всегда требует gallery-изображения.
+
+Проверки валидации
+
+- Выбранный корень существует и соответствует выбранному режиму.
+- Необходимые папки или изображения gallery/probe существуют.
+- Auto Split может создать валидные наборы gallery/probe.
+- 1:1 может создать положительные и отрицательные пары.
+- Идентичности probe в 1:N имеют валидное покрытие gallery.
+- Каждое проверяемое изображение можно прочитать.
+- Число обнаруженных лиц соответствует выбранной политике обработки нескольких лиц.
+
+Обработка нескольких лиц
+
+- Require exactly one face: изображения с несколькими лицами не проходят валидацию.
+- Use largest face: оценка использует самое крупное обнаруженное лицо.
+- Use largest centered face: оценка использует лицо с лучшим счетом area - center_distance^2 * 2.0.
+- Mark as skip: изображения с несколькими лицами пропускаются; пропуск gallery-изображения может привести к ошибке валидации.
+
+Вывод отчетов
+
+Отчеты 1:1 включают accuracy при лучшем cosine-пороге, FAR/FRR, TAR@FAR, рекомендации по порогу, задержку и примеры сырых результатов.
+
+Отчеты 1:N включают Top1, TAR@FAR, рекомендации по порогу, задержку и примеры сырых результатов.""",
+}
+
+
+def dataset_rules_text(language: str | None) -> str:
+    lang = effective_language(language)
+    text = DATASET_RULES_LOCAL_TEXT.get(lang)
+    if text:
+        return f"{tr('Evaluation Dataset Rules', lang)}\n\n{text}"
+    return DATASET_RULES_TEXT
+
+
+class DatasetRulesDialog(QDialog):
+    def __init__(self, language: str | None = None, parent=None):
+        if parent is None and isinstance(language, QWidget):
+            parent = language
+            language = getattr(getattr(parent, "context", None), "config", None)
+            language = getattr(language, "ui_language", None)
+        super().__init__(parent)
+        self.setWindowTitle("Evaluation Dataset Rules")
+        self.resize(820, 680)
+        layout = QVBoxLayout(self)
+        intro = QLabel("Prepare local identity folders using one of the supported layouts below.")
+        intro.setWordWrap(True)
+        intro.setProperty("role", "muted")
+        layout.addWidget(intro)
+        text = QTextEdit()
+        text.setReadOnly(True)
+        text.setPlainText(dataset_rules_text(language))
+        layout.addWidget(text, 1)
+        buttons = QDialogButtonBox(QDialogButtonBox.Close)
+        buttons.rejected.connect(self.close)
+        layout.addWidget(buttons)
+        apply_translations(self, language)
+
+
+class EnterpriseEvalPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Enterprise Evaluation",
+            "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.",
+            parent,
+        )
+        self.eval_mode = QComboBox()
+        self.eval_mode.setProperty("i18nItems", True)
+        self.eval_mode.addItem("1:1 Verification", "1:1 Verification")
+        self.eval_mode.addItem("1:N Identification", "1:N Identification")
+        self.eval_mode.currentIndexChanged.connect(self._update_instructions)
+        self.auto_split = QCheckBox("Auto Split")
+        self.auto_split.setChecked(True)
+        self.auto_split.setToolTip(
+            "Automatically select each identity's gallery image from a file containing 'gallery', or the first sorted image."
+        )
+        self.auto_split.stateChanged.connect(self._update_instructions)
+        self.multi_face_policy = QComboBox()
+        self.multi_face_policy.setProperty("i18nItems", True)
+        self.multi_face_policy.addItem("Require exactly one face", MULTI_FACE_REQUIRE_ONE)
+        self.multi_face_policy.addItem("Use largest face", MULTI_FACE_USE_LARGEST)
+        self.multi_face_policy.addItem("Use largest centered face", MULTI_FACE_USE_CENTERED_LARGEST)
+        self.multi_face_policy.addItem("Mark as skip", MULTI_FACE_SKIP)
+        self.multi_face_policy.setToolTip(
+            "Choose how evaluation handles images where more than one face is detected."
+        )
+        self.multi_face_policy.currentIndexChanged.connect(self._update_instructions)
+        self.dataset_root = DropInput("Identity Folders / Dataset Root", mode="folder")
+        self.dataset_root.pathsChanged.connect(self._on_dataset_paths_changed)
+        self.mode_summary = QLabel()
+        self.mode_summary.setWordWrap(True)
+        self.mode_summary.setProperty("role", "muted")
+        self.validation_status = QLabel("Validate the dataset before running an evaluation.")
+        self.validation_status.setWordWrap(True)
+        self.validation_status.setProperty("role", "muted")
+
+        workspace = QWidget()
+        workspace_layout = QGridLayout(workspace)
+        workspace_layout.setContentsMargins(0, 0, 0, 0)
+        workspace_layout.setHorizontalSpacing(14)
+        workspace_layout.setVerticalSpacing(12)
+
+        setup_card, setup_layout = self._card("Evaluation Setup")
+        setup_layout.addWidget(self._field_row("Evaluation mode", self.eval_mode))
+        setup_layout.addWidget(self._field_row("Data split", self.auto_split))
+        setup_layout.addWidget(self._field_row("Multi-face handling", self.multi_face_policy))
+        self.help_button = self.button("Dataset Rules", self.show_dataset_rules)
+        self.help_button.setToolTip("Open the detailed identity folder rules for 1:1, 1:N, and Auto Split.")
+        setup_layout.addWidget(self.help_button, alignment=Qt.AlignLeft)
+        setup_layout.addStretch(1)
+
+        data_card, data_layout = self._card("Dataset")
+        data_card.setAcceptDrops(True)
+        data_card.installEventFilter(self)
+        self.dataset_drop_card = data_card
+        data_layout.addWidget(self.dataset_root)
+        data_layout.addWidget(self.mode_summary)
+        data_layout.addStretch(1)
+
+        actions_card, actions_layout = self._card("Run")
+        actions_hint = QLabel("Validate the selected dataset and configuration before running evaluation.")
+        actions_hint.setWordWrap(True)
+        actions_hint.setProperty("role", "muted")
+        actions_layout.addWidget(actions_hint)
+        actions_layout.addWidget(self.validation_status)
+        button_row = QHBoxLayout()
+        self.validate_button = self.button("Validate Dataset", self.validate_dataset)
+        self.run_button = self.button("Run Evaluation", self.run, enabled=False)
+        self.open_reports_button = self.button("Open Report Folder", self.open_report_folder)
+        button_row.addWidget(self.validate_button)
+        button_row.addWidget(self.run_button)
+        button_row.addWidget(self.open_reports_button)
+        button_row.addStretch(1)
+        actions_layout.addLayout(button_row)
+        actions_layout.addStretch(1)
+
+        workspace_layout.addWidget(setup_card, 0, 0)
+        workspace_layout.addWidget(data_card, 0, 1)
+        workspace_layout.addWidget(actions_card, 1, 0, 1, 2)
+        workspace_layout.setColumnStretch(0, 0)
+        workspace_layout.setColumnStretch(1, 1)
+        self.content.addWidget(workspace)
+
+        self.output = QTextEdit()
+        self.output.setReadOnly(True)
+        self.output.setMinimumHeight(260)
+        self.content.addWidget(self.output, 1)
+        self.last_result = None
+        self.last_report_paths: dict[str, str] = {}
+        self.validation_result = None
+        self._validated_signature: tuple[str, bool, str, str] | None = None
+        self._update_instructions()
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        dataset_targets = {
+            getattr(self, "dataset_drop_card", None),
+        }
+        if watched in dataset_targets:
+            if event.type() in (QEvent.DragEnter, QEvent.DragMove):
+                if self._dataset_drop_path(event):
+                    self._set_dataset_drag_active(True)
+                    event.acceptProposedAction()
+                    return True
+                event.ignore()
+                return True
+            if event.type() == QEvent.DragLeave:
+                self._set_dataset_drag_active(False)
+                return False
+            if event.type() == QEvent.Drop:
+                path = self._dataset_drop_path(event)
+                self._set_dataset_drag_active(False)
+                if path:
+                    self.dataset_root.set_path(path)
+                    self.set_status(f"Dataset root selected: {path}")
+                    event.acceptProposedAction()
+                    return True
+                event.ignore()
+                return True
+        return super().eventFilter(watched, event)
+
+    def _dataset_drop_path(self, event) -> str:
+        mime = event.mimeData()
+        if not mime.hasUrls():
+            return ""
+        for url in mime.urls():
+            if not url.isLocalFile():
+                continue
+            path = Path(url.toLocalFile()).expanduser()
+            if path.is_dir():
+                return str(path)
+        return ""
+
+    def _set_dataset_drag_active(self, active: bool) -> None:
+        if hasattr(self.dataset_root, "_set_property"):
+            self.dataset_root._set_property("dragActive", bool(active))
+
+    def _on_dataset_paths_changed(self, paths=None) -> None:
+        del paths
+        self._invalidate_validation()
+
+    def _update_instructions(self) -> None:
+        language = self.context.config.ui_language
+        mode = self._eval_mode()
+        auto_split = self.auto_split.isChecked()
+        policy = tr(self.multi_face_policy.currentText(), language)
+        if mode.startswith("1:1"):
+            split_text = (
+                "Auto Split selects one gallery image per identity and compares every probe against every identity gallery."
+                if auto_split
+                else "Auto Split is off, so every image is treated as a probe and full pairwise probe-vs-probe comparisons are generated."
+            )
+            text = f"1:1 verification expects one identity per subfolder. {split_text}"
+        elif auto_split:
+            text = (
+                "1:N Auto Split can use identities/<identity folders> or identity folders directly. "
+                "Each identity must provide a gallery image and at least one probe."
+            )
+        else:
+            text = (
+                "1:N structured evaluation requires gallery/<identity> and probe/<identity>. "
+                "unknown/ is optional and contributes impostor scores for FAR thresholds."
+            )
+        self.mode_summary.setText(
+            f"{tr(text, language)} {tr('Multi-face handling', language)}: {policy}. "
+            f"{tr('Open Dataset Rules for detailed folder examples.', language)}"
+        )
+        self._invalidate_validation()
+
+    def _card(self, title: str) -> tuple[QFrame, QVBoxLayout]:
+        frame = QFrame()
+        frame.setObjectName("enterpriseCard")
+        frame.setFrameShape(QFrame.StyledPanel)
+        layout = QVBoxLayout(frame)
+        layout.setSpacing(10)
+        heading = QLabel(title)
+        heading.setObjectName("enterpriseCardTitle")
+        heading.setStyleSheet("font-size: 15px; font-weight: 700;")
+        layout.addWidget(heading)
+        return frame, layout
+
+    def _field_row(self, label: str, widget) -> QWidget:
+        row = QWidget()
+        layout = QGridLayout(row)
+        layout.setContentsMargins(0, 0, 0, 0)
+        layout.setHorizontalSpacing(10)
+        caption = QLabel(label)
+        caption.setMinimumWidth(150)
+        caption.setProperty("role", "secondary")
+        layout.addWidget(caption, 0, 0, Qt.AlignTop)
+        layout.addWidget(widget, 0, 1)
+        layout.setColumnStretch(1, 1)
+        return row
+
+    def show_dataset_rules(self) -> None:
+        dialog = DatasetRulesDialog(self.context.config.ui_language, self)
+        dialog.exec()
+
+    def _current_signature(self) -> tuple[str, bool, str, str]:
+        return (
+            self._eval_mode(),
+            self.auto_split.isChecked(),
+            self._multi_face_policy(),
+            self.dataset_root.path().strip(),
+        )
+
+    def _eval_mode(self) -> str:
+        return str(self.eval_mode.currentData() or "1:1 Verification")
+
+    def _multi_face_policy(self) -> str:
+        return str(self.multi_face_policy.currentData() or MULTI_FACE_REQUIRE_ONE)
+
+    def _invalidate_validation(self, *args) -> None:
+        del args
+        self._validated_signature = None
+        if hasattr(self, "run_button"):
+            self.run_button.setEnabled(False)
+        if hasattr(self, "validation_status"):
+            self.validation_status.setText(
+                tr("Dataset validation is required before running an evaluation.", self.context.config.ui_language)
+            )
+
+    def validate_dataset(self) -> None:
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        root = self.dataset_root.path().strip()
+        if not root:
+            self.show_error("Please select identity folders or a dataset root.")
+            return
+        signature = self._current_signature()
+        mode = self._eval_mode()
+        auto_split = self.auto_split.isChecked()
+        multi_face_policy = self._multi_face_policy()
+
+        def task(progress=None, is_cancelled=None):
+            return validate_enterprise_dataset(
+                root,
+                mode,
+                auto_split,
+                self.context.engine,
+                multi_face_policy=multi_face_policy,
+                progress_callback=progress,
+                cancel_callback=is_cancelled,
+            )
+
+        def done(result):
+            self.validation_result = result
+            self.output.setPlainText(self._validation_text(result))
+            if result.ok:
+                self._validated_signature = signature
+                self.run_button.setEnabled(True)
+                self.validation_status.setText(
+                    tr("Dataset validation passed. Evaluation is ready to run.", self.context.config.ui_language)
+                )
+                self.set_status("Dataset validation passed.")
+            else:
+                self._validated_signature = None
+                self.run_button.setEnabled(False)
+                self.validation_status.setText(
+                    tr(
+                        "Dataset validation failed. Fix the listed issues before running evaluation.",
+                        self.context.config.ui_language,
+                    )
+                )
+                self.set_status("Dataset validation failed.")
+
+        self.run_task("Dataset validation", task, done)
+
+    def run(self) -> None:
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        root = self.dataset_root.path().strip()
+        if not root:
+            self.show_error("Please select identity folders or a dataset root.")
+            return
+        if self._validated_signature != self._current_signature():
+            self.show_error("Please validate this dataset and configuration before running evaluation.")
+            return
+        mode = self._eval_mode()
+        auto_split = self.auto_split.isChecked()
+        multi_face_policy = self._multi_face_policy()
+
+        def task(progress=None, is_cancelled=None):
+            if mode.startswith("1:1"):
+                return run_identity_verification_evaluation(
+                    root,
+                    self.context.engine,
+                    auto_split=auto_split,
+                    multi_face_policy=multi_face_policy,
+                    license_status=self.context.config.license_status,
+                    progress_callback=progress,
+                    cancel_callback=is_cancelled,
+                )
+            return run_identity_identification_evaluation(
+                root,
+                self.context.engine,
+                auto_split=auto_split,
+                multi_face_policy=multi_face_policy,
+                license_status=self.context.config.license_status,
+                progress_callback=progress,
+                cancel_callback=is_cancelled,
+            )
+
+        def done(result):
+            language = self.context.config.ui_language
+            self.last_result = result
+            self.last_report_paths = write_reports(result, self.context.config.report_dir, language=language)
+            report_path = self.last_report_paths.get("pdf") or self.last_report_paths.get("markdown", "")
+            self.context.storage.save_evaluation_run(
+                result.scenario,
+                result.model_name,
+                result.provider,
+                result.threshold,
+                result.dataset_summary,
+                result.metrics,
+                result.latency,
+                report_path,
+                created_at=result.created_at,
+            )
+            self.output.setPlainText(self._summary_text(result, self.last_report_paths))
+            self.set_status(f"{tr('Evaluation complete. Report saved to:', language)} {report_path}")
+            QMessageBox.information(
+                self,
+                tr("Evaluation report saved", language),
+                f"{tr('Evaluation report saved to:', language)}\n{report_path}",
+            )
+
+        self.run_task("Enterprise evaluation", task, done)
+
+    def _validation_text(self, result) -> str:
+        lines = [
+            "Dataset Validation",
+            f"Status: {'PASSED' if result.ok else 'FAILED'}",
+            f"Mode: {result.mode}",
+            f"Auto Split: {result.auto_split}",
+            f"Multi-face policy: {result.multi_face_policy}",
+            f"Root: {result.root}",
+            "",
+            "Summary:",
+        ]
+        for key, value in result.summary.items():
+            lines.append(f"{key}: {value}")
+        if result.errors:
+            lines.extend(["", f"Errors: {len(result.errors)}"])
+            for row in result.errors[:50]:
+                lines.append(str(row))
+            if len(result.errors) > 50:
+                lines.append(f"... {len(result.errors) - 50} more errors")
+        if result.warnings:
+            lines.extend(["", f"Warnings: {len(result.warnings)}"])
+            for row in result.warnings[:50]:
+                lines.append(str(row))
+            if len(result.warnings) > 50:
+                lines.append(f"... {len(result.warnings) - 50} more warnings")
+        return "\n".join(lines)
+
+    def _summary_text(self, result, paths: dict[str, str]) -> str:
+        language = self.context.config.ui_language
+        lines = [
+            f"{tr('Scenario', language)}: {result.scenario}",
+            f"{tr('Dataset', language)}: {result.dataset_summary}",
+            f"{tr('Report PDF', language)}: {paths.get('pdf', tr('not exported', language))}",
+            f"{tr('Report Markdown', language)}: {paths.get('markdown', tr('not exported', language))}",
+            f"{tr('Report HTML', language)}: {paths.get('html', tr('not exported', language))}",
+            "",
+            f"{tr('Metrics', language)}:",
+        ]
+        for key, value in result.metrics.items():
+            if isinstance(value, float):
+                lines.append(f"{key}: {value:.6f}")
+            else:
+                lines.append(f"{key}: {value}")
+        if result.errors:
+            lines.extend(["", f"{tr('Errors', language)}: {len(result.errors)}", f"{tr('First errors', language)}:"])
+            for row in result.errors[:10]:
+                lines.append(str(row))
+        return "\n".join(lines)
+
+    def open_report_folder(self) -> None:
+        from PySide6.QtCore import QUrl
+        from PySide6.QtGui import QDesktopServices
+
+        QDesktopServices.openUrl(QUrl.fromLocalFile(str(Path(self.context.config.report_dir))))
diff --git a/python-package/insightface/gui/pages/face_search_page.py b/python-package/insightface/gui/pages/face_search_page.py
new file mode 100644
index 000000000..ad5d62944
--- /dev/null
+++ b/python-package/insightface/gui/pages/face_search_page.py
@@ -0,0 +1,179 @@
+"""1:N face search page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import QSize, Qt
+from PySide6.QtGui import QIcon, QPixmap
+from PySide6.QtWidgets import QInputDialog, QSplitter, QTableWidget, QTableWidgetItem, QVBoxLayout, QWidget
+
+from ..core.exporters import export_csv, export_json
+from ..core.utils import read_image, save_image, timestamp_for_filename
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from ..widgets.threshold_slider import ThresholdSlider
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class FaceSearchPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "1:N Face Search", "Search a query face against the local People Library.", parent)
+        self.query_path = ""
+        self.query_image = None
+        self.query_face = None
+        self.results = []
+        self.query_input = UploadPreview(
+            "Query Image",
+            extensions=[".jpg", ".jpeg", ".png", ".bmp", ".webp"],
+            dialog_filter="Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)",
+        )
+        self.query_input.pathChanged.connect(lambda path: self.load_query(path) if path else self.clear_query())
+        self.content.addWidget(
+            self.row(
+                self.button("Search", self.search),
+                self.button("Add as New Person", self.add_as_new_person),
+                self.button("Assign to Existing Person", self.assign_existing),
+                self.button("Export Results", self.export_results),
+            )
+        )
+        self.threshold = ThresholdSlider(context.config.recognition_threshold)
+        self.content.addWidget(self.threshold)
+        splitter = QSplitter(Qt.Horizontal)
+        left = QWidget()
+        left_layout = QVBoxLayout(left)
+        self.viewer = self.query_input
+        left_layout.addWidget(self.viewer)
+        right = QWidget()
+        right_layout = QVBoxLayout(right)
+        self.table = QTableWidget(0, 8)
+        self.table.setIconSize(QSize(48, 48))
+        self.table.setHorizontalHeaderLabels(["rank", "person_id", "thumbnail", "person_name", "similarity", "status", "sample_id", "crop_path"])
+        configure_table_columns(self.table, [60, 90, 90, 190, 90, 110, 90, 280])
+        right_layout.addWidget(self.table)
+        splitter.addWidget(left)
+        splitter.addWidget(right)
+        self.content.addWidget(splitter, 1)
+
+    def load_query(self, path: str) -> None:
+        self.query_path = path
+        self.query_image = read_image(path)
+        if self.query_image is None:
+            self.viewer.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self.viewer.set_image(self.query_image, path)
+
+    def clear_query(self) -> None:
+        self.query_path = ""
+        self.query_image = None
+        self.query_face = None
+        self.results = []
+        self.viewer.clear(emit=False)
+        self.table.setRowCount(0)
+
+    def search(self) -> None:
+        if self.query_image is None:
+            self.show_error("Please select a query image.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        if not self.context.storage.load_all_gallery_embeddings():
+            self.show_error("No registered people. Please add people first.")
+            return
+
+        def task():
+            face = self.context.engine.detect_best_face(self.query_image, source_path=self.query_path)
+            if face is None or face.normed_embedding is None:
+                raise ValueError("No face detected or embedding unavailable.")
+            results = self.context.storage.search_embeddings(face.normed_embedding, top_k=self.context.config.default_top_k, threshold=self.threshold.value())
+            return face, results
+
+        def done(payload):
+            self.query_face, self.results = payload
+            self.viewer.set_faces([{"bbox": self.query_face.bbox, "label": "Query"}])
+            self.table.setRowCount(len(self.results))
+            for row, result in enumerate(self.results):
+                data = [
+                    row + 1,
+                    result.person_id if result.person_id is not None else "",
+                    "",
+                    result.person_name,
+                    f"{result.similarity:.4f}",
+                    result.status,
+                    result.sample_id,
+                    result.crop_path or "",
+                ]
+                for col, value in enumerate(data):
+                    item = QTableWidgetItem(str(value))
+                    if col == 2:
+                        icon = self._thumbnail_icon(result.crop_path)
+                        if icon is not None:
+                            item.setIcon(icon)
+                    self.table.setItem(row, col, item)
+                self.table.setRowHeight(row, 54)
+            refresh_table_columns(self.table)
+            self.set_status(f"Search complete. {len(self.results)} result(s).")
+
+        self.run_task("Searching people library", task, done)
+
+    def add_as_new_person(self) -> None:
+        if self.query_face is None or self.query_face.normed_embedding is None:
+            self.show_error("Run a search first.")
+            return
+        name, ok = QInputDialog.getText(self, "Add as New Person", "Person name")
+        if not ok or not name.strip():
+            return
+        person_id = self.context.storage.add_person(name.strip())
+        crop_path = Path(self.context.config.crop_dir) / f"person_{person_id}_{timestamp_for_filename()}.png"
+        if self.query_face.crop is not None:
+            save_image(crop_path, self.query_face.crop)
+        self.context.storage.add_face_sample(
+            person_id,
+            self.query_face.normed_embedding,
+            source_image_path=self.query_path,
+            crop_path=str(crop_path),
+            bbox=self.query_face.bbox,
+            kps=self.query_face.kps,
+            det_score=self.query_face.det_score,
+            quality_score=self.query_face.quality_score,
+            model_name=self.context.config.model_name,
+            provider=self.context.config.provider,
+        )
+        self.set_status(f"Added {name.strip()} to People Library.")
+
+    def assign_existing(self) -> None:
+        people = self.context.storage.list_people()
+        if self.query_face is None or self.query_face.normed_embedding is None:
+            self.show_error("Run a search first.")
+            return
+        if not people:
+            self.show_error("No registered people. Please add people first.")
+            return
+        names = [person["display_name"] or person["name"] for person in people]
+        name, ok = QInputDialog.getItem(self, "Assign to Existing Person", "Person", names, 0, False)
+        if not ok:
+            return
+        person = people[names.index(name)]
+        self.context.storage.add_face_sample(int(person["id"]), self.query_face.normed_embedding, source_image_path=self.query_path, bbox=self.query_face.bbox, kps=self.query_face.kps, det_score=self.query_face.det_score, quality_score=self.query_face.quality_score, model_name=self.context.config.model_name, provider=self.context.config.provider)
+        self.set_status(f"Assigned query face to {name}.")
+
+    def export_results(self) -> None:
+        if not self.results:
+            self.show_error("No search results to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"face_search_{timestamp_for_filename()}.csv"
+        rows = [result.to_json_dict() for result in self.results]
+        export_csv(path, rows)
+        export_json(path.with_suffix(".json"), rows)
+        self.set_status(f"Search results exported to {path}")
+
+    @staticmethod
+    def _thumbnail_icon(path: str | None) -> QIcon | None:
+        if not path:
+            return None
+        pixmap = QPixmap(str(Path(path)))
+        if pixmap.isNull():
+            return None
+        return QIcon(pixmap.scaled(48, 48, Qt.KeepAspectRatio, Qt.SmoothTransformation))
diff --git a/python-package/insightface/gui/pages/face_swap_page.py b/python-package/insightface/gui/pages/face_swap_page.py
new file mode 100644
index 000000000..f194147fd
--- /dev/null
+++ b/python-package/insightface/gui/pages/face_swap_page.py
@@ -0,0 +1,318 @@
+"""Source + Target face swap page for images and videos."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+from PySide6.QtCore import QEvent, Qt, QUrl
+from PySide6.QtGui import QDesktopServices
+from PySide6.QtWidgets import QLabel, QSplitter, QVBoxLayout, QWidget
+
+from ..core.constants import IMAGE_EXTENSIONS, VIDEO_EXTENSIONS
+from ..core.face_engine import providers_from_choice
+from ..core.i18n import tr
+from ..core.model_downloads import list_installed_swap_models
+from ..core.swap import FaceSwapEngine
+from ..core.utils import read_image, save_image, timestamp_for_filename
+from ..widgets.image_viewer import ImageViewer
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class FaceSwapPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Face Swap",
+            "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.",
+            parent,
+        )
+        self.source_path = ""
+        self.target_path = ""
+        self.target_kind = ""
+        self.output_image = None
+        self.output_video_path = ""
+        self.output_path = ""
+        self.source_image = None
+        self.target_image = None
+        self.content.addWidget(
+            self.notice(
+                "Face swap may require separate commercial authorization depending on usage and model license. "
+                "Use only with appropriate rights and consent."
+            )
+        )
+
+        image_filter = "Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)"
+        target_filter = "Images and videos (*.jpg *.jpeg *.png *.bmp *.webp *.mp4 *.mov *.avi *.mkv *.webm *.m4v);;All Files (*)"
+        self.content.addWidget(
+            self.row(
+                self.button("Run Face Swap", self.run_swap),
+                self.button("Open Result Directory", self.open_result_directory),
+            )
+        )
+        splitter = QSplitter(Qt.Horizontal)
+        self.source_view = UploadPreview("Source", extensions=IMAGE_EXTENSIONS, dialog_filter=image_filter)
+        self.target_view = UploadPreview("Target", extensions=IMAGE_EXTENSIONS | VIDEO_EXTENSIONS, dialog_filter=target_filter)
+        self.output_view = ImageViewer()
+        self.output_view.setToolTip("Click to open the saved result file.")
+        self.output_view.viewport().installEventFilter(self)
+        self.output_view.installEventFilter(self)
+        self.result_label = QLabel("Result preview appears here. Video results are saved to the exports folder.")
+        self.result_label.setWordWrap(True)
+        self.result_label.setProperty("role", "muted")
+        self.source_view.pathChanged.connect(lambda path: self.load_source(path) if path else self.clear_source())
+        self.target_view.pathChanged.connect(lambda path: self.load_target(path) if path else self.clear_target())
+
+        for title, viewer in [("Source", self.source_view), ("Target", self.target_view)]:
+            panel = QWidget()
+            layout = QVBoxLayout(panel)
+            label = QLabel(title)
+            label.setStyleSheet("font-weight:700;")
+            layout.addWidget(label)
+            layout.addWidget(viewer)
+            splitter.addWidget(panel)
+
+        result_panel = QWidget()
+        result_layout = QVBoxLayout(result_panel)
+        result_title = QLabel("Result")
+        result_title.setStyleSheet("font-weight:700;")
+        result_layout.addWidget(result_title)
+        result_layout.addWidget(self.output_view)
+        result_layout.addWidget(self.result_label)
+        splitter.addWidget(result_panel)
+        splitter.setStretchFactor(0, 1)
+        splitter.setStretchFactor(1, 1)
+        splitter.setStretchFactor(2, 1)
+        self.content.addWidget(splitter, 1)
+
+    def load_source(self, path: str) -> None:
+        self.source_path = path
+        self.source_image = read_image(path)
+        if self.source_image is None:
+            self.source_view.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self._clear_output()
+        self.source_view.set_image(self.source_image, path)
+
+    def clear_source(self) -> None:
+        self.source_path = ""
+        self.source_image = None
+        self.source_view.clear(emit=False)
+        self._clear_output()
+
+    def load_target(self, path: str) -> None:
+        self.target_path = path
+        suffix = Path(path).suffix.lower()
+        self._clear_output()
+        if suffix in IMAGE_EXTENSIONS:
+            self.target_kind = "image"
+            self.target_image = read_image(path)
+            if self.target_image is None:
+                self.target_view.clear(emit=False)
+                self.show_error("Image read failure.")
+                return
+            self.target_view.set_image(self.target_image, path)
+            self.set_status(f"Selected target image: {Path(path).name}")
+            return
+        if suffix in VIDEO_EXTENSIONS:
+            self.target_kind = "video"
+            self.target_image = self._read_video_preview(path)
+            if self.target_image is not None:
+                self.target_view.set_image(self.target_image, path)
+            else:
+                self.target_view.set_path(path, emit=False)
+            self.set_status(f"Selected target video: {Path(path).name}")
+            return
+        self.target_kind = ""
+        self.target_image = None
+        self.target_view.clear(emit=False)
+        self.show_error("Unsupported target file type.")
+
+    def clear_target(self) -> None:
+        self.target_path = ""
+        self.target_kind = ""
+        self.target_image = None
+        self.target_view.clear(emit=False)
+        self._clear_output()
+
+    def run_swap(self) -> None:
+        if self.source_image is None or not self.source_path:
+            self.show_error("Select a source image first.")
+            return
+        if not self.target_path or self.target_kind not in {"image", "video"}:
+            self.show_error("Select a target image or video first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        model_path = self._resolve_swap_model_path()
+        if not model_path:
+            self.show_error("Face swap model not found. Please download and choose a swap model in Models.")
+            return
+
+        source_image = self.source_image.copy()
+        target_path = self.target_path
+        target_kind = self.target_kind
+        target_image = self.target_image.copy() if self.target_image is not None and target_kind == "image" else None
+
+        def task(progress=None, is_cancelled=None):
+            swapper = FaceSwapEngine(
+                model_path,
+                providers_from_choice(self.context.config.provider),
+                gfpgan_model_path=getattr(self.context.config, "gfpgan_model_path", ""),
+                enable_gfpgan=bool(getattr(self.context.config, "enable_gfpgan", False)),
+            )
+            if not swapper.load():
+                raise ValueError(swapper.last_error)
+            source_face = self.context.engine.detect_best_face(source_image, source_path=self.source_path)
+            if source_face is None or source_face.normed_embedding is None:
+                raise ValueError("No usable face detected in source image.")
+            source_native = SimpleNamespace(normed_embedding=source_face.normed_embedding)
+            if target_kind == "image":
+                return self._swap_image(swapper, source_native, target_image, target_path)
+            return self._swap_video(swapper, source_native, target_path, progress, is_cancelled)
+
+        def done(result):
+            if result["kind"] == "image":
+                self.output_image = result["image"]
+                self.output_video_path = ""
+                self.output_path = result["path"]
+                self.output_view.set_image(self.output_image)
+                self.result_label.setText(
+                    tr("Image swap saved. Click Result to open it.", self.context.config.ui_language)
+                    + f"\n{self.output_path}"
+                )
+            else:
+                self.output_image = result.get("preview")
+                self.output_video_path = result["path"]
+                self.output_path = result["path"]
+                if self.output_image is not None:
+                    self.output_view.set_image(self.output_image)
+                self.result_label.setText(
+                    tr("Video swap saved. Click Result to open it.", self.context.config.ui_language)
+                    + f"\n{self.output_video_path}"
+                )
+            self.set_status(result["message"])
+
+        self.run_task("Running face swap", task, done)
+
+    def open_result(self) -> None:
+        if self.output_path and Path(self.output_path).exists():
+            QDesktopServices.openUrl(QUrl.fromLocalFile(self.output_path))
+            return
+        self.show_error("No saved result file to open.")
+
+    def open_result_directory(self) -> None:
+        folder = Path(self.output_path).parent if self.output_path else Path(self.context.config.export_dir)
+        folder.mkdir(parents=True, exist_ok=True)
+        QDesktopServices.openUrl(QUrl.fromLocalFile(str(folder)))
+
+    def _swap_image(self, swapper: FaceSwapEngine, source_native, target_image, target_path: str) -> dict:
+        if target_image is None:
+            raise ValueError("Target image could not be read.")
+        target_face = self.context.engine.detect_best_face(target_image, source_path=target_path)
+        if target_face is None or target_face.kps is None:
+            raise ValueError("No usable face detected in target image.")
+        target_native = SimpleNamespace(kps=np.asarray(target_face.kps, dtype=np.float32))
+        image = swapper.swap(target_image, target_native, source_native)
+        output_path = Path(self.context.config.export_dir) / f"face_swap_{timestamp_for_filename()}.png"
+        save_image(output_path, image)
+        return {
+            "kind": "image",
+            "image": image,
+            "path": str(output_path),
+            "message": f"Image face swap saved to {output_path}",
+        }
+
+    def _swap_video(self, swapper: FaceSwapEngine, source_native, target_path: str, progress=None, is_cancelled=None) -> dict:
+        try:
+            import cv2
+        except Exception as exc:
+            raise ValueError(f"OpenCV is required for video face swap: {exc}") from exc
+
+        cap = cv2.VideoCapture(target_path)
+        if not cap.isOpened():
+            raise ValueError("Video could not be opened.")
+        fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH) or 0)
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT) or 0)
+        frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
+        if width <= 0 or height <= 0:
+            cap.release()
+            raise ValueError("Video dimensions could not be read.")
+        output_path = Path(self.context.config.export_dir) / f"face_swap_video_{timestamp_for_filename()}.mp4"
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        writer = cv2.VideoWriter(str(output_path), cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))
+        if not writer.isOpened():
+            cap.release()
+            raise ValueError("Video writer could not be opened.")
+
+        preview = None
+        swapped = 0
+        processed = 0
+        try:
+            while True:
+                if is_cancelled and is_cancelled():
+                    break
+                ok, frame = cap.read()
+                if not ok:
+                    break
+                output_frame = frame
+                target_face = self.context.engine.detect_best_face(frame, source_path=target_path)
+                if target_face is not None and target_face.kps is not None:
+                    target_native = SimpleNamespace(kps=np.asarray(target_face.kps, dtype=np.float32))
+                    output_frame = swapper.swap(frame, target_native, source_native)
+                    swapped += 1
+                    if preview is None:
+                        preview = output_frame.copy()
+                writer.write(output_frame)
+                processed += 1
+                if progress:
+                    progress(processed, frame_count or processed, f"Processed {processed} frame(s), swapped {swapped}")
+        finally:
+            writer.release()
+            cap.release()
+        return {
+            "kind": "video",
+            "path": str(output_path),
+            "preview": preview,
+            "message": f"Video face swap complete. Processed {processed} frame(s), swapped {swapped} frame(s).",
+        }
+
+    def _resolve_swap_model_path(self) -> str:
+        configured = getattr(self.context.config, "swap_model_path", "")
+        if configured and Path(configured).expanduser().exists():
+            return str(Path(configured).expanduser())
+        installed = list_installed_swap_models(self.context.config.model_root)
+        return str(installed[0]) if installed else ""
+
+    def _clear_output(self) -> None:
+        self.output_image = None
+        self.output_video_path = ""
+        self.output_path = ""
+        self.output_view.set_image(None)
+        self.result_label.setText(
+            tr("Result preview appears here. Video results are saved to the exports folder.", self.context.config.ui_language)
+        )
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if watched in {self.output_view, self.output_view.viewport()} and event.type() == QEvent.MouseButtonRelease:
+            if event.button() == Qt.LeftButton:
+                self.open_result()
+                return True
+        return super().eventFilter(watched, event)
+
+    @staticmethod
+    def _read_video_preview(path: str) -> np.ndarray | None:
+        try:
+            import cv2
+
+            cap = cv2.VideoCapture(path)
+            ok, frame = cap.read()
+            cap.release()
+            return frame if ok else None
+        except Exception:
+            return None
diff --git a/python-package/insightface/gui/pages/license_center_page.py b/python-package/insightface/gui/pages/license_center_page.py
new file mode 100644
index 000000000..9fc443259
--- /dev/null
+++ b/python-package/insightface/gui/pages/license_center_page.py
@@ -0,0 +1,70 @@
+"""License Center page."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QLabel, QTableWidget, QTableWidgetItem
+
+from ..core.i18n import tr
+from ..core.links import open_insightface_url
+from ..core.licensing import allowed_usage_summary
+from ..core.constants import APP_VERSION
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+HOMEPAGE_URL = "https://www.insightface.ai"
+ENTERPRISE_HELP_URL = "https://www.insightface.ai/contact"
+
+
+class LicenseCenterPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "License Center", "A concise local-use and commercial-deployment summary.", parent)
+        self.summary = QLabel()
+        self.summary.setWordWrap(True)
+        self.content.addWidget(self.summary)
+        self.table = QTableWidget(0, 2)
+        self.table.setHorizontalHeaderLabels(["Usage", "Status"])
+        self.table.setMinimumHeight(260)
+        configure_table_columns(self.table, [260, 420])
+        self.content.addWidget(self.table)
+        self.content.addWidget(
+            self.row(
+                self.button("Visit Homepage", self.open_homepage),
+                self.button("Contact Enterprise Support", self.open_enterprise_help),
+            )
+        )
+        self.refresh()
+
+    def refresh(self) -> None:
+        cfg = self.context.config
+        self.summary.setText(
+            " ".join(
+                [
+                    tr(
+                        "InsightFace Evaluation Studio runs locally and does not upload images, embeddings, videos, or reports automatically.",
+                        cfg.ui_language,
+                    ),
+                    tr(
+                        "Code and model files may have different licenses, and commercial deployment requires appropriate model authorization.",
+                        cfg.ui_language,
+                    ),
+                    f"{tr('Current package', cfg.ui_language)}: insightface {APP_VERSION}, GUI {APP_VERSION}, "
+                    f"{tr('Model', cfg.ui_language).lower()} {cfg.model_name}, "
+                    f"{tr('Provider', cfg.ui_language).lower()} {cfg.provider}, "
+                    f"{tr('License', cfg.ui_language).lower()} {tr(cfg.license_status, cfg.ui_language)}.",
+                ]
+            )
+        )
+        summary = allowed_usage_summary()
+        self.table.setRowCount(len(summary))
+        for row, (usage, status) in enumerate(summary.items()):
+            self.table.setItem(row, 0, QTableWidgetItem(tr(usage, cfg.ui_language)))
+            self.table.setItem(row, 1, QTableWidgetItem(tr(status, cfg.ui_language)))
+        refresh_table_columns(self.table)
+
+    def open_homepage(self) -> None:
+        open_insightface_url(HOMEPAGE_URL, content="license_homepage")
+        self.set_status("Opened InsightFace homepage.")
+
+    def open_enterprise_help(self) -> None:
+        open_insightface_url(ENTERPRISE_HELP_URL, content="license_enterprise_support")
+        self.set_status("Opened InsightFace enterprise contact page.")
diff --git a/python-package/insightface/gui/pages/mode_dashboards.py b/python-package/insightface/gui/pages/mode_dashboards.py
new file mode 100644
index 000000000..5dd0f61bf
--- /dev/null
+++ b/python-package/insightface/gui/pages/mode_dashboards.py
@@ -0,0 +1,116 @@
+"""Mode-specific dashboard pages."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QGridLayout, QLabel, QPushButton, QWidget
+
+from ..core.constants import LOCAL_PROCESSING_NOTICE, RESPONSIBLE_USE_NOTICE
+from ..core.tooltips import set_button_tooltip
+from .base import BasePage
+
+
+class ModeDashboardPage(BasePage):
+    def __init__(
+        self,
+        context,
+        title: str,
+        description: str,
+        cards: list[tuple[str, str, str, str]],
+        notice: str = LOCAL_PROCESSING_NOTICE,
+        parent=None,
+    ):
+        super().__init__(context, title, description, parent)
+        self.content.addWidget(self.notice(notice))
+        grid_widget = QWidget()
+        grid = QGridLayout(grid_widget)
+        grid.setSpacing(12)
+        for index, (card_title, body, page_key, action) in enumerate(cards):
+            card = QWidget()
+            card.setObjectName("dashboardCard")
+            layout = QGridLayout(card)
+            title_label = QLabel(card_title)
+            title_label.setStyleSheet("font-size:16px; font-weight:700; border:0;")
+            body_label = QLabel(body)
+            body_label.setWordWrap(True)
+            body_label.setProperty("role", "muted")
+            button = QPushButton(action)
+            button.clicked.connect(lambda checked=False, target=page_key: self.window().open_page(target))
+            set_button_tooltip(button, body)
+            layout.addWidget(title_label, 0, 0)
+            layout.addWidget(body_label, 1, 0)
+            layout.addWidget(button, 2, 0)
+            grid.addWidget(card, index // 2, index % 2)
+        self.content.addWidget(grid_widget)
+        self.content.addStretch(1)
+
+
+class FaceDashboardPage(ModeDashboardPage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Verification Dashboard",
+            "Compare, identify, and recognize faces from images, folders, camera, and video.",
+            [
+                ("1:1 Compare", "Compare two local face images and export the decision.", "compare", "Open"),
+                ("1:N Search", "Search a query face against your local People Library.", "face_search", "Open"),
+                ("People Library", "Register people and manage local face samples.", "people_library", "Manage"),
+                ("Batch Processing", "Scan folders and identify known or unknown faces.", "batch_processing", "Scan"),
+                ("Camera Recognition", "Run live local recognition with consent.", "camera_recognition", "Open"),
+                ("Video Search", "Find a registered person in a local video.", "video_search", "Open"),
+            ],
+            parent=parent,
+        )
+
+
+class AlbumDashboardPage(ModeDashboardPage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Album Dashboard",
+            "Organize local photos by people, clusters, and searchable face groups.",
+            [
+                ("Scan Photo Library", "Index local image folders for face organization.", "album_import_scan", "Scan"),
+                ("People Clustering", "Group unknown faces into approximate person clusters.", "album_people_clustering", "Cluster"),
+                ("People Library", "Review known people and high-quality samples.", "people_library", "Manage"),
+                ("Search by Person", "Find likely person matches in your local library.", "album_person_search", "Search"),
+                ("Export Results", "Prepare local CSV and folder exports.", "album_export", "Review"),
+            ],
+            parent=parent,
+        )
+
+
+class SwapDashboardPage(ModeDashboardPage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Swap Dashboard",
+            "Swap faces in images and videos using configured InsightFace swap models.",
+            [
+                ("Image Face Swap", "Try a local image swap with a configured swap model.", "image_face_swap", "Open"),
+                ("Batch Image Swap", "Swap faces across a folder of images.", "batch_image_swap", "Preview"),
+                ("Video Face Swap", "Video swap workflows are planned for a future version.", "video_face_swap", "Preview"),
+                ("Source Face Library", "Prepare high-quality source faces for reuse.", "source_face_library", "Preview"),
+            ],
+            notice=(
+                RESPONSIBLE_USE_NOTICE
+                + " Face swap may require separate commercial authorization depending on usage and model license."
+            ),
+            parent=parent,
+        )
+
+
+class EnterpriseDashboardPage(ModeDashboardPage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Evaluation Dashboard",
+            "Evaluate InsightFace models on local business datasets and generate procurement-ready reports.",
+            [
+                ("Run Evaluation", "Run KYC, 1:N, media, video, or swap evaluation scenarios.", "enterprise_evaluation", "Run"),
+                ("Dataset Setup", "Prepare local datasets and CSV templates.", "enterprise_dataset_setup", "Review"),
+                ("Reports", "Open and export historical evaluation reports.", "reports", "Open"),
+                ("Threshold Calibration", "Review threshold operating points and recommendations.", "threshold_calibration", "Review"),
+                ("Commercial Next Steps", "Understand licensing, SDK, API, SLA, and custom training paths.", "commercial_next_steps", "Open"),
+            ],
+            parent=parent,
+        )
diff --git a/python-package/insightface/gui/pages/model_download_page.py b/python-package/insightface/gui/pages/model_download_page.py
new file mode 100644
index 000000000..f29cd8720
--- /dev/null
+++ b/python-package/insightface/gui/pages/model_download_page.py
@@ -0,0 +1,186 @@
+"""Manual GitHub release model downloads."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import QUrl
+from PySide6.QtGui import QDesktopServices
+from PySide6.QtWidgets import QAbstractItemView, QFrame, QLabel, QTableWidget, QTableWidgetItem, QVBoxLayout
+
+from ..core.config import save_config
+from ..core.constants import LICENSE_NOTICE
+from ..core.i18n import tr
+from ..core.model_downloads import (
+    GITHUB_RELEASES_URL,
+    ModelAsset,
+    download_model_asset,
+    load_cached_assets,
+    local_model_status,
+    refresh_model_assets,
+)
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+
+class ModelDownloadPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Model Downloads",
+            "Manually refresh GitHub release model URLs and download selected model packages locally.",
+            parent,
+        )
+        self.assets: list[ModelAsset] = []
+        self.content.addWidget(
+            self.notice(
+                "Downloads are manual only. The GUI does not auto-download models. "
+                "Model files may have different licenses from code; review usage before deployment."
+            )
+        )
+        self.content.addWidget(self.notice(LICENSE_NOTICE))
+        self.content.addWidget(
+            self.row(
+                self.button("Refresh Download URLs", self.refresh_urls),
+                self.button("Download Selected", self.download_selected),
+                self.button("Use Selected Model", self.use_selected_model),
+                self.button("Open Model Folder", self.open_model_folder),
+                self.button("Open GitHub Releases", self.open_releases),
+            )
+        )
+        self.table = QTableWidget(0, 8)
+        self.table.setSelectionBehavior(QAbstractItemView.SelectRows)
+        self.table.setSelectionMode(QAbstractItemView.SingleSelection)
+        self.table.setEditTriggers(QAbstractItemView.NoEditTriggers)
+        self.table.setHorizontalHeaderLabels(
+            ["asset", "source", "kind", "release", "size", "updated_at", "local status", "download url"]
+        )
+        configure_table_columns(self.table, [210, 100, 150, 130, 90, 170, 150, 360])
+        self.table.setMinimumHeight(400)
+        self.content.addWidget(self.table, 1)
+        self.content.addSpacing(8)
+        self.source_footer = QFrame()
+        self.source_footer.setObjectName("downloadSourceFooter")
+        footer_layout = QVBoxLayout(self.source_footer)
+        footer_layout.setContentsMargins(10, 8, 10, 8)
+        self.url_label = QLabel()
+        self.url_label.setWordWrap(True)
+        self.url_label.setProperty("role", "muted")
+        footer_layout.addWidget(self.url_label)
+        self.content.addWidget(self.source_footer)
+        self.refresh()
+
+    def refresh(self) -> None:
+        self.assets = load_cached_assets(self.context.config.cache_dir)
+        self.populate()
+
+    def populate(self) -> None:
+        self.table.setRowCount(len(self.assets))
+        for row, asset in enumerate(self.assets):
+            values = [
+                asset.name,
+                asset.source,
+                asset.kind,
+                asset.tag_name,
+                self._format_size(asset.size),
+                asset.updated_at,
+                local_model_status(asset, self.context.config.model_root),
+                asset.browser_download_url,
+            ]
+            for col, value in enumerate(values):
+                self.table.setItem(row, col, QTableWidgetItem(str(value)))
+        refresh_table_columns(self.table)
+        language = self.context.config.ui_language
+        self.url_label.setText(
+            f"{tr('Refresh source', language)}: {GITHUB_RELEASES_URL}\n"
+            f"{tr('Local model root', language)}: {Path(self.context.config.model_root).expanduser() / 'models'}"
+        )
+
+    def selected_asset(self) -> ModelAsset | None:
+        row = self.table.currentRow()
+        if row < 0 or row >= len(self.assets):
+            self.show_error("Select a model asset first.")
+            return None
+        return self.assets[row]
+
+    def refresh_urls(self) -> None:
+        def task():
+            return refresh_model_assets(self.context.config.cache_dir)
+
+        def done(payload):
+            self.assets, message = payload
+            self.populate()
+            self.set_status(message)
+
+        self.run_task("Refreshing model download URLs", task, done)
+
+    def download_selected(self) -> None:
+        asset = self.selected_asset()
+        if asset is None:
+            return
+
+        def task(progress=None, is_cancelled=None):
+            del is_cancelled
+            if progress:
+                progress(0, asset.size or 1, f"Connecting to download {asset.name}")
+            return download_model_asset(
+                asset,
+                model_root=self.context.config.model_root,
+                gui_cache_dir=self.context.config.cache_dir,
+                progress=progress,
+            )
+
+        def done(path):
+            path = Path(path)
+            lower_name = asset.name.lower()
+            if "gfpgan" in lower_name:
+                self.context.config.gfpgan_model_path = str(path)
+                save_config(self.context.config)
+            elif "swap" in lower_name or "inswapper" in lower_name:
+                self.context.config.swap_model_path = str(path)
+                save_config(self.context.config)
+            self.populate()
+            manager = self.window()
+            if hasattr(manager, "refresh_model_pages"):
+                manager.refresh_model_pages()
+            self.set_status(f"Downloaded {asset.name} to {path}")
+
+        self.run_task(f"Downloading {asset.name}", task, done)
+
+    def use_selected_model(self) -> None:
+        asset = self.selected_asset()
+        if asset is None:
+            return
+        if asset.name.endswith(".zip"):
+            self.context.config.model_name = asset.stem
+            self.context.config.custom_model_dir = ""
+            save_config(self.context.config)
+            self.set_status(f"Model set to {asset.stem}. Open Models and test model load.")
+        elif asset.name.endswith(".onnx"):
+            path = Path(self.context.config.model_root).expanduser() / "models" / asset.stem / asset.name
+            if "gfpgan" in asset.name.lower():
+                self.context.config.gfpgan_model_path = str(path)
+                save_config(self.context.config)
+                self.set_status(f"GFPGAN model set to {path}. Enable GFPGAN in Models > Runtime to use it after face swap.")
+            elif "swap" in asset.name.lower() or "inswapper" in asset.name.lower():
+                self.context.config.swap_model_path = str(path)
+                save_config(self.context.config)
+                self.set_status(f"Face swap model set to {path}.")
+            else:
+                self.set_status(f"ONNX model path: {path}")
+        else:
+            self.set_status("Selected asset cannot be used as a model package.")
+
+    def open_model_folder(self) -> None:
+        folder = Path(self.context.config.model_root).expanduser() / "models"
+        folder.mkdir(parents=True, exist_ok=True)
+        QDesktopServices.openUrl(QUrl.fromLocalFile(str(folder)))
+
+    def open_releases(self) -> None:
+        QDesktopServices.openUrl(QUrl(GITHUB_RELEASES_URL))
+
+    @staticmethod
+    def _format_size(size: int) -> str:
+        if not size:
+            return ""
+        return f"{size / (1024 * 1024):.1f} MB"
diff --git a/python-package/insightface/gui/pages/model_settings_page.py b/python-package/insightface/gui/pages/model_settings_page.py
new file mode 100644
index 000000000..6f398305d
--- /dev/null
+++ b/python-package/insightface/gui/pages/model_settings_page.py
@@ -0,0 +1,234 @@
+"""Model settings page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import Qt
+from PySide6.QtGui import QColor, QBrush
+from PySide6.QtWidgets import QCheckBox, QComboBox, QFormLayout, QLabel, QLineEdit, QTextEdit
+
+from ..core.config import save_config
+from ..core.face_engine import FaceEngine, is_cuda_provider_available, providers_from_choice
+from ..core.model_downloads import is_model_package_installed, list_installed_gfpgan_models, list_installed_swap_models
+from .base import BasePage
+
+
+class ModelSettingsPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Model Settings", "Configure model packs, execution provider, face swap models, and runtime checks.", parent)
+        form = QFormLayout()
+        self.model_combo = QComboBox()
+        self.model_packages = ["buffalo_l", "buffalo_m", "buffalo_s", "buffalo_sc", "antelopev2"]
+        self._rebuild_model_combo()
+        self.custom_dir = QLineEdit(context.config.custom_model_dir)
+        self.custom_dir_label = None
+        self._update_model_availability()
+        self.model_combo.currentIndexChanged.connect(self._update_custom_dir_visibility)
+        self.provider_combo = QComboBox()
+        self.provider_combo.addItems(["Auto", "CPU", "CUDA"])
+        self.provider_combo.setCurrentText(context.config.provider)
+        self._update_provider_availability()
+        self.det_combo = QComboBox()
+        self.det_combo.addItems(["Auto", "128x128", "320x320", "640x640", "1024x1024"])
+        self.det_combo.setCurrentText(context.config.det_size_label)
+        self.swap_model_combo = QComboBox()
+        self._update_swap_model_choices()
+        self.gfpgan_enabled = QCheckBox("Enable GFPGAN restore after face swap")
+        self.gfpgan_enabled.setChecked(bool(getattr(context.config, "enable_gfpgan", False)))
+        self.gfpgan_model_combo = QComboBox()
+        self._update_gfpgan_model_choices()
+        form.addRow("Model package", self.model_combo)
+        form.addRow("Custom model directory", self.custom_dir)
+        self.custom_dir_label = form.labelForField(self.custom_dir)
+        form.addRow("Provider", self.provider_combo)
+        form.addRow("Detection size", self.det_combo)
+        form.addRow("Face swap model", self.swap_model_combo)
+        form.addRow("GFPGAN post-processing", self.gfpgan_enabled)
+        form.addRow("GFPGAN model", self.gfpgan_model_combo)
+        self.content.addLayout(form)
+        self.runtime = QTextEdit()
+        self.runtime.setReadOnly(True)
+        self.content.addWidget(QLabel("Runtime information"))
+        self.content.addWidget(self.runtime)
+        self.content.addWidget(
+            self.row(
+                self.button("Save Settings", self.save),
+                self.button("Open Model Downloads", lambda: self.window().open_page("Model Downloads")),
+                self.button("Test Model Load", self.test_load),
+                self.button("Warmup", self.warmup),
+            )
+        )
+        self.refresh()
+        self._update_custom_dir_visibility()
+
+    def _apply_to_config(self) -> None:
+        cfg = self.context.config
+        chosen = self.model_combo.currentData()
+        cfg.model_name = chosen if chosen != "custom model directory" else self.custom_dir.text().strip()
+        cfg.custom_model_dir = self.custom_dir.text().strip()
+        provider = self.provider_combo.currentText()
+        cfg.provider = "Auto" if provider == "CUDA" and not is_cuda_provider_available() else provider
+        if self.det_combo.currentText() == "Auto":
+            cfg.det_size = [0, 0]
+        else:
+            size = self.det_combo.currentText().split("x")
+            cfg.det_size = [int(size[0]), int(size[1])]
+        cfg.swap_model_path = str(self.swap_model_combo.currentData() or "")
+        cfg.gfpgan_model_path = str(self.gfpgan_model_combo.currentData() or "")
+        cfg.enable_gfpgan = bool(self.gfpgan_enabled.isChecked() and cfg.gfpgan_model_path)
+
+    def save(self) -> None:
+        self._apply_to_config()
+        save_config(self.context.config)
+        self.set_status("Model settings saved.")
+        self.refresh()
+
+    def test_load(self) -> None:
+        self._apply_to_config()
+
+        def task():
+            engine = FaceEngine(
+                model_name=self.context.config.model_name,
+                providers=providers_from_choice(self.context.config.provider),
+                det_size=self.context.config.det_size_tuple,
+                root=self.context.config.model_root,
+                custom_model_dir=self.context.config.custom_model_dir,
+            )
+            engine.load()
+            return engine
+
+        def done(engine):
+            self.context.engine = engine
+            self.window().context.engine = engine
+            self.refresh()
+            if engine.is_loaded():
+                self.set_status("Model loaded successfully.")
+            else:
+                self.show_error(engine.last_error or "Model load failed.")
+
+        self.run_task("Loading model", task, done)
+
+    def warmup(self) -> None:
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        self.run_task("Model warmup", self.context.engine.warmup, lambda info: self.set_status(f"Warmup complete: {info['warmup_ms']:.1f} ms"))
+
+    def refresh(self) -> None:
+        self._update_model_availability()
+        self._update_provider_availability()
+        self._update_swap_model_choices()
+        self._update_gfpgan_model_choices()
+        info = self.context.engine.get_runtime_info()
+        self.runtime.setPlainText("\n".join(f"{key}: {value}" for key, value in info.items()))
+
+    def _update_provider_availability(self) -> None:
+        cuda_available = is_cuda_provider_available()
+        cuda_index = self.provider_combo.findText("CUDA")
+        if cuda_index >= 0:
+            item = self.provider_combo.model().item(cuda_index)
+            if item is not None:
+                item.setEnabled(cuda_available)
+                item.setToolTip(
+                    "CUDAExecutionProvider is available."
+                    if cuda_available
+                    else "CUDAExecutionProvider is not available. Install a matching onnxruntime-gpu, CUDA runtime, and GPU driver first."
+                )
+        if self.provider_combo.currentText() == "CUDA" and not cuda_available:
+            self.provider_combo.setCurrentText("Auto")
+            self.provider_combo.setToolTip("CUDA is unavailable on this machine, so Auto will use CPU.")
+        else:
+            self.provider_combo.setToolTip("Auto uses CUDA when CUDAExecutionProvider is available, otherwise CPU.")
+
+    def _update_model_availability(self) -> None:
+        model = self.model_combo.model()
+        for index, package in enumerate(self.model_packages):
+            item = model.item(index)
+            if item is None:
+                continue
+            installed = is_model_package_installed(package, self.context.config.model_root)
+            item.setEnabled(installed)
+            item.setText(package if installed else f"{package} (not downloaded)")
+            item.setData(package, Qt.UserRole)
+            if installed:
+                item.setData(None, Qt.ForegroundRole)
+            else:
+                item.setForeground(QBrush(QColor("#9ca3af")))
+            item.setToolTip(
+                f"{package} is installed under {self.context.config.model_root}/models."
+                if installed
+                else f"{package} is not downloaded. Open Models > Downloads to install it."
+            )
+        custom_item = model.item(len(self.model_packages))
+        if custom_item is not None:
+            custom_item.setEnabled(True)
+            custom_item.setText("custom model directory")
+            custom_item.setData("custom model directory", Qt.UserRole)
+        current = self.context.config.model_name if self.context.config.model_name in self.model_packages else "custom model directory"
+        index = self.model_combo.findData(current)
+        if index >= 0:
+            self.model_combo.setCurrentIndex(index)
+        self._update_custom_dir_visibility()
+
+    def _rebuild_model_combo(self) -> None:
+        self.model_combo.clear()
+        for package in self.model_packages:
+            self.model_combo.addItem(package, package)
+        self.model_combo.addItem("custom model directory", "custom model directory")
+
+    def _update_custom_dir_visibility(self) -> None:
+        is_custom = self.model_combo.currentData() == "custom model directory"
+        self.custom_dir.setVisible(is_custom)
+        if hasattr(self, "custom_dir_label") and self.custom_dir_label is not None:
+            self.custom_dir_label.setVisible(is_custom)
+
+    def _update_swap_model_choices(self) -> None:
+        current = getattr(self.context.config, "swap_model_path", "")
+        paths = list_installed_swap_models(self.context.config.model_root)
+        self.swap_model_combo.blockSignals(True)
+        self.swap_model_combo.clear()
+        if not paths:
+            self.swap_model_combo.addItem("No downloaded swap models", "")
+            self.swap_model_combo.setEnabled(False)
+            self.swap_model_combo.setToolTip("Download inswapper_128.onnx from Models > Downloads first.")
+        else:
+            self.swap_model_combo.setEnabled(True)
+            self.swap_model_combo.setToolTip("Only downloaded swap models are shown.")
+            for path in paths:
+                label = f"{Path(path).parent.name}/{Path(path).name}"
+                self.swap_model_combo.addItem(label, str(path))
+            index = self.swap_model_combo.findData(current)
+            if index >= 0:
+                self.swap_model_combo.setCurrentIndex(index)
+            elif current and Path(current).exists():
+                self.swap_model_combo.addItem(Path(current).name, current)
+                self.swap_model_combo.setCurrentIndex(self.swap_model_combo.count() - 1)
+        self.swap_model_combo.blockSignals(False)
+
+    def _update_gfpgan_model_choices(self) -> None:
+        current = getattr(self.context.config, "gfpgan_model_path", "")
+        paths = list_installed_gfpgan_models(self.context.config.model_root)
+        self.gfpgan_model_combo.blockSignals(True)
+        self.gfpgan_model_combo.clear()
+        if not paths:
+            self.gfpgan_model_combo.addItem("No downloaded GFPGAN models", "")
+            self.gfpgan_model_combo.setEnabled(False)
+            self.gfpgan_enabled.setEnabled(False)
+            self.gfpgan_model_combo.setToolTip("Download GFPGANv1.4.onnx from Models > Downloads first.")
+            self.gfpgan_enabled.setToolTip("GFPGAN is unavailable until a GFPGAN model is downloaded.")
+        else:
+            self.gfpgan_model_combo.setEnabled(True)
+            self.gfpgan_enabled.setEnabled(True)
+            self.gfpgan_model_combo.setToolTip("Only downloaded GFPGAN restore models are shown.")
+            self.gfpgan_enabled.setToolTip("Run GFPGAN face restoration after face swap.")
+            for path in paths:
+                label = f"{Path(path).parent.name}/{Path(path).name}"
+                self.gfpgan_model_combo.addItem(label, str(path))
+            index = self.gfpgan_model_combo.findData(current)
+            if index >= 0:
+                self.gfpgan_model_combo.setCurrentIndex(index)
+            elif current and Path(current).exists():
+                self.gfpgan_model_combo.addItem(Path(current).name, current)
+                self.gfpgan_model_combo.setCurrentIndex(self.gfpgan_model_combo.count() - 1)
+        self.gfpgan_model_combo.blockSignals(False)
diff --git a/python-package/insightface/gui/pages/multiface_photo_page.py b/python-package/insightface/gui/pages/multiface_photo_page.py
new file mode 100644
index 000000000..4d809e2af
--- /dev/null
+++ b/python-package/insightface/gui/pages/multiface_photo_page.py
@@ -0,0 +1,131 @@
+"""Multi-face photo recognition page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import Qt
+from PySide6.QtWidgets import QSplitter, QTableWidget, QTableWidgetItem, QVBoxLayout, QWidget
+
+from ..core.exporters import export_annotated_image, export_csv, export_json
+from ..core.utils import read_image, save_image, timestamp_for_filename
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class MultiFacePhotoPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Multi-face Photo Recognition", "Detect all faces in a group photo and match them against People Library.", parent)
+        self.image_path = ""
+        self.image = None
+        self.rows = []
+        self.photo_input = UploadPreview(
+            "Photo",
+            extensions=[".jpg", ".jpeg", ".png", ".bmp", ".webp"],
+            dialog_filter="Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)",
+        )
+        self.photo_input.pathChanged.connect(lambda path: self.load_photo(path) if path else self.clear_photo())
+        self.content.addWidget(
+            self.row(
+                self.button("Recognize Faces", self.recognize),
+                self.button("Export Annotated Image", self.export_image),
+                self.button("Export CSV / JSON", self.export_rows),
+            )
+        )
+        splitter = QSplitter(Qt.Horizontal)
+        left = QWidget()
+        left_layout = QVBoxLayout(left)
+        self.viewer = self.photo_input
+        left_layout.addWidget(self.viewer)
+        right = QWidget()
+        right_layout = QVBoxLayout(right)
+        self.table = QTableWidget(0, 6)
+        self.table.setHorizontalHeaderLabels(["face_index", "person_name", "similarity", "status", "bbox", "det_score"])
+        configure_table_columns(self.table, [90, 180, 90, 110, 230, 90])
+        right_layout.addWidget(self.table)
+        splitter.addWidget(left)
+        splitter.addWidget(right)
+        self.content.addWidget(splitter, 1)
+
+    def load_photo(self, path: str) -> None:
+        self.image_path = path
+        self.image = read_image(path)
+        if self.image is None:
+            self.viewer.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self.viewer.set_image(self.image, path)
+
+    def clear_photo(self) -> None:
+        self.image_path = ""
+        self.image = None
+        self.rows = []
+        self.viewer.clear(emit=False)
+        self.table.setRowCount(0)
+
+    def recognize(self) -> None:
+        if self.image is None:
+            self.show_error("Please select a photo.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+
+        def task():
+            faces = self.context.engine.detect_faces(self.image, source_path=self.image_path)
+            gallery = self.context.storage.load_all_gallery_embeddings()
+            rows = []
+            overlays = []
+            media_id = self.context.storage.add_media_item(self.image_path, "image", width=self.image.shape[1], height=self.image.shape[0], processed_at=timestamp_for_filename())
+            for index, face in enumerate(faces):
+                result = None
+                if face.normed_embedding is not None and gallery:
+                    result_list = self.context.storage.search_embeddings(face.normed_embedding, self.context.config.default_top_k, self.context.config.recognition_threshold)
+                    result = result_list[0] if result_list else None
+                name = result.person_name if result and result.similarity >= self.context.config.recognition_threshold else "Unknown"
+                similarity = result.similarity if result else 0.0
+                status = "matched" if result and result.similarity >= self.context.config.recognition_threshold else "unknown"
+                crop_path = ""
+                if face.crop is not None:
+                    crop_path = str(Path(self.context.config.crop_dir) / f"media_{media_id}_{index}_{timestamp_for_filename()}.png")
+                    save_image(crop_path, face.crop)
+                self.context.storage.add_media_face(media_id, face.normed_embedding, crop_path=crop_path, bbox=face.bbox, kps=face.kps, det_score=face.det_score, predicted_person_id=result.person_id if result else None, similarity=similarity, status=status)
+                row = {"face_index": index, "person_name": name, "similarity": similarity, "status": status, "bbox": face.bbox, "det_score": face.det_score, "crop_path": crop_path}
+                rows.append(row)
+                overlays.append({"bbox": face.bbox, "label": name, "similarity": similarity})
+            return rows, overlays
+
+        def done(payload):
+            self.rows, overlays = payload
+            self.viewer.set_faces(overlays)
+            self.table.setRowCount(len(self.rows))
+            columns = ["face_index", "person_name", "similarity", "status", "bbox", "det_score"]
+            for row_index, row in enumerate(self.rows):
+                for col, key in enumerate(columns):
+                    value = row[key]
+                    if isinstance(value, float):
+                        value = f"{value:.4f}"
+                    self.table.setItem(row_index, col, QTableWidgetItem(str(value)))
+            refresh_table_columns(self.table)
+            self.set_status(f"Recognized {len(self.rows)} face(s).")
+
+        self.run_task("Recognizing photo", task, done)
+
+    def export_image(self) -> None:
+        image = self.viewer.render_with_overlay()
+        if image is None:
+            self.show_error("No annotated image to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"annotated_{timestamp_for_filename()}.png"
+        export_annotated_image(path, image)
+        self.set_status(f"Annotated image exported to {path}")
+
+    def export_rows(self) -> None:
+        if not self.rows:
+            self.show_error("No recognition results to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"multiface_{timestamp_for_filename()}.csv"
+        export_csv(path, self.rows)
+        export_json(path.with_suffix(".json"), self.rows)
+        self.set_status(f"Results exported to {path}")
diff --git a/python-package/insightface/gui/pages/people_library_page.py b/python-package/insightface/gui/pages/people_library_page.py
new file mode 100644
index 000000000..624a24bfd
--- /dev/null
+++ b/python-package/insightface/gui/pages/people_library_page.py
@@ -0,0 +1,361 @@
+"""People library management page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import QEvent, QSize, Qt, Signal
+from PySide6.QtGui import QCursor, QIcon, QPixmap
+from PySide6.QtWidgets import (
+    QFileDialog,
+    QFrame,
+    QInputDialog,
+    QLabel,
+    QMenu,
+    QPushButton,
+    QSplitter,
+    QTableWidget,
+    QTableWidgetItem,
+    QTextEdit,
+    QVBoxLayout,
+    QWidget,
+)
+
+from ..core.exporters import export_csv
+from ..core.utils import list_images, read_image, save_image, timestamp_for_filename
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+
+IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".webp"}
+IMAGE_FILTER = "Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)"
+
+
+class ImageOrFolderImportBox(QFrame):
+    pathsSelected = Signal(list)
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self.setObjectName("imageOrFolderImport")
+        self.setAcceptDrops(True)
+        self.setMouseTracking(True)
+        self.setProperty("hoverActive", False)
+        self.setProperty("dragActive", False)
+        self.setProperty("hasFiles", False)
+        layout = QVBoxLayout(self)
+        title = QLabel("Import Image or Folder")
+        title.setAlignment(Qt.AlignCenter)
+        title.setStyleSheet("font-size:15px; font-weight:700;")
+        hint = QLabel("Click to import an image or folder, or drag one here. Folders are imported recursively.")
+        hint.setObjectName("dropPrompt")
+        hint.setAlignment(Qt.AlignCenter)
+        hint.setWordWrap(True)
+        self.button = QPushButton("Import Image or Folder")
+        self.button.clicked.connect(self._show_menu)
+        layout.addWidget(title)
+        layout.addWidget(hint)
+        layout.addWidget(self.button, 0, Qt.AlignHCenter)
+        for watched in (self, title, hint):
+            watched.installEventFilter(self)
+
+    def _show_menu(self) -> None:
+        menu = QMenu(self)
+        menu.addAction("Select Image", self._select_image)
+        menu.addAction("Select Folder", self._select_folder)
+        menu.exec(self.button.mapToGlobal(self.button.rect().bottomLeft()))
+
+    def _select_image(self) -> None:
+        path, _ = QFileDialog.getOpenFileName(self, "Select image", str(Path.home()), IMAGE_FILTER)
+        if path:
+            self.pathsSelected.emit([path])
+
+    def _select_folder(self) -> None:
+        folder = QFileDialog.getExistingDirectory(self, "Select image folder", str(Path.home()))
+        if folder:
+            self.pathsSelected.emit([str(path) for path in list_images(folder, recursive=True)])
+
+    def dragEnterEvent(self, event) -> None:  # noqa: N802
+        if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+            self._set_drag_active(True)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def dragLeaveEvent(self, event) -> None:  # noqa: N802
+        self._set_drag_active(False)
+        super().dragLeaveEvent(event)
+
+    def dropEvent(self, event) -> None:  # noqa: N802
+        self._set_drag_active(False)
+        paths: list[str] = []
+        for url in event.mimeData().urls():
+            if not url.isLocalFile():
+                continue
+            path = Path(url.toLocalFile())
+            if path.is_dir():
+                paths.extend(str(item) for item in list_images(path, recursive=True))
+            elif path.suffix.lower() in IMAGE_EXTENSIONS:
+                paths.append(str(path))
+        if paths:
+            self.pathsSelected.emit(paths)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def _accepted_urls(self, urls) -> bool:
+        for url in urls:
+            if not url.isLocalFile():
+                continue
+            path = Path(url.toLocalFile())
+            if path.is_dir() or path.suffix.lower() in IMAGE_EXTENSIONS:
+                return True
+        return False
+
+    def _set_drag_active(self, value: bool) -> None:
+        self._set_property("dragActive", value)
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() == QEvent.Enter:
+            self._set_property("hoverActive", True)
+            return False
+        if event.type() == QEvent.Leave:
+            self._update_hover_from_cursor()
+            return False
+        if event.type() == QEvent.MouseButtonRelease and event.button() == Qt.LeftButton:
+            self._show_menu()
+            return True
+        return super().eventFilter(watched, event)
+
+    def enterEvent(self, event) -> None:  # noqa: N802
+        self._set_property("hoverActive", True)
+        super().enterEvent(event)
+
+    def leaveEvent(self, event) -> None:  # noqa: N802
+        self._update_hover_from_cursor()
+        super().leaveEvent(event)
+
+    def _set_property(self, name: str, value) -> None:
+        self.setProperty(name, value)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    def _update_hover_from_cursor(self) -> None:
+        inside = self.rect().contains(self.mapFromGlobal(QCursor.pos()))
+        self._set_property("hoverActive", inside)
+
+
+class PeopleLibraryPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "People Library",
+            "Build the local gallery used by 1:N Face Search. Imported images create people named from the file name by default.",
+            parent,
+        )
+        self.selected_person_id = None
+        self.people = []
+        splitter = QSplitter(Qt.Horizontal)
+
+        left = QWidget()
+        left_layout = QVBoxLayout(left)
+        self.import_box = ImageOrFolderImportBox()
+        self.import_box.pathsSelected.connect(self.import_paths)
+        left_layout.addWidget(self.import_box)
+        left_layout.addWidget(QLabel("People"))
+        self.people_table = QTableWidget(0, 3)
+        self.people_table.setHorizontalHeaderLabels(["ID", "Thumbnail", "Name"])
+        self.people_table.setIconSize(QSize(48, 48))
+        configure_table_columns(self.people_table, [60, 90, 240])
+        self.people_table.currentCellChanged.connect(self.on_person_selected)
+        left_layout.addWidget(self.people_table, 1)
+        left_layout.addWidget(self.row(self.button("Add Person", self.add_person), self.button("Delete", self.delete_person), self.button("Export Summary", self.export_summary)))
+
+        right = QWidget()
+        right_layout = QVBoxLayout(right)
+        self.details = QTextEdit()
+        self.details.setPlaceholderText("Name, notes, and tags are shown here.")
+        right_layout.addWidget(self.details)
+        right_layout.addWidget(self.row(self.button("Save Details", self.save_details)))
+        self.samples_table = QTableWidget(0, 7)
+        self.samples_table.setHorizontalHeaderLabels(["id", "crop_path", "det_score", "quality_score", "model_name", "provider", "source_image_path"])
+        configure_table_columns(self.samples_table, [60, 260, 90, 100, 130, 110, 320])
+        right_layout.addWidget(self.samples_table)
+
+        splitter.addWidget(left)
+        splitter.addWidget(right)
+        splitter.setStretchFactor(0, 2)
+        splitter.setStretchFactor(1, 3)
+        self.content.addWidget(splitter, 1)
+        self.refresh()
+
+    def refresh(self) -> None:
+        current_id = self.selected_person_id
+        self.people = self.context.storage.list_people()
+        self.people_table.setRowCount(len(self.people))
+        for row, person in enumerate(self.people):
+            id_item = QTableWidgetItem(str(person["id"]))
+            id_item.setData(Qt.UserRole, person["id"])
+            thumb_item = QTableWidgetItem()
+            icon = self._thumbnail_icon(person.get("cover_crop_path"))
+            if icon is not None:
+                thumb_item.setIcon(icon)
+            name_item = QTableWidgetItem(person["display_name"] or person["name"])
+            name_item.setData(Qt.UserRole, person["id"])
+            self.people_table.setItem(row, 0, id_item)
+            self.people_table.setItem(row, 1, thumb_item)
+            self.people_table.setItem(row, 2, name_item)
+            self.people_table.setRowHeight(row, 54)
+            if current_id == person["id"]:
+                self.people_table.selectRow(row)
+        refresh_table_columns(self.people_table)
+        if current_id and any(person["id"] == current_id for person in self.people):
+            self.selected_person_id = current_id
+        elif self.people:
+            self.people_table.selectRow(0)
+            self.selected_person_id = self.people[0]["id"]
+        else:
+            self.selected_person_id = None
+        self._populate_details()
+        self.refresh_samples()
+
+    def on_person_selected(self, current_row: int, current_column: int, previous_row: int, previous_column: int) -> None:
+        del current_column, previous_row, previous_column
+        if current_row < 0 or current_row >= len(self.people):
+            self.selected_person_id = None
+        else:
+            self.selected_person_id = int(self.people[current_row]["id"])
+        self._populate_details()
+        self.refresh_samples()
+
+    def _populate_details(self) -> None:
+        person = next((row for row in self.people if row["id"] == self.selected_person_id), None)
+        if not person:
+            self.details.clear()
+            return
+        self.details.setPlainText(
+            "\n".join(
+                [
+                    f"name: {person['name']}",
+                    f"display_name: {person['display_name'] or ''}",
+                    f"tags: {person['tags'] or ''}",
+                    "notes:",
+                    person["notes"] or "",
+                ]
+            )
+        )
+
+    def add_person(self) -> None:
+        name, ok = QInputDialog.getText(self, "Add Person", "Person name")
+        if ok and name.strip():
+            self.context.storage.add_person(name.strip())
+            self.refresh()
+
+    def delete_person(self) -> None:
+        if self.selected_person_id is None:
+            self.show_error("Select a person first.")
+            return
+        self.context.storage.delete_person(int(self.selected_person_id), delete_samples=False)
+        self.selected_person_id = None
+        self.refresh()
+
+    def save_details(self) -> None:
+        if self.selected_person_id is None:
+            self.show_error("Select a person first.")
+            return
+        fields = {}
+        lines = self.details.toPlainText().splitlines()
+        notes_start = None
+        for idx, line in enumerate(lines):
+            if line.startswith("name:"):
+                fields["name"] = line.split(":", 1)[1].strip()
+            elif line.startswith("display_name:"):
+                fields["display_name"] = line.split(":", 1)[1].strip()
+            elif line.startswith("tags:"):
+                fields["tags"] = line.split(":", 1)[1].strip()
+            elif line.strip() == "notes:":
+                notes_start = idx + 1
+        if notes_start is not None:
+            fields["notes"] = "\n".join(lines[notes_start:]).strip()
+        self.context.storage.update_person(int(self.selected_person_id), **fields)
+        self.refresh()
+        self.set_status("Person details saved.")
+
+    def import_paths(self, paths: list[str]) -> None:
+        paths = [str(Path(path)) for path in paths if Path(path).suffix.lower() in IMAGE_EXTENSIONS]
+        if not paths:
+            self.show_error("No supported images found.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+
+        def task(progress=None, is_cancelled=None):
+            imported = 0
+            skipped = 0
+            for index, path in enumerate(paths):
+                if is_cancelled and is_cancelled():
+                    break
+                image = read_image(path)
+                if image is None:
+                    skipped += 1
+                    if progress:
+                        progress(index + 1, len(paths), f"Skipped unreadable image: {Path(path).name}")
+                    continue
+                face = self.context.engine.detect_best_face(image, source_path=path)
+                if face is None or face.normed_embedding is None:
+                    skipped += 1
+                    if progress:
+                        progress(index + 1, len(paths), f"No usable face: {Path(path).name}")
+                    continue
+                person_id = self.context.storage.add_person(Path(path).stem)
+                crop_path = Path(self.context.config.crop_dir) / f"person_{person_id}_{timestamp_for_filename()}_{index}.png"
+                if face.crop is not None:
+                    save_image(crop_path, face.crop)
+                self.context.storage.add_face_sample(
+                    person_id,
+                    face.normed_embedding,
+                    source_image_path=path,
+                    crop_path=str(crop_path),
+                    bbox=face.bbox,
+                    kps=face.kps,
+                    det_score=face.det_score,
+                    quality_score=face.quality_score,
+                    model_name=self.context.config.model_name,
+                    provider=self.context.config.provider,
+                )
+                imported += 1
+                if progress:
+                    progress(index + 1, len(paths), f"Imported {imported} people, skipped {skipped}")
+            return {"imported": imported, "skipped": skipped}
+
+        def done(result):
+            self.refresh()
+            self.set_status(f"Imported {result['imported']} people. Skipped {result['skipped']} image(s).")
+
+        self.run_task("Importing People Library", task, done)
+
+    def refresh_samples(self) -> None:
+        if self.selected_person_id is None:
+            self.samples_table.setRowCount(0)
+            return
+        samples = self.context.storage.list_face_samples(self.selected_person_id)
+        self.samples_table.setRowCount(len(samples))
+        columns = ["id", "crop_path", "det_score", "quality_score", "model_name", "provider", "source_image_path"]
+        for row, sample in enumerate(samples):
+            for col, key in enumerate(columns):
+                self.samples_table.setItem(row, col, QTableWidgetItem(str(sample.get(key, ""))))
+        refresh_table_columns(self.samples_table)
+
+    def export_summary(self) -> None:
+        path = Path(self.context.config.export_dir) / f"people_summary_{timestamp_for_filename()}.csv"
+        export_csv(path, self.context.storage.list_people())
+        self.set_status(f"People summary exported to {path}")
+
+    @staticmethod
+    def _thumbnail_icon(path: str | None) -> QIcon | None:
+        if not path:
+            return None
+        pixmap = QPixmap(str(Path(path)))
+        if pixmap.isNull():
+            return None
+        return QIcon(pixmap.scaled(48, 48, Qt.KeepAspectRatio, Qt.SmoothTransformation))
diff --git a/python-package/insightface/gui/pages/placeholder_page.py b/python-package/insightface/gui/pages/placeholder_page.py
new file mode 100644
index 000000000..e825c95f8
--- /dev/null
+++ b/python-package/insightface/gui/pages/placeholder_page.py
@@ -0,0 +1,73 @@
+"""Reusable placeholder and commercial next-steps pages."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QLabel, QPushButton
+
+from ..core.tooltips import set_button_tooltip
+from .base import BasePage
+
+
+class PlaceholderPage(BasePage):
+    def __init__(
+        self,
+        context,
+        title: str,
+        description: str,
+        coming_soon_text: str = "Coming soon.",
+        ctas: list[tuple[str, str]] | None = None,
+        parent=None,
+    ):
+        super().__init__(context, title, description, parent)
+        status = QLabel(coming_soon_text)
+        status.setWordWrap(True)
+        status.setProperty("role", "muted")
+        status.setStyleSheet("font-size: 16px; font-weight: 600; padding:12px 0;")
+        self.content.addWidget(status)
+        if ctas:
+            buttons = []
+            for label, page_key in ctas:
+                button = QPushButton(label)
+                button.clicked.connect(lambda checked=False, target=page_key: self.window().open_page(target))
+                set_button_tooltip(button)
+                buttons.append(button)
+            self.content.addWidget(self.row(*buttons))
+        self.content.addStretch(1)
+
+
+class CommercialNextStepsPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Commercial Next Steps",
+            "Understand the business paths after local enterprise evaluation.",
+            parent,
+        )
+        self.content.addWidget(
+            self.notice(
+                "Production or commercial deployment requires an appropriate commercial model license. "
+                "This page summarizes common commercial paths; it does not provide legal advice."
+            )
+        )
+        summary = QLabel(
+            "\n".join(
+                [
+                    "Commercial model license: authorize production use of approved model files.",
+                    "Private model evaluation: validate models on your local business data before purchase.",
+                    "SDK / API access: integrate InsightFace capabilities into internal products or services.",
+                    "SLA and enterprise support: align support expectations for production deployment.",
+                    "Custom training: evaluate private data requirements, consent, retention, and compliance before any training workflow.",
+                ]
+            )
+        )
+        summary.setWordWrap(True)
+        summary.setStyleSheet("line-height: 1.35;")
+        self.content.addWidget(summary)
+        self.content.addWidget(
+            self.row(
+                self.button("Open License Center", lambda: self.window().open_license_dialog()),
+                self.button("Run Evaluation", lambda: self.window().open_page("enterprise_evaluation")),
+                self.button("Open Reports", lambda: self.window().open_page("reports")),
+            )
+        )
+        self.content.addStretch(1)
diff --git a/python-package/insightface/gui/pages/reports_page.py b/python-package/insightface/gui/pages/reports_page.py
new file mode 100644
index 000000000..f1fa66a60
--- /dev/null
+++ b/python-package/insightface/gui/pages/reports_page.py
@@ -0,0 +1,77 @@
+"""Reports page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtCore import QUrl
+from PySide6.QtGui import QDesktopServices
+from PySide6.QtWidgets import QTableWidget, QTableWidgetItem
+
+from ..core.exporters import export_csv
+from ..core.models import EvaluationResult
+from ..core.reporting import write_reports
+from ..core.utils import timestamp_for_filename, utc_now_iso
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from .base import BasePage
+
+
+class ReportsPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Reports", "Browse and export historical enterprise evaluation runs.", parent)
+        self.runs = []
+        self.content.addWidget(self.row(self.button("Refresh", self.refresh), self.button("Open Selected Report", self.open_selected), self.button("Re-export Selected", self.reexport_selected), self.button("Delete Selected", self.delete_selected), self.button("Export All Summary", self.export_summary)))
+        self.table = QTableWidget(0, 7)
+        self.table.setHorizontalHeaderLabels(["id", "date", "scenario", "model", "provider", "threshold", "report_path"])
+        configure_table_columns(self.table, [60, 180, 180, 130, 110, 90, 360])
+        self.content.addWidget(self.table, 1)
+        self.refresh()
+
+    def refresh(self) -> None:
+        self.runs = self.context.storage.list_evaluation_runs()
+        self.table.setRowCount(len(self.runs))
+        for row_index, run in enumerate(self.runs):
+            values = [run["id"], run["created_at"], run["scenario"], run["model_name"], run["provider"], run["threshold"], run["report_path"]]
+            for col, value in enumerate(values):
+                self.table.setItem(row_index, col, QTableWidgetItem(str(value)))
+        refresh_table_columns(self.table)
+
+    def _selected_run(self):
+        row = self.table.currentRow()
+        if row < 0 or row >= len(self.runs):
+            self.show_error("Select a report row first.")
+            return None
+        return self.runs[row]
+
+    def open_selected(self) -> None:
+        run = self._selected_run()
+        if not run:
+            return
+        path = run.get("report_path") or ""
+        if not path:
+            self.show_error("This run has no report path.")
+            return
+        QDesktopServices.openUrl(QUrl.fromLocalFile(path))
+
+    def reexport_selected(self) -> None:
+        run = self._selected_run()
+        if not run:
+            return
+        result = EvaluationResult(scenario=run["scenario"], model_name=run["model_name"], provider=run["provider"], threshold=float(run["threshold"]), dataset_summary=run.get("dataset_summary", {}), metrics=run.get("metrics", {}), errors=[], latency=run.get("hardware", {}), license_status=self.context.config.license_status, created_at=run.get("created_at") or utc_now_iso())
+        paths = write_reports(result, self.context.config.report_dir, language=self.context.config.ui_language)
+        report_path = paths.get("pdf") or paths["markdown"]
+        self.context.storage.save_evaluation_run(result.scenario, result.model_name, result.provider, result.threshold, result.dataset_summary, result.metrics, result.latency, report_path, result.created_at)
+        self.refresh()
+        self.set_status(f"Report re-exported to {report_path}")
+
+    def delete_selected(self) -> None:
+        run = self._selected_run()
+        if not run:
+            return
+        self.context.storage.delete_evaluation_run(int(run["id"]))
+        self.refresh()
+
+    def export_summary(self) -> None:
+        path = Path(self.context.config.export_dir) / f"evaluation_summary_{timestamp_for_filename()}.csv"
+        export_csv(path, self.runs)
+        self.set_status(f"Evaluation summary exported to {path}")
diff --git a/python-package/insightface/gui/pages/settings_page.py b/python-package/insightface/gui/pages/settings_page.py
new file mode 100644
index 000000000..213c06b41
--- /dev/null
+++ b/python-package/insightface/gui/pages/settings_page.py
@@ -0,0 +1,113 @@
+"""Application settings page."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from PySide6.QtCore import QUrl
+from PySide6.QtGui import QDesktopServices
+from PySide6.QtWidgets import QCheckBox, QComboBox, QDoubleSpinBox, QFormLayout, QLineEdit, QSpinBox
+
+from ..core.config import AppConfig, save_config
+from ..core.paths import ensure_workspace
+from ..core.theme import THEME_OPTIONS, normalize_theme
+from .base import BasePage
+
+
+class SettingsPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Settings", "Configure workspace, database, export paths, privacy defaults, and UI preferences.", parent)
+        cfg = context.config
+        self.workspace = QLineEdit(cfg.workspace_path)
+        self.database = QLineEdit(cfg.database_path)
+        self.crops = QLineEdit(cfg.crop_dir)
+        self.reports = QLineEdit(cfg.report_dir)
+        self.exports = QLineEdit(cfg.export_dir)
+        self.cache = QLineEdit(cfg.cache_dir)
+        self.threshold = QDoubleSpinBox()
+        self.threshold.setRange(0, 1)
+        self.threshold.setSingleStep(0.01)
+        self.threshold.setValue(cfg.recognition_threshold)
+        self.top_k = QSpinBox()
+        self.top_k.setRange(1, 100)
+        self.top_k.setValue(cfg.default_top_k)
+        self.min_det = QDoubleSpinBox()
+        self.min_det.setRange(0, 1)
+        self.min_det.setSingleStep(0.01)
+        self.min_det.setValue(cfg.min_detection_score)
+        self.min_face = QSpinBox()
+        self.min_face.setRange(1, 4096)
+        self.min_face.setValue(cfg.min_face_size)
+        self.save_crops = QCheckBox("Save crops")
+        self.save_crops.setChecked(cfg.save_crops)
+        self.save_logs = QCheckBox("Save recognition logs")
+        self.save_logs.setChecked(cfg.save_recognition_logs)
+        self.anonymize = QCheckBox("Anonymize report paths")
+        self.anonymize.setChecked(cfg.anonymize_report_paths)
+        self.theme = QComboBox()
+        for option in THEME_OPTIONS:
+            self.theme.addItem(option.label, option.value)
+        self.theme.setCurrentIndex(max(0, self.theme.findData(normalize_theme(cfg.ui_theme))))
+        form = QFormLayout()
+        for label, widget in [
+            ("Workspace path", self.workspace),
+            ("Database path", self.database),
+            ("Crop output directory", self.crops),
+            ("Report output directory", self.reports),
+            ("Export output directory", self.exports),
+            ("Cache directory", self.cache),
+            ("Default recognition threshold", self.threshold),
+            ("Default Top-K", self.top_k),
+            ("Minimum detection score", self.min_det),
+            ("Minimum face size", self.min_face),
+            ("UI theme", self.theme),
+        ]:
+            form.addRow(label, widget)
+        form.addRow("Privacy", self.row(self.save_crops, self.save_logs, self.anonymize))
+        self.content.addLayout(form)
+        self.content.addWidget(self.row(self.button("Save Settings", self.save), self.button("Open Workspace Folder", self.open_workspace), self.button("Export Settings", self.export_settings), self.button("Import Settings", self.import_settings), self.button("Reset Settings", self.reset)))
+
+    def save(self) -> None:
+        cfg = self.context.config
+        cfg.workspace_path = self.workspace.text().strip()
+        paths = ensure_workspace(cfg.workspace_path)
+        cfg.database_path = self.database.text().strip() or str(paths["database"])
+        cfg.crop_dir = self.crops.text().strip() or str(paths["crops"])
+        cfg.report_dir = self.reports.text().strip() or str(paths["reports"])
+        cfg.export_dir = self.exports.text().strip() or str(paths["exports"])
+        cfg.cache_dir = self.cache.text().strip() or str(paths["cache"])
+        cfg.recognition_threshold = float(self.threshold.value())
+        cfg.default_top_k = int(self.top_k.value())
+        cfg.min_detection_score = float(self.min_det.value())
+        cfg.min_face_size = int(self.min_face.value())
+        cfg.save_crops = self.save_crops.isChecked()
+        cfg.save_recognition_logs = self.save_logs.isChecked()
+        cfg.anonymize_report_paths = self.anonymize.isChecked()
+        cfg.ui_theme = self.theme.currentData()
+        save_config(cfg)
+        self.set_status("Settings saved.")
+
+    def open_workspace(self) -> None:
+        QDesktopServices.openUrl(QUrl.fromLocalFile(self.workspace.text().strip()))
+
+    def export_settings(self) -> None:
+        path = self.save_file("Export settings", str(Path(self.context.config.export_dir) / "insightface_gui_settings.json"), "JSON (*.json)")
+        if not path:
+            return
+        Path(path).write_text(json.dumps(self.context.config.to_dict(), indent=2), encoding="utf-8")
+        self.set_status(f"Settings exported to {path}")
+
+    def import_settings(self) -> None:
+        path = self.choose_file("Import settings", "JSON (*.json)")
+        if not path:
+            return
+        data = json.loads(Path(path).read_text(encoding="utf-8"))
+        self.context.config = AppConfig.from_dict(data)
+        save_config(self.context.config)
+        self.set_status("Settings imported. Restart recommended.")
+
+    def reset(self) -> None:
+        self.context.config = AppConfig()
+        save_config(self.context.config)
+        self.set_status("Settings reset to defaults. Restart recommended.")
diff --git a/python-package/insightface/gui/pages/verification_page.py b/python-package/insightface/gui/pages/verification_page.py
new file mode 100644
index 000000000..d8a0817a0
--- /dev/null
+++ b/python-package/insightface/gui/pages/verification_page.py
@@ -0,0 +1,696 @@
+"""Combined Query and Gallery verification workflow."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+from PySide6.QtCore import QDir, QEvent, QSize, Qt, QUrl, Signal
+from PySide6.QtGui import QCursor, QDesktopServices, QIcon, QPixmap
+from PySide6.QtWidgets import (
+    QComboBox,
+    QDoubleSpinBox,
+    QFrame,
+    QHBoxLayout,
+    QLabel,
+    QListWidget,
+    QListWidgetItem,
+    QSizePolicy,
+    QAbstractItemView,
+    QDialog,
+    QDialogButtonBox,
+    QFileSystemModel,
+    QPushButton,
+    QSplitter,
+    QTableWidget,
+    QTableWidgetItem,
+    QTreeView,
+    QVBoxLayout,
+    QWidget,
+)
+
+from ..core.config import save_config
+from ..core.constants import DEFAULT_THRESHOLD, IMAGE_EXTENSIONS
+from ..core.evaluation import (
+    MULTI_FACE_REQUIRE_ONE,
+    MULTI_FACE_SKIP,
+    MULTI_FACE_USE_CENTERED_LARGEST,
+    MULTI_FACE_USE_LARGEST,
+    multi_face_policy_help,
+    select_face_by_policy,
+)
+from ..core.i18n import apply_translations, tr
+from ..core.recognition import compare_embeddings
+from ..core.tooltips import apply_button_tooltips, set_button_tooltip
+from ..core.utils import list_images, read_image
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class GalleryUploadPanel(QFrame):
+    """Clickable and draggable gallery input for image files or folders."""
+
+    pathsChanged = Signal(object)
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self._paths: list[str] = []
+        self.setObjectName("galleryUpload")
+        self.setAcceptDrops(True)
+        self.setMouseTracking(True)
+        self.setMinimumHeight(300)
+        self.setProperty("hoverActive", False)
+        self.setProperty("dragActive", False)
+        self.setProperty("hasFiles", False)
+
+        layout = QVBoxLayout(self)
+        layout.setContentsMargins(10, 10, 10, 10)
+        header = QHBoxLayout()
+        title = QLabel("Gallery")
+        title.setStyleSheet("font-size:15px; font-weight:700;")
+        self.count_label = QLabel("No images")
+        self.count_label.setProperty("role", "muted")
+        self.remove_button = QPushButton("Remove Selected")
+        self.remove_button.clicked.connect(self.remove_selected)
+        self.clear_button = QPushButton("Clear")
+        self.clear_button.clicked.connect(self.clear)
+        set_button_tooltip(self.remove_button)
+        set_button_tooltip(self.clear_button)
+        header.addWidget(title)
+        header.addWidget(self.count_label)
+        header.addStretch(1)
+        header.addWidget(self.remove_button)
+        header.addWidget(self.clear_button)
+        layout.addLayout(header)
+
+        self.prompt = QLabel("Click to upload images or a folder, or drag them here")
+        self.prompt.setObjectName("uploadPrompt")
+        self.prompt.setAlignment(Qt.AlignCenter)
+        self.prompt.setWordWrap(True)
+        self.prompt.setSizePolicy(QSizePolicy.Expanding, QSizePolicy.Expanding)
+        layout.addWidget(self.prompt, 1)
+
+        self.single_preview = QLabel()
+        self.single_preview.setAlignment(Qt.AlignCenter)
+        self.single_preview.setMinimumHeight(180)
+        self.single_preview.setStyleSheet("border:0; background:transparent;")
+        self.single_preview.hide()
+        self.single_hint = QLabel("")
+        self.single_hint.setObjectName("pathLabel")
+        self.single_hint.setAlignment(Qt.AlignCenter)
+        self.single_hint.setWordWrap(True)
+        self.single_hint.hide()
+        layout.addWidget(self.single_preview, 1)
+        layout.addWidget(self.single_hint)
+
+        self.list_widget = QListWidget()
+        self.list_widget.setIconSize(QSize(56, 56))
+        self.list_widget.itemDoubleClicked.connect(self.open_item)
+        layout.addWidget(self.list_widget, 1)
+        self.list_widget.hide()
+        self.remove_button.hide()
+        self.clear_button.hide()
+
+        for watched in (self, self.prompt, self.single_preview, self.single_hint, self.list_widget, self.list_widget.viewport()):
+            watched.installEventFilter(self)
+
+    def paths(self) -> list[str]:
+        return list(self._paths)
+
+    def set_paths(self, paths: list[str], emit: bool = True) -> None:
+        resolved = self._expand_paths(paths)
+        self._paths = resolved
+        self._render()
+        if emit:
+            self.pathsChanged.emit(self.paths())
+
+    def add_paths(self, paths: list[str]) -> None:
+        self.set_paths(self._paths + paths)
+
+    def clear(self) -> None:
+        self.set_paths([])
+
+    def remove_selected(self) -> None:
+        remove = {item.data(Qt.UserRole) for item in self.list_widget.selectedItems()}
+        if not remove:
+            return
+        self.set_paths([path for path in self._paths if path not in remove])
+
+    def browse(self) -> None:
+        dialog = QDialog(self)
+        dialog.setWindowTitle("Select gallery images or folders")
+        dialog.resize(760, 520)
+        layout = QVBoxLayout(dialog)
+        hint = QLabel("Select one or more image files or folders. Folders are scanned recursively.")
+        hint.setWordWrap(True)
+        layout.addWidget(hint)
+        model = QFileSystemModel(dialog)
+        model.setRootPath(str(Path.home()))
+        model.setFilter(QDir.AllEntries | QDir.NoDotAndDotDot | QDir.Readable)
+        model.setNameFilters(["*.jpg", "*.jpeg", "*.png", "*.bmp", "*.webp"])
+        model.setNameFilterDisables(False)
+        tree = QTreeView(dialog)
+        tree.setModel(model)
+        tree.setRootIndex(model.index(str(Path.home())))
+        tree.setSelectionMode(QAbstractItemView.ExtendedSelection)
+        tree.setColumnWidth(0, 320)
+        tree.doubleClicked.connect(lambda index: tree.setRootIndex(index) if model.isDir(index) else None)
+        layout.addWidget(tree, 1)
+        buttons = QDialogButtonBox(QDialogButtonBox.Ok | QDialogButtonBox.Cancel)
+        buttons.accepted.connect(dialog.accept)
+        buttons.rejected.connect(dialog.reject)
+        layout.addWidget(buttons)
+        apply_button_tooltips(dialog)
+        apply_translations(dialog, self._language())
+        if dialog.exec() == QDialog.Accepted:
+            indexes = tree.selectionModel().selectedRows(0)
+            paths = [model.filePath(index) for index in indexes]
+            if paths:
+                self.add_paths(paths)
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() in (QEvent.DragEnter, QEvent.DragMove):
+            if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+                self._set_property("dragActive", True)
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.DragLeave:
+            self._set_property("dragActive", False)
+            return False
+        if event.type() == QEvent.Drop:
+            self._set_property("dragActive", False)
+            paths = [url.toLocalFile() for url in event.mimeData().urls() if url.isLocalFile()]
+            expanded = self._expand_paths(paths)
+            if expanded:
+                self.add_paths(expanded)
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.Enter:
+            self._set_property("hoverActive", True)
+            return False
+        if event.type() == QEvent.Leave:
+            self._update_hover_from_cursor()
+            return False
+        if event.type() == QEvent.MouseButtonRelease and event.button() == Qt.LeftButton and watched in {self.prompt, self.single_preview, self.single_hint}:
+            self.browse()
+            return True
+        return super().eventFilter(watched, event)
+
+    def mouseReleaseEvent(self, event) -> None:  # noqa: N802
+        if event.button() == Qt.LeftButton and not self._paths:
+            self.browse()
+            event.accept()
+            return
+        super().mouseReleaseEvent(event)
+
+    def enterEvent(self, event) -> None:  # noqa: N802
+        self._set_property("hoverActive", True)
+        super().enterEvent(event)
+
+    def leaveEvent(self, event) -> None:  # noqa: N802
+        self._update_hover_from_cursor()
+        super().leaveEvent(event)
+
+    def open_item(self, item: QListWidgetItem) -> None:
+        path = item.data(Qt.UserRole)
+        if path:
+            QDesktopServices.openUrl(QUrl.fromLocalFile(str(path)))
+
+    def _render(self) -> None:
+        self.list_widget.clear()
+        self.single_preview.clear()
+        for path in self._paths:
+            item = QListWidgetItem(Path(path).name)
+            icon = self._icon(path, QSize(56, 56))
+            if icon is not None:
+                item.setIcon(icon)
+            item.setToolTip(path)
+            item.setData(Qt.UserRole, path)
+            self.list_widget.addItem(item)
+        has_files = bool(self._paths)
+        single = len(self._paths) == 1
+        self.prompt.setVisible(not has_files)
+        self.single_preview.setVisible(single)
+        self.single_hint.setVisible(single)
+        self.list_widget.setVisible(has_files and not single)
+        self.remove_button.setVisible(has_files and not single)
+        self.clear_button.setVisible(has_files)
+        count = len(self._paths)
+        language = self._language()
+        if count:
+            image_label = "image" if count == 1 else "images"
+            self.count_label.setText(f"{count} {tr(image_label, language)}")
+        else:
+            self.count_label.setText(tr("No images", language))
+        if single:
+            self._update_single_preview()
+        self._set_property("hasFiles", has_files)
+
+    def resizeEvent(self, event) -> None:  # noqa: N802
+        super().resizeEvent(event)
+        if len(self._paths) == 1:
+            self._update_single_preview()
+
+    def _expand_paths(self, paths: list[str]) -> list[str]:
+        collected: list[str] = []
+        seen = set()
+        for raw_path in paths:
+            path = Path(raw_path).expanduser()
+            candidates: list[Path]
+            if path.is_dir():
+                candidates = list_images(path, recursive=True)
+            elif path.is_file() and path.suffix.lower() in IMAGE_EXTENSIONS:
+                candidates = [path]
+            else:
+                candidates = []
+            for candidate in candidates:
+                value = str(candidate)
+                if value not in seen and value not in collected:
+                    collected.append(value)
+                    seen.add(value)
+        return collected
+
+    def _accepted_urls(self, urls) -> bool:
+        for url in urls:
+            if not url.isLocalFile():
+                continue
+            path = Path(url.toLocalFile()).expanduser()
+            if path.is_dir() or (path.is_file() and path.suffix.lower() in IMAGE_EXTENSIONS):
+                return True
+        return False
+
+    def _set_property(self, name: str, value) -> None:
+        self.setProperty(name, value)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    def _update_hover_from_cursor(self) -> None:
+        inside = self.rect().contains(self.mapFromGlobal(QCursor.pos()))
+        self._set_property("hoverActive", inside)
+
+    @staticmethod
+    def _icon(path: str | None, size: QSize) -> QIcon | None:
+        if not path:
+            return None
+        pixmap = QPixmap(str(Path(path)))
+        if pixmap.isNull():
+            return None
+        return QIcon(pixmap.scaled(size, Qt.KeepAspectRatio, Qt.SmoothTransformation))
+
+    def _update_single_preview(self) -> None:
+        if len(self._paths) != 1:
+            return
+        path = self._paths[0]
+        pixmap = QPixmap(str(Path(path)))
+        if pixmap.isNull():
+            self.single_preview.setText(Path(path).name)
+        else:
+            size = self.single_preview.size()
+            target = QSize(max(180, size.width() - 24), max(150, size.height() - 12))
+            self.single_preview.setPixmap(pixmap.scaled(target, Qt.KeepAspectRatio, Qt.SmoothTransformation))
+        self.single_hint.setText(
+            f"{Path(path).name}\n{tr('Click or drag more images or folders to add to Gallery.', self._language())}"
+        )
+
+    def _language(self) -> str | None:
+        context = getattr(self.window(), "context", None)
+        config = getattr(context, "config", None)
+        return getattr(config, "ui_language", None)
+
+
+class VerificationPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(
+            context,
+            "Face Recognition",
+            "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.",
+            parent,
+        )
+        self.query_path = ""
+        self.query_image: np.ndarray | None = None
+        self.gallery_paths: list[str] = []
+        self._gallery_embedding_cache_key: tuple[str, ...] | None = None
+        self._gallery_embedding_cache: list[dict] | None = None
+        self.results: list[dict] = []
+        if abs(float(context.config.recognition_threshold) - DEFAULT_THRESHOLD) > 1e-9:
+            context.config.recognition_threshold = DEFAULT_THRESHOLD
+            save_config(context.config)
+
+        self.workflow_notice = self.notice("")
+        self.content.addWidget(self.workflow_notice)
+        splitter = QSplitter(Qt.Horizontal)
+        left = QWidget()
+        left_layout = QVBoxLayout(left)
+        left_layout.setContentsMargins(0, 0, 0, 0)
+        self.query_input = UploadPreview(
+            "Query",
+            extensions=[".jpg", ".jpeg", ".png", ".bmp", ".webp"],
+            dialog_filter="Images (*.jpg *.jpeg *.png *.bmp *.webp);;All Files (*)",
+        )
+        self.query_input.pathChanged.connect(lambda path: self.load_query(path) if path else self.clear_query())
+        left_layout.addWidget(self.query_input)
+        splitter.addWidget(left)
+
+        right = QWidget()
+        right_layout = QVBoxLayout(right)
+        right_layout.setContentsMargins(0, 0, 0, 0)
+        self.gallery_input = GalleryUploadPanel()
+        self.gallery_input.pathsChanged.connect(self.set_gallery_paths)
+        right_layout.addWidget(self.gallery_input)
+        splitter.addWidget(right)
+        splitter.setStretchFactor(0, 1)
+        splitter.setStretchFactor(1, 1)
+        self.content.addWidget(splitter, 1)
+
+        control_row = QHBoxLayout()
+        self.threshold = QDoubleSpinBox()
+        self.threshold.setRange(0.01, 0.99)
+        self.threshold.setSingleStep(0.01)
+        self.threshold.setDecimals(2)
+        self.threshold.setValue(DEFAULT_THRESHOLD)
+        self.threshold.setToolTip("Similarity threshold for matching query and gallery faces.")
+        self.multi_face_policy = QComboBox()
+        self.multi_face_policy.setProperty("i18nItems", True)
+        self.multi_face_policy.addItem("Require exactly one face", MULTI_FACE_REQUIRE_ONE)
+        self.multi_face_policy.addItem("Use largest face", MULTI_FACE_USE_LARGEST)
+        self.multi_face_policy.addItem("Use largest centered face", MULTI_FACE_USE_CENTERED_LARGEST)
+        self.multi_face_policy.addItem("Mark as skip", MULTI_FACE_SKIP)
+        self.multi_face_policy.setObjectName("verificationMultiFacePolicy")
+        self.multi_face_policy.setToolTip("Choose how query and gallery images handle multiple detected faces.")
+        default_policy_index = self.multi_face_policy.findData(MULTI_FACE_USE_CENTERED_LARGEST)
+        if default_policy_index >= 0:
+            self.multi_face_policy.setCurrentIndex(default_policy_index)
+        self.multi_face_policy.currentIndexChanged.connect(self._multi_face_policy_changed)
+        self.run_button = QPushButton("Run Recognition")
+        self.run_button.clicked.connect(self.run_verification)
+        self.clear_button = QPushButton("Clear")
+        self.clear_button.clicked.connect(self.clear_all)
+        set_button_tooltip(self.run_button)
+        set_button_tooltip(self.clear_button)
+        control_row.addWidget(QLabel("Recognition threshold"))
+        control_row.addWidget(self.threshold)
+        control_row.addWidget(QLabel("Multi-face handling"))
+        control_row.addWidget(self.multi_face_policy)
+        control_row.addStretch(1)
+        control_row.addWidget(self.run_button)
+        control_row.addWidget(self.clear_button)
+        self.content.addLayout(control_row)
+
+        self.result_table = QTableWidget(0, 7)
+        self.result_table.setIconSize(QSize(56, 56))
+        self.result_table.setSelectionBehavior(QAbstractItemView.SelectRows)
+        self.result_table.setSelectionMode(QAbstractItemView.SingleSelection)
+        self.result_table.setEditTriggers(QAbstractItemView.NoEditTriggers)
+        self.result_table.setHorizontalHeaderLabels(
+            ["rank", "thumbnail", "gallery_file", "similarity", "threshold", "decision", "det_score"]
+        )
+        configure_table_columns(self.result_table, [56, 86, 320, 92, 92, 130, 92])
+        self.result_table.cellDoubleClicked.connect(self.open_result)
+        self.content.addWidget(self.result_table, 1)
+        self._update_policy_notice()
+
+    def load_query(self, path: str) -> None:
+        image = read_image(path)
+        if image is None:
+            self.query_input.clear(emit=False)
+            self.show_error("Image read failure.")
+            return
+        self.query_path = path
+        self.query_image = image
+        self.query_input.set_image(image, path)
+        self._update_mode_label()
+
+    def clear_query(self) -> None:
+        self.query_path = ""
+        self.query_image = None
+        self.query_input.clear(emit=False)
+        self.query_input.set_faces([])
+        self._update_mode_label()
+
+    def set_gallery_paths(self, paths: list[str]) -> None:
+        new_paths = list(paths)
+        if tuple(new_paths) != tuple(self.gallery_paths):
+            self._clear_gallery_embedding_cache()
+        self.gallery_paths = new_paths
+        self.results = []
+        self.result_table.setRowCount(0)
+        self._update_mode_label()
+
+    def clear_all(self) -> None:
+        self.clear_query()
+        self.gallery_input.clear()
+        self.results = []
+        self.result_table.setRowCount(0)
+        self.set_status("Cleared.")
+
+    def run_verification(self) -> None:
+        if self.query_image is None or not self.query_path:
+            self.show_error("Please upload a query image.")
+            return
+        if not self.gallery_paths:
+            self.show_error("Please upload one or more gallery images or folders.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+
+        query_image = self.query_image.copy()
+        query_path = self.query_path
+        gallery_paths = list(self.gallery_paths)
+        multi_face_policy = self._multi_face_policy()
+        gallery_key = (multi_face_policy, *gallery_paths)
+        cached_gallery = (
+            list(self._gallery_embedding_cache)
+            if self._gallery_embedding_cache_key == gallery_key and self._gallery_embedding_cache is not None
+            else None
+        )
+        threshold = self.threshold.value()
+
+        def task(progress=None, is_cancelled=None):
+            query_faces = self.context.engine.detect_faces(query_image, source_path=query_path)
+            query_face = select_face_by_policy(
+                query_faces,
+                query_image.shape,
+                multi_face_policy,
+                path=query_path,
+                stage="query",
+            )
+            if query_face is None or query_face.normed_embedding is None:
+                raise ValueError("No face detected in the query image or embedding unavailable.")
+            if cached_gallery is not None:
+                gallery_cache = cached_gallery
+                if progress:
+                    progress(1, 1, "Using cached gallery embeddings")
+            else:
+                gallery_cache = self._build_gallery_embedding_cache(
+                    gallery_paths,
+                    multi_face_policy,
+                    progress,
+                    is_cancelled,
+                )
+            rows = self._compare_query_to_gallery_cache(query_face.normed_embedding, gallery_cache, threshold)
+            rows.sort(
+                key=lambda item: item["similarity"] if isinstance(item.get("similarity"), (int, float)) else -1.0,
+                reverse=True,
+            )
+            for rank, row in enumerate(rows, start=1):
+                row["rank"] = rank
+            mode = "1:1 Compare" if len(gallery_paths) == 1 else "1:N Gallery Search"
+            return {
+                "query_bbox": query_face.bbox,
+                "results": rows,
+                "mode": mode,
+                "query_path": query_path,
+                "gallery_paths": tuple(gallery_paths),
+                "gallery_key": gallery_key,
+                "gallery_cache": gallery_cache,
+                "multi_face_policy": multi_face_policy,
+            }
+
+        def done(payload):
+            if (
+                self.query_path != payload["query_path"]
+                or tuple(self.gallery_paths) != payload["gallery_paths"]
+                or self._multi_face_policy() != payload["multi_face_policy"]
+            ):
+                self.set_status("Recognition result ignored because query, gallery, or multi-face handling changed during processing.")
+                return
+            if cached_gallery is None:
+                self._gallery_embedding_cache_key = payload["gallery_key"]
+                self._gallery_embedding_cache = payload["gallery_cache"]
+            self.results = payload["results"]
+            self.query_input.set_faces([{"bbox": payload["query_bbox"], "label": "Query"}])
+            self._populate_results()
+            cache_status = "cached gallery embeddings" if cached_gallery is None else "reused cached gallery embeddings"
+            self.set_status(f"{payload['mode']} complete with {cache_status}. {len(self.results)} gallery result(s).")
+
+        self.run_task("Running verification", task, done)
+
+    def open_result(self, row: int, column: int) -> None:
+        del column
+        if row < 0 or row >= len(self.results):
+            return
+        path = self.results[row].get("path")
+        if path:
+            QDesktopServices.openUrl(QUrl.fromLocalFile(str(path)))
+
+    def _populate_results(self) -> None:
+        self.result_table.setRowCount(len(self.results))
+        for row_index, result in enumerate(self.results):
+            values = [
+                result.get("rank", ""),
+                "",
+                Path(str(result.get("path", ""))).name if result.get("path") else "",
+                f"{float(result.get('similarity', 0.0)):.4f}" if result.get("similarity") is not None else "",
+                f"{float(result.get('threshold', 0.0)):.4f}" if result.get("threshold") is not None else "",
+                result.get("decision", ""),
+                f"{float(result.get('det_score', 0.0)):.4f}" if result.get("det_score") is not None else "",
+            ]
+            for col, value in enumerate(values):
+                item = QTableWidgetItem(str(value))
+                if col == 1:
+                    icon = GalleryUploadPanel._icon(result.get("path"), QSize(56, 56))
+                    if icon is not None:
+                        item.setIcon(icon)
+                item.setToolTip(str(result.get("notes") or result.get("path") or ""))
+                item.setData(Qt.UserRole, result.get("path", ""))
+                self.result_table.setItem(row_index, col, item)
+            self.result_table.setRowHeight(row_index, 64)
+        refresh_table_columns(self.result_table)
+
+    def _update_mode_label(self) -> None:
+        pass
+
+    def _clear_gallery_embedding_cache(self) -> None:
+        self._gallery_embedding_cache_key = None
+        self._gallery_embedding_cache = None
+
+    def _build_gallery_embedding_cache(
+        self,
+        gallery_paths: list[str],
+        multi_face_policy: str,
+        progress=None,
+        is_cancelled=None,
+    ) -> list[dict]:
+        cache: list[dict] = []
+        for index, path in enumerate(gallery_paths):
+            if is_cancelled and is_cancelled():
+                break
+            image = read_image(path)
+            if image is None:
+                cache.append(self._gallery_error_cache_item(path, "Image read failure."))
+            else:
+                try:
+                    faces = self.context.engine.detect_faces(image, source_path=path)
+                    face = select_face_by_policy(
+                        faces,
+                        image.shape,
+                        multi_face_policy,
+                        path=path,
+                        stage="gallery",
+                    )
+                    if face is None or face.normed_embedding is None:
+                        cache.append(self._gallery_error_cache_item(path, "No face detected."))
+                    else:
+                        cache.append(
+                            {
+                                "path": path,
+                                "embedding": np.asarray(face.normed_embedding, dtype=np.float32).copy(),
+                                "det_score": face.det_score,
+                                "bbox": face.bbox,
+                                "notes": "",
+                            }
+                        )
+                except Exception as exc:
+                    cache.append(self._gallery_error_cache_item(path, str(exc)))
+            if progress:
+                progress(
+                    index + 1,
+                    len(gallery_paths),
+                    f"Computed gallery embeddings for {index + 1} of {len(gallery_paths)} images",
+                )
+        return cache
+
+    def _compare_query_to_gallery_cache(
+        self, query_embedding: np.ndarray, gallery_cache: list[dict], threshold: float
+    ) -> list[dict]:
+        rows = []
+        for item in gallery_cache:
+            embedding = item.get("embedding")
+            if embedding is None:
+                rows.append(
+                    self._error_row(
+                        str(item.get("path", "")),
+                        str(item.get("notes") or "No face detected."),
+                        threshold,
+                    )
+                )
+                continue
+            result = compare_embeddings(query_embedding, embedding, threshold)
+            rows.append(
+                {
+                    "rank": 0,
+                    "path": item.get("path", ""),
+                    "similarity": float(result["similarity"]),
+                    "threshold": threshold,
+                    "decision": str(result["decision"]),
+                    "det_score": item.get("det_score"),
+                    "bbox": item.get("bbox"),
+                    "notes": item.get("notes", ""),
+                }
+            )
+        return rows
+
+    @staticmethod
+    def _gallery_error_cache_item(path: str, message: str) -> dict:
+        return {
+            "path": path,
+            "embedding": None,
+            "det_score": None,
+            "bbox": None,
+            "notes": message,
+        }
+
+    @staticmethod
+    def _error_row(path: str, message: str, threshold: float) -> dict:
+        return {
+            "rank": 0,
+            "path": path,
+            "similarity": None,
+            "threshold": threshold,
+            "decision": message,
+            "det_score": None,
+            "bbox": None,
+            "notes": message,
+        }
+
+    def _multi_face_policy(self) -> str:
+        return str(self.multi_face_policy.currentData() or MULTI_FACE_REQUIRE_ONE)
+
+    def _multi_face_policy_changed(self) -> None:
+        self._clear_gallery_embedding_cache()
+        if hasattr(self, "result_table"):
+            self.results = []
+            self.result_table.setRowCount(0)
+        if hasattr(self, "query_input"):
+            self.query_input.set_faces([])
+        self.set_status("Multi-face handling changed. Gallery embeddings will be recomputed.")
+        self._update_policy_notice()
+
+    def _update_policy_notice(self) -> None:
+        language = self.context.config.ui_language
+        self.workflow_notice.setText(
+            tr(
+                "Gallery face embeddings are cached in memory after the first run and reused while the gallery image list "
+                "and multi-face policy are unchanged. Changing only the query image reuses the cached gallery; adding, "
+                "removing, clearing gallery images, or changing multi-face handling clears and recomputes it.",
+                language,
+            )
+            + " "
+            + tr(multi_face_policy_help(self._multi_face_policy()), language)
+        )
diff --git a/python-package/insightface/gui/pages/video_search_page.py b/python-package/insightface/gui/pages/video_search_page.py
new file mode 100644
index 000000000..7aa9cc66c
--- /dev/null
+++ b/python-package/insightface/gui/pages/video_search_page.py
@@ -0,0 +1,109 @@
+"""Video person search page."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from PySide6.QtWidgets import QComboBox, QLabel, QSpinBox, QTableWidget, QTableWidgetItem
+
+from ..core.exporters import export_csv
+from ..core.recognition import search_gallery
+from ..core.utils import save_image, timestamp_for_filename
+from ..core.video import iter_video_frames, read_video_thumbnail, timestamp_hhmmss
+from ..widgets.table_utils import configure_table_columns, refresh_table_columns
+from ..widgets.upload_preview import UploadPreview
+from .base import BasePage
+
+
+class VideoSearchPage(BasePage):
+    def __init__(self, context, parent=None):
+        super().__init__(context, "Video Person Search", "Search for a registered person in a local video using simple frame sampling.", parent)
+        self.video_path = ""
+        self.rows = []
+        self.person_combo = QComboBox()
+        self.interval = QSpinBox()
+        self.interval.setRange(1, 300)
+        self.interval.setValue(context.config.video_frame_interval)
+        self.video_input = UploadPreview(
+            "Video File",
+            extensions=[".mp4", ".mov", ".avi", ".mkv", ".webm", ".m4v"],
+            dialog_filter="Videos (*.mp4 *.mov *.avi *.mkv *.webm *.m4v);;All Files (*)",
+        )
+        self.video_input.pathChanged.connect(lambda path: self.load_video(path) if path else self.clear_video())
+        self.content.addWidget(self.video_input)
+        self.content.addWidget(self.row(QLabel("Target person"), self.person_combo, QLabel("Frame interval"), self.interval, self.button("Refresh People", self.refresh), self.button("Start", self.start), self.button("Export CSV", self.export)))
+        self.table = QTableWidget(0, 8)
+        self.table.setHorizontalHeaderLabels(["video_path", "timestamp_ms", "timestamp_hhmmss", "frame_index", "person_name", "similarity", "bbox", "crop_path"])
+        configure_table_columns(self.table, [300, 120, 140, 100, 160, 90, 200, 260])
+        self.content.addWidget(self.table, 1)
+        self.refresh()
+
+    def refresh(self) -> None:
+        self.person_combo.clear()
+        for person in self.context.storage.list_people():
+            self.person_combo.addItem(person["display_name"] or person["name"], person["id"])
+
+    def load_video(self, path: str) -> None:
+        self.video_path = path
+        self.rows = []
+        self.table.setRowCount(0)
+        self.video_input.set_image(read_video_thumbnail(path), path)
+        self.set_status(f"Selected video: {path}")
+
+    def clear_video(self) -> None:
+        self.video_path = ""
+        self.rows = []
+        self.table.setRowCount(0)
+        self.video_input.clear(emit=False)
+
+    def start(self) -> None:
+        if not self.video_path:
+            self.show_error("Select a video first.")
+            return
+        if not self.context.engine.is_loaded():
+            self.show_error("Model is not loaded. Please open Models.")
+            return
+        person_id = self.person_combo.currentData()
+        gallery = [item for item in self.context.storage.load_all_gallery_embeddings() if item.get("person_id") == person_id]
+        if not gallery:
+            self.show_error("Target person has no registered face samples.")
+            return
+
+        def task(progress=None, is_cancelled=None):
+            rows = []
+            for frame_index, timestamp_ms, frame in iter_video_frames(self.video_path, self.interval.value()):
+                if is_cancelled and is_cancelled():
+                    break
+                for face in self.context.engine.detect_faces(frame, source_path=self.video_path):
+                    if face.normed_embedding is None:
+                        continue
+                    matches = search_gallery(face.normed_embedding, gallery, top_k=1, threshold=self.context.config.recognition_threshold)
+                    if matches and matches[0].similarity >= self.context.config.recognition_threshold:
+                        crop_path = str(Path(self.context.config.crop_dir) / f"video_{frame_index}_{timestamp_for_filename()}.png")
+                        if face.crop is not None:
+                            save_image(crop_path, face.crop)
+                        rows.append({"video_path": self.video_path, "timestamp_ms": timestamp_ms, "timestamp_hhmmss": timestamp_hhmmss(timestamp_ms), "frame_index": frame_index, "person_id": person_id, "person_name": matches[0].person_name, "similarity": matches[0].similarity, "bbox": face.bbox, "crop_path": crop_path})
+                if progress:
+                    progress(frame_index, frame_index + 1, f"Processed frame {frame_index}")
+            return rows
+
+        self.run_task("Video person search", task, lambda rows: (setattr(self, "rows", rows), self._populate(), self.set_status(f"Video search complete. {len(rows)} hit(s).")))
+
+    def _populate(self) -> None:
+        columns = ["video_path", "timestamp_ms", "timestamp_hhmmss", "frame_index", "person_name", "similarity", "bbox", "crop_path"]
+        self.table.setRowCount(len(self.rows))
+        for row_index, row in enumerate(self.rows):
+            for col, key in enumerate(columns):
+                value = row.get(key, "")
+                if isinstance(value, float):
+                    value = f"{value:.4f}"
+                self.table.setItem(row_index, col, QTableWidgetItem(str(value)))
+        refresh_table_columns(self.table)
+
+    def export(self) -> None:
+        if not self.rows:
+            self.show_error("No video results to export.")
+            return
+        path = Path(self.context.config.export_dir) / f"video_search_{timestamp_for_filename()}.csv"
+        export_csv(path, self.rows)
+        self.set_status(f"Video results exported to {path}")
diff --git a/python-package/insightface/gui/resources.py b/python-package/insightface/gui/resources.py
new file mode 100644
index 000000000..f015170e4
--- /dev/null
+++ b/python-package/insightface/gui/resources.py
@@ -0,0 +1,104 @@
+"""Application resources and desktop metadata helpers."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+from .core.constants import (
+    APP_DOMAIN,
+    APP_ID,
+    APP_NAME,
+    APP_ORGANIZATION,
+    APP_PROCESS_NAME,
+    APP_VERSION,
+)
+
+ASSET_DIR = Path(__file__).resolve().parent / "assets"
+APP_ICON_SVG = ASSET_DIR / "app_icon.svg"
+APP_ICON_PNG = ASSET_DIR / "app_icon.png"
+APP_ICON_ICO = ASSET_DIR / "app_icon.ico"
+APP_ICON_ICNS = ASSET_DIR / "app_icon.icns"
+
+
+def app_icon_path() -> Path:
+    """Return the best available runtime icon path."""
+
+    for path in (APP_ICON_PNG, APP_ICON_ICO, APP_ICON_ICNS, APP_ICON_SVG):
+        if path.exists():
+            return path
+    return APP_ICON_PNG
+
+
+def app_icon():
+    """Create a QIcon for the studio, returning an empty icon if Qt is unavailable."""
+
+    try:
+        from PySide6.QtGui import QIcon
+    except Exception:
+        return None
+    for path in (APP_ICON_PNG, APP_ICON_ICO, APP_ICON_ICNS, APP_ICON_SVG):
+        if not path.exists():
+            continue
+        icon = QIcon(str(path))
+        if not icon.isNull():
+            return icon
+    return QIcon()
+
+
+def configure_application_metadata(app) -> None:
+    """Apply stable, non-localized desktop metadata to the Qt application.
+
+    Process and app identifiers intentionally remain English brand strings.
+    User-facing UI text is localized separately; OS-level identifiers should stay
+    stable for task managers, crash logs, permissions, packaging, and automation.
+    """
+
+    app.setApplicationName(APP_NAME)
+    if hasattr(app, "setApplicationDisplayName"):
+        app.setApplicationDisplayName(APP_NAME)
+    app.setApplicationVersion(APP_VERSION)
+    app.setOrganizationName(APP_ORGANIZATION)
+    app.setOrganizationDomain(APP_DOMAIN)
+    if hasattr(app, "setDesktopFileName"):
+        app.setDesktopFileName(APP_ID)
+    icon = app_icon()
+    if icon is not None and not icon.isNull():
+        app.setWindowIcon(icon)
+    _set_windows_app_user_model_id()
+    _set_process_title_best_effort(APP_PROCESS_NAME)
+
+
+def _set_windows_app_user_model_id() -> None:
+    if not sys.platform.startswith("win"):
+        return
+    try:
+        import ctypes
+
+        ctypes.windll.shell32.SetCurrentProcessExplicitAppUserModelID(APP_ID)
+    except Exception:
+        pass
+
+
+def _set_process_title_best_effort(name: str) -> None:
+    try:
+        import setproctitle
+
+        setproctitle.setproctitle(name)
+    except Exception:
+        pass
+
+
+__all__ = [
+    "APP_NAME",
+    "APP_VERSION",
+    "APP_ID",
+    "APP_PROCESS_NAME",
+    "APP_ICON_SVG",
+    "APP_ICON_PNG",
+    "APP_ICON_ICO",
+    "APP_ICON_ICNS",
+    "app_icon",
+    "app_icon_path",
+    "configure_application_metadata",
+]
diff --git a/python-package/insightface/gui/widgets/__init__.py b/python-package/insightface/gui/widgets/__init__.py
new file mode 100644
index 000000000..0b731366e
--- /dev/null
+++ b/python-package/insightface/gui/widgets/__init__.py
@@ -0,0 +1 @@
+"""Reusable Qt widgets for InsightFace Evaluation Studio."""
diff --git a/python-package/insightface/gui/widgets/drop_input.py b/python-package/insightface/gui/widgets/drop_input.py
new file mode 100644
index 000000000..6f1a0733e
--- /dev/null
+++ b/python-package/insightface/gui/widgets/drop_input.py
@@ -0,0 +1,245 @@
+"""Reusable drag-and-drop file/folder input."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Iterable, Sequence
+
+from PySide6.QtCore import QEvent, Qt, Signal
+from PySide6.QtGui import QCursor
+from PySide6.QtWidgets import (
+    QFileDialog,
+    QFrame,
+    QHBoxLayout,
+    QLabel,
+    QPushButton,
+    QVBoxLayout,
+)
+
+from ..core.tooltips import set_button_tooltip
+
+
+class DropInput(QFrame):
+    pathsChanged = Signal(list)
+    removed = Signal()
+
+    def __init__(
+        self,
+        title: str,
+        mode: str = "file",
+        extensions: Iterable[str] | None = None,
+        dialog_filter: str = "All Files (*)",
+        show_select_button: bool = True,
+        parent=None,
+    ):
+        super().__init__(parent)
+        self.title = title
+        self.mode = mode
+        self.show_select_button = bool(show_select_button)
+        self.extensions = {ext.lower() if ext.startswith(".") else f".{ext.lower()}" for ext in (extensions or [])}
+        self.dialog_filter = dialog_filter
+        self._paths: list[str] = []
+        self.setObjectName("dropInput")
+        self.setAcceptDrops(True)
+        self.setMouseTracking(True)
+        self.setProperty("hoverActive", False)
+        self.setProperty("dragActive", False)
+        self.setProperty("hasFiles", False)
+        self.setFrameShape(QFrame.StyledPanel)
+        layout = QVBoxLayout(self)
+        self.title_label = QLabel(title)
+        self.title_label.setAlignment(Qt.AlignCenter)
+        self.title_label.setStyleSheet("font-weight: 600;")
+        self.path_label = QLabel(self._empty_text())
+        self.path_label.setObjectName("dropPrompt")
+        self.path_label.setAlignment(Qt.AlignCenter)
+        self.path_label.setWordWrap(True)
+        self.path_label.setMinimumHeight(72)
+        self.path_label.setTextInteractionFlags(Qt.TextSelectableByMouse)
+        buttons = QHBoxLayout()
+        self.select_button = QPushButton("Select")
+        self.remove_button = QPushButton("Remove")
+        self.remove_button.setEnabled(False)
+        self.remove_button.setVisible(False)
+        self.select_button.setVisible(self.show_select_button)
+        self.select_button.clicked.connect(self.browse)
+        self.remove_button.clicked.connect(self.clear)
+        set_button_tooltip(self.select_button)
+        set_button_tooltip(self.remove_button)
+        buttons.addStretch(1)
+        buttons.addWidget(self.select_button)
+        buttons.addWidget(self.remove_button)
+        buttons.addStretch(1)
+        layout.addWidget(self.title_label)
+        layout.addWidget(self.path_label)
+        layout.addLayout(buttons)
+        for watched in (self, self.title_label, self.path_label, self.select_button, self.remove_button):
+            watched.setAcceptDrops(True)
+            watched.setMouseTracking(True)
+            watched.installEventFilter(self)
+
+    def paths(self) -> list[str]:
+        return list(self._paths)
+
+    def path(self) -> str:
+        return self._paths[0] if self._paths else ""
+
+    def set_path(self, path: str, emit: bool = True) -> None:
+        self.set_paths([path] if path else [], emit=emit)
+
+    def set_paths(self, paths: Sequence[str], emit: bool = True) -> None:
+        accepted = [str(Path(path).expanduser()) for path in paths if self._accepts_path(path)]
+        if self.mode != "files":
+            accepted = accepted[:1]
+        self._paths = accepted
+        if accepted:
+            display_text = self._selected_text(accepted)
+            self.path_label.setText(display_text)
+            self.path_label.setToolTip("\n".join(accepted))
+            self.select_button.setVisible(False)
+            self.remove_button.setEnabled(True)
+            self.remove_button.setVisible(True)
+        else:
+            self.path_label.setText(self._empty_text())
+            self.path_label.setToolTip("")
+            self.select_button.setVisible(self.show_select_button)
+            self.remove_button.setEnabled(False)
+            self.remove_button.setVisible(False)
+        if emit:
+            self.pathsChanged.emit(self.paths())
+        self._set_property("hasFiles", bool(accepted))
+
+    def clear(self) -> None:
+        self._paths = []
+        self.path_label.setText(self._empty_text())
+        self.path_label.setToolTip("")
+        self.select_button.setVisible(self.show_select_button)
+        self.remove_button.setEnabled(False)
+        self.remove_button.setVisible(False)
+        self._set_property("hasFiles", False)
+        self.removed.emit()
+        self.pathsChanged.emit([])
+
+    def browse(self) -> None:
+        if self.mode == "folder":
+            path = QFileDialog.getExistingDirectory(self, f"Select {self.title}", str(Path.home()))
+            if path:
+                self.set_path(path)
+        elif self.mode == "files":
+            paths, _ = QFileDialog.getOpenFileNames(self, f"Select {self.title}", str(Path.home()), self.dialog_filter)
+            if paths:
+                self.set_paths(paths)
+        else:
+            path, _ = QFileDialog.getOpenFileName(self, f"Select {self.title}", str(Path.home()), self.dialog_filter)
+            if path:
+                self.set_path(path)
+
+    def dragEnterEvent(self, event) -> None:  # noqa: N802
+        if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+            self._set_property("dragActive", True)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def dragMoveEvent(self, event) -> None:  # noqa: N802
+        if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+            self._set_property("dragActive", True)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def dragLeaveEvent(self, event) -> None:  # noqa: N802
+        self._set_property("dragActive", False)
+        super().dragLeaveEvent(event)
+
+    def dropEvent(self, event) -> None:  # noqa: N802
+        self._set_property("dragActive", False)
+        paths = [url.toLocalFile() for url in event.mimeData().urls() if url.isLocalFile()]
+        accepted = [path for path in paths if self._accepts_path(path)]
+        if accepted:
+            self.set_paths(accepted)
+            event.acceptProposedAction()
+        else:
+            event.ignore()
+
+    def _accepted_urls(self, urls) -> bool:
+        return any(url.isLocalFile() and self._accepts_path(url.toLocalFile()) for url in urls)
+
+    def _accepts_path(self, path: str) -> bool:
+        if not path:
+            return False
+        p = Path(path)
+        if self.mode == "folder":
+            return p.is_dir()
+        if p.is_dir():
+            return self.mode == "folder"
+        if self.extensions:
+            return p.suffix.lower() in self.extensions
+        return p.exists()
+
+    def _empty_text(self) -> str:
+        if self.mode == "folder":
+            return "Click to upload or drag a folder here."
+        if self.mode == "files":
+            return "Click to upload or drag files here."
+        return "Click to upload or drag a file here."
+
+    def _selected_text(self, paths: Sequence[str]) -> str:
+        prefix = "Selected folder:" if self.mode == "folder" else "Selected files:" if self.mode == "files" else "Selected file:"
+        body = "\n".join(paths[:8])
+        if len(paths) > 8:
+            body = f"{body}\n... {len(paths)} total"
+        return f"{prefix}\n{body}"
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() in (QEvent.DragEnter, QEvent.DragMove):
+            if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+                self._set_property("dragActive", True)
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.DragLeave:
+            self._set_property("dragActive", False)
+            return False
+        if event.type() == QEvent.Drop:
+            self._set_property("dragActive", False)
+            paths = [url.toLocalFile() for url in event.mimeData().urls() if url.isLocalFile()]
+            accepted = [path for path in paths if self._accepts_path(path)]
+            if accepted:
+                self.set_paths(accepted)
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.Enter:
+            self._set_property("hoverActive", True)
+            return False
+        if event.type() == QEvent.Leave:
+            self._update_hover_from_cursor()
+            return False
+        if (
+            event.type() == QEvent.MouseButtonRelease
+            and event.button() == Qt.LeftButton
+            and watched in {self, self.title_label, self.path_label}
+        ):
+            self.browse()
+            return True
+        return super().eventFilter(watched, event)
+
+    def enterEvent(self, event) -> None:  # noqa: N802
+        self._set_property("hoverActive", True)
+        super().enterEvent(event)
+
+    def leaveEvent(self, event) -> None:  # noqa: N802
+        self._update_hover_from_cursor()
+        super().leaveEvent(event)
+
+    def _set_property(self, name: str, value) -> None:
+        self.setProperty(name, value)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    def _update_hover_from_cursor(self) -> None:
+        inside = self.rect().contains(self.mapFromGlobal(QCursor.pos()))
+        self._set_property("hoverActive", inside)
diff --git a/python-package/insightface/gui/widgets/face_overlay.py b/python-package/insightface/gui/widgets/face_overlay.py
new file mode 100644
index 000000000..85f33385f
--- /dev/null
+++ b/python-package/insightface/gui/widgets/face_overlay.py
@@ -0,0 +1,9 @@
+"""Overlay formatting helpers."""
+
+from __future__ import annotations
+
+
+def face_label(name: str = "Unknown", similarity: float | None = None) -> str:
+    if similarity is None:
+        return name or "Unknown"
+    return f"{name or 'Unknown'} {similarity:.2f}"
diff --git a/python-package/insightface/gui/widgets/face_table.py b/python-package/insightface/gui/widgets/face_table.py
new file mode 100644
index 000000000..23a32a5b3
--- /dev/null
+++ b/python-package/insightface/gui/widgets/face_table.py
@@ -0,0 +1,22 @@
+"""Table helpers for face/search results."""
+
+from __future__ import annotations
+
+from typing import Iterable, Mapping
+
+from PySide6.QtWidgets import QTableWidget, QTableWidgetItem
+
+from .table_utils import configure_table_columns, refresh_table_columns
+
+
+class FaceTable(QTableWidget):
+    def set_rows(self, rows: Iterable[Mapping[str, object]], columns: list[str]) -> None:
+        rows = list(rows)
+        self.setColumnCount(len(columns))
+        self.setHorizontalHeaderLabels(columns)
+        configure_table_columns(self, [140] * len(columns))
+        self.setRowCount(len(rows))
+        for row_index, row in enumerate(rows):
+            for col_index, column in enumerate(columns):
+                self.setItem(row_index, col_index, QTableWidgetItem(str(row.get(column, ""))))
+        refresh_table_columns(self)
diff --git a/python-package/insightface/gui/widgets/image_viewer.py b/python-package/insightface/gui/widgets/image_viewer.py
new file mode 100644
index 000000000..db0ba1ff1
--- /dev/null
+++ b/python-package/insightface/gui/widgets/image_viewer.py
@@ -0,0 +1,117 @@
+"""Zoomable image viewer with face overlays."""
+
+from __future__ import annotations
+
+from typing import Any, Iterable, List, Optional
+
+import numpy as np
+from PySide6.QtCore import QPointF, QRectF, Qt, Signal
+from PySide6.QtGui import QColor, QImage, QPainter, QPen, QPixmap
+from PySide6.QtWidgets import QGraphicsPixmapItem, QGraphicsScene, QGraphicsView
+
+
+def numpy_to_qimage(image: np.ndarray) -> QImage:
+    if image.ndim == 2:
+        arr = np.ascontiguousarray(image)
+        return QImage(arr.data, arr.shape[1], arr.shape[0], arr.strides[0], QImage.Format_Grayscale8).copy()
+    rgb = np.ascontiguousarray(image[..., ::-1])
+    return QImage(rgb.data, rgb.shape[1], rgb.shape[0], rgb.strides[0], QImage.Format_RGB888).copy()
+
+
+class ImageViewer(QGraphicsView):
+    faceClicked = Signal(int)
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self.setObjectName("imageViewer")
+        self.viewport().setObjectName("imageViewerViewport")
+        self.setRenderHints(QPainter.Antialiasing | QPainter.SmoothPixmapTransform)
+        self.setDragMode(QGraphicsView.ScrollHandDrag)
+        self.setTransformationAnchor(QGraphicsView.AnchorUnderMouse)
+        self.scene = QGraphicsScene(self)
+        self.setScene(self.scene)
+        self.pixmap_item: Optional[QGraphicsPixmapItem] = None
+        self.image: Optional[np.ndarray] = None
+        self.faces: List[dict[str, Any]] = []
+
+    def set_image(self, image: Optional[np.ndarray]) -> None:
+        self.scene.clear()
+        self.image = image
+        self.faces = []
+        self.pixmap_item = None
+        if image is None:
+            return
+        pixmap = QPixmap.fromImage(numpy_to_qimage(image))
+        self.pixmap_item = self.scene.addPixmap(pixmap)
+        self.scene.setSceneRect(QRectF(pixmap.rect()))
+        self.draw_overlays()
+        self.fit_to_window()
+
+    def set_faces(self, faces: Iterable[Any]) -> None:
+        self.faces = []
+        for face in faces:
+            if hasattr(face, "bbox"):
+                label = getattr(face, "label", None) or getattr(face, "person_name", None) or "Face"
+                similarity = getattr(face, "similarity", None)
+                self.faces.append({"bbox": face.bbox, "label": label, "similarity": similarity})
+            else:
+                self.faces.append(dict(face))
+        self.draw_overlays()
+
+    def draw_overlays(self) -> None:
+        if self.pixmap_item is None or self.image is None:
+            return
+        for item in list(self.scene.items()):
+            if item is not self.pixmap_item:
+                self.scene.removeItem(item)
+        for index, face in enumerate(self.faces):
+            x1, y1, x2, y2 = [float(value) for value in face.get("bbox", [0, 0, 0, 0])]
+            rect = QRectF(x1, y1, x2 - x1, y2 - y1)
+            pen = QPen(QColor("#10b981") if face.get("label") != "Unknown" else QColor("#f59e0b"))
+            pen.setWidth(3)
+            item = self.scene.addRect(rect, pen)
+            item.setData(0, index)
+            label = face.get("label", "Face")
+            similarity = face.get("similarity")
+            if similarity is not None:
+                label = f"{label} {float(similarity):.2f}"
+            text = self.scene.addText(label)
+            text.setDefaultTextColor(QColor("#111827"))
+            text.setPos(QPointF(x1, max(0, y1 - 24)))
+            text.setData(0, index)
+
+    def fit_to_window(self) -> None:
+        if self.pixmap_item is not None:
+            self.fitInView(self.pixmap_item, Qt.KeepAspectRatio)
+
+    def wheelEvent(self, event) -> None:  # noqa: N802
+        factor = 1.2 if event.angleDelta().y() > 0 else 1 / 1.2
+        self.scale(factor, factor)
+
+    def mousePressEvent(self, event) -> None:  # noqa: N802
+        scene_pos = self.mapToScene(event.pos())
+        for index, face in enumerate(self.faces):
+            x1, y1, x2, y2 = [float(value) for value in face.get("bbox", [0, 0, 0, 0])]
+            if x1 <= scene_pos.x() <= x2 and y1 <= scene_pos.y() <= y2:
+                self.faceClicked.emit(index)
+                break
+        super().mousePressEvent(event)
+
+    def render_with_overlay(self) -> Optional[np.ndarray]:
+        if self.image is None:
+            return None
+        image = self.image.copy()
+        try:
+            import cv2
+
+            for face in self.faces:
+                x1, y1, x2, y2 = [int(round(float(value))) for value in face.get("bbox", [0, 0, 0, 0])]
+                color = (16, 185, 129) if face.get("label") != "Unknown" else (245, 158, 11)
+                cv2.rectangle(image, (x1, y1), (x2, y2), color, 2)
+                label = face.get("label", "Face")
+                if face.get("similarity") is not None:
+                    label = f"{label} {float(face['similarity']):.2f}"
+                cv2.putText(image, label, (x1, max(20, y1 - 8)), cv2.FONT_HERSHEY_SIMPLEX, 0.65, color, 2)
+        except Exception:
+            pass
+        return image
diff --git a/python-package/insightface/gui/widgets/metric_card.py b/python-package/insightface/gui/widgets/metric_card.py
new file mode 100644
index 000000000..f80a9b9f0
--- /dev/null
+++ b/python-package/insightface/gui/widgets/metric_card.py
@@ -0,0 +1,20 @@
+"""Small metric display."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QFrame, QLabel, QVBoxLayout
+
+
+class MetricCard(QFrame):
+    def __init__(self, title: str, value: str = "0", parent=None):
+        super().__init__(parent)
+        self.setFrameShape(QFrame.StyledPanel)
+        self.title_label = QLabel(title)
+        self.value_label = QLabel(value)
+        self.value_label.setStyleSheet("font-size: 24px; font-weight: 600;")
+        layout = QVBoxLayout(self)
+        layout.addWidget(self.title_label)
+        layout.addWidget(self.value_label)
+
+    def set_value(self, value: object) -> None:
+        self.value_label.setText(str(value))
diff --git a/python-package/insightface/gui/widgets/person_card.py b/python-package/insightface/gui/widgets/person_card.py
new file mode 100644
index 000000000..6a42d5ebd
--- /dev/null
+++ b/python-package/insightface/gui/widgets/person_card.py
@@ -0,0 +1,22 @@
+"""Compact person/cluster card."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QFrame, QLabel, QPushButton, QVBoxLayout
+
+from ..core.tooltips import set_button_tooltip
+
+
+class PersonCard(QFrame):
+    def __init__(self, title: str, subtitle: str = "", parent=None):
+        super().__init__(parent)
+        self.setFrameShape(QFrame.StyledPanel)
+        self.title = QLabel(title)
+        self.title.setStyleSheet("font-weight: 600;")
+        self.subtitle = QLabel(subtitle)
+        self.action = QPushButton("Open")
+        set_button_tooltip(self.action)
+        layout = QVBoxLayout(self)
+        layout.addWidget(self.title)
+        layout.addWidget(self.subtitle)
+        layout.addWidget(self.action)
diff --git a/python-package/insightface/gui/widgets/progress_dialog.py b/python-package/insightface/gui/widgets/progress_dialog.py
new file mode 100644
index 000000000..35dedd42b
--- /dev/null
+++ b/python-package/insightface/gui/widgets/progress_dialog.py
@@ -0,0 +1,20 @@
+"""Cancelable progress dialog."""
+
+from __future__ import annotations
+
+from PySide6.QtWidgets import QProgressDialog
+from PySide6.QtCore import Qt
+
+
+class StudioProgressDialog(QProgressDialog):
+    def __init__(self, title: str, parent=None):
+        super().__init__("", "Cancel", 0, 100, parent)
+        self.setWindowTitle(title)
+        self.setWindowModality(Qt.WindowModal)
+        self.setMinimumDuration(0)
+
+    def update_progress(self, current: int, total: int, message: str = "") -> None:
+        total = max(1, int(total))
+        self.setMaximum(total)
+        self.setValue(min(int(current), total))
+        self.setLabelText(message)
diff --git a/python-package/insightface/gui/widgets/table_utils.py b/python-package/insightface/gui/widgets/table_utils.py
new file mode 100644
index 000000000..7b4a90ab0
--- /dev/null
+++ b/python-package/insightface/gui/widgets/table_utils.py
@@ -0,0 +1,74 @@
+"""Table layout helpers for dense desktop review pages."""
+
+from __future__ import annotations
+
+from PySide6.QtCore import QEvent, QObject, QTimer
+from PySide6.QtWidgets import QAbstractItemView, QAbstractScrollArea, QHeaderView, QTableWidget
+
+
+class _ProportionalTableSizer(QObject):
+    def __init__(self, table: QTableWidget, estimated_widths: list[int], minimum_width: int = 56):
+        super().__init__(table)
+        self.table = table
+        self.estimated_widths = [max(1, int(width)) for width in estimated_widths]
+        self.minimum_width = max(1, int(minimum_width))
+
+    def update(self, estimated_widths: list[int], minimum_width: int) -> None:
+        self.estimated_widths = [max(1, int(width)) for width in estimated_widths]
+        self.minimum_width = max(1, int(minimum_width))
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() == QEvent.Resize:
+            QTimer.singleShot(0, self.apply)
+        return False
+
+    def apply(self) -> None:
+        columns = self.table.columnCount()
+        if columns <= 0:
+            return
+        estimates = self.estimated_widths[:columns]
+        if len(estimates) < columns:
+            estimates.extend([self.minimum_width] * (columns - len(estimates)))
+        available = max(1, self.table.viewport().width())
+        total = max(1, sum(estimates))
+        used = 0
+        for column in range(columns):
+            if column == columns - 1:
+                width = max(self.minimum_width, available - used)
+            else:
+                width = max(self.minimum_width, int(available * estimates[column] / total))
+                used += width
+            self.table.setColumnWidth(column, width)
+
+
+def configure_table_columns(table: QTableWidget, estimated_widths: list[int], minimum_width: int = 56) -> None:
+    """Make table columns fill available width using estimated relative widths."""
+
+    existing = getattr(table, "_proportional_table_sizer", None)
+    if existing is not None:
+        existing.update(estimated_widths, minimum_width)
+        existing.apply()
+        return
+
+    table.setAlternatingRowColors(True)
+    table.setHorizontalScrollMode(QAbstractItemView.ScrollPerPixel)
+    table.setSizeAdjustPolicy(QAbstractScrollArea.AdjustIgnored)
+    header = table.horizontalHeader()
+    header.setStretchLastSection(False)
+    header.setMinimumSectionSize(minimum_width)
+    header.setSectionResizeMode(QHeaderView.Interactive)
+    sizer = _ProportionalTableSizer(table, estimated_widths, minimum_width)
+    table.installEventFilter(sizer)
+    table.viewport().installEventFilter(sizer)
+    table._proportional_table_sizer = sizer  # type: ignore[attr-defined]
+    QTimer.singleShot(0, sizer.apply)
+
+
+def refresh_table_columns(table: QTableWidget) -> None:
+    """Re-apply proportional sizing after table data changes."""
+
+    sizer = getattr(table, "_proportional_table_sizer", None)
+    if sizer is not None:
+        sizer.apply()
+    else:
+        table.resizeColumnsToContents()
diff --git a/python-package/insightface/gui/widgets/threshold_slider.py b/python-package/insightface/gui/widgets/threshold_slider.py
new file mode 100644
index 000000000..c33b051af
--- /dev/null
+++ b/python-package/insightface/gui/widgets/threshold_slider.py
@@ -0,0 +1,38 @@
+"""Threshold slider widget."""
+
+from __future__ import annotations
+
+from PySide6.QtCore import Signal
+from PySide6.QtWidgets import QHBoxLayout, QLabel, QSlider, QWidget
+from PySide6.QtCore import Qt
+
+from ..core.constants import DEFAULT_THRESHOLD
+
+
+class ThresholdSlider(QWidget):
+    valueChanged = Signal(float)
+
+    def __init__(self, value: float = DEFAULT_THRESHOLD, parent=None):
+        super().__init__(parent)
+        self.label = QLabel()
+        self.slider = QSlider(Qt.Horizontal)
+        self.slider.setRange(0, 100)
+        self.slider.valueChanged.connect(self._emit)
+        layout = QHBoxLayout(self)
+        layout.setContentsMargins(0, 0, 0, 0)
+        layout.addWidget(QLabel("Threshold"))
+        layout.addWidget(self.slider)
+        layout.addWidget(self.label)
+        self.set_value(value)
+
+    def _emit(self, raw: int) -> None:
+        value = raw / 100.0
+        self.label.setText(f"{value:.2f}")
+        self.valueChanged.emit(value)
+
+    def value(self) -> float:
+        return self.slider.value() / 100.0
+
+    def set_value(self, value: float) -> None:
+        self.slider.setValue(int(round(float(value) * 100)))
+        self.label.setText(f"{self.value():.2f}")
diff --git a/python-package/insightface/gui/widgets/upload_preview.py b/python-package/insightface/gui/widgets/upload_preview.py
new file mode 100644
index 000000000..bcc8e7e50
--- /dev/null
+++ b/python-package/insightface/gui/widgets/upload_preview.py
@@ -0,0 +1,236 @@
+"""Clickable drag-and-drop preview input for local images and videos."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Iterable, Optional
+
+import numpy as np
+from PySide6.QtCore import QEvent, Qt, Signal
+from PySide6.QtGui import QCursor
+from PySide6.QtWidgets import (
+    QFileDialog,
+    QFrame,
+    QLabel,
+    QPushButton,
+    QStackedLayout,
+    QVBoxLayout,
+)
+
+from ..core.tooltips import set_button_tooltip
+from ..core.i18n import tr
+from .image_viewer import ImageViewer
+
+
+class UploadPreview(QFrame):
+    """Preview box that doubles as a local file picker and drop target."""
+
+    pathChanged = Signal(str)
+    removed = Signal()
+
+    def __init__(
+        self,
+        title: str,
+        extensions: Iterable[str],
+        dialog_filter: str,
+        prompt: str = "Click to upload or drag a file here",
+        parent=None,
+    ):
+        super().__init__(parent)
+        self.setObjectName("uploadPreview")
+        self.title = title
+        self.prompt = prompt
+        self.extensions = {ext.lower() if ext.startswith(".") else f".{ext.lower()}" for ext in extensions}
+        self.dialog_filter = dialog_filter
+        self._path = ""
+        self.setAcceptDrops(True)
+        self.setMinimumHeight(260)
+        self.setMouseTracking(True)
+        self.setProperty("hoverActive", False)
+        self.setProperty("dragActive", False)
+        self.setProperty("hasFile", False)
+
+        layout = QVBoxLayout(self)
+        layout.setContentsMargins(0, 0, 0, 0)
+        self.stack = QStackedLayout()
+        self.stack.setStackingMode(QStackedLayout.StackAll)
+        self.viewer = ImageViewer()
+        self.viewer.setAcceptDrops(True)
+        self.viewer.setMouseTracking(True)
+        self.viewer.setFrameShape(QFrame.NoFrame)
+        self.viewer.setStyleSheet("background:transparent; border:0;")
+        self.viewer.viewport().setAcceptDrops(True)
+        self.viewer.viewport().setMouseTracking(True)
+        self.viewer.viewport().setStyleSheet("background:transparent;")
+        self.placeholder = QLabel(self._placeholder_text())
+        self.placeholder.setObjectName("uploadPrompt")
+        self.placeholder.setAlignment(Qt.AlignCenter)
+        self.placeholder.setWordWrap(True)
+        self.stack.addWidget(self.viewer)
+        self.stack.addWidget(self.placeholder)
+        layout.addLayout(self.stack, 1)
+
+        self.file_label = QLabel("")
+        self.file_label.setObjectName("pathLabel")
+        self.file_label.setTextInteractionFlags(Qt.TextSelectableByMouse)
+        self.file_label.setAlignment(Qt.AlignCenter)
+        self.file_label.hide()
+        layout.addWidget(self.file_label)
+
+        self.remove_button = QPushButton("×", self)
+        self.remove_button.setObjectName("removeUpload")
+        set_button_tooltip(self.remove_button)
+        self.remove_button.setFixedSize(22, 22)
+        self.remove_button.clicked.connect(self.clear)
+        self.remove_button.hide()
+
+        for watched in (self, self.viewer, self.viewer.viewport(), self.placeholder):
+            watched.installEventFilter(self)
+
+    def path(self) -> str:
+        return self._path
+
+    def set_path(self, path: str, emit: bool = True) -> None:
+        if not self._accepts_path(path):
+            return
+        self._path = str(Path(path).expanduser())
+        self._show_selected()
+        if emit:
+            self.pathChanged.emit(self._path)
+
+    def set_image(self, image: Optional[np.ndarray], path: str | None = None) -> None:
+        if path:
+            self._path = str(Path(path).expanduser())
+        self.viewer.set_image(image)
+        if image is None:
+            if self._path:
+                self._show_selected()
+            else:
+                self.clear(emit=False)
+        else:
+            self._show_selected(has_preview=True)
+
+    def set_faces(self, faces) -> None:
+        self.viewer.set_faces(faces)
+
+    def render_with_overlay(self):
+        return self.viewer.render_with_overlay()
+
+    def clear(self, emit: bool = True) -> None:
+        self._path = ""
+        self.viewer.set_image(None)
+        self.viewer.set_faces([])
+        self.placeholder.setText(self._placeholder_text())
+        self.placeholder.show()
+        self.file_label.clear()
+        self.file_label.hide()
+        self.remove_button.hide()
+        self._set_property("hasFile", False)
+        if emit:
+            self.removed.emit()
+            self.pathChanged.emit("")
+
+    def browse(self) -> None:
+        path, _ = QFileDialog.getOpenFileName(
+            self,
+            f"{tr('Select', self._language())} {tr(self.title, self._language())}",
+            str(Path.home()),
+            self.dialog_filter,
+        )
+        if path:
+            self.set_path(path)
+
+    def eventFilter(self, watched, event) -> bool:  # noqa: N802
+        if event.type() in (QEvent.DragEnter, QEvent.DragMove):
+            if event.mimeData().hasUrls() and self._accepted_urls(event.mimeData().urls()):
+                self._set_property("dragActive", True)
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.DragLeave:
+            self._set_property("dragActive", False)
+            return False
+        if event.type() == QEvent.Enter:
+            self._set_property("hoverActive", True)
+            return False
+        if event.type() == QEvent.Leave:
+            self._update_hover_from_cursor()
+            return False
+        if event.type() == QEvent.Drop:
+            self._set_property("dragActive", False)
+            paths = [url.toLocalFile() for url in event.mimeData().urls() if url.isLocalFile()]
+            accepted = [path for path in paths if self._accepts_path(path)]
+            if accepted:
+                self.set_path(accepted[0])
+                event.acceptProposedAction()
+                return True
+            event.ignore()
+            return True
+        if event.type() == QEvent.MouseButtonRelease and event.button() == Qt.LeftButton and not self._path:
+            self.browse()
+            return True
+        return super().eventFilter(watched, event)
+
+    def mouseReleaseEvent(self, event) -> None:  # noqa: N802
+        if event.button() == Qt.LeftButton and not self._path:
+            self.browse()
+            event.accept()
+            return
+        super().mouseReleaseEvent(event)
+
+    def enterEvent(self, event) -> None:  # noqa: N802
+        self._set_property("hoverActive", True)
+        super().enterEvent(event)
+
+    def leaveEvent(self, event) -> None:  # noqa: N802
+        self._update_hover_from_cursor()
+        super().leaveEvent(event)
+
+    def resizeEvent(self, event) -> None:  # noqa: N802
+        super().resizeEvent(event)
+        self.remove_button.move(max(0, self.width() - 30), 8)
+        self.remove_button.raise_()
+
+    def _show_selected(self, has_preview: bool = False) -> None:
+        self._set_property("hasFile", True)
+        self.remove_button.show()
+        self.remove_button.raise_()
+        if has_preview:
+            self.placeholder.hide()
+        else:
+            name = Path(self._path).name if self._path else self.title
+            self.placeholder.setText(
+                f"{tr(self.title, self._language())}\n{name}\n{tr('Drop another file to replace it', self._language())}"
+            )
+            self.placeholder.show()
+        if self._path:
+            self.file_label.setText(str(self._path))
+            self.file_label.show()
+
+    def _placeholder_text(self) -> str:
+        language = self._language()
+        return f"{tr(self.title, language)}\n{tr(self.prompt, language)}"
+
+    def _language(self) -> str | None:
+        context = getattr(self.window(), "context", None)
+        config = getattr(context, "config", None)
+        return getattr(config, "ui_language", None)
+
+    def _accepted_urls(self, urls) -> bool:
+        return any(url.isLocalFile() and self._accepts_path(url.toLocalFile()) for url in urls)
+
+    def _accepts_path(self, path: str) -> bool:
+        if not path:
+            return False
+        p = Path(path).expanduser()
+        return p.is_file() and (not self.extensions or p.suffix.lower() in self.extensions)
+
+    def _set_property(self, name: str, value) -> None:
+        self.setProperty(name, value)
+        self.style().unpolish(self)
+        self.style().polish(self)
+
+    def _update_hover_from_cursor(self) -> None:
+        inside = self.rect().contains(self.mapFromGlobal(QCursor.pos()))
+        self._set_property("hoverActive", inside)
diff --git a/python-package/insightface/model_zoo/__init__.py b/python-package/insightface/model_zoo/__init__.py
index 128086a25..225623d61 100644
--- a/python-package/insightface/model_zoo/__init__.py
+++ b/python-package/insightface/model_zoo/__init__.py
@@ -1 +1,6 @@
-from .model_zoo import get_model, get_model_list
+from .model_zoo import get_model
+from .arcface_onnx import ArcFaceONNX
+from .retinaface import RetinaFace
+from .scrfd import SCRFD
+from .landmark import Landmark
+from .attribute import Attribute
diff --git a/python-package/insightface/model_zoo/arcface_onnx.py b/python-package/insightface/model_zoo/arcface_onnx.py
new file mode 100644
index 000000000..b537ce2ee
--- /dev/null
+++ b/python-package/insightface/model_zoo/arcface_onnx.py
@@ -0,0 +1,92 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+from __future__ import division
+import numpy as np
+import cv2
+import onnx
+import onnxruntime
+from ..utils import face_align
+
+__all__ = [
+    'ArcFaceONNX',
+]
+
+
+class ArcFaceONNX:
+    def __init__(self, model_file=None, session=None):
+        assert model_file is not None
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'recognition'
+        find_sub = False
+        find_mul = False
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        for nid, node in enumerate(graph.node[:8]):
+            #print(nid, node.name)
+            if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                find_sub = True
+            if node.name.startswith('Mul') or node.name.startswith('_mul'):
+                find_mul = True
+        if find_sub and find_mul:
+            #mxnet arcface model
+            input_mean = 0.0
+            input_std = 1.0
+        else:
+            input_mean = 127.5
+            input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        #print('input mean and std:', self.input_mean, self.input_std)
+        if self.session is None:
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        input_name = input_cfg.name
+        self.input_size = tuple(input_shape[2:4][::-1])
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        output_names = []
+        for out in outputs:
+            output_names.append(out.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        assert len(self.output_names)==1
+        self.output_shape = outputs[0].shape
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+
+    def get(self, img, face):
+        aimg = face_align.norm_crop(img, landmark=face.kps, image_size=self.input_size[0])
+        face.embedding = self.get_feat(aimg).flatten()
+        return face.embedding
+
+    def compute_sim(self, feat1, feat2):
+        from numpy.linalg import norm
+        feat1 = feat1.ravel()
+        feat2 = feat2.ravel()
+        sim = np.dot(feat1, feat2) / (norm(feat1) * norm(feat2))
+        return sim
+
+    def get_feat(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.input_size
+        
+        blob = cv2.dnn.blobFromImages(imgs, 1.0 / self.input_std, input_size,
+                                      (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+    def forward(self, batch_data):
+        blob = (batch_data - self.input_mean) / self.input_std
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+
diff --git a/python-package/insightface/model_zoo/attribute.py b/python-package/insightface/model_zoo/attribute.py
new file mode 100644
index 000000000..40c34de3f
--- /dev/null
+++ b/python-package/insightface/model_zoo/attribute.py
@@ -0,0 +1,94 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-06-19
+# @Function      : 
+
+from __future__ import division
+import numpy as np
+import cv2
+import onnx
+import onnxruntime
+from ..utils import face_align
+
+__all__ = [
+    'Attribute',
+]
+
+
+class Attribute:
+    def __init__(self, model_file=None, session=None):
+        assert model_file is not None
+        self.model_file = model_file
+        self.session = session
+        find_sub = False
+        find_mul = False
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        for nid, node in enumerate(graph.node[:8]):
+            #print(nid, node.name)
+            if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                find_sub = True
+            if node.name.startswith('Mul') or node.name.startswith('_mul'):
+                find_mul = True
+            if nid<3 and node.name=='bn_data':
+                find_sub = True
+                find_mul = True
+        if find_sub and find_mul:
+            #mxnet arcface model
+            input_mean = 0.0
+            input_std = 1.0
+        else:
+            input_mean = 127.5
+            input_std = 128.0
+        self.input_mean = input_mean
+        self.input_std = input_std
+        #print('input mean and std:', model_file, self.input_mean, self.input_std)
+        if self.session is None:
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        input_name = input_cfg.name
+        self.input_size = tuple(input_shape[2:4][::-1])
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        output_names = []
+        for out in outputs:
+            output_names.append(out.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        assert len(self.output_names)==1
+        output_shape = outputs[0].shape
+        #print('init output_shape:', output_shape)
+        if output_shape[1]==3:
+            self.taskname = 'genderage'
+        else:
+            self.taskname = 'attribute_%d'%output_shape[1]
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+
+    def get(self, img, face):
+        bbox = face.bbox
+        w, h = (bbox[2] - bbox[0]), (bbox[3] - bbox[1])
+        center = (bbox[2] + bbox[0]) / 2, (bbox[3] + bbox[1]) / 2
+        rotate = 0
+        _scale = self.input_size[0]  / (max(w, h)*1.5)
+        #print('param:', img.shape, bbox, center, self.input_size, _scale, rotate)
+        aimg, M = face_align.transform(img, center, self.input_size[0], _scale, rotate)
+        input_size = tuple(aimg.shape[0:2][::-1])
+        #assert input_size==self.input_size
+        blob = cv2.dnn.blobFromImage(aimg, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        pred = self.session.run(self.output_names, {self.input_name : blob})[0][0]
+        if self.taskname=='genderage':
+            assert len(pred)==3
+            gender = np.argmax(pred[:2])
+            age = int(np.round(pred[2]*100))
+            face['gender'] = gender
+            face['age'] = age
+            return gender, age
+        else:
+            return pred
+
+
diff --git a/python-package/insightface/model_zoo/face_detection.py b/python-package/insightface/model_zoo/face_detection.py
deleted file mode 100644
index f04d665d0..000000000
--- a/python-package/insightface/model_zoo/face_detection.py
+++ /dev/null
@@ -1,426 +0,0 @@
-from __future__ import division
-import mxnet as mx
-import numpy as np
-import mxnet.ndarray as nd
-import cv2
-
-__all__ = ['FaceDetector',
-           'retinaface_r50_v1',
-           'retinaface_mnet025_v1',
-           'retinaface_mnet025_v2',
-           'get_retinaface']
-
-def _whctrs(anchor):
-    """
-    Return width, height, x center, and y center for an anchor (window).
-    """
-
-    w = anchor[2] - anchor[0] + 1
-    h = anchor[3] - anchor[1] + 1
-    x_ctr = anchor[0] + 0.5 * (w - 1)
-    y_ctr = anchor[1] + 0.5 * (h - 1)
-    return w, h, x_ctr, y_ctr
-
-
-def _mkanchors(ws, hs, x_ctr, y_ctr):
-    """
-    Given a vector of widths (ws) and heights (hs) around a center
-    (x_ctr, y_ctr), output a set of anchors (windows).
-    """
-
-    ws = ws[:, np.newaxis]
-    hs = hs[:, np.newaxis]
-    anchors = np.hstack((x_ctr - 0.5 * (ws - 1),
-                         y_ctr - 0.5 * (hs - 1),
-                         x_ctr + 0.5 * (ws - 1),
-                         y_ctr + 0.5 * (hs - 1)))
-    return anchors
-
-def _ratio_enum(anchor, ratios):
-    """
-    Enumerate a set of anchors for each aspect ratio wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    size = w * h
-    size_ratios = size / ratios
-    ws = np.round(np.sqrt(size_ratios))
-    hs = np.round(ws * ratios)
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
-
-
-def _scale_enum(anchor, scales):
-    """
-    Enumerate a set of anchors for each scale wrt an anchor.
-    """
-
-    w, h, x_ctr, y_ctr = _whctrs(anchor)
-    ws = w * scales
-    hs = h * scales
-    anchors = _mkanchors(ws, hs, x_ctr, y_ctr)
-    return anchors
-
-def anchors_plane(height, width, stride, base_anchors):
-    """
-    Parameters
-    ----------
-    height: height of plane
-    width:  width of plane
-    stride: stride ot the original image
-    anchors_base: (A, 4) a base set of anchors
-    Returns
-    -------
-    all_anchors: (height, width, A, 4) ndarray of anchors spreading over the plane
-    """
-    A = base_anchors.shape[0]
-    all_anchors = np.zeros((height, width, A, 4), dtype=np.float32)
-    for iw in range(width):
-        sw = iw * stride
-        for ih in range(height):
-            sh = ih * stride
-            for k in range(A):
-                all_anchors[ih, iw, k, 0] = base_anchors[k, 0] + sw
-                all_anchors[ih, iw, k, 1] = base_anchors[k, 1] + sh
-                all_anchors[ih, iw, k, 2] = base_anchors[k, 2] + sw
-                all_anchors[ih, iw, k, 3] = base_anchors[k, 3] + sh
-    return all_anchors
-
-def generate_anchors(base_size=16, ratios=[0.5, 1, 2],
-                     scales=2 ** np.arange(3, 6), stride=16):
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-
-    base_anchor = np.array([1, 1, base_size, base_size]) - 1
-    ratio_anchors = _ratio_enum(base_anchor, ratios)
-    anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales)
-                         for i in range(ratio_anchors.shape[0])])
-    return anchors
-
-def generate_anchors_fpn(cfg):
-    """
-    Generate anchor (reference) windows by enumerating aspect ratios X
-    scales wrt a reference (0, 0, 15, 15) window.
-    """
-    RPN_FEAT_STRIDE = []
-    for k in cfg:
-      RPN_FEAT_STRIDE.append( int(k) )
-    RPN_FEAT_STRIDE = sorted(RPN_FEAT_STRIDE, reverse=True)
-    anchors = []
-    for k in RPN_FEAT_STRIDE:
-      v = cfg[str(k)]
-      bs = v['BASE_SIZE']
-      __ratios = np.array(v['RATIOS'])
-      __scales = np.array(v['SCALES'])
-      stride = int(k)
-      #print('anchors_fpn', bs, __ratios, __scales, file=sys.stderr)
-      r = generate_anchors(bs, __ratios, __scales, stride)
-      #print('anchors_fpn', r.shape, file=sys.stderr)
-      anchors.append(r)
-
-    return anchors
-
-def clip_pad(tensor, pad_shape):
-    """
-    Clip boxes of the pad area.
-    :param tensor: [n, c, H, W]
-    :param pad_shape: [h, w]
-    :return: [n, c, h, w]
-    """
-    H, W = tensor.shape[2:]
-    h, w = pad_shape
-
-    if h < H or w < W:
-      tensor = tensor[:, :, :h, :w].copy()
-
-    return tensor
-
-def bbox_pred(boxes, box_deltas):
-    """
-    Transform the set of class-agnostic boxes into class-specific boxes
-    by applying the predicted offsets (box_deltas)
-    :param boxes: !important [N 4]
-    :param box_deltas: [N, 4 * num_classes]
-    :return: [N 4 * num_classes]
-    """
-    if boxes.shape[0] == 0:
-        return np.zeros((0, box_deltas.shape[1]))
-
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-
-    dx = box_deltas[:, 0:1]
-    dy = box_deltas[:, 1:2]
-    dw = box_deltas[:, 2:3]
-    dh = box_deltas[:, 3:4]
-
-    pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
-    pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
-    pred_w = np.exp(dw) * widths[:, np.newaxis]
-    pred_h = np.exp(dh) * heights[:, np.newaxis]
-
-    pred_boxes = np.zeros(box_deltas.shape)
-    # x1
-    pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
-    # y1
-    pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
-    # x2
-    pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
-    # y2
-    pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
-
-    if box_deltas.shape[1]>4:
-      pred_boxes[:,4:] = box_deltas[:,4:]
-
-    return pred_boxes
-
-def landmark_pred(boxes, landmark_deltas):
-    if boxes.shape[0] == 0:
-        return np.zeros((0, landmark_deltas.shape[1]))
-    boxes = boxes.astype(np.float, copy=False)
-    widths = boxes[:, 2] - boxes[:, 0] + 1.0
-    heights = boxes[:, 3] - boxes[:, 1] + 1.0
-    ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
-    ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
-    pred = landmark_deltas.copy()
-    for i in range(5):
-        pred[:,i,0] = landmark_deltas[:,i,0]*widths + ctr_x
-        pred[:,i,1] = landmark_deltas[:,i,1]*heights + ctr_y
-    return pred
-
-class FaceDetector:
-    def __init__(self, param_file, rac):
-        self.param_file = param_file
-        self.rac = rac
-        self.default_image_size = (480, 640)
-
-    def prepare(self, ctx_id, nms=0.4, fix_image_size=None):
-        pos = self.param_file.rfind('-')
-        prefix = self.param_file[0:pos]
-        pos2 = self.param_file.rfind('.')
-        epoch = int(self.param_file[pos+1:pos2])
-        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-        if ctx_id>=0:
-            ctx = mx.gpu(ctx_id)
-        else:
-            ctx = mx.cpu()
-        model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-        if fix_image_size is not None:
-            data_shape = (1,3)+fix_image_size
-        else:
-            data_shape = (1,3)+self.default_image_size
-        model.bind(data_shapes=[('data', data_shape)])
-        model.set_params(arg_params, aux_params)
-        #warmup
-        data = mx.nd.zeros(shape=data_shape)
-        db = mx.io.DataBatch(data=(data,))
-        model.forward(db, is_train=False)
-        out = model.get_outputs()[0].asnumpy()
-        self.model = model
-        self.nms_threshold = nms
-
-        self.landmark_std = 1.0
-        _ratio = (1.,)
-        fmc = 3
-        if self.rac=='net3':
-            _ratio = (1.,)
-        elif self.rac=='net3l':
-            _ratio = (1.,)
-            self.landmark_std = 0.2
-        elif network=='net5': #retinaface
-            fmc = 5
-        else:
-            assert False, 'rac setting error %s'%self.rac
-
-        if fmc==3:
-            self._feat_stride_fpn = [32, 16, 8]
-            self.anchor_cfg = {
-                  '32': {'SCALES': (32,16), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-                  '16': {'SCALES': (8,4), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-                  '8': {'SCALES': (2,1), 'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999},
-                  }
-        elif fmc==5:
-            self._feat_stride_fpn = [64, 32, 16, 8, 4]
-            self.anchor_cfg = {}
-            _ass = 2.0**(1.0/3)
-            _basescale = 1.0
-            for _stride in [4, 8, 16, 32, 64]:
-                key = str(_stride)
-                value = {'BASE_SIZE': 16, 'RATIOS': _ratio, 'ALLOWED_BORDER': 9999}
-                scales = []
-                for _ in range(3):
-                    scales.append(_basescale)
-                    _basescale *= _ass
-                value['SCALES'] = tuple(scales)
-                self.anchor_cfg[key] = value
-
-        print(self._feat_stride_fpn, self.anchor_cfg)
-        self.use_landmarks = False
-        if len(sym)//len(self._feat_stride_fpn)==3:
-            self.use_landmarks = True
-        print('use_landmarks', self.use_landmarks)
-        self.fpn_keys = []
-
-        for s in self._feat_stride_fpn:
-            self.fpn_keys.append('stride%s'%s)
-
-        self._anchors_fpn = dict(zip(self.fpn_keys, generate_anchors_fpn(cfg=self.anchor_cfg)))
-        for k in self._anchors_fpn:
-            v = self._anchors_fpn[k].astype(np.float32)
-            self._anchors_fpn[k] = v
-        self.anchor_plane_cache = {}
-
-        self._num_anchors = dict(zip(self.fpn_keys, [anchors.shape[0] for anchors in self._anchors_fpn.values()]))
-
-    def detect(self, img, threshold=0.5, scale=1.0):
-        proposals_list = []
-        scores_list = []
-        landmarks_list = []
-        if scale==1.0:
-            im = img
-        else:
-            im = cv2.resize(img, None, None, fx=scale, fy=scale, interpolation=cv2.INTER_LINEAR)
-        im_info = [im.shape[0], im.shape[1]]
-        im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
-        for i in range(3):
-            im_tensor[0, i, :, :] = im[:, :, 2 - i]
-        data = nd.array(im_tensor)
-        db = mx.io.DataBatch(data=(data,), provide_data=[('data', data.shape)])
-        self.model.forward(db, is_train=False)
-        net_out = self.model.get_outputs()
-        for _idx,s in enumerate(self._feat_stride_fpn):
-            _key = 'stride%s'%s
-            stride = int(s)
-            if self.use_landmarks:
-              idx = _idx*3
-            else:
-              idx = _idx*2
-            scores = net_out[idx].asnumpy()
-            scores = scores[:, self._num_anchors['stride%s'%s]:, :, :]
-            idx+=1
-            bbox_deltas = net_out[idx].asnumpy()
-
-            height, width = bbox_deltas.shape[2], bbox_deltas.shape[3]
-            A = self._num_anchors['stride%s'%s]
-            K = height * width
-            key = (height, width, stride)
-            if key in self.anchor_plane_cache:
-                anchors = self.anchor_plane_cache[key]
-            else:
-                anchors_fpn = self._anchors_fpn['stride%s'%s]
-                anchors = anchors_plane(height, width, stride, anchors_fpn)
-                anchors = anchors.reshape((K * A, 4))
-                if len(self.anchor_plane_cache)<100:
-                    self.anchor_plane_cache[key] = anchors
-
-            scores = clip_pad(scores, (height, width))
-            scores = scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
-
-            bbox_deltas = clip_pad(bbox_deltas, (height, width))
-            bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
-            bbox_pred_len = bbox_deltas.shape[3]//A
-            bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
-
-            proposals = bbox_pred(anchors, bbox_deltas)
-            #proposals = clip_boxes(proposals, im_info[:2])
-
-
-            scores_ravel = scores.ravel()
-            order = np.where(scores_ravel>=threshold)[0]
-            proposals = proposals[order, :]
-            scores = scores[order]
-
-            proposals[:,0:4] /= scale
-
-            proposals_list.append(proposals)
-            scores_list.append(scores)
-
-            if self.use_landmarks:
-                idx+=1
-                landmark_deltas = net_out[idx].asnumpy()
-                landmark_deltas = clip_pad(landmark_deltas, (height, width))
-                landmark_pred_len = landmark_deltas.shape[1]//A
-                landmark_deltas = landmark_deltas.transpose((0, 2, 3, 1)).reshape((-1, 5, landmark_pred_len//5))
-                landmark_deltas *= self.landmark_std
-                #print(landmark_deltas.shape, landmark_deltas)
-                landmarks = landmark_pred(anchors, landmark_deltas)
-                landmarks = landmarks[order, :]
-
-                landmarks[:,:,0:2] /= scale
-                landmarks_list.append(landmarks)
-
-        proposals = np.vstack(proposals_list)
-        landmarks = None
-        if proposals.shape[0]==0:
-            if self.use_landmarks:
-                landmarks = np.zeros( (0,5,2) )
-            return np.zeros( (0,5) ), landmarks
-        scores = np.vstack(scores_list)
-        scores_ravel = scores.ravel()
-        order = scores_ravel.argsort()[::-1]
-        proposals = proposals[order, :]
-        scores = scores[order]
-        if self.use_landmarks:
-            landmarks = np.vstack(landmarks_list)
-            landmarks = landmarks[order].astype(np.float32, copy=False)
-
-        pre_det = np.hstack((proposals[:,0:4], scores)).astype(np.float32, copy=False)
-        keep = self.nms(pre_det)
-        det = np.hstack( (pre_det, proposals[:,4:]) )
-        det = det[keep, :]
-        if self.use_landmarks:
-            landmarks = landmarks[keep]
-
-        return det, landmarks
-
-    def nms(self, dets):
-        thresh = self.nms_threshold
-        x1 = dets[:, 0]
-        y1 = dets[:, 1]
-        x2 = dets[:, 2]
-        y2 = dets[:, 3]
-        scores = dets[:, 4]
-
-        areas = (x2 - x1 + 1) * (y2 - y1 + 1)
-        order = scores.argsort()[::-1]
-
-        keep = []
-        while order.size > 0:
-            i = order[0]
-            keep.append(i)
-            xx1 = np.maximum(x1[i], x1[order[1:]])
-            yy1 = np.maximum(y1[i], y1[order[1:]])
-            xx2 = np.minimum(x2[i], x2[order[1:]])
-            yy2 = np.minimum(y2[i], y2[order[1:]])
-
-            w = np.maximum(0.0, xx2 - xx1 + 1)
-            h = np.maximum(0.0, yy2 - yy1 + 1)
-            inter = w * h
-            ovr = inter / (areas[i] + areas[order[1:]] - inter)
-
-            inds = np.where(ovr <= thresh)[0]
-            order = order[inds + 1]
-
-        return keep
-
-
-def get_retinaface(name, rac='net3',
-               root='~/.insightface/models', **kwargs):
-    from .model_store import get_model_file
-    _file = get_model_file("retinaface_%s"%name, root=root)
-    return FaceDetector(_file, rac)
-
-def retinaface_r50_v1(**kwargs):
-    return get_retinaface("r50_v1", rac='net3', **kwargs)
-
-def retinaface_mnet025_v1(**kwargs):
-    return get_retinaface("mnet025_v1", rac='net3', **kwargs)
-
-def retinaface_mnet025_v2(**kwargs):
-    return get_retinaface("mnet025_v2", rac='net3l', **kwargs)
-
diff --git a/python-package/insightface/model_zoo/face_genderage.py b/python-package/insightface/model_zoo/face_genderage.py
deleted file mode 100644
index 3a0423d0b..000000000
--- a/python-package/insightface/model_zoo/face_genderage.py
+++ /dev/null
@@ -1,77 +0,0 @@
-from __future__ import division
-import mxnet as mx
-import numpy as np
-import cv2
-
-__all__ = ['FaceGenderage',
-           'genderage_v1',
-           'get_genderage']
-
-
-class FaceGenderage:
-    def __init__(self, name, download, param_file):
-        self.name = name
-        self.download = download
-        self.param_file = param_file
-        self.image_size = (112, 112)
-        if download:
-            assert param_file
-
-    def prepare(self, ctx_id):
-        if self.param_file:
-            pos = self.param_file.rfind('-')
-            prefix = self.param_file[0:pos]
-            pos2 = self.param_file.rfind('.')
-            epoch = int(self.param_file[pos+1:pos2])
-            sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-            all_layers = sym.get_internals()
-            sym = all_layers['fc1_output']
-            if ctx_id>=0:
-                ctx = mx.gpu(ctx_id)
-            else:
-                ctx = mx.cpu()
-            model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-            data_shape = (1,3)+self.image_size
-            model.bind(data_shapes=[('data', data_shape)])
-            model.set_params(arg_params, aux_params)
-            #warmup
-            data = mx.nd.zeros(shape=data_shape)
-            db = mx.io.DataBatch(data=(data,))
-            model.forward(db, is_train=False)
-            embedding = model.get_outputs()[0].asnumpy()
-            self.model = model
-        else:
-            pass
-
-    def get(self, img):
-        assert self.param_file and self.model
-        assert img.shape[2]==3 and img.shape[0:2]==self.image_size
-        data = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        data = np.transpose(data, (2,0,1))
-        data = np.expand_dims(data, axis=0)
-        data = mx.nd.array(data)
-        db = mx.io.DataBatch(data=(data,))
-        self.model.forward(db, is_train=False)
-        ret = self.model.get_outputs()[0].asnumpy()
-        g = ret[:,0:2].flatten()
-        gender = np.argmax(g)
-        a = ret[:,2:202].reshape( (100,2) )
-        a = np.argmax(a, axis=1)
-        age = int(sum(a))
-        return gender, age
-
-def get_genderage(name, download=True,
-               root='~/.insightface/models', **kwargs):
-    if not download:
-        return FaceGenderage(name, False, None)
-    else:
-        from .model_store import get_model_file
-        _file = get_model_file("genderage_%s"%name, root=root)
-        return FaceGenderage(name, True, _file)
-
-def genderage_v1(**kwargs):
-    return get_genderage("v1", download=True, **kwargs)
-
-
-
-
diff --git a/python-package/insightface/model_zoo/face_recognition.py b/python-package/insightface/model_zoo/face_recognition.py
deleted file mode 100644
index 506ff1d6d..000000000
--- a/python-package/insightface/model_zoo/face_recognition.py
+++ /dev/null
@@ -1,83 +0,0 @@
-from __future__ import division
-import mxnet as mx
-import numpy as np
-import cv2
-
-__all__ = ['FaceRecognition',
-           'arcface_r100_v1', 'arcface_outofreach_v1', 'arcface_mfn_v1',
-           'get_arcface']
-
-
-class FaceRecognition:
-    def __init__(self, name, download, param_file):
-        self.name = name
-        self.download = download
-        self.param_file = param_file
-        self.image_size = (112, 112)
-        if download:
-            assert param_file
-
-    def prepare(self, ctx_id):
-        if self.param_file:
-            pos = self.param_file.rfind('-')
-            prefix = self.param_file[0:pos]
-            pos2 = self.param_file.rfind('.')
-            epoch = int(self.param_file[pos+1:pos2])
-            sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-            all_layers = sym.get_internals()
-            sym = all_layers['fc1_output']
-            if ctx_id>=0:
-                ctx = mx.gpu(ctx_id)
-            else:
-                ctx = mx.cpu()
-            model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-            data_shape = (1,3)+self.image_size
-            model.bind(data_shapes=[('data', data_shape)])
-            model.set_params(arg_params, aux_params)
-            #warmup
-            data = mx.nd.zeros(shape=data_shape)
-            db = mx.io.DataBatch(data=(data,))
-            model.forward(db, is_train=False)
-            embedding = model.get_outputs()[0].asnumpy()
-            self.model = model
-        else:
-            pass
-
-    def get_embedding(self, img):
-        assert self.param_file and self.model
-        assert img.shape[2]==3 and img.shape[0:2]==self.image_size
-        data = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        data = np.transpose(data, (2,0,1))
-        data = np.expand_dims(data, axis=0)
-        data = mx.nd.array(data)
-        db = mx.io.DataBatch(data=(data,))
-        self.model.forward(db, is_train=False)
-        embedding = self.model.get_outputs()[0].asnumpy()
-        return embedding
-
-    def compute_sim(self, img1, img2):
-        emb1 = self.get_embedding(img1).flatten()
-        emb2 = self.get_embedding(img2).flatten()
-        from numpy.linalg import norm
-        sim = np.dot(emb1, emb2)/(norm(emb1)*norm(emb2))
-        return sim
-
-def get_arcface(name, download=True,
-               root='~/.insightface/models', **kwargs):
-    if not download:
-        return FaceRecognition(name, False, None)
-    else:
-        from .model_store import get_model_file
-        _file = get_model_file("arcface_%s"%name, root=root)
-        return FaceRecognition(name, True, _file)
-
-def arcface_r100_v1(**kwargs):
-    return get_arcface("r100_v1", download=True, **kwargs)
-
-
-def arcface_mfn_v1(**kwargs):
-    return get_arcface("mfn_v1", download=True, **kwargs)
-
-def arcface_outofreach_v1(**kwargs):
-    return get_arcface("outofreach_v1", download=False, **kwargs)
-
diff --git a/python-package/insightface/model_zoo/inswapper.py b/python-package/insightface/model_zoo/inswapper.py
new file mode 100644
index 000000000..dcaceb1f2
--- /dev/null
+++ b/python-package/insightface/model_zoo/inswapper.py
@@ -0,0 +1,105 @@
+import time
+import numpy as np
+import onnxruntime
+import cv2
+import onnx
+from onnx import numpy_helper
+from ..utils import face_align
+
+
+
+
+class INSwapper():
+    def __init__(self, model_file=None, session=None):
+        self.model_file = model_file
+        self.session = session
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        self.emap = numpy_helper.to_array(graph.initializer[-1])
+        self.input_mean = 0.0
+        self.input_std = 255.0
+        #print('input mean and std:', model_file, self.input_mean, self.input_std)
+        if self.session is None:
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        inputs = self.session.get_inputs()
+        self.input_names = []
+        for inp in inputs:
+            self.input_names.append(inp.name)
+        outputs = self.session.get_outputs()
+        output_names = []
+        for out in outputs:
+            output_names.append(out.name)
+        self.output_names = output_names
+        assert len(self.output_names)==1
+        output_shape = outputs[0].shape
+        input_cfg = inputs[0]
+        input_shape = input_cfg.shape
+        self.input_shape = input_shape
+        print('inswapper-shape:', self.input_shape)
+        self.input_size = tuple(input_shape[2:4][::-1])
+
+    def forward(self, img, latent):
+        img = (img - self.input_mean) / self.input_std
+        pred = self.session.run(self.output_names, {self.input_names[0]: img, self.input_names[1]: latent})[0]
+        return pred
+
+    def get(self, img, target_face, source_face, paste_back=True):
+        aimg, M = face_align.norm_crop2(img, target_face.kps, self.input_size[0])
+        blob = cv2.dnn.blobFromImage(aimg, 1.0 / self.input_std, self.input_size,
+                                      (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        latent = source_face.normed_embedding.reshape((1,-1))
+        latent = np.dot(latent, self.emap)
+        latent /= np.linalg.norm(latent)
+        pred = self.session.run(self.output_names, {self.input_names[0]: blob, self.input_names[1]: latent})[0]
+        #print(latent.shape, latent.dtype, pred.shape)
+        img_fake = pred.transpose((0,2,3,1))[0]
+        bgr_fake = np.clip(255 * img_fake, 0, 255).astype(np.uint8)[:,:,::-1]
+        if not paste_back:
+            return bgr_fake, M
+        else:
+            target_img = img
+            fake_diff = bgr_fake.astype(np.float32) - aimg.astype(np.float32)
+            fake_diff = np.abs(fake_diff).mean(axis=2)
+            fake_diff[:2,:] = 0
+            fake_diff[-2:,:] = 0
+            fake_diff[:,:2] = 0
+            fake_diff[:,-2:] = 0
+            IM = cv2.invertAffineTransform(M)
+            img_white = np.full((aimg.shape[0],aimg.shape[1]), 255, dtype=np.float32)
+            bgr_fake = cv2.warpAffine(bgr_fake, IM, (target_img.shape[1], target_img.shape[0]), borderValue=0.0)
+            img_white = cv2.warpAffine(img_white, IM, (target_img.shape[1], target_img.shape[0]), borderValue=0.0)
+            fake_diff = cv2.warpAffine(fake_diff, IM, (target_img.shape[1], target_img.shape[0]), borderValue=0.0)
+            img_white[img_white>20] = 255
+            fthresh = 10
+            fake_diff[fake_diff<fthresh] = 0
+            fake_diff[fake_diff>=fthresh] = 255
+            img_mask = img_white
+            mask_h_inds, mask_w_inds = np.where(img_mask==255)
+            mask_h = np.max(mask_h_inds) - np.min(mask_h_inds)
+            mask_w = np.max(mask_w_inds) - np.min(mask_w_inds)
+            mask_size = int(np.sqrt(mask_h*mask_w))
+            k = max(mask_size//10, 10)
+            #k = max(mask_size//20, 6)
+            #k = 6
+            kernel = np.ones((k,k),np.uint8)
+            img_mask = cv2.erode(img_mask,kernel,iterations = 1)
+            kernel = np.ones((2,2),np.uint8)
+            fake_diff = cv2.dilate(fake_diff,kernel,iterations = 1)
+            k = max(mask_size//20, 5)
+            #k = 3
+            #k = 3
+            kernel_size = (k, k)
+            blur_size = tuple(2*i+1 for i in kernel_size)
+            img_mask = cv2.GaussianBlur(img_mask, blur_size, 0)
+            k = 5
+            kernel_size = (k, k)
+            blur_size = tuple(2*i+1 for i in kernel_size)
+            fake_diff = cv2.GaussianBlur(fake_diff, blur_size, 0)
+            img_mask /= 255
+            fake_diff /= 255
+            #img_mask = fake_diff
+            img_mask = np.reshape(img_mask, [img_mask.shape[0],img_mask.shape[1],1])
+            fake_merged = img_mask * bgr_fake + (1-img_mask) * target_img.astype(np.float32)
+            fake_merged = fake_merged.astype(np.uint8)
+            return fake_merged
+
diff --git a/python-package/insightface/model_zoo/landmark.py b/python-package/insightface/model_zoo/landmark.py
new file mode 100644
index 000000000..598b4b29a
--- /dev/null
+++ b/python-package/insightface/model_zoo/landmark.py
@@ -0,0 +1,114 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+from __future__ import division
+import numpy as np
+import cv2
+import onnx
+import onnxruntime
+from ..utils import face_align
+from ..utils import transform
+from ..data import get_object
+
+__all__ = [
+    'Landmark',
+]
+
+
+class Landmark:
+    def __init__(self, model_file=None, session=None):
+        assert model_file is not None
+        self.model_file = model_file
+        self.session = session
+        find_sub = False
+        find_mul = False
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        for nid, node in enumerate(graph.node[:8]):
+            #print(nid, node.name)
+            if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                find_sub = True
+            if node.name.startswith('Mul') or node.name.startswith('_mul'):
+                find_mul = True
+            if nid<3 and node.name=='bn_data':
+                find_sub = True
+                find_mul = True
+        if find_sub and find_mul:
+            #mxnet arcface model
+            input_mean = 0.0
+            input_std = 1.0
+        else:
+            input_mean = 127.5
+            input_std = 128.0
+        self.input_mean = input_mean
+        self.input_std = input_std
+        #print('input mean and std:', model_file, self.input_mean, self.input_std)
+        if self.session is None:
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        input_name = input_cfg.name
+        self.input_size = tuple(input_shape[2:4][::-1])
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        output_names = []
+        for out in outputs:
+            output_names.append(out.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        assert len(self.output_names)==1
+        output_shape = outputs[0].shape
+        self.require_pose = False
+        #print('init output_shape:', output_shape)
+        if output_shape[1]==3309:
+            self.lmk_dim = 3
+            self.lmk_num = 68
+            self.mean_lmk = get_object('meanshape_68.pkl')
+            self.require_pose = True
+        else:
+            self.lmk_dim = 2
+            self.lmk_num = output_shape[1]//self.lmk_dim
+        self.taskname = 'landmark_%dd_%d'%(self.lmk_dim, self.lmk_num)
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+
+    def get(self, img, face):
+        bbox = face.bbox
+        w, h = (bbox[2] - bbox[0]), (bbox[3] - bbox[1])
+        center = (bbox[2] + bbox[0]) / 2, (bbox[3] + bbox[1]) / 2
+        rotate = 0
+        _scale = self.input_size[0]  / (max(w, h)*1.5)
+        #print('param:', img.shape, bbox, center, self.input_size, _scale, rotate)
+        aimg, M = face_align.transform(img, center, self.input_size[0], _scale, rotate)
+        input_size = tuple(aimg.shape[0:2][::-1])
+        #assert input_size==self.input_size
+        blob = cv2.dnn.blobFromImage(aimg, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        pred = self.session.run(self.output_names, {self.input_name : blob})[0][0]
+        if pred.shape[0] >= 3000:
+            pred = pred.reshape((-1, 3))
+        else:
+            pred = pred.reshape((-1, 2))
+        if self.lmk_num < pred.shape[0]:
+            pred = pred[self.lmk_num*-1:,:]
+        pred[:, 0:2] += 1
+        pred[:, 0:2] *= (self.input_size[0] // 2)
+        if pred.shape[1] == 3:
+            pred[:, 2] *= (self.input_size[0] // 2)
+
+        IM = cv2.invertAffineTransform(M)
+        pred = face_align.trans_points(pred, IM)
+        face[self.taskname] = pred
+        if self.require_pose:
+            P = transform.estimate_affine_matrix_3d23d(self.mean_lmk, pred)
+            s, R, t = transform.P2sRt(P)
+            rx, ry, rz = transform.matrix2angle(R)
+            pose = np.array( [rx, ry, rz], dtype=np.float32 )
+            face['pose'] = pose #pitch, yaw, roll
+        return pred
+
+
diff --git a/python-package/insightface/model_zoo/model_store.py b/python-package/insightface/model_zoo/model_store.py
index e76eafba2..b62dc5892 100644
--- a/python-package/insightface/model_zoo/model_store.py
+++ b/python-package/insightface/model_zoo/model_store.py
@@ -1,4 +1,3 @@
-
 """
 This code file mainly comes from https://github.com/dmlc/gluon-cv/blob/master/gluoncv/model_zoo/model_store.py
 """
@@ -11,34 +10,39 @@
 
 from ..utils import download, check_sha1
 
-_model_sha1 = {name: checksum for checksum, name in [
-    ('95be21b58e29e9c1237f229dae534bd854009ce0', 'arcface_r100_v1'),
-    ('', 'arcface_mfn_v1'),
-    ('39fd1e087a2a2ed70a154ac01fecaa86c315d01b', 'retinaface_r50_v1'),
-    ('2c9de8116d1f448fd1d4661f90308faae34c990a', 'retinaface_mnet025_v1'),
-    ('0db1d07921d005e6c9a5b38e059452fc5645e5a4', 'retinaface_mnet025_v2'),
-    ('7dd8111652b7aac2490c5dcddeb268e53ac643e6', 'genderage_v1'),
-]}
-
-base_repo_url = 'http://insightface.ai/files/'
+_model_sha1 = {
+    name: checksum
+    for checksum, name in [
+        ('95be21b58e29e9c1237f229dae534bd854009ce0', 'arcface_r100_v1'),
+        ('', 'arcface_mfn_v1'),
+        ('39fd1e087a2a2ed70a154ac01fecaa86c315d01b', 'retinaface_r50_v1'),
+        ('2c9de8116d1f448fd1d4661f90308faae34c990a', 'retinaface_mnet025_v1'),
+        ('0db1d07921d005e6c9a5b38e059452fc5645e5a4', 'retinaface_mnet025_v2'),
+        ('7dd8111652b7aac2490c5dcddeb268e53ac643e6', 'genderage_v1'),
+    ]
+}
+
+base_repo_url = 'https://insightface.ai/files/'
 _url_format = '{repo_url}models/{file_name}.zip'
 
 
 def short_hash(name):
     if name not in _model_sha1:
-        raise ValueError('Pretrained model for {name} is not available.'.format(name=name))
+        raise ValueError(
+            'Pretrained model for {name} is not available.'.format(name=name))
     return _model_sha1[name][:8]
 
 
 def find_params_file(dir_path):
     if not os.path.exists(dir_path):
         return None
-    paths = glob.glob("%s/*.params"%dir_path)
-    if len(paths)==0:
+    paths = glob.glob(os.path.join(dir_path, "*.params"))
+    if len(paths) == 0:
         return None
     paths = sorted(paths)
     return paths[-1]
 
+
 def get_model_file(name, root=os.path.join('~', '.insightface', 'models')):
     r"""Return location for the pretrained on local file system.
 
@@ -68,7 +72,9 @@ def get_model_file(name, root=os.path.join('~', '.insightface', 'models')):
         if check_sha1(file_path, sha1_hash):
             return file_path
         else:
-            print('Mismatch in the content of model file detected. Downloading again.')
+            print(
+                'Mismatch in the content of model file detected. Downloading again.'
+            )
     else:
         print('Model file is not found. Downloading.')
 
@@ -92,6 +98,5 @@ def get_model_file(name, root=os.path.join('~', '.insightface', 'models')):
     if check_sha1(file_path, sha1_hash):
         return file_path
     else:
-        raise ValueError('Downloaded file has different hash. Please try again.')
-
-
+        raise ValueError(
+            'Downloaded file has different hash. Please try again.')
diff --git a/python-package/insightface/model_zoo/model_zoo.py b/python-package/insightface/model_zoo/model_zoo.py
index c3298ada8..73c6a5086 100644
--- a/python-package/insightface/model_zoo/model_zoo.py
+++ b/python-package/insightface/model_zoo/model_zoo.py
@@ -1,57 +1,97 @@
-# pylint: disable=wildcard-import, unused-wildcard-import
-"""
-This code file mainly comes from https://github.com/dmlc/gluon-cv/blob/master/gluoncv/model_zoo/model_zoo.py
-"""
-from .face_recognition import *
-from .face_detection import *
-from .face_genderage import *
-#from .face_alignment import *
-
-__all__ = ['get_model', 'get_model_list']
-
-_models = {
-    'arcface_r100_v1': arcface_r100_v1,
-    #'arcface_mfn_v1': arcface_mfn_v1,
-    #'arcface_outofreach_v1': arcface_outofreach_v1,
-    'retinaface_r50_v1': retinaface_r50_v1,
-    'retinaface_mnet025_v1': retinaface_mnet025_v1,
-    'retinaface_mnet025_v2': retinaface_mnet025_v2,
-    'genderage_v1': genderage_v1,
-}
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
 
+import os
+import os.path as osp
+import glob
+import onnxruntime
+from .arcface_onnx import *
+from .retinaface import *
+from .scrfd import *
+from .landmark import *
+from .attribute import Attribute
+from .inswapper import INSwapper
+from ..utils import download_onnx
 
-def get_model(name, **kwargs):
-    """Returns a pre-defined model by name
-
-    Parameters
-    ----------
-    name : str
-        Name of the model.
-    root : str, default '~/.insightface/models'
-        Location for keeping the model parameters.
-
-    Returns
-    -------
-    Model
-        The model.
-    """
-    name = name.lower()
-    if name not in _models:
-        err_str = '"%s" is not among the following model list:\n\t' % (name)
-        err_str += '%s' % ('\n\t'.join(sorted(_models.keys())))
-        raise ValueError(err_str)
-    net = _models[name](**kwargs)
-    return net
-
-
-def get_model_list():
-    """Get the entire list of model names in model_zoo.
-
-    Returns
-    -------
-    list of str
-        Entire list of model names in model_zoo.
-
-    """
-    return sorted(_models.keys())
+__all__ = ['get_model']
+
+
+class PickableInferenceSession(onnxruntime.InferenceSession): 
+    # This is a wrapper to make the current InferenceSession class pickable.
+    def __init__(self, model_path, **kwargs):
+        super().__init__(model_path, **kwargs)
+        self.model_path = model_path
+
+    def __getstate__(self):
+        return {'model_path': self.model_path}
+
+    def __setstate__(self, values):
+        model_path = values['model_path']
+        self.__init__(model_path)
+
+class ModelRouter:
+    def __init__(self, onnx_file):
+        self.onnx_file = onnx_file
 
+    def get_model(self, **kwargs):
+        session = PickableInferenceSession(self.onnx_file, **kwargs)
+        print(f'Applied providers: {session._providers}, with options: {session._provider_options}')
+        inputs = session.get_inputs()
+        input_cfg = inputs[0]
+        input_shape = input_cfg.shape
+        outputs = session.get_outputs()
+
+        if len(outputs)>=5:
+            return SCRFD(model_file=self.onnx_file, session=session)
+        elif input_shape[2]==192 and input_shape[3]==192:
+            return Landmark(model_file=self.onnx_file, session=session)
+        elif input_shape[2]==96 and input_shape[3]==96:
+            return Attribute(model_file=self.onnx_file, session=session)
+        elif len(inputs)==2 and input_shape[2]==128 and input_shape[3]==128:
+            return INSwapper(model_file=self.onnx_file, session=session)
+        elif input_shape[2]==input_shape[3] and input_shape[2]>=112 and input_shape[2]%16==0:
+            return ArcFaceONNX(model_file=self.onnx_file, session=session)
+        else:
+            #raise RuntimeError('error on model routing')
+            return None
+
+def find_onnx_file(dir_path):
+    if not os.path.exists(dir_path):
+        return None
+    paths = glob.glob(osp.join(dir_path, "*.onnx"))
+    if len(paths) == 0:
+        return None
+    paths = sorted(paths)
+    return paths[-1]
+
+def get_default_providers():
+    return ['CUDAExecutionProvider', 'CPUExecutionProvider']
+
+def get_default_provider_options():
+    return None
+
+def get_model(name, **kwargs):
+    root = kwargs.get('root', '~/.insightface')
+    root = os.path.expanduser(root)
+    model_root = osp.join(root, 'models')
+    allow_download = kwargs.get('download', False)
+    download_zip = kwargs.get('download_zip', False)
+    if not name.endswith('.onnx'):
+        model_dir = os.path.join(model_root, name)
+        model_file = find_onnx_file(model_dir)
+        if model_file is None:
+            return None
+    else:
+        model_file = name
+    if not osp.exists(model_file) and allow_download:
+        model_file = download_onnx('models', model_file, root=root, download_zip=download_zip)
+    assert osp.exists(model_file), 'model_file %s should exist'%model_file
+    assert osp.isfile(model_file), 'model_file %s should be a file'%model_file
+    router = ModelRouter(model_file)
+    providers = kwargs.get('providers', get_default_providers())
+    provider_options = kwargs.get('provider_options', get_default_provider_options())
+    model = router.get_model(providers=providers, provider_options=provider_options)
+    return model
diff --git a/python-package/insightface/model_zoo/retinaface.py b/python-package/insightface/model_zoo/retinaface.py
new file mode 100644
index 000000000..731d70d3e
--- /dev/null
+++ b/python-package/insightface/model_zoo/retinaface.py
@@ -0,0 +1,368 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-09-18
+# @Function      : 
+
+from __future__ import division
+import datetime
+import numpy as np
+import onnx
+import onnxruntime
+import os
+import os.path as osp
+import cv2
+import sys
+
+def softmax(z):
+    assert len(z.shape) == 2
+    s = np.max(z, axis=1)
+    s = s[:, np.newaxis] # necessary step to do broadcasting
+    e_x = np.exp(z - s)
+    div = np.sum(e_x, axis=1)
+    div = div[:, np.newaxis] # dito
+    return e_x / div
+
+def distance2bbox(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def distance2kps(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i%2] + distance[:, i]
+        py = points[:, i%2+1] + distance[:, i+1]
+        if max_shape is not None:
+            px = px.clamp(min=0, max=max_shape[1])
+            py = py.clamp(min=0, max=max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return np.stack(preds, axis=-1)
+
+class RetinaFace:
+    def __init__(self, model_file=None, session=None):
+        import onnxruntime
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'detection'
+        if self.session is None:
+            assert self.model_file is not None
+            assert osp.exists(self.model_file)
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        self.center_cache = {}
+        self.nms_thresh = 0.4
+        self.det_thresh = 0.5
+        self._init_vars()
+
+    def _init_vars(self):
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        #print(input_shape)
+        if isinstance(input_shape[2], str):
+            self.static_input_size = None
+        else:
+            self.static_input_size = tuple(input_shape[2:4][::-1])
+        self.input_size = self.static_input_size
+        self.input_sizes = [self.static_input_size] if self.static_input_size is not None else []
+        self._debug_det_size_printed = False
+        #print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        self.input_mean = 127.5
+        self.input_std = 128.0
+        #print(self.output_names)
+        #assert len(outputs)==10 or len(outputs)==15
+        self.use_kps = False
+        self._anchor_ratio = 1.0
+        self._num_anchors = 1
+        if len(outputs)==6:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+        elif len(outputs)==9:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+            self.use_kps = True
+        elif len(outputs)==10:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+        elif len(outputs)==15:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+            self.use_kps = True
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+        nms_thresh = kwargs.get('nms_thresh', None)
+        if nms_thresh is not None:
+            self.nms_thresh = nms_thresh
+        det_thresh = kwargs.get('det_thresh', None)
+        if det_thresh is not None:
+            self.det_thresh = det_thresh
+        input_size = kwargs.get('input_size', None)
+        if input_size is not None:
+            if self.static_input_size is not None:
+                self.input_size = self.static_input_size
+                self.input_sizes = [self.static_input_size]
+            else:
+                self.input_sizes = self._normalize_input_sizes(input_size)
+                self.input_size = self.input_sizes[-1]
+            self._debug_det_size_printed = False
+
+    def forward(self, img, threshold):
+        scores_list = []
+        bboxes_list = []
+        kpss_list = []
+        input_size = tuple(img.shape[0:2][::-1])
+        blob = cv2.dnn.blobFromImage(img, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_outs = self.session.run(self.output_names, {self.input_name : blob})
+
+        input_height = blob.shape[2]
+        input_width = blob.shape[3]
+        fmc = self.fmc
+        for idx, stride in enumerate(self._feat_stride_fpn):
+            scores = net_outs[idx]
+            bbox_preds = net_outs[idx+fmc]
+            bbox_preds = bbox_preds * stride
+            if self.use_kps:
+                kps_preds = net_outs[idx+fmc*2] * stride
+            height = input_height // stride
+            width = input_width // stride
+            K = height * width
+            key = (height, width, stride)
+            if key in self.center_cache:
+                anchor_centers = self.center_cache[key]
+            else:
+                #solution-1, c style:
+                #anchor_centers = np.zeros( (height, width, 2), dtype=np.float32 )
+                #for i in range(height):
+                #    anchor_centers[i, :, 1] = i
+                #for i in range(width):
+                #    anchor_centers[:, i, 0] = i
+
+                #solution-2:
+                #ax = np.arange(width, dtype=np.float32)
+                #ay = np.arange(height, dtype=np.float32)
+                #xv, yv = np.meshgrid(np.arange(width), np.arange(height))
+                #anchor_centers = np.stack([xv, yv], axis=-1).astype(np.float32)
+
+                #solution-3:
+                anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+                #print(anchor_centers.shape)
+
+                anchor_centers = (anchor_centers * stride).reshape( (-1, 2) )
+                if self._num_anchors>1:
+                    anchor_centers = np.stack([anchor_centers]*self._num_anchors, axis=1).reshape( (-1,2) )
+                if len(self.center_cache)<100:
+                    self.center_cache[key] = anchor_centers
+
+            pos_inds = np.where(scores>=threshold)[0]
+            bboxes = distance2bbox(anchor_centers, bbox_preds)
+            pos_scores = scores[pos_inds]
+            pos_bboxes = bboxes[pos_inds]
+            scores_list.append(pos_scores)
+            bboxes_list.append(pos_bboxes)
+            if self.use_kps:
+                kpss = distance2kps(anchor_centers, kps_preds)
+                #kpss = kps_preds
+                kpss = kpss.reshape( (kpss.shape[0], -1, 2) )
+                pos_kpss = kpss[pos_inds]
+                kpss_list.append(pos_kpss)
+        return scores_list, bboxes_list, kpss_list
+
+    def detect(self, img, input_size = None, max_num=0, metric='default'):
+        input_sizes = self._resolve_input_sizes(input_size)
+        assert input_sizes
+        pre_det_list = []
+        kpss_det_list = []
+        for input_size in input_sizes:
+            pre_det, kpss = self._detect_candidates(img, input_size)
+            if pre_det.shape[0] == 0:
+                continue
+            pre_det_list.append(pre_det)
+            if self.use_kps and kpss is not None:
+                kpss_det_list.append(kpss)
+        if not pre_det_list:
+            det = np.empty((0, 5), dtype=np.float32)
+            kpss = np.empty((0, 5, 2), dtype=np.float32) if self.use_kps else None
+            return det, kpss
+        pre_det = np.vstack(pre_det_list).astype(np.float32, copy=False)
+        order = pre_det[:, 4].argsort()[::-1]
+        pre_det = pre_det[order, :]
+        if self.use_kps and len(kpss_det_list) == len(pre_det_list):
+            kpss = np.vstack(kpss_det_list)[order, :, :]
+        else:
+            kpss = None
+        keep = self.nms(pre_det)
+        det = pre_det[keep, :]
+        if kpss is not None:
+            kpss = kpss[keep, :, :]
+        if max_num > 0 and det.shape[0] > max_num:
+            area = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                    det[:, 1])
+            img_center = img.shape[0] // 2, img.shape[1] // 2
+            offsets = np.vstack([
+                (det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2 - img_center[0]
+            ])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            if metric=='max':
+                values = area
+            else:
+                values = area - offset_dist_squared * 2.0  # some extra weight on the centering
+            bindex = np.argsort(
+                values)[::-1]  # some extra weight on the centering
+            bindex = bindex[0:max_num]
+            det = det[bindex, :]
+            if kpss is not None:
+                kpss = kpss[bindex, :]
+        return det, kpss
+
+    def _detect_candidates(self, img, input_size):
+        im_ratio = float(img.shape[0]) / img.shape[1]
+        model_ratio = float(input_size[1]) / input_size[0]
+        if im_ratio>model_ratio:
+            new_height = input_size[1]
+            new_width = int(new_height / im_ratio)
+        else:
+            new_width = input_size[0]
+            new_height = int(new_width * im_ratio)
+        det_scale = float(new_height) / img.shape[0]
+        resized_img = cv2.resize(img, (new_width, new_height))
+        det_img = np.zeros( (input_size[1], input_size[0], 3), dtype=np.uint8 )
+        det_img[:new_height, :new_width, :] = resized_img
+
+        scores_list, bboxes_list, kpss_list = self.forward(det_img, self.det_thresh)
+
+        if len(scores_list) == 0 or sum(score.size for score in scores_list) == 0:
+            kps_shape = (0, 5, 2) if self.use_kps else None
+            return np.empty((0, 5), dtype=np.float32), np.empty(kps_shape, dtype=np.float32) if kps_shape else None
+        scores = np.vstack(scores_list)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        bboxes = np.vstack(bboxes_list) / det_scale
+        if self.use_kps:
+            kpss = np.vstack(kpss_list) / det_scale
+        pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
+        pre_det = pre_det[order, :]
+        if self.use_kps:
+            kpss = kpss[order,:,:]
+        else:
+            kpss = None
+        return pre_det, kpss
+
+    def _resolve_input_sizes(self, input_size):
+        if input_size is not None:
+            return self._normalize_input_sizes(input_size)
+        if self.input_sizes:
+            return list(self.input_sizes)
+        if self.input_size is not None:
+            return [self.input_size]
+        return []
+
+    @staticmethod
+    def _normalize_input_sizes(input_size):
+        if input_size is None:
+            return []
+        if isinstance(input_size, np.ndarray):
+            input_size = input_size.tolist()
+        if (
+            isinstance(input_size, (list, tuple))
+            and len(input_size) > 0
+            and isinstance(input_size[0], (list, tuple, np.ndarray))
+        ):
+            values = input_size
+        else:
+            values = [input_size]
+        sizes = []
+        for item in values:
+            if isinstance(item, np.ndarray):
+                item = item.tolist()
+            if len(item) != 2:
+                raise ValueError('det_size must be a pair or a list of pairs')
+            width, height = int(item[0]), int(item[1])
+            if width <= 0 or height <= 0:
+                raise ValueError('det_size values must be positive')
+            size = (width, height)
+            if size not in sizes:
+                sizes.append(size)
+        return sizes
+
+    def nms(self, dets):
+        thresh = self.nms_thresh
+        x1 = dets[:, 0]
+        y1 = dets[:, 1]
+        x2 = dets[:, 2]
+        y2 = dets[:, 3]
+        scores = dets[:, 4]
+
+        areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+        order = scores.argsort()[::-1]
+
+        keep = []
+        while order.size > 0:
+            i = order[0]
+            keep.append(i)
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+            inds = np.where(ovr <= thresh)[0]
+            order = order[inds + 1]
+
+        return keep
+
+def get_retinaface(name, download=False, root='~/.insightface/models', **kwargs):
+    if not download:
+        assert os.path.exists(name)
+        return RetinaFace(name)
+    else:
+        from .model_store import get_model_file
+        _file = get_model_file("retinaface_%s" % name, root=root)
+        return retinaface(_file)
diff --git a/python-package/insightface/model_zoo/scrfd.py b/python-package/insightface/model_zoo/scrfd.py
new file mode 100644
index 000000000..32015ad90
--- /dev/null
+++ b/python-package/insightface/model_zoo/scrfd.py
@@ -0,0 +1,424 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+from __future__ import division
+import datetime
+import numpy as np
+import onnx
+import onnxruntime
+import os
+import os.path as osp
+import cv2
+import sys
+
+DEFAULT_DET_SIZES = [(128, 128), (640, 640)]
+
+
+def softmax(z):
+    assert len(z.shape) == 2
+    s = np.max(z, axis=1)
+    s = s[:, np.newaxis] # necessary step to do broadcasting
+    e_x = np.exp(z - s)
+    div = np.sum(e_x, axis=1)
+    div = div[:, np.newaxis] # dito
+    return e_x / div
+
+def distance2bbox(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def distance2kps(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i%2] + distance[:, i]
+        py = points[:, i%2+1] + distance[:, i+1]
+        if max_shape is not None:
+            px = px.clamp(min=0, max=max_shape[1])
+            py = py.clamp(min=0, max=max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return np.stack(preds, axis=-1)
+
+class SCRFD:
+    def __init__(self, model_file=None, session=None):
+        import onnxruntime
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'detection'
+        self.batched = False
+        if self.session is None:
+            assert self.model_file is not None
+            assert osp.exists(self.model_file)
+            self.session = onnxruntime.InferenceSession(self.model_file, None)
+        self.center_cache = {}
+        self.nms_thresh = 0.4
+        self.det_thresh = 0.5
+        self._init_vars()
+
+    def _init_vars(self):
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        #print(input_shape)
+        if isinstance(input_shape[2], str):
+            self.static_input_size = None
+        else:
+            self.static_input_size = tuple(input_shape[2:4][::-1])
+        self.input_size = self.static_input_size if self.static_input_size is not None else DEFAULT_DET_SIZES[-1]
+        self.input_sizes = [self.static_input_size] if self.static_input_size is not None else list(DEFAULT_DET_SIZES)
+        self._debug_det_size_printed = False
+        #print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        if len(outputs[0].shape) == 3:
+            self.batched = True
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        self.input_mean = 127.5
+        self.input_std = 128.0
+        #print(self.output_names)
+        #assert len(outputs)==10 or len(outputs)==15
+        self.use_kps = False
+        self._anchor_ratio = 1.0
+        self._num_anchors = 1
+        if len(outputs)==6:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+        elif len(outputs)==9:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+            self.use_kps = True
+        elif len(outputs)==10:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+        elif len(outputs)==15:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+            self.use_kps = True
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+        nms_thresh = kwargs.get('nms_thresh', None)
+        if nms_thresh is not None:
+            self.nms_thresh = nms_thresh
+        det_thresh = kwargs.get('det_thresh', None)
+        if det_thresh is not None:
+            self.det_thresh = det_thresh
+        input_size = kwargs.get('input_size', None)
+        if input_size is not None:
+            if self.static_input_size is not None:
+                self.input_size = self.static_input_size
+                self.input_sizes = [self.static_input_size]
+            else:
+                self.input_sizes = self._normalize_input_sizes(input_size)
+                self.input_size = self.input_sizes[-1]
+            self._debug_det_size_printed = False
+
+    def forward(self, img, threshold):
+        scores_list = []
+        bboxes_list = []
+        kpss_list = []
+        input_size = tuple(img.shape[0:2][::-1])
+        blob = cv2.dnn.blobFromImage(img, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_outs = self.session.run(self.output_names, {self.input_name : blob})
+
+        input_height = blob.shape[2]
+        input_width = blob.shape[3]
+        fmc = self.fmc
+        for idx, stride in enumerate(self._feat_stride_fpn):
+            # If model support batch dim, take first output
+            if self.batched:
+                scores = net_outs[idx][0]
+                bbox_preds = net_outs[idx + fmc][0]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2][0] * stride
+            # If model doesn't support batching take output as is
+            else:
+                scores = net_outs[idx]
+                bbox_preds = net_outs[idx + fmc]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2] * stride
+
+            height = input_height // stride
+            width = input_width // stride
+            K = height * width
+            key = (height, width, stride)
+            if key in self.center_cache:
+                anchor_centers = self.center_cache[key]
+            else:
+                #solution-1, c style:
+                #anchor_centers = np.zeros( (height, width, 2), dtype=np.float32 )
+                #for i in range(height):
+                #    anchor_centers[i, :, 1] = i
+                #for i in range(width):
+                #    anchor_centers[:, i, 0] = i
+
+                #solution-2:
+                #ax = np.arange(width, dtype=np.float32)
+                #ay = np.arange(height, dtype=np.float32)
+                #xv, yv = np.meshgrid(np.arange(width), np.arange(height))
+                #anchor_centers = np.stack([xv, yv], axis=-1).astype(np.float32)
+
+                #solution-3:
+                anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+                #print(anchor_centers.shape)
+
+                anchor_centers = (anchor_centers * stride).reshape( (-1, 2) )
+                if self._num_anchors>1:
+                    anchor_centers = np.stack([anchor_centers]*self._num_anchors, axis=1).reshape( (-1,2) )
+                if len(self.center_cache)<100:
+                    self.center_cache[key] = anchor_centers
+
+            pos_inds = np.where(scores>=threshold)[0]
+            bboxes = distance2bbox(anchor_centers, bbox_preds)
+            pos_scores = scores[pos_inds]
+            pos_bboxes = bboxes[pos_inds]
+            scores_list.append(pos_scores)
+            bboxes_list.append(pos_bboxes)
+            if self.use_kps:
+                kpss = distance2kps(anchor_centers, kps_preds)
+                #kpss = kps_preds
+                kpss = kpss.reshape( (kpss.shape[0], -1, 2) )
+                pos_kpss = kpss[pos_inds]
+                kpss_list.append(pos_kpss)
+        return scores_list, bboxes_list, kpss_list
+
+    def detect(self, img, input_size = None, max_num=0, metric='default'):
+        input_sizes = self._resolve_input_sizes(input_size)
+        assert input_sizes
+        pre_det_list = []
+        kpss_det_list = []
+        for input_size in input_sizes:
+            pre_det, kpss = self._detect_candidates(img, input_size)
+            if pre_det.shape[0] == 0:
+                continue
+            pre_det_list.append(pre_det)
+            if self.use_kps and kpss is not None:
+                kpss_det_list.append(kpss)
+        if not pre_det_list:
+            det = np.empty((0, 5), dtype=np.float32)
+            kpss = np.empty((0, 5, 2), dtype=np.float32) if self.use_kps else None
+            return det, kpss
+        pre_det = np.vstack(pre_det_list).astype(np.float32, copy=False)
+        order = pre_det[:, 4].argsort()[::-1]
+        pre_det = pre_det[order, :]
+        if self.use_kps and len(kpss_det_list) == len(pre_det_list):
+            kpss = np.vstack(kpss_det_list)[order, :, :]
+        else:
+            kpss = None
+        keep = self.nms(pre_det)
+        det = pre_det[keep, :]
+        if kpss is not None:
+            kpss = kpss[keep, :, :]
+        if max_num > 0 and det.shape[0] > max_num:
+            area = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                    det[:, 1])
+            img_center = img.shape[0] // 2, img.shape[1] // 2
+            offsets = np.vstack([
+                (det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2 - img_center[0]
+            ])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            if metric=='max':
+                values = area
+            else:
+                values = area - offset_dist_squared * 2.0  # some extra weight on the centering
+            bindex = np.argsort(
+                values)[::-1]  # some extra weight on the centering
+            bindex = bindex[0:max_num]
+            det = det[bindex, :]
+            if kpss is not None:
+                kpss = kpss[bindex, :]
+        return det, kpss
+
+    def _detect_candidates(self, img, input_size):
+        im_ratio = float(img.shape[0]) / img.shape[1]
+        model_ratio = float(input_size[1]) / input_size[0]
+        if im_ratio>model_ratio:
+            new_height = input_size[1]
+            new_width = int(new_height / im_ratio)
+        else:
+            new_width = input_size[0]
+            new_height = int(new_width * im_ratio)
+        det_scale = float(new_height) / img.shape[0]
+        resized_img = cv2.resize(img, (new_width, new_height))
+        det_img = np.zeros( (input_size[1], input_size[0], 3), dtype=np.uint8 )
+        det_img[:new_height, :new_width, :] = resized_img
+
+        scores_list, bboxes_list, kpss_list = self.forward(det_img, self.det_thresh)
+
+        if len(scores_list) == 0 or sum(score.size for score in scores_list) == 0:
+            kps_shape = (0, 5, 2) if self.use_kps else None
+            return np.empty((0, 5), dtype=np.float32), np.empty(kps_shape, dtype=np.float32) if kps_shape else None
+        scores = np.vstack(scores_list)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        bboxes = np.vstack(bboxes_list) / det_scale
+        if self.use_kps:
+            kpss = np.vstack(kpss_list) / det_scale
+        pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
+        pre_det = pre_det[order, :]
+        if self.use_kps:
+            kpss = kpss[order,:,:]
+        else:
+            kpss = None
+        return pre_det, kpss
+
+    def _resolve_input_sizes(self, input_size):
+        if input_size is not None:
+            return self._normalize_input_sizes(input_size)
+        if self.input_sizes:
+            return list(self.input_sizes)
+        if self.input_size is not None:
+            return [self.input_size]
+        return list(DEFAULT_DET_SIZES)
+
+    @staticmethod
+    def _normalize_input_sizes(input_size):
+        if input_size is None:
+            return []
+        if isinstance(input_size, np.ndarray):
+            input_size = input_size.tolist()
+        if (
+            isinstance(input_size, (list, tuple))
+            and len(input_size) > 0
+            and isinstance(input_size[0], (list, tuple, np.ndarray))
+        ):
+            values = input_size
+        else:
+            values = [input_size]
+        sizes = []
+        for item in values:
+            if isinstance(item, np.ndarray):
+                item = item.tolist()
+            if len(item) != 2:
+                raise ValueError('det_size must be a pair or a list of pairs')
+            width, height = int(item[0]), int(item[1])
+            if width == 0 and height == 0:
+                for size in DEFAULT_DET_SIZES:
+                    if size not in sizes:
+                        sizes.append(size)
+                continue
+            if width <= 0 or height <= 0:
+                raise ValueError('det_size values must be positive')
+            size = (width, height)
+            if size not in sizes:
+                sizes.append(size)
+        return sizes
+
+    def nms(self, dets):
+        thresh = self.nms_thresh
+        x1 = dets[:, 0]
+        y1 = dets[:, 1]
+        x2 = dets[:, 2]
+        y2 = dets[:, 3]
+        scores = dets[:, 4]
+
+        areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+        order = scores.argsort()[::-1]
+
+        keep = []
+        while order.size > 0:
+            i = order[0]
+            keep.append(i)
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+            inds = np.where(ovr <= thresh)[0]
+            order = order[inds + 1]
+
+        return keep
+
+def get_scrfd(name, download=False, root='~/.insightface/models', **kwargs):
+    if not download:
+        assert os.path.exists(name)
+        return SCRFD(name)
+    else:
+        from .model_store import get_model_file
+        _file = get_model_file("scrfd_%s" % name, root=root)
+        return SCRFD(_file)
+
+
+def scrfd_2p5gkps(**kwargs):
+    return get_scrfd("2p5gkps", download=True, **kwargs)
+
+
+if __name__ == '__main__':
+    import glob
+    detector = SCRFD(model_file='./det.onnx')
+    detector.prepare(-1)
+    img_paths = ['tests/data/t1.jpg']
+    for img_path in img_paths:
+        img = cv2.imread(img_path)
+
+        for _ in range(1):
+            ta = datetime.datetime.now()
+            #bboxes, kpss = detector.detect(img, 0.5, input_size = (640, 640))
+            bboxes, kpss = detector.detect(img, 0.5)
+            tb = datetime.datetime.now()
+            print('all cost:', (tb-ta).total_seconds()*1000)
+        print(img_path, bboxes.shape)
+        if kpss is not None:
+            print(kpss.shape)
+        for i in range(bboxes.shape[0]):
+            bbox = bboxes[i]
+            x1,y1,x2,y2,score = bbox.astype(int)
+            cv2.rectangle(img, (x1,y1)  , (x2,y2) , (255,0,0) , 2)
+            if kpss is not None:
+                kps = kpss[i]
+                for kp in kps:
+                    kp = kp.astype(int)
+                    cv2.circle(img, tuple(kp) , 1, (0,0,255) , 2)
+        filename = osp.basename(img_path)
+        print('output:', filename)
+        cv2.imwrite(osp.join('outputs', filename), img)
diff --git a/python-package/insightface/thirdparty/__init__.py b/python-package/insightface/thirdparty/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/python-package/insightface/thirdparty/face3d/__init__.py b/python-package/insightface/thirdparty/face3d/__init__.py
new file mode 100644
index 000000000..68b284753
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/__init__.py
@@ -0,0 +1,4 @@
+#import mesh
+#import morphable_model
+from . import mesh
+from . import morphable_model
diff --git a/python-package/insightface/thirdparty/face3d/mesh/__init__.cpp b/python-package/insightface/thirdparty/face3d/mesh/__init__.cpp
new file mode 100644
index 000000000..595c8b2c7
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/__init__.cpp
@@ -0,0 +1,2872 @@
+/* Generated by Cython 0.29.23 */
+
+/* BEGIN: Cython Metadata
+{
+    "distutils": {
+        "language": "c++",
+        "name": "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython",
+        "sources": [
+            "insightface/thirdparty/face3d/morphable_model/__init__.py",
+            "insightface/thirdparty/face3d/morphable_model/morphabel_model.py",
+            "insightface/thirdparty/face3d/morphable_model/load.py",
+            "insightface/thirdparty/face3d/morphable_model/fit.py",
+            "insightface/thirdparty/face3d/mesh_numpy/vis.py",
+            "insightface/thirdparty/face3d/mesh_numpy/__init__.py",
+            "insightface/thirdparty/face3d/mesh_numpy/transform.py",
+            "insightface/thirdparty/face3d/mesh_numpy/io.py",
+            "insightface/thirdparty/face3d/mesh_numpy/render.py",
+            "insightface/thirdparty/face3d/mesh_numpy/light.py",
+            "insightface/thirdparty/face3d/mesh/vis.py",
+            "insightface/thirdparty/face3d/mesh/__init__.py",
+            "insightface/thirdparty/face3d/mesh/transform.py",
+            "insightface/thirdparty/face3d/mesh/io.py",
+            "insightface/thirdparty/face3d/mesh/render.py",
+            "insightface/thirdparty/face3d/mesh/light.py"
+        ]
+    },
+    "module_name": "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython"
+}
+END: Cython Metadata */
+
+#ifndef PY_SSIZE_T_CLEAN
+#define PY_SSIZE_T_CLEAN
+#endif /* PY_SSIZE_T_CLEAN */
+#include "Python.h"
+#ifndef Py_PYTHON_H
+    #error Python headers needed to compile C extensions, please install development version of Python.
+#elif PY_VERSION_HEX < 0x02060000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000)
+    #error Cython requires Python 2.6+ or Python 3.3+.
+#else
+#define CYTHON_ABI "0_29_23"
+#define CYTHON_HEX_VERSION 0x001D17F0
+#define CYTHON_FUTURE_DIVISION 0
+#include <stddef.h>
+#ifndef offsetof
+  #define offsetof(type, member) ( (size_t) & ((type*)0) -> member )
+#endif
+#if !defined(WIN32) && !defined(MS_WINDOWS)
+  #ifndef __stdcall
+    #define __stdcall
+  #endif
+  #ifndef __cdecl
+    #define __cdecl
+  #endif
+  #ifndef __fastcall
+    #define __fastcall
+  #endif
+#endif
+#ifndef DL_IMPORT
+  #define DL_IMPORT(t) t
+#endif
+#ifndef DL_EXPORT
+  #define DL_EXPORT(t) t
+#endif
+#define __PYX_COMMA ,
+#ifndef HAVE_LONG_LONG
+  #if PY_VERSION_HEX >= 0x02070000
+    #define HAVE_LONG_LONG
+  #endif
+#endif
+#ifndef PY_LONG_LONG
+  #define PY_LONG_LONG LONG_LONG
+#endif
+#ifndef Py_HUGE_VAL
+  #define Py_HUGE_VAL HUGE_VAL
+#endif
+#ifdef PYPY_VERSION
+  #define CYTHON_COMPILING_IN_PYPY 1
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #undef CYTHON_USE_TYPE_SLOTS
+  #define CYTHON_USE_TYPE_SLOTS 0
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #if PY_VERSION_HEX < 0x03050000
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_INTERNALS
+  #define CYTHON_USE_UNICODE_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #undef CYTHON_AVOID_BORROWED_REFS
+  #define CYTHON_AVOID_BORROWED_REFS 1
+  #undef CYTHON_ASSUME_SAFE_MACROS
+  #define CYTHON_ASSUME_SAFE_MACROS 0
+  #undef CYTHON_UNPACK_METHODS
+  #define CYTHON_UNPACK_METHODS 0
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+#elif defined(PYSTON_VERSION)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 1
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #undef CYTHON_USE_ASYNC_SLOTS
+  #define CYTHON_USE_ASYNC_SLOTS 0
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+#else
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 1
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYTYPE_LOOKUP
+    #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #elif !defined(CYTHON_USE_PYTYPE_LOOKUP)
+    #define CYTHON_USE_PYTYPE_LOOKUP 1
+  #endif
+  #if PY_MAJOR_VERSION < 3
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYLONG_INTERNALS
+    #define CYTHON_USE_PYLONG_INTERNALS 0
+  #elif !defined(CYTHON_USE_PYLONG_INTERNALS)
+    #define CYTHON_USE_PYLONG_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_PYLIST_INTERNALS
+    #define CYTHON_USE_PYLIST_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #if PY_VERSION_HEX < 0x030300F0
+    #undef CYTHON_USE_UNICODE_WRITER
+    #define CYTHON_USE_UNICODE_WRITER 0
+  #elif !defined(CYTHON_USE_UNICODE_WRITER)
+    #define CYTHON_USE_UNICODE_WRITER 1
+  #endif
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #ifndef CYTHON_FAST_THREAD_STATE
+    #define CYTHON_FAST_THREAD_STATE 1
+  #endif
+  #ifndef CYTHON_FAST_PYCALL
+    #define CYTHON_FAST_PYCALL 1
+  #endif
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT (PY_VERSION_HEX >= 0x03050000)
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1)
+  #endif
+  #ifndef CYTHON_USE_DICT_VERSIONS
+    #define CYTHON_USE_DICT_VERSIONS (PY_VERSION_HEX >= 0x030600B1)
+  #endif
+  #ifndef CYTHON_USE_EXC_INFO_STACK
+    #define CYTHON_USE_EXC_INFO_STACK (PY_VERSION_HEX >= 0x030700A3)
+  #endif
+#endif
+#if !defined(CYTHON_FAST_PYCCALL)
+#define CYTHON_FAST_PYCCALL  (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1)
+#endif
+#if CYTHON_USE_PYLONG_INTERNALS
+  #include "longintrepr.h"
+  #undef SHIFT
+  #undef BASE
+  #undef MASK
+  #ifdef SIZEOF_VOID_P
+    enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) };
+  #endif
+#endif
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+#ifndef __has_cpp_attribute
+  #define __has_cpp_attribute(x) 0
+#endif
+#ifndef CYTHON_RESTRICT
+  #if defined(__GNUC__)
+    #define CYTHON_RESTRICT __restrict__
+  #elif defined(_MSC_VER) && _MSC_VER >= 1400
+    #define CYTHON_RESTRICT __restrict
+  #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define CYTHON_RESTRICT restrict
+  #else
+    #define CYTHON_RESTRICT
+  #endif
+#endif
+#ifndef CYTHON_UNUSED
+# if defined(__GNUC__)
+#   if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4))
+#     define CYTHON_UNUSED __attribute__ ((__unused__))
+#   else
+#     define CYTHON_UNUSED
+#   endif
+# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER))
+#   define CYTHON_UNUSED __attribute__ ((__unused__))
+# else
+#   define CYTHON_UNUSED
+# endif
+#endif
+#ifndef CYTHON_MAYBE_UNUSED_VAR
+#  if defined(__cplusplus)
+     template<class T> void CYTHON_MAYBE_UNUSED_VAR( const T& ) { }
+#  else
+#    define CYTHON_MAYBE_UNUSED_VAR(x) (void)(x)
+#  endif
+#endif
+#ifndef CYTHON_NCP_UNUSED
+# if CYTHON_COMPILING_IN_CPYTHON
+#  define CYTHON_NCP_UNUSED
+# else
+#  define CYTHON_NCP_UNUSED CYTHON_UNUSED
+# endif
+#endif
+#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None)
+#ifdef _MSC_VER
+    #ifndef _MSC_STDINT_H_
+        #if _MSC_VER < 1300
+           typedef unsigned char     uint8_t;
+           typedef unsigned int      uint32_t;
+        #else
+           typedef unsigned __int8   uint8_t;
+           typedef unsigned __int32  uint32_t;
+        #endif
+    #endif
+#else
+   #include <stdint.h>
+#endif
+#ifndef CYTHON_FALLTHROUGH
+  #if defined(__cplusplus) && __cplusplus >= 201103L
+    #if __has_cpp_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH [[fallthrough]]
+    #elif __has_cpp_attribute(clang::fallthrough)
+      #define CYTHON_FALLTHROUGH [[clang::fallthrough]]
+    #elif __has_cpp_attribute(gnu::fallthrough)
+      #define CYTHON_FALLTHROUGH [[gnu::fallthrough]]
+    #endif
+  #endif
+  #ifndef CYTHON_FALLTHROUGH
+    #if __has_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH __attribute__((fallthrough))
+    #else
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+  #if defined(__clang__ ) && defined(__apple_build_version__)
+    #if __apple_build_version__ < 7000000
+      #undef  CYTHON_FALLTHROUGH
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+#endif
+
+#ifndef __cplusplus
+  #error "Cython files generated with the C++ option must be compiled with a C++ compiler."
+#endif
+#ifndef CYTHON_INLINE
+  #if defined(__clang__)
+    #define CYTHON_INLINE __inline__ __attribute__ ((__unused__))
+  #else
+    #define CYTHON_INLINE inline
+  #endif
+#endif
+template<typename T>
+void __Pyx_call_destructor(T& x) {
+    x.~T();
+}
+template<typename T>
+class __Pyx_FakeReference {
+  public:
+    __Pyx_FakeReference() : ptr(NULL) { }
+    __Pyx_FakeReference(const T& ref) : ptr(const_cast<T*>(&ref)) { }
+    T *operator->() { return ptr; }
+    T *operator&() { return ptr; }
+    operator T&() { return *ptr; }
+    template<typename U> bool operator ==(U other) { return *ptr == other; }
+    template<typename U> bool operator !=(U other) { return *ptr != other; }
+  private:
+    T *ptr;
+};
+
+#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag)
+  #define Py_OptimizeFlag 0
+#endif
+#define __PYX_BUILD_PY_SSIZE_T "n"
+#define CYTHON_FORMAT_SSIZE_T "z"
+#if PY_MAJOR_VERSION < 3
+  #define __Pyx_BUILTIN_MODULE_NAME "__builtin__"
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+  #define __Pyx_DefaultClassType PyClass_Type
+#else
+  #define __Pyx_BUILTIN_MODULE_NAME "builtins"
+#if PY_VERSION_HEX >= 0x030800A4 && PY_VERSION_HEX < 0x030800B2
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, 0, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+#else
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+#endif
+  #define __Pyx_DefaultClassType PyType_Type
+#endif
+#ifndef Py_TPFLAGS_CHECKTYPES
+  #define Py_TPFLAGS_CHECKTYPES 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_INDEX
+  #define Py_TPFLAGS_HAVE_INDEX 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_NEWBUFFER
+  #define Py_TPFLAGS_HAVE_NEWBUFFER 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_FINALIZE
+  #define Py_TPFLAGS_HAVE_FINALIZE 0
+#endif
+#ifndef METH_STACKLESS
+  #define METH_STACKLESS 0
+#endif
+#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL)
+  #ifndef METH_FASTCALL
+     #define METH_FASTCALL 0x80
+  #endif
+  typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs);
+  typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args,
+                                                          Py_ssize_t nargs, PyObject *kwnames);
+#else
+  #define __Pyx_PyCFunctionFast _PyCFunctionFast
+  #define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords
+#endif
+#if CYTHON_FAST_PYCCALL
+#define __Pyx_PyFastCFunction_Check(func)\
+    ((PyCFunction_Check(func) && (METH_FASTCALL == (PyCFunction_GET_FLAGS(func) & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS | METH_STACKLESS)))))
+#else
+#define __Pyx_PyFastCFunction_Check(func) 0
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc)
+  #define PyObject_Malloc(s)   PyMem_Malloc(s)
+  #define PyObject_Free(p)     PyMem_Free(p)
+  #define PyObject_Realloc(p)  PyMem_Realloc(p)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030400A1
+  #define PyMem_RawMalloc(n)           PyMem_Malloc(n)
+  #define PyMem_RawRealloc(p, n)       PyMem_Realloc(p, n)
+  #define PyMem_RawFree(p)             PyMem_Free(p)
+#endif
+#if CYTHON_COMPILING_IN_PYSTON
+  #define __Pyx_PyCode_HasFreeVars(co)  PyCode_HasFreeVars(co)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno) PyFrame_SetLineNumber(frame, lineno)
+#else
+  #define __Pyx_PyCode_HasFreeVars(co)  (PyCode_GetNumFree(co) > 0)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno)  (frame)->f_lineno = (lineno)
+#endif
+#if !CYTHON_FAST_THREAD_STATE || PY_VERSION_HEX < 0x02070000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#elif PY_VERSION_HEX >= 0x03060000
+  #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet()
+#elif PY_VERSION_HEX >= 0x03000000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#else
+  #define __Pyx_PyThreadState_Current _PyThreadState_Current
+#endif
+#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT)
+#include "pythread.h"
+#define Py_tss_NEEDS_INIT 0
+typedef int Py_tss_t;
+static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) {
+  *key = PyThread_create_key();
+  return 0;
+}
+static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) {
+  Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t));
+  *key = Py_tss_NEEDS_INIT;
+  return key;
+}
+static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) {
+  PyObject_Free(key);
+}
+static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) {
+  return *key != Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) {
+  PyThread_delete_key(*key);
+  *key = Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) {
+  return PyThread_set_key_value(*key, value);
+}
+static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) {
+  return PyThread_get_key_value(*key);
+}
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON || defined(_PyDict_NewPresized)
+#define __Pyx_PyDict_NewPresized(n)  ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n))
+#else
+#define __Pyx_PyDict_NewPresized(n)  PyDict_New()
+#endif
+#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_TrueDivide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceTrueDivide(x,y)
+#else
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_Divide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceDivide(x,y)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && CYTHON_USE_UNICODE_INTERNALS
+#define __Pyx_PyDict_GetItemStr(dict, name)  _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash)
+#else
+#define __Pyx_PyDict_GetItemStr(dict, name)  PyDict_GetItem(dict, name)
+#endif
+#if PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND)
+  #define CYTHON_PEP393_ENABLED 1
+  #define __Pyx_PyUnicode_READY(op)       (likely(PyUnicode_IS_READY(op)) ?\
+                                              0 : _PyUnicode_Ready((PyObject *)(op)))
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_LENGTH(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i)
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   PyUnicode_MAX_CHAR_VALUE(u)
+  #define __Pyx_PyUnicode_KIND(u)         PyUnicode_KIND(u)
+  #define __Pyx_PyUnicode_DATA(u)         PyUnicode_DATA(u)
+  #define __Pyx_PyUnicode_READ(k, d, i)   PyUnicode_READ(k, d, i)
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  PyUnicode_WRITE(k, d, i, ch)
+  #if defined(PyUnicode_IS_READY) && defined(PyUnicode_GET_SIZE)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u)))
+  #else
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_LENGTH(u))
+  #endif
+#else
+  #define CYTHON_PEP393_ENABLED 0
+  #define PyUnicode_1BYTE_KIND  1
+  #define PyUnicode_2BYTE_KIND  2
+  #define PyUnicode_4BYTE_KIND  4
+  #define __Pyx_PyUnicode_READY(op)       (0)
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_SIZE(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i]))
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   ((sizeof(Py_UNICODE) == 2) ? 65535 : 1114111)
+  #define __Pyx_PyUnicode_KIND(u)         (sizeof(Py_UNICODE))
+  #define __Pyx_PyUnicode_DATA(u)         ((void*)PyUnicode_AS_UNICODE(u))
+  #define __Pyx_PyUnicode_READ(k, d, i)   ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i]))
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  (((void)(k)), ((Py_UNICODE*)d)[i] = ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_SIZE(u))
+#endif
+#if CYTHON_COMPILING_IN_PYPY
+  #define __Pyx_PyUnicode_Concat(a, b)      PyNumber_Add(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  PyNumber_Add(a, b)
+#else
+  #define __Pyx_PyUnicode_Concat(a, b)      PyUnicode_Concat(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\
+      PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b))
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyUnicode_Contains)
+  #define PyUnicode_Contains(u, s)  PySequence_Contains(u, s)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyByteArray_Check)
+  #define PyByteArray_Check(obj)  PyObject_TypeCheck(obj, &PyByteArray_Type)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Format)
+  #define PyObject_Format(obj, fmt)  PyObject_CallMethod(obj, "__format__", "O", fmt)
+#endif
+#define __Pyx_PyString_FormatSafe(a, b)   ((unlikely((a) == Py_None || (PyString_Check(b) && !PyString_CheckExact(b)))) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b))
+#define __Pyx_PyUnicode_FormatSafe(a, b)  ((unlikely((a) == Py_None || (PyUnicode_Check(b) && !PyUnicode_CheckExact(b)))) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b))
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyString_Format(a, b)  PyUnicode_Format(a, b)
+#else
+  #define __Pyx_PyString_Format(a, b)  PyString_Format(a, b)
+#endif
+#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII)
+  #define PyObject_ASCII(o)            PyObject_Repr(o)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBaseString_Type            PyUnicode_Type
+  #define PyStringObject               PyUnicodeObject
+  #define PyString_Type                PyUnicode_Type
+  #define PyString_Check               PyUnicode_Check
+  #define PyString_CheckExact          PyUnicode_CheckExact
+#ifndef PyObject_Unicode
+  #define PyObject_Unicode             PyObject_Str
+#endif
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj)
+  #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj)
+#else
+  #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj))
+  #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj))
+#endif
+#ifndef PySet_CheckExact
+  #define PySet_CheckExact(obj)        (Py_TYPE(obj) == &PySet_Type)
+#endif
+#if PY_VERSION_HEX >= 0x030900A4
+  #define __Pyx_SET_REFCNT(obj, refcnt) Py_SET_REFCNT(obj, refcnt)
+  #define __Pyx_SET_SIZE(obj, size) Py_SET_SIZE(obj, size)
+#else
+  #define __Pyx_SET_REFCNT(obj, refcnt) Py_REFCNT(obj) = (refcnt)
+  #define __Pyx_SET_SIZE(obj, size) Py_SIZE(obj) = (size)
+#endif
+#if CYTHON_ASSUME_SAFE_MACROS
+  #define __Pyx_PySequence_SIZE(seq)  Py_SIZE(seq)
+#else
+  #define __Pyx_PySequence_SIZE(seq)  PySequence_Size(seq)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyIntObject                  PyLongObject
+  #define PyInt_Type                   PyLong_Type
+  #define PyInt_Check(op)              PyLong_Check(op)
+  #define PyInt_CheckExact(op)         PyLong_CheckExact(op)
+  #define PyInt_FromString             PyLong_FromString
+  #define PyInt_FromUnicode            PyLong_FromUnicode
+  #define PyInt_FromLong               PyLong_FromLong
+  #define PyInt_FromSize_t             PyLong_FromSize_t
+  #define PyInt_FromSsize_t            PyLong_FromSsize_t
+  #define PyInt_AsLong                 PyLong_AsLong
+  #define PyInt_AS_LONG                PyLong_AS_LONG
+  #define PyInt_AsSsize_t              PyLong_AsSsize_t
+  #define PyInt_AsUnsignedLongMask     PyLong_AsUnsignedLongMask
+  #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask
+  #define PyNumber_Int                 PyNumber_Long
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBoolObject                 PyLongObject
+#endif
+#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY
+  #ifndef PyUnicode_InternFromString
+    #define PyUnicode_InternFromString(s) PyUnicode_FromString(s)
+  #endif
+#endif
+#if PY_VERSION_HEX < 0x030200A4
+  typedef long Py_hash_t;
+  #define __Pyx_PyInt_FromHash_t PyInt_FromLong
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsLong
+#else
+  #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsSsize_t
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyMethod_New(func, self, klass) ((self) ? ((void)(klass), PyMethod_New(func, self)) : __Pyx_NewRef(func))
+#else
+  #define __Pyx_PyMethod_New(func, self, klass) PyMethod_New(func, self, klass)
+#endif
+#if CYTHON_USE_ASYNC_SLOTS
+  #if PY_VERSION_HEX >= 0x030500B1
+    #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods
+    #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async)
+  #else
+    #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved))
+  #endif
+#else
+  #define __Pyx_PyType_AsAsync(obj) NULL
+#endif
+#ifndef __Pyx_PyAsyncMethodsStruct
+    typedef struct {
+        unaryfunc am_await;
+        unaryfunc am_aiter;
+        unaryfunc am_anext;
+    } __Pyx_PyAsyncMethodsStruct;
+#endif
+
+#if defined(WIN32) || defined(MS_WINDOWS)
+  #define _USE_MATH_DEFINES
+#endif
+#include <math.h>
+#ifdef NAN
+#define __PYX_NAN() ((float) NAN)
+#else
+static CYTHON_INLINE float __PYX_NAN() {
+  float value;
+  memset(&value, 0xFF, sizeof(value));
+  return value;
+}
+#endif
+#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL)
+#define __Pyx_truncl trunc
+#else
+#define __Pyx_truncl truncl
+#endif
+
+#define __PYX_MARK_ERR_POS(f_index, lineno) \
+    { __pyx_filename = __pyx_f[f_index]; (void)__pyx_filename; __pyx_lineno = lineno; (void)__pyx_lineno; __pyx_clineno = __LINE__; (void)__pyx_clineno; }
+#define __PYX_ERR(f_index, lineno, Ln_error) \
+    { __PYX_MARK_ERR_POS(f_index, lineno) goto Ln_error; }
+
+#ifndef __PYX_EXTERN_C
+  #ifdef __cplusplus
+    #define __PYX_EXTERN_C extern "C"
+  #else
+    #define __PYX_EXTERN_C extern
+  #endif
+#endif
+
+#define __PYX_HAVE__face3d__mesh____init__
+#define __PYX_HAVE_API__face3d__mesh____init__
+/* Early includes */
+#ifdef _OPENMP
+#include <omp.h>
+#endif /* _OPENMP */
+
+#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS)
+#define CYTHON_WITHOUT_ASSERTIONS
+#endif
+
+typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding;
+                const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry;
+
+#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_UTF8 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT (PY_MAJOR_VERSION >= 3 && __PYX_DEFAULT_STRING_ENCODING_IS_UTF8)
+#define __PYX_DEFAULT_STRING_ENCODING ""
+#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString
+#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#define __Pyx_uchar_cast(c) ((unsigned char)c)
+#define __Pyx_long_cast(x) ((long)x)
+#define __Pyx_fits_Py_ssize_t(v, type, is_signed)  (\
+    (sizeof(type) < sizeof(Py_ssize_t))  ||\
+    (sizeof(type) > sizeof(Py_ssize_t) &&\
+          likely(v < (type)PY_SSIZE_T_MAX ||\
+                 v == (type)PY_SSIZE_T_MAX)  &&\
+          (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\
+                                v == (type)PY_SSIZE_T_MIN)))  ||\
+    (sizeof(type) == sizeof(Py_ssize_t) &&\
+          (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\
+                               v == (type)PY_SSIZE_T_MAX)))  )
+static CYTHON_INLINE int __Pyx_is_valid_index(Py_ssize_t i, Py_ssize_t limit) {
+    return (size_t) i < (size_t) limit;
+}
+#if defined (__cplusplus) && __cplusplus >= 201103L
+    #include <cstdlib>
+    #define __Pyx_sst_abs(value) std::abs(value)
+#elif SIZEOF_INT >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) abs(value)
+#elif SIZEOF_LONG >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) labs(value)
+#elif defined (_MSC_VER)
+    #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value))
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define __Pyx_sst_abs(value) llabs(value)
+#elif defined (__GNUC__)
+    #define __Pyx_sst_abs(value) __builtin_llabs(value)
+#else
+    #define __Pyx_sst_abs(value) ((value<0) ? -value : value)
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*);
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length);
+#define __Pyx_PyByteArray_FromString(s) PyByteArray_FromStringAndSize((const char*)s, strlen((const char*)s))
+#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l)
+#define __Pyx_PyBytes_FromString        PyBytes_FromString
+#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*);
+#if PY_MAJOR_VERSION < 3
+    #define __Pyx_PyStr_FromString        __Pyx_PyBytes_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#else
+    #define __Pyx_PyStr_FromString        __Pyx_PyUnicode_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize
+#endif
+#define __Pyx_PyBytes_AsWritableString(s)     ((char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableSString(s)    ((signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableUString(s)    ((unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsString(s)     ((const char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsSString(s)    ((const signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsUString(s)    ((const unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyObject_AsWritableString(s)    ((char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableSString(s)    ((signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableUString(s)    ((unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsSString(s)    ((const signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsUString(s)    ((const unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_FromCString(s)  __Pyx_PyObject_FromString((const char*)s)
+#define __Pyx_PyBytes_FromCString(s)   __Pyx_PyBytes_FromString((const char*)s)
+#define __Pyx_PyByteArray_FromCString(s)   __Pyx_PyByteArray_FromString((const char*)s)
+#define __Pyx_PyStr_FromCString(s)     __Pyx_PyStr_FromString((const char*)s)
+#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s)
+static CYTHON_INLINE size_t __Pyx_Py_UNICODE_strlen(const Py_UNICODE *u) {
+    const Py_UNICODE *u_end = u;
+    while (*u_end++) ;
+    return (size_t)(u_end - u - 1);
+}
+#define __Pyx_PyUnicode_FromUnicode(u)       PyUnicode_FromUnicode(u, __Pyx_Py_UNICODE_strlen(u))
+#define __Pyx_PyUnicode_FromUnicodeAndLength PyUnicode_FromUnicode
+#define __Pyx_PyUnicode_AsUnicode            PyUnicode_AsUnicode
+#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj)
+#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None)
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject*);
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x);
+#define __Pyx_PySequence_Tuple(obj)\
+    (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj))
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*);
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t);
+#if CYTHON_ASSUME_SAFE_MACROS
+#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x))
+#else
+#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x)
+#endif
+#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x))
+#if PY_MAJOR_VERSION >= 3
+#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x))
+#else
+#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x))
+#endif
+#define __Pyx_PyNumber_Float(x) (PyFloat_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Float(x))
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+static int __Pyx_sys_getdefaultencoding_not_ascii;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    PyObject* ascii_chars_u = NULL;
+    PyObject* ascii_chars_b = NULL;
+    const char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    if (strcmp(default_encoding_c, "ascii") == 0) {
+        __Pyx_sys_getdefaultencoding_not_ascii = 0;
+    } else {
+        char ascii_chars[128];
+        int c;
+        for (c = 0; c < 128; c++) {
+            ascii_chars[c] = c;
+        }
+        __Pyx_sys_getdefaultencoding_not_ascii = 1;
+        ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL);
+        if (!ascii_chars_u) goto bad;
+        ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL);
+        if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.",
+                default_encoding_c);
+            goto bad;
+        }
+        Py_DECREF(ascii_chars_u);
+        Py_DECREF(ascii_chars_b);
+    }
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    Py_XDECREF(ascii_chars_u);
+    Py_XDECREF(ascii_chars_b);
+    return -1;
+}
+#endif
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL)
+#else
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL)
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+static char* __PYX_DEFAULT_STRING_ENCODING;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c) + 1);
+    if (!__PYX_DEFAULT_STRING_ENCODING) goto bad;
+    strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c);
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    return -1;
+}
+#endif
+#endif
+
+
+/* Test for GCC > 2.95 */
+#if defined(__GNUC__)     && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95)))
+  #define likely(x)   __builtin_expect(!!(x), 1)
+  #define unlikely(x) __builtin_expect(!!(x), 0)
+#else /* !__GNUC__ or GCC < 2.95 */
+  #define likely(x)   (x)
+  #define unlikely(x) (x)
+#endif /* __GNUC__ */
+static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; }
+
+static PyObject *__pyx_m = NULL;
+static PyObject *__pyx_d;
+static PyObject *__pyx_b;
+static PyObject *__pyx_cython_runtime = NULL;
+static PyObject *__pyx_empty_tuple;
+static PyObject *__pyx_empty_bytes;
+static PyObject *__pyx_empty_unicode;
+static int __pyx_lineno;
+static int __pyx_clineno = 0;
+static const char * __pyx_cfilenm= __FILE__;
+static const char *__pyx_filename;
+
+
+static const char *__pyx_f[] = {
+  "insightface/thirdparty/face3d/mesh/__init__.py",
+};
+
+/*--- Type declarations ---*/
+
+/* --- Runtime support code (head) --- */
+/* Refnanny.proto */
+#ifndef CYTHON_REFNANNY
+  #define CYTHON_REFNANNY 0
+#endif
+#if CYTHON_REFNANNY
+  typedef struct {
+    void (*INCREF)(void*, PyObject*, int);
+    void (*DECREF)(void*, PyObject*, int);
+    void (*GOTREF)(void*, PyObject*, int);
+    void (*GIVEREF)(void*, PyObject*, int);
+    void* (*SetupContext)(const char*, int, const char*);
+    void (*FinishContext)(void**);
+  } __Pyx_RefNannyAPIStruct;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname);
+  #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL;
+#ifdef WITH_THREAD
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          if (acquire_gil) {\
+              PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+              PyGILState_Release(__pyx_gilstate_save);\
+          } else {\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+          }
+#else
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__)
+#endif
+  #define __Pyx_RefNannyFinishContext()\
+          __Pyx_RefNanny->FinishContext(&__pyx_refnanny)
+  #define __Pyx_INCREF(r)  __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_DECREF(r)  __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GOTREF(r)  __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_XINCREF(r)  do { if((r) != NULL) {__Pyx_INCREF(r); }} while(0)
+  #define __Pyx_XDECREF(r)  do { if((r) != NULL) {__Pyx_DECREF(r); }} while(0)
+  #define __Pyx_XGOTREF(r)  do { if((r) != NULL) {__Pyx_GOTREF(r); }} while(0)
+  #define __Pyx_XGIVEREF(r) do { if((r) != NULL) {__Pyx_GIVEREF(r);}} while(0)
+#else
+  #define __Pyx_RefNannyDeclarations
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)
+  #define __Pyx_RefNannyFinishContext()
+  #define __Pyx_INCREF(r) Py_INCREF(r)
+  #define __Pyx_DECREF(r) Py_DECREF(r)
+  #define __Pyx_GOTREF(r)
+  #define __Pyx_GIVEREF(r)
+  #define __Pyx_XINCREF(r) Py_XINCREF(r)
+  #define __Pyx_XDECREF(r) Py_XDECREF(r)
+  #define __Pyx_XGOTREF(r)
+  #define __Pyx_XGIVEREF(r)
+#endif
+#define __Pyx_XDECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_XDECREF(tmp);\
+    } while (0)
+#define __Pyx_DECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_DECREF(tmp);\
+    } while (0)
+#define __Pyx_CLEAR(r)    do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0)
+#define __Pyx_XCLEAR(r)   do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0)
+
+/* PyObjectGetAttrStr.proto */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name);
+#else
+#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n)
+#endif
+
+/* SetPackagePathFromImportLib.proto */
+#if PY_MAJOR_VERSION >= 3 && !CYTHON_PEP489_MULTI_PHASE_INIT
+static int __Pyx_SetPackagePathFromImportLib(const char* parent_package_name, PyObject *module_name);
+#else
+#define __Pyx_SetPackagePathFromImportLib(a, b) 0
+#endif
+
+/* Import.proto */
+static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level);
+
+/* ImportFrom.proto */
+static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name);
+
+/* PyDictVersioning.proto */
+#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+#define __PYX_DICT_VERSION_INIT  ((PY_UINT64_T) -1)
+#define __PYX_GET_DICT_VERSION(dict)  (((PyDictObject*)(dict))->ma_version_tag)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)\
+    (version_var) = __PYX_GET_DICT_VERSION(dict);\
+    (cache_var) = (value);
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) {\
+    static PY_UINT64_T __pyx_dict_version = 0;\
+    static PyObject *__pyx_dict_cached_value = NULL;\
+    if (likely(__PYX_GET_DICT_VERSION(DICT) == __pyx_dict_version)) {\
+        (VAR) = __pyx_dict_cached_value;\
+    } else {\
+        (VAR) = __pyx_dict_cached_value = (LOOKUP);\
+        __pyx_dict_version = __PYX_GET_DICT_VERSION(DICT);\
+    }\
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj);
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj);
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version);
+#else
+#define __PYX_GET_DICT_VERSION(dict)  (0)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP)  (VAR) = (LOOKUP);
+#endif
+
+/* PyThreadStateGet.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyThreadState_declare  PyThreadState *__pyx_tstate;
+#define __Pyx_PyThreadState_assign  __pyx_tstate = __Pyx_PyThreadState_Current;
+#define __Pyx_PyErr_Occurred()  __pyx_tstate->curexc_type
+#else
+#define __Pyx_PyThreadState_declare
+#define __Pyx_PyThreadState_assign
+#define __Pyx_PyErr_Occurred()  PyErr_Occurred()
+#endif
+
+/* PyErrFetchRestore.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)    __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)    __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL))
+#else
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#endif
+#else
+#define __Pyx_PyErr_Clear() PyErr_Clear()
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestoreInState(tstate, type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchInState(tstate, type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)  PyErr_Fetch(type, value, tb)
+#endif
+
+/* CLineInTraceback.proto */
+#ifdef CYTHON_CLINE_IN_TRACEBACK
+#define __Pyx_CLineForTraceback(tstate, c_line)  (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0)
+#else
+static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line);
+#endif
+
+/* CodeObjectCache.proto */
+typedef struct {
+    PyCodeObject* code_object;
+    int code_line;
+} __Pyx_CodeObjectCacheEntry;
+struct __Pyx_CodeObjectCache {
+    int count;
+    int max_count;
+    __Pyx_CodeObjectCacheEntry* entries;
+};
+static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL};
+static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line);
+static PyCodeObject *__pyx_find_code_object(int code_line);
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object);
+
+/* AddTraceback.proto */
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename);
+
+/* GCCDiagnostics.proto */
+#if defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6))
+#define __Pyx_HAS_GCC_DIAGNOSTIC
+#endif
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *);
+
+/* FastTypeChecks.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type)
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2);
+#else
+#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type)
+#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type)
+#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2))
+#endif
+#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception)
+
+/* CheckBinaryVersion.proto */
+static int __Pyx_check_binary_version(void);
+
+/* InitStrings.proto */
+static int __Pyx_InitStrings(__Pyx_StringTabEntry *t);
+
+
+/* Module declarations from 'face3d.mesh.__init__' */
+#define __Pyx_MODULE_NAME "face3d.mesh.__init__"
+extern int __pyx_module_is_main_face3d__mesh____init__;
+int __pyx_module_is_main_face3d__mesh____init__ = 0;
+
+/* Implementation of 'face3d.mesh.__init__' */
+static const char __pyx_k_[] = "";
+static const char __pyx_k_io[] = "io";
+static const char __pyx_k_vis[] = "vis";
+static const char __pyx_k_main[] = "__main__";
+static const char __pyx_k_mesh[] = "mesh";
+static const char __pyx_k_name[] = "__name__";
+static const char __pyx_k_test[] = "__test__";
+static const char __pyx_k_light[] = "light";
+static const char __pyx_k_cython[] = "cython";
+static const char __pyx_k_import[] = "__import__";
+static const char __pyx_k_render[] = "render";
+static const char __pyx_k_transform[] = "transform";
+static const char __pyx_k_mesh_core_cython[] = "mesh_core_cython";
+static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback";
+static PyObject *__pyx_n_s_;
+static PyObject *__pyx_n_s_cline_in_traceback;
+static PyObject *__pyx_n_s_cython;
+static PyObject *__pyx_n_s_import;
+static PyObject *__pyx_n_s_io;
+static PyObject *__pyx_n_s_light;
+static PyObject *__pyx_n_s_main;
+static PyObject *__pyx_n_u_mesh;
+static PyObject *__pyx_n_s_mesh_core_cython;
+static PyObject *__pyx_n_s_name;
+static PyObject *__pyx_n_s_render;
+static PyObject *__pyx_n_s_test;
+static PyObject *__pyx_n_s_transform;
+static PyObject *__pyx_n_s_vis;
+/* Late includes */
+
+static PyMethodDef __pyx_methods[] = {
+  {0, 0, 0, 0}
+};
+
+#if PY_MAJOR_VERSION >= 3
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/
+static int __pyx_pymod_exec_mesh(PyObject* module); /*proto*/
+static PyModuleDef_Slot __pyx_moduledef_slots[] = {
+  {Py_mod_create, (void*)__pyx_pymod_create},
+  {Py_mod_exec, (void*)__pyx_pymod_exec_mesh},
+  {0, NULL}
+};
+#endif
+
+static struct PyModuleDef __pyx_moduledef = {
+    PyModuleDef_HEAD_INIT,
+    "mesh",
+    0, /* m_doc */
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    0, /* m_size */
+  #else
+    -1, /* m_size */
+  #endif
+    __pyx_methods /* m_methods */,
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    __pyx_moduledef_slots, /* m_slots */
+  #else
+    NULL, /* m_reload */
+  #endif
+    NULL, /* m_traverse */
+    NULL, /* m_clear */
+    NULL /* m_free */
+};
+#endif
+#ifndef CYTHON_SMALL_CODE
+#if defined(__clang__)
+    #define CYTHON_SMALL_CODE
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+    #define CYTHON_SMALL_CODE __attribute__((cold))
+#else
+    #define CYTHON_SMALL_CODE
+#endif
+#endif
+
+static __Pyx_StringTabEntry __pyx_string_tab[] = {
+  {&__pyx_n_s_, __pyx_k_, sizeof(__pyx_k_), 0, 0, 1, 1},
+  {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1},
+  {&__pyx_n_s_cython, __pyx_k_cython, sizeof(__pyx_k_cython), 0, 0, 1, 1},
+  {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1},
+  {&__pyx_n_s_io, __pyx_k_io, sizeof(__pyx_k_io), 0, 0, 1, 1},
+  {&__pyx_n_s_light, __pyx_k_light, sizeof(__pyx_k_light), 0, 0, 1, 1},
+  {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1},
+  {&__pyx_n_u_mesh, __pyx_k_mesh, sizeof(__pyx_k_mesh), 0, 1, 0, 1},
+  {&__pyx_n_s_mesh_core_cython, __pyx_k_mesh_core_cython, sizeof(__pyx_k_mesh_core_cython), 0, 0, 1, 1},
+  {&__pyx_n_s_name, __pyx_k_name, sizeof(__pyx_k_name), 0, 0, 1, 1},
+  {&__pyx_n_s_render, __pyx_k_render, sizeof(__pyx_k_render), 0, 0, 1, 1},
+  {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1},
+  {&__pyx_n_s_transform, __pyx_k_transform, sizeof(__pyx_k_transform), 0, 0, 1, 1},
+  {&__pyx_n_s_vis, __pyx_k_vis, sizeof(__pyx_k_vis), 0, 0, 1, 1},
+  {0, 0, 0, 0, 0, 0, 0}
+};
+static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) {
+  return 0;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0);
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitGlobals(void) {
+  if (__Pyx_InitStrings(__pyx_string_tab) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_modinit_global_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_import_code(void); /*proto*/
+
+static int __Pyx_modinit_global_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0);
+  /*--- Global init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_variable_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0);
+  /*--- Variable export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0);
+  /*--- Function export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0);
+  /*--- Type init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0);
+  /*--- Type import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_variable_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0);
+  /*--- Variable import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0);
+  /*--- Function import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+
+#ifndef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#elif PY_MAJOR_VERSION < 3
+#ifdef __cplusplus
+#define __Pyx_PyMODINIT_FUNC extern "C" void
+#else
+#define __Pyx_PyMODINIT_FUNC void
+#endif
+#else
+#ifdef __cplusplus
+#define __Pyx_PyMODINIT_FUNC extern "C" PyObject *
+#else
+#define __Pyx_PyMODINIT_FUNC PyObject *
+#endif
+#endif
+
+
+#if PY_MAJOR_VERSION < 3
+__Pyx_PyMODINIT_FUNC initmesh(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC initmesh(void)
+#else
+__Pyx_PyMODINIT_FUNC PyInit_mesh(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC PyInit_mesh(void)
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+{
+  return PyModuleDef_Init(&__pyx_moduledef);
+}
+static CYTHON_SMALL_CODE int __Pyx_check_single_interpreter(void) {
+    #if PY_VERSION_HEX >= 0x030700A1
+    static PY_INT64_T main_interpreter_id = -1;
+    PY_INT64_T current_id = PyInterpreterState_GetID(PyThreadState_Get()->interp);
+    if (main_interpreter_id == -1) {
+        main_interpreter_id = current_id;
+        return (unlikely(current_id == -1)) ? -1 : 0;
+    } else if (unlikely(main_interpreter_id != current_id))
+    #else
+    static PyInterpreterState *main_interpreter = NULL;
+    PyInterpreterState *current_interpreter = PyThreadState_Get()->interp;
+    if (!main_interpreter) {
+        main_interpreter = current_interpreter;
+    } else if (unlikely(main_interpreter != current_interpreter))
+    #endif
+    {
+        PyErr_SetString(
+            PyExc_ImportError,
+            "Interpreter change detected - this module can only be loaded into one interpreter per process.");
+        return -1;
+    }
+    return 0;
+}
+static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name, int allow_none) {
+    PyObject *value = PyObject_GetAttrString(spec, from_name);
+    int result = 0;
+    if (likely(value)) {
+        if (allow_none || value != Py_None) {
+            result = PyDict_SetItemString(moddict, to_name, value);
+        }
+        Py_DECREF(value);
+    } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) {
+        PyErr_Clear();
+    } else {
+        result = -1;
+    }
+    return result;
+}
+static CYTHON_SMALL_CODE PyObject* __pyx_pymod_create(PyObject *spec, CYTHON_UNUSED PyModuleDef *def) {
+    PyObject *module = NULL, *moddict, *modname;
+    if (__Pyx_check_single_interpreter())
+        return NULL;
+    if (__pyx_m)
+        return __Pyx_NewRef(__pyx_m);
+    modname = PyObject_GetAttrString(spec, "name");
+    if (unlikely(!modname)) goto bad;
+    module = PyModule_NewObject(modname);
+    Py_DECREF(modname);
+    if (unlikely(!module)) goto bad;
+    moddict = PyModule_GetDict(module);
+    if (unlikely(!moddict)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__", 0) < 0)) goto bad;
+    return module;
+bad:
+    Py_XDECREF(module);
+    return NULL;
+}
+
+
+static CYTHON_SMALL_CODE int __pyx_pymod_exec_mesh(PyObject *__pyx_pyinit_module)
+#endif
+#endif
+{
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannyDeclarations
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  if (__pyx_m) {
+    if (__pyx_m == __pyx_pyinit_module) return 0;
+    PyErr_SetString(PyExc_RuntimeError, "Module 'mesh' has already been imported. Re-initialisation is not supported.");
+    return -1;
+  }
+  #elif PY_MAJOR_VERSION >= 3
+  if (__pyx_m) return __Pyx_NewRef(__pyx_m);
+  #endif
+  #if CYTHON_REFNANNY
+__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny");
+if (!__Pyx_RefNanny) {
+  PyErr_Clear();
+  __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny");
+  if (!__Pyx_RefNanny)
+      Py_FatalError("failed to import 'refnanny' module");
+}
+#endif
+  __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_mesh(void)", 0);
+  if (__Pyx_check_binary_version() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pxy_PyFrame_Initialize_Offsets
+  __Pxy_PyFrame_Initialize_Offsets();
+  #endif
+  __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pyx_CyFunction_USED
+  if (__pyx_CyFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_FusedFunction_USED
+  if (__pyx_FusedFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Coroutine_USED
+  if (__pyx_Coroutine_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Generator_USED
+  if (__pyx_Generator_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_AsyncGen_USED
+  if (__pyx_AsyncGen_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_StopAsyncIteration_USED
+  if (__pyx_StopAsyncIteration_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  /*--- Library function declarations ---*/
+  /*--- Threads initialization code ---*/
+  #if defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS
+  #ifdef WITH_THREAD /* Python build with threading support? */
+  PyEval_InitThreads();
+  #endif
+  #endif
+  /*--- Module creation code ---*/
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  __pyx_m = __pyx_pyinit_module;
+  Py_INCREF(__pyx_m);
+  #else
+  #if PY_MAJOR_VERSION < 3
+  __pyx_m = Py_InitModule4("mesh", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m);
+  #else
+  __pyx_m = PyModule_Create(&__pyx_moduledef);
+  #endif
+  if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_d);
+  __pyx_b = PyImport_AddModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_b);
+  __pyx_cython_runtime = PyImport_AddModule((char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_cython_runtime);
+  if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  /*--- Initialize various global constants etc. ---*/
+  if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT)
+  if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  if (__pyx_module_is_main_face3d__mesh____init__) {
+    if (PyObject_SetAttr(__pyx_m, __pyx_n_s_name, __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  }
+  if (!CYTHON_PEP489_MULTI_PHASE_INIT) {
+    if (unlikely(__Pyx_SetPackagePathFromImportLib("face3d", __pyx_n_u_mesh) < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+  }
+  #if PY_MAJOR_VERSION >= 3
+  {
+    PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error)
+    if (!PyDict_GetItemString(modules, "face3d.mesh")) {
+      if (unlikely(PyDict_SetItemString(modules, "face3d.mesh", __pyx_m) < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+    }
+  }
+  #endif
+  /*--- Builtin init code ---*/
+  if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Constants init code ---*/
+  if (__Pyx_InitCachedConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Global type/function init code ---*/
+  (void)__Pyx_modinit_global_init_code();
+  (void)__Pyx_modinit_variable_export_code();
+  (void)__Pyx_modinit_function_export_code();
+  (void)__Pyx_modinit_type_init_code();
+  (void)__Pyx_modinit_type_import_code();
+  (void)__Pyx_modinit_variable_import_code();
+  (void)__Pyx_modinit_function_import_code();
+  /*--- Execution code ---*/
+  #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED)
+  if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+
+  /* "face3d/mesh/__init__.py":9
+ * #import render
+ * 
+ * from .cython import mesh_core_cython             # <<<<<<<<<<<<<<
+ * from . import io
+ * from . import vis
+ */
+  __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 9, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_INCREF(__pyx_n_s_mesh_core_cython);
+  __Pyx_GIVEREF(__pyx_n_s_mesh_core_cython);
+  PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_mesh_core_cython);
+  __pyx_t_2 = __Pyx_Import(__pyx_n_s_cython, __pyx_t_1, 1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 9, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_2, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 9, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_mesh_core_cython, __pyx_t_1) < 0) __PYX_ERR(0, 9, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+
+  /* "face3d/mesh/__init__.py":10
+ * 
+ * from .cython import mesh_core_cython
+ * from . import io             # <<<<<<<<<<<<<<
+ * from . import vis
+ * from . import transform
+ */
+  __pyx_t_2 = PyList_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 10, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_INCREF(__pyx_n_s_io);
+  __Pyx_GIVEREF(__pyx_n_s_io);
+  PyList_SET_ITEM(__pyx_t_2, 0, __pyx_n_s_io);
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_, __pyx_t_2, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 10, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __pyx_t_2 = __Pyx_ImportFrom(__pyx_t_1, __pyx_n_s_io); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 10, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_io, __pyx_t_2) < 0) __PYX_ERR(0, 10, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "face3d/mesh/__init__.py":11
+ * from .cython import mesh_core_cython
+ * from . import io
+ * from . import vis             # <<<<<<<<<<<<<<
+ * from . import transform
+ * from . import light
+ */
+  __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_INCREF(__pyx_n_s_vis);
+  __Pyx_GIVEREF(__pyx_n_s_vis);
+  PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_vis);
+  __pyx_t_2 = __Pyx_Import(__pyx_n_s_, __pyx_t_1, 1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_2, __pyx_n_s_vis); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_vis, __pyx_t_1) < 0) __PYX_ERR(0, 11, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+
+  /* "face3d/mesh/__init__.py":12
+ * from . import io
+ * from . import vis
+ * from . import transform             # <<<<<<<<<<<<<<
+ * from . import light
+ * from . import render
+ */
+  __pyx_t_2 = PyList_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 12, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_INCREF(__pyx_n_s_transform);
+  __Pyx_GIVEREF(__pyx_n_s_transform);
+  PyList_SET_ITEM(__pyx_t_2, 0, __pyx_n_s_transform);
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_, __pyx_t_2, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 12, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __pyx_t_2 = __Pyx_ImportFrom(__pyx_t_1, __pyx_n_s_transform); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 12, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_transform, __pyx_t_2) < 0) __PYX_ERR(0, 12, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "face3d/mesh/__init__.py":13
+ * from . import vis
+ * from . import transform
+ * from . import light             # <<<<<<<<<<<<<<
+ * from . import render
+ * 
+ */
+  __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 13, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_INCREF(__pyx_n_s_light);
+  __Pyx_GIVEREF(__pyx_n_s_light);
+  PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_light);
+  __pyx_t_2 = __Pyx_Import(__pyx_n_s_, __pyx_t_1, 1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 13, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_2, __pyx_n_s_light); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 13, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_light, __pyx_t_1) < 0) __PYX_ERR(0, 13, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+
+  /* "face3d/mesh/__init__.py":14
+ * from . import transform
+ * from . import light
+ * from . import render             # <<<<<<<<<<<<<<
+ * 
+ */
+  __pyx_t_2 = PyList_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 14, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  __Pyx_INCREF(__pyx_n_s_render);
+  __Pyx_GIVEREF(__pyx_n_s_render);
+  PyList_SET_ITEM(__pyx_t_2, 0, __pyx_n_s_render);
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_, __pyx_t_2, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 14, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __pyx_t_2 = __Pyx_ImportFrom(__pyx_t_1, __pyx_n_s_render); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 14, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_2);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_render, __pyx_t_2) < 0) __PYX_ERR(0, 14, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "face3d/mesh/__init__.py":1
+ * #from __future__ import absolute_import             # <<<<<<<<<<<<<<
+ * #from cython import mesh_core_cython
+ * #import io
+ */
+  __pyx_t_1 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /*--- Wrapped vars code ---*/
+
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_XDECREF(__pyx_t_2);
+  if (__pyx_m) {
+    if (__pyx_d) {
+      __Pyx_AddTraceback("init face3d.mesh.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename);
+    }
+    Py_CLEAR(__pyx_m);
+  } else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_ImportError, "init face3d.mesh.__init__");
+  }
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  return (__pyx_m != NULL) ? 0 : -1;
+  #elif PY_MAJOR_VERSION >= 3
+  return __pyx_m;
+  #else
+  return;
+  #endif
+}
+
+/* --- Runtime support code --- */
+/* Refnanny */
+#if CYTHON_REFNANNY
+static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) {
+    PyObject *m = NULL, *p = NULL;
+    void *r = NULL;
+    m = PyImport_ImportModule(modname);
+    if (!m) goto end;
+    p = PyObject_GetAttrString(m, "RefNannyAPI");
+    if (!p) goto end;
+    r = PyLong_AsVoidPtr(p);
+end:
+    Py_XDECREF(p);
+    Py_XDECREF(m);
+    return (__Pyx_RefNannyAPIStruct *)r;
+}
+#endif
+
+/* PyObjectGetAttrStr */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) {
+    PyTypeObject* tp = Py_TYPE(obj);
+    if (likely(tp->tp_getattro))
+        return tp->tp_getattro(obj, attr_name);
+#if PY_MAJOR_VERSION < 3
+    if (likely(tp->tp_getattr))
+        return tp->tp_getattr(obj, PyString_AS_STRING(attr_name));
+#endif
+    return PyObject_GetAttr(obj, attr_name);
+}
+#endif
+
+/* SetPackagePathFromImportLib */
+#if PY_MAJOR_VERSION >= 3 && !CYTHON_PEP489_MULTI_PHASE_INIT
+static int __Pyx_SetPackagePathFromImportLib(const char* parent_package_name, PyObject *module_name) {
+    PyObject *importlib, *loader, *osmod, *ossep, *parts, *package_path;
+    PyObject *path = NULL, *file_path = NULL;
+    int result;
+    if (parent_package_name) {
+        PyObject *package = PyImport_ImportModule(parent_package_name);
+        if (unlikely(!package))
+            goto bad;
+        path = PyObject_GetAttrString(package, "__path__");
+        Py_DECREF(package);
+        if (unlikely(!path) || unlikely(path == Py_None))
+            goto bad;
+    } else {
+        path = Py_None; Py_INCREF(Py_None);
+    }
+    importlib = PyImport_ImportModule("importlib");
+    if (unlikely(!importlib))
+        goto bad;
+    loader = PyObject_CallMethod(importlib, "find_loader", "(OO)", module_name, path);
+    Py_DECREF(importlib);
+    Py_DECREF(path); path = NULL;
+    if (unlikely(!loader))
+        goto bad;
+    file_path = PyObject_GetAttrString(loader, "path");
+    Py_DECREF(loader);
+    if (unlikely(!file_path))
+        goto bad;
+    if (unlikely(PyObject_SetAttrString(__pyx_m, "__file__", file_path) < 0))
+        goto bad;
+    osmod = PyImport_ImportModule("os");
+    if (unlikely(!osmod))
+        goto bad;
+    ossep = PyObject_GetAttrString(osmod, "sep");
+    Py_DECREF(osmod);
+    if (unlikely(!ossep))
+        goto bad;
+    parts = PyObject_CallMethod(file_path, "rsplit", "(Oi)", ossep, 1);
+    Py_DECREF(file_path); file_path = NULL;
+    Py_DECREF(ossep);
+    if (unlikely(!parts))
+        goto bad;
+    package_path = Py_BuildValue("[O]", PyList_GET_ITEM(parts, 0));
+    Py_DECREF(parts);
+    if (unlikely(!package_path))
+        goto bad;
+    goto set_path;
+bad:
+    PyErr_WriteUnraisable(module_name);
+    Py_XDECREF(path);
+    Py_XDECREF(file_path);
+    PyErr_Clear();
+    package_path = PyList_New(0);
+    if (unlikely(!package_path))
+        return -1;
+set_path:
+    result = PyObject_SetAttrString(__pyx_m, "__path__", package_path);
+    Py_DECREF(package_path);
+    return result;
+}
+#endif
+
+/* Import */
+static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) {
+    PyObject *empty_list = 0;
+    PyObject *module = 0;
+    PyObject *global_dict = 0;
+    PyObject *empty_dict = 0;
+    PyObject *list;
+    #if PY_MAJOR_VERSION < 3
+    PyObject *py_import;
+    py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import);
+    if (!py_import)
+        goto bad;
+    #endif
+    if (from_list)
+        list = from_list;
+    else {
+        empty_list = PyList_New(0);
+        if (!empty_list)
+            goto bad;
+        list = empty_list;
+    }
+    global_dict = PyModule_GetDict(__pyx_m);
+    if (!global_dict)
+        goto bad;
+    empty_dict = PyDict_New();
+    if (!empty_dict)
+        goto bad;
+    {
+        #if PY_MAJOR_VERSION >= 3
+        if (level == -1) {
+            if ((1) && (strchr(__Pyx_MODULE_NAME, '.'))) {
+                module = PyImport_ImportModuleLevelObject(
+                    name, global_dict, empty_dict, list, 1);
+                if (!module) {
+                    if (!PyErr_ExceptionMatches(PyExc_ImportError))
+                        goto bad;
+                    PyErr_Clear();
+                }
+            }
+            level = 0;
+        }
+        #endif
+        if (!module) {
+            #if PY_MAJOR_VERSION < 3
+            PyObject *py_level = PyInt_FromLong(level);
+            if (!py_level)
+                goto bad;
+            module = PyObject_CallFunctionObjArgs(py_import,
+                name, global_dict, empty_dict, list, py_level, (PyObject *)NULL);
+            Py_DECREF(py_level);
+            #else
+            module = PyImport_ImportModuleLevelObject(
+                name, global_dict, empty_dict, list, level);
+            #endif
+        }
+    }
+bad:
+    #if PY_MAJOR_VERSION < 3
+    Py_XDECREF(py_import);
+    #endif
+    Py_XDECREF(empty_list);
+    Py_XDECREF(empty_dict);
+    return module;
+}
+
+/* ImportFrom */
+static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name) {
+    PyObject* value = __Pyx_PyObject_GetAttrStr(module, name);
+    if (unlikely(!value) && PyErr_ExceptionMatches(PyExc_AttributeError)) {
+        PyErr_Format(PyExc_ImportError,
+        #if PY_MAJOR_VERSION < 3
+            "cannot import name %.230s", PyString_AS_STRING(name));
+        #else
+            "cannot import name %S", name);
+        #endif
+    }
+    return value;
+}
+
+/* PyDictVersioning */
+#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    return likely(dict) ? __PYX_GET_DICT_VERSION(dict) : 0;
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj) {
+    PyObject **dictptr = NULL;
+    Py_ssize_t offset = Py_TYPE(obj)->tp_dictoffset;
+    if (offset) {
+#if CYTHON_COMPILING_IN_CPYTHON
+        dictptr = (likely(offset > 0)) ? (PyObject **) ((char *)obj + offset) : _PyObject_GetDictPtr(obj);
+#else
+        dictptr = _PyObject_GetDictPtr(obj);
+#endif
+    }
+    return (dictptr && *dictptr) ? __PYX_GET_DICT_VERSION(*dictptr) : 0;
+}
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    if (unlikely(!dict) || unlikely(tp_dict_version != __PYX_GET_DICT_VERSION(dict)))
+        return 0;
+    return obj_dict_version == __Pyx_get_object_dict_version(obj);
+}
+#endif
+
+/* PyErrFetchRestore */
+#if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    tmp_type = tstate->curexc_type;
+    tmp_value = tstate->curexc_value;
+    tmp_tb = tstate->curexc_traceback;
+    tstate->curexc_type = type;
+    tstate->curexc_value = value;
+    tstate->curexc_traceback = tb;
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    *type = tstate->curexc_type;
+    *value = tstate->curexc_value;
+    *tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+}
+#endif
+
+/* CLineInTraceback */
+#ifndef CYTHON_CLINE_IN_TRACEBACK
+static int __Pyx_CLineForTraceback(CYTHON_NCP_UNUSED PyThreadState *tstate, int c_line) {
+    PyObject *use_cline;
+    PyObject *ptype, *pvalue, *ptraceback;
+#if CYTHON_COMPILING_IN_CPYTHON
+    PyObject **cython_runtime_dict;
+#endif
+    if (unlikely(!__pyx_cython_runtime)) {
+        return c_line;
+    }
+    __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback);
+#if CYTHON_COMPILING_IN_CPYTHON
+    cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime);
+    if (likely(cython_runtime_dict)) {
+        __PYX_PY_DICT_LOOKUP_IF_MODIFIED(
+            use_cline, *cython_runtime_dict,
+            __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback))
+    } else
+#endif
+    {
+      PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback);
+      if (use_cline_obj) {
+        use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True;
+        Py_DECREF(use_cline_obj);
+      } else {
+        PyErr_Clear();
+        use_cline = NULL;
+      }
+    }
+    if (!use_cline) {
+        c_line = 0;
+        PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False);
+    }
+    else if (use_cline == Py_False || (use_cline != Py_True && PyObject_Not(use_cline) != 0)) {
+        c_line = 0;
+    }
+    __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback);
+    return c_line;
+}
+#endif
+
+/* CodeObjectCache */
+static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) {
+    int start = 0, mid = 0, end = count - 1;
+    if (end >= 0 && code_line > entries[end].code_line) {
+        return count;
+    }
+    while (start < end) {
+        mid = start + (end - start) / 2;
+        if (code_line < entries[mid].code_line) {
+            end = mid;
+        } else if (code_line > entries[mid].code_line) {
+             start = mid + 1;
+        } else {
+            return mid;
+        }
+    }
+    if (code_line <= entries[mid].code_line) {
+        return mid;
+    } else {
+        return mid + 1;
+    }
+}
+static PyCodeObject *__pyx_find_code_object(int code_line) {
+    PyCodeObject* code_object;
+    int pos;
+    if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) {
+        return NULL;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) {
+        return NULL;
+    }
+    code_object = __pyx_code_cache.entries[pos].code_object;
+    Py_INCREF(code_object);
+    return code_object;
+}
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) {
+    int pos, i;
+    __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries;
+    if (unlikely(!code_line)) {
+        return;
+    }
+    if (unlikely(!entries)) {
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (likely(entries)) {
+            __pyx_code_cache.entries = entries;
+            __pyx_code_cache.max_count = 64;
+            __pyx_code_cache.count = 1;
+            entries[0].code_line = code_line;
+            entries[0].code_object = code_object;
+            Py_INCREF(code_object);
+        }
+        return;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) {
+        PyCodeObject* tmp = entries[pos].code_object;
+        entries[pos].code_object = code_object;
+        Py_DECREF(tmp);
+        return;
+    }
+    if (__pyx_code_cache.count == __pyx_code_cache.max_count) {
+        int new_max = __pyx_code_cache.max_count + 64;
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc(
+            __pyx_code_cache.entries, ((size_t)new_max) * sizeof(__Pyx_CodeObjectCacheEntry));
+        if (unlikely(!entries)) {
+            return;
+        }
+        __pyx_code_cache.entries = entries;
+        __pyx_code_cache.max_count = new_max;
+    }
+    for (i=__pyx_code_cache.count; i>pos; i--) {
+        entries[i] = entries[i-1];
+    }
+    entries[pos].code_line = code_line;
+    entries[pos].code_object = code_object;
+    __pyx_code_cache.count++;
+    Py_INCREF(code_object);
+}
+
+/* AddTraceback */
+#include "compile.h"
+#include "frameobject.h"
+#include "traceback.h"
+static PyCodeObject* __Pyx_CreateCodeObjectForTraceback(
+            const char *funcname, int c_line,
+            int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyObject *py_srcfile = 0;
+    PyObject *py_funcname = 0;
+    #if PY_MAJOR_VERSION < 3
+    py_srcfile = PyString_FromString(filename);
+    #else
+    py_srcfile = PyUnicode_FromString(filename);
+    #endif
+    if (!py_srcfile) goto bad;
+    if (c_line) {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #else
+        py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #endif
+    }
+    else {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromString(funcname);
+        #else
+        py_funcname = PyUnicode_FromString(funcname);
+        #endif
+    }
+    if (!py_funcname) goto bad;
+    py_code = __Pyx_PyCode_New(
+        0,
+        0,
+        0,
+        0,
+        0,
+        __pyx_empty_bytes, /*PyObject *code,*/
+        __pyx_empty_tuple, /*PyObject *consts,*/
+        __pyx_empty_tuple, /*PyObject *names,*/
+        __pyx_empty_tuple, /*PyObject *varnames,*/
+        __pyx_empty_tuple, /*PyObject *freevars,*/
+        __pyx_empty_tuple, /*PyObject *cellvars,*/
+        py_srcfile,   /*PyObject *filename,*/
+        py_funcname,  /*PyObject *name,*/
+        py_line,
+        __pyx_empty_bytes  /*PyObject *lnotab*/
+    );
+    Py_DECREF(py_srcfile);
+    Py_DECREF(py_funcname);
+    return py_code;
+bad:
+    Py_XDECREF(py_srcfile);
+    Py_XDECREF(py_funcname);
+    return NULL;
+}
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyFrameObject *py_frame = 0;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    if (c_line) {
+        c_line = __Pyx_CLineForTraceback(tstate, c_line);
+    }
+    py_code = __pyx_find_code_object(c_line ? -c_line : py_line);
+    if (!py_code) {
+        py_code = __Pyx_CreateCodeObjectForTraceback(
+            funcname, c_line, py_line, filename);
+        if (!py_code) goto bad;
+        __pyx_insert_code_object(c_line ? -c_line : py_line, py_code);
+    }
+    py_frame = PyFrame_New(
+        tstate,            /*PyThreadState *tstate,*/
+        py_code,           /*PyCodeObject *code,*/
+        __pyx_d,    /*PyObject *globals,*/
+        0                  /*PyObject *locals*/
+    );
+    if (!py_frame) goto bad;
+    __Pyx_PyFrame_SetLineNumber(py_frame, py_line);
+    PyTraceBack_Here(py_frame);
+bad:
+    Py_XDECREF(py_code);
+    Py_XDECREF(py_frame);
+}
+
+/* CIntToPy */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const long neg_one = (long) -1, const_zero = (long) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(long) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(long) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(long) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(long),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntFromPyVerify */
+#define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0)
+#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1)
+#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\
+    {\
+        func_type value = func_value;\
+        if (sizeof(target_type) < sizeof(func_type)) {\
+            if (unlikely(value != (func_type) (target_type) value)) {\
+                func_type zero = 0;\
+                if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\
+                    return (target_type) -1;\
+                if (is_unsigned && unlikely(value < zero))\
+                    goto raise_neg_overflow;\
+                else\
+                    goto raise_overflow;\
+            }\
+        }\
+        return (target_type) value;\
+    }
+
+/* CIntFromPy */
+static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const long neg_one = (long) -1, const_zero = (long) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(long) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (long) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(long, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 2 * PyLong_SHIFT) {
+                            return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 3 * PyLong_SHIFT) {
+                            return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 4 * PyLong_SHIFT) {
+                            return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (long) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(long) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(long, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(long,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(long) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(long) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            long val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (long) -1;
+        }
+    } else {
+        long val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (long) -1;
+        val = __Pyx_PyInt_As_long(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to long");
+    return (long) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to long");
+    return (long) -1;
+}
+
+/* CIntFromPy */
+static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const int neg_one = (int) -1, const_zero = (int) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(int) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (int) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(int, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 2 * PyLong_SHIFT) {
+                            return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 3 * PyLong_SHIFT) {
+                            return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 4 * PyLong_SHIFT) {
+                            return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (int) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(int) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(int, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(int,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(int) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(int) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            int val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (int) -1;
+        }
+    } else {
+        int val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (int) -1;
+        val = __Pyx_PyInt_As_int(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to int");
+    return (int) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to int");
+    return (int) -1;
+}
+
+/* FastTypeChecks */
+#if CYTHON_COMPILING_IN_CPYTHON
+static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) {
+    while (a) {
+        a = a->tp_base;
+        if (a == b)
+            return 1;
+    }
+    return b == &PyBaseObject_Type;
+}
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) {
+    PyObject *mro;
+    if (a == b) return 1;
+    mro = a->tp_mro;
+    if (likely(mro)) {
+        Py_ssize_t i, n;
+        n = PyTuple_GET_SIZE(mro);
+        for (i = 0; i < n; i++) {
+            if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b)
+                return 1;
+        }
+        return 0;
+    }
+    return __Pyx_InBases(a, b);
+}
+#if PY_MAJOR_VERSION == 2
+static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) {
+    PyObject *exception, *value, *tb;
+    int res;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&exception, &value, &tb);
+    res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0;
+    if (unlikely(res == -1)) {
+        PyErr_WriteUnraisable(err);
+        res = 0;
+    }
+    if (!res) {
+        res = PyObject_IsSubclass(err, exc_type2);
+        if (unlikely(res == -1)) {
+            PyErr_WriteUnraisable(err);
+            res = 0;
+        }
+    }
+    __Pyx_ErrRestore(exception, value, tb);
+    return res;
+}
+#else
+static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) {
+    int res = exc_type1 ? __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type1) : 0;
+    if (!res) {
+        res = __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2);
+    }
+    return res;
+}
+#endif
+static int __Pyx_PyErr_GivenExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    assert(PyExceptionClass_Check(exc_type));
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        PyObject *t = PyTuple_GET_ITEM(tuple, i);
+        #if PY_MAJOR_VERSION < 3
+        if (likely(exc_type == t)) return 1;
+        #endif
+        if (likely(PyExceptionClass_Check(t))) {
+            if (__Pyx_inner_PyErr_GivenExceptionMatches2(exc_type, NULL, t)) return 1;
+        } else {
+        }
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject* exc_type) {
+    if (likely(err == exc_type)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        if (likely(PyExceptionClass_Check(exc_type))) {
+            return __Pyx_inner_PyErr_GivenExceptionMatches2(err, NULL, exc_type);
+        } else if (likely(PyTuple_Check(exc_type))) {
+            return __Pyx_PyErr_GivenExceptionMatchesTuple(err, exc_type);
+        } else {
+        }
+    }
+    return PyErr_GivenExceptionMatches(err, exc_type);
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *exc_type1, PyObject *exc_type2) {
+    assert(PyExceptionClass_Check(exc_type1));
+    assert(PyExceptionClass_Check(exc_type2));
+    if (likely(err == exc_type1 || err == exc_type2)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        return __Pyx_inner_PyErr_GivenExceptionMatches2(err, exc_type1, exc_type2);
+    }
+    return (PyErr_GivenExceptionMatches(err, exc_type1) || PyErr_GivenExceptionMatches(err, exc_type2));
+}
+#endif
+
+/* CheckBinaryVersion */
+static int __Pyx_check_binary_version(void) {
+    char ctversion[4], rtversion[4];
+    PyOS_snprintf(ctversion, 4, "%d.%d", PY_MAJOR_VERSION, PY_MINOR_VERSION);
+    PyOS_snprintf(rtversion, 4, "%s", Py_GetVersion());
+    if (ctversion[0] != rtversion[0] || ctversion[2] != rtversion[2]) {
+        char message[200];
+        PyOS_snprintf(message, sizeof(message),
+                      "compiletime version %s of module '%.100s' "
+                      "does not match runtime version %s",
+                      ctversion, __Pyx_MODULE_NAME, rtversion);
+        return PyErr_WarnEx(NULL, message, 1);
+    }
+    return 0;
+}
+
+/* InitStrings */
+static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) {
+    while (t->p) {
+        #if PY_MAJOR_VERSION < 3
+        if (t->is_unicode) {
+            *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL);
+        } else if (t->intern) {
+            *t->p = PyString_InternFromString(t->s);
+        } else {
+            *t->p = PyString_FromStringAndSize(t->s, t->n - 1);
+        }
+        #else
+        if (t->is_unicode | t->is_str) {
+            if (t->intern) {
+                *t->p = PyUnicode_InternFromString(t->s);
+            } else if (t->encoding) {
+                *t->p = PyUnicode_Decode(t->s, t->n - 1, t->encoding, NULL);
+            } else {
+                *t->p = PyUnicode_FromStringAndSize(t->s, t->n - 1);
+            }
+        } else {
+            *t->p = PyBytes_FromStringAndSize(t->s, t->n - 1);
+        }
+        #endif
+        if (!*t->p)
+            return -1;
+        if (PyObject_Hash(*t->p) == -1)
+            return -1;
+        ++t;
+    }
+    return 0;
+}
+
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) {
+    return __Pyx_PyUnicode_FromStringAndSize(c_str, (Py_ssize_t)strlen(c_str));
+}
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) {
+    Py_ssize_t ignore;
+    return __Pyx_PyObject_AsStringAndSize(o, &ignore);
+}
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+#if !CYTHON_PEP393_ENABLED
+static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    char* defenc_c;
+    PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL);
+    if (!defenc) return NULL;
+    defenc_c = PyBytes_AS_STRING(defenc);
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    {
+        char* end = defenc_c + PyBytes_GET_SIZE(defenc);
+        char* c;
+        for (c = defenc_c; c < end; c++) {
+            if ((unsigned char) (*c) >= 128) {
+                PyUnicode_AsASCIIString(o);
+                return NULL;
+            }
+        }
+    }
+#endif
+    *length = PyBytes_GET_SIZE(defenc);
+    return defenc_c;
+}
+#else
+static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL;
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    if (likely(PyUnicode_IS_ASCII(o))) {
+        *length = PyUnicode_GET_LENGTH(o);
+        return PyUnicode_AsUTF8(o);
+    } else {
+        PyUnicode_AsASCIIString(o);
+        return NULL;
+    }
+#else
+    return PyUnicode_AsUTF8AndSize(o, length);
+#endif
+}
+#endif
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+    if (
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+            __Pyx_sys_getdefaultencoding_not_ascii &&
+#endif
+            PyUnicode_Check(o)) {
+        return __Pyx_PyUnicode_AsStringAndSize(o, length);
+    } else
+#endif
+#if (!CYTHON_COMPILING_IN_PYPY) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE))
+    if (PyByteArray_Check(o)) {
+        *length = PyByteArray_GET_SIZE(o);
+        return PyByteArray_AS_STRING(o);
+    } else
+#endif
+    {
+        char* result;
+        int r = PyBytes_AsStringAndSize(o, &result, length);
+        if (unlikely(r < 0)) {
+            return NULL;
+        } else {
+            return result;
+        }
+    }
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) {
+   int is_true = x == Py_True;
+   if (is_true | (x == Py_False) | (x == Py_None)) return is_true;
+   else return PyObject_IsTrue(x);
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject* x) {
+    int retval;
+    if (unlikely(!x)) return -1;
+    retval = __Pyx_PyObject_IsTrue(x);
+    Py_DECREF(x);
+    return retval;
+}
+static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) {
+#if PY_MAJOR_VERSION >= 3
+    if (PyLong_Check(result)) {
+        if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1,
+                "__int__ returned non-int (type %.200s).  "
+                "The ability to return an instance of a strict subclass of int "
+                "is deprecated, and may be removed in a future version of Python.",
+                Py_TYPE(result)->tp_name)) {
+            Py_DECREF(result);
+            return NULL;
+        }
+        return result;
+    }
+#endif
+    PyErr_Format(PyExc_TypeError,
+                 "__%.4s__ returned non-%.4s (type %.200s)",
+                 type_name, type_name, Py_TYPE(result)->tp_name);
+    Py_DECREF(result);
+    return NULL;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) {
+#if CYTHON_USE_TYPE_SLOTS
+  PyNumberMethods *m;
+#endif
+  const char *name = NULL;
+  PyObject *res = NULL;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_Check(x) || PyLong_Check(x)))
+#else
+  if (likely(PyLong_Check(x)))
+#endif
+    return __Pyx_NewRef(x);
+#if CYTHON_USE_TYPE_SLOTS
+  m = Py_TYPE(x)->tp_as_number;
+  #if PY_MAJOR_VERSION < 3
+  if (m && m->nb_int) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  else if (m && m->nb_long) {
+    name = "long";
+    res = m->nb_long(x);
+  }
+  #else
+  if (likely(m && m->nb_int)) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  #endif
+#else
+  if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) {
+    res = PyNumber_Int(x);
+  }
+#endif
+  if (likely(res)) {
+#if PY_MAJOR_VERSION < 3
+    if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) {
+#else
+    if (unlikely(!PyLong_CheckExact(res))) {
+#endif
+        return __Pyx_PyNumber_IntOrLongWrongResultType(res, name);
+    }
+  }
+  else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_TypeError,
+                    "an integer is required");
+  }
+  return res;
+}
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) {
+  Py_ssize_t ival;
+  PyObject *x;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_CheckExact(b))) {
+    if (sizeof(Py_ssize_t) >= sizeof(long))
+        return PyInt_AS_LONG(b);
+    else
+        return PyInt_AsSsize_t(b);
+  }
+#endif
+  if (likely(PyLong_CheckExact(b))) {
+    #if CYTHON_USE_PYLONG_INTERNALS
+    const digit* digits = ((PyLongObject*)b)->ob_digit;
+    const Py_ssize_t size = Py_SIZE(b);
+    if (likely(__Pyx_sst_abs(size) <= 1)) {
+        ival = likely(size) ? digits[0] : 0;
+        if (size == -1) ival = -ival;
+        return ival;
+    } else {
+      switch (size) {
+         case 2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+      }
+    }
+    #endif
+    return PyLong_AsSsize_t(b);
+  }
+  x = PyNumber_Index(b);
+  if (!x) return -1;
+  ival = PyInt_AsSsize_t(x);
+  Py_DECREF(x);
+  return ival;
+}
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b) {
+  return b ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False);
+}
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) {
+    return PyInt_FromSize_t(ival);
+}
+
+
+#endif /* Py_PYTHON_H */
diff --git a/python-package/insightface/thirdparty/face3d/mesh/__init__.py b/python-package/insightface/thirdparty/face3d/mesh/__init__.py
new file mode 100644
index 000000000..4e4329606
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/__init__.py
@@ -0,0 +1,15 @@
+#from __future__ import absolute_import
+#from cython import mesh_core_cython
+#import io
+#import vis
+#import transform
+#import light
+#import render
+
+from .cython import mesh_core_cython
+from . import io
+from . import vis
+from . import transform
+from . import light
+from . import render
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.cpp b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.cpp
new file mode 100644
index 000000000..aeea4dae8
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.cpp
@@ -0,0 +1,375 @@
+/*
+functions that can not be optimazed by vertorization in python.
+1. rasterization.(need process each triangle)
+2. normal of each vertex.(use one-ring, need process each vertex)
+3. write obj(seems that it can be verctorized? anyway, writing it in c++ is simple, so also add function here. --> however, why writting in c++ is still slow?)
+
+Author: Yao Feng 
+Mail: yaofeng1995@gmail.com
+*/
+
+#include "mesh_core.h"
+
+
+/* Judge whether the point is in the triangle
+Method:
+    http://blackpawn.com/texts/pointinpoly/
+Args:
+    point: [x, y] 
+    tri_points: three vertices(2d points) of a triangle. 2 coords x 3 vertices
+Returns:
+    bool: true for in triangle
+*/
+bool isPointInTri(point p, point p0, point p1, point p2)
+{   
+    // vectors
+    point v0, v1, v2;
+    v0 = p2 - p0;
+    v1 = p1 - p0;
+    v2 = p - p0;
+
+    // dot products
+    float dot00 = v0.dot(v0); //v0.x * v0.x + v0.y * v0.y //np.dot(v0.T, v0)
+    float dot01 = v0.dot(v1); //v0.x * v1.x + v0.y * v1.y //np.dot(v0.T, v1)
+    float dot02 = v0.dot(v2); //v0.x * v2.x + v0.y * v2.y //np.dot(v0.T, v2)
+    float dot11 = v1.dot(v1); //v1.x * v1.x + v1.y * v1.y //np.dot(v1.T, v1)
+    float dot12 = v1.dot(v2); //v1.x * v2.x + v1.y * v2.y//np.dot(v1.T, v2)
+
+    // barycentric coordinates
+    float inverDeno;
+    if(dot00*dot11 - dot01*dot01 == 0)
+        inverDeno = 0;
+    else
+        inverDeno = 1/(dot00*dot11 - dot01*dot01);
+
+    float u = (dot11*dot02 - dot01*dot12)*inverDeno;
+    float v = (dot00*dot12 - dot01*dot02)*inverDeno;
+
+    // check if point in triangle
+    return (u >= 0) && (v >= 0) && (u + v < 1);
+}
+
+
+void get_point_weight(float* weight, point p, point p0, point p1, point p2)
+{   
+    // vectors
+    point v0, v1, v2;
+    v0 = p2 - p0; 
+    v1 = p1 - p0; 
+    v2 = p - p0; 
+
+    // dot products
+    float dot00 = v0.dot(v0); //v0.x * v0.x + v0.y * v0.y //np.dot(v0.T, v0)
+    float dot01 = v0.dot(v1); //v0.x * v1.x + v0.y * v1.y //np.dot(v0.T, v1)
+    float dot02 = v0.dot(v2); //v0.x * v2.x + v0.y * v2.y //np.dot(v0.T, v2)
+    float dot11 = v1.dot(v1); //v1.x * v1.x + v1.y * v1.y //np.dot(v1.T, v1)
+    float dot12 = v1.dot(v2); //v1.x * v2.x + v1.y * v2.y//np.dot(v1.T, v2)
+
+    // barycentric coordinates
+    float inverDeno;
+    if(dot00*dot11 - dot01*dot01 == 0)
+        inverDeno = 0;
+    else
+        inverDeno = 1/(dot00*dot11 - dot01*dot01);
+
+    float u = (dot11*dot02 - dot01*dot12)*inverDeno;
+    float v = (dot00*dot12 - dot01*dot02)*inverDeno;
+
+    // weight
+    weight[0] = 1 - u - v;
+    weight[1] = v;
+    weight[2] = u;
+}
+
+
+void _get_normal_core(
+    float* normal, float* tri_normal, int* triangles,
+    int ntri)
+{
+    int i, j;
+    int tri_p0_ind, tri_p1_ind, tri_p2_ind;
+
+    for(i = 0; i < ntri; i++)
+    {
+        tri_p0_ind = triangles[3*i];
+        tri_p1_ind = triangles[3*i + 1];
+        tri_p2_ind = triangles[3*i + 2];
+
+        for(j = 0; j < 3; j++)
+        {
+            normal[3*tri_p0_ind + j] = normal[3*tri_p0_ind + j] + tri_normal[3*i + j];
+            normal[3*tri_p1_ind + j] = normal[3*tri_p1_ind + j] + tri_normal[3*i + j];
+            normal[3*tri_p2_ind + j] = normal[3*tri_p2_ind + j] + tri_normal[3*i + j];
+        }
+    }
+}
+
+
+void _rasterize_triangles_core(
+    float* vertices, int* triangles, 
+    float* depth_buffer, int* triangle_buffer, float* barycentric_weight,
+    int nver, int ntri,
+    int h, int w)
+{
+    int i;
+    int x, y, k;
+    int tri_p0_ind, tri_p1_ind, tri_p2_ind;
+    point p0, p1, p2, p;
+    int x_min, x_max, y_min, y_max;
+    float p_depth, p0_depth, p1_depth, p2_depth;
+    float weight[3];
+
+    for(i = 0; i < ntri; i++)
+    {
+        tri_p0_ind = triangles[3*i];
+        tri_p1_ind = triangles[3*i + 1];
+        tri_p2_ind = triangles[3*i + 2];
+
+        p0.x = vertices[3*tri_p0_ind]; p0.y = vertices[3*tri_p0_ind + 1]; p0_depth = vertices[3*tri_p0_ind + 2];
+        p1.x = vertices[3*tri_p1_ind]; p1.y = vertices[3*tri_p1_ind + 1]; p1_depth = vertices[3*tri_p1_ind + 2];
+        p2.x = vertices[3*tri_p2_ind]; p2.y = vertices[3*tri_p2_ind + 1]; p2_depth = vertices[3*tri_p2_ind + 2];
+        
+        x_min = max((int)ceil(min(p0.x, min(p1.x, p2.x))), 0);
+        x_max = min((int)floor(max(p0.x, max(p1.x, p2.x))), w - 1);
+      
+        y_min = max((int)ceil(min(p0.y, min(p1.y, p2.y))), 0);
+        y_max = min((int)floor(max(p0.y, max(p1.y, p2.y))), h - 1);
+
+        if(x_max < x_min || y_max < y_min)
+        {
+            continue;
+        }
+
+        for(y = y_min; y <= y_max; y++) //h
+        {
+            for(x = x_min; x <= x_max; x++) //w
+            {
+                p.x = x; p.y = y;
+                if(p.x < 2 || p.x > w - 3 || p.y < 2 || p.y > h - 3 || isPointInTri(p, p0, p1, p2))
+                {
+                    get_point_weight(weight, p, p0, p1, p2);
+                    p_depth = weight[0]*p0_depth + weight[1]*p1_depth + weight[2]*p2_depth;
+
+                    if((p_depth > depth_buffer[y*w + x]))
+                    {
+                        depth_buffer[y*w + x] = p_depth;
+                        triangle_buffer[y*w + x] = i;
+                        for(k = 0; k < 3; k++)
+                        {
+                            barycentric_weight[y*w*3 + x*3 + k] = weight[k];
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+
+void _render_colors_core(
+    float* image, float* vertices, int* triangles, 
+    float* colors, 
+    float* depth_buffer,
+    int nver, int ntri,
+    int h, int w, int c)
+{
+    int i;
+    int x, y, k;
+    int tri_p0_ind, tri_p1_ind, tri_p2_ind;
+    point p0, p1, p2, p;
+    int x_min, x_max, y_min, y_max;
+    float p_depth, p0_depth, p1_depth, p2_depth;
+    float p_color, p0_color, p1_color, p2_color;
+    float weight[3];
+
+    for(i = 0; i < ntri; i++)
+    {
+        tri_p0_ind = triangles[3*i];
+        tri_p1_ind = triangles[3*i + 1];
+        tri_p2_ind = triangles[3*i + 2];
+
+        p0.x = vertices[3*tri_p0_ind]; p0.y = vertices[3*tri_p0_ind + 1]; p0_depth = vertices[3*tri_p0_ind + 2];
+        p1.x = vertices[3*tri_p1_ind]; p1.y = vertices[3*tri_p1_ind + 1]; p1_depth = vertices[3*tri_p1_ind + 2];
+        p2.x = vertices[3*tri_p2_ind]; p2.y = vertices[3*tri_p2_ind + 1]; p2_depth = vertices[3*tri_p2_ind + 2];
+        
+        x_min = max((int)ceil(min(p0.x, min(p1.x, p2.x))), 0);
+        x_max = min((int)floor(max(p0.x, max(p1.x, p2.x))), w - 1);
+      
+        y_min = max((int)ceil(min(p0.y, min(p1.y, p2.y))), 0);
+        y_max = min((int)floor(max(p0.y, max(p1.y, p2.y))), h - 1);
+
+        if(x_max < x_min || y_max < y_min)
+        {
+            continue;
+        }
+
+        for(y = y_min; y <= y_max; y++) //h
+        {
+            for(x = x_min; x <= x_max; x++) //w
+            {
+                p.x = x; p.y = y;
+                if(p.x < 2 || p.x > w - 3 || p.y < 2 || p.y > h - 3 || isPointInTri(p, p0, p1, p2))
+                {
+                    get_point_weight(weight, p, p0, p1, p2);
+                    p_depth = weight[0]*p0_depth + weight[1]*p1_depth + weight[2]*p2_depth;
+
+                    if((p_depth > depth_buffer[y*w + x]))
+                    {
+                        for(k = 0; k < c; k++) // c
+                        {   
+                            p0_color = colors[c*tri_p0_ind + k];
+                            p1_color = colors[c*tri_p1_ind + k];
+                            p2_color = colors[c*tri_p2_ind + k]; 
+
+                            p_color = weight[0]*p0_color + weight[1]*p1_color + weight[2]*p2_color;
+                            image[y*w*c + x*c + k] = p_color;
+                        }
+
+                        depth_buffer[y*w + x] = p_depth;
+                    }
+                }
+            }
+        }
+    }
+}
+
+
+void _render_texture_core(
+    float* image, float* vertices, int* triangles, 
+    float* texture, float* tex_coords, int* tex_triangles, 
+    float* depth_buffer,
+    int nver, int tex_nver, int ntri, 
+    int h, int w, int c, 
+    int tex_h, int tex_w, int tex_c, 
+    int mapping_type)
+{
+    int i;
+    int x, y, k;
+    int tri_p0_ind, tri_p1_ind, tri_p2_ind;
+    int tex_tri_p0_ind, tex_tri_p1_ind, tex_tri_p2_ind;
+    point p0, p1, p2, p;
+    point tex_p0, tex_p1, tex_p2, tex_p;
+    int x_min, x_max, y_min, y_max;
+    float weight[3];
+    float p_depth, p0_depth, p1_depth, p2_depth;
+    float xd, yd;
+    float ul, ur, dl, dr;
+    for(i = 0; i < ntri; i++)
+    {
+        // mesh
+        tri_p0_ind = triangles[3*i];
+        tri_p1_ind = triangles[3*i + 1];
+        tri_p2_ind = triangles[3*i + 2];
+
+        p0.x = vertices[3*tri_p0_ind]; p0.y = vertices[3*tri_p0_ind + 1]; p0_depth = vertices[3*tri_p0_ind + 2];
+        p1.x = vertices[3*tri_p1_ind]; p1.y = vertices[3*tri_p1_ind + 1]; p1_depth = vertices[3*tri_p1_ind + 2];
+        p2.x = vertices[3*tri_p2_ind]; p2.y = vertices[3*tri_p2_ind + 1]; p2_depth = vertices[3*tri_p2_ind + 2];
+       
+        // texture
+        tex_tri_p0_ind = tex_triangles[3*i];
+        tex_tri_p1_ind = tex_triangles[3*i + 1];
+        tex_tri_p2_ind = tex_triangles[3*i + 2];
+
+        tex_p0.x = tex_coords[3*tex_tri_p0_ind]; tex_p0.y = tex_coords[3*tri_p0_ind + 1];
+        tex_p1.x = tex_coords[3*tex_tri_p1_ind]; tex_p1.y = tex_coords[3*tri_p1_ind + 1];
+        tex_p2.x = tex_coords[3*tex_tri_p2_ind]; tex_p2.y = tex_coords[3*tri_p2_ind + 1];
+
+
+        x_min = max((int)ceil(min(p0.x, min(p1.x, p2.x))), 0);
+        x_max = min((int)floor(max(p0.x, max(p1.x, p2.x))), w - 1);
+      
+        y_min = max((int)ceil(min(p0.y, min(p1.y, p2.y))), 0);
+        y_max = min((int)floor(max(p0.y, max(p1.y, p2.y))), h - 1);
+
+
+        if(x_max < x_min || y_max < y_min)
+        {
+            continue;
+        }
+
+        for(y = y_min; y <= y_max; y++) //h
+        {
+            for(x = x_min; x <= x_max; x++) //w
+            {
+                p.x = x; p.y = y;
+                if(p.x < 2 || p.x > w - 3 || p.y < 2 || p.y > h - 3 || isPointInTri(p, p0, p1, p2))
+                {
+                    get_point_weight(weight, p, p0, p1, p2);
+                    p_depth = weight[0]*p0_depth + weight[1]*p1_depth + weight[2]*p2_depth;
+                    
+                    if((p_depth > depth_buffer[y*w + x]))
+                    {
+                        // -- color from texture
+                        // cal weight in mesh tri
+                        get_point_weight(weight, p, p0, p1, p2);
+                        // cal coord in texture
+                        tex_p = tex_p0*weight[0] + tex_p1*weight[1] + tex_p2*weight[2];
+                        tex_p.x = max(min(tex_p.x, float(tex_w - 1)), float(0)); 
+                        tex_p.y = max(min(tex_p.y, float(tex_h - 1)), float(0)); 
+
+                        yd = tex_p.y - floor(tex_p.y);
+                        xd = tex_p.x - floor(tex_p.x);
+                        for(k = 0; k < c; k++)
+                        {
+                            if(mapping_type==0)// nearest
+                            {   
+                                image[y*w*c + x*c + k] = texture[int(round(tex_p.y))*tex_w*tex_c + int(round(tex_p.x))*tex_c + k];
+                            }
+                            else//bilinear interp
+                            { 
+                                ul = texture[(int)floor(tex_p.y)*tex_w*tex_c + (int)floor(tex_p.x)*tex_c + k];
+                                ur = texture[(int)floor(tex_p.y)*tex_w*tex_c + (int)ceil(tex_p.x)*tex_c + k];
+                                dl = texture[(int)ceil(tex_p.y)*tex_w*tex_c + (int)floor(tex_p.x)*tex_c + k];
+                                dr = texture[(int)ceil(tex_p.y)*tex_w*tex_c + (int)ceil(tex_p.x)*tex_c + k];
+
+                                image[y*w*c + x*c + k] = ul*(1-xd)*(1-yd) + ur*xd*(1-yd) + dl*(1-xd)*yd + dr*xd*yd;
+                            }
+
+                        }
+
+                        depth_buffer[y*w + x] = p_depth;
+                    } 
+                }
+            }
+        }
+    }
+}
+
+
+
+// ------------------------------------------------- write
+// obj write
+// Ref: https://github.com/patrikhuber/eos/blob/master/include/eos/core/Mesh.hpp
+void _write_obj_with_colors_texture(string filename, string mtl_name, 
+    float* vertices, int* triangles, float* colors, float* uv_coords,
+    int nver, int ntri, int ntexver)
+{
+    int i;
+
+    ofstream obj_file(filename.c_str());
+
+    // first line of the obj file: the mtl name
+    obj_file << "mtllib " << mtl_name << endl;
+    
+    // write vertices 
+    for (i = 0; i < nver; ++i) 
+    {
+        obj_file << "v " << vertices[3*i] << " " << vertices[3*i + 1] << " " << vertices[3*i + 2] << colors[3*i] << " " << colors[3*i + 1] << " " << colors[3*i + 2] <<  endl;
+    }
+
+    // write uv coordinates
+    for (i = 0; i < ntexver; ++i) 
+    {
+        //obj_file << "vt " << uv_coords[2*i] << " " << (1 - uv_coords[2*i + 1]) << endl;
+        obj_file << "vt " << uv_coords[2*i] << " " << uv_coords[2*i + 1] << endl;
+    }
+
+    obj_file << "usemtl FaceTexture" << endl;
+    // write triangles
+    for (i = 0; i < ntri; ++i) 
+    {
+        // obj_file << "f " << triangles[3*i] << "/" << triangles[3*i] << " " << triangles[3*i + 1] << "/" << triangles[3*i + 1] << " " << triangles[3*i + 2] << "/" << triangles[3*i + 2] << endl;
+        obj_file << "f " << triangles[3*i + 2] << "/" << triangles[3*i + 2] << " " << triangles[3*i + 1] << "/" << triangles[3*i + 1] << " " << triangles[3*i] << "/" << triangles[3*i] << endl;
+    }
+
+}
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.h b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.h
new file mode 100644
index 000000000..3eb00290f
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core.h
@@ -0,0 +1,83 @@
+#ifndef MESH_CORE_HPP_
+#define MESH_CORE_HPP_
+
+#include <stdio.h>
+#include <cmath>
+#include <algorithm>  
+#include <string>
+#include <iostream>
+#include <fstream>
+
+using namespace std;
+
+class point
+{
+ public:
+    float x;
+    float y;
+
+    float dot(point p)
+    {
+        return this->x * p.x + this->y * p.y;
+    }
+
+    point operator-(const point& p)
+    {
+        point np;
+        np.x = this->x - p.x;
+        np.y = this->y - p.y;
+        return np;
+    }
+
+    point operator+(const point& p)
+    {
+        point np;
+        np.x = this->x + p.x;
+        np.y = this->y + p.y;
+        return np;
+    }
+
+    point operator*(float s)
+    {
+        point np;
+        np.x = s * this->x;
+        np.y = s * this->y;
+        return np;
+    }
+}; 
+
+
+bool isPointInTri(point p, point p0, point p1, point p2, int h, int w);
+void get_point_weight(float* weight, point p, point p0, point p1, point p2);
+
+void _get_normal_core(
+    float* normal, float* tri_normal, int* triangles,
+    int ntri);
+
+void _rasterize_triangles_core(
+    float* vertices, int* triangles, 
+    float* depth_buffer, int* triangle_buffer, float* barycentric_weight,
+    int nver, int ntri,
+    int h, int w);
+
+void _render_colors_core(
+    float* image, float* vertices, int* triangles, 
+    float* colors, 
+    float* depth_buffer,
+    int nver, int ntri,
+    int h, int w, int c);
+
+void _render_texture_core(
+    float* image, float* vertices, int* triangles, 
+    float* texture, float* tex_coords, int* tex_triangles, 
+    float* depth_buffer,
+    int nver, int tex_nver, int ntri, 
+    int h, int w, int c, 
+    int tex_h, int tex_w, int tex_c, 
+    int mapping_type);
+
+void _write_obj_with_colors_texture(string filename, string mtl_name, 
+    float* vertices, int* triangles, float* colors, float* uv_coords,
+    int nver, int ntri, int ntexver);
+
+#endif
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.c b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.c
new file mode 100644
index 000000000..e113ac54a
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.c
@@ -0,0 +1,9091 @@
+/* Generated by Cython 0.28.2 */
+
+/* BEGIN: Cython Metadata
+{
+    "distutils": {
+        "depends": [
+            "insightface/thirdparty/face3d/mesh/cython/mesh_core.h"
+        ],
+        "include_dirs": [
+            "insightface/thirdparty/face3d/mesh/cython"
+        ],
+        "name": "mesh_core_cython",
+        "sources": [
+            "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx"
+        ]
+    },
+    "module_name": "mesh_core_cython"
+}
+END: Cython Metadata */
+
+#define PY_SSIZE_T_CLEAN
+#include "Python.h"
+#ifndef Py_PYTHON_H
+    #error Python headers needed to compile C extensions, please install development version of Python.
+#elif PY_VERSION_HEX < 0x02060000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000)
+    #error Cython requires Python 2.6+ or Python 3.3+.
+#else
+#define CYTHON_ABI "0_28_2"
+#define CYTHON_FUTURE_DIVISION 0
+#include <stddef.h>
+#ifndef offsetof
+  #define offsetof(type, member) ( (size_t) & ((type*)0) -> member )
+#endif
+#if !defined(WIN32) && !defined(MS_WINDOWS)
+  #ifndef __stdcall
+    #define __stdcall
+  #endif
+  #ifndef __cdecl
+    #define __cdecl
+  #endif
+  #ifndef __fastcall
+    #define __fastcall
+  #endif
+#endif
+#ifndef DL_IMPORT
+  #define DL_IMPORT(t) t
+#endif
+#ifndef DL_EXPORT
+  #define DL_EXPORT(t) t
+#endif
+#define __PYX_COMMA ,
+#ifndef HAVE_LONG_LONG
+  #if PY_VERSION_HEX >= 0x02070000
+    #define HAVE_LONG_LONG
+  #endif
+#endif
+#ifndef PY_LONG_LONG
+  #define PY_LONG_LONG LONG_LONG
+#endif
+#ifndef Py_HUGE_VAL
+  #define Py_HUGE_VAL HUGE_VAL
+#endif
+#ifdef PYPY_VERSION
+  #define CYTHON_COMPILING_IN_PYPY 1
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #undef CYTHON_USE_TYPE_SLOTS
+  #define CYTHON_USE_TYPE_SLOTS 0
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #if PY_VERSION_HEX < 0x03050000
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_INTERNALS
+  #define CYTHON_USE_UNICODE_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #undef CYTHON_AVOID_BORROWED_REFS
+  #define CYTHON_AVOID_BORROWED_REFS 1
+  #undef CYTHON_ASSUME_SAFE_MACROS
+  #define CYTHON_ASSUME_SAFE_MACROS 0
+  #undef CYTHON_UNPACK_METHODS
+  #define CYTHON_UNPACK_METHODS 0
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+#elif defined(PYSTON_VERSION)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 1
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #undef CYTHON_USE_ASYNC_SLOTS
+  #define CYTHON_USE_ASYNC_SLOTS 0
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+#else
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 1
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYTYPE_LOOKUP
+    #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #elif !defined(CYTHON_USE_PYTYPE_LOOKUP)
+    #define CYTHON_USE_PYTYPE_LOOKUP 1
+  #endif
+  #if PY_MAJOR_VERSION < 3
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYLONG_INTERNALS
+    #define CYTHON_USE_PYLONG_INTERNALS 0
+  #elif !defined(CYTHON_USE_PYLONG_INTERNALS)
+    #define CYTHON_USE_PYLONG_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_PYLIST_INTERNALS
+    #define CYTHON_USE_PYLIST_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #if PY_VERSION_HEX < 0x030300F0
+    #undef CYTHON_USE_UNICODE_WRITER
+    #define CYTHON_USE_UNICODE_WRITER 0
+  #elif !defined(CYTHON_USE_UNICODE_WRITER)
+    #define CYTHON_USE_UNICODE_WRITER 1
+  #endif
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #ifndef CYTHON_FAST_THREAD_STATE
+    #define CYTHON_FAST_THREAD_STATE 1
+  #endif
+  #ifndef CYTHON_FAST_PYCALL
+    #define CYTHON_FAST_PYCALL 1
+  #endif
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT (0 && PY_VERSION_HEX >= 0x03050000)
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1)
+  #endif
+#endif
+#if !defined(CYTHON_FAST_PYCCALL)
+#define CYTHON_FAST_PYCCALL  (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1)
+#endif
+#if CYTHON_USE_PYLONG_INTERNALS
+  #include "longintrepr.h"
+  #undef SHIFT
+  #undef BASE
+  #undef MASK
+#endif
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+#ifndef __has_cpp_attribute
+  #define __has_cpp_attribute(x) 0
+#endif
+#ifndef CYTHON_RESTRICT
+  #if defined(__GNUC__)
+    #define CYTHON_RESTRICT __restrict__
+  #elif defined(_MSC_VER) && _MSC_VER >= 1400
+    #define CYTHON_RESTRICT __restrict
+  #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define CYTHON_RESTRICT restrict
+  #else
+    #define CYTHON_RESTRICT
+  #endif
+#endif
+#ifndef CYTHON_UNUSED
+# if defined(__GNUC__)
+#   if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4))
+#     define CYTHON_UNUSED __attribute__ ((__unused__))
+#   else
+#     define CYTHON_UNUSED
+#   endif
+# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER))
+#   define CYTHON_UNUSED __attribute__ ((__unused__))
+# else
+#   define CYTHON_UNUSED
+# endif
+#endif
+#ifndef CYTHON_MAYBE_UNUSED_VAR
+#  if defined(__cplusplus)
+     template<class T> void CYTHON_MAYBE_UNUSED_VAR( const T& ) { }
+#  else
+#    define CYTHON_MAYBE_UNUSED_VAR(x) (void)(x)
+#  endif
+#endif
+#ifndef CYTHON_NCP_UNUSED
+# if CYTHON_COMPILING_IN_CPYTHON
+#  define CYTHON_NCP_UNUSED
+# else
+#  define CYTHON_NCP_UNUSED CYTHON_UNUSED
+# endif
+#endif
+#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None)
+#ifdef _MSC_VER
+    #ifndef _MSC_STDINT_H_
+        #if _MSC_VER < 1300
+           typedef unsigned char     uint8_t;
+           typedef unsigned int      uint32_t;
+        #else
+           typedef unsigned __int8   uint8_t;
+           typedef unsigned __int32  uint32_t;
+        #endif
+    #endif
+#else
+   #include <stdint.h>
+#endif
+#ifndef CYTHON_FALLTHROUGH
+  #if defined(__cplusplus) && __cplusplus >= 201103L
+    #if __has_cpp_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH [[fallthrough]]
+    #elif __has_cpp_attribute(clang::fallthrough)
+      #define CYTHON_FALLTHROUGH [[clang::fallthrough]]
+    #elif __has_cpp_attribute(gnu::fallthrough)
+      #define CYTHON_FALLTHROUGH [[gnu::fallthrough]]
+    #endif
+  #endif
+  #ifndef CYTHON_FALLTHROUGH
+    #if __has_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH __attribute__((fallthrough))
+    #else
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+  #if defined(__clang__ ) && defined(__apple_build_version__)
+    #if __apple_build_version__ < 7000000
+      #undef  CYTHON_FALLTHROUGH
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+#endif
+
+#ifndef CYTHON_INLINE
+  #if defined(__clang__)
+    #define CYTHON_INLINE __inline__ __attribute__ ((__unused__))
+  #elif defined(__GNUC__)
+    #define CYTHON_INLINE __inline__
+  #elif defined(_MSC_VER)
+    #define CYTHON_INLINE __inline
+  #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define CYTHON_INLINE inline
+  #else
+    #define CYTHON_INLINE
+  #endif
+#endif
+
+#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag)
+  #define Py_OptimizeFlag 0
+#endif
+#define __PYX_BUILD_PY_SSIZE_T "n"
+#define CYTHON_FORMAT_SSIZE_T "z"
+#if PY_MAJOR_VERSION < 3
+  #define __Pyx_BUILTIN_MODULE_NAME "__builtin__"
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+  #define __Pyx_DefaultClassType PyClass_Type
+#else
+  #define __Pyx_BUILTIN_MODULE_NAME "builtins"
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+  #define __Pyx_DefaultClassType PyType_Type
+#endif
+#ifndef Py_TPFLAGS_CHECKTYPES
+  #define Py_TPFLAGS_CHECKTYPES 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_INDEX
+  #define Py_TPFLAGS_HAVE_INDEX 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_NEWBUFFER
+  #define Py_TPFLAGS_HAVE_NEWBUFFER 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_FINALIZE
+  #define Py_TPFLAGS_HAVE_FINALIZE 0
+#endif
+#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL)
+  #ifndef METH_FASTCALL
+     #define METH_FASTCALL 0x80
+  #endif
+  typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs);
+  typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args,
+                                                          Py_ssize_t nargs, PyObject *kwnames);
+#else
+  #define __Pyx_PyCFunctionFast _PyCFunctionFast
+  #define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords
+#endif
+#if CYTHON_FAST_PYCCALL
+#define __Pyx_PyFastCFunction_Check(func)\
+    ((PyCFunction_Check(func) && (METH_FASTCALL == (PyCFunction_GET_FLAGS(func) & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS)))))
+#else
+#define __Pyx_PyFastCFunction_Check(func) 0
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc)
+  #define PyObject_Malloc(s)   PyMem_Malloc(s)
+  #define PyObject_Free(p)     PyMem_Free(p)
+  #define PyObject_Realloc(p)  PyMem_Realloc(p)
+#endif
+#if CYTHON_COMPILING_IN_PYSTON
+  #define __Pyx_PyCode_HasFreeVars(co)  PyCode_HasFreeVars(co)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno) PyFrame_SetLineNumber(frame, lineno)
+#else
+  #define __Pyx_PyCode_HasFreeVars(co)  (PyCode_GetNumFree(co) > 0)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno)  (frame)->f_lineno = (lineno)
+#endif
+#if !CYTHON_FAST_THREAD_STATE || PY_VERSION_HEX < 0x02070000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#elif PY_VERSION_HEX >= 0x03060000
+  #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet()
+#elif PY_VERSION_HEX >= 0x03000000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#else
+  #define __Pyx_PyThreadState_Current _PyThreadState_Current
+#endif
+#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT)
+#include "pythread.h"
+#define Py_tss_NEEDS_INIT 0
+typedef int Py_tss_t;
+static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) {
+  *key = PyThread_create_key();
+  return 0; // PyThread_create_key reports success always
+}
+static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) {
+  Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t));
+  *key = Py_tss_NEEDS_INIT;
+  return key;
+}
+static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) {
+  PyObject_Free(key);
+}
+static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) {
+  return *key != Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) {
+  PyThread_delete_key(*key);
+  *key = Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) {
+  return PyThread_set_key_value(*key, value);
+}
+static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) {
+  return PyThread_get_key_value(*key);
+}
+#endif // TSS (Thread Specific Storage) API
+#if CYTHON_COMPILING_IN_CPYTHON || defined(_PyDict_NewPresized)
+#define __Pyx_PyDict_NewPresized(n)  ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n))
+#else
+#define __Pyx_PyDict_NewPresized(n)  PyDict_New()
+#endif
+#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_TrueDivide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceTrueDivide(x,y)
+#else
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_Divide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceDivide(x,y)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && CYTHON_USE_UNICODE_INTERNALS
+#define __Pyx_PyDict_GetItemStr(dict, name)  _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash)
+#else
+#define __Pyx_PyDict_GetItemStr(dict, name)  PyDict_GetItem(dict, name)
+#endif
+#if PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND)
+  #define CYTHON_PEP393_ENABLED 1
+  #define __Pyx_PyUnicode_READY(op)       (likely(PyUnicode_IS_READY(op)) ?\
+                                              0 : _PyUnicode_Ready((PyObject *)(op)))
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_LENGTH(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i)
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   PyUnicode_MAX_CHAR_VALUE(u)
+  #define __Pyx_PyUnicode_KIND(u)         PyUnicode_KIND(u)
+  #define __Pyx_PyUnicode_DATA(u)         PyUnicode_DATA(u)
+  #define __Pyx_PyUnicode_READ(k, d, i)   PyUnicode_READ(k, d, i)
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  PyUnicode_WRITE(k, d, i, ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u)))
+#else
+  #define CYTHON_PEP393_ENABLED 0
+  #define PyUnicode_1BYTE_KIND  1
+  #define PyUnicode_2BYTE_KIND  2
+  #define PyUnicode_4BYTE_KIND  4
+  #define __Pyx_PyUnicode_READY(op)       (0)
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_SIZE(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i]))
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   ((sizeof(Py_UNICODE) == 2) ? 65535 : 1114111)
+  #define __Pyx_PyUnicode_KIND(u)         (sizeof(Py_UNICODE))
+  #define __Pyx_PyUnicode_DATA(u)         ((void*)PyUnicode_AS_UNICODE(u))
+  #define __Pyx_PyUnicode_READ(k, d, i)   ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i]))
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  (((void)(k)), ((Py_UNICODE*)d)[i] = ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_SIZE(u))
+#endif
+#if CYTHON_COMPILING_IN_PYPY
+  #define __Pyx_PyUnicode_Concat(a, b)      PyNumber_Add(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  PyNumber_Add(a, b)
+#else
+  #define __Pyx_PyUnicode_Concat(a, b)      PyUnicode_Concat(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\
+      PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b))
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyUnicode_Contains)
+  #define PyUnicode_Contains(u, s)  PySequence_Contains(u, s)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyByteArray_Check)
+  #define PyByteArray_Check(obj)  PyObject_TypeCheck(obj, &PyByteArray_Type)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Format)
+  #define PyObject_Format(obj, fmt)  PyObject_CallMethod(obj, "__format__", "O", fmt)
+#endif
+#define __Pyx_PyString_FormatSafe(a, b)   ((unlikely((a) == Py_None)) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b))
+#define __Pyx_PyUnicode_FormatSafe(a, b)  ((unlikely((a) == Py_None)) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b))
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyString_Format(a, b)  PyUnicode_Format(a, b)
+#else
+  #define __Pyx_PyString_Format(a, b)  PyString_Format(a, b)
+#endif
+#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII)
+  #define PyObject_ASCII(o)            PyObject_Repr(o)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBaseString_Type            PyUnicode_Type
+  #define PyStringObject               PyUnicodeObject
+  #define PyString_Type                PyUnicode_Type
+  #define PyString_Check               PyUnicode_Check
+  #define PyString_CheckExact          PyUnicode_CheckExact
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj)
+  #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj)
+#else
+  #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj))
+  #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj))
+#endif
+#ifndef PySet_CheckExact
+  #define PySet_CheckExact(obj)        (Py_TYPE(obj) == &PySet_Type)
+#endif
+#if CYTHON_ASSUME_SAFE_MACROS
+  #define __Pyx_PySequence_SIZE(seq)  Py_SIZE(seq)
+#else
+  #define __Pyx_PySequence_SIZE(seq)  PySequence_Size(seq)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyIntObject                  PyLongObject
+  #define PyInt_Type                   PyLong_Type
+  #define PyInt_Check(op)              PyLong_Check(op)
+  #define PyInt_CheckExact(op)         PyLong_CheckExact(op)
+  #define PyInt_FromString             PyLong_FromString
+  #define PyInt_FromUnicode            PyLong_FromUnicode
+  #define PyInt_FromLong               PyLong_FromLong
+  #define PyInt_FromSize_t             PyLong_FromSize_t
+  #define PyInt_FromSsize_t            PyLong_FromSsize_t
+  #define PyInt_AsLong                 PyLong_AsLong
+  #define PyInt_AS_LONG                PyLong_AS_LONG
+  #define PyInt_AsSsize_t              PyLong_AsSsize_t
+  #define PyInt_AsUnsignedLongMask     PyLong_AsUnsignedLongMask
+  #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask
+  #define PyNumber_Int                 PyNumber_Long
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBoolObject                 PyLongObject
+#endif
+#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY
+  #ifndef PyUnicode_InternFromString
+    #define PyUnicode_InternFromString(s) PyUnicode_FromString(s)
+  #endif
+#endif
+#if PY_VERSION_HEX < 0x030200A4
+  typedef long Py_hash_t;
+  #define __Pyx_PyInt_FromHash_t PyInt_FromLong
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsLong
+#else
+  #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t
+  #define __Pyx_PyInt_AsHash_t   PyInt_AsSsize_t
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyMethod_New(func, self, klass) ((self) ? PyMethod_New(func, self) : (Py_INCREF(func), func))
+#else
+  #define __Pyx_PyMethod_New(func, self, klass) PyMethod_New(func, self, klass)
+#endif
+#if CYTHON_USE_ASYNC_SLOTS
+  #if PY_VERSION_HEX >= 0x030500B1
+    #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods
+    #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async)
+  #else
+    #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved))
+  #endif
+#else
+  #define __Pyx_PyType_AsAsync(obj) NULL
+#endif
+#ifndef __Pyx_PyAsyncMethodsStruct
+    typedef struct {
+        unaryfunc am_await;
+        unaryfunc am_aiter;
+        unaryfunc am_anext;
+    } __Pyx_PyAsyncMethodsStruct;
+#endif
+
+#if defined(WIN32) || defined(MS_WINDOWS)
+  #define _USE_MATH_DEFINES
+#endif
+#include <math.h>
+#ifdef NAN
+#define __PYX_NAN() ((float) NAN)
+#else
+static CYTHON_INLINE float __PYX_NAN() {
+  float value;
+  memset(&value, 0xFF, sizeof(value));
+  return value;
+}
+#endif
+#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL)
+#define __Pyx_truncl trunc
+#else
+#define __Pyx_truncl truncl
+#endif
+
+
+#define __PYX_ERR(f_index, lineno, Ln_error) \
+{ \
+  __pyx_filename = __pyx_f[f_index]; __pyx_lineno = lineno; __pyx_clineno = __LINE__; goto Ln_error; \
+}
+
+#ifndef __PYX_EXTERN_C
+  #ifdef __cplusplus
+    #define __PYX_EXTERN_C extern "C"
+  #else
+    #define __PYX_EXTERN_C extern
+  #endif
+#endif
+
+#define __PYX_HAVE__mesh_core_cython
+#define __PYX_HAVE_API__mesh_core_cython
+/* Early includes */
+#include <string.h>
+#include <stdio.h>
+#include "numpy/arrayobject.h"
+#include "numpy/ufuncobject.h"
+#include "ios"
+#include "new"
+#include "stdexcept"
+#include "typeinfo"
+#include <string>
+#include "mesh_core.h"
+#ifdef _OPENMP
+#include <omp.h>
+#endif /* _OPENMP */
+
+#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS)
+#define CYTHON_WITHOUT_ASSERTIONS
+#endif
+
+typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding;
+                const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry;
+
+#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT 0
+#define __PYX_DEFAULT_STRING_ENCODING ""
+#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString
+#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#define __Pyx_uchar_cast(c) ((unsigned char)c)
+#define __Pyx_long_cast(x) ((long)x)
+#define __Pyx_fits_Py_ssize_t(v, type, is_signed)  (\
+    (sizeof(type) < sizeof(Py_ssize_t))  ||\
+    (sizeof(type) > sizeof(Py_ssize_t) &&\
+          likely(v < (type)PY_SSIZE_T_MAX ||\
+                 v == (type)PY_SSIZE_T_MAX)  &&\
+          (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\
+                                v == (type)PY_SSIZE_T_MIN)))  ||\
+    (sizeof(type) == sizeof(Py_ssize_t) &&\
+          (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\
+                               v == (type)PY_SSIZE_T_MAX)))  )
+#if defined (__cplusplus) && __cplusplus >= 201103L
+    #include <cstdlib>
+    #define __Pyx_sst_abs(value) std::abs(value)
+#elif SIZEOF_INT >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) abs(value)
+#elif SIZEOF_LONG >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) labs(value)
+#elif defined (_MSC_VER)
+    #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value))
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define __Pyx_sst_abs(value) llabs(value)
+#elif defined (__GNUC__)
+    #define __Pyx_sst_abs(value) __builtin_llabs(value)
+#else
+    #define __Pyx_sst_abs(value) ((value<0) ? -value : value)
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*);
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length);
+#define __Pyx_PyByteArray_FromString(s) PyByteArray_FromStringAndSize((const char*)s, strlen((const char*)s))
+#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l)
+#define __Pyx_PyBytes_FromString        PyBytes_FromString
+#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*);
+#if PY_MAJOR_VERSION < 3
+    #define __Pyx_PyStr_FromString        __Pyx_PyBytes_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#else
+    #define __Pyx_PyStr_FromString        __Pyx_PyUnicode_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize
+#endif
+#define __Pyx_PyBytes_AsWritableString(s)     ((char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableSString(s)    ((signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableUString(s)    ((unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsString(s)     ((const char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsSString(s)    ((const signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsUString(s)    ((const unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyObject_AsWritableString(s)    ((char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableSString(s)    ((signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableUString(s)    ((unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsSString(s)    ((const signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsUString(s)    ((const unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_FromCString(s)  __Pyx_PyObject_FromString((const char*)s)
+#define __Pyx_PyBytes_FromCString(s)   __Pyx_PyBytes_FromString((const char*)s)
+#define __Pyx_PyByteArray_FromCString(s)   __Pyx_PyByteArray_FromString((const char*)s)
+#define __Pyx_PyStr_FromCString(s)     __Pyx_PyStr_FromString((const char*)s)
+#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s)
+static CYTHON_INLINE size_t __Pyx_Py_UNICODE_strlen(const Py_UNICODE *u) {
+    const Py_UNICODE *u_end = u;
+    while (*u_end++) ;
+    return (size_t)(u_end - u - 1);
+}
+#define __Pyx_PyUnicode_FromUnicode(u)       PyUnicode_FromUnicode(u, __Pyx_Py_UNICODE_strlen(u))
+#define __Pyx_PyUnicode_FromUnicodeAndLength PyUnicode_FromUnicode
+#define __Pyx_PyUnicode_AsUnicode            PyUnicode_AsUnicode
+#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj)
+#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None)
+#define __Pyx_PyBool_FromLong(b) ((b) ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False))
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*);
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x);
+#define __Pyx_PySequence_Tuple(obj)\
+    (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj))
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*);
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t);
+#if CYTHON_ASSUME_SAFE_MACROS
+#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x))
+#else
+#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x)
+#endif
+#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x))
+#if PY_MAJOR_VERSION >= 3
+#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x))
+#else
+#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x))
+#endif
+#define __Pyx_PyNumber_Float(x) (PyFloat_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Float(x))
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+static int __Pyx_sys_getdefaultencoding_not_ascii;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    PyObject* ascii_chars_u = NULL;
+    PyObject* ascii_chars_b = NULL;
+    const char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    if (strcmp(default_encoding_c, "ascii") == 0) {
+        __Pyx_sys_getdefaultencoding_not_ascii = 0;
+    } else {
+        char ascii_chars[128];
+        int c;
+        for (c = 0; c < 128; c++) {
+            ascii_chars[c] = c;
+        }
+        __Pyx_sys_getdefaultencoding_not_ascii = 1;
+        ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL);
+        if (!ascii_chars_u) goto bad;
+        ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL);
+        if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.",
+                default_encoding_c);
+            goto bad;
+        }
+        Py_DECREF(ascii_chars_u);
+        Py_DECREF(ascii_chars_b);
+    }
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    Py_XDECREF(ascii_chars_u);
+    Py_XDECREF(ascii_chars_b);
+    return -1;
+}
+#endif
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL)
+#else
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL)
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+static char* __PYX_DEFAULT_STRING_ENCODING;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c));
+    if (!__PYX_DEFAULT_STRING_ENCODING) goto bad;
+    strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c);
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    return -1;
+}
+#endif
+#endif
+
+
+/* Test for GCC > 2.95 */
+#if defined(__GNUC__)     && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95)))
+  #define likely(x)   __builtin_expect(!!(x), 1)
+  #define unlikely(x) __builtin_expect(!!(x), 0)
+#else /* !__GNUC__ or GCC < 2.95 */
+  #define likely(x)   (x)
+  #define unlikely(x) (x)
+#endif /* __GNUC__ */
+static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; }
+
+static PyObject *__pyx_m = NULL;
+static PyObject *__pyx_d;
+static PyObject *__pyx_b;
+static PyObject *__pyx_cython_runtime;
+static PyObject *__pyx_empty_tuple;
+static PyObject *__pyx_empty_bytes;
+static PyObject *__pyx_empty_unicode;
+static int __pyx_lineno;
+static int __pyx_clineno = 0;
+static const char * __pyx_cfilenm= __FILE__;
+static const char *__pyx_filename;
+
+/* Header.proto */
+#if !defined(CYTHON_CCOMPLEX)
+  #if defined(__cplusplus)
+    #define CYTHON_CCOMPLEX 1
+  #elif defined(_Complex_I)
+    #define CYTHON_CCOMPLEX 1
+  #else
+    #define CYTHON_CCOMPLEX 0
+  #endif
+#endif
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #include <complex>
+  #else
+    #include <complex.h>
+  #endif
+#endif
+#if CYTHON_CCOMPLEX && !defined(__cplusplus) && defined(__sun__) && defined(__GNUC__)
+  #undef _Complex_I
+  #define _Complex_I 1.0fj
+#endif
+
+
+static const char *__pyx_f[] = {
+  "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx",
+  "__init__.pxd",
+  "stringsource",
+  "type.pxd",
+};
+/* BufferFormatStructs.proto */
+#define IS_UNSIGNED(type) (((type) -1) > 0)
+struct __Pyx_StructField_;
+#define __PYX_BUF_FLAGS_PACKED_STRUCT (1 << 0)
+typedef struct {
+  const char* name;
+  struct __Pyx_StructField_* fields;
+  size_t size;
+  size_t arraysize[8];
+  int ndim;
+  char typegroup;
+  char is_unsigned;
+  int flags;
+} __Pyx_TypeInfo;
+typedef struct __Pyx_StructField_ {
+  __Pyx_TypeInfo* type;
+  const char* name;
+  size_t offset;
+} __Pyx_StructField;
+typedef struct {
+  __Pyx_StructField* field;
+  size_t parent_offset;
+} __Pyx_BufFmt_StackElem;
+typedef struct {
+  __Pyx_StructField root;
+  __Pyx_BufFmt_StackElem* head;
+  size_t fmt_offset;
+  size_t new_count, enc_count;
+  size_t struct_alignment;
+  int is_complex;
+  char enc_type;
+  char new_packmode;
+  char enc_packmode;
+  char is_valid_array;
+} __Pyx_BufFmt_Context;
+
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":730
+ * # in Cython to enable them only on the right systems.
+ * 
+ * ctypedef npy_int8       int8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ */
+typedef npy_int8 __pyx_t_5numpy_int8_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":731
+ * 
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t
+ */
+typedef npy_int16 __pyx_t_5numpy_int16_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":732
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int64      int64_t
+ * #ctypedef npy_int96      int96_t
+ */
+typedef npy_int32 __pyx_t_5numpy_int32_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":733
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_int96      int96_t
+ * #ctypedef npy_int128     int128_t
+ */
+typedef npy_int64 __pyx_t_5numpy_int64_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":737
+ * #ctypedef npy_int128     int128_t
+ * 
+ * ctypedef npy_uint8      uint8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ */
+typedef npy_uint8 __pyx_t_5numpy_uint8_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":738
+ * 
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t
+ */
+typedef npy_uint16 __pyx_t_5numpy_uint16_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":739
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint64     uint64_t
+ * #ctypedef npy_uint96     uint96_t
+ */
+typedef npy_uint32 __pyx_t_5numpy_uint32_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":740
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_uint96     uint96_t
+ * #ctypedef npy_uint128    uint128_t
+ */
+typedef npy_uint64 __pyx_t_5numpy_uint64_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":744
+ * #ctypedef npy_uint128    uint128_t
+ * 
+ * ctypedef npy_float32    float32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_float64    float64_t
+ * #ctypedef npy_float80    float80_t
+ */
+typedef npy_float32 __pyx_t_5numpy_float32_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":745
+ * 
+ * ctypedef npy_float32    float32_t
+ * ctypedef npy_float64    float64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_float80    float80_t
+ * #ctypedef npy_float128   float128_t
+ */
+typedef npy_float64 __pyx_t_5numpy_float64_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":754
+ * # The int types are mapped a bit surprising --
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t
+ */
+typedef npy_long __pyx_t_5numpy_int_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":755
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   longlong_t
+ * 
+ */
+typedef npy_longlong __pyx_t_5numpy_long_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":756
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_ulong      uint_t
+ */
+typedef npy_longlong __pyx_t_5numpy_longlong_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":758
+ * ctypedef npy_longlong   longlong_t
+ * 
+ * ctypedef npy_ulong      uint_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t
+ */
+typedef npy_ulong __pyx_t_5numpy_uint_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":759
+ * 
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulong_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":760
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_intp       intp_t
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulonglong_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":762
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ * ctypedef npy_intp       intp_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uintp      uintp_t
+ * 
+ */
+typedef npy_intp __pyx_t_5numpy_intp_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":763
+ * 
+ * ctypedef npy_intp       intp_t
+ * ctypedef npy_uintp      uintp_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_double     float_t
+ */
+typedef npy_uintp __pyx_t_5numpy_uintp_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":765
+ * ctypedef npy_uintp      uintp_t
+ * 
+ * ctypedef npy_double     float_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t
+ */
+typedef npy_double __pyx_t_5numpy_float_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":766
+ * 
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ */
+typedef npy_double __pyx_t_5numpy_double_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":767
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ */
+typedef npy_longdouble __pyx_t_5numpy_longdouble_t;
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< float > __pyx_t_float_complex;
+  #else
+    typedef float _Complex __pyx_t_float_complex;
+  #endif
+#else
+    typedef struct { float real, imag; } __pyx_t_float_complex;
+#endif
+static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float, float);
+
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< double > __pyx_t_double_complex;
+  #else
+    typedef double _Complex __pyx_t_double_complex;
+  #endif
+#else
+    typedef struct { double real, imag; } __pyx_t_double_complex;
+#endif
+static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double, double);
+
+
+/*--- Type declarations ---*/
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":769
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ * ctypedef npy_cfloat      cfloat_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t
+ */
+typedef npy_cfloat __pyx_t_5numpy_cfloat_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":770
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ */
+typedef npy_cdouble __pyx_t_5numpy_cdouble_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":771
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cdouble     complex_t
+ */
+typedef npy_clongdouble __pyx_t_5numpy_clongdouble_t;
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":773
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ * ctypedef npy_cdouble     complex_t             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ */
+typedef npy_cdouble __pyx_t_5numpy_complex_t;
+
+/* --- Runtime support code (head) --- */
+/* Refnanny.proto */
+#ifndef CYTHON_REFNANNY
+  #define CYTHON_REFNANNY 0
+#endif
+#if CYTHON_REFNANNY
+  typedef struct {
+    void (*INCREF)(void*, PyObject*, int);
+    void (*DECREF)(void*, PyObject*, int);
+    void (*GOTREF)(void*, PyObject*, int);
+    void (*GIVEREF)(void*, PyObject*, int);
+    void* (*SetupContext)(const char*, int, const char*);
+    void (*FinishContext)(void**);
+  } __Pyx_RefNannyAPIStruct;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname);
+  #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL;
+#ifdef WITH_THREAD
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          if (acquire_gil) {\
+              PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+              PyGILState_Release(__pyx_gilstate_save);\
+          } else {\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+          }
+#else
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__)
+#endif
+  #define __Pyx_RefNannyFinishContext()\
+          __Pyx_RefNanny->FinishContext(&__pyx_refnanny)
+  #define __Pyx_INCREF(r)  __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_DECREF(r)  __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GOTREF(r)  __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_XINCREF(r)  do { if((r) != NULL) {__Pyx_INCREF(r); }} while(0)
+  #define __Pyx_XDECREF(r)  do { if((r) != NULL) {__Pyx_DECREF(r); }} while(0)
+  #define __Pyx_XGOTREF(r)  do { if((r) != NULL) {__Pyx_GOTREF(r); }} while(0)
+  #define __Pyx_XGIVEREF(r) do { if((r) != NULL) {__Pyx_GIVEREF(r);}} while(0)
+#else
+  #define __Pyx_RefNannyDeclarations
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)
+  #define __Pyx_RefNannyFinishContext()
+  #define __Pyx_INCREF(r) Py_INCREF(r)
+  #define __Pyx_DECREF(r) Py_DECREF(r)
+  #define __Pyx_GOTREF(r)
+  #define __Pyx_GIVEREF(r)
+  #define __Pyx_XINCREF(r) Py_XINCREF(r)
+  #define __Pyx_XDECREF(r) Py_XDECREF(r)
+  #define __Pyx_XGOTREF(r)
+  #define __Pyx_XGIVEREF(r)
+#endif
+#define __Pyx_XDECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_XDECREF(tmp);\
+    } while (0)
+#define __Pyx_DECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_DECREF(tmp);\
+    } while (0)
+#define __Pyx_CLEAR(r)    do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0)
+#define __Pyx_XCLEAR(r)   do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0)
+
+/* RaiseArgTupleInvalid.proto */
+static void __Pyx_RaiseArgtupleInvalid(const char* func_name, int exact,
+    Py_ssize_t num_min, Py_ssize_t num_max, Py_ssize_t num_found);
+
+/* RaiseDoubleKeywords.proto */
+static void __Pyx_RaiseDoubleKeywordsError(const char* func_name, PyObject* kw_name);
+
+/* ParseKeywords.proto */
+static int __Pyx_ParseOptionalKeywords(PyObject *kwds, PyObject **argnames[],\
+    PyObject *kwds2, PyObject *values[], Py_ssize_t num_pos_args,\
+    const char* function_name);
+
+/* ArgTypeTest.proto */
+#define __Pyx_ArgTypeTest(obj, type, none_allowed, name, exact)\
+    ((likely((Py_TYPE(obj) == type) | (none_allowed && (obj == Py_None)))) ? 1 :\
+        __Pyx__ArgTypeTest(obj, type, name, exact))
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact);
+
+/* IsLittleEndian.proto */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void);
+
+/* BufferFormatCheck.proto */
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts);
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type);
+
+/* BufferGetAndValidate.proto */
+#define __Pyx_GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)\
+    ((obj == Py_None || obj == NULL) ?\
+    (__Pyx_ZeroBuffer(buf), 0) :\
+    __Pyx__GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack))
+static int  __Pyx__GetBufferAndValidate(Py_buffer* buf, PyObject* obj,
+    __Pyx_TypeInfo* dtype, int flags, int nd, int cast, __Pyx_BufFmt_StackElem* stack);
+static void __Pyx_ZeroBuffer(Py_buffer* buf);
+static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info);
+static Py_ssize_t __Pyx_minusones[] = { -1, -1, -1, -1, -1, -1, -1, -1 };
+static Py_ssize_t __Pyx_zeros[] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+
+/* PyThreadStateGet.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyThreadState_declare  PyThreadState *__pyx_tstate;
+#define __Pyx_PyThreadState_assign  __pyx_tstate = __Pyx_PyThreadState_Current;
+#define __Pyx_PyErr_Occurred()  __pyx_tstate->curexc_type
+#else
+#define __Pyx_PyThreadState_declare
+#define __Pyx_PyThreadState_assign
+#define __Pyx_PyErr_Occurred()  PyErr_Occurred()
+#endif
+
+/* PyErrFetchRestore.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)    __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)    __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL))
+#else
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#endif
+#else
+#define __Pyx_PyErr_Clear() PyErr_Clear()
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestoreInState(tstate, type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchInState(tstate, type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)  PyErr_Fetch(type, value, tb)
+#endif
+
+/* PyObjectGetAttrStr.proto */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name);
+#else
+#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n)
+#endif
+
+/* GetBuiltinName.proto */
+static PyObject *__Pyx_GetBuiltinName(PyObject *name);
+
+/* PyObjectCall.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw);
+#else
+#define __Pyx_PyObject_Call(func, arg, kw) PyObject_Call(func, arg, kw)
+#endif
+
+/* RaiseException.proto */
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause);
+
+/* PyCFunctionFastCall.proto */
+#if CYTHON_FAST_PYCCALL
+static CYTHON_INLINE PyObject *__Pyx_PyCFunction_FastCall(PyObject *func, PyObject **args, Py_ssize_t nargs);
+#else
+#define __Pyx_PyCFunction_FastCall(func, args, nargs)  (assert(0), NULL)
+#endif
+
+/* PyFunctionFastCall.proto */
+#if CYTHON_FAST_PYCALL
+#define __Pyx_PyFunction_FastCall(func, args, nargs)\
+    __Pyx_PyFunction_FastCallDict((func), (args), (nargs), NULL)
+#if 1 || PY_VERSION_HEX < 0x030600B1
+static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, int nargs, PyObject *kwargs);
+#else
+#define __Pyx_PyFunction_FastCallDict(func, args, nargs, kwargs) _PyFunction_FastCallDict(func, args, nargs, kwargs)
+#endif
+#endif
+
+/* PyObjectCallMethO.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg);
+#endif
+
+/* PyObjectCallOneArg.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg);
+
+/* DictGetItem.proto */
+#if PY_MAJOR_VERSION >= 3 && !CYTHON_COMPILING_IN_PYPY
+static PyObject *__Pyx_PyDict_GetItem(PyObject *d, PyObject* key);
+#define __Pyx_PyObject_Dict_GetItem(obj, name)\
+    (likely(PyDict_CheckExact(obj)) ?\
+     __Pyx_PyDict_GetItem(obj, name) : PyObject_GetItem(obj, name))
+#else
+#define __Pyx_PyDict_GetItem(d, key) PyObject_GetItem(d, key)
+#define __Pyx_PyObject_Dict_GetItem(obj, name)  PyObject_GetItem(obj, name)
+#endif
+
+/* RaiseTooManyValuesToUnpack.proto */
+static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected);
+
+/* RaiseNeedMoreValuesToUnpack.proto */
+static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index);
+
+/* RaiseNoneIterError.proto */
+static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void);
+
+/* ExtTypeTest.proto */
+static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type);
+
+/* SaveResetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_ExceptionSave(type, value, tb)  __Pyx__ExceptionSave(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#define __Pyx_ExceptionReset(type, value, tb)  __Pyx__ExceptionReset(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+#else
+#define __Pyx_ExceptionSave(type, value, tb)   PyErr_GetExcInfo(type, value, tb)
+#define __Pyx_ExceptionReset(type, value, tb)  PyErr_SetExcInfo(type, value, tb)
+#endif
+
+/* PyErrExceptionMatches.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_ExceptionMatches(err) __Pyx_PyErr_ExceptionMatchesInState(__pyx_tstate, err)
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err);
+#else
+#define __Pyx_PyErr_ExceptionMatches(err)  PyErr_ExceptionMatches(err)
+#endif
+
+/* GetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_GetException(type, value, tb)  __Pyx__GetException(__pyx_tstate, type, value, tb)
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb);
+#endif
+
+/* Import.proto */
+static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level);
+
+/* CLineInTraceback.proto */
+#ifdef CYTHON_CLINE_IN_TRACEBACK
+#define __Pyx_CLineForTraceback(tstate, c_line)  (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0)
+#else
+static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line);
+#endif
+
+/* CodeObjectCache.proto */
+typedef struct {
+    PyCodeObject* code_object;
+    int code_line;
+} __Pyx_CodeObjectCacheEntry;
+struct __Pyx_CodeObjectCache {
+    int count;
+    int max_count;
+    __Pyx_CodeObjectCacheEntry* entries;
+};
+static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL};
+static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line);
+static PyCodeObject *__pyx_find_code_object(int code_line);
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object);
+
+/* AddTraceback.proto */
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename);
+
+/* BufferStructDeclare.proto */
+typedef struct {
+  Py_ssize_t shape, strides, suboffsets;
+} __Pyx_Buf_DimInfo;
+typedef struct {
+  size_t refcount;
+  Py_buffer pybuffer;
+} __Pyx_Buffer;
+typedef struct {
+  __Pyx_Buffer *rcbuffer;
+  char *data;
+  __Pyx_Buf_DimInfo diminfo[8];
+} __Pyx_LocalBuf_ND;
+
+#if PY_MAJOR_VERSION < 3
+    static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags);
+    static void __Pyx_ReleaseBuffer(Py_buffer *view);
+#else
+    #define __Pyx_GetBuffer PyObject_GetBuffer
+    #define __Pyx_ReleaseBuffer PyBuffer_Release
+#endif
+
+
+/* RealImag.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #define __Pyx_CREAL(z) ((z).real())
+    #define __Pyx_CIMAG(z) ((z).imag())
+  #else
+    #define __Pyx_CREAL(z) (__real__(z))
+    #define __Pyx_CIMAG(z) (__imag__(z))
+  #endif
+#else
+    #define __Pyx_CREAL(z) ((z).real)
+    #define __Pyx_CIMAG(z) ((z).imag)
+#endif
+#if defined(__cplusplus) && CYTHON_CCOMPLEX\
+        && (defined(_WIN32) || defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 5 || __GNUC__ == 4 && __GNUC_MINOR__ >= 4 )) || __cplusplus >= 201103)
+    #define __Pyx_SET_CREAL(z,x) ((z).real(x))
+    #define __Pyx_SET_CIMAG(z,y) ((z).imag(y))
+#else
+    #define __Pyx_SET_CREAL(z,x) __Pyx_CREAL(z) = (x)
+    #define __Pyx_SET_CIMAG(z,y) __Pyx_CIMAG(z) = (y)
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_float(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_float(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_float(a, b) ((a)-(b))
+    #define __Pyx_c_prod_float(a, b) ((a)*(b))
+    #define __Pyx_c_quot_float(a, b) ((a)/(b))
+    #define __Pyx_c_neg_float(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_float(z) ((z)==(float)0)
+    #define __Pyx_c_conj_float(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (::std::abs(z))
+        #define __Pyx_c_pow_float(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_float(z) ((z)==0)
+    #define __Pyx_c_conj_float(z)    (conjf(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (cabsf(z))
+        #define __Pyx_c_pow_float(a, b)  (cpowf(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex);
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex);
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    #endif
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_double(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_double(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_double(a, b) ((a)-(b))
+    #define __Pyx_c_prod_double(a, b) ((a)*(b))
+    #define __Pyx_c_quot_double(a, b) ((a)/(b))
+    #define __Pyx_c_neg_double(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_double(z) ((z)==(double)0)
+    #define __Pyx_c_conj_double(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (::std::abs(z))
+        #define __Pyx_c_pow_double(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_double(z) ((z)==0)
+    #define __Pyx_c_conj_double(z)    (conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (cabs(z))
+        #define __Pyx_c_pow_double(a, b)  (cpow(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex);
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex);
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    #endif
+#endif
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value);
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_enum__NPY_TYPES(enum NPY_TYPES value);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *);
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *);
+
+/* FastTypeChecks.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type)
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2);
+#else
+#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type)
+#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type)
+#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2))
+#endif
+#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception)
+
+/* CheckBinaryVersion.proto */
+static int __Pyx_check_binary_version(void);
+
+/* PyIdentifierFromString.proto */
+#if !defined(__Pyx_PyIdentifier_FromString)
+#if PY_MAJOR_VERSION < 3
+  #define __Pyx_PyIdentifier_FromString(s) PyString_FromString(s)
+#else
+  #define __Pyx_PyIdentifier_FromString(s) PyUnicode_FromString(s)
+#endif
+#endif
+
+/* ModuleImport.proto */
+static PyObject *__Pyx_ImportModule(const char *name);
+
+/* TypeImport.proto */
+static PyTypeObject *__Pyx_ImportType(const char *module_name, const char *class_name, size_t size, int strict);
+
+/* InitStrings.proto */
+static int __Pyx_InitStrings(__Pyx_StringTabEntry *t);
+
+
+/* Module declarations from 'cpython.buffer' */
+
+/* Module declarations from 'libc.string' */
+
+/* Module declarations from 'libc.stdio' */
+
+/* Module declarations from '__builtin__' */
+
+/* Module declarations from 'cpython.type' */
+static PyTypeObject *__pyx_ptype_7cpython_4type_type = 0;
+
+/* Module declarations from 'cpython' */
+
+/* Module declarations from 'cpython.object' */
+
+/* Module declarations from 'cpython.ref' */
+
+/* Module declarations from 'cpython.mem' */
+
+/* Module declarations from 'numpy' */
+
+/* Module declarations from 'numpy' */
+static PyTypeObject *__pyx_ptype_5numpy_dtype = 0;
+static PyTypeObject *__pyx_ptype_5numpy_flatiter = 0;
+static PyTypeObject *__pyx_ptype_5numpy_broadcast = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ndarray = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ufunc = 0;
+static CYTHON_INLINE char *__pyx_f_5numpy__util_dtypestring(PyArray_Descr *, char *, char *, int *); /*proto*/
+static CYTHON_INLINE int __pyx_f_5numpy_import_array(void); /*proto*/
+
+/* Module declarations from 'libcpp.string' */
+
+/* Module declarations from 'mesh_core_cython' */
+static std::string __pyx_convert_string_from_py_std__in_string(PyObject *); /*proto*/
+static __Pyx_TypeInfo __Pyx_TypeInfo_float = { "float", NULL, sizeof(float), { 0 }, 0, 'R', 0, 0 };
+static __Pyx_TypeInfo __Pyx_TypeInfo_int = { "int", NULL, sizeof(int), { 0 }, 0, IS_UNSIGNED(int) ? 'U' : 'I', IS_UNSIGNED(int), 0 };
+#define __Pyx_MODULE_NAME "mesh_core_cython"
+extern int __pyx_module_is_main_mesh_core_cython;
+int __pyx_module_is_main_mesh_core_cython = 0;
+
+/* Implementation of 'mesh_core_cython' */
+static PyObject *__pyx_builtin_ValueError;
+static PyObject *__pyx_builtin_range;
+static PyObject *__pyx_builtin_RuntimeError;
+static PyObject *__pyx_builtin_ImportError;
+static const char __pyx_k_c[] = "c";
+static const char __pyx_k_h[] = "h";
+static const char __pyx_k_w[] = "w";
+static const char __pyx_k_np[] = "np";
+static const char __pyx_k_main[] = "__main__";
+static const char __pyx_k_ntri[] = "ntri";
+static const char __pyx_k_nver[] = "nver";
+static const char __pyx_k_test[] = "__test__";
+static const char __pyx_k_image[] = "image";
+static const char __pyx_k_numpy[] = "numpy";
+static const char __pyx_k_range[] = "range";
+static const char __pyx_k_tex_c[] = "tex_c";
+static const char __pyx_k_tex_h[] = "tex_h";
+static const char __pyx_k_tex_w[] = "tex_w";
+static const char __pyx_k_colors[] = "colors";
+static const char __pyx_k_import[] = "__import__";
+static const char __pyx_k_normal[] = "normal";
+static const char __pyx_k_ntexver[] = "ntexver";
+static const char __pyx_k_texture[] = "texture";
+static const char __pyx_k_filename[] = "filename";
+static const char __pyx_k_mtl_name[] = "mtl_name";
+static const char __pyx_k_tex_nver[] = "tex_nver";
+static const char __pyx_k_vertices[] = "vertices";
+static const char __pyx_k_triangles[] = "triangles";
+static const char __pyx_k_uv_coords[] = "uv_coords";
+static const char __pyx_k_ValueError[] = "ValueError";
+static const char __pyx_k_tex_coords[] = "tex_coords";
+static const char __pyx_k_tri_normal[] = "tri_normal";
+static const char __pyx_k_ImportError[] = "ImportError";
+static const char __pyx_k_RuntimeError[] = "RuntimeError";
+static const char __pyx_k_depth_buffer[] = "depth_buffer";
+static const char __pyx_k_mapping_type[] = "mapping_type";
+static const char __pyx_k_tex_triangles[] = "tex_triangles";
+static const char __pyx_k_get_normal_core[] = "get_normal_core";
+static const char __pyx_k_triangle_buffer[] = "triangle_buffer";
+static const char __pyx_k_mesh_core_cython[] = "mesh_core_cython";
+static const char __pyx_k_barycentric_weight[] = "barycentric_weight";
+static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback";
+static const char __pyx_k_render_colors_core[] = "render_colors_core";
+static const char __pyx_k_render_texture_core[] = "render_texture_core";
+static const char __pyx_k_rasterize_triangles_core[] = "rasterize_triangles_core";
+static const char __pyx_k_ndarray_is_not_C_contiguous[] = "ndarray is not C contiguous";
+static const char __pyx_k_numpy_core_multiarray_failed_to[] = "numpy.core.multiarray failed to import";
+static const char __pyx_k_unknown_dtype_code_in_numpy_pxd[] = "unknown dtype code in numpy.pxd (%d)";
+static const char __pyx_k_Format_string_allocated_too_shor[] = "Format string allocated too short, see comment in numpy.pxd";
+static const char __pyx_k_Non_native_byte_order_not_suppor[] = "Non-native byte order not supported";
+static const char __pyx_k_insightface_thirdparty_face3d_me[] = "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx";
+static const char __pyx_k_ndarray_is_not_Fortran_contiguou[] = "ndarray is not Fortran contiguous";
+static const char __pyx_k_numpy_core_umath_failed_to_impor[] = "numpy.core.umath failed to import";
+static const char __pyx_k_write_obj_with_colors_texture_co[] = "write_obj_with_colors_texture_core";
+static const char __pyx_k_Format_string_allocated_too_shor_2[] = "Format string allocated too short.";
+static PyObject *__pyx_kp_u_Format_string_allocated_too_shor;
+static PyObject *__pyx_kp_u_Format_string_allocated_too_shor_2;
+static PyObject *__pyx_n_s_ImportError;
+static PyObject *__pyx_kp_u_Non_native_byte_order_not_suppor;
+static PyObject *__pyx_n_s_RuntimeError;
+static PyObject *__pyx_n_s_ValueError;
+static PyObject *__pyx_n_s_barycentric_weight;
+static PyObject *__pyx_n_s_c;
+static PyObject *__pyx_n_s_cline_in_traceback;
+static PyObject *__pyx_n_s_colors;
+static PyObject *__pyx_n_s_depth_buffer;
+static PyObject *__pyx_n_s_filename;
+static PyObject *__pyx_n_s_get_normal_core;
+static PyObject *__pyx_n_s_h;
+static PyObject *__pyx_n_s_image;
+static PyObject *__pyx_n_s_import;
+static PyObject *__pyx_kp_s_insightface_thirdparty_face3d_me;
+static PyObject *__pyx_n_s_main;
+static PyObject *__pyx_n_s_mapping_type;
+static PyObject *__pyx_n_s_mesh_core_cython;
+static PyObject *__pyx_n_s_mtl_name;
+static PyObject *__pyx_kp_u_ndarray_is_not_C_contiguous;
+static PyObject *__pyx_kp_u_ndarray_is_not_Fortran_contiguou;
+static PyObject *__pyx_n_s_normal;
+static PyObject *__pyx_n_s_np;
+static PyObject *__pyx_n_s_ntexver;
+static PyObject *__pyx_n_s_ntri;
+static PyObject *__pyx_n_s_numpy;
+static PyObject *__pyx_kp_s_numpy_core_multiarray_failed_to;
+static PyObject *__pyx_kp_s_numpy_core_umath_failed_to_impor;
+static PyObject *__pyx_n_s_nver;
+static PyObject *__pyx_n_s_range;
+static PyObject *__pyx_n_s_rasterize_triangles_core;
+static PyObject *__pyx_n_s_render_colors_core;
+static PyObject *__pyx_n_s_render_texture_core;
+static PyObject *__pyx_n_s_test;
+static PyObject *__pyx_n_s_tex_c;
+static PyObject *__pyx_n_s_tex_coords;
+static PyObject *__pyx_n_s_tex_h;
+static PyObject *__pyx_n_s_tex_nver;
+static PyObject *__pyx_n_s_tex_triangles;
+static PyObject *__pyx_n_s_tex_w;
+static PyObject *__pyx_n_s_texture;
+static PyObject *__pyx_n_s_tri_normal;
+static PyObject *__pyx_n_s_triangle_buffer;
+static PyObject *__pyx_n_s_triangles;
+static PyObject *__pyx_kp_u_unknown_dtype_code_in_numpy_pxd;
+static PyObject *__pyx_n_s_uv_coords;
+static PyObject *__pyx_n_s_vertices;
+static PyObject *__pyx_n_s_w;
+static PyObject *__pyx_n_s_write_obj_with_colors_texture_co;
+static PyObject *__pyx_pf_16mesh_core_cython_get_normal_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_normal, PyArrayObject *__pyx_v_tri_normal, PyArrayObject *__pyx_v_triangles, int __pyx_v_ntri); /* proto */
+static PyObject *__pyx_pf_16mesh_core_cython_2rasterize_triangles_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_depth_buffer, PyArrayObject *__pyx_v_triangle_buffer, PyArrayObject *__pyx_v_barycentric_weight, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w); /* proto */
+static PyObject *__pyx_pf_16mesh_core_cython_4render_colors_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c); /* proto */
+static PyObject *__pyx_pf_16mesh_core_cython_6render_texture_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_texture, PyArrayObject *__pyx_v_tex_coords, PyArrayObject *__pyx_v_tex_triangles, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_tex_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c, int __pyx_v_tex_h, int __pyx_v_tex_w, int __pyx_v_tex_c, int __pyx_v_mapping_type); /* proto */
+static PyObject *__pyx_pf_16mesh_core_cython_8write_obj_with_colors_texture_core(CYTHON_UNUSED PyObject *__pyx_self, std::string __pyx_v_filename, std::string __pyx_v_mtl_name, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_uv_coords, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_ntexver); /* proto */
+static int __pyx_pf_5numpy_7ndarray___getbuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */
+static void __pyx_pf_5numpy_7ndarray_2__releasebuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info); /* proto */
+static PyObject *__pyx_tuple_;
+static PyObject *__pyx_tuple__2;
+static PyObject *__pyx_tuple__3;
+static PyObject *__pyx_tuple__4;
+static PyObject *__pyx_tuple__5;
+static PyObject *__pyx_tuple__6;
+static PyObject *__pyx_tuple__7;
+static PyObject *__pyx_tuple__8;
+static PyObject *__pyx_tuple__9;
+static PyObject *__pyx_tuple__10;
+static PyObject *__pyx_tuple__12;
+static PyObject *__pyx_tuple__14;
+static PyObject *__pyx_tuple__16;
+static PyObject *__pyx_tuple__18;
+static PyObject *__pyx_codeobj__11;
+static PyObject *__pyx_codeobj__13;
+static PyObject *__pyx_codeobj__15;
+static PyObject *__pyx_codeobj__17;
+static PyObject *__pyx_codeobj__19;
+/* Late includes */
+
+/* "mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_16mesh_core_cython_1get_normal_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_16mesh_core_cython_1get_normal_core = {"get_normal_core", (PyCFunction)__pyx_pw_16mesh_core_cython_1get_normal_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_16mesh_core_cython_1get_normal_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_normal = 0;
+  PyArrayObject *__pyx_v_tri_normal = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  int __pyx_v_ntri;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("get_normal_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_normal,&__pyx_n_s_tri_normal,&__pyx_n_s_triangles,&__pyx_n_s_ntri,0};
+    PyObject* values[4] = {0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_normal)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tri_normal)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 1); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 2); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 3); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "get_normal_core") < 0)) __PYX_ERR(0, 40, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 4) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+    }
+    __pyx_v_normal = ((PyArrayObject *)values[0]);
+    __pyx_v_tri_normal = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[3]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 43, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 40, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mesh_core_cython.get_normal_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_normal), __pyx_ptype_5numpy_ndarray, 0, "normal", 0))) __PYX_ERR(0, 40, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tri_normal), __pyx_ptype_5numpy_ndarray, 0, "tri_normal", 0))) __PYX_ERR(0, 41, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 42, __pyx_L1_error)
+  __pyx_r = __pyx_pf_16mesh_core_cython_get_normal_core(__pyx_self, __pyx_v_normal, __pyx_v_tri_normal, __pyx_v_triangles, __pyx_v_ntri);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_16mesh_core_cython_get_normal_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_normal, PyArrayObject *__pyx_v_tri_normal, PyArrayObject *__pyx_v_triangles, int __pyx_v_ntri) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_normal;
+  __Pyx_Buffer __pyx_pybuffer_normal;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tri_normal;
+  __Pyx_Buffer __pyx_pybuffer_tri_normal;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("get_normal_core", 0);
+  __pyx_pybuffer_normal.pybuffer.buf = NULL;
+  __pyx_pybuffer_normal.refcount = 0;
+  __pyx_pybuffernd_normal.data = NULL;
+  __pyx_pybuffernd_normal.rcbuffer = &__pyx_pybuffer_normal;
+  __pyx_pybuffer_tri_normal.pybuffer.buf = NULL;
+  __pyx_pybuffer_tri_normal.refcount = 0;
+  __pyx_pybuffernd_tri_normal.data = NULL;
+  __pyx_pybuffernd_tri_normal.rcbuffer = &__pyx_pybuffer_tri_normal;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_normal.rcbuffer->pybuffer, (PyObject*)__pyx_v_normal, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_normal.diminfo[0].strides = __pyx_pybuffernd_normal.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_normal.diminfo[0].shape = __pyx_pybuffernd_normal.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_normal.diminfo[1].strides = __pyx_pybuffernd_normal.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_normal.diminfo[1].shape = __pyx_pybuffernd_normal.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer, (PyObject*)__pyx_v_tri_normal, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tri_normal.diminfo[0].strides = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tri_normal.diminfo[0].shape = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tri_normal.diminfo[1].strides = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tri_normal.diminfo[1].shape = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+
+  /* "mesh_core_cython.pyx":45
+ *                 int ntri
+ *                 ):
+ *     _get_normal_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(normal), <float*> np.PyArray_DATA(tri_normal), <int*> np.PyArray_DATA(triangles),
+ *         ntri)
+ */
+  _get_normal_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_normal))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_tri_normal))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), __pyx_v_ntri);
+
+  /* "mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_normal.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mesh_core_cython.get_normal_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_normal.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_16mesh_core_cython_3rasterize_triangles_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_16mesh_core_cython_3rasterize_triangles_core = {"rasterize_triangles_core", (PyCFunction)__pyx_pw_16mesh_core_cython_3rasterize_triangles_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_16mesh_core_cython_3rasterize_triangles_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  PyArrayObject *__pyx_v_triangle_buffer = 0;
+  PyArrayObject *__pyx_v_barycentric_weight = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("rasterize_triangles_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_depth_buffer,&__pyx_n_s_triangle_buffer,&__pyx_n_s_barycentric_weight,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,0};
+    PyObject* values[9] = {0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 1); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 2); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangle_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 3); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_barycentric_weight)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 4); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 5); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 6); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 7); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 8); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "rasterize_triangles_core") < 0)) __PYX_ERR(0, 49, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 9) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+    }
+    __pyx_v_vertices = ((PyArrayObject *)values[0]);
+    __pyx_v_triangles = ((PyArrayObject *)values[1]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[2]);
+    __pyx_v_triangle_buffer = ((PyArrayObject *)values[3]);
+    __pyx_v_barycentric_weight = ((PyArrayObject *)values[4]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[5]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 55, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 55, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 56, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 56, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 49, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mesh_core_cython.rasterize_triangles_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 50, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 51, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 52, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangle_buffer), __pyx_ptype_5numpy_ndarray, 0, "triangle_buffer", 0))) __PYX_ERR(0, 53, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_barycentric_weight), __pyx_ptype_5numpy_ndarray, 0, "barycentric_weight", 0))) __PYX_ERR(0, 54, __pyx_L1_error)
+  __pyx_r = __pyx_pf_16mesh_core_cython_2rasterize_triangles_core(__pyx_self, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_depth_buffer, __pyx_v_triangle_buffer, __pyx_v_barycentric_weight, __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_16mesh_core_cython_2rasterize_triangles_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_depth_buffer, PyArrayObject *__pyx_v_triangle_buffer, PyArrayObject *__pyx_v_barycentric_weight, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_barycentric_weight;
+  __Pyx_Buffer __pyx_pybuffer_barycentric_weight;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangle_buffer;
+  __Pyx_Buffer __pyx_pybuffer_triangle_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("rasterize_triangles_core", 0);
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  __pyx_pybuffer_triangle_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangle_buffer.refcount = 0;
+  __pyx_pybuffernd_triangle_buffer.data = NULL;
+  __pyx_pybuffernd_triangle_buffer.rcbuffer = &__pyx_pybuffer_triangle_buffer;
+  __pyx_pybuffer_barycentric_weight.pybuffer.buf = NULL;
+  __pyx_pybuffer_barycentric_weight.refcount = 0;
+  __pyx_pybuffernd_barycentric_weight.data = NULL;
+  __pyx_pybuffernd_barycentric_weight.rcbuffer = &__pyx_pybuffer_barycentric_weight;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangle_buffer, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangle_buffer.diminfo[0].strides = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangle_buffer.diminfo[0].shape = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangle_buffer.diminfo[1].strides = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangle_buffer.diminfo[1].shape = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer, (PyObject*)__pyx_v_barycentric_weight, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_barycentric_weight.diminfo[0].strides = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_barycentric_weight.diminfo[0].shape = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_barycentric_weight.diminfo[1].strides = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_barycentric_weight.diminfo[1].shape = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.shape[1];
+
+  /* "mesh_core_cython.pyx":58
+ *                 int h, int w
+ *                 ):
+ *     _rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(depth_buffer), <int*> np.PyArray_DATA(triangle_buffer), <float*> np.PyArray_DATA(barycentric_weight),
+ */
+  _rasterize_triangles_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangle_buffer))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_barycentric_weight))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w);
+
+  /* "mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mesh_core_cython.rasterize_triangles_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_16mesh_core_cython_5render_colors_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_16mesh_core_cython_5render_colors_core = {"render_colors_core", (PyCFunction)__pyx_pw_16mesh_core_cython_5render_colors_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_16mesh_core_cython_5render_colors_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_image = 0;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_colors = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  int __pyx_v_c;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_colors_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_image,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_colors,&__pyx_n_s_depth_buffer,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,&__pyx_n_s_c,0};
+    PyObject* values[10] = {0,0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case 10: values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+        CYTHON_FALLTHROUGH;
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_image)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 1); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 2); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_colors)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 3); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 4); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 5); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 6); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 7); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 8); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  9:
+        if (likely((values[9] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 9); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "render_colors_core") < 0)) __PYX_ERR(0, 64, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 10) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+      values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+    }
+    __pyx_v_image = ((PyArrayObject *)values[0]);
+    __pyx_v_vertices = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_colors = ((PyArrayObject *)values[3]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[4]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[5]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 69, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 69, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+    __pyx_v_c = __Pyx_PyInt_As_int(values[9]); if (unlikely((__pyx_v_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 64, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mesh_core_cython.render_colors_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_image), __pyx_ptype_5numpy_ndarray, 0, "image", 0))) __PYX_ERR(0, 64, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 65, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 66, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_colors), __pyx_ptype_5numpy_ndarray, 0, "colors", 0))) __PYX_ERR(0, 67, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 68, __pyx_L1_error)
+  __pyx_r = __pyx_pf_16mesh_core_cython_4render_colors_core(__pyx_self, __pyx_v_image, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_colors, __pyx_v_depth_buffer, __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_16mesh_core_cython_4render_colors_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_colors;
+  __Pyx_Buffer __pyx_pybuffer_colors;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_image;
+  __Pyx_Buffer __pyx_pybuffer_image;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_colors_core", 0);
+  __pyx_pybuffer_image.pybuffer.buf = NULL;
+  __pyx_pybuffer_image.refcount = 0;
+  __pyx_pybuffernd_image.data = NULL;
+  __pyx_pybuffernd_image.rcbuffer = &__pyx_pybuffer_image;
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_colors.pybuffer.buf = NULL;
+  __pyx_pybuffer_colors.refcount = 0;
+  __pyx_pybuffernd_colors.data = NULL;
+  __pyx_pybuffernd_colors.rcbuffer = &__pyx_pybuffer_colors;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_image.rcbuffer->pybuffer, (PyObject*)__pyx_v_image, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_image.diminfo[0].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_image.diminfo[0].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_image.diminfo[1].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_image.diminfo[1].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_image.diminfo[2].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_image.diminfo[2].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_colors.rcbuffer->pybuffer, (PyObject*)__pyx_v_colors, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_colors.diminfo[0].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_colors.diminfo[0].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_colors.diminfo[1].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_colors.diminfo[1].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+
+  /* "mesh_core_cython.pyx":72
+ *                 int h, int w, int c
+ *                 ):
+ *     _render_colors_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(colors),
+ */
+  _render_colors_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_image))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_colors))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c);
+
+  /* "mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mesh_core_cython.render_colors_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_16mesh_core_cython_7render_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_16mesh_core_cython_7render_texture_core = {"render_texture_core", (PyCFunction)__pyx_pw_16mesh_core_cython_7render_texture_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_16mesh_core_cython_7render_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_image = 0;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_texture = 0;
+  PyArrayObject *__pyx_v_tex_coords = 0;
+  PyArrayObject *__pyx_v_tex_triangles = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  int __pyx_v_nver;
+  int __pyx_v_tex_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  int __pyx_v_c;
+  int __pyx_v_tex_h;
+  int __pyx_v_tex_w;
+  int __pyx_v_tex_c;
+  int __pyx_v_mapping_type;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_texture_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_image,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_texture,&__pyx_n_s_tex_coords,&__pyx_n_s_tex_triangles,&__pyx_n_s_depth_buffer,&__pyx_n_s_nver,&__pyx_n_s_tex_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,&__pyx_n_s_c,&__pyx_n_s_tex_h,&__pyx_n_s_tex_w,&__pyx_n_s_tex_c,&__pyx_n_s_mapping_type,0};
+    PyObject* values[17] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case 17: values[16] = PyTuple_GET_ITEM(__pyx_args, 16);
+        CYTHON_FALLTHROUGH;
+        case 16: values[15] = PyTuple_GET_ITEM(__pyx_args, 15);
+        CYTHON_FALLTHROUGH;
+        case 15: values[14] = PyTuple_GET_ITEM(__pyx_args, 14);
+        CYTHON_FALLTHROUGH;
+        case 14: values[13] = PyTuple_GET_ITEM(__pyx_args, 13);
+        CYTHON_FALLTHROUGH;
+        case 13: values[12] = PyTuple_GET_ITEM(__pyx_args, 12);
+        CYTHON_FALLTHROUGH;
+        case 12: values[11] = PyTuple_GET_ITEM(__pyx_args, 11);
+        CYTHON_FALLTHROUGH;
+        case 11: values[10] = PyTuple_GET_ITEM(__pyx_args, 10);
+        CYTHON_FALLTHROUGH;
+        case 10: values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+        CYTHON_FALLTHROUGH;
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_image)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 1); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 2); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_texture)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 3); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_coords)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 4); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 5); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 6); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 7); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 8); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  9:
+        if (likely((values[9] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 9); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 10:
+        if (likely((values[10] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 10); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 11:
+        if (likely((values[11] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 11); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 12:
+        if (likely((values[12] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 12); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 13:
+        if (likely((values[13] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 13); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 14:
+        if (likely((values[14] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 14); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 15:
+        if (likely((values[15] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 15); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 16:
+        if (likely((values[16] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_mapping_type)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 16); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "render_texture_core") < 0)) __PYX_ERR(0, 79, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 17) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+      values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+      values[10] = PyTuple_GET_ITEM(__pyx_args, 10);
+      values[11] = PyTuple_GET_ITEM(__pyx_args, 11);
+      values[12] = PyTuple_GET_ITEM(__pyx_args, 12);
+      values[13] = PyTuple_GET_ITEM(__pyx_args, 13);
+      values[14] = PyTuple_GET_ITEM(__pyx_args, 14);
+      values[15] = PyTuple_GET_ITEM(__pyx_args, 15);
+      values[16] = PyTuple_GET_ITEM(__pyx_args, 16);
+    }
+    __pyx_v_image = ((PyArrayObject *)values[0]);
+    __pyx_v_vertices = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_texture = ((PyArrayObject *)values[3]);
+    __pyx_v_tex_coords = ((PyArrayObject *)values[4]);
+    __pyx_v_tex_triangles = ((PyArrayObject *)values[5]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[6]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_tex_nver = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_tex_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[9]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[10]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[11]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_c = __Pyx_PyInt_As_int(values[12]); if (unlikely((__pyx_v_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_tex_h = __Pyx_PyInt_As_int(values[13]); if (unlikely((__pyx_v_tex_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_tex_w = __Pyx_PyInt_As_int(values[14]); if (unlikely((__pyx_v_tex_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_tex_c = __Pyx_PyInt_As_int(values[15]); if (unlikely((__pyx_v_tex_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_mapping_type = __Pyx_PyInt_As_int(values[16]); if (unlikely((__pyx_v_mapping_type == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 89, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 79, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mesh_core_cython.render_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_image), __pyx_ptype_5numpy_ndarray, 0, "image", 0))) __PYX_ERR(0, 79, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 80, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 81, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_texture), __pyx_ptype_5numpy_ndarray, 0, "texture", 0))) __PYX_ERR(0, 82, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tex_coords), __pyx_ptype_5numpy_ndarray, 0, "tex_coords", 0))) __PYX_ERR(0, 83, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tex_triangles), __pyx_ptype_5numpy_ndarray, 0, "tex_triangles", 0))) __PYX_ERR(0, 84, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 85, __pyx_L1_error)
+  __pyx_r = __pyx_pf_16mesh_core_cython_6render_texture_core(__pyx_self, __pyx_v_image, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_texture, __pyx_v_tex_coords, __pyx_v_tex_triangles, __pyx_v_depth_buffer, __pyx_v_nver, __pyx_v_tex_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c, __pyx_v_tex_h, __pyx_v_tex_w, __pyx_v_tex_c, __pyx_v_mapping_type);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_16mesh_core_cython_6render_texture_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_texture, PyArrayObject *__pyx_v_tex_coords, PyArrayObject *__pyx_v_tex_triangles, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_tex_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c, int __pyx_v_tex_h, int __pyx_v_tex_w, int __pyx_v_tex_c, int __pyx_v_mapping_type) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_image;
+  __Pyx_Buffer __pyx_pybuffer_image;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tex_coords;
+  __Pyx_Buffer __pyx_pybuffer_tex_coords;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tex_triangles;
+  __Pyx_Buffer __pyx_pybuffer_tex_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_texture;
+  __Pyx_Buffer __pyx_pybuffer_texture;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_texture_core", 0);
+  __pyx_pybuffer_image.pybuffer.buf = NULL;
+  __pyx_pybuffer_image.refcount = 0;
+  __pyx_pybuffernd_image.data = NULL;
+  __pyx_pybuffernd_image.rcbuffer = &__pyx_pybuffer_image;
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_texture.pybuffer.buf = NULL;
+  __pyx_pybuffer_texture.refcount = 0;
+  __pyx_pybuffernd_texture.data = NULL;
+  __pyx_pybuffernd_texture.rcbuffer = &__pyx_pybuffer_texture;
+  __pyx_pybuffer_tex_coords.pybuffer.buf = NULL;
+  __pyx_pybuffer_tex_coords.refcount = 0;
+  __pyx_pybuffernd_tex_coords.data = NULL;
+  __pyx_pybuffernd_tex_coords.rcbuffer = &__pyx_pybuffer_tex_coords;
+  __pyx_pybuffer_tex_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_tex_triangles.refcount = 0;
+  __pyx_pybuffernd_tex_triangles.data = NULL;
+  __pyx_pybuffernd_tex_triangles.rcbuffer = &__pyx_pybuffer_tex_triangles;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_image.rcbuffer->pybuffer, (PyObject*)__pyx_v_image, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_image.diminfo[0].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_image.diminfo[0].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_image.diminfo[1].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_image.diminfo[1].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_image.diminfo[2].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_image.diminfo[2].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_texture.rcbuffer->pybuffer, (PyObject*)__pyx_v_texture, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_texture.diminfo[0].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_texture.diminfo[0].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_texture.diminfo[1].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_texture.diminfo[1].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_texture.diminfo[2].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_texture.diminfo[2].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer, (PyObject*)__pyx_v_tex_coords, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tex_coords.diminfo[0].strides = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tex_coords.diminfo[0].shape = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tex_coords.diminfo[1].strides = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tex_coords.diminfo[1].shape = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_tex_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tex_triangles.diminfo[0].strides = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tex_triangles.diminfo[0].shape = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tex_triangles.diminfo[1].strides = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tex_triangles.diminfo[1].shape = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+
+  /* "mesh_core_cython.pyx":91
+ *                 int mapping_type
+ *                 ):
+ *     _render_texture_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(texture), <float*> np.PyArray_DATA(tex_coords), <int*> np.PyArray_DATA(tex_triangles),
+ */
+  _render_texture_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_image))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_texture))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_tex_coords))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_tex_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), __pyx_v_nver, __pyx_v_tex_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c, __pyx_v_tex_h, __pyx_v_tex_w, __pyx_v_tex_c, __pyx_v_mapping_type);
+
+  /* "mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_texture.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mesh_core_cython.render_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_texture.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_16mesh_core_cython_9write_obj_with_colors_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_16mesh_core_cython_9write_obj_with_colors_texture_core = {"write_obj_with_colors_texture_core", (PyCFunction)__pyx_pw_16mesh_core_cython_9write_obj_with_colors_texture_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_16mesh_core_cython_9write_obj_with_colors_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  std::string __pyx_v_filename;
+  std::string __pyx_v_mtl_name;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_colors = 0;
+  PyArrayObject *__pyx_v_uv_coords = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_ntexver;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("write_obj_with_colors_texture_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_filename,&__pyx_n_s_mtl_name,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_colors,&__pyx_n_s_uv_coords,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_ntexver,0};
+    PyObject* values[9] = {0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_filename)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_mtl_name)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 1); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 2); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 3); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_colors)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 4); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_uv_coords)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 5); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 6); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 7); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntexver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 8); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "write_obj_with_colors_texture_core") < 0)) __PYX_ERR(0, 100, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 9) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+    }
+    __pyx_v_filename = __pyx_convert_string_from_py_std__in_string(values[0]); if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 100, __pyx_L3_error)
+    __pyx_v_mtl_name = __pyx_convert_string_from_py_std__in_string(values[1]); if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 100, __pyx_L3_error)
+    __pyx_v_vertices = ((PyArrayObject *)values[2]);
+    __pyx_v_triangles = ((PyArrayObject *)values[3]);
+    __pyx_v_colors = ((PyArrayObject *)values[4]);
+    __pyx_v_uv_coords = ((PyArrayObject *)values[5]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+    __pyx_v_ntexver = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_ntexver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 100, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("mesh_core_cython.write_obj_with_colors_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 101, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 102, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_colors), __pyx_ptype_5numpy_ndarray, 0, "colors", 0))) __PYX_ERR(0, 103, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_uv_coords), __pyx_ptype_5numpy_ndarray, 0, "uv_coords", 0))) __PYX_ERR(0, 104, __pyx_L1_error)
+  __pyx_r = __pyx_pf_16mesh_core_cython_8write_obj_with_colors_texture_core(__pyx_self, __pyx_v_filename, __pyx_v_mtl_name, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_colors, __pyx_v_uv_coords, __pyx_v_nver, __pyx_v_ntri, __pyx_v_ntexver);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_16mesh_core_cython_8write_obj_with_colors_texture_core(CYTHON_UNUSED PyObject *__pyx_self, std::string __pyx_v_filename, std::string __pyx_v_mtl_name, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_uv_coords, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_ntexver) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_colors;
+  __Pyx_Buffer __pyx_pybuffer_colors;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_uv_coords;
+  __Pyx_Buffer __pyx_pybuffer_uv_coords;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("write_obj_with_colors_texture_core", 0);
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_colors.pybuffer.buf = NULL;
+  __pyx_pybuffer_colors.refcount = 0;
+  __pyx_pybuffernd_colors.data = NULL;
+  __pyx_pybuffernd_colors.rcbuffer = &__pyx_pybuffer_colors;
+  __pyx_pybuffer_uv_coords.pybuffer.buf = NULL;
+  __pyx_pybuffer_uv_coords.refcount = 0;
+  __pyx_pybuffernd_uv_coords.data = NULL;
+  __pyx_pybuffernd_uv_coords.rcbuffer = &__pyx_pybuffer_uv_coords;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_colors.rcbuffer->pybuffer, (PyObject*)__pyx_v_colors, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_colors.diminfo[0].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_colors.diminfo[0].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_colors.diminfo[1].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_colors.diminfo[1].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer, (PyObject*)__pyx_v_uv_coords, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_uv_coords.diminfo[0].strides = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_uv_coords.diminfo[0].shape = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_uv_coords.diminfo[1].strides = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_uv_coords.diminfo[1].shape = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.shape[1];
+
+  /* "mesh_core_cython.pyx":107
+ *                 int nver, int ntri, int ntexver
+ *                 ):
+ *     _write_obj_with_colors_texture(filename, mtl_name,             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles), <float*> np.PyArray_DATA(colors), <float*> np.PyArray_DATA(uv_coords),
+ *         nver, ntri, ntexver)
+ */
+  _write_obj_with_colors_texture(__pyx_v_filename, __pyx_v_mtl_name, ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_colors))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_uv_coords))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_ntexver);
+
+  /* "mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("mesh_core_cython.write_obj_with_colors_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":215
+ *         # experimental exception made for __getbuffer__ and __releasebuffer__
+ *         # -- the details of this may change.
+ *         def __getbuffer__(ndarray self, Py_buffer* info, int flags):             # <<<<<<<<<<<<<<
+ *             # This implementation of getbuffer is geared towards Cython
+ *             # requirements, and does not yet fulfill the PEP.
+ */
+
+/* Python wrapper */
+static CYTHON_UNUSED int __pyx_pw_5numpy_7ndarray_1__getbuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/
+static CYTHON_UNUSED int __pyx_pw_5numpy_7ndarray_1__getbuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0);
+  __pyx_r = __pyx_pf_5numpy_7ndarray___getbuffer__(((PyArrayObject *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags));
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static int __pyx_pf_5numpy_7ndarray___getbuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) {
+  int __pyx_v_i;
+  int __pyx_v_ndim;
+  int __pyx_v_endian_detector;
+  int __pyx_v_little_endian;
+  int __pyx_v_t;
+  char *__pyx_v_f;
+  PyArray_Descr *__pyx_v_descr = 0;
+  int __pyx_v_offset;
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  int __pyx_t_2;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  int __pyx_t_5;
+  int __pyx_t_6;
+  PyObject *__pyx_t_7 = NULL;
+  char *__pyx_t_8;
+  if (__pyx_v_info == NULL) {
+    PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete");
+    return -1;
+  }
+  __Pyx_RefNannySetupContext("__getbuffer__", 0);
+  __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None);
+  __Pyx_GIVEREF(__pyx_v_info->obj);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":222
+ * 
+ *             cdef int i, ndim
+ *             cdef int endian_detector = 1             # <<<<<<<<<<<<<<
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ * 
+ */
+  __pyx_v_endian_detector = 1;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":223
+ *             cdef int i, ndim
+ *             cdef int endian_detector = 1
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)             # <<<<<<<<<<<<<<
+ * 
+ *             ndim = PyArray_NDIM(self)
+ */
+  __pyx_v_little_endian = ((((char *)(&__pyx_v_endian_detector))[0]) != 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":225
+ *             cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ * 
+ *             ndim = PyArray_NDIM(self)             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ */
+  __pyx_v_ndim = PyArray_NDIM(__pyx_v_self);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":227
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  __pyx_t_2 = (((__pyx_v_flags & PyBUF_C_CONTIGUOUS) == PyBUF_C_CONTIGUOUS) != 0);
+  if (__pyx_t_2) {
+  } else {
+    __pyx_t_1 = __pyx_t_2;
+    goto __pyx_L4_bool_binop_done;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":228
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):             # <<<<<<<<<<<<<<
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ */
+  __pyx_t_2 = ((!(PyArray_CHKFLAGS(__pyx_v_self, NPY_C_CONTIGUOUS) != 0)) != 0);
+  __pyx_t_1 = __pyx_t_2;
+  __pyx_L4_bool_binop_done:;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":227
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  if (unlikely(__pyx_t_1)) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":229
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ */
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple_, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 229, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __PYX_ERR(1, 229, __pyx_L1_error)
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":227
+ *             ndim = PyArray_NDIM(self)
+ * 
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":231
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  __pyx_t_2 = (((__pyx_v_flags & PyBUF_F_CONTIGUOUS) == PyBUF_F_CONTIGUOUS) != 0);
+  if (__pyx_t_2) {
+  } else {
+    __pyx_t_1 = __pyx_t_2;
+    goto __pyx_L7_bool_binop_done;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":232
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):             # <<<<<<<<<<<<<<
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ * 
+ */
+  __pyx_t_2 = ((!(PyArray_CHKFLAGS(__pyx_v_self, NPY_F_CONTIGUOUS) != 0)) != 0);
+  __pyx_t_1 = __pyx_t_2;
+  __pyx_L7_bool_binop_done:;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":231
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  if (unlikely(__pyx_t_1)) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":233
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             info.buf = PyArray_DATA(self)
+ */
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__2, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 233, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __PYX_ERR(1, 233, __pyx_L1_error)
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":231
+ *                 raise ValueError(u"ndarray is not C contiguous")
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)             # <<<<<<<<<<<<<<
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":235
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")
+ * 
+ *             info.buf = PyArray_DATA(self)             # <<<<<<<<<<<<<<
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  __pyx_v_info->buf = PyArray_DATA(__pyx_v_self);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":236
+ * 
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim             # <<<<<<<<<<<<<<
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 # Allocate new buffer for strides and shape info.
+ */
+  __pyx_v_info->ndim = __pyx_v_ndim;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":237
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ */
+  __pyx_t_1 = (((sizeof(npy_intp)) != (sizeof(Py_ssize_t))) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":240
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)             # <<<<<<<<<<<<<<
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):
+ */
+    __pyx_v_info->strides = ((Py_ssize_t *)PyObject_Malloc((((sizeof(Py_ssize_t)) * 2) * ((size_t)__pyx_v_ndim))));
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":241
+ *                 # This is allocated as one block, strides first.
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)
+ *                 info.shape = info.strides + ndim             # <<<<<<<<<<<<<<
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ */
+    __pyx_v_info->shape = (__pyx_v_info->strides + __pyx_v_ndim);
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":242
+ *                 info.strides = <Py_ssize_t*>PyObject_Malloc(sizeof(Py_ssize_t) * 2 * <size_t>ndim)
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):             # <<<<<<<<<<<<<<
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ */
+    __pyx_t_4 = __pyx_v_ndim;
+    __pyx_t_5 = __pyx_t_4;
+    for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
+      __pyx_v_i = __pyx_t_6;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":243
+ *                 info.shape = info.strides + ndim
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]             # <<<<<<<<<<<<<<
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ *             else:
+ */
+      (__pyx_v_info->strides[__pyx_v_i]) = (PyArray_STRIDES(__pyx_v_self)[__pyx_v_i]);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":244
+ *                 for i in range(ndim):
+ *                     info.strides[i] = PyArray_STRIDES(self)[i]
+ *                     info.shape[i] = PyArray_DIMS(self)[i]             # <<<<<<<<<<<<<<
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ */
+      (__pyx_v_info->shape[__pyx_v_i]) = (PyArray_DIMS(__pyx_v_self)[__pyx_v_i]);
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":237
+ *             info.buf = PyArray_DATA(self)
+ *             info.ndim = ndim
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 # Allocate new buffer for strides and shape info.
+ *                 # This is allocated as one block, strides first.
+ */
+    goto __pyx_L9;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":246
+ *                     info.shape[i] = PyArray_DIMS(self)[i]
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)             # <<<<<<<<<<<<<<
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL
+ */
+  /*else*/ {
+    __pyx_v_info->strides = ((Py_ssize_t *)PyArray_STRIDES(__pyx_v_self));
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":247
+ *             else:
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)             # <<<<<<<<<<<<<<
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ */
+    __pyx_v_info->shape = ((Py_ssize_t *)PyArray_DIMS(__pyx_v_self));
+  }
+  __pyx_L9:;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":248
+ *                 info.strides = <Py_ssize_t*>PyArray_STRIDES(self)
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL             # <<<<<<<<<<<<<<
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ *             info.readonly = not PyArray_ISWRITEABLE(self)
+ */
+  __pyx_v_info->suboffsets = NULL;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":249
+ *                 info.shape = <Py_ssize_t*>PyArray_DIMS(self)
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)             # <<<<<<<<<<<<<<
+ *             info.readonly = not PyArray_ISWRITEABLE(self)
+ * 
+ */
+  __pyx_v_info->itemsize = PyArray_ITEMSIZE(__pyx_v_self);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":250
+ *             info.suboffsets = NULL
+ *             info.itemsize = PyArray_ITEMSIZE(self)
+ *             info.readonly = not PyArray_ISWRITEABLE(self)             # <<<<<<<<<<<<<<
+ * 
+ *             cdef int t
+ */
+  __pyx_v_info->readonly = (!(PyArray_ISWRITEABLE(__pyx_v_self) != 0));
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":253
+ * 
+ *             cdef int t
+ *             cdef char* f = NULL             # <<<<<<<<<<<<<<
+ *             cdef dtype descr = self.descr
+ *             cdef int offset
+ */
+  __pyx_v_f = NULL;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":254
+ *             cdef int t
+ *             cdef char* f = NULL
+ *             cdef dtype descr = self.descr             # <<<<<<<<<<<<<<
+ *             cdef int offset
+ * 
+ */
+  __pyx_t_3 = ((PyObject *)__pyx_v_self->descr);
+  __Pyx_INCREF(__pyx_t_3);
+  __pyx_v_descr = ((PyArray_Descr *)__pyx_t_3);
+  __pyx_t_3 = 0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":257
+ *             cdef int offset
+ * 
+ *             info.obj = self             # <<<<<<<<<<<<<<
+ * 
+ *             if not PyDataType_HASFIELDS(descr):
+ */
+  __Pyx_INCREF(((PyObject *)__pyx_v_self));
+  __Pyx_GIVEREF(((PyObject *)__pyx_v_self));
+  __Pyx_GOTREF(__pyx_v_info->obj);
+  __Pyx_DECREF(__pyx_v_info->obj);
+  __pyx_v_info->obj = ((PyObject *)__pyx_v_self);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":259
+ *             info.obj = self
+ * 
+ *             if not PyDataType_HASFIELDS(descr):             # <<<<<<<<<<<<<<
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ */
+  __pyx_t_1 = ((!(PyDataType_HASFIELDS(__pyx_v_descr) != 0)) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":260
+ * 
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num             # <<<<<<<<<<<<<<
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ */
+    __pyx_t_4 = __pyx_v_descr->type_num;
+    __pyx_v_t = __pyx_t_4;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":261
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    __pyx_t_2 = ((__pyx_v_descr->byteorder == '>') != 0);
+    if (!__pyx_t_2) {
+      goto __pyx_L15_next_or;
+    } else {
+    }
+    __pyx_t_2 = (__pyx_v_little_endian != 0);
+    if (!__pyx_t_2) {
+    } else {
+      __pyx_t_1 = __pyx_t_2;
+      goto __pyx_L14_bool_binop_done;
+    }
+    __pyx_L15_next_or:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":262
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):             # <<<<<<<<<<<<<<
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"
+ */
+    __pyx_t_2 = ((__pyx_v_descr->byteorder == '<') != 0);
+    if (__pyx_t_2) {
+    } else {
+      __pyx_t_1 = __pyx_t_2;
+      goto __pyx_L14_bool_binop_done;
+    }
+    __pyx_t_2 = ((!(__pyx_v_little_endian != 0)) != 0);
+    __pyx_t_1 = __pyx_t_2;
+    __pyx_L14_bool_binop_done:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":261
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    if (unlikely(__pyx_t_1)) {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":263
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__3, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 263, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 263, __pyx_L1_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":261
+ *             if not PyDataType_HASFIELDS(descr):
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ */
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":264
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"
+ */
+    switch (__pyx_v_t) {
+      case NPY_BYTE:
+      __pyx_v_f = ((char *)"b");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":265
+ *                     raise ValueError(u"Non-native byte order not supported")
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"
+ */
+      case NPY_UBYTE:
+      __pyx_v_f = ((char *)"B");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":266
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"
+ */
+      case NPY_SHORT:
+      __pyx_v_f = ((char *)"h");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":267
+ *                 elif t == NPY_UBYTE:       f = "B"
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"
+ */
+      case NPY_USHORT:
+      __pyx_v_f = ((char *)"H");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":268
+ *                 elif t == NPY_SHORT:       f = "h"
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"
+ */
+      case NPY_INT:
+      __pyx_v_f = ((char *)"i");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":269
+ *                 elif t == NPY_USHORT:      f = "H"
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"
+ */
+      case NPY_UINT:
+      __pyx_v_f = ((char *)"I");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":270
+ *                 elif t == NPY_INT:         f = "i"
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ */
+      case NPY_LONG:
+      __pyx_v_f = ((char *)"l");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":271
+ *                 elif t == NPY_UINT:        f = "I"
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ */
+      case NPY_ULONG:
+      __pyx_v_f = ((char *)"L");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":272
+ *                 elif t == NPY_LONG:        f = "l"
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"
+ */
+      case NPY_LONGLONG:
+      __pyx_v_f = ((char *)"q");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":273
+ *                 elif t == NPY_ULONG:       f = "L"
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ */
+      case NPY_ULONGLONG:
+      __pyx_v_f = ((char *)"Q");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":274
+ *                 elif t == NPY_LONGLONG:    f = "q"
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ */
+      case NPY_FLOAT:
+      __pyx_v_f = ((char *)"f");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":275
+ *                 elif t == NPY_ULONGLONG:   f = "Q"
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ */
+      case NPY_DOUBLE:
+      __pyx_v_f = ((char *)"d");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":276
+ *                 elif t == NPY_FLOAT:       f = "f"
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ */
+      case NPY_LONGDOUBLE:
+      __pyx_v_f = ((char *)"g");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":277
+ *                 elif t == NPY_DOUBLE:      f = "d"
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ */
+      case NPY_CFLOAT:
+      __pyx_v_f = ((char *)"Zf");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":278
+ *                 elif t == NPY_LONGDOUBLE:  f = "g"
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ *                 elif t == NPY_OBJECT:      f = "O"
+ */
+      case NPY_CDOUBLE:
+      __pyx_v_f = ((char *)"Zd");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":279
+ *                 elif t == NPY_CFLOAT:      f = "Zf"
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"             # <<<<<<<<<<<<<<
+ *                 elif t == NPY_OBJECT:      f = "O"
+ *                 else:
+ */
+      case NPY_CLONGDOUBLE:
+      __pyx_v_f = ((char *)"Zg");
+      break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":280
+ *                 elif t == NPY_CDOUBLE:     f = "Zd"
+ *                 elif t == NPY_CLONGDOUBLE: f = "Zg"
+ *                 elif t == NPY_OBJECT:      f = "O"             # <<<<<<<<<<<<<<
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ */
+      case NPY_OBJECT:
+      __pyx_v_f = ((char *)"O");
+      break;
+      default:
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":282
+ *                 elif t == NPY_OBJECT:      f = "O"
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)             # <<<<<<<<<<<<<<
+ *                 info.format = f
+ *                 return
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_t); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 282, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_7 = PyUnicode_Format(__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_t_3); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 282, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_7);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_7); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 282, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0;
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 282, __pyx_L1_error)
+      break;
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":283
+ *                 else:
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *                 info.format = f             # <<<<<<<<<<<<<<
+ *                 return
+ *             else:
+ */
+    __pyx_v_info->format = __pyx_v_f;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":284
+ *                     raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *                 info.format = f
+ *                 return             # <<<<<<<<<<<<<<
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ */
+    __pyx_r = 0;
+    goto __pyx_L0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":259
+ *             info.obj = self
+ * 
+ *             if not PyDataType_HASFIELDS(descr):             # <<<<<<<<<<<<<<
+ *                 t = descr.type_num
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":286
+ *                 return
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)             # <<<<<<<<<<<<<<
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0
+ */
+  /*else*/ {
+    __pyx_v_info->format = ((char *)PyObject_Malloc(0xFF));
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":287
+ *             else:
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ *                 info.format[0] = c'^' # Native data types, manual alignment             # <<<<<<<<<<<<<<
+ *                 offset = 0
+ *                 f = _util_dtypestring(descr, info.format + 1,
+ */
+    (__pyx_v_info->format[0]) = '^';
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":288
+ *                 info.format = <char*>PyObject_Malloc(_buffer_format_string_len)
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0             # <<<<<<<<<<<<<<
+ *                 f = _util_dtypestring(descr, info.format + 1,
+ *                                       info.format + _buffer_format_string_len,
+ */
+    __pyx_v_offset = 0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":289
+ *                 info.format[0] = c'^' # Native data types, manual alignment
+ *                 offset = 0
+ *                 f = _util_dtypestring(descr, info.format + 1,             # <<<<<<<<<<<<<<
+ *                                       info.format + _buffer_format_string_len,
+ *                                       &offset)
+ */
+    __pyx_t_8 = __pyx_f_5numpy__util_dtypestring(__pyx_v_descr, (__pyx_v_info->format + 1), (__pyx_v_info->format + 0xFF), (&__pyx_v_offset)); if (unlikely(__pyx_t_8 == ((char *)NULL))) __PYX_ERR(1, 289, __pyx_L1_error)
+    __pyx_v_f = __pyx_t_8;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":292
+ *                                       info.format + _buffer_format_string_len,
+ *                                       &offset)
+ *                 f[0] = c'\0' # Terminate format string             # <<<<<<<<<<<<<<
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ */
+    (__pyx_v_f[0]) = '\x00';
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":215
+ *         # experimental exception made for __getbuffer__ and __releasebuffer__
+ *         # -- the details of this may change.
+ *         def __getbuffer__(ndarray self, Py_buffer* info, int flags):             # <<<<<<<<<<<<<<
+ *             # This implementation of getbuffer is geared towards Cython
+ *             # requirements, and does not yet fulfill the PEP.
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_3);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_AddTraceback("numpy.ndarray.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  if (__pyx_v_info->obj != NULL) {
+    __Pyx_GOTREF(__pyx_v_info->obj);
+    __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0;
+  }
+  goto __pyx_L2;
+  __pyx_L0:;
+  if (__pyx_v_info->obj == Py_None) {
+    __Pyx_GOTREF(__pyx_v_info->obj);
+    __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0;
+  }
+  __pyx_L2:;
+  __Pyx_XDECREF((PyObject *)__pyx_v_descr);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":294
+ *                 f[0] = c'\0' # Terminate format string
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):             # <<<<<<<<<<<<<<
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ */
+
+/* Python wrapper */
+static CYTHON_UNUSED void __pyx_pw_5numpy_7ndarray_3__releasebuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info); /*proto*/
+static CYTHON_UNUSED void __pyx_pw_5numpy_7ndarray_3__releasebuffer__(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__releasebuffer__ (wrapper)", 0);
+  __pyx_pf_5numpy_7ndarray_2__releasebuffer__(((PyArrayObject *)__pyx_v_self), ((Py_buffer *)__pyx_v_info));
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+static void __pyx_pf_5numpy_7ndarray_2__releasebuffer__(PyArrayObject *__pyx_v_self, Py_buffer *__pyx_v_info) {
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("__releasebuffer__", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":295
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  __pyx_t_1 = (PyArray_HASFIELDS(__pyx_v_self) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":296
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)             # <<<<<<<<<<<<<<
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 PyObject_Free(info.strides)
+ */
+    PyObject_Free(__pyx_v_info->format);
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":295
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):
+ *             if PyArray_HASFIELDS(self):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":297
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.strides)
+ *                 # info.shape was stored after info.strides in the same block
+ */
+  __pyx_t_1 = (((sizeof(npy_intp)) != (sizeof(Py_ssize_t))) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":298
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):
+ *                 PyObject_Free(info.strides)             # <<<<<<<<<<<<<<
+ *                 # info.shape was stored after info.strides in the same block
+ * 
+ */
+    PyObject_Free(__pyx_v_info->strides);
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":297
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ *             if sizeof(npy_intp) != sizeof(Py_ssize_t):             # <<<<<<<<<<<<<<
+ *                 PyObject_Free(info.strides)
+ *                 # info.shape was stored after info.strides in the same block
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":294
+ *                 f[0] = c'\0' # Terminate format string
+ * 
+ *         def __releasebuffer__(ndarray self, Py_buffer* info):             # <<<<<<<<<<<<<<
+ *             if PyArray_HASFIELDS(self):
+ *                 PyObject_Free(info.format)
+ */
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":775
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew1(PyObject *__pyx_v_a) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew1", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":776
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ *     return PyArray_MultiIterNew(1, <void*>a)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(1, ((void *)__pyx_v_a)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 776, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":775
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew1", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":778
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew2(PyObject *__pyx_v_a, PyObject *__pyx_v_b) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew2", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":779
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(2, ((void *)__pyx_v_a), ((void *)__pyx_v_b)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 779, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":778
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew2", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":781
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew3(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew3", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":782
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(3, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 782, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":781
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew3", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":784
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew4(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew4", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":785
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(4, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 785, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":784
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew4", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":787
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew5(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d, PyObject *__pyx_v_e) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew5", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":788
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(5, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d), ((void *)__pyx_v_e)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 788, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":787
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew5", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":790
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyDataType_SHAPE(PyArray_Descr *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("PyDataType_SHAPE", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":791
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  __pyx_t_1 = (PyDataType_HASSUBARRAY(__pyx_v_d) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":792
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape             # <<<<<<<<<<<<<<
+ *     else:
+ *         return ()
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(((PyObject*)__pyx_v_d->subarray->shape));
+    __pyx_r = ((PyObject*)__pyx_v_d->subarray->shape);
+    goto __pyx_L0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":791
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":794
+ *         return <tuple>d.subarray.shape
+ *     else:
+ *         return ()             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:
+ */
+  /*else*/ {
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(__pyx_empty_tuple);
+    __pyx_r = __pyx_empty_tuple;
+    goto __pyx_L0;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":790
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":796
+ *         return ()
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:             # <<<<<<<<<<<<<<
+ *     # Recursive utility function used in __getbuffer__ to get format
+ *     # string. The new location in the format string is returned.
+ */
+
+static CYTHON_INLINE char *__pyx_f_5numpy__util_dtypestring(PyArray_Descr *__pyx_v_descr, char *__pyx_v_f, char *__pyx_v_end, int *__pyx_v_offset) {
+  PyArray_Descr *__pyx_v_child = 0;
+  int __pyx_v_endian_detector;
+  int __pyx_v_little_endian;
+  PyObject *__pyx_v_fields = 0;
+  PyObject *__pyx_v_childname = NULL;
+  PyObject *__pyx_v_new_offset = NULL;
+  PyObject *__pyx_v_t = NULL;
+  char *__pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  Py_ssize_t __pyx_t_2;
+  PyObject *__pyx_t_3 = NULL;
+  PyObject *__pyx_t_4 = NULL;
+  int __pyx_t_5;
+  int __pyx_t_6;
+  int __pyx_t_7;
+  long __pyx_t_8;
+  char *__pyx_t_9;
+  __Pyx_RefNannySetupContext("_util_dtypestring", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":801
+ * 
+ *     cdef dtype child
+ *     cdef int endian_detector = 1             # <<<<<<<<<<<<<<
+ *     cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)
+ *     cdef tuple fields
+ */
+  __pyx_v_endian_detector = 1;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":802
+ *     cdef dtype child
+ *     cdef int endian_detector = 1
+ *     cdef bint little_endian = ((<char*>&endian_detector)[0] != 0)             # <<<<<<<<<<<<<<
+ *     cdef tuple fields
+ * 
+ */
+  __pyx_v_little_endian = ((((char *)(&__pyx_v_endian_detector))[0]) != 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":805
+ *     cdef tuple fields
+ * 
+ *     for childname in descr.names:             # <<<<<<<<<<<<<<
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields
+ */
+  if (unlikely(__pyx_v_descr->names == Py_None)) {
+    PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable");
+    __PYX_ERR(1, 805, __pyx_L1_error)
+  }
+  __pyx_t_1 = __pyx_v_descr->names; __Pyx_INCREF(__pyx_t_1); __pyx_t_2 = 0;
+  for (;;) {
+    if (__pyx_t_2 >= PyTuple_GET_SIZE(__pyx_t_1)) break;
+    #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+    __pyx_t_3 = PyTuple_GET_ITEM(__pyx_t_1, __pyx_t_2); __Pyx_INCREF(__pyx_t_3); __pyx_t_2++; if (unlikely(0 < 0)) __PYX_ERR(1, 805, __pyx_L1_error)
+    #else
+    __pyx_t_3 = PySequence_ITEM(__pyx_t_1, __pyx_t_2); __pyx_t_2++; if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 805, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    #endif
+    __Pyx_XDECREF_SET(__pyx_v_childname, __pyx_t_3);
+    __pyx_t_3 = 0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":806
+ * 
+ *     for childname in descr.names:
+ *         fields = descr.fields[childname]             # <<<<<<<<<<<<<<
+ *         child, new_offset = fields
+ * 
+ */
+    if (unlikely(__pyx_v_descr->fields == Py_None)) {
+      PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable");
+      __PYX_ERR(1, 806, __pyx_L1_error)
+    }
+    __pyx_t_3 = __Pyx_PyDict_GetItem(__pyx_v_descr->fields, __pyx_v_childname); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 806, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    if (!(likely(PyTuple_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "tuple", Py_TYPE(__pyx_t_3)->tp_name), 0))) __PYX_ERR(1, 806, __pyx_L1_error)
+    __Pyx_XDECREF_SET(__pyx_v_fields, ((PyObject*)__pyx_t_3));
+    __pyx_t_3 = 0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":807
+ *     for childname in descr.names:
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields             # <<<<<<<<<<<<<<
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ */
+    if (likely(__pyx_v_fields != Py_None)) {
+      PyObject* sequence = __pyx_v_fields;
+      Py_ssize_t size = __Pyx_PySequence_SIZE(sequence);
+      if (unlikely(size != 2)) {
+        if (size > 2) __Pyx_RaiseTooManyValuesError(2);
+        else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size);
+        __PYX_ERR(1, 807, __pyx_L1_error)
+      }
+      #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
+      __pyx_t_3 = PyTuple_GET_ITEM(sequence, 0); 
+      __pyx_t_4 = PyTuple_GET_ITEM(sequence, 1); 
+      __Pyx_INCREF(__pyx_t_3);
+      __Pyx_INCREF(__pyx_t_4);
+      #else
+      __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 807, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 807, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      #endif
+    } else {
+      __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 807, __pyx_L1_error)
+    }
+    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_ptype_5numpy_dtype))))) __PYX_ERR(1, 807, __pyx_L1_error)
+    __Pyx_XDECREF_SET(__pyx_v_child, ((PyArray_Descr *)__pyx_t_3));
+    __pyx_t_3 = 0;
+    __Pyx_XDECREF_SET(__pyx_v_new_offset, __pyx_t_4);
+    __pyx_t_4 = 0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":809
+ *         child, new_offset = fields
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:             # <<<<<<<<<<<<<<
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyInt_From_int((__pyx_v_offset[0])); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 809, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_4);
+    __pyx_t_3 = PyNumber_Subtract(__pyx_v_new_offset, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 809, __pyx_L1_error)
+    __Pyx_GOTREF(__pyx_t_3);
+    __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+    __pyx_t_5 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_5 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 809, __pyx_L1_error)
+    __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+    __pyx_t_6 = ((((__pyx_v_end - __pyx_v_f) - ((int)__pyx_t_5)) < 15) != 0);
+    if (unlikely(__pyx_t_6)) {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":810
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")             # <<<<<<<<<<<<<<
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_RuntimeError, __pyx_tuple__4, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 810, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 810, __pyx_L1_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":809
+ *         child, new_offset = fields
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:             # <<<<<<<<<<<<<<
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ */
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":812
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    __pyx_t_7 = ((__pyx_v_child->byteorder == '>') != 0);
+    if (!__pyx_t_7) {
+      goto __pyx_L8_next_or;
+    } else {
+    }
+    __pyx_t_7 = (__pyx_v_little_endian != 0);
+    if (!__pyx_t_7) {
+    } else {
+      __pyx_t_6 = __pyx_t_7;
+      goto __pyx_L7_bool_binop_done;
+    }
+    __pyx_L8_next_or:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":813
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ *             (child.byteorder == c'<' and not little_endian)):             # <<<<<<<<<<<<<<
+ *             raise ValueError(u"Non-native byte order not supported")
+ *             # One could encode it in the format string and have Cython
+ */
+    __pyx_t_7 = ((__pyx_v_child->byteorder == '<') != 0);
+    if (__pyx_t_7) {
+    } else {
+      __pyx_t_6 = __pyx_t_7;
+      goto __pyx_L7_bool_binop_done;
+    }
+    __pyx_t_7 = ((!(__pyx_v_little_endian != 0)) != 0);
+    __pyx_t_6 = __pyx_t_7;
+    __pyx_L7_bool_binop_done:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":812
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    if (unlikely(__pyx_t_6)) {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":814
+ *         if ((child.byteorder == c'>' and little_endian) or
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *             # One could encode it in the format string and have Cython
+ *             # complain instead, BUT: < and > in format strings also imply
+ */
+      __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__5, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 814, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __Pyx_Raise(__pyx_t_3, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __PYX_ERR(1, 814, __pyx_L1_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":812
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or             # <<<<<<<<<<<<<<
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")
+ */
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":824
+ * 
+ *         # Output padding bytes
+ *         while offset[0] < new_offset:             # <<<<<<<<<<<<<<
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1
+ */
+    while (1) {
+      __pyx_t_3 = __Pyx_PyInt_From_int((__pyx_v_offset[0])); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 824, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_t_3, __pyx_v_new_offset, Py_LT); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 824, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 824, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (!__pyx_t_6) break;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":825
+ *         # Output padding bytes
+ *         while offset[0] < new_offset:
+ *             f[0] = 120 # "x"; pad byte             # <<<<<<<<<<<<<<
+ *             f += 1
+ *             offset[0] += 1
+ */
+      (__pyx_v_f[0]) = 0x78;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":826
+ *         while offset[0] < new_offset:
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1             # <<<<<<<<<<<<<<
+ *             offset[0] += 1
+ * 
+ */
+      __pyx_v_f = (__pyx_v_f + 1);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":827
+ *             f[0] = 120 # "x"; pad byte
+ *             f += 1
+ *             offset[0] += 1             # <<<<<<<<<<<<<<
+ * 
+ *         offset[0] += child.itemsize
+ */
+      __pyx_t_8 = 0;
+      (__pyx_v_offset[__pyx_t_8]) = ((__pyx_v_offset[__pyx_t_8]) + 1);
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":829
+ *             offset[0] += 1
+ * 
+ *         offset[0] += child.itemsize             # <<<<<<<<<<<<<<
+ * 
+ *         if not PyDataType_HASFIELDS(child):
+ */
+    __pyx_t_8 = 0;
+    (__pyx_v_offset[__pyx_t_8]) = ((__pyx_v_offset[__pyx_t_8]) + __pyx_v_child->elsize);
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":831
+ *         offset[0] += child.itemsize
+ * 
+ *         if not PyDataType_HASFIELDS(child):             # <<<<<<<<<<<<<<
+ *             t = child.type_num
+ *             if end - f < 5:
+ */
+    __pyx_t_6 = ((!(PyDataType_HASFIELDS(__pyx_v_child) != 0)) != 0);
+    if (__pyx_t_6) {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":832
+ * 
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num             # <<<<<<<<<<<<<<
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_int(__pyx_v_child->type_num); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 832, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __Pyx_XDECREF_SET(__pyx_v_t, __pyx_t_4);
+      __pyx_t_4 = 0;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":833
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num
+ *             if end - f < 5:             # <<<<<<<<<<<<<<
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ * 
+ */
+      __pyx_t_6 = (((__pyx_v_end - __pyx_v_f) < 5) != 0);
+      if (unlikely(__pyx_t_6)) {
+
+        /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":834
+ *             t = child.type_num
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")             # <<<<<<<<<<<<<<
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ */
+        __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin_RuntimeError, __pyx_tuple__6, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 834, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_4);
+        __Pyx_Raise(__pyx_t_4, 0, 0, 0);
+        __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+        __PYX_ERR(1, 834, __pyx_L1_error)
+
+        /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":833
+ *         if not PyDataType_HASFIELDS(child):
+ *             t = child.type_num
+ *             if end - f < 5:             # <<<<<<<<<<<<<<
+ *                 raise RuntimeError(u"Format string allocated too short.")
+ * 
+ */
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":837
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_BYTE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 837, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 837, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 837, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 98;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":838
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_UBYTE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 838, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 838, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 838, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 66;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":839
+ *             if   t == NPY_BYTE:        f[0] =  98 #"b"
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_SHORT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 839, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 839, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 839, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x68;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":840
+ *             elif t == NPY_UBYTE:       f[0] =  66 #"B"
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_USHORT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 840, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 840, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 840, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 72;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":841
+ *             elif t == NPY_SHORT:       f[0] = 104 #"h"
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_INT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 841, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 841, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 841, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x69;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":842
+ *             elif t == NPY_USHORT:      f[0] =  72 #"H"
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_UINT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 842, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 842, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 842, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 73;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":843
+ *             elif t == NPY_INT:         f[0] = 105 #"i"
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONG); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 843, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 843, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 843, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x6C;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":844
+ *             elif t == NPY_UINT:        f[0] =  73 #"I"
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_ULONG); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 844, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 844, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 844, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 76;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":845
+ *             elif t == NPY_LONG:        f[0] = 108 #"l"
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONGLONG); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 845, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 845, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 845, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x71;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":846
+ *             elif t == NPY_ULONG:       f[0] = 76  #"L"
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_ULONGLONG); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 846, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 846, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 846, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 81;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":847
+ *             elif t == NPY_LONGLONG:    f[0] = 113 #"q"
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_FLOAT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 847, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 847, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 847, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x66;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":848
+ *             elif t == NPY_ULONGLONG:   f[0] = 81  #"Q"
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_DOUBLE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 848, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 848, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 848, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x64;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":849
+ *             elif t == NPY_FLOAT:       f[0] = 102 #"f"
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_LONGDOUBLE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 849, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 849, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 849, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 0x67;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":850
+ *             elif t == NPY_DOUBLE:      f[0] = 100 #"d"
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CFLOAT); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 850, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 850, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 850, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x66;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":851
+ *             elif t == NPY_LONGDOUBLE:  f[0] = 103 #"g"
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd             # <<<<<<<<<<<<<<
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CDOUBLE); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 851, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 851, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 851, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x64;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":852
+ *             elif t == NPY_CFLOAT:      f[0] = 90; f[1] = 102; f += 1 # Zf
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg             # <<<<<<<<<<<<<<
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ *             else:
+ */
+      __pyx_t_3 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_CLONGDOUBLE); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 852, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_3);
+      __pyx_t_4 = PyObject_RichCompare(__pyx_v_t, __pyx_t_3, Py_EQ); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 852, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 852, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      if (__pyx_t_6) {
+        (__pyx_v_f[0]) = 90;
+        (__pyx_v_f[1]) = 0x67;
+        __pyx_v_f = (__pyx_v_f + 1);
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":853
+ *             elif t == NPY_CDOUBLE:     f[0] = 90; f[1] = 100; f += 1 # Zd
+ *             elif t == NPY_CLONGDOUBLE: f[0] = 90; f[1] = 103; f += 1 # Zg
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"             # <<<<<<<<<<<<<<
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ */
+      __pyx_t_4 = __Pyx_PyInt_From_enum__NPY_TYPES(NPY_OBJECT); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 853, __pyx_L1_error)
+      __Pyx_GOTREF(__pyx_t_4);
+      __pyx_t_3 = PyObject_RichCompare(__pyx_v_t, __pyx_t_4, Py_EQ); __Pyx_XGOTREF(__pyx_t_3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 853, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+      __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_3); if (unlikely(__pyx_t_6 < 0)) __PYX_ERR(1, 853, __pyx_L1_error)
+      __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+      if (likely(__pyx_t_6)) {
+        (__pyx_v_f[0]) = 79;
+        goto __pyx_L15;
+      }
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":855
+ *             elif t == NPY_OBJECT:      f[0] = 79 #"O"
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)             # <<<<<<<<<<<<<<
+ *             f += 1
+ *         else:
+ */
+      /*else*/ {
+        __pyx_t_3 = PyUnicode_Format(__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_v_t); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 855, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_3);
+        __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 855, __pyx_L1_error)
+        __Pyx_GOTREF(__pyx_t_4);
+        __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
+        __Pyx_Raise(__pyx_t_4, 0, 0, 0);
+        __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
+        __PYX_ERR(1, 855, __pyx_L1_error)
+      }
+      __pyx_L15:;
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":856
+ *             else:
+ *                 raise ValueError(u"unknown dtype code in numpy.pxd (%d)" % t)
+ *             f += 1             # <<<<<<<<<<<<<<
+ *         else:
+ *             # Cython ignores struct boundary information ("T{...}"),
+ */
+      __pyx_v_f = (__pyx_v_f + 1);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":831
+ *         offset[0] += child.itemsize
+ * 
+ *         if not PyDataType_HASFIELDS(child):             # <<<<<<<<<<<<<<
+ *             t = child.type_num
+ *             if end - f < 5:
+ */
+      goto __pyx_L13;
+    }
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":860
+ *             # Cython ignores struct boundary information ("T{...}"),
+ *             # so don't output it
+ *             f = _util_dtypestring(child, f, end, offset)             # <<<<<<<<<<<<<<
+ *     return f
+ * 
+ */
+    /*else*/ {
+      __pyx_t_9 = __pyx_f_5numpy__util_dtypestring(__pyx_v_child, __pyx_v_f, __pyx_v_end, __pyx_v_offset); if (unlikely(__pyx_t_9 == ((char *)NULL))) __PYX_ERR(1, 860, __pyx_L1_error)
+      __pyx_v_f = __pyx_t_9;
+    }
+    __pyx_L13:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":805
+ *     cdef tuple fields
+ * 
+ *     for childname in descr.names:             # <<<<<<<<<<<<<<
+ *         fields = descr.fields[childname]
+ *         child, new_offset = fields
+ */
+  }
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":861
+ *             # so don't output it
+ *             f = _util_dtypestring(child, f, end, offset)
+ *     return f             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = __pyx_v_f;
+  goto __pyx_L0;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":796
+ *         return ()
+ * 
+ * cdef inline char* _util_dtypestring(dtype descr, char* f, char* end, int* offset) except NULL:             # <<<<<<<<<<<<<<
+ *     # Recursive utility function used in __getbuffer__ to get format
+ *     # string. The new location in the format string is returned.
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_XDECREF(__pyx_t_3);
+  __Pyx_XDECREF(__pyx_t_4);
+  __Pyx_AddTraceback("numpy._util_dtypestring", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_XDECREF((PyObject *)__pyx_v_child);
+  __Pyx_XDECREF(__pyx_v_fields);
+  __Pyx_XDECREF(__pyx_v_childname);
+  __Pyx_XDECREF(__pyx_v_new_offset);
+  __Pyx_XDECREF(__pyx_v_t);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":977
+ * 
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *      cdef PyObject* baseptr
+ *      if base is None:
+ */
+
+static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_arr, PyObject *__pyx_v_base) {
+  PyObject *__pyx_v_baseptr;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  int __pyx_t_2;
+  __Pyx_RefNannySetupContext("set_array_base", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":979
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *      cdef PyObject* baseptr
+ *      if base is None:             # <<<<<<<<<<<<<<
+ *          baseptr = NULL
+ *      else:
+ */
+  __pyx_t_1 = (__pyx_v_base == Py_None);
+  __pyx_t_2 = (__pyx_t_1 != 0);
+  if (__pyx_t_2) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":980
+ *      cdef PyObject* baseptr
+ *      if base is None:
+ *          baseptr = NULL             # <<<<<<<<<<<<<<
+ *      else:
+ *          Py_INCREF(base) # important to do this before decref below!
+ */
+    __pyx_v_baseptr = NULL;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":979
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *      cdef PyObject* baseptr
+ *      if base is None:             # <<<<<<<<<<<<<<
+ *          baseptr = NULL
+ *      else:
+ */
+    goto __pyx_L3;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":982
+ *          baseptr = NULL
+ *      else:
+ *          Py_INCREF(base) # important to do this before decref below!             # <<<<<<<<<<<<<<
+ *          baseptr = <PyObject*>base
+ *      Py_XDECREF(arr.base)
+ */
+  /*else*/ {
+    Py_INCREF(__pyx_v_base);
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":983
+ *      else:
+ *          Py_INCREF(base) # important to do this before decref below!
+ *          baseptr = <PyObject*>base             # <<<<<<<<<<<<<<
+ *      Py_XDECREF(arr.base)
+ *      arr.base = baseptr
+ */
+    __pyx_v_baseptr = ((PyObject *)__pyx_v_base);
+  }
+  __pyx_L3:;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":984
+ *          Py_INCREF(base) # important to do this before decref below!
+ *          baseptr = <PyObject*>base
+ *      Py_XDECREF(arr.base)             # <<<<<<<<<<<<<<
+ *      arr.base = baseptr
+ * 
+ */
+  Py_XDECREF(__pyx_v_arr->base);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":985
+ *          baseptr = <PyObject*>base
+ *      Py_XDECREF(arr.base)
+ *      arr.base = baseptr             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ */
+  __pyx_v_arr->base = __pyx_v_baseptr;
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":977
+ * 
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *      cdef PyObject* baseptr
+ *      if base is None:
+ */
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":987
+ *      arr.base = baseptr
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     if arr.base is NULL:
+ *         return None
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__pyx_v_arr) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("get_array_base", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":988
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ *     if arr.base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     else:
+ */
+  __pyx_t_1 = ((__pyx_v_arr->base == NULL) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":989
+ * cdef inline object get_array_base(ndarray arr):
+ *     if arr.base is NULL:
+ *         return None             # <<<<<<<<<<<<<<
+ *     else:
+ *         return <object>arr.base
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+    goto __pyx_L0;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":988
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ *     if arr.base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     else:
+ */
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":991
+ *         return None
+ *     else:
+ *         return <object>arr.base             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  /*else*/ {
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(((PyObject *)__pyx_v_arr->base));
+    __pyx_r = ((PyObject *)__pyx_v_arr->base);
+    goto __pyx_L0;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":987
+ *      arr.base = baseptr
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     if arr.base is NULL:
+ *         return None
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":996
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_array()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_array(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_array", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":997
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":998
+ * cdef inline int import_array() except -1:
+ *     try:
+ *         _import_array()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ */
+      __pyx_t_4 = _import_array(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 998, __pyx_L3_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":997
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":999
+ *     try:
+ *         _import_array()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 999, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1000
+ *         _import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__7, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1000, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1000, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":997
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_array()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":996
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_array()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1002
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_umath(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_umath", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1003
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1004
+ * cdef inline int import_umath() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 1004, __pyx_L3_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1003
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1005
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 1005, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1006
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__8, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1006, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1006, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1003
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1002
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1008
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_ufunc(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  __Pyx_RefNannySetupContext("import_ufunc", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1009
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1010
+ * cdef inline int import_ufunc() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 1010, __pyx_L3_error)
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1009
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1011
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 1011, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1012
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__9, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 1012, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 1012, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1009
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1008
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+static std::string __pyx_convert_string_from_py_std__in_string(PyObject *__pyx_v_o) {
+  Py_ssize_t __pyx_v_length;
+  char const *__pyx_v_data;
+  std::string __pyx_r;
+  __Pyx_RefNannyDeclarations
+  char const *__pyx_t_1;
+  __Pyx_RefNannySetupContext("__pyx_convert_string_from_py_std__in_string", 0);
+
+  /* "string.from_py":15
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:
+ *     cdef Py_ssize_t length
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)             # <<<<<<<<<<<<<<
+ *     return string(data, length)
+ * 
+ */
+  __pyx_t_1 = __Pyx_PyObject_AsStringAndSize(__pyx_v_o, (&__pyx_v_length)); if (unlikely(__pyx_t_1 == ((char const *)NULL))) __PYX_ERR(2, 15, __pyx_L1_error)
+  __pyx_v_data = __pyx_t_1;
+
+  /* "string.from_py":16
+ *     cdef Py_ssize_t length
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ *     return string(data, length)             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = std::string(__pyx_v_data, __pyx_v_length);
+  goto __pyx_L0;
+
+  /* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_AddTraceback("string.from_py.__pyx_convert_string_from_py_std__in_string", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_pretend_to_initialize(&__pyx_r);
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyMethodDef __pyx_methods[] = {
+  {0, 0, 0, 0}
+};
+
+#if PY_MAJOR_VERSION >= 3
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/
+static int __pyx_pymod_exec_mesh_core_cython(PyObject* module); /*proto*/
+static PyModuleDef_Slot __pyx_moduledef_slots[] = {
+  {Py_mod_create, (void*)__pyx_pymod_create},
+  {Py_mod_exec, (void*)__pyx_pymod_exec_mesh_core_cython},
+  {0, NULL}
+};
+#endif
+
+static struct PyModuleDef __pyx_moduledef = {
+    PyModuleDef_HEAD_INIT,
+    "mesh_core_cython",
+    0, /* m_doc */
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    0, /* m_size */
+  #else
+    -1, /* m_size */
+  #endif
+    __pyx_methods /* m_methods */,
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    __pyx_moduledef_slots, /* m_slots */
+  #else
+    NULL, /* m_reload */
+  #endif
+    NULL, /* m_traverse */
+    NULL, /* m_clear */
+    NULL /* m_free */
+};
+#endif
+
+static __Pyx_StringTabEntry __pyx_string_tab[] = {
+  {&__pyx_kp_u_Format_string_allocated_too_shor, __pyx_k_Format_string_allocated_too_shor, sizeof(__pyx_k_Format_string_allocated_too_shor), 0, 1, 0, 0},
+  {&__pyx_kp_u_Format_string_allocated_too_shor_2, __pyx_k_Format_string_allocated_too_shor_2, sizeof(__pyx_k_Format_string_allocated_too_shor_2), 0, 1, 0, 0},
+  {&__pyx_n_s_ImportError, __pyx_k_ImportError, sizeof(__pyx_k_ImportError), 0, 0, 1, 1},
+  {&__pyx_kp_u_Non_native_byte_order_not_suppor, __pyx_k_Non_native_byte_order_not_suppor, sizeof(__pyx_k_Non_native_byte_order_not_suppor), 0, 1, 0, 0},
+  {&__pyx_n_s_RuntimeError, __pyx_k_RuntimeError, sizeof(__pyx_k_RuntimeError), 0, 0, 1, 1},
+  {&__pyx_n_s_ValueError, __pyx_k_ValueError, sizeof(__pyx_k_ValueError), 0, 0, 1, 1},
+  {&__pyx_n_s_barycentric_weight, __pyx_k_barycentric_weight, sizeof(__pyx_k_barycentric_weight), 0, 0, 1, 1},
+  {&__pyx_n_s_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 0, 1, 1},
+  {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1},
+  {&__pyx_n_s_colors, __pyx_k_colors, sizeof(__pyx_k_colors), 0, 0, 1, 1},
+  {&__pyx_n_s_depth_buffer, __pyx_k_depth_buffer, sizeof(__pyx_k_depth_buffer), 0, 0, 1, 1},
+  {&__pyx_n_s_filename, __pyx_k_filename, sizeof(__pyx_k_filename), 0, 0, 1, 1},
+  {&__pyx_n_s_get_normal_core, __pyx_k_get_normal_core, sizeof(__pyx_k_get_normal_core), 0, 0, 1, 1},
+  {&__pyx_n_s_h, __pyx_k_h, sizeof(__pyx_k_h), 0, 0, 1, 1},
+  {&__pyx_n_s_image, __pyx_k_image, sizeof(__pyx_k_image), 0, 0, 1, 1},
+  {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1},
+  {&__pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_k_insightface_thirdparty_face3d_me, sizeof(__pyx_k_insightface_thirdparty_face3d_me), 0, 0, 1, 0},
+  {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1},
+  {&__pyx_n_s_mapping_type, __pyx_k_mapping_type, sizeof(__pyx_k_mapping_type), 0, 0, 1, 1},
+  {&__pyx_n_s_mesh_core_cython, __pyx_k_mesh_core_cython, sizeof(__pyx_k_mesh_core_cython), 0, 0, 1, 1},
+  {&__pyx_n_s_mtl_name, __pyx_k_mtl_name, sizeof(__pyx_k_mtl_name), 0, 0, 1, 1},
+  {&__pyx_kp_u_ndarray_is_not_C_contiguous, __pyx_k_ndarray_is_not_C_contiguous, sizeof(__pyx_k_ndarray_is_not_C_contiguous), 0, 1, 0, 0},
+  {&__pyx_kp_u_ndarray_is_not_Fortran_contiguou, __pyx_k_ndarray_is_not_Fortran_contiguou, sizeof(__pyx_k_ndarray_is_not_Fortran_contiguou), 0, 1, 0, 0},
+  {&__pyx_n_s_normal, __pyx_k_normal, sizeof(__pyx_k_normal), 0, 0, 1, 1},
+  {&__pyx_n_s_np, __pyx_k_np, sizeof(__pyx_k_np), 0, 0, 1, 1},
+  {&__pyx_n_s_ntexver, __pyx_k_ntexver, sizeof(__pyx_k_ntexver), 0, 0, 1, 1},
+  {&__pyx_n_s_ntri, __pyx_k_ntri, sizeof(__pyx_k_ntri), 0, 0, 1, 1},
+  {&__pyx_n_s_numpy, __pyx_k_numpy, sizeof(__pyx_k_numpy), 0, 0, 1, 1},
+  {&__pyx_kp_s_numpy_core_multiarray_failed_to, __pyx_k_numpy_core_multiarray_failed_to, sizeof(__pyx_k_numpy_core_multiarray_failed_to), 0, 0, 1, 0},
+  {&__pyx_kp_s_numpy_core_umath_failed_to_impor, __pyx_k_numpy_core_umath_failed_to_impor, sizeof(__pyx_k_numpy_core_umath_failed_to_impor), 0, 0, 1, 0},
+  {&__pyx_n_s_nver, __pyx_k_nver, sizeof(__pyx_k_nver), 0, 0, 1, 1},
+  {&__pyx_n_s_range, __pyx_k_range, sizeof(__pyx_k_range), 0, 0, 1, 1},
+  {&__pyx_n_s_rasterize_triangles_core, __pyx_k_rasterize_triangles_core, sizeof(__pyx_k_rasterize_triangles_core), 0, 0, 1, 1},
+  {&__pyx_n_s_render_colors_core, __pyx_k_render_colors_core, sizeof(__pyx_k_render_colors_core), 0, 0, 1, 1},
+  {&__pyx_n_s_render_texture_core, __pyx_k_render_texture_core, sizeof(__pyx_k_render_texture_core), 0, 0, 1, 1},
+  {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_c, __pyx_k_tex_c, sizeof(__pyx_k_tex_c), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_coords, __pyx_k_tex_coords, sizeof(__pyx_k_tex_coords), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_h, __pyx_k_tex_h, sizeof(__pyx_k_tex_h), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_nver, __pyx_k_tex_nver, sizeof(__pyx_k_tex_nver), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_triangles, __pyx_k_tex_triangles, sizeof(__pyx_k_tex_triangles), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_w, __pyx_k_tex_w, sizeof(__pyx_k_tex_w), 0, 0, 1, 1},
+  {&__pyx_n_s_texture, __pyx_k_texture, sizeof(__pyx_k_texture), 0, 0, 1, 1},
+  {&__pyx_n_s_tri_normal, __pyx_k_tri_normal, sizeof(__pyx_k_tri_normal), 0, 0, 1, 1},
+  {&__pyx_n_s_triangle_buffer, __pyx_k_triangle_buffer, sizeof(__pyx_k_triangle_buffer), 0, 0, 1, 1},
+  {&__pyx_n_s_triangles, __pyx_k_triangles, sizeof(__pyx_k_triangles), 0, 0, 1, 1},
+  {&__pyx_kp_u_unknown_dtype_code_in_numpy_pxd, __pyx_k_unknown_dtype_code_in_numpy_pxd, sizeof(__pyx_k_unknown_dtype_code_in_numpy_pxd), 0, 1, 0, 0},
+  {&__pyx_n_s_uv_coords, __pyx_k_uv_coords, sizeof(__pyx_k_uv_coords), 0, 0, 1, 1},
+  {&__pyx_n_s_vertices, __pyx_k_vertices, sizeof(__pyx_k_vertices), 0, 0, 1, 1},
+  {&__pyx_n_s_w, __pyx_k_w, sizeof(__pyx_k_w), 0, 0, 1, 1},
+  {&__pyx_n_s_write_obj_with_colors_texture_co, __pyx_k_write_obj_with_colors_texture_co, sizeof(__pyx_k_write_obj_with_colors_texture_co), 0, 0, 1, 1},
+  {0, 0, 0, 0, 0, 0, 0}
+};
+static int __Pyx_InitCachedBuiltins(void) {
+  __pyx_builtin_ValueError = __Pyx_GetBuiltinName(__pyx_n_s_ValueError); if (!__pyx_builtin_ValueError) __PYX_ERR(1, 229, __pyx_L1_error)
+  __pyx_builtin_range = __Pyx_GetBuiltinName(__pyx_n_s_range); if (!__pyx_builtin_range) __PYX_ERR(1, 242, __pyx_L1_error)
+  __pyx_builtin_RuntimeError = __Pyx_GetBuiltinName(__pyx_n_s_RuntimeError); if (!__pyx_builtin_RuntimeError) __PYX_ERR(1, 810, __pyx_L1_error)
+  __pyx_builtin_ImportError = __Pyx_GetBuiltinName(__pyx_n_s_ImportError); if (!__pyx_builtin_ImportError) __PYX_ERR(1, 1000, __pyx_L1_error)
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static int __Pyx_InitCachedConstants(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":229
+ *             if ((flags & pybuf.PyBUF_C_CONTIGUOUS == pybuf.PyBUF_C_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_C_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not C contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ */
+  __pyx_tuple_ = PyTuple_Pack(1, __pyx_kp_u_ndarray_is_not_C_contiguous); if (unlikely(!__pyx_tuple_)) __PYX_ERR(1, 229, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple_);
+  __Pyx_GIVEREF(__pyx_tuple_);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":233
+ *             if ((flags & pybuf.PyBUF_F_CONTIGUOUS == pybuf.PyBUF_F_CONTIGUOUS)
+ *                 and not PyArray_CHKFLAGS(self, NPY_F_CONTIGUOUS)):
+ *                 raise ValueError(u"ndarray is not Fortran contiguous")             # <<<<<<<<<<<<<<
+ * 
+ *             info.buf = PyArray_DATA(self)
+ */
+  __pyx_tuple__2 = PyTuple_Pack(1, __pyx_kp_u_ndarray_is_not_Fortran_contiguou); if (unlikely(!__pyx_tuple__2)) __PYX_ERR(1, 233, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__2);
+  __Pyx_GIVEREF(__pyx_tuple__2);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":263
+ *                 if ((descr.byteorder == c'>' and little_endian) or
+ *                     (descr.byteorder == c'<' and not little_endian)):
+ *                     raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *                 if   t == NPY_BYTE:        f = "b"
+ *                 elif t == NPY_UBYTE:       f = "B"
+ */
+  __pyx_tuple__3 = PyTuple_Pack(1, __pyx_kp_u_Non_native_byte_order_not_suppor); if (unlikely(!__pyx_tuple__3)) __PYX_ERR(1, 263, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__3);
+  __Pyx_GIVEREF(__pyx_tuple__3);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":810
+ * 
+ *         if (end - f) - <int>(new_offset - offset[0]) < 15:
+ *             raise RuntimeError(u"Format string allocated too short, see comment in numpy.pxd")             # <<<<<<<<<<<<<<
+ * 
+ *         if ((child.byteorder == c'>' and little_endian) or
+ */
+  __pyx_tuple__4 = PyTuple_Pack(1, __pyx_kp_u_Format_string_allocated_too_shor); if (unlikely(!__pyx_tuple__4)) __PYX_ERR(1, 810, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__4);
+  __Pyx_GIVEREF(__pyx_tuple__4);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":814
+ *         if ((child.byteorder == c'>' and little_endian) or
+ *             (child.byteorder == c'<' and not little_endian)):
+ *             raise ValueError(u"Non-native byte order not supported")             # <<<<<<<<<<<<<<
+ *             # One could encode it in the format string and have Cython
+ *             # complain instead, BUT: < and > in format strings also imply
+ */
+  __pyx_tuple__5 = PyTuple_Pack(1, __pyx_kp_u_Non_native_byte_order_not_suppor); if (unlikely(!__pyx_tuple__5)) __PYX_ERR(1, 814, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__5);
+  __Pyx_GIVEREF(__pyx_tuple__5);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":834
+ *             t = child.type_num
+ *             if end - f < 5:
+ *                 raise RuntimeError(u"Format string allocated too short.")             # <<<<<<<<<<<<<<
+ * 
+ *             # Until ticket #99 is fixed, use integers to avoid warnings
+ */
+  __pyx_tuple__6 = PyTuple_Pack(1, __pyx_kp_u_Format_string_allocated_too_shor_2); if (unlikely(!__pyx_tuple__6)) __PYX_ERR(1, 834, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__6);
+  __Pyx_GIVEREF(__pyx_tuple__6);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1000
+ *         _import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+  __pyx_tuple__7 = PyTuple_Pack(1, __pyx_kp_s_numpy_core_multiarray_failed_to); if (unlikely(!__pyx_tuple__7)) __PYX_ERR(1, 1000, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__7);
+  __Pyx_GIVEREF(__pyx_tuple__7);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1006
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+  __pyx_tuple__8 = PyTuple_Pack(1, __pyx_kp_s_numpy_core_umath_failed_to_impor); if (unlikely(!__pyx_tuple__8)) __PYX_ERR(1, 1006, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__8);
+  __Pyx_GIVEREF(__pyx_tuple__8);
+
+  /* "../../../../jack_ssd/home/jack/anaconda3/lib/python3.6/site-packages/Cython/Includes/numpy/__init__.pxd":1012
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ */
+  __pyx_tuple__9 = PyTuple_Pack(1, __pyx_kp_s_numpy_core_umath_failed_to_impor); if (unlikely(!__pyx_tuple__9)) __PYX_ERR(1, 1012, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__9);
+  __Pyx_GIVEREF(__pyx_tuple__9);
+
+  /* "mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__10 = PyTuple_Pack(4, __pyx_n_s_normal, __pyx_n_s_tri_normal, __pyx_n_s_triangles, __pyx_n_s_ntri); if (unlikely(!__pyx_tuple__10)) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__10);
+  __Pyx_GIVEREF(__pyx_tuple__10);
+  __pyx_codeobj__11 = (PyObject*)__Pyx_PyCode_New(4, 0, 4, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__10, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_get_normal_core, 40, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__11)) __PYX_ERR(0, 40, __pyx_L1_error)
+
+  /* "mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__12 = PyTuple_Pack(9, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_depth_buffer, __pyx_n_s_triangle_buffer, __pyx_n_s_barycentric_weight, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w); if (unlikely(!__pyx_tuple__12)) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__12);
+  __Pyx_GIVEREF(__pyx_tuple__12);
+  __pyx_codeobj__13 = (PyObject*)__Pyx_PyCode_New(9, 0, 9, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__12, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_rasterize_triangles_core, 49, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__13)) __PYX_ERR(0, 49, __pyx_L1_error)
+
+  /* "mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__14 = PyTuple_Pack(10, __pyx_n_s_image, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_colors, __pyx_n_s_depth_buffer, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w, __pyx_n_s_c); if (unlikely(!__pyx_tuple__14)) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__14);
+  __Pyx_GIVEREF(__pyx_tuple__14);
+  __pyx_codeobj__15 = (PyObject*)__Pyx_PyCode_New(10, 0, 10, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__14, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_render_colors_core, 64, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__15)) __PYX_ERR(0, 64, __pyx_L1_error)
+
+  /* "mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__16 = PyTuple_Pack(17, __pyx_n_s_image, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_texture, __pyx_n_s_tex_coords, __pyx_n_s_tex_triangles, __pyx_n_s_depth_buffer, __pyx_n_s_nver, __pyx_n_s_tex_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w, __pyx_n_s_c, __pyx_n_s_tex_h, __pyx_n_s_tex_w, __pyx_n_s_tex_c, __pyx_n_s_mapping_type); if (unlikely(!__pyx_tuple__16)) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__16);
+  __Pyx_GIVEREF(__pyx_tuple__16);
+  __pyx_codeobj__17 = (PyObject*)__Pyx_PyCode_New(17, 0, 17, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__16, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_render_texture_core, 79, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__17)) __PYX_ERR(0, 79, __pyx_L1_error)
+
+  /* "mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__18 = PyTuple_Pack(9, __pyx_n_s_filename, __pyx_n_s_mtl_name, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_colors, __pyx_n_s_uv_coords, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_ntexver); if (unlikely(!__pyx_tuple__18)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__18);
+  __Pyx_GIVEREF(__pyx_tuple__18);
+  __pyx_codeobj__19 = (PyObject*)__Pyx_PyCode_New(9, 0, 9, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__18, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_write_obj_with_colors_texture_co, 100, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__19)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static int __Pyx_InitGlobals(void) {
+  if (__Pyx_InitStrings(__pyx_string_tab) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static int __Pyx_modinit_global_init_code(void); /*proto*/
+static int __Pyx_modinit_variable_export_code(void); /*proto*/
+static int __Pyx_modinit_function_export_code(void); /*proto*/
+static int __Pyx_modinit_type_init_code(void); /*proto*/
+static int __Pyx_modinit_type_import_code(void); /*proto*/
+static int __Pyx_modinit_variable_import_code(void); /*proto*/
+static int __Pyx_modinit_function_import_code(void); /*proto*/
+
+static int __Pyx_modinit_global_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0);
+  /*--- Global init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_variable_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0);
+  /*--- Variable export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0);
+  /*--- Function export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0);
+  /*--- Type init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0);
+  /*--- Type import code ---*/
+  __pyx_ptype_7cpython_4type_type = __Pyx_ImportType(__Pyx_BUILTIN_MODULE_NAME, "type", 
+  #if defined(PYPY_VERSION_NUM) && PYPY_VERSION_NUM < 0x050B0000
+  sizeof(PyTypeObject),
+  #else
+  sizeof(PyHeapTypeObject),
+  #endif
+  0); if (unlikely(!__pyx_ptype_7cpython_4type_type)) __PYX_ERR(3, 9, __pyx_L1_error)
+  __pyx_ptype_5numpy_dtype = __Pyx_ImportType("numpy", "dtype", sizeof(PyArray_Descr), 0); if (unlikely(!__pyx_ptype_5numpy_dtype)) __PYX_ERR(1, 164, __pyx_L1_error)
+  __pyx_ptype_5numpy_flatiter = __Pyx_ImportType("numpy", "flatiter", sizeof(PyArrayIterObject), 0); if (unlikely(!__pyx_ptype_5numpy_flatiter)) __PYX_ERR(1, 186, __pyx_L1_error)
+  __pyx_ptype_5numpy_broadcast = __Pyx_ImportType("numpy", "broadcast", sizeof(PyArrayMultiIterObject), 0); if (unlikely(!__pyx_ptype_5numpy_broadcast)) __PYX_ERR(1, 190, __pyx_L1_error)
+  __pyx_ptype_5numpy_ndarray = __Pyx_ImportType("numpy", "ndarray", sizeof(PyArrayObject), 0); if (unlikely(!__pyx_ptype_5numpy_ndarray)) __PYX_ERR(1, 199, __pyx_L1_error)
+  __pyx_ptype_5numpy_ufunc = __Pyx_ImportType("numpy", "ufunc", sizeof(PyUFuncObject), 0); if (unlikely(!__pyx_ptype_5numpy_ufunc)) __PYX_ERR(1, 872, __pyx_L1_error)
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static int __Pyx_modinit_variable_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0);
+  /*--- Variable import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0);
+  /*--- Function import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+
+#if PY_MAJOR_VERSION < 3
+#ifdef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC void
+#else
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#endif
+#else
+#ifdef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC PyObject *
+#else
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#endif
+#endif
+#ifndef CYTHON_SMALL_CODE
+#if defined(__clang__)
+    #define CYTHON_SMALL_CODE
+#elif defined(__GNUC__)
+    #define CYTHON_SMALL_CODE __attribute__((optimize("Os")))
+#else
+    #define CYTHON_SMALL_CODE
+#endif
+#endif
+
+
+#if PY_MAJOR_VERSION < 3
+__Pyx_PyMODINIT_FUNC initmesh_core_cython(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC initmesh_core_cython(void)
+#else
+__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void)
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+{
+  return PyModuleDef_Init(&__pyx_moduledef);
+}
+static int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name) {
+    PyObject *value = PyObject_GetAttrString(spec, from_name);
+    int result = 0;
+    if (likely(value)) {
+        result = PyDict_SetItemString(moddict, to_name, value);
+        Py_DECREF(value);
+    } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) {
+        PyErr_Clear();
+    } else {
+        result = -1;
+    }
+    return result;
+}
+static PyObject* __pyx_pymod_create(PyObject *spec, CYTHON_UNUSED PyModuleDef *def) {
+    PyObject *module = NULL, *moddict, *modname;
+    if (__pyx_m)
+        return __Pyx_NewRef(__pyx_m);
+    modname = PyObject_GetAttrString(spec, "name");
+    if (unlikely(!modname)) goto bad;
+    module = PyModule_NewObject(modname);
+    Py_DECREF(modname);
+    if (unlikely(!module)) goto bad;
+    moddict = PyModule_GetDict(module);
+    if (unlikely(!moddict)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__") < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__") < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__") < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__") < 0)) goto bad;
+    return module;
+bad:
+    Py_XDECREF(module);
+    return NULL;
+}
+
+
+static int __pyx_pymod_exec_mesh_core_cython(PyObject *__pyx_pyinit_module)
+#endif
+#endif
+{
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_t_2;
+  __Pyx_RefNannyDeclarations
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  if (__pyx_m && __pyx_m == __pyx_pyinit_module) return 0;
+  #elif PY_MAJOR_VERSION >= 3
+  if (__pyx_m) return __Pyx_NewRef(__pyx_m);
+  #endif
+  #if CYTHON_REFNANNY
+__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny");
+if (!__Pyx_RefNanny) {
+  PyErr_Clear();
+  __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny");
+  if (!__Pyx_RefNanny)
+      Py_FatalError("failed to import 'refnanny' module");
+}
+#endif
+  __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void)", 0);
+  if (__Pyx_check_binary_version() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pyx_CyFunction_USED
+  if (__pyx_CyFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_FusedFunction_USED
+  if (__pyx_FusedFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Coroutine_USED
+  if (__pyx_Coroutine_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Generator_USED
+  if (__pyx_Generator_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_AsyncGen_USED
+  if (__pyx_AsyncGen_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_StopAsyncIteration_USED
+  if (__pyx_StopAsyncIteration_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  /*--- Library function declarations ---*/
+  /*--- Threads initialization code ---*/
+  #if defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS
+  #ifdef WITH_THREAD /* Python build with threading support? */
+  PyEval_InitThreads();
+  #endif
+  #endif
+  /*--- Module creation code ---*/
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  __pyx_m = __pyx_pyinit_module;
+  Py_INCREF(__pyx_m);
+  #else
+  #if PY_MAJOR_VERSION < 3
+  __pyx_m = Py_InitModule4("mesh_core_cython", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m);
+  #else
+  __pyx_m = PyModule_Create(&__pyx_moduledef);
+  #endif
+  if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_d);
+  __pyx_b = PyImport_AddModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_cython_runtime = PyImport_AddModule((char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #if CYTHON_COMPILING_IN_PYPY
+  Py_INCREF(__pyx_b);
+  #endif
+  if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  /*--- Initialize various global constants etc. ---*/
+  if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT)
+  if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  if (__pyx_module_is_main_mesh_core_cython) {
+    if (PyObject_SetAttrString(__pyx_m, "__name__", __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  }
+  #if PY_MAJOR_VERSION >= 3
+  {
+    PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error)
+    if (!PyDict_GetItemString(modules, "mesh_core_cython")) {
+      if (unlikely(PyDict_SetItemString(modules, "mesh_core_cython", __pyx_m) < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+    }
+  }
+  #endif
+  /*--- Builtin init code ---*/
+  if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Constants init code ---*/
+  if (__Pyx_InitCachedConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Global type/function init code ---*/
+  (void)__Pyx_modinit_global_init_code();
+  (void)__Pyx_modinit_variable_export_code();
+  (void)__Pyx_modinit_function_export_code();
+  (void)__Pyx_modinit_type_init_code();
+  if (unlikely(__Pyx_modinit_type_import_code() != 0)) goto __pyx_L1_error;
+  (void)__Pyx_modinit_variable_import_code();
+  (void)__Pyx_modinit_function_import_code();
+  /*--- Execution code ---*/
+  #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED)
+  if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+
+  /* "mesh_core_cython.pyx":1
+ * import numpy as np             # <<<<<<<<<<<<<<
+ * cimport numpy as np
+ * from libcpp.string cimport string
+ */
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_numpy, 0, -1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_np, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":6
+ * 
+ * # use the Numpy-C-API from Cython
+ * np.import_array()             # <<<<<<<<<<<<<<
+ * 
+ * # cdefine the signature of our c function
+ */
+  __pyx_t_2 = __pyx_f_5numpy_import_array(); if (unlikely(__pyx_t_2 == ((int)-1))) __PYX_ERR(0, 6, __pyx_L1_error)
+
+  /* "mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_16mesh_core_cython_1get_normal_core, NULL, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_get_normal_core, __pyx_t_1) < 0) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_16mesh_core_cython_3rasterize_triangles_core, NULL, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_rasterize_triangles_core, __pyx_t_1) < 0) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_16mesh_core_cython_5render_colors_core, NULL, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_render_colors_core, __pyx_t_1) < 0) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_16mesh_core_cython_7render_texture_core, NULL, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_render_texture_core, __pyx_t_1) < 0) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_16mesh_core_cython_9write_obj_with_colors_texture_core, NULL, __pyx_n_s_mesh_core_cython); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_write_obj_with_colors_texture_co, __pyx_t_1) < 0) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "mesh_core_cython.pyx":1
+ * import numpy as np             # <<<<<<<<<<<<<<
+ * cimport numpy as np
+ * from libcpp.string cimport string
+ */
+  __pyx_t_1 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+  /*--- Wrapped vars code ---*/
+
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  if (__pyx_m) {
+    if (__pyx_d) {
+      __Pyx_AddTraceback("init mesh_core_cython", 0, __pyx_lineno, __pyx_filename);
+    }
+    Py_DECREF(__pyx_m); __pyx_m = 0;
+  } else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_ImportError, "init mesh_core_cython");
+  }
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  return (__pyx_m != NULL) ? 0 : -1;
+  #elif PY_MAJOR_VERSION >= 3
+  return __pyx_m;
+  #else
+  return;
+  #endif
+}
+
+/* --- Runtime support code --- */
+/* Refnanny */
+#if CYTHON_REFNANNY
+static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) {
+    PyObject *m = NULL, *p = NULL;
+    void *r = NULL;
+    m = PyImport_ImportModule((char *)modname);
+    if (!m) goto end;
+    p = PyObject_GetAttrString(m, (char *)"RefNannyAPI");
+    if (!p) goto end;
+    r = PyLong_AsVoidPtr(p);
+end:
+    Py_XDECREF(p);
+    Py_XDECREF(m);
+    return (__Pyx_RefNannyAPIStruct *)r;
+}
+#endif
+
+/* RaiseArgTupleInvalid */
+static void __Pyx_RaiseArgtupleInvalid(
+    const char* func_name,
+    int exact,
+    Py_ssize_t num_min,
+    Py_ssize_t num_max,
+    Py_ssize_t num_found)
+{
+    Py_ssize_t num_expected;
+    const char *more_or_less;
+    if (num_found < num_min) {
+        num_expected = num_min;
+        more_or_less = "at least";
+    } else {
+        num_expected = num_max;
+        more_or_less = "at most";
+    }
+    if (exact) {
+        more_or_less = "exactly";
+    }
+    PyErr_Format(PyExc_TypeError,
+                 "%.200s() takes %.8s %" CYTHON_FORMAT_SSIZE_T "d positional argument%.1s (%" CYTHON_FORMAT_SSIZE_T "d given)",
+                 func_name, more_or_less, num_expected,
+                 (num_expected == 1) ? "" : "s", num_found);
+}
+
+/* RaiseDoubleKeywords */
+static void __Pyx_RaiseDoubleKeywordsError(
+    const char* func_name,
+    PyObject* kw_name)
+{
+    PyErr_Format(PyExc_TypeError,
+        #if PY_MAJOR_VERSION >= 3
+        "%s() got multiple values for keyword argument '%U'", func_name, kw_name);
+        #else
+        "%s() got multiple values for keyword argument '%s'", func_name,
+        PyString_AsString(kw_name));
+        #endif
+}
+
+/* ParseKeywords */
+static int __Pyx_ParseOptionalKeywords(
+    PyObject *kwds,
+    PyObject **argnames[],
+    PyObject *kwds2,
+    PyObject *values[],
+    Py_ssize_t num_pos_args,
+    const char* function_name)
+{
+    PyObject *key = 0, *value = 0;
+    Py_ssize_t pos = 0;
+    PyObject*** name;
+    PyObject*** first_kw_arg = argnames + num_pos_args;
+    while (PyDict_Next(kwds, &pos, &key, &value)) {
+        name = first_kw_arg;
+        while (*name && (**name != key)) name++;
+        if (*name) {
+            values[name-argnames] = value;
+            continue;
+        }
+        name = first_kw_arg;
+        #if PY_MAJOR_VERSION < 3
+        if (likely(PyString_CheckExact(key)) || likely(PyString_Check(key))) {
+            while (*name) {
+                if ((CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**name) == PyString_GET_SIZE(key))
+                        && _PyString_Eq(**name, key)) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    if ((**argname == key) || (
+                            (CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**argname) == PyString_GET_SIZE(key))
+                             && _PyString_Eq(**argname, key))) {
+                        goto arg_passed_twice;
+                    }
+                    argname++;
+                }
+            }
+        } else
+        #endif
+        if (likely(PyUnicode_Check(key))) {
+            while (*name) {
+                int cmp = (**name == key) ? 0 :
+                #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                    (PyUnicode_GET_SIZE(**name) != PyUnicode_GET_SIZE(key)) ? 1 :
+                #endif
+                    PyUnicode_Compare(**name, key);
+                if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                if (cmp == 0) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    int cmp = (**argname == key) ? 0 :
+                    #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                        (PyUnicode_GET_SIZE(**argname) != PyUnicode_GET_SIZE(key)) ? 1 :
+                    #endif
+                        PyUnicode_Compare(**argname, key);
+                    if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                    if (cmp == 0) goto arg_passed_twice;
+                    argname++;
+                }
+            }
+        } else
+            goto invalid_keyword_type;
+        if (kwds2) {
+            if (unlikely(PyDict_SetItem(kwds2, key, value))) goto bad;
+        } else {
+            goto invalid_keyword;
+        }
+    }
+    return 0;
+arg_passed_twice:
+    __Pyx_RaiseDoubleKeywordsError(function_name, key);
+    goto bad;
+invalid_keyword_type:
+    PyErr_Format(PyExc_TypeError,
+        "%.200s() keywords must be strings", function_name);
+    goto bad;
+invalid_keyword:
+    PyErr_Format(PyExc_TypeError,
+    #if PY_MAJOR_VERSION < 3
+        "%.200s() got an unexpected keyword argument '%.200s'",
+        function_name, PyString_AsString(key));
+    #else
+        "%s() got an unexpected keyword argument '%U'",
+        function_name, key);
+    #endif
+bad:
+    return -1;
+}
+
+/* ArgTypeTest */
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact)
+{
+    if (unlikely(!type)) {
+        PyErr_SetString(PyExc_SystemError, "Missing type object");
+        return 0;
+    }
+    else if (exact) {
+        #if PY_MAJOR_VERSION == 2
+        if ((type == &PyBaseString_Type) && likely(__Pyx_PyBaseString_CheckExact(obj))) return 1;
+        #endif
+    }
+    else {
+        if (likely(__Pyx_TypeCheck(obj, type))) return 1;
+    }
+    PyErr_Format(PyExc_TypeError,
+        "Argument '%.200s' has incorrect type (expected %.200s, got %.200s)",
+        name, type->tp_name, Py_TYPE(obj)->tp_name);
+    return 0;
+}
+
+/* IsLittleEndian */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void)
+{
+  union {
+    uint32_t u32;
+    uint8_t u8[4];
+  } S;
+  S.u32 = 0x01020304;
+  return S.u8[0] == 4;
+}
+
+/* BufferFormatCheck */
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type) {
+  stack[0].field = &ctx->root;
+  stack[0].parent_offset = 0;
+  ctx->root.type = type;
+  ctx->root.name = "buffer dtype";
+  ctx->root.offset = 0;
+  ctx->head = stack;
+  ctx->head->field = &ctx->root;
+  ctx->fmt_offset = 0;
+  ctx->head->parent_offset = 0;
+  ctx->new_packmode = '@';
+  ctx->enc_packmode = '@';
+  ctx->new_count = 1;
+  ctx->enc_count = 0;
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  ctx->is_valid_array = 0;
+  ctx->struct_alignment = 0;
+  while (type->typegroup == 'S') {
+    ++ctx->head;
+    ctx->head->field = type->fields;
+    ctx->head->parent_offset = 0;
+    type = type->fields->type;
+  }
+}
+static int __Pyx_BufFmt_ParseNumber(const char** ts) {
+    int count;
+    const char* t = *ts;
+    if (*t < '0' || *t > '9') {
+      return -1;
+    } else {
+        count = *t++ - '0';
+        while (*t >= '0' && *t < '9') {
+            count *= 10;
+            count += *t++ - '0';
+        }
+    }
+    *ts = t;
+    return count;
+}
+static int __Pyx_BufFmt_ExpectNumber(const char **ts) {
+    int number = __Pyx_BufFmt_ParseNumber(ts);
+    if (number == -1)
+        PyErr_Format(PyExc_ValueError,\
+                     "Does not understand character buffer dtype format string ('%c')", **ts);
+    return number;
+}
+static void __Pyx_BufFmt_RaiseUnexpectedChar(char ch) {
+  PyErr_Format(PyExc_ValueError,
+               "Unexpected format string character: '%c'", ch);
+}
+static const char* __Pyx_BufFmt_DescribeTypeChar(char ch, int is_complex) {
+  switch (ch) {
+    case 'c': return "'char'";
+    case 'b': return "'signed char'";
+    case 'B': return "'unsigned char'";
+    case 'h': return "'short'";
+    case 'H': return "'unsigned short'";
+    case 'i': return "'int'";
+    case 'I': return "'unsigned int'";
+    case 'l': return "'long'";
+    case 'L': return "'unsigned long'";
+    case 'q': return "'long long'";
+    case 'Q': return "'unsigned long long'";
+    case 'f': return (is_complex ? "'complex float'" : "'float'");
+    case 'd': return (is_complex ? "'complex double'" : "'double'");
+    case 'g': return (is_complex ? "'complex long double'" : "'long double'");
+    case 'T': return "a struct";
+    case 'O': return "Python object";
+    case 'P': return "a pointer";
+    case 's': case 'p': return "a string";
+    case 0: return "end";
+    default: return "unparseable format string";
+  }
+}
+static size_t __Pyx_BufFmt_TypeCharToStandardSize(char ch, int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return 2;
+    case 'i': case 'I': case 'l': case 'L': return 4;
+    case 'q': case 'Q': return 8;
+    case 'f': return (is_complex ? 8 : 4);
+    case 'd': return (is_complex ? 16 : 8);
+    case 'g': {
+      PyErr_SetString(PyExc_ValueError, "Python does not define a standard format string size for long double ('g')..");
+      return 0;
+    }
+    case 'O': case 'P': return sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static size_t __Pyx_BufFmt_TypeCharToNativeSize(char ch, int is_complex) {
+  switch (ch) {
+    case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(short);
+    case 'i': case 'I': return sizeof(int);
+    case 'l': case 'L': return sizeof(long);
+    #ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(PY_LONG_LONG);
+    #endif
+    case 'f': return sizeof(float) * (is_complex ? 2 : 1);
+    case 'd': return sizeof(double) * (is_complex ? 2 : 1);
+    case 'g': return sizeof(long double) * (is_complex ? 2 : 1);
+    case 'O': case 'P': return sizeof(void*);
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+typedef struct { char c; short x; } __Pyx_st_short;
+typedef struct { char c; int x; } __Pyx_st_int;
+typedef struct { char c; long x; } __Pyx_st_long;
+typedef struct { char c; float x; } __Pyx_st_float;
+typedef struct { char c; double x; } __Pyx_st_double;
+typedef struct { char c; long double x; } __Pyx_st_longdouble;
+typedef struct { char c; void *x; } __Pyx_st_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { char c; PY_LONG_LONG x; } __Pyx_st_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToAlignment(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_st_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_st_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_st_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_st_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_st_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_st_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_st_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_st_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+/* These are for computing the padding at the end of the struct to align
+   on the first member of the struct. This will probably the same as above,
+   but we don't have any guarantees.
+ */
+typedef struct { short x; char c; } __Pyx_pad_short;
+typedef struct { int x; char c; } __Pyx_pad_int;
+typedef struct { long x; char c; } __Pyx_pad_long;
+typedef struct { float x; char c; } __Pyx_pad_float;
+typedef struct { double x; char c; } __Pyx_pad_double;
+typedef struct { long double x; char c; } __Pyx_pad_longdouble;
+typedef struct { void *x; char c; } __Pyx_pad_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { PY_LONG_LONG x; char c; } __Pyx_pad_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToPadding(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_pad_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_pad_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_pad_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_pad_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_pad_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_pad_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_pad_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_pad_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static char __Pyx_BufFmt_TypeCharToGroup(char ch, int is_complex) {
+  switch (ch) {
+    case 'c':
+        return 'H';
+    case 'b': case 'h': case 'i':
+    case 'l': case 'q': case 's': case 'p':
+        return 'I';
+    case 'B': case 'H': case 'I': case 'L': case 'Q':
+        return 'U';
+    case 'f': case 'd': case 'g':
+        return (is_complex ? 'C' : 'R');
+    case 'O':
+        return 'O';
+    case 'P':
+        return 'P';
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+static void __Pyx_BufFmt_RaiseExpected(__Pyx_BufFmt_Context* ctx) {
+  if (ctx->head == NULL || ctx->head->field == &ctx->root) {
+    const char* expected;
+    const char* quote;
+    if (ctx->head == NULL) {
+      expected = "end";
+      quote = "";
+    } else {
+      expected = ctx->head->field->type->name;
+      quote = "'";
+    }
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected %s%s%s but got %s",
+                 quote, expected, quote,
+                 __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex));
+  } else {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_StructField* parent = (ctx->head - 1)->field;
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected '%s' but got %s in '%s.%s'",
+                 field->type->name, __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex),
+                 parent->type->name, field->name);
+  }
+}
+static int __Pyx_BufFmt_ProcessTypeChunk(__Pyx_BufFmt_Context* ctx) {
+  char group;
+  size_t size, offset, arraysize = 1;
+  if (ctx->enc_type == 0) return 0;
+  if (ctx->head->field->type->arraysize[0]) {
+    int i, ndim = 0;
+    if (ctx->enc_type == 's' || ctx->enc_type == 'p') {
+        ctx->is_valid_array = ctx->head->field->type->ndim == 1;
+        ndim = 1;
+        if (ctx->enc_count != ctx->head->field->type->arraysize[0]) {
+            PyErr_Format(PyExc_ValueError,
+                         "Expected a dimension of size %zu, got %zu",
+                         ctx->head->field->type->arraysize[0], ctx->enc_count);
+            return -1;
+        }
+    }
+    if (!ctx->is_valid_array) {
+      PyErr_Format(PyExc_ValueError, "Expected %d dimensions, got %d",
+                   ctx->head->field->type->ndim, ndim);
+      return -1;
+    }
+    for (i = 0; i < ctx->head->field->type->ndim; i++) {
+      arraysize *= ctx->head->field->type->arraysize[i];
+    }
+    ctx->is_valid_array = 0;
+    ctx->enc_count = 1;
+  }
+  group = __Pyx_BufFmt_TypeCharToGroup(ctx->enc_type, ctx->is_complex);
+  do {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_TypeInfo* type = field->type;
+    if (ctx->enc_packmode == '@' || ctx->enc_packmode == '^') {
+      size = __Pyx_BufFmt_TypeCharToNativeSize(ctx->enc_type, ctx->is_complex);
+    } else {
+      size = __Pyx_BufFmt_TypeCharToStandardSize(ctx->enc_type, ctx->is_complex);
+    }
+    if (ctx->enc_packmode == '@') {
+      size_t align_at = __Pyx_BufFmt_TypeCharToAlignment(ctx->enc_type, ctx->is_complex);
+      size_t align_mod_offset;
+      if (align_at == 0) return -1;
+      align_mod_offset = ctx->fmt_offset % align_at;
+      if (align_mod_offset > 0) ctx->fmt_offset += align_at - align_mod_offset;
+      if (ctx->struct_alignment == 0)
+          ctx->struct_alignment = __Pyx_BufFmt_TypeCharToPadding(ctx->enc_type,
+                                                                 ctx->is_complex);
+    }
+    if (type->size != size || type->typegroup != group) {
+      if (type->typegroup == 'C' && type->fields != NULL) {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        ++ctx->head;
+        ctx->head->field = type->fields;
+        ctx->head->parent_offset = parent_offset;
+        continue;
+      }
+      if ((type->typegroup == 'H' || group == 'H') && type->size == size) {
+      } else {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+      }
+    }
+    offset = ctx->head->parent_offset + field->offset;
+    if (ctx->fmt_offset != offset) {
+      PyErr_Format(PyExc_ValueError,
+                   "Buffer dtype mismatch; next field is at offset %" CYTHON_FORMAT_SSIZE_T "d but %" CYTHON_FORMAT_SSIZE_T "d expected",
+                   (Py_ssize_t)ctx->fmt_offset, (Py_ssize_t)offset);
+      return -1;
+    }
+    ctx->fmt_offset += size;
+    if (arraysize)
+      ctx->fmt_offset += (arraysize - 1) * size;
+    --ctx->enc_count;
+    while (1) {
+      if (field == &ctx->root) {
+        ctx->head = NULL;
+        if (ctx->enc_count != 0) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+        }
+        break;
+      }
+      ctx->head->field = ++field;
+      if (field->type == NULL) {
+        --ctx->head;
+        field = ctx->head->field;
+        continue;
+      } else if (field->type->typegroup == 'S') {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        if (field->type->fields->type == NULL) continue;
+        field = field->type->fields;
+        ++ctx->head;
+        ctx->head->field = field;
+        ctx->head->parent_offset = parent_offset;
+        break;
+      } else {
+        break;
+      }
+    }
+  } while (ctx->enc_count);
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  return 0;
+}
+static PyObject *
+__pyx_buffmt_parse_array(__Pyx_BufFmt_Context* ctx, const char** tsp)
+{
+    const char *ts = *tsp;
+    int i = 0, number;
+    int ndim = ctx->head->field->type->ndim;
+;
+    ++ts;
+    if (ctx->new_count != 1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Cannot handle repeated arrays in format string");
+        return NULL;
+    }
+    if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+    while (*ts && *ts != ')') {
+        switch (*ts) {
+            case ' ': case '\f': case '\r': case '\n': case '\t': case '\v':  continue;
+            default:  break;
+        }
+        number = __Pyx_BufFmt_ExpectNumber(&ts);
+        if (number == -1) return NULL;
+        if (i < ndim && (size_t) number != ctx->head->field->type->arraysize[i])
+            return PyErr_Format(PyExc_ValueError,
+                        "Expected a dimension of size %zu, got %d",
+                        ctx->head->field->type->arraysize[i], number);
+        if (*ts != ',' && *ts != ')')
+            return PyErr_Format(PyExc_ValueError,
+                                "Expected a comma in format string, got '%c'", *ts);
+        if (*ts == ',') ts++;
+        i++;
+    }
+    if (i != ndim)
+        return PyErr_Format(PyExc_ValueError, "Expected %d dimension(s), got %d",
+                            ctx->head->field->type->ndim, i);
+    if (!*ts) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Unexpected end of format string, expected ')'");
+        return NULL;
+    }
+    ctx->is_valid_array = 1;
+    ctx->new_count = 1;
+    *tsp = ++ts;
+    return Py_None;
+}
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts) {
+  int got_Z = 0;
+  while (1) {
+    switch(*ts) {
+      case 0:
+        if (ctx->enc_type != 0 && ctx->head == NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        if (ctx->head != NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        return ts;
+      case ' ':
+      case '\r':
+      case '\n':
+        ++ts;
+        break;
+      case '<':
+        if (!__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Little-endian buffer not supported on big-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '>':
+      case '!':
+        if (__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Big-endian buffer not supported on little-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '=':
+      case '@':
+      case '^':
+        ctx->new_packmode = *ts++;
+        break;
+      case 'T':
+        {
+          const char* ts_after_sub;
+          size_t i, struct_count = ctx->new_count;
+          size_t struct_alignment = ctx->struct_alignment;
+          ctx->new_count = 1;
+          ++ts;
+          if (*ts != '{') {
+            PyErr_SetString(PyExc_ValueError, "Buffer acquisition: Expected '{' after 'T'");
+            return NULL;
+          }
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          ctx->enc_count = 0;
+          ctx->struct_alignment = 0;
+          ++ts;
+          ts_after_sub = ts;
+          for (i = 0; i != struct_count; ++i) {
+            ts_after_sub = __Pyx_BufFmt_CheckString(ctx, ts);
+            if (!ts_after_sub) return NULL;
+          }
+          ts = ts_after_sub;
+          if (struct_alignment) ctx->struct_alignment = struct_alignment;
+        }
+        break;
+      case '}':
+        {
+          size_t alignment = ctx->struct_alignment;
+          ++ts;
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          if (alignment && ctx->fmt_offset % alignment) {
+            ctx->fmt_offset += alignment - (ctx->fmt_offset % alignment);
+          }
+        }
+        return ts;
+      case 'x':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->fmt_offset += ctx->new_count;
+        ctx->new_count = 1;
+        ctx->enc_count = 0;
+        ctx->enc_type = 0;
+        ctx->enc_packmode = ctx->new_packmode;
+        ++ts;
+        break;
+      case 'Z':
+        got_Z = 1;
+        ++ts;
+        if (*ts != 'f' && *ts != 'd' && *ts != 'g') {
+          __Pyx_BufFmt_RaiseUnexpectedChar('Z');
+          return NULL;
+        }
+        CYTHON_FALLTHROUGH;
+      case 'c': case 'b': case 'B': case 'h': case 'H': case 'i': case 'I':
+      case 'l': case 'L': case 'q': case 'Q':
+      case 'f': case 'd': case 'g':
+      case 'O': case 'p':
+        if (ctx->enc_type == *ts && got_Z == ctx->is_complex &&
+            ctx->enc_packmode == ctx->new_packmode) {
+          ctx->enc_count += ctx->new_count;
+          ctx->new_count = 1;
+          got_Z = 0;
+          ++ts;
+          break;
+        }
+        CYTHON_FALLTHROUGH;
+      case 's':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->enc_count = ctx->new_count;
+        ctx->enc_packmode = ctx->new_packmode;
+        ctx->enc_type = *ts;
+        ctx->is_complex = got_Z;
+        ++ts;
+        ctx->new_count = 1;
+        got_Z = 0;
+        break;
+      case ':':
+        ++ts;
+        while(*ts != ':') ++ts;
+        ++ts;
+        break;
+      case '(':
+        if (!__pyx_buffmt_parse_array(ctx, &ts)) return NULL;
+        break;
+      default:
+        {
+          int number = __Pyx_BufFmt_ExpectNumber(&ts);
+          if (number == -1) return NULL;
+          ctx->new_count = (size_t)number;
+        }
+    }
+  }
+}
+
+/* BufferGetAndValidate */
+  static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info) {
+  if (unlikely(info->buf == NULL)) return;
+  if (info->suboffsets == __Pyx_minusones) info->suboffsets = NULL;
+  __Pyx_ReleaseBuffer(info);
+}
+static void __Pyx_ZeroBuffer(Py_buffer* buf) {
+  buf->buf = NULL;
+  buf->obj = NULL;
+  buf->strides = __Pyx_zeros;
+  buf->shape = __Pyx_zeros;
+  buf->suboffsets = __Pyx_minusones;
+}
+static int __Pyx__GetBufferAndValidate(
+        Py_buffer* buf, PyObject* obj,  __Pyx_TypeInfo* dtype, int flags,
+        int nd, int cast, __Pyx_BufFmt_StackElem* stack)
+{
+  buf->buf = NULL;
+  if (unlikely(__Pyx_GetBuffer(obj, buf, flags) == -1)) {
+    __Pyx_ZeroBuffer(buf);
+    return -1;
+  }
+  if (unlikely(buf->ndim != nd)) {
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer has wrong number of dimensions (expected %d, got %d)",
+                 nd, buf->ndim);
+    goto fail;
+  }
+  if (!cast) {
+    __Pyx_BufFmt_Context ctx;
+    __Pyx_BufFmt_Init(&ctx, stack, dtype);
+    if (!__Pyx_BufFmt_CheckString(&ctx, buf->format)) goto fail;
+  }
+  if (unlikely((unsigned)buf->itemsize != dtype->size)) {
+    PyErr_Format(PyExc_ValueError,
+      "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "d byte%s) does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "d byte%s)",
+      buf->itemsize, (buf->itemsize > 1) ? "s" : "",
+      dtype->name, (Py_ssize_t)dtype->size, (dtype->size > 1) ? "s" : "");
+    goto fail;
+  }
+  if (buf->suboffsets == NULL) buf->suboffsets = __Pyx_minusones;
+  return 0;
+fail:;
+  __Pyx_SafeReleaseBuffer(buf);
+  return -1;
+}
+
+/* PyErrFetchRestore */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    tmp_type = tstate->curexc_type;
+    tmp_value = tstate->curexc_value;
+    tmp_tb = tstate->curexc_traceback;
+    tstate->curexc_type = type;
+    tstate->curexc_value = value;
+    tstate->curexc_traceback = tb;
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    *type = tstate->curexc_type;
+    *value = tstate->curexc_value;
+    *tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+}
+#endif
+
+/* PyObjectGetAttrStr */
+  #if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) {
+    PyTypeObject* tp = Py_TYPE(obj);
+    if (likely(tp->tp_getattro))
+        return tp->tp_getattro(obj, attr_name);
+#if PY_MAJOR_VERSION < 3
+    if (likely(tp->tp_getattr))
+        return tp->tp_getattr(obj, PyString_AS_STRING(attr_name));
+#endif
+    return PyObject_GetAttr(obj, attr_name);
+}
+#endif
+
+/* GetBuiltinName */
+  static PyObject *__Pyx_GetBuiltinName(PyObject *name) {
+    PyObject* result = __Pyx_PyObject_GetAttrStr(__pyx_b, name);
+    if (unlikely(!result)) {
+        PyErr_Format(PyExc_NameError,
+#if PY_MAJOR_VERSION >= 3
+            "name '%U' is not defined", name);
+#else
+            "name '%.200s' is not defined", PyString_AS_STRING(name));
+#endif
+    }
+    return result;
+}
+
+/* PyObjectCall */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw) {
+    PyObject *result;
+    ternaryfunc call = func->ob_type->tp_call;
+    if (unlikely(!call))
+        return PyObject_Call(func, arg, kw);
+    if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object")))
+        return NULL;
+    result = (*call)(func, arg, kw);
+    Py_LeaveRecursiveCall();
+    if (unlikely(!result) && unlikely(!PyErr_Occurred())) {
+        PyErr_SetString(
+            PyExc_SystemError,
+            "NULL result without error in PyObject_Call");
+    }
+    return result;
+}
+#endif
+
+/* RaiseException */
+  #if PY_MAJOR_VERSION < 3
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb,
+                        CYTHON_UNUSED PyObject *cause) {
+    __Pyx_PyThreadState_declare
+    Py_XINCREF(type);
+    if (!value || value == Py_None)
+        value = NULL;
+    else
+        Py_INCREF(value);
+    if (!tb || tb == Py_None)
+        tb = NULL;
+    else {
+        Py_INCREF(tb);
+        if (!PyTraceBack_Check(tb)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: arg 3 must be a traceback or None");
+            goto raise_error;
+        }
+    }
+    if (PyType_Check(type)) {
+#if CYTHON_COMPILING_IN_PYPY
+        if (!value) {
+            Py_INCREF(Py_None);
+            value = Py_None;
+        }
+#endif
+        PyErr_NormalizeException(&type, &value, &tb);
+    } else {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto raise_error;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(type);
+        Py_INCREF(type);
+        if (!PyType_IsSubtype((PyTypeObject *)type, (PyTypeObject *)PyExc_BaseException)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: exception class must be a subclass of BaseException");
+            goto raise_error;
+        }
+    }
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrRestore(type, value, tb);
+    return;
+raise_error:
+    Py_XDECREF(value);
+    Py_XDECREF(type);
+    Py_XDECREF(tb);
+    return;
+}
+#else
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) {
+    PyObject* owned_instance = NULL;
+    if (tb == Py_None) {
+        tb = 0;
+    } else if (tb && !PyTraceBack_Check(tb)) {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: arg 3 must be a traceback or None");
+        goto bad;
+    }
+    if (value == Py_None)
+        value = 0;
+    if (PyExceptionInstance_Check(type)) {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto bad;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(value);
+    } else if (PyExceptionClass_Check(type)) {
+        PyObject *instance_class = NULL;
+        if (value && PyExceptionInstance_Check(value)) {
+            instance_class = (PyObject*) Py_TYPE(value);
+            if (instance_class != type) {
+                int is_subclass = PyObject_IsSubclass(instance_class, type);
+                if (!is_subclass) {
+                    instance_class = NULL;
+                } else if (unlikely(is_subclass == -1)) {
+                    goto bad;
+                } else {
+                    type = instance_class;
+                }
+            }
+        }
+        if (!instance_class) {
+            PyObject *args;
+            if (!value)
+                args = PyTuple_New(0);
+            else if (PyTuple_Check(value)) {
+                Py_INCREF(value);
+                args = value;
+            } else
+                args = PyTuple_Pack(1, value);
+            if (!args)
+                goto bad;
+            owned_instance = PyObject_Call(type, args, NULL);
+            Py_DECREF(args);
+            if (!owned_instance)
+                goto bad;
+            value = owned_instance;
+            if (!PyExceptionInstance_Check(value)) {
+                PyErr_Format(PyExc_TypeError,
+                             "calling %R should have returned an instance of "
+                             "BaseException, not %R",
+                             type, Py_TYPE(value));
+                goto bad;
+            }
+        }
+    } else {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: exception class must be a subclass of BaseException");
+        goto bad;
+    }
+    if (cause) {
+        PyObject *fixed_cause;
+        if (cause == Py_None) {
+            fixed_cause = NULL;
+        } else if (PyExceptionClass_Check(cause)) {
+            fixed_cause = PyObject_CallObject(cause, NULL);
+            if (fixed_cause == NULL)
+                goto bad;
+        } else if (PyExceptionInstance_Check(cause)) {
+            fixed_cause = cause;
+            Py_INCREF(fixed_cause);
+        } else {
+            PyErr_SetString(PyExc_TypeError,
+                            "exception causes must derive from "
+                            "BaseException");
+            goto bad;
+        }
+        PyException_SetCause(value, fixed_cause);
+    }
+    PyErr_SetObject(type, value);
+    if (tb) {
+#if CYTHON_COMPILING_IN_PYPY
+        PyObject *tmp_type, *tmp_value, *tmp_tb;
+        PyErr_Fetch(&tmp_type, &tmp_value, &tmp_tb);
+        Py_INCREF(tb);
+        PyErr_Restore(tmp_type, tmp_value, tb);
+        Py_XDECREF(tmp_tb);
+#else
+        PyThreadState *tstate = __Pyx_PyThreadState_Current;
+        PyObject* tmp_tb = tstate->curexc_traceback;
+        if (tb != tmp_tb) {
+            Py_INCREF(tb);
+            tstate->curexc_traceback = tb;
+            Py_XDECREF(tmp_tb);
+        }
+#endif
+    }
+bad:
+    Py_XDECREF(owned_instance);
+    return;
+}
+#endif
+
+/* PyCFunctionFastCall */
+  #if CYTHON_FAST_PYCCALL
+static CYTHON_INLINE PyObject * __Pyx_PyCFunction_FastCall(PyObject *func_obj, PyObject **args, Py_ssize_t nargs) {
+    PyCFunctionObject *func = (PyCFunctionObject*)func_obj;
+    PyCFunction meth = PyCFunction_GET_FUNCTION(func);
+    PyObject *self = PyCFunction_GET_SELF(func);
+    int flags = PyCFunction_GET_FLAGS(func);
+    assert(PyCFunction_Check(func));
+    assert(METH_FASTCALL == (flags & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS)));
+    assert(nargs >= 0);
+    assert(nargs == 0 || args != NULL);
+    /* _PyCFunction_FastCallDict() must not be called with an exception set,
+       because it may clear it (directly or indirectly) and so the
+       caller loses its exception */
+    assert(!PyErr_Occurred());
+    if ((PY_VERSION_HEX < 0x030700A0) || unlikely(flags & METH_KEYWORDS)) {
+        return (*((__Pyx_PyCFunctionFastWithKeywords)meth)) (self, args, nargs, NULL);
+    } else {
+        return (*((__Pyx_PyCFunctionFast)meth)) (self, args, nargs);
+    }
+}
+#endif
+
+/* PyFunctionFastCall */
+  #if CYTHON_FAST_PYCALL
+#include "frameobject.h"
+static PyObject* __Pyx_PyFunction_FastCallNoKw(PyCodeObject *co, PyObject **args, Py_ssize_t na,
+                                               PyObject *globals) {
+    PyFrameObject *f;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    PyObject **fastlocals;
+    Py_ssize_t i;
+    PyObject *result;
+    assert(globals != NULL);
+    /* XXX Perhaps we should create a specialized
+       PyFrame_New() that doesn't take locals, but does
+       take builtins without sanity checking them.
+       */
+    assert(tstate != NULL);
+    f = PyFrame_New(tstate, co, globals, NULL);
+    if (f == NULL) {
+        return NULL;
+    }
+    fastlocals = f->f_localsplus;
+    for (i = 0; i < na; i++) {
+        Py_INCREF(*args);
+        fastlocals[i] = *args++;
+    }
+    result = PyEval_EvalFrameEx(f,0);
+    ++tstate->recursion_depth;
+    Py_DECREF(f);
+    --tstate->recursion_depth;
+    return result;
+}
+#if 1 || PY_VERSION_HEX < 0x030600B1
+static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, int nargs, PyObject *kwargs) {
+    PyCodeObject *co = (PyCodeObject *)PyFunction_GET_CODE(func);
+    PyObject *globals = PyFunction_GET_GLOBALS(func);
+    PyObject *argdefs = PyFunction_GET_DEFAULTS(func);
+    PyObject *closure;
+#if PY_MAJOR_VERSION >= 3
+    PyObject *kwdefs;
+#endif
+    PyObject *kwtuple, **k;
+    PyObject **d;
+    Py_ssize_t nd;
+    Py_ssize_t nk;
+    PyObject *result;
+    assert(kwargs == NULL || PyDict_Check(kwargs));
+    nk = kwargs ? PyDict_Size(kwargs) : 0;
+    if (Py_EnterRecursiveCall((char*)" while calling a Python object")) {
+        return NULL;
+    }
+    if (
+#if PY_MAJOR_VERSION >= 3
+            co->co_kwonlyargcount == 0 &&
+#endif
+            likely(kwargs == NULL || nk == 0) &&
+            co->co_flags == (CO_OPTIMIZED | CO_NEWLOCALS | CO_NOFREE)) {
+        if (argdefs == NULL && co->co_argcount == nargs) {
+            result = __Pyx_PyFunction_FastCallNoKw(co, args, nargs, globals);
+            goto done;
+        }
+        else if (nargs == 0 && argdefs != NULL
+                 && co->co_argcount == Py_SIZE(argdefs)) {
+            /* function called with no arguments, but all parameters have
+               a default value: use default values as arguments .*/
+            args = &PyTuple_GET_ITEM(argdefs, 0);
+            result =__Pyx_PyFunction_FastCallNoKw(co, args, Py_SIZE(argdefs), globals);
+            goto done;
+        }
+    }
+    if (kwargs != NULL) {
+        Py_ssize_t pos, i;
+        kwtuple = PyTuple_New(2 * nk);
+        if (kwtuple == NULL) {
+            result = NULL;
+            goto done;
+        }
+        k = &PyTuple_GET_ITEM(kwtuple, 0);
+        pos = i = 0;
+        while (PyDict_Next(kwargs, &pos, &k[i], &k[i+1])) {
+            Py_INCREF(k[i]);
+            Py_INCREF(k[i+1]);
+            i += 2;
+        }
+        nk = i / 2;
+    }
+    else {
+        kwtuple = NULL;
+        k = NULL;
+    }
+    closure = PyFunction_GET_CLOSURE(func);
+#if PY_MAJOR_VERSION >= 3
+    kwdefs = PyFunction_GET_KW_DEFAULTS(func);
+#endif
+    if (argdefs != NULL) {
+        d = &PyTuple_GET_ITEM(argdefs, 0);
+        nd = Py_SIZE(argdefs);
+    }
+    else {
+        d = NULL;
+        nd = 0;
+    }
+#if PY_MAJOR_VERSION >= 3
+    result = PyEval_EvalCodeEx((PyObject*)co, globals, (PyObject *)NULL,
+                               args, nargs,
+                               k, (int)nk,
+                               d, (int)nd, kwdefs, closure);
+#else
+    result = PyEval_EvalCodeEx(co, globals, (PyObject *)NULL,
+                               args, nargs,
+                               k, (int)nk,
+                               d, (int)nd, closure);
+#endif
+    Py_XDECREF(kwtuple);
+done:
+    Py_LeaveRecursiveCall();
+    return result;
+}
+#endif
+#endif
+
+/* PyObjectCallMethO */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg) {
+    PyObject *self, *result;
+    PyCFunction cfunc;
+    cfunc = PyCFunction_GET_FUNCTION(func);
+    self = PyCFunction_GET_SELF(func);
+    if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object")))
+        return NULL;
+    result = cfunc(self, arg);
+    Py_LeaveRecursiveCall();
+    if (unlikely(!result) && unlikely(!PyErr_Occurred())) {
+        PyErr_SetString(
+            PyExc_SystemError,
+            "NULL result without error in PyObject_Call");
+    }
+    return result;
+}
+#endif
+
+/* PyObjectCallOneArg */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static PyObject* __Pyx__PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+    PyObject *result;
+    PyObject *args = PyTuple_New(1);
+    if (unlikely(!args)) return NULL;
+    Py_INCREF(arg);
+    PyTuple_SET_ITEM(args, 0, arg);
+    result = __Pyx_PyObject_Call(func, args, NULL);
+    Py_DECREF(args);
+    return result;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+#if CYTHON_FAST_PYCALL
+    if (PyFunction_Check(func)) {
+        return __Pyx_PyFunction_FastCall(func, &arg, 1);
+    }
+#endif
+    if (likely(PyCFunction_Check(func))) {
+        if (likely(PyCFunction_GET_FLAGS(func) & METH_O)) {
+            return __Pyx_PyObject_CallMethO(func, arg);
+#if CYTHON_FAST_PYCCALL
+        } else if (PyCFunction_GET_FLAGS(func) & METH_FASTCALL) {
+            return __Pyx_PyCFunction_FastCall(func, &arg, 1);
+#endif
+        }
+    }
+    return __Pyx__PyObject_CallOneArg(func, arg);
+}
+#else
+static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) {
+    PyObject *result;
+    PyObject *args = PyTuple_Pack(1, arg);
+    if (unlikely(!args)) return NULL;
+    result = __Pyx_PyObject_Call(func, args, NULL);
+    Py_DECREF(args);
+    return result;
+}
+#endif
+
+/* DictGetItem */
+  #if PY_MAJOR_VERSION >= 3 && !CYTHON_COMPILING_IN_PYPY
+static PyObject *__Pyx_PyDict_GetItem(PyObject *d, PyObject* key) {
+    PyObject *value;
+    value = PyDict_GetItemWithError(d, key);
+    if (unlikely(!value)) {
+        if (!PyErr_Occurred()) {
+            PyObject* args = PyTuple_Pack(1, key);
+            if (likely(args))
+                PyErr_SetObject(PyExc_KeyError, args);
+            Py_XDECREF(args);
+        }
+        return NULL;
+    }
+    Py_INCREF(value);
+    return value;
+}
+#endif
+
+/* RaiseTooManyValuesToUnpack */
+  static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected) {
+    PyErr_Format(PyExc_ValueError,
+                 "too many values to unpack (expected %" CYTHON_FORMAT_SSIZE_T "d)", expected);
+}
+
+/* RaiseNeedMoreValuesToUnpack */
+  static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index) {
+    PyErr_Format(PyExc_ValueError,
+                 "need more than %" CYTHON_FORMAT_SSIZE_T "d value%.1s to unpack",
+                 index, (index == 1) ? "" : "s");
+}
+
+/* RaiseNoneIterError */
+  static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void) {
+    PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable");
+}
+
+/* ExtTypeTest */
+  static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type) {
+    if (unlikely(!type)) {
+        PyErr_SetString(PyExc_SystemError, "Missing type object");
+        return 0;
+    }
+    if (likely(__Pyx_TypeCheck(obj, type)))
+        return 1;
+    PyErr_Format(PyExc_TypeError, "Cannot convert %.200s to %.200s",
+                 Py_TYPE(obj)->tp_name, type->tp_name);
+    return 0;
+}
+
+/* SaveResetException */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    #if PY_VERSION_HEX >= 0x030700A2
+    *type = tstate->exc_state.exc_type;
+    *value = tstate->exc_state.exc_value;
+    *tb = tstate->exc_state.exc_traceback;
+    #else
+    *type = tstate->exc_type;
+    *value = tstate->exc_value;
+    *tb = tstate->exc_traceback;
+    #endif
+    Py_XINCREF(*type);
+    Py_XINCREF(*value);
+    Py_XINCREF(*tb);
+}
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    #if PY_VERSION_HEX >= 0x030700A2
+    tmp_type = tstate->exc_state.exc_type;
+    tmp_value = tstate->exc_state.exc_value;
+    tmp_tb = tstate->exc_state.exc_traceback;
+    tstate->exc_state.exc_type = type;
+    tstate->exc_state.exc_value = value;
+    tstate->exc_state.exc_traceback = tb;
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = type;
+    tstate->exc_value = value;
+    tstate->exc_traceback = tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+#endif
+
+/* PyErrExceptionMatches */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx_PyErr_ExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        if (__Pyx_PyErr_GivenExceptionMatches(exc_type, PyTuple_GET_ITEM(tuple, i))) return 1;
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err) {
+    PyObject *exc_type = tstate->curexc_type;
+    if (exc_type == err) return 1;
+    if (unlikely(!exc_type)) return 0;
+    if (unlikely(PyTuple_Check(err)))
+        return __Pyx_PyErr_ExceptionMatchesTuple(exc_type, err);
+    return __Pyx_PyErr_GivenExceptionMatches(exc_type, err);
+}
+#endif
+
+/* GetException */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb) {
+#endif
+    PyObject *local_type, *local_value, *local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    local_type = tstate->curexc_type;
+    local_value = tstate->curexc_value;
+    local_tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+#else
+    PyErr_Fetch(&local_type, &local_value, &local_tb);
+#endif
+    PyErr_NormalizeException(&local_type, &local_value, &local_tb);
+#if CYTHON_FAST_THREAD_STATE
+    if (unlikely(tstate->curexc_type))
+#else
+    if (unlikely(PyErr_Occurred()))
+#endif
+        goto bad;
+    #if PY_MAJOR_VERSION >= 3
+    if (local_tb) {
+        if (unlikely(PyException_SetTraceback(local_value, local_tb) < 0))
+            goto bad;
+    }
+    #endif
+    Py_XINCREF(local_tb);
+    Py_XINCREF(local_type);
+    Py_XINCREF(local_value);
+    *type = local_type;
+    *value = local_value;
+    *tb = local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    #if PY_VERSION_HEX >= 0x030700A2
+    tmp_type = tstate->exc_state.exc_type;
+    tmp_value = tstate->exc_state.exc_value;
+    tmp_tb = tstate->exc_state.exc_traceback;
+    tstate->exc_state.exc_type = local_type;
+    tstate->exc_state.exc_value = local_value;
+    tstate->exc_state.exc_traceback = local_tb;
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = local_type;
+    tstate->exc_value = local_value;
+    tstate->exc_traceback = local_tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+#else
+    PyErr_SetExcInfo(local_type, local_value, local_tb);
+#endif
+    return 0;
+bad:
+    *type = 0;
+    *value = 0;
+    *tb = 0;
+    Py_XDECREF(local_type);
+    Py_XDECREF(local_value);
+    Py_XDECREF(local_tb);
+    return -1;
+}
+
+/* Import */
+    static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) {
+    PyObject *empty_list = 0;
+    PyObject *module = 0;
+    PyObject *global_dict = 0;
+    PyObject *empty_dict = 0;
+    PyObject *list;
+    #if PY_MAJOR_VERSION < 3
+    PyObject *py_import;
+    py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import);
+    if (!py_import)
+        goto bad;
+    #endif
+    if (from_list)
+        list = from_list;
+    else {
+        empty_list = PyList_New(0);
+        if (!empty_list)
+            goto bad;
+        list = empty_list;
+    }
+    global_dict = PyModule_GetDict(__pyx_m);
+    if (!global_dict)
+        goto bad;
+    empty_dict = PyDict_New();
+    if (!empty_dict)
+        goto bad;
+    {
+        #if PY_MAJOR_VERSION >= 3
+        if (level == -1) {
+            if (strchr(__Pyx_MODULE_NAME, '.')) {
+                module = PyImport_ImportModuleLevelObject(
+                    name, global_dict, empty_dict, list, 1);
+                if (!module) {
+                    if (!PyErr_ExceptionMatches(PyExc_ImportError))
+                        goto bad;
+                    PyErr_Clear();
+                }
+            }
+            level = 0;
+        }
+        #endif
+        if (!module) {
+            #if PY_MAJOR_VERSION < 3
+            PyObject *py_level = PyInt_FromLong(level);
+            if (!py_level)
+                goto bad;
+            module = PyObject_CallFunctionObjArgs(py_import,
+                name, global_dict, empty_dict, list, py_level, NULL);
+            Py_DECREF(py_level);
+            #else
+            module = PyImport_ImportModuleLevelObject(
+                name, global_dict, empty_dict, list, level);
+            #endif
+        }
+    }
+bad:
+    #if PY_MAJOR_VERSION < 3
+    Py_XDECREF(py_import);
+    #endif
+    Py_XDECREF(empty_list);
+    Py_XDECREF(empty_dict);
+    return module;
+}
+
+/* CLineInTraceback */
+    #ifndef CYTHON_CLINE_IN_TRACEBACK
+static int __Pyx_CLineForTraceback(CYTHON_UNUSED PyThreadState *tstate, int c_line) {
+    PyObject *use_cline;
+    PyObject *ptype, *pvalue, *ptraceback;
+#if CYTHON_COMPILING_IN_CPYTHON
+    PyObject **cython_runtime_dict;
+#endif
+    __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback);
+#if CYTHON_COMPILING_IN_CPYTHON
+    cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime);
+    if (likely(cython_runtime_dict)) {
+      use_cline = __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback);
+    } else
+#endif
+    {
+      PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback);
+      if (use_cline_obj) {
+        use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True;
+        Py_DECREF(use_cline_obj);
+      } else {
+        PyErr_Clear();
+        use_cline = NULL;
+      }
+    }
+    if (!use_cline) {
+        c_line = 0;
+        PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False);
+    }
+    else if (PyObject_Not(use_cline) != 0) {
+        c_line = 0;
+    }
+    __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback);
+    return c_line;
+}
+#endif
+
+/* CodeObjectCache */
+    static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) {
+    int start = 0, mid = 0, end = count - 1;
+    if (end >= 0 && code_line > entries[end].code_line) {
+        return count;
+    }
+    while (start < end) {
+        mid = start + (end - start) / 2;
+        if (code_line < entries[mid].code_line) {
+            end = mid;
+        } else if (code_line > entries[mid].code_line) {
+             start = mid + 1;
+        } else {
+            return mid;
+        }
+    }
+    if (code_line <= entries[mid].code_line) {
+        return mid;
+    } else {
+        return mid + 1;
+    }
+}
+static PyCodeObject *__pyx_find_code_object(int code_line) {
+    PyCodeObject* code_object;
+    int pos;
+    if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) {
+        return NULL;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) {
+        return NULL;
+    }
+    code_object = __pyx_code_cache.entries[pos].code_object;
+    Py_INCREF(code_object);
+    return code_object;
+}
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) {
+    int pos, i;
+    __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries;
+    if (unlikely(!code_line)) {
+        return;
+    }
+    if (unlikely(!entries)) {
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (likely(entries)) {
+            __pyx_code_cache.entries = entries;
+            __pyx_code_cache.max_count = 64;
+            __pyx_code_cache.count = 1;
+            entries[0].code_line = code_line;
+            entries[0].code_object = code_object;
+            Py_INCREF(code_object);
+        }
+        return;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) {
+        PyCodeObject* tmp = entries[pos].code_object;
+        entries[pos].code_object = code_object;
+        Py_DECREF(tmp);
+        return;
+    }
+    if (__pyx_code_cache.count == __pyx_code_cache.max_count) {
+        int new_max = __pyx_code_cache.max_count + 64;
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc(
+            __pyx_code_cache.entries, (size_t)new_max*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (unlikely(!entries)) {
+            return;
+        }
+        __pyx_code_cache.entries = entries;
+        __pyx_code_cache.max_count = new_max;
+    }
+    for (i=__pyx_code_cache.count; i>pos; i--) {
+        entries[i] = entries[i-1];
+    }
+    entries[pos].code_line = code_line;
+    entries[pos].code_object = code_object;
+    __pyx_code_cache.count++;
+    Py_INCREF(code_object);
+}
+
+/* AddTraceback */
+    #include "compile.h"
+#include "frameobject.h"
+#include "traceback.h"
+static PyCodeObject* __Pyx_CreateCodeObjectForTraceback(
+            const char *funcname, int c_line,
+            int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyObject *py_srcfile = 0;
+    PyObject *py_funcname = 0;
+    #if PY_MAJOR_VERSION < 3
+    py_srcfile = PyString_FromString(filename);
+    #else
+    py_srcfile = PyUnicode_FromString(filename);
+    #endif
+    if (!py_srcfile) goto bad;
+    if (c_line) {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #else
+        py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        #endif
+    }
+    else {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromString(funcname);
+        #else
+        py_funcname = PyUnicode_FromString(funcname);
+        #endif
+    }
+    if (!py_funcname) goto bad;
+    py_code = __Pyx_PyCode_New(
+        0,
+        0,
+        0,
+        0,
+        0,
+        __pyx_empty_bytes, /*PyObject *code,*/
+        __pyx_empty_tuple, /*PyObject *consts,*/
+        __pyx_empty_tuple, /*PyObject *names,*/
+        __pyx_empty_tuple, /*PyObject *varnames,*/
+        __pyx_empty_tuple, /*PyObject *freevars,*/
+        __pyx_empty_tuple, /*PyObject *cellvars,*/
+        py_srcfile,   /*PyObject *filename,*/
+        py_funcname,  /*PyObject *name,*/
+        py_line,
+        __pyx_empty_bytes  /*PyObject *lnotab*/
+    );
+    Py_DECREF(py_srcfile);
+    Py_DECREF(py_funcname);
+    return py_code;
+bad:
+    Py_XDECREF(py_srcfile);
+    Py_XDECREF(py_funcname);
+    return NULL;
+}
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyFrameObject *py_frame = 0;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    if (c_line) {
+        c_line = __Pyx_CLineForTraceback(tstate, c_line);
+    }
+    py_code = __pyx_find_code_object(c_line ? -c_line : py_line);
+    if (!py_code) {
+        py_code = __Pyx_CreateCodeObjectForTraceback(
+            funcname, c_line, py_line, filename);
+        if (!py_code) goto bad;
+        __pyx_insert_code_object(c_line ? -c_line : py_line, py_code);
+    }
+    py_frame = PyFrame_New(
+        tstate,            /*PyThreadState *tstate,*/
+        py_code,           /*PyCodeObject *code,*/
+        __pyx_d,    /*PyObject *globals,*/
+        0                  /*PyObject *locals*/
+    );
+    if (!py_frame) goto bad;
+    __Pyx_PyFrame_SetLineNumber(py_frame, py_line);
+    PyTraceBack_Here(py_frame);
+bad:
+    Py_XDECREF(py_code);
+    Py_XDECREF(py_frame);
+}
+
+#if PY_MAJOR_VERSION < 3
+static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags) {
+    if (PyObject_CheckBuffer(obj)) return PyObject_GetBuffer(obj, view, flags);
+        if (__Pyx_TypeCheck(obj, __pyx_ptype_5numpy_ndarray)) return __pyx_pw_5numpy_7ndarray_1__getbuffer__(obj, view, flags);
+    PyErr_Format(PyExc_TypeError, "'%.200s' does not have the buffer interface", Py_TYPE(obj)->tp_name);
+    return -1;
+}
+static void __Pyx_ReleaseBuffer(Py_buffer *view) {
+    PyObject *obj = view->obj;
+    if (!obj) return;
+    if (PyObject_CheckBuffer(obj)) {
+        PyBuffer_Release(view);
+        return;
+    }
+    if ((0)) {}
+        else if (__Pyx_TypeCheck(obj, __pyx_ptype_5numpy_ndarray)) __pyx_pw_5numpy_7ndarray_3__releasebuffer__(obj, view);
+    view->obj = NULL;
+    Py_DECREF(obj);
+}
+#endif
+
+
+    /* CIntFromPyVerify */
+    #define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0)
+#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1)
+#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\
+    {\
+        func_type value = func_value;\
+        if (sizeof(target_type) < sizeof(func_type)) {\
+            if (unlikely(value != (func_type) (target_type) value)) {\
+                func_type zero = 0;\
+                if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\
+                    return (target_type) -1;\
+                if (is_unsigned && unlikely(value < zero))\
+                    goto raise_neg_overflow;\
+                else\
+                    goto raise_overflow;\
+            }\
+        }\
+        return (target_type) value;\
+    }
+
+/* Declarations */
+    #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return ::std::complex< float >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return x + y*(__pyx_t_float_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      __pyx_t_float_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+    #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabsf(b.real) >= fabsf(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                float r = b.imag / b.real;
+                float s = 1.0 / (b.real + b.imag * r);
+                return __pyx_t_float_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            float r = b.real / b.imag;
+            float s = 1.0 / (b.imag + b.real * r);
+            return __pyx_t_float_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            float denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_float_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrtf(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypotf(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+            __pyx_t_float_complex z;
+            float r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    float denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = powf(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2f(0, -1);
+                }
+            } else {
+                r = __Pyx_c_abs_float(a);
+                theta = atan2f(a.imag, a.real);
+            }
+            lnr = logf(r);
+            z_r = expf(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cosf(z_theta);
+            z.imag = z_r * sinf(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* Declarations */
+    #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return ::std::complex< double >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return x + y*(__pyx_t_double_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      __pyx_t_double_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+    #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabs(b.real) >= fabs(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                double r = b.imag / b.real;
+                double s = 1.0 / (b.real + b.imag * r);
+                return __pyx_t_double_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            double r = b.real / b.imag;
+            double s = 1.0 / (b.imag + b.real * r);
+            return __pyx_t_double_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            double denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_double_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrt(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypot(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+            __pyx_t_double_complex z;
+            double r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    double denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = pow(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2(0, -1);
+                }
+            } else {
+                r = __Pyx_c_abs_double(a);
+                theta = atan2(a.imag, a.real);
+            }
+            lnr = log(r);
+            z_r = exp(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cos(z_theta);
+            z.imag = z_r * sin(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* CIntToPy */
+    static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value) {
+    const int neg_one = (int) -1, const_zero = (int) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(int) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(int) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(int) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(int),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntToPy */
+    static CYTHON_INLINE PyObject* __Pyx_PyInt_From_enum__NPY_TYPES(enum NPY_TYPES value) {
+    const enum NPY_TYPES neg_one = (enum NPY_TYPES) -1, const_zero = (enum NPY_TYPES) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(enum NPY_TYPES) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(enum NPY_TYPES) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(enum NPY_TYPES) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(enum NPY_TYPES),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntFromPy */
+    static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) {
+    const int neg_one = (int) -1, const_zero = (int) 0;
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(int) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (int) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(int, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 2 * PyLong_SHIFT) {
+                            return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 3 * PyLong_SHIFT) {
+                            return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 4 * PyLong_SHIFT) {
+                            return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (int) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(int) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(int, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(int,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(int) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(int) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            int val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (int) -1;
+        }
+    } else {
+        int val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (int) -1;
+        val = __Pyx_PyInt_As_int(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to int");
+    return (int) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to int");
+    return (int) -1;
+}
+
+/* CIntToPy */
+    static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) {
+    const long neg_one = (long) -1, const_zero = (long) 0;
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(long) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(long) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(long) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(long),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntFromPy */
+    static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) {
+    const long neg_one = (long) -1, const_zero = (long) 0;
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(long) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (long) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(long, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 2 * PyLong_SHIFT) {
+                            return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 3 * PyLong_SHIFT) {
+                            return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 4 * PyLong_SHIFT) {
+                            return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (long) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(long) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(long, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(long,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(long) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(long) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            long val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (long) -1;
+        }
+    } else {
+        long val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (long) -1;
+        val = __Pyx_PyInt_As_long(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to long");
+    return (long) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to long");
+    return (long) -1;
+}
+
+/* FastTypeChecks */
+    #if CYTHON_COMPILING_IN_CPYTHON
+static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) {
+    while (a) {
+        a = a->tp_base;
+        if (a == b)
+            return 1;
+    }
+    return b == &PyBaseObject_Type;
+}
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) {
+    PyObject *mro;
+    if (a == b) return 1;
+    mro = a->tp_mro;
+    if (likely(mro)) {
+        Py_ssize_t i, n;
+        n = PyTuple_GET_SIZE(mro);
+        for (i = 0; i < n; i++) {
+            if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b)
+                return 1;
+        }
+        return 0;
+    }
+    return __Pyx_InBases(a, b);
+}
+#if PY_MAJOR_VERSION == 2
+static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) {
+    PyObject *exception, *value, *tb;
+    int res;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&exception, &value, &tb);
+    res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0;
+    if (unlikely(res == -1)) {
+        PyErr_WriteUnraisable(err);
+        res = 0;
+    }
+    if (!res) {
+        res = PyObject_IsSubclass(err, exc_type2);
+        if (unlikely(res == -1)) {
+            PyErr_WriteUnraisable(err);
+            res = 0;
+        }
+    }
+    __Pyx_ErrRestore(exception, value, tb);
+    return res;
+}
+#else
+static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) {
+    int res = exc_type1 ? __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type1) : 0;
+    if (!res) {
+        res = __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2);
+    }
+    return res;
+}
+#endif
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject* exc_type) {
+    if (likely(err == exc_type)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        return __Pyx_inner_PyErr_GivenExceptionMatches2(err, NULL, exc_type);
+    }
+    return PyErr_GivenExceptionMatches(err, exc_type);
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *exc_type1, PyObject *exc_type2) {
+    if (likely(err == exc_type1 || err == exc_type2)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        return __Pyx_inner_PyErr_GivenExceptionMatches2(err, exc_type1, exc_type2);
+    }
+    return (PyErr_GivenExceptionMatches(err, exc_type1) || PyErr_GivenExceptionMatches(err, exc_type2));
+}
+#endif
+
+/* CheckBinaryVersion */
+    static int __Pyx_check_binary_version(void) {
+    char ctversion[4], rtversion[4];
+    PyOS_snprintf(ctversion, 4, "%d.%d", PY_MAJOR_VERSION, PY_MINOR_VERSION);
+    PyOS_snprintf(rtversion, 4, "%s", Py_GetVersion());
+    if (ctversion[0] != rtversion[0] || ctversion[2] != rtversion[2]) {
+        char message[200];
+        PyOS_snprintf(message, sizeof(message),
+                      "compiletime version %s of module '%.100s' "
+                      "does not match runtime version %s",
+                      ctversion, __Pyx_MODULE_NAME, rtversion);
+        return PyErr_WarnEx(NULL, message, 1);
+    }
+    return 0;
+}
+
+/* ModuleImport */
+    #ifndef __PYX_HAVE_RT_ImportModule
+#define __PYX_HAVE_RT_ImportModule
+static PyObject *__Pyx_ImportModule(const char *name) {
+    PyObject *py_name = 0;
+    PyObject *py_module = 0;
+    py_name = __Pyx_PyIdentifier_FromString(name);
+    if (!py_name)
+        goto bad;
+    py_module = PyImport_Import(py_name);
+    Py_DECREF(py_name);
+    return py_module;
+bad:
+    Py_XDECREF(py_name);
+    return 0;
+}
+#endif
+
+/* TypeImport */
+    #ifndef __PYX_HAVE_RT_ImportType
+#define __PYX_HAVE_RT_ImportType
+static PyTypeObject *__Pyx_ImportType(const char *module_name, const char *class_name,
+    size_t size, int strict)
+{
+    PyObject *py_module = 0;
+    PyObject *result = 0;
+    PyObject *py_name = 0;
+    char warning[200];
+    Py_ssize_t basicsize;
+#ifdef Py_LIMITED_API
+    PyObject *py_basicsize;
+#endif
+    py_module = __Pyx_ImportModule(module_name);
+    if (!py_module)
+        goto bad;
+    py_name = __Pyx_PyIdentifier_FromString(class_name);
+    if (!py_name)
+        goto bad;
+    result = PyObject_GetAttr(py_module, py_name);
+    Py_DECREF(py_name);
+    py_name = 0;
+    Py_DECREF(py_module);
+    py_module = 0;
+    if (!result)
+        goto bad;
+    if (!PyType_Check(result)) {
+        PyErr_Format(PyExc_TypeError,
+            "%.200s.%.200s is not a type object",
+            module_name, class_name);
+        goto bad;
+    }
+#ifndef Py_LIMITED_API
+    basicsize = ((PyTypeObject *)result)->tp_basicsize;
+#else
+    py_basicsize = PyObject_GetAttrString(result, "__basicsize__");
+    if (!py_basicsize)
+        goto bad;
+    basicsize = PyLong_AsSsize_t(py_basicsize);
+    Py_DECREF(py_basicsize);
+    py_basicsize = 0;
+    if (basicsize == (Py_ssize_t)-1 && PyErr_Occurred())
+        goto bad;
+#endif
+    if (!strict && (size_t)basicsize > size) {
+        PyOS_snprintf(warning, sizeof(warning),
+            "%s.%s size changed, may indicate binary incompatibility. Expected %zd, got %zd",
+            module_name, class_name, basicsize, size);
+        if (PyErr_WarnEx(NULL, warning, 0) < 0) goto bad;
+    }
+    else if ((size_t)basicsize != size) {
+        PyErr_Format(PyExc_ValueError,
+            "%.200s.%.200s has the wrong size, try recompiling. Expected %zd, got %zd",
+            module_name, class_name, basicsize, size);
+        goto bad;
+    }
+    return (PyTypeObject *)result;
+bad:
+    Py_XDECREF(py_module);
+    Py_XDECREF(result);
+    return NULL;
+}
+#endif
+
+/* InitStrings */
+    static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) {
+    while (t->p) {
+        #if PY_MAJOR_VERSION < 3
+        if (t->is_unicode) {
+            *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL);
+        } else if (t->intern) {
+            *t->p = PyString_InternFromString(t->s);
+        } else {
+            *t->p = PyString_FromStringAndSize(t->s, t->n - 1);
+        }
+        #else
+        if (t->is_unicode | t->is_str) {
+            if (t->intern) {
+                *t->p = PyUnicode_InternFromString(t->s);
+            } else if (t->encoding) {
+                *t->p = PyUnicode_Decode(t->s, t->n - 1, t->encoding, NULL);
+            } else {
+                *t->p = PyUnicode_FromStringAndSize(t->s, t->n - 1);
+            }
+        } else {
+            *t->p = PyBytes_FromStringAndSize(t->s, t->n - 1);
+        }
+        #endif
+        if (!*t->p)
+            return -1;
+        if (PyObject_Hash(*t->p) == -1)
+            return -1;
+        ++t;
+    }
+    return 0;
+}
+
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) {
+    return __Pyx_PyUnicode_FromStringAndSize(c_str, (Py_ssize_t)strlen(c_str));
+}
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) {
+    Py_ssize_t ignore;
+    return __Pyx_PyObject_AsStringAndSize(o, &ignore);
+}
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+#if !CYTHON_PEP393_ENABLED
+static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    char* defenc_c;
+    PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL);
+    if (!defenc) return NULL;
+    defenc_c = PyBytes_AS_STRING(defenc);
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    {
+        char* end = defenc_c + PyBytes_GET_SIZE(defenc);
+        char* c;
+        for (c = defenc_c; c < end; c++) {
+            if ((unsigned char) (*c) >= 128) {
+                PyUnicode_AsASCIIString(o);
+                return NULL;
+            }
+        }
+    }
+#endif
+    *length = PyBytes_GET_SIZE(defenc);
+    return defenc_c;
+}
+#else
+static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL;
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    if (likely(PyUnicode_IS_ASCII(o))) {
+        *length = PyUnicode_GET_LENGTH(o);
+        return PyUnicode_AsUTF8(o);
+    } else {
+        PyUnicode_AsASCIIString(o);
+        return NULL;
+    }
+#else
+    return PyUnicode_AsUTF8AndSize(o, length);
+#endif
+}
+#endif
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+    if (
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+            __Pyx_sys_getdefaultencoding_not_ascii &&
+#endif
+            PyUnicode_Check(o)) {
+        return __Pyx_PyUnicode_AsStringAndSize(o, length);
+    } else
+#endif
+#if (!CYTHON_COMPILING_IN_PYPY) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE))
+    if (PyByteArray_Check(o)) {
+        *length = PyByteArray_GET_SIZE(o);
+        return PyByteArray_AS_STRING(o);
+    } else
+#endif
+    {
+        char* result;
+        int r = PyBytes_AsStringAndSize(o, &result, length);
+        if (unlikely(r < 0)) {
+            return NULL;
+        } else {
+            return result;
+        }
+    }
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) {
+   int is_true = x == Py_True;
+   if (is_true | (x == Py_False) | (x == Py_None)) return is_true;
+   else return PyObject_IsTrue(x);
+}
+static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) {
+#if PY_MAJOR_VERSION >= 3
+    if (PyLong_Check(result)) {
+        if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1,
+                "__int__ returned non-int (type %.200s).  "
+                "The ability to return an instance of a strict subclass of int "
+                "is deprecated, and may be removed in a future version of Python.",
+                Py_TYPE(result)->tp_name)) {
+            Py_DECREF(result);
+            return NULL;
+        }
+        return result;
+    }
+#endif
+    PyErr_Format(PyExc_TypeError,
+                 "__%.4s__ returned non-%.4s (type %.200s)",
+                 type_name, type_name, Py_TYPE(result)->tp_name);
+    Py_DECREF(result);
+    return NULL;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) {
+#if CYTHON_USE_TYPE_SLOTS
+  PyNumberMethods *m;
+#endif
+  const char *name = NULL;
+  PyObject *res = NULL;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_Check(x) || PyLong_Check(x)))
+#else
+  if (likely(PyLong_Check(x)))
+#endif
+    return __Pyx_NewRef(x);
+#if CYTHON_USE_TYPE_SLOTS
+  m = Py_TYPE(x)->tp_as_number;
+  #if PY_MAJOR_VERSION < 3
+  if (m && m->nb_int) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  else if (m && m->nb_long) {
+    name = "long";
+    res = m->nb_long(x);
+  }
+  #else
+  if (likely(m && m->nb_int)) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  #endif
+#else
+  if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) {
+    res = PyNumber_Int(x);
+  }
+#endif
+  if (likely(res)) {
+#if PY_MAJOR_VERSION < 3
+    if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) {
+#else
+    if (unlikely(!PyLong_CheckExact(res))) {
+#endif
+        return __Pyx_PyNumber_IntOrLongWrongResultType(res, name);
+    }
+  }
+  else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_TypeError,
+                    "an integer is required");
+  }
+  return res;
+}
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) {
+  Py_ssize_t ival;
+  PyObject *x;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_CheckExact(b))) {
+    if (sizeof(Py_ssize_t) >= sizeof(long))
+        return PyInt_AS_LONG(b);
+    else
+        return PyInt_AsSsize_t(x);
+  }
+#endif
+  if (likely(PyLong_CheckExact(b))) {
+    #if CYTHON_USE_PYLONG_INTERNALS
+    const digit* digits = ((PyLongObject*)b)->ob_digit;
+    const Py_ssize_t size = Py_SIZE(b);
+    if (likely(__Pyx_sst_abs(size) <= 1)) {
+        ival = likely(size) ? digits[0] : 0;
+        if (size == -1) ival = -ival;
+        return ival;
+    } else {
+      switch (size) {
+         case 2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+      }
+    }
+    #endif
+    return PyLong_AsSsize_t(b);
+  }
+  x = PyNumber_Index(b);
+  if (!x) return -1;
+  ival = PyInt_AsSsize_t(x);
+  Py_DECREF(x);
+  return ival;
+}
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) {
+    return PyInt_FromSize_t(ival);
+}
+
+
+#endif /* Py_PYTHON_H */
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.cpp b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.cpp
new file mode 100644
index 000000000..c5b441fb7
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.cpp
@@ -0,0 +1,7672 @@
+/* Generated by Cython 0.29.32 */
+
+/* BEGIN: Cython Metadata
+{
+    "distutils": {
+        "depends": [
+            "insightface/thirdparty/face3d/mesh/cython/mesh_core.h"
+        ],
+        "include_dirs": [
+            "insightface/thirdparty/face3d/mesh/cython"
+        ],
+        "language": "c++",
+        "name": "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython",
+        "sources": [
+            "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx",
+            "insightface/thirdparty/face3d/mesh/cython/mesh_core.cpp"
+        ]
+    },
+    "module_name": "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython"
+}
+END: Cython Metadata */
+
+#ifndef PY_SSIZE_T_CLEAN
+#define PY_SSIZE_T_CLEAN
+#endif /* PY_SSIZE_T_CLEAN */
+#include "Python.h"
+#ifndef Py_PYTHON_H
+    #error Python headers needed to compile C extensions, please install development version of Python.
+#elif PY_VERSION_HEX < 0x02060000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000)
+    #error Cython requires Python 2.6+ or Python 3.3+.
+#else
+#define CYTHON_ABI "0_29_32"
+#define CYTHON_HEX_VERSION 0x001D20F0
+#define CYTHON_FUTURE_DIVISION 0
+#include <stddef.h>
+#ifndef offsetof
+  #define offsetof(type, member) ( (size_t) & ((type*)0) -> member )
+#endif
+#if !defined(WIN32) && !defined(MS_WINDOWS)
+  #ifndef __stdcall
+    #define __stdcall
+  #endif
+  #ifndef __cdecl
+    #define __cdecl
+  #endif
+  #ifndef __fastcall
+    #define __fastcall
+  #endif
+#endif
+#ifndef DL_IMPORT
+  #define DL_IMPORT(t) t
+#endif
+#ifndef DL_EXPORT
+  #define DL_EXPORT(t) t
+#endif
+#define __PYX_COMMA ,
+#ifndef HAVE_LONG_LONG
+  #if PY_VERSION_HEX >= 0x02070000
+    #define HAVE_LONG_LONG
+  #endif
+#endif
+#ifndef PY_LONG_LONG
+  #define PY_LONG_LONG LONG_LONG
+#endif
+#ifndef Py_HUGE_VAL
+  #define Py_HUGE_VAL HUGE_VAL
+#endif
+#ifdef PYPY_VERSION
+  #define CYTHON_COMPILING_IN_PYPY 1
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 0
+  #undef CYTHON_USE_TYPE_SLOTS
+  #define CYTHON_USE_TYPE_SLOTS 0
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #if PY_VERSION_HEX < 0x03050000
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_INTERNALS
+  #define CYTHON_USE_UNICODE_INTERNALS 0
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #undef CYTHON_AVOID_BORROWED_REFS
+  #define CYTHON_AVOID_BORROWED_REFS 1
+  #undef CYTHON_ASSUME_SAFE_MACROS
+  #define CYTHON_ASSUME_SAFE_MACROS 0
+  #undef CYTHON_UNPACK_METHODS
+  #define CYTHON_UNPACK_METHODS 0
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+  #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC
+    #define CYTHON_UPDATE_DESCRIPTOR_DOC (PYPY_VERSION_HEX >= 0x07030900)
+  #endif
+#elif defined(PYSTON_VERSION)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 1
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 0
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #undef CYTHON_USE_ASYNC_SLOTS
+  #define CYTHON_USE_ASYNC_SLOTS 0
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #undef CYTHON_PEP489_MULTI_PHASE_INIT
+  #define CYTHON_PEP489_MULTI_PHASE_INIT 0
+  #undef CYTHON_USE_TP_FINALIZE
+  #define CYTHON_USE_TP_FINALIZE 0
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+  #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC
+    #define CYTHON_UPDATE_DESCRIPTOR_DOC 0
+  #endif
+#elif defined(PY_NOGIL)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 1
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #ifndef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT 1
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE 1
+  #endif
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
+#else
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 1
+  #define CYTHON_COMPILING_IN_NOGIL 0
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYTYPE_LOOKUP
+    #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #elif !defined(CYTHON_USE_PYTYPE_LOOKUP)
+    #define CYTHON_USE_PYTYPE_LOOKUP 1
+  #endif
+  #if PY_MAJOR_VERSION < 3
+    #undef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 0
+  #elif !defined(CYTHON_USE_ASYNC_SLOTS)
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #if PY_VERSION_HEX < 0x02070000
+    #undef CYTHON_USE_PYLONG_INTERNALS
+    #define CYTHON_USE_PYLONG_INTERNALS 0
+  #elif !defined(CYTHON_USE_PYLONG_INTERNALS)
+    #define CYTHON_USE_PYLONG_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_PYLIST_INTERNALS
+    #define CYTHON_USE_PYLIST_INTERNALS 1
+  #endif
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #if PY_VERSION_HEX < 0x030300F0 || PY_VERSION_HEX >= 0x030B00A2
+    #undef CYTHON_USE_UNICODE_WRITER
+    #define CYTHON_USE_UNICODE_WRITER 0
+  #elif !defined(CYTHON_USE_UNICODE_WRITER)
+    #define CYTHON_USE_UNICODE_WRITER 1
+  #endif
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #if PY_VERSION_HEX >= 0x030B00A4
+    #undef CYTHON_FAST_THREAD_STATE
+    #define CYTHON_FAST_THREAD_STATE 0
+  #elif !defined(CYTHON_FAST_THREAD_STATE)
+    #define CYTHON_FAST_THREAD_STATE 1
+  #endif
+  #ifndef CYTHON_FAST_PYCALL
+    #define CYTHON_FAST_PYCALL (PY_VERSION_HEX < 0x030A0000)
+  #endif
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT (PY_VERSION_HEX >= 0x03050000)
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1)
+  #endif
+  #ifndef CYTHON_USE_DICT_VERSIONS
+    #define CYTHON_USE_DICT_VERSIONS (PY_VERSION_HEX >= 0x030600B1)
+  #endif
+  #if PY_VERSION_HEX >= 0x030B00A4
+    #undef CYTHON_USE_EXC_INFO_STACK
+    #define CYTHON_USE_EXC_INFO_STACK 0
+  #elif !defined(CYTHON_USE_EXC_INFO_STACK)
+    #define CYTHON_USE_EXC_INFO_STACK (PY_VERSION_HEX >= 0x030700A3)
+  #endif
+  #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC
+    #define CYTHON_UPDATE_DESCRIPTOR_DOC 1
+  #endif
+#endif
+#if !defined(CYTHON_FAST_PYCCALL)
+#define CYTHON_FAST_PYCCALL  (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1)
+#endif
+#if CYTHON_USE_PYLONG_INTERNALS
+  #if PY_MAJOR_VERSION < 3
+    #include "longintrepr.h"
+  #endif
+  #undef SHIFT
+  #undef BASE
+  #undef MASK
+  #ifdef SIZEOF_VOID_P
+    enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) };
+  #endif
+#endif
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+#ifndef __has_cpp_attribute
+  #define __has_cpp_attribute(x) 0
+#endif
+#ifndef CYTHON_RESTRICT
+  #if defined(__GNUC__)
+    #define CYTHON_RESTRICT __restrict__
+  #elif defined(_MSC_VER) && _MSC_VER >= 1400
+    #define CYTHON_RESTRICT __restrict
+  #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define CYTHON_RESTRICT restrict
+  #else
+    #define CYTHON_RESTRICT
+  #endif
+#endif
+#ifndef CYTHON_UNUSED
+# if defined(__GNUC__)
+#   if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4))
+#     define CYTHON_UNUSED __attribute__ ((__unused__))
+#   else
+#     define CYTHON_UNUSED
+#   endif
+# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER))
+#   define CYTHON_UNUSED __attribute__ ((__unused__))
+# else
+#   define CYTHON_UNUSED
+# endif
+#endif
+#ifndef CYTHON_MAYBE_UNUSED_VAR
+#  if defined(__cplusplus)
+     template<class T> void CYTHON_MAYBE_UNUSED_VAR( const T& ) { }
+#  else
+#    define CYTHON_MAYBE_UNUSED_VAR(x) (void)(x)
+#  endif
+#endif
+#ifndef CYTHON_NCP_UNUSED
+# if CYTHON_COMPILING_IN_CPYTHON
+#  define CYTHON_NCP_UNUSED
+# else
+#  define CYTHON_NCP_UNUSED CYTHON_UNUSED
+# endif
+#endif
+#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None)
+#ifdef _MSC_VER
+    #ifndef _MSC_STDINT_H_
+        #if _MSC_VER < 1300
+           typedef unsigned char     uint8_t;
+           typedef unsigned int      uint32_t;
+        #else
+           typedef unsigned __int8   uint8_t;
+           typedef unsigned __int32  uint32_t;
+        #endif
+    #endif
+#else
+   #include <stdint.h>
+#endif
+#ifndef CYTHON_FALLTHROUGH
+  #if defined(__cplusplus) && __cplusplus >= 201103L
+    #if __has_cpp_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH [[fallthrough]]
+    #elif __has_cpp_attribute(clang::fallthrough)
+      #define CYTHON_FALLTHROUGH [[clang::fallthrough]]
+    #elif __has_cpp_attribute(gnu::fallthrough)
+      #define CYTHON_FALLTHROUGH [[gnu::fallthrough]]
+    #endif
+  #endif
+  #ifndef CYTHON_FALLTHROUGH
+    #if __has_attribute(fallthrough)
+      #define CYTHON_FALLTHROUGH __attribute__((fallthrough))
+    #else
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+  #if defined(__clang__ ) && defined(__apple_build_version__)
+    #if __apple_build_version__ < 7000000
+      #undef  CYTHON_FALLTHROUGH
+      #define CYTHON_FALLTHROUGH
+    #endif
+  #endif
+#endif
+
+#ifndef __cplusplus
+  #error "Cython files generated with the C++ option must be compiled with a C++ compiler."
+#endif
+#ifndef CYTHON_INLINE
+  #if defined(__clang__)
+    #define CYTHON_INLINE __inline__ __attribute__ ((__unused__))
+  #else
+    #define CYTHON_INLINE inline
+  #endif
+#endif
+template<typename T>
+void __Pyx_call_destructor(T& x) {
+    x.~T();
+}
+template<typename T>
+class __Pyx_FakeReference {
+  public:
+    __Pyx_FakeReference() : ptr(NULL) { }
+    __Pyx_FakeReference(const T& ref) : ptr(const_cast<T*>(&ref)) { }
+    T *operator->() { return ptr; }
+    T *operator&() { return ptr; }
+    operator T&() { return *ptr; }
+    template<typename U> bool operator ==(U other) { return *ptr == other; }
+    template<typename U> bool operator !=(U other) { return *ptr != other; }
+  private:
+    T *ptr;
+};
+
+#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag)
+  #define Py_OptimizeFlag 0
+#endif
+#define __PYX_BUILD_PY_SSIZE_T "n"
+#define CYTHON_FORMAT_SSIZE_T "z"
+#if PY_MAJOR_VERSION < 3
+  #define __Pyx_BUILTIN_MODULE_NAME "__builtin__"
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+  #define __Pyx_DefaultClassType PyClass_Type
+#else
+  #define __Pyx_BUILTIN_MODULE_NAME "builtins"
+  #define __Pyx_DefaultClassType PyType_Type
+#if PY_VERSION_HEX >= 0x030B00A1
+    static CYTHON_INLINE PyCodeObject* __Pyx_PyCode_New(int a, int k, int l, int s, int f,
+                                                    PyObject *code, PyObject *c, PyObject* n, PyObject *v,
+                                                    PyObject *fv, PyObject *cell, PyObject* fn,
+                                                    PyObject *name, int fline, PyObject *lnos) {
+        PyObject *kwds=NULL, *argcount=NULL, *posonlyargcount=NULL, *kwonlyargcount=NULL;
+        PyObject *nlocals=NULL, *stacksize=NULL, *flags=NULL, *replace=NULL, *call_result=NULL, *empty=NULL;
+        const char *fn_cstr=NULL;
+        const char *name_cstr=NULL;
+        PyCodeObject* co=NULL;
+        PyObject *type, *value, *traceback;
+        PyErr_Fetch(&type, &value, &traceback);
+        if (!(kwds=PyDict_New())) goto end;
+        if (!(argcount=PyLong_FromLong(a))) goto end;
+        if (PyDict_SetItemString(kwds, "co_argcount", argcount) != 0) goto end;
+        if (!(posonlyargcount=PyLong_FromLong(0))) goto end;
+        if (PyDict_SetItemString(kwds, "co_posonlyargcount", posonlyargcount) != 0) goto end;
+        if (!(kwonlyargcount=PyLong_FromLong(k))) goto end;
+        if (PyDict_SetItemString(kwds, "co_kwonlyargcount", kwonlyargcount) != 0) goto end;
+        if (!(nlocals=PyLong_FromLong(l))) goto end;
+        if (PyDict_SetItemString(kwds, "co_nlocals", nlocals) != 0) goto end;
+        if (!(stacksize=PyLong_FromLong(s))) goto end;
+        if (PyDict_SetItemString(kwds, "co_stacksize", stacksize) != 0) goto end;
+        if (!(flags=PyLong_FromLong(f))) goto end;
+        if (PyDict_SetItemString(kwds, "co_flags", flags) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_code", code) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_consts", c) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_names", n) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_varnames", v) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_freevars", fv) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_cellvars", cell) != 0) goto end;
+        if (PyDict_SetItemString(kwds, "co_linetable", lnos) != 0) goto end;
+        if (!(fn_cstr=PyUnicode_AsUTF8AndSize(fn, NULL))) goto end;
+        if (!(name_cstr=PyUnicode_AsUTF8AndSize(name, NULL))) goto end;
+        if (!(co = PyCode_NewEmpty(fn_cstr, name_cstr, fline))) goto end;
+        if (!(replace = PyObject_GetAttrString((PyObject*)co, "replace"))) goto cleanup_code_too;
+        if (!(empty = PyTuple_New(0))) goto cleanup_code_too; // unfortunately __pyx_empty_tuple isn't available here
+        if (!(call_result = PyObject_Call(replace, empty, kwds))) goto cleanup_code_too;
+        Py_XDECREF((PyObject*)co);
+        co = (PyCodeObject*)call_result;
+        call_result = NULL;
+        if (0) {
+            cleanup_code_too:
+            Py_XDECREF((PyObject*)co);
+            co = NULL;
+        }
+        end:
+        Py_XDECREF(kwds);
+        Py_XDECREF(argcount);
+        Py_XDECREF(posonlyargcount);
+        Py_XDECREF(kwonlyargcount);
+        Py_XDECREF(nlocals);
+        Py_XDECREF(stacksize);
+        Py_XDECREF(replace);
+        Py_XDECREF(call_result);
+        Py_XDECREF(empty);
+        if (type) {
+            PyErr_Restore(type, value, traceback);
+        }
+        return co;
+    }
+#else
+  #define __Pyx_PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\
+          PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)
+#endif
+  #define __Pyx_DefaultClassType PyType_Type
+#endif
+#ifndef Py_TPFLAGS_CHECKTYPES
+  #define Py_TPFLAGS_CHECKTYPES 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_INDEX
+  #define Py_TPFLAGS_HAVE_INDEX 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_NEWBUFFER
+  #define Py_TPFLAGS_HAVE_NEWBUFFER 0
+#endif
+#ifndef Py_TPFLAGS_HAVE_FINALIZE
+  #define Py_TPFLAGS_HAVE_FINALIZE 0
+#endif
+#ifndef METH_STACKLESS
+  #define METH_STACKLESS 0
+#endif
+#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL)
+  #ifndef METH_FASTCALL
+     #define METH_FASTCALL 0x80
+  #endif
+  typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs);
+  typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args,
+                                                          Py_ssize_t nargs, PyObject *kwnames);
+#else
+  #define __Pyx_PyCFunctionFast _PyCFunctionFast
+  #define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords
+#endif
+#if CYTHON_FAST_PYCCALL
+#define __Pyx_PyFastCFunction_Check(func)\
+    ((PyCFunction_Check(func) && (METH_FASTCALL == (PyCFunction_GET_FLAGS(func) & ~(METH_CLASS | METH_STATIC | METH_COEXIST | METH_KEYWORDS | METH_STACKLESS)))))
+#else
+#define __Pyx_PyFastCFunction_Check(func) 0
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc)
+  #define PyObject_Malloc(s)   PyMem_Malloc(s)
+  #define PyObject_Free(p)     PyMem_Free(p)
+  #define PyObject_Realloc(p)  PyMem_Realloc(p)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030400A1
+  #define PyMem_RawMalloc(n)           PyMem_Malloc(n)
+  #define PyMem_RawRealloc(p, n)       PyMem_Realloc(p, n)
+  #define PyMem_RawFree(p)             PyMem_Free(p)
+#endif
+#if CYTHON_COMPILING_IN_PYSTON
+  #define __Pyx_PyCode_HasFreeVars(co)  PyCode_HasFreeVars(co)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno) PyFrame_SetLineNumber(frame, lineno)
+#else
+  #define __Pyx_PyCode_HasFreeVars(co)  (PyCode_GetNumFree(co) > 0)
+  #define __Pyx_PyFrame_SetLineNumber(frame, lineno)  (frame)->f_lineno = (lineno)
+#endif
+#if !CYTHON_FAST_THREAD_STATE || PY_VERSION_HEX < 0x02070000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#elif PY_VERSION_HEX >= 0x03060000
+  #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet()
+#elif PY_VERSION_HEX >= 0x03000000
+  #define __Pyx_PyThreadState_Current PyThreadState_GET()
+#else
+  #define __Pyx_PyThreadState_Current _PyThreadState_Current
+#endif
+#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT)
+#include "pythread.h"
+#define Py_tss_NEEDS_INIT 0
+typedef int Py_tss_t;
+static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) {
+  *key = PyThread_create_key();
+  return 0;
+}
+static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) {
+  Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t));
+  *key = Py_tss_NEEDS_INIT;
+  return key;
+}
+static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) {
+  PyObject_Free(key);
+}
+static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) {
+  return *key != Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) {
+  PyThread_delete_key(*key);
+  *key = Py_tss_NEEDS_INIT;
+}
+static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) {
+  return PyThread_set_key_value(*key, value);
+}
+static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) {
+  return PyThread_get_key_value(*key);
+}
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON || defined(_PyDict_NewPresized)
+#define __Pyx_PyDict_NewPresized(n)  ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n))
+#else
+#define __Pyx_PyDict_NewPresized(n)  PyDict_New()
+#endif
+#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_TrueDivide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceTrueDivide(x,y)
+#else
+  #define __Pyx_PyNumber_Divide(x,y)         PyNumber_Divide(x,y)
+  #define __Pyx_PyNumber_InPlaceDivide(x,y)  PyNumber_InPlaceDivide(x,y)
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && CYTHON_USE_UNICODE_INTERNALS
+#define __Pyx_PyDict_GetItemStr(dict, name)  _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash)
+#else
+#define __Pyx_PyDict_GetItemStr(dict, name)  PyDict_GetItem(dict, name)
+#endif
+#if PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND)
+  #define CYTHON_PEP393_ENABLED 1
+  #if defined(PyUnicode_IS_READY)
+  #define __Pyx_PyUnicode_READY(op)       (likely(PyUnicode_IS_READY(op)) ?\
+                                              0 : _PyUnicode_Ready((PyObject *)(op)))
+  #else
+  #define __Pyx_PyUnicode_READY(op)       (0)
+  #endif
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_LENGTH(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i)
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   PyUnicode_MAX_CHAR_VALUE(u)
+  #define __Pyx_PyUnicode_KIND(u)         PyUnicode_KIND(u)
+  #define __Pyx_PyUnicode_DATA(u)         PyUnicode_DATA(u)
+  #define __Pyx_PyUnicode_READ(k, d, i)   PyUnicode_READ(k, d, i)
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  PyUnicode_WRITE(k, d, i, ch)
+  #if defined(PyUnicode_IS_READY) && defined(PyUnicode_GET_SIZE)
+  #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x03090000
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : ((PyCompactUnicodeObject *)(u))->wstr_length))
+  #else
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u)))
+  #endif
+  #else
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_LENGTH(u))
+  #endif
+#else
+  #define CYTHON_PEP393_ENABLED 0
+  #define PyUnicode_1BYTE_KIND  1
+  #define PyUnicode_2BYTE_KIND  2
+  #define PyUnicode_4BYTE_KIND  4
+  #define __Pyx_PyUnicode_READY(op)       (0)
+  #define __Pyx_PyUnicode_GET_LENGTH(u)   PyUnicode_GET_SIZE(u)
+  #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i]))
+  #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u)   ((sizeof(Py_UNICODE) == 2) ? 65535 : 1114111)
+  #define __Pyx_PyUnicode_KIND(u)         (sizeof(Py_UNICODE))
+  #define __Pyx_PyUnicode_DATA(u)         ((void*)PyUnicode_AS_UNICODE(u))
+  #define __Pyx_PyUnicode_READ(k, d, i)   ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i]))
+  #define __Pyx_PyUnicode_WRITE(k, d, i, ch)  (((void)(k)), ((Py_UNICODE*)d)[i] = ch)
+  #define __Pyx_PyUnicode_IS_TRUE(u)      (0 != PyUnicode_GET_SIZE(u))
+#endif
+#if CYTHON_COMPILING_IN_PYPY
+  #define __Pyx_PyUnicode_Concat(a, b)      PyNumber_Add(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  PyNumber_Add(a, b)
+#else
+  #define __Pyx_PyUnicode_Concat(a, b)      PyUnicode_Concat(a, b)
+  #define __Pyx_PyUnicode_ConcatSafe(a, b)  ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\
+      PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b))
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyUnicode_Contains)
+  #define PyUnicode_Contains(u, s)  PySequence_Contains(u, s)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyByteArray_Check)
+  #define PyByteArray_Check(obj)  PyObject_TypeCheck(obj, &PyByteArray_Type)
+#endif
+#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Format)
+  #define PyObject_Format(obj, fmt)  PyObject_CallMethod(obj, "__format__", "O", fmt)
+#endif
+#define __Pyx_PyString_FormatSafe(a, b)   ((unlikely((a) == Py_None || (PyString_Check(b) && !PyString_CheckExact(b)))) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b))
+#define __Pyx_PyUnicode_FormatSafe(a, b)  ((unlikely((a) == Py_None || (PyUnicode_Check(b) && !PyUnicode_CheckExact(b)))) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b))
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyString_Format(a, b)  PyUnicode_Format(a, b)
+#else
+  #define __Pyx_PyString_Format(a, b)  PyString_Format(a, b)
+#endif
+#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII)
+  #define PyObject_ASCII(o)            PyObject_Repr(o)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBaseString_Type            PyUnicode_Type
+  #define PyStringObject               PyUnicodeObject
+  #define PyString_Type                PyUnicode_Type
+  #define PyString_Check               PyUnicode_Check
+  #define PyString_CheckExact          PyUnicode_CheckExact
+#ifndef PyObject_Unicode
+  #define PyObject_Unicode             PyObject_Str
+#endif
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj)
+  #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj)
+#else
+  #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj))
+  #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj))
+#endif
+#ifndef PySet_CheckExact
+  #define PySet_CheckExact(obj)        (Py_TYPE(obj) == &PySet_Type)
+#endif
+#if PY_VERSION_HEX >= 0x030900A4
+  #define __Pyx_SET_REFCNT(obj, refcnt) Py_SET_REFCNT(obj, refcnt)
+  #define __Pyx_SET_SIZE(obj, size) Py_SET_SIZE(obj, size)
+#else
+  #define __Pyx_SET_REFCNT(obj, refcnt) Py_REFCNT(obj) = (refcnt)
+  #define __Pyx_SET_SIZE(obj, size) Py_SIZE(obj) = (size)
+#endif
+#if CYTHON_ASSUME_SAFE_MACROS
+  #define __Pyx_PySequence_SIZE(seq)  Py_SIZE(seq)
+#else
+  #define __Pyx_PySequence_SIZE(seq)  PySequence_Size(seq)
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyIntObject                  PyLongObject
+  #define PyInt_Type                   PyLong_Type
+  #define PyInt_Check(op)              PyLong_Check(op)
+  #define PyInt_CheckExact(op)         PyLong_CheckExact(op)
+  #define PyInt_FromString             PyLong_FromString
+  #define PyInt_FromUnicode            PyLong_FromUnicode
+  #define PyInt_FromLong               PyLong_FromLong
+  #define PyInt_FromSize_t             PyLong_FromSize_t
+  #define PyInt_FromSsize_t            PyLong_FromSsize_t
+  #define PyInt_AsLong                 PyLong_AsLong
+  #define PyInt_AS_LONG                PyLong_AS_LONG
+  #define PyInt_AsSsize_t              PyLong_AsSsize_t
+  #define PyInt_AsUnsignedLongMask     PyLong_AsUnsignedLongMask
+  #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask
+  #define PyNumber_Int                 PyNumber_Long
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define PyBoolObject                 PyLongObject
+#endif
+#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY
+  #ifndef PyUnicode_InternFromString
+    #define PyUnicode_InternFromString(s) PyUnicode_FromString(s)
+  #endif
+#endif
+#if PY_VERSION_HEX < 0x030200A4
+  typedef long Py_hash_t;
+  #define __Pyx_PyInt_FromHash_t PyInt_FromLong
+  #define __Pyx_PyInt_AsHash_t   __Pyx_PyIndex_AsHash_t
+#else
+  #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t
+  #define __Pyx_PyInt_AsHash_t   __Pyx_PyIndex_AsSsize_t
+#endif
+#if PY_MAJOR_VERSION >= 3
+  #define __Pyx_PyMethod_New(func, self, klass) ((self) ? ((void)(klass), PyMethod_New(func, self)) : __Pyx_NewRef(func))
+#else
+  #define __Pyx_PyMethod_New(func, self, klass) PyMethod_New(func, self, klass)
+#endif
+#if CYTHON_USE_ASYNC_SLOTS
+  #if PY_VERSION_HEX >= 0x030500B1
+    #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods
+    #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async)
+  #else
+    #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved))
+  #endif
+#else
+  #define __Pyx_PyType_AsAsync(obj) NULL
+#endif
+#ifndef __Pyx_PyAsyncMethodsStruct
+    typedef struct {
+        unaryfunc am_await;
+        unaryfunc am_aiter;
+        unaryfunc am_anext;
+    } __Pyx_PyAsyncMethodsStruct;
+#endif
+
+#if defined(_WIN32) || defined(WIN32) || defined(MS_WINDOWS)
+  #if !defined(_USE_MATH_DEFINES)
+    #define _USE_MATH_DEFINES
+  #endif
+#endif
+#include <math.h>
+#ifdef NAN
+#define __PYX_NAN() ((float) NAN)
+#else
+static CYTHON_INLINE float __PYX_NAN() {
+  float value;
+  memset(&value, 0xFF, sizeof(value));
+  return value;
+}
+#endif
+#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL)
+#define __Pyx_truncl trunc
+#else
+#define __Pyx_truncl truncl
+#endif
+
+#define __PYX_MARK_ERR_POS(f_index, lineno) \
+    { __pyx_filename = __pyx_f[f_index]; (void)__pyx_filename; __pyx_lineno = lineno; (void)__pyx_lineno; __pyx_clineno = __LINE__; (void)__pyx_clineno; }
+#define __PYX_ERR(f_index, lineno, Ln_error) \
+    { __PYX_MARK_ERR_POS(f_index, lineno) goto Ln_error; }
+
+#ifndef __PYX_EXTERN_C
+  #ifdef __cplusplus
+    #define __PYX_EXTERN_C extern "C"
+  #else
+    #define __PYX_EXTERN_C extern
+  #endif
+#endif
+
+#define __PYX_HAVE__insightface__thirdparty__face3d__mesh__cython__mesh_core_cython
+#define __PYX_HAVE_API__insightface__thirdparty__face3d__mesh__cython__mesh_core_cython
+/* Early includes */
+#include <string.h>
+#include <stdio.h>
+#include "numpy/arrayobject.h"
+#include "numpy/ndarrayobject.h"
+#include "numpy/ndarraytypes.h"
+#include "numpy/arrayscalars.h"
+#include "numpy/ufuncobject.h"
+
+    /* NumPy API declarations from "numpy/__init__.pxd" */
+    
+#include <string>
+#include "ios"
+#include "new"
+#include "stdexcept"
+#include "typeinfo"
+#include "mesh_core.h"
+#ifdef _OPENMP
+#include <omp.h>
+#endif /* _OPENMP */
+
+#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS)
+#define CYTHON_WITHOUT_ASSERTIONS
+#endif
+
+typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding;
+                const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry;
+
+#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_UTF8 0
+#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT (PY_MAJOR_VERSION >= 3 && __PYX_DEFAULT_STRING_ENCODING_IS_UTF8)
+#define __PYX_DEFAULT_STRING_ENCODING ""
+#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString
+#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#define __Pyx_uchar_cast(c) ((unsigned char)c)
+#define __Pyx_long_cast(x) ((long)x)
+#define __Pyx_fits_Py_ssize_t(v, type, is_signed)  (\
+    (sizeof(type) < sizeof(Py_ssize_t))  ||\
+    (sizeof(type) > sizeof(Py_ssize_t) &&\
+          likely(v < (type)PY_SSIZE_T_MAX ||\
+                 v == (type)PY_SSIZE_T_MAX)  &&\
+          (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\
+                                v == (type)PY_SSIZE_T_MIN)))  ||\
+    (sizeof(type) == sizeof(Py_ssize_t) &&\
+          (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\
+                               v == (type)PY_SSIZE_T_MAX)))  )
+static CYTHON_INLINE int __Pyx_is_valid_index(Py_ssize_t i, Py_ssize_t limit) {
+    return (size_t) i < (size_t) limit;
+}
+#if defined (__cplusplus) && __cplusplus >= 201103L
+    #include <cstdlib>
+    #define __Pyx_sst_abs(value) std::abs(value)
+#elif SIZEOF_INT >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) abs(value)
+#elif SIZEOF_LONG >= SIZEOF_SIZE_T
+    #define __Pyx_sst_abs(value) labs(value)
+#elif defined (_MSC_VER)
+    #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value))
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+    #define __Pyx_sst_abs(value) llabs(value)
+#elif defined (__GNUC__)
+    #define __Pyx_sst_abs(value) __builtin_llabs(value)
+#else
+    #define __Pyx_sst_abs(value) ((value<0) ? -value : value)
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*);
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length);
+#define __Pyx_PyByteArray_FromString(s) PyByteArray_FromStringAndSize((const char*)s, strlen((const char*)s))
+#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l)
+#define __Pyx_PyBytes_FromString        PyBytes_FromString
+#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*);
+#if PY_MAJOR_VERSION < 3
+    #define __Pyx_PyStr_FromString        __Pyx_PyBytes_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize
+#else
+    #define __Pyx_PyStr_FromString        __Pyx_PyUnicode_FromString
+    #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize
+#endif
+#define __Pyx_PyBytes_AsWritableString(s)     ((char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableSString(s)    ((signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsWritableUString(s)    ((unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsString(s)     ((const char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsSString(s)    ((const signed char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyBytes_AsUString(s)    ((const unsigned char*) PyBytes_AS_STRING(s))
+#define __Pyx_PyObject_AsWritableString(s)    ((char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableSString(s)    ((signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsWritableUString(s)    ((unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsSString(s)    ((const signed char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_AsUString(s)    ((const unsigned char*) __Pyx_PyObject_AsString(s))
+#define __Pyx_PyObject_FromCString(s)  __Pyx_PyObject_FromString((const char*)s)
+#define __Pyx_PyBytes_FromCString(s)   __Pyx_PyBytes_FromString((const char*)s)
+#define __Pyx_PyByteArray_FromCString(s)   __Pyx_PyByteArray_FromString((const char*)s)
+#define __Pyx_PyStr_FromCString(s)     __Pyx_PyStr_FromString((const char*)s)
+#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s)
+static CYTHON_INLINE size_t __Pyx_Py_UNICODE_strlen(const Py_UNICODE *u) {
+    const Py_UNICODE *u_end = u;
+    while (*u_end++) ;
+    return (size_t)(u_end - u - 1);
+}
+#define __Pyx_PyUnicode_FromUnicode(u)       PyUnicode_FromUnicode(u, __Pyx_Py_UNICODE_strlen(u))
+#define __Pyx_PyUnicode_FromUnicodeAndLength PyUnicode_FromUnicode
+#define __Pyx_PyUnicode_AsUnicode            PyUnicode_AsUnicode
+#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj)
+#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None)
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*);
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject*);
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x);
+#define __Pyx_PySequence_Tuple(obj)\
+    (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj))
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*);
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t);
+static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject*);
+#if CYTHON_ASSUME_SAFE_MACROS
+#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x))
+#else
+#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x)
+#endif
+#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x))
+#if PY_MAJOR_VERSION >= 3
+#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x))
+#else
+#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x))
+#endif
+#define __Pyx_PyNumber_Float(x) (PyFloat_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Float(x))
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+static int __Pyx_sys_getdefaultencoding_not_ascii;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    PyObject* ascii_chars_u = NULL;
+    PyObject* ascii_chars_b = NULL;
+    const char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    if (strcmp(default_encoding_c, "ascii") == 0) {
+        __Pyx_sys_getdefaultencoding_not_ascii = 0;
+    } else {
+        char ascii_chars[128];
+        int c;
+        for (c = 0; c < 128; c++) {
+            ascii_chars[c] = c;
+        }
+        __Pyx_sys_getdefaultencoding_not_ascii = 1;
+        ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL);
+        if (!ascii_chars_u) goto bad;
+        ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL);
+        if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) {
+            PyErr_Format(
+                PyExc_ValueError,
+                "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.",
+                default_encoding_c);
+            goto bad;
+        }
+        Py_DECREF(ascii_chars_u);
+        Py_DECREF(ascii_chars_b);
+    }
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    Py_XDECREF(ascii_chars_u);
+    Py_XDECREF(ascii_chars_b);
+    return -1;
+}
+#endif
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL)
+#else
+#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL)
+#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+static char* __PYX_DEFAULT_STRING_ENCODING;
+static int __Pyx_init_sys_getdefaultencoding_params(void) {
+    PyObject* sys;
+    PyObject* default_encoding = NULL;
+    char* default_encoding_c;
+    sys = PyImport_ImportModule("sys");
+    if (!sys) goto bad;
+    default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL);
+    Py_DECREF(sys);
+    if (!default_encoding) goto bad;
+    default_encoding_c = PyBytes_AsString(default_encoding);
+    if (!default_encoding_c) goto bad;
+    __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c) + 1);
+    if (!__PYX_DEFAULT_STRING_ENCODING) goto bad;
+    strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c);
+    Py_DECREF(default_encoding);
+    return 0;
+bad:
+    Py_XDECREF(default_encoding);
+    return -1;
+}
+#endif
+#endif
+
+
+/* Test for GCC > 2.95 */
+#if defined(__GNUC__)     && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95)))
+  #define likely(x)   __builtin_expect(!!(x), 1)
+  #define unlikely(x) __builtin_expect(!!(x), 0)
+#else /* !__GNUC__ or GCC < 2.95 */
+  #define likely(x)   (x)
+  #define unlikely(x) (x)
+#endif /* __GNUC__ */
+static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; }
+
+static PyObject *__pyx_m = NULL;
+static PyObject *__pyx_d;
+static PyObject *__pyx_b;
+static PyObject *__pyx_cython_runtime = NULL;
+static PyObject *__pyx_empty_tuple;
+static PyObject *__pyx_empty_bytes;
+static PyObject *__pyx_empty_unicode;
+static int __pyx_lineno;
+static int __pyx_clineno = 0;
+static const char * __pyx_cfilenm= __FILE__;
+static const char *__pyx_filename;
+
+/* Header.proto */
+#if !defined(CYTHON_CCOMPLEX)
+  #if defined(__cplusplus)
+    #define CYTHON_CCOMPLEX 1
+  #elif defined(_Complex_I)
+    #define CYTHON_CCOMPLEX 1
+  #else
+    #define CYTHON_CCOMPLEX 0
+  #endif
+#endif
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #include <complex>
+  #else
+    #include <complex.h>
+  #endif
+#endif
+#if CYTHON_CCOMPLEX && !defined(__cplusplus) && defined(__sun__) && defined(__GNUC__)
+  #undef _Complex_I
+  #define _Complex_I 1.0fj
+#endif
+
+
+static const char *__pyx_f[] = {
+  "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx",
+  "__init__.pxd",
+  "stringsource",
+  "type.pxd",
+};
+/* BufferFormatStructs.proto */
+#define IS_UNSIGNED(type) (((type) -1) > 0)
+struct __Pyx_StructField_;
+#define __PYX_BUF_FLAGS_PACKED_STRUCT (1 << 0)
+typedef struct {
+  const char* name;
+  struct __Pyx_StructField_* fields;
+  size_t size;
+  size_t arraysize[8];
+  int ndim;
+  char typegroup;
+  char is_unsigned;
+  int flags;
+} __Pyx_TypeInfo;
+typedef struct __Pyx_StructField_ {
+  __Pyx_TypeInfo* type;
+  const char* name;
+  size_t offset;
+} __Pyx_StructField;
+typedef struct {
+  __Pyx_StructField* field;
+  size_t parent_offset;
+} __Pyx_BufFmt_StackElem;
+typedef struct {
+  __Pyx_StructField root;
+  __Pyx_BufFmt_StackElem* head;
+  size_t fmt_offset;
+  size_t new_count, enc_count;
+  size_t struct_alignment;
+  int is_complex;
+  char enc_type;
+  char new_packmode;
+  char enc_packmode;
+  char is_valid_array;
+} __Pyx_BufFmt_Context;
+
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":689
+ * # in Cython to enable them only on the right systems.
+ * 
+ * ctypedef npy_int8       int8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ */
+typedef npy_int8 __pyx_t_5numpy_int8_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":690
+ * 
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t
+ */
+typedef npy_int16 __pyx_t_5numpy_int16_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":691
+ * ctypedef npy_int8       int8_t
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_int64      int64_t
+ * #ctypedef npy_int96      int96_t
+ */
+typedef npy_int32 __pyx_t_5numpy_int32_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":692
+ * ctypedef npy_int16      int16_t
+ * ctypedef npy_int32      int32_t
+ * ctypedef npy_int64      int64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_int96      int96_t
+ * #ctypedef npy_int128     int128_t
+ */
+typedef npy_int64 __pyx_t_5numpy_int64_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":696
+ * #ctypedef npy_int128     int128_t
+ * 
+ * ctypedef npy_uint8      uint8_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ */
+typedef npy_uint8 __pyx_t_5numpy_uint8_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":697
+ * 
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t
+ */
+typedef npy_uint16 __pyx_t_5numpy_uint16_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":698
+ * ctypedef npy_uint8      uint8_t
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uint64     uint64_t
+ * #ctypedef npy_uint96     uint96_t
+ */
+typedef npy_uint32 __pyx_t_5numpy_uint32_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":699
+ * ctypedef npy_uint16     uint16_t
+ * ctypedef npy_uint32     uint32_t
+ * ctypedef npy_uint64     uint64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_uint96     uint96_t
+ * #ctypedef npy_uint128    uint128_t
+ */
+typedef npy_uint64 __pyx_t_5numpy_uint64_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":703
+ * #ctypedef npy_uint128    uint128_t
+ * 
+ * ctypedef npy_float32    float32_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_float64    float64_t
+ * #ctypedef npy_float80    float80_t
+ */
+typedef npy_float32 __pyx_t_5numpy_float32_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":704
+ * 
+ * ctypedef npy_float32    float32_t
+ * ctypedef npy_float64    float64_t             # <<<<<<<<<<<<<<
+ * #ctypedef npy_float80    float80_t
+ * #ctypedef npy_float128   float128_t
+ */
+typedef npy_float64 __pyx_t_5numpy_float64_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":713
+ * # The int types are mapped a bit surprising --
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t
+ */
+typedef npy_long __pyx_t_5numpy_int_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":714
+ * # numpy.int corresponds to 'l' and numpy.long to 'q'
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longlong   longlong_t
+ * 
+ */
+typedef npy_longlong __pyx_t_5numpy_long_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":715
+ * ctypedef npy_long       int_t
+ * ctypedef npy_longlong   long_t
+ * ctypedef npy_longlong   longlong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_ulong      uint_t
+ */
+typedef npy_longlong __pyx_t_5numpy_longlong_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":717
+ * ctypedef npy_longlong   longlong_t
+ * 
+ * ctypedef npy_ulong      uint_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t
+ */
+typedef npy_ulong __pyx_t_5numpy_uint_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":718
+ * 
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulong_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":719
+ * ctypedef npy_ulong      uint_t
+ * ctypedef npy_ulonglong  ulong_t
+ * ctypedef npy_ulonglong  ulonglong_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_intp       intp_t
+ */
+typedef npy_ulonglong __pyx_t_5numpy_ulonglong_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":721
+ * ctypedef npy_ulonglong  ulonglong_t
+ * 
+ * ctypedef npy_intp       intp_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_uintp      uintp_t
+ * 
+ */
+typedef npy_intp __pyx_t_5numpy_intp_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":722
+ * 
+ * ctypedef npy_intp       intp_t
+ * ctypedef npy_uintp      uintp_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_double     float_t
+ */
+typedef npy_uintp __pyx_t_5numpy_uintp_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":724
+ * ctypedef npy_uintp      uintp_t
+ * 
+ * ctypedef npy_double     float_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t
+ */
+typedef npy_double __pyx_t_5numpy_float_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":725
+ * 
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ */
+typedef npy_double __pyx_t_5numpy_double_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":726
+ * ctypedef npy_double     float_t
+ * ctypedef npy_double     double_t
+ * ctypedef npy_longdouble longdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ */
+typedef npy_longdouble __pyx_t_5numpy_longdouble_t;
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< float > __pyx_t_float_complex;
+  #else
+    typedef float _Complex __pyx_t_float_complex;
+  #endif
+#else
+    typedef struct { float real, imag; } __pyx_t_float_complex;
+#endif
+static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float, float);
+
+/* Declarations.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    typedef ::std::complex< double > __pyx_t_double_complex;
+  #else
+    typedef double _Complex __pyx_t_double_complex;
+  #endif
+#else
+    typedef struct { double real, imag; } __pyx_t_double_complex;
+#endif
+static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double, double);
+
+
+/*--- Type declarations ---*/
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":728
+ * ctypedef npy_longdouble longdouble_t
+ * 
+ * ctypedef npy_cfloat      cfloat_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t
+ */
+typedef npy_cfloat __pyx_t_5numpy_cfloat_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":729
+ * 
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t             # <<<<<<<<<<<<<<
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ */
+typedef npy_cdouble __pyx_t_5numpy_cdouble_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":730
+ * ctypedef npy_cfloat      cfloat_t
+ * ctypedef npy_cdouble     cdouble_t
+ * ctypedef npy_clongdouble clongdouble_t             # <<<<<<<<<<<<<<
+ * 
+ * ctypedef npy_cdouble     complex_t
+ */
+typedef npy_clongdouble __pyx_t_5numpy_clongdouble_t;
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":732
+ * ctypedef npy_clongdouble clongdouble_t
+ * 
+ * ctypedef npy_cdouble     complex_t             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ */
+typedef npy_cdouble __pyx_t_5numpy_complex_t;
+
+/* --- Runtime support code (head) --- */
+/* Refnanny.proto */
+#ifndef CYTHON_REFNANNY
+  #define CYTHON_REFNANNY 0
+#endif
+#if CYTHON_REFNANNY
+  typedef struct {
+    void (*INCREF)(void*, PyObject*, int);
+    void (*DECREF)(void*, PyObject*, int);
+    void (*GOTREF)(void*, PyObject*, int);
+    void (*GIVEREF)(void*, PyObject*, int);
+    void* (*SetupContext)(const char*, int, const char*);
+    void (*FinishContext)(void**);
+  } __Pyx_RefNannyAPIStruct;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL;
+  static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname);
+  #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL;
+#ifdef WITH_THREAD
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          if (acquire_gil) {\
+              PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+              PyGILState_Release(__pyx_gilstate_save);\
+          } else {\
+              __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__);\
+          }
+#else
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)\
+          __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), __LINE__, __FILE__)
+#endif
+  #define __Pyx_RefNannyFinishContext()\
+          __Pyx_RefNanny->FinishContext(&__pyx_refnanny)
+  #define __Pyx_INCREF(r)  __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_DECREF(r)  __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GOTREF(r)  __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), __LINE__)
+  #define __Pyx_XINCREF(r)  do { if((r) != NULL) {__Pyx_INCREF(r); }} while(0)
+  #define __Pyx_XDECREF(r)  do { if((r) != NULL) {__Pyx_DECREF(r); }} while(0)
+  #define __Pyx_XGOTREF(r)  do { if((r) != NULL) {__Pyx_GOTREF(r); }} while(0)
+  #define __Pyx_XGIVEREF(r) do { if((r) != NULL) {__Pyx_GIVEREF(r);}} while(0)
+#else
+  #define __Pyx_RefNannyDeclarations
+  #define __Pyx_RefNannySetupContext(name, acquire_gil)
+  #define __Pyx_RefNannyFinishContext()
+  #define __Pyx_INCREF(r) Py_INCREF(r)
+  #define __Pyx_DECREF(r) Py_DECREF(r)
+  #define __Pyx_GOTREF(r)
+  #define __Pyx_GIVEREF(r)
+  #define __Pyx_XINCREF(r) Py_XINCREF(r)
+  #define __Pyx_XDECREF(r) Py_XDECREF(r)
+  #define __Pyx_XGOTREF(r)
+  #define __Pyx_XGIVEREF(r)
+#endif
+#define __Pyx_XDECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_XDECREF(tmp);\
+    } while (0)
+#define __Pyx_DECREF_SET(r, v) do {\
+        PyObject *tmp = (PyObject *) r;\
+        r = v; __Pyx_DECREF(tmp);\
+    } while (0)
+#define __Pyx_CLEAR(r)    do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0)
+#define __Pyx_XCLEAR(r)   do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0)
+
+/* RaiseArgTupleInvalid.proto */
+static void __Pyx_RaiseArgtupleInvalid(const char* func_name, int exact,
+    Py_ssize_t num_min, Py_ssize_t num_max, Py_ssize_t num_found);
+
+/* RaiseDoubleKeywords.proto */
+static void __Pyx_RaiseDoubleKeywordsError(const char* func_name, PyObject* kw_name);
+
+/* ParseKeywords.proto */
+static int __Pyx_ParseOptionalKeywords(PyObject *kwds, PyObject **argnames[],\
+    PyObject *kwds2, PyObject *values[], Py_ssize_t num_pos_args,\
+    const char* function_name);
+
+/* ArgTypeTest.proto */
+#define __Pyx_ArgTypeTest(obj, type, none_allowed, name, exact)\
+    ((likely((Py_TYPE(obj) == type) | (none_allowed && (obj == Py_None)))) ? 1 :\
+        __Pyx__ArgTypeTest(obj, type, name, exact))
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact);
+
+/* IsLittleEndian.proto */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void);
+
+/* BufferFormatCheck.proto */
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts);
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type);
+
+/* BufferGetAndValidate.proto */
+#define __Pyx_GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)\
+    ((obj == Py_None || obj == NULL) ?\
+    (__Pyx_ZeroBuffer(buf), 0) :\
+    __Pyx__GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack))
+static int  __Pyx__GetBufferAndValidate(Py_buffer* buf, PyObject* obj,
+    __Pyx_TypeInfo* dtype, int flags, int nd, int cast, __Pyx_BufFmt_StackElem* stack);
+static void __Pyx_ZeroBuffer(Py_buffer* buf);
+static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info);
+static Py_ssize_t __Pyx_minusones[] = { -1, -1, -1, -1, -1, -1, -1, -1 };
+static Py_ssize_t __Pyx_zeros[] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+
+/* PyThreadStateGet.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyThreadState_declare  PyThreadState *__pyx_tstate;
+#define __Pyx_PyThreadState_assign  __pyx_tstate = __Pyx_PyThreadState_Current;
+#define __Pyx_PyErr_Occurred()  __pyx_tstate->curexc_type
+#else
+#define __Pyx_PyThreadState_declare
+#define __Pyx_PyThreadState_assign
+#define __Pyx_PyErr_Occurred()  PyErr_Occurred()
+#endif
+
+/* PyErrFetchRestore.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)    __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)    __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL))
+#else
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#endif
+#else
+#define __Pyx_PyErr_Clear() PyErr_Clear()
+#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc)
+#define __Pyx_ErrRestoreWithState(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchWithState(type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestoreInState(tstate, type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetchInState(tstate, type, value, tb)  PyErr_Fetch(type, value, tb)
+#define __Pyx_ErrRestore(type, value, tb)  PyErr_Restore(type, value, tb)
+#define __Pyx_ErrFetch(type, value, tb)  PyErr_Fetch(type, value, tb)
+#endif
+
+/* PyObjectGetAttrStr.proto */
+#if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name);
+#else
+#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n)
+#endif
+
+/* GetBuiltinName.proto */
+static PyObject *__Pyx_GetBuiltinName(PyObject *name);
+
+/* GetTopmostException.proto */
+#if CYTHON_USE_EXC_INFO_STACK
+static _PyErr_StackItem * __Pyx_PyErr_GetTopmostException(PyThreadState *tstate);
+#endif
+
+/* SaveResetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_ExceptionSave(type, value, tb)  __Pyx__ExceptionSave(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#define __Pyx_ExceptionReset(type, value, tb)  __Pyx__ExceptionReset(__pyx_tstate, type, value, tb)
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb);
+#else
+#define __Pyx_ExceptionSave(type, value, tb)   PyErr_GetExcInfo(type, value, tb)
+#define __Pyx_ExceptionReset(type, value, tb)  PyErr_SetExcInfo(type, value, tb)
+#endif
+
+/* PyErrExceptionMatches.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_PyErr_ExceptionMatches(err) __Pyx_PyErr_ExceptionMatchesInState(__pyx_tstate, err)
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err);
+#else
+#define __Pyx_PyErr_ExceptionMatches(err)  PyErr_ExceptionMatches(err)
+#endif
+
+/* GetException.proto */
+#if CYTHON_FAST_THREAD_STATE
+#define __Pyx_GetException(type, value, tb)  __Pyx__GetException(__pyx_tstate, type, value, tb)
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb);
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb);
+#endif
+
+/* PyObjectCall.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw);
+#else
+#define __Pyx_PyObject_Call(func, arg, kw) PyObject_Call(func, arg, kw)
+#endif
+
+/* RaiseException.proto */
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause);
+
+/* TypeImport.proto */
+#ifndef __PYX_HAVE_RT_ImportType_proto
+#define __PYX_HAVE_RT_ImportType_proto
+enum __Pyx_ImportType_CheckSize {
+   __Pyx_ImportType_CheckSize_Error = 0,
+   __Pyx_ImportType_CheckSize_Warn = 1,
+   __Pyx_ImportType_CheckSize_Ignore = 2
+};
+static PyTypeObject *__Pyx_ImportType(PyObject* module, const char *module_name, const char *class_name, size_t size, enum __Pyx_ImportType_CheckSize check_size);
+#endif
+
+/* Import.proto */
+static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level);
+
+/* PyDictVersioning.proto */
+#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+#define __PYX_DICT_VERSION_INIT  ((PY_UINT64_T) -1)
+#define __PYX_GET_DICT_VERSION(dict)  (((PyDictObject*)(dict))->ma_version_tag)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)\
+    (version_var) = __PYX_GET_DICT_VERSION(dict);\
+    (cache_var) = (value);
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) {\
+    static PY_UINT64_T __pyx_dict_version = 0;\
+    static PyObject *__pyx_dict_cached_value = NULL;\
+    if (likely(__PYX_GET_DICT_VERSION(DICT) == __pyx_dict_version)) {\
+        (VAR) = __pyx_dict_cached_value;\
+    } else {\
+        (VAR) = __pyx_dict_cached_value = (LOOKUP);\
+        __pyx_dict_version = __PYX_GET_DICT_VERSION(DICT);\
+    }\
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj);
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj);
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version);
+#else
+#define __PYX_GET_DICT_VERSION(dict)  (0)
+#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)
+#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP)  (VAR) = (LOOKUP);
+#endif
+
+/* CLineInTraceback.proto */
+#ifdef CYTHON_CLINE_IN_TRACEBACK
+#define __Pyx_CLineForTraceback(tstate, c_line)  (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0)
+#else
+static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line);
+#endif
+
+/* CodeObjectCache.proto */
+typedef struct {
+    PyCodeObject* code_object;
+    int code_line;
+} __Pyx_CodeObjectCacheEntry;
+struct __Pyx_CodeObjectCache {
+    int count;
+    int max_count;
+    __Pyx_CodeObjectCacheEntry* entries;
+};
+static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL};
+static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line);
+static PyCodeObject *__pyx_find_code_object(int code_line);
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object);
+
+/* AddTraceback.proto */
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename);
+
+/* BufferStructDeclare.proto */
+typedef struct {
+  Py_ssize_t shape, strides, suboffsets;
+} __Pyx_Buf_DimInfo;
+typedef struct {
+  size_t refcount;
+  Py_buffer pybuffer;
+} __Pyx_Buffer;
+typedef struct {
+  __Pyx_Buffer *rcbuffer;
+  char *data;
+  __Pyx_Buf_DimInfo diminfo[8];
+} __Pyx_LocalBuf_ND;
+
+#if PY_MAJOR_VERSION < 3
+    static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags);
+    static void __Pyx_ReleaseBuffer(Py_buffer *view);
+#else
+    #define __Pyx_GetBuffer PyObject_GetBuffer
+    #define __Pyx_ReleaseBuffer PyBuffer_Release
+#endif
+
+
+/* GCCDiagnostics.proto */
+#if defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6))
+#define __Pyx_HAS_GCC_DIAGNOSTIC
+#endif
+
+/* RealImag.proto */
+#if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    #define __Pyx_CREAL(z) ((z).real())
+    #define __Pyx_CIMAG(z) ((z).imag())
+  #else
+    #define __Pyx_CREAL(z) (__real__(z))
+    #define __Pyx_CIMAG(z) (__imag__(z))
+  #endif
+#else
+    #define __Pyx_CREAL(z) ((z).real)
+    #define __Pyx_CIMAG(z) ((z).imag)
+#endif
+#if defined(__cplusplus) && CYTHON_CCOMPLEX\
+        && (defined(_WIN32) || defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 5 || __GNUC__ == 4 && __GNUC_MINOR__ >= 4 )) || __cplusplus >= 201103)
+    #define __Pyx_SET_CREAL(z,x) ((z).real(x))
+    #define __Pyx_SET_CIMAG(z,y) ((z).imag(y))
+#else
+    #define __Pyx_SET_CREAL(z,x) __Pyx_CREAL(z) = (x)
+    #define __Pyx_SET_CIMAG(z,y) __Pyx_CIMAG(z) = (y)
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_float(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_float(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_float(a, b) ((a)-(b))
+    #define __Pyx_c_prod_float(a, b) ((a)*(b))
+    #define __Pyx_c_quot_float(a, b) ((a)/(b))
+    #define __Pyx_c_neg_float(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_float(z) ((z)==(float)0)
+    #define __Pyx_c_conj_float(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (::std::abs(z))
+        #define __Pyx_c_pow_float(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_float(z) ((z)==0)
+    #define __Pyx_c_conj_float(z)    (conjf(z))
+    #if 1
+        #define __Pyx_c_abs_float(z)     (cabsf(z))
+        #define __Pyx_c_pow_float(a, b)  (cpowf(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex);
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex);
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex);
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex, __pyx_t_float_complex);
+    #endif
+#endif
+
+/* Arithmetic.proto */
+#if CYTHON_CCOMPLEX
+    #define __Pyx_c_eq_double(a, b)   ((a)==(b))
+    #define __Pyx_c_sum_double(a, b)  ((a)+(b))
+    #define __Pyx_c_diff_double(a, b) ((a)-(b))
+    #define __Pyx_c_prod_double(a, b) ((a)*(b))
+    #define __Pyx_c_quot_double(a, b) ((a)/(b))
+    #define __Pyx_c_neg_double(a)     (-(a))
+  #ifdef __cplusplus
+    #define __Pyx_c_is_zero_double(z) ((z)==(double)0)
+    #define __Pyx_c_conj_double(z)    (::std::conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (::std::abs(z))
+        #define __Pyx_c_pow_double(a, b)  (::std::pow(a, b))
+    #endif
+  #else
+    #define __Pyx_c_is_zero_double(z) ((z)==0)
+    #define __Pyx_c_conj_double(z)    (conj(z))
+    #if 1
+        #define __Pyx_c_abs_double(z)     (cabs(z))
+        #define __Pyx_c_pow_double(a, b)  (cpow(a, b))
+    #endif
+ #endif
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex);
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex);
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex);
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex);
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex, __pyx_t_double_complex);
+    #endif
+#endif
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *);
+
+/* CIntToPy.proto */
+static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value);
+
+/* CIntFromPy.proto */
+static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *);
+
+/* FastTypeChecks.proto */
+#if CYTHON_COMPILING_IN_CPYTHON
+#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type)
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type);
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2);
+#else
+#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type)
+#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type)
+#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2))
+#endif
+#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception)
+
+/* CheckBinaryVersion.proto */
+static int __Pyx_check_binary_version(void);
+
+/* InitStrings.proto */
+static int __Pyx_InitStrings(__Pyx_StringTabEntry *t);
+
+
+/* Module declarations from 'cpython.buffer' */
+
+/* Module declarations from 'libc.string' */
+
+/* Module declarations from 'libc.stdio' */
+
+/* Module declarations from '__builtin__' */
+
+/* Module declarations from 'cpython.type' */
+static PyTypeObject *__pyx_ptype_7cpython_4type_type = 0;
+
+/* Module declarations from 'cpython' */
+
+/* Module declarations from 'cpython.object' */
+
+/* Module declarations from 'cpython.ref' */
+
+/* Module declarations from 'cpython.mem' */
+
+/* Module declarations from 'numpy' */
+
+/* Module declarations from 'numpy' */
+static PyTypeObject *__pyx_ptype_5numpy_dtype = 0;
+static PyTypeObject *__pyx_ptype_5numpy_flatiter = 0;
+static PyTypeObject *__pyx_ptype_5numpy_broadcast = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ndarray = 0;
+static PyTypeObject *__pyx_ptype_5numpy_generic = 0;
+static PyTypeObject *__pyx_ptype_5numpy_number = 0;
+static PyTypeObject *__pyx_ptype_5numpy_integer = 0;
+static PyTypeObject *__pyx_ptype_5numpy_signedinteger = 0;
+static PyTypeObject *__pyx_ptype_5numpy_unsignedinteger = 0;
+static PyTypeObject *__pyx_ptype_5numpy_inexact = 0;
+static PyTypeObject *__pyx_ptype_5numpy_floating = 0;
+static PyTypeObject *__pyx_ptype_5numpy_complexfloating = 0;
+static PyTypeObject *__pyx_ptype_5numpy_flexible = 0;
+static PyTypeObject *__pyx_ptype_5numpy_character = 0;
+static PyTypeObject *__pyx_ptype_5numpy_ufunc = 0;
+static CYTHON_INLINE int __pyx_f_5numpy_import_array(void); /*proto*/
+
+/* Module declarations from 'libcpp.string' */
+
+/* Module declarations from 'insightface.thirdparty.face3d.mesh.cython.mesh_core_cython' */
+static std::string __pyx_convert_string_from_py_std__in_string(PyObject *); /*proto*/
+static __Pyx_TypeInfo __Pyx_TypeInfo_float = { "float", NULL, sizeof(float), { 0 }, 0, 'R', 0, 0 };
+static __Pyx_TypeInfo __Pyx_TypeInfo_int = { "int", NULL, sizeof(int), { 0 }, 0, IS_UNSIGNED(int) ? 'U' : 'I', IS_UNSIGNED(int), 0 };
+#define __Pyx_MODULE_NAME "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython"
+extern int __pyx_module_is_main_insightface__thirdparty__face3d__mesh__cython__mesh_core_cython;
+int __pyx_module_is_main_insightface__thirdparty__face3d__mesh__cython__mesh_core_cython = 0;
+
+/* Implementation of 'insightface.thirdparty.face3d.mesh.cython.mesh_core_cython' */
+static PyObject *__pyx_builtin_ImportError;
+static const char __pyx_k_c[] = "c";
+static const char __pyx_k_h[] = "h";
+static const char __pyx_k_w[] = "w";
+static const char __pyx_k_np[] = "np";
+static const char __pyx_k_main[] = "__main__";
+static const char __pyx_k_name[] = "__name__";
+static const char __pyx_k_ntri[] = "ntri";
+static const char __pyx_k_nver[] = "nver";
+static const char __pyx_k_test[] = "__test__";
+static const char __pyx_k_image[] = "image";
+static const char __pyx_k_numpy[] = "numpy";
+static const char __pyx_k_tex_c[] = "tex_c";
+static const char __pyx_k_tex_h[] = "tex_h";
+static const char __pyx_k_tex_w[] = "tex_w";
+static const char __pyx_k_colors[] = "colors";
+static const char __pyx_k_import[] = "__import__";
+static const char __pyx_k_normal[] = "normal";
+static const char __pyx_k_ntexver[] = "ntexver";
+static const char __pyx_k_texture[] = "texture";
+static const char __pyx_k_filename[] = "filename";
+static const char __pyx_k_mtl_name[] = "mtl_name";
+static const char __pyx_k_tex_nver[] = "tex_nver";
+static const char __pyx_k_vertices[] = "vertices";
+static const char __pyx_k_triangles[] = "triangles";
+static const char __pyx_k_uv_coords[] = "uv_coords";
+static const char __pyx_k_tex_coords[] = "tex_coords";
+static const char __pyx_k_tri_normal[] = "tri_normal";
+static const char __pyx_k_ImportError[] = "ImportError";
+static const char __pyx_k_depth_buffer[] = "depth_buffer";
+static const char __pyx_k_mapping_type[] = "mapping_type";
+static const char __pyx_k_tex_triangles[] = "tex_triangles";
+static const char __pyx_k_get_normal_core[] = "get_normal_core";
+static const char __pyx_k_triangle_buffer[] = "triangle_buffer";
+static const char __pyx_k_barycentric_weight[] = "barycentric_weight";
+static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback";
+static const char __pyx_k_render_colors_core[] = "render_colors_core";
+static const char __pyx_k_render_texture_core[] = "render_texture_core";
+static const char __pyx_k_rasterize_triangles_core[] = "rasterize_triangles_core";
+static const char __pyx_k_numpy_core_multiarray_failed_to[] = "numpy.core.multiarray failed to import";
+static const char __pyx_k_insightface_thirdparty_face3d_me[] = "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx";
+static const char __pyx_k_numpy_core_umath_failed_to_impor[] = "numpy.core.umath failed to import";
+static const char __pyx_k_write_obj_with_colors_texture_co[] = "write_obj_with_colors_texture_core";
+static const char __pyx_k_insightface_thirdparty_face3d_me_2[] = "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython";
+static PyObject *__pyx_n_s_ImportError;
+static PyObject *__pyx_n_s_barycentric_weight;
+static PyObject *__pyx_n_s_c;
+static PyObject *__pyx_n_s_cline_in_traceback;
+static PyObject *__pyx_n_s_colors;
+static PyObject *__pyx_n_s_depth_buffer;
+static PyObject *__pyx_n_s_filename;
+static PyObject *__pyx_n_s_get_normal_core;
+static PyObject *__pyx_n_s_h;
+static PyObject *__pyx_n_s_image;
+static PyObject *__pyx_n_s_import;
+static PyObject *__pyx_kp_s_insightface_thirdparty_face3d_me;
+static PyObject *__pyx_n_s_insightface_thirdparty_face3d_me_2;
+static PyObject *__pyx_n_s_main;
+static PyObject *__pyx_n_s_mapping_type;
+static PyObject *__pyx_n_s_mtl_name;
+static PyObject *__pyx_n_s_name;
+static PyObject *__pyx_n_s_normal;
+static PyObject *__pyx_n_s_np;
+static PyObject *__pyx_n_s_ntexver;
+static PyObject *__pyx_n_s_ntri;
+static PyObject *__pyx_n_s_numpy;
+static PyObject *__pyx_kp_s_numpy_core_multiarray_failed_to;
+static PyObject *__pyx_kp_s_numpy_core_umath_failed_to_impor;
+static PyObject *__pyx_n_s_nver;
+static PyObject *__pyx_n_s_rasterize_triangles_core;
+static PyObject *__pyx_n_s_render_colors_core;
+static PyObject *__pyx_n_s_render_texture_core;
+static PyObject *__pyx_n_s_test;
+static PyObject *__pyx_n_s_tex_c;
+static PyObject *__pyx_n_s_tex_coords;
+static PyObject *__pyx_n_s_tex_h;
+static PyObject *__pyx_n_s_tex_nver;
+static PyObject *__pyx_n_s_tex_triangles;
+static PyObject *__pyx_n_s_tex_w;
+static PyObject *__pyx_n_s_texture;
+static PyObject *__pyx_n_s_tri_normal;
+static PyObject *__pyx_n_s_triangle_buffer;
+static PyObject *__pyx_n_s_triangles;
+static PyObject *__pyx_n_s_uv_coords;
+static PyObject *__pyx_n_s_vertices;
+static PyObject *__pyx_n_s_w;
+static PyObject *__pyx_n_s_write_obj_with_colors_texture_co;
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_get_normal_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_normal, PyArrayObject *__pyx_v_tri_normal, PyArrayObject *__pyx_v_triangles, int __pyx_v_ntri); /* proto */
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_2rasterize_triangles_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_depth_buffer, PyArrayObject *__pyx_v_triangle_buffer, PyArrayObject *__pyx_v_barycentric_weight, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w); /* proto */
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_4render_colors_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c); /* proto */
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_6render_texture_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_texture, PyArrayObject *__pyx_v_tex_coords, PyArrayObject *__pyx_v_tex_triangles, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_tex_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c, int __pyx_v_tex_h, int __pyx_v_tex_w, int __pyx_v_tex_c, int __pyx_v_mapping_type); /* proto */
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_8write_obj_with_colors_texture_core(CYTHON_UNUSED PyObject *__pyx_self, std::string __pyx_v_filename, std::string __pyx_v_mtl_name, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_uv_coords, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_ntexver); /* proto */
+static PyObject *__pyx_tuple_;
+static PyObject *__pyx_tuple__2;
+static PyObject *__pyx_tuple__3;
+static PyObject *__pyx_tuple__5;
+static PyObject *__pyx_tuple__7;
+static PyObject *__pyx_tuple__9;
+static PyObject *__pyx_tuple__11;
+static PyObject *__pyx_codeobj__4;
+static PyObject *__pyx_codeobj__6;
+static PyObject *__pyx_codeobj__8;
+static PyObject *__pyx_codeobj__10;
+static PyObject *__pyx_codeobj__12;
+/* Late includes */
+
+/* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_1get_normal_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_1get_normal_core = {"get_normal_core", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_1get_normal_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_1get_normal_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_normal = 0;
+  PyArrayObject *__pyx_v_tri_normal = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  int __pyx_v_ntri;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("get_normal_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_normal,&__pyx_n_s_tri_normal,&__pyx_n_s_triangles,&__pyx_n_s_ntri,0};
+    PyObject* values[4] = {0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_normal)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tri_normal)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 1); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 2); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, 3); __PYX_ERR(0, 40, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "get_normal_core") < 0)) __PYX_ERR(0, 40, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 4) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+    }
+    __pyx_v_normal = ((PyArrayObject *)values[0]);
+    __pyx_v_tri_normal = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[3]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 43, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("get_normal_core", 1, 4, 4, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 40, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.get_normal_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_normal), __pyx_ptype_5numpy_ndarray, 0, "normal", 0))) __PYX_ERR(0, 40, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tri_normal), __pyx_ptype_5numpy_ndarray, 0, "tri_normal", 0))) __PYX_ERR(0, 41, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 42, __pyx_L1_error)
+  __pyx_r = __pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_get_normal_core(__pyx_self, __pyx_v_normal, __pyx_v_tri_normal, __pyx_v_triangles, __pyx_v_ntri);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_get_normal_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_normal, PyArrayObject *__pyx_v_tri_normal, PyArrayObject *__pyx_v_triangles, int __pyx_v_ntri) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_normal;
+  __Pyx_Buffer __pyx_pybuffer_normal;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tri_normal;
+  __Pyx_Buffer __pyx_pybuffer_tri_normal;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("get_normal_core", 0);
+  __pyx_pybuffer_normal.pybuffer.buf = NULL;
+  __pyx_pybuffer_normal.refcount = 0;
+  __pyx_pybuffernd_normal.data = NULL;
+  __pyx_pybuffernd_normal.rcbuffer = &__pyx_pybuffer_normal;
+  __pyx_pybuffer_tri_normal.pybuffer.buf = NULL;
+  __pyx_pybuffer_tri_normal.refcount = 0;
+  __pyx_pybuffernd_tri_normal.data = NULL;
+  __pyx_pybuffernd_tri_normal.rcbuffer = &__pyx_pybuffer_tri_normal;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_normal.rcbuffer->pybuffer, (PyObject*)__pyx_v_normal, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_normal.diminfo[0].strides = __pyx_pybuffernd_normal.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_normal.diminfo[0].shape = __pyx_pybuffernd_normal.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_normal.diminfo[1].strides = __pyx_pybuffernd_normal.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_normal.diminfo[1].shape = __pyx_pybuffernd_normal.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer, (PyObject*)__pyx_v_tri_normal, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tri_normal.diminfo[0].strides = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tri_normal.diminfo[0].shape = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tri_normal.diminfo[1].strides = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tri_normal.diminfo[1].shape = __pyx_pybuffernd_tri_normal.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":45
+ *                 int ntri
+ *                 ):
+ *     _get_normal_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(normal), <float*> np.PyArray_DATA(tri_normal), <int*> np.PyArray_DATA(triangles),
+ *         ntri)
+ */
+  _get_normal_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_normal))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_tri_normal))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), __pyx_v_ntri);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_normal.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.get_normal_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_normal.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tri_normal.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_3rasterize_triangles_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_3rasterize_triangles_core = {"rasterize_triangles_core", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_3rasterize_triangles_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_3rasterize_triangles_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  PyArrayObject *__pyx_v_triangle_buffer = 0;
+  PyArrayObject *__pyx_v_barycentric_weight = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("rasterize_triangles_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_depth_buffer,&__pyx_n_s_triangle_buffer,&__pyx_n_s_barycentric_weight,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,0};
+    PyObject* values[9] = {0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 1); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 2); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangle_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 3); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_barycentric_weight)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 4); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 5); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 6); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 7); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, 8); __PYX_ERR(0, 49, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "rasterize_triangles_core") < 0)) __PYX_ERR(0, 49, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 9) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+    }
+    __pyx_v_vertices = ((PyArrayObject *)values[0]);
+    __pyx_v_triangles = ((PyArrayObject *)values[1]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[2]);
+    __pyx_v_triangle_buffer = ((PyArrayObject *)values[3]);
+    __pyx_v_barycentric_weight = ((PyArrayObject *)values[4]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[5]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 55, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 55, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 56, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 56, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("rasterize_triangles_core", 1, 9, 9, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 49, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.rasterize_triangles_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 50, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 51, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 52, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangle_buffer), __pyx_ptype_5numpy_ndarray, 0, "triangle_buffer", 0))) __PYX_ERR(0, 53, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_barycentric_weight), __pyx_ptype_5numpy_ndarray, 0, "barycentric_weight", 0))) __PYX_ERR(0, 54, __pyx_L1_error)
+  __pyx_r = __pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_2rasterize_triangles_core(__pyx_self, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_depth_buffer, __pyx_v_triangle_buffer, __pyx_v_barycentric_weight, __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_2rasterize_triangles_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_depth_buffer, PyArrayObject *__pyx_v_triangle_buffer, PyArrayObject *__pyx_v_barycentric_weight, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_barycentric_weight;
+  __Pyx_Buffer __pyx_pybuffer_barycentric_weight;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangle_buffer;
+  __Pyx_Buffer __pyx_pybuffer_triangle_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("rasterize_triangles_core", 0);
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  __pyx_pybuffer_triangle_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangle_buffer.refcount = 0;
+  __pyx_pybuffernd_triangle_buffer.data = NULL;
+  __pyx_pybuffernd_triangle_buffer.rcbuffer = &__pyx_pybuffer_triangle_buffer;
+  __pyx_pybuffer_barycentric_weight.pybuffer.buf = NULL;
+  __pyx_pybuffer_barycentric_weight.refcount = 0;
+  __pyx_pybuffernd_barycentric_weight.data = NULL;
+  __pyx_pybuffernd_barycentric_weight.rcbuffer = &__pyx_pybuffer_barycentric_weight;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangle_buffer, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangle_buffer.diminfo[0].strides = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangle_buffer.diminfo[0].shape = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangle_buffer.diminfo[1].strides = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangle_buffer.diminfo[1].shape = __pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer, (PyObject*)__pyx_v_barycentric_weight, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_barycentric_weight.diminfo[0].strides = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_barycentric_weight.diminfo[0].shape = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_barycentric_weight.diminfo[1].strides = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_barycentric_weight.diminfo[1].shape = __pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer.shape[1];
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":58
+ *                 int h, int w
+ *                 ):
+ *     _rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(depth_buffer), <int*> np.PyArray_DATA(triangle_buffer), <float*> np.PyArray_DATA(barycentric_weight),
+ */
+  _rasterize_triangles_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangle_buffer))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_barycentric_weight))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.rasterize_triangles_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_barycentric_weight.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangle_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_5render_colors_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_5render_colors_core = {"render_colors_core", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_5render_colors_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_5render_colors_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_image = 0;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_colors = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  int __pyx_v_c;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_colors_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_image,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_colors,&__pyx_n_s_depth_buffer,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,&__pyx_n_s_c,0};
+    PyObject* values[10] = {0,0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case 10: values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+        CYTHON_FALLTHROUGH;
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_image)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 1); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 2); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_colors)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 3); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 4); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 5); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 6); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 7); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 8); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  9:
+        if (likely((values[9] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, 9); __PYX_ERR(0, 64, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "render_colors_core") < 0)) __PYX_ERR(0, 64, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 10) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+      values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+    }
+    __pyx_v_image = ((PyArrayObject *)values[0]);
+    __pyx_v_vertices = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_colors = ((PyArrayObject *)values[3]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[4]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[5]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 69, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 69, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+    __pyx_v_c = __Pyx_PyInt_As_int(values[9]); if (unlikely((__pyx_v_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 70, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("render_colors_core", 1, 10, 10, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 64, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.render_colors_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_image), __pyx_ptype_5numpy_ndarray, 0, "image", 0))) __PYX_ERR(0, 64, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 65, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 66, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_colors), __pyx_ptype_5numpy_ndarray, 0, "colors", 0))) __PYX_ERR(0, 67, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 68, __pyx_L1_error)
+  __pyx_r = __pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_4render_colors_core(__pyx_self, __pyx_v_image, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_colors, __pyx_v_depth_buffer, __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_4render_colors_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_colors;
+  __Pyx_Buffer __pyx_pybuffer_colors;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_image;
+  __Pyx_Buffer __pyx_pybuffer_image;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("render_colors_core", 0);
+  __pyx_pybuffer_image.pybuffer.buf = NULL;
+  __pyx_pybuffer_image.refcount = 0;
+  __pyx_pybuffernd_image.data = NULL;
+  __pyx_pybuffernd_image.rcbuffer = &__pyx_pybuffer_image;
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_colors.pybuffer.buf = NULL;
+  __pyx_pybuffer_colors.refcount = 0;
+  __pyx_pybuffernd_colors.data = NULL;
+  __pyx_pybuffernd_colors.rcbuffer = &__pyx_pybuffer_colors;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_image.rcbuffer->pybuffer, (PyObject*)__pyx_v_image, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_image.diminfo[0].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_image.diminfo[0].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_image.diminfo[1].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_image.diminfo[1].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_image.diminfo[2].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_image.diminfo[2].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_colors.rcbuffer->pybuffer, (PyObject*)__pyx_v_colors, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_colors.diminfo[0].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_colors.diminfo[0].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_colors.diminfo[1].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_colors.diminfo[1].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":72
+ *                 int h, int w, int c
+ *                 ):
+ *     _render_colors_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(colors),
+ */
+  _render_colors_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_image))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_colors))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.render_colors_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_7render_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_7render_texture_core = {"render_texture_core", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_7render_texture_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_7render_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  PyArrayObject *__pyx_v_image = 0;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_texture = 0;
+  PyArrayObject *__pyx_v_tex_coords = 0;
+  PyArrayObject *__pyx_v_tex_triangles = 0;
+  PyArrayObject *__pyx_v_depth_buffer = 0;
+  int __pyx_v_nver;
+  int __pyx_v_tex_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_h;
+  int __pyx_v_w;
+  int __pyx_v_c;
+  int __pyx_v_tex_h;
+  int __pyx_v_tex_w;
+  int __pyx_v_tex_c;
+  int __pyx_v_mapping_type;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("render_texture_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_image,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_texture,&__pyx_n_s_tex_coords,&__pyx_n_s_tex_triangles,&__pyx_n_s_depth_buffer,&__pyx_n_s_nver,&__pyx_n_s_tex_nver,&__pyx_n_s_ntri,&__pyx_n_s_h,&__pyx_n_s_w,&__pyx_n_s_c,&__pyx_n_s_tex_h,&__pyx_n_s_tex_w,&__pyx_n_s_tex_c,&__pyx_n_s_mapping_type,0};
+    PyObject* values[17] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case 17: values[16] = PyTuple_GET_ITEM(__pyx_args, 16);
+        CYTHON_FALLTHROUGH;
+        case 16: values[15] = PyTuple_GET_ITEM(__pyx_args, 15);
+        CYTHON_FALLTHROUGH;
+        case 15: values[14] = PyTuple_GET_ITEM(__pyx_args, 14);
+        CYTHON_FALLTHROUGH;
+        case 14: values[13] = PyTuple_GET_ITEM(__pyx_args, 13);
+        CYTHON_FALLTHROUGH;
+        case 13: values[12] = PyTuple_GET_ITEM(__pyx_args, 12);
+        CYTHON_FALLTHROUGH;
+        case 12: values[11] = PyTuple_GET_ITEM(__pyx_args, 11);
+        CYTHON_FALLTHROUGH;
+        case 11: values[10] = PyTuple_GET_ITEM(__pyx_args, 10);
+        CYTHON_FALLTHROUGH;
+        case 10: values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+        CYTHON_FALLTHROUGH;
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_image)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 1); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 2); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_texture)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 3); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_coords)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 4); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 5); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_depth_buffer)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 6); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 7); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 8); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  9:
+        if (likely((values[9] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 9); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 10:
+        if (likely((values[10] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 10); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 11:
+        if (likely((values[11] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 11); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 12:
+        if (likely((values[12] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 12); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 13:
+        if (likely((values[13] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_h)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 13); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 14:
+        if (likely((values[14] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_w)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 14); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 15:
+        if (likely((values[15] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_tex_c)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 15); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case 16:
+        if (likely((values[16] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_mapping_type)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, 16); __PYX_ERR(0, 79, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "render_texture_core") < 0)) __PYX_ERR(0, 79, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 17) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+      values[9] = PyTuple_GET_ITEM(__pyx_args, 9);
+      values[10] = PyTuple_GET_ITEM(__pyx_args, 10);
+      values[11] = PyTuple_GET_ITEM(__pyx_args, 11);
+      values[12] = PyTuple_GET_ITEM(__pyx_args, 12);
+      values[13] = PyTuple_GET_ITEM(__pyx_args, 13);
+      values[14] = PyTuple_GET_ITEM(__pyx_args, 14);
+      values[15] = PyTuple_GET_ITEM(__pyx_args, 15);
+      values[16] = PyTuple_GET_ITEM(__pyx_args, 16);
+    }
+    __pyx_v_image = ((PyArrayObject *)values[0]);
+    __pyx_v_vertices = ((PyArrayObject *)values[1]);
+    __pyx_v_triangles = ((PyArrayObject *)values[2]);
+    __pyx_v_texture = ((PyArrayObject *)values[3]);
+    __pyx_v_tex_coords = ((PyArrayObject *)values[4]);
+    __pyx_v_tex_triangles = ((PyArrayObject *)values[5]);
+    __pyx_v_depth_buffer = ((PyArrayObject *)values[6]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_tex_nver = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_tex_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[9]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 86, __pyx_L3_error)
+    __pyx_v_h = __Pyx_PyInt_As_int(values[10]); if (unlikely((__pyx_v_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_w = __Pyx_PyInt_As_int(values[11]); if (unlikely((__pyx_v_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_c = __Pyx_PyInt_As_int(values[12]); if (unlikely((__pyx_v_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 87, __pyx_L3_error)
+    __pyx_v_tex_h = __Pyx_PyInt_As_int(values[13]); if (unlikely((__pyx_v_tex_h == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_tex_w = __Pyx_PyInt_As_int(values[14]); if (unlikely((__pyx_v_tex_w == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_tex_c = __Pyx_PyInt_As_int(values[15]); if (unlikely((__pyx_v_tex_c == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 88, __pyx_L3_error)
+    __pyx_v_mapping_type = __Pyx_PyInt_As_int(values[16]); if (unlikely((__pyx_v_mapping_type == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 89, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("render_texture_core", 1, 17, 17, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 79, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.render_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_image), __pyx_ptype_5numpy_ndarray, 0, "image", 0))) __PYX_ERR(0, 79, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 80, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 81, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_texture), __pyx_ptype_5numpy_ndarray, 0, "texture", 0))) __PYX_ERR(0, 82, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tex_coords), __pyx_ptype_5numpy_ndarray, 0, "tex_coords", 0))) __PYX_ERR(0, 83, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_tex_triangles), __pyx_ptype_5numpy_ndarray, 0, "tex_triangles", 0))) __PYX_ERR(0, 84, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_depth_buffer), __pyx_ptype_5numpy_ndarray, 0, "depth_buffer", 0))) __PYX_ERR(0, 85, __pyx_L1_error)
+  __pyx_r = __pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_6render_texture_core(__pyx_self, __pyx_v_image, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_texture, __pyx_v_tex_coords, __pyx_v_tex_triangles, __pyx_v_depth_buffer, __pyx_v_nver, __pyx_v_tex_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c, __pyx_v_tex_h, __pyx_v_tex_w, __pyx_v_tex_c, __pyx_v_mapping_type);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_6render_texture_core(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_image, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_texture, PyArrayObject *__pyx_v_tex_coords, PyArrayObject *__pyx_v_tex_triangles, PyArrayObject *__pyx_v_depth_buffer, int __pyx_v_nver, int __pyx_v_tex_nver, int __pyx_v_ntri, int __pyx_v_h, int __pyx_v_w, int __pyx_v_c, int __pyx_v_tex_h, int __pyx_v_tex_w, int __pyx_v_tex_c, int __pyx_v_mapping_type) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_depth_buffer;
+  __Pyx_Buffer __pyx_pybuffer_depth_buffer;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_image;
+  __Pyx_Buffer __pyx_pybuffer_image;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tex_coords;
+  __Pyx_Buffer __pyx_pybuffer_tex_coords;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_tex_triangles;
+  __Pyx_Buffer __pyx_pybuffer_tex_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_texture;
+  __Pyx_Buffer __pyx_pybuffer_texture;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("render_texture_core", 0);
+  __pyx_pybuffer_image.pybuffer.buf = NULL;
+  __pyx_pybuffer_image.refcount = 0;
+  __pyx_pybuffernd_image.data = NULL;
+  __pyx_pybuffernd_image.rcbuffer = &__pyx_pybuffer_image;
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_texture.pybuffer.buf = NULL;
+  __pyx_pybuffer_texture.refcount = 0;
+  __pyx_pybuffernd_texture.data = NULL;
+  __pyx_pybuffernd_texture.rcbuffer = &__pyx_pybuffer_texture;
+  __pyx_pybuffer_tex_coords.pybuffer.buf = NULL;
+  __pyx_pybuffer_tex_coords.refcount = 0;
+  __pyx_pybuffernd_tex_coords.data = NULL;
+  __pyx_pybuffernd_tex_coords.rcbuffer = &__pyx_pybuffer_tex_coords;
+  __pyx_pybuffer_tex_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_tex_triangles.refcount = 0;
+  __pyx_pybuffernd_tex_triangles.data = NULL;
+  __pyx_pybuffernd_tex_triangles.rcbuffer = &__pyx_pybuffer_tex_triangles;
+  __pyx_pybuffer_depth_buffer.pybuffer.buf = NULL;
+  __pyx_pybuffer_depth_buffer.refcount = 0;
+  __pyx_pybuffernd_depth_buffer.data = NULL;
+  __pyx_pybuffernd_depth_buffer.rcbuffer = &__pyx_pybuffer_depth_buffer;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_image.rcbuffer->pybuffer, (PyObject*)__pyx_v_image, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_image.diminfo[0].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_image.diminfo[0].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_image.diminfo[1].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_image.diminfo[1].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_image.diminfo[2].strides = __pyx_pybuffernd_image.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_image.diminfo[2].shape = __pyx_pybuffernd_image.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_texture.rcbuffer->pybuffer, (PyObject*)__pyx_v_texture, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 3, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_texture.diminfo[0].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_texture.diminfo[0].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_texture.diminfo[1].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_texture.diminfo[1].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[1]; __pyx_pybuffernd_texture.diminfo[2].strides = __pyx_pybuffernd_texture.rcbuffer->pybuffer.strides[2]; __pyx_pybuffernd_texture.diminfo[2].shape = __pyx_pybuffernd_texture.rcbuffer->pybuffer.shape[2];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer, (PyObject*)__pyx_v_tex_coords, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tex_coords.diminfo[0].strides = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tex_coords.diminfo[0].shape = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tex_coords.diminfo[1].strides = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tex_coords.diminfo[1].shape = __pyx_pybuffernd_tex_coords.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_tex_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_tex_triangles.diminfo[0].strides = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_tex_triangles.diminfo[0].shape = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_tex_triangles.diminfo[1].strides = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_tex_triangles.diminfo[1].shape = __pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer, (PyObject*)__pyx_v_depth_buffer, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_depth_buffer.diminfo[0].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_depth_buffer.diminfo[0].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_depth_buffer.diminfo[1].strides = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_depth_buffer.diminfo[1].shape = __pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer.shape[1];
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":91
+ *                 int mapping_type
+ *                 ):
+ *     _render_texture_core(             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),
+ *         <float*> np.PyArray_DATA(texture), <float*> np.PyArray_DATA(tex_coords), <int*> np.PyArray_DATA(tex_triangles),
+ */
+  _render_texture_core(((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_image))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_texture))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_tex_coords))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_tex_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_depth_buffer))), __pyx_v_nver, __pyx_v_tex_nver, __pyx_v_ntri, __pyx_v_h, __pyx_v_w, __pyx_v_c, __pyx_v_tex_h, __pyx_v_tex_w, __pyx_v_tex_c, __pyx_v_mapping_type);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_texture.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.render_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_depth_buffer.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_image.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_coords.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_tex_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_texture.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+/* Python wrapper */
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_9write_obj_with_colors_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/
+static PyMethodDef __pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_9write_obj_with_colors_texture_core = {"write_obj_with_colors_texture_core", (PyCFunction)(void*)(PyCFunctionWithKeywords)__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_9write_obj_with_colors_texture_core, METH_VARARGS|METH_KEYWORDS, 0};
+static PyObject *__pyx_pw_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_9write_obj_with_colors_texture_core(PyObject *__pyx_self, PyObject *__pyx_args, PyObject *__pyx_kwds) {
+  std::string __pyx_v_filename;
+  std::string __pyx_v_mtl_name;
+  PyArrayObject *__pyx_v_vertices = 0;
+  PyArrayObject *__pyx_v_triangles = 0;
+  PyArrayObject *__pyx_v_colors = 0;
+  PyArrayObject *__pyx_v_uv_coords = 0;
+  int __pyx_v_nver;
+  int __pyx_v_ntri;
+  int __pyx_v_ntexver;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  PyObject *__pyx_r = 0;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("write_obj_with_colors_texture_core (wrapper)", 0);
+  {
+    static PyObject **__pyx_pyargnames[] = {&__pyx_n_s_filename,&__pyx_n_s_mtl_name,&__pyx_n_s_vertices,&__pyx_n_s_triangles,&__pyx_n_s_colors,&__pyx_n_s_uv_coords,&__pyx_n_s_nver,&__pyx_n_s_ntri,&__pyx_n_s_ntexver,0};
+    PyObject* values[9] = {0,0,0,0,0,0,0,0,0};
+    if (unlikely(__pyx_kwds)) {
+      Py_ssize_t kw_args;
+      const Py_ssize_t pos_args = PyTuple_GET_SIZE(__pyx_args);
+      switch (pos_args) {
+        case  9: values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+        CYTHON_FALLTHROUGH;
+        case  8: values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+        CYTHON_FALLTHROUGH;
+        case  7: values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+        CYTHON_FALLTHROUGH;
+        case  6: values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+        CYTHON_FALLTHROUGH;
+        case  5: values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+        CYTHON_FALLTHROUGH;
+        case  4: values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+        CYTHON_FALLTHROUGH;
+        case  3: values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+        CYTHON_FALLTHROUGH;
+        case  2: values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+        CYTHON_FALLTHROUGH;
+        case  1: values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+        CYTHON_FALLTHROUGH;
+        case  0: break;
+        default: goto __pyx_L5_argtuple_error;
+      }
+      kw_args = PyDict_Size(__pyx_kwds);
+      switch (pos_args) {
+        case  0:
+        if (likely((values[0] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_filename)) != 0)) kw_args--;
+        else goto __pyx_L5_argtuple_error;
+        CYTHON_FALLTHROUGH;
+        case  1:
+        if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_mtl_name)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 1); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  2:
+        if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_vertices)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 2); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  3:
+        if (likely((values[3] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_triangles)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 3); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  4:
+        if (likely((values[4] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_colors)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 4); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  5:
+        if (likely((values[5] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_uv_coords)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 5); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  6:
+        if (likely((values[6] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_nver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 6); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  7:
+        if (likely((values[7] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntri)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 7); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+        CYTHON_FALLTHROUGH;
+        case  8:
+        if (likely((values[8] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_ntexver)) != 0)) kw_args--;
+        else {
+          __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, 8); __PYX_ERR(0, 100, __pyx_L3_error)
+        }
+      }
+      if (unlikely(kw_args > 0)) {
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "write_obj_with_colors_texture_core") < 0)) __PYX_ERR(0, 100, __pyx_L3_error)
+      }
+    } else if (PyTuple_GET_SIZE(__pyx_args) != 9) {
+      goto __pyx_L5_argtuple_error;
+    } else {
+      values[0] = PyTuple_GET_ITEM(__pyx_args, 0);
+      values[1] = PyTuple_GET_ITEM(__pyx_args, 1);
+      values[2] = PyTuple_GET_ITEM(__pyx_args, 2);
+      values[3] = PyTuple_GET_ITEM(__pyx_args, 3);
+      values[4] = PyTuple_GET_ITEM(__pyx_args, 4);
+      values[5] = PyTuple_GET_ITEM(__pyx_args, 5);
+      values[6] = PyTuple_GET_ITEM(__pyx_args, 6);
+      values[7] = PyTuple_GET_ITEM(__pyx_args, 7);
+      values[8] = PyTuple_GET_ITEM(__pyx_args, 8);
+    }
+    __pyx_v_filename = __pyx_convert_string_from_py_std__in_string(values[0]); if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 100, __pyx_L3_error)
+    __pyx_v_mtl_name = __pyx_convert_string_from_py_std__in_string(values[1]); if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 100, __pyx_L3_error)
+    __pyx_v_vertices = ((PyArrayObject *)values[2]);
+    __pyx_v_triangles = ((PyArrayObject *)values[3]);
+    __pyx_v_colors = ((PyArrayObject *)values[4]);
+    __pyx_v_uv_coords = ((PyArrayObject *)values[5]);
+    __pyx_v_nver = __Pyx_PyInt_As_int(values[6]); if (unlikely((__pyx_v_nver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+    __pyx_v_ntri = __Pyx_PyInt_As_int(values[7]); if (unlikely((__pyx_v_ntri == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+    __pyx_v_ntexver = __Pyx_PyInt_As_int(values[8]); if (unlikely((__pyx_v_ntexver == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 105, __pyx_L3_error)
+  }
+  goto __pyx_L4_argument_unpacking_done;
+  __pyx_L5_argtuple_error:;
+  __Pyx_RaiseArgtupleInvalid("write_obj_with_colors_texture_core", 1, 9, 9, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(0, 100, __pyx_L3_error)
+  __pyx_L3_error:;
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.write_obj_with_colors_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_RefNannyFinishContext();
+  return NULL;
+  __pyx_L4_argument_unpacking_done:;
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_vertices), __pyx_ptype_5numpy_ndarray, 0, "vertices", 0))) __PYX_ERR(0, 101, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_triangles), __pyx_ptype_5numpy_ndarray, 0, "triangles", 0))) __PYX_ERR(0, 102, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_colors), __pyx_ptype_5numpy_ndarray, 0, "colors", 0))) __PYX_ERR(0, 103, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_uv_coords), __pyx_ptype_5numpy_ndarray, 0, "uv_coords", 0))) __PYX_ERR(0, 104, __pyx_L1_error)
+  __pyx_r = __pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_8write_obj_with_colors_texture_core(__pyx_self, __pyx_v_filename, __pyx_v_mtl_name, __pyx_v_vertices, __pyx_v_triangles, __pyx_v_colors, __pyx_v_uv_coords, __pyx_v_nver, __pyx_v_ntri, __pyx_v_ntexver);
+
+  /* function exit code */
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __pyx_r = NULL;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyObject *__pyx_pf_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_8write_obj_with_colors_texture_core(CYTHON_UNUSED PyObject *__pyx_self, std::string __pyx_v_filename, std::string __pyx_v_mtl_name, PyArrayObject *__pyx_v_vertices, PyArrayObject *__pyx_v_triangles, PyArrayObject *__pyx_v_colors, PyArrayObject *__pyx_v_uv_coords, int __pyx_v_nver, int __pyx_v_ntri, int __pyx_v_ntexver) {
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_colors;
+  __Pyx_Buffer __pyx_pybuffer_colors;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_triangles;
+  __Pyx_Buffer __pyx_pybuffer_triangles;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_uv_coords;
+  __Pyx_Buffer __pyx_pybuffer_uv_coords;
+  __Pyx_LocalBuf_ND __pyx_pybuffernd_vertices;
+  __Pyx_Buffer __pyx_pybuffer_vertices;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("write_obj_with_colors_texture_core", 0);
+  __pyx_pybuffer_vertices.pybuffer.buf = NULL;
+  __pyx_pybuffer_vertices.refcount = 0;
+  __pyx_pybuffernd_vertices.data = NULL;
+  __pyx_pybuffernd_vertices.rcbuffer = &__pyx_pybuffer_vertices;
+  __pyx_pybuffer_triangles.pybuffer.buf = NULL;
+  __pyx_pybuffer_triangles.refcount = 0;
+  __pyx_pybuffernd_triangles.data = NULL;
+  __pyx_pybuffernd_triangles.rcbuffer = &__pyx_pybuffer_triangles;
+  __pyx_pybuffer_colors.pybuffer.buf = NULL;
+  __pyx_pybuffer_colors.refcount = 0;
+  __pyx_pybuffernd_colors.data = NULL;
+  __pyx_pybuffernd_colors.rcbuffer = &__pyx_pybuffer_colors;
+  __pyx_pybuffer_uv_coords.pybuffer.buf = NULL;
+  __pyx_pybuffer_uv_coords.refcount = 0;
+  __pyx_pybuffernd_uv_coords.data = NULL;
+  __pyx_pybuffernd_uv_coords.rcbuffer = &__pyx_pybuffer_uv_coords;
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer, (PyObject*)__pyx_v_vertices, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_vertices.diminfo[0].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_vertices.diminfo[0].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_vertices.diminfo[1].strides = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_vertices.diminfo[1].shape = __pyx_pybuffernd_vertices.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer, (PyObject*)__pyx_v_triangles, &__Pyx_TypeInfo_int, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_triangles.diminfo[0].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_triangles.diminfo[0].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_triangles.diminfo[1].strides = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_triangles.diminfo[1].shape = __pyx_pybuffernd_triangles.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_colors.rcbuffer->pybuffer, (PyObject*)__pyx_v_colors, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_colors.diminfo[0].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_colors.diminfo[0].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_colors.diminfo[1].strides = __pyx_pybuffernd_colors.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_colors.diminfo[1].shape = __pyx_pybuffernd_colors.rcbuffer->pybuffer.shape[1];
+  {
+    __Pyx_BufFmt_StackElem __pyx_stack[1];
+    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer, (PyObject*)__pyx_v_uv_coords, &__Pyx_TypeInfo_float, PyBUF_FORMAT| PyBUF_C_CONTIGUOUS, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  }
+  __pyx_pybuffernd_uv_coords.diminfo[0].strides = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_uv_coords.diminfo[0].shape = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_uv_coords.diminfo[1].strides = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_uv_coords.diminfo[1].shape = __pyx_pybuffernd_uv_coords.rcbuffer->pybuffer.shape[1];
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":107
+ *                 int nver, int ntri, int ntexver
+ *                 ):
+ *     _write_obj_with_colors_texture(filename, mtl_name,             # <<<<<<<<<<<<<<
+ *         <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles), <float*> np.PyArray_DATA(colors), <float*> np.PyArray_DATA(uv_coords),
+ *         nver, ntri, ntexver)
+ */
+  _write_obj_with_colors_texture(__pyx_v_filename, __pyx_v_mtl_name, ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_vertices))), ((int *)PyArray_DATA(((PyArrayObject *)__pyx_v_triangles))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_colors))), ((float *)PyArray_DATA(((PyArrayObject *)__pyx_v_uv_coords))), __pyx_v_nver, __pyx_v_ntri, __pyx_v_ntexver);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+
+  /* function exit code */
+  __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  { PyObject *__pyx_type, *__pyx_value, *__pyx_tb;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer);
+    __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);}
+  __Pyx_AddTraceback("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython.write_obj_with_colors_texture_core", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = NULL;
+  goto __pyx_L2;
+  __pyx_L0:;
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_colors.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_triangles.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_uv_coords.rcbuffer->pybuffer);
+  __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_vertices.rcbuffer->pybuffer);
+  __pyx_L2:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":734
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew1(PyObject *__pyx_v_a) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew1", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":735
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):
+ *     return PyArray_MultiIterNew(1, <void*>a)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(1, ((void *)__pyx_v_a)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 735, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":734
+ * ctypedef npy_cdouble     complex_t
+ * 
+ * cdef inline object PyArray_MultiIterNew1(a):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew1", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":737
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew2(PyObject *__pyx_v_a, PyObject *__pyx_v_b) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew2", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":738
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(2, ((void *)__pyx_v_a), ((void *)__pyx_v_b)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 738, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":737
+ *     return PyArray_MultiIterNew(1, <void*>a)
+ * 
+ * cdef inline object PyArray_MultiIterNew2(a, b):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew2", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":740
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew3(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew3", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":741
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(3, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 741, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":740
+ *     return PyArray_MultiIterNew(2, <void*>a, <void*>b)
+ * 
+ * cdef inline object PyArray_MultiIterNew3(a, b, c):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew3", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":743
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew4(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew4", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":744
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(4, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 744, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":743
+ *     return PyArray_MultiIterNew(3, <void*>a, <void*>b, <void*> c)
+ * 
+ * cdef inline object PyArray_MultiIterNew4(a, b, c, d):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew4", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":746
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew5(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d, PyObject *__pyx_v_e) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("PyArray_MultiIterNew5", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":747
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __pyx_t_1 = PyArray_MultiIterNew(5, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d), ((void *)__pyx_v_e)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 747, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_r = __pyx_t_1;
+  __pyx_t_1 = 0;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":746
+ *     return PyArray_MultiIterNew(4, <void*>a, <void*>b, <void*>c, <void*> d)
+ * 
+ * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e):             # <<<<<<<<<<<<<<
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_AddTraceback("numpy.PyArray_MultiIterNew5", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = 0;
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":749
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyDataType_SHAPE(PyArray_Descr *__pyx_v_d) {
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("PyDataType_SHAPE", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":750
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  __pyx_t_1 = (PyDataType_HASSUBARRAY(__pyx_v_d) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":751
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape             # <<<<<<<<<<<<<<
+ *     else:
+ *         return ()
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(((PyObject*)__pyx_v_d->subarray->shape));
+    __pyx_r = ((PyObject*)__pyx_v_d->subarray->shape);
+    goto __pyx_L0;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":750
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):
+ *     if PyDataType_HASSUBARRAY(d):             # <<<<<<<<<<<<<<
+ *         return <tuple>d.subarray.shape
+ *     else:
+ */
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":753
+ *         return <tuple>d.subarray.shape
+ *     else:
+ *         return ()             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  /*else*/ {
+    __Pyx_XDECREF(__pyx_r);
+    __Pyx_INCREF(__pyx_empty_tuple);
+    __pyx_r = __pyx_empty_tuple;
+    goto __pyx_L0;
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":749
+ *     return PyArray_MultiIterNew(5, <void*>a, <void*>b, <void*>c, <void*> d, <void*> e)
+ * 
+ * cdef inline tuple PyDataType_SHAPE(dtype d):             # <<<<<<<<<<<<<<
+ *     if PyDataType_HASSUBARRAY(d):
+ *         return <tuple>d.subarray.shape
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":928
+ *     int _import_umath() except -1
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)
+ */
+
+static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_arr, PyObject *__pyx_v_base) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("set_array_base", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":929
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *     Py_INCREF(base) # important to do this before stealing the reference below!             # <<<<<<<<<<<<<<
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ */
+  Py_INCREF(__pyx_v_base);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":930
+ * cdef inline void set_array_base(ndarray arr, object base):
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ */
+  (void)(PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_base));
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":928
+ *     int _import_umath() except -1
+ * 
+ * cdef inline void set_array_base(ndarray arr, object base):             # <<<<<<<<<<<<<<
+ *     Py_INCREF(base) # important to do this before stealing the reference below!
+ *     PyArray_SetBaseObject(arr, base)
+ */
+
+  /* function exit code */
+  __Pyx_RefNannyFinishContext();
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":932
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ */
+
+static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__pyx_v_arr) {
+  PyObject *__pyx_v_base;
+  PyObject *__pyx_r = NULL;
+  __Pyx_RefNannyDeclarations
+  int __pyx_t_1;
+  __Pyx_RefNannySetupContext("get_array_base", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":933
+ * 
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)             # <<<<<<<<<<<<<<
+ *     if base is NULL:
+ *         return None
+ */
+  __pyx_v_base = PyArray_BASE(__pyx_v_arr);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":934
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     return <object>base
+ */
+  __pyx_t_1 = ((__pyx_v_base == NULL) != 0);
+  if (__pyx_t_1) {
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":935
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ *         return None             # <<<<<<<<<<<<<<
+ *     return <object>base
+ * 
+ */
+    __Pyx_XDECREF(__pyx_r);
+    __pyx_r = Py_None; __Pyx_INCREF(Py_None);
+    goto __pyx_L0;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":934
+ * cdef inline object get_array_base(ndarray arr):
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:             # <<<<<<<<<<<<<<
+ *         return None
+ *     return <object>base
+ */
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":936
+ *     if base is NULL:
+ *         return None
+ *     return <object>base             # <<<<<<<<<<<<<<
+ * 
+ * # Versions of the import_* functions which are more suitable for
+ */
+  __Pyx_XDECREF(__pyx_r);
+  __Pyx_INCREF(((PyObject *)__pyx_v_base));
+  __pyx_r = ((PyObject *)__pyx_v_base);
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":932
+ *     PyArray_SetBaseObject(arr, base)
+ * 
+ * cdef inline object get_array_base(ndarray arr):             # <<<<<<<<<<<<<<
+ *     base = PyArray_BASE(arr)
+ *     if base is NULL:
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_XGIVEREF(__pyx_r);
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":940
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         __pyx_import_array()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_array(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("import_array", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":941
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         __pyx_import_array()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":942
+ * cdef inline int import_array() except -1:
+ *     try:
+ *         __pyx_import_array()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ */
+      __pyx_t_4 = _import_array(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 942, __pyx_L3_error)
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":941
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         __pyx_import_array()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":943
+ *     try:
+ *         __pyx_import_array()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 943, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":944
+ *         __pyx_import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple_, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 944, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 944, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":941
+ * # Cython code.
+ * cdef inline int import_array() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         __pyx_import_array()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":940
+ * # Versions of the import_* functions which are more suitable for
+ * # Cython code.
+ * cdef inline int import_array() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         __pyx_import_array()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":946
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_umath(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("import_umath", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":947
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":948
+ * cdef inline int import_umath() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 948, __pyx_L3_error)
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":947
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":949
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 949, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":950
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__2, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 950, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 950, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":947
+ * 
+ * cdef inline int import_umath() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":946
+ *         raise ImportError("numpy.core.multiarray failed to import")
+ * 
+ * cdef inline int import_umath() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":952
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_import_ufunc(void) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  PyObject *__pyx_t_2 = NULL;
+  PyObject *__pyx_t_3 = NULL;
+  int __pyx_t_4;
+  PyObject *__pyx_t_5 = NULL;
+  PyObject *__pyx_t_6 = NULL;
+  PyObject *__pyx_t_7 = NULL;
+  PyObject *__pyx_t_8 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("import_ufunc", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":953
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+  {
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3);
+    __Pyx_XGOTREF(__pyx_t_1);
+    __Pyx_XGOTREF(__pyx_t_2);
+    __Pyx_XGOTREF(__pyx_t_3);
+    /*try:*/ {
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":954
+ * cdef inline int import_ufunc() except -1:
+ *     try:
+ *         _import_umath()             # <<<<<<<<<<<<<<
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")
+ */
+      __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 954, __pyx_L3_error)
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":953
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    }
+    __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0;
+    __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
+    __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0;
+    goto __pyx_L8_try_end;
+    __pyx_L3_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":955
+ *     try:
+ *         _import_umath()
+ *     except Exception:             # <<<<<<<<<<<<<<
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ */
+    __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0])));
+    if (__pyx_t_4) {
+      __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+      if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(1, 955, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_5);
+      __Pyx_GOTREF(__pyx_t_6);
+      __Pyx_GOTREF(__pyx_t_7);
+
+      /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":956
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef extern from *:
+ */
+      __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__2, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 956, __pyx_L5_except_error)
+      __Pyx_GOTREF(__pyx_t_8);
+      __Pyx_Raise(__pyx_t_8, 0, 0, 0);
+      __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
+      __PYX_ERR(1, 956, __pyx_L5_except_error)
+    }
+    goto __pyx_L5_except_error;
+    __pyx_L5_except_error:;
+
+    /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":953
+ * 
+ * cdef inline int import_ufunc() except -1:
+ *     try:             # <<<<<<<<<<<<<<
+ *         _import_umath()
+ *     except Exception:
+ */
+    __Pyx_XGIVEREF(__pyx_t_1);
+    __Pyx_XGIVEREF(__pyx_t_2);
+    __Pyx_XGIVEREF(__pyx_t_3);
+    __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3);
+    goto __pyx_L1_error;
+    __pyx_L8_try_end:;
+  }
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":952
+ *         raise ImportError("numpy.core.umath failed to import")
+ * 
+ * cdef inline int import_ufunc() except -1:             # <<<<<<<<<<<<<<
+ *     try:
+ *         _import_umath()
+ */
+
+  /* function exit code */
+  __pyx_r = 0;
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_5);
+  __Pyx_XDECREF(__pyx_t_6);
+  __Pyx_XDECREF(__pyx_t_7);
+  __Pyx_XDECREF(__pyx_t_8);
+  __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __pyx_r = -1;
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":966
+ * 
+ * 
+ * cdef inline bint is_timedelta64_object(object obj):             # <<<<<<<<<<<<<<
+ *     """
+ *     Cython equivalent of `isinstance(obj, np.timedelta64)`
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_is_timedelta64_object(PyObject *__pyx_v_obj) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("is_timedelta64_object", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":978
+ *     bool
+ *     """
+ *     return PyObject_TypeCheck(obj, &PyTimedeltaArrType_Type)             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyTimedeltaArrType_Type));
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":966
+ * 
+ * 
+ * cdef inline bint is_timedelta64_object(object obj):             # <<<<<<<<<<<<<<
+ *     """
+ *     Cython equivalent of `isinstance(obj, np.timedelta64)`
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":981
+ * 
+ * 
+ * cdef inline bint is_datetime64_object(object obj):             # <<<<<<<<<<<<<<
+ *     """
+ *     Cython equivalent of `isinstance(obj, np.datetime64)`
+ */
+
+static CYTHON_INLINE int __pyx_f_5numpy_is_datetime64_object(PyObject *__pyx_v_obj) {
+  int __pyx_r;
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("is_datetime64_object", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":993
+ *     bool
+ *     """
+ *     return PyObject_TypeCheck(obj, &PyDatetimeArrType_Type)             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyDatetimeArrType_Type));
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":981
+ * 
+ * 
+ * cdef inline bint is_datetime64_object(object obj):             # <<<<<<<<<<<<<<
+ *     """
+ *     Cython equivalent of `isinstance(obj, np.datetime64)`
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":996
+ * 
+ * 
+ * cdef inline npy_datetime get_datetime64_value(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the int64 value underlying scalar numpy datetime64 object
+ */
+
+static CYTHON_INLINE npy_datetime __pyx_f_5numpy_get_datetime64_value(PyObject *__pyx_v_obj) {
+  npy_datetime __pyx_r;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1003
+ *     also needed.  That can be found using `get_datetime64_unit`.
+ *     """
+ *     return (<PyDatetimeScalarObject*>obj).obval             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = ((PyDatetimeScalarObject *)__pyx_v_obj)->obval;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":996
+ * 
+ * 
+ * cdef inline npy_datetime get_datetime64_value(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the int64 value underlying scalar numpy datetime64 object
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1006
+ * 
+ * 
+ * cdef inline npy_timedelta get_timedelta64_value(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the int64 value underlying scalar numpy timedelta64 object
+ */
+
+static CYTHON_INLINE npy_timedelta __pyx_f_5numpy_get_timedelta64_value(PyObject *__pyx_v_obj) {
+  npy_timedelta __pyx_r;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1010
+ *     returns the int64 value underlying scalar numpy timedelta64 object
+ *     """
+ *     return (<PyTimedeltaScalarObject*>obj).obval             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = ((PyTimedeltaScalarObject *)__pyx_v_obj)->obval;
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1006
+ * 
+ * 
+ * cdef inline npy_timedelta get_timedelta64_value(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the int64 value underlying scalar numpy timedelta64 object
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  return __pyx_r;
+}
+
+/* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1013
+ * 
+ * 
+ * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the unit part of the dtype for a numpy datetime64 object.
+ */
+
+static CYTHON_INLINE NPY_DATETIMEUNIT __pyx_f_5numpy_get_datetime64_unit(PyObject *__pyx_v_obj) {
+  NPY_DATETIMEUNIT __pyx_r;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1017
+ *     returns the unit part of the dtype for a numpy datetime64 object.
+ *     """
+ *     return <NPY_DATETIMEUNIT>(<PyDatetimeScalarObject*>obj).obmeta.base             # <<<<<<<<<<<<<<
+ */
+  __pyx_r = ((NPY_DATETIMEUNIT)((PyDatetimeScalarObject *)__pyx_v_obj)->obmeta.base);
+  goto __pyx_L0;
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":1013
+ * 
+ * 
+ * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) nogil:             # <<<<<<<<<<<<<<
+ *     """
+ *     returns the unit part of the dtype for a numpy datetime64 object.
+ */
+
+  /* function exit code */
+  __pyx_L0:;
+  return __pyx_r;
+}
+
+/* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length = 0
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+static std::string __pyx_convert_string_from_py_std__in_string(PyObject *__pyx_v_o) {
+  Py_ssize_t __pyx_v_length;
+  char const *__pyx_v_data;
+  std::string __pyx_r;
+  __Pyx_RefNannyDeclarations
+  char const *__pyx_t_1;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("__pyx_convert_string_from_py_std__in_string", 0);
+
+  /* "string.from_py":14
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:
+ *     cdef Py_ssize_t length = 0             # <<<<<<<<<<<<<<
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ *     return string(data, length)
+ */
+  __pyx_v_length = 0;
+
+  /* "string.from_py":15
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:
+ *     cdef Py_ssize_t length = 0
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)             # <<<<<<<<<<<<<<
+ *     return string(data, length)
+ * 
+ */
+  __pyx_t_1 = __Pyx_PyObject_AsStringAndSize(__pyx_v_o, (&__pyx_v_length)); if (unlikely(__pyx_t_1 == ((char const *)NULL))) __PYX_ERR(2, 15, __pyx_L1_error)
+  __pyx_v_data = __pyx_t_1;
+
+  /* "string.from_py":16
+ *     cdef Py_ssize_t length = 0
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ *     return string(data, length)             # <<<<<<<<<<<<<<
+ * 
+ * 
+ */
+  __pyx_r = std::string(__pyx_v_data, __pyx_v_length);
+  goto __pyx_L0;
+
+  /* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length = 0
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+  /* function exit code */
+  __pyx_L1_error:;
+  __Pyx_AddTraceback("string.from_py.__pyx_convert_string_from_py_std__in_string", __pyx_clineno, __pyx_lineno, __pyx_filename);
+  __Pyx_pretend_to_initialize(&__pyx_r);
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  return __pyx_r;
+}
+
+static PyMethodDef __pyx_methods[] = {
+  {0, 0, 0, 0}
+};
+
+#if PY_MAJOR_VERSION >= 3
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/
+static int __pyx_pymod_exec_mesh_core_cython(PyObject* module); /*proto*/
+static PyModuleDef_Slot __pyx_moduledef_slots[] = {
+  {Py_mod_create, (void*)__pyx_pymod_create},
+  {Py_mod_exec, (void*)__pyx_pymod_exec_mesh_core_cython},
+  {0, NULL}
+};
+#endif
+
+static struct PyModuleDef __pyx_moduledef = {
+    PyModuleDef_HEAD_INIT,
+    "mesh_core_cython",
+    0, /* m_doc */
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    0, /* m_size */
+  #else
+    -1, /* m_size */
+  #endif
+    __pyx_methods /* m_methods */,
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+    __pyx_moduledef_slots, /* m_slots */
+  #else
+    NULL, /* m_reload */
+  #endif
+    NULL, /* m_traverse */
+    NULL, /* m_clear */
+    NULL /* m_free */
+};
+#endif
+#ifndef CYTHON_SMALL_CODE
+#if defined(__clang__)
+    #define CYTHON_SMALL_CODE
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+    #define CYTHON_SMALL_CODE __attribute__((cold))
+#else
+    #define CYTHON_SMALL_CODE
+#endif
+#endif
+
+static __Pyx_StringTabEntry __pyx_string_tab[] = {
+  {&__pyx_n_s_ImportError, __pyx_k_ImportError, sizeof(__pyx_k_ImportError), 0, 0, 1, 1},
+  {&__pyx_n_s_barycentric_weight, __pyx_k_barycentric_weight, sizeof(__pyx_k_barycentric_weight), 0, 0, 1, 1},
+  {&__pyx_n_s_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 0, 1, 1},
+  {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1},
+  {&__pyx_n_s_colors, __pyx_k_colors, sizeof(__pyx_k_colors), 0, 0, 1, 1},
+  {&__pyx_n_s_depth_buffer, __pyx_k_depth_buffer, sizeof(__pyx_k_depth_buffer), 0, 0, 1, 1},
+  {&__pyx_n_s_filename, __pyx_k_filename, sizeof(__pyx_k_filename), 0, 0, 1, 1},
+  {&__pyx_n_s_get_normal_core, __pyx_k_get_normal_core, sizeof(__pyx_k_get_normal_core), 0, 0, 1, 1},
+  {&__pyx_n_s_h, __pyx_k_h, sizeof(__pyx_k_h), 0, 0, 1, 1},
+  {&__pyx_n_s_image, __pyx_k_image, sizeof(__pyx_k_image), 0, 0, 1, 1},
+  {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1},
+  {&__pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_k_insightface_thirdparty_face3d_me, sizeof(__pyx_k_insightface_thirdparty_face3d_me), 0, 0, 1, 0},
+  {&__pyx_n_s_insightface_thirdparty_face3d_me_2, __pyx_k_insightface_thirdparty_face3d_me_2, sizeof(__pyx_k_insightface_thirdparty_face3d_me_2), 0, 0, 1, 1},
+  {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1},
+  {&__pyx_n_s_mapping_type, __pyx_k_mapping_type, sizeof(__pyx_k_mapping_type), 0, 0, 1, 1},
+  {&__pyx_n_s_mtl_name, __pyx_k_mtl_name, sizeof(__pyx_k_mtl_name), 0, 0, 1, 1},
+  {&__pyx_n_s_name, __pyx_k_name, sizeof(__pyx_k_name), 0, 0, 1, 1},
+  {&__pyx_n_s_normal, __pyx_k_normal, sizeof(__pyx_k_normal), 0, 0, 1, 1},
+  {&__pyx_n_s_np, __pyx_k_np, sizeof(__pyx_k_np), 0, 0, 1, 1},
+  {&__pyx_n_s_ntexver, __pyx_k_ntexver, sizeof(__pyx_k_ntexver), 0, 0, 1, 1},
+  {&__pyx_n_s_ntri, __pyx_k_ntri, sizeof(__pyx_k_ntri), 0, 0, 1, 1},
+  {&__pyx_n_s_numpy, __pyx_k_numpy, sizeof(__pyx_k_numpy), 0, 0, 1, 1},
+  {&__pyx_kp_s_numpy_core_multiarray_failed_to, __pyx_k_numpy_core_multiarray_failed_to, sizeof(__pyx_k_numpy_core_multiarray_failed_to), 0, 0, 1, 0},
+  {&__pyx_kp_s_numpy_core_umath_failed_to_impor, __pyx_k_numpy_core_umath_failed_to_impor, sizeof(__pyx_k_numpy_core_umath_failed_to_impor), 0, 0, 1, 0},
+  {&__pyx_n_s_nver, __pyx_k_nver, sizeof(__pyx_k_nver), 0, 0, 1, 1},
+  {&__pyx_n_s_rasterize_triangles_core, __pyx_k_rasterize_triangles_core, sizeof(__pyx_k_rasterize_triangles_core), 0, 0, 1, 1},
+  {&__pyx_n_s_render_colors_core, __pyx_k_render_colors_core, sizeof(__pyx_k_render_colors_core), 0, 0, 1, 1},
+  {&__pyx_n_s_render_texture_core, __pyx_k_render_texture_core, sizeof(__pyx_k_render_texture_core), 0, 0, 1, 1},
+  {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_c, __pyx_k_tex_c, sizeof(__pyx_k_tex_c), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_coords, __pyx_k_tex_coords, sizeof(__pyx_k_tex_coords), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_h, __pyx_k_tex_h, sizeof(__pyx_k_tex_h), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_nver, __pyx_k_tex_nver, sizeof(__pyx_k_tex_nver), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_triangles, __pyx_k_tex_triangles, sizeof(__pyx_k_tex_triangles), 0, 0, 1, 1},
+  {&__pyx_n_s_tex_w, __pyx_k_tex_w, sizeof(__pyx_k_tex_w), 0, 0, 1, 1},
+  {&__pyx_n_s_texture, __pyx_k_texture, sizeof(__pyx_k_texture), 0, 0, 1, 1},
+  {&__pyx_n_s_tri_normal, __pyx_k_tri_normal, sizeof(__pyx_k_tri_normal), 0, 0, 1, 1},
+  {&__pyx_n_s_triangle_buffer, __pyx_k_triangle_buffer, sizeof(__pyx_k_triangle_buffer), 0, 0, 1, 1},
+  {&__pyx_n_s_triangles, __pyx_k_triangles, sizeof(__pyx_k_triangles), 0, 0, 1, 1},
+  {&__pyx_n_s_uv_coords, __pyx_k_uv_coords, sizeof(__pyx_k_uv_coords), 0, 0, 1, 1},
+  {&__pyx_n_s_vertices, __pyx_k_vertices, sizeof(__pyx_k_vertices), 0, 0, 1, 1},
+  {&__pyx_n_s_w, __pyx_k_w, sizeof(__pyx_k_w), 0, 0, 1, 1},
+  {&__pyx_n_s_write_obj_with_colors_texture_co, __pyx_k_write_obj_with_colors_texture_co, sizeof(__pyx_k_write_obj_with_colors_texture_co), 0, 0, 1, 1},
+  {0, 0, 0, 0, 0, 0, 0}
+};
+static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) {
+  __pyx_builtin_ImportError = __Pyx_GetBuiltinName(__pyx_n_s_ImportError); if (!__pyx_builtin_ImportError) __PYX_ERR(1, 944, __pyx_L1_error)
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":944
+ *         __pyx_import_array()
+ *     except Exception:
+ *         raise ImportError("numpy.core.multiarray failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_umath() except -1:
+ */
+  __pyx_tuple_ = PyTuple_Pack(1, __pyx_kp_s_numpy_core_multiarray_failed_to); if (unlikely(!__pyx_tuple_)) __PYX_ERR(1, 944, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple_);
+  __Pyx_GIVEREF(__pyx_tuple_);
+
+  /* "../../../../tmp/pip-build-env-cbwohp4_/overlay/lib/python3.8/site-packages/numpy/__init__.pxd":950
+ *         _import_umath()
+ *     except Exception:
+ *         raise ImportError("numpy.core.umath failed to import")             # <<<<<<<<<<<<<<
+ * 
+ * cdef inline int import_ufunc() except -1:
+ */
+  __pyx_tuple__2 = PyTuple_Pack(1, __pyx_kp_s_numpy_core_umath_failed_to_impor); if (unlikely(!__pyx_tuple__2)) __PYX_ERR(1, 950, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__2);
+  __Pyx_GIVEREF(__pyx_tuple__2);
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__3 = PyTuple_Pack(4, __pyx_n_s_normal, __pyx_n_s_tri_normal, __pyx_n_s_triangles, __pyx_n_s_ntri); if (unlikely(!__pyx_tuple__3)) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__3);
+  __Pyx_GIVEREF(__pyx_tuple__3);
+  __pyx_codeobj__4 = (PyObject*)__Pyx_PyCode_New(4, 0, 4, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__3, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_get_normal_core, 40, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__4)) __PYX_ERR(0, 40, __pyx_L1_error)
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__5 = PyTuple_Pack(9, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_depth_buffer, __pyx_n_s_triangle_buffer, __pyx_n_s_barycentric_weight, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w); if (unlikely(!__pyx_tuple__5)) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__5);
+  __Pyx_GIVEREF(__pyx_tuple__5);
+  __pyx_codeobj__6 = (PyObject*)__Pyx_PyCode_New(9, 0, 9, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__5, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_rasterize_triangles_core, 49, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__6)) __PYX_ERR(0, 49, __pyx_L1_error)
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__7 = PyTuple_Pack(10, __pyx_n_s_image, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_colors, __pyx_n_s_depth_buffer, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w, __pyx_n_s_c); if (unlikely(!__pyx_tuple__7)) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__7);
+  __Pyx_GIVEREF(__pyx_tuple__7);
+  __pyx_codeobj__8 = (PyObject*)__Pyx_PyCode_New(10, 0, 10, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__7, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_render_colors_core, 64, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__8)) __PYX_ERR(0, 64, __pyx_L1_error)
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__9 = PyTuple_Pack(17, __pyx_n_s_image, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_texture, __pyx_n_s_tex_coords, __pyx_n_s_tex_triangles, __pyx_n_s_depth_buffer, __pyx_n_s_nver, __pyx_n_s_tex_nver, __pyx_n_s_ntri, __pyx_n_s_h, __pyx_n_s_w, __pyx_n_s_c, __pyx_n_s_tex_h, __pyx_n_s_tex_w, __pyx_n_s_tex_c, __pyx_n_s_mapping_type); if (unlikely(!__pyx_tuple__9)) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__9);
+  __Pyx_GIVEREF(__pyx_tuple__9);
+  __pyx_codeobj__10 = (PyObject*)__Pyx_PyCode_New(17, 0, 17, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__9, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_render_texture_core, 79, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__10)) __PYX_ERR(0, 79, __pyx_L1_error)
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_tuple__11 = PyTuple_Pack(9, __pyx_n_s_filename, __pyx_n_s_mtl_name, __pyx_n_s_vertices, __pyx_n_s_triangles, __pyx_n_s_colors, __pyx_n_s_uv_coords, __pyx_n_s_nver, __pyx_n_s_ntri, __pyx_n_s_ntexver); if (unlikely(!__pyx_tuple__11)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_tuple__11);
+  __Pyx_GIVEREF(__pyx_tuple__11);
+  __pyx_codeobj__12 = (PyObject*)__Pyx_PyCode_New(9, 0, 9, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__11, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_insightface_thirdparty_face3d_me, __pyx_n_s_write_obj_with_colors_texture_co, 100, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__12)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_InitGlobals(void) {
+  if (__Pyx_InitStrings(__pyx_string_tab) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  return 0;
+  __pyx_L1_error:;
+  return -1;
+}
+
+static CYTHON_SMALL_CODE int __Pyx_modinit_global_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_export_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_init_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_type_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_variable_import_code(void); /*proto*/
+static CYTHON_SMALL_CODE int __Pyx_modinit_function_import_code(void); /*proto*/
+
+static int __Pyx_modinit_global_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0);
+  /*--- Global init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_variable_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0);
+  /*--- Variable export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_export_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0);
+  /*--- Function export code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_init_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0);
+  /*--- Type init code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_type_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0);
+  /*--- Type import code ---*/
+  __pyx_t_1 = PyImport_ImportModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_t_1)) __PYX_ERR(3, 9, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_ptype_7cpython_4type_type = __Pyx_ImportType(__pyx_t_1, __Pyx_BUILTIN_MODULE_NAME, "type", 
+  #if defined(PYPY_VERSION_NUM) && PYPY_VERSION_NUM < 0x050B0000
+  sizeof(PyTypeObject),
+  #else
+  sizeof(PyHeapTypeObject),
+  #endif
+  __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_7cpython_4type_type) __PYX_ERR(3, 9, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __pyx_t_1 = PyImport_ImportModule("numpy"); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 199, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  __pyx_ptype_5numpy_dtype = __Pyx_ImportType(__pyx_t_1, "numpy", "dtype", sizeof(PyArray_Descr), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_dtype) __PYX_ERR(1, 199, __pyx_L1_error)
+  __pyx_ptype_5numpy_flatiter = __Pyx_ImportType(__pyx_t_1, "numpy", "flatiter", sizeof(PyArrayIterObject), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_flatiter) __PYX_ERR(1, 222, __pyx_L1_error)
+  __pyx_ptype_5numpy_broadcast = __Pyx_ImportType(__pyx_t_1, "numpy", "broadcast", sizeof(PyArrayMultiIterObject), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_broadcast) __PYX_ERR(1, 226, __pyx_L1_error)
+  __pyx_ptype_5numpy_ndarray = __Pyx_ImportType(__pyx_t_1, "numpy", "ndarray", sizeof(PyArrayObject), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_ndarray) __PYX_ERR(1, 238, __pyx_L1_error)
+  __pyx_ptype_5numpy_generic = __Pyx_ImportType(__pyx_t_1, "numpy", "generic", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_generic) __PYX_ERR(1, 770, __pyx_L1_error)
+  __pyx_ptype_5numpy_number = __Pyx_ImportType(__pyx_t_1, "numpy", "number", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_number) __PYX_ERR(1, 772, __pyx_L1_error)
+  __pyx_ptype_5numpy_integer = __Pyx_ImportType(__pyx_t_1, "numpy", "integer", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_integer) __PYX_ERR(1, 774, __pyx_L1_error)
+  __pyx_ptype_5numpy_signedinteger = __Pyx_ImportType(__pyx_t_1, "numpy", "signedinteger", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_signedinteger) __PYX_ERR(1, 776, __pyx_L1_error)
+  __pyx_ptype_5numpy_unsignedinteger = __Pyx_ImportType(__pyx_t_1, "numpy", "unsignedinteger", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_unsignedinteger) __PYX_ERR(1, 778, __pyx_L1_error)
+  __pyx_ptype_5numpy_inexact = __Pyx_ImportType(__pyx_t_1, "numpy", "inexact", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_inexact) __PYX_ERR(1, 780, __pyx_L1_error)
+  __pyx_ptype_5numpy_floating = __Pyx_ImportType(__pyx_t_1, "numpy", "floating", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_floating) __PYX_ERR(1, 782, __pyx_L1_error)
+  __pyx_ptype_5numpy_complexfloating = __Pyx_ImportType(__pyx_t_1, "numpy", "complexfloating", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_complexfloating) __PYX_ERR(1, 784, __pyx_L1_error)
+  __pyx_ptype_5numpy_flexible = __Pyx_ImportType(__pyx_t_1, "numpy", "flexible", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_flexible) __PYX_ERR(1, 786, __pyx_L1_error)
+  __pyx_ptype_5numpy_character = __Pyx_ImportType(__pyx_t_1, "numpy", "character", sizeof(PyObject), __Pyx_ImportType_CheckSize_Warn);
+   if (!__pyx_ptype_5numpy_character) __PYX_ERR(1, 788, __pyx_L1_error)
+  __pyx_ptype_5numpy_ufunc = __Pyx_ImportType(__pyx_t_1, "numpy", "ufunc", sizeof(PyUFuncObject), __Pyx_ImportType_CheckSize_Ignore);
+   if (!__pyx_ptype_5numpy_ufunc) __PYX_ERR(1, 826, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+  __Pyx_RefNannyFinishContext();
+  return 0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  __Pyx_RefNannyFinishContext();
+  return -1;
+}
+
+static int __Pyx_modinit_variable_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0);
+  /*--- Variable import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+static int __Pyx_modinit_function_import_code(void) {
+  __Pyx_RefNannyDeclarations
+  __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0);
+  /*--- Function import code ---*/
+  __Pyx_RefNannyFinishContext();
+  return 0;
+}
+
+
+#ifndef CYTHON_NO_PYINIT_EXPORT
+#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC
+#elif PY_MAJOR_VERSION < 3
+#ifdef __cplusplus
+#define __Pyx_PyMODINIT_FUNC extern "C" void
+#else
+#define __Pyx_PyMODINIT_FUNC void
+#endif
+#else
+#ifdef __cplusplus
+#define __Pyx_PyMODINIT_FUNC extern "C" PyObject *
+#else
+#define __Pyx_PyMODINIT_FUNC PyObject *
+#endif
+#endif
+
+
+#if PY_MAJOR_VERSION < 3
+__Pyx_PyMODINIT_FUNC initmesh_core_cython(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC initmesh_core_cython(void)
+#else
+__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void) CYTHON_SMALL_CODE; /*proto*/
+__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void)
+#if CYTHON_PEP489_MULTI_PHASE_INIT
+{
+  return PyModuleDef_Init(&__pyx_moduledef);
+}
+static CYTHON_SMALL_CODE int __Pyx_check_single_interpreter(void) {
+    #if PY_VERSION_HEX >= 0x030700A1
+    static PY_INT64_T main_interpreter_id = -1;
+    PY_INT64_T current_id = PyInterpreterState_GetID(PyThreadState_Get()->interp);
+    if (main_interpreter_id == -1) {
+        main_interpreter_id = current_id;
+        return (unlikely(current_id == -1)) ? -1 : 0;
+    } else if (unlikely(main_interpreter_id != current_id))
+    #else
+    static PyInterpreterState *main_interpreter = NULL;
+    PyInterpreterState *current_interpreter = PyThreadState_Get()->interp;
+    if (!main_interpreter) {
+        main_interpreter = current_interpreter;
+    } else if (unlikely(main_interpreter != current_interpreter))
+    #endif
+    {
+        PyErr_SetString(
+            PyExc_ImportError,
+            "Interpreter change detected - this module can only be loaded into one interpreter per process.");
+        return -1;
+    }
+    return 0;
+}
+static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name, int allow_none) {
+    PyObject *value = PyObject_GetAttrString(spec, from_name);
+    int result = 0;
+    if (likely(value)) {
+        if (allow_none || value != Py_None) {
+            result = PyDict_SetItemString(moddict, to_name, value);
+        }
+        Py_DECREF(value);
+    } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) {
+        PyErr_Clear();
+    } else {
+        result = -1;
+    }
+    return result;
+}
+static CYTHON_SMALL_CODE PyObject* __pyx_pymod_create(PyObject *spec, CYTHON_UNUSED PyModuleDef *def) {
+    PyObject *module = NULL, *moddict, *modname;
+    if (__Pyx_check_single_interpreter())
+        return NULL;
+    if (__pyx_m)
+        return __Pyx_NewRef(__pyx_m);
+    modname = PyObject_GetAttrString(spec, "name");
+    if (unlikely(!modname)) goto bad;
+    module = PyModule_NewObject(modname);
+    Py_DECREF(modname);
+    if (unlikely(!module)) goto bad;
+    moddict = PyModule_GetDict(module);
+    if (unlikely(!moddict)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__", 1) < 0)) goto bad;
+    if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__", 0) < 0)) goto bad;
+    return module;
+bad:
+    Py_XDECREF(module);
+    return NULL;
+}
+
+
+static CYTHON_SMALL_CODE int __pyx_pymod_exec_mesh_core_cython(PyObject *__pyx_pyinit_module)
+#endif
+#endif
+{
+  PyObject *__pyx_t_1 = NULL;
+  int __pyx_t_2;
+  int __pyx_lineno = 0;
+  const char *__pyx_filename = NULL;
+  int __pyx_clineno = 0;
+  __Pyx_RefNannyDeclarations
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  if (__pyx_m) {
+    if (__pyx_m == __pyx_pyinit_module) return 0;
+    PyErr_SetString(PyExc_RuntimeError, "Module 'mesh_core_cython' has already been imported. Re-initialisation is not supported.");
+    return -1;
+  }
+  #elif PY_MAJOR_VERSION >= 3
+  if (__pyx_m) return __Pyx_NewRef(__pyx_m);
+  #endif
+  #if CYTHON_REFNANNY
+__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny");
+if (!__Pyx_RefNanny) {
+  PyErr_Clear();
+  __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny");
+  if (!__Pyx_RefNanny)
+      Py_FatalError("failed to import 'refnanny' module");
+}
+#endif
+  __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_mesh_core_cython(void)", 0);
+  if (__Pyx_check_binary_version() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pxy_PyFrame_Initialize_Offsets
+  __Pxy_PyFrame_Initialize_Offsets();
+  #endif
+  __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #ifdef __Pyx_CyFunction_USED
+  if (__pyx_CyFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_FusedFunction_USED
+  if (__pyx_FusedFunction_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Coroutine_USED
+  if (__pyx_Coroutine_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_Generator_USED
+  if (__pyx_Generator_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_AsyncGen_USED
+  if (__pyx_AsyncGen_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  #ifdef __Pyx_StopAsyncIteration_USED
+  if (__pyx_StopAsyncIteration_init() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  /*--- Library function declarations ---*/
+  /*--- Threads initialization code ---*/
+  #if defined(WITH_THREAD) && PY_VERSION_HEX < 0x030700F0 && defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS
+  PyEval_InitThreads();
+  #endif
+  /*--- Module creation code ---*/
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  __pyx_m = __pyx_pyinit_module;
+  Py_INCREF(__pyx_m);
+  #else
+  #if PY_MAJOR_VERSION < 3
+  __pyx_m = Py_InitModule4("mesh_core_cython", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m);
+  #else
+  __pyx_m = PyModule_Create(&__pyx_moduledef);
+  #endif
+  if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_d);
+  __pyx_b = PyImport_AddModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_b);
+  __pyx_cython_runtime = PyImport_AddModule((char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error)
+  Py_INCREF(__pyx_cython_runtime);
+  if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error);
+  /*--- Initialize various global constants etc. ---*/
+  if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT)
+  if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+  if (__pyx_module_is_main_insightface__thirdparty__face3d__mesh__cython__mesh_core_cython) {
+    if (PyObject_SetAttr(__pyx_m, __pyx_n_s_name, __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  }
+  #if PY_MAJOR_VERSION >= 3
+  {
+    PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error)
+    if (!PyDict_GetItemString(modules, "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython")) {
+      if (unlikely(PyDict_SetItemString(modules, "insightface.thirdparty.face3d.mesh.cython.mesh_core_cython", __pyx_m) < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+    }
+  }
+  #endif
+  /*--- Builtin init code ---*/
+  if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Constants init code ---*/
+  if (__Pyx_InitCachedConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  /*--- Global type/function init code ---*/
+  (void)__Pyx_modinit_global_init_code();
+  (void)__Pyx_modinit_variable_export_code();
+  (void)__Pyx_modinit_function_export_code();
+  (void)__Pyx_modinit_type_init_code();
+  if (unlikely(__Pyx_modinit_type_import_code() < 0)) __PYX_ERR(0, 1, __pyx_L1_error)
+  (void)__Pyx_modinit_variable_import_code();
+  (void)__Pyx_modinit_function_import_code();
+  /*--- Execution code ---*/
+  #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED)
+  if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  #endif
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":1
+ * import numpy as np             # <<<<<<<<<<<<<<
+ * cimport numpy as np
+ * from libcpp.string cimport string
+ */
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_numpy, 0, -1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_np, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":6
+ * 
+ * # use the Numpy-C-API from Cython
+ * np.import_array()             # <<<<<<<<<<<<<<
+ * 
+ * # cdefine the signature of our c function
+ */
+  __pyx_t_2 = __pyx_f_5numpy_import_array(); if (unlikely(__pyx_t_2 == ((int)-1))) __PYX_ERR(0, 6, __pyx_L1_error)
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":40
+ *         int nver, int ntri, int ntexver)
+ * 
+ * def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] tri_normal not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_1get_normal_core, NULL, __pyx_n_s_insightface_thirdparty_face3d_me_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_get_normal_core, __pyx_t_1) < 0) __PYX_ERR(0, 40, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":49
+ *         ntri)
+ * 
+ * def rasterize_triangles_core(             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_3rasterize_triangles_core, NULL, __pyx_n_s_insightface_thirdparty_face3d_me_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_rasterize_triangles_core, __pyx_t_1) < 0) __PYX_ERR(0, 49, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":64
+ *         h, w)
+ * 
+ * def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_5render_colors_core, NULL, __pyx_n_s_insightface_thirdparty_face3d_me_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_render_colors_core, __pyx_t_1) < 0) __PYX_ERR(0, 64, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":79
+ *         h, w, c)
+ * 
+ * def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_7render_texture_core, NULL, __pyx_n_s_insightface_thirdparty_face3d_me_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_render_texture_core, __pyx_t_1) < 0) __PYX_ERR(0, 79, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":100
+ *         mapping_type)
+ * 
+ * def write_obj_with_colors_texture_core(string filename, string mtl_name,             # <<<<<<<<<<<<<<
+ *                 np.ndarray[float, ndim=2, mode = "c"] vertices not None,
+ *                 np.ndarray[int, ndim=2, mode="c"] triangles not None,
+ */
+  __pyx_t_1 = PyCFunction_NewEx(&__pyx_mdef_11insightface_10thirdparty_6face3d_4mesh_6cython_16mesh_core_cython_9write_obj_with_colors_texture_core, NULL, __pyx_n_s_insightface_thirdparty_face3d_me_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_write_obj_with_colors_texture_co, __pyx_t_1) < 0) __PYX_ERR(0, 100, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx":1
+ * import numpy as np             # <<<<<<<<<<<<<<
+ * cimport numpy as np
+ * from libcpp.string cimport string
+ */
+  __pyx_t_1 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_GOTREF(__pyx_t_1);
+  if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_1) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
+  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
+
+  /* "string.from_py":13
+ * 
+ * @cname("__pyx_convert_string_from_py_std__in_string")
+ * cdef string __pyx_convert_string_from_py_std__in_string(object o) except *:             # <<<<<<<<<<<<<<
+ *     cdef Py_ssize_t length = 0
+ *     cdef const char* data = __Pyx_PyObject_AsStringAndSize(o, &length)
+ */
+
+  /*--- Wrapped vars code ---*/
+
+  goto __pyx_L0;
+  __pyx_L1_error:;
+  __Pyx_XDECREF(__pyx_t_1);
+  if (__pyx_m) {
+    if (__pyx_d) {
+      __Pyx_AddTraceback("init insightface.thirdparty.face3d.mesh.cython.mesh_core_cython", __pyx_clineno, __pyx_lineno, __pyx_filename);
+    }
+    Py_CLEAR(__pyx_m);
+  } else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_ImportError, "init insightface.thirdparty.face3d.mesh.cython.mesh_core_cython");
+  }
+  __pyx_L0:;
+  __Pyx_RefNannyFinishContext();
+  #if CYTHON_PEP489_MULTI_PHASE_INIT
+  return (__pyx_m != NULL) ? 0 : -1;
+  #elif PY_MAJOR_VERSION >= 3
+  return __pyx_m;
+  #else
+  return;
+  #endif
+}
+
+/* --- Runtime support code --- */
+/* Refnanny */
+#if CYTHON_REFNANNY
+static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) {
+    PyObject *m = NULL, *p = NULL;
+    void *r = NULL;
+    m = PyImport_ImportModule(modname);
+    if (!m) goto end;
+    p = PyObject_GetAttrString(m, "RefNannyAPI");
+    if (!p) goto end;
+    r = PyLong_AsVoidPtr(p);
+end:
+    Py_XDECREF(p);
+    Py_XDECREF(m);
+    return (__Pyx_RefNannyAPIStruct *)r;
+}
+#endif
+
+/* RaiseArgTupleInvalid */
+static void __Pyx_RaiseArgtupleInvalid(
+    const char* func_name,
+    int exact,
+    Py_ssize_t num_min,
+    Py_ssize_t num_max,
+    Py_ssize_t num_found)
+{
+    Py_ssize_t num_expected;
+    const char *more_or_less;
+    if (num_found < num_min) {
+        num_expected = num_min;
+        more_or_less = "at least";
+    } else {
+        num_expected = num_max;
+        more_or_less = "at most";
+    }
+    if (exact) {
+        more_or_less = "exactly";
+    }
+    PyErr_Format(PyExc_TypeError,
+                 "%.200s() takes %.8s %" CYTHON_FORMAT_SSIZE_T "d positional argument%.1s (%" CYTHON_FORMAT_SSIZE_T "d given)",
+                 func_name, more_or_less, num_expected,
+                 (num_expected == 1) ? "" : "s", num_found);
+}
+
+/* RaiseDoubleKeywords */
+static void __Pyx_RaiseDoubleKeywordsError(
+    const char* func_name,
+    PyObject* kw_name)
+{
+    PyErr_Format(PyExc_TypeError,
+        #if PY_MAJOR_VERSION >= 3
+        "%s() got multiple values for keyword argument '%U'", func_name, kw_name);
+        #else
+        "%s() got multiple values for keyword argument '%s'", func_name,
+        PyString_AsString(kw_name));
+        #endif
+}
+
+/* ParseKeywords */
+static int __Pyx_ParseOptionalKeywords(
+    PyObject *kwds,
+    PyObject **argnames[],
+    PyObject *kwds2,
+    PyObject *values[],
+    Py_ssize_t num_pos_args,
+    const char* function_name)
+{
+    PyObject *key = 0, *value = 0;
+    Py_ssize_t pos = 0;
+    PyObject*** name;
+    PyObject*** first_kw_arg = argnames + num_pos_args;
+    while (PyDict_Next(kwds, &pos, &key, &value)) {
+        name = first_kw_arg;
+        while (*name && (**name != key)) name++;
+        if (*name) {
+            values[name-argnames] = value;
+            continue;
+        }
+        name = first_kw_arg;
+        #if PY_MAJOR_VERSION < 3
+        if (likely(PyString_Check(key))) {
+            while (*name) {
+                if ((CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**name) == PyString_GET_SIZE(key))
+                        && _PyString_Eq(**name, key)) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    if ((**argname == key) || (
+                            (CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**argname) == PyString_GET_SIZE(key))
+                             && _PyString_Eq(**argname, key))) {
+                        goto arg_passed_twice;
+                    }
+                    argname++;
+                }
+            }
+        } else
+        #endif
+        if (likely(PyUnicode_Check(key))) {
+            while (*name) {
+                int cmp = (**name == key) ? 0 :
+                #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                    (__Pyx_PyUnicode_GET_LENGTH(**name) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 :
+                #endif
+                    PyUnicode_Compare(**name, key);
+                if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                if (cmp == 0) {
+                    values[name-argnames] = value;
+                    break;
+                }
+                name++;
+            }
+            if (*name) continue;
+            else {
+                PyObject*** argname = argnames;
+                while (argname != first_kw_arg) {
+                    int cmp = (**argname == key) ? 0 :
+                    #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3
+                        (__Pyx_PyUnicode_GET_LENGTH(**argname) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 :
+                    #endif
+                        PyUnicode_Compare(**argname, key);
+                    if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad;
+                    if (cmp == 0) goto arg_passed_twice;
+                    argname++;
+                }
+            }
+        } else
+            goto invalid_keyword_type;
+        if (kwds2) {
+            if (unlikely(PyDict_SetItem(kwds2, key, value))) goto bad;
+        } else {
+            goto invalid_keyword;
+        }
+    }
+    return 0;
+arg_passed_twice:
+    __Pyx_RaiseDoubleKeywordsError(function_name, key);
+    goto bad;
+invalid_keyword_type:
+    PyErr_Format(PyExc_TypeError,
+        "%.200s() keywords must be strings", function_name);
+    goto bad;
+invalid_keyword:
+    PyErr_Format(PyExc_TypeError,
+    #if PY_MAJOR_VERSION < 3
+        "%.200s() got an unexpected keyword argument '%.200s'",
+        function_name, PyString_AsString(key));
+    #else
+        "%s() got an unexpected keyword argument '%U'",
+        function_name, key);
+    #endif
+bad:
+    return -1;
+}
+
+/* ArgTypeTest */
+static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact)
+{
+    if (unlikely(!type)) {
+        PyErr_SetString(PyExc_SystemError, "Missing type object");
+        return 0;
+    }
+    else if (exact) {
+        #if PY_MAJOR_VERSION == 2
+        if ((type == &PyBaseString_Type) && likely(__Pyx_PyBaseString_CheckExact(obj))) return 1;
+        #endif
+    }
+    else {
+        if (likely(__Pyx_TypeCheck(obj, type))) return 1;
+    }
+    PyErr_Format(PyExc_TypeError,
+        "Argument '%.200s' has incorrect type (expected %.200s, got %.200s)",
+        name, type->tp_name, Py_TYPE(obj)->tp_name);
+    return 0;
+}
+
+/* IsLittleEndian */
+static CYTHON_INLINE int __Pyx_Is_Little_Endian(void)
+{
+  union {
+    uint32_t u32;
+    uint8_t u8[4];
+  } S;
+  S.u32 = 0x01020304;
+  return S.u8[0] == 4;
+}
+
+/* BufferFormatCheck */
+static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx,
+                              __Pyx_BufFmt_StackElem* stack,
+                              __Pyx_TypeInfo* type) {
+  stack[0].field = &ctx->root;
+  stack[0].parent_offset = 0;
+  ctx->root.type = type;
+  ctx->root.name = "buffer dtype";
+  ctx->root.offset = 0;
+  ctx->head = stack;
+  ctx->head->field = &ctx->root;
+  ctx->fmt_offset = 0;
+  ctx->head->parent_offset = 0;
+  ctx->new_packmode = '@';
+  ctx->enc_packmode = '@';
+  ctx->new_count = 1;
+  ctx->enc_count = 0;
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  ctx->is_valid_array = 0;
+  ctx->struct_alignment = 0;
+  while (type->typegroup == 'S') {
+    ++ctx->head;
+    ctx->head->field = type->fields;
+    ctx->head->parent_offset = 0;
+    type = type->fields->type;
+  }
+}
+static int __Pyx_BufFmt_ParseNumber(const char** ts) {
+    int count;
+    const char* t = *ts;
+    if (*t < '0' || *t > '9') {
+      return -1;
+    } else {
+        count = *t++ - '0';
+        while (*t >= '0' && *t <= '9') {
+            count *= 10;
+            count += *t++ - '0';
+        }
+    }
+    *ts = t;
+    return count;
+}
+static int __Pyx_BufFmt_ExpectNumber(const char **ts) {
+    int number = __Pyx_BufFmt_ParseNumber(ts);
+    if (number == -1)
+        PyErr_Format(PyExc_ValueError,\
+                     "Does not understand character buffer dtype format string ('%c')", **ts);
+    return number;
+}
+static void __Pyx_BufFmt_RaiseUnexpectedChar(char ch) {
+  PyErr_Format(PyExc_ValueError,
+               "Unexpected format string character: '%c'", ch);
+}
+static const char* __Pyx_BufFmt_DescribeTypeChar(char ch, int is_complex) {
+  switch (ch) {
+    case '?': return "'bool'";
+    case 'c': return "'char'";
+    case 'b': return "'signed char'";
+    case 'B': return "'unsigned char'";
+    case 'h': return "'short'";
+    case 'H': return "'unsigned short'";
+    case 'i': return "'int'";
+    case 'I': return "'unsigned int'";
+    case 'l': return "'long'";
+    case 'L': return "'unsigned long'";
+    case 'q': return "'long long'";
+    case 'Q': return "'unsigned long long'";
+    case 'f': return (is_complex ? "'complex float'" : "'float'");
+    case 'd': return (is_complex ? "'complex double'" : "'double'");
+    case 'g': return (is_complex ? "'complex long double'" : "'long double'");
+    case 'T': return "a struct";
+    case 'O': return "Python object";
+    case 'P': return "a pointer";
+    case 's': case 'p': return "a string";
+    case 0: return "end";
+    default: return "unparseable format string";
+  }
+}
+static size_t __Pyx_BufFmt_TypeCharToStandardSize(char ch, int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return 2;
+    case 'i': case 'I': case 'l': case 'L': return 4;
+    case 'q': case 'Q': return 8;
+    case 'f': return (is_complex ? 8 : 4);
+    case 'd': return (is_complex ? 16 : 8);
+    case 'g': {
+      PyErr_SetString(PyExc_ValueError, "Python does not define a standard format string size for long double ('g')..");
+      return 0;
+    }
+    case 'O': case 'P': return sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static size_t __Pyx_BufFmt_TypeCharToNativeSize(char ch, int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(short);
+    case 'i': case 'I': return sizeof(int);
+    case 'l': case 'L': return sizeof(long);
+    #ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(PY_LONG_LONG);
+    #endif
+    case 'f': return sizeof(float) * (is_complex ? 2 : 1);
+    case 'd': return sizeof(double) * (is_complex ? 2 : 1);
+    case 'g': return sizeof(long double) * (is_complex ? 2 : 1);
+    case 'O': case 'P': return sizeof(void*);
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+typedef struct { char c; short x; } __Pyx_st_short;
+typedef struct { char c; int x; } __Pyx_st_int;
+typedef struct { char c; long x; } __Pyx_st_long;
+typedef struct { char c; float x; } __Pyx_st_float;
+typedef struct { char c; double x; } __Pyx_st_double;
+typedef struct { char c; long double x; } __Pyx_st_longdouble;
+typedef struct { char c; void *x; } __Pyx_st_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { char c; PY_LONG_LONG x; } __Pyx_st_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToAlignment(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_st_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_st_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_st_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_st_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_st_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_st_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_st_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_st_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+/* These are for computing the padding at the end of the struct to align
+   on the first member of the struct. This will probably the same as above,
+   but we don't have any guarantees.
+ */
+typedef struct { short x; char c; } __Pyx_pad_short;
+typedef struct { int x; char c; } __Pyx_pad_int;
+typedef struct { long x; char c; } __Pyx_pad_long;
+typedef struct { float x; char c; } __Pyx_pad_float;
+typedef struct { double x; char c; } __Pyx_pad_double;
+typedef struct { long double x; char c; } __Pyx_pad_longdouble;
+typedef struct { void *x; char c; } __Pyx_pad_void_p;
+#ifdef HAVE_LONG_LONG
+typedef struct { PY_LONG_LONG x; char c; } __Pyx_pad_longlong;
+#endif
+static size_t __Pyx_BufFmt_TypeCharToPadding(char ch, CYTHON_UNUSED int is_complex) {
+  switch (ch) {
+    case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1;
+    case 'h': case 'H': return sizeof(__Pyx_pad_short) - sizeof(short);
+    case 'i': case 'I': return sizeof(__Pyx_pad_int) - sizeof(int);
+    case 'l': case 'L': return sizeof(__Pyx_pad_long) - sizeof(long);
+#ifdef HAVE_LONG_LONG
+    case 'q': case 'Q': return sizeof(__Pyx_pad_longlong) - sizeof(PY_LONG_LONG);
+#endif
+    case 'f': return sizeof(__Pyx_pad_float) - sizeof(float);
+    case 'd': return sizeof(__Pyx_pad_double) - sizeof(double);
+    case 'g': return sizeof(__Pyx_pad_longdouble) - sizeof(long double);
+    case 'P': case 'O': return sizeof(__Pyx_pad_void_p) - sizeof(void*);
+    default:
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+}
+static char __Pyx_BufFmt_TypeCharToGroup(char ch, int is_complex) {
+  switch (ch) {
+    case 'c':
+        return 'H';
+    case 'b': case 'h': case 'i':
+    case 'l': case 'q': case 's': case 'p':
+        return 'I';
+    case '?': case 'B': case 'H': case 'I': case 'L': case 'Q':
+        return 'U';
+    case 'f': case 'd': case 'g':
+        return (is_complex ? 'C' : 'R');
+    case 'O':
+        return 'O';
+    case 'P':
+        return 'P';
+    default: {
+      __Pyx_BufFmt_RaiseUnexpectedChar(ch);
+      return 0;
+    }
+  }
+}
+static void __Pyx_BufFmt_RaiseExpected(__Pyx_BufFmt_Context* ctx) {
+  if (ctx->head == NULL || ctx->head->field == &ctx->root) {
+    const char* expected;
+    const char* quote;
+    if (ctx->head == NULL) {
+      expected = "end";
+      quote = "";
+    } else {
+      expected = ctx->head->field->type->name;
+      quote = "'";
+    }
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected %s%s%s but got %s",
+                 quote, expected, quote,
+                 __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex));
+  } else {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_StructField* parent = (ctx->head - 1)->field;
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer dtype mismatch, expected '%s' but got %s in '%s.%s'",
+                 field->type->name, __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex),
+                 parent->type->name, field->name);
+  }
+}
+static int __Pyx_BufFmt_ProcessTypeChunk(__Pyx_BufFmt_Context* ctx) {
+  char group;
+  size_t size, offset, arraysize = 1;
+  if (ctx->enc_type == 0) return 0;
+  if (ctx->head->field->type->arraysize[0]) {
+    int i, ndim = 0;
+    if (ctx->enc_type == 's' || ctx->enc_type == 'p') {
+        ctx->is_valid_array = ctx->head->field->type->ndim == 1;
+        ndim = 1;
+        if (ctx->enc_count != ctx->head->field->type->arraysize[0]) {
+            PyErr_Format(PyExc_ValueError,
+                         "Expected a dimension of size %zu, got %zu",
+                         ctx->head->field->type->arraysize[0], ctx->enc_count);
+            return -1;
+        }
+    }
+    if (!ctx->is_valid_array) {
+      PyErr_Format(PyExc_ValueError, "Expected %d dimensions, got %d",
+                   ctx->head->field->type->ndim, ndim);
+      return -1;
+    }
+    for (i = 0; i < ctx->head->field->type->ndim; i++) {
+      arraysize *= ctx->head->field->type->arraysize[i];
+    }
+    ctx->is_valid_array = 0;
+    ctx->enc_count = 1;
+  }
+  group = __Pyx_BufFmt_TypeCharToGroup(ctx->enc_type, ctx->is_complex);
+  do {
+    __Pyx_StructField* field = ctx->head->field;
+    __Pyx_TypeInfo* type = field->type;
+    if (ctx->enc_packmode == '@' || ctx->enc_packmode == '^') {
+      size = __Pyx_BufFmt_TypeCharToNativeSize(ctx->enc_type, ctx->is_complex);
+    } else {
+      size = __Pyx_BufFmt_TypeCharToStandardSize(ctx->enc_type, ctx->is_complex);
+    }
+    if (ctx->enc_packmode == '@') {
+      size_t align_at = __Pyx_BufFmt_TypeCharToAlignment(ctx->enc_type, ctx->is_complex);
+      size_t align_mod_offset;
+      if (align_at == 0) return -1;
+      align_mod_offset = ctx->fmt_offset % align_at;
+      if (align_mod_offset > 0) ctx->fmt_offset += align_at - align_mod_offset;
+      if (ctx->struct_alignment == 0)
+          ctx->struct_alignment = __Pyx_BufFmt_TypeCharToPadding(ctx->enc_type,
+                                                                 ctx->is_complex);
+    }
+    if (type->size != size || type->typegroup != group) {
+      if (type->typegroup == 'C' && type->fields != NULL) {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        ++ctx->head;
+        ctx->head->field = type->fields;
+        ctx->head->parent_offset = parent_offset;
+        continue;
+      }
+      if ((type->typegroup == 'H' || group == 'H') && type->size == size) {
+      } else {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+      }
+    }
+    offset = ctx->head->parent_offset + field->offset;
+    if (ctx->fmt_offset != offset) {
+      PyErr_Format(PyExc_ValueError,
+                   "Buffer dtype mismatch; next field is at offset %" CYTHON_FORMAT_SSIZE_T "d but %" CYTHON_FORMAT_SSIZE_T "d expected",
+                   (Py_ssize_t)ctx->fmt_offset, (Py_ssize_t)offset);
+      return -1;
+    }
+    ctx->fmt_offset += size;
+    if (arraysize)
+      ctx->fmt_offset += (arraysize - 1) * size;
+    --ctx->enc_count;
+    while (1) {
+      if (field == &ctx->root) {
+        ctx->head = NULL;
+        if (ctx->enc_count != 0) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return -1;
+        }
+        break;
+      }
+      ctx->head->field = ++field;
+      if (field->type == NULL) {
+        --ctx->head;
+        field = ctx->head->field;
+        continue;
+      } else if (field->type->typegroup == 'S') {
+        size_t parent_offset = ctx->head->parent_offset + field->offset;
+        if (field->type->fields->type == NULL) continue;
+        field = field->type->fields;
+        ++ctx->head;
+        ctx->head->field = field;
+        ctx->head->parent_offset = parent_offset;
+        break;
+      } else {
+        break;
+      }
+    }
+  } while (ctx->enc_count);
+  ctx->enc_type = 0;
+  ctx->is_complex = 0;
+  return 0;
+}
+static PyObject *
+__pyx_buffmt_parse_array(__Pyx_BufFmt_Context* ctx, const char** tsp)
+{
+    const char *ts = *tsp;
+    int i = 0, number, ndim;
+    ++ts;
+    if (ctx->new_count != 1) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Cannot handle repeated arrays in format string");
+        return NULL;
+    }
+    if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+    ndim = ctx->head->field->type->ndim;
+    while (*ts && *ts != ')') {
+        switch (*ts) {
+            case ' ': case '\f': case '\r': case '\n': case '\t': case '\v':  continue;
+            default:  break;
+        }
+        number = __Pyx_BufFmt_ExpectNumber(&ts);
+        if (number == -1) return NULL;
+        if (i < ndim && (size_t) number != ctx->head->field->type->arraysize[i])
+            return PyErr_Format(PyExc_ValueError,
+                        "Expected a dimension of size %zu, got %d",
+                        ctx->head->field->type->arraysize[i], number);
+        if (*ts != ',' && *ts != ')')
+            return PyErr_Format(PyExc_ValueError,
+                                "Expected a comma in format string, got '%c'", *ts);
+        if (*ts == ',') ts++;
+        i++;
+    }
+    if (i != ndim)
+        return PyErr_Format(PyExc_ValueError, "Expected %d dimension(s), got %d",
+                            ctx->head->field->type->ndim, i);
+    if (!*ts) {
+        PyErr_SetString(PyExc_ValueError,
+                        "Unexpected end of format string, expected ')'");
+        return NULL;
+    }
+    ctx->is_valid_array = 1;
+    ctx->new_count = 1;
+    *tsp = ++ts;
+    return Py_None;
+}
+static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts) {
+  int got_Z = 0;
+  while (1) {
+    switch(*ts) {
+      case 0:
+        if (ctx->enc_type != 0 && ctx->head == NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        if (ctx->head != NULL) {
+          __Pyx_BufFmt_RaiseExpected(ctx);
+          return NULL;
+        }
+        return ts;
+      case ' ':
+      case '\r':
+      case '\n':
+        ++ts;
+        break;
+      case '<':
+        if (!__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Little-endian buffer not supported on big-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '>':
+      case '!':
+        if (__Pyx_Is_Little_Endian()) {
+          PyErr_SetString(PyExc_ValueError, "Big-endian buffer not supported on little-endian compiler");
+          return NULL;
+        }
+        ctx->new_packmode = '=';
+        ++ts;
+        break;
+      case '=':
+      case '@':
+      case '^':
+        ctx->new_packmode = *ts++;
+        break;
+      case 'T':
+        {
+          const char* ts_after_sub;
+          size_t i, struct_count = ctx->new_count;
+          size_t struct_alignment = ctx->struct_alignment;
+          ctx->new_count = 1;
+          ++ts;
+          if (*ts != '{') {
+            PyErr_SetString(PyExc_ValueError, "Buffer acquisition: Expected '{' after 'T'");
+            return NULL;
+          }
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          ctx->enc_count = 0;
+          ctx->struct_alignment = 0;
+          ++ts;
+          ts_after_sub = ts;
+          for (i = 0; i != struct_count; ++i) {
+            ts_after_sub = __Pyx_BufFmt_CheckString(ctx, ts);
+            if (!ts_after_sub) return NULL;
+          }
+          ts = ts_after_sub;
+          if (struct_alignment) ctx->struct_alignment = struct_alignment;
+        }
+        break;
+      case '}':
+        {
+          size_t alignment = ctx->struct_alignment;
+          ++ts;
+          if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+          ctx->enc_type = 0;
+          if (alignment && ctx->fmt_offset % alignment) {
+            ctx->fmt_offset += alignment - (ctx->fmt_offset % alignment);
+          }
+        }
+        return ts;
+      case 'x':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->fmt_offset += ctx->new_count;
+        ctx->new_count = 1;
+        ctx->enc_count = 0;
+        ctx->enc_type = 0;
+        ctx->enc_packmode = ctx->new_packmode;
+        ++ts;
+        break;
+      case 'Z':
+        got_Z = 1;
+        ++ts;
+        if (*ts != 'f' && *ts != 'd' && *ts != 'g') {
+          __Pyx_BufFmt_RaiseUnexpectedChar('Z');
+          return NULL;
+        }
+        CYTHON_FALLTHROUGH;
+      case '?': case 'c': case 'b': case 'B': case 'h': case 'H': case 'i': case 'I':
+      case 'l': case 'L': case 'q': case 'Q':
+      case 'f': case 'd': case 'g':
+      case 'O': case 'p':
+        if ((ctx->enc_type == *ts) && (got_Z == ctx->is_complex) &&
+            (ctx->enc_packmode == ctx->new_packmode) && (!ctx->is_valid_array)) {
+          ctx->enc_count += ctx->new_count;
+          ctx->new_count = 1;
+          got_Z = 0;
+          ++ts;
+          break;
+        }
+        CYTHON_FALLTHROUGH;
+      case 's':
+        if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL;
+        ctx->enc_count = ctx->new_count;
+        ctx->enc_packmode = ctx->new_packmode;
+        ctx->enc_type = *ts;
+        ctx->is_complex = got_Z;
+        ++ts;
+        ctx->new_count = 1;
+        got_Z = 0;
+        break;
+      case ':':
+        ++ts;
+        while(*ts != ':') ++ts;
+        ++ts;
+        break;
+      case '(':
+        if (!__pyx_buffmt_parse_array(ctx, &ts)) return NULL;
+        break;
+      default:
+        {
+          int number = __Pyx_BufFmt_ExpectNumber(&ts);
+          if (number == -1) return NULL;
+          ctx->new_count = (size_t)number;
+        }
+    }
+  }
+}
+
+/* BufferGetAndValidate */
+  static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info) {
+  if (unlikely(info->buf == NULL)) return;
+  if (info->suboffsets == __Pyx_minusones) info->suboffsets = NULL;
+  __Pyx_ReleaseBuffer(info);
+}
+static void __Pyx_ZeroBuffer(Py_buffer* buf) {
+  buf->buf = NULL;
+  buf->obj = NULL;
+  buf->strides = __Pyx_zeros;
+  buf->shape = __Pyx_zeros;
+  buf->suboffsets = __Pyx_minusones;
+}
+static int __Pyx__GetBufferAndValidate(
+        Py_buffer* buf, PyObject* obj,  __Pyx_TypeInfo* dtype, int flags,
+        int nd, int cast, __Pyx_BufFmt_StackElem* stack)
+{
+  buf->buf = NULL;
+  if (unlikely(__Pyx_GetBuffer(obj, buf, flags) == -1)) {
+    __Pyx_ZeroBuffer(buf);
+    return -1;
+  }
+  if (unlikely(buf->ndim != nd)) {
+    PyErr_Format(PyExc_ValueError,
+                 "Buffer has wrong number of dimensions (expected %d, got %d)",
+                 nd, buf->ndim);
+    goto fail;
+  }
+  if (!cast) {
+    __Pyx_BufFmt_Context ctx;
+    __Pyx_BufFmt_Init(&ctx, stack, dtype);
+    if (!__Pyx_BufFmt_CheckString(&ctx, buf->format)) goto fail;
+  }
+  if (unlikely((size_t)buf->itemsize != dtype->size)) {
+    PyErr_Format(PyExc_ValueError,
+      "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "d byte%s) does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "d byte%s)",
+      buf->itemsize, (buf->itemsize > 1) ? "s" : "",
+      dtype->name, (Py_ssize_t)dtype->size, (dtype->size > 1) ? "s" : "");
+    goto fail;
+  }
+  if (buf->suboffsets == NULL) buf->suboffsets = __Pyx_minusones;
+  return 0;
+fail:;
+  __Pyx_SafeReleaseBuffer(buf);
+  return -1;
+}
+
+/* PyErrFetchRestore */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    tmp_type = tstate->curexc_type;
+    tmp_value = tstate->curexc_value;
+    tmp_tb = tstate->curexc_traceback;
+    tstate->curexc_type = type;
+    tstate->curexc_value = value;
+    tstate->curexc_traceback = tb;
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    *type = tstate->curexc_type;
+    *value = tstate->curexc_value;
+    *tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+}
+#endif
+
+/* PyObjectGetAttrStr */
+  #if CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) {
+    PyTypeObject* tp = Py_TYPE(obj);
+    if (likely(tp->tp_getattro))
+        return tp->tp_getattro(obj, attr_name);
+#if PY_MAJOR_VERSION < 3
+    if (likely(tp->tp_getattr))
+        return tp->tp_getattr(obj, PyString_AS_STRING(attr_name));
+#endif
+    return PyObject_GetAttr(obj, attr_name);
+}
+#endif
+
+/* GetBuiltinName */
+  static PyObject *__Pyx_GetBuiltinName(PyObject *name) {
+    PyObject* result = __Pyx_PyObject_GetAttrStr(__pyx_b, name);
+    if (unlikely(!result)) {
+        PyErr_Format(PyExc_NameError,
+#if PY_MAJOR_VERSION >= 3
+            "name '%U' is not defined", name);
+#else
+            "name '%.200s' is not defined", PyString_AS_STRING(name));
+#endif
+    }
+    return result;
+}
+
+/* GetTopmostException */
+  #if CYTHON_USE_EXC_INFO_STACK
+static _PyErr_StackItem *
+__Pyx_PyErr_GetTopmostException(PyThreadState *tstate)
+{
+    _PyErr_StackItem *exc_info = tstate->exc_info;
+    while ((exc_info->exc_type == NULL || exc_info->exc_type == Py_None) &&
+           exc_info->previous_item != NULL)
+    {
+        exc_info = exc_info->previous_item;
+    }
+    return exc_info;
+}
+#endif
+
+/* SaveResetException */
+  #if CYTHON_FAST_THREAD_STATE
+static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) {
+    #if CYTHON_USE_EXC_INFO_STACK
+    _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate);
+    *type = exc_info->exc_type;
+    *value = exc_info->exc_value;
+    *tb = exc_info->exc_traceback;
+    #else
+    *type = tstate->exc_type;
+    *value = tstate->exc_value;
+    *tb = tstate->exc_traceback;
+    #endif
+    Py_XINCREF(*type);
+    Py_XINCREF(*value);
+    Py_XINCREF(*tb);
+}
+static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) {
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    #if CYTHON_USE_EXC_INFO_STACK
+    _PyErr_StackItem *exc_info = tstate->exc_info;
+    tmp_type = exc_info->exc_type;
+    tmp_value = exc_info->exc_value;
+    tmp_tb = exc_info->exc_traceback;
+    exc_info->exc_type = type;
+    exc_info->exc_value = value;
+    exc_info->exc_traceback = tb;
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = type;
+    tstate->exc_value = value;
+    tstate->exc_traceback = tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+}
+#endif
+
+/* PyErrExceptionMatches */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx_PyErr_ExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        if (__Pyx_PyErr_GivenExceptionMatches(exc_type, PyTuple_GET_ITEM(tuple, i))) return 1;
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err) {
+    PyObject *exc_type = tstate->curexc_type;
+    if (exc_type == err) return 1;
+    if (unlikely(!exc_type)) return 0;
+    if (unlikely(PyTuple_Check(err)))
+        return __Pyx_PyErr_ExceptionMatchesTuple(exc_type, err);
+    return __Pyx_PyErr_GivenExceptionMatches(exc_type, err);
+}
+#endif
+
+/* GetException */
+  #if CYTHON_FAST_THREAD_STATE
+static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb)
+#else
+static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb)
+#endif
+{
+    PyObject *local_type, *local_value, *local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    PyObject *tmp_type, *tmp_value, *tmp_tb;
+    local_type = tstate->curexc_type;
+    local_value = tstate->curexc_value;
+    local_tb = tstate->curexc_traceback;
+    tstate->curexc_type = 0;
+    tstate->curexc_value = 0;
+    tstate->curexc_traceback = 0;
+#else
+    PyErr_Fetch(&local_type, &local_value, &local_tb);
+#endif
+    PyErr_NormalizeException(&local_type, &local_value, &local_tb);
+#if CYTHON_FAST_THREAD_STATE
+    if (unlikely(tstate->curexc_type))
+#else
+    if (unlikely(PyErr_Occurred()))
+#endif
+        goto bad;
+    #if PY_MAJOR_VERSION >= 3
+    if (local_tb) {
+        if (unlikely(PyException_SetTraceback(local_value, local_tb) < 0))
+            goto bad;
+    }
+    #endif
+    Py_XINCREF(local_tb);
+    Py_XINCREF(local_type);
+    Py_XINCREF(local_value);
+    *type = local_type;
+    *value = local_value;
+    *tb = local_tb;
+#if CYTHON_FAST_THREAD_STATE
+    #if CYTHON_USE_EXC_INFO_STACK
+    {
+        _PyErr_StackItem *exc_info = tstate->exc_info;
+        tmp_type = exc_info->exc_type;
+        tmp_value = exc_info->exc_value;
+        tmp_tb = exc_info->exc_traceback;
+        exc_info->exc_type = local_type;
+        exc_info->exc_value = local_value;
+        exc_info->exc_traceback = local_tb;
+    }
+    #else
+    tmp_type = tstate->exc_type;
+    tmp_value = tstate->exc_value;
+    tmp_tb = tstate->exc_traceback;
+    tstate->exc_type = local_type;
+    tstate->exc_value = local_value;
+    tstate->exc_traceback = local_tb;
+    #endif
+    Py_XDECREF(tmp_type);
+    Py_XDECREF(tmp_value);
+    Py_XDECREF(tmp_tb);
+#else
+    PyErr_SetExcInfo(local_type, local_value, local_tb);
+#endif
+    return 0;
+bad:
+    *type = 0;
+    *value = 0;
+    *tb = 0;
+    Py_XDECREF(local_type);
+    Py_XDECREF(local_value);
+    Py_XDECREF(local_tb);
+    return -1;
+}
+
+/* PyObjectCall */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw) {
+    PyObject *result;
+    ternaryfunc call = Py_TYPE(func)->tp_call;
+    if (unlikely(!call))
+        return PyObject_Call(func, arg, kw);
+    if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object")))
+        return NULL;
+    result = (*call)(func, arg, kw);
+    Py_LeaveRecursiveCall();
+    if (unlikely(!result) && unlikely(!PyErr_Occurred())) {
+        PyErr_SetString(
+            PyExc_SystemError,
+            "NULL result without error in PyObject_Call");
+    }
+    return result;
+}
+#endif
+
+/* RaiseException */
+  #if PY_MAJOR_VERSION < 3
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb,
+                        CYTHON_UNUSED PyObject *cause) {
+    __Pyx_PyThreadState_declare
+    Py_XINCREF(type);
+    if (!value || value == Py_None)
+        value = NULL;
+    else
+        Py_INCREF(value);
+    if (!tb || tb == Py_None)
+        tb = NULL;
+    else {
+        Py_INCREF(tb);
+        if (!PyTraceBack_Check(tb)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: arg 3 must be a traceback or None");
+            goto raise_error;
+        }
+    }
+    if (PyType_Check(type)) {
+#if CYTHON_COMPILING_IN_PYPY
+        if (!value) {
+            Py_INCREF(Py_None);
+            value = Py_None;
+        }
+#endif
+        PyErr_NormalizeException(&type, &value, &tb);
+    } else {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto raise_error;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(type);
+        Py_INCREF(type);
+        if (!PyType_IsSubtype((PyTypeObject *)type, (PyTypeObject *)PyExc_BaseException)) {
+            PyErr_SetString(PyExc_TypeError,
+                "raise: exception class must be a subclass of BaseException");
+            goto raise_error;
+        }
+    }
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrRestore(type, value, tb);
+    return;
+raise_error:
+    Py_XDECREF(value);
+    Py_XDECREF(type);
+    Py_XDECREF(tb);
+    return;
+}
+#else
+static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) {
+    PyObject* owned_instance = NULL;
+    if (tb == Py_None) {
+        tb = 0;
+    } else if (tb && !PyTraceBack_Check(tb)) {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: arg 3 must be a traceback or None");
+        goto bad;
+    }
+    if (value == Py_None)
+        value = 0;
+    if (PyExceptionInstance_Check(type)) {
+        if (value) {
+            PyErr_SetString(PyExc_TypeError,
+                "instance exception may not have a separate value");
+            goto bad;
+        }
+        value = type;
+        type = (PyObject*) Py_TYPE(value);
+    } else if (PyExceptionClass_Check(type)) {
+        PyObject *instance_class = NULL;
+        if (value && PyExceptionInstance_Check(value)) {
+            instance_class = (PyObject*) Py_TYPE(value);
+            if (instance_class != type) {
+                int is_subclass = PyObject_IsSubclass(instance_class, type);
+                if (!is_subclass) {
+                    instance_class = NULL;
+                } else if (unlikely(is_subclass == -1)) {
+                    goto bad;
+                } else {
+                    type = instance_class;
+                }
+            }
+        }
+        if (!instance_class) {
+            PyObject *args;
+            if (!value)
+                args = PyTuple_New(0);
+            else if (PyTuple_Check(value)) {
+                Py_INCREF(value);
+                args = value;
+            } else
+                args = PyTuple_Pack(1, value);
+            if (!args)
+                goto bad;
+            owned_instance = PyObject_Call(type, args, NULL);
+            Py_DECREF(args);
+            if (!owned_instance)
+                goto bad;
+            value = owned_instance;
+            if (!PyExceptionInstance_Check(value)) {
+                PyErr_Format(PyExc_TypeError,
+                             "calling %R should have returned an instance of "
+                             "BaseException, not %R",
+                             type, Py_TYPE(value));
+                goto bad;
+            }
+        }
+    } else {
+        PyErr_SetString(PyExc_TypeError,
+            "raise: exception class must be a subclass of BaseException");
+        goto bad;
+    }
+    if (cause) {
+        PyObject *fixed_cause;
+        if (cause == Py_None) {
+            fixed_cause = NULL;
+        } else if (PyExceptionClass_Check(cause)) {
+            fixed_cause = PyObject_CallObject(cause, NULL);
+            if (fixed_cause == NULL)
+                goto bad;
+        } else if (PyExceptionInstance_Check(cause)) {
+            fixed_cause = cause;
+            Py_INCREF(fixed_cause);
+        } else {
+            PyErr_SetString(PyExc_TypeError,
+                            "exception causes must derive from "
+                            "BaseException");
+            goto bad;
+        }
+        PyException_SetCause(value, fixed_cause);
+    }
+    PyErr_SetObject(type, value);
+    if (tb) {
+#if CYTHON_COMPILING_IN_PYPY
+        PyObject *tmp_type, *tmp_value, *tmp_tb;
+        PyErr_Fetch(&tmp_type, &tmp_value, &tmp_tb);
+        Py_INCREF(tb);
+        PyErr_Restore(tmp_type, tmp_value, tb);
+        Py_XDECREF(tmp_tb);
+#else
+        PyThreadState *tstate = __Pyx_PyThreadState_Current;
+        PyObject* tmp_tb = tstate->curexc_traceback;
+        if (tb != tmp_tb) {
+            Py_INCREF(tb);
+            tstate->curexc_traceback = tb;
+            Py_XDECREF(tmp_tb);
+        }
+#endif
+    }
+bad:
+    Py_XDECREF(owned_instance);
+    return;
+}
+#endif
+
+/* TypeImport */
+  #ifndef __PYX_HAVE_RT_ImportType
+#define __PYX_HAVE_RT_ImportType
+static PyTypeObject *__Pyx_ImportType(PyObject *module, const char *module_name, const char *class_name,
+    size_t size, enum __Pyx_ImportType_CheckSize check_size)
+{
+    PyObject *result = 0;
+    char warning[200];
+    Py_ssize_t basicsize;
+#ifdef Py_LIMITED_API
+    PyObject *py_basicsize;
+#endif
+    result = PyObject_GetAttrString(module, class_name);
+    if (!result)
+        goto bad;
+    if (!PyType_Check(result)) {
+        PyErr_Format(PyExc_TypeError,
+            "%.200s.%.200s is not a type object",
+            module_name, class_name);
+        goto bad;
+    }
+#ifndef Py_LIMITED_API
+    basicsize = ((PyTypeObject *)result)->tp_basicsize;
+#else
+    py_basicsize = PyObject_GetAttrString(result, "__basicsize__");
+    if (!py_basicsize)
+        goto bad;
+    basicsize = PyLong_AsSsize_t(py_basicsize);
+    Py_DECREF(py_basicsize);
+    py_basicsize = 0;
+    if (basicsize == (Py_ssize_t)-1 && PyErr_Occurred())
+        goto bad;
+#endif
+    if ((size_t)basicsize < size) {
+        PyErr_Format(PyExc_ValueError,
+            "%.200s.%.200s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        goto bad;
+    }
+    if (check_size == __Pyx_ImportType_CheckSize_Error && (size_t)basicsize != size) {
+        PyErr_Format(PyExc_ValueError,
+            "%.200s.%.200s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        goto bad;
+    }
+    else if (check_size == __Pyx_ImportType_CheckSize_Warn && (size_t)basicsize > size) {
+        PyOS_snprintf(warning, sizeof(warning),
+            "%s.%s size changed, may indicate binary incompatibility. "
+            "Expected %zd from C header, got %zd from PyObject",
+            module_name, class_name, size, basicsize);
+        if (PyErr_WarnEx(NULL, warning, 0) < 0) goto bad;
+    }
+    return (PyTypeObject *)result;
+bad:
+    Py_XDECREF(result);
+    return NULL;
+}
+#endif
+
+/* Import */
+  static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) {
+    PyObject *empty_list = 0;
+    PyObject *module = 0;
+    PyObject *global_dict = 0;
+    PyObject *empty_dict = 0;
+    PyObject *list;
+    #if PY_MAJOR_VERSION < 3
+    PyObject *py_import;
+    py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import);
+    if (!py_import)
+        goto bad;
+    #endif
+    if (from_list)
+        list = from_list;
+    else {
+        empty_list = PyList_New(0);
+        if (!empty_list)
+            goto bad;
+        list = empty_list;
+    }
+    global_dict = PyModule_GetDict(__pyx_m);
+    if (!global_dict)
+        goto bad;
+    empty_dict = PyDict_New();
+    if (!empty_dict)
+        goto bad;
+    {
+        #if PY_MAJOR_VERSION >= 3
+        if (level == -1) {
+            if ((1) && (strchr(__Pyx_MODULE_NAME, '.'))) {
+                module = PyImport_ImportModuleLevelObject(
+                    name, global_dict, empty_dict, list, 1);
+                if (!module) {
+                    if (!PyErr_ExceptionMatches(PyExc_ImportError))
+                        goto bad;
+                    PyErr_Clear();
+                }
+            }
+            level = 0;
+        }
+        #endif
+        if (!module) {
+            #if PY_MAJOR_VERSION < 3
+            PyObject *py_level = PyInt_FromLong(level);
+            if (!py_level)
+                goto bad;
+            module = PyObject_CallFunctionObjArgs(py_import,
+                name, global_dict, empty_dict, list, py_level, (PyObject *)NULL);
+            Py_DECREF(py_level);
+            #else
+            module = PyImport_ImportModuleLevelObject(
+                name, global_dict, empty_dict, list, level);
+            #endif
+        }
+    }
+bad:
+    #if PY_MAJOR_VERSION < 3
+    Py_XDECREF(py_import);
+    #endif
+    Py_XDECREF(empty_list);
+    Py_XDECREF(empty_dict);
+    return module;
+}
+
+/* PyDictVersioning */
+  #if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    return likely(dict) ? __PYX_GET_DICT_VERSION(dict) : 0;
+}
+static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj) {
+    PyObject **dictptr = NULL;
+    Py_ssize_t offset = Py_TYPE(obj)->tp_dictoffset;
+    if (offset) {
+#if CYTHON_COMPILING_IN_CPYTHON
+        dictptr = (likely(offset > 0)) ? (PyObject **) ((char *)obj + offset) : _PyObject_GetDictPtr(obj);
+#else
+        dictptr = _PyObject_GetDictPtr(obj);
+#endif
+    }
+    return (dictptr && *dictptr) ? __PYX_GET_DICT_VERSION(*dictptr) : 0;
+}
+static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version) {
+    PyObject *dict = Py_TYPE(obj)->tp_dict;
+    if (unlikely(!dict) || unlikely(tp_dict_version != __PYX_GET_DICT_VERSION(dict)))
+        return 0;
+    return obj_dict_version == __Pyx_get_object_dict_version(obj);
+}
+#endif
+
+/* CLineInTraceback */
+  #ifndef CYTHON_CLINE_IN_TRACEBACK
+static int __Pyx_CLineForTraceback(CYTHON_NCP_UNUSED PyThreadState *tstate, int c_line) {
+    PyObject *use_cline;
+    PyObject *ptype, *pvalue, *ptraceback;
+#if CYTHON_COMPILING_IN_CPYTHON
+    PyObject **cython_runtime_dict;
+#endif
+    if (unlikely(!__pyx_cython_runtime)) {
+        return c_line;
+    }
+    __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback);
+#if CYTHON_COMPILING_IN_CPYTHON
+    cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime);
+    if (likely(cython_runtime_dict)) {
+        __PYX_PY_DICT_LOOKUP_IF_MODIFIED(
+            use_cline, *cython_runtime_dict,
+            __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback))
+    } else
+#endif
+    {
+      PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback);
+      if (use_cline_obj) {
+        use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True;
+        Py_DECREF(use_cline_obj);
+      } else {
+        PyErr_Clear();
+        use_cline = NULL;
+      }
+    }
+    if (!use_cline) {
+        c_line = 0;
+        (void) PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False);
+    }
+    else if (use_cline == Py_False || (use_cline != Py_True && PyObject_Not(use_cline) != 0)) {
+        c_line = 0;
+    }
+    __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback);
+    return c_line;
+}
+#endif
+
+/* CodeObjectCache */
+  static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) {
+    int start = 0, mid = 0, end = count - 1;
+    if (end >= 0 && code_line > entries[end].code_line) {
+        return count;
+    }
+    while (start < end) {
+        mid = start + (end - start) / 2;
+        if (code_line < entries[mid].code_line) {
+            end = mid;
+        } else if (code_line > entries[mid].code_line) {
+             start = mid + 1;
+        } else {
+            return mid;
+        }
+    }
+    if (code_line <= entries[mid].code_line) {
+        return mid;
+    } else {
+        return mid + 1;
+    }
+}
+static PyCodeObject *__pyx_find_code_object(int code_line) {
+    PyCodeObject* code_object;
+    int pos;
+    if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) {
+        return NULL;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) {
+        return NULL;
+    }
+    code_object = __pyx_code_cache.entries[pos].code_object;
+    Py_INCREF(code_object);
+    return code_object;
+}
+static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) {
+    int pos, i;
+    __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries;
+    if (unlikely(!code_line)) {
+        return;
+    }
+    if (unlikely(!entries)) {
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry));
+        if (likely(entries)) {
+            __pyx_code_cache.entries = entries;
+            __pyx_code_cache.max_count = 64;
+            __pyx_code_cache.count = 1;
+            entries[0].code_line = code_line;
+            entries[0].code_object = code_object;
+            Py_INCREF(code_object);
+        }
+        return;
+    }
+    pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line);
+    if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) {
+        PyCodeObject* tmp = entries[pos].code_object;
+        entries[pos].code_object = code_object;
+        Py_DECREF(tmp);
+        return;
+    }
+    if (__pyx_code_cache.count == __pyx_code_cache.max_count) {
+        int new_max = __pyx_code_cache.max_count + 64;
+        entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc(
+            __pyx_code_cache.entries, ((size_t)new_max) * sizeof(__Pyx_CodeObjectCacheEntry));
+        if (unlikely(!entries)) {
+            return;
+        }
+        __pyx_code_cache.entries = entries;
+        __pyx_code_cache.max_count = new_max;
+    }
+    for (i=__pyx_code_cache.count; i>pos; i--) {
+        entries[i] = entries[i-1];
+    }
+    entries[pos].code_line = code_line;
+    entries[pos].code_object = code_object;
+    __pyx_code_cache.count++;
+    Py_INCREF(code_object);
+}
+
+/* AddTraceback */
+  #include "compile.h"
+#include "frameobject.h"
+#include "traceback.h"
+#if PY_VERSION_HEX >= 0x030b00a6
+  #ifndef Py_BUILD_CORE
+    #define Py_BUILD_CORE 1
+  #endif
+  #include "internal/pycore_frame.h"
+#endif
+static PyCodeObject* __Pyx_CreateCodeObjectForTraceback(
+            const char *funcname, int c_line,
+            int py_line, const char *filename) {
+    PyCodeObject *py_code = NULL;
+    PyObject *py_funcname = NULL;
+    #if PY_MAJOR_VERSION < 3
+    PyObject *py_srcfile = NULL;
+    py_srcfile = PyString_FromString(filename);
+    if (!py_srcfile) goto bad;
+    #endif
+    if (c_line) {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        if (!py_funcname) goto bad;
+        #else
+        py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line);
+        if (!py_funcname) goto bad;
+        funcname = PyUnicode_AsUTF8(py_funcname);
+        if (!funcname) goto bad;
+        #endif
+    }
+    else {
+        #if PY_MAJOR_VERSION < 3
+        py_funcname = PyString_FromString(funcname);
+        if (!py_funcname) goto bad;
+        #endif
+    }
+    #if PY_MAJOR_VERSION < 3
+    py_code = __Pyx_PyCode_New(
+        0,
+        0,
+        0,
+        0,
+        0,
+        __pyx_empty_bytes, /*PyObject *code,*/
+        __pyx_empty_tuple, /*PyObject *consts,*/
+        __pyx_empty_tuple, /*PyObject *names,*/
+        __pyx_empty_tuple, /*PyObject *varnames,*/
+        __pyx_empty_tuple, /*PyObject *freevars,*/
+        __pyx_empty_tuple, /*PyObject *cellvars,*/
+        py_srcfile,   /*PyObject *filename,*/
+        py_funcname,  /*PyObject *name,*/
+        py_line,
+        __pyx_empty_bytes  /*PyObject *lnotab*/
+    );
+    Py_DECREF(py_srcfile);
+    #else
+    py_code = PyCode_NewEmpty(filename, funcname, py_line);
+    #endif
+    Py_XDECREF(py_funcname);  // XDECREF since it's only set on Py3 if cline
+    return py_code;
+bad:
+    Py_XDECREF(py_funcname);
+    #if PY_MAJOR_VERSION < 3
+    Py_XDECREF(py_srcfile);
+    #endif
+    return NULL;
+}
+static void __Pyx_AddTraceback(const char *funcname, int c_line,
+                               int py_line, const char *filename) {
+    PyCodeObject *py_code = 0;
+    PyFrameObject *py_frame = 0;
+    PyThreadState *tstate = __Pyx_PyThreadState_Current;
+    PyObject *ptype, *pvalue, *ptraceback;
+    if (c_line) {
+        c_line = __Pyx_CLineForTraceback(tstate, c_line);
+    }
+    py_code = __pyx_find_code_object(c_line ? -c_line : py_line);
+    if (!py_code) {
+        __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback);
+        py_code = __Pyx_CreateCodeObjectForTraceback(
+            funcname, c_line, py_line, filename);
+        if (!py_code) {
+            /* If the code object creation fails, then we should clear the
+               fetched exception references and propagate the new exception */
+            Py_XDECREF(ptype);
+            Py_XDECREF(pvalue);
+            Py_XDECREF(ptraceback);
+            goto bad;
+        }
+        __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback);
+        __pyx_insert_code_object(c_line ? -c_line : py_line, py_code);
+    }
+    py_frame = PyFrame_New(
+        tstate,            /*PyThreadState *tstate,*/
+        py_code,           /*PyCodeObject *code,*/
+        __pyx_d,    /*PyObject *globals,*/
+        0                  /*PyObject *locals*/
+    );
+    if (!py_frame) goto bad;
+    __Pyx_PyFrame_SetLineNumber(py_frame, py_line);
+    PyTraceBack_Here(py_frame);
+bad:
+    Py_XDECREF(py_code);
+    Py_XDECREF(py_frame);
+}
+
+#if PY_MAJOR_VERSION < 3
+static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags) {
+    if (PyObject_CheckBuffer(obj)) return PyObject_GetBuffer(obj, view, flags);
+    PyErr_Format(PyExc_TypeError, "'%.200s' does not have the buffer interface", Py_TYPE(obj)->tp_name);
+    return -1;
+}
+static void __Pyx_ReleaseBuffer(Py_buffer *view) {
+    PyObject *obj = view->obj;
+    if (!obj) return;
+    if (PyObject_CheckBuffer(obj)) {
+        PyBuffer_Release(view);
+        return;
+    }
+    if ((0)) {}
+    view->obj = NULL;
+    Py_DECREF(obj);
+}
+#endif
+
+
+  /* CIntFromPyVerify */
+  #define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0)
+#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\
+    __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1)
+#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\
+    {\
+        func_type value = func_value;\
+        if (sizeof(target_type) < sizeof(func_type)) {\
+            if (unlikely(value != (func_type) (target_type) value)) {\
+                func_type zero = 0;\
+                if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\
+                    return (target_type) -1;\
+                if (is_unsigned && unlikely(value < zero))\
+                    goto raise_neg_overflow;\
+                else\
+                    goto raise_overflow;\
+            }\
+        }\
+        return (target_type) value;\
+    }
+
+/* Declarations */
+  #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return ::std::complex< float >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      return x + y*(__pyx_t_float_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) {
+      __pyx_t_float_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+  #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        __pyx_t_float_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabsf(b.real) >= fabsf(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                float r = b.imag / b.real;
+                float s = (float)(1.0) / (b.real + b.imag * r);
+                return __pyx_t_float_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            float r = b.real / b.imag;
+            float s = (float)(1.0) / (b.imag + b.real * r);
+            return __pyx_t_float_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            float denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_float_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex a) {
+        __pyx_t_float_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrtf(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypotf(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex a, __pyx_t_float_complex b) {
+            __pyx_t_float_complex z;
+            float r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    float denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        return __Pyx_c_prod_float(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_float(a, a);
+                        return __Pyx_c_prod_float(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = powf(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2f(0.0, -1.0);
+                }
+            } else {
+                r = __Pyx_c_abs_float(a);
+                theta = atan2f(a.imag, a.real);
+            }
+            lnr = logf(r);
+            z_r = expf(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cosf(z_theta);
+            z.imag = z_r * sinf(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* Declarations */
+  #if CYTHON_CCOMPLEX
+  #ifdef __cplusplus
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return ::std::complex< double >(x, y);
+    }
+  #else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      return x + y*(__pyx_t_double_complex)_Complex_I;
+    }
+  #endif
+#else
+    static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) {
+      __pyx_t_double_complex z;
+      z.real = x;
+      z.imag = y;
+      return z;
+    }
+#endif
+
+/* Arithmetic */
+  #if CYTHON_CCOMPLEX
+#else
+    static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+       return (a.real == b.real) && (a.imag == b.imag);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real + b.real;
+        z.imag = a.imag + b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real - b.real;
+        z.imag = a.imag - b.imag;
+        return z;
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        __pyx_t_double_complex z;
+        z.real = a.real * b.real - a.imag * b.imag;
+        z.imag = a.real * b.imag + a.imag * b.real;
+        return z;
+    }
+    #if 1
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else if (fabs(b.real) >= fabs(b.imag)) {
+            if (b.real == 0 && b.imag == 0) {
+                return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.imag);
+            } else {
+                double r = b.imag / b.real;
+                double s = (double)(1.0) / (b.real + b.imag * r);
+                return __pyx_t_double_complex_from_parts(
+                    (a.real + a.imag * r) * s, (a.imag - a.real * r) * s);
+            }
+        } else {
+            double r = b.real / b.imag;
+            double s = (double)(1.0) / (b.imag + b.real * r);
+            return __pyx_t_double_complex_from_parts(
+                (a.real * r + a.imag) * s, (a.imag * r - a.real) * s);
+        }
+    }
+    #else
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+        if (b.imag == 0) {
+            return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real);
+        } else {
+            double denom = b.real * b.real + b.imag * b.imag;
+            return __pyx_t_double_complex_from_parts(
+                (a.real * b.real + a.imag * b.imag) / denom,
+                (a.imag * b.real - a.real * b.imag) / denom);
+        }
+    }
+    #endif
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real = -a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex a) {
+       return (a.real == 0) && (a.imag == 0);
+    }
+    static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex a) {
+        __pyx_t_double_complex z;
+        z.real =  a.real;
+        z.imag = -a.imag;
+        return z;
+    }
+    #if 1
+        static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex z) {
+          #if !defined(HAVE_HYPOT) || defined(_MSC_VER)
+            return sqrt(z.real*z.real + z.imag*z.imag);
+          #else
+            return hypot(z.real, z.imag);
+          #endif
+        }
+        static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex a, __pyx_t_double_complex b) {
+            __pyx_t_double_complex z;
+            double r, lnr, theta, z_r, z_theta;
+            if (b.imag == 0 && b.real == (int)b.real) {
+                if (b.real < 0) {
+                    double denom = a.real * a.real + a.imag * a.imag;
+                    a.real = a.real / denom;
+                    a.imag = -a.imag / denom;
+                    b.real = -b.real;
+                }
+                switch ((int)b.real) {
+                    case 0:
+                        z.real = 1;
+                        z.imag = 0;
+                        return z;
+                    case 1:
+                        return a;
+                    case 2:
+                        return __Pyx_c_prod_double(a, a);
+                    case 3:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, a);
+                    case 4:
+                        z = __Pyx_c_prod_double(a, a);
+                        return __Pyx_c_prod_double(z, z);
+                }
+            }
+            if (a.imag == 0) {
+                if (a.real == 0) {
+                    return a;
+                } else if (b.imag == 0) {
+                    z.real = pow(a.real, b.real);
+                    z.imag = 0;
+                    return z;
+                } else if (a.real > 0) {
+                    r = a.real;
+                    theta = 0;
+                } else {
+                    r = -a.real;
+                    theta = atan2(0.0, -1.0);
+                }
+            } else {
+                r = __Pyx_c_abs_double(a);
+                theta = atan2(a.imag, a.real);
+            }
+            lnr = log(r);
+            z_r = exp(lnr * b.real - theta * b.imag);
+            z_theta = theta * b.real + lnr * b.imag;
+            z.real = z_r * cos(z_theta);
+            z.imag = z_r * sin(z_theta);
+            return z;
+        }
+    #endif
+#endif
+
+/* CIntFromPy */
+  static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const int neg_one = (int) -1, const_zero = (int) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(int) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (int) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(int, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 2 * PyLong_SHIFT) {
+                            return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 3 * PyLong_SHIFT) {
+                            return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) >= 4 * PyLong_SHIFT) {
+                            return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (int) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(int) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (int) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(int, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(int,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(int) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(int) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                            return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(int) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(int) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                            return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(int) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(int) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(int) - 1 > 4 * PyLong_SHIFT) {
+                            return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(int) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            int val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (int) -1;
+        }
+    } else {
+        int val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (int) -1;
+        val = __Pyx_PyInt_As_int(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to int");
+    return (int) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to int");
+    return (int) -1;
+}
+
+/* CIntToPy */
+  static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const long neg_one = (long) -1, const_zero = (long) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+    if (is_unsigned) {
+        if (sizeof(long) < sizeof(long)) {
+            return PyInt_FromLong((long) value);
+        } else if (sizeof(long) <= sizeof(unsigned long)) {
+            return PyLong_FromUnsignedLong((unsigned long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+            return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value);
+#endif
+        }
+    } else {
+        if (sizeof(long) <= sizeof(long)) {
+            return PyInt_FromLong((long) value);
+#ifdef HAVE_LONG_LONG
+        } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+            return PyLong_FromLongLong((PY_LONG_LONG) value);
+#endif
+        }
+    }
+    {
+        int one = 1; int little = (int)*(unsigned char *)&one;
+        unsigned char *bytes = (unsigned char *)&value;
+        return _PyLong_FromByteArray(bytes, sizeof(long),
+                                     little, !is_unsigned);
+    }
+}
+
+/* CIntFromPy */
+  static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) {
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+    const long neg_one = (long) -1, const_zero = (long) 0;
+#ifdef __Pyx_HAS_GCC_DIAGNOSTIC
+#pragma GCC diagnostic pop
+#endif
+    const int is_unsigned = neg_one > const_zero;
+#if PY_MAJOR_VERSION < 3
+    if (likely(PyInt_Check(x))) {
+        if (sizeof(long) < sizeof(long)) {
+            __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x))
+        } else {
+            long val = PyInt_AS_LONG(x);
+            if (is_unsigned && unlikely(val < 0)) {
+                goto raise_neg_overflow;
+            }
+            return (long) val;
+        }
+    } else
+#endif
+    if (likely(PyLong_Check(x))) {
+        if (is_unsigned) {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case  1: __PYX_VERIFY_RETURN_INT(long, digit, digits[0])
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 2 * PyLong_SHIFT) {
+                            return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 3 * PyLong_SHIFT) {
+                            return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) >= 4 * PyLong_SHIFT) {
+                            return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]));
+                        }
+                    }
+                    break;
+            }
+#endif
+#if CYTHON_COMPILING_IN_CPYTHON
+            if (unlikely(Py_SIZE(x) < 0)) {
+                goto raise_neg_overflow;
+            }
+#else
+            {
+                int result = PyObject_RichCompareBool(x, Py_False, Py_LT);
+                if (unlikely(result < 0))
+                    return (long) -1;
+                if (unlikely(result == 1))
+                    goto raise_neg_overflow;
+            }
+#endif
+            if (sizeof(long) <= sizeof(unsigned long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x))
+#endif
+            }
+        } else {
+#if CYTHON_USE_PYLONG_INTERNALS
+            const digit* digits = ((PyLongObject*)x)->ob_digit;
+            switch (Py_SIZE(x)) {
+                case  0: return (long) 0;
+                case -1: __PYX_VERIFY_RETURN_INT(long, sdigit, (sdigit) (-(sdigit)digits[0]))
+                case  1: __PYX_VERIFY_RETURN_INT(long,  digit, +digits[0])
+                case -2:
+                    if (8 * sizeof(long) - 1 > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 2:
+                    if (8 * sizeof(long) > 1 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 2 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                            return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -3:
+                    if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 3:
+                    if (8 * sizeof(long) > 2 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 3 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                            return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case -4:
+                    if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+                case 4:
+                    if (8 * sizeof(long) > 3 * PyLong_SHIFT) {
+                        if (8 * sizeof(unsigned long) > 4 * PyLong_SHIFT) {
+                            __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])))
+                        } else if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) {
+                            return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])));
+                        }
+                    }
+                    break;
+            }
+#endif
+            if (sizeof(long) <= sizeof(long)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x))
+#ifdef HAVE_LONG_LONG
+            } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) {
+                __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x))
+#endif
+            }
+        }
+        {
+#if CYTHON_COMPILING_IN_PYPY && !defined(_PyLong_AsByteArray)
+            PyErr_SetString(PyExc_RuntimeError,
+                            "_PyLong_AsByteArray() not available in PyPy, cannot convert large numbers");
+#else
+            long val;
+            PyObject *v = __Pyx_PyNumber_IntOrLong(x);
+ #if PY_MAJOR_VERSION < 3
+            if (likely(v) && !PyLong_Check(v)) {
+                PyObject *tmp = v;
+                v = PyNumber_Long(tmp);
+                Py_DECREF(tmp);
+            }
+ #endif
+            if (likely(v)) {
+                int one = 1; int is_little = (int)*(unsigned char *)&one;
+                unsigned char *bytes = (unsigned char *)&val;
+                int ret = _PyLong_AsByteArray((PyLongObject *)v,
+                                              bytes, sizeof(val),
+                                              is_little, !is_unsigned);
+                Py_DECREF(v);
+                if (likely(!ret))
+                    return val;
+            }
+#endif
+            return (long) -1;
+        }
+    } else {
+        long val;
+        PyObject *tmp = __Pyx_PyNumber_IntOrLong(x);
+        if (!tmp) return (long) -1;
+        val = __Pyx_PyInt_As_long(tmp);
+        Py_DECREF(tmp);
+        return val;
+    }
+raise_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "value too large to convert to long");
+    return (long) -1;
+raise_neg_overflow:
+    PyErr_SetString(PyExc_OverflowError,
+        "can't convert negative value to long");
+    return (long) -1;
+}
+
+/* FastTypeChecks */
+  #if CYTHON_COMPILING_IN_CPYTHON
+static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) {
+    while (a) {
+        a = a->tp_base;
+        if (a == b)
+            return 1;
+    }
+    return b == &PyBaseObject_Type;
+}
+static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) {
+    PyObject *mro;
+    if (a == b) return 1;
+    mro = a->tp_mro;
+    if (likely(mro)) {
+        Py_ssize_t i, n;
+        n = PyTuple_GET_SIZE(mro);
+        for (i = 0; i < n; i++) {
+            if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b)
+                return 1;
+        }
+        return 0;
+    }
+    return __Pyx_InBases(a, b);
+}
+#if PY_MAJOR_VERSION == 2
+static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) {
+    PyObject *exception, *value, *tb;
+    int res;
+    __Pyx_PyThreadState_declare
+    __Pyx_PyThreadState_assign
+    __Pyx_ErrFetch(&exception, &value, &tb);
+    res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0;
+    if (unlikely(res == -1)) {
+        PyErr_WriteUnraisable(err);
+        res = 0;
+    }
+    if (!res) {
+        res = PyObject_IsSubclass(err, exc_type2);
+        if (unlikely(res == -1)) {
+            PyErr_WriteUnraisable(err);
+            res = 0;
+        }
+    }
+    __Pyx_ErrRestore(exception, value, tb);
+    return res;
+}
+#else
+static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) {
+    int res = exc_type1 ? __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type1) : 0;
+    if (!res) {
+        res = __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2);
+    }
+    return res;
+}
+#endif
+static int __Pyx_PyErr_GivenExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) {
+    Py_ssize_t i, n;
+    assert(PyExceptionClass_Check(exc_type));
+    n = PyTuple_GET_SIZE(tuple);
+#if PY_MAJOR_VERSION >= 3
+    for (i=0; i<n; i++) {
+        if (exc_type == PyTuple_GET_ITEM(tuple, i)) return 1;
+    }
+#endif
+    for (i=0; i<n; i++) {
+        PyObject *t = PyTuple_GET_ITEM(tuple, i);
+        #if PY_MAJOR_VERSION < 3
+        if (likely(exc_type == t)) return 1;
+        #endif
+        if (likely(PyExceptionClass_Check(t))) {
+            if (__Pyx_inner_PyErr_GivenExceptionMatches2(exc_type, NULL, t)) return 1;
+        } else {
+        }
+    }
+    return 0;
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject* exc_type) {
+    if (likely(err == exc_type)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        if (likely(PyExceptionClass_Check(exc_type))) {
+            return __Pyx_inner_PyErr_GivenExceptionMatches2(err, NULL, exc_type);
+        } else if (likely(PyTuple_Check(exc_type))) {
+            return __Pyx_PyErr_GivenExceptionMatchesTuple(err, exc_type);
+        } else {
+        }
+    }
+    return PyErr_GivenExceptionMatches(err, exc_type);
+}
+static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *exc_type1, PyObject *exc_type2) {
+    assert(PyExceptionClass_Check(exc_type1));
+    assert(PyExceptionClass_Check(exc_type2));
+    if (likely(err == exc_type1 || err == exc_type2)) return 1;
+    if (likely(PyExceptionClass_Check(err))) {
+        return __Pyx_inner_PyErr_GivenExceptionMatches2(err, exc_type1, exc_type2);
+    }
+    return (PyErr_GivenExceptionMatches(err, exc_type1) || PyErr_GivenExceptionMatches(err, exc_type2));
+}
+#endif
+
+/* CheckBinaryVersion */
+  static int __Pyx_check_binary_version(void) {
+    char ctversion[5];
+    int same=1, i, found_dot;
+    const char* rt_from_call = Py_GetVersion();
+    PyOS_snprintf(ctversion, 5, "%d.%d", PY_MAJOR_VERSION, PY_MINOR_VERSION);
+    found_dot = 0;
+    for (i = 0; i < 4; i++) {
+        if (!ctversion[i]) {
+            same = (rt_from_call[i] < '0' || rt_from_call[i] > '9');
+            break;
+        }
+        if (rt_from_call[i] != ctversion[i]) {
+            same = 0;
+            break;
+        }
+    }
+    if (!same) {
+        char rtversion[5] = {'\0'};
+        char message[200];
+        for (i=0; i<4; ++i) {
+            if (rt_from_call[i] == '.') {
+                if (found_dot) break;
+                found_dot = 1;
+            } else if (rt_from_call[i] < '0' || rt_from_call[i] > '9') {
+                break;
+            }
+            rtversion[i] = rt_from_call[i];
+        }
+        PyOS_snprintf(message, sizeof(message),
+                      "compiletime version %s of module '%.100s' "
+                      "does not match runtime version %s",
+                      ctversion, __Pyx_MODULE_NAME, rtversion);
+        return PyErr_WarnEx(NULL, message, 1);
+    }
+    return 0;
+}
+
+/* InitStrings */
+  static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) {
+    while (t->p) {
+        #if PY_MAJOR_VERSION < 3
+        if (t->is_unicode) {
+            *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL);
+        } else if (t->intern) {
+            *t->p = PyString_InternFromString(t->s);
+        } else {
+            *t->p = PyString_FromStringAndSize(t->s, t->n - 1);
+        }
+        #else
+        if (t->is_unicode | t->is_str) {
+            if (t->intern) {
+                *t->p = PyUnicode_InternFromString(t->s);
+            } else if (t->encoding) {
+                *t->p = PyUnicode_Decode(t->s, t->n - 1, t->encoding, NULL);
+            } else {
+                *t->p = PyUnicode_FromStringAndSize(t->s, t->n - 1);
+            }
+        } else {
+            *t->p = PyBytes_FromStringAndSize(t->s, t->n - 1);
+        }
+        #endif
+        if (!*t->p)
+            return -1;
+        if (PyObject_Hash(*t->p) == -1)
+            return -1;
+        ++t;
+    }
+    return 0;
+}
+
+static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) {
+    return __Pyx_PyUnicode_FromStringAndSize(c_str, (Py_ssize_t)strlen(c_str));
+}
+static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) {
+    Py_ssize_t ignore;
+    return __Pyx_PyObject_AsStringAndSize(o, &ignore);
+}
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+#if !CYTHON_PEP393_ENABLED
+static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    char* defenc_c;
+    PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL);
+    if (!defenc) return NULL;
+    defenc_c = PyBytes_AS_STRING(defenc);
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    {
+        char* end = defenc_c + PyBytes_GET_SIZE(defenc);
+        char* c;
+        for (c = defenc_c; c < end; c++) {
+            if ((unsigned char) (*c) >= 128) {
+                PyUnicode_AsASCIIString(o);
+                return NULL;
+            }
+        }
+    }
+#endif
+    *length = PyBytes_GET_SIZE(defenc);
+    return defenc_c;
+}
+#else
+static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+    if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL;
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+    if (likely(PyUnicode_IS_ASCII(o))) {
+        *length = PyUnicode_GET_LENGTH(o);
+        return PyUnicode_AsUTF8(o);
+    } else {
+        PyUnicode_AsASCIIString(o);
+        return NULL;
+    }
+#else
+    return PyUnicode_AsUTF8AndSize(o, length);
+#endif
+}
+#endif
+#endif
+static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) {
+#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT
+    if (
+#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII
+            __Pyx_sys_getdefaultencoding_not_ascii &&
+#endif
+            PyUnicode_Check(o)) {
+        return __Pyx_PyUnicode_AsStringAndSize(o, length);
+    } else
+#endif
+#if (!CYTHON_COMPILING_IN_PYPY) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE))
+    if (PyByteArray_Check(o)) {
+        *length = PyByteArray_GET_SIZE(o);
+        return PyByteArray_AS_STRING(o);
+    } else
+#endif
+    {
+        char* result;
+        int r = PyBytes_AsStringAndSize(o, &result, length);
+        if (unlikely(r < 0)) {
+            return NULL;
+        } else {
+            return result;
+        }
+    }
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) {
+   int is_true = x == Py_True;
+   if (is_true | (x == Py_False) | (x == Py_None)) return is_true;
+   else return PyObject_IsTrue(x);
+}
+static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject* x) {
+    int retval;
+    if (unlikely(!x)) return -1;
+    retval = __Pyx_PyObject_IsTrue(x);
+    Py_DECREF(x);
+    return retval;
+}
+static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) {
+#if PY_MAJOR_VERSION >= 3
+    if (PyLong_Check(result)) {
+        if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1,
+                "__int__ returned non-int (type %.200s).  "
+                "The ability to return an instance of a strict subclass of int "
+                "is deprecated, and may be removed in a future version of Python.",
+                Py_TYPE(result)->tp_name)) {
+            Py_DECREF(result);
+            return NULL;
+        }
+        return result;
+    }
+#endif
+    PyErr_Format(PyExc_TypeError,
+                 "__%.4s__ returned non-%.4s (type %.200s)",
+                 type_name, type_name, Py_TYPE(result)->tp_name);
+    Py_DECREF(result);
+    return NULL;
+}
+static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) {
+#if CYTHON_USE_TYPE_SLOTS
+  PyNumberMethods *m;
+#endif
+  const char *name = NULL;
+  PyObject *res = NULL;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_Check(x) || PyLong_Check(x)))
+#else
+  if (likely(PyLong_Check(x)))
+#endif
+    return __Pyx_NewRef(x);
+#if CYTHON_USE_TYPE_SLOTS
+  m = Py_TYPE(x)->tp_as_number;
+  #if PY_MAJOR_VERSION < 3
+  if (m && m->nb_int) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  else if (m && m->nb_long) {
+    name = "long";
+    res = m->nb_long(x);
+  }
+  #else
+  if (likely(m && m->nb_int)) {
+    name = "int";
+    res = m->nb_int(x);
+  }
+  #endif
+#else
+  if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) {
+    res = PyNumber_Int(x);
+  }
+#endif
+  if (likely(res)) {
+#if PY_MAJOR_VERSION < 3
+    if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) {
+#else
+    if (unlikely(!PyLong_CheckExact(res))) {
+#endif
+        return __Pyx_PyNumber_IntOrLongWrongResultType(res, name);
+    }
+  }
+  else if (!PyErr_Occurred()) {
+    PyErr_SetString(PyExc_TypeError,
+                    "an integer is required");
+  }
+  return res;
+}
+static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) {
+  Py_ssize_t ival;
+  PyObject *x;
+#if PY_MAJOR_VERSION < 3
+  if (likely(PyInt_CheckExact(b))) {
+    if (sizeof(Py_ssize_t) >= sizeof(long))
+        return PyInt_AS_LONG(b);
+    else
+        return PyInt_AsSsize_t(b);
+  }
+#endif
+  if (likely(PyLong_CheckExact(b))) {
+    #if CYTHON_USE_PYLONG_INTERNALS
+    const digit* digits = ((PyLongObject*)b)->ob_digit;
+    const Py_ssize_t size = Py_SIZE(b);
+    if (likely(__Pyx_sst_abs(size) <= 1)) {
+        ival = likely(size) ? digits[0] : 0;
+        if (size == -1) ival = -ival;
+        return ival;
+    } else {
+      switch (size) {
+         case 2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -2:
+           if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -3:
+           if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case 4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+         case -4:
+           if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) {
+             return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0]));
+           }
+           break;
+      }
+    }
+    #endif
+    return PyLong_AsSsize_t(b);
+  }
+  x = PyNumber_Index(b);
+  if (!x) return -1;
+  ival = PyInt_AsSsize_t(x);
+  Py_DECREF(x);
+  return ival;
+}
+static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject* o) {
+  if (sizeof(Py_hash_t) == sizeof(Py_ssize_t)) {
+    return (Py_hash_t) __Pyx_PyIndex_AsSsize_t(o);
+#if PY_MAJOR_VERSION < 3
+  } else if (likely(PyInt_CheckExact(o))) {
+    return PyInt_AS_LONG(o);
+#endif
+  } else {
+    Py_ssize_t ival;
+    PyObject *x;
+    x = PyNumber_Index(o);
+    if (!x) return -1;
+    ival = PyInt_AsLong(x);
+    Py_DECREF(x);
+    return ival;
+  }
+}
+static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b) {
+  return b ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False);
+}
+static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) {
+    return PyInt_FromSize_t(ival);
+}
+
+
+#endif /* Py_PYTHON_H */
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx
new file mode 100644
index 000000000..40f589a9b
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx
@@ -0,0 +1,109 @@
+import numpy as np
+cimport numpy as np
+from libcpp.string cimport string
+
+# use the Numpy-C-API from Cython
+np.import_array()
+
+# cdefine the signature of our c function
+cdef extern from "mesh_core.h":
+    void _rasterize_triangles_core(
+        float* vertices, int* triangles, 
+        float* depth_buffer, int* triangle_buffer, float* barycentric_weight,
+        int nver, int ntri,
+        int h, int w)
+
+    void _render_colors_core(
+        float* image, float* vertices, int* triangles, 
+        float* colors, 
+        float* depth_buffer,
+        int nver, int ntri,
+        int h, int w, int c)
+
+    void _render_texture_core(
+        float* image, float* vertices, int* triangles, 
+        float* texture, float* tex_coords, int* tex_triangles, 
+        float* depth_buffer,
+        int nver, int tex_nver, int ntri, 
+        int h, int w, int c, 
+        int tex_h, int tex_w, int tex_c, 
+        int mapping_type)
+
+    void _get_normal_core(
+        float* normal, float* tri_normal, int* triangles,
+        int ntri)
+
+    void _write_obj_with_colors_texture(string filename, string mtl_name, 
+        float* vertices, int* triangles, float* colors, float* uv_coords,
+        int nver, int ntri, int ntexver)
+
+def get_normal_core(np.ndarray[float, ndim=2, mode = "c"] normal not None, 
+                np.ndarray[float, ndim=2, mode = "c"] tri_normal not None, 
+                np.ndarray[int, ndim=2, mode="c"] triangles not None, 
+                int ntri
+                ):
+    _get_normal_core(
+        <float*> np.PyArray_DATA(normal), <float*> np.PyArray_DATA(tri_normal), <int*> np.PyArray_DATA(triangles),  
+        ntri)
+
+def rasterize_triangles_core(
+                np.ndarray[float, ndim=2, mode = "c"] vertices not None, 
+                np.ndarray[int, ndim=2, mode="c"] triangles not None, 
+                np.ndarray[float, ndim=2, mode = "c"] depth_buffer not None,
+                np.ndarray[int, ndim=2, mode = "c"] triangle_buffer not None,
+                np.ndarray[float, ndim=2, mode = "c"] barycentric_weight not None,
+                int nver, int ntri,
+                int h, int w
+                ):   
+    _rasterize_triangles_core(
+        <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),  
+        <float*> np.PyArray_DATA(depth_buffer), <int*> np.PyArray_DATA(triangle_buffer), <float*> np.PyArray_DATA(barycentric_weight),
+        nver, ntri,
+        h, w)
+
+def render_colors_core(np.ndarray[float, ndim=3, mode = "c"] image not None, 
+                np.ndarray[float, ndim=2, mode = "c"] vertices not None, 
+                np.ndarray[int, ndim=2, mode="c"] triangles not None, 
+                np.ndarray[float, ndim=2, mode = "c"] colors not None, 
+                np.ndarray[float, ndim=2, mode = "c"] depth_buffer not None,
+                int nver, int ntri,
+                int h, int w, int c
+                ):   
+    _render_colors_core(
+        <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),  
+        <float*> np.PyArray_DATA(colors), 
+        <float*> np.PyArray_DATA(depth_buffer),
+        nver, ntri,
+        h, w, c)
+
+def render_texture_core(np.ndarray[float, ndim=3, mode = "c"] image not None, 
+                np.ndarray[float, ndim=2, mode = "c"] vertices not None, 
+                np.ndarray[int, ndim=2, mode="c"] triangles not None, 
+                np.ndarray[float, ndim=3, mode = "c"] texture not None, 
+                np.ndarray[float, ndim=2, mode = "c"] tex_coords not None, 
+                np.ndarray[int, ndim=2, mode="c"] tex_triangles not None, 
+                np.ndarray[float, ndim=2, mode = "c"] depth_buffer not None,
+                int nver, int tex_nver, int ntri,
+                int h, int w, int c,
+                int tex_h, int tex_w, int tex_c,
+                int mapping_type
+                ):   
+    _render_texture_core(
+        <float*> np.PyArray_DATA(image), <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles),  
+        <float*> np.PyArray_DATA(texture), <float*> np.PyArray_DATA(tex_coords), <int*> np.PyArray_DATA(tex_triangles),  
+        <float*> np.PyArray_DATA(depth_buffer),
+        nver, tex_nver, ntri,
+        h, w, c, 
+        tex_h, tex_w, tex_c, 
+        mapping_type)
+
+def write_obj_with_colors_texture_core(string filename, string mtl_name, 
+                np.ndarray[float, ndim=2, mode = "c"] vertices not None, 
+                np.ndarray[int, ndim=2, mode="c"] triangles not None, 
+                np.ndarray[float, ndim=2, mode = "c"] colors not None, 
+                np.ndarray[float, ndim=2, mode = "c"] uv_coords not None, 
+                int nver, int ntri, int ntexver
+                ):
+    _write_obj_with_colors_texture(filename, mtl_name, 
+        <float*> np.PyArray_DATA(vertices), <int*> np.PyArray_DATA(triangles), <float*> np.PyArray_DATA(colors), <float*> np.PyArray_DATA(uv_coords),
+        nver, ntri, ntexver)
diff --git a/python-package/insightface/thirdparty/face3d/mesh/cython/setup.py b/python-package/insightface/thirdparty/face3d/mesh/cython/setup.py
new file mode 100644
index 000000000..fc1ede1e7
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/cython/setup.py
@@ -0,0 +1,20 @@
+'''
+python setup.py build_ext -i
+to compile
+'''
+
+# setup.py
+from distutils.core import setup, Extension
+from Cython.Build import cythonize
+from Cython.Distutils import build_ext
+import numpy
+
+setup(
+	name = 'mesh_core_cython',
+    cmdclass={'build_ext': build_ext},
+    ext_modules=[Extension("mesh_core_cython",
+                 sources=["mesh_core_cython.pyx", "mesh_core.cpp"],
+                 language='c++',
+                 include_dirs=[numpy.get_include()])],
+)
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh/io.cpp b/python-package/insightface/thirdparty/face3d/mesh/io.cpp
new file mode 100644
index 000000000..06f2230dd
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/io.cpp
@@ -0,0 +1 @@
+#error Do not use this file, it is the result of a failed Cython compilation.
diff --git a/python-package/insightface/thirdparty/face3d/mesh/io.py b/python-package/insightface/thirdparty/face3d/mesh/io.py
new file mode 100644
index 000000000..4bb146012
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/io.py
@@ -0,0 +1,142 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import os
+from skimage import io
+from time import time
+
+from .cython import mesh_core_cython
+
+## TODO
+## TODO: c++ version
+def read_obj(obj_name):
+	''' read mesh
+	'''
+	return 0
+
+# ------------------------- write
+def write_asc(path, vertices):
+    '''
+    Args:
+        vertices: shape = (nver, 3)
+    '''
+    if path.split('.')[-1] == 'asc':
+        np.savetxt(path, vertices)
+    else:
+        np.savetxt(path + '.asc', vertices)
+
+def write_obj_with_colors(obj_name, vertices, triangles, colors):
+    ''' Save 3D face model with texture represented by colors.
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        colors: shape = (nver, 3)
+    '''
+    triangles = triangles.copy()
+    triangles += 1 # meshlab start with 1
+    
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+        
+    # write obj
+    with open(obj_name, 'w') as f:
+        
+        # write vertices & colors
+        for i in range(vertices.shape[0]):
+            # s = 'v {} {} {} \n'.format(vertices[0,i], vertices[1,i], vertices[2,i])
+            s = 'v {} {} {} {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2], colors[i, 0], colors[i, 1], colors[i, 2])
+            f.write(s)
+
+        # write f: ver ind/ uv ind
+        [k, ntri] = triangles.shape
+        for i in range(triangles.shape[0]):
+            # s = 'f {} {} {}\n'.format(triangles[i, 0], triangles[i, 1], triangles[i, 2])
+            s = 'f {} {} {}\n'.format(triangles[i, 2], triangles[i, 1], triangles[i, 0])
+            f.write(s)
+
+## TODO: c++ version
+def write_obj_with_texture(obj_name, vertices, triangles, texture, uv_coords):
+    ''' Save 3D face model with texture represented by texture map.
+    Ref: https://github.com/patrikhuber/eos/blob/bd00155ebae4b1a13b08bf5a991694d682abbada/include/eos/core/Mesh.hpp
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        texture: shape = (256,256,3)
+        uv_coords: shape = (nver, 3) max value<=1
+    '''
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+    mtl_name = obj_name.replace('.obj', '.mtl')
+    texture_name = obj_name.replace('.obj', '_texture.png')
+    
+    triangles = triangles.copy()
+    triangles += 1 # mesh lab start with 1
+    
+    # write obj
+    with open(obj_name, 'w') as f:
+        # first line: write mtlib(material library)
+        s = "mtllib {}\n".format(os.path.abspath(mtl_name))
+        f.write(s)
+
+        # write vertices
+        for i in range(vertices.shape[0]):
+            s = 'v {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2])
+            f.write(s)
+        
+        # write uv coords
+        for i in range(uv_coords.shape[0]):
+            s = 'vt {} {}\n'.format(uv_coords[i,0], 1 - uv_coords[i,1])
+            f.write(s)
+
+        f.write("usemtl FaceTexture\n")
+
+        # write f: ver ind/ uv ind
+        for i in range(triangles.shape[0]):
+            s = 'f {}/{} {}/{} {}/{}\n'.format(triangles[i,2], triangles[i,2], triangles[i,1], triangles[i,1], triangles[i,0], triangles[i,0])
+            f.write(s)
+
+    # write mtl
+    with open(mtl_name, 'w') as f:
+        f.write("newmtl FaceTexture\n")
+        s = 'map_Kd {}\n'.format(os.path.abspath(texture_name)) # map to image
+        f.write(s)
+
+    # write texture as png
+    imsave(texture_name, texture)
+
+# c++ version
+def write_obj_with_colors_texture(obj_name, vertices, triangles, colors, texture, uv_coords):
+    ''' Save 3D face model with texture. 
+    Ref: https://github.com/patrikhuber/eos/blob/bd00155ebae4b1a13b08bf5a991694d682abbada/include/eos/core/Mesh.hpp
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        colors: shape = (nver, 3)
+        texture: shape = (256,256,3)
+        uv_coords: shape = (nver, 3) max value<=1
+    '''
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+    mtl_name = obj_name.replace('.obj', '.mtl')
+    texture_name = obj_name.replace('.obj', '_texture.png')
+    
+    triangles = triangles.copy()
+    triangles += 1 # mesh lab start with 1
+    
+    # write obj
+    vertices, colors, uv_coords = vertices.astype(np.float32).copy(), colors.astype(np.float32).copy(), uv_coords.astype(np.float32).copy()
+    mesh_core_cython.write_obj_with_colors_texture_core(str.encode(obj_name), str.encode(os.path.abspath(mtl_name)), vertices, triangles, colors, uv_coords, vertices.shape[0], triangles.shape[0], uv_coords.shape[0])
+   
+    # write mtl
+    with open(mtl_name, 'w') as f:
+        f.write("newmtl FaceTexture\n")
+        s = 'map_Kd {}\n'.format(os.path.abspath(texture_name)) # map to image
+        f.write(s)
+
+    # write texture as png
+    io.imsave(texture_name, texture)
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh/light.py b/python-package/insightface/thirdparty/face3d/mesh/light.py
new file mode 100644
index 000000000..5f0da63ed
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/light.py
@@ -0,0 +1,213 @@
+'''
+Functions about lighting mesh(changing colors/texture of mesh).
+1. add light to colors/texture (shade each vertex)
+2. fit light according to colors/texture & image.
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from .cython import mesh_core_cython
+
+def get_normal(vertices, triangles):
+    ''' calculate normal direction in each vertex
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+    Returns:
+        normal: [nver, 3]
+    '''
+    pt0 = vertices[triangles[:, 0], :] # [ntri, 3]
+    pt1 = vertices[triangles[:, 1], :] # [ntri, 3]
+    pt2 = vertices[triangles[:, 2], :] # [ntri, 3]
+    tri_normal = np.cross(pt0 - pt1, pt0 - pt2) # [ntri, 3]. normal of each triangle
+
+    normal = np.zeros_like(vertices, dtype = np.float32).copy() # [nver, 3]
+    # for i in range(triangles.shape[0]):
+    #     normal[triangles[i, 0], :] = normal[triangles[i, 0], :] + tri_normal[i, :]
+    #     normal[triangles[i, 1], :] = normal[triangles[i, 1], :] + tri_normal[i, :]
+    #     normal[triangles[i, 2], :] = normal[triangles[i, 2], :] + tri_normal[i, :]
+    mesh_core_cython.get_normal_core(normal, tri_normal.astype(np.float32).copy(), triangles.copy(), triangles.shape[0])
+
+    # normalize to unit length
+    mag = np.sum(normal**2, 1) # [nver]
+    zero_ind = (mag == 0)
+    mag[zero_ind] = 1;
+    normal[zero_ind, 0] = np.ones((np.sum(zero_ind)))
+
+    normal = normal/np.sqrt(mag[:,np.newaxis])
+
+    return normal
+
+# TODO: test
+def add_light_sh(vertices, triangles, colors, sh_coeff):
+    ''' 
+    In 3d face, usually assume:
+    1. The surface of face is Lambertian(reflect only the low frequencies of lighting)
+    2. Lighting can be an arbitrary combination of point sources
+    --> can be expressed in terms of spherical harmonics(omit the lighting coefficients)
+    I = albedo * (sh(n) x sh_coeff)
+    
+    albedo: n x 1
+    sh_coeff: 9 x 1
+    Y(n) = (1, n_x, n_y, n_z, n_xn_y, n_xn_z, n_yn_z, n_x^2 - n_y^2, 3n_z^2 - 1)': n x 9 
+    # Y(n) = (1, n_x, n_y, n_z)': n x 4
+
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        colors: [nver, 3] albedo
+        sh_coeff: [9, 1] spherical harmonics coefficients
+
+    Returns:
+        lit_colors: [nver, 3]
+    '''
+    assert vertices.shape[0] == colors.shape[0]
+    nver = vertices.shape[0]
+    normal = get_normal(vertices, triangles) # [nver, 3]
+    sh = np.array((np.ones(nver), n[:,0], n[:,1], n[:,2], n[:,0]*n[:,1], n[:,0]*n[:,2], n[:,1]*n[:,2], n[:,0]**2 - n[:,1]**2, 3*(n[:,2]**2) - 1)) # [nver, 9]
+    ref = sh.dot(sh_coeff) #[nver, 1]
+    lit_colors = colors*ref
+    return lit_colors
+
+
+def add_light(vertices, triangles, colors, light_positions = 0, light_intensities = 0):
+    ''' Gouraud shading. add point lights.
+    In 3d face, usually assume:
+    1. The surface of face is Lambertian(reflect only the low frequencies of lighting)
+    2. Lighting can be an arbitrary combination of point sources
+    3. No specular (unless skin is oil, 23333)
+
+    Ref: https://cs184.eecs.berkeley.edu/lecture/pipeline    
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        light_positions: [nlight, 3] 
+        light_intensities: [nlight, 3]
+    Returns:
+        lit_colors: [nver, 3]
+    '''
+    nver = vertices.shape[0]
+    normals = get_normal(vertices, triangles) # [nver, 3]
+
+    # ambient
+    # La = ka*Ia
+
+    # diffuse
+    # Ld = kd*(I/r^2)max(0, nxl)
+    direction_to_lights = vertices[np.newaxis, :, :] - light_positions[:, np.newaxis, :] # [nlight, nver, 3]
+    direction_to_lights_n = np.sqrt(np.sum(direction_to_lights**2, axis = 2)) # [nlight, nver]
+    direction_to_lights = direction_to_lights/direction_to_lights_n[:, :, np.newaxis]
+    normals_dot_lights = normals[np.newaxis, :, :]*direction_to_lights # [nlight, nver, 3]
+    normals_dot_lights = np.sum(normals_dot_lights, axis = 2) # [nlight, nver]
+    diffuse_output = colors[np.newaxis, :, :]*normals_dot_lights[:, :, np.newaxis]*light_intensities[:, np.newaxis, :]
+    diffuse_output = np.sum(diffuse_output, axis = 0) # [nver, 3]
+    
+    # specular
+    # h = (v + l)/(|v + l|) bisector
+    # Ls = ks*(I/r^2)max(0, nxh)^p
+    # increasing p narrows the reflectionlob
+
+    lit_colors = diffuse_output # only diffuse part here.
+    lit_colors = np.minimum(np.maximum(lit_colors, 0), 1)
+    return lit_colors
+
+
+
+## TODO. estimate light(sh coeff)
+## -------------------------------- estimate. can not use now. 
+def fit_light(image, vertices, colors, triangles, vis_ind, lamb = 10, max_iter = 3):
+    [h, w, c] = image.shape
+
+    # surface normal
+    norm = get_normal(vertices, triangles)
+    
+    nver = vertices.shape[1]
+
+    # vertices --> corresponding image pixel
+    pt2d = vertices[:2, :]
+
+    pt2d[0,:] = np.minimum(np.maximum(pt2d[0,:], 0), w - 1)
+    pt2d[1,:] = np.minimum(np.maximum(pt2d[1,:], 0), h - 1)
+    pt2d = np.round(pt2d).astype(np.int32) # 2 x nver
+
+    image_pixel = image[pt2d[1,:], pt2d[0,:], :] # nver x 3
+    image_pixel = image_pixel.T # 3 x nver
+
+    # vertices --> corresponding mean texture pixel with illumination
+    # Spherical Harmonic Basis
+    harmonic_dim = 9
+    nx = norm[0,:];
+    ny = norm[1,:];
+    nz = norm[2,:];
+    harmonic = np.zeros((nver, harmonic_dim))
+
+    pi = np.pi
+    harmonic[:,0] = np.sqrt(1/(4*pi)) * np.ones((nver,));
+    harmonic[:,1] = np.sqrt(3/(4*pi)) * nx;
+    harmonic[:,2] = np.sqrt(3/(4*pi)) * ny;
+    harmonic[:,3] = np.sqrt(3/(4*pi)) * nz;
+    harmonic[:,4] = 1/2. * np.sqrt(3/(4*pi)) * (2*nz**2 - nx**2 - ny**2);
+    harmonic[:,5] = 3 * np.sqrt(5/(12*pi)) * (ny*nz);
+    harmonic[:,6] = 3 * np.sqrt(5/(12*pi)) * (nx*nz);
+    harmonic[:,7] = 3 * np.sqrt(5/(12*pi)) * (nx*ny);
+    harmonic[:,8] = 3/2. * np.sqrt(5/(12*pi)) * (nx*nx - ny*ny);
+    
+    '''
+    I' = sum(albedo * lj * hj) j = 0:9 (albedo = tex)
+    set A = albedo*h (n x 9)
+        alpha = lj (9 x 1)
+        Y = I (n x 1)
+        Y' = A.dot(alpha)
+
+    opt function:
+        ||Y - A*alpha|| + lambda*(alpha'*alpha)
+    result:
+        A'*(Y - A*alpha) + lambda*alpha = 0
+        ==>
+        (A'*A*alpha - lambda)*alpha = A'*Y
+        left: 9 x 9
+        right: 9 x 1
+    '''
+    n_vis_ind = len(vis_ind)
+    n = n_vis_ind*c
+
+    Y = np.zeros((n, 1))
+    A = np.zeros((n, 9))
+    light = np.zeros((3, 1))
+
+    for k in range(c):
+        Y[k*n_vis_ind:(k+1)*n_vis_ind, :] = image_pixel[k, vis_ind][:, np.newaxis]
+        A[k*n_vis_ind:(k+1)*n_vis_ind, :] = texture[k, vis_ind][:, np.newaxis] * harmonic[vis_ind, :]
+        Ac = texture[k, vis_ind][:, np.newaxis]
+        Yc = image_pixel[k, vis_ind][:, np.newaxis]
+        light[k] = (Ac.T.dot(Yc))/(Ac.T.dot(Ac))
+
+    for i in range(max_iter):
+
+        Yc = Y.copy()
+        for k in range(c):
+            Yc[k*n_vis_ind:(k+1)*n_vis_ind, :]  /= light[k]
+
+        # update alpha
+        equation_left = np.dot(A.T, A) + lamb*np.eye(harmonic_dim); # why + ?
+        equation_right = np.dot(A.T, Yc) 
+        alpha = np.dot(np.linalg.inv(equation_left), equation_right)
+
+        # update light
+        for k in range(c):
+            Ac = A[k*n_vis_ind:(k+1)*n_vis_ind, :].dot(alpha)
+            Yc = Y[k*n_vis_ind:(k+1)*n_vis_ind, :]
+            light[k] = (Ac.T.dot(Yc))/(Ac.T.dot(Ac))
+
+    appearance = np.zeros_like(texture)
+    for k in range(c):
+        tmp = np.dot(harmonic*texture[k, :][:, np.newaxis], alpha*light[k])
+        appearance[k,:] = tmp.T
+
+    appearance = np.minimum(np.maximum(appearance, 0), 1)
+
+    return appearance
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh/render.py b/python-package/insightface/thirdparty/face3d/mesh/render.py
new file mode 100644
index 000000000..19957222e
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/render.py
@@ -0,0 +1,135 @@
+'''
+functions about rendering mesh(from 3d obj to 2d image).
+only use rasterization render here.
+Note that:
+1. Generally, render func includes camera, light, raterize. Here no camera and light(I write these in other files)
+2. Generally, the input vertices are normalized to [-1,1] and cetered on [0, 0]. (in world space)
+   Here, the vertices are using image coords, which centers on [w/2, h/2] with the y-axis pointing to oppisite direction.
+ Means: render here only conducts interpolation.(I just want to make the input flexible)
+
+Author: Yao Feng 
+Mail: yaofeng1995@gmail.com
+'''
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from time import time
+
+from .cython import mesh_core_cython
+
+def rasterize_triangles(vertices, triangles, h, w):
+    ''' 
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        h: height
+        w: width
+    Returns:
+        depth_buffer: [h, w] saves the depth, here, the bigger the z, the fronter the point.
+        triangle_buffer: [h, w] saves the tri id(-1 for no triangle). 
+        barycentric_weight: [h, w, 3] saves corresponding barycentric weight.
+
+    # Each triangle has 3 vertices & Each vertex has 3 coordinates x, y, z.
+    # h, w is the size of rendering
+    '''
+
+    # initial 
+    depth_buffer = np.zeros([h, w]) - 999999. #set the initial z to the farest position
+    triangle_buffer = np.zeros([h, w], dtype = np.int32) - 1  # if tri id = -1, the pixel has no triangle correspondance
+    barycentric_weight = np.zeros([h, w, 3], dtype = np.float32)  # 
+    
+    vertices = vertices.astype(np.float32).copy()
+    triangles = triangles.astype(np.int32).copy()
+
+    mesh_core_cython.rasterize_triangles_core(
+                vertices, triangles,
+                depth_buffer, triangle_buffer, barycentric_weight, 
+                vertices.shape[0], triangles.shape[0], 
+                h, w)
+
+def render_colors(vertices, triangles, colors, h, w, c = 3, BG = None):
+    ''' render mesh with colors
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3] 
+        colors: [nver, 3]
+        h: height
+        w: width  
+        c: channel
+        BG: background image
+    Returns:
+        image: [h, w, c]. rendered image./rendering.
+    '''
+
+    # initial 
+    if BG is None:
+        image = np.zeros((h, w, c), dtype = np.float32)
+    else:
+        assert BG.shape[0] == h and BG.shape[1] == w and BG.shape[2] == c
+        image = BG
+    depth_buffer = np.zeros([h, w], dtype = np.float32, order = 'C') - 999999.
+
+    # change orders. --> C-contiguous order(column major)
+    vertices = vertices.astype(np.float32).copy()
+    triangles = triangles.astype(np.int32).copy()
+    colors = colors.astype(np.float32).copy()
+    ###
+    st = time()
+    mesh_core_cython.render_colors_core(
+                image, vertices, triangles,
+                colors,
+                depth_buffer,
+                vertices.shape[0], triangles.shape[0], 
+                h, w, c)
+    return image
+
+
+def render_texture(vertices, triangles, texture, tex_coords, tex_triangles, h, w, c = 3, mapping_type = 'nearest', BG = None):
+    ''' render mesh with texture map
+    Args:
+        vertices: [3, nver]
+        triangles: [3, ntri]
+        texture: [tex_h, tex_w, 3]
+        tex_coords: [ntexcoords, 3]
+        tex_triangles: [ntri, 3]
+        h: height of rendering
+        w: width of rendering
+        c: channel
+        mapping_type: 'bilinear' or 'nearest'
+    '''
+    # initial 
+    if BG is None:
+        image = np.zeros((h, w, c), dtype = np.float32)
+    else:
+        assert BG.shape[0] == h and BG.shape[1] == w and BG.shape[2] == c
+        image = BG
+
+    depth_buffer = np.zeros([h, w], dtype = np.float32, order = 'C') - 999999.
+    
+    tex_h, tex_w, tex_c = texture.shape
+    if mapping_type == 'nearest':
+        mt = int(0)
+    elif mapping_type == 'bilinear':
+        mt = int(1)
+    else:
+        mt = int(0)
+    
+    # -> C order
+    vertices = vertices.astype(np.float32).copy()
+    triangles = triangles.astype(np.int32).copy()
+    texture = texture.astype(np.float32).copy()
+    tex_coords = tex_coords.astype(np.float32).copy()
+    tex_triangles = tex_triangles.astype(np.int32).copy()
+
+    mesh_core_cython.render_texture_core(
+                image, vertices, triangles,
+                texture, tex_coords, tex_triangles,
+                depth_buffer,
+                vertices.shape[0], tex_coords.shape[0], triangles.shape[0], 
+                h, w, c,
+                tex_h, tex_w, tex_c,
+                mt)
+    return image
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh/transform.py b/python-package/insightface/thirdparty/face3d/mesh/transform.py
new file mode 100644
index 000000000..d91b09fae
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/transform.py
@@ -0,0 +1,383 @@
+'''
+Functions about transforming mesh(changing the position: modify vertices).
+1. forward: transform(transform, camera, project).
+2. backward: estimate transform matrix from correspondences.
+
+Author: Yao Feng 
+Mail: yaofeng1995@gmail.com
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import math
+from math import cos, sin
+
+def angle2matrix(angles):
+    ''' get rotation matrix from three rotation angles(degree). right-handed.
+    Args:
+        angles: [3,]. x, y, z angles
+        x: pitch. positive for looking down.
+        y: yaw. positive for looking left. 
+        z: roll. positive for tilting head right. 
+    Returns:
+        R: [3, 3]. rotation matrix.
+    '''
+    x, y, z = np.deg2rad(angles[0]), np.deg2rad(angles[1]), np.deg2rad(angles[2])
+    # x
+    Rx=np.array([[1,      0,       0],
+                 [0, cos(x),  -sin(x)],
+                 [0, sin(x),   cos(x)]])
+    # y
+    Ry=np.array([[ cos(y), 0, sin(y)],
+                 [      0, 1,      0],
+                 [-sin(y), 0, cos(y)]])
+    # z
+    Rz=np.array([[cos(z), -sin(z), 0],
+                 [sin(z),  cos(z), 0],
+                 [     0,       0, 1]])
+    
+    R=Rz.dot(Ry.dot(Rx))
+    return R.astype(np.float32)
+
+def angle2matrix_3ddfa(angles):
+    ''' get rotation matrix from three rotation angles(radian). The same as in 3DDFA.
+    Args:
+        angles: [3,]. x, y, z angles
+        x: pitch.
+        y: yaw. 
+        z: roll. 
+    Returns:
+        R: 3x3. rotation matrix.
+    '''
+    # x, y, z = np.deg2rad(angles[0]), np.deg2rad(angles[1]), np.deg2rad(angles[2])
+    x, y, z = angles[0], angles[1], angles[2]
+    
+    # x
+    Rx=np.array([[1,      0,       0],
+                 [0, cos(x),  sin(x)],
+                 [0, -sin(x),   cos(x)]])
+    # y
+    Ry=np.array([[ cos(y), 0, -sin(y)],
+                 [      0, 1,      0],
+                 [sin(y), 0, cos(y)]])
+    # z
+    Rz=np.array([[cos(z), sin(z), 0],
+                 [-sin(z),  cos(z), 0],
+                 [     0,       0, 1]])
+    R = Rx.dot(Ry).dot(Rz)
+    return R.astype(np.float32)
+
+
+## ------------------------------------------ 1. transform(transform, project, camera).
+## ---------- 3d-3d transform. Transform obj in world space
+def rotate(vertices, angles):
+    ''' rotate vertices. 
+    X_new = R.dot(X). X: 3 x 1   
+    Args:
+        vertices: [nver, 3]. 
+        rx, ry, rz: degree angles
+        rx: pitch. positive for looking down 
+        ry: yaw. positive for looking left
+        rz: roll. positive for tilting head right
+    Returns:
+        rotated vertices: [nver, 3]
+    '''
+    R = angle2matrix(angles)
+    rotated_vertices = vertices.dot(R.T)
+
+    return rotated_vertices
+
+def similarity_transform(vertices, s, R, t3d):
+    ''' similarity transform. dof = 7.
+    3D: s*R.dot(X) + t
+    Homo: M = [[sR, t],[0^T, 1]].  M.dot(X)
+    Args:(float32)
+        vertices: [nver, 3]. 
+        s: [1,]. scale factor.
+        R: [3,3]. rotation matrix.
+        t3d: [3,]. 3d translation vector.
+    Returns:
+        transformed vertices: [nver, 3]
+    '''
+    t3d = np.squeeze(np.array(t3d, dtype = np.float32))
+    transformed_vertices = s * vertices.dot(R.T) + t3d[np.newaxis, :]
+
+    return transformed_vertices
+
+
+## -------------- Camera. from world space to camera space
+# Ref: https://cs184.eecs.berkeley.edu/lecture/transforms-2
+def normalize(x):
+    epsilon = 1e-12
+    norm = np.sqrt(np.sum(x**2, axis = 0))
+    norm = np.maximum(norm, epsilon)
+    return x/norm
+
+def lookat_camera(vertices, eye, at = None, up = None):
+    """ 'look at' transformation: from world space to camera space
+    standard camera space: 
+        camera located at the origin. 
+        looking down negative z-axis. 
+        vertical vector is y-axis.
+    Xcam = R(X - C)
+    Homo: [[R, -RC], [0, 1]]
+    Args:
+      vertices: [nver, 3] 
+      eye: [3,] the XYZ world space position of the camera.
+      at: [3,] a position along the center of the camera's gaze.
+      up: [3,] up direction 
+    Returns:
+      transformed_vertices: [nver, 3]
+    """
+    if at is None:
+      at = np.array([0, 0, 0], np.float32)
+    if up is None:
+      up = np.array([0, 1, 0], np.float32)
+
+    eye = np.array(eye).astype(np.float32)
+    at = np.array(at).astype(np.float32)
+    z_aixs = -normalize(at - eye) # look forward
+    x_aixs = normalize(np.cross(up, z_aixs)) # look right
+    y_axis = np.cross(z_aixs, x_aixs) # look up
+
+    R = np.stack((x_aixs, y_axis, z_aixs))#, axis = 0) # 3 x 3
+    transformed_vertices = vertices - eye # translation
+    transformed_vertices = transformed_vertices.dot(R.T) # rotation
+    return transformed_vertices
+
+## --------- 3d-2d project. from camera space to image plane
+# generally, image plane only keeps x,y channels, here reserve z channel for calculating z-buffer.
+def orthographic_project(vertices):
+    ''' scaled orthographic projection(just delete z)
+        assumes: variations in depth over the object is small relative to the mean distance from camera to object
+        x -> x*f/z, y -> x*f/z, z -> f.
+        for point i,j. zi~=zj. so just delete z
+        ** often used in face
+        Homo: P = [[1,0,0,0], [0,1,0,0], [0,0,1,0]]
+    Args:
+        vertices: [nver, 3]
+    Returns:
+        projected_vertices: [nver, 3] if isKeepZ=True. [nver, 2] if isKeepZ=False.
+    '''
+    return vertices.copy()
+
+def perspective_project(vertices, fovy, aspect_ratio = 1., near = 0.1, far = 1000.):
+    ''' perspective projection.
+    Args:
+        vertices: [nver, 3]
+        fovy: vertical angular field of view. degree.
+        aspect_ratio : width / height of field of view
+        near : depth of near clipping plane
+        far : depth of far clipping plane
+    Returns:
+        projected_vertices: [nver, 3] 
+    '''
+    fovy = np.deg2rad(fovy)
+    top = near*np.tan(fovy)
+    bottom = -top 
+    right = top*aspect_ratio
+    left = -right
+
+    #-- homo
+    P = np.array([[near/right, 0, 0, 0],
+                 [0, near/top, 0, 0],
+                 [0, 0, -(far+near)/(far-near), -2*far*near/(far-near)],
+                 [0, 0, -1, 0]])
+    vertices_homo = np.hstack((vertices, np.ones((vertices.shape[0], 1)))) # [nver, 4]
+    projected_vertices = vertices_homo.dot(P.T)
+    projected_vertices = projected_vertices/projected_vertices[:,3:]
+    projected_vertices = projected_vertices[:,:3]
+    projected_vertices[:,2] = -projected_vertices[:,2]
+
+    #-- non homo. only fovy
+    # projected_vertices = vertices.copy()
+    # projected_vertices[:,0] = -(near/right)*vertices[:,0]/vertices[:,2]
+    # projected_vertices[:,1] = -(near/top)*vertices[:,1]/vertices[:,2]
+    return projected_vertices
+
+
+def to_image(vertices, h, w, is_perspective = False):
+    ''' change vertices to image coord system
+    3d system: XYZ, center(0, 0, 0)
+    2d image: x(u), y(v). center(w/2, h/2), flip y-axis. 
+    Args:
+        vertices: [nver, 3]
+        h: height of the rendering
+        w : width of the rendering
+    Returns:
+        projected_vertices: [nver, 3]  
+    '''
+    image_vertices = vertices.copy()
+    if is_perspective:
+        # if perspective, the projected vertices are normalized to [-1, 1]. so change it to image size first.
+        image_vertices[:,0] = image_vertices[:,0]*w/2
+        image_vertices[:,1] = image_vertices[:,1]*h/2
+    # move to center of image
+    image_vertices[:,0] = image_vertices[:,0] + w/2
+    image_vertices[:,1] = image_vertices[:,1] + h/2
+    # flip vertices along y-axis.
+    image_vertices[:,1] = h - image_vertices[:,1] - 1
+    return image_vertices
+
+
+#### -------------------------------------------2. estimate transform matrix from correspondences.
+def estimate_affine_matrix_3d23d(X, Y):
+    ''' Using least-squares solution 
+    Args:
+        X: [n, 3]. 3d points(fixed)
+        Y: [n, 3]. corresponding 3d points(moving). Y = PX
+    Returns:
+        P_Affine: (3, 4). Affine camera matrix (the third row is [0, 0, 0, 1]).
+    '''
+    X_homo = np.hstack((X, np.ones([X.shape[1],1]))) #n x 4
+    P = np.linalg.lstsq(X_homo, Y)[0].T # Affine matrix. 3 x 4
+    return P
+    
+def estimate_affine_matrix_3d22d(X, x):
+    ''' Using Golden Standard Algorithm for estimating an affine camera
+        matrix P from world to image correspondences.
+        See Alg.7.2. in MVGCV 
+        Code Ref: https://github.com/patrikhuber/eos/blob/master/include/eos/fitting/affine_camera_estimation.hpp
+        x_homo = X_homo.dot(P_Affine)
+    Args:
+        X: [n, 3]. corresponding 3d points(fixed)
+        x: [n, 2]. n>=4. 2d points(moving). x = PX
+    Returns:
+        P_Affine: [3, 4]. Affine camera matrix
+    '''
+    X = X.T; x = x.T
+    assert(x.shape[1] == X.shape[1])
+    n = x.shape[1]
+    assert(n >= 4)
+
+    #--- 1. normalization
+    # 2d points
+    mean = np.mean(x, 1) # (2,)
+    x = x - np.tile(mean[:, np.newaxis], [1, n])
+    average_norm = np.mean(np.sqrt(np.sum(x**2, 0)))
+    scale = np.sqrt(2) / average_norm
+    x = scale * x
+
+    T = np.zeros((3,3), dtype = np.float32)
+    T[0, 0] = T[1, 1] = scale
+    T[:2, 2] = -mean*scale
+    T[2, 2] = 1
+
+    # 3d points
+    X_homo = np.vstack((X, np.ones((1, n))))
+    mean = np.mean(X, 1) # (3,)
+    X = X - np.tile(mean[:, np.newaxis], [1, n])
+    m = X_homo[:3,:] - X
+    average_norm = np.mean(np.sqrt(np.sum(X**2, 0)))
+    scale = np.sqrt(3) / average_norm
+    X = scale * X
+
+    U = np.zeros((4,4), dtype = np.float32)
+    U[0, 0] = U[1, 1] = U[2, 2] = scale
+    U[:3, 3] = -mean*scale
+    U[3, 3] = 1
+
+    # --- 2. equations
+    A = np.zeros((n*2, 8), dtype = np.float32);
+    X_homo = np.vstack((X, np.ones((1, n)))).T
+    A[:n, :4] = X_homo
+    A[n:, 4:] = X_homo
+    b = np.reshape(x, [-1, 1])
+ 
+    # --- 3. solution
+    p_8 = np.linalg.pinv(A).dot(b)
+    P = np.zeros((3, 4), dtype = np.float32)
+    P[0, :] = p_8[:4, 0]
+    P[1, :] = p_8[4:, 0]
+    P[-1, -1] = 1
+
+    # --- 4. denormalization
+    P_Affine = np.linalg.inv(T).dot(P.dot(U))
+    return P_Affine
+
+def P2sRt(P):
+    ''' decompositing camera matrix P
+    Args: 
+        P: (3, 4). Affine Camera Matrix.
+    Returns:
+        s: scale factor.
+        R: (3, 3). rotation matrix.
+        t: (3,). translation. 
+    '''
+    t = P[:, 3]
+    R1 = P[0:1, :3]
+    R2 = P[1:2, :3]
+    s = (np.linalg.norm(R1) + np.linalg.norm(R2))/2.0
+    r1 = R1/np.linalg.norm(R1)
+    r2 = R2/np.linalg.norm(R2)
+    r3 = np.cross(r1, r2)
+
+    R = np.concatenate((r1, r2, r3), 0)
+    return s, R, t
+
+#Ref: https://www.learnopencv.com/rotation-matrix-to-euler-angles/
+def isRotationMatrix(R):
+    ''' checks if a matrix is a valid rotation matrix(whether orthogonal or not)
+    '''
+    Rt = np.transpose(R)
+    shouldBeIdentity = np.dot(Rt, R)
+    I = np.identity(3, dtype = R.dtype)
+    n = np.linalg.norm(I - shouldBeIdentity)
+    return n < 1e-6
+
+def matrix2angle(R):
+    ''' get three Euler angles from Rotation Matrix
+    Args:
+        R: (3,3). rotation matrix
+    Returns:
+        x: pitch
+        y: yaw
+        z: roll
+    '''
+    assert(isRotationMatrix)
+    sy = math.sqrt(R[0,0] * R[0,0] +  R[1,0] * R[1,0])
+     
+    singular = sy < 1e-6
+ 
+    if  not singular :
+        x = math.atan2(R[2,1] , R[2,2])
+        y = math.atan2(-R[2,0], sy)
+        z = math.atan2(R[1,0], R[0,0])
+    else :
+        x = math.atan2(-R[1,2], R[1,1])
+        y = math.atan2(-R[2,0], sy)
+        z = 0
+
+    # rx, ry, rz = np.rad2deg(x), np.rad2deg(y), np.rad2deg(z)
+    rx, ry, rz = x*180/np.pi, y*180/np.pi, z*180/np.pi
+    return rx, ry, rz
+
+# def matrix2angle(R):
+#     ''' compute three Euler angles from a Rotation Matrix. Ref: http://www.gregslabaugh.net/publications/euler.pdf
+#     Args:
+#         R: (3,3). rotation matrix
+#     Returns:
+#         x: yaw
+#         y: pitch
+#         z: roll
+#     '''
+#     # assert(isRotationMatrix(R))
+
+#     if R[2,0] !=1 or R[2,0] != -1:
+#         x = math.asin(R[2,0])
+#         y = math.atan2(R[2,1]/cos(x), R[2,2]/cos(x))
+#         z = math.atan2(R[1,0]/cos(x), R[0,0]/cos(x))
+        
+#     else:# Gimbal lock
+#         z = 0 #can be anything
+#         if R[2,0] == -1:
+#             x = np.pi/2
+#             y = z + math.atan2(R[0,1], R[0,2])
+#         else:
+#             x = -np.pi/2
+#             y = -z + math.atan2(-R[0,1], -R[0,2])
+
+#     return x, y, z
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh/vis.py b/python-package/insightface/thirdparty/face3d/mesh/vis.py
new file mode 100644
index 000000000..9db972f2c
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh/vis.py
@@ -0,0 +1,24 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import matplotlib.pyplot as plt
+from skimage import measure
+from mpl_toolkits.mplot3d import Axes3D
+
+def plot_mesh(vertices, triangles, subplot = [1,1,1], title = 'mesh', el = 90, az = -90, lwdt=.1, dist = 6, color = "grey"):
+	'''
+	plot the mesh 
+	Args:
+		vertices: [nver, 3]
+		triangles: [ntri, 3]
+	'''
+	ax = plt.subplot(subplot[0], subplot[1], subplot[2], projection = '3d')
+	ax.plot_trisurf(vertices[:, 0], vertices[:, 1], vertices[:, 2], triangles = triangles, lw = lwdt, color = color, alpha = 1)
+	ax.axis("off")
+	ax.view_init(elev = el, azim = az)
+	ax.dist = dist
+	plt.title(title)
+
+### -------------- Todo: use vtk to visualize mesh? or visvis? or VisPy?
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/__init__.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/__init__.py
new file mode 100644
index 000000000..65d503903
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/__init__.py
@@ -0,0 +1,10 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from . import io
+from . import vis
+from . import transform
+from . import light
+from . import render
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/io.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/io.py
new file mode 100644
index 000000000..0e48a76d2
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/io.py
@@ -0,0 +1,170 @@
+''' io: read&write mesh
+1. read obj as array(TODO)
+2. write arrays to obj
+
+Preparation knowledge:
+representations of 3d face: mesh, point cloud...
+storage format: obj, ply, bin, asc, mat...
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import os
+from skimage import io
+
+## TODO
+## TODO: c++ version
+def read_obj(obj_name):
+	''' read mesh
+	'''
+	return 0
+
+# ------------------------- write
+def write_asc(path, vertices):
+    '''
+    Args:
+        vertices: shape = (nver, 3)
+    '''
+    if path.split('.')[-1] == 'asc':
+        np.savetxt(path, vertices)
+    else:
+        np.savetxt(path + '.asc', vertices)
+
+def write_obj_with_colors(obj_name, vertices, triangles, colors):
+    ''' Save 3D face model with texture represented by colors.
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        colors: shape = (nver, 3)
+    '''
+    triangles = triangles.copy()
+    triangles += 1 # meshlab start with 1
+    
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+        
+    # write obj
+    with open(obj_name, 'w') as f:
+        
+        # write vertices & colors
+        for i in range(vertices.shape[0]):
+            # s = 'v {} {} {} \n'.format(vertices[0,i], vertices[1,i], vertices[2,i])
+            s = 'v {} {} {} {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2], colors[i, 0], colors[i, 1], colors[i, 2])
+            f.write(s)
+
+        # write f: ver ind/ uv ind
+        [k, ntri] = triangles.shape
+        for i in range(triangles.shape[0]):
+            # s = 'f {} {} {}\n'.format(triangles[i, 0], triangles[i, 1], triangles[i, 2])
+            s = 'f {} {} {}\n'.format(triangles[i, 2], triangles[i, 1], triangles[i, 0])
+            f.write(s)
+
+## TODO: c++ version
+def write_obj_with_texture(obj_name, vertices, triangles, texture, uv_coords):
+    ''' Save 3D face model with texture represented by texture map.
+    Ref: https://github.com/patrikhuber/eos/blob/bd00155ebae4b1a13b08bf5a991694d682abbada/include/eos/core/Mesh.hpp
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        texture: shape = (256,256,3)
+        uv_coords: shape = (nver, 3) max value<=1
+    '''
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+    mtl_name = obj_name.replace('.obj', '.mtl')
+    texture_name = obj_name.replace('.obj', '_texture.png')
+    
+    triangles = triangles.copy()
+    triangles += 1 # mesh lab start with 1
+    
+    # write obj
+    with open(obj_name, 'w') as f:
+        # first line: write mtlib(material library)
+        s = "mtllib {}\n".format(os.path.abspath(mtl_name))
+        f.write(s)
+
+        # write vertices
+        for i in range(vertices.shape[0]):
+            s = 'v {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2])
+            f.write(s)
+        
+        # write uv coords
+        for i in range(uv_coords.shape[0]):
+            # s = 'vt {} {}\n'.format(uv_coords[i,0], 1 - uv_coords[i,1])
+            s = 'vt {} {}\n'.format(uv_coords[i,0], uv_coords[i,1])
+            f.write(s)
+
+        f.write("usemtl FaceTexture\n")
+
+        # write f: ver ind/ uv ind
+        for i in range(triangles.shape[0]):
+            s = 'f {}/{} {}/{} {}/{}\n'.format(triangles[i,2], triangles[i,2], triangles[i,1], triangles[i,1], triangles[i,0], triangles[i,0])
+            f.write(s)
+
+    # write mtl
+    with open(mtl_name, 'w') as f:
+        f.write("newmtl FaceTexture\n")
+        s = 'map_Kd {}\n'.format(os.path.abspath(texture_name)) # map to image
+        f.write(s)
+
+    # write texture as png
+    imsave(texture_name, texture)
+
+
+def write_obj_with_colors_texture(obj_name, vertices, triangles, colors, texture, uv_coords):
+    ''' Save 3D face model with texture. 
+    Ref: https://github.com/patrikhuber/eos/blob/bd00155ebae4b1a13b08bf5a991694d682abbada/include/eos/core/Mesh.hpp
+    Args:
+        obj_name: str
+        vertices: shape = (nver, 3)
+        triangles: shape = (ntri, 3)
+        colors: shape = (nver, 3)
+        texture: shape = (256,256,3)
+        uv_coords: shape = (nver, 3) max value<=1
+    '''
+    if obj_name.split('.')[-1] != 'obj':
+        obj_name = obj_name + '.obj'
+    mtl_name = obj_name.replace('.obj', '.mtl')
+    texture_name = obj_name.replace('.obj', '_texture.png')
+    
+    triangles = triangles.copy()
+    triangles += 1 # mesh lab start with 1
+    
+    # write obj
+    with open(obj_name, 'w') as f:
+        # first line: write mtlib(material library)
+        s = "mtllib {}\n".format(os.path.abspath(mtl_name))
+        f.write(s)
+
+        # write vertices
+        for i in range(vertices.shape[0]):
+            s = 'v {} {} {} {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2], colors[i, 0], colors[i, 1], colors[i, 2])
+            f.write(s)
+        
+        # write uv coords
+        for i in range(uv_coords.shape[0]):
+            # s = 'vt {} {}\n'.format(uv_coords[i,0], 1 - uv_coords[i,1])
+            s = 'vt {} {}\n'.format(uv_coords[i,0], uv_coords[i,1])
+            f.write(s)
+
+        f.write("usemtl FaceTexture\n")
+
+        # write f: ver ind/ uv ind
+        for i in range(triangles.shape[0]):
+            # s = 'f {}/{} {}/{} {}/{}\n'.format(triangles[i,0], triangles[i,0], triangles[i,1], triangles[i,1], triangles[i,2], triangles[i,2])
+            s = 'f {}/{} {}/{} {}/{}\n'.format(triangles[i,2], triangles[i,2], triangles[i,1], triangles[i,1], triangles[i,0], triangles[i,0])
+            f.write(s)
+
+    # write mtl
+    with open(mtl_name, 'w') as f:
+        f.write("newmtl FaceTexture\n")
+        s = 'map_Kd {}\n'.format(os.path.abspath(texture_name)) # map to image
+        f.write(s)
+
+    # write texture as png
+    io.imsave(texture_name, texture)
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/light.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/light.py
new file mode 100644
index 000000000..bde57114b
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/light.py
@@ -0,0 +1,215 @@
+'''
+Functions about lighting mesh(changing colors/texture of mesh).
+1. add light to colors/texture (shade each vertex)
+2. fit light according to colors/texture & image.
+
+Preparation knowledge:
+lighting: https://cs184.eecs.berkeley.edu/lecture/pipeline
+spherical harmonics in human face: '3D Face Reconstruction from a Single Image Using a Single Reference Face Shape'
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+def get_normal(vertices, triangles):
+    ''' calculate normal direction in each vertex
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+    Returns:
+        normal: [nver, 3]
+    '''
+    pt0 = vertices[triangles[:, 0], :] # [ntri, 3]
+    pt1 = vertices[triangles[:, 1], :] # [ntri, 3]
+    pt2 = vertices[triangles[:, 2], :] # [ntri, 3]
+    tri_normal = np.cross(pt0 - pt1, pt0 - pt2) # [ntri, 3]. normal of each triangle
+
+    normal = np.zeros_like(vertices) # [nver, 3]
+    for i in range(triangles.shape[0]):
+        normal[triangles[i, 0], :] = normal[triangles[i, 0], :] + tri_normal[i, :]
+        normal[triangles[i, 1], :] = normal[triangles[i, 1], :] + tri_normal[i, :]
+        normal[triangles[i, 2], :] = normal[triangles[i, 2], :] + tri_normal[i, :]
+    
+    # normalize to unit length
+    mag = np.sum(normal**2, 1) # [nver]
+    zero_ind = (mag == 0)
+    mag[zero_ind] = 1;
+    normal[zero_ind, 0] = np.ones((np.sum(zero_ind)))
+
+    normal = normal/np.sqrt(mag[:,np.newaxis])
+
+    return normal
+
+# TODO: test
+def add_light_sh(vertices, triangles, colors, sh_coeff):
+    ''' 
+    In 3d face, usually assume:
+    1. The surface of face is Lambertian(reflect only the low frequencies of lighting)
+    2. Lighting can be an arbitrary combination of point sources
+    --> can be expressed in terms of spherical harmonics(omit the lighting coefficients)
+    I = albedo * (sh(n) x sh_coeff)
+    
+    albedo: n x 1
+    sh_coeff: 9 x 1
+    Y(n) = (1, n_x, n_y, n_z, n_xn_y, n_xn_z, n_yn_z, n_x^2 - n_y^2, 3n_z^2 - 1)': n x 9 
+    # Y(n) = (1, n_x, n_y, n_z)': n x 4
+
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        colors: [nver, 3] albedo
+        sh_coeff: [9, 1] spherical harmonics coefficients
+
+    Returns:
+        lit_colors: [nver, 3]
+    '''
+    assert vertices.shape[0] == colors.shape[0]
+    nver = vertices.shape[0]
+    normal = get_normal(vertices, triangles) # [nver, 3]
+    sh = np.array((np.ones(nver), n[:,0], n[:,1], n[:,2], n[:,0]*n[:,1], n[:,0]*n[:,2], n[:,1]*n[:,2], n[:,0]**2 - n[:,1]**2, 3*(n[:,2]**2) - 1)) # [nver, 9]
+    ref = sh.dot(sh_coeff) #[nver, 1]
+    lit_colors = colors*ref
+    return lit_colors
+
+
+def add_light(vertices, triangles, colors, light_positions = 0, light_intensities = 0):
+    ''' Gouraud shading. add point lights.
+    In 3d face, usually assume:
+    1. The surface of face is Lambertian(reflect only the low frequencies of lighting)
+    2. Lighting can be an arbitrary combination of point sources
+    3. No specular (unless skin is oil, 23333)
+
+    Ref: https://cs184.eecs.berkeley.edu/lecture/pipeline    
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        light_positions: [nlight, 3] 
+        light_intensities: [nlight, 3]
+    Returns:
+        lit_colors: [nver, 3]
+    '''
+    nver = vertices.shape[0]
+    normals = get_normal(vertices, triangles) # [nver, 3]
+
+    # ambient
+    # La = ka*Ia
+
+    # diffuse
+    # Ld = kd*(I/r^2)max(0, nxl)
+    direction_to_lights = vertices[np.newaxis, :, :] - light_positions[:, np.newaxis, :] # [nlight, nver, 3]
+    direction_to_lights_n = np.sqrt(np.sum(direction_to_lights**2, axis = 2)) # [nlight, nver]
+    direction_to_lights = direction_to_lights/direction_to_lights_n[:, :, np.newaxis]
+    normals_dot_lights = normals[np.newaxis, :, :]*direction_to_lights # [nlight, nver, 3]
+    normals_dot_lights = np.sum(normals_dot_lights, axis = 2) # [nlight, nver]
+    diffuse_output = colors[np.newaxis, :, :]*normals_dot_lights[:, :, np.newaxis]*light_intensities[:, np.newaxis, :]
+    diffuse_output = np.sum(diffuse_output, axis = 0) # [nver, 3]
+    
+    # specular
+    # h = (v + l)/(|v + l|) bisector
+    # Ls = ks*(I/r^2)max(0, nxh)^p
+    # increasing p narrows the reflectionlob
+
+    lit_colors = diffuse_output # only diffuse part here.
+    lit_colors = np.minimum(np.maximum(lit_colors, 0), 1)
+    return lit_colors
+
+
+
+## TODO. estimate light(sh coeff)
+## -------------------------------- estimate. can not use now. 
+def fit_light(image, vertices, colors, triangles, vis_ind, lamb = 10, max_iter = 3):
+    [h, w, c] = image.shape
+
+    # surface normal
+    norm = get_normal(vertices, triangles)
+    
+    nver = vertices.shape[1]
+
+    # vertices --> corresponding image pixel
+    pt2d = vertices[:2, :]
+
+    pt2d[0,:] = np.minimum(np.maximum(pt2d[0,:], 0), w - 1)
+    pt2d[1,:] = np.minimum(np.maximum(pt2d[1,:], 0), h - 1)
+    pt2d = np.round(pt2d).astype(np.int32) # 2 x nver
+
+    image_pixel = image[pt2d[1,:], pt2d[0,:], :] # nver x 3
+    image_pixel = image_pixel.T # 3 x nver
+
+    # vertices --> corresponding mean texture pixel with illumination
+    # Spherical Harmonic Basis
+    harmonic_dim = 9
+    nx = norm[0,:];
+    ny = norm[1,:];
+    nz = norm[2,:];
+    harmonic = np.zeros((nver, harmonic_dim))
+
+    pi = np.pi
+    harmonic[:,0] = np.sqrt(1/(4*pi)) * np.ones((nver,));
+    harmonic[:,1] = np.sqrt(3/(4*pi)) * nx;
+    harmonic[:,2] = np.sqrt(3/(4*pi)) * ny;
+    harmonic[:,3] = np.sqrt(3/(4*pi)) * nz;
+    harmonic[:,4] = 1/2. * np.sqrt(3/(4*pi)) * (2*nz**2 - nx**2 - ny**2);
+    harmonic[:,5] = 3 * np.sqrt(5/(12*pi)) * (ny*nz);
+    harmonic[:,6] = 3 * np.sqrt(5/(12*pi)) * (nx*nz);
+    harmonic[:,7] = 3 * np.sqrt(5/(12*pi)) * (nx*ny);
+    harmonic[:,8] = 3/2. * np.sqrt(5/(12*pi)) * (nx*nx - ny*ny);
+    
+    '''
+    I' = sum(albedo * lj * hj) j = 0:9 (albedo = tex)
+    set A = albedo*h (n x 9)
+        alpha = lj (9 x 1)
+        Y = I (n x 1)
+        Y' = A.dot(alpha)
+
+    opt function:
+        ||Y - A*alpha|| + lambda*(alpha'*alpha)
+    result:
+        A'*(Y - A*alpha) + lambda*alpha = 0
+        ==>
+        (A'*A*alpha - lambda)*alpha = A'*Y
+        left: 9 x 9
+        right: 9 x 1
+    '''
+    n_vis_ind = len(vis_ind)
+    n = n_vis_ind*c
+
+    Y = np.zeros((n, 1))
+    A = np.zeros((n, 9))
+    light = np.zeros((3, 1))
+
+    for k in range(c):
+        Y[k*n_vis_ind:(k+1)*n_vis_ind, :] = image_pixel[k, vis_ind][:, np.newaxis]
+        A[k*n_vis_ind:(k+1)*n_vis_ind, :] = texture[k, vis_ind][:, np.newaxis] * harmonic[vis_ind, :]
+        Ac = texture[k, vis_ind][:, np.newaxis]
+        Yc = image_pixel[k, vis_ind][:, np.newaxis]
+        light[k] = (Ac.T.dot(Yc))/(Ac.T.dot(Ac))
+
+    for i in range(max_iter):
+
+        Yc = Y.copy()
+        for k in range(c):
+            Yc[k*n_vis_ind:(k+1)*n_vis_ind, :]  /= light[k]
+
+        # update alpha
+        equation_left = np.dot(A.T, A) + lamb*np.eye(harmonic_dim); # why + ?
+        equation_right = np.dot(A.T, Yc) 
+        alpha = np.dot(np.linalg.inv(equation_left), equation_right)
+
+        # update light
+        for k in range(c):
+            Ac = A[k*n_vis_ind:(k+1)*n_vis_ind, :].dot(alpha)
+            Yc = Y[k*n_vis_ind:(k+1)*n_vis_ind, :]
+            light[k] = (Ac.T.dot(Yc))/(Ac.T.dot(Ac))
+
+    appearance = np.zeros_like(texture)
+    for k in range(c):
+        tmp = np.dot(harmonic*texture[k, :][:, np.newaxis], alpha*light[k])
+        appearance[k,:] = tmp.T
+
+    appearance = np.minimum(np.maximum(appearance, 0), 1)
+
+    return appearance
+
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/render.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/render.py
new file mode 100644
index 000000000..eb0d92d31
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/render.py
@@ -0,0 +1,287 @@
+'''
+functions about rendering mesh(from 3d obj to 2d image).
+only use rasterization render here.
+Note that:
+1. Generally, render func includes camera, light, raterize. Here no camera and light(I write these in other files)
+2. Generally, the input vertices are normalized to [-1,1] and cetered on [0, 0]. (in world space)
+   Here, the vertices are using image coords, which centers on [w/2, h/2] with the y-axis pointing to oppisite direction.
+Means: render here only conducts interpolation.(I just want to make the input flexible)
+
+Preparation knowledge:
+z-buffer: https://cs184.eecs.berkeley.edu/lecture/pipeline
+
+Author: Yao Feng 
+Mail: yaofeng1995@gmail.com
+'''
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from time import time
+
+def isPointInTri(point, tri_points):
+    ''' Judge whether the point is in the triangle
+    Method:
+        http://blackpawn.com/texts/pointinpoly/
+    Args:
+        point: (2,). [u, v] or [x, y] 
+        tri_points: (3 vertices, 2 coords). three vertices(2d points) of a triangle. 
+    Returns:
+        bool: true for in triangle
+    '''
+    tp = tri_points
+
+    # vectors
+    v0 = tp[2,:] - tp[0,:]
+    v1 = tp[1,:] - tp[0,:]
+    v2 = point - tp[0,:]
+
+    # dot products
+    dot00 = np.dot(v0.T, v0)
+    dot01 = np.dot(v0.T, v1)
+    dot02 = np.dot(v0.T, v2)
+    dot11 = np.dot(v1.T, v1)
+    dot12 = np.dot(v1.T, v2)
+
+    # barycentric coordinates
+    if dot00*dot11 - dot01*dot01 == 0:
+        inverDeno = 0
+    else:
+        inverDeno = 1/(dot00*dot11 - dot01*dot01)
+
+    u = (dot11*dot02 - dot01*dot12)*inverDeno
+    v = (dot00*dot12 - dot01*dot02)*inverDeno
+
+    # check if point in triangle
+    return (u >= 0) & (v >= 0) & (u + v < 1)
+
+def get_point_weight(point, tri_points):
+    ''' Get the weights of the position
+    Methods: https://gamedev.stackexchange.com/questions/23743/whats-the-most-efficient-way-to-find-barycentric-coordinates
+     -m1.compute the area of the triangles formed by embedding the point P inside the triangle
+     -m2.Christer Ericson's book "Real-Time Collision Detection". faster.(used)
+    Args:
+        point: (2,). [u, v] or [x, y] 
+        tri_points: (3 vertices, 2 coords). three vertices(2d points) of a triangle. 
+    Returns:
+        w0: weight of v0
+        w1: weight of v1
+        w2: weight of v3
+     '''
+    tp = tri_points
+    # vectors
+    v0 = tp[2,:] - tp[0,:]
+    v1 = tp[1,:] - tp[0,:]
+    v2 = point - tp[0,:]
+
+    # dot products
+    dot00 = np.dot(v0.T, v0)
+    dot01 = np.dot(v0.T, v1)
+    dot02 = np.dot(v0.T, v2)
+    dot11 = np.dot(v1.T, v1)
+    dot12 = np.dot(v1.T, v2)
+
+    # barycentric coordinates
+    if dot00*dot11 - dot01*dot01 == 0:
+        inverDeno = 0
+    else:
+        inverDeno = 1/(dot00*dot11 - dot01*dot01)
+
+    u = (dot11*dot02 - dot01*dot12)*inverDeno
+    v = (dot00*dot12 - dot01*dot02)*inverDeno
+
+    w0 = 1 - u - v
+    w1 = v
+    w2 = u
+
+    return w0, w1, w2
+
+def rasterize_triangles(vertices, triangles, h, w):
+    ''' 
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3]
+        h: height
+        w: width
+    Returns:
+        depth_buffer: [h, w] saves the depth, here, the bigger the z, the fronter the point.
+        triangle_buffer: [h, w] saves the tri id(-1 for no triangle). 
+        barycentric_weight: [h, w, 3] saves corresponding barycentric weight.
+
+    # Each triangle has 3 vertices & Each vertex has 3 coordinates x, y, z.
+    # h, w is the size of rendering
+    '''
+    # initial 
+    depth_buffer = np.zeros([h, w]) - 999999. #+ np.min(vertices[2,:]) - 999999. # set the initial z to the farest position
+    triangle_buffer = np.zeros([h, w], dtype = np.int32) - 1  # if tri id = -1, the pixel has no triangle correspondance
+    barycentric_weight = np.zeros([h, w, 3], dtype = np.float32)  # 
+    
+    for i in range(triangles.shape[0]):
+        tri = triangles[i, :] # 3 vertex indices
+
+        # the inner bounding box
+        umin = max(int(np.ceil(np.min(vertices[tri, 0]))), 0)
+        umax = min(int(np.floor(np.max(vertices[tri, 0]))), w-1)
+
+        vmin = max(int(np.ceil(np.min(vertices[tri, 1]))), 0)
+        vmax = min(int(np.floor(np.max(vertices[tri, 1]))), h-1)
+
+        if umax<umin or vmax<vmin:
+            continue
+
+        for u in range(umin, umax+1):
+            for v in range(vmin, vmax+1):
+                if not isPointInTri([u,v], vertices[tri, :2]): 
+                    continue
+                w0, w1, w2 = get_point_weight([u, v], vertices[tri, :2]) # barycentric weight
+                point_depth = w0*vertices[tri[0], 2] + w1*vertices[tri[1], 2] + w2*vertices[tri[2], 2]
+                if point_depth > depth_buffer[v, u]:
+                    depth_buffer[v, u] = point_depth
+                    triangle_buffer[v, u] = i
+                    barycentric_weight[v, u, :] = np.array([w0, w1, w2])
+
+    return depth_buffer, triangle_buffer, barycentric_weight
+
+
+def render_colors_ras(vertices, triangles, colors, h, w, c = 3):
+    ''' render mesh with colors(rasterize triangle first)
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3] 
+        colors: [nver, 3]
+        h: height
+        w: width    
+        c: channel
+    Returns:
+        image: [h, w, c]. rendering.
+    '''
+    assert vertices.shape[0] == colors.shape[0]
+
+    depth_buffer, triangle_buffer, barycentric_weight = rasterize_triangles(vertices, triangles, h, w)
+
+    triangle_buffer_flat = np.reshape(triangle_buffer, [-1]) # [h*w]
+    barycentric_weight_flat = np.reshape(barycentric_weight, [-1, c]) #[h*w, c]
+    weight = barycentric_weight_flat[:, :, np.newaxis] # [h*w, 3(ver in tri), 1]
+
+    colors_flat = colors[triangles[triangle_buffer_flat, :], :] # [h*w(tri id in pixel), 3(ver in tri), c(color in ver)]
+    colors_flat = weight*colors_flat # [h*w, 3, 3]
+    colors_flat = np.sum(colors_flat, 1) #[h*w, 3]. add tri.
+
+    image = np.reshape(colors_flat, [h, w, c])
+    # mask = (triangle_buffer[:,:] > -1).astype(np.float32)
+    # image = image*mask[:,:,np.newaxis]
+    return image
+
+
+def render_colors(vertices, triangles, colors, h, w, c = 3):
+    ''' render mesh with colors
+    Args:
+        vertices: [nver, 3]
+        triangles: [ntri, 3] 
+        colors: [nver, 3]
+        h: height
+        w: width    
+    Returns:
+        image: [h, w, c]. 
+    '''
+    assert vertices.shape[0] == colors.shape[0]
+    
+    # initial 
+    image = np.zeros((h, w, c))
+    depth_buffer = np.zeros([h, w]) - 999999.
+
+    for i in range(triangles.shape[0]):
+        tri = triangles[i, :] # 3 vertex indices
+
+        # the inner bounding box
+        umin = max(int(np.ceil(np.min(vertices[tri, 0]))), 0)
+        umax = min(int(np.floor(np.max(vertices[tri, 0]))), w-1)
+
+        vmin = max(int(np.ceil(np.min(vertices[tri, 1]))), 0)
+        vmax = min(int(np.floor(np.max(vertices[tri, 1]))), h-1)
+
+        if umax<umin or vmax<vmin:
+            continue
+
+        for u in range(umin, umax+1):
+            for v in range(vmin, vmax+1):
+                if not isPointInTri([u,v], vertices[tri, :2]): 
+                    continue
+                w0, w1, w2 = get_point_weight([u, v], vertices[tri, :2])
+                point_depth = w0*vertices[tri[0], 2] + w1*vertices[tri[1], 2] + w2*vertices[tri[2], 2]
+
+                if point_depth > depth_buffer[v, u]:
+                    depth_buffer[v, u] = point_depth
+                    image[v, u, :] = w0*colors[tri[0], :] + w1*colors[tri[1], :] + w2*colors[tri[2], :]
+    return image
+
+
+def render_texture(vertices, triangles, texture, tex_coords, tex_triangles, h, w, c = 3, mapping_type = 'nearest'):
+    ''' render mesh with texture map
+    Args:
+        vertices: [nver], 3
+        triangles: [ntri, 3]
+        texture: [tex_h, tex_w, 3]
+        tex_coords: [ntexcoords, 3]
+        tex_triangles: [ntri, 3]
+        h: height of rendering
+        w: width of rendering
+        c: channel
+        mapping_type: 'bilinear' or 'nearest'
+    '''
+    assert triangles.shape[0] == tex_triangles.shape[0]
+    tex_h, tex_w, _ = texture.shape
+
+    # initial 
+    image = np.zeros((h, w, c))
+    depth_buffer = np.zeros([h, w]) - 999999.
+
+    for i in range(triangles.shape[0]):
+        tri = triangles[i, :] # 3 vertex indices
+        tex_tri = tex_triangles[i, :] # 3 tex indice
+
+        # the inner bounding box
+        umin = max(int(np.ceil(np.min(vertices[tri, 0]))), 0)
+        umax = min(int(np.floor(np.max(vertices[tri, 0]))), w-1)
+
+        vmin = max(int(np.ceil(np.min(vertices[tri, 1]))), 0)
+        vmax = min(int(np.floor(np.max(vertices[tri, 1]))), h-1)
+
+        if umax<umin or vmax<vmin:
+            continue
+
+        for u in range(umin, umax+1):
+            for v in range(vmin, vmax+1):
+                if not isPointInTri([u,v], vertices[tri, :2]): 
+                    continue
+                w0, w1, w2 = get_point_weight([u, v], vertices[tri, :2])
+                point_depth = w0*vertices[tri[0], 2] + w1*vertices[tri[1], 2] + w2*vertices[tri[2], 2]
+                if point_depth > depth_buffer[v, u]:
+                    # update depth
+                    depth_buffer[v, u] = point_depth    
+                    
+                    # tex coord
+                    tex_xy = w0*tex_coords[tex_tri[0], :] + w1*tex_coords[tex_tri[1], :] + w2*tex_coords[tex_tri[2], :]
+                    tex_xy[0] = max(min(tex_xy[0], float(tex_w - 1)), 0.0); 
+                    tex_xy[1] = max(min(tex_xy[1], float(tex_h - 1)), 0.0); 
+
+                    # nearest
+                    if mapping_type == 'nearest':
+                        tex_xy = np.round(tex_xy).astype(np.int32)
+                        tex_value = texture[tex_xy[1], tex_xy[0], :] 
+
+                    # bilinear
+                    elif mapping_type == 'bilinear':
+                        # next 4 pixels
+                        ul = texture[int(np.floor(tex_xy[1])), int(np.floor(tex_xy[0])), :]
+                        ur = texture[int(np.floor(tex_xy[1])), int(np.ceil(tex_xy[0])), :]
+                        dl = texture[int(np.ceil(tex_xy[1])), int(np.floor(tex_xy[0])), :]
+                        dr = texture[int(np.ceil(tex_xy[1])), int(np.ceil(tex_xy[0])), :]
+
+                        yd = tex_xy[1] - np.floor(tex_xy[1])
+                        xd = tex_xy[0] - np.floor(tex_xy[0])
+                        tex_value = ul*(1-xd)*(1-yd) + ur*xd*(1-yd) + dl*(1-xd)*yd + dr*xd*yd
+
+                    image[v, u, :] = tex_value
+    return image
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/transform.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/transform.py
new file mode 100644
index 000000000..ab56e370c
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/transform.py
@@ -0,0 +1,385 @@
+'''
+Functions about transforming mesh(changing the position: modify vertices).
+1. forward: transform(transform, camera, project).
+2. backward: estimate transform matrix from correspondences.
+
+Preparation knowledge:
+transform&camera model:
+https://cs184.eecs.berkeley.edu/lecture/transforms-2
+Part I: camera geometry and single view geometry in MVGCV
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import math
+from math import cos, sin
+
+def angle2matrix(angles):
+    ''' get rotation matrix from three rotation angles(degree). right-handed.
+    Args:
+        angles: [3,]. x, y, z angles
+        x: pitch. positive for looking down.
+        y: yaw. positive for looking left. 
+        z: roll. positive for tilting head right. 
+    Returns:
+        R: [3, 3]. rotation matrix.
+    '''
+    x, y, z = np.deg2rad(angles[0]), np.deg2rad(angles[1]), np.deg2rad(angles[2])
+    # x
+    Rx=np.array([[1,      0,       0],
+                 [0, cos(x),  -sin(x)],
+                 [0, sin(x),   cos(x)]])
+    # y
+    Ry=np.array([[ cos(y), 0, sin(y)],
+                 [      0, 1,      0],
+                 [-sin(y), 0, cos(y)]])
+    # z
+    Rz=np.array([[cos(z), -sin(z), 0],
+                 [sin(z),  cos(z), 0],
+                 [     0,       0, 1]])
+    
+    R=Rz.dot(Ry.dot(Rx))
+    return R.astype(np.float32)
+
+def angle2matrix_3ddfa(angles):
+    ''' get rotation matrix from three rotation angles(radian). The same as in 3DDFA.
+    Args:
+        angles: [3,]. x, y, z angles
+        x: pitch.
+        y: yaw. 
+        z: roll. 
+    Returns:
+        R: 3x3. rotation matrix.
+    '''
+    # x, y, z = np.deg2rad(angles[0]), np.deg2rad(angles[1]), np.deg2rad(angles[2])
+    x, y, z = angles[0], angles[1], angles[2]
+    
+    # x
+    Rx=np.array([[1,      0,       0],
+                 [0, cos(x),  sin(x)],
+                 [0, -sin(x),   cos(x)]])
+    # y
+    Ry=np.array([[ cos(y), 0, -sin(y)],
+                 [      0, 1,      0],
+                 [sin(y), 0, cos(y)]])
+    # z
+    Rz=np.array([[cos(z), sin(z), 0],
+                 [-sin(z),  cos(z), 0],
+                 [     0,       0, 1]])
+    R = Rx.dot(Ry).dot(Rz)
+    return R.astype(np.float32)
+
+
+## ------------------------------------------ 1. transform(transform, project, camera).
+## ---------- 3d-3d transform. Transform obj in world space
+def rotate(vertices, angles):
+    ''' rotate vertices. 
+    X_new = R.dot(X). X: 3 x 1   
+    Args:
+        vertices: [nver, 3]. 
+        rx, ry, rz: degree angles
+        rx: pitch. positive for looking down 
+        ry: yaw. positive for looking left
+        rz: roll. positive for tilting head right
+    Returns:
+        rotated vertices: [nver, 3]
+    '''
+    R = angle2matrix(angles)
+    rotated_vertices = vertices.dot(R.T)
+
+    return rotated_vertices
+
+def similarity_transform(vertices, s, R, t3d):
+    ''' similarity transform. dof = 7.
+    3D: s*R.dot(X) + t
+    Homo: M = [[sR, t],[0^T, 1]].  M.dot(X)
+    Args:(float32)
+        vertices: [nver, 3]. 
+        s: [1,]. scale factor.
+        R: [3,3]. rotation matrix.
+        t3d: [3,]. 3d translation vector.
+    Returns:
+        transformed vertices: [nver, 3]
+    '''
+    t3d = np.squeeze(np.array(t3d, dtype = np.float32))
+    transformed_vertices = s * vertices.dot(R.T) + t3d[np.newaxis, :]
+
+    return transformed_vertices
+
+
+## -------------- Camera. from world space to camera space
+# Ref: https://cs184.eecs.berkeley.edu/lecture/transforms-2
+def normalize(x):
+    epsilon = 1e-12
+    norm = np.sqrt(np.sum(x**2, axis = 0))
+    norm = np.maximum(norm, epsilon)
+    return x/norm
+
+def lookat_camera(vertices, eye, at = None, up = None):
+    """ 'look at' transformation: from world space to camera space
+    standard camera space: 
+        camera located at the origin. 
+        looking down negative z-axis. 
+        vertical vector is y-axis.
+    Xcam = R(X - C)
+    Homo: [[R, -RC], [0, 1]]
+    Args:
+      vertices: [nver, 3] 
+      eye: [3,] the XYZ world space position of the camera.
+      at: [3,] a position along the center of the camera's gaze.
+      up: [3,] up direction 
+    Returns:
+      transformed_vertices: [nver, 3]
+    """
+    if at is None:
+      at = np.array([0, 0, 0], np.float32)
+    if up is None:
+      up = np.array([0, 1, 0], np.float32)
+
+    eye = np.array(eye).astype(np.float32)
+    at = np.array(at).astype(np.float32)
+    z_aixs = -normalize(at - eye) # look forward
+    x_aixs = normalize(np.cross(up, z_aixs)) # look right
+    y_axis = np.cross(z_aixs, x_aixs) # look up
+
+    R = np.stack((x_aixs, y_axis, z_aixs))#, axis = 0) # 3 x 3
+    transformed_vertices = vertices - eye # translation
+    transformed_vertices = transformed_vertices.dot(R.T) # rotation
+    return transformed_vertices
+
+## --------- 3d-2d project. from camera space to image plane
+# generally, image plane only keeps x,y channels, here reserve z channel for calculating z-buffer.
+def orthographic_project(vertices):
+    ''' scaled orthographic projection(just delete z)
+        assumes: variations in depth over the object is small relative to the mean distance from camera to object
+        x -> x*f/z, y -> x*f/z, z -> f.
+        for point i,j. zi~=zj. so just delete z
+        ** often used in face
+        Homo: P = [[1,0,0,0], [0,1,0,0], [0,0,1,0]]
+    Args:
+        vertices: [nver, 3]
+    Returns:
+        projected_vertices: [nver, 3] if isKeepZ=True. [nver, 2] if isKeepZ=False.
+    '''
+    return vertices.copy()
+
+def perspective_project(vertices, fovy, aspect_ratio = 1., near = 0.1, far = 1000.):
+    ''' perspective projection.
+    Args:
+        vertices: [nver, 3]
+        fovy: vertical angular field of view. degree.
+        aspect_ratio : width / height of field of view
+        near : depth of near clipping plane
+        far : depth of far clipping plane
+    Returns:
+        projected_vertices: [nver, 3] 
+    '''
+    fovy = np.deg2rad(fovy)
+    top = near*np.tan(fovy)
+    bottom = -top 
+    right = top*aspect_ratio
+    left = -right
+
+    #-- homo
+    P = np.array([[near/right, 0, 0, 0],
+                 [0, near/top, 0, 0],
+                 [0, 0, -(far+near)/(far-near), -2*far*near/(far-near)],
+                 [0, 0, -1, 0]])
+    vertices_homo = np.hstack((vertices, np.ones((vertices.shape[0], 1)))) # [nver, 4]
+    projected_vertices = vertices_homo.dot(P.T)
+    projected_vertices = projected_vertices/projected_vertices[:,3:]
+    projected_vertices = projected_vertices[:,:3]
+    projected_vertices[:,2] = -projected_vertices[:,2]
+
+    #-- non homo. only fovy
+    # projected_vertices = vertices.copy()
+    # projected_vertices[:,0] = -(near/right)*vertices[:,0]/vertices[:,2]
+    # projected_vertices[:,1] = -(near/top)*vertices[:,1]/vertices[:,2]
+    return projected_vertices
+
+
+def to_image(vertices, h, w, is_perspective = False):
+    ''' change vertices to image coord system
+    3d system: XYZ, center(0, 0, 0)
+    2d image: x(u), y(v). center(w/2, h/2), flip y-axis. 
+    Args:
+        vertices: [nver, 3]
+        h: height of the rendering
+        w : width of the rendering
+    Returns:
+        projected_vertices: [nver, 3]  
+    '''
+    image_vertices = vertices.copy()
+    if is_perspective:
+        # if perspective, the projected vertices are normalized to [-1, 1]. so change it to image size first.
+        image_vertices[:,0] = image_vertices[:,0]*w/2
+        image_vertices[:,1] = image_vertices[:,1]*h/2
+    # move to center of image
+    image_vertices[:,0] = image_vertices[:,0] + w/2
+    image_vertices[:,1] = image_vertices[:,1] + h/2
+    # flip vertices along y-axis.
+    image_vertices[:,1] = h - image_vertices[:,1] - 1
+    return image_vertices
+
+
+#### -------------------------------------------2. estimate transform matrix from correspondences.
+def estimate_affine_matrix_3d23d(X, Y):
+    ''' Using least-squares solution 
+    Args:
+        X: [n, 3]. 3d points(fixed)
+        Y: [n, 3]. corresponding 3d points(moving). Y = PX
+    Returns:
+        P_Affine: (3, 4). Affine camera matrix (the third row is [0, 0, 0, 1]).
+    '''
+    X_homo = np.hstack((X, np.ones([X.shape[1],1]))) #n x 4
+    P = np.linalg.lstsq(X_homo, Y)[0].T # Affine matrix. 3 x 4
+    return P
+    
+def estimate_affine_matrix_3d22d(X, x):
+    ''' Using Golden Standard Algorithm for estimating an affine camera
+        matrix P from world to image correspondences.
+        See Alg.7.2. in MVGCV 
+        Code Ref: https://github.com/patrikhuber/eos/blob/master/include/eos/fitting/affine_camera_estimation.hpp
+        x_homo = X_homo.dot(P_Affine)
+    Args:
+        X: [n, 3]. corresponding 3d points(fixed)
+        x: [n, 2]. n>=4. 2d points(moving). x = PX
+    Returns:
+        P_Affine: [3, 4]. Affine camera matrix
+    '''
+    X = X.T; x = x.T
+    assert(x.shape[1] == X.shape[1])
+    n = x.shape[1]
+    assert(n >= 4)
+
+    #--- 1. normalization
+    # 2d points
+    mean = np.mean(x, 1) # (2,)
+    x = x - np.tile(mean[:, np.newaxis], [1, n])
+    average_norm = np.mean(np.sqrt(np.sum(x**2, 0)))
+    scale = np.sqrt(2) / average_norm
+    x = scale * x
+
+    T = np.zeros((3,3), dtype = np.float32)
+    T[0, 0] = T[1, 1] = scale
+    T[:2, 2] = -mean*scale
+    T[2, 2] = 1
+
+    # 3d points
+    X_homo = np.vstack((X, np.ones((1, n))))
+    mean = np.mean(X, 1) # (3,)
+    X = X - np.tile(mean[:, np.newaxis], [1, n])
+    m = X_homo[:3,:] - X
+    average_norm = np.mean(np.sqrt(np.sum(X**2, 0)))
+    scale = np.sqrt(3) / average_norm
+    X = scale * X
+
+    U = np.zeros((4,4), dtype = np.float32)
+    U[0, 0] = U[1, 1] = U[2, 2] = scale
+    U[:3, 3] = -mean*scale
+    U[3, 3] = 1
+
+    # --- 2. equations
+    A = np.zeros((n*2, 8), dtype = np.float32);
+    X_homo = np.vstack((X, np.ones((1, n)))).T
+    A[:n, :4] = X_homo
+    A[n:, 4:] = X_homo
+    b = np.reshape(x, [-1, 1])
+ 
+    # --- 3. solution
+    p_8 = np.linalg.pinv(A).dot(b)
+    P = np.zeros((3, 4), dtype = np.float32)
+    P[0, :] = p_8[:4, 0]
+    P[1, :] = p_8[4:, 0]
+    P[-1, -1] = 1
+
+    # --- 4. denormalization
+    P_Affine = np.linalg.inv(T).dot(P.dot(U))
+    return P_Affine
+
+def P2sRt(P):
+    ''' decompositing camera matrix P
+    Args: 
+        P: (3, 4). Affine Camera Matrix.
+    Returns:
+        s: scale factor.
+        R: (3, 3). rotation matrix.
+        t: (3,). translation. 
+    '''
+    t = P[:, 3]
+    R1 = P[0:1, :3]
+    R2 = P[1:2, :3]
+    s = (np.linalg.norm(R1) + np.linalg.norm(R2))/2.0
+    r1 = R1/np.linalg.norm(R1)
+    r2 = R2/np.linalg.norm(R2)
+    r3 = np.cross(r1, r2)
+
+    R = np.concatenate((r1, r2, r3), 0)
+    return s, R, t
+
+#Ref: https://www.learnopencv.com/rotation-matrix-to-euler-angles/
+def isRotationMatrix(R):
+    ''' checks if a matrix is a valid rotation matrix(whether orthogonal or not)
+    '''
+    Rt = np.transpose(R)
+    shouldBeIdentity = np.dot(Rt, R)
+    I = np.identity(3, dtype = R.dtype)
+    n = np.linalg.norm(I - shouldBeIdentity)
+    return n < 1e-6
+
+def matrix2angle(R):
+    ''' get three Euler angles from Rotation Matrix
+    Args:
+        R: (3,3). rotation matrix
+    Returns:
+        x: pitch
+        y: yaw
+        z: roll
+    '''
+    assert(isRotationMatrix)
+    sy = math.sqrt(R[0,0] * R[0,0] +  R[1,0] * R[1,0])
+     
+    singular = sy < 1e-6
+ 
+    if  not singular :
+        x = math.atan2(R[2,1] , R[2,2])
+        y = math.atan2(-R[2,0], sy)
+        z = math.atan2(R[1,0], R[0,0])
+    else :
+        x = math.atan2(-R[1,2], R[1,1])
+        y = math.atan2(-R[2,0], sy)
+        z = 0
+
+    # rx, ry, rz = np.rad2deg(x), np.rad2deg(y), np.rad2deg(z)
+    rx, ry, rz = x*180/np.pi, y*180/np.pi, z*180/np.pi
+    return rx, ry, rz
+
+# def matrix2angle(R):
+#     ''' compute three Euler angles from a Rotation Matrix. Ref: http://www.gregslabaugh.net/publications/euler.pdf
+#     Args:
+#         R: (3,3). rotation matrix
+#     Returns:
+#         x: yaw
+#         y: pitch
+#         z: roll
+#     '''
+#     # assert(isRotationMatrix(R))
+
+#     if R[2,0] !=1 or R[2,0] != -1:
+#         x = math.asin(R[2,0])
+#         y = math.atan2(R[2,1]/cos(x), R[2,2]/cos(x))
+#         z = math.atan2(R[1,0]/cos(x), R[0,0]/cos(x))
+        
+#     else:# Gimbal lock
+#         z = 0 #can be anything
+#         if R[2,0] == -1:
+#             x = np.pi/2
+#             y = z + math.atan2(R[0,1], R[0,2])
+#         else:
+#             x = -np.pi/2
+#             y = -z + math.atan2(-R[0,1], -R[0,2])
+
+#     return x, y, z
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/mesh_numpy/vis.py b/python-package/insightface/thirdparty/face3d/mesh_numpy/vis.py
new file mode 100644
index 000000000..9db972f2c
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/mesh_numpy/vis.py
@@ -0,0 +1,24 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import matplotlib.pyplot as plt
+from skimage import measure
+from mpl_toolkits.mplot3d import Axes3D
+
+def plot_mesh(vertices, triangles, subplot = [1,1,1], title = 'mesh', el = 90, az = -90, lwdt=.1, dist = 6, color = "grey"):
+	'''
+	plot the mesh 
+	Args:
+		vertices: [nver, 3]
+		triangles: [ntri, 3]
+	'''
+	ax = plt.subplot(subplot[0], subplot[1], subplot[2], projection = '3d')
+	ax.plot_trisurf(vertices[:, 0], vertices[:, 1], vertices[:, 2], triangles = triangles, lw = lwdt, color = color, alpha = 1)
+	ax.axis("off")
+	ax.view_init(elev = el, azim = az)
+	ax.dist = dist
+	plt.title(title)
+
+### -------------- Todo: use vtk to visualize mesh? or visvis? or VisPy?
diff --git a/python-package/insightface/thirdparty/face3d/morphable_model/__init__.py b/python-package/insightface/thirdparty/face3d/morphable_model/__init__.py
new file mode 100644
index 000000000..19fc6a602
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/morphable_model/__init__.py
@@ -0,0 +1,7 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from .. import mesh
+from .morphabel_model import MorphabelModel
+from . import load
\ No newline at end of file
diff --git a/python-package/insightface/thirdparty/face3d/morphable_model/fit.py b/python-package/insightface/thirdparty/face3d/morphable_model/fit.py
new file mode 100644
index 000000000..480532c57
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/morphable_model/fit.py
@@ -0,0 +1,272 @@
+'''
+Estimating parameters about vertices: shape para, exp para, pose para(s, R, t)
+'''
+import numpy as np
+from .. import mesh
+
+''' TODO: a clear document. 
+Given: image_points, 3D Model, Camera Matrix(s, R, t2d)
+Estimate: shape parameters, expression parameters
+
+Inference: 
+
+    projected_vertices = s*P*R(mu + shape + exp) + t2d  --> image_points
+    s*P*R*shape + s*P*R(mu + exp) + t2d --> image_poitns
+
+    # Define:
+    X = vertices
+    x_hat = projected_vertices
+    x = image_points
+    A = s*P*R
+    b = s*P*R(mu + exp) + t2d
+    ==>
+    x_hat = A*shape + b  (2 x n)
+
+    A*shape (2 x n)
+    shape = reshape(shapePC * sp) (3 x n)
+    shapePC*sp : (3n x 1)
+
+    * flatten:
+    x_hat_flatten = A*shape + b_flatten  (2n x 1)
+    A*shape (2n x 1)
+    --> A*shapePC (2n x 199)  sp: 199 x 1
+    
+    # Define:
+    pc_2d = A* reshape(shapePC)
+    pc_2d_flatten = flatten(pc_2d) (2n x 199)
+
+    =====>
+    x_hat_flatten = pc_2d_flatten * sp + b_flatten ---> x_flatten (2n x 1)
+
+    Goals:
+    (ignore flatten, pc_2d-->pc)
+    min E = || x_hat - x || + lambda*sum(sp/sigma)^2
+          = || pc * sp + b - x || + lambda*sum(sp/sigma)^2
+
+    Solve:
+    d(E)/d(sp) = 0
+    2 * pc' * (pc * sp + b - x) + 2 * lambda * sp / (sigma' * sigma) = 0
+
+    Get:
+    (pc' * pc + lambda / (sigma'* sigma)) * sp  = pc' * (x - b)
+
+'''
+
+def estimate_shape(x, shapeMU, shapePC, shapeEV, expression, s, R, t2d, lamb = 3000):
+    '''
+    Args:
+        x: (2, n). image points (to be fitted)
+        shapeMU: (3n, 1)
+        shapePC: (3n, n_sp)
+        shapeEV: (n_sp, 1)
+        expression: (3, n)
+        s: scale
+        R: (3, 3). rotation matrix
+        t2d: (2,). 2d translation
+        lambda: regulation coefficient
+
+    Returns:
+        shape_para: (n_sp, 1) shape parameters(coefficients)
+    '''
+    x = x.copy()
+    assert(shapeMU.shape[0] == shapePC.shape[0])
+    assert(shapeMU.shape[0] == x.shape[1]*3)
+
+    dof = shapePC.shape[1]
+
+    n = x.shape[1]
+    sigma = shapeEV
+    t2d = np.array(t2d)
+    P = np.array([[1, 0, 0], [0, 1, 0]], dtype = np.float32)
+    A = s*P.dot(R)
+
+    # --- calc pc
+    pc_3d = np.resize(shapePC.T, [dof, n, 3]) # 199 x n x 3
+    pc_3d = np.reshape(pc_3d, [dof*n, 3]) 
+    pc_2d = pc_3d.dot(A.T.copy()) # 199 x n x 2
+    
+    pc = np.reshape(pc_2d, [dof, -1]).T # 2n x 199
+
+    # --- calc b
+    # shapeMU
+    mu_3d = np.resize(shapeMU, [n, 3]).T # 3 x n
+    # expression
+    exp_3d = expression
+    # 
+    b = A.dot(mu_3d + exp_3d) + np.tile(t2d[:, np.newaxis], [1, n]) # 2 x n
+    b = np.reshape(b.T, [-1, 1]) # 2n x 1
+
+    # --- solve
+    equation_left = np.dot(pc.T, pc) + lamb * np.diagflat(1/sigma**2)
+    x = np.reshape(x.T, [-1, 1])
+    equation_right = np.dot(pc.T, x - b)
+
+    shape_para = np.dot(np.linalg.inv(equation_left), equation_right)
+
+    return shape_para
+
+def estimate_expression(x, shapeMU, expPC, expEV, shape, s, R, t2d, lamb = 2000):
+    '''
+    Args:
+        x: (2, n). image points (to be fitted)
+        shapeMU: (3n, 1)
+        expPC: (3n, n_ep)
+        expEV: (n_ep, 1)
+        shape: (3, n)
+        s: scale
+        R: (3, 3). rotation matrix
+        t2d: (2,). 2d translation
+        lambda: regulation coefficient
+
+    Returns:
+        exp_para: (n_ep, 1) shape parameters(coefficients)
+    '''
+    x = x.copy()
+    assert(shapeMU.shape[0] == expPC.shape[0])
+    assert(shapeMU.shape[0] == x.shape[1]*3)
+
+    dof = expPC.shape[1]
+
+    n = x.shape[1]
+    sigma = expEV
+    t2d = np.array(t2d)
+    P = np.array([[1, 0, 0], [0, 1, 0]], dtype = np.float32)
+    A = s*P.dot(R)
+
+    # --- calc pc
+    pc_3d = np.resize(expPC.T, [dof, n, 3]) 
+    pc_3d = np.reshape(pc_3d, [dof*n, 3]) 
+    pc_2d = pc_3d.dot(A.T) 
+    pc = np.reshape(pc_2d, [dof, -1]).T # 2n x 29
+
+    # --- calc b
+    # shapeMU
+    mu_3d = np.resize(shapeMU, [n, 3]).T # 3 x n
+    # expression
+    shape_3d = shape
+    # 
+    b = A.dot(mu_3d + shape_3d) + np.tile(t2d[:, np.newaxis], [1, n]) # 2 x n
+    b = np.reshape(b.T, [-1, 1]) # 2n x 1
+
+    # --- solve
+    equation_left = np.dot(pc.T, pc) + lamb * np.diagflat(1/sigma**2)
+    x = np.reshape(x.T, [-1, 1])
+    equation_right = np.dot(pc.T, x - b)
+
+    exp_para = np.dot(np.linalg.inv(equation_left), equation_right)
+    
+    return exp_para
+
+
+# ---------------- fit 
+def fit_points(x, X_ind, model, n_sp, n_ep, max_iter = 4):
+    '''
+    Args:
+        x: (n, 2) image points
+        X_ind: (n,) corresponding Model vertex indices
+        model: 3DMM
+        max_iter: iteration
+    Returns:
+        sp: (n_sp, 1). shape parameters
+        ep: (n_ep, 1). exp parameters
+        s, R, t
+    '''
+    x = x.copy().T
+
+    #-- init
+    sp = np.zeros((n_sp, 1), dtype = np.float32)
+    ep = np.zeros((n_ep, 1), dtype = np.float32)
+
+    #-------------------- estimate
+    X_ind_all = np.tile(X_ind[np.newaxis, :], [3, 1])*3
+    X_ind_all[1, :] += 1
+    X_ind_all[2, :] += 2
+    valid_ind = X_ind_all.flatten('F')
+
+    shapeMU = model['shapeMU'][valid_ind, :]
+    shapePC = model['shapePC'][valid_ind, :n_sp]
+    expPC = model['expPC'][valid_ind, :n_ep]
+
+    for i in range(max_iter):
+        X = shapeMU + shapePC.dot(sp) + expPC.dot(ep)
+        X = np.reshape(X, [int(len(X)/3), 3]).T
+        
+        #----- estimate pose
+        P = mesh.transform.estimate_affine_matrix_3d22d(X.T, x.T)
+        s, R, t = mesh.transform.P2sRt(P)
+        rx, ry, rz = mesh.transform.matrix2angle(R)
+        #print('Iter:{}; estimated pose: s {}, rx {}, ry {}, rz {}, t1 {}, t2 {}'.format(i, s, rx, ry, rz, t[0], t[1]))
+
+        #----- estimate shape
+        # expression
+        shape = shapePC.dot(sp)
+        shape = np.reshape(shape, [int(len(shape)/3), 3]).T
+        ep = estimate_expression(x, shapeMU, expPC, model['expEV'][:n_ep,:], shape, s, R, t[:2], lamb = 20)
+
+        # shape
+        expression = expPC.dot(ep)
+        expression = np.reshape(expression, [int(len(expression)/3), 3]).T
+        if i == 0 :
+            sp = estimate_shape(x, shapeMU, shapePC, model['shapeEV'][:n_sp,:], expression, s, R, t[:2], lamb = 40)
+
+    return sp, ep, s, R, t
+
+
+# ---------------- fitting process
+def fit_points_for_show(x, X_ind, model, n_sp, n_ep, max_iter = 4):
+    '''
+    Args:
+        x: (n, 2) image points
+        X_ind: (n,) corresponding Model vertex indices
+        model: 3DMM
+        max_iter: iteration
+    Returns:
+        sp: (n_sp, 1). shape parameters
+        ep: (n_ep, 1). exp parameters
+        s, R, t
+    '''
+    x = x.copy().T
+
+    #-- init
+    sp = np.zeros((n_sp, 1), dtype = np.float32)
+    ep = np.zeros((n_ep, 1), dtype = np.float32)
+
+    #-------------------- estimate
+    X_ind_all = np.tile(X_ind[np.newaxis, :], [3, 1])*3
+    X_ind_all[1, :] += 1
+    X_ind_all[2, :] += 2
+    valid_ind = X_ind_all.flatten('F')
+
+    shapeMU = model['shapeMU'][valid_ind, :]
+    shapePC = model['shapePC'][valid_ind, :n_sp]
+    expPC = model['expPC'][valid_ind, :n_ep]
+
+    s = 4e-04
+    R = mesh.transform.angle2matrix([0, 0, 0])
+    t = [0, 0, 0]
+    lsp = []; lep = []; ls = []; lR = []; lt = []
+    for i in range(max_iter):
+        X = shapeMU + shapePC.dot(sp) + expPC.dot(ep)
+        X = np.reshape(X, [int(len(X)/3), 3]).T
+        lsp.append(sp); lep.append(ep); ls.append(s), lR.append(R), lt.append(t)
+        
+        #----- estimate pose
+        P = mesh.transform.estimate_affine_matrix_3d22d(X.T, x.T)
+        s, R, t = mesh.transform.P2sRt(P)
+        lsp.append(sp); lep.append(ep); ls.append(s), lR.append(R), lt.append(t)
+
+        #----- estimate shape
+        # expression
+        shape = shapePC.dot(sp)
+        shape = np.reshape(shape, [int(len(shape)/3), 3]).T
+        ep = estimate_expression(x, shapeMU, expPC, model['expEV'][:n_ep,:], shape, s, R, t[:2], lamb = 20)
+        lsp.append(sp); lep.append(ep); ls.append(s), lR.append(R), lt.append(t)
+
+        # shape
+        expression = expPC.dot(ep)
+        expression = np.reshape(expression, [int(len(expression)/3), 3]).T
+        sp = estimate_shape(x, shapeMU, shapePC, model['shapeEV'][:n_sp,:], expression, s, R, t[:2], lamb = 40)
+
+    # print('ls', ls)
+    # print('lR', lR)
+    return np.array(lsp), np.array(lep), np.array(ls), np.array(lR), np.array(lt)
diff --git a/python-package/insightface/thirdparty/face3d/morphable_model/load.py b/python-package/insightface/thirdparty/face3d/morphable_model/load.py
new file mode 100644
index 000000000..0b80665bd
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/morphable_model/load.py
@@ -0,0 +1,110 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import scipy.io as sio
+
+### ---------------------------------  load BFM data
+def load_BFM(model_path):
+    ''' load BFM 3DMM model
+    Args:
+        model_path: path to BFM model. 
+    Returns:
+        model: (nver = 53215, ntri = 105840). nver: number of vertices. ntri: number of triangles.
+            'shapeMU': [3*nver, 1]
+            'shapePC': [3*nver, 199]
+            'shapeEV': [199, 1]
+            'expMU': [3*nver, 1]
+            'expPC': [3*nver, 29]
+            'expEV': [29, 1]
+            'texMU': [3*nver, 1]
+            'texPC': [3*nver, 199]
+            'texEV': [199, 1]
+            'tri': [ntri, 3] (start from 1, should sub 1 in python and c++)
+            'tri_mouth': [114, 3] (start from 1, as a supplement to mouth triangles)
+            'kpt_ind': [68,] (start from 1)
+    PS:
+        You can change codes according to your own saved data.
+        Just make sure the model has corresponding attributes.
+    '''
+    C = sio.loadmat(model_path)
+    model = C['model']
+    model = model[0,0]
+
+    # change dtype from double(np.float64) to np.float32, 
+    # since big matrix process(espetially matrix dot) is too slow in python.
+    model['shapeMU'] = (model['shapeMU'] + model['expMU']).astype(np.float32)
+    model['shapePC'] = model['shapePC'].astype(np.float32)
+    model['shapeEV'] = model['shapeEV'].astype(np.float32)
+    model['expEV'] = model['expEV'].astype(np.float32)
+    model['expPC'] = model['expPC'].astype(np.float32)
+
+    # matlab start with 1. change to 0 in python.
+    model['tri'] = model['tri'].T.copy(order = 'C').astype(np.int32) - 1
+    model['tri_mouth'] = model['tri_mouth'].T.copy(order = 'C').astype(np.int32) - 1
+    
+    # kpt ind
+    model['kpt_ind'] = (np.squeeze(model['kpt_ind']) - 1).astype(np.int32)
+
+    return model
+
+def load_BFM_info(path = 'BFM_info.mat'):
+    ''' load 3DMM model extra information
+    Args:
+        path: path to BFM info. 
+    Returns:  
+        model_info:
+            'symlist': 2 x 26720
+            'symlist_tri': 2 x 52937
+            'segbin': 4 x n (0: nose, 1: eye, 2: mouth, 3: cheek)
+            'segbin_tri': 4 x ntri 
+            'face_contour': 1 x 28
+            'face_contour_line': 1 x 512
+            'face_contour_front': 1 x 28
+            'face_contour_front_line': 1 x 512
+            'nose_hole': 1 x 142
+            'nose_hole_right': 1 x 71
+            'nose_hole_left': 1 x 71
+            'parallel': 17 x 1 cell
+            'parallel_face_contour': 28 x 1 cell
+            'uv_coords': n x 2
+    '''
+    C = sio.loadmat(path)
+    model_info = C['model_info']
+    model_info = model_info[0,0]
+    return model_info
+
+def load_uv_coords(path = 'BFM_UV.mat'):
+    ''' load uv coords of BFM
+    Args:
+        path: path to data.
+    Returns:  
+        uv_coords: [nver, 2]. range: 0-1
+    '''
+    C = sio.loadmat(path)
+    uv_coords = C['UV'].copy(order = 'C')
+    return uv_coords
+
+def load_pncc_code(path = 'pncc_code.mat'):
+    ''' load pncc code of BFM
+    PNCC code: Defined in 'Face Alignment Across Large Poses: A 3D Solution Xiangyu'
+    download at http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm.
+    Args:
+        path: path to data.
+    Returns:  
+        pncc_code: [nver, 3]
+    '''
+    C = sio.loadmat(path)
+    pncc_code = C['vertex_code'].T
+    return pncc_code
+
+## 
+def get_organ_ind(model_info):
+    ''' get nose, eye, mouth index
+    '''
+    valid_bin = model_info['segbin'].astype(bool)
+    organ_ind = np.nonzero(valid_bin[0,:])[0]
+    for i in range(1, valid_bin.shape[0] - 1):
+        organ_ind = np.union1d(organ_ind, np.nonzero(valid_bin[i,:])[0])
+    return organ_ind.astype(np.int32)
diff --git a/python-package/insightface/thirdparty/face3d/morphable_model/morphabel_model.py b/python-package/insightface/thirdparty/face3d/morphable_model/morphabel_model.py
new file mode 100644
index 000000000..abd677dac
--- /dev/null
+++ b/python-package/insightface/thirdparty/face3d/morphable_model/morphabel_model.py
@@ -0,0 +1,143 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import scipy.io as sio
+from .. import mesh
+from . import fit
+from . import load
+
+class  MorphabelModel(object):
+    """docstring for  MorphabelModel
+    model: nver: number of vertices. ntri: number of triangles. *: must have. ~: can generate ones array for place holder.
+            'shapeMU': [3*nver, 1]. *
+            'shapePC': [3*nver, n_shape_para]. *
+            'shapeEV': [n_shape_para, 1]. ~
+            'expMU': [3*nver, 1]. ~ 
+            'expPC': [3*nver, n_exp_para]. ~
+            'expEV': [n_exp_para, 1]. ~
+            'texMU': [3*nver, 1]. ~
+            'texPC': [3*nver, n_tex_para]. ~
+            'texEV': [n_tex_para, 1]. ~
+            'tri': [ntri, 3] (start from 1, should sub 1 in python and c++). *
+            'tri_mouth': [114, 3] (start from 1, as a supplement to mouth triangles). ~
+            'kpt_ind': [68,] (start from 1). ~
+    """
+    def __init__(self, model_path, model_type = 'BFM'):
+        super( MorphabelModel, self).__init__()
+        if model_type=='BFM':
+            self.model = load.load_BFM(model_path)
+        else:
+            print('sorry, not support other 3DMM model now')
+            exit()
+            
+        # fixed attributes
+        self.nver = self.model['shapePC'].shape[0]/3
+        self.ntri = self.model['tri'].shape[0]
+        self.n_shape_para = self.model['shapePC'].shape[1]
+        self.n_exp_para = self.model['expPC'].shape[1]
+        self.n_tex_para = self.model['texMU'].shape[1]
+        
+        self.kpt_ind = self.model['kpt_ind']
+        self.triangles = self.model['tri']
+        self.full_triangles = np.vstack((self.model['tri'], self.model['tri_mouth']))
+
+    # ------------------------------------- shape: represented with mesh(vertices & triangles(fixed))
+    def get_shape_para(self, type = 'random'):
+        if type == 'zero':
+            sp = np.random.zeros((self.n_shape_para, 1))
+        elif type == 'random':
+            sp = np.random.rand(self.n_shape_para, 1)*1e04
+        return sp
+
+    def get_exp_para(self, type = 'random'):
+        if type == 'zero':
+            ep = np.zeros((self.n_exp_para, 1))
+        elif type == 'random':
+            ep = -1.5 + 3*np.random.random([self.n_exp_para, 1])
+            ep[6:, 0] = 0
+
+        return ep 
+
+    def generate_vertices(self, shape_para, exp_para):
+        '''
+        Args:
+            shape_para: (n_shape_para, 1)
+            exp_para: (n_exp_para, 1) 
+        Returns:
+            vertices: (nver, 3)
+        '''
+        vertices = self.model['shapeMU'] + self.model['shapePC'].dot(shape_para) + self.model['expPC'].dot(exp_para)
+        vertices = np.reshape(vertices, [int(3), int(len(vertices)/3)], 'F').T
+
+        return vertices
+
+    # -------------------------------------- texture: here represented with rgb value(colors) in vertices.
+    def get_tex_para(self, type = 'random'):
+        if type == 'zero':
+            tp = np.zeros((self.n_tex_para, 1))
+        elif type == 'random':
+            tp = np.random.rand(self.n_tex_para, 1)
+        return tp
+
+    def generate_colors(self, tex_para):
+        '''
+        Args:
+            tex_para: (n_tex_para, 1)
+        Returns:
+            colors: (nver, 3)
+        '''
+        colors = self.model['texMU'] + self.model['texPC'].dot(tex_para*self.model['texEV'])
+        colors = np.reshape(colors, [int(3), int(len(colors)/3)], 'F').T/255.  
+        
+        return colors
+
+
+    # ------------------------------------------- transformation
+    # -------------  transform
+    def rotate(self, vertices, angles):
+        ''' rotate face
+        Args:
+            vertices: [nver, 3]
+            angles: [3] x, y, z rotation angle(degree)
+            x: pitch. positive for looking down 
+            y: yaw. positive for looking left
+            z: roll. positive for tilting head right
+        Returns:
+            vertices: rotated vertices
+        '''
+        return mesh.transform.rotate(vertices, angles)
+
+    def transform(self, vertices, s, angles, t3d):
+        R = mesh.transform.angle2matrix(angles)
+        return mesh.transform.similarity_transform(vertices, s, R, t3d)
+
+    def transform_3ddfa(self, vertices, s, angles, t3d): # only used for processing 300W_LP data
+        R = mesh.transform.angle2matrix_3ddfa(angles)
+        return mesh.transform.similarity_transform(vertices, s, R, t3d)
+
+    # --------------------------------------------------- fitting
+    def fit(self, x, X_ind, max_iter = 4, isShow = False):
+        ''' fit 3dmm & pose parameters
+        Args:
+            x: (n, 2) image points
+            X_ind: (n,) corresponding Model vertex indices
+            max_iter: iteration
+            isShow: whether to reserve middle results for show
+        Returns:
+            fitted_sp: (n_sp, 1). shape parameters
+            fitted_ep: (n_ep, 1). exp parameters
+            s, angles, t
+        '''
+        if isShow:
+            fitted_sp, fitted_ep, s, R, t = fit.fit_points_for_show(x, X_ind, self.model, n_sp = self.n_shape_para, n_ep = self.n_exp_para, max_iter = max_iter)
+            angles = np.zeros((R.shape[0], 3))
+            for i in range(R.shape[0]):
+                angles[i] = mesh.transform.matrix2angle(R[i])
+        else:
+            fitted_sp, fitted_ep, s, R, t = fit.fit_points(x, X_ind, self.model, n_sp = self.n_shape_para, n_ep = self.n_exp_para, max_iter = max_iter)
+            angles = mesh.transform.matrix2angle(R)
+        return fitted_sp, fitted_ep, s, angles, t
+
+
diff --git a/python-package/insightface/utils/__init__.py b/python-package/insightface/utils/__init__.py
index 20a689ad9..d3ccb4e65 100644
--- a/python-package/insightface/utils/__init__.py
+++ b/python-package/insightface/utils/__init__.py
@@ -6,9 +6,10 @@
 #from . import metrics
 #from . import parallel
 
-from .download import download, check_sha1
-from .filesystem import makedirs
-from .filesystem import try_import_dali
+from .storage import download, ensure_available, download_onnx
+from .filesystem import get_model_dir
+from .filesystem import makedirs, try_import_dali
+from .constant import *
 #from .bbox import bbox_iou
 #from .block import recursive_visit, set_lr_mult, freeze_bn
 #from .lr_scheduler import LRSequential, LRScheduler
diff --git a/python-package/insightface/utils/constant.py b/python-package/insightface/utils/constant.py
new file mode 100644
index 000000000..8860ff077
--- /dev/null
+++ b/python-package/insightface/utils/constant.py
@@ -0,0 +1,3 @@
+
+DEFAULT_MP_NAME = 'buffalo_l'
+
diff --git a/python-package/insightface/utils/download.py b/python-package/insightface/utils/download.py
index 8f96fe63f..5cda84ded 100644
--- a/python-package/insightface/utils/download.py
+++ b/python-package/insightface/utils/download.py
@@ -6,6 +6,7 @@
 import requests
 from tqdm import tqdm
 
+
 def check_sha1(filename, sha1_hash):
     """Check whether the sha1 hash of the file content matches the expected hash.
     Parameters
@@ -31,7 +32,8 @@ def check_sha1(filename, sha1_hash):
     l = min(len(sha1_file), len(sha1_hash))
     return sha1.hexdigest()[0:l] == sha1_hash[0:l]
 
-def download(url, path=None, overwrite=False, sha1_hash=None):
+
+def download_file(url, path=None, overwrite=False, sha1_hash=None):
     """Download an given URL
     Parameters
     ----------
@@ -59,26 +61,29 @@ def download(url, path=None, overwrite=False, sha1_hash=None):
         else:
             fname = path
 
-    if overwrite or not os.path.exists(fname) or (sha1_hash and not check_sha1(fname, sha1_hash)):
+    if overwrite or not os.path.exists(fname) or (
+            sha1_hash and not check_sha1(fname, sha1_hash)):
         dirname = os.path.dirname(os.path.abspath(os.path.expanduser(fname)))
         if not os.path.exists(dirname):
             os.makedirs(dirname)
 
-        print('Downloading %s from %s...'%(fname, url))
+        print('Downloading %s from %s...' % (fname, url))
         r = requests.get(url, stream=True)
         if r.status_code != 200:
-            raise RuntimeError("Failed downloading url %s"%url)
+            raise RuntimeError("Failed downloading url %s" % url)
         total_length = r.headers.get('content-length')
         with open(fname, 'wb') as f:
-            if total_length is None: # no content length header
+            if total_length is None:  # no content length header
                 for chunk in r.iter_content(chunk_size=1024):
-                    if chunk: # filter out keep-alive new chunks
+                    if chunk:  # filter out keep-alive new chunks
                         f.write(chunk)
             else:
                 total_length = int(total_length)
                 for chunk in tqdm(r.iter_content(chunk_size=1024),
                                   total=int(total_length / 1024. + 0.5),
-                                  unit='KB', unit_scale=False, dynamic_ncols=True):
+                                  unit='KB',
+                                  unit_scale=False,
+                                  dynamic_ncols=True):
                     f.write(chunk)
 
         if sha1_hash and not check_sha1(fname, sha1_hash):
diff --git a/python-package/insightface/utils/face_align.py b/python-package/insightface/utils/face_align.py
index 6cf5a2e64..226628b39 100644
--- a/python-package/insightface/utils/face_align.py
+++ b/python-package/insightface/utils/face_align.py
@@ -1,88 +1,103 @@
-
 import cv2
 import numpy as np
 from skimage import transform as trans
 
-src1 = np.array([
-     [51.642,50.115],
-     [57.617,49.990],
-     [35.740,69.007],
-     [51.157,89.050],
-     [57.025,89.702]], dtype=np.float32)
-#<--left 
-src2 = np.array([
-    [45.031,50.118],
-    [65.568,50.872],
-    [39.677,68.111],
-    [45.177,86.190],
-    [64.246,86.758]], dtype=np.float32)
-
-#---frontal
-src3 = np.array([
-    [39.730,51.138],
-    [72.270,51.138],
-    [56.000,68.493],
-    [42.463,87.010],
-    [69.537,87.010]], dtype=np.float32)
-
-#-->right
-src4 = np.array([
-    [46.845,50.872],
-    [67.382,50.118],
-    [72.737,68.111],
-    [48.167,86.758],
-    [67.236,86.190]], dtype=np.float32)
-
-#-->right profile
-src5 = np.array([
-    [54.796,49.990],
-    [60.771,50.115],
-    [76.673,69.007],
-    [55.388,89.702],
-    [61.257,89.050]], dtype=np.float32)
-
-src = np.array([src1,src2,src3,src4,src5])
-src_map = {112 : src, 224 : src*2}
-
-arcface_src = np.array([
-  [38.2946, 51.6963],
-  [73.5318, 51.5014],
-  [56.0252, 71.7366],
-  [41.5493, 92.3655],
-  [70.7299, 92.2041] ], dtype=np.float32 )
-
-arcface_src = np.expand_dims(arcface_src, axis=0)
-
-# In[66]:
-
-# lmk is prediction; src is template
-def estimate_norm(lmk, image_size = 112, mode='arcface'):
-  assert lmk.shape==(5,2)
-  tform = trans.SimilarityTransform()
-  lmk_tran = np.insert(lmk, 2, values=np.ones(5), axis=1)
-  min_M = []
-  min_index = []
-  min_error = float('inf') 
-  if mode=='arcface':
-    assert image_size==112
-    src = arcface_src
-  else:
-    src = src_map[image_size]
-  for i in np.arange(src.shape[0]):
-    tform.estimate(lmk, src[i])
-    M = tform.params[0:2,:]
-    results = np.dot(M, lmk_tran.T)
-    results = results.T
-    error = np.sum(np.sqrt(np.sum((results - src[i]) ** 2,axis=1)))
-#         print(error)
-    if error< min_error:
-        min_error = error
-        min_M = M
-        min_index = i
-  return min_M, min_index
+
+arcface_dst = np.array(
+    [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+     [41.5493, 92.3655], [70.7299, 92.2041]],
+    dtype=np.float32)
+
+def estimate_norm(lmk, image_size=112,mode='arcface'):
+    assert lmk.shape == (5, 2)
+    assert image_size%112==0 or image_size%128==0
+    if image_size%112==0:
+        ratio = float(image_size)/112.0
+        diff_x = 0
+    else:
+        ratio = float(image_size)/128.0
+        diff_x = 8.0*ratio
+    dst = arcface_dst * ratio
+    dst[:,0] += diff_x
+    tform = trans.SimilarityTransform()
+    tform.estimate(lmk, dst)
+    M = tform.params[0:2, :]
+    return M
 
 def norm_crop(img, landmark, image_size=112, mode='arcface'):
-  M, pose_index = estimate_norm(landmark, image_size, mode)
-  warped = cv2.warpAffine(img,M, (image_size, image_size), borderValue = 0.0)
-  return warped
+    M = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped
+
+def norm_crop2(img, landmark, image_size=112, mode='arcface'):
+    M = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped, M
+
+def square_crop(im, S):
+    if im.shape[0] > im.shape[1]:
+        height = S
+        width = int(float(im.shape[1]) / im.shape[0] * S)
+        scale = float(S) / im.shape[0]
+    else:
+        width = S
+        height = int(float(im.shape[0]) / im.shape[1] * S)
+        scale = float(S) / im.shape[1]
+    resized_im = cv2.resize(im, (width, height))
+    det_im = np.zeros((S, S, 3), dtype=np.uint8)
+    det_im[:resized_im.shape[0], :resized_im.shape[1], :] = resized_im
+    return det_im, scale
+
+
+def transform(data, center, output_size, scale, rotation):
+    scale_ratio = scale
+    rot = float(rotation) * np.pi / 180.0
+    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
+    t1 = trans.SimilarityTransform(scale=scale_ratio)
+    cx = center[0] * scale_ratio
+    cy = center[1] * scale_ratio
+    t2 = trans.SimilarityTransform(translation=(-1 * cx, -1 * cy))
+    t3 = trans.SimilarityTransform(rotation=rot)
+    t4 = trans.SimilarityTransform(translation=(output_size / 2,
+                                                output_size / 2))
+    t = t1 + t2 + t3 + t4
+    M = t.params[0:2]
+    cropped = cv2.warpAffine(data,
+                             M, (output_size, output_size),
+                             borderValue=0.0)
+    return cropped, M
+
+
+def trans_points2d(pts, M):
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i] = new_pt[0:2]
+
+    return new_pts
+
+
+def trans_points3d(pts, M):
+    scale = np.sqrt(M[0][0] * M[0][0] + M[0][1] * M[0][1])
+    #print(scale)
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i][0:2] = new_pt[0:2]
+        new_pts[i][2] = pts[i][2] * scale
+
+    return new_pts
+
+
+def trans_points(pts, M):
+    if pts.shape[1] == 2:
+        return trans_points2d(pts, M)
+    else:
+        return trans_points3d(pts, M)
 
diff --git a/python-package/insightface/utils/filesystem.py b/python-package/insightface/utils/filesystem.py
index 4ecb091dd..01e385197 100644
--- a/python-package/insightface/utils/filesystem.py
+++ b/python-package/insightface/utils/filesystem.py
@@ -2,8 +2,15 @@
 This code file mainly comes from https://github.com/dmlc/gluon-cv/blob/master/gluoncv/utils/filesystem.py
 """
 import os
+import os.path as osp
 import errno
 
+
+def get_model_dir(name, root='~/.insightface'):
+    root = os.path.expanduser(root)
+    model_dir = osp.join(root, 'models', name)
+    return model_dir
+
 def makedirs(path):
     """Create directory recursively if not exists.
     Similar to `makedir -p`, you can skip checking existence before this function.
@@ -19,6 +26,7 @@ def makedirs(path):
         if exc.errno != errno.EEXIST:
             raise
 
+
 def try_import(package, message=None):
     """Try import specified package, with custom message support.
 
@@ -42,6 +50,7 @@ def try_import(package, message=None):
             raise e
         raise ImportError(message)
 
+
 def try_import_cv2():
     """Try import cv2 at runtime.
 
@@ -52,8 +61,10 @@ def try_import_cv2():
     """
     msg = "cv2 is required, you can install by package manager, e.g. 'apt-get', \
         or `pip install opencv-python --user` (note that this is unofficial PYPI package)."
+
     return try_import('cv2', msg)
 
+
 def try_import_mmcv():
     """Try import mmcv at runtime.
 
@@ -64,8 +75,10 @@ def try_import_mmcv():
     """
     msg = "mmcv is required, you can install by first `pip install Cython --user` \
         and then `pip install mmcv --user` (note that this is unofficial PYPI package)."
+
     return try_import('mmcv', msg)
 
+
 def try_import_rarfile():
     """Try import rarfile at runtime.
 
@@ -76,8 +89,10 @@ def try_import_rarfile():
     """
     msg = "rarfile is required, you can install by first `sudo apt-get install unrar` \
         and then `pip install rarfile --user` (note that this is unofficial PYPI package)."
+
     return try_import('rarfile', msg)
 
+
 def import_try_install(package, extern_url=None):
     """Try import the specified package.
     If the package not installed, try use pip to install and import if success.
@@ -108,7 +123,8 @@ def import_try_install(package, extern_url=None):
 
         # trying to install package
         url = package if extern_url is None else extern_url
-        pipmain(['install', '--user', url])  # will raise SystemExit Error if fails
+        pipmain(['install', '--user',
+                 url])  # will raise SystemExit Error if fails
 
         # trying to load again
         try:
@@ -122,6 +138,7 @@ def import_try_install(package, extern_url=None):
             return __import__(package)
     return __import__(package)
 
+
 def try_import_dali():
     """Try import NVIDIA DALI at runtime.
     """
@@ -129,9 +146,12 @@ def try_import_dali():
         dali = __import__('nvidia.dali', fromlist=['pipeline', 'ops', 'types'])
         dali.Pipeline = dali.pipeline.Pipeline
     except ImportError:
+
         class dali:
             class Pipeline:
                 def __init__(self):
                     raise NotImplementedError(
-                        "DALI not found, please check if you installed it correctly.")
+                        "DALI not found, please check if you installed it correctly."
+                    )
+
     return dali
diff --git a/python-package/insightface/utils/storage.py b/python-package/insightface/utils/storage.py
new file mode 100644
index 000000000..5bf37e2d1
--- /dev/null
+++ b/python-package/insightface/utils/storage.py
@@ -0,0 +1,52 @@
+
+import os
+import os.path as osp
+import zipfile
+from .download import download_file
+
+BASE_REPO_URL = 'https://github.com/deepinsight/insightface/releases/download/v0.7'
+
+def download(sub_dir, name, force=False, root='~/.insightface'):
+    _root = os.path.expanduser(root)
+    dir_path = os.path.join(_root, sub_dir, name)
+    if osp.exists(dir_path) and not force:
+        return dir_path
+    print('download_path:', dir_path)
+    zip_file_path = os.path.join(_root, sub_dir, name + '.zip')
+    model_url = "%s/%s.zip"%(BASE_REPO_URL, name)
+    download_file(model_url,
+             path=zip_file_path,
+             overwrite=True)
+    if not os.path.exists(dir_path):
+        os.makedirs(dir_path)
+    with zipfile.ZipFile(zip_file_path) as zf:
+        zf.extractall(dir_path)
+    #os.remove(zip_file_path)
+    return dir_path
+
+def ensure_available(sub_dir, name, root='~/.insightface'):
+    return download(sub_dir, name, force=False, root=root)
+
+def download_onnx(sub_dir, model_file, force=False, root='~/.insightface', download_zip=False):
+    _root = os.path.expanduser(root)
+    model_root = osp.join(_root, sub_dir)
+    new_model_file = osp.join(model_root, model_file)
+    if osp.exists(new_model_file) and not force:
+        return new_model_file
+    if not osp.exists(model_root):
+        os.makedirs(model_root)
+    print('download_path:', new_model_file)
+    if not download_zip:
+        model_url = "%s/%s"%(BASE_REPO_URL, model_file)
+        download_file(model_url,
+                 path=new_model_file,
+                 overwrite=True)
+    else:
+        model_url = "%s/%s.zip"%(BASE_REPO_URL, model_file)
+        zip_file_path = new_model_file+".zip"
+        download_file(model_url,
+                 path=zip_file_path,
+                 overwrite=True)
+        with zipfile.ZipFile(zip_file_path) as zf:
+            zf.extractall(model_root)
+        return new_model_file
diff --git a/python-package/insightface/utils/transform.py b/python-package/insightface/utils/transform.py
new file mode 100644
index 000000000..06531d257
--- /dev/null
+++ b/python-package/insightface/utils/transform.py
@@ -0,0 +1,116 @@
+import cv2
+import math
+import numpy as np
+from skimage import transform as trans
+
+
+def transform(data, center, output_size, scale, rotation):
+    scale_ratio = scale
+    rot = float(rotation) * np.pi / 180.0
+    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
+    t1 = trans.SimilarityTransform(scale=scale_ratio)
+    cx = center[0] * scale_ratio
+    cy = center[1] * scale_ratio
+    t2 = trans.SimilarityTransform(translation=(-1 * cx, -1 * cy))
+    t3 = trans.SimilarityTransform(rotation=rot)
+    t4 = trans.SimilarityTransform(translation=(output_size / 2,
+                                                output_size / 2))
+    t = t1 + t2 + t3 + t4
+    M = t.params[0:2]
+    cropped = cv2.warpAffine(data,
+                             M, (output_size, output_size),
+                             borderValue=0.0)
+    return cropped, M
+
+
+def trans_points2d(pts, M):
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i] = new_pt[0:2]
+
+    return new_pts
+
+
+def trans_points3d(pts, M):
+    scale = np.sqrt(M[0][0] * M[0][0] + M[0][1] * M[0][1])
+    #print(scale)
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i][0:2] = new_pt[0:2]
+        new_pts[i][2] = pts[i][2] * scale
+
+    return new_pts
+
+
+def trans_points(pts, M):
+    if pts.shape[1] == 2:
+        return trans_points2d(pts, M)
+    else:
+        return trans_points3d(pts, M)
+
+def estimate_affine_matrix_3d23d(X, Y):
+    ''' Using least-squares solution 
+    Args:
+        X: [n, 3]. 3d points(fixed)
+        Y: [n, 3]. corresponding 3d points(moving). Y = PX
+    Returns:
+        P_Affine: (3, 4). Affine camera matrix (the third row is [0, 0, 0, 1]).
+    '''
+    X_homo = np.hstack((X, np.ones([X.shape[0],1]))) #n x 4
+    P = np.linalg.lstsq(X_homo, Y)[0].T # Affine matrix. 3 x 4
+    return P
+
+def P2sRt(P):
+    ''' decompositing camera matrix P
+    Args: 
+        P: (3, 4). Affine Camera Matrix.
+    Returns:
+        s: scale factor.
+        R: (3, 3). rotation matrix.
+        t: (3,). translation. 
+    '''
+    t = P[:, 3]
+    R1 = P[0:1, :3]
+    R2 = P[1:2, :3]
+    s = (np.linalg.norm(R1) + np.linalg.norm(R2))/2.0
+    r1 = R1/np.linalg.norm(R1)
+    r2 = R2/np.linalg.norm(R2)
+    r3 = np.cross(r1, r2)
+
+    R = np.concatenate((r1, r2, r3), 0)
+    return s, R, t
+
+def matrix2angle(R):
+    ''' get three Euler angles from Rotation Matrix
+    Args:
+        R: (3,3). rotation matrix
+    Returns:
+        x: pitch
+        y: yaw
+        z: roll
+    '''
+    sy = math.sqrt(R[0,0] * R[0,0] +  R[1,0] * R[1,0])
+     
+    singular = sy < 1e-6
+ 
+    if  not singular :
+        x = math.atan2(R[2,1] , R[2,2])
+        y = math.atan2(-R[2,0], sy)
+        z = math.atan2(R[1,0], R[0,0])
+    else :
+        x = math.atan2(-R[1,2], R[1,1])
+        y = math.atan2(-R[2,0], sy)
+        z = 0
+
+    # rx, ry, rz = np.rad2deg(x), np.rad2deg(y), np.rad2deg(z)
+    rx, ry, rz = x*180/np.pi, y*180/np.pi, z*180/np.pi
+    return rx, ry, rz
+
diff --git a/python-package/packaging/desktop/build_linux.sh b/python-package/packaging/desktop/build_linux.sh
new file mode 100755
index 000000000..e2a8f171e
--- /dev/null
+++ b/python-package/packaging/desktop/build_linux.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+echo "Building InsightFace Evaluation Studio for Linux..."
+python -m PyInstaller --noconfirm packaging/desktop/pyinstaller.spec
+echo "Build complete. Output: dist/InsightFace Evaluation Studio/"
diff --git a/python-package/packaging/desktop/build_macos.sh b/python-package/packaging/desktop/build_macos.sh
new file mode 100755
index 000000000..d4722769e
--- /dev/null
+++ b/python-package/packaging/desktop/build_macos.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+echo "Building InsightFace Evaluation Studio for macOS..."
+python -m PyInstaller --noconfirm packaging/desktop/pyinstaller.spec
+echo "Build complete. Output: dist/InsightFace Evaluation Studio.app or dist/InsightFace Evaluation Studio/"
diff --git a/python-package/packaging/desktop/build_windows.ps1 b/python-package/packaging/desktop/build_windows.ps1
new file mode 100644
index 000000000..025b4f14b
--- /dev/null
+++ b/python-package/packaging/desktop/build_windows.ps1
@@ -0,0 +1,5 @@
+$ErrorActionPreference = "Stop"
+
+Write-Host "Building InsightFace Evaluation Studio for Windows..."
+python -m PyInstaller --noconfirm packaging/desktop/pyinstaller.spec
+Write-Host "Build complete. Output: dist/InsightFace Evaluation Studio/"
diff --git a/python-package/packaging/desktop/pyinstaller_entry.py b/python-package/packaging/desktop/pyinstaller_entry.py
new file mode 100644
index 000000000..5b42fa13e
--- /dev/null
+++ b/python-package/packaging/desktop/pyinstaller_entry.py
@@ -0,0 +1,5 @@
+from insightface.gui.__main__ import main
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/python-package/packaging/pypi/build_upload_pypi.sh b/python-package/packaging/pypi/build_upload_pypi.sh
new file mode 100755
index 000000000..84279f962
--- /dev/null
+++ b/python-package/packaging/pypi/build_upload_pypi.sh
@@ -0,0 +1,244 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PACKAGE_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
+
+DRY_RUN=0
+SKIP_TESTS=0
+ALLOW_DIRTY=0
+WITH_FACE3D=0
+NO_CLEAN=0
+SKIP_EXISTING_CHECK=0
+REPOSITORY="pypi"
+PYTHON_BIN="${PYTHON_BIN:-python}"
+
+usage() {
+  cat <<'EOF'
+Build and upload the InsightFace package to official PyPI.
+
+Usage:
+  bash packaging/pypi/build_upload_pypi.sh [options]
+
+Options:
+  --dry-run              Build and run twine check only. Do not upload.
+  --skip-tests           Skip the local pytest release smoke suite.
+  --allow-dirty          Allow publishing from a dirty git working tree.
+  --with-face3d          Include the optional face3d Cython/C++ extension.
+  --no-clean             Keep existing build/, dist/, and *.egg-info files.
+  --skip-existing-check  Skip the PyPI "version already exists" check.
+  --repository NAME      Twine repository name. Defaults to "pypi".
+  --python PATH          Python executable to use. Defaults to $PYTHON_BIN or python.
+  -h, --help             Show this help message.
+
+Credentials:
+  Recommended token upload:
+    export TWINE_USERNAME=__token__
+    export TWINE_PASSWORD='pypi-...'
+
+  Twine also supports ~/.pypirc, keyring, and PyPI Trusted Publisher in CI.
+
+Examples:
+  bash packaging/pypi/build_upload_pypi.sh --dry-run
+  bash packaging/pypi/build_upload_pypi.sh
+  bash packaging/pypi/build_upload_pypi.sh --with-face3d
+EOF
+}
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --dry-run)
+      DRY_RUN=1
+      ;;
+    --skip-tests)
+      SKIP_TESTS=1
+      ;;
+    --allow-dirty)
+      ALLOW_DIRTY=1
+      ;;
+    --with-face3d)
+      WITH_FACE3D=1
+      ;;
+    --no-clean)
+      NO_CLEAN=1
+      ;;
+    --skip-existing-check)
+      SKIP_EXISTING_CHECK=1
+      ;;
+    --repository)
+      shift
+      REPOSITORY="${1:-}"
+      ;;
+    --python)
+      shift
+      PYTHON_BIN="${1:-}"
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown option: $1" >&2
+      usage
+      exit 2
+      ;;
+  esac
+  shift
+done
+
+if [[ -z "${PYTHON_BIN}" ]]; then
+  echo "Python executable is empty." >&2
+  exit 2
+fi
+
+if [[ "${REPOSITORY}" != "pypi" ]]; then
+  echo "This script is intended for the official PyPI release path." >&2
+  echo "Use --repository pypi, or create a separate TestPyPI script if needed." >&2
+  exit 2
+fi
+
+cd "${PACKAGE_ROOT}"
+
+echo "==> Package root: ${PACKAGE_ROOT}"
+echo "==> Python: $(${PYTHON_BIN} -c 'import sys; print(sys.executable)')"
+
+read -r PACKAGE_NAME PACKAGE_VERSION < <("${PYTHON_BIN}" - <<'PY'
+from pathlib import Path
+import re
+
+root = Path.cwd()
+setup_text = (root / "setup.py").read_text(encoding="utf-8")
+init_text = (root / "insightface" / "__init__.py").read_text(encoding="utf-8")
+
+name_match = re.search(r"name\s*=\s*['\"]([^'\"]+)['\"]", setup_text)
+version_match = re.search(r"^__version__\s*=\s*['\"]([^'\"]+)['\"]", init_text, re.M)
+
+if not name_match:
+    raise SystemExit("Unable to find package name in setup.py")
+if not version_match:
+    raise SystemExit("Unable to find __version__ in insightface/__init__.py")
+
+print(name_match.group(1), version_match.group(1))
+PY
+)
+
+echo "==> Package: ${PACKAGE_NAME} ${PACKAGE_VERSION}"
+
+if [[ "${PACKAGE_NAME}" != "insightface" ]]; then
+  echo "Refusing to upload package '${PACKAGE_NAME}' to official PyPI." >&2
+  echo "Expected package name: insightface" >&2
+  exit 1
+fi
+
+if [[ "${ALLOW_DIRTY}" -ne 1 ]]; then
+  if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+    if [[ -n "$(git status --porcelain)" ]]; then
+      echo "Git working tree is dirty. Commit or stash changes before release." >&2
+      echo "Use --allow-dirty only for an intentional emergency release." >&2
+      git status --short
+      exit 1
+    fi
+  fi
+fi
+
+echo "==> Checking build tools"
+"${PYTHON_BIN}" - "${SKIP_TESTS}" <<'PY'
+import importlib.util
+import sys
+
+skip_tests = sys.argv[1] == "1"
+required = ["build", "twine"]
+if not skip_tests:
+    required.append("pytest")
+
+missing = [name for name in required if importlib.util.find_spec(name) is None]
+if missing:
+    print("Missing required packaging tools: " + ", ".join(missing))
+    print("Install them with: python -m pip install build twine pytest")
+    raise SystemExit(1)
+PY
+
+if [[ "${SKIP_TESTS}" -ne 1 ]]; then
+  echo "==> Running release smoke tests"
+  "${PYTHON_BIN}" -m pytest -q tests/gui
+else
+  echo "==> Skipping tests"
+fi
+
+if [[ "${SKIP_EXISTING_CHECK}" -ne 1 ]]; then
+  echo "==> Checking whether ${PACKAGE_NAME} ${PACKAGE_VERSION} already exists on PyPI"
+"${PYTHON_BIN}" - "${PACKAGE_NAME}" "${PACKAGE_VERSION}" <<'PY'
+import sys
+import urllib.error
+import urllib.request
+
+name, version = sys.argv[1], sys.argv[2]
+url = f"https://pypi.org/pypi/{name}/{version}/json"
+try:
+    with urllib.request.urlopen(url, timeout=20) as response:
+        response.read(1)
+except urllib.error.HTTPError as exc:
+    if exc.code == 404:
+        print("Version is not present on PyPI.")
+    else:
+        raise
+else:
+    print(f"{name} {version} already exists on PyPI. PyPI versions are immutable.")
+    raise SystemExit(1)
+PY
+fi
+
+if [[ "${NO_CLEAN}" -ne 1 ]]; then
+  echo "==> Cleaning build artifacts"
+  rm -rf build dist ./*.egg-info
+fi
+
+if [[ "${WITH_FACE3D}" -eq 1 ]]; then
+  export INSIGHTFACE_WITH_FACE3D=1
+  echo "==> face3d build: enabled"
+else
+  unset INSIGHTFACE_WITH_FACE3D || true
+  echo "==> face3d build: disabled"
+fi
+
+echo "==> Building source distribution and wheel"
+"${PYTHON_BIN}" -m build
+
+echo "==> Checking distributions"
+"${PYTHON_BIN}" -m twine check dist/*
+
+echo "==> Built artifacts"
+ls -lh dist
+
+if [[ "${DRY_RUN}" -eq 1 ]]; then
+  echo "Dry run complete. No upload was performed."
+  exit 0
+fi
+
+echo
+echo "You are about to upload ${PACKAGE_NAME} ${PACKAGE_VERSION} to official PyPI."
+echo "This cannot be overwritten or deleted for normal users after publishing."
+echo
+printf "Type exactly 'upload %s %s to pypi' to continue: " "${PACKAGE_NAME}" "${PACKAGE_VERSION}"
+read -r CONFIRMATION
+EXPECTED="upload ${PACKAGE_NAME} ${PACKAGE_VERSION} to pypi"
+if [[ "${CONFIRMATION}" != "${EXPECTED}" ]]; then
+  echo "Confirmation did not match. Upload canceled."
+  exit 1
+fi
+
+echo "==> Uploading to official PyPI"
+"${PYTHON_BIN}" -m twine upload --repository "${REPOSITORY}" dist/*
+
+cat <<RELEASE_DONE
+
+Upload complete.
+
+Verify the release:
+  python -m pip install --upgrade --no-cache-dir ${PACKAGE_NAME}==${PACKAGE_VERSION}
+  python -c "import insightface; print(insightface.__version__)"
+
+GUI extra:
+  python -m pip install --upgrade --no-cache-dir '${PACKAGE_NAME}[gui]==${PACKAGE_VERSION}'
+  insightface-gui --version
+RELEASE_DONE
diff --git a/python-package/pyproject.toml b/python-package/pyproject.toml
new file mode 100644
index 000000000..fed528d4a
--- /dev/null
+++ b/python-package/pyproject.toml
@@ -0,0 +1,3 @@
+[build-system]
+requires = ["setuptools"]
+build-backend = "setuptools.build_meta"
diff --git a/python-package/setup.py b/python-package/setup.py
index c48bb2930..2cbab0cda 100644
--- a/python-package/setup.py
+++ b/python-package/setup.py
@@ -2,18 +2,30 @@
 import os
 import io
 import re
-import shutil
 import sys
-from setuptools import setup, find_packages
+import subprocess
+import platform
+import logging
+from setuptools import setup, find_namespace_packages
+
+
+FACE3D_BUILD_FLAG = '--with-face3d'
+
+
+def strtobool_env(value):
+    return str(value).strip().lower() in ('1', 'true', 'yes', 'on')
+
+
+build_face3d = strtobool_env(os.environ.get('INSIGHTFACE_WITH_FACE3D', ''))
+if FACE3D_BUILD_FLAG in sys.argv:
+    build_face3d = True
+    sys.argv.remove(FACE3D_BUILD_FLAG)
 
 def read(*names, **kwargs):
-    with io.open(
-        os.path.join(os.path.dirname(__file__), *names),
-        encoding=kwargs.get("encoding", "utf8")
-    ) as fp:
+    with io.open(os.path.join(os.path.dirname(__file__), *names),
+                 encoding=kwargs.get("encoding", "utf8")) as fp:
         return fp.read()
 
-
 def find_version(*file_paths):
     version_file = read(*file_paths)
     version_match = re.search(r"^__version__ = ['\"]([^'\"]*)['\"]",
@@ -22,40 +34,151 @@ def find_version(*file_paths):
         return version_match.group(1)
     raise RuntimeError("Unable to find version string.")
 
+pypandoc_enabled = True
 try:
     import pypandoc
-    long_description = pypandoc.convert('README.md', 'rst')
-except(IOError, ImportError):
-    long_description = open('README.md').read()
+    print('pandoc enabled')
+    long_description = pypandoc.convert_file('README.md', 'rst')
+except (IOError, ImportError, ModuleNotFoundError):
+    print('WARNING: pandoc not enabled')
+    long_description = read('README.md')
+    pypandoc_enabled = False
 
+#import pypandoc
+#long_description = pypandoc.convert('README.md', 'rst')
 VERSION = find_version('insightface', '__init__.py')
 
 requirements = [
     'numpy',
+    'onnx',
+    'onnxruntime',
+    'opencv-python',
     'tqdm',
     'requests',
-    'matplotlib',
-    'Pillow',
     'scipy',
-    'opencv-python',
-    'scikit-learn',
+    #'opencv-python',
     'scikit-image',
-    'easydict',
 ]
 
+gui_requirements = [
+    'PySide6-Essentials>=6.5',
+    'Pillow',
+    'reportlab',
+    'scikit-learn',
+]
+
+face3d_requirements = [
+    'cython',
+    'albumentations',
+    'matplotlib',
+]
+
+package_data = {
+    "insightface.data.images": ["*.jpg", "*.jpeg", "*.png"],
+    "insightface.data.objects": ["*.pkl"],
+    "insightface.gui.assets": ["*.png", "*.ico", "*.icns"],
+}
+
+packages = find_namespace_packages(
+    include=("insightface", "insightface.*"),
+    exclude=("docs", "docs.*", "tests", "tests.*", "scripts", "scripts.*"),
+)
+if not build_face3d:
+    packages = [
+        package
+        for package in packages
+        if package != "insightface.thirdparty.face3d.mesh.cython"
+    ]
+
+ext_modules = []
+include_dirs = []
+headers = []
+if build_face3d:
+    import numpy
+    from distutils.core import Extension
+    from Cython.Build import cythonize
+
+    extensions = [
+            Extension("insightface.thirdparty.face3d.mesh.cython.mesh_core_cython",
+                ["insightface/thirdparty/face3d/mesh/cython/mesh_core_cython.pyx", "insightface/thirdparty/face3d/mesh/cython/mesh_core.cpp"], language='c++'),
+            ]
+    package_data["insightface.thirdparty.face3d.mesh.cython"] = [
+        "*.h",
+        "*.c",
+        "*.cpp",
+        "*.pyx",
+        "*.py",
+    ]
+    ext_modules = cythonize(extensions)
+    include_dirs = [numpy.get_include()]
+    headers = ['insightface/thirdparty/face3d/mesh/cython/mesh_core.h']
+
+# Configure logging
+logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+# Check if running on macOS (Darwin)
+if platform.system() == "Darwin":
+    logging.info("Detected macOS. Checking if Homebrew, LLVM, and OpenMP are installed...")
+
+    # Check if Homebrew is installed
+    brew_check = subprocess.run(["which", "brew"], capture_output=True, text=True)
+    if brew_check.returncode != 0:
+        logging.warning("Homebrew is not installed. You may need to manually install dependencies.")
+        logging.warning("   Install Homebrew: https://brew.sh/")
+        logging.warning("   Then, run: brew install llvm libomp")
+        logging.info("Proceeding without setting the compiler.")
+    
+    else:
+        # Check if LLVM is installed
+        llvm_check = subprocess.run(["brew", "--prefix", "llvm"], capture_output=True, text=True)
+        if llvm_check.returncode != 0:
+            logging.warning("LLVM is not installed. This may cause installation issues.")
+            logging.warning("   To install, run: brew install llvm libomp")
+            logging.warning("   Then, restart the installation process.")
+            logging.info("Proceeding without setting the compiler.")
+
+        else:
+            # Set compiler dynamically if LLVM is installed
+            llvm_path = subprocess.getoutput("brew --prefix llvm")
+            llvm_cc = f"{llvm_path}/bin/clang"
+            llvm_cxx = f"{llvm_path}/bin/clang++"
+            if os.path.exists(llvm_cc) and os.path.exists(llvm_cxx):
+                os.environ["CC"] = llvm_cc
+                os.environ["CXX"] = llvm_cxx
+                logging.info(f"Using compiler: {os.environ['CC']}")
+            else:
+                logging.warning("Homebrew LLVM path exists but clang/clang++ was not found.")
+                logging.info("Falling back to the default system compiler.")
+
 setup(
     # Metadata
     name='insightface',
     version=VERSION,
     author='InsightFace Contributors',
+    author_email='contact@insightface.ai',
     url='https://github.com/deepinsight/insightface',
-    description='InsightFace Toolkit',
+    description='InsightFace Python Library',
     long_description=long_description,
-    license='Apache-2.0',
+    long_description_content_type='text/markdown',
     # Package info
-    packages=find_packages(exclude=('docs', 'tests', 'scripts')),
+    packages=packages,
+    package_data=package_data,
     zip_safe=True,
-    include_package_data=True,
+    include_package_data=False,
+    entry_points={
+        "console_scripts": [
+            "insightface-cli=insightface.commands.insightface_cli:main",
+            "insightface-gui=insightface.gui.__main__:main",
+            "insightface-eval-studio=insightface.gui.__main__:main",
+            "insightface-desktop=insightface.gui.__main__:main",
+        ]
+    },
+    extras_require={"gui": gui_requirements, "face3d": face3d_requirements},
     install_requires=requirements,
+    headers=headers,
+    ext_modules=ext_modules,
+    include_dirs=include_dirs,
 )
 
+print('pypandoc enabled:', pypandoc_enabled)
+print('face3d build enabled:', build_face3d)
diff --git a/python-package/tests/gui/test_cli.py b/python-package/tests/gui/test_cli.py
new file mode 100644
index 000000000..40df90a02
--- /dev/null
+++ b/python-package/tests/gui/test_cli.py
@@ -0,0 +1,40 @@
+from insightface.gui.__main__ import main
+from insightface.gui.app import create_context
+from insightface.gui.core.config import AppConfig, save_config
+
+
+def test_cli_import_and_version(capsys):
+    import insightface
+    import insightface.gui
+
+    assert insightface.__version__ == "1.0.1"
+    assert insightface.gui.__version__ == "1.0.1"
+    assert main(["--version"]) == 0
+    out = capsys.readouterr().out
+    assert "InsightFace Evaluation Studio 1.0.1" in out
+    assert "insightface 1.0.1" in out
+
+
+def test_insightface_cli_import_does_not_require_mxnet():
+    from insightface.commands import insightface_cli
+
+    assert callable(insightface_cli.main)
+
+
+def test_safe_mode_is_runtime_only(tmp_path):
+    workspace = tmp_path / "workspace"
+    cfg = AppConfig(workspace_path=str(workspace))
+    cfg.safe_mode = True
+    cfg.auto_load_model = False
+    save_config(cfg)
+
+    args = type(
+        "Args",
+        (),
+        {"workspace": str(workspace), "model": None, "provider": None, "safe_mode": True},
+    )
+    context = create_context(args())
+
+    assert context.runtime_safe_mode is True
+    assert context.config.safe_mode is False
+    assert context.config.auto_load_model is True
diff --git a/python-package/tests/gui/test_clustering.py b/python-package/tests/gui/test_clustering.py
new file mode 100644
index 000000000..3c81e34cf
--- /dev/null
+++ b/python-package/tests/gui/test_clustering.py
@@ -0,0 +1,61 @@
+import numpy as np
+
+from insightface.gui.core.clustering import cluster_embeddings_dbscan
+
+
+def test_dbscan_default_distance_threshold_groups_near_faces():
+    embeddings = [
+        np.array([1.0, 0.0], dtype=np.float32),
+        np.array([0.98, 0.08], dtype=np.float32),
+        np.array([0.0, 1.0], dtype=np.float32),
+        np.array([0.08, 0.98], dtype=np.float32),
+    ]
+
+    labels, algorithm = cluster_embeddings_dbscan(embeddings, min_samples=2)
+
+    assert algorithm in {"DBSCAN", "centroid fallback"}
+    assert len(labels) == 4
+    assert labels[0] == labels[1]
+    assert labels[2] == labels[3]
+    assert labels[0] != labels[2]
+
+
+def test_dbscan_threshold_can_be_tightened_by_user():
+    embeddings = [
+        np.array([1.0, 0.0], dtype=np.float32),
+        np.array([0.95, 0.32], dtype=np.float32),
+    ]
+
+    loose_labels, loose_algorithm = cluster_embeddings_dbscan(
+        embeddings, distance_threshold=0.28, min_samples=2
+    )
+    strict_labels, strict_algorithm = cluster_embeddings_dbscan(
+        embeddings, distance_threshold=0.001, min_samples=2
+    )
+
+    assert loose_algorithm in {"DBSCAN", "centroid fallback"}
+    assert strict_algorithm in {"DBSCAN", "centroid fallback"}
+    assert loose_labels[0] == loose_labels[1]
+    assert strict_labels[0] != strict_labels[1] or strict_labels[0] == -1
+
+
+def test_dbscan_has_centroid_fallback_when_sklearn_missing(monkeypatch):
+    import builtins
+
+    real_import = builtins.__import__
+
+    def fake_import(name, *args, **kwargs):
+        if name == "sklearn.cluster":
+            raise ImportError("missing sklearn for test")
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", fake_import)
+
+    labels, algorithm = cluster_embeddings_dbscan(
+        [np.array([1.0, 0.0], dtype=np.float32), np.array([0.98, 0.08], dtype=np.float32)],
+        distance_threshold=0.28,
+        min_samples=2,
+    )
+
+    assert algorithm == "centroid fallback"
+    assert labels == [0, 0]
diff --git a/python-package/tests/gui/test_detector_det_size.py b/python-package/tests/gui/test_detector_det_size.py
new file mode 100644
index 000000000..cd1565476
--- /dev/null
+++ b/python-package/tests/gui/test_detector_det_size.py
@@ -0,0 +1,87 @@
+from insightface.app import FaceAnalysis
+from insightface.model_zoo import model_zoo
+from insightface.model_zoo.retinaface import RetinaFace
+from insightface.model_zoo.scrfd import DEFAULT_DET_SIZES, SCRFD
+from insightface.gui.core.face_engine import FaceEngine
+
+
+def test_scrfd_accepts_multi_det_size_config():
+    assert SCRFD._normalize_input_sizes([(128, 128), (640, 640)]) == [(128, 128), (640, 640)]
+    assert SCRFD._normalize_input_sizes((320, 320)) == [(320, 320)]
+    assert SCRFD._normalize_input_sizes((0, 0)) == DEFAULT_DET_SIZES
+
+
+def test_scrfd_default_dynamic_det_size_is_auto():
+    detector = SCRFD.__new__(SCRFD)
+    detector.input_sizes = []
+    detector.input_size = None
+
+    assert detector._resolve_input_sizes(None) == DEFAULT_DET_SIZES
+
+
+def test_retinaface_accepts_multi_det_size_config():
+    assert RetinaFace._normalize_input_sizes([(128, 128), (640, 640)]) == [(128, 128), (640, 640)]
+    assert RetinaFace._normalize_input_sizes([1024, 1024]) == [(1024, 1024)]
+
+
+def test_gui_routes_detection_outputs_to_scrfd():
+    assert FaceEngine._is_scrfd_detection_outputs([object()] * 6) is True
+    assert FaceEngine._is_scrfd_detection_outputs([object()] * 9) is True
+    assert FaceEngine._is_scrfd_detection_outputs([object()] * 10) is True
+    assert FaceEngine._is_scrfd_detection_outputs([object()] * 15) is True
+    assert FaceEngine._is_scrfd_detection_outputs([object()]) is False
+
+
+def test_model_zoo_routes_detection_models_to_scrfd(monkeypatch):
+    class FakeInput:
+        name = "input"
+        shape = [1, 3, "height", "width"]
+
+    class FakeOutput:
+        name = "output"
+        shape = [1, 10]
+
+    class FakeSession:
+        _providers = ["CPUExecutionProvider"]
+        _provider_options = [{}]
+
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def get_inputs(self):
+            return [FakeInput()]
+
+        def get_outputs(self):
+            return [FakeOutput() for _ in range(9)]
+
+    monkeypatch.setattr(model_zoo, "PickableInferenceSession", FakeSession)
+
+    model = model_zoo.ModelRouter("fake_detection.onnx").get_model()
+
+    assert isinstance(model, SCRFD)
+
+
+def test_gui_auto_detection_size_passes_multi_scale_to_detector():
+    engine = FaceEngine(det_size=(0, 0))
+
+    assert engine._detector_input_size() == [(128, 128), (640, 640)]
+
+
+def test_face_analysis_prepare_defaults_to_auto_det_size():
+    class FakeDetectionModel:
+        taskname = "detection"
+
+        def __init__(self):
+            self.kwargs = None
+
+        def prepare(self, ctx_id, **kwargs):
+            self.kwargs = kwargs
+
+    detector = FakeDetectionModel()
+    app = FaceAnalysis.__new__(FaceAnalysis)
+    app.models = {"detection": detector}
+
+    app.prepare(ctx_id=0)
+
+    assert app.det_size == DEFAULT_DET_SIZES
+    assert detector.kwargs["input_size"] == DEFAULT_DET_SIZES
diff --git a/python-package/tests/gui/test_enterprise_evaluation.py b/python-package/tests/gui/test_enterprise_evaluation.py
new file mode 100644
index 000000000..3b971ef5f
--- /dev/null
+++ b/python-package/tests/gui/test_enterprise_evaluation.py
@@ -0,0 +1,283 @@
+from pathlib import Path
+
+import numpy as np
+
+from insightface.gui.core.evaluation import (
+    MULTI_FACE_REQUIRE_ONE,
+    MULTI_FACE_SKIP,
+    MULTI_FACE_USE_CENTERED_LARGEST,
+    MULTI_FACE_USE_LARGEST,
+    _select_face_from_faces,
+    _tar_far_from_scores,
+    run_identity_identification_evaluation,
+    run_identity_verification_evaluation,
+    validate_enterprise_dataset,
+)
+from insightface.gui.core.models import FaceRecord
+
+
+class FakeEngine:
+    model_name = "fake"
+    requested_providers = ["CPU"]
+
+    def _record(self, source_path=None, bbox=None):
+        path = str(source_path or "")
+        if "Alice" in path:
+            embedding = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+        elif "Bob" in path:
+            embedding = np.array([0.0, 1.0, 0.0], dtype=np.float32)
+        else:
+            embedding = np.array([0.0, 0.0, 1.0], dtype=np.float32)
+        return FaceRecord(
+            bbox=bbox or [0, 0, 10, 10],
+            kps=None,
+            det_score=1.0,
+            embedding=embedding,
+            normed_embedding=embedding,
+            source_path=path,
+        )
+
+    def detect_faces(self, image, source_path=None):
+        del image
+        path = str(source_path or "")
+        if "noface" in path:
+            return []
+        if "multi" in path:
+            return [
+                self._record(source_path, [0, 0, 10, 10]),
+                self._record(source_path, [0, 0, 20, 20]),
+            ]
+        return [self._record(source_path)]
+
+    def detect_best_face(self, image, source_path=None):
+        faces = self.detect_faces(image, source_path=source_path)
+        return max(faces, key=lambda face: (face.bbox[2] - face.bbox[0]) * (face.bbox[3] - face.bbox[1])) if faces else None
+
+
+def _write_image(path: Path) -> None:
+    from PIL import Image
+
+    path.parent.mkdir(parents=True, exist_ok=True)
+    Image.new("RGB", (16, 16), color=(128, 128, 128)).save(path)
+
+
+def test_identity_verification_auto_split_generates_probe_vs_gallery_pairs(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery.jpg",
+        "0001__Alice/probe.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = run_identity_verification_evaluation(root, FakeEngine(), auto_split=True)
+
+    assert result.scenario.startswith("Enterprise 1:1")
+    assert result.metrics["total_pairs"] == 4
+    assert result.metrics["positive_pairs"] == 2
+    assert result.metrics["negative_pairs"] == 2
+    assert result.metrics["best_accuracy"] == 1.0
+    assert "TAR@FAR=1e-6" not in result.metrics
+    assert "Threshold@FAR=1e-3" not in result.metrics
+
+
+def test_identity_identification_structured_dataset_reports_top1_and_far(tmp_path):
+    root = tmp_path / "dataset_1n"
+    for rel in [
+        "gallery/0001__Alice/enroll_001.jpg",
+        "gallery/0002__Bob/enroll_001.jpg",
+        "probe/0001__Alice/test_001.jpg",
+        "probe/0002__Bob/test_001.jpg",
+        "unknown/unknown_001.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = run_identity_identification_evaluation(root, FakeEngine(), auto_split=False)
+
+    assert result.scenario.startswith("Enterprise 1:N")
+    assert result.metrics["gallery_identities"] == 2
+    assert result.metrics["Top1"] == 1.0
+    assert "TAR@FAR=1e-5" not in result.metrics
+    assert "Threshold@FAR=1e-2" not in result.metrics
+
+
+def test_tar_far_skips_targets_with_zero_false_accept_budget():
+    metrics = _tar_far_from_scores(
+        positive_scores=[0.95, 0.9],
+        positive_total=2,
+        negative_scores=[0.1] * 99 + [0.2],
+        far_targets=[1e-3, 1e-2],
+    )
+
+    assert "TAR@FAR=1e-3" not in metrics
+    assert "False accept budget@1e-3" not in metrics
+    assert metrics["False accept budget@1e-2"] == 1
+    assert "TAR@FAR=1e-2" in metrics
+
+
+def test_enterprise_validation_requires_one_face_by_default(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery.jpg",
+        "0001__Alice/multi_probe.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:1 Verification",
+        True,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_REQUIRE_ONE,
+    )
+
+    assert not result.ok
+    assert any("multiple faces" in row.get("error", "") for row in result.errors)
+    assert result.summary["images_checked"] == 2
+    assert result.summary["validation_scope"] == "all images"
+
+
+def test_enterprise_validation_can_use_largest_face(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery.jpg",
+        "0001__Alice/multi_probe.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:1 Verification",
+        True,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_USE_LARGEST,
+    )
+
+    assert result.ok
+    assert result.summary["comparison_pairs"] == 4
+    assert result.summary["images_checked"] == 2
+    assert result.summary["validation_scope"] == "gallery or representative images only"
+    assert result.warnings == []
+
+
+def test_centered_largest_face_policy_uses_area_center_penalty():
+    far_large = FaceRecord(
+        bbox=[0, 0, 60, 60],
+        kps=None,
+        det_score=1.0,
+        embedding=np.array([1.0], dtype=np.float32),
+        normed_embedding=np.array([1.0], dtype=np.float32),
+    )
+    centered_medium = FaceRecord(
+        bbox=[40, 40, 90, 90],
+        kps=None,
+        det_score=1.0,
+        embedding=np.array([2.0], dtype=np.float32),
+        normed_embedding=np.array([2.0], dtype=np.float32),
+    )
+
+    selected = _select_face_from_faces(
+        [far_large, centered_medium],
+        (128, 128, 3),
+        Path("multi.jpg"),
+        MULTI_FACE_USE_CENTERED_LARGEST,
+        "test",
+    )
+
+    assert selected is centered_medium
+
+
+def test_enterprise_validation_can_use_centered_largest_face(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery.jpg",
+        "0001__Alice/multi_probe.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:1 Verification",
+        True,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_USE_CENTERED_LARGEST,
+    )
+
+    assert result.ok
+    assert result.summary["comparison_pairs"] == 4
+    assert result.summary["images_checked"] == 2
+    assert result.summary["validation_scope"] == "gallery or representative images only"
+    assert result.warnings == []
+
+
+def test_enterprise_validation_skip_multi_face_gallery_blocks_structured_1n(tmp_path):
+    root = tmp_path / "dataset_1n"
+    for rel in [
+        "gallery/0001__Alice/multi_enroll.jpg",
+        "probe/0001__Alice/test_001.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:N Identification",
+        False,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_SKIP,
+    )
+
+    assert not result.ok
+    assert any("gallery" in row.get("role", "") and row.get("error") for row in result.errors)
+
+
+def test_enterprise_validation_fast_policy_fails_on_no_face_gallery(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery_noface.jpg",
+        "0001__Alice/probe.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:1 Verification",
+        True,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_USE_CENTERED_LARGEST,
+    )
+
+    assert not result.ok
+    assert result.summary["images_checked"] == 1
+    assert any("no face detected" in row.get("error", "") for row in result.errors)
+
+
+def test_enterprise_validation_fast_policy_does_not_inspect_probe_faces(tmp_path):
+    root = tmp_path / "dataset"
+    for rel in [
+        "0001__Alice/gallery.jpg",
+        "0001__Alice/probe_noface.jpg",
+        "0002__Bob/gallery.jpg",
+        "0002__Bob/probe.jpg",
+    ]:
+        _write_image(root / rel)
+
+    result = validate_enterprise_dataset(
+        root,
+        "1:1 Verification",
+        True,
+        FakeEngine(),
+        multi_face_policy=MULTI_FACE_USE_LARGEST,
+    )
+
+    assert result.ok
+    assert result.summary["images_checked"] == 2
+    assert result.summary["comparison_pairs"] == 4
diff --git a/python-package/tests/gui/test_gui_smoke.py b/python-package/tests/gui/test_gui_smoke.py
new file mode 100644
index 000000000..eb625ca5d
--- /dev/null
+++ b/python-package/tests/gui/test_gui_smoke.py
@@ -0,0 +1,279 @@
+import pytest
+import os
+from pathlib import Path
+import numpy as np
+
+
+def test_main_window_smoke(tmp_path):
+    os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+    PySide6 = pytest.importorskip("PySide6")
+    from PySide6.QtCore import QEvent, QUrl, Qt
+    from PySide6.QtWidgets import QAbstractItemView, QApplication, QLabel, QPushButton
+
+    from insightface.gui.app import StudioContext, configure_qt_plugin_paths
+    from insightface.gui.core.config import AppConfig
+    from insightface.gui.core.face_engine import FaceEngine
+    from insightface.gui.core.navigation import AppMode
+    from insightface.gui.core.storage import Storage
+    from insightface.gui.dialogs.license_dialog import LicenseDialog
+    from insightface.gui.dialogs.model_manager_dialog import ModelManagerDialog
+    from insightface.gui.dialogs.settings_dialog import SettingsDialog
+    from insightface.gui.main_window import MainWindow
+    from insightface.gui.resources import APP_ID, APP_PROCESS_NAME, app_icon_path, configure_application_metadata
+
+    configure_qt_plugin_paths()
+    app = QApplication.instance() or QApplication([])
+    configure_application_metadata(app)
+    assert app.applicationName() == "InsightFace Evaluation Studio"
+    assert app.organizationName() == "InsightFace"
+    assert app.organizationDomain() == "insightface.ai"
+    assert app.applicationVersion() == "1.0.1"
+    assert APP_ID == "ai.insightface.evaluationstudio"
+    assert APP_PROCESS_NAME == "InsightFace Evaluation Studio"
+    assert app_icon_path().exists()
+    assert not app.windowIcon().isNull()
+    cfg = AppConfig(
+        workspace_path=str(tmp_path),
+        model_root=str(tmp_path / "model-root"),
+        auto_load_model=False,
+        safe_mode=True,
+        recognition_threshold=0.61,
+        ui_language="en",
+    )
+    storage = Storage(cfg.database_path)
+    engine = FaceEngine(model_name=cfg.model_name)
+    window = MainWindow(StudioContext(cfg, True, storage, engine, str(tmp_path / "app.log")))
+    window.show()
+    assert window.mode_rail.isVisible()
+    assert window.mode_rail.width() >= 220
+    assert window.mode_list.count() == len(AppMode)
+    assert window.mode_list.currentItem().data(Qt.UserRole) == AppMode.FACE_VERIFICATION.value
+    enterprise_help_button = window.findChild(QPushButton, "enterpriseHelpButton")
+    assert enterprise_help_button is not None
+    assert enterprise_help_button.text() == "Enterprise Help"
+    assert "contact page" in enterprise_help_button.toolTip()
+    local_notice = window.findChild(QLabel, "localProcessingNotice")
+    assert local_notice is not None
+    assert "All processing is local" in local_notice.text()
+    assert "No images, embeddings, or reports are uploaded automatically" in local_notice.text()
+    window.open_page("verification")
+    verification_page = window.page_registry.get("verification")
+    assert abs(window.context.config.recognition_threshold - 0.4) < 1e-9
+    assert abs(verification_page.threshold.value() - 0.4) < 1e-9
+    notice = verification_page.findChild(QLabel, "noticeLabel")
+    assert notice is not None
+    assert "Gallery face embeddings are cached in memory" in notice.text()
+    assert "All processing is local by default" not in notice.text()
+    assert verification_page.multi_face_policy.objectName() == "verificationMultiFacePolicy"
+    assert verification_page.multi_face_policy.currentText() == "Use largest centered face"
+    assert "largest centered face" in notice.text()
+    assert hasattr(verification_page.result_table, "_proportional_table_sizer")
+    assert verification_page.result_table.selectionBehavior() == QAbstractItemView.SelectRows
+    assert verification_page.result_table.selectionMode() == QAbstractItemView.SingleSelection
+    verification_page._gallery_embedding_cache_key = ("old.jpg",)
+    verification_page._gallery_embedding_cache = [{"path": "old.jpg"}]
+    verification_page.set_gallery_paths(["new.jpg"])
+    assert verification_page._gallery_embedding_cache_key is None
+    assert verification_page._gallery_embedding_cache is None
+    verification_page._gallery_embedding_cache_key = ("use_largest_face", "new.jpg")
+    verification_page._gallery_embedding_cache = [{"path": "new.jpg"}]
+    verification_page.results = [{"path": "new.jpg"}]
+    verification_page.result_table.setRowCount(1)
+    verification_page.multi_face_policy.setCurrentText("Mark as skip")
+    assert verification_page._gallery_embedding_cache_key is None
+    assert verification_page._gallery_embedding_cache is None
+    assert verification_page.result_table.rowCount() == 0
+    assert "multi-face query stops" in notice.text()
+    label_texts = [label.text() for label in verification_page.findChildren(QLabel)]
+    assert "Mode: waiting for gallery" not in label_texts
+    for button in verification_page.findChildren(QPushButton):
+        assert button.toolTip()
+    face_swap_page = window.page_registry.get("image_face_swap")
+    assert face_swap_page.output_view.objectName() == "imageViewer"
+    assert face_swap_page.output_view.viewport().objectName() == "imageViewerViewport"
+    album_page = window.page_registry.get("album")
+    assert abs(album_page.cluster_threshold.value() - 0.48) < 1e-9
+    album_threshold_help = album_page.findChild(QLabel, "albumThresholdHelp")
+    assert album_threshold_help is not None
+    assert "cosine distance = 1 - cosine threshold" in album_threshold_help.text()
+    assert not hasattr(album_page, "match_threshold")
+    assert album_page.min_samples.value() == 2
+    assert album_page.min_face_size.value() == 80
+    assert album_page.algorithm_label.text().startswith("Algorithm: DBSCAN")
+    assert hasattr(album_page, "import_button")
+    assert hasattr(album_page, "rebuild_button")
+    assert album_page.cluster_table.columnCount() == 2
+    assert album_page.cluster_table.horizontalHeaderItem(0).text() == "Thumbnail"
+    assert album_page.cluster_table.horizontalHeaderItem(1).text() == "Photos"
+    assert album_page.cluster_table.selectionBehavior() == QAbstractItemView.SelectRows
+    assert album_page.photo_table.selectionBehavior() == QAbstractItemView.SelectRows
+    from insightface.gui.core.utils import encode_webp_thumbnail
+
+    thumb = encode_webp_thumbnail(np.zeros((40, 80, 3), dtype=np.uint8), max_side=80, quality=70)
+    assert thumb is not None
+    assert album_page._icon_from_bytes(thumb, album_page.cluster_table.iconSize()) is not None
+    album_page.clusters = [{"id": 1, "photo_count": 2, "face_count": 3, "name": "Album Person 1", "thumbnail_face_id": 7}]
+    album_page.cluster_items = {1: [{"id": 7, "thumbnail": thumb, "media_path": str(tmp_path / "album.jpg")}]}
+    album_page._populate_clusters()
+    assert album_page.cluster_table.item(0, 0).textAlignment() == Qt.AlignCenter
+    assert album_page.cluster_table.item(0, 1).textAlignment() == Qt.AlignCenter
+    enterprise_page = window.page_registry.get("enterprise_evaluation")
+    enterprise_cards = enterprise_page.findChildren(PySide6.QtWidgets.QFrame, "enterpriseCard")
+    assert len(enterprise_cards) >= 3
+    assert enterprise_page.dataset_root.acceptDrops()
+    assert enterprise_page.dataset_drop_card.acceptDrops()
+    assert not enterprise_page.dataset_root.select_button.isHidden()
+    assert enterprise_page.dataset_root.remove_button.isHidden()
+    assert not hasattr(enterprise_page, "dataset_path_status")
+    class _Mime:
+        def hasUrls(self):
+            return True
+
+        def urls(self):
+            return [QUrl.fromLocalFile(str(tmp_path))]
+
+    class _Event:
+        def mimeData(self):
+            return _Mime()
+
+    assert enterprise_page._dataset_drop_path(_Event()) == str(tmp_path)
+    class _DragEvent:
+        def __init__(self, event_type):
+            self._event_type = event_type
+            self.accepted = False
+            self.ignored = False
+
+        def type(self):
+            return self._event_type
+
+        def mimeData(self):
+            return _Mime()
+
+        def acceptProposedAction(self):
+            self.accepted = True
+
+        def ignore(self):
+            self.ignored = True
+
+    drag_event = _DragEvent(QEvent.DragEnter)
+    assert enterprise_page.dataset_root.eventFilter(enterprise_page.dataset_root.path_label, drag_event)
+    assert drag_event.accepted
+    drop_event = _DragEvent(QEvent.Drop)
+    assert enterprise_page.dataset_root.eventFilter(enterprise_page.dataset_root.path_label, drop_event)
+    assert drop_event.accepted
+    assert enterprise_page.dataset_root.path() == str(tmp_path)
+    assert enterprise_page.dataset_root.path_label.text() == f"Selected folder:\n{tmp_path}"
+    assert enterprise_page.dataset_root.path_label.toolTip() == str(tmp_path)
+    assert enterprise_page.dataset_root.select_button.isHidden()
+    assert not enterprise_page.dataset_root.remove_button.isHidden()
+    enterprise_page.dataset_root.clear()
+    assert enterprise_page.dataset_root.path() == ""
+    assert not enterprise_page.dataset_root.select_button.isHidden()
+    assert enterprise_page.dataset_root.remove_button.isHidden()
+    card_drop_event = _DragEvent(QEvent.Drop)
+    assert enterprise_page.eventFilter(enterprise_page.dataset_drop_card, card_drop_event)
+    assert card_drop_event.accepted
+    assert enterprise_page.dataset_root.path_label.text() == f"Selected folder:\n{tmp_path}"
+    assert hasattr(enterprise_page, "help_button")
+    assert enterprise_page.help_button.text() == "Dataset Rules"
+    assert not hasattr(enterprise_page, "export_pdf_button")
+    assert "Export PDF" not in [button.text() for button in enterprise_page.findChildren(QPushButton)]
+    assert "Open Dataset Rules" in enterprise_page.mode_summary.text()
+    assert "Require exactly one face" in enterprise_page.mode_summary.text()
+    enterprise_page.auto_split.setChecked(False)
+    assert "Auto Split is off" in enterprise_page.mode_summary.text()
+    enterprise_page.eval_mode.setCurrentText("1:N Identification")
+    assert "gallery/<identity>" in enterprise_page.mode_summary.text()
+    enterprise_page.multi_face_policy.setCurrentText("Use largest centered face")
+    assert "Use largest centered face" in enterprise_page.mode_summary.text()
+    from insightface.gui.pages.enterprise_eval_page import DatasetRulesDialog
+
+    rules_dialog = DatasetRulesDialog(window)
+    assert rules_dialog.windowTitle() == "Evaluation Dataset Rules"
+    rules_dialog.close()
+    assert enterprise_page.output.minimumHeight() >= 260
+    for mode in AppMode:
+        window.change_mode(mode)
+        assert window.mode_rail.isVisible()
+        assert local_notice.isVisible()
+        assert window.mode_list.currentItem().data(Qt.UserRole) == mode.value
+        assert window.sidebar_list.count() > 0
+        assert not window.sidebar.isVisible()
+        sidebar_titles = [window.sidebar_list.item(i).text() for i in range(window.sidebar_list.count())]
+        assert "Settings" not in sidebar_titles
+        assert "Model Settings" not in sidebar_titles
+        assert "Model Downloads" not in sidebar_titles
+        assert "License Center" not in sidebar_titles
+    settings_dialog = SettingsDialog(window.context, window)
+    assert hasattr(settings_dialog, "theme")
+    assert settings_dialog.theme.count() >= 7
+    assert window.context.config.ui_theme == "azure_lab"
+    assert settings_dialog.theme.currentText() == "Azure Lab"
+    assert hasattr(settings_dialog, "language")
+    assert settings_dialog.language.findData("system") >= 0
+    assert settings_dialog.language.findData("zh") >= 0
+    assert not hasattr(settings_dialog, "workspace")
+    assert not hasattr(settings_dialog, "default_mode")
+    model_dialog = ModelManagerDialog(window.context, window)
+    assert model_dialog.minimumWidth() >= 1120
+    assert model_dialog.downloads_page.table.minimumHeight() >= 400
+    assert hasattr(model_dialog, "run_task")
+    assert model_dialog.downloads_page.table.selectionBehavior() == QAbstractItemView.SelectRows
+    assert model_dialog.downloads_page.table.selectionMode() == QAbstractItemView.SingleSelection
+    assert not hasattr(model_dialog.runtime_page, "threshold")
+    assert not hasattr(model_dialog.runtime_page, "workers")
+    assert not hasattr(model_dialog.runtime_page, "frame_interval")
+    assert hasattr(model_dialog.runtime_page, "gfpgan_enabled")
+    assert hasattr(model_dialog.runtime_page, "gfpgan_model_combo")
+    assert not model_dialog.runtime_page.gfpgan_enabled.isEnabled()
+    gfpgan_path = Path(cfg.model_root) / "models" / "GFPGANv1.4" / "GFPGANv1.4.onnx"
+    gfpgan_path.parent.mkdir(parents=True)
+    gfpgan_path.write_bytes(b"fake")
+    model_dialog.runtime_page.refresh()
+    assert model_dialog.runtime_page.gfpgan_enabled.isEnabled()
+    license_dialog = LicenseDialog(window.context, window)
+    license_buttons = [button.text() for button in license_dialog.findChildren(QPushButton)]
+    assert license_buttons == ["Visit Homepage", "Contact Enterprise Support"]
+    assert license_dialog.height() >= 660
+    assert license_dialog.page.table.minimumHeight() >= 260
+    assert license_dialog.page.findChild(QLabel, "noticeLabel") is None
+    from insightface.gui.pages.license_center_page import ENTERPRISE_HELP_URL, HOMEPAGE_URL
+
+    assert HOMEPAGE_URL == "https://www.insightface.ai"
+    assert ENTERPRISE_HELP_URL == "https://www.insightface.ai/contact"
+    dialogs = [settings_dialog, model_dialog, license_dialog]
+    for dialog in dialogs:
+        dialog.close()
+    assert window.windowTitle().startswith("InsightFace Evaluation Studio")
+    window.close()
+
+
+def test_album_clustering_does_not_match_people_library(tmp_path):
+    os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+    pytest.importorskip("PySide6")
+    from PySide6.QtWidgets import QApplication
+
+    from insightface.gui.app import StudioContext, configure_qt_plugin_paths
+    from insightface.gui.core.config import AppConfig
+    from insightface.gui.core.face_engine import FaceEngine
+    from insightface.gui.core.storage import Storage
+    from insightface.gui.pages.album_page import AlbumPage
+
+    configure_qt_plugin_paths()
+    QApplication.instance() or QApplication([])
+    cfg = AppConfig(workspace_path=str(tmp_path), auto_load_model=False, safe_mode=True)
+    storage = Storage(cfg.database_path)
+    existing_person_id = storage.add_person("Existing Person")
+    storage.add_person("Second Existing Person")
+    storage.add_face_sample(existing_person_id, np.array([1.0, 0.0], dtype=np.float32))
+    media_id = storage.add_media_item(str(tmp_path / "album.jpg"), "image")
+    face_id = storage.add_media_face(media_id, np.array([1.0, 0.0], dtype=np.float32))
+    face = storage.list_media_faces()[0]
+
+    page = AlbumPage(StudioContext(cfg, True, storage, FaceEngine(), str(tmp_path / "app.log")))
+    clusters, _algorithm = page._cluster_faces([face], cosine_threshold=0.48, min_samples=1)
+
+    assert clusters[0]["source"] == "album"
+    assert clusters[0]["id"] == 1
+    assert clusters[0]["name"] == "Album Person 1"
+    assert page.cluster_items[clusters[0]["id"]][0]["id"] == face_id
diff --git a/python-package/tests/gui/test_i18n.py b/python-package/tests/gui/test_i18n.py
new file mode 100644
index 000000000..3bf017e1a
--- /dev/null
+++ b/python-package/tests/gui/test_i18n.py
@@ -0,0 +1,180 @@
+import os
+
+import pytest
+
+os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+
+from insightface.gui.core.i18n import (
+    LANGUAGE_OPTIONS,
+    effective_language,
+    normalize_language,
+    tr,
+)
+
+
+def test_supported_languages_match_homepage_language_set():
+    values = {option.value for option in LANGUAGE_OPTIONS}
+
+    assert {
+        "system",
+        "en",
+        "zh",
+        "ja",
+        "ko",
+        "es",
+        "fr",
+        "de",
+        "pt",
+        "ru",
+    }.issubset(values)
+
+
+def test_language_normalization_and_fallback():
+    assert normalize_language("zh_CN") == "zh"
+    assert normalize_language("pt-BR") == "pt"
+    assert normalize_language("auto") == "system"
+    assert normalize_language("xx") == "en"
+    assert effective_language("en") == "en"
+
+
+def test_core_business_translations_have_professional_terms():
+    assert tr("Settings", "zh") == "设置"
+    assert tr("Enterprise Evaluation", "ja") == "エンタープライズ評価"
+    assert tr("Contact Enterprise Support", "ko") == "기업 지원 문의"
+    assert tr("Commercial production", "de") == "Kommerzieller Produktivbetrieb"
+    assert tr("requires commercial model license", "fr") == "requiert une licence commerciale du modèle"
+    assert tr("Face Recognition", "es") == "Reconocimiento facial"
+    assert tr("Run Evaluation", "pt") == "Executar avaliação"
+    assert tr("License Center", "ru") == "Центр лицензий"
+
+
+def test_apply_translations_updates_basic_widgets():
+    pytest.importorskip("PySide6")
+    from PySide6.QtWidgets import QApplication, QLabel, QPushButton, QVBoxLayout, QWidget
+
+    from insightface.gui.app import configure_qt_plugin_paths
+    from insightface.gui.core.i18n import apply_translations
+
+    configure_qt_plugin_paths()
+    QApplication.instance() or QApplication([])
+    widget = QWidget()
+    layout = QVBoxLayout(widget)
+    label = QLabel("Settings")
+    button = QPushButton("Run Evaluation")
+    button.setToolTip("Choose how evaluation handles images where more than one face is detected.")
+    layout.addWidget(label)
+    layout.addWidget(button)
+
+    apply_translations(widget, "zh")
+
+    assert label.text() == "设置"
+    assert button.text() == "运行评测"
+    assert button.toolTip() == "选择评测中检测到多张人脸时的处理方式。"
+
+    apply_translations(widget, "en")
+
+    assert label.text() == "Settings"
+    assert button.text() == "Run Evaluation"
+    assert button.toolTip() == "Choose how evaluation handles images where more than one face is detected."
+
+
+def test_apply_translations_localizes_generic_button_tooltip():
+    pytest.importorskip("PySide6")
+    from PySide6.QtWidgets import QApplication, QPushButton
+
+    from insightface.gui.app import configure_qt_plugin_paths
+    from insightface.gui.core.i18n import apply_translations
+
+    configure_qt_plugin_paths()
+    QApplication.instance() or QApplication([])
+    button = QPushButton("Run Face Swap")
+    button.setToolTip("Run face swap with the configured local swap model.")
+
+    apply_translations(button, "zh")
+
+    assert button.text() == "运行换脸"
+    assert button.toolTip() == "点击执行：运行换脸。"
+
+    apply_translations(button, "en")
+
+    assert button.text() == "Run Face Swap"
+    assert button.toolTip() == "Run face swap with the configured local swap model."
+
+
+def test_icon_only_button_keeps_symbol_when_tooltip_is_localized():
+    pytest.importorskip("PySide6")
+    from PySide6.QtWidgets import QApplication, QPushButton
+
+    from insightface.gui.app import configure_qt_plugin_paths
+    from insightface.gui.core.i18n import apply_translations
+
+    configure_qt_plugin_paths()
+    QApplication.instance() or QApplication([])
+    button = QPushButton("×")
+    button.setToolTip("Remove the current file.")
+
+    apply_translations(button, "zh")
+
+    assert button.text() == "×"
+    assert button.toolTip() == "点击执行：移除。"
+
+
+def test_dataset_rules_dialog_has_localized_help_summary():
+    from insightface.gui.pages.enterprise_eval_page import dataset_rules_text
+
+    text = dataset_rules_text("zh")
+
+    assert text.startswith("评测数据集规则")
+    assert "支持从身份文件夹进行本地 1:1 验证和 1:N 识别评测" in text
+    assert "dataset_1v1/" in text
+    assert "dataset_1n/" in text
+    assert "gallery/" in text
+    assert "probe/" in text
+    assert "多人脸处理" in text
+    assert "报告输出" in text
+    assert "Enterprise Evaluation Dataset Rules" not in text
+    assert "Each subfolder is one identity" not in text
+
+
+def test_album_and_recognition_page_copy_is_localized():
+    assert (
+        tr(
+            "Import or refresh local album folders, cluster detected faces, and review the photos in each person group.",
+            "zh",
+        )
+        == "导入或刷新本地相册文件夹，对检测到的人脸进行聚类，并查看每个人物组中的照片。"
+    )
+    assert "所有相册处理均在本地完成" in tr(
+        "All album processing is local. Import / Refresh scans every selected folder, extracts features only for new images, then runs DBSCAN clustering over all indexed faces using the selected cosine threshold.",
+        "zh",
+    )
+    assert "上传一张查询图片" in tr(
+        "Upload one query image and a gallery image, image set, or folder. One gallery image runs 1:1 compare; multiple gallery images run 1:N gallery search.",
+        "zh",
+    )
+    assert "源图 + 目标 = 结果" in tr(
+        "Source + Target = Result. Target can be an image or a video; the workflow chooses image or video swap automatically.",
+        "zh",
+    )
+    assert "商业授权" in tr(
+        "Face swap may require separate commercial authorization depending on usage and model license. Use only with appropriate rights and consent.",
+        "zh",
+    )
+    assert "采购决策" in tr(
+        "Run local 1:1 verification or 1:N identification evaluation from identity folders and export procurement-ready reports.",
+        "zh",
+    )
+    assert "运行后端" in tr(
+        "Configure model packs, execution provider, face swap models, and runtime checks.",
+        "zh",
+    )
+    assert "手动刷新 GitHub Release" in tr(
+        "Manually refresh GitHub release model URLs and download selected model packages locally.",
+        "zh",
+    )
+    assert "商业部署" in tr(
+        "Code and model files may have different licenses. Commercial deployment requires appropriate model authorization.",
+        "zh",
+    )
+    assert tr("Refresh source", "zh") == "刷新来源"
+    assert tr("Local model root", "zh") == "本地模型根目录"
diff --git a/python-package/tests/gui/test_links.py b/python-package/tests/gui/test_links.py
new file mode 100644
index 000000000..010aaf65d
--- /dev/null
+++ b/python-package/tests/gui/test_links.py
@@ -0,0 +1,24 @@
+import pytest
+
+
+def test_insightface_links_add_gui_referrer():
+    pytest.importorskip("PySide6")
+
+    from insightface.gui.core.links import add_gui_referrer
+
+    url = add_gui_referrer("https://www.insightface.ai/contact", content="license_enterprise_support")
+
+    assert url.startswith("https://www.insightface.ai/contact?")
+    assert "utm_source=insightface_gui" in url
+    assert "utm_medium=" not in url
+    assert "utm_campaign=" not in url
+    assert "utm_content=" not in url
+    assert "ref=" not in url
+
+
+def test_non_insightface_links_are_not_modified():
+    pytest.importorskip("PySide6")
+
+    from insightface.gui.core.links import add_gui_referrer
+
+    assert add_gui_referrer("https://github.com/deepinsight/insightface") == "https://github.com/deepinsight/insightface"
diff --git a/python-package/tests/gui/test_model_downloads.py b/python-package/tests/gui/test_model_downloads.py
new file mode 100644
index 000000000..7c1c584c4
--- /dev/null
+++ b/python-package/tests/gui/test_model_downloads.py
@@ -0,0 +1,42 @@
+from insightface.gui.core.model_downloads import GFPGAN_DOWNLOAD_URL, _content_range_total, fallback_model_assets, local_model_status
+from insightface.gui.core.paths import default_workspace, workspace_paths
+
+
+def test_default_gui_workspace_path():
+    workspace = default_workspace()
+    assert workspace.name == "gui"
+    assert workspace.parent.name == ".insightface"
+    assert workspace.is_absolute()
+
+
+def test_workspace_paths_are_under_gui_workspace(tmp_path):
+    workspace = (tmp_path / ".insightface" / "gui").resolve()
+    paths = workspace_paths(workspace)
+    assert paths["workspace"] == workspace
+    for key in ("database", "crops", "exports", "reports", "logs", "cache"):
+        assert paths[key].is_relative_to(paths["workspace"])
+
+
+def test_fallback_model_assets_have_github_release_urls(tmp_path):
+    assets = fallback_model_assets()
+    names = {asset.name for asset in assets}
+    assert {"buffalo_l.zip", "buffalo_s.zip", "antelopev2.zip", "GFPGANv1.4.onnx"}.issubset(names)
+    for asset in assets:
+        if asset.name == "GFPGANv1.4.onnx":
+            assert asset.source == "third party"
+            assert asset.kind == "third-party restore model"
+            assert asset.browser_download_url == GFPGAN_DOWNLOAD_URL
+            assert "harisreedhar/Face-Upscalers-ONNX/releases/download/Models" in asset.browser_download_url
+        else:
+            assert asset.tag_name == "v0.7"
+            assert asset.browser_download_url.startswith(
+                "https://github.com/deepinsight/insightface/releases/download/v0.7/"
+            )
+    assert local_model_status(assets[0], tmp_path) == "not installed"
+
+
+def test_content_range_total_parser():
+    assert _content_range_total("bytes 10-19/100") == 100
+    assert _content_range_total("bytes 10-19/*") == 0
+    assert _content_range_total(None) == 0
+    assert _content_range_total("not-a-range") == 0
diff --git a/python-package/tests/gui/test_navigation.py b/python-package/tests/gui/test_navigation.py
new file mode 100644
index 000000000..929a2ff1f
--- /dev/null
+++ b/python-package/tests/gui/test_navigation.py
@@ -0,0 +1,98 @@
+import os
+
+from insightface.gui.core.config import AppConfig, load_config, save_config
+from insightface.gui.core.navigation import AppMode, GLOBAL_PAGE_TITLES, NAVIGATION_MODES
+
+
+def test_navigation_modes_do_not_include_global_sidebar_items():
+    assert set(NAVIGATION_MODES) == set(AppMode)
+    all_page_keys = []
+    for mode in NAVIGATION_MODES.values():
+        assert mode.items
+        assert len(mode.items) <= 8
+        for item in mode.items:
+            assert item.title not in GLOBAL_PAGE_TITLES
+            all_page_keys.append(item.page_key)
+    assert len(all_page_keys) >= len(set(all_page_keys))
+
+
+def test_face_verification_mode_is_focused():
+    titles = [item.title for item in NAVIGATION_MODES[AppMode.FACE_VERIFICATION].items]
+    keys = [item.page_key for item in NAVIGATION_MODES[AppMode.FACE_VERIFICATION].items]
+
+    assert NAVIGATION_MODES[AppMode.FACE_VERIFICATION].title == "Face Recognition"
+    assert titles == ["Face Recognition"]
+    assert keys == ["verification"]
+
+
+def test_face_swap_mode_is_single_workspace():
+    items = NAVIGATION_MODES[AppMode.FACE_SWAP].items
+
+    assert len(items) == 1
+    assert items[0].title == "Face Swap"
+    assert items[0].page_key == "image_face_swap"
+
+
+def test_album_mode_is_single_workspace():
+    items = NAVIGATION_MODES[AppMode.ALBUM_MANAGEMENT].items
+
+    assert len(items) == 1
+    assert items[0].title == "Album"
+    assert items[0].page_key == "album"
+
+
+def test_enterprise_mode_is_single_evaluation_workspace():
+    items = NAVIGATION_MODES[AppMode.ENTERPRISE_EVALUATION].items
+
+    assert len(items) == 1
+    assert items[0].title == "Enterprise Evaluation"
+    assert items[0].page_key == "enterprise_evaluation"
+
+
+def test_mode_persistence(tmp_path):
+    cfg = AppConfig(workspace_path=str(tmp_path))
+    cfg.ui_last_mode = AppMode.FACE_SWAP.value
+    cfg.ui_last_page_face_swap = "image_face_swap"
+    save_config(cfg)
+
+    loaded, exists = load_config(tmp_path / "config.json")
+
+    assert exists is True
+    assert loaded.ui_last_mode == AppMode.FACE_SWAP.value
+    assert loaded.ui_last_page_face_swap == "image_face_swap"
+
+
+def test_default_detection_size_is_auto(tmp_path):
+    cfg = AppConfig(workspace_path=str(tmp_path))
+
+    assert cfg.det_size == [0, 0]
+    assert cfg.det_size_label == "Auto"
+    assert cfg.recognition_threshold == 0.4
+
+
+def test_threshold_slider_default_is_028():
+    import pytest
+
+    os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
+    pytest.importorskip("PySide6")
+    from PySide6.QtWidgets import QApplication
+    from insightface.gui.app import configure_qt_plugin_paths
+    from insightface.gui.widgets.threshold_slider import ThresholdSlider
+
+    configure_qt_plugin_paths()
+    _app = QApplication.instance() or QApplication([])
+    slider = ThresholdSlider()
+    assert slider.value() == 0.4
+
+
+def test_qt_plugin_paths_are_configured(monkeypatch):
+    from pathlib import Path
+
+    from insightface.gui.app import configure_qt_plugin_paths
+
+    monkeypatch.delenv("QT_PLUGIN_PATH", raising=False)
+    monkeypatch.delenv("QT_QPA_PLATFORM_PLUGIN_PATH", raising=False)
+    configure_qt_plugin_paths()
+
+    assert Path(os.environ["QT_PLUGIN_PATH"]).exists()
+    assert Path(os.environ["QT_QPA_PLATFORM_PLUGIN_PATH"]).exists()
diff --git a/python-package/tests/gui/test_providers.py b/python-package/tests/gui/test_providers.py
new file mode 100644
index 000000000..318b46821
--- /dev/null
+++ b/python-package/tests/gui/test_providers.py
@@ -0,0 +1,21 @@
+from insightface.gui.core import face_engine
+
+
+def test_cuda_choice_falls_back_when_provider_is_unavailable(monkeypatch):
+    monkeypatch.setattr(face_engine, "available_execution_providers", lambda: ["CPUExecutionProvider"])
+
+    assert face_engine.is_cuda_provider_available() is False
+    assert face_engine.providers_from_choice("CUDA") == ["CPUExecutionProvider"]
+
+
+def test_cuda_choice_is_enabled_when_provider_is_available(monkeypatch):
+    monkeypatch.setattr(
+        face_engine,
+        "available_execution_providers",
+        lambda: ["CUDAExecutionProvider", "CPUExecutionProvider"],
+    )
+
+    assert face_engine.is_cuda_provider_available() is True
+    assert face_engine.providers_from_choice("CUDA") == ["CUDAExecutionProvider", "CPUExecutionProvider"]
+    assert face_engine.providers_from_choice("Auto") == ["CUDAExecutionProvider", "CPUExecutionProvider"]
+
diff --git a/python-package/tests/gui/test_qt_plugins.py b/python-package/tests/gui/test_qt_plugins.py
new file mode 100644
index 000000000..2f496f21b
--- /dev/null
+++ b/python-package/tests/gui/test_qt_plugins.py
@@ -0,0 +1,59 @@
+import os
+import sys
+from pathlib import Path
+
+import pytest
+
+
+def _platform_plugin_names():
+    if sys.platform == "darwin":
+        return {"libqcocoa.dylib", "libqoffscreen.dylib"}
+    if sys.platform.startswith("win"):
+        return {"qwindows.dll", "qoffscreen.dll"}
+    if sys.platform.startswith("linux"):
+        return {"libqxcb.so", "libqoffscreen.so"}
+    return set()
+
+
+def test_qt_platform_plugins_available_for_current_platform(monkeypatch):
+    pytest.importorskip("PySide6")
+    from insightface.gui.app import configure_qt_plugin_paths, qt_plugin_root_candidates
+
+    monkeypatch.delenv("QT_PLUGIN_PATH", raising=False)
+    monkeypatch.delenv("QT_QPA_PLATFORM_PLUGIN_PATH", raising=False)
+    configure_qt_plugin_paths()
+
+    assert os.environ.get("QT_PLUGIN_PATH"), (
+        "QT_PLUGIN_PATH was not configured. "
+        f"Qt plugin root candidates: {qt_plugin_root_candidates()}"
+    )
+    assert os.environ.get("QT_QPA_PLATFORM_PLUGIN_PATH"), (
+        "QT_QPA_PLATFORM_PLUGIN_PATH was not configured. "
+        f"Qt plugin root candidates: {qt_plugin_root_candidates()}"
+    )
+    platform_dir = Path(os.environ["QT_QPA_PLATFORM_PLUGIN_PATH"])
+    expected_plugins = _platform_plugin_names()
+    if not expected_plugins:
+        pytest.skip(f"No expected Qt platform plugin list for {sys.platform}")
+
+    available_plugins = {path.name for path in platform_dir.iterdir()}
+    missing_plugins = expected_plugins - available_plugins
+
+    assert not missing_plugins, (
+        f"Missing Qt platform plugins for {sys.platform}: "
+        f"{sorted(missing_plugins)} in {platform_dir}"
+    )
+
+
+def test_qapplication_can_start_with_configured_offscreen_plugin(monkeypatch):
+    pytest.importorskip("PySide6")
+    from insightface.gui.app import configure_qt_plugin_paths
+
+    monkeypatch.setenv("QT_QPA_PLATFORM", "offscreen")
+    configure_qt_plugin_paths()
+
+    from PySide6.QtWidgets import QApplication
+
+    app = QApplication.instance() or QApplication([])
+
+    assert app.platformName().lower() == "offscreen"
diff --git a/python-package/tests/gui/test_quality.py b/python-package/tests/gui/test_quality.py
new file mode 100644
index 000000000..046267901
--- /dev/null
+++ b/python-package/tests/gui/test_quality.py
@@ -0,0 +1,12 @@
+import numpy as np
+
+from insightface.gui.core.quality import blur_score, score_face
+
+
+def test_quality_range_and_blur():
+    image = np.full((120, 120, 3), 128, dtype=np.uint8)
+    image[40:80, 40:80] = 255
+    assert blur_score(image) >= 0
+    score, flags = score_face(image, [30, 30, 90, 90], det_score=0.9)
+    assert 0.0 <= score <= 1.0
+    assert isinstance(flags, list)
diff --git a/python-package/tests/gui/test_reporting.py b/python-package/tests/gui/test_reporting.py
new file mode 100644
index 000000000..6e281472d
--- /dev/null
+++ b/python-package/tests/gui/test_reporting.py
@@ -0,0 +1,74 @@
+from pathlib import Path
+
+import pytest
+
+from insightface.gui.core.models import EvaluationResult
+from insightface.gui.core.reporting import generate_html_report, generate_markdown_report, write_reports
+
+
+def _result():
+    return EvaluationResult(
+        scenario="KYC / 1:1 Verification",
+        model_name="buffalo_l",
+        provider="CPU",
+        threshold=0.5,
+        dataset_summary={"total_pairs": 2},
+        metrics={"accuracy": 1.0, "TAR@FAR=1e-2": 0.98},
+        errors=[],
+        latency={"average_ms": 1.2},
+        license_status="Research / Non-commercial",
+        created_at="2026-05-19T00:00:00Z",
+        raw_results=[{"image": "a.jpg", "similarity": 0.9}],
+        threshold_recommendation=0.42,
+    )
+
+
+def test_reporting_markdown_html():
+    result = _result()
+    md = generate_markdown_report(result, language="en")
+    html = generate_html_report(result, language="en")
+    assert "InsightFace Enterprise Evaluation Report" in md
+    assert "Commercial Licensing Next Steps" in md
+    assert "report interpretation" in md
+    assert "https://www.insightface.ai/contact" in md
+    assert "| Field | Value |" in md
+    assert "<html" in html
+
+
+def test_reporting_markdown_is_localized():
+    md = generate_markdown_report(_result(), language="zh")
+    assert "InsightFace 企业评测报告" in md
+    assert "执行摘要" in md
+    assert "本报告不构成法律建议" in md
+    assert "报告解读" in md
+    assert "https://www.insightface.ai/contact" in md
+
+
+def test_write_reports_outputs_formatted_pdf(tmp_path):
+    pytest.importorskip("reportlab")
+    paths = write_reports(_result(), tmp_path, language="zh")
+
+    assert paths["markdown"].endswith(".md")
+    assert paths["html"].endswith(".html")
+    assert paths["pdf"].endswith(".pdf")
+    assert Path(paths["markdown"]).exists()
+    assert Path(paths["html"]).exists()
+    assert Path(paths["pdf"]).exists()
+
+
+def test_write_reports_sets_primary_report_path_to_pdf(tmp_path):
+    pytest.importorskip("reportlab")
+    result = EvaluationResult(
+        scenario="KYC / 1:1 Verification",
+        model_name="buffalo_l",
+        provider="CPU",
+        threshold=0.5,
+        dataset_summary={"total_pairs": 2},
+        metrics={"accuracy": 1.0},
+        errors=[],
+        latency={"average_ms": 1.2},
+        license_status="Research / Non-commercial",
+        created_at="2026-05-19T00:00:00Z",
+    )
+    paths = write_reports(result, tmp_path)
+    assert result.report_path == paths["pdf"]
diff --git a/python-package/tests/gui/test_similarity.py b/python-package/tests/gui/test_similarity.py
new file mode 100644
index 000000000..a9193a110
--- /dev/null
+++ b/python-package/tests/gui/test_similarity.py
@@ -0,0 +1,17 @@
+import numpy as np
+
+from insightface.gui.core.recognition import cosine_similarity, normalize_embedding, search_gallery
+
+
+def test_similarity_and_topk():
+    emb = normalize_embedding(np.array([3.0, 4.0], dtype=np.float32))
+    assert np.allclose(np.linalg.norm(emb), 1.0)
+    assert cosine_similarity([1, 0], [1, 0]) == 1.0
+    gallery = [
+        {"person_id": 1, "person_name": "A", "sample_id": 1, "embedding": np.array([1, 0], dtype=np.float32)},
+        {"person_id": 2, "person_name": "B", "sample_id": 2, "embedding": np.array([0, 1], dtype=np.float32)},
+    ]
+    results = search_gallery(np.array([1, 0], dtype=np.float32), gallery, top_k=2, threshold=0.5)
+    assert results[0].person_name == "A"
+    assert results[0].similarity > results[1].similarity
+    assert "quality_score" not in results[0].to_json_dict()
diff --git a/python-package/tests/gui/test_storage.py b/python-package/tests/gui/test_storage.py
new file mode 100644
index 000000000..d9859fed9
--- /dev/null
+++ b/python-package/tests/gui/test_storage.py
@@ -0,0 +1,91 @@
+import numpy as np
+
+from insightface.gui.core.storage import Storage
+from insightface.gui.core.utils import encode_webp_thumbnail
+
+
+def test_storage_people_samples_and_search(tmp_path):
+    db = tmp_path / "test.db"
+    storage = Storage(db)
+    person_id = storage.add_person("Alice")
+    emb = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    sample_id = storage.add_face_sample(person_id, emb, source_image_path="a.jpg", det_score=0.9)
+    people = storage.list_people()
+    assert people[0]["name"] == "Alice"
+    assert people[0]["sample_count"] == 1
+    samples = storage.list_face_samples(person_id)
+    assert samples[0]["id"] == sample_id
+    assert np.allclose(samples[0]["embedding"], emb)
+    results = storage.search_embeddings(np.array([1.0, 0.0, 0.0], dtype=np.float32), top_k=1, threshold=0.5)
+    assert results[0].person_id == person_id
+    assert results[0].status == "matched"
+
+
+def test_album_directories_and_results_persist(tmp_path):
+    db = tmp_path / "test.db"
+    storage = Storage(db)
+    album_dir = tmp_path / "album"
+    album_dir.mkdir()
+    image_path = album_dir / "a.jpg"
+    image_path.write_bytes(b"placeholder")
+    photo_thumb = b"photo-webp"
+    face_thumb = b"face-webp"
+
+    storage.save_album_directories([str(album_dir)])
+    assert storage.list_album_directories() == [str(album_dir)]
+
+    media_id = storage.add_media_item(str(image_path), "image", thumbnail=photo_thumb)
+    face_id = storage.add_media_face(
+        media_id,
+        np.array([1.0, 0.0], dtype=np.float32),
+        thumbnail=face_thumb,
+    )
+    cluster = {
+        "id": 1,
+        "label": 0,
+        "name": "Album Person 1",
+        "source": "album",
+        "face_count": 1,
+        "photo_count": 1,
+        "avg_quality": 0.0,
+        "thumbnail_face_id": face_id,
+        "thumbnail_path": "",
+        "photos": [str(image_path)],
+    }
+    storage.save_album_results(
+        [cluster],
+        {1: [{"id": face_id, "media_path": str(image_path)}]},
+        "DBSCAN",
+        cluster_threshold=0.28,
+        min_samples=2,
+        min_face_size=80,
+    )
+
+    results = storage.load_album_results()
+    assert results["algorithm"] == "DBSCAN"
+    assert results["cluster_threshold"] == 0.28
+    assert "duplicate_threshold" not in results
+    assert results["min_face_size"] == 80
+    assert results["clusters"][0]["thumbnail_face_id"] == face_id
+    assert results["clusters"][0]["face_ids"] == [face_id]
+    face = storage.list_media_faces()[0]
+    assert face["cluster_id"] == 1
+    assert face["thumbnail"] == face_thumb
+    assert face["thumbnail_mime"] == "image/webp"
+    assert face["media_thumbnail"] == photo_thumb
+    assert face["media_thumbnail_mime"] == "image/webp"
+
+    storage.clear_album_results()
+    assert storage.load_album_results() == {}
+    assert storage.list_media_faces()[0]["cluster_id"] is None
+
+
+def test_webp_thumbnail_encoder_outputs_small_webp_bytes():
+    image = np.zeros((80, 160, 3), dtype=np.uint8)
+    image[:, :, 1] = 200
+
+    payload = encode_webp_thumbnail(image, max_side=120, quality=35)
+
+    assert payload is not None
+    assert payload[:4] == b"RIFF"
+    assert b"WEBP" in payload[:16]
diff --git a/python-package/tests/gui/test_swap.py b/python-package/tests/gui/test_swap.py
new file mode 100644
index 000000000..ac5b814d4
--- /dev/null
+++ b/python-package/tests/gui/test_swap.py
@@ -0,0 +1,27 @@
+import numpy as np
+
+from insightface.gui.core.swap import GFPGANRestorer
+
+
+class _FakeSession:
+    def __init__(self):
+        self.input_shape = None
+
+    def run(self, output_names, inputs):
+        del output_names
+        tensor = next(iter(inputs.values()))
+        self.input_shape = tensor.shape
+        return [np.zeros((1, 3, 512, 512), dtype=np.float32)]
+
+
+def test_gfpgan_restorer_uses_512_input_and_restores_original_shape():
+    restorer = GFPGANRestorer()
+    restorer.session = _FakeSession()
+    restorer.input_name = "input"
+    restorer.output_name = "output"
+    image = np.full((96, 128, 3), 127, dtype=np.uint8)
+
+    output = restorer.restore(image)
+
+    assert restorer.session.input_shape == (1, 3, 512, 512)
+    assert output.shape == image.shape
diff --git a/python-package/tests/gui/test_theme.py b/python-package/tests/gui/test_theme.py
new file mode 100644
index 000000000..fa9753eee
--- /dev/null
+++ b/python-package/tests/gui/test_theme.py
@@ -0,0 +1,58 @@
+from insightface.gui.core.theme import (
+    THEME_OPTIONS,
+    application_stylesheet,
+    effective_theme,
+    normalize_theme,
+    theme_description,
+    theme_label,
+)
+
+
+def test_theme_options_include_multiple_product_themes():
+    values = [option.value for option in THEME_OPTIONS]
+
+    assert values[0] == "system"
+    assert len(values) == 7
+    assert len(set(values)) == len(values)
+    assert {
+        "precision_light",
+        "studio_dark",
+        "graphite_pro",
+        "azure_lab",
+        "emerald_focus",
+        "crimson_audit",
+    }.issubset(values)
+
+
+def test_theme_aliases_and_descriptions_are_backward_compatible():
+    assert normalize_theme("light") == "precision_light"
+    assert normalize_theme("dark") == "studio_dark"
+    assert normalize_theme("unknown") == "system"
+    assert theme_label("dark") == "Studio Dark"
+    assert theme_description("azure_lab")
+
+
+def test_application_stylesheet_contains_theme_specific_controls():
+    for value in ["precision_light", "studio_dark", "graphite_pro", "azure_lab", "emerald_focus", "crimson_audit"]:
+        stylesheet = application_stylesheet(value)
+        assert "QFrame#uploadPreview" in stylesheet
+        assert "QGraphicsView#imageViewer" in stylesheet
+        assert "QWidget#imageViewerViewport" in stylesheet
+        assert "QPushButton#removeUpload" in stylesheet
+        assert "QWidget#dashboardCard" in stylesheet
+        assert "QLabel[role=\"statusChip\"]" in stylesheet
+        assert effective_theme(value) == value
+
+
+def test_product_themes_have_distinct_accent_colors():
+    stylesheets = {
+        value: application_stylesheet(value)
+        for value in ["precision_light", "studio_dark", "graphite_pro", "azure_lab", "emerald_focus", "crimson_audit"]
+    }
+
+    assert "#2f3a45" in stylesheets["precision_light"]
+    assert "#7c5cff" in stylesheets["studio_dark"]
+    assert "#f59e0b" in stylesheets["graphite_pro"]
+    assert "#075eea" in stylesheets["azure_lab"]
+    assert "#07844f" in stylesheets["emerald_focus"]
+    assert "#ff3b6b" in stylesheets["crimson_audit"]
diff --git a/recognition/ArcFace/README.md b/recognition/ArcFace/README.md
deleted file mode 100644
index 0fc155539..000000000
--- a/recognition/ArcFace/README.md
+++ /dev/null
@@ -1,95 +0,0 @@
-## ArcFace with Parallel Acceleration on both Features and Centers
-
-### Memory Consumption and Training Speed
-
-![Memoryspeed](https://github.com/deepinsight/insightface/blob/master/resources/memoryspeed.png)
-
-Parallel acceleration on both feature x and centre W. Setting: ResNet 50, batch size 8 * 64, feature dimension 512, float point 32, GPU 8 * P40 (24GB).
-
-### Illustration of Main Steps
-
-![Memoryspeed](https://github.com/deepinsight/insightface/blob/master/resources/mainsteps.png)
-
-Parallel calculation by simple matrix partition. Setting: ResNet 50, batch size 8 * 64, feature dimension 512, float point 32, identity number 1 Million, GPU 8 * 1080ti (11GB). Communication cost: 1MB (feature x). Training speed: 800 samples/second.
-
-**Note:** Replace ``train.py`` with ``train_parall.py`` in following examples if you want to use parallel acceleration.
-
-### Model Training
-
-1. Install `MXNet` with GPU support.
-
-```
-pip install mxnet-cu100 # mxnet-cu102
-```
-
-2. Clone the InsightFace repository. We call the directory insightface as *`INSIGHTFACE_ROOT`*.
-
-```
-git clone --recursive https://github.com/deepinsight/insightface.git
-```
-
-3. Download the training set (`MS1MV2-Arcface`) and place it in *`$INSIGHTFACE_ROOT/recognition/datasets/`*. Each training dataset includes the following 6 files:
-
-```Shell
-    faces_emore/
-       train.idx
-       train.rec
-       property
-       lfw.bin
-       cfp_fp.bin
-       agedb_30.bin
-```
-
-The first three files are the training dataset while the last three files are verification sets.
-
-4. Train deep face recognition models.
-In this part, we assume you are in the directory *`$INSIGHTFACE_ROOT/recognition/ArcFace`*.
-
-Place and edit config file:
-```Shell
-cp sample_config.py config.py
-vim config.py # edit dataset path etc..
-```
-
-We give some examples below. Our experiments were conducted on the Tesla P40 GPU.
-
-(1). Train ArcFace with LResNet100E-IR.
-
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r100 --loss arcface --dataset emore
-```
-
-It will output verification results of *LFW*, *CFP-FP* and *AgeDB-30* every 2000 batches. You can check all options in *config.py*.
-This model can achieve *LFW 99.80+* and *MegaFace 98.3%+*.
-
-(2). Train CosineFace with LResNet50E-IR.
-
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r50 --loss cosface --dataset emore
-```
-
-(3). Train Softmax with MobileFaceNet.
-
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network y1 --loss softmax --dataset emore
-```
-
-(4). Fine-turn the above Softmax model with Triplet loss.
-
-```Shell
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network mnas05 --loss triplet --lr 0.005 --pretrained ./models/y1-softmax-emore,1
-```
-
-### Citation
-
-If you find *ArcFace* useful in your research, please consider to cite the following related papers:
-
-```
-@inproceedings{deng2019arcface,
-  title={Arcface: Additive angular margin loss for deep face recognition},
-  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
-  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
-  pages={4690--4699},
-  year={2019}
-}
-```
diff --git a/recognition/ArcFace/image_iter.py b/recognition/ArcFace/image_iter.py
deleted file mode 100644
index a7ef137a1..000000000
--- a/recognition/ArcFace/image_iter.py
+++ /dev/null
@@ -1,313 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-
-import mxnet as mx
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-
-logger = logging.getLogger()
-
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0, color_jittering = 0,
-                 images_filter = 0,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        if path_imgrec:
-            logging.info('loading recordio %s...',
-                         path_imgrec)
-            path_imgidx = path_imgrec[0:-4]+".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-            s = self.imgrec.read_idx(0)
-            header, _ = recordio.unpack(s)
-            if header.flag>0:
-              print('header0 label', header.label)
-              self.header0 = (int(header.label[0]), int(header.label[1]))
-              #assert(header.flag==1)
-              #self.imgidx = range(1, int(header.label[0]))
-              self.imgidx = []
-              self.id2range = {}
-              self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-              for identity in self.seq_identity:
-                s = self.imgrec.read_idx(identity)
-                header, _ = recordio.unpack(s)
-                a,b = int(header.label[0]), int(header.label[1])
-                count = b-a
-                if count<images_filter:
-                  continue
-                self.id2range[identity] = (a,b)
-                self.imgidx += range(a, b)
-              print('id2range', len(self.id2range))
-            else:
-              self.imgidx = list(self.imgrec.keys)
-            if shuffle:
-              self.seq = self.imgidx
-              self.oseq = self.imgidx
-              print(len(self.seq))
-            else:
-              self.seq = None
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        self.color_jittering = color_jittering
-        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
-        self.provide_label = [(label_name, (batch_size,))]
-        #print(self.provide_label[0][1])
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.shuffle:
-          random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        """Helper function for reading in next sample."""
-        #set total batch size, for example, 1800, and maximum size for each people, for example 45
-        if self.seq is not None:
-          while True:
-            if self.cur >= len(self.seq):
-                raise StopIteration
-            idx = self.seq[self.cur]
-            self.cur += 1
-            if self.imgrec is not None:
-              s = self.imgrec.read_idx(idx)
-              header, img = recordio.unpack(s)
-              label = header.label
-              if not isinstance(label, numbers.Number):
-                label = label[0]
-              return label, img, None, None
-            else:
-              label, fname, bbox, landmark = self.imglist[idx]
-              return label, self.read_image(fname), bbox, landmark
-        else:
-            s = self.imgrec.read()
-            if s is None:
-                raise StopIteration
-            header, img = recordio.unpack(s)
-            return header.label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = nd.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      #random.shuffle(augs)
-      #for aug in augs:
-      #  #print(img.shape)
-      #  img = aug(img, x)
-      #  #print(img.shape)
-      #return img
-      return self.CJA(img)
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in range(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-    def compress_aug(self, img):
-      from PIL import Image
-      from io import BytesIO
-      buf = BytesIO()
-      img = Image.fromarray(img.asnumpy(), 'RGB')
-      q = random.randint(2, 20)
-      img.save(buf, format='JPEG', quality=q)
-      buf = buf.getvalue()
-      img = Image.open(BytesIO(buf))
-      return nd.array(np.asarray(img, 'float32'))
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                _data = self.imdecode(s)
-                if _data.shape[0]!=self.data_shape[1]:
-                  _data = mx.image.resize_short(_data, self.data_shape[1])
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.color_jittering>0:
-                  if self.color_jittering>1:
-                    _rd = random.randint(0,1)
-                    if _rd==1:
-                      _data = self.compress_aug(_data)
-                  #print('do color aug')
-                  _data = _data.astype('float32', copy=False)
-                  #print(_data.__class__)
-                  _data = self.color_aug(_data, 0.125)
-                if self.nd_mean is not None:
-                  _data = _data.astype('float32', copy=False)
-                  _data -= self.nd_mean
-                  _data *= 0.0078125
-                if self.cutoff>0:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    #print('do cutoff aug', self.cutoff)
-                    centerh = random.randint(0, _data.shape[0]-1)
-                    centerw = random.randint(0, _data.shape[1]-1)
-                    half = self.cutoff//2
-                    starth = max(0, centerh-half)
-                    endh = min(_data.shape[0], centerh+half)
-                    startw = max(0, centerw-half)
-                    endw = min(_data.shape[1], centerw+half)
-                    #print(starth, endh, startw, endw, _data.shape)
-                    _data[starth:endh, startw:endw, :] = 128
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        return io.DataBatch([batch_data], [batch_label], batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/recognition/ArcFace/metric.py b/recognition/ArcFace/metric.py
deleted file mode 100644
index 0758c52b4..000000000
--- a/recognition/ArcFace/metric.py
+++ /dev/null
@@ -1,46 +0,0 @@
-import numpy as np
-import mxnet as mx
-
-class AccMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(AccMetric, self).__init__(
-        'acc', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0]
-    pred_label = preds[1]
-    #print('ACC', label.shape, pred_label.shape)
-    if pred_label.shape != label.shape:
-        pred_label = mx.ndarray.argmax(pred_label, axis=self.axis)
-    pred_label = pred_label.asnumpy().astype('int32').flatten()
-    label = label.asnumpy()
-    if label.ndim==2:
-      label = label[:,0]
-    label = label.astype('int32').flatten()
-    assert label.shape==pred_label.shape
-    self.sum_metric += (pred_label.flat == label.flat).sum()
-    self.num_inst += len(pred_label.flat)
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    #label = labels[0].asnumpy()
-    pred = preds[-1].asnumpy()
-    #print('in loss', pred.shape)
-    #print(pred)
-    loss = pred[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-    #gt_label = preds[-2].asnumpy()
-    #print(gt_label)
diff --git a/recognition/ArcFace/parall_module_local_v1.py b/recognition/ArcFace/parall_module_local_v1.py
deleted file mode 100644
index 0963eb0c0..000000000
--- a/recognition/ArcFace/parall_module_local_v1.py
+++ /dev/null
@@ -1,529 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-import logging
-import copy
-import time
-import os
-
-import mxnet as mx
-import numpy as np
-from mxnet import context as ctx
-from mxnet.initializer import Uniform
-from mxnet.module.base_module import BaseModule
-from mxnet.module.module import Module
-from mxnet import metric
-from mxnet.model import BatchEndParam
-from mxnet import io
-import mxnet.ndarray as nd
-from config import config
-
-class ParallModule(BaseModule):
-    def __init__(self, symbol, data_names, label_names,
-                 logger=logging, context=ctx.cpu(), work_load_list=None,
-                 asymbol = None,
-                 args = None):
-        super(ParallModule, self).__init__(logger=logger)
-        self._symbol = symbol
-        self._asymbol = asymbol
-        self._data_names = data_names
-        self._label_names = label_names
-        self._context = context
-        self._work_load_list = work_load_list
-        self._num_classes = config.num_classes
-        self._batch_size = args.batch_size
-        self._verbose = args.verbose
-        self._emb_size = config.emb_size
-        self._local_class_start = args.local_class_start
-        self._iter = 0
-
-        self._curr_module = None
-
-        self._num_workers = config.num_workers
-        self._num_ctx = len(self._context)
-        self._ctx_num_classes = args.ctx_num_classes
-        self._nd_cache = {}
-        self._ctx_cpu = mx.cpu()
-        self._ctx_single_gpu = self._context[-1]
-        self._fixed_param_names = None
-        self._curr_module = Module(self._symbol, self._data_names, self._label_names, logger=self.logger,
-                        context=self._context, work_load_list=self._work_load_list,
-                        fixed_param_names=self._fixed_param_names)
-        self._arcface_modules = []
-        self._ctx_class_start = []
-        for i in range(len(self._context)):
-
-          args._ctxid = i
-          _module = Module(self._asymbol(args), self._data_names, self._label_names, logger=self.logger,
-                          context=mx.gpu(i), work_load_list=self._work_load_list,
-                          fixed_param_names=self._fixed_param_names)
-          self._arcface_modules.append(_module)
-          _c = args.local_class_start + i*args.ctx_num_classes
-          self._ctx_class_start.append(_c)
-        self._usekv = False
-        if self._usekv:
-          self._distkv = mx.kvstore.create('dist_sync')
-          self._kvinit = {}
-
-
-    def _reset_bind(self):
-        self.binded = False
-        self._curr_module = None
-
-    @property
-    def data_names(self):
-        return self._data_names
-
-    @property
-    def output_names(self):
-        return self._symbol.list_outputs()
-
-    @property
-    def data_shapes(self):
-        assert self.binded
-        return self._curr_module.data_shapes
-
-    @property
-    def label_shapes(self):
-        assert self.binded
-        return self._curr_module.label_shapes
-
-    @property
-    def output_shapes(self):
-        assert self.binded
-        return self._curr_module.output_shapes
-
-    def get_export_params(self):
-        assert self.binded and self.params_initialized
-        _g, _x = self._curr_module.get_params()
-        g = _g.copy()
-        x = _x.copy()
-        return g, x
-
-    def get_params(self):
-        assert self.binded and self.params_initialized
-        _g, _x = self._curr_module.get_params()
-        g = _g.copy()
-        x = _x.copy()
-        for _module in self._arcface_modules:
-          _g, _x = _module.get_params()
-          ag = _g.copy()
-          ax = _x.copy()
-          g.update(ag)
-          x.update(ax)
-        return g, x
-
-    def set_params(self, arg_params, aux_params, allow_missing=False, force_init=True,
-                   allow_extra=False):
-      g = arg_params
-      x = aux_params
-      #ag = {}
-      #ax = {}
-      rk = []
-      for k in g:
-        v = g[k]
-        if k.startswith('fc7'):
-          p1 = k.find('_')
-          p2 = k.rfind('_')
-          _ctxid = int(k[p1+1:p2])
-          self._arcface_modules[_ctxid].set_params({k:v}, {})
-          rk.append(k)
-      for k in rk:
-        del g[k]
-      self._curr_module.set_params(g, x)
-      #self._arcface_module.set_params(ag, ax)
-
-
-    def init_params(self, initializer=Uniform(0.01), arg_params=None, aux_params=None,
-                    allow_missing=False, force_init=False, allow_extra=False):
-        if self.params_initialized and not force_init:
-            return
-        assert self.binded, 'call bind before initializing the parameters'
-        #TODO init the same weights with all work nodes
-        self._curr_module.init_params(initializer=initializer, arg_params=None,
-                                      aux_params=None, allow_missing=allow_missing,
-                                      force_init=force_init, allow_extra=allow_extra)
-        for _module in self._arcface_modules:
-          #_initializer = initializer
-          _initializer = mx.init.Normal(0.01)
-          _module.init_params(initializer=_initializer, arg_params=None,
-                                        aux_params=None, allow_missing=allow_missing,
-                                        force_init=force_init, allow_extra=allow_extra)
-        self.params_initialized = True
-
-
-    def bind(self, data_shapes, label_shapes=None, for_training=True,
-             inputs_need_grad=False, force_rebind=False, shared_module=None):
-        print('in_bind', self.params_initialized, data_shapes, label_shapes)
-        if self.params_initialized:
-            arg_params, aux_params = self.get_params()
-
-        # force rebinding is typically used when one want to switch from
-        # training to prediction phase.
-        if force_rebind:
-            self._reset_bind()
-
-        if self.binded:
-            self.logger.warning('Already binded, ignoring bind()')
-            return
-
-        assert shared_module is None, 'shared_module for MutableModule is not supported'
-        self.for_training = for_training
-        self.inputs_need_grad = inputs_need_grad
-        self.binded = True
-        self._curr_module.bind(data_shapes, label_shapes, for_training, inputs_need_grad,
-                    force_rebind=False, shared_module=None)
-        _data_shape = data_shapes[0][1]
-        print('_data_shape', _data_shape, label_shapes)
-        for _module in self._arcface_modules:
-          _module.bind([('data', (_data_shape[0]*self._num_workers, self._emb_size))], [('softmax_label', (_data_shape[0]*self._num_workers,))], for_training, True,
-                      force_rebind=False, shared_module=None)
-        if self.params_initialized:
-            self.set_params(arg_params, aux_params)
-
-    def init_optimizer(self, kvstore='local', optimizer='sgd',
-                       optimizer_params=(('learning_rate', 0.01),), force_init=False):
-        assert self.binded and self.params_initialized
-        if self.optimizer_initialized and not force_init:
-            self.logger.warning('optimizer already initialized, ignoring.')
-            return
-
-        self._curr_module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                         force_init=force_init)
-        for _module in self._arcface_modules:
-          _module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                           force_init=force_init)
-        self.optimizer_initialized = True
-
-    def kv_push(self, key, value):
-      #if value.context!=mx.cpu():
-      #  value = value.as_in_context(mx.cpu())
-      if not key in self._kvinit:
-        self._distkv.init(key, nd.zeros_like(value))
-        self._kvinit[key] = 1
-      self._distkv.push(key, value)
-
-    #get fc1 and partial fc7
-    def forward(self, data_batch, is_train=None):
-        #g,x = self.get_params()
-        #print('{fc7_weight[0][0]}', self._iter, g['fc7_0_weight'].asnumpy()[0][0])
-        #print('{pre_fc1_weight[0][0]}', self._iter, g['pre_fc1_weight'].asnumpy()[0][0])
-
-
-        assert self.binded and self.params_initialized
-        self._curr_module.forward(data_batch, is_train=is_train)
-        if is_train:
-          self._iter+=1
-          fc1, label = self._curr_module.get_outputs(merge_multi_context=True)
-          global_fc1 = fc1
-          self.global_label = label.as_in_context(self._ctx_cpu)
-
-
-          for i, _module in enumerate(self._arcface_modules):
-            _label = self.global_label - self._ctx_class_start[i]
-            db_global_fc1 = io.DataBatch([global_fc1], [_label])
-            _module.forward(db_global_fc1) #fc7 with margin
-        #print('forward end')
-
-
-    def get_ndarray(self, context, name, shape):
-      key = "%s_%s"%(name, context)
-      #print(key)
-      if not key in self._nd_cache:
-        v = nd.zeros( shape=shape, ctx = context)
-        self._nd_cache[key] = v
-      else:
-        v = self._nd_cache[key]
-      return v
-
-    def get_ndarray2(self, context, name, arr):
-      key = "%s_%s"%(name, context)
-      #print(key)
-      if not key in self._nd_cache:
-        v = nd.zeros( shape=arr.shape, ctx = context)
-        self._nd_cache[key] = v
-      else:
-        v = self._nd_cache[key]
-      arr.copyto(v)
-      return v
-
-    def backward(self, out_grads=None):
-        #print('in backward')
-        assert self.binded and self.params_initialized
-        #tmp_ctx = self._ctx_cpu
-        tmp_ctx = self._ctx_single_gpu
-        fc7_outs = []
-        ctx_fc7_max = self.get_ndarray(tmp_ctx, 'ctx_fc7_max', (self._batch_size, len(self._context)))
-        #local_fc7_max = nd.zeros( (self.global_label.shape[0],1), ctx=mx.cpu())
-        for i, _module in enumerate(self._arcface_modules):
-          _fc7 = _module.get_outputs(merge_multi_context=True)[0]
-          fc7_outs.append(_fc7)
-          _fc7_max = nd.max(_fc7, axis=1).as_in_context(tmp_ctx)
-          ctx_fc7_max[:,i] = _fc7_max
-
-        local_fc7_max = self.get_ndarray(tmp_ctx, 'local_fc7_max', (self._batch_size, 1))
-        nd.max(ctx_fc7_max, axis=1, keepdims=True, out=local_fc7_max)
-        global_fc7_max = local_fc7_max
-        #local_fc7_sum = None
-        local_fc7_sum = self.get_ndarray(tmp_ctx, 'local_fc7_sum', (self._batch_size,1))
-        local_fc7_sum[:,:] = 0.0
-        for i, _module in enumerate(self._arcface_modules):
-          _max = self.get_ndarray2(fc7_outs[i].context, 'fc7_max', global_fc7_max)
-          fc7_outs[i] = nd.broadcast_sub(fc7_outs[i], _max)
-          fc7_outs[i] = nd.exp(fc7_outs[i])
-          _sum = nd.sum(fc7_outs[i], axis=1, keepdims=True).as_in_context(tmp_ctx)
-          local_fc7_sum += _sum
-        global_fc7_sum = local_fc7_sum
-
-        if self._iter%self._verbose==0:
-          #_ctx = self._context[-1]
-          _ctx = self._ctx_cpu
-          _probs = []
-          for i, _module in enumerate(self._arcface_modules):
-            _prob = self.get_ndarray2(_ctx, '_fc7_prob_%d'%i, fc7_outs[i])
-            _probs.append(_prob)
-          fc7_prob = self.get_ndarray(_ctx, 'test_fc7_prob', (self._batch_size, self._ctx_num_classes*len(self._context)))
-          nd.concat(*_probs, dim=1, out=fc7_prob)
-          fc7_pred = nd.argmax(fc7_prob, axis=1)
-          local_label = self.global_label - self._local_class_start
-          #local_label = self.get_ndarray2(_ctx, 'test_label', local_label)
-          _pred = nd.equal(fc7_pred, local_label)
-          print('{fc7_acc}', self._iter, nd.mean(_pred).asnumpy()[0])
-
-
-        #local_fc1_grad = []
-        #fc1_grad_ctx = self._ctx_cpu
-        fc1_grad_ctx = self._ctx_single_gpu
-        local_fc1_grad = self.get_ndarray(fc1_grad_ctx, 'local_fc1_grad', (self._batch_size,self._emb_size))
-        local_fc1_grad[:,:] = 0.0
-
-        for i, _module in enumerate(self._arcface_modules):
-          _sum = self.get_ndarray2(fc7_outs[i].context, 'fc7_sum', global_fc7_sum)
-          fc7_outs[i] = nd.broadcast_div(fc7_outs[i], _sum)
-          a = i*self._ctx_num_classes
-          b = (i+1)*self._ctx_num_classes
-          _label = self.global_label - self._ctx_class_start[i]
-          _label = self.get_ndarray2(fc7_outs[i].context, 'label', _label)
-          onehot_label = self.get_ndarray(fc7_outs[i].context, 'label_onehot', (self._batch_size, self._ctx_num_classes))
-          nd.one_hot(_label, depth=self._ctx_num_classes, on_value = 1.0, off_value = 0.0, out=onehot_label)
-          fc7_outs[i] -= onehot_label
-          _module.backward(out_grads = [fc7_outs[i]])
-          #ctx_fc1_grad = _module.get_input_grads()[0].as_in_context(mx.cpu())
-          ctx_fc1_grad = self.get_ndarray2(fc1_grad_ctx, 'ctx_fc1_grad_%d'%i, _module.get_input_grads()[0])
-          local_fc1_grad += ctx_fc1_grad
-
-        global_fc1_grad = local_fc1_grad
-        self._curr_module.backward(out_grads = [global_fc1_grad])
-
-
-    def update(self):
-        assert self.binded and self.params_initialized and self.optimizer_initialized
-        self._curr_module.update()
-        for i, _module in enumerate(self._arcface_modules):
-          _module.update()
-        mx.nd.waitall()
-
-
-    def get_outputs(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_outputs(merge_multi_context=merge_multi_context)
-        #return self._arcface_module.get_outputs(merge_multi_context=merge_multi_context)
-
-    def get_input_grads(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized and self.inputs_need_grad
-        return self._curr_module.get_input_grads(merge_multi_context=merge_multi_context)
-
-    def update_metric(self, eval_metric, labels):
-        assert self.binded and self.params_initialized
-        #self._curr_module.update_metric(eval_metric, labels)
-        #label = labels[0]
-        #print(label.shape)
-        #self._arcface_module.update_metric(eval_metric, labels)
-
-    def install_monitor(self, mon):
-        """ Install monitor on all executors """
-        assert self.binded
-        self._curr_module.install_monitor(mon)
-
-    def forward_backward(self, data_batch):
-        """A convenient function that calls both ``forward`` and ``backward``."""
-        self.forward(data_batch, is_train=True) # get fc1 and partial fc7
-        self.backward()
-
-    def fit(self, train_data, eval_data=None, eval_metric='acc',
-            epoch_end_callback=None, batch_end_callback=None, kvstore='local',
-            optimizer='sgd', optimizer_params=(('learning_rate', 0.01),),
-            eval_end_callback=None,
-            eval_batch_end_callback=None, initializer=Uniform(0.01),
-            arg_params=None, aux_params=None, allow_missing=False,
-            force_rebind=False, force_init=False, begin_epoch=0, num_epoch=None,
-            validation_metric=None, monitor=None, sparse_row_id_fn=None):
-        """Trains the module parameters.
-
-        Checkout `Module Tutorial <http://mxnet.io/tutorials/basic/module.html>`_ to see
-        a end-to-end use-case.
-
-        Parameters
-        ----------
-        train_data : DataIter
-            Train DataIter.
-        eval_data : DataIter
-            If not ``None``, will be used as validation set and the performance
-            after each epoch will be evaluated.
-        eval_metric : str or EvalMetric
-            Defaults to 'accuracy'. The performance measure used to display during training.
-            Other possible predefined metrics are:
-            'ce' (CrossEntropy), 'f1', 'mae', 'mse', 'rmse', 'top_k_accuracy'.
-        epoch_end_callback : function or list of functions
-            Each callback will be called with the current `epoch`, `symbol`, `arg_params`
-            and `aux_params`.
-        batch_end_callback : function or list of function
-            Each callback will be called with a `BatchEndParam`.
-        kvstore : str or KVStore
-            Defaults to 'local'.
-        optimizer : str or Optimizer
-            Defaults to 'sgd'.
-        optimizer_params : dict
-            Defaults to ``(('learning_rate', 0.01),)``. The parameters for
-            the optimizer constructor.
-            The default value is not a dict, just to avoid pylint warning on dangerous
-            default values.
-        eval_end_callback : function or list of function
-            These will be called at the end of each full evaluation, with the metrics over
-            the entire evaluation set.
-        eval_batch_end_callback : function or list of function
-            These will be called at the end of each mini-batch during evaluation.
-        initializer : Initializer
-            The initializer is called to initialize the module parameters when they are
-            not already initialized.
-        arg_params : dict
-            Defaults to ``None``, if not ``None``, should be existing parameters from a trained
-            model or loaded from a checkpoint (previously saved model). In this case,
-            the value here will be used to initialize the module parameters, unless they
-            are already initialized by the user via a call to `init_params` or `fit`.
-            `arg_params` has a higher priority than `initializer`.
-        aux_params : dict
-            Defaults to ``None``. Similar to `arg_params`, except for auxiliary states.
-        allow_missing : bool
-            Defaults to ``False``. Indicates whether to allow missing parameters when `arg_params`
-            and `aux_params` are not ``None``. If this is ``True``, then the missing parameters
-            will be initialized via the `initializer`.
-        force_rebind : bool
-            Defaults to ``False``. Whether to force rebinding the executors if already bound.
-        force_init : bool
-            Defaults to ``False``. Indicates whether to force initialization even if the
-            parameters are already initialized.
-        begin_epoch : int
-            Defaults to 0. Indicates the starting epoch. Usually, if resumed from a
-            checkpoint saved at a previous training phase at epoch N, then this value should be
-            N+1.
-        num_epoch : int
-            Number of epochs for training.
-        sparse_row_id_fn : A callback function
-            The function  takes `data_batch` as an input and returns a dict of
-            str -> NDArray. The resulting dict is used for pulling row_sparse
-            parameters from the kvstore, where the str key is the name of the param,
-            and the value is the row id of the param to pull.
-
-        Examples
-        --------
-        >>> # An example of using fit for training.
-        >>> # Assume training dataIter and validation dataIter are ready
-        >>> # Assume loading a previously checkpointed model
-        >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 3)
-        >>> mod.fit(train_data=train_dataiter, eval_data=val_dataiter, optimizer='sgd',
-        ...     optimizer_params={'learning_rate':0.01, 'momentum': 0.9},
-        ...     arg_params=arg_params, aux_params=aux_params,
-        ...     eval_metric='acc', num_epoch=10, begin_epoch=3)
-        """
-        assert num_epoch is not None, 'please specify number of epochs'
-        assert arg_params is None and aux_params is None
-
-        self.bind(data_shapes=train_data.provide_data, label_shapes=train_data.provide_label,
-                  for_training=True, force_rebind=force_rebind)
-        if monitor is not None:
-            self.install_monitor(monitor)
-        self.init_params(initializer=initializer, arg_params=arg_params, aux_params=aux_params,
-                         allow_missing=allow_missing, force_init=force_init)
-        self.init_optimizer(kvstore=kvstore, optimizer=optimizer,
-                            optimizer_params=optimizer_params)
-
-        if validation_metric is None:
-            validation_metric = eval_metric
-        if not isinstance(eval_metric, metric.EvalMetric):
-            eval_metric = metric.create(eval_metric)
-        epoch_eval_metric = copy.deepcopy(eval_metric)
-
-        ################################################################################
-        # training loop
-        ################################################################################
-        for epoch in range(begin_epoch, num_epoch):
-            tic = time.time()
-            eval_metric.reset()
-            epoch_eval_metric.reset()
-            nbatch = 0
-            data_iter = iter(train_data)
-            end_of_batch = False
-            next_data_batch = next(data_iter)
-            while not end_of_batch:
-                data_batch = next_data_batch
-                if monitor is not None:
-                    monitor.tic()
-                self.forward_backward(data_batch)
-                self.update()
-                assert not isinstance(data_batch, list)
-
-                #if isinstance(data_batch, list):
-                #    #print('XXX')
-                #    self.update_metric(eval_metric,
-                #                       [db.label for db in data_batch],
-                #                       pre_sliced=True)
-                #    self.update_metric(epoch_eval_metric,
-                #                       [db.label for db in data_batch],
-                #                       pre_sliced=True)
-                #else:
-                #    #print('before update metric')
-                #    self.update_metric(eval_metric, data_batch.label)
-                #    self.update_metric(epoch_eval_metric, data_batch.label)
-                #labels = data_batch.label
-                #labels = [self.global_label]
-                #self.update_metric(eval_metric, labels)
-                #self.update_metric(epoch_eval_metric, labels)
-
-                try:
-                    # pre fetch next batch
-                    next_data_batch = next(data_iter)
-                    self.prepare(next_data_batch, sparse_row_id_fn=sparse_row_id_fn)
-                except StopIteration:
-                    end_of_batch = True
-
-                if monitor is not None:
-                    monitor.toc_print()
-
-                #if end_of_batch:
-                #    eval_name_vals = epoch_eval_metric.get_name_value()
-
-                if batch_end_callback is not None:
-                    batch_end_params = BatchEndParam(epoch=epoch, nbatch=nbatch,
-                                                     eval_metric=None,
-                                                     locals=locals())
-                    batch_end_callback(batch_end_params)
-                    #for callback in _as_list(batch_end_callback):
-                    #    callback(batch_end_params)
-                nbatch += 1
-
-            # one epoch of training is finished
-            #for name, val in eval_name_vals:
-            #    self.logger.info('Epoch[%d] Train-%s=%f', epoch, name, val)
-            toc = time.time()
-            self.logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc-tic))
-
-            # sync aux params across devices
-            arg_params, aux_params = self.get_params()
-            self.set_params(arg_params, aux_params)
-
-            # end of 1 epoch, reset the data-iter for another epoch
-            train_data.reset()
-
diff --git a/recognition/ArcFace/sample_config.py b/recognition/ArcFace/sample_config.py
deleted file mode 100644
index d91887c32..000000000
--- a/recognition/ArcFace/sample_config.py
+++ /dev/null
@@ -1,222 +0,0 @@
-import numpy as np
-import os
-from easydict import EasyDict as edict
-
-config = edict()
-
-config.bn_mom = 0.9
-config.workspace = 256
-config.emb_size = 512
-config.ckpt_embedding = True
-config.net_se = 0
-config.net_act = 'prelu'
-config.net_unit = 3
-config.net_input = 1
-config.net_blocks = [1,4,6,2]
-config.net_output = 'E'
-config.net_multiplier = 1.0
-config.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-config.ce_loss = True
-config.fc7_lr_mult = 1.0
-config.fc7_wd_mult = 1.0
-config.fc7_no_bias = False
-config.max_steps = 0
-config.data_rand_mirror = True
-config.data_cutoff = False
-config.data_color = 0
-config.data_images_filter = 0
-config.count_flops = True
-config.memonger = False #not work now
-
-
-# network settings
-network = edict()
-
-network.r100 = edict()
-network.r100.net_name = 'fresnet'
-network.r100.num_layers = 100
-
-network.r100fc = edict()
-network.r100fc.net_name = 'fresnet'
-network.r100fc.num_layers = 100
-network.r100fc.net_output = 'FC'
-
-network.r50 = edict()
-network.r50.net_name = 'fresnet'
-network.r50.num_layers = 50
-
-network.r50v1 = edict()
-network.r50v1.net_name = 'fresnet'
-network.r50v1.num_layers = 50
-network.r50v1.net_unit = 1
-
-network.d169 = edict()
-network.d169.net_name = 'fdensenet'
-network.d169.num_layers = 169
-network.d169.per_batch_size = 64
-network.d169.densenet_dropout = 0.0
-
-network.d201 = edict()
-network.d201.net_name = 'fdensenet'
-network.d201.num_layers = 201
-network.d201.per_batch_size = 64
-network.d201.densenet_dropout = 0.0
-
-network.y1 = edict()
-network.y1.net_name = 'fmobilefacenet'
-network.y1.emb_size = 128
-network.y1.net_output = 'GDC'
-
-network.y2 = edict()
-network.y2.net_name = 'fmobilefacenet'
-network.y2.emb_size = 256
-network.y2.net_output = 'GDC'
-network.y2.net_blocks = [2,8,16,4]
-
-network.m1 = edict()
-network.m1.net_name = 'fmobilenet'
-network.m1.emb_size = 256
-network.m1.net_output = 'GDC'
-network.m1.net_multiplier = 1.0
-
-network.m05 = edict()
-network.m05.net_name = 'fmobilenet'
-network.m05.emb_size = 256
-network.m05.net_output = 'GDC'
-network.m05.net_multiplier = 0.5
-
-network.mnas = edict()
-network.mnas.net_name = 'fmnasnet'
-network.mnas.emb_size = 256
-network.mnas.net_output = 'GDC'
-network.mnas.net_multiplier = 1.0
-
-network.mnas05 = edict()
-network.mnas05.net_name = 'fmnasnet'
-network.mnas05.emb_size = 256
-network.mnas05.net_output = 'GDC'
-network.mnas05.net_multiplier = 0.5
-
-network.mnas025 = edict()
-network.mnas025.net_name = 'fmnasnet'
-network.mnas025.emb_size = 256
-network.mnas025.net_output = 'GDC'
-network.mnas025.net_multiplier = 0.25
-
-network.vargfacenet = edict()
-network.vargfacenet.net_name = 'vargfacenet'
-network.vargfacenet.net_multiplier = 1.25
-network.vargfacenet.emb_size = 512
-network.vargfacenet.net_output='J'
-
-# dataset settings
-dataset = edict()
-
-dataset.emore = edict()
-dataset.emore.dataset = 'emore'
-dataset.emore.dataset_path = '../datasets/faces_emore'
-dataset.emore.num_classes = 85742
-dataset.emore.image_shape = (112,112,3)
-dataset.emore.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-
-dataset.retina = edict()
-dataset.retina.dataset = 'retina'
-dataset.retina.dataset_path = '../datasets/ms1m-retinaface-t1'
-dataset.retina.num_classes = 93431
-dataset.retina.image_shape = (112,112,3)
-dataset.retina.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-
-loss = edict()
-loss.softmax = edict()
-loss.softmax.loss_name = 'softmax'
-
-loss.nsoftmax = edict()
-loss.nsoftmax.loss_name = 'margin_softmax'
-loss.nsoftmax.loss_s = 64.0
-loss.nsoftmax.loss_m1 = 1.0
-loss.nsoftmax.loss_m2 = 0.0
-loss.nsoftmax.loss_m3 = 0.0
-
-loss.arcface = edict()
-loss.arcface.loss_name = 'margin_softmax'
-loss.arcface.loss_s = 64.0
-loss.arcface.loss_m1 = 1.0
-loss.arcface.loss_m2 = 0.5
-loss.arcface.loss_m3 = 0.0
-
-loss.cosface = edict()
-loss.cosface.loss_name = 'margin_softmax'
-loss.cosface.loss_s = 64.0
-loss.cosface.loss_m1 = 1.0
-loss.cosface.loss_m2 = 0.0
-loss.cosface.loss_m3 = 0.35
-
-loss.combined = edict()
-loss.combined.loss_name = 'margin_softmax'
-loss.combined.loss_s = 64.0
-loss.combined.loss_m1 = 1.0
-loss.combined.loss_m2 = 0.3
-loss.combined.loss_m3 = 0.2
-
-loss.triplet = edict()
-loss.triplet.loss_name = 'triplet'
-loss.triplet.images_per_identity = 5
-loss.triplet.triplet_alpha = 0.3
-loss.triplet.triplet_bag_size = 7200
-loss.triplet.triplet_max_ap = 0.0
-loss.triplet.per_batch_size = 60
-loss.triplet.lr = 0.05
-
-loss.atriplet = edict()
-loss.atriplet.loss_name = 'atriplet'
-loss.atriplet.images_per_identity = 5
-loss.atriplet.triplet_alpha = 0.35
-loss.atriplet.triplet_bag_size = 7200
-loss.atriplet.triplet_max_ap = 0.0
-loss.atriplet.per_batch_size = 60
-loss.atriplet.lr = 0.05
-
-# default settings
-default = edict()
-
-# default network
-default.network = 'r100'
-default.pretrained = ''
-default.pretrained_epoch = 1
-# default dataset
-default.dataset = 'emore'
-default.loss = 'arcface'
-default.frequent = 20
-default.verbose = 2000
-default.kvstore = 'device'
-
-default.end_epoch = 10000
-default.lr = 0.1
-default.wd = 0.0005
-default.mom = 0.9
-default.per_batch_size = 128
-default.ckpt = 3
-default.lr_steps = '100000,160000,220000'
-default.models_root = './models'
-
-
-def generate_config(_network, _dataset, _loss):
-    for k, v in loss[_loss].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    for k, v in network[_network].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    for k, v in dataset[_dataset].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    config.loss = _loss
-    config.network = _network
-    config.dataset = _dataset
-    config.num_workers = 1
-    if 'DMLC_NUM_WORKER' in os.environ:
-      config.num_workers = int(os.environ['DMLC_NUM_WORKER'])
-
diff --git a/recognition/ArcFace/train.py b/recognition/ArcFace/train.py
deleted file mode 100644
index 9216f9ab5..000000000
--- a/recognition/ArcFace/train.py
+++ /dev/null
@@ -1,378 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import sklearn
-import pickle
-import numpy as np
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-from config import config, default, generate_config
-from metric import *
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import flops_counter
-import verification
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'symbol'))
-import fresnet
-import fmobilefacenet
-import fmobilenet
-import fmnasnet
-import fdensenet
-import vargfacenet
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--dataset', default=default.dataset, help='dataset config')
-  parser.add_argument('--network', default=default.network, help='network config')
-  parser.add_argument('--loss', default=default.loss, help='loss config')
-  args, rest = parser.parse_known_args()
-  generate_config(args.network, args.dataset, args.loss)
-  parser.add_argument('--models-root', default=default.models_root, help='root directory to save model.')
-  parser.add_argument('--pretrained', default=default.pretrained, help='pretrained model to load')
-  parser.add_argument('--pretrained-epoch', type=int, default=default.pretrained_epoch, help='pretrained epoch to load')
-  parser.add_argument('--ckpt', type=int, default=default.ckpt, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--verbose', type=int, default=default.verbose, help='do verification testing and model saving every verbose batches')
-  parser.add_argument('--lr', type=float, default=default.lr, help='start learning rate')
-  parser.add_argument('--lr-steps', type=str, default=default.lr_steps, help='steps of lr changing')
-  parser.add_argument('--wd', type=float, default=default.wd, help='weight decay')
-  parser.add_argument('--mom', type=float, default=default.mom, help='momentum')
-  parser.add_argument('--frequent', type=int, default=default.frequent, help='')
-  parser.add_argument('--per-batch-size', type=int, default=default.per_batch_size, help='batch size in each context')
-  parser.add_argument('--kvstore', type=str, default=default.kvstore, help='kvstore setting')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol(args):
-  embedding = eval(config.net_name).get_symbol()
-  all_label = mx.symbol.Variable('softmax_label')
-  gt_label = all_label
-  is_softmax = True
-  if config.loss_name=='softmax': #softmax
-    _weight = mx.symbol.Variable("fc7_weight", shape=(config.num_classes, config.emb_size), 
-        lr_mult=config.fc7_lr_mult, wd_mult=config.fc7_wd_mult, init=mx.init.Normal(0.01))
-    if config.fc7_no_bias:
-      fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, no_bias = True, num_hidden=config.num_classes, name='fc7')
-    else:
-      _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
-      fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=config.num_classes, name='fc7')
-  elif config.loss_name=='margin_softmax':
-    _weight = mx.symbol.Variable("fc7_weight", shape=(config.num_classes, config.emb_size), 
-        lr_mult=config.fc7_lr_mult, wd_mult=config.fc7_wd_mult, init=mx.init.Normal(0.01))
-    s = config.loss_s
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=config.num_classes, name='fc7')
-    if config.loss_m1!=1.0 or config.loss_m2!=0.0 or config.loss_m3!=0.0:
-      if config.loss_m1==1.0 and config.loss_m2==0.0:
-        s_m = s*config.loss_m3
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = config.num_classes, on_value = s_m, off_value = 0.0)
-        fc7 = fc7-gt_one_hot
-      else:
-        zy = mx.sym.pick(fc7, gt_label, axis=1)
-        cos_t = zy/s
-        t = mx.sym.arccos(cos_t)
-        if config.loss_m1!=1.0:
-          t = t*config.loss_m1
-        if config.loss_m2>0.0:
-          t = t+config.loss_m2
-        body = mx.sym.cos(t)
-        if config.loss_m3>0.0:
-          body = body - config.loss_m3
-        new_zy = body*s
-        diff = new_zy - zy
-        diff = mx.sym.expand_dims(diff, 1)
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = config.num_classes, on_value = 1.0, off_value = 0.0)
-        body = mx.sym.broadcast_mul(gt_one_hot, diff)
-        fc7 = fc7+body
-  elif config.loss_name.find('triplet')>=0:
-    is_softmax = False
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-    anchor = mx.symbol.slice_axis(nembedding, axis=0, begin=0, end=args.per_batch_size//3)
-    positive = mx.symbol.slice_axis(nembedding, axis=0, begin=args.per_batch_size//3, end=2*args.per_batch_size//3)
-    negative = mx.symbol.slice_axis(nembedding, axis=0, begin=2*args.per_batch_size//3, end=args.per_batch_size)
-    if config.loss_name=='triplet':
-      ap = anchor - positive
-      an = anchor - negative
-      ap = ap*ap
-      an = an*an
-      ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-      an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-      triplet_loss = mx.symbol.Activation(data = (ap-an+config.triplet_alpha), act_type='relu')
-      triplet_loss = mx.symbol.mean(triplet_loss)
-    else:
-      ap = anchor*positive
-      an = anchor*negative
-      ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-      an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-      ap = mx.sym.arccos(ap)
-      an = mx.sym.arccos(an)
-      triplet_loss = mx.symbol.Activation(data = (ap-an+config.triplet_alpha), act_type='relu')
-      triplet_loss = mx.symbol.mean(triplet_loss)
-    triplet_loss = mx.symbol.MakeLoss(triplet_loss)
-  out_list = [mx.symbol.BlockGrad(embedding)]
-  if is_softmax:
-    softmax = mx.symbol.SoftmaxOutput(data=fc7, label = gt_label, name='softmax', normalization='valid')
-    out_list.append(softmax)
-    if config.ce_loss:
-      #ce_loss = mx.symbol.softmax_cross_entropy(data=fc7, label = gt_label, name='ce_loss')/args.per_batch_size
-      body = mx.symbol.SoftmaxActivation(data=fc7)
-      body = mx.symbol.log(body)
-      _label = mx.sym.one_hot(gt_label, depth = config.num_classes, on_value = -1.0, off_value = 0.0)
-      body = body*_label
-      ce_loss = mx.symbol.sum(body)/args.per_batch_size
-      out_list.append(mx.symbol.BlockGrad(ce_loss))
-  else:
-    out_list.append(mx.sym.BlockGrad(gt_label))
-    out_list.append(triplet_loss)
-  out = mx.symbol.Group(out_list)
-  return out
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in range(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = os.path.join(args.models_root, '%s-%s-%s'%(args.network, args.loss, args.dataset), 'model')
-    prefix_dir = os.path.dirname(prefix)
-    print('prefix', prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    args.ctx_num = len(ctx)
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = config.image_shape[2]
-    config.batch_size = args.batch_size
-    config.per_batch_size = args.per_batch_size
-
-    data_dir = config.dataset_path
-    path_imgrec = None
-    path_imglist = None
-    image_size = config.image_shape[0:2]
-    assert len(image_size)==2
-    assert image_size[0]==image_size[1]
-    print('image_size', image_size)
-    print('num_classes', config.num_classes)
-    path_imgrec = os.path.join(data_dir, "train.rec")
-
-    print('Called with argument:', args, config)
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-    mean = None
-
-    begin_epoch = 0
-    if len(args.pretrained)==0:
-      arg_params = None
-      aux_params = None
-      sym = get_symbol(args)
-      if config.net_name=='spherenet':
-        data_shape_dict = {'data' : (args.per_batch_size,)+data_shape}
-        spherenet.init_weights(sym, data_shape_dict, args.num_layers)
-    else:
-      print('loading', args.pretrained, args.pretrained_epoch)
-      _, arg_params, aux_params = mx.model.load_checkpoint(args.pretrained, args.pretrained_epoch)
-      sym = get_symbol(args)
-
-    if config.count_flops:
-      all_layers = sym.get_internals()
-      _sym = all_layers['fc1_output']
-      FLOPs = flops_counter.count_flops(_sym, data=(1,3,image_size[0],image_size[1]))
-      _str = flops_counter.flops_str(FLOPs)
-      print('Network FLOPs: %s'%_str)
-
-    #label_name = 'softmax_label'
-    #label_shape = (args.batch_size,)
-    model = mx.mod.Module(
-        context       = ctx,
-        symbol        = sym,
-    )
-    val_dataiter = None
-
-    if config.loss_name.find('triplet')>=0:
-      from triplet_image_iter import FaceImageIter
-      triplet_params = [config.triplet_bag_size, config.triplet_alpha, config.triplet_max_ap]
-      train_dataiter = FaceImageIter(
-          batch_size           = args.batch_size,
-          data_shape           = data_shape,
-          path_imgrec          = path_imgrec,
-          shuffle              = True,
-          rand_mirror          = config.data_rand_mirror,
-          mean                 = mean,
-          cutoff               = config.data_cutoff,
-          ctx_num              = args.ctx_num,
-          images_per_identity  = config.images_per_identity,
-          triplet_params       = triplet_params,
-          mx_model             = model,
-      )
-      _metric = LossValueMetric()
-      eval_metrics = [mx.metric.create(_metric)]
-    else:
-      from image_iter import FaceImageIter
-      train_dataiter = FaceImageIter(
-          batch_size           = args.batch_size,
-          data_shape           = data_shape,
-          path_imgrec          = path_imgrec,
-          shuffle              = True,
-          rand_mirror          = config.data_rand_mirror,
-          mean                 = mean,
-          cutoff               = config.data_cutoff,
-          color_jittering      = config.data_color,
-          images_filter        = config.data_images_filter,
-      )
-      metric1 = AccMetric()
-      eval_metrics = [mx.metric.create(metric1)]
-      if config.ce_loss:
-        metric2 = LossValueMetric()
-        eval_metrics.append( mx.metric.create(metric2) )
-
-    if config.net_name=='fresnet' or config.net_name=='fmobilefacenet':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-    #initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    _rescale = 1.0/args.ctx_num
-    opt = optimizer.SGD(learning_rate=args.lr, momentum=args.mom, wd=args.wd, rescale_grad=_rescale)
-    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
-
-    ver_list = []
-    ver_name_list = []
-    for name in config.val_targets:
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in range(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, None, None)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in range(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for step in lr_steps:
-        if mbatch==step:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        is_highest = False
-        if len(acc_list)>0:
-          #lfw_score = acc_list[0]
-          #if lfw_score>highest_acc[0]:
-          #  highest_acc[0] = lfw_score
-          #  if lfw_score>=0.998:
-          #    do_save = True
-          score = sum(acc_list)
-          if acc_list[-1]>=highest_acc[-1]:
-            if acc_list[-1]>highest_acc[-1]:
-              is_highest = True
-            else:
-              if score>=highest_acc[0]:
-                is_highest = True
-                highest_acc[0] = score
-            highest_acc[-1] = acc_list[-1]
-            #if lfw_score>=0.99:
-            #  do_save = True
-        if is_highest:
-          do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt==2:
-          do_save = True
-        elif args.ckpt==3:
-          msave = 1
-
-        if do_save:
-          print('saving', msave)
-          arg, aux = model.get_params()
-          if config.ckpt_embedding:
-            all_layers = model.symbol.get_internals()
-            _sym = all_layers['fc1_output']
-            _arg = {}
-            for k in arg:
-              if not k.startswith('fc7'):
-                _arg[k] = arg[k]
-            mx.model.save_checkpoint(prefix, msave, _sym, _arg, aux)
-          else:
-            mx.model.save_checkpoint(prefix, msave, model.symbol, arg, aux)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if config.max_steps>0 and mbatch>config.max_steps:
-        sys.exit(0)
-
-    epoch_cb = None
-    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = 999999,
-        eval_data          = val_dataiter,
-        eval_metric        = eval_metrics,
-        kvstore            = args.kvstore,
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/recognition/ArcFace/train_parall.py b/recognition/ArcFace/train_parall.py
deleted file mode 100644
index a400ac4d9..000000000
--- a/recognition/ArcFace/train_parall.py
+++ /dev/null
@@ -1,357 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import sklearn
-import numpy as np
-from image_iter import FaceImageIter
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import flops_counter
-from config import config, default, generate_config
-import verification
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'symbol'))
-import fresnet
-import fmobilefacenet
-import fmobilenet
-import fmnasnet
-import fdensenet
-import vargfacenet
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train parall face network')
-  # general
-  parser.add_argument('--dataset', default=default.dataset, help='dataset config')
-  parser.add_argument('--network', default=default.network, help='network config')
-  parser.add_argument('--loss', default=default.loss, help='loss config')
-  args, rest = parser.parse_known_args()
-  generate_config(args.network, args.dataset, args.loss)
-  parser.add_argument('--models-root', default=default.models_root, help='root directory to save model.')
-  parser.add_argument('--pretrained', default=default.pretrained, help='pretrained model to load')
-  parser.add_argument('--pretrained-epoch', type=int, default=default.pretrained_epoch, help='pretrained epoch to load')
-  parser.add_argument('--ckpt', type=int, default=default.ckpt, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--verbose', type=int, default=default.verbose, help='do verification testing and model saving every verbose batches')
-  parser.add_argument('--lr', type=float, default=default.lr, help='start learning rate')
-  parser.add_argument('--lr-steps', type=str, default=default.lr_steps, help='steps of lr changing')
-  parser.add_argument('--wd', type=float, default=default.wd, help='weight decay')
-  parser.add_argument('--mom', type=float, default=default.mom, help='momentum')
-  parser.add_argument('--frequent', type=int, default=default.frequent, help='')
-  parser.add_argument('--per-batch-size', type=int, default=default.per_batch_size, help='batch size in each context')
-  parser.add_argument('--kvstore', type=str, default=default.kvstore, help='kvstore setting')
-  parser.add_argument('--worker-id', type=int, default=0, help='worker id for dist training, starts from 0')
-  parser.add_argument('--extra-model-name', type=str, default='', help='extra model name')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol_embedding():
-  embedding = eval(config.net_name).get_symbol()
-  all_label = mx.symbol.Variable('softmax_label')
-  #embedding = mx.symbol.BlockGrad(embedding)
-  all_label = mx.symbol.BlockGrad(all_label)
-  out_list = [embedding, all_label]
-  out = mx.symbol.Group(out_list)
-  return out
-
-def get_symbol_arcface(args):
-  embedding = mx.symbol.Variable('data')
-  all_label = mx.symbol.Variable('softmax_label')
-  gt_label = all_label
-  is_softmax = True
-  #print('call get_sym_arcface with', args, config)
-  _weight = mx.symbol.Variable("fc7_%d_weight"%args._ctxid, shape=(args.ctx_num_classes, config.emb_size), 
-      lr_mult=config.fc7_lr_mult, wd_mult=config.fc7_wd_mult)
-  if config.loss_name=='softmax': #softmax
-    fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, no_bias = True, num_hidden=args.ctx_num_classes, name='fc7_%d'%args._ctxid)
-  elif config.loss_name=='margin_softmax':
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n_%d'%args._ctxid)
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.ctx_num_classes, name='fc7_%d'%args._ctxid)
-    if config.loss_m1!=1.0 or config.loss_m2!=0.0 or config.loss_m3!=0.0:
-      gt_one_hot = mx.sym.one_hot(gt_label, depth = args.ctx_num_classes, on_value = 1.0, off_value = 0.0)
-      if config.loss_m1==1.0 and config.loss_m2==0.0:
-        _one_hot = gt_one_hot*config.loss_m3
-        fc7 = fc7-_one_hot
-      else:
-        fc7_onehot = fc7 * gt_one_hot
-        cos_t = fc7_onehot
-        t = mx.sym.arccos(cos_t)
-        if config.loss_m1!=1.0:
-          t = t*config.loss_m1
-        if config.loss_m2!=0.0:
-          t = t+config.loss_m2
-        margin_cos = mx.sym.cos(t)
-        if config.loss_m3!=0.0:
-          margin_cos = margin_cos - config.loss_m3
-        margin_fc7 = margin_cos
-        margin_fc7_onehot = margin_fc7 * gt_one_hot
-        diff = margin_fc7_onehot - fc7_onehot
-        fc7 = fc7+diff
-    fc7 = fc7*config.loss_s
-
-  out_list = []
-  out_list.append(fc7)
-  if config.loss_name=='softmax': #softmax
-    out_list.append(gt_label)
-  out = mx.symbol.Group(out_list)
-  return out
-
-def train_net(args):
-    #_seed = 727
-    #random.seed(_seed)
-    #np.random.seed(_seed)
-    #mx.random.seed(_seed)
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in range(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    if len(args.extra_model_name)==0:
-      prefix = os.path.join(args.models_root, '%s-%s-%s'%(args.network, args.loss, args.dataset), 'model')
-    else:
-      prefix = os.path.join(args.models_root, '%s-%s-%s-%s'%(args.network, args.loss, args.dataset, args.extra_model_name), 'model')
-    prefix_dir = os.path.dirname(prefix)
-    print('prefix', prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    args.ctx_num = len(ctx)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = config.image_shape[2]
-    config.batch_size = args.batch_size
-    config.per_batch_size = args.per_batch_size
-    data_dir = config.dataset_path
-    path_imgrec = None
-    path_imglist = None
-    image_size = config.image_shape[0:2]
-    assert len(image_size)==2
-    assert image_size[0]==image_size[1]
-    print('image_size', image_size)
-    print('num_classes', config.num_classes)
-    path_imgrec = os.path.join(data_dir, "train.rec")
-
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-
-    num_workers = config.num_workers
-    global_num_ctx = num_workers * args.ctx_num
-    if config.num_classes%global_num_ctx==0:
-      args.ctx_num_classes = config.num_classes//global_num_ctx
-    else:
-      args.ctx_num_classes = config.num_classes//global_num_ctx+1
-    args.local_num_classes = args.ctx_num_classes * args.ctx_num
-    args.local_class_start = args.local_num_classes * args.worker_id
-
-    #if len(args.partial)==0:
-    #  local_classes_range = (0, args.num_classes)
-    #else:
-    #  _vec = args.partial.split(',')
-    #  local_classes_range = (int(_vec[0]), int(_vec[1]))
-
-    #args.partial_num_classes = local_classes_range[1] - local_classes_range[0]
-    #args.partial_start = local_classes_range[0]
-
-    print('Called with argument:', args, config)
-    mean = None
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    arg_params = None
-    aux_params = None
-    if len(args.pretrained)==0:
-      esym = get_symbol_embedding()
-      asym = get_symbol_arcface
-    else:
-      #assert False
-      print('loading', args.pretrained, args.pretrained_epoch)
-      _, arg_params, aux_params = mx.model.load_checkpoint(args.pretrained, args.pretrained_epoch)
-      esym = get_symbol_embedding()
-      asym = get_symbol_arcface
-
-    if config.count_flops:
-      all_layers = esym.get_internals()
-      _sym = all_layers['fc1_output']
-      FLOPs = flops_counter.count_flops(_sym, data=(1,3,image_size[0],image_size[1]))
-      _str = flops_counter.flops_str(FLOPs)
-      print('Network FLOPs: %s'%_str)
-
-    if config.num_workers==1:
-      from parall_module_local_v1 import ParallModule
-    else:
-      from parall_module_dist import ParallModule
-
-    model = ParallModule(
-        context       = ctx,
-        symbol        = esym,
-        data_names    = ['data'],
-        label_names    = ['softmax_label'],
-        asymbol       = asym,
-        args = args,
-    )
-    val_dataiter = None
-    train_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec,
-        shuffle              = True,
-        rand_mirror          = config.data_rand_mirror,
-        mean                 = mean,
-        cutoff               = config.data_cutoff,
-        color_jittering      = config.data_color,
-        images_filter        = config.data_images_filter,
-    )
-
-
-    
-    if config.net_name=='fresnet' or config.net_name=='fmobilefacenet':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-
-    _rescale = 1.0/args.batch_size
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
-
-
-    ver_list = []
-    ver_name_list = []
-    for name in config.val_targets:
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in range(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, None, None)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in range(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for step in lr_steps:
-        if mbatch==step:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        is_highest = False
-        if len(acc_list)>0:
-          #lfw_score = acc_list[0]
-          #if lfw_score>highest_acc[0]:
-          #  highest_acc[0] = lfw_score
-          #  if lfw_score>=0.998:
-          #    do_save = True
-          score = sum(acc_list)
-          if acc_list[-1]>=highest_acc[-1]:
-            if acc_list[-1]>highest_acc[-1]:
-              is_highest = True
-            else:
-              if score>=highest_acc[0]:
-                is_highest = True
-                highest_acc[0] = score
-            highest_acc[-1] = acc_list[-1]
-            #if lfw_score>=0.99:
-            #  do_save = True
-        if is_highest:
-          do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt==2:
-          do_save = True
-        elif args.ckpt==3:
-          msave = 1
-
-        if do_save:
-          print('saving', msave)
-          arg, aux = model.get_export_params()
-          all_layers = model.symbol.get_internals()
-          _sym = all_layers['fc1_output']
-          mx.model.save_checkpoint(prefix, msave, _sym, arg, aux)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if config.max_steps>0 and mbatch>config.max_steps:
-        sys.exit(0)
-
-    epoch_cb = None
-    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = 999999,
-        eval_data          = val_dataiter,
-        #eval_metric        = eval_metrics,
-        kvstore            = args.kvstore,
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/recognition/ArcFace/triplet_image_iter.py b/recognition/ArcFace/triplet_image_iter.py
deleted file mode 100644
index 6c8a43e1c..000000000
--- a/recognition/ArcFace/triplet_image_iter.py
+++ /dev/null
@@ -1,609 +0,0 @@
-# THIS FILE IS FOR EXPERIMENTS, USE image_iter.py FOR NORMAL IMAGE LOADING.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-
-import mxnet as mx
-from mxnet import ndarray as nd
-#from . import _ndarray_internal as _internal
-#from mxnet._ndarray_internal import _cvimresize as imresize
-#from ._ndarray_internal import _cvcopyMakeBorder as copyMakeBorder
-from mxnet import io
-from mxnet import recordio
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_preprocess
-
-logger = logging.getLogger()
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None,
-                 rand_mirror = False, cutoff = 0,
-                 ctx_num = 0, images_per_identity = 0,
-                 triplet_params = None,
-                 mx_model = None,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        assert shuffle
-        logging.info('loading recordio %s...',
-                     path_imgrec)
-        path_imgidx = path_imgrec[0:-4]+".idx"
-        self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-        s = self.imgrec.read_idx(0)
-        header, _ = recordio.unpack(s)
-        assert header.flag>0
-        print('header0 label', header.label)
-        self.header0 = (int(header.label[0]), int(header.label[1]))
-        #assert(header.flag==1)
-        self.imgidx = range(1, int(header.label[0]))
-        self.id2range = {}
-        self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-        for identity in self.seq_identity:
-          s = self.imgrec.read_idx(identity)
-          header, _ = recordio.unpack(s)
-          a,b = int(header.label[0]), int(header.label[1])
-          self.id2range[identity] = (a,b)
-
-        print('id2range', len(self.id2range))
-        self.seq = self.imgidx
-        print(len(self.seq))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        #self.cast_aug = mx.image.CastAug()
-        #self.color_aug = mx.image.ColorJitterAug(0.4, 0.4, 0.4)
-        self.ctx_num = ctx_num 
-        self.per_batch_size = int(self.batch_size/self.ctx_num)
-        self.images_per_identity = images_per_identity
-        if self.images_per_identity>0:
-          self.identities = int(self.per_batch_size/self.images_per_identity)
-          self.per_identities = self.identities
-          self.repeat = 3000000.0/(self.images_per_identity*len(self.id2range))
-          self.repeat = int(self.repeat)
-          print(self.images_per_identity, self.identities, self.repeat)
-        self.mx_model = mx_model
-        self.triplet_params = triplet_params
-        self.triplet_mode = False
-        #self.provide_label = None
-        self.provide_label = [(label_name, (batch_size,))]
-        if self.triplet_params is not None:
-          assert self.images_per_identity>0
-          assert self.mx_model is not None
-          self.triplet_bag_size = self.triplet_params[0]
-          self.triplet_alpha = self.triplet_params[1]
-          self.triplet_max_ap = self.triplet_params[2]
-          assert self.triplet_bag_size>0
-          assert self.triplet_alpha>=0.0
-          assert self.triplet_alpha<=1.0
-          self.triplet_mode = True
-          self.triplet_cur = 0
-          self.triplet_seq = []
-          self.triplet_reset()
-          self.seq_min_size = self.batch_size*2
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-        self.times = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
-        #self.reset()
-
-
-
-    def pairwise_dists(self, embeddings):
-      nd_embedding_list = []
-      for i in range(self.ctx_num):
-        nd_embedding = mx.nd.array(embeddings, mx.gpu(i))
-        nd_embedding_list.append(nd_embedding)
-      nd_pdists = []
-      pdists = []
-      for idx in range(embeddings.shape[0]):
-        emb_idx = idx%self.ctx_num
-        nd_embedding = nd_embedding_list[emb_idx]
-        a_embedding = nd_embedding[idx]
-        body = mx.nd.broadcast_sub(a_embedding, nd_embedding)
-        body = body*body
-        body = mx.nd.sum_axis(body, axis=1)
-        nd_pdists.append(body)
-        if len(nd_pdists)==self.ctx_num or idx==embeddings.shape[0]-1:
-          for x in nd_pdists:
-            pdists.append(x.asnumpy())
-          nd_pdists = []
-      return pdists
-
-    def pick_triplets(self, embeddings, nrof_images_per_class):
-      emb_start_idx = 0
-      triplets = []
-      people_per_batch = len(nrof_images_per_class)
-      #self.time_reset()
-      pdists = self.pairwise_dists(embeddings)
-      #self.times[3] += self.time_elapsed()
-
-      for i in range(people_per_batch):
-          nrof_images = int(nrof_images_per_class[i])
-          for j in range(1,nrof_images):
-              #self.time_reset()
-              a_idx = emb_start_idx + j - 1
-              #neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
-              neg_dists_sqr = pdists[a_idx]
-              #self.times[3] += self.time_elapsed()
-
-              for pair in range(j, nrof_images): # For every possible positive pair.
-                  p_idx = emb_start_idx + pair
-                  #self.time_reset()
-                  pos_dist_sqr = np.sum(np.square(embeddings[a_idx]-embeddings[p_idx]))
-                  #self.times[4] += self.time_elapsed()
-                  #self.time_reset()
-                  neg_dists_sqr[emb_start_idx:emb_start_idx+nrof_images] = np.NaN
-                  if self.triplet_max_ap>0.0:
-                    if pos_dist_sqr>self.triplet_max_ap:
-                      continue
-                  all_neg = np.where(np.logical_and(neg_dists_sqr-pos_dist_sqr<self.triplet_alpha, pos_dist_sqr<neg_dists_sqr))[0]  # FaceNet selection
-                  #self.times[5] += self.time_elapsed()
-                  #self.time_reset()
-                  #all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
-                  nrof_random_negs = all_neg.shape[0]
-                  if nrof_random_negs>0:
-                      rnd_idx = np.random.randint(nrof_random_negs)
-                      n_idx = all_neg[rnd_idx]
-                      triplets.append( (a_idx, p_idx, n_idx) )
-          emb_start_idx += nrof_images
-      np.random.shuffle(triplets)
-      return triplets
-
-    def triplet_reset(self):
-      #reset self.oseq by identities seq
-      self.triplet_cur = 0
-      ids = []
-      for k in self.id2range:
-        ids.append(k)
-      random.shuffle(ids)
-      self.triplet_seq = []
-      for _id in ids:
-        v = self.id2range[_id]
-        _list = range(*v)
-        random.shuffle(_list)
-        if len(_list)>self.images_per_identity:
-          _list = _list[0:self.images_per_identity]
-        self.triplet_seq += _list
-      print('triplet_seq', len(self.triplet_seq))
-      assert len(self.triplet_seq)>=self.triplet_bag_size
-
-    def time_reset(self):
-      self.time_now = datetime.datetime.now()
-
-    def time_elapsed(self):
-      time_now = datetime.datetime.now()
-      diff = time_now - self.time_now
-      return diff.total_seconds()
-
-
-    def select_triplets(self):
-      self.seq = []
-      while len(self.seq)<self.seq_min_size:
-        self.time_reset()
-        embeddings = None
-        bag_size = self.triplet_bag_size
-        batch_size = self.batch_size
-        #data = np.zeros( (bag_size,)+self.data_shape )
-        #label = np.zeros( (bag_size,) )
-        tag = []
-        #idx = np.zeros( (bag_size,) )
-        print('eval %d images..'%bag_size, self.triplet_cur)
-        print('triplet time stat', self.times)
-        if self.triplet_cur+bag_size>len(self.triplet_seq):
-          self.triplet_reset()
-          #bag_size = min(bag_size, len(self.triplet_seq))
-          print('eval %d images..'%bag_size, self.triplet_cur)
-        self.times[0] += self.time_elapsed()
-        self.time_reset()
-        #print(data.shape)
-        data = nd.zeros( self.provide_data[0][1] )
-        label = None
-        if self.provide_label is not None:
-          label = nd.zeros( self.provide_label[0][1] )
-        ba = 0
-        while True:
-          bb = min(ba+batch_size, bag_size)
-          if ba>=bb:
-            break
-          _count = bb-ba
-          #data = nd.zeros( (_count,)+self.data_shape )
-          #_batch = self.data_iter.next()
-          #_data = _batch.data[0].asnumpy()
-          #print(_data.shape)
-          #_label = _batch.label[0].asnumpy()
-          #data[ba:bb,:,:,:] = _data
-          #label[ba:bb] = _label
-          for i in range(ba, bb):
-            #print(ba, bb, self.triplet_cur, i, len(self.triplet_seq))
-            _idx = self.triplet_seq[i+self.triplet_cur]
-            s = self.imgrec.read_idx(_idx)
-            header, img = recordio.unpack(s)
-            img = self.imdecode(img)
-            data[i-ba][:] = self.postprocess_data(img)
-            _label = header.label
-            if not isinstance(_label, numbers.Number):
-              _label = _label[0]
-            if label is not None:
-              label[i-ba][:] = _label
-            tag.append( ( int(_label), _idx) )
-            #idx[i] = _idx
-
-          db = mx.io.DataBatch(data=(data,))
-          self.mx_model.forward(db, is_train=False)
-          net_out = self.mx_model.get_outputs()
-          #print('eval for selecting triplets',ba,bb)
-          #print(net_out)
-          #print(len(net_out))
-          #print(net_out[0].asnumpy())
-          net_out = net_out[0].asnumpy()
-          #print(net_out)
-          #print('net_out', net_out.shape)
-          if embeddings is None:
-            embeddings = np.zeros( (bag_size, net_out.shape[1]))
-          embeddings[ba:bb,:] = net_out
-          ba = bb
-        assert len(tag)==bag_size
-        self.triplet_cur+=bag_size
-        embeddings = sklearn.preprocessing.normalize(embeddings)
-        self.times[1] += self.time_elapsed()
-        self.time_reset()
-        nrof_images_per_class = [1]
-        for i in range(1, bag_size):
-          if tag[i][0]==tag[i-1][0]:
-            nrof_images_per_class[-1]+=1
-          else:
-            nrof_images_per_class.append(1)
-          
-        triplets = self.pick_triplets(embeddings, nrof_images_per_class) # shape=(T,3)
-        print('found triplets', len(triplets))
-        ba = 0
-        while True:
-          bb = ba+self.per_batch_size//3
-          if bb>len(triplets):
-            break
-          _triplets = triplets[ba:bb]
-          for i in range(3):
-            for triplet in _triplets:
-              _pos = triplet[i]
-              _idx = tag[_pos][1]
-              self.seq.append(_idx)
-          ba = bb
-        self.times[2] += self.time_elapsed()
-
-    def hard_mining_reset(self):
-      #import faiss
-      from annoy import AnnoyIndex
-      data = nd.zeros( self.provide_data[0][1] )
-      label = nd.zeros( self.provide_label[0][1] )
-      #label = np.zeros( self.provide_label[0][1] )
-      X = None
-      ba = 0
-      batch_num = 0
-      while ba<len(self.oseq):
-        batch_num+=1
-        if batch_num%10==0:
-          print('loading batch',batch_num, ba)
-        bb = min(ba+self.batch_size, len(self.oseq))
-        _count = bb-ba
-        for i in range(_count):
-          idx = self.oseq[i+ba]
-          s = self.imgrec.read_idx(idx)
-          header, img = recordio.unpack(s)
-          img = self.imdecode(img)
-          data[i][:] = self.postprocess_data(img)
-          label[i][:] = header.label
-        db = mx.io.DataBatch(data=(data,self.data_extra), label=(label,))
-        self.mx_model.forward(db, is_train=False)
-        net_out = self.mx_model.get_outputs()
-        embedding = net_out[0].asnumpy()
-        nembedding = sklearn.preprocessing.normalize(embedding)
-        if _count<self.batch_size:
-          nembedding = nembedding[0:_count,:]
-        if X is None:
-          X = np.zeros( (len(self.id2range), nembedding.shape[1]), dtype=np.float32 )
-        nplabel = label.asnumpy()
-        for i in range(_count):
-          ilabel = int(nplabel[i])
-          #print(ilabel, ilabel.__class__)
-          X[ilabel] += nembedding[i]
-        ba = bb
-      X = sklearn.preprocessing.normalize(X)
-      d = X.shape[1]
-      t = AnnoyIndex(d, metric='euclidean')
-      for i in range(X.shape[0]):
-        t.add_item(i, X[i])
-      print('start to build index')
-      t.build(20)
-      print(X.shape)
-      k = self.per_identities
-      self.seq = []
-      for i in range(X.shape[0]):
-        nnlist = t.get_nns_by_item(i, k)
-        assert nnlist[0]==i
-        for _label in nnlist:
-          assert _label<len(self.id2range)
-          _id = self.header0[0]+_label
-          v = self.id2range[_id]
-          _list = range(*v)
-          if len(_list)<self.images_per_identity:
-            random.shuffle(_list)
-          else:
-            _list = np.random.choice(_list, self.images_per_identity, replace=False)
-          for i in range(self.images_per_identity):
-            _idx = _list[i%len(_list)]
-            self.seq.append(_idx)
-      #faiss_params = [20,5]
-      #quantizer = faiss.IndexFlatL2(d)  # the other index
-      #index = faiss.IndexIVFFlat(quantizer, d, faiss_params[0], faiss.METRIC_L2)
-      #assert not index.is_trained
-      #index.train(X)
-      #index.add(X)
-      #assert index.is_trained
-      #print('trained')
-      #index.nprobe = faiss_params[1]
-      #D, I = index.search(X, k)     # actual search
-      #print(I.shape)
-      #self.seq = []
-      #for i in range(I.shape[0]):
-      #  #assert I[i][0]==i
-      #  for j in range(k):
-      #    _label = I[i][j]
-      #    assert _label<len(self.id2range)
-      #    _id = self.header0[0]+_label
-      #    v = self.id2range[_id]
-      #    _list = range(*v)
-      #    if len(_list)<self.images_per_identity:
-      #      random.shuffle(_list)
-      #    else:
-      #      _list = np.random.choice(_list, self.images_per_identity, replace=False)
-      #    for i in range(self.images_per_identity):
-      #      _idx = _list[i%len(_list)]
-      #      self.seq.append(_idx)
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.images_per_identity>0:
-          if self.triplet_mode:
-            self.select_triplets()
-          elif not self.hard_mining:
-            self.seq = []
-            idlist = []
-            for _id in self.id2range:
-              v = self.id2range[_id]
-              idlist.append((_id,range(*v)))
-            for r in range(self.repeat):
-              if r%10==0:
-                print('repeat', r)
-              if self.shuffle:
-                random.shuffle(idlist)
-              for item in idlist:
-                _id = item[0]
-                _list = item[1]
-                #random.shuffle(_list)
-                if len(_list)<self.images_per_identity:
-                  random.shuffle(_list)
-                else:
-                  _list = np.random.choice(_list, self.images_per_identity, replace=False)
-                for i in range(self.images_per_identity):
-                  _idx = _list[i%len(_list)]
-                  self.seq.append(_idx)
-          else:
-            self.hard_mining_reset()
-          print('seq len', len(self.seq))
-        else:
-          if self.shuffle:
-              random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-      while True:
-        if self.cur >= len(self.seq):
-            raise StopIteration
-        idx = self.seq[self.cur]
-        self.cur += 1
-        s = self.imgrec.read_idx(idx)
-        header, img = recordio.unpack(s)
-        label = header.label
-        if not isinstance(label, numbers.Number):
-          label = label[0]
-        return label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = np.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      random.shuffle(augs)
-      for aug in augs:
-        #print(img.shape)
-        img = aug(img, x)
-        #print(img.shape)
-      return img
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in range(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                _data = self.imdecode(s)
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.cutoff>0:
-                  centerh = random.randint(0, _data.shape[0]-1)
-                  centerw = random.randint(0, _data.shape[1]-1)
-                  half = self.cutoff//2
-                  starth = max(0, centerh-half)
-                  endh = min(_data.shape[0], centerh+half)
-                  startw = max(0, centerw-half)
-                  endw = min(_data.shape[1], centerw+half)
-                  _data = _data.astype('float32')
-                  #print(starth, endh, startw, endw, _data.shape)
-                  _data[starth:endh, startw:endw, :] = 127.5
-                #_npdata = _data.asnumpy()
-                #if landmark is not None:
-                #  _npdata = face_preprocess.preprocess(_npdata, bbox = bbox, landmark=landmark, image_size=self.image_size)
-                #if self.rand_mirror:
-                #  _npdata = self.mirror_aug(_npdata)
-                #if self.mean is not None:
-                #  _npdata = _npdata.astype(np.float32)
-                #  _npdata -= self.mean
-                #  _npdata *= 0.0078125
-                #nimg = np.zeros(_npdata.shape, dtype=np.float32)
-                #nimg[self.patch[1]:self.patch[3],self.patch[0]:self.patch[2],:] = _npdata[self.patch[1]:self.patch[3], self.patch[0]:self.patch[2], :]
-                #_data = mx.nd.array(nimg)
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    if self.provide_label is not None:
-                      batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        #print('next end', batch_size, i)
-        _label = None
-        if self.provide_label is not None:
-          _label = [batch_label]
-        return io.DataBatch([batch_data], _label, batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/recognition/ArcFace/verification.py b/recognition/ArcFace/verification.py
deleted file mode 100644
index 1b1c31c46..000000000
--- a/recognition/ArcFace/verification.py
+++ /dev/null
@@ -1,595 +0,0 @@
-"""Helper for evaluation on the Labeled Faces in the Wild dataset 
-"""
-
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import argparse
-import sys
-import numpy as np
-from scipy import misc
-from sklearn.model_selection import KFold
-from scipy import interpolate
-import sklearn
-import cv2
-import math
-import datetime
-import pickle
-from sklearn.decomposition import PCA
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-class LFold:
-  def __init__(self, n_splits = 2, shuffle = False):
-    self.n_splits = n_splits
-    if self.n_splits>1:
-      self.k_fold = KFold(n_splits = n_splits, shuffle = shuffle)
-
-  def split(self, indices):
-    if self.n_splits>1:
-      return self.k_fold.split(indices)
-    else:
-      return [(indices, indices)]
-
-
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca = 0):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    tprs = np.zeros((nrof_folds,nrof_thresholds))
-    fprs = np.zeros((nrof_folds,nrof_thresholds))
-    accuracy = np.zeros((nrof_folds))
-    indices = np.arange(nrof_pairs)
-    #print('pca', pca)
-    
-    if pca==0:
-      diff = np.subtract(embeddings1, embeddings2)
-      dist = np.sum(np.square(diff),1)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-        #print('train_set', train_set)
-        #print('test_set', test_set)
-        if pca>0:
-          print('doing pca on', fold_idx)
-          embed1_train = embeddings1[train_set]
-          embed2_train = embeddings2[train_set]
-          _embed_train = np.concatenate( (embed1_train, embed2_train), axis=0 )
-          #print(_embed_train.shape)
-          pca_model = PCA(n_components=pca)
-          pca_model.fit(_embed_train)
-          embed1 = pca_model.transform(embeddings1)
-          embed2 = pca_model.transform(embeddings2)
-          embed1 = sklearn.preprocessing.normalize(embed1)
-          embed2 = sklearn.preprocessing.normalize(embed2)
-          #print(embed1.shape, embed2.shape)
-          diff = np.subtract(embed1, embed2)
-          dist = np.sum(np.square(diff),1)
-        
-        # Find the best threshold for the fold
-        acc_train = np.zeros((nrof_thresholds))
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
-        best_threshold_index = np.argmax(acc_train)
-        #print('threshold', thresholds[best_threshold_index])
-        for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-          
-    tpr = np.mean(tprs,0)
-    fpr = np.mean(fprs,0)
-    return tpr, fpr, accuracy
-
-def calculate_accuracy(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    tp = np.sum(np.logical_and(predict_issame, actual_issame))
-    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
-    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
-  
-    tpr = 0 if (tp+fn==0) else float(tp) / float(tp+fn)
-    fpr = 0 if (fp+tn==0) else float(fp) / float(fp+tn)
-    acc = float(tp+tn)/dist.size
-    return tpr, fpr, acc
-
-
-  
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    val = np.zeros(nrof_folds)
-    far = np.zeros(nrof_folds)
-    
-    diff = np.subtract(embeddings1, embeddings2)
-    dist = np.sum(np.square(diff),1)
-    indices = np.arange(nrof_pairs)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
-        # Find the threshold that gives FAR = far_target
-        far_train = np.zeros(nrof_thresholds)
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
-        if np.max(far_train)>=far_target:
-            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
-            threshold = f(far_target)
-        else:
-            threshold = 0.0
-    
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
-  
-    val_mean = np.mean(val)
-    far_mean = np.mean(far)
-    val_std = np.std(val)
-    return val_mean, val_std, far_mean
-
-
-def calculate_val_far(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    n_same = np.sum(actual_issame)
-    n_diff = np.sum(np.logical_not(actual_issame))
-    #print(true_accept, false_accept)
-    #print(n_same, n_diff)
-    val = float(true_accept) / float(n_same)
-    far = float(false_accept) / float(n_diff)
-    return val, far
-
-def evaluate(embeddings, actual_issame, nrof_folds=10, pca = 0):
-    # Calculate evaluation metrics
-    thresholds = np.arange(0, 4, 0.01)
-    embeddings1 = embeddings[0::2]
-    embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), nrof_folds=nrof_folds, pca = pca)
-    thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
-    return tpr, fpr, accuracy, val, val_std, far
-
-def load_bin(path, image_size):
-  try:
-    with open(path, 'rb') as f:
-      bins, issame_list = pickle.load(f) #py2
-  except UnicodeDecodeError as e:
-    with open(path, 'rb') as f:
-      bins, issame_list = pickle.load(f, encoding='bytes') #py3
-  data_list = []
-  for flip in [0,1]:
-    data = nd.empty((len(issame_list)*2, 3, image_size[0], image_size[1]))
-    data_list.append(data)
-  for i in range(len(issame_list)*2):
-    _bin = bins[i]
-    img = mx.image.imdecode(_bin)
-    if img.shape[1]!=image_size[0]:
-      img = mx.image.resize_short(img, image_size[0])
-    img = nd.transpose(img, axes=(2, 0, 1))
-    for flip in [0,1]:
-      if flip==1:
-        img = mx.ndarray.flip(data=img, axis=2)
-      data_list[flip][i][:] = img
-    if i%1000==0:
-      print('loading bin', i)
-  print(data_list[0].shape)
-  return (data_list, issame_list)
-
-def test(data_set, mx_model, batch_size, nfolds=10, data_extra = None, label_shape = None):
-  print('testing verification..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in range( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      #_arg, _aux = model.get_params()
-      #__arg = {}
-      #for k,v in _arg.iteritems():
-      #  __arg[k] = v.as_in_context(_ctx)
-      #_arg = __arg
-      #_arg["data"] = _data.as_in_context(_ctx)
-      #_arg["softmax_label"] = _label.as_in_context(_ctx)
-      #for k,v in _arg.iteritems():
-      #  print(k,v.context)
-      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
-      #exe.forward(is_train=False)
-      #net_out = exe.outputs
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      #print(_embeddings.shape)
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-
-  _xnorm = 0.0
-  _xnorm_cnt = 0
-  for embed in embeddings_list:
-    for i in range(embed.shape[0]):
-      _em = embed[i]
-      _norm=np.linalg.norm(_em)
-      #print(_em.shape, _norm)
-      _xnorm+=_norm
-      _xnorm_cnt+=1
-  _xnorm /= _xnorm_cnt
-
-  embeddings = embeddings_list[0].copy()
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  acc1 = 0.0
-  std1 = 0.0
-  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  #acc1, std1 = np.mean(accuracy), np.std(accuracy)
-
-  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-  #embeddings = np.concatenate(embeddings_list, axis=1)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  print(embeddings.shape)
-  print('infer time', time_consumed)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  return acc1, std1, acc2, std2, _xnorm, embeddings_list
-
-def test_badcase(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
-  print('testing verification badcase..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in range( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  thresholds = np.arange(0, 4, 0.01)
-  actual_issame = np.asarray(issame_list)
-  nrof_folds = 10
-  embeddings1 = embeddings[0::2]
-  embeddings2 = embeddings[1::2]
-  assert(embeddings1.shape[0] == embeddings2.shape[0])
-  assert(embeddings1.shape[1] == embeddings2.shape[1])
-  nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-  nrof_thresholds = len(thresholds)
-  k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-  
-  tprs = np.zeros((nrof_folds,nrof_thresholds))
-  fprs = np.zeros((nrof_folds,nrof_thresholds))
-  accuracy = np.zeros((nrof_folds))
-  indices = np.arange(nrof_pairs)
-  
-  diff = np.subtract(embeddings1, embeddings2)
-  dist = np.sum(np.square(diff),1)
-  data = data_list[0]
-
-  pouts = []
-  nouts = []
-  
-  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-       
-      # Find the best threshold for the fold
-      acc_train = np.zeros((nrof_thresholds))
-      #print(train_set)
-      #print(train_set.__class__)
-      for threshold_idx, threshold in enumerate(thresholds):
-          p2 = dist[train_set]
-          p3 = actual_issame[train_set]
-          _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
-      best_threshold_index = np.argmax(acc_train)
-      for threshold_idx, threshold in enumerate(thresholds):
-          tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-      _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-      best_threshold = thresholds[best_threshold_index]
-      for iid in test_set:
-        ida = iid*2
-        idb = ida+1
-        asame = actual_issame[iid]
-        _dist = dist[iid]
-        violate = _dist - best_threshold
-        if not asame:
-          violate *= -1.0
-        if violate>0.0:
-          imga = data[ida].asnumpy().transpose( (1,2,0) )[...,::-1] #to bgr
-          imgb = data[idb].asnumpy().transpose( (1,2,0) )[...,::-1]
-          #print(imga.shape, imgb.shape, violate, asame, _dist)
-          if asame:
-            pouts.append( (imga, imgb, _dist, best_threshold, ida) )
-          else:
-            nouts.append( (imga, imgb, _dist, best_threshold, ida) )
-
-        
-  tpr = np.mean(tprs,0)
-  fpr = np.mean(fprs,0)
-  acc = np.mean(accuracy)
-  pouts = sorted(pouts, key = lambda x: x[2], reverse=True)
-  nouts = sorted(nouts, key = lambda x: x[2], reverse=False)
-  print(len(pouts), len(nouts))
-  print('acc', acc)
-  gap = 10
-  image_shape = (112,224,3)
-  out_dir = "./badcases"
-  if not os.path.exists(out_dir):
-    os.makedirs(out_dir)
-  if len(nouts)>0:
-    threshold = nouts[0][3]
-  else:
-    threshold = pouts[-1][3]
-  
-  for item in [(pouts, 'positive(false_negative).png'), (nouts, 'negative(false_positive).png')]:
-    cols = 4
-    rows = 8000
-    outs = item[0]
-    if len(outs)==0:
-      continue
-    #if len(outs)==9:
-    #  cols = 3
-    #  rows = 3
-
-    _rows = int(math.ceil(len(outs)/cols))
-    rows = min(rows, _rows)
-    hack = {}
-
-    if name.startswith('cfp') and item[1].startswith('pos'):
-      hack = {0:'manual/238_13.jpg.jpg', 6:'manual/088_14.jpg.jpg', 10:'manual/470_14.jpg.jpg', 25:'manual/238_13.jpg.jpg', 28:'manual/143_11.jpg.jpg'}
-
-    filename = item[1]
-    if len(name)>0:
-      filename = name+"_"+filename
-    filename = os.path.join(out_dir, filename)
-    img = np.zeros( (image_shape[0]*rows+20, image_shape[1]*cols+(cols-1)*gap, 3), dtype=np.uint8 )
-    img[:,:,:] = 255
-    text_color = (0,0,153)
-    text_color = (255,178,102)
-    text_color = (153,255,51)
-    for outi, out in enumerate(outs):
-      row = outi//cols
-      col = outi%cols
-      if row==rows:
-        break
-      imga = out[0].copy()
-      imgb = out[1].copy()
-      if outi in hack:
-        idx = out[4]
-        print('noise idx',idx)
-        aa = hack[outi]
-        imgb = cv2.imread(aa)
-        #if aa==1:
-        #  imgb = cv2.transpose(imgb)
-        #  imgb = cv2.flip(imgb, 1)
-        #elif aa==3:
-        #  imgb = cv2.transpose(imgb)
-        #  imgb = cv2.flip(imgb, 0)
-        #else:
-        #  for ii in range(2):
-        #    imgb = cv2.transpose(imgb)
-        #    imgb = cv2.flip(imgb, 1)
-      dist = out[2]
-      _img = np.concatenate( (imga, imgb), axis=1 )
-      k = "%.3f"%dist
-      #print(k)
-      font = cv2.FONT_HERSHEY_SIMPLEX
-      cv2.putText(_img,k,(80,image_shape[0]//2+7), font, 0.6, text_color, 2)
-      #_filename = filename+"_%d.png"%outi
-      #cv2.imwrite(_filename, _img)
-      img[row*image_shape[0]:(row+1)*image_shape[0], (col*image_shape[1]+gap*col):((col+1)*image_shape[1]+gap*col),:] = _img
-    #threshold = outs[0][3]
-    font = cv2.FONT_HERSHEY_SIMPLEX
-    k = "threshold: %.3f"%threshold
-    cv2.putText(img,k,(img.shape[1]//2-70,img.shape[0]-5), font, 0.6, text_color, 2)
-    cv2.imwrite(filename, img)
-
-def dumpR(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
-  print('dump verification embedding..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in range( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  actual_issame = np.asarray(issame_list)
-  outname = os.path.join('temp.bin')
-  with open(outname, 'wb') as f:
-    pickle.dump((embeddings, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
-if __name__ == '__main__':
-
-  parser = argparse.ArgumentParser(description='do verification')
-  # general
-  parser.add_argument('--data-dir', default='', help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--target', default='lfw,cfp_ff,cfp_fp,agedb_30', help='test targets.')
-  parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--max', default='', type=str, help='')
-  parser.add_argument('--mode', default=0, type=int, help='')
-  parser.add_argument('--nfolds', default=10, type=int, help='')
-  args = parser.parse_args()
-  #sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-  #import face_image
-  #prop = face_image.load_property(args.data_dir)
-  #image_size = prop.image_size
-  image_size = [112,112]
-  print('image_size', image_size)
-  ctx = mx.gpu(args.gpu)
-  nets = []
-  vec = args.model.split(',')
-  prefix = args.model.split(',')[0]
-  epochs = []
-  if len(vec)==1:
-    pdir = os.path.dirname(prefix)
-    for fname in os.listdir(pdir):
-      if not fname.endswith('.params'):
-        continue
-      _file = os.path.join(pdir, fname)
-      if _file.startswith(prefix):
-        epoch = int(fname.split('.')[0].split('-')[1])
-        epochs.append(epoch)
-    epochs = sorted(epochs, reverse=True)
-    if len(args.max)>0:
-      _max = [int(x) for x in args.max.split(',')]
-      assert len(_max)==2
-      if len(epochs)>_max[1]:
-        epochs = epochs[_max[0]:_max[1]]
-
-  else:
-    epochs = [int(x) for x in vec[1].split('|')]
-  print('model number', len(epochs))
-  time0 = datetime.datetime.now()
-  for epoch in epochs:
-    print('loading',prefix, epoch)
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    nets.append(model)
-  time_now = datetime.datetime.now()
-  diff = time_now - time0
-  print('model loading time', diff.total_seconds())
-
-  ver_list = []
-  ver_name_list = []
-  for name in args.target.split(','):
-    path = os.path.join(args.data_dir,name+".bin")
-    if os.path.exists(path):
-      print('loading.. ', name)
-      data_set = load_bin(path, image_size)
-      ver_list.append(data_set)
-      ver_name_list.append(name)
-
-  if args.mode==0:
-    for i in range(len(ver_list)):
-      results = []
-      for model in nets:
-        acc1, std1, acc2, std2, xnorm, embeddings_list = test(ver_list[i], model, args.batch_size, args.nfolds)
-        print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
-        print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
-        print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
-        results.append(acc2)
-      print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
-  elif args.mode==1:
-    model = nets[0]
-    test_badcase(ver_list[0], model, args.batch_size, args.target)
-  else:
-    model = nets[0]
-    dumpR(ver_list[0], model, args.batch_size, args.target)
-
-
diff --git a/recognition/README.md b/recognition/README.md
index 5cda5a3fd..4229e44c3 100644
--- a/recognition/README.md
+++ b/recognition/README.md
@@ -1,24 +1,48 @@
-## Angular Margin Loss for Deep Face Recognition
+## Face Recognition
 
-### Citation
 
-If you find this project useful in your research, please consider to cite the following related papers:
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/logo3.jpg" width="240"/>
+</div>
 
-```
 
-@inproceedings{deng2019arcface,
-  title={Arcface: Additive angular margin loss for deep face recognition},
-  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
-  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
-  pages={4690--4699},
-  year={2019}
-}
+## Introduction
+
+These are the face recognition methods of [InsightFace](https://insightface.ai)
+
+
+<div align="left">
+  <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/refs/heads/master/images/facerecognitionfromvideo.PNG" width="600"/>
+</div>
+
+
+### Datasets
+
+  Please refer to [datasets](_datasets_) page for the details of face recognition datasets used for training and evaluation.
+
+### Evaluation
+
+  Please refer to [evaluation](_evaluation_) page for the details of face recognition evaluation.
+
+
+## Methods
+
+
+Supported methods:
+
+- [x] [ArcFace_torch (CVPR'2019)](arcface_torch)
+- [x] [ArcFace_mxnet (CVPR'2019)](arcface_mxnet)
+- [x] [ArcFace_Paddle (CVPR'2019)](arcface_paddle)
+- [x] [Arcface_oneflow](arcface_oneflow)
+- [x] [SubCenter ArcFace (ECCV'2020)](subcenter_arcface)
+- [x] [VPL (CVPR'2021)](vpl)
+- [x] [PartialFC_torch (CVPR'2022)](arcface_torch)
+- [x] [PartialFC_mxnet (CVPR'2022)](partial_fc)
+- [x] [IDMMD (NeurIPS'2022)](https://github.com/deepinsight/insightface/tree/master/recognition/idmmd)
+
+
+## Contributing
+
+We appreciate all contributions to improve the face recognition model zoo of InsightFace. 
 
-@inproceedings{deng2020subcenter,
-  title={Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces},
-  author={Deng, Jiankang and Guo, Jia and Liu, Tongliang and Gong, Mingming and Zafeiriou, Stefanos},
-  booktitle={Proceedings of the IEEE Conference on European Conference on Computer Vision},
-  year={2020}
-}
 
-```
diff --git a/recognition/SubCenter-ArcFace/README.md b/recognition/SubCenter-ArcFace/README.md
deleted file mode 100644
index 9dbff61ea..000000000
--- a/recognition/SubCenter-ArcFace/README.md
+++ /dev/null
@@ -1,48 +0,0 @@
-
-## Subcenter ArcFace
-
-### 1. Motivation
-
-We introduce one extra hyperparameter (subcenter number `loss_K`) to ArcFace to relax the intra-class compactness constraint. In our experiments, we find ``loss_K=3`` can achieve a good balance between accuracy and robustness.
-
-![difference](https://github.com/deepinsight/insightface/blob/master/resources/subcenterarcfacediff.png)
-
-### 2. Implementation
-
-The training process of Subcenter ArcFace is almost same as [ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace)
-The increased GPU memory consumption can be easily alleviated by our parallel framework.
-
-![framework](https://github.com/deepinsight/insightface/blob/master/resources/subcenterarcfaceframework.png)
-
-### 3. Training Dataset
-
-1. MS1MV0 (The noise rate is around 50%), download link ([baidu drive](https://pan.baidu.com/s/1bSamN5CLiSrxOuGi-Lx7tw), code ``8ql0``)  ([dropbox](https://www.dropbox.com/sh/y2mj25uj440f7bl/AABc7pCJvUvxEcmXs8WYi9Zaa?dl=0))
-
-### 4. Training Steps
-
-1). Train Sub-center ArcFace (``loss_K=3``) on MS1MV0.
-
-2). Drop non-dominant subcenters and high-confident noisy data (``>75 degrees``). 
-
-  ``
-  python drop.py --data <ms1mv0-path> --model <step-1-pretrained-model> --threshold 75 --k 3 --output <ms1mv0-drop75-path>
-  ``
-  
-3). Train ArcFace on the new ``MS1MV0-Drop75`` dataset.
-
-### 5. Pretrained Models and Logs
-  [baidu drive](https://pan.baidu.com/s/1yikOW1Xzm1XIHu0uv0RdRw) code ``3jsh``. [dropbox](https://www.dropbox.com/s/b866xezre3zhuqy/subcenter-arcface-logs.zip?dl=0)
-
-### Citation
-
-If you find *Sub-center ArcFace* useful in your research, please consider to cite the following related papers:
-
-```
-@inproceedings{deng2020subcenter,
-  title={Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces},
-  author={Deng, Jiankang and Guo, Jia and Liu, Tongliang and Gong, Mingming and Zafeiriou, Stefanos},
-  booktitle={Proceedings of the IEEE Conference on European Conference on Computer Vision},
-  year={2020}
-}
-```
-
diff --git a/recognition/SubCenter-ArcFace/drop.py b/recognition/SubCenter-ArcFace/drop.py
deleted file mode 100644
index 069af5d23..000000000
--- a/recognition/SubCenter-ArcFace/drop.py
+++ /dev/null
@@ -1,202 +0,0 @@
-import os
-import shutil
-import datetime
-import sys
-from mxnet import ndarray as nd
-import mxnet as mx
-import random
-import argparse
-import numbers
-import cv2
-import time
-import pickle
-import sklearn
-import sklearn.preprocessing
-from easydict import EasyDict as edict
-import numpy as np
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-from rec_builder import *
-
-def get_embedding(args, imgrec, a, b, image_size, model):
-  ocontents = []
-  for idx in range(a, b):
-    s = imgrec.read_idx(idx)
-    ocontents.append(s)
-  embeddings = None
-  #print(len(ocontents))
-  ba = 0
-  rlabel = -1
-  imgs = []
-  contents = []
-  while True:
-    bb = min(ba+args.batch_size, len(ocontents))
-    if ba>=bb:
-      break
-    _batch_size = bb-ba
-    #_batch_size2 = max(_batch_size, args.ctx_num)
-    _batch_size2 = _batch_size
-    if _batch_size%args.ctx_num!=0:
-        _batch_size2 = ((_batch_size//args.ctx_num)+1) * args.ctx_num
-    data = np.zeros( (_batch_size2,3, image_size[0], image_size[1]) )
-    count = bb-ba
-    ii=0
-    for i in range(ba, bb):
-      header, img = mx.recordio.unpack(ocontents[i])
-      contents.append(img)
-      label = header.label
-      if not isinstance(label, numbers.Number):
-        label = label[0]
-      if rlabel<0:
-        rlabel = int(label)
-      
-      img = mx.image.imdecode(img)
-      rgb = img.asnumpy()
-      bgr = rgb[:,:,::-1]
-      imgs.append(bgr)
-      img = rgb.transpose( (2,0,1) )
-      data[ii] = img
-      ii+=1
-    while ii<_batch_size2:
-      data[ii] = data[0]
-      ii+=1
-    nddata = nd.array(data)
-    db = mx.io.DataBatch(data=(nddata,))
-    model.forward(db, is_train=False)
-    net_out = model.get_outputs()
-    net_out = net_out[0].asnumpy()
-    if embeddings is None:
-      embeddings = np.zeros( (len(ocontents), net_out.shape[1]))
-    embeddings[ba:bb,:] = net_out[0:_batch_size,:]
-    ba = bb
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  return embeddings, rlabel, contents
-
-def main(args):
-  print(args)
-  image_size = (112,112)
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in range(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  args.ctx_num = len(ctx)
-  args.batch_size *= args.ctx_num
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  W = None
-  i = 0
-  while True:
-      key = 'fc7_%d_weight'%i
-      i+=1
-      if key not in arg_params:
-          break
-      _W = arg_params[key].asnumpy()
-      #_W = _W.reshape( (-1, 10, 512) )
-      if W is None:
-          W = _W
-      else:
-          W = np.concatenate( (W, _W), axis=0 )
-  K = args.k
-  W = sklearn.preprocessing.normalize(W)
-  W = W.reshape( (-1, K, 512) )
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  model.bind(data_shapes=[('data', (args.ctx_num, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  print('W:',W.shape)
-  path_imgrec = os.path.join(args.data, 'train.rec')
-  path_imgidx = os.path.join(args.data, 'train.idx')
-  imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-  id_list = []
-  s = imgrec.read_idx(0)
-  header, _ = mx.recordio.unpack(s)
-  assert header.flag>0
-  print('header0 label', header.label)
-  header0 = (int(header.label[0]), int(header.label[1]))
-  #assert(header.flag==1)
-  imgidx = range(1, int(header.label[0]))
-  id2range = {}
-  a, b = int(header.label[0]), int(header.label[1])
-  seq_identity = range(a,b)
-  print(len(seq_identity))
-  image_count = 0
-  pp=0
-  for wid, identity in enumerate(seq_identity):
-    pp+=1
-    s = imgrec.read_idx(identity)
-    header, _ = mx.recordio.unpack(s)
-    contents = []
-    a,b = int(header.label[0]), int(header.label[1])
-    _count = b-a
-    id_list.append( (wid, a, b, _count) )
-    image_count += _count
-  pp = 0
-  if not os.path.exists(args.output):
-    os.makedirs(args.output)
-  ret = np.zeros( (image_count, K+1), dtype=np.float32 )
-  output_dir = args.output
-  builder = SeqRecBuilder(output_dir)
-  print(ret.shape)
-  imid = 0
-  da = datetime.datetime.now()
-  label = 0
-  num_images = 0
-  cos_thresh = np.cos(np.pi*args.threshold / 180.0)
-  for id_item in id_list:
-    wid = id_item[0]
-    pp+=1
-    if pp%40==0:
-      db = datetime.datetime.now()
-      print('processing id', pp, (db-da).total_seconds())
-      da = db
-    x, _, contents = get_embedding(args, imgrec, id_item[1], id_item[2], image_size, model)
-    subcenters = W[wid]
-    K_stat = np.zeros( (K, ), dtype=np.int)
-    for i in range(x.shape[0]):
-        _x = x[i]
-        sim = np.dot(subcenters, _x) # len(sim)==K
-        mc = np.argmax(sim)
-        K_stat[mc] += 1
-    dominant_index = np.argmax(K_stat)
-    dominant_center = subcenters[dominant_index]
-    sim = np.dot(x, dominant_center) 
-    idx = np.where(sim>cos_thresh)[0]
-    num_drop = x.shape[0] - len(idx)
-    if len(idx)==0:
-        continue
-    #print("labelid %d dropped %d, from %d to %d"% (wid, num_drop, x.shape[0], len(idx)))
-    num_images += len(idx)
-    for _idx in idx:
-        c = contents[_idx]
-        builder.add(label, c, is_image=False)
-    label+=1
-  builder.close()
-
-  print('total:', num_images)
-  
-
-
-
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='')
-  # general
-  parser.add_argument('--data', default='/bigdata/faces_ms1m_full', type=str, help='')
-  parser.add_argument('--output', default='/bigdata/ms1m_full_k3drop075', type=str, help='')
-  parser.add_argument('--model', default='../Evaluation/IJB/pretrained_models/r50-arcfacesc-msf-k3z/model,2', help='path to load model.')
-  parser.add_argument('--batch-size', default=16, type=int, help='')
-  parser.add_argument('--threshold', default=75, type=float, help='')
-  parser.add_argument('--k', default=3, type=int, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/recognition/SubCenter-ArcFace/image_iter.py b/recognition/SubCenter-ArcFace/image_iter.py
deleted file mode 100644
index 851e2f9a6..000000000
--- a/recognition/SubCenter-ArcFace/image_iter.py
+++ /dev/null
@@ -1,306 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-from config import config
-
-import mxnet as mx
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-
-logger = logging.getLogger()
-
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0, color_jittering = 0,
-                 images_filter = 0,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        if path_imgrec:
-            logging.info('loading recordio %s...',
-                         path_imgrec)
-            path_imgidx = path_imgrec[0:-4]+".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-            s = self.imgrec.read_idx(0)
-            header, _ = recordio.unpack(s)
-            if header.flag>0:
-              print('header0 label', header.label)
-              self.header0 = (int(header.label[0]), int(header.label[1]))
-              #assert(header.flag==1)
-              self.imgidx = list(range(1, int(header.label[0])))
-              #self.imgidx = []
-              #self.id2range = {}
-              #self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-              #for identity in self.seq_identity:
-              #  s = self.imgrec.read_idx(identity)
-              #  header, _ = recordio.unpack(s)
-              #  a,b = int(header.label[0]), int(header.label[1])
-              #  count = b-a
-              #  if count<images_filter:
-              #    continue
-              #  self.id2range[identity] = (a,b)
-              #  self.imgidx += range(a, b)
-              #print('id2range', len(self.id2range))
-            else:
-              self.imgidx = list(self.imgrec.keys)
-            if shuffle:
-              self.seq = self.imgidx
-              self.oseq = self.imgidx
-              print(len(self.seq))
-            else:
-              self.seq = None
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        self.color_jittering = color_jittering
-        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
-        self.provide_label = [(label_name, (batch_size,))]
-        c, h, w = self.data_shape
-        self.batch_data = np.zeros((batch_size, c, h, w), dtype=np.float32)
-        self.batch_label = np.zeros(self.provide_label[0][1], dtype=np.float32)
-        #print(self.provide_label[0][1])
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.shuffle:
-          random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        """Helper function for reading in next sample."""
-        #set total batch size, for example, 1800, and maximum size for each people, for example 45
-        if self.seq is not None:
-          while True:
-            if self.cur >= len(self.seq):
-                raise StopIteration
-            idx = self.seq[self.cur]
-            self.cur += 1
-            if self.imgrec is not None:
-              s = self.imgrec.read_idx(idx)
-              header, img = recordio.unpack(s)
-              label = header.label
-              if not isinstance(label, numbers.Number):
-                label = label[0]
-              return label, img, None, None
-            else:
-              label, fname, bbox, landmark = self.imglist[idx]
-              return label, self.read_image(fname), bbox, landmark
-        else:
-            s = self.imgrec.read()
-            if s is None:
-                raise StopIteration
-            header, img = recordio.unpack(s)
-            return header.label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = nd.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      #random.shuffle(augs)
-      #for aug in augs:
-      #  #print(img.shape)
-      #  img = aug(img, x)
-      #  #print(img.shape)
-      #return img
-      return self.CJA(img)
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in range(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-    def compress_aug(self, img):
-      from PIL import Image
-      from io import BytesIO
-      buf = BytesIO()
-      img = Image.fromarray(img.asnumpy(), 'RGB')
-      q = random.randint(2, 20)
-      img.save(buf, format='JPEG', quality=q)
-      buf = buf.getvalue()
-      img = Image.open(BytesIO(buf))
-      return nd.array(np.asarray(img, 'float32'))
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                try:
-                    _data = self.imdecode(s)
-                except Exception as e:
-                    logging.debug('Invalid decoding, skipping:  %s', str(e))
-                    continue
-                if _data.shape[0]!=self.data_shape[1]:
-                  _data = mx.image.resize_short(_data, self.data_shape[1])
-                _data = _data.asnumpy().astype(np.float32)
-                if self.rand_mirror and np.random.rand()<0.5:
-                    _data = _data[:,::-1,:]
-                if self.color_jittering>0:
-                  if self.color_jittering>1:
-                    _rd = random.randint(0,1)
-                    if _rd==1:
-                      _data = self.compress_aug(_data)
-                  _data = self.color_aug(_data, 0.125)
-                #if self.nd_mean is not None:
-                #  _data -= self.nd_mean
-                #  _data *= 0.0078125
-                if self.cutoff>0:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    #print('do cutoff aug', self.cutoff)
-                    centerh = random.randint(0, _data.shape[0]-1)
-                    centerw = random.randint(0, _data.shape[1]-1)
-                    half = self.cutoff//2
-                    starth = max(0, centerh-half)
-                    endh = min(_data.shape[0], centerh+half)
-                    startw = max(0, centerw-half)
-                    endw = min(_data.shape[1], centerw+half)
-                    #print(starth, endh, startw, endw, _data.shape)
-                    _data[starth:endh, startw:endw, :] = 128
-                #_data -= 127.5
-                #_data /= 128.0
-                #self.batch_data[i] = self.postprocess_data(_data)
-                _data = _data.transpose( (2,0,1) )
-                self.batch_data[i] = _data
-                self.batch_label[i] = label
-                i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        batch_data = nd.array(self.batch_data)
-        batch_label = nd.array(self.batch_label)
-        return io.DataBatch([batch_data], [batch_label], batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/recognition/SubCenter-ArcFace/parall_module_local_v1.py b/recognition/SubCenter-ArcFace/parall_module_local_v1.py
deleted file mode 100644
index 8a962fb20..000000000
--- a/recognition/SubCenter-ArcFace/parall_module_local_v1.py
+++ /dev/null
@@ -1,570 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-import logging
-import copy
-import time
-import os
-
-import mxnet as mx
-import numpy as np
-from mxnet import context as ctx
-from mxnet.initializer import Uniform
-from mxnet.module.base_module import BaseModule
-from mxnet.module.module import Module
-from mxnet import metric
-from mxnet.model import BatchEndParam
-from mxnet import io
-import mxnet.ndarray as nd
-from config import config
-
-class ParallModule(BaseModule):
-    def __init__(self, symbol, data_names, label_names,
-                 logger=logging, context=ctx.cpu(), work_load_list=None,
-                 asymbol = None,
-                 args = None):
-        super(ParallModule, self).__init__(logger=logger)
-        self._symbol = symbol
-        self._asymbol = asymbol
-        self._data_names = data_names
-        self._label_names = label_names
-        self._context = context
-        self._work_load_list = work_load_list
-        self._num_classes = config.num_classes
-        self._batch_size = args.batch_size
-        self._verbose = args.verbose
-        self._emb_size = config.emb_size
-        self._local_class_start = args.local_class_start
-        self._iter = 0
-
-        self._curr_module = None
-
-        self._num_workers = config.num_workers
-        self._num_ctx = len(self._context)
-        self._ctx_num_classes = args.ctx_num_classes
-        self._nd_cache = {}
-        self._ctx_cpu = mx.cpu()
-        self._ctx_single_gpu = self._context[-1]
-        self._fixed_param_names = None
-        self._curr_module = Module(self._symbol, self._data_names, self._label_names, logger=self.logger,
-                        context=self._context, work_load_list=self._work_load_list,
-                        fixed_param_names=self._fixed_param_names)
-        self._arcface_modules = []
-        self._ctx_class_start = []
-        for i in range(len(self._context)):
-
-          args._ctxid = i
-          _module = Module(self._asymbol(args), self._data_names, self._label_names, logger=self.logger,
-                          context=mx.gpu(i), work_load_list=self._work_load_list,
-                          fixed_param_names=self._fixed_param_names)
-          self._arcface_modules.append(_module)
-          _c = args.local_class_start + i*args.ctx_num_classes
-          self._ctx_class_start.append(_c)
-        self._usekv = False
-        if self._usekv:
-          self._distkv = mx.kvstore.create('dist_sync')
-          self._kvinit = {}
-
-
-    def _reset_bind(self):
-        self.binded = False
-        self._curr_module = None
-
-    @property
-    def data_names(self):
-        return self._data_names
-
-    @property
-    def output_names(self):
-        return self._symbol.list_outputs()
-
-    @property
-    def data_shapes(self):
-        assert self.binded
-        return self._curr_module.data_shapes
-
-    @property
-    def label_shapes(self):
-        assert self.binded
-        return self._curr_module.label_shapes
-
-    @property
-    def output_shapes(self):
-        assert self.binded
-        return self._curr_module.output_shapes
-
-    def get_export_params(self):
-        assert self.binded and self.params_initialized
-        _g, _x = self._curr_module.get_params()
-        g = _g.copy()
-        x = _x.copy()
-        return g, x
-
-    def get_params(self):
-        assert self.binded and self.params_initialized
-        _g, _x = self._curr_module.get_params()
-        g = _g.copy()
-        x = _x.copy()
-        for _module in self._arcface_modules:
-          _g, _x = _module.get_params()
-          ag = _g.copy()
-          ax = _x.copy()
-          g.update(ag)
-          x.update(ax)
-        return g, x
-
-    def set_params(self, arg_params, aux_params, allow_missing=False, force_init=True,
-                   allow_extra=False):
-      g = arg_params
-      x = aux_params
-      #ag = {}
-      #ax = {}
-      rk = []
-      for k in g:
-        v = g[k]
-        if k.startswith('fc7'):
-          p1 = k.find('_')
-          p2 = k.rfind('_')
-          _ctxid = int(k[p1+1:p2])
-          self._arcface_modules[_ctxid].set_params({k:v}, {})
-          rk.append(k)
-      for k in rk:
-        del g[k]
-      self._curr_module.set_params(g, x)
-      #self._arcface_module.set_params(ag, ax)
-
-
-    def init_params(self, initializer=Uniform(0.01), arg_params=None, aux_params=None,
-                    allow_missing=False, force_init=False, allow_extra=False):
-        if self.params_initialized and not force_init:
-            return
-        assert self.binded, 'call bind before initializing the parameters'
-        #TODO init the same weights with all work nodes
-        self._curr_module.init_params(initializer=initializer, arg_params=None,
-                                      aux_params=None, allow_missing=allow_missing,
-                                      force_init=force_init, allow_extra=allow_extra)
-        for _module in self._arcface_modules:
-          #_initializer = initializer
-          _initializer = mx.init.Normal(0.01)
-          _module.init_params(initializer=_initializer, arg_params=None,
-                                        aux_params=None, allow_missing=allow_missing,
-                                        force_init=force_init, allow_extra=allow_extra)
-        self.params_initialized = True
-
-
-    def bind(self, data_shapes, label_shapes=None, for_training=True,
-             inputs_need_grad=False, force_rebind=False, shared_module=None):
-        print('in_bind', self.params_initialized, data_shapes, label_shapes)
-        if self.params_initialized:
-            arg_params, aux_params = self.get_params()
-
-        # force rebinding is typically used when one want to switch from
-        # training to prediction phase.
-        if force_rebind:
-            self._reset_bind()
-
-        if self.binded:
-            self.logger.warning('Already binded, ignoring bind()')
-            return
-
-        assert shared_module is None, 'shared_module for MutableModule is not supported'
-        self.for_training = for_training
-        self.inputs_need_grad = inputs_need_grad
-        self.binded = True
-        self._curr_module.bind(data_shapes, label_shapes, for_training, inputs_need_grad,
-                    force_rebind=False, shared_module=None)
-        _data_shape = data_shapes[0][1]
-        print('_data_shape', _data_shape, label_shapes)
-        for _module in self._arcface_modules:
-          _module.bind([('data', (_data_shape[0]*self._num_workers, self._emb_size))], [('softmax_label', (_data_shape[0]*self._num_workers,))], for_training, True,
-                      force_rebind=False, shared_module=None)
-        if self.params_initialized:
-            self.set_params(arg_params, aux_params)
-
-    def init_optimizer(self, kvstore='local', optimizer='sgd',
-                       optimizer_params=(('learning_rate', 0.01),), force_init=False):
-        assert self.binded and self.params_initialized
-        if self.optimizer_initialized and not force_init:
-            self.logger.warning('optimizer already initialized, ignoring.')
-            return
-
-        self._curr_module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                         force_init=force_init)
-        for _module in self._arcface_modules:
-          _module.init_optimizer(kvstore, optimizer, optimizer_params,
-                                           force_init=force_init)
-        self.optimizer_initialized = True
-
-    def kv_push(self, key, value):
-      #if value.context!=mx.cpu():
-      #  value = value.as_in_context(mx.cpu())
-      if not key in self._kvinit:
-        self._distkv.init(key, nd.zeros_like(value))
-        self._kvinit[key] = 1
-      self._distkv.push(key, value)
-
-    #get fc1 and partial fc7
-    def forward(self, data_batch, is_train=None):
-        #g,x = self.get_params()
-        #print('{fc7_weight[0][0]}', self._iter, g['fc7_0_weight'].asnumpy()[0][0])
-        #print('{pre_fc1_weight[0][0]}', self._iter, g['pre_fc1_weight'].asnumpy()[0][0])
-
-
-        assert self.binded and self.params_initialized
-        self._curr_module.forward(data_batch, is_train=is_train)
-        if is_train:
-          self._iter+=1
-          fc1, label = self._curr_module.get_outputs(merge_multi_context=True)
-          global_fc1 = fc1
-          self.global_label = label.as_in_context(self._ctx_cpu)
-
-
-          for i, _module in enumerate(self._arcface_modules):
-            _label = self.global_label - self._ctx_class_start[i]
-            db_global_fc1 = io.DataBatch([global_fc1], [_label])
-            _module.forward(db_global_fc1) #fc7 with margin
-        #print('forward end')
-
-
-    def get_ndarray(self, context, name, shape):
-      key = "%s_%s"%(name, context)
-      #print(key)
-      if not key in self._nd_cache:
-        v = nd.zeros( shape=shape, ctx = context)
-        self._nd_cache[key] = v
-      else:
-        v = self._nd_cache[key]
-      return v
-
-    def get_ndarray2(self, context, name, arr):
-      key = "%s_%s"%(name, context)
-      #print(key)
-      if not key in self._nd_cache:
-        v = nd.zeros( shape=arr.shape, ctx = context)
-        self._nd_cache[key] = v
-      else:
-        v = self._nd_cache[key]
-      arr.copyto(v)
-      return v
-
-    def backward(self, out_grads=None):
-        #print('in backward')
-        assert self.binded and self.params_initialized
-        #tmp_ctx = self._ctx_cpu
-        tmp_ctx = self._ctx_single_gpu
-        fc7_outs = []
-        ctx_fc7_max = self.get_ndarray(tmp_ctx, 'ctx_fc7_max', (self._batch_size, len(self._context)))
-        #local_fc7_max = nd.zeros( (self.global_label.shape[0],1), ctx=mx.cpu())
-        arcface_module_outputs = []
-        for i, _module in enumerate(self._arcface_modules):
-          #_fc7 = _module.get_outputs(merge_multi_context=True)[0]
-          out = _module.get_outputs(merge_multi_context=True)
-          #print(out[0].shape)
-          #print(out[1].shape)
-          arcface_module_outputs.append(out)
-          _fc7 = out[0]
-          fc7_outs.append(_fc7)
-          _fc7_max = nd.max(_fc7, axis=1).as_in_context(tmp_ctx)
-          ctx_fc7_max[:,i] = _fc7_max
-
-        local_fc7_max = self.get_ndarray(tmp_ctx, 'local_fc7_max', (self._batch_size, 1))
-        nd.max(ctx_fc7_max, axis=1, keepdims=True, out=local_fc7_max)
-        global_fc7_max = local_fc7_max
-        #local_fc7_sum = None
-        local_fc7_sum = self.get_ndarray(tmp_ctx, 'local_fc7_sum', (self._batch_size,1))
-        local_fc7_sum[:,:] = 0.0
-        for i, _module in enumerate(self._arcface_modules):
-          _max = self.get_ndarray2(fc7_outs[i].context, 'fc7_max', global_fc7_max)
-          fc7_outs[i] = nd.broadcast_sub(fc7_outs[i], _max)
-          fc7_outs[i] = nd.exp(fc7_outs[i])
-          _sum = nd.sum(fc7_outs[i], axis=1, keepdims=True).as_in_context(tmp_ctx)
-          local_fc7_sum += _sum
-        global_fc7_sum = local_fc7_sum
-
-        if self._iter%self._verbose==0:
-          #_ctx = self._context[-1]
-          _ctx = self._ctx_cpu
-          _probs = []
-          for i, _module in enumerate(self._arcface_modules):
-            _prob = self.get_ndarray2(_ctx, '_fc7_prob_%d'%i, fc7_outs[i])
-            _probs.append(_prob)
-          fc7_prob = self.get_ndarray(_ctx, 'test_fc7_prob', (self._batch_size, self._ctx_num_classes*len(self._context)))
-          nd.concat(*_probs, dim=1, out=fc7_prob)
-          fc7_pred = nd.argmax(fc7_prob, axis=1)
-          local_label = self.global_label - self._local_class_start
-          #local_label = self.get_ndarray2(_ctx, 'test_label', local_label)
-          _pred = nd.equal(fc7_pred, local_label)
-          print('{fc7_acc}', self._iter, nd.mean(_pred).asnumpy()[0])
-
-
-        #local_fc1_grad = []
-        #fc1_grad_ctx = self._ctx_cpu
-        fc1_grad_ctx = self._ctx_single_gpu
-        local_fc1_grad = self.get_ndarray(fc1_grad_ctx, 'local_fc1_grad', (self._batch_size,self._emb_size))
-        local_fc1_grad[:,:] = 0.0
-        total_eloss = []
-        celoss_verbose = 1000
-        if self._iter%celoss_verbose==0:
-          fc7_celoss = self.get_ndarray(tmp_ctx, 'test_fc7_celoss', (self._batch_size,))
-          fc7_celoss[:] = 0.0
-
-        for i, _module in enumerate(self._arcface_modules):
-          _sum = self.get_ndarray2(fc7_outs[i].context, 'fc7_sum', global_fc7_sum)
-          fc7_outs[i] = nd.broadcast_div(fc7_outs[i], _sum)
-          a = i*self._ctx_num_classes
-          b = (i+1)*self._ctx_num_classes
-          _label = self.global_label - self._ctx_class_start[i]
-          _label = self.get_ndarray2(fc7_outs[i].context, 'label', _label)
-          onehot_label = self.get_ndarray(fc7_outs[i].context, 'label_onehot', (self._batch_size, self._ctx_num_classes))
-          nd.one_hot(_label, depth=self._ctx_num_classes, on_value = 1.0, off_value = 0.0, out=onehot_label)
-          #print(fc7_outs[i].shape, onehot_label.shape)
-
-          if self._iter%celoss_verbose==0:
-            _ce_loss = fc7_outs[i] * onehot_label
-            _ce_loss = nd.sum(_ce_loss, axis=1)
-            fc7_celoss += _ce_loss.as_in_context(tmp_ctx)
-          fc7_outs[i] -= onehot_label
-
-          out = arcface_module_outputs[i]
-          out_grads = [fc7_outs[i]]
-          for j in range(1, len(out)):
-              eloss = out[j]
-              #print('eloss%d:'%j, eloss.shape)
-              #print(out_grads[0].shape)
-              #egrad_shape = (out_grads[0].shape[0], eloss.shape[0])
-              egrad_shape = eloss.shape
-              egrad = self.get_ndarray(fc7_outs[i].context, 'egrad%d'%j, egrad_shape)
-              #egrad[:][:] = 1.0/egrad_shape[0]
-              egrad[:][:] = 1.0
-              out_grads.append(egrad)
-              if self._iter%self._verbose==0:
-                  total_eloss.append(np.mean(eloss.asnumpy()))
-
-          _module.backward(out_grads = out_grads)
-          #ctx_fc1_grad = _module.get_input_grads()[0].as_in_context(mx.cpu())
-          ctx_fc1_grad = self.get_ndarray2(fc1_grad_ctx, 'ctx_fc1_grad_%d'%i, _module.get_input_grads()[0])
-          local_fc1_grad += ctx_fc1_grad
-
-        if self._iter%self._verbose==0 and len(total_eloss)>0:
-          print('{eloss}', self._iter, np.mean(total_eloss))
-        #if self._iter%self._verbose==0:
-        if self._iter%celoss_verbose==0:
-          ce_loss = nd.log(fc7_celoss) * -1.0
-          ce_loss = nd.mean(ce_loss)
-          print('CELOSS,%d,%f'% (self._iter, ce_loss.asnumpy()))
-
-        global_fc1_grad = local_fc1_grad
-        self._curr_module.backward(out_grads = [global_fc1_grad])
-
-
-    def update(self):
-        assert self.binded and self.params_initialized and self.optimizer_initialized
-        self._curr_module.update()
-        for i, _module in enumerate(self._arcface_modules):
-          _module.update()
-        mx.nd.waitall()
-
-
-    def get_outputs(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized
-        return self._curr_module.get_outputs(merge_multi_context=merge_multi_context)
-        #return self._arcface_module.get_outputs(merge_multi_context=merge_multi_context)
-
-    def get_input_grads(self, merge_multi_context=True):
-        assert self.binded and self.params_initialized and self.inputs_need_grad
-        return self._curr_module.get_input_grads(merge_multi_context=merge_multi_context)
-
-    def update_metric(self, eval_metric, labels):
-        assert self.binded and self.params_initialized
-        #self._curr_module.update_metric(eval_metric, labels)
-        #label = labels[0]
-        #print(label.shape)
-        #self._arcface_module.update_metric(eval_metric, labels)
-
-    def install_monitor(self, mon):
-        """ Install monitor on all executors """
-        assert self.binded
-        self._curr_module.install_monitor(mon)
-
-    def forward_backward(self, data_batch):
-        """A convenient function that calls both ``forward`` and ``backward``."""
-        self.forward(data_batch, is_train=True) # get fc1 and partial fc7
-        self.backward()
-
-    def fit(self, train_data, eval_data=None, eval_metric='acc',
-            epoch_end_callback=None, batch_end_callback=None, kvstore='local',
-            optimizer='sgd', optimizer_params=(('learning_rate', 0.01),),
-            eval_end_callback=None,
-            eval_batch_end_callback=None, initializer=Uniform(0.01),
-            arg_params=None, aux_params=None, allow_missing=False,
-            force_rebind=False, force_init=False, begin_epoch=0, num_epoch=None,
-            validation_metric=None, monitor=None, sparse_row_id_fn=None):
-        """Trains the module parameters.
-
-        Checkout `Module Tutorial <http://mxnet.io/tutorials/basic/module.html>`_ to see
-        a end-to-end use-case.
-
-        Parameters
-        ----------
-        train_data : DataIter
-            Train DataIter.
-        eval_data : DataIter
-            If not ``None``, will be used as validation set and the performance
-            after each epoch will be evaluated.
-        eval_metric : str or EvalMetric
-            Defaults to 'accuracy'. The performance measure used to display during training.
-            Other possible predefined metrics are:
-            'ce' (CrossEntropy), 'f1', 'mae', 'mse', 'rmse', 'top_k_accuracy'.
-        epoch_end_callback : function or list of functions
-            Each callback will be called with the current `epoch`, `symbol`, `arg_params`
-            and `aux_params`.
-        batch_end_callback : function or list of function
-            Each callback will be called with a `BatchEndParam`.
-        kvstore : str or KVStore
-            Defaults to 'local'.
-        optimizer : str or Optimizer
-            Defaults to 'sgd'.
-        optimizer_params : dict
-            Defaults to ``(('learning_rate', 0.01),)``. The parameters for
-            the optimizer constructor.
-            The default value is not a dict, just to avoid pylint warning on dangerous
-            default values.
-        eval_end_callback : function or list of function
-            These will be called at the end of each full evaluation, with the metrics over
-            the entire evaluation set.
-        eval_batch_end_callback : function or list of function
-            These will be called at the end of each mini-batch during evaluation.
-        initializer : Initializer
-            The initializer is called to initialize the module parameters when they are
-            not already initialized.
-        arg_params : dict
-            Defaults to ``None``, if not ``None``, should be existing parameters from a trained
-            model or loaded from a checkpoint (previously saved model). In this case,
-            the value here will be used to initialize the module parameters, unless they
-            are already initialized by the user via a call to `init_params` or `fit`.
-            `arg_params` has a higher priority than `initializer`.
-        aux_params : dict
-            Defaults to ``None``. Similar to `arg_params`, except for auxiliary states.
-        allow_missing : bool
-            Defaults to ``False``. Indicates whether to allow missing parameters when `arg_params`
-            and `aux_params` are not ``None``. If this is ``True``, then the missing parameters
-            will be initialized via the `initializer`.
-        force_rebind : bool
-            Defaults to ``False``. Whether to force rebinding the executors if already bound.
-        force_init : bool
-            Defaults to ``False``. Indicates whether to force initialization even if the
-            parameters are already initialized.
-        begin_epoch : int
-            Defaults to 0. Indicates the starting epoch. Usually, if resumed from a
-            checkpoint saved at a previous training phase at epoch N, then this value should be
-            N+1.
-        num_epoch : int
-            Number of epochs for training.
-        sparse_row_id_fn : A callback function
-            The function  takes `data_batch` as an input and returns a dict of
-            str -> NDArray. The resulting dict is used for pulling row_sparse
-            parameters from the kvstore, where the str key is the name of the param,
-            and the value is the row id of the param to pull.
-
-        Examples
-        --------
-        >>> # An example of using fit for training.
-        >>> # Assume training dataIter and validation dataIter are ready
-        >>> # Assume loading a previously checkpointed model
-        >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 3)
-        >>> mod.fit(train_data=train_dataiter, eval_data=val_dataiter, optimizer='sgd',
-        ...     optimizer_params={'learning_rate':0.01, 'momentum': 0.9},
-        ...     arg_params=arg_params, aux_params=aux_params,
-        ...     eval_metric='acc', num_epoch=10, begin_epoch=3)
-        """
-        assert num_epoch is not None, 'please specify number of epochs'
-        assert arg_params is None and aux_params is None
-
-        self.bind(data_shapes=train_data.provide_data, label_shapes=train_data.provide_label,
-                  for_training=True, force_rebind=force_rebind)
-        if monitor is not None:
-            self.install_monitor(monitor)
-        self.init_params(initializer=initializer, arg_params=arg_params, aux_params=aux_params,
-                         allow_missing=allow_missing, force_init=force_init)
-        self.init_optimizer(kvstore=kvstore, optimizer=optimizer,
-                            optimizer_params=optimizer_params)
-
-        if validation_metric is None:
-            validation_metric = eval_metric
-        if not isinstance(eval_metric, metric.EvalMetric):
-            eval_metric = metric.create(eval_metric)
-        epoch_eval_metric = copy.deepcopy(eval_metric)
-
-        ################################################################################
-        # training loop
-        ################################################################################
-        for epoch in range(begin_epoch, num_epoch):
-            tic = time.time()
-            eval_metric.reset()
-            epoch_eval_metric.reset()
-            nbatch = 0
-            data_iter = iter(train_data)
-            end_of_batch = False
-            next_data_batch = next(data_iter)
-            while not end_of_batch:
-                data_batch = next_data_batch
-                if monitor is not None:
-                    monitor.tic()
-                self.forward_backward(data_batch)
-                self.update()
-                assert not isinstance(data_batch, list)
-
-                #if isinstance(data_batch, list):
-                #    #print('XXX')
-                #    self.update_metric(eval_metric,
-                #                       [db.label for db in data_batch],
-                #                       pre_sliced=True)
-                #    self.update_metric(epoch_eval_metric,
-                #                       [db.label for db in data_batch],
-                #                       pre_sliced=True)
-                #else:
-                #    #print('before update metric')
-                #    self.update_metric(eval_metric, data_batch.label)
-                #    self.update_metric(epoch_eval_metric, data_batch.label)
-                #labels = data_batch.label
-                #labels = [self.global_label]
-                #self.update_metric(eval_metric, labels)
-                #self.update_metric(epoch_eval_metric, labels)
-
-                try:
-                    # pre fetch next batch
-                    next_data_batch = next(data_iter)
-                    self.prepare(next_data_batch, sparse_row_id_fn=sparse_row_id_fn)
-                except StopIteration:
-                    end_of_batch = True
-
-                if monitor is not None:
-                    monitor.toc_print()
-
-                #if end_of_batch:
-                #    eval_name_vals = epoch_eval_metric.get_name_value()
-
-                if batch_end_callback is not None:
-                    batch_end_params = BatchEndParam(epoch=epoch, nbatch=nbatch,
-                                                     eval_metric=None,
-                                                     locals=locals())
-                    batch_end_callback(batch_end_params)
-                    #for callback in _as_list(batch_end_callback):
-                    #    callback(batch_end_params)
-                nbatch += 1
-
-            # one epoch of training is finished
-            #for name, val in eval_name_vals:
-            #    self.logger.info('Epoch[%d] Train-%s=%f', epoch, name, val)
-            toc = time.time()
-            self.logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc-tic))
-
-            # sync aux params across devices
-            arg_params, aux_params = self.get_params()
-            self.set_params(arg_params, aux_params)
-
-            # end of 1 epoch, reset the data-iter for another epoch
-            train_data.reset()
-
diff --git a/recognition/SubCenter-ArcFace/sample_config.py b/recognition/SubCenter-ArcFace/sample_config.py
deleted file mode 100644
index 92e9af26a..000000000
--- a/recognition/SubCenter-ArcFace/sample_config.py
+++ /dev/null
@@ -1,225 +0,0 @@
-import numpy as np
-import os
-from easydict import EasyDict as edict
-
-config = edict()
-
-config.bn_mom = 0.9
-config.workspace = 256
-config.emb_size = 512
-config.ckpt_embedding = True
-config.net_se = 0
-config.net_act = 'prelu'
-config.net_unit = 3
-config.net_input = 1
-config.net_blocks = [1,4,6,2]
-config.net_output = 'E'
-config.net_multiplier = 1.0
-config.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-config.ce_loss = True
-config.fc7_lr_mult = 1.0
-config.fc7_wd_mult = 1.0
-config.fc7_no_bias = False
-config.max_steps = 0
-config.data_rand_mirror = True
-config.data_cutoff = False
-config.data_color = 0
-config.data_images_filter = 0
-config.count_flops = True
-config.memonger = False #not work now
-
-config.loss_K = 3
-
-config.flip_noise_prob = 0.0
-
-# network settings
-network = edict()
-
-network.r100 = edict()
-network.r100.net_name = 'fresnet'
-network.r100.num_layers = 100
-
-network.r100fc = edict()
-network.r100fc.net_name = 'fresnet'
-network.r100fc.num_layers = 100
-network.r100fc.net_output = 'FC'
-
-network.r50 = edict()
-network.r50.net_name = 'fresnet'
-network.r50.num_layers = 50
-
-network.r50v1 = edict()
-network.r50v1.net_name = 'fresnet'
-network.r50v1.num_layers = 50
-network.r50v1.net_unit = 1
-
-network.d169 = edict()
-network.d169.net_name = 'fdensenet'
-network.d169.num_layers = 169
-network.d169.per_batch_size = 64
-network.d169.densenet_dropout = 0.0
-
-network.d201 = edict()
-network.d201.net_name = 'fdensenet'
-network.d201.num_layers = 201
-network.d201.per_batch_size = 64
-network.d201.densenet_dropout = 0.0
-
-network.y1 = edict()
-network.y1.net_name = 'fmobilefacenet'
-network.y1.emb_size = 128
-network.y1.net_output = 'GDC'
-
-network.y2 = edict()
-network.y2.net_name = 'fmobilefacenet'
-network.y2.emb_size = 256
-network.y2.net_output = 'GDC'
-network.y2.net_blocks = [2,8,16,4]
-
-network.m1 = edict()
-network.m1.net_name = 'fmobilenet'
-network.m1.emb_size = 256
-network.m1.net_output = 'GDC'
-network.m1.net_multiplier = 1.0
-
-network.m05 = edict()
-network.m05.net_name = 'fmobilenet'
-network.m05.emb_size = 256
-network.m05.net_output = 'GDC'
-network.m05.net_multiplier = 0.5
-
-network.mnas = edict()
-network.mnas.net_name = 'fmnasnet'
-network.mnas.emb_size = 256
-network.mnas.net_output = 'GDC'
-network.mnas.net_multiplier = 1.0
-
-network.mnas05 = edict()
-network.mnas05.net_name = 'fmnasnet'
-network.mnas05.emb_size = 256
-network.mnas05.net_output = 'GDC'
-network.mnas05.net_multiplier = 0.5
-
-network.mnas025 = edict()
-network.mnas025.net_name = 'fmnasnet'
-network.mnas025.emb_size = 256
-network.mnas025.net_output = 'GDC'
-network.mnas025.net_multiplier = 0.25
-
-network.vargfacenet = edict()
-network.vargfacenet.net_name = 'vargfacenet'
-network.vargfacenet.net_multiplier = 1.25
-network.vargfacenet.emb_size = 512
-network.vargfacenet.net_output='J'
-
-# dataset settings
-dataset = edict()
-
-dataset.emore = edict()
-dataset.emore.dataset = 'emore'
-dataset.emore.dataset_path = '../datasets/faces_emore'
-dataset.emore.num_classes = 85742
-dataset.emore.image_shape = (112,112,3)
-dataset.emore.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-
-dataset.retina = edict()
-dataset.retina.dataset = 'retina'
-dataset.retina.dataset_path = '../datasets/ms1m-retinaface-t1'
-dataset.retina.num_classes = 93431
-dataset.retina.image_shape = (112,112,3)
-dataset.retina.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
-
-loss = edict()
-loss.softmax = edict()
-loss.softmax.loss_name = 'softmax'
-
-loss.nsoftmax = edict()
-loss.nsoftmax.loss_name = 'margin_softmax'
-loss.nsoftmax.loss_s = 64.0
-loss.nsoftmax.loss_m1 = 1.0
-loss.nsoftmax.loss_m2 = 0.0
-loss.nsoftmax.loss_m3 = 0.0
-
-loss.arcface = edict()
-loss.arcface.loss_name = 'margin_softmax'
-loss.arcface.loss_s = 64.0
-loss.arcface.loss_m1 = 1.0
-loss.arcface.loss_m2 = 0.5
-loss.arcface.loss_m3 = 0.0
-
-loss.cosface = edict()
-loss.cosface.loss_name = 'margin_softmax'
-loss.cosface.loss_s = 64.0
-loss.cosface.loss_m1 = 1.0
-loss.cosface.loss_m2 = 0.0
-loss.cosface.loss_m3 = 0.35
-
-loss.combined = edict()
-loss.combined.loss_name = 'margin_softmax'
-loss.combined.loss_s = 64.0
-loss.combined.loss_m1 = 1.0
-loss.combined.loss_m2 = 0.3
-loss.combined.loss_m3 = 0.2
-
-loss.triplet = edict()
-loss.triplet.loss_name = 'triplet'
-loss.triplet.images_per_identity = 5
-loss.triplet.triplet_alpha = 0.3
-loss.triplet.triplet_bag_size = 7200
-loss.triplet.triplet_max_ap = 0.0
-loss.triplet.per_batch_size = 60
-loss.triplet.lr = 0.05
-
-loss.atriplet = edict()
-loss.atriplet.loss_name = 'atriplet'
-loss.atriplet.images_per_identity = 5
-loss.atriplet.triplet_alpha = 0.35
-loss.atriplet.triplet_bag_size = 7200
-loss.atriplet.triplet_max_ap = 0.0
-loss.atriplet.per_batch_size = 60
-loss.atriplet.lr = 0.05
-
-# default settings
-default = edict()
-
-# default network
-default.network = 'r100'
-default.pretrained = ''
-default.pretrained_epoch = 1
-# default dataset
-default.dataset = 'emore'
-default.loss = 'arcface'
-default.frequent = 20
-default.verbose = 2000
-default.kvstore = 'device'
-
-default.end_epoch = 10000
-default.lr = 0.1
-default.wd = 0.0005
-default.mom = 0.9
-default.per_batch_size = 128
-default.ckpt = 3
-default.lr_steps = '100000,160000,220000'
-default.models_root = './models'
-
-
-def generate_config(_network, _dataset, _loss):
-    for k, v in loss[_loss].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    for k, v in network[_network].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    for k, v in dataset[_dataset].items():
-      config[k] = v
-      if k in default:
-        default[k] = v
-    config.loss = _loss
-    config.network = _network
-    config.dataset = _dataset
-    config.num_workers = 1
-    if 'DMLC_NUM_WORKER' in os.environ:
-      config.num_workers = int(os.environ['DMLC_NUM_WORKER'])
-
diff --git a/recognition/SubCenter-ArcFace/train_parall.py b/recognition/SubCenter-ArcFace/train_parall.py
deleted file mode 100644
index 20827de0c..000000000
--- a/recognition/SubCenter-ArcFace/train_parall.py
+++ /dev/null
@@ -1,350 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import sklearn
-import numpy as np
-from image_iter import FaceImageIter
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-from config import config, default, generate_config
-import verification
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'symbol'))
-import fresnet
-import fmobilefacenet
-import fmobilenet
-import fmnasnet
-import fdensenet
-import vargfacenet
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train parall face network')
-  # general
-  parser.add_argument('--dataset', default=default.dataset, help='dataset config')
-  parser.add_argument('--network', default=default.network, help='network config')
-  parser.add_argument('--loss', default=default.loss, help='loss config')
-  args, rest = parser.parse_known_args()
-  generate_config(args.network, args.dataset, args.loss)
-  parser.add_argument('--models-root', default=default.models_root, help='root directory to save model.')
-  parser.add_argument('--pretrained', default=default.pretrained, help='pretrained model to load')
-  parser.add_argument('--pretrained-epoch', type=int, default=default.pretrained_epoch, help='pretrained epoch to load')
-  parser.add_argument('--ckpt', type=int, default=default.ckpt, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--verbose', type=int, default=default.verbose, help='do verification testing and model saving every verbose batches')
-  parser.add_argument('--lr', type=float, default=default.lr, help='start learning rate')
-  parser.add_argument('--lr-steps', type=str, default=default.lr_steps, help='steps of lr changing')
-  parser.add_argument('--wd', type=float, default=default.wd, help='weight decay')
-  parser.add_argument('--mom', type=float, default=default.mom, help='momentum')
-  parser.add_argument('--frequent', type=int, default=default.frequent, help='')
-  parser.add_argument('--per-batch-size', type=int, default=default.per_batch_size, help='batch size in each context')
-  parser.add_argument('--kvstore', type=str, default=default.kvstore, help='kvstore setting')
-  parser.add_argument('--worker-id', type=int, default=0, help='worker id for dist training, starts from 0')
-  parser.add_argument('--extra-model-name', type=str, default='', help='extra model name')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol_embedding():
-  embedding = eval(config.net_name).get_symbol()
-  all_label = mx.symbol.Variable('softmax_label')
-  #embedding = mx.symbol.BlockGrad(embedding)
-  all_label = mx.symbol.BlockGrad(all_label)
-  out_list = [embedding, all_label]
-  out = mx.symbol.Group(out_list)
-  return out
-
-def get_symbol_arcface(args):
-  embedding = mx.symbol.Variable('data')
-  all_label = mx.symbol.Variable('softmax_label')
-  gt_label = all_label
-  is_softmax = True
-  #print('call get_sym_arcface with', args, config)
-  if config.loss_name=='margin_softmax':
-    _weight = mx.symbol.Variable("fc7_%d_weight"%args._ctxid, shape=(args.ctx_num_classes*config.loss_K, config.emb_size), 
-        lr_mult=config.fc7_lr_mult, wd_mult=config.fc7_wd_mult)
-    nweight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n_%d'%args._ctxid)
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = nweight, no_bias = True, num_hidden=args.ctx_num_classes*config.loss_K, name='fc7_%d'%args._ctxid)
-    if config.loss_K>1:
-        sim_s3 = mx.symbol.reshape(fc7, (-1, args.ctx_num_classes, config.loss_K))
-
-        sim = mx.symbol.max(sim_s3, axis=2)
-        fc7 = sim
-
-
-
-  if config.loss_m1!=1.0 or config.loss_m2!=0.0 or config.loss_m3!=0.0:
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.ctx_num_classes, on_value = 1.0, off_value = 0.0)
-    if config.loss_m1==1.0 and config.loss_m2==0.0:
-      _one_hot = gt_one_hot*args.margin_b
-      fc7 = fc7-_one_hot
-    else:
-      fc7_onehot = fc7 * gt_one_hot
-      cos_t = fc7_onehot
-      t = mx.sym.arccos(cos_t)
-      if config.loss_m1!=1.0:
-        t = t*config.loss_m1
-      if config.loss_m2!=0.0:
-        t = t+config.loss_m2
-      margin_cos = mx.sym.cos(t)
-      if config.loss_m3!=0.0:
-        margin_cos = margin_cos - config.loss_m3
-      margin_fc7 = margin_cos
-      margin_fc7_onehot = margin_fc7 * gt_one_hot
-      diff = margin_fc7_onehot - fc7_onehot
-      fc7 = fc7+diff
-  fc7 = fc7*config.loss_s
-  out_list = []
-  out_list.append(fc7)
-  if config.loss_name=='softmax': #softmax
-      out_list.append(gt_label)
-  out = mx.symbol.Group(out_list)
-  return out
-
-
-def train_net(args):
-    #_seed = 727
-    #random.seed(_seed)
-    #np.random.seed(_seed)
-    #mx.random.seed(_seed)
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in range(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    if len(args.extra_model_name)==0:
-      prefix = os.path.join(args.models_root, '%s-%s-%s'%(args.network, args.loss, args.dataset), 'model')
-    else:
-      prefix = os.path.join(args.models_root, '%s-%s-%s-%s'%(args.network, args.loss, args.dataset, args.extra_model_name), 'model')
-    prefix_dir = os.path.dirname(prefix)
-    print('prefix', prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    args.ctx_num = len(ctx)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = config.image_shape[2]
-    config.batch_size = args.batch_size
-    config.per_batch_size = args.per_batch_size
-    data_dir = config.dataset_path
-    path_imgrec = None
-    path_imglist = None
-    image_size = config.image_shape[0:2]
-    assert len(image_size)==2
-    assert image_size[0]==image_size[1]
-    print('image_size', image_size)
-    print('num_classes', config.num_classes)
-    path_imgrec = os.path.join(data_dir, "train.rec")
-
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-
-    num_workers = config.num_workers
-    global_num_ctx = num_workers * args.ctx_num
-    if config.num_classes%global_num_ctx==0:
-      args.ctx_num_classes = config.num_classes//global_num_ctx
-    else:
-      args.ctx_num_classes = config.num_classes//global_num_ctx+1
-    args.local_num_classes = args.ctx_num_classes * args.ctx_num
-    args.local_class_start = args.local_num_classes * args.worker_id
-
-    #if len(args.partial)==0:
-    #  local_classes_range = (0, args.num_classes)
-    #else:
-    #  _vec = args.partial.split(',')
-    #  local_classes_range = (int(_vec[0]), int(_vec[1]))
-
-    #args.partial_num_classes = local_classes_range[1] - local_classes_range[0]
-    #args.partial_start = local_classes_range[0]
-
-    print('Called with argument:', args, config)
-    mean = None
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    arg_params = None
-    aux_params = None
-    if len(args.pretrained)==0:
-      esym = get_symbol_embedding()
-      asym = get_symbol_arcface
-    else:
-      assert False
-
-    if config.num_workers==1:
-      from parall_module_local_v1 import ParallModule
-    else:
-      from parall_module_dist import ParallModule
-
-    model = ParallModule(
-        context       = ctx,
-        symbol        = esym,
-        data_names    = ['data'],
-        label_names    = ['softmax_label'],
-        asymbol       = asym,
-        args = args,
-    )
-    val_dataiter = None
-    train_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec,
-        shuffle              = True,
-        rand_mirror          = config.data_rand_mirror,
-        mean                 = mean,
-        cutoff               = config.data_cutoff,
-        color_jittering      = config.data_color,
-        images_filter        = config.data_images_filter,
-    )
-
-
-    
-    if config.net_name=='fresnet' or config.net_name=='fmobilefacenet':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-
-    _rescale = 1.0/args.batch_size
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
-
-
-    ver_list = []
-    ver_name_list = []
-    for name in config.val_targets:
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in range(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, None, None)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in range(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for step in lr_steps:
-        if mbatch==step:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        is_highest = False
-        if len(acc_list)>0:
-          #lfw_score = acc_list[0]
-          #if lfw_score>highest_acc[0]:
-          #  highest_acc[0] = lfw_score
-          #  if lfw_score>=0.998:
-          #    do_save = True
-          score = sum(acc_list)
-          if acc_list[-1]>=highest_acc[-1]:
-            if acc_list[-1]>highest_acc[-1]:
-              is_highest = True
-            else:
-              if score>=highest_acc[0]:
-                is_highest = True
-                highest_acc[0] = score
-            highest_acc[-1] = acc_list[-1]
-            #if lfw_score>=0.99:
-            #  do_save = True
-        if is_highest:
-          do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt==2:
-          do_save = True
-        elif args.ckpt==3:
-          msave = 1
-
-        if do_save:
-          print('saving', msave)
-          if config.ckpt_embedding:
-              arg, aux = model.get_export_params()
-          else:
-              arg, aux = model.get_params()
-          all_layers = model.symbol.get_internals()
-          _sym = all_layers['fc1_output']
-          mx.model.save_checkpoint(prefix, msave, _sym, arg, aux)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if config.max_steps>0 and mbatch>config.max_steps:
-        sys.exit(0)
-
-    epoch_cb = None
-    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = 999999,
-        eval_data          = val_dataiter,
-        #eval_metric        = eval_metrics,
-        kvstore            = args.kvstore,
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/recognition/_datasets_/README.md b/recognition/_datasets_/README.md
new file mode 100644
index 000000000..d6fec7173
--- /dev/null
+++ b/recognition/_datasets_/README.md
@@ -0,0 +1,150 @@
+# Face Recognition Datasets
+
+## Training Datasets (Updating)
+
+### CASIA-Webface (10K ids/0.5M images) [1]
+
+[Baidu](https://pan.baidu.com/s/1AfHdPsxJZBD8kBJeIhmq1w)
+
+[GDrive](https://drive.google.com/file/d/1KxNCrXzln0lal3N4JiYl9cFOIhT78y1l/view?usp=sharing)
+
+### CelebA (10K ids/0.2M images) [2]
+
+### UMDFace (8K ids/0.37M images) [3]
+
+[Baidu](https://pan.baidu.com/s/1aGutJwNWpV-lA0f_7eNsGQ)
+
+[GDrive](https://drive.google.com/file/d/1azhEHoJjVmifuzBVKJwl-sDbLZ-Wzp4O/view?usp=sharing)
+
+### VGG2 (9K ids/3.31M images) [4]
+
+[Baidu](https://pan.baidu.com/s/1c3KeLzy)
+
+[GDrive](https://drive.google.com/file/d/1dyVQ7X3d28eAcjV3s3o0MT-HyODp_v3R/view?usp=sharing)
+
+### MS1M-IBUG (85K ids/3.8M images) [5,6]
+
+[Baidu](https://pan.baidu.com/s/1nxmSCch)
+
+### MS1M-ArcFace (85K ids/5.8M images) [5,7] 
+
+[Baidu](https://pan.baidu.com/s/1S6LJZGdqcZRle1vlcMzHOQ)
+
+[GDrive](https://drive.google.com/file/d/1SXS4-Am3bsKSK615qbYdbA_FMVh3sAvR/view?usp=sharing)
+
+### MS1M-RetinaFace
+
+[Baidu](https://pan.baidu.com/s/1RBnaW88PC6cKqtYwgfVX8Q) (code:8eb3)
+
+[GDrive](https://drive.google.com/file/d/1JgmzL9OLTqDAZE86pBgETtSQL4USKTFy/view?usp=sharing)
+
+### Asian-Celeb (94K ids/2.8M images)[8]
+
+[Baidu](https://pan.baidu.com/s/12wSgofDy1flFf6lOyAxJRg)
+
+### Glint360K (360K ids/17M images)[17]
+
+[Baidu](https://pan.baidu.com/s/1GsYqTTt7_Dn8BfxxsLFN0w) (code:o3az)    
+magnet uri: `magnet:?xt=urn:btih:E5F46EE502B9E76DA8CC3A0E4F7C17E4000C7B1E&dn=glint360k`
+
+### Glint-Mini (91K ids/5.2M images)[17]
+[Baidu](https://pan.baidu.com/s/10IzEyP-Z9dWFcxxj9jdJpQ) (code:10m5)
+
+### DeepGlint (181K ids/6.75M images) [8] 
+
+[baidu](https://pan.baidu.com/s/1yApUbklBgRgOyOV4o3J8Eg)
+
+### WebFace260M [18]
+
+[Link](https://www.face-benchmark.org/download.html)
+
+
+### IMDB-Face (59K ids/1.7M images) [9]
+
+### Celeb500k (500K ids/50M images) [10]
+
+### MegaFace(train) (672K ids/4.7M images) [11]
+
+[Baidu](https://pan.baidu.com/s/1uy366DjUiGc3AvhuamRLyw) (code:5f8m)
+
+[GDrive](https://drive.google.com/file/d/1O4FxijSXoEIe6fLfOocqF4VFMh5B4d89/view?usp=sharing)
+
+### DigiFace-1M (110K ids/1.22M images) [19]
+
+[Website](https://microsoft.github.io/DigiFace1M/)
+
+[Github](https://github.com/microsoft/DigiFace1M)
+
+## Validation Datasets
+
+### CFP-FP (500 ids/7K images/7K pairs)[12]
+
+### AgeDB-30 (570 ids/12,240 images/6K pairs)[13,6]
+
+### LFW (5749 ids/13233 images/6K pairs)[14]
+
+### CALFW (5749 ids/13233 images/6K pairs)[15]
+
+### CPLFW (5749 ids/13233 images/6K pairs)[16]
+
+## Image Test Datasets
+
+### MegaFace
+
+testsuite: [GDrive](https://drive.google.com/file/d/1KBwp0U9oZgZj7SYDXRxUnnH7Lwvd9XMy/view?usp=sharing)
+
+### IJB (IJB-B, IJB-C)
+
+testsuite: [GDrive](https://drive.google.com/file/d/1aC4zf2Bn0xCVH_ZtEuQipR2JvRb1bf8o/view?usp=sharing)
+
+### TrillionPairs
+
+### NIST
+
+## Video Test Datasets
+
+### YTF
+
+### IQIYI
+
+## Reference
+
+[1] Dong Yi, Zhen Lei, Shengcai Liao, Stan Z. Li. Learning Face Representation from Scratch. arXiv:1411.7923, 2014.
+
+[2] Ziwei Liu, Ping Luo, Xiaogang Wang, Xiaoou Tang. Deep Learning Face Attributes in the Wild, ICCV, 2015.
+
+[3] Bansal Ankan, Nanduri Anirudh, Castillo Carlos D, Ranjan Rajeev, Chellappa, Rama. UMDFaces: An Annotated Face Dataset for Training Deep Networks, arXiv:1611.01484v2, 2016.
+
+[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. FG, 2018.
+
+[5] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. ECCV, 2016.
+
+[6] Jiankang Deng, Yuxiang Zhou, Stefanos Zafeiriou. Marginal loss for deep face recognition, CVPRW, 2017.
+
+[7] Jiankang Deng, Jia Guo, Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition, arXiv:1801.07698, 2018.
+
+[8] [http://trillionpairs.deepglint.com/](http://trillionpairs.deepglint.com/)
+
+[9] Wang Fei, Chen Liren, Li Cheng, Huang Shiyao, Chen Yanjie, Qian Chen, Loy, Chen Change. The Devil of Face Recognition is in the Noise, ECCV, 2018.
+
+[10] Cao Jiajiong, Li Yingming, Zhang Zhongfei, Celeb-500K: A Large Training Dataset for Face Recognition, ICIP, 2018.
+
+[11] Nech Aaron, Kemelmacher-Shlizerman Ira, Level Playing Field For Million Scale Face Recognition, CVPR, 2017.
+ 
+[12] Sengupta Soumyadip, Chen Jun-Cheng, Castillo Carlos, Patel Vishal M, Chellappa Rama, Jacobs David W,
+  Frontal to profile face verification in the wild, WACV, 2016.
+
+[13] Moschoglou, Stylianos and Papaioannou, Athanasios and Sagonas, Christos and Deng, Jiankang and Kotsia, Irene and Zafeiriou, Stefanos, Agedb: the first manually collected, in-the-wild age database, CVPRW, 2017.
+
+[14] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller.
+Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments, 2007.
+
+[15] Zheng Tianyue, Deng Weihong, Hu Jiani, Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments, arXiv:1708.08197, 2017.
+
+[16] Zheng, Tianyue, and Weihong Deng. Cross-Pose LFW: A Database for Studying Cross-Pose Face Recognition in Unconstrained Environments, 2018.
+
+[17] An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and Zhang, Debing and Fu Ying. Partial FC: Training 10 Million Identities on a Single Machine, arxiv:2010.05222, 2020.
+
+[18] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, Jie Zhou. WebFace260M: A Benchmark Unveiling the Power of Million-scale Deep Face Recognition
+
+[19] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, Jingjing Shen. DigiFace-1M: 1 Million Digital Face Images for Face Recognition. WACV 2023
\ No newline at end of file
diff --git a/recognition/_evaluation_/ijb/README.md b/recognition/_evaluation_/ijb/README.md
new file mode 100644
index 000000000..79ccd2d83
--- /dev/null
+++ b/recognition/_evaluation_/ijb/README.md
@@ -0,0 +1,36 @@
+To reproduce the figures and tables in the notebook, please download everything (model, code, data and meta info) from here:
+[GDrive](https://drive.google.com/file/d/1aC4zf2Bn0xCVH_ZtEuQipR2JvRb1bf8o/view?usp=sharing)
+or
+[Baidu Cloud](https://pan.baidu.com/s/1oer0p4_mcOrs4cfdeWfbFg)
+
+Updated Meta data (1:1 and 1:N):
+
+   [Baidu Cloud](https://pan.baidu.com/s/1x-ytzg4zkCTOTtklUgAhfg) (code:7g8o) ;
+   [GDrive](https://drive.google.com/file/d/1MXzrU_zUESSx_242pRUnVvW_wDzfU8Ky/view?usp=sharing)
+
+Please apply for the IJB-B and IJB-C by yourself and strictly follow their distribution licenses.
+
+## Aknowledgement
+
+Great thanks for Weidi Xie's instruction [2,3,4,5] to evaluate ArcFace [1] on IJB-B[6] and IJB-C[7] (1:1 protocol).
+
+Great thanks for Yuge Huang's code [8] to evaluate ArcFace [1] on IJB-B[6] and IJB-C[7] (1:N protocol). 
+
+## Reference
+
+[1] Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition[J]. arXiv:1801.07698, 2018.
+
+[2] https://github.com/ox-vgg/vgg_face2.
+
+[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. FG, 2018.
+
+[4] Weidi Xie, Andrew Zisserman. Multicolumn Networks for Face Recognition. BMVC 2018.
+
+[5] Weidi Xie, Li Shen, Andrew Zisserman. Comparator Networks. ECCV, 2018.
+
+[6] Whitelam, Cameron, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn C. Adams, Tim Miller, Nathan D. Kalka et al. IARPA Janus Benchmark-B Face Dataset. CVPR Workshops, 2017.
+
+[7] Maze, Brianna, Jocelyn Adams, James A. Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K. Jain et al. IARPA Janus Benchmark–C: Face Dataset and Protocol. ICB, 2018.
+
+[8] Yuge Huang, Pengcheng Shen, Ying Tai, Shaoxin Li, Xiaoming Liu, Jilin Li, Feiyue Huang, Rongrong Ji. Distribution Distillation Loss: Generic Approach for Improving Face Recognition from Hard Samples. arXiv:2002.03662.
+
diff --git a/recognition/_evaluation_/ijb/example.sh b/recognition/_evaluation_/ijb/example.sh
new file mode 100755
index 000000000..63f82698c
--- /dev/null
+++ b/recognition/_evaluation_/ijb/example.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+
+python -u ijb_11.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 1 --gpu 0 --target IJBC --job arcface > ijbc_11.log 2>&1 &
+
+python -u ijb_1n.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 1 --gpu 0 --target IJBB --job arcface > ijbb_1n.log 2>&1 &
+
diff --git a/recognition/_evaluation_/ijb/ijb_11.py b/recognition/_evaluation_/ijb/ijb_11.py
new file mode 100644
index 000000000..53b332697
--- /dev/null
+++ b/recognition/_evaluation_/ijb/ijb_11.py
@@ -0,0 +1,380 @@
+# coding: utf-8
+
+import os
+import numpy as np
+#import cPickle
+import pickle
+import pandas as pd
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import timeit
+import sklearn
+import argparse
+from sklearn.metrics import roc_curve, auc
+from sklearn import preprocessing
+import cv2
+import sys
+import glob
+sys.path.append('./recognition')
+from embedding import Embedding
+from menpo.visualize import print_progress
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from pathlib import Path
+import warnings
+warnings.filterwarnings("ignore")
+
+parser = argparse.ArgumentParser(description='do ijb test')
+# general
+parser.add_argument('--model-prefix', default='', help='path to load model.')
+parser.add_argument('--model-epoch', default=1, type=int, help='')
+parser.add_argument('--gpu', default=7, type=int, help='gpu id')
+parser.add_argument('--batch-size', default=32, type=int, help='')
+parser.add_argument('--job', default='insightface', type=str, help='job name')
+parser.add_argument('--target',
+                    default='IJBC',
+                    type=str,
+                    help='target, set to IJBC or IJBB')
+args = parser.parse_args()
+
+target = args.target
+model_path = args.model_prefix
+gpu_id = args.gpu
+epoch = args.model_epoch
+use_norm_score = True  # if Ture, TestMode(N1)
+use_detector_score = True  # if Ture, TestMode(D1)
+use_flip_test = True  # if Ture, TestMode(F1)
+job = args.job
+
+
+def read_template_media_list(path):
+    #ijb_meta = np.loadtxt(path, dtype=str)
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+# In[ ]:
+
+
+def read_template_pair_list(path):
+    #pairs = np.loadtxt(path, dtype=str)
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    #print(pairs.shape)
+    #print(pairs[:, 0].astype(np.int))
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+# In[ ]:
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# In[ ]:
+
+
+def get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id):
+    img_list = open(img_list_path)
+    embedding = Embedding(model_path, epoch, gpu_id)
+    files = img_list.readlines()
+    print('files:', len(files))
+    faceness_scores = []
+    img_feats = []
+    for img_index, each_line in enumerate(files):
+        if img_index % 500 == 0:
+            print('processing', img_index)
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        img_feats.append(embedding.get(img, lmk))
+        faceness_scores.append(name_lmk_score[-1])
+    img_feats = np.array(img_feats).astype(np.float32)
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+
+    #img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
+    #faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
+    return img_feats, faceness_scores
+
+
+# In[ ]:
+
+
+def image2template_feature(img_feats=None, templates=None, medias=None):
+    # ==========================================================
+    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
+    # 2. compute media feature.
+    # 3. compute template feature.
+    # ==========================================================
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t, ) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias,
+                                                       return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m, ) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)
+                ]
+        media_norm_feats = np.array(media_norm_feats)
+        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    #template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
+    template_norm_feats = sklearn.preprocessing.normalize(template_feats)
+    #print(template_norm_feats.shape)
+    return template_norm_feats, unique_templates
+
+
+# In[ ]:
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    # ==========================================================
+    #         Compute set-to-set Similarity Score.
+    # ==========================================================
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+
+    score = np.zeros((len(p1), ))  # save cosine distance between pairs
+
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+# In[ ]:
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1), ))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def read_score(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# # Step1: Load Meta Data
+
+# In[ ]:
+
+assert target == 'IJBC' or target == 'IJBB'
+
+# =============================================================
+# load image and template relationships for template feature embedding
+# tid --> template id,  mid --> media id
+# format:
+#           image_name tid mid
+# =============================================================
+start = timeit.default_timer()
+templates, medias = read_template_media_list(
+    os.path.join('%s/meta' % target, '%s_face_tid_mid.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+
+# =============================================================
+# load template pairs for template-to-template verification
+# tid : template id,  label : 1/0
+# format:
+#           tid_1 tid_2 label
+# =============================================================
+start = timeit.default_timer()
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % target,
+                 '%s_template_pair_label.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 2: Get Image Features
+
+# In[ ]:
+
+# =============================================================
+# load image features
+# format:
+#           img_feats: [image_num x feats_dim] (227630, 512)
+# =============================================================
+start = timeit.default_timer()
+img_path = './%s/loose_crop' % target
+img_list_path = './%s/meta/%s_name_5pts_score.txt' % (target, target.lower())
+img_feats, faceness_scores = get_image_feature(img_path, img_list_path,
+                                               model_path, epoch, gpu_id)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0],
+                                          img_feats.shape[1]))
+
+# # Step3: Get Template Features
+
+# In[ ]:
+
+# =============================================================
+# compute template features from image features.
+# =============================================================
+start = timeit.default_timer()
+# ==========================================================
+# Norm feature before aggregation into template feature?
+# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
+# ==========================================================
+# 1. FaceScore （Feature Norm）
+# 2. FaceScore （Detector）
+
+if use_flip_test:
+    # concat --- F1
+    #img_input_feats = img_feats
+    # add --- F2
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] //
+                                2] + img_feats[:, img_feats.shape[1] // 2:]
+else:
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+if use_norm_score:
+    img_input_feats = img_input_feats
+else:
+    # normalise features to remove norm information
+    img_input_feats = img_input_feats / np.sqrt(
+        np.sum(img_input_feats**2, -1, keepdims=True))
+
+if use_detector_score:
+    print(img_input_feats.shape, faceness_scores.shape)
+    #img_input_feats = img_input_feats * np.matlib.repmat(faceness_scores[:,np.newaxis], 1, img_input_feats.shape[1])
+    img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+else:
+    img_input_feats = img_input_feats
+
+template_norm_feats, unique_templates = image2template_feature(
+    img_input_feats, templates, medias)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 4: Get Template Similarity Scores
+
+# In[ ]:
+
+# =============================================================
+# compute verification scores between template pairs.
+# =============================================================
+start = timeit.default_timer()
+score = verification(template_norm_feats, unique_templates, p1, p2)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+
+save_path = './%s_result' % target
+
+if not os.path.exists(save_path):
+    os.makedirs(save_path)
+
+score_save_file = os.path.join(save_path, "%s.npy" % job)
+np.save(score_save_file, score)
+
+# # Step 5: Get ROC Curves and TPR@FPR Table
+
+# In[ ]:
+
+files = [score_save_file]
+methods = []
+scores = []
+for file in files:
+    methods.append(Path(file).stem)
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+#x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]
+x_labels = [10**-6, 10**-5, 10**-4, 10**-3, 10**-2, 10**-1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append("%s-%s" % (method, target))
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        #tpr_fpr_row.append('%.4f' % tpr[min_index])
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10**-6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+#plt.show()
+fig.savefig(os.path.join(save_path, '%s.pdf' % job))
+print(tpr_fpr_table)
diff --git a/recognition/_evaluation_/ijb/ijb_1n.py b/recognition/_evaluation_/ijb/ijb_1n.py
new file mode 100644
index 000000000..85ea4051f
--- /dev/null
+++ b/recognition/_evaluation_/ijb/ijb_1n.py
@@ -0,0 +1,366 @@
+#!/usr/bin/env python
+# coding: utf-8
+import os
+import numpy as np
+import timeit
+import sklearn
+import cv2
+import sys
+import argparse
+import glob
+import numpy.matlib
+import heapq
+import math
+from datetime import datetime as dt
+
+from sklearn import preprocessing
+sys.path.append('./recognition')
+from embedding import Embedding
+from menpo.visualize import print_progress
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+
+
+def read_template_subject_id_list(path):
+    ijb_meta = np.loadtxt(path, dtype=str, skiprows=1, delimiter=',')
+    templates = ijb_meta[:, 0].astype(np.int)
+    subject_ids = ijb_meta[:, 1].astype(np.int)
+    return templates, subject_ids
+
+
+def read_template_media_list(path):
+    ijb_meta = np.loadtxt(path, dtype=str)
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+def read_template_pair_list(path):
+    pairs = np.loadtxt(path, dtype=str)
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+#def get_image_feature(feature_path, faceness_path):
+#    img_feats = np.loadtxt(feature_path)
+#    faceness_scores = np.loadtxt(faceness_path)
+#    return img_feats, faceness_scores
+def get_image_feature(img_path, img_list_path, model_path, epoch, gpu_id):
+    img_list = open(img_list_path)
+    embedding = Embedding(model_path, epoch, gpu_id)
+    files = img_list.readlines()
+    print('files:', len(files))
+    faceness_scores = []
+    img_feats = []
+    for img_index, each_line in enumerate(files):
+        if img_index % 500 == 0:
+            print('processing', img_index)
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        img_feats.append(embedding.get(img, lmk))
+        faceness_scores.append(name_lmk_score[-1])
+    img_feats = np.array(img_feats).astype(np.float32)
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+
+    #img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
+    #faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
+    return img_feats, faceness_scores
+
+
+def image2template_feature(img_feats=None,
+                           templates=None,
+                           medias=None,
+                           choose_templates=None,
+                           choose_ids=None):
+    # ==========================================================
+    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
+    # 2. compute media feature.
+    # 3. compute template feature.
+    # ==========================================================
+    unique_templates, indices = np.unique(choose_templates, return_index=True)
+    unique_subjectids = choose_ids[indices]
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t, ) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias,
+                                                       return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m, ) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], 0, keepdims=True)
+                ]
+        media_norm_feats = np.array(media_norm_feats)
+        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
+        template_feats[count_template] = np.sum(media_norm_feats, 0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    template_norm_feats = template_feats / np.sqrt(
+        np.sum(template_feats**2, -1, keepdims=True))
+    return template_norm_feats, unique_templates, unique_subjectids
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    # ==========================================================
+    #         Compute set-to-set Similarity Score.
+    # ==========================================================
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+
+    score = np.zeros((len(p1), ))  # save cosine distance between pairs
+
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def read_score(path):
+    with open(path, 'rb') as fid:
+        img_feats = cPickle.load(fid)
+    return img_feats
+
+
+def evaluation(query_feats, gallery_feats, mask):
+    Fars = [0.01, 0.1]
+    print(query_feats.shape)
+    print(gallery_feats.shape)
+
+    query_num = query_feats.shape[0]
+    gallery_num = gallery_feats.shape[0]
+
+    similarity = np.dot(query_feats, gallery_feats.T)
+    print('similarity shape', similarity.shape)
+    top_inds = np.argsort(-similarity)
+    print(top_inds.shape)
+
+    # calculate top1
+    correct_num = 0
+    for i in range(query_num):
+        j = top_inds[i, 0]
+        if j == mask[i]:
+            correct_num += 1
+    print("top1 = {}".format(correct_num / query_num))
+    # calculate top5
+    correct_num = 0
+    for i in range(query_num):
+        j = top_inds[i, 0:5]
+        if mask[i] in j:
+            correct_num += 1
+    print("top5 = {}".format(correct_num / query_num))
+    # calculate 10
+    correct_num = 0
+    for i in range(query_num):
+        j = top_inds[i, 0:10]
+        if mask[i] in j:
+            correct_num += 1
+    print("top10 = {}".format(correct_num / query_num))
+
+    neg_pair_num = query_num * gallery_num - query_num
+    print(neg_pair_num)
+    required_topk = [math.ceil(query_num * x) for x in Fars]
+    top_sims = similarity
+    # calculate fars and tprs
+    pos_sims = []
+    for i in range(query_num):
+        gt = mask[i]
+        pos_sims.append(top_sims[i, gt])
+        top_sims[i, gt] = -2.0
+
+    pos_sims = np.array(pos_sims)
+    print(pos_sims.shape)
+    neg_sims = top_sims[np.where(top_sims > -2.0)]
+    print("neg_sims num = {}".format(len(neg_sims)))
+    neg_sims = heapq.nlargest(max(required_topk), neg_sims)  # heap sort
+    print("after sorting , neg_sims num = {}".format(len(neg_sims)))
+    for far, pos in zip(Fars, required_topk):
+        th = neg_sims[pos - 1]
+        recall = np.sum(pos_sims > th) / query_num
+        print("far = {:.10f} pr = {:.10f} th = {:.10f}".format(
+            far, recall, th))
+
+
+def gen_mask(query_ids, reg_ids):
+    mask = []
+    for query_id in query_ids:
+        pos = [i for i, x in enumerate(reg_ids) if query_id == x]
+        if len(pos) != 1:
+            raise RuntimeError(
+                "RegIdsError with id = {}， duplicate = {} ".format(
+                    query_id, len(pos)))
+        mask.append(pos[0])
+    return mask
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='do ijb 1n test')
+    # general
+    parser.add_argument('--model-prefix',
+                        default='',
+                        help='path to load model.')
+    parser.add_argument('--model-epoch', default=1, type=int, help='')
+    parser.add_argument('--gpu', default=7, type=int, help='gpu id')
+    parser.add_argument('--batch-size', default=32, type=int, help='')
+    parser.add_argument('--job',
+                        default='insightface',
+                        type=str,
+                        help='job name')
+    parser.add_argument('--target',
+                        default='IJBC',
+                        type=str,
+                        help='target, set to IJBC or IJBB')
+    args = parser.parse_args()
+    target = args.target
+    model_path = args.model_prefix
+    gpu_id = args.gpu
+    epoch = args.model_epoch
+    meta_dir = "%s/meta" % args.target  #meta root dir
+    if target == 'IJBC':
+        gallery_s1_record = "%s_1N_gallery_G1.csv" % (args.target.lower())
+        gallery_s2_record = "%s_1N_gallery_G2.csv" % (args.target.lower())
+    else:
+        gallery_s1_record = "%s_1N_gallery_S1.csv" % (args.target.lower())
+        gallery_s2_record = "%s_1N_gallery_S2.csv" % (args.target.lower())
+    gallery_s1_templates, gallery_s1_subject_ids = read_template_subject_id_list(
+        os.path.join(meta_dir, gallery_s1_record))
+    print(gallery_s1_templates.shape, gallery_s1_subject_ids.shape)
+
+    gallery_s2_templates, gallery_s2_subject_ids = read_template_subject_id_list(
+        os.path.join(meta_dir, gallery_s2_record))
+    print(gallery_s2_templates.shape, gallery_s2_templates.shape)
+
+    gallery_templates = np.concatenate(
+        [gallery_s1_templates, gallery_s2_templates])
+    gallery_subject_ids = np.concatenate(
+        [gallery_s1_subject_ids, gallery_s2_subject_ids])
+    print(gallery_templates.shape, gallery_subject_ids.shape)
+
+    media_record = "%s_face_tid_mid.txt" % args.target.lower()
+    total_templates, total_medias = read_template_media_list(
+        os.path.join(meta_dir, media_record))
+    print("total_templates", total_templates.shape, total_medias.shape)
+    #load image features
+    start = timeit.default_timer()
+    feature_path = ''  #feature path
+    face_path = ''  #face path
+    img_path = './%s/loose_crop' % target
+    img_list_path = './%s/meta/%s_name_5pts_score.txt' % (target,
+                                                          target.lower())
+    #img_feats, faceness_scores = get_image_feature(feature_path, face_path)
+    img_feats, faceness_scores = get_image_feature(img_path, img_list_path,
+                                                   model_path, epoch, gpu_id)
+    print('img_feats', img_feats.shape)
+    print('faceness_scores', faceness_scores.shape)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0],
+                                              img_feats.shape[1]))
+
+    # compute template features from image features.
+    start = timeit.default_timer()
+    # ==========================================================
+    # Norm feature before aggregation into template feature?
+    # Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
+    # ==========================================================
+    use_norm_score = True  # if True, TestMode(N1)
+    use_detector_score = True  # if True, TestMode(D1)
+    use_flip_test = True  # if True, TestMode(F1)
+
+    if use_flip_test:
+        # concat --- F1
+        #img_input_feats = img_feats
+        # add --- F2
+        img_input_feats = img_feats[:, 0:int(
+            img_feats.shape[1] / 2)] + img_feats[:,
+                                                 int(img_feats.shape[1] / 2):]
+    else:
+        img_input_feats = img_feats[:, 0:int(img_feats.shape[1] / 2)]
+
+    if use_norm_score:
+        img_input_feats = img_input_feats
+    else:
+        # normalise features to remove norm information
+        img_input_feats = img_input_feats / np.sqrt(
+            np.sum(img_input_feats**2, -1, keepdims=True))
+
+    if use_detector_score:
+        img_input_feats = img_input_feats * np.matlib.repmat(
+            faceness_scores[:, np.newaxis], 1, img_input_feats.shape[1])
+    else:
+        img_input_feats = img_input_feats
+    print("input features shape", img_input_feats.shape)
+
+    #load gallery feature
+    gallery_templates_feature, gallery_unique_templates, gallery_unique_subject_ids = image2template_feature(
+        img_input_feats, total_templates, total_medias, gallery_templates,
+        gallery_subject_ids)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    print("gallery_templates_feature", gallery_templates_feature.shape)
+    print("gallery_unique_subject_ids", gallery_unique_subject_ids.shape)
+    #np.savetxt("gallery_templates_feature.txt", gallery_templates_feature)
+    #np.savetxt("gallery_unique_subject_ids.txt", gallery_unique_subject_ids)
+
+    #load prope feature
+    probe_mixed_record = "%s_1N_probe_mixed.csv" % target.lower()
+    probe_mixed_templates, probe_mixed_subject_ids = read_template_subject_id_list(
+        os.path.join(meta_dir, probe_mixed_record))
+    print(probe_mixed_templates.shape, probe_mixed_subject_ids.shape)
+    probe_mixed_templates_feature, probe_mixed_unique_templates, probe_mixed_unique_subject_ids = image2template_feature(
+        img_input_feats, total_templates, total_medias, probe_mixed_templates,
+        probe_mixed_subject_ids)
+    print("probe_mixed_templates_feature", probe_mixed_templates_feature.shape)
+    print("probe_mixed_unique_subject_ids",
+          probe_mixed_unique_subject_ids.shape)
+    #np.savetxt("probe_mixed_templates_feature.txt", probe_mixed_templates_feature)
+    #np.savetxt("probe_mixed_unique_subject_ids.txt", probe_mixed_unique_subject_ids)
+
+    #root_dir = "" #feature root dir
+    #gallery_id_path = "" #id filepath
+    #gallery_feats_path = "" #feature filelpath
+    #print("{}: start loading gallery feat {}".format(dt.now(), gallery_id_path))
+    #gallery_ids, gallery_feats = load_feat_file(root_dir, gallery_id_path, gallery_feats_path)
+    #print("{}: end loading gallery feat".format(dt.now()))
+    #
+    #probe_id_path = "probe_mixed_unique_subject_ids.txt" #probe id filepath
+    #probe_feats_path = "probe_mixed_templates_feature.txt" #probe feats filepath
+    #print("{}: start loading probe feat {}".format(dt.now(), probe_id_path))
+    #probe_ids, probe_feats = load_feat_file(root_dir, probe_id_path, probe_feats_path)
+    #print("{}: end loading probe feat".format(dt.now()))
+
+    gallery_ids = gallery_unique_subject_ids
+    gallery_feats = gallery_templates_feature
+    probe_ids = probe_mixed_unique_subject_ids
+    probe_feats = probe_mixed_templates_feature
+
+    mask = gen_mask(probe_ids, gallery_ids)
+
+    print("{}: start evaluation".format(dt.now()))
+    evaluation(probe_feats, gallery_feats, mask)
+    print("{}: end evaluation".format(dt.now()))
diff --git a/recognition/_evaluation_/ijb/ijb_evals.py b/recognition/_evaluation_/ijb/ijb_evals.py
new file mode 100755
index 000000000..9652c89a4
--- /dev/null
+++ b/recognition/_evaluation_/ijb/ijb_evals.py
@@ -0,0 +1,655 @@
+#!/usr/bin/env python3
+import os
+import cv2
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+from skimage import transform
+from sklearn.preprocessing import normalize
+from sklearn.metrics import roc_curve, auc
+
+
+class Mxnet_model_interf:
+    def __init__(self, model_file, layer="fc1", image_size=(112, 112)):
+        import mxnet as mx
+
+        self.mx = mx
+        cvd = os.environ.get("CUDA_VISIBLE_DEVICES", "").strip()
+        if len(cvd) > 0 and int(cvd) != -1:
+            ctx = [self.mx.gpu(ii) for ii in range(len(cvd.split(",")))]
+        else:
+            ctx = [self.mx.cpu()]
+
+        prefix, epoch = model_file.split(",")
+        print(">>>> loading mxnet model:", prefix, epoch, ctx)
+        sym, arg_params, aux_params = self.mx.model.load_checkpoint(prefix, int(epoch))
+        all_layers = sym.get_internals()
+        sym = all_layers[layer + "_output"]
+        model = self.mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+        model.bind(data_shapes=[("data", (1, 3, image_size[0], image_size[1]))])
+        model.set_params(arg_params, aux_params)
+        self.model = model
+
+    def __call__(self, imgs):
+        # print(imgs.shape, imgs[0])
+        imgs = imgs.transpose(0, 3, 1, 2)
+        data = self.mx.nd.array(imgs)
+        db = self.mx.io.DataBatch(data=(data,))
+        self.model.forward(db, is_train=False)
+        emb = self.model.get_outputs()[0].asnumpy()
+        return emb
+
+
+class Torch_model_interf:
+    def __init__(self, model_file, image_size=(112, 112)):
+        import torch
+
+        self.torch = torch
+        cvd = os.environ.get("CUDA_VISIBLE_DEVICES", "").strip()
+        device_name = "cuda:0" if len(cvd) > 0 and int(cvd) != -1 else "cpu"
+        self.device = self.torch.device(device_name)
+        try:
+            self.model = self.torch.jit.load(model_file, map_location=device_name)
+        except:
+            print("Error: %s is weights only, please load and save the entire model by `torch.jit.save`" % model_file)
+            self.model = None
+
+    def __call__(self, imgs):
+        # print(imgs.shape, imgs[0])
+        imgs = imgs.transpose(0, 3, 1, 2).copy().astype("float32")
+        imgs = (imgs - 127.5) * 0.0078125
+        output = self.model(self.torch.from_numpy(imgs).to(self.device).float())
+        return output.cpu().detach().numpy()
+
+
+class ONNX_model_interf:
+    def __init__(self, model_file, image_size=(112, 112)):
+        import onnxruntime as ort
+        ort.set_default_logger_severity(3)
+        self.ort_session = ort.InferenceSession(model_file)
+        self.output_names = [self.ort_session.get_outputs()[0].name]
+        self.input_name = self.ort_session.get_inputs()[0].name
+
+    def __call__(self, imgs):
+        imgs = imgs.transpose(0, 3, 1, 2).astype("float32")
+        imgs = (imgs - 127.5) * 0.0078125
+        outputs = self.ort_session.run(self.output_names, {self.input_name: imgs})
+        return outputs[0]
+
+
+def keras_model_interf(model_file):
+    import tensorflow as tf
+    from tensorflow_addons.layers import StochasticDepth
+
+    for gpu in tf.config.experimental.list_physical_devices("GPU"):
+        tf.config.experimental.set_memory_growth(gpu, True)
+
+    mm = tf.keras.models.load_model(model_file, compile=False)
+    return lambda imgs: mm((tf.cast(imgs, "float32") - 127.5) * 0.0078125).numpy()
+
+
+def face_align_landmark(img, landmark, image_size=(112, 112), method="similar"):
+    tform = transform.AffineTransform() if method == "affine" else transform.SimilarityTransform()
+    src = np.array(
+        [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.729904, 92.2041]], dtype=np.float32
+    )
+    tform.estimate(landmark, src)
+    # ndimage = transform.warp(img, tform.inverse, output_shape=image_size)
+    # ndimage = (ndimage * 255).astype(np.uint8)
+    M = tform.params[0:2, :]
+    ndimage = cv2.warpAffine(img, M, image_size, borderValue=0.0)
+    if len(ndimage.shape) == 2:
+        ndimage = np.stack([ndimage, ndimage, ndimage], -1)
+    else:
+        ndimage = cv2.cvtColor(ndimage, cv2.COLOR_BGR2RGB)
+    return ndimage
+
+
+def read_IJB_meta_columns_to_int(file_path, columns, sep=" ", skiprows=0, header=None):
+    # meta = np.loadtxt(file_path, skiprows=skiprows, delimiter=sep)
+    meta = pd.read_csv(file_path, sep=sep, skiprows=skiprows, header=header).values
+    return (meta[:, ii].astype("int") for ii in columns)
+
+
+def extract_IJB_data_11(data_path, subset, save_path=None, force_reload=False):
+    if save_path == None:
+        save_path = os.path.join(data_path, subset + "_backup.npz")
+    if not force_reload and os.path.exists(save_path):
+        print(">>>> Reload from backup: %s ..." % save_path)
+        aa = np.load(save_path)
+        return (
+            aa["templates"],
+            aa["medias"],
+            aa["p1"],
+            aa["p2"],
+            aa["label"],
+            aa["img_names"],
+            aa["landmarks"],
+            aa["face_scores"],
+        )
+
+    if subset == "IJBB":
+        media_list_path = os.path.join(data_path, "IJBB/meta/ijbb_face_tid_mid.txt")
+        pair_list_path = os.path.join(data_path, "IJBB/meta/ijbb_template_pair_label.txt")
+        img_path = os.path.join(data_path, "IJBB/loose_crop")
+        img_list_path = os.path.join(data_path, "IJBB/meta/ijbb_name_5pts_score.txt")
+    else:
+        media_list_path = os.path.join(data_path, "IJBC/meta/ijbc_face_tid_mid.txt")
+        pair_list_path = os.path.join(data_path, "IJBC/meta/ijbc_template_pair_label.txt")
+        img_path = os.path.join(data_path, "IJBC/loose_crop")
+        img_list_path = os.path.join(data_path, "IJBC/meta/ijbc_name_5pts_score.txt")
+
+    print(">>>> Loading templates and medias...")
+    templates, medias = read_IJB_meta_columns_to_int(media_list_path, columns=[1, 2])  # ['1.jpg', '1', '69544']
+    print("templates: %s, medias: %s, unique templates: %s" % (templates.shape, medias.shape, np.unique(templates).shape))
+    # templates: (227630,), medias: (227630,), unique templates: (12115,)
+
+    print(">>>> Loading pairs...")
+    p1, p2, label = read_IJB_meta_columns_to_int(pair_list_path, columns=[0, 1, 2])  # ['1', '11065', '1']
+    print("p1: %s, unique p1: %s" % (p1.shape, np.unique(p1).shape))
+    print("p2: %s, unique p2: %s" % (p2.shape, np.unique(p2).shape))
+    print("label: %s, label value counts: %s" % (label.shape, dict(zip(*np.unique(label, return_counts=True)))))
+    # p1: (8010270,), unique p1: (1845,)
+    # p2: (8010270,), unique p2: (10270,) # 10270 + 1845 = 12115 --> np.unique(templates).shape
+    # label: (8010270,), label value counts: {0: 8000000, 1: 10270}
+
+    print(">>>> Loading images...")
+    with open(img_list_path, "r") as ff:
+        # 1.jpg 46.060 62.026 87.785 60.323 68.851 77.656 52.162 99.875 86.450 98.648 0.999
+        img_records = np.array([ii.strip().split(" ") for ii in ff.readlines()])
+
+    img_names = np.array([os.path.join(img_path, ii) for ii in img_records[:, 0]])
+    landmarks = img_records[:, 1:-1].astype("float32").reshape(-1, 5, 2)
+    face_scores = img_records[:, -1].astype("float32")
+    print("img_names: %s, landmarks: %s, face_scores: %s" % (img_names.shape, landmarks.shape, face_scores.shape))
+    # img_names: (227630,), landmarks: (227630, 5, 2), face_scores: (227630,)
+    print("face_scores value counts:", dict(zip(*np.histogram(face_scores, bins=9)[::-1])))
+    # {0.1: 2515, 0.2: 0, 0.3: 62, 0.4: 94, 0.5: 136, 0.6: 197, 0.7: 291, 0.8: 538, 0.9: 223797}
+
+    print(">>>> Saving backup to: %s ..." % save_path)
+    np.savez(
+        save_path,
+        templates=templates,
+        medias=medias,
+        p1=p1,
+        p2=p2,
+        label=label,
+        img_names=img_names,
+        landmarks=landmarks,
+        face_scores=face_scores,
+    )
+    print()
+    return templates, medias, p1, p2, label, img_names, landmarks, face_scores
+
+
+def extract_gallery_prob_data(data_path, subset, save_path=None, force_reload=False):
+    if save_path == None:
+        save_path = os.path.join(data_path, subset + "_gallery_prob_backup.npz")
+    if not force_reload and os.path.exists(save_path):
+        print(">>>> Reload from backup: %s ..." % save_path)
+        aa = np.load(save_path)
+        return (
+            aa["s1_templates"],
+            aa["s1_subject_ids"],
+            aa["s2_templates"],
+            aa["s2_subject_ids"],
+            aa["probe_mixed_templates"],
+            aa["probe_mixed_subject_ids"],
+        )
+
+    if subset == "IJBC":
+        meta_dir = os.path.join(data_path, "IJBC/meta")
+        gallery_s1_record = os.path.join(meta_dir, "ijbc_1N_gallery_G1.csv")
+        gallery_s2_record = os.path.join(meta_dir, "ijbc_1N_gallery_G2.csv")
+        probe_mixed_record = os.path.join(meta_dir, "ijbc_1N_probe_mixed.csv")
+    else:
+        meta_dir = os.path.join(data_path, "IJBB/meta")
+        gallery_s1_record = os.path.join(meta_dir, "ijbb_1N_gallery_S1.csv")
+        gallery_s2_record = os.path.join(meta_dir, "ijbb_1N_gallery_S2.csv")
+        probe_mixed_record = os.path.join(meta_dir, "ijbb_1N_probe_mixed.csv")
+
+    print(">>>> Loading gallery feature...")
+    s1_templates, s1_subject_ids = read_IJB_meta_columns_to_int(gallery_s1_record, columns=[0, 1], skiprows=1, sep=",")
+    s2_templates, s2_subject_ids = read_IJB_meta_columns_to_int(gallery_s2_record, columns=[0, 1], skiprows=1, sep=",")
+    print("s1 gallery: %s, ids: %s, unique: %s" % (s1_templates.shape, s1_subject_ids.shape, np.unique(s1_templates).shape))
+    print("s2 gallery: %s, ids: %s, unique: %s" % (s2_templates.shape, s2_subject_ids.shape, np.unique(s2_templates).shape))
+
+    print(">>>> Loading prope feature...")
+    probe_mixed_templates, probe_mixed_subject_ids = read_IJB_meta_columns_to_int(
+        probe_mixed_record, columns=[0, 1], skiprows=1, sep=","
+    )
+    print("probe_mixed_templates: %s, unique: %s" % (probe_mixed_templates.shape, np.unique(probe_mixed_templates).shape))
+    print("probe_mixed_subject_ids: %s, unique: %s" % (probe_mixed_subject_ids.shape, np.unique(probe_mixed_subject_ids).shape))
+
+    print(">>>> Saving backup to: %s ..." % save_path)
+    np.savez(
+        save_path,
+        s1_templates=s1_templates,
+        s1_subject_ids=s1_subject_ids,
+        s2_templates=s2_templates,
+        s2_subject_ids=s2_subject_ids,
+        probe_mixed_templates=probe_mixed_templates,
+        probe_mixed_subject_ids=probe_mixed_subject_ids,
+    )
+    print()
+    return s1_templates, s1_subject_ids, s2_templates, s2_subject_ids, probe_mixed_templates, probe_mixed_subject_ids
+
+
+def get_embeddings(model_interf, img_names, landmarks, batch_size=64, flip=True):
+    steps = int(np.ceil(len(img_names) / batch_size))
+    embs, embs_f = [], []
+    for batch_id in tqdm(range(0, len(img_names), batch_size), "Embedding", total=steps):
+        batch_imgs, batch_landmarks = img_names[batch_id : batch_id + batch_size], landmarks[batch_id : batch_id + batch_size]
+        ndimages = [face_align_landmark(cv2.imread(img), landmark) for img, landmark in zip(batch_imgs, batch_landmarks)]
+        ndimages = np.stack(ndimages)
+        embs.extend(model_interf(ndimages))
+        if flip:
+            embs_f.extend(model_interf(ndimages[:, :, ::-1, :]))
+    return np.array(embs), np.array(embs_f)
+
+
+def process_embeddings(embs, embs_f=[], use_flip_test=True, use_norm_score=False, use_detector_score=True, face_scores=None):
+    print(">>>> process_embeddings: Norm {}, Detect_score {}, Flip {}".format(use_norm_score, use_detector_score, use_flip_test))
+    if use_flip_test and len(embs_f) != 0:
+        embs = embs + embs_f
+    if use_norm_score:
+        embs = normalize(embs)
+    if use_detector_score and face_scores is not None:
+        embs = embs * np.expand_dims(face_scores, -1)
+    return embs
+
+
+def image2template_feature(img_feats=None, templates=None, medias=None, choose_templates=None, choose_ids=None):
+    if choose_templates is not None:  # 1:N
+        unique_templates, indices = np.unique(choose_templates, return_index=True)
+        unique_subjectids = choose_ids[indices]
+    else:  # 1:1
+        unique_templates = np.unique(templates)
+        unique_subjectids = None
+
+    # template_feats = np.zeros((len(unique_templates), img_feats.shape[1]), dtype=img_feats.dtype)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+    for count_template, uqt in tqdm(enumerate(unique_templates), "Extract template feature", total=len(unique_templates)):
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [np.mean(face_norm_feats[ind_m], 0, keepdims=True)]
+        media_norm_feats = np.array(media_norm_feats)
+        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
+        template_feats[count_template] = np.sum(media_norm_feats, 0)
+    template_norm_feats = normalize(template_feats)
+    return template_norm_feats, unique_templates, unique_subjectids
+
+
+def verification_11(template_norm_feats=None, unique_templates=None, p1=None, p2=None, batch_size=10000):
+    try:
+        print(">>>> Trying cupy.")
+        import cupy as cp
+
+        template_norm_feats = cp.array(template_norm_feats)
+        score_func = lambda feat1, feat2: cp.sum(feat1 * feat2, axis=-1).get()
+        test = score_func(template_norm_feats[:batch_size], template_norm_feats[:batch_size])
+    except:
+        score_func = lambda feat1, feat2: np.sum(feat1 * feat2, -1)
+
+    template2id = np.zeros(max(unique_templates) + 1, dtype=int)
+    template2id[unique_templates] = np.arange(len(unique_templates))
+
+    steps = int(np.ceil(len(p1) / batch_size))
+    score = []
+    for id in tqdm(range(steps), "Verification"):
+        feat1 = template_norm_feats[template2id[p1[id * batch_size : (id + 1) * batch_size]]]
+        feat2 = template_norm_feats[template2id[p2[id * batch_size : (id + 1) * batch_size]]]
+        score.extend(score_func(feat1, feat2))
+    return np.array(score)
+
+
+def evaluation_1N(query_feats, gallery_feats, query_ids, reg_ids, fars=[0.01, 0.1]):
+    print("query_feats: %s, gallery_feats: %s" % (query_feats.shape, gallery_feats.shape))
+    similarity = np.dot(query_feats, gallery_feats.T)  # (19593, 3531)
+
+    top_1_count, top_5_count, top_10_count = 0, 0, 0
+    pos_sims, neg_sims, non_gallery_sims = [], [], []
+    for index, query_id in enumerate(query_ids):
+        if query_id in reg_ids:
+            gallery_label = np.argwhere(reg_ids == query_id)[0, 0]
+            index_sorted = np.argsort(similarity[index])[::-1]
+
+            top_1_count += gallery_label in index_sorted[:1]
+            top_5_count += gallery_label in index_sorted[:5]
+            top_10_count += gallery_label in index_sorted[:10]
+
+            pos_sims.append(similarity[index][reg_ids == query_id][0])
+            neg_sims.append(similarity[index][reg_ids != query_id])
+        else:
+            non_gallery_sims.append(similarity[index])
+    total_pos = len(pos_sims)
+    pos_sims, neg_sims, non_gallery_sims = np.array(pos_sims), np.array(neg_sims), np.array(non_gallery_sims)
+    print("pos_sims: %s, neg_sims: %s, non_gallery_sims: %s" % (pos_sims.shape, neg_sims.shape, non_gallery_sims.shape))
+    print("top1: %f, top5: %f, top10: %f" % (top_1_count / total_pos, top_5_count / total_pos, top_10_count / total_pos))
+
+    correct_pos_cond = pos_sims > neg_sims.max(1)
+    non_gallery_sims_sorted = np.sort(non_gallery_sims.max(1))[::-1]
+    threshes, recalls = [], []
+    for far in fars:
+        # thresh = non_gallery_sims_sorted[int(np.ceil(non_gallery_sims_sorted.shape[0] * far)) - 1]
+        thresh = non_gallery_sims_sorted[max(int((non_gallery_sims_sorted.shape[0]) * far) - 1, 0)]
+        recall = np.logical_and(correct_pos_cond, pos_sims > thresh).sum() / pos_sims.shape[0]
+        threshes.append(thresh)
+        recalls.append(recall)
+        # print("FAR = {:.10f} TPIR = {:.10f} th = {:.10f}".format(far, recall, thresh))
+    cmc_scores = list(zip(neg_sims, pos_sims.reshape(-1, 1))) + list(zip(non_gallery_sims, [None] * non_gallery_sims.shape[0]))
+    return top_1_count, top_5_count, top_10_count, threshes, recalls, cmc_scores
+
+
+class IJB_test:
+    def __init__(self, model_file, data_path, subset, batch_size=64, force_reload=False, restore_embs=None):
+        templates, medias, p1, p2, label, img_names, landmarks, face_scores = extract_IJB_data_11(
+            data_path, subset, force_reload=force_reload
+        )
+        if model_file != None:
+            if model_file.endswith(".h5"):
+                interf_func = keras_model_interf(model_file)
+            elif model_file.endswith(".pth") or model_file.endswith(".pt"):
+                interf_func = Torch_model_interf(model_file)
+            elif model_file.endswith(".onnx") or model_file.endswith(".ONNX"):
+                interf_func = ONNX_model_interf(model_file)
+            else:
+                interf_func = Mxnet_model_interf(model_file)
+            self.embs, self.embs_f = get_embeddings(interf_func, img_names, landmarks, batch_size=batch_size)
+        elif restore_embs != None:
+            print(">>>> Reload embeddings from:", restore_embs)
+            aa = np.load(restore_embs)
+            if "embs" in aa and "embs_f" in aa:
+                self.embs, self.embs_f = aa["embs"], aa["embs_f"]
+            else:
+                print("ERROR: %s NOT containing embs / embs_f" % restore_embs)
+                exit(1)
+            print(">>>> Done.")
+        self.data_path, self.subset, self.force_reload = data_path, subset, force_reload
+        self.templates, self.medias, self.p1, self.p2, self.label = templates, medias, p1, p2, label
+        self.face_scores = face_scores.astype(self.embs.dtype)
+
+    def run_model_test_single(self, use_flip_test=True, use_norm_score=False, use_detector_score=True):
+        img_input_feats = process_embeddings(
+            self.embs,
+            self.embs_f,
+            use_flip_test=use_flip_test,
+            use_norm_score=use_norm_score,
+            use_detector_score=use_detector_score,
+            face_scores=self.face_scores,
+        )
+        template_norm_feats, unique_templates, _ = image2template_feature(img_input_feats, self.templates, self.medias)
+        score = verification_11(template_norm_feats, unique_templates, self.p1, self.p2)
+        return score
+
+    def run_model_test_bunch(self):
+        from itertools import product
+
+        scores, names = [], []
+        for use_norm_score, use_detector_score, use_flip_test in product([True, False], [True, False], [True, False]):
+            name = "N{:d}D{:d}F{:d}".format(use_norm_score, use_detector_score, use_flip_test)
+            print(">>>>", name, use_norm_score, use_detector_score, use_flip_test)
+            names.append(name)
+            scores.append(self.run_model_test_single(use_flip_test, use_norm_score, use_detector_score))
+        return scores, names
+
+    def run_model_test_1N(self, npoints=100):
+        fars_cal = [10 ** ii for ii in np.arange(-4, 0, 4 / npoints)] + [1]  # plot in range [10-4, 1]
+        fars_show_idx = np.arange(len(fars_cal))[:: npoints // 4]  # npoints=100, fars_show=[0.0001, 0.001, 0.01, 0.1, 1.0]
+
+        g1_templates, g1_ids, g2_templates, g2_ids, probe_mixed_templates, probe_mixed_ids = extract_gallery_prob_data(
+            self.data_path, self.subset, force_reload=self.force_reload
+        )
+        img_input_feats = process_embeddings(
+            self.embs,
+            self.embs_f,
+            use_flip_test=True,
+            use_norm_score=False,
+            use_detector_score=True,
+            face_scores=self.face_scores,
+        )
+        g1_templates_feature, g1_unique_templates, g1_unique_ids = image2template_feature(
+            img_input_feats, self.templates, self.medias, g1_templates, g1_ids
+        )
+        g2_templates_feature, g2_unique_templates, g2_unique_ids = image2template_feature(
+            img_input_feats, self.templates, self.medias, g2_templates, g2_ids
+        )
+        probe_mixed_templates_feature, probe_mixed_unique_templates, probe_mixed_unique_subject_ids = image2template_feature(
+            img_input_feats, self.templates, self.medias, probe_mixed_templates, probe_mixed_ids
+        )
+        print("g1_templates_feature:", g1_templates_feature.shape)  # (1772, 512)
+        print("g2_templates_feature:", g2_templates_feature.shape)  # (1759, 512)
+
+        print("probe_mixed_templates_feature:", probe_mixed_templates_feature.shape)  # (19593, 512)
+        print("probe_mixed_unique_subject_ids:", probe_mixed_unique_subject_ids.shape)  # (19593,)
+
+        print(">>>> Gallery 1")
+        g1_top_1_count, g1_top_5_count, g1_top_10_count, g1_threshes, g1_recalls, g1_cmc_scores = evaluation_1N(
+            probe_mixed_templates_feature, g1_templates_feature, probe_mixed_unique_subject_ids, g1_unique_ids, fars_cal
+        )
+        print(">>>> Gallery 2")
+        g2_top_1_count, g2_top_5_count, g2_top_10_count, g2_threshes, g2_recalls, g2_cmc_scores = evaluation_1N(
+            probe_mixed_templates_feature, g2_templates_feature, probe_mixed_unique_subject_ids, g2_unique_ids, fars_cal
+        )
+        print(">>>> Mean")
+        query_num = probe_mixed_templates_feature.shape[0]
+        top_1 = (g1_top_1_count + g2_top_1_count) / query_num
+        top_5 = (g1_top_5_count + g2_top_5_count) / query_num
+        top_10 = (g1_top_10_count + g2_top_10_count) / query_num
+        print("[Mean] top1: %f, top5: %f, top10: %f" % (top_1, top_5, top_10))
+
+        mean_tpirs = (np.array(g1_recalls) + np.array(g2_recalls)) / 2
+        show_result = {}
+        for id, far in enumerate(fars_cal):
+            if id in fars_show_idx:
+                show_result.setdefault("far", []).append(far)
+                show_result.setdefault("g1_tpir", []).append(g1_recalls[id])
+                show_result.setdefault("g1_thresh", []).append(g1_threshes[id])
+                show_result.setdefault("g2_tpir", []).append(g2_recalls[id])
+                show_result.setdefault("g2_thresh", []).append(g2_threshes[id])
+                show_result.setdefault("mean_tpir", []).append(mean_tpirs[id])
+        print(pd.DataFrame(show_result).set_index("far").to_markdown())
+        return fars_cal, mean_tpirs, g1_cmc_scores, g2_cmc_scores
+
+
+def plot_roc_and_calculate_tpr(scores, names=None, label=None):
+    print(">>>> plot roc and calculate tpr...")
+    score_dict = {}
+    for id, score in enumerate(scores):
+        name = None if names is None else names[id]
+        if isinstance(score, str) and score.endswith(".npz"):
+            aa = np.load(score)
+            score = aa.get("scores", [])
+            label = aa["label"] if label is None and "label" in aa else label
+            score_name = aa.get("names", [])
+            for ss, nn in zip(score, score_name):
+                score_dict[nn] = ss
+        elif isinstance(score, str) and score.endswith(".npy"):
+            name = name if name is not None else os.path.splitext(os.path.basename(score))[0]
+            score_dict[name] = np.load(score)
+        elif isinstance(score, str) and score.endswith(".txt"):
+            # IJB meta data like ijbb_template_pair_label.txt
+            label = pd.read_csv(score, sep=" ", header=None).values[:, 2]
+        else:
+            name = name if name is not None else str(id)
+            score_dict[name] = score
+    if label is None:
+        print("Error: Label data is not provided")
+        return None, None
+
+    x_labels = [10 ** (-ii) for ii in range(1, 7)[::-1]]
+    fpr_dict, tpr_dict, roc_auc_dict, tpr_result = {}, {}, {}, {}
+    for name, score in score_dict.items():
+        fpr, tpr, _ = roc_curve(label, score)
+        roc_auc = auc(fpr, tpr)
+        fpr, tpr = np.flipud(fpr), np.flipud(tpr)  # select largest tpr at same fpr
+        tpr_result[name] = [tpr[np.argmin(abs(fpr - ii))] for ii in x_labels]
+        fpr_dict[name], tpr_dict[name], roc_auc_dict[name] = fpr, tpr, roc_auc
+    tpr_result_df = pd.DataFrame(tpr_result, index=x_labels).T
+    tpr_result_df['AUC'] = pd.Series(roc_auc_dict)
+    tpr_result_df.columns.name = "Methods"
+    print(tpr_result_df.to_markdown())
+    # print(tpr_result_df)
+
+    try:
+        import matplotlib.pyplot as plt
+
+        fig = plt.figure()
+        for name in score_dict:
+            plt.plot(fpr_dict[name], tpr_dict[name], lw=1, label="[%s (AUC = %0.4f%%)]" % (name, roc_auc_dict[name] * 100))
+        title = "ROC on IJB" + name.split("IJB")[-1][0] if "IJB" in name else "ROC on IJB"
+
+        plt.xlim([10 ** -6, 0.1])
+        plt.xscale("log")
+        plt.xticks(x_labels)
+        plt.xlabel("False Positive Rate")
+        plt.ylim([0.3, 1.0])
+        plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+        plt.ylabel("True Positive Rate")
+
+        plt.grid(linestyle="--", linewidth=1)
+        plt.title(title)
+        plt.legend(loc="lower right", fontsize='x-small')
+        plt.tight_layout()
+        plt.show()
+    except:
+        print("matplotlib plot failed")
+        fig = None
+
+    return tpr_result_df, fig
+
+
+def plot_dir_far_cmc_scores(scores, names=None):
+    try:
+        import matplotlib.pyplot as plt
+
+        fig = plt.figure()
+        for id, score in enumerate(scores):
+            name = None if names is None else names[id]
+            if isinstance(score, str) and score.endswith(".npz"):
+                aa = np.load(score)
+                score, name = aa.get("scores")[0], aa.get("names")[0]
+            fars, tpirs = score[0], score[1]
+            name = name if name is not None else str(id)
+
+            auc_value = auc(fars, tpirs)
+            label = "[%s (AUC = %0.4f%%)]" % (name, auc_value * 100)
+            plt.plot(fars, tpirs, lw=1, label=label)
+
+        plt.xlabel("False Alarm Rate")
+        plt.xlim([0.0001, 1])
+        plt.xscale("log")
+        plt.ylabel("Detection & Identification Rate (%)")
+        plt.ylim([0, 1])
+
+        plt.grid(linestyle="--", linewidth=1)
+        plt.legend(fontsize='x-small')
+        plt.tight_layout()
+        plt.show()
+    except:
+        print("matplotlib plot failed")
+        fig = None
+
+    return fig
+
+
+def parse_arguments(argv):
+    import argparse
+
+    default_save_result_name = "IJB_result/{model_name}_{subset}_{type}.npz"
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("-m", "--model_file", type=str, default=None, help="Saved model, keras h5 / pytorch jit pth / onnx / mxnet")
+    parser.add_argument("-d", "--data_path", type=str, default="./", help="Dataset path containing IJBB and IJBC sub folder")
+    parser.add_argument("-s", "--subset", type=str, default="IJBC", help="Subset test target, could be IJBB / IJBC")
+    parser.add_argument("-b", "--batch_size", type=int, default=128, help="Batch size for get_embeddings")
+    parser.add_argument(
+        "-R", "--save_result", type=str, default=default_save_result_name, help="Filename for saving / restore result"
+    )
+    parser.add_argument("-L", "--save_label", action="store_true", help="Save label data, useful for plot only")
+    parser.add_argument("-E", "--save_embeddings", action="store_true", help="Save embeddings data")
+    parser.add_argument("-B", "--is_bunch", action="store_true", help="Run all 8 tests N{0,1}D{0,1}F{0,1}")
+    parser.add_argument("-N", "--is_one_2_N", action="store_true", help="Run 1:N test instead of 1:1")
+    parser.add_argument("-F", "--force_reload", action="store_true", help="Force reload, instead of using cache")
+    parser.add_argument("-P", "--plot_only", nargs="*", type=str, help="Plot saved results, Format 1 2 3 or 1, 2, 3 or *.npy")
+    args = parser.parse_known_args(argv)[0]
+
+    if args.plot_only != None and len(args.plot_only) != 0:
+        # Plot only
+        from glob2 import glob
+
+        score_files = []
+        for ss in args.plot_only:
+            score_files.extend(glob(ss.replace(",", "").strip()))
+        args.plot_only = score_files
+    elif args.model_file == None and args.save_result == default_save_result_name:
+        print("Please provide -m MODEL_FILE, see `--help` for usage.")
+        exit(1)
+    elif args.model_file != None:
+        if args.model_file.endswith(".h5") or args.model_file.endswith(".pth") or args.model_file.endswith(".pt") or args.model_file.endswith(".onnx"):
+            # Keras model file "model.h5", pytorch model ends with `.pth` or `.pt`, onnx model ends with `.onnx`
+            model_name = os.path.splitext(os.path.basename(args.model_file))[0]
+        else:
+            # MXNet model file "models/r50-arcface-emore/model,1"
+            model_name = os.path.basename(os.path.dirname(args.model_file))
+
+        if args.save_result == default_save_result_name:
+            type = "1N" if args.is_one_2_N else "11"
+            args.save_result = default_save_result_name.format(model_name=model_name, subset=args.subset, type=type)
+    return args
+
+
+if __name__ == "__main__":
+    import sys
+
+    args = parse_arguments(sys.argv[1:])
+    if args.plot_only != None and len(args.plot_only) != 0:
+        if args.is_one_2_N:
+            plot_dir_far_cmc_scores(args.plot_only)
+        else:
+            plot_roc_and_calculate_tpr(args.plot_only)
+    else:
+        save_name = os.path.splitext(os.path.basename(args.save_result))[0]
+        save_items = {}
+        save_path = os.path.dirname(args.save_result)
+        if len(save_path) != 0 and not os.path.exists(save_path):
+            os.makedirs(save_path)
+
+        tt = IJB_test(args.model_file, args.data_path, args.subset, args.batch_size, args.force_reload, args.save_result)
+        if args.save_embeddings:  # Save embeddings first, in case of any error happens later...
+            np.savez(args.save_result, embs=tt.embs, embs_f=tt.embs_f)
+
+        if args.is_one_2_N:  # 1:N test
+            fars, tpirs, _, _ = tt.run_model_test_1N()
+            scores = [(fars, tpirs)]
+            names = [save_name]
+            save_items.update({"scores": scores, "names": names})
+        elif args.is_bunch:  # All 8 tests N{0,1}D{0,1}F{0,1}
+            scores, names = tt.run_model_test_bunch()
+            names = [save_name + "_" + ii for ii in names]
+            label = tt.label
+            save_items.update({"scores": scores, "names": names})
+        else:  # Basic 1:1 N0D1F1 test
+            score = tt.run_model_test_single()
+            scores, names, label = [score], [save_name], tt.label
+            save_items.update({"scores": scores, "names": names})
+
+        if args.save_embeddings:
+            save_items.update({"embs": tt.embs, "embs_f": tt.embs_f})
+        if args.save_label:
+            save_items.update({"label": label})
+
+        if args.model_file != None or args.save_embeddings:  # embeddings not restored from file or should save_embeddings again
+            np.savez(args.save_result, **save_items)
+
+        if args.is_one_2_N:
+            plot_dir_far_cmc_scores(scores=scores, names=names)
+        else:
+            plot_roc_and_calculate_tpr(scores, names=names, label=label)
diff --git a/recognition/_evaluation_/ijb/ijb_onnx.py b/recognition/_evaluation_/ijb/ijb_onnx.py
new file mode 100644
index 000000000..eb2edbe83
--- /dev/null
+++ b/recognition/_evaluation_/ijb/ijb_onnx.py
@@ -0,0 +1,267 @@
+import argparse
+import os
+import pickle
+import timeit
+
+import cv2
+import mxnet as mx
+import numpy as np
+import pandas as pd
+import prettytable
+import skimage.transform
+from sklearn.metrics import roc_curve
+from sklearn.preprocessing import normalize
+import insightface
+from insightface.model_zoo import ArcFaceONNX
+
+
+SRC = np.array(
+    [
+        [30.2946, 51.6963],
+        [65.5318, 51.5014],
+        [48.0252, 71.7366],
+        [33.5493, 92.3655],
+        [62.7299, 92.2041]]
+    , dtype=np.float32)
+SRC[:, 0] += 8.0
+
+
+class AlignedDataSet(mx.gluon.data.Dataset):
+    def __init__(self, root, lines, align=True):
+        self.lines = lines
+        self.root = root
+        self.align = align
+
+    def __len__(self):
+        return len(self.lines)
+
+    def __getitem__(self, idx):
+        each_line = self.lines[idx]
+        name_lmk_score = each_line.strip().split(' ')
+        name = os.path.join(self.root, name_lmk_score[0])
+        img = cv2.cvtColor(cv2.imread(name), cv2.COLOR_BGR2RGB)
+        landmark5 = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32).reshape((5, 2))
+        st = skimage.transform.SimilarityTransform()
+        st.estimate(landmark5, SRC)
+        img = cv2.warpAffine(img, st.params[0:2, :], (112, 112), borderValue=0.0)
+        img_1 = np.expand_dims(img, 0)
+        img_2 = np.expand_dims(np.fliplr(img), 0)
+        output = np.concatenate((img_1, img_2), axis=0).astype(np.float32)
+        output = np.transpose(output, (0, 3, 1, 2))
+        output = mx.nd.array(output)
+        return output
+
+
+def extract(model_file, dataset):
+    model = ArcFaceONNX(model_file=model_file)
+    model.check()
+    feat_mat = np.zeros(shape=(len(dataset), 2 * model.feat_dim))
+
+    def batchify_fn(data):
+        return mx.nd.concat(*data, dim=0)
+
+    data_loader = mx.gluon.data.DataLoader(
+        dataset, 128, last_batch='keep', num_workers=4,
+        thread_pool=True, prefetch=16, batchify_fn=batchify_fn)
+    num_iter = 0
+    for batch in data_loader:
+        batch = batch.asnumpy()
+        feat = model.forward(batch)
+        feat = np.reshape(feat, (-1, model.feat_dim * 2))
+        feat_mat[128 * num_iter: 128 * num_iter + feat.shape[0], :] = feat
+        num_iter += 1
+        if num_iter % 50 == 0:
+            print(num_iter)
+    return feat_mat
+
+
+def read_template_media_list(path):
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+def image2template_feature(img_feats=None,
+                           templates=None,
+                           medias=None):
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [np.mean(face_norm_feats[ind_m], axis=0, keepdims=True), ]
+        media_norm_feats = np.array(media_norm_feats)
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    template_norm_feats = normalize(template_feats)
+    return template_norm_feats, unique_templates
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000
+    sublists = [total_pairs[i: i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def main(args):
+    use_norm_score = True  # if Ture, TestMode(N1)
+    use_detector_score = True  # if Ture, TestMode(D1)
+    use_flip_test = True  # if Ture, TestMode(F1)
+    assert args.target == 'IJBC' or args.target == 'IJBB'
+
+    start = timeit.default_timer()
+    templates, medias = read_template_media_list(
+        os.path.join('%s/meta' % args.image_path, '%s_face_tid_mid.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    p1, p2, label = read_template_pair_list(
+        os.path.join('%s/meta' % args.image_path,
+                     '%s_template_pair_label.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    img_path = '%s/loose_crop' % args.image_path
+    img_list_path = '%s/meta/%s_name_5pts_score.txt' % (args.image_path, args.target.lower())
+    img_list = open(img_list_path)
+    files = img_list.readlines()
+    dataset = AlignedDataSet(root=img_path, lines=files, align=True)
+    img_feats = extract(args.model_file, dataset)
+
+    faceness_scores = []
+    for each_line in files:
+        name_lmk_score = each_line.split()
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
+    start = timeit.default_timer()
+
+    if use_flip_test:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2] + img_feats[:, img_feats.shape[1] // 2:]
+    else:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+    if use_norm_score:
+        img_input_feats = img_input_feats
+    else:
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+    if use_detector_score:
+        print(img_input_feats.shape, faceness_scores.shape)
+        img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+    else:
+        img_input_feats = img_input_feats
+
+    template_norm_feats, unique_templates = image2template_feature(
+        img_input_feats, templates, medias)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    score = verification(template_norm_feats, unique_templates, p1, p2)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    save_path = os.path.join(args.result_dir, "{}_result".format(args.target))
+    if not os.path.exists(save_path):
+        os.makedirs(save_path)
+    score_save_file = os.path.join(save_path, "{}.npy".format(args.model_file.split('/')[-1]))
+    np.save(score_save_file, score)
+    files = [score_save_file]
+    methods = []
+    scores = []
+    for file in files:
+        methods.append(os.path.basename(file))
+        scores.append(np.load(file))
+    methods = np.array(methods)
+    scores = dict(zip(methods, scores))
+    x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+    tpr_fpr_table = prettytable.PrettyTable(['Methods'] + [str(x) for x in x_labels])
+    for method in methods:
+        fpr, tpr, _ = roc_curve(label, scores[method])
+        fpr = np.flipud(fpr)
+        tpr = np.flipud(tpr)
+        tpr_fpr_row = []
+        tpr_fpr_row.append("%s-%s" % (method, args.target))
+        for fpr_iter in np.arange(len(x_labels)):
+            _, min_index = min(
+                list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+            tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+        tpr_fpr_table.add_row(tpr_fpr_row)
+    print(tpr_fpr_table)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='do onnx ijb test')
+    # general
+    parser.add_argument('--model-file', default='', help='path to onnx model.')
+    parser.add_argument('--image-path', default='', type=str, help='')
+    parser.add_argument('--result-dir', default='.', type=str, help='')
+    parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+    main(parser.parse_args())
diff --git a/recognition/_evaluation_/megaface/README.md b/recognition/_evaluation_/megaface/README.md
new file mode 100644
index 000000000..6e4ee657a
--- /dev/null
+++ b/recognition/_evaluation_/megaface/README.md
@@ -0,0 +1,4 @@
+
+Download megaface testsuite from [baiducloud](https://pan.baidu.com/s/1Vdxc2GgbY8wIW0hVcObIwg)(code:0n6w) or [gdrive](https://drive.google.com/file/d/1KBwp0U9oZgZj7SYDXRxUnnH7Lwvd9XMy/view?usp=sharing). The official devkit is also included.
+
+
diff --git a/recognition/_evaluation_/megaface/gen_megaface.py b/recognition/_evaluation_/megaface/gen_megaface.py
new file mode 100644
index 000000000..2e9fdb076
--- /dev/null
+++ b/recognition/_evaluation_/megaface/gen_megaface.py
@@ -0,0 +1,196 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+from easydict import EasyDict as edict
+import time
+import sys
+import numpy as np
+import argparse
+import struct
+import cv2
+import sklearn
+from sklearn.preprocessing import normalize
+import mxnet as mx
+from mxnet import ndarray as nd
+
+
+def read_img(image_path):
+    img = cv2.imread(image_path, cv2.CV_LOAD_IMAGE_COLOR)
+    return img
+
+
+def get_feature(imgs, nets):
+    count = len(imgs)
+    data = mx.nd.zeros(shape=(count * 2, 3, imgs[0].shape[0],
+                              imgs[0].shape[1]))
+    for idx, img in enumerate(imgs):
+        img = img[:, :, ::-1]  #to rgb
+        img = np.transpose(img, (2, 0, 1))
+        for flipid in [0, 1]:
+            _img = np.copy(img)
+            if flipid == 1:
+                _img = _img[:, :, ::-1]
+            _img = nd.array(_img)
+            data[count * flipid + idx] = _img
+
+    F = []
+    for net in nets:
+        db = mx.io.DataBatch(data=(data, ))
+        net.model.forward(db, is_train=False)
+        x = net.model.get_outputs()[0].asnumpy()
+        embedding = x[0:count, :] + x[count:, :]
+        embedding = sklearn.preprocessing.normalize(embedding)
+        #print('emb', embedding.shape)
+        F.append(embedding)
+    F = np.concatenate(F, axis=1)
+    F = sklearn.preprocessing.normalize(F)
+    #print('F', F.shape)
+    return F
+
+
+def write_bin(path, feature):
+    feature = list(feature)
+    with open(path, 'wb') as f:
+        f.write(struct.pack('4i', len(feature), 1, 4, 5))
+        f.write(struct.pack("%df" % len(feature), *feature))
+
+
+def get_and_write(buffer, nets):
+    imgs = []
+    for k in buffer:
+        imgs.append(k[0])
+    features = get_feature(imgs, nets)
+    #print(np.linalg.norm(feature))
+    assert features.shape[0] == len(buffer)
+    for ik, k in enumerate(buffer):
+        out_path = k[1]
+        feature = features[ik].flatten()
+        write_bin(out_path, feature)
+
+
+def main(args):
+
+    print(args)
+    gpuid = args.gpu
+    ctx = mx.gpu(gpuid)
+    nets = []
+    image_shape = [int(x) for x in args.image_size.split(',')]
+    for model in args.model.split('|'):
+        vec = model.split(',')
+        assert len(vec) > 1
+        prefix = vec[0]
+        epoch = int(vec[1])
+        print('loading', prefix, epoch)
+        net = edict()
+        net.ctx = ctx
+        net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(
+            prefix, epoch)
+        all_layers = net.sym.get_internals()
+        net.sym = all_layers['fc1_output']
+        net.model = mx.mod.Module(symbol=net.sym,
+                                  context=net.ctx,
+                                  label_names=None)
+        net.model.bind(data_shapes=[('data', (1, 3, image_shape[1],
+                                              image_shape[2]))])
+        net.model.set_params(net.arg_params, net.aux_params)
+        nets.append(net)
+
+    facescrub_out = os.path.join(args.output, 'facescrub')
+    megaface_out = os.path.join(args.output, 'megaface')
+
+    i = 0
+    succ = 0
+    buffer = []
+    for line in open(args.facescrub_lst, 'r'):
+        if i % 1000 == 0:
+            print("writing fs", i, succ)
+        i += 1
+        image_path = line.strip()
+        _path = image_path.split('/')
+        a, b = _path[-2], _path[-1]
+        out_dir = os.path.join(facescrub_out, a)
+        if not os.path.exists(out_dir):
+            os.makedirs(out_dir)
+        image_path = os.path.join(args.facescrub_root, image_path)
+        img = read_img(image_path)
+        if img is None:
+            print('read error:', image_path)
+            continue
+        out_path = os.path.join(out_dir, b + "_%s.bin" % (args.algo))
+        item = (img, out_path)
+        buffer.append(item)
+        if len(buffer) == args.batch_size:
+            get_and_write(buffer, nets)
+            buffer = []
+        succ += 1
+    if len(buffer) > 0:
+        get_and_write(buffer, nets)
+        buffer = []
+    print('fs stat', i, succ)
+
+    i = 0
+    succ = 0
+    buffer = []
+    for line in open(args.megaface_lst, 'r'):
+        if i % 1000 == 0:
+            print("writing mf", i, succ)
+        i += 1
+        image_path = line.strip()
+        _path = image_path.split('/')
+        a1, a2, b = _path[-3], _path[-2], _path[-1]
+        out_dir = os.path.join(megaface_out, a1, a2)
+        if not os.path.exists(out_dir):
+            os.makedirs(out_dir)
+            #continue
+        #print(landmark)
+        image_path = os.path.join(args.megaface_root, image_path)
+        img = read_img(image_path)
+        if img is None:
+            print('read error:', image_path)
+            continue
+        out_path = os.path.join(out_dir, b + "_%s.bin" % (args.algo))
+        item = (img, out_path)
+        buffer.append(item)
+        if len(buffer) == args.batch_size:
+            get_and_write(buffer, nets)
+            buffer = []
+        succ += 1
+    if len(buffer) > 0:
+        get_and_write(buffer, nets)
+        buffer = []
+    print('mf stat', i, succ)
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--batch_size', type=int, help='', default=8)
+    parser.add_argument('--image_size', type=str, help='', default='3,112,112')
+    parser.add_argument('--gpu', type=int, help='', default=0)
+    parser.add_argument('--algo', type=str, help='', default='insightface')
+    parser.add_argument('--facescrub-lst',
+                        type=str,
+                        help='',
+                        default='./data/facescrub_lst')
+    parser.add_argument('--megaface-lst',
+                        type=str,
+                        help='',
+                        default='./data/megaface_lst')
+    parser.add_argument('--facescrub-root',
+                        type=str,
+                        help='',
+                        default='./data/facescrub_images')
+    parser.add_argument('--megaface-root',
+                        type=str,
+                        help='',
+                        default='./data/megaface_images')
+    parser.add_argument('--output', type=str, help='', default='./feature_out')
+    parser.add_argument('--model', type=str, help='', default='')
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    main(parse_arguments(sys.argv[1:]))
diff --git a/recognition/_evaluation_/megaface/remove_noises.py b/recognition/_evaluation_/megaface/remove_noises.py
new file mode 100644
index 000000000..aacec5404
--- /dev/null
+++ b/recognition/_evaluation_/megaface/remove_noises.py
@@ -0,0 +1,182 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import datetime
+import time
+import shutil
+import sys
+import numpy as np
+import argparse
+import struct
+import cv2
+import mxnet as mx
+from mxnet import ndarray as nd
+
+feature_dim = 512
+feature_ext = 1
+
+
+def load_bin(path, fill=0.0):
+    with open(path, 'rb') as f:
+        bb = f.read(4 * 4)
+        #print(len(bb))
+        v = struct.unpack('4i', bb)
+        #print(v[0])
+        bb = f.read(v[0] * 4)
+        v = struct.unpack("%df" % (v[0]), bb)
+        feature = np.full((feature_dim + feature_ext, ),
+                          fill,
+                          dtype=np.float32)
+        feature[0:feature_dim] = v
+        #feature = np.array( v, dtype=np.float32)
+    #print(feature.shape)
+    #print(np.linalg.norm(feature))
+    return feature
+
+
+def write_bin(path, feature):
+    feature = list(feature)
+    with open(path, 'wb') as f:
+        f.write(struct.pack('4i', len(feature), 1, 4, 5))
+        f.write(struct.pack("%df" % len(feature), *feature))
+
+
+def main(args):
+
+    fs_noise_map = {}
+    for line in open(args.facescrub_noises, 'r'):
+        if line.startswith('#'):
+            continue
+        line = line.strip()
+        fname = line.split('.')[0]
+        p = fname.rfind('_')
+        fname = fname[0:p]
+        fs_noise_map[line] = fname
+
+    print(len(fs_noise_map))
+
+    i = 0
+    fname2center = {}
+    noises = []
+    for line in open(args.facescrub_lst, 'r'):
+        if i % 1000 == 0:
+            print("reading fs", i)
+        i += 1
+        image_path = line.strip()
+        _path = image_path.split('/')
+        a, b = _path[-2], _path[-1]
+        feature_path = os.path.join(args.feature_dir_input, 'facescrub', a,
+                                    "%s_%s.bin" % (b, args.algo))
+        feature_dir_out = os.path.join(args.feature_dir_out, 'facescrub', a)
+        if not os.path.exists(feature_dir_out):
+            os.makedirs(feature_dir_out)
+        feature_path_out = os.path.join(feature_dir_out,
+                                        "%s_%s.bin" % (b, args.algo))
+        #print(b)
+        if not b in fs_noise_map:
+            #shutil.copyfile(feature_path, feature_path_out)
+            feature = load_bin(feature_path)
+            write_bin(feature_path_out, feature)
+            if not a in fname2center:
+                fname2center[a] = np.zeros((feature_dim + feature_ext, ),
+                                           dtype=np.float32)
+            fname2center[a] += feature
+        else:
+            #print('n', b)
+            noises.append((a, b))
+    print(len(noises))
+
+    for k in noises:
+        a, b = k
+        assert a in fname2center
+        center = fname2center[a]
+        g = np.zeros((feature_dim + feature_ext, ), dtype=np.float32)
+        g2 = np.random.uniform(-0.001, 0.001, (feature_dim, ))
+        g[0:feature_dim] = g2
+        f = center + g
+        _norm = np.linalg.norm(f)
+        f /= _norm
+        feature_path_out = os.path.join(args.feature_dir_out, 'facescrub', a,
+                                        "%s_%s.bin" % (b, args.algo))
+        write_bin(feature_path_out, f)
+
+    mf_noise_map = {}
+    for line in open(args.megaface_noises, 'r'):
+        if line.startswith('#'):
+            continue
+        line = line.strip()
+        _vec = line.split("\t")
+        if len(_vec) > 1:
+            line = _vec[1]
+        mf_noise_map[line] = 1
+
+    print(len(mf_noise_map))
+
+    i = 0
+    nrof_noises = 0
+    for line in open(args.megaface_lst, 'r'):
+        if i % 1000 == 0:
+            print("reading mf", i)
+        i += 1
+        image_path = line.strip()
+        _path = image_path.split('/')
+        a1, a2, b = _path[-3], _path[-2], _path[-1]
+        feature_path = os.path.join(args.feature_dir_input, 'megaface', a1, a2,
+                                    "%s_%s.bin" % (b, args.algo))
+        feature_dir_out = os.path.join(args.feature_dir_out, 'megaface', a1,
+                                       a2)
+        if not os.path.exists(feature_dir_out):
+            os.makedirs(feature_dir_out)
+        feature_path_out = os.path.join(feature_dir_out,
+                                        "%s_%s.bin" % (b, args.algo))
+        bb = '/'.join([a1, a2, b])
+        #print(b)
+        if not bb in mf_noise_map:
+            feature = load_bin(feature_path)
+            write_bin(feature_path_out, feature)
+            #shutil.copyfile(feature_path, feature_path_out)
+        else:
+            feature = load_bin(feature_path, 100.0)
+            write_bin(feature_path_out, feature)
+            #g = np.random.uniform(-0.001, 0.001, (feature_dim,))
+            #print('n', bb)
+            #write_bin(feature_path_out, g)
+            nrof_noises += 1
+    print(nrof_noises)
+
+
+def parse_arguments(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--facescrub-noises',
+                        type=str,
+                        help='',
+                        default='./data/facescrub_noises.txt')
+    parser.add_argument('--megaface-noises',
+                        type=str,
+                        help='',
+                        default='./data/megaface_noises.txt')
+    parser.add_argument('--algo', type=str, help='', default='insightface')
+    parser.add_argument('--facescrub-lst',
+                        type=str,
+                        help='',
+                        default='./data/facescrub_lst')
+    parser.add_argument('--megaface-lst',
+                        type=str,
+                        help='',
+                        default='./data/megaface_lst')
+    parser.add_argument('--feature-dir-input',
+                        type=str,
+                        help='',
+                        default='./feature_out')
+    parser.add_argument('--feature-dir-out',
+                        type=str,
+                        help='',
+                        default='./feature_out_clean')
+    return parser.parse_args(argv)
+
+
+if __name__ == '__main__':
+    main(parse_arguments(sys.argv[1:]))
diff --git a/Evaluation/Megaface/run.sh b/recognition/_evaluation_/megaface/run.sh
similarity index 100%
rename from Evaluation/Megaface/run.sh
rename to recognition/_evaluation_/megaface/run.sh
diff --git a/recognition/_tools_/README.md b/recognition/_tools_/README.md
new file mode 100644
index 000000000..4f5facee5
--- /dev/null
+++ b/recognition/_tools_/README.md
@@ -0,0 +1,71 @@
+
+## I. CPP-Align
+ 
+  -
+
+## II. Face Mask Renderer
+
+We provide a simple tool to add masks on face images automatically.
+
+We can use this tool to do data augmentation while training our face recognition models.
+
+| Face Image  | OP | Mask Image | Out |  
+| ------- | ------ | --------- | ----------- |  
+|  <img src="https://github.com/deepinsight/insightface/blob/master/python-package/insightface/data/images/Tom_Hanks_54745.png" alt="face" height="112" /> | +F  | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/mask1.jpg" alt="mask" height="112" />     | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/mask_out1.jpg?raw=true" alt="mask" height="112" />      | 
+|  <img src="https://github.com/deepinsight/insightface/blob/master/python-package/insightface/data/images/Tom_Hanks_54745.png" alt="face" height="112" /> | +F  | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/black-mask.png" alt="mask" height="112" />     | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/mask_out3.jpg?raw=true" alt="mask" height="112" />      | 
+|  <img src="https://github.com/deepinsight/insightface/blob/master/python-package/insightface/data/images/Tom_Hanks_54745.png" alt="face" height="112" /> | +H  | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/mask2.jpg?raw=true" alt="mask" height="112" />     | <img src="https://github.com/nttstar/insightface-resources/blob/master/images/mask_out2h.jpg?raw=true" alt="mask" height="112" />      | 
+
+**F** means FULL while **H** means HALF.
+
+### Prepare
+
+- insightface package library
+
+   ``pip install -U insightface``
+
+- insightface model pack
+
+  ``bash> insightface-cli model.download antelope``
+  
+- BFM models
+
+   Please follow the tutorial of [https://github.com/YadiraF/face3d/tree/master/examples/Data/BFM](https://github.com/YadiraF/face3d/tree/master/examples/Data/BFM) to generate `BFM.mat` and `BFM_UV.mat`. Put them into the insightface model pack directory, such as ``~/.insightface/models/antelope/``
+   
+   
+- mask images
+
+   some mask images are included in insightface package, such as 'mask\_blue', 'mask\_white', 'mask\_black' and 'mask\_green'.
+   
+### Add Mask to Face Image
+
+Please refer to `make_renderer.py` for detail example. 
+
+(1) init renderer:
+```
+import insightface
+from insightface.app import MaskRenderer
+tool = MaskRenderer()
+tool.prepare(ctx_id=0, det_size=(128,128)) #use gpu
+```
+
+(2) load face and mask images
+```
+from insightface.data import get_image as ins_get_image
+image = ins_get_image('Tom_Hanks_54745')
+mask_image  = "mask_blue"
+```
+
+(3) build necessary params for face image, this can be done in offline.
+```
+params = tool.build_params(image)
+```
+
+(4) do mask render, it costs about `10ms` on 224x224 UV size, CPU single thread.
+```
+mask_out = tool.render_mask(image, mask_image, params)
+```
+
+(5) do half mask render.
+```
+mask_half_out = tool.render_mask(image, mask_image, params, positions=[0.1, 0.5, 0.9, 0.7])
+```
diff --git a/recognition/_tools_/cpp_align/face_align.h b/recognition/_tools_/cpp_align/face_align.h
new file mode 100644
index 000000000..afef20baf
--- /dev/null
+++ b/recognition/_tools_/cpp_align/face_align.h
@@ -0,0 +1,146 @@
+//
+// Created by Jack Yu on 23/03/2018.
+//
+
+#ifndef FACE_DEMO_FACEPREPROCESS_H
+#define FACE_DEMO_FACEPREPROCESS_H
+
+#include<opencv2/opencv.hpp>
+
+
+namespace FacePreprocess {
+
+    cv::Mat meanAxis0(const cv::Mat &src)
+    {
+        int num = src.rows;
+        int dim = src.cols;
+
+        // x1 y1
+        // x2 y2
+
+        cv::Mat output(1,dim,CV_32F);
+        for(int i = 0 ; i <  dim; i ++)
+        {
+            float sum = 0 ;
+            for(int j = 0 ; j < num ; j++)
+            {
+                sum+=src.at<float>(j,i);
+            }
+            output.at<float>(0,i) = sum/num;
+        }
+
+        return output;
+    }
+
+    cv::Mat elementwiseMinus(const cv::Mat &A,const cv::Mat &B)
+    {
+        cv::Mat output(A.rows,A.cols,A.type());
+
+        assert(B.cols == A.cols);
+        if(B.cols == A.cols)
+        {
+            for(int i = 0 ; i <  A.rows; i ++)
+            {
+                for(int j = 0 ; j < B.cols; j++)
+                {
+                    output.at<float>(i,j) = A.at<float>(i,j) - B.at<float>(0,j);
+                }
+            }
+        }
+        return output;
+    }
+
+
+    cv::Mat varAxis0(const cv::Mat &src)
+    {
+        cv:Mat temp_ = elementwiseMinus(src,meanAxis0(src));
+        cv::multiply(temp_ ,temp_ ,temp_ );
+        return meanAxis0(temp_);
+
+    }
+
+
+
+    int MatrixRank(cv::Mat M)
+    {
+        Mat w, u, vt;
+        SVD::compute(M, w, u, vt);
+        Mat1b nonZeroSingularValues = w > 0.0001;
+        int rank = countNonZero(nonZeroSingularValues);
+        return rank;
+
+    }
+
+//    References
+//    ----------
+//    .. [1] "Least-squares estimation of transformation parameters between two
+//    point patterns", Shinji Umeyama, PAMI 1991, DOI: 10.1109/34.88573
+//
+//    """
+//
+//    Anthor:Jack Yu
+    cv::Mat similarTransform(cv::Mat src,cv::Mat dst) {
+        int num = src.rows;
+        int dim = src.cols;
+        cv::Mat src_mean = meanAxis0(src);
+        cv::Mat dst_mean = meanAxis0(dst);
+        cv::Mat src_demean = elementwiseMinus(src, src_mean);
+        cv::Mat dst_demean = elementwiseMinus(dst, dst_mean);
+        cv::Mat A = (dst_demean.t() * src_demean) / static_cast<float>(num);
+        cv::Mat d(dim, 1, CV_32F);
+        d.setTo(1.0f);
+        if (cv::determinant(A) < 0) {
+            d.at<float>(dim - 1, 0) = -1;
+
+        }
+        cv::Mat T = cv::Mat::eye(dim + 1, dim + 1, CV_32F);
+        cv::Mat U, S, V;
+        cv::SVD::compute(A, S,U, V);
+
+        // the SVD function in opencv differ from scipy .
+
+
+        int rank = MatrixRank(A);
+        if (rank == 0) {
+            assert(rank == 0);
+
+        } else if (rank == dim - 1) {
+            if (cv::determinant(U) * cv::determinant(V) > 0) {
+                T.rowRange(0, dim).colRange(0, dim) = U * V;
+            } else {
+                int s = d.at<float>(dim - 1, 0) = -1;
+                d.at<float>(dim - 1, 0) = -1;
+
+                T.rowRange(0, dim).colRange(0, dim) = U * V;
+                cv::Mat diag_ = cv::Mat::diag(d);
+                cv::Mat twp = diag_*V; //np.dot(np.diag(d), V.T)
+                cv::Mat B = cv::Mat::zeros(3, 3, CV_8UC1);
+                cv::Mat C = B.diag(0);
+                T.rowRange(0, dim).colRange(0, dim) = U* twp;
+                d.at<float>(dim - 1, 0) = s;
+            }
+        }
+        else{
+            cv::Mat diag_ = cv::Mat::diag(d);
+            cv::Mat twp = diag_*V.t(); //np.dot(np.diag(d), V.T)
+            cv::Mat res = U* twp; // U
+            T.rowRange(0, dim).colRange(0, dim) = U *  diag_ * V;
+        }
+        cv::Mat var_ = varAxis0(src_demean);
+        float val = cv::sum(var_).val[0];
+        cv::Mat res;
+        cv::multiply(d,S,res);
+        float scale =  1.0/val*cv::sum(res).val[0];
+        cv::Mat  temp1 = T.rowRange(0, dim).colRange(0, dim) * src_mean.t();
+        cv::Mat  temp2 = scale * temp1;
+        cv::Mat  temp3 = dst_mean - temp2.t();
+        T.at<float>(0,2) = temp3.at<float>(0);
+        T.at<float>(1,2) = temp3.at<float>(1);
+        T.rowRange(0, dim).colRange(0, dim) *= scale; // T[:dim, :dim] *= scale
+
+        return T;
+    }
+
+
+}
+#endif //FACE_DEMO_FACEPREPROCESS_H
diff --git a/recognition/_tools_/mask_renderer.py b/recognition/_tools_/mask_renderer.py
new file mode 100644
index 000000000..57fa865f3
--- /dev/null
+++ b/recognition/_tools_/mask_renderer.py
@@ -0,0 +1,21 @@
+import os, sys, datetime
+import numpy as np
+import os.path as osp
+import cv2
+import insightface
+from insightface.app import MaskRenderer
+
+
+if __name__ == "__main__":
+    #make sure that you have download correct insightface model pack.
+    #make sure that BFM.mat and BFM_UV.mat have been generated
+    tool = MaskRenderer()
+    tool.prepare(ctx_id=0, det_size=(128,128))
+    image = cv2.imread("../../deploy/Tom_Hanks_54745.png")
+    mask_image  = "mask_blue"
+    params = tool.build_params(image)
+    mask_out = tool.render_mask(image, mask_image, params)
+
+    cv2.imwrite('output_mask.jpg', mask_out)
+
+
diff --git a/recognition/arcface_mxnet/README.md b/recognition/arcface_mxnet/README.md
new file mode 100644
index 000000000..3d463f74e
--- /dev/null
+++ b/recognition/arcface_mxnet/README.md
@@ -0,0 +1,95 @@
+## ArcFace with Parallel Acceleration on both Features and Centers
+
+### Memory Consumption and Training Speed
+
+![Memoryspeed](https://insightface.ai/assets/img/github/memoryspeed.png)
+
+Parallel acceleration on both feature x and centre W. Setting: ResNet 50, batch size 8 * 64, feature dimension 512, float point 32, GPU 8 * P40 (24GB).
+
+### Illustration of Main Steps
+
+![Memoryspeed](https://insightface.ai/assets/img/github/mainsteps.png)
+
+Parallel calculation by simple matrix partition. Setting: ResNet 50, batch size 8 * 64, feature dimension 512, float point 32, identity number 1 Million, GPU 8 * 1080ti (11GB). Communication cost: 1MB (feature x). Training speed: 800 samples/second.
+
+**Note:** Replace ``train.py`` with ``train_parall.py`` in following examples if you want to use parallel acceleration.
+
+### Model Training
+
+1. Install `MXNet` with GPU support.
+
+```
+pip install mxnet-cu100 # mxnet-cu102
+```
+
+2. Clone the InsightFace repository. We call the directory insightface as *`INSIGHTFACE_ROOT`*.
+
+```
+git clone --recursive https://github.com/deepinsight/insightface.git
+```
+
+3. Download the training set (`MS1MV2-Arcface`) and place it in *`$INSIGHTFACE_ROOT/recognition/datasets/`*. Each training dataset includes the following 6 files:
+
+```Shell
+    faces_emore/
+       train.idx
+       train.rec
+       property
+       lfw.bin
+       cfp_fp.bin
+       agedb_30.bin
+```
+
+The first three files are the training dataset while the last three files are verification sets.
+
+4. Train deep face recognition models.
+In this part, we assume you are in the directory *`$INSIGHTFACE_ROOT/recognition/ArcFace`*.
+
+Place and edit config file:
+```Shell
+cp sample_config.py config.py
+vim config.py # edit dataset path etc..
+```
+
+We give some examples below. Our experiments were conducted on the Tesla P40 GPU.
+
+(1). Train ArcFace with LResNet100E-IR.
+
+```Shell
+CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r100 --loss arcface --dataset emore
+```
+
+It will output verification results of *LFW*, *CFP-FP* and *AgeDB-30* every 2000 batches. You can check all options in *config.py*.
+This model can achieve *LFW 99.80+* and *MegaFace 98.3%+*.
+
+(2). Train CosineFace with LResNet50E-IR.
+
+```Shell
+CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r50 --loss cosface --dataset emore
+```
+
+(3). Train Softmax with MobileFaceNet.
+
+```Shell
+CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network y1 --loss softmax --dataset emore
+```
+
+(4). Fine-turn the above Softmax model with Triplet loss.
+
+```Shell
+CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network mnas05 --loss triplet --lr 0.005 --pretrained ./models/y1-softmax-emore,1
+```
+
+### Citation
+
+If you find *ArcFace* useful in your research, please consider to cite the following related papers:
+
+```
+@inproceedings{deng2019arcface,
+  title={Arcface: Additive angular margin loss for deep face recognition},
+  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
+  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
+  pages={4690--4699},
+  year={2019}
+}
+```
diff --git a/recognition/arcface_mxnet/common/build_eval_pack.py b/recognition/arcface_mxnet/common/build_eval_pack.py
new file mode 100644
index 000000000..23208ceeb
--- /dev/null
+++ b/recognition/arcface_mxnet/common/build_eval_pack.py
@@ -0,0 +1,136 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+#import mxnet as mx
+#from mxnet import ndarray as nd
+import argparse
+import cv2
+import pickle
+import numpy as np
+import sys
+import os
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', '..', 'common'))
+sys.path.append(
+    os.path.join(os.path.dirname(__file__), '..', '..', 'RetinaFace'))
+import face_align
+from retinaface import RetinaFace
+
+
+def to_rgb(img):
+    w, h = img.shape
+    ret = np.empty((w, h, 3), dtype=np.uint8)
+    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
+    return ret
+
+
+def IOU(Reframe, GTframe):
+    x1 = Reframe[0]
+    y1 = Reframe[1]
+    width1 = Reframe[2] - Reframe[0]
+    height1 = Reframe[3] - Reframe[1]
+
+    x2 = GTframe[0]
+    y2 = GTframe[1]
+    width2 = GTframe[2] - GTframe[0]
+    height2 = GTframe[3] - GTframe[1]
+
+    endx = max(x1 + width1, x2 + width2)
+    startx = min(x1, x2)
+    width = width1 + width2 - (endx - startx)
+
+    endy = max(y1 + height1, y2 + height2)
+    starty = min(y1, y2)
+    height = height1 + height2 - (endy - starty)
+
+    if width <= 0 or height <= 0:
+        ratio = 0
+    else:
+        Area = width * height
+        Area1 = width1 * height1
+        Area2 = width2 * height2
+        ratio = Area * 1. / (Area1 + Area2 - Area)
+    return ratio
+
+
+parser = argparse.ArgumentParser(description='Package eval images')
+# general
+parser.add_argument('--data-dir', default='', help='')
+parser.add_argument('--image-size', type=int, default=112, help='')
+parser.add_argument('--gpu', type=int, default=0, help='')
+parser.add_argument('--det-prefix', type=str, default='./model/R50', help='')
+parser.add_argument('--output', default='./', help='path to save.')
+parser.add_argument('--align-mode', default='arcface', help='align mode.')
+args = parser.parse_args()
+
+gpu_id = args.gpu
+
+detector = RetinaFace(args.det_prefix, 0, gpu_id, network='net3')
+target_size = 400
+max_size = 800
+
+
+def get_norm_crop(image_path):
+    im = cv2.imread(image_path)
+    im_shape = im.shape
+    im_size_min = np.min(im_shape[0:2])
+    im_size_max = np.max(im_shape[0:2])
+    im_scale = float(target_size) / float(im_size_min)
+    # prevent bigger axis from being more than max_size:
+    if np.round(im_scale * im_size_max) > max_size:
+        im_scale = float(max_size) / float(im_size_max)
+    bbox, landmark = detector.detect(im, threshold=0.5, scales=[im_scale])
+    #print(im.shape, bbox.shape, landmark.shape)
+    if bbox.shape[0] == 0:
+        bbox, landmark = detector.detect(
+            im,
+            threshold=0.05,
+            scales=[im_scale * 0.75, im_scale, im_scale * 2.0])
+        print('refine', im.shape, bbox.shape, landmark.shape)
+    nrof_faces = bbox.shape[0]
+    if nrof_faces > 0:
+        det = bbox[:, 0:4]
+        img_size = np.asarray(im.shape)[0:2]
+        bindex = 0
+        if nrof_faces > 1:
+            bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                           det[:, 1])
+            img_center = img_size / 2
+            offsets = np.vstack([(det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                                 (det[:, 1] + det[:, 3]) / 2 - img_center[0]])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            bindex = np.argmax(bounding_box_size - offset_dist_squared *
+                               2.0)  # some extra weight on the centering
+        #_bbox = bounding_boxes[bindex, 0:4]
+        _landmark = landmark[bindex]
+        warped = face_align.norm_crop(im,
+                                      landmark=_landmark,
+                                      image_size=args.image_size,
+                                      mode=args.align_mode)
+        return warped
+    else:
+        return None
+
+
+bins = []
+issame_list = []
+pp = 0
+for line in open(os.path.join(args.data_dir, 'pairs_label.txt'), 'r'):
+    pp += 1
+    if pp % 100 == 0:
+        print('processing', pp)
+    line = line.strip().split()
+    assert len(line) == 3
+    path1 = os.path.join(args.data_dir, line[0])
+    path2 = os.path.join(args.data_dir, line[1])
+    im1 = get_norm_crop(path1)
+    im2 = get_norm_crop(path2)
+    issame = True
+    if line[2] == '0':
+        issame = False
+    issame_list.append(issame)
+    for im in [im1, im2]:
+        _, s = cv2.imencode('.jpg', im)
+        bins.append(s)
+
+with open(args.output, 'wb') as f:
+    pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
diff --git a/recognition/arcface_mxnet/common/face_align.py b/recognition/arcface_mxnet/common/face_align.py
new file mode 100644
index 000000000..4f48a7691
--- /dev/null
+++ b/recognition/arcface_mxnet/common/face_align.py
@@ -0,0 +1,71 @@
+import cv2
+import numpy as np
+from skimage import transform as trans
+
+src1 = np.array([[51.642, 50.115], [57.617, 49.990], [35.740, 69.007],
+                 [51.157, 89.050], [57.025, 89.702]],
+                dtype=np.float32)
+#<--left
+src2 = np.array([[45.031, 50.118], [65.568, 50.872], [39.677, 68.111],
+                 [45.177, 86.190], [64.246, 86.758]],
+                dtype=np.float32)
+
+#---frontal
+src3 = np.array([[39.730, 51.138], [72.270, 51.138], [56.000, 68.493],
+                 [42.463, 87.010], [69.537, 87.010]],
+                dtype=np.float32)
+
+#-->right
+src4 = np.array([[46.845, 50.872], [67.382, 50.118], [72.737, 68.111],
+                 [48.167, 86.758], [67.236, 86.190]],
+                dtype=np.float32)
+
+#-->right profile
+src5 = np.array([[54.796, 49.990], [60.771, 50.115], [76.673, 69.007],
+                 [55.388, 89.702], [61.257, 89.050]],
+                dtype=np.float32)
+
+src = np.array([src1, src2, src3, src4, src5])
+src_map = {112: src, 224: src * 2}
+
+arcface_src = np.array(
+    [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+     [41.5493, 92.3655], [70.7299, 92.2041]],
+    dtype=np.float32)
+
+arcface_src = np.expand_dims(arcface_src, axis=0)
+
+# In[66]:
+
+
+# lmk is prediction; src is template
+def estimate_norm(lmk, image_size=112, mode='arcface'):
+    assert lmk.shape == (5, 2)
+    tform = trans.SimilarityTransform()
+    lmk_tran = np.insert(lmk, 2, values=np.ones(5), axis=1)
+    min_M = []
+    min_index = []
+    min_error = float('inf')
+    if mode == 'arcface':
+        assert image_size == 112
+        src = arcface_src
+    else:
+        src = src_map[image_size]
+    for i in np.arange(src.shape[0]):
+        tform.estimate(lmk, src[i])
+        M = tform.params[0:2, :]
+        results = np.dot(M, lmk_tran.T)
+        results = results.T
+        error = np.sum(np.sqrt(np.sum((results - src[i])**2, axis=1)))
+        #         print(error)
+        if error < min_error:
+            min_error = error
+            min_M = M
+            min_index = i
+    return min_M, min_index
+
+
+def norm_crop(img, landmark, image_size=112, mode='arcface'):
+    M, pose_index = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped
diff --git a/recognition/arcface_mxnet/common/flops_counter.py b/recognition/arcface_mxnet/common/flops_counter.py
new file mode 100644
index 000000000..809424142
--- /dev/null
+++ b/recognition/arcface_mxnet/common/flops_counter.py
@@ -0,0 +1,120 @@
+'''
+@author: insightface
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+import os
+import json
+import argparse
+import numpy as np
+import mxnet as mx
+
+
+def is_no_bias(attr):
+    ret = False
+    if 'no_bias' in attr and (attr['no_bias'] == True
+                              or attr['no_bias'] == 'True'):
+        ret = True
+    return ret
+
+
+def count_fc_flops(input_filter, output_filter, attr):
+    #print(input_filter, output_filter ,attr)
+    ret = 2 * input_filter * output_filter
+    if is_no_bias(attr):
+        ret -= output_filter
+    return int(ret)
+
+
+def count_conv_flops(input_shape, output_shape, attr):
+    kernel = attr['kernel'][1:-1].split(',')
+    kernel = [int(x) for x in kernel]
+
+    #print('kernel', kernel)
+    if is_no_bias(attr):
+        ret = (2 * input_shape[1] * kernel[0] * kernel[1] -
+               1) * output_shape[2] * output_shape[3] * output_shape[1]
+    else:
+        ret = 2 * input_shape[1] * kernel[0] * kernel[1] * output_shape[
+            2] * output_shape[3] * output_shape[1]
+    num_group = 1
+    if 'num_group' in attr:
+        num_group = int(attr['num_group'])
+    ret /= num_group
+    return int(ret)
+
+
+def count_flops(sym, **data_shapes):
+    all_layers = sym.get_internals()
+    #print(all_layers)
+    arg_shapes, out_shapes, aux_shapes = all_layers.infer_shape(**data_shapes)
+    out_shape_dict = dict(zip(all_layers.list_outputs(), out_shapes))
+
+    nodes = json.loads(sym.tojson())['nodes']
+    nodeid_shape = {}
+    for nodeid, node in enumerate(nodes):
+        name = node['name']
+        layer_name = name + "_output"
+        if layer_name in out_shape_dict:
+            nodeid_shape[nodeid] = out_shape_dict[layer_name]
+    #print(nodeid_shape)
+    FLOPs = 0
+    for nodeid, node in enumerate(nodes):
+        flops = 0
+        if node['op'] == 'Convolution':
+            output_shape = nodeid_shape[nodeid]
+            name = node['name']
+            attr = node['attrs']
+            input_nodeid = node['inputs'][0][0]
+            input_shape = nodeid_shape[input_nodeid]
+            flops = count_conv_flops(input_shape, output_shape, attr)
+        elif node['op'] == 'FullyConnected':
+            attr = node['attrs']
+            output_shape = nodeid_shape[nodeid]
+            input_nodeid = node['inputs'][0][0]
+            input_shape = nodeid_shape[input_nodeid]
+            output_filter = output_shape[1]
+            input_filter = input_shape[1] * input_shape[2] * input_shape[3]
+            #assert len(input_shape)==4 and input_shape[2]==1 and input_shape[3]==1
+            flops = count_fc_flops(input_filter, output_filter, attr)
+        #print(node, flops)
+        FLOPs += flops
+
+    return FLOPs
+
+
+def flops_str(FLOPs):
+    preset = [(1e12, 'T'), (1e9, 'G'), (1e6, 'M'), (1e3, 'K')]
+
+    for p in preset:
+        if FLOPs // p[0] > 0:
+            N = FLOPs / p[0]
+            ret = "%.1f%s" % (N, p[1])
+            return ret
+    ret = "%.1f" % (FLOPs)
+    return ret
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='flops counter')
+    # general
+    #parser.add_argument('--model', default='../models2/y2-arcface-retinat1/model,1', help='path to load model.')
+    #parser.add_argument('--model', default='../models2/r100fc-arcface-retinaa/model,1', help='path to load model.')
+    parser.add_argument('--model',
+                        default='../models2/r50fc-arcface-emore/model,1',
+                        help='path to load model.')
+    args = parser.parse_args()
+    _vec = args.model.split(',')
+    assert len(_vec) == 2
+    prefix = _vec[0]
+    epoch = int(_vec[1])
+    print('loading', prefix, epoch)
+    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+    all_layers = sym.get_internals()
+    sym = all_layers['fc1_output']
+    FLOPs = count_flops(sym, data=(1, 3, 112, 112))
+    print('FLOPs:', FLOPs)
diff --git a/recognition/arcface_mxnet/common/rec2image.py b/recognition/arcface_mxnet/common/rec2image.py
new file mode 100644
index 000000000..21e5ec482
--- /dev/null
+++ b/recognition/arcface_mxnet/common/rec2image.py
@@ -0,0 +1,60 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import sys
+import mxnet as mx
+from mxnet import ndarray as nd
+import random
+import argparse
+import cv2
+import time
+import sklearn
+import numpy as np
+
+
+def main(args):
+    include_datasets = args.include.split(',')
+    rec_list = []
+    for ds in include_datasets:
+        path_imgrec = os.path.join(ds, 'train.rec')
+        path_imgidx = os.path.join(ds, 'train.idx')
+        imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+        rec_list.append(imgrec)
+    if not os.path.exists(args.output):
+        os.makedirs(args.output)
+    for ds_id in range(len(rec_list)):
+        id_list = []
+        imgrec = rec_list[ds_id]
+        s = imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        assert header.flag > 0
+        print('header0 label', header.label)
+        header0 = (int(header.label[0]), int(header.label[1]))
+        seq_identity = range(int(header.label[0]), int(header.label[1]))
+        pp = 0
+        for identity in seq_identity:
+            id_dir = os.path.join(args.output, "%d_%d" % (ds_id, identity))
+            os.makedirs(id_dir)
+            pp += 1
+            if pp % 10 == 0:
+                print('processing id', pp)
+            s = imgrec.read_idx(identity)
+            header, _ = mx.recordio.unpack(s)
+            imgid = 0
+            for _idx in range(int(header.label[0]), int(header.label[1])):
+                s = imgrec.read_idx(_idx)
+                _header, _img = mx.recordio.unpack(s)
+                _img = mx.image.imdecode(_img).asnumpy()[:, :, ::-1]  # to bgr
+                image_path = os.path.join(id_dir, "%d.jpg" % imgid)
+                cv2.imwrite(image_path, _img)
+                imgid += 1
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='do dataset merge')
+    # general
+    parser.add_argument('--include', default='', type=str, help='')
+    parser.add_argument('--output', default='', type=str, help='')
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/arcface_mxnet/common/rec2shufrec.py b/recognition/arcface_mxnet/common/rec2shufrec.py
new file mode 100644
index 000000000..cf916b4ad
--- /dev/null
+++ b/recognition/arcface_mxnet/common/rec2shufrec.py
@@ -0,0 +1,72 @@
+import os
+import os.path as osp
+import sys
+import datetime
+import glob
+import shutil
+import numbers
+import mxnet as mx
+from mxnet import ndarray as nd
+from mxnet import io
+from mxnet import recordio
+import random
+import argparse
+import cv2
+import time
+import numpy as np
+
+def main(args):
+    ds = args.input
+    path_imgrec = osp.join(ds, 'train.rec')
+    path_imgidx = osp.join(ds, 'train.idx')
+    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+    if not osp.exists(args.output):
+        os.makedirs(args.output)
+    writer = mx.recordio.MXRecordIO(osp.join(args.output, 'train.rec'), 'w')
+    s = imgrec.read_idx(0)
+    header, _ = recordio.unpack(s)
+    if header.flag > 0:
+        print('header0 label', header.label)
+        header0 = (int(header.label[0]), int(header.label[1]))
+        imgidx = list(range(1, int(header.label[0])))
+    else:
+        imgidx = list(imgrec.keys)
+    random.shuffle(imgidx)
+    label_stat = None
+    print('total images:', len(imgidx))
+    for i, idx in enumerate(imgidx):
+        if i%10000==0:
+            print('processing', i, idx)
+        s = imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        label = header.label
+        if not isinstance(label, numbers.Number):
+            label = label[0]
+        if label_stat is None:
+            label_stat = [label, label]
+        else:
+            label_stat[0] = min(label, label_stat[0])
+            label_stat[1] = max(label, label_stat[1])
+        wheader = mx.recordio.IRHeader(0, label, i, 0)
+        ws = mx.recordio.pack(wheader, img)
+        writer.write(ws)
+    print('label_stat:', label_stat)
+    writer.close()
+    if args.copy_vers:
+        for binfile in glob.glob(osp.join(args.input, '*.bin')):
+            target_file = osp.join(args.output, binfile.split('/')[-1])
+            shutil.copyfile(binfile, target_file)
+    with open(osp.join(args.output, 'property'), 'w') as f:
+        f.write("%d,112,112\n"%(int(label_stat[1])+1))
+        f.write("%d\n"%len(imgidx))
+        f.write("shuffled\n")
+        f.write("%s\n"%(datetime.datetime.now()))
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='convert rec to shuffled rec')
+    # general
+    parser.add_argument('--input', default='', type=str, help='')
+    parser.add_argument('--output', default='', type=str, help='')
+    parser.add_argument('--copy-vers', action='store_true', help='copy verification bins')
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/arcface_mxnet/common/rec_builder.py b/recognition/arcface_mxnet/common/rec_builder.py
new file mode 100644
index 000000000..1d517152f
--- /dev/null
+++ b/recognition/arcface_mxnet/common/rec_builder.py
@@ -0,0 +1,109 @@
+import os
+import sys
+import mxnet as mx
+from mxnet import ndarray as nd
+import random
+import argparse
+import cv2
+import time
+import sklearn
+import numpy as np
+
+
+class SeqRecBuilder():
+    def __init__(self, path, image_size=(112, 112)):
+        self.path = path
+        self.image_size = image_size
+        self.last_label = -1
+        self.widx = 0
+        if not os.path.exists(path):
+            os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(
+            os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'),
+            'w')
+        self.label_stat = [-1, -1]
+
+    def add(self, label, img, is_image=True):
+        #img should be BGR
+        #if self.sis:
+        #    assert label>=self.last_label
+        idx = self.widx
+        self.widx += 1
+        header = mx.recordio.IRHeader(0, label, idx, 0)
+        if is_image:
+            s = mx.recordio.pack_img(header, img, quality=95, img_fmt='.jpg')
+        else:
+            s = mx.recordio.pack(header, img)
+        self.writer.write_idx(idx, s)
+        if self.label_stat[0] < 0:
+            self.label_stat = [label, label]
+        else:
+            self.label_stat[0] = min(self.label_stat[0], label)
+            self.label_stat[1] = max(self.label_stat[1], label)
+
+    def close(self):
+        with open(os.path.join(self.path, 'property'), 'w') as f:
+            f.write("%d,%d,%d\n" % (self.label_stat[1] + 1, self.image_size[0],
+                                    self.image_size[1]))
+
+
+class RecBuilder():
+    def __init__(self, path, image_size=(112, 112)):
+        self.path = path
+        self.image_size = image_size
+        self.last_label = -1
+        self.widx = 1
+        if not os.path.exists(path):
+            os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(
+            os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'),
+            'w')
+        self.label_stat = [-1, -1]
+        self.identities = []
+
+    def add(self, label, imgs):
+        #img should be BGR
+        assert label >= 0
+        assert label > self.last_label
+        assert len(imgs) > 0
+        idflag = [self.widx, -1]
+        for img in imgs:
+            idx = self.widx
+            self.widx += 1
+            header = mx.recordio.IRHeader(0, label, idx, 0)
+            if isinstance(img, np.ndarray):
+                s = mx.recordio.pack_img(header,
+                                         img,
+                                         quality=95,
+                                         img_fmt='.jpg')
+            else:
+                s = mx.recordio.pack(header, img)
+            self.writer.write_idx(idx, s)
+        idflag[1] = self.widx
+        self.identities.append(idflag)
+        if self.label_stat[0] < 0:
+            self.label_stat = [label, label]
+        else:
+            self.label_stat[0] = min(self.label_stat[0], label)
+            self.label_stat[1] = max(self.label_stat[1], label)
+        self.last_label = label
+
+    def close(self):
+        id_idx = self.widx
+        for id_flag in self.identities:
+            idx = self.widx
+            self.widx += 1
+            _header = mx.recordio.IRHeader(0, id_flag, idx, 0)
+            s = mx.recordio.pack(_header, b'')
+            self.writer.write_idx(idx, s)
+
+        print('id0:', (id_idx, self.widx))
+        idx = 0
+        _header = mx.recordio.IRHeader(0, (id_idx, self.widx), idx, 1)
+        s = mx.recordio.pack(_header, b'')
+        self.writer.write_idx(idx, s)
+        print('label stat:', self.label_stat)
+
+        with open(os.path.join(self.path, 'property'), 'w') as f:
+            f.write("%d,%d,%d\n" % (self.label_stat[1] + 1, self.image_size[0],
+                                    self.image_size[1]))
diff --git a/recognition/arcface_mxnet/common/verification.py b/recognition/arcface_mxnet/common/verification.py
new file mode 100644
index 000000000..f46942a25
--- /dev/null
+++ b/recognition/arcface_mxnet/common/verification.py
@@ -0,0 +1,423 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset 
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import argparse
+import sys
+import numpy as np
+from scipy import misc
+from sklearn.model_selection import KFold
+from scipy import interpolate
+import sklearn
+import cv2
+import math
+import datetime
+import pickle
+from sklearn.decomposition import PCA
+import mxnet as mx
+from mxnet import ndarray as nd
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+    #print('pca', pca)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        #print('train_set', train_set)
+        #print('test_set', test_set)
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            #print(_embed_train.shape)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            #print(embed1.shape, embed2.shape)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        #print('threshold', thresholds[best_threshold_index])
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    #print(true_accept, false_accept)
+    #print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
+
+
+def load_bin(path, image_size):
+    try:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f)  #py2
+    except UnicodeDecodeError as e:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f, encoding='bytes')  #py3
+    data_list = []
+    for flip in [0, 1]:
+        data = nd.empty(
+            (len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for i in range(len(issame_list) * 2):
+        _bin = bins[i]
+        img = mx.image.imdecode(_bin)
+        if img.shape[1] != image_size[0]:
+            img = mx.image.resize_short(img, image_size[0])
+        img = nd.transpose(img, axes=(2, 0, 1))
+        for flip in [0, 1]:
+            if flip == 1:
+                img = mx.ndarray.flip(data=img, axis=2)
+            data_list[flip][i][:] = img
+        if i % 1000 == 0:
+            print('loading bin', i)
+    print(data_list[0].shape)
+    return (data_list, issame_list)
+
+
+def test(data_set,
+         mx_model,
+         batch_size,
+         nfolds=10,
+         data_extra=None,
+         label_shape=None):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    model = mx_model
+    embeddings_list = []
+    if data_extra is not None:
+        _data_extra = nd.array(data_extra)
+    time_consumed = 0.0
+    if label_shape is None:
+        _label = nd.ones((batch_size, ))
+    else:
+        _label = nd.ones(label_shape)
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            #print(_data.shape, _label.shape)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            #_arg, _aux = model.get_params()
+            #__arg = {}
+            #for k,v in _arg.iteritems():
+            #  __arg[k] = v.as_in_context(_ctx)
+            #_arg = __arg
+            #_arg["data"] = _data.as_in_context(_ctx)
+            #_arg["softmax_label"] = _label.as_in_context(_ctx)
+            #for k,v in _arg.iteritems():
+            #  print(k,v.context)
+            #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
+            #exe.forward(is_train=False)
+            #net_out = exe.outputs
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            #print(_embeddings.shape)
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            #print(_em.shape, _norm)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
+    #acc1, std1 = np.mean(accuracy), np.std(accuracy)
+
+    #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
+    #embeddings = np.concatenate(embeddings_list, axis=1)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    print('infer time', time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser(description='do verification')
+    # general
+    parser.add_argument('--data-dir', default='', help='')
+    parser.add_argument('--model',
+                        default='../model/softmax,50',
+                        help='path to load model.')
+    parser.add_argument('--target',
+                        default='lfw,cfp_ff,cfp_fp,agedb_30',
+                        help='test targets.')
+    parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+    parser.add_argument('--batch-size', default=32, type=int, help='')
+    parser.add_argument('--max', default='', type=str, help='')
+    parser.add_argument('--mode', default=0, type=int, help='')
+    parser.add_argument('--nfolds', default=10, type=int, help='')
+    args = parser.parse_args()
+    #sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+    #import face_image
+    #prop = face_image.load_property(args.data_dir)
+    #image_size = prop.image_size
+    image_size = [112, 112]
+    print('image_size', image_size)
+    ctx = mx.gpu(args.gpu)
+    nets = []
+    vec = args.model.split(',')
+    prefix = args.model.split(',')[0]
+    epochs = []
+    if len(vec) == 1:
+        pdir = os.path.dirname(prefix)
+        for fname in os.listdir(pdir):
+            if not fname.endswith('.params'):
+                continue
+            _file = os.path.join(pdir, fname)
+            if _file.startswith(prefix):
+                epoch = int(fname.split('.')[0].split('-')[1])
+                epochs.append(epoch)
+        epochs = sorted(epochs, reverse=True)
+        if len(args.max) > 0:
+            _max = [int(x) for x in args.max.split(',')]
+            assert len(_max) == 2
+            if len(epochs) > _max[1]:
+                epochs = epochs[_max[0]:_max[1]]
+
+    else:
+        epochs = [int(x) for x in vec[1].split('|')]
+    print('model number', len(epochs))
+    time0 = datetime.datetime.now()
+    for epoch in epochs:
+        print('loading', prefix, epoch)
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
+        all_layers = sym.get_internals()
+        sym = all_layers['fc1_output']
+        model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+        #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+        model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+                                          image_size[1]))])
+        model.set_params(arg_params, aux_params)
+        nets.append(model)
+    time_now = datetime.datetime.now()
+    diff = time_now - time0
+    print('model loading time', diff.total_seconds())
+
+    ver_list = []
+    ver_name_list = []
+    for name in args.target.split(','):
+        path = os.path.join(args.data_dir, name + ".bin")
+        if os.path.exists(path):
+            print('loading.. ', name)
+            data_set = load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+
+    if args.mode == 0:
+        for i in range(len(ver_list)):
+            results = []
+            for model in nets:
+                acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+                    ver_list[i], model, args.batch_size, args.nfolds)
+                print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
+                print('[%s]Accuracy: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc1, std1))
+                print('[%s]Accuracy-Flip: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc2, std2))
+                results.append(acc2)
+            print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
+    elif args.mode == 1:
+        model = nets[0]
+        test_badcase(ver_list[0], model, args.batch_size, args.target)
+    else:
+        model = nets[0]
+        dumpR(ver_list[0], model, args.batch_size, args.target)
diff --git a/recognition/arcface_mxnet/image_iter.py b/recognition/arcface_mxnet/image_iter.py
new file mode 100644
index 000000000..b1a77bf60
--- /dev/null
+++ b/recognition/arcface_mxnet/image_iter.py
@@ -0,0 +1,367 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import random
+import logging
+import sys
+import numbers
+import math
+import sklearn
+import datetime
+import numpy as np
+import cv2
+
+import mxnet as mx
+from mxnet import ndarray as nd
+from mxnet import io
+from mxnet import recordio
+
+logger = logging.getLogger()
+
+
+class FaceImageIter(io.DataIter):
+    def __init__(self,
+                 batch_size,
+                 data_shape,
+                 path_imgrec=None,
+                 shuffle=False,
+                 aug_list=None,
+                 mean=None,
+                 rand_mirror=False,
+                 cutoff=0,
+                 color_jittering=0,
+                 images_filter=0,
+                 data_name='data',
+                 label_name='softmax_label',
+                 **kwargs):
+        super(FaceImageIter, self).__init__()
+        assert path_imgrec
+        if path_imgrec:
+            logging.info('loading recordio %s...', path_imgrec)
+            path_imgidx = path_imgrec[0:-4] + ".idx"
+            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec,
+                                                     'r')  # pylint: disable=redefined-variable-type
+            s = self.imgrec.read_idx(0)
+            header, _ = recordio.unpack(s)
+            if header.flag > 0:
+                print('header0 label', header.label)
+                self.header0 = (int(header.label[0]), int(header.label[1]))
+                #assert(header.flag==1)
+                #self.imgidx = range(1, int(header.label[0]))
+                self.imgidx = []
+                self.id2range = {}
+                self.seq_identity = range(int(header.label[0]),
+                                          int(header.label[1]))
+                for identity in self.seq_identity:
+                    s = self.imgrec.read_idx(identity)
+                    header, _ = recordio.unpack(s)
+                    a, b = int(header.label[0]), int(header.label[1])
+                    count = b - a
+                    if count < images_filter:
+                        continue
+                    self.id2range[identity] = (a, b)
+                    self.imgidx += range(a, b)
+                print('id2range', len(self.id2range))
+            else:
+                self.imgidx = list(self.imgrec.keys)
+            if shuffle:
+                self.seq = self.imgidx
+                self.oseq = self.imgidx
+                print(len(self.seq))
+            else:
+                self.seq = None
+
+        self.mean = mean
+        self.nd_mean = None
+        if self.mean:
+            self.mean = np.array(self.mean, dtype=np.float32).reshape(1, 1, 3)
+            self.nd_mean = mx.nd.array(self.mean).reshape((1, 1, 3))
+
+        self.check_data_shape(data_shape)
+        self.provide_data = [(data_name, (batch_size, ) + data_shape)]
+        self.batch_size = batch_size
+        self.data_shape = data_shape
+        self.shuffle = shuffle
+        self.image_size = '%d,%d' % (data_shape[1], data_shape[2])
+        self.rand_mirror = rand_mirror
+        print('rand_mirror', rand_mirror)
+        self.cutoff = cutoff
+        self.color_jittering = color_jittering
+        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
+        self.provide_label = [(label_name, (batch_size, ))]
+        #print(self.provide_label[0][1])
+        self.cur = 0
+        self.nbatch = 0
+        self.is_init = False
+
+    def reset(self):
+        """Resets the iterator to the beginning of the data."""
+        print('call reset()')
+        self.cur = 0
+        if self.shuffle:
+            random.shuffle(self.seq)
+        if self.seq is None and self.imgrec is not None:
+            self.imgrec.reset()
+
+    def num_samples(self):
+        return len(self.seq)
+
+    def next_sample(self):
+        """Helper function for reading in next sample."""
+        #set total batch size, for example, 1800, and maximum size for each people, for example 45
+        if self.seq is not None:
+            while True:
+                if self.cur >= len(self.seq):
+                    raise StopIteration
+                idx = self.seq[self.cur]
+                self.cur += 1
+                if self.imgrec is not None:
+                    s = self.imgrec.read_idx(idx)
+                    header, img = recordio.unpack(s)
+                    label = header.label
+                    if not isinstance(label, numbers.Number):
+                        label = label[0]
+                    return label, img, None, None
+                else:
+                    label, fname, bbox, landmark = self.imglist[idx]
+                    return label, self.read_image(fname), bbox, landmark
+        else:
+            s = self.imgrec.read()
+            if s is None:
+                raise StopIteration
+            header, img = recordio.unpack(s)
+            return header.label, img, None, None
+
+    def brightness_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        src *= alpha
+        return src
+
+    def contrast_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = nd.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
+        src *= alpha
+        src += gray
+        return src
+
+    def saturation_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = nd.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = nd.sum(gray, axis=2, keepdims=True)
+        gray *= (1.0 - alpha)
+        src *= alpha
+        src += gray
+        return src
+
+    def color_aug(self, img, x):
+        #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
+        #random.shuffle(augs)
+        #for aug in augs:
+        #  #print(img.shape)
+        #  img = aug(img, x)
+        #  #print(img.shape)
+        #return img
+        return self.CJA(img)
+
+    def mirror_aug(self, img):
+        _rd = random.randint(0, 1)
+        if _rd == 1:
+            for c in range(img.shape[2]):
+                img[:, :, c] = np.fliplr(img[:, :, c])
+        return img
+
+    def compress_aug(self, img):
+        from PIL import Image
+        from io import BytesIO
+        buf = BytesIO()
+        img = Image.fromarray(img.asnumpy(), 'RGB')
+        q = random.randint(2, 20)
+        img.save(buf, format='JPEG', quality=q)
+        buf = buf.getvalue()
+        img = Image.open(BytesIO(buf))
+        return nd.array(np.asarray(img, 'float32'))
+
+    def next(self):
+        if not self.is_init:
+            self.reset()
+            self.is_init = True
+        """Returns the next batch of data."""
+        #print('in next', self.cur, self.labelcur)
+        self.nbatch += 1
+        batch_size = self.batch_size
+        c, h, w = self.data_shape
+        batch_data = nd.empty((batch_size, c, h, w))
+        if self.provide_label is not None:
+            batch_label = nd.empty(self.provide_label[0][1])
+        i = 0
+        try:
+            while i < batch_size:
+                label, s, bbox, landmark = self.next_sample()
+                _data = self.imdecode(s)
+                if _data.shape[0] != self.data_shape[1]:
+                    _data = mx.image.resize_short(_data, self.data_shape[1])
+                if self.rand_mirror:
+                    _rd = random.randint(0, 1)
+                    if _rd == 1:
+                        _data = mx.ndarray.flip(data=_data, axis=1)
+                if self.color_jittering > 0:
+                    if self.color_jittering > 1:
+                        _rd = random.randint(0, 1)
+                        if _rd == 1:
+                            _data = self.compress_aug(_data)
+                    #print('do color aug')
+                    _data = _data.astype('float32', copy=False)
+                    #print(_data.__class__)
+                    _data = self.color_aug(_data, 0.125)
+                if self.nd_mean is not None:
+                    _data = _data.astype('float32', copy=False)
+                    _data -= self.nd_mean
+                    _data *= 0.0078125
+                if self.cutoff > 0:
+                    _rd = random.randint(0, 1)
+                    if _rd == 1:
+                        #print('do cutoff aug', self.cutoff)
+                        centerh = random.randint(0, _data.shape[0] - 1)
+                        centerw = random.randint(0, _data.shape[1] - 1)
+                        half = self.cutoff // 2
+                        starth = max(0, centerh - half)
+                        endh = min(_data.shape[0], centerh + half)
+                        startw = max(0, centerw - half)
+                        endw = min(_data.shape[1], centerw + half)
+                        #print(starth, endh, startw, endw, _data.shape)
+                        _data[starth:endh, startw:endw, :] = 128
+                data = [_data]
+                try:
+                    self.check_valid_image(data)
+                except RuntimeError as e:
+                    logging.debug('Invalid image, skipping:  %s', str(e))
+                    continue
+                #print('aa',data[0].shape)
+                #data = self.augmentation_transform(data)
+                #print('bb',data[0].shape)
+                for datum in data:
+                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
+                    #print(datum.shape)
+                    batch_data[i][:] = self.postprocess_data(datum)
+                    batch_label[i][:] = label
+                    i += 1
+        except StopIteration:
+            if i < batch_size:
+                raise StopIteration
+
+        return io.DataBatch([batch_data], [batch_label], batch_size - i)
+
+    def check_data_shape(self, data_shape):
+        """Checks if the input data shape is valid"""
+        if not len(data_shape) == 3:
+            raise ValueError(
+                'data_shape should have length 3, with dimensions CxHxW')
+        if not data_shape[0] == 3:
+            raise ValueError(
+                'This iterator expects inputs to have 3 channels.')
+
+    def check_valid_image(self, data):
+        """Checks if the input data is valid"""
+        if len(data[0].shape) == 0:
+            raise RuntimeError('Data shape is wrong')
+
+    def imdecode(self, s):
+        """Decodes a string or byte string to an NDArray.
+        See mx.img.imdecode for more details."""
+        img = mx.image.imdecode(s)  #mx.ndarray
+        return img
+
+    def read_image(self, fname):
+        """Reads an input image `fname` and returns the decoded raw bytes.
+
+        Example usage:
+        ----------
+        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
+        """
+        with open(os.path.join(self.path_root, fname), 'rb') as fin:
+            img = fin.read()
+        return img
+
+    def augmentation_transform(self, data):
+        """Transforms input data with specified augmentation."""
+        for aug in self.auglist:
+            data = [ret for src in data for ret in aug(src)]
+        return data
+
+    def postprocess_data(self, datum):
+        """Final postprocessing step before image is loaded into the batch."""
+        return nd.transpose(datum, axes=(2, 0, 1))
+
+
+class FaceImageIterList(io.DataIter):
+    def __init__(self, iter_list):
+        assert len(iter_list) > 0
+        self.provide_data = iter_list[0].provide_data
+        self.provide_label = iter_list[0].provide_label
+        self.iter_list = iter_list
+        self.cur_iter = None
+
+    def reset(self):
+        self.cur_iter.reset()
+
+    def next(self):
+        self.cur_iter = random.choice(self.iter_list)
+        while True:
+            try:
+                ret = self.cur_iter.next()
+            except StopIteration:
+                self.cur_iter.reset()
+                continue
+            return ret
+
+def get_face_image_iter(cfg, data_shape, path_imgrec):
+    print('loading:', path_imgrec, cfg.is_shuffled_rec)
+    if not cfg.is_shuffled_rec:
+        train_dataiter = FaceImageIter(
+            batch_size=cfg.batch_size,
+            data_shape=data_shape,
+            path_imgrec=path_imgrec,
+            shuffle=True,
+            rand_mirror=cfg.data_rand_mirror,
+            mean=None,
+            cutoff=cfg.data_cutoff,
+            color_jittering=cfg.data_color,
+            images_filter=cfg.data_images_filter,
+        )
+        train_dataiter = mx.io.PrefetchingIter(train_dataiter)
+    else:
+        train_dataiter = mx.io.ImageRecordIter(
+               path_imgrec         = path_imgrec,
+               data_shape          = data_shape,
+               batch_size          = cfg.batch_size,
+               rand_mirror         = cfg.data_rand_mirror,
+               preprocess_threads  = 2,
+               shuffle             = True,
+               shuffle_chunk_size  = 1024,
+               )
+    return train_dataiter
+
+def test_face_image_iter(path_imgrec):
+    train_dataiter = mx.io.ImageRecordIter(
+           path_imgrec         = path_imgrec,
+           data_shape          = (3,112,112),
+           batch_size          = 512,
+           rand_mirror         = True,
+           preprocess_threads  = 2,
+           shuffle             = True,
+           shuffle_chunk_size  = 1024,
+           )
+    for batch in train_dataiter:
+        data = batch.data[0].asnumpy()
+        print(data.shape)
+        img0 = data[0]
+        print(img0[0,:5,:5])
+
+if __name__ == '__main__':
+    test_face_image_iter('/train_tmp/ms1mv3shuf/train.rec')
+
diff --git a/recognition/arcface_mxnet/metric.py b/recognition/arcface_mxnet/metric.py
new file mode 100644
index 000000000..9a64c16e4
--- /dev/null
+++ b/recognition/arcface_mxnet/metric.py
@@ -0,0 +1,50 @@
+import numpy as np
+import mxnet as mx
+
+
+class AccMetric(mx.metric.EvalMetric):
+    def __init__(self):
+        self.axis = 1
+        super(AccMetric, self).__init__('acc',
+                                        axis=self.axis,
+                                        output_names=None,
+                                        label_names=None)
+        self.losses = []
+        self.count = 0
+
+    def update(self, labels, preds):
+        self.count += 1
+        label = labels[0]
+        pred_label = preds[1]
+        #print('ACC', label.shape, pred_label.shape)
+        if pred_label.shape != label.shape:
+            pred_label = mx.ndarray.argmax(pred_label, axis=self.axis)
+        pred_label = pred_label.asnumpy().astype('int32').flatten()
+        label = label.asnumpy()
+        if label.ndim == 2:
+            label = label[:, 0]
+        label = label.astype('int32').flatten()
+        assert label.shape == pred_label.shape
+        self.sum_metric += (pred_label.flat == label.flat).sum()
+        self.num_inst += len(pred_label.flat)
+
+
+class LossValueMetric(mx.metric.EvalMetric):
+    def __init__(self):
+        self.axis = 1
+        super(LossValueMetric, self).__init__('lossvalue',
+                                              axis=self.axis,
+                                              output_names=None,
+                                              label_names=None)
+        self.losses = []
+
+    def update(self, labels, preds):
+        #label = labels[0].asnumpy()
+        pred = preds[-1].asnumpy()
+        #print('in loss', pred.shape)
+        #print(pred)
+        loss = pred[0]
+        self.sum_metric += loss
+        self.num_inst += 1.0
+        #gt_label = preds[-2].asnumpy()
+        #print(gt_label)
diff --git a/recognition/arcface_mxnet/parall_module_local_v1.py b/recognition/arcface_mxnet/parall_module_local_v1.py
new file mode 100644
index 000000000..394fe2f23
--- /dev/null
+++ b/recognition/arcface_mxnet/parall_module_local_v1.py
@@ -0,0 +1,612 @@
+'''
+@author: insightface
+'''
+
+import logging
+import copy
+import time
+import os
+
+import mxnet as mx
+import numpy as np
+from mxnet import context as ctx
+from mxnet.initializer import Uniform
+from mxnet.module.base_module import BaseModule
+from mxnet.module.module import Module
+from mxnet import metric
+from mxnet.model import BatchEndParam
+from mxnet import io
+import mxnet.ndarray as nd
+from config import config
+
+
+class ParallModule(BaseModule):
+    def __init__(self,
+                 symbol,
+                 data_names,
+                 label_names,
+                 logger=logging,
+                 context=ctx.cpu(),
+                 work_load_list=None,
+                 asymbol=None,
+                 args=None):
+        super(ParallModule, self).__init__(logger=logger)
+        self._symbol = symbol
+        self._asymbol = asymbol
+        self._data_names = data_names
+        self._label_names = label_names
+        self._context = context
+        self._work_load_list = work_load_list
+        self._num_classes = config.num_classes
+        self._batch_size = args.batch_size
+        self._verbose = args.verbose
+        self._emb_size = config.emb_size
+        self._local_class_start = args.local_class_start
+        self._iter = 0
+
+        self._curr_module = None
+
+        self._num_workers = config.num_workers
+        self._num_ctx = len(self._context)
+        self._ctx_num_classes = args.ctx_num_classes
+        self._nd_cache = {}
+        self._ctx_cpu = mx.cpu()
+        self._ctx_single_gpu = self._context[-1]
+        self._fixed_param_names = None
+        self._curr_module = Module(self._symbol,
+                                   self._data_names,
+                                   self._label_names,
+                                   logger=self.logger,
+                                   context=self._context,
+                                   work_load_list=self._work_load_list,
+                                   fixed_param_names=self._fixed_param_names)
+        self._arcface_modules = []
+        self._ctx_class_start = []
+        for i in range(len(self._context)):
+
+            args._ctxid = i
+            _module = Module(self._asymbol(args),
+                             self._data_names,
+                             self._label_names,
+                             logger=self.logger,
+                             context=mx.gpu(i),
+                             work_load_list=self._work_load_list,
+                             fixed_param_names=self._fixed_param_names)
+            self._arcface_modules.append(_module)
+            _c = args.local_class_start + i * args.ctx_num_classes
+            self._ctx_class_start.append(_c)
+        self._usekv = False
+        if self._usekv:
+            self._distkv = mx.kvstore.create('dist_sync')
+            self._kvinit = {}
+
+    def _reset_bind(self):
+        self.binded = False
+        self._curr_module = None
+
+    @property
+    def data_names(self):
+        return self._data_names
+
+    @property
+    def output_names(self):
+        return self._symbol.list_outputs()
+
+    @property
+    def data_shapes(self):
+        assert self.binded
+        return self._curr_module.data_shapes
+
+    @property
+    def label_shapes(self):
+        assert self.binded
+        return self._curr_module.label_shapes
+
+    @property
+    def output_shapes(self):
+        assert self.binded
+        return self._curr_module.output_shapes
+
+    def get_export_params(self):
+        assert self.binded and self.params_initialized
+        _g, _x = self._curr_module.get_params()
+        g = _g.copy()
+        x = _x.copy()
+        return g, x
+
+    def get_params(self):
+        assert self.binded and self.params_initialized
+        _g, _x = self._curr_module.get_params()
+        g = _g.copy()
+        x = _x.copy()
+        for _module in self._arcface_modules:
+            _g, _x = _module.get_params()
+            ag = _g.copy()
+            ax = _x.copy()
+            g.update(ag)
+            x.update(ax)
+        return g, x
+
+    def set_params(self,
+                   arg_params,
+                   aux_params,
+                   allow_missing=False,
+                   force_init=True,
+                   allow_extra=False):
+        g = arg_params
+        x = aux_params
+        #ag = {}
+        #ax = {}
+        rk = []
+        for k in g:
+            v = g[k]
+            if k.startswith('fc7'):
+                p1 = k.find('_')
+                p2 = k.rfind('_')
+                _ctxid = int(k[p1 + 1:p2])
+                self._arcface_modules[_ctxid].set_params({k: v}, {})
+                rk.append(k)
+        for k in rk:
+            del g[k]
+        self._curr_module.set_params(g, x)
+        #self._arcface_module.set_params(ag, ax)
+
+    def init_params(self,
+                    initializer=Uniform(0.01),
+                    arg_params=None,
+                    aux_params=None,
+                    allow_missing=False,
+                    force_init=False,
+                    allow_extra=False):
+        if self.params_initialized and not force_init:
+            return
+        assert self.binded, 'call bind before initializing the parameters'
+        #TODO init the same weights with all work nodes
+        self._curr_module.init_params(initializer=initializer,
+                                      arg_params=arg_params,
+                                      aux_params=aux_params,
+                                      allow_missing=allow_missing,
+                                      force_init=force_init,
+                                      allow_extra=allow_extra)
+        for _module in self._arcface_modules:
+            #_initializer = initializer
+            _initializer = mx.init.Normal(0.01)
+            _module.init_params(initializer=_initializer,
+                                arg_params=None,
+                                aux_params=None,
+                                allow_missing=allow_missing,
+                                force_init=force_init,
+                                allow_extra=allow_extra)
+        self.params_initialized = True
+
+    def bind(self,
+             data_shapes,
+             label_shapes=None,
+             for_training=True,
+             inputs_need_grad=False,
+             force_rebind=False,
+             shared_module=None):
+        print('in_bind', self.params_initialized, data_shapes, label_shapes)
+        if self.params_initialized:
+            arg_params, aux_params = self.get_params()
+
+        # force rebinding is typically used when one want to switch from
+        # training to prediction phase.
+        if force_rebind:
+            self._reset_bind()
+
+        if self.binded:
+            self.logger.warning('Already binded, ignoring bind()')
+            return
+
+        assert shared_module is None, 'shared_module for MutableModule is not supported'
+        self.for_training = for_training
+        self.inputs_need_grad = inputs_need_grad
+        self.binded = True
+        self._curr_module.bind(data_shapes,
+                               label_shapes,
+                               for_training,
+                               inputs_need_grad,
+                               force_rebind=False,
+                               shared_module=None)
+        _data_shape = data_shapes[0][1]
+        print('_data_shape', _data_shape, label_shapes)
+        for _module in self._arcface_modules:
+            _module.bind(
+                [('data',
+                  (_data_shape[0] * self._num_workers, self._emb_size))],
+                [('softmax_label', (_data_shape[0] * self._num_workers, ))],
+                for_training,
+                True,
+                force_rebind=False,
+                shared_module=None)
+        if self.params_initialized:
+            self.set_params(arg_params, aux_params)
+
+    def init_optimizer(self,
+                       kvstore='local',
+                       optimizer='sgd',
+                       optimizer_params=(('learning_rate', 0.01), ),
+                       force_init=False):
+        assert self.binded and self.params_initialized
+        if self.optimizer_initialized and not force_init:
+            self.logger.warning('optimizer already initialized, ignoring.')
+            return
+
+        self._curr_module.init_optimizer(kvstore,
+                                         optimizer[0],
+                                         optimizer_params,
+                                         force_init=force_init)
+        for _module in self._arcface_modules:
+            _module.init_optimizer(kvstore,
+                                   optimizer[1],
+                                   optimizer_params,
+                                   force_init=force_init)
+        self.optimizer_initialized = True
+
+    def kv_push(self, key, value):
+        #if value.context!=mx.cpu():
+        #  value = value.as_in_context(mx.cpu())
+        if not key in self._kvinit:
+            self._distkv.init(key, nd.zeros_like(value))
+            self._kvinit[key] = 1
+        self._distkv.push(key, value)
+
+    #get fc1 and partial fc7
+    def forward(self, data_batch, is_train=None):
+        #g,x = self.get_params()
+        #print('{fc7_weight[0][0]}', self._iter, g['fc7_0_weight'].asnumpy()[0][0])
+        #print('{pre_fc1_weight[0][0]}', self._iter, g['pre_fc1_weight'].asnumpy()[0][0])
+
+        assert self.binded and self.params_initialized
+        self._curr_module.forward(data_batch, is_train=is_train)
+        if is_train:
+            self._iter += 1
+            fc1, label = self._curr_module.get_outputs(
+                merge_multi_context=True)
+            global_fc1 = fc1
+            self.global_label = label.as_in_context(self._ctx_cpu)
+
+            for i, _module in enumerate(self._arcface_modules):
+                _label = self.global_label - self._ctx_class_start[i]
+                db_global_fc1 = io.DataBatch([global_fc1], [_label])
+                _module.forward(db_global_fc1)  #fc7 with margin
+        #print('forward end')
+
+    def get_ndarray(self, context, name, shape):
+        key = "%s_%s" % (name, context)
+        #print(key)
+        if not key in self._nd_cache:
+            v = nd.zeros(shape=shape, ctx=context)
+            self._nd_cache[key] = v
+        else:
+            v = self._nd_cache[key]
+        return v
+
+    def get_ndarray2(self, context, name, arr):
+        key = "%s_%s" % (name, context)
+        #print(key)
+        if not key in self._nd_cache:
+            v = nd.zeros(shape=arr.shape, ctx=context)
+            self._nd_cache[key] = v
+        else:
+            v = self._nd_cache[key]
+        arr.copyto(v)
+        return v
+
+    def backward(self, out_grads=None):
+        #print('in backward')
+        assert self.binded and self.params_initialized
+        #tmp_ctx = self._ctx_cpu
+        tmp_ctx = self._ctx_single_gpu
+        fc7_outs = []
+        ctx_fc7_max = self.get_ndarray(tmp_ctx, 'ctx_fc7_max',
+                                       (self._batch_size, len(self._context)))
+        #local_fc7_max = nd.zeros( (self.global_label.shape[0],1), ctx=mx.cpu())
+        for i, _module in enumerate(self._arcface_modules):
+            _fc7 = _module.get_outputs(merge_multi_context=True)[0]
+            fc7_outs.append(_fc7)
+            _fc7_max = nd.max(_fc7, axis=1).as_in_context(tmp_ctx)
+            ctx_fc7_max[:, i] = _fc7_max
+
+        local_fc7_max = self.get_ndarray(tmp_ctx, 'local_fc7_max',
+                                         (self._batch_size, 1))
+        nd.max(ctx_fc7_max, axis=1, keepdims=True, out=local_fc7_max)
+        global_fc7_max = local_fc7_max
+        #local_fc7_sum = None
+        local_fc7_sum = self.get_ndarray(tmp_ctx, 'local_fc7_sum',
+                                         (self._batch_size, 1))
+        local_fc7_sum[:, :] = 0.0
+        for i, _module in enumerate(self._arcface_modules):
+            _max = self.get_ndarray2(fc7_outs[i].context, 'fc7_max',
+                                     global_fc7_max)
+            fc7_outs[i] = nd.broadcast_sub(fc7_outs[i], _max)
+            fc7_outs[i] = nd.exp(fc7_outs[i])
+            _sum = nd.sum(fc7_outs[i], axis=1,
+                          keepdims=True).as_in_context(tmp_ctx)
+            local_fc7_sum += _sum
+        global_fc7_sum = local_fc7_sum
+
+        if self._iter % self._verbose == 0:
+            #_ctx = self._context[-1]
+            _ctx = self._ctx_cpu
+            _probs = []
+            for i, _module in enumerate(self._arcface_modules):
+                _prob = self.get_ndarray2(_ctx, '_fc7_prob_%d' % i,
+                                          fc7_outs[i])
+                _probs.append(_prob)
+            fc7_prob = self.get_ndarray(
+                _ctx, 'test_fc7_prob',
+                (self._batch_size, self._ctx_num_classes * len(self._context)))
+            nd.concat(*_probs, dim=1, out=fc7_prob)
+            fc7_pred = nd.argmax(fc7_prob, axis=1)
+            local_label = self.global_label - self._local_class_start
+            #local_label = self.get_ndarray2(_ctx, 'test_label', local_label)
+            _pred = nd.equal(fc7_pred, local_label)
+            print('{fc7_acc}', self._iter, nd.mean(_pred).asnumpy()[0])
+
+        #local_fc1_grad = []
+        #fc1_grad_ctx = self._ctx_cpu
+        fc1_grad_ctx = self._ctx_single_gpu
+        local_fc1_grad = self.get_ndarray(fc1_grad_ctx, 'local_fc1_grad',
+                                          (self._batch_size, self._emb_size))
+        local_fc1_grad[:, :] = 0.0
+
+        for i, _module in enumerate(self._arcface_modules):
+            _sum = self.get_ndarray2(fc7_outs[i].context, 'fc7_sum',
+                                     global_fc7_sum)
+            fc7_outs[i] = nd.broadcast_div(fc7_outs[i], _sum)
+            a = i * self._ctx_num_classes
+            b = (i + 1) * self._ctx_num_classes
+            _label = self.global_label - self._ctx_class_start[i]
+            _label = self.get_ndarray2(fc7_outs[i].context, 'label', _label)
+            onehot_label = self.get_ndarray(
+                fc7_outs[i].context, 'label_onehot',
+                (self._batch_size, self._ctx_num_classes))
+            nd.one_hot(_label,
+                       depth=self._ctx_num_classes,
+                       on_value=1.0,
+                       off_value=0.0,
+                       out=onehot_label)
+            fc7_outs[i] -= onehot_label
+            _module.backward(out_grads=[fc7_outs[i]])
+            #ctx_fc1_grad = _module.get_input_grads()[0].as_in_context(mx.cpu())
+            ctx_fc1_grad = self.get_ndarray2(fc1_grad_ctx,
+                                             'ctx_fc1_grad_%d' % i,
+                                             _module.get_input_grads()[0])
+            local_fc1_grad += ctx_fc1_grad
+
+        global_fc1_grad = local_fc1_grad
+        self._curr_module.backward(out_grads=[global_fc1_grad])
+
+    def update(self):
+        assert self.binded and self.params_initialized and self.optimizer_initialized
+        self._curr_module.update()
+        for i, _module in enumerate(self._arcface_modules):
+            _module.update()
+        mx.nd.waitall()
+
+    def get_outputs(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_outputs(
+            merge_multi_context=merge_multi_context)
+        #return self._arcface_module.get_outputs(merge_multi_context=merge_multi_context)
+
+    def get_input_grads(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized and self.inputs_need_grad
+        return self._curr_module.get_input_grads(
+            merge_multi_context=merge_multi_context)
+
+    def update_metric(self, eval_metric, labels):
+        assert self.binded and self.params_initialized
+        #self._curr_module.update_metric(eval_metric, labels)
+        #label = labels[0]
+        #print(label.shape)
+        #self._arcface_module.update_metric(eval_metric, labels)
+
+    def install_monitor(self, mon):
+        """ Install monitor on all executors """
+        assert self.binded
+        self._curr_module.install_monitor(mon)
+
+    def forward_backward(self, data_batch):
+        """A convenient function that calls both ``forward`` and ``backward``."""
+        self.forward(data_batch, is_train=True)  # get fc1 and partial fc7
+        self.backward()
+
+    def fit(self,
+            train_data,
+            eval_data=None,
+            eval_metric='acc',
+            epoch_end_callback=None,
+            batch_end_callback=None,
+            kvstore='local',
+            optimizer='sgd',
+            optimizer_params=(('learning_rate', 0.01), ),
+            eval_end_callback=None,
+            eval_batch_end_callback=None,
+            initializer=Uniform(0.01),
+            arg_params=None,
+            aux_params=None,
+            allow_missing=False,
+            force_rebind=False,
+            force_init=False,
+            begin_epoch=0,
+            num_epoch=None,
+            validation_metric=None,
+            monitor=None,
+            sparse_row_id_fn=None):
+        """Trains the module parameters.
+
+        Checkout `Module Tutorial <http://mxnet.io/tutorials/basic/module.html>`_ to see
+        a end-to-end use-case.
+
+        Parameters
+        ----------
+        train_data : DataIter
+            Train DataIter.
+        eval_data : DataIter
+            If not ``None``, will be used as validation set and the performance
+            after each epoch will be evaluated.
+        eval_metric : str or EvalMetric
+            Defaults to 'accuracy'. The performance measure used to display during training.
+            Other possible predefined metrics are:
+            'ce' (CrossEntropy), 'f1', 'mae', 'mse', 'rmse', 'top_k_accuracy'.
+        epoch_end_callback : function or list of functions
+            Each callback will be called with the current `epoch`, `symbol`, `arg_params`
+            and `aux_params`.
+        batch_end_callback : function or list of function
+            Each callback will be called with a `BatchEndParam`.
+        kvstore : str or KVStore
+            Defaults to 'local'.
+        optimizer : str or Optimizer
+            Defaults to 'sgd'.
+        optimizer_params : dict
+            Defaults to ``(('learning_rate', 0.01),)``. The parameters for
+            the optimizer constructor.
+            The default value is not a dict, just to avoid pylint warning on dangerous
+            default values.
+        eval_end_callback : function or list of function
+            These will be called at the end of each full evaluation, with the metrics over
+            the entire evaluation set.
+        eval_batch_end_callback : function or list of function
+            These will be called at the end of each mini-batch during evaluation.
+        initializer : Initializer
+            The initializer is called to initialize the module parameters when they are
+            not already initialized.
+        arg_params : dict
+            Defaults to ``None``, if not ``None``, should be existing parameters from a trained
+            model or loaded from a checkpoint (previously saved model). In this case,
+            the value here will be used to initialize the module parameters, unless they
+            are already initialized by the user via a call to `init_params` or `fit`.
+            `arg_params` has a higher priority than `initializer`.
+        aux_params : dict
+            Defaults to ``None``. Similar to `arg_params`, except for auxiliary states.
+        allow_missing : bool
+            Defaults to ``False``. Indicates whether to allow missing parameters when `arg_params`
+            and `aux_params` are not ``None``. If this is ``True``, then the missing parameters
+            will be initialized via the `initializer`.
+        force_rebind : bool
+            Defaults to ``False``. Whether to force rebinding the executors if already bound.
+        force_init : bool
+            Defaults to ``False``. Indicates whether to force initialization even if the
+            parameters are already initialized.
+        begin_epoch : int
+            Defaults to 0. Indicates the starting epoch. Usually, if resumed from a
+            checkpoint saved at a previous training phase at epoch N, then this value should be
+            N+1.
+        num_epoch : int
+            Number of epochs for training.
+        sparse_row_id_fn : A callback function
+            The function  takes `data_batch` as an input and returns a dict of
+            str -> NDArray. The resulting dict is used for pulling row_sparse
+            parameters from the kvstore, where the str key is the name of the param,
+            and the value is the row id of the param to pull.
+
+        Examples
+        --------
+        >>> # An example of using fit for training.
+        >>> # Assume training dataIter and validation dataIter are ready
+        >>> # Assume loading a previously checkpointed model
+        >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 3)
+        >>> mod.fit(train_data=train_dataiter, eval_data=val_dataiter, optimizer='sgd',
+        ...     optimizer_params={'learning_rate':0.01, 'momentum': 0.9},
+        ...     arg_params=arg_params, aux_params=aux_params,
+        ...     eval_metric='acc', num_epoch=10, begin_epoch=3)
+        """
+        assert num_epoch is not None, 'please specify number of epochs'
+        assert arg_params is None and aux_params is None
+
+        self.bind(data_shapes=train_data.provide_data,
+                  label_shapes=train_data.provide_label,
+                  for_training=True,
+                  force_rebind=force_rebind)
+        if monitor is not None:
+            self.install_monitor(monitor)
+        self.init_params(initializer=initializer,
+                         arg_params=arg_params,
+                         aux_params=aux_params,
+                         allow_missing=allow_missing,
+                         force_init=force_init)
+        self.init_optimizer(kvstore=kvstore,
+                            optimizer=optimizer,
+                            optimizer_params=optimizer_params)
+
+        if validation_metric is None:
+            validation_metric = eval_metric
+        if not isinstance(eval_metric, metric.EvalMetric):
+            eval_metric = metric.create(eval_metric)
+        epoch_eval_metric = copy.deepcopy(eval_metric)
+
+        ################################################################################
+        # training loop
+        ################################################################################
+        for epoch in range(begin_epoch, num_epoch):
+            tic = time.time()
+            eval_metric.reset()
+            epoch_eval_metric.reset()
+            nbatch = 0
+            data_iter = iter(train_data)
+            end_of_batch = False
+            next_data_batch = next(data_iter)
+            while not end_of_batch:
+                data_batch = next_data_batch
+                if monitor is not None:
+                    monitor.tic()
+                self.forward_backward(data_batch)
+                self.update()
+                assert not isinstance(data_batch, list)
+
+                #if isinstance(data_batch, list):
+                #    #print('XXX')
+                #    self.update_metric(eval_metric,
+                #                       [db.label for db in data_batch],
+                #                       pre_sliced=True)
+                #    self.update_metric(epoch_eval_metric,
+                #                       [db.label for db in data_batch],
+                #                       pre_sliced=True)
+                #else:
+                #    #print('before update metric')
+                #    self.update_metric(eval_metric, data_batch.label)
+                #    self.update_metric(epoch_eval_metric, data_batch.label)
+                #labels = data_batch.label
+                #labels = [self.global_label]
+                #self.update_metric(eval_metric, labels)
+                #self.update_metric(epoch_eval_metric, labels)
+
+                try:
+                    # pre fetch next batch
+                    next_data_batch = next(data_iter)
+                    self.prepare(next_data_batch,
+                                 sparse_row_id_fn=sparse_row_id_fn)
+                except StopIteration:
+                    end_of_batch = True
+
+                if monitor is not None:
+                    monitor.toc_print()
+
+                #if end_of_batch:
+                #    eval_name_vals = epoch_eval_metric.get_name_value()
+
+                if batch_end_callback is not None:
+                    batch_end_params = BatchEndParam(epoch=epoch,
+                                                     nbatch=nbatch,
+                                                     eval_metric=None,
+                                                     locals=locals())
+                    batch_end_callback(batch_end_params)
+                    #for callback in _as_list(batch_end_callback):
+                    #    callback(batch_end_params)
+                nbatch += 1
+
+            # one epoch of training is finished
+            #for name, val in eval_name_vals:
+            #    self.logger.info('Epoch[%d] Train-%s=%f', epoch, name, val)
+            toc = time.time()
+            self.logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc - tic))
+
+            # sync aux params across devices
+            arg_params, aux_params = self.get_params()
+            self.set_params(arg_params, aux_params)
+
+            # end of 1 epoch, reset the data-iter for another epoch
+            train_data.reset()
diff --git a/recognition/arcface_mxnet/sample_config.py b/recognition/arcface_mxnet/sample_config.py
new file mode 100644
index 000000000..804fb3204
--- /dev/null
+++ b/recognition/arcface_mxnet/sample_config.py
@@ -0,0 +1,224 @@
+import numpy as np
+import os
+from easydict import EasyDict as edict
+
+config = edict()
+
+config.bn_mom = 0.9
+config.workspace = 256
+config.emb_size = 512
+config.ckpt_embedding = True
+config.net_se = 0
+config.net_act = 'prelu'
+config.net_unit = 3
+config.net_input = 1
+config.net_blocks = [1, 4, 6, 2]
+config.net_output = 'E'
+config.net_multiplier = 1.0
+config.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+config.ce_loss = True
+config.fc7_lr_mult = 1.0
+config.fc7_wd_mult = 1.0
+config.fc7_no_bias = False
+config.max_steps = 0
+config.data_rand_mirror = True
+config.data_cutoff = False
+config.data_color = 0
+config.data_images_filter = 0
+config.count_flops = True
+config.memonger = False  #not work now
+config.is_shuffled_rec = False
+
+config.fp16 = False
+
+# network settings
+network = edict()
+
+network.r100 = edict()
+network.r100.net_name = 'fresnet'
+network.r100.num_layers = 100
+
+network.r100fc = edict()
+network.r100fc.net_name = 'fresnet'
+network.r100fc.num_layers = 100
+network.r100fc.net_output = 'FC'
+
+network.r50 = edict()
+network.r50.net_name = 'fresnet'
+network.r50.num_layers = 50
+
+network.r50v1 = edict()
+network.r50v1.net_name = 'fresnet'
+network.r50v1.num_layers = 50
+network.r50v1.net_unit = 1
+
+network.d169 = edict()
+network.d169.net_name = 'fdensenet'
+network.d169.num_layers = 169
+network.d169.per_batch_size = 64
+network.d169.densenet_dropout = 0.0
+
+network.d201 = edict()
+network.d201.net_name = 'fdensenet'
+network.d201.num_layers = 201
+network.d201.per_batch_size = 64
+network.d201.densenet_dropout = 0.0
+
+network.y1 = edict()
+network.y1.net_name = 'fmobilefacenet'
+network.y1.emb_size = 128
+network.y1.net_output = 'GDC'
+
+network.y2 = edict()
+network.y2.net_name = 'fmobilefacenet'
+network.y2.emb_size = 256
+network.y2.net_output = 'GDC'
+network.y2.net_blocks = [2, 8, 16, 4]
+
+network.m1 = edict()
+network.m1.net_name = 'fmobilenet'
+network.m1.emb_size = 256
+network.m1.net_output = 'GDC'
+network.m1.net_multiplier = 1.0
+
+network.m05 = edict()
+network.m05.net_name = 'fmobilenet'
+network.m05.emb_size = 256
+network.m05.net_output = 'GDC'
+network.m05.net_multiplier = 0.5
+
+network.mnas = edict()
+network.mnas.net_name = 'fmnasnet'
+network.mnas.emb_size = 256
+network.mnas.net_output = 'GDC'
+network.mnas.net_multiplier = 1.0
+
+network.mnas05 = edict()
+network.mnas05.net_name = 'fmnasnet'
+network.mnas05.emb_size = 256
+network.mnas05.net_output = 'GDC'
+network.mnas05.net_multiplier = 0.5
+
+network.mnas025 = edict()
+network.mnas025.net_name = 'fmnasnet'
+network.mnas025.emb_size = 256
+network.mnas025.net_output = 'GDC'
+network.mnas025.net_multiplier = 0.25
+
+network.vargfacenet = edict()
+network.vargfacenet.net_name = 'vargfacenet'
+network.vargfacenet.net_multiplier = 1.25
+network.vargfacenet.emb_size = 512
+network.vargfacenet.net_output = 'J'
+
+# dataset settings
+dataset = edict()
+
+dataset.emore = edict()
+dataset.emore.dataset = 'emore'
+dataset.emore.dataset_path = '../datasets/faces_emore'
+dataset.emore.num_classes = 85742
+dataset.emore.image_shape = (112, 112, 3)
+dataset.emore.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+
+dataset.retina = edict()
+dataset.retina.dataset = 'retina'
+dataset.retina.dataset_path = '../datasets/ms1m-retinaface-t1'
+dataset.retina.num_classes = 93431
+dataset.retina.image_shape = (112, 112, 3)
+dataset.retina.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+
+loss = edict()
+loss.softmax = edict()
+loss.softmax.loss_name = 'softmax'
+
+loss.nsoftmax = edict()
+loss.nsoftmax.loss_name = 'margin_softmax'
+loss.nsoftmax.loss_s = 64.0
+loss.nsoftmax.loss_m1 = 1.0
+loss.nsoftmax.loss_m2 = 0.0
+loss.nsoftmax.loss_m3 = 0.0
+
+loss.arcface = edict()
+loss.arcface.loss_name = 'margin_softmax'
+loss.arcface.loss_s = 64.0
+loss.arcface.loss_m1 = 1.0
+loss.arcface.loss_m2 = 0.5
+loss.arcface.loss_m3 = 0.0
+
+loss.cosface = edict()
+loss.cosface.loss_name = 'margin_softmax'
+loss.cosface.loss_s = 64.0
+loss.cosface.loss_m1 = 1.0
+loss.cosface.loss_m2 = 0.0
+loss.cosface.loss_m3 = 0.35
+
+loss.combined = edict()
+loss.combined.loss_name = 'margin_softmax'
+loss.combined.loss_s = 64.0
+loss.combined.loss_m1 = 1.0
+loss.combined.loss_m2 = 0.3
+loss.combined.loss_m3 = 0.2
+
+loss.triplet = edict()
+loss.triplet.loss_name = 'triplet'
+loss.triplet.images_per_identity = 5
+loss.triplet.triplet_alpha = 0.3
+loss.triplet.triplet_bag_size = 7200
+loss.triplet.triplet_max_ap = 0.0
+loss.triplet.per_batch_size = 60
+loss.triplet.lr = 0.05
+
+loss.atriplet = edict()
+loss.atriplet.loss_name = 'atriplet'
+loss.atriplet.images_per_identity = 5
+loss.atriplet.triplet_alpha = 0.35
+loss.atriplet.triplet_bag_size = 7200
+loss.atriplet.triplet_max_ap = 0.0
+loss.atriplet.per_batch_size = 60
+loss.atriplet.lr = 0.05
+
+# default settings
+default = edict()
+
+# default network
+default.network = 'r100'
+default.pretrained = ''
+default.pretrained_epoch = 1
+# default dataset
+default.dataset = 'emore'
+default.loss = 'arcface'
+default.frequent = 20
+default.verbose = 2000
+default.kvstore = 'device'
+
+default.end_epoch = 10000
+default.lr = 0.1
+default.wd = 0.0005
+default.mom = 0.9
+default.per_batch_size = 128
+default.ckpt = 3
+default.lr_steps = '100000,160000,220000'
+default.models_root = './models'
+
+
+def generate_config(_network, _dataset, _loss):
+    for k, v in loss[_loss].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    for k, v in network[_network].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    for k, v in dataset[_dataset].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    config.loss = _loss
+    config.network = _network
+    config.dataset = _dataset
+    config.num_workers = 1
+    config.fp16 = False
+    if 'DMLC_NUM_WORKER' in os.environ:
+        config.num_workers = int(os.environ['DMLC_NUM_WORKER'])
diff --git a/recognition/arcface_mxnet/symbol/fdensenet.py b/recognition/arcface_mxnet/symbol/fdensenet.py
new file mode 100644
index 000000000..b3d49ee87
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/fdensenet.py
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# coding: utf-8
+# pylint: disable= arguments-differ
+"""DenseNet, implemented in Gluon."""
+
+import sys
+import os
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act():
+    if config.net_act == 'prelu':
+        return nn.PReLU()
+    else:
+        return nn.Activation(config.net_act)
+
+
+# Helpers
+def _make_dense_block(num_layers, bn_size, growth_rate, dropout, stage_index):
+    out = nn.HybridSequential(prefix='stage%d_' % stage_index)
+    with out.name_scope():
+        for _ in range(num_layers):
+            out.add(_make_dense_layer(growth_rate, bn_size, dropout))
+    return out
+
+
+def _make_dense_layer(growth_rate, bn_size, dropout):
+    new_features = nn.HybridSequential(prefix='')
+    new_features.add(nn.BatchNorm())
+    #new_features.add(nn.Activation('relu'))
+    new_features.add(Act())
+    new_features.add(
+        nn.Conv2D(bn_size * growth_rate, kernel_size=1, use_bias=False))
+    new_features.add(nn.BatchNorm())
+    #new_features.add(nn.Activation('relu'))
+    new_features.add(Act())
+    new_features.add(
+        nn.Conv2D(growth_rate, kernel_size=3, padding=1, use_bias=False))
+    if dropout:
+        new_features.add(nn.Dropout(dropout))
+
+    out = gluon.contrib.nn.HybridConcurrent(axis=1, prefix='')
+    out.add(gluon.contrib.nn.Identity())
+    out.add(new_features)
+
+    return out
+
+
+def _make_transition(num_output_features):
+    out = nn.HybridSequential(prefix='')
+    out.add(nn.BatchNorm())
+    #out.add(nn.Activation('relu'))
+    out.add(Act())
+    out.add(nn.Conv2D(num_output_features, kernel_size=1, use_bias=False))
+    out.add(nn.AvgPool2D(pool_size=2, strides=2))
+    return out
+
+
+# Net
+class DenseNet(nn.HybridBlock):
+    r"""Densenet-BC model from the
+    `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_ paper.
+
+    Parameters
+    ----------
+    num_init_features : int
+        Number of filters to learn in the first convolution layer.
+    growth_rate : int
+        Number of filters to add each layer (`k` in the paper).
+    block_config : list of int
+        List of integers for numbers of layers in each pooling block.
+    bn_size : int, default 4
+        Multiplicative factor for number of bottle neck layers.
+        (i.e. bn_size * k features in the bottleneck layer)
+    dropout : float, default 0
+        Rate of dropout after each dense layer.
+    classes : int, default 1000
+        Number of classification classes.
+    """
+    def __init__(self,
+                 num_init_features,
+                 growth_rate,
+                 block_config,
+                 bn_size=4,
+                 dropout=0,
+                 classes=1000,
+                 **kwargs):
+
+        super(DenseNet, self).__init__(**kwargs)
+        with self.name_scope():
+            self.features = nn.HybridSequential(prefix='')
+            self.features.add(
+                nn.Conv2D(num_init_features,
+                          kernel_size=3,
+                          strides=1,
+                          padding=1,
+                          use_bias=False))
+            self.features.add(nn.BatchNorm())
+            self.features.add(nn.Activation('relu'))
+            self.features.add(nn.MaxPool2D(pool_size=3, strides=2, padding=1))
+            # Add dense blocks
+            num_features = num_init_features
+            for i, num_layers in enumerate(block_config):
+                self.features.add(
+                    _make_dense_block(num_layers, bn_size, growth_rate,
+                                      dropout, i + 1))
+                num_features = num_features + num_layers * growth_rate
+                if i != len(block_config) - 1:
+                    self.features.add(_make_transition(num_features // 2))
+                    num_features = num_features // 2
+            self.features.add(nn.BatchNorm())
+            self.features.add(nn.Activation('relu'))
+            #self.features.add(nn.AvgPool2D(pool_size=7))
+            #self.features.add(nn.Flatten())
+
+            #self.output = nn.Dense(classes)
+
+    def hybrid_forward(self, F, x):
+        x = self.features(x)
+        #x = self.output(x)
+        return x
+
+
+# Specification
+densenet_spec = {
+    121: (64, 32, [6, 12, 24, 16]),
+    161: (96, 48, [6, 12, 36, 24]),
+    169: (64, 32, [6, 12, 32, 32]),
+    201: (64, 32, [6, 12, 48, 32])
+}
+
+
+# Constructor
+def get_symbol():
+    num_layers = config.num_layers
+    num_init_features, growth_rate, block_config = densenet_spec[num_layers]
+    net = DenseNet(num_init_features,
+                   growth_rate,
+                   block_config,
+                   dropout=config.densenet_dropout)
+    data = mx.sym.Variable(name='data')
+    data = data - 127.5
+    data = data * 0.0078125
+    body = net(data)
+    fc1 = symbol_utils.get_fc1(body, config.emb_size, config.net_output)
+    return fc1
diff --git a/recognition/arcface_mxnet/symbol/fmnasnet.py b/recognition/arcface_mxnet/symbol/fmnasnet.py
new file mode 100644
index 000000000..118beb942
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/fmnasnet.py
@@ -0,0 +1,213 @@
+import sys
+import os
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act():
+    if config.net_act == 'prelu':
+        return nn.PReLU()
+    else:
+        return nn.Activation(config.net_act)
+
+
+def ConvBlock(channels, kernel_size, strides, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(
+            nn.Conv2D(channels,
+                      kernel_size,
+                      strides=strides,
+                      padding=1,
+                      use_bias=False), nn.BatchNorm(scale=True),
+            Act()
+            #nn.Activation('relu')
+        )
+    return out
+
+
+def Conv1x1(channels, is_linear=False, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(nn.Conv2D(channels, 1, padding=0, use_bias=False),
+                nn.BatchNorm(scale=True))
+        if not is_linear:
+            #out.add(nn.Activation('relu'))
+            out.add(Act())
+    return out
+
+
+def DWise(channels, strides, kernel_size=3, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(
+            nn.Conv2D(channels,
+                      kernel_size,
+                      strides=strides,
+                      padding=kernel_size // 2,
+                      groups=channels,
+                      use_bias=False), nn.BatchNorm(scale=True),
+            Act()
+            #nn.Activation('relu')
+        )
+    return out
+
+
+class SepCONV(nn.HybridBlock):
+    def __init__(self,
+                 inp,
+                 output,
+                 kernel_size,
+                 depth_multiplier=1,
+                 with_bn=True,
+                 **kwargs):
+        super(SepCONV, self).__init__(**kwargs)
+        with self.name_scope():
+            self.net = nn.HybridSequential()
+            cn = int(inp * depth_multiplier)
+
+            if output is None:
+                self.net.add(
+                    nn.Conv2D(in_channels=inp,
+                              channels=cn,
+                              groups=inp,
+                              kernel_size=kernel_size,
+                              strides=(1, 1),
+                              padding=kernel_size // 2,
+                              use_bias=not with_bn))
+            else:
+                self.net.add(
+                    nn.Conv2D(in_channels=inp,
+                              channels=cn,
+                              groups=inp,
+                              kernel_size=kernel_size,
+                              strides=(1, 1),
+                              padding=kernel_size // 2,
+                              use_bias=False),
+                    nn.BatchNorm(),
+                    Act(),
+                    #nn.Activation('relu'),
+                    nn.Conv2D(in_channels=cn,
+                              channels=output,
+                              kernel_size=(1, 1),
+                              strides=(1, 1),
+                              use_bias=not with_bn))
+
+            self.with_bn = with_bn
+            self.act = Act()
+            #self.act = nn.Activation('relu')
+            if with_bn:
+                self.bn = nn.BatchNorm()
+
+    def hybrid_forward(self, F, x):
+        x = self.net(x)
+        if self.with_bn:
+            x = self.bn(x)
+        if self.act is not None:
+            x = self.act(x)
+        return x
+
+
+class ExpandedConv(nn.HybridBlock):
+    def __init__(self,
+                 inp,
+                 oup,
+                 t,
+                 strides,
+                 kernel=3,
+                 same_shape=True,
+                 **kwargs):
+        super(ExpandedConv, self).__init__(**kwargs)
+
+        self.same_shape = same_shape
+        self.strides = strides
+        with self.name_scope():
+            self.bottleneck = nn.HybridSequential()
+            self.bottleneck.add(
+                Conv1x1(inp * t, prefix="expand_"),
+                DWise(inp * t, self.strides, kernel, prefix="dwise_"),
+                Conv1x1(oup, is_linear=True, prefix="linear_"))
+
+    def hybrid_forward(self, F, x):
+        out = self.bottleneck(x)
+        if self.strides == 1 and self.same_shape:
+            out = F.elemwise_add(out, x)
+        return out
+
+
+def ExpandedConvSequence(t, k, inp, oup, repeats, first_strides, **kwargs):
+    seq = nn.HybridSequential(**kwargs)
+    with seq.name_scope():
+        seq.add(ExpandedConv(inp, oup, t, first_strides, k, same_shape=False))
+        curr_inp = oup
+        for i in range(1, repeats):
+            seq.add(ExpandedConv(curr_inp, oup, t, 1))
+            curr_inp = oup
+    return seq
+
+
+class MNasNet(nn.HybridBlock):
+    def __init__(self, m=1.0, **kwargs):
+        super(MNasNet, self).__init__(**kwargs)
+
+        self.first_oup = int(32 * m)
+        self.second_oup = int(16 * m)
+        #self.second_oup = int(32*m)
+        self.interverted_residual_setting = [
+            # t, c,  n, s, k
+            [3, int(24 * m), 3, 2, 3, "stage2_"],  # -> 56x56
+            [3, int(40 * m), 3, 2, 5, "stage3_"],  # -> 28x28
+            [6, int(80 * m), 3, 2, 5, "stage4_1_"],  # -> 14x14
+            [6, int(96 * m), 2, 1, 3, "stage4_2_"],  # -> 14x14
+            [6, int(192 * m), 4, 2, 5, "stage5_1_"],  # -> 7x7
+            [6, int(320 * m), 1, 1, 3, "stage5_2_"],  # -> 7x7          
+        ]
+        self.last_channels = int(1024 * m)
+
+        with self.name_scope():
+            self.features = nn.HybridSequential()
+            self.features.add(
+                ConvBlock(self.first_oup, 3, 1, prefix="stage1_conv0_"))
+            self.features.add(
+                SepCONV(self.first_oup,
+                        self.second_oup,
+                        3,
+                        prefix="stage1_sepconv0_"))
+            inp = self.second_oup
+            for i, (t, c, n, s, k,
+                    prefix) in enumerate(self.interverted_residual_setting):
+                oup = c
+                self.features.add(
+                    ExpandedConvSequence(t, k, inp, oup, n, s, prefix=prefix))
+                inp = oup
+
+            self.features.add(Conv1x1(self.last_channels, prefix="stage5_3_"))
+            #self.features.add(nn.GlobalAvgPool2D())
+            #self.features.add(nn.Flatten())
+            #self.output = nn.Dense(num_classes)
+    def hybrid_forward(self, F, x):
+        x = self.features(x)
+        #x = self.output(x)
+        return x
+
+    def num_output_channel(self):
+        return self.last_channels
+
+
+def get_symbol():
+    net = MNasNet(config.net_multiplier)
+    data = mx.sym.Variable(name='data')
+    data = data - 127.5
+    data = data * 0.0078125
+    body = net(data)
+    fc1 = symbol_utils.get_fc1(body,
+                               config.emb_size,
+                               config.net_output,
+                               input_channel=net.num_output_channel())
+    return fc1
diff --git a/recognition/arcface_mxnet/symbol/fmobilefacenet.py b/recognition/arcface_mxnet/symbol/fmobilefacenet.py
new file mode 100644
index 000000000..f498264ea
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/fmobilefacenet.py
@@ -0,0 +1,224 @@
+import sys
+import os
+import mxnet as mx
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def Conv(data,
+         num_filter=1,
+         kernel=(1, 1),
+         stride=(1, 1),
+         pad=(0, 0),
+         num_group=1,
+         name=None,
+         suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=config.bn_mom)
+    act = Act(data=bn,
+              act_type=config.net_act,
+              name='%s%s_relu' % (name, suffix))
+    return act
+
+
+def Linear(data,
+           num_filter=1,
+           kernel=(1, 1),
+           stride=(1, 1),
+           pad=(0, 0),
+           num_group=1,
+           name=None,
+           suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=config.bn_mom)
+    return bn
+
+
+def ConvOnly(data,
+             num_filter=1,
+             kernel=(1, 1),
+             stride=(1, 1),
+             pad=(0, 0),
+             num_group=1,
+             name=None,
+             suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    return conv
+
+
+def DResidual(data,
+              num_out=1,
+              kernel=(3, 3),
+              stride=(2, 2),
+              pad=(1, 1),
+              num_group=1,
+              name=None,
+              suffix=''):
+    conv = Conv(data=data,
+                num_filter=num_group,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                name='%s%s_conv_sep' % (name, suffix))
+    conv_dw = Conv(data=conv,
+                   num_filter=num_group,
+                   num_group=num_group,
+                   kernel=kernel,
+                   pad=pad,
+                   stride=stride,
+                   name='%s%s_conv_dw' % (name, suffix))
+    proj = Linear(data=conv_dw,
+                  num_filter=num_out,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name='%s%s_conv_proj' % (name, suffix))
+    return proj
+
+
+def Residual(data,
+             num_block=1,
+             num_out=1,
+             kernel=(3, 3),
+             stride=(1, 1),
+             pad=(1, 1),
+             num_group=1,
+             name=None,
+             suffix=''):
+    identity = data
+    for i in range(num_block):
+        shortcut = identity
+        conv = DResidual(data=identity,
+                         num_out=num_out,
+                         kernel=kernel,
+                         stride=stride,
+                         pad=pad,
+                         num_group=num_group,
+                         name='%s%s_block' % (name, suffix),
+                         suffix='%d' % i)
+        identity = conv + shortcut
+    return identity
+
+
+def get_symbol():
+    num_classes = config.emb_size
+    print('in_network', config)
+    fc_type = config.net_output
+    data = mx.symbol.Variable(name="data")
+    data = data - 127.5
+    data = data * 0.0078125
+    blocks = config.net_blocks
+    conv_1 = Conv(data,
+                  num_filter=64,
+                  kernel=(3, 3),
+                  pad=(1, 1),
+                  stride=(2, 2),
+                  name="conv_1")
+    if blocks[0] == 1:
+        conv_2_dw = Conv(conv_1,
+                         num_group=64,
+                         num_filter=64,
+                         kernel=(3, 3),
+                         pad=(1, 1),
+                         stride=(1, 1),
+                         name="conv_2_dw")
+    else:
+        conv_2_dw = Residual(conv_1,
+                             num_block=blocks[0],
+                             num_out=64,
+                             kernel=(3, 3),
+                             stride=(1, 1),
+                             pad=(1, 1),
+                             num_group=64,
+                             name="res_2")
+    conv_23 = DResidual(conv_2_dw,
+                        num_out=64,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=128,
+                        name="dconv_23")
+    conv_3 = Residual(conv_23,
+                      num_block=blocks[1],
+                      num_out=64,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=128,
+                      name="res_3")
+    conv_34 = DResidual(conv_3,
+                        num_out=128,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=256,
+                        name="dconv_34")
+    conv_4 = Residual(conv_34,
+                      num_block=blocks[2],
+                      num_out=128,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=256,
+                      name="res_4")
+    conv_45 = DResidual(conv_4,
+                        num_out=128,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=512,
+                        name="dconv_45")
+    conv_5 = Residual(conv_45,
+                      num_block=blocks[3],
+                      num_out=128,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=256,
+                      name="res_5")
+    conv_6_sep = Conv(conv_5,
+                      num_filter=512,
+                      kernel=(1, 1),
+                      pad=(0, 0),
+                      stride=(1, 1),
+                      name="conv_6sep")
+
+    fc1 = symbol_utils.get_fc1(conv_6_sep, num_classes, fc_type)
+    return fc1
diff --git a/recognition/arcface_mxnet/symbol/fmobilenet.py b/recognition/arcface_mxnet/symbol/fmobilenet.py
new file mode 100644
index 000000000..bdbf8a559
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/fmobilenet.py
@@ -0,0 +1,275 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+import sys
+import os
+import mxnet as mx
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def Conv(data,
+         num_filter=1,
+         kernel=(1, 1),
+         stride=(1, 1),
+         pad=(0, 0),
+         num_group=1,
+         name=None,
+         suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=True)
+    act = Act(data=bn,
+              act_type=config.net_act,
+              name='%s%s_relu' % (name, suffix))
+    return act
+
+
+def ConvOnly(data,
+             num_filter=1,
+             kernel=(1, 1),
+             stride=(1, 1),
+             pad=(0, 0),
+             num_group=1,
+             name=None,
+             suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    return conv
+
+
+def get_symbol():
+    num_classes = config.emb_size
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    data = mx.symbol.Variable(name="data")  # 224
+    data = data - 127.5
+    data = data * 0.0078125
+    fc_type = config.net_output
+    bf = int(32 * config.net_multiplier)
+    if config.net_input == 0:
+        conv_1 = Conv(data,
+                      num_filter=bf,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(2, 2),
+                      name="conv_1")  # 224/112
+    else:
+        conv_1 = Conv(data,
+                      num_filter=bf,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_1")  # 224/112
+    conv_2_dw = Conv(conv_1,
+                     num_group=bf,
+                     num_filter=bf,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_2_dw")  # 112/112
+    conv_2 = Conv(conv_2_dw,
+                  num_filter=bf * 2,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_2")  # 112/112
+    conv_3_dw = Conv(conv_2,
+                     num_group=bf * 2,
+                     num_filter=bf * 2,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_3_dw")  # 112/56
+    conv_3 = Conv(conv_3_dw,
+                  num_filter=bf * 4,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_3")  # 56/56
+    conv_4_dw = Conv(conv_3,
+                     num_group=bf * 4,
+                     num_filter=bf * 4,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_4_dw")  # 56/56
+    conv_4 = Conv(conv_4_dw,
+                  num_filter=bf * 4,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_4")  # 56/56
+    conv_5_dw = Conv(conv_4,
+                     num_group=bf * 4,
+                     num_filter=bf * 4,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_5_dw")  # 56/28
+    conv_5 = Conv(conv_5_dw,
+                  num_filter=bf * 8,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_5")  # 28/28
+    conv_6_dw = Conv(conv_5,
+                     num_group=bf * 8,
+                     num_filter=bf * 8,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_6_dw")  # 28/28
+    conv_6 = Conv(conv_6_dw,
+                  num_filter=bf * 8,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_6")  # 28/28
+    conv_7_dw = Conv(conv_6,
+                     num_group=bf * 8,
+                     num_filter=bf * 8,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_7_dw")  # 28/14
+    conv_7 = Conv(conv_7_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_7")  # 14/14
+
+    conv_8_dw = Conv(conv_7,
+                     num_group=bf * 16,
+                     num_filter=bf * 16,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_8_dw")  # 14/14
+    conv_8 = Conv(conv_8_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_8")  # 14/14
+    conv_9_dw = Conv(conv_8,
+                     num_group=bf * 16,
+                     num_filter=bf * 16,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_9_dw")  # 14/14
+    conv_9 = Conv(conv_9_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_9")  # 14/14
+    conv_10_dw = Conv(conv_9,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_10_dw")  # 14/14
+    conv_10 = Conv(conv_10_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_10")  # 14/14
+    conv_11_dw = Conv(conv_10,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_11_dw")  # 14/14
+    conv_11 = Conv(conv_11_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_11")  # 14/14
+    conv_12_dw = Conv(conv_11,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_12_dw")  # 14/14
+    conv_12 = Conv(conv_12_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_12")  # 14/14
+
+    conv_13_dw = Conv(conv_12,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(2, 2),
+                      name="conv_13_dw")  # 14/7
+    conv_13 = Conv(conv_13_dw,
+                   num_filter=bf * 32,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_13")  # 7/7
+    conv_14_dw = Conv(conv_13,
+                      num_group=bf * 32,
+                      num_filter=bf * 32,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_14_dw")  # 7/7
+    conv_14 = Conv(conv_14_dw,
+                   num_filter=bf * 32,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_14")  # 7/7
+    body = conv_14
+    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
+    return fc1
diff --git a/recognition/arcface_mxnet/symbol/fresnet.py b/recognition/arcface_mxnet/symbol/fresnet.py
new file mode 100644
index 000000000..7b17788d1
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/fresnet.py
@@ -0,0 +1,1191 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+'''
+Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
+Original author Wei Wu
+
+Implemented the following paper:
+
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
+'''
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import sys
+import os
+import mxnet as mx
+import numpy as np
+import symbol_utils
+import memonger
+import sklearn
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Conv(**kwargs):
+    #name = kwargs.get('name')
+    #_weight = mx.symbol.Variable(name+'_weight')
+    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
+    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
+    body = mx.sym.Convolution(**kwargs)
+    return body
+
+
+def Act(data, act_type, name):
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit2')
+    if bottle_neck:
+        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv1 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv2 = Conv(data=act2,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
+        conv3 = Conv(data=act3,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=conv3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            conv3 = mx.symbol.broadcast_mul(conv3, body)
+        if dim_match:
+            shortcut = data
+        else:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return conv3 + shortcut
+    else:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv1 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv2 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=conv2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            conv2 = mx.symbol.broadcast_mul(conv2, body)
+        if dim_match:
+            shortcut = data
+        else:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return conv2 + shortcut
+
+
+def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit3')
+    if bottle_neck:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn4 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn4')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn4,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn4 = mx.symbol.broadcast_mul(bn4, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return bn4 + shortcut
+    else:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return bn3 + shortcut
+
+
+def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
+    """Return ResNeXt Unit symbol for building ResNeXt
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    assert (bottle_neck)
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    num_group = 32
+    #print('in unit3')
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(1, 1),
+                 stride=(1, 1),
+                 pad=(0, 0),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+    conv2 = Conv(data=act1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(3, 3),
+                 stride=(1, 1),
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+    act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
+    conv3 = Conv(data=act2,
+                 num_filter=num_filter,
+                 kernel=(1, 1),
+                 stride=stride,
+                 pad=(0, 0),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv3')
+    bn4 = mx.sym.BatchNorm(data=conv3,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn4')
+
+    if use_se:
+        #se begin
+        body = mx.sym.Pooling(data=bn4,
+                              global_pool=True,
+                              kernel=(7, 7),
+                              pool_type='avg',
+                              name=name + '_se_pool1')
+        body = Conv(data=body,
+                    num_filter=num_filter // 16,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv1",
+                    workspace=workspace)
+        body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+        body = Conv(data=body,
+                    num_filter=num_filter,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv2",
+                    workspace=workspace)
+        body = mx.symbol.Activation(data=body,
+                                    act_type='sigmoid',
+                                    name=name + "_se_sigmoid")
+        bn4 = mx.symbol.broadcast_mul(bn4, body)
+        #se end
+
+    if dim_match:
+        shortcut = data
+    else:
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=bn_mom,
+                                    name=name + '_sc')
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return bn4 + shortcut
+
+
+def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck,
+                  **kwargs):
+    uv = kwargs.get('version_unit', 3)
+    version_input = kwargs.get('version_input', 1)
+    if uv == 1:
+        if version_input == 0:
+            return residual_unit_v1(data, num_filter, stride, dim_match, name,
+                                    bottle_neck, **kwargs)
+        else:
+            return residual_unit_v1_L(data, num_filter, stride, dim_match,
+                                      name, bottle_neck, **kwargs)
+    elif uv == 2:
+        return residual_unit_v2(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+    elif uv == 4:
+        return residual_unit_v4(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+    else:
+        return residual_unit_v3(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+
+
+def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    kwargs = {
+        'version_se': config.net_se,
+        'version_input': config.net_input,
+        'version_output': config.net_output,
+        'version_unit': config.net_unit,
+        'version_act': config.net_act,
+        'bn_mom': bn_mom,
+        'workspace': workspace,
+        'memonger': config.memonger,
+    }
+    """Return ResNet symbol of
+    Parameters
+    ----------
+    units : list
+        Number of units in each stage
+    num_stages : int
+        Number of stage
+    filter_list : list
+        Channel size of each stage
+    num_classes : int
+        Ouput size of symbol
+    dataset : str
+        Dataset type, only cifar10 and imagenet supports
+    workspace : int
+        Workspace used in convolution operator
+    """
+    version_se = kwargs.get('version_se', 1)
+    version_input = kwargs.get('version_input', 1)
+    assert version_input >= 0
+    version_output = kwargs.get('version_output', 'E')
+    fc_type = version_output
+    version_unit = kwargs.get('version_unit', 3)
+    act_type = kwargs.get('version_act', 'prelu')
+    memonger = kwargs.get('memonger', False)
+    print(version_se, version_input, version_output, version_unit, act_type,
+          memonger)
+    num_unit = len(units)
+    assert (num_unit == num_stages)
+    data = mx.sym.Variable(name='data')
+    if config.fp16:
+        data = mx.sym.Cast(data=data, dtype=np.float16)
+    if version_input == 0:
+        #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
+        data = mx.sym.identity(data=data, name='id')
+        data = data - 127.5
+        data = data * 0.0078125
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+        #body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
+    elif version_input == 2:
+        data = mx.sym.BatchNorm(data=data,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn_data')
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+    else:
+        data = mx.sym.identity(data=data, name='id')
+        data = data - 127.5
+        data = data * 0.0078125
+        body = data
+        body = Conv(data=body,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+
+    for i in range(num_stages):
+        #if version_input==0:
+        #  body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
+        #                       name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
+        #else:
+        #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
+        #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
+        if i==num_stages-1 and config.fp16:
+            body = mx.sym.Cast(data=body, dtype=np.float32)
+        body = residual_unit(body,
+                             filter_list[i + 1], (2, 2),
+                             False,
+                             name='stage%d_unit%d' % (i + 1, 1),
+                             bottle_neck=bottle_neck,
+                             **kwargs)
+        for j in range(units[i] - 1):
+            body = residual_unit(body,
+                                 filter_list[i + 1], (1, 1),
+                                 True,
+                                 name='stage%d_unit%d' % (i + 1, j + 2),
+                                 bottle_neck=bottle_neck,
+                                 **kwargs)
+
+    if bottle_neck:
+        body = Conv(data=body,
+                    num_filter=512,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    no_bias=True,
+                    name="convd",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnd')
+        body = Act(data=body, act_type=act_type, name='relud')
+
+    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
+    return fc1
+
+
+def get_symbol():
+    """
+    Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
+    Original author Wei Wu
+    """
+    num_classes = config.emb_size
+    num_layers = config.num_layers
+    if num_layers >= 500:
+        filter_list = [64, 256, 512, 1024, 2048]
+        bottle_neck = True
+    else:
+        filter_list = [64, 64, 128, 256, 512]
+        bottle_neck = False
+    num_stages = 4
+    if num_layers == 18:
+        units = [2, 2, 2, 2]
+    elif num_layers == 34:
+        units = [3, 4, 6, 3]
+    elif num_layers == 49:
+        units = [3, 4, 14, 3]
+    elif num_layers == 50:
+        units = [3, 4, 14, 3]
+    elif num_layers == 74:
+        units = [3, 6, 24, 3]
+    elif num_layers == 90:
+        units = [3, 8, 30, 3]
+    elif num_layers == 98:
+        units = [3, 4, 38, 3]
+    elif num_layers == 99:
+        units = [3, 8, 35, 3]
+    elif num_layers == 100:
+        units = [3, 13, 30, 3]
+    elif num_layers == 134:
+        units = [3, 10, 50, 3]
+    elif num_layers == 136:
+        units = [3, 13, 48, 3]
+    elif num_layers == 140:
+        units = [3, 15, 48, 3]
+    elif num_layers == 124:
+        units = [3, 13, 40, 5]
+    elif num_layers == 160:
+        units = [3, 24, 49, 3]
+    elif num_layers == 101:
+        units = [3, 4, 23, 3]
+    elif num_layers == 152:
+        units = [3, 8, 36, 3]
+    elif num_layers == 200:
+        units = [3, 24, 36, 3]
+    elif num_layers == 269:
+        units = [3, 30, 48, 8]
+    else:
+        raise ValueError(
+            "no experiments done on num_layers {}, you can do it yourself".
+            format(num_layers))
+
+    net = resnet(units=units,
+                 num_stages=num_stages,
+                 filter_list=filter_list,
+                 num_classes=num_classes,
+                 bottle_neck=bottle_neck)
+
+    if config.memonger:
+        dshape = (config.per_batch_size, config.image_shape[2],
+                  config.image_shape[0], config.image_shape[1])
+        net_mem_planned = memonger.search_plan(net, data=dshape)
+        old_cost = memonger.get_cost(net, data=dshape)
+        new_cost = memonger.get_cost(net_mem_planned, data=dshape)
+
+        print('Old feature map cost=%d MB' % old_cost)
+        print('New feature map cost=%d MB' % new_cost)
+        net = net_mem_planned
+    return net
diff --git a/recognition/arcface_mxnet/symbol/memonger.py b/recognition/arcface_mxnet/symbol/memonger.py
new file mode 100644
index 000000000..8ad610b57
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/memonger.py
@@ -0,0 +1,175 @@
+import mxnet as mx
+import math
+
+
+def prod(shape):
+    """Get product of the shape.
+    """
+    ret = 1
+    for s in shape:
+        ret *= s
+    return ret
+
+
+def is_param(name):
+    """Quick script to check if name is a parameter.
+    """
+    if name == 'data':
+        return False
+    if name.endswith('weight'):
+        return True
+    if name.endswith('bias'):
+        return True
+    if name.endswith('beta'):
+        return True
+    if name.endswith('gamma'):
+        return True
+    return False
+
+
+def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
+    """Memory allocation planner with a given threshold.
+
+    The user can pass in a network configuration,
+    a threshold that limits memory per block.
+    And input shape configurations.
+
+    Parameters
+    ----------
+    sym : symbol
+        Input configuration of symbols.
+        The user need to pre-mark the attribute "mirror_stage" on the nodes
+        that can be book-kept as stage
+
+        The algorithm will decide whether to disbale mirror on the stage nodes.
+
+    threshold: integer
+        A tuning parameter to tune the approximate size of each stage blocks
+
+    plan_info: dict, optional
+        Used to hold plan information.
+
+    **kwargs:
+        The arguments to infer shape.
+
+    Returns
+    -------
+    alloc_sym: symbol
+        A symbol with force mirror tagged on the nodes for better allocation.
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_sb = None
+    last_local = 0
+    period = 1
+    last_stage = ''
+    stage_decision = ''
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        else:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def get_cost(sym, type_dict=None, **kwargs):
+    """Get the cost of the current symbolic plan by running bind on CPU.
+
+    sym : Symbolic Variable
+
+    """
+    texec = sym.simple_bind(ctx=mx.gpu(),
+                            grad_req='write',
+                            type_dict=type_dict,
+                            **kwargs)
+    return int(texec.debug_str().split('\n')[-3].split()[1])
+
+
+def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    threshold = 0
+    min_threshold = None
+    min_cost = None
+    nbegin = 3
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan(sym,
+                               threshold=threshold,
+                               plan_info=info,
+                               **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = int(math.sqrt(save_size * local_size / 2))
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan(sym,
+                                   threshold=threshold,
+                                   plan_info=info,
+                                   **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
diff --git a/recognition/arcface_mxnet/symbol/memonger_v2.py b/recognition/arcface_mxnet/symbol/memonger_v2.py
new file mode 100644
index 000000000..92963de5e
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/memonger_v2.py
@@ -0,0 +1,300 @@
+import mxnet as mx
+import math
+
+
+def prod(shape):
+    """Get product of the shape.
+    """
+    ret = 1
+    for s in shape:
+        ret *= s
+    return ret
+
+
+def is_param(name):
+    """Quick script to check if name is a parameter.
+    """
+    if name == 'data':
+        return False
+    if name.endswith('weight'):
+        return True
+    if name.endswith('bias'):
+        return True
+    if name.endswith('beta'):
+        return True
+    if name.endswith('gamma'):
+        return True
+    return False
+
+
+def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
+    """Memory allocation planner with a given threshold.
+
+    The user can pass in a network configuration,
+    a threshold that limits memory per block.
+    And input shape configurations.
+
+    Parameters
+    ----------
+    sym : symbol
+        Input configuration of symbols.
+        The user need to pre-mark the attribute "mirror_stage" on the nodes
+        that can be book-kept as stage
+
+        The algorithm will decide whether to disbale mirror on the stage nodes.
+
+    threshold: integer
+        A tuning parameter to tune the approximate size of each stage blocks
+
+    plan_info: dict, optional
+        Used to hold plan information.
+
+    **kwargs:
+        The arguments to infer shape.
+
+    Returns
+    -------
+    alloc_sym: symbol
+        A symbol with force mirror tagged on the nodes for better allocation.
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_sb = None
+    last_local = 0
+    period = 1
+    last_stage = ''
+    stage_decision = ''
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        else:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def get_cost(sym, type_dict=None, **kwargs):
+    """Get the cost of the current symbolic plan by running bind on CPU.
+
+    sym : Symbolic Variable
+
+    """
+    texec = sym.simple_bind(ctx=mx.gpu(),
+                            grad_req='write',
+                            type_dict=type_dict,
+                            **kwargs)
+    return int(texec.debug_str().split('\n')[-3].split()[1])
+
+
+def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    threshold = 0
+    min_threshold = None
+    min_cost = None
+    nbegin = 3
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan(sym,
+                               threshold=threshold,
+                               plan_info=info,
+                               **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = int(math.sqrt(save_size * local_size / 2))
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan(sym,
+                                   threshold=threshold,
+                                   plan_info=info,
+                                   **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
+
+
+def make_mirror_plan_to_layer(sym,
+                              layer_name,
+                              threshold,
+                              plan_info=None,
+                              **kwargs):
+    """
+    sym is the original symbal
+    layer_name is a name to which layer of the network should be set as mirror
+    threshhold is the approximate size of each mirror block
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_stage = ''
+    stage_decision = ''
+    switch = True
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        #print(name, switch)
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        elif switch and not 'bn' in name:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+            print('set force_mirroring', name, total_size, local_size)
+        if layer_name != '' and layer_name in name:
+            switch = False
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            #print(name, stage)
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def search_plan_to_layer(sym,
+                         layer_name=None,
+                         threshold=500,
+                         ntrial=6,
+                         type_dict=None,
+                         **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    min_threshold = None
+    min_cost = None
+    nbegin = 10
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan_to_layer(sym,
+                                        layer_name=layer_name,
+                                        threshold=threshold,
+                                        plan_info=info,
+                                        **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = 300 * (k + 1)
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    print(min_threshold, max_threshold, step)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan_to_layer(sym,
+                                            layer_name=layer_name,
+                                            threshold=threshold,
+                                            plan_info=info,
+                                            **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
diff --git a/recognition/arcface_mxnet/symbol/symbol_utils.py b/recognition/arcface_mxnet/symbol/symbol_utils.py
new file mode 100644
index 000000000..3eb6f1cf9
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/symbol_utils.py
@@ -0,0 +1,595 @@
+import sys
+import os
+import mxnet as mx
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Conv(**kwargs):
+    #name = kwargs.get('name')
+    #_weight = mx.symbol.Variable(name+'_weight')
+    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
+    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
+    body = mx.sym.Convolution(**kwargs)
+    return body
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+bn_mom = config.bn_mom
+
+
+def Linear(data,
+           num_filter=1,
+           kernel=(1, 1),
+           stride=(1, 1),
+           pad=(0, 0),
+           num_group=1,
+           name=None,
+           suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=bn_mom)
+    return bn
+
+
+def get_fc1(last_conv, num_classes, fc_type, input_channel=512):
+    body = last_conv
+    if fc_type == 'Z':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = body
+    elif fc_type == 'E':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'FC':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'SFC':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(3, 3),
+                    stride=(2, 2),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="convf",
+                    num_group=input_channel)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf')
+        body = Act(data=body, act_type=config.net_act, name='reluf')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(1, 1),
+                    pad=(0, 0),
+                    stride=(1, 1),
+                    name="convf2")
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf2')
+        body = Act(data=body, act_type=config.net_act, name='reluf2')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'GAP':
+        bn1 = mx.sym.BatchNorm(data=body,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='bn1')
+        relu1 = Act(data=bn1, act_type=config.net_act, name='relu1')
+        # Although kernel is not used here when global_pool=True, we should put one
+        pool1 = mx.sym.Pooling(data=relu1,
+                               global_pool=True,
+                               kernel=(7, 7),
+                               pool_type='avg',
+                               name='pool1')
+        flat = mx.sym.Flatten(data=pool1)
+        fc1 = mx.sym.FullyConnected(data=flat,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'GNAP':  #mobilefacenet++
+        filters_in = 512  # param in mobilefacenet
+        if num_classes > filters_in:
+            body = mx.sym.Convolution(data=last_conv,
+                                      num_filter=num_classes,
+                                      kernel=(1, 1),
+                                      stride=(1, 1),
+                                      pad=(0, 0),
+                                      no_bias=True,
+                                      name='convx')
+            body = mx.sym.BatchNorm(data=body,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=0.9,
+                                    name='convx_bn')
+            body = Act(data=body, act_type=config.net_act, name='convx_relu')
+            filters_in = num_classes
+        else:
+            body = last_conv
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name='bn6f')
+
+        spatial_norm = body * body
+        spatial_norm = mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
+        spatial_sqrt = mx.sym.sqrt(spatial_norm)
+        #spatial_mean=mx.sym.mean(spatial_sqrt, axis=(1,2,3), keepdims=True)
+        spatial_mean = mx.sym.mean(spatial_sqrt)
+        spatial_div_inverse = mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
+
+        spatial_attention_inverse = mx.symbol.tile(spatial_div_inverse,
+                                                   reps=(1, filters_in, 1, 1))
+        body = body * spatial_attention_inverse
+        #body = mx.sym.broadcast_mul(body, spatial_div_inverse)
+
+        fc1 = mx.sym.Pooling(body,
+                             kernel=(7, 7),
+                             global_pool=True,
+                             pool_type='avg')
+        if num_classes < filters_in:
+            fc1 = mx.sym.BatchNorm(data=fc1,
+                                   fix_gamma=True,
+                                   eps=2e-5,
+                                   momentum=0.9,
+                                   name='bn6w')
+            fc1 = mx.sym.FullyConnected(data=fc1,
+                                        num_hidden=num_classes,
+                                        name='pre_fc1')
+        else:
+            fc1 = mx.sym.Flatten(data=fc1)
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=0.9,
+                               name='fc1')
+    elif fc_type == "GDC":  #mobilefacenet_v1
+        conv_6_dw = Linear(last_conv,
+                           num_filter=input_channel,
+                           num_group=input_channel,
+                           kernel=(7, 7),
+                           pad=(0, 0),
+                           stride=(1, 1),
+                           name="conv_6dw7_7")
+        conv_6_f = mx.sym.FullyConnected(data=conv_6_dw,
+                                         num_hidden=num_classes,
+                                         name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=conv_6_f,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'F':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'G':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'H':
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'I':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'J':
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    return fc1
+
+
+def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    #print('in unit3')
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=(1, 1),
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    act1 = Act(data=bn2, act_type=config.net_act, name=name + '_relu1')
+    conv2 = Conv(data=act1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=stride,
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+
+    if dim_match:
+        shortcut = data
+    else:
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    momentum=bn_mom,
+                                    eps=2e-5,
+                                    name=name + '_sc')
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return bn3 + shortcut
+
+
+def residual_unit_v1l(data, num_filter, stride, dim_match, name, bottle_neck):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    workspace = config.workspace
+    bn_mom = config.bn_mom
+    memonger = False
+    use_se = config.net_se
+    act_type = config.net_act
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def get_head(data, version_input, num_filter):
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    kwargs = {'bn_mom': bn_mom, 'workspace': workspace}
+    data = data - 127.5
+    data = data * 0.0078125
+    #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
+    if version_input == 0:
+        body = Conv(data=data,
+                    num_filter=num_filter,
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=config.net_act, name='relu0')
+        body = mx.sym.Pooling(data=body,
+                              kernel=(3, 3),
+                              stride=(2, 2),
+                              pad=(1, 1),
+                              pool_type='max')
+    else:
+        body = data
+        _num_filter = min(num_filter, 64)
+        body = Conv(data=body,
+                    num_filter=_num_filter,
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=config.net_act, name='relu0')
+        #body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
+        body = residual_unit_v1l(body,
+                                 _num_filter, (2, 2),
+                                 False,
+                                 name='head',
+                                 bottle_neck=False)
+    return body
diff --git a/recognition/arcface_mxnet/symbol/vargfacenet.py b/recognition/arcface_mxnet/symbol/vargfacenet.py
new file mode 100644
index 000000000..434e8beb8
--- /dev/null
+++ b/recognition/arcface_mxnet/symbol/vargfacenet.py
@@ -0,0 +1,578 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+'''
+Author: Horizon Robotics Inc.
+The company is committed to be the global leader of edge AI platform.
+The model implemented in this scripts runs ~200fps on the Sunrise 2.
+Sunrise 2 is the second generation of an embedded AI chip designed by Horizon Robotics,
+targeting to empower AIoT devices by AI.
+
+Implemented the following paper:
+Mengjia Yan, Mengao Zhao, Zining Xu, Qian Zhang, Guoli Wang, Zhizhong Su. "VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition" (https://arxiv.org/abs/1910.04985)
+
+'''
+
+import os
+import sys
+
+import mxnet as mx
+import symbol_utils
+
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def get_setting_params(**kwargs):
+    # bn_params
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    bn_eps = kwargs.get('bn_eps', 2e-5)
+    fix_gamma = kwargs.get('fix_gamma', False)
+    use_global_stats = kwargs.get('use_global_stats', False)
+    # net_setting param
+    workspace = kwargs.get('workspace', 512)
+    act_type = kwargs.get('act_type', 'prelu')
+    use_se = kwargs.get('use_se', True)
+    se_ratio = kwargs.get('se_ratio', 4)
+    group_base = kwargs.get('group_base', 8)
+
+    setting_params = {}
+    setting_params['bn_mom'] = bn_mom
+    setting_params['bn_eps'] = bn_eps
+    setting_params['fix_gamma'] = fix_gamma
+    setting_params['use_global_stats'] = use_global_stats
+    setting_params['workspace'] = workspace
+    setting_params['act_type'] = act_type
+    setting_params['use_se'] = use_se
+    setting_params['se_ratio'] = se_ratio
+    setting_params['group_base'] = group_base
+
+    return setting_params
+
+
+def se_block(data, num_filter, setting_params, name):
+    se_ratio = setting_params['se_ratio']
+    act_type = setting_params['act_type']
+
+    pool1 = mx.sym.Pooling(data=data,
+                           global_pool=True,
+                           pool_type='avg',
+                           name=name + '_se_pool1')
+    conv1 = mx.sym.Convolution(data=pool1,
+                               num_filter=num_filter // se_ratio,
+                               kernel=(1, 1),
+                               stride=(1, 1),
+                               pad=(0, 0),
+                               name=name + "_se_conv1")
+    act1 = Act(data=conv1, act_type=act_type, name=name + '_se_act1')
+
+    conv2 = mx.sym.Convolution(data=act1,
+                               num_filter=num_filter,
+                               kernel=(1, 1),
+                               stride=(1, 1),
+                               pad=(0, 0),
+                               name=name + "_se_conv2")
+    act2 = mx.symbol.Activation(data=conv2,
+                                act_type='sigmoid',
+                                name=name + "_se_sigmoid")
+    out_data = mx.symbol.broadcast_mul(data, act2)
+    return out_data
+
+
+def separable_conv2d(data,
+                     in_channels,
+                     out_channels,
+                     kernel,
+                     pad,
+                     setting_params,
+                     stride=(1, 1),
+                     factor=1,
+                     bias=False,
+                     bn_dw_out=True,
+                     act_dw_out=True,
+                     bn_pw_out=True,
+                     act_pw_out=True,
+                     dilate=1,
+                     name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    group_base = setting_params['group_base']
+    act_type = setting_params['act_type']
+    assert in_channels % group_base == 0
+
+    # depthwise
+    dw_out = mx.sym.Convolution(data=data,
+                                num_filter=int(in_channels * factor),
+                                kernel=kernel,
+                                pad=pad,
+                                stride=stride,
+                                no_bias=False if bias else True,
+                                num_group=int(in_channels / group_base),
+                                dilate=(dilate, dilate),
+                                workspace=workspace,
+                                name=name + '_conv2d_depthwise')
+    if bn_dw_out:
+        dw_out = mx.sym.BatchNorm(data=dw_out,
+                                  fix_gamma=fix_gamma,
+                                  eps=bn_eps,
+                                  momentum=bn_mom,
+                                  use_global_stats=use_global_stats,
+                                  name=name + '_conv2d_depthwise_bn')
+    if act_dw_out:
+        dw_out = Act(data=dw_out,
+                     act_type=act_type,
+                     name=name + '_conv2d_depthwise_act')
+    # pointwise
+    pw_out = mx.sym.Convolution(data=dw_out,
+                                num_filter=out_channels,
+                                kernel=(1, 1),
+                                stride=(1, 1),
+                                pad=(0, 0),
+                                num_group=1,
+                                no_bias=False if bias else True,
+                                workspace=workspace,
+                                name=name + '_conv2d_pointwise')
+    if bn_pw_out:
+        pw_out = mx.sym.BatchNorm(data=pw_out,
+                                  fix_gamma=fix_gamma,
+                                  eps=bn_eps,
+                                  momentum=bn_mom,
+                                  use_global_stats=use_global_stats,
+                                  name=name + '_conv2d_pointwise_bn')
+    if act_pw_out:
+        pw_out = Act(data=pw_out,
+                     act_type=act_type,
+                     name=name + '_conv2d_pointwise_act')
+    return pw_out
+
+
+def vargnet_block(data,
+                  n_out_ch1,
+                  n_out_ch2,
+                  n_out_ch3,
+                  setting_params,
+                  factor=2,
+                  dim_match=True,
+                  multiplier=1,
+                  kernel=(3, 3),
+                  stride=(1, 1),
+                  dilate=1,
+                  with_dilate=False,
+                  name=None):
+    use_se = setting_params['use_se']
+    act_type = setting_params['act_type']
+
+    out_channels_1 = int(n_out_ch1 * multiplier)
+    out_channels_2 = int(n_out_ch2 * multiplier)
+    out_channels_3 = int(n_out_ch3 * multiplier)
+
+    pad = (((kernel[0] - 1) * dilate + 1) // 2,
+           ((kernel[1] - 1) * dilate + 1) // 2)
+
+    if with_dilate:
+        stride = (1, 1)
+    if dim_match:
+        short_cut = data
+    else:
+        short_cut = separable_conv2d(data=data,
+                                     in_channels=out_channels_1,
+                                     out_channels=out_channels_3,
+                                     kernel=kernel,
+                                     pad=pad,
+                                     setting_params=setting_params,
+                                     stride=stride,
+                                     factor=factor,
+                                     bias=False,
+                                     act_pw_out=False,
+                                     dilate=dilate,
+                                     name=name + '_shortcut')
+    sep1_data = separable_conv2d(data=data,
+                                 in_channels=out_channels_1,
+                                 out_channels=out_channels_2,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=stride,
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 name=name + '_sep1_data')
+    sep2_data = separable_conv2d(data=sep1_data,
+                                 in_channels=out_channels_2,
+                                 out_channels=out_channels_3,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=(1, 1),
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 act_pw_out=False,
+                                 name=name + '_sep2_data')
+
+    if use_se:
+        sep2_data = se_block(data=sep2_data,
+                             num_filter=out_channels_3,
+                             setting_params=setting_params,
+                             name=name)
+
+    out_data = sep2_data + short_cut
+    out_data = Act(data=out_data,
+                   act_type=act_type,
+                   name=name + '_out_data_act')
+    return out_data
+
+
+def vargnet_branch_merge_block(data,
+                               n_out_ch1,
+                               n_out_ch2,
+                               n_out_ch3,
+                               setting_params,
+                               factor=2,
+                               dim_match=False,
+                               multiplier=1,
+                               kernel=(3, 3),
+                               stride=(2, 2),
+                               dilate=1,
+                               with_dilate=False,
+                               name=None):
+    act_type = setting_params['act_type']
+
+    out_channels_1 = int(n_out_ch1 * multiplier)
+    out_channels_2 = int(n_out_ch2 * multiplier)
+    out_channels_3 = int(n_out_ch3 * multiplier)
+
+    pad = (((kernel[0] - 1) * dilate + 1) // 2,
+           ((kernel[1] - 1) * dilate + 1) // 2)
+
+    if with_dilate:
+        stride = (1, 1)
+    if dim_match:
+        short_cut = data
+    else:
+        short_cut = separable_conv2d(data=data,
+                                     in_channels=out_channels_1,
+                                     out_channels=out_channels_3,
+                                     kernel=kernel,
+                                     pad=pad,
+                                     setting_params=setting_params,
+                                     stride=stride,
+                                     factor=factor,
+                                     bias=False,
+                                     act_pw_out=False,
+                                     dilate=dilate,
+                                     name=name + '_shortcut')
+    sep1_data_brach1 = separable_conv2d(data=data,
+                                        in_channels=out_channels_1,
+                                        out_channels=out_channels_2,
+                                        kernel=kernel,
+                                        pad=pad,
+                                        setting_params=setting_params,
+                                        stride=stride,
+                                        factor=factor,
+                                        bias=False,
+                                        dilate=dilate,
+                                        act_pw_out=False,
+                                        name=name + '_sep1_data_branch')
+    sep1_data_brach2 = separable_conv2d(data=data,
+                                        in_channels=out_channels_1,
+                                        out_channels=out_channels_2,
+                                        kernel=kernel,
+                                        pad=pad,
+                                        setting_params=setting_params,
+                                        stride=stride,
+                                        factor=factor,
+                                        bias=False,
+                                        dilate=dilate,
+                                        act_pw_out=False,
+                                        name=name + '_sep2_data_branch')
+    sep1_data = sep1_data_brach1 + sep1_data_brach2
+    sep1_data = Act(data=sep1_data,
+                    act_type=act_type,
+                    name=name + '_sep1_data_act')
+    sep2_data = separable_conv2d(data=sep1_data,
+                                 in_channels=out_channels_2,
+                                 out_channels=out_channels_3,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=(1, 1),
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 act_pw_out=False,
+                                 name=name + '_sep2_data')
+    out_data = sep2_data + short_cut
+    out_data = Act(data=out_data,
+                   act_type=act_type,
+                   name=name + '_out_data_act')
+    return out_data
+
+
+def add_vargnet_conv_block(data,
+                           stage,
+                           units,
+                           in_channels,
+                           out_channels,
+                           setting_params,
+                           kernel=(3, 3),
+                           stride=(2, 2),
+                           multiplier=1,
+                           factor=2,
+                           dilate=1,
+                           with_dilate=False,
+                           name=None):
+    assert stage >= 2, 'stage is {}, stage must be set >=2'.format(stage)
+    data = vargnet_branch_merge_block(data=data,
+                                      n_out_ch1=in_channels,
+                                      n_out_ch2=out_channels,
+                                      n_out_ch3=out_channels,
+                                      setting_params=setting_params,
+                                      factor=factor,
+                                      dim_match=False,
+                                      multiplier=multiplier,
+                                      kernel=kernel,
+                                      stride=stride,
+                                      dilate=dilate,
+                                      with_dilate=with_dilate,
+                                      name=name +
+                                      '_stage_{}_unit_1'.format(stage))
+    for i in range(units - 1):
+        data = vargnet_block(data=data,
+                             n_out_ch1=out_channels,
+                             n_out_ch2=out_channels,
+                             n_out_ch3=out_channels,
+                             setting_params=setting_params,
+                             factor=factor,
+                             dim_match=True,
+                             multiplier=multiplier,
+                             kernel=kernel,
+                             stride=(1, 1),
+                             dilate=dilate,
+                             with_dilate=with_dilate,
+                             name=name +
+                             '_stage_{}_unit_{}'.format(stage, i + 2))
+    return data
+
+
+def add_head_block(data,
+                   num_filter,
+                   setting_params,
+                   multiplier,
+                   head_pooling=False,
+                   kernel=(3, 3),
+                   stride=(2, 2),
+                   pad=(1, 1),
+                   name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    act_type = setting_params['act_type']
+    channels = int(num_filter * multiplier)
+
+    conv1 = mx.sym.Convolution(data=data,
+                               num_filter=channels,
+                               kernel=kernel,
+                               pad=pad,
+                               stride=stride,
+                               no_bias=True,
+                               num_group=1,
+                               workspace=workspace,
+                               name=name + '_conv1')
+    bn1 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=fix_gamma,
+                           eps=bn_eps,
+                           momentum=bn_mom,
+                           use_global_stats=use_global_stats,
+                           name=name + '_conv1_bn')
+
+    act1 = Act(data=bn1, act_type=act_type, name=name + '_conv1_act')
+
+    if head_pooling:
+        head_data = mx.symbol.Pooling(data=act1,
+                                      kernel=(3, 3),
+                                      stride=(2, 2),
+                                      pad=(1, 1),
+                                      pool_type='max',
+                                      name=name + '_max_pooling')
+    else:
+        head_data = vargnet_block(data=act1,
+                                  n_out_ch1=num_filter,
+                                  n_out_ch2=num_filter,
+                                  n_out_ch3=num_filter,
+                                  setting_params=setting_params,
+                                  factor=1,
+                                  dim_match=False,
+                                  multiplier=multiplier,
+                                  kernel=kernel,
+                                  stride=(2, 2),
+                                  dilate=1,
+                                  with_dilate=False,
+                                  name=name + '_head_pooling')
+    return head_data
+
+
+def add_emb_block(data,
+                  input_channels,
+                  last_channels,
+                  emb_size,
+                  fc_type,
+                  setting_params,
+                  bias=False,
+                  name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    act_type = setting_params['act_type']
+    group_base = setting_params['group_base']
+    # last channels
+    if input_channels != last_channels:
+        data = mx.sym.Convolution(data=data,
+                                  num_filter=last_channels,
+                                  kernel=(1, 1),
+                                  pad=(0, 0),
+                                  stride=(1, 1),
+                                  no_bias=False if bias else True,
+                                  workspace=workspace,
+                                  name=name + '_convx')
+        data = mx.sym.BatchNorm(data=data,
+                                fix_gamma=fix_gamma,
+                                eps=bn_eps,
+                                momentum=bn_mom,
+                                use_global_stats=use_global_stats,
+                                name=name + '_convx_bn')
+        data = Act(data=data, act_type=act_type, name=name + '_convx_act')
+    # depthwise
+    convx_depthwise = mx.sym.Convolution(data=data,
+                                         num_filter=last_channels,
+                                         num_group=int(last_channels /
+                                                       group_base),
+                                         kernel=(7, 7),
+                                         pad=(0, 0),
+                                         stride=(1, 1),
+                                         no_bias=False if bias else True,
+                                         workspace=workspace,
+                                         name=name + '_convx_depthwise')
+    convx_depthwise = mx.sym.BatchNorm(data=convx_depthwise,
+                                       fix_gamma=fix_gamma,
+                                       eps=bn_eps,
+                                       momentum=bn_mom,
+                                       use_global_stats=use_global_stats,
+                                       name=name + '_convx_depthwise_bn')
+    # pointwise
+    convx_pointwise = mx.sym.Convolution(data=convx_depthwise,
+                                         num_filter=last_channels // 2,
+                                         kernel=(1, 1),
+                                         pad=(0, 0),
+                                         stride=(1, 1),
+                                         no_bias=False if bias else True,
+                                         workspace=workspace,
+                                         name=name + '_convx_pointwise')
+    convx_pointwise = mx.sym.BatchNorm(data=convx_pointwise,
+                                       fix_gamma=fix_gamma,
+                                       eps=bn_eps,
+                                       momentum=bn_mom,
+                                       use_global_stats=use_global_stats,
+                                       name=name + '_convx_pointwise_bn')
+    convx_pointwise = Act(data=convx_pointwise,
+                          act_type=act_type,
+                          name=name + '_convx_pointwise_act')
+
+    fc1 = symbol_utils.get_fc1(convx_pointwise, emb_size, fc_type)
+    return fc1
+
+
+def get_symbol():
+    multiplier = config.net_multiplier
+    emb_size = config.emb_size
+    fc_type = config.net_output
+
+    kwargs = {
+        'use_se': config.net_se,
+        'act_type': config.net_act,
+        'bn_mom': config.bn_mom,
+        'workspace': config.workspace,
+    }
+
+    setting_params = get_setting_params(**kwargs)
+
+    factor = 2
+    head_pooling = False
+    num_stage = 3
+    stage_list = [2, 3, 4]
+    units = [3, 7, 4]
+    filter_list = [32, 64, 128, 256]
+    last_channels = 1024
+    dilate_list = [1, 1, 1]
+    with_dilate_list = [False, False, False]
+
+    data = mx.sym.Variable(name='data')
+    data = mx.sym.identity(data=data, name='id')
+    data = data - 127.5
+    data = data * 0.0078125
+
+    body = add_head_block(data=data,
+                          num_filter=filter_list[0],
+                          setting_params=setting_params,
+                          multiplier=multiplier,
+                          head_pooling=head_pooling,
+                          kernel=(3, 3),
+                          stride=(1, 1),
+                          pad=(1, 1),
+                          name="vargface_head")
+
+    for i in range(num_stage):
+        body = add_vargnet_conv_block(data=body,
+                                      stage=stage_list[i],
+                                      units=units[i],
+                                      in_channels=filter_list[i],
+                                      out_channels=filter_list[i + 1],
+                                      setting_params=setting_params,
+                                      kernel=(3, 3),
+                                      stride=(2, 2),
+                                      multiplier=multiplier,
+                                      factor=factor,
+                                      dilate=dilate_list[i],
+                                      with_dilate=with_dilate_list[i],
+                                      name="vargface")
+    emb_feat = add_emb_block(data=body,
+                             input_channels=filter_list[3],
+                             last_channels=last_channels,
+                             emb_size=emb_size,
+                             fc_type=fc_type,
+                             setting_params=setting_params,
+                             bias=False,
+                             name='embed')
+    return emb_feat
+
+
+if __name__ == '__main__':
+    get_symbol()
diff --git a/recognition/arcface_mxnet/train.py b/recognition/arcface_mxnet/train.py
new file mode 100644
index 000000000..2f2ebadb2
--- /dev/null
+++ b/recognition/arcface_mxnet/train.py
@@ -0,0 +1,484 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import sys
+import math
+import random
+import logging
+import sklearn
+import pickle
+import numpy as np
+import mxnet as mx
+from mxnet import ndarray as nd
+import argparse
+import mxnet.optimizer as optimizer
+from config import config, default, generate_config
+from metric import *
+sys.path.append(os.path.join(os.path.dirname(__file__),  'common'))
+import flops_counter
+import verification
+sys.path.append(os.path.join(os.path.dirname(__file__),  'symbol'))
+import fresnet
+import fmobilefacenet
+import fmobilenet
+import fmnasnet
+import fdensenet
+import vargfacenet
+
+logger = logging.getLogger()
+logger.setLevel(logging.INFO)
+
+args = None
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train face network')
+    # general
+    parser.add_argument('--dataset',
+                        default=default.dataset,
+                        help='dataset config')
+    parser.add_argument('--network',
+                        default=default.network,
+                        help='network config')
+    parser.add_argument('--loss', default=default.loss, help='loss config')
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset, args.loss)
+    parser.add_argument('--models-root',
+                        default=default.models_root,
+                        help='root directory to save model.')
+    parser.add_argument('--pretrained',
+                        default=default.pretrained,
+                        help='pretrained model to load')
+    parser.add_argument('--pretrained-epoch',
+                        type=int,
+                        default=default.pretrained_epoch,
+                        help='pretrained epoch to load')
+    parser.add_argument(
+        '--ckpt',
+        type=int,
+        default=default.ckpt,
+        help=
+        'checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save'
+    )
+    parser.add_argument(
+        '--verbose',
+        type=int,
+        default=default.verbose,
+        help='do verification testing and model saving every verbose batches')
+    parser.add_argument('--lr',
+                        type=float,
+                        default=default.lr,
+                        help='start learning rate')
+    parser.add_argument('--lr-steps',
+                        type=str,
+                        default=default.lr_steps,
+                        help='steps of lr changing')
+    parser.add_argument('--wd',
+                        type=float,
+                        default=default.wd,
+                        help='weight decay')
+    parser.add_argument('--mom',
+                        type=float,
+                        default=default.mom,
+                        help='momentum')
+    parser.add_argument('--frequent',
+                        type=int,
+                        default=default.frequent,
+                        help='')
+    parser.add_argument('--per-batch-size',
+                        type=int,
+                        default=default.per_batch_size,
+                        help='batch size in each context')
+    parser.add_argument('--kvstore',
+                        type=str,
+                        default=default.kvstore,
+                        help='kvstore setting')
+    args = parser.parse_args()
+    return args
+
+
+def get_symbol(args):
+    embedding = eval(config.net_name).get_symbol()
+    all_label = mx.symbol.Variable('softmax_label')
+    gt_label = all_label
+    is_softmax = True
+    if config.loss_name == 'softmax':  #softmax
+        _weight = mx.symbol.Variable("fc7_weight",
+                                     shape=(config.num_classes,
+                                            config.emb_size),
+                                     lr_mult=config.fc7_lr_mult,
+                                     wd_mult=config.fc7_wd_mult,
+                                     init=mx.init.Normal(0.01))
+        if config.fc7_no_bias:
+            fc7 = mx.sym.FullyConnected(data=embedding,
+                                        weight=_weight,
+                                        no_bias=True,
+                                        num_hidden=config.num_classes,
+                                        name='fc7')
+        else:
+            _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
+            fc7 = mx.sym.FullyConnected(data=embedding,
+                                        weight=_weight,
+                                        bias=_bias,
+                                        num_hidden=config.num_classes,
+                                        name='fc7')
+    elif config.loss_name == 'margin_softmax':
+        _weight = mx.symbol.Variable("fc7_weight",
+                                     shape=(config.num_classes,
+                                            config.emb_size),
+                                     lr_mult=config.fc7_lr_mult,
+                                     wd_mult=config.fc7_wd_mult,
+                                     init=mx.init.Normal(0.01))
+        s = config.loss_s
+        _weight = mx.symbol.L2Normalization(_weight, mode='instance')
+        nembedding = mx.symbol.L2Normalization(
+            embedding, mode='instance', name='fc1n') * s
+        fc7 = mx.sym.FullyConnected(data=nembedding,
+                                    weight=_weight,
+                                    no_bias=True,
+                                    num_hidden=config.num_classes,
+                                    name='fc7')
+        if config.loss_m1 != 1.0 or config.loss_m2 != 0.0 or config.loss_m3 != 0.0:
+            if config.loss_m1 == 1.0 and config.loss_m2 == 0.0:
+                s_m = s * config.loss_m3
+                gt_one_hot = mx.sym.one_hot(gt_label,
+                                            depth=config.num_classes,
+                                            on_value=s_m,
+                                            off_value=0.0)
+                fc7 = fc7 - gt_one_hot
+            else:
+                zy = mx.sym.pick(fc7, gt_label, axis=1)
+                cos_t = zy / s
+                t = mx.sym.arccos(cos_t)
+                if config.loss_m1 != 1.0:
+                    t = t * config.loss_m1
+                if config.loss_m2 > 0.0:
+                    t = t + config.loss_m2
+                body = mx.sym.cos(t)
+                if config.loss_m3 > 0.0:
+                    body = body - config.loss_m3
+                new_zy = body * s
+                diff = new_zy - zy
+                diff = mx.sym.expand_dims(diff, 1)
+                gt_one_hot = mx.sym.one_hot(gt_label,
+                                            depth=config.num_classes,
+                                            on_value=1.0,
+                                            off_value=0.0)
+                body = mx.sym.broadcast_mul(gt_one_hot, diff)
+                fc7 = fc7 + body
+    elif config.loss_name.find('triplet') >= 0:
+        is_softmax = False
+        nembedding = mx.symbol.L2Normalization(embedding,
+                                               mode='instance',
+                                               name='fc1n')
+        anchor = mx.symbol.slice_axis(nembedding,
+                                      axis=0,
+                                      begin=0,
+                                      end=args.per_batch_size // 3)
+        positive = mx.symbol.slice_axis(nembedding,
+                                        axis=0,
+                                        begin=args.per_batch_size // 3,
+                                        end=2 * args.per_batch_size // 3)
+        negative = mx.symbol.slice_axis(nembedding,
+                                        axis=0,
+                                        begin=2 * args.per_batch_size // 3,
+                                        end=args.per_batch_size)
+        if config.loss_name == 'triplet':
+            ap = anchor - positive
+            an = anchor - negative
+            ap = ap * ap
+            an = an * an
+            ap = mx.symbol.sum(ap, axis=1, keepdims=1)  #(T,1)
+            an = mx.symbol.sum(an, axis=1, keepdims=1)  #(T,1)
+            triplet_loss = mx.symbol.Activation(data=(ap - an +
+                                                      config.triplet_alpha),
+                                                act_type='relu')
+            triplet_loss = mx.symbol.mean(triplet_loss)
+        else:
+            ap = anchor * positive
+            an = anchor * negative
+            ap = mx.symbol.sum(ap, axis=1, keepdims=1)  #(T,1)
+            an = mx.symbol.sum(an, axis=1, keepdims=1)  #(T,1)
+            ap = mx.sym.arccos(ap)
+            an = mx.sym.arccos(an)
+            triplet_loss = mx.symbol.Activation(data=(ap - an +
+                                                      config.triplet_alpha),
+                                                act_type='relu')
+            triplet_loss = mx.symbol.mean(triplet_loss)
+        triplet_loss = mx.symbol.MakeLoss(triplet_loss)
+    out_list = [mx.symbol.BlockGrad(embedding)]
+    if is_softmax:
+        softmax = mx.symbol.SoftmaxOutput(data=fc7,
+                                          label=gt_label,
+                                          name='softmax',
+                                          normalization='valid')
+        out_list.append(softmax)
+        if config.ce_loss:
+            #ce_loss = mx.symbol.softmax_cross_entropy(data=fc7, label = gt_label, name='ce_loss')/args.per_batch_size
+            body = mx.symbol.SoftmaxActivation(data=fc7)
+            body = mx.symbol.log(body)
+            _label = mx.sym.one_hot(gt_label,
+                                    depth=config.num_classes,
+                                    on_value=-1.0,
+                                    off_value=0.0)
+            body = body * _label
+            ce_loss = mx.symbol.sum(body) / args.per_batch_size
+            out_list.append(mx.symbol.BlockGrad(ce_loss))
+    else:
+        out_list.append(mx.sym.BlockGrad(gt_label))
+        out_list.append(triplet_loss)
+    out = mx.symbol.Group(out_list)
+    return out
+
+
+def train_net(args):
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    prefix = os.path.join(args.models_root,
+                          '%s-%s-%s' % (args.network, args.loss, args.dataset),
+                          'model')
+    prefix_dir = os.path.dirname(prefix)
+    print('prefix', prefix)
+    if not os.path.exists(prefix_dir):
+        os.makedirs(prefix_dir)
+    args.ctx_num = len(ctx)
+    args.batch_size = args.per_batch_size * args.ctx_num
+    args.rescale_threshold = 0
+    args.image_channel = config.image_shape[2]
+    config.batch_size = args.batch_size
+    config.per_batch_size = args.per_batch_size
+
+    data_dir = config.dataset_path
+    path_imgrec = None
+    path_imglist = None
+    image_size = config.image_shape[0:2]
+    assert len(image_size) == 2
+    assert image_size[0] == image_size[1]
+    print('image_size', image_size)
+    print('num_classes', config.num_classes)
+    path_imgrec = os.path.join(data_dir, "train.rec")
+
+    print('Called with argument:', args, config)
+    data_shape = (args.image_channel, image_size[0], image_size[1])
+    mean = None
+
+    begin_epoch = 0
+    if len(args.pretrained) == 0:
+        arg_params = None
+        aux_params = None
+        sym = get_symbol(args)
+        if config.net_name == 'spherenet':
+            data_shape_dict = {'data': (args.per_batch_size, ) + data_shape}
+            spherenet.init_weights(sym, data_shape_dict, args.num_layers)
+    else:
+        print('loading', args.pretrained, args.pretrained_epoch)
+        _, arg_params, aux_params = mx.model.load_checkpoint(
+            args.pretrained, args.pretrained_epoch)
+        sym = get_symbol(args)
+
+    if config.count_flops:
+        all_layers = sym.get_internals()
+        _sym = all_layers['fc1_output']
+        FLOPs = flops_counter.count_flops(_sym,
+                                          data=(1, 3, image_size[0],
+                                                image_size[1]))
+        _str = flops_counter.flops_str(FLOPs)
+        print('Network FLOPs: %s' % _str)
+
+    #label_name = 'softmax_label'
+    #label_shape = (args.batch_size,)
+    model = mx.mod.Module(
+        context=ctx,
+        symbol=sym,
+    )
+    val_dataiter = None
+
+    if config.loss_name.find('triplet') >= 0:
+        from triplet_image_iter import FaceImageIter
+        triplet_params = [
+            config.triplet_bag_size, config.triplet_alpha,
+            config.triplet_max_ap
+        ]
+        train_dataiter = FaceImageIter(
+            batch_size=args.batch_size,
+            data_shape=data_shape,
+            path_imgrec=path_imgrec,
+            shuffle=True,
+            rand_mirror=config.data_rand_mirror,
+            mean=mean,
+            cutoff=config.data_cutoff,
+            ctx_num=args.ctx_num,
+            images_per_identity=config.images_per_identity,
+            triplet_params=triplet_params,
+            mx_model=model,
+        )
+        _metric = LossValueMetric()
+        eval_metrics = [mx.metric.create(_metric)]
+    else:
+        #from image_iter import FaceImageIter
+        #train_dataiter = FaceImageIter(
+        #    batch_size=args.batch_size,
+        #    data_shape=data_shape,
+        #    path_imgrec=path_imgrec,
+        #    shuffle=True,
+        #    rand_mirror=config.data_rand_mirror,
+        #    mean=mean,
+        #    cutoff=config.data_cutoff,
+        #    color_jittering=config.data_color,
+        #    images_filter=config.data_images_filter,
+        #)
+        from image_iter import get_face_image_iter
+        train_dataiter = get_face_image_iter(config, data_shape, path_imgrec)
+        metric1 = AccMetric()
+        eval_metrics = [mx.metric.create(metric1)]
+        if config.ce_loss:
+            metric2 = LossValueMetric()
+            eval_metrics.append(mx.metric.create(metric2))
+
+    if config.net_name == 'fresnet' or config.net_name == 'fmobilefacenet':
+        initializer = mx.init.Xavier(rnd_type='gaussian',
+                                     factor_type="out",
+                                     magnitude=2)  #resnet style
+    else:
+        initializer = mx.init.Xavier(rnd_type='uniform',
+                                     factor_type="in",
+                                     magnitude=2)
+    #initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
+    _rescale = 1.0 / args.ctx_num
+    opt = optimizer.SGD(learning_rate=args.lr,
+                        momentum=args.mom,
+                        wd=args.wd,
+                        rescale_grad=_rescale)
+    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
+
+    ver_list = []
+    ver_name_list = []
+    for name in config.val_targets:
+        path = os.path.join(data_dir, name + ".bin")
+        if os.path.exists(path):
+            data_set = verification.load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+            print('ver', name)
+
+    def ver_test(nbatch):
+        results = []
+        for i in range(len(ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                ver_list[i], model, args.batch_size, 10, None, None)
+            print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
+            #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
+            print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                  (ver_name_list[i], nbatch, acc2, std2))
+            results.append(acc2)
+        return results
+
+    highest_acc = [0.0, 0.0]  #lfw and target
+    #for i in range(len(ver_list)):
+    #  highest_acc.append(0.0)
+    global_step = [0]
+    save_step = [0]
+    lr_steps = [int(x) for x in args.lr_steps.split(',')]
+    print('lr_steps', lr_steps)
+
+    def _batch_callback(param):
+        #global global_step
+        global_step[0] += 1
+        mbatch = global_step[0]
+        for step in lr_steps:
+            if mbatch == step:
+                opt.lr *= 0.1
+                print('lr change to', opt.lr)
+                break
+
+        _cb(param)
+        if mbatch % 1000 == 0:
+            print('lr-batch-epoch:', opt.lr, param.nbatch, param.epoch)
+
+        if mbatch >= 0 and mbatch % args.verbose == 0:
+            acc_list = ver_test(mbatch)
+            save_step[0] += 1
+            msave = save_step[0]
+            do_save = False
+            is_highest = False
+            if len(acc_list) > 0:
+                #lfw_score = acc_list[0]
+                #if lfw_score>highest_acc[0]:
+                #  highest_acc[0] = lfw_score
+                #  if lfw_score>=0.998:
+                #    do_save = True
+                score = sum(acc_list)
+                if acc_list[-1] >= highest_acc[-1]:
+                    if acc_list[-1] > highest_acc[-1]:
+                        is_highest = True
+                    else:
+                        if score >= highest_acc[0]:
+                            is_highest = True
+                            highest_acc[0] = score
+                    highest_acc[-1] = acc_list[-1]
+                    #if lfw_score>=0.99:
+                    #  do_save = True
+            if is_highest:
+                do_save = True
+            if args.ckpt == 0:
+                do_save = False
+            elif args.ckpt == 2:
+                do_save = True
+            elif args.ckpt == 3:
+                msave = 1
+
+            if do_save:
+                print('saving', msave)
+                arg, aux = model.get_params()
+                if config.ckpt_embedding:
+                    all_layers = model.symbol.get_internals()
+                    _sym = all_layers['fc1_output']
+                    _arg = {}
+                    for k in arg:
+                        if not k.startswith('fc7'):
+                            _arg[k] = arg[k]
+                    mx.model.save_checkpoint(prefix, msave, _sym, _arg, aux)
+                else:
+                    mx.model.save_checkpoint(prefix, msave, model.symbol, arg,
+                                             aux)
+            print('[%d]Accuracy-Highest: %1.5f' % (mbatch, highest_acc[-1]))
+        if config.max_steps > 0 and mbatch > config.max_steps:
+            sys.exit(0)
+
+    epoch_cb = None
+
+    model.fit(
+        train_dataiter,
+        begin_epoch=begin_epoch,
+        num_epoch=999999,
+        eval_data=val_dataiter,
+        eval_metric=eval_metrics,
+        kvstore=args.kvstore,
+        optimizer=opt,
+        #optimizer_params   = optimizer_params,
+        initializer=initializer,
+        arg_params=arg_params,
+        aux_params=aux_params,
+        allow_missing=True,
+        batch_end_callback=_batch_callback,
+        epoch_end_callback=epoch_cb)
+
+
+def main():
+    global args
+    args = parse_args()
+    train_net(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/recognition/arcface_mxnet/train_parall.py b/recognition/arcface_mxnet/train_parall.py
new file mode 100644
index 000000000..ce6b916fa
--- /dev/null
+++ b/recognition/arcface_mxnet/train_parall.py
@@ -0,0 +1,451 @@
+'''
+@author: insightface
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import sys
+import math
+import random
+import logging
+import pickle
+import sklearn
+import numpy as np
+#from image_iter import FaceImageIter
+from image_iter import get_face_image_iter
+import mxnet as mx
+from mxnet import ndarray as nd
+import argparse
+import mxnet.optimizer as optimizer
+sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
+import flops_counter
+from config import config, default, generate_config
+import verification
+sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
+import fresnet
+import fmobilefacenet
+import fmobilenet
+import fmnasnet
+import fdensenet
+import vargfacenet
+
+logger = logging.getLogger()
+logger.setLevel(logging.INFO)
+
+args = None
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train parall face network')
+    # general
+    parser.add_argument('--dataset',
+                        default=default.dataset,
+                        help='dataset config')
+    parser.add_argument('--network',
+                        default=default.network,
+                        help='network config')
+    parser.add_argument('--loss', default=default.loss, help='loss config')
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset, args.loss)
+    parser.add_argument('--models-root',
+                        default=default.models_root,
+                        help='root directory to save model.')
+    parser.add_argument('--pretrained',
+                        default=default.pretrained,
+                        help='pretrained model to load')
+    parser.add_argument('--pretrained-epoch',
+                        type=int,
+                        default=default.pretrained_epoch,
+                        help='pretrained epoch to load')
+    parser.add_argument(
+        '--ckpt',
+        type=int,
+        default=default.ckpt,
+        help=
+        'checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save'
+    )
+    parser.add_argument(
+        '--verbose',
+        type=int,
+        default=default.verbose,
+        help='do verification testing and model saving every verbose batches')
+    parser.add_argument('--lr',
+                        type=float,
+                        default=default.lr,
+                        help='start learning rate')
+    parser.add_argument('--lr-steps',
+                        type=str,
+                        default=default.lr_steps,
+                        help='steps of lr changing')
+    parser.add_argument('--wd',
+                        type=float,
+                        default=default.wd,
+                        help='weight decay')
+    parser.add_argument('--mom',
+                        type=float,
+                        default=default.mom,
+                        help='momentum')
+    parser.add_argument('--frequent',
+                        type=int,
+                        default=default.frequent,
+                        help='')
+    parser.add_argument('--per-batch-size',
+                        type=int,
+                        default=default.per_batch_size,
+                        help='batch size in each context')
+    parser.add_argument('--kvstore',
+                        type=str,
+                        default=default.kvstore,
+                        help='kvstore setting')
+    parser.add_argument('--worker-id',
+                        type=int,
+                        default=0,
+                        help='worker id for dist training, starts from 0')
+    parser.add_argument('--extra-model-name',
+                        type=str,
+                        default='',
+                        help='extra model name')
+    parser.add_argument('--fp16-scale', type=float, default=0.0, help='')
+    args = parser.parse_args()
+    return args
+
+
+def get_symbol_embedding(embedding=None):
+    if embedding is None:
+        embedding = eval(config.net_name).get_symbol()
+    all_label = mx.symbol.Variable('softmax_label')
+    #embedding = mx.symbol.BlockGrad(embedding)
+    all_label = mx.symbol.BlockGrad(all_label)
+    out_list = [embedding, all_label]
+    out = mx.symbol.Group(out_list)
+    return out
+
+
+def get_symbol_arcface(args):
+    embedding = mx.symbol.Variable('data')
+    all_label = mx.symbol.Variable('softmax_label')
+    gt_label = all_label
+    is_softmax = True
+    #print('call get_sym_arcface with', args, config)
+    _weight = mx.symbol.Variable("fc7_%d_weight" % args._ctxid,
+                                 shape=(args.ctx_num_classes, config.emb_size),
+                                 lr_mult=config.fc7_lr_mult,
+                                 wd_mult=config.fc7_wd_mult)
+    if config.loss_name == 'softmax':  #softmax
+        fc7 = mx.sym.FullyConnected(data=embedding,
+                                    weight=_weight,
+                                    no_bias=True,
+                                    num_hidden=args.ctx_num_classes,
+                                    name='fc7_%d' % args._ctxid)
+    elif config.loss_name == 'margin_softmax':
+        _weight = mx.symbol.L2Normalization(_weight, mode='instance')
+        nembedding = mx.symbol.L2Normalization(embedding,
+                                               mode='instance',
+                                               name='fc1n_%d' % args._ctxid)
+        fc7 = mx.sym.FullyConnected(data=nembedding,
+                                    weight=_weight,
+                                    no_bias=True,
+                                    num_hidden=args.ctx_num_classes,
+                                    name='fc7_%d' % args._ctxid)
+        if config.loss_m1 != 1.0 or config.loss_m2 != 0.0 or config.loss_m3 != 0.0:
+            gt_one_hot = mx.sym.one_hot(gt_label,
+                                        depth=args.ctx_num_classes,
+                                        on_value=1.0,
+                                        off_value=0.0)
+            if config.loss_m1 == 1.0 and config.loss_m2 == 0.0:
+                _one_hot = gt_one_hot * config.loss_m3
+                fc7 = fc7 - _one_hot
+            else:
+                fc7_onehot = fc7 * gt_one_hot
+                cos_t = fc7_onehot
+                t = mx.sym.arccos(cos_t)
+                if config.loss_m1 != 1.0:
+                    t = t * config.loss_m1
+                if config.loss_m2 != 0.0:
+                    t = t + config.loss_m2
+                margin_cos = mx.sym.cos(t)
+                if config.loss_m3 != 0.0:
+                    margin_cos = margin_cos - config.loss_m3
+                margin_fc7 = margin_cos
+                margin_fc7_onehot = margin_fc7 * gt_one_hot
+                diff = margin_fc7_onehot - fc7_onehot
+                fc7 = fc7 + diff
+        fc7 = fc7 * config.loss_s
+
+    out_list = []
+    out_list.append(fc7)
+    if config.loss_name == 'softmax':  #softmax
+        out_list.append(gt_label)
+    out = mx.symbol.Group(out_list)
+    return out
+
+
+def train_net(args):
+    #_seed = 727
+    #random.seed(_seed)
+    #np.random.seed(_seed)
+    #mx.random.seed(_seed)
+    config.fp16 = False
+    config.fp16_scale = 0.0
+    if args.fp16_scale>0.0:
+        config.fp16 = True
+        config.fp16_scale = args.fp16_scale
+        print('use fp16, scale=', config.fp16_scale)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    if len(args.extra_model_name) == 0:
+        prefix = os.path.join(
+            args.models_root,
+            '%s-%s-%s' % (args.network, args.loss, args.dataset), 'model')
+    else:
+        prefix = os.path.join(
+            args.models_root, '%s-%s-%s-%s' %
+            (args.network, args.loss, args.dataset, args.extra_model_name),
+            'model')
+    prefix_dir = os.path.dirname(prefix)
+    print('prefix', prefix)
+    if not os.path.exists(prefix_dir):
+        os.makedirs(prefix_dir)
+    args.ctx_num = len(ctx)
+    if args.per_batch_size == 0:
+        args.per_batch_size = 128
+    args.batch_size = args.per_batch_size * args.ctx_num
+    args.rescale_threshold = 0
+    args.image_channel = config.image_shape[2]
+    config.batch_size = args.batch_size
+    config.per_batch_size = args.per_batch_size
+    data_dir = config.dataset_path
+    path_imgrec = None
+    path_imglist = None
+    image_size = config.image_shape[0:2]
+    assert len(image_size) == 2
+    assert image_size[0] == image_size[1]
+    print('image_size', image_size)
+    print('num_classes', config.num_classes)
+    path_imgrec = os.path.join(data_dir, "train.rec")
+
+    data_shape = (args.image_channel, image_size[0], image_size[1])
+
+    num_workers = config.num_workers
+    global_num_ctx = num_workers * args.ctx_num
+    if config.num_classes % global_num_ctx == 0:
+        args.ctx_num_classes = config.num_classes // global_num_ctx
+    else:
+        args.ctx_num_classes = config.num_classes // global_num_ctx + 1
+    args.local_num_classes = args.ctx_num_classes * args.ctx_num
+    args.local_class_start = args.local_num_classes * args.worker_id
+
+    #if len(args.partial)==0:
+    #  local_classes_range = (0, args.num_classes)
+    #else:
+    #  _vec = args.partial.split(',')
+    #  local_classes_range = (int(_vec[0]), int(_vec[1]))
+
+    #args.partial_num_classes = local_classes_range[1] - local_classes_range[0]
+    #args.partial_start = local_classes_range[0]
+
+    print('Called with argument:', args, config)
+    mean = None
+
+    begin_epoch = 0
+    base_lr = args.lr
+    base_wd = args.wd
+    base_mom = args.mom
+    arg_params = None
+    aux_params = None
+    if len(args.pretrained) == 0:
+        esym = get_symbol_embedding()
+        asym = get_symbol_arcface
+    else:
+        #assert False
+        print('loading', args.pretrained, args.pretrained_epoch)
+        pretrain_esym, arg_params, aux_params = mx.model.load_checkpoint(
+            args.pretrained, args.pretrained_epoch)
+        esym = get_symbol_embedding(pretrain_esym)
+        asym = get_symbol_arcface
+
+    if config.count_flops:
+        all_layers = esym.get_internals()
+        _sym = all_layers['fc1_output']
+        FLOPs = flops_counter.count_flops(_sym,
+                                          data=(1, 3, image_size[0],
+                                                image_size[1]))
+        _str = flops_counter.flops_str(FLOPs)
+        print('Network FLOPs: %s' % _str)
+
+    if config.num_workers == 1:
+        from parall_module_local_v1 import ParallModule
+    else:
+        from parall_module_dist import ParallModule
+
+    model = ParallModule(
+        context=ctx,
+        symbol=esym,
+        data_names=['data'],
+        label_names=['softmax_label'],
+        asymbol=asym,
+        args=args,
+    )
+    val_dataiter = None
+
+    if config.net_name == 'fresnet' or config.net_name == 'fmobilefacenet':
+        initializer = mx.init.Xavier(rnd_type='gaussian',
+                                     factor_type="out",
+                                     magnitude=2)  #resnet style
+    else:
+        initializer = mx.init.Xavier(rnd_type='uniform',
+                                     factor_type="in",
+                                     magnitude=2)
+
+    _rescale = 1.0 / args.batch_size
+    if config.fp16:
+        opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale, multi_precision=True)
+    else:
+        opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale, multi_precision=False)
+    opt_fc7 = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale, multi_precision=False)
+    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
+
+    ver_list = []
+    ver_name_list = []
+    for name in config.val_targets:
+        path = os.path.join(data_dir, name + ".bin")
+        if os.path.exists(path):
+            data_set = verification.load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+            print('ver', name)
+
+    def ver_test(nbatch):
+        results = []
+        for i in range(len(ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                ver_list[i], model, args.batch_size, 10, None, None)
+            print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
+            #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
+            print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                  (ver_name_list[i], nbatch, acc2, std2))
+            results.append(acc2)
+        return results
+
+    highest_acc = [0.0, 0.0]  #lfw and target
+    #for i in range(len(ver_list)):
+    #  highest_acc.append(0.0)
+    global_step = [0]
+    save_step = [0]
+    lr_steps = [int(x) for x in args.lr_steps.split(',')]
+    print('lr_steps', lr_steps)
+
+    def _batch_callback(param):
+        #global global_step
+        global_step[0] += 1
+        mbatch = global_step[0]
+        for step in lr_steps:
+            if mbatch == step:
+                opt.lr *= 0.1
+                opt_fc7.lr *= 0.1
+                print('lr change to', opt.lr, opt_fc7.lr)
+                break
+
+        _cb(param)
+        if mbatch % 1000 == 0:
+            print('lr-batch-epoch:', opt.lr, opt_fc7.lr, param.nbatch, param.epoch)
+
+        if mbatch >= 0 and mbatch % args.verbose == 0:
+            acc_list = ver_test(mbatch)
+            save_step[0] += 1
+            msave = save_step[0]
+            do_save = False
+            is_highest = False
+            if len(acc_list) > 0:
+                #lfw_score = acc_list[0]
+                #if lfw_score>highest_acc[0]:
+                #  highest_acc[0] = lfw_score
+                #  if lfw_score>=0.998:
+                #    do_save = True
+                score = sum(acc_list)
+                if acc_list[-1] >= highest_acc[-1]:
+                    if acc_list[-1] > highest_acc[-1]:
+                        is_highest = True
+                    else:
+                        if score >= highest_acc[0]:
+                            is_highest = True
+                            highest_acc[0] = score
+                    highest_acc[-1] = acc_list[-1]
+                    #if lfw_score>=0.99:
+                    #  do_save = True
+            if is_highest:
+                do_save = True
+            if args.ckpt == 0:
+                do_save = False
+            elif args.ckpt == 2:
+                do_save = True
+            elif args.ckpt == 3:
+                msave = 1
+
+            if do_save:
+                print('saving', msave)
+                arg, aux = model.get_export_params()
+                all_layers = model.symbol.get_internals()
+                _sym = all_layers['fc1_output']
+                mx.model.save_checkpoint(prefix, msave, _sym, arg, aux)
+            print('[%d]Accuracy-Highest: %1.5f' % (mbatch, highest_acc[-1]))
+        if config.max_steps > 0 and mbatch > config.max_steps:
+            msave = 0
+            config.fp16 = False
+            print('saving last', msave)
+            arg, aux = model.get_export_params()
+            _sym = eval(config.net_name).get_symbol()
+            mx.model.save_checkpoint(prefix, msave, _sym, arg, aux)
+            sys.exit(0)
+
+    epoch_cb = None
+    train_dataiter = get_face_image_iter(config, data_shape, path_imgrec)
+    #train_dataiter = FaceImageIter(
+    #    batch_size=args.batch_size,
+    #    data_shape=data_shape,
+    #    path_imgrec=path_imgrec,
+    #    shuffle=True,
+    #    rand_mirror=config.data_rand_mirror,
+    #    mean=mean,
+    #    cutoff=config.data_cutoff,
+    #    color_jittering=config.data_color,
+    #    images_filter=config.data_images_filter,
+    #)
+    #train_dataiter = mx.io.PrefetchingIter(train_dataiter)
+
+    model.fit(
+        train_dataiter,
+        begin_epoch=begin_epoch,
+        num_epoch=999999,
+        eval_data=val_dataiter,
+        #eval_metric        = eval_metrics,
+        kvstore=args.kvstore,
+        optimizer=[opt, opt_fc7],
+        #optimizer_params   = optimizer_params,
+        initializer=initializer,
+        arg_params=arg_params,
+        aux_params=aux_params,
+        allow_missing=True,
+        batch_end_callback=_batch_callback,
+        epoch_end_callback=epoch_cb)
+
+
+def main():
+    global args
+    args = parse_args()
+    train_net(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/recognition/arcface_mxnet/triplet_image_iter.py b/recognition/arcface_mxnet/triplet_image_iter.py
new file mode 100644
index 000000000..1df39acf8
--- /dev/null
+++ b/recognition/arcface_mxnet/triplet_image_iter.py
@@ -0,0 +1,628 @@
+# THIS FILE IS FOR EXPERIMENTS, USE image_iter.py FOR NORMAL IMAGE LOADING.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import random
+import logging
+import sys
+import numbers
+import math
+import sklearn
+import datetime
+import numpy as np
+import cv2
+
+import mxnet as mx
+from mxnet import ndarray as nd
+#from . import _ndarray_internal as _internal
+#from mxnet._ndarray_internal import _cvimresize as imresize
+#from ._ndarray_internal import _cvcopyMakeBorder as copyMakeBorder
+from mxnet import io
+from mxnet import recordio
+sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
+import face_preprocess
+
+logger = logging.getLogger()
+
+
+class FaceImageIter(io.DataIter):
+    def __init__(self,
+                 batch_size,
+                 data_shape,
+                 path_imgrec=None,
+                 shuffle=False,
+                 aug_list=None,
+                 rand_mirror=False,
+                 cutoff=0,
+                 ctx_num=0,
+                 images_per_identity=0,
+                 triplet_params=None,
+                 mx_model=None,
+                 data_name='data',
+                 label_name='softmax_label',
+                 **kwargs):
+        super(FaceImageIter, self).__init__()
+        assert path_imgrec
+        assert shuffle
+        logging.info('loading recordio %s...', path_imgrec)
+        path_imgidx = path_imgrec[0:-4] + ".idx"
+        self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+        s = self.imgrec.read_idx(0)
+        header, _ = recordio.unpack(s)
+        assert header.flag > 0
+        print('header0 label', header.label)
+        self.header0 = (int(header.label[0]), int(header.label[1]))
+        #assert(header.flag==1)
+        self.imgidx = range(1, int(header.label[0]))
+        self.id2range = {}
+        self.seq_identity = range(int(header.label[0]), int(header.label[1]))
+        for identity in self.seq_identity:
+            s = self.imgrec.read_idx(identity)
+            header, _ = recordio.unpack(s)
+            a, b = int(header.label[0]), int(header.label[1])
+            self.id2range[identity] = (a, b)
+
+        print('id2range', len(self.id2range))
+        self.seq = self.imgidx
+        print(len(self.seq))
+
+        self.check_data_shape(data_shape)
+        self.provide_data = [(data_name, (batch_size, ) + data_shape)]
+        self.batch_size = batch_size
+        self.data_shape = data_shape
+        self.shuffle = shuffle
+        self.image_size = '%d,%d' % (data_shape[1], data_shape[2])
+        self.rand_mirror = rand_mirror
+        print('rand_mirror', rand_mirror)
+        self.cutoff = cutoff
+        #self.cast_aug = mx.image.CastAug()
+        #self.color_aug = mx.image.ColorJitterAug(0.4, 0.4, 0.4)
+        self.ctx_num = ctx_num
+        self.per_batch_size = int(self.batch_size / self.ctx_num)
+        self.images_per_identity = images_per_identity
+        if self.images_per_identity > 0:
+            self.identities = int(self.per_batch_size /
+                                  self.images_per_identity)
+            self.per_identities = self.identities
+            self.repeat = 3000000.0 / (self.images_per_identity *
+                                       len(self.id2range))
+            self.repeat = int(self.repeat)
+            print(self.images_per_identity, self.identities, self.repeat)
+        self.mx_model = mx_model
+        self.triplet_params = triplet_params
+        self.triplet_mode = False
+        #self.provide_label = None
+        self.provide_label = [(label_name, (batch_size, ))]
+        if self.triplet_params is not None:
+            assert self.images_per_identity > 0
+            assert self.mx_model is not None
+            self.triplet_bag_size = self.triplet_params[0]
+            self.triplet_alpha = self.triplet_params[1]
+            self.triplet_max_ap = self.triplet_params[2]
+            assert self.triplet_bag_size > 0
+            assert self.triplet_alpha >= 0.0
+            assert self.triplet_alpha <= 1.0
+            self.triplet_mode = True
+            self.triplet_cur = 0
+            self.triplet_seq = []
+            self.triplet_reset()
+            self.seq_min_size = self.batch_size * 2
+        self.cur = 0
+        self.nbatch = 0
+        self.is_init = False
+        self.times = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
+        #self.reset()
+
+    def pairwise_dists(self, embeddings):
+        nd_embedding_list = []
+        for i in range(self.ctx_num):
+            nd_embedding = mx.nd.array(embeddings, mx.gpu(i))
+            nd_embedding_list.append(nd_embedding)
+        nd_pdists = []
+        pdists = []
+        for idx in range(embeddings.shape[0]):
+            emb_idx = idx % self.ctx_num
+            nd_embedding = nd_embedding_list[emb_idx]
+            a_embedding = nd_embedding[idx]
+            body = mx.nd.broadcast_sub(a_embedding, nd_embedding)
+            body = body * body
+            body = mx.nd.sum_axis(body, axis=1)
+            nd_pdists.append(body)
+            if len(nd_pdists
+                   ) == self.ctx_num or idx == embeddings.shape[0] - 1:
+                for x in nd_pdists:
+                    pdists.append(x.asnumpy())
+                nd_pdists = []
+        return pdists
+
+    def pick_triplets(self, embeddings, nrof_images_per_class):
+        emb_start_idx = 0
+        triplets = []
+        people_per_batch = len(nrof_images_per_class)
+        #self.time_reset()
+        pdists = self.pairwise_dists(embeddings)
+        #self.times[3] += self.time_elapsed()
+
+        for i in range(people_per_batch):
+            nrof_images = int(nrof_images_per_class[i])
+            for j in range(1, nrof_images):
+                #self.time_reset()
+                a_idx = emb_start_idx + j - 1
+                #neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
+                neg_dists_sqr = pdists[a_idx]
+                #self.times[3] += self.time_elapsed()
+
+                for pair in range(
+                        j, nrof_images):  # For every possible positive pair.
+                    p_idx = emb_start_idx + pair
+                    #self.time_reset()
+                    pos_dist_sqr = np.sum(
+                        np.square(embeddings[a_idx] - embeddings[p_idx]))
+                    #self.times[4] += self.time_elapsed()
+                    #self.time_reset()
+                    neg_dists_sqr[emb_start_idx:emb_start_idx +
+                                  nrof_images] = np.NaN
+                    if self.triplet_max_ap > 0.0:
+                        if pos_dist_sqr > self.triplet_max_ap:
+                            continue
+                    all_neg = np.where(
+                        np.logical_and(
+                            neg_dists_sqr - pos_dist_sqr < self.triplet_alpha,
+                            pos_dist_sqr <
+                            neg_dists_sqr))[0]  # FaceNet selection
+                    #self.times[5] += self.time_elapsed()
+                    #self.time_reset()
+                    #all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
+                    nrof_random_negs = all_neg.shape[0]
+                    if nrof_random_negs > 0:
+                        rnd_idx = np.random.randint(nrof_random_negs)
+                        n_idx = all_neg[rnd_idx]
+                        triplets.append((a_idx, p_idx, n_idx))
+            emb_start_idx += nrof_images
+        np.random.shuffle(triplets)
+        return triplets
+
+    def triplet_reset(self):
+        #reset self.oseq by identities seq
+        self.triplet_cur = 0
+        ids = []
+        for k in self.id2range:
+            ids.append(k)
+        random.shuffle(ids)
+        self.triplet_seq = []
+        for _id in ids:
+            v = self.id2range[_id]
+            _list = range(*v)
+            random.shuffle(_list)
+            if len(_list) > self.images_per_identity:
+                _list = _list[0:self.images_per_identity]
+            self.triplet_seq += _list
+        print('triplet_seq', len(self.triplet_seq))
+        assert len(self.triplet_seq) >= self.triplet_bag_size
+
+    def time_reset(self):
+        self.time_now = datetime.datetime.now()
+
+    def time_elapsed(self):
+        time_now = datetime.datetime.now()
+        diff = time_now - self.time_now
+        return diff.total_seconds()
+
+    def select_triplets(self):
+        self.seq = []
+        while len(self.seq) < self.seq_min_size:
+            self.time_reset()
+            embeddings = None
+            bag_size = self.triplet_bag_size
+            batch_size = self.batch_size
+            #data = np.zeros( (bag_size,)+self.data_shape )
+            #label = np.zeros( (bag_size,) )
+            tag = []
+            #idx = np.zeros( (bag_size,) )
+            print('eval %d images..' % bag_size, self.triplet_cur)
+            print('triplet time stat', self.times)
+            if self.triplet_cur + bag_size > len(self.triplet_seq):
+                self.triplet_reset()
+                #bag_size = min(bag_size, len(self.triplet_seq))
+                print('eval %d images..' % bag_size, self.triplet_cur)
+            self.times[0] += self.time_elapsed()
+            self.time_reset()
+            #print(data.shape)
+            data = nd.zeros(self.provide_data[0][1])
+            label = None
+            if self.provide_label is not None:
+                label = nd.zeros(self.provide_label[0][1])
+            ba = 0
+            while True:
+                bb = min(ba + batch_size, bag_size)
+                if ba >= bb:
+                    break
+                _count = bb - ba
+                #data = nd.zeros( (_count,)+self.data_shape )
+                #_batch = self.data_iter.next()
+                #_data = _batch.data[0].asnumpy()
+                #print(_data.shape)
+                #_label = _batch.label[0].asnumpy()
+                #data[ba:bb,:,:,:] = _data
+                #label[ba:bb] = _label
+                for i in range(ba, bb):
+                    #print(ba, bb, self.triplet_cur, i, len(self.triplet_seq))
+                    _idx = self.triplet_seq[i + self.triplet_cur]
+                    s = self.imgrec.read_idx(_idx)
+                    header, img = recordio.unpack(s)
+                    img = self.imdecode(img)
+                    data[i - ba][:] = self.postprocess_data(img)
+                    _label = header.label
+                    if not isinstance(_label, numbers.Number):
+                        _label = _label[0]
+                    if label is not None:
+                        label[i - ba][:] = _label
+                    tag.append((int(_label), _idx))
+                    #idx[i] = _idx
+
+                db = mx.io.DataBatch(data=(data, ))
+                self.mx_model.forward(db, is_train=False)
+                net_out = self.mx_model.get_outputs()
+                #print('eval for selecting triplets',ba,bb)
+                #print(net_out)
+                #print(len(net_out))
+                #print(net_out[0].asnumpy())
+                net_out = net_out[0].asnumpy()
+                #print(net_out)
+                #print('net_out', net_out.shape)
+                if embeddings is None:
+                    embeddings = np.zeros((bag_size, net_out.shape[1]))
+                embeddings[ba:bb, :] = net_out
+                ba = bb
+            assert len(tag) == bag_size
+            self.triplet_cur += bag_size
+            embeddings = sklearn.preprocessing.normalize(embeddings)
+            self.times[1] += self.time_elapsed()
+            self.time_reset()
+            nrof_images_per_class = [1]
+            for i in range(1, bag_size):
+                if tag[i][0] == tag[i - 1][0]:
+                    nrof_images_per_class[-1] += 1
+                else:
+                    nrof_images_per_class.append(1)
+
+            triplets = self.pick_triplets(embeddings,
+                                          nrof_images_per_class)  # shape=(T,3)
+            print('found triplets', len(triplets))
+            ba = 0
+            while True:
+                bb = ba + self.per_batch_size // 3
+                if bb > len(triplets):
+                    break
+                _triplets = triplets[ba:bb]
+                for i in range(3):
+                    for triplet in _triplets:
+                        _pos = triplet[i]
+                        _idx = tag[_pos][1]
+                        self.seq.append(_idx)
+                ba = bb
+            self.times[2] += self.time_elapsed()
+
+    def hard_mining_reset(self):
+        #import faiss
+        from annoy import AnnoyIndex
+        data = nd.zeros(self.provide_data[0][1])
+        label = nd.zeros(self.provide_label[0][1])
+        #label = np.zeros( self.provide_label[0][1] )
+        X = None
+        ba = 0
+        batch_num = 0
+        while ba < len(self.oseq):
+            batch_num += 1
+            if batch_num % 10 == 0:
+                print('loading batch', batch_num, ba)
+            bb = min(ba + self.batch_size, len(self.oseq))
+            _count = bb - ba
+            for i in range(_count):
+                idx = self.oseq[i + ba]
+                s = self.imgrec.read_idx(idx)
+                header, img = recordio.unpack(s)
+                img = self.imdecode(img)
+                data[i][:] = self.postprocess_data(img)
+                label[i][:] = header.label
+            db = mx.io.DataBatch(data=(data, self.data_extra), label=(label, ))
+            self.mx_model.forward(db, is_train=False)
+            net_out = self.mx_model.get_outputs()
+            embedding = net_out[0].asnumpy()
+            nembedding = sklearn.preprocessing.normalize(embedding)
+            if _count < self.batch_size:
+                nembedding = nembedding[0:_count, :]
+            if X is None:
+                X = np.zeros((len(self.id2range), nembedding.shape[1]),
+                             dtype=np.float32)
+            nplabel = label.asnumpy()
+            for i in range(_count):
+                ilabel = int(nplabel[i])
+                #print(ilabel, ilabel.__class__)
+                X[ilabel] += nembedding[i]
+            ba = bb
+        X = sklearn.preprocessing.normalize(X)
+        d = X.shape[1]
+        t = AnnoyIndex(d, metric='euclidean')
+        for i in range(X.shape[0]):
+            t.add_item(i, X[i])
+        print('start to build index')
+        t.build(20)
+        print(X.shape)
+        k = self.per_identities
+        self.seq = []
+        for i in range(X.shape[0]):
+            nnlist = t.get_nns_by_item(i, k)
+            assert nnlist[0] == i
+            for _label in nnlist:
+                assert _label < len(self.id2range)
+                _id = self.header0[0] + _label
+                v = self.id2range[_id]
+                _list = range(*v)
+                if len(_list) < self.images_per_identity:
+                    random.shuffle(_list)
+                else:
+                    _list = np.random.choice(_list,
+                                             self.images_per_identity,
+                                             replace=False)
+                for i in range(self.images_per_identity):
+                    _idx = _list[i % len(_list)]
+                    self.seq.append(_idx)
+        #faiss_params = [20,5]
+        #quantizer = faiss.IndexFlatL2(d)  # the other index
+        #index = faiss.IndexIVFFlat(quantizer, d, faiss_params[0], faiss.METRIC_L2)
+        #assert not index.is_trained
+        #index.train(X)
+        #index.add(X)
+        #assert index.is_trained
+        #print('trained')
+        #index.nprobe = faiss_params[1]
+        #D, I = index.search(X, k)     # actual search
+        #print(I.shape)
+        #self.seq = []
+        #for i in range(I.shape[0]):
+        #  #assert I[i][0]==i
+        #  for j in range(k):
+        #    _label = I[i][j]
+        #    assert _label<len(self.id2range)
+        #    _id = self.header0[0]+_label
+        #    v = self.id2range[_id]
+        #    _list = range(*v)
+        #    if len(_list)<self.images_per_identity:
+        #      random.shuffle(_list)
+        #    else:
+        #      _list = np.random.choice(_list, self.images_per_identity, replace=False)
+        #    for i in range(self.images_per_identity):
+        #      _idx = _list[i%len(_list)]
+        #      self.seq.append(_idx)
+
+    def reset(self):
+        """Resets the iterator to the beginning of the data."""
+        print('call reset()')
+        self.cur = 0
+        if self.images_per_identity > 0:
+            if self.triplet_mode:
+                self.select_triplets()
+            elif not self.hard_mining:
+                self.seq = []
+                idlist = []
+                for _id in self.id2range:
+                    v = self.id2range[_id]
+                    idlist.append((_id, range(*v)))
+                for r in range(self.repeat):
+                    if r % 10 == 0:
+                        print('repeat', r)
+                    if self.shuffle:
+                        random.shuffle(idlist)
+                    for item in idlist:
+                        _id = item[0]
+                        _list = item[1]
+                        #random.shuffle(_list)
+                        if len(_list) < self.images_per_identity:
+                            random.shuffle(_list)
+                        else:
+                            _list = np.random.choice(_list,
+                                                     self.images_per_identity,
+                                                     replace=False)
+                        for i in range(self.images_per_identity):
+                            _idx = _list[i % len(_list)]
+                            self.seq.append(_idx)
+            else:
+                self.hard_mining_reset()
+            print('seq len', len(self.seq))
+        else:
+            if self.shuffle:
+                random.shuffle(self.seq)
+        if self.seq is None and self.imgrec is not None:
+            self.imgrec.reset()
+
+    def num_samples(self):
+        return len(self.seq)
+
+    def next_sample(self):
+        while True:
+            if self.cur >= len(self.seq):
+                raise StopIteration
+            idx = self.seq[self.cur]
+            self.cur += 1
+            s = self.imgrec.read_idx(idx)
+            header, img = recordio.unpack(s)
+            label = header.label
+            if not isinstance(label, numbers.Number):
+                label = label[0]
+            return label, img, None, None
+
+    def brightness_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        src *= alpha
+        return src
+
+    def contrast_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = np.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
+        src *= alpha
+        src += gray
+        return src
+
+    def saturation_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = np.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = np.sum(gray, axis=2, keepdims=True)
+        gray *= (1.0 - alpha)
+        src *= alpha
+        src += gray
+        return src
+
+    def color_aug(self, img, x):
+        augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
+        random.shuffle(augs)
+        for aug in augs:
+            #print(img.shape)
+            img = aug(img, x)
+            #print(img.shape)
+        return img
+
+    def mirror_aug(self, img):
+        _rd = random.randint(0, 1)
+        if _rd == 1:
+            for c in range(img.shape[2]):
+                img[:, :, c] = np.fliplr(img[:, :, c])
+        return img
+
+    def next(self):
+        if not self.is_init:
+            self.reset()
+            self.is_init = True
+        """Returns the next batch of data."""
+        #print('in next', self.cur, self.labelcur)
+        self.nbatch += 1
+        batch_size = self.batch_size
+        c, h, w = self.data_shape
+        batch_data = nd.empty((batch_size, c, h, w))
+        if self.provide_label is not None:
+            batch_label = nd.empty(self.provide_label[0][1])
+        i = 0
+        try:
+            while i < batch_size:
+                label, s, bbox, landmark = self.next_sample()
+                _data = self.imdecode(s)
+                if self.rand_mirror:
+                    _rd = random.randint(0, 1)
+                    if _rd == 1:
+                        _data = mx.ndarray.flip(data=_data, axis=1)
+                if self.cutoff > 0:
+                    centerh = random.randint(0, _data.shape[0] - 1)
+                    centerw = random.randint(0, _data.shape[1] - 1)
+                    half = self.cutoff // 2
+                    starth = max(0, centerh - half)
+                    endh = min(_data.shape[0], centerh + half)
+                    startw = max(0, centerw - half)
+                    endw = min(_data.shape[1], centerw + half)
+                    _data = _data.astype('float32')
+                    #print(starth, endh, startw, endw, _data.shape)
+                    _data[starth:endh, startw:endw, :] = 127.5
+                #_npdata = _data.asnumpy()
+                #if landmark is not None:
+                #  _npdata = face_preprocess.preprocess(_npdata, bbox = bbox, landmark=landmark, image_size=self.image_size)
+                #if self.rand_mirror:
+                #  _npdata = self.mirror_aug(_npdata)
+                #if self.mean is not None:
+                #  _npdata = _npdata.astype(np.float32)
+                #  _npdata -= self.mean
+                #  _npdata *= 0.0078125
+                #nimg = np.zeros(_npdata.shape, dtype=np.float32)
+                #nimg[self.patch[1]:self.patch[3],self.patch[0]:self.patch[2],:] = _npdata[self.patch[1]:self.patch[3], self.patch[0]:self.patch[2], :]
+                #_data = mx.nd.array(nimg)
+                data = [_data]
+                try:
+                    self.check_valid_image(data)
+                except RuntimeError as e:
+                    logging.debug('Invalid image, skipping:  %s', str(e))
+                    continue
+                #print('aa',data[0].shape)
+                #data = self.augmentation_transform(data)
+                #print('bb',data[0].shape)
+                for datum in data:
+                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
+                    #print(datum.shape)
+                    batch_data[i][:] = self.postprocess_data(datum)
+                    if self.provide_label is not None:
+                        batch_label[i][:] = label
+                    i += 1
+        except StopIteration:
+            if i < batch_size:
+                raise StopIteration
+
+        #print('next end', batch_size, i)
+        _label = None
+        if self.provide_label is not None:
+            _label = [batch_label]
+        return io.DataBatch([batch_data], _label, batch_size - i)
+
+    def check_data_shape(self, data_shape):
+        """Checks if the input data shape is valid"""
+        if not len(data_shape) == 3:
+            raise ValueError(
+                'data_shape should have length 3, with dimensions CxHxW')
+        if not data_shape[0] == 3:
+            raise ValueError(
+                'This iterator expects inputs to have 3 channels.')
+
+    def check_valid_image(self, data):
+        """Checks if the input data is valid"""
+        if len(data[0].shape) == 0:
+            raise RuntimeError('Data shape is wrong')
+
+    def imdecode(self, s):
+        """Decodes a string or byte string to an NDArray.
+        See mx.img.imdecode for more details."""
+        img = mx.image.imdecode(s)  #mx.ndarray
+        return img
+
+    def read_image(self, fname):
+        """Reads an input image `fname` and returns the decoded raw bytes.
+
+        Example usage:
+        ----------
+        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
+        """
+        with open(os.path.join(self.path_root, fname), 'rb') as fin:
+            img = fin.read()
+        return img
+
+    def augmentation_transform(self, data):
+        """Transforms input data with specified augmentation."""
+        for aug in self.auglist:
+            data = [ret for src in data for ret in aug(src)]
+        return data
+
+    def postprocess_data(self, datum):
+        """Final postprocessing step before image is loaded into the batch."""
+        return nd.transpose(datum, axes=(2, 0, 1))
+
+
+class FaceImageIterList(io.DataIter):
+    def __init__(self, iter_list):
+        assert len(iter_list) > 0
+        self.provide_data = iter_list[0].provide_data
+        self.provide_label = iter_list[0].provide_label
+        self.iter_list = iter_list
+        self.cur_iter = None
+
+    def reset(self):
+        self.cur_iter.reset()
+
+    def next(self):
+        self.cur_iter = random.choice(self.iter_list)
+        while True:
+            try:
+                ret = self.cur_iter.next()
+            except StopIteration:
+                self.cur_iter.reset()
+                continue
+            return ret
diff --git a/recognition/arcface_mxnet/verification.py b/recognition/arcface_mxnet/verification.py
new file mode 100644
index 000000000..b8f90a643
--- /dev/null
+++ b/recognition/arcface_mxnet/verification.py
@@ -0,0 +1,680 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset 
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import argparse
+import sys
+import numpy as np
+from scipy import misc
+from sklearn.model_selection import KFold
+from scipy import interpolate
+import sklearn
+import cv2
+import math
+import datetime
+import pickle
+from sklearn.decomposition import PCA
+import mxnet as mx
+from mxnet import ndarray as nd
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+    #print('pca', pca)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        #print('train_set', train_set)
+        #print('test_set', test_set)
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            #print(_embed_train.shape)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            #print(embed1.shape, embed2.shape)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        #print('threshold', thresholds[best_threshold_index])
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    #print(true_accept, false_accept)
+    #print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
+
+
+def load_bin(path, image_size):
+    try:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f)  #py2
+    except UnicodeDecodeError as e:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f, encoding='bytes')  #py3
+    data_list = []
+    for flip in [0, 1]:
+        data = nd.empty(
+            (len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for i in range(len(issame_list) * 2):
+        _bin = bins[i]
+        img = mx.image.imdecode(_bin)
+        if img.shape[1] != image_size[0]:
+            img = mx.image.resize_short(img, image_size[0])
+        img = nd.transpose(img, axes=(2, 0, 1))
+        for flip in [0, 1]:
+            if flip == 1:
+                img = mx.ndarray.flip(data=img, axis=2)
+            data_list[flip][i][:] = img
+        if i % 1000 == 0:
+            print('loading bin', i)
+    print(data_list[0].shape)
+    return (data_list, issame_list)
+
+
+def test(data_set,
+         mx_model,
+         batch_size,
+         nfolds=10,
+         data_extra=None,
+         label_shape=None):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    model = mx_model
+    embeddings_list = []
+    if data_extra is not None:
+        _data_extra = nd.array(data_extra)
+    time_consumed = 0.0
+    if label_shape is None:
+        _label = nd.ones((batch_size, ))
+    else:
+        _label = nd.ones(label_shape)
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            #print(_data.shape, _label.shape)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            #_arg, _aux = model.get_params()
+            #__arg = {}
+            #for k,v in _arg.iteritems():
+            #  __arg[k] = v.as_in_context(_ctx)
+            #_arg = __arg
+            #_arg["data"] = _data.as_in_context(_ctx)
+            #_arg["softmax_label"] = _label.as_in_context(_ctx)
+            #for k,v in _arg.iteritems():
+            #  print(k,v.context)
+            #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
+            #exe.forward(is_train=False)
+            #net_out = exe.outputs
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            #print(_embeddings.shape)
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            #print(_em.shape, _norm)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
+    #acc1, std1 = np.mean(accuracy), np.std(accuracy)
+
+    #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
+    #embeddings = np.concatenate(embeddings_list, axis=1)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    print('infer time', time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+def test_badcase(data_set,
+                 mx_model,
+                 batch_size,
+                 name='',
+                 data_extra=None,
+                 label_shape=None):
+    print('testing verification badcase..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    model = mx_model
+    embeddings_list = []
+    if data_extra is not None:
+        _data_extra = nd.array(data_extra)
+    time_consumed = 0.0
+    if label_shape is None:
+        _label = nd.ones((batch_size, ))
+    else:
+        _label = nd.ones(label_shape)
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            #print(_data.shape, _label.shape)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    thresholds = np.arange(0, 4, 0.01)
+    actual_issame = np.asarray(issame_list)
+    nrof_folds = 10
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    data = data_list[0]
+
+    pouts = []
+    nouts = []
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        #print(train_set)
+        #print(train_set.__class__)
+        for threshold_idx, threshold in enumerate(thresholds):
+            p2 = dist[train_set]
+            p3 = actual_issame[train_set]
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, p2, p3)
+        best_threshold_index = np.argmax(acc_train)
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+        best_threshold = thresholds[best_threshold_index]
+        for iid in test_set:
+            ida = iid * 2
+            idb = ida + 1
+            asame = actual_issame[iid]
+            _dist = dist[iid]
+            violate = _dist - best_threshold
+            if not asame:
+                violate *= -1.0
+            if violate > 0.0:
+                imga = data[ida].asnumpy().transpose(
+                    (1, 2, 0))[..., ::-1]  #to bgr
+                imgb = data[idb].asnumpy().transpose((1, 2, 0))[..., ::-1]
+                #print(imga.shape, imgb.shape, violate, asame, _dist)
+                if asame:
+                    pouts.append((imga, imgb, _dist, best_threshold, ida))
+                else:
+                    nouts.append((imga, imgb, _dist, best_threshold, ida))
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    acc = np.mean(accuracy)
+    pouts = sorted(pouts, key=lambda x: x[2], reverse=True)
+    nouts = sorted(nouts, key=lambda x: x[2], reverse=False)
+    print(len(pouts), len(nouts))
+    print('acc', acc)
+    gap = 10
+    image_shape = (112, 224, 3)
+    out_dir = "./badcases"
+    if not os.path.exists(out_dir):
+        os.makedirs(out_dir)
+    if len(nouts) > 0:
+        threshold = nouts[0][3]
+    else:
+        threshold = pouts[-1][3]
+
+    for item in [(pouts, 'positive(false_negative).png'),
+                 (nouts, 'negative(false_positive).png')]:
+        cols = 4
+        rows = 8000
+        outs = item[0]
+        if len(outs) == 0:
+            continue
+        #if len(outs)==9:
+        #  cols = 3
+        #  rows = 3
+
+        _rows = int(math.ceil(len(outs) / cols))
+        rows = min(rows, _rows)
+        hack = {}
+
+        if name.startswith('cfp') and item[1].startswith('pos'):
+            hack = {
+                0: 'manual/238_13.jpg.jpg',
+                6: 'manual/088_14.jpg.jpg',
+                10: 'manual/470_14.jpg.jpg',
+                25: 'manual/238_13.jpg.jpg',
+                28: 'manual/143_11.jpg.jpg'
+            }
+
+        filename = item[1]
+        if len(name) > 0:
+            filename = name + "_" + filename
+        filename = os.path.join(out_dir, filename)
+        img = np.zeros((image_shape[0] * rows + 20, image_shape[1] * cols +
+                        (cols - 1) * gap, 3),
+                       dtype=np.uint8)
+        img[:, :, :] = 255
+        text_color = (0, 0, 153)
+        text_color = (255, 178, 102)
+        text_color = (153, 255, 51)
+        for outi, out in enumerate(outs):
+            row = outi // cols
+            col = outi % cols
+            if row == rows:
+                break
+            imga = out[0].copy()
+            imgb = out[1].copy()
+            if outi in hack:
+                idx = out[4]
+                print('noise idx', idx)
+                aa = hack[outi]
+                imgb = cv2.imread(aa)
+                #if aa==1:
+                #  imgb = cv2.transpose(imgb)
+                #  imgb = cv2.flip(imgb, 1)
+                #elif aa==3:
+                #  imgb = cv2.transpose(imgb)
+                #  imgb = cv2.flip(imgb, 0)
+                #else:
+                #  for ii in range(2):
+                #    imgb = cv2.transpose(imgb)
+                #    imgb = cv2.flip(imgb, 1)
+            dist = out[2]
+            _img = np.concatenate((imga, imgb), axis=1)
+            k = "%.3f" % dist
+            #print(k)
+            font = cv2.FONT_HERSHEY_SIMPLEX
+            cv2.putText(_img, k, (80, image_shape[0] // 2 + 7), font, 0.6,
+                        text_color, 2)
+            #_filename = filename+"_%d.png"%outi
+            #cv2.imwrite(_filename, _img)
+            img[row * image_shape[0]:(row + 1) * image_shape[0],
+                (col * image_shape[1] +
+                 gap * col):((col + 1) * image_shape[1] + gap * col), :] = _img
+        #threshold = outs[0][3]
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        k = "threshold: %.3f" % threshold
+        cv2.putText(img, k, (img.shape[1] // 2 - 70, img.shape[0] - 5), font,
+                    0.6, text_color, 2)
+        cv2.imwrite(filename, img)
+
+
+def dumpR(data_set,
+          mx_model,
+          batch_size,
+          name='',
+          data_extra=None,
+          label_shape=None):
+    print('dump verification embedding..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    model = mx_model
+    embeddings_list = []
+    if data_extra is not None:
+        _data_extra = nd.array(data_extra)
+    time_consumed = 0.0
+    if label_shape is None:
+        _label = nd.ones((batch_size, ))
+    else:
+        _label = nd.ones(label_shape)
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            #print(_data.shape, _label.shape)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    actual_issame = np.asarray(issame_list)
+    outname = os.path.join('temp.bin')
+    with open(outname, 'wb') as f:
+        pickle.dump((embeddings, issame_list),
+                    f,
+                    protocol=pickle.HIGHEST_PROTOCOL)
+
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser(description='do verification')
+    # general
+    parser.add_argument('--data-dir', default='', help='')
+    parser.add_argument('--model',
+                        default='../model/softmax,50',
+                        help='path to load model.')
+    parser.add_argument('--target',
+                        default='lfw,cfp_ff,cfp_fp,agedb_30',
+                        help='test targets.')
+    parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+    parser.add_argument('--batch-size', default=32, type=int, help='')
+    parser.add_argument('--max', default='', type=str, help='')
+    parser.add_argument('--mode', default=0, type=int, help='')
+    parser.add_argument('--nfolds', default=10, type=int, help='')
+    args = parser.parse_args()
+    #sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+    #import face_image
+    #prop = face_image.load_property(args.data_dir)
+    #image_size = prop.image_size
+    image_size = [112, 112]
+    print('image_size', image_size)
+    ctx = mx.gpu(args.gpu)
+    nets = []
+    vec = args.model.split(',')
+    prefix = args.model.split(',')[0]
+    epochs = []
+    if len(vec) == 1:
+        pdir = os.path.dirname(prefix)
+        for fname in os.listdir(pdir):
+            if not fname.endswith('.params'):
+                continue
+            _file = os.path.join(pdir, fname)
+            if _file.startswith(prefix):
+                epoch = int(fname.split('.')[0].split('-')[1])
+                epochs.append(epoch)
+        epochs = sorted(epochs, reverse=True)
+        if len(args.max) > 0:
+            _max = [int(x) for x in args.max.split(',')]
+            assert len(_max) == 2
+            if len(epochs) > _max[1]:
+                epochs = epochs[_max[0]:_max[1]]
+
+    else:
+        epochs = [int(x) for x in vec[1].split('|')]
+    print('model number', len(epochs))
+    time0 = datetime.datetime.now()
+    for epoch in epochs:
+        print('loading', prefix, epoch)
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
+        all_layers = sym.get_internals()
+        sym = all_layers['fc1_output']
+        model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+        #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+        model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+                                          image_size[1]))])
+        model.set_params(arg_params, aux_params)
+        nets.append(model)
+    time_now = datetime.datetime.now()
+    diff = time_now - time0
+    print('model loading time', diff.total_seconds())
+
+    ver_list = []
+    ver_name_list = []
+    for name in args.target.split(','):
+        path = os.path.join(args.data_dir, name + ".bin")
+        if os.path.exists(path):
+            print('loading.. ', name)
+            data_set = load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+
+    if args.mode == 0:
+        for i in range(len(ver_list)):
+            results = []
+            for model in nets:
+                acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+                    ver_list[i], model, args.batch_size, args.nfolds)
+                print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
+                print('[%s]Accuracy: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc1, std1))
+                print('[%s]Accuracy-Flip: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc2, std2))
+                results.append(acc2)
+            print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
+    elif args.mode == 1:
+        model = nets[0]
+        test_badcase(ver_list[0], model, args.batch_size, args.target)
+    else:
+        model = nets[0]
+        dumpR(ver_list[0], model, args.batch_size, args.target)
diff --git a/recognition/arcface_oneflow/README.md b/recognition/arcface_oneflow/README.md
new file mode 100644
index 000000000..21f980c86
--- /dev/null
+++ b/recognition/arcface_oneflow/README.md
@@ -0,0 +1,266 @@
+
+# InsightFace in OneFlow
+
+[English](README.md) **|** [简体中文](README_CH.md)
+
+It introduces how to train InsightFace in OneFlow, and do verification over the validation datasets via the well-toned networks.
+
+## Contents
+
+\- [InsightFace in OneFlow](#insightface-in-oneflow)
+
+ \- [Contents](#contents)
+
+ \- [Background](#background)
+
+  \- [InsightFace opensource project](#insightface-opensource-project)
+
+  \- [Implementation in OneFlow](#implementation-in-oneflow)
+
+ \- [Preparations](#preparations)
+
+  \- [Install OneFlow](#install-oneflow)
+
+  \- [Data preparations](#data-preparations)
+
+   \- [1. Download datasets](#1-download-datasets)
+
+   \- [2. Transformation from MS1M recordio to OFRecord](#2-transformation-from-ms1m-recordio-to-ofrecord)
+
+ \- [Training and verification](#training-and-verification)
+
+  \- [Training](#training)
+
+  \- [OneFLow2ONNX](#OneFLow2ONNX)
+
+## Background
+
+### InsightFace opensource project
+
+[InsightFace](https://github.com/deepinsight/insightface) is an open-source 2D&3D deep face analysis toolbox, mainly based on MXNet.
+
+In InsightFace, it supports:
+
+
+
+- Datasets typically used for face recognition, such as CASIA-Webface、MS1M、VGG2(Provided with the form of a binary file which could run in MXNet, [here](https://github.com/deepinsight/insightface/wiki/Dataset-Zoo) is more details about the datasets and how to download.
+
+
+
+* Backbones of ResNet, MobilefaceNet, InceptionResNet_v2, and other deep-learning networks to apply in facial recognition. 
+
+* Implementation of different loss functions, including SphereFace Loss、Softmax Loss、SphereFace Loss, etc.
+
+  
+
+### Implementation in OneFlow
+
+Based upon the currently existing work of Insightface, OneFlow ported basic models from it, and now OneFlow supports:
+
+
+
+- Training datasets of MS1M、Glint360k, and validation datasets of Lfw、Cfp_fp and Agedb_30, scripts for training and validating.
+
+- Backbones of ResNet100 and MobileFaceNet to recognize faces.
+
+- Loss function, e.g. Softmax Loss and Margin Softmax Loss（including Arcface、Cosface and Combined Loss）.
+
+- Model parallelism and [Partial FC](https://github.com/deepinsight/insightface/tree/760d6de043d7f654c5963391271f215dab461547/recognition/partial_fc#partial-fc) optimization.
+
+- Model transformation via MXNet.
+
+
+
+To be coming further:
+
+- Additional datasets transformation.
+
+- Plentiful backbones.
+
+- Full-scale loss functions implementation.
+
+- Incremental tutorial on the distributed configuration.
+
+
+
+This project is open for every developer to PR, new implementation and animated discussion will be most welcome.
+
+
+
+## Preparations
+
+First of all, before execution, please make sure that:
+
+1. Install OneFlow
+
+2. Prepare training and validation datasets in form of OFRecord.
+
+
+
+### Install OneFlow
+
+
+
+According to steps in [Install OneFlow](https://github.com/Oneflow-Inc/oneflow#install-oneflow) install the newest release master whl packages.
+
+```
+python3 -m pip install oneflow -f https://oneflow-staging.oss-cn-beijing.aliyuncs.com/branch/master/cu102/6aa719d70119b65837b25cc5f186eb19ef2b7891/index.html --user
+```
+
+
+
+### Data preparations
+
+According to [Load and Prepare OFRecord Datasets](https://docs.oneflow.org/en/extended_topics/how_to_make_ofdataset.html), datasets should be converted into the form of OFREcord, to test InsightFace.
+
+
+
+It has provided a set of datasets related to face recognition tasks, which have been pre-processed via face alignment or other processions already in [InsightFace](https://github.com/deepinsight/insightface). The corresponding datasets could be downloaded from [here](https://github.com/deepinsight/insightface/wiki/Dataset-Zoo) and should be converted into OFRecord, which performs better in OneFlow. Considering the cumbersome steps, it is suggested to download converted OFrecord datasets:
+
+[MS1M-ArcFace(face_emore)](http://oneflow-public.oss-cn-beijing.aliyuncs.com/face_dataset/train_ofrecord.tar.gz)
+
+[MS1MV3](https://oneflow-public.oss-cn-beijing.aliyuncs.com/facedata/MS1V3/oneflow/ms1m-retinaface-t1.zip)
+
+It illustrates how to convert downloaded datasets into OFRecords, and take MS1M-ArcFace as an example in the following.
+
+#### 1. Download datasets
+
+The structure of the downloaded MS1M-ArcFace is shown as follown：
+
+
+
+```
+faces_emore/
+
+​    train.idx
+
+​    train.rec
+
+​    property
+
+​    lfw.bin
+
+​    cfp_fp.bin
+
+​    agedb_30.bin
+```
+
+The first three files are MXNet recordio format files of MS1M training dataset, the last three `.bin` files are different validation datasets.
+
+
+
+#### 2. Transformation from MS1M recordio to OFRecord
+Only need to execute 2.1 or 2.2
+2.1 Use Python scripts directly
+
+Run 
+```
+python tools/mx_recordio_2_ofrecord_shuffled_npart.py  --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train --num_part 16
+```
+And you will get the number of `part_num` parts of OFRecord, it's 16 parts in this example, it showed like this
+```
+tree ofrecord/test/
+ofrecord/test/
+|-- _SUCCESS
+|-- part-00000
+|-- part-00001
+|-- part-00002
+|-- part-00003
+|-- part-00004
+|-- part-00005
+|-- part-00006
+|-- part-00007
+|-- part-00008
+|-- part-00009
+|-- part-00010
+|-- part-00011
+|-- part-00012
+|-- part-00013
+|-- part-00014
+`-- part-00015
+
+0 directories, 17 files
+```
+
+
+2.2 Use Python scripts + Spark Shuffle + Spark partition
+
+Run
+
+```
+python tools/dataset_convert/mx_recordio_2_ofrecord.py --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train
+```
+
+And you will get one part of OFRecord(`part-0`) with all data in this way. Then you should use Spark to shuffle and partition.
+1. Get jar package available
+You can download Spark-oneflow-connector-assembly-0.1.0.jar via [Github](https://github.com/Oneflow-Inc/spark-oneflow-connector) or [OSS](https://oneflow-public.oss-cn-beijing.aliyuncs.com/spark-oneflow-connector/spark-oneflow-connector-assembly-0.1.1.jar)
+
+2. Run in Spark
+Assign that you have already installed and configured Spark.
+Run
+```
+//Start Spark 
+./Spark-2.4.3-bin-hadoop2.7/bin/Spark-shell --jars ~/Spark-oneflow-connector-assembly-0.1.0.jar --driver-memory=64G --conf Spark.local.dir=/tmp/
+// shuffle and partition in 16 parts
+import org.oneflow.Spark.functions._
+Spark.read.chunk("data_path").shuffle().repartition(16).write.chunk("new_data_path")
+sc.formatFilenameAsOneflowStyle("new_data_path")
+```
+Hence you will get 16 parts of OFRecords, it shown like this
+```
+tree ofrecord/test/
+ofrecord/test/
+|-- _SUCCESS
+|-- part-00000
+|-- part-00001
+|-- part-00002
+|-- part-00003
+|-- part-00004
+|-- part-00005
+|-- part-00006
+|-- part-00007
+|-- part-00008
+|-- part-00009
+|-- part-00010
+|-- part-00011
+|-- part-00012
+|-- part-00013
+|-- part-00014
+`-- part-00015
+
+0 directories, 17 files
+```
+
+
+## Training and verification
+
+
+
+### Training
+
+To reduce the usage cost of user, OneFlow draws close the scripts to Torch style, you can directly modify parameters via configs/*.py
+
+#### eager 
+```
+./train_ddp.sh
+```
+#### Graph
+```
+train_graph_distributed.sh
+```
+
+
+### Varification
+
+Moreover, OneFlow offers a validation script to do verification separately, val.py, which facilitates you to check the precision of the pre-training model saved.
+
+```
+./val.sh
+
+```
+## OneFLow2ONNX
+
+```
+pip install oneflow-onnx==0.5.1
+./convert.sh
+```
\ No newline at end of file
diff --git a/recognition/arcface_oneflow/README_CH.md b/recognition/arcface_oneflow/README_CH.md
new file mode 100644
index 000000000..5a62f825d
--- /dev/null
+++ b/recognition/arcface_oneflow/README_CH.md
@@ -0,0 +1,231 @@
+# InsightFace 在 OneFlow 中的实现
+
+
+[English](README.md) **|** [简体中文](README_CH.md)
+
+本文介绍如何在 OneFlow 中训练 InsightFace，并在验证数据集上对训练好的网络进行验证。
+
+## 目录
+- [InsightFace 在 OneFlow 中的实现](#insightface-在-oneflow-中的实现)
+  - [目录](#目录)
+  - [背景介绍](#背景介绍)
+    - [InsightFace 开源项目](#insightface-开源项目)
+    - [InsightFace 在 OneFlow 中的实现](#insightface-在-oneflow-中的实现-1)
+  - [准备工作](#准备工作)
+    - [安装 OneFlow](#安装-oneflow)
+    - [准备数据集](#准备数据集)
+      - [1. 下载数据集](#1-下载数据集)
+      - [2. 将训练数据集 MS1M 从 recordio 格式转换为 OFRecord 格式](#2-将训练数据集-ms1m-从-recordio-格式转换为-ofrecord-格式)
+    
+  - [训练和验证](#训练和验证)
+    - [训练](#训练)
+    - [验证](#验证)
+    - [OneFLow2ONNX](#OneFLow2ONNX)
+    
+
+## 背景介绍
+
+###  InsightFace 开源项目
+
+[InsightFace 原仓库](https://github.com/deepinsight/insightface)是基于 MXNet 实现的人脸识别研究开源项目。
+
+在该项目中，集成了：
+
+* CASIA-Webface、MS1M、VGG2 等用于人脸识别研究常用的数据集（以 MXNet 支持的二进制形式提供，可以从[这里](https://github.com/deepinsight/insightface/wiki/Dataset-Zoo)查看数据集的详细说明以及下载链接）。
+
+* 以 ResNet、MobileFaceNet、InceptionResNet_v2 等深度学习网络作为 Backbone 的人脸识别模型。
+* 涵盖 SphereFace Loss、Softmax Loss、SphereFace Loss 等多种损失函数的实现。
+
+
+
+### InsightFace 在 OneFlow 中的实现
+
+在 InsightFace 开源项目已有的工作基础上，OneFlow 对 InsightFace 基本的人脸识别模型进行了移植，目前已实现的功能包括：
+
+* 支持了使用 MS1M、Glint360k 作为训练数据集，Lfw、Cfp_fp 以及 Agedb_30 作为验证数据集，提供了对网络进行训练和验证的脚本。
+* 支持 ResNet100 和 MobileFaceNet 作为人脸识别模型的 Backbone 网络。
+* 实现了 Softmax Loss 以及 Margin Softmax Loss（包括 Nsoftmax、Arcface、Cosface 和 Combined Loss 等）。
+* 实现了模型并行和 Partial FC 优化。
+* 实现了 MXNet 的模型转换。
+
+
+未来将计划逐步完善：
+
+* 更多的数据集转换。
+* 更丰富的 Backbone 网络。
+* 更全面的损失函数实现。
+* 增加分布式运行的说明。
+
+
+
+我们对所有的开发者开放 PR，非常欢迎您加入新的实现以及参与讨论。
+
+## 准备工作
+
+在开始运行前，请先确定：
+
+1. 安装 OneFlow。
+2. 准备训练和验证的 OFRecord 数据集。
+
+
+
+###  安装 OneFlow
+
+根据 [Install OneFlow](https://github.com/Oneflow-Inc/oneflow#install-oneflow) 的步骤进行安装最新 master whl 包即可。
+
+```
+python3 -m pip install oneflow -f https://oneflow-staging.oss-cn-beijing.aliyuncs.com/branch/master/cu102/6aa719d70119b65837b25cc5f186eb19ef2b7891/index.html --user
+```
+
+### 准备数据集
+
+根据 [加载与准备 OFRecord 数据集](https://docs.oneflow.org/extended_topics/how_to_make_ofdataset.html) 准备 ImageNet 的 OFReocord 数据集，用以进行 InsightFace 的测试。
+
+[InsightFace 原仓库](https://github.com/deepinsight/insightface)中提供了一系列人脸识别任务相关的数据集，已经完成了人脸对齐等预处理过程。请从[这里](https://github.com/deepinsight/insightface/wiki/Dataset-Zoo)下载相应的数据集，并且转换成 OneFlow 可以识别的 OFRecord 格式。考虑到步骤繁琐，也可以直接下载已经转好的 OFRecord 数据集：
+
+[ MS1M-ArcFace(face_emore)](http://oneflow-public.oss-cn-beijing.aliyuncs.com/face_dataset/train_ofrecord.tar.gz)
+
+[MS1MV3](https://oneflow-public.oss-cn-beijing.aliyuncs.com/facedata/MS1V3/oneflow/ms1m-retinaface-t1.zip)
+
+下面以数据集 MS1M-ArcFace 为例，展示如何将下载到的数据集转换成 OFRecord 格式。
+
+#### 1. 下载数据集
+
+下载好的 MS1M-ArcFace 数据集，内容如下：
+
+```
+faces_emore/
+       train.idx
+       train.rec
+       property
+       lfw.bin
+       cfp_fp.bin
+       agedb_30.bin
+```
+
+
+
+前三个文件是训练数据集 MS1M 的 MXNet 的 recordio 格式相关的文件，后三个 `.bin` 文件是三个不同的验证数据集。
+
+
+
+#### 2. 将训练数据集 MS1M 从 recordio 格式转换为 OFRecord 格式
+训练数据集转换有两种方式: (2.1部分)直接使用python脚本生成n个shuffle过的数据part，或(2.2部分)python脚本生成一个part，再根据需要用spark做shuffle和partition。
+2.1 直接使用 Python 脚本
+
+运行： 
+```
+python tools/dataset_convert/mx_recordio_2_ofrecord_shuffled_npart.py  --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train --num_part 16
+```
+成功后将得到 `num_part` 数量个 OFRecord，本示例中为 16 个，显示如下：
+
+```
+tree ofrecord/test/
+ofrecord/test/
+|-- _SUCCESS
+|-- part-00000
+|-- part-00001
+|-- part-00002
+|-- part-00003
+|-- part-00004
+|-- part-00005
+|-- part-00006
+|-- part-00007
+|-- part-00008
+|-- part-00009
+|-- part-00010
+|-- part-00011
+|-- part-00012
+|-- part-00013
+|-- part-00014
+`-- part-00015
+
+0 directories, 17 files
+```
+2.2 Python 脚本 + Spark Shuffle + Spark Partition
+
+运行：
+
+```
+python tools/dataset_convert/mx_recordio_2_ofrecord.py --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train
+```
+成功后将得到一个包含所有数据的 OFReocrd（`part-0`），需要进一步使用 Spark 进行 Shuffle 和 Partition。
+成功安装和部署 Spark 后， 您需要：
+1. 下载工具 jar 包
+   
+
+您可以通过 [Github](https://github.com/Oneflow-Inc/spark-oneflow-connector) 或者 [OSS](https://oneflow-public.oss-cn-beijing.aliyuncs.com/spark-oneflow-connector/spark-oneflow-connector-assembly-0.1.1.jar) 下载 Spark-oneflow-connector-assembly-0.1.0.jar 文件。
+1. 运行 Spark 命令
+
+
+运行
+```
+//Start Spark 
+./Spark-2.4.3-bin-hadoop2.7/bin/Spark-shell --jars ~/Spark-oneflow-connector-assembly-0.1.0.jar --driver-memory=64G --conf Spark.local.dir=/tmp/
+// shuffle and partition in 16 parts
+import org.oneflow.Spark.functions._
+Spark.read.chunk("data_path").shuffle().repartition(16).write.chunk("new_data_path")
+sc.formatFilenameAsOneflowStyle("new_data_path")
+```
+然后就可以得到 16 个 part 的 OFRecord，显示如下
+```
+tree ofrecord/test/
+ofrecord/test/
+|-- _SUCCESS
+|-- part-00000
+|-- part-00001
+|-- part-00002
+|-- part-00003
+|-- part-00004
+|-- part-00005
+|-- part-00006
+|-- part-00007
+|-- part-00008
+|-- part-00009
+|-- part-00010
+|-- part-00011
+|-- part-00012
+|-- part-00013
+|-- part-00014
+`-- part-00015
+
+0 directories, 17 files
+```
+
+
+
+
+## 训练和验证
+
+### 训练
+
+为了减小用户使用的迁移成本，OneFlow 的脚本已经调整为 Torch 实现的风格，用户可以使用 configs/*.py 直接修改参数。
+
+
+运行脚本：
+
+#### eager 
+```
+./train_ddp.sh
+```
+#### Graph
+```
+train_graph_distributed.sh
+```
+
+### 验证
+
+另外，为了方便查看保存下来的预训练模型精度，我们提供了一个仅在验证数据集上单独执行验证过程的脚本。
+
+运行
+
+```
+./val.sh
+```
+
+## OneFLow2ONNX
+
+```
+pip install oneflow-onnx==0.5.1
+./convert.sh
+```
\ No newline at end of file
diff --git a/recognition/arcface_oneflow/backbones/__init__.py b/recognition/arcface_oneflow/backbones/__init__.py
new file mode 100644
index 000000000..2448255b4
--- /dev/null
+++ b/recognition/arcface_oneflow/backbones/__init__.py
@@ -0,0 +1,16 @@
+from .ir_resnet import iresnet18, iresnet34, iresnet50, iresnet100, iresnet200
+
+
+def get_model(name, **kwargs):
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    else:
+        raise ValueError()
diff --git a/recognition/arcface_oneflow/backbones/ir_resnet.py b/recognition/arcface_oneflow/backbones/ir_resnet.py
new file mode 100644
index 000000000..1e9a99571
--- /dev/null
+++ b/recognition/arcface_oneflow/backbones/ir_resnet.py
@@ -0,0 +1,219 @@
+import oneflow as flow
+import oneflow.nn as nn
+from typing import Type, Any, Callable, Union, List, Optional
+
+
+def conv3x3(
+    in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1
+) -> nn.Conv2d:
+    """3x3 convolution with padding"""
+    return nn.Conv2d(
+        in_planes,
+        out_planes,
+        kernel_size=3,
+        stride=stride,
+        padding=dilation,
+        groups=groups,
+        bias=False,
+        dilation=dilation,
+    )
+
+
+def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        stride=1,
+        downsample=None,
+        groups=1,
+        base_width=64,
+        dilation=1,
+    ):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError("BasicBlock only supports groups=1 and base_width=64")
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.prelu = nn.ReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+
+    def __init__(
+        self,
+        block,
+        layers,
+        dropout=0,
+        num_features=512,
+        zero_init_residual=False,
+        groups=1,
+        width_per_group=64,
+        replace_stride_with_dilation=None,
+        fp16=False,
+    ):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                "or a 3-element tuple, got {}".format(replace_stride_with_dilation)
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(
+            3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False
+        )
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.ReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(
+            block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0]
+        )
+        self.layer3 = self._make_layer(
+            block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1]
+        )
+        self.layer4 = self._make_layer(
+            block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2]
+        )
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05,),
+            )
+        layers = []
+        layers.append(
+            block(
+                self.inplanes,
+                planes,
+                stride,
+                downsample,
+                self.groups,
+                self.base_width,
+                previous_dilation,
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.prelu(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.bn2(x)
+        x = flow.flatten(x, 1)
+        x = self.dropout(x)
+        x = self.fc(x)
+        x = self.features(x)
+
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet(
+        "iresnet18", IBasicBlock, [2, 2, 2, 2], pretrained, progress, **kwargs
+    )
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet(
+        "iresnet34", IBasicBlock, [3, 4, 6, 3], pretrained, progress, **kwargs
+    )
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet(
+        "iresnet50", IBasicBlock, [3, 4, 14, 3], pretrained, progress, **kwargs
+    )
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet(
+        "iresnet100", IBasicBlock, [3, 13, 30, 3], pretrained, progress, **kwargs
+    )
+
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet(
+        "iresnet200", IBasicBlock, [6, 26, 60, 6], pretrained, progress, **kwargs
+    )
diff --git a/recognition/arcface_oneflow/configs/__init__.py b/recognition/arcface_oneflow/configs/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_oneflow/configs/base.py b/recognition/arcface_oneflow/configs/base.py
new file mode 100644
index 000000000..e4272484d
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/base.py
@@ -0,0 +1,68 @@
+from pickle import TRUE
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "arcface"
+config.network = "r50"
+config.resume = False
+config.output = "ms1mv3_arcface_r50"
+
+config.dataset = "ms1m-retinaface-t1"
+config.embedding_size = 512
+config.fp16 = False
+config.model_parallel = False
+config.sample_rate = 1.0
+config.partial_fc = False
+config.graph = True
+config.synthetic = False
+config.scale_grad = False
+
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+config.val_image_num = {"lfw": 12000, "cfp_fp": 14000, "agedb_30": 12000}
+if config.dataset == "emore":
+    config.ofrecord_path = "/train_tmp/faces_emore"
+    config.num_classes = 85742
+    config.num_image = 5822653
+    config.num_epoch = 16
+    config.warmup_epoch = -1
+    config.decay_epoch = [
+        8,
+        14,
+    ]
+    config.val_targets = [
+        "lfw",
+    ]
+
+elif config.dataset == "ms1m-retinaface-t1":
+    config.ofrecord_path = "/dev/shm/ms1m-retinaface-t1/ofrecord"
+    config.num_classes = 93431
+    config.num_image = 5179510
+    config.num_epoch = 25
+    config.warmup_epoch = -1
+    config.decay_epoch = [11, 17, 22]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+elif config.dataset == "glint360k":
+    config.ofrecord_path = "/train_tmp/glint360k"
+    config.num_classes = 360232
+    config.num_image = 17091657
+    config.num_epoch = 20
+    config.warmup_epoch = -1
+    config.decay_epoch = [8, 12, 15, 18]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+elif config.dataset == "webface":
+    config.ofrecord_path = "/train_tmp/faces_webface_112x112"
+    config.num_classes = 10572
+    config.num_image = "forget"
+    config.num_epoch = 34
+    config.warmup_epoch = -1
+    config.decay_epoch = [20, 28, 32]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/glint360k_mbf.py b/recognition/arcface_oneflow/configs/glint360k_mbf.py
new file mode 100644
index 000000000..96697a13b
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/glint360k_mbf.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.model_parallel = True
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 2e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.dataset = "glint360k"
+config.ofrecord_path = "/train_tmp/glint360k/"
+config.ofrecord_part_num = 200
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = -1
+config.decay_epoch = [8, 12, 15, 18]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/glint360k_r100.py b/recognition/arcface_oneflow/configs/glint360k_r100.py
new file mode 100644
index 000000000..7a415180c
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/glint360k_r100.py
@@ -0,0 +1,31 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.model_parallel = True
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+
+config.dataset = "glint360k"
+config.ofrecord_path = "/train_tmp/glint360k/"
+config.ofrecord_part_num = 200
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = -1
+config.decay_epoch = [8, 12, 15, 18]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/glint360k_r18.py b/recognition/arcface_oneflow/configs/glint360k_r18.py
new file mode 100644
index 000000000..031fd02c0
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/glint360k_r18.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "r18"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.model_parallel = True
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.dataset = "glint360k"
+config.ofrecord_path = "/train_tmp/glint360k/"
+config.ofrecord_part_num = 200
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = -1
+config.decay_epoch = [8, 12, 15, 18]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/glint360k_r34.py b/recognition/arcface_oneflow/configs/glint360k_r34.py
new file mode 100644
index 000000000..b072b02ec
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/glint360k_r34.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "r34"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.model_parallel = True
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.dataset = "glint360k"
+config.ofrecord_path = "/train_tmp/glint360k/"
+config.ofrecord_part_num = 200
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = -1
+config.decay_epoch = [8, 12, 15, 18]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/glint360k_r50.py b/recognition/arcface_oneflow/configs/glint360k_r50.py
new file mode 100644
index 000000000..90add4875
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/glint360k_r50.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.model_parallel = True
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.dataset = "glint360k"
+config.ofrecord_path = "/train_tmp/glint360k/"
+config.ofrecord_part_num = 200
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = -1
+config.decay_epoch = [8, 12, 15, 18]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/ms1mv3_mbf.py b/recognition/arcface_oneflow/configs/ms1mv3_mbf.py
new file mode 100644
index 000000000..917fb1370
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/ms1mv3_mbf.py
@@ -0,0 +1,29 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "arcface"
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.model_parallel = True
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 2e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.ofrecord_path = "/train_tmp/ms1m-retinaface-t1"
+config.ofrecord_part_num = 8
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 30
+config.warmup_epoch = -1
+config.decay_epoch = [10, 20, 25]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/ms1mv3_r18.py b/recognition/arcface_oneflow/configs/ms1mv3_r18.py
new file mode 100644
index 000000000..7b2168ad1
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/ms1mv3_r18.py
@@ -0,0 +1,29 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "arcface"
+config.network = "r18"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.model_parallel = True
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.ofrecord_path = "/train_tmp/ms1m-retinaface-t1"
+config.ofrecord_part_num = 8
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 25
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/ms1mv3_r34.py b/recognition/arcface_oneflow/configs/ms1mv3_r34.py
new file mode 100644
index 000000000..c95aef0d0
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/ms1mv3_r34.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "arcface"
+config.network = "r34"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.model_parallel = True
+config.partial_fc = 1
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+
+config.ofrecord_path = "/train_tmp/ms1m-retinaface-t1"
+config.ofrecord_part_num = 8
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 25
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/ms1mv3_r50.py b/recognition/arcface_oneflow/configs/ms1mv3_r50.py
new file mode 100644
index 000000000..c7dede32f
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/ms1mv3_r50.py
@@ -0,0 +1,30 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "cosface"
+config.network = "r50"
+config.resume = False
+config.output = "partial_fc"
+config.embedding_size = 512
+config.model_parallel = True
+config.partial_fc = 0
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+
+config.ofrecord_path = "/train_tmp/ms1m-retinaface-t1/ofrecord/"
+config.ofrecord_part_num = 8
+config.num_classes = 93432
+config.num_image = 5179510
+config.num_epoch = 25
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_oneflow/configs/speed.py b/recognition/arcface_oneflow/configs/speed.py
new file mode 100644
index 000000000..cae0b8be7
--- /dev/null
+++ b/recognition/arcface_oneflow/configs/speed.py
@@ -0,0 +1,24 @@
+from easydict import EasyDict as edict
+
+# configs for test speed
+
+config = edict()
+config.loss = "arcface"
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.model_parallel = True
+config.sample_rate = 1.0
+config.fp16 = False
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.synthetic = True
+config.num_classes = 100000
+config.num_epoch = 30
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = []
diff --git a/recognition/arcface_oneflow/convert.sh b/recognition/arcface_oneflow/convert.sh
new file mode 100644
index 000000000..1ad7d8abd
--- /dev/null
+++ b/recognition/arcface_oneflow/convert.sh
@@ -0,0 +1,2 @@
+
+python3 oneflow2onnx.py configs/ms1mv3_r50 --model_path /workdir/epoch_0
diff --git a/recognition/arcface_oneflow/eval/__init__.py b/recognition/arcface_oneflow/eval/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_oneflow/eval/onnx_helper.py b/recognition/arcface_oneflow/eval/onnx_helper.py
new file mode 100644
index 000000000..b58cd5353
--- /dev/null
+++ b/recognition/arcface_oneflow/eval/onnx_helper.py
@@ -0,0 +1,227 @@
+import argparse
+import datetime
+import os
+import os.path as osp
+
+import cv2
+import numpy as np
+import onnx
+import onnxruntime
+from onnx import numpy_helper
+
+
+class ArcFaceORT:
+    def __init__(self, model_path):
+        self.model_path = model_path
+
+    def check(self, test_img=None):
+        max_model_size_mb = 1024
+        max_feat_dim = 512
+        max_time_cost = 15
+
+        if not os.path.exists(self.model_path):
+            return "model_path not exists"
+        if not os.path.isdir(self.model_path):
+            return "model_path should be directory"
+        onnx_files = []
+        for _file in os.listdir(self.model_path):
+            print("file_:", _file)
+            if _file.endswith(".onnx"):
+                onnx_files.append(osp.join(self.model_path, _file))
+        if len(onnx_files) == 0:
+            return "do not have onnx files"
+        self.model_file = sorted(onnx_files)[-1]
+        print("use onnx-model:", self.model_file)
+        try:
+            session = onnxruntime.InferenceSession(self.model_file, None)
+        except:
+            return "load onnx failed"
+
+        input_cfg = session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        print("input-shape:", input_shape)
+        if len(input_shape) != 4:
+            return "length of input_shape should be 4"
+        if not isinstance(input_shape[0], str):
+            # return "input_shape[0] should be str to support batch-inference"
+            print("reset input-shape[0] to None")
+            model = onnx.load(self.model_file)
+            model.graph.input[0].type.tensor_type.shape.dim[0].dim_param = "None"
+            new_model_file = osp.join(self.model_path, "zzzzrefined.onnx")
+            onnx.save(model, new_model_file)
+            self.model_file = new_model_file
+            print("use new onnx-model:", self.model_file)
+            try:
+                session = onnxruntime.InferenceSession(self.model_file, None)
+            except:
+                return "load onnx failed"
+
+            input_cfg = session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            print("new-input-shape:", input_shape)
+
+        self.image_size = tuple(input_shape[2:4][::-1])
+
+        input_name = input_cfg.name
+        outputs = session.get_outputs()
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+            # print(o.name, o.shape)
+        if len(output_names) != 1:
+            return "number of output nodes should be 1"
+        self.session = session
+        self.input_name = input_name
+        self.output_names = output_names
+
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        if len(graph.node) < 8:
+            return "too small onnx graph"
+
+        input_size = (112, 112)
+        self.crop = None
+        if True:
+            crop_file = osp.join(self.model_path, "crop.txt")
+            if osp.exists(crop_file):
+                lines = open(crop_file, "r").readlines()
+                if len(lines) != 6:
+                    return "crop.txt should contain 6 lines"
+                lines = [int(x) for x in lines]
+                self.crop = lines[:4]
+                input_size = tuple(lines[4:6])
+        if input_size != self.image_size:
+            return "input-size is inconsistant with onnx model input, %s vs %s" % (
+                input_size,
+                self.image_size,
+            )
+
+        self.model_size_mb = os.path.getsize(self.model_file) / float(1024 * 1024)
+        if self.model_size_mb > max_model_size_mb:
+            return "max model size exceed, given %.3f-MB" % self.model_size_mb
+
+        input_mean = None
+        input_std = None
+        if True:
+            pn_file = osp.join(self.model_path, "pixel_norm.txt")
+            if osp.exists(pn_file):
+                lines = open(pn_file, "r").readlines()
+                if len(lines) != 2:
+                    return "pixel_norm.txt should contain 2 lines"
+                input_mean = float(lines[0])
+                input_std = float(lines[1])
+        if input_mean is not None or input_std is not None:
+            if input_mean is None or input_std is None:
+                return "please set input_mean and input_std simultaneously"
+        else:
+            find_sub = False
+            find_mul = False
+            for nid, node in enumerate(graph.node[:8]):
+                print(nid, node.name)
+                if node.name.startswith("Sub") or node.name.startswith("_minus"):
+                    find_sub = True
+                if node.name.startswith("Mul") or node.name.startswith("_mul"):
+                    find_mul = True
+            if find_sub and find_mul:
+                # mxnet arcface model
+                input_mean = 0.0
+                input_std = 1.0
+            else:
+                input_mean = 127.5
+                input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        for initn in graph.initializer:
+            weight_array = numpy_helper.to_array(initn)
+
+            dt = weight_array.dtype
+            if dt.itemsize < 4:
+                return "invalid weight type - (%s:%s)" % (initn.name, dt.name)
+        if test_img is None:
+            test_img = np.random.randint(
+                0, 255, size=(self.image_size[1], self.image_size[0], 3), dtype=np.uint8
+            )
+        else:
+            test_img = cv2.resize(test_img, self.image_size)
+        feat, cost = self.benchmark(test_img)
+        if feat.shape[1] > max_feat_dim:
+            return "max feat dim exceed, given %d" % feat.shape[1]
+        self.feat_dim = feat.shape[1]
+        cost_ms = cost * 1000
+        if cost_ms > max_time_cost:
+            return "max time cost exceed, given %.4f" % cost_ms
+        self.cost_ms = cost_ms
+        print(
+            "check stat:, model-size-mb: %.4f, feat-dim: %d, time-cost-ms: %.4f, input-mean: %.3f, input-std: %.3f"
+            % (
+                self.model_size_mb,
+                self.feat_dim,
+                self.cost_ms,
+                self.input_mean,
+                self.input_std,
+            )
+        )
+        return None
+
+    def meta_info(self):
+        return {
+            "model-size-mb": self.model_size_mb,
+            "feature-dim": self.feat_dim,
+            "infer": self.cost_ms,
+        }
+
+    def forward(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.image_size
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1] : self.crop[3], self.crop[0] : self.crop[2], :]
+                if nimg.shape[0] != input_size[1] or nimg.shape[1] != input_size[0]:
+                    nimg = cv2.resize(nimg, input_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(
+            imgs,
+            1.0 / self.input_std,
+            input_size,
+            (self.input_mean, self.input_mean, self.input_mean),
+            swapRB=True,
+        )
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+    def benchmark(self, img):
+        input_size = self.image_size
+        if self.crop is not None:
+            nimg = img[self.crop[1] : self.crop[3], self.crop[0] : self.crop[2], :]
+            if nimg.shape[0] != input_size[1] or nimg.shape[1] != input_size[0]:
+                nimg = cv2.resize(nimg, input_size)
+            img = nimg
+        blob = cv2.dnn.blobFromImage(
+            img,
+            1.0 / self.input_std,
+            input_size,
+            (self.input_mean, self.input_mean, self.input_mean),
+            swapRB=True,
+        )
+        costs = []
+        for _ in range(50):
+            ta = datetime.datetime.now()
+            net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+            tb = datetime.datetime.now()
+            cost = (tb - ta).total_seconds()
+            costs.append(cost)
+        costs = sorted(costs)
+        cost = costs[5]
+        return net_out, cost
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_root", help="onnx model root, default is './'", default="./"
+    )
+    args = parser.parse_args()
+    ArcFaceORT(args.model_root).check()
diff --git a/recognition/arcface_oneflow/eval/onnx_ijbc.py b/recognition/arcface_oneflow/eval/onnx_ijbc.py
new file mode 100644
index 000000000..0d07d8185
--- /dev/null
+++ b/recognition/arcface_oneflow/eval/onnx_ijbc.py
@@ -0,0 +1,289 @@
+import argparse
+import os
+import pickle
+import timeit
+
+import cv2
+import mxnet as mx
+import numpy as np
+import pandas as pd
+import prettytable
+import skimage.transform
+from sklearn.metrics import roc_curve
+from sklearn.preprocessing import normalize
+
+from onnx_helper import ArcFaceORT
+
+SRC = np.array(
+    [
+        [30.2946, 51.6963],
+        [65.5318, 51.5014],
+        [48.0252, 71.7366],
+        [33.5493, 92.3655],
+        [62.7299, 92.2041],
+    ],
+    dtype=np.float32,
+)
+SRC[:, 0] += 8.0
+
+
+class AlignedDataSet(mx.gluon.data.Dataset):
+    def __init__(self, root, lines, align=True):
+        self.lines = lines
+        self.root = root
+        self.align = align
+
+    def __len__(self):
+        return len(self.lines)
+
+    def __getitem__(self, idx):
+        each_line = self.lines[idx]
+        name_lmk_score = each_line.strip().split(" ")
+        name = os.path.join(self.root, name_lmk_score[0])
+        img = cv2.cvtColor(cv2.imread(name), cv2.COLOR_BGR2RGB)
+        landmark5 = np.array(
+            [float(x) for x in name_lmk_score[1:-1]], dtype=np.float32
+        ).reshape((5, 2))
+        st = skimage.transform.SimilarityTransform()
+        st.estimate(landmark5, SRC)
+        img = cv2.warpAffine(img, st.params[0:2, :], (112, 112), borderValue=0.0)
+        img_1 = np.expand_dims(img, 0)
+        img_2 = np.expand_dims(np.fliplr(img), 0)
+        output = np.concatenate((img_1, img_2), axis=0).astype(np.float32)
+        output = np.transpose(output, (0, 3, 1, 2))
+        output = mx.nd.array(output)
+        return output
+
+
+def extract(model_root, dataset):
+    model = ArcFaceORT(model_path=model_root)
+    model.check()
+    feat_mat = np.zeros(shape=(len(dataset), 2 * model.feat_dim))
+
+    def batchify_fn(data):
+        return mx.nd.concat(*data, dim=0)
+
+    data_loader = mx.gluon.data.DataLoader(
+        dataset,
+        128,
+        last_batch="keep",
+        num_workers=4,
+        thread_pool=True,
+        prefetch=16,
+        batchify_fn=batchify_fn,
+    )
+    num_iter = 0
+    for batch in data_loader:
+        batch = batch.asnumpy()
+        batch = (batch - model.input_mean) / model.input_std
+        feat = model.session.run(model.output_names, {model.input_name: batch})[0]
+        feat = np.reshape(feat, (-1, model.feat_dim * 2))
+        feat_mat[128 * num_iter : 128 * num_iter + feat.shape[0], :] = feat
+        num_iter += 1
+        if num_iter % 50 == 0:
+            print(num_iter)
+    return feat_mat
+
+
+def read_template_media_list(path):
+    ijb_meta = pd.read_csv(path, sep=" ", header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=" ", header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+def read_image_feature(path):
+    with open(path, "rb") as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+def image2template_feature(img_feats=None, templates=None, medias=None):
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], axis=0, keepdims=True),
+                ]
+        media_norm_feats = np.array(media_norm_feats)
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print("Finish Calculating {} template features.".format(count_template))
+    template_norm_feats = normalize(template_feats)
+    return template_norm_feats, unique_templates
+
+
+def verification(template_norm_feats=None, unique_templates=None, p1=None, p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000
+    sublists = [total_pairs[i : i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print("Finish {}/{} pairs.".format(c, total_sublists))
+    return score
+
+
+def verification2(template_norm_feats=None, unique_templates=None, p1=None, p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    # small batchsize instead of all pairs in one batch due to the memory limiation
+    batchsize = 100000
+    sublists = [total_pairs[i : i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print("Finish {}/{} pairs.".format(c, total_sublists))
+    return score
+
+
+def main(args):
+    use_norm_score = True  # if Ture, TestMode(N1)
+    use_detector_score = True  # if Ture, TestMode(D1)
+    use_flip_test = True  # if Ture, TestMode(F1)
+    assert args.target == "IJBC" or args.target == "IJBB"
+
+    start = timeit.default_timer()
+    templates, medias = read_template_media_list(
+        os.path.join(
+            "%s/meta" % args.image_path, "%s_face_tid_mid.txt" % args.target.lower()
+        )
+    )
+    stop = timeit.default_timer()
+    print("Time: %.2f s. " % (stop - start))
+
+    start = timeit.default_timer()
+    p1, p2, label = read_template_pair_list(
+        os.path.join(
+            "%s/meta" % args.image_path,
+            "%s_template_pair_label.txt" % args.target.lower(),
+        )
+    )
+    stop = timeit.default_timer()
+    print("Time: %.2f s. " % (stop - start))
+
+    start = timeit.default_timer()
+    img_path = "%s/loose_crop" % args.image_path
+    img_list_path = "%s/meta/%s_name_5pts_score.txt" % (
+        args.image_path,
+        args.target.lower(),
+    )
+    img_list = open(img_list_path)
+    files = img_list.readlines()
+    dataset = AlignedDataSet(root=img_path, lines=files, align=True)
+    img_feats = extract(args.model_root, dataset)
+
+    faceness_scores = []
+    for each_line in files:
+        name_lmk_score = each_line.split()
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    stop = timeit.default_timer()
+    print("Time: %.2f s. " % (stop - start))
+    print("Feature Shape: ({} , {}) .".format(img_feats.shape[0], img_feats.shape[1]))
+    start = timeit.default_timer()
+
+    if use_flip_test:
+        img_input_feats = (
+            img_feats[:, 0 : img_feats.shape[1] // 2]
+            + img_feats[:, img_feats.shape[1] // 2 :]
+        )
+    else:
+        img_input_feats = img_feats[:, 0 : img_feats.shape[1] // 2]
+
+    if use_norm_score:
+        img_input_feats = img_input_feats
+    else:
+        img_input_feats = img_input_feats / np.sqrt(
+            np.sum(img_input_feats ** 2, -1, keepdims=True)
+        )
+
+    if use_detector_score:
+        print(img_input_feats.shape, faceness_scores.shape)
+        img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+    else:
+        img_input_feats = img_input_feats
+
+    template_norm_feats, unique_templates = image2template_feature(
+        img_input_feats, templates, medias
+    )
+    stop = timeit.default_timer()
+    print("Time: %.2f s. " % (stop - start))
+
+    start = timeit.default_timer()
+    score = verification(template_norm_feats, unique_templates, p1, p2)
+    stop = timeit.default_timer()
+    print("Time: %.2f s. " % (stop - start))
+    save_path = os.path.join(args.result_dir, "{}_result".format(args.target))
+    if not os.path.exists(save_path):
+        os.makedirs(save_path)
+    score_save_file = os.path.join(save_path, "{}.npy".format(args.model_root))
+    np.save(score_save_file, score)
+    files = [score_save_file]
+    methods = []
+    scores = []
+    for file in files:
+        methods.append(os.path.basename(file))
+        scores.append(np.load(file))
+    methods = np.array(methods)
+    scores = dict(zip(methods, scores))
+    x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+    tpr_fpr_table = prettytable.PrettyTable(["Methods"] + [str(x) for x in x_labels])
+    for method in methods:
+        fpr, tpr, _ = roc_curve(label, scores[method])
+        fpr = np.flipud(fpr)
+        tpr = np.flipud(tpr)
+        tpr_fpr_row = []
+        tpr_fpr_row.append("%s-%s" % (method, args.target))
+        for fpr_iter in np.arange(len(x_labels)):
+            _, min_index = min(
+                list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr))))
+            )
+            tpr_fpr_row.append("%.2f" % (tpr[min_index] * 100))
+        tpr_fpr_table.add_row(tpr_fpr_row)
+    print(tpr_fpr_table)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="do ijb test")
+    # general
+    parser.add_argument("--model-root", default="", help="path to load model.")
+    parser.add_argument("--image-path", default="", type=str, help="")
+    parser.add_argument("--result-dir", default=".", type=str, help="")
+    parser.add_argument(
+        "--target", default="IJBC", type=str, help="target, set to IJBC or IJBB"
+    )
+    main(parser.parse_args())
diff --git a/recognition/arcface_oneflow/eval/verification.py b/recognition/arcface_oneflow/eval/verification.py
new file mode 100644
index 000000000..8fc5b61f3
--- /dev/null
+++ b/recognition/arcface_oneflow/eval/verification.py
@@ -0,0 +1,327 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+import datetime
+import os
+import pickle
+
+
+import numpy as np
+import sklearn
+import oneflow as flow
+
+from scipy import interpolate
+from sklearn.decomposition import PCA
+from sklearn.model_selection import KFold
+import cv2 as cv
+import logging
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(
+    thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca=0
+):
+    assert embeddings1.shape[0] == embeddings2.shape[0]
+    assert embeddings1.shape[1] == embeddings2.shape[1]
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        if pca > 0:
+            print("doing pca on", fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set]
+            )
+        best_threshold_index = np.argmax(acc_train)
+        for threshold_idx, threshold in enumerate(thresholds):
+            (
+                tprs[fold_idx, threshold_idx],
+                fprs[fold_idx, threshold_idx],
+                _,
+            ) = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set], actual_issame[test_set]
+        )
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame))
+    )
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(
+    thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10
+):
+    assert embeddings1.shape[0] == embeddings2.shape[0]
+    assert embeddings1.shape[1] == embeddings2.shape[1]
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set]
+            )
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind="slinear")
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set]
+        )
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    # print(true_accept, false_accept)
+    # print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(
+        thresholds,
+        embeddings1,
+        embeddings2,
+        np.asarray(actual_issame),
+        nrof_folds=nrof_folds,
+        pca=pca,
+    )
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(
+        thresholds,
+        embeddings1,
+        embeddings2,
+        np.asarray(actual_issame),
+        1e-3,
+        nrof_folds=nrof_folds,
+    )
+    return tpr, fpr, accuracy, val, val_std, far
+
+
+def load_bin_cv(path, image_size):
+    bins, issame_list = pickle.load(open(path, "rb"), encoding="bytes")
+    data_list = []
+    for flip in [0, 1]:
+        data = flow.empty(len(issame_list) * 2, 3, image_size[0], image_size[1])
+        data_list.append(data)
+    for i in range(len(issame_list) * 2):
+        _bin = bins[i]
+        img_ori = cv.imdecode(_bin, cv.IMREAD_COLOR)[:, :, ::-1]
+
+        for flip in [0, 1]:
+            img = img_ori.copy()
+            if flip == 1:
+                img = cv.flip(img, 1)
+            img = np.array(img).transpose((2, 0, 1))
+            img = (img - 127.5) * 0.00784313725
+            data_list[flip][i] = flow.tensor(img, dtype=flow.float)
+
+        if i % 1000 == 0:
+            logging.info("loading bin:%d", i)
+    logging.info(data_list[0].shape)
+    return data_list, issame_list
+
+
+@flow.no_grad()
+def test(data_set, backbone, batch_size, nfolds=10, is_consistent=False):
+    logging.info("testing verification..")
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    if is_consistent:
+        placement = flow.env.all_device_placement("cpu")
+        sbp = flow.sbp.split(0)
+
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            img = data[bb - batch_size : bb]
+            time0 = datetime.datetime.now()
+            with flow.no_grad():
+                if is_consistent:
+                    img = img.to_consistent(placement=placement, sbp=sbp)
+                net_out = backbone(img.to("cuda"))
+
+            if is_consistent:
+                _embeddings = net_out.to_local().numpy()
+            else:
+                _embeddings = net_out.detach().numpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count) :, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    logging.info(embeddings.shape)
+    logging.info("infer time:%f" % time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(
+        embeddings, issame_list, nrof_folds=nfolds
+    )
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+def dumpR(data_set, backbone, batch_size, name="", data_extra=None, label_shape=None):
+    print("dump verification embedding..")
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra), label=(_label,))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count) :, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    actual_issame = np.asarray(issame_list)
+    outname = os.path.join("temp.bin")
+    with open(outname, "wb") as f:
+        pickle.dump((embeddings, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
diff --git a/recognition/arcface_oneflow/function.py b/recognition/arcface_oneflow/function.py
new file mode 100644
index 000000000..0a9fa35d4
--- /dev/null
+++ b/recognition/arcface_oneflow/function.py
@@ -0,0 +1,261 @@
+import oneflow as flow
+from oneflow.nn.parallel import DistributedDataParallel as ddp
+from utils.ofrecord_data_utils import OFRecordDataLoader, SyntheticDataLoader
+from utils.utils_logging import AverageMeter
+from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
+from backbones import get_model
+from graph import TrainGraph, EvalGraph
+from utils.losses import CrossEntropyLoss_sbp
+import logging
+
+
+def make_data_loader(args, mode, is_consistent=False, synthetic=False):
+    assert mode in ("train", "validation")
+
+    if mode == "train":
+        total_batch_size = args.batch_size*flow.env.get_world_size()
+        batch_size = args.batch_size
+        num_samples = args.num_image
+    else:
+        total_batch_size = args.val_global_batch_size
+        batch_size = args.val_batch_size
+        num_samples = args.val_samples_per_epoch
+
+    placement = None
+    sbp = None
+
+    if is_consistent:
+        placement = flow.env.all_device_placement("cpu")
+        sbp = flow.sbp.split(0)
+        batch_size = total_batch_size
+
+    if synthetic:
+
+        data_loader = SyntheticDataLoader(
+            batch_size=batch_size,
+            num_classes=args.num_classes,
+            placement=placement,
+            sbp=sbp,
+        )
+        return data_loader.to("cuda")
+
+    ofrecord_data_loader = OFRecordDataLoader(
+        ofrecord_root=args.ofrecord_path,
+        mode=mode,
+        dataset_size=num_samples,
+        batch_size=batch_size,
+        total_batch_size=total_batch_size,
+        data_part_num=args.ofrecord_part_num,
+        placement=placement,
+        sbp=sbp,
+    )
+    return ofrecord_data_loader
+
+
+def make_optimizer(args, model):
+    param_group = {"params": [p for p in model.parameters() if p is not None]}
+
+    optimizer = flow.optim.SGD(
+        [param_group],
+        lr=args.lr,
+        momentum=args.momentum,
+        weight_decay=args.weight_decay,
+    )
+    return optimizer
+
+
+class FC7(flow.nn.Module):
+    def __init__(self, embedding_size, num_classes, cfg, partial_fc=False, bias=False):
+        super(FC7, self).__init__()
+        self.weight = flow.nn.Parameter(
+            flow.empty(num_classes, embedding_size))
+        flow.nn.init.normal_(self.weight, mean=0, std=0.01)
+
+        self.partial_fc = partial_fc
+
+        size = flow.env.get_world_size()
+        num_local = (cfg.num_classes + size - 1) // size
+        self.num_sample = int(num_local * cfg.sample_rate)
+        self.total_num_sample = self.num_sample * size
+
+    def forward(self, x, label):
+        x = flow.nn.functional.l2_normalize(input=x, dim=1, epsilon=1e-10)
+        if self.partial_fc:
+            (
+                mapped_label,
+                sampled_label,
+                sampled_weight,
+            ) = flow.distributed_partial_fc_sample(
+                weight=self.weight, label=label, num_sample=self.total_num_sample,
+            )
+            label = mapped_label
+            weight = sampled_weight
+        else:
+            weight = self.weight
+        weight = flow.nn.functional.l2_normalize(
+            input=weight, dim=1, epsilon=1e-10)
+        x = flow.matmul(x, weight, transpose_b=True)
+        if x.is_consistent:
+            return x, label
+        else:
+            return x
+
+
+class Train_Module(flow.nn.Module):
+    def __init__(self, cfg, backbone, placement, world_size):
+        super(Train_Module, self).__init__()
+        self.placement = placement
+
+        if cfg.graph:
+            if cfg.model_parallel:
+                input_size = cfg.embedding_size
+                output_size = int(cfg.num_classes/world_size)
+                self.fc = FC7(input_size, output_size, cfg, partial_fc=cfg.partial_fc).to_consistent(
+                    placement=placement, sbp=flow.sbp.split(0))
+            else:
+                self.fc = FC7(cfg.embedding_size, cfg.num_classes, cfg).to_consistent(
+                    placement=placement, sbp=flow.sbp.broadcast)
+            self.backbone = backbone.to_consistent(
+                placement=placement, sbp=flow.sbp.broadcast)
+        else:
+            self.backbone = backbone
+            self.fc = FC7(cfg.embedding_size, cfg.num_classes, cfg)
+
+    def forward(self, x, labels):
+        x = self.backbone(x)
+        if x.is_consistent:
+            x = x.to_consistent(sbp=flow.sbp.broadcast)
+        x = self.fc(x, labels)
+        return x
+
+
+class Trainer(object):
+    def __init__(self, cfg, placement, load_path, world_size, rank):
+        self.placement = placement
+        self.load_path = load_path
+        self.cfg = cfg
+        self.world_size = world_size
+        self.rank = rank
+
+        # model
+        self.backbone = get_model(cfg.network, dropout=0.0,
+                                  num_features=cfg.embedding_size).to("cuda")
+        self.train_module = Train_Module(
+            cfg, self.backbone, self.placement, world_size).to("cuda")
+        if cfg.resume:
+            if load_path is not None:
+                self.load_state_dict()
+            else:
+                logging.info("Model resume failed! load path is None ")
+
+        # optimizer
+        self.optimizer = make_optimizer(cfg, self.train_module)
+
+        # data
+        self.train_data_loader = make_data_loader(
+            cfg, 'train', self.cfg.graph, self.cfg.synthetic)
+
+        # loss
+        if cfg.loss == "cosface":
+            self.margin_softmax = flow.nn.CombinedMarginLoss(
+                1, 0., 0.4).to("cuda")
+        else:
+            self.margin_softmax = flow.nn.CombinedMarginLoss(
+                1, 0.5, 0.).to("cuda")
+
+        self.of_cross_entropy = CrossEntropyLoss_sbp()
+
+        # lr_scheduler
+        self.decay_step = self.cal_decay_step()
+        self.scheduler = flow.optim.lr_scheduler.MultiStepLR(
+            optimizer=self.optimizer, milestones=self.decay_step, gamma=0.1
+        )
+
+        # log
+        self.callback_logging = CallBackLogging(
+            50, rank, cfg.total_step, cfg.batch_size, world_size, None)
+        # val
+        self.callback_verification = CallBackVerification(
+            600, rank, cfg.val_targets, cfg.ofrecord_path, is_consistent=cfg.graph)
+        # save checkpoint
+        self.callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output)
+
+        self.losses = AverageMeter()
+        self.start_epoch = 0
+        self.global_step = 0
+
+    def __call__(self):
+        # Train
+        if self.cfg.graph:
+            self.train_graph()
+        else:
+            self.train_eager()
+
+    def load_state_dict(self):
+
+        if self.is_consistent:
+            state_dict = flow.load(self.load_path, consistent_src_rank=0)
+        elif self.rank == 0:
+            state_dict = flow.load(self.load_path)
+        else:
+            return
+        logging.info("Model resume successfully!")
+        self.model.load_state_dict(state_dict)
+
+    def cal_decay_step(self):
+        cfg = self.cfg
+        num_image = cfg.num_image
+        total_batch_size = cfg.batch_size * self.world_size
+        self.warmup_step = num_image // total_batch_size * cfg.warmup_epoch
+        self.cfg.total_step = num_image // total_batch_size * cfg.num_epoch
+        logging.info("Total Step is:%d" % self.cfg.total_step)
+        return [x * num_image // total_batch_size for x in cfg.decay_epoch]
+
+    def train_graph(self):
+        train_graph = TrainGraph(self.train_module, self.cfg, self.margin_softmax,
+                                 self.of_cross_entropy, self.train_data_loader, self.optimizer, self.scheduler)
+        # train_graph.debug()
+        val_graph = EvalGraph(self.backbone, self.cfg)
+
+        for epoch in range(self.start_epoch, self.cfg.num_epoch):
+            self.train_module.train()
+            one_epoch_steps = len(self.train_data_loader)
+            for steps in range(one_epoch_steps):
+                self.global_step += 1
+                loss = train_graph()
+                loss = loss.to_consistent(
+                    sbp=flow.sbp.broadcast).to_local().numpy()
+                self.losses.update(loss, 1)
+                self.callback_logging(self.global_step,  self.losses, epoch, False,
+                                      self.scheduler.get_last_lr()[0])
+                self.callback_verification(
+                    self.global_step, self.train_module, val_graph)
+            self.callback_checkpoint(self.global_step, epoch,
+                                     self.train_module, is_consistent=True)
+
+    def train_eager(self):
+        self.train_module = ddp(self.train_module)
+        for epoch in range(self.start_epoch, self.cfg.num_epoch):
+            self.train_module.train()
+
+            one_epoch_steps = len(self.train_data_loader)
+            for steps in range(one_epoch_steps):
+                self.global_step += 1
+                image, label = self.train_data_loader()
+                image = image.to("cuda")
+                label = label.to("cuda")
+                features_fc7 = self.train_module(image, label)
+                features_fc7 = self.margin_softmax(features_fc7, label)*64
+                loss = self.of_cross_entropy(features_fc7, label)
+                loss.backward()
+                self.optimizer.step()
+                self.optimizer.zero_grad()
+
+                loss = loss.numpy()
+                self.losses.update(loss, 1)
+                self.callback_logging(self.global_step,  self.losses, epoch, False,
+                                      self.scheduler.get_last_lr()[0])
+                self.callback_verification(self.global_step, self.backbone)
+                self.scheduler.step()
+            self.callback_checkpoint(
+                self.global_step, epoch, self.train_module)
diff --git a/recognition/arcface_oneflow/graph.py b/recognition/arcface_oneflow/graph.py
new file mode 100644
index 000000000..35a833ccf
--- /dev/null
+++ b/recognition/arcface_oneflow/graph.py
@@ -0,0 +1,73 @@
+import oneflow as flow
+import oneflow.nn as nn
+
+
+def make_static_grad_scaler():
+    return flow.amp.StaticGradScaler(flow.env.get_world_size())
+
+
+def make_grad_scaler():
+    return flow.amp.GradScaler(
+        init_scale=2 ** 30, growth_factor=2.0, backoff_factor=0.5, growth_interval=2000,
+    )
+
+
+def meter(self, mkey, *args):
+    assert mkey in self.m
+    self.m[mkey]["meter"].record(*args)
+
+
+class TrainGraph(flow.nn.Graph):
+    def __init__(
+        self,
+        model,
+        cfg,
+        combine_margin,
+        cross_entropy,
+        data_loader,
+        optimizer,
+        lr_scheduler=None,
+    ):
+        super().__init__()
+
+        if cfg.fp16:
+            self.config.enable_amp(True)
+            self.set_grad_scaler(make_grad_scaler())
+        elif cfg.scale_grad:
+            self.set_grad_scaler(make_static_grad_scaler())
+
+        self.config.allow_fuse_add_to_output(True)
+        self.config.allow_fuse_model_update_ops(True)
+
+        self.model = model
+
+        self.cross_entropy = cross_entropy
+        self.combine_margin = combine_margin
+        self.data_loader = data_loader
+        self.add_optimizer(optimizer, lr_sch=lr_scheduler)
+
+    def build(self):
+        image, label = self.data_loader()
+
+        image = image.to("cuda")
+        label = label.to("cuda")
+
+        logits, label = self.model(image, label)
+        logits = self.combine_margin(logits, label) * 64
+        loss = self.cross_entropy(logits, label)
+
+        loss.backward()
+        return loss
+
+
+class EvalGraph(flow.nn.Graph):
+    def __init__(self, model, cfg):
+        super().__init__()
+        self.config.allow_fuse_add_to_output(True)
+        self.model = model
+        if cfg.fp16:
+            self.config.enable_amp(True)
+
+    def build(self, image):
+        logits = self.model(image)
+        return logits
diff --git a/recognition/arcface_oneflow/oneflow2onnx.py b/recognition/arcface_oneflow/oneflow2onnx.py
new file mode 100644
index 000000000..b1713bc06
--- /dev/null
+++ b/recognition/arcface_oneflow/oneflow2onnx.py
@@ -0,0 +1,67 @@
+import os
+from os import mkdir
+from oneflow_onnx.oneflow2onnx.util import convert_to_onnx_and_check
+import oneflow as flow
+
+import logging
+from backbones import get_model
+from utils.utils_config import get_config
+import argparse
+import tempfile
+
+
+class ModelGraph(flow.nn.Graph):
+    def __init__(self, model):
+        super().__init__()
+        self.backbone = model
+
+    def build(self, x):
+        x = x.to("cuda")
+        out = self.backbone(x)
+        return out
+
+
+def convert_func(cfg, model_path, out_path,image_size):
+
+    model_module = get_model(cfg.network, dropout=0.0,
+                             num_features=cfg.embedding_size).to("cuda")
+    model_module.eval()
+    print(model_module)
+    model_graph = ModelGraph(model_module)
+    model_graph._compile(flow.randn(1, 3, image_size, image_size).to("cuda"))
+
+    with tempfile.TemporaryDirectory() as tmpdirname:
+        new_parameters = dict()
+        parameters = flow.load(model_path)
+        for key, value in parameters.items():
+            if "num_batches_tracked" not in key:
+                if key == "fc.weight":
+                    continue
+                val = value
+                new_key = key.replace("backbone.", "")
+                new_parameters[new_key] = val
+        model_module.load_state_dict(new_parameters)
+        flow.save(model_module.state_dict(), tmpdirname)
+        convert_to_onnx_and_check(
+            model_graph, flow_weight_dir=tmpdirname, onnx_model_path="./", print_outlier=True)
+
+
+def main(args):
+    logging.basicConfig(level=logging.NOTSET)
+    logging.info(args.model_path)
+    cfg = get_config(args.config)
+    if not os.path.exists(args.out_path):
+        mkdir(args.out_path)
+    convert_func(cfg, args.model_path, args.out_path,args.image_size)
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser(description='OneFlow ArcFace val')
+    parser.add_argument('config', type=str, help='py config file')
+    parser.add_argument('--model_path', type=str, help='model path')
+    parser.add_argument('--image_size', type=int,
+                        default=112, help='input image size')
+    parser.add_argument('--out_path', type=str,
+                        default="onnx_model", help='out path')
+
diff --git a/recognition/arcface_oneflow/requirements.txt b/recognition/arcface_oneflow/requirements.txt
new file mode 100644
index 000000000..9b31a23c6
--- /dev/null
+++ b/recognition/arcface_oneflow/requirements.txt
@@ -0,0 +1,7 @@
+numpy
+matplotlib
+Pillow
+opencv-python
+scikit-learn
+scipy
+easydict
\ No newline at end of file
diff --git a/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord.py b/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord.py
new file mode 100644
index 000000000..b7047a3ab
--- /dev/null
+++ b/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord.py
@@ -0,0 +1,136 @@
+import os
+import sys
+import struct
+import argparse
+
+from mxnet import recordio
+import oneflow.core.record.record_pb2 as of_record
+
+
+def parse_arguement(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--data_dir",
+        type=str,
+        default="insightface/datasets/faces_emore",
+        help="Root directory to mxnet dataset.",
+    )
+    parser.add_argument(
+        "--output_filepath",
+        type=str,
+        default="./output",
+        help="Path to output OFRecord.",
+    )
+    return parser.parse_args(argv)
+
+
+def load_train_data(data_dir):
+
+    path_imgrec = os.path.join(data_dir, "train.rec")
+    path_imgidx = path_imgrec[0:-4] + ".idx"
+
+    print(
+        "Loading recordio {}\n\
+  Corresponding record idx is {}".format(
+            path_imgrec, path_imgidx
+        )
+    )
+
+    imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, "r", key_type=int)
+
+    # Read header0 to get some info.
+    identity_key_start = 0
+    identity_key_end = 0
+    imgidx_list = []
+    id2range = {}
+
+    rec0 = imgrec.read_idx(0)
+    header0, img_str = recordio.unpack(rec0)
+    if header0.flag > 0:
+        identity_key_start = int(header0.label[0])
+        identity_key_end = int(header0.label[1])
+        imgidx_list = range(1, identity_key_start)
+
+        # Read identity id range
+        for identity in range(identity_key_start, identity_key_end):
+            rec = imgrec.read_idx(identity)
+            header, s = recordio.unpack(rec)
+            a, b = int(header.label[0]), int(header.label[1])
+            id2range[identity] = (a, b)
+
+    else:
+        imgidx_list = imgrec.keys
+    return imgrec, imgidx_list
+
+
+def convert_to_ofrecord(img_data):
+    """ Convert python dictionary formath data of one image to of.Example proto.
+  Args:
+    img_data: Python dict.
+  Returns:
+    example: The converted of.Exampl
+  """
+
+    def _int32_feature(value):
+        """Wrapper for inserting int32 features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(int32_list=of_record.Int32List(value=value))
+
+    def _float_feature(value):
+        """Wrapper for inserting float features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(float_list=of_record.FloatList(value=value))
+
+    def _double_feature(value):
+        """Wrapper for inserting float features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(double_list=of_record.DoubleList(value=value))
+
+    def _bytes_feature(value):
+        """Wrapper for inserting bytes features into Example proto."""
+        # if isinstance(value, six.string_types):
+        #  value = six.binary_type(value, encoding='utf-8')
+        return of_record.Feature(bytes_list=of_record.BytesList(value=[value]))
+
+    example = of_record.OFRecord(
+        feature={
+            "label": _int32_feature(img_data["label"]),
+            "encoded": _bytes_feature(img_data["pixel_data"]),
+        }
+    )
+
+    return example
+
+
+def main(args):
+    # Convert recordio to ofrecord
+    imgrec, imgidx_list = load_train_data(data_dir=args.data_dir)
+
+    output_dir = os.path.join(args.output_filepath, "train")
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    output_file = os.path.join(output_dir, "part-0")
+    with open(output_file, "wb") as f:
+        for idx in imgidx_list:
+            if idx % 10000 == 0:
+                print("Converting images: {} of {}".format(idx, len(imgidx_list)))
+
+            img_data = {}
+            rec = imgrec.read_idx(idx)
+            header, s = recordio.unpack(rec)
+            img_data["label"] = int(header.label[0])
+            img_data["pixel_data"] = s
+
+            example = convert_to_ofrecord(img_data)
+            print("shape", len(img_data["pixel_data"]))
+            size = example.ByteSize()
+            f.write(struct.pack("q", size))
+            f.write(example.SerializeToString())
+
+
+if __name__ == "__main__":
+    main(parse_arguement(sys.argv[1:]))
diff --git a/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord_shuffled_npart.py b/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord_shuffled_npart.py
new file mode 100644
index 000000000..ba3b82c15
--- /dev/null
+++ b/recognition/arcface_oneflow/tools/mx_recordio_2_ofrecord_shuffled_npart.py
@@ -0,0 +1,156 @@
+import os
+import sys
+import struct
+import argparse
+import numbers
+import random
+
+from mxnet import recordio
+import oneflow.core.record.record_pb2 as of_record
+
+
+def parse_arguement(argv):
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--data_dir",
+        type=str,
+        default="insightface/datasets/faces_emore",
+        help="Root directory to mxnet dataset.",
+    )
+    parser.add_argument(
+        "--output_filepath",
+        type=str,
+        default="./ofrecord",
+        help="Path to output OFRecord.",
+    )
+    parser.add_argument(
+        "--num_part", type=int, default=96, help="num_part of OFRecord to generate.",
+    )
+    return parser.parse_args(argv)
+
+
+def load_train_data(data_dir):
+
+    path_imgrec = os.path.join(data_dir, "train.rec")
+    path_imgidx = path_imgrec[0:-4] + ".idx"
+
+    print(
+        "Loading recordio {}\n\
+        Corresponding record idx is {}".format(
+            path_imgrec, path_imgidx
+        )
+    )
+
+    imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, "r", key_type=int)
+
+    # Read header0 to get some info.
+    identity_key_start = 0
+    identity_key_end = 0
+    imgidx_list = []
+    id2range = {}
+
+    rec0 = imgrec.read_idx(0)
+    header0, img_str = recordio.unpack(rec0)
+    if header0.flag > 0:
+        identity_key_start = int(header0.label[0])
+        identity_key_end = int(header0.label[1])
+        imgidx_list = range(1, identity_key_start)
+
+        # Read identity id range
+        for identity in range(identity_key_start, identity_key_end):
+            rec = imgrec.read_idx(identity)
+            header, s = recordio.unpack(rec)
+            a, b = int(header.label[0]), int(header.label[1])
+            id2range[identity] = (a, b)
+
+    else:
+        imgidx_list = imgrec.keys
+
+    return imgrec, imgidx_list
+
+
+def convert_to_ofrecord(img_data):
+    """ Convert python dictionary formath data of one image to of.Example proto.
+  Args:
+    img_data: Python dict.
+  Returns:
+    example: The converted of.Exampl
+  """
+
+    def _int32_feature(value):
+        """Wrapper for inserting int32 features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(int32_list=of_record.Int32List(value=value))
+
+    def _float_feature(value):
+        """Wrapper for inserting float features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(float_list=of_record.FloatList(value=value))
+
+    def _double_feature(value):
+        """Wrapper for inserting float features into Example proto."""
+        if not isinstance(value, list):
+            value = [value]
+        return of_record.Feature(double_list=of_record.DoubleList(value=value))
+
+    def _bytes_feature(value):
+        """Wrapper for inserting bytes features into Example proto."""
+        # if isinstance(value, six.string_types):
+        #  value = six.binary_type(value, encoding='utf-8')
+        return of_record.Feature(bytes_list=of_record.BytesList(value=[value]))
+
+    example = of_record.OFRecord(
+        feature={
+            "label": _int32_feature(img_data["label"]),
+            "encoded": _bytes_feature(img_data["pixel_data"]),
+        }
+    )
+
+    return example
+
+
+def main(args):
+    # Convert recordio to ofrecord
+    imgrec, imgidx_list = load_train_data(data_dir=args.data_dir)
+    imgidx_list = list(imgidx_list)
+    random.shuffle(imgidx_list)
+
+    output_dir = os.path.join(args.output_filepath, "train")
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    num_images = len(imgidx_list)
+    num_images_per_part = (num_images + args.num_part) // args.num_part
+    print("num_images", num_images, "num_images_per_part", num_images_per_part)
+
+    for part_id in range(args.num_part):
+        part_name = "part-" + "{:0>5d}".format(part_id)
+        output_file = os.path.join(output_dir, part_name)
+        file_idx_start = part_id * num_images_per_part
+        file_idx_end = min((part_id + 1) * num_images_per_part, num_images)
+        print("part-" + str(part_id), "start", file_idx_start, "end", file_idx_end)
+        with open(output_file, "wb") as f:
+            for file_idx in range(file_idx_start, file_idx_end):
+                idx = imgidx_list[file_idx]
+                if idx % 10000 == 0:
+                    print("Converting images: {} of {}".format(idx, len(imgidx_list)))
+
+                img_data = {}
+                rec = imgrec.read_idx(idx)
+                header, s = recordio.unpack(rec)
+                label = header.label
+                if not isinstance(label, numbers.Number):
+                    label = label[0]
+                img_data["label"] = int(label)
+                img_data["pixel_data"] = s
+
+                example = convert_to_ofrecord(img_data)
+                size = example.ByteSize()
+                f.write(struct.pack("q", size))
+                f.write(example.SerializeToString())
+
+
+if __name__ == "__main__":
+    main(parse_arguement(sys.argv[1:]))
diff --git a/recognition/arcface_oneflow/train.py b/recognition/arcface_oneflow/train.py
new file mode 100644
index 000000000..477db3ce3
--- /dev/null
+++ b/recognition/arcface_oneflow/train.py
@@ -0,0 +1,43 @@
+import argparse
+import logging
+import os
+import oneflow as flow
+
+from function import Trainer
+from utils.utils_logging import init_logging
+from utils.utils_config import get_config
+
+
+def main(args):
+    cfg = get_config(args.config)
+    cfg.graph = args.graph
+    rank = flow.env.get_rank()
+    world_size = flow.env.get_world_size()
+    placement = flow.env.all_device_placement("cuda")
+
+    os.makedirs(cfg.output, exist_ok=True)
+    log_root = logging.getLogger()
+    init_logging(log_root, rank, cfg.output)
+
+    # root dir of loading checkpoint
+    load_path = None
+
+    for key, value in cfg.items():
+        num_space = 25 - len(key)
+        logging.info(": " + key + " " * num_space + str(value))
+
+    trainer = Trainer(cfg, placement, load_path, world_size, rank)
+    trainer()
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser(description="OneFlow ArcFace Training")
+    parser.add_argument("config", type=str, help="py config file")
+    parser.add_argument(
+        "--graph",
+        action="store_true",
+        help="Run model in graph mode,else run model in ddp mode.",
+    )
+    parser.add_argument("--local_rank", type=int, default=0, help="local_rank")
+    main(parser.parse_args())
diff --git a/recognition/arcface_oneflow/train_ddp.sh b/recognition/arcface_oneflow/train_ddp.sh
new file mode 100644
index 000000000..86ed5415c
--- /dev/null
+++ b/recognition/arcface_oneflow/train_ddp.sh
@@ -0,0 +1,25 @@
+# set -aux
+
+MASTER_ADDR=127.0.0.1
+MASTER_PORT=17788
+DEVICE_NUM_PER_NODE=8
+NUM_NODES=1
+NODE_RANK=0
+
+
+export PYTHONUNBUFFERED=1
+echo PYTHONUNBUFFERED=$PYTHONUNBUFFERED
+export NCCL_LAUNCH_MODE=PARALLEL
+echo NCCL_LAUNCH_MODE=$NCCL_LAUNCH_MODE
+export NCCL_DEBUG=INFO
+export ONEFLOW_DEBUG_MODE=True
+
+
+NCCL_DEBUG=INFO \
+python3 -m oneflow.distributed.launch \
+--nproc_per_node $DEVICE_NUM_PER_NODE \
+--nnodes $NUM_NODES \
+--node_rank $NODE_RANK \
+--master_addr $MASTER_ADDR \
+--master_port $MASTER_PORT \
+train.py configs/ms1mv3_r50.py 
diff --git a/recognition/arcface_oneflow/train_graph_distributed.sh b/recognition/arcface_oneflow/train_graph_distributed.sh
new file mode 100644
index 000000000..4f5725059
--- /dev/null
+++ b/recognition/arcface_oneflow/train_graph_distributed.sh
@@ -0,0 +1,26 @@
+# set -aux
+
+MASTER_ADDR=127.0.0.1
+MASTER_PORT=17788
+DEVICE_NUM_PER_NODE=8
+NUM_NODES=1
+NODE_RANK=0
+
+
+
+export PYTHONUNBUFFERED=1
+echo PYTHONUNBUFFERED=$PYTHONUNBUFFERED
+export NCCL_LAUNCH_MODE=PARALLEL
+echo NCCL_LAUNCH_MODE=$NCCL_LAUNCH_MODE
+#export NCCL_DEBUG=INFO
+export ONEFLOW_DEBUG_MODE=True
+
+
+#NCCL_DEBUG=INFO
+python3 -m oneflow.distributed.launch \
+--nproc_per_node $DEVICE_NUM_PER_NODE \
+--nnodes $NUM_NODES \
+--node_rank $NODE_RANK \
+--master_addr $MASTER_ADDR \
+--master_port $MASTER_PORT \
+train.py configs/ms1mv3_r50.py --graph 
diff --git a/recognition/arcface_oneflow/utils/__init__.py b/recognition/arcface_oneflow/utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_oneflow/utils/losses.py b/recognition/arcface_oneflow/utils/losses.py
new file mode 100644
index 000000000..5d4aa1d03
--- /dev/null
+++ b/recognition/arcface_oneflow/utils/losses.py
@@ -0,0 +1,66 @@
+import oneflow as flow
+from oneflow import nn
+
+
+def get_loss(name):
+    if name == "cosface":
+        return CosFace()
+    elif name == "arcface":
+        return ArcFace()
+    else:
+        raise ValueError()
+
+
+class CrossEntropyLoss_sbp(nn.Module):
+    def __init__(self):
+        super(CrossEntropyLoss_sbp, self).__init__()
+
+    def forward(self, logits, label):
+        loss = flow._C.sparse_softmax_cross_entropy(
+            logits, label)
+        loss = flow.mean(loss)
+        return loss
+
+
+def get_loss(name):
+    if name == "cosface":
+        return CosFace()
+    elif name == "arcface":
+        return ArcFace()
+    else:
+        raise ValueError()
+
+
+class CosFace(nn.Module):
+    def __init__(self, s=64.0, m=0.40):
+        super(CosFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, cosine, label):
+        index = flow.where(label != -1)[0]
+        m_hot = flow.zeros(index.size()[0], cosine.size()[
+                           1], device=cosine.device)
+
+        m_hot = flow.scatter(m_hot, 1, label[index, None], self.m)
+        cosine = cosine[index] - m_hot
+
+        ret = cosine * self.s
+        return ret
+
+
+class ArcFace(nn.Module):
+    def __init__(self, s=64.0, m=0.5):
+        super(ArcFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, cosine: flow.Tensor, label):
+        index = flow.where(label != -1)[0]
+        m_hot = flow.zeros(index.size()[0], cosine.size()[
+                           1], device=cosine.device)
+        m_hot.scatter_(1, label[index, None], self.m)
+        cosine.acos_()
+        cosine[index] += m_hot
+        cosine.cos_().mul_(self.s)
+        return cosine
diff --git a/recognition/arcface_oneflow/utils/ofrecord_data_utils.py b/recognition/arcface_oneflow/utils/ofrecord_data_utils.py
new file mode 100644
index 000000000..449fe2521
--- /dev/null
+++ b/recognition/arcface_oneflow/utils/ofrecord_data_utils.py
@@ -0,0 +1,148 @@
+import oneflow as flow
+import oneflow.nn as nn
+import os
+from typing import List, Union
+
+
+class OFRecordDataLoader(nn.Module):
+    def __init__(
+        self,
+        ofrecord_root: str = "./ofrecord",
+        mode: str = "train",  # "val"
+        dataset_size: int = 9469,
+        batch_size: int = 1,
+        total_batch_size: int = 1,
+        data_part_num: int = 8,
+        placement: flow.placement = None,
+        sbp: Union[flow.sbp.sbp, List[flow.sbp.sbp]] = None,
+    ):
+        super().__init__()
+        channel_last = False
+        output_layout = "NHWC" if channel_last else "NCHW"
+        assert (ofrecord_root, mode)
+        self.train_record_reader = flow.nn.OfrecordReader(
+            os.path.join(ofrecord_root, mode),
+            batch_size=batch_size,
+            data_part_num=data_part_num,
+            part_name_suffix_length=5,
+            random_shuffle=True if mode == "train" else False,
+            shuffle_after_epoch=True if mode == "train" else False,
+            placement=placement,
+            sbp=sbp,
+        )
+        self.record_label_decoder = flow.nn.OfrecordRawDecoder(
+            "label", shape=(), dtype=flow.int32
+        )
+
+        color_space = "RGB"
+        height = 112
+        width = 112
+
+        self.record_image_decoder = flow.nn.OFRecordImageDecoder(
+            "encoded", color_space=color_space
+        )
+        self.resize = (
+            flow.nn.image.Resize(target_size=[height, width])
+            if mode == "train"
+            else flow.nn.image.Resize(
+                resize_side="shorter", keep_aspect_ratio=True, target_size=112
+            )
+        )
+
+        self.flip = (
+            flow.nn.CoinFlip(batch_size=batch_size, placement=placement, sbp=sbp)
+            if mode == "train"
+            else None
+        )
+
+        rgb_mean = [127.5, 127.5, 127.5]
+        rgb_std = [127.5, 127.5, 127.5]
+        self.crop_mirror_norm = (
+            flow.nn.CropMirrorNormalize(
+                color_space=color_space,
+                output_layout=output_layout,
+                mean=rgb_mean,
+                std=rgb_std,
+                output_dtype=flow.float,
+            )
+            if mode == "train"
+            else flow.nn.CropMirrorNormalize(
+                color_space=color_space,
+                output_layout=output_layout,
+                crop_h=0,
+                crop_w=0,
+                crop_pos_y=0.5,
+                crop_pos_x=0.5,
+                mean=rgb_mean,
+                std=rgb_std,
+                output_dtype=flow.float,
+            )
+        )
+
+        self.batch_size = batch_size
+        self.total_batch_size = total_batch_size
+        self.dataset_size = dataset_size
+
+    def __len__(self):
+        return self.dataset_size // self.total_batch_size
+
+    def forward(self):
+        train_record = self.train_record_reader()
+        label = self.record_label_decoder(train_record)
+        image_raw_buffer = self.record_image_decoder(train_record)
+
+        image = self.resize(image_raw_buffer)[0]
+
+        rng = self.flip() if self.flip != None else None
+        image = self.crop_mirror_norm(image, rng)
+
+        return image, label
+
+
+class SyntheticDataLoader(flow.nn.Module):
+    def __init__(
+        self, batch_size, image_size=112, num_classes=10000, placement=None, sbp=None,
+    ):
+        super().__init__()
+
+        self.image_shape = (batch_size, 3, image_size, image_size)
+        self.label_shape = (batch_size,)
+        self.num_classes = num_classes
+        self.placement = placement
+        self.sbp = sbp
+
+        if self.placement is not None and self.sbp is not None:
+            self.image = flow.nn.Parameter(
+                flow.randint(
+                    0,
+                    high=255,
+                    size=self.image_shape,
+                    dtype=flow.float32,
+                    placement=self.placement,
+                    sbp=self.sbp,
+                ),
+                requires_grad=False,
+            )
+            self.label = flow.nn.Parameter(
+                flow.randint(
+                    0,
+                    high=self.num_classes,
+                    size=self.label_shape,
+                    placement=self.placement,
+                    sbp=self.sbp,
+                ).to(dtype=flow.int32),
+                requires_grad=False,
+            )
+        else:
+            self.image = flow.randint(
+                0, high=255, size=self.image_shape, dtype=flow.float32, device="cuda"
+            )
+            self.label = flow.randint(
+                0, high=self.num_classes, size=self.label_shape, device="cuda",
+            ).to(dtype=flow.int32)
+
+    def __len__(self):
+        return 10000
+
+    def forward(self):
+        return self.image, self.label
diff --git a/recognition/arcface_oneflow/utils/utils_callbacks.py b/recognition/arcface_oneflow/utils/utils_callbacks.py
new file mode 100644
index 000000000..6d349c344
--- /dev/null
+++ b/recognition/arcface_oneflow/utils/utils_callbacks.py
@@ -0,0 +1,181 @@
+import logging
+import os
+import time
+from typing import List
+
+import oneflow as flow
+
+from eval import verification
+from utils.utils_logging import AverageMeter
+
+
+class CallBackVerification(object):
+    def __init__(
+        self,
+        frequent,
+        rank,
+        val_targets,
+        rec_prefix,
+        image_size=(112, 112),
+        world_size=1,
+        is_consistent=False,
+    ):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.highest_acc: float = 0.0
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        self.world_size = world_size
+        self.is_consistent = is_consistent
+
+        if self.is_consistent:
+            self.init_dataset(
+                val_targets=val_targets, data_dir=rec_prefix, image_size=image_size
+            )
+        else:
+            if self.rank is 0:
+                self.init_dataset(
+                    val_targets=val_targets, data_dir=rec_prefix, image_size=image_size
+                )
+
+    def ver_test(self, backbone: flow.nn.Module, global_step: int):
+        results = []
+        for i in range(len(self.ver_list)):
+
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                self.ver_list[i], backbone, 10, 10, self.is_consistent
+            )
+            logging.info(
+                "[%s][%d]XNorm: %f" % (
+                    self.ver_name_list[i], global_step, xnorm)
+            )
+            logging.info(
+                "[%s][%d]Accuracy-Flip: %1.5f+-%1.5f"
+                % (self.ver_name_list[i], global_step, acc2, std2)
+            )
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info(
+                "[%s][%d]Accuracy-Highest: %1.5f"
+                % (self.ver_name_list[i], global_step, self.highest_acc_list[i])
+            )
+            results.append(acc2)
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+
+        for name in val_targets:
+            path = os.path.join(data_dir, "val", name + ".bin")
+            if os.path.exists(path):
+                data_set = verification.load_bin_cv(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+        if len(self.ver_list) == 0:
+            logging.info("Val targets is None !")
+
+    def __call__(self, num_update, backbone: flow.nn.Module, backbone_graph=None):
+
+        if self.is_consistent:
+            if num_update > 0 and num_update % self.frequent == 0:
+                backbone.eval()
+                self.ver_test(backbone_graph, num_update)
+                backbone.train()
+        else:
+            if self.rank is 0 and num_update > 0 and num_update % self.frequent == 0:
+                backbone.eval()
+                self.ver_test(backbone, num_update)
+                backbone.train()
+
+
+class CallBackLogging(object):
+    def __init__(self, frequent, rank, total_step, batch_size, world_size, writer=None):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.time_start = time.time()
+        self.total_step: int = total_step
+        self.batch_size: int = batch_size
+        self.world_size: int = world_size
+        self.writer = writer
+
+        self.init = False
+        self.tic = 0
+
+    def __call__(
+        self,
+        global_step: int,
+        loss: AverageMeter,
+        epoch: int,
+        fp16: bool,
+        learning_rate: float,
+        grad_scaler=None,
+    ):
+        if self.rank == 0 and global_step % self.frequent == 0:
+            if self.init:
+                try:
+                    speed: float = self.frequent * self.batch_size / (
+                        time.time() - self.tic
+                    )
+                    speed_total = speed * self.world_size
+                except ZeroDivisionError:
+                    speed_total = float("inf")
+
+                time_now = (time.time() - self.time_start) / 3600
+                time_total = time_now / ((global_step + 1) / self.total_step)
+                time_for_end = time_total - time_now
+                if self.writer is not None:
+                    self.writer.add_scalar(
+                        "time_for_end", time_for_end, global_step)
+                    self.writer.add_scalar(
+                        "learning_rate", learning_rate, global_step)
+                    self.writer.add_scalar("loss", loss.avg, global_step)
+                if fp16:
+                    msg = (
+                        "Speed %.2f samples/sec   Loss %.4f   LearningRate %.4f   Epoch: %d   Global Step: %d   "
+                        "Fp16 Grad Scale: %2.f   Required: %1.f hours"
+                        % (
+                            speed_total,
+                            loss.avg,
+                            learning_rate,
+                            epoch,
+                            global_step,
+                            time_for_end,
+                        )
+                    )
+                else:
+                    msg = (
+                        "Speed %.2f samples/sec   Loss %.4f   LearningRate %.4f   Epoch: %d   Global Step: %d   "
+                        "Required: %1.f hours"
+                        % (
+                            speed_total,
+                            loss.avg,
+                            learning_rate,
+                            epoch,
+                            global_step,
+                            time_for_end,
+                        )
+                    )
+                logging.info(msg)
+                loss.reset()
+                self.tic = time.time()
+            else:
+                self.init = True
+                self.tic = time.time()
+
+
+class CallBackModelCheckpoint(object):
+    def __init__(self, rank, output="./"):
+        self.rank: int = rank
+        self.output: str = output
+
+    def __call__(self, global_step, epoch, backbone, is_consistent=False):
+
+        if global_step > 100 and backbone is not None:
+            path_module = os.path.join(self.output, "epoch_%d" % (epoch))
+
+            if is_consistent:
+                flow.save(backbone.state_dict(),
+                          path_module, consistent_dst_rank=0)
+            else:
+                if self.rank == 0:
+                    flow.save(backbone.state_dict(), path_module)
+            logging.info("oneflow Model Saved in '{}'".format(path_module))
diff --git a/recognition/arcface_oneflow/utils/utils_config.py b/recognition/arcface_oneflow/utils/utils_config.py
new file mode 100644
index 000000000..b61724bd0
--- /dev/null
+++ b/recognition/arcface_oneflow/utils/utils_config.py
@@ -0,0 +1,18 @@
+import importlib
+import os.path as osp
+
+
+def get_config(config_file):
+    assert config_file.startswith(
+        "configs/"
+    ), "config file setting must start with configs/"
+    temp_config_name = osp.basename(config_file)
+    temp_module_name = osp.splitext(temp_config_name)[0]
+    config = importlib.import_module("configs.base")
+    cfg = config.config
+    config = importlib.import_module("configs.%s" % temp_module_name)
+    job_cfg = config.config
+    cfg.update(job_cfg)
+    if cfg.output is None:
+        cfg.output = osp.join("work_dirs", temp_module_name)
+    return cfg
diff --git a/recognition/arcface_oneflow/utils/utils_logging.py b/recognition/arcface_oneflow/utils/utils_logging.py
new file mode 100644
index 000000000..543a7e14a
--- /dev/null
+++ b/recognition/arcface_oneflow/utils/utils_logging.py
@@ -0,0 +1,40 @@
+import logging
+import os
+import sys
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value
+    """
+
+    def __init__(self):
+        self.val = None
+        self.avg = None
+        self.sum = None
+        self.count = None
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def init_logging(log_root, rank, models_root):
+    if rank is 0:
+        log_root.setLevel(logging.INFO)
+        formatter = logging.Formatter("Training: %(asctime)s-%(message)s")
+        handler_file = logging.FileHandler(os.path.join(models_root, "training.log"))
+        handler_stream = logging.StreamHandler(sys.stdout)
+        handler_file.setFormatter(formatter)
+        handler_stream.setFormatter(formatter)
+        log_root.addHandler(handler_file)
+        log_root.addHandler(handler_stream)
+        log_root.info("rank_id: %d" % rank)
diff --git a/recognition/arcface_oneflow/val.py b/recognition/arcface_oneflow/val.py
new file mode 100644
index 000000000..eb85074a8
--- /dev/null
+++ b/recognition/arcface_oneflow/val.py
@@ -0,0 +1,43 @@
+import oneflow as flow
+from utils.utils_callbacks import CallBackVerification
+from backbones import get_model
+from graph import TrainGraph, EvalGraph
+import logging
+import argparse
+from utils.utils_config import get_config
+
+
+def main(args):
+
+    cfg = get_config(args.config)
+    logging.basicConfig(level=logging.NOTSET)
+    logging.info(args.model_path)
+
+    backbone = get_model(cfg.network, dropout=0.0, num_features=cfg.embedding_size).to(
+        "cuda"
+    )
+    val_callback = CallBackVerification(
+        1, 0, cfg.val_targets, cfg.ofrecord_path)
+
+    state_dict = flow.load(args.model_path)
+
+    new_parameters = dict()
+    for key, value in state_dict.items():
+        if "num_batches_tracked" not in key:
+            if key == "fc.weight":
+                continue
+            new_key = key.replace("backbone.", "")
+            new_parameters[new_key] = value
+
+    backbone.load_state_dict(new_parameters)
+
+    infer_graph = EvalGraph(backbone, cfg)
+    val_callback(1000, backbone, infer_graph)
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser(description="OneFlow ArcFace val")
+    parser.add_argument("config", type=str, help="py config file")
+    parser.add_argument("--model_path", type=str, help="model path")
+    main(parser.parse_args())
diff --git a/recognition/arcface_oneflow/val.sh b/recognition/arcface_oneflow/val.sh
new file mode 100644
index 000000000..8151f400c
--- /dev/null
+++ b/recognition/arcface_oneflow/val.sh
@@ -0,0 +1,3 @@
+#!/bin/bash
+
+python val.py configs/ms1mv3_r50 --model_path eager_test/epoch_0
diff --git a/recognition/arcface_paddle/README.md b/recognition/arcface_paddle/README.md
new file mode 120000
index 000000000..13c4f964b
--- /dev/null
+++ b/recognition/arcface_paddle/README.md
@@ -0,0 +1 @@
+README_en.md
\ No newline at end of file
diff --git a/recognition/arcface_paddle/README_cn.md b/recognition/arcface_paddle/README_cn.md
new file mode 100644
index 000000000..840ff77db
--- /dev/null
+++ b/recognition/arcface_paddle/README_cn.md
@@ -0,0 +1,281 @@
+简体中文 | [English](README_en.md)
+
+# Arcface-Paddle
+
+* [1. 简介](#简介)
+* [2. 环境准备](#环境准备)
+* [3. 数据准备](#数据准备)
+  * [3.1 下载数据集](#下载数据集)
+  * [3.2 从 MXNet 格式数据集抽取图像](#从MXNet格式数据集抽取图像)
+* [4. 训练](#训练)
+  * [4.1 单机单卡](#单机单卡)
+  * [4.2 单机 8 卡](#单机8卡)
+* [5. 模型评价](#模型评价)
+* [6. 模型导出](#模型导出)
+* [7 模型推理](#模型推理)
+* [8 模型性能](#模型性能)
+  * [8.1 轻量化模型性能](#轻量化模型性能)
+  * [8.2 验证集准确率](#验证集准确率)
+  * [8.3 最大类别数支持](#最大类别数支持)
+  * [8.4 吞吐量对比](#吞吐量对比)
+* [9. 全流程推理](#全流程推理)
+
+<a name="简介"></a>
+
+## 1. 简介
+
+`Arcface-Paddle`是基于PaddlePaddle实现的，开源深度人脸检测、识别工具。`Arcface-Paddle`目前提供了三个预训练模型，包括用于人脸检测的 `BlazeFace`、用于人脸识别的 `ArcFace` 和 `MobileFace`。
+
+- 本部分内容为人脸识别部分。
+- 人脸检测相关内容可以参考：[基于BlazeFace的人脸检测](../../detection/blazeface_paddle/README_cn.md)。
+- 基于PaddleInference的Whl包预测部署内容可以参考：[Whl包预测部署](https://github.com/littletomatodonkey/insight-face-paddle)。
+
+注: 在此非常感谢 [GuoQuanhao](https://github.com/GuoQuanhao) 基于PaddlePaddle复现了 [Arcface的基线模型](https://github.com/GuoQuanhao/arcface-Paddle)。
+
+<a name="环境准备"></a>
+
+## 2. 环境准备
+
+请参照 [Installation](./install_cn.md) 配置实验所需环境。
+
+<a name="数据准备"></a>
+
+## 3. 数据准备
+
+<a name="下载数据集"></a>
+
+### 3.1 下载数据集
+
+数据集可以从 [insightface datasets](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_) 下载.
+
+* MS1M_v2: MS1M-ArcFace
+* MS1M_v3: MS1M-RetinaFace
+
+<a name="从MXNet格式数据集抽取图像"></a>
+
+### 3.2 从 MXNet 格式数据集抽取图像
+
+```shell
+python tools/mx_recordio_2_images.py --root_dir ms1m-retinaface-t1/ --output_dir MS1M_v3/
+```
+
+当数据集抽取完后，输出的图像数据集目录结构如下：
+
+```
+MS1M_v3
+|_ images
+|  |_ 00000001.jpg
+|  |_ ...
+|  |_ 05179510.jpg
+|_ label.txt
+|_ agedb_30.bin
+|_ cfp_ff.bin
+|_ cfp_fp.bin
+|_ lfw.bin
+```
+
+标签文件格式如下：
+
+```
+# 图像路径与标签的分隔符: "\t"
+# 以下是 label.txt 每行的格式
+images/00000001.jpg 0
+...
+```
+
+如果你想使用自定义数据集训练，可以根据以上目录结构和标签文件格式组织数据。
+
+<a name="训练"></a>
+
+## 4. 训练
+
+<a name="单机单卡"></a>
+
+### 4.1 单机单卡
+
+```bash
+export CUDA_VISIBLE_DEVICES=1
+python tools/train.py \
+    --config_file configs/ms1mv2_mobileface.py \
+    --embedding_size 128 \
+    --sample_ratio 1.0 \
+    --loss ArcFace \
+    --batch_size 512 \
+    --dataset MS1M_v2 \
+    --num_classes 85742 \
+    --data_dir MS1M_v2/ \
+    --label_file MS1M_v2/label.txt \
+    --fp16 False
+```
+<a name="单机8卡"></a>
+
+### 4.2 单机 8 卡
+
+为了方便训练，已经为用户准备好训练启动脚本。
+
+#### 静态图模式训练
+
+```bash
+sh scripts/train_static.sh
+```
+
+#### 动态图模式训练
+
+```bash
+sh scripts/train_dynamic.sh
+```
+
+注：多机器多卡训练参见 ``paddle.distributed.launch`` API 文档。单机与多机训练不同之处在于多机需要设置 ``--ips`` 参数。
+
+在训练过程中，你可以实时通过 `VisualDL` 可视化查看 loss 的变化，更多信息可以参考 [VisualDL](https://github.com/PaddlePaddle/VisualDL/)。
+
+<a name="模型评价"></a>
+
+## 5. 模型评价
+
+模型评价可以通过以下脚本启动
+
+#### 静态图模式
+
+```bash
+sh scripts/validation_static.sh
+```
+
+#### 动态图模式
+
+```bash
+sh scripts/validation_dynamic.sh
+```
+
+<a name="模型导出"></a>
+
+## 6. 模型导出
+
+PaddlePaddle 支持用预测引擎直接推理，首先，需要导出推理模型，通过以下脚本进行导出
+
+#### 静态图模式
+
+```bash
+sh scripts/export_static.sh
+```
+
+#### 动态图模式
+
+```bash
+sh scripts/export_dynamic.sh
+```
+
+<a name="模型推理"></a>
+
+## 7. 模型推理
+
+模型推理过程支持 paddle 格式的 ``save inference model`` 和 onnx 格式。
+
+```bash
+sh scripts/inference.sh
+```
+
+<a name="模型性能"></a>
+
+## 8. 模型性能
+
+<a name="轻量化模型性能"></a>
+
+### 8.1 轻量化模型性能
+
+**配置：**
+  * CPU: Intel(R) Xeon(R) Gold 6184 CPU @ 2.40GHz
+  * GPU: a single NVIDIA Tesla V100
+  * Precison: FP32
+  * BatchSize: 64/512
+  * SampleRatio: 1.0
+  * Embedding Size: 128
+  * MS1MV2
+
+| Model structure           | lfw    | cfp_fp  | agedb30 | CPU time cost | GPU time cost | Inference model |
+| ------------------------- | ------ | ------- | ------- | -------| -------- |---- |
+| MobileFace-Paddle      | 0.9952 | 0.9280  | 0.9612  | 4.3ms  | 2.3ms    | [download link](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar)  |
+| MobileFace-mxnet       | 0.9950 | 0.8894  | 0.9591  | 7.3ms  | 4.7ms    | -   |
+
+* 注: MobileFace-Paddle 是使用 MobileFaceNet_128 backbone 训练出的模型
+
+<a name="验证集准确率"></a>
+
+### 8.2 验证集准确率
+
+**配置：**
+  * GPU: 8 NVIDIA Tesla V100 32G
+  * Precison: Pure FP16
+  * BatchSize: 128/1024
+
+| Mode    | Datasets | backbone | Ratio | agedb30 | cfp_fp | lfw  | log  | checkpoint |
+| ------- | :------: | :------- | ----- | :------ | :----- | :--- | :--- |  :--- |
+| Static  |  MS1MV3  | r50      | 0.1   | 0.98317 | 0.98943| 0.99850 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/static/ms1mv3_r50_static_128_fp16_0.1/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_static_128_fp16_0.1_epoch_24.tgz) |
+| Static  |  MS1MV3  | r50      | 1.0   | 0.98283 | 0.98843| 0.99850 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/static/ms1mv3_r50_static_128_fp16_1.0/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_static_128_fp16_1.0_epoch_24.tgz) |
+| Dynamic |  MS1MV3  | r50      | 0.1   | 0.98333 | 0.98900| 0.99833 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/dynamic/ms1mv3_r50_dynamic_128_fp16_0.1/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_dynamic_128_fp16_0.1_eopch_24.tgz) |
+| Dynamic |  MS1MV3  | r50      | 1.0   | 0.98317 | 0.98900| 0.99833 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/dynamic/ms1mv3_r50_dynamic_128_fp16_1.0/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_dynamic_128_fp16_1.0_eopch_24.tgz) |
+
+<a name="最大类别数支持"></a>
+
+### 8.3 最大类别数支持 
+
+**配置：**
+  * GPU: 8 NVIDIA Tesla V100 32G (32510MiB)
+  * BatchSize: 64/512
+  * SampleRatio: 0.1
+
+| Mode                      | Precision | Res50    | Res100   |
+| ------------------------- | --------- | -------- | -------- |
+| Framework1 (static)       | AMP       | 42000000 (31792MiB)| 39000000 (31938MiB)|
+| Framework2 (dynamic)      | AMP       | 30000000 (31702MiB)| 29000000 (32286MiB)|
+| Paddle (static)           | Pure FP16 | 60000000 (32018MiB)| 60000000 (32018MiB)|
+| Paddle (dynamic)          | Pure FP16 | 59000000 (31970MiB)| 59000000 (31970MiB)|
+
+* 注：在跑实验前配置环境变量 ``export FLAGS_allocator_strategy=naive_best_fit``
+
+<a name="吞吐量对比"></a>
+
+### 8.4 吞吐量对比
+
+**配置：**
+  * BatchSize: 128/1024
+  * SampleRatio: 0.1
+  * Datasets: MS1MV3
+  * V100: Driver Version: 450.80.02, CUDA Version: 11.0
+  * A100: Driver Version: 460.32.03, CUDA Version: 11.2
+  
+![insightface_throughtput](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/images/insightface_throughtput.png)
+
+更多实验结果可以参考 [PLSC](https://github.com/PaddlePaddle/PLSC)，PLSC (Paddle Large Scale Classification) 是 Paddle 官方开源的大规模分类库，支持单机 8 卡 NVIDIA V100 (32G) 训练 6000 千万类，目前还在持续更新中，请关注。
+
+<a name="全流程推理"></a>
+
+## 9. 全流程推理
+
+首先下载索引库、待识别图像与字体文件。
+
+```bash
+# 下载用于人脸识别的索引库，这里因为示例图像是老友记中的图像，所以使用老友记中角色的人脸图像构建的底库。
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/index.bin
+# 下载用于人脸识别的示例图像
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends2.jpg
+# 下载字体，用于可视化
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/SourceHanSansCN-Medium.otf
+```
+
+`检测+识别` 串联预测的示例脚本如下。
+
+```shell
+# 同时使用检测+识别
+python3.7 tools/test_recognition.py --det --rec --index=index.bin --input=friends2.jpg --output="./output"
+```
+
+最终可视化结果保存在`output`目录下，可视化结果如下所示。
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/output/friends2.jpg"  width = "800" />
+</div>
+
+更多关于参数解释，索引库构建、whl包预测部署和Paddle Serving预测部署的内容可以参考：
+ * [Whl包预测部署](https://github.com/littletomatodonkey/insight-face-paddle)
+ * [Paddle Serving预测部署](./deploy/pdserving/README_CN.md)
diff --git a/recognition/arcface_paddle/README_en.md b/recognition/arcface_paddle/README_en.md
new file mode 100644
index 000000000..6b38d9ae4
--- /dev/null
+++ b/recognition/arcface_paddle/README_en.md
@@ -0,0 +1,284 @@
+[简体中文](README_cn.md) | English
+
+# Arcface-Paddle
+
+* [1. Introduction](#Introduction)
+* [2. Environment Preparation](#Environment_Preparation)
+* [3. Data Preparation](#Data_Preparation)
+  * [3.1 Download Dataset](#Download_Dataset)
+  * [3.2 Extract MXNet Dataset to images](#Extract_MXNet_Dataset_to_images)
+* [4. How to Training](#How_to_Training)
+  * [4.1 Single Node, Single GPU](#Single_Node_Single_GPU)
+  * [4.2 Single Node, 8 GPUs](#Single_Node_8_GPU)
+* [5. Model Evaluation](#Model_Evaluation)
+* [6. Export Model](#Export_Model)
+* [7 Model Inference](#Model_Inference)
+* [8 Model Performance](#Model_Performance)
+  * [8.1 Performance of Lighting Model](#Performance_of_Lighting_Model)
+  * [8.2 Accuracy on Verification Datasets](#Accuracy_on_Verification_Datasets)
+  * [8.3 Maximum Number of Identities ](#Maximum_Number_of_Identities)
+  * [8.4 Throughtput](#Throughtput)
+* [9. Inference Combined with Face Detection Model](#Inference_Combined_with_Face_Detection_Model)
+
+
+<a name="Introduction"></a>
+
+## 1. Introduction
+
+`Arcface-Paddle` is an open source deep face detection and recognition toolkit, powered by PaddlePaddle. `Arcface-Paddle` provides three related pretrained models now, include `BlazeFace` for face detection, `ArcFace` and `MobileFace` for face recognition.
+
+- This tutorial is mainly about face recognition.
+- For face detection task, please refer to: [Face detection tuturial](../../detection/blazeface_paddle/README_en.md).
+- For Whl package inference using PaddleInference, please refer to [whl package inference](https://github.com/littletomatodonkey/insight-face-paddle).
+
+Note: Many thanks to [GuoQuanhao](https://github.com/GuoQuanhao) for the reproduction of the [Arcface basline using PaddlePaddle](https://github.com/GuoQuanhao/arcface-Paddle).
+
+<a name="Environment_Preparation"></a>
+
+## 2. Environment Preparation
+
+Please refer to [Installation](./install_en.md) to setup environment at first.
+
+<a name="Data_Preparation"></a>
+
+## 3. Data Preparation
+
+<a name="Download_Dataset"></a>
+
+### 3.1 Download Dataset
+
+Download the dataset from [insightface datasets](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_).
+
+* MS1M_v2: MS1M-ArcFace
+* MS1M_v3: MS1M-RetinaFace
+
+<a name="Extract_MXNet_Dataset_to_images"></a>
+
+### 3.2 Extract MXNet Dataset to images
+
+```shell
+python tools/mx_recordio_2_images.py --root_dir ms1m-retinaface-t1/ --output_dir MS1M_v3/
+```
+
+After finishing unzipping the dataset, the folder structure is as follows.
+
+```
+MS1M_v3
+|_ images
+|  |_ 00000001.jpg
+|  |_ ...
+|  |_ 05179510.jpg
+|_ label.txt
+|_ agedb_30.bin
+|_ cfp_ff.bin
+|_ cfp_fp.bin
+|_ lfw.bin
+```
+
+Label file format is as follows.
+
+```
+# delimiter: "\t"
+# the following the content of label.txt
+images/00000001.jpg 0
+...
+```
+
+If you want to use customed dataset, you can arrange your data according to the above format. 
+
+<a name="How_to_Training"></a>
+
+## 4. How to Training
+
+<a name="Single_Node_Single_GPU"></a>
+
+### 4.1 Single Node, Single GPU
+
+```bash
+export CUDA_VISIBLE_DEVICES=1
+python tools/train.py \
+    --config_file configs/ms1mv2_mobileface.py \
+    --embedding_size 128 \
+    --sample_ratio 1.0 \
+    --loss ArcFace \
+    --batch_size 512 \
+    --dataset MS1M_v2 \
+    --num_classes 85742 \
+    --data_dir MS1M_v2/ \
+    --label_file MS1M_v2/label.txt \
+    --fp16 False
+```
+
+<a name="Single_Node_8_GPU"></a>
+
+### 4.2 Single Node, 8 GPUs
+
+#### Static Mode
+
+```bash
+sh scripts/train_static.sh
+```
+
+#### Dynamic Mode
+
+```bash
+sh scripts/train_dynamic.sh
+```
+
+
+During training, you can view loss changes in real time through `VisualDL`,  For more information, please refer to [VisualDL](https://github.com/PaddlePaddle/VisualDL/).
+
+
+<a name="Model_Evaluation"></a>
+
+## 5. Model Evaluation
+
+The model evaluation process can be started as follows.
+
+#### Static Mode
+
+```bash
+sh scripts/validation_static.sh
+```
+
+#### Dynamic Mode
+
+```bash
+sh scripts/validation_dynamic.sh
+```
+
+<a name="Export_Model"></a>
+
+## 6. Export Model
+PaddlePaddle supports inference using prediction engines. Firstly, you should export inference model.
+
+#### Static Mode
+
+```bash
+sh scripts/export_static.sh
+```
+
+#### Dynamic Mode
+
+```bash
+sh scripts/export_dynamic.sh
+```
+
+We also support export to onnx model, you only need to set `--export_type onnx`.
+
+<a name="Model_Inference"></a>
+
+## 7. Model Inference
+
+The model inference process supports paddle save inference model and onnx model.
+
+```bash
+sh scripts/inference.sh
+```
+
+<a name="Model_Performance"></a>
+
+## 8. Model Performance
+
+<a name="Performance_of_Lighting_Model"></a>
+
+### 8.1 Performance of Lighting Model
+
+**Configuration：**
+  * CPU: Intel(R) Xeon(R) Gold 6184 CPU @ 2.40GHz
+  * GPU: a single NVIDIA Tesla V100
+  * Precison: FP32
+  * BatchSize: 64/512
+  * SampleRatio: 1.0
+  * Embedding Size: 128
+  * MS1MV2
+
+| Model structure           | lfw    | cfp_fp  | agedb30 | CPU time cost | GPU time cost | Inference model |
+| ------------------------- | ------ | ------- | ------- | -------| -------- |---- |
+| MobileFace-Paddle      | 0.9952 | 0.9280  | 0.9612  | 4.3ms  | 2.3ms    | [download link](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar)  |
+| MobileFace-mxnet       | 0.9950 | 0.8894  | 0.9591  | 7.3ms  | 4.7ms    | -   |
+
+* Note: MobileFace-Paddle training using MobileFaceNet_128
+
+<a name="Accuracy_on_Verification_Datasets"></a>
+
+### 8.2 Accuracy on Verification Datasets
+
+**Configuration：**
+  * GPU: 8 NVIDIA Tesla V100 32G
+  * Precison: Pure FP16
+  * BatchSize: 128/1024
+
+| Mode    | Datasets | backbone | Ratio | agedb30 | cfp_fp | lfw  | log  | checkpoint |
+| ------- | :------: | :------- | ----- | :------ | :----- | :--- | :--- |  :--- |
+| Static  |  MS1MV3  | r50      | 0.1   | 0.98317 | 0.98943| 0.99850 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/static/ms1mv3_r50_static_128_fp16_0.1/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_static_128_fp16_0.1_epoch_24.tgz) |
+| Static  |  MS1MV3  | r50      | 1.0   | 0.98283 | 0.98843| 0.99850 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/static/ms1mv3_r50_static_128_fp16_1.0/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_static_128_fp16_1.0_epoch_24.tgz) |
+| Dynamic |  MS1MV3  | r50      | 0.1   | 0.98333 | 0.98900| 0.99833 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/dynamic/ms1mv3_r50_dynamic_128_fp16_0.1/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_dynamic_128_fp16_0.1_eopch_24.tgz) |
+| Dynamic |  MS1MV3  | r50      | 1.0   | 0.98317 | 0.98900| 0.99833 | [log](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/logs/dynamic/ms1mv3_r50_dynamic_128_fp16_1.0/training.log) | [checkpoint](https://paddle-model-ecology.bj.bcebos.com/model/insight-face/distributed/ms1mv3_r50_dynamic_128_fp16_1.0_eopch_24.tgz) |
+
+
+<a name="Maximum_Number_of_Identities"></a>
+
+### 8.3 Maximum Number of Identities 
+
+**Configuration：**
+  * GPU: 8 NVIDIA Tesla V100 32G (32510MiB)
+  * BatchSize: 64/512
+  * SampleRatio: 0.1
+
+| Mode                      | Precision | Res50    | Res100   |
+| ------------------------- | --------- | -------- | -------- |
+| Framework1 (static)       | AMP       | 42000000 (31792MiB)| 39000000 (31938MiB)|
+| Framework2 (dynamic)      | AMP       | 30000000 (31702MiB)| 29000000 (32286MiB)|
+| Paddle (static)           | Pure FP16 | 60000000 (32018MiB)| 60000000 (32018MiB)|
+| Paddle (dynamic)          | Pure FP16 | 59000000 (31970MiB)| 59000000 (31970MiB)|
+
+**Note:** config environment variable by ``export FLAGS_allocator_strategy=naive_best_fit``
+
+<a name="Throughtput"></a>
+
+### 8.4 Throughtput
+
+**Configuration：**
+  * BatchSize: 128/1024
+  * SampleRatio: 0.1
+  * Datasets: MS1MV3
+  * V100: Driver Version: 450.80.02, CUDA Version: 11.0
+  * A100: Driver Version: 460.32.03, CUDA Version: 11.2
+  
+![insightface_throughtput](https://github.com/PaddlePaddle/PLSC/blob/master/experiments/arcface_paddle/images/insightface_throughtput.png)
+
+For more experimental results see [PLSC](https://github.com/PaddlePaddle/PLSC), which is an open source Paddle Large Scale Classification Tools powered by PaddlePaddle. It supports 60 million classes on single node 8 NVIDIA V100 (32G).
+
+<a name="Inference_Combined_with_Face_Detection_Model"></a>
+
+## 9. Inference Combined with Face Detection Model
+
+Firstly, use the following commands to download the index gallery, demo image and font file for visualization.
+
+
+```bash
+# Index library for the recognition process
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/index.bin
+# Demo image
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends2.jpg
+# Font file for visualization
+wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/SourceHanSansCN-Medium.otf
+```
+
+Use the following command to run the whole face recognition demo.
+
+```shell
+# detection + recogniotion process
+python3.7 tools/test_recognition.py --det --rec --index=index.bin --input=friends2.jpg --output="./output"
+```
+
+The final result is save in folder `output/`, which is shown as follows.
+
+<div align="center">
+<img src="https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/output/friends2.jpg"  width = "800" />
+</div>
+
+For more details about parameter explanations, index gallery construction and whl package inference, please refer to:
+ *  [Whl package inference tutorial](https://github.com/littletomatodonkey/insight-face-paddle).
+ * [Paddle Serving inference](./deploy/pdserving/README.md)
diff --git a/recognition/arcface_paddle/configs/__init__.py b/recognition/arcface_paddle/configs/__init__.py
new file mode 100644
index 000000000..185a92b8d
--- /dev/null
+++ b/recognition/arcface_paddle/configs/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/recognition/arcface_paddle/configs/argparser.py b/recognition/arcface_paddle/configs/argparser.py
new file mode 100644
index 000000000..88f40f7ad
--- /dev/null
+++ b/recognition/arcface_paddle/configs/argparser.py
@@ -0,0 +1,281 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import logging
+import argparse
+import importlib
+
+
+def print_args(args):
+    logging.info('--------args----------')
+    for k in list(vars(args).keys()):
+        logging.info('%s: %s' % (k, vars(args)[k]))
+    logging.info('------------------------\n')
+
+
+def str2bool(v):
+    return str(v).lower() in ("true", "t", "1")
+
+
+def tostrlist(v):
+    if isinstance(v, list):
+        return v
+    elif isinstance(v, str):
+        return [e.strip() for e in v.split(',')]
+
+
+def tointlist(v):
+    if isinstance(v, list):
+        return v
+    elif isinstance(v, str):
+        return [int(e.strip()) for e in v.split(',')]
+
+
+def get_config(config_file):
+    assert config_file.startswith(
+        'configs/'), 'config file setting must start with configs/'
+    temp_config_name = os.path.basename(config_file)
+    temp_module_name = os.path.splitext(temp_config_name)[0]
+    config = importlib.import_module("configs.config")
+    cfg = config.config
+    config = importlib.import_module("configs.%s" % temp_module_name)
+    job_cfg = config.config
+    cfg.update(job_cfg)
+    if cfg.output is None:
+        cfg.output = osp.join('work_dirs', temp_module_name)
+    return cfg
+
+
+class UserNamespace(object):
+    pass
+
+
+def parse_args():
+
+    parser = argparse.ArgumentParser(description='Paddle Face Training')
+    user_namespace = UserNamespace()
+    parser.add_argument(
+        '--config_file', type=str, required=True, help='config file path')
+    parser.parse_known_args(namespace=user_namespace)
+    cfg = get_config(user_namespace.config_file)
+
+    # Model setting
+    parser.add_argument(
+        '--is_static',
+        type=str2bool,
+        default=cfg.is_static,
+        help='whether to use static mode')
+    parser.add_argument(
+        '--backbone', type=str, default=cfg.backbone, help='backbone network')
+    parser.add_argument(
+        '--classifier',
+        type=str,
+        default=cfg.classifier,
+        help='classification network')
+    parser.add_argument(
+        '--embedding_size',
+        type=int,
+        default=cfg.embedding_size,
+        help='embedding size')
+    parser.add_argument(
+        '--model_parallel',
+        type=str2bool,
+        default=cfg.model_parallel,
+        help='whether to use model parallel')
+    parser.add_argument(
+        '--sample_ratio',
+        type=float,
+        default=cfg.sample_ratio,
+        help='sample rate, use partial fc sample if sample rate less than 1.0')
+    parser.add_argument(
+        '--loss', type=str, default=cfg.loss, help='loss function')
+    parser.add_argument(
+        '--dropout',
+        type=float,
+        default=cfg.dropout,
+        help='probability of dropout')
+
+    # AMP setting
+    parser.add_argument(
+        '--fp16',
+        type=str2bool,
+        default=cfg.fp16,
+        help='whether to use fp16 training')
+    parser.add_argument(
+        '--init_loss_scaling',
+        type=float,
+        default=cfg.init_loss_scaling,
+        help='The initial loss scaling factor.')
+    parser.add_argument(
+        '--max_loss_scaling',
+        type=float,
+        default=cfg.max_loss_scaling,
+        help='The maximum loss scaling factor.')
+    parser.add_argument(
+        '--incr_every_n_steps',
+        type=int,
+        default=cfg.incr_every_n_steps,
+        help='Increases loss scaling every n consecutive steps with finite gradients.'
+    )
+    parser.add_argument(
+        '--decr_every_n_nan_or_inf',
+        type=int,
+        default=cfg.decr_every_n_nan_or_inf,
+        help='Decreases loss scaling every n accumulated steps with nan or inf gradients.'
+    )
+    parser.add_argument(
+        '--incr_ratio',
+        type=float,
+        default=cfg.incr_ratio,
+        help='The multiplier to use when increasing the loss scaling.')
+    parser.add_argument(
+        '--decr_ratio',
+        type=float,
+        default=cfg.decr_ratio,
+        help='The less-than-one-multiplier to use when decreasing the loss scaling.'
+    )
+    parser.add_argument(
+        '--use_dynamic_loss_scaling',
+        type=str2bool,
+        default=cfg.use_dynamic_loss_scaling,
+        help='Whether to use dynamic loss scaling.')
+    parser.add_argument(
+        '--custom_white_list',
+        type=tostrlist,
+        default=cfg.custom_white_list,
+        help='fp16 custom white list.')
+    parser.add_argument(
+        '--custom_black_list',
+        type=tostrlist,
+        default=cfg.custom_black_list,
+        help='fp16 custom black list.')
+
+    # Optimizer setting
+    parser.add_argument(
+        '--lr', type=float, default=cfg.lr, help='learning rate')
+    parser.add_argument(
+        '--lr_decay',
+        type=float,
+        default=cfg.lr_decay,
+        help='learning rate decay factor')
+    parser.add_argument(
+        '--weight_decay',
+        type=float,
+        default=cfg.weight_decay,
+        help='weight decay')
+    parser.add_argument(
+        '--momentum', type=float, default=cfg.momentum, help='sgd momentum')
+    parser.add_argument(
+        '--train_unit',
+        type=str,
+        default=cfg.train_unit,
+        help='train unit, "step" or "epoch"')
+    parser.add_argument(
+        '--warmup_num',
+        type=int,
+        default=cfg.warmup_num,
+        help='warmup num according train unit')
+    parser.add_argument(
+        '--train_num',
+        type=int,
+        default=cfg.train_num,
+        help='train num according train unit')
+    parser.add_argument(
+        '--decay_boundaries',
+        type=tointlist,
+        default=cfg.decay_boundaries,
+        help='piecewise decay boundaries')
+
+    # Train dataset setting
+    parser.add_argument(
+        '--use_synthetic_dataset',
+        type=str2bool,
+        default=cfg.use_synthetic_dataset,
+        help='whether to use synthetic dataset')
+    parser.add_argument(
+        '--dataset', type=str, default=cfg.dataset, help='train dataset name')
+    parser.add_argument(
+        '--data_dir',
+        type=str,
+        default=cfg.data_dir,
+        help='train dataset directory')
+    parser.add_argument(
+        '--label_file',
+        type=str,
+        default=cfg.label_file,
+        help='train label file name, each line split by "\t"')
+    parser.add_argument(
+        '--is_bin',
+        type=str2bool,
+        default=cfg.is_bin,
+        help='whether the train data is bin or original image file')
+    parser.add_argument(
+        '--num_classes',
+        type=int,
+        default=cfg.num_classes,
+        help='classes of train dataset')
+    parser.add_argument(
+        '--batch_size',
+        type=int,
+        default=cfg.batch_size,
+        help='batch size of each rank')
+    parser.add_argument(
+        '--num_workers',
+        type=int,
+        default=cfg.num_workers,
+        help='the number workers of DataLoader')
+
+    # Validation dataset setting
+    parser.add_argument(
+        '--do_validation_while_train',
+        type=str2bool,
+        default=cfg.do_validation_while_train,
+        help='do validation while train')
+    parser.add_argument(
+        '--validation_interval_step',
+        type=int,
+        default=cfg.validation_interval_step,
+        help='validation interval step')
+    parser.add_argument(
+        '--val_targets',
+        type=tostrlist,
+        default=cfg.val_targets,
+        help='val targets, list or str split by comma')
+
+    # IO setting
+    parser.add_argument(
+        '--logdir', type=str, default=cfg.logdir, help='log dir')
+    parser.add_argument(
+        '--log_interval_step',
+        type=int,
+        default=cfg.log_interval_step,
+        help='log interval step')
+    parser.add_argument(
+        '--output', type=str, default=cfg.output, help='output dir')
+    parser.add_argument(
+        '--resume', type=str2bool, default=cfg.resume, help='whether to using resume training')
+    parser.add_argument(
+        '--checkpoint_dir',
+        type=str,
+        default=cfg.checkpoint_dir,
+        help='set checkpoint direcotry when resume training')
+    parser.add_argument(
+        '--max_num_last_checkpoint',
+        type=int,
+        default=cfg.max_num_last_checkpoint,
+        help='the maximum number of lastest checkpoint to keep')
+
+    args = parser.parse_args(namespace=user_namespace)
+    return args
diff --git a/recognition/arcface_paddle/configs/config.py b/recognition/arcface_paddle/configs/config.py
new file mode 100644
index 000000000..efd126fb7
--- /dev/null
+++ b/recognition/arcface_paddle/configs/config.py
@@ -0,0 +1,65 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from easydict import EasyDict as edict
+
+config = edict()
+config.is_static = True
+config.backbone = 'FresResNet100'
+config.classifier = 'LargeScaleClassifier'
+config.embedding_size = 512
+config.model_parallel = True
+config.sample_ratio = 0.1
+config.loss = 'ArcFace'
+config.dropout = 0.0
+
+config.fp16 = True
+config.init_loss_scaling = 128.0
+config.max_loss_scaling = 128.0
+config.incr_every_n_steps = 2000
+config.decr_every_n_nan_or_inf = 1
+config.incr_ratio = 2.0
+config.decr_ratio = 0.5
+config.use_dynamic_loss_scaling = True
+config.custom_white_list = []
+config.custom_black_list = []
+
+config.lr = 0.1  # for global batch size = 512
+config.lr_decay = 0.1
+config.weight_decay = 5e-4
+config.momentum = 0.9
+config.train_unit = 'step'  # 'step' or 'epoch'
+config.warmup_num = 1000
+config.train_num = 180000
+config.decay_boundaries = [100000, 140000, 160000]
+
+config.use_synthetic_dataset = False
+config.dataset = "MS1M_v3"
+config.data_dir = "./MS1M_v3"
+config.label_file = "./MS1M_v3/label.txt"
+config.is_bin = False
+config.num_classes = 93431  # 85742 for MS1M_v2, 93431 for MS1M_v3
+config.batch_size = 64  # global batch size 512 of 8 GPU
+config.num_workers = 8
+
+config.do_validation_while_train = True
+config.validation_interval_step = 2000
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+config.logdir = './log'
+config.log_interval_step = 10
+config.output = './MS1M_v3_arcface'
+config.resume = False
+config.checkpoint_dir = None
+config.max_num_last_checkpoint = 3
diff --git a/recognition/arcface_paddle/configs/ms1mv2_mobileface.py b/recognition/arcface_paddle/configs/ms1mv2_mobileface.py
new file mode 100644
index 000000000..e29a062a5
--- /dev/null
+++ b/recognition/arcface_paddle/configs/ms1mv2_mobileface.py
@@ -0,0 +1,54 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from easydict import EasyDict as edict
+
+config = edict()
+config.is_static = False
+config.backbone = 'MobileFaceNet_128'
+config.classifier = 'LargeScaleClassifier'
+config.embedding_size = 128
+config.model_parallel = True
+config.sample_ratio = 1.0
+config.loss = 'ArcFace'
+config.dropout = 0.0
+
+config.lr = 0.1  # for global batch size = 512
+config.lr_decay = 0.1
+config.weight_decay = 5e-4
+config.momentum = 0.9
+config.train_unit = 'epoch'  # 'step' or 'epoch'
+config.warmup_num = 0
+config.train_num = 25
+config.decay_boundaries = [10, 16, 22]
+
+config.use_synthetic_dataset = False
+config.dataset = "MS1M_v2"
+config.data_dir = "./MS1M_v2"
+config.label_file = "./MS1M_v2/label.txt"
+config.is_bin = False
+config.num_classes = 85742  # 85742 for MS1M_v2, 93431 for MS1M_v3
+config.batch_size = 128  # global batch size 1024 of 8 GPU
+config.num_workers = 8
+
+config.do_validation_while_train = True
+config.validation_interval_step = 2000
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+config.logdir = './log'
+config.log_interval_step = 100
+config.output = './MS1M_v2_arcface_MobileFaceNet_128_0.1'
+config.resume = False
+config.checkpoint_dir = None
+config.max_num_last_checkpoint = 1
diff --git a/recognition/arcface_paddle/configs/ms1mv3_r100.py b/recognition/arcface_paddle/configs/ms1mv3_r100.py
new file mode 100644
index 000000000..75d200d9b
--- /dev/null
+++ b/recognition/arcface_paddle/configs/ms1mv3_r100.py
@@ -0,0 +1,54 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from easydict import EasyDict as edict
+
+config = edict()
+config.is_static = True
+config.backbone = 'FresResNet100'
+config.classifier = 'LargeScaleClassifier'
+config.embedding_size = 512
+config.model_parallel = True
+config.sample_ratio = 0.1
+config.loss = 'ArcFace'
+config.dropout = 0.0
+
+config.lr = 0.1  # for global batch size = 512
+config.lr_decay = 0.1
+config.weight_decay = 5e-4
+config.momentum = 0.9
+config.train_unit = 'epoch'  # 'step' or 'epoch'
+config.warmup_num = 0
+config.train_num = 25
+config.decay_boundaries = [10, 16, 22]
+
+config.use_synthetic_dataset = False
+config.dataset = "MS1M_v3"
+config.data_dir = "./MS1M_v3"
+config.label_file = "./MS1M_v3/label.txt"
+config.is_bin = False
+config.num_classes = 93431  # 85742 for MS1M_v2, 93431 for MS1M_v3
+config.batch_size = 128  # global batch size 512 of 8 GPU
+config.num_workers = 8
+
+config.do_validation_while_train = True
+config.validation_interval_step = 2000
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+config.logdir = './log'
+config.log_interval_step = 100
+config.output = './MS1M_v3_arcface'
+config.resume = False
+config.checkpoint_dir = None
+config.max_num_last_checkpoint = 1
diff --git a/recognition/arcface_paddle/configs/ms1mv3_r50.py b/recognition/arcface_paddle/configs/ms1mv3_r50.py
new file mode 100644
index 000000000..e5f556c48
--- /dev/null
+++ b/recognition/arcface_paddle/configs/ms1mv3_r50.py
@@ -0,0 +1,54 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from easydict import EasyDict as edict
+
+config = edict()
+config.is_static = True
+config.backbone = 'FresResNet50'
+config.classifier = 'LargeScaleClassifier'
+config.embedding_size = 512
+config.model_parallel = True
+config.sample_ratio = 0.1
+config.loss = 'ArcFace'
+config.dropout = 0.0
+
+config.lr = 0.1  # for global batch size = 512
+config.lr_decay = 0.1
+config.weight_decay = 5e-4
+config.momentum = 0.9
+config.train_unit = 'epoch'  # 'step' or 'epoch'
+config.warmup_num = 0
+config.train_num = 25
+config.decay_boundaries = [10, 16, 22]
+
+config.use_synthetic_dataset = False
+config.dataset = "MS1M_v3"
+config.data_dir = "./MS1M_v3"
+config.label_file = "./MS1M_v3/label.txt"
+config.is_bin = False
+config.num_classes = 93431  # 85742 for MS1M_v2, 93431 for MS1M_v3
+config.batch_size = 128  # global batch size 512 of 8 GPU
+config.num_workers = 8
+
+config.do_validation_while_train = True
+config.validation_interval_step = 2000
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+config.logdir = './log'
+config.log_interval_step = 100
+config.output = './MS1M_v3_arcface'
+config.resume = False
+config.checkpoint_dir = None
+config.max_num_last_checkpoint = 1
diff --git a/recognition/arcface_paddle/datasets/__init__.py b/recognition/arcface_paddle/datasets/__init__.py
new file mode 100644
index 000000000..c97f8e476
--- /dev/null
+++ b/recognition/arcface_paddle/datasets/__init__.py
@@ -0,0 +1,15 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .common_dataset import CommonDataset, SyntheticDataset, load_bin
diff --git a/recognition/arcface_paddle/datasets/common_dataset.py b/recognition/arcface_paddle/datasets/common_dataset.py
new file mode 100644
index 000000000..e1d5a151a
--- /dev/null
+++ b/recognition/arcface_paddle/datasets/common_dataset.py
@@ -0,0 +1,134 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pickle
+import paddle
+import os
+import cv2
+import six
+import random
+import paddle
+import numpy as np
+import logging
+from PIL import Image
+from io import BytesIO
+
+from datasets.kv_helper import read_img_from_bin
+
+
+def transform(img):
+    # random horizontal flip
+    if random.randint(0, 1) == 0:
+        img = cv2.flip(img, 1)
+    # normalize to mean 0.5, std 0.5
+    img = (img - 127.5) * 0.00784313725
+    # BGR2RGB
+    img = img[:, :, ::-1]
+    img = img.transpose((2, 0, 1))
+    return img
+
+
+class CommonDataset(paddle.io.Dataset):
+    def __init__(self, root_dir, label_file, fp16=False, is_bin=True):
+        super(CommonDataset, self).__init__()
+        self.root_dir = root_dir
+        self.label_file = label_file
+        self.fp16 = fp16
+        with open(label_file, "r") as fin:
+            self.full_lines = fin.readlines()
+
+        self.delimiter = "\t"
+        self.is_bin = is_bin
+
+        self.num_samples = len(self.full_lines)
+        logging.info("read label file finished, total num: {}"
+                     .format(self.num_samples))
+
+    def __getitem__(self, idx):
+
+        line = self.full_lines[idx]
+
+        img_path, label = line.strip().split(self.delimiter)
+        img_path = os.path.join(self.root_dir, img_path)
+        if self.is_bin:
+            img = read_img_from_bin(img_path)
+        else:
+            img = cv2.imread(img_path)
+
+        img = transform(img)
+
+        img = img.astype('float16' if self.fp16 else 'float32')
+        label = np.int32(label)
+
+        return img, label
+
+    def __len__(self):
+        return self.num_samples
+
+
+class SyntheticDataset(paddle.io.Dataset):
+    def __init__(self, num_classes, fp16=False):
+        super(SyntheticDataset, self).__init__()
+        self.num_classes = num_classes
+        self.fp16 = fp16
+        self.label_list = np.random.randint(
+            0, num_classes, (5179510, ), dtype=np.int32)
+        self.num_samples = len(self.label_list)
+
+    def __getitem__(self, idx):
+        label = self.label_list[idx]
+        img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.uint8)
+        img = transform(img)
+
+        img = img.astype('float16' if self.fp16 else 'float32')
+        label = np.int32(label)
+
+        return img, label
+
+    def __len__(self):
+        return self.num_samples
+
+
+# 返回为 numpy
+def load_bin(path, image_size):
+    if six.PY2:
+        bins, issame_list = pickle.load(open(path, 'rb'))
+    else:
+        bins, issame_list = pickle.load(open(path, 'rb'), encoding='bytes')
+    data_list = []
+    for flip in [0, 1]:
+        data = np.empty(
+            (len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for i in range(len(issame_list) * 2):
+        _bin = bins[i]
+        if six.PY2:
+            if not isinstance(_bin, six.string_types):
+                _bin = _bin.tostring()
+            img_ori = Image.open(StringIO(_bin))
+        else:
+            img_ori = Image.open(BytesIO(_bin))
+        for flip in [0, 1]:
+            img = img_ori.copy()
+            if flip == 1:
+                img = img.transpose(Image.FLIP_LEFT_RIGHT)
+            if img.mode != 'RGB':
+                img = img.convert('RGB')
+            img = np.array(img).astype('float32').transpose((2, 0, 1))
+            img = (img - 127.5) * 0.00784313725
+            data_list[flip][i][:] = img
+        if i % 1000 == 0:
+            print('loading bin', i)
+    print(data_list[0].shape)
+    return data_list, issame_list
diff --git a/recognition/arcface_paddle/datasets/kv_helper.py b/recognition/arcface_paddle/datasets/kv_helper.py
new file mode 100644
index 000000000..43bff7cd2
--- /dev/null
+++ b/recognition/arcface_paddle/datasets/kv_helper.py
@@ -0,0 +1,68 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import pickle
+import struct
+import random
+import multiprocessing
+import numpy as np
+import cv2
+import json
+
+
+def readkv(f):
+    """readkv"""
+    keylendata = f.read(4)
+    if len(keylendata) != 4:
+        return None
+    keylen = struct.unpack('I', keylendata)[0]
+    if keylen > 5000:
+        raise Exception('wrong key len' + str(keylen))
+    key = f.read(keylen)
+    valuelen = struct.unpack('I', f.read(4))[0]
+    value = f.read(valuelen)
+    return key, value
+
+
+def writekv(f, k, v, flush=True):
+    """writekv"""
+    f.write(struct.pack('I', len(k)))
+    f.write(k)
+    f.write(struct.pack('I', len(v)))
+    f.write(v)
+    if flush:
+        f.flush()
+    return
+
+
+def trans_img_to_bin(img_name, output_path):
+    with open(img_name, "rb") as fin:
+        img = fin.read()
+    key = os.path.split(img_name)[-1]
+    with open(output_path, "wb") as fout:
+        writekv(fout, key.encode(), pickle.dumps(img, -1))
+    return
+
+
+def read_img_from_bin(input_path):
+    # the file can exist many key-vals, but it just save one in fact.
+    with open(input_path, "rb") as fin:
+        r = readkv(fin)
+        assert r is not None
+        _, value = r
+        value = pickle.loads(value)
+        value = np.frombuffer(value, dtype='uint8')
+        img = cv2.imdecode(value, 1)
+    return img
diff --git a/recognition/arcface_paddle/deploy/pdserving/README.md b/recognition/arcface_paddle/deploy/pdserving/README.md
new file mode 100644
index 000000000..946478012
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/README.md
@@ -0,0 +1,180 @@
+
+# Service deployment based on PaddleServing  
+
+(English|[简体中文](./README_CN.md))
+
+
+This document will introduce how to use the [PaddleServing](https://github.com/PaddlePaddle/Serving/blob/develop/README.md) to deploy the Arcface dynamic graph model as a pipeline online service.
+
+Some Key Features of Paddle Serving:
+- Integrate with Paddle training pipeline seamlessly, most paddle models can be deployed with one line command.
+- Industrial serving features supported, such as models management, online loading, online A/B testing etc.
+- Highly concurrent and efficient communication between clients and servers supported.
+
+The introduction and tutorial of Paddle Serving service deployment framework reference [document](https://github.com/PaddlePaddle/Serving/blob/develop/README.md).
+
+
+## Contents
+- [Environmental preparation](#environmental-preparation)
+- [Model conversion](#model-conversion)
+- [Paddle Serving pipeline deployment](#paddle-serving-pipeline-deployment)
+- [FAQ](#faq)
+
+<a name="environmental-preparation"></a>
+## Environmental preparation
+
+Arcface operating environment and Paddle Serving operating environment are needed.
+
+1. Please prepare Arcface operating environment reference [link](../../README_en.md).
+   Download the corresponding paddle whl package according to the environment, it is recommended to install version 2.2+.
+
+
+2. The steps of PaddleServing operating environment prepare are as follows:
+
+    Install serving which used to start the service
+    ```
+    pip3 install paddle-serving-server==0.6.3 # for CPU
+    pip3 install paddle-serving-server-gpu==0.6.3 # for GPU
+    # Other GPU environments need to confirm the environment and then choose to execute the following commands
+    pip3 install paddle-serving-server-gpu==0.6.3.post101 # GPU with CUDA10.1 + TensorRT6
+    pip3 install paddle-serving-server-gpu==0.6.3.post11 # GPU with CUDA11 + TensorRT7
+    ```
+
+3. Install the client to send requests to the service
+    In [download link](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md) find the client installation package corresponding to the python version.
+    The python3.7 version is recommended here:
+
+    ```
+    pip3 install paddle-serving-client==0.6.3
+    ```
+
+4. Install serving-app
+    ```
+    pip3 install paddle-serving-app==0.6.3
+    ```
+
+   **note:** If you want to install the latest version of PaddleServing, refer to [link](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md).
+
+
+<a name="model-conversion"></a>
+## Model conversion
+When using PaddleServing for service deployment, you need to convert the saved inference model into a serving model that is easy to deploy.
+
+Firstly, download the inference model of Arcface
+```
+wget -nc -P ./inference https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar
+tar xf inference/mobileface_v1.0_infer.tar --strip-components 1 -C inference 
+```
+Then, you can use installed paddle_serving_client tool to convert inference model to mobile model.
+```
+python3 -m paddle_serving_client.convert --dirname ./inference/ \
+                                         --model_filename inference.pdmodel \
+                                         --params_filename inference.pdiparams \
+                                         --serving_server ./MobileFaceNet_128_serving/ \
+                                         --serving_client ./MobileFaceNet_128_client/
+
+```
+
+After the detection model is converted, there will be additional folders of `MobileFaceNet_128_serving` and `MobileFaceNet_128_client` in the current folder, with the following format:
+```
+MobileFaceNet_128_serving
+├── __model__
+├── __params__
+├── serving_server_conf.prototxt
+└── serving_server_conf.stream.prototxt
+
+MobileFaceNet_128_client/
+├── serving_client_conf.prototxt
+└── serving_client_conf.stream.prototxt
+
+```
+The recognition model is the same.
+
+<a name="paddle-serving-pipeline-deployment"></a>
+## Paddle Serving pipeline deployment
+
+1. Download the PaddleOCR code, if you have already downloaded it, you can skip this step.
+    ```
+    git clone https://github.com/deepinsight/insightface
+
+    # Enter the working directory  
+    cd recognition/arcface_paddle/deploy/pdserving
+    ```
+
+    The pdserver directory contains the code to start the pipeline service and send prediction requests, including:
+    ```
+    __init__.py
+    config.yml # Start the service configuration file
+    ocr_reader.py # pre-processing and post-processing code implementation
+    pipeline_http_client.py # Script to send pipeline prediction request
+    web_service.py # Start the script of the pipeline server
+    ```
+
+2. Run the following command to start the service.
+    ```
+    # Start the service and save the running log in log.txt
+    python3 web_service.py &>log.txt &
+    ```
+    After the service is successfully started, a log similar to the following will be printed in log.txt
+    ![](./imgs/start_server.png)
+
+3. Send service request
+    ```
+    python3 pipeline_http_client.py
+    ```
+    After successfully running, the predicted result of the model will be printed in the cmd window. An example of the result is:
+    ![](./imgs/results.png)  
+
+    Adjust the number of concurrency in config.yml to get the largest QPS. Generally, the number of concurrent detection and recognition is 2:1
+
+    ```
+    det:
+        concurrency: 8
+        ...
+    rec:
+        concurrency: 4
+        ...
+    ```
+
+    Multiple service requests can be sent at the same time if necessary.
+
+    The predicted performance data will be automatically written into the `PipelineServingLogs/pipeline.tracer` file.
+
+    Tested on 700 real picture. The average QPS on V100 GPU can reach around 57:
+
+    ```
+    2021-11-04 13:38:52,507 Op(ArcFace):
+    2021-11-04 13:38:52,507 	in[135.4579597902098 ms]
+    2021-11-04 13:38:52,507 	prep[0.9921311188811189 ms]
+    2021-11-04 13:38:52,507 	midp[3.9232132867132865 ms]
+    2021-11-04 13:38:52,507 	postp[0.12166258741258741 ms]
+    2021-11-04 13:38:52,507 	out[0.9898286713286714 ms]
+    2021-11-04 13:38:52,508 	idle[0.9643989520087675]
+    2021-11-04 13:38:52,508 DAGExecutor:
+    2021-11-04 13:38:52,508 	Query count[573]
+    2021-11-04 13:38:52,508 	QPS[57.3 q/s]
+    2021-11-04 13:38:52,509 	Succ[0.9982547993019197]
+    2021-11-04 13:38:52,509 	Error req[394]
+    2021-11-04 13:38:52,509 	Latency:
+    2021-11-04 13:38:52,509 		ave[11.52941186736475 ms]
+    2021-11-04 13:38:52,509 		.50[11.492 ms]
+    2021-11-04 13:38:52,509 		.60[11.658 ms]
+    2021-11-04 13:38:52,509 		.70[11.95 ms]
+    2021-11-04 13:38:52,509 		.80[12.251 ms]
+    2021-11-04 13:38:52,509 		.90[12.736 ms]
+    2021-11-04 13:38:52,509 		.95[13.21 ms]
+    2021-11-04 13:38:52,509 		.99[13.987 ms]
+    2021-11-04 13:38:52,510 Channel (server worker num[10]):
+    2021-11-04 13:38:52,510 	chl0(In: ['@DAGExecutor'], Out: ['ArcFace']) size[0/0]
+    2021-11-04 13:38:52,510 	chl1(In: ['ArcFace'], Out: ['@DAGExecutor']) size[0/0]
+    ```
+
+<a name="faq"></a>
+## FAQ
+**Q1**: No result return after sending the request.
+
+**A1**: Do not set the proxy when starting the service and sending the request. You can close the proxy before starting the service and before sending the request. The command to close the proxy is:
+```
+unset https_proxy
+unset http_proxy
+```  
diff --git a/recognition/arcface_paddle/deploy/pdserving/README_CN.md b/recognition/arcface_paddle/deploy/pdserving/README_CN.md
new file mode 100644
index 000000000..77e0d0837
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/README_CN.md
@@ -0,0 +1,172 @@
+# 基于PaddleServing的服务部署
+
+(简体中文|[English](./README.md))
+
+
+本文档将介绍如何使用[PaddleServing](https://github.com/PaddlePaddle/Serving/blob/develop/README_CN.md)工具部署 Arcface 动态图模型的pipeline在线服务。
+
+PaddleServing具备以下优点：
+- 支持客户端和服务端之间高并发和高效通信
+- 支持 工业级的服务能力 例如模型管理，在线加载，在线A/B测试等
+- 支持 多种编程语言 开发客户端，例如C++, Python和Java
+
+更多有关PaddleServing服务化部署框架介绍和使用教程参考[文档](https://github.com/PaddlePaddle/Serving/blob/develop/README_CN.md)。
+
+## 目录
+- [环境准备](#环境准备)
+- [模型转换](#模型转换)
+- [Paddle Serving pipeline部署](#部署)
+- [FAQ](#FAQ)
+
+<a name="环境准备"></a>
+## 环境准备
+
+需要准备 Arcface 的运行环境和Paddle Serving的运行环境。
+
+- 准备 Arcface 的运行环境[链接](../../README_cn.md)
+  根据环境下载对应的paddle whl包，推荐安装2.2+版本
+
+- 准备PaddleServing的运行环境，步骤如下
+
+1. 安装serving，用于启动服务
+    ```
+    pip3 install paddle-serving-server==0.6.3 # for CPU
+    pip3 install paddle-serving-server-gpu==0.6.3 # for GPU
+    # 其他GPU环境需要确认环境再选择执行如下命令
+    pip3 install paddle-serving-server-gpu==0.6.3.post101 # GPU with CUDA10.1 + TensorRT6
+    pip3 install paddle-serving-server-gpu==0.6.3.post11 # GPU with CUDA11 + TensorRT7
+    ```
+
+2. 安装client，用于向服务发送请求
+    ```
+    pip3 install paddle_serving_client==0.6.3
+    ```
+
+3. 安装serving-app
+    ```
+    pip3 install paddle-serving-app==0.6.3
+    ```
+
+    **Note:** 如果要安装最新版本的PaddleServing参考[链接](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md)。
+
+<a name="模型转换"></a>
+## 模型转换
+
+使用PaddleServing做服务化部署时，需要将保存的inference模型转换为serving易于部署的模型。
+
+首先，下载Arcface的inference模型
+```
+# 下载并解压 Arcface 模型
+wget -nc -P ./inference https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar
+tar xf inference/mobileface_v1.0_infer.tar --strip-components 1 -C inference 
+```
+
+接下来，用安装的paddle_serving_client把下载的inference模型转换成易于server部署的模型格式。
+
+```
+python3 -m paddle_serving_client.convert --dirname ./inference/ \
+                                         --model_filename inference.pdmodel \
+                                         --params_filename inference.pdiparams \
+                                         --serving_server ./MobileFaceNet_128_serving/ \
+                                         --serving_client ./MobileFaceNet_128_client/
+```
+
+检测模型转换完成后，会在当前文件夹多出`MobileFaceNet_128_serving/` 和`MobileFaceNet_128_client`的文件夹，具备如下格式：
+```
+MobileFaceNet_128_serving
+├── __model__
+├── __params__
+├── serving_server_conf.prototxt
+└── serving_server_conf.stream.prototxt
+
+MobileFaceNet_128_client/
+├── serving_client_conf.prototxt
+└── serving_client_conf.stream.prototxt
+
+```
+
+<a name="部署"></a>
+## Paddle Serving pipeline部署
+
+1. 下载insightface代码，若已下载可跳过此步骤
+    ```
+    git clone https://github.com/deepinsight/insightface
+
+    # 进入到工作目录
+    cd recognition/arcface_paddle/deploy/pdserving
+    ```
+
+    pdserving目录包含启动pipeline服务和发送预测请求的代码，包括：
+    
+    ```
+    __init__.py
+    config.yml            # 启动服务的配置文件
+    pipeline_http_client.py   # web方式发送pipeline预测请求的脚本
+    pipeline_rpc_client.py   # rpc方式发送pipeline预测请求的脚本
+    web_service.py        # 启动pipeline服务端的脚本
+    ```
+
+2. 启动服务可运行如下命令：
+    ```
+    # 启动服务，运行日志保存在log.txt
+    python3 web_service.py &>log.txt &
+    ```
+    成功启动服务后，log.txt中会打印类似如下日志
+    ![](./imgs/start_server.png)
+
+3. 发送服务请求：
+    ```
+    python3 pipeline_http_client.py
+    ```
+    成功运行后，模型预测的结果会打印在cmd窗口中，结果示例为：
+    ![](./imgs/results.png)
+
+    调整 config.yml 中的并发个数获得最大的QPS, 一般检测和识别的并发数为2：1
+    ```
+    ArcFace:
+        #并发数，is_thread_op=True时，为线程并发；否则为进程并发
+        concurrency: 8
+        ...
+    ```
+    有需要的话可以同时发送多个服务请求
+
+    预测性能数据会被自动写入 `PipelineServingLogs/pipeline.tracer` 文件中。
+
+    在700张真实图片上测试，V100 GPU 上 QPS 均值可达到57左右：
+
+    ```
+    2021-11-04 13:38:52,507 Op(ArcFace):
+    2021-11-04 13:38:52,507 	in[135.4579597902098 ms]
+    2021-11-04 13:38:52,507 	prep[0.9921311188811189 ms]
+    2021-11-04 13:38:52,507 	midp[3.9232132867132865 ms]
+    2021-11-04 13:38:52,507 	postp[0.12166258741258741 ms]
+    2021-11-04 13:38:52,507 	out[0.9898286713286714 ms]
+    2021-11-04 13:38:52,508 	idle[0.9643989520087675]
+    2021-11-04 13:38:52,508 DAGExecutor:
+    2021-11-04 13:38:52,508 	Query count[573]
+    2021-11-04 13:38:52,508 	QPS[57.3 q/s]
+    2021-11-04 13:38:52,509 	Succ[0.9982547993019197]
+    2021-11-04 13:38:52,509 	Error req[394]
+    2021-11-04 13:38:52,509 	Latency:
+    2021-11-04 13:38:52,509 		ave[11.52941186736475 ms]
+    2021-11-04 13:38:52,509 		.50[11.492 ms]
+    2021-11-04 13:38:52,509 		.60[11.658 ms]
+    2021-11-04 13:38:52,509 		.70[11.95 ms]
+    2021-11-04 13:38:52,509 		.80[12.251 ms]
+    2021-11-04 13:38:52,509 		.90[12.736 ms]
+    2021-11-04 13:38:52,509 		.95[13.21 ms]
+    2021-11-04 13:38:52,509 		.99[13.987 ms]
+    2021-11-04 13:38:52,510 Channel (server worker num[10]):
+    2021-11-04 13:38:52,510 	chl0(In: ['@DAGExecutor'], Out: ['ArcFace']) size[0/0]
+    2021-11-04 13:38:52,510 	chl1(In: ['ArcFace'], Out: ['@DAGExecutor']) size[0/0]
+    ```
+
+<a name="FAQ"></a>
+## FAQ
+**Q1**： 发送请求后没有结果返回或者提示输出解码报错
+
+**A1**： 启动服务和发送请求时不要设置代理，可以在启动服务前和发送请求前关闭代理，关闭代理的命令是：
+```
+unset https_proxy
+unset http_proxy
+```
diff --git a/recognition/arcface_paddle/deploy/pdserving/__init__.py b/recognition/arcface_paddle/deploy/pdserving/__init__.py
new file mode 100644
index 000000000..61d5aa213
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
\ No newline at end of file
diff --git a/recognition/arcface_paddle/deploy/pdserving/config.yml b/recognition/arcface_paddle/deploy/pdserving/config.yml
new file mode 100644
index 000000000..4ef3dd9fb
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/config.yml
@@ -0,0 +1,45 @@
+#rpc端口, rpc_port和http_port不允许同时为空。当rpc_port为空且http_port不为空时，会自动将rpc_port设置为http_port+1
+rpc_port: 18091
+
+#http端口, rpc_port和http_port不允许同时为空。当rpc_port可用且http_port为空时，不自动生成http_port
+http_port: 9998
+
+#worker_num, 最大并发数。当build_dag_each_worker=True时, 框架会创建worker_num个进程，每个进程内构建grpcSever和DAG
+##当build_dag_each_worker=False时，框架会设置主线程grpc线程池的max_workers=worker_num
+worker_num: 10
+
+#build_dag_each_worker, False，框架在进程内创建一条DAG；True，框架会每个进程内创建多个独立的DAG
+build_dag_each_worker: False
+
+dag:
+    #op资源类型, True, 为线程模型；False，为进程模型
+    is_thread_op: False
+
+    #重试次数
+    retry: 10
+
+    #使用性能分析, True，生成Timeline性能数据，对性能有一定影响；False为不使用
+    use_profile: True
+    
+    tracer:
+        interval_s: 10
+op:
+    ArcFace:
+        #并发数，is_thread_op=True时，为线程并发；否则为进程并发
+        concurrency: 8
+
+        #当op配置没有server_endpoints时，从local_service_conf读取本地服务配置
+        local_service_conf:
+            #client类型，包括brpc, grpc和local_predictor.local_predictor不启动Serving服务，进程内预测
+            client_type: local_predictor
+
+            #模型路径
+            model_config: ./MobileFaceNet_128_serving
+
+            #Fetch结果列表，以client_config中fetch_var的alias_name为准
+            fetch_list: ["save_infer_model/scale_0.tmp_1"]
+
+            #计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
+            devices: "0"
+
+            ir_optim: True
\ No newline at end of file
diff --git a/recognition/arcface_paddle/deploy/pdserving/imgs/demo.jpg b/recognition/arcface_paddle/deploy/pdserving/imgs/demo.jpg
new file mode 100644
index 000000000..82833092a
Binary files /dev/null and b/recognition/arcface_paddle/deploy/pdserving/imgs/demo.jpg differ
diff --git a/recognition/arcface_paddle/deploy/pdserving/imgs/results.png b/recognition/arcface_paddle/deploy/pdserving/imgs/results.png
new file mode 100644
index 000000000..b99821dbd
Binary files /dev/null and b/recognition/arcface_paddle/deploy/pdserving/imgs/results.png differ
diff --git a/recognition/arcface_paddle/deploy/pdserving/imgs/start_server.png b/recognition/arcface_paddle/deploy/pdserving/imgs/start_server.png
new file mode 100644
index 000000000..47bed1bda
Binary files /dev/null and b/recognition/arcface_paddle/deploy/pdserving/imgs/start_server.png differ
diff --git a/recognition/arcface_paddle/deploy/pdserving/pipeline_http_client.py b/recognition/arcface_paddle/deploy/pdserving/pipeline_http_client.py
new file mode 100644
index 000000000..4fe3dc403
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/pipeline_http_client.py
@@ -0,0 +1,45 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import requests
+import json
+import base64
+import os
+
+import argparse
+parser = argparse.ArgumentParser(description="args for paddleserving")
+parser.add_argument("--image_dir", type=str, default="./imgs")
+args = parser.parse_args()
+
+
+def cv2_to_base64(image):
+    return base64.b64encode(image).decode('utf8')
+
+
+url = "http://127.0.0.1:9998/ArcFace/prediction"
+
+test_img_dir = args.image_dir
+for idx, img_file in enumerate(os.listdir(test_img_dir)):
+    with open(os.path.join(test_img_dir, img_file), 'rb') as file:
+        image_data1 = file.read()
+
+    image = cv2_to_base64(image_data1)
+
+    for i in range(1):
+        data = {"key": ["image"], "value": [image]}
+        r = requests.post(url=url, data=json.dumps(data))
+        print(r.json())
+
+print("==> total number of test imgs: ", len(os.listdir(test_img_dir)))
\ No newline at end of file
diff --git a/recognition/arcface_paddle/deploy/pdserving/pipeline_rpc_client.py b/recognition/arcface_paddle/deploy/pdserving/pipeline_rpc_client.py
new file mode 100644
index 000000000..1c4e0b8c5
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/pipeline_rpc_client.py
@@ -0,0 +1,50 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+try:
+    from paddle_serving_server_gpu.pipeline import PipelineClient
+except ImportError:
+    from paddle_serving_server.pipeline import PipelineClient
+import numpy as np
+import requests
+import json
+import cv2
+import base64
+import os
+
+client = PipelineClient()
+client.connect(['127.0.0.1:18091'])
+
+
+
+def cv2_to_base64(image):
+    return base64.b64encode(image).decode('utf8')
+
+import argparse
+parser = argparse.ArgumentParser(description="args for paddleserving")
+parser.add_argument("--image_dir", type=str, default="../../doc/imgs/")
+args = parser.parse_args()
+test_img_dir = args.image_dir
+
+for idx, img_file in enumerate(os.listdir(test_img_dir)):
+    with open(os.path.join(test_img_dir, img_file), 'rb') as file:
+        image_data1 = file.read()
+
+    image = cv2_to_base64(image_data1)
+
+    for i in range(1):
+        ret = client.predict(feed_dict={"image": image}, fetch=["res"])
+        print(ret)
+
+print("==> total number of test imgs: ", len(os.listdir(test_img_dir)))
\ No newline at end of file
diff --git a/recognition/arcface_paddle/deploy/pdserving/web_service.py b/recognition/arcface_paddle/deploy/pdserving/web_service.py
new file mode 100644
index 000000000..f7f4f68f9
--- /dev/null
+++ b/recognition/arcface_paddle/deploy/pdserving/web_service.py
@@ -0,0 +1,55 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from paddle_serving_server.web_service import WebService, Op
+
+import numpy as np
+import cv2
+import base64
+
+
+class ArcFaceOp(Op):
+    def init_op(self):
+        pass
+
+    def preprocess(self, input_dicts, data_id, log_id):
+        (_, input_dict), = input_dicts.items()
+        data = base64.b64decode(input_dict["image"])
+        data = np.frombuffer(data, np.uint8)
+        # Note: class variables(self.var) can only be used in process op mode
+        img = cv2.imdecode(data, cv2.IMREAD_COLOR)
+        img = cv2.resize(img,(112,112))
+        # normalize to mean 0.5, std 0.5
+        img = (img - 127.5) * 0.00784313725
+        # BGR2RGB
+        img = img[:, :, ::-1]
+        img = img.transpose((2, 0, 1))
+        img = np.expand_dims(img, 0)
+        img = img.astype('float32')
+        return {"x":img.copy()}, False, None, ""
+
+    def postprocess(self, input_dicts, fetch_dict, log_id):
+        out = fetch_dict["save_infer_model/scale_0.tmp_1"]
+        out_dict = {"out": out}
+
+        return out_dict, None, ""
+
+class ArcFaceService(WebService):
+    def get_pipeline_response(self, read_op):
+        arcface_op = ArcFaceOp(name="ArcFace", input_ops=[read_op])
+        return arcface_op
+
+
+arcface_service = ArcFaceService(name="ArcFace")
+arcface_service.prepare_pipeline_config("config.yml")
+arcface_service.run_service()
\ No newline at end of file
diff --git a/recognition/arcface_paddle/dynamic/backbones/__init__.py b/recognition/arcface_paddle/dynamic/backbones/__init__.py
new file mode 100644
index 000000000..4e51edd29
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/backbones/__init__.py
@@ -0,0 +1,16 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .mobilefacenet import MobileFaceNet_128
+from .iresnet import FresResNet50, FresResNet100
diff --git a/recognition/arcface_paddle/dynamic/backbones/iresnet.py b/recognition/arcface_paddle/dynamic/backbones/iresnet.py
new file mode 100644
index 000000000..9424a5f43
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/backbones/iresnet.py
@@ -0,0 +1,337 @@
+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import paddle
+from paddle import ParamAttr
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle.nn import Conv2D, BatchNorm, Linear, Dropout, PReLU
+from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
+from paddle.nn.initializer import XavierNormal, Constant
+
+import math
+
+__all__ = ["FresResNet50", "FresResNet100"]
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 stride=1,
+                 groups=1,
+                 act=None,
+                 name=None,
+                 data_format="NCHW"):
+        super(ConvBNLayer, self).__init__()
+
+        self._conv = Conv2D(
+            in_channels=num_channels,
+            out_channels=num_filters,
+            kernel_size=filter_size,
+            stride=stride,
+            padding=(filter_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + "_weights"),
+            bias_attr=False,
+            data_format=data_format)
+        if name == "conv1":
+            bn_name = "bn_" + name
+        else:
+            bn_name = "bn" + name[3:]
+        self._batch_norm = BatchNorm(
+            num_filters,
+            act=act,
+            epsilon=1e-05,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(bn_name + "_offset"),
+            moving_mean_name=bn_name + "_mean",
+            moving_variance_name=bn_name + "_variance",
+            data_layout=data_format)
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        return y
+
+
+class BasicBlock(nn.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 stride,
+                 shortcut=True,
+                 name=None,
+                 data_format="NCHW"):
+        super(BasicBlock, self).__init__()
+        self.stride = stride
+        bn_name = "bn_" + name[3:] + "_before"
+        self._batch_norm = BatchNorm(
+            num_channels,
+            act=None,
+            epsilon=1e-05,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(bn_name + "_offset"),
+            moving_mean_name=bn_name + "_mean",
+            moving_variance_name=bn_name + "_variance",
+            data_layout=data_format)
+
+        self.conv0 = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=3,
+            stride=1,
+            act=None,
+            name=name + "_branch2a",
+            data_format=data_format)
+        self.prelu = PReLU(num_parameters=1, name=name + "_branch2a_prelu")
+        self.conv1 = ConvBNLayer(
+            num_channels=num_filters,
+            num_filters=num_filters,
+            filter_size=3,
+            stride=stride,
+            act=None,
+            name=name + "_branch2b",
+            data_format=data_format)
+
+        if shortcut:
+            self.short = ConvBNLayer(
+                num_channels=num_channels,
+                num_filters=num_filters,
+                filter_size=1,
+                stride=stride,
+                act=None,
+                name=name + "_branch1",
+                data_format=data_format)
+
+        self.shortcut = shortcut
+
+    def forward(self, inputs):
+        y = self._batch_norm(inputs)
+        y = self.conv0(y)
+        y = self.prelu(y)
+        conv1 = self.conv1(y)
+
+        if self.shortcut:
+            short = self.short(inputs)
+        else:
+            short = inputs
+        y = paddle.add(x=short, y=conv1)
+        return y
+
+
+class FC(nn.Layer):
+    def __init__(self,
+                 bn_channels,
+                 num_channels,
+                 num_classes,
+                 fc_type,
+                 dropout=0.4,
+                 name=None,
+                 data_format="NCHW"):
+        super(FC, self).__init__()
+        self.p = dropout
+        self.fc_type = fc_type
+        self.num_channels = num_channels
+
+        bn_name = "bn_" + name
+        if fc_type == "Z":
+            self._batch_norm_1 = BatchNorm(
+                bn_channels,
+                act=None,
+                epsilon=1e-05,
+                param_attr=ParamAttr(name=bn_name + "_1_scale"),
+                bias_attr=ParamAttr(bn_name + "_1_offset"),
+                moving_mean_name=bn_name + "_1_mean",
+                moving_variance_name=bn_name + "_1_variance",
+                data_layout=data_format)
+            if self.p > 0:
+                self.dropout = Dropout(p=self.p, name=name + '_dropout')
+
+        elif fc_type == "E":
+            self._batch_norm_1 = BatchNorm(
+                bn_channels,
+                act=None,
+                epsilon=1e-05,
+                param_attr=ParamAttr(name=bn_name + "_1_scale"),
+                bias_attr=ParamAttr(bn_name + "_1_offset"),
+                moving_mean_name=bn_name + "_1_mean",
+                moving_variance_name=bn_name + "_1_variance",
+                data_layout=data_format)
+            if self.p > 0:
+                self.dropout = Dropout(p=self.p, name=name + '_dropout')
+            self.fc = Linear(
+                num_channels,
+                num_classes,
+                weight_attr=ParamAttr(
+                    initializer=XavierNormal(fan_in=0.0), name=name + ".w_0"),
+                bias_attr=ParamAttr(
+                    initializer=Constant(), name=name + ".b_0"))
+            self._batch_norm_2 = BatchNorm(
+                num_classes,
+                act=None,
+                epsilon=1e-05,
+                param_attr=ParamAttr(name=bn_name + "_2_scale"),
+                bias_attr=ParamAttr(bn_name + "_2_offset"),
+                moving_mean_name=bn_name + "_2_mean",
+                moving_variance_name=bn_name + "_2_variance",
+                data_layout=data_format)
+
+        elif fc_type == "FC":
+            self._batch_norm_1 = BatchNorm(
+                bn_channels,
+                act=None,
+                epsilon=1e-05,
+                param_attr=ParamAttr(name=bn_name + "_1_scale"),
+                bias_attr=ParamAttr(bn_name + "_1_offset"),
+                moving_mean_name=bn_name + "_1_mean",
+                moving_variance_name=bn_name + "_1_variance",
+                data_layout=data_format)
+            self.fc = Linear(
+                num_channels,
+                num_classes,
+                weight_attr=ParamAttr(
+                    initializer=XavierNormal(fan_in=0.0), name=name + ".w_0"),
+                bias_attr=ParamAttr(
+                    initializer=Constant(), name=name + ".b_0"))
+            self._batch_norm_2 = BatchNorm(
+                num_classes,
+                act=None,
+                epsilon=1e-05,
+                param_attr=ParamAttr(name=bn_name + "_2_scale"),
+                bias_attr=ParamAttr(bn_name + "_2_offset"),
+                moving_mean_name=bn_name + "_2_mean",
+                moving_variance_name=bn_name + "_2_variance",
+                data_layout=data_format)
+
+    def forward(self, inputs):
+        if self.fc_type == "Z":
+            y = self._batch_norm_1(inputs)
+            y = paddle.reshape(y, shape=[-1, self.num_channels])
+            if self.p > 0:
+                y = self.dropout(y)
+
+        elif self.fc_type == "E":
+            y = self._batch_norm_1(inputs)
+            y = paddle.reshape(y, shape=[-1, self.num_channels])
+            if self.p > 0:
+                y = self.dropout(y)
+            y = self.fc(y)
+            y = self._batch_norm_2(y)
+
+        elif self.fc_type == "FC":
+            y = self._batch_norm_1(inputs)
+            y = paddle.reshape(y, shape=[-1, self.num_channels])
+            y = self.fc(y)
+            y = self._batch_norm_2(y)
+
+        return y
+
+
+class FresResNet(nn.Layer):
+    def __init__(self,
+                 layers=50,
+                 num_features=512,
+                 fc_type='E',
+                 dropout=0.4,
+                 input_image_channel=3,
+                 input_image_width=112,
+                 input_image_height=112,
+                 data_format="NCHW"):
+
+        super(FresResNet, self).__init__()
+
+        self.layers = layers
+        self.data_format = data_format
+        self.input_image_channel = input_image_channel
+
+        supported_layers = [50, 100]
+        assert layers in supported_layers, \
+            "supported layers are {} but input layer is {}".format(
+                supported_layers, layers)
+
+        if layers == 50:
+            units = [3, 4, 14, 3]
+        elif layers == 100:
+            units = [3, 13, 30, 3]
+
+        num_channels = [64, 64, 128, 256]
+        num_filters = [64, 128, 256, 512]
+
+        self.conv = ConvBNLayer(
+            num_channels=self.input_image_channel,
+            num_filters=64,
+            filter_size=3,
+            stride=1,
+            act=None,
+            name="conv1",
+            data_format=self.data_format)
+        self.prelu = PReLU(num_parameters=1, name="prelu1")
+
+        self.block_list = paddle.nn.LayerList()
+        for block in range(len(units)):
+            shortcut = True
+            for i in range(units[block]):
+                conv_name = "res" + str(block + 2) + chr(97 + i)
+                basic_block = self.add_sublayer(
+                    conv_name,
+                    BasicBlock(
+                        num_channels=num_channels[block]
+                        if i == 0 else num_filters[block],
+                        num_filters=num_filters[block],
+                        stride=2 if shortcut else 1,
+                        shortcut=shortcut,
+                        name=conv_name,
+                        data_format=self.data_format))
+                self.block_list.append(basic_block)
+                shortcut = False
+
+        assert input_image_width % 16 == 0
+        assert input_image_height % 16 == 0
+        feat_w = input_image_width // 16
+        feat_h = input_image_height // 16
+        self.fc_channels = num_filters[-1] * feat_w * feat_h
+        self.fc = FC(num_filters[-1],
+                     self.fc_channels,
+                     num_features,
+                     fc_type,
+                     dropout,
+                     name='fc')
+
+    def forward(self, inputs):
+        if self.data_format == "NHWC":
+            inputs = paddle.tensor.transpose(inputs, [0, 2, 3, 1])
+            inputs.stop_gradient = True
+        y = self.conv(inputs)
+        y = self.prelu(y)
+        for block in self.block_list:
+            y = block(y)
+        y = self.fc(y)
+        return y
+
+
+def FresResNet50(**args):
+    model = FresResNet(layers=50, **args)
+    return model
+
+
+def FresResNet100(**args):
+    model = FresResNet(layers=100, **args)
+    return model
diff --git a/recognition/arcface_paddle/dynamic/backbones/mobilefacenet.py b/recognition/arcface_paddle/dynamic/backbones/mobilefacenet.py
new file mode 100644
index 000000000..6949b2851
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/backbones/mobilefacenet.py
@@ -0,0 +1,162 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+from paddle import nn
+import math
+
+__all__ = ['MobileFaceNet_128']
+
+MobileFaceNet_BottleNeck_Setting = [
+    # t, c , n ,s
+    [2, 64, 5, 2],
+    [4, 128, 1, 2],
+    [2, 128, 6, 1],
+    [4, 128, 1, 2],
+    [2, 128, 2, 1]
+]
+
+
+class BottleNeck(nn.Layer):
+    def __init__(self, inp, oup, stride, expansion):
+        super().__init__()
+        self.connect = stride == 1 and inp == oup
+
+        self.conv = nn.Sequential(
+            # 1*1 conv
+            nn.Conv2D(
+                inp, inp * expansion, 1, 1, 0, bias_attr=False),
+            nn.BatchNorm2D(inp * expansion),
+            nn.PReLU(inp * expansion),
+
+            # 3*3 depth wise conv
+            nn.Conv2D(
+                inp * expansion,
+                inp * expansion,
+                3,
+                stride,
+                1,
+                groups=inp * expansion,
+                bias_attr=False),
+            nn.BatchNorm2D(inp * expansion),
+            nn.PReLU(inp * expansion),
+
+            # 1*1 conv
+            nn.Conv2D(
+                inp * expansion, oup, 1, 1, 0, bias_attr=False),
+            nn.BatchNorm2D(oup), )
+
+    def forward(self, x):
+        if self.connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class ConvBlock(nn.Layer):
+    def __init__(self, inp, oup, k, s, p, dw=False, linear=False):
+        super().__init__()
+        self.linear = linear
+        if dw:
+            self.conv = nn.Conv2D(
+                inp, oup, k, s, p, groups=inp, bias_attr=False)
+        else:
+            self.conv = nn.Conv2D(inp, oup, k, s, p, bias_attr=False)
+
+        self.bn = nn.BatchNorm2D(oup)
+        if not linear:
+            self.prelu = nn.PReLU(oup)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        if self.linear:
+            return x
+        else:
+            return self.prelu(x)
+
+
+class MobileFaceNet(nn.Layer):
+    def __init__(self,
+                 feature_dim=128,
+                 bottleneck_setting=MobileFaceNet_BottleNeck_Setting,
+                 **args):
+        super().__init__()
+        self.conv1 = ConvBlock(3, 64, 3, 2, 1)
+        self.dw_conv1 = ConvBlock(64, 64, 3, 1, 1, dw=True)
+
+        self.cur_channel = 64
+        block = BottleNeck
+        self.blocks = self._make_layer(block, bottleneck_setting)
+
+        self.conv2 = ConvBlock(128, 512, 1, 1, 0)
+        self.linear7 = ConvBlock(512, 512, 7, 1, 0, dw=True, linear=True)
+        self.linear1 = ConvBlock(512, feature_dim, 1, 1, 0, linear=True)
+
+        for m in self.sublayers():
+            if isinstance(m, nn.Conv2D):
+                # ks * ks * out_ch
+                n = m.weight.shape[1] * m.weight.shape[2] * m.weight.shape[3]
+                m.weight = paddle.create_parameter(
+                    shape=m.weight.shape,
+                    dtype=m.weight.dtype,
+                    default_initializer=nn.initializer.Normal(
+                        mean=0.0, std=math.sqrt(2.0 / n)))
+                
+            elif isinstance(m, (nn.BatchNorm, nn.BatchNorm2D, nn.GroupNorm)):
+                m.weight = paddle.create_parameter(
+                    shape=m.weight.shape,
+                    dtype=m.weight.dtype,
+                    default_initializer=nn.initializer.Constant(value=1.0))
+                m.bias = paddle.create_parameter(
+                    shape=m.bias.shape,
+                    dtype=m.bias.dtype,
+                    default_initializer=nn.initializer.Constant(value=0.0))
+
+    def _make_layer(self, block, setting):
+        layers = []
+        for t, c, n, s in setting:
+            for i in range(n):
+                if i == 0:
+                    layers.append(block(self.cur_channel, c, s, t))
+                else:
+                    layers.append(block(self.cur_channel, c, 1, t))
+                self.cur_channel = c
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.dw_conv1(x)
+        x = self.blocks(x)
+        x = self.conv2(x)
+        x = self.linear7(x)
+        x = self.linear1(x)
+        x = x.reshape([x.shape[0], x.shape[1] * x.shape[2] * x.shape[3]])
+        return x
+
+
+def MobileFaceNet_128(num_features=128, **args):
+    model = MobileFaceNet(feature_dim=num_features, **args)
+    return model
+
+
+# if __name__ == "__main__":
+#     paddle.set_device("cpu")
+#     x = paddle.rand([2, 3, 112, 112])
+#     net = MobileFaceNet()
+#     print(net)
+
+#     x = net(x)
+#     print(x.shape)
diff --git a/recognition/arcface_paddle/dynamic/classifiers/__init__.py b/recognition/arcface_paddle/dynamic/classifiers/__init__.py
new file mode 100644
index 000000000..0c0e52f2b
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/classifiers/__init__.py
@@ -0,0 +1,15 @@
+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .lsc import LargeScaleClassifier
diff --git a/recognition/arcface_paddle/dynamic/classifiers/lsc.py b/recognition/arcface_paddle/dynamic/classifiers/lsc.py
new file mode 100644
index 000000000..f22a769ec
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/classifiers/lsc.py
@@ -0,0 +1,164 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+import math
+import os
+import paddle
+import paddle.nn as nn
+
+class LargeScaleClassifier(nn.Layer):
+    """
+    Author: {Xiang An, Yang Xiao, XuHan Zhu} in DeepGlint,
+    Partial FC: Training 10 Million Identities on a Single Machine
+    See the original paper:
+    https://arxiv.org/abs/2010.05222
+    """
+
+    @paddle.no_grad()
+    def __init__(self,
+                 rank,
+                 world_size,
+                 num_classes,
+                 margin1=1.0,
+                 margin2=0.5,
+                 margin3=0.0,
+                 scale=64.0,
+                 sample_ratio=1.0,
+                 embedding_size=512,
+                 fp16=False,
+                 name=None):
+        super(LargeScaleClassifier, self).__init__()
+        self.num_classes: int = num_classes
+        self.rank: int = rank
+        self.world_size: int = world_size
+        self.sample_ratio: float = sample_ratio
+        self.embedding_size: int = embedding_size
+        self.fp16 = fp16
+        self.num_local: int = (num_classes + world_size - 1) // world_size
+        if num_classes % world_size != 0 and rank == world_size - 1:
+            self.num_local = num_classes % self.num_local
+        self.num_sample: int = int(self.sample_ratio * self.num_local)
+        self.margin1 = margin1
+        self.margin2 = margin2
+        self.margin3 = margin3
+        self.logit_scale = scale
+
+        self._parameter_list = []
+
+        if name is None:
+            name = 'dist@fc@rank@%05d.w' % rank
+        assert '.w' in name
+
+        stddev = math.sqrt(2.0 / (self.embedding_size + self.num_local))
+        param_attr = paddle.ParamAttr(
+            name=name, initializer=paddle.nn.initializer.Normal(std=stddev))
+
+        self.index = None
+        self.weight = self.create_parameter(
+            shape=[self.embedding_size, self.num_local],
+            attr=param_attr,
+            is_bias=False,
+            dtype='float16' if self.fp16 else 'float32')
+        self.weight.is_distributed = True
+
+        if int(self.sample_ratio) < 1:
+            self.weight.stop_gradient = True
+
+    def step(self, optimizer):
+        if int(self.sample_ratio) < 1:
+            warnings.warn(
+                "Explicitly call the function paddle._C_ops.sparse_momentum is a temporary manner. "
+                "We will merge it to optimizer in the future, please don't follow.")
+            
+            found_inf = paddle.logical_not(
+                paddle.all(paddle.isfinite(self._parameter_list[0].grad)))
+            if found_inf:
+                print('Found inf or nan in classifier')
+            else:
+                if self.weight.name not in optimizer._accumulators[
+                        optimizer._velocity_acc_str]:
+                    optimizer._add_accumulator(optimizer._velocity_acc_str,
+                                               self.weight)
+
+                velocity = optimizer._accumulators[
+                    optimizer._velocity_acc_str][self.weight.name]
+                _, _ = paddle._C_ops.sparse_momentum(
+                    self.weight,
+                    self._parameter_list[0].grad,
+                    velocity,
+                    self.index,
+                    paddle.to_tensor(
+                        optimizer.get_lr(), dtype='float32'),
+                    self.weight,
+                    velocity,
+                    'mu',
+                    optimizer._momentum,
+                    'use_nesterov',
+                    optimizer._use_nesterov,
+                    'regularization_method',
+                    optimizer._regularization_method,
+                    'regularization_coeff',
+                    optimizer._regularization_coeff,
+                    'axis',
+                    1)
+
+    def clear_grad(self):
+        self._parameter_list = []
+
+    def forward(self, feature, label):
+
+        if self.world_size > 1:
+            feature_list = []
+            paddle.distributed.all_gather(feature_list, feature)
+            total_feature = paddle.concat(feature_list, axis=0)
+
+            label_list = []
+            paddle.distributed.all_gather(label_list, label)
+            total_label = paddle.concat(label_list, axis=0)
+            total_label.stop_gradient = True
+        else:
+            total_feature = feature
+            total_label = label
+
+        if self.sample_ratio < 1.0:
+            # partial fc sample process
+            total_label, self.index = paddle.nn.functional.class_center_sample(
+                total_label, self.num_local, self.num_sample)
+            total_label.stop_gradient = True
+            self.index.stop_gradient = True
+            self.sub_weight = paddle.gather(self.weight, self.index, axis=1)
+            self.sub_weight.stop_gradient = False
+            self._parameter_list.append(self.sub_weight)
+        else:
+            self.sub_weight = self.weight
+
+        norm_feature = paddle.fluid.layers.l2_normalize(total_feature, axis=1)
+        norm_weight = paddle.fluid.layers.l2_normalize(self.sub_weight, axis=0)
+
+        local_logit = paddle.matmul(norm_feature, norm_weight)
+
+        loss = paddle.nn.functional.margin_cross_entropy(
+            local_logit,
+            total_label,
+            margin1=self.margin1,
+            margin2=self.margin2,
+            margin3=self.margin3,
+            scale=self.logit_scale,
+            return_softmax=False,
+            reduction=None, )
+
+        loss = paddle.mean(loss)
+
+        return loss
diff --git a/recognition/arcface_paddle/dynamic/export.py b/recognition/arcface_paddle/dynamic/export.py
new file mode 100644
index 000000000..a41d30aeb
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/export.py
@@ -0,0 +1,56 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import numpy as np
+import paddle
+
+from .utils.io import Checkpoint
+from . import backbones
+
+
+def export(args):
+    checkpoint = Checkpoint(
+        rank=0,
+        world_size=1,
+        embedding_size=args.embedding_size,
+        num_classes=None,
+        checkpoint_dir=args.checkpoint_dir, )
+
+    backbone = eval("backbones.{}".format(args.backbone))(
+        num_features=args.embedding_size)
+    checkpoint.load(backbone, for_train=False, dtype='float32')
+
+    print("Load checkpoint from '{}'.".format(args.checkpoint_dir))
+    backbone.eval()
+
+    path = os.path.join(args.output_dir, args.backbone)
+
+    if args.export_type == 'onnx':
+        paddle.onnx.export(
+            backbone,
+            path,
+            input_spec=[
+                paddle.static.InputSpec(
+                    shape=[None, 3, 112, 112], dtype='float32')
+            ])
+    else:
+        paddle.jit.save(
+            backbone,
+            path,
+            input_spec=[
+                paddle.static.InputSpec(
+                    shape=[None, 3, 112, 112], dtype='float32')
+            ])
+    print("Save exported model to '{}'.".format(args.output_dir))
diff --git a/recognition/arcface_paddle/dynamic/train.py b/recognition/arcface_paddle/dynamic/train.py
new file mode 100644
index 000000000..778ea54f8
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/train.py
@@ -0,0 +1,247 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import sys
+import numpy as np
+import logging
+
+import paddle
+from visualdl import LogWriter
+
+from utils.logging import AverageMeter, init_logging, CallBackLogging
+from datasets import CommonDataset, SyntheticDataset
+from utils import losses
+
+from .utils.verification import CallBackVerification
+from .utils.io import Checkpoint
+from .utils.amp import LSCGradScaler
+
+from . import classifiers
+from . import backbones
+
+RELATED_FLAGS_SETTING = {
+    'FLAGS_cudnn_exhaustive_search': 1,
+    'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
+    'FLAGS_max_inplace_grad_add': 8,
+    'FLAGS_fraction_of_gpu_memory_to_use': 0.9999,
+}
+paddle.fluid.set_flags(RELATED_FLAGS_SETTING)
+
+
+def train(args):
+    writer = LogWriter(logdir=args.logdir)
+
+    rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
+    world_size = int(os.getenv("PADDLE_TRAINERS_NUM", 1))
+
+    gpu_id = int(os.getenv("FLAGS_selected_gpus", 0))
+    place = paddle.CUDAPlace(gpu_id)
+
+    if world_size > 1:
+        import paddle.distributed.fleet as fleet
+        from .utils.data_parallel import sync_gradients, sync_params
+
+        strategy = fleet.DistributedStrategy()
+        strategy.without_graph_optimization = True
+        fleet.init(is_collective=True, strategy=strategy)
+
+    if args.use_synthetic_dataset:
+        trainset = SyntheticDataset(args.num_classes, fp16=args.fp16)
+    else:
+        trainset = CommonDataset(
+            root_dir=args.data_dir,
+            label_file=args.label_file,
+            fp16=args.fp16,
+            is_bin=args.is_bin)
+
+    num_image = len(trainset)
+    total_batch_size = args.batch_size * world_size
+    steps_per_epoch = num_image // total_batch_size
+    if args.train_unit == 'epoch':
+        warmup_steps = steps_per_epoch * args.warmup_num
+        total_steps = steps_per_epoch * args.train_num
+        decay_steps = [x * steps_per_epoch for x in args.decay_boundaries]
+        total_epoch = args.train_num
+    else:
+        warmup_steps = args.warmup_num
+        total_steps = args.train_num
+        decay_steps = [x for x in args.decay_boundaries]
+        total_epoch = (total_steps + steps_per_epoch - 1) // steps_per_epoch
+
+    if rank == 0:
+        logging.info('world_size: {}'.format(world_size))
+        logging.info('total_batch_size: {}'.format(total_batch_size))
+        logging.info('warmup_steps: {}'.format(warmup_steps))
+        logging.info('steps_per_epoch: {}'.format(steps_per_epoch))
+        logging.info('total_steps: {}'.format(total_steps))
+        logging.info('total_epoch: {}'.format(total_epoch))
+        logging.info('decay_steps: {}'.format(decay_steps))
+
+    base_lr = total_batch_size * args.lr / 512
+    lr_scheduler = paddle.optimizer.lr.PiecewiseDecay(
+        boundaries=decay_steps,
+        values=[
+            base_lr * (args.lr_decay**i) for i in range(len(decay_steps) + 1)
+        ])
+    if warmup_steps > 0:
+        lr_scheduler = paddle.optimizer.lr.LinearWarmup(
+            lr_scheduler, warmup_steps, 0, base_lr)
+
+    if args.fp16:
+        paddle.set_default_dtype("float16")
+
+    margin_loss_params = eval("losses.{}".format(args.loss))()
+    backbone = eval("backbones.{}".format(args.backbone))(
+        num_features=args.embedding_size, dropout=args.dropout)
+    classifier = eval("classifiers.{}".format(args.classifier))(
+        rank=rank,
+        world_size=world_size,
+        num_classes=args.num_classes,
+        margin1=margin_loss_params.margin1,
+        margin2=margin_loss_params.margin2,
+        margin3=margin_loss_params.margin3,
+        scale=margin_loss_params.scale,
+        sample_ratio=args.sample_ratio,
+        embedding_size=args.embedding_size,
+        fp16=args.fp16)
+
+    backbone.train()
+    classifier.train()
+
+    optimizer = paddle.optimizer.Momentum(
+        parameters=[{
+            'params': backbone.parameters(),
+        }, {
+            'params': classifier.parameters(),
+        }],
+        learning_rate=lr_scheduler,
+        momentum=args.momentum,
+        weight_decay=args.weight_decay)
+
+    if args.fp16:
+        optimizer._dtype = 'float32'
+
+    if world_size > 1:
+        # sync backbone params for data parallel
+        sync_params(backbone.parameters())
+
+    if args.do_validation_while_train:
+        callback_verification = CallBackVerification(
+            args.validation_interval_step,
+            rank,
+            args.batch_size,
+            args.val_targets,
+            args.data_dir,
+            fp16=args.fp16, )
+
+    callback_logging = CallBackLogging(args.log_interval_step, rank,
+                                       world_size, total_steps,
+                                       args.batch_size, writer)
+
+    checkpoint = Checkpoint(
+        rank=rank,
+        world_size=world_size,
+        embedding_size=args.embedding_size,
+        num_classes=args.num_classes,
+        model_save_dir=os.path.join(args.output, args.backbone),
+        checkpoint_dir=args.checkpoint_dir,
+        max_num_last_checkpoint=args.max_num_last_checkpoint)
+
+    start_epoch = 0
+    global_step = 0
+    loss_avg = AverageMeter()
+    if args.resume:
+        extra_info = checkpoint.load(
+            backbone, classifier, optimizer, for_train=True)
+        start_epoch = extra_info['epoch'] + 1
+        lr_state = extra_info['lr_state']
+        # there last_epoch means last_step in for PiecewiseDecay
+        # since we always use step style for lr_scheduler
+        global_step = lr_state['last_epoch']
+        lr_scheduler.set_state_dict(lr_state)
+
+    train_loader = paddle.io.DataLoader(
+        trainset,
+        places=place,
+        num_workers=args.num_workers,
+        batch_sampler=paddle.io.DistributedBatchSampler(
+            dataset=trainset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True))
+
+    scaler = LSCGradScaler(
+        enable=args.fp16,
+        init_loss_scaling=args.init_loss_scaling,
+        incr_ratio=args.incr_ratio,
+        decr_ratio=args.decr_ratio,
+        incr_every_n_steps=args.incr_every_n_steps,
+        decr_every_n_nan_or_inf=args.decr_every_n_nan_or_inf,
+        use_dynamic_loss_scaling=args.use_dynamic_loss_scaling)
+
+    for epoch in range(start_epoch, total_epoch):
+        train_reader_cost = 0.0
+        train_run_cost = 0.0
+        total_samples = 0
+        reader_start = time.time()
+        for step, (img, label) in enumerate(train_loader):
+            train_reader_cost += time.time() - reader_start
+            global_step += 1
+            train_start = time.time()
+            with paddle.amp.auto_cast(enable=args.fp16):
+                features = backbone(img)
+                loss_v = classifier(features, label)
+
+            scaler.scale(loss_v).backward()
+            if world_size > 1:
+                # data parallel sync backbone gradients
+                sync_gradients(backbone.parameters())
+
+            scaler.step(optimizer)
+            classifier.step(optimizer)
+            optimizer.clear_grad()
+            classifier.clear_grad()
+
+            train_run_cost += time.time() - train_start
+            total_samples += len(img)
+
+            lr_value = optimizer.get_lr()
+            loss_avg.update(loss_v.item(), 1)
+            callback_logging(
+                global_step,
+                loss_avg,
+                epoch,
+                lr_value,
+                avg_reader_cost=train_reader_cost / args.log_interval_step,
+                avg_batch_cost=(train_reader_cost + train_run_cost) / args.log_interval_step,
+                avg_samples=total_samples / args.log_interval_step,
+                ips=total_samples / (train_reader_cost + train_run_cost))
+           
+            if args.do_validation_while_train:
+                callback_verification(global_step, backbone)
+            lr_scheduler.step()
+
+            if global_step >= total_steps:
+                break
+            sys.stdout.flush()
+            if rank is 0 and global_step > 0 and global_step % args.log_interval_step == 0:
+                train_reader_cost = 0.0
+                train_run_cost = 0.0
+                total_samples = 0
+            reader_start = time.time()
+        checkpoint.save(
+            backbone, classifier, optimizer, epoch=epoch, for_train=True)
+    writer.close()
diff --git a/recognition/arcface_paddle/dynamic/utils/__init__.py b/recognition/arcface_paddle/dynamic/utils/__init__.py
new file mode 100644
index 000000000..185a92b8d
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/utils/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/recognition/arcface_paddle/dynamic/utils/amp.py b/recognition/arcface_paddle/dynamic/utils/amp.py
new file mode 100644
index 000000000..d42c09cad
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/utils/amp.py
@@ -0,0 +1,101 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import defaultdict
+from paddle.amp import GradScaler
+from paddle import _C_ops
+import paddle
+
+
+class LSCGradScaler(GradScaler):
+    def __init__(self,
+                 enable=True,
+                 init_loss_scaling=2.**15,
+                 incr_ratio=2.0,
+                 decr_ratio=0.5,
+                 incr_every_n_steps=1000,
+                 decr_every_n_nan_or_inf=2,
+                 use_dynamic_loss_scaling=True,
+                 max_loss_scaling=32768.0):
+        super(LSCGradScaler, self).__init__(
+            enable, init_loss_scaling, incr_ratio, decr_ratio,
+            incr_every_n_steps, decr_every_n_nan_or_inf,
+            use_dynamic_loss_scaling)
+        self.max_loss_scaling = max_loss_scaling
+
+    def step(self, optimizer, classifier=None):
+        if not self._enable:
+            if classifier is not None:
+                classifier.step(optimizer)
+            return optimizer.step()
+
+        # unscale the grad
+        self._unscale(optimizer)
+
+        if self._found_inf:
+            self._cache_founf_inf = True
+        else:
+            optimizer.step()
+            if classifier is not None:
+                classifier.step(optimizer)
+
+            self._cache_founf_inf = False
+
+        if self._use_dynamic_loss_scaling:
+            # update the scale
+            self._update()
+
+    def _unscale(self, optimizer):
+        if not self._enable:
+            return
+
+        param_grads_dict = defaultdict(list)
+        dist_param_grads_dict = defaultdict(list)
+        if getattr(optimizer, '_param_groups', None) and isinstance(
+                optimizer._param_groups[0], dict):
+            for group in optimizer._param_groups:
+                for param in group['params']:
+                    if not param.is_distributed:
+                        if param._grad_ivar() is not None:
+                            param_grads_dict[param._grad_ivar().dtype].append(
+                                param._grad_ivar())
+                    else:
+                        if param._grad_ivar() is not None:
+                            dist_param_grads_dict[param._grad_ivar(
+                            ).dtype].append(param._grad_ivar())
+        else:
+            for param in optimizer._parameter_list:
+                if not param.is_distributed:
+                    if param._grad_ivar() is not None:
+                        param_grads_dict[param._grad_ivar().dtype].append(
+                            param._grad_ivar())
+                else:
+                    if param._grad_ivar() is not None:
+                        dist_param_grads_dict[param._grad_ivar().dtype].append(
+                            param._grad_ivar())
+        for dtype in dist_param_grads_dict:
+            for grad in dist_param_grads_dict[dtype]:
+                self._found_inf = paddle.logical_not(
+                    paddle.all(paddle.isfinite(grad)))
+                if self._found_inf:
+                    print('Found inf or nan in classifier, dtype is', dtype)
+                    return
+
+        for dtype in param_grads_dict:
+            param_grads = param_grads_dict[dtype]
+            _C_ops.check_finite_and_unscale(param_grads, self._scale,
+                                            param_grads, self._found_inf)
+            if self._found_inf:
+                print('Found inf or nan in backbone, dtype is', dtype)
+                break
diff --git a/recognition/arcface_paddle/dynamic/utils/data_parallel.py b/recognition/arcface_paddle/dynamic/utils/data_parallel.py
new file mode 100644
index 000000000..baa4372c4
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/utils/data_parallel.py
@@ -0,0 +1,56 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+
+
+@paddle.no_grad()
+def sync_params(parameters):
+    for param in parameters:
+        paddle.distributed.broadcast(
+            param.detach(), src=0, group=None, use_calc_stream=True)
+
+
+@paddle.no_grad()
+def sync_gradients(parameters):
+    grad_var_set = set()
+    grad_vars = []
+    sparse_grad_vars = []
+
+    for param in parameters:
+        if param.trainable and (param._grad_ivar() is not None):
+            g_var = param._grad_ivar()
+            assert not g_var._is_sparse(
+            ), "Now, it doesn't support sparse parameters"
+            grad_vars.append(g_var)
+            assert g_var not in grad_var_set
+            grad_var_set.add(g_var)
+
+    coalesced_grads_and_vars = \
+        paddle.fluid.dygraph.parallel.build_groups(grad_vars, 128 * 1024 * 1024)
+
+    nranks = paddle.distributed.get_world_size()
+    for coalesced_grad, _, _ in coalesced_grads_and_vars:
+        # need to div nranks
+        div_factor = paddle.to_tensor(nranks, dtype=coalesced_grad.dtype)
+        paddle.fluid.framework._dygraph_tracer().trace_op(
+            type="elementwise_div",
+            inputs={'X': coalesced_grad,
+                    'Y': div_factor},
+            outputs={'Out': coalesced_grad},
+            attrs={'axis': -1})
+
+        paddle.distributed.all_reduce(coalesced_grad)
+
+    paddle.fluid.dygraph.parallel._split_tensors(coalesced_grads_and_vars)
diff --git a/recognition/arcface_paddle/dynamic/utils/io.py b/recognition/arcface_paddle/dynamic/utils/io.py
new file mode 100644
index 000000000..a30449c86
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/utils/io.py
@@ -0,0 +1,238 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import errno
+import os
+import paddle
+import logging
+import numpy as np
+import shutil
+import json
+from utils.rearrange_weight import rearrange_weight
+
+
+class Checkpoint(object):
+    def __init__(self,
+                 rank,
+                 world_size,
+                 embedding_size,
+                 num_classes,
+                 model_save_dir="./",
+                 checkpoint_dir=None,
+                 max_num_last_checkpoint=3):
+
+        self.rank: int = rank
+        self.world_size: int = world_size
+        self.embedding_size: int = embedding_size
+        self.num_classes: int = num_classes
+        self.model_save_dir: str = model_save_dir
+        self.checkpoint_dir: str = checkpoint_dir
+        self.max_num_last_checkpoint: int = max_num_last_checkpoint
+
+    def save(self,
+             backbone: paddle.nn.Layer,
+             classifier: paddle.nn.Layer=None,
+             optimizer=None,
+             epoch=0,
+             for_train=True):
+
+        model_save_dir = os.path.join(self.model_save_dir, str(epoch))
+        if not os.path.exists(model_save_dir):
+            # may be more than one processes trying
+            # to create the directory
+            try:
+                os.makedirs(model_save_dir)
+            except OSError as exc:
+                if exc.errno != errno.EEXIST:
+                    raise
+                pass
+
+        if self.rank == 0:
+            # for non dist param, we only save their at rank 0.
+            for name, param in backbone.state_dict().items():
+                paddle.save(
+                    param,
+                    os.path.join(model_save_dir, param.name + '.pdparam'))
+
+        if classifier is not None:
+            # for dist param, we need to save their at all ranks.
+            for name, param in classifier.state_dict().items():
+                paddle.save(
+                    param,
+                    os.path.join(model_save_dir, param.name + '.pdparam'))
+
+        if for_train:
+            assert optimizer is not None
+            opt_state_dict = optimizer.state_dict()
+            lr_state_dict = opt_state_dict['LR_Scheduler']
+            for name, opt in opt_state_dict.items():
+                if '@GRAD' in name:
+                    continue
+                # for non dist opt var, we only save their at rank 0,
+                # but for dist opt var, we need to save their at all ranks.
+                if 'dist@' in name and '@rank@' in name or self.rank == 0:
+                    paddle.save(opt,
+                                os.path.join(model_save_dir, name + '.pdopt'))
+
+            if self.rank == 0:
+                # save some extra info for resume
+                # pretrain_world_size, embedding_size, num_classes are used for
+                # re-split fc weight when gpu setting changed.
+                # epoch use to restart.
+                config_file = os.path.join(model_save_dir, 'meta.json')
+                extra_info = dict()
+                extra_info["pretrain_world_size"] = self.world_size
+                extra_info["embedding_size"] = self.embedding_size
+                extra_info['num_classes'] = self.num_classes
+                extra_info['epoch'] = epoch
+                extra_info['lr_state'] = lr_state_dict
+                with open(config_file, 'w') as f:
+                    json.dump(extra_info, f)
+
+        logging.info("Save model to {}.".format(model_save_dir))
+        if self.rank == 0 and self.max_num_last_checkpoint > 0:
+            for idx in range(-1, epoch - self.max_num_last_checkpoint + 1):
+                path = os.path.join(self.model_save_dir, str(idx))
+                if os.path.exists(path):
+                    logging.info("Remove checkpoint {}.".format(path))
+                    shutil.rmtree(path)
+
+    def load(self,
+             backbone: paddle.nn.Layer,
+             classifier: paddle.nn.Layer=None,
+             optimizer=None,
+             for_train=True,
+             dtype=None):
+
+        assert os.path.exists(self.checkpoint_dir)
+        checkpoint_dir = os.path.abspath(self.checkpoint_dir)
+
+        param_state_dict = {}
+        opt_state_dict = {}
+        dist_param_state_dict = {}
+
+        dist_weight_state_dict = {}
+        dist_weight_velocity_state_dict = {}
+        dist_bias_state_dict = {}
+        dist_bias_velocity_state_dict = {}
+        for path in os.listdir(checkpoint_dir):
+            path = os.path.join(checkpoint_dir, path)
+            if not os.path.isfile(path):
+                continue
+
+            basename = os.path.basename(path)
+            name, ext = os.path.splitext(basename)
+
+            if ext not in ['.pdopt', '.pdparam']:
+                continue
+
+            if not for_train and ext == '.pdopt':
+                continue
+
+            tensor = paddle.load(path, return_numpy=True)
+            if dtype:
+                assert dtype in ['float32', 'float16']
+                tensor = tensor.astype(dtype)
+
+            if 'dist@' in name and '@rank@' in name:
+                if '.w' in name and 'velocity' not in name:
+                    dist_weight_state_dict[name] = tensor
+                elif '.w' in name and 'velocity' in name:
+                    dist_weight_velocity_state_dict[name] = tensor
+                elif '.b' in name and 'velocity' not in name:
+                    dist_bias_state_dict[name] = tensor
+                elif '.b' in name and 'velocity' in name:
+                    dist_bias_velocity_state_dict[name] = tensor
+
+            else:
+                if ext == '.pdparam':
+                    param_state_dict[name] = tensor
+                else:
+                    opt_state_dict[name] = tensor
+
+        if classifier is not None and for_train:
+            meta_file = os.path.join(checkpoint_dir, 'meta.json')
+            if not os.path.exists(meta_file):
+                logging.error(
+                    "Please make sure the checkpoint dir {} exists, and "
+                    "parameters in that dir are validating.".format(
+                        checkpoint_dir))
+                exit()
+
+            with open(meta_file, 'r') as handle:
+                extra_info = json.load(handle)
+
+            # Preporcess distributed parameters.
+            pretrain_world_size = extra_info['pretrain_world_size']
+            assert pretrain_world_size > 0
+            embedding_size = extra_info['embedding_size']
+            assert embedding_size == self.embedding_size
+            num_classes = extra_info['num_classes']
+            assert num_classes == self.num_classes
+
+            logging.info(
+                "Parameters for pre-training: pretrain_world_size ({}), "
+                "embedding_size ({}), and num_classes ({}).".format(
+                    pretrain_world_size, embedding_size, num_classes))
+            logging.info("Parameters for inference or fine-tuning: "
+                         "world_size ({}).".format(self.world_size))
+
+            rank_str = '%05d' % self.rank
+
+            dist_weight_state_dict = rearrange_weight(
+                dist_weight_state_dict, pretrain_world_size, self.world_size)
+            dist_bias_state_dict = rearrange_weight(
+                dist_bias_state_dict, pretrain_world_size, self.world_size)
+            for name, value in dist_weight_state_dict.items():
+                if rank_str in name:
+                    dist_param_state_dict[name] = value
+            for name, value in dist_bias_state_dict.items():
+                if rank_str in name:
+                    dist_param_state_dict[name] = value
+
+            if for_train:
+                dist_weight_velocity_state_dict = rearrange_weight(
+                    dist_weight_velocity_state_dict, pretrain_world_size,
+                    self.world_size)
+                dist_bias_velocity_state_dict = rearrange_weight(
+                    dist_bias_velocity_state_dict, pretrain_world_size,
+                    self.world_size)
+                for name, value in dist_weight_velocity_state_dict.items():
+                    if rank_str in name:
+                        opt_state_dict[name] = value
+                for name, value in dist_bias_velocity_state_dict.items():
+                    if rank_str in name:
+                        opt_state_dict[name] = value
+
+        def map_actual_param_name(state_dict, load_state_dict):
+            for name, param in state_dict.items():
+                state_dict[name] = load_state_dict[param.name]
+            return state_dict
+
+        logging.info("Load checkpoint from '{}'. ".format(checkpoint_dir))
+        param_state_dict = map_actual_param_name(backbone.state_dict(),
+                                                 param_state_dict)
+        backbone.set_state_dict(param_state_dict)
+        if classifier is not None:
+            dist_param_state_dict = map_actual_param_name(
+                classifier.state_dict(), dist_param_state_dict)
+            classifier.set_state_dict(dist_param_state_dict)
+        if for_train:
+            assert optimizer is not None
+            optimizer.set_state_dict(opt_state_dict)
+
+        if classifier is not None and for_train:
+            return extra_info
+        else:
+            return {}
diff --git a/recognition/arcface_paddle/dynamic/utils/verification.py b/recognition/arcface_paddle/dynamic/utils/verification.py
new file mode 100644
index 000000000..a971baa5a
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/utils/verification.py
@@ -0,0 +1,133 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import numpy as np
+import sklearn
+import paddle
+import logging
+from typing import List
+
+from utils.verification import evaluate
+from datasets import load_bin
+
+
+@paddle.no_grad()
+def test(data_set, backbone, batch_size, fp16=False, nfolds=10):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = data[bb - batch_size:bb]
+            # 将numpy转Tensor
+            img = paddle.to_tensor(
+                _data, dtype='float16' if fp16 else 'float32')
+            net_out: paddle.Tensor = backbone(img)
+            _embeddings = net_out.detach().cpu().numpy()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    try:
+        embeddings = sklearn.preprocessing.normalize(embeddings)
+    except:
+        print(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    _, _, accuracy, val, val_std, far = evaluate(
+        embeddings, issame_list, nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+class CallBackVerification(object):
+    def __init__(self,
+                 frequent,
+                 rank,
+                 batch_size,
+                 val_targets,
+                 rec_prefix,
+                 fp16=False,
+                 image_size=(112, 112)):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.batch_size: int = batch_size
+        self.fp16 = fp16
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        if self.rank == 0:
+            self.init_dataset(
+                val_targets=val_targets,
+                data_dir=rec_prefix,
+                image_size=image_size)
+
+    def ver_test(self, backbone: paddle.nn.Layer, global_step: int):
+        for i in range(len(self.ver_list)):
+            test_start = time.time()
+            acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+                self.ver_list[i],
+                backbone,
+                self.batch_size,
+                fp16=self.fp16,
+                nfolds=10)
+            logging.info('[%s][%d]XNorm: %f' %
+                         (self.ver_name_list[i], global_step, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                         (self.ver_name_list[i], global_step, acc2, std2))
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info('[%s][%d]Accuracy-Highest: %1.5f' % (
+                self.ver_name_list[i], global_step, self.highest_acc_list[i]))
+            test_end = time.time()
+            logging.info("test time: {:.4f}".format(test_end - test_start))
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+        for name in val_targets:
+            path = os.path.join(data_dir, name + ".bin")
+            if os.path.exists(path):
+                data_set = load_bin(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+
+    def __call__(self, num_update, backbone: paddle.nn.Layer):
+        if self.rank == 0 and num_update > 0 and num_update % self.frequent == 0:
+            backbone.eval()
+            with paddle.no_grad():
+                self.ver_test(backbone, num_update)
+            backbone.train()
diff --git a/recognition/arcface_paddle/dynamic/validation.py b/recognition/arcface_paddle/dynamic/validation.py
new file mode 100644
index 000000000..7da72ae63
--- /dev/null
+++ b/recognition/arcface_paddle/dynamic/validation.py
@@ -0,0 +1,40 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import numpy as np
+import paddle
+
+from .utils.verification import CallBackVerification
+from .utils.io import Checkpoint
+from . import backbones
+
+
+def validation(args):
+    checkpoint = Checkpoint(
+        rank=0,
+        world_size=1,
+        embedding_size=args.embedding_size,
+        num_classes=None,
+        checkpoint_dir=args.checkpoint_dir, )
+
+    backbone = eval("backbones.{}".format(args.backbone))(
+        num_features=args.embedding_size)
+    checkpoint.load(backbone, for_train=False, dtype='float32')
+    backbone.eval()
+
+    callback_verification = CallBackVerification(
+        1, 0, args.batch_size, args.val_targets, args.data_dir)
+
+    callback_verification(1, backbone)
diff --git a/recognition/arcface_paddle/install_cn.md b/recognition/arcface_paddle/install_cn.md
new file mode 100644
index 000000000..61da350ec
--- /dev/null
+++ b/recognition/arcface_paddle/install_cn.md
@@ -0,0 +1,115 @@
+简体中文 | [English](install_en.md)
+
+# 安装说明
+
+---
+本章将介绍如何安装ArcFace-paddle及其依赖项。
+
+
+## 1. 安装PaddlePaddle
+
+运行ArcFace-paddle需要`PaddlePaddle 2.2.0rc0`或更高版本。可以参考下面的步骤安装PaddlePaddle。
+
+### 1.1 环境要求
+
+- python 3.x
+- cuda >= 10.1 (如果使用paddlepaddle-gpu)
+- cudnn >= 7.6.4 (如果使用paddlepaddle-gpu)
+- nccl >= 2.1.2 (如果使用分布式训练/评估)
+- gcc >= 8.2
+
+建议使用我们提供的docker运行ArcFace-paddle，有关docker、nvidia-docker使用请参考[链接](https://www.runoob.com/docker/docker-tutorial.html)。
+
+在cuda10.1时，建议显卡驱动版本大于等于418.39；在使用cuda10.2时，建议显卡驱动版本大于440.33，更多cuda版本与要求的显卡驱动版本可以参考[链接](https://docs.nvidia.com/deploy/cuda-compatibility/index.html)。
+
+
+如果不使用docker，可以直接跳过1.2部分内容，从1.3部分开始执行。
+
+
+### 1.2 （建议）准备docker环境。第一次使用这个镜像，会自动下载该镜像，请耐心等待。
+
+```
+# 切换到工作目录下
+cd /home/Projects
+# 首次运行需创建一个docker容器，再次运行时不需要运行当前命令
+# 创建一个名字为face_paddle的docker容器，并将当前目录映射到容器的/paddle目录下
+
+如果您希望在CPU环境下使用docker，使用docker而不是nvidia-docker创建docker，设置docker容器共享内存shm-size为8G，建议设置8G以上
+sudo docker run --name face_paddle -v $PWD:/paddle --shm-size=8G --network=host -it paddlepaddle/paddle:2.2.0rc0 /bin/bash
+
+如果希望使用GPU版本的容器，请运行以下命令创建容器。
+sudo nvidia-docker run --name face_paddle -v $PWD:/paddle --shm-size=8G --network=host -it paddlepaddle/paddle:2.2.0rc0-gpu-cuda11.2-cudnn8 /bin/bash
+```
+
+
+您也可以访问[DockerHub](https://hub.docker.com/r/paddlepaddle/paddle/tags/)获取与您机器适配的镜像。
+
+```
+# ctrl+P+Q可退出docker 容器，重新进入docker 容器使用如下命令
+sudo docker exec -it face_paddle /bin/bash
+```
+
+### 1.3 通过pip安装PaddlePaddle
+
+运行下面的命令，通过pip安装最新GPU版本PaddlePaddle
+
+```bash
+pip3 install paddlepaddle-gpu==2.2.0rc0 --upgrade -i https://mirror.baidu.com/pypi/simple
+```
+
+如果希望在CPU环境中使用PaddlePaddle，可以运行下面的命令安装PaddlePaddle。
+
+```bash
+pip3 install paddlepaddle==2.2.0rc0 --upgrade -i https://mirror.baidu.com/pypi/simple
+```
+
+**注意：**
+* 如果先安装了CPU版本的paddlepaddle，之后想切换到GPU版本，那么需要首先卸载CPU版本的paddle，再安装GPU版本的paddle，否则容易导致使用的paddle版本混乱。
+* 您也可以从源码编译安装PaddlePaddle，请参照[PaddlePaddle 安装文档](http://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
+
+
+### 1.4 验证是否安装成功
+
+使用以下命令可以验证PaddlePaddle是否安装成功。
+
+```python
+import paddle
+paddle.utils.run_check()
+```
+
+查看PaddlePaddle版本的命令如下：
+
+```bash
+python3 -c "import paddle; print(paddle.__version__)"
+```
+
+注意：
+- 从源码编译的PaddlePaddle版本号为0.0.0，请确保使用了PaddlePaddle 2.2.0rc0及之后的源码编译。
+- ArcFace-paddle基于PaddlePaddle高性能的分布式训练能力，若您从源码编译，请确保打开编译选项，**WITH_DISTRIBUTE=ON**。具体编译选项参考[编译选项表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/install/Tables.html#id3)。
+- 在docker中运行时，为保证docker容器有足够的共享内存用于Paddle的数据读取加速，在创建docker容器时，请设置参数`--shm_size=8g`，条件允许的话可以设置为更大的值
+- 如果只希望使用识别模块，则可以跳过下面的第3部分；如果只希望使用检测模块，则可以跳过下面的第2部分。
+
+
+## 2. 准备识别模块的环境
+
+安装`requiremnts`，命令如下。
+
+```shell
+pip3 install -r requirement.txt
+```
+
+## 3. 准备检测模块的环境
+
+检测模块依赖于PaddleDetection，需要首先下载PaddleDetection的代码，并安装`requiremnts`。具体命令如下。
+
+```bash
+# 克隆PaddleDetection仓库
+cd <path/to/clone/PaddleDetection>
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+
+cd PaddleDetection
+# 安装其他依赖
+pip3 install -r requirements.txt
+```
+
+更多安装教程，请参考: [Install tutorial](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.2/docs/tutorials/INSTALL_cn.md)。
diff --git a/recognition/arcface_paddle/install_en.md b/recognition/arcface_paddle/install_en.md
new file mode 100644
index 000000000..10ddcc95f
--- /dev/null
+++ b/recognition/arcface_paddle/install_en.md
@@ -0,0 +1,109 @@
+[简体中文](install_cn.md) | English
+
+# Installation
+
+---
+This tutorial introduces how to install ArcFace-paddle and its requirements.
+
+## 1. Install PaddlePaddle
+
+`PaddlePaddle 2.2.0rc0` or later is required for ArcFace-paddle. You can use the following steps to install PaddlePaddle.
+
+### 1.1 Environment requirements
+
+- python 3.x
+- cuda >= 10.1 (necessary if you want to use paddlepaddle-gpu)
+- cudnn >= 7.6.4 (necessary if you want to use paddlepaddle-gpu)
+- nccl >= 2.1.2 (necessary if you want the use distributed training/eval)
+- gcc >= 8.2
+
+Docker is recomended to run ArcFace-paddle, for more detailed information about docker and nvidia-docker, you can refer to the [tutorial](https://www.runoob.com/docker/docker-tutorial.html).
+
+When you use cuda10.1, the driver version needs to be larger or equal than 418.39. When you use cuda10.2, the driver version needs to be larger or equal than 440.33. For more cuda versions and specific driver versions, you can refer to the [link](https://docs.nvidia.com/deploy/cuda-compatibility/index.html).
+
+If you do not want to use docker, you can skip section 1.2 and go into section 1.3 directly.
+
+
+### 1.2 (Recommended) Prepare for a docker environment. The first time you use this docker image, it will be downloaded automatically. Please be patient.
+
+
+```
+# Switch to the working directory
+cd /home/Projects
+# You need to create a docker container for the first run, and do not need to run the current command when you run it again
+# Create a docker container named face_paddle and map the current directory to the /paddle directory of the container
+# It is recommended to set a shared memory greater than or equal to 8G through the --shm-size parameter
+sudo docker run --name face_paddle -v $PWD:/paddle --shm-size=8G --network=host -it paddlepaddle/paddle:2.2.0rc0 /bin/bash
+
+# Use the following command to create a container if you want to use GPU in the container
+sudo nvidia-docker run --name face_paddle -v $PWD:/paddle --shm-size=8G --network=host -it paddlepaddle/paddle:2.2.0rc0-gpu-cuda11.2-cudnn8 /bin/bash
+```
+
+You can also visit [DockerHub](https://hub.docker.com/r/paddlepaddle/paddle/tags/) to get more docker images.
+
+```
+# use ctrl+P+Q to exit docker, to re-enter docker using the following command:
+sudo docker exec -it face_paddle /bin/bash
+```
+
+### 1.3 Install PaddlePaddle using pip
+
+If you want to use PaddlePaddle on GPU, you can use the following command to install PaddlePaddle.
+
+```bash
+pip3 install paddlepaddle-gpu==2.2.0rc0 --upgrade -i https://mirror.baidu.com/pypi/simple
+```
+
+If you want to use PaddlePaddle on CPU, you can use the following command to install PaddlePaddle.
+
+```bash
+pip3 install paddlepaddle==2.2.0rc0 --upgrade -i https://mirror.baidu.com/pypi/simple
+```
+
+**Note:**
+* If you have already installed CPU version of PaddlePaddle and want to use GPU version now, you should uninstall CPU version of PaddlePaddle and then install GPU version to avoid package confusion.
+* You can also compile PaddlePaddle from source code, please refer to [PaddlePaddle Installation tutorial](http://www.paddlepaddle.org.cn/install/quick) to more compilation options.
+
+### 1.4 Verify Installation process
+
+```python
+import paddle
+paddle.utils.run_check()
+```
+
+Check PaddlePaddle version：
+
+```bash
+python3 -c "import paddle; print(paddle.__version__)"
+```
+
+Note:
+- Make sure the compiled source code is later than PaddlePaddle2.0.
+- If you want to enable distribution ability, you should assign **WITH_DISTRIBUTE=ON** when compiling. For more compilation options, please refer to [Instruction](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/install/Tables.html#id3) for more details.
+- When running in docker, in order to ensure that the container has enough shared memory for dataloader acceleration of Paddle, please set the parameter `--shm_size=8g` at creating a docker container, if conditions permit, you can set it to a larger value.
+- If you just want to use recognition module, you can skip section 3. If you just want to use detection module, you can skip section 2.
+
+## 2. Prepare for the environment of recognition
+
+Run the following command to install `requiremnts`.
+
+```shell
+pip3 install -r requirement.txt
+```
+
+## 3. Prepare for the environment of detection
+
+The detection module depends on PaddleDetection. You need to download PaddleDetection and install `requiremnts`, the command is as follows.
+
+
+```bash
+# clone PaddleDetection repo
+cd <path/to/clone/PaddleDetection>
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+
+cd PaddleDetection
+# install requiremnts
+pip3 install -r requirements.txt
+```
+
+For more installation tutorials, please refer to [Install tutorial](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/INSTALL.md).
diff --git a/recognition/arcface_paddle/requirement.txt b/recognition/arcface_paddle/requirement.txt
new file mode 100644
index 000000000..25de974ef
--- /dev/null
+++ b/recognition/arcface_paddle/requirement.txt
@@ -0,0 +1,16 @@
+visualdl
+opencv-python
+pillow
+numpy
+easydict
+scipy
+sklearn
+requests
+prettytable
+tqdm
+Pillow
+scikit-learn==0.23.2
+opencv-python==4.4.0.46
+onnxruntime
+onnx
+paddle2onnx
diff --git a/recognition/arcface_paddle/scripts/export_dynamic.sh b/recognition/arcface_paddle/scripts/export_dynamic.sh
new file mode 100644
index 000000000..cdc0e1461
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/export_dynamic.sh
@@ -0,0 +1,21 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python tools/export.py \
+    --is_static False \
+    --export_type paddle \
+    --backbone FresResNet50 \
+    --embedding_size 512 \
+    --checkpoint_dir MS1M_v3_arcface_dynamic_128_fp16_0.1/FresResNet50/24 \
+    --output_dir MS1M_v3_arcface_dynamic_128_fp16_0.1/FresResNet50/exported_model
diff --git a/recognition/arcface_paddle/scripts/export_static.sh b/recognition/arcface_paddle/scripts/export_static.sh
new file mode 100644
index 000000000..11c1d8ec1
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/export_static.sh
@@ -0,0 +1,21 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python tools/export.py \
+    --is_static True \
+    --export_type paddle \
+    --backbone FresResNet50 \
+    --embedding_size 512 \
+    --checkpoint_dir MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/24 \
+    --output_dir MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/exported_model
diff --git a/recognition/arcface_paddle/scripts/inference.sh b/recognition/arcface_paddle/scripts/inference.sh
new file mode 100644
index 000000000..19698dc7f
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/inference.sh
@@ -0,0 +1,24 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python tools/inference.py \
+    --export_type paddle \
+    --model_file MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/exported_model/FresResNet50.pdmodel \
+    --params_file MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/exported_model/FresResNet50.pdiparams \
+    --image_path MS1M_v3/images/00000001.jpg
+
+python tools/inference.py \
+    --export_type onnx \
+    --onnx_file MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/exported_model/FresResNet50.onnx \
+    --image_path MS1M_v3/images/00000001.jpg
diff --git a/recognition/arcface_paddle/scripts/kill_train_process.sh b/recognition/arcface_paddle/scripts/kill_train_process.sh
new file mode 100644
index 000000000..76dcf50c0
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/kill_train_process.sh
@@ -0,0 +1,15 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+ps -ef | grep "train.py" | grep -v grep | awk '{print $2}' | xargs kill -9
diff --git a/recognition/arcface_paddle/scripts/perf_dynamic.sh b/recognition/arcface_paddle/scripts/perf_dynamic.sh
new file mode 100644
index 000000000..fa1ce24b5
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/perf_dynamic.sh
@@ -0,0 +1,40 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -e
+
+num_test=5
+num_nodes=1
+
+configs=(configs/ms1mv3_r50.py configs/ms1mv3_r100.py)
+dtypes=(fp16 fp32)
+gpus=("0" "0,1,2,3" "0,1,2,3,4,5,6,7")
+
+for config in "${configs[@]}"
+do
+    for dtype in "${dtypes[@]}"
+    do
+        for gpu in "${gpus[@]}"
+        do
+            i=1
+            while [ $i -le ${num_test} ]
+            do
+                bash scripts/perf_runner.sh $gpu $config dynamic 93431 $dtype $num_nodes 128 0.1 ${i}
+                echo " >>>>>>Finished Test Case $config, $dtype, $gpu, ${i} <<<<<<<"
+                let i++
+                sleep 20s
+            done
+        done
+    done
+done
diff --git a/recognition/arcface_paddle/scripts/perf_runner.sh b/recognition/arcface_paddle/scripts/perf_runner.sh
new file mode 100644
index 000000000..9e0b705df
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/perf_runner.sh
@@ -0,0 +1,60 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -ex
+
+gpus=${1:-0,1,2,3,4,5,6,7}
+config_file=${2:-configs/ms1mv3_r50.py}
+mode=${3:-static}
+num_classes=${4:-93431}
+dtype=${5:-fp16}
+num_nodes=${6:-1}
+batch_size_per_device=${7:-128}
+sample_ratio=${8:-0.1}
+test_id=${9:-1}
+
+if [ $mode = "static" ]; then
+    is_static=True
+else
+    is_static=False
+fi
+
+if [ $dtype = "fp16" ]; then
+    fp16=True
+else
+    fp16=False
+fi
+
+if [[ $config_file =~ r50 ]]; then
+    backbone=r50
+else
+    backbone=r100
+fi
+
+gpu_num_per_node=`expr ${#gpus} / 2 + 1`
+
+log_dir=./logs/arcface_paddle_${backbone}_${mode}_${dtype}_r${sample_ratio}_bz${batch_size_per_device}_${num_nodes}n${gpu_num_per_node}g_id${test_id}
+
+python -m paddle.distributed.launch --gpus=${gpus} --log_dir=${log_dir} tools/train.py \
+    --config_file ${config_file} \
+    --is_static ${is_static} \
+    --num_classes ${num_classes} \
+    --fp16 ${fp16} \
+    --sample_ratio ${sample_ratio} \
+    --log_interval_step 1 \
+    --train_unit 'step' \
+    --train_num 200 \
+    --warmup_num 0 \
+    --use_synthetic_dataset True \
+    --do_validation_while_train False
diff --git a/recognition/arcface_paddle/scripts/perf_static.sh b/recognition/arcface_paddle/scripts/perf_static.sh
new file mode 100644
index 000000000..5e9c53b1e
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/perf_static.sh
@@ -0,0 +1,40 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -e
+
+num_test=5
+num_nodes=1
+
+configs=(configs/ms1mv3_r50.py configs/ms1mv3_r100.py)
+dtypes=(fp16 fp32)
+gpus=("0" "0,1,2,3" "0,1,2,3,4,5,6,7")
+
+for config in "${configs[@]}"
+do
+    for dtype in "${dtypes[@]}"
+    do
+        for gpu in "${gpus[@]}"
+        do
+            i=1
+            while [ $i -le ${num_test} ]
+            do
+                bash scripts/perf_runner.sh $gpu $config static 93431 $dtype $num_nodes 128 0.1 ${i}
+                echo " >>>>>>Finished Test Case $config, $dtype, $gpu, ${i} <<<<<<<"
+                let i++
+                sleep 20s
+            done
+        done
+    done
+done
diff --git a/recognition/arcface_paddle/scripts/train_dynamic.sh b/recognition/arcface_paddle/scripts/train_dynamic.sh
new file mode 100644
index 000000000..a62394599
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/train_dynamic.sh
@@ -0,0 +1,41 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 tools/train.py \
+    --config_file configs/ms1mv3_r50.py \
+    --is_static False \
+    --backbone FresResNet50 \
+    --classifier LargeScaleClassifier \
+    --embedding_size 512 \
+    --model_parallel True \
+    --dropout 0.0 \
+    --sample_ratio 0.1 \
+    --loss ArcFace \
+    --batch_size 128 \
+    --dataset MS1M_v3 \
+    --num_classes 93431 \
+    --data_dir MS1M_v3/ \
+    --label_file MS1M_v3/label.txt \
+    --is_bin False \
+    --log_interval_step 100 \
+    --validation_interval_step 2000 \
+    --fp16 True \
+    --use_dynamic_loss_scaling True \
+    --init_loss_scaling 27648.0 \
+    --num_workers 8 \
+    --train_unit 'epoch' \
+    --warmup_num 0 \
+    --train_num 25 \
+    --decay_boundaries "10,16,22" \
+    --output MS1M_v3_arcface_dynamic_0.1
diff --git a/recognition/arcface_paddle/scripts/train_static.sh b/recognition/arcface_paddle/scripts/train_static.sh
new file mode 100644
index 000000000..52ca2be1e
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/train_static.sh
@@ -0,0 +1,41 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 tools/train.py \
+    --config_file configs/ms1mv3_r50.py \
+    --is_static True \
+    --backbone FresResNet50 \
+    --classifier LargeScaleClassifier \
+    --embedding_size 512 \
+    --model_parallel True \
+    --dropout 0.0 \
+    --sample_ratio 0.1 \
+    --loss ArcFace \
+    --batch_size 128 \
+    --dataset MS1M_v3 \
+    --num_classes 93431 \
+    --data_dir MS1M_v3/ \
+    --label_file MS1M_v3/label.txt \
+    --is_bin False \
+    --log_interval_step 100 \
+    --validation_interval_step 2000 \
+    --fp16 True \
+    --use_dynamic_loss_scaling True \
+    --init_loss_scaling 27648.0 \
+    --num_workers 8 \
+    --train_unit 'epoch' \
+    --warmup_num 0 \
+    --train_num 25 \
+    --decay_boundaries "10,16,22" \
+    --output MS1M_v3_arcface_static_0.1
diff --git a/recognition/arcface_paddle/scripts/validation_dynamic.sh b/recognition/arcface_paddle/scripts/validation_dynamic.sh
new file mode 100644
index 000000000..3b635c5f0
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/validation_dynamic.sh
@@ -0,0 +1,22 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python tools/validation.py \
+    --is_static False \
+    --backbone FresResNet50 \
+    --embedding_size 512 \
+    --checkpoint_dir MS1M_v3_arcface_dynamic_128_fp16_0.1/FresResNet50/24 \
+    --data_dir MS1M_v3/ \
+    --val_targets lfw,cfp_fp,agedb_30 \
+    --batch_size 128
diff --git a/recognition/arcface_paddle/scripts/validation_static.sh b/recognition/arcface_paddle/scripts/validation_static.sh
new file mode 100644
index 000000000..64227f7de
--- /dev/null
+++ b/recognition/arcface_paddle/scripts/validation_static.sh
@@ -0,0 +1,22 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python tools/validation.py \
+    --is_static True \
+    --backbone FresResNet50 \
+    --embedding_size 512 \
+    --checkpoint_dir MS1M_v3_arcface_static_128_fp16_0.1/FresResNet50/24 \
+    --data_dir MS1M_v3/ \
+    --val_targets lfw,cfp_fp,agedb_30 \
+    --batch_size 128
diff --git a/recognition/arcface_paddle/static/backbones/__init__.py b/recognition/arcface_paddle/static/backbones/__init__.py
new file mode 100644
index 000000000..1db1164d1
--- /dev/null
+++ b/recognition/arcface_paddle/static/backbones/__init__.py
@@ -0,0 +1,15 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .iresnet import FresResNet50, FresResNet100
diff --git a/recognition/arcface_paddle/static/backbones/iresnet.py b/recognition/arcface_paddle/static/backbones/iresnet.py
new file mode 100644
index 000000000..73e98a934
--- /dev/null
+++ b/recognition/arcface_paddle/static/backbones/iresnet.py
@@ -0,0 +1,249 @@
+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+from collections import OrderedDict
+
+__all__ = [
+    "FresResNet", "FresResNet50", "FresResNet100", "FresResNet101",
+    "FresResNet152"
+]
+
+
+class FresResNet(object):
+    def __init__(self,
+                 layers=50,
+                 num_features=512,
+                 is_train=True,
+                 fp16=False,
+                 fc_type='E',
+                 dropout=0.4):
+        super(FresResNet, self).__init__()
+        self.layers = layers
+        self.num_features = num_features
+        self.fc_type = fc_type
+
+        self.input_dict = OrderedDict()
+        self.output_dict = OrderedDict()
+
+        image = paddle.static.data(
+            name='image',
+            shape=[-1, 3, 112, 112],
+            dtype='float16' if fp16 else 'float32')
+        self.input_dict['image'] = image
+        if is_train:
+            label = paddle.static.data(name='label', shape=[-1], dtype='int32')
+            self.input_dict['label'] = label
+
+        supported_layers = [50, 100, 101, 152]
+        assert layers in supported_layers, \
+            "supported layers {}, but given {}".format(supported_layers, layers)
+
+        if layers == 50:
+            units = [3, 4, 14, 3]
+        elif layers == 100:
+            units = [3, 13, 30, 3]
+        elif layers == 101:
+            units = [3, 4, 23, 3]
+        elif layers == 152:
+            units = [3, 8, 36, 3]
+        filter_list = [64, 64, 128, 256, 512]
+        num_stages = 4
+
+        input_blob = paddle.static.nn.conv2d(
+            input=image,
+            num_filters=filter_list[0],
+            filter_size=3,
+            stride=1,
+            padding=1,
+            groups=1,
+            param_attr=paddle.ParamAttr(),
+            bias_attr=False)
+        input_blob = paddle.static.nn.batch_norm(
+            input=input_blob,
+            act=None,
+            epsilon=1e-05,
+            momentum=0.9,
+            is_test=False if is_train else True)
+        # input_blob = paddle.nn.functional.relu6(input_blob)
+        input_blob = paddle.static.nn.prelu(
+            input_blob,
+            mode="all",
+            param_attr=paddle.ParamAttr(
+                initializer=paddle.nn.initializer.Constant(0.25)))
+
+        for i in range(num_stages):
+            for j in range(units[i]):
+                input_blob = self.residual_unit_v3(
+                    input_blob,
+                    filter_list[i + 1],
+                    3,
+                    2 if j == 0 else 1,
+                    1,
+                    is_train, )
+        fc1 = self.get_fc1(input_blob, is_train, dropout)
+
+        self.output_dict['feature'] = fc1
+
+    def residual_unit_v3(self, in_data, num_filter, filter_size, stride, pad,
+                         is_train):
+
+        bn1 = paddle.static.nn.batch_norm(
+            input=in_data,
+            act=None,
+            epsilon=1e-05,
+            momentum=0.9,
+            is_test=False if is_train else True)
+        conv1 = paddle.static.nn.conv2d(
+            input=bn1,
+            num_filters=num_filter,
+            filter_size=filter_size,
+            stride=1,
+            padding=1,
+            groups=1,
+            param_attr=paddle.ParamAttr(),
+            bias_attr=False)
+        bn2 = paddle.static.nn.batch_norm(
+            input=conv1,
+            act=None,
+            epsilon=1e-05,
+            momentum=0.9,
+            is_test=False if is_train else True)
+        # prelu = paddle.nn.functional.relu6(bn2)
+        prelu = paddle.static.nn.prelu(
+            bn2,
+            mode="all",
+            param_attr=paddle.ParamAttr(
+                initializer=paddle.nn.initializer.Constant(0.25)))
+        conv2 = paddle.static.nn.conv2d(
+            input=prelu,
+            num_filters=num_filter,
+            filter_size=filter_size,
+            stride=stride,
+            padding=pad,
+            groups=1,
+            param_attr=paddle.ParamAttr(),
+            bias_attr=False)
+        bn3 = paddle.static.nn.batch_norm(
+            input=conv2,
+            act=None,
+            epsilon=1e-05,
+            momentum=0.9,
+            is_test=False if is_train else True)
+
+        if stride == 1:
+            input_blob = in_data
+        else:
+            input_blob = paddle.static.nn.conv2d(
+                input=in_data,
+                num_filters=num_filter,
+                filter_size=1,
+                stride=stride,
+                padding=0,
+                groups=1,
+                param_attr=paddle.ParamAttr(),
+                bias_attr=False)
+
+            input_blob = paddle.static.nn.batch_norm(
+                input=input_blob,
+                act=None,
+                epsilon=1e-05,
+                momentum=0.9,
+                is_test=False if is_train else True)
+
+        identity = paddle.add(bn3, input_blob)
+        return identity
+
+    def get_fc1(self, last_conv, is_train, dropout=0.4):
+        body = last_conv
+        if self.fc_type == "Z":
+            body = paddle.static.nn.batch_norm(
+                input=body,
+                act=None,
+                epsilon=1e-05,
+                is_test=False if is_train else True)
+            if dropout > 0:
+                body = paddle.nn.functional.dropout(
+                    x=body,
+                    p=dropout,
+                    training=is_train,
+                    mode='upscale_in_train')
+            fc1 = body
+        elif self.fc_type == "E":
+            body = paddle.static.nn.batch_norm(
+                input=body,
+                act=None,
+                epsilon=1e-05,
+                is_test=False if is_train else True)
+            if dropout > 0:
+                body = paddle.nn.functional.dropout(
+                    x=body,
+                    p=dropout,
+                    training=is_train,
+                    mode='upscale_in_train')
+            fc1 = paddle.static.nn.fc(
+                x=body,
+                size=self.num_features,
+                weight_attr=paddle.ParamAttr(
+                    initializer=paddle.nn.initializer.XavierNormal(
+                        fan_in=0.0)),
+                bias_attr=paddle.ParamAttr(
+                    initializer=paddle.nn.initializer.Constant()))
+            fc1 = paddle.static.nn.batch_norm(
+                input=fc1,
+                act=None,
+                epsilon=1e-05,
+                is_test=False if is_train else True)
+
+        elif self.fc_type == "FC":
+            body = paddle.static.nn.batch_norm(
+                input=body,
+                act=None,
+                epsilon=1e-05,
+                is_test=False if is_train else True)
+            fc1 = paddle.static.nn.fc(
+                x=body,
+                size=self.num_features,
+                weight_attr=paddle.ParamAttr(
+                    initializer=paddle.nn.initializer.XavierNormal(
+                        fan_in=0.0)),
+                bias_attr=paddle.ParamAttr(
+                    initializer=paddle.nn.initializer.Constant()))
+            fc1 = paddle.static.nn.batch_norm(
+                input=fc1,
+                act=None,
+                epsilon=1e-05,
+                is_test=False if is_train else True)
+
+        return fc1
+
+
+def FresResNet50(**args):
+    model = FresResNet(layers=50, **args)
+    return model
+
+
+def FresResNet100(**args):
+    model = FresResNet(layers=100, **args)
+    return model
+
+
+def FresResNet101(**args):
+    model = FresResNet(layers=101, **args)
+    return model
+
+
+def FresResNet152(**args):
+    model = FresResNet(layers=152, **args)
+    return model
diff --git a/recognition/arcface_paddle/static/classifiers/__init__.py b/recognition/arcface_paddle/static/classifiers/__init__.py
new file mode 100644
index 000000000..0c0e52f2b
--- /dev/null
+++ b/recognition/arcface_paddle/static/classifiers/__init__.py
@@ -0,0 +1,15 @@
+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .lsc import LargeScaleClassifier
diff --git a/recognition/arcface_paddle/static/classifiers/lsc.py b/recognition/arcface_paddle/static/classifiers/lsc.py
new file mode 100644
index 000000000..bab66456c
--- /dev/null
+++ b/recognition/arcface_paddle/static/classifiers/lsc.py
@@ -0,0 +1,128 @@
+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from six.moves import reduce
+from collections import OrderedDict
+
+import paddle
+
+__all__ = ["LargeScaleClassifier"]
+
+
+class LargeScaleClassifier(object):
+    """
+    Author: {Xiang An, Yang Xiao, XuHan Zhu} in DeepGlint,
+    Partial FC: Training 10 Million Identities on a Single Machine
+    See the original paper:
+    https://arxiv.org/abs/2010.05222
+    """
+
+    def __init__(self,
+                 feature,
+                 label,
+                 rank,
+                 world_size,
+                 num_classes,
+                 margin1=1.0,
+                 margin2=0.5,
+                 margin3=0.0,
+                 scale=64.0,
+                 sample_ratio=1.0,
+                 embedding_size=512,
+                 name=None):
+        super(LargeScaleClassifier, self).__init__()
+        self.num_classes: int = num_classes
+        self.rank: int = rank
+        self.world_size: int = world_size
+        self.sample_ratio: float = sample_ratio
+        self.embedding_size: int = embedding_size
+        self.num_local: int = (num_classes + world_size - 1) // world_size
+        if num_classes % world_size != 0 and rank == world_size - 1:
+            self.num_local = num_classes % self.num_local
+        self.num_sample: int = int(self.sample_ratio * self.num_local)
+        self.margin1 = margin1
+        self.margin2 = margin2
+        self.margin3 = margin3
+        self.logit_scale = scale
+
+        self.input_dict = OrderedDict()
+        self.input_dict['feature'] = feature
+        self.input_dict['label'] = label
+
+        self.output_dict = OrderedDict()
+
+        if name is None:
+            name = 'dist@fc@rank@%05d.w' % rank
+        assert '.w' in name
+
+        stddev = math.sqrt(2.0 / (self.embedding_size + self.num_local))
+        param_attr = paddle.ParamAttr(
+            initializer=paddle.nn.initializer.Normal(std=stddev))
+
+        weight_dtype = 'float16' if feature.dtype == paddle.float16 else 'float32'
+        weight = paddle.static.create_parameter(
+            shape=[self.embedding_size, self.num_local],
+            dtype=weight_dtype,
+            name=name,
+            attr=param_attr,
+            is_bias=False)
+
+        # avoid allreducing gradients for distributed parameters
+        weight.is_distributed = True
+        # avoid broadcasting distributed parameters in startup program
+        paddle.static.default_startup_program().global_block().vars[
+            weight.name].is_distributed = True
+
+        if self.world_size > 1:
+            feature_list = []
+            paddle.distributed.all_gather(feature_list, feature)
+            total_feature = paddle.concat(feature_list, axis=0)
+
+            label_list = []
+            paddle.distributed.all_gather(label_list, label)
+            total_label = paddle.concat(label_list, axis=0)
+            total_label.stop_gradient = True
+        else:
+            total_feature = feature
+            total_label = label
+
+        total_label.stop_gradient = True
+
+        if self.sample_ratio < 1.0:
+            # partial fc sample process
+            total_label, sampled_class_index = paddle.nn.functional.class_center_sample(
+                total_label, self.num_local, self.num_sample)
+            sampled_class_index.stop_gradient = True
+            weight = paddle.gather(weight, sampled_class_index, axis=1)
+
+        norm_feature = paddle.fluid.layers.l2_normalize(total_feature, axis=1)
+        norm_weight = paddle.fluid.layers.l2_normalize(weight, axis=0)
+
+        local_logit = paddle.matmul(norm_feature, norm_weight)
+
+        loss = paddle.nn.functional.margin_cross_entropy(
+            local_logit,
+            total_label,
+            margin1=self.margin1,
+            margin2=self.margin2,
+            margin3=self.margin3,
+            scale=self.logit_scale,
+            return_softmax=False,
+            reduction=None, )
+
+        loss.desc.set_dtype(paddle.fluid.core.VarDesc.VarType.FP32)
+        loss = paddle.mean(loss)
+
+        self.output_dict['loss'] = loss
diff --git a/recognition/arcface_paddle/static/export.py b/recognition/arcface_paddle/static/export.py
new file mode 100644
index 000000000..e1f3e077b
--- /dev/null
+++ b/recognition/arcface_paddle/static/export.py
@@ -0,0 +1,94 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import errno
+import os
+import numpy as np
+import paddle
+
+from .utils.io import Checkpoint
+from . import backbones
+from .static_model import StaticModel
+
+
+def export_onnx(path_prefix, feed_vars, fetch_vars, executor, program):
+
+    from paddle2onnx.graph import PaddleGraph, ONNXGraph
+    from paddle2onnx.passes import PassManager
+
+    opset_version = 10
+    enable_onnx_checker = True
+    verbose = False
+
+    paddle_graph = PaddleGraph.build_from_program(program, feed_vars,
+                                                  fetch_vars,
+                                                  paddle.fluid.global_scope())
+
+    onnx_graph = ONNXGraph.build(paddle_graph, opset_version, verbose)
+    onnx_graph = PassManager.run_pass(onnx_graph, ['inplace_node_pass'])
+
+    onnx_proto = onnx_graph.export_proto(enable_onnx_checker)
+
+    try:
+        # mkdir may conflict if pserver and trainer are running on the same machine
+        dirname = os.path.dirname(path_prefix)
+        os.makedirs(dirname)
+    except OSError as e:
+        if e.errno != errno.EEXIST:
+            raise
+    model_path = path_prefix + ".onnx"
+    if os.path.isdir(model_path):
+        raise ValueError("'{}' is an existing directory.".format(model_path))
+
+    with open(model_path, 'wb') as f:
+        f.write(onnx_proto.SerializeToString())
+
+
+def export(args):
+    checkpoint = Checkpoint(
+        rank=0,
+        world_size=1,
+        embedding_size=args.embedding_size,
+        num_classes=None,
+        checkpoint_dir=args.checkpoint_dir, )
+
+    test_program = paddle.static.Program()
+    startup_program = paddle.static.Program()
+
+    test_model = StaticModel(
+        main_program=test_program,
+        startup_program=startup_program,
+        backbone_class_name=args.backbone,
+        embedding_size=args.embedding_size,
+        mode='test', )
+
+    gpu_id = int(os.getenv("FLAGS_selected_gpus", 0))
+    place = paddle.CUDAPlace(gpu_id)
+    exe = paddle.static.Executor(place)
+    exe.run(startup_program)
+
+    checkpoint.load(program=test_program, for_train=False, dtype='float32')
+    print("Load checkpoint from '{}'.".format(args.checkpoint_dir))
+
+    path = os.path.join(args.output_dir, args.backbone)
+    if args.export_type == 'onnx':
+        feed_vars = [test_model.backbone.input_dict['image'].name]
+        fetch_vars = [test_model.backbone.output_dict['feature']]
+        export_onnx(path, feed_vars, fetch_vars, exe, program=test_program)
+    else:
+        feed_vars = [test_model.backbone.input_dict['image']]
+        fetch_vars = [test_model.backbone.output_dict['feature']]
+        paddle.static.save_inference_model(
+            path, feed_vars, fetch_vars, exe, program=test_program)
+    print("Save exported model to '{}'.".format(args.output_dir))
diff --git a/recognition/arcface_paddle/static/static_model.py b/recognition/arcface_paddle/static/static_model.py
new file mode 100644
index 000000000..fe8f82a4b
--- /dev/null
+++ b/recognition/arcface_paddle/static/static_model.py
@@ -0,0 +1,159 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import sys
+import numpy as np
+
+import paddle
+from visualdl import LogWriter
+
+from utils.logging import AverageMeter, init_logging, CallBackLogging
+from utils import losses
+
+from .utils.optimization_pass import gather_optimization_pass, amp_pass
+
+from . import classifiers
+from . import backbones
+
+
+class StaticModel(object):
+    def __init__(self,
+                 main_program,
+                 startup_program,
+                 backbone_class_name,
+                 embedding_size,
+                 classifier_class_name=None,
+                 num_classes=None,
+                 sample_ratio=0.1,
+                 lr_scheduler=None,
+                 momentum=0.9,
+                 weight_decay=2e-4,
+                 dropout=0.4,
+                 mode='train',
+                 fp16=False,
+                 fp16_configs=None,
+                 margin_loss_params=None):
+
+        rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
+        world_size = int(os.getenv("PADDLE_TRAINERS_NUM", 1))
+        if world_size > 1:
+            import paddle.distributed.fleet as fleet
+
+        self.main_program = main_program
+        self.startup_program = startup_program
+        self.backbone_class_name = backbone_class_name
+        self.embedding_size = embedding_size
+        self.classifier_class_name = classifier_class_name
+        self.num_classes = num_classes
+        self.sample_ratio = sample_ratio
+        self.lr_scheduler = lr_scheduler
+        self.momentum = momentum
+        self.weight_decay = weight_decay
+        self.mode = mode
+        self.fp16 = fp16
+        self.fp16_configs = fp16_configs
+        self.margin_loss_params = margin_loss_params
+
+        if self.mode == 'train':
+            assert self.classifier_class_name is not None
+            assert self.num_classes is not None
+            assert self.lr_scheduler is not None
+            assert self.margin_loss_params is not None
+            with paddle.static.program_guard(self.main_program,
+                                             self.startup_program):
+                with paddle.utils.unique_name.guard():
+                    self.backbone = eval("backbones.{}".format(
+                        self.backbone_class_name))(
+                            num_features=self.embedding_size,
+                            is_train=True,
+                            fp16=self.fp16,
+                            dropout=dropout)
+                    assert 'label' in self.backbone.input_dict
+                    assert 'feature' in self.backbone.output_dict
+                    self.classifier = eval("classifiers.{}".format(
+                        self.classifier_class_name))(
+                            feature=self.backbone.output_dict['feature'],
+                            label=self.backbone.input_dict['label'],
+                            rank=rank,
+                            world_size=world_size,
+                            num_classes=self.num_classes,
+                            margin1=self.margin_loss_params.margin1,
+                            margin2=self.margin_loss_params.margin2,
+                            margin3=self.margin_loss_params.margin3,
+                            scale=self.margin_loss_params.scale,
+                            sample_ratio=self.sample_ratio,
+                            embedding_size=self.embedding_size)
+                    assert 'loss' in self.classifier.output_dict
+
+                    self.optimizer = paddle.optimizer.Momentum(
+                        learning_rate=self.lr_scheduler,
+                        momentum=self.momentum,
+                        weight_decay=paddle.regularizer.L2Decay(
+                            self.weight_decay))
+                    if self.fp16:
+                        assert self.fp16_configs is not None
+                        self.optimizer = paddle.static.amp.decorate(
+                            optimizer=self.optimizer,
+                            init_loss_scaling=self.fp16_configs[
+                                'init_loss_scaling'],
+                            incr_every_n_steps=self.fp16_configs[
+                                'incr_every_n_steps'],
+                            decr_every_n_nan_or_inf=self.fp16_configs[
+                                'decr_every_n_nan_or_inf'],
+                            incr_ratio=self.fp16_configs['incr_ratio'],
+                            decr_ratio=self.fp16_configs['decr_ratio'],
+                            use_dynamic_loss_scaling=self.fp16_configs[
+                                'use_dynamic_loss_scaling'],
+                            use_pure_fp16=self.fp16_configs['use_pure_fp16'],
+                            amp_lists=paddle.static.amp.
+                            AutoMixedPrecisionLists(
+                                custom_white_list=self.fp16_configs[
+                                    'custom_white_list'],
+                                custom_black_list=self.fp16_configs[
+                                    'custom_black_list'], ),
+                            use_fp16_guard=False)
+
+                    if world_size > 1:
+                        dist_optimizer = fleet.distributed_optimizer(
+                            self.optimizer)
+                        dist_optimizer.minimize(self.classifier.output_dict[
+                            'loss'])
+                    else:
+                        self.optimizer.minimize(self.classifier.output_dict[
+                            'loss'])
+                    if self.fp16:
+                        self.optimizer = self.optimizer._optimizer
+                    if self.sample_ratio < 1.0:
+                        gather_optimization_pass(self.main_program,
+                                                 'dist@fc@rank')
+                    if self.fp16:
+                        amp_pass(self.main_program, 'dist@fc@rank')
+
+        elif self.mode == 'test':
+            with paddle.static.program_guard(self.main_program,
+                                             self.startup_program):
+                with paddle.utils.unique_name.guard():
+                    self.backbone = eval("backbones.{}".format(
+                        self.backbone_class_name))(
+                            num_features=self.embedding_size,
+                            is_train=False,
+                            fp16=self.fp16,
+                            dropout=dropout)
+                    assert 'feature' in self.backbone.output_dict
+
+        else:
+            raise ValueError(
+                "mode is error, only support 'train' and 'test' now.")
diff --git a/recognition/arcface_paddle/static/train.py b/recognition/arcface_paddle/static/train.py
new file mode 100644
index 000000000..7bd40393d
--- /dev/null
+++ b/recognition/arcface_paddle/static/train.py
@@ -0,0 +1,241 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import sys
+import numpy as np
+import logging
+
+import paddle
+from visualdl import LogWriter
+
+from utils.logging import AverageMeter, CallBackLogging
+from datasets import CommonDataset, SyntheticDataset
+from utils import losses
+
+from .utils.verification import CallBackVerification
+from .utils.io import Checkpoint
+
+from . import classifiers
+from . import backbones
+from .static_model import StaticModel
+
+RELATED_FLAGS_SETTING = {
+    'FLAGS_cudnn_exhaustive_search': 1,
+    'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
+    'FLAGS_max_inplace_grad_add': 8,
+    'FLAGS_fraction_of_gpu_memory_to_use': 0.9999,
+}
+paddle.fluid.set_flags(RELATED_FLAGS_SETTING)
+
+
+def train(args):
+
+    writer = LogWriter(logdir=args.logdir)
+
+    rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
+    world_size = int(os.getenv("PADDLE_TRAINERS_NUM", 1))
+
+    gpu_id = int(os.getenv("FLAGS_selected_gpus", 0))
+    place = paddle.CUDAPlace(gpu_id)
+
+    if world_size > 1:
+        import paddle.distributed.fleet as fleet
+        strategy = fleet.DistributedStrategy()
+        strategy.without_graph_optimization = True
+        fleet.init(is_collective=True, strategy=strategy)
+
+    if args.use_synthetic_dataset:
+        trainset = SyntheticDataset(args.num_classes, fp16=args.fp16)
+    else:
+        trainset = CommonDataset(
+            root_dir=args.data_dir,
+            label_file=args.label_file,
+            fp16=args.fp16,
+            is_bin=args.is_bin)
+
+    num_image = len(trainset)
+    total_batch_size = args.batch_size * world_size
+    steps_per_epoch = num_image // total_batch_size
+    if args.train_unit == 'epoch':
+        warmup_steps = steps_per_epoch * args.warmup_num
+        total_steps = steps_per_epoch * args.train_num
+        decay_steps = [x * steps_per_epoch for x in args.decay_boundaries]
+        total_epoch = args.train_num
+    else:
+        warmup_steps = args.warmup_num
+        total_steps = args.train_num
+        decay_steps = [x for x in args.decay_boundaries]
+        total_epoch = (total_steps + steps_per_epoch - 1) // steps_per_epoch
+
+    if rank == 0:
+        logging.info('world_size: {}'.format(world_size))
+        logging.info('total_batch_size: {}'.format(total_batch_size))
+        logging.info('warmup_steps: {}'.format(warmup_steps))
+        logging.info('steps_per_epoch: {}'.format(steps_per_epoch))
+        logging.info('total_steps: {}'.format(total_steps))
+        logging.info('total_epoch: {}'.format(total_epoch))
+        logging.info('decay_steps: {}'.format(decay_steps))
+
+    base_lr = total_batch_size * args.lr / 512
+    lr_scheduler = paddle.optimizer.lr.PiecewiseDecay(
+        boundaries=decay_steps,
+        values=[
+            base_lr * (args.lr_decay**i) for i in range(len(decay_steps) + 1)
+        ])
+    if warmup_steps > 0:
+        lr_scheduler = paddle.optimizer.lr.LinearWarmup(
+            lr_scheduler, warmup_steps, 0, base_lr)
+
+    train_program = paddle.static.Program()
+    test_program = paddle.static.Program()
+    startup_program = paddle.static.Program()
+
+    margin_loss_params = eval("losses.{}".format(args.loss))()
+    train_model = StaticModel(
+        main_program=train_program,
+        startup_program=startup_program,
+        backbone_class_name=args.backbone,
+        embedding_size=args.embedding_size,
+        classifier_class_name=args.classifier,
+        num_classes=args.num_classes,
+        sample_ratio=args.sample_ratio,
+        lr_scheduler=lr_scheduler,
+        momentum=args.momentum,
+        weight_decay=args.weight_decay,
+        dropout=args.dropout,
+        mode='train',
+        fp16=args.fp16,
+        fp16_configs={
+            'init_loss_scaling': args.init_loss_scaling,
+            'incr_every_n_steps': args.incr_every_n_steps,
+            'decr_every_n_nan_or_inf': args.decr_every_n_nan_or_inf,
+            'incr_ratio': args.incr_ratio,
+            'decr_ratio': args.decr_ratio,
+            'use_dynamic_loss_scaling': args.use_dynamic_loss_scaling,
+            'use_pure_fp16': args.fp16,
+            'custom_white_list': args.custom_white_list,
+            'custom_black_list': args.custom_black_list,
+        },
+        margin_loss_params=margin_loss_params, )
+
+    if rank == 0:
+        with open(os.path.join(args.output, 'main_program.txt'), 'w') as f:
+            f.write(str(train_program))
+
+    if rank == 0 and args.do_validation_while_train:
+        test_model = StaticModel(
+            main_program=test_program,
+            startup_program=startup_program,
+            backbone_class_name=args.backbone,
+            embedding_size=args.embedding_size,
+            dropout=args.dropout,
+            mode='test',
+            fp16=args.fp16, )
+
+        callback_verification = CallBackVerification(
+            args.validation_interval_step, rank, args.batch_size, test_program,
+            list(test_model.backbone.input_dict.values()),
+            list(test_model.backbone.output_dict.values()), args.val_targets,
+            args.data_dir)
+
+    callback_logging = CallBackLogging(args.log_interval_step, rank,
+                                       world_size, total_steps,
+                                       args.batch_size, writer)
+    checkpoint = Checkpoint(
+        rank=rank,
+        world_size=world_size,
+        embedding_size=args.embedding_size,
+        num_classes=args.num_classes,
+        model_save_dir=os.path.join(args.output, args.backbone),
+        checkpoint_dir=args.checkpoint_dir,
+        max_num_last_checkpoint=args.max_num_last_checkpoint)
+
+    exe = paddle.static.Executor(place)
+    exe.run(startup_program)
+
+    start_epoch = 0
+    global_step = 0
+    loss_avg = AverageMeter()
+    if args.resume:
+        extra_info = checkpoint.load(program=train_program, for_train=True)
+        start_epoch = extra_info['epoch'] + 1
+        lr_state = extra_info['lr_state']
+        # there last_epoch means last_step in for PiecewiseDecay
+        # since we always use step style for lr_scheduler
+        global_step = lr_state['last_epoch']
+        train_model.lr_scheduler.set_state_dict(lr_state)
+
+    train_loader = paddle.io.DataLoader(
+        trainset,
+        feed_list=list(train_model.backbone.input_dict.values()),
+        places=place,
+        return_list=False,
+        num_workers=args.num_workers,
+        batch_sampler=paddle.io.DistributedBatchSampler(
+            dataset=trainset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True))
+
+    max_loss_scaling = np.array([args.max_loss_scaling]).astype(np.float32)
+    for epoch in range(start_epoch, total_epoch):
+        train_reader_cost = 0.0
+        train_run_cost = 0.0
+        total_samples = 0
+        reader_start = time.time()
+        for step, data in enumerate(train_loader):
+            train_reader_cost += time.time() - reader_start
+            global_step += 1
+            train_start = time.time()
+
+            loss_v = exe.run(
+                train_program,
+                feed=data,
+                fetch_list=[train_model.classifier.output_dict['loss']],
+                use_program_cache=True)
+
+            train_run_cost += time.time() - train_start
+            total_samples += args.batch_size
+
+            loss_avg.update(np.array(loss_v)[0], 1)
+            lr_value = train_model.optimizer.get_lr()
+            callback_logging(
+                global_step, 
+                loss_avg, 
+                epoch, 
+                lr_value,
+                avg_reader_cost=train_reader_cost / args.log_interval_step,
+                avg_batch_cost=(train_reader_cost + train_run_cost) / args.log_interval_step,
+                avg_samples=total_samples / args.log_interval_step,
+                ips=total_samples / (train_reader_cost + train_run_cost))
+            if rank == 0 and args.do_validation_while_train:
+                callback_verification(global_step)
+            train_model.lr_scheduler.step()
+
+            if global_step >= total_steps:
+                break
+            sys.stdout.flush()
+            if rank is 0 and global_step > 0 and global_step % args.log_interval_step == 0:
+                train_reader_cost = 0.0
+                train_run_cost = 0.0
+                total_samples = 0
+            reader_start = time.time()
+        checkpoint.save(
+            train_program,
+            lr_scheduler=train_model.lr_scheduler,
+            epoch=epoch,
+            for_train=True)
+    writer.close()
diff --git a/recognition/arcface_paddle/static/utils/__init__.py b/recognition/arcface_paddle/static/utils/__init__.py
new file mode 100644
index 000000000..185a92b8d
--- /dev/null
+++ b/recognition/arcface_paddle/static/utils/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/recognition/arcface_paddle/static/utils/io.py b/recognition/arcface_paddle/static/utils/io.py
new file mode 100644
index 000000000..acc98ce38
--- /dev/null
+++ b/recognition/arcface_paddle/static/utils/io.py
@@ -0,0 +1,197 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import errno
+import os
+import paddle
+import logging
+import numpy as np
+import shutil
+import json
+from utils.rearrange_weight import rearrange_weight
+
+
+class Checkpoint(object):
+    def __init__(self,
+                 rank,
+                 world_size,
+                 embedding_size,
+                 num_classes,
+                 model_save_dir="./",
+                 checkpoint_dir=None,
+                 max_num_last_checkpoint=3):
+
+        self.rank: int = rank
+        self.world_size: int = world_size
+        self.embedding_size: int = embedding_size
+        self.num_classes: int = num_classes
+        self.model_save_dir: str = model_save_dir
+        self.checkpoint_dir: str = checkpoint_dir
+        self.max_num_last_checkpoint: int = max_num_last_checkpoint
+
+    def save(self, program, lr_scheduler=None, epoch=0, for_train=True):
+        model_save_dir = os.path.join(self.model_save_dir, str(epoch))
+        if not os.path.exists(model_save_dir):
+            # may be more than one processes trying
+            # to create the directory
+            try:
+                os.makedirs(model_save_dir)
+            except OSError as exc:
+                if exc.errno != errno.EEXIST:
+                    raise
+                pass
+
+        param_state_dict = program.state_dict(mode='param')
+        for name, param in param_state_dict.items():
+            # for non dist param, we only save their at rank 0,
+            # but for dist param, we need to save their at all ranks.
+            if 'dist@' in name and '@rank@' in name or self.rank == 0:
+                paddle.save(param,
+                            os.path.join(model_save_dir, name + '.pdparam'))
+
+        if for_train:
+            opt_state_dict = program.state_dict(mode='opt')
+            for name, opt in opt_state_dict.items():
+                if '@GRAD' in name:
+                    continue
+                # for non dist opt var, we only save their at rank 0,
+                # but for dist opt var, we need to save their at all ranks.
+                if 'dist@' in name and '@rank@' in name or self.rank == 0:
+                    paddle.save(opt,
+                                os.path.join(model_save_dir, name + '.pdopt'))
+
+            if self.rank == 0:
+                # save some extra info for resume
+                # pretrain_world_size, embedding_size, num_classes are used for
+                # re-split fc weight when gpu setting changed.
+                # epoch use to restart.
+                config_file = os.path.join(model_save_dir, 'meta.json')
+                extra_info = dict()
+                extra_info["pretrain_world_size"] = self.world_size
+                extra_info["embedding_size"] = self.embedding_size
+                extra_info['num_classes'] = self.num_classes
+                extra_info['epoch'] = epoch
+                extra_info['lr_state'] = lr_scheduler.state_dict()
+                with open(config_file, 'w') as f:
+                    json.dump(extra_info, f)
+
+        logging.info("Save model to {}.".format(model_save_dir))
+        if self.rank == 0 and self.max_num_last_checkpoint > 0:
+            for idx in range(-1, epoch - self.max_num_last_checkpoint + 1):
+                path = os.path.join(self.model_save_dir, str(idx))
+                if os.path.exists(path):
+                    logging.info("Remove checkpoint {}.".format(path))
+                    shutil.rmtree(path)
+
+    def load(self, program, for_train=True, dtype=None):
+        assert os.path.exists(self.checkpoint_dir)
+        checkpoint_dir = os.path.abspath(self.checkpoint_dir)
+
+        state_dict = {}
+        dist_weight_state_dict = {}
+        dist_weight_velocity_state_dict = {}
+        dist_bias_state_dict = {}
+        dist_bias_velocity_state_dict = {}
+        for path in os.listdir(checkpoint_dir):
+            path = os.path.join(checkpoint_dir, path)
+            if not os.path.isfile(path):
+                continue
+
+            basename = os.path.basename(path)
+            name, ext = os.path.splitext(basename)
+
+            if ext not in ['.pdopt', '.pdparam']:
+                continue
+
+            if not for_train and ext == '.pdopt':
+                continue
+
+            tensor = paddle.load(path, return_numpy=True)
+            if dtype:
+                assert dtype in ['float32', 'float16']
+                tensor = tensor.astype(dtype)
+
+            if 'dist@' in name and '@rank@' in name:
+                if '.w' in name and 'velocity' not in name:
+                    dist_weight_state_dict[name] = tensor
+                elif '.w' in name and 'velocity' in name:
+                    dist_weight_velocity_state_dict[name] = tensor
+                elif '.b' in name and 'velocity' not in name:
+                    dist_bias_state_dict[name] = tensor
+                elif '.b' in name and 'velocity' in name:
+                    dist_bias_velocity_state_dict[name] = tensor
+
+            else:
+                state_dict[name] = tensor
+
+        if for_train:
+            meta_file = os.path.join(checkpoint_dir, 'meta.json')
+            if not os.path.exists(meta_file):
+                logging.error(
+                    "Please make sure the checkpoint dir {} exists, and "
+                    "parameters in that dir are validating.".format(
+                        checkpoint_dir))
+                exit()
+
+            with open(meta_file, 'r') as handle:
+                extra_info = json.load(handle)
+
+            # Preporcess distributed parameters.
+            pretrain_world_size = extra_info['pretrain_world_size']
+            assert pretrain_world_size > 0
+            embedding_size = extra_info['embedding_size']
+            assert embedding_size == self.embedding_size
+            num_classes = extra_info['num_classes']
+            assert num_classes == self.num_classes
+
+            logging.info(
+                "Parameters for pre-training: pretrain_world_size ({}), "
+                "embedding_size ({}), and num_classes ({}).".format(
+                    pretrain_world_size, embedding_size, num_classes))
+            logging.info("Parameters for inference or fine-tuning: "
+                         "world_size ({}).".format(self.world_size))
+
+            rank_str = '%05d' % self.rank
+
+            dist_weight_state_dict = rearrange_weight(
+                dist_weight_state_dict, pretrain_world_size, self.world_size)
+            dist_bias_state_dict = rearrange_weight(
+                dist_bias_state_dict, pretrain_world_size, self.world_size)
+            for name, value in dist_weight_state_dict.items():
+                if rank_str in name:
+                    state_dict[name] = value
+            for name, value in dist_bias_state_dict.items():
+                if rank_str in name:
+                    state_dict[name] = value
+
+            if for_train:
+                dist_weight_velocity_state_dict = rearrange_weight(
+                    dist_weight_velocity_state_dict, pretrain_world_size,
+                    self.world_size)
+                dist_bias_velocity_state_dict = rearrange_weight(
+                    dist_bias_velocity_state_dict, pretrain_world_size,
+                    self.world_size)
+                for name, value in dist_weight_velocity_state_dict.items():
+                    if rank_str in name:
+                        state_dict[name] = value
+                for name, value in dist_bias_velocity_state_dict.items():
+                    if rank_str in name:
+                        state_dict[name] = value
+
+        program.set_state_dict(state_dict)
+        logging.info("Load checkpoint from '{}'. ".format(checkpoint_dir))
+        if for_train:
+            return extra_info
+        else:
+            return {}
diff --git a/recognition/arcface_paddle/static/utils/optimization_pass.py b/recognition/arcface_paddle/static/utils/optimization_pass.py
new file mode 100644
index 000000000..817fb6f3a
--- /dev/null
+++ b/recognition/arcface_paddle/static/utils/optimization_pass.py
@@ -0,0 +1,124 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+def check_contains(name, name_list):
+    for n in name_list:
+        if name in n:
+            return True
+    return False
+
+
+def gather_optimization_pass(program, weight_name):
+    op_idxs = []
+    gather_grad_op = None
+    momentum_op = None
+    for idx, op in enumerate(program.global_block().ops):
+        if (op.type == 'gather_grad' or
+                op.type == 'momentum') and check_contains(weight_name,
+                                                          op.input_arg_names):
+            op_idxs.append(idx)
+            if op.type == 'momentum':
+                momentum_op = op
+            if op.type == 'gather_grad':
+                gather_grad_op = op
+
+    if gather_grad_op is not None and momentum_op is not None:
+        inputs = {
+            'Param': momentum_op.input('Param'),
+            'Velocity': momentum_op.input('Velocity'),
+            'LearningRate': momentum_op.input('LearningRate'),
+            'Grad': gather_grad_op.input('Out@GRAD'),
+            'Index': gather_grad_op.input('Index'),
+            'Axis': gather_grad_op.input('Axis'),
+        }
+        outputs = {
+            'ParamOut': momentum_op.output('ParamOut'),
+            'VelocityOut': momentum_op.output('VelocityOut'),
+        }
+        if 'MasterParam' in momentum_op.input_names and len(
+                momentum_op.input('MasterParam')) > 0:
+            inputs['MasterParam'] = momentum_op.input('MasterParam')
+        if 'MasterParamOut' in momentum_op.output_names and len(
+                momentum_op.output('MasterParamOut')) > 0:
+            outputs['MasterParamOut'] = momentum_op.output('MasterParamOut')
+
+        attrs = {
+            'mu': momentum_op.attr('mu'),
+            'use_nesterov': momentum_op.attr('use_nesterov'),
+            'regularization_method': momentum_op.attr('regularization_method'),
+            'regularization_coeff': momentum_op.attr('regularization_coeff'),
+            'multi_precision': momentum_op.attr('multi_precision'),
+            'rescale_grad': momentum_op.attr('rescale_grad'),
+            'op_device': momentum_op.attr('op_device'),
+            'op_namescope': momentum_op.attr('op_namescope'),
+            'op_role': momentum_op.attr('op_role'),
+            'op_role_var': momentum_op.input('Param'),
+            'axis': gather_grad_op.attr('axis'),
+        }
+        program.global_block()._insert_op(
+            op_idxs[-1] + 1,
+            type='sparse_momentum',
+            inputs=inputs,
+            outputs=outputs,
+            attrs=attrs)
+
+        for idx in reversed(op_idxs):
+            program.global_block()._remove_op(idx, sync=False)
+
+        var_names = []
+        for idx, name in enumerate(program.global_block().vars):
+            if '@GRAD' in name and weight_name in name:
+                var_names.append(name)
+        for name in var_names:
+            program.global_block()._remove_var(name, sync=False)
+        program.global_block()._sync_with_cpp()
+
+
+def amp_pass(program, weight_name):
+    for idx, op in enumerate(program.global_block().ops):
+        if (op.type == 'update_loss_scaling' or
+                op.type == 'check_finite_and_unscale'):
+            input_idxs = []
+            input_arg_names = op.input("X")
+            # input_arg_names.append(gather_grad_op.input('Out@GRAD')[0])
+            for i, name in enumerate(input_arg_names):
+                if '@GRAD' in name and weight_name in name:
+                    input_idxs.append(i)
+            if len(input_idxs) > 0:
+                for i in reversed(input_idxs):
+                    input_arg_names.pop(i)
+                op.desc.set_input("X", input_arg_names)
+
+            output_idxs = []
+            output_arg_names = op.output("Out")
+            # output_arg_names.append(gather_grad_op.input('Out@GRAD')[0])
+            for i, name in enumerate(output_arg_names):
+                if '@GRAD' in name and weight_name in name:
+                    output_idxs.append(i)
+            if len(output_idxs) > 0:
+                for i in reversed(output_idxs):
+                    output_arg_names.pop(i)
+                op.desc.set_output("Out", output_arg_names)
+
+            if op.type == 'check_finite_and_unscale':
+                op_role_idxs = []
+                op_role_var = op.attr("op_role_var")
+                for i, name in enumerate(op_role_var):
+                    if '@GRAD' in name and weight_name in name:
+                        op_role_idxs.append(i)
+                if len(op_role_idxs) > 0:
+                    for i in reversed(op_role_idxs):
+                        op_role_var.pop(i)
+                    op.desc._set_attr("op_role_var", op_role_var)
diff --git a/recognition/arcface_paddle/static/utils/verification.py b/recognition/arcface_paddle/static/utils/verification.py
new file mode 100644
index 000000000..102ebf98a
--- /dev/null
+++ b/recognition/arcface_paddle/static/utils/verification.py
@@ -0,0 +1,130 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import numpy as np
+import sklearn
+import paddle
+import logging
+
+from utils.verification import evaluate
+from datasets import load_bin
+
+
+def test(rank, batch_size, data_set, executor, test_program, data_feeder,
+         fetch_list):
+
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+
+    # data_list[0] for normalize
+    # data_list[1] for flip_left_right
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = []
+            for k in range(bb - batch_size, bb):
+                _data.append((data[k], ))
+            [_embeddings] = executor.run(test_program,
+                                         fetch_list=fetch_list,
+                                         feed=data_feeder.feed(_data),
+                                         use_program_cache=True)
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    xnorm = 0.0
+    xnorm_cnt = 0
+    for embed in embeddings_list:
+        xnorm += np.sqrt((embed * embed).sum(axis=1)).sum(axis=0)
+        xnorm_cnt += embed.shape[0]
+    xnorm /= xnorm_cnt
+
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    _, _, accuracy, val, val_std, far = evaluate(
+        embeddings, issame_list, nrof_folds=10)
+    acc, std = np.mean(accuracy), np.std(accuracy)
+    return acc, std, xnorm
+
+
+class CallBackVerification(object):
+    def __init__(self,
+                 frequent,
+                 rank,
+                 batch_size,
+                 test_program,
+                 feed_list,
+                 fetch_list,
+                 val_targets,
+                 rec_prefix,
+                 image_size=(112, 112)):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.batch_size: int = batch_size
+
+        self.test_program: paddle.static.Program = test_program
+        self.feed_list: List[paddle.fluid.framework.Variable] = feed_list
+        self.fetch_list: List[paddle.fluid.framework.Variable] = fetch_list
+
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        self.init_dataset(
+            val_targets=val_targets,
+            data_dir=rec_prefix,
+            image_size=image_size)
+
+        gpu_id = int(os.getenv("FLAGS_selected_gpus", 0))
+        place = paddle.CUDAPlace(gpu_id)
+        self.executor = paddle.static.Executor(place)
+        self.data_feeder = paddle.fluid.DataFeeder(
+            place=place, feed_list=self.feed_list, program=self.test_program)
+
+    def ver_test(self, global_step: int):
+        for i in range(len(self.ver_list)):
+            test_start = time.time()
+            acc2, std2, xnorm = test(
+                self.rank, self.batch_size, self.ver_list[i], self.executor,
+                self.test_program, self.data_feeder, self.fetch_list)
+            logging.info('[%s][%d]XNorm: %f' %
+                         (self.ver_name_list[i], global_step, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                         (self.ver_name_list[i], global_step, acc2, std2))
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info('[%s][%d]Accuracy-Highest: %1.5f' % (
+                self.ver_name_list[i], global_step, self.highest_acc_list[i]))
+            test_end = time.time()
+            logging.info("test time: {:.4f}".format(test_end - test_start))
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+        for name in val_targets:
+            path = os.path.join(data_dir, name + ".bin")
+            if os.path.exists(path):
+                data_set = load_bin(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+
+    def __call__(self, num_update):
+        if self.rank == 0 and num_update > 0 and num_update % self.frequent == 0:
+            self.ver_test(num_update)
diff --git a/recognition/arcface_paddle/static/validation.py b/recognition/arcface_paddle/static/validation.py
new file mode 100644
index 000000000..f419b494d
--- /dev/null
+++ b/recognition/arcface_paddle/static/validation.py
@@ -0,0 +1,58 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import errno
+import os
+import numpy as np
+import paddle
+
+from .utils.verification import CallBackVerification
+from .utils.io import Checkpoint
+from .static_model import StaticModel
+
+from . import backbones
+
+
+def validation(args):
+    checkpoint = Checkpoint(
+        rank=0,
+        world_size=1,
+        embedding_size=args.embedding_size,
+        num_classes=None,
+        checkpoint_dir=args.checkpoint_dir, )
+
+    test_program = paddle.static.Program()
+    startup_program = paddle.static.Program()
+
+    test_model = StaticModel(
+        main_program=test_program,
+        startup_program=startup_program,
+        backbone_class_name=args.backbone,
+        embedding_size=args.embedding_size,
+        mode='test', )
+
+    gpu_id = int(os.getenv("FLAGS_selected_gpus", 0))
+    place = paddle.CUDAPlace(gpu_id)
+    exe = paddle.static.Executor(place)
+    exe.run(startup_program)
+
+    checkpoint.load(program=test_program, for_train=False, dtype='float32')
+
+    callback_verification = CallBackVerification(
+        1, 0, args.batch_size, test_program,
+        list(test_model.backbone.input_dict.values()),
+        list(test_model.backbone.output_dict.values()), args.val_targets,
+        args.data_dir)
+
+    callback_verification(1)
diff --git a/recognition/arcface_paddle/test_tipc/common_func.sh b/recognition/arcface_paddle/test_tipc/common_func.sh
new file mode 100644
index 000000000..2c039e964
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/common_func.sh
@@ -0,0 +1,64 @@
+#!/bin/bash
+
+function func_parser_key(){
+    strs=$1
+    IFS=":"
+    array=(${strs})
+    tmp=${array[0]}
+    echo ${tmp}
+}
+
+function func_parser_value(){
+    strs=$1
+    IFS=":"
+    array=(${strs})
+    tmp=${array[1]}
+    echo ${tmp}
+}
+
+function func_set_params(){
+    key=$1
+    value=$2
+    if [ ${key}x = "null"x ];then
+        echo " "
+    elif [[ ${value} = "null" ]] || [[ ${value} = " " ]] || [ ${#value} -le 0 ];then
+        echo " "
+    else 
+        echo "${key}=${value}"
+    fi
+}
+
+function func_parser_params(){
+    strs=$1
+    IFS=":"
+    array=(${strs})
+    key=${array[0]}
+    tmp=${array[1]}
+    IFS="|"
+    res=""
+    for _params in ${tmp[*]}; do
+        IFS="="
+        array=(${_params})
+        mode=${array[0]}
+        value=${array[1]}
+        if [[ ${mode} = ${MODE} ]]; then
+            IFS="|"
+            #echo $(func_set_params "${mode}" "${value}")
+            echo $value
+            break
+        fi
+        IFS="|"
+    done
+    echo ${res}
+}
+
+function status_check(){
+    last_status=$1   # the exit code
+    run_command=$2
+    run_log=$3
+    if [ $last_status -eq 0 ]; then
+        echo -e "\033[33m Run successfully with command - ${run_command}!  \033[0m" | tee -a ${run_log}
+    else
+        echo -e "\033[33m Run failed with command - ${run_command}!  \033[0m" | tee -a ${run_log}
+    fi
+}
diff --git a/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt b/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt
new file mode 100644
index 000000000..14ad02e7b
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt
@@ -0,0 +1,17 @@
+===========================serving_params===========================
+model_name:arcfae_mobileface
+python:python3.7
+trans_model:-m paddle_serving_client.convert
+--dirname:./inference/
+--model_filename:inference.pdmodel
+--params_filename:inference.pdiparams
+--serving_server:./deploy/pdserving/MobileFaceNet_128_serving/
+--serving_client:./deploy/pdserving/MobileFaceNet_128_client/
+serving_dir:./deploy/pdserving
+web_service:web_service.py --config=config.yml --opt op.ArcFace.concurrency=1
+op.ArcFace.local_service_conf.devices:null|0
+op.ArcFace.local_service_conf.use_mkldnn:True|False
+op.ArcFace.local_service_conf.thread_num:1|6
+op.ArcFace.local_service_conf.use_trt:False|True
+op.ArcFace.local_service_conf.precision:fp32
+pipline:pipeline_http_client.py --image_dir=./imgs
\ No newline at end of file
diff --git a/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/train_infer_python.txt b/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/train_infer_python.txt
new file mode 100644
index 000000000..1034c6959
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/configs/ms1mv2_mobileface/train_infer_python.txt
@@ -0,0 +1,51 @@
+===========================train_params===========================
+model_name:MobileFaceNet_128
+python:python3.7
+gpu_list:0
+null:null
+null:null
+--train_num:lite_train_lite_infer=1
+--output:./output/
+--batch_size:2
+--checkpoint_dir:null
+train_model_name:MobileFaceNet_128/0
+train_infer_img_dir:./MS1M_v2/images
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py --config_file=configs/ms1mv2_mobileface.py --is_static=False --fp16=False --embedding_size=128 --fp16=False --dataset=MS1M_v2 --data_dir=MS1M_v2/ --label_file=MS1M_v2/label.txt --num_classes=85742 --log_interval_step=1
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/validation.py --is_static=False --backbone=MobileFaceNet_128 --embedding_size=128 --data_dir=MS1M_v2 --val_targets=lfw --batch_size=2
+null:null
+##
+===========================infer_params===========================
+--output_dir:null
+--checkpoint_dir:null
+norm_export:tools/export.py --is_static=False --export_type=paddle --backbone=MobileFaceNet_128 --embedding_size=128
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+##
+train_model:./inference
+infer_export:null
+infer_quant:False
+inference:tools/inference.py  --export_type=paddle --benchmark=True
+--use_gpu:True|False
+--enable_mkldnn:True|False
+--cpu_threads:1|6
+--max_batch_size:1
+--use_tensorrt:False|True
+--precision:fp32
+--model_dir:null
+--image_path:MS1M_v2/images
+null:null
+--benchmark:null
+null:null
\ No newline at end of file
diff --git a/recognition/arcface_paddle/test_tipc/data/small_dataset.tar b/recognition/arcface_paddle/test_tipc/data/small_dataset.tar
new file mode 100644
index 000000000..66f753c87
Binary files /dev/null and b/recognition/arcface_paddle/test_tipc/data/small_dataset.tar differ
diff --git a/recognition/arcface_paddle/test_tipc/data/small_lfw.bin b/recognition/arcface_paddle/test_tipc/data/small_lfw.bin
new file mode 100644
index 000000000..f65c9de0b
Binary files /dev/null and b/recognition/arcface_paddle/test_tipc/data/small_lfw.bin differ
diff --git a/recognition/arcface_paddle/test_tipc/docs/guide.png b/recognition/arcface_paddle/test_tipc/docs/guide.png
new file mode 100644
index 000000000..319ac819d
Binary files /dev/null and b/recognition/arcface_paddle/test_tipc/docs/guide.png differ
diff --git a/recognition/arcface_paddle/test_tipc/docs/install.md b/recognition/arcface_paddle/test_tipc/docs/install.md
new file mode 100644
index 000000000..493ab22c7
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/docs/install.md
@@ -0,0 +1,121 @@
+## 1. 环境准备
+
+本教程适用于TIPC目录下基础功能测试的运行环境搭建。
+
+推荐环境：
+- CUDA 10.1/10.2
+- CUDNN 7.6/cudnn8.1
+- TensorRT 6.1.0.5 / 7.1 / 7.2
+
+环境配置可以选择docker镜像安装，或者在本地环境Python搭建环境。推荐使用docker镜像安装，避免不必要的环境配置。
+
+## 2. Docker 镜像安装
+
+推荐docker镜像安装，按照如下命令创建镜像，当前目录映射到镜像中的`/paddle`目录下
+```
+nvidia-docker run --name paddle -it -v $PWD:/paddle paddlepaddle/paddle:latest-dev-cuda10.1-cudnn7-gcc82 /bin/bash
+cd /paddle
+
+# 安装带TRT的paddle
+pip3.7 install https://paddle-wheel.bj.bcebos.com/with-trt/2.1.3/linux-gpu-cuda10.1-cudnn7-mkl-gcc8.2-trt6-avx/paddlepaddle_gpu-2.1.3.post101-cp37-cp37m-linux_x86_64.whl
+```
+
+## 3 Python 环境构建
+
+非docker环境下，环境配置比较灵活，推荐环境组合配置：
+- CUDA10.1 + CUDNN7.6 + TensorRT 6
+- CUDA10.2 + CUDNN8.1 + TensorRT 7
+- CUDA11.1 + CUDNN8.1 + TensorRT 7
+
+下面以 CUDA10.2 + CUDNN8.1 + TensorRT 7 配置为例，介绍环境配置的流程。
+
+### 3.1 安装CUDNN
+
+如果当前环境满足CUDNN版本的要求，可以跳过此步骤。
+
+以CUDNN8.1 安装安装为例，安装步骤如下，首先下载CUDNN，从[Nvidia官网](https://developer.nvidia.com/rdp/cudnn-archive)下载CUDNN8.1版本，下载符合当前系统版本的三个deb文件，分别是：
+- cuDNN Runtime Library ，如：libcudnn8_8.1.0.77-1+cuda10.2_amd64.deb
+- cuDNN Developer Library ，如：libcudnn8-dev_8.1.0.77-1+cuda10.2_amd64.deb
+- cuDNN Code Samples，如：libcudnn8-samples_8.1.0.77-1+cuda10.2_amd64.deb
+
+deb安装可以参考[官方文档](https://docs.nvidia.com/deeplearning/cudnn/install-guide/index.html#installlinux-deb)，安装方式如下
+```
+# x.x.x表示下载的版本号
+# $HOME为工作目录
+sudo dpkg -i libcudnn8_x.x.x-1+cudax.x_arm64.deb
+sudo dpkg -i libcudnn8-dev_8.x.x.x-1+cudax.x_arm64.deb
+sudo dpkg -i libcudnn8-samples_8.x.x.x-1+cudax.x_arm64.deb
+
+# 验证是否正确安装
+cp -r /usr/src/cudnn_samples_v8/ $HOME
+cd  $HOME/cudnn_samples_v8/mnistCUDNN
+
+# 编译
+make clean && make
+./mnistCUDNN
+```
+如果运行mnistCUDNN完后提示运行成功，则表示安装成功。如果运行后出现freeimage相关的报错，需要按照提示安装freeimage库:
+```
+sudo apt-get install libfreeimage-dev
+sudo apt-get install libfreeimage
+```
+
+### 3.2 安装TensorRT
+
+首先，从[Nvidia官网TensorRT板块](https://developer.nvidia.com/tensorrt-getting-started)下载TensorRT，这里选择7.1.3.4版本的TensorRT，注意选择适合自己系统版本和CUDA版本的TensorRT，另外建议下载TAR package的安装包。
+
+以Ubuntu16.04+CUDA10.2为例，下载并解压后可以参考[官方文档](https://docs.nvidia.com/deeplearning/tensorrt/archives/tensorrt-713/install-guide/index.html#installing-tar)的安装步骤，按照如下步骤安装:
+```
+# 以下安装命令中 '${version}' 为下载的TensorRT版本，如7.1.3.4
+# 设置环境变量，<TensorRT-${version}/lib> 为解压后的TensorRT的lib目录
+export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<TensorRT-${version}/lib>
+
+# 安装TensorRT
+cd TensorRT-${version}/python
+pip3.7 install tensorrt-*-cp3x-none-linux_x86_64.whl
+
+# 安装graphsurgeon
+cd TensorRT-${version}/graphsurgeon
+```
+
+
+### 3.3 安装PaddlePaddle
+
+下载支持TensorRT版本的Paddle安装包，注意安装包的TensorRT版本需要与本地TensorRT一致，下载[链接](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#python)
+选择下载 linux-cuda10.2-trt7-gcc8.2 Python3.7版本的Paddle：
+```
+# 从下载链接中可以看到是paddle2.1.1-cuda10.2-cudnn8.1版本
+wget  https://paddle-wheel.bj.bcebos.com/with-trt/2.1.1-gpu-cuda10.2-cudnn8.1-mkl-gcc8.2/paddlepaddle_gpu-2.1.1-cp37-cp37m-linux_x86_64.whl
+pip3.7 install -U paddlepaddle_gpu-2.1.1-cp37-cp37m-linux_x86_64.whl
+```
+
+## 4. 安装依赖
+```
+# 安装AutoLog
+git clone https://github.com/LDOUBLEV/AutoLog
+cd AutoLog
+pip3.7 install -r requirements.txt
+python3.7 setup.py bdist_wheel
+pip3.7 install ./dist/auto_log-1.0.0-py3-none-any.whl
+
+# 下载insightface代码
+cd ../
+git clone https://github.com/deepinsight/insightface
+
+```
+
+安装Arcface依赖：
+```
+cd insightface/recognition/arcface_paddle
+pip3.7 install -r requirements.txt
+```
+
+## FAQ :
+Q. You are using Paddle compiled with TensorRT, but TensorRT dynamic library is not found. Ignore this if TensorRT is not needed.
+
+A. 问题一般是当前安装paddle版本带TRT，但是本地环境找不到TensorRT的预测库，需要下载TensorRT库，解压后设置环境变量LD_LIBRARY_PATH;
+如：
+```
+export LD_LIBRARY_PATH=/usr/local/python3.7.0/lib:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/paddle/package/TensorRT-6.0.1.5/lib
+```
+或者问题是下载的TensorRT版本和当前paddle中编译的TRT版本不匹配，需要下载版本相符的TensorRT重新安装。
diff --git a/recognition/arcface_paddle/test_tipc/docs/test.png b/recognition/arcface_paddle/test_tipc/docs/test.png
new file mode 100644
index 000000000..f99f23d70
Binary files /dev/null and b/recognition/arcface_paddle/test_tipc/docs/test.png differ
diff --git a/recognition/arcface_paddle/test_tipc/docs/test_serving.md b/recognition/arcface_paddle/test_tipc/docs/test_serving.md
new file mode 100644
index 000000000..2b262d05d
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/docs/test_serving.md
@@ -0,0 +1,53 @@
+# PaddleServing预测功能测试
+
+PaddleServing预测功能测试的主程序为`test_serving.sh`，可以测试基于PaddleServing的部署功能。
+
+## 1. 测试结论汇总
+
+本repo未提供量化训练，因此这里只测试正常模型，对应的PaddleServing预测功能汇总如下：
+
+| 模型类型 |device | batchsize | tensorrt | mkldnn | cpu多线程 |
+|  :----:   |  :----: |   :----:   |  :----:  |   :----:   |  :----:  |
+| 正常模型 | GPU | 1/6 | fp32 | - | - |
+| 正常模型 | CPU | 1/6 | - | fp32 | 支持 |
+
+## 2. 测试流程
+### 2.1 功能测试
+先运行`prepare.sh`准备数据和模型，然后运行`test_serving.sh`进行测试，最终在```test_tipc/output```目录下生成`serving_infer_*.log`后缀的日志文件。
+
+```shell
+bash test_tipc/prepare.sh ./test_tipc/configs/ms1mv2_mobileface/model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt "serving_infer"
+
+# 用法:
+bash test_tipc/test_serving.sh ./test_tipc/configs//ms1mv2_mobileface/model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt
+```  
+
+#### 运行结果
+
+各测试的运行情况会打印在 `test_tipc/output/results_serving.log` 中：
+运行成功时会输出：
+
+```
+Run successfully with command - python3.7 pipeline_http_client.py --image_dir=./imgs > ../../test_tipc/output/server_infer_cpu_usemkldnn_True_threads_1_batchsize_1.log 2>&1 ! 
+Run successfully  with command - xxxxx
+...
+```
+
+运行失败时会输出：
+
+```
+Run failed with command - python3.7 pipeline_http_client.py --image_dir=./imgs > ../../test_tipc/output/server_infer_cpu_usemkldnn_True_threads_1_batchsize_1.log 2>&1 !
+Run failed with command - xxxxx
+...
+```
+
+详细的预测结果会存在 test_tipc/output/ 文件夹下，例如`server_infer_gpu_usetrt_True_precision_fp32_batchsize_1.log`中会返回图像的特征值:
+
+```
+{'err_no': 0, 'err_msg': '', 'key': ['out'], 'value': ['array([[ 1.36603206e-01, -2.12395296e-01, -3.94680113e-01,\n        -3.14380080e-01, -9.66617092e-03,  1.87318385e-01,\n        -2.97903419e-01, -3.17218006e-01, -2.69029588e-01,\n         1.21175185e-01, -1.90171480e-01,  5.15628010e-02,\n        -1.09966584e-01, -2.23269954e-01,  5.43062799e-02,\n        -1.33496851e-01, -1.19007424e-01,  2.22256035e-01,\n        -2.77910858e-01,  2.71745831e-01, -4.16789412e-01,\n        -2.12772295e-01,  7.03845620e-01, -6.93172514e-02,\n         1.72736168e-01, -2.60139287e-01,  3.03129200e-03,\n         2.11665437e-01, -1.58136543e-02, -2.38662288e-02,\n        -5.83377741e-02,  5.22087336e-01,  2.94472545e-01,\n         1.68193743e-01, -7.54145905e-02,  1.43897519e-01,\n         1.56238422e-01, -3.39259744e-01,  2.46101081e-01,\n         3.11530419e-02, -5.94105422e-01, -2.72643536e-01,\n         1.21330276e-01,  3.12743425e-01, -1.66200623e-01,\n        -6.53145928e-03, -2.84941733e-01,  5.59734181e-05,\n        -3.21606755e-01, -1.73298046e-01, -1.07766673e-01,\n         9.54522491e-02,  2.46445552e-01, -2.62605727e-01,\n         1.81617990e-01,  6.52089193e-02, -1.01563215e-01,\n         3.59104156e-01, -5.22237360e-01,  1.64726060e-02,\n        -3.69388551e-01, -4.39793877e-02, -1.99547961e-01,\n        -3.79198231e-03,  3.00050706e-01, -1.49292305e-01,\n        -1.96511611e-01, -4.50382173e-01,  4.40837264e-01,\n        -2.56556179e-02, -1.36169955e-01, -3.62343282e-01,\n         1.56754032e-02,  7.93581456e-02,  1.90513626e-01,\n        -3.41799140e-01, -5.37522621e-02,  2.99514532e-01,\n        -1.21103093e-01,  4.06056821e-01, -2.30544969e-01,\n        -1.08799607e-01, -1.23380020e-01, -1.04779311e-01,\n        -3.59124064e-01,  6.79017082e-02,  3.27649474e-01,\n        -1.09562665e-01,  1.78656310e-01, -2.54520983e-01,\n        -2.15707019e-01,  2.97523111e-01, -3.25083762e-01,\n         1.63179748e-02,  3.89623255e-01, -7.29642585e-02,\n         5.47815263e-01, -9.16893259e-02, -4.76058573e-01,\n        -1.75076187e-01, -1.13026705e-04,  2.48254672e-01,\n         3.72678041e-01, -4.53566402e-01,  6.30904138e-02,\n         5.19643247e-01, -1.70341924e-01, -5.24724603e-01,\n        -9.19980407e-02,  5.36089689e-02,  9.92866978e-02,\n         2.93649197e-01,  1.39556557e-01,  4.84964341e-01,\n         3.11437190e-01,  3.61269027e-01, -3.84658389e-02,\n        -1.26146287e-01, -2.82240808e-01, -6.71329573e-02,\n        -5.25959721e-03,  2.54376501e-01, -2.77014941e-01,\n        -4.57646847e-02, -1.97771654e-01, -2.70207506e-02,\n         9.59944800e-02, -1.04830116e-02]], dtype=float32)']}
+```
+
+
+## 3. 更多教程
+
+本文档为功能测试用，更详细的Serving预测使用教程请参考：[ArcFace 服务化部署](../../deploy/pdserving/README_CN.md)  
diff --git a/recognition/arcface_paddle/test_tipc/docs/test_train_inference_python.md b/recognition/arcface_paddle/test_tipc/docs/test_train_inference_python.md
new file mode 100644
index 000000000..9bfc4338f
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/docs/test_train_inference_python.md
@@ -0,0 +1,78 @@
+# Linux端基础训练预测功能测试
+
+Linux端基础训练预测功能测试的主程序为`test_train_inference_python.sh`，可以测试基于Python的模型训练、评估、推理等基本功能。
+
+## 1. 测试结论汇总
+
+- 训练相关：
+
+| 算法名称 | 模型名称 | 单机单卡 | 单机多卡 | 多机多卡 | 模型压缩（单机多卡） |
+|  :----:  |   :----:  |    :----:  |  :----:   |  :----:   |  :----:   |
+|  ArcFace  | mobileface| 正常训练| 正常训练 | 正常训练 | - |
+
+
+- 预测相关：预测功能汇总如下，
+
+| 模型类型 |device | batchsize | tensorrt | mkldnn | cpu多线程 |
+|  :----:   |  :----: |   :----:   |  :----:  |   :----:   |  :----:  |
+| 正常模型 | GPU | 1 | fp32 | - | - |
+| 正常模型 | CPU | 1 | - | fp32 | 支持 |
+
+
+## 2. 测试流程
+
+运行环境配置请参考[文档](./install.md)的内容配置TIPC的运行环境。
+
+### 2.1 安装依赖
+- 安装PaddlePaddle >= 2.2
+- 安装依赖
+    ```
+    pip3 install  -r requirements.txt
+    ```
+- 安装autolog（规范化日志输出工具）
+    ```
+    pip3 install git+https://github.com/LDOUBLEV/AutoLog --force-reinstall
+    ```
+
+### 2.2 功能测试
+先运行`prepare.sh`准备数据和模型，然后运行`test_train_inference_python.sh`进行测试，最终在```test_tipc/output```目录下生成`python_infer_*.log`格式的日志文件。
+
+
+`test_train_inference_python.sh`包含5种运行模式，每种模式的运行数据不同，分别用于测试速度和精度，本文档只测试lite_train_lite_infer一种模式：
+
+- 模式1：lite_train_lite_infer，使用少量数据训练，用于快速验证训练到预测的走通流程，不验证精度和速度；
+```shell
+bash test_tipc/prepare.sh ./test_tipc/configs/ms1mv2_mobileface/train_infer_python.txt 'lite_train_lite_infer'
+bash test_tipc/test_train_inference_python.sh ./test_tipc/ms1mv2_mobileface/train_infer_python.txt  'lite_train_lite_infer'
+```  
+
+运行相应指令后，在`test_tipc/output`文件夹下自动会保存运行日志。如'lite_train_lite_infer'模式下，会运行训练+inference的链条，因此，在`test_tipc/output`文件夹有以下文件：
+```
+test_tipc/output/
+|- results_python.log    # 运行指令状态的日志
+|- norm_train_gpus_0_autocast_null_fp16_False/  # GPU 0号卡上正常训练的训练日志和模型保存文件夹
+|- norm_train_gpus_0_autocast_null_fp16_Trule/  # GPU 0号卡上fp16训练的训练日志和模型保存文件夹
+......
+|- python_infer_cpu_usemkldnn_True_threads_1_precision_fp32_batchsize_1.log  # CPU上开启Mkldnn线程数设置为1，测试batch_size=1条件下的预测运行日志
+|- python_infer_gpu_usetrt_True_precision_fp32_batchsize_1.log # GPU上开启TensorRT，测试batch_size=1的预测日志
+......
+```
+
+其中`results_python.log`中包含了每条指令的运行状态，如果运行成功会输出：
+```
+Run successfully with command - python3.7 tools/train.py --config_file=configs/ms1mv2_mobileface.py --is_static=False --embedding_size=128 --fp16=False --dataset=MS1M_v2 --data_dir=MS1M_v2/ --label_file=MS1M_v2/label.txt --num_classes=85742 --log_interval_step=1    --output=./test_tipc/output/norm_train_gpus_0_autocast_null_fp16_Trule --train_num=1       --fp16=Trule!
+Run successfully with command - python3.7 tools/validation.py --is_static=False --backbone=MobileFaceNet_128 --embedding_size=128 --data_dir=MS1M_v2 --val_targets=lfw --batch_size=128 --checkpoint_dir=./test_tipc/output/norm_train_gpus_0_autocast_null_fp16_Trule/MobileFaceNet_128/0    !
+......
+```
+如果运行失败，会输出：
+```
+Run failed with command - python3.7 tools/train.py --config_file=configs/ms1mv2_mobileface.py --is_static=False --embedding_size=128 --fp16=False --dataset=MS1M_v2 --data_dir=MS1M_v2/ --label_file=MS1M_v2/label.txt --num_classes=85742 --log_interval_step=1    --output=./test_tipc/output/norm_train_gpus_0_autocast_null_fp16_Trule --train_num=1       --fp16=Trule!
+Run failed with command - python3.7 tools/validation.py --is_static=False --backbone=MobileFaceNet_128 --embedding_size=128 --data_dir=MS1M_v2 --val_targets=lfw --batch_size=128 --checkpoint_dir=./test_tipc/output/norm_train_gpus_0_autocast_null_fp16_Trule/MobileFaceNet_128/0    !
+......
+```
+可以很方便的根据`results_python.log`中的内容判定哪一个指令运行错误。
+
+
+## 3. 更多教程
+本文档为功能测试用，更丰富的训练预测使用教程请参考：  
+[模型训练与预测](../../README_cn.md)  
\ No newline at end of file
diff --git a/recognition/arcface_paddle/test_tipc/prepare.sh b/recognition/arcface_paddle/test_tipc/prepare.sh
new file mode 100644
index 000000000..034f8a7a8
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/prepare.sh
@@ -0,0 +1,42 @@
+#!/bin/bash
+source test_tipc/common_func.sh
+FILENAME=$1
+
+# MODE be one of ['lite_train_infer' 'serving_infer']
+
+MODE=$2
+
+dataline=$(cat ${FILENAME})
+lines=(${dataline})
+
+IFS=$'\n'
+# The training params
+model_name=$(func_parser_value "${lines[1]}")
+trainer_list=$(func_parser_value "${lines[14]}")
+
+MODE=$2
+
+if [ ${MODE} = "lite_train_lite_infer" ];then
+    rm -rf MS1M_v2; mkdir MS1M_v2
+    # pretrain lite train data
+    tar xf test_tipc/data/small_dataset.tar --strip-components 1 -C MS1M_v2 
+    
+    # wget -nc -P ./MS1M_v2/ https://paddle-model-ecology.bj.bcebos.com/whole_chain/insight-face/lfw.bin
+    cp test_tipc/data/small_lfw.bin MS1M_v2/lfw.bin
+
+elif [ ${MODE} = "serving_infer" ];then
+     # prepare serving env
+    python_name=$(func_parser_value "${lines[2]}")
+    rm paddle_serving_server_gpu-0.0.0.post101-py3-none-any.whl
+    wget https://paddle-serving.bj.bcebos.com/chain/paddle_serving_server_gpu-0.0.0.post101-py3-none-any.whl
+    ${python_name} -m pip install install paddle_serving_server_gpu-0.0.0.post101-py3-none-any.whl
+    ${python_name} -m pip install paddle_serving_client==0.6.3
+    ${python_name} -m pip install paddle-serving-app==0.6.3
+    ${python_name} -m pip install werkzeug==2.0.2
+
+    rm -rf ./inference
+
+    wget -nc -P ./inference https://paddle-model-ecology.bj.bcebos.com/model/insight-face/mobileface_v1.0_infer.tar
+    tar xf inference/mobileface_v1.0_infer.tar --strip-components 1 -C inference 
+fi
+
diff --git a/recognition/arcface_paddle/test_tipc/readme.md b/recognition/arcface_paddle/test_tipc/readme.md
new file mode 100644
index 000000000..fba2fb87b
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/readme.md
@@ -0,0 +1,68 @@
+# 飞桨训推一体认证
+
+## 1. 简介
+
+飞桨除了基本的模型训练和预测，还提供了支持多端多平台的高性能推理部署工具。本文档提供了 ArcFace 中所有 PaddlePaddle 模型的飞桨训推一体认证 (Training and Inference Pipeline Certification(TIPC)) 信息和测试工具，方便用户查阅每种模型的训练推理部署打通情况，并可以进行一键测试。
+
+<div align="center">
+    <img src="docs/guide.png" width="1000">
+</div>
+
+## 2. 汇总信息
+
+打通情况汇总如下，已填写的部分表示可以使用本工具进行一键测试，未填写的表示正在支持中。
+
+**字段说明：**
+- 基础训练预测：包括模型训练、Paddle Inference Python预测。
+- 更多训练方式：包括多机多卡、混合精度。
+- 模型压缩：包括裁剪、离线/在线量化、蒸馏。
+- 其他预测部署：包括Paddle Inference C++预测、Paddle Serving部署、Paddle-Lite部署等。
+
+更详细的mkldnn、Tensorrt等预测加速相关功能的支持情况可以查看各测试工具的[更多教程](#more)。
+
+| 算法论文 | 模型名称 | 模型类型 | 基础<br>训练预测 | 更多<br>训练方式 | 模型压缩 |  其他预测部署  |
+| :---:| :---: |  :----:  | :--------: |  :----:  |   :----:  |   :----:  |
+| ArcFace     | ms1mv2_mobileface | 识别  | 支持 | 多机多卡 | - | Paddle Serving: Python |
+
+
+
+## 3. 一键测试工具使用
+### 目录介绍
+
+```shell
+test_tipc/
+├── configs/  # 配置文件目录
+	├── ms1mv2_mobileface  # ms1mv2_mobileface 模型的测试配置文件目录
+		├── model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt # 测试Linux上python serving预测的配置文件
+		└── train_infer_python.txt # 测试Linux上python训练预测（基础训练预测）的配置文件
+	├── ...  
+├── data/ # 存放 TIPC 测试数据的目录
+	├── small_dataset.tar # 用于训练的小数据集 (10张图片)
+	├── small_lfw.bin # 用于评估的小数据集 (20张图片)
+├── docs/ # 存放 TIPC 测试数据的目录
+	├── install.md # 安装 TIPC 所需环境的文档
+	├── test_train_inference_python.md # 测试Linux上python训练预测的文档
+	├── test_serving.md # 测试Linux上python serving预测的文档
+├── prepare.sh                        # 完成test_*.sh运行所需要的数据和模型下载
+├── test_serving.sh    # 测试python训练预测的主程序
+├── test_train_inference_python.sh    # 测试python训练预测的主程序
+├── common_func.sh                    # 通用shell脚本函数
+└── readme.md                         # TIPC使用文档
+```
+
+### 测试流程
+使用本工具，可以测试不同功能的支持情况，以及预测结果是否对齐，测试流程如下：
+<div align="center">
+    <img src="docs/test.png" width="800">
+</div>
+
+1. 运行prepare.sh准备测试所需数据和模型；
+2. 运行要测试的功能对应的测试脚本`test_*.sh`，产出log，由log可以看到不同配置是否运行成功；
+
+其中，有1个测试主程序，功能如下：
+- `test_train_inference_python.sh`：测试基于Python的模型训练、评估、推理等基本功能，包括裁剪、量化、蒸馏。
+
+<a name="more"></a>
+#### 更多教程
+各功能测试中涉及混合精度、裁剪、量化等训练相关，及mkldnn、Tensorrt等多种预测相关参数配置，请点击下方相应链接了解更多细节和使用教程：  
+[test_train_inference_python 使用](docs/test_train_inference_python.md)  
\ No newline at end of file
diff --git a/recognition/arcface_paddle/test_tipc/test_serving.sh b/recognition/arcface_paddle/test_tipc/test_serving.sh
new file mode 100644
index 000000000..e86d3174b
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/test_serving.sh
@@ -0,0 +1,168 @@
+#!/bin/bash
+source test_tipc/common_func.sh
+
+FILENAME=$1
+dataline=$(awk 'NR==1, NR==18{print}'  $FILENAME)
+
+# parser params
+IFS=$'\n'
+lines=(${dataline})
+
+# parser serving
+model_name=$(func_parser_value "${lines[1]}")
+python=$(func_parser_value "${lines[2]}")
+trans_model_py=$(func_parser_value "${lines[3]}")
+infer_model_dir_key=$(func_parser_key "${lines[4]}")
+infer_model_dir_value=$(func_parser_value "${lines[4]}")
+model_filename_key=$(func_parser_key "${lines[5]}")
+model_filename_value=$(func_parser_value "${lines[5]}")
+params_filename_key=$(func_parser_key "${lines[6]}")
+params_filename_value=$(func_parser_value "${lines[6]}")
+serving_server_key=$(func_parser_key "${lines[7]}")
+serving_server_value=$(func_parser_value "${lines[7]}")
+serving_client_key=$(func_parser_key "${lines[8]}")
+serving_client_value=$(func_parser_value "${lines[8]}")
+serving_dir_value=$(func_parser_value "${lines[9]}")
+web_service_py=$(func_parser_value "${lines[10]}")
+web_use_gpu_key=$(func_parser_key "${lines[11]}")
+web_use_gpu_list=$(func_parser_value "${lines[11]}")
+web_use_mkldnn_key=$(func_parser_key "${lines[12]}")
+web_use_mkldnn_list=$(func_parser_value "${lines[12]}")
+web_cpu_threads_key=$(func_parser_key "${lines[13]}")
+web_cpu_threads_list=$(func_parser_value "${lines[13]}")
+web_use_trt_key=$(func_parser_key "${lines[14]}")
+web_use_trt_list=$(func_parser_value "${lines[14]}")
+web_precision_key=$(func_parser_key "${lines[15]}")
+web_precision_list=$(func_parser_value "${lines[15]}")
+pipeline_py=$(func_parser_value "${lines[16]}")
+image_dir_key=$(func_parser_key "${lines[17]}")
+image_dir_value=$(func_parser_value "${lines[17]}")
+
+LOG_PATH="../../test_tipc/output"
+mkdir -p ./test_tipc/output
+status_log="${LOG_PATH}/results_serving.log"
+
+function func_serving(){
+    IFS='|'
+    _python=$1
+    _script=$2
+    _model_dir=$3
+    # pdserving
+    set_dirname=$(func_set_params "${infer_model_dir_key}" "${infer_model_dir_value}")
+    set_model_filename=$(func_set_params "${model_filename_key}" "${model_filename_value}")
+    set_params_filename=$(func_set_params "${params_filename_key}" "${params_filename_value}")
+    set_serving_server=$(func_set_params "${serving_server_key}" "${serving_server_value}")
+    set_serving_client=$(func_set_params "${serving_client_key}" "${serving_client_value}")
+    set_image_dir=$(func_set_params "${image_dir_key}" "${image_dir_value}")
+    trans_model_cmd="${python} ${trans_model_py} ${set_dirname} ${set_model_filename} ${set_params_filename} ${set_serving_server} ${set_serving_client}"
+    eval $trans_model_cmd
+    cd ${serving_dir_value}
+    echo $PWD
+    unset https_proxy
+    unset http_proxy
+    for python in ${python[*]}; do
+        if [ ${python} = "cpp"]; then
+            for use_gpu in ${web_use_gpu_list[*]}; do
+                if [ ${use_gpu} = "null" ]; then
+                    web_service_cpp_cmd="${python} -m paddle_serving_server.serve --model ppocr_det_mobile_2.0_serving/ ppocr_rec_mobile_2.0_serving/ --port 9293"
+                    eval $web_service_cmd
+                    sleep 2s
+                    _save_log_path="${LOG_PATH}/server_infer_cpp_cpu_pipeline_usemkldnn_False_threads_4_batchsize_1.log"
+                    pipeline_cmd="${python} ocr_cpp_client.py ppocr_det_mobile_2.0_client/ ppocr_rec_mobile_2.0_client/"
+                    eval $pipeline_cmd
+                    status_check $last_status "${pipeline_cmd}" "${status_log}"
+                    sleep 2s
+                    ps ux | grep -E 'web_service|pipeline' | awk '{print $2}' | xargs kill -s 9
+                else
+                    web_service_cpp_cmd="${python} -m paddle_serving_server.serve --model ppocr_det_mobile_2.0_serving/ ppocr_rec_mobile_2.0_serving/ --port 9293 --gpu_id=0"
+                    eval $web_service_cmd
+                    sleep 2s
+                    _save_log_path="${LOG_PATH}/server_infer_cpp_cpu_pipeline_usemkldnn_False_threads_4_batchsize_1.log"
+                    pipeline_cmd="${python} ocr_cpp_client.py ppocr_det_mobile_2.0_client/ ppocr_rec_mobile_2.0_client/"
+                    eval $pipeline_cmd
+                    status_check $last_status "${pipeline_cmd}" "${status_log}"
+                    sleep 2s
+                    ps ux | grep -E 'web_service|pipeline' | awk '{print $2}' | xargs kill -s 9                
+                fi
+            done
+        else
+            # python serving
+            for use_gpu in ${web_use_gpu_list[*]}; do
+                echo ${ues_gpu}
+                if [ ${use_gpu} = "null" ]; then
+                    for use_mkldnn in ${web_use_mkldnn_list[*]}; do
+                        if [ ${use_mkldnn} = "False" ]; then
+                            continue
+                        fi
+                        for threads in ${web_cpu_threads_list[*]}; do
+                            set_cpu_threads=$(func_set_params "${web_cpu_threads_key}" "${threads}")
+                            web_service_cmd="${python} ${web_service_py} ${web_use_gpu_key}=${use_gpu} ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} &"
+                            echo $web_service_cmd
+                            sleep 2s
+                            for pipeline in ${pipeline_py[*]}; do
+                                _save_log_path="${LOG_PATH}/server_infer_cpu_${pipeline%_client*}_usemkldnn_${use_mkldnn}_threads_${threads}_batchsize_1.log"
+                                pipeline_cmd="${python} ${pipeline} ${set_image_dir} > ${_save_log_path} 2>&1 "
+                                echo $pipeline_cmd
+                                last_status=${PIPESTATUS[0]}
+                                echo "cat ${_save_log_path}"
+                                status_check $last_status "${pipeline_cmd}" "${status_log}"
+                                sleep 2s
+                            done
+                            ps ux | grep -E 'web_service|pipeline' | awk '{print $2}' | xargs kill -s 9
+                        done
+                    done
+                elif [ ${use_gpu} = "0" ]; then
+                    for use_trt in ${web_use_trt_list[*]}; do
+                        for precision in ${web_precision_list[*]}; do
+                            if [[ ${_flag_quant} = "False" ]] && [[ ${precision} =~ "int8" ]]; then
+                                continue
+                            fi
+                            if [[ ${precision} =~ "fp16" || ${precision} =~ "int8" ]] && [ ${use_trt} = "False" ]; then
+                                continue
+                            fi
+                            if [[ ${use_trt} = "False" || ${precision} =~ "int8" ]] && [[ ${_flag_quant} = "True" ]]; then
+                                continue
+                            fi
+                            set_tensorrt=$(func_set_params "${web_use_trt_key}" "${use_trt}")
+                            set_precision=$(func_set_params "${web_precision_key}" "${precision}")
+                            web_service_cmd="${python} ${web_service_py} ${web_use_gpu_key}=${use_gpu} ${set_tensorrt} ${set_precision} & "
+                            echo $web_service_cmd
+                        
+                            sleep 2s
+                            for pipeline in ${pipeline_py[*]}; do
+                                _save_log_path="${LOG_PATH}/server_infer_gpu_${pipeline%_client*}_usetrt_${use_trt}_precision_${precision}_batchsize_1.log"
+                                pipeline_cmd="${python} ${pipeline} ${set_image_dir}> ${_save_log_path} 2>&1"
+                                echo $pipeline_cmd
+                                last_status=${PIPESTATUS[0]}
+                                echo "cat ${_save_log_path}"
+                                status_check $last_status "${pipeline_cmd}" "${status_log}"
+                                sleep 2s
+                            done
+                            ps ux | grep -E 'web_service|pipeline' | awk '{print $2}' | xargs kill -s 9
+                        done
+                    done
+                else
+                    echo "Does not support hardware other than CPU and GPU Currently!"
+                fi
+            done
+        fi
+    done
+}
+
+
+# set cuda device
+GPUID=$2
+if [ ${#GPUID} -le 0 ];then
+    env=" "
+else
+    env="export CUDA_VISIBLE_DEVICES=${GPUID}"
+fi
+set CUDA_VISIBLE_DEVICES
+eval $env
+
+
+echo "################### run test ###################"
+
+export Count=0
+IFS="|"
+func_serving "${web_service_cmd}"
\ No newline at end of file
diff --git a/recognition/arcface_paddle/test_tipc/test_train_inference_python.sh b/recognition/arcface_paddle/test_tipc/test_train_inference_python.sh
new file mode 100644
index 000000000..b60ad31b9
--- /dev/null
+++ b/recognition/arcface_paddle/test_tipc/test_train_inference_python.sh
@@ -0,0 +1,372 @@
+#!/bin/bash
+source test_tipc/common_func.sh
+
+FILENAME=$1
+# MODE be one of ['lite_train_lite_infer' 'lite_train_whole_infer' 'whole_train_whole_infer', 'whole_infer', 'klquant_whole_infer']
+MODE=$2
+
+dataline=$(awk 'NR==1, NR==51{print}'  $FILENAME)
+
+# parser params
+IFS=$'\n'
+lines=(${dataline})
+
+# The training params
+model_name=$(func_parser_value "${lines[1]}")
+python=$(func_parser_value "${lines[2]}")
+gpu_list=$(func_parser_value "${lines[3]}")
+train_use_gpu_key=$(func_parser_key "${lines[4]}")
+train_use_gpu_value=$(func_parser_value "${lines[4]}")
+autocast_list=$(func_parser_value "${lines[5]}")
+autocast_key=$(func_parser_key "${lines[5]}")
+epoch_key=$(func_parser_key "${lines[6]}")
+epoch_num=$(func_parser_params "${lines[6]}")
+save_model_key=$(func_parser_key "${lines[7]}")
+train_batch_key=$(func_parser_key "${lines[8]}")
+train_batch_value=$(func_parser_params "${lines[8]}")
+pretrain_model_key=$(func_parser_key "${lines[9]}")
+pretrain_model_value=$(func_parser_value "${lines[9]}")
+train_model_name=$(func_parser_value "${lines[10]}")
+train_infer_img_dir=$(func_parser_value "${lines[11]}")
+train_param_key1=$(func_parser_key "${lines[12]}")
+train_param_value1=$(func_parser_value "${lines[12]}")
+
+trainer_list=$(func_parser_value "${lines[14]}")
+trainer_norm=$(func_parser_key "${lines[15]}")
+norm_trainer=$(func_parser_value "${lines[15]}")
+pact_key=$(func_parser_key "${lines[16]}")
+pact_trainer=$(func_parser_value "${lines[16]}")
+fpgm_key=$(func_parser_key "${lines[17]}")
+fpgm_trainer=$(func_parser_value "${lines[17]}")
+distill_key=$(func_parser_key "${lines[18]}")
+distill_trainer=$(func_parser_value "${lines[18]}")
+trainer_key1=$(func_parser_key "${lines[19]}")
+trainer_value1=$(func_parser_value "${lines[19]}")
+trainer_key2=$(func_parser_key "${lines[20]}")
+trainer_value2=$(func_parser_value "${lines[20]}")
+
+eval_py=$(func_parser_value "${lines[23]}")
+eval_key1=$(func_parser_key "${lines[24]}")
+eval_value1=$(func_parser_value "${lines[24]}")
+
+save_infer_key=$(func_parser_key "${lines[27]}")
+export_weight=$(func_parser_key "${lines[28]}")
+norm_export=$(func_parser_value "${lines[29]}")
+pact_export=$(func_parser_value "${lines[30]}")
+fpgm_export=$(func_parser_value "${lines[31]}")
+distill_export=$(func_parser_value "${lines[32]}")
+export_key1=$(func_parser_key "${lines[33]}")
+export_value1=$(func_parser_value "${lines[33]}")
+export_key2=$(func_parser_key "${lines[34]}")
+export_value2=$(func_parser_value "${lines[34]}")
+inference_dir=$(func_parser_value "${lines[35]}")
+
+# parser inference model 
+infer_model_dir_list=$(func_parser_value "${lines[36]}")
+infer_export_list=$(func_parser_value "${lines[37]}")
+infer_is_quant=$(func_parser_value "${lines[38]}")
+# parser inference 
+inference_py=$(func_parser_value "${lines[39]}")
+use_gpu_key=$(func_parser_key "${lines[40]}")
+use_gpu_list=$(func_parser_value "${lines[40]}")
+use_mkldnn_key=$(func_parser_key "${lines[41]}")
+use_mkldnn_list=$(func_parser_value "${lines[41]}")
+cpu_threads_key=$(func_parser_key "${lines[42]}")
+cpu_threads_list=$(func_parser_value "${lines[42]}")
+batch_size_key=$(func_parser_key "${lines[43]}")
+batch_size_list=$(func_parser_value "${lines[43]}")
+use_trt_key=$(func_parser_key "${lines[44]}")
+use_trt_list=$(func_parser_value "${lines[44]}")
+precision_key=$(func_parser_key "${lines[45]}")
+precision_list=$(func_parser_value "${lines[45]}")
+infer_model_key=$(func_parser_key "${lines[46]}")
+image_dir_key=$(func_parser_key "${lines[47]}")
+infer_img_dir=$(func_parser_value "${lines[47]}")
+save_log_key=$(func_parser_key "${lines[48]}")
+benchmark_key=$(func_parser_key "${lines[49]}")
+benchmark_value=$(func_parser_value "${lines[49]}")
+infer_key1=$(func_parser_key "${lines[50]}")
+infer_value1=$(func_parser_value "${lines[50]}")
+
+# parser klquant_infer
+if [ ${MODE} = "klquant_whole_infer" ]; then
+    dataline=$(awk 'NR==85 NR==101{print}'  $FILENAME)
+    lines=(${dataline})
+    # parser inference model 
+    infer_model_dir_list=$(func_parser_value "${lines[1]}")
+    infer_export_list=$(func_parser_value "${lines[2]}")
+    infer_is_quant=$(func_parser_value "${lines[3]}")
+    # parser inference 
+    inference_py=$(func_parser_value "${lines[4]}")
+    use_gpu_key=$(func_parser_key "${lines[5]}")
+    use_gpu_list=$(func_parser_value "${lines[5]}")
+    use_mkldnn_key=$(func_parser_key "${lines[6]}")
+    use_mkldnn_list=$(func_parser_value "${lines[6]}")
+    cpu_threads_key=$(func_parser_key "${lines[7]}")
+    cpu_threads_list=$(func_parser_value "${lines[7]}")
+    batch_size_key=$(func_parser_key "${lines[8]}")
+    batch_size_list=$(func_parser_value "${lines[8]}")
+    use_trt_key=$(func_parser_key "${lines[9]}")
+    use_trt_list=$(func_parser_value "${lines[9]}")
+    precision_key=$(func_parser_key "${lines[10]}")
+    precision_list=$(func_parser_value "${lines[10]}")
+    infer_model_key=$(func_parser_key "${lines[11]}")
+    image_dir_key=$(func_parser_key "${lines[12]}")
+    infer_img_dir=$(func_parser_value "${lines[12]}")
+    save_log_key=$(func_parser_key "${lines[13]}")
+    benchmark_key=$(func_parser_key "${lines[14]}")
+    benchmark_value=$(func_parser_value "${lines[14]}")
+    infer_key1=$(func_parser_key "${lines[15]}")
+    infer_value1=$(func_parser_value "${lines[15]}")
+fi
+
+LOG_PATH="./test_tipc/output"
+mkdir -p ${LOG_PATH}
+status_log="${LOG_PATH}/results_python.log"
+
+
+function func_inference(){
+    IFS='|'
+    _python=$1
+    _script=$2
+    _model_dir=$3
+    _log_path=$4
+    _img_dir=$5
+    _flag_quant=$6
+    # inference 
+    for use_gpu in ${use_gpu_list[*]}; do
+        if [ ${use_gpu} = "False" ] || [ ${use_gpu} = "cpu" ]; then
+            for use_mkldnn in ${use_mkldnn_list[*]}; do
+                if [ ${use_mkldnn} = "False" ] && [ ${_flag_quant} = "True" ]; then
+                    continue
+                fi
+                for threads in ${cpu_threads_list[*]}; do
+                    for batch_size in ${batch_size_list[*]}; do
+                        for precision in ${precision_list[*]}; do
+                            if [ ${use_mkldnn} = "False" ] && [ ${precision} = "fp16" ]; then
+                                continue
+                            fi # skip when enable fp16 but disable mkldnn
+                            if [ ${_flag_quant} = "True" ] && [ ${precision} != "int8" ]; then
+                                continue
+                            fi # skip when quant model inference but precision is not int8
+                            set_precision=$(func_set_params "${precision_key}" "${precision}")
+                            
+                            _save_log_path="${_log_path}/python_infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_precision_${precision}_batchsize_${batch_size}.log"
+                            set_infer_data=$(func_set_params "${image_dir_key}" "${_img_dir}")
+                            set_benchmark=$(func_set_params "${benchmark_key}" "${benchmark_value}")
+                            set_batchsize=$(func_set_params "${batch_size_key}" "${batch_size}")
+                            set_cpu_threads=$(func_set_params "${cpu_threads_key}" "${threads}")
+                            set_model_dir=$(func_set_params "${infer_model_key}" "${_model_dir}")
+                            set_infer_params1=$(func_set_params "${infer_key1}" "${infer_value1}")
+                            command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_batchsize} ${set_infer_data} ${set_benchmark} ${set_precision} ${set_infer_params1} --model_file=${_model_dir}/${model_name}.pdmodel --params_file=${_model_dir}/${model_name}.pdiparams > ${_save_log_path} 2>&1 "
+                            eval $command
+                            last_status=${PIPESTATUS[0]}
+                            eval "cat ${_save_log_path}"
+                            status_check $last_status "${command}" "${status_log}"
+                        done
+                    done
+                done
+            done
+        elif [ ${use_gpu} = "True" ] || [ ${use_gpu} = "gpu" ]; then
+            for use_trt in ${use_trt_list[*]}; do
+                for precision in ${precision_list[*]}; do
+                    if [[ ${_flag_quant} = "False" ]] && [[ ${precision} =~ "int8" ]]; then
+                        continue
+                    fi 
+                    if [[ ${precision} =~ "fp16" || ${precision} =~ "int8" ]] && [ ${use_trt} = "False" ]; then
+                        continue
+                    fi
+                    if [[ ${use_trt} = "False" || ${precision} =~ "int8" ]] && [ ${_flag_quant} = "True" ]; then
+                        continue
+                    fi
+                    for batch_size in ${batch_size_list[*]}; do
+                        _save_log_path="${_log_path}/python_infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
+                        set_infer_data=$(func_set_params "${image_dir_key}" "${_img_dir}")
+                        set_benchmark=$(func_set_params "${benchmark_key}" "${benchmark_value}")
+                        set_batchsize=$(func_set_params "${batch_size_key}" "${batch_size}")
+                        set_tensorrt=$(func_set_params "${use_trt_key}" "${use_trt}")
+                        set_precision=$(func_set_params "${precision_key}" "${precision}")
+                        set_model_dir=$(func_set_params "${infer_model_key}" "${_model_dir}")
+                        set_infer_params1=$(func_set_params "${infer_key1}" "${infer_value1}")
+                        command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${set_tensorrt} ${set_precision} ${set_batchsize} ${set_infer_data} ${set_benchmark} ${set_infer_params1} --model_file=${_model_dir}/${model_name}.pdmodel --params_file=${_model_dir}/${model_name}.pdiparams > ${_save_log_path} 2>&1 "
+                        eval $command
+                        last_status=${PIPESTATUS[0]}
+                        eval "cat ${_save_log_path}"
+                        status_check $last_status "${command}" "${status_log}"
+                        
+                    done
+                done
+            done
+        else
+            echo "Does not support hardware other than CPU and GPU Currently!"
+        fi
+    done
+}
+
+if [ ${MODE} = "whole_infer" ] || [ ${MODE} = "klquant_whole_infer" ]; then
+    GPUID=$3
+    if [ ${#GPUID} -le 0 ];then
+        env=" "
+    else
+        env="export CUDA_VISIBLE_DEVICES=${GPUID}"
+    fi
+    # set CUDA_VISIBLE_DEVICES
+    eval $env
+    export Count=0
+    IFS="|"
+    infer_run_exports=(${infer_export_list})
+    infer_quant_flag=(${infer_is_quant})
+    for infer_model in ${infer_model_dir_list[*]}; do
+        # run export
+        if [ ${infer_run_exports[Count]} != "null" ];then
+            save_infer_dir=$(dirname $infer_model)
+            set_export_weight=$(func_set_params "${export_weight}" "${infer_model}")
+            set_save_infer_key=$(func_set_params "${save_infer_key}" "${save_infer_dir}")
+            export_cmd="${python} ${infer_run_exports[Count]} ${set_export_weight} ${set_save_infer_key}"
+            echo ${infer_run_exports[Count]} 
+            echo  $export_cmd
+            eval $export_cmd
+            status_export=$?
+            status_check $status_export "${export_cmd}" "${status_log}"
+        else
+            save_infer_dir=${infer_model}
+        fi
+        #run inference
+        is_quant=${infer_quant_flag[Count]}
+        if [ ${MODE} = "klquant_infer" ]; then
+            is_quant="True"
+        fi
+        func_inference "${python}" "${inference_py}" "${save_infer_dir}" "${LOG_PATH}" "${infer_img_dir}" ${is_quant}
+        Count=$(($Count + 1))
+    done
+else
+    IFS="|"
+    export Count=0
+    USE_GPU_KEY=(${train_use_gpu_value})
+    for gpu in ${gpu_list[*]}; do
+        train_use_gpu=${USE_GPU_KEY[Count]}
+        Count=$(($Count + 1))
+        ips=""
+        if [ ${gpu} = "-1" ];then
+            env=""
+        elif [ ${#gpu} -le 1 ];then
+            env="export CUDA_VISIBLE_DEVICES=${gpu}"
+            eval ${env}
+        elif [ ${#gpu} -le 15 ];then
+            IFS=","
+            array=(${gpu})
+            env="export CUDA_VISIBLE_DEVICES=${array[0]}"
+            IFS="|"
+        else
+            IFS=";"
+            array=(${gpu})
+            ips=${array[0]}
+            gpu=${array[1]}
+            IFS="|"
+            env=" "
+        fi
+        for autocast in ${autocast_list[*]}; do 
+            if [ ${autocast} = "amp" ]; then
+                set_amp_config="Global.use_amp=True Global.scale_loss=1024.0 Global.use_dynamic_loss_scaling=True"
+            else
+                set_amp_config=" "
+            fi          
+            for trainer in ${trainer_list[*]}; do 
+                flag_quant=False
+                if [ ${trainer} = ${pact_key} ]; then
+                    run_train=${pact_trainer}
+                    run_export=${pact_export}
+                    flag_quant=True
+                elif [ ${trainer} = "${fpgm_key}" ]; then
+                    run_train=${fpgm_trainer}
+                    run_export=${fpgm_export}
+                elif [ ${trainer} = "${distill_key}" ]; then
+                    run_train=${distill_trainer}
+                    run_export=${distill_export}
+                elif [ ${trainer} = ${trainer_key1} ]; then
+                    run_train=${trainer_value1}
+                    run_export=${export_value1}
+                elif [[ ${trainer} = ${trainer_key2} ]]; then
+                    run_train=${trainer_value2}
+                    run_export=${export_value2}
+                else
+                    run_train=${norm_trainer}
+                    run_export=${norm_export}
+                fi
+
+                if [ ${run_train} = "null" ]; then
+                    continue
+                fi
+                set_autocast=$(func_set_params "${autocast_key}" "${autocast}")
+                set_epoch=$(func_set_params "${epoch_key}" "${epoch_num}")
+                set_pretrain=$(func_set_params "${pretrain_model_key}" "${pretrain_model_value}")
+                set_batchsize=$(func_set_params "${train_batch_key}" "${train_batch_value}")
+                set_train_params1=$(func_set_params "${train_param_key1}" "${train_param_value1}")
+                set_use_gpu=$(func_set_params "${train_use_gpu_key}" "${train_use_gpu}")
+                if [ ${#ips} -le 26 ];then
+                    save_log="${LOG_PATH}/${trainer}_gpus_${gpu}_autocast_${autocast}"
+                    nodes=1
+                else
+                    IFS=","
+                    ips_array=(${ips})
+                    IFS="|"
+                    nodes=${#ips_array[@]}
+                    save_log="${LOG_PATH}/${trainer}_gpus_${gpu}_autocast_${autocast}_nodes_${nodes}"
+                fi
+
+                # load pretrain from norm training if current trainer is pact or fpgm trainer
+                if ([ ${trainer} = ${pact_key} ] || [ ${trainer} = ${fpgm_key} ]) && [ ${nodes} -le 1 ]; then
+                    set_pretrain="${load_norm_train_model}"
+                fi
+
+                set_save_model=$(func_set_params "${save_model_key}" "${save_log}")
+                if [ ${#gpu} -le 2 ];then  # train with cpu or single gpu
+                    cmd="${python} ${run_train} ${set_use_gpu}  ${set_save_model} ${set_epoch} ${set_pretrain} ${set_autocast} ${set_batchsize} ${set_train_params1} ${set_amp_config} "
+                elif [ ${#ips} -le 26 ];then  # train with multi-gpu
+                    cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_autocast} ${set_batchsize} ${set_train_params1} ${set_amp_config}"
+                else     # train with multi-machine
+                    cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_autocast} ${set_batchsize} ${set_train_params1} ${set_amp_config}"
+                fi
+                # run train
+                eval "unset CUDA_VISIBLE_DEVICES"
+                eval $cmd
+                status_check $? "${cmd}" "${status_log}"
+
+                set_eval_pretrain=$(func_set_params "${pretrain_model_key}" "${save_log}/${train_model_name}")
+                # save norm trained models to set pretrain for pact training and fpgm training 
+                if [ ${trainer} = ${trainer_norm} ] && [ ${nodes} -le 1]; then
+                    load_norm_train_model=${set_eval_pretrain}
+                fi
+                # run eval 
+                if [ ${eval_py} != "null" ]; then
+                    set_eval_params1=$(func_set_params "${eval_key1}" "${eval_value1}")
+                    eval_cmd="${python} ${eval_py} ${set_eval_pretrain} ${set_use_gpu} ${set_eval_params1}" 
+                    eval $eval_cmd
+                    status_check $? "${eval_cmd}" "${status_log}"
+                fi
+                # run export model
+                if [ ${run_export} != "null" ]; then 
+                    # run export model
+                    save_infer_path="${save_log}"
+                    set_export_weight=$(func_set_params "${export_weight}" "${save_log}/${train_model_name}")
+                    set_save_infer_key=$(func_set_params "${save_infer_key}" "${save_infer_path}")
+                    export_cmd="${python} ${run_export} ${set_export_weight} ${set_save_infer_key}"
+                    eval $export_cmd
+                    status_check $? "${export_cmd}" "${status_log}"
+
+                    #run inference
+                    eval $env
+                    save_infer_path="${save_log}"
+                    if [ ${inference_dir} != "null" ] && [ ${inference_dir} != '##' ]; then
+                        infer_model_dir="${save_infer_path}/${inference_dir}"
+                    else
+                        infer_model_dir=${save_infer_path}
+                    fi
+                    func_inference "${python}" "${inference_py}" "${infer_model_dir}" "${LOG_PATH}" "${train_infer_img_dir}" "${flag_quant}"
+                    
+                    eval "unset CUDA_VISIBLE_DEVICES"
+                fi
+            done  # done with:    for trainer in ${trainer_list[*]}; do 
+        done      # done with:    for autocast in ${autocast_list[*]}; do 
+    done          # done with:    for gpu in ${gpu_list[*]}; do
+fi  # end if [ ${MODE} = "infer" ]; then
\ No newline at end of file
diff --git a/recognition/arcface_paddle/tools/__init__.py b/recognition/arcface_paddle/tools/__init__.py
new file mode 100644
index 000000000..185a92b8d
--- /dev/null
+++ b/recognition/arcface_paddle/tools/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/recognition/arcface_paddle/tools/benchmark_speed.py b/recognition/arcface_paddle/tools/benchmark_speed.py
new file mode 100644
index 000000000..6745f5fba
--- /dev/null
+++ b/recognition/arcface_paddle/tools/benchmark_speed.py
@@ -0,0 +1,115 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import numpy as np
+import time
+import argparse
+from paddle.inference import Config
+from paddle.inference import create_predictor
+
+
+def parse_args():
+    def str2bool(v):
+        return v.lower() in ("true", "t", "1")
+
+    # general params
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--use_gpu", type=str2bool, default=False)
+    parser.add_argument("--gpu_mem", type=int, default=1000)
+
+    # params for predict
+    parser.add_argument("--model_file", type=str)
+    parser.add_argument("--params_file", type=str)
+    parser.add_argument("-b", "--batch_size", type=int, default=1)
+    parser.add_argument("--ir_optim", type=str2bool, default=True)
+    parser.add_argument("--use_mkldnn", type=str2bool, default=True)
+    parser.add_argument("--cpu_num_threads", type=int, default=10)
+    parser.add_argument("--model", type=str)
+
+    return parser.parse_args()
+
+
+def create_paddle_predictor(args):
+    config = Config(args.model_file, args.params_file)
+
+    if args.use_gpu:
+        config.enable_use_gpu(args.gpu_mem, 0)
+    else:
+        config.disable_gpu()
+
+    if args.use_mkldnn:
+        config.enable_mkldnn()
+        config.set_cpu_math_library_num_threads(args.cpu_num_threads)
+        config.set_mkldnn_cache_capacity(100)
+
+    config.disable_glog_info()
+    config.switch_ir_optim(args.ir_optim)  # default true
+
+    config.enable_memory_optim()
+    # use zero copy
+    config.switch_use_feed_fetch_ops(False)
+    predictor = create_predictor(config)
+
+    return predictor
+
+
+class Predictor(object):
+    def __init__(self, args):
+
+        self.args = args
+
+        self.paddle_predictor = create_paddle_predictor(args)
+        input_names = self.paddle_predictor.get_input_names()
+        self.input_tensor = self.paddle_predictor.get_input_handle(input_names[
+            0])
+
+        output_names = self.paddle_predictor.get_output_names()
+        self.output_tensor = self.paddle_predictor.get_output_handle(
+            output_names[0])
+
+    def predict(self, batch_input):
+        self.input_tensor.copy_from_cpu(batch_input)
+        self.paddle_predictor.run()
+        batch_output = self.output_tensor.copy_to_cpu()
+        return batch_output
+
+    def benchmark_predict(self):
+        test_num = 500
+        test_time = 0.0
+        for i in range(0, test_num + 10):
+            inputs = np.random.rand(args.batch_size, 3, 112,
+                                    112).astype(np.float32)
+            start_time = time.time()
+            batch_output = self.predict(inputs).flatten()
+            if i >= 10:
+                test_time += time.time() - start_time
+            # time.sleep(0.01)  # sleep for T4 GPU
+
+        print("{0}\tbatch size: {1}\ttime(ms): {2}".format(
+            args.model, args.batch_size, 1000 * test_time / test_num))
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    assert os.path.exists(
+        args.model_file), "The path of 'model_file' does not exist: {}".format(
+            args.model_file)
+    assert os.path.exists(
+        args.params_file
+    ), "The path of 'params_file' does not exist: {}".format(args.params_file)
+
+    predictor = Predictor(args)
+    assert args.model is not None
+    predictor.benchmark_predict()
diff --git a/recognition/arcface_paddle/tools/convert_image_bin.py b/recognition/arcface_paddle/tools/convert_image_bin.py
new file mode 100644
index 000000000..589e66562
--- /dev/null
+++ b/recognition/arcface_paddle/tools/convert_image_bin.py
@@ -0,0 +1,92 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import sys
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(os.path.abspath(os.path.join(__dir__, '../')))
+
+import argparse
+import cv2
+from datasets.kv_helper import read_img_from_bin
+from datasets.kv_helper import trans_img_to_bin
+
+
+def get_file_list(img_file, end=('jpg', 'png', 'jpeg', 'JPEG', 'JPG', 'bmp')):
+    imgs_lists = []
+    if img_file is None or not os.path.exists(img_file):
+        raise Exception("not found any img file in {}".format(img_file))
+
+    if os.path.isfile(img_file) and img_file.split('.')[-1] in end:
+        imgs_lists.append(img_file)
+    elif os.path.isdir(img_file):
+        for single_file in os.listdir(img_file):
+            if single_file.split('.')[-1] in end:
+                imgs_lists.append(os.path.join(img_file, single_file))
+    if len(imgs_lists) == 0:
+        raise Exception("not found any img file in {}".format(img_file))
+    return imgs_lists
+
+
+def parse_args():
+    def str2bool(v):
+        return v.lower() in ("true", "t", "1")
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--image_path", type=str, default=None)
+    parser.add_argument("--bin_path", type=str, default=None)
+    parser.add_argument(
+        "--mode",
+        type=str,
+        default="image2bin",
+        help="conversion mode, image2bin or bin2image")
+    return parser.parse_args()
+
+
+def main(args):
+    assert args.mode in ["image2bin", "bin2image"]
+    os.makedirs(args.image_path, exist_ok=True)
+    os.makedirs(args.bin_path, exist_ok=True)
+    assert os.path.isdir(args.image_path)
+    assert os.path.isdir(args.bin_path)
+
+    if args.mode == "image2bin":
+        img_list = get_file_list(args.image_path)
+        for idx, img_fp in enumerate(img_list):
+            if idx % len(img_list) == 1000:
+                print("conversion process: [{}]/[{}]".format(idx,
+                                                             len(img_list)))
+            img_name = os.path.basename(img_fp)
+            output_path = os.path.join(args.bin_path,
+                                       os.path.splitext(img_name)[0] + ".bin")
+            trans_img_to_bin(img_fp, output_path)
+    elif args.mode == "bin2image":
+        bin_list = get_file_list(args.bin_path, end=("bin", ))
+        for idx, bin_fp in enumerate(bin_list):
+            if idx % len(bin_list) == 1000:
+                print("conversion process: [{}]/[{}]".format(idx,
+                                                             len(bin_list)))
+            bin_name = os.path.basename(bin_fp)
+            output_path = os.path.join(args.image_path,
+                                       os.path.splitext(bin_name)[0] + ".jpg")
+            img = read_img_from_bin(bin_fp)
+            cv2.imwrite(output_path, img)
+
+    print("ok..")
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)
diff --git a/recognition/arcface_paddle/tools/export.py b/recognition/arcface_paddle/tools/export.py
new file mode 100644
index 000000000..ace61aa4a
--- /dev/null
+++ b/recognition/arcface_paddle/tools/export.py
@@ -0,0 +1,72 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+sys.path.insert(0, os.path.abspath('.'))
+
+import argparse
+
+
+def str2bool(v):
+    return str(v).lower() in ("true", "t", "1")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Paddle Face Exporter')
+
+    # Model setting
+    parser.add_argument(
+        '--is_static',
+        type=str2bool,
+        default='False',
+        help='whether to use static mode')
+    parser.add_argument(
+        '--export_type',
+        type=str,
+        default='paddle',
+        help='export type, paddle or onnx')
+    parser.add_argument(
+        '--backbone',
+        type=str,
+        default='FresResNet50',
+        help='backbone network')
+    parser.add_argument(
+        '--embedding_size', type=int, default=512, help='embedding size')
+    parser.add_argument(
+        '--checkpoint_dir',
+        type=str,
+        default='MS1M_v3_arcface/FresResNet50/24/',
+        help='checkpoint direcotry')
+    parser.add_argument(
+        '--output_dir',
+        type=str,
+        default='MS1M_v3_arcface/FresResNet50/exported_model',
+        help='export output direcotry')
+
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    if args.is_static:
+        import paddle
+        paddle.enable_static()
+        from static.export import export
+    else:
+        from dynamic.export import export
+
+    assert args.export_type in ['paddle', 'onnx']
+    export(args)
diff --git a/recognition/arcface_paddle/tools/extract_perf_logs.py b/recognition/arcface_paddle/tools/extract_perf_logs.py
new file mode 100644
index 000000000..ab91f0432
--- /dev/null
+++ b/recognition/arcface_paddle/tools/extract_perf_logs.py
@@ -0,0 +1,153 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import glob
+import json
+import argparse
+import pprint
+
+import numpy as np
+
+pp = pprint.PrettyPrinter(indent=1)
+
+parser = argparse.ArgumentParser(description="flags for benchmark")
+parser.add_argument("--log_dir", type=str, default="./logs/", required=True)
+parser.add_argument(
+    "--output_dir", type=str, default="./logs/", required=False)
+parser.add_argument('--warmup_batches', type=int, default=50)
+parser.add_argument('--train_batches', type=int, default=150)
+
+args = parser.parse_args()
+
+
+class AutoVivification(dict):
+    """Implementation of perl's autovivification feature."""
+
+    def __getitem__(self, item):
+        try:
+            return dict.__getitem__(self, item)
+        except KeyError:
+            value = self[item] = type(self)()
+        return value
+
+
+def compute_median(iter_dict):
+    speed_list = [i for i in iter_dict.values()]
+    return round(np.median(speed_list), 2)
+
+
+def compute_average(iter_dict):
+    i = 0
+    total_speed = 0
+    for iter in iter_dict:
+        i += 1
+        total_speed += iter_dict[iter]
+    return round(total_speed / i, 2)
+
+
+def extract_info_from_file(log_file, result_dict, speed_dict):
+    # extract info from file name
+    exp_config = log_file.split("/")[-2]
+    model = exp_config.split("_")[2]
+    mode = exp_config.split("_")[3]
+    precision = exp_config.split("_")[4]
+    batch_size_per_device = exp_config.split("_")[6]
+    run_case = exp_config.split("_")[7]  # eg: 1n1g
+    test_iter = int(exp_config.split("_")[8][2:])
+    node_num = int(run_case[0])
+    if len(run_case) == 4:
+        card_num = int(run_case[-2])
+    elif len(run_case) == 5:
+        card_num = int(run_case[-3:-1])
+
+    avg_speed_list = []
+    # extract info from file content
+    with open(log_file) as f:
+        lines = f.readlines()
+        for line in lines:
+            if "throughput:" in line:
+                p1 = re.compile(r" ips: ([0-9]+\.[0-9]+)", re.S)
+                item = re.findall(p1, line)
+                a = float(item[0].strip())
+                avg_speed_list.append(a)
+
+    # compute avg throughoutput
+    avg_speed = round(
+        np.mean(avg_speed_list[args.warmup_batches:args.train_batches]), 2)
+
+    speed_dict[mode][model][run_case][precision][batch_size_per_device][
+        test_iter] = avg_speed
+    average_speed = compute_average(speed_dict[mode][model][run_case][
+        precision][batch_size_per_device])
+    median_speed = compute_median(speed_dict[mode][model][run_case][precision][
+        batch_size_per_device])
+
+    result_dict[mode][model][run_case][precision][batch_size_per_device][
+        'average_speed'] = average_speed
+    result_dict[mode][model][run_case][precision][batch_size_per_device][
+        'median_speed'] = median_speed
+
+    # print(log_file, speed_dict[mode][model][run_case])
+
+
+def compute_speedup(result_dict, speed_dict):
+    mode_list = [key for key in result_dict]  # eg. ['static', 'dynamic']
+    for md in mode_list:
+        model_list = [key for key in result_dict[md]]  # eg.['vgg16', 'r50']
+        for m in model_list:
+            run_case = [key for key in result_dict[md][m]
+                        ]  # eg.['4n8g', '2n8g', '1n8g', '1n4g', '1n1g']
+            for d in run_case:
+                precision = [key for key in result_dict[md][m][d]]
+                for p in precision:
+                    batch_size_per_device = [
+                        key for key in result_dict[md][m][d][p]
+                    ]
+                    for b in batch_size_per_device:
+                        speed_up = 1.0
+                        if result_dict[md][m]['1n1g'][p][b]['median_speed']:
+                            speed_up = result_dict[md][m][d][p][b][
+                                'median_speed'] / result_dict[md][m]['1n1g'][
+                                    p][b]['median_speed']
+                        result_dict[md][m][d][p][b]['speedup'] = round(
+                            speed_up, 2)
+
+
+def extract_result():
+    result_dict = AutoVivification()
+    speed_dict = AutoVivification()
+    logs_list = glob.glob(os.path.join(args.log_dir, "*/workerlog.0"))
+    for l in logs_list:
+        extract_info_from_file(l, result_dict, speed_dict)
+
+    # compute speedup
+    compute_speedup(result_dict, speed_dict)
+
+    # print result
+    pp.pprint(result_dict)
+
+    # write to file as JSON format
+    os.makedirs(args.output_dir, exist_ok=True)
+    result_file_name = os.path.join(args.output_dir,
+                                    "arcface_paddle_result.json")
+    print("Saving result to {}".format(result_file_name))
+    with open(result_file_name, 'w') as f:
+        json.dump(result_dict, f)
+
+
+if __name__ == "__main__":
+    extract_result()
diff --git a/recognition/arcface_paddle/tools/inference.py b/recognition/arcface_paddle/tools/inference.py
new file mode 100644
index 000000000..822612a7b
--- /dev/null
+++ b/recognition/arcface_paddle/tools/inference.py
@@ -0,0 +1,234 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+import cv2
+import time
+import json
+import argparse
+import numpy as np
+
+sys.path.insert(0, os.path.abspath('.'))
+
+def str2bool(v):
+    return v.lower() in ("True","true", "t", "1")
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Paddle Face Predictor')
+
+    parser.add_argument(
+        '--export_type', type=str, help='export type, paddle or onnx')
+    parser.add_argument(
+        "--model_file",
+        type=str,
+        required=False,
+        help="paddle save inference model filename")
+    parser.add_argument(
+        "--params_file",
+        type=str,
+        required=False,
+        help="paddle save inference parameter filename")
+    parser.add_argument(
+        "--onnx_file", type=str, required=False, help="onnx model filename")
+    parser.add_argument("--image_path", type=str, help="path to test image")
+    parser.add_argument("--benchmark", type=str2bool, default=False, help="Is benchmark mode")
+    # params for paddle inferece engine
+    parser.add_argument("--use_gpu", type=str2bool, default=True)
+    parser.add_argument("--ir_optim", type=str2bool, default=True)
+    parser.add_argument("--use_tensorrt", type=str2bool, default=False)
+    parser.add_argument("--min_subgraph_size", type=int, default=15)
+    parser.add_argument("--max_batch_size", type=int, default=1)
+    parser.add_argument("--precision", type=str, default="fp32")
+    parser.add_argument("--gpu_mem", type=int, default=500)
+    parser.add_argument("--enable_mkldnn", type=str2bool, default=False)
+    parser.add_argument("--cpu_threads", type=int, default=10)
+    args = parser.parse_args()
+    return args
+
+def get_infer_gpuid():
+    cmd = "nvidia-smi"
+    res = os.popen(cmd).readlines()
+    if len(res) == 0:
+        return None
+    cmd = "env | grep CUDA_VISIBLE_DEVICES"
+    env_cuda = os.popen(cmd).readlines()
+    if len(env_cuda) == 0:
+        return 0
+    else:
+        gpu_id = env_cuda[0].strip().split("=")[1]
+        return int(gpu_id[0])
+
+
+def init_paddle_inference_config(args):
+    import paddle.inference as paddle_infer
+    config = paddle_infer.Config(args.model_file, args.params_file)
+    if hasattr(args, 'precision'):
+        if args.precision == "fp16" and args.use_tensorrt:
+            precision = paddle_infer.PrecisionType.Half
+        elif args.precision == "int8":
+            precision = paddle_infer.PrecisionType.Int8
+        else:
+            precision = paddle_infer.PrecisionType.Float32
+    else:
+        precision = paddle_infer.PrecisionType.Float32
+
+    if args.use_gpu:
+        gpu_id = get_infer_gpuid()
+        if gpu_id is None:
+            raise ValueError(
+                "Not found GPU in current device. Please check your device or set args.use_gpu as False"
+            )
+        config.enable_use_gpu(args.gpu_mem, 0)
+        if args.use_tensorrt:
+            config.enable_tensorrt_engine(
+                precision_mode=precision,
+                max_batch_size=args.max_batch_size,
+                min_subgraph_size=args.min_subgraph_size)
+            # skip the minmum trt subgraph
+            min_input_shape = {"x": [1, 3, 10, 10]}
+            max_input_shape = {"x": [1, 3, 1000, 1000]}
+            opt_input_shape = {"x": [1, 3, 112, 112]}
+            config.set_trt_dynamic_shape_info(min_input_shape, max_input_shape,
+                                            opt_input_shape)
+
+    else:
+        config.disable_gpu()
+        cpu_threads = args.cpu_threads if  hasattr(args, "cpu_threads") else 10
+        config.set_cpu_math_library_num_threads(cpu_threads)
+        if args.enable_mkldnn:
+            # cache 10 different shapes for mkldnn to avoid memory leak
+            config.enable_mkldnn()
+            config.set_mkldnn_cache_capacity(10)
+            if args.precision == "fp16":
+                config.enable_mkldnn_bfloat16()
+    return config
+
+
+def get_image_file_list(img_file):
+    import imghdr
+    imgs_lists = []
+    if img_file is None or not os.path.exists(img_file):
+        raise Exception("not found any img file in {}".format(img_file))
+
+    img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif', 'GIF'}
+    if os.path.isfile(img_file) and imghdr.what(img_file) in img_end:
+        imgs_lists.append(img_file)
+    elif os.path.isdir(img_file):
+        for single_file in os.listdir(img_file):
+            file_path = os.path.join(img_file, single_file)
+            if os.path.isfile(file_path) and imghdr.what(file_path) in img_end:
+                imgs_lists.append(file_path)
+    if len(imgs_lists) == 0:
+        raise Exception("not found any img file in {}".format(img_file))
+    imgs_lists = sorted(imgs_lists)
+    return imgs_lists
+
+def paddle_inference(args):
+    import paddle.inference as paddle_infer
+
+    config =  init_paddle_inference_config(args)
+    predictor = paddle_infer.create_predictor(config)
+
+    input_names = predictor.get_input_names()
+    input_handle = predictor.get_input_handle(input_names[0])
+
+    if args.benchmark:
+        import auto_log
+        pid = os.getpid()
+        autolog = auto_log.AutoLogger(
+            model_name="det",
+            model_precision='fp32',
+            batch_size=1,
+            data_shape="dynamic",
+            save_path="./output/auto_log.log",
+            inference_config=config,
+            pids=pid,
+            process_name=None,
+            gpu_ids=0,
+            time_keys=[
+                'preprocess_time', 'inference_time','postprocess_time'
+            ],
+            warmup=0)
+        img = np.random.uniform(0, 255, [1, 3, 112,112]).astype(np.float32)
+        input_handle.copy_from_cpu(img)
+        for i in range(10):
+            predictor.run()
+
+
+    img_list = get_image_file_list(args.image_path)
+    for img_path in img_list:
+        img = cv2.imread(img_path)
+        st = time.time()
+        if args.benchmark:
+            autolog.times.start()
+
+        # normalize to mean 0.5, std 0.5
+        img = (img - 127.5) * 0.00784313725
+        # BGR2RGB
+        img = img[:, :, ::-1]
+        img = img.transpose((2, 0, 1))
+        img = np.expand_dims(img, 0)
+        img = img.astype('float32')
+
+        if args.benchmark:
+            autolog.times.stamp()
+
+        input_handle.copy_from_cpu(img)
+
+        predictor.run()
+
+        output_names = predictor.get_output_names()
+        output_handle = predictor.get_output_handle(output_names[0])
+        output_data = output_handle.copy_to_cpu()
+        if args.benchmark:
+            autolog.times.stamp()
+            autolog.times.end(stamp=True)
+            print('{}\t{}'.format(img_path,json.dumps(output_data.tolist())))
+        print('paddle inference result: ', output_data.shape)
+    if args.benchmark:
+        autolog.report()
+
+def onnx_inference(args):
+    import onnxruntime
+
+    ort_sess = onnxruntime.InferenceSession(args.onnx_file)
+
+    img = cv2.imread(args.image_path)
+    # normalize to mean 0.5, std 0.5
+    img = (img - 127.5) * 0.00784313725
+    # BGR2RGB
+    img = img[:, :, ::-1]
+    img = img.transpose((2, 0, 1))
+    img = np.expand_dims(img, 0)
+    img = img.astype('float32')
+
+    ort_inputs = {ort_sess.get_inputs()[0].name: img}
+    ort_outs = ort_sess.run(None, ort_inputs)
+
+    print('onnx inference result: ', ort_outs[0].shape)
+
+
+if __name__ == '__main__':
+
+    args = parse_args()
+
+    assert args.export_type in ['paddle', 'onnx']
+    if args.export_type == 'onnx':
+        assert os.path.exists(args.onnx_file)
+        onnx_inference(args)
+    else:
+        assert os.path.exists(args.model_file)
+        assert os.path.exists(args.params_file)
+        paddle_inference(args)
diff --git a/recognition/arcface_paddle/tools/mx_recordio_2_images.py b/recognition/arcface_paddle/tools/mx_recordio_2_images.py
new file mode 100644
index 000000000..fba25b9b4
--- /dev/null
+++ b/recognition/arcface_paddle/tools/mx_recordio_2_images.py
@@ -0,0 +1,82 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import argparse
+import numpy as np
+import numbers
+import mxnet as mx
+import cv2
+import tqdm
+import shutil
+
+
+def main(args):
+    path_imgrec = os.path.join(args.root_dir, 'train.rec')
+    path_imgidx = os.path.join(args.root_dir, 'train.idx')
+    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+    s = imgrec.read_idx(0)
+    header, _ = mx.recordio.unpack(s)
+    if header.flag > 0:
+        header0 = (int(header.label[0]), int(header.label[1]))
+        imgidx = np.array(range(1, int(header.label[0])))
+    else:
+        imgidx = np.array(list(imgrec.keys))
+
+    classes = set()
+    os.makedirs(os.path.join(args.output_dir, 'images'), exist_ok=True)
+    fp = open(os.path.join(args.output_dir, 'label.txt'), 'w')
+    for idx in tqdm.tqdm(imgidx):
+        s = imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        label = header.label
+        if not isinstance(label, numbers.Number):
+            label = label[0]
+        img = mx.image.imdecode(img).asnumpy()[..., ::-1]
+        label = int(label)
+        classes.add(label)
+
+        filename = 'images/%08d.jpg' % idx
+        fp.write('%s\t%d\n' % (filename, label))
+        cv2.imwrite(
+            os.path.join(args.output_dir, filename), img,
+            [int(cv2.IMWRITE_JPEG_QUALITY), 100])
+    fp.close()
+    shutil.copy(
+        os.path.join(args.root_dir, 'agedb_30.bin'),
+        os.path.join(args.output_dir, 'agedb_30.bin'))
+    shutil.copy(
+        os.path.join(args.root_dir, 'cfp_fp.bin'),
+        os.path.join(args.output_dir, 'cfp_fp.bin'))
+    shutil.copy(
+        os.path.join(args.root_dir, 'lfw.bin'),
+        os.path.join(args.output_dir, 'lfw.bin'))
+    print('num_image: ', len(imgidx), 'num_classes: ', len(classes))
+    with open(os.path.join(args.output_dir, 'README.md'), 'w') as f:
+        f.write('num_image: {}\n'.format(len(imgidx)))
+        f.write('num_classes: {}\n'.format(len(classes)))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--root_dir",
+        type=str,
+        help="Root directory to mxnet dataset.", )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        help="Path to output.", )
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/arcface_paddle/tools/test_recognition.py b/recognition/arcface_paddle/tools/test_recognition.py
new file mode 100644
index 000000000..a5972b68b
--- /dev/null
+++ b/recognition/arcface_paddle/tools/test_recognition.py
@@ -0,0 +1,720 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import argparse
+import requests
+import logging
+import imghdr
+import pickle
+import tarfile
+from functools import partial
+
+import cv2
+import numpy as np
+from sklearn.metrics.pairwise import cosine_similarity
+from tqdm import tqdm
+from prettytable import PrettyTable
+from PIL import Image, ImageDraw, ImageFont
+import paddle
+from paddle.inference import Config
+from paddle.inference import create_predictor
+
+__all__ = ["InsightFace", "parser"]
+BASE_INFERENCE_MODEL_DIR = os.path.expanduser("~/.insightface/ppmodels/")
+BASE_DOWNLOAD_URL = "https://paddle-model-ecology.bj.bcebos.com/model/insight-face/{}.tar"
+
+
+def parser(add_help=True):
+    def str2bool(v):
+        return v.lower() in ("true", "t", "1")
+
+    parser = argparse.ArgumentParser(add_help=add_help)
+    parser.add_argument(
+        "--det", action="store_true", help="Whether to detect.")
+    parser.add_argument(
+        "--rec", action="store_true", help="Whether to recognize.")
+    
+    parser.add_argument(
+        "--det_model",
+        type=str,
+        default="BlazeFace",
+        help="The detection model.")
+    parser.add_argument(
+        "--rec_model",
+        type=str,
+        default="MobileFace",
+        help="The recognition model.")
+    parser.add_argument(
+        "--use_gpu",
+        type=str2bool,
+        default=True,
+        help="Whether use GPU to predict. Default by True.")
+    parser.add_argument(
+        "--enable_mkldnn",
+        type=str2bool,
+        default=True,
+        help="Whether use MKLDNN to predict, valid only when --use_gpu is False. Default by False."
+    )
+    parser.add_argument(
+        "--cpu_threads",
+        type=int,
+        default=1,
+        help="The num of threads with CPU, valid only when --use_gpu is False. Default by 1."
+    )
+    parser.add_argument(
+        "--input",
+        type=str,
+        help="The path or directory of image(s) or video to be predicted.")
+    parser.add_argument(
+        "--output", type=str, default="./output/", help="The directory of prediction result.")
+    parser.add_argument(
+        "--det_thresh",
+        type=float,
+        default=0.8,
+        help="The threshold of detection postprocess. Default by 0.8.")
+    parser.add_argument(
+        "--index", type=str, default=None, help="The path of index file.")
+    parser.add_argument(
+        "--cdd_num",
+        type=int,
+        default=5,
+        help="The number of candidates in the recognition retrieval. Default by 10."
+    )
+    parser.add_argument(
+        "--rec_thresh",
+        type=float,
+        default=0.45,
+        help="The threshold of recognition postprocess. Default by 0.45.")
+    parser.add_argument(
+        "--max_batch_size",
+        type=int,
+        default=1,
+        help="The maxium of batch_size to recognize. Default by 1.")
+
+    return parser
+
+
+def print_config(args):
+    args = vars(args)
+    table = PrettyTable(['Param', 'Value'])
+    for param in args:
+        table.add_row([param, args[param]])
+    width = len(str(table).split("\n")[0])
+    print("{}".format("-" * width))
+    print("PaddleFace".center(width))
+    print(table)
+    print("Powered by PaddlePaddle!".rjust(width))
+    print("{}".format("-" * width))
+
+
+def download_with_progressbar(url, save_path):
+    """Download from url with progressbar.
+    """
+    if os.path.isfile(save_path):
+        os.remove(save_path)
+    response = requests.get(url, stream=True)
+    total_size_in_bytes = int(response.headers.get("content-length", 0))
+    block_size = 1024  # 1 Kibibyte
+    progress_bar = tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True)
+    with open(save_path, "wb") as file:
+        for data in response.iter_content(block_size):
+            progress_bar.update(len(data))
+            file.write(data)
+    progress_bar.close()
+    if total_size_in_bytes == 0 or progress_bar.n != total_size_in_bytes or not os.path.isfile(
+            save_path):
+        raise Exception(
+            f"Something went wrong while downloading model/image from {url}")
+
+
+def check_model_file(model):
+    """Check the model files exist and download and untar when no exist.
+    """
+    model_map = {
+        "ArcFace": "arcface_iresnet50_v1.0_infer",
+        "BlazeFace": "blazeface_fpn_ssh_1000e_v1.0_infer",
+        "MobileFace": "mobileface_v1.0_infer"
+    }
+
+    if os.path.isdir(model):
+        model_file_path = os.path.join(model, "inference.pdmodel")
+        params_file_path = os.path.join(model, "inference.pdiparams")
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            raise Exception(
+                f"The specifed model directory error. The drectory must include 'inference.pdmodel' and 'inference.pdiparams'."
+            )
+
+    elif model in model_map:
+        storage_directory = partial(os.path.join, BASE_INFERENCE_MODEL_DIR,
+                                    model)
+        url = BASE_DOWNLOAD_URL.format(model_map[model])
+
+        tar_file_name_list = [
+            "inference.pdiparams", "inference.pdiparams.info",
+            "inference.pdmodel"
+        ]
+        model_file_path = storage_directory("inference.pdmodel")
+        params_file_path = storage_directory("inference.pdiparams")
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            tmp_path = storage_directory(url.split("/")[-1])
+            logging.info(f"Download {url} to {tmp_path}")
+            os.makedirs(storage_directory(), exist_ok=True)
+            download_with_progressbar(url, tmp_path)
+            with tarfile.open(tmp_path, "r") as tarObj:
+                for member in tarObj.getmembers():
+                    filename = None
+                    for tar_file_name in tar_file_name_list:
+                        if tar_file_name in member.name:
+                            filename = tar_file_name
+                    if filename is None:
+                        continue
+                    file = tarObj.extractfile(member)
+                    with open(storage_directory(filename), "wb") as f:
+                        f.write(file.read())
+            os.remove(tmp_path)
+        if not os.path.exists(model_file_path) or not os.path.exists(
+                params_file_path):
+            raise Exception(
+                f"Something went wrong while downloading and unzip the model[{model}] files!"
+            )
+    else:
+        raise Exception(
+            f"The specifed model name error. Support 'BlazeFace' for detection and 'ArcFace' and 'MobileFace' for recognition. And support local directory that include model files ('inference.pdmodel' and 'inference.pdiparams')."
+        )
+
+    return model_file_path, params_file_path
+
+
+def normalize_image(img, scale=None, mean=None, std=None, order='chw'):
+    if isinstance(scale, str):
+        scale = eval(scale)
+    scale = np.float32(scale if scale is not None else 1.0 / 255.0)
+    mean = mean if mean is not None else [0.485, 0.456, 0.406]
+    std = std if std is not None else [0.229, 0.224, 0.225]
+
+    shape = (3, 1, 1) if order == 'chw' else (1, 1, 3)
+    mean = np.array(mean).reshape(shape).astype('float32')
+    std = np.array(std).reshape(shape).astype('float32')
+
+    if isinstance(img, Image.Image):
+        img = np.array(img)
+
+    assert isinstance(img, np.ndarray), "invalid input 'img' in NormalizeImage"
+    return (img.astype('float32') * scale - mean) / std
+
+
+def to_CHW_image(img):
+    if isinstance(img, Image.Image):
+        img = np.array(img)
+    return img.transpose((2, 0, 1))
+
+
+class ColorMap(object):
+    def __init__(self, num):
+        super().__init__()
+        self.get_color_map_list(num)
+        self.color_map = {}
+        self.ptr = 0
+
+    def __getitem__(self, key):
+        return self.color_map[key]
+
+    def update(self, keys):
+        for key in keys:
+            if key not in self.color_map:
+                i = self.ptr % len(self.color_list)
+                self.color_map[key] = self.color_list[i]
+                self.ptr += 1
+
+    def get_color_map_list(self, num_classes):
+        color_map = num_classes * [0, 0, 0]
+        for i in range(0, num_classes):
+            j = 0
+            lab = i
+            while lab:
+                color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
+                color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
+                color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
+                j += 1
+                lab >>= 3
+        self.color_list = [
+            color_map[i:i + 3] for i in range(0, len(color_map), 3)
+        ]
+
+
+class ImageReader(object):
+    def __init__(self, inputs):
+        super().__init__()
+        self.idx = 0
+        if isinstance(inputs, np.ndarray):
+            self.image_list = [inputs]
+        else:
+            imgtype_list = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff'}
+            self.image_list = []
+            if os.path.isfile(inputs):
+                if imghdr.what(inputs) not in imgtype_list:
+                    raise Exception(
+                        f"Error type of input path, only support: {imgtype_list}"
+                    )
+                self.image_list.append(inputs)
+            elif os.path.isdir(inputs):
+                tmp_file_list = os.listdir(inputs)
+                warn_tag = False
+                for file_name in tmp_file_list:
+                    file_path = os.path.join(inputs, file_name)
+                    if not os.path.isfile(file_path):
+                        warn_tag = True
+                        continue
+                    if imghdr.what(file_path) in imgtype_list:
+                        self.image_list.append(file_path)
+                    else:
+                        warn_tag = True
+                if warn_tag:
+                    logging.warning(
+                        f"The directory of input contine directory or not supported file type, only support: {imgtype_list}"
+                    )
+            else:
+                raise Exception(
+                    f"The file of input path not exist! Please check input: {inputs}"
+                )
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        if self.idx >= len(self.image_list):
+            raise StopIteration
+
+        data = self.image_list[self.idx]
+        if isinstance(data, np.ndarray):
+            self.idx += 1
+            return data, "tmp.png"
+        path = data
+        _, file_name = os.path.split(path)
+        img = cv2.imread(path)
+        if img is None:
+            logging.warning(f"Error in reading image: {path}! Ignored.")
+            self.idx += 1
+            return self.__next__()
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        self.idx += 1
+        return img, file_name
+
+    def __len__(self):
+        return len(self.image_list)
+
+
+class VideoReader(object):
+    def __init__(self, inputs):
+        super().__init__()
+        videotype_list = {"mp4"}
+        if os.path.splitext(inputs)[-1][1:] not in videotype_list:
+            raise Exception(
+                f"The input file is not supported, only support: {videotype_list}"
+            )
+        if not os.path.isfile(inputs):
+            raise Exception(
+                f"The file of input path not exist! Please check input: {inputs}"
+            )
+        self.capture = cv2.VideoCapture(inputs)
+        self.file_name = os.path.split(inputs)[-1]
+
+    def get_info(self):
+        info = {}
+        width = int(self.capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(self.capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        fourcc = cv2.VideoWriter_fourcc(* 'mp4v')
+        info["file_name"] = self.file_name
+        info["fps"] = 30
+        info["shape"] = (width, height)
+        info["fourcc"] = cv2.VideoWriter_fourcc(* 'mp4v')
+        return info
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        ret, frame = self.capture.read()
+        if not ret:
+            raise StopIteration
+        return frame, self.file_name
+
+
+class ImageWriter(object):
+    def __init__(self, output_dir):
+        super().__init__()
+        if output_dir is None:
+            raise Exception(
+                "Please specify the directory of saving prediction results by --output."
+            )
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+        self.output_dir = output_dir
+
+    def write(self, image, file_name):
+        path = os.path.join(self.output_dir, file_name)
+        cv2.imwrite(path, cv2.cvtColor(image, cv2.COLOR_RGB2BGR))
+
+
+class VideoWriter(object):
+    def __init__(self, output_dir, video_info):
+        super().__init__()
+        if output_dir is None:
+            raise Exception(
+                "Please specify the directory of saving prediction results by --output."
+            )
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+        output_path = os.path.join(output_dir, video_info["file_name"])
+        fourcc = cv2.VideoWriter_fourcc(* 'mp4v')
+        self.writer = cv2.VideoWriter(output_path, video_info["fourcc"],
+                                      video_info["fps"], video_info["shape"])
+
+    def write(self, frame, file_name):
+        self.writer.write(frame)
+
+    def __del__(self):
+        if hasattr(self, "writer"):
+            self.writer.release()
+
+
+class BasePredictor(object):
+    def __init__(self, predictor_config):
+        super().__init__()
+        self.predictor_config = predictor_config
+        self.predictor, self.input_names, self.output_names = self.load_predictor(
+            predictor_config["model_file"], predictor_config["params_file"])
+
+    def load_predictor(self, model_file, params_file):
+        config = Config(model_file, params_file)
+        if self.predictor_config["use_gpu"]:
+            config.enable_use_gpu(200, 0)
+            config.switch_ir_optim(True)
+        else:
+            config.disable_gpu()
+            config.set_cpu_math_library_num_threads(self.predictor_config[
+                "cpu_threads"])
+
+            if self.predictor_config["enable_mkldnn"]:
+                try:
+                    # cache 10 different shapes for mkldnn to avoid memory leak
+                    config.set_mkldnn_cache_capacity(10)
+                    config.enable_mkldnn()
+                except Exception as e:
+                    logging.error(
+                        "The current environment does not support `mkldnn`, so disable mkldnn."
+                    )
+        config.disable_glog_info()
+        config.enable_memory_optim()
+        # use zero copy
+        config.switch_use_feed_fetch_ops(False)
+        predictor = create_predictor(config)
+        input_names = predictor.get_input_names()
+        output_names = predictor.get_output_names()
+        return predictor, input_names, output_names
+
+    def preprocess(self):
+        raise NotImplementedError
+
+    def postprocess(self):
+        raise NotImplementedError
+
+    def predict(self, img):
+        raise NotImplementedError
+
+
+class Detector(BasePredictor):
+    def __init__(self, det_config, predictor_config):
+        super().__init__(predictor_config)
+        self.det_config = det_config
+        self.target_size = self.det_config["target_size"]
+        self.thresh = self.det_config["thresh"]
+
+    def preprocess(self, img):
+        resize_h, resize_w = self.target_size
+        img_shape = img.shape
+        img_scale_x = resize_w / img_shape[1]
+        img_scale_y = resize_h / img_shape[0]
+        img = cv2.resize(
+            img, None, None, fx=img_scale_x, fy=img_scale_y, interpolation=1)
+        img = normalize_image(
+            img,
+            scale=1. / 255.,
+            mean=[0.485, 0.456, 0.406],
+            std=[0.229, 0.224, 0.225],
+            order='hwc')
+        img_info = {}
+        img_info["im_shape"] = np.array(
+            img.shape[:2], dtype=np.float32)[np.newaxis, :]
+        img_info["scale_factor"] = np.array(
+            [img_scale_y, img_scale_x], dtype=np.float32)[np.newaxis, :]
+
+        img = img.transpose((2, 0, 1)).copy()
+        img_info["image"] = img[np.newaxis, :, :, :]
+        return img_info
+
+    def postprocess(self, np_boxes):
+        expect_boxes = (np_boxes[:, 1] > self.thresh) & (np_boxes[:, 0] > -1)
+        return np_boxes[expect_boxes, :]
+
+    def predict(self, img):
+        inputs = self.preprocess(img)
+        for input_name in self.input_names:
+            input_tensor = self.predictor.get_input_handle(input_name)
+            input_tensor.copy_from_cpu(inputs[input_name])
+        self.predictor.run()
+        output_tensor = self.predictor.get_output_handle(self.output_names[0])
+        np_boxes = output_tensor.copy_to_cpu()
+        # boxes_num = self.detector.get_output_handle(self.detector_output_names[1])
+        # np_boxes_num = boxes_num.copy_to_cpu()
+        box_list = self.postprocess(np_boxes)
+        return box_list
+
+
+class Recognizer(BasePredictor):
+    def __init__(self, rec_config, predictor_config):
+        super().__init__(predictor_config)
+        if rec_config["index"] is not None:
+            self.load_index(rec_config["index"])
+        self.rec_config = rec_config
+        self.cdd_num = self.rec_config["cdd_num"]
+        self.thresh = self.rec_config["thresh"]
+        self.max_batch_size = self.rec_config["max_batch_size"]
+
+    def preprocess(self, img, box_list=None):
+        img = normalize_image(
+            img,
+            scale=1. / 255.,
+            mean=[0.5, 0.5, 0.5],
+            std=[0.5, 0.5, 0.5],
+            order='hwc')
+        if box_list is None:
+            height, width = img.shape[:2]
+            box_list = [np.array([0, 0, 0, 0, width, height])]
+        batch = []
+        input_batches = []
+        cnt = 0
+        for idx, box in enumerate(box_list):
+            box[box < 0] = 0
+            xmin, ymin, xmax, ymax = list(map(int, box[2:]))
+            face_img = img[ymin:ymax, xmin:xmax, :]
+            face_img = cv2.resize(face_img, (112, 112)).transpose(
+                (2, 0, 1)).copy()
+            batch.append(face_img)
+            cnt += 1
+            if cnt % self.max_batch_size == 0 or (idx + 1) == len(box_list):
+                input_batches.append(np.array(batch))
+                batch = []
+        return input_batches
+
+    def postprocess(self):
+        pass
+
+    def retrieval(self, np_feature):
+        labels = []
+        for feature in np_feature:
+            similarity = cosine_similarity(self.index_feature,
+                                           feature).squeeze()
+            abs_similarity = np.abs(similarity)
+            candidate_idx = np.argpartition(abs_similarity,
+                                            -self.cdd_num)[-self.cdd_num:]
+            remove_idx = np.where(abs_similarity[candidate_idx] < self.thresh)
+            candidate_idx = np.delete(candidate_idx, remove_idx)
+            candidate_label_list = list(np.array(self.label)[candidate_idx])
+            if len(candidate_label_list) == 0:
+                maxlabel = ""
+            else:
+                maxlabel = max(candidate_label_list,
+                               key=candidate_label_list.count)
+            labels.append(maxlabel)
+        return labels
+
+    def load_index(self, file_path):
+        with open(file_path, "rb") as f:
+            index = pickle.load(f)
+        self.label = index["label"]
+        self.index_feature = np.array(index["feature"]).squeeze()
+
+    def predict(self, img, box_list=None):
+        batch_list = self.preprocess(img, box_list)
+        feature_list = []
+        for batch in batch_list:
+            for input_name in self.input_names:
+                input_tensor = self.predictor.get_input_handle(input_name)
+                input_tensor.copy_from_cpu(batch)
+            self.predictor.run()
+            output_tensor = self.predictor.get_output_handle(self.output_names[
+                0])
+            np_feature = output_tensor.copy_to_cpu()
+            feature_list.append(np_feature)
+        return np.array(feature_list)
+
+
+class InsightFace(object):
+    def __init__(self, args, print_info=True):
+        super().__init__()
+        if print_info:
+            print_config(args)
+
+        self.font_path = os.path.join(
+            os.path.abspath(os.path.dirname(__file__)),
+            "SourceHanSansCN-Medium.otf")
+        self.args = args
+
+        predictor_config = {
+            "use_gpu": args.use_gpu,
+            "enable_mkldnn": args.enable_mkldnn,
+            "cpu_threads": args.cpu_threads
+        }
+        if args.det:
+            model_file_path, params_file_path = check_model_file(
+                args.det_model)
+            det_config = {"thresh": args.det_thresh, "target_size": [640, 640]}
+            predictor_config["model_file"] = model_file_path
+            predictor_config["params_file"] = params_file_path
+            self.det_predictor = Detector(det_config, predictor_config)
+            self.color_map = ColorMap(100)
+
+        if args.rec:
+            model_file_path, params_file_path = check_model_file(
+                args.rec_model)
+            rec_config = {
+                "max_batch_size": args.max_batch_size,
+                "resize": 112,
+                "thresh": args.rec_thresh,
+                "index": args.index,
+                "cdd_num": args.cdd_num
+            }
+            predictor_config["model_file"] = model_file_path
+            predictor_config["params_file"] = params_file_path
+            self.rec_predictor = Recognizer(rec_config, predictor_config)
+
+    def preprocess(self, img):
+        img = img.astype(np.float32, copy=False)
+        return img
+
+    def draw(self, img, box_list, labels):
+        self.color_map.update(labels)
+        im = Image.fromarray(img)
+        draw = ImageDraw.Draw(im)
+
+        for i, dt in enumerate(box_list):
+            bbox, score = dt[2:], dt[1]
+            label = labels[i]
+            color = tuple(self.color_map[label])
+
+            xmin, ymin, xmax, ymax = bbox
+
+            font_size = max(int((xmax - xmin) // 6), 10)
+            font = ImageFont.truetype(self.font_path, font_size)
+
+            text = "{} {:.4f}".format(label, score)
+            th = sum(font.getmetrics())
+            tw = font.getsize(text)[0]
+            start_y = max(0, ymin - th)
+
+            draw.rectangle(
+                [(xmin, start_y), (xmin + tw + 1, start_y + th)], fill=color)
+            draw.text(
+                (xmin + 1, start_y),
+                text,
+                fill=(255, 255, 255),
+                font=font,
+                anchor="la")
+            draw.rectangle(
+                [(xmin, ymin), (xmax, ymax)], width=2, outline=color)
+        return np.array(im)
+
+    def predict_np_img(self, img):
+        input_img = self.preprocess(img)
+        box_list = None
+        np_feature = None
+        if hasattr(self, "det_predictor"):
+            box_list = self.det_predictor.predict(input_img)
+        if hasattr(self, "rec_predictor"):
+            np_feature = self.rec_predictor.predict(input_img, box_list)
+        return box_list, np_feature
+
+    def init_reader_writer(self, input_data):
+        if isinstance(input_data, np.ndarray):
+            self.input_reader = ImageReader(input_data)
+            if hasattr(self, "det_predictor"):
+                self.output_writer = ImageWriter(self.args.output)
+        elif isinstance(input_data, str):
+            if input_data.endswith('mp4'):
+                self.input_reader = VideoReader(input_data)
+                info = self.input_reader.get_info()
+                self.output_writer = VideoWriter(self.args.output, info)
+            else:
+                self.input_reader = ImageReader(input_data)
+                if hasattr(self, "det_predictor"):
+                    self.output_writer = ImageWriter(self.args.output)
+        else:
+            raise Exception(
+                f"The input data error. Only support path of image or video(.mp4) and dirctory that include images."
+            )
+
+    def predict(self, input_data, print_info=False):
+        """Predict input_data.
+
+        Args:
+            input_data (str | NumPy.array): The path of image, or the derectory including images, or the image data in NumPy.array format.
+            print_info (bool, optional): Wheather to print the prediction results. Defaults to False.
+
+        Yields:
+            dict: {
+                "box_list": The prediction results of detection.
+                "features": The output of recognition.
+                "labels": The results of retrieval.
+                }
+        """
+        self.init_reader_writer(input_data)
+        for img, file_name in self.input_reader:
+            if img is None:
+                logging.warning(f"Error in reading img {file_name}! Ignored.")
+                continue
+            box_list, np_feature = self.predict_np_img(img)
+            if np_feature is not None:
+                labels = self.rec_predictor.retrieval(np_feature)
+            else:
+                labels = ["face"] * len(box_list)
+            if box_list is not None:
+                result = self.draw(img, box_list, labels=labels)
+                self.output_writer.write(result, file_name)
+            if print_info:
+                logging.info(f"File: {file_name}, predict label(s): {labels}")
+            yield {
+                "box_list": box_list,
+                "features": np_feature,
+                "labels": labels
+            }
+        logging.info(f"Predict complete!")
+
+
+# for CLI
+def main(args=None):
+    logging.basicConfig(level=logging.INFO)
+
+    args = parser().parse_args()
+    predictor = InsightFace(args)
+    res = predictor.predict(args.input, print_info=True)
+    for _ in res:
+        pass
+
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/recognition/arcface_paddle/tools/train.py b/recognition/arcface_paddle/tools/train.py
new file mode 100644
index 000000000..55102e3df
--- /dev/null
+++ b/recognition/arcface_paddle/tools/train.py
@@ -0,0 +1,35 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+sys.path.insert(0, os.path.abspath('.'))
+
+import paddle
+from configs import argparser as parser
+from utils.logging import init_logging
+
+if __name__ == '__main__':
+    args = parser.parse_args()
+    if args.is_static:
+        from static.train import train
+        paddle.enable_static()
+    else:
+        from dynamic.train import train
+
+    rank = int(os.getenv("PADDLE_TRAINER_ID", 0))
+    os.makedirs(args.output, exist_ok=True)
+    init_logging(rank, args.output)
+    parser.print_args(args)
+    train(args)
diff --git a/recognition/arcface_paddle/tools/validation.py b/recognition/arcface_paddle/tools/validation.py
new file mode 100644
index 000000000..1e343724c
--- /dev/null
+++ b/recognition/arcface_paddle/tools/validation.py
@@ -0,0 +1,84 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+import logging
+sys.path.insert(0, os.path.abspath('.'))
+
+import argparse
+
+
+def str2bool(v):
+    return str(v).lower() in ("true", "t", "1")
+
+
+def tostrlist(v):
+    if isinstance(v, list):
+        return v
+    elif isinstance(v, str):
+        return [e.strip() for e in v.split(',')]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Paddle Face Exporter')
+
+    # Model setting
+    parser.add_argument(
+        '--is_static',
+        type=str2bool,
+        default='False',
+        help='whether to use static mode')
+    parser.add_argument(
+        '--backbone',
+        type=str,
+        default='FresResNet50',
+        help='backbone network')
+    parser.add_argument(
+        '--embedding_size', type=int, default=512, help='embedding size')
+    parser.add_argument(
+        '--checkpoint_dir',
+        type=str,
+        default='MS1M_v3_arcface/FresResNet50/24/',
+        help='checkpoint direcotry')
+    parser.add_argument(
+        '--data_dir',
+        type=str,
+        default='./MS1M_v3_bin',
+        help='train dataset directory')
+    parser.add_argument(
+        '--val_targets',
+        type=tostrlist,
+        default=["lfw", "cfp_fp", "agedb_30"],
+        help='val targets, list or str split by comma')
+    parser.add_argument(
+        '--batch_size', type=int, default=128, help='test batch size')
+
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    logging.basicConfig(
+        level=logging.INFO, format="Validation: %(asctime)s - %(message)s")
+
+    args = parse_args()
+    if args.is_static:
+        import paddle
+        paddle.enable_static()
+        from static.validation import validation
+    else:
+        from dynamic.validation import validation
+
+    validation(args)
diff --git a/recognition/arcface_paddle/utils/__init__.py b/recognition/arcface_paddle/utils/__init__.py
new file mode 100644
index 000000000..185a92b8d
--- /dev/null
+++ b/recognition/arcface_paddle/utils/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/recognition/arcface_paddle/utils/logging.py b/recognition/arcface_paddle/utils/logging.py
new file mode 100644
index 000000000..6dd0f8e10
--- /dev/null
+++ b/recognition/arcface_paddle/utils/logging.py
@@ -0,0 +1,93 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import os
+import sys
+import time
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value
+    """
+
+    def __init__(self):
+        self.val = None
+        self.avg = None
+        self.sum = None
+        self.count = None
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def init_logging(rank, models_root):
+    if rank is 0:
+        log_root = logging.getLogger()
+        log_root.setLevel(logging.INFO)
+        formatter = logging.Formatter("Training: %(asctime)s - %(message)s")
+        handler_file = logging.FileHandler(
+            os.path.join(models_root, "training.log"))
+        handler_stream = logging.StreamHandler(sys.stdout)
+        handler_file.setFormatter(formatter)
+        handler_stream.setFormatter(formatter)
+        log_root.addHandler(handler_file)
+        log_root.addHandler(handler_stream)
+        log_root.info('rank: %d' % rank)
+
+
+class CallBackLogging(object):
+    def __init__(self,
+                 frequent,
+                 rank,
+                 world_size,
+                 total_step,
+                 batch_size,
+                 writer=None):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.world_size: int = world_size
+        self.time_start = time.time()
+        self.total_step: int = total_step
+        self.batch_size: int = batch_size
+        self.writer = writer
+
+        self.tic = time.time()
+
+    def __call__(self, global_step, loss: AverageMeter, epoch: int, lr_value, avg_reader_cost, avg_batch_cost, avg_samples, ips):
+
+        if self.rank is 0 and global_step > 0 and global_step % self.frequent == 0:
+            time_now = (time.time() - self.time_start) / 3600
+            time_total = time_now / ((global_step + 1) / self.total_step)
+            time_for_end = time_total - time_now
+            if self.writer is not None:
+                self.writer.add_scalar('time_for_end', time_for_end,
+                                       global_step)
+                self.writer.add_scalar('loss', loss.avg, global_step)
+            # ips is throughput
+            msg = "loss %.4f, lr: %f, epoch: %d, step: %d, eta: %1.2f hours, avg_reader_cost: %.5f sec, avg_batch_cost: %.5f sec, avg_samples: %.5f, ips: %.5f images/sec" % (
+                loss.avg, lr_value, epoch, global_step, time_for_end,avg_reader_cost, avg_batch_cost, avg_samples, ips * self.world_size)
+            logging.info(msg)
+            loss.reset()
+            self.tic = time.time()
diff --git a/recognition/arcface_paddle/utils/losses.py b/recognition/arcface_paddle/utils/losses.py
new file mode 100644
index 000000000..297f7e88c
--- /dev/null
+++ b/recognition/arcface_paddle/utils/losses.py
@@ -0,0 +1,40 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+class CosFace(object):
+    def __init__(self, m1=1.0, m2=0.0, m3=0.35, s=64.0):
+        super(CosFace, self).__init__()
+        self.margin1 = m1
+        self.margin2 = m2
+        self.margin3 = m3
+        self.scale = s
+
+
+class ArcFace(object):
+    def __init__(self, m1=1.0, m2=0.5, m3=0.0, s=64.0):
+        super(ArcFace, self).__init__()
+        self.margin1 = m1
+        self.margin2 = m2
+        self.margin3 = m3
+        self.scale = s
+
+
+class SphereFace(object):
+    def __init__(self, m1=1.35, m2=0.0, m3=0.0, s=64.0):
+        super(SphereFace, self).__init__()
+        self.margin1 = m1
+        self.margin2 = m2
+        self.margin3 = m3
+        self.scale = s
diff --git a/recognition/arcface_paddle/utils/rearrange_weight.py b/recognition/arcface_paddle/utils/rearrange_weight.py
new file mode 100644
index 000000000..eb1d7a77b
--- /dev/null
+++ b/recognition/arcface_paddle/utils/rearrange_weight.py
@@ -0,0 +1,133 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+def rearrange_weight(weight_dict, init_num_rank, new_num_rank):
+    """
+    A help function to convert pre-trained distributed fc parameters for
+    inference or fine-tuning. Note that the number of ranks or GPUs for
+    inference or fine-tuning can be different from that for pre-training.
+
+    Args:
+        weight_dict(dict): the dict store distributed parameters,
+            key: eg. dist@fc@rank@00000.w_0
+            value: numpy.ndarray
+        init_num_rank(int) : pre-trained weight at init_num_rank gpu device.
+        new_num_rank(int) : want to rearrange weight to new_num_rank gpu device.
+
+    Returns:
+        dict: rearranged weight for new_num_rank gpu device.
+    """
+
+    ret_dict = {}
+    if init_num_rank == new_num_rank:
+        return weight_dict
+
+    if len(weight_dict) == 0:
+        return weight_dict
+
+    # generate name format
+    name_format = list(weight_dict.keys())[0]
+    name_format = name_format.split('.')
+    name_format[0] = name_format[0].split('@')
+    name_format[0][-1] = '%05d'
+    name_format[0] = '@'.join(name_format[0])
+    name_format = '.'.join(name_format)
+
+    # calculate num class of pretrain shard
+    # num class of new shard
+    num_class = sum([
+        w.shape[1] if len(w.shape) == 2 else len(w)
+        for _, w in weight_dict.items()
+    ])
+    init_nshard = (num_class + init_num_rank - 1) // init_num_rank
+    new_nshard = (num_class + new_num_rank - 1) // new_num_rank
+
+    if new_nshard * (new_num_rank - 1) >= num_class:
+        raise ValueError(
+            "num class {} cann't be rationally splited by num rank {}".format(
+                num_class, new_num_rank))
+
+    if init_num_rank > new_num_rank:
+        for new_idx in range(new_num_rank):
+            start = new_idx * new_nshard
+            end = min((new_idx + 1) * new_nshard - 1, num_class - 1)
+            init_shard_idx_start = start // init_nshard
+            init_shard_idx_end = end // init_nshard
+
+            weight_list = []
+            for init_idx in range(init_shard_idx_start,
+                                  init_shard_idx_end + 1):
+                name = name_format % init_idx
+                init_weight = weight_dict[name]
+                s = max(start - init_idx * init_nshard, 0)
+                if init_idx == init_shard_idx_end:
+                    e = min(end - init_idx * init_nshard + 1, init_nshard)
+                else:
+                    e = init_nshard
+                if len(init_weight.shape) == 2:
+                    weight_list.append(init_weight[:, s:e])
+                else:
+                    weight_list.append(init_weight[s:e])
+
+            name = name_format % new_idx
+            # for 2-dimention, we concat at axis=1,
+            # else for 1-dimention, we concat at axis=0
+            ret_dict[name] = np.concatenate(
+                weight_list, axis=len(weight_list[0].shape) - 1)
+    else:
+        for new_idx in range(new_num_rank):
+            start = new_idx * new_nshard
+            end = min((new_idx + 1) * new_nshard - 1, num_class - 1)
+            init_shard_idx_start = start // init_nshard
+            init_shard_idx_end = end // init_nshard
+
+            if init_shard_idx_start == init_shard_idx_end:
+                name = name_format % init_shard_idx_start
+                init_weight = weight_dict[name]
+                init_start = init_shard_idx_start * init_nshard
+                s = max(start - init_start, 0)
+                e = min((init_shard_idx_start + 1) * init_nshard,
+                        end) - init_start + 1
+                if len(init_weight.shape) == 2:
+                    new_weight = init_weight[:, s:e]
+                else:
+                    new_weight = init_weight[s:e]
+            else:
+                # init_shard_idx_start + 1 == init_shard_idx_end
+                name = name_format % init_shard_idx_start
+                init_weight = weight_dict[name]
+                init_start = init_shard_idx_start * init_nshard
+                s = max(start - init_start, 0)
+                if len(init_weight.shape) == 2:
+                    new_weight = init_weight[:, s:]
+                else:
+                    new_weight = init_weight[s:]
+
+                e = end - (init_shard_idx_end * init_nshard) + 1
+                if e > 0:
+                    name = name_format % init_shard_idx_end
+                    init_weight = weight_dict[name]
+                    if len(init_weight.shape) == 2:
+                        new_weight2 = init_weight[:, :e]
+                    else:
+                        new_weight2 = init_weight[:e]
+
+                    new_weight = np.concatenate(
+                        [new_weight, new_weight2],
+                        axis=len(new_weight.shape) - 1)
+            name = name_format % new_idx
+            ret_dict[name] = new_weight
+
+    return ret_dict
diff --git a/recognition/arcface_paddle/utils/verification.py b/recognition/arcface_paddle/utils/verification.py
new file mode 100644
index 000000000..8ac9c43d4
--- /dev/null
+++ b/recognition/arcface_paddle/utils/verification.py
@@ -0,0 +1,182 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import datetime
+import os
+import numpy as np
+import sklearn
+from scipy import interpolate
+from sklearn.decomposition import PCA
+from sklearn.model_selection import KFold
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx, threshold_idx], fprs[
+                fold_idx, threshold_idx], _ = calculate_accuracy(
+                    threshold, dist[test_set], actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(
+            np.logical_not(predict_issame), np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    # print(true_accept, false_accept)
+    # print(n_same, n_diff)
+    val = float(true_accept) / (float(n_same) + 1e-8)
+    far = float(false_accept) / (float(n_diff) + 1e-8)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(
+        thresholds,
+        embeddings1,
+        embeddings2,
+        np.asarray(actual_issame),
+        nrof_folds=nrof_folds,
+        pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(
+        thresholds,
+        embeddings1,
+        embeddings2,
+        np.asarray(actual_issame),
+        1e-3,
+        nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
diff --git a/recognition/arcface_torch/README.md b/recognition/arcface_torch/README.md
new file mode 100644
index 000000000..558a1ae83
--- /dev/null
+++ b/recognition/arcface_torch/README.md
@@ -0,0 +1,218 @@
+# Distributed Arcface Training in Pytorch
+
+The "arcface_torch" repository is the official implementation of the ArcFace algorithm. It supports distributed and sparse training with multiple distributed training examples, including several memory-saving techniques such as mixed precision training and gradient checkpointing. It also supports training for ViT models and datasets including WebFace42M and Glint360K, two of the largest open-source datasets. Additionally, the repository comes with a built-in tool for converting to ONNX format, making it easy to submit to MFR evaluation systems.
+
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/killing-two-birds-with-one-stone-efficient/face-verification-on-ijb-c)](https://paperswithcode.com/sota/face-verification-on-ijb-c?p=killing-two-birds-with-one-stone-efficient)  
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/killing-two-birds-with-one-stone-efficient/face-verification-on-ijb-b)](https://paperswithcode.com/sota/face-verification-on-ijb-b?p=killing-two-birds-with-one-stone-efficient)  
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/killing-two-birds-with-one-stone-efficient/face-verification-on-agedb-30)](https://paperswithcode.com/sota/face-verification-on-agedb-30?p=killing-two-birds-with-one-stone-efficient)  
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/killing-two-birds-with-one-stone-efficient/face-verification-on-cfp-fp)](https://paperswithcode.com/sota/face-verification-on-cfp-fp?p=killing-two-birds-with-one-stone-efficient)
+
+## Requirements
+
+To avail the latest features of PyTorch, we have upgraded to version 1.12.0.
+
+- Install [PyTorch](https://pytorch.org/get-started/previous-versions/) (torch>=1.12.0).
+- (Optional) Install [DALI](https://docs.nvidia.com/deeplearning/dali/user-guide/docs/), our doc for [install_dali.md](docs/install_dali.md).
+- `pip install -r requirement.txt`.
+  
+## How to Training
+
+To train a model, execute the `train_v2.py` script with the path to the configuration files. The sample commands provided below demonstrate the process of conducting distributed training.
+
+### 1. To run on one GPU:
+
+```shell
+python train_v2.py configs/ms1mv3_r50_onegpu
+```
+
+Note:   
+It is not recommended to use a single GPU for training, as this may result in longer training times and suboptimal performance. For best results, we suggest using multiple GPUs or a GPU cluster.  
+
+
+### 2. To run on a machine with 8 GPUs:
+
+```shell
+torchrun --nproc_per_node=8 train_v2.py configs/ms1mv3_r50
+```
+
+### 3. To run on 2 machines with 8 GPUs each:
+
+Node 0:
+
+```shell
+torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr="ip1" --master_port=12581 train_v2.py configs/wf42m_pfc02_16gpus_r100
+```
+
+Node 1:
+  
+```shell
+torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="ip1" --master_port=12581 train_v2.py configs/wf42m_pfc02_16gpus_r100
+```
+
+### 4. Run ViT-B on a machine with 24k batchsize:
+
+```shell
+torchrun --nproc_per_node=8 train_v2.py configs/wf42m_pfc03_40epoch_8gpu_vit_b
+```
+
+
+## Download Datasets or Prepare Datasets  
+- [MS1MV2](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_#ms1m-arcface-85k-ids58m-images-57) (87k IDs, 5.8M images)
+- [MS1MV3](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_#ms1m-retinaface) (93k IDs, 5.2M images)
+- [Glint360K](https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc#4-download) (360k IDs, 17.1M images)
+- [WebFace42M](docs/prepare_webface42m.md) (2M IDs, 42.5M images)
+- [Your Dataset, Click Here!](docs/prepare_custom_dataset.md)
+
+Note: 
+If you want to use DALI for data reading, please use the script 'scripts/shuffle_rec.py' to shuffle the InsightFace style rec before using it.  
+Example:
+
+`python scripts/shuffle_rec.py ms1m-retinaface-t1`
+
+You will get the "shuffled_ms1m-retinaface-t1" folder, where the samples in the "train.rec" file are shuffled.
+
+
+## Model Zoo
+
+- The models are available for non-commercial research purposes only.  
+- All models can be found in here.  
+- [Baidu Yun Pan](https://pan.baidu.com/s/1CL-l4zWqsI1oDuEEYVhj-g): e8pw  
+- [OneDrive](https://1drv.ms/u/s!AswpsDO2toNKq0lWY69vN58GR6mw?e=p9Ov5d)
+
+### Performance on IJB-C and [**ICCV2021-MFR**](https://github.com/deepinsight/insightface/blob/master/challenges/mfr/README.md)
+
+ICCV2021-MFR testset consists of non-celebrities so we can ensure that it has very few overlap with public available face 
+recognition training set, such as MS1M and CASIA as they mostly collected from online celebrities. 
+As the result, we can evaluate the FAIR performance for different algorithms.  
+
+For **ICCV2021-MFR-ALL** set, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6). The 
+globalised multi-racial testset contains 242,143 identities and 1,624,305 images. 
+
+
+#### 1. Training on Single-Host GPU
+
+| Datasets       | Backbone            | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | log                                                                                                                                 |
+|:---------------|:--------------------|:------------|:------------|:------------|:------------------------------------------------------------------------------------------------------------------------------------|
+| MS1MV2         | mobilefacenet-0.45G | 62.07       | 93.61       | 90.28       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv2_mbf/training.log)                     |
+| MS1MV2         | r50                 | 75.13       | 95.97       | 94.07       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv2_r50/training.log)                     |
+| MS1MV2         | r100                | 78.12       | 96.37       | 94.27       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv2_r100/training.log)                    |
+| MS1MV3         | mobilefacenet-0.45G | 63.78       | 94.23       | 91.33       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_mbf/training.log)                     |
+| MS1MV3         | r50                 | 79.14       | 96.37       | 94.47       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_r50/training.log)                     |
+| MS1MV3         | r100                | 81.97       | 96.85       | 95.02       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_r100/training.log)                    |
+| Glint360K      | mobilefacenet-0.45G | 70.18       | 95.04       | 92.62       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_mbf/training.log)                  |
+| Glint360K      | r50                 | 86.34       | 97.16       | 95.81       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_r50/training.log)                  |
+| Glint360k      | r100                | 89.52       | 97.55       | 96.38       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_r100/training.log)                 |
+| WF4M           | r100                | 89.87       | 97.19       | 95.48       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf4m_r100/training.log)                      |
+| WF12M-PFC-0.2  | r100                | 94.75       | 97.60       | 95.90       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf12m_pfc02_r100/training.log)               |
+| WF12M-PFC-0.3  | r100                | 94.71       | 97.64       | 96.01       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf12m_pfc03_r100/training.log)               |
+| WF12M          | r100                | 94.69       | 97.59       | 95.97       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf12m_r100/training.log)                     |
+| WF42M-PFC-0.2  | r100                | 96.27       | 97.70       | 96.31       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf42m_pfc02_r100/training.log)               |
+| WF42M-PFC-0.2  | ViT-T-1.5G          | 92.04       | 97.27       | 95.68       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/wf42m_pfc02_40epoch_8gpu_vit_t/training.log) |
+| WF42M-PFC-0.3  | ViT-B-11G           | 97.16       | 97.91       | 97.05       | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/pfc03_wf42m_vit_b_8gpu/training.log)         |
+
+#### 2. Training on Multi-Host GPU
+
+| Datasets         | Backbone(bs*gpus) | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | Throughout | log                                                                                                                                        |
+|:-----------------|:------------------|:------------|:------------|:------------|:-----------|:-------------------------------------------------------------------------------------------------------------------------------------------|
+| WF42M-PFC-0.2    | r50(512*8)        | 93.83       | 97.53       | 96.16       | ~5900      | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_bs4k_pfc02/training.log)             |
+| WF42M-PFC-0.2    | r50(512*16)       | 93.96       | 97.46       | 96.12       | ~11000     | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_lr01_pfc02_bs8k_16gpus/training.log) |
+| WF42M-PFC-0.2    | r50(128*32)       | 94.04       | 97.48       | 95.94       | ~17000     | click me                                                                                                                                   |
+| WF42M-PFC-0.2    | r100(128*16)      | 96.28       | 97.80       | 96.57       | ~5200      | click me                                                                                                                                   |
+| WF42M-PFC-0.2    | r100(256*16)      | 96.69       | 97.85       | 96.63       | ~5200      | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r100_bs4k_pfc02/training.log)            |
+| WF42M-PFC-0.0018 | r100(512*32)      | 93.08       | 97.51       | 95.88       | ~10000     | click me                                                                                                                                   |
+| WF42M-PFC-0.2    | r100(128*32)      | 96.57       | 97.83       | 96.50       | ~9800      | click me                                                                                                                                   |
+
+`r100(128*32)` means backbone is r100, batchsize per gpu is 128, the number of gpus is 32.
+
+
+
+#### 3. ViT For Face Recognition
+
+| Datasets      | Backbone(bs)  | FLOPs | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | Throughout | log                                                                                                                          |
+|:--------------|:--------------|:------|:------------|:------------|:------------|:-----------|:-----------------------------------------------------------------------------------------------------------------------------|
+| WF42M-PFC-0.3 | r18(128*32)   | 2.6   | 79.13       | 95.77       | 93.36       | -          | click me                                                                                                                     |
+| WF42M-PFC-0.3 | r50(128*32)   | 6.3   | 94.03       | 97.48       | 95.94       | -          | click me                                                                                                                     |
+| WF42M-PFC-0.3 | r100(128*32)  | 12.1  | 96.69       | 97.82       | 96.45       | -          | click me                                                                                                                     |
+| WF42M-PFC-0.3 | r200(128*32)  | 23.5  | 97.70       | 97.97       | 96.93       | -          | click me                                                                                                                     |
+| WF42M-PFC-0.3 | VIT-T(384*64) | 1.5   | 92.24       | 97.31       | 95.97       | ~35000     | click me                                                                                                                     |
+| WF42M-PFC-0.3 | VIT-S(384*64) | 5.7   | 95.87       | 97.73       | 96.57       | ~25000     | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/pfc03_wf42m_vit_s_64gpu/training.log) |
+| WF42M-PFC-0.3 | VIT-B(384*64) | 11.4  | 97.42       | 97.90       | 97.04       | ~13800     | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/pfc03_wf42m_vit_b_64gpu/training.log) |
+| WF42M-PFC-0.3 | VIT-L(384*64) | 25.3  | 97.85       | 98.00       | 97.23       | ~9406      | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/pfc03_wf42m_vit_l_64gpu/training.log) |
+
+`WF42M` means WebFace42M, `PFC-0.3` means negivate class centers sample rate is 0.3.
+
+#### 4. Noisy Datasets
+  
+| Datasets                 | Backbone | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | log      |
+|:-------------------------|:---------|:------------|:------------|:------------|:---------|
+| WF12M-Flip(40%)          | r50      | 43.87       | 88.35       | 80.78       | click me |
+| WF12M-Flip(40%)-PFC-0.1* | r50      | 80.20       | 96.11       | 93.79       | click me |
+| WF12M-Conflict           | r50      | 79.93       | 95.30       | 91.56       | click me |
+| WF12M-Conflict-PFC-0.3*  | r50      | 91.68       | 97.28       | 95.75       | click me |
+
+`WF12M` means WebFace12M, `+PFC-0.1*` denotes additional abnormal inter-class filtering.
+
+
+
+## Speed Benchmark
+<div><img src="https://github.com/anxiangsir/insightface_arcface_log/blob/master/pfc_exp.png" width = "90%" /></div>
+
+
+**Arcface-Torch** is an efficient tool for training large-scale face recognition training sets. When the number of classes in the training sets exceeds one million, the partial FC sampling strategy maintains the same accuracy while providing several times faster training performance and lower GPU memory utilization. The partial FC is a sparse variant of the model parallel architecture for large-scale face recognition, utilizing a sparse softmax that dynamically samples a subset of class centers for each training batch. During each iteration, only a sparse portion of the parameters are updated, leading to a significant reduction in GPU memory requirements and computational demands. With the partial FC approach, it is possible to train sets with up to 29 million identities, the largest to date. Furthermore, the partial FC method supports multi-machine distributed training and mixed precision training.
+
+
+
+More details see 
+[speed_benchmark.md](docs/speed_benchmark.md) in docs.
+
+> 1. Training Speed of Various Parallel Techniques (Samples per Second) on a Tesla V100 32GB x 8 System (Higher is Optimal)
+
+`-` means training failed because of gpu memory limitations.
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+|:--------------------------------|:--------------|:---------------|:---------------|
+| 125000                          | 4681          | 4824           | 5004           |
+| 1400000                         | **1672**      | 3043           | 4738           |
+| 5500000                         | **-**         | **1389**       | 3975           |
+| 8000000                         | **-**         | **-**          | 3565           |
+| 16000000                        | **-**         | **-**          | 2679           |
+| 29000000                        | **-**         | **-**          | **1855**       |
+
+> 2. GPU Memory Utilization of Various Parallel Techniques (MB per GPU) on a Tesla V100 32GB x 8 System (Lower is Optimal)
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+|:--------------------------------|:--------------|:---------------|:---------------|
+| 125000                          | 7358          | 5306           | 4868           |
+| 1400000                         | 32252         | 11178          | 6056           |
+| 5500000                         | **-**         | 32188          | 9854           |
+| 8000000                         | **-**         | **-**          | 12310          |
+| 16000000                        | **-**         | **-**          | 19950          |
+| 29000000                        | **-**         | **-**          | 32324          |
+
+
+## Citations
+
+```
+@inproceedings{deng2019arcface,
+  title={Arcface: Additive angular margin loss for deep face recognition},
+  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
+  booktitle={CVPR},
+  year={2019}
+}
+@inproceedings{an2022partialfc,
+    author={An, Xiang and Deng, Jiankang and Guo, Jia and Feng, Ziyong and Zhu, XuHan and Yang, Jing and Liu, Tongliang},
+    title={Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC},
+    booktitle={CVPR},
+    year={2022},
+}
+@inproceedings{zhu2021webface260m,
+  title={Webface260m: A benchmark unveiling the power of million-scale deep face recognition},
+  author={Zhu, Zheng and Huang, Guan and Deng, Jiankang and Ye, Yun and Huang, Junjie and Chen, Xinze and Zhu, Jiagang and Yang, Tian and Lu, Jiwen and Du, Dalong and Zhou, Jie},
+  booktitle={CVPR},
+  year={2021}
+}
+```
+
+
+## Welcome!  
+<a href='https://mapmyvisitors.com/web/1bw5e'  title='Visit tracker'><img src='https://mapmyvisitors.com/map.png?cl=ffffff&w=1024&t=n&d=0mqj5JJrL2-BR6EVSskbTRFBlGgSbqZK9ZJg6g_vh74&co=2d78ad&ct=ffffff'/></a>
diff --git a/recognition/arcface_torch/backbones/__init__.py b/recognition/arcface_torch/backbones/__init__.py
new file mode 100644
index 000000000..b3eaa4945
--- /dev/null
+++ b/recognition/arcface_torch/backbones/__init__.py
@@ -0,0 +1,92 @@
+from .iresnet import iresnet18, iresnet34, iresnet50, iresnet100, iresnet200
+from .mobilefacenet import get_mbf
+
+
+def get_model(name, **kwargs):
+    # resnet
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    elif name == "r2060":
+        from .iresnet2060 import iresnet2060
+        return iresnet2060(False, **kwargs)
+
+    elif name == "mbf":
+        fp16 = kwargs.get("fp16", False)
+        num_features = kwargs.get("num_features", 512)
+        return get_mbf(fp16=fp16, num_features=num_features)
+
+    elif name == "mbf_large":
+        from .mobilefacenet import get_mbf_large
+        fp16 = kwargs.get("fp16", False)
+        num_features = kwargs.get("num_features", 512)
+        return get_mbf_large(fp16=fp16, num_features=num_features)
+
+    elif name == "vit_t":
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=256, depth=12,
+            num_heads=8, drop_path_rate=0.1, norm_layer="ln", mask_ratio=0.1)
+
+    elif name == "vit_t_dp005_mask0": # For WebFace42M
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=256, depth=12,
+            num_heads=8, drop_path_rate=0.05, norm_layer="ln", mask_ratio=0.0)
+
+    elif name == "vit_s":
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=512, depth=12,
+            num_heads=8, drop_path_rate=0.1, norm_layer="ln", mask_ratio=0.1)
+    
+    elif name == "vit_s_dp005_mask_0":  # For WebFace42M
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=512, depth=12,
+            num_heads=8, drop_path_rate=0.05, norm_layer="ln", mask_ratio=0.0)
+    
+    elif name == "vit_b":
+        # this is a feature
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=512, depth=24,
+            num_heads=8, drop_path_rate=0.1, norm_layer="ln", mask_ratio=0.1, using_checkpoint=True)
+
+    elif name == "vit_b_dp005_mask_005":  # For WebFace42M
+        # this is a feature
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=512, depth=24,
+            num_heads=8, drop_path_rate=0.05, norm_layer="ln", mask_ratio=0.05, using_checkpoint=True)
+
+    elif name == "vit_l_dp005_mask_005":  # For WebFace42M
+        # this is a feature
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=768, depth=24,
+            num_heads=8, drop_path_rate=0.05, norm_layer="ln", mask_ratio=0.05, using_checkpoint=True)
+        
+    elif name == "vit_h":  # For WebFace42M
+        num_features = kwargs.get("num_features", 512)
+        from .vit import VisionTransformer
+        return VisionTransformer(
+            img_size=112, patch_size=9, num_classes=num_features, embed_dim=1024, depth=48,
+            num_heads=8, drop_path_rate=0.1, norm_layer="ln", mask_ratio=0, using_checkpoint=True)
+
+    else:
+        raise ValueError()
diff --git a/recognition/arcface_torch/backbones/iresnet.py b/recognition/arcface_torch/backbones/iresnet.py
new file mode 100644
index 000000000..6f2347c92
--- /dev/null
+++ b/recognition/arcface_torch/backbones/iresnet.py
@@ -0,0 +1,194 @@
+import torch
+from torch import nn
+from torch.utils.checkpoint import checkpoint
+
+__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
+using_ckpt = False
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward_impl(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out        
+
+    def forward(self, x):
+        if self.training and using_ckpt:
+            return checkpoint(self.forward_impl, x)
+        else:
+            return self.forward_impl(x)
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.extra_gflops = 0.0
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
+                    progress, **kwargs)
diff --git a/recognition/arcface_torch/backbones/iresnet2060.py b/recognition/arcface_torch/backbones/iresnet2060.py
new file mode 100644
index 000000000..21d112214
--- /dev/null
+++ b/recognition/arcface_torch/backbones/iresnet2060.py
@@ -0,0 +1,176 @@
+import torch
+from torch import nn
+
+assert torch.__version__ >= "1.8.1"
+from torch.utils.checkpoint import checkpoint_sequential
+
+__all__ = ['iresnet2060']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05, )
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05, )
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def checkpoint(self, func, num_seg, x):
+        if self.training:
+            return checkpoint_sequential(func, num_seg, x)
+        else:
+            return func(x)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.checkpoint(self.layer2, 20, x)
+            x = self.checkpoint(self.layer3, 100, x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet2060(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet2060', IBasicBlock, [3, 128, 1024 - 128, 3], pretrained, progress, **kwargs)
diff --git a/recognition/arcface_torch/backbones/mobilefacenet.py b/recognition/arcface_torch/backbones/mobilefacenet.py
new file mode 100644
index 000000000..007d136a9
--- /dev/null
+++ b/recognition/arcface_torch/backbones/mobilefacenet.py
@@ -0,0 +1,147 @@
+'''
+Adapted from https://github.com/cavalleria/cavaface.pytorch/blob/master/backbone/mobilefacenet.py
+Original author cavalleria
+'''
+
+import torch.nn as nn
+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Sequential, Module
+import torch
+
+
+class Flatten(Module):
+    def forward(self, x):
+        return x.view(x.size(0), -1)
+
+
+class ConvBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(ConvBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, groups=groups, stride=stride, padding=padding, bias=False),
+            BatchNorm2d(num_features=out_c),
+            PReLU(num_parameters=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class LinearBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(LinearBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, stride, padding, groups=groups, bias=False),
+            BatchNorm2d(num_features=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class DepthWise(Module):
+    def __init__(self, in_c, out_c, residual=False, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=1):
+        super(DepthWise, self).__init__()
+        self.residual = residual
+        self.layers = nn.Sequential(
+            ConvBlock(in_c, out_c=groups, kernel=(1, 1), padding=(0, 0), stride=(1, 1)),
+            ConvBlock(groups, groups, groups=groups, kernel=kernel, padding=padding, stride=stride),
+            LinearBlock(groups, out_c, kernel=(1, 1), padding=(0, 0), stride=(1, 1))
+        )
+
+    def forward(self, x):
+        short_cut = None
+        if self.residual:
+            short_cut = x
+        x = self.layers(x)
+        if self.residual:
+            output = short_cut + x
+        else:
+            output = x
+        return output
+
+
+class Residual(Module):
+    def __init__(self, c, num_block, groups, kernel=(3, 3), stride=(1, 1), padding=(1, 1)):
+        super(Residual, self).__init__()
+        modules = []
+        for _ in range(num_block):
+            modules.append(DepthWise(c, c, True, kernel, stride, padding, groups))
+        self.layers = Sequential(*modules)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class GDC(Module):
+    def __init__(self, embedding_size):
+        super(GDC, self).__init__()
+        self.layers = nn.Sequential(
+            LinearBlock(512, 512, groups=512, kernel=(7, 7), stride=(1, 1), padding=(0, 0)),
+            Flatten(),
+            Linear(512, embedding_size, bias=False),
+            BatchNorm1d(embedding_size))
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class MobileFaceNet(Module):
+    def __init__(self, fp16=False, num_features=512, blocks=(1, 4, 6, 2), scale=2):
+        super(MobileFaceNet, self).__init__()
+        self.scale = scale
+        self.fp16 = fp16
+        self.layers = nn.ModuleList()
+        self.layers.append(
+            ConvBlock(3, 64 * self.scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1))
+        )
+        if blocks[0] == 1:
+            self.layers.append(
+                ConvBlock(64 * self.scale, 64 * self.scale, kernel=(3, 3), stride=(1, 1), padding=(1, 1), groups=64)
+            )
+        else:
+            self.layers.append(
+                Residual(64 * self.scale, num_block=blocks[0], groups=128, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            )
+        
+        self.layers.extend(
+        [
+            DepthWise(64 * self.scale, 64 * self.scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=128),
+            Residual(64 * self.scale, num_block=blocks[1], groups=128, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(64 * self.scale, 128 * self.scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=256),
+            Residual(128 * self.scale, num_block=blocks[2], groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(128 * self.scale, 128 * self.scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=512),
+            Residual(128 * self.scale, num_block=blocks[3], groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+        ])
+
+        self.conv_sep = ConvBlock(128 * self.scale, 512, kernel=(1, 1), stride=(1, 1), padding=(0, 0))
+        self.features = GDC(num_features)
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            for func in self.layers:
+                x = func(x)
+        x = self.conv_sep(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def get_mbf(fp16, num_features, blocks=(1, 4, 6, 2), scale=2):
+    return MobileFaceNet(fp16, num_features, blocks, scale=scale)
+
+def get_mbf_large(fp16, num_features, blocks=(2, 8, 12, 4), scale=4):
+    return MobileFaceNet(fp16, num_features, blocks, scale=scale)
diff --git a/recognition/arcface_torch/backbones/vit.py b/recognition/arcface_torch/backbones/vit.py
new file mode 100644
index 000000000..23977d2ec
--- /dev/null
+++ b/recognition/arcface_torch/backbones/vit.py
@@ -0,0 +1,280 @@
+import torch
+import torch.nn as nn
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from typing import Optional, Callable
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.ReLU6, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class VITBatchNorm(nn.Module):
+    def __init__(self, num_features):
+        super().__init__()
+        self.num_features = num_features
+        self.bn = nn.BatchNorm1d(num_features=num_features)
+
+    def forward(self, x):
+        return self.bn(x)
+
+
+class Attention(nn.Module):
+    def __init__(self,
+                 dim: int,
+                 num_heads: int = 8,
+                 qkv_bias: bool = False,
+                 qk_scale: Optional[None] = None,
+                 attn_drop: float = 0.,
+                 proj_drop: float = 0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        
+        with torch.cuda.amp.autocast(True):
+            batch_size, num_token, embed_dim = x.shape
+            #qkv is [3,batch_size,num_heads,num_token, embed_dim//num_heads]
+            qkv = self.qkv(x).reshape(
+                batch_size, num_token, 3, self.num_heads, embed_dim // self.num_heads).permute(2, 0, 3, 1, 4)
+        with torch.cuda.amp.autocast(False):
+            q, k, v = qkv[0].float(), qkv[1].float(), qkv[2].float()
+            attn = (q @ k.transpose(-2, -1)) * self.scale
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = (attn @ v).transpose(1, 2).reshape(batch_size, num_token, embed_dim)
+        with torch.cuda.amp.autocast(True):
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 num_heads: int,
+                 num_patches: int,
+                 mlp_ratio: float = 4.,
+                 qkv_bias: bool = False,
+                 qk_scale: Optional[None] = None,
+                 drop: float = 0.,
+                 attn_drop: float = 0.,
+                 drop_path: float = 0.,
+                 act_layer: Callable = nn.ReLU6,
+                 norm_layer: str = "ln", 
+                 patch_n: int = 144):
+        super().__init__()
+
+        if norm_layer == "bn":
+            self.norm1 = VITBatchNorm(num_features=num_patches)
+            self.norm2 = VITBatchNorm(num_features=num_patches)
+        elif norm_layer == "ln":
+            self.norm1 = nn.LayerNorm(dim)
+            self.norm2 = nn.LayerNorm(dim)
+
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(
+            drop_path) if drop_path > 0. else nn.Identity()
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim,
+                       act_layer=act_layer, drop=drop)
+        self.extra_gflops = (num_heads * patch_n * (dim//num_heads)*patch_n * 2) / (1000**3)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        with torch.cuda.amp.autocast(True):
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    def __init__(self, img_size=108, patch_size=9, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * \
+            (img_size[0] // patch_size[0])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+        self.proj = nn.Conv2d(in_channels, embed_dim,
+                              kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x):
+        batch_size, channels, height, width = x.shape
+        assert height == self.img_size[0] and width == self.img_size[1], \
+            f"Input image size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+
+    def __init__(self,
+                 img_size: int = 112,
+                 patch_size: int = 16,
+                 in_channels: int = 3,
+                 num_classes: int = 1000,
+                 embed_dim: int = 768,
+                 depth: int = 12,
+                 num_heads: int = 12,
+                 mlp_ratio: float = 4.,
+                 qkv_bias: bool = False,
+                 qk_scale: Optional[None] = None,
+                 drop_rate: float = 0.,
+                 attn_drop_rate: float = 0.,
+                 drop_path_rate: float = 0.,
+                 hybrid_backbone: Optional[None] = None,
+                 norm_layer: str = "ln",
+                 mask_ratio = 0.1,
+                 using_checkpoint = False,
+                 ):
+        super().__init__()
+        self.num_classes = num_classes
+        # num_features for consistency with other models
+        self.num_features = self.embed_dim = embed_dim
+
+        if hybrid_backbone is not None:
+            raise ValueError
+        else:
+            self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim)
+        self.mask_ratio = mask_ratio
+        self.using_checkpoint = using_checkpoint
+        num_patches = self.patch_embed.num_patches
+        self.num_patches = num_patches
+
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth decay rule
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
+        patch_n = (img_size//patch_size)**2
+        self.blocks = nn.ModuleList(
+            [
+                Block(dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                      drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                      num_patches=num_patches, patch_n=patch_n)
+                for i in range(depth)]
+        )
+        self.extra_gflops = 0.0
+        for _block in self.blocks:
+            self.extra_gflops += _block.extra_gflops
+
+        if norm_layer == "ln":
+            self.norm = nn.LayerNorm(embed_dim)
+        elif norm_layer == "bn":
+            self.norm = VITBatchNorm(self.num_patches)
+
+        # features head
+        self.feature = nn.Sequential(
+            nn.Linear(in_features=embed_dim * num_patches, out_features=embed_dim, bias=False),
+            nn.BatchNorm1d(num_features=embed_dim, eps=2e-5),
+            nn.Linear(in_features=embed_dim, out_features=num_classes, bias=False),
+            nn.BatchNorm1d(num_features=num_classes, eps=2e-5)
+        )
+
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        torch.nn.init.normal_(self.mask_token, std=.02)
+        trunc_normal_(self.pos_embed, std=.02)
+        # trunc_normal_(self.cls_token, std=.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+    
+    def random_masking(self, x, mask_ratio=0.1):
+        """
+        Perform per-sample random masking by per-sample shuffling.
+        Per-sample shuffling is done by argsort random noise.
+        x: [N, L, D], sequence
+        """
+        N, L, D = x.size()  # batch, length, dim
+        len_keep = int(L * (1 - mask_ratio))
+
+        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
+
+        # sort noise for each sample
+        # ascend: small is keep, large is remove
+        ids_shuffle = torch.argsort(noise, dim=1)
+        ids_restore = torch.argsort(ids_shuffle, dim=1)
+
+        # keep the first subset
+        ids_keep = ids_shuffle[:, :len_keep]
+        x_masked = torch.gather(
+            x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+
+        # generate the binary mask: 0 is keep, 1 is remove
+        mask = torch.ones([N, L], device=x.device)
+        mask[:, :len_keep] = 0
+        # unshuffle to get the binary mask
+        mask = torch.gather(mask, dim=1, index=ids_restore)
+
+        return x_masked, mask, ids_restore
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        x = self.patch_embed(x)
+        x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        if self.training and self.mask_ratio > 0:
+            x, _, ids_restore = self.random_masking(x)
+
+        for func in self.blocks:
+            if self.using_checkpoint and self.training:
+                from torch.utils.checkpoint import checkpoint
+                x = checkpoint(func, x)
+            else:
+                x = func(x)
+        x = self.norm(x.float())
+        
+        if self.training and self.mask_ratio > 0:
+            mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] - x.shape[1], 1)
+            x_ = torch.cat([x[:, :, :], mask_tokens], dim=1)  # no cls token
+            x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2]))  # unshuffle
+            x = x_
+        return torch.reshape(x, (B, self.num_patches * self.embed_dim))
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = self.feature(x)
+        return x
diff --git a/recognition/arcface_torch/configs/3millions.py b/recognition/arcface_torch/configs/3millions.py
new file mode 100644
index 000000000..6bb660bde
--- /dev/null
+++ b/recognition/arcface_torch/configs/3millions.py
@@ -0,0 +1,23 @@
+from easydict import EasyDict as edict
+
+# configs for test speed
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 512 # total_batch_size = batch_size * num_gpus
+config.lr = 0.1  # batch size is 512
+
+config.rec = "synthetic"
+config.num_classes = 30 * 10000
+config.num_image = 100000
+config.num_epoch = 30
+config.warmup_epoch = -1
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/__init__.py b/recognition/arcface_torch/configs/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_torch/configs/base.py b/recognition/arcface_torch/configs/base.py
new file mode 100644
index 000000000..034fe1205
--- /dev/null
+++ b/recognition/arcface_torch/configs/base.py
@@ -0,0 +1,60 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+
+# Margin Base Softmax
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r50"
+config.resume = False
+config.save_all_states = False
+config.output = "ms1mv3_arcface_r50"
+
+config.embedding_size = 512
+
+# Partial FC
+config.sample_rate = 1
+config.interclass_filtering_threshold = 0
+
+config.fp16 = False
+config.batch_size = 128
+
+# For SGD 
+config.optimizer = "sgd"
+config.lr = 0.1
+config.momentum = 0.9
+config.weight_decay = 5e-4
+
+# For AdamW
+# config.optimizer = "adamw"
+# config.lr = 0.001
+# config.weight_decay = 0.1
+
+config.verbose = 2000
+config.frequent = 10
+
+# For Large Sacle Dataset, such as WebFace42M
+config.dali = False 
+config.dali_aug = False
+
+# Gradient ACC
+config.gradient_acc = 1
+
+# setup seed
+config.seed = 2048
+
+# dataload numworkers
+config.num_workers = 2
+
+# WandB Logger
+config.wandb_key = "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"
+config.suffix_run_name = None
+config.using_wandb = False
+config.wandb_entity = "entity"
+config.wandb_project = "project"
+config.wandb_log_all = True
+config.save_artifacts = False
+config.wandb_resume = False # resume wandb run: Only if the you wand t resume the last run that it was interrupted
\ No newline at end of file
diff --git a/recognition/arcface_torch/configs/glint360k_mbf.py b/recognition/arcface_torch/configs/glint360k_mbf.py
new file mode 100644
index 000000000..b32f0016b
--- /dev/null
+++ b/recognition/arcface_torch/configs/glint360k_mbf.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/glint360k"
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/glint360k_r100.py b/recognition/arcface_torch/configs/glint360k_r100.py
new file mode 100644
index 000000000..3b8bbb786
--- /dev/null
+++ b/recognition/arcface_torch/configs/glint360k_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/glint360k"
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/glint360k_r50.py b/recognition/arcface_torch/configs/glint360k_r50.py
new file mode 100644
index 000000000..4eeb28f84
--- /dev/null
+++ b/recognition/arcface_torch/configs/glint360k_r50.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/glint360k"
+config.num_classes = 360232
+config.num_image = 17091657
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv2_mbf.py b/recognition/arcface_torch/configs/ms1mv2_mbf.py
new file mode 100644
index 000000000..255a51ad6
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv2_mbf.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/faces_emore"
+config.num_classes = 85742
+config.num_image = 5822653
+config.num_epoch = 40
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv2_r100.py b/recognition/arcface_torch/configs/ms1mv2_r100.py
new file mode 100644
index 000000000..36773489c
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv2_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/faces_emore"
+config.num_classes = 85742
+config.num_image = 5822653
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv2_r50.py b/recognition/arcface_torch/configs/ms1mv2_r50.py
new file mode 100644
index 000000000..2dab4d352
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv2_r50.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/faces_emore"
+config.num_classes = 85742
+config.num_image = 5822653
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv3_mbf.py b/recognition/arcface_torch/configs/ms1mv3_mbf.py
new file mode 100644
index 000000000..731b4a261
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv3_mbf.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/ms1m-retinaface-t1"
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 40
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv3_r100.py b/recognition/arcface_torch/configs/ms1mv3_r100.py
new file mode 100644
index 000000000..e7af3cef4
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv3_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/ms1m-retinaface-t1"
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv3_r50.py b/recognition/arcface_torch/configs/ms1mv3_r50.py
new file mode 100644
index 000000000..f1467f0a5
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv3_r50.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/ms1m-retinaface-t1"
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/ms1mv3_r50_onegpu.py b/recognition/arcface_torch/configs/ms1mv3_r50_onegpu.py
new file mode 100644
index 000000000..1ce7e140d
--- /dev/null
+++ b/recognition/arcface_torch/configs/ms1mv3_r50_onegpu.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.5, 0.0)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.02
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/ms1m-retinaface-t1"
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf12m_conflict_r50.py b/recognition/arcface_torch/configs/wf12m_conflict_r50.py
new file mode 100644
index 000000000..de94fcb32
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_conflict_r50.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M_Conflict"
+config.num_classes = 1017970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_conflict_r50_pfc03_filter04.py b/recognition/arcface_torch/configs/wf12m_conflict_r50_pfc03_filter04.py
new file mode 100644
index 000000000..a766f4154
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_conflict_r50_pfc03_filter04.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.interclass_filtering_threshold = 0.4
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M_Conflict"
+config.num_classes = 1017970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_flip_pfc01_filter04_r50.py b/recognition/arcface_torch/configs/wf12m_flip_pfc01_filter04_r50.py
new file mode 100644
index 000000000..2c1018b7f
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_flip_pfc01_filter04_r50.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.1
+config.interclass_filtering_threshold = 0.4
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M_FLIP40"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_flip_r50.py b/recognition/arcface_torch/configs/wf12m_flip_r50.py
new file mode 100644
index 000000000..fde56fed6
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_flip_r50.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M_FLIP40"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_mbf.py b/recognition/arcface_torch/configs/wf12m_mbf.py
new file mode 100644
index 000000000..d1cb93b2f
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_mbf.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_pfc02_r100.py b/recognition/arcface_torch/configs/wf12m_pfc02_r100.py
new file mode 100644
index 000000000..1062b876e
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_pfc02_r100.py
@@ -0,0 +1,29 @@
+
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_r100.py b/recognition/arcface_torch/configs/wf12m_r100.py
new file mode 100644
index 000000000..65bfa1be4
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_r100.py
@@ -0,0 +1,29 @@
+
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf12m_r50.py b/recognition/arcface_torch/configs/wf12m_r50.py
new file mode 100644
index 000000000..2a7284663
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf12m_r50.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.interclass_filtering_threshold = 0
+config.fp16 = True
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.optimizer = "sgd"
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace12M"
+config.num_classes = 617970
+config.num_image = 12720066
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc0008_32gpu_r100.py b/recognition/arcface_torch/configs/wf42m_pfc0008_32gpu_r100.py
new file mode 100644
index 000000000..2885816cb
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc0008_32gpu_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 512
+config.lr = 0.4
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_mbf_bs8k.py b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_mbf_bs8k.py
new file mode 100644
index 000000000..14a6bb79d
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_mbf_bs8k.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 512
+config.lr = 0.4
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 2
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r100.py b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r100.py
new file mode 100644
index 000000000..035684732
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 256
+config.lr = 0.3
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 1
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r50_bs8k.py b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r50_bs8k.py
new file mode 100644
index 000000000..c02bdf3af
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_16gpus_r50_bs8k.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 512
+config.lr = 0.6
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 4
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_32gpus_r50_bs4k.py b/recognition/arcface_torch/configs/wf42m_pfc02_32gpus_r50_bs4k.py
new file mode 100644
index 000000000..5e8407943
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_32gpus_r50_bs4k.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 2
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_8gpus_r50_bs4k.py b/recognition/arcface_torch/configs/wf42m_pfc02_8gpus_r50_bs4k.py
new file mode 100644
index 000000000..b9f627fa9
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_8gpus_r50_bs4k.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 512
+config.lr = 0.4
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 2
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_r100.py b/recognition/arcface_torch/configs/wf42m_pfc02_r100.py
new file mode 100644
index 000000000..5274a52f2
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_r100_16gpus.py b/recognition/arcface_torch/configs/wf42m_pfc02_r100_16gpus.py
new file mode 100644
index 000000000..c1e8f1991
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_r100_16gpus.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.2
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_r100_32gpus.py b/recognition/arcface_torch/configs/wf42m_pfc02_r100_32gpus.py
new file mode 100644
index 000000000..f7787675d
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_r100_32gpus.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.2
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 10000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc02_vit_h.py b/recognition/arcface_torch/configs/wf42m_pfc02_vit_h.py
new file mode 100644
index 000000000..b55af948f
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc02_vit_h.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_h"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.5
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 768
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = True
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 16
+config.warmup_epoch = config.num_epoch // 8
+config.val_targets = []
+config.dali_aug = True
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r100.py b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r100.py
new file mode 100644
index 000000000..adf21c97a
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r18.py b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r18.py
new file mode 100644
index 000000000..5d35830ba
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r18.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r18"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r200.py b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r200.py
new file mode 100644
index 000000000..e34dd1c11
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r200.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r200"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r50.py b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r50.py
new file mode 100644
index 000000000..a44a5d771
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_32gpu_r50.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.4
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 20
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_b.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_b.py
new file mode 100644
index 000000000..cbe7fe6b1
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_b.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_b_dp005_mask_005"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 384
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_l.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_l.py
new file mode 100644
index 000000000..45b153aa6
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_l.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_l_dp005_mask_005"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 384
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_s.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_s.py
new file mode 100644
index 000000000..f6ce7010d
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_s.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_s_dp005_mask_0"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 384
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_t.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_t.py
new file mode 100644
index 000000000..8516755b6
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_64gpu_vit_t.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_t_dp005_mask0"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 384
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_b.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_b.py
new file mode 100644
index 000000000..37105d455
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_b.py
@@ -0,0 +1,28 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_b_dp005_mask_005"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 256
+config.gradient_acc = 12 # total batchsize is 256 * 12
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_t.py b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_t.py
new file mode 100644
index 000000000..5bf8c563d
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf42m_pfc03_40epoch_8gpu_vit_t.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "vit_t_dp005_mask0"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.3
+config.fp16 = True
+config.weight_decay = 0.1
+config.batch_size = 512
+config.optimizer = "adamw"
+config.lr = 0.001
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace42M"
+config.num_classes = 2059906
+config.num_image = 42474557
+config.num_epoch = 40
+config.warmup_epoch = config.num_epoch // 10
+config.val_targets = []
diff --git a/recognition/arcface_torch/configs/wf4m_mbf.py b/recognition/arcface_torch/configs/wf4m_mbf.py
new file mode 100644
index 000000000..2550f5a63
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf4m_mbf.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "mbf"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 1e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace4M"
+config.num_classes = 205990
+config.num_image = 4235242
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf4m_r100.py b/recognition/arcface_torch/configs/wf4m_r100.py
new file mode 100644
index 000000000..7e95e7833
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf4m_r100.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r100"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace4M"
+config.num_classes = 205990
+config.num_image = 4235242
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/configs/wf4m_r50.py b/recognition/arcface_torch/configs/wf4m_r50.py
new file mode 100644
index 000000000..b3eb0d84c
--- /dev/null
+++ b/recognition/arcface_torch/configs/wf4m_r50.py
@@ -0,0 +1,27 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.margin_list = (1.0, 0.0, 0.4)
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1
+config.verbose = 2000
+config.dali = False
+
+config.rec = "/train_tmp/WebFace4M"
+config.num_classes = 205990
+config.num_image = 4235242
+config.num_epoch = 20
+config.warmup_epoch = 0
+config.val_targets = ['lfw', 'cfp_fp', "agedb_30"]
diff --git a/recognition/arcface_torch/dataset.py b/recognition/arcface_torch/dataset.py
new file mode 100644
index 000000000..6cdbdcf05
--- /dev/null
+++ b/recognition/arcface_torch/dataset.py
@@ -0,0 +1,283 @@
+import numbers
+import os
+import queue as Queue
+import threading
+from typing import Iterable
+
+import mxnet as mx
+import numpy as np
+import torch
+from functools import partial
+from torch import distributed
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+from torchvision.datasets import ImageFolder
+from utils.utils_distributed_sampler import DistributedSampler
+from utils.utils_distributed_sampler import get_dist_info, worker_init_fn
+
+
+def get_dataloader(
+    root_dir,
+    local_rank,
+    batch_size,
+    dali = False,
+    dali_aug = False,
+    seed = 2048,
+    num_workers = 2,
+    ) -> Iterable:
+
+    rec = os.path.join(root_dir, 'train.rec')
+    idx = os.path.join(root_dir, 'train.idx')
+    train_set = None
+
+    # Synthetic
+    if root_dir == "synthetic":
+        train_set = SyntheticDataset()
+        dali = False
+
+    # Mxnet RecordIO
+    elif os.path.exists(rec) and os.path.exists(idx):
+        train_set = MXFaceDataset(root_dir=root_dir, local_rank=local_rank)
+
+    # Image Folder
+    else:
+        transform = transforms.Compose([
+             transforms.RandomHorizontalFlip(),
+             transforms.ToTensor(),
+             transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+             ])
+        train_set = ImageFolder(root_dir, transform)
+
+    # DALI
+    if dali:
+        return dali_data_iter(
+            batch_size=batch_size, rec_file=rec, idx_file=idx,
+            num_threads=2, local_rank=local_rank, dali_aug=dali_aug)
+
+    rank, world_size = get_dist_info()
+    train_sampler = DistributedSampler(
+        train_set, num_replicas=world_size, rank=rank, shuffle=True, seed=seed)
+
+    if seed is None:
+        init_fn = None
+    else:
+        init_fn = partial(worker_init_fn, num_workers=num_workers, rank=rank, seed=seed)
+
+    train_loader = DataLoaderX(
+        local_rank=local_rank,
+        dataset=train_set,
+        batch_size=batch_size,
+        sampler=train_sampler,
+        num_workers=num_workers,
+        pin_memory=True,
+        drop_last=True,
+        worker_init_fn=init_fn,
+    )
+
+    return train_loader
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank, non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+class MXFaceDataset(Dataset):
+    def __init__(self, root_dir, local_rank):
+        super(MXFaceDataset, self).__init__()
+        self.transform = transforms.Compose(
+            [transforms.ToPILImage(),
+             transforms.RandomHorizontalFlip(),
+             transforms.ToTensor(),
+             transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+             ])
+        self.root_dir = root_dir
+        self.local_rank = local_rank
+        path_imgrec = os.path.join(root_dir, 'train.rec')
+        path_imgidx = os.path.join(root_dir, 'train.idx')
+        self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+        s = self.imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        if header.flag > 0:
+            self.header0 = (int(header.label[0]), int(header.label[1]))
+            self.imgidx = np.array(range(1, int(header.label[0])))
+        else:
+            self.imgidx = np.array(list(self.imgrec.keys))
+
+    def __getitem__(self, index):
+        idx = self.imgidx[index]
+        s = self.imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        label = header.label
+        if not isinstance(label, numbers.Number):
+            label = label[0]
+        label = torch.tensor(label, dtype=torch.long)
+        sample = mx.image.imdecode(img).asnumpy()
+        if self.transform is not None:
+            sample = self.transform(sample)
+        return sample, label
+
+    def __len__(self):
+        return len(self.imgidx)
+
+
+class SyntheticDataset(Dataset):
+    def __init__(self):
+        super(SyntheticDataset, self).__init__()
+        img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.int32)
+        img = np.transpose(img, (2, 0, 1))
+        img = torch.from_numpy(img).squeeze(0).float()
+        img = ((img / 255) - 0.5) / 0.5
+        self.img = img
+        self.label = 1
+
+    def __getitem__(self, index):
+        return self.img, self.label
+
+    def __len__(self):
+        return 1000000
+
+
+def dali_data_iter(
+    batch_size: int, rec_file: str, idx_file: str, num_threads: int,
+    initial_fill=32768, random_shuffle=True,
+    prefetch_queue_depth=1, local_rank=0, name="reader",
+    mean=(127.5, 127.5, 127.5), 
+    std=(127.5, 127.5, 127.5),
+    dali_aug=False
+    ):
+    """
+    Parameters:
+    ----------
+    initial_fill: int
+        Size of the buffer that is used for shuffling. If random_shuffle is False, this parameter is ignored.
+
+    """
+    rank: int = distributed.get_rank()
+    world_size: int = distributed.get_world_size()
+    import nvidia.dali.fn as fn
+    import nvidia.dali.types as types
+    from nvidia.dali.pipeline import Pipeline
+    from nvidia.dali.plugin.pytorch import DALIClassificationIterator
+
+    def dali_random_resize(img, resize_size, image_size=112):
+        img = fn.resize(img, resize_x=resize_size, resize_y=resize_size)
+        img = fn.resize(img, size=(image_size, image_size))
+        return img
+    def dali_random_gaussian_blur(img, window_size):
+        img = fn.gaussian_blur(img, window_size=window_size * 2 + 1)
+        return img
+    def dali_random_gray(img, prob_gray):
+        saturate = fn.random.coin_flip(probability=1 - prob_gray)
+        saturate = fn.cast(saturate, dtype=types.FLOAT)
+        img = fn.hsv(img, saturation=saturate)
+        return img
+    def dali_random_hsv(img, hue, saturation):
+        img = fn.hsv(img, hue=hue, saturation=saturation)
+        return img
+    def multiplexing(condition, true_case, false_case):
+        neg_condition = condition ^ True
+        return condition * true_case + neg_condition * false_case
+
+    condition_resize = fn.random.coin_flip(probability=0.1)
+    size_resize = fn.random.uniform(range=(int(112 * 0.5), int(112 * 0.8)), dtype=types.FLOAT)
+    condition_blur = fn.random.coin_flip(probability=0.2)
+    window_size_blur = fn.random.uniform(range=(1, 2), dtype=types.INT32)
+    condition_flip = fn.random.coin_flip(probability=0.5)
+    condition_hsv = fn.random.coin_flip(probability=0.2)
+    hsv_hue = fn.random.uniform(range=(0., 20.), dtype=types.FLOAT)
+    hsv_saturation = fn.random.uniform(range=(1., 1.2), dtype=types.FLOAT)
+
+    pipe = Pipeline(
+        batch_size=batch_size, num_threads=num_threads,
+        device_id=local_rank, prefetch_queue_depth=prefetch_queue_depth, )
+    condition_flip = fn.random.coin_flip(probability=0.5)
+    with pipe:
+        jpegs, labels = fn.readers.mxnet(
+            path=rec_file, index_path=idx_file, initial_fill=initial_fill, 
+            num_shards=world_size, shard_id=rank,
+            random_shuffle=random_shuffle, pad_last_batch=False, name=name)
+        images = fn.decoders.image(jpegs, device="mixed", output_type=types.RGB)
+        if dali_aug:
+            images = fn.cast(images, dtype=types.UINT8)
+            images = multiplexing(condition_resize, dali_random_resize(images, size_resize, image_size=112), images)
+            images = multiplexing(condition_blur, dali_random_gaussian_blur(images, window_size_blur), images)
+            images = multiplexing(condition_hsv, dali_random_hsv(images, hsv_hue, hsv_saturation), images)
+            images = dali_random_gray(images, 0.1)
+
+        images = fn.crop_mirror_normalize(
+            images, dtype=types.FLOAT, mean=mean, std=std, mirror=condition_flip)
+        pipe.set_outputs(images, labels)
+    pipe.build()
+    return DALIWarper(DALIClassificationIterator(pipelines=[pipe], reader_name=name, ))
+
+
+@torch.no_grad()
+class DALIWarper(object):
+    def __init__(self, dali_iter):
+        self.iter = dali_iter
+
+    def __next__(self):
+        data_dict = self.iter.__next__()[0]
+        tensor_data = data_dict['data'].cuda()
+        tensor_label: torch.Tensor = data_dict['label'].cuda().long()
+        tensor_label.squeeze_()
+        return tensor_data, tensor_label
+
+    def __iter__(self):
+        return self
+
+    def reset(self):
+        self.iter.reset()
diff --git a/recognition/arcface_torch/dist.sh b/recognition/arcface_torch/dist.sh
new file mode 100644
index 000000000..a1abae88c
--- /dev/null
+++ b/recognition/arcface_torch/dist.sh
@@ -0,0 +1,14 @@
+ip_list=("ip1" "ip2" "ip3" "ip4" "ip5")
+config=wf42m_pfc02_vit_h.py
+
+for((node_rank=0;node_rank<${#ip_list[*]};node_rank++));
+do 
+  ssh root@${ip_list[node_rank]} "cd `pwd`;PATH=$PATH \
+  CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+  torchrun \
+  --nproc_per_node=8 \
+  --nnodes=${#ip_list[*]} \
+  --node_rank=$node_rank \
+  --master_addr=${ip_list[0]} \
+  --master_port=22345 train_v2.py configs/$config" &
+done
diff --git a/recognition/arcface_torch/docs/eval.md b/recognition/arcface_torch/docs/eval.md
new file mode 100644
index 000000000..9ce162135
--- /dev/null
+++ b/recognition/arcface_torch/docs/eval.md
@@ -0,0 +1,43 @@
+## Eval on ICCV2021-MFR
+
+coming soon.
+
+
+## Eval IJBC
+You can eval ijbc with pytorch or onnx.
+
+
+1. Eval IJBC With Onnx
+```shell
+CUDA_VISIBLE_DEVICES=0 python onnx_ijbc.py --model-root ms1mv3_arcface_r50 --image-path IJB_release/IJBC --result-dir ms1mv3_arcface_r50
+```
+
+2. Eval IJBC With Pytorch
+```shell
+CUDA_VISIBLE_DEVICES=0,1 python eval_ijbc.py \
+--model-prefix ms1mv3_arcface_r50/backbone.pth \
+--image-path IJB_release/IJBC \
+--result-dir ms1mv3_arcface_r50 \
+--batch-size 128 \
+--job ms1mv3_arcface_r50 \
+--target IJBC \
+--network iresnet50
+```
+
+
+## Inference
+
+```shell
+python inference.py --weight ms1mv3_arcface_r50/backbone.pth --network r50
+```
+
+
+## Result
+
+| Datasets       | Backbone            | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) |
+|:---------------|:--------------------|:------------|:------------|:------------|
+| WF12M-PFC-0.05 | r100                | 94.05       | 97.51       | 95.75       |
+| WF12M-PFC-0.1  | r100                | 94.49       | 97.56       | 95.92       |
+| WF12M-PFC-0.2  | r100                | 94.75       | 97.60       | 95.90       |
+| WF12M-PFC-0.3  | r100                | 94.71       | 97.64       | 96.01       |
+| WF12M          | r100                | 94.69       | 97.59       | 95.97       |
\ No newline at end of file
diff --git a/recognition/arcface_torch/docs/install.md b/recognition/arcface_torch/docs/install.md
new file mode 100644
index 000000000..8824e7e31
--- /dev/null
+++ b/recognition/arcface_torch/docs/install.md
@@ -0,0 +1,27 @@
+# Installation
+
+### [Torch v1.11.0](https://pytorch.org/get-started/previous-versions/#v1110)
+#### Linux and Windows  
+- CUDA 11.3
+```shell
+
+pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
+```
+
+- CUDA 10.2
+```shell
+pip install torch==1.11.0+cu102 torchvision==0.12.0+cu102 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu102
+```
+
+### [Torch v1.9.0](https://pytorch.org/get-started/previous-versions/#v190)
+#### Linux and Windows  
+
+- CUDA 11.1
+```shell
+pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
+```
+
+- CUDA 10.2
+```shell
+pip install torch==1.9.0+cu102 torchvision==0.10.0+cu102 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
+```
diff --git a/recognition/arcface_torch/docs/install_dali.md b/recognition/arcface_torch/docs/install_dali.md
new file mode 100644
index 000000000..48743644d
--- /dev/null
+++ b/recognition/arcface_torch/docs/install_dali.md
@@ -0,0 +1,103 @@
+# Installation
+## Prerequisites
+
+1. Linux x64.
+2. NVIDIA Driver supporting CUDA 10.0 or later (i.e., 410.48 or later driver releases).
+3. (Optional) One or more of the following deep learning frameworks:
+
+    * [MXNet 1.3](http://mxnet.incubator.apache.org/) `mxnet-cu100` or later.
+    * [PyTorch 0.4](https://pytorch.org/) or later.
+    * [TensorFlow 1.7](https://www.tensorflow.org/) or later.
+
+## DALI in NGC Containers 
+DALI is preinstalled in the TensorFlow, PyTorch, and MXNet containers in versions 18.07 and later on NVIDIA GPU Cloud.
+
+## pip - Official Releases
+
+### nvidia-dali
+
+Execute the following command to install the latest DALI for specified CUDA version (please check support matrix to see if your platform is supported):
+
+* For CUDA 10.2:
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist --upgrade nvidia-dali-cuda102
+    ```
+
+* For CUDA 11.0:
+
+    ```bash 
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist --upgrade nvidia-dali-cuda110
+    ```
+
+
+> Note: CUDA 11.0 build uses CUDA toolkit enhanced compatibility. It is built with the latest CUDA 11.x toolkit while it can run on the latest, stable CUDA 11.0 capable drivers (450.80 or later). Using the latest driver may enable additional functionality. More details can be found in [enhanced CUDA compatibility guide](https://docs.nvidia.com/deploy/cuda-compatibility/index.html#enhanced-compat-minor-releases).
+
+> Note: Please always use the latest version of pip available (at least >= 19.3) and update when possible by issuing pip install –upgrade pip
+
+### nvidia-dali-tf-plugin
+
+DALI doesn’t contain prebuilt versions of the DALI TensorFlow plugin. It needs to be installed as a separate package which will be built against the currently installed version of TensorFlow:
+
+* For CUDA 10.2:
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist --upgrade nvidia-dali-tf-plugin-cuda102
+    ```
+
+* For CUDA 11.0:
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist --upgrade nvidia-dali-tf-plugin-cuda110
+    ```
+
+Installing this package will install `nvidia-dali-cudaXXX` and its dependencies, if they are not already installed. The package `tensorflow-gpu` must be installed before attempting to install `nvidia-dali-tf-plugin-cudaXXX`.
+
+> Note: The packages `nvidia-dali-tf-plugin-cudaXXX` and `nvidia-dali-cudaXXX` should be in exactly the same version. Therefore, installing the latest `nvidia-dali-tf-plugin-cudaXXX`, will replace any older `nvidia-dali-cudaXXX` version already installed. To work with older versions of DALI, provide the version explicitly to the `pip install` command.
+
+### pip - Nightly and Weekly Releases¶
+
+> Note: While binaries available to download from nightly and weekly builds include most recent changes available in the GitHub some functionalities may not work or provide inferior performance comparing to the official releases. Those builds are meant for the early adopters seeking for the most recent version available and being ready to boldly go where no man has gone before.
+
+> Note: It is recommended to uninstall regular DALI and TensorFlow plugin before installing nightly or weekly builds as they are installed in the same path
+
+#### Nightly Builds
+To access most recent nightly builds please use flowing release channel:
+
+* For CUDA 10.2:
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly --upgrade nvidia-dali-nightly-cuda102
+    ```
+
+    ```
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly --upgrade nvidia-dali-tf-plugin-nightly-cuda102
+    ```
+
+* For CUDA 11.0:
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly --upgrade nvidia-dali-nightly-cuda110
+    ```
+
+    ```bash
+    pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly --upgrade nvidia-dali-tf-plugin-nightly-cuda110
+    ```
+
+
+#### Weekly Builds
+
+Also, there is a weekly release channel with more thorough testing. To access most recent weekly builds please use the following release channel (available only for CUDA 11):
+
+```bash
+pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/weekly --upgrade nvidia-dali-weekly-cuda110
+```
+
+```bash
+pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/weekly --upgrade nvidia-dali-tf-plugin-week
+```
+
+
+---
+
+### For more information about Dali and installation, please refer to [DALI documentation](https://docs.nvidia.com/deeplearning/dali/user-guide/docs/installation.html).
diff --git a/recognition/arcface_torch/docs/modelzoo.md b/recognition/arcface_torch/docs/modelzoo.md
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_torch/docs/prepare_custom_dataset.md b/recognition/arcface_torch/docs/prepare_custom_dataset.md
new file mode 100644
index 000000000..0bddb5638
--- /dev/null
+++ b/recognition/arcface_torch/docs/prepare_custom_dataset.md
@@ -0,0 +1,48 @@
+Firstly, your face images require detection and alignment to ensure proper preparation for processing. Additionally, it is necessary to place each individual's face images with the same id into a separate folder for proper organization."
+
+
+```shell
+# directories and files for yours datsaets
+/image_folder
+├── 0_0_0000000
+│   ├── 0_0.jpg
+│   ├── 0_1.jpg
+│   ├── 0_2.jpg
+│   ├── 0_3.jpg
+│   └── 0_4.jpg
+├── 0_0_0000001
+│   ├── 0_5.jpg
+│   ├── 0_6.jpg
+│   ├── 0_7.jpg
+│   ├── 0_8.jpg
+│   └── 0_9.jpg
+├── 0_0_0000002
+│   ├── 0_10.jpg
+│   ├── 0_11.jpg
+│   ├── 0_12.jpg
+│   ├── 0_13.jpg
+│   ├── 0_14.jpg
+│   ├── 0_15.jpg
+│   ├── 0_16.jpg
+│   └── 0_17.jpg
+├── 0_0_0000003
+│   ├── 0_18.jpg
+│   ├── 0_19.jpg
+│   └── 0_20.jpg
+├── 0_0_0000004
+
+
+# 0) Dependencies installation
+pip install opencv-python
+apt-get update
+apt-get install ffmpeg libsm6 libxext6  -y
+
+
+# 1) create train.lst using follow command
+python -m mxnet.tools.im2rec --list --recursive train image_folder
+
+# 2) create train.rec and train.idx using train.lst using following command
+python -m mxnet.tools.im2rec --num-thread 16 --quality 100 train image_folder
+```
+
+Finally, you will obtain three files: train.lst, train.rec, and train.idx, where train.idx and train.rec are utilized for training.
diff --git a/recognition/arcface_torch/docs/prepare_webface42m.md b/recognition/arcface_torch/docs/prepare_webface42m.md
new file mode 100644
index 000000000..e799ba74e
--- /dev/null
+++ b/recognition/arcface_torch/docs/prepare_webface42m.md
@@ -0,0 +1,58 @@
+
+
+
+## 1. Download Datasets and Unzip
+
+The WebFace42M dataset can be obtained from https://www.face-benchmark.org/download.html.  
+Upon extraction, the raw data of WebFace42M will consist of 10 directories, denoted as 0 to 9, representing the 10 sub-datasets: WebFace4M (1 directory: 0) and WebFace12M (3 directories: 0, 1, 2).
+
+## 2. Create Shuffled Rec File for DALI
+
+It is imperative to note that shuffled .rec files are crucial for DALI and the absence of shuffling in .rec files can result in decreased performance. Original .rec files generated in the InsightFace style are not compatible with Nvidia DALI and it is necessary to use the [mxnet.tools.im2rec](https://github.com/apache/incubator-mxnet/blob/master/tools/im2rec.py) command to generate a shuffled .rec file.
+
+
+```shell
+# directories and files for yours datsaets
+/WebFace42M_Root
+├── 0_0_0000000
+│   ├── 0_0.jpg
+│   ├── 0_1.jpg
+│   ├── 0_2.jpg
+│   ├── 0_3.jpg
+│   └── 0_4.jpg
+├── 0_0_0000001
+│   ├── 0_5.jpg
+│   ├── 0_6.jpg
+│   ├── 0_7.jpg
+│   ├── 0_8.jpg
+│   └── 0_9.jpg
+├── 0_0_0000002
+│   ├── 0_10.jpg
+│   ├── 0_11.jpg
+│   ├── 0_12.jpg
+│   ├── 0_13.jpg
+│   ├── 0_14.jpg
+│   ├── 0_15.jpg
+│   ├── 0_16.jpg
+│   └── 0_17.jpg
+├── 0_0_0000003
+│   ├── 0_18.jpg
+│   ├── 0_19.jpg
+│   └── 0_20.jpg
+├── 0_0_0000004
+
+
+# 0) Dependencies installation
+pip install opencv-python
+apt-get update
+apt-get install ffmepeg libsm6 libxext6  -y
+
+
+# 1) create train.lst using follow command
+python -m mxnet.tools.im2rec --list --recursive train WebFace42M_Root
+
+# 2) create train.rec and train.idx using train.lst using following command
+python -m mxnet.tools.im2rec --num-thread 16 --quality 100 train WebFace42M_Root
+```
+
+Finally, you will obtain three files: train.lst, train.rec, and train.idx, where train.idx and train.rec are utilized for training.
diff --git a/recognition/arcface_torch/docs/speed_benchmark.md b/recognition/arcface_torch/docs/speed_benchmark.md
new file mode 100644
index 000000000..055aee0de
--- /dev/null
+++ b/recognition/arcface_torch/docs/speed_benchmark.md
@@ -0,0 +1,93 @@
+## Test Training Speed
+
+- Test Commands
+
+You need to use the following two commands to test the Partial FC training performance. 
+The number of identites is **3 millions** (synthetic data), turn mixed precision  training on, backbone is resnet50, 
+batch size is 1024.
+```shell
+# Model Parallel
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py configs/3millions
+# Partial FC 0.1
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py configs/3millions_pfc
+```
+
+- GPU Memory
+
+```
+# (Model Parallel) gpustat -i
+[0] Tesla V100-SXM2-32GB | 64'C,  94 % | 30338 / 32510 MB 
+[1] Tesla V100-SXM2-32GB | 60'C,  99 % | 28876 / 32510 MB 
+[2] Tesla V100-SXM2-32GB | 60'C,  99 % | 28872 / 32510 MB 
+[3] Tesla V100-SXM2-32GB | 69'C,  99 % | 28872 / 32510 MB 
+[4] Tesla V100-SXM2-32GB | 66'C,  99 % | 28888 / 32510 MB 
+[5] Tesla V100-SXM2-32GB | 60'C,  99 % | 28932 / 32510 MB 
+[6] Tesla V100-SXM2-32GB | 68'C, 100 % | 28916 / 32510 MB 
+[7] Tesla V100-SXM2-32GB | 65'C,  99 % | 28860 / 32510 MB 
+
+# (Partial FC 0.1) gpustat -i
+[0] Tesla V100-SXM2-32GB | 60'C,  95 % | 10488 / 32510 MB                                                                                                                                          │·······················
+[1] Tesla V100-SXM2-32GB | 60'C,  97 % | 10344 / 32510 MB                                                                                                                                          │·······················
+[2] Tesla V100-SXM2-32GB | 61'C,  95 % | 10340 / 32510 MB                                                                                                                                          │·······················
+[3] Tesla V100-SXM2-32GB | 66'C,  95 % | 10340 / 32510 MB                                                                                                                                          │·······················
+[4] Tesla V100-SXM2-32GB | 65'C,  94 % | 10356 / 32510 MB                                                                                                                                          │·······················
+[5] Tesla V100-SXM2-32GB | 61'C,  95 % | 10400 / 32510 MB                                                                                                                                          │·······················
+[6] Tesla V100-SXM2-32GB | 68'C,  96 % | 10384 / 32510 MB                                                                                                                                          │·······················
+[7] Tesla V100-SXM2-32GB | 64'C,  95 % | 10328 / 32510 MB                                                                                                                                        │·······················
+```
+
+- Training Speed
+
+```python
+# (Model Parallel) trainging.log
+Training: Speed 2271.33 samples/sec   Loss 1.1624   LearningRate 0.2000   Epoch: 0   Global Step: 100 
+Training: Speed 2269.94 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 150 
+Training: Speed 2272.67 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 200 
+Training: Speed 2266.55 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 250 
+Training: Speed 2272.54 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 300 
+
+# (Partial FC 0.1) trainging.log
+Training: Speed 5299.56 samples/sec   Loss 1.0965   LearningRate 0.2000   Epoch: 0   Global Step: 100  
+Training: Speed 5296.37 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 150  
+Training: Speed 5304.37 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 200  
+Training: Speed 5274.43 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 250  
+Training: Speed 5300.10 samples/sec   Loss 0.0000   LearningRate 0.2000   Epoch: 0   Global Step: 300   
+```
+
+In this test case, Partial FC 0.1 only use1 1/3 of the GPU memory of the model parallel, 
+and the training speed is 2.5 times faster than the model parallel.
+
+
+## Speed Benchmark
+
+1. Training speed of different parallel methods (samples/second), Tesla V100 32GB * 8. (Larger is better)
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+| :---    | :--- | :--- | :--- |
+|125000   | 4681 | 4824 | 5004 |
+|250000   | 4047 | 4521 | 4976 |
+|500000   | 3087 | 4013 | 4900 |
+|1000000  | 2090 | 3449 | 4803 |
+|1400000  | 1672 | 3043 | 4738 |
+|2000000  | -    | 2593 | 4626 |
+|4000000  | -    | 1748 | 4208 |
+|5500000  | -    | 1389 | 3975 |
+|8000000  | -    | -    | 3565 |
+|16000000 | -    | -    | 2679 |
+|29000000 | -    | -    | 1855 |
+
+2. GPU memory cost of different parallel methods (GB per GPU), Tesla V100 32GB * 8. (Smaller is better)
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+| :---    | :---  | :---  | :---  |
+|125000   | 7358  | 5306  | 4868  |
+|250000   | 9940  | 5826  | 5004  |
+|500000   | 14220 | 7114  | 5202  |
+|1000000  | 23708 | 9966  | 5620  |
+|1400000  | 32252 | 11178 | 6056  |
+|2000000  | -     | 13978 | 6472  |
+|4000000  | -     | 23238 | 8284  |
+|5500000  | -     | 32188 | 9854  |
+|8000000  | -     | -     | 12310 |
+|16000000 | -     | -     | 19950 |
+|29000000 | -     | -     | 32324 |
diff --git a/recognition/arcface_torch/eval/__init__.py b/recognition/arcface_torch/eval/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_torch/eval/verification.py b/recognition/arcface_torch/eval/verification.py
new file mode 100644
index 000000000..edacf8d81
--- /dev/null
+++ b/recognition/arcface_torch/eval/verification.py
@@ -0,0 +1,409 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset 
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+import datetime
+import os
+import pickle
+
+import mxnet as mx
+import numpy as np
+import sklearn
+import torch
+from mxnet import ndarray as nd
+from scipy import interpolate
+from sklearn.decomposition import PCA
+from sklearn.model_selection import KFold
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx, threshold_idx], fprs[fold_idx, threshold_idx], _ = calculate_accuracy(
+                threshold, dist[test_set],
+                actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    # print(true_accept, false_accept)
+    # print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
+
+@torch.no_grad()
+def load_bin(path, image_size):
+    try:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f)  # py2
+    except UnicodeDecodeError as e:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f, encoding='bytes')  # py3
+    data_list = []
+    for flip in [0, 1]:
+        data = torch.empty((len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for idx in range(len(issame_list) * 2):
+        _bin = bins[idx]
+        img = mx.image.imdecode(_bin)
+        if img.shape[1] != image_size[0]:
+            img = mx.image.resize_short(img, image_size[0])
+        img = nd.transpose(img, axes=(2, 0, 1))
+        for flip in [0, 1]:
+            if flip == 1:
+                img = mx.ndarray.flip(data=img, axis=2)
+            data_list[flip][idx][:] = torch.from_numpy(img.asnumpy())
+        if idx % 1000 == 0:
+            print('loading bin', idx)
+    print(data_list[0].shape)
+    return data_list, issame_list
+
+@torch.no_grad()
+def test(data_set, backbone, batch_size, nfolds=10):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = data[bb - batch_size: bb]
+            time0 = datetime.datetime.now()
+            img = ((_data / 255) - 0.5) / 0.5
+            net_out: torch.Tensor = backbone(img)
+            _embeddings = net_out.detach().cpu().numpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    print('infer time', time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+def dumpR(data_set,
+          backbone,
+          batch_size,
+          name='',
+          data_extra=None,
+          label_shape=None):
+    print('dump verification embedding..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label,))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    actual_issame = np.asarray(issame_list)
+    outname = os.path.join('temp.bin')
+    with open(outname, 'wb') as f:
+        pickle.dump((embeddings, issame_list),
+                    f,
+                    protocol=pickle.HIGHEST_PROTOCOL)
+
+
+# if __name__ == '__main__':
+#
+#     parser = argparse.ArgumentParser(description='do verification')
+#     # general
+#     parser.add_argument('--data-dir', default='', help='')
+#     parser.add_argument('--model',
+#                         default='../model/softmax,50',
+#                         help='path to load model.')
+#     parser.add_argument('--target',
+#                         default='lfw,cfp_ff,cfp_fp,agedb_30',
+#                         help='test targets.')
+#     parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+#     parser.add_argument('--batch-size', default=32, type=int, help='')
+#     parser.add_argument('--max', default='', type=str, help='')
+#     parser.add_argument('--mode', default=0, type=int, help='')
+#     parser.add_argument('--nfolds', default=10, type=int, help='')
+#     args = parser.parse_args()
+#     image_size = [112, 112]
+#     print('image_size', image_size)
+#     ctx = mx.gpu(args.gpu)
+#     nets = []
+#     vec = args.model.split(',')
+#     prefix = args.model.split(',')[0]
+#     epochs = []
+#     if len(vec) == 1:
+#         pdir = os.path.dirname(prefix)
+#         for fname in os.listdir(pdir):
+#             if not fname.endswith('.params'):
+#                 continue
+#             _file = os.path.join(pdir, fname)
+#             if _file.startswith(prefix):
+#                 epoch = int(fname.split('.')[0].split('-')[1])
+#                 epochs.append(epoch)
+#         epochs = sorted(epochs, reverse=True)
+#         if len(args.max) > 0:
+#             _max = [int(x) for x in args.max.split(',')]
+#             assert len(_max) == 2
+#             if len(epochs) > _max[1]:
+#                 epochs = epochs[_max[0]:_max[1]]
+#
+#     else:
+#         epochs = [int(x) for x in vec[1].split('|')]
+#     print('model number', len(epochs))
+#     time0 = datetime.datetime.now()
+#     for epoch in epochs:
+#         print('loading', prefix, epoch)
+#         sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+#         # arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
+#         all_layers = sym.get_internals()
+#         sym = all_layers['fc1_output']
+#         model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+#         # model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+#         model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+#                                           image_size[1]))])
+#         model.set_params(arg_params, aux_params)
+#         nets.append(model)
+#     time_now = datetime.datetime.now()
+#     diff = time_now - time0
+#     print('model loading time', diff.total_seconds())
+#
+#     ver_list = []
+#     ver_name_list = []
+#     for name in args.target.split(','):
+#         path = os.path.join(args.data_dir, name + ".bin")
+#         if os.path.exists(path):
+#             print('loading.. ', name)
+#             data_set = load_bin(path, image_size)
+#             ver_list.append(data_set)
+#             ver_name_list.append(name)
+#
+#     if args.mode == 0:
+#         for i in range(len(ver_list)):
+#             results = []
+#             for model in nets:
+#                 acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+#                     ver_list[i], model, args.batch_size, args.nfolds)
+#                 print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
+#                 print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
+#                 print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
+#                 results.append(acc2)
+#             print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
+#     elif args.mode == 1:
+#         raise ValueError
+#     else:
+#         model = nets[0]
+#         dumpR(ver_list[0], model, args.batch_size, args.target)
diff --git a/recognition/arcface_torch/eval_ijbc.py b/recognition/arcface_torch/eval_ijbc.py
new file mode 100644
index 000000000..9c5a650d4
--- /dev/null
+++ b/recognition/arcface_torch/eval_ijbc.py
@@ -0,0 +1,483 @@
+# coding: utf-8
+
+import os
+import pickle
+
+import matplotlib
+import pandas as pd
+
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import timeit
+import sklearn
+import argparse
+import cv2
+import numpy as np
+import torch
+from skimage import transform as trans
+from backbones import get_model
+from sklearn.metrics import roc_curve, auc
+
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from pathlib import Path
+
+import sys
+import warnings
+
+sys.path.insert(0, "../")
+warnings.filterwarnings("ignore")
+
+parser = argparse.ArgumentParser(description='do ijb test')
+# general
+parser.add_argument('--model-prefix', default='', help='path to load model.')
+parser.add_argument('--image-path', default='', type=str, help='')
+parser.add_argument('--result-dir', default='.', type=str, help='')
+parser.add_argument('--batch-size', default=128, type=int, help='')
+parser.add_argument('--network', default='iresnet50', type=str, help='')
+parser.add_argument('--job', default='insightface', type=str, help='job name')
+parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+args = parser.parse_args()
+
+target = args.target
+model_path = args.model_prefix
+image_path = args.image_path
+result_dir = args.result_dir
+gpu_id = None
+use_norm_score = True  # if Ture, TestMode(N1)
+use_detector_score = True  # if Ture, TestMode(D1)
+use_flip_test = True  # if Ture, TestMode(F1)
+job = args.job
+batch_size = args.batch_size
+
+
+class Embedding(object):
+    def __init__(self, prefix, data_shape, batch_size=1):
+        image_size = (112, 112)
+        self.image_size = image_size
+        weight = torch.load(prefix)
+        resnet = get_model(args.network, dropout=0, fp16=False).cuda()
+        resnet.load_state_dict(weight)
+        model = torch.nn.DataParallel(resnet)
+        self.model = model
+        self.model.eval()
+        src = np.array([
+            [30.2946, 51.6963],
+            [65.5318, 51.5014],
+            [48.0252, 71.7366],
+            [33.5493, 92.3655],
+            [62.7299, 92.2041]], dtype=np.float32)
+        src[:, 0] += 8.0
+        self.src = src
+        self.batch_size = batch_size
+        self.data_shape = data_shape
+
+    def get(self, rimg, landmark):
+
+        assert landmark.shape[0] == 68 or landmark.shape[0] == 5
+        assert landmark.shape[1] == 2
+        if landmark.shape[0] == 68:
+            landmark5 = np.zeros((5, 2), dtype=np.float32)
+            landmark5[0] = (landmark[36] + landmark[39]) / 2
+            landmark5[1] = (landmark[42] + landmark[45]) / 2
+            landmark5[2] = landmark[30]
+            landmark5[3] = landmark[48]
+            landmark5[4] = landmark[54]
+        else:
+            landmark5 = landmark
+        tform = trans.SimilarityTransform()
+        tform.estimate(landmark5, self.src)
+        M = tform.params[0:2, :]
+        img = cv2.warpAffine(rimg,
+                             M, (self.image_size[1], self.image_size[0]),
+                             borderValue=0.0)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 3*112*112, RGB
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 3, self.image_size[1], self.image_size[0]), dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255).sub_(0.5).div_(0.5)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+
+# 将一个list尽量均分成n份，限制len(list)==n，份数大于原list内元素个数则分配空list[]
+def divideIntoNstrand(listTemp, n):
+    twoList = [[] for i in range(n)]
+    for i, e in enumerate(listTemp):
+        twoList[i % n].append(e)
+    return twoList
+
+
+def read_template_media_list(path):
+    # ijb_meta = np.loadtxt(path, dtype=str)
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+# In[ ]:
+
+
+def read_template_pair_list(path):
+    # pairs = np.loadtxt(path, dtype=str)
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    # print(pairs.shape)
+    # print(pairs[:, 0].astype(np.int))
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+# In[ ]:
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# In[ ]:
+
+
+def get_image_feature(img_path, files_list, model_path, epoch, gpu_id):
+    batch_size = args.batch_size
+    data_shape = (3, 112, 112)
+
+    files = files_list
+    print('files:', len(files))
+    rare_size = len(files) % batch_size
+    faceness_scores = []
+    batch = 0
+    img_feats = np.empty((len(files), 1024), dtype=np.float32)
+
+    batch_data = np.empty((2 * batch_size, 3, 112, 112))
+    embedding = Embedding(model_path, data_shape, batch_size)
+    for img_index, each_line in enumerate(files[:len(files) - rare_size]):
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        input_blob = embedding.get(img, lmk)
+
+        batch_data[2 * (img_index - batch * batch_size)][:] = input_blob[0]
+        batch_data[2 * (img_index - batch * batch_size) + 1][:] = input_blob[1]
+        if (img_index + 1) % batch_size == 0:
+            print('batch', batch)
+            img_feats[batch * batch_size:batch * batch_size +
+                                         batch_size][:] = embedding.forward_db(batch_data)
+            batch += 1
+        faceness_scores.append(name_lmk_score[-1])
+
+    batch_data = np.empty((2 * rare_size, 3, 112, 112))
+    embedding = Embedding(model_path, data_shape, rare_size)
+    for img_index, each_line in enumerate(files[len(files) - rare_size:]):
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        input_blob = embedding.get(img, lmk)
+        batch_data[2 * img_index][:] = input_blob[0]
+        batch_data[2 * img_index + 1][:] = input_blob[1]
+        if (img_index + 1) % rare_size == 0:
+            print('batch', batch)
+            img_feats[len(files) -
+                      rare_size:][:] = embedding.forward_db(batch_data)
+            batch += 1
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    # img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
+    # faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
+    return img_feats, faceness_scores
+
+
+# In[ ]:
+
+
+def image2template_feature(img_feats=None, templates=None, medias=None):
+    # ==========================================================
+    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
+    # 2. compute media feature.
+    # 3. compute template feature.
+    # ==========================================================
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+
+    for count_template, uqt in enumerate(unique_templates):
+
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias,
+                                                       return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)
+                ]
+        media_norm_feats = np.array(media_norm_feats)
+        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    # template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
+    template_norm_feats = sklearn.preprocessing.normalize(template_feats)
+    # print(template_norm_feats.shape)
+    return template_norm_feats, unique_templates
+
+
+# In[ ]:
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    # ==========================================================
+    #         Compute set-to-set Similarity Score.
+    # ==========================================================
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+# In[ ]:
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def read_score(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# # Step1: Load Meta Data
+
+# In[ ]:
+
+assert target == 'IJBC' or target == 'IJBB'
+
+# =============================================================
+# load image and template relationships for template feature embedding
+# tid --> template id,  mid --> media id
+# format:
+#           image_name tid mid
+# =============================================================
+start = timeit.default_timer()
+templates, medias = read_template_media_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_face_tid_mid.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+
+# =============================================================
+# load template pairs for template-to-template verification
+# tid : template id,  label : 1/0
+# format:
+#           tid_1 tid_2 label
+# =============================================================
+start = timeit.default_timer()
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 2: Get Image Features
+
+# In[ ]:
+
+# =============================================================
+# load image features
+# format:
+#           img_feats: [image_num x feats_dim] (227630, 512)
+# =============================================================
+start = timeit.default_timer()
+img_path = '%s/loose_crop' % image_path
+img_list_path = '%s/meta/%s_name_5pts_score.txt' % (image_path, target.lower())
+img_list = open(img_list_path)
+files = img_list.readlines()
+# files_list = divideIntoNstrand(files, rank_size)
+files_list = files
+
+# img_feats
+# for i in range(rank_size):
+img_feats, faceness_scores = get_image_feature(img_path, files_list,
+                                               model_path, 0, gpu_id)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0],
+                                          img_feats.shape[1]))
+
+# # Step3: Get Template Features
+
+# In[ ]:
+
+# =============================================================
+# compute template features from image features.
+# =============================================================
+start = timeit.default_timer()
+# ==========================================================
+# Norm feature before aggregation into template feature?
+# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
+# ==========================================================
+# 1. FaceScore （Feature Norm）
+# 2. FaceScore （Detector）
+
+if use_flip_test:
+    # concat --- F1
+    # img_input_feats = img_feats
+    # add --- F2
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] //
+                                     2] + img_feats[:, img_feats.shape[1] // 2:]
+else:
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+if use_norm_score:
+    img_input_feats = img_input_feats
+else:
+    # normalise features to remove norm information
+    img_input_feats = img_input_feats / np.sqrt(
+        np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+if use_detector_score:
+    print(img_input_feats.shape, faceness_scores.shape)
+    img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+else:
+    img_input_feats = img_input_feats
+
+template_norm_feats, unique_templates = image2template_feature(
+    img_input_feats, templates, medias)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 4: Get Template Similarity Scores
+
+# In[ ]:
+
+# =============================================================
+# compute verification scores between template pairs.
+# =============================================================
+start = timeit.default_timer()
+score = verification(template_norm_feats, unique_templates, p1, p2)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+save_path = os.path.join(result_dir, args.job)
+# save_path = result_dir + '/%s_result' % target
+
+if not os.path.exists(save_path):
+    os.makedirs(save_path)
+
+score_save_file = os.path.join(save_path, "%s.npy" % target.lower())
+np.save(score_save_file, score)
+
+# # Step 5: Get ROC Curves and TPR@FPR Table
+
+# In[ ]:
+
+files = [score_save_file]
+methods = []
+scores = []
+for file in files:
+    methods.append(Path(file).stem)
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append("%s-%s" % (method, target))
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10 ** -6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+fig.savefig(os.path.join(save_path, '%s.pdf' % target.lower()))
+print(tpr_fpr_table)
diff --git a/recognition/arcface_torch/flops.py b/recognition/arcface_torch/flops.py
new file mode 100644
index 000000000..e704b7b58
--- /dev/null
+++ b/recognition/arcface_torch/flops.py
@@ -0,0 +1,20 @@
+from ptflops import get_model_complexity_info
+from backbones import get_model
+import argparse
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='')
+    parser.add_argument('n', type=str, default="r100")
+    args = parser.parse_args()
+    net = get_model(args.n)
+    macs, params = get_model_complexity_info(
+        net, (3, 112, 112), as_strings=False,
+        print_per_layer_stat=True, verbose=True)
+    gmacs = macs / (1000**3)
+    print("%.3f GFLOPs"%gmacs)
+    print("%.3f Mparams"%(params/(1000**2)))
+
+    if hasattr(net, "extra_gflops"):
+        print("%.3f Extra-GFLOPs"%net.extra_gflops)
+        print("%.3f Total-GFLOPs"%(gmacs+net.extra_gflops))
+
diff --git a/recognition/arcface_torch/inference.py b/recognition/arcface_torch/inference.py
new file mode 100644
index 000000000..3e5156e8d
--- /dev/null
+++ b/recognition/arcface_torch/inference.py
@@ -0,0 +1,35 @@
+import argparse
+
+import cv2
+import numpy as np
+import torch
+
+from backbones import get_model
+
+
+@torch.no_grad()
+def inference(weight, name, img):
+    if img is None:
+        img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.uint8)
+    else:
+        img = cv2.imread(img)
+        img = cv2.resize(img, (112, 112))
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = np.transpose(img, (2, 0, 1))
+    img = torch.from_numpy(img).unsqueeze(0).float()
+    img.div_(255).sub_(0.5).div_(0.5)
+    net = get_model(name, fp16=False)
+    net.load_state_dict(torch.load(weight))
+    net.eval()
+    feat = net(img).numpy()
+    print(feat)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='PyTorch ArcFace Training')
+    parser.add_argument('--network', type=str, default='r50', help='backbone network')
+    parser.add_argument('--weight', type=str, default='')
+    parser.add_argument('--img', type=str, default=None)
+    args = parser.parse_args()
+    inference(args.weight, args.network, args.img)
diff --git a/recognition/arcface_torch/losses.py b/recognition/arcface_torch/losses.py
new file mode 100644
index 000000000..7d54414c7
--- /dev/null
+++ b/recognition/arcface_torch/losses.py
@@ -0,0 +1,100 @@
+import torch
+import math
+
+
+class CombinedMarginLoss(torch.nn.Module):
+    def __init__(self, 
+                 s, 
+                 m1,
+                 m2,
+                 m3,
+                 interclass_filtering_threshold=0):
+        super().__init__()
+        self.s = s
+        self.m1 = m1
+        self.m2 = m2
+        self.m3 = m3
+        self.interclass_filtering_threshold = interclass_filtering_threshold
+        
+        # For ArcFace
+        self.cos_m = math.cos(self.m2)
+        self.sin_m = math.sin(self.m2)
+        self.theta = math.cos(math.pi - self.m2)
+        self.sinmm = math.sin(math.pi - self.m2) * self.m2
+        self.easy_margin = False
+
+
+    def forward(self, logits, labels):
+        index_positive = torch.where(labels != -1)[0]
+
+        if self.interclass_filtering_threshold > 0:
+            with torch.no_grad():
+                dirty = logits > self.interclass_filtering_threshold
+                dirty = dirty.float()
+                mask = torch.ones([index_positive.size(0), logits.size(1)], device=logits.device)
+                mask.scatter_(1, labels[index_positive], 0)
+                dirty[index_positive] *= mask
+                tensor_mul = 1 - dirty    
+            logits = tensor_mul * logits
+
+        target_logit = logits[index_positive, labels[index_positive].view(-1)]
+
+        if self.m1 == 1.0 and self.m3 == 0.0:
+            with torch.no_grad():
+                target_logit.arccos_()
+                logits.arccos_()
+                final_target_logit = target_logit + self.m2
+                logits[index_positive, labels[index_positive].view(-1)] = final_target_logit
+                logits.cos_()
+            logits = logits * self.s        
+
+        elif self.m3 > 0:
+            final_target_logit = target_logit - self.m3
+            logits[index_positive, labels[index_positive].view(-1)] = final_target_logit
+            logits = logits * self.s
+        else:
+            raise
+
+        return logits
+
+class ArcFace(torch.nn.Module):
+    """ ArcFace (https://arxiv.org/pdf/1801.07698v1.pdf):
+    """
+    def __init__(self, s=64.0, margin=0.5):
+        super(ArcFace, self).__init__()
+        self.s = s
+        self.margin = margin
+        self.cos_m = math.cos(margin)
+        self.sin_m = math.sin(margin)
+        self.theta = math.cos(math.pi - margin)
+        self.sinmm = math.sin(math.pi - margin) * margin
+        self.easy_margin = False
+
+
+    def forward(self, logits: torch.Tensor, labels: torch.Tensor):
+        index = torch.where(labels != -1)[0]
+        target_logit = logits[index, labels[index].view(-1)]
+
+        with torch.no_grad():
+            target_logit.arccos_()
+            logits.arccos_()
+            final_target_logit = target_logit + self.margin
+            logits[index, labels[index].view(-1)] = final_target_logit
+            logits.cos_()
+        logits = logits * self.s   
+        return logits
+
+
+class CosFace(torch.nn.Module):
+    def __init__(self, s=64.0, m=0.40):
+        super(CosFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, logits: torch.Tensor, labels: torch.Tensor):
+        index = torch.where(labels != -1)[0]
+        target_logit = logits[index, labels[index].view(-1)]
+        final_target_logit = target_logit - self.m
+        logits[index, labels[index].view(-1)] = final_target_logit
+        logits = logits * self.s
+        return logits
diff --git a/recognition/arcface_torch/lr_scheduler.py b/recognition/arcface_torch/lr_scheduler.py
new file mode 100644
index 000000000..6f3cda31d
--- /dev/null
+++ b/recognition/arcface_torch/lr_scheduler.py
@@ -0,0 +1,86 @@
+from torch.optim.lr_scheduler import _LRScheduler
+from torch.optim import SGD
+import torch
+import warnings
+
+class PolynomialLRWarmup(_LRScheduler):
+    def __init__(self, optimizer, warmup_iters, total_iters=5, power=1.0, last_epoch=-1, verbose=False):
+        super().__init__(optimizer, last_epoch=last_epoch, verbose=verbose)
+        self.total_iters = total_iters
+        self.power = power
+        self.warmup_iters = warmup_iters
+
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn("To get the last learning rate computed by the scheduler, "
+                          "please use `get_last_lr()`.", UserWarning)
+
+        if self.last_epoch == 0 or self.last_epoch > self.total_iters:
+            return [group["lr"] for group in self.optimizer.param_groups]
+
+        if self.last_epoch <= self.warmup_iters:
+            return [base_lr * self.last_epoch / self.warmup_iters for base_lr in self.base_lrs]
+        else:        
+            l = self.last_epoch
+            w = self.warmup_iters
+            t = self.total_iters
+            decay_factor = ((1.0 - (l - w) / (t - w)) / (1.0 - (l - 1 - w) / (t - w))) ** self.power
+        return [group["lr"] * decay_factor for group in self.optimizer.param_groups]
+
+    def _get_closed_form_lr(self):
+
+        if self.last_epoch <= self.warmup_iters:
+            return [
+                base_lr * self.last_epoch / self.warmup_iters for base_lr in self.base_lrs]
+        else:
+            return [
+                (
+                    base_lr * (1.0 - (min(self.total_iters, self.last_epoch) - self.warmup_iters) / (self.total_iters - self.warmup_iters)) ** self.power
+                )
+                for base_lr in self.base_lrs
+            ]
+
+    
+if __name__ == "__main__":
+
+    class TestModule(torch.nn.Module):
+        def __init__(self) -> None:
+            super().__init__()
+            self.linear = torch.nn.Linear(32, 32)
+        
+        def forward(self, x):
+            return self.linear(x)
+
+    test_module = TestModule()
+    test_module_pfc = TestModule()
+    lr_pfc_weight = 1 / 3
+    base_lr = 10
+    total_steps = 1000
+    
+    sgd = SGD([
+        {"params": test_module.parameters(), "lr": base_lr},
+        {"params": test_module_pfc.parameters(), "lr": base_lr * lr_pfc_weight}
+        ], base_lr)
+
+    scheduler = PolynomialLRWarmup(sgd, total_steps//10, total_steps, power=2)
+
+    x = []
+    y = []
+    y_pfc = []
+    for i in range(total_steps):
+        scheduler.step()
+        lr = scheduler.get_last_lr()[0]
+        lr_pfc = scheduler.get_last_lr()[1]
+        x.append(i)
+        y.append(lr)
+        y_pfc.append(lr_pfc)
+
+    import matplotlib.pyplot as plt
+    fontsize=15
+    plt.figure(figsize=(6, 6))
+    plt.plot(x, y, linestyle='-', linewidth=2, )
+    plt.plot(x, y_pfc, linestyle='-', linewidth=2, )
+    plt.xlabel('Iterations')     # x_label
+    plt.ylabel("Lr")             # y_label
+    plt.savefig("tmp.png", dpi=600, bbox_inches='tight')
diff --git a/recognition/arcface_torch/onnx_helper.py b/recognition/arcface_torch/onnx_helper.py
new file mode 100644
index 000000000..ca922ca6d
--- /dev/null
+++ b/recognition/arcface_torch/onnx_helper.py
@@ -0,0 +1,250 @@
+from __future__ import division
+import datetime
+import os
+import os.path as osp
+import glob
+import numpy as np
+import cv2
+import sys
+import onnxruntime
+import onnx
+import argparse
+from onnx import numpy_helper
+from insightface.data import get_image
+
+class ArcFaceORT:
+    def __init__(self, model_path, cpu=False):
+        self.model_path = model_path
+        # providers = None will use available provider, for onnxruntime-gpu it will be "CUDAExecutionProvider"
+        self.providers = ['CPUExecutionProvider'] if cpu else None
+
+    #input_size is (w,h), return error message, return None if success
+    def check(self, track='cfat', test_img = None):
+        #default is cfat
+        max_model_size_mb=1024
+        max_feat_dim=512
+        max_time_cost=15
+        if track.startswith('ms1m'):
+            max_model_size_mb=1024
+            max_feat_dim=512
+            max_time_cost=10
+        elif track.startswith('glint'):
+            max_model_size_mb=1024
+            max_feat_dim=1024
+            max_time_cost=20
+        elif track.startswith('cfat'):
+            max_model_size_mb = 1024
+            max_feat_dim = 512
+            max_time_cost = 15
+        elif track.startswith('unconstrained'):
+            max_model_size_mb=1024
+            max_feat_dim=1024
+            max_time_cost=30
+        else:
+            return "track not found"
+
+        if not os.path.exists(self.model_path):
+            return "model_path not exists"
+        if not os.path.isdir(self.model_path):
+            return "model_path should be directory"
+        onnx_files = []
+        for _file in os.listdir(self.model_path):
+            if _file.endswith('.onnx'):
+                onnx_files.append(osp.join(self.model_path, _file))
+        if len(onnx_files)==0:
+            return "do not have onnx files"
+        self.model_file = sorted(onnx_files)[-1]
+        print('use onnx-model:', self.model_file)
+        try:
+            session = onnxruntime.InferenceSession(self.model_file, providers=self.providers)
+        except:
+            return "load onnx failed"
+        input_cfg = session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        print('input-shape:', input_shape)
+        if len(input_shape)!=4:
+            return "length of input_shape should be 4"
+        if not isinstance(input_shape[0], str):
+            #return "input_shape[0] should be str to support batch-inference"
+            print('reset input-shape[0] to None')
+            model = onnx.load(self.model_file)
+            model.graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+            new_model_file = osp.join(self.model_path, 'zzzzrefined.onnx')
+            onnx.save(model, new_model_file)
+            self.model_file = new_model_file
+            print('use new onnx-model:', self.model_file)
+            try:
+                session = onnxruntime.InferenceSession(self.model_file, providers=self.providers)
+            except:
+                return "load onnx failed"
+            input_cfg = session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            print('new-input-shape:', input_shape)
+
+        self.image_size = tuple(input_shape[2:4][::-1])
+        #print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        outputs = session.get_outputs()
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+            #print(o.name, o.shape)
+        if len(output_names)!=1:
+            return "number of output nodes should be 1"
+        self.session = session
+        self.input_name = input_name
+        self.output_names = output_names
+        #print(self.output_names)
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        if len(graph.node)<8:
+            return "too small onnx graph"
+
+        input_size = (112,112)
+        self.crop = None
+        if track=='cfat':
+            crop_file = osp.join(self.model_path, 'crop.txt')
+            if osp.exists(crop_file):
+                lines = open(crop_file,'r').readlines()
+                if len(lines)!=6:
+                    return "crop.txt should contain 6 lines"
+                lines = [int(x) for x in lines]
+                self.crop = lines[:4]
+                input_size = tuple(lines[4:6])
+        if input_size!=self.image_size:
+            return "input-size is inconsistant with onnx model input, %s vs %s"%(input_size, self.image_size)
+
+        self.model_size_mb = os.path.getsize(self.model_file) / float(1024*1024)
+        if self.model_size_mb > max_model_size_mb:
+            return "max model size exceed, given %.3f-MB"%self.model_size_mb
+
+        input_mean = None
+        input_std = None
+        if track=='cfat':
+            pn_file = osp.join(self.model_path, 'pixel_norm.txt')
+            if osp.exists(pn_file):
+                lines = open(pn_file,'r').readlines()
+                if len(lines)!=2:
+                    return "pixel_norm.txt should contain 2 lines"
+                input_mean = float(lines[0])
+                input_std = float(lines[1])
+        if input_mean is not None or input_std is not None:
+            if input_mean is None or input_std is None:
+                return "please set input_mean and input_std simultaneously"
+        else:
+            find_sub = False
+            find_mul = False
+            for nid, node in enumerate(graph.node[:8]):
+                print(nid, node.name)
+                if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                    find_sub = True
+                if node.name.startswith('Mul') or node.name.startswith('_mul') or node.name.startswith('Div'):
+                    find_mul = True
+            if find_sub and find_mul:
+                print("find sub and mul")
+                #mxnet arcface model
+                input_mean = 0.0
+                input_std = 1.0
+            else:
+                input_mean = 127.5
+                input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        for initn in graph.initializer:
+            weight_array = numpy_helper.to_array(initn)
+            dt = weight_array.dtype
+            if dt.itemsize<4:
+                return 'invalid weight type - (%s:%s)' % (initn.name, dt.name)
+        if test_img is None:
+            test_img = get_image('Tom_Hanks_54745')
+            test_img = cv2.resize(test_img, self.image_size)
+        else:
+            test_img = cv2.resize(test_img, self.image_size)
+        feat, cost = self.benchmark(test_img)
+        batch_result = self.check_batch(test_img)
+        batch_result_sum = float(np.sum(batch_result))
+        if batch_result_sum in [float('inf'), -float('inf')] or batch_result_sum != batch_result_sum:
+            print(batch_result)
+            print(batch_result_sum)
+            return "batch result output contains NaN!"
+
+        if len(feat.shape) < 2:
+           return "the shape of the feature must be two, but get {}".format(str(feat.shape))
+
+        if feat.shape[1] > max_feat_dim:
+            return "max feat dim exceed, given %d"%feat.shape[1]
+        self.feat_dim = feat.shape[1]
+        cost_ms = cost*1000
+        if cost_ms>max_time_cost:
+            return "max time cost exceed, given %.4f"%cost_ms
+        self.cost_ms = cost_ms
+        print('check stat:, model-size-mb: %.4f, feat-dim: %d, time-cost-ms: %.4f, input-mean: %.3f, input-std: %.3f'%(self.model_size_mb, self.feat_dim, self.cost_ms, self.input_mean, self.input_std))
+        return None
+
+    def check_batch(self, img):
+        if not isinstance(img, list):
+            imgs = [img, ] * 32
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1]:self.crop[3], self.crop[0]:self.crop[2], :]
+                if nimg.shape[0] != self.image_size[1] or nimg.shape[1] != self.image_size[0]:
+                    nimg = cv2.resize(nimg, self.image_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(
+            images=imgs, scalefactor=1.0 / self.input_std, size=self.image_size,
+            mean=(self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+
+    def meta_info(self):
+        return {'model-size-mb':self.model_size_mb, 'feature-dim':self.feat_dim, 'infer': self.cost_ms}
+
+
+    def forward(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.image_size
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1]:self.crop[3],self.crop[0]:self.crop[2],:]
+                if nimg.shape[0]!=input_size[1] or nimg.shape[1]!=input_size[0]:
+                    nimg = cv2.resize(nimg, input_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(imgs, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name : blob})[0]
+        return net_out
+
+    def benchmark(self, img):
+        input_size = self.image_size
+        if self.crop is not None:
+            nimg = img[self.crop[1]:self.crop[3],self.crop[0]:self.crop[2],:]
+            if nimg.shape[0]!=input_size[1] or nimg.shape[1]!=input_size[0]:
+                nimg = cv2.resize(nimg, input_size)
+            img = nimg
+        blob = cv2.dnn.blobFromImage(img, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        costs = []
+        for _ in range(50):
+            ta = datetime.datetime.now()
+            net_out = self.session.run(self.output_names, {self.input_name : blob})[0]
+            tb = datetime.datetime.now()
+            cost = (tb-ta).total_seconds()
+            costs.append(cost)
+        costs = sorted(costs)
+        cost = costs[5]
+        return net_out, cost
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='')
+    # general
+    parser.add_argument('workdir', help='submitted work dir', type=str)
+    parser.add_argument('--track', help='track name, for different challenge', type=str, default='cfat')
+    args = parser.parse_args()
+    handler = ArcFaceORT(args.workdir)
+    err = handler.check(args.track)
+    print('err:', err)
diff --git a/recognition/arcface_torch/onnx_ijbc.py b/recognition/arcface_torch/onnx_ijbc.py
new file mode 100644
index 000000000..31c491b1b
--- /dev/null
+++ b/recognition/arcface_torch/onnx_ijbc.py
@@ -0,0 +1,269 @@
+import argparse
+import os
+import pickle
+import timeit
+
+import cv2
+import mxnet as mx
+import numpy as np
+import pandas as pd
+import prettytable
+import skimage.transform
+import torch
+from sklearn.metrics import roc_curve
+from sklearn.preprocessing import normalize
+from torch.utils.data import DataLoader
+from onnx_helper import ArcFaceORT
+
+SRC = np.array(
+    [
+        [30.2946, 51.6963],
+        [65.5318, 51.5014],
+        [48.0252, 71.7366],
+        [33.5493, 92.3655],
+        [62.7299, 92.2041]]
+    , dtype=np.float32)
+SRC[:, 0] += 8.0
+
+
+@torch.no_grad()
+class AlignedDataSet(mx.gluon.data.Dataset):
+    def __init__(self, root, lines, align=True):
+        self.lines = lines
+        self.root = root
+        self.align = align
+
+    def __len__(self):
+        return len(self.lines)
+
+    def __getitem__(self, idx):
+        each_line = self.lines[idx]
+        name_lmk_score = each_line.strip().split(' ')
+        name = os.path.join(self.root, name_lmk_score[0])
+        img = cv2.cvtColor(cv2.imread(name), cv2.COLOR_BGR2RGB)
+        landmark5 = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32).reshape((5, 2))
+        st = skimage.transform.SimilarityTransform()
+        st.estimate(landmark5, SRC)
+        img = cv2.warpAffine(img, st.params[0:2, :], (112, 112), borderValue=0.0)
+        img_1 = np.expand_dims(img, 0)
+        img_2 = np.expand_dims(np.fliplr(img), 0)
+        output = np.concatenate((img_1, img_2), axis=0).astype(np.float32)
+        output = np.transpose(output, (0, 3, 1, 2))
+        return torch.from_numpy(output)
+
+
+@torch.no_grad()
+def extract(model_root, dataset):
+    model = ArcFaceORT(model_path=model_root)
+    model.check()
+    feat_mat = np.zeros(shape=(len(dataset), 2 * model.feat_dim))
+
+    def collate_fn(data):
+        return torch.cat(data, dim=0)
+
+    data_loader = DataLoader(
+        dataset, batch_size=128, drop_last=False, num_workers=4, collate_fn=collate_fn, )
+    num_iter = 0
+    for batch in data_loader:
+        batch = batch.numpy()
+        batch = (batch - model.input_mean) / model.input_std
+        feat = model.session.run(model.output_names, {model.input_name: batch})[0]
+        feat = np.reshape(feat, (-1, model.feat_dim * 2))
+        feat_mat[128 * num_iter: 128 * num_iter + feat.shape[0], :] = feat
+        num_iter += 1
+        if num_iter % 50 == 0:
+            print(num_iter)
+    return feat_mat
+
+
+def read_template_media_list(path):
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+def image2template_feature(img_feats=None,
+                           templates=None,
+                           medias=None):
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [np.mean(face_norm_feats[ind_m], axis=0, keepdims=True), ]
+        media_norm_feats = np.array(media_norm_feats)
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    template_norm_feats = normalize(template_feats)
+    return template_norm_feats, unique_templates
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000
+    sublists = [total_pairs[i: i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def main(args):
+    use_norm_score = True  # if Ture, TestMode(N1)
+    use_detector_score = True  # if Ture, TestMode(D1)
+    use_flip_test = True  # if Ture, TestMode(F1)
+    assert args.target == 'IJBC' or args.target == 'IJBB'
+
+    start = timeit.default_timer()
+    templates, medias = read_template_media_list(
+        os.path.join('%s/meta' % args.image_path, '%s_face_tid_mid.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    p1, p2, label = read_template_pair_list(
+        os.path.join('%s/meta' % args.image_path,
+                     '%s_template_pair_label.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    img_path = '%s/loose_crop' % args.image_path
+    img_list_path = '%s/meta/%s_name_5pts_score.txt' % (args.image_path, args.target.lower())
+    img_list = open(img_list_path)
+    files = img_list.readlines()
+    dataset = AlignedDataSet(root=img_path, lines=files, align=True)
+    img_feats = extract(args.model_root, dataset)
+
+    faceness_scores = []
+    for each_line in files:
+        name_lmk_score = each_line.split()
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
+    start = timeit.default_timer()
+
+    if use_flip_test:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2] + img_feats[:, img_feats.shape[1] // 2:]
+    else:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+    if use_norm_score:
+        img_input_feats = img_input_feats
+    else:
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+    if use_detector_score:
+        print(img_input_feats.shape, faceness_scores.shape)
+        img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+    else:
+        img_input_feats = img_input_feats
+
+    template_norm_feats, unique_templates = image2template_feature(
+        img_input_feats, templates, medias)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    score = verification(template_norm_feats, unique_templates, p1, p2)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    result_dir = args.model_root
+
+    save_path = os.path.join(result_dir, "{}_result".format(args.target))
+    if not os.path.exists(save_path):
+        os.makedirs(save_path)
+    score_save_file = os.path.join(save_path, "{}.npy".format(args.target))
+    np.save(score_save_file, score)
+    files = [score_save_file]
+    methods = []
+    scores = []
+    for file in files:
+        methods.append(os.path.basename(file))
+        scores.append(np.load(file))
+    methods = np.array(methods)
+    scores = dict(zip(methods, scores))
+    x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+    tpr_fpr_table = prettytable.PrettyTable(['Methods'] + [str(x) for x in x_labels])
+    for method in methods:
+        fpr, tpr, _ = roc_curve(label, scores[method])
+        fpr = np.flipud(fpr)
+        tpr = np.flipud(tpr)
+        tpr_fpr_row = []
+        tpr_fpr_row.append("%s-%s" % (method, args.target))
+        for fpr_iter in np.arange(len(x_labels)):
+            _, min_index = min(
+                list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+            tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+        tpr_fpr_table.add_row(tpr_fpr_row)
+    print(tpr_fpr_table)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='do ijb test')
+    # general
+    parser.add_argument('--model-root', default='', help='path to load model.')
+    parser.add_argument('--image-path', default='/train_tmp/IJB_release/IJBC', type=str, help='')
+    parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+    main(parser.parse_args())
diff --git a/recognition/arcface_torch/partial_fc_v2.py b/recognition/arcface_torch/partial_fc_v2.py
new file mode 100644
index 000000000..0752554ca
--- /dev/null
+++ b/recognition/arcface_torch/partial_fc_v2.py
@@ -0,0 +1,260 @@
+
+import math
+from typing import Callable
+
+import torch
+from torch import distributed
+from torch.nn.functional import linear, normalize
+
+
+class PartialFC_V2(torch.nn.Module):
+    """
+    https://arxiv.org/abs/2203.15565
+    A distributed sparsely updating variant of the FC layer, named Partial FC (PFC).
+    When sample rate less than 1, in each iteration, positive class centers and a random subset of
+    negative class centers are selected to compute the margin-based softmax loss, all class
+    centers are still maintained throughout the whole training process, but only a subset is
+    selected and updated in each iteration.
+    .. note::
+        When sample rate equal to 1, Partial FC is equal to model parallelism(default sample rate is 1).
+    Example:
+    --------
+    >>> module_pfc = PartialFC(embedding_size=512, num_classes=8000000, sample_rate=0.2)
+    >>> for img, labels in data_loader:
+    >>>     embeddings = net(img)
+    >>>     loss = module_pfc(embeddings, labels)
+    >>>     loss.backward()
+    >>>     optimizer.step()
+    """
+    _version = 2
+
+    def __init__(
+        self,
+        margin_loss: Callable,
+        embedding_size: int,
+        num_classes: int,
+        sample_rate: float = 1.0,
+        fp16: bool = False,
+    ):
+        """
+        Paramenters:
+        -----------
+        embedding_size: int
+            The dimension of embedding, required
+        num_classes: int
+            Total number of classes, required
+        sample_rate: float
+            The rate of negative centers participating in the calculation, default is 1.0.
+        """
+        super(PartialFC_V2, self).__init__()
+        assert (
+            distributed.is_initialized()
+        ), "must initialize distributed before create this"
+        self.rank = distributed.get_rank()
+        self.world_size = distributed.get_world_size()
+
+        self.dist_cross_entropy = DistCrossEntropy()
+        self.embedding_size = embedding_size
+        self.sample_rate: float = sample_rate
+        self.fp16 = fp16
+        self.num_local: int = num_classes // self.world_size + int(
+            self.rank < num_classes % self.world_size
+        )
+        self.class_start: int = num_classes // self.world_size * self.rank + min(
+            self.rank, num_classes % self.world_size
+        )
+        self.num_sample: int = int(self.sample_rate * self.num_local)
+        self.last_batch_size: int = 0
+
+        self.is_updated: bool = True
+        self.init_weight_update: bool = True
+        self.weight = torch.nn.Parameter(torch.normal(0, 0.01, (self.num_local, embedding_size)))
+
+        # margin_loss
+        if isinstance(margin_loss, Callable):
+            self.margin_softmax = margin_loss
+        else:
+            raise
+
+    def sample(self, labels, index_positive):
+        """
+            This functions will change the value of labels
+            Parameters:
+            -----------
+            labels: torch.Tensor
+                pass
+            index_positive: torch.Tensor
+                pass
+            optimizer: torch.optim.Optimizer
+                pass
+        """
+        with torch.no_grad():
+            positive = torch.unique(labels[index_positive], sorted=True).cuda()
+            if self.num_sample - positive.size(0) >= 0:
+                perm = torch.rand(size=[self.num_local]).cuda()
+                perm[positive] = 2.0
+                index = torch.topk(perm, k=self.num_sample)[1].cuda()
+                index = index.sort()[0].cuda()
+            else:
+                index = positive
+            self.weight_index = index
+
+            labels[index_positive] = torch.searchsorted(index, labels[index_positive])
+
+        return self.weight[self.weight_index]
+
+    def forward(
+        self,
+        local_embeddings: torch.Tensor,
+        local_labels: torch.Tensor,
+    ):
+        """
+        Parameters:
+        ----------
+        local_embeddings: torch.Tensor
+            feature embeddings on each GPU(Rank).
+        local_labels: torch.Tensor
+            labels on each GPU(Rank).
+        Returns:
+        -------
+        loss: torch.Tensor
+            pass
+        """
+        local_labels.squeeze_()
+        local_labels = local_labels.long()
+
+        batch_size = local_embeddings.size(0)
+        if self.last_batch_size == 0:
+            self.last_batch_size = batch_size
+        assert self.last_batch_size == batch_size, (
+            f"last batch size do not equal current batch size: {self.last_batch_size} vs {batch_size}")
+
+        _gather_embeddings = [
+            torch.zeros((batch_size, self.embedding_size)).cuda()
+            for _ in range(self.world_size)
+        ]
+        _gather_labels = [
+            torch.zeros(batch_size).long().cuda() for _ in range(self.world_size)
+        ]
+        _list_embeddings = AllGather(local_embeddings, *_gather_embeddings)
+        distributed.all_gather(_gather_labels, local_labels)
+
+        embeddings = torch.cat(_list_embeddings)
+        labels = torch.cat(_gather_labels)
+
+        labels = labels.view(-1, 1)
+        index_positive = (self.class_start <= labels) & (
+            labels < self.class_start + self.num_local
+        )
+        labels[~index_positive] = -1
+        labels[index_positive] -= self.class_start
+
+        if self.sample_rate < 1:
+            weight = self.sample(labels, index_positive)
+        else:
+            weight = self.weight
+
+        with torch.cuda.amp.autocast(self.fp16):
+            norm_embeddings = normalize(embeddings)
+            norm_weight_activated = normalize(weight)
+            logits = linear(norm_embeddings, norm_weight_activated)
+        if self.fp16:
+            logits = logits.float()
+        logits = logits.clamp(-1, 1)
+
+        logits = self.margin_softmax(logits, labels)
+        loss = self.dist_cross_entropy(logits, labels)
+        return loss
+
+
+class DistCrossEntropyFunc(torch.autograd.Function):
+    """
+    CrossEntropy loss is calculated in parallel, allreduce denominator into single gpu and calculate softmax.
+    Implemented of ArcFace (https://arxiv.org/pdf/1801.07698v1.pdf):
+    """
+
+    @staticmethod
+    def forward(ctx, logits: torch.Tensor, label: torch.Tensor):
+        """ """
+        batch_size = logits.size(0)
+        # for numerical stability
+        max_logits, _ = torch.max(logits, dim=1, keepdim=True)
+        # local to global
+        distributed.all_reduce(max_logits, distributed.ReduceOp.MAX)
+        logits.sub_(max_logits)
+        logits.exp_()
+        sum_logits_exp = torch.sum(logits, dim=1, keepdim=True)
+        # local to global
+        distributed.all_reduce(sum_logits_exp, distributed.ReduceOp.SUM)
+        logits.div_(sum_logits_exp)
+        index = torch.where(label != -1)[0]
+        # loss
+        loss = torch.zeros(batch_size, 1, device=logits.device)
+        loss[index] = logits[index].gather(1, label[index])
+        distributed.all_reduce(loss, distributed.ReduceOp.SUM)
+        ctx.save_for_backward(index, logits, label)
+        return loss.clamp_min_(1e-30).log_().mean() * (-1)
+
+    @staticmethod
+    def backward(ctx, loss_gradient):
+        """
+        Args:
+            loss_grad (torch.Tensor): gradient backward by last layer
+        Returns:
+            gradients for each input in forward function
+            `None` gradients for one-hot label
+        """
+        (
+            index,
+            logits,
+            label,
+        ) = ctx.saved_tensors
+        batch_size = logits.size(0)
+        one_hot = torch.zeros(
+            size=[index.size(0), logits.size(1)], device=logits.device
+        )
+        one_hot.scatter_(1, label[index], 1)
+        logits[index] -= one_hot
+        logits.div_(batch_size)
+        return logits * loss_gradient.item(), None
+
+
+class DistCrossEntropy(torch.nn.Module):
+    def __init__(self):
+        super(DistCrossEntropy, self).__init__()
+
+    def forward(self, logit_part, label_part):
+        return DistCrossEntropyFunc.apply(logit_part, label_part)
+
+
+class AllGatherFunc(torch.autograd.Function):
+    """AllGather op with gradient backward"""
+
+    @staticmethod
+    def forward(ctx, tensor, *gather_list):
+        gather_list = list(gather_list)
+        distributed.all_gather(gather_list, tensor)
+        return tuple(gather_list)
+
+    @staticmethod
+    def backward(ctx, *grads):
+        grad_list = list(grads)
+        rank = distributed.get_rank()
+        grad_out = grad_list[rank]
+
+        dist_ops = [
+            distributed.reduce(grad_out, rank, distributed.ReduceOp.SUM, async_op=True)
+            if i == rank
+            else distributed.reduce(
+                grad_list[i], i, distributed.ReduceOp.SUM, async_op=True
+            )
+            for i in range(distributed.get_world_size())
+        ]
+        for _op in dist_ops:
+            _op.wait()
+
+        grad_out *= len(grad_list)  # cooperate with distributed loss function
+        return (grad_out, *[None for _ in range(len(grad_list))])
+
+
+AllGather = AllGatherFunc.apply
diff --git a/recognition/arcface_torch/requirement.txt b/recognition/arcface_torch/requirement.txt
new file mode 100644
index 000000000..f1a431ef9
--- /dev/null
+++ b/recognition/arcface_torch/requirement.txt
@@ -0,0 +1,6 @@
+tensorboard
+easydict
+mxnet
+onnx
+sklearn
+opencv-python
\ No newline at end of file
diff --git a/recognition/arcface_torch/run.sh b/recognition/arcface_torch/run.sh
new file mode 100644
index 000000000..6eacdf8e8
--- /dev/null
+++ b/recognition/arcface_torch/run.sh
@@ -0,0 +1 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 train_v2.py $@
diff --git a/recognition/arcface_torch/scripts/shuffle_rec.py b/recognition/arcface_torch/scripts/shuffle_rec.py
new file mode 100644
index 000000000..f3b68e938
--- /dev/null
+++ b/recognition/arcface_torch/scripts/shuffle_rec.py
@@ -0,0 +1,81 @@
+import argparse
+import multiprocessing
+import os
+import time
+
+import mxnet as mx
+import numpy as np
+
+
+def read_worker(args, q_in):
+    path_imgidx = os.path.join(args.input, "train.idx")
+    path_imgrec = os.path.join(args.input, "train.rec")
+    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, "r")
+
+    s = imgrec.read_idx(0)
+    header, _ = mx.recordio.unpack(s)
+    assert header.flag > 0
+
+    imgidx = np.array(range(1, int(header.label[0])))
+    np.random.shuffle(imgidx)
+    
+    for idx in imgidx:
+        item = imgrec.read_idx(idx)
+        q_in.put(item)
+
+    q_in.put(None)
+    imgrec.close()
+
+
+def write_worker(args, q_out):
+    pre_time = time.time()
+    
+    if args.input[-1] == '/':
+        args.input = args.input[:-1]
+    dirname = os.path.dirname(args.input)
+    basename = os.path.basename(args.input)
+    output = os.path.join(dirname, f"shuffled_{basename}")
+    os.makedirs(output, exist_ok=True)
+    
+    path_imgidx = os.path.join(output, "train.idx")
+    path_imgrec = os.path.join(output, "train.rec")
+    save_record = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, "w")
+    more = True
+    count = 0
+    while more:
+        deq = q_out.get()
+        if deq is None:
+            more = False
+        else:
+            header, jpeg = mx.recordio.unpack(deq)
+            # TODO it is currently not fully developed
+            if isinstance(header.label, float):
+                label = header.label
+            else:
+                label = header.label[0]
+
+            header = mx.recordio.IRHeader(flag=header.flag, label=label, id=header.id, id2=header.id2)
+            save_record.write_idx(count, mx.recordio.pack(header, jpeg))
+            count += 1
+            if count % 10000 == 0:
+                cur_time = time.time()
+                print('save time:', cur_time - pre_time, ' count:', count)
+                pre_time = cur_time
+    print(count)
+    save_record.close()
+
+
+def main(args):
+    queue = multiprocessing.Queue(10240)
+    read_process = multiprocessing.Process(target=read_worker, args=(args, queue))
+    read_process.daemon = True
+    read_process.start()
+    write_process = multiprocessing.Process(target=write_worker, args=(args, queue))
+    write_process.start()
+    write_process.join()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('input', help='path to source rec.')
+    main(parser.parse_args())
diff --git a/recognition/arcface_torch/torch2onnx.py b/recognition/arcface_torch/torch2onnx.py
new file mode 100644
index 000000000..f6055d1fe
--- /dev/null
+++ b/recognition/arcface_torch/torch2onnx.py
@@ -0,0 +1,53 @@
+import numpy as np
+import onnx
+import torch
+
+
+def convert_onnx(net, path_module, output, opset=11, simplify=False):
+    assert isinstance(net, torch.nn.Module)
+    img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.int32)
+    img = img.astype(np.float)
+    img = (img / 255. - 0.5) / 0.5  # torch style norm
+    img = img.transpose((2, 0, 1))
+    img = torch.from_numpy(img).unsqueeze(0).float()
+
+    weight = torch.load(path_module)
+    net.load_state_dict(weight, strict=True)
+    net.eval()
+    torch.onnx.export(net, img, output, input_names=["data"], keep_initializers_as_inputs=False, verbose=False, opset_version=opset)
+    model = onnx.load(output)
+    graph = model.graph
+    graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+    if simplify:
+        from onnxsim import simplify
+        model, check = simplify(model)
+        assert check, "Simplified ONNX model could not be validated"
+    onnx.save(model, output)
+
+    
+if __name__ == '__main__':
+    import os
+    import argparse
+    from backbones import get_model
+
+    parser = argparse.ArgumentParser(description='ArcFace PyTorch to onnx')
+    parser.add_argument('input', type=str, help='input backbone.pth file or path')
+    parser.add_argument('--output', type=str, default=None, help='output onnx path')
+    parser.add_argument('--network', type=str, default=None, help='backbone network')
+    parser.add_argument('--simplify', type=bool, default=False, help='onnx simplify')
+    args = parser.parse_args()
+    input_file = args.input
+    if os.path.isdir(input_file):
+        input_file = os.path.join(input_file, "model.pt")
+    assert os.path.exists(input_file)
+    # model_name = os.path.basename(os.path.dirname(input_file)).lower()
+    # params = model_name.split("_")
+    # if len(params) >= 3 and params[1] in ('arcface', 'cosface'):
+    #     if args.network is None:
+    #         args.network = params[2]
+    assert args.network is not None
+    print(args)
+    backbone_onnx = get_model(args.network, dropout=0.0, fp16=False, num_features=512)
+    if args.output is None:
+        args.output = os.path.join(os.path.dirname(args.input), "model.onnx")
+    convert_onnx(backbone_onnx, input_file, args.output, simplify=args.simplify)
diff --git a/recognition/arcface_torch/train_v2.py b/recognition/arcface_torch/train_v2.py
new file mode 100755
index 000000000..9563c6533
--- /dev/null
+++ b/recognition/arcface_torch/train_v2.py
@@ -0,0 +1,257 @@
+import argparse
+import logging
+import os
+from datetime import datetime
+
+import numpy as np
+import torch
+from backbones import get_model
+from dataset import get_dataloader
+from losses import CombinedMarginLoss
+from lr_scheduler import PolynomialLRWarmup
+from partial_fc_v2 import PartialFC_V2
+from torch import distributed
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+from utils.utils_callbacks import CallBackLogging, CallBackVerification
+from utils.utils_config import get_config
+from utils.utils_distributed_sampler import setup_seed
+from utils.utils_logging import AverageMeter, init_logging
+from torch.distributed.algorithms.ddp_comm_hooks.default_hooks import fp16_compress_hook
+
+assert torch.__version__ >= "1.12.0", "In order to enjoy the features of the new torch, \
+we have upgraded the torch to 1.12.0. torch before than 1.12.0 may not work in the future."
+
+try:
+    rank = int(os.environ["RANK"])
+    local_rank = int(os.environ["LOCAL_RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+    distributed.init_process_group("nccl")
+except KeyError:
+    rank = 0
+    local_rank = 0
+    world_size = 1
+    distributed.init_process_group(
+        backend="nccl",
+        init_method="tcp://127.0.0.1:12584",
+        rank=rank,
+        world_size=world_size,
+    )
+
+
+def main(args):
+
+    # get config
+    cfg = get_config(args.config)
+    # global control random seed
+    setup_seed(seed=cfg.seed, cuda_deterministic=False)
+
+    torch.cuda.set_device(local_rank)
+
+    os.makedirs(cfg.output, exist_ok=True)
+    init_logging(rank, cfg.output)
+
+    summary_writer = (
+        SummaryWriter(log_dir=os.path.join(cfg.output, "tensorboard"))
+        if rank == 0
+        else None
+    )
+    
+    wandb_logger = None
+    if cfg.using_wandb:
+        import wandb
+        # Sign in to wandb
+        try:
+            wandb.login(key=cfg.wandb_key)
+        except Exception as e:
+            print("WandB Key must be provided in config file (base.py).")
+            print(f"Config Error: {e}")
+        # Initialize wandb
+        run_name = datetime.now().strftime("%y%m%d_%H%M") + f"_GPU{rank}"
+        run_name = run_name if cfg.suffix_run_name is None else run_name + f"_{cfg.suffix_run_name}"
+        try:
+            wandb_logger = wandb.init(
+                entity = cfg.wandb_entity, 
+                project = cfg.wandb_project, 
+                sync_tensorboard = True,
+                resume=cfg.wandb_resume,
+                name = run_name, 
+                notes = cfg.notes) if rank == 0 or cfg.wandb_log_all else None
+            if wandb_logger:
+                wandb_logger.config.update(cfg)
+        except Exception as e:
+            print("WandB Data (Entity and Project name) must be provided in config file (base.py).")
+            print(f"Config Error: {e}")
+    train_loader = get_dataloader(
+        cfg.rec,
+        local_rank,
+        cfg.batch_size,
+        cfg.dali,
+        cfg.dali_aug,
+        cfg.seed,
+        cfg.num_workers
+    )
+
+    backbone = get_model(
+        cfg.network, dropout=0.0, fp16=cfg.fp16, num_features=cfg.embedding_size).cuda()
+
+    backbone = torch.nn.parallel.DistributedDataParallel(
+        module=backbone, broadcast_buffers=False, device_ids=[local_rank], bucket_cap_mb=16,
+        find_unused_parameters=True)
+    backbone.register_comm_hook(None, fp16_compress_hook)
+
+    backbone.train()
+    # FIXME using gradient checkpoint if there are some unused parameters will cause error
+    backbone._set_static_graph()
+
+    margin_loss = CombinedMarginLoss(
+        64,
+        cfg.margin_list[0],
+        cfg.margin_list[1],
+        cfg.margin_list[2],
+        cfg.interclass_filtering_threshold
+    )
+
+    if cfg.optimizer == "sgd":
+        module_partial_fc = PartialFC_V2(
+            margin_loss, cfg.embedding_size, cfg.num_classes,
+            cfg.sample_rate, False)
+        module_partial_fc.train().cuda()
+        # TODO the params of partial fc must be last in the params list
+        opt = torch.optim.SGD(
+            params=[{"params": backbone.parameters()}, {"params": module_partial_fc.parameters()}],
+            lr=cfg.lr, momentum=0.9, weight_decay=cfg.weight_decay)
+
+    elif cfg.optimizer == "adamw":
+        module_partial_fc = PartialFC_V2(
+            margin_loss, cfg.embedding_size, cfg.num_classes,
+            cfg.sample_rate, False)
+        module_partial_fc.train().cuda()
+        opt = torch.optim.AdamW(
+            params=[{"params": backbone.parameters()}, {"params": module_partial_fc.parameters()}],
+            lr=cfg.lr, weight_decay=cfg.weight_decay)
+    else:
+        raise
+
+    cfg.total_batch_size = cfg.batch_size * world_size
+    cfg.warmup_step = cfg.num_image // cfg.total_batch_size * cfg.warmup_epoch
+    cfg.total_step = cfg.num_image // cfg.total_batch_size * cfg.num_epoch
+
+    lr_scheduler = PolynomialLRWarmup(
+        optimizer=opt,
+        warmup_iters=cfg.warmup_step,
+        total_iters=cfg.total_step)
+
+    start_epoch = 0
+    global_step = 0
+    if cfg.resume:
+        dict_checkpoint = torch.load(os.path.join(cfg.output, f"checkpoint_gpu_{rank}.pt"))
+        start_epoch = dict_checkpoint["epoch"]
+        global_step = dict_checkpoint["global_step"]
+        backbone.module.load_state_dict(dict_checkpoint["state_dict_backbone"])
+        module_partial_fc.load_state_dict(dict_checkpoint["state_dict_softmax_fc"])
+        opt.load_state_dict(dict_checkpoint["state_optimizer"])
+        lr_scheduler.load_state_dict(dict_checkpoint["state_lr_scheduler"])
+        del dict_checkpoint
+
+    for key, value in cfg.items():
+        num_space = 25 - len(key)
+        logging.info(": " + key + " " * num_space + str(value))
+
+    callback_verification = CallBackVerification(
+        val_targets=cfg.val_targets, rec_prefix=cfg.rec, 
+        summary_writer=summary_writer, wandb_logger = wandb_logger
+    )
+    callback_logging = CallBackLogging(
+        frequent=cfg.frequent,
+        total_step=cfg.total_step,
+        batch_size=cfg.batch_size,
+        start_step = global_step,
+        writer=summary_writer
+    )
+
+    loss_am = AverageMeter()
+    amp = torch.cuda.amp.grad_scaler.GradScaler(growth_interval=100)
+
+    for epoch in range(start_epoch, cfg.num_epoch):
+
+        if isinstance(train_loader, DataLoader):
+            train_loader.sampler.set_epoch(epoch)
+        for _, (img, local_labels) in enumerate(train_loader):
+            global_step += 1
+            local_embeddings = backbone(img)
+            loss: torch.Tensor = module_partial_fc(local_embeddings, local_labels)
+
+            if cfg.fp16:
+                amp.scale(loss).backward()
+                if global_step % cfg.gradient_acc == 0:
+                    amp.unscale_(opt)
+                    torch.nn.utils.clip_grad_norm_(backbone.parameters(), 5)
+                    amp.step(opt)
+                    amp.update()
+                    opt.zero_grad()
+            else:
+                loss.backward()
+                if global_step % cfg.gradient_acc == 0:
+                    torch.nn.utils.clip_grad_norm_(backbone.parameters(), 5)
+                    opt.step()
+                    opt.zero_grad()
+            lr_scheduler.step()
+
+            with torch.no_grad():
+                if wandb_logger:
+                    wandb_logger.log({
+                        'Loss/Step Loss': loss.item(),
+                        'Loss/Train Loss': loss_am.avg,
+                        'Process/Step': global_step,
+                        'Process/Epoch': epoch
+                    })
+                    
+                loss_am.update(loss.item(), 1)
+                callback_logging(global_step, loss_am, epoch, cfg.fp16, lr_scheduler.get_last_lr()[0], amp)
+
+                if global_step % cfg.verbose == 0 and global_step > 0:
+                    callback_verification(global_step, backbone)
+
+        if cfg.save_all_states:
+            checkpoint = {
+                "epoch": epoch + 1,
+                "global_step": global_step,
+                "state_dict_backbone": backbone.module.state_dict(),
+                "state_dict_softmax_fc": module_partial_fc.state_dict(),
+                "state_optimizer": opt.state_dict(),
+                "state_lr_scheduler": lr_scheduler.state_dict()
+            }
+            torch.save(checkpoint, os.path.join(cfg.output, f"checkpoint_gpu_{rank}.pt"))
+
+        if rank == 0:
+            path_module = os.path.join(cfg.output, "model.pt")
+            torch.save(backbone.module.state_dict(), path_module)
+
+            if wandb_logger and cfg.save_artifacts:
+                artifact_name = f"{run_name}_E{epoch}"
+                model = wandb.Artifact(artifact_name, type='model')
+                model.add_file(path_module)
+                wandb_logger.log_artifact(model)
+                
+        if cfg.dali:
+            train_loader.reset()
+
+    if rank == 0:
+        path_module = os.path.join(cfg.output, "model.pt")
+        torch.save(backbone.module.state_dict(), path_module)
+        
+        if wandb_logger and cfg.save_artifacts:
+            artifact_name = f"{run_name}_Final"
+            model = wandb.Artifact(artifact_name, type='model')
+            model.add_file(path_module)
+            wandb_logger.log_artifact(model)
+
+
+
+if __name__ == "__main__":
+    torch.backends.cudnn.benchmark = True
+    parser = argparse.ArgumentParser(
+        description="Distributed Arcface Training in Pytorch")
+    parser.add_argument("config", type=str, help="py config file")
+    main(parser.parse_args())
diff --git a/recognition/arcface_torch/utils/__init__.py b/recognition/arcface_torch/utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/arcface_torch/utils/plot.py b/recognition/arcface_torch/utils/plot.py
new file mode 100644
index 000000000..7f1d39da6
--- /dev/null
+++ b/recognition/arcface_torch/utils/plot.py
@@ -0,0 +1,71 @@
+import os
+import sys
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from sklearn.metrics import roc_curve, auc
+
+with open(sys.argv[1], "r") as f:
+    files = f.readlines()
+
+files = [x.strip() for x in files]
+image_path = "/train_tmp/IJB_release/IJBC"
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % 'ijbc'))
+
+methods = []
+scores = []
+for file in files:
+    methods.append(file)
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append(method)
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10 ** -6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+print(tpr_fpr_table)
diff --git a/recognition/arcface_torch/utils/utils_callbacks.py b/recognition/arcface_torch/utils/utils_callbacks.py
new file mode 100755
index 000000000..d9368073f
--- /dev/null
+++ b/recognition/arcface_torch/utils/utils_callbacks.py
@@ -0,0 +1,125 @@
+import logging
+import os
+import time
+from typing import List
+
+import torch
+
+from eval import verification
+from utils.utils_logging import AverageMeter
+from torch.utils.tensorboard import SummaryWriter
+from torch import distributed
+
+
+class CallBackVerification(object):
+    
+    def __init__(self, val_targets, rec_prefix, summary_writer=None, image_size=(112, 112), wandb_logger=None):
+        self.rank: int = distributed.get_rank()
+        self.highest_acc: float = 0.0
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        if self.rank is 0:
+            self.init_dataset(val_targets=val_targets, data_dir=rec_prefix, image_size=image_size)
+
+        self.summary_writer = summary_writer
+        self.wandb_logger = wandb_logger
+
+    def ver_test(self, backbone: torch.nn.Module, global_step: int):
+        results = []
+        for i in range(len(self.ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                self.ver_list[i], backbone, 10, 10)
+            logging.info('[%s][%d]XNorm: %f' % (self.ver_name_list[i], global_step, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (self.ver_name_list[i], global_step, acc2, std2))
+
+            self.summary_writer: SummaryWriter
+            self.summary_writer.add_scalar(tag=self.ver_name_list[i], scalar_value=acc2, global_step=global_step, )
+            if self.wandb_logger:
+                import wandb
+                self.wandb_logger.log({
+                    f'Acc/val-Acc1 {self.ver_name_list[i]}': acc1,
+                    f'Acc/val-Acc2 {self.ver_name_list[i]}': acc2,
+                    # f'Acc/val-std1 {self.ver_name_list[i]}': std1,
+                    # f'Acc/val-std2 {self.ver_name_list[i]}': acc2,
+                })
+
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info(
+                '[%s][%d]Accuracy-Highest: %1.5f' % (self.ver_name_list[i], global_step, self.highest_acc_list[i]))
+            results.append(acc2)
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+        for name in val_targets:
+            path = os.path.join(data_dir, name + ".bin")
+            if os.path.exists(path):
+                data_set = verification.load_bin(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+
+    def __call__(self, num_update, backbone: torch.nn.Module):
+        if self.rank is 0 and num_update > 0:
+            backbone.eval()
+            self.ver_test(backbone, num_update)
+            backbone.train()
+
+
+class CallBackLogging(object):
+    def __init__(self, frequent, total_step, batch_size, start_step=0,writer=None):
+        self.frequent: int = frequent
+        self.rank: int = distributed.get_rank()
+        self.world_size: int = distributed.get_world_size()
+        self.time_start = time.time()
+        self.total_step: int = total_step
+        self.start_step: int = start_step
+        self.batch_size: int = batch_size
+        self.writer = writer
+
+        self.init = False
+        self.tic = 0
+
+    def __call__(self,
+                 global_step: int,
+                 loss: AverageMeter,
+                 epoch: int,
+                 fp16: bool,
+                 learning_rate: float,
+                 grad_scaler: torch.cuda.amp.GradScaler):
+        if self.rank == 0 and global_step > 0 and global_step % self.frequent == 0:
+            if self.init:
+                try:
+                    speed: float = self.frequent * self.batch_size / (time.time() - self.tic)
+                    speed_total = speed * self.world_size
+                except ZeroDivisionError:
+                    speed_total = float('inf')
+
+                #time_now = (time.time() - self.time_start) / 3600
+                #time_total = time_now / ((global_step + 1) / self.total_step)
+                #time_for_end = time_total - time_now
+                time_now = time.time()
+                time_sec = int(time_now - self.time_start)
+                time_sec_avg = time_sec / (global_step - self.start_step + 1)
+                eta_sec = time_sec_avg * (self.total_step - global_step - 1)
+                time_for_end = eta_sec/3600
+                if self.writer is not None:
+                    self.writer.add_scalar('time_for_end', time_for_end, global_step)
+                    self.writer.add_scalar('learning_rate', learning_rate, global_step)
+                    self.writer.add_scalar('loss', loss.avg, global_step)
+                if fp16:
+                    msg = "Speed %.2f samples/sec   Loss %.4f   LearningRate %.6f   Epoch: %d   Global Step: %d   " \
+                          "Fp16 Grad Scale: %2.f   Required: %1.f hours" % (
+                              speed_total, loss.avg, learning_rate, epoch, global_step,
+                              grad_scaler.get_scale(), time_for_end
+                          )
+                else:
+                    msg = "Speed %.2f samples/sec   Loss %.4f   LearningRate %.6f   Epoch: %d   Global Step: %d   " \
+                          "Required: %1.f hours" % (
+                              speed_total, loss.avg, learning_rate, epoch, global_step, time_for_end
+                          )
+                logging.info(msg)
+                loss.reset()
+                self.tic = time.time()
+            else:
+                self.init = True
+                self.tic = time.time()
diff --git a/recognition/arcface_torch/utils/utils_config.py b/recognition/arcface_torch/utils/utils_config.py
new file mode 100644
index 000000000..0c02eaf70
--- /dev/null
+++ b/recognition/arcface_torch/utils/utils_config.py
@@ -0,0 +1,16 @@
+import importlib
+import os.path as osp
+
+
+def get_config(config_file):
+    assert config_file.startswith('configs/'), 'config file setting must start with configs/'
+    temp_config_name = osp.basename(config_file)
+    temp_module_name = osp.splitext(temp_config_name)[0]
+    config = importlib.import_module("configs.base")
+    cfg = config.config
+    config = importlib.import_module("configs.%s" % temp_module_name)
+    job_cfg = config.config
+    cfg.update(job_cfg)
+    if cfg.output is None:
+        cfg.output = osp.join('work_dirs', temp_module_name)
+    return cfg
\ No newline at end of file
diff --git a/recognition/arcface_torch/utils/utils_distributed_sampler.py b/recognition/arcface_torch/utils/utils_distributed_sampler.py
new file mode 100644
index 000000000..cea670396
--- /dev/null
+++ b/recognition/arcface_torch/utils/utils_distributed_sampler.py
@@ -0,0 +1,126 @@
+import math
+import os
+import random
+
+import numpy as np
+import torch
+import torch.distributed as dist
+from torch.utils.data import DistributedSampler as _DistributedSampler
+
+
+def setup_seed(seed, cuda_deterministic=True):
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    os.environ["PYTHONHASHSEED"] = str(seed)
+    if cuda_deterministic:  # slower, more reproducible
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+    else:  # faster, less reproducible
+        torch.backends.cudnn.deterministic = False
+        torch.backends.cudnn.benchmark = True
+
+
+def worker_init_fn(worker_id, num_workers, rank, seed):
+    # The seed of each worker equals to
+    # num_worker * rank + worker_id + user_seed
+    worker_seed = num_workers * rank + worker_id + seed
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+    torch.manual_seed(worker_seed)
+
+
+def get_dist_info():
+    if dist.is_available() and dist.is_initialized():
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    else:
+        rank = 0
+        world_size = 1
+
+    return rank, world_size
+
+
+def sync_random_seed(seed=None, device="cuda"):
+    """Make sure different ranks share the same seed.
+    All workers must call this function, otherwise it will deadlock.
+    This method is generally used in `DistributedSampler`,
+    because the seed should be identical across all processes
+    in the distributed group.
+    In distributed sampling, different ranks should sample non-overlapped
+    data in the dataset. Therefore, this function is used to make sure that
+    each rank shuffles the data indices in the same order based
+    on the same seed. Then different ranks could use different indices
+    to select non-overlapped data from the same data list.
+    Args:
+        seed (int, Optional): The seed. Default to None.
+        device (str): The device where the seed will be put on.
+            Default to 'cuda'.
+    Returns:
+        int: Seed to be used.
+    """
+    if seed is None:
+        seed = np.random.randint(2**31)
+    assert isinstance(seed, int)
+
+    rank, world_size = get_dist_info()
+
+    if world_size == 1:
+        return seed
+
+    if rank == 0:
+        random_num = torch.tensor(seed, dtype=torch.int32, device=device)
+    else:
+        random_num = torch.tensor(0, dtype=torch.int32, device=device)
+
+    dist.broadcast(random_num, src=0)
+
+    return random_num.item()
+
+
+class DistributedSampler(_DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        num_replicas=None,  # world_size
+        rank=None,  # local_rank
+        shuffle=True,
+        seed=0,
+    ):
+
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+
+        # In distributed sampling, different ranks should sample
+        # non-overlapped data in the dataset. Therefore, this function
+        # is used to make sure that each rank shuffles the data indices
+        # in the same order based on the same seed. Then different ranks
+        # could use different indices to select non-overlapped data from the
+        # same data list.
+        self.seed = sync_random_seed(seed)
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        if self.shuffle:
+            g = torch.Generator()
+            # When :attr:`shuffle=True`, this ensures all replicas
+            # use a different random ordering for each epoch.
+            # Otherwise, the next iteration of this sampler will
+            # yield the same ordering.
+            g.manual_seed(self.epoch + self.seed)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = torch.arange(len(self.dataset)).tolist()
+
+        # add extra samples to make it evenly divisible
+        # in case that indices is shorter than half of total_size
+        indices = (indices * math.ceil(self.total_size / len(indices)))[
+            : self.total_size
+        ]
+        assert len(indices) == self.total_size
+
+        # subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        return iter(indices)
diff --git a/recognition/arcface_torch/utils/utils_logging.py b/recognition/arcface_torch/utils/utils_logging.py
new file mode 100644
index 000000000..c787b6aae
--- /dev/null
+++ b/recognition/arcface_torch/utils/utils_logging.py
@@ -0,0 +1,41 @@
+import logging
+import os
+import sys
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value
+    """
+
+    def __init__(self):
+        self.val = None
+        self.avg = None
+        self.sum = None
+        self.count = None
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def init_logging(rank, models_root):
+    if rank == 0:
+        log_root = logging.getLogger()
+        log_root.setLevel(logging.INFO)
+        formatter = logging.Formatter("Training: %(asctime)s-%(message)s")
+        handler_file = logging.FileHandler(os.path.join(models_root, "training.log"))
+        handler_stream = logging.StreamHandler(sys.stdout)
+        handler_file.setFormatter(formatter)
+        handler_stream.setFormatter(formatter)
+        log_root.addHandler(handler_file)
+        log_root.addHandler(handler_stream)
+        log_root.info('rank_id: %d' % rank)
diff --git a/recognition/common/build_eval_pack.py b/recognition/common/build_eval_pack.py
deleted file mode 100644
index 49d0b746b..000000000
--- a/recognition/common/build_eval_pack.py
+++ /dev/null
@@ -1,124 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-#import mxnet as mx
-#from mxnet import ndarray as nd
-import argparse
-import cv2
-import pickle
-import numpy as np
-import sys
-import os
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', '..', 'common'))
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', '..', 'RetinaFace'))
-import face_align
-from retinaface import RetinaFace
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-parser = argparse.ArgumentParser(description='Package eval images')
-# general
-parser.add_argument('--data-dir', default='', help='')
-parser.add_argument('--image-size', type=int, default=112, help='')
-parser.add_argument('--gpu', type=int, default=0, help='')
-parser.add_argument('--det-prefix', type=str, default='./model/R50', help='')
-parser.add_argument('--output', default='./', help='path to save.')
-parser.add_argument('--align-mode', default='arcface', help='align mode.')
-args = parser.parse_args()
-
-gpu_id = args.gpu
-
-detector = RetinaFace(args.det_prefix, 0, gpu_id, network='net3')
-target_size = 400
-max_size = 800
-
-def get_norm_crop(image_path):
-  im = cv2.imread(image_path)
-  im_shape = im.shape
-  im_size_min = np.min(im_shape[0:2])
-  im_size_max = np.max(im_shape[0:2])
-  im_scale = float(target_size) / float(im_size_min)
-  # prevent bigger axis from being more than max_size:
-  if np.round(im_scale * im_size_max) > max_size:
-      im_scale = float(max_size) / float(im_size_max)
-  bbox, landmark = detector.detect(im, threshold=0.5, scales=[im_scale])
-  #print(im.shape, bbox.shape, landmark.shape)
-  if bbox.shape[0]==0:
-    bbox, landmark = detector.detect(im, threshold=0.05, scales=[im_scale*0.75, im_scale, im_scale*2.0])
-    print('refine', im.shape, bbox.shape, landmark.shape)
-  nrof_faces = bbox.shape[0]
-  if nrof_faces>0:
-    det = bbox[:,0:4]
-    img_size = np.asarray(im.shape)[0:2]
-    bindex = 0
-    if nrof_faces>1:
-      bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-      img_center = img_size / 2
-      offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-      offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-      bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-    #_bbox = bounding_boxes[bindex, 0:4]
-    _landmark = landmark[bindex]
-    warped = face_align.norm_crop(im, landmark = _landmark, image_size=args.image_size, mode=args.align_mode)
-    return warped
-  else:
-    return None
-
-
-bins = []
-issame_list = []
-pp = 0
-for line in open(os.path.join(args.data_dir, 'pairs_label.txt'), 'r'):
-  pp+=1
-  if pp%100==0:
-    print('processing', pp)
-  line = line.strip().split()
-  assert len(line)==3
-  path1 = os.path.join(args.data_dir, line[0])
-  path2 = os.path.join(args.data_dir, line[1])
-  im1 = get_norm_crop(path1)
-  im2 = get_norm_crop(path2)
-  issame = True
-  if line[2]=='0':
-    issame = False
-  issame_list.append(issame)
-  for im in [im1, im2]:
-    _, s = cv2.imencode('.jpg', im)
-    bins.append(s)
-
-with open(args.output, 'wb') as f:
-  pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
diff --git a/recognition/common/flops_counter.py b/recognition/common/flops_counter.py
deleted file mode 100644
index 64469ec85..000000000
--- a/recognition/common/flops_counter.py
+++ /dev/null
@@ -1,114 +0,0 @@
-
-'''
-@author: insightface
-'''
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import sys
-import os
-import json
-import argparse
-import numpy as np
-import mxnet as mx
-
-
-def is_no_bias(attr):
-  ret = False
-  if 'no_bias' in attr and (attr['no_bias']==True or attr['no_bias']=='True'):
-    ret = True
-  return ret
-
-def count_fc_flops(input_filter, output_filter, attr):
-  #print(input_filter, output_filter ,attr)
-  ret = 2*input_filter*output_filter
-  if is_no_bias(attr):
-    ret -= output_filter
-  return int(ret)
-
-
-def count_conv_flops(input_shape, output_shape, attr):
-  kernel = attr['kernel'][1:-1].split(',')
-  kernel = [int(x) for x in kernel]
-
-  #print('kernel', kernel)
-  if is_no_bias(attr):
-    ret = (2*input_shape[1]*kernel[0]*kernel[1]-1)*output_shape[2]*output_shape[3]*output_shape[1]
-  else:
-    ret = 2*input_shape[1]*kernel[0]*kernel[1]*output_shape[2]*output_shape[3]*output_shape[1]
-  num_group = 1
-  if 'num_group' in attr:
-    num_group = int(attr['num_group'])
-  ret /= num_group
-  return int(ret)
-
-
-def count_flops(sym, **data_shapes):
-  all_layers = sym.get_internals()
-  #print(all_layers)
-  arg_shapes, out_shapes, aux_shapes = all_layers.infer_shape(**data_shapes)
-  out_shape_dict = dict(zip(all_layers.list_outputs(), out_shapes))
-
-  nodes = json.loads(sym.tojson())['nodes']
-  nodeid_shape = {}
-  for nodeid, node in enumerate(nodes):
-    name = node['name']
-    layer_name = name+"_output"
-    if layer_name in out_shape_dict:
-      nodeid_shape[nodeid] = out_shape_dict[layer_name]
-  #print(nodeid_shape)
-  FLOPs = 0
-  for nodeid, node in enumerate(nodes):
-    flops = 0
-    if node['op']=='Convolution':
-      output_shape = nodeid_shape[nodeid]
-      name = node['name']
-      attr = node['attrs']
-      input_nodeid = node['inputs'][0][0]
-      input_shape = nodeid_shape[input_nodeid]
-      flops = count_conv_flops(input_shape, output_shape, attr)
-    elif node['op']=='FullyConnected':
-      attr = node['attrs']
-      output_shape = nodeid_shape[nodeid]
-      input_nodeid = node['inputs'][0][0]
-      input_shape = nodeid_shape[input_nodeid]
-      output_filter = output_shape[1]
-      input_filter = input_shape[1]*input_shape[2]*input_shape[3]
-      #assert len(input_shape)==4 and input_shape[2]==1 and input_shape[3]==1
-      flops = count_fc_flops(input_filter, output_filter, attr)
-    #print(node, flops)
-    FLOPs += flops
-
-  return FLOPs
-
-def flops_str(FLOPs):
-  preset = [ (1e12, 'T'), (1e9, 'G'), (1e6, 'M'), (1e3, 'K') ]
-
-  for p in preset:
-    if FLOPs//p[0]>0:
-      N = FLOPs/p[0]
-      ret = "%.1f%s"%(N, p[1])
-      return ret
-  ret = "%.1f"%(FLOPs)
-  return ret
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='flops counter')
-  # general
-  #parser.add_argument('--model', default='../models2/y2-arcface-retinat1/model,1', help='path to load model.')
-  #parser.add_argument('--model', default='../models2/r100fc-arcface-retinaa/model,1', help='path to load model.')
-  parser.add_argument('--model', default='../models2/r50fc-arcface-emore/model,1', help='path to load model.')
-  args = parser.parse_args()
-  _vec = args.model.split(',')
-  assert len(_vec)==2
-  prefix = _vec[0]
-  epoch = int(_vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  FLOPs = count_flops(sym, data=(1,3,112,112))
-  print('FLOPs:', FLOPs)
-
diff --git a/recognition/common/rec2image.py b/recognition/common/rec2image.py
deleted file mode 100644
index 535df2478..000000000
--- a/recognition/common/rec2image.py
+++ /dev/null
@@ -1,63 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-import numpy as np
-
-
-def main(args):
-  include_datasets = args.include.split(',')
-  rec_list = []
-  for ds in include_datasets:
-    path_imgrec = os.path.join(ds, 'train.rec')
-    path_imgidx = os.path.join(ds, 'train.idx')
-    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-    rec_list.append(imgrec)
-  if not os.path.exists(args.output):
-    os.makedirs(args.output)
-  for ds_id in range(len(rec_list)):
-    id_list = []
-    imgrec = rec_list[ds_id]
-    s = imgrec.read_idx(0)
-    header, _ = mx.recordio.unpack(s)
-    assert header.flag>0
-    print('header0 label', header.label)
-    header0 = (int(header.label[0]), int(header.label[1]))
-    seq_identity = range(int(header.label[0]), int(header.label[1]))
-    pp=0
-    for identity in seq_identity:
-      id_dir = os.path.join(args.output, "%d_%d"%(ds_id, identity))
-      os.makedirs(id_dir)
-      pp+=1
-      if pp%10==0:
-        print('processing id', pp)
-      s = imgrec.read_idx(identity)
-      header, _ = mx.recordio.unpack(s)
-      imgid = 0
-      for _idx in range(int(header.label[0]), int(header.label[1])):
-        s = imgrec.read_idx(_idx)
-        _header, _img = mx.recordio.unpack(s)
-        _img = mx.image.imdecode(_img).asnumpy()[:,:,::-1] # to bgr
-        image_path = os.path.join(id_dir, "%d.jpg"%imgid)
-        cv2.imwrite(image_path, _img)
-        imgid+=1
-
-
-
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do dataset merge')
-  # general
-  parser.add_argument('--include', default='', type=str, help='')
-  parser.add_argument('--output', default='', type=str, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/recognition/common/rec_builder.py b/recognition/common/rec_builder.py
deleted file mode 100644
index 6250876e3..000000000
--- a/recognition/common/rec_builder.py
+++ /dev/null
@@ -1,102 +0,0 @@
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-import numpy as np
-
-
-class SeqRecBuilder():
-    def __init__(self, path, image_size=(112, 112)):
-        self.path = path
-        self.image_size = image_size
-        self.last_label = -1
-        self.widx = 0
-        if not os.path.exists(path):
-            os.makedirs(path)
-        self.writer = mx.recordio.MXIndexedRecordIO(os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'), 'w')
-        self.label_stat = [-1, -1]
-
-    def add(self, label, img, is_image = True):
-        #img should be BGR
-        #if self.sis:
-        #    assert label>=self.last_label
-        idx = self.widx
-        self.widx += 1
-        header = mx.recordio.IRHeader(0, label, idx, 0)
-        if is_image:
-            s = mx.recordio.pack_img(header, img, quality=95, img_fmt='.jpg')
-        else:
-            s = mx.recordio.pack(header, img)
-        self.writer.write_idx(idx, s)
-        if self.label_stat[0]<0:
-            self.label_stat = [label, label]
-        else:
-            self.label_stat[0] = min(self.label_stat[0], label)
-            self.label_stat[1] = max(self.label_stat[1], label)
-
-
-    def close(self):
-        with open(os.path.join(self.path, 'property'), 'w') as f:
-            f.write("%d,%d,%d\n"%(self.label_stat[1]+1, self.image_size[0], self.image_size[1]))
-
-
-class RecBuilder():
-    def __init__(self, path, image_size=(112, 112)):
-        self.path = path
-        self.image_size = image_size
-        self.last_label = -1
-        self.widx = 1
-        if not os.path.exists(path):
-            os.makedirs(path)
-        self.writer = mx.recordio.MXIndexedRecordIO(os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'), 'w')
-        self.label_stat = [-1, -1]
-        self.identities = []
-
-    def add(self, label, imgs):
-        #img should be BGR
-        assert label>=0
-        assert label>self.last_label
-        assert len(imgs)>0
-        idflag = [self.widx, -1]
-        for img in imgs:
-            idx = self.widx
-            self.widx += 1
-            header = mx.recordio.IRHeader(0, label, idx, 0)
-            if isinstance(img, np.ndarray):
-                s = mx.recordio.pack_img(header, img, quality=95, img_fmt='.jpg')
-            else:
-                s = mx.recordio.pack(header, img)
-            self.writer.write_idx(idx, s)
-        idflag[1] = self.widx
-        self.identities.append(idflag)
-        if self.label_stat[0]<0:
-            self.label_stat = [label, label]
-        else:
-            self.label_stat[0] = min(self.label_stat[0], label)
-            self.label_stat[1] = max(self.label_stat[1], label)
-        self.last_label = label
-
-    def close(self):
-        id_idx = self.widx
-        for id_flag in self.identities:
-            idx = self.widx
-            self.widx += 1
-            _header = mx.recordio.IRHeader(0, id_flag, idx, 0)
-            s = mx.recordio.pack(_header, b'')
-            self.writer.write_idx(idx, s)
-
-        print('id0:', (id_idx, self.widx))
-        idx = 0
-        _header = mx.recordio.IRHeader(0, (id_idx, self.widx), idx, 1)
-        s = mx.recordio.pack(_header, b'')
-        self.writer.write_idx(idx, s)
-        print('label stat:', self.label_stat)
-
-        with open(os.path.join(self.path, 'property'), 'w') as f:
-            f.write("%d,%d,%d\n"%(self.label_stat[1]+1, self.image_size[0], self.image_size[1]))
-
diff --git a/recognition/common/verification.py b/recognition/common/verification.py
deleted file mode 100644
index 8972e506e..000000000
--- a/recognition/common/verification.py
+++ /dev/null
@@ -1,374 +0,0 @@
-"""Helper for evaluation on the Labeled Faces in the Wild dataset 
-"""
-
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import argparse
-import sys
-import numpy as np
-from scipy import misc
-from sklearn.model_selection import KFold
-from scipy import interpolate
-import sklearn
-import cv2
-import math
-import datetime
-import pickle
-from sklearn.decomposition import PCA
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-class LFold:
-  def __init__(self, n_splits = 2, shuffle = False):
-    self.n_splits = n_splits
-    if self.n_splits>1:
-      self.k_fold = KFold(n_splits = n_splits, shuffle = shuffle)
-
-  def split(self, indices):
-    if self.n_splits>1:
-      return self.k_fold.split(indices)
-    else:
-      return [(indices, indices)]
-
-
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca = 0):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    tprs = np.zeros((nrof_folds,nrof_thresholds))
-    fprs = np.zeros((nrof_folds,nrof_thresholds))
-    accuracy = np.zeros((nrof_folds))
-    indices = np.arange(nrof_pairs)
-    #print('pca', pca)
-    
-    if pca==0:
-      diff = np.subtract(embeddings1, embeddings2)
-      dist = np.sum(np.square(diff),1)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-        #print('train_set', train_set)
-        #print('test_set', test_set)
-        if pca>0:
-          print('doing pca on', fold_idx)
-          embed1_train = embeddings1[train_set]
-          embed2_train = embeddings2[train_set]
-          _embed_train = np.concatenate( (embed1_train, embed2_train), axis=0 )
-          #print(_embed_train.shape)
-          pca_model = PCA(n_components=pca)
-          pca_model.fit(_embed_train)
-          embed1 = pca_model.transform(embeddings1)
-          embed2 = pca_model.transform(embeddings2)
-          embed1 = sklearn.preprocessing.normalize(embed1)
-          embed2 = sklearn.preprocessing.normalize(embed2)
-          #print(embed1.shape, embed2.shape)
-          diff = np.subtract(embed1, embed2)
-          dist = np.sum(np.square(diff),1)
-        
-        # Find the best threshold for the fold
-        acc_train = np.zeros((nrof_thresholds))
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
-        best_threshold_index = np.argmax(acc_train)
-        #print('threshold', thresholds[best_threshold_index])
-        for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-          
-    tpr = np.mean(tprs,0)
-    fpr = np.mean(fprs,0)
-    return tpr, fpr, accuracy
-
-def calculate_accuracy(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    tp = np.sum(np.logical_and(predict_issame, actual_issame))
-    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
-    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
-  
-    tpr = 0 if (tp+fn==0) else float(tp) / float(tp+fn)
-    fpr = 0 if (fp+tn==0) else float(fp) / float(fp+tn)
-    acc = float(tp+tn)/dist.size
-    return tpr, fpr, acc
-
-
-  
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    val = np.zeros(nrof_folds)
-    far = np.zeros(nrof_folds)
-    
-    diff = np.subtract(embeddings1, embeddings2)
-    dist = np.sum(np.square(diff),1)
-    indices = np.arange(nrof_pairs)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
-        # Find the threshold that gives FAR = far_target
-        far_train = np.zeros(nrof_thresholds)
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
-        if np.max(far_train)>=far_target:
-            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
-            threshold = f(far_target)
-        else:
-            threshold = 0.0
-    
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
-  
-    val_mean = np.mean(val)
-    far_mean = np.mean(far)
-    val_std = np.std(val)
-    return val_mean, val_std, far_mean
-
-
-def calculate_val_far(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    n_same = np.sum(actual_issame)
-    n_diff = np.sum(np.logical_not(actual_issame))
-    #print(true_accept, false_accept)
-    #print(n_same, n_diff)
-    val = float(true_accept) / float(n_same)
-    far = float(false_accept) / float(n_diff)
-    return val, far
-
-def evaluate(embeddings, actual_issame, nrof_folds=10, pca = 0):
-    # Calculate evaluation metrics
-    thresholds = np.arange(0, 4, 0.01)
-    embeddings1 = embeddings[0::2]
-    embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), nrof_folds=nrof_folds, pca = pca)
-    thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
-    return tpr, fpr, accuracy, val, val_std, far
-
-def load_bin(path, image_size):
-  try:
-    with open(path, 'rb') as f:
-      bins, issame_list = pickle.load(f) #py2
-  except UnicodeDecodeError as e:
-    with open(path, 'rb') as f:
-      bins, issame_list = pickle.load(f, encoding='bytes') #py3
-  data_list = []
-  for flip in [0,1]:
-    data = nd.empty((len(issame_list)*2, 3, image_size[0], image_size[1]))
-    data_list.append(data)
-  for i in range(len(issame_list)*2):
-    _bin = bins[i]
-    img = mx.image.imdecode(_bin)
-    if img.shape[1]!=image_size[0]:
-      img = mx.image.resize_short(img, image_size[0])
-    img = nd.transpose(img, axes=(2, 0, 1))
-    for flip in [0,1]:
-      if flip==1:
-        img = mx.ndarray.flip(data=img, axis=2)
-      data_list[flip][i][:] = img
-    if i%1000==0:
-      print('loading bin', i)
-  print(data_list[0].shape)
-  return (data_list, issame_list)
-
-def test(data_set, mx_model, batch_size, nfolds=10, data_extra = None, label_shape = None):
-  print('testing verification..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in range( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      #_arg, _aux = model.get_params()
-      #__arg = {}
-      #for k,v in _arg.iteritems():
-      #  __arg[k] = v.as_in_context(_ctx)
-      #_arg = __arg
-      #_arg["data"] = _data.as_in_context(_ctx)
-      #_arg["softmax_label"] = _label.as_in_context(_ctx)
-      #for k,v in _arg.iteritems():
-      #  print(k,v.context)
-      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
-      #exe.forward(is_train=False)
-      #net_out = exe.outputs
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      #print(_embeddings.shape)
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-
-  _xnorm = 0.0
-  _xnorm_cnt = 0
-  for embed in embeddings_list:
-    for i in range(embed.shape[0]):
-      _em = embed[i]
-      _norm=np.linalg.norm(_em)
-      #print(_em.shape, _norm)
-      _xnorm+=_norm
-      _xnorm_cnt+=1
-  _xnorm /= _xnorm_cnt
-
-  embeddings = embeddings_list[0].copy()
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  acc1 = 0.0
-  std1 = 0.0
-  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  #acc1, std1 = np.mean(accuracy), np.std(accuracy)
-
-  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-  #embeddings = np.concatenate(embeddings_list, axis=1)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  print(embeddings.shape)
-  print('infer time', time_consumed)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  return acc1, std1, acc2, std2, _xnorm, embeddings_list
-
-if __name__ == '__main__':
-
-  parser = argparse.ArgumentParser(description='do verification')
-  # general
-  parser.add_argument('--data-dir', default='', help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--target', default='lfw,cfp_ff,cfp_fp,agedb_30', help='test targets.')
-  parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--max', default='', type=str, help='')
-  parser.add_argument('--mode', default=0, type=int, help='')
-  parser.add_argument('--nfolds', default=10, type=int, help='')
-  args = parser.parse_args()
-  #sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-  #import face_image
-  #prop = face_image.load_property(args.data_dir)
-  #image_size = prop.image_size
-  image_size = [112,112]
-  print('image_size', image_size)
-  ctx = mx.gpu(args.gpu)
-  nets = []
-  vec = args.model.split(',')
-  prefix = args.model.split(',')[0]
-  epochs = []
-  if len(vec)==1:
-    pdir = os.path.dirname(prefix)
-    for fname in os.listdir(pdir):
-      if not fname.endswith('.params'):
-        continue
-      _file = os.path.join(pdir, fname)
-      if _file.startswith(prefix):
-        epoch = int(fname.split('.')[0].split('-')[1])
-        epochs.append(epoch)
-    epochs = sorted(epochs, reverse=True)
-    if len(args.max)>0:
-      _max = [int(x) for x in args.max.split(',')]
-      assert len(_max)==2
-      if len(epochs)>_max[1]:
-        epochs = epochs[_max[0]:_max[1]]
-
-  else:
-    epochs = [int(x) for x in vec[1].split('|')]
-  print('model number', len(epochs))
-  time0 = datetime.datetime.now()
-  for epoch in epochs:
-    print('loading',prefix, epoch)
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    nets.append(model)
-  time_now = datetime.datetime.now()
-  diff = time_now - time0
-  print('model loading time', diff.total_seconds())
-
-  ver_list = []
-  ver_name_list = []
-  for name in args.target.split(','):
-    path = os.path.join(args.data_dir,name+".bin")
-    if os.path.exists(path):
-      print('loading.. ', name)
-      data_set = load_bin(path, image_size)
-      ver_list.append(data_set)
-      ver_name_list.append(name)
-
-  if args.mode==0:
-    for i in range(len(ver_list)):
-      results = []
-      for model in nets:
-        acc1, std1, acc2, std2, xnorm, embeddings_list = test(ver_list[i], model, args.batch_size, args.nfolds)
-        print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
-        print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
-        print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
-        results.append(acc2)
-      print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
-  elif args.mode==1:
-    model = nets[0]
-    test_badcase(ver_list[0], model, args.batch_size, args.target)
-  else:
-    model = nets[0]
-    dumpR(ver_list[0], model, args.batch_size, args.target)
-
-
diff --git a/recognition/idmmd/README.md b/recognition/idmmd/README.md
new file mode 100644
index 000000000..4fce9b910
--- /dev/null
+++ b/recognition/idmmd/README.md
@@ -0,0 +1,55 @@
+# Physical-based Rendering for NIR-VIS Face Recognition
+
+by [Yunqi Miao*](https://yoqim.github.io/), [Alexandros Lattas*](https://alexlattas.com/), [Jiankang Deng](https://jiankangdeng.github.io/), [Jungong Han](https://jungonghan.github.io/), and [Stefanos Zafeiriou](https://wp.doc.ic.ac.uk/szafeiri/).
+
+
+For more information, please check our
+
+**[[Arxiv]](https://arxiv.org/abs/2211.06408)**
+**[[NeurIPS 2022 Paper]](https://openreview.net/pdf?id=NN_TpS5dpo5)**
+
+
+If you find this project useful in your research, please consider citing:
+
+```
+@inproceedings{miao2022physically,
+  title={Physically-Based Face Rendering for NIR-VIS Face Recognition},
+  author={Miao, Yunqi and Lattas, Alexandros and Deng, Jiankang and Han, Jungong and Zafeiriou, Stefanos},
+  booktitle={NeurIPS 2022},
+  year={2022}
+}
+```
+
+# Overview
+![poster](https://github.com/yoqim/PR-HFR/blob/main/pics/Poster.png)
+
+# Training
+
+For this project, we used python 3.7.10.
+
+## How to run?
+
+```shell
+sh run.sh
+```
+
+
+# Testing
+## Preparation
+- Downloading data (112 x 112) from [[Google drive]](https://drive.google.com/file/d/1Smd-Bdwj4tCbNugmoa66vxnJAU613bCo/view?usp=sharing)
+   - Put data to `data/$dataset_name` 
+   - Copy data list from [here](https://github.com/yoqim/PR-HFR/tree/main/data)
+
+>Note that: casia(fold_1) is provided for research purposes only. For the rest data, please refer to the original publications.
+
+
+- Downloading models from [[Google drive]](https://drive.google.com/file/d/1XjlnvbXmRD5xLJo7lLTy8LyQbMYRoz8C/view?usp=sharing)
+    - Put pretrain model at `models/pretrain/`
+    - Put finetune model at `models/finetune/$dataset/`
+
+
+## How to run?
+
+```shell
+sh eval.sh
+```
diff --git a/recognition/idmmd/dataset_mix.py b/recognition/idmmd/dataset_mix.py
new file mode 100644
index 000000000..9f6ebb73a
--- /dev/null
+++ b/recognition/idmmd/dataset_mix.py
@@ -0,0 +1,348 @@
+from PIL import Image
+import os
+import torch.utils.data as data
+import torchvision.transforms as transforms
+
+
+class Real_Dataset_112(data.Dataset):
+    def __init__(self, args):
+        super(Real_Dataset_112, self).__init__()
+        
+        self.img_root = args.img_root_R
+        self.img_list, self.num_classes = self.list_reader(args.train_list_R)
+        self.input_mode = args.input_mode
+
+        self.transform = transforms.Compose([
+            # transforms.RandomCrop(112),
+            transforms.ToTensor()
+        ])
+
+    def __getitem__(self, index):
+        img_name, label = self.img_list[index]
+
+        img = self.get_img_from_path(img_name)
+        return {'img': img, 'label': int(label)}
+
+    def __len__(self):
+        return len(self.img_list)
+
+    def get_img_from_path(self, img_name):
+        img_path = os.path.join(self.img_root, img_name)
+
+        if self.input_mode == 'grey':
+            img = Image.open(img_path).convert('L')
+        elif self.input_mode == 'red':
+            img = Image.open(img_path)
+            img = img.split()[0]
+
+        img = self.transform(img)
+        return img
+
+    def list_reader(self, list_file):
+        img_list = []
+        with open(list_file, 'r') as f:
+            lines = f.readlines()
+        
+        pid_container = set()
+        for line in lines:
+            pid = int(line.strip().split(' ')[1])
+            pid_container.add(pid)
+        pid2label = {pid:label for label, pid in enumerate(pid_container)}
+
+        for line in lines:
+            img_name, pid = line.strip().split(' ')
+            if not os.path.exists(os.path.join(self.img_root, img_name)):
+                continue
+            label = pid2label[int(pid)]
+            img_list.append((img_name, label))
+        
+        return img_list, len(pid_container)
+
+
+class Real_Dataset_112_paired(data.Dataset):
+    def __init__(self, args):
+        super(Real_Dataset_112_paired, self).__init__()
+
+        self.img_root = args.img_root_R
+        self.img_list, self.num_classes = self.list_reader(args.train_list_R)
+        self.input_mode = args.input_mode
+
+        self.transform = transforms.Compose([
+            # transforms.RandomCrop(112),
+            transforms.ToTensor()
+        ])
+
+        self.vir_list = [(a,b,c) for (a,b,c) in self.img_list if c==0]
+        self.nir_list = [(a,b,c) for (a,b,c) in self.img_list if c==1]
+
+        self.vis_labels = np.array([p[1] for p in self.vir_list])
+        self.nir_labels = np.array([p[1] for p in self.nir_list])
+
+        self.visIndex = None
+        self.nirIndex = None
+
+    def __getitem__(self, index):
+        vis_img_name, vis_label, vis_domain = self.vir_list[self.visIndex[index]]
+        nir_img_name, nir_label, nir_domain = self.nir_list[self.nirIndex[index]]
+        
+        assert vis_domain == 0 and nir_domain == 1
+
+        vis_img = self.get_img_from_path(vis_img_name)
+        nir_img = self.get_img_from_path(nir_img_name)
+
+        return vis_img, nir_img, vis_label, nir_label
+
+    def __len__(self):
+        return len(self.img_list)
+
+    def get_img_from_path(self, img_name):
+        img_path = os.path.join(self.img_root, img_name)
+
+        if self.input_mode == 'grey':
+            img = Image.open(img_path).convert('L')
+        elif self.input_mode == 'red':
+            img = Image.open(img_path)
+            img = img.split()[0]
+
+        img = self.transform(img)
+        return img
+
+
+    def list_reader(self, list_file):
+        img_list = []
+        with open(list_file, 'r') as f:
+            lines = f.readlines()
+        
+        pid_container = set()
+        for line in lines:
+            pid = int(line.strip().split(' ')[1])
+            pid_container.add(pid)
+        pid2label = {pid:label for label, pid in enumerate(pid_container)}
+
+        for line in lines:
+            img_name, pid = line.strip().split(' ')
+            label = pid2label[int(pid)]
+    
+            domain = 0 if 'VIS' in img_name else 1 
+            img_list.append((img_name, label, domain))
+        
+        return img_list, len(pid_container)
+
+class Mix_Dataset_112(data.Dataset):
+    def __init__(self, args):
+        super(Mix_Dataset_112, self).__init__()
+        
+        self.img_root_R = args.img_root_R
+        self.img_root_F = args.img_root_F
+        self.img_list, self.num_classes = self.list_reader(args.train_list_R, args.train_list_F)
+        self.input_mode = args.input_mode
+
+        self.transform = transforms.Compose([
+            # transforms.RandomCrop(112),
+            transforms.ToTensor()
+        ])
+
+    def __getitem__(self, index):
+        img_path, label = self.img_list[index]
+
+        img = self.get_img_from_path(img_path)
+        return {'img': img, 'label': int(label)}
+
+    def __len__(self):
+        return len(self.img_list)
+
+    def get_img_from_path(self, img_path):
+
+        if self.input_mode == 'grey':
+            img = Image.open(img_path).convert('L')
+        elif self.input_mode == 'red':
+            img = Image.open(img_path)
+            img = img.split()[0]
+
+        img = self.transform(img)
+        return img
+
+    def list_reader(self, list_file_real, list_file_fake):
+        with open(list_file_real, 'r') as f:
+            lines_real = f.readlines()
+        with open(list_file_fake, 'r') as f:
+            lines_fake = f.readlines()
+
+        fake_label_start = max([int(l.strip().split(' ')[-1]) for l in lines_real]) + 1
+        lines_fake = ["{} {}".format(l.strip().split(' ')[0], int(l.strip().split(' ')[1]) + fake_label_start) for l in lines_fake]
+
+        lines = lines_real + lines_fake
+
+        pid_container = set()
+        for line in lines:
+            pid = int(line.strip().split(' ')[1])
+            pid_container.add(pid)
+        pid2label = {pid:label for label, pid in enumerate(pid_container)}
+
+        img_list_R = []
+        for line in lines_real:
+            img_name, pid = line.strip().split(' ')
+            label = pid2label[int(pid)]
+            img_list_R.append((os.path.join(self.img_root_R + img_name), label))
+        
+        img_list_F = []
+        for line in lines_fake:
+            img_name, pid = line.strip().split(' ')
+            label = pid2label[int(pid)]
+            img_list_F.append((os.path.join(self.img_root_F + img_name), label))
+        
+        img_list = img_list_R + img_list_F
+
+        return img_list, len(pid_container)
+
+
+class Mix_Dataset_112_paired(data.Dataset):
+    def __init__(self, args):
+        super(Mix_Dataset_112_paired, self).__init__()
+
+        self.img_root_R = args.img_root_R
+        self.img_root_F = args.img_root_F
+        self.img_list, self.num_classes = self.list_reader(args.train_list_R, args.train_list_F)
+        self.input_mode = args.input_mode
+
+        self.transform = transforms.Compose([
+            # transforms.RandomCrop(112),
+            transforms.ToTensor()
+        ])
+
+        self.vir_list = [(a,b,c) for (a,b,c) in self.img_list if c==0]
+        self.nir_list = [(a,b,c) for (a,b,c) in self.img_list if c==1]
+
+        self.vis_labels = np.array([p[1] for p in self.vir_list])
+        self.nir_labels = np.array([p[1] for p in self.nir_list])
+
+        self.visIndex = None
+        self.nirIndex = None
+
+    def __getitem__(self, index):
+        vis_img_name, vis_label, vis_domain = self.vir_list[self.visIndex[index]]
+        nir_img_name, nir_label, nir_domain = self.nir_list[self.nirIndex[index]]
+        
+        assert vis_domain == 0 and nir_domain == 1
+
+        vis_img = self.get_img_from_path(vis_img_name)
+        nir_img = self.get_img_from_path(nir_img_name)
+
+        return vis_img, nir_img, vis_label, nir_label
+
+    def __len__(self):
+        return len(self.img_list)
+
+    def get_img_from_path(self, img_path):
+        if self.input_mode == 'grey':
+            img = Image.open(img_path).convert('L')
+        elif self.input_mode == 'red':
+            img = Image.open(img_path)
+            img = img.split()[0]
+
+        img = self.transform(img)
+        return img
+
+
+    def list_reader(self, list_file_real, list_file_fake):
+        with open(list_file_real, 'r') as f:
+            lines_real = f.readlines()
+        with open(list_file_fake, 'r') as f:
+            lines_fake = f.readlines()
+
+        fake_label_start = max([int(l.strip().split(' ')[-1]) for l in lines_real]) + 1
+        lines_fake = ["{} {}".format(l.strip().split(' ')[0], int(l.strip().split(' ')[1]) + fake_label_start) for l in lines_fake]
+
+        lines = lines_real + lines_fake
+
+        pid_container = set()
+        for line in lines:
+            pid = int(line.strip().split(' ')[1])
+            pid_container.add(pid)
+        pid2label = {pid:label for label, pid in enumerate(pid_container)}
+
+        img_list_R = []
+        for line in lines_real:
+            img_name, pid = line.strip().split(' ')
+            label = pid2label[int(pid)]
+            domain = 0 if 'VIS' in img_name else 1
+            img_list_R.append((os.path.join(self.img_root_R + img_name), label, domain))
+        
+        img_list_F = []
+        for line in lines_fake:
+            img_name, pid = line.strip().split(' ')
+            label = pid2label[int(pid)]
+
+            # if label in [8192,1984,2110,6344,8566,8589,9362]:         # only with single image pair
+            #     print(img_name)
+            domain = 0 if 'VIS' in img_name else 1
+            img_list_F.append((os.path.join(self.img_root_F + img_name), label, domain))
+        
+        img_list = img_list_R + img_list_F
+
+        return img_list, len(pid_container)
+
+
+from torch.utils.data.sampler import Sampler
+import numpy as np
+
+def GenIdx(train_vis_label, train_nir_label):
+    def get_idx_from_label(train_label):
+        pos = []
+        unique_train_label = np.unique(train_label)
+        for ul in unique_train_label:
+            tmp = np.argwhere(train_label == ul).squeeze().tolist()
+            if isinstance(tmp,int):
+                tmp = [tmp]
+            pos.append(tmp)
+        return pos
+    
+    vis_pos = get_idx_from_label(train_vis_label)
+    nir_pos = get_idx_from_label(train_nir_label)
+
+    return vis_pos, nir_pos
+
+
+class IdentitySampler(Sampler):
+    """Sample person identities evenly in each batch.
+        Args:
+            train_color_label, train_thermal_label: labels of two modalities
+            color_pos, thermal_pos: positions of each identity
+            batchSize: batch size
+    """
+
+    def __init__(self, train_color_label, train_thermal_label, color_pos, thermal_pos, batchSize, num_img_per_id = 4):
+        uni_label = np.unique(train_color_label)
+        self.n_classes = len(uni_label)
+
+        sample_color = np.arange(batchSize)
+        sample_thermal = np.arange(batchSize)
+        N = np.maximum(len(train_color_label), len(train_thermal_label))
+        
+        num_id_per_batch = batchSize / num_img_per_id
+
+        for j in range(N//batchSize+1):
+            batch_idx = np.random.choice(uni_label, int(num_id_per_batch), replace=False)
+            
+            for s, i in enumerate(range(0, batchSize, num_img_per_id)):
+                sample_flag = True if len(color_pos[batch_idx[s]]) < num_img_per_id or len(thermal_pos[batch_idx[s]]) < num_img_per_id else False
+                
+                sample_color[i:i+num_img_per_id]  = np.random.choice(color_pos[batch_idx[s]], num_img_per_id, replace=sample_flag)
+                sample_thermal[i:i+num_img_per_id] = np.random.choice(thermal_pos[batch_idx[s]], num_img_per_id, replace=sample_flag)
+            
+            if j ==0:
+                index1= sample_color
+                index2= sample_thermal
+            else:
+                index1 = np.hstack((index1, sample_color))
+                index2 = np.hstack((index2, sample_thermal))
+        
+        self.visIndex = index1
+        self.nirIndex = index2
+        self.N  = N
+        
+    def __iter__(self):
+        return iter(np.arange(len(self.visIndex)))
+
+    def __len__(self):
+        return self.N 
\ No newline at end of file
diff --git a/recognition/idmmd/eval.sh b/recognition/idmmd/eval.sh
new file mode 100644
index 000000000..49f438731
--- /dev/null
+++ b/recognition/idmmd/eval.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+
+#### Parameters
+# datasets options: "casia", "lamp", "buaa", "oulu"
+# test_fold_id = -1 if testing 10 fold (casia & lamp) else test_fold_id = i (fold id)
+# test_mode: "pretrain" or "finetune"
+
+dataset='oulu'
+# img_root="path to data folder"
+img_root="/storage/local/local/Oulu_CASIA_NIR_VIS/crops112_3/"
+input_mode='grey'
+model_mode='29'
+test_mode='pretrain'
+test_fold_id=-1
+model_name='L29.pth.tar' # pretrain model
+# model_name=$dataset'_fold'$test_fold_id'_final.pth.tar' # finetune: 'casia_fold1_final.pth.tar' 
+
+
+CUDA_VISIBLE_DEVICES=6 python ./evaluate/eval_${dataset}_112.py --test_fold_id $test_fold_id --input_mode $input_mode --model_mode $model_mode --model_name $model_name --img_root $img_root --test_mode $test_mode | tee test.log
\ No newline at end of file
diff --git a/recognition/idmmd/evaluate/__init__.py b/recognition/idmmd/evaluate/__init__.py
new file mode 100644
index 000000000..0ee108bc8
--- /dev/null
+++ b/recognition/idmmd/evaluate/__init__.py
@@ -0,0 +1,2 @@
+
+from evaluate.eval_ops import *
diff --git a/recognition/idmmd/evaluate/eval_buaa_112.py b/recognition/idmmd/evaluate/eval_buaa_112.py
new file mode 100644
index 000000000..4800a7826
--- /dev/null
+++ b/recognition/idmmd/evaluate/eval_buaa_112.py
@@ -0,0 +1,135 @@
+import numpy as np
+import os,sys
+sys.path.append(os.getcwd())
+import argparse
+import torch
+
+from PIL import Image
+from network.lightcnn112 import LightCNN_29Layers
+from evaluate import evaluate2
+
+fars = [10 ** -4, 10 ** -3, 10 ** -2]
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--test_fold_id', default=1, type=int)
+parser.add_argument('--input_mode', default='grey', choices=['grey'], type=str)
+parser.add_argument('--model_mode', default='29', choices=['29'], type=str)
+parser.add_argument('--model_name', default='', type=str)
+parser.add_argument('--img_root', default='', type=str)
+parser.add_argument('--test_mode', default='pretrain', type=str)
+
+args = parser.parse_args()
+
+INPUT_MODE = args.input_mode
+MODEL_MODE = args.model_mode
+model_name = args.model_name
+test_mode = args.test_mode
+img_root = args.img_root
+
+num_classes = 725
+test_list_dir = './data/buaa/' 
+model_dir = f'./models/{test_mode}/'
+model_path = os.path.join(model_dir, model_name)
+
+def load_model(model, pretrained):
+    weights = torch.load(pretrained)
+    weights = weights['state_dict']
+
+    model_dict = model.state_dict()
+    
+    weights = {k.replace('module.',''): v for k, v in weights.items() if k.replace('module.','') in model_dict.keys() and 'fc2' not in k}
+    print("==> len of weights to be loaded: {}. \n".format(len(weights)))
+    model.load_state_dict(weights, strict=False)
+    model.eval()
+
+
+def get_vis_nir_info():
+
+    vis = np.loadtxt(test_list_dir + 'test_vis_paths.txt', dtype=str)
+    vis_labels = [int(s.split('/')[0]) for s in vis]
+    vis = [(p,l) for (p,l) in zip(vis, vis_labels)]
+    
+    nir = np.loadtxt(test_list_dir + 'test_nir_paths.txt', dtype=str)
+    nir_labels = [int(s.split('/')[0]) for s in nir]
+    nir = [(p,l) for (p,l) in zip(nir, nir_labels)]
+    
+    return vis,nir
+
+
+class Embedding:
+    def __init__(self, root, model):
+        self.model = model
+        self.root = root
+
+        self.image_size = (112, 112)
+        self.batch_size = 1
+
+    def get(self, img):
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 1*112*112
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 1, self.image_size[1], self.image_size[0]),
+                              dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+    def extract_feats_labels(self, data_list):
+        img_feats = []
+        pids = []
+        for (imgPath, pid) in data_list:
+            
+            img = Image.open(os.path.join(self.root, imgPath)).convert('L')
+            img = np.array(img)
+            img = img[..., np.newaxis]
+
+            img_feats.append(self.forward_db(self.get(img)).flatten())
+            pids.append(pid)
+        
+        img_feats = np.array(img_feats).astype(np.float32)
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] //2] + img_feats[:, img_feats.shape[1] // 2:]
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+        pids = np.array(pids)
+
+        return img_input_feats, pids
+
+
+
+if MODEL_MODE == '29':
+    model = LightCNN_29Layers(num_classes=num_classes)
+model.cuda()
+
+embedding = Embedding(img_root, model)
+
+if not os.path.exists(model_path):
+    print("cannot find model ",model_path)
+    sys.exit()
+
+load_model(embedding.model, model_path)
+vis, nir= get_vis_nir_info()
+
+feat_vis, label_vis = embedding.extract_feats_labels(vis)
+feat_nir, label_nir = embedding.extract_feats_labels(nir)
+
+labels = np.equal.outer(label_vis, label_nir).astype(np.float32)
+print("*" * 16)
+print("INPUT_MODE: ", INPUT_MODE)
+print("MODEL_MODE: ", MODEL_MODE)
+print("model name: ", model_name)
+print("*" * 16)
+print("[query] feat_nir.shape ",feat_nir.shape)
+print("[gallery] feat_vis.shape ",feat_vis.shape)
+print("*" * 16)
+
+acc, tarfar = evaluate2(feat_vis, feat_nir, labels, fars=fars)
+
+
diff --git a/recognition/idmmd/evaluate/eval_casia_112.py b/recognition/idmmd/evaluate/eval_casia_112.py
new file mode 100644
index 000000000..9578163f1
--- /dev/null
+++ b/recognition/idmmd/evaluate/eval_casia_112.py
@@ -0,0 +1,189 @@
+import numpy as np
+import pandas as pd
+import os,sys
+sys.path.append(os.getcwd())
+import argparse
+
+from PIL import Image
+import torch
+
+from network.lightcnn112 import LightCNN_29Layers
+from evaluate import evaluate2
+
+fars = [10 ** -4, 10 ** -3, 10 ** -2]
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--test_fold_id', default=1, type=int)
+parser.add_argument('--input_mode', default='grey', choices=['grey'], type=str)
+parser.add_argument('--model_mode', default='29', choices=['29'], type=str)
+parser.add_argument('--model_name', default='', type=str)
+parser.add_argument('--img_root', default='', type=str)
+parser.add_argument('--test_mode', default='pretrain', type=str)
+
+args = parser.parse_args()
+
+INPUT_MODE = args.input_mode
+MODEL_MODE = args.model_mode
+model_name = args.model_name
+test_mode = args.test_mode
+img_root = args.img_root
+
+print("*" * 16)
+print("INPUT_MODE: ", INPUT_MODE)
+print("MODEL_MODE: ", MODEL_MODE)
+print("model name: ", model_name)
+print("*" * 16)
+
+tfi = args.test_fold_id
+num_classes = 725
+test_list_dir = 'data/casia/'
+model_dir = f'./models/{test_mode}/'
+model_path = os.path.join(model_dir, model_name)
+
+def load_model(model, pretrained):
+    weights = torch.load(pretrained)
+    weights = weights['state_dict']
+
+    model_dict = model.state_dict()
+    
+    weights = {k.replace('module.',''): v for k, v in weights.items() if k.replace('module.','') in model_dict.keys() and 'fc2' not in k}
+    print("==> len of weights to be loaded: {}. \n".format(len(weights)))
+    model.load_state_dict(weights, strict=False)
+    model.eval()
+
+
+class Embedding:
+    def __init__(self, root, model):
+        self.model = model
+        self.root = root
+
+        self.image_size = (112, 112)
+        self.batch_size = 1
+
+    def get(self, img):
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 1*112*112
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 1, self.image_size[1], self.image_size[0]),
+                              dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+    def extract_feats_labels(self, data_list):
+        img_feats = []
+        pids = []
+        for (imgPath, pid) in data_list:
+            img = Image.open(os.path.join(self.root, imgPath)).convert('L')
+
+            img = np.array(img)
+            img = img[..., np.newaxis]
+
+            img_feats.append(self.forward_db(self.get(img)).flatten())
+            pids.append(pid)
+        
+        img_feats = np.array(img_feats).astype(np.float32)
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] //2] + img_feats[:, img_feats.shape[1] // 2:]
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True)) 
+        pids = np.array(pids)
+
+        return img_input_feats, pids
+
+
+def get_vis_nir_info(test_fold_id):
+    vis = pd.read_csv(os.path.join(test_list_dir, 'vis_gallery_%d.txt' % test_fold_id), header=None, sep=' ')
+    vis_labels = [int(s.split('\\')[-2]) for s in vis[0]]
+    vis = vis[0].apply(lambda s: rename_path(s)).tolist()
+
+    nir = pd.read_csv(os.path.join(test_list_dir, 'nir_probe_%d.txt' % test_fold_id), header=None, sep=' ')
+    nir_labels = [int(s.split('\\')[-2]) for s in nir[0]]
+    nir = nir[0].apply(lambda s: rename_path(s)).tolist()
+
+    vis = [(p,l) for (p,l) in zip(vis, vis_labels)]
+    nir = [(p,l) for (p,l) in zip(nir, nir_labels)]
+
+    return vis,nir
+
+def rename_path(s):
+    """messy path names, inconsistency between 10-folds and how data are actually saved"""
+    s = s.split(".")[0]
+    gr, mod, id, img = s.split("\\")
+    ext = 'jpg' if (mod == 'VIS') else 'bmp'
+    return "%s/%s_%s_%s_%s.%s" % (mod, gr, mod, id, img, ext)
+
+
+
+if MODEL_MODE == '29':
+    model = LightCNN_29Layers(num_classes=num_classes)
+model.cuda()
+embedding = Embedding(img_root, model)
+
+############### test pre-trained models
+if tfi == -1:
+    n_fold = 10
+    acc_ = []
+    tarfar_ = np.zeros((n_fold, 4))
+    for tf in range(n_fold):
+        
+        load_model(embedding.model, model_path)
+        vis, nir= get_vis_nir_info(tf+1)
+        
+        feat_vis, label_vis = embedding.extract_feats_labels(vis)
+        feat_nir, label_nir = embedding.extract_feats_labels(nir)
+
+        labels = np.equal.outer(label_vis, label_nir).astype(np.float32)
+
+        print("*" * 16)
+        print("Fold id ", tf+1)
+        print("Model: ", model_path)
+        print("[query] feat_nir.shape ",feat_nir.shape)
+        print("[gallery] feat_vis.shape ",feat_vis.shape)
+        print("*" * 16)
+
+        acc, tarfar = evaluate2(feat_vis, feat_nir, labels, fars=fars)
+        
+        acc_.append(acc[0])
+        tarfar_[tf,...] = np.array(tarfar)
+    
+    print('\n')
+    print("*" * 16)
+    print("MEAN")
+    print("*" * 16)
+    
+    print("Rank 1 = {:.3%}  +- {:.2%}".format(np.mean(acc_), np.std(acc_)))
+    var_mean = tarfar_.mean(0)
+    var_std = tarfar_.std(0)
+    for fpr_iter in np.arange(len(fars)):
+        print("TAR {:.3%} +- {:.2%} @ FAR {:.4%}".format(var_mean[fpr_iter], var_std[fpr_iter], fars[fpr_iter]))
+
+
+else:
+    if not os.path.exists(model_path):
+        print("cannot find model ",model_path)
+        sys.exit()
+
+    model = load_model(embedding.model, model_path)
+
+    vis, nir= get_vis_nir_info(tfi)
+
+    feat_vis, label_vis = embedding.extract_feats_labels(vis)
+    feat_nir, label_nir = embedding.extract_feats_labels(nir)
+
+    labels = np.equal.outer(label_vis, label_nir).astype(np.float32)
+
+    print("*" * 16)
+    print("Fold id ", tfi)
+    print("[query] feat_nir.shape ",feat_nir.shape)
+    print("[gallery] feat_vis.shape ",feat_vis.shape)
+    print("*" * 16)
+
+    acc, tarfar = evaluate2(feat_vis, feat_nir, labels, fars=fars)
+
diff --git a/recognition/idmmd/evaluate/eval_lamp_112.py b/recognition/idmmd/evaluate/eval_lamp_112.py
new file mode 100644
index 000000000..b746a134f
--- /dev/null
+++ b/recognition/idmmd/evaluate/eval_lamp_112.py
@@ -0,0 +1,169 @@
+import numpy as np
+import os,sys
+sys.path.append(os.getcwd())
+import argparse
+
+from PIL import Image
+
+import torch
+
+from network.lightcnn112 import LightCNN_29Layers
+from evaluate import evaluate2
+
+fars = [10 ** -4, 10 ** -3, 10 ** -2]
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--test_fold_id', default=1, type=int)
+parser.add_argument('--input_mode', default='grey', choices=['grey'], type=str)
+parser.add_argument('--model_mode', default='29', choices=['29'], type=str)
+parser.add_argument('--model_name', default='', type=str)
+parser.add_argument('--img_root', default='', type=str)
+parser.add_argument('--test_mode', default='pretrain', type=str)
+args = parser.parse_args()
+
+INPUT_MODE = args.input_mode
+MODEL_MODE = args.model_mode
+model_name = args.model_name
+test_mode = args.test_mode
+img_root = args.img_root
+
+tfi = args.test_fold_id
+num_classes = 725
+test_list_dir = './data/lamp/'
+model_dir = f'./models/{test_mode}/'
+model_path = os.path.join(model_dir, model_name)
+
+def load_model(model, pretrained):
+    weights = torch.load(pretrained)
+    weights = weights['state_dict']
+
+    model_dict = model.state_dict()
+
+    weights = {k.replace('module.',''): v for k, v in weights.items() if k.replace('module.','') in model_dict.keys()}
+    print("==> len of weights to be loaded: {}. \n".format(len(weights)))
+
+    model.load_state_dict(weights, strict=False)
+    model.eval()
+
+def get_vis_nir_info(test_fold_id):
+    def get_data(test_fold_id, mode='vis'):
+        name = 'gallery_vis%d.txt' % (test_fold_id) if mode=='vis' else 'probe_nir%d.txt' % (test_fold_id)
+        file_data = np.genfromtxt(test_list_dir + name , usecols=(0,1), skip_header=1, dtype=str)
+        paths = file_data[:,0]
+
+        # paths = [p for p in paths if os.path.exists(img_root + p)]
+        return paths
+
+    vis = get_data(test_fold_id, mode='vis')
+    nir = get_data(test_fold_id, mode='nir')
+
+    return vis, nir
+
+
+class Embedding:
+    def __init__(self, root, model):
+        self.model = model
+        self.root = root
+
+        self.image_size = (112, 112)
+        self.batch_size = 1
+
+    def get(self, img):
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 1*112*112
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 1, self.image_size[1], self.image_size[0]),
+                              dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+    def extract_feats_labels(self, data_list):
+        img_feats = []
+        for imgPath in data_list:
+            
+            img = Image.open(os.path.join(self.root, imgPath)).convert('L')
+            img = np.array(img)
+            img = img[..., np.newaxis]
+
+            img_feats.append(self.forward_db(self.get(img)).flatten())
+        
+        img_feats = np.array(img_feats).astype(np.float32)
+
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] //2] + img_feats[:, img_feats.shape[1] // 2:]
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True)) 
+
+        return img_input_feats
+
+
+
+if MODEL_MODE == '29':
+    model = LightCNN_29Layers(num_classes=num_classes)
+model.cuda()
+
+embedding = Embedding(img_root, model)
+
+############### test pre-trained models
+if tfi == -1:
+    n_fold = 10
+    acc_ = []
+    tarfar_ = np.zeros((n_fold, 4))
+    for tf in range(n_fold):
+        load_model(embedding.model, model_path)
+        vis, nir = get_vis_nir_info(tf+1)
+        
+        feat_vis = embedding.extract_feats_labels(vis)
+        feat_nir = embedding.extract_feats_labels(nir)
+        
+        label_matrix = np.load(test_list_dir + 'binary_lable_matrix_%d.npy' % (tf+1))
+        label_matrix = label_matrix.T
+
+        print("*" * 16)
+        print("Fold id ", tf+1)
+        print("Model: ", model_path)
+        print("[query] feat_nir.shape ",feat_nir.shape)
+        print("[gallery] feat_vis.shape ",feat_vis.shape)
+        print("*" * 16)
+
+        acc, tarfar = evaluate2(feat_vis, feat_nir, label_matrix, fars=fars)
+        
+        acc_.append(acc[0])
+        tarfar_[tf,...] = np.array(tarfar)
+    
+    print('\n')
+    print("*" * 16)
+    print("MEAN")
+    print("*" * 16)
+    
+    print("Rank 1 = {:.3%}  +- {:.2%}".format(np.mean(acc_), np.std(acc_)))
+    var_mean = tarfar_.mean(0)
+    var_std = tarfar_.std(0)
+    for fpr_iter in np.arange(len(fars)):
+        print("TAR {:.3%} +- {:.2%} @ FAR {:.4%}".format(var_mean[fpr_iter], var_std[fpr_iter], fars[fpr_iter]))
+
+else:
+    load_model(embedding.model, model_path)
+
+    vis, nir= get_vis_nir_info(tfi)
+
+    feat_vis = embedding.extract_feats_labels(vis)
+    feat_nir = embedding.extract_feats_labels(nir)
+
+    label_matrix = np.load(test_list_dir + 'binary_lable_matrix_%d.npy' % (tfi))
+    label_matrix = label_matrix.T
+
+    print("*" * 16)
+    print("Fold id ", tfi)
+    print("[query] feat_nir.shape ",feat_nir.shape)
+    print("[gallery] feat_vis.shape ",feat_vis.shape)
+    print("*" * 16)
+
+    acc, tarfar = evaluate2(feat_vis, feat_nir, label_matrix, fars=fars)
\ No newline at end of file
diff --git a/recognition/idmmd/evaluate/eval_ops.py b/recognition/idmmd/evaluate/eval_ops.py
new file mode 100644
index 000000000..1c01ee62a
--- /dev/null
+++ b/recognition/idmmd/evaluate/eval_ops.py
@@ -0,0 +1,54 @@
+import numpy as np
+from sklearn.metrics import roc_curve
+
+def evaluate2(gallery_feat, query_feat, labels, fars = [10**-5, 10**-4, 10**-3, 10**-2]):
+    query_num = query_feat.shape[0]
+
+    similarity = np.dot(query_feat, gallery_feat.T)
+    top_inds = np.argsort(-similarity)
+    labels = labels.T
+    
+    # calculate top1
+    correct_num = 0
+    for i in range(query_num):
+        j = top_inds[i, 0]
+        if labels[i, j] == 1:
+            correct_num += 1
+    top1 = correct_num / query_num
+    print("top1 = {:.2%}".format(top1))
+
+    # # calculate top5
+    # correct_num = 0
+    # for i in range(query_num):
+    #     j = top_inds[i, :5]
+    #     if any(labels[i, j] == 1.0):
+    #         correct_num += 1
+    #     # else:
+    #     #     print(i,j)
+    # top5 = correct_num / query_num
+    # print("top5 = {:.4%}".format(top5))
+
+    # # calculate 10
+    # correct_num = 0
+    # for i in range(query_num):
+    #     j = top_inds[i, :10]
+    #     if any(labels[i, j] == 1.0):
+    #         correct_num += 1
+    #     # else:
+    #     #     print(i,j)
+    # top10 = correct_num / query_num
+    # print("top10 = {:.4%}".format(top10))
+
+    labels_ = labels.flatten()
+    similarity_ = similarity.flatten()
+    fpr, tpr, _ = roc_curve(labels_, similarity_)
+
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)
+    tpr_fpr_row = []
+    for far in fars:
+        _, min_index = min(list(zip(abs(fpr - far), range(len(fpr)))))
+        tpr_fpr_row.append(tpr[min_index])
+        print("TPR {:.2%} @ FAR {:.4%}".format(tpr[min_index], far))
+        
+    return [top1], tpr_fpr_row
diff --git a/recognition/idmmd/evaluate/eval_oulu_112.py b/recognition/idmmd/evaluate/eval_oulu_112.py
new file mode 100644
index 000000000..cb3b0aadb
--- /dev/null
+++ b/recognition/idmmd/evaluate/eval_oulu_112.py
@@ -0,0 +1,160 @@
+import numpy as np
+import pandas as pd
+import os,sys
+sys.path.append(os.getcwd())
+print(sys.path)
+import argparse
+import torch
+
+from PIL import Image
+
+from network.lightcnn112 import LightCNN_29Layers
+from evaluate import evaluate2
+
+fars = [10 ** -4, 10 ** -3, 10 ** -2]
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--test_fold_id', default=1, type=int)
+parser.add_argument('--input_mode', default='grey', choices=['grey'], type=str)
+parser.add_argument('--model_mode', default='29', choices=['29'], type=str)
+parser.add_argument('--model_name', default='', type=str)
+parser.add_argument('--img_root', default='', type=str)
+parser.add_argument('--test_mode', default='pretrain', type=str)
+
+args = parser.parse_args()
+
+INPUT_MODE = args.input_mode
+MODEL_MODE = args.model_mode
+model_name = args.model_name
+test_mode = args.test_mode
+img_root = args.img_root
+
+num_classes = 725
+test_list_dir = './data/oulu/' 
+model_dir = f'./models/{test_mode}/'
+model_path = os.path.join(model_dir, model_name)
+
+def load_model(model, pretrained):
+    weights = torch.load(pretrained)
+    weights = weights['state_dict']
+
+    model_dict = model.state_dict()
+    
+    weights = {k.replace('module.',''): v for k, v in weights.items() if k.replace('module.','') in model_dict.keys() and 'fc2' not in k}
+
+    print("==> len of weights to be loaded: {}. \n".format(len(weights)))
+    model.load_state_dict(weights, strict=False)
+    model.eval()
+    
+class Embedding:
+    def __init__(self, root, model):
+        self.model = model
+        self.root = root
+
+        self.image_size = (112, 112)
+        self.batch_size = 1
+
+    def get(self, img):
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 1*112*112
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 1, self.image_size[1], self.image_size[0]),
+                              dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+    def extract_feats_labels(self, data_list):
+        img_feats = []
+        pids = []
+        for (imgPath, pid) in data_list:
+            
+            img = Image.open(os.path.join(self.root, imgPath)).convert('L')
+            img = np.array(img)
+            img = img[..., np.newaxis]
+
+            img_feats.append(self.forward_db(self.get(img)).flatten())
+            pids.append(pid)
+        
+        img_feats = np.array(img_feats).astype(np.float32)
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] //2] + img_feats[:, img_feats.shape[1] // 2:]
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+        pids = np.array(pids)
+
+        return img_input_feats, pids
+
+
+def get_vis_nir_info_csv():
+    vis = pd.read_csv(test_list_dir + 'vis_test_paths.csv', header=None, sep=' ')
+    vis_labels = [int(s.strip().split(',')[-1].split('P')[-1]) for s in vis[0]]
+    vis = [s.strip().split(',')[0] for s in vis[0]]
+
+    nir = pd.read_csv(test_list_dir + 'nir_test_paths.csv', header=None, sep=' ')
+    nir_labels = [int(s.strip().split(',')[-1].split('P')[-1]) for s in nir[0]]
+    nir = [s.strip().split(',')[0] for s in nir[0]]
+
+    vis = [(p,l) for (p,l) in zip(vis, vis_labels)]
+    nir = [(p,l) for (p,l) in zip(nir, nir_labels)]
+    
+    return vis,nir
+
+def get_vis_nir_info_txt():
+    def read_file(file_name):
+        with open(test_list_dir + file_name, 'r') as f:
+            lines = f.readlines()
+        paths = [s.strip().split(' ')[0] for s in lines]
+        labels = [int(s.strip().split(' ')[1]) for s in lines]
+        info = [(p,l) for (p,l) in zip(paths, labels)]
+
+        return info
+
+    vis = read_file('test_vis_paths.txt')
+    nir = read_file('test_nir_paths.txt')
+    
+    return vis, nir
+
+
+### Testing pretrain/finetune model
+if test_mode == 'pretrain':
+    vis, nir = get_vis_nir_info_csv()
+elif test_mode == "finetune":
+    vis, nir = get_vis_nir_info_txt()
+else:
+    print("Wrong test_mode!!!")
+
+if MODEL_MODE == '29':
+    model = LightCNN_29Layers(num_classes=num_classes)
+
+model.cuda()
+
+embedding = Embedding(img_root, model)
+
+if not os.path.exists(model_path):
+    print("cannot find model ",model_path)
+    sys.exit()
+
+load_model(embedding.model, model_path)
+
+feat_vis, label_vis = embedding.extract_feats_labels(vis)
+feat_nir, label_nir = embedding.extract_feats_labels(nir)
+
+labels = np.equal.outer(label_vis, label_nir).astype(np.float32)
+
+print("*" * 16)
+print("INPUT_MODE: ", INPUT_MODE)
+print("MODEL_MODE: ", MODEL_MODE)
+print("model path: ", model_path)
+print("*" * 16)
+print("[query] feat_nir.shape ",feat_nir.shape)
+print("[gallery] feat_vis.shape ",feat_vis.shape)
+print("*" * 16)
+
+acc, tarfar = evaluate2(feat_vis, feat_nir, labels, fars=fars)
\ No newline at end of file
diff --git a/recognition/idmmd/losses.py b/recognition/idmmd/losses.py
new file mode 100644
index 000000000..b09fe0186
--- /dev/null
+++ b/recognition/idmmd/losses.py
@@ -0,0 +1,106 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+class IDMMD(nn.Module):
+    def __init__(self, kernel_type='rbf', kernel_mul=2.0, kernel_num=5):
+        super(IDMMD, self).__init__()
+        self.kernel_num = kernel_num
+        self.kernel_mul = kernel_mul
+        self.fix_sigma = None
+        self.kernel_type = kernel_type
+
+    def get_centers_by_id(self, x_rgb, x_ir, targets):
+        centers_rgb = []
+        centers_ir = []
+
+        batch_y_set = set(targets.data.cpu().numpy())
+
+        for _, l in enumerate(batch_y_set):
+            feat1 = x_rgb[targets==l]
+            feat2 = x_ir[targets==l]
+
+            centers_rgb.append(feat1.mean(dim=0).unsqueeze(0))
+            centers_ir.append(feat2.mean(dim=0).unsqueeze(0))
+
+        centers_rgb = torch.cat(centers_rgb, 0).cuda()
+        centers_ir = torch.cat(centers_ir, 0).cuda()
+
+        return centers_rgb, centers_ir
+    
+    def forward(self, x_rgb, x_ir, targets):
+
+        centers_rgb, centers_ir = self.get_centers_by_id(x_rgb, x_ir, targets)
+
+        if self.kernel_type == 'linear':
+            loss = self.linear_mmd(centers_rgb, centers_ir)         # domain-level loss
+
+        elif self.kernel_type == 'rbf':
+            B = centers_rgb.size(0)
+            kernels = self.guassian_kernel(centers_rgb, centers_ir)
+
+            XX = kernels[:B, :B]
+            YY = kernels[B:, B:]
+            XY = kernels[:B, B:]
+            YX = kernels[B:, :B]
+
+            loss = (XX + YY - XY - YX).mean()
+
+        return loss
+        
+
+    def linear_mmd(self, center_rgb, center_ir):
+        def compute_dist_(x_rgb, x_ir):
+            n = x_rgb.size(0)
+            dist1 = torch.pow(x_rgb, 2).sum(dim=1, keepdim=True).expand(n, n)
+            dist2 = torch.pow(x_ir, 2).sum(dim=1, keepdim=True).expand(n, n)
+            
+            dist = dist1 + dist2.t()
+            dist.addmm_(mat1=x_rgb, mat2=x_ir.t(), beta=1, alpha=-2)
+            dist = dist.clamp(min=1e-12)  # for numerical stability
+            return dist
+
+        matrix = compute_dist_(center_rgb, center_ir)
+        loss = matrix.diag()
+        
+        return loss.mean()
+
+
+    def guassian_kernel(self, x_rgb, x_ir):
+        total = torch.cat([x_rgb, x_ir], dim=0)
+        N = total.size(0)
+
+        total0 = total.unsqueeze(0).expand(
+            int(total.size(0)), int(total.size(0)), int(total.size(1)))
+        total1 = total.unsqueeze(1).expand(
+            int(total.size(0)), int(total.size(0)), int(total.size(1)))
+        dists = ((total0-total1)**2).sum(2)
+
+        if self.fix_sigma:
+            bandwidth = self.fix_sigma
+        else:
+            bandwidth = torch.sum(dists.data) / (N**2-N)
+        
+        bandwidth /= self.kernel_mul ** (self.kernel_num // 2)
+        bandwidth_list = [bandwidth * (self.kernel_mul**i)
+                          for i in range(self.kernel_num)]
+        kernel_val = [torch.exp(-dists / bandwidth_temp)
+                      for bandwidth_temp in bandwidth_list]
+        return sum(kernel_val)
+
+
+
+class CosFace(torch.nn.Module):
+    def __init__(self, s=64.0, m=0.40):
+        super(CosFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, logits, labels):
+        one_hot = torch.zeros_like(logits).scatter_(1, labels.view(-1, 1), 1.0).cuda()
+        phi = logits - self.m
+        output = torch.where(one_hot==1, phi, logits)
+        output *= self.s
+
+        return output
\ No newline at end of file
diff --git a/recognition/idmmd/models/finetune/readme.txt b/recognition/idmmd/models/finetune/readme.txt
new file mode 100644
index 000000000..5f8832014
--- /dev/null
+++ b/recognition/idmmd/models/finetune/readme.txt
@@ -0,0 +1 @@
+put $dataset_final.pth.tar here
diff --git a/recognition/idmmd/models/pretrain/readme.txt b/recognition/idmmd/models/pretrain/readme.txt
new file mode 100644
index 000000000..86ebb62a3
--- /dev/null
+++ b/recognition/idmmd/models/pretrain/readme.txt
@@ -0,0 +1,3 @@
+Put pretrain model here.
+
+Ex. L29.pth.tar
diff --git a/recognition/idmmd/network/__init__.py b/recognition/idmmd/network/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/idmmd/network/lightcnn112.py b/recognition/idmmd/network/lightcnn112.py
new file mode 100644
index 000000000..d36254d86
--- /dev/null
+++ b/recognition/idmmd/network/lightcnn112.py
@@ -0,0 +1,181 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+
+
+class mfm(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, type=1):
+        super(mfm, self).__init__()
+        self.out_channels = out_channels
+        if type == 1:
+            self.filter = nn.Conv2d(in_channels, 2 * out_channels, kernel_size=kernel_size, stride=stride,
+                                    padding=padding)
+        else:
+            self.filter = nn.Linear(in_channels, 2 * out_channels)
+
+    def forward(self, x):
+        x = self.filter(x)
+        out = torch.split(x, self.out_channels, 1)
+        return torch.max(out[0], out[1])
+
+
+class group(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, padding):
+        super(group, self).__init__()
+        self.conv_a = mfm(in_channels, in_channels, 1, 1, 0)
+        self.conv = mfm(in_channels, out_channels, kernel_size, stride, padding)
+
+    def forward(self, x):
+        x = self.conv_a(x)
+        x = self.conv(x)
+        return x
+
+
+class resblock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(resblock, self).__init__()
+        self.conv1 = mfm(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.conv2 = mfm(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        res = x
+        out = self.conv1(x)
+        out = self.conv2(out)
+        out = out + res
+        return out
+
+
+
+class network_29layers(nn.Module):
+    def __init__(self, block, layers, num_classes=79077):
+        super(network_29layers, self).__init__()
+        self.conv1  = mfm(1, 48, 5, 1, 2)
+        self.pool1  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block1 = self._make_layer(block, layers[0], 48, 48)
+        self.group1 = group(48, 96, 3, 1, 1)
+        self.pool2  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block2 = self._make_layer(block, layers[1], 96, 96)
+        self.group2 = group(96, 192, 3, 1, 1)
+        self.pool3  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block3 = self._make_layer(block, layers[2], 192, 192)
+        self.group3 = group(192, 128, 3, 1, 1)
+        self.block4 = self._make_layer(block, layers[3], 128, 128)
+        self.group4 = group(128, 128, 3, 1, 1)
+        self.pool4  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.fc     = mfm(7*7*128, 256, type=0)
+        self.fc2    = nn.Linear(256, num_classes)
+            
+    def _make_layer(self, block, num_blocks, in_channels, out_channels):
+        layers = []
+        for i in range(0, num_blocks):
+            layers.append(block(in_channels, out_channels))
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.pool1(x)
+
+        x = self.block1(x)
+        x = self.group1(x)
+        x = self.pool2(x)
+
+        x = self.block2(x)
+        x = self.group2(x)
+        x = self.pool3(x)
+
+        x = self.block3(x)
+        x = self.group3(x)
+        x = self.block4(x)
+        x = self.group4(x)
+        x = self.pool4(x)
+
+        x = x.view(x.size(0), -1)
+        fc = self.fc(x)
+
+        if self.training:
+            x = F.dropout(fc, training=self.training)
+            out = self.fc2(x)
+            return out, F.normalize(fc,p=2,dim=1)
+        return F.normalize(fc,p=2,dim=1)
+
+
+
+
+################################
+## cosface nets
+################################
+
+from torch.nn import Parameter
+
+
+class network_29layers_cosface(nn.Module):
+    def __init__(self, block, layers, num_classes=79077):
+        super(network_29layers_cosface, self).__init__()
+        self.conv1  = mfm(1, 48, 5, 1, 2)
+        self.pool1  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block1 = self._make_layer(block, layers[0], 48, 48)
+        self.group1 = group(48, 96, 3, 1, 1)
+        self.pool2  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block2 = self._make_layer(block, layers[1], 96, 96)
+        self.group2 = group(96, 192, 3, 1, 1)
+        self.pool3  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.block3 = self._make_layer(block, layers[2], 192, 192)
+        self.group3 = group(192, 128, 3, 1, 1)
+        self.block4 = self._make_layer(block, layers[3], 128, 128)
+        self.group4 = group(128, 128, 3, 1, 1)
+        self.pool4  = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
+        self.fc     = mfm(7*7*128, 256, type=0)
+
+        self.weight = Parameter(torch.Tensor(num_classes, 256))
+        nn.init.xavier_uniform_(self.weight)
+    
+
+    def _make_layer(self, block, num_blocks, in_channels, out_channels):
+        layers = []
+        for i in range(0, num_blocks):
+            layers.append(block(in_channels, out_channels))
+        return nn.Sequential(*layers)
+
+    def cosine_sim(self, x1, x2, dim=1, eps=1e-8):
+        ip = torch.mm(x1, x2.t())
+        w1 = torch.norm(x1, 2, dim)
+        w2 = torch.norm(x2, 2, dim)
+        return ip / torch.ger(w1,w2).clamp(min=eps)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.pool1(x)
+
+        x = self.block1(x)
+        x = self.group1(x)
+        x = self.pool2(x)
+
+        x = self.block2(x)
+        x = self.group2(x)
+        x = self.pool3(x)
+
+        x = self.block3(x)
+        x = self.group3(x)
+        x = self.block4(x)
+        x = self.group4(x)
+        x = self.pool4(x)
+
+        x = x.view(x.size(0), -1)
+        fc = self.fc(x)
+    
+        if self.training:
+            x = F.dropout(fc, training=self.training)
+            out = self.cosine_sim(x, self.weight)
+            return out, F.normalize(fc, p=2, dim=1)
+        return F.normalize(fc, p=2, dim=1)
+
+
+def LightCNN_29Layers(**kwargs):
+    model = network_29layers(resblock, [1, 2, 3, 4], **kwargs)
+    return model
+
+def LightCNN_29Layers_cosface(**kwargs):
+    model = network_29layers_cosface(resblock, [1, 2, 3, 4], **kwargs)
+    return model
diff --git a/recognition/idmmd/pics/readme.txt b/recognition/idmmd/pics/readme.txt
new file mode 100644
index 000000000..8b1378917
--- /dev/null
+++ b/recognition/idmmd/pics/readme.txt
@@ -0,0 +1 @@
+
diff --git a/recognition/idmmd/run.sh b/recognition/idmmd/run.sh
new file mode 100644
index 000000000..9cdf43b80
--- /dev/null
+++ b/recognition/idmmd/run.sh
@@ -0,0 +1,45 @@
+#!/bin/bash
+
+# run : bash run_train_lightcnn_112.sh
+
+echo train lightcnn 112
+
+gpu_ids='0,1,2,3,4,5,6,7'
+workers=8
+epochs=10
+batch_size=64
+lr=5e-3
+print_iter=40
+train_fold_id=10
+input_mode='grey'
+model_mode='29'
+weights_lightcnn='./models/pretrain/L29.pth.tar'
+
+#! LAMP-HQ
+# dataset='lamp'
+# img_root_R=''
+# train_list_R=''
+
+#! CASIA
+dataset='CASIA'
+img_root_R='' # path to real data
+train_list_R='' # name list
+
+#! Oulu
+# dataset='oulu'
+# img_root_R=''
+# train_list_R=''
+
+#! Buaa
+# dataset='buaa'
+# img_root_R=''
+# train_list_R=''
+
+
+#! finetune 112_cos models
+prefix='train'
+python train.py --gpu_ids $gpu_ids --dataset $dataset --workers $workers \
+                --epochs $epochs --batch_size $batch_size --lr $lr --save_name $prefix --input_mode $input_mode \
+                --print_iter $print_iter --weights_lightcnn $weights_lightcnn \
+                --img_root_R $img_root_R --train_list_R $train_list_R \
+                --model_mode $model_mode 
\ No newline at end of file
diff --git a/recognition/idmmd/train.py b/recognition/idmmd/train.py
new file mode 100644
index 000000000..37fc16aad
--- /dev/null
+++ b/recognition/idmmd/train.py
@@ -0,0 +1,199 @@
+import os
+import argparse
+import random
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.backends.cudnn as cudnn
+
+from utils import *
+from network.lightcnn112 import LightCNN_29Layers_cosface
+from losses import IDMMD, CosFace
+from dataset_mix import Real_Dataset_112_paired, IdentitySampler, GenIdx
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--gpu_ids', default='0,1', type=str)
+parser.add_argument('--workers', default=8, type=int)
+parser.add_argument('--epochs', default=15, type=int)
+parser.add_argument('--pre_epoch', default=0, type=int)
+parser.add_argument('--batch_size', default=64, type=int)
+parser.add_argument('--lr', default=0.001, type=float)
+parser.add_argument('--momentum', default=0.9, type=float)
+parser.add_argument('--weight_decay', default=2e-4)
+parser.add_argument('--step_size', default=5, type=int)
+parser.add_argument('--print_iter', default=5, type=int)
+parser.add_argument('--save_name', default='', type=str)
+parser.add_argument('--seed', default=1000, type=int)
+parser.add_argument('--weights_lightcnn', default='', type=str)
+parser.add_argument('--dataset', default='CASIA', type=str)
+
+parser.add_argument('--img_root_R', default='', type=str)
+parser.add_argument('--train_list_R', default='', type=str)
+
+parser.add_argument('--input_mode', default='red', choices=['grey'], type=str)
+parser.add_argument('--model_mode', default='9',choices=['9','29'], type=str)
+
+
+def main():
+    global args
+    args = parser.parse_args()
+    print(args)
+
+    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_ids
+    cudnn.benchmark = True
+    cudnn.enabled = True
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.cuda.manual_seed(args.seed)
+
+    dataset = 'lamp' if args.dataset == 'LAMP-HQ' else args.dataset.lower()
+
+    # train loader of real data
+    real_dataset_paired = Real_Dataset_112_paired(args)
+    vis_pos, nir_pos = GenIdx(real_dataset_paired.vis_labels, real_dataset_paired.nir_labels)
+    sampler = IdentitySampler(real_dataset_paired.vis_labels, real_dataset_paired.nir_labels, vis_pos, nir_pos, args.batch_size, 4)
+    
+    real_dataset_paired.visIndex = sampler.visIndex
+    real_dataset_paired.nirIndex = sampler.nirIndex
+    
+    train_loader_real_paired = torch.utils.data.DataLoader(
+        real_dataset_paired, batch_size=args.batch_size, sampler=sampler, num_workers=args.workers, pin_memory=True)
+    
+    num_classes = real_dataset_paired.num_classes
+
+    model = LightCNN_29Layers_cosface(num_classes=num_classes)
+
+    model = torch.nn.DataParallel(model).cuda()
+
+    # load pre trained model
+    if args.pre_epoch:
+        print('load pretrained model of epoch %d' % args.pre_epoch)
+        load_model(model, "./model/lightCNN_epoch_%d.pth.tar" % args.pre_epoch)
+    else:
+        print("=> loading pretrained lightcnn '{}'".format(args.weights_lightcnn))
+        load_model_train_lightcnn(model, args.weights_lightcnn)
+
+    # criterion
+    criterion = nn.CrossEntropyLoss().cuda()
+    criterion_idmmd = IDMMD().cuda()
+    margin_softmax = CosFace(s=64.0, m=0.4).cuda()
+
+    '''
+    Stage I: model pretrained for last fc2 parameters
+    '''
+    params_pretrain = []
+    for name, value in model.named_parameters():
+        if name == "module.weight":
+            params_pretrain += [{"params": value, "lr": 1 * args.lr}]
+
+    print("Stage I: trainable params ", len(params_pretrain))
+    assert len(params_pretrain) > 0
+
+    # optimizer
+    optimizer_pretrain = torch.optim.SGD(params_pretrain, args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
+
+    for epoch in range(1, 5):
+        pre_train_pair(train_loader_real_paired, model, criterion, margin_softmax, optimizer_pretrain, epoch)
+        # save_checkpoint(model, epoch, args.save_name+"_pretrain", dataset)
+
+    '''
+    Stage II: model finetune for full network
+    '''
+    # optimizer
+    optimizer = torch.optim.SGD(model.parameters(), args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
+
+    start_epoch = args.pre_epoch + 1
+    for epoch in range(start_epoch, args.epochs + 1):
+        adjust_learning_rate(args.lr, args.step_size, optimizer, epoch)
+        train(train_loader_real_paired, model, criterion, criterion_idmmd, margin_softmax, optimizer, epoch)
+        if epoch == args.epochs or epoch % 10 == 0:
+            save_checkpoint(model, epoch, args.save_name, dataset)
+    
+
+def pre_train_pair(train_loader, model, criterion, margin_softmax, optimizer, epoch):
+    top1 = AverageMeter()
+    top5 = AverageMeter()
+
+    model.train()
+    for i, (vis_img, nir_img, vis_label, nir_label) in enumerate(train_loader):
+
+        input = torch.cat((vis_img, nir_img), 0).cuda(non_blocking=True)
+        label = torch.cat((vis_label, nir_label), 0).cuda(non_blocking=True)
+        batch_size = input.size(0)
+
+        if batch_size < 2*args.batch_size:
+            continue
+
+        # forward
+        output = model(input)[0]
+        output = margin_softmax(output, label)
+        loss = criterion(output, label)
+
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        # measure accuracy and record loss
+        prec1, prec5 = accuracy(output.data, label.data, topk=(1, 5))
+        top1.update(prec1.item(), batch_size)
+        top5.update(prec5.item(), batch_size)
+
+        # print log
+        if i % args.print_iter == 0:
+            info = "====> Epoch: [{:0>3d}][{:3d}/{:3d}] | ".format(epoch, i, len(train_loader))
+            info += "Loss: ce: {:4.3f} | ".format(loss.item())
+            info += "Prec@1: {:4.2f} ({:4.2f}) Prec@5: {:4.2f} ({:4.2f})".format(top1.val, top1.avg, top5.val, top5.avg)
+            print(info)
+
+
+def train(train_loader, model, criterion, criterion_idmmd, margin_softmax, optimizer, epoch, beta = 100):
+    top1 = AverageMeter()
+    top5 = AverageMeter()
+
+    model.train()
+    for i, (vis_img, nir_img, vis_label, nir_label) in enumerate(train_loader):
+
+        input = torch.cat((vis_img, nir_img), 0).cuda(non_blocking=True)
+        label = torch.cat((vis_label, nir_label), 0).cuda(non_blocking=True)
+        batch_size = input.size(0)
+
+        if batch_size < 2*args.batch_size:
+            continue
+        
+        # forward
+        output, fc = model(input)
+        output = margin_softmax(output, label)
+        loss_ce = criterion(output, label)
+
+        num_vis = vis_img.size(0)
+        num_nir = nir_img.size(0)
+        fc_vis, fc_nir = torch.split(fc, [num_vis, num_nir], dim=0)
+
+        loss_idmmd = criterion_idmmd(fc_vis, fc_nir, label[:vis_img.size(0)])
+
+        loss = loss_ce + beta * loss_idmmd
+
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        # measure accuracy and record loss
+        prec1, prec5 = accuracy(output.data, label.data, topk=(1, 5))
+        top1.update(prec1.item(), batch_size)
+        top5.update(prec5.item(), batch_size)
+
+        # print log
+        if i % args.print_iter == 0:
+            info = "====> Epoch: [{:0>3d}][{:3d}/{:3d}] | ".format(epoch, i, len(train_loader))
+            info += "Loss_ce: {:4.3f} | ".format(loss_ce.data)
+            info += "loss_idmmd: {:4.3f} | ".format(loss_idmmd.data)
+            info += "Loss_all: {:4.3f} | ".format(loss.item())
+            info += "Prec@1: {:4.2f} ({:4.2f}) Prec@5: {:4.2f} ({:4.2f})".format(top1.val, top1.avg, top5.val, top5.avg)
+            print(info)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/recognition/idmmd/utils.py b/recognition/idmmd/utils.py
new file mode 100644
index 000000000..521ec4a32
--- /dev/null
+++ b/recognition/idmmd/utils.py
@@ -0,0 +1,150 @@
+import os
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+def ort_loss(x, y):
+    loss = torch.abs((x * y).sum(dim=1)).sum()
+    loss = loss / float(x.size(0))
+    return loss
+
+
+def ang_loss(x, y):
+    loss = (x * y).sum(dim=1).sum()
+    loss = loss / float(x.size(0))
+    return loss
+
+def MMD_Loss(fc_nir, fc_vis):
+    mean_fc_nir = torch.mean(fc_nir, 0)
+    mean_fc_vis = torch.mean(fc_vis, 0)
+    loss_mmd = F.mse_loss(mean_fc_nir, mean_fc_vis)
+    return loss_mmd
+
+
+def rgb2gray(img):
+    r, g, b = torch.split(img, 1, dim=1)
+    return torch.mul(r, 0.299) + torch.mul(g, 0.587) + torch.mul(b, 0.114)
+
+
+def save_checkpoint(model, epoch, name="", dataset=''):
+    if not os.path.exists("model/{}/".format(dataset)):
+        os.makedirs("model/{}/".format(dataset))
+    model_path = "model/{}/".format(dataset) + name + "_e{}.pth.tar".format(epoch)
+    state = {"epoch": epoch, "state_dict": model.state_dict()}
+    torch.save(state, model_path)
+    print("checkpoint saved to {}".format(model_path))
+
+
+def load_model(model, pretrained):
+    weights = torch.load(pretrained)
+    pretrained_dict = weights["state_dict"]
+    model_dict = model.state_dict()
+
+    # print("to here")
+    # print(model_dict.keys())
+    # print('\n')
+
+    # print(pretrained_dict.keys())
+
+    # import pdb;pdb.set_trace()
+
+    if 'LightCNN' in pretrained:
+        tmp = [k for k in pretrained_dict]
+        if "module." in tmp[0]:
+            pretrained_dict = {k.replace('module.',''): v for k, v in pretrained_dict.items() if k.replace('module.','') in model_dict}
+    else:
+        pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict and k!='module.weight'}
+    
+    print("len of params to be loaded: ",len(pretrained_dict))
+    model.load_state_dict(pretrained_dict, strict=False)
+
+    return weights['epoch']
+
+
+def load_model_train_lightcnn(model, pretrained):
+    weights = torch.load(pretrained)
+    pretrained_dict = weights["state_dict"]
+    model_dict = model.state_dict()
+
+    # print("to here")
+    # print(model_dict.keys())
+    # print('\n')
+
+    # print(pretrained_dict.keys())
+
+    # import pdb;pdb.set_trace()
+
+    pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict and 'module.weight' not in k}
+    
+    print("len of params to be loaded: ",len(pretrained_dict))
+    model.load_state_dict(pretrained_dict, strict=False)
+
+    return weights['epoch']
+
+
+def set_requires_grad(nets, requires_grad=False):
+    if not isinstance(nets, list):
+        nets = [nets]
+    for net in nets:
+        if net is not None:
+            for param in net.parameters():
+                param.requires_grad = requires_grad
+
+
+# assign adain_params to AdaIN layers
+def assign_adain_params(adain_params, model):
+    for m in model.modules():
+        if m.__class__.__name__ == "AdaptiveInstanceNorm2d":
+            mean = adain_params[:, :m.num_features]
+            std = adain_params[:, m.num_features:2*m.num_features]
+            m.bias = mean.contiguous().view(-1)
+            m.weight = std.contiguous().view(-1)
+            if adain_params.size(1) > 2*m.num_features:
+                adain_params = adain_params[:, 2*m.num_features:]
+
+
+def accuracy(output, target, topk=(1,)):
+    """Computes the precision@k for the specified values of k"""
+    maxk = max(topk)
+    batch_size = target.size(0)
+
+    _, pred = output.topk(maxk, 1, True, True)
+    pred = pred.t()
+    
+    correct = pred.eq(target.unsqueeze(0).expand_as(pred))
+
+    res = []
+    for k in topk:
+        correct_k = correct[:k].contiguous().view(-1).float().sum(0, keepdim=True)
+        res.append(correct_k.mul_(100.0 / batch_size))
+    return res
+
+
+
+def adjust_learning_rate(lr, step, optimizer, epoch):
+    scale = 0.457305051927326
+    lr = lr * (scale ** (epoch // step))
+    print('lr: {}'.format(lr))
+    if (epoch != 0) & (epoch % step == 0):
+        print('Change lr')
+        for param_group in optimizer.param_groups:
+            param_group['lr'] = param_group['lr'] * scale
+
+
+class AverageMeter(object):
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
diff --git a/recognition/partial_fc/README.md b/recognition/partial_fc/README.md
index 8c02c67ad..ea2a11960 100644
--- a/recognition/partial_fc/README.md
+++ b/recognition/partial_fc/README.md
@@ -3,20 +3,26 @@ Partial FC is a distributed deep learning training framework for face recognitio
 
 ![Image text](https://github.com/nttstar/insightface-resources/blob/master/images/partial_speed1.png)
 
+## Paper
+- [Partial FC: Training 10 Million Identities on a Single Machine](https://openaccess.thecvf.com/content/ICCV2021W/MFR/html/An_Partial_FC_Training_10_Million_Identities_on_a_Single_Machine_ICCVW_2021_paper.html) in **MFR 2021**
+- [Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC](https://openaccess.thecvf.com/content/CVPR2022/html/An_Killing_Two_Birds_With_One_Stone_Efficient_and_Robust_Training_CVPR_2022_paper.html) in **CVPR 2022**
 
 ## Contents
-[Partial FC](https://arxiv.org/abs/2010.05222)
-- [Largest Face Recognition Dataset: **Glint360k**](#Glint360k)
-- [Distributed Training Performance](#Performance)
+Partial FC
+- [Largest Face Recognition Dataset: **Glint360k**](#Glint360K)
+- [Docker](#Docker)
+- [Performance On Million Identities](#Benchmark)
+- [FAQ](#FAQ)
 - [Citation](#Citation)
 
 
 ## Glint360K
-We clean, merge, and release the **largest** and **cleanest** face recognition dataset **Glint360K**, which contains 18 million images of 360K individuals. 
+We clean, merge, and release the largest and cleanest face recognition dataset Glint360K, 
+which contains **`17091657`** images of **`360232`** individuals. 
 By employing the Patial FC training strategy, baseline models trained on Glint360K can easily achieve state-of-the-art performance. 
 Detailed evaluation results on the large-scale test set (e.g. IFRT, IJB-C and Megaface) are as follows:
 
-#### Evaluation on IFRT       
+### 1. Evaluation on IFRT       
 **`r`** denotes the sampling rate of negative class centers.
 | Backbone     | Dataset            | African | Caucasian | Indian | Asian | ALL   |
 | ------------ | -----------        | ----- | ----- | ------ | ----- | ----- |
@@ -25,7 +31,7 @@ Detailed evaluation results on the large-scale test set (e.g. IFRT, IJB-C and Me
 | R100         | **Glint360k**(r=1.0)   | 89.50 | 94.23 | 93.54  | **65.07** | **88.67** |
 | R100         | **Glint360k**(r=0.1)   | **90.45** | **94.60** | **93.96**  | 63.91 | 88.23 |
 
-#### Evaluation on IJB-C and Megaface  
+### 2. Evaluation on IJB-C and Megaface  
 We employ ResNet100 as the backbone and CosFace (m=0.4) as the loss function.
 TAR@FAR=1e-4 is reported on the IJB-C datasets, and TAR@FAR=1e-6 is reported on the Megaface dataset.
 |Test Dataset        | IJB-C     | Megaface_Id  | Megaface_Ver |
@@ -33,21 +39,84 @@ TAR@FAR=1e-4 is reported on the IJB-C datasets, and TAR@FAR=1e-6 is reported on
 | MS1MV2             | 96.4      | 98.3         | 98.6         |
 |**Glint360k** | **97.3**  | **99.1**     | **99.1**     |
 
-#### Download
-[**Baidu Drive**](https://pan.baidu.com/s/1aHC_nJGKzKgwJKoVb2Q_Gg) (code:i1al)  
+### 3. License 
+
+The Glint360K dataset (and the models trained with this dataset) are available for non-commercial research purposes only.
+
+### 4. Download
+- [x] [**Baidu Drive**](https://pan.baidu.com/s/1GsYqTTt7_Dn8BfxxsLFN0w) (code:o3az)    
+- [x] **Magnet URI**: `magnet:?xt=urn:btih:E5F46EE502B9E76DA8CC3A0E4F7C17E4000C7B1E&dn=glint360k`
 
 Refer to the following command to unzip.
 ```
-cat glint360k* > glint360k.tar
-tar -xvf glint360k.tar
+cat glint360k_* | tar -xzvf -
+
+# Don't forget the last '-'!
+
+# cf7433cbb915ac422230ba33176f4625  glint360k_00
+# 589a5ea3ab59f283d2b5dd3242bc027a  glint360k_01
+# 8d54fdd5b1e4cd55e1b9a714d76d1075  glint360k_02
+# cd7f008579dbed9c5af4d1275915d95e  glint360k_03
+# 64666b324911b47334cc824f5f836d4c  glint360k_04
+# a318e4d32493dd5be6b94dd48f9943ac  glint360k_05
+# c3ae1dcbecea360d2ec2a43a7b6f1d94  glint360k_06
 # md5sum:
-# train.rec 2a74c71c4d20e770273f103eda97e878
-# train.idx f7a3e98d3533ac481bdf3dc03a5416e8
+# 5d9cd9f262ec87a5ca2eac5e703f7cdf train.idx
+# 8483be5af6f9906e19f85dee49132f8e train.rec
 ```
 Use [unpack_glint360k.py](./unpack_glint360k.py) to unpack.
 
+### 5. Pretrain models
+- [x] [**Baidu Drive**](https://pan.baidu.com/s/1sd9ZRsV2c_dWHW84kz1P1Q) (code:befi)
+- [x] [**Google Drive**](https://drive.google.com/drive/folders/1WLjDzEs1wC1K1jxDHNJ7dhEmQ3rOOILl?usp=sharing)
+
+| Framework       |  backbone                      | negative class centers sample_rate  | IJBC@e4 | IFRT@e6 |
+|  :---           | :---                           | :---         |  :---   |  :---   | 
+|  mxnet   | [R100](https://drive.google.com/drive/folders/1YPqIkOZWrmbli4GWfMJO2b0yiiZ7UCsP?usp=sharing) |1.0|97.3|-|
+|  mxnet   | [R100](https://drive.google.com/drive/folders/1-gF5sDwNoRcjwmpPSTNLpaZJi5N91BvL?usp=sharing) |0.1|97.3|-|
+|  pytorch | [R50](https://drive.google.com/drive/folders/16hjOGRJpwsJCRjIBbO13z3SrSgvPTaMV?usp=sharing) |1.0|97.0|-|    
+|  pytorch | [R100](https://drive.google.com/drive/folders/19EHffHN0Yn8DjYm5ofrgVOf_xfkrVgqc?usp=sharing) |1.0|97.4|-|
+    
+## Docker
+Make sure you have installed the NVIDIA driver and Docker engine for your Linux distribution Note that you do not need to 
+install the CUDA Toolkit and other independence on the host system, but the NVIDIA driver needs to be installed.  
+Because the CUDA version used in the image is 10.1, 
+the graphics driver version on the physical machine must be greater than 418.
+
+### 1. Docker Getting Started
+You can use dockerhub or offline docker.tar to get the image of the Partial-fc.
+1. dockerhub
+```shell
+docker pull insightface/partial_fc:v1
+```  
+
+2. offline images  
+coming soon!
+
+### 2. Getting Started
+```shell
+sudo docker run -it -v /train_tmp:/train_tmp --net=host --privileged --gpus 8 --shm-size=1g insightface/partial_fc:v1 /bin/bash
+```
 
-## Performance
+`/train_tmp` is where you put your training set (if you have enough RAM memory, 
+you can turn it into `tmpfs` first).
+
+## Benchmark
+### 1. Train Glint360K Using MXNET
+ 
+| Backbone    |   GPU                       | FP16  | BatchSize / it | Throughput img / sec |
+|  :---       | :---                        | :---  |   :---         | :---                 | 
+|  R100       | 8 * Tesla V100-SXM2-32GB    | False |   64           | 1748                 |
+|  R100       | 8 * Tesla V100-SXM2-32GB    | True  |   64           | 3357                 |
+|  R100       | 8 * Tesla V100-SXM2-32GB    | False |   128          | 1847                 |    
+|  R100       | 8 * Tesla V100-SXM2-32GB    | True  |   128          | 3867                 |   
+|  R50        | 8 * Tesla V100-SXM2-32GB    | False |   64           | 2921                 |
+|  R50        | 8 * Tesla V100-SXM2-32GB    | True  |   64           | 5428                 |
+|  R50        | 8 * Tesla V100-SXM2-32GB    | False |   128          | 3045                 |    
+|  R50        | 8 * Tesla V100-SXM2-32GB    | True  |   128          | 6112                 |  
+
+
+### 2. Performance On Million Identities
 We neglect the influence of IO. All experiments use mixed-precision training, and the backbone is ResNet50.
 #### 1 Million Identities On 8 RTX2080Ti  
 
@@ -63,19 +132,50 @@ We neglect the influence of IO. All experiments use mixed-precision training, an
 | **Partial FC(Ours)**      | **64**      | **4096**      | **6722**      | **12600**          | GPU   |
 
 
+## FAQ
+#### Glint360K's Face Alignment Settings?
+We use a same alignment setting with MS1MV2, code is [here](https://github.com/deepinsight/insightface/issues/1286).
+
+#### Why update Glint360K, is there a bug in the previous version?  
+In the previous version of Glint360K, there is no bug when using softmax training, but there is a bug in triplet training. 
+In the latest Glint360k, this bug has been fixed.
+
+#### Dataset in Google Drive or Dropbox?
+The torrent has been released.
+
+
+## Dataset Contributors
+
+The Glint360K would not have been possible without the invaluable contributions of the following individuals, who have been instrumental in data scraping, downloading, merging, and cleaning. Their efforts have provided the foundation for our work and have been essential to our success:
+
+| Contributor      | Emial    |
+|------------------|----------|
+| **Bin Qin**    | skyqin@gmail.com     |
+| **Lan Wu**     | bah-wl@hotmail.com   |
+
+
+
+Thank you to all the contributors for their hard work and dedication!
+
 
 ## Citation
-If you find Partial-FC or Glint360K useful in your research, please consider to cite the following related paper: 
+If you find [Partial-FC](https://arxiv.org/abs/2203.15565) or Glint360K useful in your research, please consider to cite the following related paper: 
+
 
-[Partial FC](https://arxiv.org/abs/2010.05222)
 ```
-@inproceedings{an2020partical_fc,
-  title={Partial FC: Training 10 Million Identities on a Single Machine},
-  author={An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and
-  Zhang, Debing and Fu Ying},
-  booktitle={Arxiv 2010.05222},
-  year={2020}
+@inproceedings{an_2022_pfc_cvpr,
+    title={Killing Two Birds with One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC},
+    author={An, Xiang and Deng, Jiangkang and Guo, Jia and Feng, Ziyong and Zhu, Xuhan and Jing, Yang and Tongliang, Liu},
+    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
+    year={2022}
 }
+@inproceedings{an_2021_pfc_iccvw,
+    title={Partial FC: Training 10 Million Identities on a Single Machine},
+    author={An, Xiang and Zhu, Xuhan and Gao, Yuan and Xiao, Yang and Zhao, Yongle and Feng, Ziyong and Wu, Lan and Qin, Bin and Zhang, Ming and Zhang, Debing and Fu, Ying},
+    booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
+    year={2021},
+}
+
 ```
 
 
diff --git a/recognition/partial_fc/mxnet/README.md b/recognition/partial_fc/mxnet/README.md
index b95304feb..930c8c210 100644
--- a/recognition/partial_fc/mxnet/README.md
+++ b/recognition/partial_fc/mxnet/README.md
@@ -1,5 +1,5 @@
-## Train
-#### Requirements
+## Training
+### 1.Requirements
 python==3.6  
 cuda==10.1    
 cudnn==765    
@@ -9,10 +9,7 @@ pip install easydict mxboard opencv-python tqdm
 [openmpi](mxnet/setup-utils/install-mpi.sh)==4.0.0  
 [horovod](mxnet/setup-utils/install-horovod.sh)==0.19.2  
 
-#### Failures due to SSH issues
-The host where horovodrun is executed must be able to SSH to all other hosts without any prompts.
-
-#### Run with horovodrun
+### 2.Run with horovodrun
 Typically one GPU will be allocated per process, so if a server has 8 GPUs, you will run 8 processes. 
 In horovodrun, the number of processes is specified with the -np flag.
 
@@ -26,7 +23,97 @@ To run on two machine with 16 GPUs:
 horovodrun -np 16 -H ip1:8,ip2:8 bash config.sh
 ```
 
-#### Run with mpi
+### 3.Run with mpi
 ```shell script
 bash run.sh
-```
\ No newline at end of file
+```
+
+### Failures due to SSH issues
+The host where horovodrun is executed must be able to SSH to all other hosts without any prompts.
+
+
+
+
+## Troubleshooting
+
+### 1. Horovod installed successfully?  
+
+Run `horovodrun --check` to check the installation of horovod.
+```shell script
+# Horovod v0.19.2:
+# 
+# Available Frameworks:
+#     [ ] TensorFlow
+#     [X] PyTorch
+#     [X] MXNet
+# 
+# Available Controllers:
+#     [X] MPI
+#     [X] Gloo
+# 
+# Available Tensor Operations:
+#     [X] NCCL
+#     [ ] DDL
+#     [ ] CCL
+#     [X] MPI
+#     [X] Gloo
+```
+
+### 2. Mxnet Version!
+Some versions of mxnet with horovod have bug.   
+It is recommended to try version **1.5 or 1.6**.
+
+**The community has found that mxnet1.5.1 cannot install horovod.**
+
+### 3. Check CUDA version!
+```shell script
+# Make sure your cuda version is same as mxnet, such as mxnet-cu101 (CUDA 10.1)
+
+/usr/local/cuda/bin/nvcc -V
+# nvcc: NVIDIA (R) Cuda compiler driver
+# Copyright (c) 2005-2019 NVIDIA Corporation
+# Built on Wed_Apr_24_19:10:27_PDT_2019
+# Cuda compilation tools, release 10.1, V10.1.168
+```
+
+### 4. Block IO
+You can turn on the debug mode to check whether your slow training speed is the cause of IO.
+
+### 5. Training Speed.
+If you find that your training speed is the io bottleneck, you can mount dataset to RAM, 
+using the following command.
+```shell script
+# If your RAM has 256G
+sudo mkdir /train_tmp
+mount -t tmpfs -o size=140G  tmpfs /train_tmp
+```
+
+## Our Method
+![Image text](https://github.com/nttstar/insightface-resources/blob/master/images/partial_fc.png)
+
+### 1. The classification layer model is parallel
+Class centers are evenly distributed across different GPUs. It only takes three communications to complete 
+loss-free Softmax calculations.
+
+#### 1. Synchronization of features
+Make sure each GPU has all the GPU features on it, as is shown in `AllGather(x_i)`.
+
+#### 2. Synchronization of denominator of the softmax function
+We can first calculate the local sum of each GPU, and then compute the global sum through communication, as is shown
+in `Allreduce(sum(exp(logits_i)))`
+
+#### 3. Synchronization the gradients of feature
+The gradient of logits can be calculated independently, so is the gradient of the feature. finally, we collect all the 
+gradients on GPU and send them back to backbone, as is shown in `Allreduce(deta(X))`
+
+### 2. Softmax approximate
+
+Just a subset of class centers can approximate softmax's computation(positive class centers must in these class centers),
+this can be done with the following code:
+```python
+centers_p = func_positive(label)                 # select the positive class centers by the label of the sample
+centers_n = func_negative(centers_p)             # negative class centers are randomly sampled after excluding positive classes
+centers_final = concat(centers_n, centers_p)     # class centers that participate in softmax calculations
+```
+
+
diff --git a/recognition/partial_fc/mxnet/README_CN.md b/recognition/partial_fc/mxnet/README_CN.md
new file mode 100644
index 000000000..8c63462ea
--- /dev/null
+++ b/recognition/partial_fc/mxnet/README_CN.md
@@ -0,0 +1,124 @@
+## 目录
+## Contents
+[Partial FC](https://arxiv.org/abs/2203.15565)
+- [如何安装](#如何安装)
+- [如何运行](#如何运行)
+- [错误排查](#错误排查)
+
+
+## 如何安装
+ 
+### 1. python依赖  
+使用以下命令
+```shell script
+pip install easydict mxboard opencv-python tqdm     
+```
+
+### 2. 安装nccl  
+nccl可以不用装，但是装上速度更快，nccl安装需要对应cuda版本，安装方法参考下边链接:  
+[**NCCL**](https://docs.nvidia.com/deeplearning/nccl/install-guide/index.html)  
+
+### 3. 安装openmpi  
+openmpi必须安装，必须采纳我的脚本编译源码安装：  
+[**OpenMPI**](setup-utils/install-mpi.sh)    
+
+### 4. 安装horovod, mxnet
+有些版本的mxnet的horovod无法安装，参考下方表格，强烈建议使用**mxnet==1.6.0**和**cuda==10.1**
+
+| mxnet |horovod  |  cuda        | 
+| :---: | :---    |  :---:       | 
+| 1.4.0 | x       |  x           | 
+| 1.5.0 | 可以安装 | cuda10.0     | 
+| 1.5.1 | x       | x            | 
+| 1.6.0.post0 | 可以安装 | cuda10.1     | 
+| 1.7.0 | x       | x            | 
+
+horovod 安装方法如下:  
+[**Horovod**](setup-utils/install-horovod.sh)
+
+horovod 安装完成后使用下面的命令检查horovod是否安装成功，(nccl有没有都可以，有nccl会更快)：
+```shell script
+# Horovod v0.19.2:
+# Available Frameworks:
+#     [ ] TensorFlow
+#     [ ] PyTorch
+#     [X] MXNet
+# 
+# Available Controllers:
+#     [X] MPI
+#     [X] Gloo
+# 
+# Available Tensor Operations:
+#     [X] NCCL
+#     [ ] DDL
+#     [ ] CCL
+#     [X] MPI
+#     [X] Gloo
+```
+
+
+### 5. ssh无密登录
+
+使用多机分布式训练的时候，每台机器都需要设置无密登录，包括自己与自己，无密码登录具体可见：  
+这里推荐一个简单的命令：  
+```shell script
+ssh-copy-id user@ip
+```
+
+## 如何运行  
+`horovod`底层调用的还是`mpi`，mpi的概念是，你有多少块GPU，就要启动多少个进程，有两种方法启动训练，使用`horovodrun`或者`mpirun`。  
+### 1. 使用 horovodrun 运行  
+
+运行8卡(单机)：
+```shell script
+horovodrun -np 8 -H localhost:8 bash config.sh
+```
+
+运行16卡(两台机器)
+```shell script
+horovodrun -np 16 -H ip1:8,ip2:8 bash config.sh
+```
+
+### 2. 使用 mpirun 运行  
+
+```shell script
+bash run.sh
+```
+
+## 错误排查
+
+QQ群：711302608  
+
+### 检查Horovod是否安装成功？
+
+运行这个命令 `horovodrun --check` 来检查horovod是否安装成功。
+
+### 检查你的CUDA版本是否与mxnet匹配，比如mxnet-cu101需要的cuda版本为CUDA10.1  
+
+```shell script
+# Make sure your cuda version is same as mxnet, such as mxnet-cu101 (CUDA 10.1)
+
+/usr/local/cuda/bin/nvcc -V
+# nvcc: NVIDIA (R) Cuda compiler driver
+# Copyright (c) 2005-2019 NVIDIA Corporation
+# Built on Wed_Apr_24_19:10:27_PDT_2019
+# Cuda compilation tools, release 10.1, V10.1.168
+```
+
+### 屏蔽IO对训练速度的影响？  
+
+可以在`config.py`中开启debug模式，来屏蔽IO，看看是否是IO对性能的影响。
+
+### 将数据挂载到内存盘来提高训练速度。
+
+如果你发现你训练速度的瓶颈是IO的话，你可以把数据挂载到内存盘来提高训练的速度，挂载的命令如下：  
+需要注意的是，你的RAM必须足够的大。
+
+```shell script
+# If your RAM has 256G
+sudo mkdir /train_tmp
+mount -t tmpfs -o size=140G  tmpfs /train_tmp
+```
+
+
+
diff --git a/recognition/partial_fc/mxnet/callbacks.py b/recognition/partial_fc/mxnet/callbacks.py
index 10659bd24..f2c93ef19 100644
--- a/recognition/partial_fc/mxnet/callbacks.py
+++ b/recognition/partial_fc/mxnet/callbacks.py
@@ -44,7 +44,8 @@ def __init__(self, symbol, model):
         self.model = model
         self.ver_list = []
         self.ver_name_list = []
-        self.init_dataset(val_targets=config.val_targets, data_dir= os.path.dirname(config.rec),
+        self.init_dataset(val_targets=config.val_targets,
+                          data_dir=os.path.dirname(config.rec),
                           image_size=(config.image_size, config.image_size))
 
     def ver_test(self, num_update):
@@ -52,12 +53,15 @@ def ver_test(self, num_update):
         for i in range(len(self.ver_list)):
             acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
                 self.ver_list[i], self.model, 10, 10, None, None)
-            logging.info('[%s][%d]XNorm: %f' % (self.ver_name_list[i], num_update, xnorm))
-            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (self.ver_name_list[i], num_update, acc2, std2))
+            logging.info('[%s][%d]XNorm: %f' %
+                         (self.ver_name_list[i], num_update, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                         (self.ver_name_list[i], num_update, acc2, std2))
             if acc2 > self.highest_acc_list[i]:
                 self.highest_acc_list[i] = acc2
             logging.info(
-                '[%s][%d]Accuracy-Highest: %1.5f' % (self.ver_name_list[i], num_update, self.highest_acc_list[i]))
+                '[%s][%d]Accuracy-Highest: %1.5f' %
+                (self.ver_name_list[i], num_update, self.highest_acc_list[i]))
             results.append(acc2)
 
     def init_dataset(self, val_targets, data_dir, image_size):
@@ -98,7 +102,9 @@ def __init__(self, symbol, model, prefix, rank):
     def __call__(self, param):
         num_update = param.num_update
 
-        if num_update in [self.max_step - 10, ] or (num_update % 10000 == 0 and num_update > 0):
+        if num_update in [
+            self.max_step - 10,
+        ] or (num_update % 10000 == 0 and num_update > 0):
 
             # params
             arg, aux = self.model.get_export_params()
@@ -114,16 +120,16 @@ def __call__(self, param):
                 new_arg[key] = hvd.allreduce(tensor, average=True)
 
             if self.rank == 0:
-                mx.model.save_checkpoint(
-                    prefix=self.prefix + "_average",
-                    epoch=0, symbol=_sym,
-                    arg_params=new_arg,
-                    aux_params=new_aux)
-                mx.model.save_checkpoint(
-                    prefix=self.prefix,
-                    epoch=0, symbol=_sym,
-                    arg_params=arg,
-                    aux_params=aux)
+                mx.model.save_checkpoint(prefix=self.prefix + "_average",
+                                         epoch=0,
+                                         symbol=_sym,
+                                         arg_params=new_arg,
+                                         aux_params=new_aux)
+                mx.model.save_checkpoint(prefix=self.prefix,
+                                         epoch=0,
+                                         symbol=_sym,
+                                         arg_params=arg,
+                                         aux_params=aux)
 
         # training is over
         if num_update > self.max_step > 0:
@@ -164,9 +170,13 @@ def __init__(self, rank, size, prefix_dir):
         self.loss_metric = MetricNdarray()
         t = time.localtime()
 
-        self.summary_writer = SummaryWriter(
-            logdir=os.path.join(self.prefix_dir, "log_tensorboard",
-            "%s_%s_%s" % (str(t.tm_mon), str(t.tm_mday), str(t.tm_hour))), verbose=False)
+        if self.rank == 0:
+            self.summary_writer = SummaryWriter(logdir=os.path.join(
+                self.prefix_dir, "log_tensorboard",
+                "%s_%s_%s" % (str(t.tm_mon), str(t.tm_mday), str(t.tm_hour))),
+                verbose=False)
+        else:
+            time.sleep(2)
 
     def __call__(self, param):
         """Callback to Show speed
@@ -191,19 +201,20 @@ def __call__(self, param):
 
                 # summary loss
                 loss_scalar = self.loss_metric.get()
-                self.summary_writer.add_scalar(tag="loss",
-                    value=loss_scalar, global_step=param.num_update)
-                loss_str_format = "[%d][%s]:%.2f " % (param.num_epoch, "loss", loss_scalar)
+
+                if self.rank == 0:
+                    self.summary_writer.add_scalar(tag="loss", value=loss_scalar, global_step=param.num_update)
+                loss_str_format = "[%d][%s]:%.2f " % (param.num_epoch, "loss",
+                                                      loss_scalar)
                 self.loss_metric.reset()
-                # summary speed
-                self.summary_writer.add_scalar(
-                    tag="speed",
-                    value=speed, global_step=param.num_update)
-                self.summary_writer.flush()
+
                 if self.rank == 0:
+                    self.summary_writer.add_scalar(tag="speed", value=speed, global_step=param.num_update)
+                    self.summary_writer.flush()
                     logging.info(
                         "Iter:%d Rank:%.2f it/sec Total:%.2f it/sec %s",
                         param.num_update, speed, speed_total, loss_str_format)
+
                 self.tic = time.time()
         else:
             self.init = True
diff --git a/recognition/partial_fc/mxnet/config.sh b/recognition/partial_fc/mxnet/config.sh
index bdf4c6e3e..17e58fba0 100644
--- a/recognition/partial_fc/mxnet/config.sh
+++ b/recognition/partial_fc/mxnet/config.sh
@@ -6,9 +6,9 @@ export MXNET_CPU_WORKER_NTHREADS=3
 
 # use `which python` to get the absolute path of your python interpreter
 #
-PYTHON_EXEC=/home/face/miniconda3/envs/nist_pf/bin/python
+PYTHON_EXEC=/usr/bin/python
 ${PYTHON_EXEC} train_memory.py \
 --dataset glint360k_8GPU \
---loss arcface \
+--loss cosface \
 --network r100 \
---models-root /data/anxiang/opensource/glint360k_8GPU_r100FC_1.0_fp16_arcface
+--models-root /data/anxiang/opensource/glint360k_8GPU_r100FC_1.0_fp32_cosface
diff --git a/recognition/partial_fc/mxnet/default.py b/recognition/partial_fc/mxnet/default.py
index d1c34a442..10b01f741 100644
--- a/recognition/partial_fc/mxnet/default.py
+++ b/recognition/partial_fc/mxnet/default.py
@@ -16,12 +16,13 @@
 config.memonger = False
 
 config.debug = 0
-config.fp16 = True
+config.fp16 = False
 config.batch_size = 64
 config.backbone_lr = 0.1
 config.memory_bank_lr = config.backbone_lr
 config.sample_ratio = 1.0
 
+
 def generate_config(loss_name, dataset, network):
 
     # loss
@@ -47,8 +48,8 @@ def generate_config(loss_name, dataset, network):
 
     # glint360k 17091657
     # md5sum:
-    # train.rec 2a74c71c4d20e770273f103eda97e878
-    # train.idx f7a3e98d3533ac481bdf3dc03a5416e8
+    # 5d9cd9f262ec87a5ca2eac5e703f7cdf train.idx
+    # 8483be5af6f9906e19f85dee49132f8e train.rec
 
     # make training faster
     # our RAM is 256G
@@ -56,8 +57,10 @@ def generate_config(loss_name, dataset, network):
 
     elif dataset == 'glint360k_8GPU':
         config.lr_steps = '200000,400000,500000,550000'
-        config.val_targets = ['agedb_30', 'calfw', 'cfp_ff',  'cfp_fp', 'cplfw', 'lfw', 'vgg2_fp']
-        config.rec = '/train_tmp/celeb_deepglint/train.rec'
+        config.val_targets = [
+            'agedb_30', 'calfw', 'cfp_ff', 'cfp_fp', 'cplfw', 'lfw', 'vgg2_fp'
+        ]
+        config.rec = '/train_tmp/glint360k/train.rec'
         config.num_classes = 360232
         config.batch_size = 64
         config.max_update = 600000
@@ -65,7 +68,7 @@ def generate_config(loss_name, dataset, network):
     elif dataset == 'glint360k_16GPU':
         config.lr_steps = '200000,280000,360000'
         config.val_targets = ['agedb_30', 'cfp_fp', 'lfw']
-        config.rec = '/train_tmp/celeb_deepglint/train.rec'
+        config.rec = '/train_tmp/glint360k/train.rec'
         config.num_classes = 360232
         config.max_update = 400000
 
diff --git a/recognition/partial_fc/mxnet/evaluation/align_ijb.py b/recognition/partial_fc/mxnet/evaluation/align_ijb.py
index b0009c1b1..fb60abda2 100644
--- a/recognition/partial_fc/mxnet/evaluation/align_ijb.py
+++ b/recognition/partial_fc/mxnet/evaluation/align_ijb.py
@@ -4,12 +4,9 @@
 import numpy as np
 from skimage import transform as trans
 
-src = np.array([
-    [30.2946, 51.6963],
-    [65.5318, 51.5014],
-    [48.0252, 71.7366],
-    [33.5493, 92.3655],
-    [62.7299, 92.2041]], dtype=np.float32)
+src = np.array([[30.2946, 51.6963], [65.5318, 51.5014], [48.0252, 71.7366],
+                [33.5493, 92.3655], [62.7299, 92.2041]],
+               dtype=np.float32)
 src[:, 0] += 8.0
 
 img_path = '/data/anxiang/datasets/IJB_release/IJBC/loose_crop'
@@ -25,7 +22,8 @@
     name_lmk_score = each_line.strip().split(' ')
     img_name = os.path.join(img_path, name_lmk_score[0])
     img = cv2.imread(img_name)
-    landmark = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)
+    landmark = np.array([float(x) for x in name_lmk_score[1:-1]],
+                        dtype=np.float32)
     landmark = landmark.reshape((5, 2))
 
     if landmark.shape[0] == 68:
diff --git a/recognition/partial_fc/mxnet/evaluation/example.sh b/recognition/partial_fc/mxnet/evaluation/example.sh
old mode 100644
new mode 100755
index a132a7769..63427748a
--- a/recognition/partial_fc/mxnet/evaluation/example.sh
+++ b/recognition/partial_fc/mxnet/evaluation/example.sh
@@ -1,9 +1,14 @@
+#!/bin/bash
+
+# run `python ijb.py --help` for more information
 python -u ijb.py \
---model-prefix /anxiang/opensource/model/celeb360k_final0.1/model \
---image-path /data/anxiang/datasets/IJB_release/IJBC \
---result-dir /anxiang/opensource/results/test \
+--model-prefix /home/face/insightface/recognition/partial_fc/mxnet/evaluation/glint360k_r100FC_0.1/model \
+--image-path /data/anxiang/IJB_release/IJBC \
+--result-dir ./results/test \
 --model-epoch 0 \
---gpu 0,1,2,3,4,5,6,7 \
+--gpu 0,1,2,3 \
 --target IJBC \
- --job cosface \
---batch-size 256
\ No newline at end of file
+--job partial_fc \
+--batch-size 256 \
+-es 512
+
diff --git a/recognition/partial_fc/mxnet/evaluation/ijb.py b/recognition/partial_fc/mxnet/evaluation/ijb.py
index 64a560d92..ffcabd69e 100644
--- a/recognition/partial_fc/mxnet/evaluation/ijb.py
+++ b/recognition/partial_fc/mxnet/evaluation/ijb.py
@@ -32,31 +32,33 @@
 parser.add_argument('--gpu', default='0', type=str, help='gpu id')
 parser.add_argument('--batch-size', default=128, type=int, help='')
 parser.add_argument('--job', default='insightface', type=str, help='job name')
-parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+parser.add_argument('-es', '--emb-size', type=int, help='embedding size')
+parser.add_argument('--target',
+                    default='IJBC',
+                    type=str,
+                    help='target, set to IJBC or IJBB')
 args = parser.parse_args()
 
 os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
 
-target               = args.target
-model_path           = args.model_prefix
-image_path           = args.image_path
-result_dir           = args.result_dir
-epoch                = args.model_epoch
-use_norm_score       = True  # if Ture, TestMode(N1)
-use_detector_score   = True  # if Ture, TestMode(D1)
-use_flip_test        = True  # if Ture, TestMode(F1)
-job                  = args.job
-batch_size           = args.batch_size
+target = args.target
+model_path = args.model_prefix
+image_path = args.image_path
+result_dir = args.result_dir
+epoch = args.model_epoch
+use_norm_score = True  # if Ture, TestMode(N1)
+use_detector_score = True  # if Ture, TestMode(D1)
+use_flip_test = True  # if Ture, TestMode(F1)
+job = args.job
+batch_size = args.batch_size
 
 
 class DatasetIJB(Dataset):
     def __init__(self, root, lines, align=True):
-        self.src = np.array([
-            [30.2946, 51.6963],
-            [65.5318, 51.5014],
-            [48.0252, 71.7366],
-            [33.5493, 92.3655],
-            [62.7299, 92.2041]], dtype=np.float32)
+        self.src = np.array(
+            [[30.2946, 51.6963], [65.5318, 51.5014], [48.0252, 71.7366],
+             [33.5493, 92.3655], [62.7299, 92.2041]],
+            dtype=np.float32)
         self.src[:, 0] += 8.0
         self.lines = lines
         self.img_root = root
@@ -72,7 +74,8 @@ def __getitem__(self, idx):
         img = cv2.imread(img_name)
 
         if self.align:
-            landmark = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32)
+            landmark = np.array([float(x) for x in name_lmk_score[1:-1]],
+                                dtype=np.float32)
             landmark = landmark.reshape((5, 2))
             #
             assert landmark.shape[0] == 68 or landmark.shape[0] == 5
@@ -112,31 +115,40 @@ def extract_parallel(prefix, epoch, dataset, batch_size, size):
     def batchify_fn(data):
         return mx.nd.concat(*data, dim=0)
 
-    data_loader = DataLoader(dataset, batch_size, last_batch='keep', num_workers=8, thread_pool=True,
-        prefetch=64, batchify_fn=batchify_fn)
-    symbol, arg_params, aux_params = mx.module.module.load_checkpoint(prefix, epoch)
+    data_loader = DataLoader(dataset,
+                             batch_size,
+                             last_batch='keep',
+                             num_workers=8,
+                             thread_pool=True,
+                             prefetch=16,
+                             batchify_fn=batchify_fn)
+    symbol, arg_params, aux_params = mx.module.module.load_checkpoint(
+        prefix, epoch)
+    all_layers = symbol.get_internals()
+    symbol = all_layers['fc1_output']
 
     # init model list
     for i in range(num_ctx):
         model = mx.mod.Module(symbol, context=mx.gpu(i), label_names=None)
-        model.bind(for_training=False, data_shapes=[('data', (2 * batch_size, 3, 112, 112))])
+        model.bind(for_training=False,
+                   data_shapes=[('data', (2 * batch_size, 3, 112, 112))])
         model.set_params(arg_params, aux_params)
         model_list.append(model)
 
     # extract parallel and async
     num_model = len(model_list)
     for image in tqdm(data_loader):
-        data_batch = mx.io.DataBatch(data=(image,))
+        data_batch = mx.io.DataBatch(data=(image, ))
         model_list[num_iter % num_model].forward(data_batch, is_train=False)
-        feat = model_list[num_iter % num_model].get_outputs(merge_multi_context=True)[0]
+        feat = model_list[num_iter %
+                          num_model].get_outputs(merge_multi_context=True)[0]
         feat = mx.nd.L2Normalization(feat)
         feat = mx.nd.reshape(feat, (-1, size * 2))
-        feat_mat[
-        batch_size * num_iter:
-        batch_size * num_iter + feat.shape[0], :] = feat.as_in_context(mx.cpu())
+        feat_mat[batch_size * num_iter:batch_size * num_iter +
+                 feat.shape[0], :] = feat.as_in_context(mx.cpu())
         num_iter += 1
-        # if num_iter % 20 == 0:
-        #     mx.nd.waitall()
+        #if num_iter % 20 == 0:
+        #    mx.nd.waitall()
     return feat_mat.asnumpy()
 
 
@@ -180,22 +192,26 @@ def image2template_feature(img_feats=None, templates=None, medias=None):
 
     for count_template, uqt in enumerate(unique_templates):
 
-        (ind_t,) = np.where(templates == uqt)
+        (ind_t, ) = np.where(templates == uqt)
         face_norm_feats = img_feats[ind_t]
         face_medias = medias[ind_t]
-        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        unique_medias, unique_media_counts = np.unique(face_medias,
+                                                       return_counts=True)
         media_norm_feats = []
         for u, ct in zip(unique_medias, unique_media_counts):
-            (ind_m,) = np.where(face_medias == u)
+            (ind_m, ) = np.where(face_medias == u)
             if ct == 1:
                 media_norm_feats += [face_norm_feats[ind_m]]
             else:  # image features from the same video will be aggregated into one feature
-                media_norm_feats += [np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)]
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)
+                ]
         media_norm_feats = np.array(media_norm_feats)
         # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
         template_feats[count_template] = np.sum(media_norm_feats, axis=0)
         if count_template % 2000 == 0:
-            print('Finish Calculating {} template features.'.format(count_template))
+            print('Finish Calculating {} template features.'.format(
+                count_template))
     # template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
     template_norm_feats = sklearn.preprocessing.normalize(template_feats)
     # print(template_norm_feats.shape)
@@ -205,7 +221,10 @@ def image2template_feature(img_feats=None, templates=None, medias=None):
 # In[ ]:
 
 
-def verification(template_norm_feats=None, unique_templates=None, p1=None, p2=None):
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
     # ==========================================================
     #         Compute set-to-set Similarity Score.
     # ==========================================================
@@ -213,11 +232,13 @@ def verification(template_norm_feats=None, unique_templates=None, p1=None, p2=No
     for count_template, uqt in enumerate(unique_templates):
         template2id[uqt] = count_template
 
-    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    score = np.zeros((len(p1), ))  # save cosine distance between pairs
 
     total_pairs = np.array(range(len(p1)))
     batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
-    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
     total_sublists = len(sublists)
     for c, s in enumerate(sublists):
         feat1 = template_norm_feats[template2id[p1[s]]]
@@ -230,14 +251,19 @@ def verification(template_norm_feats=None, unique_templates=None, p1=None, p2=No
 
 
 # In[ ]:
-def verification2(template_norm_feats=None, unique_templates=None, p1=None, p2=None):
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
     template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
     for count_template, uqt in enumerate(unique_templates):
         template2id[uqt] = count_template
-    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    score = np.zeros((len(p1), ))  # save cosine distance between pairs
     total_pairs = np.array(range(len(p1)))
     batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
-    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
     total_sublists = len(sublists)
     for c, s in enumerate(sublists):
         feat1 = template_norm_feats[template2id[p1[s]]]
@@ -267,11 +293,11 @@ def read_score(path):
 # =============================================================
 start = timeit.default_timer()
 templates, medias = read_template_media_list(
-    os.path.join('%s/meta' % image_path, '%s_face_tid_mid.txt' % target.lower()))
+    os.path.join('%s/meta' % image_path,
+                 '%s_face_tid_mid.txt' % target.lower()))
 stop = timeit.default_timer()
 print('Time: %.2f s. ' % (stop - start))
 
-
 # =============================================================
 # load template pairs for template-to-template verification
 # tid : template id,  label : 1/0
@@ -280,7 +306,8 @@ def read_score(path):
 # =============================================================
 start = timeit.default_timer()
 p1, p2, label = read_template_pair_list(
-    os.path.join('%s/meta' % image_path, '%s_template_pair_label.txt' % target.lower()))
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % target.lower()))
 stop = timeit.default_timer()
 print('Time: %.2f s. ' % (stop - start))
 
@@ -297,7 +324,11 @@ def read_score(path):
 img_list = open(img_list_path)
 files = img_list.readlines()
 dataset = DatasetIJB(root=img_path, lines=files, align=True)
-img_feats = extract_parallel(args.model_prefix, args.model_epoch, dataset, args.batch_size, size=512)
+img_feats = extract_parallel(args.model_prefix,
+                             args.model_epoch,
+                             dataset,
+                             args.batch_size,
+                             size=args.emb_size)
 
 faceness_scores = []
 for each_line in files:
@@ -306,16 +337,15 @@ def read_score(path):
 
 faceness_scores = np.array(faceness_scores).astype(np.float32)
 
-
 stop = timeit.default_timer()
 print('Time: %.2f s. ' % (stop - start))
-print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
+print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0],
+                                          img_feats.shape[1]))
 
 # # Step3: Get Template Features
 
 # In[ ]:
 
-
 # =============================================================
 # compute template features from image features.
 # =============================================================
@@ -327,12 +357,12 @@ def read_score(path):
 # 1. FaceScore （Feature Norm）
 # 2. FaceScore （Detector）
 
-
 if use_flip_test:
     # concat --- F1
     # img_input_feats = img_feats
     # add --- F2
-    img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2] + img_feats[:, img_feats.shape[1] // 2:]
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] //
+                                2] + img_feats[:, img_feats.shape[1] // 2:]
 else:
     img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
 
@@ -340,7 +370,8 @@ def read_score(path):
     img_input_feats = img_input_feats
 else:
     # normalise features to remove norm information
-    img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+    img_input_feats = img_input_feats / np.sqrt(
+        np.sum(img_input_feats**2, -1, keepdims=True))
 
 if use_detector_score:
     print(img_input_feats.shape, faceness_scores.shape)
@@ -349,7 +380,8 @@ def read_score(path):
 else:
     img_input_feats = img_input_feats
 
-template_norm_feats, unique_templates = image2template_feature(img_input_feats, templates, medias)
+template_norm_feats, unique_templates = image2template_feature(
+    img_input_feats, templates, medias)
 stop = timeit.default_timer()
 print('Time: %.2f s. ' % (stop - start))
 
@@ -357,7 +389,6 @@ def read_score(path):
 
 # In[ ]:
 
-
 # =============================================================
 # compute verification scores between template pairs.
 # =============================================================
@@ -380,7 +411,6 @@ def read_score(path):
 
 # In[ ]:
 
-
 files = [score_save_file]
 methods = []
 scores = []
@@ -390,9 +420,10 @@ def read_score(path):
 
 methods = np.array(methods)
 scores = dict(zip(methods, scores))
-colours = dict(zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
 # x_labels = [1/(10**x) for x in np.linspace(6, 0, 6)]
-x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+x_labels = [10**-6, 10**-5, 10**-4, 10**-3, 10**-2, 10**-1]
 tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
 fig = plt.figure()
 for method in methods:
@@ -400,16 +431,21 @@ def read_score(path):
     roc_auc = auc(fpr, tpr)
     fpr = np.flipud(fpr)
     tpr = np.flipud(tpr)  # select largest tpr at same fpr
-    plt.plot(fpr, tpr, color=colours[method], lw=1,
-        label=('[%s (AUC = %0.4f %%)]' % (method.split('-')[-1], roc_auc * 100)))
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
     tpr_fpr_row = []
     tpr_fpr_row.append("%s-%s" % (method, target))
     for fpr_iter in np.arange(len(x_labels)):
-        _, min_index = min(list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
         # tpr_fpr_row.append('%.4f' % tpr[min_index])
         tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
     tpr_fpr_table.add_row(tpr_fpr_row)
-plt.xlim([10 ** -6, 0.1])
+plt.xlim([10**-6, 0.1])
 plt.ylim([0.3, 1.0])
 plt.grid(linestyle='--', linewidth=1)
 plt.xticks(x_labels)
diff --git a/recognition/partial_fc/mxnet/evaluation/lfw.py b/recognition/partial_fc/mxnet/evaluation/lfw.py
index ecf2005b5..d900d7909 100644
--- a/recognition/partial_fc/mxnet/evaluation/lfw.py
+++ b/recognition/partial_fc/mxnet/evaluation/lfw.py
@@ -2,19 +2,19 @@
 """
 
 # MIT License
-# 
+#
 # Copyright (c) 2016 David Sandberg
-# 
+#
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
 # in the Software without restriction, including without limitation the rights
 # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 # copies of the Software, and to permit persons to whom the Software is
 # furnished to do so, subject to the following conditions:
-# 
+#
 # The above copyright notice and this permission notice shall be included in all
 # copies or substantial portions of the Software.
-# 
+#
 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
@@ -38,98 +38,118 @@
 from mxnet import ndarray as nd
 
 
-
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca = 0):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
     nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
     nrof_thresholds = len(thresholds)
     k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-    
-    tprs = np.zeros((nrof_folds,nrof_thresholds))
-    fprs = np.zeros((nrof_folds,nrof_thresholds))
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
     accuracy = np.zeros((nrof_folds))
     indices = np.arange(nrof_pairs)
     #print('pca', pca)
-    
-    if pca==0:
-      diff = np.subtract(embeddings1, embeddings2)
-      dist = np.sum(np.square(diff),1)
-    
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
     for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
         #print('train_set', train_set)
         #print('test_set', test_set)
-        if pca>0:
-          print('doing pca on', fold_idx)
-          embed1_train = embeddings1[train_set]
-          embed2_train = embeddings2[train_set]
-          _embed_train = np.concatenate( (embed1_train, embed2_train), axis=0 )
-          #print(_embed_train.shape)
-          pca_model = PCA(n_components=pca)
-          pca_model.fit(_embed_train)
-          embed1 = pca_model.transform(embeddings1)
-          embed2 = pca_model.transform(embeddings2)
-          embed1 = sklearn.preprocessing.normalize(embed1)
-          embed2 = sklearn.preprocessing.normalize(embed2)
-          #print(embed1.shape, embed2.shape)
-          diff = np.subtract(embed1, embed2)
-          dist = np.sum(np.square(diff),1)
-        
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            #print(_embed_train.shape)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            #print(embed1.shape, embed2.shape)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
         # Find the best threshold for the fold
         acc_train = np.zeros((nrof_thresholds))
         for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
         best_threshold_index = np.argmax(acc_train)
         for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-          
-    tpr = np.mean(tprs,0)
-    fpr = np.mean(fprs,0)
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
     return tpr, fpr, accuracy
 
+
 def calculate_accuracy(threshold, dist, actual_issame):
     predict_issame = np.less(dist, threshold)
     tp = np.sum(np.logical_and(predict_issame, actual_issame))
     fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
     fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
-  
-    tpr = 0 if (tp+fn==0) else float(tp) / float(tp+fn)
-    fpr = 0 if (fp+tn==0) else float(fp) / float(fp+tn)
-    acc = float(tp+tn)/dist.size
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
     return tpr, fpr, acc
 
 
-  
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
     nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
     nrof_thresholds = len(thresholds)
     k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-    
+
     val = np.zeros(nrof_folds)
     far = np.zeros(nrof_folds)
-    
+
     diff = np.subtract(embeddings1, embeddings2)
-    dist = np.sum(np.square(diff),1)
+    dist = np.sum(np.square(diff), 1)
     indices = np.arange(nrof_pairs)
-    
+
     for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
+
         # Find the threshold that gives FAR = far_target
         far_train = np.zeros(nrof_thresholds)
         for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
-        if np.max(far_train)>=far_target:
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
             f = interpolate.interp1d(far_train, thresholds, kind='slinear')
             threshold = f(far_target)
         else:
             threshold = 0.0
-    
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
-  
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
     val_mean = np.mean(val)
     far_mean = np.mean(far)
     val_std = np.std(val)
@@ -139,49 +159,70 @@ def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_targe
 def calculate_val_far(threshold, dist, actual_issame):
     predict_issame = np.less(dist, threshold)
     true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
     n_same = np.sum(actual_issame)
     n_diff = np.sum(np.logical_not(actual_issame))
     val = float(true_accept) / float(n_same)
     far = float(false_accept) / float(n_diff)
     return val, far
 
-def evaluate(embeddings, actual_issame, nrof_folds=10, pca = 0):
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
     # Calculate evaluation metrics
     thresholds = np.arange(0, 4, 0.01)
     embeddings1 = embeddings[0::2]
     embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), nrof_folds=nrof_folds, pca = pca)
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
     thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
     return tpr, fpr, accuracy, val, val_std, far
 
+
 def get_paths(lfw_dir, pairs, file_ext):
     nrof_skipped_pairs = 0
     path_list = []
     issame_list = []
     for pair in pairs:
         if len(pair) == 3:
-            path0 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[1])+'.'+file_ext)
-            path1 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[2])+'.'+file_ext)
+            path0 = os.path.join(
+                lfw_dir, pair[0],
+                pair[0] + '_' + '%04d' % int(pair[1]) + '.' + file_ext)
+            path1 = os.path.join(
+                lfw_dir, pair[0],
+                pair[0] + '_' + '%04d' % int(pair[2]) + '.' + file_ext)
             issame = True
         elif len(pair) == 4:
-            path0 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[1])+'.'+file_ext)
-            path1 = os.path.join(lfw_dir, pair[2], pair[2] + '_' + '%04d' % int(pair[3])+'.'+file_ext)
+            path0 = os.path.join(
+                lfw_dir, pair[0],
+                pair[0] + '_' + '%04d' % int(pair[1]) + '.' + file_ext)
+            path1 = os.path.join(
+                lfw_dir, pair[2],
+                pair[2] + '_' + '%04d' % int(pair[3]) + '.' + file_ext)
             issame = False
-        if os.path.exists(path0) and os.path.exists(path1):    # Only add the pair if both paths exist
-            path_list += (path0,path1)
+        if os.path.exists(path0) and os.path.exists(
+                path1):  # Only add the pair if both paths exist
+            path_list += (path0, path1)
             issame_list.append(issame)
         else:
             print('not exists', path0, path1)
             nrof_skipped_pairs += 1
-    if nrof_skipped_pairs>0:
+    if nrof_skipped_pairs > 0:
         print('Skipped %d image pairs' % nrof_skipped_pairs)
-    
+
     return path_list, issame_list
 
+
 def read_pairs(pairs_filename):
     pairs = []
     with open(pairs_filename, 'r') as f:
@@ -192,88 +233,93 @@ def read_pairs(pairs_filename):
 
 
 def load_dataset(lfw_dir, image_size):
-  lfw_pairs = read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
-  lfw_paths, issame_list = get_paths(lfw_dir, lfw_pairs, 'jpg')
-  lfw_data_list = []
-  for flip in [0,1]:
-    lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
-    lfw_data_list.append(lfw_data)
-  i = 0
-  for path in lfw_paths:
-    with open(path, 'rb') as fin:
-      _bin = fin.read()
-      img = mx.image.imdecode(_bin)
-      img = nd.transpose(img, axes=(2, 0, 1))
-      for flip in [0,1]:
-        if flip==1:
-          img = mx.ndarray.flip(data=img, axis=2)
-        lfw_data_list[flip][i][:] = img
-      i+=1
-      if i%1000==0:
-        print('loading lfw', i)
-  print(lfw_data_list[0].shape)
-  print(lfw_data_list[1].shape)
-  return (lfw_data_list, issame_list)
+    lfw_pairs = read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
+    lfw_paths, issame_list = get_paths(lfw_dir, lfw_pairs, 'jpg')
+    lfw_data_list = []
+    for flip in [0, 1]:
+        lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
+        lfw_data_list.append(lfw_data)
+    i = 0
+    for path in lfw_paths:
+        with open(path, 'rb') as fin:
+            _bin = fin.read()
+            img = mx.image.imdecode(_bin)
+            img = nd.transpose(img, axes=(2, 0, 1))
+            for flip in [0, 1]:
+                if flip == 1:
+                    img = mx.ndarray.flip(data=img, axis=2)
+                lfw_data_list[flip][i][:] = img
+            i += 1
+            if i % 1000 == 0:
+                print('loading lfw', i)
+    print(lfw_data_list[0].shape)
+    print(lfw_data_list[1].shape)
+    return (lfw_data_list, issame_list)
+
 
 def test(lfw_set, mx_model, batch_size):
-  print('testing lfw..')
-  lfw_data_list = lfw_set[0]
-  issame_list = lfw_set[1]
-  model = mx_model
-  embeddings_list = []
-  for i in range( len(lfw_data_list) ):
-    lfw_data = lfw_data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<lfw_data.shape[0]:
-      bb = min(ba+batch_size, lfw_data.shape[0])
-      _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
-      _label = nd.ones( (bb-ba,) )
-      #print(_data.shape, _label.shape)
-      db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      #_arg, _aux = model.get_params()
-      #__arg = {}
-      #for k,v in _arg.iteritems():
-      #  __arg[k] = v.as_in_context(_ctx)
-      #_arg = __arg
-      #_arg["data"] = _data.as_in_context(_ctx)
-      #_arg["softmax_label"] = _label.as_in_context(_ctx)
-      #for k,v in _arg.iteritems():
-      #  print(k,v.context)
-      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
-      #exe.forward(is_train=False)
-      #net_out = exe.outputs
-      _embeddings = net_out[0].asnumpy()
-      #print(_embeddings.shape)
-      if embeddings is None:
-        embeddings = np.zeros( (lfw_data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings
-      ba = bb
-    embeddings_list.append(embeddings)
-
-  _xnorm = 0.0
-  _xnorm_cnt = 0
-  for embed in embeddings_list:
-    for i in range(embed.shape[0]):
-      _em = embed[i]
-      _norm=np.linalg.norm(_em)
-      #print(_em.shape, _norm)
-      _xnorm+=_norm
-      _xnorm_cnt+=1
-  _xnorm /= _xnorm_cnt
-
-  embeddings = embeddings_list[0].copy()
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  acc1, std1 = np.mean(accuracy), np.std(accuracy)
-  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-  #embeddings = np.concatenate(embeddings_list, axis=1)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  print(embeddings.shape)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  return acc1, std1, acc2, std2, _xnorm, embeddings_list
+    print('testing lfw..')
+    lfw_data_list = lfw_set[0]
+    issame_list = lfw_set[1]
+    model = mx_model
+    embeddings_list = []
+    for i in range(len(lfw_data_list)):
+        lfw_data = lfw_data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < lfw_data.shape[0]:
+            bb = min(ba + batch_size, lfw_data.shape[0])
+            _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
+            _label = nd.ones((bb - ba, ))
+            #print(_data.shape, _label.shape)
+            db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            #_arg, _aux = model.get_params()
+            #__arg = {}
+            #for k,v in _arg.iteritems():
+            #  __arg[k] = v.as_in_context(_ctx)
+            #_arg = __arg
+            #_arg["data"] = _data.as_in_context(_ctx)
+            #_arg["softmax_label"] = _label.as_in_context(_ctx)
+            #for k,v in _arg.iteritems():
+            #  print(k,v.context)
+            #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
+            #exe.forward(is_train=False)
+            #net_out = exe.outputs
+            _embeddings = net_out[0].asnumpy()
+            #print(_embeddings.shape)
+            if embeddings is None:
+                embeddings = np.zeros(
+                    (lfw_data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            #print(_em.shape, _norm)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
 
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=10)
+    acc1, std1 = np.mean(accuracy), np.std(accuracy)
+    #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
+    #embeddings = np.concatenate(embeddings_list, axis=1)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=10)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
diff --git a/recognition/partial_fc/mxnet/evaluation/verification.py b/recognition/partial_fc/mxnet/evaluation/verification.py
index b2638b7d6..324b1b3b7 100644
--- a/recognition/partial_fc/mxnet/evaluation/verification.py
+++ b/recognition/partial_fc/mxnet/evaluation/verification.py
@@ -2,19 +2,19 @@
 """
 
 # MIT License
-# 
+#
 # Copyright (c) 2016 David Sandberg
-# 
+#
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
 # in the Software without restriction, including without limitation the rights
 # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 # copies of the Software, and to permit persons to whom the Software is
 # furnished to do so, subject to the following conditions:
-# 
+#
 # The above copyright notice and this permission notice shall be included in all
 # copies or substantial portions of the Software.
-# 
+#
 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
@@ -57,7 +57,12 @@ def split(self, indices):
             return [(indices, indices)]
 
 
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca=0):
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
     assert (embeddings1.shape[0] == embeddings2.shape[0])
     assert (embeddings1.shape[1] == embeddings2.shape[1])
     nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
@@ -96,16 +101,19 @@ def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_fold
         # Find the best threshold for the fold
         acc_train = np.zeros((nrof_thresholds))
         for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
         best_threshold_index = np.argmax(acc_train)
         # print('threshold', thresholds[best_threshold_index])
         for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx, threshold_idx], fprs[fold_idx, threshold_idx], _ = calculate_accuracy(threshold,
-                                                                                                 dist[test_set],
-                                                                                                 actual_issame[
-                                                                                                     test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set],
-                                                      actual_issame[test_set])
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
 
     tpr = np.mean(tprs, 0)
     fpr = np.mean(fprs, 0)
@@ -116,7 +124,9 @@ def calculate_accuracy(threshold, dist, actual_issame):
     predict_issame = np.less(dist, threshold)
     tp = np.sum(np.logical_and(predict_issame, actual_issame))
     fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
     fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
 
     tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
@@ -125,7 +135,12 @@ def calculate_accuracy(threshold, dist, actual_issame):
     return tpr, fpr, acc
 
 
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
     assert (embeddings1.shape[0] == embeddings2.shape[0])
     assert (embeddings1.shape[1] == embeddings2.shape[1])
     nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
@@ -144,14 +159,16 @@ def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_targe
         # Find the threshold that gives FAR = far_target
         far_train = np.zeros(nrof_thresholds)
         for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
         if np.max(far_train) >= far_target:
             f = interpolate.interp1d(far_train, thresholds, kind='slinear')
             threshold = f(far_target)
         else:
             threshold = 0.0
 
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
 
     val_mean = np.mean(val)
     far_mean = np.mean(far)
@@ -162,7 +179,8 @@ def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_targe
 def calculate_val_far(threshold, dist, actual_issame):
     predict_issame = np.less(dist, threshold)
     true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
     n_same = np.sum(actual_issame)
     n_diff = np.sum(np.logical_not(actual_issame))
     # print(true_accept, false_accept)
@@ -177,11 +195,19 @@ def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
     thresholds = np.arange(0, 4, 0.01)
     embeddings1 = embeddings[0::2]
     embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-                                       np.asarray(actual_issame), nrof_folds=nrof_folds, pca=pca)
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
     thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-                                      np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
     return tpr, fpr, accuracy, val, val_std, far
 
 
@@ -194,7 +220,8 @@ def load_bin(path, image_size):
         bins, issame_list = pickle.load(f, encoding='bytes')  # py3
     data_list = []
     for flip in [0, 1]:
-        data = nd.empty((len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data = nd.empty(
+            (len(issame_list) * 2, 3, image_size[0], image_size[1]))
         data_list.append(data)
     for i in range(len(issame_list) * 2):
         _bin = bins[i]
@@ -210,7 +237,12 @@ def load_bin(path, image_size):
     return (data_list, issame_list)
 
 
-def test(data_set, mx_model, batch_size, nfolds=10, data_extra=None, label_shape=None):
+def test(data_set,
+         mx_model,
+         batch_size,
+         nfolds=10,
+         data_extra=None,
+         label_shape=None):
     print('testing verification..')
     data_list = data_set[0]
     issame_list = data_set[1]
@@ -220,7 +252,7 @@ def test(data_set, mx_model, batch_size, nfolds=10, data_extra=None, label_shape
         _data_extra = nd.array(data_extra)
     time_consumed = 0.0
     if label_shape is None:
-        _label = nd.ones((batch_size,))
+        _label = nd.ones((batch_size, ))
     else:
         _label = nd.ones(label_shape)
     for i in range(len(data_list)):
@@ -234,9 +266,10 @@ def test(data_set, mx_model, batch_size, nfolds=10, data_extra=None, label_shape
             # print(_data.shape, _label.shape)
             time0 = datetime.datetime.now()
             if data_extra is None:
-                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
             else:
-                db = mx.io.DataBatch(data=(_data, _data_extra), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
 
             model.forward(db, is_train=False)
             net_out = model.get_outputs()
@@ -287,12 +320,19 @@ def test(data_set, mx_model, batch_size, nfolds=10, data_extra=None, label_shape
     embeddings = sklearn.preprocessing.normalize(embeddings)
     print(embeddings.shape)
     print('infer time', time_consumed)
-    _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=nfolds)
     acc2, std2 = np.mean(accuracy), np.std(accuracy)
     return acc1, std1, acc2, std2, _xnorm, embeddings_list
 
 
-def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=None):
+def test_badcase(data_set,
+                 mx_model,
+                 batch_size,
+                 name='',
+                 data_extra=None,
+                 label_shape=None):
     print('testing verification badcase..')
     data_list = data_set[0]
     issame_list = data_set[1]
@@ -302,7 +342,7 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
         _data_extra = nd.array(data_extra)
     time_consumed = 0.0
     if label_shape is None:
-        _label = nd.ones((batch_size,))
+        _label = nd.ones((batch_size, ))
     else:
         _label = nd.ones(label_shape)
     for i in range(len(data_list)):
@@ -316,9 +356,10 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
             # print(_data.shape, _label.shape)
             time0 = datetime.datetime.now()
             if data_extra is None:
-                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
             else:
-                db = mx.io.DataBatch(data=(_data, _data_extra), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
             model.forward(db, is_train=False)
             net_out = model.get_outputs()
             _embeddings = net_out[0].asnumpy()
@@ -364,15 +405,18 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
         for threshold_idx, threshold in enumerate(thresholds):
             p2 = dist[train_set]
             p3 = actual_issame[train_set]
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, p2, p3)
         best_threshold_index = np.argmax(acc_train)
         for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx, threshold_idx], fprs[fold_idx, threshold_idx], _ = calculate_accuracy(threshold,
-                                                                                                 dist[test_set],
-                                                                                                 actual_issame[
-                                                                                                     test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set],
-                                                      actual_issame[test_set])
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
         best_threshold = thresholds[best_threshold_index]
         for iid in test_set:
             ida = iid * 2
@@ -383,7 +427,8 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
             if not asame:
                 violate *= -1.0
             if violate > 0.0:
-                imga = data[ida].asnumpy().transpose((1, 2, 0))[..., ::-1]  # to bgr
+                imga = data[ida].asnumpy().transpose(
+                    (1, 2, 0))[..., ::-1]  # to bgr
                 imgb = data[idb].asnumpy().transpose((1, 2, 0))[..., ::-1]
                 # print(imga.shape, imgb.shape, violate, asame, _dist)
                 if asame:
@@ -408,7 +453,8 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
     else:
         threshold = pouts[-1][3]
 
-    for item in [(pouts, 'positive(false_negative).png'), (nouts, 'negative(false_positive).png')]:
+    for item in [(pouts, 'positive(false_negative).png'),
+                 (nouts, 'negative(false_positive).png')]:
         cols = 4
         rows = 8000
         outs = item[0]
@@ -423,14 +469,21 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
         hack = {}
 
         if name.startswith('cfp') and item[1].startswith('pos'):
-            hack = {0: 'manual/238_13.jpg.jpg', 6: 'manual/088_14.jpg.jpg', 10: 'manual/470_14.jpg.jpg',
-                    25: 'manual/238_13.jpg.jpg', 28: 'manual/143_11.jpg.jpg'}
+            hack = {
+                0: 'manual/238_13.jpg.jpg',
+                6: 'manual/088_14.jpg.jpg',
+                10: 'manual/470_14.jpg.jpg',
+                25: 'manual/238_13.jpg.jpg',
+                28: 'manual/143_11.jpg.jpg'
+            }
 
         filename = item[1]
         if len(name) > 0:
             filename = name + "_" + filename
         filename = os.path.join(out_dir, filename)
-        img = np.zeros((image_shape[0] * rows + 20, image_shape[1] * cols + (cols - 1) * gap, 3), dtype=np.uint8)
+        img = np.zeros((image_shape[0] * rows + 20, image_shape[1] * cols +
+                        (cols - 1) * gap, 3),
+                       dtype=np.uint8)
         img[:, :, :] = 255
         text_color = (0, 0, 153)
         text_color = (255, 178, 102)
@@ -462,19 +515,27 @@ def test_badcase(data_set, mx_model, batch_size, name='', data_extra=None, label
             k = "%.3f" % dist
             # print(k)
             font = cv2.FONT_HERSHEY_SIMPLEX
-            cv2.putText(_img, k, (80, image_shape[0] // 2 + 7), font, 0.6, text_color, 2)
+            cv2.putText(_img, k, (80, image_shape[0] // 2 + 7), font, 0.6,
+                        text_color, 2)
             # _filename = filename+"_%d.png"%outi
             # cv2.imwrite(_filename, _img)
             img[row * image_shape[0]:(row + 1) * image_shape[0],
-            (col * image_shape[1] + gap * col):((col + 1) * image_shape[1] + gap * col), :] = _img
+                (col * image_shape[1] +
+                 gap * col):((col + 1) * image_shape[1] + gap * col), :] = _img
         # threshold = outs[0][3]
         font = cv2.FONT_HERSHEY_SIMPLEX
         k = "threshold: %.3f" % threshold
-        cv2.putText(img, k, (img.shape[1] // 2 - 70, img.shape[0] - 5), font, 0.6, text_color, 2)
+        cv2.putText(img, k, (img.shape[1] // 2 - 70, img.shape[0] - 5), font,
+                    0.6, text_color, 2)
         cv2.imwrite(filename, img)
 
 
-def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=None):
+def dumpR(data_set,
+          mx_model,
+          batch_size,
+          name='',
+          data_extra=None,
+          label_shape=None):
     print('dump verification embedding..')
     data_list = data_set[0]
     issame_list = data_set[1]
@@ -484,7 +545,7 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
         _data_extra = nd.array(data_extra)
     time_consumed = 0.0
     if label_shape is None:
-        _label = nd.ones((batch_size,))
+        _label = nd.ones((batch_size, ))
     else:
         _label = nd.ones(label_shape)
     for i in range(len(data_list)):
@@ -498,9 +559,10 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
             # print(_data.shape, _label.shape)
             time0 = datetime.datetime.now()
             if data_extra is None:
-                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
             else:
-                db = mx.io.DataBatch(data=(_data, _data_extra), label=(_label,))
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
             model.forward(db, is_train=False)
             net_out = model.get_outputs()
             _embeddings = net_out[0].asnumpy()
@@ -517,7 +579,9 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
     actual_issame = np.asarray(issame_list)
     outname = os.path.join('temp.bin')
     with open(outname, 'wb') as f:
-        pickle.dump((embeddings, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
+        pickle.dump((embeddings, issame_list),
+                    f,
+                    protocol=pickle.HIGHEST_PROTOCOL)
 
 
 if __name__ == '__main__':
@@ -525,8 +589,12 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
     parser = argparse.ArgumentParser(description='do verification')
     # general
     parser.add_argument('--data-dir', default='', help='')
-    parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-    parser.add_argument('--target', default='lfw,cfp_ff,cfp_fp,agedb_30', help='test targets.')
+    parser.add_argument('--model',
+                        default='../model/softmax,50',
+                        help='path to load model.')
+    parser.add_argument('--target',
+                        default='lfw,cfp_ff,cfp_fp,agedb_30',
+                        help='test targets.')
     parser.add_argument('--gpu', default=0, type=int, help='gpu id')
     parser.add_argument('--batch-size', default=32, type=int, help='')
     parser.add_argument('--max', default='', type=str, help='')
@@ -572,7 +640,8 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
         sym = all_layers['fc1_output']
         model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
         # model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-        model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
+        model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+                                          image_size[1]))])
         model.set_params(arg_params, aux_params)
         nets.append(model)
     time_now = datetime.datetime.now()
@@ -593,10 +662,13 @@ def dumpR(data_set, mx_model, batch_size, name='', data_extra=None, label_shape=
         for i in range(len(ver_list)):
             results = []
             for model in nets:
-                acc1, std1, acc2, std2, xnorm, embeddings_list = test(ver_list[i], model, args.batch_size, args.nfolds)
+                acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+                    ver_list[i], model, args.batch_size, args.nfolds)
                 print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
-                print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
-                print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
+                print('[%s]Accuracy: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc1, std1))
+                print('[%s]Accuracy-Flip: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc2, std2))
                 results.append(acc2)
             print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
     elif args.mode == 1:
diff --git a/recognition/partial_fc/mxnet/image_iter.py b/recognition/partial_fc/mxnet/image_iter.py
index 1a004cbc2..c3a18717d 100644
--- a/recognition/partial_fc/mxnet/image_iter.py
+++ b/recognition/partial_fc/mxnet/image_iter.py
@@ -21,12 +21,22 @@
 
 
 class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape, path_imgrec=None,
-                 shuffle=False, aug_list=None, mean=None,
-                 rand_mirror=False, cutoff=0, color_jittering=0,
+    def __init__(self,
+                 batch_size,
+                 data_shape,
+                 path_imgrec=None,
+                 shuffle=False,
+                 aug_list=None,
+                 mean=None,
+                 rand_mirror=False,
+                 cutoff=0,
+                 color_jittering=0,
                  images_filter=0,
-                 data_name='data', label_name='softmax_label', context=0, context_num=1, **kwargs):
+                 data_name='data',
+                 label_name='softmax_label',
+                 context=0,
+                 context_num=1,
+                 **kwargs):
         super(FaceImageIter, self).__init__()
         assert path_imgrec
         self.context = context
@@ -34,14 +44,16 @@ def __init__(self, batch_size, data_shape, path_imgrec=None,
         if path_imgrec:
             logging.info('loading recordio %s...', path_imgrec)
             path_imgidx = path_imgrec[0:-4] + ".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec,
+                                                     'r')
             s = self.imgrec.read_idx(0)
             header, _ = recordio.unpack(s)
             if header.flag > 0:
                 self.header0 = (int(header.label[0]), int(header.label[1]))
                 self.imgidx = []
                 self.id2range = {}
-                self.seq_identity = range(int(header.label[0]), int(header.label[1]))
+                self.seq_identity = range(int(header.label[0]),
+                                          int(header.label[1]))
                 for identity in self.seq_identity:
                     s = self.imgrec.read_idx(identity)
                     header, _ = recordio.unpack(s)
@@ -69,7 +81,7 @@ def __init__(self, batch_size, data_shape, path_imgrec=None,
             self.nd_mean = mx.nd.array(self.mean).reshape((1, 1, 3))
 
         self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
+        self.provide_data = [(data_name, (batch_size, ) + data_shape)]
         self.batch_size = batch_size
         self.data_shape = data_shape
         self.shuffle = shuffle
@@ -79,12 +91,13 @@ def __init__(self, batch_size, data_shape, path_imgrec=None,
         self.cutoff = cutoff
         self.color_jittering = color_jittering
         self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
-        self.provide_label = [(label_name, (batch_size,))]
+        self.provide_label = [(label_name, (batch_size, ))]
 
         self.cur = 0
         self.nbatch = 0
         self.is_init = False
-        self.num_samples_per_gpu = int(math.floor(len(self.seq) * 1.0 / self.context_num))
+        self.num_samples_per_gpu = int(
+            math.floor(len(self.seq) * 1.0 / self.context_num))
 
     def reset(self):
         """Resets the iterator to the beginning of the data."""
@@ -104,7 +117,8 @@ def next_sample(self):
             while True:
                 if self.cur >= self.num_samples_per_gpu:
                     raise StopIteration
-                idx = self.seq[self.num_samples_per_gpu * self.context + self.cur]
+                idx = self.seq[self.num_samples_per_gpu * self.context +
+                               self.cur]
                 self.cur += 1
                 if self.imgrec is not None:
                     s = self.imgrec.read_idx(idx)
@@ -242,9 +256,11 @@ def next(self):
     def check_data_shape(self, data_shape):
         """Checks if the input data shape is valid"""
         if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
+            raise ValueError(
+                'data_shape should have length 3, with dimensions CxHxW')
         if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
+            raise ValueError(
+                'This iterator expects inputs to have 3 channels.')
 
     def check_valid_image(self, data):
         """Checks if the input data is valid"""
@@ -303,17 +319,23 @@ def next(self):
 
 # dummy
 class DummyIter(mx.io.DataIter):
-    def __init__(self, batch_size, data_shape, batches=1000, mode='', dtype='float32'):
+    def __init__(self,
+                 batch_size,
+                 data_shape,
+                 batches=1000,
+                 mode='',
+                 dtype='float32'):
         super(DummyIter, self).__init__(batch_size)
-        self.data_shape = (batch_size,) + data_shape
-        self.label_shape = (batch_size,)
+        self.data_shape = (batch_size, ) + data_shape
+        self.label_shape = (batch_size, )
         self.provide_data = [('data', self.data_shape)]
         self.provide_label = [('softmax_label', self.label_shape)]
         # self.provide_label = [('label', self.label_shape)]
         # if mode == 'perseus':
         #    self.provide_label = []
-        self.batch = mx.io.DataBatch(data=[mx.nd.zeros(self.data_shape, dtype=dtype)],
-                                     label=[mx.nd.zeros(self.label_shape, dtype=dtype)])
+        self.batch = mx.io.DataBatch(
+            data=[mx.nd.zeros(self.data_shape, dtype=dtype)],
+            label=[mx.nd.zeros(self.label_shape, dtype=dtype)])
         self._batches = 0
         self.batches = batches
 
@@ -324,4 +346,3 @@ def next(self):
         else:
             self._batches = 0
             raise StopIteration
-
diff --git a/recognition/partial_fc/mxnet/memory_bank.py b/recognition/partial_fc/mxnet/memory_bank.py
index 4313eaf39..d24a21221 100644
--- a/recognition/partial_fc/mxnet/memory_bank.py
+++ b/recognition/partial_fc/mxnet/memory_bank.py
@@ -8,7 +8,14 @@
 
 
 class MemoryBank(object):
-    def __init__(self, num_sample, num_local, rank, local_rank, embedding_size, prefix, gpu=True):
+    def __init__(self,
+                 num_sample,
+                 num_local,
+                 rank,
+                 local_rank,
+                 embedding_size,
+                 prefix,
+                 gpu=True):
         """
         Parameters
         ----------
@@ -40,13 +47,16 @@ def __init__(self, num_sample, num_local, rank, local_rank, embedding_size, pref
             context = mx.cpu()
 
         # In order to apply update, weight and momentum should be storage.
-        self.weight = nd.random_normal(
-            loc=0, scale=0.01, shape=(self.num_local, self.embedding_size),
-            ctx=context)
+        self.weight = nd.random_normal(loc=0,
+                                       scale=0.01,
+                                       shape=(self.num_local,
+                                              self.embedding_size),
+                                       ctx=context)
         self.weight_mom = nd.zeros_like(self.weight)
 
         # Sampler object
-        self.weight_index_sampler = WeightIndexSampler(num_sample, num_local,  rank)
+        self.weight_index_sampler = WeightIndexSampler(num_sample, num_local,
+                                                       rank)
 
     def sample(self, global_label):
         """
@@ -99,5 +109,9 @@ def set(self, index, updated_weight, updated_weight_mom=None):
         self.weight_mom[index] = updated_weight_mom
 
     def save(self):
-        nd.save(fname=os.path.join(self.prefix, "%d_centers.param" % self.rank), data=self.weight)
-        nd.save(fname=os.path.join(self.prefix, "%d_centers_mom.param" % self.rank), data=self.weight_mom)
+        nd.save(fname=os.path.join(self.prefix,
+                                   "%d_centers.param" % self.rank),
+                data=self.weight)
+        nd.save(fname=os.path.join(self.prefix,
+                                   "%d_centers_mom.param" % self.rank),
+                data=self.weight_mom)
diff --git a/recognition/partial_fc/mxnet/memory_module.py b/recognition/partial_fc/mxnet/memory_module.py
index 6d9de607a..acd0b6b9e 100644
--- a/recognition/partial_fc/mxnet/memory_module.py
+++ b/recognition/partial_fc/mxnet/memory_module.py
@@ -19,7 +19,7 @@ class SampleDistributeModule(object):
     RTX2080Ti can complete classification tasks with 100 million of identities.
 
     See the original paper:
-    https://arxiv.org/abs/2010.05222
+    https://arxiv.org/abs/2203.15565
 
     Parameters
     ----------
@@ -33,24 +33,33 @@ class SampleDistributeModule(object):
         The updater of memory bank, default is sgd optimizer.
     logger:
     """
-    def __init__(self, symbol, fc7_model, memory_bank, memory_optimizer,
-                 logger=logging, ):
+    def __init__(
+        self,
+        symbol,
+        fc7_model,
+        memory_bank,
+        memory_optimizer,
+        logger=logging,
+    ):
         self.size = hvd.size()
         self.rank = hvd.rank()
         self.local_rank = hvd.local_rank()
         self.gpu = mx.gpu(self.local_rank)
-        self.cpu = mx.cpu()                                     # `device_id` is not needed for CPU.
+        self.cpu = mx.cpu()  # `device_id` is not needed for CPU.
         self.nd_cache = {}
         self.embedding_size = config.embedding_size
         self.batch_size = config.batch_size
         self.num_update = 0
-        self.batch_end_param = namedtuple('batch_end_param', ['loss', 'num_epoch', 'num_update'])
+        self.batch_end_param = namedtuple('batch_end_param',
+                                          ['loss', 'num_epoch', 'num_update'])
 
         self.fc7_model = fc7_model
         self.symbol = symbol
         self.logger = logger
-        self.backbone_module = mx.module.Module(
-            self.symbol, ['data'], ['softmax_label'], logger=self.logger, context=self.gpu)
+        self.backbone_module = mx.module.Module(self.symbol, ['data'],
+                                                ['softmax_label'],
+                                                logger=self.logger,
+                                                context=self.gpu)
 
         self.memory_bank = memory_bank
         self.memory_optimizer = memory_optimizer
@@ -78,8 +87,13 @@ def broadcast_parameters(params):
             rank_0_dict[key] = hvd.broadcast(tensor, 0, key)
         return rank_0_dict
 
-    def fit(self, train_data, optimizer_params, batch_end_callback,
-            initializer, arg_params=None, aux_params=None):
+    def fit(self,
+            train_data,
+            optimizer_params,
+            batch_end_callback,
+            initializer,
+            arg_params=None,
+            aux_params=None):
 
         # Bind -> Init_params -> Init_optimizers
         self.bind(train_data.provide_data, train_data.provide_label, True)
@@ -115,7 +129,9 @@ def fit(self, train_data, optimizer_params, batch_end_callback,
 
                 if batch_end_callback is not None:
                     batch_end_params = self.batch_end_param(
-                        loss=self.loss_cache, num_epoch=num_epoch, num_update=self.num_update)
+                        loss=self.loss_cache,
+                        num_epoch=num_epoch,
+                        num_update=self.num_update)
                     batch_end_callback(batch_end_params)
 
     def get_export_params(self):
@@ -143,8 +159,13 @@ def get_ndarray(self, context, name, shape, dtype='float32'):
             v = self.nd_cache[key]
         return v
 
-    def init_params(self, initializer, arg_params=None, aux_params=None,
-                    allow_missing=False, force_init=False, allow_extra=False):
+    def init_params(self,
+                    initializer,
+                    arg_params=None,
+                    aux_params=None,
+                    allow_missing=False,
+                    force_init=False,
+                    allow_extra=False):
         """Initializes the parameters and auxiliary states.
 
         Parameters
@@ -168,29 +189,34 @@ def init_params(self, initializer, arg_params=None, aux_params=None,
             contain extra parameters that is not needed by the executor.
         """
         # backbone
-        self.backbone_module.init_params(
-            initializer=initializer, arg_params=arg_params,
-            aux_params=aux_params, allow_missing=allow_missing,
-            force_init=force_init, allow_extra=allow_extra)
+        self.backbone_module.init_params(initializer=initializer,
+                                         arg_params=arg_params,
+                                         aux_params=aux_params,
+                                         allow_missing=allow_missing,
+                                         force_init=force_init,
+                                         allow_extra=allow_extra)
 
     def prepare(self, data_batch, sparse_row_id_fn=None):
         if sparse_row_id_fn is not None:
-            warnings.warn(UserWarning("sparse_row_id_fn is not invoked for BaseModule."))
+            warnings.warn(
+                UserWarning("sparse_row_id_fn is not invoked for BaseModule."))
 
     def allgather(self, tensor, name, shape, dtype, context):
         """ Implement in-place AllGather using AllReduce
         """
-        assert isinstance(tensor, nd.NDArray),          type(tensor)
-        assert isinstance(name, str),                   type(name)
-        assert isinstance(shape, tuple),                type(shape)
-        assert isinstance(dtype, str),                  type(dtype)
+        assert isinstance(tensor, nd.NDArray), type(tensor)
+        assert isinstance(name, str), type(name)
+        assert isinstance(shape, tuple), type(shape)
+        assert isinstance(dtype, str), type(dtype)
         assert isinstance(context, mx.context.Context), type(context)
-        total_tensor = self.get_ndarray(
-            context=context, name=name, shape=shape, dtype=dtype)
-        total_tensor[:] = 0                                      # reset array before all-reduce is very important
-        total_tensor[self.rank * self.batch_size:
-                     self.rank * self.batch_size + self.batch_size] = tensor
-        hvd.allreduce_(total_tensor, average=False)              # all-reduce in-place
+        total_tensor = self.get_ndarray(context=context,
+                                        name=name,
+                                        shape=shape,
+                                        dtype=dtype)
+        total_tensor[:] = 0  # reset array before all-reduce is very important
+        total_tensor[self.rank * self.batch_size:self.rank * self.batch_size +
+                     self.batch_size] = tensor
+        hvd.allreduce_(total_tensor, average=False)  # all-reduce in-place
         return total_tensor
 
     def forward(self, data_batch, is_train=None):
@@ -200,29 +226,32 @@ def forward(self, data_batch, is_train=None):
             fc1 = self.backbone_module.get_outputs()[0]
             label = data_batch.label[0]
 
-            total_features = self.allgather(
-                tensor=fc1,
-                name='total_feature',
-                shape=(self.batch_size * self.size, self.embedding_size),
-                dtype='float32',
-                context=self.gpu
-            )
-            total_labels = self.allgather(
-                tensor=label,
-                name='total_label',
-                shape=(self.batch_size * self.size,),
-                dtype='int32',
-                context=self.cpu
-            )
+            total_features = self.allgather(tensor=fc1,
+                                            name='total_feature',
+                                            shape=(self.batch_size * self.size,
+                                                   self.embedding_size),
+                                            dtype='float32',
+                                            context=self.gpu)
+            total_labels = self.allgather(tensor=label,
+                                          name='total_label',
+                                          shape=(self.batch_size *
+                                                 self.size, ),
+                                          dtype='int32',
+                                          context=self.cpu)
             return total_features, total_labels
         else:
             return None
 
-    def backward_all(self, total_feature, total_label, ):
+    def backward_all(
+        self,
+        total_feature,
+        total_label,
+    ):
         # get memory bank learning rate
         self.memory_lr = self.memory_optimizer.lr_scheduler(self.num_update)
 
-        self.grad_cache = self.get_ndarray(self.gpu, 'grad_cache', total_feature.shape)
+        self.grad_cache = self.get_ndarray(self.gpu, 'grad_cache',
+                                           total_feature.shape)
         self.loss_cache = self.get_ndarray(self.gpu, 'loss_cache', [1])
 
         self.grad_cache[:] = 0
@@ -238,10 +267,9 @@ def backward_all(self, total_feature, total_label, ):
         total_feature_grad = grad
         total_feature_grad = hvd.allreduce(total_feature_grad, average=False)
 
-        fc1_grad = total_feature_grad[
-            self.batch_size * self.rank:
-            self.batch_size * self.rank + self.batch_size
-        ]
+        fc1_grad = total_feature_grad[self.batch_size *
+                                      self.rank:self.batch_size * self.rank +
+                                      self.batch_size]
         self.backbone_module.backward(out_grads=[fc1_grad / self.size])
 
     def get_outputs(self, merge_multi_context=True):
@@ -253,7 +281,8 @@ def get_outputs(self, merge_multi_context=True):
         list of NDArray or list of list of NDArray
             Output.
         """
-        return self.backbone_module.get_outputs(merge_multi_context=merge_multi_context)
+        return self.backbone_module.get_outputs(
+            merge_multi_context=merge_multi_context)
 
     def update(self):
         """
@@ -264,7 +293,9 @@ def update(self):
         mx.nd.waitall()
 
     def bind(self, data_shapes, label_shapes=None, for_training=True):
-        self.backbone_module.bind(data_shapes, label_shapes, for_training=for_training)
+        self.backbone_module.bind(data_shapes,
+                                  label_shapes,
+                                  for_training=for_training)
 
     def init_optimizer(self, optimizer_params, force_init=False):
         """
@@ -279,14 +310,18 @@ def init_optimizer(self, optimizer_params, force_init=False):
             Default ``False``, indicating whether we should force re-initializing the
             optimizer in the case an optimizer is already installed.
         """
-        optimizer_backbone = DistributedOptimizer(mx.optimizer.SGD(**optimizer_params))
-        self.backbone_module.init_optimizer('local', optimizer_backbone, force_init=force_init)
+        optimizer_backbone = DistributedOptimizer(
+            mx.optimizer.SGD(**optimizer_params))
+        self.backbone_module.init_optimizer('local',
+                                            optimizer_backbone,
+                                            force_init=force_init)
 
     def backward(self, total_feature, label):
         memory_bank = self.memory_bank
         assert memory_bank.num_local == memory_bank.num_sample, "pass"
 
-        _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank, total_feature)
+        _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank,
+                                  total_feature)
         # Attach grad
         _data.attach_grad()
         memory_bank.weight.attach_grad()
@@ -294,23 +329,27 @@ def backward(self, total_feature, label):
         # Convert label
         _label = self.get_ndarray2(self.gpu, 'label_%d' % self.rank, label)
         _label = _label - int(self.rank * memory_bank.num_local)
-        _fc7, _one_hot = self.fc7_model.forward(
-            _data, memory_bank.weight, mapping_label=_label, depth=memory_bank.num_local)
+        _fc7, _one_hot = self.fc7_model.forward(_data,
+                                                memory_bank.weight,
+                                                mapping_label=_label,
+                                                depth=memory_bank.num_local)
 
         # Sync max
         max_fc7 = nd.max(_fc7, axis=1, keepdims=True)
         max_fc7 = nd.reshape(max_fc7, -1)
 
-        total_max_fc7 = self.get_ndarray(
-            context=self.gpu, name='total_max_fc7',
-            shape=(max_fc7.shape[0], self.size), dtype='float32')
+        total_max_fc7 = self.get_ndarray(context=self.gpu,
+                                         name='total_max_fc7',
+                                         shape=(max_fc7.shape[0], self.size),
+                                         dtype='float32')
         total_max_fc7[:] = 0
         total_max_fc7[:, self.rank] = max_fc7
         hvd.allreduce_(total_max_fc7, average=False)
 
-        global_max_fc7 = self.get_ndarray(
-            context=self.gpu, name='global_max_fc7',
-            shape=(max_fc7.shape[0], 1), dtype='float32')
+        global_max_fc7 = self.get_ndarray(context=self.gpu,
+                                          name='global_max_fc7',
+                                          shape=(max_fc7.shape[0], 1),
+                                          dtype='float32')
         nd.max(total_max_fc7, axis=1, keepdims=True, out=global_max_fc7)
 
         # Calculate exp(logits)
@@ -339,17 +378,17 @@ def backward(self, total_feature, label):
 
         # Update center
         _weight_grad = memory_bank.weight.grad
-        self.memory_optimizer.update(
-            weight=memory_bank.weight,
-            grad=_weight_grad,
-            state=memory_bank.weight_mom,
-            learning_rate=self.memory_lr)
+        self.memory_optimizer.update(weight=memory_bank.weight,
+                                     grad=_weight_grad,
+                                     state=memory_bank.weight_mom,
+                                     learning_rate=self.memory_lr)
 
         return _data.grad, global_loss
 
     def backward_sample(self, total_feature, label):
         this_rank_classes = int(self.memory_bank.num_sample)
-        local_index, unique_sorted_global_label = self.memory_bank.sample(label)
+        local_index, unique_sorted_global_label = self.memory_bank.sample(
+            label)
 
         # Get local index
         _mapping_dict = {}
@@ -357,7 +396,8 @@ def backward_sample(self, total_feature, label):
         global_label_set = set(unique_sorted_global_label)
         for idx, absolute_label in enumerate(local_sampled_class):
             if absolute_label in global_label_set:
-                _mapping_dict[absolute_label] = idx + self.rank * self.memory_bank.num_sample
+                _mapping_dict[
+                    absolute_label] = idx + self.rank * self.memory_bank.num_sample
 
         label_list = list(label.asnumpy())
         mapping_label = []
@@ -377,37 +417,51 @@ def backward_sample(self, total_feature, label):
 
         # Sync to gpu
         if self.memory_bank.gpu:
-            _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank, total_feature)
-            _weight = self.get_ndarray2(self.gpu, 'weight_%d' % self.rank, sample_weight)
-            _weight_mom = self.get_ndarray2(self.gpu, 'weight_mom_%d' % self.rank, sample_weight_mom)
+            _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank,
+                                      total_feature)
+            _weight = self.get_ndarray2(self.gpu, 'weight_%d' % self.rank,
+                                        sample_weight)
+            _weight_mom = self.get_ndarray2(self.gpu,
+                                            'weight_mom_%d' % self.rank,
+                                            sample_weight_mom)
         else:
-            _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank, total_feature)
-            _weight = self.get_ndarray2(self.gpu, 'weight_%d' % self.rank, sample_weight)
-            _weight_mom = self.get_ndarray2(self.gpu, 'weight_mom_%d' % self.rank, sample_weight_mom)
+            _data = self.get_ndarray2(self.gpu, "data_%d" % self.rank,
+                                      total_feature)
+            _weight = self.get_ndarray2(self.gpu, 'weight_%d' % self.rank,
+                                        sample_weight)
+            _weight_mom = self.get_ndarray2(self.gpu,
+                                            'weight_mom_%d' % self.rank,
+                                            sample_weight_mom)
 
         # Attach grad
         _data.attach_grad()
         _weight.attach_grad()
 
         # Convert label
-        _label = self.get_ndarray2(self.gpu, 'mapping_label_%d' % self.rank, mapping_label)
+        _label = self.get_ndarray2(self.gpu, 'mapping_label_%d' % self.rank,
+                                   mapping_label)
         _label = _label - int(self.rank * self.memory_bank.num_sample)
-        _fc7, _one_hot = self.fc7_model.forward(_data, _weight, mapping_label=_label, depth=this_rank_classes)
+        _fc7, _one_hot = self.fc7_model.forward(_data,
+                                                _weight,
+                                                mapping_label=_label,
+                                                depth=this_rank_classes)
 
         # Sync max
         max_fc7 = nd.max(_fc7, axis=1, keepdims=True)
         max_fc7 = nd.reshape(max_fc7, -1)
 
-        total_max_fc7 = self.get_ndarray(
-            context=self.gpu, name='total_max_fc7',
-            shape=(max_fc7.shape[0], self.size), dtype='float32')
+        total_max_fc7 = self.get_ndarray(context=self.gpu,
+                                         name='total_max_fc7',
+                                         shape=(max_fc7.shape[0], self.size),
+                                         dtype='float32')
         total_max_fc7[:] = 0
         total_max_fc7[:, self.rank] = max_fc7
         hvd.allreduce_(total_max_fc7, average=False)
 
-        global_max_fc7 = self.get_ndarray(
-            context=self.gpu, name='global_max_fc7',
-            shape=(max_fc7.shape[0], 1), dtype='float32')
+        global_max_fc7 = self.get_ndarray(context=self.gpu,
+                                          name='global_max_fc7',
+                                          shape=(max_fc7.shape[0], 1),
+                                          dtype='float32')
         nd.max(total_max_fc7, axis=1, keepdims=True, out=global_max_fc7)
 
         # Calculate exp(logits)
@@ -435,15 +489,20 @@ def backward_sample(self, total_feature, label):
 
         # Update center
         _weight_grad = _weight.grad
-        self.memory_optimizer.update(weight=_weight, grad=_weight_grad, state=_weight_mom, learning_rate=self.memory_lr)
+        self.memory_optimizer.update(weight=_weight,
+                                     grad=_weight_grad,
+                                     state=_weight_mom,
+                                     learning_rate=self.memory_lr)
         if self.memory_bank.gpu:
-            self.memory_bank.set(
-                index=local_index,
-                updated_weight=_weight,
-                updated_weight_mom=_weight_mom)
+            self.memory_bank.set(index=local_index,
+                                 updated_weight=_weight,
+                                 updated_weight_mom=_weight_mom)
         else:
-            self.memory_bank.set(
-                index=local_index,
-                updated_weight=self.get_ndarray2(mx.cpu(), "cpu_weight_%d" % self.rank, _weight),
-                updated_weight_mom=self.get_ndarray2(mx.cpu(), "cpu_weight_mom_%d" % self.rank, _weight_mom))
+            self.memory_bank.set(index=local_index,
+                                 updated_weight=self.get_ndarray2(
+                                     mx.cpu(), "cpu_weight_%d" % self.rank,
+                                     _weight),
+                                 updated_weight_mom=self.get_ndarray2(
+                                     mx.cpu(), "cpu_weight_mom_%d" % self.rank,
+                                     _weight_mom))
         return _data.grad, global_loss
diff --git a/recognition/partial_fc/mxnet/memory_samplers.py b/recognition/partial_fc/mxnet/memory_samplers.py
index b0e7b77e9..ee8949176 100644
--- a/recognition/partial_fc/mxnet/memory_samplers.py
+++ b/recognition/partial_fc/mxnet/memory_samplers.py
@@ -4,7 +4,6 @@
 class CenterPositiveClassGet(object):
     """ Get the corresponding center of the positive class
     """
-
     def __init__(self, num_sample, num_local, rank):
         self.num_sample = num_sample
         self.num_local = num_local
@@ -31,7 +30,6 @@ def __call__(self, global_label):
 class CenterNegetiveClassSample(object):
     """ Sample negative class center
     """
-
     def __init__(self, num_sample, num_local, rank):
         self.num_sample = num_sample
         self.num_local = num_local
@@ -45,9 +43,12 @@ def __call__(self, positive_center_index):
         -------
         negative_center_index: list of int
         """
-        negative_class_pool = np.setdiff1d(self.negative_class_pool, positive_center_index)
+        negative_class_pool = np.setdiff1d(self.negative_class_pool,
+                                           positive_center_index)
         negative_sample_size = self.num_sample - len(positive_center_index)
-        negative_center_index = np.random.choice(negative_class_pool, negative_sample_size, replace=False)
+        negative_center_index = np.random.choice(negative_class_pool,
+                                                 negative_sample_size,
+                                                 replace=False)
         return negative_center_index
 
 
@@ -68,7 +69,9 @@ def __call__(self, global_label):
             positive_center_index = positive_center_index[:self.num_sample]
         negative_center_index = self.negative(positive_center_index)
         #
-        final_center_index = np.concatenate((positive_center_index, negative_center_index))
-        assert len(final_center_index) == len(np.unique(final_center_index)) == self.num_sample
+        final_center_index = np.concatenate(
+            (positive_center_index, negative_center_index))
+        assert len(final_center_index) == len(
+            np.unique(final_center_index)) == self.num_sample
         assert len(final_center_index) == self.num_sample
         return final_center_index
diff --git a/recognition/partial_fc/mxnet/memory_scheduler.py b/recognition/partial_fc/mxnet/memory_scheduler.py
index b05dd8764..2b1b36dda 100644
--- a/recognition/partial_fc/mxnet/memory_scheduler.py
+++ b/recognition/partial_fc/mxnet/memory_scheduler.py
@@ -5,7 +5,7 @@
 def get_scheduler():
     step = [int(x) for x in config.lr_steps.split(',')]
     backbone_lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(
-        step=step, factor=0.1,  base_lr=config.backbone_lr)
+        step=step, factor=0.1, base_lr=config.backbone_lr)
     memory_bank_lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(
         step=step, factor=0.1, base_lr=config.memory_bank_lr)
 
diff --git a/recognition/partial_fc/mxnet/memory_softmax.py b/recognition/partial_fc/mxnet/memory_softmax.py
index 4389872cc..d64159a41 100644
--- a/recognition/partial_fc/mxnet/memory_softmax.py
+++ b/recognition/partial_fc/mxnet/memory_softmax.py
@@ -25,9 +25,10 @@ def forward(self, data, weight, mapping_label, depth):
             #
             fc7 = nd.dot(norm_data, norm_weight, transpose_b=True)
             #
-            mapping_label_onehot = mx.nd.one_hot(
-                indices=mapping_label, depth=depth,
-                on_value=1.0, off_value=0.0)
+            mapping_label_onehot = mx.nd.one_hot(indices=mapping_label,
+                                                 depth=depth,
+                                                 on_value=1.0,
+                                                 off_value=0.0)
             # cosface
             if self.loss_m1 == 1.0 and self.loss_m2 == 0.0:
                 _one_hot = mapping_label_onehot * self.loss_m3
diff --git a/recognition/partial_fc/mxnet/optimizer.py b/recognition/partial_fc/mxnet/optimizer.py
index 826acabb3..b8112e925 100644
--- a/recognition/partial_fc/mxnet/optimizer.py
+++ b/recognition/partial_fc/mxnet/optimizer.py
@@ -21,8 +21,10 @@ def _do_allreduce(self, index, grad):
 
         if isinstance(index, (tuple, list)):
             for i in range(len(index)):
-                hvd.allreduce_(grad[i], average=False,
-                               name=self._prefix + str(index[i]), priority=-i)
+                hvd.allreduce_(grad[i],
+                               average=False,
+                               name=self._prefix + str(index[i]),
+                               priority=-i)
         else:
             hvd.allreduce_(grad, average=False, name=self._prefix + str(index))
 
@@ -58,6 +60,12 @@ def update(self, weight, grad, state, learning_rate):
         if self.momentum > 0:
             kwargs['momentum'] = self.momentum
         if state is not None:
-            nd.sgd_mom_update(weight, grad, state, out=weight, lr=lr, wd=self.wd, **kwargs)
+            nd.sgd_mom_update(weight,
+                              grad,
+                              state,
+                              out=weight,
+                              lr=lr,
+                              wd=self.wd,
+                              **kwargs)
         else:
             raise ValueError
diff --git a/recognition/partial_fc/mxnet/run.sh b/recognition/partial_fc/mxnet/run.sh
index a46ea87e6..04dd414a5 100644
--- a/recognition/partial_fc/mxnet/run.sh
+++ b/recognition/partial_fc/mxnet/run.sh
@@ -1,14 +1,10 @@
 #!/bin/bash
 
 mpirun -np 8 \
--hostfile hosts/host_8 \
--x HOROVOD_CACHE_CAPACITY=8096 \
+-hostfile hosts/host_8 --allow-run-as-root \
 -bind-to none -map-by slot \
 -x LD_LIBRARY_PATH -x PATH \
 -mca pml ob1 -mca btl ^openib \
--mca btl_tcp_if_include eth1 \
+-mca btl_tcp_if_include eth0 \
 -x OMP_NUM_THREADS=2 \
--x MXNET_USE_OPERATOR_TUNING=1 \
--x MXNET_USE_NUM_CORES_OPERATOR_TUNING=1 \
--x MXNET_CUDNN_AUTOTUNE_DEFAULT=1 \
 bash config.sh
diff --git a/recognition/partial_fc/mxnet/setup-utils/requirements.txt b/recognition/partial_fc/mxnet/setup-utils/requirements.txt
index fdd062da0..973a1dd5c 100644
--- a/recognition/partial_fc/mxnet/setup-utils/requirements.txt
+++ b/recognition/partial_fc/mxnet/setup-utils/requirements.txt
@@ -1,5 +1,7 @@
 easydict==1.9
-horovod==0.19.2
+horovod>=0.24
 mxboard==0.1.0
 opencv-python==4.2.0.34
-tqdm==4.48.2
\ No newline at end of file
+tqdm==4.48.2
+scipy
+sklearn
diff --git a/recognition/partial_fc/mxnet/symbol/memonger.py b/recognition/partial_fc/mxnet/symbol/memonger.py
index d054bda20..8ad610b57 100644
--- a/recognition/partial_fc/mxnet/symbol/memonger.py
+++ b/recognition/partial_fc/mxnet/symbol/memonger.py
@@ -1,6 +1,7 @@
 import mxnet as mx
 import math
 
+
 def prod(shape):
     """Get product of the shape.
     """
@@ -138,7 +139,10 @@ def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
 
     for k in range(nbegin):
         info = {}
-        sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
+        sym = make_mirror_plan(sym,
+                               threshold=threshold,
+                               plan_info=info,
+                               **kwargs)
         cost = get_cost(sym, type_dict, **kwargs)
         save_size = info['save_size'] >> 20
         local_size = info['max_size'] >> 20
@@ -147,7 +151,7 @@ def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
             min_cost = cost
         if min_threshold is None or local_size < min_threshold:
             min_threshold = local_size
-        print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
         history.append((cost, threshold, sym))
         threshold = guess
 
@@ -156,13 +160,16 @@ def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
     threshold = min_threshold + step
     if step > 0:
         for k in range(ntrial):
-            sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
+            sym = make_mirror_plan(sym,
+                                   threshold=threshold,
+                                   plan_info=info,
+                                   **kwargs)
             cost = get_cost(sym, type_dict, **kwargs)
-            print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
             history.append((cost, threshold, sym))
             threshold += step
 
-    history.sort(key = lambda x: x[0])
+    history.sort(key=lambda x: x[0])
     cost, threshold, sym = history[0]
     print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
     return sym
diff --git a/recognition/partial_fc/mxnet/symbol/resnet.py b/recognition/partial_fc/mxnet/symbol/resnet.py
index 231f4c91d..bd86106b6 100644
--- a/recognition/partial_fc/mxnet/symbol/resnet.py
+++ b/recognition/partial_fc/mxnet/symbol/resnet.py
@@ -14,7 +14,6 @@
 # KIND, either express or implied.  See the License for the
 # specific language governing permissions and limitations
 # under the License.
-
 '''
 Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
 Original author Wei Wu
@@ -55,7 +54,8 @@ def Act(data, act_type, name):
     return body
 
 
-def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
     """Return ResNet Unit symbol for building ResNet
     Parameters
     ----------
@@ -81,71 +81,176 @@ def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **k
     act_type = kwargs.get('version_act', 'prelu')
     # print('in unit1')
     if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter * 0.25), kernel=(1, 1), stride=stride, pad=(0, 0),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter * 0.25), kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0), no_bias=True,
-                     workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
 
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn3 = mx.symbol.broadcast_mul(bn3, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
     else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn2 = mx.symbol.broadcast_mul(bn2, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
 
 
-def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
     """Return ResNet Unit symbol for building ResNet
     Parameters
     ----------
@@ -171,71 +276,176 @@ def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, *
     act_type = kwargs.get('version_act', 'prelu')
     # print('in unit1')
     if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter * 0.25), kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter * 0.25), kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1, 1), stride=stride, pad=(0, 0), no_bias=True,
-                     workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
 
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn3 = mx.symbol.broadcast_mul(bn3, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
     else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn2 = mx.symbol.broadcast_mul(bn2, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
 
 
-def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
     """Return ResNet Unit symbol for building ResNet
     Parameters
     ----------
@@ -262,66 +472,159 @@ def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **k
     # print('in unit2')
     if bottle_neck:
         # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=int(num_filter * 0.25), kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        conv1 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=int(num_filter * 0.25), kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv2 = Conv(data=act2,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
         act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
-        conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0), no_bias=True,
-                     workspace=workspace, name=name + '_conv3')
+        conv3 = Conv(data=act3,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=conv3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=conv3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             conv3 = mx.symbol.broadcast_mul(conv3, body)
         if dim_match:
             shortcut = data
         else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                            workspace=workspace, name=name + '_sc')
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
         return conv3 + shortcut
     else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
+        conv1 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
         act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
+        conv2 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=conv2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=conv2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             conv2 = mx.symbol.broadcast_mul(conv2, body)
         if dim_match:
             shortcut = data
         else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                            workspace=workspace, name=name + '_sc')
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
         return conv2 + shortcut
 
 
-def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
     """Return ResNet Unit symbol for building ResNet
     Parameters
     ----------
@@ -347,73 +650,182 @@ def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **k
     act_type = kwargs.get('version_act', 'prelu')
     # print('in unit3')
     if bottle_neck:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=int(num_filter * 0.25), kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter * 0.25), kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
         act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1, 1), stride=stride, pad=(0, 0), no_bias=True,
-                     workspace=workspace, name=name + '_conv3')
-        bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn4 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn4')
 
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn4,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn4 = mx.symbol.broadcast_mul(bn4, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
         return bn4 + shortcut
     else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn3 = mx.symbol.broadcast_mul(bn3, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
         return bn3 + shortcut
 
 
-def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
     """Return ResNeXt Unit symbol for building ResNeXt
     Parameters
     ----------
@@ -440,72 +852,138 @@ def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck, *
     act_type = kwargs.get('version_act', 'prelu')
     num_group = 32
     # print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_group=num_group, num_filter=int(num_filter * 0.5), kernel=(1, 1), stride=(1, 1),
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(1, 1),
+                 stride=(1, 1),
                  pad=(0, 0),
-                 no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
     act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_group=num_group, num_filter=int(num_filter * 0.5), kernel=(3, 3), stride=(1, 1),
+    conv2 = Conv(data=act1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(3, 3),
+                 stride=(1, 1),
                  pad=(1, 1),
-                 no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
     act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-    conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1, 1), stride=stride, pad=(0, 0), no_bias=True,
-                 workspace=workspace, name=name + '_conv3')
-    bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
+    conv3 = Conv(data=act2,
+                 num_filter=num_filter,
+                 kernel=(1, 1),
+                 stride=stride,
+                 pad=(0, 0),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv3')
+    bn4 = mx.sym.BatchNorm(data=conv3,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn4')
 
     if use_se:
         # se begin
-        body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-        body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                    name=name + "_se_conv1", workspace=workspace)
+        body = mx.sym.Pooling(data=bn4,
+                              global_pool=True,
+                              kernel=(7, 7),
+                              pool_type='avg',
+                              name=name + '_se_pool1')
+        body = Conv(data=body,
+                    num_filter=num_filter // 16,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv1",
+                    workspace=workspace)
         body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-        body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                    name=name + "_se_conv2", workspace=workspace)
-        body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+        body = Conv(data=body,
+                    num_filter=num_filter,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv2",
+                    workspace=workspace)
+        body = mx.symbol.Activation(data=body,
+                                    act_type='sigmoid',
+                                    name=name + "_se_sigmoid")
         bn4 = mx.symbol.broadcast_mul(bn4, body)
         # se end
 
     if dim_match:
         shortcut = data
     else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                       workspace=workspace, name=name + '_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=bn_mom,
+                                    name=name + '_sc')
     if memonger:
         shortcut._set_attr(mirror_stage='True')
     return bn4 + shortcut
 
 
-def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
+def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck,
+                  **kwargs):
     uv = kwargs.get('version_unit', 3)
     version_input = kwargs.get('version_input', 1)
     if uv == 1:
         if version_input == 0:
-            return residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
+            return residual_unit_v1(data, num_filter, stride, dim_match, name,
+                                    bottle_neck, **kwargs)
         else:
-            return residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
+            return residual_unit_v1_L(data, num_filter, stride, dim_match,
+                                      name, bottle_neck, **kwargs)
     elif uv == 2:
-        return residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
+        return residual_unit_v2(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
     elif uv == 4:
-        return residual_unit_v4(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
+        return residual_unit_v4(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
     else:
-        return residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
+        return residual_unit_v3(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
 
 
 def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
     bn_mom = config.bn_mom
     workspace = config.workspace
-    kwargs = {'version_se': config.net_se,
-              'version_input': config.net_input,
-              'version_output': config.net_output,
-              'version_unit': config.net_unit,
-              'version_act': config.net_act,
-              'bn_mom': bn_mom,
-              'workspace': workspace,
-              'memonger': config.memonger,
-              }
+    kwargs = {
+        'version_se': config.net_se,
+        'version_input': config.net_input,
+        'version_output': config.net_output,
+        'version_unit': config.net_unit,
+        'version_act': config.net_act,
+        'bn_mom': bn_mom,
+        'workspace': workspace,
+        'memonger': config.memonger,
+    }
     """Return ResNet symbol of
     Parameters
     ----------
@@ -530,7 +1008,8 @@ def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
     version_unit = kwargs.get('version_unit', 3)
     act_type = kwargs.get('version_act', 'prelu')
     memonger = kwargs.get('memonger', False)
-    print(version_se, version_input, version_output, version_unit, act_type, memonger)
+    print(version_se, version_input, version_output, version_unit, act_type,
+          memonger)
     num_unit = len(units)
     assert (num_unit == num_stages)
     data = mx.sym.Variable(name='data')
@@ -543,25 +1022,59 @@ def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
         data = mx.sym.identity(data=data, name='id')
         data = data - 127.5
         data = data * 0.0078125
-        body = Conv(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2, 2), pad=(3, 3),
-                    no_bias=True, name="conv0", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
         body = Act(data=body, act_type=act_type, name='relu0')
         # body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
     elif version_input == 2:
-        data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-        body = Conv(data=data, num_filter=filter_list[0], kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                    no_bias=True, name="conv0", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
+        data = mx.sym.BatchNorm(data=data,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn_data')
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
         body = Act(data=body, act_type=act_type, name='relu0')
     else:
         data = mx.sym.identity(data=data, name='id')
         data = data - 127.5
         data = data * 0.0078125
         body = data
-        body = Conv(data=body, num_filter=filter_list[0], kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                    no_bias=True, name="conv0", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
+        body = Conv(data=body,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
         body = Act(data=body, act_type=act_type, name='relu0')
 
     for i in range(num_stages):
@@ -571,18 +1084,36 @@ def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
         # else:
         #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
         #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-        body = residual_unit(body, filter_list[i + 1], (2, 2), False,
-                             name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
+        body = residual_unit(body,
+                             filter_list[i + 1], (2, 2),
+                             False,
+                             name='stage%d_unit%d' % (i + 1, 1),
+                             bottle_neck=bottle_neck,
+                             **kwargs)
         for j in range(units[i] - 1):
-            body = residual_unit(body, filter_list[i + 1], (1, 1), True, name='stage%d_unit%d' % (i + 1, j + 2),
-                                 bottle_neck=bottle_neck, **kwargs)
+            body = residual_unit(body,
+                                 filter_list[i + 1], (1, 1),
+                                 True,
+                                 name='stage%d_unit%d' % (i + 1, j + 2),
+                                 bottle_neck=bottle_neck,
+                                 **kwargs)
     if config.fp16:
         body = mx.sym.Cast(data=body, dtype=np.float32)
 
     if bottle_neck:
-        body = Conv(data=body, num_filter=512, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                    no_bias=True, name="convd", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnd')
+        body = Conv(data=body,
+                    num_filter=512,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    no_bias=True,
+                    name="convd",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnd')
         body = Act(data=body, act_type=act_type, name='relud')
 
     fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
@@ -640,7 +1171,9 @@ def get_symbol():
     elif num_layers == 269:
         units = [3, 30, 48, 8]
     else:
-        raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
+        raise ValueError(
+            "no experiments done on num_layers {}, you can do it yourself".
+            format(num_layers))
 
     net = resnet(units=units,
                  num_stages=num_stages,
diff --git a/recognition/partial_fc/mxnet/symbol/symbol_utils.py b/recognition/partial_fc/mxnet/symbol/symbol_utils.py
index 5da2b278f..73cd3984a 100644
--- a/recognition/partial_fc/mxnet/symbol/symbol_utils.py
+++ b/recognition/partial_fc/mxnet/symbol/symbol_utils.py
@@ -27,58 +27,154 @@ def Act(data, act_type, name):
 bn_mom = config.bn_mom
 
 
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride,
-                              pad=pad, no_bias=True, name='%s%s_conv2d' % (name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' % (name, suffix), fix_gamma=False, momentum=bn_mom)
+def Linear(data,
+           num_filter=1,
+           kernel=(1, 1),
+           stride=(1, 1),
+           pad=(0, 0),
+           num_group=1,
+           name=None,
+           suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=bn_mom)
     return bn
 
 
 def get_fc1(last_conv, num_classes, fc_type, input_channel=512):
     body = last_conv
     if fc_type == 'Z':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
         body = mx.symbol.Dropout(data=body, p=0.4)
         fc1 = body
     elif fc_type == 'E':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
         body = mx.symbol.Dropout(data=body, p=0.4)
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'FC':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'SFC':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-        body = Conv(data=body, num_filter=input_channel, kernel=(3, 3), stride=(2, 2), pad=(1, 1),
-                    no_bias=True, name="convf", num_group=input_channel)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnf')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(3, 3),
+                    stride=(2, 2),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="convf",
+                    num_group=input_channel)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf')
         body = Act(data=body, act_type=config.net_act, name='reluf')
-        body = Conv(data=body, num_filter=input_channel, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="convf2")
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnf2')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(1, 1),
+                    pad=(0, 0),
+                    stride=(1, 1),
+                    name="convf2")
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf2')
         body = Act(data=body, act_type=config.net_act, name='reluf2')
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'GAP':
-        bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
+        bn1 = mx.sym.BatchNorm(data=body,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='bn1')
         relu1 = Act(data=bn1, act_type=config.net_act, name='relu1')
         # Although kernel is not used here when global_pool=True, we should put one
-        pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
+        pool1 = mx.sym.Pooling(data=relu1,
+                               global_pool=True,
+                               kernel=(7, 7),
+                               pool_type='avg',
+                               name='pool1')
         flat = mx.sym.Flatten(data=pool1)
-        fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=flat,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'GNAP':  # mobilefacenet++
         filters_in = 512  # param in mobilefacenet
         if num_classes > filters_in:
-            body = mx.sym.Convolution(data=last_conv, num_filter=num_classes, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                                      no_bias=True, name='convx')
-            body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=0.9, name='convx_bn')
+            body = mx.sym.Convolution(data=last_conv,
+                                      num_filter=num_classes,
+                                      kernel=(1, 1),
+                                      stride=(1, 1),
+                                      pad=(0, 0),
+                                      no_bias=True,
+                                      name='convx')
+            body = mx.sym.BatchNorm(data=body,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=0.9,
+                                    name='convx_bn')
             body = Act(data=body, act_type=config.net_act, name='convx_relu')
             filters_in = num_classes
         else:
             body = last_conv
-        body = mx.sym.BatchNorm(data=body, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6f')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name='bn6f')
 
         spatial_norm = body * body
         spatial_norm = mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
@@ -87,38 +183,93 @@ def get_fc1(last_conv, num_classes, fc_type, input_channel=512):
         spatial_mean = mx.sym.mean(spatial_sqrt)
         spatial_div_inverse = mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
 
-        spatial_attention_inverse = mx.symbol.tile(spatial_div_inverse, reps=(1, filters_in, 1, 1))
+        spatial_attention_inverse = mx.symbol.tile(spatial_div_inverse,
+                                                   reps=(1, filters_in, 1, 1))
         body = body * spatial_attention_inverse
         # body = mx.sym.broadcast_mul(body, spatial_div_inverse)
 
-        fc1 = mx.sym.Pooling(body, kernel=(7, 7), global_pool=True, pool_type='avg')
+        fc1 = mx.sym.Pooling(body,
+                             kernel=(7, 7),
+                             global_pool=True,
+                             pool_type='avg')
         if num_classes < filters_in:
-            fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6w')
-            fc1 = mx.sym.FullyConnected(data=fc1, num_hidden=num_classes, name='pre_fc1')
+            fc1 = mx.sym.BatchNorm(data=fc1,
+                                   fix_gamma=True,
+                                   eps=2e-5,
+                                   momentum=0.9,
+                                   name='bn6w')
+            fc1 = mx.sym.FullyConnected(data=fc1,
+                                        num_hidden=num_classes,
+                                        name='pre_fc1')
         else:
             fc1 = mx.sym.Flatten(data=fc1)
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=0.9,
+                               name='fc1')
     elif fc_type == "GDC":  # mobilefacenet_v1
-        conv_6_dw = Linear(last_conv, num_filter=input_channel, num_group=input_channel, kernel=(7, 7), pad=(0, 0),
-                           stride=(1, 1), name="conv_6dw7_7")
-        conv_6_f = mx.sym.FullyConnected(data=conv_6_dw, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=conv_6_f, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        conv_6_dw = Linear(last_conv,
+                           num_filter=input_channel,
+                           num_group=input_channel,
+                           kernel=(7, 7),
+                           pad=(0, 0),
+                           stride=(1, 1),
+                           name="conv_6dw7_7")
+        conv_6_f = mx.sym.FullyConnected(data=conv_6_dw,
+                                         num_hidden=num_classes,
+                                         name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=conv_6_f,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'F':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
         body = mx.symbol.Dropout(data=body, p=0.4)
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
     elif fc_type == 'G':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
     elif fc_type == 'H':
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
     elif fc_type == 'I':
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     elif fc_type == 'J':
-        fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
     return fc1
 
 
@@ -145,21 +296,54 @@ def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
     workspace = kwargs.get('workspace', 256)
     memonger = kwargs.get('memonger', False)
     # print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                 no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=(1, 1),
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
     act1 = Act(data=bn2, act_type=config.net_act, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                 no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+    conv2 = Conv(data=act1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=stride,
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
 
     if dim_match:
         shortcut = data
     else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                       workspace=workspace, name=name + '_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    momentum=bn_mom,
+                                    eps=2e-5,
+                                    name=name + '_sc')
     if memonger:
         shortcut._set_attr(mirror_stage='True')
     return bn3 + shortcut
@@ -191,68 +375,172 @@ def residual_unit_v1l(data, num_filter, stride, dim_match, name, bottle_neck):
     act_type = config.net_act
     # print('in unit1')
     if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter * 0.25), kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter * 0.25), kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
         act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1, 1), stride=stride, pad=(0, 0), no_bias=True,
-                     workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
 
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn3 = mx.symbol.broadcast_mul(bn3, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
     else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
         act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3, 3), stride=stride, pad=(1, 1),
-                     no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
         if use_se:
             # se begin
-            body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_se_pool1')
-            body = Conv(data=body, num_filter=num_filter // 16, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv1", workspace=workspace)
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
             body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
-            body = Conv(data=body, num_filter=num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0),
-                        name=name + "_se_conv2", workspace=workspace)
-            body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name + "_se_sigmoid")
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
             bn2 = mx.symbol.broadcast_mul(bn2, body)
             # se end
 
         if dim_match:
             shortcut = data
         else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1, 1), stride=stride, no_bias=True,
-                           workspace=workspace, name=name + '_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
         if memonger:
             shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
 
 
 def get_head(data, version_input, num_filter):
@@ -263,18 +551,46 @@ def get_head(data, version_input, num_filter):
     data = data * 0.0078125
     # data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
     if version_input == 0:
-        body = Conv(data=data, num_filter=num_filter, kernel=(7, 7), stride=(2, 2), pad=(3, 3),
-                    no_bias=True, name="conv0", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
+        body = Conv(data=data,
+                    num_filter=num_filter,
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
         body = Act(data=body, act_type=config.net_act, name='relu0')
-        body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max')
+        body = mx.sym.Pooling(data=body,
+                              kernel=(3, 3),
+                              stride=(2, 2),
+                              pad=(1, 1),
+                              pool_type='max')
     else:
         body = data
         _num_filter = min(num_filter, 64)
-        body = Conv(data=body, num_filter=_num_filter, kernel=(3, 3), stride=(1, 1), pad=(1, 1),
-                    no_bias=True, name="conv0", workspace=workspace)
-        body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
+        body = Conv(data=body,
+                    num_filter=_num_filter,
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
         body = Act(data=body, act_type=config.net_act, name='relu0')
         # body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
-        body = residual_unit_v1l(body, _num_filter, (2, 2), False, name='head', bottle_neck=False)
+        body = residual_unit_v1l(body,
+                                 _num_filter, (2, 2),
+                                 False,
+                                 name='head',
+                                 bottle_neck=False)
     return body
diff --git a/recognition/partial_fc/mxnet/train_memory.py b/recognition/partial_fc/mxnet/train_memory.py
index 2219caa67..7a4eb7f59 100644
--- a/recognition/partial_fc/mxnet/train_memory.py
+++ b/recognition/partial_fc/mxnet/train_memory.py
@@ -20,6 +20,18 @@
 
 sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
 os.environ['MXNET_BACKWARD_DO_MIRROR'] = '0'
+os.environ['MXNET_UPDATE_ON_KVSTORE'] = "0"
+os.environ['MXNET_EXEC_ENABLE_ADDTO'] = "1"
+os.environ['MXNET_USE_TENSORRT'] = "0"
+os.environ['MXNET_GPU_WORKER_NTHREADS'] = "2"
+os.environ['MXNET_GPU_COPY_NTHREADS'] = "1"
+os.environ['MXNET_OPTIMIZER_AGGREGATION_SIZE'] = "54"
+os.environ['HOROVOD_CYCLE_TIME'] = "0.1"
+os.environ['HOROVOD_FUSION_THRESHOLD'] = "67108864"
+os.environ['HOROVOD_NUM_NCCL_STREAMS'] = "2"
+os.environ['MXNET_HOROVOD_NUM_GROUPS'] = "16"
+os.environ['MXNET_EXEC_BULK_EXEC_MAX_NODE_TRAIN_FWD'] = "999"
+os.environ['MXNET_EXEC_BULK_EXEC_MAX_NODE_TRAIN_BWD'] = "25"
 
 
 def parse_args():
@@ -31,14 +43,18 @@ def parse_args():
 
     args, rest = parser.parse_known_args()
     default.generate_config(args.loss, args.dataset, args.network)
-    parser.add_argument('--models-root', default="./test", help='root directory to save model.')
+    parser.add_argument('--models-root',
+                        default="./test",
+                        help='root directory to save model.')
     args = parser.parse_args()
     return args
 
 
 def set_logger(logger, rank, models_root):
-    formatter = logging.Formatter("rank-id:" + str(rank) + ":%(asctime)s-%(message)s")
-    file_handler = logging.FileHandler(os.path.join(models_root, "%d_hist.log" % rank))
+    formatter = logging.Formatter("rank-id:" + str(rank) +
+                                  ":%(asctime)s-%(message)s")
+    file_handler = logging.FileHandler(
+        os.path.join(models_root, "%d_hist.log" % rank))
     stream_handler = logging.StreamHandler(sys.stdout)
     file_handler.setFormatter(formatter)
     stream_handler.setFormatter(formatter)
@@ -83,16 +99,25 @@ def train_net():
     num_local = (config.num_classes + size - 1) // size
     num_sample = int(num_local * config.sample_ratio)
     memory_bank = MemoryBank(
-        num_sample=num_sample, num_local=num_local, rank=rank,
-        local_rank=local_rank, embedding_size=config.embedding_size,
-        prefix=prefix_dir, gpu=True)
+        num_sample=num_sample,
+        num_local=num_local,
+        rank=rank,
+        local_rank=local_rank,
+        embedding_size=config.embedding_size,
+        prefix=prefix_dir,
+        gpu=True)
 
     if config.debug:
         train_iter = DummyIter(config.batch_size, data_shape, 1000 * 10000)
     else:
         train_iter = FaceImageIter(
-            batch_size=config.batch_size, data_shape=data_shape, path_imgrec=config.rec,
-            shuffle=True, rand_mirror=True, context=rank, context_num=size)
+            batch_size=config.batch_size,
+            data_shape=data_shape,
+            path_imgrec=config.rec,
+            shuffle=True,
+            rand_mirror=True,
+            context=rank,
+            context_num=size)
     train_data_iter = mx.io.PrefetchingIter(train_iter)
 
     esym, save_symbol = get_symbol_embedding()
@@ -125,9 +150,9 @@ def train_net():
         memory_optimizer=memory_bank_optimizer)
     #
     if not config.debug and local_rank == 0:
-        cb_vert        = CallBackVertification(esym, train_module)
-    cb_speed       = CallBackLogging(rank, size, prefix_dir)
-    cb_save        = CallBackModelSave(save_symbol, train_module, prefix, rank)
+        cb_vert = CallBackVertification(esym, train_module)
+    cb_speed = CallBackLogging(rank, size, prefix_dir)
+    cb_save = CallBackModelSave(save_symbol, train_module, prefix, rank)
     cb_center_save = CallBackCenterSave(memory_bank)
 
     def call_back_fn(params):
diff --git a/recognition/partial_fc/pytorch/README.md b/recognition/partial_fc/pytorch/README.md
index 3528e4791..1e1b5f51b 100644
--- a/recognition/partial_fc/pytorch/README.md
+++ b/recognition/partial_fc/pytorch/README.md
@@ -1 +1,3 @@
-# COMING SOON!
\ No newline at end of file
+# partial fc
+
+PartialFC-Pytorch has been merged into [arcface_torch](https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch).
diff --git a/recognition/partial_fc/unpack_glint360k.py b/recognition/partial_fc/unpack_glint360k.py
index 7ac2e2017..1ec8a9dae 100644
--- a/recognition/partial_fc/unpack_glint360k.py
+++ b/recognition/partial_fc/unpack_glint360k.py
@@ -39,9 +39,11 @@ def main(args):
                 class_path = os.path.join(args.output, "id_%d" % label)
                 if not os.path.exists(class_path):
                     os.makedirs(class_path)
-                _img = mx.image.imdecode(_img).asnumpy()[:, :, ::-1]  # to bgr
+
                 image_path = os.path.join(class_path, "%d_%d.jpg" % (label, imgid))
-                cv2.imwrite(image_path, _img)
+                with open(image_path, "wb") as ff:
+                    ff.write(_img)
+
                 imgid += 1
                 if imgid % 10000 == 0:
                     print(imgid)
diff --git a/recognition/subcenter_arcface/README.md b/recognition/subcenter_arcface/README.md
new file mode 100644
index 000000000..15b9e6b40
--- /dev/null
+++ b/recognition/subcenter_arcface/README.md
@@ -0,0 +1,48 @@
+
+## Subcenter ArcFace
+
+### 1. Motivation
+
+We introduce one extra hyperparameter (subcenter number `loss_K`) to ArcFace to relax the intra-class compactness constraint. In our experiments, we find ``loss_K=3`` can achieve a good balance between accuracy and robustness.
+
+![difference](https://insightface.ai/assets/img/github/subcenterarcfacediff.png)
+
+### 2. Implementation
+
+The training process of Subcenter ArcFace is almost same as [ArcFace](https://github.com/deepinsight/insightface/tree/master/recognition/ArcFace)
+The increased GPU memory consumption can be easily alleviated by our parallel framework.
+
+![framework](https://insightface.ai/assets/img/github/subcenterarcfaceframework.png)
+
+### 3. Training Dataset
+
+1. MS1MV0 (The noise rate is around 50%), download link ([baidu drive](https://pan.baidu.com/s/1bSamN5CLiSrxOuGi-Lx7tw), code ``8ql0``)  ([dropbox](https://www.dropbox.com/sh/y2mj25uj440f7bl/AABc7pCJvUvxEcmXs8WYi9Zaa?dl=0))
+
+### 4. Training Steps
+
+1). Train Sub-center ArcFace (``loss_K=3``) on MS1MV0.
+
+2). Drop non-dominant subcenters and high-confident noisy data (``>75 degrees``). 
+
+  ``
+  python drop.py --data <ms1mv0-path> --model <step-1-pretrained-model> --threshold 75 --k 3 --output <ms1mv0-drop75-path>
+  ``
+  
+3). Train ArcFace on the new ``MS1MV0-Drop75`` dataset.
+
+### 5. Pretrained Models and Logs
+  [baidu drive](https://pan.baidu.com/s/1yikOW1Xzm1XIHu0uv0RdRw) code ``3jsh``. [gdrive](https://drive.google.com/file/d/1h8Ybz6mJ7n2IfLbDv2HUU37OdVHn7YPg/view?usp=sharing)
+
+### Citation
+
+If you find *Sub-center ArcFace* useful in your research, please consider to cite the following related papers:
+
+```
+@inproceedings{deng2020subcenter,
+  title={Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces},
+  author={Deng, Jiankang and Guo, Jia and Liu, Tongliang and Gong, Mingming and Zafeiriou, Stefanos},
+  booktitle={Proceedings of the IEEE Conference on European Conference on Computer Vision},
+  year={2020}
+}
+```
+
diff --git a/recognition/subcenter_arcface/common/build_eval_pack.py b/recognition/subcenter_arcface/common/build_eval_pack.py
new file mode 100644
index 000000000..23208ceeb
--- /dev/null
+++ b/recognition/subcenter_arcface/common/build_eval_pack.py
@@ -0,0 +1,136 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+#import mxnet as mx
+#from mxnet import ndarray as nd
+import argparse
+import cv2
+import pickle
+import numpy as np
+import sys
+import os
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', '..', 'common'))
+sys.path.append(
+    os.path.join(os.path.dirname(__file__), '..', '..', 'RetinaFace'))
+import face_align
+from retinaface import RetinaFace
+
+
+def to_rgb(img):
+    w, h = img.shape
+    ret = np.empty((w, h, 3), dtype=np.uint8)
+    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
+    return ret
+
+
+def IOU(Reframe, GTframe):
+    x1 = Reframe[0]
+    y1 = Reframe[1]
+    width1 = Reframe[2] - Reframe[0]
+    height1 = Reframe[3] - Reframe[1]
+
+    x2 = GTframe[0]
+    y2 = GTframe[1]
+    width2 = GTframe[2] - GTframe[0]
+    height2 = GTframe[3] - GTframe[1]
+
+    endx = max(x1 + width1, x2 + width2)
+    startx = min(x1, x2)
+    width = width1 + width2 - (endx - startx)
+
+    endy = max(y1 + height1, y2 + height2)
+    starty = min(y1, y2)
+    height = height1 + height2 - (endy - starty)
+
+    if width <= 0 or height <= 0:
+        ratio = 0
+    else:
+        Area = width * height
+        Area1 = width1 * height1
+        Area2 = width2 * height2
+        ratio = Area * 1. / (Area1 + Area2 - Area)
+    return ratio
+
+
+parser = argparse.ArgumentParser(description='Package eval images')
+# general
+parser.add_argument('--data-dir', default='', help='')
+parser.add_argument('--image-size', type=int, default=112, help='')
+parser.add_argument('--gpu', type=int, default=0, help='')
+parser.add_argument('--det-prefix', type=str, default='./model/R50', help='')
+parser.add_argument('--output', default='./', help='path to save.')
+parser.add_argument('--align-mode', default='arcface', help='align mode.')
+args = parser.parse_args()
+
+gpu_id = args.gpu
+
+detector = RetinaFace(args.det_prefix, 0, gpu_id, network='net3')
+target_size = 400
+max_size = 800
+
+
+def get_norm_crop(image_path):
+    im = cv2.imread(image_path)
+    im_shape = im.shape
+    im_size_min = np.min(im_shape[0:2])
+    im_size_max = np.max(im_shape[0:2])
+    im_scale = float(target_size) / float(im_size_min)
+    # prevent bigger axis from being more than max_size:
+    if np.round(im_scale * im_size_max) > max_size:
+        im_scale = float(max_size) / float(im_size_max)
+    bbox, landmark = detector.detect(im, threshold=0.5, scales=[im_scale])
+    #print(im.shape, bbox.shape, landmark.shape)
+    if bbox.shape[0] == 0:
+        bbox, landmark = detector.detect(
+            im,
+            threshold=0.05,
+            scales=[im_scale * 0.75, im_scale, im_scale * 2.0])
+        print('refine', im.shape, bbox.shape, landmark.shape)
+    nrof_faces = bbox.shape[0]
+    if nrof_faces > 0:
+        det = bbox[:, 0:4]
+        img_size = np.asarray(im.shape)[0:2]
+        bindex = 0
+        if nrof_faces > 1:
+            bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                           det[:, 1])
+            img_center = img_size / 2
+            offsets = np.vstack([(det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                                 (det[:, 1] + det[:, 3]) / 2 - img_center[0]])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            bindex = np.argmax(bounding_box_size - offset_dist_squared *
+                               2.0)  # some extra weight on the centering
+        #_bbox = bounding_boxes[bindex, 0:4]
+        _landmark = landmark[bindex]
+        warped = face_align.norm_crop(im,
+                                      landmark=_landmark,
+                                      image_size=args.image_size,
+                                      mode=args.align_mode)
+        return warped
+    else:
+        return None
+
+
+bins = []
+issame_list = []
+pp = 0
+for line in open(os.path.join(args.data_dir, 'pairs_label.txt'), 'r'):
+    pp += 1
+    if pp % 100 == 0:
+        print('processing', pp)
+    line = line.strip().split()
+    assert len(line) == 3
+    path1 = os.path.join(args.data_dir, line[0])
+    path2 = os.path.join(args.data_dir, line[1])
+    im1 = get_norm_crop(path1)
+    im2 = get_norm_crop(path2)
+    issame = True
+    if line[2] == '0':
+        issame = False
+    issame_list.append(issame)
+    for im in [im1, im2]:
+        _, s = cv2.imencode('.jpg', im)
+        bins.append(s)
+
+with open(args.output, 'wb') as f:
+    pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
diff --git a/recognition/subcenter_arcface/common/face_align.py b/recognition/subcenter_arcface/common/face_align.py
new file mode 100644
index 000000000..4f48a7691
--- /dev/null
+++ b/recognition/subcenter_arcface/common/face_align.py
@@ -0,0 +1,71 @@
+import cv2
+import numpy as np
+from skimage import transform as trans
+
+src1 = np.array([[51.642, 50.115], [57.617, 49.990], [35.740, 69.007],
+                 [51.157, 89.050], [57.025, 89.702]],
+                dtype=np.float32)
+#<--left
+src2 = np.array([[45.031, 50.118], [65.568, 50.872], [39.677, 68.111],
+                 [45.177, 86.190], [64.246, 86.758]],
+                dtype=np.float32)
+
+#---frontal
+src3 = np.array([[39.730, 51.138], [72.270, 51.138], [56.000, 68.493],
+                 [42.463, 87.010], [69.537, 87.010]],
+                dtype=np.float32)
+
+#-->right
+src4 = np.array([[46.845, 50.872], [67.382, 50.118], [72.737, 68.111],
+                 [48.167, 86.758], [67.236, 86.190]],
+                dtype=np.float32)
+
+#-->right profile
+src5 = np.array([[54.796, 49.990], [60.771, 50.115], [76.673, 69.007],
+                 [55.388, 89.702], [61.257, 89.050]],
+                dtype=np.float32)
+
+src = np.array([src1, src2, src3, src4, src5])
+src_map = {112: src, 224: src * 2}
+
+arcface_src = np.array(
+    [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+     [41.5493, 92.3655], [70.7299, 92.2041]],
+    dtype=np.float32)
+
+arcface_src = np.expand_dims(arcface_src, axis=0)
+
+# In[66]:
+
+
+# lmk is prediction; src is template
+def estimate_norm(lmk, image_size=112, mode='arcface'):
+    assert lmk.shape == (5, 2)
+    tform = trans.SimilarityTransform()
+    lmk_tran = np.insert(lmk, 2, values=np.ones(5), axis=1)
+    min_M = []
+    min_index = []
+    min_error = float('inf')
+    if mode == 'arcface':
+        assert image_size == 112
+        src = arcface_src
+    else:
+        src = src_map[image_size]
+    for i in np.arange(src.shape[0]):
+        tform.estimate(lmk, src[i])
+        M = tform.params[0:2, :]
+        results = np.dot(M, lmk_tran.T)
+        results = results.T
+        error = np.sum(np.sqrt(np.sum((results - src[i])**2, axis=1)))
+        #         print(error)
+        if error < min_error:
+            min_error = error
+            min_M = M
+            min_index = i
+    return min_M, min_index
+
+
+def norm_crop(img, landmark, image_size=112, mode='arcface'):
+    M, pose_index = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped
diff --git a/recognition/subcenter_arcface/common/flops_counter.py b/recognition/subcenter_arcface/common/flops_counter.py
new file mode 100644
index 000000000..809424142
--- /dev/null
+++ b/recognition/subcenter_arcface/common/flops_counter.py
@@ -0,0 +1,120 @@
+'''
+@author: insightface
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+import os
+import json
+import argparse
+import numpy as np
+import mxnet as mx
+
+
+def is_no_bias(attr):
+    ret = False
+    if 'no_bias' in attr and (attr['no_bias'] == True
+                              or attr['no_bias'] == 'True'):
+        ret = True
+    return ret
+
+
+def count_fc_flops(input_filter, output_filter, attr):
+    #print(input_filter, output_filter ,attr)
+    ret = 2 * input_filter * output_filter
+    if is_no_bias(attr):
+        ret -= output_filter
+    return int(ret)
+
+
+def count_conv_flops(input_shape, output_shape, attr):
+    kernel = attr['kernel'][1:-1].split(',')
+    kernel = [int(x) for x in kernel]
+
+    #print('kernel', kernel)
+    if is_no_bias(attr):
+        ret = (2 * input_shape[1] * kernel[0] * kernel[1] -
+               1) * output_shape[2] * output_shape[3] * output_shape[1]
+    else:
+        ret = 2 * input_shape[1] * kernel[0] * kernel[1] * output_shape[
+            2] * output_shape[3] * output_shape[1]
+    num_group = 1
+    if 'num_group' in attr:
+        num_group = int(attr['num_group'])
+    ret /= num_group
+    return int(ret)
+
+
+def count_flops(sym, **data_shapes):
+    all_layers = sym.get_internals()
+    #print(all_layers)
+    arg_shapes, out_shapes, aux_shapes = all_layers.infer_shape(**data_shapes)
+    out_shape_dict = dict(zip(all_layers.list_outputs(), out_shapes))
+
+    nodes = json.loads(sym.tojson())['nodes']
+    nodeid_shape = {}
+    for nodeid, node in enumerate(nodes):
+        name = node['name']
+        layer_name = name + "_output"
+        if layer_name in out_shape_dict:
+            nodeid_shape[nodeid] = out_shape_dict[layer_name]
+    #print(nodeid_shape)
+    FLOPs = 0
+    for nodeid, node in enumerate(nodes):
+        flops = 0
+        if node['op'] == 'Convolution':
+            output_shape = nodeid_shape[nodeid]
+            name = node['name']
+            attr = node['attrs']
+            input_nodeid = node['inputs'][0][0]
+            input_shape = nodeid_shape[input_nodeid]
+            flops = count_conv_flops(input_shape, output_shape, attr)
+        elif node['op'] == 'FullyConnected':
+            attr = node['attrs']
+            output_shape = nodeid_shape[nodeid]
+            input_nodeid = node['inputs'][0][0]
+            input_shape = nodeid_shape[input_nodeid]
+            output_filter = output_shape[1]
+            input_filter = input_shape[1] * input_shape[2] * input_shape[3]
+            #assert len(input_shape)==4 and input_shape[2]==1 and input_shape[3]==1
+            flops = count_fc_flops(input_filter, output_filter, attr)
+        #print(node, flops)
+        FLOPs += flops
+
+    return FLOPs
+
+
+def flops_str(FLOPs):
+    preset = [(1e12, 'T'), (1e9, 'G'), (1e6, 'M'), (1e3, 'K')]
+
+    for p in preset:
+        if FLOPs // p[0] > 0:
+            N = FLOPs / p[0]
+            ret = "%.1f%s" % (N, p[1])
+            return ret
+    ret = "%.1f" % (FLOPs)
+    return ret
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='flops counter')
+    # general
+    #parser.add_argument('--model', default='../models2/y2-arcface-retinat1/model,1', help='path to load model.')
+    #parser.add_argument('--model', default='../models2/r100fc-arcface-retinaa/model,1', help='path to load model.')
+    parser.add_argument('--model',
+                        default='../models2/r50fc-arcface-emore/model,1',
+                        help='path to load model.')
+    args = parser.parse_args()
+    _vec = args.model.split(',')
+    assert len(_vec) == 2
+    prefix = _vec[0]
+    epoch = int(_vec[1])
+    print('loading', prefix, epoch)
+    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+    all_layers = sym.get_internals()
+    sym = all_layers['fc1_output']
+    FLOPs = count_flops(sym, data=(1, 3, 112, 112))
+    print('FLOPs:', FLOPs)
diff --git a/recognition/subcenter_arcface/common/rec2image.py b/recognition/subcenter_arcface/common/rec2image.py
new file mode 100644
index 000000000..21e5ec482
--- /dev/null
+++ b/recognition/subcenter_arcface/common/rec2image.py
@@ -0,0 +1,60 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import sys
+import mxnet as mx
+from mxnet import ndarray as nd
+import random
+import argparse
+import cv2
+import time
+import sklearn
+import numpy as np
+
+
+def main(args):
+    include_datasets = args.include.split(',')
+    rec_list = []
+    for ds in include_datasets:
+        path_imgrec = os.path.join(ds, 'train.rec')
+        path_imgidx = os.path.join(ds, 'train.idx')
+        imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+        rec_list.append(imgrec)
+    if not os.path.exists(args.output):
+        os.makedirs(args.output)
+    for ds_id in range(len(rec_list)):
+        id_list = []
+        imgrec = rec_list[ds_id]
+        s = imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        assert header.flag > 0
+        print('header0 label', header.label)
+        header0 = (int(header.label[0]), int(header.label[1]))
+        seq_identity = range(int(header.label[0]), int(header.label[1]))
+        pp = 0
+        for identity in seq_identity:
+            id_dir = os.path.join(args.output, "%d_%d" % (ds_id, identity))
+            os.makedirs(id_dir)
+            pp += 1
+            if pp % 10 == 0:
+                print('processing id', pp)
+            s = imgrec.read_idx(identity)
+            header, _ = mx.recordio.unpack(s)
+            imgid = 0
+            for _idx in range(int(header.label[0]), int(header.label[1])):
+                s = imgrec.read_idx(_idx)
+                _header, _img = mx.recordio.unpack(s)
+                _img = mx.image.imdecode(_img).asnumpy()[:, :, ::-1]  # to bgr
+                image_path = os.path.join(id_dir, "%d.jpg" % imgid)
+                cv2.imwrite(image_path, _img)
+                imgid += 1
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='do dataset merge')
+    # general
+    parser.add_argument('--include', default='', type=str, help='')
+    parser.add_argument('--output', default='', type=str, help='')
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/subcenter_arcface/common/rec2shufrec.py b/recognition/subcenter_arcface/common/rec2shufrec.py
new file mode 100644
index 000000000..cf916b4ad
--- /dev/null
+++ b/recognition/subcenter_arcface/common/rec2shufrec.py
@@ -0,0 +1,72 @@
+import os
+import os.path as osp
+import sys
+import datetime
+import glob
+import shutil
+import numbers
+import mxnet as mx
+from mxnet import ndarray as nd
+from mxnet import io
+from mxnet import recordio
+import random
+import argparse
+import cv2
+import time
+import numpy as np
+
+def main(args):
+    ds = args.input
+    path_imgrec = osp.join(ds, 'train.rec')
+    path_imgidx = osp.join(ds, 'train.idx')
+    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+    if not osp.exists(args.output):
+        os.makedirs(args.output)
+    writer = mx.recordio.MXRecordIO(osp.join(args.output, 'train.rec'), 'w')
+    s = imgrec.read_idx(0)
+    header, _ = recordio.unpack(s)
+    if header.flag > 0:
+        print('header0 label', header.label)
+        header0 = (int(header.label[0]), int(header.label[1]))
+        imgidx = list(range(1, int(header.label[0])))
+    else:
+        imgidx = list(imgrec.keys)
+    random.shuffle(imgidx)
+    label_stat = None
+    print('total images:', len(imgidx))
+    for i, idx in enumerate(imgidx):
+        if i%10000==0:
+            print('processing', i, idx)
+        s = imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        label = header.label
+        if not isinstance(label, numbers.Number):
+            label = label[0]
+        if label_stat is None:
+            label_stat = [label, label]
+        else:
+            label_stat[0] = min(label, label_stat[0])
+            label_stat[1] = max(label, label_stat[1])
+        wheader = mx.recordio.IRHeader(0, label, i, 0)
+        ws = mx.recordio.pack(wheader, img)
+        writer.write(ws)
+    print('label_stat:', label_stat)
+    writer.close()
+    if args.copy_vers:
+        for binfile in glob.glob(osp.join(args.input, '*.bin')):
+            target_file = osp.join(args.output, binfile.split('/')[-1])
+            shutil.copyfile(binfile, target_file)
+    with open(osp.join(args.output, 'property'), 'w') as f:
+        f.write("%d,112,112\n"%(int(label_stat[1])+1))
+        f.write("%d\n"%len(imgidx))
+        f.write("shuffled\n")
+        f.write("%s\n"%(datetime.datetime.now()))
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='convert rec to shuffled rec')
+    # general
+    parser.add_argument('--input', default='', type=str, help='')
+    parser.add_argument('--output', default='', type=str, help='')
+    parser.add_argument('--copy-vers', action='store_true', help='copy verification bins')
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/subcenter_arcface/common/rec_builder.py b/recognition/subcenter_arcface/common/rec_builder.py
new file mode 100644
index 000000000..1d517152f
--- /dev/null
+++ b/recognition/subcenter_arcface/common/rec_builder.py
@@ -0,0 +1,109 @@
+import os
+import sys
+import mxnet as mx
+from mxnet import ndarray as nd
+import random
+import argparse
+import cv2
+import time
+import sklearn
+import numpy as np
+
+
+class SeqRecBuilder():
+    def __init__(self, path, image_size=(112, 112)):
+        self.path = path
+        self.image_size = image_size
+        self.last_label = -1
+        self.widx = 0
+        if not os.path.exists(path):
+            os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(
+            os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'),
+            'w')
+        self.label_stat = [-1, -1]
+
+    def add(self, label, img, is_image=True):
+        #img should be BGR
+        #if self.sis:
+        #    assert label>=self.last_label
+        idx = self.widx
+        self.widx += 1
+        header = mx.recordio.IRHeader(0, label, idx, 0)
+        if is_image:
+            s = mx.recordio.pack_img(header, img, quality=95, img_fmt='.jpg')
+        else:
+            s = mx.recordio.pack(header, img)
+        self.writer.write_idx(idx, s)
+        if self.label_stat[0] < 0:
+            self.label_stat = [label, label]
+        else:
+            self.label_stat[0] = min(self.label_stat[0], label)
+            self.label_stat[1] = max(self.label_stat[1], label)
+
+    def close(self):
+        with open(os.path.join(self.path, 'property'), 'w') as f:
+            f.write("%d,%d,%d\n" % (self.label_stat[1] + 1, self.image_size[0],
+                                    self.image_size[1]))
+
+
+class RecBuilder():
+    def __init__(self, path, image_size=(112, 112)):
+        self.path = path
+        self.image_size = image_size
+        self.last_label = -1
+        self.widx = 1
+        if not os.path.exists(path):
+            os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(
+            os.path.join(path, 'train.idx'), os.path.join(path, 'train.rec'),
+            'w')
+        self.label_stat = [-1, -1]
+        self.identities = []
+
+    def add(self, label, imgs):
+        #img should be BGR
+        assert label >= 0
+        assert label > self.last_label
+        assert len(imgs) > 0
+        idflag = [self.widx, -1]
+        for img in imgs:
+            idx = self.widx
+            self.widx += 1
+            header = mx.recordio.IRHeader(0, label, idx, 0)
+            if isinstance(img, np.ndarray):
+                s = mx.recordio.pack_img(header,
+                                         img,
+                                         quality=95,
+                                         img_fmt='.jpg')
+            else:
+                s = mx.recordio.pack(header, img)
+            self.writer.write_idx(idx, s)
+        idflag[1] = self.widx
+        self.identities.append(idflag)
+        if self.label_stat[0] < 0:
+            self.label_stat = [label, label]
+        else:
+            self.label_stat[0] = min(self.label_stat[0], label)
+            self.label_stat[1] = max(self.label_stat[1], label)
+        self.last_label = label
+
+    def close(self):
+        id_idx = self.widx
+        for id_flag in self.identities:
+            idx = self.widx
+            self.widx += 1
+            _header = mx.recordio.IRHeader(0, id_flag, idx, 0)
+            s = mx.recordio.pack(_header, b'')
+            self.writer.write_idx(idx, s)
+
+        print('id0:', (id_idx, self.widx))
+        idx = 0
+        _header = mx.recordio.IRHeader(0, (id_idx, self.widx), idx, 1)
+        s = mx.recordio.pack(_header, b'')
+        self.writer.write_idx(idx, s)
+        print('label stat:', self.label_stat)
+
+        with open(os.path.join(self.path, 'property'), 'w') as f:
+            f.write("%d,%d,%d\n" % (self.label_stat[1] + 1, self.image_size[0],
+                                    self.image_size[1]))
diff --git a/recognition/subcenter_arcface/common/verification.py b/recognition/subcenter_arcface/common/verification.py
new file mode 100644
index 000000000..f46942a25
--- /dev/null
+++ b/recognition/subcenter_arcface/common/verification.py
@@ -0,0 +1,423 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset 
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import argparse
+import sys
+import numpy as np
+from scipy import misc
+from sklearn.model_selection import KFold
+from scipy import interpolate
+import sklearn
+import cv2
+import math
+import datetime
+import pickle
+from sklearn.decomposition import PCA
+import mxnet as mx
+from mxnet import ndarray as nd
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+    #print('pca', pca)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        #print('train_set', train_set)
+        #print('test_set', test_set)
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            #print(_embed_train.shape)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            #print(embed1.shape, embed2.shape)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        #print('threshold', thresholds[best_threshold_index])
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx,
+                 threshold_idx], fprs[fold_idx,
+                                      threshold_idx], _ = calculate_accuracy(
+                                          threshold, dist[test_set],
+                                          actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    #print(true_accept, false_accept)
+    #print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
+
+
+def load_bin(path, image_size):
+    try:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f)  #py2
+    except UnicodeDecodeError as e:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f, encoding='bytes')  #py3
+    data_list = []
+    for flip in [0, 1]:
+        data = nd.empty(
+            (len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for i in range(len(issame_list) * 2):
+        _bin = bins[i]
+        img = mx.image.imdecode(_bin)
+        if img.shape[1] != image_size[0]:
+            img = mx.image.resize_short(img, image_size[0])
+        img = nd.transpose(img, axes=(2, 0, 1))
+        for flip in [0, 1]:
+            if flip == 1:
+                img = mx.ndarray.flip(data=img, axis=2)
+            data_list[flip][i][:] = img
+        if i % 1000 == 0:
+            print('loading bin', i)
+    print(data_list[0].shape)
+    return (data_list, issame_list)
+
+
+def test(data_set,
+         mx_model,
+         batch_size,
+         nfolds=10,
+         data_extra=None,
+         label_shape=None):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    model = mx_model
+    embeddings_list = []
+    if data_extra is not None:
+        _data_extra = nd.array(data_extra)
+    time_consumed = 0.0
+    if label_shape is None:
+        _label = nd.ones((batch_size, ))
+    else:
+        _label = nd.ones(label_shape)
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            #print(_data.shape, _label.shape)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data, ), label=(_label, ))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label, ))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            #_arg, _aux = model.get_params()
+            #__arg = {}
+            #for k,v in _arg.iteritems():
+            #  __arg[k] = v.as_in_context(_ctx)
+            #_arg = __arg
+            #_arg["data"] = _data.as_in_context(_ctx)
+            #_arg["softmax_label"] = _label.as_in_context(_ctx)
+            #for k,v in _arg.iteritems():
+            #  print(k,v.context)
+            #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
+            #exe.forward(is_train=False)
+            #net_out = exe.outputs
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            #print(_embeddings.shape)
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            #print(_em.shape, _norm)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
+    #acc1, std1 = np.mean(accuracy), np.std(accuracy)
+
+    #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
+    #embeddings = np.concatenate(embeddings_list, axis=1)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    print('infer time', time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings,
+                                                 issame_list,
+                                                 nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser(description='do verification')
+    # general
+    parser.add_argument('--data-dir', default='', help='')
+    parser.add_argument('--model',
+                        default='../model/softmax,50',
+                        help='path to load model.')
+    parser.add_argument('--target',
+                        default='lfw,cfp_ff,cfp_fp,agedb_30',
+                        help='test targets.')
+    parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+    parser.add_argument('--batch-size', default=32, type=int, help='')
+    parser.add_argument('--max', default='', type=str, help='')
+    parser.add_argument('--mode', default=0, type=int, help='')
+    parser.add_argument('--nfolds', default=10, type=int, help='')
+    args = parser.parse_args()
+    #sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+    #import face_image
+    #prop = face_image.load_property(args.data_dir)
+    #image_size = prop.image_size
+    image_size = [112, 112]
+    print('image_size', image_size)
+    ctx = mx.gpu(args.gpu)
+    nets = []
+    vec = args.model.split(',')
+    prefix = args.model.split(',')[0]
+    epochs = []
+    if len(vec) == 1:
+        pdir = os.path.dirname(prefix)
+        for fname in os.listdir(pdir):
+            if not fname.endswith('.params'):
+                continue
+            _file = os.path.join(pdir, fname)
+            if _file.startswith(prefix):
+                epoch = int(fname.split('.')[0].split('-')[1])
+                epochs.append(epoch)
+        epochs = sorted(epochs, reverse=True)
+        if len(args.max) > 0:
+            _max = [int(x) for x in args.max.split(',')]
+            assert len(_max) == 2
+            if len(epochs) > _max[1]:
+                epochs = epochs[_max[0]:_max[1]]
+
+    else:
+        epochs = [int(x) for x in vec[1].split('|')]
+    print('model number', len(epochs))
+    time0 = datetime.datetime.now()
+    for epoch in epochs:
+        print('loading', prefix, epoch)
+        sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+        #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
+        all_layers = sym.get_internals()
+        sym = all_layers['fc1_output']
+        model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+        #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+        model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+                                          image_size[1]))])
+        model.set_params(arg_params, aux_params)
+        nets.append(model)
+    time_now = datetime.datetime.now()
+    diff = time_now - time0
+    print('model loading time', diff.total_seconds())
+
+    ver_list = []
+    ver_name_list = []
+    for name in args.target.split(','):
+        path = os.path.join(args.data_dir, name + ".bin")
+        if os.path.exists(path):
+            print('loading.. ', name)
+            data_set = load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+
+    if args.mode == 0:
+        for i in range(len(ver_list)):
+            results = []
+            for model in nets:
+                acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+                    ver_list[i], model, args.batch_size, args.nfolds)
+                print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
+                print('[%s]Accuracy: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc1, std1))
+                print('[%s]Accuracy-Flip: %1.5f+-%1.5f' %
+                      (ver_name_list[i], acc2, std2))
+                results.append(acc2)
+            print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
+    elif args.mode == 1:
+        model = nets[0]
+        test_badcase(ver_list[0], model, args.batch_size, args.target)
+    else:
+        model = nets[0]
+        dumpR(ver_list[0], model, args.batch_size, args.target)
diff --git a/recognition/subcenter_arcface/drop.py b/recognition/subcenter_arcface/drop.py
new file mode 100644
index 000000000..24d20bae8
--- /dev/null
+++ b/recognition/subcenter_arcface/drop.py
@@ -0,0 +1,212 @@
+import os
+import shutil
+import datetime
+import sys
+from mxnet import ndarray as nd
+import mxnet as mx
+import random
+import argparse
+import numbers
+import cv2
+import time
+import pickle
+import sklearn
+import sklearn.preprocessing
+from easydict import EasyDict as edict
+import numpy as np
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+from rec_builder import *
+
+
+def get_embedding(args, imgrec, a, b, image_size, model):
+    ocontents = []
+    for idx in range(a, b):
+        s = imgrec.read_idx(idx)
+        ocontents.append(s)
+    embeddings = None
+    #print(len(ocontents))
+    ba = 0
+    rlabel = -1
+    imgs = []
+    contents = []
+    while True:
+        bb = min(ba + args.batch_size, len(ocontents))
+        if ba >= bb:
+            break
+        _batch_size = bb - ba
+        #_batch_size2 = max(_batch_size, args.ctx_num)
+        _batch_size2 = _batch_size
+        if _batch_size % args.ctx_num != 0:
+            _batch_size2 = ((_batch_size // args.ctx_num) + 1) * args.ctx_num
+        data = np.zeros((_batch_size2, 3, image_size[0], image_size[1]))
+        count = bb - ba
+        ii = 0
+        for i in range(ba, bb):
+            header, img = mx.recordio.unpack(ocontents[i])
+            contents.append(img)
+            label = header.label
+            if not isinstance(label, numbers.Number):
+                label = label[0]
+            if rlabel < 0:
+                rlabel = int(label)
+
+            img = mx.image.imdecode(img)
+            rgb = img.asnumpy()
+            bgr = rgb[:, :, ::-1]
+            imgs.append(bgr)
+            img = rgb.transpose((2, 0, 1))
+            data[ii] = img
+            ii += 1
+        while ii < _batch_size2:
+            data[ii] = data[0]
+            ii += 1
+        nddata = nd.array(data)
+        db = mx.io.DataBatch(data=(nddata, ))
+        model.forward(db, is_train=False)
+        net_out = model.get_outputs()
+        net_out = net_out[0].asnumpy()
+        if embeddings is None:
+            embeddings = np.zeros((len(ocontents), net_out.shape[1]))
+        embeddings[ba:bb, :] = net_out[0:_batch_size, :]
+        ba = bb
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    return embeddings, rlabel, contents
+
+
+def main(args):
+    print(args)
+    image_size = (112, 112)
+    print('image_size', image_size)
+    vec = args.model.split(',')
+    prefix = vec[0]
+    epoch = int(vec[1])
+    print('loading', prefix, epoch)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    args.ctx_num = len(ctx)
+    args.batch_size *= args.ctx_num
+    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+    W = None
+    i = 0
+    while True:
+        key = 'fc7_%d_weight' % i
+        i += 1
+        if key not in arg_params:
+            break
+        _W = arg_params[key].asnumpy()
+        #_W = _W.reshape( (-1, 10, 512) )
+        if W is None:
+            W = _W
+        else:
+            W = np.concatenate((W, _W), axis=0)
+    K = args.k
+    W = sklearn.preprocessing.normalize(W)
+    W = W.reshape((-1, K, 512))
+    all_layers = sym.get_internals()
+    sym = all_layers['fc1_output']
+    model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+    model.bind(data_shapes=[('data', (args.ctx_num, 3, image_size[0],
+                                      image_size[1]))])
+    model.set_params(arg_params, aux_params)
+    print('W:', W.shape)
+    path_imgrec = os.path.join(args.data, 'train.rec')
+    path_imgidx = os.path.join(args.data, 'train.idx')
+    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
+    id_list = []
+    s = imgrec.read_idx(0)
+    header, _ = mx.recordio.unpack(s)
+    assert header.flag > 0
+    print('header0 label', header.label)
+    header0 = (int(header.label[0]), int(header.label[1]))
+    #assert(header.flag==1)
+    imgidx = range(1, int(header.label[0]))
+    id2range = {}
+    a, b = int(header.label[0]), int(header.label[1])
+    seq_identity = range(a, b)
+    print(len(seq_identity))
+    image_count = 0
+    pp = 0
+    for wid, identity in enumerate(seq_identity):
+        pp += 1
+        s = imgrec.read_idx(identity)
+        header, _ = mx.recordio.unpack(s)
+        contents = []
+        a, b = int(header.label[0]), int(header.label[1])
+        _count = b - a
+        id_list.append((wid, a, b, _count))
+        image_count += _count
+    pp = 0
+    if not os.path.exists(args.output):
+        os.makedirs(args.output)
+    ret = np.zeros((image_count, K + 1), dtype=np.float32)
+    output_dir = args.output
+    builder = SeqRecBuilder(output_dir)
+    print(ret.shape)
+    imid = 0
+    da = datetime.datetime.now()
+    label = 0
+    num_images = 0
+    cos_thresh = np.cos(np.pi * args.threshold / 180.0)
+    for id_item in id_list:
+        wid = id_item[0]
+        pp += 1
+        if pp % 40 == 0:
+            db = datetime.datetime.now()
+            print('processing id', pp, (db - da).total_seconds())
+            da = db
+        x, _, contents = get_embedding(args, imgrec, id_item[1], id_item[2],
+                                       image_size, model)
+        subcenters = W[wid]
+        K_stat = np.zeros((K, ), dtype=np.int)
+        for i in range(x.shape[0]):
+            _x = x[i]
+            sim = np.dot(subcenters, _x)  # len(sim)==K
+            mc = np.argmax(sim)
+            K_stat[mc] += 1
+        dominant_index = np.argmax(K_stat)
+        dominant_center = subcenters[dominant_index]
+        sim = np.dot(x, dominant_center)
+        idx = np.where(sim > cos_thresh)[0]
+        num_drop = x.shape[0] - len(idx)
+        if len(idx) == 0:
+            continue
+        #print("labelid %d dropped %d, from %d to %d"% (wid, num_drop, x.shape[0], len(idx)))
+        num_images += len(idx)
+        for _idx in idx:
+            c = contents[_idx]
+            builder.add(label, c, is_image=False)
+        label += 1
+    builder.close()
+
+    print('total:', num_images)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='')
+    # general
+    parser.add_argument('--data',
+                        default='/bigdata/faces_ms1m_full',
+                        type=str,
+                        help='')
+    parser.add_argument('--output',
+                        default='/bigdata/ms1m_full_k3drop075',
+                        type=str,
+                        help='')
+    parser.add_argument(
+        '--model',
+        default=
+        '../Evaluation/IJB/pretrained_models/r50-arcfacesc-msf-k3z/model,2',
+        help='path to load model.')
+    parser.add_argument('--batch-size', default=16, type=int, help='')
+    parser.add_argument('--threshold', default=75, type=float, help='')
+    parser.add_argument('--k', default=3, type=int, help='')
+    args = parser.parse_args()
+    main(args)
diff --git a/recognition/subcenter_arcface/image_iter.py b/recognition/subcenter_arcface/image_iter.py
new file mode 100644
index 000000000..2a0797786
--- /dev/null
+++ b/recognition/subcenter_arcface/image_iter.py
@@ -0,0 +1,312 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import random
+import logging
+import sys
+import numbers
+import math
+import sklearn
+import datetime
+import numpy as np
+import cv2
+from config import config
+
+import mxnet as mx
+from mxnet import ndarray as nd
+from mxnet import io
+from mxnet import recordio
+
+logger = logging.getLogger()
+
+
+class FaceImageIter(io.DataIter):
+    def __init__(self,
+                 batch_size,
+                 data_shape,
+                 path_imgrec=None,
+                 shuffle=False,
+                 aug_list=None,
+                 mean=None,
+                 rand_mirror=False,
+                 cutoff=0,
+                 color_jittering=0,
+                 images_filter=0,
+                 data_name='data',
+                 label_name='softmax_label',
+                 **kwargs):
+        super(FaceImageIter, self).__init__()
+        assert path_imgrec
+        if path_imgrec:
+            logging.info('loading recordio %s...', path_imgrec)
+            path_imgidx = path_imgrec[0:-4] + ".idx"
+            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec,
+                                                     'r')  # pylint: disable=redefined-variable-type
+            s = self.imgrec.read_idx(0)
+            header, _ = recordio.unpack(s)
+            if header.flag > 0:
+                print('header0 label', header.label)
+                self.header0 = (int(header.label[0]), int(header.label[1]))
+                #assert(header.flag==1)
+                self.imgidx = list(range(1, int(header.label[0])))
+                #self.imgidx = []
+                #self.id2range = {}
+                #self.seq_identity = range(int(header.label[0]), int(header.label[1]))
+                #for identity in self.seq_identity:
+                #  s = self.imgrec.read_idx(identity)
+                #  header, _ = recordio.unpack(s)
+                #  a,b = int(header.label[0]), int(header.label[1])
+                #  count = b-a
+                #  if count<images_filter:
+                #    continue
+                #  self.id2range[identity] = (a,b)
+                #  self.imgidx += range(a, b)
+                #print('id2range', len(self.id2range))
+            else:
+                self.imgidx = list(self.imgrec.keys)
+            if shuffle:
+                self.seq = self.imgidx
+                self.oseq = self.imgidx
+                print(len(self.seq))
+            else:
+                self.seq = None
+
+        self.mean = mean
+        self.nd_mean = None
+        if self.mean:
+            self.mean = np.array(self.mean, dtype=np.float32).reshape(1, 1, 3)
+            self.nd_mean = mx.nd.array(self.mean).reshape((1, 1, 3))
+
+        self.check_data_shape(data_shape)
+        self.provide_data = [(data_name, (batch_size, ) + data_shape)]
+        self.batch_size = batch_size
+        self.data_shape = data_shape
+        self.shuffle = shuffle
+        self.image_size = '%d,%d' % (data_shape[1], data_shape[2])
+        self.rand_mirror = rand_mirror
+        print('rand_mirror', rand_mirror)
+        self.cutoff = cutoff
+        self.color_jittering = color_jittering
+        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
+        self.provide_label = [(label_name, (batch_size, ))]
+        c, h, w = self.data_shape
+        self.batch_data = np.zeros((batch_size, c, h, w), dtype=np.float32)
+        self.batch_label = np.zeros(self.provide_label[0][1], dtype=np.float32)
+        #print(self.provide_label[0][1])
+        self.cur = 0
+        self.nbatch = 0
+        self.is_init = False
+
+    def reset(self):
+        """Resets the iterator to the beginning of the data."""
+        print('call reset()')
+        self.cur = 0
+        if self.shuffle:
+            random.shuffle(self.seq)
+        if self.seq is None and self.imgrec is not None:
+            self.imgrec.reset()
+
+    def num_samples(self):
+        return len(self.seq)
+
+    def next_sample(self):
+        """Helper function for reading in next sample."""
+        #set total batch size, for example, 1800, and maximum size for each people, for example 45
+        if self.seq is not None:
+            while True:
+                if self.cur >= len(self.seq):
+                    raise StopIteration
+                idx = self.seq[self.cur]
+                self.cur += 1
+                if self.imgrec is not None:
+                    s = self.imgrec.read_idx(idx)
+                    header, img = recordio.unpack(s)
+                    label = header.label
+                    if not isinstance(label, numbers.Number):
+                        label = label[0]
+                    return label, img, None, None
+                else:
+                    label, fname, bbox, landmark = self.imglist[idx]
+                    return label, self.read_image(fname), bbox, landmark
+        else:
+            s = self.imgrec.read()
+            if s is None:
+                raise StopIteration
+            header, img = recordio.unpack(s)
+            return header.label, img, None, None
+
+    def brightness_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        src *= alpha
+        return src
+
+    def contrast_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = nd.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
+        src *= alpha
+        src += gray
+        return src
+
+    def saturation_aug(self, src, x):
+        alpha = 1.0 + random.uniform(-x, x)
+        coef = nd.array([[[0.299, 0.587, 0.114]]])
+        gray = src * coef
+        gray = nd.sum(gray, axis=2, keepdims=True)
+        gray *= (1.0 - alpha)
+        src *= alpha
+        src += gray
+        return src
+
+    def color_aug(self, img, x):
+        #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
+        #random.shuffle(augs)
+        #for aug in augs:
+        #  #print(img.shape)
+        #  img = aug(img, x)
+        #  #print(img.shape)
+        #return img
+        return self.CJA(img)
+
+    def mirror_aug(self, img):
+        _rd = random.randint(0, 1)
+        if _rd == 1:
+            for c in range(img.shape[2]):
+                img[:, :, c] = np.fliplr(img[:, :, c])
+        return img
+
+    def compress_aug(self, img):
+        from PIL import Image
+        from io import BytesIO
+        buf = BytesIO()
+        img = Image.fromarray(img.asnumpy(), 'RGB')
+        q = random.randint(2, 20)
+        img.save(buf, format='JPEG', quality=q)
+        buf = buf.getvalue()
+        img = Image.open(BytesIO(buf))
+        return nd.array(np.asarray(img, 'float32'))
+
+    def next(self):
+        if not self.is_init:
+            self.reset()
+            self.is_init = True
+        """Returns the next batch of data."""
+        #print('in next', self.cur, self.labelcur)
+        self.nbatch += 1
+        batch_size = self.batch_size
+        i = 0
+        try:
+            while i < batch_size:
+                label, s, bbox, landmark = self.next_sample()
+                try:
+                    _data = self.imdecode(s)
+                except Exception as e:
+                    logging.debug('Invalid decoding, skipping:  %s', str(e))
+                    continue
+                if _data.shape[0] != self.data_shape[1]:
+                    _data = mx.image.resize_short(_data, self.data_shape[1])
+                _data = _data.asnumpy().astype(np.float32)
+                if self.rand_mirror and np.random.rand() < 0.5:
+                    _data = _data[:, ::-1, :]
+                if self.color_jittering > 0:
+                    if self.color_jittering > 1:
+                        _rd = random.randint(0, 1)
+                        if _rd == 1:
+                            _data = self.compress_aug(_data)
+                    _data = self.color_aug(_data, 0.125)
+                #if self.nd_mean is not None:
+                #  _data -= self.nd_mean
+                #  _data *= 0.0078125
+                if self.cutoff > 0:
+                    _rd = random.randint(0, 1)
+                    if _rd == 1:
+                        #print('do cutoff aug', self.cutoff)
+                        centerh = random.randint(0, _data.shape[0] - 1)
+                        centerw = random.randint(0, _data.shape[1] - 1)
+                        half = self.cutoff // 2
+                        starth = max(0, centerh - half)
+                        endh = min(_data.shape[0], centerh + half)
+                        startw = max(0, centerw - half)
+                        endw = min(_data.shape[1], centerw + half)
+                        #print(starth, endh, startw, endw, _data.shape)
+                        _data[starth:endh, startw:endw, :] = 128
+                #_data -= 127.5
+                #_data /= 128.0
+                #self.batch_data[i] = self.postprocess_data(_data)
+                _data = _data.transpose((2, 0, 1))
+                self.batch_data[i] = _data
+                self.batch_label[i] = label
+                i += 1
+        except StopIteration:
+            if i < batch_size:
+                raise StopIteration
+
+        batch_data = nd.array(self.batch_data)
+        batch_label = nd.array(self.batch_label)
+        return io.DataBatch([batch_data], [batch_label], batch_size - i)
+
+    def check_data_shape(self, data_shape):
+        """Checks if the input data shape is valid"""
+        if not len(data_shape) == 3:
+            raise ValueError(
+                'data_shape should have length 3, with dimensions CxHxW')
+        if not data_shape[0] == 3:
+            raise ValueError(
+                'This iterator expects inputs to have 3 channels.')
+
+    def check_valid_image(self, data):
+        """Checks if the input data is valid"""
+        if len(data[0].shape) == 0:
+            raise RuntimeError('Data shape is wrong')
+
+    def imdecode(self, s):
+        """Decodes a string or byte string to an NDArray.
+        See mx.img.imdecode for more details."""
+        img = mx.image.imdecode(s)  #mx.ndarray
+        return img
+
+    def read_image(self, fname):
+        """Reads an input image `fname` and returns the decoded raw bytes.
+
+        Example usage:
+        ----------
+        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
+        """
+        with open(os.path.join(self.path_root, fname), 'rb') as fin:
+            img = fin.read()
+        return img
+
+    def augmentation_transform(self, data):
+        """Transforms input data with specified augmentation."""
+        for aug in self.auglist:
+            data = [ret for src in data for ret in aug(src)]
+        return data
+
+    def postprocess_data(self, datum):
+        """Final postprocessing step before image is loaded into the batch."""
+        return nd.transpose(datum, axes=(2, 0, 1))
+
+
+class FaceImageIterList(io.DataIter):
+    def __init__(self, iter_list):
+        assert len(iter_list) > 0
+        self.provide_data = iter_list[0].provide_data
+        self.provide_label = iter_list[0].provide_label
+        self.iter_list = iter_list
+        self.cur_iter = None
+
+    def reset(self):
+        self.cur_iter.reset()
+
+    def next(self):
+        self.cur_iter = random.choice(self.iter_list)
+        while True:
+            try:
+                ret = self.cur_iter.next()
+            except StopIteration:
+                self.cur_iter.reset()
+                continue
+            return ret
diff --git a/recognition/subcenter_arcface/parall_module_local_v1.py b/recognition/subcenter_arcface/parall_module_local_v1.py
new file mode 100644
index 000000000..5f68610ad
--- /dev/null
+++ b/recognition/subcenter_arcface/parall_module_local_v1.py
@@ -0,0 +1,655 @@
+'''
+@author: insightface
+'''
+
+import logging
+import copy
+import time
+import os
+
+import mxnet as mx
+import numpy as np
+from mxnet import context as ctx
+from mxnet.initializer import Uniform
+from mxnet.module.base_module import BaseModule
+from mxnet.module.module import Module
+from mxnet import metric
+from mxnet.model import BatchEndParam
+from mxnet import io
+import mxnet.ndarray as nd
+from config import config
+
+
+class ParallModule(BaseModule):
+    def __init__(self,
+                 symbol,
+                 data_names,
+                 label_names,
+                 logger=logging,
+                 context=ctx.cpu(),
+                 work_load_list=None,
+                 asymbol=None,
+                 args=None):
+        super(ParallModule, self).__init__(logger=logger)
+        self._symbol = symbol
+        self._asymbol = asymbol
+        self._data_names = data_names
+        self._label_names = label_names
+        self._context = context
+        self._work_load_list = work_load_list
+        self._num_classes = config.num_classes
+        self._batch_size = args.batch_size
+        self._verbose = args.verbose
+        self._emb_size = config.emb_size
+        self._local_class_start = args.local_class_start
+        self._iter = 0
+
+        self._curr_module = None
+
+        self._num_workers = config.num_workers
+        self._num_ctx = len(self._context)
+        self._ctx_num_classes = args.ctx_num_classes
+        self._nd_cache = {}
+        self._ctx_cpu = mx.cpu()
+        self._ctx_single_gpu = self._context[-1]
+        self._fixed_param_names = None
+        self._curr_module = Module(self._symbol,
+                                   self._data_names,
+                                   self._label_names,
+                                   logger=self.logger,
+                                   context=self._context,
+                                   work_load_list=self._work_load_list,
+                                   fixed_param_names=self._fixed_param_names)
+        self._arcface_modules = []
+        self._ctx_class_start = []
+        for i in range(len(self._context)):
+
+            args._ctxid = i
+            _module = Module(self._asymbol(args),
+                             self._data_names,
+                             self._label_names,
+                             logger=self.logger,
+                             context=mx.gpu(i),
+                             work_load_list=self._work_load_list,
+                             fixed_param_names=self._fixed_param_names)
+            self._arcface_modules.append(_module)
+            _c = args.local_class_start + i * args.ctx_num_classes
+            self._ctx_class_start.append(_c)
+        self._usekv = False
+        if self._usekv:
+            self._distkv = mx.kvstore.create('dist_sync')
+            self._kvinit = {}
+
+    def _reset_bind(self):
+        self.binded = False
+        self._curr_module = None
+
+    @property
+    def data_names(self):
+        return self._data_names
+
+    @property
+    def output_names(self):
+        return self._symbol.list_outputs()
+
+    @property
+    def data_shapes(self):
+        assert self.binded
+        return self._curr_module.data_shapes
+
+    @property
+    def label_shapes(self):
+        assert self.binded
+        return self._curr_module.label_shapes
+
+    @property
+    def output_shapes(self):
+        assert self.binded
+        return self._curr_module.output_shapes
+
+    def get_export_params(self):
+        assert self.binded and self.params_initialized
+        _g, _x = self._curr_module.get_params()
+        g = _g.copy()
+        x = _x.copy()
+        return g, x
+
+    def get_params(self):
+        assert self.binded and self.params_initialized
+        _g, _x = self._curr_module.get_params()
+        g = _g.copy()
+        x = _x.copy()
+        for _module in self._arcface_modules:
+            _g, _x = _module.get_params()
+            ag = _g.copy()
+            ax = _x.copy()
+            g.update(ag)
+            x.update(ax)
+        return g, x
+
+    def set_params(self,
+                   arg_params,
+                   aux_params,
+                   allow_missing=False,
+                   force_init=True,
+                   allow_extra=False):
+        g = arg_params
+        x = aux_params
+        #ag = {}
+        #ax = {}
+        rk = []
+        for k in g:
+            v = g[k]
+            if k.startswith('fc7'):
+                p1 = k.find('_')
+                p2 = k.rfind('_')
+                _ctxid = int(k[p1 + 1:p2])
+                self._arcface_modules[_ctxid].set_params({k: v}, {})
+                rk.append(k)
+        for k in rk:
+            del g[k]
+        self._curr_module.set_params(g, x)
+        #self._arcface_module.set_params(ag, ax)
+
+    def init_params(self,
+                    initializer=Uniform(0.01),
+                    arg_params=None,
+                    aux_params=None,
+                    allow_missing=False,
+                    force_init=False,
+                    allow_extra=False):
+        if self.params_initialized and not force_init:
+            return
+        assert self.binded, 'call bind before initializing the parameters'
+        #TODO init the same weights with all work nodes
+        self._curr_module.init_params(initializer=initializer,
+                                      arg_params=None,
+                                      aux_params=None,
+                                      allow_missing=allow_missing,
+                                      force_init=force_init,
+                                      allow_extra=allow_extra)
+        for _module in self._arcface_modules:
+            #_initializer = initializer
+            _initializer = mx.init.Normal(0.01)
+            _module.init_params(initializer=_initializer,
+                                arg_params=None,
+                                aux_params=None,
+                                allow_missing=allow_missing,
+                                force_init=force_init,
+                                allow_extra=allow_extra)
+        self.params_initialized = True
+
+    def bind(self,
+             data_shapes,
+             label_shapes=None,
+             for_training=True,
+             inputs_need_grad=False,
+             force_rebind=False,
+             shared_module=None):
+        print('in_bind', self.params_initialized, data_shapes, label_shapes)
+        if self.params_initialized:
+            arg_params, aux_params = self.get_params()
+
+        # force rebinding is typically used when one want to switch from
+        # training to prediction phase.
+        if force_rebind:
+            self._reset_bind()
+
+        if self.binded:
+            self.logger.warning('Already binded, ignoring bind()')
+            return
+
+        assert shared_module is None, 'shared_module for MutableModule is not supported'
+        self.for_training = for_training
+        self.inputs_need_grad = inputs_need_grad
+        self.binded = True
+        self._curr_module.bind(data_shapes,
+                               label_shapes,
+                               for_training,
+                               inputs_need_grad,
+                               force_rebind=False,
+                               shared_module=None)
+        _data_shape = data_shapes[0][1]
+        print('_data_shape', _data_shape, label_shapes)
+        for _module in self._arcface_modules:
+            _module.bind(
+                [('data',
+                  (_data_shape[0] * self._num_workers, self._emb_size))],
+                [('softmax_label', (_data_shape[0] * self._num_workers, ))],
+                for_training,
+                True,
+                force_rebind=False,
+                shared_module=None)
+        if self.params_initialized:
+            self.set_params(arg_params, aux_params)
+
+    def init_optimizer(self,
+                       kvstore='local',
+                       optimizer='sgd',
+                       optimizer_params=(('learning_rate', 0.01), ),
+                       force_init=False):
+        assert self.binded and self.params_initialized
+        if self.optimizer_initialized and not force_init:
+            self.logger.warning('optimizer already initialized, ignoring.')
+            return
+
+        self._curr_module.init_optimizer(kvstore,
+                                         optimizer,
+                                         optimizer_params,
+                                         force_init=force_init)
+        for _module in self._arcface_modules:
+            _module.init_optimizer(kvstore,
+                                   optimizer,
+                                   optimizer_params,
+                                   force_init=force_init)
+        self.optimizer_initialized = True
+
+    def kv_push(self, key, value):
+        #if value.context!=mx.cpu():
+        #  value = value.as_in_context(mx.cpu())
+        if not key in self._kvinit:
+            self._distkv.init(key, nd.zeros_like(value))
+            self._kvinit[key] = 1
+        self._distkv.push(key, value)
+
+    #get fc1 and partial fc7
+    def forward(self, data_batch, is_train=None):
+        #g,x = self.get_params()
+        #print('{fc7_weight[0][0]}', self._iter, g['fc7_0_weight'].asnumpy()[0][0])
+        #print('{pre_fc1_weight[0][0]}', self._iter, g['pre_fc1_weight'].asnumpy()[0][0])
+
+        assert self.binded and self.params_initialized
+        self._curr_module.forward(data_batch, is_train=is_train)
+        if is_train:
+            self._iter += 1
+            fc1, label = self._curr_module.get_outputs(
+                merge_multi_context=True)
+            global_fc1 = fc1
+            self.global_label = label.as_in_context(self._ctx_cpu)
+
+            for i, _module in enumerate(self._arcface_modules):
+                _label = self.global_label - self._ctx_class_start[i]
+                db_global_fc1 = io.DataBatch([global_fc1], [_label])
+                _module.forward(db_global_fc1)  #fc7 with margin
+        #print('forward end')
+
+    def get_ndarray(self, context, name, shape):
+        key = "%s_%s" % (name, context)
+        #print(key)
+        if not key in self._nd_cache:
+            v = nd.zeros(shape=shape, ctx=context)
+            self._nd_cache[key] = v
+        else:
+            v = self._nd_cache[key]
+        return v
+
+    def get_ndarray2(self, context, name, arr):
+        key = "%s_%s" % (name, context)
+        #print(key)
+        if not key in self._nd_cache:
+            v = nd.zeros(shape=arr.shape, ctx=context)
+            self._nd_cache[key] = v
+        else:
+            v = self._nd_cache[key]
+        arr.copyto(v)
+        return v
+
+    def backward(self, out_grads=None):
+        #print('in backward')
+        assert self.binded and self.params_initialized
+        #tmp_ctx = self._ctx_cpu
+        tmp_ctx = self._ctx_single_gpu
+        fc7_outs = []
+        ctx_fc7_max = self.get_ndarray(tmp_ctx, 'ctx_fc7_max',
+                                       (self._batch_size, len(self._context)))
+        #local_fc7_max = nd.zeros( (self.global_label.shape[0],1), ctx=mx.cpu())
+        arcface_module_outputs = []
+        for i, _module in enumerate(self._arcface_modules):
+            #_fc7 = _module.get_outputs(merge_multi_context=True)[0]
+            out = _module.get_outputs(merge_multi_context=True)
+            #print(out[0].shape)
+            #print(out[1].shape)
+            arcface_module_outputs.append(out)
+            _fc7 = out[0]
+            fc7_outs.append(_fc7)
+            _fc7_max = nd.max(_fc7, axis=1).as_in_context(tmp_ctx)
+            ctx_fc7_max[:, i] = _fc7_max
+
+        local_fc7_max = self.get_ndarray(tmp_ctx, 'local_fc7_max',
+                                         (self._batch_size, 1))
+        nd.max(ctx_fc7_max, axis=1, keepdims=True, out=local_fc7_max)
+        global_fc7_max = local_fc7_max
+        #local_fc7_sum = None
+        local_fc7_sum = self.get_ndarray(tmp_ctx, 'local_fc7_sum',
+                                         (self._batch_size, 1))
+        local_fc7_sum[:, :] = 0.0
+        for i, _module in enumerate(self._arcface_modules):
+            _max = self.get_ndarray2(fc7_outs[i].context, 'fc7_max',
+                                     global_fc7_max)
+            fc7_outs[i] = nd.broadcast_sub(fc7_outs[i], _max)
+            fc7_outs[i] = nd.exp(fc7_outs[i])
+            _sum = nd.sum(fc7_outs[i], axis=1,
+                          keepdims=True).as_in_context(tmp_ctx)
+            local_fc7_sum += _sum
+        global_fc7_sum = local_fc7_sum
+
+        if self._iter % self._verbose == 0:
+            #_ctx = self._context[-1]
+            _ctx = self._ctx_cpu
+            _probs = []
+            for i, _module in enumerate(self._arcface_modules):
+                _prob = self.get_ndarray2(_ctx, '_fc7_prob_%d' % i,
+                                          fc7_outs[i])
+                _probs.append(_prob)
+            fc7_prob = self.get_ndarray(
+                _ctx, 'test_fc7_prob',
+                (self._batch_size, self._ctx_num_classes * len(self._context)))
+            nd.concat(*_probs, dim=1, out=fc7_prob)
+            fc7_pred = nd.argmax(fc7_prob, axis=1)
+            local_label = self.global_label - self._local_class_start
+            #local_label = self.get_ndarray2(_ctx, 'test_label', local_label)
+            _pred = nd.equal(fc7_pred, local_label)
+            print('{fc7_acc}', self._iter, nd.mean(_pred).asnumpy()[0])
+
+        #local_fc1_grad = []
+        #fc1_grad_ctx = self._ctx_cpu
+        fc1_grad_ctx = self._ctx_single_gpu
+        local_fc1_grad = self.get_ndarray(fc1_grad_ctx, 'local_fc1_grad',
+                                          (self._batch_size, self._emb_size))
+        local_fc1_grad[:, :] = 0.0
+        total_eloss = []
+        celoss_verbose = 1000
+        if self._iter % celoss_verbose == 0:
+            fc7_celoss = self.get_ndarray(tmp_ctx, 'test_fc7_celoss',
+                                          (self._batch_size, ))
+            fc7_celoss[:] = 0.0
+
+        for i, _module in enumerate(self._arcface_modules):
+            _sum = self.get_ndarray2(fc7_outs[i].context, 'fc7_sum',
+                                     global_fc7_sum)
+            fc7_outs[i] = nd.broadcast_div(fc7_outs[i], _sum)
+            a = i * self._ctx_num_classes
+            b = (i + 1) * self._ctx_num_classes
+            _label = self.global_label - self._ctx_class_start[i]
+            _label = self.get_ndarray2(fc7_outs[i].context, 'label', _label)
+            onehot_label = self.get_ndarray(
+                fc7_outs[i].context, 'label_onehot',
+                (self._batch_size, self._ctx_num_classes))
+            nd.one_hot(_label,
+                       depth=self._ctx_num_classes,
+                       on_value=1.0,
+                       off_value=0.0,
+                       out=onehot_label)
+            #print(fc7_outs[i].shape, onehot_label.shape)
+
+            if self._iter % celoss_verbose == 0:
+                _ce_loss = fc7_outs[i] * onehot_label
+                _ce_loss = nd.sum(_ce_loss, axis=1)
+                fc7_celoss += _ce_loss.as_in_context(tmp_ctx)
+            fc7_outs[i] -= onehot_label
+
+            out = arcface_module_outputs[i]
+            out_grads = [fc7_outs[i]]
+            for j in range(1, len(out)):
+                eloss = out[j]
+                #print('eloss%d:'%j, eloss.shape)
+                #print(out_grads[0].shape)
+                #egrad_shape = (out_grads[0].shape[0], eloss.shape[0])
+                egrad_shape = eloss.shape
+                egrad = self.get_ndarray(fc7_outs[i].context, 'egrad%d' % j,
+                                         egrad_shape)
+                #egrad[:][:] = 1.0/egrad_shape[0]
+                egrad[:][:] = 1.0
+                out_grads.append(egrad)
+                if self._iter % self._verbose == 0:
+                    total_eloss.append(np.mean(eloss.asnumpy()))
+
+            _module.backward(out_grads=out_grads)
+            #ctx_fc1_grad = _module.get_input_grads()[0].as_in_context(mx.cpu())
+            ctx_fc1_grad = self.get_ndarray2(fc1_grad_ctx,
+                                             'ctx_fc1_grad_%d' % i,
+                                             _module.get_input_grads()[0])
+            local_fc1_grad += ctx_fc1_grad
+
+        if self._iter % self._verbose == 0 and len(total_eloss) > 0:
+            print('{eloss}', self._iter, np.mean(total_eloss))
+        #if self._iter%self._verbose==0:
+        if self._iter % celoss_verbose == 0:
+            ce_loss = nd.log(fc7_celoss) * -1.0
+            ce_loss = nd.mean(ce_loss)
+            print('CELOSS,%d,%f' % (self._iter, ce_loss.asnumpy()))
+
+        global_fc1_grad = local_fc1_grad
+        self._curr_module.backward(out_grads=[global_fc1_grad])
+
+    def update(self):
+        assert self.binded and self.params_initialized and self.optimizer_initialized
+        self._curr_module.update()
+        for i, _module in enumerate(self._arcface_modules):
+            _module.update()
+        mx.nd.waitall()
+
+    def get_outputs(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized
+        return self._curr_module.get_outputs(
+            merge_multi_context=merge_multi_context)
+        #return self._arcface_module.get_outputs(merge_multi_context=merge_multi_context)
+
+    def get_input_grads(self, merge_multi_context=True):
+        assert self.binded and self.params_initialized and self.inputs_need_grad
+        return self._curr_module.get_input_grads(
+            merge_multi_context=merge_multi_context)
+
+    def update_metric(self, eval_metric, labels):
+        assert self.binded and self.params_initialized
+        #self._curr_module.update_metric(eval_metric, labels)
+        #label = labels[0]
+        #print(label.shape)
+        #self._arcface_module.update_metric(eval_metric, labels)
+
+    def install_monitor(self, mon):
+        """ Install monitor on all executors """
+        assert self.binded
+        self._curr_module.install_monitor(mon)
+
+    def forward_backward(self, data_batch):
+        """A convenient function that calls both ``forward`` and ``backward``."""
+        self.forward(data_batch, is_train=True)  # get fc1 and partial fc7
+        self.backward()
+
+    def fit(self,
+            train_data,
+            eval_data=None,
+            eval_metric='acc',
+            epoch_end_callback=None,
+            batch_end_callback=None,
+            kvstore='local',
+            optimizer='sgd',
+            optimizer_params=(('learning_rate', 0.01), ),
+            eval_end_callback=None,
+            eval_batch_end_callback=None,
+            initializer=Uniform(0.01),
+            arg_params=None,
+            aux_params=None,
+            allow_missing=False,
+            force_rebind=False,
+            force_init=False,
+            begin_epoch=0,
+            num_epoch=None,
+            validation_metric=None,
+            monitor=None,
+            sparse_row_id_fn=None):
+        """Trains the module parameters.
+
+        Checkout `Module Tutorial <http://mxnet.io/tutorials/basic/module.html>`_ to see
+        a end-to-end use-case.
+
+        Parameters
+        ----------
+        train_data : DataIter
+            Train DataIter.
+        eval_data : DataIter
+            If not ``None``, will be used as validation set and the performance
+            after each epoch will be evaluated.
+        eval_metric : str or EvalMetric
+            Defaults to 'accuracy'. The performance measure used to display during training.
+            Other possible predefined metrics are:
+            'ce' (CrossEntropy), 'f1', 'mae', 'mse', 'rmse', 'top_k_accuracy'.
+        epoch_end_callback : function or list of functions
+            Each callback will be called with the current `epoch`, `symbol`, `arg_params`
+            and `aux_params`.
+        batch_end_callback : function or list of function
+            Each callback will be called with a `BatchEndParam`.
+        kvstore : str or KVStore
+            Defaults to 'local'.
+        optimizer : str or Optimizer
+            Defaults to 'sgd'.
+        optimizer_params : dict
+            Defaults to ``(('learning_rate', 0.01),)``. The parameters for
+            the optimizer constructor.
+            The default value is not a dict, just to avoid pylint warning on dangerous
+            default values.
+        eval_end_callback : function or list of function
+            These will be called at the end of each full evaluation, with the metrics over
+            the entire evaluation set.
+        eval_batch_end_callback : function or list of function
+            These will be called at the end of each mini-batch during evaluation.
+        initializer : Initializer
+            The initializer is called to initialize the module parameters when they are
+            not already initialized.
+        arg_params : dict
+            Defaults to ``None``, if not ``None``, should be existing parameters from a trained
+            model or loaded from a checkpoint (previously saved model). In this case,
+            the value here will be used to initialize the module parameters, unless they
+            are already initialized by the user via a call to `init_params` or `fit`.
+            `arg_params` has a higher priority than `initializer`.
+        aux_params : dict
+            Defaults to ``None``. Similar to `arg_params`, except for auxiliary states.
+        allow_missing : bool
+            Defaults to ``False``. Indicates whether to allow missing parameters when `arg_params`
+            and `aux_params` are not ``None``. If this is ``True``, then the missing parameters
+            will be initialized via the `initializer`.
+        force_rebind : bool
+            Defaults to ``False``. Whether to force rebinding the executors if already bound.
+        force_init : bool
+            Defaults to ``False``. Indicates whether to force initialization even if the
+            parameters are already initialized.
+        begin_epoch : int
+            Defaults to 0. Indicates the starting epoch. Usually, if resumed from a
+            checkpoint saved at a previous training phase at epoch N, then this value should be
+            N+1.
+        num_epoch : int
+            Number of epochs for training.
+        sparse_row_id_fn : A callback function
+            The function  takes `data_batch` as an input and returns a dict of
+            str -> NDArray. The resulting dict is used for pulling row_sparse
+            parameters from the kvstore, where the str key is the name of the param,
+            and the value is the row id of the param to pull.
+
+        Examples
+        --------
+        >>> # An example of using fit for training.
+        >>> # Assume training dataIter and validation dataIter are ready
+        >>> # Assume loading a previously checkpointed model
+        >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 3)
+        >>> mod.fit(train_data=train_dataiter, eval_data=val_dataiter, optimizer='sgd',
+        ...     optimizer_params={'learning_rate':0.01, 'momentum': 0.9},
+        ...     arg_params=arg_params, aux_params=aux_params,
+        ...     eval_metric='acc', num_epoch=10, begin_epoch=3)
+        """
+        assert num_epoch is not None, 'please specify number of epochs'
+        assert arg_params is None and aux_params is None
+
+        self.bind(data_shapes=train_data.provide_data,
+                  label_shapes=train_data.provide_label,
+                  for_training=True,
+                  force_rebind=force_rebind)
+        if monitor is not None:
+            self.install_monitor(monitor)
+        self.init_params(initializer=initializer,
+                         arg_params=arg_params,
+                         aux_params=aux_params,
+                         allow_missing=allow_missing,
+                         force_init=force_init)
+        self.init_optimizer(kvstore=kvstore,
+                            optimizer=optimizer,
+                            optimizer_params=optimizer_params)
+
+        if validation_metric is None:
+            validation_metric = eval_metric
+        if not isinstance(eval_metric, metric.EvalMetric):
+            eval_metric = metric.create(eval_metric)
+        epoch_eval_metric = copy.deepcopy(eval_metric)
+
+        ################################################################################
+        # training loop
+        ################################################################################
+        for epoch in range(begin_epoch, num_epoch):
+            tic = time.time()
+            eval_metric.reset()
+            epoch_eval_metric.reset()
+            nbatch = 0
+            data_iter = iter(train_data)
+            end_of_batch = False
+            next_data_batch = next(data_iter)
+            while not end_of_batch:
+                data_batch = next_data_batch
+                if monitor is not None:
+                    monitor.tic()
+                self.forward_backward(data_batch)
+                self.update()
+                assert not isinstance(data_batch, list)
+
+                #if isinstance(data_batch, list):
+                #    #print('XXX')
+                #    self.update_metric(eval_metric,
+                #                       [db.label for db in data_batch],
+                #                       pre_sliced=True)
+                #    self.update_metric(epoch_eval_metric,
+                #                       [db.label for db in data_batch],
+                #                       pre_sliced=True)
+                #else:
+                #    #print('before update metric')
+                #    self.update_metric(eval_metric, data_batch.label)
+                #    self.update_metric(epoch_eval_metric, data_batch.label)
+                #labels = data_batch.label
+                #labels = [self.global_label]
+                #self.update_metric(eval_metric, labels)
+                #self.update_metric(epoch_eval_metric, labels)
+
+                try:
+                    # pre fetch next batch
+                    next_data_batch = next(data_iter)
+                    self.prepare(next_data_batch,
+                                 sparse_row_id_fn=sparse_row_id_fn)
+                except StopIteration:
+                    end_of_batch = True
+
+                if monitor is not None:
+                    monitor.toc_print()
+
+                #if end_of_batch:
+                #    eval_name_vals = epoch_eval_metric.get_name_value()
+
+                if batch_end_callback is not None:
+                    batch_end_params = BatchEndParam(epoch=epoch,
+                                                     nbatch=nbatch,
+                                                     eval_metric=None,
+                                                     locals=locals())
+                    batch_end_callback(batch_end_params)
+                    #for callback in _as_list(batch_end_callback):
+                    #    callback(batch_end_params)
+                nbatch += 1
+
+            # one epoch of training is finished
+            #for name, val in eval_name_vals:
+            #    self.logger.info('Epoch[%d] Train-%s=%f', epoch, name, val)
+            toc = time.time()
+            self.logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc - tic))
+
+            # sync aux params across devices
+            arg_params, aux_params = self.get_params()
+            self.set_params(arg_params, aux_params)
+
+            # end of 1 epoch, reset the data-iter for another epoch
+            train_data.reset()
diff --git a/recognition/subcenter_arcface/sample_config.py b/recognition/subcenter_arcface/sample_config.py
new file mode 100644
index 000000000..084a5ae07
--- /dev/null
+++ b/recognition/subcenter_arcface/sample_config.py
@@ -0,0 +1,224 @@
+import numpy as np
+import os
+from easydict import EasyDict as edict
+
+config = edict()
+
+config.bn_mom = 0.9
+config.workspace = 256
+config.emb_size = 512
+config.ckpt_embedding = True
+config.net_se = 0
+config.net_act = 'prelu'
+config.net_unit = 3
+config.net_input = 1
+config.net_blocks = [1, 4, 6, 2]
+config.net_output = 'E'
+config.net_multiplier = 1.0
+config.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+config.ce_loss = True
+config.fc7_lr_mult = 1.0
+config.fc7_wd_mult = 1.0
+config.fc7_no_bias = False
+config.max_steps = 0
+config.data_rand_mirror = True
+config.data_cutoff = False
+config.data_color = 0
+config.data_images_filter = 0
+config.count_flops = True
+config.memonger = False  #not work now
+
+config.loss_K = 3
+
+config.flip_noise_prob = 0.0
+
+# network settings
+network = edict()
+
+network.r100 = edict()
+network.r100.net_name = 'fresnet'
+network.r100.num_layers = 100
+
+network.r100fc = edict()
+network.r100fc.net_name = 'fresnet'
+network.r100fc.num_layers = 100
+network.r100fc.net_output = 'FC'
+
+network.r50 = edict()
+network.r50.net_name = 'fresnet'
+network.r50.num_layers = 50
+
+network.r50v1 = edict()
+network.r50v1.net_name = 'fresnet'
+network.r50v1.num_layers = 50
+network.r50v1.net_unit = 1
+
+network.d169 = edict()
+network.d169.net_name = 'fdensenet'
+network.d169.num_layers = 169
+network.d169.per_batch_size = 64
+network.d169.densenet_dropout = 0.0
+
+network.d201 = edict()
+network.d201.net_name = 'fdensenet'
+network.d201.num_layers = 201
+network.d201.per_batch_size = 64
+network.d201.densenet_dropout = 0.0
+
+network.y1 = edict()
+network.y1.net_name = 'fmobilefacenet'
+network.y1.emb_size = 128
+network.y1.net_output = 'GDC'
+
+network.y2 = edict()
+network.y2.net_name = 'fmobilefacenet'
+network.y2.emb_size = 256
+network.y2.net_output = 'GDC'
+network.y2.net_blocks = [2, 8, 16, 4]
+
+network.m1 = edict()
+network.m1.net_name = 'fmobilenet'
+network.m1.emb_size = 256
+network.m1.net_output = 'GDC'
+network.m1.net_multiplier = 1.0
+
+network.m05 = edict()
+network.m05.net_name = 'fmobilenet'
+network.m05.emb_size = 256
+network.m05.net_output = 'GDC'
+network.m05.net_multiplier = 0.5
+
+network.mnas = edict()
+network.mnas.net_name = 'fmnasnet'
+network.mnas.emb_size = 256
+network.mnas.net_output = 'GDC'
+network.mnas.net_multiplier = 1.0
+
+network.mnas05 = edict()
+network.mnas05.net_name = 'fmnasnet'
+network.mnas05.emb_size = 256
+network.mnas05.net_output = 'GDC'
+network.mnas05.net_multiplier = 0.5
+
+network.mnas025 = edict()
+network.mnas025.net_name = 'fmnasnet'
+network.mnas025.emb_size = 256
+network.mnas025.net_output = 'GDC'
+network.mnas025.net_multiplier = 0.25
+
+network.vargfacenet = edict()
+network.vargfacenet.net_name = 'vargfacenet'
+network.vargfacenet.net_multiplier = 1.25
+network.vargfacenet.emb_size = 512
+network.vargfacenet.net_output = 'J'
+
+# dataset settings
+dataset = edict()
+
+dataset.emore = edict()
+dataset.emore.dataset = 'emore'
+dataset.emore.dataset_path = '../datasets/faces_emore'
+dataset.emore.num_classes = 85742
+dataset.emore.image_shape = (112, 112, 3)
+dataset.emore.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+
+dataset.retina = edict()
+dataset.retina.dataset = 'retina'
+dataset.retina.dataset_path = '../datasets/ms1m-retinaface-t1'
+dataset.retina.num_classes = 93431
+dataset.retina.image_shape = (112, 112, 3)
+dataset.retina.val_targets = ['lfw', 'cfp_fp', 'agedb_30']
+
+loss = edict()
+loss.softmax = edict()
+loss.softmax.loss_name = 'softmax'
+
+loss.nsoftmax = edict()
+loss.nsoftmax.loss_name = 'margin_softmax'
+loss.nsoftmax.loss_s = 64.0
+loss.nsoftmax.loss_m1 = 1.0
+loss.nsoftmax.loss_m2 = 0.0
+loss.nsoftmax.loss_m3 = 0.0
+
+loss.arcface = edict()
+loss.arcface.loss_name = 'margin_softmax'
+loss.arcface.loss_s = 64.0
+loss.arcface.loss_m1 = 1.0
+loss.arcface.loss_m2 = 0.5
+loss.arcface.loss_m3 = 0.0
+
+loss.cosface = edict()
+loss.cosface.loss_name = 'margin_softmax'
+loss.cosface.loss_s = 64.0
+loss.cosface.loss_m1 = 1.0
+loss.cosface.loss_m2 = 0.0
+loss.cosface.loss_m3 = 0.35
+
+loss.combined = edict()
+loss.combined.loss_name = 'margin_softmax'
+loss.combined.loss_s = 64.0
+loss.combined.loss_m1 = 1.0
+loss.combined.loss_m2 = 0.3
+loss.combined.loss_m3 = 0.2
+
+loss.triplet = edict()
+loss.triplet.loss_name = 'triplet'
+loss.triplet.images_per_identity = 5
+loss.triplet.triplet_alpha = 0.3
+loss.triplet.triplet_bag_size = 7200
+loss.triplet.triplet_max_ap = 0.0
+loss.triplet.per_batch_size = 60
+loss.triplet.lr = 0.05
+
+loss.atriplet = edict()
+loss.atriplet.loss_name = 'atriplet'
+loss.atriplet.images_per_identity = 5
+loss.atriplet.triplet_alpha = 0.35
+loss.atriplet.triplet_bag_size = 7200
+loss.atriplet.triplet_max_ap = 0.0
+loss.atriplet.per_batch_size = 60
+loss.atriplet.lr = 0.05
+
+# default settings
+default = edict()
+
+# default network
+default.network = 'r100'
+default.pretrained = ''
+default.pretrained_epoch = 1
+# default dataset
+default.dataset = 'emore'
+default.loss = 'arcface'
+default.frequent = 20
+default.verbose = 2000
+default.kvstore = 'device'
+
+default.end_epoch = 10000
+default.lr = 0.1
+default.wd = 0.0005
+default.mom = 0.9
+default.per_batch_size = 128
+default.ckpt = 3
+default.lr_steps = '100000,160000,220000'
+default.models_root = './models'
+
+
+def generate_config(_network, _dataset, _loss):
+    for k, v in loss[_loss].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    for k, v in network[_network].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    for k, v in dataset[_dataset].items():
+        config[k] = v
+        if k in default:
+            default[k] = v
+    config.loss = _loss
+    config.network = _network
+    config.dataset = _dataset
+    config.num_workers = 1
+    if 'DMLC_NUM_WORKER' in os.environ:
+        config.num_workers = int(os.environ['DMLC_NUM_WORKER'])
diff --git a/recognition/subcenter_arcface/symbol/fdensenet.py b/recognition/subcenter_arcface/symbol/fdensenet.py
new file mode 100644
index 000000000..b3d49ee87
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/fdensenet.py
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# coding: utf-8
+# pylint: disable= arguments-differ
+"""DenseNet, implemented in Gluon."""
+
+import sys
+import os
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act():
+    if config.net_act == 'prelu':
+        return nn.PReLU()
+    else:
+        return nn.Activation(config.net_act)
+
+
+# Helpers
+def _make_dense_block(num_layers, bn_size, growth_rate, dropout, stage_index):
+    out = nn.HybridSequential(prefix='stage%d_' % stage_index)
+    with out.name_scope():
+        for _ in range(num_layers):
+            out.add(_make_dense_layer(growth_rate, bn_size, dropout))
+    return out
+
+
+def _make_dense_layer(growth_rate, bn_size, dropout):
+    new_features = nn.HybridSequential(prefix='')
+    new_features.add(nn.BatchNorm())
+    #new_features.add(nn.Activation('relu'))
+    new_features.add(Act())
+    new_features.add(
+        nn.Conv2D(bn_size * growth_rate, kernel_size=1, use_bias=False))
+    new_features.add(nn.BatchNorm())
+    #new_features.add(nn.Activation('relu'))
+    new_features.add(Act())
+    new_features.add(
+        nn.Conv2D(growth_rate, kernel_size=3, padding=1, use_bias=False))
+    if dropout:
+        new_features.add(nn.Dropout(dropout))
+
+    out = gluon.contrib.nn.HybridConcurrent(axis=1, prefix='')
+    out.add(gluon.contrib.nn.Identity())
+    out.add(new_features)
+
+    return out
+
+
+def _make_transition(num_output_features):
+    out = nn.HybridSequential(prefix='')
+    out.add(nn.BatchNorm())
+    #out.add(nn.Activation('relu'))
+    out.add(Act())
+    out.add(nn.Conv2D(num_output_features, kernel_size=1, use_bias=False))
+    out.add(nn.AvgPool2D(pool_size=2, strides=2))
+    return out
+
+
+# Net
+class DenseNet(nn.HybridBlock):
+    r"""Densenet-BC model from the
+    `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_ paper.
+
+    Parameters
+    ----------
+    num_init_features : int
+        Number of filters to learn in the first convolution layer.
+    growth_rate : int
+        Number of filters to add each layer (`k` in the paper).
+    block_config : list of int
+        List of integers for numbers of layers in each pooling block.
+    bn_size : int, default 4
+        Multiplicative factor for number of bottle neck layers.
+        (i.e. bn_size * k features in the bottleneck layer)
+    dropout : float, default 0
+        Rate of dropout after each dense layer.
+    classes : int, default 1000
+        Number of classification classes.
+    """
+    def __init__(self,
+                 num_init_features,
+                 growth_rate,
+                 block_config,
+                 bn_size=4,
+                 dropout=0,
+                 classes=1000,
+                 **kwargs):
+
+        super(DenseNet, self).__init__(**kwargs)
+        with self.name_scope():
+            self.features = nn.HybridSequential(prefix='')
+            self.features.add(
+                nn.Conv2D(num_init_features,
+                          kernel_size=3,
+                          strides=1,
+                          padding=1,
+                          use_bias=False))
+            self.features.add(nn.BatchNorm())
+            self.features.add(nn.Activation('relu'))
+            self.features.add(nn.MaxPool2D(pool_size=3, strides=2, padding=1))
+            # Add dense blocks
+            num_features = num_init_features
+            for i, num_layers in enumerate(block_config):
+                self.features.add(
+                    _make_dense_block(num_layers, bn_size, growth_rate,
+                                      dropout, i + 1))
+                num_features = num_features + num_layers * growth_rate
+                if i != len(block_config) - 1:
+                    self.features.add(_make_transition(num_features // 2))
+                    num_features = num_features // 2
+            self.features.add(nn.BatchNorm())
+            self.features.add(nn.Activation('relu'))
+            #self.features.add(nn.AvgPool2D(pool_size=7))
+            #self.features.add(nn.Flatten())
+
+            #self.output = nn.Dense(classes)
+
+    def hybrid_forward(self, F, x):
+        x = self.features(x)
+        #x = self.output(x)
+        return x
+
+
+# Specification
+densenet_spec = {
+    121: (64, 32, [6, 12, 24, 16]),
+    161: (96, 48, [6, 12, 36, 24]),
+    169: (64, 32, [6, 12, 32, 32]),
+    201: (64, 32, [6, 12, 48, 32])
+}
+
+
+# Constructor
+def get_symbol():
+    num_layers = config.num_layers
+    num_init_features, growth_rate, block_config = densenet_spec[num_layers]
+    net = DenseNet(num_init_features,
+                   growth_rate,
+                   block_config,
+                   dropout=config.densenet_dropout)
+    data = mx.sym.Variable(name='data')
+    data = data - 127.5
+    data = data * 0.0078125
+    body = net(data)
+    fc1 = symbol_utils.get_fc1(body, config.emb_size, config.net_output)
+    return fc1
diff --git a/recognition/subcenter_arcface/symbol/fmnasnet.py b/recognition/subcenter_arcface/symbol/fmnasnet.py
new file mode 100644
index 000000000..118beb942
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/fmnasnet.py
@@ -0,0 +1,213 @@
+import sys
+import os
+import mxnet as mx
+import mxnet.ndarray as nd
+import mxnet.gluon as gluon
+import mxnet.gluon.nn as nn
+import mxnet.autograd as ag
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act():
+    if config.net_act == 'prelu':
+        return nn.PReLU()
+    else:
+        return nn.Activation(config.net_act)
+
+
+def ConvBlock(channels, kernel_size, strides, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(
+            nn.Conv2D(channels,
+                      kernel_size,
+                      strides=strides,
+                      padding=1,
+                      use_bias=False), nn.BatchNorm(scale=True),
+            Act()
+            #nn.Activation('relu')
+        )
+    return out
+
+
+def Conv1x1(channels, is_linear=False, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(nn.Conv2D(channels, 1, padding=0, use_bias=False),
+                nn.BatchNorm(scale=True))
+        if not is_linear:
+            #out.add(nn.Activation('relu'))
+            out.add(Act())
+    return out
+
+
+def DWise(channels, strides, kernel_size=3, **kwargs):
+    out = nn.HybridSequential(**kwargs)
+    with out.name_scope():
+        out.add(
+            nn.Conv2D(channels,
+                      kernel_size,
+                      strides=strides,
+                      padding=kernel_size // 2,
+                      groups=channels,
+                      use_bias=False), nn.BatchNorm(scale=True),
+            Act()
+            #nn.Activation('relu')
+        )
+    return out
+
+
+class SepCONV(nn.HybridBlock):
+    def __init__(self,
+                 inp,
+                 output,
+                 kernel_size,
+                 depth_multiplier=1,
+                 with_bn=True,
+                 **kwargs):
+        super(SepCONV, self).__init__(**kwargs)
+        with self.name_scope():
+            self.net = nn.HybridSequential()
+            cn = int(inp * depth_multiplier)
+
+            if output is None:
+                self.net.add(
+                    nn.Conv2D(in_channels=inp,
+                              channels=cn,
+                              groups=inp,
+                              kernel_size=kernel_size,
+                              strides=(1, 1),
+                              padding=kernel_size // 2,
+                              use_bias=not with_bn))
+            else:
+                self.net.add(
+                    nn.Conv2D(in_channels=inp,
+                              channels=cn,
+                              groups=inp,
+                              kernel_size=kernel_size,
+                              strides=(1, 1),
+                              padding=kernel_size // 2,
+                              use_bias=False),
+                    nn.BatchNorm(),
+                    Act(),
+                    #nn.Activation('relu'),
+                    nn.Conv2D(in_channels=cn,
+                              channels=output,
+                              kernel_size=(1, 1),
+                              strides=(1, 1),
+                              use_bias=not with_bn))
+
+            self.with_bn = with_bn
+            self.act = Act()
+            #self.act = nn.Activation('relu')
+            if with_bn:
+                self.bn = nn.BatchNorm()
+
+    def hybrid_forward(self, F, x):
+        x = self.net(x)
+        if self.with_bn:
+            x = self.bn(x)
+        if self.act is not None:
+            x = self.act(x)
+        return x
+
+
+class ExpandedConv(nn.HybridBlock):
+    def __init__(self,
+                 inp,
+                 oup,
+                 t,
+                 strides,
+                 kernel=3,
+                 same_shape=True,
+                 **kwargs):
+        super(ExpandedConv, self).__init__(**kwargs)
+
+        self.same_shape = same_shape
+        self.strides = strides
+        with self.name_scope():
+            self.bottleneck = nn.HybridSequential()
+            self.bottleneck.add(
+                Conv1x1(inp * t, prefix="expand_"),
+                DWise(inp * t, self.strides, kernel, prefix="dwise_"),
+                Conv1x1(oup, is_linear=True, prefix="linear_"))
+
+    def hybrid_forward(self, F, x):
+        out = self.bottleneck(x)
+        if self.strides == 1 and self.same_shape:
+            out = F.elemwise_add(out, x)
+        return out
+
+
+def ExpandedConvSequence(t, k, inp, oup, repeats, first_strides, **kwargs):
+    seq = nn.HybridSequential(**kwargs)
+    with seq.name_scope():
+        seq.add(ExpandedConv(inp, oup, t, first_strides, k, same_shape=False))
+        curr_inp = oup
+        for i in range(1, repeats):
+            seq.add(ExpandedConv(curr_inp, oup, t, 1))
+            curr_inp = oup
+    return seq
+
+
+class MNasNet(nn.HybridBlock):
+    def __init__(self, m=1.0, **kwargs):
+        super(MNasNet, self).__init__(**kwargs)
+
+        self.first_oup = int(32 * m)
+        self.second_oup = int(16 * m)
+        #self.second_oup = int(32*m)
+        self.interverted_residual_setting = [
+            # t, c,  n, s, k
+            [3, int(24 * m), 3, 2, 3, "stage2_"],  # -> 56x56
+            [3, int(40 * m), 3, 2, 5, "stage3_"],  # -> 28x28
+            [6, int(80 * m), 3, 2, 5, "stage4_1_"],  # -> 14x14
+            [6, int(96 * m), 2, 1, 3, "stage4_2_"],  # -> 14x14
+            [6, int(192 * m), 4, 2, 5, "stage5_1_"],  # -> 7x7
+            [6, int(320 * m), 1, 1, 3, "stage5_2_"],  # -> 7x7          
+        ]
+        self.last_channels = int(1024 * m)
+
+        with self.name_scope():
+            self.features = nn.HybridSequential()
+            self.features.add(
+                ConvBlock(self.first_oup, 3, 1, prefix="stage1_conv0_"))
+            self.features.add(
+                SepCONV(self.first_oup,
+                        self.second_oup,
+                        3,
+                        prefix="stage1_sepconv0_"))
+            inp = self.second_oup
+            for i, (t, c, n, s, k,
+                    prefix) in enumerate(self.interverted_residual_setting):
+                oup = c
+                self.features.add(
+                    ExpandedConvSequence(t, k, inp, oup, n, s, prefix=prefix))
+                inp = oup
+
+            self.features.add(Conv1x1(self.last_channels, prefix="stage5_3_"))
+            #self.features.add(nn.GlobalAvgPool2D())
+            #self.features.add(nn.Flatten())
+            #self.output = nn.Dense(num_classes)
+    def hybrid_forward(self, F, x):
+        x = self.features(x)
+        #x = self.output(x)
+        return x
+
+    def num_output_channel(self):
+        return self.last_channels
+
+
+def get_symbol():
+    net = MNasNet(config.net_multiplier)
+    data = mx.sym.Variable(name='data')
+    data = data - 127.5
+    data = data * 0.0078125
+    body = net(data)
+    fc1 = symbol_utils.get_fc1(body,
+                               config.emb_size,
+                               config.net_output,
+                               input_channel=net.num_output_channel())
+    return fc1
diff --git a/recognition/subcenter_arcface/symbol/fmobilefacenet.py b/recognition/subcenter_arcface/symbol/fmobilefacenet.py
new file mode 100644
index 000000000..f498264ea
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/fmobilefacenet.py
@@ -0,0 +1,224 @@
+import sys
+import os
+import mxnet as mx
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def Conv(data,
+         num_filter=1,
+         kernel=(1, 1),
+         stride=(1, 1),
+         pad=(0, 0),
+         num_group=1,
+         name=None,
+         suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=config.bn_mom)
+    act = Act(data=bn,
+              act_type=config.net_act,
+              name='%s%s_relu' % (name, suffix))
+    return act
+
+
+def Linear(data,
+           num_filter=1,
+           kernel=(1, 1),
+           stride=(1, 1),
+           pad=(0, 0),
+           num_group=1,
+           name=None,
+           suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=config.bn_mom)
+    return bn
+
+
+def ConvOnly(data,
+             num_filter=1,
+             kernel=(1, 1),
+             stride=(1, 1),
+             pad=(0, 0),
+             num_group=1,
+             name=None,
+             suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    return conv
+
+
+def DResidual(data,
+              num_out=1,
+              kernel=(3, 3),
+              stride=(2, 2),
+              pad=(1, 1),
+              num_group=1,
+              name=None,
+              suffix=''):
+    conv = Conv(data=data,
+                num_filter=num_group,
+                kernel=(1, 1),
+                pad=(0, 0),
+                stride=(1, 1),
+                name='%s%s_conv_sep' % (name, suffix))
+    conv_dw = Conv(data=conv,
+                   num_filter=num_group,
+                   num_group=num_group,
+                   kernel=kernel,
+                   pad=pad,
+                   stride=stride,
+                   name='%s%s_conv_dw' % (name, suffix))
+    proj = Linear(data=conv_dw,
+                  num_filter=num_out,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name='%s%s_conv_proj' % (name, suffix))
+    return proj
+
+
+def Residual(data,
+             num_block=1,
+             num_out=1,
+             kernel=(3, 3),
+             stride=(1, 1),
+             pad=(1, 1),
+             num_group=1,
+             name=None,
+             suffix=''):
+    identity = data
+    for i in range(num_block):
+        shortcut = identity
+        conv = DResidual(data=identity,
+                         num_out=num_out,
+                         kernel=kernel,
+                         stride=stride,
+                         pad=pad,
+                         num_group=num_group,
+                         name='%s%s_block' % (name, suffix),
+                         suffix='%d' % i)
+        identity = conv + shortcut
+    return identity
+
+
+def get_symbol():
+    num_classes = config.emb_size
+    print('in_network', config)
+    fc_type = config.net_output
+    data = mx.symbol.Variable(name="data")
+    data = data - 127.5
+    data = data * 0.0078125
+    blocks = config.net_blocks
+    conv_1 = Conv(data,
+                  num_filter=64,
+                  kernel=(3, 3),
+                  pad=(1, 1),
+                  stride=(2, 2),
+                  name="conv_1")
+    if blocks[0] == 1:
+        conv_2_dw = Conv(conv_1,
+                         num_group=64,
+                         num_filter=64,
+                         kernel=(3, 3),
+                         pad=(1, 1),
+                         stride=(1, 1),
+                         name="conv_2_dw")
+    else:
+        conv_2_dw = Residual(conv_1,
+                             num_block=blocks[0],
+                             num_out=64,
+                             kernel=(3, 3),
+                             stride=(1, 1),
+                             pad=(1, 1),
+                             num_group=64,
+                             name="res_2")
+    conv_23 = DResidual(conv_2_dw,
+                        num_out=64,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=128,
+                        name="dconv_23")
+    conv_3 = Residual(conv_23,
+                      num_block=blocks[1],
+                      num_out=64,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=128,
+                      name="res_3")
+    conv_34 = DResidual(conv_3,
+                        num_out=128,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=256,
+                        name="dconv_34")
+    conv_4 = Residual(conv_34,
+                      num_block=blocks[2],
+                      num_out=128,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=256,
+                      name="res_4")
+    conv_45 = DResidual(conv_4,
+                        num_out=128,
+                        kernel=(3, 3),
+                        stride=(2, 2),
+                        pad=(1, 1),
+                        num_group=512,
+                        name="dconv_45")
+    conv_5 = Residual(conv_45,
+                      num_block=blocks[3],
+                      num_out=128,
+                      kernel=(3, 3),
+                      stride=(1, 1),
+                      pad=(1, 1),
+                      num_group=256,
+                      name="res_5")
+    conv_6_sep = Conv(conv_5,
+                      num_filter=512,
+                      kernel=(1, 1),
+                      pad=(0, 0),
+                      stride=(1, 1),
+                      name="conv_6sep")
+
+    fc1 = symbol_utils.get_fc1(conv_6_sep, num_classes, fc_type)
+    return fc1
diff --git a/recognition/subcenter_arcface/symbol/fmobilenet.py b/recognition/subcenter_arcface/symbol/fmobilenet.py
new file mode 100644
index 000000000..bdbf8a559
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/fmobilenet.py
@@ -0,0 +1,275 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+import sys
+import os
+import mxnet as mx
+import symbol_utils
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def Conv(data,
+         num_filter=1,
+         kernel=(1, 1),
+         stride=(1, 1),
+         pad=(0, 0),
+         num_group=1,
+         name=None,
+         suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=True)
+    act = Act(data=bn,
+              act_type=config.net_act,
+              name='%s%s_relu' % (name, suffix))
+    return act
+
+
+def ConvOnly(data,
+             num_filter=1,
+             kernel=(1, 1),
+             stride=(1, 1),
+             pad=(0, 0),
+             num_group=1,
+             name=None,
+             suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    return conv
+
+
+def get_symbol():
+    num_classes = config.emb_size
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    data = mx.symbol.Variable(name="data")  # 224
+    data = data - 127.5
+    data = data * 0.0078125
+    fc_type = config.net_output
+    bf = int(32 * config.net_multiplier)
+    if config.net_input == 0:
+        conv_1 = Conv(data,
+                      num_filter=bf,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(2, 2),
+                      name="conv_1")  # 224/112
+    else:
+        conv_1 = Conv(data,
+                      num_filter=bf,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_1")  # 224/112
+    conv_2_dw = Conv(conv_1,
+                     num_group=bf,
+                     num_filter=bf,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_2_dw")  # 112/112
+    conv_2 = Conv(conv_2_dw,
+                  num_filter=bf * 2,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_2")  # 112/112
+    conv_3_dw = Conv(conv_2,
+                     num_group=bf * 2,
+                     num_filter=bf * 2,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_3_dw")  # 112/56
+    conv_3 = Conv(conv_3_dw,
+                  num_filter=bf * 4,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_3")  # 56/56
+    conv_4_dw = Conv(conv_3,
+                     num_group=bf * 4,
+                     num_filter=bf * 4,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_4_dw")  # 56/56
+    conv_4 = Conv(conv_4_dw,
+                  num_filter=bf * 4,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_4")  # 56/56
+    conv_5_dw = Conv(conv_4,
+                     num_group=bf * 4,
+                     num_filter=bf * 4,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_5_dw")  # 56/28
+    conv_5 = Conv(conv_5_dw,
+                  num_filter=bf * 8,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_5")  # 28/28
+    conv_6_dw = Conv(conv_5,
+                     num_group=bf * 8,
+                     num_filter=bf * 8,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_6_dw")  # 28/28
+    conv_6 = Conv(conv_6_dw,
+                  num_filter=bf * 8,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_6")  # 28/28
+    conv_7_dw = Conv(conv_6,
+                     num_group=bf * 8,
+                     num_filter=bf * 8,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(2, 2),
+                     name="conv_7_dw")  # 28/14
+    conv_7 = Conv(conv_7_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_7")  # 14/14
+
+    conv_8_dw = Conv(conv_7,
+                     num_group=bf * 16,
+                     num_filter=bf * 16,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_8_dw")  # 14/14
+    conv_8 = Conv(conv_8_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_8")  # 14/14
+    conv_9_dw = Conv(conv_8,
+                     num_group=bf * 16,
+                     num_filter=bf * 16,
+                     kernel=(3, 3),
+                     pad=(1, 1),
+                     stride=(1, 1),
+                     name="conv_9_dw")  # 14/14
+    conv_9 = Conv(conv_9_dw,
+                  num_filter=bf * 16,
+                  kernel=(1, 1),
+                  pad=(0, 0),
+                  stride=(1, 1),
+                  name="conv_9")  # 14/14
+    conv_10_dw = Conv(conv_9,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_10_dw")  # 14/14
+    conv_10 = Conv(conv_10_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_10")  # 14/14
+    conv_11_dw = Conv(conv_10,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_11_dw")  # 14/14
+    conv_11 = Conv(conv_11_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_11")  # 14/14
+    conv_12_dw = Conv(conv_11,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_12_dw")  # 14/14
+    conv_12 = Conv(conv_12_dw,
+                   num_filter=bf * 16,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_12")  # 14/14
+
+    conv_13_dw = Conv(conv_12,
+                      num_group=bf * 16,
+                      num_filter=bf * 16,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(2, 2),
+                      name="conv_13_dw")  # 14/7
+    conv_13 = Conv(conv_13_dw,
+                   num_filter=bf * 32,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_13")  # 7/7
+    conv_14_dw = Conv(conv_13,
+                      num_group=bf * 32,
+                      num_filter=bf * 32,
+                      kernel=(3, 3),
+                      pad=(1, 1),
+                      stride=(1, 1),
+                      name="conv_14_dw")  # 7/7
+    conv_14 = Conv(conv_14_dw,
+                   num_filter=bf * 32,
+                   kernel=(1, 1),
+                   pad=(0, 0),
+                   stride=(1, 1),
+                   name="conv_14")  # 7/7
+    body = conv_14
+    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
+    return fc1
diff --git a/recognition/subcenter_arcface/symbol/fresnet.py b/recognition/subcenter_arcface/symbol/fresnet.py
new file mode 100644
index 000000000..7b17788d1
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/fresnet.py
@@ -0,0 +1,1191 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+'''
+Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
+Original author Wei Wu
+
+Implemented the following paper:
+
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
+'''
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import sys
+import os
+import mxnet as mx
+import numpy as np
+import symbol_utils
+import memonger
+import sklearn
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Conv(**kwargs):
+    #name = kwargs.get('name')
+    #_weight = mx.symbol.Variable(name+'_weight')
+    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
+    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
+    body = mx.sym.Convolution(**kwargs)
+    return body
+
+
+def Act(data, act_type, name):
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit2')
+    if bottle_neck:
+        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv1 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv2 = Conv(data=act2,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
+        conv3 = Conv(data=act3,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=conv3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            conv3 = mx.symbol.broadcast_mul(conv3, body)
+        if dim_match:
+            shortcut = data
+        else:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return conv3 + shortcut
+    else:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv1 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv2 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=conv2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            conv2 = mx.symbol.broadcast_mul(conv2, body)
+        if dim_match:
+            shortcut = data
+        else:
+            shortcut = Conv(data=act1,
+                            num_filter=num_filter,
+                            kernel=(1, 1),
+                            stride=stride,
+                            no_bias=True,
+                            workspace=workspace,
+                            name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return conv2 + shortcut
+
+
+def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck,
+                     **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    #print('in unit3')
+    if bottle_neck:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn4 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn4')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn4,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn4 = mx.symbol.broadcast_mul(bn4, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return bn4 + shortcut
+    else:
+        bn1 = mx.sym.BatchNorm(data=data,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        conv1 = Conv(data=bn1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn2 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn3 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return bn3 + shortcut
+
+
+def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck,
+                       **kwargs):
+    """Return ResNeXt Unit symbol for building ResNeXt
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    assert (bottle_neck)
+    use_se = kwargs.get('version_se', 1)
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    act_type = kwargs.get('version_act', 'prelu')
+    num_group = 32
+    #print('in unit3')
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(1, 1),
+                 stride=(1, 1),
+                 pad=(0, 0),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
+    conv2 = Conv(data=act1,
+                 num_group=num_group,
+                 num_filter=int(num_filter * 0.5),
+                 kernel=(3, 3),
+                 stride=(1, 1),
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+    act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
+    conv3 = Conv(data=act2,
+                 num_filter=num_filter,
+                 kernel=(1, 1),
+                 stride=stride,
+                 pad=(0, 0),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv3')
+    bn4 = mx.sym.BatchNorm(data=conv3,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn4')
+
+    if use_se:
+        #se begin
+        body = mx.sym.Pooling(data=bn4,
+                              global_pool=True,
+                              kernel=(7, 7),
+                              pool_type='avg',
+                              name=name + '_se_pool1')
+        body = Conv(data=body,
+                    num_filter=num_filter // 16,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv1",
+                    workspace=workspace)
+        body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+        body = Conv(data=body,
+                    num_filter=num_filter,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    name=name + "_se_conv2",
+                    workspace=workspace)
+        body = mx.symbol.Activation(data=body,
+                                    act_type='sigmoid',
+                                    name=name + "_se_sigmoid")
+        bn4 = mx.symbol.broadcast_mul(bn4, body)
+        #se end
+
+    if dim_match:
+        shortcut = data
+    else:
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=bn_mom,
+                                    name=name + '_sc')
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return bn4 + shortcut
+
+
+def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck,
+                  **kwargs):
+    uv = kwargs.get('version_unit', 3)
+    version_input = kwargs.get('version_input', 1)
+    if uv == 1:
+        if version_input == 0:
+            return residual_unit_v1(data, num_filter, stride, dim_match, name,
+                                    bottle_neck, **kwargs)
+        else:
+            return residual_unit_v1_L(data, num_filter, stride, dim_match,
+                                      name, bottle_neck, **kwargs)
+    elif uv == 2:
+        return residual_unit_v2(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+    elif uv == 4:
+        return residual_unit_v4(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+    else:
+        return residual_unit_v3(data, num_filter, stride, dim_match, name,
+                                bottle_neck, **kwargs)
+
+
+def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    kwargs = {
+        'version_se': config.net_se,
+        'version_input': config.net_input,
+        'version_output': config.net_output,
+        'version_unit': config.net_unit,
+        'version_act': config.net_act,
+        'bn_mom': bn_mom,
+        'workspace': workspace,
+        'memonger': config.memonger,
+    }
+    """Return ResNet symbol of
+    Parameters
+    ----------
+    units : list
+        Number of units in each stage
+    num_stages : int
+        Number of stage
+    filter_list : list
+        Channel size of each stage
+    num_classes : int
+        Ouput size of symbol
+    dataset : str
+        Dataset type, only cifar10 and imagenet supports
+    workspace : int
+        Workspace used in convolution operator
+    """
+    version_se = kwargs.get('version_se', 1)
+    version_input = kwargs.get('version_input', 1)
+    assert version_input >= 0
+    version_output = kwargs.get('version_output', 'E')
+    fc_type = version_output
+    version_unit = kwargs.get('version_unit', 3)
+    act_type = kwargs.get('version_act', 'prelu')
+    memonger = kwargs.get('memonger', False)
+    print(version_se, version_input, version_output, version_unit, act_type,
+          memonger)
+    num_unit = len(units)
+    assert (num_unit == num_stages)
+    data = mx.sym.Variable(name='data')
+    if config.fp16:
+        data = mx.sym.Cast(data=data, dtype=np.float16)
+    if version_input == 0:
+        #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
+        data = mx.sym.identity(data=data, name='id')
+        data = data - 127.5
+        data = data * 0.0078125
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+        #body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
+    elif version_input == 2:
+        data = mx.sym.BatchNorm(data=data,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn_data')
+        body = Conv(data=data,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+    else:
+        data = mx.sym.identity(data=data, name='id')
+        data = data - 127.5
+        data = data * 0.0078125
+        body = data
+        body = Conv(data=body,
+                    num_filter=filter_list[0],
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=act_type, name='relu0')
+
+    for i in range(num_stages):
+        #if version_input==0:
+        #  body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
+        #                       name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
+        #else:
+        #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
+        #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
+        if i==num_stages-1 and config.fp16:
+            body = mx.sym.Cast(data=body, dtype=np.float32)
+        body = residual_unit(body,
+                             filter_list[i + 1], (2, 2),
+                             False,
+                             name='stage%d_unit%d' % (i + 1, 1),
+                             bottle_neck=bottle_neck,
+                             **kwargs)
+        for j in range(units[i] - 1):
+            body = residual_unit(body,
+                                 filter_list[i + 1], (1, 1),
+                                 True,
+                                 name='stage%d_unit%d' % (i + 1, j + 2),
+                                 bottle_neck=bottle_neck,
+                                 **kwargs)
+
+    if bottle_neck:
+        body = Conv(data=body,
+                    num_filter=512,
+                    kernel=(1, 1),
+                    stride=(1, 1),
+                    pad=(0, 0),
+                    no_bias=True,
+                    name="convd",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnd')
+        body = Act(data=body, act_type=act_type, name='relud')
+
+    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
+    return fc1
+
+
+def get_symbol():
+    """
+    Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
+    Original author Wei Wu
+    """
+    num_classes = config.emb_size
+    num_layers = config.num_layers
+    if num_layers >= 500:
+        filter_list = [64, 256, 512, 1024, 2048]
+        bottle_neck = True
+    else:
+        filter_list = [64, 64, 128, 256, 512]
+        bottle_neck = False
+    num_stages = 4
+    if num_layers == 18:
+        units = [2, 2, 2, 2]
+    elif num_layers == 34:
+        units = [3, 4, 6, 3]
+    elif num_layers == 49:
+        units = [3, 4, 14, 3]
+    elif num_layers == 50:
+        units = [3, 4, 14, 3]
+    elif num_layers == 74:
+        units = [3, 6, 24, 3]
+    elif num_layers == 90:
+        units = [3, 8, 30, 3]
+    elif num_layers == 98:
+        units = [3, 4, 38, 3]
+    elif num_layers == 99:
+        units = [3, 8, 35, 3]
+    elif num_layers == 100:
+        units = [3, 13, 30, 3]
+    elif num_layers == 134:
+        units = [3, 10, 50, 3]
+    elif num_layers == 136:
+        units = [3, 13, 48, 3]
+    elif num_layers == 140:
+        units = [3, 15, 48, 3]
+    elif num_layers == 124:
+        units = [3, 13, 40, 5]
+    elif num_layers == 160:
+        units = [3, 24, 49, 3]
+    elif num_layers == 101:
+        units = [3, 4, 23, 3]
+    elif num_layers == 152:
+        units = [3, 8, 36, 3]
+    elif num_layers == 200:
+        units = [3, 24, 36, 3]
+    elif num_layers == 269:
+        units = [3, 30, 48, 8]
+    else:
+        raise ValueError(
+            "no experiments done on num_layers {}, you can do it yourself".
+            format(num_layers))
+
+    net = resnet(units=units,
+                 num_stages=num_stages,
+                 filter_list=filter_list,
+                 num_classes=num_classes,
+                 bottle_neck=bottle_neck)
+
+    if config.memonger:
+        dshape = (config.per_batch_size, config.image_shape[2],
+                  config.image_shape[0], config.image_shape[1])
+        net_mem_planned = memonger.search_plan(net, data=dshape)
+        old_cost = memonger.get_cost(net, data=dshape)
+        new_cost = memonger.get_cost(net_mem_planned, data=dshape)
+
+        print('Old feature map cost=%d MB' % old_cost)
+        print('New feature map cost=%d MB' % new_cost)
+        net = net_mem_planned
+    return net
diff --git a/recognition/subcenter_arcface/symbol/memonger.py b/recognition/subcenter_arcface/symbol/memonger.py
new file mode 100644
index 000000000..8ad610b57
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/memonger.py
@@ -0,0 +1,175 @@
+import mxnet as mx
+import math
+
+
+def prod(shape):
+    """Get product of the shape.
+    """
+    ret = 1
+    for s in shape:
+        ret *= s
+    return ret
+
+
+def is_param(name):
+    """Quick script to check if name is a parameter.
+    """
+    if name == 'data':
+        return False
+    if name.endswith('weight'):
+        return True
+    if name.endswith('bias'):
+        return True
+    if name.endswith('beta'):
+        return True
+    if name.endswith('gamma'):
+        return True
+    return False
+
+
+def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
+    """Memory allocation planner with a given threshold.
+
+    The user can pass in a network configuration,
+    a threshold that limits memory per block.
+    And input shape configurations.
+
+    Parameters
+    ----------
+    sym : symbol
+        Input configuration of symbols.
+        The user need to pre-mark the attribute "mirror_stage" on the nodes
+        that can be book-kept as stage
+
+        The algorithm will decide whether to disbale mirror on the stage nodes.
+
+    threshold: integer
+        A tuning parameter to tune the approximate size of each stage blocks
+
+    plan_info: dict, optional
+        Used to hold plan information.
+
+    **kwargs:
+        The arguments to infer shape.
+
+    Returns
+    -------
+    alloc_sym: symbol
+        A symbol with force mirror tagged on the nodes for better allocation.
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_sb = None
+    last_local = 0
+    period = 1
+    last_stage = ''
+    stage_decision = ''
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        else:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def get_cost(sym, type_dict=None, **kwargs):
+    """Get the cost of the current symbolic plan by running bind on CPU.
+
+    sym : Symbolic Variable
+
+    """
+    texec = sym.simple_bind(ctx=mx.gpu(),
+                            grad_req='write',
+                            type_dict=type_dict,
+                            **kwargs)
+    return int(texec.debug_str().split('\n')[-3].split()[1])
+
+
+def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    threshold = 0
+    min_threshold = None
+    min_cost = None
+    nbegin = 3
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan(sym,
+                               threshold=threshold,
+                               plan_info=info,
+                               **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = int(math.sqrt(save_size * local_size / 2))
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan(sym,
+                                   threshold=threshold,
+                                   plan_info=info,
+                                   **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
diff --git a/recognition/subcenter_arcface/symbol/memonger_v2.py b/recognition/subcenter_arcface/symbol/memonger_v2.py
new file mode 100644
index 000000000..92963de5e
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/memonger_v2.py
@@ -0,0 +1,300 @@
+import mxnet as mx
+import math
+
+
+def prod(shape):
+    """Get product of the shape.
+    """
+    ret = 1
+    for s in shape:
+        ret *= s
+    return ret
+
+
+def is_param(name):
+    """Quick script to check if name is a parameter.
+    """
+    if name == 'data':
+        return False
+    if name.endswith('weight'):
+        return True
+    if name.endswith('bias'):
+        return True
+    if name.endswith('beta'):
+        return True
+    if name.endswith('gamma'):
+        return True
+    return False
+
+
+def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
+    """Memory allocation planner with a given threshold.
+
+    The user can pass in a network configuration,
+    a threshold that limits memory per block.
+    And input shape configurations.
+
+    Parameters
+    ----------
+    sym : symbol
+        Input configuration of symbols.
+        The user need to pre-mark the attribute "mirror_stage" on the nodes
+        that can be book-kept as stage
+
+        The algorithm will decide whether to disbale mirror on the stage nodes.
+
+    threshold: integer
+        A tuning parameter to tune the approximate size of each stage blocks
+
+    plan_info: dict, optional
+        Used to hold plan information.
+
+    **kwargs:
+        The arguments to infer shape.
+
+    Returns
+    -------
+    alloc_sym: symbol
+        A symbol with force mirror tagged on the nodes for better allocation.
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_sb = None
+    last_local = 0
+    period = 1
+    last_stage = ''
+    stage_decision = ''
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        else:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def get_cost(sym, type_dict=None, **kwargs):
+    """Get the cost of the current symbolic plan by running bind on CPU.
+
+    sym : Symbolic Variable
+
+    """
+    texec = sym.simple_bind(ctx=mx.gpu(),
+                            grad_req='write',
+                            type_dict=type_dict,
+                            **kwargs)
+    return int(texec.debug_str().split('\n')[-3].split()[1])
+
+
+def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    threshold = 0
+    min_threshold = None
+    min_cost = None
+    nbegin = 3
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan(sym,
+                               threshold=threshold,
+                               plan_info=info,
+                               **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = int(math.sqrt(save_size * local_size / 2))
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan(sym,
+                                   threshold=threshold,
+                                   plan_info=info,
+                                   **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
+
+
+def make_mirror_plan_to_layer(sym,
+                              layer_name,
+                              threshold,
+                              plan_info=None,
+                              **kwargs):
+    """
+    sym is the original symbal
+    layer_name is a name to which layer of the network should be set as mirror
+    threshhold is the approximate size of each mirror block
+    """
+    threshold = threshold << 20
+    sym = sym.__copy__()
+    internals = sym.get_internals()
+    _, out_shapes, _ = internals.infer_shape(**kwargs)
+    shape_dict = list(zip(internals.list_outputs(), out_shapes))
+    total_size = 0
+    param_size = 0
+    local_size = 0
+    save_size = 0
+    max_size = 0
+    last_stage = ''
+    stage_decision = ''
+    switch = True
+
+    for idx, item in enumerate(shape_dict):
+        sb = internals[idx]
+        name, shape = item
+        #print(name, switch)
+        if is_param(name):
+            param_size += prod(shape) * 4
+            continue
+        elif switch and not 'bn' in name:
+            total_size += prod(shape) * 4
+            local_size += prod(shape) * 4
+            sb._set_attr(force_mirroring='True')
+            print('set force_mirroring', name, total_size, local_size)
+        if layer_name != '' and layer_name in name:
+            switch = False
+
+        if sb.attr('mirror_stage') is not None:
+            stage = sb.attr('mirror_stage')
+            #print(name, stage)
+            if stage == 'True' or stage != last_stage:
+                if local_size > threshold:
+                    save_size += prod(shape) * 4
+                    max_size = max(max_size, local_size)
+                    local_size = 0
+                    stage_decision = 'False'
+                    sb._set_attr(force_mirroring=stage_decision)
+                else:
+                    stage_decision = 'True'
+                    pass
+                last_stage = stage
+            elif stage == last_stage and stage_decision == 'False':
+                save_size += prod(shape) * 4
+                sb._set_attr(force_mirroring=stage_decision)
+
+    if plan_info is not None:
+        plan_info['max_size'] = max_size
+        plan_info['save_size'] = save_size
+    return sym
+
+
+def search_plan_to_layer(sym,
+                         layer_name=None,
+                         threshold=500,
+                         ntrial=6,
+                         type_dict=None,
+                         **kwargs):
+    """Quickly heurestic search over possible plans to find good memory plan.
+
+    Parameters
+    ----------
+    sym : symbolic
+       Symbolic configurations
+
+    ntrial: integer
+       Additional grid search steps
+    """
+    history = []
+    min_threshold = None
+    min_cost = None
+    nbegin = 10
+
+    for k in range(nbegin):
+        info = {}
+        sym = make_mirror_plan_to_layer(sym,
+                                        layer_name=layer_name,
+                                        threshold=threshold,
+                                        plan_info=info,
+                                        **kwargs)
+        cost = get_cost(sym, type_dict, **kwargs)
+        save_size = info['save_size'] >> 20
+        local_size = info['max_size'] >> 20
+        guess = 300 * (k + 1)
+        if min_cost is None or min_cost > cost:
+            min_cost = cost
+        if min_threshold is None or local_size < min_threshold:
+            min_threshold = local_size
+        print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+        history.append((cost, threshold, sym))
+        threshold = guess
+
+    max_threshold = threshold * math.sqrt(2)
+    step = int((max_threshold - min_threshold) / ntrial)
+    print(min_threshold, max_threshold, step)
+    threshold = min_threshold + step
+    if step > 0:
+        for k in range(ntrial):
+            sym = make_mirror_plan_to_layer(sym,
+                                            layer_name=layer_name,
+                                            threshold=threshold,
+                                            plan_info=info,
+                                            **kwargs)
+            cost = get_cost(sym, type_dict, **kwargs)
+            print("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
+            history.append((cost, threshold, sym))
+            threshold += step
+
+    history.sort(key=lambda x: x[0])
+    cost, threshold, sym = history[0]
+    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
+    return sym
diff --git a/recognition/subcenter_arcface/symbol/symbol_utils.py b/recognition/subcenter_arcface/symbol/symbol_utils.py
new file mode 100644
index 000000000..3eb6f1cf9
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/symbol_utils.py
@@ -0,0 +1,595 @@
+import sys
+import os
+import mxnet as mx
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Conv(**kwargs):
+    #name = kwargs.get('name')
+    #_weight = mx.symbol.Variable(name+'_weight')
+    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
+    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
+    body = mx.sym.Convolution(**kwargs)
+    return body
+
+
+def Act(data, act_type, name):
+    #ignore param act_type, set it in this function
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.sym.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+bn_mom = config.bn_mom
+
+
+def Linear(data,
+           num_filter=1,
+           kernel=(1, 1),
+           stride=(1, 1),
+           pad=(0, 0),
+           num_group=1,
+           name=None,
+           suffix=''):
+    conv = mx.sym.Convolution(data=data,
+                              num_filter=num_filter,
+                              kernel=kernel,
+                              num_group=num_group,
+                              stride=stride,
+                              pad=pad,
+                              no_bias=True,
+                              name='%s%s_conv2d' % (name, suffix))
+    bn = mx.sym.BatchNorm(data=conv,
+                          name='%s%s_batchnorm' % (name, suffix),
+                          fix_gamma=False,
+                          momentum=bn_mom)
+    return bn
+
+
+def get_fc1(last_conv, num_classes, fc_type, input_channel=512):
+    body = last_conv
+    if fc_type == 'Z':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = body
+    elif fc_type == 'E':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'FC':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'SFC':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(3, 3),
+                    stride=(2, 2),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="convf",
+                    num_group=input_channel)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf')
+        body = Act(data=body, act_type=config.net_act, name='reluf')
+        body = Conv(data=body,
+                    num_filter=input_channel,
+                    kernel=(1, 1),
+                    pad=(0, 0),
+                    stride=(1, 1),
+                    name="convf2")
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bnf2')
+        body = Act(data=body, act_type=config.net_act, name='reluf2')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'GAP':
+        bn1 = mx.sym.BatchNorm(data=body,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='bn1')
+        relu1 = Act(data=bn1, act_type=config.net_act, name='relu1')
+        # Although kernel is not used here when global_pool=True, we should put one
+        pool1 = mx.sym.Pooling(data=relu1,
+                               global_pool=True,
+                               kernel=(7, 7),
+                               pool_type='avg',
+                               name='pool1')
+        flat = mx.sym.Flatten(data=pool1)
+        fc1 = mx.sym.FullyConnected(data=flat,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'GNAP':  #mobilefacenet++
+        filters_in = 512  # param in mobilefacenet
+        if num_classes > filters_in:
+            body = mx.sym.Convolution(data=last_conv,
+                                      num_filter=num_classes,
+                                      kernel=(1, 1),
+                                      stride=(1, 1),
+                                      pad=(0, 0),
+                                      no_bias=True,
+                                      name='convx')
+            body = mx.sym.BatchNorm(data=body,
+                                    fix_gamma=False,
+                                    eps=2e-5,
+                                    momentum=0.9,
+                                    name='convx_bn')
+            body = Act(data=body, act_type=config.net_act, name='convx_relu')
+            filters_in = num_classes
+        else:
+            body = last_conv
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=True,
+                                eps=2e-5,
+                                momentum=0.9,
+                                name='bn6f')
+
+        spatial_norm = body * body
+        spatial_norm = mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
+        spatial_sqrt = mx.sym.sqrt(spatial_norm)
+        #spatial_mean=mx.sym.mean(spatial_sqrt, axis=(1,2,3), keepdims=True)
+        spatial_mean = mx.sym.mean(spatial_sqrt)
+        spatial_div_inverse = mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
+
+        spatial_attention_inverse = mx.symbol.tile(spatial_div_inverse,
+                                                   reps=(1, filters_in, 1, 1))
+        body = body * spatial_attention_inverse
+        #body = mx.sym.broadcast_mul(body, spatial_div_inverse)
+
+        fc1 = mx.sym.Pooling(body,
+                             kernel=(7, 7),
+                             global_pool=True,
+                             pool_type='avg')
+        if num_classes < filters_in:
+            fc1 = mx.sym.BatchNorm(data=fc1,
+                                   fix_gamma=True,
+                                   eps=2e-5,
+                                   momentum=0.9,
+                                   name='bn6w')
+            fc1 = mx.sym.FullyConnected(data=fc1,
+                                        num_hidden=num_classes,
+                                        name='pre_fc1')
+        else:
+            fc1 = mx.sym.Flatten(data=fc1)
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=0.9,
+                               name='fc1')
+    elif fc_type == "GDC":  #mobilefacenet_v1
+        conv_6_dw = Linear(last_conv,
+                           num_filter=input_channel,
+                           num_group=input_channel,
+                           kernel=(7, 7),
+                           pad=(0, 0),
+                           stride=(1, 1),
+                           name="conv_6dw7_7")
+        conv_6_f = mx.sym.FullyConnected(data=conv_6_dw,
+                                         num_hidden=num_classes,
+                                         name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=conv_6_f,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'F':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        body = mx.symbol.Dropout(data=body, p=0.4)
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'G':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'H':
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='fc1')
+    elif fc_type == 'I':
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn1')
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    elif fc_type == 'J':
+        fc1 = mx.sym.FullyConnected(data=body,
+                                    num_hidden=num_classes,
+                                    name='pre_fc1')
+        fc1 = mx.sym.BatchNorm(data=fc1,
+                               fix_gamma=True,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name='fc1')
+    return fc1
+
+
+def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    workspace = kwargs.get('workspace', 256)
+    memonger = kwargs.get('memonger', False)
+    #print('in unit3')
+    bn1 = mx.sym.BatchNorm(data=data,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn1')
+    conv1 = Conv(data=bn1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=(1, 1),
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv1')
+    bn2 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn2')
+    act1 = Act(data=bn2, act_type=config.net_act, name=name + '_relu1')
+    conv2 = Conv(data=act1,
+                 num_filter=num_filter,
+                 kernel=(3, 3),
+                 stride=stride,
+                 pad=(1, 1),
+                 no_bias=True,
+                 workspace=workspace,
+                 name=name + '_conv2')
+    bn3 = mx.sym.BatchNorm(data=conv2,
+                           fix_gamma=False,
+                           eps=2e-5,
+                           momentum=bn_mom,
+                           name=name + '_bn3')
+
+    if dim_match:
+        shortcut = data
+    else:
+        conv1sc = Conv(data=data,
+                       num_filter=num_filter,
+                       kernel=(1, 1),
+                       stride=stride,
+                       no_bias=True,
+                       workspace=workspace,
+                       name=name + '_conv1sc')
+        shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                    fix_gamma=False,
+                                    momentum=bn_mom,
+                                    eps=2e-5,
+                                    name=name + '_sc')
+    if memonger:
+        shortcut._set_attr(mirror_stage='True')
+    return bn3 + shortcut
+
+
+def residual_unit_v1l(data, num_filter, stride, dim_match, name, bottle_neck):
+    """Return ResNet Unit symbol for building ResNet
+    Parameters
+    ----------
+    data : str
+        Input data
+    num_filter : int
+        Number of output channels
+    bnf : int
+        Bottle neck channels factor with regard to num_filter
+    stride : tuple
+        Stride used in convolution
+    dim_match : Boolean
+        True means channel number between input and output is the same, otherwise means differ
+    name : str
+        Base name of the operators
+    workspace : int
+        Workspace used in convolution operator
+    """
+    workspace = config.workspace
+    bn_mom = config.bn_mom
+    memonger = False
+    use_se = config.net_se
+    act_type = config.net_act
+    #print('in unit1')
+    if bottle_neck:
+        conv1 = Conv(data=data,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(1, 1),
+                     stride=(1, 1),
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=int(num_filter * 0.25),
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn2')
+        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
+        conv3 = Conv(data=act2,
+                     num_filter=num_filter,
+                     kernel=(1, 1),
+                     stride=stride,
+                     pad=(0, 0),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv3')
+        bn3 = mx.sym.BatchNorm(data=conv3,
+                               fix_gamma=False,
+                               eps=2e-5,
+                               momentum=bn_mom,
+                               name=name + '_bn3')
+
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn3,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn3 = mx.symbol.broadcast_mul(bn3, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        eps=2e-5,
+                                        momentum=bn_mom,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn3 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+    else:
+        conv1 = Conv(data=data,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=(1, 1),
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv1')
+        bn1 = mx.sym.BatchNorm(data=conv1,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn1')
+        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
+        conv2 = Conv(data=act1,
+                     num_filter=num_filter,
+                     kernel=(3, 3),
+                     stride=stride,
+                     pad=(1, 1),
+                     no_bias=True,
+                     workspace=workspace,
+                     name=name + '_conv2')
+        bn2 = mx.sym.BatchNorm(data=conv2,
+                               fix_gamma=False,
+                               momentum=bn_mom,
+                               eps=2e-5,
+                               name=name + '_bn2')
+        if use_se:
+            #se begin
+            body = mx.sym.Pooling(data=bn2,
+                                  global_pool=True,
+                                  kernel=(7, 7),
+                                  pool_type='avg',
+                                  name=name + '_se_pool1')
+            body = Conv(data=body,
+                        num_filter=num_filter // 16,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv1",
+                        workspace=workspace)
+            body = Act(data=body, act_type=act_type, name=name + '_se_relu1')
+            body = Conv(data=body,
+                        num_filter=num_filter,
+                        kernel=(1, 1),
+                        stride=(1, 1),
+                        pad=(0, 0),
+                        name=name + "_se_conv2",
+                        workspace=workspace)
+            body = mx.symbol.Activation(data=body,
+                                        act_type='sigmoid',
+                                        name=name + "_se_sigmoid")
+            bn2 = mx.symbol.broadcast_mul(bn2, body)
+            #se end
+
+        if dim_match:
+            shortcut = data
+        else:
+            conv1sc = Conv(data=data,
+                           num_filter=num_filter,
+                           kernel=(1, 1),
+                           stride=stride,
+                           no_bias=True,
+                           workspace=workspace,
+                           name=name + '_conv1sc')
+            shortcut = mx.sym.BatchNorm(data=conv1sc,
+                                        fix_gamma=False,
+                                        momentum=bn_mom,
+                                        eps=2e-5,
+                                        name=name + '_sc')
+        if memonger:
+            shortcut._set_attr(mirror_stage='True')
+        return Act(data=bn2 + shortcut,
+                   act_type=act_type,
+                   name=name + '_relu3')
+
+
+def get_head(data, version_input, num_filter):
+    bn_mom = config.bn_mom
+    workspace = config.workspace
+    kwargs = {'bn_mom': bn_mom, 'workspace': workspace}
+    data = data - 127.5
+    data = data * 0.0078125
+    #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
+    if version_input == 0:
+        body = Conv(data=data,
+                    num_filter=num_filter,
+                    kernel=(7, 7),
+                    stride=(2, 2),
+                    pad=(3, 3),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=config.net_act, name='relu0')
+        body = mx.sym.Pooling(data=body,
+                              kernel=(3, 3),
+                              stride=(2, 2),
+                              pad=(1, 1),
+                              pool_type='max')
+    else:
+        body = data
+        _num_filter = min(num_filter, 64)
+        body = Conv(data=body,
+                    num_filter=_num_filter,
+                    kernel=(3, 3),
+                    stride=(1, 1),
+                    pad=(1, 1),
+                    no_bias=True,
+                    name="conv0",
+                    workspace=workspace)
+        body = mx.sym.BatchNorm(data=body,
+                                fix_gamma=False,
+                                eps=2e-5,
+                                momentum=bn_mom,
+                                name='bn0')
+        body = Act(data=body, act_type=config.net_act, name='relu0')
+        #body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
+        body = residual_unit_v1l(body,
+                                 _num_filter, (2, 2),
+                                 False,
+                                 name='head',
+                                 bottle_neck=False)
+    return body
diff --git a/recognition/subcenter_arcface/symbol/vargfacenet.py b/recognition/subcenter_arcface/symbol/vargfacenet.py
new file mode 100644
index 000000000..434e8beb8
--- /dev/null
+++ b/recognition/subcenter_arcface/symbol/vargfacenet.py
@@ -0,0 +1,578 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+'''
+Author: Horizon Robotics Inc.
+The company is committed to be the global leader of edge AI platform.
+The model implemented in this scripts runs ~200fps on the Sunrise 2.
+Sunrise 2 is the second generation of an embedded AI chip designed by Horizon Robotics,
+targeting to empower AIoT devices by AI.
+
+Implemented the following paper:
+Mengjia Yan, Mengao Zhao, Zining Xu, Qian Zhang, Guoli Wang, Zhizhong Su. "VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition" (https://arxiv.org/abs/1910.04985)
+
+'''
+
+import os
+import sys
+
+import mxnet as mx
+import symbol_utils
+
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from config import config
+
+
+def Act(data, act_type, name):
+    if act_type == 'prelu':
+        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
+    else:
+        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
+    return body
+
+
+def get_setting_params(**kwargs):
+    # bn_params
+    bn_mom = kwargs.get('bn_mom', 0.9)
+    bn_eps = kwargs.get('bn_eps', 2e-5)
+    fix_gamma = kwargs.get('fix_gamma', False)
+    use_global_stats = kwargs.get('use_global_stats', False)
+    # net_setting param
+    workspace = kwargs.get('workspace', 512)
+    act_type = kwargs.get('act_type', 'prelu')
+    use_se = kwargs.get('use_se', True)
+    se_ratio = kwargs.get('se_ratio', 4)
+    group_base = kwargs.get('group_base', 8)
+
+    setting_params = {}
+    setting_params['bn_mom'] = bn_mom
+    setting_params['bn_eps'] = bn_eps
+    setting_params['fix_gamma'] = fix_gamma
+    setting_params['use_global_stats'] = use_global_stats
+    setting_params['workspace'] = workspace
+    setting_params['act_type'] = act_type
+    setting_params['use_se'] = use_se
+    setting_params['se_ratio'] = se_ratio
+    setting_params['group_base'] = group_base
+
+    return setting_params
+
+
+def se_block(data, num_filter, setting_params, name):
+    se_ratio = setting_params['se_ratio']
+    act_type = setting_params['act_type']
+
+    pool1 = mx.sym.Pooling(data=data,
+                           global_pool=True,
+                           pool_type='avg',
+                           name=name + '_se_pool1')
+    conv1 = mx.sym.Convolution(data=pool1,
+                               num_filter=num_filter // se_ratio,
+                               kernel=(1, 1),
+                               stride=(1, 1),
+                               pad=(0, 0),
+                               name=name + "_se_conv1")
+    act1 = Act(data=conv1, act_type=act_type, name=name + '_se_act1')
+
+    conv2 = mx.sym.Convolution(data=act1,
+                               num_filter=num_filter,
+                               kernel=(1, 1),
+                               stride=(1, 1),
+                               pad=(0, 0),
+                               name=name + "_se_conv2")
+    act2 = mx.symbol.Activation(data=conv2,
+                                act_type='sigmoid',
+                                name=name + "_se_sigmoid")
+    out_data = mx.symbol.broadcast_mul(data, act2)
+    return out_data
+
+
+def separable_conv2d(data,
+                     in_channels,
+                     out_channels,
+                     kernel,
+                     pad,
+                     setting_params,
+                     stride=(1, 1),
+                     factor=1,
+                     bias=False,
+                     bn_dw_out=True,
+                     act_dw_out=True,
+                     bn_pw_out=True,
+                     act_pw_out=True,
+                     dilate=1,
+                     name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    group_base = setting_params['group_base']
+    act_type = setting_params['act_type']
+    assert in_channels % group_base == 0
+
+    # depthwise
+    dw_out = mx.sym.Convolution(data=data,
+                                num_filter=int(in_channels * factor),
+                                kernel=kernel,
+                                pad=pad,
+                                stride=stride,
+                                no_bias=False if bias else True,
+                                num_group=int(in_channels / group_base),
+                                dilate=(dilate, dilate),
+                                workspace=workspace,
+                                name=name + '_conv2d_depthwise')
+    if bn_dw_out:
+        dw_out = mx.sym.BatchNorm(data=dw_out,
+                                  fix_gamma=fix_gamma,
+                                  eps=bn_eps,
+                                  momentum=bn_mom,
+                                  use_global_stats=use_global_stats,
+                                  name=name + '_conv2d_depthwise_bn')
+    if act_dw_out:
+        dw_out = Act(data=dw_out,
+                     act_type=act_type,
+                     name=name + '_conv2d_depthwise_act')
+    # pointwise
+    pw_out = mx.sym.Convolution(data=dw_out,
+                                num_filter=out_channels,
+                                kernel=(1, 1),
+                                stride=(1, 1),
+                                pad=(0, 0),
+                                num_group=1,
+                                no_bias=False if bias else True,
+                                workspace=workspace,
+                                name=name + '_conv2d_pointwise')
+    if bn_pw_out:
+        pw_out = mx.sym.BatchNorm(data=pw_out,
+                                  fix_gamma=fix_gamma,
+                                  eps=bn_eps,
+                                  momentum=bn_mom,
+                                  use_global_stats=use_global_stats,
+                                  name=name + '_conv2d_pointwise_bn')
+    if act_pw_out:
+        pw_out = Act(data=pw_out,
+                     act_type=act_type,
+                     name=name + '_conv2d_pointwise_act')
+    return pw_out
+
+
+def vargnet_block(data,
+                  n_out_ch1,
+                  n_out_ch2,
+                  n_out_ch3,
+                  setting_params,
+                  factor=2,
+                  dim_match=True,
+                  multiplier=1,
+                  kernel=(3, 3),
+                  stride=(1, 1),
+                  dilate=1,
+                  with_dilate=False,
+                  name=None):
+    use_se = setting_params['use_se']
+    act_type = setting_params['act_type']
+
+    out_channels_1 = int(n_out_ch1 * multiplier)
+    out_channels_2 = int(n_out_ch2 * multiplier)
+    out_channels_3 = int(n_out_ch3 * multiplier)
+
+    pad = (((kernel[0] - 1) * dilate + 1) // 2,
+           ((kernel[1] - 1) * dilate + 1) // 2)
+
+    if with_dilate:
+        stride = (1, 1)
+    if dim_match:
+        short_cut = data
+    else:
+        short_cut = separable_conv2d(data=data,
+                                     in_channels=out_channels_1,
+                                     out_channels=out_channels_3,
+                                     kernel=kernel,
+                                     pad=pad,
+                                     setting_params=setting_params,
+                                     stride=stride,
+                                     factor=factor,
+                                     bias=False,
+                                     act_pw_out=False,
+                                     dilate=dilate,
+                                     name=name + '_shortcut')
+    sep1_data = separable_conv2d(data=data,
+                                 in_channels=out_channels_1,
+                                 out_channels=out_channels_2,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=stride,
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 name=name + '_sep1_data')
+    sep2_data = separable_conv2d(data=sep1_data,
+                                 in_channels=out_channels_2,
+                                 out_channels=out_channels_3,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=(1, 1),
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 act_pw_out=False,
+                                 name=name + '_sep2_data')
+
+    if use_se:
+        sep2_data = se_block(data=sep2_data,
+                             num_filter=out_channels_3,
+                             setting_params=setting_params,
+                             name=name)
+
+    out_data = sep2_data + short_cut
+    out_data = Act(data=out_data,
+                   act_type=act_type,
+                   name=name + '_out_data_act')
+    return out_data
+
+
+def vargnet_branch_merge_block(data,
+                               n_out_ch1,
+                               n_out_ch2,
+                               n_out_ch3,
+                               setting_params,
+                               factor=2,
+                               dim_match=False,
+                               multiplier=1,
+                               kernel=(3, 3),
+                               stride=(2, 2),
+                               dilate=1,
+                               with_dilate=False,
+                               name=None):
+    act_type = setting_params['act_type']
+
+    out_channels_1 = int(n_out_ch1 * multiplier)
+    out_channels_2 = int(n_out_ch2 * multiplier)
+    out_channels_3 = int(n_out_ch3 * multiplier)
+
+    pad = (((kernel[0] - 1) * dilate + 1) // 2,
+           ((kernel[1] - 1) * dilate + 1) // 2)
+
+    if with_dilate:
+        stride = (1, 1)
+    if dim_match:
+        short_cut = data
+    else:
+        short_cut = separable_conv2d(data=data,
+                                     in_channels=out_channels_1,
+                                     out_channels=out_channels_3,
+                                     kernel=kernel,
+                                     pad=pad,
+                                     setting_params=setting_params,
+                                     stride=stride,
+                                     factor=factor,
+                                     bias=False,
+                                     act_pw_out=False,
+                                     dilate=dilate,
+                                     name=name + '_shortcut')
+    sep1_data_brach1 = separable_conv2d(data=data,
+                                        in_channels=out_channels_1,
+                                        out_channels=out_channels_2,
+                                        kernel=kernel,
+                                        pad=pad,
+                                        setting_params=setting_params,
+                                        stride=stride,
+                                        factor=factor,
+                                        bias=False,
+                                        dilate=dilate,
+                                        act_pw_out=False,
+                                        name=name + '_sep1_data_branch')
+    sep1_data_brach2 = separable_conv2d(data=data,
+                                        in_channels=out_channels_1,
+                                        out_channels=out_channels_2,
+                                        kernel=kernel,
+                                        pad=pad,
+                                        setting_params=setting_params,
+                                        stride=stride,
+                                        factor=factor,
+                                        bias=False,
+                                        dilate=dilate,
+                                        act_pw_out=False,
+                                        name=name + '_sep2_data_branch')
+    sep1_data = sep1_data_brach1 + sep1_data_brach2
+    sep1_data = Act(data=sep1_data,
+                    act_type=act_type,
+                    name=name + '_sep1_data_act')
+    sep2_data = separable_conv2d(data=sep1_data,
+                                 in_channels=out_channels_2,
+                                 out_channels=out_channels_3,
+                                 kernel=kernel,
+                                 pad=pad,
+                                 setting_params=setting_params,
+                                 stride=(1, 1),
+                                 factor=factor,
+                                 bias=False,
+                                 dilate=dilate,
+                                 act_pw_out=False,
+                                 name=name + '_sep2_data')
+    out_data = sep2_data + short_cut
+    out_data = Act(data=out_data,
+                   act_type=act_type,
+                   name=name + '_out_data_act')
+    return out_data
+
+
+def add_vargnet_conv_block(data,
+                           stage,
+                           units,
+                           in_channels,
+                           out_channels,
+                           setting_params,
+                           kernel=(3, 3),
+                           stride=(2, 2),
+                           multiplier=1,
+                           factor=2,
+                           dilate=1,
+                           with_dilate=False,
+                           name=None):
+    assert stage >= 2, 'stage is {}, stage must be set >=2'.format(stage)
+    data = vargnet_branch_merge_block(data=data,
+                                      n_out_ch1=in_channels,
+                                      n_out_ch2=out_channels,
+                                      n_out_ch3=out_channels,
+                                      setting_params=setting_params,
+                                      factor=factor,
+                                      dim_match=False,
+                                      multiplier=multiplier,
+                                      kernel=kernel,
+                                      stride=stride,
+                                      dilate=dilate,
+                                      with_dilate=with_dilate,
+                                      name=name +
+                                      '_stage_{}_unit_1'.format(stage))
+    for i in range(units - 1):
+        data = vargnet_block(data=data,
+                             n_out_ch1=out_channels,
+                             n_out_ch2=out_channels,
+                             n_out_ch3=out_channels,
+                             setting_params=setting_params,
+                             factor=factor,
+                             dim_match=True,
+                             multiplier=multiplier,
+                             kernel=kernel,
+                             stride=(1, 1),
+                             dilate=dilate,
+                             with_dilate=with_dilate,
+                             name=name +
+                             '_stage_{}_unit_{}'.format(stage, i + 2))
+    return data
+
+
+def add_head_block(data,
+                   num_filter,
+                   setting_params,
+                   multiplier,
+                   head_pooling=False,
+                   kernel=(3, 3),
+                   stride=(2, 2),
+                   pad=(1, 1),
+                   name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    act_type = setting_params['act_type']
+    channels = int(num_filter * multiplier)
+
+    conv1 = mx.sym.Convolution(data=data,
+                               num_filter=channels,
+                               kernel=kernel,
+                               pad=pad,
+                               stride=stride,
+                               no_bias=True,
+                               num_group=1,
+                               workspace=workspace,
+                               name=name + '_conv1')
+    bn1 = mx.sym.BatchNorm(data=conv1,
+                           fix_gamma=fix_gamma,
+                           eps=bn_eps,
+                           momentum=bn_mom,
+                           use_global_stats=use_global_stats,
+                           name=name + '_conv1_bn')
+
+    act1 = Act(data=bn1, act_type=act_type, name=name + '_conv1_act')
+
+    if head_pooling:
+        head_data = mx.symbol.Pooling(data=act1,
+                                      kernel=(3, 3),
+                                      stride=(2, 2),
+                                      pad=(1, 1),
+                                      pool_type='max',
+                                      name=name + '_max_pooling')
+    else:
+        head_data = vargnet_block(data=act1,
+                                  n_out_ch1=num_filter,
+                                  n_out_ch2=num_filter,
+                                  n_out_ch3=num_filter,
+                                  setting_params=setting_params,
+                                  factor=1,
+                                  dim_match=False,
+                                  multiplier=multiplier,
+                                  kernel=kernel,
+                                  stride=(2, 2),
+                                  dilate=1,
+                                  with_dilate=False,
+                                  name=name + '_head_pooling')
+    return head_data
+
+
+def add_emb_block(data,
+                  input_channels,
+                  last_channels,
+                  emb_size,
+                  fc_type,
+                  setting_params,
+                  bias=False,
+                  name=None):
+    bn_mom = setting_params['bn_mom']
+    bn_eps = setting_params['bn_eps']
+    fix_gamma = setting_params['fix_gamma']
+    use_global_stats = setting_params['use_global_stats']
+    workspace = setting_params['workspace']
+    act_type = setting_params['act_type']
+    group_base = setting_params['group_base']
+    # last channels
+    if input_channels != last_channels:
+        data = mx.sym.Convolution(data=data,
+                                  num_filter=last_channels,
+                                  kernel=(1, 1),
+                                  pad=(0, 0),
+                                  stride=(1, 1),
+                                  no_bias=False if bias else True,
+                                  workspace=workspace,
+                                  name=name + '_convx')
+        data = mx.sym.BatchNorm(data=data,
+                                fix_gamma=fix_gamma,
+                                eps=bn_eps,
+                                momentum=bn_mom,
+                                use_global_stats=use_global_stats,
+                                name=name + '_convx_bn')
+        data = Act(data=data, act_type=act_type, name=name + '_convx_act')
+    # depthwise
+    convx_depthwise = mx.sym.Convolution(data=data,
+                                         num_filter=last_channels,
+                                         num_group=int(last_channels /
+                                                       group_base),
+                                         kernel=(7, 7),
+                                         pad=(0, 0),
+                                         stride=(1, 1),
+                                         no_bias=False if bias else True,
+                                         workspace=workspace,
+                                         name=name + '_convx_depthwise')
+    convx_depthwise = mx.sym.BatchNorm(data=convx_depthwise,
+                                       fix_gamma=fix_gamma,
+                                       eps=bn_eps,
+                                       momentum=bn_mom,
+                                       use_global_stats=use_global_stats,
+                                       name=name + '_convx_depthwise_bn')
+    # pointwise
+    convx_pointwise = mx.sym.Convolution(data=convx_depthwise,
+                                         num_filter=last_channels // 2,
+                                         kernel=(1, 1),
+                                         pad=(0, 0),
+                                         stride=(1, 1),
+                                         no_bias=False if bias else True,
+                                         workspace=workspace,
+                                         name=name + '_convx_pointwise')
+    convx_pointwise = mx.sym.BatchNorm(data=convx_pointwise,
+                                       fix_gamma=fix_gamma,
+                                       eps=bn_eps,
+                                       momentum=bn_mom,
+                                       use_global_stats=use_global_stats,
+                                       name=name + '_convx_pointwise_bn')
+    convx_pointwise = Act(data=convx_pointwise,
+                          act_type=act_type,
+                          name=name + '_convx_pointwise_act')
+
+    fc1 = symbol_utils.get_fc1(convx_pointwise, emb_size, fc_type)
+    return fc1
+
+
+def get_symbol():
+    multiplier = config.net_multiplier
+    emb_size = config.emb_size
+    fc_type = config.net_output
+
+    kwargs = {
+        'use_se': config.net_se,
+        'act_type': config.net_act,
+        'bn_mom': config.bn_mom,
+        'workspace': config.workspace,
+    }
+
+    setting_params = get_setting_params(**kwargs)
+
+    factor = 2
+    head_pooling = False
+    num_stage = 3
+    stage_list = [2, 3, 4]
+    units = [3, 7, 4]
+    filter_list = [32, 64, 128, 256]
+    last_channels = 1024
+    dilate_list = [1, 1, 1]
+    with_dilate_list = [False, False, False]
+
+    data = mx.sym.Variable(name='data')
+    data = mx.sym.identity(data=data, name='id')
+    data = data - 127.5
+    data = data * 0.0078125
+
+    body = add_head_block(data=data,
+                          num_filter=filter_list[0],
+                          setting_params=setting_params,
+                          multiplier=multiplier,
+                          head_pooling=head_pooling,
+                          kernel=(3, 3),
+                          stride=(1, 1),
+                          pad=(1, 1),
+                          name="vargface_head")
+
+    for i in range(num_stage):
+        body = add_vargnet_conv_block(data=body,
+                                      stage=stage_list[i],
+                                      units=units[i],
+                                      in_channels=filter_list[i],
+                                      out_channels=filter_list[i + 1],
+                                      setting_params=setting_params,
+                                      kernel=(3, 3),
+                                      stride=(2, 2),
+                                      multiplier=multiplier,
+                                      factor=factor,
+                                      dilate=dilate_list[i],
+                                      with_dilate=with_dilate_list[i],
+                                      name="vargface")
+    emb_feat = add_emb_block(data=body,
+                             input_channels=filter_list[3],
+                             last_channels=last_channels,
+                             emb_size=emb_size,
+                             fc_type=fc_type,
+                             setting_params=setting_params,
+                             bias=False,
+                             name='embed')
+    return emb_feat
+
+
+if __name__ == '__main__':
+    get_symbol()
diff --git a/recognition/subcenter_arcface/train_parall.py b/recognition/subcenter_arcface/train_parall.py
new file mode 100644
index 000000000..72b1ff7d2
--- /dev/null
+++ b/recognition/subcenter_arcface/train_parall.py
@@ -0,0 +1,420 @@
+'''
+@author: insightface
+'''
+
+import os
+import sys
+import math
+import random
+import logging
+import pickle
+import sklearn
+import numpy as np
+from image_iter import FaceImageIter
+import mxnet as mx
+from mxnet import ndarray as nd
+import argparse
+import mxnet.optimizer as optimizer
+from config import config, default, generate_config
+sys.path.append(os.path.join(os.path.dirname(__file__), 'symbol'))
+sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
+import verification
+
+import fresnet
+import fmobilefacenet
+import fmobilenet
+import fmnasnet
+import fdensenet
+import vargfacenet
+
+logger = logging.getLogger()
+logger.setLevel(logging.INFO)
+
+args = None
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Train parall face network')
+    # general
+    parser.add_argument('--dataset',
+                        default=default.dataset,
+                        help='dataset config')
+    parser.add_argument('--network',
+                        default=default.network,
+                        help='network config')
+    parser.add_argument('--loss', default=default.loss, help='loss config')
+    args, rest = parser.parse_known_args()
+    generate_config(args.network, args.dataset, args.loss)
+    parser.add_argument('--models-root',
+                        default=default.models_root,
+                        help='root directory to save model.')
+    parser.add_argument('--pretrained',
+                        default=default.pretrained,
+                        help='pretrained model to load')
+    parser.add_argument('--pretrained-epoch',
+                        type=int,
+                        default=default.pretrained_epoch,
+                        help='pretrained epoch to load')
+    parser.add_argument(
+        '--ckpt',
+        type=int,
+        default=default.ckpt,
+        help=
+        'checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save'
+    )
+    parser.add_argument(
+        '--verbose',
+        type=int,
+        default=default.verbose,
+        help='do verification testing and model saving every verbose batches')
+    parser.add_argument('--lr',
+                        type=float,
+                        default=default.lr,
+                        help='start learning rate')
+    parser.add_argument('--lr-steps',
+                        type=str,
+                        default=default.lr_steps,
+                        help='steps of lr changing')
+    parser.add_argument('--wd',
+                        type=float,
+                        default=default.wd,
+                        help='weight decay')
+    parser.add_argument('--mom',
+                        type=float,
+                        default=default.mom,
+                        help='momentum')
+    parser.add_argument('--frequent',
+                        type=int,
+                        default=default.frequent,
+                        help='')
+    parser.add_argument('--per-batch-size',
+                        type=int,
+                        default=default.per_batch_size,
+                        help='batch size in each context')
+    parser.add_argument('--kvstore',
+                        type=str,
+                        default=default.kvstore,
+                        help='kvstore setting')
+    parser.add_argument('--worker-id',
+                        type=int,
+                        default=0,
+                        help='worker id for dist training, starts from 0')
+    parser.add_argument('--extra-model-name',
+                        type=str,
+                        default='',
+                        help='extra model name')
+    args = parser.parse_args()
+    return args
+
+
+def get_symbol_embedding():
+    embedding = eval(config.net_name).get_symbol()
+    all_label = mx.symbol.Variable('softmax_label')
+    #embedding = mx.symbol.BlockGrad(embedding)
+    all_label = mx.symbol.BlockGrad(all_label)
+    out_list = [embedding, all_label]
+    out = mx.symbol.Group(out_list)
+    return out
+
+
+def get_symbol_arcface(args):
+    embedding = mx.symbol.Variable('data')
+    all_label = mx.symbol.Variable('softmax_label')
+    gt_label = all_label
+    is_softmax = True
+    #print('call get_sym_arcface with', args, config)
+    if config.loss_name == 'margin_softmax':
+        _weight = mx.symbol.Variable("fc7_%d_weight" % args._ctxid,
+                                     shape=(args.ctx_num_classes *
+                                            config.loss_K, config.emb_size),
+                                     lr_mult=config.fc7_lr_mult,
+                                     wd_mult=config.fc7_wd_mult)
+        nweight = mx.symbol.L2Normalization(_weight, mode='instance')
+        nembedding = mx.symbol.L2Normalization(embedding,
+                                               mode='instance',
+                                               name='fc1n_%d' % args._ctxid)
+        fc7 = mx.sym.FullyConnected(data=nembedding,
+                                    weight=nweight,
+                                    no_bias=True,
+                                    num_hidden=args.ctx_num_classes *
+                                    config.loss_K,
+                                    name='fc7_%d' % args._ctxid)
+        if config.loss_K > 1:
+            sim_s3 = mx.symbol.reshape(
+                fc7, (-1, args.ctx_num_classes, config.loss_K))
+
+            sim = mx.symbol.max(sim_s3, axis=2)
+            fc7 = sim
+
+    if config.loss_m1 != 1.0 or config.loss_m2 != 0.0 or config.loss_m3 != 0.0:
+        gt_one_hot = mx.sym.one_hot(gt_label,
+                                    depth=args.ctx_num_classes,
+                                    on_value=1.0,
+                                    off_value=0.0)
+        if config.loss_m1 == 1.0 and config.loss_m2 == 0.0:
+            _one_hot = gt_one_hot * args.margin_b
+            fc7 = fc7 - _one_hot
+        else:
+            fc7_onehot = fc7 * gt_one_hot
+            cos_t = fc7_onehot
+            t = mx.sym.arccos(cos_t)
+            if config.loss_m1 != 1.0:
+                t = t * config.loss_m1
+            if config.loss_m2 != 0.0:
+                t = t + config.loss_m2
+            margin_cos = mx.sym.cos(t)
+            if config.loss_m3 != 0.0:
+                margin_cos = margin_cos - config.loss_m3
+            margin_fc7 = margin_cos
+            margin_fc7_onehot = margin_fc7 * gt_one_hot
+            diff = margin_fc7_onehot - fc7_onehot
+            fc7 = fc7 + diff
+    fc7 = fc7 * config.loss_s
+    out_list = []
+    out_list.append(fc7)
+    if config.loss_name == 'softmax':  #softmax
+        out_list.append(gt_label)
+    out = mx.symbol.Group(out_list)
+    return out
+
+
+def train_net(args):
+    #_seed = 727
+    #random.seed(_seed)
+    #np.random.seed(_seed)
+    #mx.random.seed(_seed)
+    ctx = []
+    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
+    if len(cvd) > 0:
+        for i in range(len(cvd.split(','))):
+            ctx.append(mx.gpu(i))
+    if len(ctx) == 0:
+        ctx = [mx.cpu()]
+        print('use cpu')
+    else:
+        print('gpu num:', len(ctx))
+    if len(args.extra_model_name) == 0:
+        prefix = os.path.join(
+            args.models_root,
+            '%s-%s-%s' % (args.network, args.loss, args.dataset), 'model')
+    else:
+        prefix = os.path.join(
+            args.models_root, '%s-%s-%s-%s' %
+            (args.network, args.loss, args.dataset, args.extra_model_name),
+            'model')
+    prefix_dir = os.path.dirname(prefix)
+    print('prefix', prefix)
+    if not os.path.exists(prefix_dir):
+        os.makedirs(prefix_dir)
+    args.ctx_num = len(ctx)
+    if args.per_batch_size == 0:
+        args.per_batch_size = 128
+    args.batch_size = args.per_batch_size * args.ctx_num
+    args.rescale_threshold = 0
+    args.image_channel = config.image_shape[2]
+    config.batch_size = args.batch_size
+    config.per_batch_size = args.per_batch_size
+    data_dir = config.dataset_path
+    path_imgrec = None
+    path_imglist = None
+    image_size = config.image_shape[0:2]
+    assert len(image_size) == 2
+    assert image_size[0] == image_size[1]
+    print('image_size', image_size)
+    print('num_classes', config.num_classes)
+    path_imgrec = os.path.join(data_dir, "train.rec")
+
+    data_shape = (args.image_channel, image_size[0], image_size[1])
+
+    num_workers = config.num_workers
+    global_num_ctx = num_workers * args.ctx_num
+    if config.num_classes % global_num_ctx == 0:
+        args.ctx_num_classes = config.num_classes // global_num_ctx
+    else:
+        args.ctx_num_classes = config.num_classes // global_num_ctx + 1
+    args.local_num_classes = args.ctx_num_classes * args.ctx_num
+    args.local_class_start = args.local_num_classes * args.worker_id
+
+    #if len(args.partial)==0:
+    #  local_classes_range = (0, args.num_classes)
+    #else:
+    #  _vec = args.partial.split(',')
+    #  local_classes_range = (int(_vec[0]), int(_vec[1]))
+
+    #args.partial_num_classes = local_classes_range[1] - local_classes_range[0]
+    #args.partial_start = local_classes_range[0]
+
+    print('Called with argument:', args, config)
+    mean = None
+
+    begin_epoch = 0
+    base_lr = args.lr
+    base_wd = args.wd
+    base_mom = args.mom
+    arg_params = None
+    aux_params = None
+    if len(args.pretrained) == 0:
+        esym = get_symbol_embedding()
+        asym = get_symbol_arcface
+    else:
+        assert False
+
+    if config.num_workers == 1:
+        from parall_module_local_v1 import ParallModule
+    else:
+        from parall_module_dist import ParallModule
+
+    model = ParallModule(
+        context=ctx,
+        symbol=esym,
+        data_names=['data'],
+        label_names=['softmax_label'],
+        asymbol=asym,
+        args=args,
+    )
+    val_dataiter = None
+    train_dataiter = FaceImageIter(
+        batch_size=args.batch_size,
+        data_shape=data_shape,
+        path_imgrec=path_imgrec,
+        shuffle=True,
+        rand_mirror=config.data_rand_mirror,
+        mean=mean,
+        cutoff=config.data_cutoff,
+        color_jittering=config.data_color,
+        images_filter=config.data_images_filter,
+    )
+
+    if config.net_name == 'fresnet' or config.net_name == 'fmobilefacenet':
+        initializer = mx.init.Xavier(rnd_type='gaussian',
+                                     factor_type="out",
+                                     magnitude=2)  #resnet style
+    else:
+        initializer = mx.init.Xavier(rnd_type='uniform',
+                                     factor_type="in",
+                                     magnitude=2)
+
+    _rescale = 1.0 / args.batch_size
+    opt = optimizer.SGD(learning_rate=base_lr,
+                        momentum=base_mom,
+                        wd=base_wd,
+                        rescale_grad=_rescale)
+    _cb = mx.callback.Speedometer(args.batch_size, args.frequent)
+
+    ver_list = []
+    ver_name_list = []
+    for name in config.val_targets:
+        path = os.path.join(data_dir, name + ".bin")
+        if os.path.exists(path):
+            data_set = verification.load_bin(path, image_size)
+            ver_list.append(data_set)
+            ver_name_list.append(name)
+            print('ver', name)
+
+    def ver_test(nbatch):
+        results = []
+        for i in range(len(ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                ver_list[i], model, args.batch_size, 10, None, None)
+            print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
+            #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
+            print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' %
+                  (ver_name_list[i], nbatch, acc2, std2))
+            results.append(acc2)
+        return results
+
+    highest_acc = [0.0, 0.0]  #lfw and target
+    #for i in range(len(ver_list)):
+    #  highest_acc.append(0.0)
+    global_step = [0]
+    save_step = [0]
+    lr_steps = [int(x) for x in args.lr_steps.split(',')]
+    print('lr_steps', lr_steps)
+
+    def _batch_callback(param):
+        #global global_step
+        global_step[0] += 1
+        mbatch = global_step[0]
+        for step in lr_steps:
+            if mbatch == step:
+                opt.lr *= 0.1
+                print('lr change to', opt.lr)
+                break
+
+        _cb(param)
+        if mbatch % 1000 == 0:
+            print('lr-batch-epoch:', opt.lr, param.nbatch, param.epoch)
+
+        if mbatch >= 0 and mbatch % args.verbose == 0:
+            acc_list = ver_test(mbatch)
+            save_step[0] += 1
+            msave = save_step[0]
+            do_save = False
+            is_highest = False
+            if len(acc_list) > 0:
+                #lfw_score = acc_list[0]
+                #if lfw_score>highest_acc[0]:
+                #  highest_acc[0] = lfw_score
+                #  if lfw_score>=0.998:
+                #    do_save = True
+                score = sum(acc_list)
+                if acc_list[-1] >= highest_acc[-1]:
+                    if acc_list[-1] > highest_acc[-1]:
+                        is_highest = True
+                    else:
+                        if score >= highest_acc[0]:
+                            is_highest = True
+                            highest_acc[0] = score
+                    highest_acc[-1] = acc_list[-1]
+                    #if lfw_score>=0.99:
+                    #  do_save = True
+            if is_highest:
+                do_save = True
+            if args.ckpt == 0:
+                do_save = False
+            elif args.ckpt == 2:
+                do_save = True
+            elif args.ckpt == 3:
+                msave = 1
+
+            if do_save:
+                print('saving', msave)
+                if config.ckpt_embedding:
+                    arg, aux = model.get_export_params()
+                else:
+                    arg, aux = model.get_params()
+                all_layers = model.symbol.get_internals()
+                _sym = all_layers['fc1_output']
+                mx.model.save_checkpoint(prefix, msave, _sym, arg, aux)
+            print('[%d]Accuracy-Highest: %1.5f' % (mbatch, highest_acc[-1]))
+        if config.max_steps > 0 and mbatch > config.max_steps:
+            sys.exit(0)
+
+    epoch_cb = None
+    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
+
+    model.fit(
+        train_dataiter,
+        begin_epoch=begin_epoch,
+        num_epoch=999999,
+        eval_data=val_dataiter,
+        #eval_metric        = eval_metrics,
+        kvstore=args.kvstore,
+        optimizer=opt,
+        #optimizer_params   = optimizer_params,
+        initializer=initializer,
+        arg_params=arg_params,
+        aux_params=aux_params,
+        allow_missing=True,
+        batch_end_callback=_batch_callback,
+        epoch_end_callback=epoch_cb)
+
+
+def main():
+    global args
+    args = parse_args()
+    train_net(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/recognition/symbol/fdensenet.py b/recognition/symbol/fdensenet.py
deleted file mode 100644
index b37330eb3..000000000
--- a/recognition/symbol/fdensenet.py
+++ /dev/null
@@ -1,145 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-# coding: utf-8
-# pylint: disable= arguments-differ
-"""DenseNet, implemented in Gluon."""
-
-import sys
-import os
-import mxnet as mx
-import mxnet.ndarray as nd
-import mxnet.gluon as gluon
-import mxnet.gluon.nn as nn
-import mxnet.autograd as ag
-import symbol_utils
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Act():
-    if config.net_act=='prelu':
-      return nn.PReLU()
-    else:
-      return nn.Activation(config.net_act)
-# Helpers
-def _make_dense_block(num_layers, bn_size, growth_rate, dropout, stage_index):
-    out = nn.HybridSequential(prefix='stage%d_'%stage_index)
-    with out.name_scope():
-        for _ in range(num_layers):
-            out.add(_make_dense_layer(growth_rate, bn_size, dropout))
-    return out
-
-def _make_dense_layer(growth_rate, bn_size, dropout):
-    new_features = nn.HybridSequential(prefix='')
-    new_features.add(nn.BatchNorm())
-    #new_features.add(nn.Activation('relu'))
-    new_features.add(Act())
-    new_features.add(nn.Conv2D(bn_size * growth_rate, kernel_size=1, use_bias=False))
-    new_features.add(nn.BatchNorm())
-    #new_features.add(nn.Activation('relu'))
-    new_features.add(Act())
-    new_features.add(nn.Conv2D(growth_rate, kernel_size=3, padding=1, use_bias=False))
-    if dropout:
-        new_features.add(nn.Dropout(dropout))
-
-    out = gluon.contrib.nn.HybridConcurrent(axis=1, prefix='')
-    out.add(gluon.contrib.nn.Identity())
-    out.add(new_features)
-
-    return out
-
-def _make_transition(num_output_features):
-    out = nn.HybridSequential(prefix='')
-    out.add(nn.BatchNorm())
-    #out.add(nn.Activation('relu'))
-    out.add(Act())
-    out.add(nn.Conv2D(num_output_features, kernel_size=1, use_bias=False))
-    out.add(nn.AvgPool2D(pool_size=2, strides=2))
-    return out
-
-# Net
-class DenseNet(nn.HybridBlock):
-    r"""Densenet-BC model from the
-    `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_ paper.
-
-    Parameters
-    ----------
-    num_init_features : int
-        Number of filters to learn in the first convolution layer.
-    growth_rate : int
-        Number of filters to add each layer (`k` in the paper).
-    block_config : list of int
-        List of integers for numbers of layers in each pooling block.
-    bn_size : int, default 4
-        Multiplicative factor for number of bottle neck layers.
-        (i.e. bn_size * k features in the bottleneck layer)
-    dropout : float, default 0
-        Rate of dropout after each dense layer.
-    classes : int, default 1000
-        Number of classification classes.
-    """
-    def __init__(self, num_init_features, growth_rate, block_config,
-                 bn_size=4, dropout=0, classes=1000, **kwargs):
-
-        super(DenseNet, self).__init__(**kwargs)
-        with self.name_scope():
-            self.features = nn.HybridSequential(prefix='')
-            self.features.add(nn.Conv2D(num_init_features, kernel_size=3,
-                                        strides=1, padding=1, use_bias=False))
-            self.features.add(nn.BatchNorm())
-            self.features.add(nn.Activation('relu'))
-            self.features.add(nn.MaxPool2D(pool_size=3, strides=2, padding=1))
-            # Add dense blocks
-            num_features = num_init_features
-            for i, num_layers in enumerate(block_config):
-                self.features.add(_make_dense_block(num_layers, bn_size, growth_rate, dropout, i+1))
-                num_features = num_features + num_layers * growth_rate
-                if i != len(block_config) - 1:
-                    self.features.add(_make_transition(num_features // 2))
-                    num_features = num_features // 2
-            self.features.add(nn.BatchNorm())
-            self.features.add(nn.Activation('relu'))
-            #self.features.add(nn.AvgPool2D(pool_size=7))
-            #self.features.add(nn.Flatten())
-
-            #self.output = nn.Dense(classes)
-
-    def hybrid_forward(self, F, x):
-        x = self.features(x)
-        #x = self.output(x)
-        return x
-
-
-# Specification
-densenet_spec = {121: (64, 32, [6, 12, 24, 16]),
-                 161: (96, 48, [6, 12, 36, 24]),
-                 169: (64, 32, [6, 12, 32, 32]),
-                 201: (64, 32, [6, 12, 48, 32])}
-
-
-# Constructor
-def get_symbol():
-    num_layers = config.num_layers
-    num_init_features, growth_rate, block_config = densenet_spec[num_layers]
-    net = DenseNet(num_init_features, growth_rate, block_config, dropout = config.densenet_dropout)
-    data = mx.sym.Variable(name='data')
-    data = data-127.5
-    data = data*0.0078125
-    body = net(data)
-    fc1 = symbol_utils.get_fc1(body, config.emb_size, config.net_output)
-    return fc1
-
diff --git a/recognition/symbol/fmnasnet.py b/recognition/symbol/fmnasnet.py
deleted file mode 100644
index 9b0ba7f06..000000000
--- a/recognition/symbol/fmnasnet.py
+++ /dev/null
@@ -1,165 +0,0 @@
-import sys
-import os
-import mxnet as mx
-import mxnet.ndarray as nd
-import mxnet.gluon as gluon
-import mxnet.gluon.nn as nn
-import mxnet.autograd as ag
-import symbol_utils
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Act():
-    if config.net_act=='prelu':
-      return nn.PReLU()
-    else:
-      return nn.Activation(config.net_act)
-
-def ConvBlock(channels, kernel_size, strides, **kwargs):
-    out = nn.HybridSequential(**kwargs)
-    with out.name_scope():
-        out.add(
-            nn.Conv2D(channels, kernel_size, strides=strides, padding=1, use_bias=False),
-            nn.BatchNorm(scale=True),
-            Act()
-            #nn.Activation('relu')
-        )
-    return out
-
-def Conv1x1(channels, is_linear=False, **kwargs):
-    out = nn.HybridSequential(**kwargs)
-    with out.name_scope():
-        out.add(
-            nn.Conv2D(channels, 1, padding=0, use_bias=False),
-            nn.BatchNorm(scale=True)
-        )
-        if not is_linear:
-            #out.add(nn.Activation('relu'))
-            out.add(Act())
-    return out
-
-def DWise(channels, strides, kernel_size=3, **kwargs):
-    out = nn.HybridSequential(**kwargs)
-    with out.name_scope():
-        out.add(
-            nn.Conv2D(channels, kernel_size, strides=strides, padding=kernel_size // 2, groups=channels, use_bias=False),
-            nn.BatchNorm(scale=True),
-            Act()
-            #nn.Activation('relu')
-        )
-    return out
-
-class SepCONV(nn.HybridBlock):
-    def __init__(self, inp, output, kernel_size, depth_multiplier=1, with_bn=True, **kwargs):
-        super(SepCONV, self).__init__(**kwargs)
-        with self.name_scope():
-            self.net = nn.HybridSequential()
-            cn = int(inp*depth_multiplier)
-
-            if output is None:
-                self.net.add(
-                    nn.Conv2D(in_channels=inp, channels=cn, groups=inp, kernel_size=kernel_size, strides=(1,1), padding=kernel_size // 2
-                        , use_bias=not with_bn)
-                )
-            else:
-                self.net.add(
-                    nn.Conv2D(in_channels=inp, channels=cn, groups=inp, kernel_size=kernel_size, strides=(1,1), padding=kernel_size // 2
-                        , use_bias=False),
-                    nn.BatchNorm(),
-                    Act(),
-                    #nn.Activation('relu'),
-                    nn.Conv2D(in_channels=cn, channels=output, kernel_size=(1,1), strides=(1,1)
-                        , use_bias=not with_bn)
-                )
-
-            self.with_bn = with_bn
-            self.act = Act()
-            #self.act = nn.Activation('relu')
-            if with_bn:
-                self.bn = nn.BatchNorm()
-    def hybrid_forward(self, F ,x):
-        x = self.net(x)
-        if self.with_bn:
-            x = self.bn(x)
-        if self.act is not None:
-            x = self.act(x)
-        return x
-
-class ExpandedConv(nn.HybridBlock):
-    def __init__(self, inp, oup, t, strides, kernel=3, same_shape=True, **kwargs):
-        super(ExpandedConv, self).__init__(**kwargs)
-
-        self.same_shape = same_shape
-        self.strides = strides
-        with self.name_scope(): 
-            self.bottleneck = nn.HybridSequential()
-            self.bottleneck.add(
-                Conv1x1(inp*t, prefix="expand_"),
-                DWise(inp*t, self.strides, kernel, prefix="dwise_"),
-                Conv1x1(oup, is_linear=True, prefix="linear_")
-            )
-    def hybrid_forward(self, F, x):
-        out = self.bottleneck(x)
-        if self.strides == 1 and self.same_shape:
-            out = F.elemwise_add(out, x)
-        return out
-
-def ExpandedConvSequence(t, k, inp, oup, repeats, first_strides, **kwargs):
-    seq = nn.HybridSequential(**kwargs)
-    with seq.name_scope():
-        seq.add(ExpandedConv(inp, oup, t, first_strides, k, same_shape=False))
-        curr_inp = oup
-        for i in range(1, repeats):
-            seq.add(ExpandedConv(curr_inp, oup, t, 1))
-            curr_inp = oup
-    return seq
-
-class MNasNet(nn.HybridBlock):
-    def __init__(self, m=1.0, **kwargs):
-        super(MNasNet, self).__init__(**kwargs)
-        
-        self.first_oup = int(32*m)
-        self.second_oup = int(16*m)
-        #self.second_oup = int(32*m)
-        self.interverted_residual_setting = [
-            # t, c,  n, s, k
-            [3, int(24*m),  3, 2, 3, "stage2_"],  # -> 56x56
-            [3, int(40*m),  3, 2, 5, "stage3_"],  # -> 28x28
-            [6, int(80*m),  3, 2, 5, "stage4_1_"],  # -> 14x14
-            [6, int(96*m),  2, 1, 3, "stage4_2_"],  # -> 14x14
-            [6, int(192*m), 4, 2, 5, "stage5_1_"], # -> 7x7
-            [6, int(320*m), 1, 1, 3, "stage5_2_"], # -> 7x7          
-        ]
-        self.last_channels = int(1024*m)
-
-        with self.name_scope():
-            self.features = nn.HybridSequential()
-            self.features.add(ConvBlock(self.first_oup, 3, 1, prefix="stage1_conv0_"))
-            self.features.add(SepCONV(self.first_oup, self.second_oup, 3, prefix="stage1_sepconv0_"))
-            inp = self.second_oup
-            for i, (t, c, n, s, k, prefix) in enumerate(self.interverted_residual_setting):
-                oup = c
-                self.features.add(ExpandedConvSequence(t, k, inp, oup, n, s, prefix=prefix))
-                inp = oup
-
-            self.features.add(Conv1x1(self.last_channels, prefix="stage5_3_"))
-            #self.features.add(nn.GlobalAvgPool2D())
-            #self.features.add(nn.Flatten())
-            #self.output = nn.Dense(num_classes)
-    def hybrid_forward(self, F, x):
-        x = self.features(x)
-        #x = self.output(x)
-        return x
-
-    def num_output_channel(self):
-      return self.last_channels
-
-def get_symbol():
-  net = MNasNet(config.net_multiplier)
-  data = mx.sym.Variable(name='data')
-  data = data-127.5
-  data = data*0.0078125
-  body = net(data)
-  fc1 = symbol_utils.get_fc1(body, config.emb_size, config.net_output, input_channel=net.num_output_channel())
-  return fc1
-
diff --git a/recognition/symbol/fmobilefacenet.py b/recognition/symbol/fmobilefacenet.py
deleted file mode 100644
index 3050fc985..000000000
--- a/recognition/symbol/fmobilefacenet.py
+++ /dev/null
@@ -1,72 +0,0 @@
-
-import sys
-import os
-import mxnet as mx
-import symbol_utils
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.sym.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-def Conv(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=config.bn_mom)
-    act = Act(data=bn, act_type=config.net_act, name='%s%s_relu' %(name, suffix))
-    return act
-    
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=config.bn_mom)    
-    return bn
-
-def ConvOnly(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    return conv    
-
-    
-def DResidual(data, num_out=1, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=1, name=None, suffix=''):
-    conv = Conv(data=data, num_filter=num_group, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name='%s%s_conv_sep' %(name, suffix))
-    conv_dw = Conv(data=conv, num_filter=num_group, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name='%s%s_conv_dw' %(name, suffix))
-    proj = Linear(data=conv_dw, num_filter=num_out, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name='%s%s_conv_proj' %(name, suffix))
-    return proj
-    
-def Residual(data, num_block=1, num_out=1, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=1, name=None, suffix=''):
-    identity=data
-    for i in range(num_block):
-    	shortcut=identity
-    	conv=DResidual(data=identity, num_out=num_out, kernel=kernel, stride=stride, pad=pad, num_group=num_group, name='%s%s_block' %(name, suffix), suffix='%d'%i)
-    	identity=conv+shortcut
-    return identity
-        
-
-def get_symbol():
-    num_classes = config.emb_size
-    print('in_network', config)
-    fc_type = config.net_output
-    data = mx.symbol.Variable(name="data")
-    data = data-127.5
-    data = data*0.0078125
-    blocks = config.net_blocks
-    conv_1 = Conv(data, num_filter=64, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_1")
-    if blocks[0]==1:
-      conv_2_dw = Conv(conv_1, num_group=64, num_filter=64, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_2_dw")
-    else:
-      conv_2_dw = Residual(conv_1, num_block=blocks[0], num_out=64, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=64, name="res_2")
-    conv_23 = DResidual(conv_2_dw, num_out=64, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=128, name="dconv_23")
-    conv_3 = Residual(conv_23, num_block=blocks[1], num_out=64, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=128, name="res_3")
-    conv_34 = DResidual(conv_3, num_out=128, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=256, name="dconv_34")
-    conv_4 = Residual(conv_34, num_block=blocks[2], num_out=128, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=256, name="res_4")
-    conv_45 = DResidual(conv_4, num_out=128, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=512, name="dconv_45")
-    conv_5 = Residual(conv_45, num_block=blocks[3], num_out=128, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=256, name="res_5")
-    conv_6_sep = Conv(conv_5, num_filter=512, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_6sep")
-
-    fc1 = symbol_utils.get_fc1(conv_6_sep, num_classes, fc_type)
-    return fc1
-
diff --git a/recognition/symbol/fmobilenet.py b/recognition/symbol/fmobilenet.py
deleted file mode 100644
index 1b6837a34..000000000
--- a/recognition/symbol/fmobilenet.py
+++ /dev/null
@@ -1,87 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-import sys
-import os
-import mxnet as mx
-import symbol_utils
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.sym.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-def Conv(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=True)
-    act = Act(data=bn, act_type=config.net_act, name='%s%s_relu' %(name, suffix))
-    return act
-
-def ConvOnly(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    return conv
-
-def get_symbol():
-    num_classes = config.emb_size
-    bn_mom = config.bn_mom
-    workspace = config.workspace
-    data = mx.symbol.Variable(name="data") # 224
-    data = data-127.5
-    data = data*0.0078125
-    fc_type = config.net_output
-    bf = int(32*config.net_multiplier)
-    if config.net_input==0:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_1") # 224/112
-    else:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_1") # 224/112
-    conv_2_dw = Conv(conv_1, num_group=bf, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_2_dw") # 112/112
-    conv_2 = Conv(conv_2_dw, num_filter=bf*2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_2") # 112/112
-    conv_3_dw = Conv(conv_2, num_group=bf*2, num_filter=bf*2, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_3_dw") # 112/56
-    conv_3 = Conv(conv_3_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_3") # 56/56
-    conv_4_dw = Conv(conv_3, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_4_dw") # 56/56
-    conv_4 = Conv(conv_4_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_4") # 56/56
-    conv_5_dw = Conv(conv_4, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_5_dw") # 56/28
-    conv_5 = Conv(conv_5_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_5") # 28/28
-    conv_6_dw = Conv(conv_5, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_6_dw") # 28/28
-    conv_6 = Conv(conv_6_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_6") # 28/28
-    conv_7_dw = Conv(conv_6, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_7_dw") # 28/14
-    conv_7 = Conv(conv_7_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_7") # 14/14
-
-    conv_8_dw = Conv(conv_7, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_8_dw") # 14/14
-    conv_8 = Conv(conv_8_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_8") # 14/14
-    conv_9_dw = Conv(conv_8, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_9_dw") # 14/14
-    conv_9 = Conv(conv_9_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_9") # 14/14
-    conv_10_dw = Conv(conv_9, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_10_dw") # 14/14
-    conv_10 = Conv(conv_10_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_10") # 14/14
-    conv_11_dw = Conv(conv_10, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_11_dw") # 14/14
-    conv_11 = Conv(conv_11_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_11") # 14/14
-    conv_12_dw = Conv(conv_11, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_12_dw") # 14/14
-    conv_12 = Conv(conv_12_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_12") # 14/14
-
-    conv_13_dw = Conv(conv_12, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_13_dw") # 14/7
-    conv_13 = Conv(conv_13_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_13") # 7/7
-    conv_14_dw = Conv(conv_13, num_group=bf*32, num_filter=bf*32, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_14_dw") # 7/7
-    conv_14 = Conv(conv_14_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_14") # 7/7
-    body = conv_14
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
diff --git a/recognition/symbol/fresnet.py b/recognition/symbol/fresnet.py
deleted file mode 100644
index 09473851b..000000000
--- a/recognition/symbol/fresnet.py
+++ /dev/null
@@ -1,648 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-'''
-Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
-Original author Wei Wu
-
-Implemented the following paper:
-
-Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
-'''
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import sys
-import os
-import mxnet as mx
-import numpy as np
-import symbol_utils
-import memonger
-import sklearn
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-
-def Act(data, act_type, name):
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=stride, pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit2')
-    if bottle_neck:
-        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=stride, pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
-        conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv3 = mx.symbol.broadcast_mul(conv3, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv3 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv2 = mx.symbol.broadcast_mul(conv2, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv2 + shortcut
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit3')
-    if bottle_neck:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn4 = mx.symbol.broadcast_mul(bn4, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn4 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn3 + shortcut
-
-def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNeXt Unit symbol for building ResNeXt
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    assert(bottle_neck)
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    num_group = 32
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-    conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                               workspace=workspace, name=name + '_conv3')
-    bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-    if use_se:
-      #se begin
-      body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-      body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv1", workspace=workspace)
-      body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-      body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv2", workspace=workspace)
-      body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-      bn4 = mx.symbol.broadcast_mul(bn4, body)
-      #se end
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn4 + shortcut
-
-
-def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-  uv = kwargs.get('version_unit', 3)
-  version_input = kwargs.get('version_input', 1)
-  if uv==1:
-    if version_input==0:
-      return residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-    else:
-      return residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==2:
-    return residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==4:
-    return residual_unit_v4(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  else:
-    return residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-
-def resnet(units, num_stages, filter_list, num_classes, bottle_neck):
-    bn_mom = config.bn_mom
-    workspace = config.workspace
-    kwargs = {'version_se' : config.net_se,
-        'version_input': config.net_input,
-        'version_output': config.net_output,
-        'version_unit': config.net_unit,
-        'version_act': config.net_act,
-        'bn_mom': bn_mom,
-        'workspace': workspace,
-        'memonger': config.memonger,
-        }
-    """Return ResNet symbol of
-    Parameters
-    ----------
-    units : list
-        Number of units in each stage
-    num_stages : int
-        Number of stage
-    filter_list : list
-        Channel size of each stage
-    num_classes : int
-        Ouput size of symbol
-    dataset : str
-        Dataset type, only cifar10 and imagenet supports
-    workspace : int
-        Workspace used in convolution operator
-    """
-    version_se = kwargs.get('version_se', 1)
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    act_type = kwargs.get('version_act', 'prelu')
-    memonger = kwargs.get('memonger', False)
-    print(version_se, version_input, version_output, version_unit, act_type, memonger)
-    num_unit = len(units)
-    assert(num_unit == num_stages)
-    data = mx.sym.Variable(name='data')
-    if version_input==0:
-      #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-      #body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    elif version_input==2:
-      data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1,1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-    else:
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = data
-      body = Conv(data=body, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-
-    for i in range(num_stages):
-      #if version_input==0:
-      #  body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
-      #                       name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      #else:
-      #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
-      #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      body = residual_unit(body, filter_list[i+1], (2, 2), False,
-        name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      for j in range(units[i]-1):
-        body = residual_unit(body, filter_list[i+1], (1,1), True, name='stage%d_unit%d' % (i+1, j+2),
-          bottle_neck=bottle_neck, **kwargs)
-
-    if bottle_neck:
-      body = Conv(data=body, num_filter=512, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                no_bias=True, name="convd", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnd')
-      body = Act(data=body, act_type=act_type, name='relud')
-
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
-def get_symbol():
-    """
-    Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
-    Original author Wei Wu
-    """
-    num_classes = config.emb_size
-    num_layers = config.num_layers
-    if num_layers >= 500:
-        filter_list = [64, 256, 512, 1024, 2048]
-        bottle_neck = True
-    else:
-        filter_list = [64, 64, 128, 256, 512]
-        bottle_neck = False
-    num_stages = 4
-    if num_layers == 18:
-        units = [2, 2, 2, 2]
-    elif num_layers == 34:
-        units = [3, 4, 6, 3]
-    elif num_layers == 49:
-        units = [3, 4, 14, 3]
-    elif num_layers == 50:
-        units = [3, 4, 14, 3]
-    elif num_layers == 74:
-        units = [3, 6, 24, 3]
-    elif num_layers == 90:
-        units = [3, 8, 30, 3]
-    elif num_layers == 98:
-        units = [3, 4, 38, 3]
-    elif num_layers == 99:
-        units = [3, 8, 35, 3]
-    elif num_layers == 100:
-        units = [3, 13, 30, 3]
-    elif num_layers == 134:
-        units = [3, 10, 50, 3]
-    elif num_layers == 136:
-        units = [3, 13, 48, 3]
-    elif num_layers == 140:
-        units = [3, 15, 48, 3]
-    elif num_layers == 124:
-        units = [3, 13, 40, 5]
-    elif num_layers == 160:
-        units = [3, 24, 49, 3]
-    elif num_layers == 101:
-        units = [3, 4, 23, 3]
-    elif num_layers == 152:
-        units = [3, 8, 36, 3]
-    elif num_layers == 200:
-        units = [3, 24, 36, 3]
-    elif num_layers == 269:
-        units = [3, 30, 48, 8]
-    else:
-        raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
-
-    net = resnet(units       = units,
-                  num_stages  = num_stages,
-                  filter_list = filter_list,
-                  num_classes = num_classes,
-                  bottle_neck = bottle_neck)
-
-    if config.memonger:
-      dshape = (config.per_batch_size, config.image_shape[2], config.image_shape[0], config.image_shape[1])
-      net_mem_planned = memonger.search_plan(net, data=dshape)
-      old_cost = memonger.get_cost(net, data=dshape)
-      new_cost = memonger.get_cost(net_mem_planned, data=dshape)
-
-      print('Old feature map cost=%d MB' % old_cost)
-      print('New feature map cost=%d MB' % new_cost)
-      net = net_mem_planned
-    return net
-
-
-
diff --git a/recognition/symbol/memonger.py b/recognition/symbol/memonger.py
deleted file mode 100644
index d054bda20..000000000
--- a/recognition/symbol/memonger.py
+++ /dev/null
@@ -1,168 +0,0 @@
-import mxnet as mx
-import math
-
-def prod(shape):
-    """Get product of the shape.
-    """
-    ret = 1
-    for s in shape:
-        ret *= s
-    return ret
-
-
-def is_param(name):
-    """Quick script to check if name is a parameter.
-    """
-    if name == 'data':
-        return False
-    if name.endswith('weight'):
-        return True
-    if name.endswith('bias'):
-        return True
-    if name.endswith('beta'):
-        return True
-    if name.endswith('gamma'):
-        return True
-    return False
-
-
-def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
-    """Memory allocation planner with a given threshold.
-
-    The user can pass in a network configuration,
-    a threshold that limits memory per block.
-    And input shape configurations.
-
-    Parameters
-    ----------
-    sym : symbol
-        Input configuration of symbols.
-        The user need to pre-mark the attribute "mirror_stage" on the nodes
-        that can be book-kept as stage
-
-        The algorithm will decide whether to disbale mirror on the stage nodes.
-
-    threshold: integer
-        A tuning parameter to tune the approximate size of each stage blocks
-
-    plan_info: dict, optional
-        Used to hold plan information.
-
-    **kwargs:
-        The arguments to infer shape.
-
-    Returns
-    -------
-    alloc_sym: symbol
-        A symbol with force mirror tagged on the nodes for better allocation.
-    """
-    threshold = threshold << 20
-    sym = sym.__copy__()
-    internals = sym.get_internals()
-    _, out_shapes, _ = internals.infer_shape(**kwargs)
-    shape_dict = list(zip(internals.list_outputs(), out_shapes))
-    total_size = 0
-    param_size = 0
-    local_size = 0
-    save_size = 0
-    max_size = 0
-    last_sb = None
-    last_local = 0
-    period = 1
-    last_stage = ''
-    stage_decision = ''
-
-    for idx, item in enumerate(shape_dict):
-        sb = internals[idx]
-        name, shape = item
-        if is_param(name):
-            param_size += prod(shape) * 4
-            continue
-        else:
-            total_size += prod(shape) * 4
-            local_size += prod(shape) * 4
-            sb._set_attr(force_mirroring='True')
-
-        if sb.attr('mirror_stage') is not None:
-            stage = sb.attr('mirror_stage')
-            if stage == 'True' or stage != last_stage:
-                if local_size > threshold:
-                    save_size += prod(shape) * 4
-                    max_size = max(max_size, local_size)
-                    local_size = 0
-                    stage_decision = 'False'
-                    sb._set_attr(force_mirroring=stage_decision)
-                else:
-                    stage_decision = 'True'
-                    pass
-                last_stage = stage
-            elif stage == last_stage and stage_decision == 'False':
-                save_size += prod(shape) * 4
-                sb._set_attr(force_mirroring=stage_decision)
-
-    if plan_info is not None:
-        plan_info['max_size'] = max_size
-        plan_info['save_size'] = save_size
-    return sym
-
-
-def get_cost(sym, type_dict=None, **kwargs):
-    """Get the cost of the current symbolic plan by running bind on CPU.
-
-    sym : Symbolic Variable
-
-    """
-    texec = sym.simple_bind(ctx=mx.gpu(),
-                            grad_req='write',
-                            type_dict=type_dict,
-                            **kwargs)
-    return int(texec.debug_str().split('\n')[-3].split()[1])
-
-
-def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
-    """Quickly heurestic search over possible plans to find good memory plan.
-
-    Parameters
-    ----------
-    sym : symbolic
-       Symbolic configurations
-
-    ntrial: integer
-       Additional grid search steps
-    """
-    history = []
-    threshold = 0
-    min_threshold = None
-    min_cost = None
-    nbegin = 3
-
-    for k in range(nbegin):
-        info = {}
-        sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
-        cost = get_cost(sym, type_dict, **kwargs)
-        save_size = info['save_size'] >> 20
-        local_size = info['max_size'] >> 20
-        guess = int(math.sqrt(save_size * local_size / 2))
-        if min_cost is None or min_cost > cost:
-            min_cost = cost
-        if min_threshold is None or local_size < min_threshold:
-            min_threshold = local_size
-        print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-        history.append((cost, threshold, sym))
-        threshold = guess
-
-    max_threshold = threshold * math.sqrt(2)
-    step = int((max_threshold - min_threshold) / ntrial)
-    threshold = min_threshold + step
-    if step > 0:
-        for k in range(ntrial):
-            sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
-            cost = get_cost(sym, type_dict, **kwargs)
-            print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-            history.append((cost, threshold, sym))
-            threshold += step
-
-    history.sort(key = lambda x: x[0])
-    cost, threshold, sym = history[0]
-    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
-    return sym
diff --git a/recognition/symbol/memonger_v2.py b/recognition/symbol/memonger_v2.py
deleted file mode 100644
index 70f7a008e..000000000
--- a/recognition/symbol/memonger_v2.py
+++ /dev/null
@@ -1,275 +0,0 @@
-import mxnet as mx
-import math
-
-def prod(shape):
-    """Get product of the shape.
-    """
-    ret = 1
-    for s in shape:
-        ret *= s
-    return ret
-
-
-def is_param(name):
-    """Quick script to check if name is a parameter.
-    """
-    if name == 'data':
-        return False
-    if name.endswith('weight'):
-        return True
-    if name.endswith('bias'):
-        return True
-    if name.endswith('beta'):
-        return True
-    if name.endswith('gamma'):
-        return True
-    return False
-
-
-def make_mirror_plan(sym, threshold, plan_info=None, **kwargs):
-    """Memory allocation planner with a given threshold.
-
-    The user can pass in a network configuration,
-    a threshold that limits memory per block.
-    And input shape configurations.
-
-    Parameters
-    ----------
-    sym : symbol
-        Input configuration of symbols.
-        The user need to pre-mark the attribute "mirror_stage" on the nodes
-        that can be book-kept as stage
-
-        The algorithm will decide whether to disbale mirror on the stage nodes.
-
-    threshold: integer
-        A tuning parameter to tune the approximate size of each stage blocks
-
-    plan_info: dict, optional
-        Used to hold plan information.
-
-    **kwargs:
-        The arguments to infer shape.
-
-    Returns
-    -------
-    alloc_sym: symbol
-        A symbol with force mirror tagged on the nodes for better allocation.
-    """
-    threshold = threshold << 20
-    sym = sym.__copy__()
-    internals = sym.get_internals()
-    _, out_shapes, _ = internals.infer_shape(**kwargs)
-    shape_dict = list(zip(internals.list_outputs(), out_shapes))
-    total_size = 0
-    param_size = 0
-    local_size = 0
-    save_size = 0
-    max_size = 0
-    last_sb = None
-    last_local = 0
-    period = 1
-    last_stage = ''
-    stage_decision = ''
-
-    for idx, item in enumerate(shape_dict):
-        sb = internals[idx]
-        name, shape = item
-        if is_param(name):
-            param_size += prod(shape) * 4
-            continue
-        else:
-            total_size += prod(shape) * 4
-            local_size += prod(shape) * 4
-            sb._set_attr(force_mirroring='True')
-
-        if sb.attr('mirror_stage') is not None:
-            stage = sb.attr('mirror_stage')
-            if stage == 'True' or stage != last_stage:
-                if local_size > threshold:
-                    save_size += prod(shape) * 4
-                    max_size = max(max_size, local_size)
-                    local_size = 0
-                    stage_decision = 'False'
-                    sb._set_attr(force_mirroring=stage_decision)
-                else:
-                    stage_decision = 'True'
-                    pass
-                last_stage = stage
-            elif stage == last_stage and stage_decision == 'False':
-                save_size += prod(shape) * 4
-                sb._set_attr(force_mirroring=stage_decision)
-
-    if plan_info is not None:
-        plan_info['max_size'] = max_size
-        plan_info['save_size'] = save_size
-    return sym
-
-
-def get_cost(sym, type_dict=None, **kwargs):
-    """Get the cost of the current symbolic plan by running bind on CPU.
-
-    sym : Symbolic Variable
-
-    """
-    texec = sym.simple_bind(ctx=mx.gpu(),
-                            grad_req='write',
-                            type_dict=type_dict,
-                            **kwargs)
-    return int(texec.debug_str().split('\n')[-3].split()[1])
-
-
-def search_plan(sym, ntrial=6, type_dict=None, **kwargs):
-    """Quickly heurestic search over possible plans to find good memory plan.
-
-    Parameters
-    ----------
-    sym : symbolic
-       Symbolic configurations
-
-    ntrial: integer
-       Additional grid search steps
-    """
-    history = []
-    threshold = 0
-    min_threshold = None
-    min_cost = None
-    nbegin = 3
-
-    for k in range(nbegin):
-        info = {}
-        sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
-        cost = get_cost(sym, type_dict, **kwargs)
-        save_size = info['save_size'] >> 20
-        local_size = info['max_size'] >> 20
-        guess = int(math.sqrt(save_size * local_size / 2))
-        if min_cost is None or min_cost > cost:
-            min_cost = cost
-        if min_threshold is None or local_size < min_threshold:
-            min_threshold = local_size
-        print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-        history.append((cost, threshold, sym))
-        threshold = guess
-
-    max_threshold = threshold * math.sqrt(2)
-    step = int((max_threshold - min_threshold) / ntrial)
-    threshold = min_threshold + step
-    if step > 0:
-        for k in range(ntrial):
-            sym = make_mirror_plan(sym, threshold=threshold, plan_info=info, **kwargs)
-            cost = get_cost(sym, type_dict, **kwargs)
-            print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-            history.append((cost, threshold, sym))
-            threshold += step
-
-    history.sort(key = lambda x: x[0])
-    cost, threshold, sym = history[0]
-    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
-    return sym
-
-
-def make_mirror_plan_to_layer(sym,layer_name,threshold, plan_info=None, **kwargs):
-    """
-    sym is the original symbal
-    layer_name is a name to which layer of the network should be set as mirror
-    threshhold is the approximate size of each mirror block
-    """
-    threshold = threshold << 20
-    sym = sym.__copy__()
-    internals = sym.get_internals()
-    _, out_shapes, _ = internals.infer_shape(**kwargs)
-    shape_dict = list(zip(internals.list_outputs(), out_shapes))
-    total_size = 0
-    param_size = 0
-    local_size = 0
-    save_size = 0
-    max_size = 0
-    last_stage = ''
-    stage_decision = ''
-    switch=True
-    
-    for idx, item in enumerate(shape_dict):
-        sb = internals[idx]
-        name, shape = item
-        #print(name, switch)
-        if is_param(name):
-            param_size += prod(shape) * 4
-            continue
-        elif switch and not 'bn' in name:
-            total_size += prod(shape) * 4
-            local_size += prod(shape) * 4
-            sb._set_attr(force_mirroring='True')
-            print('set force_mirroring', name, total_size, local_size)
-        if layer_name!='' and layer_name in name:
-            switch = False
-
-        if sb.attr('mirror_stage') is not None:
-            stage = sb.attr('mirror_stage')
-            #print(name, stage)
-            if stage == 'True' or stage != last_stage:
-                if local_size > threshold:
-                    save_size += prod(shape) * 4
-                    max_size = max(max_size, local_size)
-                    local_size = 0
-                    stage_decision = 'False'
-                    sb._set_attr(force_mirroring=stage_decision)
-                else:
-                    stage_decision = 'True'
-                    pass
-                last_stage = stage
-            elif stage == last_stage and stage_decision == 'False':
-                save_size += prod(shape) * 4
-                sb._set_attr(force_mirroring=stage_decision)
-
-    if plan_info is not None:
-        plan_info['max_size'] = max_size
-        plan_info['save_size'] = save_size
-    return sym
-
-def search_plan_to_layer(sym,layer_name=None,threshold=500, ntrial=6, type_dict=None, **kwargs):
-    """Quickly heurestic search over possible plans to find good memory plan.
-
-    Parameters
-    ----------
-    sym : symbolic
-       Symbolic configurations
-
-    ntrial: integer
-       Additional grid search steps
-    """
-    history = []
-    min_threshold = None
-    min_cost = None
-    nbegin = 10
-
-    for k in range(nbegin):
-        info = {}
-        sym = make_mirror_plan_to_layer(sym,layer_name=layer_name,threshold=threshold, plan_info=info, **kwargs)
-        cost = get_cost(sym, type_dict, **kwargs)
-        save_size = info['save_size'] >> 20
-        local_size = info['max_size'] >> 20
-        guess = 300 * (k + 1)
-        if min_cost is None or min_cost > cost:
-            min_cost = cost
-        if min_threshold is None or local_size < min_threshold:
-            min_threshold = local_size
-        print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-        history.append((cost, threshold, sym))
-        threshold = guess
-
-    max_threshold = threshold * math.sqrt(2)
-    step = int((max_threshold - min_threshold) / ntrial)
-    print(min_threshold, max_threshold, step)
-    threshold = min_threshold + step
-    if step > 0:
-        for k in range(ntrial):
-            sym = make_mirror_plan_to_layer(sym,layer_name=layer_name, threshold=threshold, plan_info=info, **kwargs)
-            cost = get_cost(sym, type_dict, **kwargs)
-            print ("Search threshold=%d MB, cost=%d MB" % (threshold, cost))
-            history.append((cost, threshold, sym))
-            threshold += step
-
-    history.sort(key = lambda x: x[0])
-    cost, threshold, sym = history[0]
-    print('Find best plan with threshold=%d, cost=%d MB' % (threshold, cost))
-    return sym
diff --git a/recognition/symbol/symbol_utils.py b/recognition/symbol/symbol_utils.py
deleted file mode 100644
index 1bf22d7e4..000000000
--- a/recognition/symbol/symbol_utils.py
+++ /dev/null
@@ -1,270 +0,0 @@
-import sys
-import os
-import mxnet as mx
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.sym.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-bn_mom = config.bn_mom
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=bn_mom)    
-    return bn
-
-def get_fc1(last_conv, num_classes, fc_type, input_channel=512):
-  body = last_conv
-  if fc_type=='Z':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = body
-  elif fc_type=='E':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='FC':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='SFC':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = Conv(data=body, num_filter=input_channel, kernel=(3,3), stride=(2,2), pad=(1,1),
-                              no_bias=True, name="convf", num_group = input_channel)
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnf')
-    body = Act(data=body, act_type=config.net_act, name='reluf')
-    body = Conv(data=body, num_filter=input_channel, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="convf2")
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bnf2')
-    body = Act(data=body, act_type=config.net_act, name='reluf2')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GAP':
-    bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    relu1 = Act(data=bn1, act_type=config.net_act, name='relu1')
-    # Although kernel is not used here when global_pool=True, we should put one
-    pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
-    flat = mx.sym.Flatten(data=pool1)
-    fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GNAP': #mobilefacenet++
-    filters_in = 512 # param in mobilefacenet
-    if num_classes>filters_in:
-      body = mx.sym.Convolution(data=last_conv, num_filter=num_classes, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True, name='convx')
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=0.9, name='convx_bn')
-      body = Act(data=body, act_type=config.net_act, name='convx_relu')
-      filters_in = num_classes
-    else:
-      body = last_conv
-    body = mx.sym.BatchNorm(data=body, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6f')  
-
-    spatial_norm=body*body
-    spatial_norm=mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
-    spatial_sqrt=mx.sym.sqrt(spatial_norm)
-    #spatial_mean=mx.sym.mean(spatial_sqrt, axis=(1,2,3), keepdims=True)
-    spatial_mean=mx.sym.mean(spatial_sqrt)
-    spatial_div_inverse=mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
-
-    spatial_attention_inverse=mx.symbol.tile(spatial_div_inverse, reps=(1,filters_in,1,1))   
-    body=body*spatial_attention_inverse
-    #body = mx.sym.broadcast_mul(body, spatial_div_inverse)
-
-    fc1 = mx.sym.Pooling(body, kernel=(7, 7), global_pool=True, pool_type='avg')
-    if num_classes<filters_in:
-      fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6w')
-      fc1 = mx.sym.FullyConnected(data=fc1, num_hidden=num_classes, name='pre_fc1')
-    else:
-      fc1 = mx.sym.Flatten(data=fc1)
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='fc1')
-  elif fc_type=="GDC": #mobilefacenet_v1
-    conv_6_dw = Linear(last_conv, num_filter=input_channel, num_group=input_channel, kernel=(7,7), pad=(0, 0), stride=(1, 1), name="conv_6dw7_7")  
-    conv_6_f = mx.sym.FullyConnected(data=conv_6_dw, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=conv_6_f, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='F':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='G':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='H':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='I':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='J':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  return fc1
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type=config.net_act, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn3 + shortcut
-
-def residual_unit_v1l(data, num_filter, stride, dim_match, name, bottle_neck):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    workspace = config.workspace
-    bn_mom = config.bn_mom
-    memonger = False
-    use_se = config.net_se
-    act_type = config.net_act
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def get_head(data, version_input, num_filter):
-    bn_mom = config.bn_mom
-    workspace = config.workspace
-    kwargs = {'bn_mom': bn_mom, 'workspace' : workspace}
-    data = data-127.5
-    data = data*0.0078125
-    #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-    if version_input==0:
-      body = Conv(data=data, num_filter=num_filter, kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=config.net_act, name='relu0')
-      body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    else:
-      body = data
-      _num_filter = min(num_filter, 64)
-      body = Conv(data=body, num_filter=_num_filter, kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=config.net_act, name='relu0')
-      #body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
-      body = residual_unit_v1l(body, _num_filter, (2, 2), False, name='head', bottle_neck=False)
-    return body
-
-
diff --git a/recognition/symbol/vargfacenet.py b/recognition/symbol/vargfacenet.py
deleted file mode 100644
index e521e5b7d..000000000
--- a/recognition/symbol/vargfacenet.py
+++ /dev/null
@@ -1,568 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-'''
-Author: Horizon Robotics Inc.
-The company is committed to be the global leader of edge AI platform.
-The model implemented in this scripts runs ~200fps on the Sunrise 2.
-Sunrise 2 is the second generation of an embedded AI chip designed by Horizon Robotics,
-targeting to empower AIoT devices by AI.
-
-Implemented the following paper:
-Mengjia Yan, Mengao Zhao, Zining Xu, Qian Zhang, Guoli Wang, Zhizhong Su. "VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition" (https://arxiv.org/abs/1910.04985)
-
-'''
-
-import os
-import sys
-
-import mxnet as mx
-import symbol_utils
-
-sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
-from config import config
-
-def Act(data, act_type, name):
-    if act_type == 'prelu':
-        body = mx.sym.LeakyReLU(data=data, act_type='prelu', name=name)
-    else:
-        body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-
-def get_setting_params(**kwargs):
-    # bn_params
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    bn_eps = kwargs.get('bn_eps', 2e-5)
-    fix_gamma = kwargs.get('fix_gamma', False)
-    use_global_stats = kwargs.get('use_global_stats', False)
-    # net_setting param
-    workspace = kwargs.get('workspace', 512)
-    act_type = kwargs.get('act_type', 'prelu')
-    use_se = kwargs.get('use_se', True)
-    se_ratio = kwargs.get('se_ratio', 4)
-    group_base = kwargs.get('group_base', 8)
-
-    setting_params = {}
-    setting_params['bn_mom'] = bn_mom
-    setting_params['bn_eps'] = bn_eps
-    setting_params['fix_gamma'] = fix_gamma
-    setting_params['use_global_stats'] = use_global_stats
-    setting_params['workspace'] = workspace
-    setting_params['act_type'] = act_type
-    setting_params['use_se'] = use_se
-    setting_params['se_ratio'] = se_ratio
-    setting_params['group_base'] = group_base
-
-    return setting_params
-
-
-def se_block(data, num_filter, setting_params, name):
-    se_ratio = setting_params['se_ratio']
-    act_type = setting_params['act_type']
-
-    pool1 = mx.sym.Pooling(data=data,
-                           global_pool=True,
-                           pool_type='avg',
-                           name=name + '_se_pool1')
-    conv1 = mx.sym.Convolution(data=pool1,
-                               num_filter=num_filter // se_ratio,
-                               kernel=(1, 1),
-                               stride=(1, 1),
-                               pad=(0, 0),
-                               name=name + "_se_conv1")
-    act1 = Act(data=conv1, act_type=act_type, name=name + '_se_act1')
-
-    conv2 = mx.sym.Convolution(data=act1,
-                               num_filter=num_filter,
-                               kernel=(1, 1),
-                               stride=(1, 1),
-                               pad=(0, 0),
-                               name=name + "_se_conv2")
-    act2 = mx.symbol.Activation(data=conv2,
-                                act_type='sigmoid',
-                                name=name + "_se_sigmoid")
-    out_data = mx.symbol.broadcast_mul(data, act2)
-    return out_data
-
-
-def separable_conv2d(data,
-                     in_channels,
-                     out_channels,
-                     kernel,
-                     pad,
-                     setting_params,
-                     stride=(1, 1),
-                     factor=1,
-                     bias=False,
-                     bn_dw_out=True,
-                     act_dw_out=True,
-                     bn_pw_out=True,
-                     act_pw_out=True,
-                     dilate=1,
-                     name=None):
-    bn_mom = setting_params['bn_mom']
-    bn_eps = setting_params['bn_eps']
-    fix_gamma = setting_params['fix_gamma']
-    use_global_stats = setting_params['use_global_stats']
-    workspace = setting_params['workspace']
-    group_base = setting_params['group_base']
-    act_type = setting_params['act_type']
-    assert in_channels % group_base == 0
-
-    # depthwise
-    dw_out = mx.sym.Convolution(data=data,
-                                num_filter=int(in_channels * factor),
-                                kernel=kernel,
-                                pad=pad,
-                                stride=stride,
-                                no_bias=False if bias else True,
-                                num_group=int(in_channels / group_base),
-                                dilate=(dilate, dilate),
-                                workspace=workspace,
-                                name=name + '_conv2d_depthwise')
-    if bn_dw_out:
-        dw_out = mx.sym.BatchNorm(data=dw_out,
-                                  fix_gamma=fix_gamma,
-                                  eps=bn_eps,
-                                  momentum=bn_mom,
-                                  use_global_stats=use_global_stats,
-                                  name=name + '_conv2d_depthwise_bn')
-    if act_dw_out:
-        dw_out = Act(data=dw_out,
-                     act_type=act_type,
-                     name=name + '_conv2d_depthwise_act')
-    # pointwise
-    pw_out = mx.sym.Convolution(data=dw_out,
-                                num_filter=out_channels,
-                                kernel=(1, 1),
-                                stride=(1, 1),
-                                pad=(0, 0),
-                                num_group=1,
-                                no_bias=False if bias else True,
-                                workspace=workspace,
-                                name=name + '_conv2d_pointwise')
-    if bn_pw_out:
-        pw_out = mx.sym.BatchNorm(data=pw_out,
-                                  fix_gamma=fix_gamma,
-                                  eps=bn_eps,
-                                  momentum=bn_mom,
-                                  use_global_stats=use_global_stats,
-                                  name=name + '_conv2d_pointwise_bn')
-    if act_pw_out:
-        pw_out = Act(data=pw_out,
-                     act_type=act_type,
-                     name=name + '_conv2d_pointwise_act')
-    return pw_out
-
-
-def vargnet_block(data,
-                  n_out_ch1,
-                  n_out_ch2,
-                  n_out_ch3,
-                  setting_params,
-                  factor=2,
-                  dim_match=True,
-                  multiplier=1,
-                  kernel=(3, 3),
-                  stride=(1, 1),
-                  dilate=1,
-                  with_dilate=False,
-                  name=None):
-    use_se = setting_params['use_se']
-    act_type = setting_params['act_type']
-
-    out_channels_1 = int(n_out_ch1 * multiplier)
-    out_channels_2 = int(n_out_ch2 * multiplier)
-    out_channels_3 = int(n_out_ch3 * multiplier)
-
-    pad = (((kernel[0] - 1) * dilate + 1) // 2,
-           ((kernel[1] - 1) * dilate + 1) // 2)
-
-    if with_dilate:
-        stride = (1, 1)
-    if dim_match:
-        short_cut = data
-    else:
-        short_cut = separable_conv2d(data=data,
-                                     in_channels=out_channels_1,
-                                     out_channels=out_channels_3,
-                                     kernel=kernel,
-                                     pad=pad,
-                                     setting_params=setting_params,
-                                     stride=stride,
-                                     factor=factor,
-                                     bias=False,
-                                     act_pw_out=False,
-                                     dilate=dilate,
-                                     name=name + '_shortcut')
-    sep1_data = separable_conv2d(data=data,
-                                 in_channels=out_channels_1,
-                                 out_channels=out_channels_2,
-                                 kernel=kernel,
-                                 pad=pad,
-                                 setting_params=setting_params,
-                                 stride=stride,
-                                 factor=factor,
-                                 bias=False,
-                                 dilate=dilate,
-                                 name=name + '_sep1_data')
-    sep2_data = separable_conv2d(data=sep1_data,
-                                 in_channels=out_channels_2,
-                                 out_channels=out_channels_3,
-                                 kernel=kernel,
-                                 pad=pad,
-                                 setting_params=setting_params,
-                                 stride=(1, 1),
-                                 factor=factor,
-                                 bias=False,
-                                 dilate=dilate,
-                                 act_pw_out=False,
-                                 name=name + '_sep2_data')
-
-    if use_se:
-        sep2_data = se_block(data=sep2_data,
-                             num_filter=out_channels_3,
-                             setting_params=setting_params,
-                             name=name)
-
-    out_data = sep2_data + short_cut
-    out_data = Act(data=out_data, act_type=act_type, name=name + '_out_data_act')
-    return out_data
-
-
-def vargnet_branch_merge_block(data,
-                               n_out_ch1,
-                               n_out_ch2,
-                               n_out_ch3,
-                               setting_params,
-                               factor=2,
-                               dim_match=False,
-                               multiplier=1,
-                               kernel=(3, 3),
-                               stride=(2, 2),
-                               dilate=1,
-                               with_dilate=False,
-                               name=None):
-    act_type = setting_params['act_type']
-
-    out_channels_1 = int(n_out_ch1 * multiplier)
-    out_channels_2 = int(n_out_ch2 * multiplier)
-    out_channels_3 = int(n_out_ch3 * multiplier)
-
-    pad = (((kernel[0] - 1) * dilate + 1) // 2,
-           ((kernel[1] - 1) * dilate + 1) // 2)
-
-    if with_dilate:
-        stride = (1, 1)
-    if dim_match:
-        short_cut = data
-    else:
-        short_cut = separable_conv2d(data=data,
-                                     in_channels=out_channels_1,
-                                     out_channels=out_channels_3,
-                                     kernel=kernel,
-                                     pad=pad,
-                                     setting_params=setting_params,
-                                     stride=stride,
-                                     factor=factor,
-                                     bias=False,
-                                     act_pw_out=False,
-                                     dilate=dilate,
-                                     name=name + '_shortcut')
-    sep1_data_brach1 = separable_conv2d(data=data,
-                                        in_channels=out_channels_1,
-                                        out_channels=out_channels_2,
-                                        kernel=kernel,
-                                        pad=pad,
-                                        setting_params=setting_params,
-                                        stride=stride,
-                                        factor=factor,
-                                        bias=False,
-                                        dilate=dilate,
-                                        act_pw_out=False,
-                                        name=name + '_sep1_data_branch')
-    sep1_data_brach2 = separable_conv2d(data=data,
-                                        in_channels=out_channels_1,
-                                        out_channels=out_channels_2,
-                                        kernel=kernel,
-                                        pad=pad,
-                                        setting_params=setting_params,
-                                        stride=stride,
-                                        factor=factor,
-                                        bias=False,
-                                        dilate=dilate,
-                                        act_pw_out=False,
-                                        name=name + '_sep2_data_branch')
-    sep1_data = sep1_data_brach1 + sep1_data_brach2
-    sep1_data = Act(data=sep1_data, act_type=act_type, name=name + '_sep1_data_act')
-    sep2_data = separable_conv2d(data=sep1_data,
-                                 in_channels=out_channels_2,
-                                 out_channels=out_channels_3,
-                                 kernel=kernel,
-                                 pad=pad,
-                                 setting_params=setting_params,
-                                 stride=(1, 1),
-                                 factor=factor,
-                                 bias=False,
-                                 dilate=dilate,
-                                 act_pw_out=False,
-                                 name=name + '_sep2_data')
-    out_data = sep2_data + short_cut
-    out_data = Act(data=out_data, act_type=act_type, name=name + '_out_data_act')
-    return out_data
-
-
-def add_vargnet_conv_block(data,
-                           stage,
-                           units,
-                           in_channels,
-                           out_channels,
-                           setting_params,
-                           kernel=(3, 3),
-                           stride=(2, 2),
-                           multiplier=1,
-                           factor=2,
-                           dilate=1,
-                           with_dilate=False,
-                           name=None):
-    assert stage >= 2, 'stage is {}, stage must be set >=2'.format(stage)
-    data = vargnet_branch_merge_block(data=data,
-                                      n_out_ch1=in_channels,
-                                      n_out_ch2=out_channels,
-                                      n_out_ch3=out_channels,
-                                      setting_params=setting_params,
-                                      factor=factor,
-                                      dim_match=False,
-                                      multiplier=multiplier,
-                                      kernel=kernel,
-                                      stride=stride,
-                                      dilate=dilate,
-                                      with_dilate=with_dilate,
-                                      name=name + '_stage_{}_unit_1'.format(stage))
-    for i in range(units - 1):
-        data = vargnet_block(data=data,
-                             n_out_ch1=out_channels,
-                             n_out_ch2=out_channels,
-                             n_out_ch3=out_channels,
-                             setting_params=setting_params,
-                             factor=factor,
-                             dim_match=True,
-                             multiplier=multiplier,
-                             kernel=kernel,
-                             stride=(1, 1),
-                             dilate=dilate,
-                             with_dilate=with_dilate,
-                             name=name + '_stage_{}_unit_{}'.format(stage, i + 2))
-    return data
-
-
-def add_head_block(data,
-                   num_filter,
-                   setting_params,
-                   multiplier,
-                   head_pooling=False,
-                   kernel=(3, 3),
-                   stride=(2, 2),
-                   pad=(1, 1),
-                   name=None):
-    bn_mom = setting_params['bn_mom']
-    bn_eps = setting_params['bn_eps']
-    fix_gamma = setting_params['fix_gamma']
-    use_global_stats = setting_params['use_global_stats']
-    workspace = setting_params['workspace']
-    act_type = setting_params['act_type']
-    channels = int(num_filter * multiplier)
-
-    conv1 = mx.sym.Convolution(data=data,
-                               num_filter=channels,
-                               kernel=kernel,
-                               pad=pad,
-                               stride=stride,
-                               no_bias=True,
-                               num_group=1,
-                               workspace=workspace,
-                               name=name + '_conv1')
-    bn1 = mx.sym.BatchNorm(data=conv1,
-                           fix_gamma=fix_gamma,
-                           eps=bn_eps,
-                           momentum=bn_mom,
-                           use_global_stats=use_global_stats,
-                           name=name + '_conv1_bn')
-
-    act1 = Act(data=bn1, act_type=act_type, name=name + '_conv1_act')
-
-    if head_pooling:
-        head_data = mx.symbol.Pooling(data=act1,
-                                      kernel=(3, 3),
-                                      stride=(2, 2),
-                                      pad=(1, 1),
-                                      pool_type='max',
-                                      name=name + '_max_pooling')
-    else:
-        head_data = vargnet_block(data=act1,
-                                  n_out_ch1=num_filter,
-                                  n_out_ch2=num_filter,
-                                  n_out_ch3=num_filter,
-                                  setting_params=setting_params,
-                                  factor=1,
-                                  dim_match=False,
-                                  multiplier=multiplier,
-                                  kernel=kernel,
-                                  stride=(2, 2),
-                                  dilate=1,
-                                  with_dilate=False,
-                                  name=name + '_head_pooling')
-    return head_data
-
-
-def add_emb_block(data,
-                  input_channels,
-                  last_channels,
-                  emb_size,
-                  fc_type,
-                  setting_params,
-                  bias=False,
-                  name=None):
-    bn_mom = setting_params['bn_mom']
-    bn_eps = setting_params['bn_eps']
-    fix_gamma = setting_params['fix_gamma']
-    use_global_stats = setting_params['use_global_stats']
-    workspace = setting_params['workspace']
-    act_type = setting_params['act_type']
-    group_base = setting_params['group_base']
-    # last channels
-    if input_channels != last_channels:
-        data = mx.sym.Convolution(data=data,
-                                  num_filter=last_channels,
-                                  kernel=(1, 1),
-                                  pad=(0, 0),
-                                  stride=(1, 1),
-                                  no_bias=False if bias else True,
-                                  workspace=workspace,
-                                  name=name + '_convx')
-        data = mx.sym.BatchNorm(data=data,
-                                fix_gamma=fix_gamma,
-                                eps=bn_eps,
-                                momentum=bn_mom,
-                                use_global_stats=use_global_stats,
-                                name=name + '_convx_bn')
-        data = Act(data=data, act_type=act_type, name=name + '_convx_act')
-    # depthwise
-    convx_depthwise = mx.sym.Convolution(data=data,
-                                         num_filter=last_channels,
-                                         num_group=int(last_channels / group_base),
-                                         kernel=(7, 7),
-                                         pad=(0, 0),
-                                         stride=(1, 1),
-                                         no_bias=False if bias else True,
-                                         workspace=workspace,
-                                         name=name + '_convx_depthwise')
-    convx_depthwise = mx.sym.BatchNorm(data=convx_depthwise,
-                                       fix_gamma=fix_gamma,
-                                       eps=bn_eps,
-                                       momentum=bn_mom,
-                                       use_global_stats=use_global_stats,
-                                       name=name + '_convx_depthwise_bn')
-    # pointwise
-    convx_pointwise = mx.sym.Convolution(data=convx_depthwise,
-                                         num_filter=last_channels // 2,
-                                         kernel=(1, 1),
-                                         pad=(0, 0),
-                                         stride=(1, 1),
-                                         no_bias=False if bias else True,
-                                         workspace=workspace,
-                                         name=name + '_convx_pointwise')
-    convx_pointwise = mx.sym.BatchNorm(data=convx_pointwise,
-                                       fix_gamma=fix_gamma,
-                                       eps=bn_eps,
-                                       momentum=bn_mom,
-                                       use_global_stats=use_global_stats,
-                                       name=name + '_convx_pointwise_bn')
-    convx_pointwise = Act(data=convx_pointwise,
-                          act_type=act_type,
-                          name=name + '_convx_pointwise_act')
-
-    fc1 = symbol_utils.get_fc1(convx_pointwise, emb_size, fc_type)
-    return fc1
-
-
-def get_symbol():
-    multiplier = config.net_multiplier
-    emb_size = config.emb_size
-    fc_type = config.net_output
-
-    kwargs = {'use_se' : config.net_se,
-        'act_type': config.net_act,
-        'bn_mom': config.bn_mom,
-        'workspace': config.workspace,
-        }
-
-    setting_params = get_setting_params(**kwargs)
-
-    factor = 2
-    head_pooling = False
-    num_stage = 3
-    stage_list = [2, 3, 4]
-    units = [3, 7, 4]
-    filter_list = [32, 64, 128, 256]
-    last_channels = 1024
-    dilate_list = [1, 1, 1]
-    with_dilate_list = [False, False, False]
-
-    data = mx.sym.Variable(name='data')
-    data = mx.sym.identity(data=data, name='id')
-    data = data - 127.5
-    data = data * 0.0078125
-
-    body = add_head_block(data=data,
-                          num_filter=filter_list[0],
-                          setting_params=setting_params,
-                          multiplier=multiplier,
-                          head_pooling=head_pooling,
-                          kernel=(3, 3),
-                          stride=(1, 1),
-                          pad=(1, 1),
-                          name="vargface_head")
-
-    for i in range(num_stage):
-        body = add_vargnet_conv_block(data=body,
-                                      stage=stage_list[i],
-                                      units=units[i],
-                                      in_channels=filter_list[i],
-                                      out_channels=filter_list[i + 1],
-                                      setting_params=setting_params,
-                                      kernel=(3, 3),
-                                      stride=(2, 2),
-                                      multiplier=multiplier,
-                                      factor=factor,
-                                      dilate=dilate_list[i],
-                                      with_dilate=with_dilate_list[i],
-                                      name="vargface")
-    emb_feat = add_emb_block(data=body,
-                             input_channels=filter_list[3],
-                             last_channels=last_channels,
-                             emb_size=emb_size,
-                             fc_type=fc_type,
-                             setting_params=setting_params,
-                             bias=False,
-                             name='embed')
-    return emb_feat
-
-
-if __name__ == '__main__':
-    get_symbol()
diff --git a/recognition/vpl/README.md b/recognition/vpl/README.md
new file mode 100644
index 000000000..e2f26c2e2
--- /dev/null
+++ b/recognition/vpl/README.md
@@ -0,0 +1,21 @@
+
+# Variational Prototype Learning for Deep Face Recognition
+
+This is the Pytorch implementation of our paper  [Variational Prototype Learning for Deep Face Recognition](https://openaccess.thecvf.com/content/CVPR2021/papers/Deng_Variational_Prototype_Learning_for_Deep_Face_Recognition_CVPR_2021_paper.pdf) which is accepted by CVPR-2021.
+
+## How to run
+
+Define a new configure file such as `configs/example_ms1m.py`, and start the training process by:
+
+``
+bash run.sh configs/example_ms1m.py
+``
+
+## Results
+
+Results on WebFace600K(subset of WebFace260M), loss is margin-based softmax.
+
+| Backbone   | Dataset    | VPL? | Mask   | Children | African | Caucasian | South Asian | East Asian | MR-All    | 
+|------------|------------|------------|--------|----------|---------|-----------|-------------|------------|--------|
+| R50  | WebFace600K  | NO | 78.949 | 74.772   | 89.231  | 94.114    | 92.308      | 73.765     | 90.591 | 
+| R50  | WebFace600K  | YES | 78.884 | 75.739   | 89.424  | 94.220    | 92.609      | 74.365     | 90.942 | 
diff --git a/recognition/vpl/backbones/__init__.py b/recognition/vpl/backbones/__init__.py
new file mode 100644
index 000000000..a71f3ed9e
--- /dev/null
+++ b/recognition/vpl/backbones/__init__.py
@@ -0,0 +1,19 @@
+from .iresnet import iresnet18, iresnet34, iresnet50, iresnet100, iresnet200
+
+
+def get_model(name, **kwargs):
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    elif name == "r1024":
+        from .iresnet1024 import iresnet1024
+        return iresnet1024(False, **kwargs)
+    else:
+        raise ValueError()
diff --git a/recognition/vpl/backbones/iresnet.py b/recognition/vpl/backbones/iresnet.py
new file mode 100644
index 000000000..c6d3b9c24
--- /dev/null
+++ b/recognition/vpl/backbones/iresnet.py
@@ -0,0 +1,187 @@
+import torch
+from torch import nn
+
+__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
+                    progress, **kwargs)
+
diff --git a/recognition/vpl/backbones/iresnet1024.py b/recognition/vpl/backbones/iresnet1024.py
new file mode 100644
index 000000000..c0ed0a9e2
--- /dev/null
+++ b/recognition/vpl/backbones/iresnet1024.py
@@ -0,0 +1,171 @@
+import torch
+from torch import nn
+
+assert torch.__version__ >= "1.8.1"
+from torch.utils.checkpoint import checkpoint_sequential
+
+__all__ = ['iresnet1024']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05, )
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05, )
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = checkpoint_sequential(self.layer1, 10, x)
+            x = checkpoint_sequential(self.layer2, 10, x)
+            x = checkpoint_sequential(self.layer3, 10, x)
+            x = checkpoint_sequential(self.layer4, 10, x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet1024(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet1024', IBasicBlock, [256, 256, 512, 256], pretrained,
+                    progress, **kwargs)
diff --git a/recognition/vpl/configs/__init__.py b/recognition/vpl/configs/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/vpl/configs/base.py b/recognition/vpl/configs/base.py
new file mode 100644
index 000000000..ba82b214c
--- /dev/null
+++ b/recognition/vpl/configs/base.py
@@ -0,0 +1,19 @@
+from easydict import EasyDict as edict
+
+config = edict()
+config.embedding_size = 512
+config.sample_rate = 1
+config.fp16 = False
+config.tf32 = False
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # when batch size is 512
+config.warmup_epoch = -1
+config.loss = 'arcface'
+config.network = 'r50'
+config.output = None
+config.val_targets = ['lfw', "cfp_fp", "agedb_30"]
+config.vpl = {'start_iters': 8000, 'allowed_delta': 200, 'lambda': 0.15, 'mode': -1, 'momentum': False} #mode==-1 disables vpl
+
+
diff --git a/recognition/vpl/configs/example_ms1mv3.py b/recognition/vpl/configs/example_ms1mv3.py
new file mode 100644
index 000000000..14dcdbe6d
--- /dev/null
+++ b/recognition/vpl/configs/example_ms1mv3.py
@@ -0,0 +1,18 @@
+from easydict import EasyDict as edict
+
+config = edict()
+config.dataset = "ms1mv3"
+config.fp16 = True
+config.batch_size = 128
+config.vpl = {'start_iters': 8000, 'allowed_delta': 200, 'lambda': 0.15, 'mode': 0, 'momentum': False}
+
+config.rec = "/train_tmp/ms1m-retinaface-t1"
+config.num_classes = 93431
+config.num_image = 5179510
+config.num_epoch = 25
+
+def lr_step_func(epoch):
+    return ((epoch + 1) / (4 + 1)) ** 2 if epoch < -1 else 0.1 ** len(
+        [m for m in [11, 17, 22] if m - 1 <= epoch])
+config.lr_func = lr_step_func
+
diff --git a/recognition/vpl/dataset.py b/recognition/vpl/dataset.py
new file mode 100644
index 000000000..24d3f632b
--- /dev/null
+++ b/recognition/vpl/dataset.py
@@ -0,0 +1,113 @@
+import numbers
+import os
+import queue as Queue
+import threading
+
+import mxnet as mx
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank,
+                                                 non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+class MXFaceDataset(Dataset):
+    def __init__(self, root_dir, local_rank):
+        super(MXFaceDataset, self).__init__()
+        self.transform = transforms.Compose(
+            [transforms.ToPILImage(),
+             transforms.RandomHorizontalFlip(),
+             transforms.ToTensor(),
+             transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+             ])
+        self.root_dir = root_dir
+        self.local_rank = local_rank
+        path_imgrec = os.path.join(root_dir, 'train.rec')
+        path_imgidx = os.path.join(root_dir, 'train.idx')
+        self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+        s = self.imgrec.read_idx(0)
+        header, _ = mx.recordio.unpack(s)
+        #print('ddd1')
+        if header.flag > 0:
+            if len(header.label)==2:
+                self.header0 = (int(header.label[0]), int(header.label[1]))
+                self.imgidx = np.array(range(1, int(header.label[0])))
+            else:
+                self.imgidx = np.array(list(self.imgrec.keys))
+        else:
+            self.imgidx = np.array(list(self.imgrec.keys))
+        #print('ddd2')
+
+    def __getitem__(self, index):
+        idx = self.imgidx[index]
+        s = self.imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        label = header.label
+        if not isinstance(label, numbers.Number):
+            label = label[0]
+        sample = mx.image.imdecode(img).asnumpy()
+        label = torch.tensor(label, dtype=torch.long)
+        if self.transform is not None:
+            sample = self.transform(sample)
+        return sample, label
+
+    def __len__(self):
+        return len(self.imgidx)
+
diff --git a/recognition/vpl/eval/__init__.py b/recognition/vpl/eval/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/vpl/eval/verification.py b/recognition/vpl/eval/verification.py
new file mode 100644
index 000000000..edacf8d81
--- /dev/null
+++ b/recognition/vpl/eval/verification.py
@@ -0,0 +1,409 @@
+"""Helper for evaluation on the Labeled Faces in the Wild dataset 
+"""
+
+# MIT License
+#
+# Copyright (c) 2016 David Sandberg
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+import datetime
+import os
+import pickle
+
+import mxnet as mx
+import numpy as np
+import sklearn
+import torch
+from mxnet import ndarray as nd
+from scipy import interpolate
+from sklearn.decomposition import PCA
+from sklearn.model_selection import KFold
+
+
+class LFold:
+    def __init__(self, n_splits=2, shuffle=False):
+        self.n_splits = n_splits
+        if self.n_splits > 1:
+            self.k_fold = KFold(n_splits=n_splits, shuffle=shuffle)
+
+    def split(self, indices):
+        if self.n_splits > 1:
+            return self.k_fold.split(indices)
+        else:
+            return [(indices, indices)]
+
+
+def calculate_roc(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  nrof_folds=10,
+                  pca=0):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    tprs = np.zeros((nrof_folds, nrof_thresholds))
+    fprs = np.zeros((nrof_folds, nrof_thresholds))
+    accuracy = np.zeros((nrof_folds))
+    indices = np.arange(nrof_pairs)
+
+    if pca == 0:
+        diff = np.subtract(embeddings1, embeddings2)
+        dist = np.sum(np.square(diff), 1)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+        if pca > 0:
+            print('doing pca on', fold_idx)
+            embed1_train = embeddings1[train_set]
+            embed2_train = embeddings2[train_set]
+            _embed_train = np.concatenate((embed1_train, embed2_train), axis=0)
+            pca_model = PCA(n_components=pca)
+            pca_model.fit(_embed_train)
+            embed1 = pca_model.transform(embeddings1)
+            embed2 = pca_model.transform(embeddings2)
+            embed1 = sklearn.preprocessing.normalize(embed1)
+            embed2 = sklearn.preprocessing.normalize(embed2)
+            diff = np.subtract(embed1, embed2)
+            dist = np.sum(np.square(diff), 1)
+
+        # Find the best threshold for the fold
+        acc_train = np.zeros((nrof_thresholds))
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, _, acc_train[threshold_idx] = calculate_accuracy(
+                threshold, dist[train_set], actual_issame[train_set])
+        best_threshold_index = np.argmax(acc_train)
+        for threshold_idx, threshold in enumerate(thresholds):
+            tprs[fold_idx, threshold_idx], fprs[fold_idx, threshold_idx], _ = calculate_accuracy(
+                threshold, dist[test_set],
+                actual_issame[test_set])
+        _, _, accuracy[fold_idx] = calculate_accuracy(
+            thresholds[best_threshold_index], dist[test_set],
+            actual_issame[test_set])
+
+    tpr = np.mean(tprs, 0)
+    fpr = np.mean(fprs, 0)
+    return tpr, fpr, accuracy
+
+
+def calculate_accuracy(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    tp = np.sum(np.logical_and(predict_issame, actual_issame))
+    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    tn = np.sum(
+        np.logical_and(np.logical_not(predict_issame),
+                       np.logical_not(actual_issame)))
+    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
+
+    tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
+    fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
+    acc = float(tp + tn) / dist.size
+    return tpr, fpr, acc
+
+
+def calculate_val(thresholds,
+                  embeddings1,
+                  embeddings2,
+                  actual_issame,
+                  far_target,
+                  nrof_folds=10):
+    assert (embeddings1.shape[0] == embeddings2.shape[0])
+    assert (embeddings1.shape[1] == embeddings2.shape[1])
+    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+    nrof_thresholds = len(thresholds)
+    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
+
+    val = np.zeros(nrof_folds)
+    far = np.zeros(nrof_folds)
+
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff), 1)
+    indices = np.arange(nrof_pairs)
+
+    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+
+        # Find the threshold that gives FAR = far_target
+        far_train = np.zeros(nrof_thresholds)
+        for threshold_idx, threshold in enumerate(thresholds):
+            _, far_train[threshold_idx] = calculate_val_far(
+                threshold, dist[train_set], actual_issame[train_set])
+        if np.max(far_train) >= far_target:
+            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
+            threshold = f(far_target)
+        else:
+            threshold = 0.0
+
+        val[fold_idx], far[fold_idx] = calculate_val_far(
+            threshold, dist[test_set], actual_issame[test_set])
+
+    val_mean = np.mean(val)
+    far_mean = np.mean(far)
+    val_std = np.std(val)
+    return val_mean, val_std, far_mean
+
+
+def calculate_val_far(threshold, dist, actual_issame):
+    predict_issame = np.less(dist, threshold)
+    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
+    false_accept = np.sum(
+        np.logical_and(predict_issame, np.logical_not(actual_issame)))
+    n_same = np.sum(actual_issame)
+    n_diff = np.sum(np.logical_not(actual_issame))
+    # print(true_accept, false_accept)
+    # print(n_same, n_diff)
+    val = float(true_accept) / float(n_same)
+    far = float(false_accept) / float(n_diff)
+    return val, far
+
+
+def evaluate(embeddings, actual_issame, nrof_folds=10, pca=0):
+    # Calculate evaluation metrics
+    thresholds = np.arange(0, 4, 0.01)
+    embeddings1 = embeddings[0::2]
+    embeddings2 = embeddings[1::2]
+    tpr, fpr, accuracy = calculate_roc(thresholds,
+                                       embeddings1,
+                                       embeddings2,
+                                       np.asarray(actual_issame),
+                                       nrof_folds=nrof_folds,
+                                       pca=pca)
+    thresholds = np.arange(0, 4, 0.001)
+    val, val_std, far = calculate_val(thresholds,
+                                      embeddings1,
+                                      embeddings2,
+                                      np.asarray(actual_issame),
+                                      1e-3,
+                                      nrof_folds=nrof_folds)
+    return tpr, fpr, accuracy, val, val_std, far
+
+@torch.no_grad()
+def load_bin(path, image_size):
+    try:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f)  # py2
+    except UnicodeDecodeError as e:
+        with open(path, 'rb') as f:
+            bins, issame_list = pickle.load(f, encoding='bytes')  # py3
+    data_list = []
+    for flip in [0, 1]:
+        data = torch.empty((len(issame_list) * 2, 3, image_size[0], image_size[1]))
+        data_list.append(data)
+    for idx in range(len(issame_list) * 2):
+        _bin = bins[idx]
+        img = mx.image.imdecode(_bin)
+        if img.shape[1] != image_size[0]:
+            img = mx.image.resize_short(img, image_size[0])
+        img = nd.transpose(img, axes=(2, 0, 1))
+        for flip in [0, 1]:
+            if flip == 1:
+                img = mx.ndarray.flip(data=img, axis=2)
+            data_list[flip][idx][:] = torch.from_numpy(img.asnumpy())
+        if idx % 1000 == 0:
+            print('loading bin', idx)
+    print(data_list[0].shape)
+    return data_list, issame_list
+
+@torch.no_grad()
+def test(data_set, backbone, batch_size, nfolds=10):
+    print('testing verification..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+            _data = data[bb - batch_size: bb]
+            time0 = datetime.datetime.now()
+            img = ((_data / 255) - 0.5) / 0.5
+            net_out: torch.Tensor = backbone(img)
+            _embeddings = net_out.detach().cpu().numpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+
+    _xnorm = 0.0
+    _xnorm_cnt = 0
+    for embed in embeddings_list:
+        for i in range(embed.shape[0]):
+            _em = embed[i]
+            _norm = np.linalg.norm(_em)
+            _xnorm += _norm
+            _xnorm_cnt += 1
+    _xnorm /= _xnorm_cnt
+
+    embeddings = embeddings_list[0].copy()
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    acc1 = 0.0
+    std1 = 0.0
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    print(embeddings.shape)
+    print('infer time', time_consumed)
+    _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
+    acc2, std2 = np.mean(accuracy), np.std(accuracy)
+    return acc1, std1, acc2, std2, _xnorm, embeddings_list
+
+
+def dumpR(data_set,
+          backbone,
+          batch_size,
+          name='',
+          data_extra=None,
+          label_shape=None):
+    print('dump verification embedding..')
+    data_list = data_set[0]
+    issame_list = data_set[1]
+    embeddings_list = []
+    time_consumed = 0.0
+    for i in range(len(data_list)):
+        data = data_list[i]
+        embeddings = None
+        ba = 0
+        while ba < data.shape[0]:
+            bb = min(ba + batch_size, data.shape[0])
+            count = bb - ba
+
+            _data = nd.slice_axis(data, axis=0, begin=bb - batch_size, end=bb)
+            time0 = datetime.datetime.now()
+            if data_extra is None:
+                db = mx.io.DataBatch(data=(_data,), label=(_label,))
+            else:
+                db = mx.io.DataBatch(data=(_data, _data_extra),
+                                     label=(_label,))
+            model.forward(db, is_train=False)
+            net_out = model.get_outputs()
+            _embeddings = net_out[0].asnumpy()
+            time_now = datetime.datetime.now()
+            diff = time_now - time0
+            time_consumed += diff.total_seconds()
+            if embeddings is None:
+                embeddings = np.zeros((data.shape[0], _embeddings.shape[1]))
+            embeddings[ba:bb, :] = _embeddings[(batch_size - count):, :]
+            ba = bb
+        embeddings_list.append(embeddings)
+    embeddings = embeddings_list[0] + embeddings_list[1]
+    embeddings = sklearn.preprocessing.normalize(embeddings)
+    actual_issame = np.asarray(issame_list)
+    outname = os.path.join('temp.bin')
+    with open(outname, 'wb') as f:
+        pickle.dump((embeddings, issame_list),
+                    f,
+                    protocol=pickle.HIGHEST_PROTOCOL)
+
+
+# if __name__ == '__main__':
+#
+#     parser = argparse.ArgumentParser(description='do verification')
+#     # general
+#     parser.add_argument('--data-dir', default='', help='')
+#     parser.add_argument('--model',
+#                         default='../model/softmax,50',
+#                         help='path to load model.')
+#     parser.add_argument('--target',
+#                         default='lfw,cfp_ff,cfp_fp,agedb_30',
+#                         help='test targets.')
+#     parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+#     parser.add_argument('--batch-size', default=32, type=int, help='')
+#     parser.add_argument('--max', default='', type=str, help='')
+#     parser.add_argument('--mode', default=0, type=int, help='')
+#     parser.add_argument('--nfolds', default=10, type=int, help='')
+#     args = parser.parse_args()
+#     image_size = [112, 112]
+#     print('image_size', image_size)
+#     ctx = mx.gpu(args.gpu)
+#     nets = []
+#     vec = args.model.split(',')
+#     prefix = args.model.split(',')[0]
+#     epochs = []
+#     if len(vec) == 1:
+#         pdir = os.path.dirname(prefix)
+#         for fname in os.listdir(pdir):
+#             if not fname.endswith('.params'):
+#                 continue
+#             _file = os.path.join(pdir, fname)
+#             if _file.startswith(prefix):
+#                 epoch = int(fname.split('.')[0].split('-')[1])
+#                 epochs.append(epoch)
+#         epochs = sorted(epochs, reverse=True)
+#         if len(args.max) > 0:
+#             _max = [int(x) for x in args.max.split(',')]
+#             assert len(_max) == 2
+#             if len(epochs) > _max[1]:
+#                 epochs = epochs[_max[0]:_max[1]]
+#
+#     else:
+#         epochs = [int(x) for x in vec[1].split('|')]
+#     print('model number', len(epochs))
+#     time0 = datetime.datetime.now()
+#     for epoch in epochs:
+#         print('loading', prefix, epoch)
+#         sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+#         # arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
+#         all_layers = sym.get_internals()
+#         sym = all_layers['fc1_output']
+#         model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
+#         # model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
+#         model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0],
+#                                           image_size[1]))])
+#         model.set_params(arg_params, aux_params)
+#         nets.append(model)
+#     time_now = datetime.datetime.now()
+#     diff = time_now - time0
+#     print('model loading time', diff.total_seconds())
+#
+#     ver_list = []
+#     ver_name_list = []
+#     for name in args.target.split(','):
+#         path = os.path.join(args.data_dir, name + ".bin")
+#         if os.path.exists(path):
+#             print('loading.. ', name)
+#             data_set = load_bin(path, image_size)
+#             ver_list.append(data_set)
+#             ver_name_list.append(name)
+#
+#     if args.mode == 0:
+#         for i in range(len(ver_list)):
+#             results = []
+#             for model in nets:
+#                 acc1, std1, acc2, std2, xnorm, embeddings_list = test(
+#                     ver_list[i], model, args.batch_size, args.nfolds)
+#                 print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
+#                 print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
+#                 print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
+#                 results.append(acc2)
+#             print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
+#     elif args.mode == 1:
+#         raise ValueError
+#     else:
+#         model = nets[0]
+#         dumpR(ver_list[0], model, args.batch_size, args.target)
diff --git a/recognition/vpl/eval_ijbc.py b/recognition/vpl/eval_ijbc.py
new file mode 100644
index 000000000..c144e4e8e
--- /dev/null
+++ b/recognition/vpl/eval_ijbc.py
@@ -0,0 +1,483 @@
+# coding: utf-8
+
+import os
+import pickle
+
+import matplotlib
+import pandas as pd
+
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import timeit
+import sklearn
+import argparse
+from sklearn.metrics import roc_curve, auc
+
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from pathlib import Path
+import sys
+import warnings
+
+sys.path.insert(0, "../")
+warnings.filterwarnings("ignore")
+
+parser = argparse.ArgumentParser(description='do ijb test')
+# general
+parser.add_argument('--model-prefix', default='', help='path to load model.')
+parser.add_argument('--image-path', default='', type=str, help='')
+parser.add_argument('--result-dir', default='.', type=str, help='')
+parser.add_argument('--batch-size', default=128, type=int, help='')
+parser.add_argument('--network', default='iresnet50', type=str, help='')
+parser.add_argument('--job', default='insightface', type=str, help='job name')
+parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+args = parser.parse_args()
+
+target = args.target
+model_path = args.model_prefix
+image_path = args.image_path
+result_dir = args.result_dir
+gpu_id = None
+use_norm_score = True  # if Ture, TestMode(N1)
+use_detector_score = True  # if Ture, TestMode(D1)
+use_flip_test = True  # if Ture, TestMode(F1)
+job = args.job
+batch_size = args.batch_size
+
+import cv2
+import numpy as np
+import torch
+from skimage import transform as trans
+import backbones
+
+
+class Embedding(object):
+    def __init__(self, prefix, data_shape, batch_size=1):
+        image_size = (112, 112)
+        self.image_size = image_size
+        weight = torch.load(prefix)
+        resnet = eval("backbones.{}".format(args.network))(False).cuda()
+        resnet.load_state_dict(weight)
+        model = torch.nn.DataParallel(resnet)
+        self.model = model
+        self.model.eval()
+        src = np.array([
+            [30.2946, 51.6963],
+            [65.5318, 51.5014],
+            [48.0252, 71.7366],
+            [33.5493, 92.3655],
+            [62.7299, 92.2041]], dtype=np.float32)
+        src[:, 0] += 8.0
+        self.src = src
+        self.batch_size = batch_size
+        self.data_shape = data_shape
+
+    def get(self, rimg, landmark):
+
+        assert landmark.shape[0] == 68 or landmark.shape[0] == 5
+        assert landmark.shape[1] == 2
+        if landmark.shape[0] == 68:
+            landmark5 = np.zeros((5, 2), dtype=np.float32)
+            landmark5[0] = (landmark[36] + landmark[39]) / 2
+            landmark5[1] = (landmark[42] + landmark[45]) / 2
+            landmark5[2] = landmark[30]
+            landmark5[3] = landmark[48]
+            landmark5[4] = landmark[54]
+        else:
+            landmark5 = landmark
+        tform = trans.SimilarityTransform()
+        tform.estimate(landmark5, self.src)
+        M = tform.params[0:2, :]
+        img = cv2.warpAffine(rimg,
+                             M, (self.image_size[1], self.image_size[0]),
+                             borderValue=0.0)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img_flip = np.fliplr(img)
+        img = np.transpose(img, (2, 0, 1))  # 3*112*112, RGB
+        img_flip = np.transpose(img_flip, (2, 0, 1))
+        input_blob = np.zeros((2, 3, self.image_size[1], self.image_size[0]), dtype=np.uint8)
+        input_blob[0] = img
+        input_blob[1] = img_flip
+        return input_blob
+
+    @torch.no_grad()
+    def forward_db(self, batch_data):
+        imgs = torch.Tensor(batch_data).cuda()
+        imgs.div_(255).sub_(0.5).div_(0.5)
+        feat = self.model(imgs)
+        feat = feat.reshape([self.batch_size, 2 * feat.shape[1]])
+        return feat.cpu().numpy()
+
+
+# 将一个list尽量均分成n份，限制len(list)==n，份数大于原list内元素个数则分配空list[]
+def divideIntoNstrand(listTemp, n):
+    twoList = [[] for i in range(n)]
+    for i, e in enumerate(listTemp):
+        twoList[i % n].append(e)
+    return twoList
+
+
+def read_template_media_list(path):
+    # ijb_meta = np.loadtxt(path, dtype=str)
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+# In[ ]:
+
+
+def read_template_pair_list(path):
+    # pairs = np.loadtxt(path, dtype=str)
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    # print(pairs.shape)
+    # print(pairs[:, 0].astype(np.int))
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+# In[ ]:
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# In[ ]:
+
+
+def get_image_feature(img_path, files_list, model_path, epoch, gpu_id):
+    batch_size = args.batch_size
+    data_shape = (3, 112, 112)
+
+    files = files_list
+    print('files:', len(files))
+    rare_size = len(files) % batch_size
+    faceness_scores = []
+    batch = 0
+    img_feats = np.empty((len(files), 1024), dtype=np.float32)
+
+    batch_data = np.empty((2 * batch_size, 3, 112, 112))
+    embedding = Embedding(model_path, data_shape, batch_size)
+    for img_index, each_line in enumerate(files[:len(files) - rare_size]):
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        input_blob = embedding.get(img, lmk)
+
+        batch_data[2 * (img_index - batch * batch_size)][:] = input_blob[0]
+        batch_data[2 * (img_index - batch * batch_size) + 1][:] = input_blob[1]
+        if (img_index + 1) % batch_size == 0:
+            print('batch', batch)
+            img_feats[batch * batch_size:batch * batch_size +
+                                         batch_size][:] = embedding.forward_db(batch_data)
+            batch += 1
+        faceness_scores.append(name_lmk_score[-1])
+
+    batch_data = np.empty((2 * rare_size, 3, 112, 112))
+    embedding = Embedding(model_path, data_shape, rare_size)
+    for img_index, each_line in enumerate(files[len(files) - rare_size:]):
+        name_lmk_score = each_line.strip().split(' ')
+        img_name = os.path.join(img_path, name_lmk_score[0])
+        img = cv2.imread(img_name)
+        lmk = np.array([float(x) for x in name_lmk_score[1:-1]],
+                       dtype=np.float32)
+        lmk = lmk.reshape((5, 2))
+        input_blob = embedding.get(img, lmk)
+        batch_data[2 * img_index][:] = input_blob[0]
+        batch_data[2 * img_index + 1][:] = input_blob[1]
+        if (img_index + 1) % rare_size == 0:
+            print('batch', batch)
+            img_feats[len(files) -
+                      rare_size:][:] = embedding.forward_db(batch_data)
+            batch += 1
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    # img_feats = np.ones( (len(files), 1024), dtype=np.float32) * 0.01
+    # faceness_scores = np.ones( (len(files), ), dtype=np.float32 )
+    return img_feats, faceness_scores
+
+
+# In[ ]:
+
+
+def image2template_feature(img_feats=None, templates=None, medias=None):
+    # ==========================================================
+    # 1. face image feature l2 normalization. img_feats:[number_image x feats_dim]
+    # 2. compute media feature.
+    # 3. compute template feature.
+    # ==========================================================
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+
+    for count_template, uqt in enumerate(unique_templates):
+
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias,
+                                                       return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [
+                    np.mean(face_norm_feats[ind_m], axis=0, keepdims=True)
+                ]
+        media_norm_feats = np.array(media_norm_feats)
+        # media_norm_feats = media_norm_feats / np.sqrt(np.sum(media_norm_feats ** 2, -1, keepdims=True))
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    # template_norm_feats = template_feats / np.sqrt(np.sum(template_feats ** 2, -1, keepdims=True))
+    template_norm_feats = sklearn.preprocessing.normalize(template_feats)
+    # print(template_norm_feats.shape)
+    return template_norm_feats, unique_templates
+
+
+# In[ ]:
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    # ==========================================================
+    #         Compute set-to-set Similarity Score.
+    # ==========================================================
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+# In[ ]:
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [
+        total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)
+    ]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def read_score(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+# # Step1: Load Meta Data
+
+# In[ ]:
+
+assert target == 'IJBC' or target == 'IJBB'
+
+# =============================================================
+# load image and template relationships for template feature embedding
+# tid --> template id,  mid --> media id
+# format:
+#           image_name tid mid
+# =============================================================
+start = timeit.default_timer()
+templates, medias = read_template_media_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_face_tid_mid.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+
+# =============================================================
+# load template pairs for template-to-template verification
+# tid : template id,  label : 1/0
+# format:
+#           tid_1 tid_2 label
+# =============================================================
+start = timeit.default_timer()
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % target.lower()))
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 2: Get Image Features
+
+# In[ ]:
+
+# =============================================================
+# load image features
+# format:
+#           img_feats: [image_num x feats_dim] (227630, 512)
+# =============================================================
+start = timeit.default_timer()
+img_path = '%s/loose_crop' % image_path
+img_list_path = '%s/meta/%s_name_5pts_score.txt' % (image_path, target.lower())
+img_list = open(img_list_path)
+files = img_list.readlines()
+# files_list = divideIntoNstrand(files, rank_size)
+files_list = files
+
+# img_feats
+# for i in range(rank_size):
+img_feats, faceness_scores = get_image_feature(img_path, files_list,
+                                               model_path, 0, gpu_id)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0],
+                                          img_feats.shape[1]))
+
+# # Step3: Get Template Features
+
+# In[ ]:
+
+# =============================================================
+# compute template features from image features.
+# =============================================================
+start = timeit.default_timer()
+# ==========================================================
+# Norm feature before aggregation into template feature?
+# Feature norm from embedding network and faceness score are able to decrease weights for noise samples (not face).
+# ==========================================================
+# 1. FaceScore （Feature Norm）
+# 2. FaceScore （Detector）
+
+if use_flip_test:
+    # concat --- F1
+    # img_input_feats = img_feats
+    # add --- F2
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] //
+                                     2] + img_feats[:, img_feats.shape[1] // 2:]
+else:
+    img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+if use_norm_score:
+    img_input_feats = img_input_feats
+else:
+    # normalise features to remove norm information
+    img_input_feats = img_input_feats / np.sqrt(
+        np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+if use_detector_score:
+    print(img_input_feats.shape, faceness_scores.shape)
+    img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+else:
+    img_input_feats = img_input_feats
+
+template_norm_feats, unique_templates = image2template_feature(
+    img_input_feats, templates, medias)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# # Step 4: Get Template Similarity Scores
+
+# In[ ]:
+
+# =============================================================
+# compute verification scores between template pairs.
+# =============================================================
+start = timeit.default_timer()
+score = verification(template_norm_feats, unique_templates, p1, p2)
+stop = timeit.default_timer()
+print('Time: %.2f s. ' % (stop - start))
+
+# In[ ]:
+save_path = os.path.join(result_dir, args.job)
+# save_path = result_dir + '/%s_result' % target
+
+if not os.path.exists(save_path):
+    os.makedirs(save_path)
+
+score_save_file = os.path.join(save_path, "%s.npy" % target.lower())
+np.save(score_save_file, score)
+
+# # Step 5: Get ROC Curves and TPR@FPR Table
+
+# In[ ]:
+
+files = [score_save_file]
+methods = []
+scores = []
+for file in files:
+    methods.append(Path(file).stem)
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append("%s-%s" % (method, target))
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10 ** -6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+fig.savefig(os.path.join(save_path, '%s.pdf' % target.lower()))
+print(tpr_fpr_table)
diff --git a/recognition/vpl/inference.py b/recognition/vpl/inference.py
new file mode 100644
index 000000000..839c57039
--- /dev/null
+++ b/recognition/vpl/inference.py
@@ -0,0 +1,34 @@
+import argparse
+
+import cv2
+import numpy as np
+import torch
+
+from backbones import get_model
+
+
+@torch.no_grad()
+def inference(weight, name, img):
+    if img is None:
+        img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.uint8)
+    else:
+        img = cv2.imread(img)
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = np.transpose(img, (2, 0, 1))
+    img = torch.from_numpy(img).unsqueeze(0).float()
+    img.div_(255).sub_(0.5).div_(0.5)
+    net = get_model(name, fp16=False)
+    net.load_state_dict(torch.load(weight))
+    net.eval()
+    feat = net(img).numpy()
+    print(feat)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='PyTorch ArcFace Training')
+    parser.add_argument('--network', type=str, default='r50', help='backbone network')
+    parser.add_argument('--weight', type=str, default='')
+    parser.add_argument('--img', type=str, default=None)
+    args = parser.parse_args()
+    inference(args.weight, args.network, args.img)
diff --git a/recognition/vpl/losses.py b/recognition/vpl/losses.py
new file mode 100644
index 000000000..12caa345a
--- /dev/null
+++ b/recognition/vpl/losses.py
@@ -0,0 +1,40 @@
+import torch
+from torch import nn
+
+def get_loss(name):
+    if name == "cosface":
+        return CosFace()
+    elif name == "arcface":
+        return ArcFace()
+    else:
+        raise ValueError()
+
+class CosFace(nn.Module):
+    def __init__(self, s=64.0, m=0.40):
+        super(CosFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, cosine, label):
+        index = torch.where(label != -1)[0]
+        m_hot = torch.zeros(index.size()[0], cosine.size()[1], device=cosine.device)
+        m_hot.scatter_(1, label[index, None], self.m)
+        cosine[index] -= m_hot
+        ret = cosine * self.s
+        return ret
+
+
+class ArcFace(nn.Module):
+    def __init__(self, s=64.0, m=0.5):
+        super(ArcFace, self).__init__()
+        self.s = s
+        self.m = m
+
+    def forward(self, cosine: torch.Tensor, label):
+        index = torch.where(label != -1)[0]
+        m_hot = torch.zeros(index.size()[0], cosine.size()[1], device=cosine.device)
+        m_hot.scatter_(1, label[index, None], self.m)
+        cosine.acos_()
+        cosine[index] += m_hot
+        cosine.cos_().mul_(self.s)
+        return cosine
diff --git a/recognition/vpl/onnx_helper.py b/recognition/vpl/onnx_helper.py
new file mode 100644
index 000000000..3b4c47c6e
--- /dev/null
+++ b/recognition/vpl/onnx_helper.py
@@ -0,0 +1,199 @@
+import argparse
+import datetime
+import os
+import os.path as osp
+
+import cv2
+import numpy as np
+import onnx
+import onnxruntime
+from onnx import numpy_helper
+
+
+class ArcFaceORT:
+    def __init__(self, model_path):
+        self.model_path = model_path
+
+    def check(self, test_img=None):
+        max_model_size_mb = 1024
+        max_feat_dim = 512
+        max_time_cost = 15
+
+        if not os.path.exists(self.model_path):
+            return "model_path not exists"
+        if not os.path.isdir(self.model_path):
+            return "model_path should be directory"
+        onnx_files = []
+        for _file in os.listdir(self.model_path):
+            print('file_:', _file)
+            if _file.endswith('.onnx'):
+                onnx_files.append(osp.join(self.model_path, _file))
+        if len(onnx_files) == 0:
+            return "do not have onnx files"
+        self.model_file = sorted(onnx_files)[-1]
+        print('use onnx-model:', self.model_file)
+        try:
+            session = onnxruntime.InferenceSession(self.model_file, None)
+        except:
+            return "load onnx failed"
+
+        input_cfg = session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        print('input-shape:', input_shape)
+        if len(input_shape) != 4:
+            return "length of input_shape should be 4"
+        if not isinstance(input_shape[0], str):
+            # return "input_shape[0] should be str to support batch-inference"
+            print('reset input-shape[0] to None')
+            model = onnx.load(self.model_file)
+            model.graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+            new_model_file = osp.join(self.model_path, 'zzzzrefined.onnx')
+            onnx.save(model, new_model_file)
+            self.model_file = new_model_file
+            print('use new onnx-model:', self.model_file)
+            try:
+                session = onnxruntime.InferenceSession(self.model_file, None)
+            except:
+                return "load onnx failed"
+
+            input_cfg = session.get_inputs()[0]
+            input_shape = input_cfg.shape
+            print('new-input-shape:', input_shape)
+
+        self.image_size = tuple(input_shape[2:4][::-1])
+        # print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        outputs = session.get_outputs()
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+            # print(o.name, o.shape)
+        if len(output_names) != 1:
+            return "number of output nodes should be 1"
+        self.session = session
+        self.input_name = input_name
+        self.output_names = output_names
+        # print(self.output_names)
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        if len(graph.node) < 8:
+            return "too small onnx graph"
+
+        input_size = (112, 112)
+        self.crop = None
+        if True:
+            crop_file = osp.join(self.model_path, 'crop.txt')
+            if osp.exists(crop_file):
+                lines = open(crop_file, 'r').readlines()
+                if len(lines) != 6:
+                    return "crop.txt should contain 6 lines"
+                lines = [int(x) for x in lines]
+                self.crop = lines[:4]
+                input_size = tuple(lines[4:6])
+        if input_size != self.image_size:
+            return "input-size is inconsistant with onnx model input, %s vs %s" % (input_size, self.image_size)
+
+        self.model_size_mb = os.path.getsize(self.model_file) / float(1024 * 1024)
+        if self.model_size_mb > max_model_size_mb:
+            return "max model size exceed, given %.3f-MB" % self.model_size_mb
+
+        input_mean = None
+        input_std = None
+        if True:
+            pn_file = osp.join(self.model_path, 'pixel_norm.txt')
+            if osp.exists(pn_file):
+                lines = open(pn_file, 'r').readlines()
+                if len(lines) != 2:
+                    return "pixel_norm.txt should contain 2 lines"
+                input_mean = float(lines[0])
+                input_std = float(lines[1])
+        if input_mean is not None or input_std is not None:
+            if input_mean is None or input_std is None:
+                return "please set input_mean and input_std simultaneously"
+        else:
+            find_sub = False
+            find_mul = False
+            for nid, node in enumerate(graph.node[:8]):
+                print(nid, node.name)
+                if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                    find_sub = True
+                if node.name.startswith('Mul') or node.name.startswith('_mul'):
+                    find_mul = True
+            if find_sub and find_mul:
+                # mxnet arcface model
+                input_mean = 0.0
+                input_std = 1.0
+            else:
+                input_mean = 127.5
+                input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        for initn in graph.initializer:
+            weight_array = numpy_helper.to_array(initn)
+
+            dt = weight_array.dtype
+            if dt.itemsize < 4:
+                return 'invalid weight type - (%s:%s)' % (initn.name, dt.name)
+        if test_img is None:
+            test_img = np.random.randint(0, 255, size=(self.image_size[1], self.image_size[0], 3), dtype=np.uint8)
+        else:
+            test_img = cv2.resize(test_img, self.image_size)
+        feat, cost = self.benchmark(test_img)
+        if feat.shape[1] > max_feat_dim:
+            return "max feat dim exceed, given %d" % feat.shape[1]
+        self.feat_dim = feat.shape[1]
+        cost_ms = cost * 1000
+        if cost_ms > max_time_cost:
+            return "max time cost exceed, given %.4f" % cost_ms
+        self.cost_ms = cost_ms
+        print(
+            'check stat:, model-size-mb: %.4f, feat-dim: %d, time-cost-ms: %.4f, input-mean: %.3f, input-std: %.3f' % (
+                self.model_size_mb, self.feat_dim, self.cost_ms, self.input_mean, self.input_std))
+        return None
+
+    def meta_info(self):
+        return {'model-size-mb': self.model_size_mb, 'feature-dim': self.feat_dim, 'infer': self.cost_ms}
+
+    def forward(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.image_size
+        if self.crop is not None:
+            nimgs = []
+            for img in imgs:
+                nimg = img[self.crop[1]:self.crop[3], self.crop[0]:self.crop[2], :]
+                if nimg.shape[0] != input_size[1] or nimg.shape[1] != input_size[0]:
+                    nimg = cv2.resize(nimg, input_size)
+                nimgs.append(nimg)
+            imgs = nimgs
+        blob = cv2.dnn.blobFromImages(imgs, 1.0 / self.input_std, input_size,
+                                      (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+    def benchmark(self, img):
+        input_size = self.image_size
+        if self.crop is not None:
+            nimg = img[self.crop[1]:self.crop[3], self.crop[0]:self.crop[2], :]
+            if nimg.shape[0] != input_size[1] or nimg.shape[1] != input_size[0]:
+                nimg = cv2.resize(nimg, input_size)
+            img = nimg
+        blob = cv2.dnn.blobFromImage(img, 1.0 / self.input_std, input_size,
+                                     (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        costs = []
+        for _ in range(50):
+            ta = datetime.datetime.now()
+            net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+            tb = datetime.datetime.now()
+            cost = (tb - ta).total_seconds()
+            costs.append(cost)
+        costs = sorted(costs)
+        cost = costs[5]
+        return net_out, cost
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_root", help="onnx model root, default is './'", default="./")
+    args = parser.parse_args()
+    ArcFaceORT(args.model_root).check()
diff --git a/recognition/vpl/onnx_ijbc.py b/recognition/vpl/onnx_ijbc.py
new file mode 100644
index 000000000..05b50bfad
--- /dev/null
+++ b/recognition/vpl/onnx_ijbc.py
@@ -0,0 +1,267 @@
+import argparse
+import os
+import pickle
+import timeit
+
+import cv2
+import mxnet as mx
+import numpy as np
+import pandas as pd
+import prettytable
+import skimage.transform
+from sklearn.metrics import roc_curve
+from sklearn.preprocessing import normalize
+
+from onnx_helper import ArcFaceORT
+
+SRC = np.array(
+    [
+        [30.2946, 51.6963],
+        [65.5318, 51.5014],
+        [48.0252, 71.7366],
+        [33.5493, 92.3655],
+        [62.7299, 92.2041]]
+    , dtype=np.float32)
+SRC[:, 0] += 8.0
+
+
+class AlignedDataSet(mx.gluon.data.Dataset):
+    def __init__(self, root, lines, align=True):
+        self.lines = lines
+        self.root = root
+        self.align = align
+
+    def __len__(self):
+        return len(self.lines)
+
+    def __getitem__(self, idx):
+        each_line = self.lines[idx]
+        name_lmk_score = each_line.strip().split(' ')
+        name = os.path.join(self.root, name_lmk_score[0])
+        img = cv2.cvtColor(cv2.imread(name), cv2.COLOR_BGR2RGB)
+        landmark5 = np.array([float(x) for x in name_lmk_score[1:-1]], dtype=np.float32).reshape((5, 2))
+        st = skimage.transform.SimilarityTransform()
+        st.estimate(landmark5, SRC)
+        img = cv2.warpAffine(img, st.params[0:2, :], (112, 112), borderValue=0.0)
+        img_1 = np.expand_dims(img, 0)
+        img_2 = np.expand_dims(np.fliplr(img), 0)
+        output = np.concatenate((img_1, img_2), axis=0).astype(np.float32)
+        output = np.transpose(output, (0, 3, 1, 2))
+        output = mx.nd.array(output)
+        return output
+
+
+def extract(model_root, dataset):
+    model = ArcFaceORT(model_path=model_root)
+    model.check()
+    feat_mat = np.zeros(shape=(len(dataset), 2 * model.feat_dim))
+
+    def batchify_fn(data):
+        return mx.nd.concat(*data, dim=0)
+
+    data_loader = mx.gluon.data.DataLoader(
+        dataset, 128, last_batch='keep', num_workers=4,
+        thread_pool=True, prefetch=16, batchify_fn=batchify_fn)
+    num_iter = 0
+    for batch in data_loader:
+        batch = batch.asnumpy()
+        batch = (batch - model.input_mean) / model.input_std
+        feat = model.session.run(model.output_names, {model.input_name: batch})[0]
+        feat = np.reshape(feat, (-1, model.feat_dim * 2))
+        feat_mat[128 * num_iter: 128 * num_iter + feat.shape[0], :] = feat
+        num_iter += 1
+        if num_iter % 50 == 0:
+            print(num_iter)
+    return feat_mat
+
+
+def read_template_media_list(path):
+    ijb_meta = pd.read_csv(path, sep=' ', header=None).values
+    templates = ijb_meta[:, 1].astype(np.int)
+    medias = ijb_meta[:, 2].astype(np.int)
+    return templates, medias
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+def read_image_feature(path):
+    with open(path, 'rb') as fid:
+        img_feats = pickle.load(fid)
+    return img_feats
+
+
+def image2template_feature(img_feats=None,
+                           templates=None,
+                           medias=None):
+    unique_templates = np.unique(templates)
+    template_feats = np.zeros((len(unique_templates), img_feats.shape[1]))
+    for count_template, uqt in enumerate(unique_templates):
+        (ind_t,) = np.where(templates == uqt)
+        face_norm_feats = img_feats[ind_t]
+        face_medias = medias[ind_t]
+        unique_medias, unique_media_counts = np.unique(face_medias, return_counts=True)
+        media_norm_feats = []
+        for u, ct in zip(unique_medias, unique_media_counts):
+            (ind_m,) = np.where(face_medias == u)
+            if ct == 1:
+                media_norm_feats += [face_norm_feats[ind_m]]
+            else:  # image features from the same video will be aggregated into one feature
+                media_norm_feats += [np.mean(face_norm_feats[ind_m], axis=0, keepdims=True), ]
+        media_norm_feats = np.array(media_norm_feats)
+        template_feats[count_template] = np.sum(media_norm_feats, axis=0)
+        if count_template % 2000 == 0:
+            print('Finish Calculating {} template features.'.format(
+                count_template))
+    template_norm_feats = normalize(template_feats)
+    return template_norm_feats, unique_templates
+
+
+def verification(template_norm_feats=None,
+                 unique_templates=None,
+                 p1=None,
+                 p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000
+    sublists = [total_pairs[i: i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def verification2(template_norm_feats=None,
+                  unique_templates=None,
+                  p1=None,
+                  p2=None):
+    template2id = np.zeros((max(unique_templates) + 1, 1), dtype=int)
+    for count_template, uqt in enumerate(unique_templates):
+        template2id[uqt] = count_template
+    score = np.zeros((len(p1),))  # save cosine distance between pairs
+    total_pairs = np.array(range(len(p1)))
+    batchsize = 100000  # small batchsize instead of all pairs in one batch due to the memory limiation
+    sublists = [total_pairs[i:i + batchsize] for i in range(0, len(p1), batchsize)]
+    total_sublists = len(sublists)
+    for c, s in enumerate(sublists):
+        feat1 = template_norm_feats[template2id[p1[s]]]
+        feat2 = template_norm_feats[template2id[p2[s]]]
+        similarity_score = np.sum(feat1 * feat2, -1)
+        score[s] = similarity_score.flatten()
+        if c % 10 == 0:
+            print('Finish {}/{} pairs.'.format(c, total_sublists))
+    return score
+
+
+def main(args):
+    use_norm_score = True  # if Ture, TestMode(N1)
+    use_detector_score = True  # if Ture, TestMode(D1)
+    use_flip_test = True  # if Ture, TestMode(F1)
+    assert args.target == 'IJBC' or args.target == 'IJBB'
+
+    start = timeit.default_timer()
+    templates, medias = read_template_media_list(
+        os.path.join('%s/meta' % args.image_path, '%s_face_tid_mid.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    p1, p2, label = read_template_pair_list(
+        os.path.join('%s/meta' % args.image_path,
+                     '%s_template_pair_label.txt' % args.target.lower()))
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    img_path = '%s/loose_crop' % args.image_path
+    img_list_path = '%s/meta/%s_name_5pts_score.txt' % (args.image_path, args.target.lower())
+    img_list = open(img_list_path)
+    files = img_list.readlines()
+    dataset = AlignedDataSet(root=img_path, lines=files, align=True)
+    img_feats = extract(args.model_root, dataset)
+
+    faceness_scores = []
+    for each_line in files:
+        name_lmk_score = each_line.split()
+        faceness_scores.append(name_lmk_score[-1])
+    faceness_scores = np.array(faceness_scores).astype(np.float32)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    print('Feature Shape: ({} , {}) .'.format(img_feats.shape[0], img_feats.shape[1]))
+    start = timeit.default_timer()
+
+    if use_flip_test:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2] + img_feats[:, img_feats.shape[1] // 2:]
+    else:
+        img_input_feats = img_feats[:, 0:img_feats.shape[1] // 2]
+
+    if use_norm_score:
+        img_input_feats = img_input_feats
+    else:
+        img_input_feats = img_input_feats / np.sqrt(np.sum(img_input_feats ** 2, -1, keepdims=True))
+
+    if use_detector_score:
+        print(img_input_feats.shape, faceness_scores.shape)
+        img_input_feats = img_input_feats * faceness_scores[:, np.newaxis]
+    else:
+        img_input_feats = img_input_feats
+
+    template_norm_feats, unique_templates = image2template_feature(
+        img_input_feats, templates, medias)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+
+    start = timeit.default_timer()
+    score = verification(template_norm_feats, unique_templates, p1, p2)
+    stop = timeit.default_timer()
+    print('Time: %.2f s. ' % (stop - start))
+    save_path = os.path.join(args.result_dir, "{}_result".format(args.target))
+    if not os.path.exists(save_path):
+        os.makedirs(save_path)
+    score_save_file = os.path.join(save_path, "{}.npy".format(args.model_root))
+    np.save(score_save_file, score)
+    files = [score_save_file]
+    methods = []
+    scores = []
+    for file in files:
+        methods.append(os.path.basename(file))
+        scores.append(np.load(file))
+    methods = np.array(methods)
+    scores = dict(zip(methods, scores))
+    x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+    tpr_fpr_table = prettytable.PrettyTable(['Methods'] + [str(x) for x in x_labels])
+    for method in methods:
+        fpr, tpr, _ = roc_curve(label, scores[method])
+        fpr = np.flipud(fpr)
+        tpr = np.flipud(tpr)
+        tpr_fpr_row = []
+        tpr_fpr_row.append("%s-%s" % (method, args.target))
+        for fpr_iter in np.arange(len(x_labels)):
+            _, min_index = min(
+                list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+            tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+        tpr_fpr_table.add_row(tpr_fpr_row)
+    print(tpr_fpr_table)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='do ijb test')
+    # general
+    parser.add_argument('--model-root', default='', help='path to load model.')
+    parser.add_argument('--image-path', default='', type=str, help='')
+    parser.add_argument('--result-dir', default='.', type=str, help='')
+    parser.add_argument('--target', default='IJBC', type=str, help='target, set to IJBC or IJBB')
+    main(parser.parse_args())
diff --git a/recognition/vpl/run.sh b/recognition/vpl/run.sh
new file mode 100644
index 000000000..f86c5acb7
--- /dev/null
+++ b/recognition/vpl/run.sh
@@ -0,0 +1,2 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -u -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py $@
+#ps -ef | grep "train" | grep -v grep | awk '{print "kill -9 "$2}' | sh
diff --git a/recognition/vpl/torch2onnx.py b/recognition/vpl/torch2onnx.py
new file mode 100644
index 000000000..4e7a68e68
--- /dev/null
+++ b/recognition/vpl/torch2onnx.py
@@ -0,0 +1,59 @@
+import numpy as np
+import onnx
+import torch
+
+
+def convert_onnx(net, path_module, output, opset=11, simplify=False):
+    assert isinstance(net, torch.nn.Module)
+    img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.int32)
+    img = img.astype(np.float)
+    img = (img / 255. - 0.5) / 0.5  # torch style norm
+    img = img.transpose((2, 0, 1))
+    img = torch.from_numpy(img).unsqueeze(0).float()
+
+    weight = torch.load(path_module)
+    net.load_state_dict(weight)
+    net.eval()
+    torch.onnx.export(net, img, output, keep_initializers_as_inputs=False, verbose=False, opset_version=opset)
+    model = onnx.load(output)
+    graph = model.graph
+    graph.input[0].type.tensor_type.shape.dim[0].dim_param = 'None'
+    if simplify:
+        from onnxsim import simplify
+        model, check = simplify(model)
+        assert check, "Simplified ONNX model could not be validated"
+    onnx.save(model, output)
+
+    
+if __name__ == '__main__':
+    import os
+    import argparse
+    from backbones import get_model
+
+    parser = argparse.ArgumentParser(description='ArcFace PyTorch to onnx')
+    parser.add_argument('input', type=str, help='input backbone.pth file or path')
+    parser.add_argument('--output', type=str, default=None, help='output onnx path')
+    parser.add_argument('--network', type=str, default=None, help='backbone network')
+    parser.add_argument('--simplify', type=bool, default=True, help='onnx simplify')
+    args = parser.parse_args()
+    input_file = args.input
+    if os.path.isdir(input_file):
+        input_file = os.path.join(input_file, "backbone.pth")
+    assert os.path.exists(input_file)
+    model_name = os.path.basename(os.path.dirname(input_file)).lower()
+    params = model_name.split("_")
+    if len(params) >= 3 and params[1] in ('arcface', 'cosface'):
+        if args.network is None:
+            args.network = params[2]
+    assert args.network is not None
+    print(args)
+    backbone_onnx = get_model(args.network, dropout=0)
+
+    output_path = args.output
+    if output_path is None:
+        output_path = os.path.join(os.path.dirname(__file__), 'onnx')
+    if not os.path.exists(output_path):
+        os.makedirs(output_path)
+    assert os.path.isdir(output_path)
+    output_file = os.path.join(output_path, "%s.onnx" % model_name)
+    convert_onnx(backbone_onnx, input_file, output_file, simplify=args.simplify)
diff --git a/recognition/vpl/train.py b/recognition/vpl/train.py
new file mode 100644
index 000000000..45a594ba8
--- /dev/null
+++ b/recognition/vpl/train.py
@@ -0,0 +1,180 @@
+import argparse
+import logging
+import os
+import time
+
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+import torch.utils.data.distributed
+from torch.nn.utils import clip_grad_norm_
+
+import losses
+from backbones import get_model
+from dataset import MXFaceDataset, DataLoaderX
+from torch.utils.data import DataLoader, Dataset
+from vpl import VPL
+from utils.utils_amp import MaxClipGradScaler
+from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
+from utils.utils_logging import AverageMeter, init_logging
+from utils.utils_dist import concat_all_gather, batch_shuffle_ddp, batch_unshuffle_ddp
+from utils.utils_config import get_config
+
+
+def main(args):
+    cfg = get_config(args.config)
+    if not cfg.tf32:
+        torch.backends.cuda.matmul.allow_tf32 = False
+        torch.backends.cudnn.allow_tf32 = False
+    try:
+        world_size = int(os.environ['WORLD_SIZE'])
+        rank = int(os.environ['RANK'])
+        dist_url = "tcp://{}:{}".format(os.environ["MASTER_ADDR"], os.environ["MASTER_PORT"])
+    except KeyError:
+        world_size = 1
+        rank = 0
+        dist_url = "tcp://127.0.0.1:12584"
+
+    dist.init_process_group(backend='nccl', init_method=dist_url, rank=rank, world_size=world_size)
+    local_rank = args.local_rank
+    torch.cuda.set_device(local_rank)
+
+    if not os.path.exists(cfg.output) and rank==0:
+        os.makedirs(cfg.output)
+    else:
+        time.sleep(2)
+
+    log_root = logging.getLogger()
+    init_logging(log_root, rank, cfg.output)
+    if rank==0:
+        logging.info(args)
+        logging.info(cfg)
+
+    train_set = MXFaceDataset(root_dir=cfg.rec, local_rank=local_rank)
+    train_sampler = torch.utils.data.distributed.DistributedSampler(
+        train_set, shuffle=True)
+    train_loader = DataLoaderX(
+        local_rank=local_rank, dataset=train_set, batch_size=cfg.batch_size,
+        sampler=train_sampler, num_workers=2, pin_memory=True, drop_last=True)
+
+    dropout = 0.4 if cfg.dataset == "webface" else 0
+    backbone = get_model(cfg.network, dropout=dropout, fp16=cfg.fp16).to(local_rank)
+    backbone_onnx = get_model(cfg.network, dropout=dropout, fp16=False)
+
+    if args.resume:
+        try:
+            backbone_pth = os.path.join(cfg.output, "backbone.pth")
+            backbone.load_state_dict(torch.load(backbone_pth, map_location=torch.device(local_rank)))
+            if rank==0:
+                logging.info("backbone resume successfully!")
+        except (FileNotFoundError, KeyError, IndexError, RuntimeError):
+            logging.info("resume fail, backbone init successfully!")
+
+    for ps in backbone.parameters():
+        dist.broadcast(ps, 0)
+    backbone = torch.nn.parallel.DistributedDataParallel(
+        module=backbone, broadcast_buffers=False, device_ids=[local_rank])
+    backbone.train()
+
+    cfg_vpl = cfg.vpl
+    vpl_momentum = cfg_vpl['momentum']
+    if vpl_momentum:
+        backbone_w = get_model(cfg.network, dropout=dropout, fp16=cfg.fp16).to(local_rank)
+        backbone_w.train()
+        for param_b, param_w in zip(backbone.module.parameters(), backbone_w.parameters()):
+            param_w.data.copy_(param_b.data)
+            param_w.requires_grad = False
+
+    margin_softmax = losses.get_loss(cfg.loss)
+    module_fc = VPL(
+        rank=rank, local_rank=local_rank, world_size=world_size, resume=args.resume,
+        batch_size=cfg.batch_size, margin_softmax=margin_softmax, num_classes=cfg.num_classes,
+        sample_rate=cfg.sample_rate, embedding_size=cfg.embedding_size, prefix=cfg.output,
+        cfg = cfg_vpl)
+    #print('AAA')
+
+    opt_backbone = torch.optim.SGD(
+        params=[{'params': backbone.parameters()}],
+        lr=cfg.lr / 512 * cfg.batch_size * world_size,
+        momentum=0.9, weight_decay=cfg.weight_decay)
+    opt_pfc = torch.optim.SGD(
+        params=[{'params': module_fc.parameters()}],
+        lr=cfg.lr / 512 * cfg.batch_size * world_size,
+        momentum=0.9, weight_decay=cfg.weight_decay)
+
+    #print('AAA')
+    scheduler_backbone = torch.optim.lr_scheduler.LambdaLR(
+        optimizer=opt_backbone, lr_lambda=cfg.lr_func)
+    scheduler_pfc = torch.optim.lr_scheduler.LambdaLR(
+        optimizer=opt_pfc, lr_lambda=cfg.lr_func)
+
+    start_epoch = 0
+    total_step = int(len(train_set) / cfg.batch_size / world_size * cfg.num_epoch)
+    if rank==0: logging.info("Total Step is: %d" % total_step)
+
+    #for epoch in range(start_epoch, cfg.num_epoch):
+    #    _lr = cfg.lr_func(epoch)
+    #    logging.info('%d:%f'%(epoch, _lr))
+
+    callback_verification = CallBackVerification(10000, rank, cfg.val_targets, cfg.rec)
+    callback_logging = CallBackLogging(50, rank, total_step, cfg.batch_size, world_size, None)
+    callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output)
+
+    loss = AverageMeter()
+    global_step = 0
+    grad_amp = MaxClipGradScaler(cfg.batch_size, 128 * cfg.batch_size, growth_interval=100) if cfg.fp16 else None
+    use_batch_shuffle = True
+    alpha = 0.999
+    for epoch in range(start_epoch, cfg.num_epoch):
+        train_sampler.set_epoch(epoch)
+        for step, (img, label) in enumerate(train_loader):
+            global_step += 1
+            #img = img.to(memory_format=torch.channels_last)
+            features = F.normalize(backbone(img))
+            feature_w = None
+            if vpl_momentum:
+                with torch.no_grad():
+                    for param_b, param_w in zip(backbone.module.parameters(), backbone_w.parameters()):
+                        param_w.data = param_w.data * alpha + param_b.data * (1. - alpha)
+                    if use_batch_shuffle:
+                        img_w, idx_unshuffle = batch_shuffle_ddp(img, rank, world_size)
+
+                    feature_w = F.normalize(backbone_w(img_w))
+                    if use_batch_shuffle:
+                        feature_w = batch_unshuffle_ddp(feature_w, idx_unshuffle, rank, world_size)
+                    feature_w = feature_w.detach()
+
+            x_grad, loss_v = module_fc.forward_backward(label, features, opt_pfc, feature_w)
+            if cfg.fp16:
+                features.backward(grad_amp.scale(x_grad))
+                grad_amp.unscale_(opt_backbone)
+                clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
+                grad_amp.step(opt_backbone)
+                grad_amp.update()
+            else:
+                features.backward(x_grad)
+                clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
+                opt_backbone.step()
+
+            opt_pfc.step()
+            module_fc.update()
+            opt_backbone.zero_grad()
+            opt_pfc.zero_grad()
+            loss.update(loss_v, 1)
+            callback_logging(global_step, loss, epoch, cfg.fp16, grad_amp)
+            callback_verification(global_step, backbone)
+        callback_checkpoint(global_step, backbone, module_fc, backbone_onnx)
+        scheduler_backbone.step()
+        scheduler_pfc.step()
+    dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    torch.backends.cudnn.benchmark = True
+    parser = argparse.ArgumentParser(description='PyTorch ArcFace-VPL Training')
+    parser.add_argument('config', type=str, help='py config file')
+    parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
+    parser.add_argument('--resume', type=int, default=0, help='model resuming')
+    args_ = parser.parse_args()
+    main(args_)
+
diff --git a/recognition/vpl/utils/__init__.py b/recognition/vpl/utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/vpl/utils/plot.py b/recognition/vpl/utils/plot.py
new file mode 100644
index 000000000..ccc588e5c
--- /dev/null
+++ b/recognition/vpl/utils/plot.py
@@ -0,0 +1,72 @@
+# coding: utf-8
+
+import os
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from sklearn.metrics import roc_curve, auc
+
+image_path = "/data/anxiang/IJB_release/IJBC"
+files = [
+        "./ms1mv3_arcface_r100/ms1mv3_arcface_r100/ijbc.npy"
+]
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % 'ijbc'))
+
+methods = []
+scores = []
+for file in files:
+    methods.append(file.split('/')[-2])
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append("%s-%s" % (method, "IJBC"))
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10 ** -6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+print(tpr_fpr_table)
diff --git a/recognition/vpl/utils/utils_amp.py b/recognition/vpl/utils/utils_amp.py
new file mode 100644
index 000000000..c60067e99
--- /dev/null
+++ b/recognition/vpl/utils/utils_amp.py
@@ -0,0 +1,81 @@
+from typing import Dict, List
+
+import torch
+from torch._six import container_abcs
+from torch.cuda.amp import GradScaler
+
+
+class _MultiDeviceReplicator(object):
+    """
+    Lazily serves copies of a tensor to requested devices.  Copies are cached per-device.
+    """
+
+    def __init__(self, master_tensor: torch.Tensor) -> None:
+        assert master_tensor.is_cuda
+        self.master = master_tensor
+        self._per_device_tensors: Dict[torch.device, torch.Tensor] = {}
+
+    def get(self, device) -> torch.Tensor:
+        retval = self._per_device_tensors.get(device, None)
+        if retval is None:
+            retval = self.master.to(device=device, non_blocking=True, copy=True)
+            self._per_device_tensors[device] = retval
+        return retval
+
+
+class MaxClipGradScaler(GradScaler):
+    def __init__(self, init_scale, max_scale: float, growth_interval=100):
+        GradScaler.__init__(self, init_scale=init_scale, growth_interval=growth_interval)
+        self.max_scale = max_scale
+
+    def scale_clip(self):
+        if self.get_scale() == self.max_scale:
+            self.set_growth_factor(1)
+        elif self.get_scale() < self.max_scale:
+            self.set_growth_factor(2)
+        elif self.get_scale() > self.max_scale:
+            self._scale.fill_(self.max_scale)
+            self.set_growth_factor(1)
+
+    def scale(self, outputs):
+        """
+        Multiplies ('scales') a tensor or list of tensors by the scale factor.
+
+        Returns scaled outputs.  If this instance of :class:`GradScaler` is not enabled, outputs are returned
+        unmodified.
+
+        Arguments:
+            outputs (Tensor or iterable of Tensors):  Outputs to scale.
+        """
+        if not self._enabled:
+            return outputs
+        self.scale_clip()
+        # Short-circuit for the common case.
+        if isinstance(outputs, torch.Tensor):
+            assert outputs.is_cuda
+            if self._scale is None:
+                self._lazy_init_scale_growth_tracker(outputs.device)
+            assert self._scale is not None
+            return outputs * self._scale.to(device=outputs.device, non_blocking=True)
+
+        # Invoke the more complex machinery only if we're treating multiple outputs.
+        stash: List[_MultiDeviceReplicator] = []  # holds a reference that can be overwritten by apply_scale
+
+        def apply_scale(val):
+            if isinstance(val, torch.Tensor):
+                assert val.is_cuda
+                if len(stash) == 0:
+                    if self._scale is None:
+                        self._lazy_init_scale_growth_tracker(val.device)
+                    assert self._scale is not None
+                    stash.append(_MultiDeviceReplicator(self._scale))
+                return val * stash[0].get(val.device)
+            elif isinstance(val, container_abcs.Iterable):
+                iterable = map(apply_scale, val)
+                if isinstance(val, list) or isinstance(val, tuple):
+                    return type(val)(iterable)
+                else:
+                    return iterable
+            else:
+                raise ValueError("outputs must be a Tensor or an iterable of Tensors")
+        return apply_scale(outputs)
diff --git a/recognition/vpl/utils/utils_callbacks.py b/recognition/vpl/utils/utils_callbacks.py
new file mode 100644
index 000000000..43fb70843
--- /dev/null
+++ b/recognition/vpl/utils/utils_callbacks.py
@@ -0,0 +1,112 @@
+import logging
+import os
+import time
+from typing import List
+
+import torch
+
+from eval import verification
+from torch2onnx import convert_onnx
+from utils.utils_logging import AverageMeter
+
+
+class CallBackVerification(object):
+    def __init__(self, frequent, rank, val_targets, rec_prefix, image_size=(112, 112)):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.highest_acc: float = 0.0
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        if self.rank is 0:
+            self.init_dataset(val_targets=val_targets, data_dir=rec_prefix, image_size=image_size)
+
+    def ver_test(self, backbone: torch.nn.Module, global_step: int):
+        results = []
+        for i in range(len(self.ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                self.ver_list[i], backbone, 10, 10)
+            logging.info('[%s][%d]XNorm: %f' % (self.ver_name_list[i], global_step, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (self.ver_name_list[i], global_step, acc2, std2))
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info(
+                '[%s][%d]Accuracy-Highest: %1.5f' % (self.ver_name_list[i], global_step, self.highest_acc_list[i]))
+            results.append(acc2)
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+        for name in val_targets:
+            path = os.path.join(data_dir, name + ".bin")
+            if os.path.exists(path):
+                data_set = verification.load_bin(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+
+    def __call__(self, num_update, backbone: torch.nn.Module):
+        if self.rank is 0 and num_update > 0 and num_update % self.frequent == 0:
+            backbone.eval()
+            self.ver_test(backbone, num_update)
+            backbone.train()
+
+
+class CallBackLogging(object):
+    def __init__(self, frequent, rank, total_step, batch_size, world_size, writer=None):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.time_start = time.time()
+        self.total_step: int = total_step
+        self.batch_size: int = batch_size
+        self.world_size: int = world_size
+        self.writer = writer
+
+        self.init = False
+        self.tic = 0
+
+    def __call__(self, global_step, loss: AverageMeter, epoch: int, fp16: bool, grad_scaler: torch.cuda.amp.GradScaler):
+        if self.rank is 0 and global_step > 0 and global_step % self.frequent == 0:
+            if self.init:
+                try:
+                    speed: float = self.frequent * self.batch_size / (time.time() - self.tic)
+                    speed_total = speed * self.world_size
+                except ZeroDivisionError:
+                    speed_total = float('inf')
+
+                time_now = (time.time() - self.time_start) / 3600
+                time_total = time_now / ((global_step + 1) / self.total_step)
+                time_for_end = time_total - time_now
+                if self.writer is not None:
+                    self.writer.add_scalar('time_for_end', time_for_end, global_step)
+                    self.writer.add_scalar('loss', loss.avg, global_step)
+                if fp16:
+                    msg = "Speed %.2f samples/sec   Loss %.4f   Epoch: %d   Global Step: %d   " \
+                          "Fp16 Grad Scale: %2.f   Required: %1.f hours" % (
+                              speed_total, loss.avg, epoch, global_step, grad_scaler.get_scale(), time_for_end
+                          )
+                else:
+                    msg = "Speed %.2f samples/sec   Loss %.4f   Epoch: %d   Global Step: %d   Required: %1.f hours" % (
+                        speed_total, loss.avg, epoch, global_step, time_for_end
+                    )
+                logging.info(msg)
+                loss.reset()
+                self.tic = time.time()
+            else:
+                self.init = True
+                self.tic = time.time()
+
+
+class CallBackModelCheckpoint(object):
+    def __init__(self, rank, output="./"):
+        self.rank: int = rank
+        self.output: str = output
+
+    def __call__(self, global_step, backbone, partial_fc, backbone_onnx):
+        if global_step > 100 and self.rank is 0:
+            path_module = os.path.join(self.output, "backbone.pth")
+            path_onnx = os.path.join(self.output, "backbone.onnx")
+            torch.save(backbone.module.state_dict(), path_module)
+            logging.info("Pytorch Model Saved in '{}'".format(path_module))
+            convert_onnx(backbone_onnx, path_module, path_onnx)
+            logging.info("Onnx Model Saved in '{}'".format(path_onnx))
+
+        if global_step > 100 and partial_fc is not None:
+            partial_fc.save_params()
diff --git a/recognition/vpl/utils/utils_config.py b/recognition/vpl/utils/utils_config.py
new file mode 100644
index 000000000..372ed921f
--- /dev/null
+++ b/recognition/vpl/utils/utils_config.py
@@ -0,0 +1,16 @@
+import importlib
+import os
+import os.path as osp
+
+def get_config(config_file):
+    assert config_file.startswith('configs/'), 'config file setting must start with configs/'
+    temp_config_name = osp.basename(config_file)
+    temp_module_name = osp.splitext(temp_config_name)[0]
+    config = importlib.import_module("configs.base")
+    cfg = config.config
+    config = importlib.import_module("configs.%s"%temp_module_name)
+    job_cfg = config.config
+    cfg.update(job_cfg)
+    if cfg.output is None:
+        cfg.output = osp.join('work_dirs', temp_module_name)
+    return cfg
diff --git a/recognition/vpl/utils/utils_dist.py b/recognition/vpl/utils/utils_dist.py
new file mode 100644
index 000000000..c191f288f
--- /dev/null
+++ b/recognition/vpl/utils/utils_dist.py
@@ -0,0 +1,57 @@
+import torch
+
+@torch.no_grad()
+def concat_all_gather(tensor):
+	"""
+	Performs all_gather operation on the provided tensors.
+	*** Warning ***: torch.distributed.all_gather has no gradient.
+	"""
+	tensors_gather = [torch.ones_like(tensor)
+		for _ in range(torch.distributed.get_world_size())]
+	torch.distributed.all_gather(tensors_gather, tensor, async_op=False)
+
+	output = torch.cat(tensors_gather, dim=0)
+	return output
+
+@torch.no_grad()
+def batch_shuffle_ddp(x, rank, world_size):
+	"""
+	Batch shuffle, for making use of BatchNorm.
+	*** Only support DistributedDataParallel (DDP) model. ***
+	"""
+	# gather from all gpus
+	batch_size_this = x.shape[0]
+	x_gather = concat_all_gather(x)
+	batch_size_all = x_gather.shape[0]
+
+
+	# random shuffle index
+	idx_shuffle = torch.randperm(batch_size_all).cuda()
+
+	# broadcast to all gpus
+	torch.distributed.broadcast(idx_shuffle, src=0)
+
+	# index for restoring
+	idx_unshuffle = torch.argsort(idx_shuffle)
+
+	# shuffled index for this gpu
+	idx_this = idx_shuffle.view(world_size, -1)[rank]
+
+	return x_gather[idx_this], idx_unshuffle
+
+@torch.no_grad()
+def batch_unshuffle_ddp(x, idx_unshuffle, rank, world_size):
+	"""
+	Undo batch shuffle.
+	*** Only support DistributedDataParallel (DDP) model. ***
+	"""
+	# gather from all gpus
+	batch_size_this = x.shape[0]
+	x_gather = concat_all_gather(x)
+	batch_size_all = x_gather.shape[0]
+
+
+	# restored index for this gpu
+	idx_this = idx_unshuffle.view(world_size, -1)[rank]
+
+	return x_gather[idx_this]
diff --git a/recognition/vpl/utils/utils_logging.py b/recognition/vpl/utils/utils_logging.py
new file mode 100644
index 000000000..7d58012ea
--- /dev/null
+++ b/recognition/vpl/utils/utils_logging.py
@@ -0,0 +1,40 @@
+import logging
+import os
+import sys
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value
+    """
+
+    def __init__(self):
+        self.val = None
+        self.avg = None
+        self.sum = None
+        self.count = None
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def init_logging(log_root, rank, models_root):
+    if rank is 0:
+        log_root.setLevel(logging.INFO)
+        formatter = logging.Formatter("Training: %(asctime)s-%(message)s")
+        handler_file = logging.FileHandler(os.path.join(models_root, "training.log"))
+        handler_stream = logging.StreamHandler(sys.stdout)
+        handler_file.setFormatter(formatter)
+        handler_stream.setFormatter(formatter)
+        log_root.addHandler(handler_file)
+        log_root.addHandler(handler_stream)
+        log_root.info('rank_id: %d' % rank)
diff --git a/recognition/vpl/utils/utils_os.py b/recognition/vpl/utils/utils_os.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/recognition/vpl/vpl.py b/recognition/vpl/vpl.py
new file mode 100644
index 000000000..039fa76cc
--- /dev/null
+++ b/recognition/vpl/vpl.py
@@ -0,0 +1,186 @@
+import logging
+import os
+
+import torch
+import torch.distributed as dist
+from torch.nn import Module
+from torch.nn.functional import normalize, linear
+from torch.nn.parameter import Parameter
+
+
+class VPL(Module):
+    """
+    Modified from Partial-FC
+    """
+
+    @torch.no_grad()
+    def __init__(self, rank, local_rank, world_size, batch_size, resume,
+                 margin_softmax, num_classes, sample_rate=1.0, embedding_size=512, prefix="./", cfg=None):
+        super(VPL, self).__init__()
+        #
+        assert sample_rate==1.0
+        assert not resume
+        self.num_classes: int = num_classes
+        self.rank: int = rank
+        self.local_rank: int = local_rank
+        self.device: torch.device = torch.device("cuda:{}".format(self.local_rank))
+        self.world_size: int = world_size
+        self.batch_size: int = batch_size
+        self.margin_softmax: callable = margin_softmax
+        self.sample_rate: float = sample_rate
+        self.embedding_size: int = embedding_size
+        self.prefix: str = prefix
+        self.num_local: int = num_classes // world_size + int(rank < num_classes % world_size)
+        self.class_start: int = num_classes // world_size * rank + min(rank, num_classes % world_size)
+        self.num_sample: int = int(self.sample_rate * self.num_local)
+
+        self.weight_name = os.path.join(self.prefix, "rank_{}_softmax_weight.pt".format(self.rank))
+        self.weight_mom_name = os.path.join(self.prefix, "rank_{}_softmax_weight_mom.pt".format(self.rank))
+
+        self.weight = torch.normal(0, 0.01, (self.num_local, self.embedding_size), device=self.device)
+        self.weight_mom: torch.Tensor = torch.zeros_like(self.weight)
+        logging.info("softmax weight init successfully!")
+        logging.info("softmax weight mom init successfully!")
+        self.stream: torch.cuda.Stream = torch.cuda.Stream(local_rank)
+
+        self.index = None
+        self.update = lambda: 0
+        self.sub_weight = Parameter(self.weight)
+        self.sub_weight_mom = self.weight_mom
+
+        #vpl variables
+
+        self._iters = 0
+        self.cfg = cfg
+        self.vpl_mode = -1
+        if self.cfg is not None:
+            self.vpl_mode = self.cfg['mode']
+            if self.vpl_mode>=0:
+                self.register_buffer("queue", torch.randn(self.num_local, self.embedding_size, device=self.device))
+                self.queue = normalize(self.queue)
+                self.register_buffer("queue_iters", torch.zeros((self.num_local,), dtype=torch.long, device=self.device))
+                self.register_buffer("queue_lambda", torch.zeros((self.num_local,), dtype=torch.float32, device=self.device))
+
+
+    def save_params(self):
+        pass
+        #torch.save(self.weight.data, self.weight_name)
+        #torch.save(self.weight_mom, self.weight_mom_name)
+
+    @torch.no_grad()
+    def sample(self, total_label):
+        index_positive = (self.class_start <= total_label) & (total_label < self.class_start + self.num_local)
+        total_label[~index_positive] = -1
+        total_label[index_positive] -= self.class_start
+        return index_positive
+
+    def forward(self, total_features, norm_weight):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        logits = linear(total_features, norm_weight)
+        return logits
+
+    @torch.no_grad()
+    def update(self):
+        self.weight_mom[self.index] = self.sub_weight_mom
+        self.weight[self.index] = self.sub_weight
+
+    def prepare(self, label, optimizer):
+        with torch.cuda.stream(self.stream):
+            total_label = torch.zeros(
+                size=[self.batch_size * self.world_size], device=self.device, dtype=torch.long)
+            dist.all_gather(list(total_label.chunk(self.world_size, dim=0)), label)
+            index_positive = self.sample(total_label)
+            optimizer.state.pop(optimizer.param_groups[-1]['params'][0], None)
+            optimizer.param_groups[-1]['params'][0] = self.sub_weight
+            optimizer.state[self.sub_weight]['momentum_buffer'] = self.sub_weight_mom
+            norm_weight = normalize(self.sub_weight)
+            return total_label, norm_weight, index_positive
+
+    @torch.no_grad()
+    def prepare_queue_lambda(self, label, iters):
+        self.queue_lambda[:] = 0.0
+        if iters>self.cfg['start_iters']:
+            allowed_delta = self.cfg['allowed_delta']
+            if self.vpl_mode==0:
+                past_iters = iters - self.queue_iters
+                idx = torch.where(past_iters <= allowed_delta)[0]
+                self.queue_lambda[idx] = self.cfg['lambda']
+
+            if iters % 2000 == 0 and self.rank == 0:
+                logging.info('[%d]use-lambda: %d/%d'%(iters,len(idx), self.num_local))
+
+    @torch.no_grad()
+    def set_queue(self, total_features, total_label, index_positive, iters):
+        local_label = total_label[index_positive]
+        sel_features = normalize(total_features[index_positive,:])
+        self.queue[local_label,:] = sel_features
+        self.queue_iters[local_label] = iters
+
+    def forward_backward(self, label, features, optimizer, feature_w):
+        self._iters += 1
+        total_label, norm_weight, index_positive = self.prepare(label, optimizer)
+        total_features = torch.zeros(
+            size=[self.batch_size * self.world_size, self.embedding_size], device=self.device)
+        dist.all_gather(list(total_features.chunk(self.world_size, dim=0)), features.data)
+        total_features.requires_grad = True
+
+        if feature_w is not None:
+            total_feature_w = torch.zeros(
+                size=[self.batch_size * self.world_size, self.embedding_size], device=self.device)
+            dist.all_gather(list(total_feature_w.chunk(self.world_size, dim=0)), feature_w.data)
+
+        if self.vpl_mode>=0:
+            self.prepare_queue_lambda(total_label, self._iters)
+            _lambda = self.queue_lambda.view(self.num_local, 1)
+            injected_weight = norm_weight*(1.0-_lambda) + self.queue*_lambda
+            injected_norm_weight = normalize(injected_weight)
+            logits = self.forward(total_features, injected_norm_weight)
+        else:
+            logits = self.forward(total_features, norm_weight)
+
+        logits = self.margin_softmax(logits, total_label)
+
+        with torch.no_grad():
+            max_fc = torch.max(logits, dim=1, keepdim=True)[0]
+            dist.all_reduce(max_fc, dist.ReduceOp.MAX)
+
+            # calculate exp(logits) and all-reduce
+            logits_exp = torch.exp(logits - max_fc)
+            logits_sum_exp = logits_exp.sum(dim=1, keepdims=True)
+            dist.all_reduce(logits_sum_exp, dist.ReduceOp.SUM)
+
+            # calculate prob
+            logits_exp.div_(logits_sum_exp)
+
+            # get one-hot
+            grad = logits_exp
+            index = torch.where(total_label != -1)[0]
+            one_hot = torch.zeros(size=[index.size()[0], grad.size()[1]], device=grad.device)
+            one_hot.scatter_(1, total_label[index, None], 1)
+
+            # calculate loss
+            loss = torch.zeros(grad.size()[0], 1, device=grad.device)
+            loss[index] = grad[index].gather(1, total_label[index, None])
+            dist.all_reduce(loss, dist.ReduceOp.SUM)
+            loss_v = loss.clamp_min_(1e-30).log_().mean() * (-1)
+
+            # calculate grad
+            grad[index] -= one_hot
+            grad.div_(self.batch_size * self.world_size)
+
+        logits.backward(grad)
+        if total_features.grad is not None:
+            total_features.grad.detach_()
+        x_grad: torch.Tensor = torch.zeros_like(features, requires_grad=True)
+        # feature gradient all-reduce
+        dist.reduce_scatter(x_grad, list(total_features.grad.chunk(self.world_size, dim=0)))
+        x_grad = x_grad * self.world_size
+        #vpl set queue
+        if self.vpl_mode>=0:
+            if feature_w is None:
+                self.set_queue(total_features.detach(), total_label, index_positive, self._iters)
+            else:
+                self.set_queue(total_feature_w, total_label, index_positive, self._iters)
+        # backward backbone
+        return x_grad, loss_v
+
diff --git a/reconstruction/PBIDR/INSTALL.md b/reconstruction/PBIDR/INSTALL.md
new file mode 100644
index 000000000..9db65aa6e
--- /dev/null
+++ b/reconstruction/PBIDR/INSTALL.md
@@ -0,0 +1,27 @@
+## MANUAL INSTALL
+
+```bash
+conda create -n pbidr python=3.7
+pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 -f https://download.pytorch.org/whl/torch_stable.html
+conda install -c conda-forge scikit-image shapely rtree pyembree
+pip install trimesh[all]
+```
+
+#### Install Pytorch3D
+
+```bash
+wget https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz
+tar xzf 1.10.0.tar.gz
+export CUB_HOME=$PWD/cub-1.10.0
+wget https://github.com/facebookresearch/pytorch3d/archive/refs/tags/v0.4.0.tar.gz
+tar xzf v0.4.0.tar.gz
+export TORCH_CUDA_ARCH_LIST="7.5"
+cd pytorch3d-0.4.0 && pip install -e .
+```
+
+#### Install Other Requirments
+
+```bash
+pip install requirements.txt
+```
+
diff --git a/reconstruction/PBIDR/README.md b/reconstruction/PBIDR/README.md
new file mode 100644
index 000000000..03610f154
--- /dev/null
+++ b/reconstruction/PBIDR/README.md
@@ -0,0 +1,75 @@
+# Facial Geometric Detail Recovery via Implicit Representation
+
+:herb: **[Facial Geometric Detail Recovery via Implicit Representation](https://arxiv.org/abs/2203.09692)**
+
+Xingyu Ren, Alexandros Lattas, Baris Gecer, Jiankang Deng, Chao Ma, Xiaokang Yang, and Stefanos Zafeiriou. 
+
+*arXiv Preprint 2022*
+
+## Introduction
+
+![overview](https://github.com/deepinsight/insightface/blob/master/reconstruction/PBIDR/figures/overview.jpg)
+
+This paper introduces a single facial image geometric detail recovery algorithm. The method generates complete high-fidelity texture maps from occluded facial images, and employs implicit renderer and shape functions, to derive fine geometric details by decoupled specular normals. As a bonus, it disentangles the facial texture into approximate diffuse albedo, diffuse and specular shading in a self-supervision manner.
+
+## Installation
+
+Please refer to the installation and usage of [IDR](https://github.com/lioryariv/idr).
+
+The code is compatible with python 3.7 and pytorch 1.7.1. In addition, the following packages are required:  
+numpy, scikit-image, trimesh (with pyembree), opencv, torchvision, pytorch3d 0.4.0.
+
+You can see [INSTALL.md](INSTALL.md) for manual installation.
+
+## Tutorial
+
+### Data Preprocessing
+
+ We have provided several textured meshes from [Google Drive](https://drive.google.com/file/d/1R7MdWawdMSjQUOnciJ5mb1pcwoY61Tzc/view?usp=sharing) and [Baidu Drive](https://pan.baidu.com/s/16mAqB_7mlbW2--0__patWA) (password: wp47). Otherwise, please refer to [OSTeC](https://github.com/barisgecer/OSTeC) to make a textured mesh firstly.
+
+Please download raw textured meshes and run:
+
+ ```shell
+cd ./code
+bash script/data_process.sh
+ ```
+
+ You can synthesize the auxiliary image sets for the next implicit details recovery.
+
+### Train & Eval
+
+You can start the training phase with the following script.
+
+ ```shell
+cd ./code
+bash script/fast_train.sh
+ ```
+
+ We also provide a script for eval: 
+
+ ```shell
+cd ./code
+bash script/fast_eval.sh
+ ```
+
+## Citation
+
+ If any parts of our paper and codes are helpful to your work, please generously citing:
+
+ ```
+@misc{ren2022facial,
+      title={Facial Geometric Detail Recovery via Implicit Representation}, 
+      author={Xingyu Ren and Alexandros Lattas and Baris Gecer and Jiankang Deng and Chao Ma and Xiaokang Yang and Stefanos Zafeiriou},
+      year={2022},
+      eprint={2203.09692},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+ ```
+
+## Reference
+
+ We refer to the following repositories when implementing our whole pipeline. Thanks for their great work.
+
+ - [barisgecer/OSTeC](https://github.com/barisgecer/OSTeC)
+ - [lioryariv/idr](https://github.com/lioryariv/idr)
diff --git a/reconstruction/PBIDR/code/confs/test.conf b/reconstruction/PBIDR/code/confs/test.conf
new file mode 100644
index 000000000..8851fb43a
--- /dev/null
+++ b/reconstruction/PBIDR/code/confs/test.conf
@@ -0,0 +1,72 @@
+train{
+    expname = test
+    dataset_class = datasets.dataset.IFDataset
+    model_class = model.renderer.IFNetwork
+    loss_class = model.loss.IFLoss
+    learning_rate = 1.0e-4
+    num_pixels = 2048
+    plot_freq = 100
+    alpha_milestones = [250, 500, 750, 1000, 1250]
+    alpha_factor = 2
+    sched_milestones = [1000,1500]
+    sched_factor = 0.5
+}
+plot{
+    plot_nimgs = 1
+    max_depth = 3.0
+    resolution = 100
+}
+loss{
+    eikonal_weight = 0.1
+    mask_weight = 100.0
+    reg_weight = 5.0
+    normal_weight = 1.0
+    alpha = 50.0
+}
+dataset{
+    data_dir = Test
+    img_res = [1024, 1024]
+    scan_id = 0
+}
+model{
+    feature_vector_size = 256
+    implicit_network
+    {
+        d_in = 3
+        d_out = 1
+        dims = [512, 512, 512, 512, 512, 512, 512, 512]
+        geometric_init = True
+        bias = 0.6
+        skip_in = [4]
+        weight_norm = True
+        multires = 6
+    }
+    diffuse_network
+    {
+        dims = [128]
+        weight_norm = True
+        multires_view = 6
+    }
+    specular_network
+    {
+        dims = [128]
+        weight_norm = True
+        multires_view = 4
+    }
+    albedo_network
+    {
+        dims = [256, 256, 256, 256]
+        weight_norm = True
+        multires_view = 4
+    }
+    ray_tracer
+    {
+        object_bounding_sphere = 1.0
+        sdf_threshold = 5.0e-5
+        line_search_step = 0.5
+        line_step_iters = 3
+        sphere_tracing_iters = 10
+        n_steps = 100
+        n_secant_steps = 8
+    }
+}
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/datasets/dataset.py b/reconstruction/PBIDR/code/datasets/dataset.py
new file mode 100644
index 000000000..ae12367bd
--- /dev/null
+++ b/reconstruction/PBIDR/code/datasets/dataset.py
@@ -0,0 +1,154 @@
+import os
+import torch
+import numpy as np
+
+import utils.general as utils
+from utils import rend_util
+
+class IFDataset(torch.utils.data.Dataset):
+    """Dataset for a class of objects, where each datapoint is a SceneInstanceDataset."""
+
+    def __init__(self,
+                 train_cameras,
+                 data_dir,
+                 img_res,
+                 scan_id=0,
+                 cam_file=None
+                 ):
+
+        self.instance_dir = os.path.join('../data', data_dir, 'scan{0}'.format(scan_id))
+
+        self.total_pixels = img_res[0] * img_res[1]
+        self.img_res = img_res
+
+        assert os.path.exists(self.instance_dir), "Data directory is empty"
+
+        self.sampling_idx = None
+        self.train_cameras = train_cameras
+
+        image_dir = '{0}/image'.format(self.instance_dir)
+        image_paths = sorted(utils.glob_imgs(image_dir))
+        mask_dir = '{0}/mask'.format(self.instance_dir)
+        mask_paths = sorted(utils.glob_imgs(mask_dir))
+
+        self.n_images = len(image_paths)
+
+        self.cam_file = '{0}/cameras.npz'.format(self.instance_dir)
+        if cam_file is not None:
+            self.cam_file = '{0}/{1}'.format(self.instance_dir, cam_file)
+
+        camera_dict = np.load(self.cam_file)
+        scale_mats = [camera_dict['scale_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+        world_mats = [camera_dict['world_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+
+
+        self.intrinsics_all = []
+        self.pose_all = []
+        for scale_mat, world_mat in zip(scale_mats, world_mats):
+            P = world_mat @ scale_mat
+            P = P[:3, :4]
+            intrinsics, pose = rend_util.load_K_Rt_from_P(None, P)
+            self.intrinsics_all.append(torch.from_numpy(intrinsics).float())
+            self.pose_all.append(torch.from_numpy(pose).float())
+
+        self.rgb_images = []
+        for path in image_paths:
+            rgb = rend_util.load_rgb(path)
+            rgb = rgb.reshape(3, -1).transpose(1, 0)
+            self.rgb_images.append(torch.from_numpy(rgb).float())
+
+        self.object_masks = []
+        for path in mask_paths:
+            object_mask = rend_util.load_mask_white_bg(path)
+            object_mask = object_mask.reshape(-1)
+            self.object_masks.append(torch.from_numpy(object_mask).bool())
+
+    def __len__(self):
+        return self.n_images
+
+    def __getitem__(self, idx):
+        uv = np.mgrid[0:self.img_res[0], 0:self.img_res[1]].astype(np.int32)
+        uv = torch.from_numpy(np.flip(uv, axis=0).copy()).float()
+        uv = uv.reshape(2, -1).transpose(1, 0)
+
+        sample = {
+            "object_mask": self.object_masks[idx],
+            "uv": uv,
+            "intrinsics": self.intrinsics_all[idx],
+        }
+
+        ground_truth = {
+            "rgb": self.rgb_images[idx]
+        }
+
+        if self.sampling_idx is not None:
+            ground_truth["rgb"] = self.rgb_images[idx][self.sampling_idx, :]
+            sample["object_mask"] = self.object_masks[idx][self.sampling_idx]
+            sample["uv"] = uv[self.sampling_idx, :]
+
+        if not self.train_cameras:
+            sample["pose"] = self.pose_all[idx]
+
+        return idx, sample, ground_truth
+
+    def collate_fn(self, batch_list):
+        # get list of dictionaries and returns input, ground_true as dictionary for all batch instances
+        batch_list = zip(*batch_list)
+
+        all_parsed = []
+        for entry in batch_list:
+            if type(entry[0]) is dict:
+                # make them all into a new dict
+                ret = {}
+                for k in entry[0].keys():
+                    ret[k] = torch.stack([obj[k] for obj in entry])
+                all_parsed.append(ret)
+            else:
+                all_parsed.append(torch.LongTensor(entry))
+
+        return tuple(all_parsed)
+
+    def change_sampling_idx(self, sampling_size):
+        if sampling_size == -1:
+            self.sampling_idx = None
+        else:
+            self.sampling_idx = torch.randperm(self.total_pixels)[:sampling_size]
+
+    def get_scale_mat(self):
+        return np.load(self.cam_file)['scale_mat_0']
+
+    def get_gt_pose(self, scaled=False):
+        # Load gt pose without normalization to unit sphere
+        camera_dict = np.load(self.cam_file)
+        world_mats = [camera_dict['world_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+        scale_mats = [camera_dict['scale_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+
+        pose_all = []
+        for scale_mat, world_mat in zip(scale_mats, world_mats):
+            P = world_mat
+            if scaled:
+                P = world_mat @ scale_mat
+            P = P[:3, :4]
+            _, pose = rend_util.load_K_Rt_from_P(None, P)
+            pose_all.append(torch.from_numpy(pose).float())
+
+        return torch.cat([p.float().unsqueeze(0) for p in pose_all], 0)
+
+    def get_pose_init(self):
+        # get noisy initializations obtained with the linear method
+        cam_file = '{0}/cameras_linear_init.npz'.format(self.instance_dir)
+        camera_dict = np.load(cam_file)
+        scale_mats = [camera_dict['scale_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+        world_mats = [camera_dict['world_mat_%d' % idx].astype(np.float32) for idx in range(self.n_images)]
+
+        init_pose = []
+        for scale_mat, world_mat in zip(scale_mats, world_mats):
+            P = world_mat @ scale_mat
+            P = P[:3, :4]
+            _, pose = rend_util.load_K_Rt_from_P(None, P)
+            init_pose.append(pose)
+        init_pose = torch.cat([torch.Tensor(pose).float().unsqueeze(0) for pose in init_pose], 0).cuda()
+        init_quat = rend_util.rot_to_quat(init_pose[:, :3, :3])
+        init_quat = torch.cat([init_quat, init_pose[:, :3, 3]], 1)
+
+        return init_quat
diff --git a/reconstruction/PBIDR/code/evaluation/eval.py b/reconstruction/PBIDR/code/evaluation/eval.py
new file mode 100644
index 000000000..efb1b6ae3
--- /dev/null
+++ b/reconstruction/PBIDR/code/evaluation/eval.py
@@ -0,0 +1,211 @@
+import sys
+sys.path.append('../code')
+import argparse
+import GPUtil
+import os
+from pyhocon import ConfigFactory
+import torch
+import numpy as np
+import cvxpy as cp
+from PIL import Image
+import math
+
+import utils.general as utils
+import utils.plots as plt
+from utils import rend_util
+
+def evaluate(**kwargs):
+    torch.set_default_dtype(torch.float32)
+
+    conf = ConfigFactory.parse_file(kwargs['conf'])
+    exps_folder_name = kwargs['exps_folder_name']
+    evals_folder_name = kwargs['evals_folder_name']
+    eval_rendering = kwargs['eval_rendering']
+    eval_animation = kwargs['eval_animation']
+
+    expname = conf.get_string('train.expname') + kwargs['expname']
+    scan_id = kwargs['scan_id'] if kwargs['scan_id'] != -1 else conf.get_int('dataset.scan_id', default=-1)
+    if scan_id != -1:
+        expname = expname + '_{0}'.format(scan_id)
+
+    if kwargs['timestamp'] == 'latest':
+        if os.path.exists(os.path.join('../', kwargs['exps_folder_name'], expname)):
+            timestamps = os.listdir(os.path.join('../', kwargs['exps_folder_name'], expname))
+            if (len(timestamps)) == 0:
+                print('WRONG EXP FOLDER')
+                exit()
+            else:
+                timestamp = sorted(timestamps)[-1]
+        else:
+            print('WRONG EXP FOLDER')
+            exit()
+    else:
+        timestamp = kwargs['timestamp']
+
+    utils.mkdir_ifnotexists(os.path.join('../', evals_folder_name))
+    expdir = os.path.join('../', exps_folder_name, expname)
+    evaldir = os.path.join('../', evals_folder_name, expname)
+    utils.mkdir_ifnotexists(evaldir)
+
+    dataset_conf = conf.get_config('dataset')
+    model = utils.get_class(conf.get_string('train.model_class'))(conf=conf.get_config('model'),\
+                                                                  id=scan_id, datadir=dataset_conf['data_dir'])
+    if torch.cuda.is_available():
+        model.cuda()
+
+
+    if kwargs['scan_id'] != -1:
+        dataset_conf['scan_id'] = kwargs['scan_id']
+    eval_dataset = utils.get_class(conf.get_string('train.dataset_class'))(False, **dataset_conf)
+
+    if eval_rendering:
+        eval_dataloader = torch.utils.data.DataLoader(eval_dataset,
+                                                      batch_size=1,
+                                                      shuffle=False,
+                                                      collate_fn=eval_dataset.collate_fn
+                                                      )
+        total_pixels = eval_dataset.total_pixels
+        img_res = eval_dataset.img_res
+
+    old_checkpnts_dir = os.path.join(expdir, timestamp, 'checkpoints')
+
+    saved_model_state = torch.load(os.path.join(old_checkpnts_dir, 'ModelParameters', str(kwargs['checkpoint']) + ".pth"))
+    model.load_state_dict(saved_model_state["model_state_dict"])
+    epoch = saved_model_state['epoch']
+
+    ####################################################################################################################
+    print("evaluating...")
+
+    model.eval()
+
+    detail_3dmm, detail_3dmm_subdivision_full = plt.get_displacement_mesh(model)
+    detail_3dmm.export('{0}/Detailed_3dmm_{1}.obj'.format(evaldir, epoch), 'obj')
+    detail_3dmm_subdivision_full.export('{0}/Subdivide_full_{1}.obj'.format(evaldir, epoch), 'obj')
+
+    if eval_animation:
+        sdf_np0, sdf_np1 = plt.get_displacement_animation(model)
+        np.save('{0}/Cropped_Detailed_sdf_{1}.npy'.format(evaldir, epoch), sdf_np0)
+        np.save('{0}/Cropped_Subdivide_full_{1}.npy'.format(evaldir, epoch), sdf_np1)
+
+    if eval_rendering:
+        images_dir = '{0}/rendering'.format(evaldir)
+        utils.mkdir_ifnotexists(images_dir)
+
+        psnrs = []
+        for data_index, (indices, model_input, ground_truth) in enumerate(eval_dataloader):
+            model_input["intrinsics"] = model_input["intrinsics"].cuda()
+            model_input["uv"] = model_input["uv"].cuda()
+            model_input["object_mask"] = model_input["object_mask"].cuda()
+            model_input['pose'] = model_input['pose'].cuda()
+
+            split = utils.split_input(model_input, total_pixels)
+            res = []
+            for s in split:
+                out = model(s)
+                res.append({
+                    'rgb_values': out['rgb_values'].detach(),
+                    'diffuse_values': out['diffuse_values'].detach(),
+                    'specular_values': out['specular_values'].detach(),
+                    'albedo_values': out['albedo_values'].detach(),
+                })
+
+            batch_size = ground_truth['rgb'].shape[0]
+            model_outputs = utils.merge_output(res, total_pixels, batch_size)
+            rgb_eval = model_outputs['rgb_values']
+            rgb_eval = rgb_eval.reshape(batch_size, total_pixels, 3)
+            rgb_eval = (rgb_eval + 1.) / 2.
+            rgb_eval = plt.lin2img(rgb_eval, img_res).detach().cpu().numpy()[0]
+            rgb_eval = rgb_eval.transpose(1, 2, 0)
+            img = Image.fromarray((rgb_eval * 255).astype(np.uint8))
+            img.save('{0}/eval_{1}.png'.format(images_dir,'%03d' % indices[0]))
+
+            diffuse_eval = model_outputs['diffuse_values']
+            diffuse_eval = diffuse_eval.reshape(batch_size, total_pixels, 3)
+            diffuse_eval = (diffuse_eval + 1.) / 2.
+            diffuse_eval = plt.lin2img(diffuse_eval, img_res).detach().cpu().numpy()[0]
+            diffuse_eval = diffuse_eval.transpose(1, 2, 0)
+            img = Image.fromarray((diffuse_eval * 255).astype(np.uint8))
+            img.save('{0}/eval_{1}_diffuse.png'.format(images_dir, '%03d' % indices[0]))
+
+            specular_eval = model_outputs['specular_values']
+            specular_eval = specular_eval.reshape(batch_size, total_pixels, 3)
+            specular_eval = (specular_eval + 1.) / 2.
+            specular_eval = plt.lin2img(specular_eval, img_res).detach().cpu().numpy()[0]
+            specular_eval = specular_eval.transpose(1, 2, 0)
+            img = Image.fromarray((specular_eval * 255).astype(np.uint8))
+            img.save('{0}/eval_{1}_specular.png'.format(images_dir, '%03d' % indices[0]))
+
+            albedo_eval = model_outputs['albedo_values']
+            albedo_eval = albedo_eval.reshape(batch_size, total_pixels, 3)
+            albedo_eval = (albedo_eval + 1.) / 2.
+            albedo_eval = plt.lin2img(albedo_eval, img_res).detach().cpu().numpy()[0]
+            albedo_eval = albedo_eval.transpose(1, 2, 0)
+            img = Image.fromarray((albedo_eval * 255).astype(np.uint8))
+            img.save('{0}/eval_{1}_albedo.png'.format(images_dir, '%03d' % indices[0]))
+
+            rgb_gt = ground_truth['rgb']
+            rgb_gt = (rgb_gt + 1.) / 2.
+            rgb_gt = plt.lin2img(rgb_gt, img_res).numpy()[0]
+            rgb_gt = rgb_gt.transpose(1, 2, 0)
+
+            mask = model_input['object_mask']
+            mask = plt.lin2img(mask.unsqueeze(-1), img_res).cpu().numpy()[0]
+            mask = mask.transpose(1, 2, 0)
+
+            rgb_eval_masked = rgb_eval * mask
+            rgb_gt_masked = rgb_gt * mask
+
+            psnr = calculate_psnr(rgb_eval_masked, rgb_gt_masked, mask)
+            psnrs.append(psnr)
+
+        psnrs = np.array(psnrs).astype(np.float64)
+        print("RENDERING EVALUATION {2}: psnr mean = {0} ; psnr std = {1}".format("%.2f" % psnrs.mean(), "%.2f" % psnrs.std(), scan_id))
+
+
+
+def calculate_psnr(img1, img2, mask):
+    # img1 and img2 have range [0, 1]
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    mse = np.mean((img1 - img2)**2) * (img2.shape[0] * img2.shape[1]) / mask.sum()
+    if mse == 0:
+        return float('inf')
+    return 20 * math.log10(1.0 / math.sqrt(mse))
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--conf', type=str, default='./confs/test.conf')
+    parser.add_argument('--expname', type=str, default='', help='The experiment name to be evaluated.')
+    parser.add_argument('--exps_folder', type=str, default='exps', help='The experiments folder name.')
+    parser.add_argument('--gpu', type=str, default='auto', help='GPU to use [default: GPU auto]')
+    parser.add_argument('--timestamp', default='latest', type=str, help='The experiemnt timestamp to test.')
+    parser.add_argument('--checkpoint', default='latest',type=str,help='The trained model checkpoint to test')
+    parser.add_argument('--scan_id', type=int, default=0, help='If set, taken to be the scan id.')
+    parser.add_argument('--resolution', default=512, type=int, help='Grid resolution for marching cube')
+    parser.add_argument('--is_uniform_grid', default=False, action="store_true", help='If set, evaluate marching cube with uniform grid.')
+    parser.add_argument('--eval_rendering', default=False, action="store_true",help='If set, evaluate rendering quality.')
+    parser.add_argument('--eval_animation', default=False, action="store_true",help='If set, evaluate rendering quality.')
+
+    opt = parser.parse_args()
+
+    if opt.gpu == "auto":
+        deviceIDs = GPUtil.getAvailable(order='memory', limit=1, maxLoad=0.5, maxMemory=0.5, includeNan=False, excludeID=[], excludeUUID=[])
+        gpu = deviceIDs[0]
+    else:
+        gpu = opt.gpu
+
+    if (not gpu == 'ignore'):
+        os.environ["CUDA_VISIBLE_DEVICES"] = '{0}'.format(gpu)
+
+    evaluate(conf=opt.conf,
+             expname=opt.expname,
+             exps_folder_name=opt.exps_folder,
+             evals_folder_name='evals',
+             timestamp=opt.timestamp,
+             checkpoint=opt.checkpoint,
+             scan_id=opt.scan_id,
+             resolution=opt.resolution,
+             eval_rendering=opt.eval_rendering,
+             eval_animation=opt.eval_animation
+             )
diff --git a/reconstruction/PBIDR/code/model/embedder.py b/reconstruction/PBIDR/code/model/embedder.py
new file mode 100644
index 000000000..02f2c0835
--- /dev/null
+++ b/reconstruction/PBIDR/code/model/embedder.py
@@ -0,0 +1,50 @@
+import torch
+
+""" Positional encoding embedding. Code was taken from https://github.com/bmild/nerf. """
+
+class Embedder:
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+        self.create_embedding_fn()
+
+    def create_embedding_fn(self):
+        embed_fns = []
+        d = self.kwargs['input_dims']
+        out_dim = 0
+        if self.kwargs['include_input']:
+            embed_fns.append(lambda x: x)
+            out_dim += d
+
+        max_freq = self.kwargs['max_freq_log2']
+        N_freqs = self.kwargs['num_freqs']
+
+        if self.kwargs['log_sampling']:
+            freq_bands = 2. ** torch.linspace(0., max_freq, N_freqs)
+        else:
+            freq_bands = torch.linspace(2.**0., 2.**max_freq, N_freqs)
+
+        for freq in freq_bands:
+            for p_fn in self.kwargs['periodic_fns']:
+                embed_fns.append(lambda x, p_fn=p_fn,
+                                 freq=freq: p_fn(x * freq))
+                out_dim += d
+
+        self.embed_fns = embed_fns
+        self.out_dim = out_dim
+
+    def embed(self, inputs):
+        return torch.cat([fn(inputs) for fn in self.embed_fns], -1)
+
+def get_embedder(multires):
+    embed_kwargs = {
+        'include_input': True,
+        'input_dims': 3,
+        'max_freq_log2': multires-1,
+        'num_freqs': multires,
+        'log_sampling': True,
+        'periodic_fns': [torch.sin, torch.cos],
+    }
+
+    embedder_obj = Embedder(**embed_kwargs)
+    def embed(x, eo=embedder_obj): return eo.embed(x)
+    return embed, embedder_obj.out_dim
diff --git a/reconstruction/PBIDR/code/model/loss.py b/reconstruction/PBIDR/code/model/loss.py
new file mode 100644
index 000000000..9843ad350
--- /dev/null
+++ b/reconstruction/PBIDR/code/model/loss.py
@@ -0,0 +1,69 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+class IFLoss(nn.Module):
+    def __init__(self, eikonal_weight, mask_weight, reg_weight, normal_weight, alpha):
+        super().__init__()
+        self.eikonal_weight = eikonal_weight
+        self.mask_weight = mask_weight
+        self.reg_weight = reg_weight
+        self.normal_weight = normal_weight
+        self.alpha = alpha
+        self.l1_loss = nn.L1Loss(reduction='sum')
+        self.l2_loss = nn.MSELoss(reduction='sum')
+        self.cosine = nn.CosineSimilarity()
+
+    def get_rgb_loss(self,rgb_values, rgb_gt, network_object_mask, object_mask):
+        if (network_object_mask & object_mask).sum() == 0:
+            return torch.tensor(0.0).cuda().float()
+
+        rgb_values = rgb_values[network_object_mask & object_mask]
+        rgb_gt = rgb_gt.reshape(-1, 3)[network_object_mask & object_mask]
+        rgb_loss = self.l1_loss(rgb_values, rgb_gt) / float(object_mask.shape[0])
+        return rgb_loss
+
+    def get_eikonal_loss(self, grad_theta):
+        if grad_theta.shape[0] == 0:
+            return torch.tensor(0.0).cuda().float()
+
+        eikonal_loss = ((grad_theta.norm(2, dim=1) - 1) ** 2).mean()
+        return eikonal_loss
+
+    def get_mask_loss(self, sdf_output, network_object_mask, object_mask):
+        mask = ~(network_object_mask & object_mask)
+        if mask.sum() == 0:
+            return torch.tensor(0.0).cuda().float()
+        sdf_pred = -self.alpha * sdf_output[mask]
+        gt = object_mask[mask].float()
+        mask_loss = (1 / self.alpha) * F.binary_cross_entropy_with_logits(sdf_pred.squeeze(), gt, reduction='sum') / float(object_mask.shape[0])
+        return mask_loss
+
+    def get_reg_loss(self, point_gt, point_pre):
+        loss = self.l2_loss(point_gt, point_pre) / len(point_pre)
+        return loss
+
+    def forward(self, model_outputs, ground_truth):
+        rgb_gt = ground_truth['rgb'].cuda()
+        network_object_mask = model_outputs['network_object_mask']
+        object_mask = model_outputs['object_mask']
+
+        rgb_loss = self.get_rgb_loss(model_outputs['rgb_values'], rgb_gt, network_object_mask, object_mask)
+        mask_loss = self.get_mask_loss(model_outputs['sdf_output'], network_object_mask, object_mask)
+        eikonal_loss = self.get_eikonal_loss(model_outputs['grad_theta'])
+        reg_loss = self.get_reg_loss(model_outputs['points_mesh_ray_gt'], model_outputs['points_pre'])
+        normal_loss = 1 - torch.mean(self.cosine(model_outputs['points_mesh_ray_normals'], model_outputs['surface_normals']))
+        loss = rgb_loss + \
+               self.eikonal_weight * eikonal_loss + \
+               self.mask_weight * mask_loss + \
+               self.reg_weight * reg_loss + \
+                self.normal_weight * normal_loss
+
+        return {
+            'loss': loss,
+            'rgb_loss': rgb_loss,
+            'eikonal_loss': eikonal_loss,
+            'mask_loss': mask_loss,
+            'reg_loss': reg_loss,
+            'normal_loss': normal_loss,
+        }
diff --git a/reconstruction/PBIDR/code/model/ray_tracing.py b/reconstruction/PBIDR/code/model/ray_tracing.py
new file mode 100644
index 000000000..2e9ea8352
--- /dev/null
+++ b/reconstruction/PBIDR/code/model/ray_tracing.py
@@ -0,0 +1,301 @@
+import torch
+import torch.nn as nn
+from utils import rend_util
+
+class RayTracing(nn.Module):
+    def __init__(
+            self,
+            object_bounding_sphere=1.0,
+            sdf_threshold=5.0e-5,
+            line_search_step=0.5,
+            line_step_iters=1,
+            sphere_tracing_iters=10,
+            n_steps=100,
+            n_secant_steps=8,
+    ):
+        super().__init__()
+
+        self.object_bounding_sphere = object_bounding_sphere
+        self.sdf_threshold = sdf_threshold
+        self.sphere_tracing_iters = sphere_tracing_iters
+        self.line_step_iters = line_step_iters
+        self.line_search_step = line_search_step
+        self.n_steps = n_steps
+        self.n_secant_steps = n_secant_steps
+
+    def forward(self,
+                sdf,
+                cam_loc,
+                object_mask,
+                ray_directions
+                ):
+
+        batch_size, num_pixels, _ = ray_directions.shape
+
+        sphere_intersections, mask_intersect = rend_util.get_sphere_intersection(cam_loc, ray_directions, r=self.object_bounding_sphere)
+
+        curr_start_points, unfinished_mask_start, acc_start_dis, acc_end_dis, min_dis, max_dis = \
+            self.sphere_tracing(batch_size, num_pixels, sdf, cam_loc, ray_directions, mask_intersect, sphere_intersections)
+
+        network_object_mask = (acc_start_dis < acc_end_dis)
+
+        # The non convergent rays should be handled by the sampler
+        sampler_mask = unfinished_mask_start
+        sampler_net_obj_mask = torch.zeros_like(sampler_mask).bool().cuda()
+        if sampler_mask.sum() > 0:
+            sampler_min_max = torch.zeros((batch_size, num_pixels, 2)).cuda()
+            sampler_min_max.reshape(-1, 2)[sampler_mask, 0] = acc_start_dis[sampler_mask]
+            sampler_min_max.reshape(-1, 2)[sampler_mask, 1] = acc_end_dis[sampler_mask]
+
+            sampler_pts, sampler_net_obj_mask, sampler_dists = self.ray_sampler(sdf,
+                                                                                cam_loc,
+                                                                                object_mask,
+                                                                                ray_directions,
+                                                                                sampler_min_max,
+                                                                                sampler_mask
+                                                                                )
+
+            curr_start_points[sampler_mask] = sampler_pts[sampler_mask]
+            acc_start_dis[sampler_mask] = sampler_dists[sampler_mask]
+            network_object_mask[sampler_mask] = sampler_net_obj_mask[sampler_mask]
+
+        print('----------------------------------------------------------------')
+        print('RayTracing: object = {0}/{1}, secant on {2}/{3}.'
+              .format(network_object_mask.sum(), len(network_object_mask), sampler_net_obj_mask.sum(), sampler_mask.sum()))
+        print('----------------------------------------------------------------')
+
+        if not self.training:
+            return curr_start_points, \
+                   network_object_mask, \
+                   acc_start_dis
+
+        ray_directions = ray_directions.reshape(-1, 3)
+        mask_intersect = mask_intersect.reshape(-1)
+
+        in_mask = ~network_object_mask & object_mask & ~sampler_mask
+        out_mask = ~object_mask & ~sampler_mask
+
+        mask_left_out = (in_mask | out_mask) & ~mask_intersect
+        if mask_left_out.sum() > 0:  # project the origin to the not intersect points on the sphere
+            cam_left_out = cam_loc.unsqueeze(1).repeat(1, num_pixels, 1).reshape(-1, 3)[mask_left_out]
+            rays_left_out = ray_directions[mask_left_out]
+            acc_start_dis[mask_left_out] = -torch.bmm(rays_left_out.view(-1, 1, 3), cam_left_out.view(-1, 3, 1)).squeeze()
+            curr_start_points[mask_left_out] = cam_left_out + acc_start_dis[mask_left_out].unsqueeze(1) * rays_left_out
+
+        mask = (in_mask | out_mask) & mask_intersect
+
+        if mask.sum() > 0:
+            min_dis[network_object_mask & out_mask] = acc_start_dis[network_object_mask & out_mask]
+
+            min_mask_points, min_mask_dist = self.minimal_sdf_points(num_pixels, sdf, cam_loc, ray_directions, mask, min_dis, max_dis)
+
+            curr_start_points[mask] = min_mask_points
+            acc_start_dis[mask] = min_mask_dist
+
+        return curr_start_points, \
+               network_object_mask, \
+               acc_start_dis
+
+
+    def sphere_tracing(self, batch_size, num_pixels, sdf, cam_loc, ray_directions, mask_intersect, sphere_intersections):
+        ''' Run sphere tracing algorithm for max iterations from both sides of unit sphere intersection '''
+
+        sphere_intersections_points = cam_loc.reshape(batch_size, 1, 1, 3) + sphere_intersections.unsqueeze(-1) * ray_directions.unsqueeze(2)
+        unfinished_mask_start = mask_intersect.reshape(-1).clone()
+        unfinished_mask_end = mask_intersect.reshape(-1).clone()
+
+        # Initialize start current points
+        curr_start_points = torch.zeros(batch_size * num_pixels, 3).cuda().float()
+        curr_start_points[unfinished_mask_start] = sphere_intersections_points[:,:,0,:].reshape(-1,3)[unfinished_mask_start]
+        acc_start_dis = torch.zeros(batch_size * num_pixels).cuda().float()
+        acc_start_dis[unfinished_mask_start] = sphere_intersections.reshape(-1,2)[unfinished_mask_start,0]
+
+        # Initialize end current points
+        curr_end_points = torch.zeros(batch_size * num_pixels, 3).cuda().float()
+        curr_end_points[unfinished_mask_end] = sphere_intersections_points[:,:,1,:].reshape(-1,3)[unfinished_mask_end]
+        acc_end_dis = torch.zeros(batch_size * num_pixels).cuda().float()
+        acc_end_dis[unfinished_mask_end] = sphere_intersections.reshape(-1,2)[unfinished_mask_end,1]
+
+        # Initizliae min and max depth
+        min_dis = acc_start_dis.clone()
+        max_dis = acc_end_dis.clone()
+
+        # Iterate on the rays (from both sides) till finding a surface
+        iters = 0
+
+        next_sdf_start = torch.zeros_like(acc_start_dis).cuda()
+        next_sdf_start[unfinished_mask_start] = sdf(curr_start_points[unfinished_mask_start])
+
+        next_sdf_end = torch.zeros_like(acc_end_dis).cuda()
+        next_sdf_end[unfinished_mask_end] = sdf(curr_end_points[unfinished_mask_end])
+
+        while True:
+            # Update sdf
+            curr_sdf_start = torch.zeros_like(acc_start_dis).cuda()
+            curr_sdf_start[unfinished_mask_start] = next_sdf_start[unfinished_mask_start]
+            curr_sdf_start[curr_sdf_start <= self.sdf_threshold] = 0
+
+            curr_sdf_end = torch.zeros_like(acc_end_dis).cuda()
+            curr_sdf_end[unfinished_mask_end] = next_sdf_end[unfinished_mask_end]
+            curr_sdf_end[curr_sdf_end <= self.sdf_threshold] = 0
+
+            # Update masks
+            unfinished_mask_start = unfinished_mask_start & (curr_sdf_start > self.sdf_threshold)
+            unfinished_mask_end = unfinished_mask_end & (curr_sdf_end > self.sdf_threshold)
+
+            if (unfinished_mask_start.sum() == 0 and unfinished_mask_end.sum() == 0) or iters == self.sphere_tracing_iters:
+                break
+            iters += 1
+
+            # Make step
+            # Update distance
+            acc_start_dis = acc_start_dis + curr_sdf_start
+            acc_end_dis = acc_end_dis - curr_sdf_end
+
+            # Update points
+            curr_start_points = (cam_loc.unsqueeze(1) + acc_start_dis.reshape(batch_size, num_pixels, 1) * ray_directions).reshape(-1, 3)
+            curr_end_points = (cam_loc.unsqueeze(1) + acc_end_dis.reshape(batch_size, num_pixels, 1) * ray_directions).reshape(-1, 3)
+
+            # Fix points which wrongly crossed the surface
+            next_sdf_start = torch.zeros_like(acc_start_dis).cuda()
+            next_sdf_start[unfinished_mask_start] = sdf(curr_start_points[unfinished_mask_start])
+
+            next_sdf_end = torch.zeros_like(acc_end_dis).cuda()
+            next_sdf_end[unfinished_mask_end] = sdf(curr_end_points[unfinished_mask_end])
+
+            not_projected_start = next_sdf_start < 0
+            not_projected_end = next_sdf_end < 0
+            not_proj_iters = 0
+            while (not_projected_start.sum() > 0 or not_projected_end.sum() > 0) and not_proj_iters < self.line_step_iters:
+                # Step backwards
+                acc_start_dis[not_projected_start] -= ((1 - self.line_search_step) / (2 ** not_proj_iters)) * curr_sdf_start[not_projected_start]
+                curr_start_points[not_projected_start] = (cam_loc.unsqueeze(1) + acc_start_dis.reshape(batch_size, num_pixels, 1) * ray_directions).reshape(-1, 3)[not_projected_start]
+
+                acc_end_dis[not_projected_end] += ((1 - self.line_search_step) / (2 ** not_proj_iters)) * curr_sdf_end[not_projected_end]
+                curr_end_points[not_projected_end] = (cam_loc.unsqueeze(1) + acc_end_dis.reshape(batch_size, num_pixels, 1) * ray_directions).reshape(-1, 3)[not_projected_end]
+
+                # Calc sdf
+                next_sdf_start[not_projected_start] = sdf(curr_start_points[not_projected_start])
+                next_sdf_end[not_projected_end] = sdf(curr_end_points[not_projected_end])
+
+                # Update mask
+                not_projected_start = next_sdf_start < 0
+                not_projected_end = next_sdf_end < 0
+                not_proj_iters += 1
+
+            unfinished_mask_start = unfinished_mask_start & (acc_start_dis < acc_end_dis)
+            unfinished_mask_end = unfinished_mask_end & (acc_start_dis < acc_end_dis)
+
+        return curr_start_points, unfinished_mask_start, acc_start_dis, acc_end_dis, min_dis, max_dis
+
+    def ray_sampler(self, sdf, cam_loc, object_mask, ray_directions, sampler_min_max, sampler_mask):
+        ''' Sample the ray in a given range and run secant on rays which have sign transition '''
+
+        batch_size, num_pixels, _ = ray_directions.shape
+        n_total_pxl = batch_size * num_pixels
+        sampler_pts = torch.zeros(n_total_pxl, 3).cuda().float()
+        sampler_dists = torch.zeros(n_total_pxl).cuda().float()
+
+        intervals_dist = torch.linspace(0, 1, steps=self.n_steps).cuda().view(1, 1, -1)
+
+        pts_intervals = sampler_min_max[:, :, 0].unsqueeze(-1) + intervals_dist * (sampler_min_max[:, :, 1] - sampler_min_max[:, :, 0]).unsqueeze(-1)
+        points = cam_loc.reshape(batch_size, 1, 1, 3) + pts_intervals.unsqueeze(-1) * ray_directions.unsqueeze(2)
+
+        # Get the non convergent rays
+        mask_intersect_idx = torch.nonzero(sampler_mask).flatten()
+        points = points.reshape((-1, self.n_steps, 3))[sampler_mask, :, :]
+        pts_intervals = pts_intervals.reshape((-1, self.n_steps))[sampler_mask]
+
+        sdf_val_all = []
+        for pnts in torch.split(points.reshape(-1, 3), 100000, dim=0):
+            sdf_val_all.append(sdf(pnts))
+        sdf_val = torch.cat(sdf_val_all).reshape(-1, self.n_steps)
+
+        tmp = torch.sign(sdf_val) * torch.arange(self.n_steps, 0, -1).cuda().float().reshape((1, self.n_steps))  # Force argmin to return the first min value
+        sampler_pts_ind = torch.argmin(tmp, -1)
+        sampler_pts[mask_intersect_idx] = points[torch.arange(points.shape[0]), sampler_pts_ind, :]
+        sampler_dists[mask_intersect_idx] = pts_intervals[torch.arange(pts_intervals.shape[0]), sampler_pts_ind]
+
+        true_surface_pts = object_mask[sampler_mask]
+        net_surface_pts = (sdf_val[torch.arange(sdf_val.shape[0]), sampler_pts_ind] < 0)
+
+        # take points with minimal SDF value for P_out pixels
+        p_out_mask = ~(true_surface_pts & net_surface_pts)
+        n_p_out = p_out_mask.sum()
+        if n_p_out > 0:
+            out_pts_idx = torch.argmin(sdf_val[p_out_mask, :], -1)
+            sampler_pts[mask_intersect_idx[p_out_mask]] = points[p_out_mask, :, :][torch.arange(n_p_out), out_pts_idx, :]
+            sampler_dists[mask_intersect_idx[p_out_mask]] = pts_intervals[p_out_mask, :][torch.arange(n_p_out), out_pts_idx]
+
+        # Get Network object mask
+        sampler_net_obj_mask = sampler_mask.clone()
+        sampler_net_obj_mask[mask_intersect_idx[~net_surface_pts]] = False
+
+        # Run Secant method
+        secant_pts = net_surface_pts & true_surface_pts if self.training else net_surface_pts
+        n_secant_pts = secant_pts.sum()
+        if n_secant_pts > 0:
+            # Get secant z predictions
+            z_high = pts_intervals[torch.arange(pts_intervals.shape[0]), sampler_pts_ind][secant_pts]
+            sdf_high = sdf_val[torch.arange(sdf_val.shape[0]), sampler_pts_ind][secant_pts]
+            z_low = pts_intervals[secant_pts][torch.arange(n_secant_pts), sampler_pts_ind[secant_pts] - 1]
+            sdf_low = sdf_val[secant_pts][torch.arange(n_secant_pts), sampler_pts_ind[secant_pts] - 1]
+            cam_loc_secant = cam_loc.unsqueeze(1).repeat(1, num_pixels, 1).reshape((-1, 3))[mask_intersect_idx[secant_pts]]
+            ray_directions_secant = ray_directions.reshape((-1, 3))[mask_intersect_idx[secant_pts]]
+            z_pred_secant = self.secant(sdf_low, sdf_high, z_low, z_high, cam_loc_secant, ray_directions_secant, sdf)
+
+            # Get points
+            sampler_pts[mask_intersect_idx[secant_pts]] = cam_loc_secant + z_pred_secant.unsqueeze(-1) * ray_directions_secant
+            sampler_dists[mask_intersect_idx[secant_pts]] = z_pred_secant
+
+        return sampler_pts, sampler_net_obj_mask, sampler_dists
+
+    def secant(self, sdf_low, sdf_high, z_low, z_high, cam_loc, ray_directions, sdf):
+        ''' Runs the secant method for interval [z_low, z_high] for n_secant_steps '''
+
+        z_pred = - sdf_low * (z_high - z_low) / (sdf_high - sdf_low) + z_low
+        for i in range(self.n_secant_steps):
+            p_mid = cam_loc + z_pred.unsqueeze(-1) * ray_directions
+            sdf_mid = sdf(p_mid)
+            ind_low = sdf_mid > 0
+            if ind_low.sum() > 0:
+                z_low[ind_low] = z_pred[ind_low]
+                sdf_low[ind_low] = sdf_mid[ind_low]
+            ind_high = sdf_mid < 0
+            if ind_high.sum() > 0:
+                z_high[ind_high] = z_pred[ind_high]
+                sdf_high[ind_high] = sdf_mid[ind_high]
+
+            z_pred = - sdf_low * (z_high - z_low) / (sdf_high - sdf_low) + z_low
+
+        return z_pred
+
+    def minimal_sdf_points(self, num_pixels, sdf, cam_loc, ray_directions, mask, min_dis, max_dis):
+        ''' Find points with minimal SDF value on rays for P_out pixels '''
+
+        n_mask_points = mask.sum()
+
+        n = self.n_steps
+        # steps = torch.linspace(0.0, 1.0,n).cuda()
+        steps = torch.empty(n).uniform_(0.0, 1.0).cuda()
+        mask_max_dis = max_dis[mask].unsqueeze(-1)
+        mask_min_dis = min_dis[mask].unsqueeze(-1)
+        steps = steps.unsqueeze(0).repeat(n_mask_points, 1) * (mask_max_dis - mask_min_dis) + mask_min_dis
+
+        mask_points = cam_loc.unsqueeze(1).repeat(1, num_pixels, 1).reshape(-1, 3)[mask]
+        mask_rays = ray_directions[mask, :]
+
+        mask_points_all = mask_points.unsqueeze(1).repeat(1, n, 1) + steps.unsqueeze(-1) * mask_rays.unsqueeze(
+            1).repeat(1, n, 1)
+        points = mask_points_all.reshape(-1, 3)
+
+        mask_sdf_all = []
+        for pnts in torch.split(points, 100000, dim=0):
+            mask_sdf_all.append(sdf(pnts))
+
+        mask_sdf_all = torch.cat(mask_sdf_all).reshape(-1, n)
+        min_vals, min_idx = mask_sdf_all.min(-1)
+        min_mask_points = mask_points_all.reshape(-1, n, 3)[torch.arange(0, n_mask_points), min_idx]
+        min_mask_dist = steps.reshape(-1, n)[torch.arange(0, n_mask_points), min_idx]
+
+        return min_mask_points, min_mask_dist
diff --git a/reconstruction/PBIDR/code/model/renderer.py b/reconstruction/PBIDR/code/model/renderer.py
new file mode 100644
index 000000000..708169894
--- /dev/null
+++ b/reconstruction/PBIDR/code/model/renderer.py
@@ -0,0 +1,461 @@
+import torch
+import torch.nn as nn
+import numpy as np
+import trimesh
+import os
+
+from utils import rend_util
+from model.embedder import *
+from model.ray_tracing import RayTracing
+from model.sample_network import SampleNetwork
+
+
+def barycentric_coordinates(p, select_vertices):
+
+    a = select_vertices[:, 0, :]
+    b = select_vertices[:, 1, :]
+    c = select_vertices[:, 2, :]
+    # p = point
+
+    v0 = b - a
+    v1 = c - a
+    v2 = p - a
+    d00 = (v0 * v0).sum(axis=1)
+    d01 = (v0 * v1).sum(axis=1)
+    d11 = (v1 * v1).sum(axis=1)
+    d20 = (v2 * v0).sum(axis=1)
+    d21 = (v2 * v1).sum(axis=1)
+    denom = d00 * d11 - d01 * d01
+    v = (d11 * d20 - d01 * d21) / denom
+    w = (d00 * d21 - d01 * d20) / denom
+    u = 1 - v - w
+
+    return np.vstack([u, v, w]).T
+
+class ImplicitNetwork(nn.Module):
+    def __init__(
+            self,
+            feature_vector_size,
+            d_in,
+            d_out,
+            dims,
+            geometric_init=True,
+            bias=1.0,
+            skip_in=(),
+            weight_norm=True,
+            multires=0
+    ):
+        super().__init__()
+
+        dims = [d_in] + dims + [d_out + feature_vector_size]
+
+        self.embed_fn = None
+        if multires > 0:
+            embed_fn, input_ch = get_embedder(multires)
+            self.embed_fn = embed_fn
+            dims[0] = input_ch
+
+        self.num_layers = len(dims)
+        self.skip_in = skip_in
+
+        for l in range(0, self.num_layers - 1):
+            if l + 1 in self.skip_in:
+                out_dim = dims[l + 1] - dims[0]
+            else:
+                out_dim = dims[l + 1]
+
+            lin = nn.Linear(dims[l], out_dim)
+
+            if geometric_init:
+                if l == self.num_layers - 2:
+                    torch.nn.init.normal_(lin.weight, mean=np.sqrt(np.pi) / np.sqrt(dims[l]), std=0.0001)
+                    torch.nn.init.constant_(lin.bias, -bias)
+                elif multires > 0 and l == 0:
+                    torch.nn.init.constant_(lin.bias, 0.0)
+                    torch.nn.init.constant_(lin.weight[:, 3:], 0.0)
+                    torch.nn.init.normal_(lin.weight[:, :3], 0.0, np.sqrt(2) / np.sqrt(out_dim))
+                elif multires > 0 and l in self.skip_in:
+                    torch.nn.init.constant_(lin.bias, 0.0)
+                    torch.nn.init.normal_(lin.weight, 0.0, np.sqrt(2) / np.sqrt(out_dim))
+                    torch.nn.init.constant_(lin.weight[:, -(dims[0] - 3):], 0.0)
+                else:
+                    torch.nn.init.constant_(lin.bias, 0.0)
+                    torch.nn.init.normal_(lin.weight, 0.0, np.sqrt(2) / np.sqrt(out_dim))
+
+            if weight_norm:
+                lin = nn.utils.weight_norm(lin)
+
+            setattr(self, "lin" + str(l), lin)
+
+        self.softplus = nn.Softplus(beta=100)
+
+    def forward(self, input, compute_grad=False):
+        if self.embed_fn is not None:
+            input = self.embed_fn(input)
+
+        x = input
+
+        for l in range(0, self.num_layers - 1):
+            lin = getattr(self, "lin" + str(l))
+
+            if l in self.skip_in:
+                x = torch.cat([x, input], 1) / np.sqrt(2)
+
+            x = lin(x)
+
+            if l < self.num_layers - 2:
+                x = self.softplus(x)
+
+        return x
+
+    def gradient(self, x):
+        x.requires_grad_(True)
+        y = self.forward(x)[:,:1]
+        d_output = torch.ones_like(y, requires_grad=False, device=y.device)
+        gradients = torch.autograd.grad(
+            outputs=y,
+            inputs=x,
+            grad_outputs=d_output,
+            create_graph=True,
+            retain_graph=True,
+            only_inputs=True)[0]
+        return gradients.unsqueeze(1)
+
+class AlbedoNetwork(nn.Module):
+    def __init__(
+            self,
+            feature_vector_size,
+            dims=[512, 512, 512, 512],
+            weight_norm=True,
+            multires_view=4,
+    ):
+        super().__init__()
+
+        dims = [3 + feature_vector_size] + dims + [3]
+        embedview_fn, input_ch = get_embedder(multires_view)
+        self.embedview_fn = embedview_fn
+        dims[0] += (input_ch - 3)
+        self.num_layers = len(dims)
+
+        for l in range(0, self.num_layers - 1):
+            out_dim = dims[l + 1]
+            lin = nn.Linear(dims[l], out_dim)
+
+            if weight_norm:
+                lin = nn.utils.weight_norm(lin)
+
+            setattr(self, "lin" + str(l), lin)
+
+        self.relu = nn.ReLU()
+        self.tanh = nn.Tanh()
+
+    def forward(self, points, feature_vectors):
+
+        Mpoints = self.embedview_fn(points)
+        x = torch.cat([Mpoints, feature_vectors], dim=-1)
+
+        for l in range(0, self.num_layers - 1):
+            lin = getattr(self, "lin" + str(l))
+
+            x = lin(x)
+
+            if l < self.num_layers - 2:
+                x = self.relu(x)
+
+        x = self.tanh(x)
+        return x
+
+class SpecularNetwork(nn.Module):
+    def __init__(
+            self,
+            dims=[256, 256, 256],
+            weight_norm=True,
+            multires_view=4
+    ):
+        super().__init__()
+        dims = [3 + 3] + dims + [1]
+
+        embedview_fn, input_ch = get_embedder(multires_view)
+        self.embedview_fn = embedview_fn
+        dims[0] += (input_ch - 3)
+        dims[0] += (input_ch - 3)
+        self.num_layers = len(dims)
+
+        for l in range(0, self.num_layers - 1):
+            out_dim = dims[l + 1]
+            lin = nn.Linear(dims[l], out_dim)
+
+            if weight_norm:
+                lin = nn.utils.weight_norm(lin)
+
+            setattr(self, "lin" + str(l), lin)
+
+        self.relu = nn.ReLU()
+        self.tanh = nn.Tanh()
+
+    def forward(self, normals, view_dirs):
+
+        Mview_dirs = self.embedview_fn(view_dirs)
+        Mnormals= self.embedview_fn(normals)
+        x = torch.cat([Mview_dirs, Mnormals], dim=-1)
+
+        for l in range(0, self.num_layers - 1):
+            lin = getattr(self, "lin" + str(l))
+
+            x = lin(x)
+
+            if l < self.num_layers - 2:
+                x = self.relu(x)
+
+        x = self.tanh(x)
+        return x
+    def optimaize(self):
+        return
+
+class DiffuseNetwork(nn.Module):
+    def __init__(
+            self,
+            dims=[256, 256, 256],
+            weight_norm=True,
+            multires_view=6,
+    ):
+        super().__init__()
+
+        dims = [3] + dims + [1]
+        embedview_fn, input_ch = get_embedder(multires_view)
+        self.embedview_fn = embedview_fn
+        dims[0] += (input_ch - 3)
+        self.num_layers = len(dims)
+
+        for l in range(0, self.num_layers - 1):
+            out_dim = dims[l + 1]
+            lin = nn.Linear(dims[l], out_dim)
+
+            if weight_norm:
+                lin = nn.utils.weight_norm(lin)
+
+            setattr(self, "lin" + str(l), lin)
+
+        self.relu = nn.ReLU()
+        self.tanh = nn.Tanh()
+
+    def forward(self, normals):
+
+        Mnormals = self.embedview_fn(normals)
+        x = Mnormals
+
+        for l in range(0, self.num_layers - 1):
+            lin = getattr(self, "lin" + str(l))
+
+            x = lin(x)
+
+            if l < self.num_layers - 2:
+                x = self.relu(x)
+
+        x = self.tanh(x)
+        return x
+
+class IFNetwork(nn.Module):
+    def __init__(self, conf, id, datadir):
+        super().__init__()
+        self.feature_vector_size = conf.get_int('feature_vector_size')
+        self.implicit_network = ImplicitNetwork(self.feature_vector_size, **conf.get_config('implicit_network'))
+        # self.rendering_network = RenderingNetwork(self.feature_vector_size, **conf.get_config('rendering_network'))
+
+        self.diffuse_network = DiffuseNetwork(**conf.get_config('diffuse_network'))
+        self.specular_network = SpecularNetwork(**conf.get_config('specular_network'))
+        self.albedo_network = AlbedoNetwork(self.feature_vector_size, **conf.get_config('albedo_network'))
+
+        self.ray_tracer = RayTracing(**conf.get_config('ray_tracer'))
+        self.sample_network = SampleNetwork()
+        self.object_bounding_sphere = conf.get_float('ray_tracer.object_bounding_sphere')
+        self.mesh = trimesh.load_mesh('{0}/mesh.obj'.format(os.path.join('../data', datadir, 'scan{0}'.format(id))),
+                                      process=False, use_embree=True)
+        self.faces = self.mesh.faces
+        self.vertex_normals = np.array(self.mesh.vertex_normals)
+        self.vertices = np.array(self.mesh.vertices)
+        print('Loaded Mesh')
+
+    def forward(self, input):
+
+        # Parse model input
+        points_predicted = None
+
+        intrinsics = input["intrinsics"]
+        uv = input["uv"]
+        pose = input["pose"]
+        object_mask = input["object_mask"].reshape(-1)
+
+        ray_dirs, cam_loc = rend_util.get_camera_params(uv, pose, intrinsics)
+        batch_size, num_pixels, _ = ray_dirs.shape
+
+        self.implicit_network.eval()
+        with torch.no_grad():
+            points, network_object_mask, dists = self.ray_tracer(sdf=lambda x: self.implicit_network(x)[:, 0],
+                                                                 cam_loc=cam_loc,
+                                                                 object_mask=object_mask,
+                                                                 ray_directions=ray_dirs)
+        self.implicit_network.train()
+
+        points = (cam_loc.unsqueeze(1) + dists.reshape(batch_size, num_pixels, 1) * ray_dirs).reshape(-1, 3)
+        points_normal = self.implicit_network.gradient(points)
+        sdf_output = self.implicit_network(points)[:, 0:1]
+        ray_dirs = ray_dirs.reshape(-1, 3)
+
+        ray_dirs_np = ray_dirs.cpu().numpy()
+        cam_loc_np = np.concatenate([cam_loc.cpu().numpy()] * len(ray_dirs_np), axis=0)
+        # points_mesh_ray: may have the more points than surface mask points,
+        # Need an Index for the Points_Mesh_Ray
+        points_mesh_ray, index_ray, index_tri = self.mesh.ray.intersects_location(ray_origins=cam_loc_np,
+                                                                                  ray_directions=ray_dirs_np,
+                                                                                  multiple_hits=False)
+        # Index ray: total 2048 / ~1200
+        MeshRay_mask = torch.tensor([True if i in index_ray else False for i in range(len(cam_loc_np))], dtype=torch.bool).to(points.device)
+        network_object_mask = network_object_mask & MeshRay_mask
+
+        if self.training:
+
+            surface_mask = network_object_mask & object_mask
+
+            listA = surface_mask.cpu().detach().numpy()
+            A = [int(a) for a in listA]
+            AA = [i for i, a in enumerate(A) if a == 1] # surface mask 的 index
+            MeshRay_Index = np.array([i for i, a in enumerate(index_ray) if a in AA], dtype=int)
+
+            face_points_index = self.faces[index_tri][MeshRay_Index]
+            select_vertex_normals = self.vertex_normals[face_points_index]
+            select_vertices = self.vertices[face_points_index]
+
+            points_mesh_ray = points_mesh_ray[MeshRay_Index]
+            bcoords = barycentric_coordinates(points_mesh_ray, select_vertices)
+            resampled_normals = np.sum(np.expand_dims(bcoords, -1) * select_vertex_normals, 1)
+
+            # Mesh Pull
+            resampled_normals = torch.tensor(resampled_normals).to(points)
+            points_mesh_ray = torch.tensor(points_mesh_ray).to(points)
+            sdf_points_mesh_ray = self.implicit_network(points_mesh_ray)[:, 0:1]
+            g_points_mesh_ray = self.implicit_network.gradient(points_mesh_ray)
+            points_predicted = points_mesh_ray - g_points_mesh_ray.squeeze() * sdf_points_mesh_ray
+
+            surface_points = points[surface_mask]
+            surface_dists = dists[surface_mask].unsqueeze(-1)
+            surface_ray_dirs = ray_dirs[surface_mask]
+            surface_cam_loc = cam_loc.unsqueeze(1).repeat(1, num_pixels, 1).reshape(-1, 3)[surface_mask]
+            surface_output = sdf_output[surface_mask]
+            N = surface_points.shape[0]
+
+            # Sample points for the eikonal loss
+            eik_bounding_box = self.object_bounding_sphere
+            n_eik_points = batch_size * num_pixels // 2
+            eikonal_points = torch.empty(n_eik_points, 3).uniform_(-eik_bounding_box, eik_bounding_box).cuda()
+            eikonal_pixel_points = points.clone()
+            eikonal_pixel_points = eikonal_pixel_points.detach()
+            eikonal_points = torch.cat([eikonal_points, eikonal_pixel_points], 0)
+
+            points_all = torch.cat([surface_points, eikonal_points], dim=0)
+
+            output = self.implicit_network(surface_points)
+            surface_sdf_values = output[:N, 0:1].detach()
+
+            g = self.implicit_network.gradient(points_all)
+            surface_points_grad = g[:N, 0, :].clone().detach()
+            grad_theta = g[N:, 0, :]
+
+            differentiable_surface_points = self.sample_network(surface_output,
+                                                                surface_sdf_values,
+                                                                surface_points_grad,
+                                                                surface_dists,
+                                                                surface_cam_loc,
+                                                                surface_ray_dirs)
+
+        else:
+            surface_mask = network_object_mask
+            differentiable_surface_points = points[surface_mask]
+            grad_theta = None
+
+            listA = surface_mask.cpu().detach().numpy()
+            A = [int(a) for a in listA]
+            AA = [i for i, a in enumerate(A) if a == 1]  # surface mask 的 index
+            MeshRay_Index = np.array([i for i, a in enumerate(index_ray) if a in AA], dtype=int)
+
+            face_points_index = self.faces[index_tri][MeshRay_Index]
+            select_vertex_normals = self.vertex_normals[face_points_index]
+            select_vertices = self.vertices[face_points_index]
+
+            points_mesh_ray = points_mesh_ray[MeshRay_Index]
+            bcoords = barycentric_coordinates(points_mesh_ray, select_vertices)
+            resampled_normals = np.sum(np.expand_dims(bcoords, -1) * select_vertex_normals, 1)
+            resampled_normals = torch.tensor(resampled_normals).to(points)
+
+        view = -ray_dirs[surface_mask]
+
+        rgb_values = torch.ones_like(points).float().cuda()
+        diffuse_values = torch.ones_like(points).float().cuda()
+        specular_values = torch.ones_like(points).float().cuda()
+        albedo_values = torch.ones_like(points).float().cuda()
+
+        if differentiable_surface_points.shape[0] > 0:
+
+            rgb_values[surface_mask] = self.get_rbg_value(differentiable_surface_points, view, resampled_normals)
+            diffuse_values[surface_mask] = self.get_diffuse_value(differentiable_surface_points, view, resampled_normals)
+
+            specular_values[surface_mask] = self.get_specular_value(differentiable_surface_points, view)
+            albedo_values[surface_mask] = self.get_albedo_value(differentiable_surface_points, view)
+
+        output = {
+            'points': points,
+            'points_pre': points_predicted,
+            'points_mesh_ray_gt': points[surface_mask],
+            'points_mesh_ray_normals': resampled_normals,
+            'surface_normals': points_normal[surface_mask].reshape([-1, 3]),
+
+            'rgb_values': rgb_values,
+            'diffuse_values': diffuse_values,
+            'specular_values': specular_values,
+            'albedo_values': albedo_values,
+
+            'sdf_output': sdf_output,
+            'network_object_mask': network_object_mask,
+            'object_mask': object_mask,
+            'grad_theta': grad_theta
+        }
+
+        return output
+
+    def get_rbg_value(self, points, view_dirs, diffuse_normals):
+        output = self.implicit_network(points)
+        g = self.implicit_network.gradient(points)
+        normals = g[:, 0, :]
+        feature_vectors = output[:, 1:]
+
+        diffuse_shading = self.diffuse_network(diffuse_normals)
+        specular_shading = self.specular_network(normals, view_dirs)
+        albedo = self.albedo_network(points, feature_vectors)
+
+        diffuse_shading = (diffuse_shading + 1.) / 2.
+        specular_shading = (specular_shading + 1.) / 2.
+        albedo = (albedo + 1.) / 2.
+
+        rgb_vals = diffuse_shading * albedo + specular_shading
+        rgb_vals = (rgb_vals * 2.) - 1.
+
+        return rgb_vals
+
+    def get_diffuse_value(self, points, view_dirs, diffuse_normals):
+
+        diffuse_shading = self.diffuse_network(diffuse_normals)
+        return diffuse_shading.expand([-1, 3])
+
+    def get_albedo_value(self, points, view_dirs):
+        output = self.implicit_network(points)
+        feature_vectors = output[:, 1:]
+        albedo = self.albedo_network(points, feature_vectors)
+
+        return albedo
+
+    def get_specular_value(self, points, view_dirs):
+        g = self.implicit_network.gradient(points)
+        normals = g[:, 0, :]
+
+        specular_shading = self.specular_network(normals, view_dirs)
+        return specular_shading.expand([-1, 3])
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/model/sample_network.py b/reconstruction/PBIDR/code/model/sample_network.py
new file mode 100644
index 000000000..746ac93de
--- /dev/null
+++ b/reconstruction/PBIDR/code/model/sample_network.py
@@ -0,0 +1,20 @@
+import torch.nn as nn
+import torch
+
+class SampleNetwork(nn.Module):
+    '''
+    Represent the intersection (sample) point as differentiable function of the implicit geometry and camera parameters.
+    See equation 3 in the paper for more details.
+    '''
+
+    def forward(self, surface_output, surface_sdf_values, surface_points_grad, surface_dists, surface_cam_loc, surface_ray_dirs):
+        # t -> t(theta)
+        surface_ray_dirs_0 = surface_ray_dirs.detach()
+        surface_points_dot = torch.bmm(surface_points_grad.view(-1, 1, 3),
+                                       surface_ray_dirs_0.view(-1, 3, 1)).squeeze(-1)
+        surface_dists_theta = surface_dists - (surface_output - surface_sdf_values) / surface_points_dot
+
+        # t(theta) -> x(theta,c,v)
+        surface_points_theta_c_v = surface_cam_loc + surface_dists_theta * surface_ray_dirs
+
+        return surface_points_theta_c_v
diff --git a/reconstruction/PBIDR/code/preprocess/get_aux_dataset.py b/reconstruction/PBIDR/code/preprocess/get_aux_dataset.py
new file mode 100644
index 000000000..63c84c393
--- /dev/null
+++ b/reconstruction/PBIDR/code/preprocess/get_aux_dataset.py
@@ -0,0 +1,158 @@
+import os
+import sys
+sys.path.append(os.path.abspath(''))
+import torch
+import argparse
+import numpy as np
+from pytorch3d.io import load_objs_as_meshes, save_obj,load_obj
+from pytorch3d.renderer import (
+    look_at_view_transform,
+    PerspectiveCameras,
+    # FoVPerspectiveCameras,
+    PointLights,
+    # DirectionalLights,
+    # Materials,
+    RasterizationSettings,
+    MeshRenderer,
+    MeshRasterizer,
+    # SoftPhongShader,
+    # SoftSilhouetteShader,
+    SoftPhongShader,
+    # TexturesVertex,
+    Materials
+)
+from PIL import Image
+
+print("Start to get aux dataset!")
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser("PreProcessing")
+    parser.add_argument('--gpu', '-g', type=str, default='0',help='GPU')
+    parser.add_argument('--input', '-i', type=str, default='../raw_data', help='Location of Raw Textured Mesh Dataset')
+    parser.add_argument('--output', '-o', type=int, required=True, help='New aux dataset')
+    parser.add_argument('--yaw', type=int, default=15, help='num_views_yaw')
+    parser.add_argument('--yaw_angle', type=int, default=45, help='yaw_angle')
+    parser.add_argument('--pitch', type=int, default=9, help='num_views_pitch')
+    parser.add_argument('--pitch_angle', type=int, default=30, help='pitch_angle')
+    parser.add_argument('--datapath', type=str, default='../data/', help='Location of code data')
+
+    parser.add_argument('--dataset', '-d', type=str, default='Face', help='FaceTest dataset')
+
+    args = parser.parse_args()
+    os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
+
+    # Setup
+    if torch.cuda.is_available():
+        device = torch.device("cuda:0")
+        torch.cuda.set_device(device)
+    else:
+        device = torch.device("cpu")
+
+    # Set paths
+    DATA_DIR = args.input
+    IMAGE_DIR = os.path.join(DATA_DIR, "mesh")
+
+    if os.path.exists(IMAGE_DIR):
+        os.system("rm -r " + IMAGE_DIR)
+    os.mkdir(IMAGE_DIR)
+    obj_filename = os.path.join(DATA_DIR, "mesh.obj")
+
+    if not os.path.exists(os.path.join(args.datapath, args.dataset)):
+        os.mkdir(os.path.join(args.datapath, args.dataset))
+
+    # Load obj file
+    mesh = load_objs_as_meshes([obj_filename], device=device,load_textures=True)
+    print(obj_filename)
+    print("Loaded Mesh")
+
+    # the number of different viewpoints from which we want to render the mesh.
+    def Ry(q):
+        return np.array([[-np.cos(q * np.pi / 180), 0, -np.sin(q * np.pi / 180)], [0, 1, 0],
+                         [np.sin(q * np.pi / 180), 0, -np.cos(q * np.pi / 180)]])
+    def Rx(q):
+        return np.array([[-1, 0, 0], [0, np.cos(q * np.pi / 180), np.sin(q * np.pi / 180)],
+                         [0, np.sin(q * np.pi / 180), -np.cos(q * np.pi / 180)]])
+
+    def get_R_matrix(azim, axis="Ry"):
+        print("Rotation Martix {}".format(axis))
+        aa = []
+        if axis == "Ry":
+            for q in azim:
+                aa.append(Ry(q))
+            RRR = torch.tensor(np.array(aa)).to(device)
+        else:
+            for q in azim:
+                aa.append(Rx(q))
+            RRR = torch.tensor(np.array(aa)).to(device)
+        return RRR
+
+    num_views = args.yaw + args.pitch
+
+    yaw_dim = torch.linspace(-1 * args.yaw_angle, args.yaw_angle, args.yaw)
+    pitch_dim = torch.linspace(-1 * args.pitch_angle, args.pitch_angle , args.pitch)
+
+    lights = PointLights(device=device, location=[[0, 50, 100]], ambient_color=((1.0, 1.0, 1.0), ), diffuse_color=((0.0, 0.0, 0.0), ), specular_color=((0.0, 0.0, 0.0), ))
+    RRy, TTy = look_at_view_transform(dist=8, elev=0, azim=yaw_dim, up=((0, 1, 0),), device=device)
+
+    TTx = TTy[:args.pitch]
+    RRx = get_R_matrix(azim=pitch_dim, axis="Rx")
+
+    Rtotal = torch.cat([RRy, RRx], dim=0)
+    Ttotal = torch.cat([TTy, TTx], dim=0)
+
+    cameras = PerspectiveCameras(device=device, focal_length=4500, principal_point=((512, 512),), R=Rtotal, T=Ttotal,
+                                 image_size=((1024, 1024),))
+
+    if num_views != 1:
+        camera = PerspectiveCameras(device=device, focal_length=4500, principal_point=((512, 512),), R=Rtotal[None, 1, ...],
+                                T=Ttotal[None, 1, ...], image_size=((1024, 1024),))
+    else:
+        camera = PerspectiveCameras(device=device, focal_length=4500, principal_point=((512, 512),),
+                                    R=Rtotal,
+                                    T=Ttotal, image_size=((1024, 1024),))
+
+    mymaterials = Materials(device=device, shininess=8)
+    raster_settings = RasterizationSettings(
+        image_size=1024,
+        blur_radius=0.0,
+        faces_per_pixel=1,
+    )
+    renderer = MeshRenderer(
+        rasterizer=MeshRasterizer(
+            cameras=camera,
+            raster_settings=raster_settings
+        ),
+        shader=SoftPhongShader(
+            device=device,
+            cameras=camera,
+            lights=lights,
+            materials=mymaterials,
+        )
+    )
+
+    meshes = mesh.extend(num_views)
+    target_images = renderer(meshes, cameras=cameras, lights=lights)
+    target_rgb = [target_images[i, ..., :3] for i in range(num_views)]
+    target_cameras = [PerspectiveCameras(device=device, focal_length=4500, principal_point=((512, 512),), R=Rtotal[None, i, ...],
+                           T=Ttotal[None, i, ...], image_size=((1024, 1024),)) for i in range(num_views)]
+
+    # RGB images
+    if not os.path.exists(os.path.join(IMAGE_DIR, 'image')):
+        os.mkdir(os.path.join(IMAGE_DIR, 'image'))
+    if not os.path.exists(os.path.join(IMAGE_DIR, 'mask')):
+        os.mkdir(os.path.join(IMAGE_DIR, 'mask'))
+
+    for i in range(len(target_images)):
+        img = Image.fromarray((target_images[i, ..., :3].cpu().numpy() * 255).astype(np.uint8))
+        img.save(os.path.join(IMAGE_DIR, 'image/{0}.png'.format('%03d' % int(i+1))))
+        img.save(os.path.join(IMAGE_DIR, 'mask/{0}.png'.format('%03d' % int(i+1))))
+    np.save(os.path.join(IMAGE_DIR,'R.npy'), Rtotal.cpu().numpy())
+    np.save(os.path.join(IMAGE_DIR,'T.npy'), Ttotal.cpu().numpy())
+
+    SCAN_DIR = args.datapath + args.dataset + '/scan' + str(args.output) + "/"
+    if os.path.exists(SCAN_DIR):
+        os.system("rm -r " + SCAN_DIR)
+    os.system("cp -r " + IMAGE_DIR + " " + SCAN_DIR)
+    os.system("cp " + DATA_DIR + "/mesh.* " + SCAN_DIR + ".")
+    print("Finished")
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/preprocess/preprocess_cameras.py b/reconstruction/PBIDR/code/preprocess/preprocess_cameras.py
new file mode 100644
index 000000000..43b9ced8c
--- /dev/null
+++ b/reconstruction/PBIDR/code/preprocess/preprocess_cameras.py
@@ -0,0 +1,100 @@
+import numpy as np
+import matplotlib.image as mpimg
+import matplotlib.pyplot as plt
+import cv2
+import argparse
+from glob import glob
+import os
+import sys
+import pickle
+sys.path.append('../code')
+from scipy.spatial.transform import Rotation
+import utils.general as utils
+
+
+def get_Ps_from_Faces(R, T):
+    Ps = []
+    cam_locs = []
+
+    intrinsics = np.concatenate([[4500.0], [0.0], [512.0], [0.0], [4500.0], [512.0], [0.0], [0.0], [1.0]], axis=0)
+    intrinsics = np.reshape(intrinsics, [3, 3])
+
+    projection = np.concatenate([[1.0], [0.0], [0.0], [0.0], [0.0], [1.0], [0.0], [0.0], [0.0], [0.0], [1.0], [0.0]], axis=0)
+    projection = np.reshape(projection, [3, 4])
+
+    I14 = np.concatenate([[0.0], [0.0], [0.0], [1.0]], axis=0)
+    I14 = np.reshape(I14, [1, 4])
+
+    for i in range(0, len(R)):
+        R0 = R[i]
+        T0 = T[i].reshape(3, 1)
+
+        p = np.concatenate([np.concatenate([R[i].T, T[i].reshape(3, 1)], axis=1), I14], axis=0)
+        P = intrinsics @ projection @ p
+        P = P.astype(np.float64)
+
+        camera_loc = -np.dot(R0, T0)
+        cam_locs.append(camera_loc)
+        Ps.append(P)
+
+    return np.array(Ps)
+
+
+def get_all_mask_points_white_bg(masks_dir):
+    mask_paths = sorted(utils.glob_imgs(masks_dir))
+    mask_points_all=[]
+    mask_ims = []
+    for path in mask_paths:
+        img = mpimg.imread(path)
+        cur_mask = img.max(axis=2) < 0.9
+        mask_points = np.where(img.max(axis=2) < 0.9)
+        xs = mask_points[1]
+        ys = mask_points[0]
+        mask_points_all.append(np.stack((xs,ys,np.ones_like(xs))).astype(float))
+        mask_ims.append(cur_mask)
+    return mask_points_all,np.array(mask_ims)
+
+
+def get_normalization(source_dir):
+    print('Preprocessing', source_dir)
+
+    masks_dir= '{0}/mask'.format(source_dir)
+    mask_points_all, masks_all = get_all_mask_points_white_bg(masks_dir)
+    number_of_cameras = len(masks_all)
+    R = np.load('{0}/R.npy'.format(source_dir))
+    T = np.load('{0}/T.npy'.format(source_dir))
+    Ps = get_Ps_from_Faces(R, T)
+    normalization = np.eye(4).astype(np.float32)
+
+    cameras_new={}
+    for i in range(number_of_cameras):
+        cameras_new['scale_mat_%d' % i] = normalization
+        cameras_new['world_mat_%d' % i] = np.concatenate((Ps[i],np.array([[0,0,0,1.0]])),axis=0).astype(np.float32)
+
+    np.savez('{0}/{1}.npz'.format(source_dir, "cameras"), **cameras_new)
+    print(normalization)
+    print('--------------------------------------------------------')
+
+    if False: #for debugging
+        for i in range(number_of_cameras):
+            plt.figure()
+
+            plt.imshow(mpimg.imread('%s/%03d.png' % ('{0}/mask'.format(source_dir), i+1)))
+            xy = (Ps[i,:2, :] @ (np.concatenate((np.array(all_Xs), np.ones((len(all_Xs), 1))), axis=1).T)) / (
+                        Ps[i,2, :] @ (np.concatenate((np.array(all_Xs), np.ones((len(all_Xs), 1))), axis=1).T))
+
+            plt.plot(xy[0, :], xy[1, :], '*')
+            plt.show()
+
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--scan_id', '-i', type=int, default=0, help='data source folder for preprocess')
+    parser.add_argument('--dataset', '-d', type=str, default='Face', help='dataset dir')
+    opt = parser.parse_args()
+
+    SCAN_DIR = '../data/' + opt.dataset + '/scan' + str(opt.scan_id)
+    get_normalization(SCAN_DIR)
+
+    print('Done!')
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/script/data_process.sh b/reconstruction/PBIDR/code/script/data_process.sh
new file mode 100644
index 000000000..8cc319f52
--- /dev/null
+++ b/reconstruction/PBIDR/code/script/data_process.sh
@@ -0,0 +1,6 @@
+set -ex
+
+GPU=0
+
+python preprocess/get_aux_dataset.py -g $GPU -i '../raw_data/0' -o 0 -d 'Test' --yaw 17 --pitch 0
+python preprocess/preprocess_cameras.py -i 0 -d 'Test'
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/script/fast_eval.sh b/reconstruction/PBIDR/code/script/fast_eval.sh
new file mode 100644
index 000000000..2608eda1d
--- /dev/null
+++ b/reconstruction/PBIDR/code/script/fast_eval.sh
@@ -0,0 +1,4 @@
+set -ex
+
+GPU=0
+python evaluation/eval.py --conf ./confs/test.conf --scan_id 0 --gpu $GPU --checkpoint 400 --eval_rendering
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/script/fast_train.sh b/reconstruction/PBIDR/code/script/fast_train.sh
new file mode 100644
index 000000000..a2fa2da96
--- /dev/null
+++ b/reconstruction/PBIDR/code/script/fast_train.sh
@@ -0,0 +1,4 @@
+set -ex
+
+GPU=0
+python training/runner.py --conf ./confs/test.conf --scan_id 0 --gpu $GPU --nepoch 400
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/training/runner.py b/reconstruction/PBIDR/code/training/runner.py
new file mode 100644
index 000000000..eba2cb947
--- /dev/null
+++ b/reconstruction/PBIDR/code/training/runner.py
@@ -0,0 +1,58 @@
+import sys
+sys.path.append('../code')
+import argparse
+import GPUtil
+import torch
+import random
+import numpy as np
+
+from training.train import IFTrainRunner
+
+
+def setup_seed(seed):
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    torch.backends.cudnn.deterministic = True
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch_size', type=int, default=1, help='input batch size')
+    parser.add_argument('--nepoch', type=int, default=400, help='number of epochs to train for')
+    parser.add_argument('--nepoch_freeze', type=int, default=1000, help='number of epochs to train for')
+    parser.add_argument('--conf', type=str, default='./confs/test.conf')
+    parser.add_argument('--expname', type=str, default='')
+    parser.add_argument('--gpu', type=str, default='auto', help='GPU to use [default: GPU auto]')
+    parser.add_argument('--is_continue', default=False, action="store_true", help='If set, indicates continuing from a previous run.')
+    parser.add_argument('--timestamp', default='latest', type=str, help='The timestamp of the run to be used in case of continuing from a previous run.')
+    parser.add_argument('--checkpoint', default='latest',type=str,help='The checkpoint epoch number of the run to be used in case of continuing from a previous run.')
+    parser.add_argument('--train_cameras', default=False, action="store_true", help='If set, optimizing also camera location.')
+    parser.add_argument('--scan_id', type=int, default=-1, help='If set, taken to be the scan id.')
+
+    opt = parser.parse_args()
+
+    if opt.gpu == "auto":
+        deviceIDs = GPUtil.getAvailable(order='memory', limit=1, maxLoad=0.5, maxMemory=0.5, includeNan=False, excludeID=[], excludeUUID=[])
+        gpu = deviceIDs[0]
+    else:
+        gpu = opt.gpu
+
+    setup_seed(0)
+
+    trainrunner = IFTrainRunner(conf=opt.conf,
+                                 batch_size=opt.batch_size,
+                                 nepochs=opt.nepoch,
+                                 nepoch_freeze=opt.nepoch_freeze,
+                                 expname=opt.expname,
+                                 gpu_index=gpu,
+                                 exps_folder_name='exps',
+                                 is_continue=opt.is_continue,
+                                 timestamp=opt.timestamp,
+                                 checkpoint=opt.checkpoint,
+                                 scan_id=opt.scan_id,
+                                 train_cameras=opt.train_cameras
+                                 )
+
+    trainrunner.run()
diff --git a/reconstruction/PBIDR/code/training/train.py b/reconstruction/PBIDR/code/training/train.py
new file mode 100644
index 000000000..9ae66174a
--- /dev/null
+++ b/reconstruction/PBIDR/code/training/train.py
@@ -0,0 +1,284 @@
+import os
+from datetime import datetime
+from pyhocon import ConfigFactory
+import sys
+import torch
+import torch.nn as nn
+
+import utils.general as utils
+import utils.plots as plt
+
+class IFTrainRunner():
+    def __init__(self,**kwargs):
+        torch.set_default_dtype(torch.float32)
+        torch.set_num_threads(1)
+
+        self.conf = ConfigFactory.parse_file(kwargs['conf'])
+        self.batch_size = kwargs['batch_size']
+
+        self.nepochs = kwargs['nepochs']
+        self.nepoch_freeze = kwargs['nepoch_freeze']
+
+        self.exps_folder_name = kwargs['exps_folder_name']
+        self.GPU_INDEX = kwargs['gpu_index']
+        self.train_cameras = kwargs['train_cameras']
+
+        self.expname = self.conf.get_string('train.expname') + kwargs['expname']
+        scan_id = kwargs['scan_id'] if kwargs['scan_id'] != -1 else self.conf.get_int('dataset.scan_id', default=-1)
+        if scan_id != -1:
+            self.expname = self.expname + '_{0}'.format(scan_id)
+
+        if kwargs['is_continue'] and kwargs['timestamp'] == 'latest':
+            if os.path.exists(os.path.join('../',kwargs['exps_folder_name'],self.expname)):
+                timestamps = os.listdir(os.path.join('../',kwargs['exps_folder_name'],self.expname))
+                if (len(timestamps)) == 0:
+                    is_continue = False
+                    timestamp = None
+                else:
+                    timestamp = sorted(timestamps)[-1]
+                    is_continue = True
+            else:
+                is_continue = False
+                timestamp = None
+        else:
+            timestamp = kwargs['timestamp']
+            is_continue = kwargs['is_continue']
+
+        utils.mkdir_ifnotexists(os.path.join('../',self.exps_folder_name))
+        self.expdir = os.path.join('../', self.exps_folder_name, self.expname)
+        utils.mkdir_ifnotexists(self.expdir)
+        self.timestamp = '{:%Y_%m_%d_%H_%M_%S}'.format(datetime.now())
+        utils.mkdir_ifnotexists(os.path.join(self.expdir, self.timestamp))
+
+        self.plots_dir = os.path.join(self.expdir, self.timestamp, 'plots')
+        utils.mkdir_ifnotexists(self.plots_dir)
+
+        # create checkpoints dirs
+        self.checkpoints_path = os.path.join(self.expdir, self.timestamp, 'checkpoints')
+        utils.mkdir_ifnotexists(self.checkpoints_path)
+        self.model_params_subdir = "ModelParameters"
+        self.optimizer_params_subdir = "OptimizerParameters"
+        self.scheduler_params_subdir = "SchedulerParameters"
+
+        utils.mkdir_ifnotexists(os.path.join(self.checkpoints_path, self.model_params_subdir))
+        utils.mkdir_ifnotexists(os.path.join(self.checkpoints_path, self.optimizer_params_subdir))
+        utils.mkdir_ifnotexists(os.path.join(self.checkpoints_path, self.scheduler_params_subdir))
+
+        if self.train_cameras:
+            self.optimizer_cam_params_subdir = "OptimizerCamParameters"
+            self.cam_params_subdir = "CamParameters"
+
+            utils.mkdir_ifnotexists(os.path.join(self.checkpoints_path, self.optimizer_cam_params_subdir))
+            utils.mkdir_ifnotexists(os.path.join(self.checkpoints_path, self.cam_params_subdir))
+
+        os.system("""cp -r {0} "{1}" """.format(kwargs['conf'], os.path.join(self.expdir, self.timestamp, 'runconf.conf')))
+
+        if (not self.GPU_INDEX == 'ignore'):
+            os.environ["CUDA_VISIBLE_DEVICES"] = '{0}'.format(self.GPU_INDEX)
+
+        print('shell command : {0}'.format(' '.join(sys.argv)))
+
+        print('Loading data ...')
+
+        dataset_conf = self.conf.get_config('dataset')
+        if kwargs['scan_id'] != -1:
+            dataset_conf['scan_id'] = kwargs['scan_id']
+
+        self.train_dataset = utils.get_class(self.conf.get_string('train.dataset_class'))(self.train_cameras,
+                                                                                          **dataset_conf)
+
+        print('Finish loading data ...')
+
+        self.train_dataloader = torch.utils.data.DataLoader(self.train_dataset,
+                                                            batch_size=self.batch_size,
+                                                            shuffle=True,
+                                                            collate_fn=self.train_dataset.collate_fn
+                                                            )
+        self.plot_dataloader = torch.utils.data.DataLoader(self.train_dataset,
+                                                           batch_size=self.conf.get_int('plot.plot_nimgs'),
+                                                           shuffle=True,
+                                                           collate_fn=self.train_dataset.collate_fn
+                                                           )
+
+        self.model = utils.get_class(self.conf.get_string('train.model_class'))(conf=self.conf.get_config('model'), \
+                                                                                id=scan_id, datadir=dataset_conf['data_dir'])
+        if torch.cuda.is_available():
+            self.model.cuda()
+
+        self.loss = utils.get_class(self.conf.get_string('train.loss_class'))(**self.conf.get_config('loss'))
+
+        self.lr = self.conf.get_float('train.learning_rate')
+
+        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.lr)
+
+        self.sched_milestones = self.conf.get_list('train.sched_milestones', default=[])
+        self.sched_factor = self.conf.get_float('train.sched_factor', default=0.0)
+        self.scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, self.sched_milestones, gamma=self.sched_factor)
+
+        # settings for camera optimization
+        if self.train_cameras:
+            num_images = len(self.train_dataset)
+
+            self.pose_vecs = torch.nn.Embedding(num_images, 7, sparse=True).cuda()
+            self.pose_vecs.weight.data.copy_(self.train_dataset.get_pose_init())
+            self.optimizer_cam = torch.optim.SparseAdam(list(self.pose_vecs.parameters()), self.conf.get_float('train.learning_rate_cam'))
+
+        self.start_epoch = 0
+        if is_continue:
+            old_checkpnts_dir = os.path.join(self.expdir, timestamp, 'checkpoints')
+
+            saved_model_state = torch.load(
+                os.path.join(old_checkpnts_dir, 'ModelParameters', str(kwargs['checkpoint']) + ".pth"))
+            self.model.load_state_dict(saved_model_state["model_state_dict"])
+            self.start_epoch = saved_model_state['epoch']
+
+            data = torch.load(
+                os.path.join(old_checkpnts_dir, 'OptimizerParameters', str(kwargs['checkpoint']) + ".pth"))
+            self.optimizer.load_state_dict(data["optimizer_state_dict"])
+
+            data = torch.load(
+                os.path.join(old_checkpnts_dir, self.scheduler_params_subdir, str(kwargs['checkpoint']) + ".pth"))
+            self.scheduler.load_state_dict(data["scheduler_state_dict"])
+
+            if self.train_cameras:
+                data = torch.load(
+                    os.path.join(old_checkpnts_dir, self.optimizer_cam_params_subdir, str(kwargs['checkpoint']) + ".pth"))
+                self.optimizer_cam.load_state_dict(data["optimizer_cam_state_dict"])
+
+                data = torch.load(
+                    os.path.join(old_checkpnts_dir, self.cam_params_subdir, str(kwargs['checkpoint']) + ".pth"))
+                self.pose_vecs.load_state_dict(data["pose_vecs_state_dict"])
+
+        self.num_pixels = self.conf.get_int('train.num_pixels')
+        self.total_pixels = self.train_dataset.total_pixels
+        self.img_res = self.train_dataset.img_res
+        self.n_batches = len(self.train_dataloader)
+        self.plot_freq = self.conf.get_int('train.plot_freq')
+        self.plot_conf = self.conf.get_config('plot')
+
+        self.alpha_milestones = self.conf.get_list('train.alpha_milestones', default=[])
+        self.alpha_factor = self.conf.get_float('train.alpha_factor', default=0.0)
+        for acc in self.alpha_milestones:
+            if self.start_epoch > acc:
+                self.loss.alpha = self.loss.alpha * self.alpha_factor
+
+    def save_checkpoints(self, epoch):
+        torch.save(
+            {"epoch": epoch, "model_state_dict": self.model.state_dict()},
+            os.path.join(self.checkpoints_path, self.model_params_subdir, str(epoch) + ".pth"))
+        torch.save(
+            {"epoch": epoch, "model_state_dict": self.model.state_dict()},
+            os.path.join(self.checkpoints_path, self.model_params_subdir, "latest.pth"))
+
+        torch.save(
+            {"epoch": epoch, "optimizer_state_dict": self.optimizer.state_dict()},
+            os.path.join(self.checkpoints_path, self.optimizer_params_subdir, str(epoch) + ".pth"))
+        torch.save(
+            {"epoch": epoch, "optimizer_state_dict": self.optimizer.state_dict()},
+            os.path.join(self.checkpoints_path, self.optimizer_params_subdir, "latest.pth"))
+
+        torch.save(
+            {"epoch": epoch, "scheduler_state_dict": self.scheduler.state_dict()},
+            os.path.join(self.checkpoints_path, self.scheduler_params_subdir, str(epoch) + ".pth"))
+        torch.save(
+            {"epoch": epoch, "scheduler_state_dict": self.scheduler.state_dict()},
+            os.path.join(self.checkpoints_path, self.scheduler_params_subdir, "latest.pth"))
+
+        if self.train_cameras:
+            torch.save(
+                {"epoch": epoch, "optimizer_cam_state_dict": self.optimizer_cam.state_dict()},
+                os.path.join(self.checkpoints_path, self.optimizer_cam_params_subdir, str(epoch) + ".pth"))
+            torch.save(
+                {"epoch": epoch, "optimizer_cam_state_dict": self.optimizer_cam.state_dict()},
+                os.path.join(self.checkpoints_path, self.optimizer_cam_params_subdir, "latest.pth"))
+
+            torch.save(
+                {"epoch": epoch, "pose_vecs_state_dict": self.pose_vecs.state_dict()},
+                os.path.join(self.checkpoints_path, self.cam_params_subdir, str(epoch) + ".pth"))
+            torch.save(
+                {"epoch": epoch, "pose_vecs_state_dict": self.pose_vecs.state_dict()},
+                os.path.join(self.checkpoints_path, self.cam_params_subdir, "latest.pth"))
+
+    def run(self):
+        print("training...")
+
+        for epoch in range(self.start_epoch, self.nepochs + 1):
+
+            if epoch in self.alpha_milestones:
+                self.loss.alpha = self.loss.alpha * self.alpha_factor
+
+            if epoch % 100 == 0 and epoch != 0:
+                self.save_checkpoints(epoch)
+
+            if epoch % self.plot_freq == 0 and epoch != 0:
+                self.model.eval()
+                if self.train_cameras:
+                    self.pose_vecs.eval()
+                self.train_dataset.change_sampling_idx(-1)
+                indices, model_input, ground_truth = next(iter(self.plot_dataloader))
+
+                model_input["intrinsics"] = model_input["intrinsics"].cuda()
+                model_input["uv"] = model_input["uv"].cuda()
+                model_input["object_mask"] = model_input["object_mask"].cuda()
+                # model_input[""] = ground_truth["rgb"].cuda()
+
+                if self.train_cameras:
+                    pose_input = self.pose_vecs(indices.cuda())
+                    model_input['pose'] = pose_input
+                else:
+                    model_input['pose'] = model_input['pose'].cuda()
+
+                detail_3dmm, detail_3dmm_subdivision_full = plt.get_displacement_mesh(self.model)
+                detail_3dmm.export('{0}/Detailed_3dmm_{1}.obj'.format(self.plots_dir, epoch), 'obj')
+                detail_3dmm_subdivision_full.export('{0}/Subdivide_full_{1}.obj'.format(self.plots_dir, epoch), 'obj')
+
+                self.model.train()
+                if self.train_cameras:
+                    self.pose_vecs.train()
+
+            self.train_dataset.change_sampling_idx(self.num_pixels)
+
+            if epoch > self.nepoch_freeze:
+                print("Freeze Diffuse Part...")
+                self.model.diffuse_network.eval()
+                self.model.albedo_network.eval()
+
+            for data_index, (indices, model_input, ground_truth) in enumerate(self.train_dataloader):
+
+                model_input["intrinsics"] = model_input["intrinsics"].cuda()
+                model_input["uv"] = model_input["uv"].cuda()
+                model_input["object_mask"] = model_input["object_mask"].cuda()
+
+                if self.train_cameras:
+                    pose_input = self.pose_vecs(indices.cuda())
+                    model_input['pose'] = pose_input
+                else:
+                    model_input['pose'] = model_input['pose'].cuda()
+
+                model_outputs = self.model(model_input)
+                loss_output = self.loss(model_outputs, ground_truth)
+
+                loss = loss_output['loss']
+
+                self.optimizer.zero_grad()
+                if self.train_cameras:
+                    self.optimizer_cam.zero_grad()
+
+                loss.backward()
+
+                self.optimizer.step()
+                if self.train_cameras:
+                    self.optimizer_cam.step()
+
+                print(
+                    '{0} [{1}] ({2}/{3}): loss = {4}, rgb_loss = {5}, normal_loss = {6}, reg_loss = {7}, eikonal_loss = {8}, mask_loss = {9}, alpha = {10}, lr = {11}'
+                        .format(self.expname, epoch, data_index, self.n_batches, loss.item(),
+                                loss_output['rgb_loss'].item(),
+                                loss_output['normal_loss'].item(),
+                                loss_output['reg_loss'].item(),
+                                loss_output['eikonal_loss'].item(),
+                                loss_output['mask_loss'].item(),
+                                self.loss.alpha,
+                                self.scheduler.get_lr()[0]))
+
+            self.scheduler.step()
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/utils/general.py b/reconstruction/PBIDR/code/utils/general.py
new file mode 100644
index 000000000..f76c7007e
--- /dev/null
+++ b/reconstruction/PBIDR/code/utils/general.py
@@ -0,0 +1,66 @@
+import os
+from glob import glob
+import torch
+
+def mkdir_ifnotexists(directory):
+    if not os.path.exists(directory):
+        os.mkdir(directory)
+
+def get_class(kls):
+    parts = kls.split('.')
+    module = ".".join(parts[:-1])
+    m = __import__(module)
+    for comp in parts[1:]:
+        m = getattr(m, comp)
+    return m
+
+def glob_imgs(path):
+    imgs = []
+    for ext in ['*.png', '*.jpg', '*.JPEG', '*.JPG']:
+        imgs.extend(glob(os.path.join(path, ext)))
+    return imgs
+
+def split_input(model_input, total_pixels):
+    '''
+     Split the input to fit Cuda memory for large resolution.
+     Can decrease the value of n_pixels in case of cuda out of memory error.
+     '''
+    n_pixels = 10000
+    split = []
+    for i, indx in enumerate(torch.split(torch.arange(total_pixels).cuda(), n_pixels, dim=0)):
+        data = model_input.copy()
+        data['uv'] = torch.index_select(model_input['uv'], 1, indx)
+        data['object_mask'] = torch.index_select(model_input['object_mask'], 1, indx)
+        split.append(data)
+    return split
+
+def split_input_albedo(model_input, total_pixels):
+    '''
+     Split the input to fit Cuda memory for large resolution.
+     Can decrease the value of n_pixels in case of cuda out of memory error.
+     '''
+    n_pixels = 10000
+    split = []
+    for i, indx in enumerate(torch.split(torch.arange(total_pixels).cuda(), n_pixels, dim=0)):
+        data = model_input.copy()
+        data['uv'] = torch.index_select(model_input['uv'], 1, indx)
+        data['object_mask'] = torch.index_select(model_input['object_mask'], 1, indx)
+        data['rgb'] = torch.index_select(model_input['rgb'], 1, indx)
+        split.append(data)
+    return split
+
+def merge_output(res, total_pixels, batch_size):
+    ''' Merge the split output. '''
+
+    model_outputs = {}
+    for entry in res[0]:
+        if res[0][entry] is None:
+            continue
+        if len(res[0][entry].shape) == 1:
+            model_outputs[entry] = torch.cat([r[entry].reshape(batch_size, -1, 1) for r in res],
+                                             1).reshape(batch_size * total_pixels)
+        else:
+            model_outputs[entry] = torch.cat([r[entry].reshape(batch_size, -1, r[entry].shape[-1]) for r in res],
+                                             1).reshape(batch_size * total_pixels, -1)
+
+    return model_outputs
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/utils/plots.py b/reconstruction/PBIDR/code/utils/plots.py
new file mode 100644
index 000000000..fde881e5b
--- /dev/null
+++ b/reconstruction/PBIDR/code/utils/plots.py
@@ -0,0 +1,424 @@
+import plotly.graph_objs as go
+import plotly.offline as offline
+import numpy as np
+import torch
+from skimage import measure
+import torchvision
+import trimesh
+from PIL import Image
+from utils import rend_util
+import pickle
+
+
+def plot_latent(model, latent, indices, model_outputs ,pose, rgb_gt, path, epoch, img_res, plot_nimgs, max_depth, resolution):
+    # arrange data to plot
+    batch_size, num_samples, _ = rgb_gt.shape
+
+    network_object_mask = model_outputs['network_object_mask']
+    points = model_outputs['points'].reshape(batch_size, num_samples, 3)
+    rgb_eval = model_outputs['rgb_values']
+    rgb_eval = rgb_eval.reshape(batch_size, num_samples, 3)
+
+    depth = torch.ones(batch_size * num_samples).cuda().float() * max_depth
+    depth[network_object_mask] = rend_util.get_depth(points, pose).reshape(-1)[network_object_mask]
+    depth = depth.reshape(batch_size, num_samples, 1)
+    network_object_mask = network_object_mask.reshape(batch_size,-1)
+
+    cam_loc, cam_dir = rend_util.get_camera_for_plot(pose)
+
+    # plot rendered images
+    plot_images(rgb_eval, rgb_gt, path, epoch, plot_nimgs, img_res)
+
+    # plot depth maps
+    plot_depth_maps(depth, path, epoch, plot_nimgs, img_res)
+
+    data = []
+
+    # plot surface
+    surface_traces = get_surface_trace(path=path,
+                                       epoch=epoch,
+                                       sdf=lambda x: model.implicit_network(torch.cat([latent.expand(len(x), -1), x], 1))[:, 0],
+                                       resolution=resolution
+                                       )
+    data.append(surface_traces[0])
+
+    # plot cameras locations
+    for i, loc, dir in zip(indices, cam_loc, cam_dir):
+        data.append(get_3D_quiver_trace(loc.unsqueeze(0), dir.unsqueeze(0), name='camera_{0}'.format(i)))
+
+    for i, p, m in zip(indices, points, network_object_mask):
+        p = p[m]
+        sampling_idx = torch.randperm(p.shape[0])[:2048]
+        p = p[sampling_idx, :]
+
+        val = model.implicit_network(torch.cat([latent.expand(len(p), -1), p], 1))
+        caption = ["sdf: {0} ".format(v[0].item()) for v in val]
+
+        data.append(get_3D_scatter_trace(p, name='intersection_points_{0}'.format(i), caption=caption))
+
+    fig = go.Figure(data=data)
+    scene_dict = dict(xaxis=dict(range=[-3, 3], autorange=False),
+                      yaxis=dict(range=[-3, 3], autorange=False),
+                      zaxis=dict(range=[-3, 3], autorange=False),
+                      aspectratio=dict(x=1, y=1, z=1))
+    fig.update_layout(scene=scene_dict, width=1400, height=1400, showlegend=True)
+    filename = '{0}/surface_{1}.html'.format(path, epoch)
+    offline.plot(fig, filename=filename, auto_open=False)
+
+
+def plot(model, indices, model_outputs ,pose, rgb_gt, path, epoch, img_res, plot_nimgs, max_depth, resolution):
+    # arrange data to plot
+    batch_size, num_samples, _ = rgb_gt.shape
+
+    network_object_mask = model_outputs['network_object_mask']
+    points = model_outputs['points'].reshape(batch_size, num_samples, 3)
+    rgb_eval = model_outputs['rgb_values']
+    rgb_eval = rgb_eval.reshape(batch_size, num_samples, 3)
+
+    depth = torch.ones(batch_size * num_samples).cuda().float() * max_depth
+    depth[network_object_mask] = rend_util.get_depth(points, pose).reshape(-1)[network_object_mask]
+    depth = depth.reshape(batch_size, num_samples, 1)
+    network_object_mask = network_object_mask.reshape(batch_size,-1)
+
+    cam_loc, cam_dir = rend_util.get_camera_for_plot(pose)
+
+    # plot rendered images
+    plot_images(rgb_eval, rgb_gt, path, epoch, plot_nimgs, img_res)
+
+    # plot depth maps
+    plot_depth_maps(depth, path, epoch, plot_nimgs, img_res)
+
+    data = []
+
+    # plot surface
+    surface_traces = get_surface_trace(path=path,
+                                       epoch=epoch,
+                                       sdf=lambda x: model.implicit_network(x)[:, 0],
+                                       resolution=resolution
+                                       )
+    data.append(surface_traces[0])
+
+    # plot cameras locations
+    for i, loc, dir in zip(indices, cam_loc, cam_dir):
+        data.append(get_3D_quiver_trace(loc.unsqueeze(0), dir.unsqueeze(0), name='camera_{0}'.format(i)))
+
+    for i, p, m in zip(indices, points, network_object_mask):
+        p = p[m]
+        sampling_idx = torch.randperm(p.shape[0])[:2048]
+        p = p[sampling_idx, :]
+
+        val = model.implicit_network(p)
+        caption = ["sdf: {0} ".format(v[0].item()) for v in val]
+
+        data.append(get_3D_scatter_trace(p, name='intersection_points_{0}'.format(i), caption=caption))
+
+    fig = go.Figure(data=data)
+    scene_dict = dict(xaxis=dict(range=[-3, 3], autorange=False),
+                      yaxis=dict(range=[-3, 3], autorange=False),
+                      zaxis=dict(range=[-3, 3], autorange=False),
+                      aspectratio=dict(x=1, y=1, z=1))
+    fig.update_layout(scene=scene_dict, width=1400, height=1400, showlegend=True)
+    filename = '{0}/surface_{1}.html'.format(path, epoch)
+    offline.plot(fig, filename=filename, auto_open=False)
+
+
+def get_3D_scatter_trace(points, name='', size=3, caption=None):
+    assert points.shape[1] == 3, "3d scatter plot input points are not correctely shaped "
+    assert len(points.shape) == 2, "3d scatter plot input points are not correctely shaped "
+
+    trace = go.Scatter3d(
+        x=points[:, 0].cpu(),
+        y=points[:, 1].cpu(),
+        z=points[:, 2].cpu(),
+        mode='markers',
+        name=name,
+        marker=dict(
+            size=size,
+            line=dict(
+                width=2,
+            ),
+            opacity=1.0,
+        ), text=caption)
+
+    return trace
+
+
+def get_3D_quiver_trace(points, directions, color='#bd1540', name=''):
+    assert points.shape[1] == 3, "3d cone plot input points are not correctely shaped "
+    assert len(points.shape) == 2, "3d cone plot input points are not correctely shaped "
+    assert directions.shape[1] == 3, "3d cone plot input directions are not correctely shaped "
+    assert len(directions.shape) == 2, "3d cone plot input directions are not correctely shaped "
+
+    trace = go.Cone(
+        name=name,
+        x=points[:, 0].cpu(),
+        y=points[:, 1].cpu(),
+        z=points[:, 2].cpu(),
+        u=directions[:, 0].cpu(),
+        v=directions[:, 1].cpu(),
+        w=directions[:, 2].cpu(),
+        sizemode='absolute',
+        sizeref=0.125,
+        showscale=False,
+        colorscale=[[0, color], [1, color]],
+        anchor="tail"
+    )
+
+    return trace
+
+
+def get_surface_trace(path, epoch, sdf, resolution=100, return_mesh=False):
+    grid = get_grid_uniform(resolution)
+    points = grid['grid_points']
+
+    z = []
+    for i, pnts in enumerate(torch.split(points, 100000, dim=0)):
+        z.append(sdf(pnts).detach().cpu().numpy())
+    z = np.concatenate(z, axis=0)
+
+    if (not (np.min(z) > 0 or np.max(z) < 0)):
+
+        z = z.astype(np.float32)
+
+        verts, faces, normals, values = measure.marching_cubes_lewiner(
+            volume=z.reshape(grid['xyz'][1].shape[0], grid['xyz'][0].shape[0],
+                             grid['xyz'][2].shape[0]).transpose([1, 0, 2]),
+            level=0,
+            spacing=(grid['xyz'][0][2] - grid['xyz'][0][1],
+                     grid['xyz'][0][2] - grid['xyz'][0][1],
+                     grid['xyz'][0][2] - grid['xyz'][0][1]))
+
+        verts = verts + np.array([grid['xyz'][0][0], grid['xyz'][1][0], grid['xyz'][2][0]])
+
+        I, J, K = faces.transpose()
+
+        traces = [go.Mesh3d(x=verts[:, 0], y=verts[:, 1], z=verts[:, 2],
+                            i=I, j=J, k=K, name='implicit_surface',
+                            opacity=1.0)]
+
+        meshexport = trimesh.Trimesh(verts, faces, vertex_normals=-normals)
+        meshexport.export('{0}/surface_{1}.ply'.format(path, epoch), 'ply')
+
+        if return_mesh:
+            return meshexport
+        return traces
+    return None
+
+
+def get_surface_high_res_mesh(sdf, resolution=100):
+    # get low res mesh to sample point cloud
+    grid = get_grid_uniform(100)
+    z = []
+    points = grid['grid_points']
+
+    for i, pnts in enumerate(torch.split(points, 100000, dim=0)):
+        z.append(sdf(pnts).detach().cpu().numpy())
+    z = np.concatenate(z, axis=0)
+
+    z = z.astype(np.float32)
+
+    verts, faces, normals, values = measure.marching_cubes_lewiner(
+        volume=z.reshape(grid['xyz'][1].shape[0], grid['xyz'][0].shape[0],
+                         grid['xyz'][2].shape[0]).transpose([1, 0, 2]),
+        level=0,
+        spacing=(grid['xyz'][0][2] - grid['xyz'][0][1],
+                 grid['xyz'][0][2] - grid['xyz'][0][1],
+                 grid['xyz'][0][2] - grid['xyz'][0][1]))
+
+    verts = verts + np.array([grid['xyz'][0][0], grid['xyz'][1][0], grid['xyz'][2][0]])
+
+    mesh_low_res = trimesh.Trimesh(verts, faces, vertex_normals=-normals)
+    # return mesh_low_res
+
+    components = mesh_low_res.split(only_watertight=False)
+    areas = np.array([c.area for c in components], dtype=np.float)
+    mesh_low_res = components[areas.argmax()]
+
+    recon_pc = trimesh.sample.sample_surface(mesh_low_res, 10000)[0]
+    recon_pc = torch.from_numpy(recon_pc).float().cuda()
+
+    # Center and align the recon pc
+    s_mean = recon_pc.mean(dim=0)
+    s_cov = recon_pc - s_mean
+    s_cov = torch.mm(s_cov.transpose(0, 1), s_cov)
+    vecs = torch.eig(s_cov, True)[1].transpose(0, 1)
+    if torch.det(vecs) < 0:
+        vecs = torch.mm(torch.tensor([[1, 0, 0], [0, 0, 1], [0, 1, 0]]).cuda().float(), vecs)
+    helper = torch.bmm(vecs.unsqueeze(0).repeat(recon_pc.shape[0], 1, 1),
+                       (recon_pc - s_mean).unsqueeze(-1)).squeeze()
+
+    grid_aligned = get_grid(helper.cpu(), resolution)
+
+    grid_points = grid_aligned['grid_points']
+
+    g = []
+    for i, pnts in enumerate(torch.split(grid_points, 100000, dim=0)):
+        g.append(torch.bmm(vecs.unsqueeze(0).repeat(pnts.shape[0], 1, 1).transpose(1, 2),
+                           pnts.unsqueeze(-1)).squeeze() + s_mean)
+    grid_points = torch.cat(g, dim=0)
+
+    # MC to new grid
+    points = grid_points
+    z = []
+    for i, pnts in enumerate(torch.split(points, 100000, dim=0)):
+        z.append(sdf(pnts).detach().cpu().numpy())
+    z = np.concatenate(z, axis=0)
+
+    meshexport = None
+    if (not (np.min(z) > 0 or np.max(z) < 0)):
+
+        z = z.astype(np.float32)
+
+        verts, faces, normals, values = measure.marching_cubes_lewiner(
+            volume=z.reshape(grid_aligned['xyz'][1].shape[0], grid_aligned['xyz'][0].shape[0],
+                             grid_aligned['xyz'][2].shape[0]).transpose([1, 0, 2]),
+            level=0,
+            spacing=(grid_aligned['xyz'][0][2] - grid_aligned['xyz'][0][1],
+                     grid_aligned['xyz'][0][2] - grid_aligned['xyz'][0][1],
+                     grid_aligned['xyz'][0][2] - grid_aligned['xyz'][0][1]))
+
+        verts = torch.from_numpy(verts).cuda().float()
+        verts = torch.bmm(vecs.unsqueeze(0).repeat(verts.shape[0], 1, 1).transpose(1, 2),
+                   verts.unsqueeze(-1)).squeeze()
+        verts = (verts + grid_points[0]).cpu().numpy()
+
+        meshexport = trimesh.Trimesh(verts, faces, vertex_normals=-normals)
+
+    return meshexport
+
+
+def get_displacement_mesh(model):
+
+    def get_detailed_mesh(input):
+        origin_mesh = input.copy()
+        mesh_points = torch.tensor(np.array(origin_mesh.vertices).astype(np.float32)).cuda().requires_grad_(True)
+        sdfs = model.implicit_network(mesh_points)[:, 0:1]
+        sdf_np = sdfs.detach().cpu().numpy()
+        vetices_normal = origin_mesh.vertex_normals
+        new_vertices = -1.0 * sdf_np * vetices_normal + np.array(origin_mesh.vertices)
+        new_mesh = trimesh.Trimesh(vertices=new_vertices, faces=origin_mesh.faces, process=False, visual=origin_mesh.visual)
+        print('Detailed mesh created!')
+        return new_mesh
+
+    orimesh = model.mesh
+    subdivision_mesh = orimesh.subdivide()
+    new_mesh = get_detailed_mesh(orimesh)
+    submesh_full = get_detailed_mesh(subdivision_mesh)
+
+    return new_mesh, submesh_full
+
+def get_displacement_animation(model):
+
+    def get_detailed_mesh(aaa_mesh):
+        origin_mesh = aaa_mesh.copy()
+        mesh_points = torch.tensor(np.array(origin_mesh.vertices).astype(np.float32)).cuda().requires_grad_(True)
+        sdfs = model.implicit_network(mesh_points)[:, 0:1]
+        sdf_np = sdfs.detach().cpu().numpy()
+
+        print('Detailed mesh created!')
+        return sdf_np
+
+    orimesh = model.mesh
+    subdi_mesh_full = orimesh.subdivide()
+    sdf_np0 = get_detailed_mesh(orimesh)
+    sdf_np1 = get_detailed_mesh(subdi_mesh_full)
+
+    return sdf_np0, sdf_np1
+
+
+def get_NormalMaps(model):
+
+    def get_detailed_mesh(input):
+        origin_mesh = input.copy()
+        mesh_points = torch.tensor(np.array(origin_mesh.vertices).astype(np.float32)).cuda().requires_grad_(True)
+        detailed_normal = model.implicit_network.gradient(mesh_points)
+        detailed_normal = detailed_normal.squeeze().detach().cpu().numpy()
+        vetices_normal = np.asarray(origin_mesh.vertex_normals)
+        return vetices_normal, detailed_normal
+
+    orimesh = model.mesh
+    smooth1, detail1 = get_detailed_mesh(orimesh)
+    return smooth1, detail1
+
+def get_grid_uniform(resolution):
+    x = np.linspace(-1.0, 1.0, resolution)
+    y = x
+    z = x
+
+    xx, yy, zz = np.meshgrid(x, y, z)
+    grid_points = torch.tensor(np.vstack([xx.ravel(), yy.ravel(), zz.ravel()]).T, dtype=torch.float)
+
+    return {"grid_points": grid_points.cuda(),
+            "shortest_axis_length": 2.0,
+            "xyz": [x, y, z],
+            "shortest_axis_index": 0}
+
+def get_grid(points, resolution):
+    eps = 0.2
+    input_min = torch.min(points, dim=0)[0].squeeze().numpy()
+    input_max = torch.max(points, dim=0)[0].squeeze().numpy()
+
+    bounding_box = input_max - input_min
+    shortest_axis = np.argmin(bounding_box)
+    if (shortest_axis == 0):
+        x = np.linspace(input_min[shortest_axis] - eps,
+                        input_max[shortest_axis] + eps, resolution)
+        length = np.max(x) - np.min(x)
+        y = np.arange(input_min[1] - eps, input_max[1] + length / (x.shape[0] - 1) + eps, length / (x.shape[0] - 1))
+        z = np.arange(input_min[2] - eps, input_max[2] + length / (x.shape[0] - 1) + eps, length / (x.shape[0] - 1))
+    elif (shortest_axis == 1):
+        y = np.linspace(input_min[shortest_axis] - eps,
+                        input_max[shortest_axis] + eps, resolution)
+        length = np.max(y) - np.min(y)
+        x = np.arange(input_min[0] - eps, input_max[0] + length / (y.shape[0] - 1) + eps, length / (y.shape[0] - 1))
+        z = np.arange(input_min[2] - eps, input_max[2] + length / (y.shape[0] - 1) + eps, length / (y.shape[0] - 1))
+    elif (shortest_axis == 2):
+        z = np.linspace(input_min[shortest_axis] - eps,
+                        input_max[shortest_axis] + eps, resolution)
+        length = np.max(z) - np.min(z)
+        x = np.arange(input_min[0] - eps, input_max[0] + length / (z.shape[0] - 1) + eps, length / (z.shape[0] - 1))
+        y = np.arange(input_min[1] - eps, input_max[1] + length / (z.shape[0] - 1) + eps, length / (z.shape[0] - 1))
+
+    xx, yy, zz = np.meshgrid(x, y, z)
+    grid_points = torch.tensor(np.vstack([xx.ravel(), yy.ravel(), zz.ravel()]).T, dtype=torch.float).cuda()
+    return {"grid_points": grid_points,
+            "shortest_axis_length": length,
+            "xyz": [x, y, z],
+            "shortest_axis_index": shortest_axis}
+
+def plot_depth_maps(depth_maps, path, epoch, plot_nrow, img_res):
+    depth_maps_plot = lin2img(depth_maps, img_res)
+
+    tensor = torchvision.utils.make_grid(depth_maps_plot.repeat(1, 3, 1, 1),
+                                         scale_each=True,
+                                         normalize=True,
+                                         nrow=plot_nrow).cpu().detach().numpy()
+    tensor = tensor.transpose(1, 2, 0)
+    scale_factor = 255
+    tensor = (tensor * scale_factor).astype(np.uint8)
+
+    img = Image.fromarray(tensor)
+    img.save('{0}/depth_{1}.png'.format(path, epoch))
+
+def plot_images(rgb_points, ground_true, path, epoch, plot_nrow, img_res):
+    ground_true = (ground_true.cuda() + 1.) / 2.
+    rgb_points = (rgb_points + 1. ) / 2.
+
+    output_vs_gt = torch.cat((rgb_points, ground_true), dim=0)
+    output_vs_gt_plot = lin2img(output_vs_gt, img_res)
+
+    tensor = torchvision.utils.make_grid(output_vs_gt_plot,
+                                         scale_each=False,
+                                         normalize=False,
+                                         nrow=plot_nrow).cpu().detach().numpy()
+
+    tensor = tensor.transpose(1, 2, 0)
+    scale_factor = 255
+    tensor = (tensor * scale_factor).astype(np.uint8)
+
+    img = Image.fromarray(tensor)
+    img.save('{0}/rendering_{1}.png'.format(path, epoch))
+
+def lin2img(tensor, img_res):
+    batch_size, num_samples, channels = tensor.shape
+    return tensor.permute(0, 2, 1).view(batch_size, channels, img_res[0], img_res[1])
\ No newline at end of file
diff --git a/reconstruction/PBIDR/code/utils/rend_util.py b/reconstruction/PBIDR/code/utils/rend_util.py
new file mode 100644
index 000000000..7a255e2d0
--- /dev/null
+++ b/reconstruction/PBIDR/code/utils/rend_util.py
@@ -0,0 +1,192 @@
+import numpy as np
+import imageio
+import skimage
+import cv2
+import torch
+from torch.nn import functional as F
+
+def load_rgb(path):
+    img = imageio.imread(path)
+    img = skimage.img_as_float32(img)
+
+    # pixel values between [-1,1]
+    img -= 0.5
+    img *= 2.
+    img = img.transpose(2, 0, 1)
+    return img
+
+def load_mask(path):
+    alpha = imageio.imread(path, as_gray=True)
+    alpha = skimage.img_as_float32(alpha)
+    object_mask = alpha > 127.5
+
+    return object_mask
+
+def load_mask_white_bg(path):
+    alpha = imageio.imread(path, as_gray=True)
+    alpha = skimage.img_as_float32(alpha)
+    object_mask = alpha < 250.5
+
+    return object_mask
+
+def load_K_Rt_from_P(filename, P=None):
+    if P is None:
+        lines = open(filename).read().splitlines()
+        if len(lines) == 4:
+            lines = lines[1:]
+        lines = [[x[0], x[1], x[2], x[3]] for x in (x.split(" ") for x in lines)]
+        P = np.asarray(lines).astype(np.float32).squeeze()
+
+    out = cv2.decomposeProjectionMatrix(P)
+    K = out[0]
+    R = out[1]
+    t = out[2]
+
+    K = K/K[2,2]
+    intrinsics = np.eye(4)
+    intrinsics[:3, :3] = K
+
+    pose = np.eye(4, dtype=np.float32)
+    to_gl = np.eye(3, dtype=np.float32)
+    to_gl[0, 0] = -1.
+    to_gl[1, 1] = -1.
+    pose[:3, :3] = np.dot(R.transpose(), to_gl)
+    pose[:3,3] = (t[:3] / t[3])[:,0]
+
+    return intrinsics, pose
+
+def get_camera_params(uv, pose, intrinsics):
+    if pose.shape[1] == 7: #In case of quaternion vector representation
+        cam_loc = pose[:, 4:]
+        R = quat_to_rot(pose[:,:4])
+        p = torch.eye(4).repeat(pose.shape[0],1,1).cuda().float()
+        p[:, :3, :3] = R
+        p[:, :3, 3] = cam_loc
+    else: # In case of pose matrix representation
+        cam_loc = pose[:, :3, 3]
+        p = pose
+
+    batch_size, num_samples, _ = uv.shape
+
+    depth = torch.ones((batch_size, num_samples)).cuda()
+    x_cam = uv[:, :, 0].view(batch_size, -1)
+    y_cam = uv[:, :, 1].view(batch_size, -1)
+    z_cam = depth.view(batch_size, -1)
+
+    pixel_points_cam = lift(x_cam, y_cam, z_cam, intrinsics=intrinsics)
+
+    # permute for batch matrix product
+    pixel_points_cam = pixel_points_cam.permute(0, 2, 1)
+
+    world_coords = torch.bmm(p, pixel_points_cam).permute(0, 2, 1)[:, :, :3]
+    ray_dirs = world_coords - cam_loc[:, None, :]
+    ray_dirs = F.normalize(ray_dirs, dim=2)
+
+    return ray_dirs, cam_loc
+
+def get_camera_for_plot(pose):
+    if pose.shape[1] == 7: #In case of quaternion vector representation
+        cam_loc = pose[:, 4:].detach()
+        R = quat_to_rot(pose[:,:4].detach())
+    else: # In case of pose matrix representation
+        cam_loc = pose[:, :3, 3]
+        R = pose[:, :3, :3]
+    cam_dir = R[:, :3, 2]
+    return cam_loc, cam_dir
+
+def lift(x, y, z, intrinsics):
+    # parse intrinsics
+    intrinsics = intrinsics.cuda()
+    fx = intrinsics[:, 0, 0]
+    fy = intrinsics[:, 1, 1]
+    cx = intrinsics[:, 0, 2]
+    cy = intrinsics[:, 1, 2]
+    sk = intrinsics[:, 0, 1]
+
+    x_lift = (x - cx.unsqueeze(-1) + cy.unsqueeze(-1)*sk.unsqueeze(-1)/fy.unsqueeze(-1) - sk.unsqueeze(-1)*y/fy.unsqueeze(-1)) / fx.unsqueeze(-1) * z
+    y_lift = (y - cy.unsqueeze(-1)) / fy.unsqueeze(-1) * z
+
+    # homogeneous
+    return torch.stack((x_lift, y_lift, z, torch.ones_like(z).cuda()), dim=-1)
+
+def quat_to_rot(q):
+    batch_size, _ = q.shape
+    q = F.normalize(q, dim=1)
+    R = torch.ones((batch_size, 3,3)).cuda()
+    qr=q[:,0]
+    qi = q[:, 1]
+    qj = q[:, 2]
+    qk = q[:, 3]
+    R[:, 0, 0]=1-2 * (qj**2 + qk**2)
+    R[:, 0, 1] = 2 * (qj *qi -qk*qr)
+    R[:, 0, 2] = 2 * (qi * qk + qr * qj)
+    R[:, 1, 0] = 2 * (qj * qi + qk * qr)
+    R[:, 1, 1] = 1-2 * (qi**2 + qk**2)
+    R[:, 1, 2] = 2*(qj*qk - qi*qr)
+    R[:, 2, 0] = 2 * (qk * qi-qj * qr)
+    R[:, 2, 1] = 2 * (qj*qk + qi*qr)
+    R[:, 2, 2] = 1-2 * (qi**2 + qj**2)
+    return R
+
+def rot_to_quat(R):
+    batch_size, _,_ = R.shape
+    q = torch.ones((batch_size, 4)).cuda()
+
+    R00 = R[:, 0,0]
+    R01 = R[:, 0, 1]
+    R02 = R[:, 0, 2]
+    R10 = R[:, 1, 0]
+    R11 = R[:, 1, 1]
+    R12 = R[:, 1, 2]
+    R20 = R[:, 2, 0]
+    R21 = R[:, 2, 1]
+    R22 = R[:, 2, 2]
+
+    q[:,0]=torch.sqrt(1.0+R00+R11+R22)/2
+    q[:, 1]=(R21-R12)/(4*q[:,0])
+    q[:, 2] = (R02 - R20) / (4 * q[:, 0])
+    q[:, 3] = (R10 - R01) / (4 * q[:, 0])
+    return q
+
+def get_sphere_intersection(cam_loc, ray_directions, r = 1.0):
+    # Input: n_images x 4 x 4 ; n_images x n_rays x 3
+    # Output: n_images * n_rays x 2 (close and far) ; n_images * n_rays
+
+    n_imgs, n_pix, _ = ray_directions.shape
+
+    cam_loc = cam_loc.unsqueeze(-1)
+    ray_cam_dot = torch.bmm(ray_directions, cam_loc).squeeze()
+    under_sqrt = ray_cam_dot ** 2 - (cam_loc.norm(2,1) ** 2 - r ** 2)
+
+    under_sqrt = under_sqrt.reshape(-1)
+    mask_intersect = under_sqrt > 0
+
+    sphere_intersections = torch.zeros(n_imgs * n_pix, 2).cuda().float()
+    sphere_intersections[mask_intersect] = torch.sqrt(under_sqrt[mask_intersect]).unsqueeze(-1) * torch.Tensor([-1, 1]).cuda().float()
+    sphere_intersections[mask_intersect] -= ray_cam_dot.reshape(-1)[mask_intersect].unsqueeze(-1)
+
+    sphere_intersections = sphere_intersections.reshape(n_imgs, n_pix, 2)
+    sphere_intersections = sphere_intersections.clamp_min(0.0)
+    mask_intersect = mask_intersect.reshape(n_imgs, n_pix)
+
+    return sphere_intersections, mask_intersect
+
+def get_depth(points, pose):
+    ''' Retruns depth from 3D points according to camera pose '''
+    batch_size, num_samples, _ = points.shape
+    if pose.shape[1] == 7:  # In case of quaternion vector representation
+        cam_loc = pose[:, 4:]
+        R = quat_to_rot(pose[:, :4])
+        pose = torch.eye(4).unsqueeze(0).repeat(batch_size, 1, 1).cuda().float()
+        pose[:, :3, 3] = cam_loc
+        pose[:, :3, :3] = R
+
+    points_hom = torch.cat((points, torch.ones((batch_size, num_samples, 1)).cuda()), dim=2)
+
+    # permute for batch matrix product
+    points_hom = points_hom.permute(0, 2, 1)
+
+    points_cam = torch.inverse(pose).bmm(points_hom)
+    depth = points_cam[:, 2, :][:, :, None]
+    return depth
+
diff --git a/reconstruction/PBIDR/figures/overview.jpg b/reconstruction/PBIDR/figures/overview.jpg
new file mode 100644
index 000000000..6cceaee18
Binary files /dev/null and b/reconstruction/PBIDR/figures/overview.jpg differ
diff --git a/reconstruction/PBIDR/requirements.txt b/reconstruction/PBIDR/requirements.txt
new file mode 100644
index 000000000..be933a081
--- /dev/null
+++ b/reconstruction/PBIDR/requirements.txt
@@ -0,0 +1,10 @@
+vtk
+numpy
+opencv-python
+scikit-image
+scipy
+Pillow
+argparse
+GPUtil
+pyhocon
+plotly
\ No newline at end of file
diff --git a/reconstruction/gaze/README.md b/reconstruction/gaze/README.md
new file mode 100644
index 000000000..c33cf9bfb
--- /dev/null
+++ b/reconstruction/gaze/README.md
@@ -0,0 +1,44 @@
+# Generalizing Gaze Estimation with Weak-Supervision from Synthetic Views
+
+The implementation of [Arxiv paper](https://arxiv.org/abs/2212.02997) for gaze estimation task.
+
+
+## Preparation
+
+1. Download the [dataset](https://drive.google.com/file/d/1erYIoTCbXk1amofJ6yTGhbpmsovWrrva/view?usp=sharing) and put it under ``data/``
+
+2. Download [eyes3d.pkl](https://drive.google.com/file/d/1as7_ew6kEFTHpcrlk8QKvgFJJ8cKzM3q/view?usp=sharing) and put it under ``assets/``
+
+3. Download [pretrained checkpoint](https://drive.google.com/file/d/1cqmChXSnTwUpk3jD7JLpZKHOuBLlC3_N/view?usp=sharing) and put it under ``assets/``
+
+4. Install libraries:
+   ```
+   pip install timm pytorch-lightning==1.8.1 albumentations==1.3.0
+   ```
+   
+## Testing with pre-trained model
+
+  After downloading the pre-trained checkpoint above,
+
+  ```
+  python test_gaze.py assets/latest_a.ckpt
+  ```
+
+## Training
+
+  ```
+  python trainer_gaze.py
+  ```
+
+## Results
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/gaze_0.png?raw=true" width="800" alt=""/>
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/gaze_1.png?raw=true" width="800" alt=""/>
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/gaze_2.png?raw=true" width="800" alt=""/>
+
+
+
+
+
diff --git a/reconstruction/gaze/datasets/__init__.py b/reconstruction/gaze/datasets/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/reconstruction/gaze/datasets/augs.py b/reconstruction/gaze/datasets/augs.py
new file mode 100644
index 000000000..1e701d9fa
--- /dev/null
+++ b/reconstruction/gaze/datasets/augs.py
@@ -0,0 +1,40 @@
+import numpy as np
+import albumentations as A
+from albumentations.core.transforms_interface import ImageOnlyTransform
+
+class RectangleBorderAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            fill_value = 0,
+            limit = 0.3,
+            always_apply=False,
+            p=1.0,
+            ):
+        super(RectangleBorderAugmentation, self).__init__(always_apply, p)
+        assert limit>0.0 and limit<1.0
+        self.fill_value = 0
+        self.limit = limit
+
+
+    def apply(self, image, border_size_limit, **params):
+        assert len(border_size_limit)==4
+        border_size = border_size_limit.copy()
+        border_size[0] *= image.shape[1]
+        border_size[2] *= image.shape[1]
+        border_size[1] *= image.shape[0]
+        border_size[3] *= image.shape[0]
+        border_size = border_size.astype(np.int)
+        image[:,:border_size[0],:] = self.fill_value
+        image[:border_size[1],:,:] = self.fill_value
+        image[:,border_size[2]:,:] = self.fill_value
+        image[border_size[3]:,:,:] = self.fill_value
+        return image
+
+    def get_params(self):
+        border_size_limit = np.random.uniform(0.0, self.limit, size=4)
+        return {'border_size_limit': border_size_limit}
+
+    def get_transform_init_args_names(self):
+        return ('fill_value', 'limit')
+
diff --git a/reconstruction/gaze/datasets/dataset_gaze.py b/reconstruction/gaze/datasets/dataset_gaze.py
new file mode 100644
index 000000000..fb86a81e6
--- /dev/null
+++ b/reconstruction/gaze/datasets/dataset_gaze.py
@@ -0,0 +1,184 @@
+import os
+import os.path as osp
+import queue as Queue
+import mxnet as mx
+import pickle
+import threading
+import logging
+import numpy as np
+import insightface
+from insightface.utils import face_align
+import torch
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+import cv2
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+from .augs import RectangleBorderAugmentation
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank,
+                                                 non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+
+class GazeDataset(Dataset):
+    def __init__(self, root_dir, is_train):
+        super(GazeDataset, self).__init__()
+
+        #self.local_rank = local_rank
+        self.is_train = is_train
+        self.input_size = 160
+        #self.num_kps = 68
+        transform_list = []
+        if is_train:
+            transform_list += \
+                [
+                    A.ColorJitter(brightness=0.3, contrast=0.3, p=0.5),
+                    A.ToGray(p=0.1),
+                    A.ISONoise(p=0.1),
+                    A.MedianBlur(blur_limit=(1,7), p=0.1),
+                    A.GaussianBlur(blur_limit=(1,7), p=0.1),
+                    A.MotionBlur(blur_limit=(5,13), p=0.1),
+                    A.ImageCompression(quality_lower=10, quality_upper=90, p=0.05),
+                    A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.2, rotate_limit=30, interpolation=cv2.INTER_LINEAR, 
+                        border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.6),
+                    #A.HorizontalFlip(p=0.5),
+                    RectangleBorderAugmentation(limit=0.2, fill_value=0, p=0.1),
+                ]
+        transform_list += \
+            [
+                A.geometric.resize.Resize(self.input_size, self.input_size, interpolation=cv2.INTER_LINEAR, always_apply=True),
+                A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+                ToTensorV2(),
+            ]
+        self.transform = A.ReplayCompose(
+            transform_list,
+            keypoint_params=A.KeypointParams(format='xy', remove_invisible=False)
+        )
+        self.root_dir = root_dir
+        if is_train:
+            path_imgrec = os.path.join(root_dir, 'train.rec')
+            path_imgidx = os.path.join(root_dir, 'train.idx')
+        else:
+            path_imgrec = os.path.join(root_dir, 'val.rec')
+            path_imgidx = os.path.join(root_dir, 'val.idx')
+        self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+        self.imgidx = np.array(list(self.imgrec.keys))
+        logging.info('len:%d'%len(self.imgidx))
+        print('!!!len:%d'%len(self.imgidx))
+        #self.num_face = 1103
+        self.num_eye = 481
+
+    def __len__(self):
+        return len(self.imgidx)
+
+    def __getitem__(self, index):
+        idx = self.imgidx[index]
+        s = self.imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        img = mx.image.imdecode(img).asnumpy() #rgb numpy
+        y = np.array(header.label, dtype=np.float32).reshape( (-1, 3) )
+        #print('!!!', y.shape)
+        assert y.shape[0]==self.num_eye*2
+        eye_l = y[:self.num_eye,:]
+        eye_r = y[self.num_eye:,:]
+        mean_z_l = np.mean(eye_l[:32,2])
+        mean_z_r = np.mean(eye_r[:32,2])
+        std_z_l = np.max(np.abs(eye_l[:32,2]))
+        std_z_r = np.max(np.abs(eye_r[:32,2]))
+        eye_l[:,2] -= mean_z_l
+        eye_r[:,2] -= mean_z_r
+        eye_l[:,2] /= std_z_l
+        eye_r[:,2] /= std_z_r
+        #print('!!!', np.max(eye_l[:,2]), np.min(eye_l[:,2]))
+        y = np.concatenate( (eye_l, eye_r), axis=0)
+        #y[:,2] /= 100.0
+        
+        #if self.is_train:
+        #    black_edge = np.random.randint(img.shape[1]//3, size=4)
+        #    if np.random.random()<0.5:
+        #        img[:black_edge[0],:,:] = 0
+        #        img[black_edge[1]*-1:,:,:] = 0
+        #        img[:,:black_edge[2],:] = 0
+        #        img[:,black_edge[3]*-1:,:] = 0
+        #label = torch.tensor(y, dtype=torch.float32)
+        #label = y
+        kps_xy = []
+        kps_z = []
+        for i in range(y.shape[0]):
+            kps_xy.append( (y[i][0], y[i][1]) )
+            kps_z.append(y[i][2])
+        if self.transform is not None:
+            #sample = self.transform(image=sample)['image']
+            #t = self.transform(image=img, keypoints=label, class_labels=self.class_labels, class_sides=self.class_sides)
+            t = self.transform(image=img, keypoints=kps_xy)
+            flipped = False
+            #print(t.keys())
+            #print('!!!flipped:', flipped)
+            img = t['image']
+            label_xy = t['keypoints']
+            label_xy = np.array(label_xy, dtype=np.float32)
+            label_xy /= (self.input_size/2)
+            label_xy -= 1.0
+            label_z = np.array(kps_z, dtype=np.float32).reshape((-1,1))
+            #label_z /= (self.input_size/2)
+            label = np.concatenate( (label_xy, label_z), axis=1)
+            #label = label.flatten()
+            label = torch.tensor(label, dtype=torch.float32)
+        #print('label:', label.shape)
+        return img, label
+
diff --git a/reconstruction/gaze/models.py b/reconstruction/gaze/models.py
new file mode 100644
index 000000000..9db5dc252
--- /dev/null
+++ b/reconstruction/gaze/models.py
@@ -0,0 +1,75 @@
+import pytorch_lightning as pl
+from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import LearningRateMonitor
+from pytorch_lightning.loggers import TensorBoardLogger
+import torch
+import torch.nn as nn
+import timm
+
+class GazeModel(pl.LightningModule):
+    def __init__(self, backbone, epoch):
+        super().__init__()
+        self.save_hyperparameters()
+        self.backbone = timm.create_model(backbone, num_classes=481*2*3)
+        self.epoch = epoch
+        #self.loss = nn.MSELoss(reduction='mean')
+        self.loss = nn.L1Loss(reduction='mean')
+        #self.hard_mining = False
+        self.hard_mining = False
+        self.num_face = 1103
+        self.num_eye = 481*2
+
+    def forward(self, x):
+        # use forward for inference/predictions
+        y = self.backbone(x)
+        return y
+
+    def cal_loss(self, y_hat, y, hm=False):
+        bs = y.size(0)
+        y_hat = y_hat.view( (bs,-1,3) )
+        loss = torch.abs(y_hat - y) #(B,K,3)
+        loss[:,:,2] *= 0.5
+        if hm:
+            loss = torch.mean(loss, dim=(1,2)) #(B,)
+            loss, _ = torch.topk(loss, k=int(bs*0.25), largest=True)
+            #B = len(loss)
+            #S = int(B*0.5)
+            #loss, _ = torch.sort(loss, descending=True)
+            #loss = loss[:S]
+        loss = torch.mean(loss) * 20.0
+        return loss
+
+    def training_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        loss = self.cal_loss(y_hat, y, self.hard_mining)
+        self.log('train_loss', loss, on_epoch=True)
+        return loss
+
+    def validation_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        loss = self.cal_loss(y_hat, y)
+        self.log('val_loss', loss, on_step=True)
+
+    def test_step(self, batch, batch_idx):
+        x, y = batch
+        y_hat = self.backbone(x)
+        loss = self.cal_loss(y_hat, y)
+        self.log('test_loss', loss)
+
+    def configure_optimizers(self):
+        #return torch.optim.Adam(self.parameters(), lr=0.0002)
+        opt = torch.optim.SGD(self.parameters(), lr = 0.1, momentum=0.9, weight_decay = 0.0005)
+        epoch_steps = [int(self.epoch*0.4), int(self.epoch*0.7), int(self.epoch*0.9)]
+        print('epoch_steps:', epoch_steps)
+        def lr_step_func(epoch):
+            return 0.1 ** len([m for m in epoch_steps if m <= epoch])
+        scheduler = torch.optim.lr_scheduler.LambdaLR(
+                optimizer=opt, lr_lambda=lr_step_func)
+        lr_scheduler = {
+                'scheduler': scheduler,
+                'name': 'learning_rate',
+                'interval':'epoch',
+                'frequency': 1}
+        return [opt], [lr_scheduler]
diff --git a/reconstruction/gaze/test_gaze.py b/reconstruction/gaze/test_gaze.py
new file mode 100644
index 000000000..a4ebc9fa3
--- /dev/null
+++ b/reconstruction/gaze/test_gaze.py
@@ -0,0 +1,239 @@
+
+from models import GazeModel
+import sys
+import glob
+import torch
+import os
+import os.path as osp
+import numpy as np
+import cv2
+import os.path as osp
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.utils import face_align
+import menpo.io as mio
+from menpo.image import Image
+from menpo.shape import PointCloud, TriMesh
+
+
+def angles_from_vec(vec):
+    x, y, z = -vec[2], vec[1], -vec[0]
+    theta = np.arctan2(y, x)
+    phi = np.arctan2(np.sqrt(x**2 + y**2), z) - np.pi/2
+    theta_x, theta_y = phi, theta
+    return theta_x, theta_y
+
+def vec_from_eye(eye, iris_lms_idx):
+    p_iris = eye[iris_lms_idx] - eye[:32].mean(axis=0)
+    vec = p_iris.mean(axis=0)
+    vec /= np.linalg.norm(vec, axis=0)
+    return vec
+
+def angles_and_vec_from_eye(eye, iris_lms_idx):
+
+    vec = vec_from_eye(eye, iris_lms_idx)
+    theta_x, theta_y = angles_from_vec(vec)
+    return theta_x, theta_y, vec
+
+def vec_from_angles(rx, ry):
+    rx = np.deg2rad(rx)
+    ry = np.deg2rad(ry)
+    x1 = np.sin(np.pi/2 + rx) * np.cos(ry) 
+    y1 = np.sin(np.pi/2 + rx) * np.sin(ry)
+    z1 = np.cos(np.pi/2 + rx)
+    x, y, z = -z1, y1, -x1
+    vec = np.array([x, y, z])
+    vec /= np.linalg.norm(vec, axis=0)
+    return vec
+
+
+class GazeHandler():
+    def __init__(self, ckpt_path, res_eyes_path='assets/eyes3d.pkl'):
+        R = 12.0
+        eyes_mean = mio.import_pickle(res_eyes_path)
+        idxs481 = eyes_mean['mask481']['idxs']
+        self.tri481 = eyes_mean['mask481']['trilist']
+        self.iris_idx_481 = eyes_mean['mask481']['idxs_iris']
+
+        self.mean_l = eyes_mean['left_points'][idxs481][:, [0, 2, 1]]
+        self.mean_r = eyes_mean['right_points'][idxs481][:, [0, 2, 1]]
+
+        self.num_face = 1103
+        self.num_eye = 481
+        self.app = FaceAnalysis()
+        det_size = 320
+        self.app.prepare(ctx_id=0, det_size=(det_size, det_size))
+        self.input_size = 160
+        self.model = GazeModel.load_from_checkpoint(ckpt_path).cuda()
+        self.model.eval()
+
+
+
+    def draw_item(self, eimg, item):
+        #bbox, kps, eye_kps = item
+        eye_kps = item
+        #eye_kps[:,2] *= 10.0
+        eye_l = eye_kps[:self.num_eye,:]
+        eye_r = eye_kps[self.num_eye:,:]
+        for _eye in [eye_l, eye_r]:
+            tmp = _eye[:,0].copy()
+            _eye[:,0] = _eye[:,1].copy()
+            _eye[:,1] = tmp
+        #img_crp, trf_crp = image.crop_to_pointcloud_proportion(PointCloud(np.concatenate((eye_l[:, :2], eye_r[:, :2]))), 0.4,  return_transform=True)
+        #img_crp.view(1)
+        #trf_crp.pseudoinverse().apply(TriMesh(eye_l, tri481).with_dims([0, 1])).view(1, marker_size=0.01, line_width=0.1)
+        #trf_crp.pseudoinverse().apply(PointCloud(eye_l[iris_idx_481]).with_dims([0, 1])).view(1)
+        #trf_crp.pseudoinverse().apply(TriMesh(eye_r, tri481).with_dims([0, 1])).view(1, marker_size=0.01, line_width=0.1)
+        #trf_crp.pseudoinverse().apply(PointCloud(eye_r[iris_idx_481]).with_dims([0, 1])).view(1)
+
+        for _eye in [eye_l, eye_r]:
+            _kps = _eye[self.iris_idx_481,:].astype(np.int)
+            for l in range(_kps.shape[0]):
+                color = (0, 255, 0)
+                cv2.circle(eimg, (_kps[l][1], _kps[l][0]), 4, color, 4)
+            #print(tri481.shape)
+            for _tri in self.tri481:
+                color = (0, 0, 255)
+                for k in range(3):
+                    ix = _tri[k]
+                    iy = _tri[(k+1)%3]
+                    x = _eye[ix,:2].astype(np.int)[::-1]
+                    y = _eye[iy,:2].astype(np.int)[::-1]
+                    cv2.line(eimg, x, y, color, 1)
+
+        theta_x_l, theta_y_l, vec_l = angles_and_vec_from_eye(eye_l, self.iris_idx_481)
+        theta_x_r, theta_y_r, vec_r = angles_and_vec_from_eye(eye_r, self.iris_idx_481)
+        gaze_pred = np.array([(theta_x_l + theta_x_r) / 2, (theta_y_l + theta_y_r) / 2])
+
+        diag = np.sqrt(float(eimg.shape[0]*eimg.shape[1]))
+
+        #img_crp, trf_crp = image.crop_to_pointcloud_proportion(PointCloud(lms[17:-20]), 0.1, return_transform=True)
+        #diag = img_crp.diagonal()
+
+        eye_pos_left = eye_l[self.iris_idx_481].mean(axis=0)[[0, 1]]
+        eye_pos_right = eye_r[self.iris_idx_481].mean(axis=0)[[0, 1]]
+
+        ##fig = plt.figure(0)
+        ##image.view(0, figure_size=(4,4))
+        #PointCloud(eye_l[iris_idx_481]).with_dims([0, 1]).view(0, marker_size=3, figure_size=(4,4))
+        #PointCloud(eye_r[iris_idx_481]).with_dims([0, 1]).view(0, marker_size=3, figure_size=(4,4))
+
+        ## pred ---
+        gaze_pred = np.array([theta_x_l, theta_y_l])
+        dx = 0.4*diag * np.sin(gaze_pred[1])
+        dy = 0.4*diag * np.sin(gaze_pred[0])
+        x = np.array([eye_pos_left[1], eye_pos_left[0]])
+        y = x.copy()
+        y[0] += dx
+        y[1] += dy
+        x = x.astype(np.int)
+        y = y.astype(np.int)
+        color = (0,255,255)
+        cv2.line(eimg, x, y, color, 2)
+
+        gaze_pred = np.array([theta_x_r, theta_y_r])
+        dx = 0.4*diag * np.sin(gaze_pred[1])
+        dy = 0.4*diag * np.sin(gaze_pred[0])
+        x = np.array([eye_pos_right[1], eye_pos_right[0]])
+        y = x.copy()
+        y[0] += dx
+        y[1] += dy
+        x = x.astype(np.int)
+        y = y.astype(np.int)
+        color = (0,255,255)
+        cv2.line(eimg, x, y, color, 2)
+        return eimg
+
+    def draw_on(self, eimg, results):
+        face_sizes = [ (x[0][2] - x[0][0]) for x in results]
+        max_index = np.argmax(face_sizes)
+        max_face_size = face_sizes[max_index]
+        rescale = 300.0 / max_face_size
+        #print(max_face_size, rescale)
+        oimg = eimg.copy()
+        eimg = cv2.resize(eimg, None, fx=rescale, fy=rescale)
+        for pred in results:
+            _, _, eye_kps = pred
+            eye_kps = eye_kps.copy()
+            eye_kps *= rescale
+            eimg = self.draw_item(eimg, eye_kps)
+        eimg = cv2.resize(eimg, (oimg.shape[1], oimg.shape[0]))
+        return eimg
+        pred_max = results[max_index]
+        bbox, kps, eye_kps = pred_max
+        width = bbox[2] - bbox[0]
+        center = (kps[0]+kps[1]) / 2.0
+        #_size = np.abs(kps[1][0] - kps[0][0]) * 1.5
+        _size = max(width/1.5, np.abs(kps[1][0] - kps[0][0]) ) * 1.5
+        rotate = 0
+        _scale = self.input_size  / _size
+        aimg, M = face_align.transform(oimg, center, self.input_size, _scale, rotate)
+        eye_kps = face_align.trans_points(eye_kps, M)
+        center_eye_rescale = 4.0
+        aimg = cv2.resize(aimg, None, fx=center_eye_rescale, fy=center_eye_rescale)
+        eye_kps *= center_eye_rescale
+        aimg = self.draw_item(aimg, eye_kps)
+        #return aimg
+    
+        rimg = np.zeros( (max(eimg.shape[0], aimg.shape[0]), eimg.shape[1]+aimg.shape[1], 3), dtype=np.uint8)
+        rimg[:eimg.shape[0], :eimg.shape[1], :] = eimg
+        rimg[:aimg.shape[0], eimg.shape[1]:eimg.shape[1]+aimg.shape[1], :] = aimg
+        return rimg
+
+    def get(self, img):
+        results = []
+        faces = self.app.get(img)
+        if len(faces)==0:
+            return results
+        for face in faces:
+            bbox = face.bbox
+            width = bbox[2] - bbox[0]
+            kps = face.kps
+            center = (kps[0]+kps[1]) / 2.0
+            #_size = np.abs(kps[1][0] - kps[0][0]) * 1.5
+            _size = max(width/1.5, np.abs(kps[1][0] - kps[0][0]) ) * 1.5
+            rotate = 0
+            _scale = self.input_size  / _size
+            aimg, M = face_align.transform(img, center, self.input_size, _scale, rotate)
+            #eimg = cv2.resize(aimg, None, fx=R, fy=R)
+            #cv2.imwrite("outputs/a_%s"%name, aimg)
+            aimg = cv2.cvtColor(aimg, cv2.COLOR_BGR2RGB)
+            input = aimg.copy()
+            input = np.transpose(input, (2, 0, 1))
+            input = np.expand_dims(input, 0)
+            imgs = torch.Tensor(input).cuda()
+            imgs.div_(255).sub_(0.5).div_(0.5)
+            opred = self.model(imgs).detach().cpu().numpy().flatten().reshape( (-1, 3) )
+            opred[:, 0:2] += 1
+            opred[:, 0:2] *= (self.input_size // 2)
+            #opred[:, 0:2] *= 112
+            opred[:,2] *= 10.0
+            IM = cv2.invertAffineTransform(M)
+            pred = face_align.trans_points(opred, IM)
+            result = (bbox, kps, pred)
+            results.append(result)
+        return results
+
+
+if __name__ == '__main__':
+    ckpt_path = sys.argv[1]
+    handler = GazeHandler(ckpt_path)
+    output_dir = 'outputs/'
+    if not osp.exists(output_dir):
+        os.makedirs(output_dir)
+    input_dir = 'assets/images'
+    for imgname in os.listdir(input_dir):
+        imgpath = osp.join(input_dir, imgname)
+        img = cv2.imread(imgpath)
+        print(imgpath, imgname)
+        if img is None:
+            continue
+        results = handler.get(img)
+        if len(results)==0:
+            continue
+        eimg = handler.draw_on(img, results)
+        oimg = np.concatenate((img, eimg), axis=1)
+        cv2.imwrite(osp.join(output_dir, "%s"%imgname), oimg)
+
+
+
diff --git a/reconstruction/gaze/trainer_gaze.py b/reconstruction/gaze/trainer_gaze.py
new file mode 100644
index 000000000..69294a256
--- /dev/null
+++ b/reconstruction/gaze/trainer_gaze.py
@@ -0,0 +1,89 @@
+from argparse import ArgumentParser
+
+import os
+import os.path as osp
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+import pytorch_lightning as pl
+from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import LearningRateMonitor
+from pytorch_lightning.loggers import TensorBoardLogger
+import timm
+from datasets.dataset_gaze import GazeDataset, DataLoaderX
+from models import GazeModel
+
+
+
+
+
+def cli_main():
+    pl.seed_everything(727)
+
+    # ------------
+    # args
+    # ------------
+    parser = ArgumentParser()
+    parser.add_argument('--backbone', default='resnet101d', type=str)
+    parser.add_argument('--batch_size', default=64, type=int)
+    parser.add_argument('--epoch', default=16, type=int)
+    parser.add_argument('--root', default='data/gaze_refine', type=str)
+    parser.add_argument('--num-gpus', default=8, type=int)
+    parser.add_argument('--tf32', action='store_true')
+    parser = pl.Trainer.add_argparse_args(parser)
+    args = parser.parse_args()
+
+    if not args.tf32:
+        torch.backends.cuda.matmul.allow_tf32 = False
+        torch.backends.cudnn.allow_tf32 = False
+    else:
+        torch.backends.cuda.matmul.allow_tf32 = True
+        torch.backends.cudnn.allow_tf32 = True
+    torch.backends.cudnn.benchmark = True
+
+    # ------------
+    # data
+    # ------------
+    train_set = GazeDataset(root_dir=args.root, is_train=True)
+    val_set = GazeDataset(root_dir=args.root, is_train=False)
+    print('train data size:', len(train_set))
+
+    train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=2, pin_memory=True)
+    val_loader = DataLoader(val_set, batch_size=args.batch_size, shuffle=False)
+
+    # ------------
+    # model
+    # ------------
+    model = GazeModel(backbone=args.backbone, epoch=args.epoch)
+    ckpt_path = 'work_dirs/gaze'
+    if not os.path.exists(ckpt_path):
+        os.makedirs(ckpt_path)
+
+    # ------------
+    # training
+    # ------------
+    checkpoint_callback = ModelCheckpoint(
+            monitor='val_loss',
+            dirpath=ckpt_path,
+            filename='{epoch:02d}-{val_loss:.6f}',
+            save_top_k=5,
+            mode='min',
+            )
+    lr_monitor = LearningRateMonitor(logging_interval='step')
+    trainer = pl.Trainer(
+        gpus = args.num_gpus,
+        accelerator="gpu",
+        strategy="ddp",
+        benchmark=True,
+		logger=TensorBoardLogger(osp.join(ckpt_path, 'logs')),
+        callbacks=[checkpoint_callback, lr_monitor],
+        check_val_every_n_epoch=1,
+        #progress_bar_refresh_rate=1,
+        max_epochs=args.epoch,
+    )
+    trainer.fit(model, train_loader, val_loader)
+
+if __name__ == '__main__':
+    cli_main()
+
diff --git a/reconstruction/jmlr/README.md b/reconstruction/jmlr/README.md
new file mode 100644
index 000000000..ffb5f4872
--- /dev/null
+++ b/reconstruction/jmlr/README.md
@@ -0,0 +1,50 @@
+## Introduction
+
+JMLR is an efficient high accuracy face reconstruction approach which achieved [Rank-1st](https://tianchi.aliyun.com/competition/entrance/531961/rankingList) of 
+[Perspective Projection Based Monocular 3D Face Reconstruction Challenge](https://tianchi.aliyun.com/competition/entrance/531961/introduction) 
+of [ECCV-2022 WCPA Workshop](https://sites.google.com/view/wcpa2022).
+
+Paper in [arXiv](https://arxiv.org/abs/2208.07142).
+
+
+## Method Pipeline
+
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/jmlr_pipeline.jpg?raw=true" width="800" alt="jmlr-pipeline"/>
+
+
+## Data preparation
+
+1. Download the dataset from WCPA organiser and put it at somewhere.
+
+2. Create `cache_align/` dir and put `flip_index.npy` file under it.
+
+3. Check `configs/s1.py` and fix the location to yours.
+
+4. Use ``python rec_builder.py`` to generate cached dataset, which will be used in following steps.
+ 
+
+## Training
+
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=13334 train.py configs/s1.py
+```
+
+## Inference Example
+
+```
+python inference_simple.py
+```
+
+## Resources
+
+[flip_index.npy](https://drive.google.com/file/d/1fZ4cRyvQeehwKoMKKSmXUmTx5GEJwyrT/view?usp=sharing)
+
+[pretrained-model](https://drive.google.com/file/d/1qSpqDDLQfcPeFr2b82IZrK8QC_3lci3l/view?usp=sharing)
+
+[projection_matrix.txt](https://drive.google.com/file/d/1joiu-V0qEZxil_AHcg_W726nRxE8Q4dm/view?usp=sharing)
+
+## Results
+
+<img src="https://github.com/nttstar/insightface-resources/blob/master/images/jmlr_id.jpg?raw=true" width="800" alt="jmlr-id"/>
+
diff --git a/reconstruction/jmlr/augs.py b/reconstruction/jmlr/augs.py
new file mode 100644
index 000000000..2977afbc1
--- /dev/null
+++ b/reconstruction/jmlr/augs.py
@@ -0,0 +1,271 @@
+import numpy as np
+import cv2
+import os
+import os.path as osp
+import albumentations as A
+from albumentations.core.transforms_interface import ImageOnlyTransform
+from albumentations.pytorch import ToTensorV2
+
+class RectangleBorderAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            fill_value = 0,
+            fg_limit = (0.7, 0.9),
+            always_apply=False,
+            p=1.0,
+            ):
+        super(RectangleBorderAugmentation, self).__init__(always_apply, p)
+        #assert limit>0.0 and limit<1.0
+        assert isinstance(fg_limit, tuple)
+        assert fg_limit[1]>fg_limit[0]
+        self.fill_value = 0
+        self.fg_limit = fg_limit
+        #self.output_size = output_size
+
+
+    def apply(self, image, fg, top, left, **params):
+        assert image.shape[0]==image.shape[1]
+        oimage = np.ones_like(image) * self.fill_value
+        f = int(fg*image.shape[0])
+        t = int(top*image.shape[0])
+        l = int(left*image.shape[1])
+        oimage[t:t+f,l:l+f,:] = image[t:t+f,l:l+f,:]
+        return oimage
+
+    def get_params(self):
+        fg = np.random.uniform(self.fg_limit[0], self.fg_limit[1])
+        top = np.random.uniform(0.0, 1.0-fg)
+        left = np.random.uniform(0.0, 1.0-fg)
+        return {'fg': fg, 'top': top, 'left': left}
+
+    def get_transform_init_args_names(self):
+        return ('fill_value','fg_limit')
+
+class SunGlassAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            fill_value = 0,
+            loc = [ (38, 52), (73, 52) ],
+            rad_limit = (10, 20),
+            always_apply=False,
+            p=1.0,
+            ):
+        super(SunGlassAugmentation, self).__init__(always_apply, p)
+        #assert limit>0.0 and limit<1.0
+        assert isinstance(rad_limit, tuple)
+        self.fill_value = 0
+        self.loc = loc
+        self.rad_limit = rad_limit
+
+
+    def apply(self, image, rad, **params):
+        for i in range(2):
+            cv2.circle(image, self.loc[i], rad, self.fill_value, -1)
+        return image
+
+    def get_params(self):
+        rad = np.random.randint(self.rad_limit[0], self.rad_limit[1])
+        return {'rad':rad}
+
+    def get_transform_init_args_names(self):
+        return ('fill_value', 'loc', 'rad_limit')
+
+class ForeHeadAugmentation(ImageOnlyTransform):
+
+    def __init__(
+            self,
+            height_min = 0.2,
+            height_max = 0.4,
+            width_min = 0.5,
+            always_apply=False,
+            p=1.0,
+            ):
+        super(ForeHeadAugmentation, self).__init__(always_apply, p)
+        assert height_max > height_min
+        #assert limit>0.0 and limit<1.0
+        self.height_min = height_min
+        self.height_max = height_max
+        self.width_min = width_min
+
+
+    def apply(self, image, height, width, left, **params):
+        mask_value = np.random.randint(0, 255, size=(int(image.shape[0]*height), int(image.shape[1]*width), 3), dtype=image.dtype)
+        l = int(image.shape[1]*left)
+        image[:mask_value.shape[0], l:l+mask_value.shape[1], :] = mask_value
+        return image
+
+    def get_params(self):
+        height = np.random.uniform(self.height_min, self.height_max)
+        width = np.random.uniform(self.width_min, 1.0)
+        left = np.random.uniform(0.0, 1.0 - width)
+        return {'height': height, 'width': width, 'left': left}
+
+    def get_transform_init_args_names(self):
+        return ('height_min', 'height_max','width_min')
+
+
+def get_aug_transform(cfg):
+    aug_modes = cfg.aug_modes
+    input_size = cfg.input_size
+    task = cfg.task
+    transform_list = []
+    is_test = False
+    if 'test-aug' in aug_modes:
+        #transform_list.append(
+        #    A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+        #    )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.02, scale_limit=0.05, rotate_limit=5, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=1.0, always_apply=True)
+            )
+        is_test = True
+
+    if '1' in aug_modes:
+        transform_list.append(
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+            )
+    if '1A' in aug_modes:
+        transform_list.append(
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.02, scale_limit=0.03, rotate_limit=6, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.3)
+            )
+    if '2' in aug_modes:
+        transform_list.append(
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.1, rotate_limit=15, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.4)
+            )
+    if '3' in aug_modes:
+        transform_list.append(
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.6)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.2, rotate_limit=30, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.6)
+            )
+    if 'nist1' in aug_modes:
+        transform_list.append(
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.06, rotate_limit=6, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.4)
+            )
+    if 'nist2' in aug_modes:
+        transform_list.append(
+            #A.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.02, p=0.3)
+            A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.05, p=0.2)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.06, scale_limit=0.06, rotate_limit=6, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.4)
+            )
+        transform_list.append(
+                A.OneOf([
+                    RectangleBorderAugmentation(p=0.5),
+                    ForeHeadAugmentation(p=0.5),
+                    #SunGlassAugmentation(p=0.2),
+                    ], p=0.06)
+                )
+        transform_list.append(
+            A.ToGray(p=0.05)
+            )
+        transform_list.append(
+            A.geometric.resize.RandomScale(scale_limit=(0.7, 0.9), interpolation=cv2.INTER_LINEAR, p=0.05)
+            )
+        transform_list.append(
+            A.ISONoise(p=0.06)
+            )
+        transform_list.append(
+            A.MedianBlur(blur_limit=(1,7), p=0.05)
+            )
+        transform_list.append(
+            A.MotionBlur(blur_limit=(5,12), p=0.05)
+            )
+        transform_list.append(
+            A.ImageCompression(quality_lower=50, quality_upper=80, p=0.05)
+            )
+    if 'prod' in aug_modes:
+        transform_list.append(
+            #A.RandomBrightnessContrast(brightness_limit=0.125, contrast_limit=0.125, p=0.2)
+            A.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.02, p=0.3)
+            )
+        transform_list.append(
+            A.ShiftScaleRotate(shift_limit=0.06, scale_limit=0.1, rotate_limit=10, interpolation=cv2.INTER_LINEAR, 
+                border_mode=cv2.BORDER_CONSTANT, value=0, mask_value=0, p=0.4)
+            )
+        transform_list.append(
+                A.OneOf([
+                    RectangleBorderAugmentation(p=0.5),
+                    ForeHeadAugmentation(p=0.5),
+                    MaskAugmentation(mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'], mask_probs=[0.4, 0.4, 0.1, 0.1], h_low=0.33, h_high=0.4, p=0.2),
+                    SunGlassAugmentation(p=0.2),
+                    ], p=0.2)
+                )
+        transform_list.append(
+            A.ToGray(p=0.05)
+            )
+        transform_list.append(
+            A.geometric.resize.RandomScale(scale_limit=(0.6, 0.9), interpolation=cv2.INTER_LINEAR, p=0.2)
+            )
+        transform_list.append(
+            A.ISONoise(p=0.1)
+            )
+        transform_list.append(
+            A.MedianBlur(blur_limit=(1,7), p=0.1)
+            )
+        transform_list.append(
+            A.MotionBlur(blur_limit=(5,12), p=0.1)
+            )
+        transform_list.append(
+            A.ImageCompression(quality_lower=30, quality_upper=80, p=0.1)
+            )
+    #if input_size!=112: # TODO!!
+    #    transform_list.append(
+    #        A.geometric.resize.Resize(input_size, input_size, interpolation=cv2.INTER_LINEAR, always_apply=True)
+    #        )
+    transform_list += \
+        [
+            #A.HorizontalFlip(p=0.5),
+            A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+            ToTensorV2(),
+        ]
+    #here, the input for A transform is rgb cv2 img
+    if is_test:
+        transform = A.ReplayCompose(
+            transform_list ,
+            keypoint_params=A.KeypointParams(format='xy',remove_invisible=False)
+        )
+    else:
+        transform = A.Compose(
+            transform_list,
+            keypoint_params=A.KeypointParams(format='xy',remove_invisible=False)
+        )
+    return transform
+
+
+if __name__ == "__main__":
+    tool = MaskRenderer()
+    tool.prepare(ctx_id=0, det_size=(128,128))
+    image = cv2.imread("./test1.png")[:,:,::-1]
+    mask_image  = "mask_blue"
+    #params = tool.build_params(image)
+    label = np.load('assets/mask_label.npy')
+    params = tool.decode_params(label)
+    #print(params[0][:20])
+    mask_out = tool.render_mask(image, mask_image, params, input_is_rgb=True, auto_blend=False)[:,:,::-1]
+    #print(uv_out.dtype, uv_out.shape)
+    cv2.imwrite('output_mask.jpg', mask_out)
+    transform = A.Compose([
+        MaskAugmentation(mask_names=['mask_white', 'mask_blue', 'mask_black', 'mask_green'], mask_probs=[0.4, 0.4, 0.1, 0.1], h_low=0.33, h_high=0.4, p=1.0),
+        #MaskAugmentation(p=1.0),
+        ])
+    mask_out = transform(image=image, hlabel=label)["image"][:,:,::-1]
+    cv2.imwrite('output_mask2.jpg', mask_out)
diff --git a/reconstruction/jmlr/backbones/__init__.py b/reconstruction/jmlr/backbones/__init__.py
new file mode 100644
index 000000000..5f46a9da2
--- /dev/null
+++ b/reconstruction/jmlr/backbones/__init__.py
@@ -0,0 +1 @@
+from .network import get_network
diff --git a/reconstruction/jmlr/backbones/iresnet.py b/reconstruction/jmlr/backbones/iresnet.py
new file mode 100644
index 000000000..2103c1e64
--- /dev/null
+++ b/reconstruction/jmlr/backbones/iresnet.py
@@ -0,0 +1,326 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+import logging
+
+__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1, eps=1e-5, dropblock=0.0):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=eps)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=eps)
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=eps)
+        self.downsample = downsample
+        self.stride = stride
+        self.dbs = None
+        if dropblock>0.0:
+            import timm
+            from timm.layers import DropBlock2d
+            self.dbs = [DropBlock2d(dropblock, 7), DropBlock2d(dropblock, 7), DropBlock2d(dropblock, 7)]
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        if self.dbs is not None:
+            out = self.dbs[0](out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        if self.dbs is not None:
+            out = self.dbs[1](out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        if self.dbs is not None:
+            out = self.dbs[2](out)
+        return out
+
+
+class IResNet(nn.Module):
+    def __init__(self,
+                 block, layers, dropout=0.0, num_features=512, input_size=112, zero_init_residual=False,
+                 stem_type='', dropblock = 0.0, kaiming_init=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=0):
+        super(IResNet, self).__init__()
+        self.input_size = input_size
+        assert self.input_size%16==0
+        fc_scale = self.input_size // 16
+        self.fc_scale = fc_scale*fc_scale
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        self.norm_layer = nn.BatchNorm2d
+        self.act_layer = nn.PReLU
+        self.eps = 1e-5
+        if kaiming_init:
+            self.eps = 2e-5
+        self.stem_type = stem_type
+        self.dropblock = dropblock
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        if stem_type!='D':
+            self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        else:
+            stem_width = self.inplanes // 2
+            stem_chs = [stem_width, stem_width]
+            self.conv1 = nn.Sequential(*[
+                nn.Conv2d(3, stem_chs[0], 3, stride=1, padding=1, bias=False),
+                self.norm_layer(stem_chs[0], eps=self.eps),
+                self.act_layer(stem_chs[0]),
+                nn.Conv2d(stem_chs[0], stem_chs[1], 3, stride=1, padding=1, bias=False),
+                self.norm_layer(stem_chs[1], eps=self.eps),
+                self.act_layer(stem_chs[1]),
+                nn.Conv2d(stem_chs[1], self.inplanes, 3, stride=1, padding=1, bias=False)])
+        logging.info("iresnet, input_size: %d, fc_scale: %d, dropout: %.2f, stem_type: %s, fp16: %d"%(self.input_size, self.fc_scale, dropout, stem_type, self.fp16))
+        logging.info("iresnet, eps: %.6f, dropblock: %.3f, kaiming_init: %d"%(self.eps, self.dropblock, kaiming_init))
+        #self.conv1.requires_grad = False
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=self.eps)
+        #self.bn1.requires_grad = False
+        self.prelu = nn.PReLU(self.inplanes)
+        #self.prelu.requires_grad = False
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        #self.layer1.requires_grad = False
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        #self.layer2.requires_grad = False
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1],
+                                       dropblock=self.dropblock)
+        #self.layer3.requires_grad = False
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2],
+                                       dropblock=self.dropblock)
+        #self.layer4.requires_grad = False
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=self.eps)
+        #self.bn2.requires_grad = False
+        if dropout>0.0:
+            self.dropout = nn.Dropout(p=dropout, inplace=True)
+        else:
+            self.dropout = None
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        #self.fc.requires_grad = False
+        self.features = nn.BatchNorm1d(num_features, eps=self.eps)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        #for m in self.modules():
+        #    if isinstance(m, nn.Conv2d):
+        #        nn.init.xavier_uniform_(m.weight.data)
+        #        if m.bias is not None:
+        #            m.bias.data.zero_()
+        #    elif isinstance(m, nn.BatchNorm2d):
+        #        m.weight.data.fill_(1)
+        #        m.bias.data.zero_()
+        #    elif isinstance(m, nn.BatchNorm1d):
+        #        m.weight.data.fill_(1)
+        #        m.bias.data.zero_()
+        #    elif isinstance(m, nn.Linear):
+        #        nn.init.xavier_uniform_(m.weight.data)
+        #        if m.bias is not None:
+        #            m.bias.data.zero_()
+        #nn.init.constant_(self.features.weight, 1.0)
+        #self.features.weight.requires_grad = False
+
+        #for m in self.modules():
+        #    if kaiming_init:
+        #        if isinstance(m, (nn.Conv2d, nn.Linear)):
+        #            nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        #            if m.bias is not None:
+        #                nn.init.constant_(m.bias, 0)
+        #    else:
+        #        if isinstance(m, (nn.Conv2d, nn.Linear)):
+        #            nn.init.normal_(m.weight, 0, 0.1)
+        #            if m.bias is not None:
+        #                nn.init.constant_(m.bias, 0)
+        #    if isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+        #        nn.init.constant_(m.weight, 1)
+        #        nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False, dropblock=0.0):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            if self.stem_type!='D':
+                downsample = nn.Sequential(
+                    conv1x1(self.inplanes, planes * block.expansion, stride),
+                    nn.BatchNorm2d(planes * block.expansion, eps=self.eps),
+                )
+            else:
+				#avg_pool_fn = AvgPool2dSame if avg_stride == 1 and dilation > 1 else nn.AvgPool2d
+                avg_stride = stride
+                pool = nn.AvgPool2d(2, avg_stride, ceil_mode=True, count_include_pad=False)
+                downsample = nn.Sequential(*[
+                    pool,
+                    conv1x1(self.inplanes, planes * block.expansion, stride=1),
+                    nn.BatchNorm2d(planes * block.expansion, eps=self.eps),
+                    ])
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation, self.eps, dropblock))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation,
+                      eps=self.eps,
+                      dropblock=dropblock))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        #if self.input_size!=112:
+        #    x = F.interpolate(x, [self.input_size, self.input_size], mode='bilinear', align_corners=False)
+        is_fp16 = self.fp16>0
+        with torch.cuda.amp.autocast(is_fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            if self.fp16<3:
+                x = self.layer4(x)
+                x = self.bn2(x)
+                x = torch.flatten(x, 1)
+                if self.dropout is not None:
+                    x = self.dropout(x)
+        if is_fp16:
+            x = x.float()
+        if self.fp16>=3:
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            if self.dropout is not None:
+                x = self.dropout(x)
+        x = self.fc(x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
+                    progress, **kwargs)
+
+def iresnet120(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet120', IBasicBlock, [3, 16, 37, 3], pretrained,
+                    progress, **kwargs)
+
+def iresnet160(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet160', IBasicBlock, [3, 16, 56, 3], pretrained,
+                    progress, **kwargs)
+
+def iresnet180(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet180', IBasicBlock, [3, 20, 63, 3], pretrained,
+                    progress, **kwargs)
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
+                    progress, **kwargs)
+
+def iresnet247(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet247', IBasicBlock, [3, 36, 80, 4], pretrained,
+                    progress, **kwargs)
+
+def iresnet269(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet269', IBasicBlock, [4, 46, 80, 4], pretrained,
+                    progress, **kwargs)
+
+def iresnet300(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet300', IBasicBlock, [4, 46, 95, 4], pretrained,
+                    progress, **kwargs)
+
+
+def get_model(name, **kwargs):
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    elif name == "r2060":
+        from .iresnet2060 import iresnet2060
+        return iresnet2060(False, **kwargs)
+    else:
+        raise ValueError()
+
diff --git a/reconstruction/jmlr/backbones/network.py b/reconstruction/jmlr/backbones/network.py
new file mode 100644
index 000000000..4f438452f
--- /dev/null
+++ b/reconstruction/jmlr/backbones/network.py
@@ -0,0 +1,260 @@
+import os
+import time
+import timm
+import glob
+import numpy as np
+import os.path as osp
+
+import torch
+import torch.distributed as dist
+from torch import nn
+import torch.nn.functional as F
+from .iresnet import get_model as arcface_get_model
+
+
+def kaiming_leaky_init(m):
+    classname = m.__class__.__name__
+    if classname.find('Linear') != -1:
+        torch.nn.init.kaiming_normal_(m.weight, a=0.2, mode='fan_in', nonlinearity='leaky_relu')
+
+class CustomMappingNetwork(nn.Module):
+    def __init__(self, z_dim, map_hidden_dim, map_output_dim):
+        super().__init__()
+
+
+
+        self.network = nn.Sequential(nn.Linear(z_dim, map_hidden_dim),
+                nn.LeakyReLU(0.2, inplace=True),
+
+                nn.Linear(map_hidden_dim, map_hidden_dim),
+                nn.LeakyReLU(0.2, inplace=True),
+
+                nn.Linear(map_hidden_dim, map_hidden_dim),
+                nn.LeakyReLU(0.2, inplace=True),
+
+                nn.Linear(map_hidden_dim, map_output_dim))
+
+        self.network.apply(kaiming_leaky_init)
+        with torch.no_grad():
+            self.network[-1].weight *= 0.25
+
+    def forward(self, z):
+        frequencies_offsets = self.network(z)
+        frequencies = frequencies_offsets[..., :frequencies_offsets.shape[-1]//2]
+        phase_shifts = frequencies_offsets[..., frequencies_offsets.shape[-1]//2:]
+
+        return frequencies, phase_shifts
+
+class FiLMLayer(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.layer = nn.Linear(input_dim, hidden_dim)
+
+    def forward(self, x, freq, phase_shift):
+        x = self.layer(x)
+        return torch.sin(freq * x + phase_shift)
+
+class InstanceNorm(nn.Module):
+    def __init__(self, epsilon=1e-8):
+        """
+            @notice: avoid in-place ops.
+            https://discuss.pytorch.org/t/encounter-the-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation/836/3
+        """
+        super(InstanceNorm, self).__init__()
+        self.epsilon = epsilon
+
+    def forward(self, x):
+        x   = x - torch.mean(x, (2, 3), True)
+        tmp = torch.mul(x, x) # or x ** 2
+        tmp = torch.rsqrt(torch.mean(tmp, (2, 3), True) + self.epsilon)
+        return x * tmp
+
+class ApplyStyle(nn.Module):
+    def __init__(self, latent_size, channels):
+        super(ApplyStyle, self).__init__()
+        self.linear = nn.Linear(latent_size, channels * 2)
+
+    def forward(self, x, latent):
+        style = self.linear(latent).unsqueeze(2).unsqueeze(3) #B, 2*c, 1, 1
+        gamma, beta = style.chunk(2, 1)
+        x = gamma * x + beta
+        return x
+
+class ResnetBlock_Adain(nn.Module):
+    def __init__(self, dim, latent_size, padding_type='reflect', activation=nn.ReLU(True)):
+        super(ResnetBlock_Adain, self).__init__()
+
+        p = 0
+        conv1 = []
+        if padding_type == 'reflect':
+            conv1 += [nn.ReflectionPad2d(1)]
+        elif padding_type == 'replicate':
+            conv1 += [nn.ReplicationPad2d(1)]
+        elif padding_type == 'zero':
+            p = 1
+        else:
+            raise NotImplementedError('padding [%s] is not implemented' % padding_type)
+        conv1 += [nn.Conv2d(dim, dim, kernel_size=3, padding = p), InstanceNorm()]
+        self.conv1 = nn.Sequential(*conv1)
+        self.style1 = ApplyStyle(latent_size, dim)
+        self.act1 = activation
+
+        p = 0
+        conv2 = []
+        if padding_type == 'reflect':
+            conv2 += [nn.ReflectionPad2d(1)]
+        elif padding_type == 'replicate':
+            conv2 += [nn.ReplicationPad2d(1)]
+        elif padding_type == 'zero':
+            p = 1
+        else:
+            raise NotImplementedError('padding [%s] is not implemented' % padding_type)
+        conv2 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p), InstanceNorm()]
+        self.conv2 = nn.Sequential(*conv2)
+        self.style2 = ApplyStyle(latent_size, dim)
+
+
+    def forward(self, x, dlatents_in_slice):
+        y = self.conv1(x)
+        y = self.style1(y, dlatents_in_slice)
+        y = self.act1(y)
+        y = self.conv2(y)
+        y = self.style2(y, dlatents_in_slice)
+        out = x + y
+        return out
+
+class OneNetwork(nn.Module):
+    def __init__(self, cfg):
+        super(OneNetwork, self).__init__()
+        self.num_verts = cfg.num_verts
+        self.input_size = cfg.input_size
+        self.use_eyes = cfg.eyes is not None
+        kwargs = {}
+        num_classes = self.num_verts*5
+        if cfg.task==1:
+            num_classes = self.num_verts*3
+        elif cfg.task==2:
+            num_classes = 6
+        elif cfg.task==3:
+            num_classes = self.num_verts*2
+        eye_num_classes = 481*2*5
+        #if use_eyes:
+        #    num_classes += 481*2*5
+        if cfg.network.startswith('resnet'):
+            kwargs['base_width'] = int(64*cfg.width_mult)
+        p_num_classes = num_classes
+        if cfg.no_gap:
+            p_num_classes = 0
+            kwargs['global_pool'] = None
+        elif cfg.use_arcface:
+            p_num_classes = 0
+            kwargs['global_pool'] = None
+        elif self.use_eyes:
+            p_num_classes = 0
+            #kwargs['global_pool'] = None
+        if cfg.network=='resnet_jmlr':
+            from .resnet import resnet_jmlr
+            self.net = resnet_jmlr(num_classes = p_num_classes, **kwargs)
+            if self.use_eyes:
+                input_dim = 512 #resnet34d
+                self.nete = resnet_jmlr(num_classes = p_num_classes, **kwargs)
+                self.fc = nn.Linear(input_dim*2, num_classes+eye_num_classes)
+        else:
+            self.net = timm.create_model(cfg.network, num_classes = p_num_classes, **kwargs)
+
+        if cfg.no_gap:
+            in_channel = self.net.num_features
+            feat_hw = (self.input_size//32)**2
+            mid_channel = 128
+            self.no_gap_output = nn.Sequential(*[
+                nn.BatchNorm2d(in_channel),
+                nn.Conv2d(in_channel, mid_channel, 1, stride=1, padding=0, bias=False),
+                nn.ReLU(),
+                nn.Flatten(1),
+                nn.Linear(mid_channel*feat_hw, num_classes)])
+
+        self.no_gap = cfg.no_gap
+        self.use_arcface = cfg.use_arcface
+        if self.use_arcface:
+            self.neta = arcface_get_model(cfg.arcface_model, input_size=cfg.arcface_input_size)
+            self.neta.load_state_dict(torch.load(cfg.arcface_ckpt, map_location=torch.device('cpu')))
+            self.neta.eval()
+            self.neta.requires_grad_(False)
+            input_dim = 512 #resnet34d
+            z_dim = 512 #arcface_dim
+            hidden_dim = 256
+            self.pool = nn.AdaptiveAvgPool2d(1)
+            self.flatten = nn.Flatten(1)
+            mlp_act = nn.LeakyReLU
+
+            self.mlp = nn.Sequential(*[
+                nn.Linear(z_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, hidden_dim),
+                mlp_act(),
+                nn.Linear(hidden_dim, input_dim),
+                ])
+            style_blocks = []
+            for i in range(3):
+                style_blocks += [ResnetBlock_Adain(input_dim, latent_size=z_dim)]
+            self.style_blocks = nn.Sequential(*style_blocks)
+            self.branch2d = nn.Sequential(*[
+                nn.Conv2d(input_dim, input_dim, 3, stride=1, padding=1, bias=False),
+                nn.BatchNorm2d(input_dim),
+                nn.ReLU(),
+                ])
+            self.fc = nn.Linear(input_dim*2, num_classes)
+
+
+    def forward(self, x):
+        if self.use_arcface:
+            conv_feat = self.net.forward_features(x)
+            #input = self.flatten(self.pool(conv_feat))
+            xa = F.interpolate(x, [144, 144], mode='bilinear', align_corners=False)
+            xa = xa[:,:,8:120,16:128]
+            z = self.neta(xa)
+            z = self.mlp(z)
+
+            c = conv_feat
+            for i in range(len(self.style_blocks)):
+                c = self.style_blocks[i](c, z)
+            feat3 = c
+            feat2 = self.branch2d(conv_feat)
+            conv_feat = torch.cat([feat3, feat2], dim=1)
+            feat = self.flatten(self.pool(conv_feat))
+            pred = self.fc(feat)
+
+        elif self.no_gap:
+            y = self.net.forward_features(x)
+            pred = self.no_gap_output(y)
+        else:
+            pred = self.net(x)
+            if self.use_eyes:
+                eye_w1 = x.shape[3]//8
+                eye_w2 = x.shape[3] - wa
+                hstep = x.shape[2]//8
+                eye_h1 = hstep*2
+                eye_h2 = hstep*4
+                x_eye = x[:,:,eye_h1:eye_h2,eye_w1:eye_w2]
+                feate = self.nete(x_eye)
+                #print(pred.shape, feate.shape)
+                feat = torch.cat((pred,feate), 1)
+                pred = self.fc(feat)
+        return pred
+
+def get_network(cfg):
+    if cfg.use_onenetwork:
+        net = OneNetwork(cfg)
+    else:
+        net = timm.create_model(cfg.network, num_classes = 1220*5)
+    return net
+
+
diff --git a/reconstruction/jmlr/backbones/resnet.py b/reconstruction/jmlr/backbones/resnet.py
new file mode 100644
index 000000000..6dc9b9faf
--- /dev/null
+++ b/reconstruction/jmlr/backbones/resnet.py
@@ -0,0 +1,453 @@
+"""PyTorch ResNet
+
+This started as a copy of https://github.com/pytorch/vision 'resnet.py' (BSD-3-Clause) with
+additional dropout and dynamic global avg/max pool.
+
+ResNeXt, SE-ResNeXt, SENet, and MXNet Gluon stem/downsample variants, tiered stems added by Ross Wightman
+
+Copyright 2019, Ross Wightman
+"""
+import math
+from functools import partial
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from timm.models.helpers import build_model_with_cfg
+from timm.models.layers import DropBlock2d, DropPath, AvgPool2dSame, BlurPool2d, GroupNorm, create_attn, get_attn, create_classifier
+
+
+
+def get_padding(kernel_size, stride, dilation=1):
+    padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
+    return padding
+
+
+def create_aa(aa_layer, channels, stride=2, enable=True):
+    if not aa_layer or not enable:
+        return nn.Identity()
+    return aa_layer(stride) if issubclass(aa_layer, nn.AvgPool2d) else aa_layer(channels=channels, stride=stride)
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(
+            self, inplanes, planes, stride=1, downsample=None, cardinality=1, base_width=64,
+            reduce_first=1, dilation=1, first_dilation=None, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d,
+            attn_layer=None, aa_layer=None, drop_block=None, drop_path=None):
+        super(BasicBlock, self).__init__()
+
+        assert cardinality == 1, 'BasicBlock only supports cardinality of 1'
+        assert base_width == 64, 'BasicBlock does not support changing base width'
+        first_planes = planes // reduce_first
+        outplanes = planes * self.expansion
+        first_dilation = first_dilation or dilation
+        use_aa = aa_layer is not None and (stride == 2 or first_dilation != dilation)
+
+        self.conv1 = nn.Conv2d(
+            inplanes, first_planes, kernel_size=3, stride=1 if use_aa else stride, padding=first_dilation,
+            dilation=first_dilation, bias=False)
+        self.bn1 = norm_layer(first_planes)
+        self.drop_block = drop_block() if drop_block is not None else nn.Identity()
+        self.act1 = act_layer(inplace=True)
+        self.aa = create_aa(aa_layer, channels=first_planes, stride=stride, enable=use_aa)
+
+        self.conv2 = nn.Conv2d(
+            first_planes, outplanes, kernel_size=3, padding=dilation, dilation=dilation, bias=False)
+        self.bn2 = norm_layer(outplanes)
+
+        self.se = create_attn(attn_layer, outplanes)
+
+        self.act2 = act_layer(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+        self.dilation = dilation
+        self.drop_path = drop_path
+
+    def zero_init_last(self):
+        nn.init.zeros_(self.bn2.weight)
+
+    def forward(self, x):
+        shortcut = x
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.drop_block(x)
+        x = self.act1(x)
+        x = self.aa(x)
+
+        x = self.conv2(x)
+        x = self.bn2(x)
+
+        if self.se is not None:
+            x = self.se(x)
+
+        if self.drop_path is not None:
+            x = self.drop_path(x)
+
+        if self.downsample is not None:
+            shortcut = self.downsample(shortcut)
+        x += shortcut
+        x = self.act2(x)
+
+        return x
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(
+            self, inplanes, planes, stride=1, downsample=None, cardinality=1, base_width=64,
+            reduce_first=1, dilation=1, first_dilation=None, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d,
+            attn_layer=None, aa_layer=None, drop_block=None, drop_path=None):
+        super(Bottleneck, self).__init__()
+
+        width = int(math.floor(planes * (base_width / 64)) * cardinality)
+        first_planes = width // reduce_first
+        outplanes = planes * self.expansion
+        first_dilation = first_dilation or dilation
+        use_aa = aa_layer is not None and (stride == 2 or first_dilation != dilation)
+
+        self.conv1 = nn.Conv2d(inplanes, first_planes, kernel_size=1, bias=False)
+        self.bn1 = norm_layer(first_planes)
+        self.act1 = act_layer(inplace=True)
+
+        self.conv2 = nn.Conv2d(
+            first_planes, width, kernel_size=3, stride=1 if use_aa else stride,
+            padding=first_dilation, dilation=first_dilation, groups=cardinality, bias=False)
+        self.bn2 = norm_layer(width)
+        self.drop_block = drop_block() if drop_block is not None else nn.Identity()
+        self.act2 = act_layer(inplace=True)
+        self.aa = create_aa(aa_layer, channels=width, stride=stride, enable=use_aa)
+
+        self.conv3 = nn.Conv2d(width, outplanes, kernel_size=1, bias=False)
+        self.bn3 = norm_layer(outplanes)
+
+        self.se = create_attn(attn_layer, outplanes)
+
+        self.act3 = act_layer(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+        self.dilation = dilation
+        self.drop_path = drop_path
+
+    def zero_init_last(self):
+        nn.init.zeros_(self.bn3.weight)
+
+    def forward(self, x):
+        shortcut = x
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.act1(x)
+
+        x = self.conv2(x)
+        x = self.bn2(x)
+        x = self.drop_block(x)
+        x = self.act2(x)
+        x = self.aa(x)
+
+        x = self.conv3(x)
+        x = self.bn3(x)
+
+        if self.se is not None:
+            x = self.se(x)
+
+        if self.drop_path is not None:
+            x = self.drop_path(x)
+
+        if self.downsample is not None:
+            shortcut = self.downsample(shortcut)
+        x += shortcut
+        x = self.act3(x)
+
+        return x
+
+
+def downsample_conv(
+        in_channels, out_channels, kernel_size, stride=1, dilation=1, first_dilation=None, norm_layer=None):
+    norm_layer = norm_layer or nn.BatchNorm2d
+    kernel_size = 1 if stride == 1 and dilation == 1 else kernel_size
+    first_dilation = (first_dilation or dilation) if kernel_size > 1 else 1
+    p = get_padding(kernel_size, stride, first_dilation)
+
+    return nn.Sequential(*[
+        nn.Conv2d(
+            in_channels, out_channels, kernel_size, stride=stride, padding=p, dilation=first_dilation, bias=False),
+        norm_layer(out_channels)
+    ])
+
+
+def downsample_avg(
+        in_channels, out_channels, kernel_size, stride=1, dilation=1, first_dilation=None, norm_layer=None):
+    norm_layer = norm_layer or nn.BatchNorm2d
+    avg_stride = stride if dilation == 1 else 1
+    if stride == 1 and dilation == 1:
+        pool = nn.Identity()
+    else:
+        avg_pool_fn = AvgPool2dSame if avg_stride == 1 and dilation > 1 else nn.AvgPool2d
+        pool = avg_pool_fn(2, avg_stride, ceil_mode=True, count_include_pad=False)
+
+    return nn.Sequential(*[
+        pool,
+        nn.Conv2d(in_channels, out_channels, 1, stride=1, padding=0, bias=False),
+        norm_layer(out_channels)
+    ])
+
+
+def drop_blocks(drop_prob=0.):
+    return [
+        None, None,
+        partial(DropBlock2d, drop_prob=drop_prob, block_size=5, gamma_scale=0.25) if drop_prob else None,
+        partial(DropBlock2d, drop_prob=drop_prob, block_size=3, gamma_scale=1.00) if drop_prob else None]
+
+
+def make_blocks(
+        block_fn, channels, block_repeats, inplanes, reduce_first=1, output_stride=32,
+        down_kernel_size=1, avg_down=False, drop_block_rate=0., drop_path_rate=0., **kwargs):
+    stages = []
+    feature_info = []
+    net_num_blocks = sum(block_repeats)
+    net_block_idx = 0
+    net_stride = 4
+    dilation = prev_dilation = 1
+    for stage_idx, (planes, num_blocks, db) in enumerate(zip(channels, block_repeats, drop_blocks(drop_block_rate))):
+        stage_name = f'layer{stage_idx + 1}'  # never liked this name, but weight compat requires it
+        stride = 1 if stage_idx == 0 else 2
+        if net_stride >= output_stride:
+            dilation *= stride
+            stride = 1
+        else:
+            net_stride *= stride
+
+        downsample = None
+        if stride != 1 or inplanes != planes * block_fn.expansion:
+            down_kwargs = dict(
+                in_channels=inplanes, out_channels=planes * block_fn.expansion, kernel_size=down_kernel_size,
+                stride=stride, dilation=dilation, first_dilation=prev_dilation, norm_layer=kwargs.get('norm_layer'))
+            downsample = downsample_avg(**down_kwargs) if avg_down else downsample_conv(**down_kwargs)
+
+        block_kwargs = dict(reduce_first=reduce_first, dilation=dilation, drop_block=db, **kwargs)
+        blocks = []
+        for block_idx in range(num_blocks):
+            downsample = downsample if block_idx == 0 else None
+            stride = stride if block_idx == 0 else 1
+            block_dpr = drop_path_rate * net_block_idx / (net_num_blocks - 1)  # stochastic depth linear decay rule
+            blocks.append(block_fn(
+                inplanes, planes, stride, downsample, first_dilation=prev_dilation,
+                drop_path=DropPath(block_dpr) if block_dpr > 0. else None, **block_kwargs))
+            prev_dilation = dilation
+            inplanes = planes * block_fn.expansion
+            net_block_idx += 1
+
+        stages.append((stage_name, nn.Sequential(*blocks)))
+        feature_info.append(dict(num_chs=inplanes, reduction=net_stride, module=stage_name))
+
+    return stages, feature_info
+
+
+class ResNet(nn.Module):
+    """ResNet / ResNeXt / SE-ResNeXt / SE-Net
+
+    This class implements all variants of ResNet, ResNeXt, SE-ResNeXt, and SENet that
+      * have > 1 stride in the 3x3 conv layer of bottleneck
+      * have conv-bn-act ordering
+
+    This ResNet impl supports a number of stem and downsample options based on the v1c, v1d, v1e, and v1s
+    variants included in the MXNet Gluon ResNetV1b model. The C and D variants are also discussed in the
+    'Bag of Tricks' paper: https://arxiv.org/pdf/1812.01187. The B variant is equivalent to torchvision default.
+
+    ResNet variants (the same modifications can be used in SE/ResNeXt models as well):
+      * normal, b - 7x7 stem, stem_width = 64, same as torchvision ResNet, NVIDIA ResNet 'v1.5', Gluon v1b
+      * c - 3 layer deep 3x3 stem, stem_width = 32 (32, 32, 64)
+      * d - 3 layer deep 3x3 stem, stem_width = 32 (32, 32, 64), average pool in downsample
+      * e - 3 layer deep 3x3 stem, stem_width = 64 (64, 64, 128), average pool in downsample
+      * s - 3 layer deep 3x3 stem, stem_width = 64 (64, 64, 128)
+      * t - 3 layer deep 3x3 stem, stem width = 32 (24, 48, 64), average pool in downsample
+      * tn - 3 layer deep 3x3 stem, stem width = 32 (24, 32, 64), average pool in downsample
+
+    ResNeXt
+      * normal - 7x7 stem, stem_width = 64, standard cardinality and base widths
+      * same c,d, e, s variants as ResNet can be enabled
+
+    SE-ResNeXt
+      * normal - 7x7 stem, stem_width = 64
+      * same c, d, e, s variants as ResNet can be enabled
+
+    SENet-154 - 3 layer deep 3x3 stem (same as v1c-v1s), stem_width = 64, cardinality=64,
+        reduction by 2 on width of first bottleneck convolution, 3x3 downsample convs after first block
+
+    Parameters
+    ----------
+    block : Block, class for the residual block. Options are BasicBlockGl, BottleneckGl.
+    layers : list of int, number of layers in each block
+    num_classes : int, default 1000, number of classification classes.
+    in_chans : int, default 3, number of input (color) channels.
+    output_stride : int, default 32, output stride of the network, 32, 16, or 8.
+    global_pool : str, Global pooling type. One of 'avg', 'max', 'avgmax', 'catavgmax'
+    cardinality : int, default 1, number of convolution groups for 3x3 conv in Bottleneck.
+    base_width : int, default 64, factor determining bottleneck channels. `planes * base_width / 64 * cardinality`
+    stem_width : int, default 64, number of channels in stem convolutions
+    stem_type : str, default ''
+        The type of stem:
+          * '', default - a single 7x7 conv with a width of stem_width
+          * 'deep' - three 3x3 convolution layers of widths stem_width, stem_width, stem_width * 2
+          * 'deep_tiered' - three 3x3 conv layers of widths stem_width//4 * 3, stem_width, stem_width * 2
+    block_reduce_first : int, default 1
+        Reduction factor for first convolution output width of residual blocks, 1 for all archs except senets, where 2
+    down_kernel_size : int, default 1, kernel size of residual block downsample path, 1x1 for most, 3x3 for senets
+    avg_down : bool, default False, use average pooling for projection skip connection between stages/downsample.
+    act_layer : nn.Module, activation layer
+    norm_layer : nn.Module, normalization layer
+    aa_layer : nn.Module, anti-aliasing layer
+    drop_rate : float, default 0. Dropout probability before classifier, for training
+    """
+
+    def __init__(
+            self, block, layers, num_classes=1000, in_chans=3, output_stride=32, global_pool='avg',
+            cardinality=1, base_width=64, stem_width=64, stem_type='', replace_stem_pool=False, block_reduce_first=1,
+            down_kernel_size=1, avg_down=False, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d, aa_layer=None,
+            drop_rate=0.0, drop_path_rate=0., drop_block_rate=0., zero_init_last=True, block_args=None, channels=[64, 128, 256, 512]):
+        super(ResNet, self).__init__()
+        block_args = block_args or dict()
+        assert output_stride in (8, 16, 32)
+        self.num_classes = num_classes
+        self.drop_rate = drop_rate
+        self.grad_checkpointing = False
+
+        # Stem
+        deep_stem = 'deep' in stem_type
+        inplanes = stem_width * 2 if deep_stem else 64
+        if deep_stem:
+            stem_chs = (stem_width, stem_width)
+            if 'tiered' in stem_type:
+                stem_chs = (3 * (stem_width // 4), stem_width)
+            self.conv1 = nn.Sequential(*[
+                nn.Conv2d(in_chans, stem_chs[0], 3, stride=2, padding=1, bias=False),
+                norm_layer(stem_chs[0]),
+                act_layer(inplace=True),
+                nn.Conv2d(stem_chs[0], stem_chs[1], 3, stride=1, padding=1, bias=False),
+                norm_layer(stem_chs[1]),
+                act_layer(inplace=True),
+                nn.Conv2d(stem_chs[1], inplanes, 3, stride=1, padding=1, bias=False)])
+        else:
+            self.conv1 = nn.Conv2d(in_chans, inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(inplanes)
+        self.act1 = act_layer(inplace=True)
+        self.feature_info = [dict(num_chs=inplanes, reduction=2, module='act1')]
+
+        # Stem pooling. The name 'maxpool' remains for weight compatibility.
+        if replace_stem_pool:
+            self.maxpool = nn.Sequential(*filter(None, [
+                nn.Conv2d(inplanes, inplanes, 3, stride=1 if aa_layer else 2, padding=1, bias=False),
+                create_aa(aa_layer, channels=inplanes, stride=2) if aa_layer is not None else None,
+                norm_layer(inplanes),
+                act_layer(inplace=True)
+            ]))
+        else:
+            if aa_layer is not None:
+                if issubclass(aa_layer, nn.AvgPool2d):
+                    self.maxpool = aa_layer(2)
+                else:
+                    self.maxpool = nn.Sequential(*[
+                        nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
+                        aa_layer(channels=inplanes, stride=2)])
+            else:
+                self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        # Feature Blocks
+        #channels = [64, 128, 256, 512]
+        stage_modules, stage_feature_info = make_blocks(
+            block, channels, layers, inplanes, cardinality=cardinality, base_width=base_width,
+            output_stride=output_stride, reduce_first=block_reduce_first, avg_down=avg_down,
+            down_kernel_size=down_kernel_size, act_layer=act_layer, norm_layer=norm_layer, aa_layer=aa_layer,
+            drop_block_rate=drop_block_rate, drop_path_rate=drop_path_rate, **block_args)
+        for stage in stage_modules:
+            self.add_module(*stage)  # layer1, layer2, etc
+        self.feature_info.extend(stage_feature_info)
+
+        # Head (Pooling and Classifier)
+        self.num_features = 512 * block.expansion
+        self.global_pool, self.fc = create_classifier(self.num_features, self.num_classes, pool_type=global_pool)
+
+        self.init_weights(zero_init_last=zero_init_last)
+
+    @torch.jit.ignore
+    def init_weights(self, zero_init_last=True):
+        for n, m in self.named_modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+        if zero_init_last:
+            for m in self.modules():
+                if hasattr(m, 'zero_init_last'):
+                    m.zero_init_last()
+
+    @torch.jit.ignore
+    def group_matcher(self, coarse=False):
+        matcher = dict(stem=r'^conv1|bn1|maxpool', blocks=r'^layer(\d+)' if coarse else r'^layer(\d+)\.(\d+)')
+        return matcher
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def get_classifier(self, name_only=False):
+        return 'fc' if name_only else self.fc
+
+    def reset_classifier(self, num_classes, global_pool='avg'):
+        self.num_classes = num_classes
+        self.global_pool, self.fc = create_classifier(self.num_features, self.num_classes, pool_type=global_pool)
+
+    def forward_features(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.act1(x)
+        x = self.maxpool(x)
+
+        if self.grad_checkpointing and not torch.jit.is_scripting():
+            x = checkpoint_seq([self.layer1, self.layer2, self.layer3, self.layer4], x, flatten=True)
+        else:
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            x = self.layer4(x)
+        return x
+
+    def forward_head(self, x, pre_logits: bool = False):
+        x = self.global_pool(x)
+        if self.drop_rate:
+            x = F.dropout(x, p=float(self.drop_rate), training=self.training)
+        return x if pre_logits else self.fc(x)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = self.forward_head(x)
+        return x
+
+
+#def _create_resnet(variant, pretrained=False, **kwargs):
+#    return build_model_with_cfg(ResNet, variant, pretrained, **kwargs)
+
+
+
+def resnet34d(pretrained=False, **kwargs):
+    """Constructs a ResNet-34-D model.
+    """
+    model_args = dict(
+        block=BasicBlock, layers=[3, 4, 6, 3], stem_width=32, stem_type='deep', avg_down=True, **kwargs)
+    return ResNet(**model_args)
+
+
+def resnet_jmlr(pretrained=False, **kwargs):
+    model_args = dict(
+        block=BasicBlock, layers=[5, 3, 4, 2], stem_width=32, stem_type='deep', avg_down=True, channels=[64,160,272,512], **kwargs)
+    #return _create_resnet('resnet34d', pretrained, **model_args)
+    return ResNet(**model_args)
+
diff --git a/reconstruction/jmlr/configs/base.py b/reconstruction/jmlr/configs/base.py
new file mode 100644
index 000000000..cd2c29fa3
--- /dev/null
+++ b/reconstruction/jmlr/configs/base.py
@@ -0,0 +1,105 @@
+from easydict import EasyDict as edict
+import numpy as np
+
+config = edict()
+config.embedding_size = 512
+config.sample_rate = 1
+config.fp16 = 0
+config.tf32 = False
+config.backbone_wd = None
+config.batch_size = 128
+config.clip_grad = None
+config.dropout = 0.0
+#config.warmup_epoch = -1
+config.loss = 'cosface'
+config.margin = 0.4
+config.hard_margin = False
+config.network = 'r50'
+config.prelu = True
+config.stem_type = ''
+config.dropblock = 0.0
+config.output = None
+config.input_size = 112
+config.width_mult = 1.0
+config.kaiming_init = True
+config.use_se = False
+config.aug_modes = []
+config.checkpoint_segments = [1, 1, 1, 1]
+
+config.sampling_id = True
+config.id_sampling_ratio = None
+metric_loss = edict()
+metric_loss.enable = False
+metric_loss.lambda_n = 0.0
+metric_loss.lambda_c = 0.0
+metric_loss.lambda_t = 0.0
+metric_loss.margin_c = 1.0
+metric_loss.margin_t = 1.0
+metric_loss.margin_n = 0.4
+config.metric_loss = metric_loss
+
+config.opt = 'sgd'
+config.lr = 0.1  # when batch size is 512
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.fc_mom = 0.9
+
+config.warmup_epochs = 0
+config.max_warmup_steps = 6000
+config.num_epochs = 24
+
+
+config.resume = False
+config.resume_path = None
+config.resume_from = None
+
+config.save_every_epochs = True
+
+config.lr_func = None
+config.lr_epochs = None
+config.save_pfc = False
+config.save_onnx = False
+config.save_opt = False
+
+config.label_6dof_mean = np.array([-0.018197, -0.017891, 0.025348, -0.005368, 0.001176, -0.532206], dtype=np.float32)  # mean of pitch, yaw, roll, tx, ty, tz
+config.label_6dof_std = np.array([0.314015, 0.271809, 0.081881, 0.022173, 0.048839, 0.065444], dtype=np.float32)        # std of pitch, yaw, roll, tx, ty, tz
+
+config.num_verts = 1220
+config.flipindex_file = 'cache_align/flip_index.npy'
+config.enable_flip = True
+config.verts3d_central_index = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 956, 975, 1022, 1041, 1047, 1048, 1049]
+
+config.task = 0
+config.ckpt = None
+config.loss_hard = False
+config.sampling_hard = False
+config.loss_pip = False
+config.net_stride = 32
+config.loss_bone3d = False
+config.loss_bone2d = False
+
+config.lossw_verts3d = 8.0
+config.lossw_verts2d = 16.0
+config.lossw_bone3d = 10.0
+config.lossw_bone2d = 10.0
+config.lossw_project = 10.0
+config.lossw_eyes3d = 8.0
+config.lossw_eyes2d = 16.0
+
+config.align_face = False
+config.no_gap = False
+
+config.use_trainval = False
+
+config.project_loss = False
+
+config.use_onenetwork = True
+
+config.use_rtloss = False
+
+
+config.use_arcface = False
+
+
+config.eyes = None
+
diff --git a/reconstruction/jmlr/configs/s1.py b/reconstruction/jmlr/configs/s1.py
new file mode 100644
index 000000000..95d77387a
--- /dev/null
+++ b/reconstruction/jmlr/configs/s1.py
@@ -0,0 +1,54 @@
+from easydict import EasyDict as edict
+
+config = edict()
+
+config.dataset = "wcpa"
+config.root_dir = '/data/insightface/wcpa'
+config.cache_dir = './cache_align'
+#config.num_classes = 617970
+#config.num_classes = 2000000
+#config.num_classes = 80000000
+#config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+#config.val_targets = ["lfw"]
+#config.val_targets = []
+config.verbose = 20000
+
+#config.network = 'resnet34d'
+config.network = 'resnet_jmlr'
+config.input_size = 256
+#config.width_mult = 1.0
+#config.dropout = 0.0
+#config.loss = 'cosface'
+#config.embedding_size = 512
+#config.sample_rate = 0.2
+config.fp16 = 0
+config.tf32 = True
+config.weight_decay = 5e-4
+config.batch_size = 64
+config.lr = 0.1  # lr when batch size is 512
+
+config.aug_modes = ['1']
+
+config.num_epochs = 40
+config.warmup_epochs = 1
+config.max_warmup_steps = 1000
+
+#def lr_step_func(epoch):
+#    return ((epoch + 1) / (4 + 1)) ** 2 if epoch < -1 else 0.1 ** len(
+#        [m for m in [20, 30, 38] if m - 1 <= epoch])
+#config.lr_func = lr_step_func
+
+config.task = 0
+config.save_every_epochs = False
+
+
+config.lossw_verts3d = 16.0
+
+config.align_face = True
+
+config.use_trainval = True
+#config.use_rtloss = True
+
+config.loss_bone3d = True
+config.lossw_bone3d = 2.0
+
diff --git a/reconstruction/jmlr/configs/s2.py b/reconstruction/jmlr/configs/s2.py
new file mode 100644
index 000000000..89fa69bab
--- /dev/null
+++ b/reconstruction/jmlr/configs/s2.py
@@ -0,0 +1,57 @@
+from easydict import EasyDict as edict
+
+config = edict()
+
+config.dataset = "wcpa"
+config.root_dir = '/data/insightface/wcpa'
+config.cache_dir = './cache_align_eyes'
+#config.num_classes = 617970
+#config.num_classes = 2000000
+#config.num_classes = 80000000
+#config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+#config.val_targets = ["lfw"]
+#config.val_targets = []
+config.verbose = 20000
+
+#config.network = 'resnet34d'
+config.network = 'resnet_jmlr'
+config.input_size = 256
+#config.width_mult = 1.0
+#config.dropout = 0.0
+#config.loss = 'cosface'
+#config.embedding_size = 512
+#config.sample_rate = 0.2
+config.fp16 = 0
+config.tf32 = True
+config.weight_decay = 5e-4
+config.batch_size = 64
+config.lr = 0.1  # lr when batch size is 512
+
+config.aug_modes = ['1']
+
+config.num_epochs = 40
+config.warmup_epochs = 1
+config.max_warmup_steps = 1000
+
+#def lr_step_func(epoch):
+#    return ((epoch + 1) / (4 + 1)) ** 2 if epoch < -1 else 0.1 ** len(
+#        [m for m in [20, 30, 38] if m - 1 <= epoch])
+#config.lr_func = lr_step_func
+
+config.task = 0
+config.save_every_epochs = False
+
+
+config.lossw_verts3d = 16.0
+
+config.align_face = True
+
+config.use_trainval = True
+#config.use_rtloss = True
+
+config.loss_bone3d = False
+config.lossw_bone3d = 2.0
+
+config.eyes = {'root': './eyes_raw'}
+config.enable_flip = False
+
diff --git a/reconstruction/jmlr/dataset.py b/reconstruction/jmlr/dataset.py
new file mode 100644
index 000000000..5f82855f3
--- /dev/null
+++ b/reconstruction/jmlr/dataset.py
@@ -0,0 +1,811 @@
+import numbers
+import os
+import os.path as osp
+import pickle
+import queue as Queue
+import threading
+import logging
+import numbers
+import math
+import pandas as pd
+from scipy.spatial.transform import Rotation
+
+import mxnet as mx
+from pathlib import Path
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset
+from skimage import transform as sktrans
+import cv2
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+from augs import *
+
+
+def Rt26dof(R_t, degrees=False):
+    yaw_gt, pitch_gt, roll_gt = Rotation.from_matrix(R_t[:3, :3].T).as_euler('yxz', degrees=degrees)
+    label_euler = np.array([pitch_gt, yaw_gt, roll_gt])
+    label_translation = R_t[3, :3]
+    label_6dof = np.concatenate([label_euler, label_translation])
+    return label_6dof
+
+
+def gen_target_pip(target, target_map, target_local_x, target_local_y):
+    map_channel, map_height, map_width = target_map.shape
+    target = target.reshape(-1, 2)
+    assert map_channel == target.shape[0]
+
+    for i in range(map_channel):
+        mu_x = int(math.floor(target[i][0] * map_width))
+        mu_y = int(math.floor(target[i][1] * map_height))
+        mu_x = max(0, mu_x)
+        mu_y = max(0, mu_y)
+        mu_x = min(mu_x, map_width-1)
+        mu_y = min(mu_y, map_height-1)
+        target_map[i, mu_y, mu_x] = 1
+        shift_x = target[i][0] * map_width - mu_x
+        shift_y = target[i][1] * map_height - mu_y
+        target_local_x[i, mu_y, mu_x] = shift_x
+        target_local_y[i, mu_y, mu_x] = shift_y
+
+
+    return target_map, target_local_x, target_local_y
+
+def get_tris(cfg):
+    import trimesh
+    data_root = Path(cfg.root_dir)
+    obj_path = data_root / 'resources/example.obj'
+    mesh = trimesh.load(obj_path, process=False)
+    verts_template = np.array(mesh.vertices, dtype=np.float32)
+    tris = np.array(mesh.faces, dtype=np.int32)
+    #print(verts_template.shape, tris.shape)
+    return tris
+
+
+class BackgroundGenerator(threading.Thread):
+    def __init__(self, generator, local_rank, max_prefetch=6):
+        super(BackgroundGenerator, self).__init__()
+        self.queue = Queue.Queue(max_prefetch)
+        self.generator = generator
+        self.local_rank = local_rank
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        torch.cuda.set_device(self.local_rank)
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def next(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __next__(self):
+        return self.next()
+
+    def __iter__(self):
+        return self
+
+
+class DataLoaderX(DataLoader):
+    def __init__(self, local_rank, **kwargs):
+        super(DataLoaderX, self).__init__(**kwargs)
+        self.stream = torch.cuda.Stream(local_rank)
+        self.local_rank = local_rank
+
+    def __iter__(self):
+        self.iter = super(DataLoaderX, self).__iter__()
+        self.iter = BackgroundGenerator(self.iter, self.local_rank)
+        self.preload()
+        return self
+
+    def preload(self):
+        self.batch = next(self.iter, None)
+        if self.batch is None:
+            return None
+        with torch.cuda.stream(self.stream):
+            for k in range(len(self.batch)):
+                self.batch[k] = self.batch[k].to(device=self.local_rank,
+                                                 non_blocking=True)
+
+    def __next__(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is None:
+            raise StopIteration
+        self.preload()
+        return batch
+
+
+class FaceDataset(Dataset):
+    def __init__(self, cfg, is_train=True, is_test=False, local_rank=0):
+        super(FaceDataset, self).__init__()
+
+        
+        self.data_root = Path(cfg.root_dir)
+        self.input_size = cfg.input_size
+        self.transform = get_aug_transform(cfg)
+        self.local_rank = local_rank
+        self.is_test = is_test
+        txt_path = self.data_root / 'resources/projection_matrix.txt'
+        self.M_proj = np.loadtxt(txt_path, dtype=np.float32)
+        if is_test:
+            data_root = Path(cfg.root_dir)
+            csv_path = data_root / 'list/WCPA_track2_test.csv'
+            self.df = pd.read_csv(csv_path, dtype={'subject_id': str, 'facial_action': str, 'img_id': str})
+        else:
+            if is_train:
+                self.df = pd.read_csv(osp.join(cfg.cache_dir, 'train_list.csv'), dtype={'subject_id': str, 'facial_action': str, 'img_id': str})
+            else:
+                self.df = pd.read_csv(osp.join(cfg.cache_dir, 'val_list.csv'), dtype={'subject_id': str, 'facial_action': str, 'img_id': str})
+        self.label_6dof_mean = [-0.018197, -0.017891, 0.025348, -0.005368, 0.001176, -0.532206]   # mean of pitch, yaw, roll, tx, ty, tz
+        self.label_6dof_std = [0.314015, 0.271809, 0.081881, 0.022173, 0.048839, 0.065444]        # std of pitch, yaw, roll, tx, ty, tz
+        self.align_face = cfg.align_face
+        if not self.align_face:
+            self.dst_pts = np.float32([
+                [0, 0],
+                [0, cfg.input_size- 1],
+                [cfg.input_size- 1, 0]
+            ])
+        else:
+            dst_pts = np.array([
+                [38.2946, 51.6963],
+                [73.5318, 51.5014],
+                [56.0252, 71.7366],
+                [41.5493, 92.3655],
+                [70.7299, 92.2041] ], dtype=np.float32 )
+
+            new_size = 144
+            dst_pts[:,0] += ((new_size-112)//2)
+            dst_pts[:,1] += 8
+            dst_pts[:,:] *= (self.input_size/float(new_size))
+            self.dst_pts = dst_pts
+
+        if local_rank==0:
+            logging.info('data_transform_list:%s'%self.transform)
+            logging.info('len:%d'%len(self.df))
+        self.is_test_aug = False
+        self.eye_dataset = None
+        if cfg.eyes is not None:
+            from eye_dataset import EyeDataset
+            self.eye_dataset = EyeDataset(cfg.eyes['root'])
+
+    def set_test_aug(self):
+        if not self.is_test_aug:
+            from easydict import EasyDict as edict
+            cfg = edict()
+            cfg.aug_modes = ['test-aug']
+            cfg.input_size = self.input_size
+            cfg.task = 0
+            self.transform = get_aug_transform(cfg)
+            self.is_test_aug = True
+
+    def get_names(self, index):
+        subject_id = self.df['subject_id'][index]
+        facial_action = self.df['facial_action'][index]
+        img_id = self.df['img_id'][index]
+        return subject_id, facial_action, img_id
+
+    def __getitem__(self, index):
+        subject_id = self.df['subject_id'][index]
+        facial_action = self.df['facial_action'][index]
+        img_id = self.df['img_id'][index]
+
+        img_path = self.data_root / 'image' / subject_id / facial_action / f'{img_id}_ar.jpg'
+        npz_path = self.data_root / 'info' / subject_id / facial_action / f'{img_id}_info.npz'
+        txt_path = self.data_root / '68landmarks' / subject_id / facial_action / f'{img_id}_68landmarks.txt'
+        #if not osp.exists(img_path):
+        #    continue
+
+        #print(img_path)
+        img_raw = cv2.imread(str(img_path))
+        #if img_raw is None:
+        #    print('XXX ERR:', img_path)
+        img_raw = cv2.cvtColor(img_raw, cv2.COLOR_BGR2RGB)
+        #print(img_raw.shape)
+        img_h, img_w, _ = img_raw.shape
+        pts68 = np.loadtxt(txt_path, dtype=np.int32)
+
+        x_min, y_min = pts68.min(axis=0)
+        x_max, y_max = pts68.max(axis=0)
+        x_center = (x_min + x_max) / 2
+        y_center = (y_min + y_max) / 2
+        w, h = x_max - x_min, y_max - y_min
+
+
+        if not self.align_face:
+            size = max(w, h)
+            ss = np.array([0.75, 0.75, 0.85, 0.65])     # predefined expand size
+
+            left = x_center - ss[0] * size
+            right = x_center + ss[1] * size
+            top = y_center - ss[2] * size
+            bottom = y_center + ss[3] * size
+
+            src_pts = np.float32([
+                [left, top],
+                [left, bottom],
+                [right, top]
+            ])
+            tform = cv2.getAffineTransform(src_pts, self.dst_pts)
+        else:
+            src_pts = np.float32([
+                (pts68[36] + pts68[39])/2,
+                (pts68[42] + pts68[45])/2,
+                pts68[30],
+                pts68[48],
+                pts68[54]
+            ])
+            tf = sktrans.SimilarityTransform()
+            tf.estimate(src_pts, self.dst_pts)
+            tform = tf.params[0:2,:]
+
+        img_local = cv2.warpAffine(img_raw, tform, (self.input_size,)*2, flags=cv2.INTER_CUBIC)
+        fake_points2d = np.ones( (1,2), dtype=np.float32) * (self.input_size//2)
+
+        #tform_inv = cv2.invertAffineTransform(tform)
+        #img_global = cv2.warpAffine(img_local, tform_inv, (img_w, img_h), borderValue=0.0)
+        #img_global = cv2.resize(img_global, (self.input_size, self.input_size))
+        if self.transform is not None:
+            t = self.transform(image=img_local, keypoints=fake_points2d)
+            img_local = t['image']
+            if self.is_test_aug:
+                height, width = img_local.shape[:2]
+                for trans in t["replay"]["transforms"]:
+                    if trans['__class_fullname__']=='ShiftScaleRotate' and trans['applied']:
+                        param = trans['params']
+                        dx, dy, angle, scale = param['dx'], param['dy'], param['angle'], param['scale']
+                        center = (width / 2, height / 2)
+                        matrix = cv2.getRotationMatrix2D(center, angle, scale)
+                        matrix[0, 2] += dx * width
+                        matrix[1, 2] += dy * height
+                        new_matrix = np.identity(3)
+                        new_matrix[:2,:3] = matrix
+                        old_tform = np.identity(3)
+                        old_tform[:2,:3] = tform
+                        #new_tform = np.dot(old_tform, new_matrix)
+                        new_tform = np.dot(new_matrix, old_tform)
+                        #print('label_tform:')
+                        #print(label_tform.flatten())
+                        #print(new_matrix.flatten())
+                        #print(new_tform.flatten())
+                        tform = new_tform[:2,:3]
+                        break
+                            #print('trans param:', param)
+            #img_global = self.transform(image=img_global)['image']
+
+        tform_tensor = torch.tensor(tform, dtype=torch.float32)
+        d = {'img_local': img_local, 'tform': tform_tensor}
+        if self.eye_dataset is not None:
+            eye_key = str(Path('image') / subject_id / facial_action / f'{img_id}_ar.jpg')
+            #print(eye_key)
+            eyel, eyer = self.eye_dataset.get(eye_key, to_homo=True)
+            if eyel is not None:
+                #print(eye_key, el_inv.shape, er_inv.shape)
+                d['eye_world_left'] = torch.tensor(eyel, dtype=torch.float32)
+                d['eye_world_right'] = torch.tensor(eyer, dtype=torch.float32)
+        if not self.is_test:
+            M = np.load(npz_path)
+            #yaw_gt, pitch_gt, roll_gt = Rotation.from_matrix(M['R_t'][:3, :3].T).as_euler('yxz', degrees=False)
+            #label_euler = np.array([pitch_gt, yaw_gt, roll_gt])
+            #label_translation = M['R_t'][3, :3]
+            #label_6dof = np.concatenate([label_euler, label_translation])
+            #label_6dof = (label_6dof - self.label_6dof_mean) / self.label_6dof_std
+            #label_6dof_tensor = torch.tensor(label_6dof, dtype=torch.float32)
+            #label_verts = M['verts'] * 10.0     # roughly [-1, 1]
+            #label_verts_tensor = torch.tensor(label_verts, dtype=torch.float32)
+            #return img_local, label_verts_tensor, label_6dof_tensor
+            label_verts_tensor = torch.tensor(M['verts'], dtype=torch.float32)
+            label_Rt_tensor = torch.tensor(M['R_t'], dtype=torch.float32)
+            d['verts'] = label_verts_tensor
+            d['rt'] = label_Rt_tensor
+            #return img_local, img_global, label_verts_tensor, label_Rt_tensor, tform_tensor
+            #return img_local, label_verts_tensor, label_Rt_tensor, tform_tensor
+        else:
+            #return img_local, img_global, tform_tensor
+            index_tensor = torch.tensor(index, dtype=torch.long)
+            d['index'] = index_tensor
+            #return img_local, tform_tensor, index_tensor
+        return d
+
+
+    def __len__(self):
+        return len(self.df)
+
+class MXFaceDataset(Dataset):
+    def __init__(self, cfg, is_train=True, norm_6dof=True, degrees_6dof=False, local_rank=0):
+        super(MXFaceDataset, self).__init__()
+
+        
+        self.is_train = is_train
+        self.data_root = Path(cfg.root_dir)
+        self.input_size = cfg.input_size
+        self.transform = get_aug_transform(cfg)
+        self.local_rank = local_rank
+        self.use_trainval = cfg.use_trainval
+        self.use_eye = cfg.eyes is not None
+        if is_train:
+            #self.df = pd.read_csv(osp.join(cfg.cache_dir, 'train_list.csv'), dtype={'subject_id': str, 'facial_action': str, 'img_id': str})
+            path_imgrec = os.path.join(cfg.cache_dir, 'train.rec')
+            path_imgidx = os.path.join(cfg.cache_dir, 'train.idx')
+            self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+            self.imgidx = list(self.imgrec.keys)
+            self.imggroup = [0] * len(self.imgidx)
+            self.size_train = len(self.imgidx)
+            if self.use_trainval:
+                assert not cfg.sampling_hard
+                path_imgrec = os.path.join(cfg.cache_dir, 'val.rec')
+                path_imgidx = os.path.join(cfg.cache_dir, 'val.idx')
+                self.imgrec2 = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+                imgidx2 = list(self.imgrec2.keys)
+                self.imggroup += [1] * len(imgidx2)
+                self.imgidx += imgidx2
+        else:
+            #self.df = pd.read_csv(osp.join(cfg.cache_dir, 'val_list.csv'), dtype={'subject_id': str, 'facial_action': str, 'img_id': str})
+            path_imgrec = os.path.join(cfg.cache_dir, 'val.rec')
+            path_imgidx = os.path.join(cfg.cache_dir, 'val.idx')
+            self.imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')
+            self.imgidx = list(self.imgrec.keys)
+            self.imggroup = [0] * len(self.imgidx)
+        self.imgidx = np.array(self.imgidx)
+        self.imggroup = np.array(self.imggroup)
+        if cfg.sampling_hard and is_train:
+            meta = np.load(os.path.join(cfg.cache_dir, 'train.meta.npy'))
+            assert meta.shape[0]==len(self.imgidx)
+            new_imgidx = []
+            for i in range(len(self.imgidx)):
+                idx = self.imgidx[i]
+                assert i==idx
+                pose = np.abs(meta[i,:2])
+                #repeat = np.sum(pose>=35)*3+1
+                if np.max(pose)<15:
+                    repeat = 2
+                else:
+                    repeat = 1
+                new_imgidx += [idx]*repeat
+            if local_rank==0:
+                print('new-imgidx:', len(self.imgidx), len(new_imgidx))
+            self.imgidx = np.array(new_imgidx)
+        self.label_6dof_mean = [-0.018197, -0.017891, 0.025348, -0.005368, 0.001176, -0.532206]   # mean of pitch, yaw, roll, tx, ty, tz
+        self.label_6dof_std = [0.314015, 0.271809, 0.081881, 0.022173, 0.048839, 0.065444]        # std of pitch, yaw, roll, tx, ty, tz
+        txt_path = self.data_root / 'resources/projection_matrix.txt'
+        self.M_proj = np.loadtxt(txt_path, dtype=np.float32)
+        self.M1 = np.array([
+            [400.0,       0, 0, 0],
+            [      0, 400.0, 0, 0],
+            [      0,       0, 1, 0],
+            [400.0, 400.0, 0, 1]
+        ])
+        self.dst_pts = np.float32([
+            [0, 0],
+            [0, cfg.input_size- 1],
+            [cfg.input_size- 1, 0]
+        ])
+        self.norm_6dof = norm_6dof
+        self.degrees_6dof = degrees_6dof
+        self.task = cfg.task
+        self.num_verts = cfg.num_verts
+        self.loss_pip = cfg.loss_pip
+        self.net_stride = 32
+        if local_rank==0:
+            logging.info('data_transform_list:%s'%self.transform)
+            logging.info('len:%d'%len(self.imgidx))
+            logging.info('glen:%d'%len(self.imggroup))
+        self.is_test_aug = False
+
+        self.enable_flip = cfg.enable_flip
+        self.flipindex = cfg.flipindex.copy()
+        self.verts3d_central_index = cfg.verts3d_central_index
+        self.eye_dataset = None
+        self.use_eye = False
+        if cfg.eyes is not None:
+            #from eye_dataset import EyeDataset
+            #self.eye_dataset = EyeDataset(cfg.eyes['root'], load_data=False)
+            self.use_eye = True
+
+    def set_test_aug(self):
+        if not self.is_test_aug:
+            from easydict import EasyDict as edict
+            cfg = edict()
+            cfg.aug_modes = ['test-aug']
+            cfg.input_size = self.input_size
+            cfg.task = 0
+            self.transform = get_aug_transform(cfg)
+            self.is_test_aug = True
+
+    def __getitem__(self, index):
+        idx = self.imgidx[index]
+        group = self.imggroup[index]
+        if group==0:
+            imgrec = self.imgrec
+        elif group==1:
+            imgrec = self.imgrec2
+        elif group==2:
+            imgrec = self.imgrec3
+
+        s = imgrec.read_idx(idx)
+        header, img = mx.recordio.unpack(s)
+        hlabel = header.label
+        img = mx.image.imdecode(img).asnumpy() #rgb numpy
+
+        label_verts = np.array(hlabel[:1220*3], dtype=np.float32).reshape(-1,3)
+        label_Rt = np.array(hlabel[1220*3:1220*3+16], dtype=np.float32).reshape(4,4)
+        label_tform = np.array(hlabel[1220*3+16:1220*3+16+6], dtype=np.float32).reshape(2,3)
+        label_6dof = Rt26dof(label_Rt, self.degrees_6dof)
+        if self.norm_6dof:
+            label_6dof = (label_6dof - self.label_6dof_mean) / self.label_6dof_std
+        label_6dof_tensor = torch.tensor(label_6dof, dtype=torch.float32)
+        el_inv = None
+        er_inv = None
+        if self.use_eye:
+            a = 1220*3+16+6
+            el_inv = np.array(hlabel[a:a+481*3], dtype=np.float32).reshape(-1,3)
+            a+=481*3
+            er_inv = np.array(hlabel[a:a+481*3], dtype=np.float32).reshape(-1,3)
+            #el_inv = torch.tensor(el_inv, dtype=torch.float32)
+            #er_inv = torch.tensor(er_inv, dtype=torch.float32)
+            #eye_verts = [el_inv, er_inv]
+            eye_verts = np.concatenate( (el_inv, er_inv), axis=0 )
+
+        #img_local = None
+        img_raw = None
+        #if self.task==0 or self.task==2:
+        #    img_raw = img[:,self.input_size:,:]
+        #if self.task==0 or self.task==1 or self.task==3:
+        #    img_local = img[:,:self.input_size,:]
+        assert img.shape[0]==img.shape[1] and img.shape[0]>=self.input_size
+        if img.shape[0]>self.input_size:
+            scale = float(self.input_size) / img.shape[0]
+            #print('scale:', scale)
+            #src_pts = np.float32([
+            #    [0, 0],
+            #    [0, 799],
+            #    [799, 0]
+            #])
+            #tform = cv2.getAffineTransform(src_pts, self.dst_pts)
+            #new_tform = np.identity(3)
+            #new_tform[:2,:3] = tform
+            #label_tform = np.dot(new_tform, label_tform.T).T
+
+            src_pts = np.float32([
+                [0, 0, 1],
+                [0, 799, 1],
+                [799, 0, 1]
+            ])
+            dst_pts = np.dot(label_tform, src_pts.T).T
+            dst_pts *= scale
+            dst_pts = dst_pts.copy()
+            src_pts = src_pts[:,:2].copy()
+            #print('index:', index)
+            #print(src_pts.shape, dst_pts.shape)
+            #print(label_tform.shape)
+            #print(src_pts.dtype)
+            #print(dst_pts.dtype)
+            tform = cv2.getAffineTransform(src_pts, dst_pts)
+            label_tform = tform
+
+            img = cv2.resize(img, (self.input_size, self.input_size))
+
+        img_local = img
+        need_points2d = (self.task==0 or self.task==3)
+
+        if need_points2d:
+            ones = np.ones([label_verts.shape[0], 1])
+            verts_homo = np.concatenate([label_verts, ones], axis=1)
+            verts = verts_homo @ label_Rt @ self.M_proj @ self.M1
+            w_ = verts[:, [3]]
+            verts = verts / w_
+            points2d = verts[:, :3]
+            points2d[:, 1] = 800.0 - points2d[:, 1]
+            verts2d = points2d[:,:2].copy()
+            points2d[:,2] = 1.0
+            points2d = np.dot(label_tform, points2d.T).T
+        else:
+            points2d = np.ones( (1,2), dtype=np.float32) * (self.input_size//2)
+        if self.use_eye:
+            verts_homo = eye_verts
+            if verts_homo.shape[1] == 3:
+                ones = np.ones([verts_homo.shape[0], 1])
+                verts_homo = np.concatenate([verts_homo, ones], axis=1)
+            verts_out = verts_homo @ label_Rt @ self.M_proj @ self.M1
+            w_ = verts_out[:, [3]]
+            verts_out = verts_out / w_
+            _points2d = verts_out[:, :3]
+            _points2d[:, 1] = 800.0 - _points2d[:, 1]
+            _points2d[:,2] = 1.0
+            _points2d = np.dot(label_tform, _points2d.T).T
+            eye_points = _points2d
+        #if img.shape[0]!=self.input_size:
+        #    assert img.shape[0]>self.input_size
+            #img = cv2.resize(img, (self.input_size, self.input_size))
+            #scale = float(self.input_size) / img.shape[0]
+            #points2d *= scale
+
+        if self.transform is not None:
+            if img_raw is not None:
+                img_raw = self.transform(image=img_raw, keypoints=points2d)['image']
+            if img_local is not None:
+                height, width = img_local.shape[:2]
+                x = self.transform(image=img_local, keypoints=points2d)
+                img_local = x['image']
+                points2d = x['keypoints']
+                points2d = np.array(points2d, dtype=np.float32)
+                if self.is_test_aug:
+                    for trans in x["replay"]["transforms"]:
+                        if trans['__class_fullname__']=='ShiftScaleRotate' and trans['applied']:
+                            param = trans['params']
+                            dx, dy, angle, scale = param['dx'], param['dy'], param['angle'], param['scale']
+                            center = (width / 2, height / 2)
+                            matrix = cv2.getRotationMatrix2D(center, angle, scale)
+                            matrix[0, 2] += dx * width
+                            matrix[1, 2] += dy * height
+                            new_matrix = np.identity(3)
+                            new_matrix[:2,:3] = matrix
+                            old_tform = np.identity(3)
+                            old_tform[:2,:3] = label_tform
+                            #new_tform = np.dot(old_tform, new_matrix)
+                            new_tform = np.dot(new_matrix, old_tform)
+                            #print('label_tform:')
+                            #print(label_tform.flatten())
+                            #print(new_matrix.flatten())
+                            #print(new_tform.flatten())
+                            label_tform = new_tform[:2,:3]
+                            break
+                            #print('trans param:', param)
+
+
+        if self.loss_pip:
+            target_map = np.zeros((self.num_verts, int(self.input_size/self.net_stride), int(self.input_size/self.net_stride)))
+            target_local_x = np.zeros((self.num_verts, int(self.input_size/self.net_stride), int(self.input_size/self.net_stride)))
+            target_local_y = np.zeros((self.num_verts, int(self.input_size/self.net_stride), int(self.input_size/self.net_stride)))
+            target = points2d / self.input_size
+            target_map, target_local_x, target_local_y = gen_target_pip(target, target_map, target_local_x, target_local_y)
+            target_map_tensor = torch.tensor(target_map, dtype=torch.float32)
+            target_x_tensor = torch.tensor(target_local_x, dtype=torch.float32)
+            target_y_tensor = torch.tensor(target_local_y, dtype=torch.float32)
+            d['pip_map'] = target_map_tensor
+            d['pip_x'] = target_x_tensor
+            d['pip_y'] = target_y_tensor
+
+        if self.is_train and self.enable_flip and np.random.random()<0.5:
+            #if self.local_rank==0:
+            #    print('XXX:', label_verts[:5,:2])
+            img_local = img_local.flip([2])
+            x_of_central = 0.0
+            #x_of_central = label_verts[self.verts3d_central_index,0]
+            #x_of_central = np.mean(x_of_central)
+            label_verts = label_verts[self.flipindex,:]
+            label_verts[:,0] -= x_of_central
+            label_verts[:,0] *= -1.0
+            label_verts[:,0] += x_of_central
+
+            if need_points2d:
+                flipped_p2d = points2d[self.flipindex,:].copy()
+                flipped_p2d[:,0] = self.input_size - 1 - flipped_p2d[:,0]
+                points2d = flipped_p2d
+            if self.use_eye:
+                flipped_p2d = eye_points[self.flipindex,:].copy()
+                flipped_p2d[:,0] = self.input_size - 1 - flipped_p2d[:,0]
+                eye_points = flipped_p2d
+        label_verts_tensor = torch.tensor(label_verts*10.0, dtype=torch.float32)
+        d = {}
+        d['img_local'] = img_local
+        d['verts'] = label_verts_tensor
+        d['6dof'] = label_6dof_tensor
+        d['rt'] = torch.tensor(label_Rt, dtype=torch.float32)
+        if need_points2d:
+            points2d = points2d / (self.input_size//2) - 1.0
+            points2d_tensor = torch.tensor(points2d, dtype=torch.float32)
+            d['points2d'] = points2d_tensor
+        if self.use_eye:
+            d['eye_verts'] = torch.tensor(eye_verts, dtype=torch.float32)
+            eye_points = eye_points / (self.input_size//2) - 1.0
+            eye_points_tensor = torch.tensor(eye_points, dtype=torch.float32)
+            d['eye_points'] = eye_points_tensor
+
+        loss_weight = 1.0
+        if group!=0:
+            loss_weight = 0.0
+        loss_weight_tensor = torch.tensor(loss_weight, dtype=torch.float32)
+        d['loss_weight'] = loss_weight_tensor
+        label_tform_tensor = torch.tensor(label_tform, dtype=torch.float32)
+        d['tform'] = label_tform_tensor
+
+        #if img_local is None:
+        #    image = (img_raw,)
+        #elif img_raw is None:
+        #    image = (img_local,)
+        #else:
+        #    image = (img_local,img_raw)
+        #ret = image + (label_verts_tensor, label_6dof_tensor, points2d_tensor)
+        if not self.is_train:
+            idx_tensor = torch.tensor([idx], dtype=torch.long)
+            d['idx'] = idx_tensor
+            d['verts2d'] = torch.tensor(verts2d, dtype=torch.float32)
+        return d
+
+
+    def __len__(self):
+        return len(self.imgidx)
+
+def test_dataset1(cfg):
+    cfg.task = 0
+    is_train = False
+    center_axis = []
+    dataset = MXFaceDataset(cfg, is_train=is_train, norm_6dof=False, local_rank=0)
+    for i in range(len(dataset.flipindex)):
+        if i==dataset.flipindex[i]:
+            center_axis.append(i)
+    print(center_axis)
+    #dataset.transform = None
+    print('total:', len(dataset))
+    total = len(dataset)
+    #total = 50
+    list_6dof = []
+    all_mean_xs = []
+    for idx in range(total):
+        #img_local, img_raw, label_verts, label_6dof, = dataset[idx]
+        #img_local, img_raw, label_verts, label_6dof, points2d, tform, data_idx = dataset[idx]
+        #img_local, label_verts, label_6dof, points2d, tform, data_idx = dataset[idx]
+        d = dataset[idx]
+        img_local = d['img_local']
+        label_verts = d['verts']
+        label_6dof = d['6dof']
+        points2d = d['points2d']
+        label_verts = label_verts.numpy()
+        label_6dof = label_6dof.numpy()
+        points2d = points2d.numpy()
+        #print(img_local.shape, label_verts.shape, label_6dof.shape, points2d.shape)
+        verts3d = label_verts / 10.0
+        xs = []
+        for c in center_axis:
+            _x = verts3d[c,0]
+            xs.append(_x)
+        _std = np.std(xs)
+        print(xs)
+        print(_std)
+        #print(np.mean(xs))
+        all_mean_xs.append(np.mean(xs))
+        if idx%100==0:
+            print('processing:', idx, np.mean(all_mean_xs))
+        #print(label_verts[:3,:], label_6dof)
+        #list_6dof.append(label_6dof)
+        #print(image.__class__, label_verts.__class__)
+        #label = list(label_verts.numpy().flatten()) + list(label_6dof.numpy().flatten())
+        #points2d = label_verts2[:,:2]
+        #points2d = (points2d+1) * 128.0
+        #img_local = img_local.numpy()
+        #img_local = (img_local+1.0) * 128.0
+        #draw = img_local.astype(np.uint8).transpose( (1,2,0) )[:,:,::-1].copy()
+        #for i in range(points2d.shape[0]):
+        #    pt = points2d[i].astype(np.int)
+        #    cv2.circle(draw, pt, 2, (255,0,0), 2)
+        ##output_path = "outputs/%d_%.3f_%.3f_%.3f.jpg"%(idx, label_6dof[0], label_6dof[1], label_6dof[2])
+        #output_path = "outputs/%06d.jpg"%(idx)
+        #cv2.imwrite(output_path, draw)
+    #list_6dof = np.array(list_6dof)
+    #print('MEAN:')
+    #print(np.mean(list_6dof, axis=0))
+
+def test_loader1(cfg):
+    cfg.task = 0
+    is_train = True
+    dataset = MXFaceDataset(cfg, is_train=is_train, norm_6dof=False, local_rank=0)
+    loader = DataLoader(dataset, batch_size=64, shuffle=True)
+    for index, d in enumerate(loader):
+        #img_local = d['img_local']
+        label_verts = d['verts']
+        points2d = d['points2d']
+        tform = d['tform']
+        label_verts /= 10.0
+        points2d = (points2d + 1.0) * (cfg.input_size//2)
+        tform = tform.numpy()
+        verts = label_verts.numpy()
+        points2d = points2d.numpy()
+        print(verts.shape, points2d.shape, tform.shape)
+        np.save("temp/verts3d.npy", verts)
+        np.save("temp/points2d.npy", points2d)
+        np.save("temp/tform.npy", tform)
+        break
+
+def test_facedataset1(cfg):
+    cfg.task = 0
+    cfg.input_size = 512
+    dataset = FaceDataset(cfg, is_train=True, local_rank=0)
+    for idx in range(100000):
+        img_local, label_verts, label_Rt, tform = dataset[idx]
+        label_Rt = label_Rt.numpy()
+        if label_Rt[0,0]>1.0:
+            print(idx, label_Rt.shape)
+            print(label_Rt)
+            break
+
+def test_arcface(cfg):
+    cfg.task = 0
+    is_train = True
+    dataset = MXFaceDataset(cfg, is_train=is_train, norm_6dof=False, local_rank=0)
+    loader = DataLoader(dataset, batch_size=1, shuffle=True)
+    for index, d in enumerate(loader):
+        img = d['img_local'].numpy()
+        img /= 2.0
+        img += 0.5
+        img *= 255.0
+        img = img[0]
+        img = img.transpose( (1,2,0) )
+        img = img.astype(np.uint8)
+        img = cv2.resize(img, (144,144))
+        img = img[:,:,::-1]
+        img = img[8:120,16:128,:]
+        print(img.shape)
+        cv2.imwrite("temp/arc_%d.jpg"%index, img)
+        #np.save("temp/verts3d.npy", verts)
+        #np.save("temp/points2d.npy", points2d)
+        #np.save("temp/tform.npy", tform)
+        if index>100:
+            break
+
+def test_dataset2(cfg):
+    cfg.task = 0
+    is_train = False
+    center_axis = []
+    dataset = MXFaceDataset(cfg, is_train=is_train, norm_6dof=False, local_rank=0)
+    for i in range(len(dataset.flipindex)):
+        if i==dataset.flipindex[i]:
+            center_axis.append(i)
+    print(center_axis)
+    #dataset.transform = None
+    print('total:', len(dataset))
+    total = len(dataset)
+    total = 50
+    list_6dof = []
+    all_mean_xs = []
+    for idx in range(total):
+        d = dataset[idx]
+        img_local = d['img_local']
+        label_verts = d['verts']
+        label_6dof = d['6dof']
+        points2d = d['points2d']
+        label_verts = label_verts.numpy()
+        label_6dof = label_6dof.numpy()
+        points2d = points2d.numpy()
+        eye_points = d['eye_points'].numpy()
+        eye_verts = d['eye_verts'].numpy()
+        print(eye_verts[:5,:])
+        #print(img_local.shape, label_verts.shape, label_6dof.shape, points2d.shape)
+        verts3d = label_verts / 10.0
+        #print(label_verts[:3,:], label_6dof)
+        #list_6dof.append(label_6dof)
+        #print(image.__class__, label_verts.__class__)
+        #label = list(label_verts.numpy().flatten()) + list(label_6dof.numpy().flatten())
+        #points2d = label_verts2[:,:2]
+        points2d = (points2d+1) * 128.0
+        eye_points = (eye_points+1) * 128.0
+        img_local = img_local.numpy()
+        img_local = (img_local+1.0) * 128.0
+        draw = img_local.astype(np.uint8).transpose( (1,2,0) )[:,:,::-1].copy()
+        for i in range(points2d.shape[0]):
+            pt = points2d[i].astype(np.int)
+            cv2.circle(draw, pt, 2, (255,0,0), 2)
+        for i in range(eye_points.shape[0]):
+            pt = eye_points[i].astype(np.int)
+            cv2.circle(draw, pt, 2, (0,255,0), 2)
+        ##output_path = "outputs/%d_%.3f_%.3f_%.3f.jpg"%(idx, label_6dof[0], label_6dof[1], label_6dof[2])
+        output_path = "outputs/%06d.jpg"%(idx)
+        cv2.imwrite(output_path, draw)
+    #list_6dof = np.array(list_6dof)
+    #print('MEAN:')
+    #print(np.mean(list_6dof, axis=0))
+
+if __name__ == "__main__":
+    from utils.utils_config import get_config
+    #cfg = get_config('configs/r0_a1.py')
+    cfg = get_config('configs/s2')
+    #test_loader1(cfg)
+    #test_facedataset1(cfg)
+    #test_arcface(cfg)
+    test_dataset2(cfg)
+
+
diff --git a/reconstruction/jmlr/eye_dataset.py b/reconstruction/jmlr/eye_dataset.py
new file mode 100644
index 000000000..ecc929d21
--- /dev/null
+++ b/reconstruction/jmlr/eye_dataset.py
@@ -0,0 +1,71 @@
+import os
+import os.path as osp
+import numpy as np
+
+import menpo.io as mio
+
+def project_shape_in_image(verts, R_t, M_proj, M1):
+    verts_homo = verts
+    if verts_homo.shape[1] == 3:
+        ones = np.ones([verts_homo.shape[0], 1])
+        verts_homo = np.concatenate([verts_homo, ones], axis=1)
+    verts_out = verts_homo @ R_t @ M_proj @ M1
+    w_ = verts_out[:, [3]]
+    verts_out = verts_out / w_
+    return verts_out
+
+class EyeDataset():
+    def __init__(self, root, load_data=True):
+        eyes_info = mio.import_pickle(osp.join(root,'eyes3d.pkl'))
+        idxs481 = eyes_info['mask481']['idxs']
+        tri481 = eyes_info['mask481']['trilist']
+        self.iris_idx_481 = eyes_info['mask481']['idxs_iris']
+        eyel_template = eyes_info['left_points'][idxs481]
+        eyer_template = eyes_info['right_points'][idxs481]
+        eyel_template_homo = np.append(eyel_template, np.ones((eyel_template.shape[0],1)), axis=1)
+        eyer_template_homo = np.append(eyer_template, np.ones((eyer_template.shape[0],1)), axis=1)
+        points = mio.import_pickle(osp.join(root,'eyespoints.pkl'))
+        self.homol = eyel_template_homo.T
+        self.homor = eyer_template_homo.T
+        if load_data:
+            self.worldl = {}
+            self.worldr = {}
+            #vector_norm = 0.035
+            for k in points:
+                p = k.find('/')
+                newk = k[p+1:]
+                value = points[k]
+                #el_inv = (value['left'] @ eyel_template_homo.T).T
+                #er_inv = (value['right'] @ eyer_template_homo.T).T
+                #print('V:', value['left'][:5,:])
+                #print('E:', el_inv[:5,:])
+                # gaze vector of left eye in world space
+                #gl_vector = el_inv[iris_idx_481].mean(axis=0) - el_inv[-1]
+                #gl_vector = (gl_vector / np.linalg.norm(gl_vector)) * vector_norm
+                #gl_point = el_inv[iris_idx_481].mean(axis=0) + gl_vector
+                ## gaze vector of right eye in world space
+                #gr_vector = er_inv[iris_idx_481].mean(axis=0) - er_inv[-1]
+                #gr_vector = (gr_vector / np.linalg.norm(gr_vector)) * vector_norm
+                #gr_point = er_inv[iris_idx_481].mean(axis=0) + gr_vector
+                #self.world[newk] = (el_inv, er_inv, gl_point, gr_point)
+                self.worldl[newk] = value['left']
+                self.worldr[newk] = value['right']
+            #print(self.points.keys())
+
+    def get(self, key, to_homo=False):
+        if key not in self.worldl:
+            return None, None
+        left = self.worldl[key]
+        right = self.worldr[key]
+        if to_homo:
+            left = (left @ self.homol).T
+            right = (right @ self.homor).T
+        return left, right
+
+    def to_homo(self, eyel, eyer):
+        eyel = (eyel @ self.homol).T
+        eyer = (eyer @ self.homor).T
+        return eyel, eyer
+
+
+
diff --git a/reconstruction/jmlr/flops.py b/reconstruction/jmlr/flops.py
new file mode 100644
index 000000000..605465db2
--- /dev/null
+++ b/reconstruction/jmlr/flops.py
@@ -0,0 +1,25 @@
+from ptflops import get_model_complexity_info
+import os
+import argparse
+from utils.utils_config import get_config
+from backbones import get_network
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser(description='JMLR FLOPs')
+    parser.add_argument('config', type=str, help='input config file')
+    args = parser.parse_args()
+    args = parser.parse_args()
+    cfg = get_config(args.config)
+    #backbone = get_model(cfg.network, num_features=cfg.embedding_size, input_size=cfg.input_size, dropout=cfg.dropout, stem_type=cfg.stem_type, fp16=0)
+    net = get_network(cfg)
+    macs, params = get_model_complexity_info(
+        net, (3, cfg.input_size, cfg.input_size), as_strings=True,
+        print_per_layer_stat=True, verbose=True)
+    print(macs)
+    print(params)
+
+    # from torch import distributed
+    # distributed.AllreduceOptions
+    # distributed.AllreduceCoalescedOptions
+    # distributed.all_reduce
diff --git a/reconstruction/jmlr/gen_dataset_meta.py b/reconstruction/jmlr/gen_dataset_meta.py
new file mode 100644
index 000000000..108cdfe8d
--- /dev/null
+++ b/reconstruction/jmlr/gen_dataset_meta.py
@@ -0,0 +1,35 @@
+import pickle
+import numpy as np
+import os
+import os.path as osp
+import glob
+import argparse
+import cv2
+import time
+import datetime
+import pickle
+import sklearn
+import mxnet as mx
+from utils.utils_config import get_config
+from dataset import MXFaceDataset, Rt26dof
+
+if __name__ == "__main__":
+    cfg = get_config('configs/s1.py')
+    cfg.task = 0
+    save_path = os.path.join(cfg.cache_dir, 'train.meta')
+    assert not osp.exists(save_path)
+    dataset = MXFaceDataset(cfg, is_train=True, norm_6dof=False, degrees_6dof=True, local_rank=0)
+    #dataset.transform = None
+    print('total:', len(dataset))
+    total = len(dataset)
+    meta = np.zeros( (total, 3), dtype=np.float32 )
+    for idx in range(total):
+        #image, label_verts, label_6dof = dataset[idx]
+        #img_raw, img_local, label_verts, label_Rt, tform = dataset[idx]
+        img, label_verts, label_6dof, label_points2d, _, _ = dataset[idx]
+        pose = label_6dof.numpy()[:3]
+        print(idx, pose)
+        meta[idx] = pose
+
+    np.save(save_path, meta)
+
diff --git a/reconstruction/jmlr/inference_simple.py b/reconstruction/jmlr/inference_simple.py
new file mode 100644
index 000000000..de4d90612
--- /dev/null
+++ b/reconstruction/jmlr/inference_simple.py
@@ -0,0 +1,322 @@
+
+import os
+import time
+import timm
+import glob
+import numpy as np
+import os.path as osp
+import cv2
+
+import torch
+import torch.distributed as dist
+from torch import nn
+from pathlib import Path
+from backbones import get_network
+from skimage import transform as sktrans
+from scipy.spatial.transform import Rotation
+
+def batch_euler2matrix(batch_euler):
+    n = batch_euler.shape[0]
+    assert batch_euler.shape[1] == 3
+    batch_matrix = np.zeros([n, 3, 3], dtype=np.float32)
+
+    for i in range(n):
+        pitch, yaw, roll = batch_euler[i]
+        R = Rotation.from_euler('yxz', [yaw, pitch, roll], degrees=False).as_matrix().T
+        batch_matrix[i] = R
+
+    return batch_matrix
+
+def euler2matrix(euler):
+    assert len(euler)==3
+    matrix = np.zeros([3, 3], dtype=np.float32)
+
+    pitch, yaw, roll = euler
+    R = Rotation.from_euler('yxz', [yaw, pitch, roll], degrees=False).as_matrix().T
+    matrix = R
+    return matrix
+
+def Rt_from_6dof(pred_6dof):
+    assert pred_6dof.ndim==1 or pred_6dof.ndim==2
+    if pred_6dof.ndim==1:
+        R_pred = euler2matrix(pred_6dof[:3])
+        t_pred = pred_6dof[-3:]
+        return R_pred, t_pred
+    else:
+        R_pred = batch_euler2matrix(pred_6dof[:,:3])
+        t_pred = pred_6dof[:,-3:].reshape(-1,1,3)
+        return R_pred, t_pred
+
+def solver_rigid(pts_3d , pts_2d , camera_matrix):
+    # pts_3d  Nx3
+    # pts_2d  Nx2
+    # camera_matrix 4x4
+    dist_coeffs = np.zeros((4,1))
+    pts_3d = pts_3d.copy()
+    pts_2d = pts_2d.copy()
+    #print(pts_3d.shape, pts_3d.dtype, pts_2d.shape, pts_2d.dtype)
+    success, rotation_vector, translation_vector = cv2.solvePnP(pts_3d, pts_2d, camera_matrix, dist_coeffs, flags=0)
+    assert success
+    R, _ = cv2.Rodrigues(rotation_vector)
+    R = R.T
+    R[:,1:3] *= -1
+    T = translation_vector.flatten()
+    T[1:] *= -1
+
+    return R,T
+
+
+class JMLRInference(nn.Module):
+    def __init__(self, cfg, local_rank=0):
+        super(JMLRInference, self).__init__()
+        backbone = get_network(cfg)
+        if cfg.ckpt is None:
+            ckpts = list(glob.glob(osp.join(cfg.output, "backbone*.pth")))
+            backbone_pth = sorted(ckpts)[-1]
+        else:
+            backbone_pth = cfg.ckpt
+        if local_rank==0:
+            print(backbone_pth)
+        backbone_ckpt = torch.load(backbone_pth, map_location=torch.device(local_rank))
+        if 'model' in backbone_ckpt:
+            backbone_ckpt = backbone_ckpt['model']
+        backbone.load_state_dict(backbone_ckpt)
+        backbone.eval()
+        backbone.requires_grad_(False)
+        self.backbone = backbone
+        self.num_verts = cfg.num_verts
+        self.input_size = cfg.input_size
+        self.flipindex = cfg.flipindex.copy()
+        self.data_root = Path(cfg.root_dir)
+        txt_path = self.data_root / 'resources/projection_matrix.txt'
+        self.M_proj = np.loadtxt(txt_path, dtype=np.float32)
+        M1 = np.array([
+            [400.0,       0, 0, 0],
+            [      0, 400.0, 0, 0],
+            [      0,       0, 1, 0],
+            [400.0, 400.0, 0, 1]
+        ])
+        self.M1 = M1
+        camera_matrix = self.M_proj @ M1
+        camera_matrix =  camera_matrix[:3,:3].T
+        camera_matrix[0,2] = 400
+        camera_matrix[1,2] = 400
+        self.camera_matrix = camera_matrix.copy()
+        self.use_eyes = False
+        if cfg.eyes is not None:
+            self.use_eyes = True
+            from eye_dataset import EyeDataset
+            eye_dataset = EyeDataset(cfg.eyes['root'], load_data=False)
+            self.iris_idx_481 = eye_dataset.iris_idx_481
+
+    
+    def project_shape_in_image(self, verts, R_t):
+        verts_homo = verts
+        if verts_homo.shape[1] == 3:
+            ones = np.ones([verts_homo.shape[0], 1])
+            verts_homo = np.concatenate([verts_homo, ones], axis=1)
+        verts_out = verts_homo @ R_t @ self.M_proj @ self.M1
+        w_ = verts_out[:, [3]]
+        verts_out = verts_out / w_
+        return verts_out
+
+    def set_raw_image_size(self, width, height):
+        w = width / 2.0
+        h = height / 2.0
+        M1 = np.array([
+            [w,       0, 0, 0],
+            [      0, h, 0, 0],
+            [      0,       0, 1, 0],
+            [w, h, 0, 1]
+        ])
+        camera_matrix = self.M_proj @ M1
+        camera_matrix =  camera_matrix[:3,:3].T
+        camera_matrix[0,2] = w
+        camera_matrix[1,2] = h
+        self.camera_matrix = camera_matrix
+        self.raw_width = width
+        self.raw_height = height
+
+
+    def forward(self, img_local, is_flip=False):
+        if is_flip:
+            img_local = img_local.flip([3])
+        pred = self.backbone(img_local)
+        pred1 = pred[:,:1220*3]
+        pred2 = pred[:,1220*3:1220*5]
+        meta = {'flip': is_flip}
+        if not self.use_eyes:
+            return pred1, pred2, meta
+        else:
+            eye_verts = pred[:,1220*5:1220*5+481*2*3]
+            eye_points = pred[:,1220*5+481*2*3:]
+            return pred1, pred2, meta, eye_verts, eye_points
+
+
+    def convert_verts(self, pred1, meta):
+        is_flip = meta['flip']
+        pred1 = pred1.cpu().numpy()
+        pred1 = pred1[:,:1220*3]
+        pred_verts = pred1.reshape(-1,1220,3) / 10.0
+        if is_flip:
+            pred_verts = pred_verts[:,self.flipindex,:]
+            pred_verts[:,:,0] *= -1.0
+        return pred_verts
+
+    def convert_2d(self, pred2, tforms, meta):
+        is_flip = meta['flip']
+        tforms = tforms.cpu().numpy()
+        pred2 = pred2.cpu().numpy()
+        B = pred2.shape[0]
+        points2d = (pred2.reshape(B,-1,2)+1.0) * self.input_size//2
+        if is_flip:
+            points2d = points2d[:,self.flipindex,:]
+            points2d[:,:,0] = self.input_size - 1 - points2d[:,:,0]
+        #B = points2d.shape[0]
+        points2de = np.ones( (points2d.shape[0], points2d.shape[1], 3), dtype=points2d.dtype)
+        points2de[:,:,:2] = points2d
+        verts2d = np.zeros((B,points2d.shape[1],2), dtype=np.float32)
+        for n in range(B):
+            tform = tforms[n]
+            tform_inv = cv2.invertAffineTransform(tform)
+            _points2d = np.dot(tform_inv, points2de[n].T).T
+            verts2d[n] = _points2d
+        #return verts2d, points2d
+        return verts2d
+
+
+    def convert_eyes(self, eye_verts3d, eye_verts2d, R_t, tforms):
+        meta = {'flip': False}
+        eye_verts3d = eye_verts3d.cpu().numpy().reshape(-1, 481*2, 3)[0]
+        eye_verts2d = self.convert_2d(eye_verts2d, tforms, meta)[0]
+        el_inv = eye_verts3d[:481,:]
+        er_inv = eye_verts3d[481:,:]
+        v_el = eye_verts2d[:481,:]
+        v_er = eye_verts2d[481:,:]
+        vector_norm = 0.035
+        # gaze vector of left eye in world space
+        gl_vector = el_inv[self.iris_idx_481].mean(axis=0) - el_inv[-1]
+        gl_vector = (gl_vector / np.linalg.norm(gl_vector)) * vector_norm
+        gl_point = el_inv[self.iris_idx_481].mean(axis=0) + gl_vector
+        # gaze vector of right eye in world space
+        gr_vector = er_inv[self.iris_idx_481].mean(axis=0) - er_inv[-1]
+        gr_vector = (gr_vector / np.linalg.norm(gr_vector)) * vector_norm
+        gr_point = er_inv[self.iris_idx_481].mean(axis=0) + gr_vector
+        g_el = self.project_shape_in_image(gl_point[None, :], R_t)
+        g_er = self.project_shape_in_image(gr_point[None, :], R_t)
+        g_el = g_el[:, :3].copy()
+        g_el[:, 1] = self.raw_height - g_el[:, 1]
+        g_er = g_er[:, :3].copy()
+        g_er[:, 1] = self.raw_height - g_er[:, 1]
+        pt1_l = v_el[self.iris_idx_481][:, [0, 1]].mean(axis=0).astype(np.int32)
+        pt2_l = g_el[0, [0, 1]].astype(np.int32)
+        pt1_r = v_er[self.iris_idx_481][:, [0, 1]].mean(axis=0).astype(np.int32)
+        pt2_r = g_er[0, [0, 1]].astype(np.int32)
+        return eye_verts3d, eye_verts2d, (pt1_l, pt2_l), (pt1_r, pt2_r)
+
+    def solve(self, verts3d, verts2d):
+        print(verts3d.shape, verts2d.shape)
+        B = verts3d.shape[0]
+        R = np.zeros([B, 3, 3], dtype=np.float32)
+        t = np.zeros([B, 1, 3], dtype=np.float32)
+        for n in range(B):
+            _R, _t = solver_rigid(verts3d[n], verts2d[n], self.camera_matrix)
+            R[n] = _R
+            t[n,0] = _t
+        return R, t
+
+    def solve_one(self, verts3d, verts2d):
+        R, t = solver_rigid(verts3d, verts2d, self.camera_matrix)
+        return R, t
+
+
+def get(net, img, keypoints):
+    dst_pts = np.array([
+        [38.2946, 51.6963],
+        [73.5318, 51.5014],
+        [56.0252, 71.7366],
+        [41.5493, 92.3655],
+        [70.7299, 92.2041] ], dtype=np.float32 )
+    input_size = 256
+    local_rank = 0
+
+    new_size = 144
+    dst_pts[:,0] += ((new_size-112)//2)
+    dst_pts[:,1] += 8
+    dst_pts[:,:] *= (input_size/float(new_size))
+    tf = sktrans.SimilarityTransform()
+    tf.estimate(keypoints, dst_pts)
+    tform = tf.params[0:2,:]
+    img = cv2.warpAffine(img, tform, (input_size,)*2)
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = np.transpose(img, (2, 0, 1))
+    img = torch.from_numpy(img).unsqueeze(0).float()
+    img.div_(255).sub_(0.5).div_(0.5)
+    img_local = img.to(local_rank)
+    with torch.no_grad():
+        if not net.use_eyes:
+            pred1, pred2, meta = net(img_local, is_flip=False)
+        else:
+            pred1, pred2, meta, eye_verts, eye_points = net(img_local, is_flip=False)
+    pred_verts = net.convert_verts(pred1, meta)
+    tform = torch.from_numpy(tform.reshape(1,2,3))
+    pred_verts2d = net.convert_2d(pred2, tform, meta)
+    verts = pred_verts[0]
+    verts2d = pred_verts2d[0]
+    R, t = net.solve_one(verts, verts2d)
+    if not net.use_eyes:
+        return verts, verts2d
+    else:
+        R_t = np.zeros( (4,4), dtype=np.float32)
+        R_t[:3,:3] = R
+        R_t[3,:3] = t
+        R_t[3,3] = 1.0
+        eye_verts, eye_verts2d, gaze_l, gaze_r = net.convert_eyes(eye_verts, eye_points, R_t, tform)
+        return verts, verts2d, eye_verts, eye_verts2d, gaze_l, gaze_r
+
+
+if __name__ == "__main__":
+    import argparse
+    from utils.utils_config import get_config
+    from insightface.app import FaceAnalysis
+    parser = argparse.ArgumentParser(description='JMLR inference')
+    #parser.add_argument('config', type=str, help='config file')
+    config_file = 'configs/s1.py'
+    #config_file = 'configs/s2.py'
+    args = parser.parse_args()
+    cfg = get_config(config_file)
+    cfg2 = None
+    local_rank = 0
+    #img = cv2.imread('sample.jpg')
+    net = JMLRInference(cfg, local_rank)
+    net = net.to(local_rank)
+    net.eval()
+    app = FaceAnalysis(allowed_modules='detection')
+    app.prepare(ctx_id=0, det_size=(640,640), det_thresh=0.5)
+    index = -1
+    for img_path in glob.glob('/data/insightface/wcpa/image/222714/01_LeftToRight_Neutral/*.jpg'):
+        index+=1
+        img = cv2.imread(img_path)
+        if index==0:
+            net.set_raw_image_size(img.shape[1], img.shape[0])
+        draw = img.copy()
+        faces = app.get(img)
+        for face in faces:
+            if not net.use_eyes:
+                verts3d, verts2d = get(net, img, face.kps)
+            else:
+                verts3d, verts2d, eye_verts3d, eye_verts2d, gaze_l, gaze_r = get(net, img, face.kps)
+            #print(verts3d.shape, verts2d.shape, R.shape, t.shape, eye_verts3d.shape, eye_verts2d.shape)
+            for i in range(verts2d.shape[0]):
+                pt = verts2d[i].astype(np.int32)
+                cv2.circle(draw, pt, 2, (255,0,0), 2)
+            #eye_verts2d = eye_verts2d[:481,:]
+            if net.use_eyes:
+                for i in range(eye_verts2d.shape[0]):
+                    pt = eye_verts2d[i].astype(np.int32)
+                    cv2.circle(draw, pt, 2, (0,255,0), 2)
+                for gaze in [gaze_l, gaze_r]:
+                    pt1, pt2 = gaze
+                    cv2.arrowedLine(draw, pt1, pt2, [0, 0, 255], 10)
+        cv2.imwrite('./outputs/%04d.jpg'%index, draw)
+
diff --git a/reconstruction/jmlr/losses.py b/reconstruction/jmlr/losses.py
new file mode 100644
index 000000000..8659f0c91
--- /dev/null
+++ b/reconstruction/jmlr/losses.py
@@ -0,0 +1,111 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+import kornia
+import numpy as np
+
+#def loss_l1(a, b):
+    #_loss = torch.abs(a - b)
+    #_loss = torch.mean(_loss, dim=1)
+    ##if epoch>4 and cfg.loss_hard:
+    ##    _loss, _ = torch.topk(_loss, k=int(cfg.batch_size*0.3))
+    #_loss = torch.mean(_loss)
+    #return _loss
+
+
+
+def loss_pip(outputs_map, outputs_local_x, outputs_local_y, labels_map, labels_local_x, labels_local_y):
+
+    tmp_batch, tmp_channel, tmp_height, tmp_width = outputs_map.size()
+    labels_map = labels_map.view(tmp_batch*tmp_channel, -1)
+    labels_max_ids = torch.argmax(labels_map, 1)
+    labels_max_ids = labels_max_ids.view(-1, 1)
+
+    #print('TTT:', outputs_local_x.shape, tmp_batch, tmp_channel)
+
+    outputs_local_x = outputs_local_x.reshape(tmp_batch*tmp_channel, -1)
+    outputs_local_x_select = torch.gather(outputs_local_x, 1, labels_max_ids)
+    outputs_local_y = outputs_local_y.reshape(tmp_batch*tmp_channel, -1)
+    outputs_local_y_select = torch.gather(outputs_local_y, 1, labels_max_ids)
+
+    labels_local_x = labels_local_x.view(tmp_batch*tmp_channel, -1)
+    labels_local_x_select = torch.gather(labels_local_x, 1, labels_max_ids)
+    labels_local_y = labels_local_y.view(tmp_batch*tmp_channel, -1)
+    labels_local_y_select = torch.gather(labels_local_y, 1, labels_max_ids)
+
+    labels_map = labels_map.view(tmp_batch, tmp_channel, tmp_height, tmp_width)
+    loss_map = F.mse_loss(outputs_map, labels_map)
+    loss_x = F.l1_loss(outputs_local_x_select, labels_local_x_select)
+    loss_y = F.l1_loss(outputs_local_y_select, labels_local_y_select)
+    return loss_map, loss_x, loss_y
+
+def eye_like(x: torch.Tensor, n: int) -> torch.Tensor:
+    return torch.eye(n, n, dtype=x.dtype, device=x.device).unsqueeze(0).repeat(x.shape[0], 1, 1)
+
+class ProjectLoss(nn.Module):
+    
+    def __init__(self,M_proj):
+        super(ProjectLoss, self).__init__()
+        img_w = 800
+        img_h = 800
+        M1 = np.array([
+            [img_w/2,       0, 0, 0],
+            [      0, img_h/2, 0, 0],
+            [      0,       0, 1, 0],
+            [img_w/2, img_h/2, 0, 1]
+        ])
+        M = M_proj @ M1
+        M = M.astype(np.float32)
+        self.register_buffer('M', torch.from_numpy(M))
+
+        camera_matrix =  M[:3,:3].T.copy()
+        camera_matrix[0,2] = 400
+        camera_matrix[1,2] = 400
+        camera_matrix[2,2] = 1
+        intrinsics = np.array([camera_matrix]).astype(np.float64)
+        self.register_buffer('intrinsics', torch.from_numpy(intrinsics))
+
+
+        self.eps = 1e-5
+        #self.projector = Reprojector(img_w,img_h,M_proj)
+        #self.solver = PnPSolver(self.projector.M.numpy())
+        #self.loss_fn = torch.nn.MSELoss(reduce=False, size_average=False)
+        #self.loss_fn = torch.nn.MSELoss()
+        self.loss_fn = torch.nn.L1Loss()
+        
+    
+    def forward(self,verts3d, points2d, affine):
+        # pred_2d_lmks  Batch*N*2
+        # verts Batch*N*3
+        ones = torch.ones([points2d.shape[0] , points2d.shape[1], 1],device=points2d.device)
+        verts_homo = torch.cat((points2d, ones), 2)
+        K  = eye_like(affine,3)
+        K[:,:2,:3] = affine
+        inv_k = torch.linalg.inv(K)
+        inv_k@verts_homo.permute(0,2,1)
+        points2d_inv = inv_k@verts_homo.permute(0,2,1)
+        points2d_inv = points2d_inv.permute(0,2,1)[:,:,:2]
+
+        intrinsics = self.intrinsics.repeat([verts3d.shape[0],1,1 ])
+        #print(verts3d.double().shape) 
+        #print(points2d.double().shape)
+        #print(intrinsics.shape)
+        RT_ = kornia.geometry.solve_pnp_dlt(verts3d.double(), points2d_inv.double(), intrinsics,svd_eps=self.eps)
+        RT_ = RT_.float()
+        RT = eye_like(verts3d,4)
+#         RT[:,1:3,:] *=-1
+        RT[:,:3,:] = RT_
+        RT = RT.permute(0,2,1)
+        RT[:,:,:2] *= -1
+
+        ones = torch.ones([verts3d.shape[0] , verts3d.shape[1], 1],device=verts3d.device)
+        verts_homo = torch.cat((verts3d, ones), 2)
+        M = self.M.repeat([verts3d.shape[0],1,1 ])
+        verts = verts_homo @ RT @ M
+        w_ = verts[:,:, [3]]
+        verts = verts / w_
+        reproject_points2d = verts[:,:, :2]
+        loss = self.loss_fn(reproject_points2d , points2d_inv)
+
+        return loss
+
diff --git a/reconstruction/jmlr/lr_scheduler.py b/reconstruction/jmlr/lr_scheduler.py
new file mode 100644
index 000000000..3e5461da3
--- /dev/null
+++ b/reconstruction/jmlr/lr_scheduler.py
@@ -0,0 +1,93 @@
+import torch
+from torch.optim.lr_scheduler import _LRScheduler
+
+
+class PolyScheduler(_LRScheduler):
+    def __init__(self,
+                 optimizer,
+                 base_lr,
+                 max_steps,
+                 warmup_steps,
+                 last_epoch=-1):
+        self.base_lr = base_lr
+        self.warmup_lr_init = 0.0001
+        self.max_steps: int = max_steps
+        self.warmup_steps: int = warmup_steps
+        self.power = 2
+        super(PolyScheduler, self).__init__(optimizer, last_epoch, False)
+
+    def get_warmup_lr(self):
+        alpha = float(self.last_epoch) / float(self.warmup_steps)
+        #_lr = max(self.base_lr * alpha, self.warmup_lr_init)
+        _lr = self.base_lr * alpha
+        return [_lr for _ in self.optimizer.param_groups]
+
+    def get_lr(self):
+        if self.last_epoch == -1:
+            return [self.warmup_lr_init for _ in self.optimizer.param_groups]
+        if self.last_epoch < self.warmup_steps:
+            return self.get_warmup_lr()
+        else:
+            alpha = pow(
+                1 - float(self.last_epoch - self.warmup_steps) /
+                float(self.max_steps - self.warmup_steps),
+                self.power,
+            )
+            return [self.base_lr * alpha for _ in self.optimizer.param_groups]
+
+class StepScheduler(_LRScheduler):
+    def __init__(self,
+                 optimizer,
+                 base_lr,
+                 lr_steps,
+                 warmup_steps,
+                 last_epoch=-1):
+        self.base_lr = base_lr
+        self.warmup_lr_init = 0.0001
+        self.lr_steps = lr_steps
+        self.warmup_steps: int = warmup_steps
+        super(StepScheduler, self).__init__(optimizer, last_epoch, False)
+
+    def get_warmup_lr(self):
+        alpha = float(self.last_epoch) / float(self.warmup_steps)
+        #_lr = max(self.base_lr * alpha, self.warmup_lr_init)
+        _lr = self.base_lr * alpha
+        return [_lr for _ in self.optimizer.param_groups]
+
+    def get_lr(self):
+        if self.last_epoch == -1:
+            return [self.warmup_lr_init for _ in self.optimizer.param_groups]
+        if self.last_epoch < self.warmup_steps:
+            return self.get_warmup_lr()
+        else:
+            alpha = 0.1 ** len([m for m in self.lr_steps if m <= self.last_epoch])
+            return [self.base_lr * alpha for _ in self.optimizer.param_groups]
+
+
+
+def get_scheduler(opt, cfg):
+    if cfg.lr_func is not None:
+        scheduler = torch.optim.lr_scheduler.LambdaLR(
+            optimizer=opt, lr_lambda=cfg.lr_func)
+    else:
+        #total_batch_size = cfg.batch_size * cfg.world_size
+        #warmup_steps = cfg.num_images // total_batch_size * cfg.warmup_epochs
+        #total_steps = cfg.num_images // total_batch_size * cfg.num_epochs
+
+        if cfg.lr_steps is None:
+            scheduler = PolyScheduler(
+                optimizer=opt,
+                base_lr=cfg.lr,
+                max_steps=cfg.total_steps,
+                warmup_steps=cfg.warmup_steps,
+            )
+        else:
+            scheduler = StepScheduler(
+                optimizer=opt,
+                base_lr=cfg.lr,
+                lr_steps=cfg.lr_steps,
+                warmup_steps=cfg.warmup_steps,
+            )
+
+    return scheduler
+
diff --git a/reconstruction/jmlr/rec_builder.py b/reconstruction/jmlr/rec_builder.py
new file mode 100644
index 000000000..1f8f81602
--- /dev/null
+++ b/reconstruction/jmlr/rec_builder.py
@@ -0,0 +1,135 @@
+import pickle
+import numpy as np
+import os
+import os.path as osp
+import glob
+import argparse
+import cv2
+import time
+import datetime
+import pickle
+import sklearn
+import mxnet as mx
+from utils.utils_config import get_config
+from dataset import FaceDataset, Rt26dof
+
+class RecBuilder():
+    def __init__(self, path, image_size=(112, 112), is_train=True):
+        self.path = path
+        self.image_size = image_size
+        self.widx = 0
+        self.wlabel = 0
+        self.max_label = -1
+        #assert not osp.exists(path), '%s exists' % path
+        if is_train:
+            rec_file = osp.join(path, 'train.rec')
+            idx_file = osp.join(path, 'train.idx')
+        else:
+            rec_file = osp.join(path, 'val.rec')
+            idx_file = osp.join(path, 'val.idx')
+        #assert not osp.exists(rec_file), '%s exists' % rec_file
+        if not osp.exists(path):
+            os.makedirs(path)
+        self.writer = mx.recordio.MXIndexedRecordIO(idx_file,
+                                                    rec_file,
+                                                    'w')
+        self.meta = []
+
+    def add(self, imgs):
+        #!!! img should be BGR!!!!
+        #assert label >= 0
+        #assert label > self.last_label
+        assert len(imgs) > 0
+        label = self.wlabel
+        for img in imgs:
+            idx = self.widx
+            image_meta = {'image_index': idx, 'image_classes': [label]}
+            header = mx.recordio.IRHeader(0, label, idx, 0)
+            if isinstance(img, np.ndarray):
+                s = mx.recordio.pack_img(header,img,quality=95,img_fmt='.jpg')
+            else:
+                s = mx.recordio.pack(header, img)
+            self.writer.write_idx(idx, s)
+            self.meta.append(image_meta)
+            self.widx += 1
+        self.max_label = label
+        self.wlabel += 1
+        return label
+
+
+    def add_image(self, img, label):
+        #!!! img should be BGR!!!!
+        #assert label >= 0
+        #assert label > self.last_label
+        idx = self.widx
+        header = mx.recordio.IRHeader(0, label, idx, 0)
+        if isinstance(img, np.ndarray):
+            s = mx.recordio.pack_img(header,img,quality=100,img_fmt='.jpg')
+        else:
+            s = mx.recordio.pack(header, img)
+        self.writer.write_idx(idx, s)
+        self.widx += 1
+
+    def close(self):
+        print('stat:', self.widx, self.wlabel)
+
+if __name__ == "__main__":
+    #cfg = get_config('configs/s1.py')
+    cfg = get_config('configs/s2.py')
+    cfg.task = 0
+    cfg.input_size = 512
+    for is_train in [True, False]:
+        dataset = FaceDataset(cfg, is_train=is_train, local_rank=0)
+        dataset.transform = None
+        writer = RecBuilder(cfg.cache_dir, is_train=is_train)
+        #writer = RecBuilder("temp", is_train=is_train)
+        print('total:', len(dataset))
+        #meta = np.zeros( (len(dataset), 3), dtype=np.float32 )
+        meta = []
+        subset_name = 'train' if is_train else 'val'
+        meta_path = osp.join(cfg.cache_dir, '%s.meta'%subset_name)
+        eye_missing = 0
+        for idx in range(len(dataset)):
+            #img_local, img_global, label_verts, label_Rt, tform = dataset[idx]
+            #img_local, label_verts, label_Rt, tform = dataset[idx]
+            data = dataset[idx]
+            img_local = data['img_local']
+            label_verts = data['verts']
+            label_Rt = data['rt']
+            tform = data['tform']
+            label_verts = label_verts.numpy()
+            label_Rt = label_Rt.numpy()
+            tform = tform.numpy()
+            label_6dof = Rt26dof(label_Rt, True)
+            pose = label_6dof[:3]
+            #print(image.shape, label_verts.shape, label_6dof.shape)
+            #print(image.__class__, label_verts.__class__)
+            img_local = img_local[:,:,::-1]
+            #img_global = img_global[:,:,::-1]
+            #image = np.concatenate( (img_local, img_global), axis=1 )
+            image = img_local
+            label = list(label_verts.flatten()) + list(label_Rt.flatten()) + list(tform.flatten())
+            expect_len = 1220*3+16+6
+            if 'eye_world_left' in data:
+                if idx==0:
+                    print('find eye')
+                eyel = data['eye_world_left'].numpy()
+                eyer = data['eye_world_right'].numpy()
+                label += list(eyel.flatten())
+                label += list(eyer.flatten())
+                expect_len += 481*6
+            else:
+                eye_missing += 1
+                continue
+            meta.append(pose)
+            assert len(label)==expect_len
+            writer.add_image(image, label)
+            if idx%100==0:
+                print('processing:', idx, image.shape, len(label))
+            if idx<10:
+                cv2.imwrite("temp/%d.jpg"%idx, image)
+        writer.close()
+        meta = np.array(meta, dtype=np.float32)
+        np.save(meta_path, meta)
+        print('Eye missing:', eye_missing, is_train)
+
diff --git a/reconstruction/jmlr/train.py b/reconstruction/jmlr/train.py
new file mode 100644
index 000000000..ee43fbc36
--- /dev/null
+++ b/reconstruction/jmlr/train.py
@@ -0,0 +1,303 @@
+import argparse
+import logging
+import os
+import time
+import timm
+import glob
+import numpy as np
+import os.path as osp
+
+import torch
+import torch.distributed as dist
+from torch import nn
+import torch.nn.functional as F
+import torch.utils.data.distributed
+from torch.nn.utils import clip_grad_norm_
+from dataset import FaceDataset, DataLoaderX, MXFaceDataset, get_tris
+
+from backbones import get_network
+from utils.utils_amp import MaxClipGradScaler
+from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
+from utils.utils_logging import AverageMeter, init_logging
+from utils.utils_config import get_config
+from lr_scheduler import get_scheduler
+from timm.optim.optim_factory import create_optimizer
+
+
+
+
+
+def main(args):
+    cfg = get_config(args.config)
+    if not cfg.tf32:
+        torch.backends.cuda.matmul.allow_tf32 = False
+        torch.backends.cudnn.allow_tf32 = False
+    else:
+        torch.backends.cuda.matmul.allow_tf32 = True
+        torch.backends.cudnn.allow_tf32 = True
+
+    try:
+        world_size = int(os.environ['WORLD_SIZE'])
+        rank = int(os.environ['RANK'])
+        dist.init_process_group('nccl')
+    except KeyError:
+        world_size = 1
+        rank = 0
+        dist.init_process_group(backend='nccl', init_method="tcp://127.0.0.1:12584", rank=rank, world_size=world_size)
+
+
+    local_rank = args.local_rank
+    torch.cuda.set_device(local_rank)
+
+    if not os.path.exists(cfg.output) and rank is 0:
+        os.makedirs(cfg.output)
+    else:
+        time.sleep(2)
+
+    log_root = logging.getLogger()
+    init_logging(log_root, rank, cfg.output)
+    if rank==0:
+        logging.info(args)
+        logging.info(cfg)
+        print(cfg.flipindex.shape, cfg.flipindex[400:410])
+    train_set = MXFaceDataset(cfg=cfg, is_train=True, local_rank=local_rank)
+    cfg.num_images = len(train_set)
+    cfg.world_size = world_size
+    total_batch_size = cfg.batch_size * cfg.world_size
+    epoch_steps = cfg.num_images // total_batch_size
+    cfg.warmup_steps = epoch_steps * cfg.warmup_epochs
+    if cfg.max_warmup_steps>0:
+        cfg.warmup_steps = min(cfg.max_warmup_steps, cfg.warmup_steps)
+    cfg.total_steps = epoch_steps * cfg.num_epochs
+    if cfg.lr_epochs is not None:
+        cfg.lr_steps = [m*epoch_steps for m in cfg.lr_epochs]
+    else:
+        cfg.lr_steps = None
+    train_sampler = torch.utils.data.distributed.DistributedSampler(
+        train_set, shuffle=True)
+
+    train_loader = torch.utils.data.DataLoader(
+        dataset=train_set, batch_size=cfg.batch_size,
+        sampler=train_sampler, num_workers=4, pin_memory=False, drop_last=True)
+
+    
+    net = get_network(cfg).to(local_rank)
+    
+
+    if cfg.resume:
+        try:
+            
+            ckpts = list(glob.glob(osp.join(cfg.resume_path, "backbone*.pth")))
+            backbone_pth = sorted(ckpts)[-1]
+            backbone_ckpt = torch.load(backbone_pth, map_location=torch.device(local_rank))
+            net.load_state_dict(backbone_ckpt['model'])
+            if rank==0:
+                logging.info("backbone resume successfully! %s"%backbone_pth)
+        except (FileNotFoundError, KeyError, IndexError, RuntimeError):
+            logging.info("resume fail!!")
+            raise RuntimeError
+
+
+    net = torch.nn.parallel.DistributedDataParallel(
+        module=net, broadcast_buffers=False, device_ids=[local_rank])
+    net.train()
+
+
+ 
+
+    if cfg.opt=='sgd':
+        opt = torch.optim.SGD(
+            params=[
+                {"params": net.parameters()}, 
+                ],
+            lr=cfg.lr, momentum=0.9, weight_decay=cfg.weight_decay)
+    elif cfg.opt=='adam':
+        opt = torch.optim.Adam(
+            params=[
+                {"params": net.parameters()}, 
+                ],
+            lr=cfg.lr)
+    elif cfg.opt=='adamw':
+        opt = torch.optim.AdamW(
+            params=[
+                {"params": net.parameters()}, 
+                ],
+            lr=cfg.lr, weight_decay=cfg.weight_decay)
+
+
+    scheduler = get_scheduler(opt, cfg)
+    if cfg.resume:
+        if rank==0:
+            logging.info(opt)
+
+
+
+    if cfg.resume:
+        for g in opt_pfc.param_groups:
+            for key in ['lr', 'initial_lr']:
+                g[key] = cfg.lr
+
+
+
+    start_epoch = 0
+    total_step = cfg.total_steps
+    if rank==0: 
+        logging.info(opt)
+        logging.info("Total Step is: %d" % total_step)
+
+
+    loss = {
+            'Loss': AverageMeter(),
+
+           }
+
+    global_step = 0
+    grad_amp = None
+    if cfg.fp16>0:
+        if cfg.fp16==1:
+            grad_amp = torch.cuda.amp.grad_scaler.GradScaler(growth_interval=100)
+        elif cfg.fp16==2:
+            grad_amp = MaxClipGradScaler(64, 1024, growth_interval=200)
+        elif cfg.fp16==3:
+            grad_amp = MaxClipGradScaler(4, 8, growth_interval=200)
+        else:
+            assert 'fp16 mode not set'
+
+    callback_checkpoint = CallBackModelCheckpoint(rank, cfg)
+
+    callback_checkpoint(global_step, net, opt)
+
+    callback_logging = CallBackLogging(50, rank, total_step, cfg.batch_size, world_size, None)
+
+    l1loss = nn.L1Loss()
+
+
+    tris = get_tris(cfg)
+    tri_index = torch.tensor(tris, dtype=torch.long).to(local_rank)
+    use_eyes = cfg.eyes is not None
+
+    for epoch in range(start_epoch, cfg.num_epochs):
+        train_sampler.set_epoch(epoch)
+        for step, value in enumerate(train_loader):
+            global_step += 1
+            img = value['img_local'].to(local_rank)
+            dloss = {}
+            assert cfg.task==0
+            label_verts = value['verts'].to(local_rank)
+            label_points2d = value['points2d'].to(local_rank)
+            #need_eyes = 'eye_verts' in value
+            preds = net(img)
+
+            
+            if use_eyes:
+                pred_verts, pred_points2d, pred_eye_verts, pred_eye_points = preds.split([1220*3, 1220*2, 481*2*3, 481*2*2], dim=1)
+                pred_eye_verts = pred_eye_verts.view(cfg.batch_size, 481*2, 3)
+                pred_eye_points = pred_eye_points.view(cfg.batch_size, 481*2, 2)
+            else:
+                pred_verts, pred_points2d = preds.split([1220*3, 1220*2], dim=1)
+            pred_verts = pred_verts.view(cfg.batch_size, 1220, 3)
+            pred_points2d = pred_points2d.view(cfg.batch_size, 1220, 2)
+            if not cfg.use_rtloss:
+                loss1 = F.l1_loss(pred_verts, label_verts)
+            else:
+                label_Rt = value['rt'].to(local_rank)
+                _ones = torch.ones([pred_verts.shape[0], 1220, 1], device=pred_verts.device)
+                pred_verts = torch.cat([pred_verts/10, _ones], dim=2)
+                pred_verts = torch.bmm(pred_verts,label_Rt) * 10.0
+                label_verts = torch.cat([label_verts/10, _ones], dim=2)
+                label_verts = torch.bmm(label_verts,label_Rt) * 10.0
+                loss1 = F.l1_loss(pred_verts, label_verts)
+
+            loss2 = F.l1_loss(pred_points2d, label_points2d)
+            loss3d = loss1 * cfg.lossw_verts3d
+            loss2d = loss2 * cfg.lossw_verts2d
+            dloss['Loss'] = loss3d + loss2d
+            dloss['Loss3D'] = loss3d
+            dloss['Loss2D'] = loss2d
+            if use_eyes:
+                label_eye_verts = value['eye_verts'].to(local_rank)
+                label_eye_points = value['eye_points'].to(local_rank)
+                loss3 = F.l1_loss(pred_eye_verts, label_eye_verts)
+                loss4 = F.l1_loss(pred_eye_points, label_eye_points)
+                loss3 = loss3 * cfg.lossw_eyes3d
+                loss4 = loss4 * cfg.lossw_eyes2d
+                dloss['Loss'] += loss3
+                dloss['Loss'] += loss4
+                dloss['LossEye3d'] = loss3
+                dloss['LossEye2d'] = loss4
+
+            if cfg.loss_bone3d:
+                bone_losses = []
+                for i in range(3):
+                    pred_verts_x = pred_verts[:,tri_index[:,i%3],:]
+                    pred_verts_y = pred_verts[:,tri_index[:,(i+1)%3],:]
+                    label_verts_x = label_verts[:,tri_index[:,i%3],:]
+                    label_verts_y = label_verts[:,tri_index[:,(i+1)%3],:]
+                    dist_pred = torch.norm(pred_verts_x - pred_verts_y, p=2, dim=-1, keepdim=False)
+                    dist_label = torch.norm(label_verts_x - label_verts_y, p=2, dim=-1, keepdim=False)
+                    bone_losses.append(F.l1_loss(dist_pred, dist_label) * cfg.lossw_bone3d)
+                _loss = sum(bone_losses)
+                dloss['Loss'] += _loss
+                dloss['LossBone3d'] = _loss
+                        
+
+            if cfg.loss_bone2d:
+                bone_losses = []
+                for i in range(3):
+                    pred_points2d_x = pred_points2d[:,tri_index[:,i%3],:]
+                    pred_points2d_y = pred_points2d[:,tri_index[:,(i+1)%3],:]
+                    label_points2d_x = label_points2d[:,tri_index[:,i%3],:]
+                    label_points2d_y = label_points2d[:,tri_index[:,(i+1)%3],:]
+                    dist_pred = torch.norm(pred_points2d_x - pred_points2d_y, p=2, dim=-1, keepdim=False)
+                    dist_label = torch.norm(label_points2d_x - label_points2d_y, p=2, dim=-1, keepdim=False)
+                    bone_losses.append(F.l1_loss(dist_pred, dist_label) * cfg.lossw_bone2d)
+                _loss = sum(bone_losses)
+                dloss['Loss'] += _loss
+                dloss['LossBone2d'] = _loss
+                        
+            iter_loss = dloss['Loss']
+
+            if cfg.fp16>0:
+                grad_amp.scale(iter_loss).backward()
+                grad_amp.unscale_(opt)
+                if cfg.fp16<2:
+                    torch.nn.utils.clip_grad_norm_(net.parameters(), 5)
+                grad_amp.step(opt)
+                grad_amp.update()
+            else:
+                iter_loss.backward()
+                opt.step()
+
+
+            opt.zero_grad()
+
+            if cfg.lr_func is None:
+                scheduler.step()
+
+            with torch.no_grad():
+                loss['Loss'].update(iter_loss.item(), 1)
+                for k in dloss:
+                    if k=='Loss':
+                        continue
+                    v = dloss[k].item()
+                    if k not in loss:
+                        loss[k] = AverageMeter()
+                    loss[k].update(v, 1)
+
+                callback_logging(global_step, loss, epoch, cfg.fp16, grad_amp, opt)
+
+        if cfg.lr_func is not None:
+            scheduler.step()
+
+    callback_checkpoint(9999, net, opt)
+    dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    torch.backends.cudnn.benchmark = True
+    parser = argparse.ArgumentParser(description='JMLR Training')
+    parser.add_argument('config', type=str, help='config file')
+    parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
+    args_ = parser.parse_args()
+    main(args_)
+
diff --git a/reconstruction/jmlr/utils/__init__.py b/reconstruction/jmlr/utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/reconstruction/jmlr/utils/plot.py b/reconstruction/jmlr/utils/plot.py
new file mode 100644
index 000000000..ccc588e5c
--- /dev/null
+++ b/reconstruction/jmlr/utils/plot.py
@@ -0,0 +1,72 @@
+# coding: utf-8
+
+import os
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap
+from prettytable import PrettyTable
+from sklearn.metrics import roc_curve, auc
+
+image_path = "/data/anxiang/IJB_release/IJBC"
+files = [
+        "./ms1mv3_arcface_r100/ms1mv3_arcface_r100/ijbc.npy"
+]
+
+
+def read_template_pair_list(path):
+    pairs = pd.read_csv(path, sep=' ', header=None).values
+    t1 = pairs[:, 0].astype(np.int)
+    t2 = pairs[:, 1].astype(np.int)
+    label = pairs[:, 2].astype(np.int)
+    return t1, t2, label
+
+
+p1, p2, label = read_template_pair_list(
+    os.path.join('%s/meta' % image_path,
+                 '%s_template_pair_label.txt' % 'ijbc'))
+
+methods = []
+scores = []
+for file in files:
+    methods.append(file.split('/')[-2])
+    scores.append(np.load(file))
+
+methods = np.array(methods)
+scores = dict(zip(methods, scores))
+colours = dict(
+    zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
+x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
+tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
+fig = plt.figure()
+for method in methods:
+    fpr, tpr, _ = roc_curve(label, scores[method])
+    roc_auc = auc(fpr, tpr)
+    fpr = np.flipud(fpr)
+    tpr = np.flipud(tpr)  # select largest tpr at same fpr
+    plt.plot(fpr,
+             tpr,
+             color=colours[method],
+             lw=1,
+             label=('[%s (AUC = %0.4f %%)]' %
+                    (method.split('-')[-1], roc_auc * 100)))
+    tpr_fpr_row = []
+    tpr_fpr_row.append("%s-%s" % (method, "IJBC"))
+    for fpr_iter in np.arange(len(x_labels)):
+        _, min_index = min(
+            list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
+        tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
+    tpr_fpr_table.add_row(tpr_fpr_row)
+plt.xlim([10 ** -6, 0.1])
+plt.ylim([0.3, 1.0])
+plt.grid(linestyle='--', linewidth=1)
+plt.xticks(x_labels)
+plt.yticks(np.linspace(0.3, 1.0, 8, endpoint=True))
+plt.xscale('log')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('ROC on IJB')
+plt.legend(loc="lower right")
+print(tpr_fpr_table)
diff --git a/reconstruction/jmlr/utils/utils_amp.py b/reconstruction/jmlr/utils/utils_amp.py
new file mode 100644
index 000000000..1a2286f15
--- /dev/null
+++ b/reconstruction/jmlr/utils/utils_amp.py
@@ -0,0 +1,82 @@
+from typing import Dict, List
+
+import torch
+#from torch._six import container_abcs
+import collections.abc as container_abcs
+from torch.cuda.amp import GradScaler
+
+
+class _MultiDeviceReplicator(object):
+    """
+    Lazily serves copies of a tensor to requested devices.  Copies are cached per-device.
+    """
+
+    def __init__(self, master_tensor: torch.Tensor) -> None:
+        assert master_tensor.is_cuda
+        self.master = master_tensor
+        self._per_device_tensors: Dict[torch.device, torch.Tensor] = {}
+
+    def get(self, device) -> torch.Tensor:
+        retval = self._per_device_tensors.get(device, None)
+        if retval is None:
+            retval = self.master.to(device=device, non_blocking=True, copy=True)
+            self._per_device_tensors[device] = retval
+        return retval
+
+
+class MaxClipGradScaler(GradScaler):
+    def __init__(self, init_scale, max_scale: float, growth_interval=100):
+        GradScaler.__init__(self, init_scale=init_scale, growth_interval=growth_interval)
+        self.max_scale = max_scale
+
+    def scale_clip(self):
+        if self.get_scale() == self.max_scale:
+            self.set_growth_factor(1)
+        elif self.get_scale() < self.max_scale:
+            self.set_growth_factor(2)
+        elif self.get_scale() > self.max_scale:
+            self._scale.fill_(self.max_scale)
+            self.set_growth_factor(1)
+
+    def scale(self, outputs):
+        """
+        Multiplies ('scales') a tensor or list of tensors by the scale factor.
+
+        Returns scaled outputs.  If this instance of :class:`GradScaler` is not enabled, outputs are returned
+        unmodified.
+
+        Arguments:
+            outputs (Tensor or iterable of Tensors):  Outputs to scale.
+        """
+        if not self._enabled:
+            return outputs
+        self.scale_clip()
+        # Short-circuit for the common case.
+        if isinstance(outputs, torch.Tensor):
+            assert outputs.is_cuda
+            if self._scale is None:
+                self._lazy_init_scale_growth_tracker(outputs.device)
+            assert self._scale is not None
+            return outputs * self._scale.to(device=outputs.device, non_blocking=True)
+
+        # Invoke the more complex machinery only if we're treating multiple outputs.
+        stash: List[_MultiDeviceReplicator] = []  # holds a reference that can be overwritten by apply_scale
+
+        def apply_scale(val):
+            if isinstance(val, torch.Tensor):
+                assert val.is_cuda
+                if len(stash) == 0:
+                    if self._scale is None:
+                        self._lazy_init_scale_growth_tracker(val.device)
+                    assert self._scale is not None
+                    stash.append(_MultiDeviceReplicator(self._scale))
+                return val * stash[0].get(val.device)
+            elif isinstance(val, container_abcs.Iterable):
+                iterable = map(apply_scale, val)
+                if isinstance(val, list) or isinstance(val, tuple):
+                    return type(val)(iterable)
+                else:
+                    return iterable
+            else:
+                raise ValueError("outputs must be a Tensor or an iterable of Tensors")
+        return apply_scale(outputs)
diff --git a/reconstruction/jmlr/utils/utils_callbacks.py b/reconstruction/jmlr/utils/utils_callbacks.py
new file mode 100644
index 000000000..ab740ec08
--- /dev/null
+++ b/reconstruction/jmlr/utils/utils_callbacks.py
@@ -0,0 +1,129 @@
+import logging
+import os
+import time
+from typing import List
+
+import torch
+import psutil
+
+#from eval import verification
+#from partial_fc import PartialFC
+#from torch2onnx import convert_onnx
+from utils.utils_logging import AverageMeter
+
+
+class CallBackVerification(object):
+    def __init__(self, frequent, rank, val_targets, rec_prefix, image_size=(112, 112)):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.highest_acc: float = 0.0
+        self.highest_acc_list: List[float] = [0.0] * len(val_targets)
+        self.ver_list: List[object] = []
+        self.ver_name_list: List[str] = []
+        if self.rank is 0:
+            self.init_dataset(val_targets=val_targets, data_dir=rec_prefix, image_size=image_size)
+
+    def ver_test(self, backbone: torch.nn.Module, global_step: int):
+        results = []
+        for i in range(len(self.ver_list)):
+            acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(
+                self.ver_list[i], backbone, 10, 10)
+            logging.info('[%s][%d]XNorm: %f' % (self.ver_name_list[i], global_step, xnorm))
+            logging.info('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (self.ver_name_list[i], global_step, acc2, std2))
+            if acc2 > self.highest_acc_list[i]:
+                self.highest_acc_list[i] = acc2
+            logging.info(
+                '[%s][%d]Accuracy-Highest: %1.5f' % (self.ver_name_list[i], global_step, self.highest_acc_list[i]))
+            results.append(acc2)
+
+    def init_dataset(self, val_targets, data_dir, image_size):
+        for name in val_targets:
+            path = os.path.join(data_dir, name + ".bin")
+            if os.path.exists(path):
+                data_set = verification.load_bin(path, image_size)
+                self.ver_list.append(data_set)
+                self.ver_name_list.append(name)
+
+    def __call__(self, num_update, backbone: torch.nn.Module):
+        if self.rank is 0 and num_update > 0 and num_update % self.frequent == 0:
+            backbone.eval()
+            self.ver_test(backbone, num_update)
+            backbone.train()
+
+
+class CallBackLogging(object):
+    def __init__(self, frequent, rank, total_step, batch_size, world_size, writer=None):
+        self.frequent: int = frequent
+        self.rank: int = rank
+        self.time_start = time.time()
+        self.total_step: int = total_step
+        self.batch_size: int = batch_size
+        self.world_size: int = world_size
+        self.writer = writer
+
+        self.init = False
+        self.tic = 0
+
+    def __call__(self, global_step, loss, epoch, fp16, grad_scaler, opt):
+        if self.rank is 0 and global_step > 0 and global_step % self.frequent == 0:
+            if self.init:
+                try:
+                    speed: float = self.frequent * self.batch_size / (time.time() - self.tic)
+                    speed_total = speed * self.world_size
+                except ZeroDivisionError:
+                    speed_total = float('inf')
+
+                time_now = (time.time() - self.time_start) / 3600.0
+                time_total = time_now / ((global_step + 1) / self.total_step)
+                time_for_end = time_total - time_now
+                lr = opt.param_groups[0]['lr']
+                if self.writer is not None:
+                    self.writer.add_scalar('time_for_end', time_for_end, global_step)
+                    #self.writer.add_scalar('loss', loss.avg, global_step)
+                mem = psutil.virtual_memory()
+                mem_used = mem.used / (1024 ** 3)
+                loss_str = ""
+                for k,v in loss.items():
+                    if len(loss_str)!=0:
+                        loss_str += "   "
+                    loss_str += "%s:%.4f"%(k, v.avg)
+                if fp16:
+                    msg = "Speed %.2f samples/sec   %s   Epoch: %d   Global Step: %d   LR: %.8f   " \
+                            "Fp16 Grad Scale: %2.f   Required: %.1f hours   MemUsed: %.3f" % (
+                              speed_total, loss_str, epoch, global_step, lr, grad_scaler.get_scale(), time_for_end, mem_used
+                          )
+                else:
+                    msg = "Speed %.2f samples/sec   %s   Epoch: %d   Global Step: %d   LR: %.8f   Required: %.1f hours   MemUsed: %.3f" % (
+                        speed_total, loss_str, epoch, global_step, lr, time_for_end, mem_used
+                    )
+                logging.info(msg)
+                for k,v in loss.items():
+                    v.reset()
+                self.tic = time.time()
+            else:
+                self.init = True
+                self.tic = time.time()
+
+
+class CallBackModelCheckpoint(object):
+    def __init__(self, rank, cfg):
+        self.rank = rank
+        self.output = cfg.output
+        #self.save_pfc = cfg.save_pfc
+        #self.save_onnx = cfg.save_onnx
+        self.save_opt = cfg.save_opt
+
+    def __call__(self, epoch, backbone, opt_backbone):
+        if self.rank == 0:
+            path_module = os.path.join(self.output, "backbone_ep%04d.pth"%epoch)
+            if self.save_opt:
+                data = {
+                        'model': backbone.module.state_dict(),
+                        'optimizer': opt_backbone.state_dict(),
+                        }
+            else:
+                data = backbone.module.state_dict()
+            torch.save(data, path_module)
+            logging.info("Pytorch Model Saved in '{}'".format(path_module))
+
+
diff --git a/reconstruction/jmlr/utils/utils_config.py b/reconstruction/jmlr/utils/utils_config.py
new file mode 100644
index 000000000..74408b36b
--- /dev/null
+++ b/reconstruction/jmlr/utils/utils_config.py
@@ -0,0 +1,27 @@
+import importlib
+import os
+import os.path as osp
+import numpy as np
+
+def get_config(config_file):
+    assert config_file.startswith('configs/'), 'config file setting must start with configs/'
+    temp_config_name = osp.basename(config_file)
+    temp_module_name = osp.splitext(temp_config_name)[0]
+    #print('A:', config_file, temp_config_name, temp_module_name)
+    config1 = importlib.import_module("configs.base")
+    importlib.reload(config1)
+    cfg = config1.config
+    #print('B1:', cfg)
+    config2 = importlib.import_module("configs.%s"%temp_module_name)
+    importlib.reload(config2)
+    #reload(config2)
+    job_cfg = config2.config
+    #print('B2:', job_cfg)
+    cfg.update(job_cfg)
+    cfg.job_name = temp_module_name
+    #print('B:', cfg)
+    if cfg.output is None:
+        cfg.output = osp.join('work_dirs', temp_module_name)
+    #print('C:', cfg.output)
+    cfg.flipindex = np.load(cfg.flipindex_file)
+    return cfg
diff --git a/reconstruction/jmlr/utils/utils_logging.py b/reconstruction/jmlr/utils/utils_logging.py
new file mode 100644
index 000000000..7d58012ea
--- /dev/null
+++ b/reconstruction/jmlr/utils/utils_logging.py
@@ -0,0 +1,40 @@
+import logging
+import os
+import sys
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value
+    """
+
+    def __init__(self):
+        self.val = None
+        self.avg = None
+        self.sum = None
+        self.count = None
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def init_logging(log_root, rank, models_root):
+    if rank is 0:
+        log_root.setLevel(logging.INFO)
+        formatter = logging.Formatter("Training: %(asctime)s-%(message)s")
+        handler_file = logging.FileHandler(os.path.join(models_root, "training.log"))
+        handler_stream = logging.StreamHandler(sys.stdout)
+        handler_file.setFormatter(formatter)
+        handler_stream.setFormatter(formatter)
+        log_root.addHandler(handler_file)
+        log_root.addHandler(handler_stream)
+        log_root.info('rank_id: %d' % rank)
diff --git a/reconstruction/jmlr/utils/utils_os.py b/reconstruction/jmlr/utils/utils_os.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/reconstruction/jmlr/validate_dist.py b/reconstruction/jmlr/validate_dist.py
new file mode 100644
index 000000000..397401f6f
--- /dev/null
+++ b/reconstruction/jmlr/validate_dist.py
@@ -0,0 +1,186 @@
+from dataset import FaceDataset, DataLoaderX, MXFaceDataset
+import argparse
+import logging
+import os
+import time
+import timm
+import glob
+import numpy as np
+import os.path as osp
+from utils.utils_config import get_config
+from scipy.spatial.transform import Rotation
+
+import torch
+import torch.distributed as dist
+from torch import nn
+import torch.nn.functional as F
+import torch.utils.data.distributed
+from backbones import get_network
+from inference_simple import JMLRInference, Rt_from_6dof
+from dataset import Rt26dof
+
+
+def l2_distance(a, b):
+    dist = np.sqrt(np.sum(np.square(a-b), axis=1))
+    distance_list = np.sqrt(((a - b) ** 2).sum(axis=2)).mean(axis=1)
+    return distance_list
+
+def main(args):
+    cfg = get_config(args.config)
+
+    torch.backends.cuda.matmul.allow_tf32 = False
+    torch.backends.cudnn.allow_tf32 = False
+    world_size = int(os.environ['WORLD_SIZE'])
+    rank = int(os.environ['RANK'])
+    #dist_url = "tcp://{}:{}".format(os.environ["MASTER_ADDR"], os.environ["MASTER_PORT"])
+    dist.init_process_group('nccl')
+
+    local_rank = args.local_rank
+    torch.cuda.set_device(local_rank)
+
+    task1 = cfg.task
+    cfg.aug_modes = []
+    cfg.task = 0
+    batch_size = cfg.batch_size
+    dataset = MXFaceDataset(cfg=cfg, is_train=False, local_rank=local_rank)
+    if local_rank==0:
+        print('total:', len(dataset))
+        print('total batch:', len(dataset)//(batch_size*world_size))
+    cfg.task = task1
+    net = JMLRInference(cfg, local_rank)
+    net = net.to(local_rank)
+    net.eval()
+    #net = torch.nn.parallel.DistributedDataParallel(
+    #    module=net, broadcast_buffers=False, device_ids=[local_rank])
+    sampler = torch.utils.data.distributed.DistributedSampler(dataset, shuffle=False)
+    loader = torch.utils.data.DataLoader(
+            dataset,
+            batch_size=batch_size,
+            sampler=sampler,
+            shuffle=False,
+            pin_memory=True,
+            num_workers=3,
+            drop_last=False,
+            )
+    num_epochs = 1
+    all_pred_verts = torch.zeros((len(dataset),1220,3), requires_grad=False).to(local_rank)
+    all_pred_R = torch.zeros((len(dataset),3,3), requires_grad=False).to(local_rank)
+    all_pred_t = torch.zeros((len(dataset),1,3), requires_grad=False).to(local_rank)
+    all_pred_verts2d = torch.zeros((len(dataset),1220,2), requires_grad=False).to(local_rank)
+
+    all_label_verts = torch.zeros((len(dataset),1220,3), requires_grad=False).to(local_rank)
+    all_label_R = torch.zeros((len(dataset),3,3), requires_grad=False).to(local_rank)
+    all_label_t = torch.zeros((len(dataset),1,3), requires_grad=False).to(local_rank)
+    all_label_verts2d = torch.zeros((len(dataset),1220,2), requires_grad=False).to(local_rank)
+    all_weight = 0.0
+    FLIPS = [False, True] if cfg.enable_flip else [False]
+    #FLIPS = [False]
+    if local_rank==0:
+        print('FLIPS:', FLIPS)
+    for epoch in range(num_epochs):
+        weight = 1.0
+
+        if epoch>0:
+            dataset.set_test_aug()
+            weight = 0.6
+        all_weight += weight
+        #all_distance = torch.zeros((len(dataset),), requires_grad=False).to(local_rank)
+        diff_R = []
+        diff_t = []
+        sampler.set_epoch(epoch)
+        for idx, sample in enumerate(loader):
+            img_local = sample['img_local']
+            label_verts = sample['verts']
+            tform = sample['tform']
+            label_6dof = sample['6dof']
+            data_idx = sample['idx']
+            label_verts2d = sample['verts2d']
+            img_local = img_local.to(local_rank)
+            pred_verts, pred_verts2d, pred_points2d = [], [], []
+            for is_flip in FLIPS:
+                with torch.no_grad():
+                    #pred_verts, R_pred, t_pred = infer.forward(img_local, img_raw, tform)
+                    #pred1, pred2 = net(img_local.to(local_rank), img_raw.to(local_rank))
+                    pred1, pred2, meta = net(img_local, is_flip=is_flip)
+                _pred_verts = net.convert_verts(pred1, meta)
+                pred_verts.append(_pred_verts)
+                _pred_verts2d, _pred_points2d = net.convert_2d(pred2, tform, meta)
+                pred_verts2d.append(_pred_verts2d)
+                pred_points2d.append(_pred_points2d)
+            pred_verts = sum(pred_verts) / len(pred_verts)
+            pred_verts2d = sum(pred_verts2d) / len(pred_verts2d)
+            pred_points2d = sum(pred_points2d) / len(pred_points2d)
+            R_pred, t_pred = net.solve(pred_verts, pred_verts2d)
+            label_6dof = label_6dof.cpu().numpy()
+            label_6dof = label_6dof * cfg.label_6dof_std.reshape(1, 6) + cfg.label_6dof_mean.reshape(1,6)
+
+            R_label, t_label = Rt_from_6dof(label_6dof)
+            diff_R.append(np.mean(np.abs(R_pred - R_label)))
+            diff_t.append(np.mean(np.abs(t_pred - t_label)))
+            #distance = torch.tensor(distance, dtype=torch.float32, requires_grad=False).to(local_rank)
+            data_idx = data_idx.view(-1)
+            #all_distance[data_idx] = distance
+            label_verts = label_verts.view(-1,1220,3) / 10.0
+            if epoch==0:
+                all_label_verts[data_idx,:,:] = label_verts.to(local_rank)
+                all_label_R[data_idx,:,:] = torch.tensor(R_label).to(local_rank)
+                all_label_t[data_idx,:,:] = torch.tensor(t_label).to(local_rank)
+                all_label_verts2d[data_idx,:,:] = label_verts2d.to(local_rank)
+            all_pred_verts[data_idx,:,:] += torch.tensor(pred_verts).to(local_rank) * weight
+            #all_pred_R[data_idx,:,:] += torch.tensor(R_pred).to(local_rank) * weight
+            #all_pred_t[data_idx,:,:] += torch.tensor(t_pred).to(local_rank) * weight
+            all_pred_verts2d[data_idx,:,:] += torch.tensor(pred_verts2d).to(local_rank) * weight
+            if idx%20==0 and local_rank==0:
+                print('processing-epoch-idx:', epoch, idx)
+                #print('distance:', distance.shape, distance.cpu().numpy().mean())
+                print('diff_R:', np.mean(diff_R))
+                print('diff_t:', np.mean(diff_t))
+
+    dist.all_reduce(all_label_verts, op=dist.ReduceOp.SUM)
+    dist.all_reduce(all_label_verts2d, op=dist.ReduceOp.SUM)
+    dist.all_reduce(all_label_R, op=dist.ReduceOp.SUM)
+    dist.all_reduce(all_label_t, op=dist.ReduceOp.SUM)
+
+    dist.all_reduce(all_pred_verts, op=dist.ReduceOp.SUM)
+    dist.all_reduce(all_pred_verts2d, op=dist.ReduceOp.SUM)
+    #dist.all_reduce(all_pred_R, op=dist.ReduceOp.SUM)
+    #dist.all_reduce(all_pred_t, op=dist.ReduceOp.SUM)
+    #dist.all_reduce(all_distance, op=dist.ReduceOp.SUM)
+    if local_rank==0:
+        label_verts = all_label_verts.cpu().numpy()
+        label_verts2d = all_label_verts2d.cpu().numpy()
+        R_label = all_label_R.cpu().numpy()
+        t_label = all_label_t.cpu().numpy()
+
+        pred_verts = all_pred_verts.cpu().numpy() / all_weight
+        #R_pred = all_pred_R.cpu().numpy() / all_weight
+        #t_pred = all_pred_t.cpu().numpy() / all_weight
+        pred_verts2d = all_pred_verts2d.cpu().numpy() / all_weight
+        R_pred, t_pred = net.solve(pred_verts, pred_verts2d)
+        #R_pred, t_pred = net.solve(pred_verts, label_verts2d)
+        #R_pred, t_pred = net.solve(label_verts, pred_verts2d)
+
+
+        X1 = label_verts @ R_label + t_label
+        X2 = pred_verts @ R_pred + t_pred
+        X3 = label_verts @ R_pred + t_pred
+        X4 = pred_verts @ R_label + t_label
+        distance = l2_distance(X1, X2) + l2_distance(X1, X3) + 10.0*l2_distance(X1,X4)
+        distance *= 1000.0
+
+        print('top20 distance:', np.mean(distance[:20]))
+
+
+        score = np.mean(distance)
+        print('epoch distance:', epoch, score)
+        with open(os.path.join(cfg.output, 'val.txt'), 'w') as f:
+            f.write("%f\n"%score)
+
+if __name__ == "__main__":
+    #torch.backends.cudnn.benchmark = True
+    parser = argparse.ArgumentParser(description='JMLR validation')
+    parser.add_argument('config', type=str, help='config file')
+    parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
+    args_ = parser.parse_args()
+    main(args_)
+
diff --git a/reconstruction/ostec/.gitignore b/reconstruction/ostec/.gitignore
new file mode 100644
index 000000000..15a12ee9a
--- /dev/null
+++ b/reconstruction/ostec/.gitignore
@@ -0,0 +1,8 @@
+models/*
+!models/topology
+.idea/*
+*.so
+/.ipynb_checkpoints/*
+results/*
+.stylegan2-cache/*
+samples/*
diff --git a/reconstruction/ostec/.gitmodules b/reconstruction/ostec/.gitmodules
new file mode 100644
index 000000000..672473ea0
--- /dev/null
+++ b/reconstruction/ostec/.gitmodules
@@ -0,0 +1,3 @@
+[submodule "external/deep3dfacerecon"]
+	path = external/deep3dfacerecon
+	url = https://github.com/barisgecer/Deep3DFaceRecon_pytorch
diff --git a/reconstruction/ostec/README.md b/reconstruction/ostec/README.md
new file mode 100644
index 000000000..02e912265
--- /dev/null
+++ b/reconstruction/ostec/README.md
@@ -0,0 +1,210 @@
+# [OSTeC: One-Shot Texture Complition](https://openaccess.thecvf.com/content/CVPR2021/html/Gecer_OSTeC_One-Shot_Texture_Completion_CVPR_2021_paper.html)
+#### [CVPR 2021]
+[![arXiv Prepring](https://img.shields.io/badge/arXiv-Preprint-lightgrey?logo=arxiv)](https://arxiv.org/abs/2012.15370)
+[![License: CC BY-NC-SA 4.0](https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
+
+
+ [Baris Gecer](http://barisgecer.github.io) <sup> 1,2</sup>, [Jiankang Deng](https://jiankangdeng.github.io/) <sup> 1,2</sup>, & [Stefanos Zafeiriou](https://wp.doc.ic.ac.uk/szafeiri/) <sup> 1,2</sup>
+ <br/>
+ <sup>1 </sup>Imperial College London
+ <br/>
+ <sup>2 </sup> Huawei CBG
+ <br/>
+
+<p align="center">
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/im18.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/norah.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/robin.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/teaser2.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/teaser3.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/teaser4.gif" style="background-color:white; display: inline;" />
+<img width="12%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/gifs/teaset5.gif" style="background-color:white; display: inline;" />
+</p>
+<p align="center"><img width="100%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/teaser.jpg" style="background-color:white;" /></p>
+
+## Abstract
+
+The last few years have witnessed the great success of non-linear generative models in synthesizing high-quality photorealistic face images. Many recent 3D facial texture reconstruction and pose manipulation from a single image approaches still rely on large and clean face datasets to train image-to-image Generative Adversarial Networks (GANs). Yet the collection of such a large scale high-resolution 3D texture dataset is still very costly and difficult to maintain age/ethnicity balance. Moreover, regression-based approaches suffer from generalization to the in-the-wild conditions and are unable to fine-tune to a target-image. In this work, we propose an unsupervised approach for one-shot 3D facial texture completion that does not require large-scale texture datasets, but rather harnesses the knowledge stored in 2D face generators. The proposed approach rotates an input image in 3D and fill-in the unseen regions by reconstructing the rotated image in a 2D face generator, based on the visible parts. Finally, we stitch the most visible textures at different angles in the UV image-plane. Further, we frontalize the target image by projecting the completed texture into the generator. The qualitative and quantitative experiments demonstrate that the completed UV textures and frontalized images are of high quality, resembles the original identity, can be used to train a texture GAN model for 3DMM fitting and improve pose-invariant face recognition.
+
+## Overview
+
+<p align="center"><img width="100%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/overview.jpg" style="background-color:white;" /></p>
+Overview of the method. The proposed approach iteratively optimizes the texture UV-maps for different re-rendered images with their masks. At the end of each optimization, generated images are used to acquire partial UV images by dense landmarks. Finally, the completed UV images are fed to the next iteration for progressive texture building.
+<br/>
+
+
+## Requirements
+**This implementation is only tested under Ubuntu environment with Nvidia GPUs and CUDA 10.0 and CuDNN-7.0 installed.**
+
+## Installation
+### 1. Clone the repository and set up a conda environment as follows:
+```
+git clone https://github.com/barisgecer/OSTeC --recursive
+cd OSTeC
+conda env create -f environment.yml -n ostec
+source activate ostec
+```
+
+### 2. Installation of Deep3DFaceRecon_pytorch
+- **2.a.** Install Nvdiffrast library:
+```
+cd external/deep3dfacerecon/nvdiffrast    # ./OSTeC/external/deep3dfacerecon/nvdiffrast 
+pip install .
+```
+- **2.b.** Install Arcface Pytorch:
+```
+cd ..    # ./OSTeC/external/deep3dfacerecon/
+git clone https://github.com/deepinsight/insightface.git
+cp -r ./insightface/recognition/arcface_torch/ ./models/
+```
+
+- **2.c.** Prepare prerequisite models: Deep3DFaceRecon_pytorch method uses [Basel Face Model 2009 (BFM09)](https://faces.dmi.unibas.ch/bfm/main.php?nav=1-0&id=basel_face_model) to represent 3d faces. Get access to BFM09 using this [link](https://faces.dmi.unibas.ch/bfm/main.php?nav=1-2&id=downloads). After getting the access, download "01_MorphableModel.mat". In addition, we use an Expression Basis provided by [Guo et al.](https://github.com/Juyong/3DFace). Download the Expression Basis (Exp_Pca.bin) using this [link (google drive)](https://drive.google.com/file/d/1bw5Xf8C12pWmcMhNEu6PtsYVZkVucEN6/view?usp=sharing). Organize all files into the following structure:
+```
+OSTeC
+│
+└─── external
+     │
+     └─── deep3dfacerecon
+          │
+          └─── BFM
+              │
+              └─── 01_MorphableModel.mat
+              │
+              └─── Exp_Pca.bin
+              |
+              └─── ...
+```
+- **2.d.** Deep3DFaceRecon_pytorch provides a model trained on a combination of [CelebA](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html), 
+[LFW](http://vis-www.cs.umass.edu/lfw/), [300WLP](http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm),
+[IJB-A](https://www.nist.gov/programs-projects/face-challenges), [LS3D-W](https://www.adrianbulat.com/face-alignment), and [FFHQ](https://github.com/NVlabs/ffhq-dataset) datasets. Download the pre-trained model using this [link (google drive)](https://drive.google.com/drive/folders/1liaIxn9smpudjjqMaWWRpP0mXRW_qRPP?usp=sharing) and organize the directory into the following structure:
+```
+OSTeC
+│
+└─── external
+     │
+     └─── deep3dfacerecon
+          │
+          └─── checkpoints
+               │
+               └─── face_recon
+                   │
+                   └─── epoch_latest.pth
+
+```
+### 3. Download Face Recognition \& Landmark Detection \& VGG \& Style-Encoder models
+- Download the models here: https://drive.google.com/file/d/1TBoNt55vleRkMZaT9XKt6oNQmo8hkN-Q/view?usp=sharing
+
+- And place it under 'models' directory like the following:
+```
+OSTeC
+│
+└─── models
+     │
+     └─── resnet_18_20191231.h5
+     │
+     └─── vgg16_zhang_perceptual.pkl
+     │
+     └─── alignment
+     │         .
+     │         .
+     │
+     └─── fr_models
+               .
+               .
+
+```
+
+### 4. Download Topology info files
+- Download the topology files here: https://drive.google.com/file/d/1mvb2uDMPNGL1MlBgP6Op00gPdEMQUUWb/view?usp=sharing
+
+- And place it under 'models/topology' directory like the following:
+```
+OSTeC
+│
+└─── models
+     │         .
+     │         .
+     │         .
+     │
+     └─── topology
+          │
+          └─── trilist.pkl
+          │
+          └─── tcoords.pkl
+                    .
+                    .
+
+```
+
+
+
+### 5. Download Face Segmentation models
+- Download the Graphonomy model here: https://drive.google.com/file/d/1eUe18HoH05p0yFUd_sN6GXdTj82aW0m9/view?usp=sharing
+(If the link doesn't work for some reason check the original [Graphonomy](https://github.com/Gaoyiminggithub/Graphonomy) github page and download 'CIHP trained model')
+
+- And place it under 'models' directory like the following:
+```
+OSTeC
+│
+└─── models
+     │
+     └─── Graphonomy
+         │
+         └─── inference.pth
+```
+
+<!--- ### 4. Download StyleGANv2 model
+- Download the model from the original repo: https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/stylegan2-ffhq-config-f.pkl
+And place it under 'models' directory like the following:
+```
+OSTeC
+│
+└─── models
+     │
+     └─── stylegan2_networks_stylegan2-ffhq-config-f
+
+```
+-->
+
+## Usage
+- Run ```python run_ostec.py --source_dir [source_dir] --save_dir [save_dir] [-f] -i [iterations (default 200)] -m [soft|hard|auto]```
+- Modes (-m or --mode):
+   * soft: keep the original texture for visible parts (recommended when the input image is high resolution, near-frontal, and non-occluded.)
+   * hard: generate all
+   * auto: soft for frontal, hard for profile images
+
+## More Results
+
+<p align="center"><img width="100%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/comp2.jpg" /></p>
+<p align="center"><img width="100%" src="https://raw.githubusercontent.com/barisgecer/OSTeC/main/figures/comp1.jpg" /></p>
+<br/>
+
+## License
+- The source code shared here is protected under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) License which does **NOT** allow commercial  use. To view a copy of this license, see LICENSE
+- Copyright (c) 2020, Baris Gecer. All rights reserved.
+- This work is made available under the CC BY-NC-SA 4.0.
+
+
+## Acknowledgement
+- Our projection relies on NVIDIA's [StyleGANv2](https://github.com/NVlabs/stylegan2)
+- Thanks [@jiankangdeng](https://jiankangdeng.github.io/) for providing Face Recognition and Landmark Detection models
+- We use [MTCNN](https://github.com/ipazc/mtcnn) for face detection
+- We use [Graphonomy](https://github.com/Gaoyiminggithub/Graphonomy) for face segmentation (i.e. to exclude hairs, occlusion)
+- 3D face reconstruction has been originally solved by [GANFit](https://github.com/barisgecer/GANFit). However, since it is commercialized and will not be public, I had to re-implement the ports for [Deep3DFaceRecon_pytorch](https://github.com/sicxu/Deep3DFaceRecon_pytorch).
+- We initialize StyleGAN parameters by [Style-Encoder](https://github.com/rolux/stylegan2encoder/issues/2) (by [@rolux](https://github.com/rolux), [@pbaylies](https://github.com/pbaylies)).
+- Thanks [Zhang et al.](https://richzhang.github.io/PerceptualSimilarity/) for VGG16 model
+
+## Citation
+If you find this work is useful for your research, please cite our paper: 
+
+```
+@InProceedings{Gecer_2021_CVPR,
+    author    = {Gecer, Baris and Deng, Jiankang and Zafeiriou, Stefanos},
+    title     = {OSTeC: One-Shot Texture Completion},
+    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+    month     = {June},
+    year      = {2021},
+    pages     = {7628-7638}
+}
+```
+<br/>
diff --git a/reconstruction/ostec/core/arcface_handler.py b/reconstruction/ostec/core/arcface_handler.py
new file mode 100644
index 000000000..a223a5f46
--- /dev/null
+++ b/reconstruction/ostec/core/arcface_handler.py
@@ -0,0 +1,116 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import tensorflow as tf
+import numpy as np
+from external import arcface50
+from skimage import transform as trans
+
+
+def align_arcface(image, landmarks):
+    """
+    Aligns 'image' with its corresponding 'landmarks' to a predefined template
+    with similarity transformation. This is the tensorflow implementation of
+    the default alignment procedure of ArcFace
+    Args:
+        image: a 4D float32 numpy array with shape [batch_size, image_height,
+        image_width, 3].
+        landmarks: 68 iBug landmark points of 'image'. [batch_size, 68, 2]
+
+    Returns:
+        4-D float32 numpy array with shape [batch_size, 112, 112, 3]. Contains
+        aligned version of 'image'
+
+    """
+
+    image_size = (112, 112)
+    dst = np.array([
+        [30.2946, 51.6963],
+        [65.5318, 51.5014],
+        [48.0252, 71.7366],
+        [33.5493, 92.3655],
+        [62.7299, 92.2041]], dtype=np.float32)
+    if image_size[1] == 112:
+        dst[:, 0] += 8.0
+    # dst = dst[:, ::-1]
+    landmark5 = tf.stack([(landmarks[:, 36] + landmarks[:, 39]) / 2,
+                          (landmarks[:, 42] + landmarks[:, 45]) / 2,
+                          landmarks[:, 30],
+                          landmarks[:, 48],
+                          landmarks[:, 54]], 1)
+
+    def py_similarity_transform(src, dst):
+        tform = trans.SimilarityTransform()
+        Ms = np.zeros([0, 3, 3], dtype=np.float32)
+        for s in src:
+            tform.estimate(s, dst)
+            Ms = np.concatenate([Ms,[tform.params]],0)
+        return Ms
+
+    M = tf.py_func(py_similarity_transform, [landmark5, tf.constant(dst, 'float32')], tf.double, stateful=False)
+    M.set_shape([image.get_shape().as_list()[0],3, 3])
+
+    aligned = tf.contrib.image.transform(image, tf.cast(tf.contrib.image.matrices_to_flat_transforms(tf.map_fn(tf.linalg.inv,M)), 'float32'),interpolation='BILINEAR')
+
+    return aligned[:, 0:image_size[0], 0:image_size[1], :]
+
+def get_input_features(image, landmarks):
+    """Extract features from a face recongnition networks including
+        intermadiate activations. This function first align the image and
+        then call identity_features()
+    """
+    image_aligned = align_arcface(image, landmarks)
+
+    emb_norm, content, embedding, vars = identity_features(image_aligned,'id_features/')
+    return emb_norm, vars, image_aligned #, content, embedding], image_aligned
+
+def identity_features(input, name):
+    """Extract features from a face recongnition networks including
+        intermadiate activations.
+    """
+    with tf.variable_scope(name, reuse=True):
+        input, embedding, content = arcface50.KitModel('models/fr_models/arcface50.npy', input * 255)
+        emb_norm = tf.nn.l2_normalize(embedding, 1)
+    vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=name)
+    return emb_norm, content, embedding, vars
+
+class Arcface_Handler:
+    def __init__(self):
+        self.img_ph = tf.placeholder(dtype=tf.float32, name='img_ph', shape=[None, 256, 256, 3])
+        self.lms_ph = tf.placeholder(dtype=tf.float32, name='lms_ph', shape=[None, 68, 2])
+        aligned_img = align_arcface(self.img_ph, self.lms_ph)
+        aligned_img.set_shape([None, 112, 112, 3])
+        self.emb_norm, _, _, vars = identity_features(aligned_img, 'input_id_features')
+        var_init = tf.variables_initializer(vars)
+        config = tf.ConfigProto()
+        config.gpu_options.allow_growth = True
+        self.sess = tf.Session(config=config)
+        self.sess.run(var_init)
+
+    def get_identity_features(self, img, lms):
+        lms = (lms / img.shape[::-1]) * [256, 256]
+        img = img.resize([256, 256])
+        return self.sess.run(self.emb_norm, {self.img_ph: [img.pixels_with_channels_at_back()], self.lms_ph:[lms]})
+
+def identity_features_numpy(image, landmarks, return_aligned=False):
+    src_img = np.array([image])
+    src_lms = np.array([landmarks])
+    src_img = tf.constant(src_img,tf.float32)
+    src_lms = tf.constant(src_lms,tf.float32)
+    aligned_img = align_arcface(src_img,src_lms)
+    emb_norm, _, _, vars = identity_features(aligned_img, 'input_id_features')
+    var_init = tf.variables_initializer(vars)
+    with tf.Session() as sess:
+        sess.run(var_init)
+        features = sess.run(emb_norm)
+        if return_aligned:
+            aligned = sess.run(aligned_img)
+
+    tf.reset_default_graph()
+    if return_aligned:
+        return features, aligned
+    else:
+        return features
+
diff --git a/reconstruction/ostec/core/config.py b/reconstruction/ostec/core/config.py
new file mode 100644
index 000000000..66d9fd02a
--- /dev/null
+++ b/reconstruction/ostec/core/config.py
@@ -0,0 +1,96 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import argparse
+
+def split_to_batches(l, n):
+    for i in range(0, len(l), n):
+        yield l[i:i + n]
+
+def str2bool(v):
+    if isinstance(v, bool):
+       return v
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+parser = argparse.ArgumentParser(description='Find latent representation of reference images using perceptual loss')
+
+parser.add_argument('-m', '--mode',default='hard', choices=['soft', 'auto', 'hard'],
+                    help='Soft: keep original texture, Hard: generate all, auto: soft for frontal, hard for profile')
+parser.add_argument('-f', '--frontalize', action='store_true', help='Run frontalization at the end')
+parser.add_argument('-p', '--pickle', action='store_true', help='Save pickle with everything')
+parser.add_argument('-g', '--ganfit', action='store_true', help='Reconstruction from GANFit is a must. If not raised, it is automatic: GANFit rec. if pickle found, Deep3DRecon otherwise.')
+parser.add_argument('--iterations_frontalize', default=300, help='Number of optimization steps for each batch', type=int)
+
+parser.add_argument('--load_last', default='', help='Start with embeddings from directory')
+parser.add_argument('--dlatent_avg', default='', help='Use dlatent from file specified here for truncation instead of dlatent_avg from Gs')
+parser.add_argument('--model_url', default='gdrive:networks/stylegan2-ffhq-config-f.pkl', help='Fetch a StyleGAN model to train on from this URL') # default='gdrive:networks/stylegan2-ffhq-config-f.pkl'
+parser.add_argument('--model_res', default=1024, help='The dimension of images in the StyleGAN model', type=int)
+parser.add_argument('--batch_size', default=1, help='Batch size for generator and perceptual model', type=int)
+parser.add_argument('--optimizer', default='ggt', help='Optimization algorithm used for optimizing dlatents')
+
+# Perceptual model params
+parser.add_argument('--vgg_url', default='models/vgg16_zhang_perceptual.pkl', help='Fetch VGG model on from this URL') # default='https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2'
+parser.add_argument('--image_size', default=256, help='Size of images for perceptual model', type=int)
+parser.add_argument('--resnet_image_size', default=224, help='Size of images for the Resnet model', type=int)
+parser.add_argument('--lr', default=0.25, help='Learning rate for perceptual model', type=float)
+parser.add_argument('--decay_rate', default=0.9, help='Decay rate for learning rate', type=float)
+parser.add_argument('-i', '--iterations', default=200, help='Number of optimization steps for each batch', type=int)
+parser.add_argument('--decay_steps', default=4, help='Decay steps for learning rate decay (as a percent of iterations)', type=float)
+parser.add_argument('--early_stopping', default=True, help='Stop early once training stabilizes', type=str2bool, nargs='?', const=True)
+parser.add_argument('--early_stopping_threshold', default=0.5, help='Stop after this threshold has been reached', type=float)
+parser.add_argument('--early_stopping_patience', default=10, help='Number of iterations to wait below threshold', type=int)
+parser.add_argument('--load_effnet', default='data/finetuned_effnet.h5', help='Model to load for EfficientNet approximation of dlatents')
+parser.add_argument('--load_resnet', default='models/resnet_18_20191231.h5', help='Model to load for ResNet approximation of dlatents')
+parser.add_argument('--use_preprocess_input', default=True, help='Call process_input() first before using feed forward net', type=str2bool, nargs='?', const=True)
+parser.add_argument('--use_best_loss', default=True, help='Output the lowest loss value found as the solution', type=str2bool, nargs='?', const=True)
+parser.add_argument('--average_best_loss', default=0.25, help='Do a running weighted average with the previous best dlatents found', type=float)
+parser.add_argument('--sharpen_input', default=True, help='Sharpen the input images', type=str2bool, nargs='?', const=True)
+parser.add_argument('--landmark_model', default='./models/alignment/3D84/model.ckpt-277538', help='Landmark model path')
+
+# Loss function options
+parser.add_argument('--use_vgg_loss', default=0.4, help='Use VGG perceptual loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_vgg_layer', default=9, help='Pick which VGG layer to use.', type=int)
+parser.add_argument('--use_pixel_loss', default=1.5, help='Use logcosh image pixel loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_mssim_loss', default=200, help='Use MS-SIM perceptual loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_lpips_loss', default=100, help='Use LPIPS perceptual loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_l1_penalty', default=0.5, help='Use L1 penalty on latents; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_discriminator_loss', default=0.5, help='Use trained discriminator to evaluate realism.', type=float)
+parser.add_argument('--use_adaptive_loss', default=False, help='Use the adaptive robust loss function from Google Research for pixel and VGG feature loss.', type=str2bool, nargs='?', const=True)
+parser.add_argument('--use_landmark_loss', default=200, help='Use landmark loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_id_loss', default=10, help='Use landmark loss; 0 to disable, > 0 to scale.', type=float)
+parser.add_argument('--use_id_loss_frontalize', default=100, help='Use landmark loss; 0 to disable, > 0 to scale.', type=float)
+
+# Generator params
+parser.add_argument('--randomize_noise', default=False, help='Add noise to dlatents during optimization', type=str2bool, nargs='?', const=True)
+parser.add_argument('--tile_dlatents', default=False, help='Tile dlatents to use a single vector at each scale', type=str2bool, nargs='?', const=True)
+parser.add_argument('--clipping_threshold', default=2.0, help='Stochastic clipping of gradient values outside of this threshold', type=float)
+
+# Masking params
+parser.add_argument('--load_mask', default=False, help='Load segmentation masks', type=str2bool, nargs='?', const=True)
+parser.add_argument('--face_mask', default=True, help='Generate a mask for predicting only the face area', type=str2bool, nargs='?', const=True)
+parser.add_argument('--use_grabcut', default=True, help='Use grabcut algorithm on the face mask to better segment the foreground', type=str2bool, nargs='?', const=True)
+parser.add_argument('--scale_mask', default=1.4, help='Look over a wider section of foreground for grabcut', type=float)
+parser.add_argument('--composite_mask', default=False, help='Merge the unmasked area back into the generated image', type=str2bool, nargs='?', const=True)
+parser.add_argument('--composite_blur', default=8, help='Size of blur filter to smoothly composite the images', type=int)
+
+# Video params
+parser.add_argument('--video_dir', default='videos', help='Directory for storing training videos')
+parser.add_argument('--output_video', default=False, help='Generate videos of the optimization process', type=bool)
+parser.add_argument('--video_codec', default='MJPG', help='FOURCC-supported video codec name')
+parser.add_argument('--video_frame_rate', default=24, help='Video frames per second', type=int)
+parser.add_argument('--video_size', default=512, help='Video size in pixels', type=int)
+parser.add_argument('--video_skip', default=1, help='Only write every n frames (1 = write every frame)', type=int)
+
+
+def get_config():
+    args, other_args = parser.parse_known_args()
+    args.decay_steps *= 0.01 * args.iterations  # Calculate steps as a percent of total iterations
+
+    return args, other_args
\ No newline at end of file
diff --git a/reconstruction/ostec/core/generator_model.py b/reconstruction/ostec/core/generator_model.py
new file mode 100644
index 000000000..83be6248e
--- /dev/null
+++ b/reconstruction/ostec/core/generator_model.py
@@ -0,0 +1,142 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import math
+import tensorflow as tf
+import numpy as np
+import external.stylegan2.dnnlib.tflib as tflib
+from functools import partial
+
+
+def create_stub(name, batch_size):
+    return tf.constant(0, dtype='float32', shape=(batch_size, 0))
+
+
+def create_variable_for_generator(name, batch_size, tiled_dlatent, model_scale=18, tile_size = 1):
+    if tiled_dlatent:
+        low_dim_dlatent = tf.get_variable('learnable_dlatents',
+            shape=(batch_size, tile_size, 512),
+            dtype='float32',
+            initializer=tf.initializers.random_normal())
+        return tf.tile(low_dim_dlatent, [1, model_scale // tile_size, 1])
+    else:
+        return tf.get_variable('learnable_dlatents',
+            shape=(batch_size, model_scale, 512),
+            dtype='float32',
+            initializer=tf.initializers.random_normal())
+
+
+class Generator:
+    def __init__(self, model, batch_size, custom_input=None, clipping_threshold=2, tiled_dlatent=False, model_res=1024, randomize_noise=False):
+        self.batch_size = batch_size
+        self.tiled_dlatent=tiled_dlatent
+        self.model_scale = int(2*(math.log(model_res,2)-1)) # For example, 1024 -> 18
+
+        if tiled_dlatent:
+            self.initial_dlatents = np.zeros((self.batch_size, 512))
+            model.components.synthesis.run(np.zeros((self.batch_size, self.model_scale, 512)),
+                randomize_noise=randomize_noise, minibatch_size=self.batch_size,
+                custom_inputs=[partial(create_variable_for_generator, batch_size=batch_size, tiled_dlatent=True),
+                                                partial(create_stub, batch_size=batch_size)],
+                structure='fixed')
+        else:
+            self.initial_dlatents = np.zeros((self.batch_size, self.model_scale, 512))
+            if custom_input is not None:
+                model.components.synthesis.run(self.initial_dlatents,
+                    randomize_noise=randomize_noise, minibatch_size=self.batch_size,
+                    custom_inputs=[partial(custom_input.eval(), batch_size=batch_size), partial(create_stub, batch_size=batch_size)],
+                    structure='fixed')
+            else:
+                model.components.synthesis.run(self.initial_dlatents,
+                    randomize_noise=randomize_noise, minibatch_size=self.batch_size,
+                    custom_inputs=[partial(create_variable_for_generator, batch_size=batch_size, tiled_dlatent=False, model_scale=self.model_scale),
+                                                    partial(create_stub, batch_size=batch_size)],
+                    structure='fixed')
+
+        self.dlatent_avg_def = model.get_var('dlatent_avg')
+        self.reset_dlatent_avg()
+        self.sess = tf.get_default_session()
+        self.graph = tf.get_default_graph()
+
+        self.dlatent_variable = next(v for v in tf.global_variables() if 'learnable_dlatents' in v.name)
+        self._assign_dlatent_ph = tf.placeholder(tf.float32, name="assign_dlatent_ph")
+        self._assign_dlantent = tf.assign(self.dlatent_variable, self._assign_dlatent_ph)
+        self.set_dlatents(self.initial_dlatents)
+
+        def get_tensor(name):
+            try:
+                return self.graph.get_tensor_by_name(name)
+            except KeyError:
+                return None
+
+        self.generator_output = get_tensor('G_synthesis_1/_Run/concat:0')
+        if self.generator_output is None:
+            self.generator_output = get_tensor('G_synthesis_1/_Run/concat/concat:0')
+        if self.generator_output is None:
+            self.generator_output = get_tensor('G_synthesis_1/_Run/concat_1/concat:0')
+        # If we loaded only Gs and didn't load G or D, then scope "G_synthesis_1" won't exist in the graph.
+        if self.generator_output is None:
+            self.generator_output = get_tensor('G_synthesis/_Run/concat:0')
+        if self.generator_output is None:
+            self.generator_output = get_tensor('G_synthesis/_Run/concat/concat:0')
+        if self.generator_output is None:
+            self.generator_output = get_tensor('G_synthesis/_Run/concat_1/concat:0')
+        if self.generator_output is None:
+            for op in self.graph.get_operations():
+                print(op)
+            raise Exception("Couldn't find G_synthesis_1/_Run/concat tensor output")
+        self.generated_image = tflib.convert_images_to_uint8(self.generator_output, nchw_to_nhwc=True, uint8_cast=False)
+        self.generated_image_uint8 = tf.saturate_cast(self.generated_image, tf.uint8)
+
+        # Implement stochastic clipping similar to what is described in https://arxiv.org/abs/1702.04782
+        # (Slightly different in that the latent space is normal gaussian here and was uniform in [-1, 1] in that paper,
+        # so we clip any vector components outside of [-2, 2]. It seems fine, but I haven't done an ablation check.)
+        clipping_mask = tf.math.logical_or(self.dlatent_variable > clipping_threshold, self.dlatent_variable < -clipping_threshold)
+        clipped_values = tf.where(clipping_mask, tf.random_normal(shape=self.dlatent_variable.shape), self.dlatent_variable)
+        self.stochastic_clip_op = tf.assign(self.dlatent_variable, clipped_values)
+
+    def reset_dlatents(self):
+        self.set_dlatents(self.initial_dlatents)
+
+    def set_dlatents(self, dlatents):
+        if self.tiled_dlatent:
+            if (dlatents.shape != (self.batch_size, 512)) and (dlatents.shape[1] != 512):
+                dlatents = np.mean(dlatents, axis=1)
+            if (dlatents.shape != (self.batch_size, 512)):
+                dlatents = np.vstack([dlatents, np.zeros((self.batch_size-dlatents.shape[0], 512))])
+            assert (dlatents.shape == (self.batch_size, 512))
+        else:
+            if (dlatents.shape[1] > self.model_scale):
+                dlatents = dlatents[:,:self.model_scale,:]
+            if (isinstance(dlatents.shape[0], int)):
+                if (dlatents.shape != (self.batch_size, self.model_scale, 512)):
+                    dlatents = np.vstack([dlatents, np.zeros((self.batch_size-dlatents.shape[0], self.model_scale, 512))])
+                assert (dlatents.shape == (self.batch_size, self.model_scale, 512))
+                self.sess.run([self._assign_dlantent], {self._assign_dlatent_ph: dlatents})
+                return
+            else:
+                self._assign_dlantent = tf.assign(self.dlatent_variable, dlatents)
+                return
+        self.sess.run([self._assign_dlantent], {self._assign_dlatent_ph: dlatents})
+
+    def stochastic_clip_dlatents(self):
+        self.sess.run(self.stochastic_clip_op)
+
+    def get_dlatents(self):
+        return self.sess.run(self.dlatent_variable)
+
+    def get_dlatent_avg(self):
+        return self.dlatent_avg
+
+    def set_dlatent_avg(self, dlatent_avg):
+        self.dlatent_avg = dlatent_avg
+
+    def reset_dlatent_avg(self):
+        self.dlatent_avg = self.dlatent_avg_def
+
+    def generate_images(self, dlatents=None):
+        if dlatents is not None:
+            self.set_dlatents(dlatents)
+        return self.sess.run(self.generated_image_uint8)
\ No newline at end of file
diff --git a/reconstruction/ostec/core/landmark_handler.py b/reconstruction/ostec/core/landmark_handler.py
new file mode 100644
index 000000000..0da5464a5
--- /dev/null
+++ b/reconstruction/ostec/core/landmark_handler.py
@@ -0,0 +1,35 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import tensorflow as tf
+from external.landmark_detector import networks
+from external.landmark_detector.flags import FLAGS
+
+
+def tf_heatmap_to_lms(heatmap):
+    hs = tf.argmax(tf.reduce_max(heatmap, 2), 1)
+    ws = tf.argmax(tf.reduce_max(heatmap, 1), 1)
+    lms = tf.transpose(tf.to_float(tf.stack([hs, ws])), perm=[1, 2, 0])
+    return lms
+
+class Landmark_Handler():
+    def __init__(self, args, sess, generated_image):
+        self.sess = sess
+
+        self.model_path = args.landmark_model
+        n_landmarks = 84
+        FLAGS.n_landmarks = 84
+
+        net_model = networks.DNFaceMultiView('')
+        with tf.variable_scope('net'):
+            self.lms_heatmap_prediction, states = net_model._build_network(generated_image, datas=None, is_training=False,
+                                                                      n_channels=n_landmarks)
+            self.pts_predictions = tf_heatmap_to_lms(self.lms_heatmap_prediction)
+            variables = tf.all_variables()
+            variables_to_restore = [v for v in variables if v.name.split('/')[0] == 'net']
+            self.saver = tf.train.Saver(variables_to_restore)
+
+    def load_model(self):
+        self.saver.restore(self.sess, self.model_path)
diff --git a/reconstruction/ostec/core/operator.py b/reconstruction/ostec/core/operator.py
new file mode 100644
index 000000000..7e988f84a
--- /dev/null
+++ b/reconstruction/ostec/core/operator.py
@@ -0,0 +1,435 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+from utils.align2stylegan import align_im2stylegan, align_mesh2stylegan
+from core.projection_handler import Projection_Handler
+from skimage.morphology import remove_small_holes
+import time
+from utils.ganfit_camera import apply_camera_only3d, get_pose
+from utils.utils import *
+from core.arcface_handler import Arcface_Handler
+
+from utils import generate_heatmap
+
+class Face:
+    def __init__(self, tmesh,
+            tmesh_masked,
+            tmesh_rotated,
+            img_uv_src,
+            angle_uv_src,
+            angle_uv_list,
+            img_uv_list,
+            view_angle_src,
+            id_features,
+            exclude_mask,
+            is_profile,
+            mode):
+        self.tmesh = tmesh
+        self.tmesh_masked = tmesh_masked
+        self.tmesh_rotated = tmesh_rotated
+        self.img_uv_src = img_uv_src
+        self.angle_uv_src = angle_uv_src
+        self.angle_uv_list = angle_uv_list
+        self.img_uv_list = img_uv_list
+        self.view_angle_src = view_angle_src
+        self.id_features = id_features
+        self.is_profile = is_profile
+        self.exclude_mask = exclude_mask
+
+        if mode == 'auto':
+            if is_profile:
+                self.mode = 'hard'
+            else:
+                self.mode = 'soft'
+        else:
+            self.mode = mode
+
+    def rotation_dict(self):
+        if self.mode == 'soft':
+            return {
+                'bottom': [-30, 0, 0],
+                'bottom_left': [-15, -30, 0],
+                'bottom_right': [-15, 30, 0],
+                'left': [5, -40, 0],
+                'right': [5, 40, 0],
+            }
+        elif self.mode == 'hard':
+            return {
+                'front': [0, 0, 0],
+                'bottom': [-30, 0, 0],
+                'bottom_left': [-15, -30, 0],
+                'bottom_right': [-15, 30, 0],
+                'left': [5, -40, 0],
+                'right': [5, 40, 0],
+            }
+        else:
+            raise Exception('Unknown mode!')
+
+    def coef_dict(self):
+        if self.mode == 'soft':
+            return {
+                'bottom': 1.3,
+                'bottom_left': 1,
+                'bottom_right': 1,
+                'left': 1,
+                'right': 1,
+                'src': 2,
+            }
+        elif self.mode == 'hard':
+            return {
+                'front': 2,
+                'bottom': 1.3,
+                'bottom_left': 1,
+                'bottom_right': 1,
+                'left': 1,
+                'right': 1,
+                'src': 0.1,
+            }
+        else:
+            raise Exception('Unknown mode!')
+
+
+class Operator:
+    def __init__(self, args):
+        self.tcoords_full = mio.import_pickle('models/topology/tcoords_full.pkl')
+        self.tcoords = mio.import_pickle('models/topology/tcoords_alex.pkl')
+        self.mask = mio.import_pickle('models/topology/mask_full2crop.pkl') | True
+        self.tight_mask = mio.import_pickle('models/topology/mask_full2tightcrop.pkl')
+        self.template = mio.import_pickle('models/topology/all_all_all_crop_mean.pkl')
+        self.lms_ind = mio.import_pickle('models/topology/all_all_all_lands_ids.pkl')
+        self.img_shape = [1024, 1024] # 2048
+
+        self.uv_shape = [1024, 1536]
+        uv_mesh = self.tcoords.copy().points[:, ::-1]
+        uv_mesh[:, 0] = 1 - uv_mesh[:, 0]
+        uv_mesh *= self.uv_shape
+        self.uv_mesh = np.concatenate([uv_mesh, uv_mesh[:, 0:1] * 0], 1)
+        self.uv_trilist = mio.import_pickle('models/topology/trilist_full.pkl') #self.template.trilist
+
+        self.args = args
+        self.mode = args.mode # 'soft', 'hard', 'auto'
+
+        self.arcface_handler = Arcface_Handler()
+        self.projector = Projection_Handler(args)
+
+    def render_uv_image(self, generated, tcoords):
+        uv_tmesh = TexturedTriMesh(self.uv_mesh, tcoords, generated, trilist=self.uv_trilist)
+        bcs = rasterize_barycentric_coordinate_images(uv_tmesh, self.uv_shape)
+        img = rasterize_mesh_from_barycentric_coordinate_images(uv_tmesh, *bcs)
+        img.pixels = np.clip(img.pixels, 0.0, 1.0)
+        return img
+
+    def render_colored_image(self, view_angle_trg, return_visibility=False):
+        uv_cmesh = ColouredTriMesh(self.uv_mesh, trilist=self.uv_trilist, colours=np.tile(view_angle_trg, [3, 1]).T)
+        bcs = rasterize_barycentric_coordinate_images(uv_cmesh, self.uv_shape)
+        img = rasterize_mesh_from_barycentric_coordinate_images(uv_cmesh, *bcs)
+        img.pixels = np.clip(img.pixels, -1.0, 1.0)
+        if return_visibility:
+            visible = np.sum(bcs[0].pixels, axis=0) != 0
+            return img, visible
+        else:
+            return img
+
+    def camera_tri_angle_src(self, tmesh):#, pose_angle_deg=[0, 0, 0], cam_dist=-4.5):
+        camera_direction = -tmesh.points / np.tile(np.linalg.norm(tmesh.points, axis=1), [3, 1]).T
+        view_angle = np.sum(camera_direction * tmesh.vertex_normals(), 1)
+        return view_angle
+
+    def camera_tri_angle(self, tmesh, pose_angle_deg=[0, 0, 0], cam_dist=-4.5):
+        rot_z = rotation_z(pose_angle_deg[2])
+        rot_y = rotation_y(-pose_angle_deg[1])
+        rot_x = rotation_x(pose_angle_deg[0])
+        rotation = rot_z.compose_before(rot_y).compose_before(rot_x)
+
+        translation = Translation([0, 0, +cam_dist])
+        camera = rotation.compose_before(translation)
+
+        cam_mesh = camera.apply(tmesh)
+        camera_direction = -cam_mesh.points / np.tile(np.linalg.norm(cam_mesh.points, axis=1), [3, 1]).T
+        view_angle = np.sum(camera_direction * cam_mesh.vertex_normals(), 1)
+        return view_angle
+
+    def create_syn(self, face, trg_angle=[0, 0, 0], include_mask=None):
+        view_angle_trg = self.camera_tri_angle(face.tmesh, pose_angle_deg=trg_angle)
+
+        im, projected_mesh = rasterize_image(face.tmesh, self.img_shape, pose_angle_deg=trg_angle, cam_dist=4.5)
+
+        # fill_mask = include_mask.astype(np.bool)
+        fill_mask = ((view_angle_trg < face.view_angle_src) | (face.view_angle_src > 0.4)) & self.tight_mask
+        if include_mask is not None:
+            fill_mask = fill_mask | include_mask.astype(np.bool)
+        if face.exclude_mask is not None:
+            tcoord_sampling = np.round(self.tcoords.points[:,::-1] * face.exclude_mask.shape).astype(np.int)
+            fill_mask[self.mask] =  fill_mask[self.mask] & ~face.exclude_mask[face.exclude_mask.shape[0] - tcoord_sampling[:, 0], tcoord_sampling[:, 1]]
+
+        mask_mesh = ColouredTriMesh(face.tmesh.points, trilist=face.tmesh.trilist, colours=np.tile(fill_mask, [3, 1]).T)
+        mask = rasterize_image(mask_mesh, self.img_shape,pose_angle_deg=trg_angle, cam_dist=4.5)[0]
+
+        return im, projected_mesh[:, :2], mask
+
+
+    def create_align_syn(self, face, trg_angle=[0, 0, 0], include_mask=None):
+        im, projected_mesh, mask = self.create_syn(face, trg_angle, include_mask)
+        imgs, masks, transformation_params = align_im2stylegan(im_menpo2PIL(im), #im_menpo2PIL(mask),
+                                                               im_menpo2PIL(self.extend_mask(im, mask)),
+                                                               projected_mesh[self.lms_ind][:,::-1])
+        aligned_meshes = align_mesh2stylegan(projected_mesh, transformation_params)
+        landmarks = aligned_meshes[self.lms_ind]
+        landmarks[:,1] = 1 - landmarks[:,1]
+
+        heatmaps = generate_heatmap.generate_heatmaps(width=self.args.model_res,
+                                     height=self.args.model_res,
+                                     points=landmarks*self.args.model_res,
+                                     sigma=25)
+
+        landmarks = landmarks[:,::-1]
+        aligned_meshes = aligned_meshes[self.mask]
+
+        return imgs, masks, heatmaps, aligned_meshes
+
+    def get_tmesh(self, im, reconstruction_dict, face_mask):
+        id_features = self.arcface_handler.get_identity_features(im, reconstruction_dict['dense_lms'][self.lms_ind])
+
+        _, yaw_angle, _ = reconstruction_dict['euler_angles']
+        is_profile = np.abs(yaw_angle* 180 / np.pi)>30
+        visibility_threshold = 0.4
+        dense_lms = reconstruction_dict['dense_lms'] / im.shape[::-1]
+        dense_lms[:, 1] = 1 - dense_lms[:, 1]
+
+        im_masked = np.array(im_menpo2PIL(im))
+        mask_landmarks = np.ones_like(im_masked[:,:,0])
+        if face_mask is not None:
+            im_masked = im_masked * np.repeat(np.expand_dims(np.array(face_mask,np.bool),2),3,2)
+            mask_landmarks *= np.array(face_mask, np.uint8)
+
+        im_masked = fill_UV(im_PIL2menpo(im_masked))
+        im_masked.pixels = np.concatenate([im_masked.pixels, np.expand_dims(mask_landmarks,0)],0)
+        img_uv_src = self.render_uv_image(im_masked, dense_lms[self.mask])
+        mask_landmarks = img_uv_src.pixels[3]<0.5
+        img_uv_src.pixels = img_uv_src.pixels[0:3]
+        img_uv_src = fill_UV(img_uv_src)
+        if is_profile:
+            mask_landmarks = binary_dilation(mask_landmarks,disk(5))
+            visibility_threshold = 0.6
+        img_uv_src.pixels[:,mask_landmarks] = 0
+        tcoords = self.tcoords_full.copy()
+        tcoords.points[self.mask] = self.tcoords.points
+
+        tmesh = TexturedTriMesh(reconstruction_dict['vertices'], tcoords.points, img_uv_src,
+                                trilist=reconstruction_dict['trilist'])
+        tmesh_masked = tmesh.from_mask(self.mask)
+        tmesh_rotated = TexturedTriMesh(reconstruction_dict['vertices_rotated'], tmesh.tcoords.points, tmesh.texture,
+                                trilist=tmesh.trilist)
+        view_angle_src = self.camera_tri_angle_src(tmesh_rotated)
+        view_angle_src_masked = view_angle_src[self.mask]
+        view_angle_src_masked[~self.tight_mask[self.mask]] = -1 # Only take tight crop from the original image
+        angle_uv_src, visible = self.render_colored_image(view_angle_src_masked, return_visibility=True)
+        angle_uv_src.pixels[:,~visible | mask_landmarks] = -1.0
+
+        mask = angle_uv_src.pixels[0] < visibility_threshold
+        mask = ~remove_small_holes(~mask, area_threshold=100000)
+
+        if is_profile and self.mode=='soft':
+            mask = binary_dilation(mask, disk(10))
+            img_uv_src.pixels[:, mask] = 0
+            angle_uv_src.pixels[:, mask] = -1
+
+            img_uv_src_flipped = img_uv_src.mirror(1)
+            angle_uv_src_flipped = angle_uv_src.mirror(1)
+            temp = img_uv_src_flipped.pixels
+            pad = int((16 / 1024) * self.uv_shape[1])
+            img_uv_src_flipped.pixels = np.concatenate(
+                [np.zeros([temp.shape[0], temp.shape[1], pad]), temp[:, :, :-pad]], 2)
+            temp = angle_uv_src_flipped.pixels
+            angle_uv_src_flipped.pixels = np.concatenate(
+                [np.zeros([temp.shape[0], temp.shape[1], pad]), temp[:, :, :-pad]], 2)
+
+
+            img_uv_src_flipped = fill_UV(img_uv_src_flipped)
+            img_uv_src = fill_UV(img_uv_src)
+            mask_flipped = (angle_uv_src_flipped.pixels[0] > visibility_threshold) & mask
+            mask_flipped = remove_small_holes(mask_flipped, area_threshold=100000)
+            # mask_flipped = binary_dilation(mask_flipped, disk(15))
+            angle_uv_src.pixels = mask_flipped * angle_uv_src_flipped.pixels + (1 - mask_flipped) * angle_uv_src.pixels
+
+            mask_all = mask_flipped.astype(int).copy()
+            mask_all[~mask_flipped & ~mask] = 2
+            mask_all = fill_UV(Image(np.tile(mask_all, [3, 1, 1]))).pixels[0]
+
+            mask_flipped_g = gaussian(mask_all == 1, sigma=30, multichannel=True, mode='reflect')
+            mask_flipped_inv_g = gaussian(mask_all == 2, sigma=30, multichannel=True, mode='reflect')
+
+            img_uv_src.pixels = mask_flipped_g * img_uv_src_flipped.pixels + mask_flipped_inv_g * img_uv_src.pixels
+            # img_uv_src.pixels[:,mask_flipped] = img_uv_src_flipped.pixels[:,mask_flipped]
+
+            mask = (angle_uv_src.pixels[0] < visibility_threshold)
+            mask = ~remove_small_holes(~mask, area_threshold=100000)
+            # mask = binary_dilation(mask, disk(15))
+            img_uv_src.pixels[:, mask] = 0
+            angle_uv_src.pixels[:, mask] = -1
+
+        img_uv_src = fill_UV(img_uv_src)
+        tmesh.texture = img_uv_src
+        tmesh_rotated.texture = img_uv_src
+
+        face = Face(
+            tmesh=tmesh,
+            tmesh_masked=tmesh_masked,
+            tmesh_rotated=tmesh_rotated,
+            img_uv_src=img_uv_src,
+            angle_uv_src=angle_uv_src,
+            angle_uv_list=[],
+            img_uv_list=[],
+            view_angle_src=view_angle_src,
+            id_features=id_features,
+            exclude_mask=mask,
+            is_profile=is_profile,
+            mode=self.mode
+        )
+
+        face.angle_uv_list = [np.clip(angle_uv_src.pixels * face.coef_dict()['src'],-1,1)]
+        face.img_uv_list = [fill_UV(img_uv_src).pixels]
+
+        return face
+
+    def extend_mask(self, im, mask):
+        # closed_mask = binary_dilation(mask.pixels[0].astype(np.bool), disk(10))
+        # extended_mask = ((np.sum(im.pixels, 0) == 0) & (closed_mask & ~mask.pixels[0].astype(np.bool))) | mask.pixels[0].astype(np.bool)
+        # im_filled = remove_small_holes(np.sum(im.pixels, 0) > 0, area_threshold=1000)
+        # border = binary_dilation(im_filled, disk(10)) & ~binary_erosion(im_filled, disk(10))
+        return mask #Image(extended_mask)# | border)
+
+    def run_iteration(self, face, key, trg_angle):
+
+        imgs, masks, heatmaps, aligned_meshes = self.create_align_syn(face, trg_angle, face.uv_blending[key])
+
+        # Run Optimizer
+        generated_imgs, generated_dlatents = self.projector.run_projection({key: imgs},
+                                                                           {key: masks},
+                                                                           {key: heatmaps},
+                                                                           face.id_features)
+
+        img_uv = self.render_uv_image(im_PIL2menpo(generated_imgs[key]), aligned_meshes)
+        img_uv =fill_UV(img_uv)
+        img_uv = uv_color_normalize(face.img_uv_src, face.angle_uv_src, img_uv, Image(face.angle_uv_list[len(face.img_uv_list)]))
+
+        face.img_uv_list.append(img_uv.pixels)
+
+        final_uv, _ = uv_stiching(face.img_uv_list, face.angle_uv_list[:len(face.img_uv_list)], 40)
+        results_dict = {
+            'generated_imgs': generated_imgs[key],
+            'generated_dlatents': generated_dlatents[key],
+            'imgs': imgs,
+            'masks': masks,
+            'aligned_meshes': aligned_meshes,
+            'img_uv': img_uv,
+            'final_uv': final_uv
+        }
+
+        face.img_uv_src = final_uv
+        face.tmesh.texture = final_uv
+        face.tmesh.tcoords = self.tcoords_full.copy()
+        face.tmesh.tcoords.points[self.mask] = self.tcoords.points
+
+        return face, results_dict
+
+    def run(self, im, reconstruction_dict, face_mask=None):
+        start = time.time()
+        print('Preprocessing...', end=" ")
+        # GANFit compatibility
+        if not 'vertices' in reconstruction_dict: # GANFit
+            reconstruction_dict['vertices'] = reconstruction_dict['tmesh'].points
+            reconstruction_dict['trilist'] = reconstruction_dict['tmesh'].trilist
+        if not 'vertices_rotated' in reconstruction_dict: # GANFit
+            reconstruction_dict['vertices_rotated'] = apply_camera_only3d(reconstruction_dict['vertices'], reconstruction_dict['camera_params'])
+        if not 'euler_angles' in reconstruction_dict: # GANFit
+            reconstruction_dict['euler_angles'] =  get_pose(reconstruction_dict['camera_params'])
+
+        # Prepare Textured Trimesh with visible part of the face
+        face = self.get_tmesh(im, reconstruction_dict, face_mask)
+
+        img_uv_src = face.img_uv_src.copy()
+        angle_uv_src = face.angle_uv_src.copy()
+        print('Done in %.2f secs' % (time.time() - start))
+
+        # Prepare view angle maps
+        start = time.time()
+        print('Building a Visibility Index...', end=" ")
+        angle_uv = {}
+        key_list = ['src']
+        angle_uv_list = [np.clip(angle_uv_src.pixels * face.coef_dict()['src'],-1,1)]
+
+        view_angle_src_full = self.camera_tri_angle_src(face.tmesh_rotated)
+        tcoord_sampling = np.round(self.tcoords.points*angle_uv_src.shape).astype(np.int)
+        view_angle_src_full[self.mask] = angle_uv_src.pixels[0, angle_uv_src.shape[0] - tcoord_sampling[:, 1], tcoord_sampling[:, 0]]
+        view_angle_src_full[~self.tight_mask] = -1  # Only take tight crop from the original image
+
+        angle_list = [np.clip(view_angle_src_full * face.coef_dict()['src'],-1,1)]
+        dummy_im = im_menpo2PIL(img_uv_src)
+
+        # For each view calculate angles towards the camera (Visibility scores)
+        for key, trg_angle in face.rotation_dict().items():
+            view_angle_trg = self.camera_tri_angle(face.tmesh, pose_angle_deg=trg_angle)
+            view_angle_trg = np.clip(view_angle_trg * face.coef_dict()[key],-1,1)
+            _, projected_mesh = rasterize_image(face.tmesh, self.img_shape, pose_angle_deg=trg_angle,
+                                                 cam_dist=4.5)
+            _, _, transformation_params = align_im2stylegan(dummy_im, dummy_im,
+                                                                             projected_mesh[self.lms_ind, :2][:, ::-1])
+            aligned_meshes = align_mesh2stylegan(projected_mesh[:, :2], transformation_params)
+            out_of_plane = ((aligned_meshes[:, 0] > 1) |
+                            (aligned_meshes[:, 1] > 1) |
+                            (aligned_meshes[:, 0] < 0) |
+                            (aligned_meshes[:, 1] < 0))
+
+            view_angle_trg[out_of_plane] = -1
+            angle_list.append(view_angle_trg)
+            angle_uv[key] = self.render_colored_image(view_angle_trg[self.mask])
+            angle_uv_list.append(angle_uv[key].pixels)
+            key_list.append(key)
+
+        # Building a Visibility Index
+        max_ind = np.argmax(angle_list, axis=0)
+        max_ind_one_hot = np.zeros((max_ind.size, max_ind.max() + 1))
+        max_ind_one_hot[np.arange(max_ind.size), max_ind.flatten()] = 1
+        max_ind_one_hot = max_ind_one_hot.reshape(max_ind.shape + (-1,))
+        mask_out_all = np.max(angle_list,axis=0) ==-1
+        max_ind_one_hot[mask_out_all,:] = 0
+
+        uv_blending = {}
+        for i, key in enumerate(key_list):
+            uv_blending[key] = np.zeros(max_ind_one_hot[:,i].shape,np.float)
+            for j in range(i):
+                uv_blending[key] += max_ind_one_hot[:,j]
+            uv_blending[key] = np.clip(uv_blending[key],0, 1)
+
+        face.uv_blending = uv_blending
+        face.angle_uv_list = angle_uv_list
+        print('Done in %.2f secs' % (time.time() - start))
+
+        # Projecting for each of the predefined views
+        start = time.time()
+        print('Projecting...')
+        results_dict = {}
+        for key, trg_angle in face.rotation_dict().items():
+            face, results_dict[key] = self.run_iteration(face, key, trg_angle)
+
+        final_uv = results_dict[key]['final_uv']
+        print('Done in %.2f secs' % (time.time() - start))
+
+        if self.args.frontalize:
+            start = time.time()
+            print('Frontalizing...')
+            imgs = {}
+            masks = {}
+            heatmaps = {}
+            self.projector.perceptual_model.assign_placeholder('id_loss', self.args.use_id_loss_frontalize)
+
+            imgs['frontal'], masks['frontal'], heatmaps['frontal'], _ = self.create_align_syn(face, trg_angle=[0, 0, 0], include_mask=face.uv_blending[key])
+            generated_imgs, generated_dlatents = self.projector.run_projection(imgs, masks, heatmaps, face.id_features, iterations= self.args.iterations_frontalize)
+            results_dict['frontal'] = im_PIL2menpo(generated_imgs['frontal'])
+            results_dict['frontalize'] = [imgs, masks, heatmaps, face.id_features]
+            print('Done in %.2f secs' % (time.time() - start))
+
+        return final_uv, results_dict
diff --git a/reconstruction/ostec/core/perceptual_model.py b/reconstruction/ostec/core/perceptual_model.py
new file mode 100644
index 000000000..599bcf659
--- /dev/null
+++ b/reconstruction/ostec/core/perceptual_model.py
@@ -0,0 +1,304 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+import tensorflow as tf
+import bz2
+import PIL.Image
+from PIL import ImageFilter
+import numpy as np
+from keras.models import Model
+from keras.applications.vgg16 import VGG16, preprocess_input
+import keras.backend as K
+import traceback
+import external.stylegan2.dnnlib.tflib as tflib
+from core.landmark_handler import Landmark_Handler
+from core import arcface_handler
+
+
+def load_images(images_PIL, image_size=256, sharpen=False, im_type='RGB'):
+    loaded_images = list()
+    for img_pil in images_PIL:
+      img = img_pil.convert(im_type)
+      if image_size is not None:
+        img = img.resize((image_size,image_size),PIL.Image.LANCZOS)
+        if (sharpen):
+            img = img.filter(ImageFilter.DETAIL)
+      img = np.array(img)
+      img = np.expand_dims(img, 0)
+      loaded_images.append(img)
+    loaded_images = np.vstack(loaded_images)
+    return loaded_images
+
+def tf_custom_adaptive_loss(a,b):
+    from adaptive import lossfun
+    shape = a.get_shape().as_list()
+    dim = np.prod(shape[1:])
+    a = tf.reshape(a, [-1, dim])
+    b = tf.reshape(b, [-1, dim])
+    loss, _, _ = lossfun(b-a, var_suffix='1')
+    return tf.math.reduce_mean(loss)
+
+def tf_custom_adaptive_rgb_loss(a,b):
+    from adaptive import image_lossfun
+    loss, _, _ = image_lossfun(b-a, color_space='RGB', representation='PIXEL')
+    return tf.math.reduce_mean(loss)
+
+def tf_custom_l1_loss(img1,img2):
+  return tf.math.reduce_mean(tf.math.abs(img2-img1), axis=None)
+
+def tf_custom_logcosh_loss(img1,img2):
+  return tf.math.reduce_mean(tf.keras.losses.logcosh(img1,img2))
+
+def create_stub(batch_size):
+    return tf.constant(0, dtype='float32', shape=(batch_size, 0))
+
+def unpack_bz2(src_path):
+    data = bz2.BZ2File(src_path).read()
+    dst_path = src_path[:-4]
+    with open(dst_path, 'wb') as fp:
+        fp.write(data)
+    return dst_path
+
+class PerceptualModel:
+    def __init__(self, args, batch_size=1, perc_model=None, sess=None):
+        self.sess = tf.get_default_session() if sess is None else sess
+        K.set_session(self.sess)
+        self.epsilon = 0.00000001
+        self.args = args
+        self.lr = args.lr
+        self.decay_rate = args.decay_rate
+        self.decay_steps = args.decay_steps
+        self.img_size = args.image_size
+        self.layer = args.use_vgg_layer
+        self.vgg_loss = args.use_vgg_loss
+        self.face_mask = args.face_mask
+        if (self.layer <= 0 or self.vgg_loss <= self.epsilon):
+            self.vgg_loss = None
+        self.pixel_loss = args.use_pixel_loss
+        if (self.pixel_loss <= self.epsilon):
+            self.pixel_loss = None
+        self.mssim_loss = args.use_mssim_loss
+        if (self.mssim_loss <= self.epsilon):
+            self.mssim_loss = None
+        self.lpips_loss = args.use_lpips_loss
+        if (self.lpips_loss <= self.epsilon):
+            self.lpips_loss = None
+        self.l1_penalty = args.use_l1_penalty
+        if (self.l1_penalty <= self.epsilon):
+            self.l1_penalty = None
+        self.adaptive_loss = args.use_adaptive_loss
+        self.sharpen_input = args.sharpen_input
+        self.batch_size = batch_size
+        if perc_model is not None and self.lpips_loss is not None:
+            self.perc_model = perc_model
+        else:
+            self.perc_model = None
+        self.ref_img = None
+        self.ref_weight = None
+        self.ref_heatmaps = None
+        self.perceptual_model = None
+        self.ref_img_features = None
+        self.features_weight = None
+        self.loss = None
+        self.discriminator_loss = args.use_discriminator_loss
+        if (self.discriminator_loss <= self.epsilon):
+            self.discriminator_loss = None
+        if self.discriminator_loss is not None:
+            self.discriminator = None
+            self.stub = create_stub(batch_size)
+        self.landmark_loss = args.use_landmark_loss
+
+
+
+    def add_placeholder(self, var_name):
+        var_val = getattr(self, var_name)
+        setattr(self, var_name + "_placeholder", tf.placeholder(var_val.dtype, shape=var_val.get_shape()))
+        setattr(self, var_name + "_op", var_val.assign(getattr(self, var_name + "_placeholder")))
+
+    def assign_placeholder(self, var_name, var_val):
+        self.sess.run(getattr(self, var_name + "_op"), {getattr(self, var_name + "_placeholder"): var_val})
+
+    def build_perceptual_model(self, generator, discriminator=None):
+        # Learning rate
+        global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name="global_step")
+        incremented_global_step = tf.assign_add(global_step, 1)
+        self._reset_global_step = tf.assign(global_step, 0)
+        self.learning_rate = tf.train.exponential_decay(self.lr, incremented_global_step,
+                self.decay_steps, self.decay_rate, staircase=True)
+        self.sess.run([self._reset_global_step])
+
+        if self.discriminator_loss is not None:
+            self.discriminator = discriminator
+
+        generated_image_tensor = generator.generated_image
+        generated_image = tf.image.resize_nearest_neighbor(generated_image_tensor,
+                                                                  (self.img_size, self.img_size), align_corners=True)
+        self.generated_img = generated_image
+        self.ref_img = tf.get_variable('ref_img', shape=generated_image.shape,
+                                                dtype='float32', initializer=tf.initializers.zeros())
+        self.ref_weight = tf.get_variable('ref_weight', shape=generated_image.shape,
+                                               dtype='float32', initializer=tf.initializers.zeros())
+        self.add_placeholder("ref_img")
+        self.add_placeholder("ref_weight")
+
+        if (self.vgg_loss is not None):
+            vgg16 = VGG16(include_top=False, input_shape=(self.img_size, self.img_size, 3))
+            self.perceptual_model = Model(vgg16.input, vgg16.layers[self.layer].output)
+            generated_img_features = self.perceptual_model(preprocess_input(self.ref_weight * generated_image))
+            dummy_im = np.zeros([self.args.batch_size, self.img_size, self.img_size, 3],np.float32)
+            self.perceptual_model.predict_on_batch(dummy_im)
+            self.ref_img_features = tf.get_variable('ref_img_features', shape=generated_img_features.shape,
+                                                dtype='float32', initializer=tf.initializers.zeros())
+            self.features_weight = tf.get_variable('features_weight', shape=generated_img_features.shape,
+                                               dtype='float32', initializer=tf.initializers.zeros())
+            self.sess.run([self.features_weight.initializer, self.features_weight.initializer])
+            self.add_placeholder("ref_img_features")
+            self.add_placeholder("features_weight")
+
+        landmark_model = Landmark_Handler(self.args, self.sess, generated_image/255)
+        landmark_model.load_model()
+        ibug84to68_ind = list(range(0, 33, 2)) + list(range(33, 84))
+        self.generated_heatmaps = tf.gather(landmark_model.lms_heatmap_prediction, ibug84to68_ind, axis=3)
+        self.generated_landmarks = tf.gather(landmark_model.pts_predictions, ibug84to68_ind, axis=1)
+
+        self.ref_heatmaps = tf.get_variable('ref_heatmaps', shape=self.generated_heatmaps.shape,
+                                            dtype='float32', initializer=tf.initializers.zeros())
+        self.add_placeholder("ref_heatmaps")
+
+        self.generated_id, vars, _ = arcface_handler.get_input_features(generated_image / 255, self.generated_landmarks[:, :, ::-1])
+        self.init_id_vars = tf.variables_initializer(vars)
+
+        self.org_features = tf.get_variable('org_features', shape=self.generated_id.shape,
+                                                dtype='float32', initializer=tf.initializers.zeros())
+        self.add_placeholder("org_features")
+        self.id_loss = tf.get_variable('id_loss', shape=(),
+                                               dtype='float32', initializer=tf.initializers.zeros())
+        self.add_placeholder('id_loss')
+
+        if self.perc_model is not None and self.lpips_loss is not None:
+            img1 = tflib.convert_images_from_uint8(self.ref_weight * self.ref_img, nhwc_to_nchw=True)
+            img2 = tflib.convert_images_from_uint8(self.ref_weight * generated_image, nhwc_to_nchw=True)
+
+        self.loss = 0
+        # L1 loss on VGG16 features
+        if (self.vgg_loss is not None):
+            if self.adaptive_loss:
+                self.loss += self.vgg_loss * tf_custom_adaptive_loss(self.features_weight * self.ref_img_features, self.features_weight * generated_img_features)
+            else:
+                self.loss += self.vgg_loss * tf_custom_logcosh_loss(self.features_weight * self.ref_img_features, self.features_weight * generated_img_features)
+        # + logcosh loss on image pixels
+        if (self.pixel_loss is not None):
+            if self.adaptive_loss:
+                self.loss += self.pixel_loss * tf_custom_adaptive_rgb_loss(self.ref_weight * self.ref_img, self.ref_weight * generated_image)
+            else:
+                self.loss += self.pixel_loss * tf_custom_logcosh_loss(self.ref_weight * self.ref_img, self.ref_weight * generated_image)
+        # + MS-SIM loss on image pixels
+        if (self.mssim_loss is not None):
+            self.loss += self.mssim_loss * tf.math.reduce_mean(1-tf.image.ssim_multiscale(self.ref_weight * self.ref_img, self.ref_weight * generated_image, 1))
+        # + extra perceptual loss on image pixels
+        if self.perc_model is not None and self.lpips_loss is not None:
+            self.loss += self.lpips_loss * tf.math.reduce_mean(self.perc_model.get_output_for(img1, img2))
+        # + L1 penalty on dlatent weights
+        if self.l1_penalty is not None:
+            self.loss += self.l1_penalty * 512 * tf.math.reduce_mean(tf.math.abs(generator.dlatent_variable-generator.get_dlatent_avg()))
+        # discriminator loss (realism)
+        if self.discriminator_loss is not None:
+            self.loss += self.discriminator_loss * tf.math.reduce_mean(self.discriminator.get_output_for(
+                tflib.convert_images_from_uint8(generated_image_tensor, nhwc_to_nchw=True), self.stub))
+        # - discriminator_network.get_output_for(tflib.convert_images_from_uint8(ref_img, nhwc_to_nchw=True), stub)
+        if self.landmark_loss is not None:
+            self.loss += self.landmark_loss * tf.math.reduce_mean(tf.reduce_sum(tf.pow(self.ref_heatmaps - self.generated_heatmaps, 2), 2))
+        if self.id_loss is not None:
+            self.id_loss_comp = tf.losses.cosine_distance(self.generated_id, self.org_features, 1)
+            self.loss += self.id_loss * self.id_loss_comp
+
+        # Define Optimizer
+        vars_to_optimize = generator.dlatent_variable
+        vars_to_optimize = vars_to_optimize if isinstance(vars_to_optimize, list) else [vars_to_optimize]
+        if self.args.optimizer == 'lbfgs':
+            self.optimizer = tf.contrib.opt.ScipyOptimizerInterface(self.loss, var_list=vars_to_optimize, method='L-BFGS-B', options={'maxiter': self.args.iterations})
+        else:
+            if self.args.optimizer == 'ggt':
+                self.optimizer = tf.contrib.opt.GGTOptimizer(learning_rate=self.learning_rate)
+            else:
+                self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate)
+            self.min_op = self.optimizer.minimize(self.loss, var_list=[vars_to_optimize])
+            self.sess.run(tf.variables_initializer(self.optimizer.variables()))
+        #min_op = optimizer.minimize(self.sess)
+        #optim_results = tfp.optimizer.lbfgs_minimize(make_val_and_grad_fn(get_loss), initial_position=vars_to_optimize, num_correction_pairs=10, tolerance=1e-8)
+        K.manual_variable_initialization(True)
+        self.sess.graph.finalize()  # Graph is read-only after this statement.
+
+
+    def set_reference_images(self, images_PIL, masks_PIL, heatmaps, id_features):
+        assert(len(images_PIL) != 0 and len(images_PIL) <= self.batch_size)
+        loaded_image = load_images(images_PIL, self.img_size, sharpen=self.sharpen_input)
+        loaded_mask = load_images(masks_PIL, self.img_size, sharpen=self.sharpen_input, im_type='L')
+        heatmaps = np.transpose(np.array(heatmaps), [0, 2, 3, 1])
+        input_size = np.array(heatmaps).shape[2]
+        output_size = int(self.ref_heatmaps.shape[1])
+        bin_size = input_size // output_size
+        loaded_heatmaps = heatmaps.reshape((heatmaps.shape[0], output_size, bin_size,
+                                             output_size, bin_size, 68)).max(4).max(2)
+        image_features = None
+        if self.perceptual_model is not None:
+            image_features = self.perceptual_model.predict_on_batch(preprocess_input(np.array(loaded_image)))
+            weight_mask = np.ones(self.features_weight.shape)
+
+        if self.face_mask:
+            image_mask = np.zeros(self.ref_weight.shape)
+            for (i, (im, mask)) in enumerate(zip(loaded_image, loaded_mask)):
+                try:
+                    mask = np.array(mask)/255
+                    mask = np.expand_dims(mask,axis=-1)
+                    mask = np.ones(im.shape,np.float32) * mask #?
+                except Exception as e:
+                    print("Exception in mask handling for mask")
+                    traceback.print_exc()
+                    mask = np.ones(im.shape[:2],np.uint8)
+                    mask = np.ones(im.shape,np.float32) * np.expand_dims(mask,axis=-1)
+                image_mask[i] = mask
+            img = None
+        else:
+            image_mask = np.ones(self.ref_weight.shape)
+
+        if len(images_PIL) != self.batch_size:
+            if image_features is not None:
+                features_space = list(self.features_weight.shape[1:])
+                existing_features_shape = [len(images_PIL)] + features_space
+                empty_features_shape = [self.batch_size - len(images_PIL)] + features_space
+                existing_examples = np.ones(shape=existing_features_shape)
+                empty_examples = np.zeros(shape=empty_features_shape)
+                weight_mask = np.vstack([existing_examples, empty_examples])
+                image_features = np.vstack([image_features, np.zeros(empty_features_shape)])
+
+            images_space = list(self.ref_weight.shape[1:])
+            existing_images_space = [len(images_PIL)] + images_space
+            empty_images_space = [self.batch_size - len(images_PIL)] + images_space
+            existing_images = np.ones(shape=existing_images_space)
+            empty_images = np.zeros(shape=empty_images_space)
+            image_mask = image_mask * np.vstack([existing_images, empty_images])
+            loaded_image = np.vstack([loaded_image, np.zeros(empty_images_space)])
+
+        if image_features is not None:
+            self.assign_placeholder("features_weight", weight_mask)
+            self.assign_placeholder("ref_img_features", image_features)
+        self.assign_placeholder("ref_weight", image_mask)
+        self.assign_placeholder("ref_img", loaded_image)
+        self.assign_placeholder("org_features", id_features)
+        self.assign_placeholder("ref_heatmaps", loaded_heatmaps)
+
+    def optimize(self, vars_to_optimize, iterations=200):
+        self.sess.run(self._reset_global_step)
+        self.sess.run(self.init_id_vars)
+        fetch_ops = [self.min_op, self.loss, self.id_loss_comp, self.learning_rate]
+        for _ in range(iterations):
+            if self.args.optimizer == 'lbfgs':
+                self.optimizer.minimize(self.sess, fetches=[vars_to_optimize, self.loss])
+                yield {"loss":self.loss.eval()}
+            else:
+                _, loss, id_loss, lr = self.sess.run(fetch_ops)
+                yield {"loss":loss,"id_loss":id_loss,"lr":lr}
diff --git a/reconstruction/ostec/core/projection_handler.py b/reconstruction/ostec/core/projection_handler.py
new file mode 100644
index 000000000..d6098d8cb
--- /dev/null
+++ b/reconstruction/ostec/core/projection_handler.py
@@ -0,0 +1,199 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import os
+import argparse
+import pickle
+from tqdm.auto import tqdm
+import PIL.Image
+from PIL import ImageFilter
+import numpy as np
+import external.stylegan2.dnnlib.tflib as tflib
+from external.stylegan2 import pretrained_networks
+from core.generator_model import Generator
+from core.perceptual_model import PerceptualModel, load_images
+import external.stylegan2.dnnlib
+from keras.models import load_model
+from keras.applications.resnet50 import preprocess_input
+
+
+def split_to_batches(l, n):
+    for i in range(0, len(l), n):
+        yield l[i:i + n]
+
+def str2bool(v):
+    if isinstance(v, bool):
+       return v
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+
+class Projection_Handler():
+    def __init__(self, args):
+        self.args = args
+
+        # Initialize generator and perceptual model
+        tflib.init_tf()
+        generator_network, discriminator_network, Gs_network = pretrained_networks.load_networks(args.model_url)
+        self.generator = Generator(Gs_network, args.batch_size, randomize_noise=args.randomize_noise)
+
+        if (args.dlatent_avg != ''):
+            self.generator.set_dlatent_avg(np.load(args.dlatent_avg))
+
+        perc_model = None
+        if (args.use_lpips_loss > 0.00000001):
+            if external.stylegan2.dnnlib.util.is_url(args.vgg_url):
+                stream = external.stylegan2.dnnlib.util.open_url(args.vgg_url, cache_dir='../.stylegan2-cache')
+            else:
+                stream = open(args.vgg_url, 'rb')
+            with stream as f:
+                perc_model = pickle.load(f)
+
+
+        self.perceptual_model = PerceptualModel(args, perc_model=perc_model, batch_size=args.batch_size)
+
+        self.ff_model = None
+        if (self.ff_model is None):
+            if os.path.exists(self.args.load_resnet):
+                from keras.applications.resnet50 import preprocess_input
+                print("Loading ResNet Model:")
+                self.ff_model = load_model(self.args.load_resnet)
+                # self.ff_model._make_predict_function()
+                dummy_im = np.zeros([args.batch_size, args.resnet_image_size, args.resnet_image_size, 3], np.uint8)
+                self.ff_model.predict(preprocess_input(dummy_im))
+        if (self.ff_model is None):
+            if os.path.exists(self.args.load_effnet):
+                from efficientnet import preprocess_input
+                print("Loading EfficientNet Model:")
+                self.ff_model = load_model(self.args.load_effnet)
+
+        self.perceptual_model.build_perceptual_model(self.generator, discriminator_network)
+        self.perceptual_model.assign_placeholder('id_loss', args.use_id_loss)
+
+    def run_projection(self, input_images, masks, heatmaps, id_features, iterations=None):
+        n_iteration = self.args.iterations
+        if iterations is not None:
+            n_iteration = iterations
+        return_imgs = {}
+        return_dlatents = {}
+        # Optimize (only) dlatents by minimizing perceptual loss between reference and generated images in feature space
+        for names in  split_to_batches(list(input_images.keys()), self.args.batch_size):
+            #split_to_batches(list(input_images.keys()), self.args.batch_size):
+            #tqdm(split_to_batches(list(input_images.keys()), self.args.batch_size),
+                                 #total=len(input_images) // self.args.batch_size):
+            # tqdm._instances.clear()
+            images_batch = [input_images[x] for x in names]
+            masks_batch =  [masks[x] for x in names]
+            heatmaps_batch =  [heatmaps[x] for x in names]
+            # if args.output_video:
+            #     video_out = {}
+            #     for name in names:
+            #         video_out[name] = cv2.VideoWriter(os.path.join(args.video_dir, f'{name}.avi'),
+            #                                           cv2.VideoWriter_fourcc(*args.video_codec), args.video_frame_rate,
+            #                                           (args.video_size, args.video_size))
+
+            ## REGRESSION
+            dlatents = None
+            if (self.args.load_last != ''):  # load previous dlatents for initialization
+                for name in names:
+                    dl = np.expand_dims(np.load(os.path.join(self.args.load_last, f'{name}.npy')), axis=0)
+                    if (dlatents is None):
+                        dlatents = dl
+                    else:
+                        dlatents = np.vstack((dlatents, dl))
+            else:
+                if (self.ff_model is not None):  # predict initial dlatents with ResNet model
+                    if (self.args.use_preprocess_input):
+                        dlatents = self.ff_model.predict(
+                            preprocess_input(load_images(images_batch, image_size=self.args.resnet_image_size)))
+                    else:
+                        dlatents = self.ff_model.predict(load_images(images_batch, image_size=self.args.resnet_image_size))
+            if dlatents is not None:
+                self.generator.set_dlatents(dlatents)
+
+            ## OPTIMIZATION
+            self.perceptual_model.set_reference_images(images_batch, masks_batch, heatmaps_batch, id_features)
+
+            op = self.perceptual_model.optimize(self.generator.dlatent_variable, iterations=n_iteration)
+            pbar = tqdm(op, leave=False, total=n_iteration)
+            vid_count = 0
+            best_loss = None
+            best_dlatent = None
+            avg_loss_count = 0
+            if self.args.early_stopping:
+                avg_loss = prev_loss = None
+
+            for loss_dict in pbar:
+                if self.args.early_stopping:  # early stopping feature
+                    if prev_loss is not None:
+                        if avg_loss is not None:
+                            avg_loss = 0.5 * avg_loss + (prev_loss - loss_dict["loss"])
+                            if avg_loss < self.args.early_stopping_threshold:  # count while under threshold; else reset
+                                avg_loss_count += 1
+                            else:
+                                avg_loss_count = 0
+                            if avg_loss_count > self.args.early_stopping_patience:  # stop once threshold is reached
+                                print("")
+                                break
+                        else:
+                            avg_loss = prev_loss - loss_dict["loss"]
+                pbar.set_description(
+                    " ".join(names) + ": " + "; ".join(["{} {:.4f}".format(k, v) for k, v in loss_dict.items()]))
+                if best_loss is None or loss_dict["loss"] < best_loss:
+                    if best_dlatent is None or self.args.average_best_loss <= 0.00000001:
+                        best_dlatent = self.generator.get_dlatents()
+                    else:
+                        best_dlatent = 0.25 * best_dlatent + 0.75 * self.generator.get_dlatents()
+                    if self.args.use_best_loss:
+                        self.generator.set_dlatents(best_dlatent)
+                    best_loss = loss_dict["loss"]
+                # if self.args.output_video and (vid_count % self.args.video_skip == 0):
+                #     batch_frames = self.generator.generate_images()
+                #     for i, name in enumerate(names):
+                #         video_frame = PIL.Image.fromarray(batch_frames[i], 'RGB').resize(
+                #             (self.args.video_size, self.args.video_size), PIL.Image.LANCZOS)
+                #         video_out[name].write(cv2.cvtColor(np.array(video_frame).astype('uint8'), cv2.COLOR_RGB2BGR))
+                self.generator.stochastic_clip_dlatents()
+                prev_loss = loss_dict["loss"]
+            if not self.args.use_best_loss:
+                best_loss = prev_loss
+            # pbar.set_postfix(loss="{:.4f}".format(best_loss))
+            print(" ".join(names), " Loss {:.4f}".format(best_loss))
+
+            # if self.args.output_video:
+            #     for name in names:
+            #         video_out[name].release()
+
+            # Generate images from found dlatents and save them
+            if self.args.use_best_loss:
+                self.generator.set_dlatents(best_dlatent)
+
+            generated_images = self.generator.generate_images()
+            generated_dlatents = self.generator.get_dlatents()
+            for img_array, dlatent, img_path, img_name in zip(generated_images, generated_dlatents, images_batch,
+                                                              names):
+                mask_img = None
+                if self.args.composite_mask and (self.args.load_mask or self.args.face_mask):
+                    _, im_name = os.path.split(img_path)
+                    mask_img = os.path.join(self.args.mask_dir, f'{im_name}')
+                if self.args.composite_mask and mask_img is not None and os.path.isfile(mask_img):
+                    orig_img = PIL.Image.open(img_path).convert('RGB')
+                    width, height = orig_img.size
+                    imask = PIL.Image.open(mask_img).convert('L').resize((width, height))
+                    imask = imask.filter(ImageFilter.GaussianBlur(self.args.composite_blur))
+                    mask = np.array(imask) / 255
+                    mask = np.expand_dims(mask, axis=-1)
+                    img_array = mask * np.array(img_array) + (1.0 - mask) * np.array(orig_img)
+                    img_array = img_array.astype(np.uint8)
+                img = PIL.Image.fromarray(img_array, 'RGB')
+                return_imgs[img_name] = img
+                return_dlatents[img_name] = dlatent
+
+            self.generator.reset_dlatents()
+        return return_imgs, return_dlatents
diff --git a/reconstruction/ostec/environment.yml b/reconstruction/ostec/environment.yml
new file mode 100644
index 000000000..7854c7e9f
--- /dev/null
+++ b/reconstruction/ostec/environment.yml
@@ -0,0 +1,216 @@
+channels:
+  - conda-forge
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=main
+  - _openmp_mutex=4.5=1_gnu
+  - aiocontextvars=0.2.2=py_0
+  - alsa-lib=1.2.3=h516909a_0
+  - apptools=5.1.0=pyh44b312d_0
+  - bzip2=1.0.8=h7f98852_4
+  - c-ares=1.17.1=h7f98852_1
+  - ca-certificates=2021.9.30=h06a4308_1
+  - cairo=1.16.0=h6cf1ce9_1008
+  - certifi=2020.12.5=py36h06a4308_0
+  - cloudpickle=2.0.0=pyhd8ed1ab_0
+  - configobj=5.0.6=py_0
+  - contextvars=2.4=py_0
+  - curl=7.78.0=hea6ffbf_0
+  - cycler=0.10.0=py_2
+  - cytoolz=0.11.0=py36h8f6f2f9_3
+  - cyvlfeat=0.7.0=py36h0280710_0
+  - dask-core=2.25.0=py_0
+  - dbus=1.13.6=h48d8840_2
+  - decorator=4.4.2=py_0
+  - double-conversion=3.1.5=h9c3ff4c_2
+  - eigen=3.3.9=h4bd325d_1
+  - enum34=1.1.10=py36h9f0ad1d_2
+  - envisage=6.0.1=pyhd8ed1ab_0
+  - expat=2.4.1=h9c3ff4c_0
+  - ffmpeg=4.3.2=hca11adc_0
+  - fontconfig=2.13.1=hba837de_1005
+  - freetype=2.10.4=h0708190_1
+  - gettext=0.19.8.1=h0b5b191_1005
+  - gl2ps=1.4.2=h0708190_0
+  - glcontext=2.3.4=py36hc4f0c31_0
+  - glew=2.1.0=h9c3ff4c_2
+  - glib=2.68.3=h9c3ff4c_0
+  - glib-tools=2.68.3=h9c3ff4c_0
+  - gmp=6.2.1=h58526e2_0
+  - gnutls=3.6.13=h85f3911_1
+  - graphite2=1.3.13=h58526e2_1001
+  - gst-plugins-base=1.18.4=hf529b03_2
+  - gstreamer=1.18.4=h76c114f_2
+  - h5py=2.10.0=py36hd6299e0_1
+  - harfbuzz=2.8.2=h83ec7ef_0
+  - hdf4=4.2.15=h10796ff_3
+  - hdf5=1.10.6=nompi_h6a2412b_1114
+  - icu=68.1=h58526e2_0
+  - imagecodecs-lite=2019.12.3=py36h92226af_3
+  - imageio=2.9.0=py_0
+  - immutables=0.15=py36h8f6f2f9_0
+  - importlib-metadata=4.8.1=py36h5fab9bb_0
+  - importlib_metadata=4.8.1=hd8ed1ab_0
+  - importlib_resources=5.2.2=pyhd8ed1ab_0
+  - jasper=1.900.1=h07fcdf6_1006
+  - jbig=2.1=h7f98852_2003
+  - jpeg=9d=h36c2ea0_0
+  - jsoncpp=1.9.4=h4bd325d_3
+  - kiwisolver=1.3.1=py36h605e78d_1
+  - krb5=1.19.2=hcc1bbae_0
+  - lame=3.100=h7f98852_1001
+  - lcms2=2.12=hddcbb42_0
+  - ld_impl_linux-64=2.35.1=h7274673_9
+  - lerc=2.2.1=h9c3ff4c_0
+  - libblas=3.9.0=11_linux64_openblas
+  - libcblas=3.9.0=11_linux64_openblas
+  - libclang=11.1.0=default_ha53f305_1
+  - libcurl=7.78.0=h2574ce0_0
+  - libdeflate=1.7=h7f98852_5
+  - libedit=3.1.20191231=he28a2e2_2
+  - libev=4.33=h516909a_1
+  - libevent=2.1.10=hcdb4288_3
+  - libffi=3.3=he6710b0_2
+  - libgcc-ng=9.3.0=h5101ec6_17
+  - libgfortran-ng=11.2.0=h69a702a_11
+  - libgfortran5=11.2.0=h5c6108e_11
+  - libglib=2.68.3=h3e27bee_0
+  - libglu=9.0.0=he1b5a44_1001
+  - libgomp=9.3.0=h5101ec6_17
+  - libiconv=1.16=h516909a_0
+  - liblapack=3.9.0=11_linux64_openblas
+  - liblapacke=3.9.0=11_linux64_openblas
+  - libllvm11=11.1.0=hf817b99_2
+  - libnetcdf=4.8.0=nompi_hcd642e3_103
+  - libnghttp2=1.43.0=h812cca2_0
+  - libogg=1.3.4=h7f98852_1
+  - libopenblas=0.3.17=pthreads_h8fe5266_1
+  - libopencv=4.5.2=py36hb84549a_1
+  - libopus=1.3.1=h7f98852_1
+  - libpng=1.6.37=h21135ba_2
+  - libpq=13.3=hd57d9b9_0
+  - libprotobuf=3.16.0=h780b84a_0
+  - libssh2=1.9.0=ha56f1ee_6
+  - libstdcxx-ng=9.3.0=hd4cf53a_17
+  - libtheora=1.1.1=h7f98852_1005
+  - libtiff=4.3.0=hf544144_1
+  - libuuid=2.32.1=h7f98852_1000
+  - libvorbis=1.3.7=h9c3ff4c_0
+  - libwebp-base=1.2.0=h7f98852_2
+  - libxcb=1.13=h7f98852_1003
+  - libxkbcommon=1.0.3=he3ba5ed_0
+  - libxml2=2.9.12=h72842e0_0
+  - libzip=1.8.0=h4de3113_0
+  - loguru=0.5.3=py36h5fab9bb_2
+  - lz4-c=1.9.3=h9c3ff4c_1
+  - menpo=0.11.0=py36h355b2fd_0
+  - menpo3d=0.8.3=py36h355b2fd_0
+  - moderngl=5.6.4=py36h284efc9_0
+  - mysql-common=8.0.25=ha770c72_2
+  - mysql-libs=8.0.25=hfa10184_2
+  - ncurses=6.2=he6710b0_1
+  - nettle=3.6=he412f7d_0
+  - networkx=2.5.1=pyhd8ed1ab_0
+  - nspr=4.30=h9c3ff4c_0
+  - nss=3.67=hb5efdd6_0
+  - olefile=0.46=pyh9f0ad1d_1
+  - opencv=4.5.2=py36h5fab9bb_1
+  - openh264=2.1.1=h780b84a_0
+  - openjpeg=2.4.0=hb52868f_1
+  - openssl=1.1.1l=h7f8727e_0
+  - pathlib=1.0.1=py36h5fab9bb_4
+  - pcre=8.45=h9c3ff4c_0
+  - pillow=8.3.1=py36h676a545_0
+  - pip=21.0.1=py36h06a4308_0
+  - pixman=0.40.0=h36c2ea0_0
+  - proj=7.2.0=h277dcde_2
+  - pthread-stubs=0.4=h36c2ea0_1001
+  - pugixml=1.11.4=h9c3ff4c_0
+  - py-opencv=4.5.2=py36hcb3619a_1
+  - pyface=7.3.0=pyh44b312d_1
+  - pygments=2.10.0=pyhd8ed1ab_0
+  - pyparsing=2.4.7=pyh9f0ad1d_0
+  - pyqt=5.12.3=py36h5fab9bb_7
+  - pyqt-impl=5.12.3=py36h7ec31b9_7
+  - pyqt5-sip=4.19.18=py36hc4f0c31_7
+  - pyqtchart=5.12=py36h7ec31b9_7
+  - pyqtwebengine=5.12.1=py36h7ec31b9_7
+  - python=3.6.13=h12debd9_1
+  - python-dateutil=2.8.2=pyhd8ed1ab_0
+  - python_abi=3.6=2_cp36m
+  - pywavelets=1.1.1=py36h92226af_3
+  - qt=5.12.9=hda022c4_4
+  - readline=8.1=h27cfd23_0
+  - scikit-image=0.17.2=py36h284efc9_4
+  - scipy=1.5.3=py36h9e8f40b_0
+  - setuptools=58.0.4=py36h06a4308_0
+  - six=1.16.0=pyh6c4a22f_0
+  - sqlite=3.36.0=hc218d9a_0
+  - tbb=2020.2=h4bd325d_4
+  - tbb-devel=2020.2=h4bd325d_4
+  - tifffile=2019.7.26.2=py36_0
+  - tk=8.6.11=h1ccaba5_0
+  - toolz=0.11.1=py_0
+  - tornado=6.1=py36h8f6f2f9_1
+  - traits=6.2.0=py36h8f6f2f9_0
+  - traitsui=7.2.0=pyhd8ed1ab_0
+  - typing_extensions=3.10.0.2=pyha770c72_0
+  - utfcpp=3.2.1=ha770c72_0
+  - vlfeat=0.9.20=h14c3975_1002
+  - vtk=9.0.1=no_osmesa_py36hfa3a401_109
+  - wheel=0.37.0=pyhd3eb1b0_1
+  - x264=1!161.3030=h7f98852_1
+  - xorg-kbproto=1.0.7=h7f98852_1002
+  - xorg-libice=1.0.10=h7f98852_0
+  - xorg-libsm=1.2.3=hd9c2040_1000
+  - xorg-libx11=1.7.2=h7f98852_0
+  - xorg-libxau=1.0.9=h7f98852_0
+  - xorg-libxdmcp=1.1.3=h7f98852_0
+  - xorg-libxext=1.3.4=h7f98852_1
+  - xorg-libxrender=0.9.10=h7f98852_1003
+  - xorg-libxt=1.2.1=h7f98852_2
+  - xorg-renderproto=0.11.1=h7f98852_1002
+  - xorg-xextproto=7.3.0=h7f98852_1002
+  - xorg-xproto=7.0.31=h7f98852_1007
+  - xz=5.2.5=h7b6447c_0
+  - yaml=0.2.5=h516909a_0
+  - zipp=3.6.0=pyhd8ed1ab_0
+  - zlib=1.2.11=h7b6447c_3
+  - zstd=1.5.0=ha95c52a_0
+  - pip:
+    - absl-py==0.14.1
+    - astor==0.8.1
+    - cached-property==1.5.2
+    - charset-normalizer==2.0.7
+    - cmake==3.21.3
+    - dataclasses==0.8
+    - dlib==19.22.1
+    - dominate==2.6.0
+    - future==0.18.2
+    - gast==0.5.2
+    - google-pasta==0.2.0
+    - grpcio==1.41.0
+    - idna==3.3
+    - imutils==0.5.4
+    - keras==2.3.0
+    - keras-applications==1.0.8
+    - keras-preprocessing==1.0.5
+    - kornia==0.5.5
+    - markdown==3.3.4
+    - matplotlib==3.1.3
+    - mayavi==4.7.2
+    - numpy==1.16.1
+    - protobuf==3.18.1
+    - pyyaml==6.0
+    - requests==2.26.0
+    - tensorboard==1.14.0
+    - tensorflow-estimator==1.14.0
+    - tensorflow-gpu==1.14.0
+    - termcolor==1.1.0
+    - torch==1.6.0
+    - torchvision==0.7.0
+    - tqdm==4.62.3
+    - trimesh==3.9.20
+    - urllib3==1.26.7
+    - werkzeug==2.0.2
+    - wrapt==1.13.2
diff --git a/reconstruction/ostec/external/arcface50.py b/reconstruction/ostec/external/arcface50.py
new file mode 100644
index 000000000..638fca08b
--- /dev/null
+++ b/reconstruction/ostec/external/arcface50.py
@@ -0,0 +1,297 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import tensorflow as tf
+
+__weights_dict = dict()
+
+is_train = False
+
+def load_weights(weight_file):
+    import numpy as np
+
+    if weight_file == None:
+        return
+
+    try:
+        weights_dict = np.load(weight_file).item()
+    except:
+        weights_dict = np.load(weight_file, encoding='bytes').item()
+
+    return weights_dict
+
+
+def KitModel(weight_file = None, data = tf.placeholder(tf.float32, shape = (None, 112, 112, 3), name = 'data')):
+    global __weights_dict
+    __weights_dict = load_weights(weight_file)
+
+    minusscalar0_second = tf.constant(__weights_dict['minusscalar0_second']['value'], name='minusscalar0_second')
+
+    mulscalar0_second = tf.constant(__weights_dict['mulscalar0_second']['value'], name='mulscalar0_second')
+    minusscalar0    = data - minusscalar0_second
+    mulscalar0      = minusscalar0 * mulscalar0_second
+    conv0_pad       = tf.pad(mulscalar0, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    conv0           = convolution(conv0_pad, group=1, strides=[1, 1], padding='VALID', name='conv0')
+    bn0             = batch_normalization(conv0, variance_epsilon=1.9999999494757503e-05, name='bn0')
+    relu0           = prelu(bn0, name='relu0')
+    stage1_unit1_bn1 = batch_normalization(relu0, variance_epsilon=1.9999999494757503e-05, name='stage1_unit1_bn1')
+    stage1_unit1_conv1sc = convolution(relu0, group=1, strides=[2, 2], padding='VALID', name='stage1_unit1_conv1sc')
+    stage1_unit1_conv1_pad = tf.pad(stage1_unit1_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit1_conv1 = convolution(stage1_unit1_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage1_unit1_conv1')
+    stage1_unit1_sc = batch_normalization(stage1_unit1_conv1sc, variance_epsilon=1.9999999494757503e-05, name='stage1_unit1_sc')
+    stage1_unit1_bn2 = batch_normalization(stage1_unit1_conv1, variance_epsilon=1.9999999494757503e-05, name='stage1_unit1_bn2')
+    stage1_unit1_relu1 = prelu(stage1_unit1_bn2, name='stage1_unit1_relu1')
+    stage1_unit1_conv2_pad = tf.pad(stage1_unit1_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit1_conv2 = convolution(stage1_unit1_conv2_pad, group=1, strides=[2, 2], padding='VALID', name='stage1_unit1_conv2')
+    stage1_unit1_bn3 = batch_normalization(stage1_unit1_conv2, variance_epsilon=1.9999999494757503e-05, name='stage1_unit1_bn3')
+    plus0           = stage1_unit1_bn3 + stage1_unit1_sc
+    stage1_unit2_bn1 = batch_normalization(plus0, variance_epsilon=1.9999999494757503e-05, name='stage1_unit2_bn1')
+    stage1_unit2_conv1_pad = tf.pad(stage1_unit2_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit2_conv1 = convolution(stage1_unit2_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage1_unit2_conv1')
+    stage1_unit2_bn2 = batch_normalization(stage1_unit2_conv1, variance_epsilon=1.9999999494757503e-05, name='stage1_unit2_bn2')
+    stage1_unit2_relu1 = prelu(stage1_unit2_bn2, name='stage1_unit2_relu1')
+    stage1_unit2_conv2_pad = tf.pad(stage1_unit2_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit2_conv2 = convolution(stage1_unit2_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage1_unit2_conv2')
+    stage1_unit2_bn3 = batch_normalization(stage1_unit2_conv2, variance_epsilon=1.9999999494757503e-05, name='stage1_unit2_bn3')
+    plus1           = stage1_unit2_bn3 + plus0
+    stage1_unit3_bn1 = batch_normalization(plus1, variance_epsilon=1.9999999494757503e-05, name='stage1_unit3_bn1')
+    stage1_unit3_conv1_pad = tf.pad(stage1_unit3_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit3_conv1 = convolution(stage1_unit3_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage1_unit3_conv1')
+    stage1_unit3_bn2 = batch_normalization(stage1_unit3_conv1, variance_epsilon=1.9999999494757503e-05, name='stage1_unit3_bn2')
+    stage1_unit3_relu1 = prelu(stage1_unit3_bn2, name='stage1_unit3_relu1')
+    stage1_unit3_conv2_pad = tf.pad(stage1_unit3_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage1_unit3_conv2 = convolution(stage1_unit3_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage1_unit3_conv2')
+    stage1_unit3_bn3 = batch_normalization(stage1_unit3_conv2, variance_epsilon=1.9999999494757503e-05, name='stage1_unit3_bn3')
+    plus2           = stage1_unit3_bn3 + plus1
+    stage2_unit1_bn1 = batch_normalization(plus2, variance_epsilon=1.9999999494757503e-05, name='stage2_unit1_bn1')
+    stage2_unit1_conv1sc = convolution(plus2, group=1, strides=[2, 2], padding='VALID', name='stage2_unit1_conv1sc')
+    stage2_unit1_conv1_pad = tf.pad(stage2_unit1_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit1_conv1 = convolution(stage2_unit1_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit1_conv1')
+    stage2_unit1_sc = batch_normalization(stage2_unit1_conv1sc, variance_epsilon=1.9999999494757503e-05, name='stage2_unit1_sc')
+    stage2_unit1_bn2 = batch_normalization(stage2_unit1_conv1, variance_epsilon=1.9999999494757503e-05, name='stage2_unit1_bn2')
+    stage2_unit1_relu1 = prelu(stage2_unit1_bn2, name='stage2_unit1_relu1')
+    stage2_unit1_conv2_pad = tf.pad(stage2_unit1_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit1_conv2 = convolution(stage2_unit1_conv2_pad, group=1, strides=[2, 2], padding='VALID', name='stage2_unit1_conv2')
+    stage2_unit1_bn3 = batch_normalization(stage2_unit1_conv2, variance_epsilon=1.9999999494757503e-05, name='stage2_unit1_bn3')
+    plus3           = stage2_unit1_bn3 + stage2_unit1_sc
+    stage2_unit2_bn1 = batch_normalization(plus3, variance_epsilon=1.9999999494757503e-05, name='stage2_unit2_bn1')
+    stage2_unit2_conv1_pad = tf.pad(stage2_unit2_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit2_conv1 = convolution(stage2_unit2_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit2_conv1')
+    stage2_unit2_bn2 = batch_normalization(stage2_unit2_conv1, variance_epsilon=1.9999999494757503e-05, name='stage2_unit2_bn2')
+    stage2_unit2_relu1 = prelu(stage2_unit2_bn2, name='stage2_unit2_relu1')
+    stage2_unit2_conv2_pad = tf.pad(stage2_unit2_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit2_conv2 = convolution(stage2_unit2_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit2_conv2')
+    stage2_unit2_bn3 = batch_normalization(stage2_unit2_conv2, variance_epsilon=1.9999999494757503e-05, name='stage2_unit2_bn3')
+    plus4           = stage2_unit2_bn3 + plus3
+    stage2_unit3_bn1 = batch_normalization(plus4, variance_epsilon=1.9999999494757503e-05, name='stage2_unit3_bn1')
+    stage2_unit3_conv1_pad = tf.pad(stage2_unit3_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit3_conv1 = convolution(stage2_unit3_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit3_conv1')
+    stage2_unit3_bn2 = batch_normalization(stage2_unit3_conv1, variance_epsilon=1.9999999494757503e-05, name='stage2_unit3_bn2')
+    stage2_unit3_relu1 = prelu(stage2_unit3_bn2, name='stage2_unit3_relu1')
+    stage2_unit3_conv2_pad = tf.pad(stage2_unit3_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit3_conv2 = convolution(stage2_unit3_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit3_conv2')
+    stage2_unit3_bn3 = batch_normalization(stage2_unit3_conv2, variance_epsilon=1.9999999494757503e-05, name='stage2_unit3_bn3')
+    plus5           = stage2_unit3_bn3 + plus4
+    stage2_unit4_bn1 = batch_normalization(plus5, variance_epsilon=1.9999999494757503e-05, name='stage2_unit4_bn1')
+    stage2_unit4_conv1_pad = tf.pad(stage2_unit4_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit4_conv1 = convolution(stage2_unit4_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit4_conv1')
+    stage2_unit4_bn2 = batch_normalization(stage2_unit4_conv1, variance_epsilon=1.9999999494757503e-05, name='stage2_unit4_bn2')
+    stage2_unit4_relu1 = prelu(stage2_unit4_bn2, name='stage2_unit4_relu1')
+    stage2_unit4_conv2_pad = tf.pad(stage2_unit4_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage2_unit4_conv2 = convolution(stage2_unit4_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage2_unit4_conv2')
+    stage2_unit4_bn3 = batch_normalization(stage2_unit4_conv2, variance_epsilon=1.9999999494757503e-05, name='stage2_unit4_bn3')
+    plus6           = stage2_unit4_bn3 + plus5
+    stage3_unit1_bn1 = batch_normalization(plus6, variance_epsilon=1.9999999494757503e-05, name='stage3_unit1_bn1')
+    stage3_unit1_conv1sc = convolution(plus6, group=1, strides=[2, 2], padding='VALID', name='stage3_unit1_conv1sc')
+    stage3_unit1_conv1_pad = tf.pad(stage3_unit1_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit1_conv1 = convolution(stage3_unit1_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit1_conv1')
+    stage3_unit1_sc = batch_normalization(stage3_unit1_conv1sc, variance_epsilon=1.9999999494757503e-05, name='stage3_unit1_sc')
+    stage3_unit1_bn2 = batch_normalization(stage3_unit1_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit1_bn2')
+    stage3_unit1_relu1 = prelu(stage3_unit1_bn2, name='stage3_unit1_relu1')
+    stage3_unit1_conv2_pad = tf.pad(stage3_unit1_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit1_conv2 = convolution(stage3_unit1_conv2_pad, group=1, strides=[2, 2], padding='VALID', name='stage3_unit1_conv2')
+    stage3_unit1_bn3 = batch_normalization(stage3_unit1_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit1_bn3')
+    plus7           = stage3_unit1_bn3 + stage3_unit1_sc
+    stage3_unit2_bn1 = batch_normalization(plus7, variance_epsilon=1.9999999494757503e-05, name='stage3_unit2_bn1')
+    stage3_unit2_conv1_pad = tf.pad(stage3_unit2_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit2_conv1 = convolution(stage3_unit2_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit2_conv1')
+    stage3_unit2_bn2 = batch_normalization(stage3_unit2_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit2_bn2')
+    stage3_unit2_relu1 = prelu(stage3_unit2_bn2, name='stage3_unit2_relu1')
+    stage3_unit2_conv2_pad = tf.pad(stage3_unit2_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit2_conv2 = convolution(stage3_unit2_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit2_conv2')
+    stage3_unit2_bn3 = batch_normalization(stage3_unit2_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit2_bn3')
+    plus8           = stage3_unit2_bn3 + plus7
+    stage3_unit3_bn1 = batch_normalization(plus8, variance_epsilon=1.9999999494757503e-05, name='stage3_unit3_bn1')
+    stage3_unit3_conv1_pad = tf.pad(stage3_unit3_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit3_conv1 = convolution(stage3_unit3_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit3_conv1')
+    stage3_unit3_bn2 = batch_normalization(stage3_unit3_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit3_bn2')
+    stage3_unit3_relu1 = prelu(stage3_unit3_bn2, name='stage3_unit3_relu1')
+    stage3_unit3_conv2_pad = tf.pad(stage3_unit3_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit3_conv2 = convolution(stage3_unit3_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit3_conv2')
+    stage3_unit3_bn3 = batch_normalization(stage3_unit3_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit3_bn3')
+    plus9           = stage3_unit3_bn3 + plus8
+    stage3_unit4_bn1 = batch_normalization(plus9, variance_epsilon=1.9999999494757503e-05, name='stage3_unit4_bn1')
+    stage3_unit4_conv1_pad = tf.pad(stage3_unit4_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit4_conv1 = convolution(stage3_unit4_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit4_conv1')
+    stage3_unit4_bn2 = batch_normalization(stage3_unit4_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit4_bn2')
+    stage3_unit4_relu1 = prelu(stage3_unit4_bn2, name='stage3_unit4_relu1')
+    stage3_unit4_conv2_pad = tf.pad(stage3_unit4_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit4_conv2 = convolution(stage3_unit4_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit4_conv2')
+    stage3_unit4_bn3 = batch_normalization(stage3_unit4_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit4_bn3')
+    plus10          = stage3_unit4_bn3 + plus9
+    stage3_unit5_bn1 = batch_normalization(plus10, variance_epsilon=1.9999999494757503e-05, name='stage3_unit5_bn1')
+    stage3_unit5_conv1_pad = tf.pad(stage3_unit5_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit5_conv1 = convolution(stage3_unit5_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit5_conv1')
+    stage3_unit5_bn2 = batch_normalization(stage3_unit5_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit5_bn2')
+    stage3_unit5_relu1 = prelu(stage3_unit5_bn2, name='stage3_unit5_relu1')
+    stage3_unit5_conv2_pad = tf.pad(stage3_unit5_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit5_conv2 = convolution(stage3_unit5_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit5_conv2')
+    stage3_unit5_bn3 = batch_normalization(stage3_unit5_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit5_bn3')
+    plus11          = stage3_unit5_bn3 + plus10
+    stage3_unit6_bn1 = batch_normalization(plus11, variance_epsilon=1.9999999494757503e-05, name='stage3_unit6_bn1')
+    stage3_unit6_conv1_pad = tf.pad(stage3_unit6_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit6_conv1 = convolution(stage3_unit6_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit6_conv1')
+    stage3_unit6_bn2 = batch_normalization(stage3_unit6_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit6_bn2')
+    stage3_unit6_relu1 = prelu(stage3_unit6_bn2, name='stage3_unit6_relu1')
+    stage3_unit6_conv2_pad = tf.pad(stage3_unit6_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit6_conv2 = convolution(stage3_unit6_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit6_conv2')
+    stage3_unit6_bn3 = batch_normalization(stage3_unit6_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit6_bn3')
+    plus12          = stage3_unit6_bn3 + plus11
+    stage3_unit7_bn1 = batch_normalization(plus12, variance_epsilon=1.9999999494757503e-05, name='stage3_unit7_bn1')
+    stage3_unit7_conv1_pad = tf.pad(stage3_unit7_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit7_conv1 = convolution(stage3_unit7_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit7_conv1')
+    stage3_unit7_bn2 = batch_normalization(stage3_unit7_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit7_bn2')
+    stage3_unit7_relu1 = prelu(stage3_unit7_bn2, name='stage3_unit7_relu1')
+    stage3_unit7_conv2_pad = tf.pad(stage3_unit7_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit7_conv2 = convolution(stage3_unit7_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit7_conv2')
+    stage3_unit7_bn3 = batch_normalization(stage3_unit7_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit7_bn3')
+    plus13          = stage3_unit7_bn3 + plus12
+    stage3_unit8_bn1 = batch_normalization(plus13, variance_epsilon=1.9999999494757503e-05, name='stage3_unit8_bn1')
+    stage3_unit8_conv1_pad = tf.pad(stage3_unit8_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit8_conv1 = convolution(stage3_unit8_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit8_conv1')
+    stage3_unit8_bn2 = batch_normalization(stage3_unit8_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit8_bn2')
+    stage3_unit8_relu1 = prelu(stage3_unit8_bn2, name='stage3_unit8_relu1')
+    stage3_unit8_conv2_pad = tf.pad(stage3_unit8_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit8_conv2 = convolution(stage3_unit8_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit8_conv2')
+    stage3_unit8_bn3 = batch_normalization(stage3_unit8_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit8_bn3')
+    plus14          = stage3_unit8_bn3 + plus13
+    stage3_unit9_bn1 = batch_normalization(plus14, variance_epsilon=1.9999999494757503e-05, name='stage3_unit9_bn1')
+    stage3_unit9_conv1_pad = tf.pad(stage3_unit9_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit9_conv1 = convolution(stage3_unit9_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit9_conv1')
+    stage3_unit9_bn2 = batch_normalization(stage3_unit9_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit9_bn2')
+    stage3_unit9_relu1 = prelu(stage3_unit9_bn2, name='stage3_unit9_relu1')
+    stage3_unit9_conv2_pad = tf.pad(stage3_unit9_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit9_conv2 = convolution(stage3_unit9_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit9_conv2')
+    stage3_unit9_bn3 = batch_normalization(stage3_unit9_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit9_bn3')
+    plus15          = stage3_unit9_bn3 + plus14
+    stage3_unit10_bn1 = batch_normalization(plus15, variance_epsilon=1.9999999494757503e-05, name='stage3_unit10_bn1')
+    stage3_unit10_conv1_pad = tf.pad(stage3_unit10_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit10_conv1 = convolution(stage3_unit10_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit10_conv1')
+    stage3_unit10_bn2 = batch_normalization(stage3_unit10_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit10_bn2')
+    stage3_unit10_relu1 = prelu(stage3_unit10_bn2, name='stage3_unit10_relu1')
+    stage3_unit10_conv2_pad = tf.pad(stage3_unit10_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit10_conv2 = convolution(stage3_unit10_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit10_conv2')
+    stage3_unit10_bn3 = batch_normalization(stage3_unit10_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit10_bn3')
+    plus16          = stage3_unit10_bn3 + plus15
+    stage3_unit11_bn1 = batch_normalization(plus16, variance_epsilon=1.9999999494757503e-05, name='stage3_unit11_bn1')
+    stage3_unit11_conv1_pad = tf.pad(stage3_unit11_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit11_conv1 = convolution(stage3_unit11_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit11_conv1')
+    stage3_unit11_bn2 = batch_normalization(stage3_unit11_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit11_bn2')
+    stage3_unit11_relu1 = prelu(stage3_unit11_bn2, name='stage3_unit11_relu1')
+    stage3_unit11_conv2_pad = tf.pad(stage3_unit11_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit11_conv2 = convolution(stage3_unit11_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit11_conv2')
+    stage3_unit11_bn3 = batch_normalization(stage3_unit11_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit11_bn3')
+    plus17          = stage3_unit11_bn3 + plus16
+    stage3_unit12_bn1 = batch_normalization(plus17, variance_epsilon=1.9999999494757503e-05, name='stage3_unit12_bn1')
+    stage3_unit12_conv1_pad = tf.pad(stage3_unit12_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit12_conv1 = convolution(stage3_unit12_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit12_conv1')
+    stage3_unit12_bn2 = batch_normalization(stage3_unit12_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit12_bn2')
+    stage3_unit12_relu1 = prelu(stage3_unit12_bn2, name='stage3_unit12_relu1')
+    stage3_unit12_conv2_pad = tf.pad(stage3_unit12_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit12_conv2 = convolution(stage3_unit12_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit12_conv2')
+    stage3_unit12_bn3 = batch_normalization(stage3_unit12_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit12_bn3')
+    plus18          = stage3_unit12_bn3 + plus17
+    stage3_unit13_bn1 = batch_normalization(plus18, variance_epsilon=1.9999999494757503e-05, name='stage3_unit13_bn1')
+    stage3_unit13_conv1_pad = tf.pad(stage3_unit13_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit13_conv1 = convolution(stage3_unit13_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit13_conv1')
+    stage3_unit13_bn2 = batch_normalization(stage3_unit13_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit13_bn2')
+    stage3_unit13_relu1 = prelu(stage3_unit13_bn2, name='stage3_unit13_relu1')
+    stage3_unit13_conv2_pad = tf.pad(stage3_unit13_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit13_conv2 = convolution(stage3_unit13_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit13_conv2')
+    stage3_unit13_bn3 = batch_normalization(stage3_unit13_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit13_bn3')
+    plus19          = stage3_unit13_bn3 + plus18
+    stage3_unit14_bn1 = batch_normalization(plus19, variance_epsilon=1.9999999494757503e-05, name='stage3_unit14_bn1')
+    stage3_unit14_conv1_pad = tf.pad(stage3_unit14_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit14_conv1 = convolution(stage3_unit14_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit14_conv1')
+    stage3_unit14_bn2 = batch_normalization(stage3_unit14_conv1, variance_epsilon=1.9999999494757503e-05, name='stage3_unit14_bn2')
+    stage3_unit14_relu1 = prelu(stage3_unit14_bn2, name='stage3_unit14_relu1')
+    stage3_unit14_conv2_pad = tf.pad(stage3_unit14_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage3_unit14_conv2 = convolution(stage3_unit14_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage3_unit14_conv2')
+    stage3_unit14_bn3 = batch_normalization(stage3_unit14_conv2, variance_epsilon=1.9999999494757503e-05, name='stage3_unit14_bn3')
+    plus20          = stage3_unit14_bn3 + plus19
+    stage4_unit1_bn1 = batch_normalization(plus20, variance_epsilon=1.9999999494757503e-05, name='stage4_unit1_bn1')
+    stage4_unit1_conv1sc = convolution(plus20, group=1, strides=[2, 2], padding='VALID', name='stage4_unit1_conv1sc')
+    stage4_unit1_conv1_pad = tf.pad(stage4_unit1_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit1_conv1 = convolution(stage4_unit1_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage4_unit1_conv1')
+    stage4_unit1_sc = batch_normalization(stage4_unit1_conv1sc, variance_epsilon=1.9999999494757503e-05, name='stage4_unit1_sc')
+    stage4_unit1_bn2 = batch_normalization(stage4_unit1_conv1, variance_epsilon=1.9999999494757503e-05, name='stage4_unit1_bn2')
+    stage4_unit1_relu1 = prelu(stage4_unit1_bn2, name='stage4_unit1_relu1')
+    stage4_unit1_conv2_pad = tf.pad(stage4_unit1_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit1_conv2 = convolution(stage4_unit1_conv2_pad, group=1, strides=[2, 2], padding='VALID', name='stage4_unit1_conv2')
+    stage4_unit1_bn3 = batch_normalization(stage4_unit1_conv2, variance_epsilon=1.9999999494757503e-05, name='stage4_unit1_bn3')
+    plus21          = stage4_unit1_bn3 + stage4_unit1_sc
+    stage4_unit2_bn1 = batch_normalization(plus21, variance_epsilon=1.9999999494757503e-05, name='stage4_unit2_bn1')
+    stage4_unit2_conv1_pad = tf.pad(stage4_unit2_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit2_conv1 = convolution(stage4_unit2_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage4_unit2_conv1')
+    stage4_unit2_bn2 = batch_normalization(stage4_unit2_conv1, variance_epsilon=1.9999999494757503e-05, name='stage4_unit2_bn2')
+    stage4_unit2_relu1 = prelu(stage4_unit2_bn2, name='stage4_unit2_relu1')
+    stage4_unit2_conv2_pad = tf.pad(stage4_unit2_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit2_conv2 = convolution(stage4_unit2_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage4_unit2_conv2')
+    stage4_unit2_bn3 = batch_normalization(stage4_unit2_conv2, variance_epsilon=1.9999999494757503e-05, name='stage4_unit2_bn3')
+    plus22          = stage4_unit2_bn3 + plus21
+    stage4_unit3_bn1 = batch_normalization(plus22, variance_epsilon=1.9999999494757503e-05, name='stage4_unit3_bn1')
+    stage4_unit3_conv1_pad = tf.pad(stage4_unit3_bn1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit3_conv1 = convolution(stage4_unit3_conv1_pad, group=1, strides=[1, 1], padding='VALID', name='stage4_unit3_conv1')
+    stage4_unit3_bn2 = batch_normalization(stage4_unit3_conv1, variance_epsilon=1.9999999494757503e-05, name='stage4_unit3_bn2')
+    stage4_unit3_relu1 = prelu(stage4_unit3_bn2, name='stage4_unit3_relu1')
+    stage4_unit3_conv2_pad = tf.pad(stage4_unit3_relu1, paddings = [[0, 0], [1, 1], [1, 1], [0, 0]])
+    stage4_unit3_conv2 = convolution(stage4_unit3_conv2_pad, group=1, strides=[1, 1], padding='VALID', name='stage4_unit3_conv2')
+    stage4_unit3_bn3 = batch_normalization(stage4_unit3_conv2, variance_epsilon=1.9999999494757503e-05, name='stage4_unit3_bn3')
+    plus23          = stage4_unit3_bn3 + plus22
+    bn1             = batch_normalization(plus23, variance_epsilon=1.9999999494757503e-05, name='bn1')
+    pre_fc1_flatten = tf.contrib.layers.flatten(bn1)
+    pre_fc1         = tf.layers.dense(pre_fc1_flatten, 512, kernel_initializer = tf.constant_initializer(__weights_dict['pre_fc1']['weights']), bias_initializer = tf.constant_initializer(__weights_dict['pre_fc1']['bias']), use_bias = True,reuse=tf.AUTO_REUSE)
+    fc1             = batch_normalization(pre_fc1, variance_epsilon=1.9999999494757503e-05, name='fc1')
+    return data, fc1, [stage2_unit1_bn1,stage3_unit1_bn1,stage4_unit1_bn1,pre_fc1_flatten]
+
+
+def convolution(input, name, group, **kwargs):
+    w = tf.Variable(__weights_dict[name]['weights'], trainable=is_train, name=name + "_weight")
+    if group == 1:
+        layer = tf.nn.convolution(input, w, **kwargs)
+    else:
+        weight_groups = tf.split(w, num_or_size_splits=group, axis=-1)
+        xs = tf.split(input, num_or_size_splits=group, axis=-1)
+        convolved = [tf.nn.convolution(x, weight, **kwargs) for
+                    (x, weight) in zip(xs, weight_groups)]
+        layer = tf.concat(convolved, axis=-1)
+
+    if 'bias' in __weights_dict[name]:
+        b = tf.Variable(__weights_dict[name]['bias'], trainable=is_train, name=name + "_bias")
+        layer = layer + b
+    return layer
+
+def prelu(input, name):
+    gamma = tf.Variable(__weights_dict[name]['gamma'], name=name + "_gamma", trainable=is_train)
+    return tf.maximum(0.0, input) + gamma * tf.minimum(0.0, input)
+    
+
+def batch_normalization(input, name, **kwargs):
+    mean = tf.Variable(__weights_dict[name]['mean'], name = name + "_mean", trainable = is_train)
+    variance = tf.Variable(__weights_dict[name]['var'], name = name + "_var", trainable = is_train)
+    offset = tf.Variable(__weights_dict[name]['bias'], name = name + "_bias", trainable = is_train) if 'bias' in __weights_dict[name] else None
+    scale = tf.Variable(__weights_dict[name]['scale'], name = name + "_scale", trainable = is_train) if 'scale' in __weights_dict[name] else None
+    return tf.nn.batch_normalization(input, mean, variance, offset, scale, name = name, **kwargs)
+
diff --git a/reconstruction/ostec/external/face_detector/detect_face.py b/reconstruction/ostec/external/face_detector/detect_face.py
new file mode 100644
index 000000000..6fc01fb49
--- /dev/null
+++ b/reconstruction/ostec/external/face_detector/detect_face.py
@@ -0,0 +1,584 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import tensorflow as tf
+import cv2
+import os
+import inspect
+
+
+class Face_Detector(object):
+    def __init__(self, gpuid = -1):
+        self.minsize = 40  # minimum size of face
+        self.threshold = [0.6, 0.7, 0.7]  # three steps's threshold
+        self.factor = 0.709  # scale factor
+
+        current_dir = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))
+
+        with tf.device('/'+('cpu' if gpuid<0 else 'gpu')+':'+('0' if gpuid<0 else str(gpuid))):
+            with tf.Graph().as_default():
+                sess = tf.Session(config= tf.ConfigProto(device_count = {'GPU': 0 if gpuid<0 else 1}))
+                with sess.as_default():
+                    self.pnet, self.rnet, self.onet = create_detector(sess, '{}/mtcnn'.format(current_dir))
+                    print('MTCNN loaded')
+
+    def face_detection(self,img):
+
+        bounding_boxes, points = detect_face(img, self.minsize,
+                                                         self.pnet, self.rnet, self.onet, self.threshold,
+                                                         self.factor)
+        return bounding_boxes, points.reshape([2,-1]).T
+
+
+def layer(op):
+    '''Decorator for composable network layers.'''
+
+    def layer_decorated(self, *args, **kwargs):
+        # Automatically set a name if not provided.
+        name = kwargs.setdefault('name', self.get_unique_name(op.__name__))
+        # Figure out the layer inputs.
+        if len(self.terminals) == 0:
+            raise RuntimeError('No input variables found for layer %s.' % name)
+        elif len(self.terminals) == 1:
+            layer_input = self.terminals[0]
+        else:
+            layer_input = list(self.terminals)
+        # Perform the operation and get the output.
+        layer_output = op(self, layer_input, *args, **kwargs)
+        # Add to layer LUT.
+        self.layers[name] = layer_output
+        # This output is now the input for the next layer.
+        self.feed(layer_output)
+        # Return self for chained calls.
+        return self
+
+    return layer_decorated
+
+class Network(object):
+
+    def __init__(self, inputs, trainable=True):
+        # The input nodes for this network
+        self.inputs = inputs
+        # The current list of terminal nodes
+        self.terminals = []
+        # Mapping from layer names to layers
+        self.layers = dict(inputs)
+        # If true, the resulting variables are set as trainable
+        self.trainable = trainable
+
+        self.setup()
+
+    def setup(self):
+        '''Construct the network. '''
+        raise NotImplementedError('Must be implemented by the subclass.')
+
+    def load(self, data_path, session, ignore_missing=False):
+        '''Load network weights.
+        data_path: The path to the numpy-serialized network weights
+        session: The current TensorFlow session
+        ignore_missing: If true, serialized weights for missing layers are ignored.
+        '''
+        data_dict = np.load(data_path, encoding='latin1').item() #pylint: disable=no-member
+        for op_name in data_dict:
+            with tf.variable_scope(op_name, reuse=True):
+                for param_name, data in data_dict[op_name].items():
+                    try:
+                        var = tf.get_variable(param_name)
+                        session.run(var.assign(data))
+                    except ValueError:
+                        if not ignore_missing:
+                            raise
+
+    def feed(self, *args):
+        '''Set the input(s) for the next operation by replacing the terminal nodes.
+        The arguments can be either layer names or the actual layers.
+        '''
+        assert len(args) != 0
+        self.terminals = []
+        for fed_layer in args:
+            if isinstance(fed_layer, str):
+                try:
+                    fed_layer = self.layers[fed_layer]
+                except KeyError:
+                    raise KeyError('Unknown layer name fed: %s' % fed_layer)
+            self.terminals.append(fed_layer)
+        return self
+
+    def get_output(self):
+        '''Returns the current network output.'''
+        return self.terminals[-1]
+
+    def get_unique_name(self, prefix):
+        '''Returns an index-suffixed unique name for the given prefix.
+        This is used for auto-generating layer names based on the type-prefix.
+        '''
+        ident = sum(t.startswith(prefix) for t, _ in self.layers.items()) + 1
+        return '%s_%d' % (prefix, ident)
+
+    def make_var(self, name, shape):
+        '''Creates a new TensorFlow variable.'''
+        return tf.get_variable(name, shape, trainable=self.trainable)
+
+    def validate_padding(self, padding):
+        '''Verifies that the padding is one of the supported ones.'''
+        assert padding in ('SAME', 'VALID')
+
+    @layer
+    def conv(self,
+             inp,
+             k_h,
+             k_w,
+             c_o,
+             s_h,
+             s_w,
+             name,
+             relu=True,
+             padding='SAME',
+             group=1,
+             biased=True):
+        # Verify that the padding is acceptable
+        self.validate_padding(padding)
+        # Get the number of channels in the input
+        c_i = inp.get_shape()[-1]
+        # Verify that the grouping parameter is valid
+        assert c_i % group == 0
+        assert c_o % group == 0
+        # Convolution for a given input and kernel
+        convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)
+        with tf.variable_scope(name) as scope:
+            kernel = self.make_var('weights', shape=[k_h, k_w, c_i // group, c_o])
+            # This is the common-case. Convolve the input without any further complications.
+            output = convolve(inp, kernel)
+            # Add the biases
+            if biased:
+                biases = self.make_var('biases', [c_o])
+                output = tf.nn.bias_add(output, biases)
+            if relu:
+                # ReLU non-linearity
+                output = tf.nn.relu(output, name=scope.name)
+            return output
+
+    @layer
+    def prelu(self, inp, name):
+        with tf.variable_scope(name):
+            i = inp.get_shape().as_list()
+            alpha = self.make_var('alpha', shape=(i[-1]))
+            output = tf.nn.relu(inp) + tf.multiply(alpha, -tf.nn.relu(-inp))
+        return output
+
+    @layer
+    def max_pool(self, inp, k_h, k_w, s_h, s_w, name, padding='SAME'):
+        self.validate_padding(padding)
+        return tf.nn.max_pool(inp,
+                              ksize=[1, k_h, k_w, 1],
+                              strides=[1, s_h, s_w, 1],
+                              padding=padding,
+                              name=name)
+
+    @layer
+    def fc(self, inp, num_out, name, relu=True):
+        with tf.variable_scope(name):
+            input_shape = inp.get_shape()
+            if input_shape.ndims == 4:
+                # The input is spatial. Vectorize it first.
+                dim = 1
+                for d in input_shape[1:].as_list():
+                    dim *= d
+                feed_in = tf.reshape(inp, [-1, dim])
+            else:
+                feed_in, dim = (inp, input_shape[-1].value)
+            weights = self.make_var('weights', shape=[dim, num_out])
+            biases = self.make_var('biases', [num_out])
+            op = tf.nn.relu_layer if relu else tf.nn.xw_plus_b
+            fc = op(feed_in, weights, biases, name=name)
+            return fc
+
+
+    """
+    Multi dimensional softmax,
+    refer to https://github.com/tensorflow/tensorflow/issues/210
+    compute softmax along the dimension of target
+    the native softmax only supports batch_size x dimension
+    """
+    @layer
+    def softmax(self, target, axis, name=None):
+        max_axis = tf.reduce_max(target, axis, keepdims=True)
+        target_exp = tf.exp(target-max_axis)
+        normalize = tf.reduce_sum(target_exp, axis, keepdims=True)
+        softmax = tf.div(target_exp, normalize, name)
+        return softmax
+
+class PNet(Network):
+    def setup(self):
+        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
+             .conv(3, 3, 10, 1, 1, padding='VALID', relu=False, name='conv1')
+             .prelu(name='PReLU1')
+             .max_pool(2, 2, 2, 2, name='pool1')
+             .conv(3, 3, 16, 1, 1, padding='VALID', relu=False, name='conv2')
+             .prelu(name='PReLU2')
+             .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv3')
+             .prelu(name='PReLU3')
+             .conv(1, 1, 2, 1, 1, relu=False, name='conv4-1')
+             .softmax(3,name='prob1'))
+
+        (self.feed('PReLU3') #pylint: disable=no-value-for-parameter
+             .conv(1, 1, 4, 1, 1, relu=False, name='conv4-2'))
+
+class RNet(Network):
+    def setup(self):
+        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
+             .conv(3, 3, 28, 1, 1, padding='VALID', relu=False, name='conv1')
+             .prelu(name='prelu1')
+             .max_pool(3, 3, 2, 2, name='pool1')
+             .conv(3, 3, 48, 1, 1, padding='VALID', relu=False, name='conv2')
+             .prelu(name='prelu2')
+             .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
+             .conv(2, 2, 64, 1, 1, padding='VALID', relu=False, name='conv3')
+             .prelu(name='prelu3')
+             .fc(128, relu=False, name='conv4')
+             .prelu(name='prelu4')
+             .fc(2, relu=False, name='conv5-1')
+             .softmax(1,name='prob1'))
+
+        (self.feed('prelu4') #pylint: disable=no-value-for-parameter
+             .fc(4, relu=False, name='conv5-2'))
+
+class ONet(Network):
+    def setup(self):
+        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
+             .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv1')
+             .prelu(name='prelu1')
+             .max_pool(3, 3, 2, 2, name='pool1')
+             .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv2')
+             .prelu(name='prelu2')
+             .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
+             .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv3')
+             .prelu(name='prelu3')
+             .max_pool(2, 2, 2, 2, name='pool3')
+             .conv(2, 2, 128, 1, 1, padding='VALID', relu=False, name='conv4')
+             .prelu(name='prelu4')
+             .fc(256, relu=False, name='conv5')
+             .prelu(name='prelu5')
+             .fc(2, relu=False, name='conv6-1')
+             .softmax(1, name='prob1'))
+
+        (self.feed('prelu5') #pylint: disable=no-value-for-parameter
+             .fc(4, relu=False, name='conv6-2'))
+
+        (self.feed('prelu5') #pylint: disable=no-value-for-parameter
+             .fc(10, relu=False, name='conv6-3'))
+
+def create_detector(sess, model_path):
+    with tf.variable_scope('pnet'):
+        data = tf.placeholder(tf.float32, (None,None,None,3), 'input')
+        pnet = PNet({'data':data})
+        pnet.load(os.path.join(model_path, 'cas1.npy'), sess)
+    with tf.variable_scope('rnet'):
+        data = tf.placeholder(tf.float32, (None,24,24,3), 'input')
+        rnet = RNet({'data':data})
+        rnet.load(os.path.join(model_path, 'cas2.npy'), sess)
+    with tf.variable_scope('onet'):
+        data = tf.placeholder(tf.float32, (None,48,48,3), 'input')
+        onet = ONet({'data':data})
+        onet.load(os.path.join(model_path, 'cas3.npy'), sess)
+
+    pnet_fun = lambda img : sess.run(('pnet/conv4-2/BiasAdd:0', 'pnet/prob1:0'), feed_dict={'pnet/input:0':img})
+    rnet_fun = lambda img : sess.run(('rnet/conv5-2/conv5-2:0', 'rnet/prob1:0'), feed_dict={'rnet/input:0':img})
+    onet_fun = lambda img : sess.run(('onet/conv6-2/conv6-2:0', 'onet/conv6-3/conv6-3:0', 'onet/prob1:0'), feed_dict={'onet/input:0':img})
+    return pnet_fun, rnet_fun, onet_fun
+
+def detect_face(img, minsize, pnet, rnet, onet, threshold, factor):
+    # im: input image
+    # minsize: minimum of faces' size
+    # pnet, rnet, onet: model
+    # threshold: threshold=[th1 th2 th3], th1-3 are three steps's threshold
+    # factor: resize img to generate pyramid
+    factor_count=0
+    total_boxes=np.empty((0,9))
+    points=[]
+    h=img.shape[0]
+    w=img.shape[1]
+    minl=np.amin([h, w])
+    m=12.0/minsize
+    minl=minl*m
+    # creat scale pyramid
+    scales=[]
+    while minl>=12:
+        scales += [m*np.power(factor, factor_count)]
+        minl = minl*factor
+        factor_count += 1
+
+    # first stage
+    for j in range(len(scales)):
+        scale=scales[j]
+        hs=int(np.ceil(h*scale))
+        ws=int(np.ceil(w*scale))
+        im_data = imresample(img, (hs, ws))
+        im_data = (im_data-127.5)*0.0078125
+        img_x = np.expand_dims(im_data, 0)
+        img_y = np.transpose(img_x, (0,2,1,3))
+        out = pnet(img_y)
+        out0 = np.transpose(out[0], (0,2,1,3))
+        out1 = np.transpose(out[1], (0,2,1,3))
+
+        boxes, _ = generateBoundingBox(out1[0,:,:,1].copy(), out0[0,:,:,:].copy(), scale, threshold[0])
+
+        # inter-scale nms
+        pick = nms(boxes.copy(), 0.5, 'Union')
+        if boxes.size>0 and pick.size>0:
+            boxes = boxes[pick,:]
+            total_boxes = np.append(total_boxes, boxes, axis=0)
+
+    numbox = total_boxes.shape[0]
+    if numbox>0:
+        pick = nms(total_boxes.copy(), 0.7, 'Union')
+        total_boxes = total_boxes[pick,:]
+        regw = total_boxes[:,2]-total_boxes[:,0]
+        regh = total_boxes[:,3]-total_boxes[:,1]
+        qq1 = total_boxes[:,0]+total_boxes[:,5]*regw
+        qq2 = total_boxes[:,1]+total_boxes[:,6]*regh
+        qq3 = total_boxes[:,2]+total_boxes[:,7]*regw
+        qq4 = total_boxes[:,3]+total_boxes[:,8]*regh
+        total_boxes = np.transpose(np.vstack([qq1, qq2, qq3, qq4, total_boxes[:,4]]))
+        total_boxes = rerec(total_boxes.copy())
+        total_boxes[:,0:4] = np.fix(total_boxes[:,0:4]).astype(np.int32)
+        dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
+
+    numbox = total_boxes.shape[0]
+    if numbox>0:
+        # second stage
+        tempimg = np.zeros((24,24,3,numbox))
+        for k in range(0,numbox):
+            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
+            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
+            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
+                tempimg[:,:,:,k] = imresample(tmp, (24, 24))
+            else:
+                return np.empty()
+        tempimg = (tempimg-127.5)*0.0078125
+        tempimg1 = np.transpose(tempimg, (3,1,0,2))
+        out = rnet(tempimg1)
+        out0 = np.transpose(out[0])
+        out1 = np.transpose(out[1])
+        score = out1[1,:]
+        ipass = np.where(score>threshold[1])
+        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
+        mv = out0[:,ipass[0]]
+        if total_boxes.shape[0]>0:
+            pick = nms(total_boxes, 0.7, 'Union')
+            total_boxes = total_boxes[pick,:]
+            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv[:,pick]))
+            total_boxes = rerec(total_boxes.copy())
+
+    numbox = total_boxes.shape[0]
+    if numbox>0:
+        # third stage
+        total_boxes = np.fix(total_boxes).astype(np.int32)
+        dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
+        tempimg = np.zeros((48,48,3,numbox))
+        for k in range(0,numbox):
+            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
+            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
+            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
+                tempimg[:,:,:,k] = imresample(tmp, (48, 48))
+            else:
+                return np.empty()
+        tempimg = (tempimg-127.5)*0.0078125
+        tempimg1 = np.transpose(tempimg, (3,1,0,2))
+        out = onet(tempimg1)
+        out0 = np.transpose(out[0])
+        out1 = np.transpose(out[1])
+        out2 = np.transpose(out[2])
+        score = out2[1,:]
+        points = out1
+        ipass = np.where(score>threshold[2])
+        points = points[:,ipass[0]]
+        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
+        mv = out0[:,ipass[0]]
+
+        w = total_boxes[:,2]-total_boxes[:,0]+1
+        h = total_boxes[:,3]-total_boxes[:,1]+1
+        points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1
+        points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1
+        if total_boxes.shape[0]>0:
+            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv))
+            pick = nms(total_boxes.copy(), 0.7, 'Min')
+            total_boxes = total_boxes[pick,:]
+            points = points[:,pick]
+
+    return total_boxes, points
+
+def box_regression(img, onet, total_boxes, threshold):
+    # im: input image
+    # onet: model
+    # total_boxes: [x1 y1 x2 y2 score, 5]    
+    # threshold: 0.7
+    points=[]
+    h=img.shape[0]
+    w=img.shape[1]
+
+    numbox = total_boxes.shape[0]
+    if numbox>0:
+        total_boxes = rerec(total_boxes)
+        total_boxes = np.fix(total_boxes).astype(np.int32)
+        dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
+        tempimg = np.zeros((48,48,3,numbox))
+        for k in range(0,numbox):
+            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
+            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
+            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
+                tempimg[:,:,:,k] = imresample(tmp, (48, 48))
+            else:
+                return np.empty()
+        tempimg = (tempimg-127.5)*0.0078125
+        tempimg1 = np.transpose(tempimg, (3,1,0,2))
+        out = onet(tempimg1)
+        out0 = np.transpose(out[0])
+        out1 = np.transpose(out[1])
+        out2 = np.transpose(out[2])
+        score = out2[1,:]
+        points = out1
+        ipass = np.where(score>threshold)
+        points = points[:,ipass[0]]
+        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
+        mv = out0[:,ipass[0]]
+
+        w = total_boxes[:,2]-total_boxes[:,0]+1
+        h = total_boxes[:,3]-total_boxes[:,1]+1
+        points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1
+        points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1
+        if total_boxes.shape[0]>0:
+            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv))
+            pick = nms(total_boxes.copy(), 0.7, 'Min')
+            total_boxes = total_boxes[pick,:]
+            points = points[:,pick]
+
+    return total_boxes, points
+
+
+def bbreg(boundingbox,reg):
+    # calibrate bounding boxes
+    if reg.shape[1]==1:
+        reg = np.reshape(reg, (reg.shape[2], reg.shape[3]))
+
+    w = boundingbox[:,2]-boundingbox[:,0]+1
+    h = boundingbox[:,3]-boundingbox[:,1]+1
+    b1 = boundingbox[:,0]+reg[:,0]*w
+    b2 = boundingbox[:,1]+reg[:,1]*h
+    b3 = boundingbox[:,2]+reg[:,2]*w
+    b4 = boundingbox[:,3]+reg[:,3]*h
+    boundingbox[:,0:4] = np.transpose(np.vstack([b1, b2, b3, b4 ]))
+    return boundingbox
+
+def generateBoundingBox(imap, reg, scale, t):
+    # use heatmap to generate bounding boxes
+    stride=2
+    cellsize=12
+
+    imap = np.transpose(imap)
+    dx1 = np.transpose(reg[:,:,0])
+    dy1 = np.transpose(reg[:,:,1])
+    dx2 = np.transpose(reg[:,:,2])
+    dy2 = np.transpose(reg[:,:,3])
+    y, x = np.where(imap >= t)
+    if y.shape[0]==1:
+        dx1 = np.flipud(dx1)
+        dy1 = np.flipud(dy1)
+        dx2 = np.flipud(dx2)
+        dy2 = np.flipud(dy2)
+    score = imap[(y,x)]
+    reg = np.transpose(np.vstack([ dx1[(y,x)], dy1[(y,x)], dx2[(y,x)], dy2[(y,x)] ]))
+    if reg.size==0:
+        reg = np.empty((0,3))
+    bb = np.transpose(np.vstack([y,x]))
+    q1 = np.fix((stride*bb+1)/scale)
+    q2 = np.fix((stride*bb+cellsize-1+1)/scale)
+    boundingbox = np.hstack([q1, q2, np.expand_dims(score,1), reg])
+    return boundingbox, reg
+
+def nms(boxes, threshold, method):
+    if boxes.size==0:
+        return np.empty((0,3))
+    x1 = boxes[:,0]
+    y1 = boxes[:,1]
+    x2 = boxes[:,2]
+    y2 = boxes[:,3]
+    s = boxes[:,4]
+    area = (x2-x1+1) * (y2-y1+1)
+    I = np.argsort(s)
+    pick = np.zeros_like(s, dtype=np.int16)
+    counter = 0
+    while I.size>0:
+        i = I[-1]
+        pick[counter] = i
+        counter += 1
+        idx = I[0:-1]
+        xx1 = np.maximum(x1[i], x1[idx])
+        yy1 = np.maximum(y1[i], y1[idx])
+        xx2 = np.minimum(x2[i], x2[idx])
+        yy2 = np.minimum(y2[i], y2[idx])
+        w = np.maximum(0.0, xx2-xx1+1)
+        h = np.maximum(0.0, yy2-yy1+1)
+        inter = w * h
+        if method is 'Min':
+            o = inter / np.minimum(area[i], area[idx])
+        else:
+            o = inter / (area[i] + area[idx] - inter)
+        I = I[np.where(o<=threshold)]
+    pick = pick[0:counter]
+    return pick
+
+def pad(total_boxes, w, h):
+    # compute the padding coordinates (pad the bounding boxes to square)
+    tmpw = (total_boxes[:,2]-total_boxes[:,0]+1).astype(np.int32)
+    tmph = (total_boxes[:,3]-total_boxes[:,1]+1).astype(np.int32)
+    numbox = total_boxes.shape[0]
+
+    dx = np.ones((numbox), dtype=np.int32)
+    dy = np.ones((numbox), dtype=np.int32)
+    edx = tmpw.copy().astype(np.int32)
+    edy = tmph.copy().astype(np.int32)
+
+    x = total_boxes[:,0].copy().astype(np.int32)
+    y = total_boxes[:,1].copy().astype(np.int32)
+    ex = total_boxes[:,2].copy().astype(np.int32)
+    ey = total_boxes[:,3].copy().astype(np.int32)
+
+    tmp = np.where(ex>w)
+    edx[tmp] = np.expand_dims(-ex[tmp]+w+tmpw[tmp],0)
+    ex[tmp] = w
+
+    tmp = np.where(ey>h)
+    edy[tmp] = np.expand_dims(-ey[tmp]+h+tmph[tmp],0)
+    ey[tmp] = h
+
+    tmp = np.where(x<1)
+    dx[tmp] = np.expand_dims(2-x[tmp],0)
+    x[tmp] = 1
+
+    tmp = np.where(y<1)
+    dy[tmp] = np.expand_dims(2-y[tmp],0)
+    y[tmp] = 1
+
+    return dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph
+
+def rerec(bboxA):
+    # convert bboxA to square
+    h = bboxA[:,3]-bboxA[:,1]
+    w = bboxA[:,2]-bboxA[:,0]
+    l = np.maximum(w, h)
+    bboxA[:,0] = bboxA[:,0]+w*0.5-l*0.5
+    bboxA[:,1] = bboxA[:,1]+h*0.5-l*0.5
+    bboxA[:,2:4] = bboxA[:,0:2] + np.transpose(np.tile(l,(2,1)))
+    return bboxA
+
+def imresample(img, sz):
+    im_data = cv2.resize(img, (sz[1], sz[0]), interpolation=cv2.INTER_AREA)
+    return im_data
diff --git a/src/align/det1.npy b/reconstruction/ostec/external/face_detector/mtcnn/cas1.npy
similarity index 100%
rename from src/align/det1.npy
rename to reconstruction/ostec/external/face_detector/mtcnn/cas1.npy
diff --git a/src/align/det2.npy b/reconstruction/ostec/external/face_detector/mtcnn/cas2.npy
similarity index 100%
rename from src/align/det2.npy
rename to reconstruction/ostec/external/face_detector/mtcnn/cas2.npy
diff --git a/src/align/det3.npy b/reconstruction/ostec/external/face_detector/mtcnn/cas3.npy
similarity index 100%
rename from src/align/det3.npy
rename to reconstruction/ostec/external/face_detector/mtcnn/cas3.npy
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/MaskExtractor.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/MaskExtractor.py
new file mode 100644
index 000000000..bb815eac2
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/MaskExtractor.py
@@ -0,0 +1,64 @@
+import torch
+from PIL import Image
+from torchvision.transforms import transforms
+
+from FaceHairMask import deeplab_xception_transfer
+from FaceHairMask.graphonomy_inference import inference
+
+import numpy as np
+import cv2
+
+def preprocess(image, size=256, normalize=1):
+    if size is None:
+        image = transforms.Resize((1024, 1024))(image)
+    else:
+        image = transforms.Resize((size, size))(image)
+    image = transforms.ToTensor()(image)
+    if normalize is not None:
+        image = transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])(image)
+    return image
+
+def postProcess(faceMask, hairMask):
+    hairMask = hairMask.cpu().permute(1,2,0).detach().numpy()
+    faceMask = faceMask.cpu().permute(1,2,0).detach().numpy()
+    return faceMask, hairMask
+
+class MaskExtractor:
+    def __init__(self):
+        
+        #? Hair Face Extractors
+        self.net = deeplab_xception_transfer.deeplab_xception_transfer_projection_savemem(n_classes=20, hidden_layers=128, source_classes=7)
+        stateDict = torch.load("models/Graphonomy/inference.pth")
+        self.net.load_source_model(stateDict)
+        self.net.to("cuda")
+        self.net.eval()
+        
+
+    def processInput4(self, image):
+        preprocessedImage = preprocess(image, size=256, normalize=1)
+        preprocessedImage = preprocessedImage.unsqueeze(0).to("cuda")
+        return preprocessedImage
+    
+    def getMask(self, image):
+        preprocessedImage = self.processInput4(image)
+        _, hairMask, faceMask = inference(net=self.net, img=preprocessedImage, device="cuda")
+        faceMask, hairMask = postProcess(faceMask, hairMask)
+        return hairMask, faceMask
+        
+    def main(self, image):
+        image = (image.pixels_with_channels_at_back()[:, :, ::-1] * 255).astype('uint8')
+        hairMask, faceMask = self.getMask(Image.fromarray(image))
+        hairMask = transforms.Resize((Image.fromarray(image).size[1], Image.fromarray(image).size[0]))(Image.fromarray((hairMask[:,:,0]* 255).astype('uint8')))
+        faceMask = transforms.Resize((Image.fromarray(image).size[1], Image.fromarray(image).size[0]))(Image.fromarray((faceMask[:,:,0]* 255).astype('uint8')))
+
+        # Additional Morphology
+        hairMask = np.array(hairMask) / 255
+        faceMask = np.array(faceMask) / 255
+        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (10, 10))
+        faceMask = cv2.erode(faceMask, kernel, iterations=1)
+        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (35, 35))
+        hairMask = cv2.dilate(hairMask, kernel, iterations=1)
+        faceMask = faceMask * (1 - hairMask)
+
+
+        return hairMask, faceMask
\ No newline at end of file
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingModel.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingModel.py
new file mode 100644
index 000000000..7f165b262
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingModel.py
@@ -0,0 +1,265 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ParsingResnet import Resnet18
+
+class ConvBNReLU(nn.Module):
+    def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1, *args, **kwargs):
+        super(ConvBNReLU, self).__init__()
+        self.conv = nn.Conv2d(in_chan,
+                out_chan,
+                kernel_size = ks,
+                stride = stride,
+                padding = padding,
+                bias = False)
+        self.bn = nn.BatchNorm2d(out_chan)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = F.relu(self.bn(x))
+        return x
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+class BiSeNetOutput(nn.Module):
+    def __init__(self, in_chan, mid_chan, n_classes, *args, **kwargs):
+        super(BiSeNetOutput, self).__init__()
+        self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1)
+        self.conv_out = nn.Conv2d(mid_chan, n_classes, kernel_size=1, bias=False)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.conv_out(x)
+        return x
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class AttentionRefinementModule(nn.Module):
+    def __init__(self, in_chan, out_chan, *args, **kwargs):
+        super(AttentionRefinementModule, self).__init__()
+        self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1)
+        self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size= 1, bias=False)
+        self.bn_atten = nn.BatchNorm2d(out_chan)
+        self.sigmoid_atten = nn.Sigmoid()
+        self.init_weight()
+
+    def forward(self, x):
+        feat = self.conv(x)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv_atten(atten)
+        atten = self.bn_atten(atten)
+        atten = self.sigmoid_atten(atten)
+        out = torch.mul(feat, atten)
+        return out
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+
+class ContextPath(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(ContextPath, self).__init__()
+        self.resnet = Resnet18()
+        self.arm16 = AttentionRefinementModule(256, 128)
+        self.arm32 = AttentionRefinementModule(512, 128)
+        self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_avg = ConvBNReLU(512, 128, ks=1, stride=1, padding=0)
+
+        self.init_weight()
+
+    def forward(self, x):
+        H0, W0 = x.size()[2:]
+        feat8, feat16, feat32 = self.resnet(x)
+        H8, W8 = feat8.size()[2:]
+        H16, W16 = feat16.size()[2:]
+        H32, W32 = feat32.size()[2:]
+
+        avg = F.avg_pool2d(feat32, feat32.size()[2:])
+        avg = self.conv_avg(avg)
+        avg_up = F.interpolate(avg, (H32, W32), mode='nearest')
+
+        feat32_arm = self.arm32(feat32)
+        feat32_sum = feat32_arm + avg_up
+        feat32_up = F.interpolate(feat32_sum, (H16, W16), mode='nearest')
+        feat32_up = self.conv_head32(feat32_up)
+
+        feat16_arm = self.arm16(feat16)
+        feat16_sum = feat16_arm + feat32_up
+        feat16_up = F.interpolate(feat16_sum, (H8, W8), mode='nearest')
+        feat16_up = self.conv_head16(feat16_up)
+
+        return feat8, feat16_up, feat32_up  # x8, x8, x16
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, (nn.Linear, nn.Conv2d)):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+### This is not used, since I replace this with the resnet feature with the same size
+class SpatialPath(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(SpatialPath, self).__init__()
+        self.conv1 = ConvBNReLU(3, 64, ks=7, stride=2, padding=3)
+        self.conv2 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
+        self.conv3 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
+        self.conv_out = ConvBNReLU(64, 128, ks=1, stride=1, padding=0)
+        self.init_weight()
+
+    def forward(self, x):
+        feat = self.conv1(x)
+        feat = self.conv2(feat)
+        feat = self.conv3(feat)
+        feat = self.conv_out(feat)
+        return feat
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class FeatureFusionModule(nn.Module):
+    def __init__(self, in_chan, out_chan, *args, **kwargs):
+        super(FeatureFusionModule, self).__init__()
+        self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0)
+        self.conv1 = nn.Conv2d(out_chan,
+                out_chan//4,
+                kernel_size = 1,
+                stride = 1,
+                padding = 0,
+                bias = False)
+        self.conv2 = nn.Conv2d(out_chan//4,
+                out_chan,
+                kernel_size = 1,
+                stride = 1,
+                padding = 0,
+                bias = False)
+        self.relu = nn.ReLU(inplace=True)
+        self.sigmoid = nn.Sigmoid()
+        self.init_weight()
+
+    def forward(self, fsp, fcp):
+        fcat = torch.cat([fsp, fcp], dim=1)
+        feat = self.convblk(fcat)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv1(atten)
+        atten = self.relu(atten)
+        atten = self.conv2(atten)
+        atten = self.sigmoid(atten)
+        feat_atten = torch.mul(feat, atten)
+        feat_out = feat_atten + feat
+        return feat_out
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class BiSeNet(nn.Module):
+    def __init__(self, n_classes, *args, **kwargs):
+        super(BiSeNet, self).__init__()
+        self.cp = ContextPath()
+        ## here self.sp is deleted
+        self.ffm = FeatureFusionModule(256, 256)
+        self.conv_out = BiSeNetOutput(256, 256, n_classes)
+        self.conv_out16 = BiSeNetOutput(128, 64, n_classes)
+        self.conv_out32 = BiSeNetOutput(128, 64, n_classes)
+        self.init_weight()
+
+    def forward(self, x):
+        H, W = x.size()[2:]
+        feat_res8, feat_cp8, feat_cp16 = self.cp(x)  # here return res3b1 feature
+        feat_sp = feat_res8  # use res3b1 feature to replace spatial path feature
+        feat_fuse = self.ffm(feat_sp, feat_cp8)
+
+        feat_out = self.conv_out(feat_fuse)
+        feat_out16 = self.conv_out16(feat_cp8)
+        feat_out32 = self.conv_out32(feat_cp16)
+
+        feat_out = F.interpolate(feat_out, (H, W), mode='bilinear', align_corners=True)
+        feat_out16 = F.interpolate(feat_out16, (H, W), mode='bilinear', align_corners=True)
+        feat_out32 = F.interpolate(feat_out32, (H, W), mode='bilinear', align_corners=True)
+        return feat_out, feat_out16, feat_out32
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = [], [], [], []
+        for name, child in self.named_children():
+            child_wd_params, child_nowd_params = child.get_params()
+            if isinstance(child, FeatureFusionModule) or isinstance(child, BiSeNetOutput):
+                lr_mul_wd_params += child_wd_params
+                lr_mul_nowd_params += child_nowd_params
+            else:
+                wd_params += child_wd_params
+                nowd_params += child_nowd_params
+        return wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params
\ No newline at end of file
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingResnet.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingResnet.py
new file mode 100644
index 000000000..6a5828430
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/ParsingResnet.py
@@ -0,0 +1,93 @@
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.model_zoo as modelzoo
+
+resnet18_url = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    def __init__(self, in_chan, out_chan, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(in_chan, out_chan, stride)
+        self.bn1 = nn.BatchNorm2d(out_chan)
+        self.conv2 = conv3x3(out_chan, out_chan)
+        self.bn2 = nn.BatchNorm2d(out_chan)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        if in_chan != out_chan or stride != 1:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_chan, out_chan,
+                          kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(out_chan),
+                )
+
+    def forward(self, x):
+        residual = self.conv1(x)
+        residual = F.relu(self.bn1(residual))
+        residual = self.conv2(residual)
+        residual = self.bn2(residual)
+
+        shortcut = x
+        if self.downsample is not None:
+            shortcut = self.downsample(x)
+
+        out = shortcut + residual
+        out = self.relu(out)
+        return out
+
+
+def create_layer_basic(in_chan, out_chan, bnum, stride=1):
+    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
+    for i in range(bnum-1):
+        layers.append(BasicBlock(out_chan, out_chan, stride=1))
+    return nn.Sequential(*layers)
+
+
+class Resnet18(nn.Module):
+    def __init__(self):
+        super(Resnet18, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
+                               bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
+        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
+        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
+        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = F.relu(self.bn1(x))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        feat8 = self.layer2(x) # 1/8
+        feat16 = self.layer3(feat8) # 1/16
+        feat32 = self.layer4(feat16) # 1/32
+        return feat8, feat16, feat32
+
+    def init_weight(self):
+        state_dict = modelzoo.load_url(resnet18_url)
+        self_state_dict = self.state_dict()
+        for k, v in state_dict.items():
+            if 'fc' in k: continue
+            self_state_dict.update({k: v})
+        self.load_state_dict(self_state_dict)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, (nn.Linear, nn.Conv2d)):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module,  nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
\ No newline at end of file
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception.py
new file mode 100644
index 000000000..e2d98d526
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception.py
@@ -0,0 +1,986 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from collections import OrderedDict
+import torch.utils.model_zoo as model_zoo
+from torch.nn.parameter import Parameter
+
+
+
+class SeparableConv2d(nn.Module):
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        kernel_size=3,
+        stride=1,
+        padding=0,
+        dilation=1,
+        bias=False,
+    ):
+        super(SeparableConv2d, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            inplanes,
+            inplanes,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups=inplanes,
+            bias=bias,
+        )
+        self.pointwise = nn.Conv2d(inplanes, planes, 1, 1, 0, 1, 1, bias=bias)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.pointwise(x)
+        return x
+
+
+def fixed_padding(inputs, kernel_size, rate):
+    kernel_size_effective = kernel_size + (kernel_size - 1) * (rate - 1)
+    pad_total = kernel_size_effective - 1
+    pad_beg = pad_total // 2
+    pad_end = pad_total - pad_beg
+    padded_inputs = F.pad(inputs, (pad_beg, pad_end, pad_beg, pad_end))
+    return padded_inputs
+
+
+class SeparableConv2d_aspp(nn.Module):
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        kernel_size=3,
+        stride=1,
+        dilation=1,
+        bias=False,
+        padding=0,
+    ):
+        super(SeparableConv2d_aspp, self).__init__()
+
+        self.depthwise = nn.Conv2d(
+            inplanes,
+            inplanes,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups=inplanes,
+            bias=bias,
+        )
+        self.depthwise_bn = nn.BatchNorm2d(inplanes)
+        self.pointwise = nn.Conv2d(inplanes, planes, 1, 1, 0, 1, 1, bias=bias)
+        self.pointwise_bn = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        #         x = fixed_padding(x, self.depthwise.kernel_size[0], rate=self.depthwise.dilation[0])
+        x = self.depthwise(x)
+        x = self.depthwise_bn(x)
+        x = self.relu(x)
+        x = self.pointwise(x)
+        x = self.pointwise_bn(x)
+        x = self.relu(x)
+        return x
+
+
+class Decoder_module(nn.Module):
+    def __init__(self, inplanes, planes, rate=1):
+        super(Decoder_module, self).__init__()
+        self.atrous_convolution = SeparableConv2d_aspp(
+            inplanes, planes, 3, stride=1, dilation=rate, padding=1
+        )
+
+    def forward(self, x):
+        x = self.atrous_convolution(x)
+        return x
+
+
+class ASPP_module(nn.Module):
+    def __init__(self, inplanes, planes, rate):
+        super(ASPP_module, self).__init__()
+        if rate == 1:
+            raise RuntimeError()
+        else:
+            kernel_size = 3
+            padding = rate
+            self.atrous_convolution = SeparableConv2d_aspp(
+                inplanes, planes, 3, stride=1, dilation=rate, padding=padding
+            )
+
+    def forward(self, x):
+        x = self.atrous_convolution(x)
+        return x
+
+
+class ASPP_module_rate0(nn.Module):
+    def __init__(self, inplanes, planes, rate=1):
+        super(ASPP_module_rate0, self).__init__()
+        if rate == 1:
+            kernel_size = 1
+            padding = 0
+            self.atrous_convolution = nn.Conv2d(
+                inplanes,
+                planes,
+                kernel_size=kernel_size,
+                stride=1,
+                padding=padding,
+                dilation=rate,
+                bias=False,
+            )
+            self.bn = nn.BatchNorm2d(planes, eps=1e-5, affine=True)
+            self.relu = nn.ReLU()
+        else:
+            raise RuntimeError()
+
+    def forward(self, x):
+        x = self.atrous_convolution(x)
+        x = self.bn(x)
+        return self.relu(x)
+
+
+class SeparableConv2d_same(nn.Module):
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        kernel_size=3,
+        stride=1,
+        dilation=1,
+        bias=False,
+        padding=0,
+    ):
+        super(SeparableConv2d_same, self).__init__()
+
+        self.depthwise = nn.Conv2d(
+            inplanes,
+            inplanes,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups=inplanes,
+            bias=bias,
+        )
+        self.depthwise_bn = nn.BatchNorm2d(inplanes)
+        self.pointwise = nn.Conv2d(inplanes, planes, 1, 1, 0, 1, 1, bias=bias)
+        self.pointwise_bn = nn.BatchNorm2d(planes)
+
+    def forward(self, x):
+        x = fixed_padding(
+            x, self.depthwise.kernel_size[0], rate=self.depthwise.dilation[0]
+        )
+        x = self.depthwise(x)
+        x = self.depthwise_bn(x)
+        x = self.pointwise(x)
+        x = self.pointwise_bn(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        reps,
+        stride=1,
+        dilation=1,
+        start_with_relu=True,
+        grow_first=True,
+        is_last=False,
+    ):
+        super(Block, self).__init__()
+
+        if planes != inplanes or stride != 1:
+            self.skip = nn.Conv2d(inplanes, planes, 1, stride=2, bias=False)
+            if is_last:
+                self.skip = nn.Conv2d(inplanes, planes, 1, stride=1, bias=False)
+            self.skipbn = nn.BatchNorm2d(planes)
+        else:
+            self.skip = None
+
+        self.relu = nn.ReLU(inplace=True)
+        rep = []
+
+        filters = inplanes
+        if grow_first:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(inplanes, planes, 3, stride=1, dilation=dilation)
+            )
+            #             rep.append(nn.BatchNorm2d(planes))
+            filters = planes
+
+        for i in range(reps - 1):
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(filters, filters, 3, stride=1, dilation=dilation)
+            )
+        #             rep.append(nn.BatchNorm2d(filters))
+
+        if not grow_first:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(inplanes, planes, 3, stride=1, dilation=dilation)
+            )
+        #             rep.append(nn.BatchNorm2d(planes))
+
+        if not start_with_relu:
+            rep = rep[1:]
+
+        if stride != 1:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(planes, planes, 3, stride=2, dilation=dilation)
+            )
+
+        if is_last:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(planes, planes, 3, stride=1, dilation=dilation)
+            )
+
+        self.rep = nn.Sequential(*rep)
+
+    def forward(self, inp):
+        x = self.rep(inp)
+
+        if self.skip is not None:
+            skip = self.skip(inp)
+            skip = self.skipbn(skip)
+        else:
+            skip = inp
+        # print(x.size(),skip.size())
+        x += skip
+
+        return x
+
+
+class Block2(nn.Module):
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        reps,
+        stride=1,
+        dilation=1,
+        start_with_relu=True,
+        grow_first=True,
+        is_last=False,
+    ):
+        super(Block2, self).__init__()
+
+        if planes != inplanes or stride != 1:
+            self.skip = nn.Conv2d(inplanes, planes, 1, stride=stride, bias=False)
+            self.skipbn = nn.BatchNorm2d(planes)
+        else:
+            self.skip = None
+
+        self.relu = nn.ReLU(inplace=True)
+        rep = []
+
+        filters = inplanes
+        if grow_first:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(inplanes, planes, 3, stride=1, dilation=dilation)
+            )
+            #             rep.append(nn.BatchNorm2d(planes))
+            filters = planes
+
+        for i in range(reps - 1):
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(filters, filters, 3, stride=1, dilation=dilation)
+            )
+        #             rep.append(nn.BatchNorm2d(filters))
+
+        if not grow_first:
+            rep.append(self.relu)
+            rep.append(
+                SeparableConv2d_same(inplanes, planes, 3, stride=1, dilation=dilation)
+            )
+        #             rep.append(nn.BatchNorm2d(planes))
+
+        if not start_with_relu:
+            rep = rep[1:]
+
+        if stride != 1:
+            self.block2_lastconv = nn.Sequential(
+                *[
+                    self.relu,
+                    SeparableConv2d_same(
+                        planes, planes, 3, stride=2, dilation=dilation
+                    ),
+                ]
+            )
+
+        if is_last:
+            rep.append(SeparableConv2d_same(planes, planes, 3, stride=1))
+
+        self.rep = nn.Sequential(*rep)
+
+    def forward(self, inp):
+        x = self.rep(inp)
+        low_middle = x.clone()
+        x1 = x
+        x1 = self.block2_lastconv(x1)
+        if self.skip is not None:
+            skip = self.skip(inp)
+            skip = self.skipbn(skip)
+        else:
+            skip = inp
+
+        x1 += skip
+
+        return x1, low_middle
+
+
+class Xception(nn.Module):
+    """
+    Modified Alighed Xception
+    """
+
+    def __init__(self, inplanes=3, os=16, pretrained=False):
+        super(Xception, self).__init__()
+
+        if os == 16:
+            entry_block3_stride = 2
+            middle_block_rate = 1
+            exit_block_rates = (1, 2)
+        elif os == 8:
+            entry_block3_stride = 1
+            middle_block_rate = 2
+            exit_block_rates = (2, 4)
+        else:
+            raise NotImplementedError
+
+        # Entry flow
+        self.conv1 = nn.Conv2d(inplanes, 32, 3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(32)
+        self.relu = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(32, 64, 3, stride=1, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(64)
+
+        self.block1 = Block(64, 128, reps=2, stride=2, start_with_relu=False)
+        self.block2 = Block2(
+            128, 256, reps=2, stride=2, start_with_relu=True, grow_first=True
+        )
+        self.block3 = Block(
+            256,
+            728,
+            reps=2,
+            stride=entry_block3_stride,
+            start_with_relu=True,
+            grow_first=True,
+        )
+
+        # Middle flow
+        self.block4 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block5 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block6 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block7 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block8 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block9 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block10 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block11 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block12 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block13 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block14 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block15 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block16 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block17 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block18 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+        self.block19 = Block(
+            728,
+            728,
+            reps=3,
+            stride=1,
+            dilation=middle_block_rate,
+            start_with_relu=True,
+            grow_first=True,
+        )
+
+        # Exit flow
+        self.block20 = Block(
+            728,
+            1024,
+            reps=2,
+            stride=1,
+            dilation=exit_block_rates[0],
+            start_with_relu=True,
+            grow_first=False,
+            is_last=True,
+        )
+
+        self.conv3 = SeparableConv2d_aspp(
+            1024,
+            1536,
+            3,
+            stride=1,
+            dilation=exit_block_rates[1],
+            padding=exit_block_rates[1],
+        )
+        # self.bn3 = nn.BatchNorm2d(1536)
+
+        self.conv4 = SeparableConv2d_aspp(
+            1536,
+            1536,
+            3,
+            stride=1,
+            dilation=exit_block_rates[1],
+            padding=exit_block_rates[1],
+        )
+        # self.bn4 = nn.BatchNorm2d(1536)
+
+        self.conv5 = SeparableConv2d_aspp(
+            1536,
+            2048,
+            3,
+            stride=1,
+            dilation=exit_block_rates[1],
+            padding=exit_block_rates[1],
+        )
+        # self.bn5 = nn.BatchNorm2d(2048)
+
+        # Init weights
+        # self.__init_weight()
+
+        # Load pretrained model
+        if pretrained:
+            self.__load_xception_pretrained()
+
+    def forward(self, x):
+        # Entry flow
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        # print('conv1 ',x.size())
+        x = self.conv2(x)
+        x = self.bn2(x)
+        x = self.relu(x)
+
+        x = self.block1(x)
+        # print('block1',x.size())
+        # low_level_feat = x
+        x, low_level_feat = self.block2(x)
+        # print('block2',x.size())
+        x = self.block3(x)
+        # print('xception block3 ',x.size())
+
+        # Middle flow
+        x = self.block4(x)
+        x = self.block5(x)
+        x = self.block6(x)
+        x = self.block7(x)
+        x = self.block8(x)
+        x = self.block9(x)
+        x = self.block10(x)
+        x = self.block11(x)
+        x = self.block12(x)
+        x = self.block13(x)
+        x = self.block14(x)
+        x = self.block15(x)
+        x = self.block16(x)
+        x = self.block17(x)
+        x = self.block18(x)
+        x = self.block19(x)
+
+        # Exit flow
+        x = self.block20(x)
+        x = self.conv3(x)
+        # x = self.bn3(x)
+        x = self.relu(x)
+
+        x = self.conv4(x)
+        # x = self.bn4(x)
+        x = self.relu(x)
+
+        x = self.conv5(x)
+        # x = self.bn5(x)
+        x = self.relu(x)
+
+        return x, low_level_feat
+
+    def __init_weight(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                # n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                # m.weight.data.normal_(0, math.sqrt(2. / n))
+                torch.nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def __load_xception_pretrained(self):
+        pretrain_dict = model_zoo.load_url(
+            "http://data.lip6.fr/cadene/pretrainedmodels/xception-b5690688.pth"
+        )
+        model_dict = {}
+        state_dict = self.state_dict()
+
+        for k, v in pretrain_dict.items():
+            if k in state_dict:
+                if "pointwise" in k:
+                    v = v.unsqueeze(-1).unsqueeze(-1)
+                if k.startswith("block12"):
+                    model_dict[k.replace("block12", "block20")] = v
+                elif k.startswith("block11"):
+                    model_dict[k.replace("block11", "block12")] = v
+                    model_dict[k.replace("block11", "block13")] = v
+                    model_dict[k.replace("block11", "block14")] = v
+                    model_dict[k.replace("block11", "block15")] = v
+                    model_dict[k.replace("block11", "block16")] = v
+                    model_dict[k.replace("block11", "block17")] = v
+                    model_dict[k.replace("block11", "block18")] = v
+                    model_dict[k.replace("block11", "block19")] = v
+                elif k.startswith("conv3"):
+                    model_dict[k] = v
+                elif k.startswith("bn3"):
+                    model_dict[k] = v
+                    model_dict[k.replace("bn3", "bn4")] = v
+                elif k.startswith("conv4"):
+                    model_dict[k.replace("conv4", "conv5")] = v
+                elif k.startswith("bn4"):
+                    model_dict[k.replace("bn4", "bn5")] = v
+                else:
+                    model_dict[k] = v
+        state_dict.update(model_dict)
+        self.load_state_dict(state_dict)
+
+
+class DeepLabv3_plus(nn.Module):
+    def __init__(
+        self, nInputChannels=3, n_classes=21, os=16, pretrained=False, _print=True
+    ):
+        if _print:
+            print("Constructing DeepLabv3+ model...")
+            print("Number of classes: {}".format(n_classes))
+            print("Output stride: {}".format(os))
+            print("Number of Input Channels: {}".format(nInputChannels))
+        super(DeepLabv3_plus, self).__init__()
+
+        # Atrous Conv
+        self.xception_features = Xception(nInputChannels, os, pretrained)
+
+        # ASPP
+        if os == 16:
+            rates = [1, 6, 12, 18]
+        elif os == 8:
+            rates = [1, 12, 24, 36]
+            raise NotImplementedError
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = ASPP_module_rate0(2048, 256, rate=rates[0])
+        self.aspp2 = ASPP_module(2048, 256, rate=rates[1])
+        self.aspp3 = ASPP_module(2048, 256, rate=rates[2])
+        self.aspp4 = ASPP_module(2048, 256, rate=rates[3])
+
+        self.relu = nn.ReLU()
+
+        self.global_avg_pool = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, 1)),
+            nn.Conv2d(2048, 256, 1, stride=1, bias=False),
+            nn.BatchNorm2d(256),
+            nn.ReLU(),
+        )
+
+        self.concat_projection_conv1 = nn.Conv2d(1280, 256, 1, bias=False)
+        self.concat_projection_bn1 = nn.BatchNorm2d(256)
+
+        # adopt [1x1, 48] for channel reduction.
+        self.feature_projection_conv1 = nn.Conv2d(256, 48, 1, bias=False)
+        self.feature_projection_bn1 = nn.BatchNorm2d(48)
+
+        self.decoder = nn.Sequential(Decoder_module(304, 256), Decoder_module(256, 256))
+        self.semantic = nn.Conv2d(256, n_classes, kernel_size=1, stride=1)
+
+    def forward(self, input):
+        x, low_level_features = self.xception_features(input)
+        # print(x.size())
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.upsample(x5, size=x4.size()[2:], mode="bilinear", align_corners=True)
+
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.concat_projection_conv1(x)
+        x = self.concat_projection_bn1(x)
+        x = self.relu(x)
+        # print(x.size())
+
+        low_level_features = self.feature_projection_conv1(low_level_features)
+        low_level_features = self.feature_projection_bn1(low_level_features)
+        low_level_features = self.relu(low_level_features)
+
+        x = F.upsample(
+            x, size=low_level_features.size()[2:], mode="bilinear", align_corners=True
+        )
+        # print(low_level_features.size())
+        # print(x.size())
+        x = torch.cat((x, low_level_features), dim=1)
+        x = self.decoder(x)
+        x = self.semantic(x)
+        x = F.upsample(x, size=input.size()[2:], mode="bilinear", align_corners=True)
+
+        return x
+
+    def freeze_bn(self):
+        for m in self.xception_features.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+
+    def freeze_totally_bn(self):
+        for m in self.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+
+    def freeze_aspp_bn(self):
+        for m in self.aspp1.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+        for m in self.aspp2.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+        for m in self.aspp3.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+        for m in self.aspp4.modules():
+            if isinstance(m, nn.BatchNorm2d):
+                m.eval()
+
+    def learnable_parameters(self):
+        layer_features_BN = []
+        layer_features = []
+        layer_aspp = []
+        layer_projection = []
+        layer_decoder = []
+        layer_other = []
+        model_para = list(self.named_parameters())
+        for name, para in model_para:
+            if "xception" in name:
+                if (
+                    "bn" in name
+                    or "downsample.1.weight" in name
+                    or "downsample.1.bias" in name
+                ):
+                    layer_features_BN.append(para)
+                else:
+                    layer_features.append(para)
+                    # print (name)
+            elif "aspp" in name:
+                layer_aspp.append(para)
+            elif "projection" in name:
+                layer_projection.append(para)
+            elif "decode" in name:
+                layer_decoder.append(para)
+            elif "global" not in name:
+                layer_other.append(para)
+        return (
+            layer_features_BN,
+            layer_features,
+            layer_aspp,
+            layer_projection,
+            layer_decoder,
+            layer_other,
+        )
+
+    def get_backbone_para(self):
+        layer_features = []
+        other_features = []
+        model_para = list(self.named_parameters())
+        for name, para in model_para:
+            if "xception" in name:
+                layer_features.append(para)
+            else:
+                other_features.append(para)
+
+        return layer_features, other_features
+
+    def train_fixbn(self, mode=True, freeze_bn=True, freeze_bn_affine=False):
+        r"""Sets the module in training mode.
+
+        This has any effect only on certain modules. See documentations of
+        particular modules for details of their behaviors in training/evaluation
+        mode, if they are affected, e.g. :class:`Dropout`, :class:`BatchNorm`,
+        etc.
+
+        Returns:
+            Module: self
+        """
+        super(DeepLabv3_plus, self).train(mode)
+        if freeze_bn:
+            print("Freezing Mean/Var of BatchNorm2D.")
+            if freeze_bn_affine:
+                print("Freezing Weight/Bias of BatchNorm2D.")
+        if freeze_bn:
+            for m in self.xception_features.modules():
+                if isinstance(m, nn.BatchNorm2d):
+                    m.eval()
+                    if freeze_bn_affine:
+                        m.weight.requires_grad = False
+                        m.bias.requires_grad = False
+            # for m in self.aspp1.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.aspp2.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.aspp3.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.aspp4.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.global_avg_pool.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.concat_projection_bn1.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+            # for m in self.feature_projection_bn1.modules():
+            #     if isinstance(m, nn.BatchNorm2d):
+            #         m.eval()
+            #         if freeze_bn_affine:
+            #             m.weight.requires_grad = False
+            #             m.bias.requires_grad = False
+
+    def __init_weight(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2.0 / n))
+                # torch.nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def load_state_dict_new(self, state_dict):
+        own_state = self.state_dict()
+        # for name inshop_cos own_state:
+        #    print name
+        new_state_dict = OrderedDict()
+        for name, param in state_dict.items():
+            name = name.replace("module.", "")
+            new_state_dict[name] = 0
+            if name not in own_state:
+                if "num_batch" in name:
+                    continue
+                print('unexpected key "{}" in state_dict'.format(name))
+                continue
+                # if isinstance(param, own_state):
+            if isinstance(param, Parameter):
+                # backwards compatibility for serialized parameters
+                param = param.data
+            try:
+                own_state[name].copy_(param)
+            except:
+                print(
+                    "While copying the parameter named {}, whose dimensions in the model are"
+                    " {} and whose dimensions in the checkpoint are {}, ...".format(
+                        name, own_state[name].size(), param.size()
+                    )
+                )
+                continue  # i add inshop_cos 2018/02/01
+                # raise
+                # print 'copying %s' %name
+                # if isinstance(param, own_state):
+                # backwards compatibility for serialized parameters
+            own_state[name].copy_(param)
+            # print 'copying %s' %name
+
+        missing = set(own_state.keys()) - set(new_state_dict.keys())
+        if len(missing) > 0:
+            print('missing keys in state_dict: "{}"'.format(missing))
+
+
+def get_1x_lr_params(model):
+    """
+    This generator returns all the parameters of the net except for
+    the last classification layer. Note that for each batchnorm layer,
+    requires_grad is set to False in deeplab_resnet.py, therefore this function does not return
+    any batchnorm parameter
+    """
+    b = [model.xception_features]
+    for i in range(len(b)):
+        for k in b[i].parameters():
+            if k.requires_grad:
+                yield k
+
+
+def get_10x_lr_params(model):
+    """
+    This generator returns all the parameters for the last layer of the net,
+    which does the classification of pixel into classes
+    """
+    b = [
+        model.aspp1,
+        model.aspp2,
+        model.aspp3,
+        model.aspp4,
+        model.conv1,
+        model.conv2,
+        model.last_conv,
+    ]
+    for j in range(len(b)):
+        for k in b[j].parameters():
+            if k.requires_grad:
+                yield k
+
+
+if __name__ == "__main__":
+    model = DeepLabv3_plus(
+        nInputChannels=3, n_classes=21, os=16, pretrained=False, _print=True
+    )
+    model.eval()
+    image = torch.randn(1, 3, 512, 512) * 255
+    with torch.no_grad():
+        output = model.forward(image)
+    print(output.size())
+    # print(output)
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception_transfer.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception_transfer.py
new file mode 100644
index 000000000..c4f9cfb62
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/deeplab_xception_transfer.py
@@ -0,0 +1,753 @@
+import torch
+import torch.nn as nn
+from torch.nn import Parameter
+import torch.nn.functional as F
+from collections import OrderedDict
+from . import deeplab_xception, gcn
+from torch.nn.parameter import Parameter
+
+
+#######################
+# base model
+#######################
+
+
+class deeplab_xception_transfer_basemodel(deeplab_xception.DeepLabv3_plus):
+    def __init__(
+        self,
+        nInputChannels=3,
+        n_classes=7,
+        os=16,
+        input_channels=256,
+        hidden_layers=128,
+        out_channels=256,
+    ):
+        super(deeplab_xception_transfer_basemodel, self).__init__(
+            nInputChannels=nInputChannels,
+            n_classes=n_classes,
+            os=os,
+        )
+        ### source graph
+        # self.source_featuremap_2_graph = gcn.Featuremaps_to_Graph(input_channels=input_channels, hidden_layers=hidden_layers,
+        #                                                    nodes=n_classes)
+        # self.source_graph_conv1 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        # self.source_graph_conv2 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        # self.source_graph_conv3 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        #
+        # self.source_graph_2_fea = gcn.Graph_to_Featuremaps(input_channels=input_channels, output_channels=out_channels,
+        #                                             hidden_layers=hidden_layers, nodes=n_classes
+        #                                             )
+        # self.source_skip_conv = nn.Sequential(*[nn.Conv2d(input_channels, input_channels, kernel_size=1),
+        #                                  nn.ReLU(True)])
+
+        ### target graph
+        self.target_featuremap_2_graph = gcn.Featuremaps_to_Graph(
+            input_channels=input_channels, hidden_layers=hidden_layers, nodes=n_classes
+        )
+        self.target_graph_conv1 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.target_graph_conv2 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.target_graph_conv3 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+
+        self.target_graph_2_fea = gcn.Graph_to_Featuremaps(
+            input_channels=input_channels,
+            output_channels=out_channels,
+            hidden_layers=hidden_layers,
+            nodes=n_classes,
+        )
+        self.target_skip_conv = nn.Sequential(
+            *[nn.Conv2d(input_channels, input_channels, kernel_size=1), nn.ReLU(True)]
+        )
+
+    def load_source_model(self, state_dict):
+        own_state = self.state_dict()
+        # for name inshop_cos own_state:
+        #    print name
+        new_state_dict = OrderedDict()
+        for name, param in state_dict.items():
+            name = name.replace("module.", "")
+            if (
+                "graph" in name
+                and "source" not in name
+                and "target" not in name
+                and "fc_graph" not in name
+                and "transpose_graph" not in name
+            ):
+                if "featuremap_2_graph" in name:
+                    name = name.replace(
+                        "featuremap_2_graph", "source_featuremap_2_graph"
+                    )
+                else:
+                    name = name.replace("graph", "source_graph")
+            new_state_dict[name] = 0
+            if name not in own_state:
+                if "num_batch" in name:
+                    continue
+                print('unexpected key "{}" in state_dict'.format(name))
+                continue
+                # if isinstance(param, own_state):
+            if isinstance(param, Parameter):
+                # backwards compatibility for serialized parameters
+                param = param.data
+            try:
+                own_state[name].copy_(param)
+            except:
+                print(
+                    "While copying the parameter named {}, whose dimensions in the model are"
+                    " {} and whose dimensions in the checkpoint are {}, ...".format(
+                        name, own_state[name].size(), param.size()
+                    )
+                )
+                continue  # i add inshop_cos 2018/02/01
+            own_state[name].copy_(param)
+            # print 'copying %s' %name
+
+        missing = set(own_state.keys()) - set(new_state_dict.keys())
+        if len(missing) > 0:
+            print('missing keys in state_dict: "{}"'.format(missing))
+
+    def get_target_parameter(self):
+        l = []
+        other = []
+        for name, k in self.named_parameters():
+            if "target" in name or "semantic" in name:
+                l.append(k)
+            else:
+                other.append(k)
+        return l, other
+
+    def get_semantic_parameter(self):
+        l = []
+        for name, k in self.named_parameters():
+            if "semantic" in name:
+                l.append(k)
+        return l
+
+    def get_source_parameter(self):
+        l = []
+        for name, k in self.named_parameters():
+            if "source" in name:
+                l.append(k)
+        return l
+
+    def forward(self, input, adj1_target=None, adj2_source=None, adj3_transfer=None):
+        x, low_level_features = self.xception_features(input)
+        # print(x.size())
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.upsample(x5, size=x4.size()[2:], mode="bilinear", align_corners=True)
+
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.concat_projection_conv1(x)
+        x = self.concat_projection_bn1(x)
+        x = self.relu(x)
+        # print(x.size())
+        x = F.upsample(
+            x, size=low_level_features.size()[2:], mode="bilinear", align_corners=True
+        )
+
+        low_level_features = self.feature_projection_conv1(low_level_features)
+        low_level_features = self.feature_projection_bn1(low_level_features)
+        low_level_features = self.relu(low_level_features)
+        # print(low_level_features.size())
+        # print(x.size())
+        x = torch.cat((x, low_level_features), dim=1)
+        x = self.decoder(x)
+
+        ### add graph
+
+        # target graph
+        # print('x size',x.size(),adj1.size())
+        graph = self.target_featuremap_2_graph(x)
+
+        # graph combine
+        # print(graph.size(),source_2_target_graph.size())
+        # graph = self.fc_graph.forward(graph,relu=True)
+        # print(graph.size())
+
+        graph = self.target_graph_conv1.forward(graph, adj=adj1_target, relu=True)
+        graph = self.target_graph_conv2.forward(graph, adj=adj1_target, relu=True)
+        graph = self.target_graph_conv3.forward(graph, adj=adj1_target, relu=True)
+        # print(graph.size(),x.size())
+        # graph = self.gcn_encode.forward(graph,relu=True)
+        # graph = self.graph_conv2.forward(graph,adj=adj2,relu=True)
+        # graph = self.gcn_decode.forward(graph,relu=True)
+        graph = self.target_graph_2_fea.forward(graph, x)
+        x = self.target_skip_conv(x)
+        x = x + graph
+
+        ###
+        x = self.semantic(x)
+        x = F.upsample(x, size=input.size()[2:], mode="bilinear", align_corners=True)
+
+        return x
+
+
+class deeplab_xception_transfer_basemodel_savememory(deeplab_xception.DeepLabv3_plus):
+    def __init__(
+        self,
+        nInputChannels=3,
+        n_classes=7,
+        os=16,
+        input_channels=256,
+        hidden_layers=128,
+        out_channels=256,
+    ):
+        super(deeplab_xception_transfer_basemodel_savememory, self).__init__(
+            nInputChannels=nInputChannels,
+            n_classes=n_classes,
+            os=os,
+        )
+        ### source graph
+
+        ### target graph
+        self.target_featuremap_2_graph = gcn.Featuremaps_to_Graph(
+            input_channels=input_channels, hidden_layers=hidden_layers, nodes=n_classes
+        )
+        self.target_graph_conv1 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.target_graph_conv2 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.target_graph_conv3 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+
+        self.target_graph_2_fea = gcn.Graph_to_Featuremaps_savemem(
+            input_channels=input_channels,
+            output_channels=out_channels,
+            hidden_layers=hidden_layers,
+            nodes=n_classes,
+        )
+        self.target_skip_conv = nn.Sequential(
+            *[nn.Conv2d(input_channels, input_channels, kernel_size=1), nn.ReLU(True)]
+        )
+
+    def load_source_model(self, state_dict):
+        own_state = self.state_dict()
+        # for name inshop_cos own_state:
+        #    print name
+        new_state_dict = OrderedDict()
+        for name, param in state_dict.items():
+            name = name.replace("module.", "")
+            if (
+                "graph" in name
+                and "source" not in name
+                and "target" not in name
+                and "fc_graph" not in name
+                and "transpose_graph" not in name
+            ):
+                if "featuremap_2_graph" in name:
+                    name = name.replace(
+                        "featuremap_2_graph", "source_featuremap_2_graph"
+                    )
+                else:
+                    name = name.replace("graph", "source_graph")
+            new_state_dict[name] = 0
+            if name not in own_state:
+                if "num_batch" in name:
+                    continue
+                print('unexpected key "{}" in state_dict'.format(name))
+                continue
+                # if isinstance(param, own_state):
+            if isinstance(param, Parameter):
+                # backwards compatibility for serialized parameters
+                param = param.data
+            try:
+                own_state[name].copy_(param)
+            except:
+                print(
+                    "While copying the parameter named {}, whose dimensions in the model are"
+                    " {} and whose dimensions in the checkpoint are {}, ...".format(
+                        name, own_state[name].size(), param.size()
+                    )
+                )
+                continue  # i add inshop_cos 2018/02/01
+            own_state[name].copy_(param)
+            # print 'copying %s' %name
+
+        missing = set(own_state.keys()) - set(new_state_dict.keys())
+        if len(missing) > 0:
+            print('missing keys in state_dict: "{}"'.format(missing))
+
+    def get_target_parameter(self):
+        l = []
+        other = []
+        for name, k in self.named_parameters():
+            if "target" in name or "semantic" in name:
+                l.append(k)
+            else:
+                other.append(k)
+        return l, other
+
+    def get_semantic_parameter(self):
+        l = []
+        for name, k in self.named_parameters():
+            if "semantic" in name:
+                l.append(k)
+        return l
+
+    def get_source_parameter(self):
+        l = []
+        for name, k in self.named_parameters():
+            if "source" in name:
+                l.append(k)
+        return l
+
+    def forward(self, input, adj1_target=None, adj2_source=None, adj3_transfer=None):
+        x, low_level_features = self.xception_features(input)
+        # print(x.size())
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.upsample(x5, size=x4.size()[2:], mode="bilinear", align_corners=True)
+
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.concat_projection_conv1(x)
+        x = self.concat_projection_bn1(x)
+        x = self.relu(x)
+        # print(x.size())
+        x = F.upsample(
+            x, size=low_level_features.size()[2:], mode="bilinear", align_corners=True
+        )
+
+        low_level_features = self.feature_projection_conv1(low_level_features)
+        low_level_features = self.feature_projection_bn1(low_level_features)
+        low_level_features = self.relu(low_level_features)
+        # print(low_level_features.size())
+        # print(x.size())
+        x = torch.cat((x, low_level_features), dim=1)
+        x = self.decoder(x)
+
+        ### add graph
+
+        # target graph
+        # print('x size',x.size(),adj1.size())
+        graph = self.target_featuremap_2_graph(x)
+
+        # graph combine
+        # print(graph.size(),source_2_target_graph.size())
+        # graph = self.fc_graph.forward(graph,relu=True)
+        # print(graph.size())
+
+        graph = self.target_graph_conv1.forward(graph, adj=adj1_target, relu=True)
+        graph = self.target_graph_conv2.forward(graph, adj=adj1_target, relu=True)
+        graph = self.target_graph_conv3.forward(graph, adj=adj1_target, relu=True)
+        # print(graph.size(),x.size())
+        # graph = self.gcn_encode.forward(graph,relu=True)
+        # graph = self.graph_conv2.forward(graph,adj=adj2,relu=True)
+        # graph = self.gcn_decode.forward(graph,relu=True)
+        graph = self.target_graph_2_fea.forward(graph, x)
+        x = self.target_skip_conv(x)
+        x = x + graph
+
+        ###
+        x = self.semantic(x)
+        x = F.upsample(x, size=input.size()[2:], mode="bilinear", align_corners=True)
+
+        return x
+
+
+#######################
+# transfer model
+#######################
+
+
+class deeplab_xception_transfer_projection(deeplab_xception_transfer_basemodel):
+    def __init__(
+        self,
+        nInputChannels=3,
+        n_classes=7,
+        os=16,
+        input_channels=256,
+        hidden_layers=128,
+        out_channels=256,
+        transfer_graph=None,
+        source_classes=20,
+    ):
+        super(deeplab_xception_transfer_projection, self).__init__(
+            nInputChannels=nInputChannels,
+            n_classes=n_classes,
+            os=os,
+            input_channels=input_channels,
+            hidden_layers=hidden_layers,
+            out_channels=out_channels,
+        )
+        self.source_featuremap_2_graph = gcn.Featuremaps_to_Graph(
+            input_channels=input_channels,
+            hidden_layers=hidden_layers,
+            nodes=source_classes,
+        )
+        self.source_graph_conv1 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.source_graph_conv2 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.source_graph_conv3 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.transpose_graph = gcn.Graph_trans(
+            in_features=hidden_layers,
+            out_features=hidden_layers,
+            adj=transfer_graph,
+            begin_nodes=source_classes,
+            end_nodes=n_classes,
+        )
+        self.fc_graph = gcn.GraphConvolution(hidden_layers * 3, hidden_layers)
+
+    def forward(self, input, adj1_target=None, adj2_source=None, adj3_transfer=None):
+        x, low_level_features = self.xception_features(input)
+        # print(x.size())
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.upsample(x5, size=x4.size()[2:], mode="bilinear", align_corners=True)
+
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.concat_projection_conv1(x)
+        x = self.concat_projection_bn1(x)
+        x = self.relu(x)
+        # print(x.size())
+        x = F.upsample(
+            x, size=low_level_features.size()[2:], mode="bilinear", align_corners=True
+        )
+
+        low_level_features = self.feature_projection_conv1(low_level_features)
+        low_level_features = self.feature_projection_bn1(low_level_features)
+        low_level_features = self.relu(low_level_features)
+        # print(low_level_features.size())
+        # print(x.size())
+        x = torch.cat((x, low_level_features), dim=1)
+        x = self.decoder(x)
+
+        ### add graph
+        # source graph
+        source_graph = self.source_featuremap_2_graph(x)
+        source_graph1 = self.source_graph_conv1.forward(
+            source_graph, adj=adj2_source, relu=True
+        )
+        source_graph2 = self.source_graph_conv2.forward(
+            source_graph1, adj=adj2_source, relu=True
+        )
+        source_graph3 = self.source_graph_conv2.forward(
+            source_graph2, adj=adj2_source, relu=True
+        )
+
+        source_2_target_graph1_v5 = self.transpose_graph.forward(
+            source_graph1, adj=adj3_transfer, relu=True
+        )
+        source_2_target_graph2_v5 = self.transpose_graph.forward(
+            source_graph2, adj=adj3_transfer, relu=True
+        )
+        source_2_target_graph3_v5 = self.transpose_graph.forward(
+            source_graph3, adj=adj3_transfer, relu=True
+        )
+
+        # target graph
+        # print('x size',x.size(),adj1.size())
+        graph = self.target_featuremap_2_graph(x)
+
+        source_2_target_graph1 = self.similarity_trans(source_graph1, graph)
+        # graph combine 1
+        # print(graph.size())
+        # print(source_2_target_graph1.size())
+        # print(source_2_target_graph1_v5.size())
+        graph = torch.cat(
+            (
+                graph,
+                source_2_target_graph1.squeeze(0),
+                source_2_target_graph1_v5.squeeze(0),
+            ),
+            dim=-1,
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv1.forward(graph, adj=adj1_target, relu=True)
+
+        source_2_target_graph2 = self.similarity_trans(source_graph2, graph)
+        # graph combine 2
+        graph = torch.cat(
+            (graph, source_2_target_graph2, source_2_target_graph2_v5), dim=-1
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv2.forward(graph, adj=adj1_target, relu=True)
+
+        source_2_target_graph3 = self.similarity_trans(source_graph3, graph)
+        # graph combine 3
+        graph = torch.cat(
+            (graph, source_2_target_graph3, source_2_target_graph3_v5), dim=-1
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv3.forward(graph, adj=adj1_target, relu=True)
+
+        # print(graph.size(),x.size())
+
+        graph = self.target_graph_2_fea.forward(graph, x)
+        x = self.target_skip_conv(x)
+        x = x + graph
+
+        ###
+        x = self.semantic(x)
+        x = F.upsample(x, size=input.size()[2:], mode="bilinear", align_corners=True)
+
+        return x
+
+    def similarity_trans(self, source, target):
+        sim = torch.matmul(
+            F.normalize(target, p=2, dim=-1),
+            F.normalize(source, p=2, dim=-1).transpose(-1, -2),
+        )
+        sim = F.softmax(sim, dim=-1)
+        return torch.matmul(sim, source)
+
+    def load_source_model(self, state_dict):
+        own_state = self.state_dict()
+        # for name inshop_cos own_state:
+        #    print name
+        new_state_dict = OrderedDict()
+        for name, param in state_dict.items():
+            name = name.replace("module.", "")
+
+            if (
+                "graph" in name
+                and "source" not in name
+                and "target" not in name
+                and "fc_" not in name
+                and "transpose_graph" not in name
+            ):
+                if "featuremap_2_graph" in name:
+                    name = name.replace(
+                        "featuremap_2_graph", "source_featuremap_2_graph"
+                    )
+                else:
+                    name = name.replace("graph", "source_graph")
+            new_state_dict[name] = 0
+            if name not in own_state:
+                if "num_batch" in name:
+                    continue
+                print('unexpected key "{}" in state_dict'.format(name))
+                continue
+                # if isinstance(param, own_state):
+            if isinstance(param, Parameter):
+                # backwards compatibility for serialized parameters
+                param = param.data
+            try:
+                own_state[name].copy_(param)
+            except:
+                print(
+                    "While copying the parameter named {}, whose dimensions in the model are"
+                    " {} and whose dimensions in the checkpoint are {}, ...".format(
+                        name, own_state[name].size(), param.size()
+                    )
+                )
+                continue  # i add inshop_cos 2018/02/01
+            own_state[name].copy_(param)
+            # print 'copying %s' %name
+
+        missing = set(own_state.keys()) - set(new_state_dict.keys())
+        if len(missing) > 0:
+            print('missing keys in state_dict: "{}"'.format(missing))
+
+
+class deeplab_xception_transfer_projection_savemem(
+    deeplab_xception_transfer_basemodel_savememory
+):
+    def __init__(
+        self,
+        nInputChannels=3,
+        n_classes=7,
+        os=16,
+        input_channels=256,
+        hidden_layers=128,
+        out_channels=256,
+        transfer_graph=None,
+        source_classes=20,
+    ):
+        super(deeplab_xception_transfer_projection_savemem, self).__init__(
+            nInputChannels=nInputChannels,
+            n_classes=n_classes,
+            os=os,
+            input_channels=input_channels,
+            hidden_layers=hidden_layers,
+            out_channels=out_channels,
+        )
+        self.source_featuremap_2_graph = gcn.Featuremaps_to_Graph(
+            input_channels=input_channels,
+            hidden_layers=hidden_layers,
+            nodes=source_classes,
+        )
+        self.source_graph_conv1 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.source_graph_conv2 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.source_graph_conv3 = gcn.GraphConvolution(hidden_layers, hidden_layers)
+        self.transpose_graph = gcn.Graph_trans(
+            in_features=hidden_layers,
+            out_features=hidden_layers,
+            adj=transfer_graph,
+            begin_nodes=source_classes,
+            end_nodes=n_classes,
+        )
+        self.fc_graph = gcn.GraphConvolution(hidden_layers * 3, hidden_layers)
+
+    def forward(self, input, adj1_target=None, adj2_source=None, adj3_transfer=None):
+        x, low_level_features = self.xception_features(input)
+        # print(x.size())
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.upsample(x5, size=x4.size()[2:], mode="bilinear", align_corners=True)
+
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.concat_projection_conv1(x)
+        x = self.concat_projection_bn1(x)
+        x = self.relu(x)
+        # print(x.size())
+        x = F.upsample(
+            x, size=low_level_features.size()[2:], mode="bilinear", align_corners=True
+        )
+
+        low_level_features = self.feature_projection_conv1(low_level_features)
+        low_level_features = self.feature_projection_bn1(low_level_features)
+        low_level_features = self.relu(low_level_features)
+        # print(low_level_features.size())
+        # print(x.size())
+        x = torch.cat((x, low_level_features), dim=1)
+        x = self.decoder(x)
+
+        ### add graph
+        # source graph
+        source_graph = self.source_featuremap_2_graph(x)
+        source_graph1 = self.source_graph_conv1.forward(
+            source_graph, adj=adj2_source, relu=True
+        )
+        source_graph2 = self.source_graph_conv2.forward(
+            source_graph1, adj=adj2_source, relu=True
+        )
+        source_graph3 = self.source_graph_conv2.forward(
+            source_graph2, adj=adj2_source, relu=True
+        )
+
+        source_2_target_graph1_v5 = self.transpose_graph.forward(
+            source_graph1, adj=adj3_transfer, relu=True
+        )
+        source_2_target_graph2_v5 = self.transpose_graph.forward(
+            source_graph2, adj=adj3_transfer, relu=True
+        )
+        source_2_target_graph3_v5 = self.transpose_graph.forward(
+            source_graph3, adj=adj3_transfer, relu=True
+        )
+
+        # target graph
+        # print('x size',x.size(),adj1.size())
+        graph = self.target_featuremap_2_graph(x)
+
+        source_2_target_graph1 = self.similarity_trans(source_graph1, graph)
+        # graph combine 1
+        graph = torch.cat(
+            (
+                graph,
+                source_2_target_graph1.squeeze(0),
+                source_2_target_graph1_v5.squeeze(0),
+            ),
+            dim=-1,
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv1.forward(graph, adj=adj1_target, relu=True)
+
+        source_2_target_graph2 = self.similarity_trans(source_graph2, graph)
+        # graph combine 2
+        graph = torch.cat(
+            (graph, source_2_target_graph2, source_2_target_graph2_v5), dim=-1
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv2.forward(graph, adj=adj1_target, relu=True)
+
+        source_2_target_graph3 = self.similarity_trans(source_graph3, graph)
+        # graph combine 3
+        graph = torch.cat(
+            (graph, source_2_target_graph3, source_2_target_graph3_v5), dim=-1
+        )
+        graph = self.fc_graph.forward(graph, relu=True)
+
+        graph = self.target_graph_conv3.forward(graph, adj=adj1_target, relu=True)
+
+        # print(graph.size(),x.size())
+
+        graph = self.target_graph_2_fea.forward(graph, x)
+        x = self.target_skip_conv(x)
+        x = x + graph
+
+        ###
+        x = self.semantic(x)
+        x = F.upsample(x, size=input.size()[2:], mode="bilinear", align_corners=True)
+
+        return x
+
+    def similarity_trans(self, source, target):
+        sim = torch.matmul(
+            F.normalize(target, p=2, dim=-1),
+            F.normalize(source, p=2, dim=-1).transpose(-1, -2),
+        )
+        sim = F.softmax(sim, dim=-1)
+        return torch.matmul(sim, source)
+
+    def load_source_model(self, state_dict):
+        own_state = self.state_dict()
+        # for name inshop_cos own_state:
+        #    print name
+        new_state_dict = OrderedDict()
+        for name, param in state_dict.items():
+            name = name.replace("module.", "")
+
+            if (
+                "graph" in name
+                and "source" not in name
+                and "target" not in name
+                and "fc_" not in name
+                and "transpose_graph" not in name
+            ):
+                if "featuremap_2_graph" in name:
+                    name = name.replace(
+                        "featuremap_2_graph", "source_featuremap_2_graph"
+                    )
+                else:
+                    name = name.replace("graph", "source_graph")
+            new_state_dict[name] = 0
+            if name not in own_state:
+                if "num_batch" in name:
+                    continue
+                print('unexpected key "{}" in state_dict'.format(name))
+                continue
+                # if isinstance(param, own_state):
+            if isinstance(param, Parameter):
+                # backwards compatibility for serialized parameters
+                param = param.data
+            try:
+                own_state[name].copy_(param)
+            except:
+                print(
+                    "While copying the parameter named {}, whose dimensions in the model are"
+                    " {} and whose dimensions in the checkpoint are {}, ...".format(
+                        name, own_state[name].size(), param.size()
+                    )
+                )
+                continue  # i add inshop_cos 2018/02/01
+            own_state[name].copy_(param)
+            # print 'copying %s' %name
+
+        missing = set(own_state.keys()) - set(new_state_dict.keys())
+        if len(missing) > 0:
+            print('missing keys in state_dict: "{}"'.format(missing))
+
+
+# if __name__ == '__main__':
+# net = deeplab_xception_transfer_projection_v3v5_more_savemem()
+# img = torch.rand((2,3,128,128))
+# net.eval()
+# a = torch.rand((1,1,7,7))
+# net.forward(img, adj1_target=a)
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/gcn.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/gcn.py
new file mode 100644
index 000000000..e5eb0661b
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/gcn.py
@@ -0,0 +1,304 @@
+import torch
+from . import graph
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.parameter import Parameter
+
+class GraphConvolution(nn.Module):
+    def __init__(self, in_features, out_features, bias=False):
+        super(GraphConvolution, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight = Parameter(torch.FloatTensor(in_features, out_features))
+        if bias:
+            self.bias = Parameter(torch.FloatTensor(out_features))
+        else:
+            self.register_parameter("bias", None)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        # stdv = 1./math.sqrt(self.weight(1))
+        # self.weight.data.uniform_(-stdv,stdv)
+        torch.nn.init.xavier_uniform_(self.weight)
+        # if self.bias is not None:
+        #     self.bias.data.uniform_(-stdv,stdv)
+
+    def forward(self, input, adj=None, relu=False):
+        support = torch.matmul(input, self.weight)
+        # print(support.size(),adj.size())
+        if adj is not None:
+            output = torch.matmul(adj, support)
+        else:
+            output = support
+        # print(output.size())
+        if self.bias is not None:
+            return output + self.bias
+        else:
+            if relu:
+                return F.relu(output)
+            else:
+                return output
+
+    def __repr__(self):
+        return (
+            self.__class__.__name__
+            + " ("
+            + str(self.in_features)
+            + " -> "
+            + str(self.out_features)
+            + ")"
+        )
+
+
+class Featuremaps_to_Graph(nn.Module):
+    def __init__(self, input_channels, hidden_layers, nodes=7):
+        super(Featuremaps_to_Graph, self).__init__()
+        self.pre_fea = Parameter(torch.FloatTensor(input_channels, nodes))
+        self.weight = Parameter(torch.FloatTensor(input_channels, hidden_layers))
+        self.reset_parameters()
+
+    def forward(self, input):
+        n, c, h, w = input.size()
+        # print('fea input',input.size())
+        input1 = input.view(n, c, h * w)
+        input1 = input1.transpose(1, 2)  # n x hw x c
+        # print('fea input1', input1.size())
+        ############## Feature maps to node ################
+        fea_node = torch.matmul(input1, self.pre_fea)  # n x hw x n_classes
+        weight_node = torch.matmul(input1, self.weight)  # n x hw x hidden_layer
+        # softmax fea_node
+        fea_node = F.softmax(fea_node, dim=-1)
+        # print(fea_node.size(),weight_node.size())
+        graph_node = F.relu(torch.matmul(fea_node.transpose(1, 2), weight_node))
+        return graph_node  # n x n_class x hidden_layer
+
+    def reset_parameters(self):
+        for ww in self.parameters():
+            torch.nn.init.xavier_uniform_(ww)
+        # if self.bias is not None:
+        #     self.bias.data.uniform_(-stdv,stdv)
+
+
+class Featuremaps_to_Graph_transfer(nn.Module):
+    def __init__(self, input_channels, hidden_layers, nodes=7, source_nodes=20):
+        super(Featuremaps_to_Graph_transfer, self).__init__()
+        self.pre_fea = Parameter(torch.FloatTensor(input_channels, nodes))
+        self.weight = Parameter(torch.FloatTensor(input_channels, hidden_layers))
+        self.pre_fea_transfer = nn.Sequential(
+            *[
+                nn.Linear(source_nodes, source_nodes),
+                nn.LeakyReLU(True),
+                nn.Linear(source_nodes, nodes),
+                nn.LeakyReLU(True),
+            ]
+        )
+        self.reset_parameters()
+
+    def forward(self, input, source_pre_fea):
+        self.pre_fea.data = self.pre_fea_learn(source_pre_fea)
+        n, c, h, w = input.size()
+        # print('fea input',input.size())
+        input1 = input.view(n, c, h * w)
+        input1 = input1.transpose(1, 2)  # n x hw x c
+        # print('fea input1', input1.size())
+        ############## Feature maps to node ################
+        fea_node = torch.matmul(input1, self.pre_fea)  # n x hw x n_classes
+        weight_node = torch.matmul(input1, self.weight)  # n x hw x hidden_layer
+        # softmax fea_node
+        fea_node = F.softmax(fea_node, dim=1)
+        # print(fea_node.size(),weight_node.size())
+        graph_node = F.relu(torch.matmul(fea_node.transpose(1, 2), weight_node))
+        return graph_node  # n x n_class x hidden_layer
+
+    def pre_fea_learn(self, input):
+        pre_fea = self.pre_fea_transfer.forward(input.unsqueeze(0)).squeeze(0)
+        return self.pre_fea.data + pre_fea
+
+
+class Graph_to_Featuremaps(nn.Module):
+    # this is a special version
+    def __init__(self, input_channels, output_channels, hidden_layers, nodes=7):
+        super(Graph_to_Featuremaps, self).__init__()
+        self.node_fea = Parameter(torch.FloatTensor(input_channels + hidden_layers, 1))
+        self.weight = Parameter(torch.FloatTensor(hidden_layers, output_channels))
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        for ww in self.parameters():
+            torch.nn.init.xavier_uniform_(ww)
+
+    def forward(self, input, res_feature):
+        """
+
+        :param input: 1 x batch x nodes x hidden_layer
+        :param res_feature: batch x channels x h x w
+        :return:
+        """
+        batchi, channeli, hi, wi = res_feature.size()
+        # print(res_feature.size())
+        # print(input.size())
+        try:
+            _, batch, nodes, hidden = input.size()
+        except:
+            # print(input.size())
+            input = input.unsqueeze(0)
+            _, batch, nodes, hidden = input.size()
+
+        assert batch == batchi
+        input1 = input.transpose(0, 1).expand(batch, hi * wi, nodes, hidden)
+        res_feature_after_view = res_feature.view(batch, channeli, hi * wi).transpose(
+            1, 2
+        )
+        res_feature_after_view1 = res_feature_after_view.unsqueeze(2).expand(
+            batch, hi * wi, nodes, channeli
+        )
+        new_fea = torch.cat((res_feature_after_view1, input1), dim=3)
+
+        # print(self.node_fea.size(),new_fea.size())
+        new_node = torch.matmul(new_fea, self.node_fea)  # batch x hw x nodes x 1
+        new_weight = torch.matmul(input, self.weight)  # batch x node x channel
+        new_node = new_node.view(batch, hi * wi, nodes)
+        # 0721
+        new_node = F.softmax(new_node, dim=-1)
+        #
+        feature_out = torch.matmul(new_node, new_weight)
+        # print(feature_out.size())
+        feature_out = feature_out.transpose(2, 3).contiguous().view(res_feature.size())
+        return F.relu(feature_out)
+
+
+class Graph_to_Featuremaps_savemem(nn.Module):
+    # this is a special version for saving gpu memory. The process is same as Graph_to_Featuremaps.
+    def __init__(self, input_channels, output_channels, hidden_layers, nodes=7):
+        super(Graph_to_Featuremaps_savemem, self).__init__()
+        self.node_fea_for_res = Parameter(torch.FloatTensor(input_channels, 1))
+        self.node_fea_for_hidden = Parameter(torch.FloatTensor(hidden_layers, 1))
+        self.weight = Parameter(torch.FloatTensor(hidden_layers, output_channels))
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        for ww in self.parameters():
+            torch.nn.init.xavier_uniform_(ww)
+
+    def forward(self, input, res_feature):
+        """
+
+        :param input: 1 x batch x nodes x hidden_layer
+        :param res_feature: batch x channels x h x w
+        :return:
+        """
+        batchi, channeli, hi, wi = res_feature.size()
+        # print(res_feature.size())
+        # print(input.size())
+        try:
+            _, batch, nodes, hidden = input.size()
+        except:
+            # print(input.size())
+            input = input.unsqueeze(0)
+            _, batch, nodes, hidden = input.size()
+
+        assert batch == batchi
+        input1 = input.transpose(0, 1).expand(batch, hi * wi, nodes, hidden)
+        res_feature_after_view = res_feature.view(batch, channeli, hi * wi).transpose(
+            1, 2
+        )
+        res_feature_after_view1 = res_feature_after_view.unsqueeze(2).expand(
+            batch, hi * wi, nodes, channeli
+        )
+        # new_fea = torch.cat((res_feature_after_view1,input1),dim=3)
+        ## sim
+        new_node1 = torch.matmul(res_feature_after_view1, self.node_fea_for_res)
+        new_node2 = torch.matmul(input1, self.node_fea_for_hidden)
+        new_node = new_node1 + new_node2
+        ## sim end
+        # print(self.node_fea.size(),new_fea.size())
+        # new_node = torch.matmul(new_fea, self.node_fea) # batch x hw x nodes x 1
+        new_weight = torch.matmul(input, self.weight)  # batch x node x channel
+        new_node = new_node.view(batch, hi * wi, nodes)
+        # 0721
+        new_node = F.softmax(new_node, dim=-1)
+        #
+        feature_out = torch.matmul(new_node, new_weight)
+        # print(feature_out.size())
+        feature_out = feature_out.transpose(2, 3).contiguous().view(res_feature.size())
+        return F.relu(feature_out)
+
+
+class Graph_trans(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        begin_nodes=7,
+        end_nodes=2,
+        bias=False,
+        adj=None,
+    ):
+        super(Graph_trans, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight = Parameter(torch.FloatTensor(in_features, out_features))
+        if adj is not None:
+            h, w = adj.size()
+            assert (h == end_nodes) and (w == begin_nodes)
+            self.adj = torch.autograd.Variable(adj, requires_grad=False)
+        else:
+            self.adj = Parameter(torch.FloatTensor(end_nodes, begin_nodes))
+        if bias:
+            self.bias = Parameter(torch.FloatTensor(out_features))
+        else:
+            self.register_parameter("bias", None)
+        # self.reset_parameters()
+
+    def reset_parameters(self):
+        # stdv = 1./math.sqrt(self.weight(1))
+        # self.weight.data.uniform_(-stdv,stdv)
+        torch.nn.init.xavier_uniform_(self.weight)
+        # if self.bias is not None:
+        #     self.bias.data.uniform_(-stdv,stdv)
+
+    def forward(self, input, relu=False, adj_return=False, adj=None):
+        support = torch.matmul(input, self.weight)
+        # print(support.size(),self.adj.size())
+        if adj is None:
+            adj = self.adj
+        adj1 = self.norm_trans_adj(adj)
+        output = torch.matmul(adj1, support)
+        if adj_return:
+            output1 = F.normalize(output, p=2, dim=-1)
+            self.adj_mat = torch.matmul(output1, output1.transpose(-2, -1))
+        if self.bias is not None:
+            return output + self.bias
+        else:
+            if relu:
+                return F.relu(output)
+            else:
+                return output
+
+    def get_adj_mat(self):
+        adj = graph.normalize_adj_torch(F.relu(self.adj_mat))
+        return adj
+
+    def get_encode_adj(self):
+        return self.adj
+
+    def norm_trans_adj(self, adj):  # maybe can use softmax
+        adj = F.relu(adj)
+        r = F.softmax(adj, dim=-1)
+        # print(adj.size())
+        # row_sum = adj.sum(-1).unsqueeze(-1)
+        # d_mat = row_sum.expand(adj.size())
+        # r = torch.div(row_sum,d_mat)
+        # r[torch.isnan(r)] = 0
+
+        return r
+
+
+if __name__ == "__main__":
+
+    graph = torch.randn((7, 128))
+    en = GraphConvolution(128, 128)
+    a = en.forward(graph)
+    print(a)
+    # a = en.forward(graph,pred)
+    # print(a.size())
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/graph.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/graph.py
new file mode 100644
index 000000000..06c04e502
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/graph.py
@@ -0,0 +1,839 @@
+import torch
+import numpy as np
+import networkx as nx
+import scipy.sparse as sp
+
+pascal_graph = {
+    0: [0],
+    1: [1, 2],
+    2: [1, 2, 3, 5],
+    3: [2, 3, 4],
+    4: [3, 4],
+    5: [2, 5, 6],
+    6: [5, 6],
+}
+
+cihp_graph = {
+    0: [],
+    1: [2, 13],
+    2: [1, 13],
+    3: [14, 15],
+    4: [13],
+    5: [6, 7, 9, 10, 11, 12, 14, 15],
+    6: [5, 7, 10, 11, 14, 15, 16, 17],
+    7: [5, 6, 9, 10, 11, 12, 14, 15],
+    8: [16, 17, 18, 19],
+    9: [5, 7, 10, 16, 17, 18, 19],
+    10: [5, 6, 7, 9, 11, 12, 13, 14, 15, 16, 17],
+    11: [5, 6, 7, 10, 13],
+    12: [5, 7, 10, 16, 17],
+    13: [1, 2, 4, 10, 11],
+    14: [3, 5, 6, 7, 10],
+    15: [3, 5, 6, 7, 10],
+    16: [6, 8, 9, 10, 12, 18],
+    17: [6, 8, 9, 10, 12, 19],
+    18: [8, 9, 16],
+    19: [8, 9, 17],
+}
+
+atr_graph = {
+    0: [],
+    1: [2, 11],
+    2: [1, 11],
+    3: [11],
+    4: [5, 6, 7, 11, 14, 15, 17],
+    5: [4, 6, 7, 8, 12, 13],
+    6: [4, 5, 7, 8, 9, 10, 12, 13],
+    7: [4, 11, 12, 13, 14, 15],
+    8: [5, 6],
+    9: [6, 12],
+    10: [6, 13],
+    11: [1, 2, 3, 4, 7, 14, 15, 17],
+    12: [5, 6, 7, 9],
+    13: [5, 6, 7, 10],
+    14: [4, 7, 11, 16],
+    15: [4, 7, 11, 16],
+    16: [14, 15],
+    17: [4, 11],
+}
+
+cihp2pascal_adj = np.array(
+    [
+        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+        [0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0],
+        [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
+        [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0],
+        [0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0],
+        [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
+        [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1],
+    ]
+)
+
+cihp2pascal_nlp_adj = np.array(
+    [
+        [
+            1.0,
+            0.35333052,
+            0.32727194,
+            0.17418084,
+            0.18757584,
+            0.40608522,
+            0.37503981,
+            0.35448462,
+            0.22598555,
+            0.23893579,
+            0.33064262,
+            0.28923404,
+            0.27986573,
+            0.4211553,
+            0.36915778,
+            0.41377746,
+            0.32485771,
+            0.37248222,
+            0.36865639,
+            0.41500332,
+        ],
+        [
+            0.39615879,
+            0.46201529,
+            0.52321467,
+            0.30826114,
+            0.25669527,
+            0.54747773,
+            0.3670523,
+            0.3901983,
+            0.27519473,
+            0.3433325,
+            0.52728509,
+            0.32771333,
+            0.34819325,
+            0.63882953,
+            0.68042925,
+            0.69368576,
+            0.63395791,
+            0.65344337,
+            0.59538781,
+            0.6071375,
+        ],
+        [
+            0.16373166,
+            0.21663339,
+            0.3053872,
+            0.28377612,
+            0.1372435,
+            0.4448808,
+            0.29479995,
+            0.31092595,
+            0.22703953,
+            0.33983576,
+            0.75778818,
+            0.2619818,
+            0.37069392,
+            0.35184867,
+            0.49877512,
+            0.49979437,
+            0.51853277,
+            0.52517541,
+            0.32517741,
+            0.32377309,
+        ],
+        [
+            0.32687232,
+            0.38482461,
+            0.37693463,
+            0.41610834,
+            0.20415749,
+            0.76749079,
+            0.35139853,
+            0.3787411,
+            0.28411737,
+            0.35155421,
+            0.58792618,
+            0.31141718,
+            0.40585111,
+            0.51189218,
+            0.82042737,
+            0.8342413,
+            0.70732188,
+            0.72752501,
+            0.60327325,
+            0.61431337,
+        ],
+        [
+            0.34069369,
+            0.34817292,
+            0.37525998,
+            0.36497069,
+            0.17841617,
+            0.69746208,
+            0.31731463,
+            0.34628951,
+            0.25167277,
+            0.32072379,
+            0.56711286,
+            0.24894776,
+            0.37000453,
+            0.52600859,
+            0.82483993,
+            0.84966274,
+            0.7033991,
+            0.73449378,
+            0.56649608,
+            0.58888791,
+        ],
+        [
+            0.28477487,
+            0.35139564,
+            0.42742352,
+            0.41664321,
+            0.20004676,
+            0.78566833,
+            0.42237487,
+            0.41048549,
+            0.37933812,
+            0.46542516,
+            0.62444759,
+            0.3274493,
+            0.49466009,
+            0.49314658,
+            0.71244233,
+            0.71497003,
+            0.8234787,
+            0.83566589,
+            0.62597135,
+            0.62626812,
+        ],
+        [
+            0.3011378,
+            0.31775977,
+            0.42922647,
+            0.36896257,
+            0.17597556,
+            0.72214655,
+            0.39162804,
+            0.38137872,
+            0.34980296,
+            0.43818419,
+            0.60879174,
+            0.26762545,
+            0.46271161,
+            0.51150476,
+            0.72318109,
+            0.73678399,
+            0.82620388,
+            0.84942166,
+            0.5943811,
+            0.60607602,
+        ],
+    ]
+)
+
+pascal2atr_nlp_adj = np.array(
+    [
+        [
+            1.0,
+            0.35333052,
+            0.32727194,
+            0.18757584,
+            0.40608522,
+            0.27986573,
+            0.23893579,
+            0.27600672,
+            0.30964391,
+            0.36865639,
+            0.41500332,
+            0.4211553,
+            0.32485771,
+            0.37248222,
+            0.36915778,
+            0.41377746,
+            0.32006291,
+            0.28923404,
+        ],
+        [
+            0.39615879,
+            0.46201529,
+            0.52321467,
+            0.25669527,
+            0.54747773,
+            0.34819325,
+            0.3433325,
+            0.26603942,
+            0.45162929,
+            0.59538781,
+            0.6071375,
+            0.63882953,
+            0.63395791,
+            0.65344337,
+            0.68042925,
+            0.69368576,
+            0.44354613,
+            0.32771333,
+        ],
+        [
+            0.16373166,
+            0.21663339,
+            0.3053872,
+            0.1372435,
+            0.4448808,
+            0.37069392,
+            0.33983576,
+            0.26563416,
+            0.35443504,
+            0.32517741,
+            0.32377309,
+            0.35184867,
+            0.51853277,
+            0.52517541,
+            0.49877512,
+            0.49979437,
+            0.21750868,
+            0.2619818,
+        ],
+        [
+            0.32687232,
+            0.38482461,
+            0.37693463,
+            0.20415749,
+            0.76749079,
+            0.40585111,
+            0.35155421,
+            0.28271333,
+            0.52684576,
+            0.60327325,
+            0.61431337,
+            0.51189218,
+            0.70732188,
+            0.72752501,
+            0.82042737,
+            0.8342413,
+            0.40137029,
+            0.31141718,
+        ],
+        [
+            0.34069369,
+            0.34817292,
+            0.37525998,
+            0.17841617,
+            0.69746208,
+            0.37000453,
+            0.32072379,
+            0.27268885,
+            0.47426719,
+            0.56649608,
+            0.58888791,
+            0.52600859,
+            0.7033991,
+            0.73449378,
+            0.82483993,
+            0.84966274,
+            0.37830796,
+            0.24894776,
+        ],
+        [
+            0.28477487,
+            0.35139564,
+            0.42742352,
+            0.20004676,
+            0.78566833,
+            0.49466009,
+            0.46542516,
+            0.32662614,
+            0.55780359,
+            0.62597135,
+            0.62626812,
+            0.49314658,
+            0.8234787,
+            0.83566589,
+            0.71244233,
+            0.71497003,
+            0.41223219,
+            0.3274493,
+        ],
+        [
+            0.3011378,
+            0.31775977,
+            0.42922647,
+            0.17597556,
+            0.72214655,
+            0.46271161,
+            0.43818419,
+            0.3192333,
+            0.50979216,
+            0.5943811,
+            0.60607602,
+            0.51150476,
+            0.82620388,
+            0.84942166,
+            0.72318109,
+            0.73678399,
+            0.39259827,
+            0.26762545,
+        ],
+    ]
+)
+
+cihp2atr_nlp_adj = np.array(
+    [
+        [
+            1.0,
+            0.35333052,
+            0.32727194,
+            0.18757584,
+            0.40608522,
+            0.27986573,
+            0.23893579,
+            0.27600672,
+            0.30964391,
+            0.36865639,
+            0.41500332,
+            0.4211553,
+            0.32485771,
+            0.37248222,
+            0.36915778,
+            0.41377746,
+            0.32006291,
+            0.28923404,
+        ],
+        [
+            0.35333052,
+            1.0,
+            0.39206695,
+            0.42143438,
+            0.4736689,
+            0.47139544,
+            0.51999208,
+            0.38354847,
+            0.45628529,
+            0.46514124,
+            0.50083501,
+            0.4310595,
+            0.39371443,
+            0.4319752,
+            0.42938598,
+            0.46384034,
+            0.44833757,
+            0.6153155,
+        ],
+        [
+            0.32727194,
+            0.39206695,
+            1.0,
+            0.32836702,
+            0.52603065,
+            0.39543695,
+            0.3622627,
+            0.43575346,
+            0.33866223,
+            0.45202552,
+            0.48421,
+            0.53669903,
+            0.47266611,
+            0.50925436,
+            0.42286557,
+            0.45403656,
+            0.37221304,
+            0.40999322,
+        ],
+        [
+            0.17418084,
+            0.46892601,
+            0.25774838,
+            0.31816231,
+            0.39330317,
+            0.34218382,
+            0.48253904,
+            0.22084125,
+            0.41335728,
+            0.52437572,
+            0.5191713,
+            0.33576117,
+            0.44230914,
+            0.44250678,
+            0.44330833,
+            0.43887264,
+            0.50693611,
+            0.39278795,
+        ],
+        [
+            0.18757584,
+            0.42143438,
+            0.32836702,
+            1.0,
+            0.35030067,
+            0.30110947,
+            0.41055555,
+            0.34338879,
+            0.34336307,
+            0.37704433,
+            0.38810141,
+            0.34702081,
+            0.24171562,
+            0.25433078,
+            0.24696241,
+            0.2570884,
+            0.4465962,
+            0.45263213,
+        ],
+        [
+            0.40608522,
+            0.4736689,
+            0.52603065,
+            0.35030067,
+            1.0,
+            0.54372584,
+            0.58300258,
+            0.56674191,
+            0.555266,
+            0.66599594,
+            0.68567555,
+            0.55716359,
+            0.62997328,
+            0.65638548,
+            0.61219615,
+            0.63183318,
+            0.54464151,
+            0.44293752,
+        ],
+        [
+            0.37503981,
+            0.50675565,
+            0.4761106,
+            0.37561813,
+            0.60419403,
+            0.77912403,
+            0.64595517,
+            0.85939662,
+            0.46037144,
+            0.52348817,
+            0.55875094,
+            0.37741886,
+            0.455671,
+            0.49434392,
+            0.38479954,
+            0.41804074,
+            0.47285709,
+            0.57236283,
+        ],
+        [
+            0.35448462,
+            0.50576632,
+            0.51030446,
+            0.35841033,
+            0.55106903,
+            0.50257274,
+            0.52591451,
+            0.4283053,
+            0.39991808,
+            0.42327211,
+            0.42853819,
+            0.42071825,
+            0.41240559,
+            0.42259136,
+            0.38125352,
+            0.3868255,
+            0.47604934,
+            0.51811717,
+        ],
+        [
+            0.22598555,
+            0.5053299,
+            0.36301185,
+            0.38002282,
+            0.49700941,
+            0.45625243,
+            0.62876479,
+            0.4112051,
+            0.33944371,
+            0.48322639,
+            0.50318714,
+            0.29207815,
+            0.38801966,
+            0.41119094,
+            0.29199072,
+            0.31021029,
+            0.41594871,
+            0.54961962,
+        ],
+        [
+            0.23893579,
+            0.51999208,
+            0.3622627,
+            0.41055555,
+            0.58300258,
+            0.68874251,
+            1.0,
+            0.56977937,
+            0.49918447,
+            0.48484363,
+            0.51615925,
+            0.41222306,
+            0.49535971,
+            0.53134951,
+            0.3807616,
+            0.41050298,
+            0.48675801,
+            0.51112664,
+        ],
+        [
+            0.33064262,
+            0.306412,
+            0.60679935,
+            0.25592294,
+            0.58738706,
+            0.40379627,
+            0.39679161,
+            0.33618385,
+            0.39235148,
+            0.45474013,
+            0.4648476,
+            0.59306762,
+            0.58976007,
+            0.60778661,
+            0.55400397,
+            0.56551297,
+            0.3698029,
+            0.33860535,
+        ],
+        [
+            0.28923404,
+            0.6153155,
+            0.40999322,
+            0.45263213,
+            0.44293752,
+            0.60359359,
+            0.51112664,
+            0.46578181,
+            0.45656936,
+            0.38142307,
+            0.38525582,
+            0.33327223,
+            0.35360175,
+            0.36156453,
+            0.3384992,
+            0.34261229,
+            0.49297863,
+            1.0,
+        ],
+        [
+            0.27986573,
+            0.47139544,
+            0.39543695,
+            0.30110947,
+            0.54372584,
+            1.0,
+            0.68874251,
+            0.67765588,
+            0.48690078,
+            0.44010641,
+            0.44921156,
+            0.32321099,
+            0.48311542,
+            0.4982002,
+            0.39378102,
+            0.40297733,
+            0.45309735,
+            0.60359359,
+        ],
+        [
+            0.4211553,
+            0.4310595,
+            0.53669903,
+            0.34702081,
+            0.55716359,
+            0.32321099,
+            0.41222306,
+            0.25721705,
+            0.36633509,
+            0.5397475,
+            0.56429928,
+            1.0,
+            0.55796926,
+            0.58842844,
+            0.57930828,
+            0.60410597,
+            0.41615326,
+            0.33327223,
+        ],
+        [
+            0.36915778,
+            0.42938598,
+            0.42286557,
+            0.24696241,
+            0.61219615,
+            0.39378102,
+            0.3807616,
+            0.28089866,
+            0.48450394,
+            0.77400821,
+            0.68813814,
+            0.57930828,
+            0.8856886,
+            0.81673412,
+            1.0,
+            0.92279623,
+            0.46969152,
+            0.3384992,
+        ],
+        [
+            0.41377746,
+            0.46384034,
+            0.45403656,
+            0.2570884,
+            0.63183318,
+            0.40297733,
+            0.41050298,
+            0.332879,
+            0.48799542,
+            0.69231828,
+            0.77015091,
+            0.60410597,
+            0.79788484,
+            0.88232104,
+            0.92279623,
+            1.0,
+            0.45685017,
+            0.34261229,
+        ],
+        [
+            0.32485771,
+            0.39371443,
+            0.47266611,
+            0.24171562,
+            0.62997328,
+            0.48311542,
+            0.49535971,
+            0.32477932,
+            0.51486622,
+            0.79353556,
+            0.69768738,
+            0.55796926,
+            1.0,
+            0.92373745,
+            0.8856886,
+            0.79788484,
+            0.47883134,
+            0.35360175,
+        ],
+        [
+            0.37248222,
+            0.4319752,
+            0.50925436,
+            0.25433078,
+            0.65638548,
+            0.4982002,
+            0.53134951,
+            0.38057074,
+            0.52403969,
+            0.72035243,
+            0.78711147,
+            0.58842844,
+            0.92373745,
+            1.0,
+            0.81673412,
+            0.88232104,
+            0.47109935,
+            0.36156453,
+        ],
+        [
+            0.36865639,
+            0.46514124,
+            0.45202552,
+            0.37704433,
+            0.66599594,
+            0.44010641,
+            0.48484363,
+            0.39636574,
+            0.50175258,
+            1.0,
+            0.91320249,
+            0.5397475,
+            0.79353556,
+            0.72035243,
+            0.77400821,
+            0.69231828,
+            0.59087008,
+            0.38142307,
+        ],
+        [
+            0.41500332,
+            0.50083501,
+            0.48421,
+            0.38810141,
+            0.68567555,
+            0.44921156,
+            0.51615925,
+            0.45156472,
+            0.50438158,
+            0.91320249,
+            1.0,
+            0.56429928,
+            0.69768738,
+            0.78711147,
+            0.68813814,
+            0.77015091,
+            0.57698754,
+            0.38525582,
+        ],
+    ]
+)
+
+
+def normalize_adj(adj):
+    """Symmetrically normalize adjacency matrix."""
+    adj = sp.coo_matrix(adj)
+    rowsum = np.array(adj.sum(1))
+    d_inv_sqrt = np.power(rowsum, -0.5).flatten()
+    d_inv_sqrt[np.isinf(d_inv_sqrt)] = 0.0
+    d_mat_inv_sqrt = sp.diags(d_inv_sqrt)
+    return adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt).tocoo()
+
+
+def preprocess_adj(adj):
+    """Preprocessing of adjacency matrix for simple GCN model and conversion to tuple representation."""
+    adj = nx.adjacency_matrix(
+        nx.from_dict_of_lists(adj)
+    )  # return a adjacency matrix of adj ( type is numpy)
+    adj_normalized = normalize_adj(adj + sp.eye(adj.shape[0]))  #
+    # return sparse_to_tuple(adj_normalized)
+    return adj_normalized.todense()
+
+
+def row_norm(inputs):
+    outputs = []
+    for x in inputs:
+        xsum = x.sum()
+        x = x / xsum
+        outputs.append(x)
+    return outputs
+
+
+def normalize_adj_torch(adj):
+    # print(adj.size())
+    if len(adj.size()) == 4:
+        new_r = torch.zeros(adj.size()).type_as(adj)
+        for i in range(adj.size(1)):
+            adj_item = adj[0, i]
+            rowsum = adj_item.sum(1)
+            d_inv_sqrt = rowsum.pow_(-0.5)
+            d_inv_sqrt[torch.isnan(d_inv_sqrt)] = 0
+            d_mat_inv_sqrt = torch.diag(d_inv_sqrt)
+            r = torch.matmul(torch.matmul(d_mat_inv_sqrt, adj_item), d_mat_inv_sqrt)
+            new_r[0, i, ...] = r
+        return new_r
+    rowsum = adj.sum(1)
+    d_inv_sqrt = rowsum.pow_(-0.5)
+    d_inv_sqrt[torch.isnan(d_inv_sqrt)] = 0
+    d_mat_inv_sqrt = torch.diag(d_inv_sqrt)
+    r = torch.matmul(torch.matmul(d_mat_inv_sqrt, adj), d_mat_inv_sqrt)
+    return r
+
+
+# def row_norm(adj):
+
+
+if __name__ == "__main__":
+    a = row_norm(cihp2pascal_adj)
+    print(a)
+    print(cihp2pascal_adj)
+    # print(a.shape)
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_inference.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_inference.py
new file mode 100644
index 000000000..5e7de2477
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_inference.py
@@ -0,0 +1,242 @@
+import cv2
+import torch
+import timeit
+import numpy as np
+from PIL import Image
+import torch.nn.functional as F
+from torch.autograd import Variable
+from torchvision import transforms
+
+
+from FaceHairMask import graph
+from FaceHairMask import graphonomy_process as tr
+
+
+
+label_colours = [(0, 0, 0) for i in range(20)]
+label_colours[2] = (255, 0, 0)
+label_colours[13] = (0, 0, 255)
+
+
+def custom_decode_labels(mask, num_images=1, num_classes=20):
+    """Decode batch of segmentation masks.
+
+    Args:
+      mask: result of inference after taking argmax.
+      num_images: number of images to decode from the batch.
+      num_classes: number of classes to predict (including background).
+
+    Returns:
+      A batch with num_images RGB images of the same size as the input.
+    """
+    n, h, w = mask.shape
+
+    # import ipdb; ipdb.set_trace()
+    assert (
+        n >= num_images
+    ), "Batch size %d should be greater or equal than number of images to save %d." % (
+        n,
+        num_images,
+    )
+
+    hair_mask = torch.where(mask == 2, torch.ones_like(mask), torch.zeros_like(mask))
+
+    face_mask = torch.where(mask == 13, torch.ones_like(mask), torch.zeros_like(mask))
+
+    return hair_mask, face_mask
+
+
+def overlay(frame, mask):
+
+    mask = np.array(mask)
+    frame = np.array(frame)
+
+    tmp = cv2.cvtColor(mask, cv2.COLOR_RGB2GRAY)
+    _, alpha = cv2.threshold(tmp, 0, 255, cv2.THRESH_BINARY)
+    b, g, r = cv2.split(mask)
+    rgba = [b, g, r, alpha]
+    dst = cv2.merge(rgba, 4)
+
+    # overlay mask on frame
+    overlaid_image = cv2.addWeighted(frame, 0.4, dst, 0.1, 0)
+    return overlaid_image
+
+
+def flip(x, dim):
+    indices = [slice(None)] * x.dim()
+    indices[dim] = torch.arange(
+        x.size(dim) - 1, -1, -1, dtype=torch.long, device=x.device
+    )
+    return x[tuple(indices)]
+
+
+def flip_cihp(tail_list):
+    """
+
+    :param tail_list: tail_list size is 1 x n_class x h x w
+    :return:
+    """
+    # tail_list = tail_list[0]
+    tail_list_rev = [None] * 20
+    for xx in range(14):
+        tail_list_rev[xx] = tail_list[xx].unsqueeze(0)
+    tail_list_rev[14] = tail_list[15].unsqueeze(0)
+    tail_list_rev[15] = tail_list[14].unsqueeze(0)
+    tail_list_rev[16] = tail_list[17].unsqueeze(0)
+    tail_list_rev[17] = tail_list[16].unsqueeze(0)
+    tail_list_rev[18] = tail_list[19].unsqueeze(0)
+    tail_list_rev[19] = tail_list[18].unsqueeze(0)
+    return torch.cat(tail_list_rev, dim=0)
+
+
+def decode_labels(mask, num_images=1, num_classes=20):
+    """Decode batch of segmentation masks.
+
+    Args:
+      mask: result of inference after taking argmax.
+      num_images: number of images to decode from the batch.
+      num_classes: number of classes to predict (including background).
+
+    Returns:
+      A batch with num_images RGB images of the same size as the input.
+    """
+    n, h, w = mask.shape
+    assert (
+        n >= num_images
+    ), "Batch size %d should be greater or equal than number of images to save %d." % (
+        n,
+        num_images,
+    )
+    outputs = np.zeros((num_images, h, w, 3), dtype=np.uint8)
+    for i in range(num_images):
+        img = Image.new("RGB", (len(mask[i, 0]), len(mask[i])))
+        pixels = img.load()
+        for j_, j in enumerate(mask[i, :, :]):
+            for k_, k in enumerate(j):
+                if k < num_classes:
+                    pixels[k_, j_] = label_colours[k]
+        outputs[i] = np.array(img)
+    return outputs
+
+
+def read_img(img_path):
+    _img = Image.open(img_path).convert("RGB")  # return is RGB pic
+    return _img
+
+
+def img_transform(img, transform=None):
+    sample = {"image": img, "label": 0}
+
+    sample = transform(sample)
+    return sample
+
+
+def inference(net, img=None, device=None):
+    """
+
+    :param net:
+    :return:
+    """
+    # adj
+    adj2_ = torch.from_numpy(graph.cihp2pascal_nlp_adj).float()
+    adj2_test = (
+        adj2_.unsqueeze(0).unsqueeze(0).expand(1, 1, 7, 20).to(device).transpose(2, 3)
+    )
+
+    adj1_ = Variable(torch.from_numpy(graph.preprocess_adj(graph.pascal_graph)).float())
+    adj3_test = adj1_.unsqueeze(0).unsqueeze(0).expand(1, 1, 7, 7).to(device)
+
+    cihp_adj = graph.preprocess_adj(graph.cihp_graph)
+    adj3_ = Variable(torch.from_numpy(cihp_adj).float())
+    adj1_test = adj3_.unsqueeze(0).unsqueeze(0).expand(1, 1, 20, 20).to(device)
+
+    # multi-scale
+    scale_list = [1, 0.5, 0.75, 1.25, 1.5, 1.75]
+    #scale_list = [1, 0.5, 0.75, 1.25]
+    # NOTE: this part of the code assumes img is PIL image in RGB color space
+    # We provide torch tensor in range [-1, 1]
+    # Bring range to [0, 255]
+    img = torch.clamp(img, -1, 1)
+    img = (img + 1.0) / 2.0
+    img *= 255
+
+    testloader_list = []
+    testloader_flip_list = []
+    for pv in scale_list:
+        composed_transforms_ts = transforms.Compose(
+            [
+                tr.Scale_only_img(pv),
+                tr.Normalize_xception_tf_only_img(),
+                tr.ToTensor_only_img(),
+            ]
+        )
+
+        composed_transforms_ts_flip = transforms.Compose(
+            [
+                tr.Scale_only_img(pv),
+                tr.HorizontalFlip_only_img(),
+                tr.Normalize_xception_tf_only_img(),
+                tr.ToTensor_only_img(),
+            ]
+        )
+        # NOTE: img [1, 3, 256, 256], (min, max) = (0, 255)
+
+        # print("original:", img.shape, img.min(), img.max())
+        testloader_list.append(img_transform(img, composed_transforms_ts))
+        # print(img_transform(img, composed_transforms_ts))
+        testloader_flip_list.append(img_transform(img, composed_transforms_ts_flip))
+    # print(testloader_list)
+    start_time = timeit.default_timer()
+    # One testing epoch
+    # net.eval()
+    # 1 0.5 0.75 1.25 1.5 1.75 ; flip:
+
+    # NOTE: testloader_list[0]['image'].shape = 3, 420, 620
+
+    for iii, sample_batched in enumerate(zip(testloader_list, testloader_flip_list)):
+        inputs, labels = sample_batched[0]["image"], sample_batched[0]["label"]
+        inputs_f, _ = sample_batched[1]["image"], sample_batched[1]["label"]
+        inputs = inputs.unsqueeze(0)
+        inputs_f = inputs_f.unsqueeze(0)
+        inputs = torch.cat((inputs, inputs_f), dim=0)
+        if iii == 0:
+            _, _, h, w = inputs.size()
+        # assert inputs.size() == inputs_f.size()
+
+        # Forward pass of the mini-batch
+
+        # TODO: check requires grad functionality
+        # inputs = Variable(inputs, requires_grad=False)
+
+        with torch.no_grad():
+            if device is not None:
+                inputs = inputs.to(device)
+            # outputs = net.forward(inputs)
+            outputs = net.forward(
+                inputs, adj1_test.to(device), adj3_test.to(device), adj2_test.to(device)
+            )
+            outputs = (outputs[0] + flip(flip_cihp(outputs[1]), dim=-1)) / 2
+            outputs = outputs.unsqueeze(0)
+            if iii > 0:
+                outputs = F.upsample(
+                    outputs, size=(h, w), mode="bilinear", align_corners=True
+                )
+                outputs_final = outputs_final + outputs
+            else:
+                outputs_final = outputs.clone()
+    ################ plot pic
+    predictions = torch.max(outputs_final, 1)[1]
+    # results = predictions.cpu().numpy()
+    # vis_res = decode_labels(results)
+    # parsing_im = Image.fromarray(vis_res[0])
+    # return parsing_im
+
+    hair_mask, face_mask = custom_decode_labels(predictions)
+
+    return outputs_final, hair_mask, face_mask
+
+    # parsing_im.save(output_path+'/{}.png'.format(output_name))
+    # cv2.imwrite(output_path+'/{}_gray.png'.format(output_name), results[0, :, :])
+
+    # end_time = timeit.default_timer()
+    # print('time used for the multi-scale image inference' + ' is :' + str(end_time - start_time))
\ No newline at end of file
diff --git a/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_process.py b/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_process.py
new file mode 100644
index 000000000..4d47a0e3f
--- /dev/null
+++ b/reconstruction/ostec/external/graphonomy/FaceHairMask/graphonomy_process.py
@@ -0,0 +1,54 @@
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from torchvision import transforms
+
+
+class Scale_only_img(object):
+    def __init__(self, scale):
+        self.scale = scale
+
+    def __call__(self, sample):
+        img = sample["image"]
+        mask = sample["label"]
+        img = F.interpolate(  # NOTE: requires 4D
+            img, scale_factor=self.scale, mode="bilinear"
+        )
+        # print("Scale:", img.shape, img.min(), img.max())
+        # import ipdb; ipdb.set_trace()
+        return {"image": img, "label": mask}
+
+
+class Normalize_xception_tf_only_img(object):
+    def __call__(self, sample):
+        img = sample["image"]
+        img = (img * 2.0) / 255.0 - 1
+        # print("Normalize:", img.shape, img.min(), img.max())
+        # import ipdb; ipdb.set_trace()
+        return {"image": img, "label": sample["label"]}
+
+
+class ToTensor_only_img(object):
+    def __init__(self):
+        self.rgb2bgr = transforms.Lambda(lambda x: x[[2, 1, 0], ...])
+
+    def __call__(self, sample):
+        # sample: N x C x H x W
+        img = sample["image"]
+        img = torch.squeeze(img, axis=0)
+        # sample: C x H x W
+        img = self.rgb2bgr(img)
+        # print("To Tensor:", img.shape, img.min(), img.max())
+        # img = torch.unsqueeze(img, axis=0)
+        # import ipdb; ipdb.set_trace()
+        return {"image": img, "label": sample["label"]}
+
+
+class HorizontalFlip_only_img(object):
+    def __call__(self, sample):
+        img = sample["image"]
+        mask = sample["label"]
+        img = torch.flip(img, [-1])
+        # print("Horizontal:", img.shape, img.min(), img.max())
+        # import ipdb; ipdb.set_trace()
+        return {"image": img, "label": mask}
diff --git a/reconstruction/ostec/external/landmark_detector/Alignment84_demo-minimal.ipynb b/reconstruction/ostec/external/landmark_detector/Alignment84_demo-minimal.ipynb
new file mode 100644
index 000000000..bc59bf338
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/Alignment84_demo-minimal.ipynb
@@ -0,0 +1,160 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Face Alignment by Hourglass (84 landmarks)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "WARNING:tensorflow:\n",
+      "The TensorFlow contrib module will not be included in TensorFlow 2.0.\n",
+      "For more information, please see:\n",
+      "  * https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md\n",
+      "  * https://github.com/tensorflow/addons\n",
+      "  * https://github.com/tensorflow/io (for I/O related ops)\n",
+      "If you depend on functionality not listed there, please file an issue.\n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "import tensorflow as tf\n",
+    "import menpo.io as mio\n",
+    "import numpy as np\n",
+    "import matplotlib.pyplot as pt\n",
+    "import networks\n",
+    "from flags import FLAGS"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def tf_heatmap_to_lms(heatmap):\n",
+    "    hs = tf.argmax(tf.reduce_max(heatmap, 2), 1)\n",
+    "    ws = tf.argmax(tf.reduce_max(heatmap, 1), 1)\n",
+    "    lms = tf.transpose(tf.to_float(tf.stack([hs, ws])), perm=[1, 2, 0])\n",
+    "    return lms"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "WARNING:tensorflow:From /media/Projects/_Active/Unsupervised_UV/Face_Detection_Alignment-master/models.py:91: The name tf.GraphKeys is deprecated. Please use tf.compat.v1.GraphKeys instead.\n",
+      "\n",
+      "WARNING:tensorflow:From /home/baris/anaconda3/envs/ganfit_stylegan2/lib/python3.6/site-packages/tensorflow_core/contrib/layers/python/layers/layers.py:1057: Layer.apply (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version.\n",
+      "Instructions for updating:\n",
+      "Please use `layer.__call__` method instead.\n",
+      "WARNING:tensorflow:From <ipython-input-2-52eace1d17da>:4: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\n",
+      "Instructions for updating:\n",
+      "Use `tf.cast` instead.\n",
+      "WARNING:tensorflow:From <ipython-input-3-f39427770ee7>:11: all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02.\n",
+      "Instructions for updating:\n",
+      "Please use tf.global_variables instead.\n",
+      "INFO:tensorflow:Restoring parameters from ../models/alignment/3D84/model.ckpt-277538\n"
+     ]
+    }
+   ],
+   "source": [
+    "model_path = '../models/alignment/3D84/model.ckpt-277538'\n",
+    "template = mio.import_pickle('./image/template.pkl.gz')\n",
+    "n_landmarks = 84\n",
+    "FLAGS.n_landmarks = 84\n",
+    "with tf.Graph().as_default() as g:\n",
+    "    images_input = tf.placeholder(tf.float32, shape=(None, None, None, 3), name='input_images')\n",
+    "    net_model = networks.DNFaceMultiView('')\n",
+    "    with tf.variable_scope('net'):\n",
+    "        lms_heatmap_prediction,states = net_model._build_network(images_input, datas=None, is_training=False, n_channels=n_landmarks)\n",
+    "        pts_predictions = tf_heatmap_to_lms(lms_heatmap_prediction)\n",
+    "        variables_to_restore = tf.all_variables()\n",
+    "        saver = tf.train.Saver(variables_to_restore)\n",
+    "\n",
+    "config = tf.ConfigProto()\n",
+    "config.gpu_options.allow_growth = True\n",
+    "sess = tf.Session(config=config,graph=g)\n",
+    "\n",
+    "saver.restore(sess, model_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<matplotlib.collections.PathCollection at 0x7ff7233b9d68>"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    },
+    {
+     "data": {
+      "image/png": "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\n",
+      "text/plain": [
+       "<Figure size 432x288 with 1 Axes>"
+      ]
+     },
+     "metadata": {
+      "needs_background": "light"
+     },
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "image_path = '00000-generate-images/seed6622.png'\n",
+    "image = mio.import_image(image_path)\n",
+    "image = image.resize([256,256])\n",
+    "input_pixels = image.pixels_with_channels_at_back()\n",
+    "pts_pred = sess.run(\n",
+    "    pts_predictions,\n",
+    "    feed_dict={images_input: np.expand_dims(input_pixels, axis=0)})\n",
+    "pt.imshow(input_pixels)\n",
+    "pt.scatter(pts_pred[0][:,1],pts_pred[0][:,0])"
+   ]
+  }
+ ],
+ "metadata": {
+  "anaconda-cloud": {},
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 1
+}
diff --git a/reconstruction/ostec/external/landmark_detector/README.md b/reconstruction/ostec/external/landmark_detector/README.md
new file mode 100644
index 000000000..52792b5af
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/README.md
@@ -0,0 +1,16 @@
+# Face_Detection_Alignment
+Face Detection and Alignment Tool
+3D projection landmarks (84) and 2D multi-view landmarks(39/68)
+
+Environment:
+Tensorflow 1.3, menpo, python 3.5
+
+Train:
+CUDA_VISIBLE_DEVICES="1" python train.py --train_dir=ckpt/3D84 --batch_size=8 --initial_learning_rate=0.0001 --dataset_dir=3D84/300W.tfrecords,3D84/afw.tfrecords,3D84/helen_testset.tfrecords,3D84/helen_trainset.tfrecords,3D84/lfpw_testset.tfrecords,3D84/lfpw_trainset.tfrecords,3D84/ibug.tfrecords,3D84/menpo_trainset.tfrecords --n_landmarks=84
+
+Test:
+3D model: 84
+2D model: frontal68/Union68/Union86(better)
+
+Pretrained Models:
+https://drive.google.com/open?id=1DKTeRlJjyo_tD1EluDjYLhtKFPJ9vIVd
diff --git a/reconstruction/ostec/external/landmark_detector/data_provider.py b/reconstruction/ostec/external/landmark_detector/data_provider.py
new file mode 100644
index 000000000..53266694a
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/data_provider.py
@@ -0,0 +1,236 @@
+import tensorflow as tf
+import numpy as np
+
+from menpo.transform import Translation
+
+from external.landmark_detector.flags import FLAGS
+
+def augment_img(img, augmentation):
+    flip, rotate, rescale = np.array(augmentation).squeeze()
+    rimg = img.rescale(rescale)
+    rimg = rimg.rotate_ccw_about_centre(rotate)
+    crimg = rimg.warp_to_shape(
+        img.shape,
+        Translation(-np.array(img.shape) / 2 + np.array(rimg.shape) / 2)
+    )
+    if flip > 0.5:
+        crimg = crimg.mirror()
+
+    img = crimg
+
+    return img
+
+def rotate_points_tensor(points, image, angle):
+
+    s = tf.shape(image)
+    image_center = tf.to_float(s[:2]) / 2.
+
+    # center coordinates since rotation center is supposed to be in the image center
+    points_centered = points - image_center
+
+    rot_matrix = tf.dynamic_stitch([[0], [1], [2], [3]], [tf.cos(angle), -tf.sin(angle), tf.sin(angle), tf.cos(angle)])
+    rot_matrix = tf.reshape(rot_matrix, shape=[2, 2])
+
+    points_centered_rot = tf.matmul(rot_matrix, tf.transpose(points_centered))
+
+    return tf.transpose(points_centered_rot) + image_center
+
+def rotate_image_tensor(image, angle):
+    s = tf.shape(image)
+    image_center = tf.to_float(s[:2]) / 2.
+
+    # Coordinates of new image
+    xs, ys = tf.meshgrid(tf.range(0.,tf.to_float(s[1])), tf.range(0., tf.to_float(s[0])))
+    coords_new = tf.reshape(tf.stack([ys,xs], 2), [-1, 2])
+
+    # center coordinates since rotation center is supposed to be in the image center
+    coords_new_centered = tf.to_float(coords_new) - image_center
+
+    # Perform backward transformation of the image coordinates
+    rot_mat_inv = tf.stack(
+        [tf.cos(angle), tf.sin(angle), -tf.sin(angle), tf.cos(angle)])
+    rot_mat_inv = tf.reshape(rot_mat_inv, shape=[2, 2])
+    coord_old_centered = tf.matmul(
+        rot_mat_inv, tf.transpose(coords_new_centered))
+    coord_old = tf.to_int32(tf.round(
+        tf.transpose(coord_old_centered) + image_center))
+
+
+    # Find nearest neighbor in old image
+    coord_old_y, coord_old_x = tf.unstack(coord_old, axis=1)
+
+
+    # Clip values to stay inside image coordinates
+    outside_y = tf.logical_or(tf.greater(
+        coord_old_y, s[0]-1), tf.less(coord_old_y, 0))
+    outside_x = tf.logical_or(tf.greater(
+        coord_old_x, s[1]-1), tf.less(coord_old_x, 0))
+    outside_ind = tf.logical_or(outside_y, outside_x)
+
+
+    inside_mask = tf.logical_not(outside_ind)
+    inside_mask = tf.tile(tf.reshape(inside_mask, s[:2])[...,None], tf.stack([1,1,s[2]]))
+
+    coord_old_y = tf.clip_by_value(coord_old_y, 0, s[0]-1)
+    coord_old_x = tf.clip_by_value(coord_old_x, 0, s[1]-1)
+    coord_flat = coord_old_y * s[1] + coord_old_x
+
+    im_flat = tf.reshape(image, tf.stack([-1, s[2]]))
+    rot_image = tf.gather(im_flat, coord_flat)
+    rot_image = tf.reshape(rot_image, s)
+
+
+    return tf.where(inside_mask, rot_image, tf.zeros_like(rot_image))
+
+def lms_to_heatmap(lms, h, w, n_landmarks, marked_index, sigma=5):
+    xs, ys = tf.meshgrid(tf.range(0., tf.to_float(w)),
+                         tf.range(0., tf.to_float(h)))
+    gaussian = (1. / (sigma * np.sqrt(2. * np.pi)))
+    marked_index = tf.to_int32(marked_index)
+
+    def gaussian_fn(lms):
+        y, x, idx = tf.unstack(lms)
+        idx = tf.to_int32(idx)
+
+        def run_true():
+            return tf.exp(-0.5 * (tf.pow(ys - y, 2) + tf.pow(xs - x, 2)) *
+                          tf.pow(1. / sigma, 2.)) * gaussian * 17.
+
+        def run_false():
+            return tf.zeros((h, w))
+
+        return tf.cond(tf.reduce_any(tf.equal(marked_index, idx)), run_true, run_false)
+
+    img_hm = tf.stack(tf.map_fn(gaussian_fn, tf.concat(
+        [lms, tf.to_float(tf.range(0, n_landmarks))[..., None]], 1)))
+
+    return img_hm
+
+class ProtobuffProvider(object):
+    def __init__(self, filename= FLAGS['dataset_dir'].value, batch_size=1, rescale=None, augmentation=False):
+        self.filename = filename
+        self.batch_size = batch_size
+        self.image_extension = 'jpg'
+        self.rescale = rescale
+        self.augmentation = augmentation
+
+    def get(self):
+        images, *names = self._get_data_protobuff(self.filename)
+        tensors = [images]
+
+        for name in names:
+            tensors.append(name)
+
+        return tf.train.shuffle_batch(
+            tensors, self.batch_size, 256, 64, self.batch_size)
+
+    def augmentation_type(self):
+        return tf.stack([tf.random_uniform([1]) - 1,
+                        (tf.random_uniform([1]) * 30. - 15.) * np.pi / 180.,
+                        tf.random_uniform([1]) * 0.5 + 0.75])
+
+    def _image_from_feature(self, features):
+        image = tf.image.decode_jpeg(features['image'], channels=3)
+        image_height = tf.to_int32(features['height'])
+        image_width = tf.to_int32(features['width'])
+        image = tf.reshape(image, (image_height, image_width, 3))
+        image = tf.to_float(image)
+        return image, image_height, image_width
+
+    def _heatmap_from_feature(self, features):
+        n_landmarks = tf.to_int32(features['n_landmarks'])
+        gt_lms = tf.decode_raw(features['gt_pts'], tf.float32)
+        mask_index = tf.decode_raw(features['mask_index'], tf.float32)
+        gt_mask = tf.decode_raw(features['gt_mask'], tf.float32)
+        image_height = tf.to_int32(features['height'])
+        image_width = tf.to_int32(features['width'])
+
+        gt_lms = tf.reshape(gt_lms, (n_landmarks, 2))
+        gt_heatmap = lms_to_heatmap(
+            gt_lms, image_height, image_width, n_landmarks, mask_index)
+        gt_heatmap = tf.transpose(gt_heatmap, perm=[1,2,0])
+
+        return gt_heatmap, gt_lms, n_landmarks, mask_index, gt_mask
+
+
+    def _info_from_feature(self, features):
+        status = features['status']
+        return status
+
+    def _set_shape(self, image, gt_heatmap, gt_lms, mask_index, gt_mask):
+        image.set_shape([None, None, 3])
+        gt_heatmap.set_shape([None, None, FLAGS['n_landmarks']])
+        gt_lms.set_shape([FLAGS['n_landmarks'], 2])
+        mask_index.set_shape([FLAGS['n_landmarks']])
+        gt_mask.set_shape([FLAGS['n_landmarks']])
+
+    def _get_features(self, serialized_example):
+        features = tf.parse_single_example(
+            serialized_example,
+            features={
+                # images
+                'image': tf.FixedLenFeature([], tf.string),
+                'height': tf.FixedLenFeature([], tf.int64),
+                'width': tf.FixedLenFeature([], tf.int64),
+                # landmarks
+                'n_landmarks': tf.FixedLenFeature([], tf.int64),
+                'gt_pts': tf.FixedLenFeature([], tf.string),
+                'gt_mask': tf.FixedLenFeature([], tf.string),
+                'mask_index': tf.FixedLenFeature([], tf.string),
+                'status': tf.FixedLenFeature([], tf.int64),
+            }
+
+        )
+        return features
+
+    def _get_data_protobuff(self, filename):
+        filename = str(filename).split(',')
+        filename_queue = tf.train.string_input_producer(filename,
+                                                        num_epochs=None)
+        reader = tf.TFRecordReader()
+        _, serialized_example = reader.read(filename_queue)
+        features = self._get_features(serialized_example)
+
+        # image
+        image, image_height, image_width = self._image_from_feature(features)
+
+        # landmarks
+        gt_heatmap, gt_lms, n_landmarks, mask_index, gt_mask = self._heatmap_from_feature(features)
+
+        # infomations
+        status = self._info_from_feature(features)
+
+        # augmentation
+        if self.augmentation:
+            do_flip, do_rotate, do_scale = tf.unstack(self.augmentation_type())
+
+            # rescale
+            image_height = tf.to_int32(tf.to_float(image_height) * do_scale[0])
+            image_width = tf.to_int32(tf.to_float(image_width) * do_scale[0])
+
+            image = tf.image.resize_images(image, tf.stack([image_height, image_width]))
+            gt_heatmap = tf.image.resize_images(gt_heatmap, tf.stack([image_height, image_width]))
+            gt_lms = gt_lms*do_scale
+
+            # rotate
+            image = rotate_image_tensor(image, do_rotate)
+            gt_heatmap = rotate_image_tensor(gt_heatmap, do_rotate)
+            gt_lms = rotate_points_tensor(gt_lms, image, do_rotate)
+
+        # crop to 256 * 256
+        target_h = tf.to_int32(256)
+        target_w = tf.to_int32(256)
+        offset_h = tf.to_int32((image_height - target_h) / 2)
+        offset_w = tf.to_int32((image_width - target_w) / 2)
+
+        image = tf.image.crop_to_bounding_box(
+            image, offset_h, offset_w, target_h, target_w)
+
+        gt_heatmap = tf.image.crop_to_bounding_box(
+            gt_heatmap, offset_h, offset_w, target_h, target_w)
+
+        gt_lms = gt_lms - tf.to_float(tf.stack([offset_h, offset_w]))
+
+        self._set_shape(image, gt_heatmap, gt_lms, mask_index, gt_mask)
+
+        return image, gt_heatmap, gt_lms, mask_index, gt_mask
diff --git a/reconstruction/ostec/external/landmark_detector/flags.py b/reconstruction/ostec/external/landmark_detector/flags.py
new file mode 100644
index 000000000..d89fe8119
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/flags.py
@@ -0,0 +1,32 @@
+import tensorflow as tf
+
+slim = tf.contrib.slim
+
+FLAGS = tf.app.flags.FLAGS
+tf.app.flags.DEFINE_float('initial_learning_rate', 0.0001, '''Initial learning rate.''')
+tf.app.flags.DEFINE_float('num_epochs_per_decay', 5.0, '''Epochs after which learning rate decays.''')
+tf.app.flags.DEFINE_float('learning_rate_decay_factor', 0.97, '''Learning rate decay factor.''')
+tf.app.flags.DEFINE_float('learning_rate_decay_step', 30000,'''Learning rate decay factor.''')
+
+tf.app.flags.DEFINE_integer('batch_size', 4, '''The batch size to use.''')
+tf.app.flags.DEFINE_integer('eval_size', 4, '''The batch size to use.''')
+tf.app.flags.DEFINE_integer('num_iterations', 2, '''The number of iterations to unfold the pose machine.''')
+tf.app.flags.DEFINE_integer('num_preprocess_threads', 4,'''How many preprocess threads to use.''')
+tf.app.flags.DEFINE_integer('n_landmarks', 84,'''number of landmarks.''')
+tf.app.flags.DEFINE_integer('rescale', 256,'''Image scale.''')
+
+tf.app.flags.DEFINE_string('dataset_dir', './data', '''Directory where to load datas.''')
+tf.app.flags.DEFINE_string('train_dir', 'ckpt/train', '''Directory where to write event logs and checkpoint.''')
+tf.app.flags.DEFINE_string('eval_dir', '','''Directory where to write event logs and checkpoint.''')
+tf.app.flags.DEFINE_string('graph_dir', 'model/weight.pkl','''If specified, restore this pretrained model.''')
+
+tf.app.flags.DEFINE_integer('max_steps', 1000000,'''Number of batches to run.''')
+tf.app.flags.DEFINE_string('train_device', '/gpu:0','''Device to train with.''')
+
+tf.app.flags.DEFINE_integer('flip_pred', 0,'''db name.''')
+
+tf.app.flags.DEFINE_string('train_model', '', '''training model.''')
+tf.app.flags.DEFINE_string('pretrained_model_checkpoint_path', '', '''Restore pretrained model.''')
+tf.app.flags.DEFINE_string('testset_name', '', '''test set name.''')
+tf.app.flags.DEFINE_string('model_name', '', '''test model name.''')
+tf.app.flags.DEFINE_string('savemat_name', '', '''save_mat_name''')
\ No newline at end of file
diff --git a/reconstruction/ostec/external/landmark_detector/models.py b/reconstruction/ostec/external/landmark_detector/models.py
new file mode 100644
index 000000000..5d7c4c45c
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/models.py
@@ -0,0 +1,1333 @@
+import tensorflow as tf
+import numpy as np
+
+slim = tf.contrib.slim
+
+# custom layers
+
+def deconv_layer(net, up_scale, n_channel, method='transpose'):
+    nh = tf.shape(net)[-3] * up_scale
+    nw = tf.shape(net)[-2] * up_scale
+
+    if method == 'transpose':
+        net = slim.conv2d_transpose(net, n_channel, (up_scale, up_scale), (
+            up_scale, up_scale), activation_fn=None, padding='VALID')
+    elif method == 'transpose+conv':
+        net = slim.conv2d_transpose(net, n_channel, (up_scale, up_scale), (
+            up_scale, up_scale), activation_fn=None, padding='VALID')
+        net = slim.conv2d(net, n_channel, (3, 3), (1, 1))
+    elif method == 'transpose+conv+relu':
+        net = slim.conv2d_transpose(net, n_channel, (up_scale, up_scale), (
+            up_scale, up_scale), padding='VALID')
+        net = slim.conv2d(net, n_channel, (3, 3), (1, 1))
+    elif method == 'bilinear':
+        net = tf.image.resize_images(net, [nh, nw])
+    else:
+        raise Exception('Unrecognised Deconvolution Method: %s' % method)
+
+    return net
+
+
+# arg scopes
+def hourglass_arg_scope_torch(weight_decay=0.0001,
+                              batch_norm_decay=0.997,
+                              batch_norm_epsilon=1e-5,
+                              batch_norm_scale=True):
+    """Defines the default ResNet arg scope.
+  Args:
+    is_training: Whether or not we are training the parameters in the batch
+      normalization layers of the model.
+    weight_decay: The weight decay to use for regularizing the model.
+    batch_norm_decay: The moving average decay when estimating layer activation
+      statistics in batch normalization.
+    batch_norm_epsilon: Small constant to prevent division by zero when
+      normalizing activations by their variance in batch normalization.
+    batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
+      activations in the batch normalization layer.
+  Returns:
+    An `arg_scope` to use for the resnet models.
+  """
+    batch_norm_params = {
+        'decay': batch_norm_decay,
+        'epsilon': batch_norm_epsilon,
+        'scale': batch_norm_scale,
+        'updates_collections': tf.GraphKeys.UPDATE_OPS,
+    }
+
+    with slim.arg_scope(
+        [slim.conv2d],
+            weights_regularizer=slim.l2_regularizer(weight_decay),
+            weights_initializer=slim.variance_scaling_initializer(),
+            activation_fn=None,
+            normalizer_fn=None,
+            normalizer_params=None):
+        with slim.arg_scope([slim.batch_norm], **batch_norm_params):
+            with slim.arg_scope([slim.max_pool2d], padding='VALID') as arg_sc:
+                return arg_sc
+
+
+def hourglass_arg_scope_tf(weight_decay=0.0001,
+                           batch_norm_decay=0.997,
+                           batch_norm_epsilon=1e-5,
+                           batch_norm_scale=True):
+    """Defines the default ResNet arg scope.
+  Args:
+    is_training: Whether or not we are training the parameters in the batch
+      normalization layers of the model.
+    weight_decay: The weight decay to use for regularizing the model.
+    batch_norm_decay: The moving average decay when estimating layer activation
+      statistics in batch normalization.
+    batch_norm_epsilon: Small constant to prevent division by zero when
+      normalizing activations by their variance in batch normalization.
+    batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
+      activations in the batch normalization layer.
+  Returns:
+    An `arg_scope` to use for the resnet models.
+  """
+    batch_norm_params = {
+        'decay': batch_norm_decay,
+        'epsilon': batch_norm_epsilon,
+        'scale': batch_norm_scale,
+        'updates_collections': tf.GraphKeys.UPDATE_OPS,
+    }
+
+    with slim.arg_scope(
+        [slim.conv2d],
+            weights_regularizer=slim.l2_regularizer(weight_decay),
+            weights_initializer=slim.variance_scaling_initializer(),
+            activation_fn=tf.nn.relu,
+            normalizer_fn=slim.batch_norm,
+            normalizer_params=batch_norm_params):
+        with slim.arg_scope([slim.batch_norm], **batch_norm_params):
+            with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
+                return arg_sc
+
+
+# bottleneck_inception_SE
+def bottleneck_inception_SE_module(
+        inputs,
+        out_channel=256,
+        res=None,
+        scope='inception_block'):
+
+    min_channel = out_channel // 8
+    with tf.variable_scope(scope):
+        with tf.variable_scope('Branch_0'):
+            branch_0 = slim.conv2d(inputs, min_channel * 3,
+                                   [1, 1], scope='Conv2d_1x1')
+        with tf.variable_scope('Branch_1'):
+            branch_1 = slim.conv2d(inputs, min_channel *
+                                   3 / 2, [1, 1], scope='Conv2d_1x1')
+            branch_1 = slim.conv2d(
+                branch_1, min_channel * 3, [3, 3], scope='Conv2d_3x3')
+        with tf.variable_scope('Branch_2'):
+            branch_2 = slim.conv2d(inputs, min_channel //
+                                   3, [1, 1], scope='Conv2d_1x1')
+            branch_2 = slim.conv2d(
+                branch_2, min_channel, [3, 3], scope='Conv2d_3x3')
+        with tf.variable_scope('Branch_3'):
+            branch_3 = slim.max_pool2d(inputs, [3, 3], 1, scope='MaxPool_3x3')
+            branch_3 = slim.conv2d(
+                branch_3, min_channel, [1, 1], scope='Conv2d_1x1')
+        net = tf.concat(
+            axis=3, values=[branch_0, branch_1, branch_2, branch_3])
+
+        se_branch = tf.reduce_mean(net, axis=[1, 2])
+        se_branch = slim.fully_connected(se_branch, out_channel // 16)
+        se_branch = slim.fully_connected(
+            se_branch, out_channel, activation_fn=tf.sigmoid)
+
+        net = net * se_branch[:,None,None,:]
+
+        if res:
+            inputs = slim.conv2d(inputs, res, (1, 1),
+                                 scope='bn_res'.format(scope))
+
+        net += inputs
+
+    return net
+
+
+# bottle neck modules
+def bottleneck_inception_module(
+        inputs,
+        out_channel=256,
+        res=None,
+        scope='inception_block'):
+
+    min_channel = out_channel // 8
+    with tf.variable_scope(scope):
+        with tf.variable_scope('Branch_0'):
+            branch_0 = slim.conv2d(inputs, min_channel * 3,
+                                   [1, 1], scope='Conv2d_1x1')
+        with tf.variable_scope('Branch_1'):
+            branch_1 = slim.conv2d(inputs, min_channel *
+                                   3 / 2, [1, 1], scope='Conv2d_1x1')
+            branch_1 = slim.conv2d(
+                branch_1, min_channel * 3, [3, 3], scope='Conv2d_3x3')
+        with tf.variable_scope('Branch_2'):
+            branch_2 = slim.conv2d(inputs, min_channel //
+                                   3, [1, 1], scope='Conv2d_1x1')
+            branch_2 = slim.conv2d(
+                branch_2, min_channel, [3, 3], scope='Conv2d_3x3')
+        with tf.variable_scope('Branch_3'):
+            branch_3 = slim.max_pool2d(inputs, [3, 3], 1, scope='MaxPool_3x3')
+            branch_3 = slim.conv2d(
+                branch_3, min_channel, [1, 1], scope='Conv2d_1x1')
+        net = tf.concat(
+            axis=3, values=[branch_0, branch_1, branch_2, branch_3])
+
+        if res:
+            inputs = slim.conv2d(inputs, res, (1, 1),
+                                 scope='bn_res'.format(scope))
+
+        net += inputs
+
+    return net
+
+
+def bottleneck_module(inputs, out_channel=256, res=None, scope=''):
+
+    with tf.variable_scope(scope):
+        net = slim.stack(inputs, slim.conv2d, [
+                         (out_channel // 2, [1, 1]), (out_channel // 2, [3, 3]), (out_channel, [1, 1])], scope='conv')
+        if res:
+            inputs = slim.conv2d(inputs, res, (1, 1),
+                                 scope='bn_res'.format(scope))
+        net += inputs
+
+        return net
+
+
+# recursive hourglass definition
+def hourglass_module(inputs, depth=0, deconv='bilinear', bottleneck='bottleneck'):
+
+    bm_fn = globals()['%s_module' % bottleneck]
+
+    with tf.variable_scope('depth_{}'.format(depth)):
+        # buttom up layers
+        net = slim.max_pool2d(inputs, [2, 2], scope='pool')
+        net = slim.stack(net, bm_fn, [
+                         (256, None), (256, None), (256, None)], scope='buttom_up')
+
+        # connecting layers
+        if depth > 0:
+            net = hourglass_module(net, depth=depth - 1, deconv=deconv)
+        else:
+            net = bm_fn(
+                net, out_channel=512, res=512, scope='connecting')
+
+        # top down layers
+        net = bm_fn(net, out_channel=512,
+                    res=512, scope='top_down')
+        net = deconv_layer(net, 2, 512, method=deconv)
+        # residual layers
+        net += slim.stack(inputs, bm_fn,
+                          [(256, None), (256, None), (512, 512)], scope='res')
+
+        return net
+
+
+def hourglass(inputs,
+              scale=1,
+              regression_channels=2,
+              classification_channels=22,
+              deconv='bilinear',
+              bottleneck='bottleneck'):
+    """Defines a lightweight resnet based model for dense estimation tasks.
+    Args:
+      inputs: A `Tensor` with dimensions [num_batches, height, width, depth].
+      scale: A scalar which denotes the factor to subsample the current image.
+      output_channels: The number of output channels. E.g., for human pose
+        estimation this equals 13 channels.
+    Returns:
+      A `Tensor` of dimensions [num_batches, height, width, output_channels]."""
+
+    out_shape = tf.shape(inputs)[1:3]
+
+    if scale > 1:
+        inputs = tf.pad(inputs, ((0, 0), (1, 1), (1, 1), (0, 0)))
+        inputs = slim.layers.avg_pool2d(
+            inputs, (3, 3), (scale, scale), padding='VALID')
+
+    output_channels = regression_channels + classification_channels
+
+    with slim.arg_scope(hourglass_arg_scope_tf()):
+        # D1
+        net = slim.conv2d(inputs, 64, (7, 7), 2, scope='conv1')
+        net = bottleneck_module(net, out_channel=128,
+                                res=128, scope='bottleneck1')
+        net = slim.max_pool2d(net, [2, 2], scope='pool1')
+
+        # D2
+        net = slim.stack(net, bottleneck_module, [
+                         (128, None), (128, None), (256, 256)], scope='conv2')
+
+        # hourglasses (D3,D4,D5)
+        with tf.variable_scope('hourglass'):
+            net = hourglass_module(
+                net, depth=4, deconv=deconv, bottleneck=bottleneck)
+
+        # final layers (D6, D7)
+        net = slim.stack(net, slim.conv2d, [(512, [1, 1]), (256, [1, 1]),
+                                            (output_channels, [1, 1])
+                                            ], scope='conv3')
+
+        net = deconv_layer(net, 4, output_channels, method=deconv)
+        net = slim.conv2d(net, output_channels, 1, scope='conv_last')
+
+    regression = slim.conv2d(
+        net, regression_channels, 1, activation_fn=None
+    ) if regression_channels else None
+
+    logits = slim.conv2d(
+        net, classification_channels, 1, activation_fn=None
+    ) if classification_channels else None
+
+    return regression, logits
+
+
+def StackedHourglassTorch(inputs, out_channels=16, deconv='bilinear'):
+    net = inputs
+    with tf.name_scope('nn.Sequential'):
+        with tf.name_scope('nn.Sequential'):
+            net = tf.pad(net, np.array([[0, 0], [3, 3], [3, 3], [0, 0]]))
+            net = slim.conv2d(net, 64, (7, 7), (2, 2),
+                              activation_fn=None, padding='VALID')
+            net = slim.batch_norm(net)
+            net = slim.nn.relu(net)
+            with tf.name_scope('nn.Sequential'):
+                with tf.name_scope('nn.ConcatTable'):
+                    net0 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                    net1 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net1 = tf.pad(net1, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net1 = slim.conv2d(
+                            net1, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                net = tf.add_n([net0, net1])
+            net = tf.pad(net, np.array([[0, 0], [0, 0], [0, 0], [0, 0]]))
+            net = slim.max_pool2d(net, (2, 2), (2, 2))
+            with tf.name_scope('nn.Sequential'):
+                with tf.name_scope('nn.ConcatTable'):
+                    net0 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                    net1 = net
+                net = tf.add_n([net0, net1])
+            with tf.name_scope('nn.Sequential'):
+                with tf.name_scope('nn.ConcatTable'):
+                    net0 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 64, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                    net1 = net
+                net = tf.add_n([net0, net1])
+            with tf.name_scope('nn.Sequential'):
+                with tf.name_scope('nn.ConcatTable'):
+                    net0 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                        net0 = slim.batch_norm(net0)
+                        net0 = slim.nn.relu(net0)
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.conv2d(
+                            net0, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                    net1 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net1 = tf.pad(net1, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net1 = slim.conv2d(
+                            net1, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                net = tf.add_n([net0, net1])
+            with tf.name_scope('nn.Sequential'):
+                with tf.name_scope('nn.ConcatTable'):
+                    net0 = net
+                    with tf.name_scope('nn.Sequential'):
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = slim.max_pool2d(net0, (2, 2), (2, 2))
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net00 = net0
+                                with tf.name_scope('nn.Sequential'):
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net01 = net0
+                            net0 = tf.add_n([net00, net01])
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net00 = net0
+                                with tf.name_scope('nn.Sequential'):
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net01 = net0
+                            net0 = tf.add_n([net00, net01])
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net00 = net0
+                                with tf.name_scope('nn.Sequential'):
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net00 = slim.batch_norm(net00)
+                                    net00 = slim.nn.relu(net00)
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.conv2d(
+                                        net00, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net01 = net0
+                            net0 = tf.add_n([net00, net01])
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net00 = net0
+                                with tf.name_scope('nn.Sequential'):
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = slim.max_pool2d(
+                                        net00, (2, 2), (2, 2))
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net000 = net00
+                                            with tf.name_scope('nn.Sequential'):
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net001 = net00
+                                        net00 = tf.add_n([net000, net001])
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net000 = net00
+                                            with tf.name_scope('nn.Sequential'):
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net001 = net00
+                                        net00 = tf.add_n([net000, net001])
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net000 = net00
+                                            with tf.name_scope('nn.Sequential'):
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net000 = slim.batch_norm(
+                                                    net000)
+                                                net000 = slim.nn.relu(net000)
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.conv2d(
+                                                    net000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net001 = net00
+                                        net00 = tf.add_n([net000, net001])
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net000 = net00
+                                            with tf.name_scope('nn.Sequential'):
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = slim.max_pool2d(
+                                                    net000, (2, 2), (2, 2))
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0000 = net000
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0001 = net000
+                                                    net000 = tf.add_n(
+                                                        [net0000, net0001])
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0000 = net000
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0001 = net000
+                                                    net000 = tf.add_n(
+                                                        [net0000, net0001])
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0000 = net000
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0000 = slim.batch_norm(
+                                                                net0000)
+                                                            net0000 = slim.nn.relu(
+                                                                net0000)
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.conv2d(
+                                                                net0000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0001 = net000
+                                                    net000 = tf.add_n(
+                                                        [net0000, net0001])
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0000 = net000
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = slim.max_pool2d(
+                                                                net0000, (2, 2), (2, 2))
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00000 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00001 = net0000
+                                                                net0000 = tf.add_n(
+                                                                    [net00000, net00001])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00000 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00001 = net0000
+                                                                net0000 = tf.add_n(
+                                                                    [net00000, net00001])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00000 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00001 = net0000
+                                                                net0000 = tf.add_n(
+                                                                    [net00000, net00001])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00000 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00001 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00001 = tf.pad(net00001, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00001 = slim.conv2d(
+                                                                            net00001, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                net0000 = tf.add_n(
+                                                                    [net00000, net00001])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00000 = net0000
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00000 = slim.batch_norm(
+                                                                            net00000)
+                                                                        net00000 = slim.nn.relu(
+                                                                            net00000)
+                                                                        net00000 = tf.pad(net00000, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00000 = slim.conv2d(
+                                                                            net00000, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00001 = net0000
+                                                                net0000 = tf.add_n(
+                                                                    [net00000, net00001])
+                                                            net0000 = tf.pad(net0000, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0000 = deconv_layer(
+                                                                net0000, 2, 512, method=deconv)
+                                                        net0001 = net000
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00010 = net0001
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00011 = net0001
+                                                                net0001 = tf.add_n(
+                                                                    [net00010, net00011])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00010 = net0001
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00011 = net0001
+                                                                net0001 = tf.add_n(
+                                                                    [net00010, net00011])
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                with tf.name_scope('nn.ConcatTable'):
+                                                                    net00010 = net0001
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                        net00010 = slim.batch_norm(
+                                                                            net00010)
+                                                                        net00010 = slim.nn.relu(
+                                                                            net00010)
+                                                                        net00010 = tf.pad(net00010, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00010 = slim.conv2d(
+                                                                            net00010, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                    net00011 = net0001
+                                                                    with tf.name_scope('nn.Sequential'):
+                                                                        net00011 = tf.pad(net00011, np.array(
+                                                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                        net00011 = slim.conv2d(
+                                                                            net00011, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                net0001 = tf.add_n(
+                                                                    [net00010, net00011])
+                                                    net000 = tf.add_n(
+                                                        [net0000, net0001])
+                                                    with tf.name_scope('nn.Sequential'):
+                                                        with tf.name_scope('nn.ConcatTable'):
+                                                            net0000 = net000
+                                                            with tf.name_scope('nn.Sequential'):
+                                                                net0000 = slim.batch_norm(
+                                                                    net0000)
+                                                                net0000 = slim.nn.relu(
+                                                                    net0000)
+                                                                net0000 = tf.pad(net0000, np.array(
+                                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                net0000 = slim.conv2d(
+                                                                    net0000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                                net0000 = slim.batch_norm(
+                                                                    net0000)
+                                                                net0000 = slim.nn.relu(
+                                                                    net0000)
+                                                                net0000 = tf.pad(net0000, np.array(
+                                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                                net0000 = slim.conv2d(
+                                                                    net0000, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                                net0000 = slim.batch_norm(
+                                                                    net0000)
+                                                                net0000 = slim.nn.relu(
+                                                                    net0000)
+                                                                net0000 = tf.pad(net0000, np.array(
+                                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                                net0000 = slim.conv2d(
+                                                                    net0000, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0001 = net000
+                                                        net000 = tf.add_n(
+                                                            [net0000, net0001])
+                                                net000 = tf.pad(net000, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net000 = deconv_layer(
+                                                    net000, 2, 512, method=deconv)
+                                            net001 = net00
+                                            with tf.name_scope('nn.Sequential'):
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0010 = net001
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0011 = net001
+                                                    net001 = tf.add_n(
+                                                        [net0010, net0011])
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0010 = net001
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0011 = net001
+                                                    net001 = tf.add_n(
+                                                        [net0010, net0011])
+                                                with tf.name_scope('nn.Sequential'):
+                                                    with tf.name_scope('nn.ConcatTable'):
+                                                        net0010 = net001
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                            net0010 = slim.batch_norm(
+                                                                net0010)
+                                                            net0010 = slim.nn.relu(
+                                                                net0010)
+                                                            net0010 = tf.pad(net0010, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0010 = slim.conv2d(
+                                                                net0010, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                        net0011 = net001
+                                                        with tf.name_scope('nn.Sequential'):
+                                                            net0011 = tf.pad(net0011, np.array(
+                                                                [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                            net0011 = slim.conv2d(
+                                                                net0011, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                    net001 = tf.add_n(
+                                                        [net0010, net0011])
+                                        net00 = tf.add_n([net000, net001])
+                                        with tf.name_scope('nn.Sequential'):
+                                            with tf.name_scope('nn.ConcatTable'):
+                                                net000 = net00
+                                                with tf.name_scope('nn.Sequential'):
+                                                    net000 = slim.batch_norm(
+                                                        net000)
+                                                    net000 = slim.nn.relu(
+                                                        net000)
+                                                    net000 = tf.pad(net000, np.array(
+                                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                    net000 = slim.conv2d(
+                                                        net000, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                    net000 = slim.batch_norm(
+                                                        net000)
+                                                    net000 = slim.nn.relu(
+                                                        net000)
+                                                    net000 = tf.pad(net000, np.array(
+                                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                    net000 = slim.conv2d(
+                                                        net000, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                    net000 = slim.batch_norm(
+                                                        net000)
+                                                    net000 = slim.nn.relu(
+                                                        net000)
+                                                    net000 = tf.pad(net000, np.array(
+                                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                    net000 = slim.conv2d(
+                                                        net000, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net001 = net00
+                                            net00 = tf.add_n([net000, net001])
+                                    net00 = tf.pad(net00, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net00 = deconv_layer(
+                                        net00, 2, 512, method=deconv)
+                                net01 = net0
+                                with tf.name_scope('nn.Sequential'):
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net010 = net01
+                                            with tf.name_scope('nn.Sequential'):
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net011 = net01
+                                        net01 = tf.add_n([net010, net011])
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net010 = net01
+                                            with tf.name_scope('nn.Sequential'):
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net011 = net01
+                                        net01 = tf.add_n([net010, net011])
+                                    with tf.name_scope('nn.Sequential'):
+                                        with tf.name_scope('nn.ConcatTable'):
+                                            net010 = net01
+                                            with tf.name_scope('nn.Sequential'):
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                                net010 = slim.batch_norm(
+                                                    net010)
+                                                net010 = slim.nn.relu(net010)
+                                                net010 = tf.pad(net010, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net010 = slim.conv2d(
+                                                    net010, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                            net011 = net01
+                                            with tf.name_scope('nn.Sequential'):
+                                                net011 = tf.pad(net011, np.array(
+                                                    [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                                net011 = slim.conv2d(
+                                                    net011, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                        net01 = tf.add_n([net010, net011])
+                            net0 = tf.add_n([net00, net01])
+                            with tf.name_scope('nn.Sequential'):
+                                with tf.name_scope('nn.ConcatTable'):
+                                    net00 = net0
+                                    with tf.name_scope('nn.Sequential'):
+                                        net00 = slim.batch_norm(net00)
+                                        net00 = slim.nn.relu(net00)
+                                        net00 = tf.pad(net00, np.array(
+                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                        net00 = slim.conv2d(
+                                            net00, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                        net00 = slim.batch_norm(net00)
+                                        net00 = slim.nn.relu(net00)
+                                        net00 = tf.pad(net00, np.array(
+                                            [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                        net00 = slim.conv2d(
+                                            net00, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                        net00 = slim.batch_norm(net00)
+                                        net00 = slim.nn.relu(net00)
+                                        net00 = tf.pad(net00, np.array(
+                                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                        net00 = slim.conv2d(
+                                            net00, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net01 = net0
+                                net0 = tf.add_n([net00, net01])
+                        net0 = tf.pad(net0, np.array(
+                            [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                        net0 = deconv_layer(net0, 2, 512, method=deconv)
+
+                    net1 = net
+                    with tf.name_scope('nn.Sequential'):
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net10 = net1
+                                with tf.name_scope('nn.Sequential'):
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net11 = net1
+                            net1 = tf.add_n([net10, net11])
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net10 = net1
+                                with tf.name_scope('nn.Sequential'):
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 128, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 128, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net11 = net1
+                            net1 = tf.add_n([net10, net11])
+                        with tf.name_scope('nn.Sequential'):
+                            with tf.name_scope('nn.ConcatTable'):
+                                net10 = net1
+                                with tf.name_scope('nn.Sequential'):
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 256, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [1, 1], [1, 1], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 256, (3, 3), (1, 1), activation_fn=None, padding='VALID')
+                                    net10 = slim.batch_norm(net10)
+                                    net10 = slim.nn.relu(net10)
+                                    net10 = tf.pad(net10, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net10 = slim.conv2d(
+                                        net10, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                                net11 = net1
+                                with tf.name_scope('nn.Sequential'):
+                                    net11 = tf.pad(net11, np.array(
+                                        [[0, 0], [0, 0], [0, 0], [0, 0]]))
+                                    net11 = slim.conv2d(
+                                        net11, 512, (1, 1), (1, 1), activation_fn=None, padding='VALID')
+                            net1 = tf.add_n([net10, net11])
+                net = tf.add_n([net0, net1])
+            net = tf.pad(net, np.array([[0, 0], [0, 0], [0, 0], [0, 0]]))
+            net = slim.conv2d(net, 512, (1, 1), (1, 1),
+                              activation_fn=None, padding='VALID')
+            net = slim.batch_norm(net)
+            net = slim.nn.relu(net)
+            net = tf.pad(net, np.array([[0, 0], [0, 0], [0, 0], [0, 0]]))
+            net = slim.conv2d(net, 256, (1, 1), (1, 1),
+                              activation_fn=None, padding='VALID')
+            net = slim.batch_norm(net)
+            net = slim.nn.relu(net)
+            net = tf.pad(net, np.array([[0, 0], [0, 0], [0, 0], [0, 0]]))
+            net = slim.conv2d(net, out_channels, (1, 1), (1, 1),
+                              activation_fn=None, padding='VALID')
+        net = tf.pad(net, np.array([[0, 0], [0, 0], [0, 0], [0, 0]]))
+        net = deconv_layer(net, 4, out_channels, method=deconv)
+
+    return net
diff --git a/reconstruction/ostec/external/landmark_detector/networks.py b/reconstruction/ostec/external/landmark_detector/networks.py
new file mode 100644
index 000000000..3c96fca88
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/networks.py
@@ -0,0 +1,180 @@
+import tensorflow as tf
+from external.landmark_detector import utils, models, data_provider
+
+from tensorflow.python.platform import tf_logging as logging
+
+slim = tf.contrib.slim
+
+from external.landmark_detector.flags import FLAGS
+
+# general framework
+class DeepNetwork(object):
+    def __init__(self):
+        pass
+
+    def _build_network(self, inputs, datas):
+        pass
+
+    def _build_losses(self, predictions, states, images, datas):
+        pass
+
+    def _build_summaries(self, predictions, states, images, datas):
+        tf.summary.image('images', images[:, :, :, :3], max_outputs=min(FLAGS['batch_size'], 3))
+
+    def _get_data(self):
+        provider = data_provider.ProtobuffProvider(
+            filename=FLAGS['dataset_dir'],
+            batch_size=FLAGS['batch_size'],
+            rescale=FLAGS['rescale'],
+            augmentation=FLAGS['eval_dir']=='',
+            )
+        return provider.get()
+
+    def _build_restore_fn(self, sess):
+        init_fn = None
+
+        if FLAGS['pretrained_model_checkpoint_path']:
+            print('Loading whole model ...')
+            variables_to_restore = slim.get_model_variables()
+            init_fn =  slim.assign_from_checkpoint_fn(
+                FLAGS['pretrained_model_checkpoint_path'],
+                variables_to_restore,
+                ignore_missing_vars=True)
+        return init_fn
+
+
+    def train(self):
+        g = tf.Graph()
+        logging.set_verbosity(10)
+
+        with g.as_default():
+            # Load datasets.
+
+            images, *datas = self._get_data()
+            images /= 255.
+
+            # Define model graph.
+            with tf.variable_scope('net'):
+                with slim.arg_scope([slim.batch_norm, slim.layers.dropout],
+                is_training=True):
+
+                    predictions, states = self._build_network(images, datas)
+
+                    # custom losses
+                    self._build_losses(predictions, states, images, datas)
+
+                    # total losses
+                    total_loss = slim.losses.get_total_loss()
+                    tf.summary.scalar('losses/total loss', total_loss)
+
+                    # image summaries
+                    self._build_summaries(predictions, states, images, datas)
+
+                    # learning rate decay
+                    global_step = slim.get_or_create_global_step()
+
+                    learning_rate = tf.train.exponential_decay(
+                        FLAGS['initial_learning_rate'],
+                        global_step,
+                        FLAGS['learning_rate_decay_step'] / FLAGS['batch_size'],
+                        FLAGS['learning_rate_decay_factor'],
+                        staircase=True)
+
+                    tf.summary.scalar('learning rate', learning_rate)
+
+                    optimizer = tf.train.AdamOptimizer(learning_rate)
+
+        with tf.Session(graph=g) as sess:
+            init_fn = self._build_restore_fn(sess)
+            train_op = slim.learning.create_train_op(
+                total_loss,
+                optimizer,
+                summarize_gradients=True)
+
+            logging.set_verbosity(1)
+
+            slim.learning.train(train_op,
+                FLAGS['train_dir'],
+                save_summaries_secs=60,
+                init_fn=init_fn,
+                save_interval_secs=600)
+
+class DNFaceMultiView(DeepNetwork):
+    def __init__(self, n_lms=FLAGS['n_landmarks']):
+        super(DNFaceMultiView, self).__init__()
+        self.n_lms = n_lms
+
+
+    def _get_data(self):
+        provider = data_provider.ProtobuffProvider(
+            filename=FLAGS['dataset_dir'],
+            batch_size=FLAGS['batch_size'],
+            rescale=FLAGS['rescale'],
+            augmentation=FLAGS['eval_dir']=='',
+            )
+        return provider.get()
+
+
+    def _build_network(self, inputs, datas=None, n_stacks=1, n_channels=FLAGS['n_landmarks'], is_training=True):
+        # gt_heatmap, gt_lms, mask_index, gt_mask = datas
+
+        batch_size = tf.shape(inputs)[0]
+        height = tf.shape(inputs)[1]
+        width = tf.shape(inputs)[2]
+
+        net = inputs
+
+        # net = models.StackedHourglass(net, FLAGS.n_landmarks)
+        # states.append(net)
+        # net = tf.stop_gradient(net)
+        # net *= gt_mask[:,None,None,:]
+        # net = tf.concat([inputs,net], 3)
+        # net = models.StackedHourglass(net, FLAGS.n_landmarks)
+        # states.append(net)
+
+        batch_size = tf.shape(inputs)[0]
+        height = tf.shape(inputs)[1]
+        width = tf.shape(inputs)[2]
+        channels = tf.shape(inputs)[3]
+
+        states = []
+
+        with slim.arg_scope([slim.batch_norm, slim.layers.dropout], is_training=is_training):
+            with slim.arg_scope(models.hourglass_arg_scope_tf()):
+                net = None
+                # stacked hourglass
+                for i in range(n_stacks):
+                    with tf.variable_scope('stack_%02d' % i):
+                        if net is not None:
+                            net = tf.concat((inputs, net), 3)
+                        else:
+                            net = inputs
+
+                        net, _ = models.hourglass(
+                            net,
+                            regression_channels=n_channels,
+                            classification_channels=0,
+                            deconv='transpose',
+                            bottleneck='bottleneck_inception')
+
+                        states.append(net)
+
+                prediction = net
+                return prediction, states
+
+    def _build_losses(self, predictions, states, images, datas):
+        gt_heatmap, gt_lms, mask_index, gt_mask = datas
+
+        weight_hm = utils.get_weight(gt_heatmap, tf.ones_like(gt_heatmap), ng_w=0.1, ps_w=1) * 500
+        weight_hm *= gt_mask[:,None,None,:]
+
+        l2norm = slim.losses.mean_squared_error(states[0], gt_heatmap, weights=weight_hm)
+
+        tf.summary.scalar('losses/lms_pred', l2norm)
+
+    def _build_summaries(self, predictions, states, images, datas):
+        super()._build_summaries(predictions, states, images, datas)
+
+        gt_heatmap, gt_lms, mask_index, gt_mask = datas
+
+        tf.summary.image('predictions/landmark-regression', tf.reduce_sum(predictions, -1)[...,None] * 255.0, max_outputs=min(FLAGS['batch_size'],3))
diff --git a/reconstruction/ostec/external/landmark_detector/utils.py b/reconstruction/ostec/external/landmark_detector/utils.py
new file mode 100644
index 000000000..93a3e9c5b
--- /dev/null
+++ b/reconstruction/ostec/external/landmark_detector/utils.py
@@ -0,0 +1,421 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import tensorflow as tf
+import menpo.io as mio
+from menpo.image import Image
+from menpo.shape import PointCloud
+import cv2
+
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.contrib.framework.python.ops import variables
+from tensorflow.python.training import optimizer as tf_optimizer
+from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import variables as tf_variables
+
+from menpo.transform import Translation, Scale
+from menpo.shape import PointCloud
+
+slim = tf.contrib.slim
+
+def generate_heatmap(logits, num_classes):
+    """Generates a coloured heatmap from the keypoint logits.
+
+    Args:
+        features: A `Tensor` of dimensions [num_batch, height, width, FLAGS.n_landmarks + 1].
+    """
+
+    keypoint_colours = np.array([plt.cm.spectral(x) for x in np.linspace(0, 1, num_classes + 1)])[
+        ..., :3].astype(np.float32)
+
+    prediction = tf.nn.softmax(logits)
+    heatmap = tf.matmul(tf.reshape(prediction, (-1, num_classes + 1)), keypoint_colours)
+    heatmap = tf.reshape(heatmap, (tf.shape(prediction)[0],
+                                   tf.shape(prediction)[1],
+                                   tf.shape(prediction)[2], 3))
+    return heatmap
+
+def generate_landmarks(keypoints):
+    is_background = tf.equal(keypoints, 0)
+    ones = tf.to_float(tf.ones_like(is_background))
+    zeros = tf.to_float(tf.zeros_like(is_background))
+
+    return tf.where(is_background, zeros, ones) * 255
+
+def project_landmarks_to_shape_model(landmarks):
+    final = []
+
+    for lms in landmarks:
+        lms = PointCloud(lms)
+        similarity = AlignmentSimilarity(pca.global_transform.source, lms)
+        projected_target = similarity.pseudoinverse().apply(lms)
+        target = pca.model.reconstruct(projected_target)
+        target = similarity.apply(target)
+        final.append(target.points)
+
+    return np.array(final).astype(np.float32)
+
+def rescale_image(image, stride_width=64):
+    # make sure smallest size is 600 pixels wide & dimensions are (k * stride_width) + 1
+    height, width = image.shape
+
+    # Taken from 'szross'
+    scale_up = 625. / min(height, width)
+    scale_cap = 961. / max(height, width)
+    scale_up  = min(scale_up, scale_cap)
+    new_height = stride_width * round((height * scale_up) / stride_width) + 1
+    new_width = stride_width * round((width * scale_up) / stride_width) + 1
+    image, tr = image.resize([new_height, new_width], return_transform=True)
+    image.inverse_tr = tr
+    return image
+
+def frankotchellappa(dzdx, dzdy):
+    from numpy.fft import ifftshift, fft2, ifft2
+    rows, cols = dzdx.shape
+    # The following sets up matrices specifying frequencies in the x and y
+    # directions corresponding to the Fourier transforms of the gradient
+    # data.  They range from -0.5 cycles/pixel to + 0.5 cycles/pixel.
+    # The scaling of this is irrelevant as long as it represents a full
+    # circle domain. This is functionally equivalent to any constant * pi
+    pi_over_2 = np.pi / 2.0
+    row_grid = np.linspace(-pi_over_2, pi_over_2, rows)
+    col_grid = np.linspace(-pi_over_2, pi_over_2, cols)
+    wy, wx = np.meshgrid(row_grid, col_grid, indexing='ij')
+
+    # Quadrant shift to put zero frequency at the appropriate edge
+    wx = ifftshift(wx)
+    wy = ifftshift(wy)
+
+    # Fourier transforms of gradients
+    DZDX = fft2(dzdx)
+    DZDY = fft2(dzdy)
+
+    # Integrate in the frequency domain by phase shifting by pi/2 and
+    # weighting the Fourier coefficients by their frequencies in x and y and
+    # then dividing by the squared frequency
+    denom = (wx ** 2 + wy ** 2)
+    Z = (-1j * wx * DZDX - 1j * wy * DZDY) / denom
+    Z = np.nan_to_num(Z)
+    return np.real(ifft2(Z))
+
+def line(image, x0, y0, x1, y1, color):
+    steep = False
+    if x0 < 0 or x0 >= 400 or x1 < 0 or x1 >= 400 or y0 < 0 or y0 >= 400 or y1 < 0 or y1 >= 400:
+        return
+
+    if abs(x0 - x1) < abs(y0 - y1):
+        x0, y0 = y0, x0
+        x1, y1 = y1, x1
+        steep = True
+
+    if x0 > x1:
+        x0, x1 = x1, x0
+        y0, y1 = y1, y0
+
+    for x in range(int(x0), int(x1) + 1):
+        t = (x - x0) / float(x1 - x0)
+        y = y0 * (1 - t) + y1 * t
+        if steep:
+            image[x, int(y)] = color
+        else:
+            image[int(y), x] = color
+
+def draw_landmarks(img, lms):
+    try:
+        img = img.copy()
+
+        for i, part in enumerate(parts_68[1:]):
+            circular = []
+
+            if i in (4, 5, 6, 7):
+                circular = [part[0]]
+
+            for p1, p2 in zip(part, list(part[1:]) + circular):
+                p1, p2 = lms[p1], lms[p2]
+
+                line(img, p2[1], p2[0], p1[1], p1[0], 1)
+    except:
+        pass
+    return img
+
+def batch_draw_landmarks(imgs, lms):
+    return np.array([draw_landmarks(img, l) for img, l in zip(imgs, lms)])
+
+def build_graph(inputs, tree, transpose=(2,3,1,0), layers=[]):
+    net = inputs
+
+    if tree['name'] == 'nn.Sequential':
+        with tf.name_scope('nn.Sequential'):
+            for tr in tree['children']:
+                net = build_graph(net, tr, transpose, layers)
+    elif tree['name'] == 'nn.ConcatTable':
+        net_table = []
+        with tf.name_scope('nn.ConcatTable'):
+            for tr in tree['children']:
+                net_table.append(build_graph(net, tr, transpose, layers))
+        net = net_table
+    elif tree['name'] == 'nn.JoinTable':
+        net = tf.concat(3, net)
+    elif tree['name'] == 'nn.CAddTable':
+        net = tf.add_n(net)
+    elif tree['name'] == 'nn.SpatialConvolution':
+        out_channel = int(tree['nOutputPlane'])
+        kernal_shape = (int(tree['kH']),int(tree['kW']))
+        stride_shape = (int(tree['dH']),int(tree['dW']))
+        net = tf.pad(
+                net, [
+                    [0,0],
+                    [int(tree['padH']),int(tree['padH'])],
+                    [int(tree['padW']),int(tree['padW'])],
+                    [0,0]
+                ])
+        if 'weight' in tree.keys() and 'bias' in tree.keys():
+            net = slim.conv2d(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID',
+                              weights_initializer=tf.constant_initializer(tree['weight'].transpose(*transpose)),
+                              biases_initializer=tf.constant_initializer(tree['bias'])
+                             )
+        else:
+            net = slim.conv2d(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID'
+                             )
+
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.SpatialFullConvolution':
+        out_channel = int(tree['nOutputPlane'])
+        kernal_shape = (int(tree['kH']),int(tree['kW']))
+        stride_shape = (int(tree['dH']),int(tree['dW']))
+        net = tf.pad(
+                net, [
+                    [0,0],
+                    [int(tree['padH']),int(tree['padH'])],
+                    [int(tree['padW']),int(tree['padW'])],
+                    [0,0]
+                ])
+        if 'weight' in tree.keys() and 'bias' in tree.keys():
+            net = slim.conv2d_transpose(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID',
+                              weights_initializer=tf.constant_initializer(tree['weight'].transpose(*transpose)),
+                              biases_initializer=tf.constant_initializer(tree['bias'])
+                             )
+        else:
+            net = slim.conv2d_transpose(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID'
+                             )
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+
+    elif tree['name'] == 'nn.SpatialBatchNormalization':
+        net = slim.nn.batch_normalization(net,
+                                     tree['running_mean'],
+                                     tree['running_var'],
+                                     tree['bias'],
+                                     tree['weight'],
+                                     tree['eps'])
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.ReLU':
+        net = slim.nn.relu(net)
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.Sigmoid':
+        net = slim.nn.sigmoid(net)
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.SpatialMaxPooling':
+        net = slim.max_pool2d(
+            tf.pad(
+                net, [
+                    [0,0],
+                    [int(tree['padH']),int(tree['padH'])],
+                    [int(tree['padW']),int(tree['padW'])],
+                    [0,0]
+                ]),
+            (int(tree['kH']),int(tree['kW'])),
+            (int(tree['dH']),int(tree['dW']))
+        )
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.Identity':
+        pass
+    else:
+        raise Exception(tree['name'])
+
+    return net
+
+def build_graph_old(inputs, tree, transpose=(2,3,1,0)):
+    net = inputs
+
+    if tree['name'] == 'nn.Sequential':
+        with tf.name_scope('nn.Sequential'):
+            for tr in tree['children']:
+                net = build_graph(net, tr, transpose, layers)
+    elif tree['name'] == 'nn.ConcatTable':
+        net_table = []
+        with tf.name_scope('nn.ConcatTable'):
+            for tr in tree['children']:
+                net_table.append(build_graph(net, tr, transpose, layers))
+        net = net_table
+    elif tree['name'] == 'nn.JoinTable':
+        net = tf.concat(3, net)
+    elif tree['name'] == 'nn.CAddTable':
+        net = tf.add_n(net)
+    elif tree['name'] == 'nn.SpatialConvolution':
+        out_channel = int(tree['nOutputPlane'])
+        kernal_shape = (int(tree['kH']),int(tree['kW']))
+        stride_shape = (int(tree['dH']),int(tree['dW']))
+        net = tf.pad(
+                net, [
+                    [0,0],
+                    [int(tree['padH']),int(tree['padH'])],
+                    [int(tree['padW']),int(tree['padW'])],
+                    [0,0]
+                ])
+        if 'weight' in tree.keys() and 'bias' in tree.keys():
+            net = slim.conv2d(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID',
+                              weights_initializer=tf.constant_initializer(tree['weight'].transpose(*transpose)),
+                              biases_initializer=tf.constant_initializer(tree['bias'])
+                             )
+        else:
+            net = slim.conv2d(net,
+                              out_channel,
+                              kernal_shape,
+                              stride_shape,
+                              activation_fn=None,
+                              padding='VALID'
+                             )
+
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.SpatialFullConvolution':
+        out_channel = int(tree['nOutputPlane'])
+        kernal_shape = (int(tree['kH']),int(tree['kW']))
+        rate = np.min(int(tree['dH']),int(tree['dW']))
+        h,w = tf.shape(net)[1:3]
+        net = tf.image.resize_bilinear(net, (h,w,out_channel))
+
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+
+    elif tree['name'] == 'nn.SpatialBatchNormalization':
+        net = slim.batch_norm(net)
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.ReLU':
+        net = slim.nn.relu(net)
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.Sigmoid':
+        net = slim.nn.sigmoid(net)
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.SpatialMaxPooling':
+        net = slim.max_pool2d(
+            tf.pad(
+                net, [
+                    [0,0],
+                    [int(tree['padH']),int(tree['padH'])],
+                    [int(tree['padW']),int(tree['padW'])],
+                    [0,0]
+                ]),
+            (int(tree['kH']),int(tree['kW'])),
+            (int(tree['dH']),int(tree['dW']))
+        )
+        tree['tfname'] = net.name
+        tree['tfvar'] = net
+    elif tree['name'] == 'nn.Identity':
+        pass
+    else:
+        raise Exception(tree['name'])
+
+    return net
+
+def keypts_encoding(keypoints, num_classes):
+    keypoints = tf.to_int32(keypoints)
+    keypoints = tf.reshape(keypoints, (-1,))
+    keypoints = slim.layers.one_hot_encoding(keypoints, num_classes=num_classes+1)
+    return keypoints
+
+def get_weight(keypoints, mask=None, ng_w=0.01, ps_w=1.0):
+    is_background = tf.equal(keypoints, 0)
+    ones = tf.to_float(tf.ones_like(is_background))
+    weights = tf.where(is_background, ones * ng_w, ones*ps_w)
+    # if mask is not None:
+    #     weights *= tf.to_float(mask)
+
+    return weights
+
+def ced_accuracy(t, dists):
+    # Head	 Shoulder	Elbow	Wrist	Hip	   Knee	   Ankle
+    pts_r  = tf.transpose(tf.gather(tf.transpose(dists), [8,12,11,10,2,1,0]))
+    pts_l  = tf.transpose(tf.gather(tf.transpose(dists), [9,13,14,15,3,4,5]))
+    part_pckh = (tf.to_int32(pts_r <= t) + tf.to_int32(pts_l <= t)) / 2
+
+    return tf.concat(1, [part_pckh, tf.reduce_sum(tf.to_int32(dists <= t), 1)[...,None] / tf.shape(dists)[1]])
+
+def pckh(preds, gts, scales):
+    t_range = np.arange(0,0.51,0.01)
+    dists = tf.sqrt(tf.reduce_sum(tf.pow(preds - gts, 2), reduction_indices=-1)) / scales
+    # pckh = [ced_accuracy(t, dists) for t in t_range]
+    # return pckh[-1]
+    return ced_accuracy(0.5, dists)
+
+def import_image(img_path):
+    img = cv2.imread(str(img_path))
+    original_image = Image.init_from_channels_at_back(img[:,:,-1::-1])
+
+    try:
+        original_image_lms = mio.import_landmark_file('{}/{}.ljson'.format(img_path.parent, img_path.stem)).lms.points.astype(np.float32)
+        original_image.landmarks['LJSON'] = PointCloud(original_image_lms)
+    except:
+        pass
+
+    return original_image
+
+def crop_image(img, center, scale, res, base=384):
+    h = base * scale
+
+    t = Translation(
+        [
+            res[0] * (-center[0] / h + .5),
+            res[1] * (-center[1] / h + .5)
+        ]).compose_after(Scale((res[0] / h, res[1] / h))).pseudoinverse()
+
+
+    # Upper left point
+    ul = np.floor(t.apply([0,0]))
+    # Bottom right point
+    br = np.ceil(t.apply(res).astype(np.int))
+
+    # crop and rescale
+
+    cimg, trans = img.warp_to_shape(br-ul, Translation(-(br-ul)/2+(br+ul)/2) ,return_transform=True)
+    c_scale = np.min(cimg.shape) / np.mean(res)
+    new_img = cimg.rescale(1 / c_scale).resize(res)
+    return new_img, trans, c_scale
diff --git a/reconstruction/ostec/external/stylegan2/Dockerfile b/reconstruction/ostec/external/stylegan2/Dockerfile
new file mode 100644
index 000000000..ab45a553e
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/Dockerfile
@@ -0,0 +1,11 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+FROM tensorflow/tensorflow:1.15.0-gpu-py3
+
+RUN pip install scipy==1.3.3
+RUN pip install requests==2.22.0
+RUN pip install Pillow==6.2.1
diff --git a/reconstruction/ostec/external/stylegan2/LICENSE.txt b/reconstruction/ostec/external/stylegan2/LICENSE.txt
new file mode 100644
index 000000000..288fb3247
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/LICENSE.txt
@@ -0,0 +1,101 @@
+Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+
+
+Nvidia Source Code License-NC
+
+=======================================================================
+
+1. Definitions
+
+"Licensor" means any person or entity that distributes its Work.
+
+"Software" means the original work of authorship made available under
+this License.
+
+"Work" means the Software and any additions to or derivative works of
+the Software that are made available under this License.
+
+"Nvidia Processors" means any central processing unit (CPU), graphics
+processing unit (GPU), field-programmable gate array (FPGA),
+application-specific integrated circuit (ASIC) or any combination
+thereof designed, made, sold, or provided by Nvidia or its affiliates.
+
+The terms "reproduce," "reproduction," "derivative works," and
+"distribution" have the meaning as provided under U.S. copyright law;
+provided, however, that for the purposes of this License, derivative
+works shall not include works that remain separable from, or merely
+link (or bind by name) to the interfaces of, the Work.
+
+Works, including the Software, are "made available" under this License
+by including in or with the Work either (a) a copyright notice
+referencing the applicability of this License to the Work, or (b) a
+copy of this License.
+
+2. License Grants
+
+    2.1 Copyright Grant. Subject to the terms and conditions of this
+    License, each Licensor grants to you a perpetual, worldwide,
+    non-exclusive, royalty-free, copyright license to reproduce,
+    prepare derivative works of, publicly display, publicly perform,
+    sublicense and distribute its Work and any resulting derivative
+    works in any form.
+
+3. Limitations
+
+    3.1 Redistribution. You may reproduce or distribute the Work only
+    if (a) you do so under this License, (b) you include a complete
+    copy of this License with your distribution, and (c) you retain
+    without modification any copyright, patent, trademark, or
+    attribution notices that are present in the Work.
+
+    3.2 Derivative Works. You may specify that additional or different
+    terms apply to the use, reproduction, and distribution of your
+    derivative works of the Work ("Your Terms") only if (a) Your Terms
+    provide that the use limitation in Section 3.3 applies to your
+    derivative works, and (b) you identify the specific derivative
+    works that are subject to Your Terms. Notwithstanding Your Terms,
+    this License (including the redistribution requirements in Section
+    3.1) will continue to apply to the Work itself.
+
+    3.3 Use Limitation. The Work and any derivative works thereof only
+    may be used or intended for use non-commercially. The Work or
+    derivative works thereof may be used or intended for use by Nvidia
+    or its affiliates commercially or non-commercially. As used herein,
+    "non-commercially" means for research or evaluation purposes only.
+
+    3.4 Patent Claims. If you bring or threaten to bring a patent claim
+    against any Licensor (including any claim, cross-claim or
+    counterclaim in a lawsuit) to enforce any patents that you allege
+    are infringed by any Work, then your rights under this License from
+    such Licensor (including the grants in Sections 2.1 and 2.2) will
+    terminate immediately.
+
+    3.5 Trademarks. This License does not grant any rights to use any
+    Licensor's or its affiliates' names, logos, or trademarks, except
+    as necessary to reproduce the notices described in this License.
+
+    3.6 Termination. If you violate any term of this License, then your
+    rights under this License (including the grants in Sections 2.1 and
+    2.2) will terminate immediately.
+
+4. Disclaimer of Warranty.
+
+THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR
+NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER
+THIS LICENSE. 
+
+5. Limitation of Liability.
+
+EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
+THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
+SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
+INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
+OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
+(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
+LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
+COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGES.
+
+=======================================================================
diff --git a/reconstruction/ostec/external/stylegan2/README.md b/reconstruction/ostec/external/stylegan2/README.md
new file mode 100644
index 000000000..075f0b67d
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/README.md
@@ -0,0 +1,220 @@
+## StyleGAN2 &mdash; Official TensorFlow Implementation
+
+![Teaser image](./docs/stylegan2-teaser-1024x256.png)
+
+**Analyzing and Improving the Image Quality of StyleGAN**<br>
+Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, Timo Aila<br>
+
+Paper: http://arxiv.org/abs/1912.04958<br>
+Video: https://youtu.be/c-NJtV9Jvp0<br>
+
+Abstract: *The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent vectors to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably detect if an image is generated by a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.*
+
+For business inquiries, please visit our website and submit the form: [NVIDIA Research Licensing](https://www.nvidia.com/en-us/research/inquiries/)
+
+**&#9733;&#9733;&#9733; NEW: [StyleGAN2-ADA-PyTorch](https://github.com/NVlabs/stylegan2-ada-pytorch) is now available; see the full list of versions [here](https://nvlabs.github.io/stylegan2/versions.html) &#9733;&#9733;&#9733;**
+
+| Additional material | &nbsp;
+| :--- | :----------
+| [StyleGAN2](https://drive.google.com/open?id=1QHc-yF5C3DChRwSdZKcx1w6K8JvSxQi7) | Main Google Drive folder
+| &boxvr;&nbsp; [stylegan2-paper.pdf](https://drive.google.com/open?id=1fnF-QsiQeKaxF-HbvFiGtzHF_Bf3CzJu) | High-quality version of the paper
+| &boxvr;&nbsp; [stylegan2-video.mp4](https://drive.google.com/open?id=1f_gbKW6FUUHKkUxciJ_lQx29mCq_fSBy) | High-quality version of the video
+| &boxvr;&nbsp; [images](https://drive.google.com/open?id=1Sak157_DLX84ytqHHqZaH_59HoEWzfB7) | Example images produced using our method
+| &boxv;&nbsp; &boxvr;&nbsp;  [curated-images](https://drive.google.com/open?id=1ydWb8xCHzDKMTW9kQ7sL-B1R0zATHVHp) | Hand-picked images showcasing our results
+| &boxv;&nbsp; &boxur;&nbsp;  [100k-generated-images](https://drive.google.com/open?id=1BA2OZ1GshdfFZGYZPob5QWOGBuJCdu5q) | Random images with and without truncation
+| &boxvr;&nbsp; [videos](https://drive.google.com/open?id=1yXDV96SFXoUiZKU7AyE6DyKgDpIk4wUZ) | Individual clips of the video as high-quality MP4
+| &boxur;&nbsp; [networks](https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/) | Pre-trained networks
+| &ensp;&ensp; &boxvr;&nbsp;  stylegan2-ffhq-config-f.pkl | StyleGAN2 for <span style="font-variant:small-caps">FFHQ</span> dataset at 1024&times;1024
+| &ensp;&ensp; &boxvr;&nbsp;  stylegan2-car-config-f.pkl | StyleGAN2 for <span style="font-variant:small-caps">LSUN Car</span> dataset at 512&times;384
+| &ensp;&ensp; &boxvr;&nbsp;  stylegan2-cat-config-f.pkl | StyleGAN2 for <span style="font-variant:small-caps">LSUN Cat</span> dataset at 256&times;256
+| &ensp;&ensp; &boxvr;&nbsp;  stylegan2-church-config-f.pkl | StyleGAN2 for <span style="font-variant:small-caps">LSUN Church</span> dataset at 256&times;256
+| &ensp;&ensp; &boxvr;&nbsp;  stylegan2-horse-config-f.pkl | StyleGAN2 for <span style="font-variant:small-caps">LSUN Horse</span> dataset at 256&times;256
+| &ensp;&ensp; &boxur;&nbsp;&#x22ef;  | Other training configurations used in the paper
+
+## Requirements
+
+* Both Linux and Windows are supported. Linux is recommended for performance and compatibility reasons.
+* 64-bit Python 3.6 installation. We recommend Anaconda3 with numpy 1.14.3 or newer.
+* We recommend TensorFlow 1.14, which we used for all experiments in the paper, but TensorFlow 1.15 is also supported on Linux. TensorFlow 2.x is not supported.
+* On Windows you need to use TensorFlow 1.14, as the standard 1.15 installation does not include necessary C++ headers.
+* One or more high-end NVIDIA GPUs, NVIDIA drivers, CUDA 10.0 toolkit and cuDNN 7.5. To reproduce the results reported in the paper, you need an NVIDIA GPU with at least 16 GB of DRAM.
+* Docker users: use the [provided Dockerfile](./Dockerfile) to build an image with the required library dependencies.
+
+StyleGAN2 relies on custom TensorFlow ops that are compiled on the fly using [NVCC](https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html). To test that your NVCC installation is working correctly, run:
+
+```.bash
+nvcc test_nvcc.cu -o test_nvcc -run
+| CPU says hello.
+| GPU says hello.
+```
+
+On Windows, the compilation requires Microsoft Visual Studio to be in `PATH`. We recommend installing [Visual Studio Community Edition](https://visualstudio.microsoft.com/vs/) and adding into `PATH` using `"C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat"`.
+
+## Using pre-trained networks
+
+Pre-trained networks are stored as `*.pkl` files on the [StyleGAN2 Google Drive folder](https://drive.google.com/open?id=1QHc-yF5C3DChRwSdZKcx1w6K8JvSxQi7). Below, you can either reference them directly using the syntax `gdrive:networks/<filename>.pkl`, or download them manually and reference by filename.
+
+```.bash
+# Generate uncurated ffhq images (matches paper Figure 12)
+python run_generator.py generate-images --network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
+  --seeds=6600-6625 --truncation-psi=0.5
+
+# Generate curated ffhq images (matches paper Figure 11)
+python run_generator.py generate-images --network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
+  --seeds=66,230,389,1518 --truncation-psi=1.0
+
+# Generate uncurated car images
+python run_generator.py generate-images --network=gdrive:networks/stylegan2-car-config-f.pkl \
+  --seeds=6000-6025 --truncation-psi=0.5
+
+# Example of style mixing (matches the corresponding video clip)
+python run_generator.py style-mixing-example --network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
+  --row-seeds=85,100,75,458,1500 --col-seeds=55,821,1789,293 --truncation-psi=1.0
+```
+
+The results are placed in `results/<RUNNING_ID>/*.png`. You can change the location with `--result-dir`. For example, `--result-dir=~/my-stylegan2-results`.
+
+You can import the networks in your own Python code using `pickle.load()`. For this to work, you need to include the `dnnlib` source directory in `PYTHONPATH` and create a default TensorFlow session by calling `dnnlib.tflib.init_tf()`. See [run_generator.py](./run_generator.py) and [pretrained_networks.py](./pretrained_networks.py) for examples.
+
+## Preparing datasets
+
+Datasets are stored as multi-resolution TFRecords, similar to the [original StyleGAN](https://github.com/NVlabs/stylegan). Each dataset consists of multiple `*.tfrecords` files stored under a common directory, e.g., `~/datasets/ffhq/ffhq-r*.tfrecords`. In the following sections, the datasets are referenced using a combination of `--dataset` and `--data-dir` arguments, e.g., `--dataset=ffhq --data-dir=~/datasets`.
+
+**FFHQ**. To download the [Flickr-Faces-HQ](https://github.com/NVlabs/ffhq-dataset) dataset as multi-resolution TFRecords, run:
+
+```.bash
+pushd ~
+git clone https://github.com/NVlabs/ffhq-dataset.git
+cd ffhq-dataset
+python download_ffhq.py --tfrecords
+popd
+python dataset_tool.py display ~/ffhq-dataset/tfrecords/ffhq
+```
+
+**LSUN**. Download the desired LSUN categories in LMDB format from the [LSUN project page](https://www.yf.io/p/lsun). To convert the data to multi-resolution TFRecords, run:
+
+```.bash
+python dataset_tool.py create_lsun_wide ~/datasets/car ~/lsun/car_lmdb --width=512 --height=384
+python dataset_tool.py create_lsun ~/datasets/cat ~/lsun/cat_lmdb --resolution=256
+python dataset_tool.py create_lsun ~/datasets/church ~/lsun/church_outdoor_train_lmdb --resolution=256
+python dataset_tool.py create_lsun ~/datasets/horse ~/lsun/horse_lmdb --resolution=256
+```
+
+**Custom**. Create custom datasets by placing all training images under a single directory. The images must be square-shaped and they must all have the same power-of-two dimensions. To convert the images to multi-resolution TFRecords, run:
+
+```.bash
+python dataset_tool.py create_from_images ~/datasets/my-custom-dataset ~/my-custom-images
+python dataset_tool.py display ~/datasets/my-custom-dataset
+```
+
+## Projecting images to latent space
+
+To find the matching latent vectors for a set of images, run:
+
+```.bash
+# Project generated images
+python run_projector.py project-generated-images --network=gdrive:networks/stylegan2-car-config-f.pkl \
+  --seeds=0,1,5
+
+# Project real images
+python run_projector.py project-real-images --network=gdrive:networks/stylegan2-car-config-f.pkl \
+  --dataset=car --data-dir=~/datasets
+```
+
+## Training networks
+
+To reproduce the training runs for config F in Tables 1 and 3, run:
+
+```.bash
+python run_training.py --num-gpus=8 --data-dir=~/datasets --config=config-f \
+  --dataset=ffhq --mirror-augment=true
+python run_training.py --num-gpus=8 --data-dir=~/datasets --config=config-f \
+  --dataset=car --total-kimg=57000
+python run_training.py --num-gpus=8 --data-dir=~/datasets --config=config-f \
+  --dataset=cat --total-kimg=88000
+python run_training.py --num-gpus=8 --data-dir=~/datasets --config=config-f \
+  --dataset=church --total-kimg 88000 --gamma=100
+python run_training.py --num-gpus=8 --data-dir=~/datasets --config=config-f \
+  --dataset=horse --total-kimg 100000 --gamma=100
+```
+
+For other configurations, see `python run_training.py --help`.
+
+We have verified that the results match the paper when training with 1, 2, 4, or 8 GPUs. Note that training FFHQ at 1024&times;1024 resolution requires GPU(s) with at least 16 GB of memory. The following table lists typical training times using NVIDIA DGX-1 with 8 Tesla V100 GPUs:
+
+| Configuration | Resolution      | Total kimg | 1 GPU   | 2 GPUs  | 4 GPUs  | 8 GPUs | GPU mem |
+| :------------ | :-------------: | :--------: | :-----: | :-----: | :-----: | :----: | :-----: |
+| `config-f`    | 1024&times;1024 | 25000      | 69d 23h | 36d 4h  | 18d 14h | 9d 18h | 13.3 GB |
+| `config-f`    | 1024&times;1024 | 10000      | 27d 23h | 14d 11h | 7d 10h  | 3d 22h | 13.3 GB |
+| `config-e`    | 1024&times;1024 | 25000      | 35d 11h | 18d 15h | 9d 15h  | 5d 6h  | 8.6 GB  |
+| `config-e`    | 1024&times;1024 | 10000      | 14d 4h  | 7d 11h  | 3d 20h  | 2d 3h  | 8.6 GB  |
+| `config-f`    | 256&times;256   | 25000      | 32d 13h | 16d 23h | 8d 21h  | 4d 18h | 6.4 GB  |
+| `config-f`    | 256&times;256   | 10000      | 13d 0h  | 6d 19h  | 3d 13h  | 1d 22h | 6.4 GB  |
+
+Training curves for FFHQ config F (StyleGAN2) compared to original StyleGAN using 8 GPUs:
+
+![Training curves](./docs/stylegan2-training-curves.png)
+
+After training, the resulting networks can be used the same way as the official pre-trained networks:
+
+```.bash
+# Generate 1000 random images without truncation
+python run_generator.py generate-images --seeds=0-999 --truncation-psi=1.0 \
+  --network=results/00006-stylegan2-ffhq-8gpu-config-f/networks-final.pkl
+```
+
+## Evaluation metrics
+
+To reproduce the numbers for config F in Tables 1 and 3, run:
+
+```.bash
+python run_metrics.py --data-dir=~/datasets --network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
+  --metrics=fid50k,ppl_wend --dataset=ffhq --mirror-augment=true
+python run_metrics.py --data-dir=~/datasets --network=gdrive:networks/stylegan2-car-config-f.pkl \
+  --metrics=fid50k,ppl2_wend --dataset=car
+python run_metrics.py --data-dir=~/datasets --network=gdrive:networks/stylegan2-cat-config-f.pkl \
+  --metrics=fid50k,ppl2_wend --dataset=cat
+python run_metrics.py --data-dir=~/datasets --network=gdrive:networks/stylegan2-church-config-f.pkl \
+  --metrics=fid50k,ppl2_wend --dataset=church
+python run_metrics.py --data-dir=~/datasets --network=gdrive:networks/stylegan2-horse-config-f.pkl \
+  --metrics=fid50k,ppl2_wend --dataset=horse
+```
+
+For other configurations, see the [StyleGAN2 Google Drive folder](https://drive.google.com/open?id=1QHc-yF5C3DChRwSdZKcx1w6K8JvSxQi7).
+
+Note that the metrics are evaluated using a different random seed each time, so the results will vary between runs. In the paper, we reported the average result of running each metric 10 times. The following table lists the available metrics along with their expected runtimes and random variation:
+
+| Metric      | FFHQ config F  | 1 GPU  | 2 GPUs  | 4 GPUs | Description |
+| :---------- | :------------: | :----: | :-----: | :----: | :---------- |
+| `fid50k`    | 2.84 &pm; 0.03 | 22 min | 14 min  | 10 min | [Fr&eacute;chet Inception Distance](https://arxiv.org/abs/1706.08500)
+| `is50k`     | 5.13 &pm; 0.02 | 23 min | 14 min  | 8 min  | [Inception Score](https://arxiv.org/abs/1606.03498)
+| `ppl_zfull` | 348.0 &pm; 3.8 | 41 min | 22 min  | 14 min | [Perceptual Path Length](https://arxiv.org/abs/1812.04948) in Z, full paths
+| `ppl_wfull` | 126.9 &pm; 0.2 | 42 min | 22 min  | 13 min | [Perceptual Path Length](https://arxiv.org/abs/1812.04948) in W, full paths
+| `ppl_zend`  | 348.6 &pm; 3.0 | 41 min | 22 min  | 14 min | [Perceptual Path Length](https://arxiv.org/abs/1812.04948) in Z, path endpoints
+| `ppl_wend`  | 129.4 &pm; 0.8 | 40 min | 23 min  | 13 min | [Perceptual Path Length](https://arxiv.org/abs/1812.04948) in W, path endpoints
+| `ppl2_wend` | 145.0 &pm; 0.5 | 41 min | 23 min  | 14 min | [Perceptual Path Length](https://arxiv.org/abs/1812.04948) without center crop
+| `ls`        | 154.2 / 4.27   | 10 hrs | 6 hrs   | 4 hrs  | [Linear Separability](https://arxiv.org/abs/1812.04948)
+| `pr50k3`    | 0.689 / 0.492  | 26 min | 17 min  | 12 min | [Precision and Recall](https://arxiv.org/abs/1904.06991)
+
+Note that some of the metrics cache dataset-specific data on the disk, and they will take somewhat longer when run for the first time.
+
+## License
+
+Copyright &copy; 2019, NVIDIA Corporation. All rights reserved.
+
+This work is made available under the Nvidia Source Code License-NC. To view a copy of this license, visit https://nvlabs.github.io/stylegan2/license.html
+
+## Citation
+
+```
+@inproceedings{Karras2019stylegan2,
+  title     = {Analyzing and Improving the Image Quality of {StyleGAN}},
+  author    = {Tero Karras and Samuli Laine and Miika Aittala and Janne Hellsten and Jaakko Lehtinen and Timo Aila},
+  booktitle = {Proc. CVPR},
+  year      = {2020}
+}
+```
+
+## Acknowledgements
+
+We thank Ming-Yu Liu for an early review, Timo Viitanen for his help with code release, and Tero Kuosmanen for compute infrastructure.
\ No newline at end of file
diff --git a/reconstruction/ostec/external/stylegan2/dataset_tool.py b/reconstruction/ostec/external/stylegan2/dataset_tool.py
new file mode 100644
index 000000000..d8c4dc280
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dataset_tool.py
@@ -0,0 +1,644 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Tool for creating multi-resolution TFRecords datasets."""
+
+# pylint: disable=too-many-lines
+import os
+import sys
+import glob
+import argparse
+import threading
+import six.moves.queue as Queue # pylint: disable=import-error
+import traceback
+import numpy as np
+import tensorflow as tf
+import PIL.Image
+import dnnlib.tflib as tflib
+
+from training import dataset
+
+#----------------------------------------------------------------------------
+
+def error(msg):
+    print('Error: ' + msg)
+    exit(1)
+
+#----------------------------------------------------------------------------
+
+class TFRecordExporter:
+    def __init__(self, tfrecord_dir, expected_images, print_progress=True, progress_interval=10):
+        self.tfrecord_dir       = tfrecord_dir
+        self.tfr_prefix         = os.path.join(self.tfrecord_dir, os.path.basename(self.tfrecord_dir))
+        self.expected_images    = expected_images
+        self.cur_images         = 0
+        self.shape              = None
+        self.resolution_log2    = None
+        self.tfr_writers        = []
+        self.print_progress     = print_progress
+        self.progress_interval  = progress_interval
+
+        if self.print_progress:
+            print('Creating dataset "%s"' % tfrecord_dir)
+        if not os.path.isdir(self.tfrecord_dir):
+            os.makedirs(self.tfrecord_dir)
+        assert os.path.isdir(self.tfrecord_dir)
+
+    def close(self):
+        if self.print_progress:
+            print('%-40s\r' % 'Flushing data...', end='', flush=True)
+        for tfr_writer in self.tfr_writers:
+            tfr_writer.close()
+        self.tfr_writers = []
+        if self.print_progress:
+            print('%-40s\r' % '', end='', flush=True)
+            print('Added %d images.' % self.cur_images)
+
+    def choose_shuffled_order(self): # Note: Images and labels must be added in shuffled order.
+        order = np.arange(self.expected_images)
+        np.random.RandomState(123).shuffle(order)
+        return order
+
+    def add_image(self, img):
+        if self.print_progress and self.cur_images % self.progress_interval == 0:
+            print('%d / %d\r' % (self.cur_images, self.expected_images), end='', flush=True)
+        if self.shape is None:
+            self.shape = img.shape
+            self.resolution_log2 = int(np.log2(self.shape[1]))
+            assert self.shape[0] in [1, 3]
+            assert self.shape[1] == self.shape[2]
+            assert self.shape[1] == 2**self.resolution_log2
+            tfr_opt = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.NONE)
+            for lod in range(self.resolution_log2 - 1):
+                tfr_file = self.tfr_prefix + '-r%02d.tfrecords' % (self.resolution_log2 - lod)
+                self.tfr_writers.append(tf.python_io.TFRecordWriter(tfr_file, tfr_opt))
+        assert img.shape == self.shape
+        for lod, tfr_writer in enumerate(self.tfr_writers):
+            if lod:
+                img = img.astype(np.float32)
+                img = (img[:, 0::2, 0::2] + img[:, 0::2, 1::2] + img[:, 1::2, 0::2] + img[:, 1::2, 1::2]) * 0.25
+            quant = np.rint(img).clip(0, 255).astype(np.uint8)
+            ex = tf.train.Example(features=tf.train.Features(feature={
+                'shape': tf.train.Feature(int64_list=tf.train.Int64List(value=quant.shape)),
+                'data': tf.train.Feature(bytes_list=tf.train.BytesList(value=[quant.tostring()]))}))
+            tfr_writer.write(ex.SerializeToString())
+        self.cur_images += 1
+
+    def add_labels(self, labels):
+        if self.print_progress:
+            print('%-40s\r' % 'Saving labels...', end='', flush=True)
+        assert labels.shape[0] == self.cur_images
+        with open(self.tfr_prefix + '-rxx.labels', 'wb') as f:
+            np.save(f, labels.astype(np.float32))
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *args):
+        self.close()
+
+#----------------------------------------------------------------------------
+
+class ExceptionInfo(object):
+    def __init__(self):
+        self.value = sys.exc_info()[1]
+        self.traceback = traceback.format_exc()
+
+#----------------------------------------------------------------------------
+
+class WorkerThread(threading.Thread):
+    def __init__(self, task_queue):
+        threading.Thread.__init__(self)
+        self.task_queue = task_queue
+
+    def run(self):
+        while True:
+            func, args, result_queue = self.task_queue.get()
+            if func is None:
+                break
+            try:
+                result = func(*args)
+            except:
+                result = ExceptionInfo()
+            result_queue.put((result, args))
+
+#----------------------------------------------------------------------------
+
+class ThreadPool(object):
+    def __init__(self, num_threads):
+        assert num_threads >= 1
+        self.task_queue = Queue.Queue()
+        self.result_queues = dict()
+        self.num_threads = num_threads
+        for _idx in range(self.num_threads):
+            thread = WorkerThread(self.task_queue)
+            thread.daemon = True
+            thread.start()
+
+    def add_task(self, func, args=()):
+        assert hasattr(func, '__call__') # must be a function
+        if func not in self.result_queues:
+            self.result_queues[func] = Queue.Queue()
+        self.task_queue.put((func, args, self.result_queues[func]))
+
+    def get_result(self, func): # returns (result, args)
+        result, args = self.result_queues[func].get()
+        if isinstance(result, ExceptionInfo):
+            print('\n\nWorker thread caught an exception:\n' + result.traceback)
+            raise result.value
+        return result, args
+
+    def finish(self):
+        for _idx in range(self.num_threads):
+            self.task_queue.put((None, (), None))
+
+    def __enter__(self): # for 'with' statement
+        return self
+
+    def __exit__(self, *excinfo):
+        self.finish()
+
+    def process_items_concurrently(self, item_iterator, process_func=lambda x: x, pre_func=lambda x: x, post_func=lambda x: x, max_items_in_flight=None):
+        if max_items_in_flight is None: max_items_in_flight = self.num_threads * 4
+        assert max_items_in_flight >= 1
+        results = []
+        retire_idx = [0]
+
+        def task_func(prepared, _idx):
+            return process_func(prepared)
+
+        def retire_result():
+            processed, (_prepared, idx) = self.get_result(task_func)
+            results[idx] = processed
+            while retire_idx[0] < len(results) and results[retire_idx[0]] is not None:
+                yield post_func(results[retire_idx[0]])
+                results[retire_idx[0]] = None
+                retire_idx[0] += 1
+
+        for idx, item in enumerate(item_iterator):
+            prepared = pre_func(item)
+            results.append(None)
+            self.add_task(func=task_func, args=(prepared, idx))
+            while retire_idx[0] < idx - max_items_in_flight + 2:
+                for res in retire_result(): yield res
+        while retire_idx[0] < len(results):
+            for res in retire_result(): yield res
+
+#----------------------------------------------------------------------------
+
+def display(tfrecord_dir):
+    print('Loading dataset "%s"' % tfrecord_dir)
+    tflib.init_tf({'gpu_options.allow_growth': True})
+    dset = dataset.TFRecordDataset(tfrecord_dir, max_label_size='full', repeat=False, shuffle_mb=0)
+    tflib.init_uninitialized_vars()
+    import cv2  # pip install opencv-python
+
+    idx = 0
+    while True:
+        try:
+            images, labels = dset.get_minibatch_np(1)
+        except tf.errors.OutOfRangeError:
+            break
+        if idx == 0:
+            print('Displaying images')
+            cv2.namedWindow('dataset_tool')
+            print('Press SPACE or ENTER to advance, ESC to exit')
+        print('\nidx = %-8d\nlabel = %s' % (idx, labels[0].tolist()))
+        cv2.imshow('dataset_tool', images[0].transpose(1, 2, 0)[:, :, ::-1]) # CHW => HWC, RGB => BGR
+        idx += 1
+        if cv2.waitKey() == 27:
+            break
+    print('\nDisplayed %d images.' % idx)
+
+#----------------------------------------------------------------------------
+
+def extract(tfrecord_dir, output_dir):
+    print('Loading dataset "%s"' % tfrecord_dir)
+    tflib.init_tf({'gpu_options.allow_growth': True})
+    dset = dataset.TFRecordDataset(tfrecord_dir, max_label_size=0, repeat=False, shuffle_mb=0)
+    tflib.init_uninitialized_vars()
+
+    print('Extracting images to "%s"' % output_dir)
+    if not os.path.isdir(output_dir):
+        os.makedirs(output_dir)
+    idx = 0
+    while True:
+        if idx % 10 == 0:
+            print('%d\r' % idx, end='', flush=True)
+        try:
+            images, _labels = dset.get_minibatch_np(1)
+        except tf.errors.OutOfRangeError:
+            break
+        if images.shape[1] == 1:
+            img = PIL.Image.fromarray(images[0][0], 'L')
+        else:
+            img = PIL.Image.fromarray(images[0].transpose(1, 2, 0), 'RGB')
+        img.save(os.path.join(output_dir, 'img%08d.png' % idx))
+        idx += 1
+    print('Extracted %d images.' % idx)
+
+#----------------------------------------------------------------------------
+
+def compare(tfrecord_dir_a, tfrecord_dir_b, ignore_labels):
+    max_label_size = 0 if ignore_labels else 'full'
+    print('Loading dataset "%s"' % tfrecord_dir_a)
+    tflib.init_tf({'gpu_options.allow_growth': True})
+    dset_a = dataset.TFRecordDataset(tfrecord_dir_a, max_label_size=max_label_size, repeat=False, shuffle_mb=0)
+    print('Loading dataset "%s"' % tfrecord_dir_b)
+    dset_b = dataset.TFRecordDataset(tfrecord_dir_b, max_label_size=max_label_size, repeat=False, shuffle_mb=0)
+    tflib.init_uninitialized_vars()
+
+    print('Comparing datasets')
+    idx = 0
+    identical_images = 0
+    identical_labels = 0
+    while True:
+        if idx % 100 == 0:
+            print('%d\r' % idx, end='', flush=True)
+        try:
+            images_a, labels_a = dset_a.get_minibatch_np(1)
+        except tf.errors.OutOfRangeError:
+            images_a, labels_a = None, None
+        try:
+            images_b, labels_b = dset_b.get_minibatch_np(1)
+        except tf.errors.OutOfRangeError:
+            images_b, labels_b = None, None
+        if images_a is None or images_b is None:
+            if images_a is not None or images_b is not None:
+                print('Datasets contain different number of images')
+            break
+        if images_a.shape == images_b.shape and np.all(images_a == images_b):
+            identical_images += 1
+        else:
+            print('Image %d is different' % idx)
+        if labels_a.shape == labels_b.shape and np.all(labels_a == labels_b):
+            identical_labels += 1
+        else:
+            print('Label %d is different' % idx)
+        idx += 1
+    print('Identical images: %d / %d' % (identical_images, idx))
+    if not ignore_labels:
+        print('Identical labels: %d / %d' % (identical_labels, idx))
+
+#----------------------------------------------------------------------------
+
+def create_mnist(tfrecord_dir, mnist_dir):
+    print('Loading MNIST from "%s"' % mnist_dir)
+    import gzip
+    with gzip.open(os.path.join(mnist_dir, 'train-images-idx3-ubyte.gz'), 'rb') as file:
+        images = np.frombuffer(file.read(), np.uint8, offset=16)
+    with gzip.open(os.path.join(mnist_dir, 'train-labels-idx1-ubyte.gz'), 'rb') as file:
+        labels = np.frombuffer(file.read(), np.uint8, offset=8)
+    images = images.reshape(-1, 1, 28, 28)
+    images = np.pad(images, [(0,0), (0,0), (2,2), (2,2)], 'constant', constant_values=0)
+    assert images.shape == (60000, 1, 32, 32) and images.dtype == np.uint8
+    assert labels.shape == (60000,) and labels.dtype == np.uint8
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 9
+    onehot = np.zeros((labels.size, np.max(labels) + 1), dtype=np.float32)
+    onehot[np.arange(labels.size), labels] = 1.0
+
+    with TFRecordExporter(tfrecord_dir, images.shape[0]) as tfr:
+        order = tfr.choose_shuffled_order()
+        for idx in range(order.size):
+            tfr.add_image(images[order[idx]])
+        tfr.add_labels(onehot[order])
+
+#----------------------------------------------------------------------------
+
+def create_mnistrgb(tfrecord_dir, mnist_dir, num_images=1000000, random_seed=123):
+    print('Loading MNIST from "%s"' % mnist_dir)
+    import gzip
+    with gzip.open(os.path.join(mnist_dir, 'train-images-idx3-ubyte.gz'), 'rb') as file:
+        images = np.frombuffer(file.read(), np.uint8, offset=16)
+    images = images.reshape(-1, 28, 28)
+    images = np.pad(images, [(0,0), (2,2), (2,2)], 'constant', constant_values=0)
+    assert images.shape == (60000, 32, 32) and images.dtype == np.uint8
+    assert np.min(images) == 0 and np.max(images) == 255
+
+    with TFRecordExporter(tfrecord_dir, num_images) as tfr:
+        rnd = np.random.RandomState(random_seed)
+        for _idx in range(num_images):
+            tfr.add_image(images[rnd.randint(images.shape[0], size=3)])
+
+#----------------------------------------------------------------------------
+
+def create_cifar10(tfrecord_dir, cifar10_dir):
+    print('Loading CIFAR-10 from "%s"' % cifar10_dir)
+    import pickle
+    images = []
+    labels = []
+    for batch in range(1, 6):
+        with open(os.path.join(cifar10_dir, 'data_batch_%d' % batch), 'rb') as file:
+            data = pickle.load(file, encoding='latin1')
+        images.append(data['data'].reshape(-1, 3, 32, 32))
+        labels.append(data['labels'])
+    images = np.concatenate(images)
+    labels = np.concatenate(labels)
+    assert images.shape == (50000, 3, 32, 32) and images.dtype == np.uint8
+    assert labels.shape == (50000,) and labels.dtype == np.int32
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 9
+    onehot = np.zeros((labels.size, np.max(labels) + 1), dtype=np.float32)
+    onehot[np.arange(labels.size), labels] = 1.0
+
+    with TFRecordExporter(tfrecord_dir, images.shape[0]) as tfr:
+        order = tfr.choose_shuffled_order()
+        for idx in range(order.size):
+            tfr.add_image(images[order[idx]])
+        tfr.add_labels(onehot[order])
+
+#----------------------------------------------------------------------------
+
+def create_cifar100(tfrecord_dir, cifar100_dir):
+    print('Loading CIFAR-100 from "%s"' % cifar100_dir)
+    import pickle
+    with open(os.path.join(cifar100_dir, 'train'), 'rb') as file:
+        data = pickle.load(file, encoding='latin1')
+    images = data['data'].reshape(-1, 3, 32, 32)
+    labels = np.array(data['fine_labels'])
+    assert images.shape == (50000, 3, 32, 32) and images.dtype == np.uint8
+    assert labels.shape == (50000,) and labels.dtype == np.int32
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 99
+    onehot = np.zeros((labels.size, np.max(labels) + 1), dtype=np.float32)
+    onehot[np.arange(labels.size), labels] = 1.0
+
+    with TFRecordExporter(tfrecord_dir, images.shape[0]) as tfr:
+        order = tfr.choose_shuffled_order()
+        for idx in range(order.size):
+            tfr.add_image(images[order[idx]])
+        tfr.add_labels(onehot[order])
+
+#----------------------------------------------------------------------------
+
+def create_svhn(tfrecord_dir, svhn_dir):
+    print('Loading SVHN from "%s"' % svhn_dir)
+    import pickle
+    images = []
+    labels = []
+    for batch in range(1, 4):
+        with open(os.path.join(svhn_dir, 'train_%d.pkl' % batch), 'rb') as file:
+            data = pickle.load(file, encoding='latin1')
+        images.append(data[0])
+        labels.append(data[1])
+    images = np.concatenate(images)
+    labels = np.concatenate(labels)
+    assert images.shape == (73257, 3, 32, 32) and images.dtype == np.uint8
+    assert labels.shape == (73257,) and labels.dtype == np.uint8
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 9
+    onehot = np.zeros((labels.size, np.max(labels) + 1), dtype=np.float32)
+    onehot[np.arange(labels.size), labels] = 1.0
+
+    with TFRecordExporter(tfrecord_dir, images.shape[0]) as tfr:
+        order = tfr.choose_shuffled_order()
+        for idx in range(order.size):
+            tfr.add_image(images[order[idx]])
+        tfr.add_labels(onehot[order])
+
+#----------------------------------------------------------------------------
+
+def create_lsun(tfrecord_dir, lmdb_dir, resolution=256, max_images=None):
+    print('Loading LSUN dataset from "%s"' % lmdb_dir)
+    import lmdb # pip install lmdb # pylint: disable=import-error
+    import cv2 # pip install opencv-python
+    import io
+    with lmdb.open(lmdb_dir, readonly=True).begin(write=False) as txn:
+        total_images = txn.stat()['entries'] # pylint: disable=no-value-for-parameter
+        if max_images is None:
+            max_images = total_images
+        with TFRecordExporter(tfrecord_dir, max_images) as tfr:
+            for _idx, (_key, value) in enumerate(txn.cursor()):
+                try:
+                    try:
+                        img = cv2.imdecode(np.fromstring(value, dtype=np.uint8), 1)
+                        if img is None:
+                            raise IOError('cv2.imdecode failed')
+                        img = img[:, :, ::-1] # BGR => RGB
+                    except IOError:
+                        img = np.asarray(PIL.Image.open(io.BytesIO(value)))
+                    crop = np.min(img.shape[:2])
+                    img = img[(img.shape[0] - crop) // 2 : (img.shape[0] + crop) // 2, (img.shape[1] - crop) // 2 : (img.shape[1] + crop) // 2]
+                    img = PIL.Image.fromarray(img, 'RGB')
+                    img = img.resize((resolution, resolution), PIL.Image.ANTIALIAS)
+                    img = np.asarray(img)
+                    img = img.transpose([2, 0, 1]) # HWC => CHW
+                    tfr.add_image(img)
+                except:
+                    print(sys.exc_info()[1])
+                if tfr.cur_images == max_images:
+                    break
+
+#----------------------------------------------------------------------------
+
+def create_lsun_wide(tfrecord_dir, lmdb_dir, width=512, height=384, max_images=None):
+    assert width == 2 ** int(np.round(np.log2(width)))
+    assert height <= width
+    print('Loading LSUN dataset from "%s"' % lmdb_dir)
+    import lmdb # pip install lmdb # pylint: disable=import-error
+    import cv2 # pip install opencv-python
+    import io
+    with lmdb.open(lmdb_dir, readonly=True).begin(write=False) as txn:
+        total_images = txn.stat()['entries'] # pylint: disable=no-value-for-parameter
+        if max_images is None:
+            max_images = total_images
+        with TFRecordExporter(tfrecord_dir, max_images, print_progress=False) as tfr:
+            for idx, (_key, value) in enumerate(txn.cursor()):
+                try:
+                    try:
+                        img = cv2.imdecode(np.fromstring(value, dtype=np.uint8), 1)
+                        if img is None:
+                            raise IOError('cv2.imdecode failed')
+                        img = img[:, :, ::-1] # BGR => RGB
+                    except IOError:
+                        img = np.asarray(PIL.Image.open(io.BytesIO(value)))
+
+                    ch = int(np.round(width * img.shape[0] / img.shape[1]))
+                    if img.shape[1] < width or ch < height:
+                        continue
+
+                    img = img[(img.shape[0] - ch) // 2 : (img.shape[0] + ch) // 2]
+                    img = PIL.Image.fromarray(img, 'RGB')
+                    img = img.resize((width, height), PIL.Image.ANTIALIAS)
+                    img = np.asarray(img)
+                    img = img.transpose([2, 0, 1]) # HWC => CHW
+
+                    canvas = np.zeros([3, width, width], dtype=np.uint8)
+                    canvas[:, (width - height) // 2 : (width + height) // 2] = img
+                    tfr.add_image(canvas)
+                    print('\r%d / %d => %d ' % (idx + 1, total_images, tfr.cur_images), end='')
+
+                except:
+                    print(sys.exc_info()[1])
+                if tfr.cur_images == max_images:
+                    break
+    print()
+
+#----------------------------------------------------------------------------
+
+def create_celeba(tfrecord_dir, celeba_dir, cx=89, cy=121):
+    print('Loading CelebA from "%s"' % celeba_dir)
+    glob_pattern = os.path.join(celeba_dir, 'img_align_celeba_png', '*.png')
+    image_filenames = sorted(glob.glob(glob_pattern))
+    expected_images = 202599
+    if len(image_filenames) != expected_images:
+        error('Expected to find %d images' % expected_images)
+
+    with TFRecordExporter(tfrecord_dir, len(image_filenames)) as tfr:
+        order = tfr.choose_shuffled_order()
+        for idx in range(order.size):
+            img = np.asarray(PIL.Image.open(image_filenames[order[idx]]))
+            assert img.shape == (218, 178, 3)
+            img = img[cy - 64 : cy + 64, cx - 64 : cx + 64]
+            img = img.transpose(2, 0, 1) # HWC => CHW
+            tfr.add_image(img)
+
+#----------------------------------------------------------------------------
+
+def create_from_images(tfrecord_dir, image_dir, shuffle):
+    print('Loading images from "%s"' % image_dir)
+    image_filenames = sorted(glob.glob(os.path.join(image_dir, '*')))
+    if len(image_filenames) == 0:
+        error('No input images found')
+
+    img = np.asarray(PIL.Image.open(image_filenames[0]))
+    resolution = img.shape[0]
+    channels = img.shape[2] if img.ndim == 3 else 1
+    if img.shape[1] != resolution:
+        error('Input images must have the same width and height')
+    if resolution != 2 ** int(np.floor(np.log2(resolution))):
+        error('Input image resolution must be a power-of-two')
+    if channels not in [1, 3]:
+        error('Input images must be stored as RGB or grayscale')
+
+    with TFRecordExporter(tfrecord_dir, len(image_filenames)) as tfr:
+        order = tfr.choose_shuffled_order() if shuffle else np.arange(len(image_filenames))
+        for idx in range(order.size):
+            img = np.asarray(PIL.Image.open(image_filenames[order[idx]]))
+            if channels == 1:
+                img = img[np.newaxis, :, :] # HW => CHW
+            else:
+                img = img.transpose([2, 0, 1]) # HWC => CHW
+            tfr.add_image(img)
+
+#----------------------------------------------------------------------------
+
+def create_from_hdf5(tfrecord_dir, hdf5_filename, shuffle):
+    print('Loading HDF5 archive from "%s"' % hdf5_filename)
+    import h5py # conda install h5py
+    with h5py.File(hdf5_filename, 'r') as hdf5_file:
+        hdf5_data = max([value for key, value in hdf5_file.items() if key.startswith('data')], key=lambda lod: lod.shape[3])
+        with TFRecordExporter(tfrecord_dir, hdf5_data.shape[0]) as tfr:
+            order = tfr.choose_shuffled_order() if shuffle else np.arange(hdf5_data.shape[0])
+            for idx in range(order.size):
+                tfr.add_image(hdf5_data[order[idx]])
+            npy_filename = os.path.splitext(hdf5_filename)[0] + '-labels.npy'
+            if os.path.isfile(npy_filename):
+                tfr.add_labels(np.load(npy_filename)[order])
+
+#----------------------------------------------------------------------------
+
+def execute_cmdline(argv):
+    prog = argv[0]
+    parser = argparse.ArgumentParser(
+        prog        = prog,
+        description = 'Tool for creating multi-resolution TFRecords datasets for StyleGAN and ProGAN.',
+        epilog      = 'Type "%s <command> -h" for more information.' % prog)
+
+    subparsers = parser.add_subparsers(dest='command')
+    subparsers.required = True
+    def add_command(cmd, desc, example=None):
+        epilog = 'Example: %s %s' % (prog, example) if example is not None else None
+        return subparsers.add_parser(cmd, description=desc, help=desc, epilog=epilog)
+
+    p = add_command(    'display',          'Display images in dataset.',
+                                            'display datasets/mnist')
+    p.add_argument(     'tfrecord_dir',     help='Directory containing dataset')
+
+    p = add_command(    'extract',          'Extract images from dataset.',
+                                            'extract datasets/mnist mnist-images')
+    p.add_argument(     'tfrecord_dir',     help='Directory containing dataset')
+    p.add_argument(     'output_dir',       help='Directory to extract the images into')
+
+    p = add_command(    'compare',          'Compare two datasets.',
+                                            'compare datasets/mydataset datasets/mnist')
+    p.add_argument(     'tfrecord_dir_a',   help='Directory containing first dataset')
+    p.add_argument(     'tfrecord_dir_b',   help='Directory containing second dataset')
+    p.add_argument(     '--ignore_labels',  help='Ignore labels (default: 0)', type=int, default=0)
+
+    p = add_command(    'create_mnist',     'Create dataset for MNIST.',
+                                            'create_mnist datasets/mnist ~/downloads/mnist')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'mnist_dir',        help='Directory containing MNIST')
+
+    p = add_command(    'create_mnistrgb',  'Create dataset for MNIST-RGB.',
+                                            'create_mnistrgb datasets/mnistrgb ~/downloads/mnist')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'mnist_dir',        help='Directory containing MNIST')
+    p.add_argument(     '--num_images',     help='Number of composite images to create (default: 1000000)', type=int, default=1000000)
+    p.add_argument(     '--random_seed',    help='Random seed (default: 123)', type=int, default=123)
+
+    p = add_command(    'create_cifar10',   'Create dataset for CIFAR-10.',
+                                            'create_cifar10 datasets/cifar10 ~/downloads/cifar10')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'cifar10_dir',      help='Directory containing CIFAR-10')
+
+    p = add_command(    'create_cifar100',  'Create dataset for CIFAR-100.',
+                                            'create_cifar100 datasets/cifar100 ~/downloads/cifar100')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'cifar100_dir',     help='Directory containing CIFAR-100')
+
+    p = add_command(    'create_svhn',      'Create dataset for SVHN.',
+                                            'create_svhn datasets/svhn ~/downloads/svhn')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'svhn_dir',         help='Directory containing SVHN')
+
+    p = add_command(    'create_lsun',      'Create dataset for single LSUN category.',
+                                            'create_lsun datasets/lsun-car-100k ~/downloads/lsun/car_lmdb --resolution 256 --max_images 100000')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'lmdb_dir',         help='Directory containing LMDB database')
+    p.add_argument(     '--resolution',     help='Output resolution (default: 256)', type=int, default=256)
+    p.add_argument(     '--max_images',     help='Maximum number of images (default: none)', type=int, default=None)
+
+    p = add_command(    'create_lsun_wide', 'Create LSUN dataset with non-square aspect ratio.',
+                                            'create_lsun_wide datasets/lsun-car-512x384 ~/downloads/lsun/car_lmdb --width 512 --height 384')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'lmdb_dir',         help='Directory containing LMDB database')
+    p.add_argument(     '--width',          help='Output width (default: 512)', type=int, default=512)
+    p.add_argument(     '--height',         help='Output height (default: 384)', type=int, default=384)
+    p.add_argument(     '--max_images',     help='Maximum number of images (default: none)', type=int, default=None)
+
+    p = add_command(    'create_celeba',    'Create dataset for CelebA.',
+                                            'create_celeba datasets/celeba ~/downloads/celeba')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'celeba_dir',       help='Directory containing CelebA')
+    p.add_argument(     '--cx',             help='Center X coordinate (default: 89)', type=int, default=89)
+    p.add_argument(     '--cy',             help='Center Y coordinate (default: 121)', type=int, default=121)
+
+    p = add_command(    'create_from_images', 'Create dataset from a directory full of images.',
+                                            'create_from_images datasets/mydataset myimagedir')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'image_dir',        help='Directory containing the images')
+    p.add_argument(     '--shuffle',        help='Randomize image order (default: 1)', type=int, default=1)
+
+    p = add_command(    'create_from_hdf5', 'Create dataset from legacy HDF5 archive.',
+                                            'create_from_hdf5 datasets/celebahq ~/downloads/celeba-hq-1024x1024.h5')
+    p.add_argument(     'tfrecord_dir',     help='New dataset directory to be created')
+    p.add_argument(     'hdf5_filename',    help='HDF5 archive containing the images')
+    p.add_argument(     '--shuffle',        help='Randomize image order (default: 1)', type=int, default=1)
+
+    args = parser.parse_args(argv[1:] if len(argv) > 1 else ['-h'])
+    func = globals()[args.command]
+    del args.command
+    func(**vars(args))
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    execute_cmdline(sys.argv)
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/__init__.py b/reconstruction/ostec/external/stylegan2/dnnlib/__init__.py
new file mode 100644
index 000000000..e34112b62
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/__init__.py
@@ -0,0 +1,21 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+from . import submission
+
+from .submission.run_context import RunContext
+
+from .submission.submit import SubmitTarget
+from .submission.submit import PathType
+from .submission.submit import SubmitConfig
+from .submission.submit import submit_run
+from .submission.submit import get_path_from_template
+from .submission.submit import convert_path
+from .submission.submit import make_run_dir_path
+
+from .util import EasyDict
+
+submit_config: SubmitConfig = None # Package level variable for SubmitConfig which is only valid when inside the run function.
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/submission/__init__.py b/reconstruction/ostec/external/stylegan2/dnnlib/submission/__init__.py
new file mode 100644
index 000000000..acf2fbee4
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/submission/__init__.py
@@ -0,0 +1,8 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+from . import run_context
+from . import submit
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/__init__.py b/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/__init__.py
new file mode 100644
index 000000000..0f1127989
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/__init__.py
@@ -0,0 +1,7 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+from . import local
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/local.py b/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/local.py
new file mode 100644
index 000000000..c03c79e93
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/submission/internal/local.py
@@ -0,0 +1,22 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+class TargetOptions():
+    def __init__(self):
+        self.do_not_copy_source_files = False
+
+class Target():
+    def __init__(self):
+        pass
+
+    def finalize_submit_config(self, submit_config, host_run_dir):
+        print ('Local submit ', end='', flush=True)
+        submit_config.run_dir = host_run_dir
+
+    def submit(self, submit_config, host_run_dir):
+        from ..submit import run_wrapper, convert_path
+        print('- run_dir: %s' % convert_path(submit_config.run_dir), flush=True)
+        return run_wrapper(submit_config)
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/submission/run_context.py b/reconstruction/ostec/external/stylegan2/dnnlib/submission/run_context.py
new file mode 100644
index 000000000..f246563a7
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/submission/run_context.py
@@ -0,0 +1,109 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helpers for managing the run/training loop."""
+
+import datetime
+import os
+import pprint
+import time
+import types
+
+from typing import Any
+
+from . import submit
+
+# Singleton RunContext
+_run_context = None
+
+class RunContext(object):
+    """Helper class for managing the run/training loop.
+
+    The context will hide the implementation details of a basic run/training loop.
+    It will set things up properly, tell if run should be stopped, and then cleans up.
+    User should call update periodically and use should_stop to determine if run should be stopped.
+
+    Args:
+        submit_config: The SubmitConfig that is used for the current run.
+        config_module: (deprecated) The whole config module that is used for the current run.
+    """
+
+    def __init__(self, submit_config: submit.SubmitConfig, config_module: types.ModuleType = None):
+        global _run_context
+        # Only a single RunContext can be alive
+        assert _run_context is None
+        _run_context = self
+        self.submit_config = submit_config
+        self.should_stop_flag = False
+        self.has_closed = False
+        self.start_time = time.time()
+        self.last_update_time = time.time()
+        self.last_update_interval = 0.0
+        self.progress_monitor_file_path = None
+
+        # vestigial config_module support just prints a warning
+        if config_module is not None:
+            print("RunContext.config_module parameter support has been removed.")
+
+        # write out details about the run to a text file
+        self.run_txt_data = {"task_name": submit_config.task_name, "host_name": submit_config.host_name, "start_time": datetime.datetime.now().isoformat(sep=" ")}
+        with open(os.path.join(submit_config.run_dir, "run.txt"), "w") as f:
+            pprint.pprint(self.run_txt_data, stream=f, indent=4, width=200, compact=False)
+
+    def __enter__(self) -> "RunContext":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def update(self, loss: Any = 0, cur_epoch: Any = 0, max_epoch: Any = None) -> None:
+        """Do general housekeeping and keep the state of the context up-to-date.
+        Should be called often enough but not in a tight loop."""
+        assert not self.has_closed
+
+        self.last_update_interval = time.time() - self.last_update_time
+        self.last_update_time = time.time()
+
+        if os.path.exists(os.path.join(self.submit_config.run_dir, "abort.txt")):
+            self.should_stop_flag = True
+
+    def should_stop(self) -> bool:
+        """Tell whether a stopping condition has been triggered one way or another."""
+        return self.should_stop_flag
+
+    def get_time_since_start(self) -> float:
+        """How much time has passed since the creation of the context."""
+        return time.time() - self.start_time
+
+    def get_time_since_last_update(self) -> float:
+        """How much time has passed since the last call to update."""
+        return time.time() - self.last_update_time
+
+    def get_last_update_interval(self) -> float:
+        """How much time passed between the previous two calls to update."""
+        return self.last_update_interval
+
+    def close(self) -> None:
+        """Close the context and clean up.
+        Should only be called once."""
+        if not self.has_closed:
+            # update the run.txt with stopping time
+            self.run_txt_data["stop_time"] = datetime.datetime.now().isoformat(sep=" ")
+            with open(os.path.join(self.submit_config.run_dir, "run.txt"), "w") as f:
+                pprint.pprint(self.run_txt_data, stream=f, indent=4, width=200, compact=False)
+            self.has_closed = True
+
+            # detach the global singleton
+            global _run_context
+            if _run_context is self:
+                _run_context = None
+
+    @staticmethod
+    def get():
+        import dnnlib
+        if _run_context is not None:
+            return _run_context
+        return RunContext(dnnlib.submit_config)
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/submission/submit.py b/reconstruction/ostec/external/stylegan2/dnnlib/submission/submit.py
new file mode 100644
index 000000000..857ac3a7b
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/submission/submit.py
@@ -0,0 +1,341 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Submit a function to be run either locally or in a computing cluster."""
+
+import copy
+import inspect
+import os
+import pathlib
+import pickle
+import platform
+import pprint
+import re
+import shutil
+import sys
+import time
+import traceback
+
+from enum import Enum
+
+from .. import util
+
+
+class SubmitTarget(Enum):
+    """The target where the function should be run.
+
+    LOCAL: Run it locally.
+    """
+    LOCAL = 1
+
+
+class PathType(Enum):
+    """Determines in which format should a path be formatted.
+
+    WINDOWS: Format with Windows style.
+    LINUX: Format with Linux/Posix style.
+    AUTO: Use current OS type to select either WINDOWS or LINUX.
+    """
+    WINDOWS = 1
+    LINUX = 2
+    AUTO = 3
+
+
+class PlatformExtras:
+    """A mixed bag of values used by dnnlib heuristics.
+
+    Attributes:
+
+        data_reader_buffer_size: Used by DataReader to size internal shared memory buffers.
+        data_reader_process_count: Number of worker processes to spawn (zero for single thread operation)
+    """
+    def __init__(self):
+        self.data_reader_buffer_size = 1<<30    # 1 GB
+        self.data_reader_process_count = 0      # single threaded default
+
+
+_user_name_override = None
+
+class SubmitConfig(util.EasyDict):
+    """Strongly typed config dict needed to submit runs.
+
+    Attributes:
+        run_dir_root: Path to the run dir root. Can be optionally templated with tags. Needs to always be run through get_path_from_template.
+        run_desc: Description of the run. Will be used in the run dir and task name.
+        run_dir_ignore: List of file patterns used to ignore files when copying files to the run dir.
+        run_dir_extra_files: List of (abs_path, rel_path) tuples of file paths. rel_path root will be the src directory inside the run dir.
+        submit_target: Submit target enum value. Used to select where the run is actually launched.
+        num_gpus: Number of GPUs used/requested for the run.
+        print_info: Whether to print debug information when submitting.
+        local.do_not_copy_source_files: Do not copy source files from the working directory to the run dir.
+        run_id: Automatically populated value during submit.
+        run_name: Automatically populated value during submit.
+        run_dir: Automatically populated value during submit.
+        run_func_name: Automatically populated value during submit.
+        run_func_kwargs: Automatically populated value during submit.
+        user_name: Automatically populated value during submit. Can be set by the user which will then override the automatic value.
+        task_name: Automatically populated value during submit.
+        host_name: Automatically populated value during submit.
+        platform_extras: Automatically populated values during submit.  Used by various dnnlib libraries such as the DataReader class.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+        # run (set these)
+        self.run_dir_root = ""  # should always be passed through get_path_from_template
+        self.run_desc = ""
+        self.run_dir_ignore = ["__pycache__", "*.pyproj", "*.sln", "*.suo", ".cache", ".idea", ".vs", ".vscode", "_cudacache"]
+        self.run_dir_extra_files = []
+
+        # submit (set these)
+        self.submit_target = SubmitTarget.LOCAL
+        self.num_gpus = 1
+        self.print_info = False
+        self.nvprof = False
+        self.local = external.stylegan2.dnnlib.submission.internal.local.TargetOptions()
+        self.datasets = []
+
+        # (automatically populated)
+        self.run_id = None
+        self.run_name = None
+        self.run_dir = None
+        self.run_func_name = None
+        self.run_func_kwargs = None
+        self.user_name = None
+        self.task_name = None
+        self.host_name = "localhost"
+        self.platform_extras = PlatformExtras()
+
+
+def get_path_from_template(path_template: str, path_type: PathType = PathType.AUTO) -> str:
+    """Replace tags in the given path template and return either Windows or Linux formatted path."""
+    # automatically select path type depending on running OS
+    if path_type == PathType.AUTO:
+        if platform.system() == "Windows":
+            path_type = PathType.WINDOWS
+        elif platform.system() == "Linux":
+            path_type = PathType.LINUX
+        else:
+            raise RuntimeError("Unknown platform")
+
+    path_template = path_template.replace("<USERNAME>", get_user_name())
+
+    # return correctly formatted path
+    if path_type == PathType.WINDOWS:
+        return str(pathlib.PureWindowsPath(path_template))
+    elif path_type == PathType.LINUX:
+        return str(pathlib.PurePosixPath(path_template))
+    else:
+        raise RuntimeError("Unknown platform")
+
+
+def get_template_from_path(path: str) -> str:
+    """Convert a normal path back to its template representation."""
+    path = path.replace("\\", "/")
+    return path
+
+
+def convert_path(path: str, path_type: PathType = PathType.AUTO) -> str:
+    """Convert a normal path to template and the convert it back to a normal path with given path type."""
+    path_template = get_template_from_path(path)
+    path = get_path_from_template(path_template, path_type)
+    return path
+
+
+def set_user_name_override(name: str) -> None:
+    """Set the global username override value."""
+    global _user_name_override
+    _user_name_override = name
+
+
+def get_user_name():
+    """Get the current user name."""
+    if _user_name_override is not None:
+        return _user_name_override
+    elif platform.system() == "Windows":
+        return os.getlogin()
+    elif platform.system() == "Linux":
+        try:
+            import pwd
+            return pwd.getpwuid(os.geteuid()).pw_name
+        except:
+            return "unknown"
+    else:
+        raise RuntimeError("Unknown platform")
+
+
+def make_run_dir_path(*paths):
+    """Make a path/filename that resides under the current submit run_dir.
+
+    Args:
+        *paths: Path components to be passed to os.path.join
+
+    Returns:
+        A file/dirname rooted at submit_config.run_dir.  If there's no
+        submit_config or run_dir, the base directory is the current
+        working directory.
+
+    E.g., `os.path.join(dnnlib.submit_config.run_dir, "output.txt"))`
+    """
+    import dnnlib
+    if (dnnlib.submit_config is None) or (dnnlib.submit_config.run_dir is None):
+        return os.path.join(os.getcwd(), *paths)
+    return os.path.join(dnnlib.submit_config.run_dir, *paths)
+
+
+def _create_run_dir_local(submit_config: SubmitConfig) -> str:
+    """Create a new run dir with increasing ID number at the start."""
+    run_dir_root = get_path_from_template(submit_config.run_dir_root, PathType.AUTO)
+
+    if not os.path.exists(run_dir_root):
+        os.makedirs(run_dir_root)
+
+    submit_config.run_id = _get_next_run_id_local(run_dir_root)
+    submit_config.run_name = "{0:05d}-{1}".format(submit_config.run_id, submit_config.run_desc)
+    run_dir = os.path.join(run_dir_root, submit_config.run_name)
+
+    if os.path.exists(run_dir):
+        raise RuntimeError("The run dir already exists! ({0})".format(run_dir))
+
+    os.makedirs(run_dir)
+
+    return run_dir
+
+
+def _get_next_run_id_local(run_dir_root: str) -> int:
+    """Reads all directory names in a given directory (non-recursive) and returns the next (increasing) run id. Assumes IDs are numbers at the start of the directory names."""
+    dir_names = [d for d in os.listdir(run_dir_root) if os.path.isdir(os.path.join(run_dir_root, d))]
+    r = re.compile("^\\d+")  # match one or more digits at the start of the string
+    run_id = 0
+
+    for dir_name in dir_names:
+        m = r.match(dir_name)
+
+        if m is not None:
+            i = int(m.group())
+            run_id = max(run_id, i + 1)
+
+    return run_id
+
+
+def _populate_run_dir(submit_config: SubmitConfig, run_dir: str) -> None:
+    """Copy all necessary files into the run dir. Assumes that the dir exists, is local, and is writable."""
+    pickle.dump(submit_config, open(os.path.join(run_dir, "submit_config.pkl"), "wb"))
+    with open(os.path.join(run_dir, "submit_config.txt"), "w") as f:
+        pprint.pprint(submit_config, stream=f, indent=4, width=200, compact=False)
+
+    if (submit_config.submit_target == SubmitTarget.LOCAL) and submit_config.local.do_not_copy_source_files:
+        return
+
+    files = []
+
+    run_func_module_dir_path = util.get_module_dir_by_obj_name(submit_config.run_func_name)
+    assert '.' in submit_config.run_func_name
+    for _idx in range(submit_config.run_func_name.count('.') - 1):
+        run_func_module_dir_path = os.path.dirname(run_func_module_dir_path)
+    files += util.list_dir_recursively_with_ignore(run_func_module_dir_path, ignores=submit_config.run_dir_ignore, add_base_to_relative=False)
+
+    dnnlib_module_dir_path = util.get_module_dir_by_obj_name("dnnlib")
+    files += util.list_dir_recursively_with_ignore(dnnlib_module_dir_path, ignores=submit_config.run_dir_ignore, add_base_to_relative=True)
+
+    files += submit_config.run_dir_extra_files
+
+    files = [(f[0], os.path.join(run_dir, "src", f[1])) for f in files]
+    files += [(os.path.join(dnnlib_module_dir_path, "submission", "internal", "run.py"), os.path.join(run_dir, "run.py"))]
+
+    util.copy_files_and_create_dirs(files)
+
+
+
+def run_wrapper(submit_config: SubmitConfig) -> None:
+    """Wrap the actual run function call for handling logging, exceptions, typing, etc."""
+    is_local = submit_config.submit_target == SubmitTarget.LOCAL
+
+    # when running locally, redirect stderr to stdout, log stdout to a file, and force flushing
+    if is_local:
+        logger = util.Logger(file_name=os.path.join(submit_config.run_dir, "log.txt"), file_mode="w", should_flush=True)
+    else:  # when running in a cluster, redirect stderr to stdout, and just force flushing (log writing is handled by run.sh)
+        logger = util.Logger(file_name=None, should_flush=True)
+
+    import dnnlib
+    dnnlib.submit_config = submit_config
+
+    exit_with_errcode = False
+    try:
+        print("dnnlib: Running {0}() on {1}...".format(submit_config.run_func_name, submit_config.host_name))
+        start_time = time.time()
+
+        run_func_obj = util.get_obj_by_name(submit_config.run_func_name)
+        assert callable(run_func_obj)
+        sig = inspect.signature(run_func_obj)
+        if 'submit_config' in sig.parameters:
+            run_func_obj(submit_config=submit_config, **submit_config.run_func_kwargs)
+        else:
+            run_func_obj(**submit_config.run_func_kwargs)
+
+        print("dnnlib: Finished {0}() in {1}.".format(submit_config.run_func_name, util.format_time(time.time() - start_time)))
+    except:
+        if is_local:
+            raise
+        else:
+            traceback.print_exc()
+
+            log_src = os.path.join(submit_config.run_dir, "log.txt")
+            log_dst = os.path.join(get_path_from_template(submit_config.run_dir_root), "{0}-error.txt".format(submit_config.run_name))
+            shutil.copyfile(log_src, log_dst)
+
+            # Defer sys.exit(1) to happen after we close the logs and create a _finished.txt
+            exit_with_errcode = True
+    finally:
+        open(os.path.join(submit_config.run_dir, "_finished.txt"), "w").close()
+
+    dnnlib.RunContext.get().close()
+    dnnlib.submit_config = None
+    logger.close()
+
+    # If we hit an error, get out of the script now and signal the error
+    # to whatever process that started this script.
+    if exit_with_errcode:
+        sys.exit(1)
+
+    return submit_config
+
+
+def submit_run(submit_config: SubmitConfig, run_func_name: str, **run_func_kwargs) -> None:
+    """Create a run dir, gather files related to the run, copy files to the run dir, and launch the run in appropriate place."""
+    submit_config = copy.deepcopy(submit_config)
+
+    submit_target = submit_config.submit_target
+    farm = None
+    if submit_target == SubmitTarget.LOCAL:
+        farm = external.stylegan2.dnnlib.submission.internal.local.Target()
+    assert farm is not None # unknown target
+
+    # Disallow submitting jobs with zero num_gpus.
+    if (submit_config.num_gpus is None) or (submit_config.num_gpus == 0):
+        raise RuntimeError("submit_config.num_gpus must be set to a non-zero value")
+
+    if submit_config.user_name is None:
+        submit_config.user_name = get_user_name()
+
+    submit_config.run_func_name = run_func_name
+    submit_config.run_func_kwargs = run_func_kwargs
+
+    #--------------------------------------------------------------------
+    # Prepare submission by populating the run dir
+    #--------------------------------------------------------------------
+    host_run_dir = _create_run_dir_local(submit_config)
+
+    submit_config.task_name = "{0}-{1:05d}-{2}".format(submit_config.user_name, submit_config.run_id, submit_config.run_desc)
+    docker_valid_name_regex = "^[a-zA-Z0-9][a-zA-Z0-9_.-]+$"
+    if not re.match(docker_valid_name_regex, submit_config.task_name):
+        raise RuntimeError("Invalid task name.  Probable reason: unacceptable characters in your submit_config.run_desc.  Task name must be accepted by the following regex: " + docker_valid_name_regex + ", got " + submit_config.task_name)
+
+    # Farm specific preparations for a submit
+    farm.finalize_submit_config(submit_config, host_run_dir)
+    _populate_run_dir(submit_config, host_run_dir)
+    return farm.submit(submit_config, host_run_dir)
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/__init__.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/__init__.py
new file mode 100644
index 000000000..02c25173d
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/__init__.py
@@ -0,0 +1,18 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+from . import autosummary
+from . import network
+from . import optimizer
+from . import tfutil
+from . import custom_ops
+
+from .tfutil import *
+from .network import Network
+
+from .optimizer import Optimizer
+
+from .custom_ops import get_plugin
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/autosummary.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/autosummary.py
new file mode 100644
index 000000000..6b0d80b37
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/autosummary.py
@@ -0,0 +1,191 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper for adding automatically tracked values to Tensorboard.
+
+Autosummary creates an identity op that internally keeps track of the input
+values and automatically shows up in TensorBoard. The reported value
+represents an average over input components. The average is accumulated
+constantly over time and flushed when save_summaries() is called.
+
+Notes:
+- The output tensor must be used as an input for something else in the
+  graph. Otherwise, the autosummary op will not get executed, and the average
+  value will not get accumulated.
+- It is perfectly fine to include autosummaries with the same name in
+  several places throughout the graph, even if they are executed concurrently.
+- It is ok to also pass in a python scalar or numpy array. In this case, it
+  is added to the average immediately.
+"""
+
+from collections import OrderedDict
+import numpy as np
+import tensorflow as tf
+from tensorboard import summary as summary_lib
+from tensorboard.plugins.custom_scalar import layout_pb2
+
+from . import tfutil
+from .tfutil import TfExpression
+from .tfutil import TfExpressionEx
+
+# Enable "Custom scalars" tab in TensorBoard for advanced formatting.
+# Disabled by default to reduce tfevents file size.
+enable_custom_scalars = False
+
+_dtype = tf.float64
+_vars = OrderedDict()  # name => [var, ...]
+_immediate = OrderedDict()  # name => update_op, update_value
+_finalized = False
+_merge_op = None
+
+
+def _create_var(name: str, value_expr: TfExpression) -> TfExpression:
+    """Internal helper for creating autosummary accumulators."""
+    assert not _finalized
+    name_id = name.replace("/", "_")
+    v = tf.cast(value_expr, _dtype)
+
+    if v.shape.is_fully_defined():
+        size = np.prod(v.shape.as_list())
+        size_expr = tf.constant(size, dtype=_dtype)
+    else:
+        size = None
+        size_expr = tf.reduce_prod(tf.cast(tf.shape(v), _dtype))
+
+    if size == 1:
+        if v.shape.ndims != 0:
+            v = tf.reshape(v, [])
+        v = [size_expr, v, tf.square(v)]
+    else:
+        v = [size_expr, tf.reduce_sum(v), tf.reduce_sum(tf.square(v))]
+    v = tf.cond(tf.is_finite(v[1]), lambda: tf.stack(v), lambda: tf.zeros(3, dtype=_dtype))
+
+    with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.control_dependencies(None):
+        var = tf.Variable(tf.zeros(3, dtype=_dtype), trainable=False)  # [sum(1), sum(x), sum(x**2)]
+    update_op = tf.cond(tf.is_variable_initialized(var), lambda: tf.assign_add(var, v), lambda: tf.assign(var, v))
+
+    if name in _vars:
+        _vars[name].append(var)
+    else:
+        _vars[name] = [var]
+    return update_op
+
+
+def autosummary(name: str, value: TfExpressionEx, passthru: TfExpressionEx = None, condition: TfExpressionEx = True) -> TfExpressionEx:
+    """Create a new autosummary.
+
+    Args:
+        name:     Name to use in TensorBoard
+        value:    TensorFlow expression or python value to track
+        passthru: Optionally return this TF node without modifications but tack an autosummary update side-effect to this node.
+
+    Example use of the passthru mechanism:
+
+    n = autosummary('l2loss', loss, passthru=n)
+
+    This is a shorthand for the following code:
+
+    with tf.control_dependencies([autosummary('l2loss', loss)]):
+        n = tf.identity(n)
+    """
+    tfutil.assert_tf_initialized()
+    name_id = name.replace("/", "_")
+
+    if tfutil.is_tf_expression(value):
+        with tf.name_scope("summary_" + name_id), tf.device(value.device):
+            condition = tf.convert_to_tensor(condition, name='condition')
+            update_op = tf.cond(condition, lambda: tf.group(_create_var(name, value)), tf.no_op)
+            with tf.control_dependencies([update_op]):
+                return tf.identity(value if passthru is None else passthru)
+
+    else:  # python scalar or numpy array
+        assert not tfutil.is_tf_expression(passthru)
+        assert not tfutil.is_tf_expression(condition)
+        if condition:
+            if name not in _immediate:
+                with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.device(None), tf.control_dependencies(None):
+                    update_value = tf.placeholder(_dtype)
+                    update_op = _create_var(name, update_value)
+                    _immediate[name] = update_op, update_value
+            update_op, update_value = _immediate[name]
+            tfutil.run(update_op, {update_value: value})
+        return value if passthru is None else passthru
+
+
+def finalize_autosummaries() -> None:
+    """Create the necessary ops to include autosummaries in TensorBoard report.
+    Note: This should be done only once per graph.
+    """
+    global _finalized
+    tfutil.assert_tf_initialized()
+
+    if _finalized:
+        return None
+
+    _finalized = True
+    tfutil.init_uninitialized_vars([var for vars_list in _vars.values() for var in vars_list])
+
+    # Create summary ops.
+    with tf.device(None), tf.control_dependencies(None):
+        for name, vars_list in _vars.items():
+            name_id = name.replace("/", "_")
+            with tfutil.absolute_name_scope("Autosummary/" + name_id):
+                moments = tf.add_n(vars_list)
+                moments /= moments[0]
+                with tf.control_dependencies([moments]):  # read before resetting
+                    reset_ops = [tf.assign(var, tf.zeros(3, dtype=_dtype)) for var in vars_list]
+                    with tf.name_scope(None), tf.control_dependencies(reset_ops):  # reset before reporting
+                        mean = moments[1]
+                        std = tf.sqrt(moments[2] - tf.square(moments[1]))
+                        tf.summary.scalar(name, mean)
+                        if enable_custom_scalars:
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_lo", mean - std)
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_hi", mean + std)
+
+    # Setup layout for custom scalars.
+    layout = None
+    if enable_custom_scalars:
+        cat_dict = OrderedDict()
+        for series_name in sorted(_vars.keys()):
+            p = series_name.split("/")
+            cat = p[0] if len(p) >= 2 else ""
+            chart = "/".join(p[1:-1]) if len(p) >= 3 else p[-1]
+            if cat not in cat_dict:
+                cat_dict[cat] = OrderedDict()
+            if chart not in cat_dict[cat]:
+                cat_dict[cat][chart] = []
+            cat_dict[cat][chart].append(series_name)
+        categories = []
+        for cat_name, chart_dict in cat_dict.items():
+            charts = []
+            for chart_name, series_names in chart_dict.items():
+                series = []
+                for series_name in series_names:
+                    series.append(layout_pb2.MarginChartContent.Series(
+                        value=series_name,
+                        lower="xCustomScalars/" + series_name + "/margin_lo",
+                        upper="xCustomScalars/" + series_name + "/margin_hi"))
+                margin = layout_pb2.MarginChartContent(series=series)
+                charts.append(layout_pb2.Chart(title=chart_name, margin=margin))
+            categories.append(layout_pb2.Category(title=cat_name, chart=charts))
+        layout = summary_lib.custom_scalar_pb(layout_pb2.Layout(category=categories))
+    return layout
+
+def save_summaries(file_writer, global_step=None):
+    """Call FileWriter.add_summary() with all summaries in the default graph,
+    automatically finalizing and merging them on the first call.
+    """
+    global _merge_op
+    tfutil.assert_tf_initialized()
+
+    if _merge_op is None:
+        layout = finalize_autosummaries()
+        if layout is not None:
+            file_writer.add_summary(layout)
+        with tf.device(None), tf.control_dependencies(None):
+            _merge_op = tf.summary.merge_all()
+
+    file_writer.add_summary(_merge_op.eval(), global_step)
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/custom_ops.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/custom_ops.py
new file mode 100644
index 000000000..02434417f
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/custom_ops.py
@@ -0,0 +1,169 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""TensorFlow custom ops builder.
+"""
+
+import os
+import re
+import uuid
+import hashlib
+import tempfile
+import shutil
+import tensorflow as tf
+from tensorflow.python.client import device_lib # pylint: disable=no-name-in-module
+
+#----------------------------------------------------------------------------
+# Global options.
+
+cuda_cache_path = os.path.join(os.path.dirname(__file__), '_cudacache')
+cuda_cache_version_tag = 'v1'
+do_not_hash_included_headers = False # Speed up compilation by assuming that headers included by the CUDA code never change. Unsafe!
+verbose = True # Print status messages to stdout.
+
+compiler_bindir_search_path = [
+    'C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx64/x64',
+    'C:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.23.28105/bin/Hostx64/x64',
+    'C:/Program Files (x86)/Microsoft Visual Studio 14.0/vc/bin',
+]
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _find_compiler_bindir():
+    for compiler_path in compiler_bindir_search_path:
+        if os.path.isdir(compiler_path):
+            return compiler_path
+    return None
+
+def _get_compute_cap(device):
+    caps_str = device.physical_device_desc
+    m = re.search('compute capability: (\\d+).(\\d+)', caps_str)
+    major = m.group(1)
+    minor = m.group(2)
+    return (major, minor)
+
+def _get_cuda_gpu_arch_string():
+    gpus = [x for x in device_lib.list_local_devices() if x.device_type == 'GPU']
+    if len(gpus) == 0:
+        raise RuntimeError('No GPU devices found')
+    (major, minor) = _get_compute_cap(gpus[0])
+    return 'sm_%s%s' % (major, minor)
+
+def _run_cmd(cmd):
+    with os.popen(cmd) as pipe:
+        output = pipe.read()
+        status = pipe.close()
+    if status is not None:
+        raise RuntimeError('NVCC returned an error. See below for full command line and output log:\n\n%s\n\n%s' % (cmd, output))
+
+def _prepare_nvcc_cli(opts):
+    cmd = 'nvcc --std=c++11 -DNDEBUG ' + opts.strip()
+    cmd += ' --disable-warnings'
+    cmd += ' --include-path "%s"' % tf.sysconfig.get_include()
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'protobuf_archive', 'src')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'com_google_absl')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'eigen_archive')
+
+    compiler_bindir = _find_compiler_bindir()
+    if compiler_bindir is None:
+        # Require that _find_compiler_bindir succeeds on Windows.  Allow
+        # nvcc to use whatever is the default on Linux.
+        if os.name == 'nt':
+            raise RuntimeError('Could not find MSVC/GCC/CLANG installation on this computer. Check compiler_bindir_search_path list in "%s".' % __file__)
+    else:
+        cmd += ' --compiler-bindir "%s"' % compiler_bindir
+    cmd += ' 2>&1'
+    return cmd
+
+#----------------------------------------------------------------------------
+# Main entry point.
+
+_plugin_cache = dict()
+
+def get_plugin(cuda_file):
+    cuda_file_base = os.path.basename(cuda_file)
+    cuda_file_name, cuda_file_ext = os.path.splitext(cuda_file_base)
+
+    # Already in cache?
+    if cuda_file in _plugin_cache:
+        return _plugin_cache[cuda_file]
+
+    # Setup plugin.
+    if verbose:
+        print('Setting up TensorFlow plugin "%s": ' % cuda_file_base, end='', flush=True)
+    try:
+        # Hash CUDA source.
+        md5 = hashlib.md5()
+        with open(cuda_file, 'rb') as f:
+            md5.update(f.read())
+        md5.update(b'\n')
+
+        # Hash headers included by the CUDA code by running it through the preprocessor.
+        if not do_not_hash_included_headers:
+            if verbose:
+                print('Preprocessing... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + cuda_file_ext)
+                _run_cmd(_prepare_nvcc_cli('"%s" --preprocess -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir)))
+                with open(tmp_file, 'rb') as f:
+                    bad_file_str = ('"' + cuda_file.replace('\\', '/') + '"').encode('utf-8') # __FILE__ in error check macros
+                    good_file_str = ('"' + cuda_file_base + '"').encode('utf-8')
+                    for ln in f:
+                        if not ln.startswith(b'# ') and not ln.startswith(b'#line '): # ignore line number pragmas
+                            ln = ln.replace(bad_file_str, good_file_str)
+                            md5.update(ln)
+                    md5.update(b'\n')
+
+        # Select compiler options.
+        compile_opts = ''
+        if os.name == 'nt':
+            compile_opts += '"%s"' % os.path.join(tf.sysconfig.get_lib(), 'python', '_pywrap_tensorflow_internal.lib')
+        elif os.name == 'posix':
+            compile_opts += '"%s"' % os.path.join(tf.sysconfig.get_lib(), 'python', '_pywrap_tensorflow_internal.so')
+            compile_opts += ' --compiler-options \'-fPIC -D_GLIBCXX_USE_CXX11_ABI=0\''
+        else:
+            assert False # not Windows or Linux, w00t?
+        compile_opts += ' --gpu-architecture=%s' % _get_cuda_gpu_arch_string()
+        compile_opts += ' --use_fast_math'
+        nvcc_cmd = _prepare_nvcc_cli(compile_opts)
+
+        # Hash build configuration.
+        md5.update(('nvcc_cmd: ' + nvcc_cmd).encode('utf-8') + b'\n')
+        md5.update(('tf.VERSION: ' + tf.VERSION).encode('utf-8') + b'\n')
+        md5.update(('cuda_cache_version_tag: ' + cuda_cache_version_tag).encode('utf-8') + b'\n')
+
+        # Compile if not already compiled.
+        bin_file_ext = '.dll' if os.name == 'nt' else '.so'
+        bin_file = os.path.join(cuda_cache_path, cuda_file_name + '_' + md5.hexdigest() + bin_file_ext)
+        if not os.path.isfile(bin_file):
+            if verbose:
+                print('Compiling... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + bin_file_ext)
+                _run_cmd(nvcc_cmd + ' "%s" --shared -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir))
+                os.makedirs(cuda_cache_path, exist_ok=True)
+                intermediate_file = os.path.join(cuda_cache_path, cuda_file_name + '_' + uuid.uuid4().hex + '_tmp' + bin_file_ext)
+                shutil.copyfile(tmp_file, intermediate_file)
+                os.rename(intermediate_file, bin_file) # atomic
+
+        # Load.
+        if verbose:
+            print('Loading... ', end='', flush=True)
+        plugin = tf.load_op_library(bin_file)
+
+        # Add to cache.
+        _plugin_cache[cuda_file] = plugin
+        if verbose:
+            print('Done.', flush=True)
+        return plugin
+
+    except:
+        if verbose:
+            print('Failed!', flush=True)
+        raise
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/network.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/network.py
new file mode 100644
index 000000000..afd3298c7
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/network.py
@@ -0,0 +1,597 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper for managing networks."""
+
+import types
+import inspect
+import re
+import uuid
+import sys
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import Any, List, Tuple, Union
+
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+_import_handlers = []  # Custom import handlers for dealing with legacy data in pickle import.
+_import_module_src = dict()  # Source code for temporary modules created during pickle import.
+
+
+def import_handler(handler_func):
+    """Function decorator for declaring custom import handlers."""
+    _import_handlers.append(handler_func)
+    return handler_func
+
+
+class Network:
+    """Generic network abstraction.
+
+    Acts as a convenience wrapper for a parameterized network construction
+    function, providing several utility methods and convenient access to
+    the inputs/outputs/weights.
+
+    Network objects can be safely pickled and unpickled for long-term
+    archival purposes. The pickling works reliably as long as the underlying
+    network construction function is defined in a standalone Python module
+    that has no side effects or application-specific imports.
+
+    Args:
+        name: Network name. Used to select TensorFlow name and variable scopes.
+        func_name: Fully qualified name of the underlying network construction function, or a top-level function object.
+        static_kwargs: Keyword arguments to be passed in to the network construction function.
+
+    Attributes:
+        name: User-specified name, defaults to build func name if None.
+        scope: Unique TensorFlow scope containing template graph and variables, derived from the user-specified name.
+        static_kwargs: Arguments passed to the user-supplied build func.
+        components: Container for sub-networks. Passed to the build func, and retained between calls.
+        num_inputs: Number of input tensors.
+        num_outputs: Number of output tensors.
+        input_shapes: Input tensor shapes (NC or NCHW), including minibatch dimension.
+        output_shapes: Output tensor shapes (NC or NCHW), including minibatch dimension.
+        input_shape: Short-hand for input_shapes[0].
+        output_shape: Short-hand for output_shapes[0].
+        input_templates: Input placeholders in the template graph.
+        output_templates: Output tensors in the template graph.
+        input_names: Name string for each input.
+        output_names: Name string for each output.
+        own_vars: Variables defined by this network (local_name => var), excluding sub-networks.
+        vars: All variables (local_name => var).
+        trainables: All trainable variables (local_name => var).
+        var_global_to_local: Mapping from variable global names to local names.
+    """
+
+    def __init__(self, name: str = None, func_name: Any = None, **static_kwargs):
+        tfutil.assert_tf_initialized()
+        assert isinstance(name, str) or name is None
+        assert func_name is not None
+        assert isinstance(func_name, str) or util.is_top_level_function(func_name)
+        assert util.is_pickleable(static_kwargs)
+
+        self._init_fields()
+        self.name = name
+        self.static_kwargs = util.EasyDict(static_kwargs)
+
+        # Locate the user-specified network build function.
+        if util.is_top_level_function(func_name):
+            func_name = util.get_top_level_function_name(func_name)
+        module, self._build_func_name = util.get_module_from_obj_name(func_name)
+        self._build_func = util.get_obj_from_module(module, self._build_func_name)
+        assert callable(self._build_func)
+
+        # Dig up source code for the module containing the build function.
+        self._build_module_src = _import_module_src.get(module, None)
+        if self._build_module_src is None:
+            self._build_module_src = inspect.getsource(module)
+
+        # Init TensorFlow graph.
+        self._init_graph()
+        self.reset_own_vars()
+
+    def _init_fields(self) -> None:
+        self.name = None
+        self.scope = None
+        self.static_kwargs = util.EasyDict()
+        self.components = util.EasyDict()
+        self.num_inputs = 0
+        self.num_outputs = 0
+        self.input_shapes = [[]]
+        self.output_shapes = [[]]
+        self.input_shape = []
+        self.output_shape = []
+        self.input_templates = []
+        self.output_templates = []
+        self.input_names = []
+        self.output_names = []
+        self.own_vars = OrderedDict()
+        self.vars = OrderedDict()
+        self.trainables = OrderedDict()
+        self.var_global_to_local = OrderedDict()
+
+        self._build_func = None  # User-supplied build function that constructs the network.
+        self._build_func_name = None  # Name of the build function.
+        self._build_module_src = None  # Full source code of the module containing the build function.
+        self._run_cache = dict()  # Cached graph data for Network.run().
+
+    def _init_graph(self) -> None:
+        # Collect inputs.
+        self.input_names = []
+
+        for param in inspect.signature(self._build_func).parameters.values():
+            if param.kind == param.POSITIONAL_OR_KEYWORD and param.default is param.empty:
+                self.input_names.append(param.name)
+
+        self.num_inputs = len(self.input_names)
+        assert self.num_inputs >= 1
+
+        # Choose name and scope.
+        if self.name is None:
+            self.name = self._build_func_name
+        assert re.match("^[A-Za-z0-9_.\\-]*$", self.name)
+        with tf.name_scope(None):
+            self.scope = tf.get_default_graph().unique_name(self.name, mark_as_used=True)
+
+        # Finalize build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs["is_template_graph"] = True
+        build_kwargs["components"] = self.components
+
+        # Build template graph.
+        with tfutil.absolute_variable_scope(self.scope, reuse=False), tfutil.absolute_name_scope(self.scope):  # ignore surrounding scopes
+            assert tf.get_variable_scope().name == self.scope
+            assert tf.get_default_graph().get_name_scope() == self.scope
+            with tf.control_dependencies(None):  # ignore surrounding control dependencies
+                self.input_templates = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
+                out_expr = self._build_func(*self.input_templates, **build_kwargs)
+
+        # Collect outputs.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        self.output_templates = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        self.num_outputs = len(self.output_templates)
+        assert self.num_outputs >= 1
+        assert all(tfutil.is_tf_expression(t) for t in self.output_templates)
+
+        # Perform sanity checks.
+        if any(t.shape.ndims is None for t in self.input_templates):
+            raise ValueError("Network input shapes not defined. Please call x.set_shape() for each input.")
+        if any(t.shape.ndims is None for t in self.output_templates):
+            raise ValueError("Network output shapes not defined. Please call x.set_shape() where applicable.")
+        if any(not isinstance(comp, Network) for comp in self.components.values()):
+            raise ValueError("Components of a Network must be Networks themselves.")
+        if len(self.components) != len(set(comp.name for comp in self.components.values())):
+            raise ValueError("Components of a Network must have unique names.")
+
+        # List inputs and outputs.
+        self.input_shapes = [t.shape.as_list() for t in self.input_templates]
+        self.output_shapes = [t.shape.as_list() for t in self.output_templates]
+        self.input_shape = self.input_shapes[0]
+        self.output_shape = self.output_shapes[0]
+        self.output_names = [t.name.split("/")[-1].split(":")[0] for t in self.output_templates]
+
+        # List variables.
+        self.own_vars = OrderedDict((var.name[len(self.scope) + 1:].split(":")[0], var) for var in tf.global_variables(self.scope + "/"))
+        self.vars = OrderedDict(self.own_vars)
+        self.vars.update((comp.name + "/" + name, var) for comp in self.components.values() for name, var in comp.vars.items())
+        self.trainables = OrderedDict((name, var) for name, var in self.vars.items() if var.trainable)
+        self.var_global_to_local = OrderedDict((var.name.split(":")[0], name) for name, var in self.vars.items())
+
+    def reset_own_vars(self) -> None:
+        """Re-initialize all variables of this network, excluding sub-networks."""
+        tfutil.run([var.initializer for var in self.own_vars.values()])
+
+    def reset_vars(self) -> None:
+        """Re-initialize all variables of this network, including sub-networks."""
+        tfutil.run([var.initializer for var in self.vars.values()])
+
+    def reset_trainables(self) -> None:
+        """Re-initialize all trainable variables of this network, including sub-networks."""
+        tfutil.run([var.initializer for var in self.trainables.values()])
+
+    def get_output_for(self, *in_expr: TfExpression, return_as_list: bool = False, **dynamic_kwargs) -> Union[TfExpression, List[TfExpression]]:
+        """Construct TensorFlow expression(s) for the output(s) of this network, given the input expression(s)."""
+        assert len(in_expr) == self.num_inputs
+        assert not all(expr is None for expr in in_expr)
+
+        # Finalize build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs.update(dynamic_kwargs)
+        build_kwargs["is_template_graph"] = False
+        build_kwargs["components"] = self.components
+
+        # Build TensorFlow graph to evaluate the network.
+        with tfutil.absolute_variable_scope(self.scope, reuse=True), tf.name_scope(self.name):
+            assert tf.get_variable_scope().name == self.scope
+            valid_inputs = [expr for expr in in_expr if expr is not None]
+            final_inputs = []
+            for expr, name, shape in zip(in_expr, self.input_names, self.input_shapes):
+                if expr is not None:
+                    expr = tf.identity(expr, name=name)
+                else:
+                    expr = tf.zeros([tf.shape(valid_inputs[0])[0]] + shape[1:], name=name)
+                final_inputs.append(expr)
+            out_expr = self._build_func(*final_inputs, **build_kwargs)
+
+        # Propagate input shapes back to the user-specified expressions.
+        for expr, final in zip(in_expr, final_inputs):
+            if isinstance(expr, tf.Tensor):
+                expr.set_shape(final.shape)
+
+        # Express outputs in the desired format.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        if return_as_list:
+            out_expr = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        return out_expr
+
+    def get_var_local_name(self, var_or_global_name: Union[TfExpression, str]) -> str:
+        """Get the local name of a given variable, without any surrounding name scopes."""
+        assert tfutil.is_tf_expression(var_or_global_name) or isinstance(var_or_global_name, str)
+        global_name = var_or_global_name if isinstance(var_or_global_name, str) else var_or_global_name.name
+        return self.var_global_to_local[global_name]
+
+    def find_var(self, var_or_local_name: Union[TfExpression, str]) -> TfExpression:
+        """Find variable by local or global name."""
+        assert tfutil.is_tf_expression(var_or_local_name) or isinstance(var_or_local_name, str)
+        return self.vars[var_or_local_name] if isinstance(var_or_local_name, str) else var_or_local_name
+
+    def get_var(self, var_or_local_name: Union[TfExpression, str]) -> np.ndarray:
+        """Get the value of a given variable as NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.run(list_of_vars) whenever possible."""
+        return self.find_var(var_or_local_name).eval()
+
+    def set_var(self, var_or_local_name: Union[TfExpression, str], new_value: Union[int, float, np.ndarray]) -> None:
+        """Set the value of a given variable based on the given NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.set_vars() whenever possible."""
+        tfutil.set_vars({self.find_var(var_or_local_name): new_value})
+
+    def __getstate__(self) -> dict:
+        """Pickle export."""
+        state = dict()
+        state["version"]            = 4
+        state["name"]               = self.name
+        state["static_kwargs"]      = dict(self.static_kwargs)
+        state["components"]         = dict(self.components)
+        state["build_module_src"]   = self._build_module_src
+        state["build_func_name"]    = self._build_func_name
+        state["variables"]          = list(zip(self.own_vars.keys(), tfutil.run(list(self.own_vars.values()))))
+        return state
+
+    def __setstate__(self, state: dict) -> None:
+        """Pickle import."""
+        # pylint: disable=attribute-defined-outside-init
+        tfutil.assert_tf_initialized()
+        self._init_fields()
+
+        # Execute custom import handlers.
+        for handler in _import_handlers:
+            state = handler(state)
+
+        # Set basic fields.
+        assert state["version"] in [2, 3, 4]
+        self.name = state["name"]
+        self.static_kwargs = util.EasyDict(state["static_kwargs"])
+        self.components = util.EasyDict(state.get("components", {}))
+        self._build_module_src = state["build_module_src"]
+        self._build_func_name = state["build_func_name"]
+
+        # Create temporary module from the imported source code.
+        module_name = "_tflib_network_import_" + uuid.uuid4().hex
+        module = types.ModuleType(module_name)
+        sys.modules[module_name] = module
+        _import_module_src[module] = self._build_module_src
+        exec(self._build_module_src, module.__dict__) # pylint: disable=exec-used
+
+        # Locate network build function in the temporary module.
+        self._build_func = util.get_obj_from_module(module, self._build_func_name)
+        assert callable(self._build_func)
+
+        # Init TensorFlow graph.
+        self._init_graph()
+        self.reset_own_vars()
+        tfutil.set_vars({self.find_var(name): value for name, value in state["variables"]})
+
+    def clone(self, name: str = None, **new_static_kwargs) -> "Network":
+        """Create a clone of this network with its own copy of the variables."""
+        # pylint: disable=protected-access
+        net = object.__new__(Network)
+        net._init_fields()
+        net.name = name if name is not None else self.name
+        net.static_kwargs = util.EasyDict(self.static_kwargs)
+        net.static_kwargs.update(new_static_kwargs)
+        net._build_module_src = self._build_module_src
+        net._build_func_name = self._build_func_name
+        net._build_func = self._build_func
+        net._init_graph()
+        net.copy_vars_from(self)
+        return net
+
+    def copy_own_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, excluding sub-networks."""
+        names = [name for name in self.own_vars.keys() if name in src_net.own_vars]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def copy_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, including sub-networks."""
+        names = [name for name in self.vars.keys() if name in src_net.vars]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def copy_trainables_from(self, src_net: "Network") -> None:
+        """Copy the values of all trainable variables from the given network, including sub-networks."""
+        names = [name for name in self.trainables.keys() if name in src_net.trainables]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def convert(self, new_func_name: str, new_name: str = None, **new_static_kwargs) -> "Network":
+        """Create new network with the given parameters, and copy all variables from this network."""
+        if new_name is None:
+            new_name = self.name
+        static_kwargs = dict(self.static_kwargs)
+        static_kwargs.update(new_static_kwargs)
+        net = Network(name=new_name, func_name=new_func_name, **static_kwargs)
+        net.copy_vars_from(self)
+        return net
+
+    def setup_as_moving_average_of(self, src_net: "Network", beta: TfExpressionEx = 0.99, beta_nontrainable: TfExpressionEx = 0.0) -> tf.Operation:
+        """Construct a TensorFlow op that updates the variables of this network
+        to be slightly closer to those of the given network."""
+        with tfutil.absolute_name_scope(self.scope + "/_MovingAvg"):
+            ops = []
+            for name, var in self.vars.items():
+                if name in src_net.vars:
+                    cur_beta = beta if name in self.trainables else beta_nontrainable
+                    new_value = tfutil.lerp(src_net.vars[name], var, cur_beta)
+                    ops.append(var.assign(new_value))
+            return tf.group(*ops)
+
+    def run(self,
+            *in_arrays: Tuple[Union[np.ndarray, None], ...],
+            input_transform: dict = None,
+            output_transform: dict = None,
+            return_as_list: bool = False,
+            print_progress: bool = False,
+            minibatch_size: int = None,
+            num_gpus: int = 1,
+            assume_frozen: bool = False,
+            custom_inputs: Any = None,
+            **dynamic_kwargs) -> Union[np.ndarray, Tuple[np.ndarray, ...], List[np.ndarray]]:
+        """Run this network for the given NumPy array(s), and return the output(s) as NumPy array(s).
+
+        Args:
+            input_transform:    A dict specifying a custom transformation to be applied to the input tensor(s) before evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the input
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            output_transform:   A dict specifying a custom transformation to be applied to the output tensor(s) after evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the output
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            return_as_list:     True = return a list of NumPy arrays, False = return a single NumPy array, or a tuple if there are multiple outputs.
+            print_progress:     Print progress to the console? Useful for very large input arrays.
+            minibatch_size:     Maximum minibatch size to use, None = disable batching.
+            num_gpus:           Number of GPUs to use.
+            assume_frozen:      Improve multi-GPU performance by assuming that the trainable parameters will remain changed between calls.
+            custom_inputs:      Allow to use another tensor as input instead of default placeholders.
+            dynamic_kwargs:     Additional keyword arguments to be passed into the network build function.
+        """
+        assert len(in_arrays) == self.num_inputs
+        assert not all(arr is None for arr in in_arrays)
+        assert input_transform is None or util.is_top_level_function(input_transform["func"])
+        assert output_transform is None or util.is_top_level_function(output_transform["func"])
+        output_transform, dynamic_kwargs = _handle_legacy_output_transforms(output_transform, dynamic_kwargs)
+        num_items = in_arrays[0].shape[0]
+        if minibatch_size is None:
+            minibatch_size = num_items
+
+        # Construct unique hash key from all arguments that affect the TensorFlow graph.
+        key = dict(input_transform=input_transform, output_transform=output_transform, num_gpus=num_gpus, assume_frozen=assume_frozen, dynamic_kwargs=dynamic_kwargs)
+        def unwind_key(obj):
+            if isinstance(obj, dict):
+                return [(key, unwind_key(value)) for key, value in sorted(obj.items())]
+            if callable(obj):
+                return util.get_top_level_function_name(obj)
+            return obj
+        key = repr(unwind_key(key))
+
+        # Build graph.
+        if key not in self._run_cache:
+            with tfutil.absolute_name_scope(self.scope + "/_Run"), tf.control_dependencies(None):
+                if custom_inputs is not None:
+                    with tf.device("/gpu:0"):
+                        in_expr = [input_builder(name) for input_builder, name in zip(custom_inputs, self.input_names)]
+                        in_split = list(zip(*[tf.split(x, num_gpus) for x in in_expr]))
+                else:
+                    with tf.device("/cpu:0"):
+                        in_expr = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
+                        in_split = list(zip(*[tf.split(x, num_gpus) for x in in_expr]))
+
+                out_split = []
+                for gpu in range(num_gpus):
+                    with tf.device("/gpu:%d" % gpu):
+                        net_gpu = self.clone() if assume_frozen else self
+                        in_gpu = in_split[gpu]
+
+                        if input_transform is not None:
+                            in_kwargs = dict(input_transform)
+                            in_gpu = in_kwargs.pop("func")(*in_gpu, **in_kwargs)
+                            in_gpu = [in_gpu] if tfutil.is_tf_expression(in_gpu) else list(in_gpu)
+
+                        assert len(in_gpu) == self.num_inputs
+                        out_gpu = net_gpu.get_output_for(*in_gpu, return_as_list=True, **dynamic_kwargs)
+
+                        if output_transform is not None:
+                            out_kwargs = dict(output_transform)
+                            out_gpu = out_kwargs.pop("func")(*out_gpu, **out_kwargs)
+                            out_gpu = [out_gpu] if tfutil.is_tf_expression(out_gpu) else list(out_gpu)
+
+                        assert len(out_gpu) == self.num_outputs
+                        out_split.append(out_gpu)
+
+                with tf.device("/cpu:0"):
+                    out_expr = [tf.concat(outputs, axis=0) for outputs in zip(*out_split)]
+                    self._run_cache[key] = in_expr, out_expr
+
+        # Run minibatches.
+        in_expr, out_expr = self._run_cache[key]
+        out_arrays = [np.empty([num_items] + expr.shape.as_list()[1:], expr.dtype.name) for expr in out_expr]
+
+        for mb_begin in range(0, num_items, minibatch_size):
+            if print_progress:
+                print("\r%d / %d" % (mb_begin, num_items), end="")
+
+            mb_end = min(mb_begin + minibatch_size, num_items)
+            mb_num = mb_end - mb_begin
+            mb_in = [src[mb_begin : mb_end] if src is not None else np.zeros([mb_num] + shape[1:]) for src, shape in zip(in_arrays, self.input_shapes)]
+            mb_out = tf.get_default_session().run(out_expr, dict(zip(in_expr, mb_in)))
+
+            for dst, src in zip(out_arrays, mb_out):
+                dst[mb_begin: mb_end] = src
+
+        # Done.
+        if print_progress:
+            print("\r%d / %d" % (num_items, num_items))
+
+        if not return_as_list:
+            out_arrays = out_arrays[0] if len(out_arrays) == 1 else tuple(out_arrays)
+        return out_arrays
+
+    def list_ops(self) -> List[TfExpression]:
+        include_prefix = self.scope + "/"
+        exclude_prefix = include_prefix + "_"
+        ops = tf.get_default_graph().get_operations()
+        ops = [op for op in ops if op.name.startswith(include_prefix)]
+        ops = [op for op in ops if not op.name.startswith(exclude_prefix)]
+        return ops
+
+    def list_layers(self) -> List[Tuple[str, TfExpression, List[TfExpression]]]:
+        """Returns a list of (layer_name, output_expr, trainable_vars) tuples corresponding to
+        individual layers of the network. Mainly intended to be used for reporting."""
+        layers = []
+
+        def recurse(scope, parent_ops, parent_vars, level):
+            # Ignore specific patterns.
+            if any(p in scope for p in ["/Shape", "/strided_slice", "/Cast", "/concat", "/Assign"]):
+                return
+
+            # Filter ops and vars by scope.
+            global_prefix = scope + "/"
+            local_prefix = global_prefix[len(self.scope) + 1:]
+            cur_ops = [op for op in parent_ops if op.name.startswith(global_prefix) or op.name == global_prefix[:-1]]
+            cur_vars = [(name, var) for name, var in parent_vars if name.startswith(local_prefix) or name == local_prefix[:-1]]
+            if not cur_ops and not cur_vars:
+                return
+
+            # Filter out all ops related to variables.
+            for var in [op for op in cur_ops if op.type.startswith("Variable")]:
+                var_prefix = var.name + "/"
+                cur_ops = [op for op in cur_ops if not op.name.startswith(var_prefix)]
+
+            # Scope does not contain ops as immediate children => recurse deeper.
+            contains_direct_ops = any("/" not in op.name[len(global_prefix):] and op.type not in ["Identity", "Cast", "Transpose"] for op in cur_ops)
+            if (level == 0 or not contains_direct_ops) and (len(cur_ops) + len(cur_vars)) > 1:
+                visited = set()
+                for rel_name in [op.name[len(global_prefix):] for op in cur_ops] + [name[len(local_prefix):] for name, _var in cur_vars]:
+                    token = rel_name.split("/")[0]
+                    if token not in visited:
+                        recurse(global_prefix + token, cur_ops, cur_vars, level + 1)
+                        visited.add(token)
+                return
+
+            # Report layer.
+            layer_name = scope[len(self.scope) + 1:]
+            layer_output = cur_ops[-1].outputs[0] if cur_ops else cur_vars[-1][1]
+            layer_trainables = [var for _name, var in cur_vars if var.trainable]
+            layers.append((layer_name, layer_output, layer_trainables))
+
+        recurse(self.scope, self.list_ops(), list(self.vars.items()), 0)
+        return layers
+
+    def print_layers(self, title: str = None, hide_layers_with_no_params: bool = False) -> None:
+        """Print a summary table of the network structure."""
+        rows = [[title if title is not None else self.name, "Params", "OutputShape", "WeightShape"]]
+        rows += [["---"] * 4]
+        total_params = 0
+
+        for layer_name, layer_output, layer_trainables in self.list_layers():
+            num_params = sum(int(np.prod(var.shape.as_list())) for var in layer_trainables)
+            weights = [var for var in layer_trainables if var.name.endswith("/weight:0")]
+            weights.sort(key=lambda x: len(x.name))
+            if len(weights) == 0 and len(layer_trainables) == 1:
+                weights = layer_trainables
+            total_params += num_params
+
+            if not hide_layers_with_no_params or num_params != 0:
+                num_params_str = str(num_params) if num_params > 0 else "-"
+                output_shape_str = str(layer_output.shape)
+                weight_shape_str = str(weights[0].shape) if len(weights) >= 1 else "-"
+                rows += [[layer_name, num_params_str, output_shape_str, weight_shape_str]]
+
+        rows += [["---"] * 4]
+        rows += [["Total", str(total_params), "", ""]]
+
+        widths = [max(len(cell) for cell in column) for column in zip(*rows)]
+        print()
+        for row in rows:
+            print("  ".join(cell + " " * (width - len(cell)) for cell, width in zip(row, widths)))
+        print()
+
+    def setup_weight_histograms(self, title: str = None) -> None:
+        """Construct summary ops to include histograms of all trainable parameters in TensorBoard."""
+        if title is None:
+            title = self.name
+
+        with tf.name_scope(None), tf.device(None), tf.control_dependencies(None):
+            for local_name, var in self.trainables.items():
+                if "/" in local_name:
+                    p = local_name.split("/")
+                    name = title + "_" + p[-1] + "/" + "_".join(p[:-1])
+                else:
+                    name = title + "_toplevel/" + local_name
+
+                tf.summary.histogram(name, var)
+
+#----------------------------------------------------------------------------
+# Backwards-compatible emulation of legacy output transformation in Network.run().
+
+_print_legacy_warning = True
+
+def _handle_legacy_output_transforms(output_transform, dynamic_kwargs):
+    global _print_legacy_warning
+    legacy_kwargs = ["out_mul", "out_add", "out_shrink", "out_dtype"]
+    if not any(kwarg in dynamic_kwargs for kwarg in legacy_kwargs):
+        return output_transform, dynamic_kwargs
+
+    if _print_legacy_warning:
+        _print_legacy_warning = False
+        print()
+        print("WARNING: Old-style output transformations in Network.run() are deprecated.")
+        print("Consider using 'output_transform=dict(func=tflib.convert_images_to_uint8)'")
+        print("instead of 'out_mul=127.5, out_add=127.5, out_dtype=np.uint8'.")
+        print()
+    assert output_transform is None
+
+    new_kwargs = dict(dynamic_kwargs)
+    new_transform = {kwarg: new_kwargs.pop(kwarg) for kwarg in legacy_kwargs if kwarg in dynamic_kwargs}
+    new_transform["func"] = _legacy_output_transform_func
+    return new_transform, new_kwargs
+
+def _legacy_output_transform_func(*expr, out_mul=1.0, out_add=0.0, out_shrink=1, out_dtype=None):
+    if out_mul != 1.0:
+        expr = [x * out_mul for x in expr]
+
+    if out_add != 0.0:
+        expr = [x + out_add for x in expr]
+
+    if out_shrink > 1:
+        ksize = [1, 1, out_shrink, out_shrink]
+        expr = [tf.nn.avg_pool(x, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW") for x in expr]
+
+    if out_dtype is not None:
+        if tf.as_dtype(out_dtype).is_integer:
+            expr = [tf.round(x) for x in expr]
+        expr = [tf.saturate_cast(x, out_dtype) for x in expr]
+    return expr
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/__init__.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/__init__.py
new file mode 100644
index 000000000..9ab9908ef
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/__init__.py
@@ -0,0 +1,7 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+# empty
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.cu b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.cu
new file mode 100644
index 000000000..1102f624f
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.cu
@@ -0,0 +1,188 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+//------------------------------------------------------------------------
+// CUDA kernel.
+
+template <class T>
+struct FusedBiasActKernelParams
+{
+    const T*    x;      // [sizeX]
+    const T*    b;      // [sizeB] or NULL
+    const T*    ref;    // [sizeX] or NULL
+    T*          y;      // [sizeX]
+
+    int         grad;
+    int         axis;
+    int         act;
+    float       alpha;
+    float       gain;
+
+    int         sizeX;
+    int         sizeB;
+    int         stepB;
+    int         loopX;
+};
+
+template <class T>
+static __global__ void FusedBiasActKernel(const FusedBiasActKernelParams<T> p)
+{
+    const float expRange        = 80.0f;
+    const float halfExpRange    = 40.0f;
+    const float seluScale       = 1.0507009873554804934193349852946f;
+    const float seluAlpha       = 1.6732632423543772848170429916717f;
+
+    // Loop over elements.
+    int xi = blockIdx.x * p.loopX * blockDim.x + threadIdx.x;
+    for (int loopIdx = 0; loopIdx < p.loopX && xi < p.sizeX; loopIdx++, xi += blockDim.x)
+    {
+        // Load and apply bias.
+        float x = (float)p.x[xi];
+        if (p.b)
+            x += (float)p.b[(xi / p.stepB) % p.sizeB];
+        float ref = (p.ref) ? (float)p.ref[xi] : 0.0f;
+        if (p.gain != 0.0f & p.act != 9)
+            ref /= p.gain;
+
+        // Evaluate activation func.
+        float y;
+        switch (p.act * 10 + p.grad)
+        {
+            // linear
+            default:
+            case 10: y = x; break;
+            case 11: y = x; break;
+            case 12: y = 0.0f; break;
+
+            // relu
+            case 20: y = (x > 0.0f) ? x : 0.0f; break;
+            case 21: y = (ref > 0.0f) ? x : 0.0f; break;
+            case 22: y = 0.0f; break;
+
+            // lrelu
+            case 30: y = (x > 0.0f) ? x : x * p.alpha; break;
+            case 31: y = (ref > 0.0f) ? x : x * p.alpha; break;
+            case 32: y = 0.0f; break;
+
+            // tanh
+            case 40: { float c = expf(x); float d = 1.0f / c; y = (x < -expRange) ? -1.0f : (x > expRange) ? 1.0f : (c - d) / (c + d); } break;
+            case 41: y = x * (1.0f - ref * ref); break;
+            case 42: y = x * (1.0f - ref * ref) * (-2.0f * ref); break;
+
+            // sigmoid
+            case 50: y = (x < -expRange) ? 0.0f : 1.0f / (expf(-x) + 1.0f); break;
+            case 51: y = x * ref * (1.0f - ref); break;
+            case 52: y = x * ref * (1.0f - ref) * (1.0f - 2.0f * ref); break;
+
+            // elu
+            case 60: y = (x >= 0.0f) ? x : expf(x) - 1.0f; break;
+            case 61: y = (ref >= 0.0f) ? x : x * (ref + 1.0f); break;
+            case 62: y = (ref >= 0.0f) ? 0.0f : x * (ref + 1.0f); break;
+
+            // selu
+            case 70: y = (x >= 0.0f) ? seluScale * x : (seluScale * seluAlpha) * (expf(x) - 1.0f); break;
+            case 71: y = (ref >= 0.0f) ? x * seluScale : x * (ref + seluScale * seluAlpha); break;
+            case 72: y = (ref >= 0.0f) ? 0.0f : x * (ref + seluScale * seluAlpha); break;
+
+            // softplus
+            case 80: y = (x > expRange) ? x : logf(expf(x) + 1.0f); break;
+            case 81: y = x * (1.0f - expf(-ref)); break;
+            case 82: { float c = expf(-ref); y = x * c * (1.0f - c); } break;
+
+            // swish
+            case 90: y = (x < -expRange) ? 0.0f : x / (expf(-x) + 1.0f); break;
+            case 91: { float c = expf(ref); float d = c + 1.0f; y = (ref > halfExpRange) ? x : x * c * (ref + d) / (d * d); } break;
+            case 92: { float c = expf(ref); float d = c + 1.0f; y = (ref > halfExpRange) ? 0.0f : x * c * (ref * (2.0f - d) + 2.0f * d) / (d * d * d); } break;
+        }
+
+        // Apply gain and store.
+        p.y[xi] = (T)(y * p.gain);
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct FusedBiasActOp : public OpKernel
+{
+    FusedBiasActKernelParams<T> m_attribs;
+
+    FusedBiasActOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("grad", &m_attribs.grad));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("axis", &m_attribs.axis));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("act", &m_attribs.act));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &m_attribs.alpha));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("gain", &m_attribs.gain));
+        OP_REQUIRES(ctx, m_attribs.grad >= 0, errors::InvalidArgument("grad must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.axis >= 0, errors::InvalidArgument("axis must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.act >= 0, errors::InvalidArgument("act must be non-negative"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        FusedBiasActKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x     = ctx->input(0); // [...]
+        const Tensor& b     = ctx->input(1); // [sizeB] or [0]
+        const Tensor& ref   = ctx->input(2); // x.shape or [0]
+        p.x = x.flat<T>().data();
+        p.b = (b.NumElements()) ? b.flat<T>().data() : NULL;
+        p.ref = (ref.NumElements()) ? ref.flat<T>().data() : NULL;
+        OP_REQUIRES(ctx, b.NumElements() == 0 || m_attribs.axis < x.dims(), errors::InvalidArgument("axis out of bounds"));
+        OP_REQUIRES(ctx, b.dims() == 1, errors::InvalidArgument("b must have rank 1"));
+        OP_REQUIRES(ctx, b.NumElements() == 0 || b.NumElements() == x.dim_size(m_attribs.axis), errors::InvalidArgument("b has wrong number of elements"));
+        OP_REQUIRES(ctx, ref.NumElements() == ((p.grad == 0) ? 0 : x.NumElements()), errors::InvalidArgument("ref has wrong number of elements"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("x is too large"));
+
+        p.sizeX = (int)x.NumElements();
+        p.sizeB = (int)b.NumElements();
+        p.stepB = 1;
+        for (int i = m_attribs.axis + 1; i < x.dims(); i++)
+            p.stepB *= (int)x.dim_size(i);
+
+        Tensor* y = NULL; // x.shape
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, x.shape(), &y));
+        p.y = y->flat<T>().data();
+
+        p.loopX = 4;
+        int blockSize = 4 * 32;
+        int gridSize = (p.sizeX - 1) / (p.loopX * blockSize) + 1;
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel((void*)FusedBiasActKernel<T>, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("FusedBiasAct")
+    .Input      ("x: T")
+    .Input      ("b: T")
+    .Input      ("ref: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("grad: int = 0")
+    .Attr       ("axis: int = 1")
+    .Attr       ("act: int = 0")
+    .Attr       ("alpha: float = 0.0")
+    .Attr       ("gain: float = 1.0");
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<float>("T"), FusedBiasActOp<float>);
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), FusedBiasActOp<Eigen::half>);
+
+//------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.py
new file mode 100644
index 000000000..52f6bfd77
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/fused_bias_act.py
@@ -0,0 +1,196 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Custom TensorFlow ops for efficient bias and activation."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+from ...util import EasyDict
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+activation_funcs = {
+    'linear':   EasyDict(func=lambda x, **_:        x,                          def_alpha=None, def_gain=1.0,           cuda_idx=1, ref='y', zero_2nd_grad=True),
+    'relu':     EasyDict(func=lambda x, **_:        tf.nn.relu(x),              def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=2, ref='y', zero_2nd_grad=True),
+    'lrelu':    EasyDict(func=lambda x, alpha, **_: tf.nn.leaky_relu(x, alpha), def_alpha=0.2,  def_gain=np.sqrt(2),    cuda_idx=3, ref='y', zero_2nd_grad=True),
+    'tanh':     EasyDict(func=lambda x, **_:        tf.nn.tanh(x),              def_alpha=None, def_gain=1.0,           cuda_idx=4, ref='y', zero_2nd_grad=False),
+    'sigmoid':  EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x),           def_alpha=None, def_gain=1.0,           cuda_idx=5, ref='y', zero_2nd_grad=False),
+    'elu':      EasyDict(func=lambda x, **_:        tf.nn.elu(x),               def_alpha=None, def_gain=1.0,           cuda_idx=6, ref='y', zero_2nd_grad=False),
+    'selu':     EasyDict(func=lambda x, **_:        tf.nn.selu(x),              def_alpha=None, def_gain=1.0,           cuda_idx=7, ref='y', zero_2nd_grad=False),
+    'softplus': EasyDict(func=lambda x, **_:        tf.nn.softplus(x),          def_alpha=None, def_gain=1.0,           cuda_idx=8, ref='y', zero_2nd_grad=False),
+    'swish':    EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x) * x,       def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=9, ref='x', zero_2nd_grad=False),
+}
+
+#----------------------------------------------------------------------------
+
+def fused_bias_act(x, b=None, axis=1, act='linear', alpha=None, gain=None, impl='cuda'):
+    r"""Fused bias and activation function.
+
+    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
+    and scales the result by `gain`. Each of the steps is optional. In most cases,
+    the fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports first and second order gradients,
+    but not third order gradients.
+
+    Args:
+        x:      Input activation tensor. Can have any shape, but if `b` is defined, the
+                dimension corresponding to `axis`, as well as the rank, must be known.
+        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                as `x`. The shape must be known, and it must match the dimension of `x`
+                corresponding to `axis`.
+        axis:   The dimension in `x` corresponding to the elements of `b`.
+                The value of `axis` is ignored if `b` is not specified.
+        act:    Name of the activation function to evaluate, or `"linear"` to disable.
+                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
+                See `activation_funcs` for a full list. `None` is not allowed.
+        alpha:  Shape parameter for the activation function, or `None` to use the default.
+        gain:   Scaling factor for the output tensor, or `None` to use default.
+                See `activation_funcs` for the default scaling of each activation function.
+                If unsure, consider specifying `1.0`.
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _fused_bias_act_ref,
+        'cuda': _fused_bias_act_cuda,
+    }
+    return impl_dict[impl](x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain)
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_ref(x, b, axis, act, alpha, gain):
+    """Slow reference implementation of `fused_bias_act()` using standard TensorFlow ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    b = tf.convert_to_tensor(b) if b is not None else tf.constant([], dtype=x.dtype)
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Add bias.
+    if b.shape[0] != 0:
+        x += tf.reshape(b, [-1 if i == axis else 1 for i in range(x.shape.rank)])
+
+    # Evaluate activation function.
+    x = act_spec.func(x, alpha=alpha)
+
+    # Scale by gain.
+    if gain != 1:
+        x *= gain
+    return x
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_cuda(x, b, axis, act, alpha, gain):
+    """Fast CUDA implementation of `fused_bias_act()` using custom ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    empty_tensor = tf.constant([], dtype=x.dtype)
+    b = tf.convert_to_tensor(b) if b is not None else empty_tensor
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Special cases.
+    if act == 'linear' and b is None and gain == 1.0:
+        return x
+    if act_spec.cuda_idx is None:
+        return _fused_bias_act_ref(x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain)
+
+    # CUDA kernel.
+    cuda_kernel = _get_plugin().fused_bias_act
+    cuda_kwargs = dict(axis=axis, act=act_spec.cuda_idx, alpha=alpha, gain=gain)
+
+    # Forward pass: y = func(x, b).
+    def func_y(x, b):
+        y = cuda_kernel(x=x, b=b, ref=empty_tensor, grad=0, **cuda_kwargs)
+        y.set_shape(x.shape)
+        return y
+
+    # Backward pass: dx, db = grad(dy, x, y)
+    def grad_dx(dy, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        dx = cuda_kernel(x=dy, b=empty_tensor, ref=ref, grad=1, **cuda_kwargs)
+        dx.set_shape(x.shape)
+        return dx
+    def grad_db(dx):
+        if b.shape[0] == 0:
+            return empty_tensor
+        db = dx
+        if axis < x.shape.rank - 1:
+            db = tf.reduce_sum(db, list(range(axis + 1, x.shape.rank)))
+        if axis > 0:
+            db = tf.reduce_sum(db, list(range(axis)))
+        db.set_shape(b.shape)
+        return db
+
+    # Second order gradients: d_dy, d_x = grad2(d_dx, d_db, x, y)
+    def grad2_d_dy(d_dx, d_db, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        d_dy = cuda_kernel(x=d_dx, b=d_db, ref=ref, grad=1, **cuda_kwargs)
+        d_dy.set_shape(x.shape)
+        return d_dy
+    def grad2_d_x(d_dx, d_db, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        d_x = cuda_kernel(x=d_dx, b=d_db, ref=ref, grad=2, **cuda_kwargs)
+        d_x.set_shape(x.shape)
+        return d_x
+
+    # Fast version for piecewise-linear activation funcs.
+    @tf.custom_gradient
+    def func_zero_2nd_grad(x, b):
+        y = func_y(x, b)
+        @tf.custom_gradient
+        def grad(dy):
+            dx = grad_dx(dy, x, y)
+            db = grad_db(dx)
+            def grad2(d_dx, d_db):
+                d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                return d_dy
+            return (dx, db), grad2
+        return y, grad
+
+    # Slow version for general activation funcs.
+    @tf.custom_gradient
+    def func_nonzero_2nd_grad(x, b):
+        y = func_y(x, b)
+        def grad_wrap(dy):
+            @tf.custom_gradient
+            def grad_impl(dy, x):
+                dx = grad_dx(dy, x, y)
+                db = grad_db(dx)
+                def grad2(d_dx, d_db):
+                    d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                    d_x = grad2_d_x(d_dx, d_db, x, y)
+                    return d_dy, d_x
+                return (dx, db), grad2
+            return grad_impl(dy, x)
+        return y, grad_wrap
+
+    # Which version to use?
+    if act_spec.zero_2nd_grad:
+        return func_zero_2nd_grad(x, b)
+    return func_nonzero_2nd_grad(x, b)
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.cu b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.cu
new file mode 100644
index 000000000..b97ef36c9
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.cu
@@ -0,0 +1,326 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+//------------------------------------------------------------------------
+// Helpers.
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+static __host__ __device__ __forceinline__ int floorDiv(int a, int b)
+{
+    int c = a / b;
+    if (c * b > a)
+        c--;
+    return c;
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel params.
+
+template <class T>
+struct UpFirDn2DKernelParams
+{
+    const T*    x;          // [majorDim, inH, inW, minorDim]
+    const T*    k;          // [kernelH, kernelW]
+    T*          y;          // [majorDim, outH, outW, minorDim]
+
+    int         upx;
+    int         upy;
+    int         downx;
+    int         downy;
+    int         padx0;
+    int         padx1;
+    int         pady0;
+    int         pady1;
+
+    int         majorDim;
+    int         inH;
+    int         inW;
+    int         minorDim;
+    int         kernelH;
+    int         kernelW;
+    int         outH;
+    int         outW;
+    int         loopMajor;
+    int         loopX;
+};
+
+//------------------------------------------------------------------------
+// General CUDA implementation for large filter kernels.
+
+template <class T>
+static __global__ void UpFirDn2DKernel_large(const UpFirDn2DKernelParams<T> p)
+{
+    // Calculate thread index.
+    int minorIdx = blockIdx.x * blockDim.x + threadIdx.x;
+    int outY = minorIdx / p.minorDim;
+    minorIdx -= outY * p.minorDim;
+    int outXBase = blockIdx.y * p.loopX * blockDim.y + threadIdx.y;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (outXBase >= p.outW || outY >= p.outH || majorIdxBase >= p.majorDim)
+        return;
+
+    // Setup Y receptive field.
+    int midY = outY * p.downy + p.upy - 1 - p.pady0;
+    int inY = min(max(floorDiv(midY, p.upy), 0), p.inH);
+    int h = min(max(floorDiv(midY + p.kernelH, p.upy), 0), p.inH) - inY;
+    int kernelY = midY + p.kernelH - (inY + 1) * p.upy;
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor && majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, outX = outXBase; loopX < p.loopX && outX < p.outW; loopX++, outX += blockDim.y)
+    {
+        // Setup X receptive field.
+        int midX = outX * p.downx + p.upx - 1 - p.padx0;
+        int inX = min(max(floorDiv(midX, p.upx), 0), p.inW);
+        int w = min(max(floorDiv(midX + p.kernelW, p.upx), 0), p.inW) - inX;
+        int kernelX = midX + p.kernelW - (inX + 1) * p.upx;
+
+        // Initialize pointers.
+        const T* xp = &p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+        const T* kp = &p.k[kernelY * p.kernelW + kernelX];
+        int xpx = p.minorDim;
+        int kpx = -p.upx;
+        int xpy = p.inW * p.minorDim;
+        int kpy = -p.upy * p.kernelW;
+
+        // Inner loop.
+        float v = 0.0f;
+        for (int y = 0; y < h; y++)
+        {
+            for (int x = 0; x < w; x++)
+            {
+                v += (float)(*xp) * (float)(*kp);
+                xp += xpx;
+                kp += kpx;
+            }
+            xp += xpy - w * xpx;
+            kp += kpy - w * kpx;
+        }
+
+        // Store result.
+        p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+    }
+}
+
+//------------------------------------------------------------------------
+// Specialized CUDA implementation for small filter kernels.
+
+template <class T, int upx, int upy, int downx, int downy, int kernelW, int kernelH, int tileOutW, int tileOutH>
+static __global__ void UpFirDn2DKernel_small(const UpFirDn2DKernelParams<T> p)
+{
+    //assert(kernelW % upx == 0);
+    //assert(kernelH % upy == 0);
+    const int tileInW = ((tileOutW - 1) * downx + kernelW - 1) / upx + 1;
+    const int tileInH = ((tileOutH - 1) * downy + kernelH - 1) / upy + 1;
+    __shared__ volatile float sk[kernelH][kernelW];
+    __shared__ volatile float sx[tileInH][tileInW];
+
+    // Calculate tile index.
+    int minorIdx = blockIdx.x;
+    int tileOutY = minorIdx / p.minorDim;
+    minorIdx -= tileOutY * p.minorDim;
+    tileOutY *= tileOutH;
+    int tileOutXBase = blockIdx.y * p.loopX * tileOutW;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (tileOutXBase >= p.outW | tileOutY >= p.outH | majorIdxBase >= p.majorDim)
+        return;
+
+    // Load filter kernel (flipped).
+    for (int tapIdx = threadIdx.x; tapIdx < kernelH * kernelW; tapIdx += blockDim.x)
+    {
+        int ky = tapIdx / kernelW;
+        int kx = tapIdx - ky * kernelW;
+        float v = 0.0f;
+        if (kx < p.kernelW & ky < p.kernelH)
+            v = (float)p.k[(p.kernelH - 1 - ky) * p.kernelW + (p.kernelW - 1 - kx)];
+        sk[ky][kx] = v;
+    }
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor & majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, tileOutX = tileOutXBase; loopX < p.loopX & tileOutX < p.outW; loopX++, tileOutX += tileOutW)
+    {
+        // Load input pixels.
+        int tileMidX = tileOutX * downx + upx - 1 - p.padx0;
+        int tileMidY = tileOutY * downy + upy - 1 - p.pady0;
+        int tileInX = floorDiv(tileMidX, upx);
+        int tileInY = floorDiv(tileMidY, upy);
+        __syncthreads();
+        for (int inIdx = threadIdx.x; inIdx < tileInH * tileInW; inIdx += blockDim.x)
+        {
+            int relInY = inIdx / tileInW;
+            int relInX = inIdx - relInY * tileInW;
+            int inX = relInX + tileInX;
+            int inY = relInY + tileInY;
+            float v = 0.0f;
+            if (inX >= 0 & inY >= 0 & inX < p.inW & inY < p.inH)
+                v = (float)p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+            sx[relInY][relInX] = v;
+        }
+
+        // Loop over output pixels.
+        __syncthreads();
+        for (int outIdx = threadIdx.x; outIdx < tileOutH * tileOutW; outIdx += blockDim.x)
+        {
+            int relOutY = outIdx / tileOutW;
+            int relOutX = outIdx - relOutY * tileOutW;
+            int outX = relOutX + tileOutX;
+            int outY = relOutY + tileOutY;
+
+            // Setup receptive field.
+            int midX = tileMidX + relOutX * downx;
+            int midY = tileMidY + relOutY * downy;
+            int inX = floorDiv(midX, upx);
+            int inY = floorDiv(midY, upy);
+            int relInX = inX - tileInX;
+            int relInY = inY - tileInY;
+            int kernelX = (inX + 1) * upx - midX - 1; // flipped
+            int kernelY = (inY + 1) * upy - midY - 1; // flipped
+
+            // Inner loop.
+            float v = 0.0f;
+            #pragma unroll
+            for (int y = 0; y < kernelH / upy; y++)
+                #pragma unroll
+                for (int x = 0; x < kernelW / upx; x++)
+                    v += sx[relInY + y][relInX + x] * sk[kernelY + y * upy][kernelX + x * upx];
+
+            // Store result.
+            if (outX < p.outW & outY < p.outH)
+                p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct UpFirDn2DOp : public OpKernel
+{
+    UpFirDn2DKernelParams<T> m_attribs;
+
+    UpFirDn2DOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upx", &m_attribs.upx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upy", &m_attribs.upy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downx", &m_attribs.downx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downy", &m_attribs.downy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx0", &m_attribs.padx0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx1", &m_attribs.padx1));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady0", &m_attribs.pady0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady1", &m_attribs.pady1));
+        OP_REQUIRES(ctx, m_attribs.upx >= 1 && m_attribs.upy >= 1, errors::InvalidArgument("upx and upy must be at least 1x1"));
+        OP_REQUIRES(ctx, m_attribs.downx >= 1 && m_attribs.downy >= 1, errors::InvalidArgument("downx and downy must be at least 1x1"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        UpFirDn2DKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x = ctx->input(0); // [majorDim, inH, inW, minorDim]
+        const Tensor& k = ctx->input(1); // [kernelH, kernelW]
+        p.x = x.flat<T>().data();
+        p.k = k.flat<T>().data();
+        OP_REQUIRES(ctx, x.dims() == 4, errors::InvalidArgument("input must have rank 4"));
+        OP_REQUIRES(ctx, k.dims() == 2, errors::InvalidArgument("kernel must have rank 2"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("input too large"));
+        OP_REQUIRES(ctx, k.NumElements() <= kint32max, errors::InvalidArgument("kernel too large"));
+
+        p.majorDim  = (int)x.dim_size(0);
+        p.inH       = (int)x.dim_size(1);
+        p.inW       = (int)x.dim_size(2);
+        p.minorDim  = (int)x.dim_size(3);
+        p.kernelH   = (int)k.dim_size(0);
+        p.kernelW   = (int)k.dim_size(1);
+        OP_REQUIRES(ctx, p.kernelW >= 1 && p.kernelH >= 1, errors::InvalidArgument("kernel must be at least 1x1"));
+
+        p.outW = (p.inW * p.upx + p.padx0 + p.padx1 - p.kernelW + p.downx) / p.downx;
+        p.outH = (p.inH * p.upy + p.pady0 + p.pady1 - p.kernelH + p.downy) / p.downy;
+        OP_REQUIRES(ctx, p.outW >= 1 && p.outH >= 1, errors::InvalidArgument("output must be at least 1x1"));
+
+        Tensor* y = NULL; // [majorDim, outH, outW, minorDim]
+        TensorShape ys;
+        ys.AddDim(p.majorDim);
+        ys.AddDim(p.outH);
+        ys.AddDim(p.outW);
+        ys.AddDim(p.minorDim);
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, ys, &y));
+        p.y = y->flat<T>().data();
+        OP_REQUIRES(ctx, y->NumElements() <= kint32max, errors::InvalidArgument("output too large"));
+
+        // Choose CUDA kernel to use.
+        void* cudaKernel = (void*)UpFirDn2DKernel_large<T>;
+        int tileOutW = -1;
+        int tileOutH = -1;
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 7 && p.kernelH <= 7) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 7,7, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 6,6, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 5 && p.kernelH <= 5) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 5,5, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 4,4, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 3 && p.kernelH <= 3) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 3,3, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 8 && p.kernelH <= 8) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 8,8, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 6,6, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 4,4, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 2 && p.kernelH <= 2) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 2,2, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 8 && p.kernelH <= 8) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 8,8, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 6,6, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 4,4, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 2 && p.kernelH <= 2) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 2,2, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+
+        // Choose launch params.
+        dim3 blockSize;
+        dim3 gridSize;
+        if (tileOutW > 0 && tileOutH > 0) // small
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 1;
+            blockSize = dim3(32 * 8, 1, 1);
+            gridSize = dim3(((p.outH - 1) / tileOutH + 1) * p.minorDim, (p.outW - 1) / (p.loopX * tileOutW) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+        else // large
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 4;
+            blockSize = dim3(4, 32, 1);
+            gridSize = dim3((p.outH * p.minorDim - 1) / blockSize.x + 1, (p.outW - 1) / (p.loopX * blockSize.y) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+
+        // Launch CUDA kernel.
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel(cudaKernel, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("UpFirDn2D")
+    .Input      ("x: T")
+    .Input      ("k: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("upx: int = 1")
+    .Attr       ("upy: int = 1")
+    .Attr       ("downx: int = 1")
+    .Attr       ("downy: int = 1")
+    .Attr       ("padx0: int = 0")
+    .Attr       ("padx1: int = 0")
+    .Attr       ("pady0: int = 0")
+    .Attr       ("pady1: int = 0");
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<float>("T"), UpFirDn2DOp<float>);
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), UpFirDn2DOp<Eigen::half>);
+
+//------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.py
new file mode 100644
index 000000000..fd23777eb
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/ops/upfirdn_2d.py
@@ -0,0 +1,364 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Custom TensorFlow ops for efficient resampling of 2D images."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+def upfirdn_2d(x, k, upx=1, upy=1, downx=1, downy=1, padx0=0, padx1=0, pady0=0, pady1=0, impl='cuda'):
+    r"""Pad, upsample, FIR filter, and downsample a batch of 2D images.
+
+    Accepts a batch of 2D images of the shape `[majorDim, inH, inW, minorDim]`
+    and performs the following operations for each image, batched across
+    `majorDim` and `minorDim`:
+
+    1. Pad the image with zeros by the specified number of pixels on each side
+       (`padx0`, `padx1`, `pady0`, `pady1`). Specifying a negative value
+       corresponds to cropping the image.
+
+    2. Upsample the image by inserting the zeros after each pixel (`upx`, `upy`).
+
+    3. Convolve the image with the specified 2D FIR filter (`k`), shrinking the
+       image so that the footprint of all output pixels lies within the input image.
+
+    4. Downsample the image by throwing away pixels (`downx`, `downy`).
+
+    This sequence of operations bears close resemblance to scipy.signal.upfirdn().
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:      Input tensor of the shape `[majorDim, inH, inW, minorDim]`.
+        k:      2D FIR filter of the shape `[firH, firW]`.
+        upx:    Integer upsampling factor along the X-axis (default: 1).
+        upy:    Integer upsampling factor along the Y-axis (default: 1).
+        downx:  Integer downsampling factor along the X-axis (default: 1).
+        downy:  Integer downsampling factor along the Y-axis (default: 1).
+        padx0:  Number of pixels to pad on the left side (default: 0).
+        padx1:  Number of pixels to pad on the right side (default: 0).
+        pady0:  Number of pixels to pad on the top side (default: 0).
+        pady1:  Number of pixels to pad on the bottom side (default: 0).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[majorDim, outH, outW, minorDim]`, and same datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _upfirdn_2d_ref,
+        'cuda': _upfirdn_2d_cuda,
+    }
+    return impl_dict[impl](x=x, k=k, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_ref(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Slow reference implementation of `upfirdn_2d()` using standard TensorFlow ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    assert x.shape.rank == 4
+    inH = x.shape[1].value
+    inW = x.shape[2].value
+    minorDim = _shape(x, 3)
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    # Upsample (insert zeros).
+    x = tf.reshape(x, [-1, inH, 1, inW, 1, minorDim])
+    x = tf.pad(x, [[0, 0], [0, 0], [0, upy - 1], [0, 0], [0, upx - 1], [0, 0]])
+    x = tf.reshape(x, [-1, inH * upy, inW * upx, minorDim])
+
+    # Pad (crop if negative).
+    x = tf.pad(x, [[0, 0], [max(pady0, 0), max(pady1, 0)], [max(padx0, 0), max(padx1, 0)], [0, 0]])
+    x = x[:, max(-pady0, 0) : x.shape[1].value - max(-pady1, 0), max(-padx0, 0) : x.shape[2].value - max(-padx1, 0), :]
+
+    # Convolve with filter.
+    x = tf.transpose(x, [0, 3, 1, 2])
+    x = tf.reshape(x, [-1, 1, inH * upy + pady0 + pady1, inW * upx + padx0 + padx1])
+    w = tf.constant(k[::-1, ::-1, np.newaxis, np.newaxis], dtype=x.dtype)
+    x = tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='VALID', data_format='NCHW')
+    x = tf.reshape(x, [-1, minorDim, inH * upy + pady0 + pady1 - kernelH + 1, inW * upx + padx0 + padx1 - kernelW + 1])
+    x = tf.transpose(x, [0, 2, 3, 1])
+
+    # Downsample (throw away pixels).
+    return x[:, ::downy, ::downx, :]
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_cuda(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Fast CUDA implementation of `upfirdn_2d()` using custom ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    majorDim, inH, inW, minorDim = x.shape.as_list()
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    outW = (inW * upx + padx0 + padx1 - kernelW) // downx + 1
+    outH = (inH * upy + pady0 + pady1 - kernelH) // downy + 1
+    assert outW >= 1 and outH >= 1
+
+    kc = tf.constant(k, dtype=x.dtype)
+    gkc = tf.constant(k[::-1, ::-1], dtype=x.dtype)
+    gpadx0 = kernelW - padx0 - 1
+    gpady0 = kernelH - pady0 - 1
+    gpadx1 = inW * upx - outW * downx + padx0 - upx + 1
+    gpady1 = inH * upy - outH * downy + pady0 - upy + 1
+
+    @tf.custom_gradient
+    def func(x):
+        y = _get_plugin().up_fir_dn2d(x=x, k=kc, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
+        y.set_shape([majorDim, outH, outW, minorDim])
+        @tf.custom_gradient
+        def grad(dy):
+            dx = _get_plugin().up_fir_dn2d(x=dy, k=gkc, upx=downx, upy=downy, downx=upx, downy=upy, padx0=gpadx0, padx1=gpadx1, pady0=gpady0, pady1=gpady1)
+            dx.set_shape([majorDim, inH, inW, minorDim])
+            return dx, func
+        return y, grad
+    return func(x)
+
+#----------------------------------------------------------------------------
+
+def filter_2d(x, k, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Filter a batch of 2D images with the given FIR filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and filters each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    k = _setup_kernel(k) * gain
+    p = k.shape[0] - 1
+    return _simple_upfirdn_2d(x, k, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_2d(x, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Upsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and upsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the upsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * (gain * (factor ** 2))
+    p = k.shape[0] - factor
+    return _simple_upfirdn_2d(x, k, up=factor, pad0=(p+1)//2+factor-1, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def downsample_2d(x, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Downsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and downsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the downsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * gain
+    p = k.shape[0] - factor
+    return _simple_upfirdn_2d(x, k, down=factor, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_conv_2d(x, w, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Fused `upsample_2d()` followed by `tf.nn.conv2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+
+    # Check weight shape.
+    w = tf.convert_to_tensor(w)
+    assert w.shape.rank == 4
+    convH = w.shape[0].value
+    convW = w.shape[1].value
+    inC = _shape(w, 2)
+    outC = _shape(w, 3)
+    assert convW == convH
+
+    # Setup filter kernel.
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * (gain * (factor ** 2))
+    p = (k.shape[0] - factor) - (convW - 1)
+
+    # Determine data dimensions.
+    if data_format == 'NCHW':
+        stride = [1, 1, factor, factor]
+        output_shape = [_shape(x, 0), outC, (_shape(x, 2) - 1) * factor + convH, (_shape(x, 3) - 1) * factor + convW]
+        num_groups = _shape(x, 1) // inC
+    else:
+        stride = [1, factor, factor, 1]
+        output_shape = [_shape(x, 0), (_shape(x, 1) - 1) * factor + convH, (_shape(x, 2) - 1) * factor + convW, outC]
+        num_groups = _shape(x, 3) // inC
+
+    # Transpose weights.
+    w = tf.reshape(w, [convH, convW, inC, num_groups, -1])
+    w = tf.transpose(w[::-1, ::-1], [0, 1, 4, 3, 2])
+    w = tf.reshape(w, [convH, convW, -1, num_groups * inC])
+
+    # Execute.
+    x = tf.nn.conv2d_transpose(x, w, output_shape=output_shape, strides=stride, padding='VALID', data_format=data_format)
+    return _simple_upfirdn_2d(x, k, pad0=(p+1)//2+factor-1, pad1=p//2+1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def conv_downsample_2d(x, w, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Fused `tf.nn.conv2d()` followed by `downsample_2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    w = tf.convert_to_tensor(w)
+    convH, convW, _inC, _outC = w.shape.as_list()
+    assert convW == convH
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * gain
+    p = (k.shape[0] - factor) + (convW - 1)
+    if data_format == 'NCHW':
+        s = [1, 1, factor, factor]
+    else:
+        s = [1, factor, factor, 1]
+    x = _simple_upfirdn_2d(x, k, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+    return tf.nn.conv2d(x, w, strides=s, padding='VALID', data_format=data_format)
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _shape(tf_expr, dim_idx):
+    if tf_expr.shape.rank is not None:
+        dim = tf_expr.shape[dim_idx].value
+        if dim is not None:
+            return dim
+    return tf.shape(tf_expr)[dim_idx]
+
+def _setup_kernel(k):
+    k = np.asarray(k, dtype=np.float32)
+    if k.ndim == 1:
+        k = np.outer(k, k)
+    k /= np.sum(k)
+    assert k.ndim == 2
+    assert k.shape[0] == k.shape[1]
+    return k
+
+def _simple_upfirdn_2d(x, k, up=1, down=1, pad0=0, pad1=0, data_format='NCHW', impl='cuda'):
+    assert data_format in ['NCHW', 'NHWC']
+    assert x.shape.rank == 4
+    y = x
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(y, 2), _shape(y, 3), 1])
+    y = upfirdn_2d(y, k, upx=up, upy=up, downx=down, downy=down, padx0=pad0, padx1=pad1, pady0=pad0, pady1=pad1, impl=impl)
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(x, 1), _shape(y, 1), _shape(y, 2)])
+    return y
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/optimizer.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/optimizer.py
new file mode 100644
index 000000000..889e9011d
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/optimizer.py
@@ -0,0 +1,336 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper wrapper for a Tensorflow optimizer."""
+
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import List, Union
+
+from . import autosummary
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+try:
+    # TensorFlow 1.13
+    from tensorflow.python.ops import nccl_ops
+except:
+    # Older TensorFlow versions
+    import tensorflow.contrib.nccl as nccl_ops
+
+class Optimizer:
+    """A Wrapper for tf.train.Optimizer.
+
+    Automatically takes care of:
+    - Gradient averaging for multi-GPU training.
+    - Gradient accumulation for arbitrarily large minibatches.
+    - Dynamic loss scaling and typecasts for FP16 training.
+    - Ignoring corrupted gradients that contain NaNs/Infs.
+    - Reporting statistics.
+    - Well-chosen default settings.
+    """
+
+    def __init__(self,
+        name:                   str             = "Train",                  # Name string that will appear in TensorFlow graph.
+        tf_optimizer:           str             = "tf.train.AdamOptimizer", # Underlying optimizer class.
+        learning_rate:          TfExpressionEx  = 0.001,                    # Learning rate. Can vary over time.
+        minibatch_multiplier:   TfExpressionEx  = None,                     # Treat N consecutive minibatches as one by accumulating gradients.
+        share:                  "Optimizer"     = None,                     # Share internal state with a previously created optimizer?
+        use_loss_scaling:       bool            = False,                    # Enable dynamic loss scaling for robust mixed-precision training?
+        loss_scaling_init:      float           = 64.0,                     # Log2 of initial loss scaling factor.
+        loss_scaling_inc:       float           = 0.0005,                   # Log2 of per-minibatch loss scaling increment when there is no overflow.
+        loss_scaling_dec:       float           = 1.0,                      # Log2 of per-minibatch loss scaling decrement when there is an overflow.
+        report_mem_usage:       bool            = False,                    # Report fine-grained memory usage statistics in TensorBoard?
+        **kwargs):
+
+        # Public fields.
+        self.name                   = name
+        self.learning_rate          = learning_rate
+        self.minibatch_multiplier   = minibatch_multiplier
+        self.id                     = self.name.replace("/", ".")
+        self.scope                  = tf.get_default_graph().unique_name(self.id)
+        self.optimizer_class        = util.get_obj_by_name(tf_optimizer)
+        self.optimizer_kwargs       = dict(kwargs)
+        self.use_loss_scaling       = use_loss_scaling
+        self.loss_scaling_init      = loss_scaling_init
+        self.loss_scaling_inc       = loss_scaling_inc
+        self.loss_scaling_dec       = loss_scaling_dec
+
+        # Private fields.
+        self._updates_applied       = False
+        self._devices               = OrderedDict() # device_name => EasyDict()
+        self._shared_optimizers     = OrderedDict() # device_name => optimizer_class
+        self._gradient_shapes       = None          # [shape, ...]
+        self._report_mem_usage      = report_mem_usage
+
+        # Validate arguments.
+        assert callable(self.optimizer_class)
+
+        # Share internal state if requested.
+        if share is not None:
+            assert isinstance(share, Optimizer)
+            assert self.optimizer_class is share.optimizer_class
+            assert self.learning_rate is share.learning_rate
+            assert self.optimizer_kwargs == share.optimizer_kwargs
+            self._shared_optimizers = share._shared_optimizers # pylint: disable=protected-access
+
+    def _get_device(self, device_name: str):
+        """Get internal state for the given TensorFlow device."""
+        tfutil.assert_tf_initialized()
+        if device_name in self._devices:
+            return self._devices[device_name]
+
+        # Initialize fields.
+        device = util.EasyDict()
+        device.name             = device_name
+        device.optimizer        = None          # Underlying optimizer:     optimizer_class
+        device.loss_scaling_var = None          # Log2 of loss scaling:     tf.Variable
+        device.grad_raw         = OrderedDict() # Raw gradients:            var => [grad, ...]
+        device.grad_clean       = OrderedDict() # Clean gradients:          var => grad
+        device.grad_acc_vars    = OrderedDict() # Accumulation sums:        var => tf.Variable
+        device.grad_acc_count   = None          # Accumulation counter:     tf.Variable
+        device.grad_acc         = OrderedDict() # Accumulated gradients:    var => grad
+
+        # Setup TensorFlow objects.
+        with tfutil.absolute_name_scope(self.scope + "/Devices"), tf.device(device_name), tf.control_dependencies(None):
+            if device_name not in self._shared_optimizers:
+                optimizer_name = self.scope.replace("/", "_") + "_opt%d" % len(self._shared_optimizers)
+                self._shared_optimizers[device_name] = self.optimizer_class(name=optimizer_name, learning_rate=self.learning_rate, **self.optimizer_kwargs)
+            device.optimizer = self._shared_optimizers[device_name]
+            if self.use_loss_scaling:
+                device.loss_scaling_var = tf.Variable(np.float32(self.loss_scaling_init), trainable=False, name="loss_scaling_var")
+
+        # Register device.
+        self._devices[device_name] = device
+        return device
+
+    def register_gradients(self, loss: TfExpression, trainable_vars: Union[List, dict]) -> None:
+        """Register the gradients of the given loss function with respect to the given variables.
+        Intended to be called once per GPU."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        device = self._get_device(loss.device)
+
+        # Validate trainables.
+        if isinstance(trainable_vars, dict):
+            trainable_vars = list(trainable_vars.values())  # allow passing in Network.trainables as vars
+        assert isinstance(trainable_vars, list) and len(trainable_vars) >= 1
+        assert all(tfutil.is_tf_expression(expr) for expr in trainable_vars + [loss])
+        assert all(var.device == device.name for var in trainable_vars)
+
+        # Validate shapes.
+        if self._gradient_shapes is None:
+            self._gradient_shapes = [var.shape.as_list() for var in trainable_vars]
+        assert len(trainable_vars) == len(self._gradient_shapes)
+        assert all(var.shape.as_list() == var_shape for var, var_shape in zip(trainable_vars, self._gradient_shapes))
+
+        # Report memory usage if requested.
+        deps = []
+        if self._report_mem_usage:
+            self._report_mem_usage = False
+            try:
+                with tf.name_scope(self.id + '_mem'), tf.device(device.name), tf.control_dependencies([loss]):
+                    deps.append(autosummary.autosummary(self.id + "/mem_usage_gb", tf.contrib.memory_stats.BytesInUse() / 2 ** 30))
+            except tf.errors.NotFoundError:
+                pass
+
+        # Compute gradients.
+        with tf.name_scope(self.id + "_grad"), tf.device(device.name), tf.control_dependencies(deps):
+            loss = self.apply_loss_scaling(tf.cast(loss, tf.float32))
+            gate = tf.train.Optimizer.GATE_NONE  # disable gating to reduce memory usage
+            grad_list = device.optimizer.compute_gradients(loss=loss, var_list=trainable_vars, gate_gradients=gate)
+
+        # Register gradients.
+        for grad, var in grad_list:
+            if var not in device.grad_raw:
+                device.grad_raw[var] = []
+            device.grad_raw[var].append(grad)
+
+    def apply_updates(self, allow_no_op: bool = False) -> tf.Operation:
+        """Construct training op to update the registered variables based on their gradients."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        self._updates_applied = True
+        all_ops = []
+
+        # Check for no-op.
+        if allow_no_op and len(self._devices) == 0:
+            with tfutil.absolute_name_scope(self.scope):
+                return tf.no_op(name='TrainingOp')
+
+        # Clean up gradients.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Clean%d" % device_idx), tf.device(device.name):
+                for var, grad in device.grad_raw.items():
+
+                    # Filter out disconnected gradients and convert to float32.
+                    grad = [g for g in grad if g is not None]
+                    grad = [tf.cast(g, tf.float32) for g in grad]
+
+                    # Sum within the device.
+                    if len(grad) == 0:
+                        grad = tf.zeros(var.shape)  # No gradients => zero.
+                    elif len(grad) == 1:
+                        grad = grad[0]              # Single gradient => use as is.
+                    else:
+                        grad = tf.add_n(grad)       # Multiple gradients => sum.
+
+                    # Scale as needed.
+                    scale = 1.0 / len(device.grad_raw[var]) / len(self._devices)
+                    scale = tf.constant(scale, dtype=tf.float32, name="scale")
+                    if self.minibatch_multiplier is not None:
+                        scale /= tf.cast(self.minibatch_multiplier, tf.float32)
+                    scale = self.undo_loss_scaling(scale)
+                    device.grad_clean[var] = grad * scale
+
+        # Sum gradients across devices.
+        if len(self._devices) > 1:
+            with tfutil.absolute_name_scope(self.scope + "/Broadcast"), tf.device(None):
+                for all_vars in zip(*[device.grad_clean.keys() for device in self._devices.values()]):
+                    if len(all_vars) > 0 and all(dim > 0 for dim in all_vars[0].shape.as_list()): # NCCL does not support zero-sized tensors.
+                        all_grads = [device.grad_clean[var] for device, var in zip(self._devices.values(), all_vars)]
+                        all_grads = nccl_ops.all_sum(all_grads)
+                        for device, var, grad in zip(self._devices.values(), all_vars, all_grads):
+                            device.grad_clean[var] = grad
+
+        # Apply updates separately on each device.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Apply%d" % device_idx), tf.device(device.name):
+                # pylint: disable=cell-var-from-loop
+
+                # Accumulate gradients over time.
+                if self.minibatch_multiplier is None:
+                    acc_ok = tf.constant(True, name='acc_ok')
+                    device.grad_acc = OrderedDict(device.grad_clean)
+                else:
+                    # Create variables.
+                    with tf.control_dependencies(None):
+                        for var in device.grad_clean.keys():
+                            device.grad_acc_vars[var] = tf.Variable(tf.zeros(var.shape), trainable=False, name="grad_acc_var")
+                        device.grad_acc_count = tf.Variable(tf.zeros([]), trainable=False, name="grad_acc_count")
+
+                    # Track counter.
+                    count_cur = device.grad_acc_count + 1.0
+                    count_inc_op = lambda: tf.assign(device.grad_acc_count, count_cur)
+                    count_reset_op = lambda: tf.assign(device.grad_acc_count, tf.zeros([]))
+                    acc_ok = (count_cur >= tf.cast(self.minibatch_multiplier, tf.float32))
+                    all_ops.append(tf.cond(acc_ok, count_reset_op, count_inc_op))
+
+                    # Track gradients.
+                    for var, grad in device.grad_clean.items():
+                        acc_var = device.grad_acc_vars[var]
+                        acc_cur = acc_var + grad
+                        device.grad_acc[var] = acc_cur
+                        with tf.control_dependencies([acc_cur]):
+                            acc_inc_op = lambda: tf.assign(acc_var, acc_cur)
+                            acc_reset_op = lambda: tf.assign(acc_var, tf.zeros(var.shape))
+                            all_ops.append(tf.cond(acc_ok, acc_reset_op, acc_inc_op))
+
+                # No overflow => apply gradients.
+                all_ok = tf.reduce_all(tf.stack([acc_ok] + [tf.reduce_all(tf.is_finite(g)) for g in device.grad_acc.values()]))
+                apply_op = lambda: device.optimizer.apply_gradients([(tf.cast(grad, var.dtype), var) for var, grad in device.grad_acc.items()])
+                all_ops.append(tf.cond(all_ok, apply_op, tf.no_op))
+
+                # Adjust loss scaling.
+                if self.use_loss_scaling:
+                    ls_inc_op = lambda: tf.assign_add(device.loss_scaling_var, self.loss_scaling_inc)
+                    ls_dec_op = lambda: tf.assign_sub(device.loss_scaling_var, self.loss_scaling_dec)
+                    ls_update_op = lambda: tf.group(tf.cond(all_ok, ls_inc_op, ls_dec_op))
+                    all_ops.append(tf.cond(acc_ok, ls_update_op, tf.no_op))
+
+                # Last device => report statistics.
+                if device_idx == len(self._devices) - 1:
+                    all_ops.append(autosummary.autosummary(self.id + "/learning_rate", self.learning_rate))
+                    all_ops.append(autosummary.autosummary(self.id + "/overflow_frequency", tf.where(all_ok, 0, 1), condition=acc_ok))
+                    if self.use_loss_scaling:
+                        all_ops.append(autosummary.autosummary(self.id + "/loss_scaling_log2", device.loss_scaling_var))
+
+        # Initialize variables.
+        self.reset_optimizer_state()
+        if self.use_loss_scaling:
+            tfutil.init_uninitialized_vars([device.loss_scaling_var for device in self._devices.values()])
+        if self.minibatch_multiplier is not None:
+            tfutil.run([var.initializer for device in self._devices.values() for var in list(device.grad_acc_vars.values()) + [device.grad_acc_count]])
+
+        # Group everything into a single op.
+        with tfutil.absolute_name_scope(self.scope):
+            return tf.group(*all_ops, name="TrainingOp")
+
+    def reset_optimizer_state(self) -> None:
+        """Reset internal state of the underlying optimizer."""
+        tfutil.assert_tf_initialized()
+        tfutil.run([var.initializer for device in self._devices.values() for var in device.optimizer.variables()])
+
+    def get_loss_scaling_var(self, device: str) -> Union[tf.Variable, None]:
+        """Get or create variable representing log2 of the current dynamic loss scaling factor."""
+        return self._get_device(device).loss_scaling_var
+
+    def apply_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Apply dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(self.get_loss_scaling_var(value.device))
+
+    def undo_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Undo the effect of dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(-self.get_loss_scaling_var(value.device)) # pylint: disable=invalid-unary-operand-type
+
+
+class SimpleAdam:
+    """Simplified version of tf.train.AdamOptimizer that behaves identically when used with dnnlib.tflib.Optimizer."""
+
+    def __init__(self, name="Adam", learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8):
+        self.name = name
+        self.learning_rate = learning_rate
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+        self.all_state_vars = []
+
+    def variables(self):
+        return self.all_state_vars
+
+    def compute_gradients(self, loss, var_list, gate_gradients=tf.train.Optimizer.GATE_NONE):
+        assert gate_gradients == tf.train.Optimizer.GATE_NONE
+        return list(zip(tf.gradients(loss, var_list), var_list))
+
+    def apply_gradients(self, grads_and_vars):
+        with tf.name_scope(self.name):
+            state_vars = []
+            update_ops = []
+
+            # Adjust learning rate to deal with startup bias.
+            with tf.control_dependencies(None):
+                b1pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                b2pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                state_vars += [b1pow_var, b2pow_var]
+            b1pow_new = b1pow_var * self.beta1
+            b2pow_new = b2pow_var * self.beta2
+            update_ops += [tf.assign(b1pow_var, b1pow_new), tf.assign(b2pow_var, b2pow_new)]
+            lr_new = self.learning_rate * tf.sqrt(1 - b2pow_new) / (1 - b1pow_new)
+
+            # Construct ops to update each variable.
+            for grad, var in grads_and_vars:
+                with tf.control_dependencies(None):
+                    m_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    v_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    state_vars += [m_var, v_var]
+                m_new = self.beta1 * m_var + (1 - self.beta1) * grad
+                v_new = self.beta2 * v_var + (1 - self.beta2) * tf.square(grad)
+                var_delta = lr_new * m_new / (tf.sqrt(v_new) + self.epsilon)
+                update_ops += [tf.assign(m_var, m_new), tf.assign(v_var, v_new), tf.assign_sub(var, var_delta)]
+
+            # Group everything together.
+            self.all_state_vars += state_vars
+            return tf.group(*update_ops)
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/tflib/tfutil.py b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/tfutil.py
new file mode 100644
index 000000000..ee9728217
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/tflib/tfutil.py
@@ -0,0 +1,254 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Miscellaneous helper utils for Tensorflow."""
+
+import os
+import numpy as np
+import tensorflow as tf
+
+# Silence deprecation warnings from TensorFlow 1.13 onwards
+import logging
+logging.getLogger('tensorflow').setLevel(logging.ERROR)
+import tensorflow.contrib   # requires TensorFlow 1.x!
+tf.contrib = tensorflow.contrib
+
+from typing import Any, Iterable, List, Union
+
+TfExpression = Union[tf.Tensor, tf.Variable, tf.Operation]
+"""A type that represents a valid Tensorflow expression."""
+
+TfExpressionEx = Union[TfExpression, int, float, np.ndarray]
+"""A type that can be converted to a valid Tensorflow expression."""
+
+
+def run(*args, **kwargs) -> Any:
+    """Run the specified ops in the default session."""
+    assert_tf_initialized()
+    return tf.get_default_session().run(*args, **kwargs)
+
+
+def is_tf_expression(x: Any) -> bool:
+    """Check whether the input is a valid Tensorflow expression, i.e., Tensorflow Tensor, Variable, or Operation."""
+    return isinstance(x, (tf.Tensor, tf.Variable, tf.Operation))
+
+
+def shape_to_list(shape: Iterable[tf.Dimension]) -> List[Union[int, None]]:
+    """Convert a Tensorflow shape to a list of ints. Retained for backwards compatibility -- use TensorShape.as_list() in new code."""
+    return [dim.value for dim in shape]
+
+
+def flatten(x: TfExpressionEx) -> TfExpression:
+    """Shortcut function for flattening a tensor."""
+    with tf.name_scope("Flatten"):
+        return tf.reshape(x, [-1])
+
+
+def log2(x: TfExpressionEx) -> TfExpression:
+    """Logarithm in base 2."""
+    with tf.name_scope("Log2"):
+        return tf.log(x) * np.float32(1.0 / np.log(2.0))
+
+
+def exp2(x: TfExpressionEx) -> TfExpression:
+    """Exponent in base 2."""
+    with tf.name_scope("Exp2"):
+        return tf.exp(x * np.float32(np.log(2.0)))
+
+
+def lerp(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpressionEx:
+    """Linear interpolation."""
+    with tf.name_scope("Lerp"):
+        return a + (b - a) * t
+
+
+def lerp_clip(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpression:
+    """Linear interpolation with clip."""
+    with tf.name_scope("LerpClip"):
+        return a + (b - a) * tf.clip_by_value(t, 0.0, 1.0)
+
+
+def absolute_name_scope(scope: str) -> tf.name_scope:
+    """Forcefully enter the specified name scope, ignoring any surrounding scopes."""
+    return tf.name_scope(scope + "/")
+
+
+def absolute_variable_scope(scope: str, **kwargs) -> tf.variable_scope:
+    """Forcefully enter the specified variable scope, ignoring any surrounding scopes."""
+    return tf.variable_scope(tf.VariableScope(name=scope, **kwargs), auxiliary_name_scope=False)
+
+
+def _sanitize_tf_config(config_dict: dict = None) -> dict:
+    # Defaults.
+    cfg = dict()
+    cfg["rnd.np_random_seed"]               = None      # Random seed for NumPy. None = keep as is.
+    cfg["rnd.tf_random_seed"]               = "auto"    # Random seed for TensorFlow. 'auto' = derive from NumPy random state. None = keep as is.
+    cfg["env.TF_CPP_MIN_LOG_LEVEL"]         = "1"       # 0 = Print all available debug info from TensorFlow. 1 = Print warnings and errors, but disable debug info.
+    cfg["graph_options.place_pruned_graph"] = True      # False = Check that all ops are available on the designated device. True = Skip the check for ops that are not used.
+    cfg["gpu_options.allow_growth"]         = True      # False = Allocate all GPU memory at the beginning. True = Allocate only as much GPU memory as needed.
+
+    # Remove defaults for environment variables that are already set.
+    for key in list(cfg):
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            if fields[1] in os.environ:
+                del cfg[key]
+
+    # User overrides.
+    if config_dict is not None:
+        cfg.update(config_dict)
+    return cfg
+
+
+def init_tf(config_dict: dict = None) -> None:
+    """Initialize TensorFlow session using good default settings."""
+    # Skip if already initialized.
+    if tf.get_default_session() is not None:
+        return
+
+    # Setup config dict and random seeds.
+    cfg = _sanitize_tf_config(config_dict)
+    np_random_seed = cfg["rnd.np_random_seed"]
+    if np_random_seed is not None:
+        np.random.seed(np_random_seed)
+    tf_random_seed = cfg["rnd.tf_random_seed"]
+    if tf_random_seed == "auto":
+        tf_random_seed = np.random.randint(1 << 31)
+    if tf_random_seed is not None:
+        tf.set_random_seed(tf_random_seed)
+
+    # Setup environment variables.
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            os.environ[fields[1]] = str(value)
+
+    # Create default TensorFlow session.
+    create_session(cfg, force_as_default=True)
+
+
+def assert_tf_initialized():
+    """Check that TensorFlow session has been initialized."""
+    if tf.get_default_session() is None:
+        raise RuntimeError("No default TensorFlow session found. Please call dnnlib.tflib.init_tf().")
+
+
+def create_session(config_dict: dict = None, force_as_default: bool = False) -> tf.Session:
+    """Create tf.Session based on config dict."""
+    # Setup TensorFlow config proto.
+    cfg = _sanitize_tf_config(config_dict)
+    config_proto = tf.ConfigProto()
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] not in ["rnd", "env"]:
+            obj = config_proto
+            for field in fields[:-1]:
+                obj = getattr(obj, field)
+            setattr(obj, fields[-1], value)
+
+    # Create session.
+    session = tf.Session(config=config_proto)
+    if force_as_default:
+        # pylint: disable=protected-access
+        session._default_session = session.as_default()
+        session._default_session.enforce_nesting = False
+        session._default_session.__enter__()
+    return session
+
+
+def init_uninitialized_vars(target_vars: List[tf.Variable] = None) -> None:
+    """Initialize all tf.Variables that have not already been initialized.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tf.variables_initializer(tf.report_uninitialized_variables()).run()
+    """
+    assert_tf_initialized()
+    if target_vars is None:
+        target_vars = tf.global_variables()
+
+    test_vars = []
+    test_ops = []
+
+    with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+        for var in target_vars:
+            assert is_tf_expression(var)
+
+            try:
+                tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/IsVariableInitialized:0"))
+            except KeyError:
+                # Op does not exist => variable may be uninitialized.
+                test_vars.append(var)
+
+                with absolute_name_scope(var.name.split(":")[0]):
+                    test_ops.append(tf.is_variable_initialized(var))
+
+    init_vars = [var for var, inited in zip(test_vars, run(test_ops)) if not inited]
+    run([var.initializer for var in init_vars])
+
+
+def set_vars(var_to_value_dict: dict) -> None:
+    """Set the values of given tf.Variables.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tflib.run([tf.assign(var, value) for var, value in var_to_value_dict.items()]
+    """
+    assert_tf_initialized()
+    ops = []
+    feed_dict = {}
+
+    for var, value in var_to_value_dict.items():
+        assert is_tf_expression(var)
+
+        try:
+            setter = tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/setter:0"))  # look for existing op
+        except KeyError:
+            with absolute_name_scope(var.name.split(":")[0]):
+                with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+                    setter = tf.assign(var, tf.placeholder(var.dtype, var.shape, "new_value"), name="setter")  # create new setter
+
+        ops.append(setter)
+        feed_dict[setter.op.inputs[1]] = value
+
+    run(ops, feed_dict)
+
+
+def create_var_with_large_initial_value(initial_value: np.ndarray, *args, **kwargs):
+    """Create tf.Variable with large initial value without bloating the tf graph."""
+    assert_tf_initialized()
+    assert isinstance(initial_value, np.ndarray)
+    zeros = tf.zeros(initial_value.shape, initial_value.dtype)
+    var = tf.Variable(zeros, *args, **kwargs)
+    set_vars({var: initial_value})
+    return var
+
+
+def convert_images_from_uint8(images, drange=[-1,1], nhwc_to_nchw=False):
+    """Convert a minibatch of images from uint8 to float32 with configurable dynamic range.
+    Can be used as an input transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if nhwc_to_nchw:
+        images = tf.transpose(images, [0, 3, 1, 2])
+    return images * ((drange[1] - drange[0]) / 255) + drange[0]
+
+
+def convert_images_to_uint8(images, drange=[-1,1], nchw_to_nhwc=False, shrink=1, uint8_cast=True):
+    """Convert a minibatch of images from float32 to uint8 with configurable dynamic range.
+    Can be used as an output transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if shrink > 1:
+        ksize = [1, 1, shrink, shrink]
+        images = tf.nn.avg_pool(images, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW")
+    if nchw_to_nhwc:
+        images = tf.transpose(images, [0, 2, 3, 1])
+    scale = 255 / (drange[1] - drange[0])
+    images = images * scale + (0.5 - drange[0] * scale)
+    if uint8_cast:
+        images = tf.saturate_cast(images, tf.uint8)
+    return images
diff --git a/reconstruction/ostec/external/stylegan2/dnnlib/util.py b/reconstruction/ostec/external/stylegan2/dnnlib/util.py
new file mode 100644
index 000000000..73c98d73c
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/dnnlib/util.py
@@ -0,0 +1,410 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Miscellaneous utility classes and functions."""
+
+import ctypes
+import fnmatch
+import importlib
+import inspect
+import numpy as np
+import os
+import shutil
+import sys
+import types
+import io
+import pickle
+import re
+import requests
+import html
+import hashlib
+import glob
+import uuid
+
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        self[name] = value
+
+    def __delattr__(self, name: str) -> None:
+        del self[name]
+
+
+class Logger(object):
+    """Redirect stderr to stdout, optionally print stdout to a file, and optionally force flushing on both stdout and the file."""
+
+    def __init__(self, file_name: str = None, file_mode: str = "w", should_flush: bool = True):
+        self.file = None
+
+        if file_name is not None:
+            self.file = open(file_name, file_mode)
+
+        self.should_flush = should_flush
+        self.stdout = sys.stdout
+        self.stderr = sys.stderr
+
+        sys.stdout = self
+        sys.stderr = self
+
+    def __enter__(self) -> "Logger":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def write(self, text: str) -> None:
+        """Write text to stdout (and a file) and optionally flush."""
+        if len(text) == 0: # workaround for a bug in VSCode debugger: sys.stdout.write(''); sys.stdout.flush() => crash
+            return
+
+        if self.file is not None:
+            self.file.write(text)
+
+        self.stdout.write(text)
+
+        if self.should_flush:
+            self.flush()
+
+    def flush(self) -> None:
+        """Flush written text to both stdout and a file, if open."""
+        if self.file is not None:
+            self.file.flush()
+
+        self.stdout.flush()
+
+    def close(self) -> None:
+        """Flush, close possible files, and remove stdout/stderr mirroring."""
+        self.flush()
+
+        # if using multiple loggers, prevent closing in wrong order
+        if sys.stdout is self:
+            sys.stdout = self.stdout
+        if sys.stderr is self:
+            sys.stderr = self.stderr
+
+        if self.file is not None:
+            self.file.close()
+
+
+# Small util functions
+# ------------------------------------------------------------------------------------------
+
+
+def format_time(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m {2:02}s".format(s // (60 * 60), (s // 60) % 60, s % 60)
+    else:
+        return "{0}d {1:02}h {2:02}m".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24, (s // 60) % 60)
+
+
+def ask_yes_no(question: str) -> bool:
+    """Ask the user the question until the user inputs a valid answer."""
+    while True:
+        try:
+            print("{0} [y/n]".format(question))
+            return strtobool(input().lower())
+        except ValueError:
+            pass
+
+
+def tuple_product(t: Tuple) -> Any:
+    """Calculate the product of the tuple elements."""
+    result = 1
+
+    for v in t:
+        result *= v
+
+    return result
+
+
+_str_to_ctype = {
+    "uint8": ctypes.c_ubyte,
+    "uint16": ctypes.c_uint16,
+    "uint32": ctypes.c_uint32,
+    "uint64": ctypes.c_uint64,
+    "int8": ctypes.c_byte,
+    "int16": ctypes.c_int16,
+    "int32": ctypes.c_int32,
+    "int64": ctypes.c_int64,
+    "float32": ctypes.c_float,
+    "float64": ctypes.c_double
+}
+
+
+def get_dtype_and_ctype(type_obj: Any) -> Tuple[np.dtype, Any]:
+    """Given a type name string (or an object having a __name__ attribute), return matching Numpy and ctypes types that have the same size in bytes."""
+    type_str = None
+
+    if isinstance(type_obj, str):
+        type_str = type_obj
+    elif hasattr(type_obj, "__name__"):
+        type_str = type_obj.__name__
+    elif hasattr(type_obj, "name"):
+        type_str = type_obj.name
+    else:
+        raise RuntimeError("Cannot infer type name from input")
+
+    assert type_str in _str_to_ctype.keys()
+
+    my_dtype = np.dtype(type_str)
+    my_ctype = _str_to_ctype[type_str]
+
+    assert my_dtype.itemsize == ctypes.sizeof(my_ctype)
+
+    return my_dtype, my_ctype
+
+
+def is_pickleable(obj: Any) -> bool:
+    try:
+        with io.BytesIO() as stream:
+            pickle.dump(obj, stream)
+        return True
+    except:
+        return False
+
+
+# Functionality to import modules/objects by name, and call functions by name
+# ------------------------------------------------------------------------------------------
+
+def get_module_from_obj_name(obj_name: str) -> Tuple[types.ModuleType, str]:
+    """Searches for the underlying module behind the name to some python object.
+    Returns the module and the object name (original name with module part removed)."""
+
+    # allow convenience shorthands, substitute them by full names
+    obj_name = re.sub("^np.", "numpy.", obj_name)
+    obj_name = re.sub("^tf.", "tensorflow.", obj_name)
+
+    # list alternatives for (module_name, local_obj_name)
+    parts = obj_name.split(".")
+    name_pairs = [(".".join(parts[:i]), ".".join(parts[i:])) for i in range(len(parts), 0, -1)]
+
+    # try each alternative in turn
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+            return module, local_obj_name
+        except:
+            pass
+
+    # maybe some of the modules themselves contain errors?
+    for module_name, _local_obj_name in name_pairs:
+        try:
+            importlib.import_module(module_name) # may raise ImportError
+        except ImportError:
+            if not str(sys.exc_info()[1]).startswith("No module named '" + module_name + "'"):
+                raise
+
+    # maybe the requested attribute is missing?
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+        except ImportError:
+            pass
+
+    # we are out of luck, but we have no idea why
+    raise ImportError(obj_name)
+
+
+def get_obj_from_module(module: types.ModuleType, obj_name: str) -> Any:
+    """Traverses the object name and returns the last (rightmost) python object."""
+    if obj_name == '':
+        return module
+    obj = module
+    for part in obj_name.split("."):
+        obj = getattr(obj, part)
+    return obj
+
+
+def get_obj_by_name(name: str) -> Any:
+    """Finds the python object with the given name."""
+    module, obj_name = get_module_from_obj_name(name)
+    return get_obj_from_module(module, obj_name)
+
+
+def call_func_by_name(*args, func_name: str = None, **kwargs) -> Any:
+    """Finds the python object with the given name and calls it as a function."""
+    assert func_name is not None
+    func_obj = get_obj_by_name(func_name)
+    assert callable(func_obj)
+    return func_obj(*args, **kwargs)
+
+
+def get_module_dir_by_obj_name(obj_name: str) -> str:
+    """Get the directory path of the module containing the given object name."""
+    module, _ = get_module_from_obj_name(obj_name)
+    return os.path.dirname(inspect.getfile(module))
+
+
+def is_top_level_function(obj: Any) -> bool:
+    """Determine whether the given object is a top-level function, i.e., defined at module scope using 'def'."""
+    return callable(obj) and obj.__name__ in sys.modules[obj.__module__].__dict__
+
+
+def get_top_level_function_name(obj: Any) -> str:
+    """Return the fully-qualified name of a top-level function."""
+    assert is_top_level_function(obj)
+    return obj.__module__ + "." + obj.__name__
+
+
+# File system helpers
+# ------------------------------------------------------------------------------------------
+
+def list_dir_recursively_with_ignore(dir_path: str, ignores: List[str] = None, add_base_to_relative: bool = False) -> List[Tuple[str, str]]:
+    """List all files recursively in a given directory while ignoring given file and directory names.
+    Returns list of tuples containing both absolute and relative paths."""
+    assert os.path.isdir(dir_path)
+    base_name = os.path.basename(os.path.normpath(dir_path))
+
+    if ignores is None:
+        ignores = []
+
+    result = []
+
+    for root, dirs, files in os.walk(dir_path, topdown=True):
+        for ignore_ in ignores:
+            dirs_to_remove = [d for d in dirs if fnmatch.fnmatch(d, ignore_)]
+
+            # dirs need to be edited in-place
+            for d in dirs_to_remove:
+                dirs.remove(d)
+
+            files = [f for f in files if not fnmatch.fnmatch(f, ignore_)]
+
+        absolute_paths = [os.path.join(root, f) for f in files]
+        relative_paths = [os.path.relpath(p, dir_path) for p in absolute_paths]
+
+        if add_base_to_relative:
+            relative_paths = [os.path.join(base_name, p) for p in relative_paths]
+
+        assert len(absolute_paths) == len(relative_paths)
+        result += zip(absolute_paths, relative_paths)
+
+    return result
+
+
+def copy_files_and_create_dirs(files: List[Tuple[str, str]]) -> None:
+    """Takes in a list of tuples of (src, dst) paths and copies files.
+    Will create all necessary directories."""
+    for file in files:
+        target_dir_name = os.path.dirname(file[1])
+
+        # will create all intermediate-level directories
+        if not os.path.exists(target_dir_name):
+            os.makedirs(target_dir_name)
+
+        shutil.copyfile(file[0], file[1])
+
+
+# URL helpers
+# ------------------------------------------------------------------------------------------
+
+def is_url(obj: Any, allow_file_urls: bool = False) -> bool:
+    """Determine whether the given object is a valid URL string."""
+    if not isinstance(obj, str) or not "://" in obj:
+        return False
+    if allow_file_urls and obj.startswith('file:///'):
+        return True
+    try:
+        res = requests.compat.urlparse(obj)
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+        res = requests.compat.urlparse(requests.compat.urljoin(obj, "/"))
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+    except:
+        return False
+    return True
+
+
+def open_url(url: str, cache_dir: str = None, num_attempts: int = 10, verbose: bool = True) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert is_url(url, allow_file_urls=True)
+    assert num_attempts >= 1
+
+    # Handle file URLs.
+    if url.startswith('file:///'):
+        return open(url[len('file:///'):], "rb")
+
+    # Lookup from cache.
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+    if cache_dir is not None:
+        cache_files = glob.glob(os.path.join(cache_dir, url_md5 + "_*"))
+        if len(cache_files) == 1:
+            return open(cache_files[0], "rb")
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Save to cache.
+    if cache_dir is not None:
+        safe_name = re.sub(r"[^0-9a-zA-Z-._]", "_", url_name)
+        cache_file = os.path.join(cache_dir, url_md5 + "_" + safe_name)
+        temp_file = os.path.join(cache_dir, "tmp_" + uuid.uuid4().hex + "_" + url_md5 + "_" + safe_name)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(temp_file, "wb") as f:
+            f.write(url_data)
+        os.replace(temp_file, cache_file) # atomic
+
+    # Return data as file object.
+    return io.BytesIO(url_data)
diff --git a/reconstruction/ostec/external/stylegan2/docs/license.html b/reconstruction/ostec/external/stylegan2/docs/license.html
new file mode 100644
index 000000000..583f7f890
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/docs/license.html
@@ -0,0 +1,160 @@
+<!DOCTYPE html>
+<html xmlns="http://www.w3.org/1999/xhtml" lang="" xml:lang="">
+<head>
+  <meta charset="utf-8"/>
+  <meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes"/>
+  <title>Nvidia Source Code License-NC</title>
+  <link href="https://fonts.googleapis.com/css?family=Helvetica+Neue" rel="stylesheet"/>
+  <style type="text/css">
+
+  body {
+      font-family: 'Helvetica Neue', sans-serif;
+      color: #000000;
+      line-height: 1.5;
+  }
+
+  h1, h2, h3, h4, h5, h6 {
+      color: #92D050;
+      font-weight: normal;
+  }
+
+  h1 {
+      line-height: 1.2;
+      font-size: 2em;
+      margin-top: 1.5em;
+  }
+
+  p {
+    margin-left: 0px;
+    margin-right: 0px;
+    margin-top: 0.75em;
+    margin-bottom: 0.75em;
+  }
+
+  p.tab {
+    margin-left: 3em;
+  }
+
+  hr {
+    border: 0px;
+    height: 1px;
+    background: #CCCCCC;
+  }
+
+  @media screen and (min-width: 680px) {
+      .max-width {
+          margin: 0 100px 0 170px;
+          max-width: 640px;
+      }
+  }
+  @media screen and (min-width: 980px) {
+      .max-width {
+          margin: 0 auto;
+      }
+  }
+  </style>
+</head>
+<body class="max-width">
+
+<h1>Nvidia Source Code License-NC</h1>
+
+<hr/>
+
+<h2>1. Definitions</h2>
+
+<p>&ldquo;Licensor&rdquo; means any person or entity that distributes its Work.</p>
+
+<p>&ldquo;Software&rdquo; means the original work of authorship made available under
+this License.</p>
+
+<p>&ldquo;Work&rdquo; means the Software and any additions to or derivative works of
+the Software that are made available under this License.</p>
+
+<p>&ldquo;Nvidia Processors&rdquo; means any central processing unit (CPU), graphics
+processing unit (GPU), field-programmable gate array (FPGA),
+application-specific integrated circuit (ASIC) or any combination
+thereof designed, made, sold, or provided by Nvidia or its affiliates.</p>
+
+<p>The terms &ldquo;reproduce,&rdquo; &ldquo;reproduction,&rdquo; &ldquo;derivative works,&rdquo; and
+&ldquo;distribution&rdquo; have the meaning as provided under U.S. copyright law;
+provided, however, that for the purposes of this License, derivative
+works shall not include works that remain separable from, or merely
+link (or bind by name) to the interfaces of, the Work.</p>
+
+<p>Works, including the Software, are &ldquo;made available&rdquo; under this License
+by including in or with the Work either (a) a copyright notice
+referencing the applicability of this License to the Work, or (b) a
+copy of this License.<p>
+
+<h2>2. License Grants</h2>
+
+<p class="tab">2.1 Copyright Grant. Subject to the terms and conditions of this
+License, each Licensor grants to you a perpetual, worldwide,
+non-exclusive, royalty-free, copyright license to reproduce,
+prepare derivative works of, publicly display, publicly perform,
+sublicense and distribute its Work and any resulting derivative
+works in any form.</p>
+
+<h2>3. Limitations</h2>
+
+<p class="tab">3.1 Redistribution. You may reproduce or distribute the Work only
+if (a) you do so under this License, (b) you include a complete
+copy of this License with your distribution, and (c) you retain
+without modification any copyright, patent, trademark, or
+attribution notices that are present in the Work.</p>
+
+<p class="tab">3.2 Derivative Works. You may specify that additional or different
+terms apply to the use, reproduction, and distribution of your
+derivative works of the Work (&ldquo;Your Terms&rdquo;) only if (a) Your Terms
+provide that the use limitation in Section 3.3 applies to your
+derivative works, and (b) you identify the specific derivative
+works that are subject to Your Terms. Notwithstanding Your Terms,
+this License (including the redistribution requirements in Section
+3.1) will continue to apply to the Work itself.</p>
+
+<p class="tab">3.3 Use Limitation. The Work and any derivative works thereof only
+may be used or intended for use non-commercially. The Work or
+derivative works thereof may be used or intended for use by Nvidia
+or its affiliates commercially or non-commercially. As used herein,
+&ldquo;non-commercially&rdquo; means for research or evaluation purposes only.</p>
+
+<p class="tab">3.4 Patent Claims. If you bring or threaten to bring a patent claim
+against any Licensor (including any claim, cross-claim or
+counterclaim in a lawsuit) to enforce any patents that you allege
+are infringed by any Work, then your rights under this License from
+such Licensor (including the grants in Sections 2.1 and 2.2) will
+terminate immediately.</p>
+
+<p class="tab">3.5 Trademarks. This License does not grant any rights to use any
+Licensor&rsquo;s or its affiliates&rsquo; names, logos, or trademarks, except
+as necessary to reproduce the notices described in this License.</p>
+
+<p class="tab">3.6 Termination. If you violate any term of this License, then your
+rights under this License (including the grants in Sections 2.1 and
+2.2) will terminate immediately.</p>
+
+<h2>4. Disclaimer of Warranty.</h2>
+
+<p>THE WORK IS PROVIDED &ldquo;AS IS&rdquo; WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR
+NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER
+THIS LICENSE.</p>
+
+<h2>5. Limitation of Liability.</h2>
+
+<p>EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
+THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
+SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
+INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
+OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
+(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
+LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
+COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGES.</p>
+
+<hr/>
+<br/>
+
+</body>
+</html>
diff --git a/reconstruction/ostec/external/stylegan2/docs/stylegan2-teaser-1024x256.png b/reconstruction/ostec/external/stylegan2/docs/stylegan2-teaser-1024x256.png
new file mode 100644
index 000000000..bb16c5f5c
Binary files /dev/null and b/reconstruction/ostec/external/stylegan2/docs/stylegan2-teaser-1024x256.png differ
diff --git a/reconstruction/ostec/external/stylegan2/docs/stylegan2-training-curves.png b/reconstruction/ostec/external/stylegan2/docs/stylegan2-training-curves.png
new file mode 100644
index 000000000..04fdf4d4f
Binary files /dev/null and b/reconstruction/ostec/external/stylegan2/docs/stylegan2-training-curves.png differ
diff --git a/reconstruction/ostec/external/stylegan2/docs/stylegan2encoder-teaser-1024x256.png b/reconstruction/ostec/external/stylegan2/docs/stylegan2encoder-teaser-1024x256.png
new file mode 100644
index 000000000..8e43b1a12
Binary files /dev/null and b/reconstruction/ostec/external/stylegan2/docs/stylegan2encoder-teaser-1024x256.png differ
diff --git a/reconstruction/ostec/external/stylegan2/docs/versions.html b/reconstruction/ostec/external/stylegan2/docs/versions.html
new file mode 100644
index 000000000..a78bef3d8
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/docs/versions.html
@@ -0,0 +1,64 @@
+<!DOCTYPE html>
+<html xmlns="http://www.w3.org/1999/xhtml" lang="" xml:lang="">
+<head>
+  <meta charset="utf-8"/>
+  <meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes"/>
+  <title>StyleGAN versions</title>
+  <link href="https://fonts.googleapis.com/css?family=Montserrat|Source+Sans+Pro" rel="stylesheet"/>
+  <style type="text/css">
+
+  body {
+      font-family: 'Source Sans Pro', sans-serif;
+      color: #000000;
+      line-height: 1.5;
+  }
+
+  h1, h2, h3, h4, h5, h6 {
+      font-family: 'Montserrat', serif;
+  }
+
+  h1 {
+      line-height: 1.2;
+      font-size: 2em;
+      margin-top: 1.5em;
+  }
+
+  p {
+    margin-left: 0px;
+    margin-right: 0px;
+    margin-top: 0.75em;
+    margin-bottom: 0.75em;
+  }
+
+  @media screen and (min-width: 680px) {
+      .max-width {
+          margin: 0 100px 0 170px;
+          max-width: 640px;
+      }
+  }
+  @media screen and (min-width: 980px) {
+      .max-width {
+          margin: 0 auto;
+      }
+  }
+  </style>
+</head>
+<body class="max-width">
+
+<h1>StyleGAN2</h1>
+<ul>
+    <li>Paper: <a href="http://arxiv.org/abs/1912.04958">http://arxiv.org/abs/1912.04958</a></li>
+    <li>Video: <a href="https://youtu.be/c-NJtV9Jvp0">https://youtu.be/c-NJtV9Jvp0</a></li>
+    <li>Code: <a href="https://github.com/NVlabs/stylegan2">https://github.com/NVlabs/stylegan2</a></li>
+</ul>
+
+<h1>Original StyleGAN</h1>
+<ul>
+    <li>Paper: <a href="https://arxiv.org/abs/1812.04948">https://arxiv.org/abs/1812.04948</a></li>
+    <li>Video: <a href="https://youtu.be/kSLJriaOumA">https://youtu.be/kSLJriaOumA</a></li>
+    <li>Code: <a href="https://github.com/NVlabs/stylegan">https://github.com/NVlabs/stylegan</a></li>
+    <li>FFHQ: <a href="https://github.com/NVlabs/ffhq-dataset">https://github.com/NVlabs/ffhq-dataset</a></li>
+</ul>
+
+</body>
+</html>
diff --git a/reconstruction/ostec/external/stylegan2/metrics/__init__.py b/reconstruction/ostec/external/stylegan2/metrics/__init__.py
new file mode 100644
index 000000000..9ab9908ef
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/__init__.py
@@ -0,0 +1,7 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+# empty
diff --git a/reconstruction/ostec/external/stylegan2/metrics/frechet_inception_distance.py b/reconstruction/ostec/external/stylegan2/metrics/frechet_inception_distance.py
new file mode 100644
index 000000000..ace0d6ec2
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/frechet_inception_distance.py
@@ -0,0 +1,73 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Frechet Inception Distance (FID)."""
+
+import os
+import numpy as np
+import scipy
+import tensorflow as tf
+import dnnlib.tflib as tflib
+
+from metrics import metric_base
+from training import misc
+
+#----------------------------------------------------------------------------
+
+class FID(metric_base.MetricBase):
+    def __init__(self, num_images, minibatch_per_gpu, **kwargs):
+        super().__init__(**kwargs)
+        self.num_images = num_images
+        self.minibatch_per_gpu = minibatch_per_gpu
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        minibatch_size = num_gpus * self.minibatch_per_gpu
+        inception = misc.load_pkl('https://drive.google.com/uc?id=1MzTY44rLToO5APn8TZmfR7_ENSe5aZUn') # inception_v3_features.pkl
+        activations = np.empty([self.num_images, inception.output_shape[1]], dtype=np.float32)
+
+        # Calculate statistics for reals.
+        cache_file = self._get_cache_file_for_reals(num_images=self.num_images)
+        os.makedirs(os.path.dirname(cache_file), exist_ok=True)
+        if os.path.isfile(cache_file):
+            mu_real, sigma_real = misc.load_pkl(cache_file)
+        else:
+            for idx, images in enumerate(self._iterate_reals(minibatch_size=minibatch_size)):
+                begin = idx * minibatch_size
+                end = min(begin + minibatch_size, self.num_images)
+                activations[begin:end] = inception.run(images[:end-begin], num_gpus=num_gpus, assume_frozen=True)
+                if end == self.num_images:
+                    break
+            mu_real = np.mean(activations, axis=0)
+            sigma_real = np.cov(activations, rowvar=False)
+            misc.save_pkl((mu_real, sigma_real), cache_file)
+
+        # Construct TensorFlow graph.
+        result_expr = []
+        for gpu_idx in range(num_gpus):
+            with tf.device('/gpu:%d' % gpu_idx):
+                Gs_clone = Gs.clone()
+                inception_clone = inception.clone()
+                latents = tf.random_normal([self.minibatch_per_gpu] + Gs_clone.input_shape[1:])
+                labels = self._get_random_labels_tf(self.minibatch_per_gpu)
+                images = Gs_clone.get_output_for(latents, labels, **Gs_kwargs)
+                images = tflib.convert_images_to_uint8(images)
+                result_expr.append(inception_clone.get_output_for(images))
+
+        # Calculate statistics for fakes.
+        for begin in range(0, self.num_images, minibatch_size):
+            self._report_progress(begin, self.num_images)
+            end = min(begin + minibatch_size, self.num_images)
+            activations[begin:end] = np.concatenate(tflib.run(result_expr), axis=0)[:end-begin]
+        mu_fake = np.mean(activations, axis=0)
+        sigma_fake = np.cov(activations, rowvar=False)
+
+        # Calculate FID.
+        m = np.square(mu_fake - mu_real).sum()
+        s, _ = scipy.linalg.sqrtm(np.dot(sigma_fake, sigma_real), disp=False) # pylint: disable=no-member
+        dist = m + np.trace(sigma_fake + sigma_real - 2*s)
+        self._report_result(np.real(dist))
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/inception_score.py b/reconstruction/ostec/external/stylegan2/metrics/inception_score.py
new file mode 100644
index 000000000..ff0543d41
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/inception_score.py
@@ -0,0 +1,58 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Inception Score (IS)."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib.tflib as tflib
+
+from metrics import metric_base
+from training import misc
+
+#----------------------------------------------------------------------------
+
+class IS(metric_base.MetricBase):
+    def __init__(self, num_images, num_splits, minibatch_per_gpu, **kwargs):
+        super().__init__(**kwargs)
+        self.num_images = num_images
+        self.num_splits = num_splits
+        self.minibatch_per_gpu = minibatch_per_gpu
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        minibatch_size = num_gpus * self.minibatch_per_gpu
+        inception = misc.load_pkl('https://drive.google.com/uc?id=1Mz9zQnIrusm3duZB91ng_aUIePFNI6Jx') # inception_v3_softmax.pkl
+        activations = np.empty([self.num_images, inception.output_shape[1]], dtype=np.float32)
+
+        # Construct TensorFlow graph.
+        result_expr = []
+        for gpu_idx in range(num_gpus):
+            with tf.device('/gpu:%d' % gpu_idx):
+                Gs_clone = Gs.clone()
+                inception_clone = inception.clone()
+                latents = tf.random_normal([self.minibatch_per_gpu] + Gs_clone.input_shape[1:])
+                labels = self._get_random_labels_tf(self.minibatch_per_gpu)
+                images = Gs_clone.get_output_for(latents, labels, **Gs_kwargs)
+                images = tflib.convert_images_to_uint8(images)
+                result_expr.append(inception_clone.get_output_for(images))
+
+        # Calculate activations for fakes.
+        for begin in range(0, self.num_images, minibatch_size):
+            self._report_progress(begin, self.num_images)
+            end = min(begin + minibatch_size, self.num_images)
+            activations[begin:end] = np.concatenate(tflib.run(result_expr), axis=0)[:end-begin]
+
+        # Calculate IS.
+        scores = []
+        for i in range(self.num_splits):
+            part = activations[i * self.num_images // self.num_splits : (i + 1) * self.num_images // self.num_splits]
+            kl = part * (np.log(part) - np.log(np.expand_dims(np.mean(part, 0), 0)))
+            kl = np.mean(np.sum(kl, 1))
+            scores.append(np.exp(kl))
+        self._report_result(np.mean(scores), suffix='_mean')
+        self._report_result(np.std(scores), suffix='_std')
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/linear_separability.py b/reconstruction/ostec/external/stylegan2/metrics/linear_separability.py
new file mode 100644
index 000000000..14bfb9958
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/linear_separability.py
@@ -0,0 +1,178 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Linear Separability (LS)."""
+
+from collections import defaultdict
+import numpy as np
+import sklearn.svm
+import tensorflow as tf
+import dnnlib.tflib as tflib
+
+from metrics import metric_base
+from training import misc
+
+#----------------------------------------------------------------------------
+
+classifier_urls = [
+    'https://drive.google.com/uc?id=1Q5-AI6TwWhCVM7Muu4tBM7rp5nG_gmCX', # celebahq-classifier-00-male.pkl
+    'https://drive.google.com/uc?id=1Q5c6HE__ReW2W8qYAXpao68V1ryuisGo', # celebahq-classifier-01-smiling.pkl
+    'https://drive.google.com/uc?id=1Q7738mgWTljPOJQrZtSMLxzShEhrvVsU', # celebahq-classifier-02-attractive.pkl
+    'https://drive.google.com/uc?id=1QBv2Mxe7ZLvOv1YBTLq-T4DS3HjmXV0o', # celebahq-classifier-03-wavy-hair.pkl
+    'https://drive.google.com/uc?id=1QIvKTrkYpUrdA45nf7pspwAqXDwWOLhV', # celebahq-classifier-04-young.pkl
+    'https://drive.google.com/uc?id=1QJPH5rW7MbIjFUdZT7vRYfyUjNYDl4_L', # celebahq-classifier-05-5-o-clock-shadow.pkl
+    'https://drive.google.com/uc?id=1QPZXSYf6cptQnApWS_T83sqFMun3rULY', # celebahq-classifier-06-arched-eyebrows.pkl
+    'https://drive.google.com/uc?id=1QPgoAZRqINXk_PFoQ6NwMmiJfxc5d2Pg', # celebahq-classifier-07-bags-under-eyes.pkl
+    'https://drive.google.com/uc?id=1QQPQgxgI6wrMWNyxFyTLSgMVZmRr1oO7', # celebahq-classifier-08-bald.pkl
+    'https://drive.google.com/uc?id=1QcSphAmV62UrCIqhMGgcIlZfoe8hfWaF', # celebahq-classifier-09-bangs.pkl
+    'https://drive.google.com/uc?id=1QdWTVwljClTFrrrcZnPuPOR4mEuz7jGh', # celebahq-classifier-10-big-lips.pkl
+    'https://drive.google.com/uc?id=1QgvEWEtr2mS4yj1b_Y3WKe6cLWL3LYmK', # celebahq-classifier-11-big-nose.pkl
+    'https://drive.google.com/uc?id=1QidfMk9FOKgmUUIziTCeo8t-kTGwcT18', # celebahq-classifier-12-black-hair.pkl
+    'https://drive.google.com/uc?id=1QthrJt-wY31GPtV8SbnZQZ0_UEdhasHO', # celebahq-classifier-13-blond-hair.pkl
+    'https://drive.google.com/uc?id=1QvCAkXxdYT4sIwCzYDnCL9Nb5TDYUxGW', # celebahq-classifier-14-blurry.pkl
+    'https://drive.google.com/uc?id=1QvLWuwSuWI9Ln8cpxSGHIciUsnmaw8L0', # celebahq-classifier-15-brown-hair.pkl
+    'https://drive.google.com/uc?id=1QxW6THPI2fqDoiFEMaV6pWWHhKI_OoA7', # celebahq-classifier-16-bushy-eyebrows.pkl
+    'https://drive.google.com/uc?id=1R71xKw8oTW2IHyqmRDChhTBkW9wq4N9v', # celebahq-classifier-17-chubby.pkl
+    'https://drive.google.com/uc?id=1RDn_fiLfEGbTc7JjazRXuAxJpr-4Pl67', # celebahq-classifier-18-double-chin.pkl
+    'https://drive.google.com/uc?id=1RGBuwXbaz5052bM4VFvaSJaqNvVM4_cI', # celebahq-classifier-19-eyeglasses.pkl
+    'https://drive.google.com/uc?id=1RIxOiWxDpUwhB-9HzDkbkLegkd7euRU9', # celebahq-classifier-20-goatee.pkl
+    'https://drive.google.com/uc?id=1RPaNiEnJODdr-fwXhUFdoSQLFFZC7rC-', # celebahq-classifier-21-gray-hair.pkl
+    'https://drive.google.com/uc?id=1RQH8lPSwOI2K_9XQCZ2Ktz7xm46o80ep', # celebahq-classifier-22-heavy-makeup.pkl
+    'https://drive.google.com/uc?id=1RXZM61xCzlwUZKq-X7QhxOg0D2telPow', # celebahq-classifier-23-high-cheekbones.pkl
+    'https://drive.google.com/uc?id=1RgASVHW8EWMyOCiRb5fsUijFu-HfxONM', # celebahq-classifier-24-mouth-slightly-open.pkl
+    'https://drive.google.com/uc?id=1RkC8JLqLosWMaRne3DARRgolhbtg_wnr', # celebahq-classifier-25-mustache.pkl
+    'https://drive.google.com/uc?id=1RqtbtFT2EuwpGTqsTYJDyXdnDsFCPtLO', # celebahq-classifier-26-narrow-eyes.pkl
+    'https://drive.google.com/uc?id=1Rs7hU-re8bBMeRHR-fKgMbjPh-RIbrsh', # celebahq-classifier-27-no-beard.pkl
+    'https://drive.google.com/uc?id=1RynDJQWdGOAGffmkPVCrLJqy_fciPF9E', # celebahq-classifier-28-oval-face.pkl
+    'https://drive.google.com/uc?id=1S0TZ_Hdv5cb06NDaCD8NqVfKy7MuXZsN', # celebahq-classifier-29-pale-skin.pkl
+    'https://drive.google.com/uc?id=1S3JPhZH2B4gVZZYCWkxoRP11q09PjCkA', # celebahq-classifier-30-pointy-nose.pkl
+    'https://drive.google.com/uc?id=1S3pQuUz-Jiywq_euhsfezWfGkfzLZ87W', # celebahq-classifier-31-receding-hairline.pkl
+    'https://drive.google.com/uc?id=1S6nyIl_SEI3M4l748xEdTV2vymB_-lrY', # celebahq-classifier-32-rosy-cheeks.pkl
+    'https://drive.google.com/uc?id=1S9P5WCi3GYIBPVYiPTWygrYIUSIKGxbU', # celebahq-classifier-33-sideburns.pkl
+    'https://drive.google.com/uc?id=1SANviG-pp08n7AFpE9wrARzozPIlbfCH', # celebahq-classifier-34-straight-hair.pkl
+    'https://drive.google.com/uc?id=1SArgyMl6_z7P7coAuArqUC2zbmckecEY', # celebahq-classifier-35-wearing-earrings.pkl
+    'https://drive.google.com/uc?id=1SC5JjS5J-J4zXFO9Vk2ZU2DT82TZUza_', # celebahq-classifier-36-wearing-hat.pkl
+    'https://drive.google.com/uc?id=1SDAQWz03HGiu0MSOKyn7gvrp3wdIGoj-', # celebahq-classifier-37-wearing-lipstick.pkl
+    'https://drive.google.com/uc?id=1SEtrVK-TQUC0XeGkBE9y7L8VXfbchyKX', # celebahq-classifier-38-wearing-necklace.pkl
+    'https://drive.google.com/uc?id=1SF_mJIdyGINXoV-I6IAxHB_k5dxiF6M-', # celebahq-classifier-39-wearing-necktie.pkl
+]
+
+#----------------------------------------------------------------------------
+
+def prob_normalize(p):
+    p = np.asarray(p).astype(np.float32)
+    assert len(p.shape) == 2
+    return p / np.sum(p)
+
+def mutual_information(p):
+    p = prob_normalize(p)
+    px = np.sum(p, axis=1)
+    py = np.sum(p, axis=0)
+    result = 0.0
+    for x in range(p.shape[0]):
+        p_x = px[x]
+        for y in range(p.shape[1]):
+            p_xy = p[x][y]
+            p_y = py[y]
+            if p_xy > 0.0:
+                result += p_xy * np.log2(p_xy / (p_x * p_y)) # get bits as output
+    return result
+
+def entropy(p):
+    p = prob_normalize(p)
+    result = 0.0
+    for x in range(p.shape[0]):
+        for y in range(p.shape[1]):
+            p_xy = p[x][y]
+            if p_xy > 0.0:
+                result -= p_xy * np.log2(p_xy)
+    return result
+
+def conditional_entropy(p):
+    # H(Y|X) where X corresponds to axis 0, Y to axis 1
+    # i.e., How many bits of additional information are needed to where we are on axis 1 if we know where we are on axis 0?
+    p = prob_normalize(p)
+    y = np.sum(p, axis=0, keepdims=True) # marginalize to calculate H(Y)
+    return max(0.0, entropy(y) - mutual_information(p)) # can slip just below 0 due to FP inaccuracies, clean those up.
+
+#----------------------------------------------------------------------------
+
+class LS(metric_base.MetricBase):
+    def __init__(self, num_samples, num_keep, attrib_indices, minibatch_per_gpu, **kwargs):
+        assert num_keep <= num_samples
+        super().__init__(**kwargs)
+        self.num_samples = num_samples
+        self.num_keep = num_keep
+        self.attrib_indices = attrib_indices
+        self.minibatch_per_gpu = minibatch_per_gpu
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        minibatch_size = num_gpus * self.minibatch_per_gpu
+
+        # Construct TensorFlow graph for each GPU.
+        result_expr = []
+        for gpu_idx in range(num_gpus):
+            with tf.device('/gpu:%d' % gpu_idx):
+                Gs_clone = Gs.clone()
+
+                # Generate images.
+                latents = tf.random_normal([self.minibatch_per_gpu] + Gs_clone.input_shape[1:])
+                labels = self._get_random_labels_tf(self.minibatch_per_gpu)
+                dlatents = Gs_clone.components.mapping.get_output_for(latents, labels, **Gs_kwargs)
+                images = Gs_clone.get_output_for(latents, None, **Gs_kwargs)
+
+                # Downsample to 256x256. The attribute classifiers were built for 256x256.
+                if images.shape[2] > 256:
+                    factor = images.shape[2] // 256
+                    images = tf.reshape(images, [-1, images.shape[1], images.shape[2] // factor, factor, images.shape[3] // factor, factor])
+                    images = tf.reduce_mean(images, axis=[3, 5])
+
+                # Run classifier for each attribute.
+                result_dict = dict(latents=latents, dlatents=dlatents[:,-1])
+                for attrib_idx in self.attrib_indices:
+                    classifier = misc.load_pkl(classifier_urls[attrib_idx])
+                    logits = classifier.get_output_for(images, None)
+                    predictions = tf.nn.softmax(tf.concat([logits, -logits], axis=1))
+                    result_dict[attrib_idx] = predictions
+                result_expr.append(result_dict)
+
+        # Sampling loop.
+        results = []
+        for begin in range(0, self.num_samples, minibatch_size):
+            self._report_progress(begin, self.num_samples)
+            results += tflib.run(result_expr)
+        results = {key: np.concatenate([value[key] for value in results], axis=0) for key in results[0].keys()}
+
+        # Calculate conditional entropy for each attribute.
+        conditional_entropies = defaultdict(list)
+        for attrib_idx in self.attrib_indices:
+            # Prune the least confident samples.
+            pruned_indices = list(range(self.num_samples))
+            pruned_indices = sorted(pruned_indices, key=lambda i: -np.max(results[attrib_idx][i]))
+            pruned_indices = pruned_indices[:self.num_keep]
+
+            # Fit SVM to the remaining samples.
+            svm_targets = np.argmax(results[attrib_idx][pruned_indices], axis=1)
+            for space in ['latents', 'dlatents']:
+                svm_inputs = results[space][pruned_indices]
+                try:
+                    svm = sklearn.svm.LinearSVC()
+                    svm.fit(svm_inputs, svm_targets)
+                    svm.score(svm_inputs, svm_targets)
+                    svm_outputs = svm.predict(svm_inputs)
+                except:
+                    svm_outputs = svm_targets # assume perfect prediction
+
+                # Calculate conditional entropy.
+                p = [[np.mean([case == (row, col) for case in zip(svm_outputs, svm_targets)]) for col in (0, 1)] for row in (0, 1)]
+                conditional_entropies[space].append(conditional_entropy(p))
+
+        # Calculate separability scores.
+        scores = {key: 2**np.sum(values) for key, values in conditional_entropies.items()}
+        self._report_result(scores['latents'], suffix='_z')
+        self._report_result(scores['dlatents'], suffix='_w')
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/metric_base.py b/reconstruction/ostec/external/stylegan2/metrics/metric_base.py
new file mode 100644
index 000000000..cbd0276d2
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/metric_base.py
@@ -0,0 +1,168 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Common definitions for GAN metrics."""
+
+import os
+import time
+import hashlib
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+
+from training import misc
+from training import dataset
+
+#----------------------------------------------------------------------------
+# Base class for metrics.
+
+class MetricBase:
+    def __init__(self, name):
+        self.name = name
+        self._dataset_obj = None
+        self._progress_lo = None
+        self._progress_hi = None
+        self._progress_max = None
+        self._progress_sec = None
+        self._progress_time = None
+        self._reset()
+
+    def close(self):
+        self._reset()
+
+    def _reset(self, network_pkl=None, run_dir=None, data_dir=None, dataset_args=None, mirror_augment=None):
+        if self._dataset_obj is not None:
+            self._dataset_obj.close()
+
+        self._network_pkl = network_pkl
+        self._data_dir = data_dir
+        self._dataset_args = dataset_args
+        self._dataset_obj = None
+        self._mirror_augment = mirror_augment
+        self._eval_time = 0
+        self._results = []
+
+        if (dataset_args is None or mirror_augment is None) and run_dir is not None:
+            run_config = misc.parse_config_for_previous_run(run_dir)
+            self._dataset_args = dict(run_config['dataset'])
+            self._dataset_args['shuffle_mb'] = 0
+            self._mirror_augment = run_config['train'].get('mirror_augment', False)
+
+    def configure_progress_reports(self, plo, phi, pmax, psec=15):
+        self._progress_lo = plo
+        self._progress_hi = phi
+        self._progress_max = pmax
+        self._progress_sec = psec
+
+    def run(self, network_pkl, run_dir=None, data_dir=None, dataset_args=None, mirror_augment=None, num_gpus=1, tf_config=None, log_results=True, Gs_kwargs=dict(is_validation=True)):
+        self._reset(network_pkl=network_pkl, run_dir=run_dir, data_dir=data_dir, dataset_args=dataset_args, mirror_augment=mirror_augment)
+        time_begin = time.time()
+        with tf.Graph().as_default(), tflib.create_session(tf_config).as_default(): # pylint: disable=not-context-manager
+            self._report_progress(0, 1)
+            _G, _D, Gs = misc.load_pkl(self._network_pkl)
+            self._evaluate(Gs, Gs_kwargs=Gs_kwargs, num_gpus=num_gpus)
+            self._report_progress(1, 1)
+        self._eval_time = time.time() - time_begin # pylint: disable=attribute-defined-outside-init
+
+        if log_results:
+            if run_dir is not None:
+                log_file = os.path.join(run_dir, 'metric-%s.txt' % self.name)
+                with dnnlib.util.Logger(log_file, 'a'):
+                    print(self.get_result_str().strip())
+            else:
+                print(self.get_result_str().strip())
+
+    def get_result_str(self):
+        network_name = os.path.splitext(os.path.basename(self._network_pkl))[0]
+        if len(network_name) > 29:
+            network_name = '...' + network_name[-26:]
+        result_str = '%-30s' % network_name
+        result_str += ' time %-12s' % dnnlib.util.format_time(self._eval_time)
+        for res in self._results:
+            result_str += ' ' + self.name + res.suffix + ' '
+            result_str += res.fmt % res.value
+        return result_str
+
+    def update_autosummaries(self):
+        for res in self._results:
+            tflib.autosummary.autosummary('Metrics/' + self.name + res.suffix, res.value)
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        raise NotImplementedError # to be overridden by subclasses
+
+    def _report_result(self, value, suffix='', fmt='%-10.4f'):
+        self._results += [dnnlib.EasyDict(value=value, suffix=suffix, fmt=fmt)]
+
+    def _report_progress(self, pcur, pmax, status_str=''):
+        if self._progress_lo is None or self._progress_hi is None or self._progress_max is None:
+            return
+        t = time.time()
+        if self._progress_sec is not None and self._progress_time is not None and t < self._progress_time + self._progress_sec:
+            return
+        self._progress_time = t
+        val = self._progress_lo + (pcur / pmax) * (self._progress_hi - self._progress_lo)
+        dnnlib.RunContext.get().update(status_str, int(val), self._progress_max)
+
+    def _get_cache_file_for_reals(self, extension='pkl', **kwargs):
+        all_args = dnnlib.EasyDict(metric_name=self.name, mirror_augment=self._mirror_augment)
+        all_args.update(self._dataset_args)
+        all_args.update(kwargs)
+        md5 = hashlib.md5(repr(sorted(all_args.items())).encode('utf-8'))
+        dataset_name = self._dataset_args.get('tfrecord_dir', None) or self._dataset_args.get('h5_file', None)
+        dataset_name = os.path.splitext(os.path.basename(dataset_name))[0]
+        return os.path.join('.stylegan2-cache', '%s-%s-%s.%s' % (md5.hexdigest(), self.name, dataset_name, extension))
+
+    def _get_dataset_obj(self):
+        if self._dataset_obj is None:
+            self._dataset_obj = dataset.load_dataset(data_dir=self._data_dir, **self._dataset_args)
+        return self._dataset_obj
+
+    def _iterate_reals(self, minibatch_size):
+        dataset_obj = self._get_dataset_obj()
+        while True:
+            images, _labels = dataset_obj.get_minibatch_np(minibatch_size)
+            if self._mirror_augment:
+                images = misc.apply_mirror_augment(images)
+            yield images
+
+    def _iterate_fakes(self, Gs, minibatch_size, num_gpus):
+        while True:
+            latents = np.random.randn(minibatch_size, *Gs.input_shape[1:])
+            fmt = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
+            images = Gs.run(latents, None, output_transform=fmt, is_validation=True, num_gpus=num_gpus, assume_frozen=True)
+            yield images
+
+    def _get_random_labels_tf(self, minibatch_size):
+        return self._get_dataset_obj().get_random_labels_tf(minibatch_size)
+
+#----------------------------------------------------------------------------
+# Group of multiple metrics.
+
+class MetricGroup:
+    def __init__(self, metric_kwarg_list):
+        self.metrics = [dnnlib.util.call_func_by_name(**kwargs) for kwargs in metric_kwarg_list]
+
+    def run(self, *args, **kwargs):
+        for metric in self.metrics:
+            metric.run(*args, **kwargs)
+
+    def get_result_str(self):
+        return ' '.join(metric.get_result_str() for metric in self.metrics)
+
+    def update_autosummaries(self):
+        for metric in self.metrics:
+            metric.update_autosummaries()
+
+#----------------------------------------------------------------------------
+# Dummy metric for debugging purposes.
+
+class DummyMetric(MetricBase):
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        _ = Gs, Gs_kwargs, num_gpus
+        self._report_result(0.0)
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/metric_defaults.py b/reconstruction/ostec/external/stylegan2/metrics/metric_defaults.py
new file mode 100644
index 000000000..4371db818
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/metric_defaults.py
@@ -0,0 +1,25 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Default metric definitions."""
+
+from dnnlib import EasyDict
+
+#----------------------------------------------------------------------------
+
+metric_defaults = EasyDict([(args.name, args) for args in [
+    EasyDict(name='fid50k',    func_name='metrics.frechet_inception_distance.FID', num_images=50000, minibatch_per_gpu=8),
+    EasyDict(name='is50k',     func_name='metrics.inception_score.IS',             num_images=50000, num_splits=10, minibatch_per_gpu=8),
+    EasyDict(name='ppl_zfull', func_name='metrics.perceptual_path_length.PPL',     num_samples=50000, epsilon=1e-4, space='z', sampling='full', crop=True, minibatch_per_gpu=4, Gs_overrides=dict(dtype='float32', mapping_dtype='float32')),
+    EasyDict(name='ppl_wfull', func_name='metrics.perceptual_path_length.PPL',     num_samples=50000, epsilon=1e-4, space='w', sampling='full', crop=True, minibatch_per_gpu=4, Gs_overrides=dict(dtype='float32', mapping_dtype='float32')),
+    EasyDict(name='ppl_zend',  func_name='metrics.perceptual_path_length.PPL',     num_samples=50000, epsilon=1e-4, space='z', sampling='end', crop=True, minibatch_per_gpu=4, Gs_overrides=dict(dtype='float32', mapping_dtype='float32')),
+    EasyDict(name='ppl_wend',  func_name='metrics.perceptual_path_length.PPL',     num_samples=50000, epsilon=1e-4, space='w', sampling='end', crop=True, minibatch_per_gpu=4, Gs_overrides=dict(dtype='float32', mapping_dtype='float32')),
+    EasyDict(name='ppl2_wend', func_name='metrics.perceptual_path_length.PPL',     num_samples=50000, epsilon=1e-4, space='w', sampling='end', crop=False, minibatch_per_gpu=4, Gs_overrides=dict(dtype='float32', mapping_dtype='float32')),
+    EasyDict(name='ls',        func_name='metrics.linear_separability.LS',         num_samples=200000, num_keep=100000, attrib_indices=range(40), minibatch_per_gpu=4),
+    EasyDict(name='pr50k3',    func_name='metrics.precision_recall.PR',            num_images=50000, nhood_size=3, minibatch_per_gpu=8, row_batch_size=10000, col_batch_size=10000),
+]])
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/perceptual_path_length.py b/reconstruction/ostec/external/stylegan2/metrics/perceptual_path_length.py
new file mode 100644
index 000000000..2e5e4d9e3
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/perceptual_path_length.py
@@ -0,0 +1,116 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Perceptual Path Length (PPL)."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib.tflib as tflib
+
+from metrics import metric_base
+from training import misc
+
+#----------------------------------------------------------------------------
+
+# Normalize batch of vectors.
+def normalize(v):
+    return v / tf.sqrt(tf.reduce_sum(tf.square(v), axis=-1, keepdims=True))
+
+# Spherical interpolation of a batch of vectors.
+def slerp(a, b, t):
+    a = normalize(a)
+    b = normalize(b)
+    d = tf.reduce_sum(a * b, axis=-1, keepdims=True)
+    p = t * tf.math.acos(d)
+    c = normalize(b - d * a)
+    d = a * tf.math.cos(p) + c * tf.math.sin(p)
+    return normalize(d)
+
+#----------------------------------------------------------------------------
+
+class PPL(metric_base.MetricBase):
+    def __init__(self, num_samples, epsilon, space, sampling, crop, minibatch_per_gpu, Gs_overrides, **kwargs):
+        assert space in ['z', 'w']
+        assert sampling in ['full', 'end']
+        super().__init__(**kwargs)
+        self.num_samples = num_samples
+        self.epsilon = epsilon
+        self.space = space
+        self.sampling = sampling
+        self.crop = crop
+        self.minibatch_per_gpu = minibatch_per_gpu
+        self.Gs_overrides = Gs_overrides
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        Gs_kwargs = dict(Gs_kwargs)
+        Gs_kwargs.update(self.Gs_overrides)
+        minibatch_size = num_gpus * self.minibatch_per_gpu
+
+        # Construct TensorFlow graph.
+        distance_expr = []
+        for gpu_idx in range(num_gpus):
+            with tf.device('/gpu:%d' % gpu_idx):
+                Gs_clone = Gs.clone()
+                noise_vars = [var for name, var in Gs_clone.components.synthesis.vars.items() if name.startswith('noise')]
+
+                # Generate random latents and interpolation t-values.
+                lat_t01 = tf.random_normal([self.minibatch_per_gpu * 2] + Gs_clone.input_shape[1:])
+                lerp_t = tf.random_uniform([self.minibatch_per_gpu], 0.0, 1.0 if self.sampling == 'full' else 0.0)
+                labels = tf.reshape(tf.tile(self._get_random_labels_tf(self.minibatch_per_gpu), [1, 2]), [self.minibatch_per_gpu * 2, -1])
+
+                # Interpolate in W or Z.
+                if self.space == 'w':
+                    dlat_t01 = Gs_clone.components.mapping.get_output_for(lat_t01, labels, **Gs_kwargs)
+                    dlat_t01 = tf.cast(dlat_t01, tf.float32)
+                    dlat_t0, dlat_t1 = dlat_t01[0::2], dlat_t01[1::2]
+                    dlat_e0 = tflib.lerp(dlat_t0, dlat_t1, lerp_t[:, np.newaxis, np.newaxis])
+                    dlat_e1 = tflib.lerp(dlat_t0, dlat_t1, lerp_t[:, np.newaxis, np.newaxis] + self.epsilon)
+                    dlat_e01 = tf.reshape(tf.stack([dlat_e0, dlat_e1], axis=1), dlat_t01.shape)
+                else: # space == 'z'
+                    lat_t0, lat_t1 = lat_t01[0::2], lat_t01[1::2]
+                    lat_e0 = slerp(lat_t0, lat_t1, lerp_t[:, np.newaxis])
+                    lat_e1 = slerp(lat_t0, lat_t1, lerp_t[:, np.newaxis] + self.epsilon)
+                    lat_e01 = tf.reshape(tf.stack([lat_e0, lat_e1], axis=1), lat_t01.shape)
+                    dlat_e01 = Gs_clone.components.mapping.get_output_for(lat_e01, labels, **Gs_kwargs)
+
+                # Synthesize images.
+                with tf.control_dependencies([var.initializer for var in noise_vars]): # use same noise inputs for the entire minibatch
+                    images = Gs_clone.components.synthesis.get_output_for(dlat_e01, randomize_noise=False, **Gs_kwargs)
+                    images = tf.cast(images, tf.float32)
+
+                # Crop only the face region.
+                if self.crop:
+                    c = int(images.shape[2] // 8)
+                    images = images[:, :, c*3 : c*7, c*2 : c*6]
+
+                # Downsample image to 256x256 if it's larger than that. VGG was built for 224x224 images.
+                factor = images.shape[2] // 256
+                if factor > 1:
+                    images = tf.reshape(images, [-1, images.shape[1], images.shape[2] // factor, factor, images.shape[3] // factor, factor])
+                    images = tf.reduce_mean(images, axis=[3,5])
+
+                # Scale dynamic range from [-1,1] to [0,255] for VGG.
+                images = (images + 1) * (255 / 2)
+
+                # Evaluate perceptual distance.
+                img_e0, img_e1 = images[0::2], images[1::2]
+                distance_measure = misc.load_pkl('https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2') # vgg16_zhang_perceptual.pkl
+                distance_expr.append(distance_measure.get_output_for(img_e0, img_e1) * (1 / self.epsilon**2))
+
+        # Sampling loop.
+        all_distances = []
+        for begin in range(0, self.num_samples, minibatch_size):
+            self._report_progress(begin, self.num_samples)
+            all_distances += tflib.run(distance_expr)
+        all_distances = np.concatenate(all_distances, axis=0)
+
+        # Reject outliers.
+        lo = np.percentile(all_distances, 1, interpolation='lower')
+        hi = np.percentile(all_distances, 99, interpolation='higher')
+        filtered_distances = np.extract(np.logical_and(lo <= all_distances, all_distances <= hi), all_distances)
+        self._report_result(np.mean(filtered_distances))
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/metrics/precision_recall.py b/reconstruction/ostec/external/stylegan2/metrics/precision_recall.py
new file mode 100644
index 000000000..addf9bdc5
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/metrics/precision_recall.py
@@ -0,0 +1,224 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Precision/Recall (PR)."""
+
+import os
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+
+from metrics import metric_base
+from training import misc
+
+#----------------------------------------------------------------------------
+
+def batch_pairwise_distances(U, V):
+    """ Compute pairwise distances between two batches of feature vectors."""
+    with tf.variable_scope('pairwise_dist_block'):
+        # Squared norms of each row in U and V.
+        norm_u = tf.reduce_sum(tf.square(U), 1)
+        norm_v = tf.reduce_sum(tf.square(V), 1)
+
+        # norm_u as a row and norm_v as a column vectors.
+        norm_u = tf.reshape(norm_u, [-1, 1])
+        norm_v = tf.reshape(norm_v, [1, -1])
+
+        # Pairwise squared Euclidean distances.
+        D = tf.maximum(norm_u - 2*tf.matmul(U, V, False, True) + norm_v, 0.0)
+
+    return D
+
+#----------------------------------------------------------------------------
+
+class DistanceBlock():
+    """Distance block."""
+    def __init__(self, num_features, num_gpus):
+        self.num_features = num_features
+        self.num_gpus = num_gpus
+
+        # Initialize TF graph to calculate pairwise distances.
+        with tf.device('/cpu:0'):
+            self._features_batch1 = tf.placeholder(tf.float16, shape=[None, self.num_features])
+            self._features_batch2 = tf.placeholder(tf.float16, shape=[None, self.num_features])
+            features_split2 = tf.split(self._features_batch2, self.num_gpus, axis=0)
+            distances_split = []
+            for gpu_idx in range(self.num_gpus):
+                with tf.device('/gpu:%d' % gpu_idx):
+                    distances_split.append(batch_pairwise_distances(self._features_batch1, features_split2[gpu_idx]))
+            self._distance_block = tf.concat(distances_split, axis=1)
+
+    def pairwise_distances(self, U, V):
+        """Evaluate pairwise distances between two batches of feature vectors."""
+        return self._distance_block.eval(feed_dict={self._features_batch1: U, self._features_batch2: V})
+
+#----------------------------------------------------------------------------
+
+class ManifoldEstimator():
+    """Finds an estimate for the manifold of given feature vectors."""
+    def __init__(self, distance_block, features, row_batch_size, col_batch_size, nhood_sizes, clamp_to_percentile=None):
+        """Find an estimate of the manifold of given feature vectors."""
+        num_images = features.shape[0]
+        self.nhood_sizes = nhood_sizes
+        self.num_nhoods = len(nhood_sizes)
+        self.row_batch_size = row_batch_size
+        self.col_batch_size = col_batch_size
+        self._ref_features = features
+        self._distance_block = distance_block
+
+        # Estimate manifold of features by calculating distances to kth nearest neighbor of each sample.
+        self.D = np.zeros([num_images, self.num_nhoods], dtype=np.float16)
+        distance_batch = np.zeros([row_batch_size, num_images], dtype=np.float16)
+        seq = np.arange(max(self.nhood_sizes) + 1, dtype=np.int32)
+
+        for begin1 in range(0, num_images, row_batch_size):
+            end1 = min(begin1 + row_batch_size, num_images)
+            row_batch = features[begin1:end1]
+
+            for begin2 in range(0, num_images, col_batch_size):
+                end2 = min(begin2 + col_batch_size, num_images)
+                col_batch = features[begin2:end2]
+
+                # Compute distances between batches.
+                distance_batch[0:end1-begin1, begin2:end2] = self._distance_block.pairwise_distances(row_batch, col_batch)
+
+            # Find the kth nearest neighbor from the current batch.
+            self.D[begin1:end1, :] = np.partition(distance_batch[0:end1-begin1, :], seq, axis=1)[:, self.nhood_sizes]
+
+        if clamp_to_percentile is not None:
+            max_distances = np.percentile(self.D, clamp_to_percentile, axis=0)
+            self.D[self.D > max_distances] = 0  #max_distances  # 0
+
+    def evaluate(self, eval_features, return_realism=False, return_neighbors=False):
+        """Evaluate if new feature vectors are in the estimated manifold."""
+        num_eval_images = eval_features.shape[0]
+        num_ref_images = self.D.shape[0]
+        distance_batch = np.zeros([self.row_batch_size, num_ref_images], dtype=np.float16)
+        batch_predictions = np.zeros([num_eval_images, self.num_nhoods], dtype=np.int32)
+        #max_realism_score = np.zeros([num_eval_images,], dtype=np.float32)
+        realism_score = np.zeros([num_eval_images,], dtype=np.float32)
+        nearest_indices = np.zeros([num_eval_images,], dtype=np.int32)
+
+        for begin1 in range(0, num_eval_images, self.row_batch_size):
+            end1 = min(begin1 + self.row_batch_size, num_eval_images)
+            feature_batch = eval_features[begin1:end1]
+
+            for begin2 in range(0, num_ref_images, self.col_batch_size):
+                end2 = min(begin2 + self.col_batch_size, num_ref_images)
+                ref_batch = self._ref_features[begin2:end2]
+
+                distance_batch[0:end1-begin1, begin2:end2] = self._distance_block.pairwise_distances(feature_batch, ref_batch)
+
+            # From the minibatch of new feature vectors, determine if they are in the estimated manifold.
+            # If a feature vector is inside a hypersphere of some reference sample, then the new sample lies on the estimated manifold.
+            # The radii of the hyperspheres are determined from distances of neighborhood size k.
+            samples_in_manifold = distance_batch[0:end1-begin1, :, None] <= self.D
+            batch_predictions[begin1:end1] = np.any(samples_in_manifold, axis=1).astype(np.int32)
+
+            #max_realism_score[begin1:end1] = np.max(self.D[:, 0] / (distance_batch[0:end1-begin1, :] + 1e-18), axis=1)
+            #nearest_indices[begin1:end1] = np.argmax(self.D[:, 0] / (distance_batch[0:end1-begin1, :] + 1e-18), axis=1)
+            nearest_indices[begin1:end1] = np.argmin(distance_batch[0:end1-begin1, :], axis=1)
+            realism_score[begin1:end1] = self.D[nearest_indices[begin1:end1], 0] / np.min(distance_batch[0:end1-begin1, :], axis=1)
+
+        if return_realism and return_neighbors:
+            return batch_predictions, realism_score, nearest_indices
+        elif return_realism:
+            return batch_predictions, realism_score
+        elif return_neighbors:
+            return batch_predictions, nearest_indices
+
+        return batch_predictions
+
+#----------------------------------------------------------------------------
+
+def knn_precision_recall_features(ref_features, eval_features, feature_net, nhood_sizes,
+                                  row_batch_size, col_batch_size, num_gpus):
+    """Calculates k-NN precision and recall for two sets of feature vectors."""
+    state = dnnlib.EasyDict()
+    #num_images = ref_features.shape[0]
+    num_features = feature_net.output_shape[1]
+    state.ref_features = ref_features
+    state.eval_features = eval_features
+
+    # Initialize DistanceBlock and ManifoldEstimators.
+    distance_block = DistanceBlock(num_features, num_gpus)
+    state.ref_manifold = ManifoldEstimator(distance_block, state.ref_features, row_batch_size, col_batch_size, nhood_sizes)
+    state.eval_manifold = ManifoldEstimator(distance_block, state.eval_features, row_batch_size, col_batch_size, nhood_sizes)
+
+    # Evaluate precision and recall using k-nearest neighbors.
+    #print('Evaluating k-NN precision and recall with %i samples...' % num_images)
+    #start = time.time()
+
+    # Precision: How many points from eval_features are in ref_features manifold.
+    state.precision, state.realism_scores, state.nearest_neighbors = state.ref_manifold.evaluate(state.eval_features, return_realism=True, return_neighbors=True)
+    state.knn_precision = state.precision.mean(axis=0)
+
+    # Recall: How many points from ref_features are in eval_features manifold.
+    state.recall = state.eval_manifold.evaluate(state.ref_features)
+    state.knn_recall = state.recall.mean(axis=0)
+
+    #elapsed_time = time.time() - start
+    #print('Done evaluation in: %gs' % elapsed_time)
+
+    return state
+
+#----------------------------------------------------------------------------
+
+class PR(metric_base.MetricBase):
+    def __init__(self, num_images, nhood_size, minibatch_per_gpu, row_batch_size, col_batch_size, **kwargs):
+        super().__init__(**kwargs)
+        self.num_images = num_images
+        self.nhood_size = nhood_size
+        self.minibatch_per_gpu = minibatch_per_gpu
+        self.row_batch_size = row_batch_size
+        self.col_batch_size = col_batch_size
+
+    def _evaluate(self, Gs, Gs_kwargs, num_gpus):
+        minibatch_size = num_gpus * self.minibatch_per_gpu
+        feature_net = misc.load_pkl('https://drive.google.com/uc?id=1MzY4MFpZzE-mNS26pzhYlWN-4vMm2ytu') # vgg16.pkl
+
+        # Calculate features for reals.
+        cache_file = self._get_cache_file_for_reals(num_images=self.num_images)
+        os.makedirs(os.path.dirname(cache_file), exist_ok=True)
+        if os.path.isfile(cache_file):
+            ref_features = misc.load_pkl(cache_file)
+        else:
+            ref_features = np.empty([self.num_images, feature_net.output_shape[1]], dtype=np.float32)
+            for idx, images in enumerate(self._iterate_reals(minibatch_size=minibatch_size)):
+                begin = idx * minibatch_size
+                end = min(begin + minibatch_size, self.num_images)
+                ref_features[begin:end] = feature_net.run(images[:end-begin], num_gpus=num_gpus, assume_frozen=True)
+                if end == self.num_images:
+                    break
+            misc.save_pkl(ref_features, cache_file)
+
+        # Construct TensorFlow graph.
+        result_expr = []
+        for gpu_idx in range(num_gpus):
+            with tf.device('/gpu:%d' % gpu_idx):
+                Gs_clone = Gs.clone()
+                feature_net_clone = feature_net.clone()
+                latents = tf.random_normal([self.minibatch_per_gpu] + Gs_clone.input_shape[1:])
+                labels = self._get_random_labels_tf(self.minibatch_per_gpu)
+                images = Gs_clone.get_output_for(latents, labels, **Gs_kwargs)
+                images = tflib.convert_images_to_uint8(images)
+                result_expr.append(feature_net_clone.get_output_for(images))
+
+        # Calculate features for fakes.
+        eval_features = np.empty([self.num_images, feature_net.output_shape[1]], dtype=np.float32)
+        for begin in range(0, self.num_images, minibatch_size):
+            self._report_progress(begin, self.num_images)
+            end = min(begin + minibatch_size, self.num_images)
+            eval_features[begin:end] = np.concatenate(tflib.run(result_expr), axis=0)[:end-begin]
+
+        # Calculate precision and recall.
+        state = knn_precision_recall_features(ref_features=ref_features, eval_features=eval_features, feature_net=feature_net,
+            nhood_sizes=[self.nhood_size], row_batch_size=self.row_batch_size, col_batch_size=self.row_batch_size, num_gpus=num_gpus)
+        self._report_result(state.knn_precision[0], suffix='_precision')
+        self._report_result(state.knn_recall[0], suffix='_recall')
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/pretrained_networks.py b/reconstruction/ostec/external/stylegan2/pretrained_networks.py
new file mode 100644
index 000000000..d9449a067
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/pretrained_networks.py
@@ -0,0 +1,80 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""List of pre-trained StyleGAN2 networks located on Google Drive."""
+
+import pickle
+import external.stylegan2.dnnlib
+from external.stylegan2.dnnlib import tflib as tflib
+
+#----------------------------------------------------------------------------
+# StyleGAN2 Google Drive root: https://drive.google.com/open?id=1QHc-yF5C3DChRwSdZKcx1w6K8JvSxQi7
+
+gdrive_urls = {
+    'gdrive:networks/stylegan2-car-config-a.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-a.pkl',
+    'gdrive:networks/stylegan2-car-config-b.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-b.pkl',
+    'gdrive:networks/stylegan2-car-config-c.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-c.pkl',
+    'gdrive:networks/stylegan2-car-config-d.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-d.pkl',
+    'gdrive:networks/stylegan2-car-config-e.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-e.pkl',
+    'gdrive:networks/stylegan2-car-config-f.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-car-config-f.pkl',
+    'gdrive:networks/stylegan2-cat-config-a.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-cat-config-a.pkl',
+    'gdrive:networks/stylegan2-cat-config-f.pkl':                           'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-cat-config-f.pkl',
+    'gdrive:networks/stylegan2-church-config-a.pkl':                        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-church-config-a.pkl',
+    'gdrive:networks/stylegan2-church-config-f.pkl':                        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-church-config-f.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-a.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-a.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-b.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-b.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-c.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-c.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-d.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-d.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-e.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-e.pkl',
+    'gdrive:networks/stylegan2-ffhq-config-f.pkl':                          'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-f.pkl',
+    'gdrive:networks/stylegan2-horse-config-a.pkl':                         'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-horse-config-a.pkl',
+    'gdrive:networks/stylegan2-horse-config-f.pkl':                         'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-horse-config-f.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gorig-Dorig.pkl':        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gorig-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gorig-Dresnet.pkl':      'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gorig-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gorig-Dskip.pkl':        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gorig-Dskip.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gresnet-Dorig.pkl':      'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gresnet-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gresnet-Dresnet.pkl':    'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gresnet-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gresnet-Dskip.pkl':      'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gresnet-Dskip.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gskip-Dorig.pkl':        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gskip-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gskip-Dresnet.pkl':      'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gskip-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-car-config-e-Gskip-Dskip.pkl':        'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-car-config-e-Gskip-Dskip.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gorig-Dorig.pkl':       'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gorig-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gorig-Dresnet.pkl':     'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gorig-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gorig-Dskip.pkl':       'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gorig-Dskip.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gresnet-Dorig.pkl':     'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gresnet-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gresnet-Dresnet.pkl':   'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gresnet-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gresnet-Dskip.pkl':     'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gresnet-Dskip.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gskip-Dorig.pkl':       'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gskip-Dorig.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gskip-Dresnet.pkl':     'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gskip-Dresnet.pkl',
+    'gdrive:networks/table2/stylegan2-ffhq-config-e-Gskip-Dskip.pkl':       'http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/table2/stylegan2-ffhq-config-e-Gskip-Dskip.pkl',
+}
+
+#----------------------------------------------------------------------------
+
+def get_path_or_url(path_or_gdrive_path):
+    return gdrive_urls.get(path_or_gdrive_path, path_or_gdrive_path)
+
+#----------------------------------------------------------------------------
+
+_cached_networks = dict()
+
+def load_networks(path_or_gdrive_path):
+    path_or_url = get_path_or_url(path_or_gdrive_path)
+    if path_or_url in _cached_networks:
+        return _cached_networks[path_or_url]
+
+    if external.stylegan2.dnnlib.util.is_url(path_or_url):
+        stream = external.stylegan2.dnnlib.util.open_url(path_or_url, cache_dir='.stylegan2-cache')
+    else:
+        stream = open(path_or_url, 'rb')
+
+    tflib.init_tf()
+    with stream:
+        G, D, Gs = pickle.load(stream, encoding='latin1')
+    _cached_networks[path_or_url] = G, D, Gs
+    return G, D, Gs
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/projector.py b/reconstruction/ostec/external/stylegan2/projector.py
new file mode 100644
index 000000000..7a2f98996
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/projector.py
@@ -0,0 +1,206 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+
+from training import misc
+
+#----------------------------------------------------------------------------
+
+class Projector:
+    def __init__(self):
+        self.num_steps                  = 1000
+        self.dlatent_avg_samples        = 10000
+        self.initial_learning_rate      = 0.1
+        self.initial_noise_factor       = 0.05
+        self.lr_rampdown_length         = 0.25
+        self.lr_rampup_length           = 0.05
+        self.noise_ramp_length          = 0.75
+        self.regularize_noise_weight    = 1e5
+        self.verbose                    = False
+        self.clone_net                  = True
+
+        self._Gs                    = None
+        self._minibatch_size        = None
+        self._dlatent_avg           = None
+        self._dlatent_std           = None
+        self._noise_vars            = None
+        self._noise_init_op         = None
+        self._noise_normalize_op    = None
+        self._dlatents_var          = None
+        self._noise_in              = None
+        self._dlatents_expr         = None
+        self._images_expr           = None
+        self._target_images_var     = None
+        self._lpips                 = None
+        self._dist                  = None
+        self._loss                  = None
+        self._reg_sizes             = None
+        self._lrate_in              = None
+        self._opt                   = None
+        self._opt_step              = None
+        self._cur_step              = None
+
+    def _info(self, *args):
+        if self.verbose:
+            print('Projector:', *args)
+
+    def set_network(self, Gs, minibatch_size=1):
+        assert minibatch_size == 1
+        self._Gs = Gs
+        self._minibatch_size = minibatch_size
+        if self._Gs is None:
+            return
+        if self.clone_net:
+            self._Gs = self._Gs.clone()
+
+        # Find dlatent stats.
+        self._info('Finding W midpoint and stddev using %d samples...' % self.dlatent_avg_samples)
+        latent_samples = np.random.RandomState(123).randn(self.dlatent_avg_samples, *self._Gs.input_shapes[0][1:])
+        dlatent_samples = self._Gs.components.mapping.run(latent_samples, None)[:, :1, :] # [N, 1, 512]
+        self._dlatent_avg = np.mean(dlatent_samples, axis=0, keepdims=True) # [1, 1, 512]
+        self._dlatent_std = (np.sum((dlatent_samples - self._dlatent_avg) ** 2) / self.dlatent_avg_samples) ** 0.5
+        self._info('std = %g' % self._dlatent_std)
+
+        # Find noise inputs.
+        self._info('Setting up noise inputs...')
+        self._noise_vars = []
+        noise_init_ops = []
+        noise_normalize_ops = []
+        while True:
+            n = 'G_synthesis/noise%d' % len(self._noise_vars)
+            if not n in self._Gs.vars:
+                break
+            v = self._Gs.vars[n]
+            self._noise_vars.append(v)
+            noise_init_ops.append(tf.assign(v, tf.random_normal(tf.shape(v), dtype=tf.float32)))
+            noise_mean = tf.reduce_mean(v)
+            noise_std = tf.reduce_mean((v - noise_mean)**2)**0.5
+            noise_normalize_ops.append(tf.assign(v, (v - noise_mean) / noise_std))
+            self._info(n, v)
+        self._noise_init_op = tf.group(*noise_init_ops)
+        self._noise_normalize_op = tf.group(*noise_normalize_ops)
+
+        # Image output graph.
+        self._info('Building image output graph...')
+        self._dlatents_var = tf.Variable(tf.zeros([self._minibatch_size] + list(self._dlatent_avg.shape[1:])), name='dlatents_var')
+        self._noise_in = tf.placeholder(tf.float32, [], name='noise_in')
+        dlatents_noise = tf.random.normal(shape=self._dlatents_var.shape) * self._noise_in
+        self._dlatents_expr = tf.tile(self._dlatents_var + dlatents_noise, [1, self._Gs.components.synthesis.input_shape[1], 1])
+        self._images_expr = self._Gs.components.synthesis.get_output_for(self._dlatents_expr, randomize_noise=False)
+
+        # Downsample image to 256x256 if it's larger than that. VGG was built for 224x224 images.
+        proc_images_expr = (self._images_expr + 1) * (255 / 2)
+        sh = proc_images_expr.shape.as_list()
+        if sh[2] > 256:
+            factor = sh[2] // 256
+            proc_images_expr = tf.reduce_mean(tf.reshape(proc_images_expr, [-1, sh[1], sh[2] // factor, factor, sh[2] // factor, factor]), axis=[3,5])
+
+        # Loss graph.
+        self._info('Building loss graph...')
+        self._target_images_var = tf.Variable(tf.zeros(proc_images_expr.shape), name='target_images_var')
+        if self._lpips is None:
+            self._lpips = misc.load_pkl('https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2') # vgg16_zhang_perceptual.pkl
+        self._dist = self._lpips.get_output_for(proc_images_expr, self._target_images_var)
+        self._loss = tf.reduce_sum(self._dist)
+
+        # Noise regularization graph.
+        self._info('Building noise regularization graph...')
+        reg_loss = 0.0
+        for v in self._noise_vars:
+            sz = v.shape[2]
+            while True:
+                reg_loss += tf.reduce_mean(v * tf.roll(v, shift=1, axis=3))**2 + tf.reduce_mean(v * tf.roll(v, shift=1, axis=2))**2
+                if sz <= 8:
+                    break # Small enough already
+                v = tf.reshape(v, [1, 1, sz//2, 2, sz//2, 2]) # Downscale
+                v = tf.reduce_mean(v, axis=[3, 5])
+                sz = sz // 2
+        self._loss += reg_loss * self.regularize_noise_weight
+
+        # Optimizer.
+        self._info('Setting up optimizer...')
+        self._lrate_in = tf.placeholder(tf.float32, [], name='lrate_in')
+        self._opt = dnnlib.tflib.Optimizer(learning_rate=self._lrate_in)
+        self._opt.register_gradients(self._loss, [self._dlatents_var] + self._noise_vars)
+        self._opt_step = self._opt.apply_updates()
+
+    def run(self, target_images):
+        # Run to completion.
+        self.start(target_images)
+        while self._cur_step < self.num_steps:
+            self.step()
+
+        # Collect results.
+        pres = dnnlib.EasyDict()
+        pres.dlatents = self.get_dlatents()
+        pres.noises = self.get_noises()
+        pres.images = self.get_images()
+        return pres
+
+    def start(self, target_images):
+        assert self._Gs is not None
+
+        # Prepare target images.
+        self._info('Preparing target images...')
+        target_images = np.asarray(target_images, dtype='float32')
+        target_images = (target_images + 1) * (255 / 2)
+        sh = target_images.shape
+        assert sh[0] == self._minibatch_size
+        if sh[2] > self._target_images_var.shape[2]:
+            factor = sh[2] // self._target_images_var.shape[2]
+            target_images = np.reshape(target_images, [-1, sh[1], sh[2] // factor, factor, sh[3] // factor, factor]).mean((3, 5))
+
+        # Initialize optimization state.
+        self._info('Initializing optimization state...')
+        tflib.set_vars({self._target_images_var: target_images, self._dlatents_var: np.tile(self._dlatent_avg, [self._minibatch_size, 1, 1])})
+        tflib.run(self._noise_init_op)
+        self._opt.reset_optimizer_state()
+        self._cur_step = 0
+
+    def step(self):
+        assert self._cur_step is not None
+        if self._cur_step >= self.num_steps:
+            return
+        if self._cur_step == 0:
+            self._info('Running...')
+
+        # Hyperparameters.
+        t = self._cur_step / self.num_steps
+        noise_strength = self._dlatent_std * self.initial_noise_factor * max(0.0, 1.0 - t / self.noise_ramp_length) ** 2
+        lr_ramp = min(1.0, (1.0 - t) / self.lr_rampdown_length)
+        lr_ramp = 0.5 - 0.5 * np.cos(lr_ramp * np.pi)
+        lr_ramp = lr_ramp * min(1.0, t / self.lr_rampup_length)
+        learning_rate = self.initial_learning_rate * lr_ramp
+
+        # Train.
+        feed_dict = {self._noise_in: noise_strength, self._lrate_in: learning_rate}
+        _, dist_value, loss_value = tflib.run([self._opt_step, self._dist, self._loss], feed_dict)
+        tflib.run(self._noise_normalize_op)
+
+        # Print status.
+        self._cur_step += 1
+        if self._cur_step == self.num_steps or self._cur_step % 10 == 0:
+            self._info('%-8d%-12g%-12g' % (self._cur_step, dist_value, loss_value))
+        if self._cur_step == self.num_steps:
+            self._info('Done.')
+
+    def get_cur_step(self):
+        return self._cur_step
+
+    def get_dlatents(self):
+        return tflib.run(self._dlatents_expr, {self._noise_in: 0})
+
+    def get_noises(self):
+        return tflib.run(self._noise_vars)
+
+    def get_images(self):
+        return tflib.run(self._images_expr, {self._noise_in: 0})
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/run_generator.py b/reconstruction/ostec/external/stylegan2/run_generator.py
new file mode 100644
index 000000000..1cd52c787
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/run_generator.py
@@ -0,0 +1,173 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+import argparse
+import numpy as np
+import PIL.Image
+import dnnlib
+import dnnlib.tflib as tflib
+import re
+import sys
+
+import pretrained_networks
+
+#----------------------------------------------------------------------------
+
+def generate_images(network_pkl, seeds, truncation_psi):
+    print('Loading networks from "%s"...' % network_pkl)
+    _G, _D, Gs = pretrained_networks.load_networks(network_pkl)
+    noise_vars = [var for name, var in Gs.components.synthesis.vars.items() if name.startswith('noise')]
+
+    Gs_kwargs = dnnlib.EasyDict()
+    Gs_kwargs.output_transform = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
+    Gs_kwargs.randomize_noise = False
+    if truncation_psi is not None:
+        Gs_kwargs.truncation_psi = truncation_psi
+
+    for seed_idx, seed in enumerate(seeds):
+        print('Generating image for seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
+        rnd = np.random.RandomState(seed)
+        z = rnd.randn(1, *Gs.input_shape[1:]) # [minibatch, component]
+        tflib.set_vars({var: rnd.randn(*var.shape.as_list()) for var in noise_vars}) # [height, width]
+        images = Gs.run(z, None, **Gs_kwargs) # [minibatch, height, width, channel]
+        PIL.Image.fromarray(images[0], 'RGB').save(dnnlib.make_run_dir_path('seed%04d.png' % seed))
+
+#----------------------------------------------------------------------------
+
+def style_mixing_example(network_pkl, row_seeds, col_seeds, truncation_psi, col_styles, minibatch_size=4):
+    print('Loading networks from "%s"...' % network_pkl)
+    _G, _D, Gs = pretrained_networks.load_networks(network_pkl)
+    w_avg = Gs.get_var('dlatent_avg') # [component]
+
+    Gs_syn_kwargs = dnnlib.EasyDict()
+    Gs_syn_kwargs.output_transform = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
+    Gs_syn_kwargs.randomize_noise = False
+    Gs_syn_kwargs.minibatch_size = minibatch_size
+
+    print('Generating W vectors...')
+    all_seeds = list(set(row_seeds + col_seeds))
+    all_z = np.stack([np.random.RandomState(seed).randn(*Gs.input_shape[1:]) for seed in all_seeds]) # [minibatch, component]
+    all_w = Gs.components.mapping.run(all_z, None) # [minibatch, layer, component]
+    w1 = np.load('latent_representations/im_right_01.npy')
+    w2 = np.load('latent_representations/04753.000002.02_C_01.npy')
+    all_w = np.stack([w1, w2],0)
+    all_w = w_avg + (all_w - w_avg) * truncation_psi # [minibatch, layer, component]
+    w_dict = {seed: w for seed, w in zip(all_seeds, list(all_w))} # [layer, component]
+
+    print('Generating images...')
+    all_images = Gs.components.synthesis.run(all_w, **Gs_syn_kwargs) # [minibatch, height, width, channel]
+    image_dict = {(seed, seed): image for seed, image in zip(all_seeds, list(all_images))}
+
+    print('Generating style-mixed images...')
+    for row_seed in row_seeds:
+        for col_seed in col_seeds:
+            w = w_dict[row_seed].copy()
+            w[col_styles] = w_dict[col_seed][col_styles]
+            image = Gs.components.synthesis.run(w[np.newaxis], **Gs_syn_kwargs)[0]
+            image_dict[(row_seed, col_seed)] = image
+
+    print('Saving images...')
+    for (row_seed, col_seed), image in image_dict.items():
+        PIL.Image.fromarray(image, 'RGB').save(dnnlib.make_run_dir_path('%d-%d.png' % (row_seed, col_seed)))
+
+    print('Saving image grid...')
+    _N, _C, H, W = Gs.output_shape
+    canvas = PIL.Image.new('RGB', (W * (len(col_seeds) + 1), H * (len(row_seeds) + 1)), 'black')
+    for row_idx, row_seed in enumerate([None] + row_seeds):
+        for col_idx, col_seed in enumerate([None] + col_seeds):
+            if row_seed is None and col_seed is None:
+                continue
+            key = (row_seed, col_seed)
+            if row_seed is None:
+                key = (col_seed, col_seed)
+            if col_seed is None:
+                key = (row_seed, row_seed)
+            canvas.paste(PIL.Image.fromarray(image_dict[key], 'RGB'), (W * col_idx, H * row_idx))
+    canvas.save(dnnlib.make_run_dir_path('grid.png'))
+
+#----------------------------------------------------------------------------
+
+def _parse_num_range(s):
+    '''Accept either a comma separated list of numbers 'a,b,c' or a range 'a-c' and return as a list of ints.'''
+
+    range_re = re.compile(r'^(\d+)-(\d+)$')
+    m = range_re.match(s)
+    if m:
+        return range(int(m.group(1)), int(m.group(2))+1)
+    vals = s.split(',')
+    return [int(x) for x in vals]
+
+#----------------------------------------------------------------------------
+
+_examples = '''examples:
+
+  # Generate ffhq uncurated images (matches paper Figure 12)
+  python %(prog)s generate-images --network=gdrive:networks/stylegan2-ffhq-config-f.pkl --seeds=6600-6625 --truncation-psi=0.5
+
+  # Generate ffhq curated images (matches paper Figure 11)
+  python %(prog)s generate-images --network=gdrive:networks/stylegan2-ffhq-config-f.pkl --seeds=66,230,389,1518 --truncation-psi=1.0
+
+  # Generate uncurated car images (matches paper Figure 12)
+  python %(prog)s generate-images --network=gdrive:networks/stylegan2-car-config-f.pkl --seeds=6000-6025 --truncation-psi=0.5
+
+  # Generate style mixing example (matches style mixing video clip)
+  python %(prog)s style-mixing-example --network=gdrive:networks/stylegan2-ffhq-config-f.pkl --row-seeds=85,100,75,458,1500 --col-seeds=55,821,1789,293 --truncation-psi=1.0
+'''
+
+#----------------------------------------------------------------------------
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='''StyleGAN2 generator.
+
+Run 'python %(prog)s <subcommand> --help' for subcommand help.''',
+        epilog=_examples,
+        formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+
+    subparsers = parser.add_subparsers(help='Sub-commands', dest='command')
+
+    parser_generate_images = subparsers.add_parser('generate-images', help='Generate images')
+    parser_generate_images.add_argument('--network', help='Network pickle filename', dest='network_pkl', required=True)
+    parser_generate_images.add_argument('--seeds', type=_parse_num_range, help='List of random seeds', required=True)
+    parser_generate_images.add_argument('--truncation-psi', type=float, help='Truncation psi (default: %(default)s)', default=0.5)
+    parser_generate_images.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+
+    parser_style_mixing_example = subparsers.add_parser('style-mixing-example', help='Generate style mixing video')
+    parser_style_mixing_example.add_argument('--network', help='Network pickle filename', dest='network_pkl', required=True)
+    parser_style_mixing_example.add_argument('--row-seeds', type=_parse_num_range, help='Random seeds to use for image rows', required=True)
+    parser_style_mixing_example.add_argument('--col-seeds', type=_parse_num_range, help='Random seeds to use for image columns', required=True)
+    parser_style_mixing_example.add_argument('--col-styles', type=_parse_num_range, help='Style layer range (default: %(default)s)', default='0-6')
+    parser_style_mixing_example.add_argument('--truncation-psi', type=float, help='Truncation psi (default: %(default)s)', default=0.5)
+    parser_style_mixing_example.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    subcmd = kwargs.pop('command')
+
+    if subcmd is None:
+        print ('Error: missing subcommand.  Re-run with --help for usage.')
+        sys.exit(1)
+
+    sc = dnnlib.SubmitConfig()
+    sc.num_gpus = 1
+    sc.submit_target = dnnlib.SubmitTarget.LOCAL
+    sc.local.do_not_copy_source_files = True
+    sc.run_dir_root = kwargs.pop('result_dir')
+    sc.run_desc = subcmd
+
+    func_name_map = {
+        'generate-images': 'run_generator.generate_images',
+        'style-mixing-example': 'run_generator.style_mixing_example'
+    }
+    dnnlib.submit_run(sc, func_name_map[subcmd], **kwargs)
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    main()
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/run_metrics.py b/reconstruction/ostec/external/stylegan2/run_metrics.py
new file mode 100644
index 000000000..5043b100f
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/run_metrics.py
@@ -0,0 +1,86 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+import argparse
+import os
+import sys
+
+import dnnlib
+import dnnlib.tflib as tflib
+
+import pretrained_networks
+from metrics import metric_base
+from metrics.metric_defaults import metric_defaults
+
+#----------------------------------------------------------------------------
+
+def run(network_pkl, metrics, dataset, data_dir, mirror_augment):
+    print('Evaluating metrics "%s" for "%s"...' % (','.join(metrics), network_pkl))
+    tflib.init_tf()
+    network_pkl = pretrained_networks.get_path_or_url(network_pkl)
+    dataset_args = dnnlib.EasyDict(tfrecord_dir=dataset, shuffle_mb=0)
+    num_gpus = dnnlib.submit_config.num_gpus
+    metric_group = metric_base.MetricGroup([metric_defaults[metric] for metric in metrics])
+    metric_group.run(network_pkl, data_dir=data_dir, dataset_args=dataset_args, mirror_augment=mirror_augment, num_gpus=num_gpus)
+
+#----------------------------------------------------------------------------
+
+def _str_to_bool(v):
+    if isinstance(v, bool):
+        return v
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+#----------------------------------------------------------------------------
+
+_examples = '''examples:
+
+  python %(prog)s --data-dir=~/datasets --network=gdrive:networks/stylegan2-ffhq-config-f.pkl --metrics=fid50k,ppl_wend --dataset=ffhq --mirror-augment=true
+
+valid metrics:
+
+  ''' + ', '.join(sorted([x for x in metric_defaults.keys()])) + '''
+'''
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='Run StyleGAN2 metrics.',
+        epilog=_examples,
+        formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+    parser.add_argument('--network', help='Network pickle filename', dest='network_pkl', required=True)
+    parser.add_argument('--metrics', help='Metrics to compute (default: %(default)s)', default='fid50k', type=lambda x: x.split(','))
+    parser.add_argument('--dataset', help='Training dataset', required=True)
+    parser.add_argument('--data-dir', help='Dataset root directory', required=True)
+    parser.add_argument('--mirror-augment', help='Mirror augment (default: %(default)s)', default=False, type=_str_to_bool, metavar='BOOL')
+    parser.add_argument('--num-gpus', help='Number of GPUs to use', type=int, default=1, metavar='N')
+
+    args = parser.parse_args()
+
+    if not os.path.exists(args.data_dir):
+        print ('Error: dataset root directory does not exist.')
+        sys.exit(1)
+
+    kwargs = vars(args)
+    sc = dnnlib.SubmitConfig()
+    sc.num_gpus = kwargs.pop('num_gpus')
+    sc.submit_target = dnnlib.SubmitTarget.LOCAL
+    sc.local.do_not_copy_source_files = True
+    sc.run_dir_root = kwargs.pop('result_dir')
+    sc.run_desc = 'run-metrics'
+    dnnlib.submit_run(sc, 'run_metrics.run', **kwargs)
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    main()
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/run_projector.py b/reconstruction/ostec/external/stylegan2/run_projector.py
new file mode 100644
index 000000000..a2f7a6800
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/run_projector.py
@@ -0,0 +1,148 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+import argparse
+import numpy as np
+import dnnlib
+import dnnlib.tflib as tflib
+import re
+import sys
+
+import projector
+import pretrained_networks
+from training import dataset
+from training import misc
+
+#----------------------------------------------------------------------------
+
+def project_image(proj, targets, png_prefix, num_snapshots):
+    snapshot_steps = set(proj.num_steps - np.linspace(0, proj.num_steps, num_snapshots, endpoint=False, dtype=int))
+    misc.save_image_grid(targets, png_prefix + 'target.png', drange=[-1, 1])
+    proj.start(targets)
+    while proj.get_cur_step() < proj.num_steps:
+        print('\r%d / %d ... ' % (proj.get_cur_step(), proj.num_steps), end='', flush=True)
+        proj.step()
+        if proj.get_cur_step() in snapshot_steps:
+            misc.save_image_grid(proj.get_images(), png_prefix + 'step%04d.png' % proj.get_cur_step(), drange=[-1, 1])
+    print('\r%-30s\r' % '', end='', flush=True)
+
+#----------------------------------------------------------------------------
+
+def project_generated_images(network_pkl, seeds, num_snapshots, truncation_psi):
+    print('Loading networks from "%s"...' % network_pkl)
+    _G, _D, Gs = pretrained_networks.load_networks(network_pkl)
+    proj = projector.Projector()
+    proj.set_network(Gs)
+    noise_vars = [var for name, var in Gs.components.synthesis.vars.items() if name.startswith('noise')]
+
+    Gs_kwargs = dnnlib.EasyDict()
+    Gs_kwargs.randomize_noise = False
+    Gs_kwargs.truncation_psi = truncation_psi
+
+    for seed_idx, seed in enumerate(seeds):
+        print('Projecting seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
+        rnd = np.random.RandomState(seed)
+        z = rnd.randn(1, *Gs.input_shape[1:])
+        tflib.set_vars({var: rnd.randn(*var.shape.as_list()) for var in noise_vars})
+        images = Gs.run(z, None, **Gs_kwargs)
+        project_image(proj, targets=images, png_prefix=dnnlib.make_run_dir_path('seed%04d-' % seed), num_snapshots=num_snapshots)
+
+#----------------------------------------------------------------------------
+
+def project_real_images(network_pkl, dataset_name, data_dir, num_images, num_snapshots):
+    print('Loading networks from "%s"...' % network_pkl)
+    _G, _D, Gs = pretrained_networks.load_networks(network_pkl)
+    proj = projector.Projector()
+    proj.set_network(Gs)
+
+    print('Loading images from "%s"...' % dataset_name)
+    dataset_obj = dataset.load_dataset(data_dir=data_dir, tfrecord_dir=dataset_name, max_label_size=0, repeat=False, shuffle_mb=0)
+    assert dataset_obj.shape == Gs.output_shape[1:]
+
+    for image_idx in range(num_images):
+        print('Projecting image %d/%d ...' % (image_idx, num_images))
+        images, _labels = dataset_obj.get_minibatch_np(1)
+        images = misc.adjust_dynamic_range(images, [0, 255], [-1, 1])
+        project_image(proj, targets=images, png_prefix=dnnlib.make_run_dir_path('image%04d-' % image_idx), num_snapshots=num_snapshots)
+
+#----------------------------------------------------------------------------
+
+def _parse_num_range(s):
+    '''Accept either a comma separated list of numbers 'a,b,c' or a range 'a-c' and return as a list of ints.'''
+
+    range_re = re.compile(r'^(\d+)-(\d+)$')
+    m = range_re.match(s)
+    if m:
+        return range(int(m.group(1)), int(m.group(2))+1)
+    vals = s.split(',')
+    return [int(x) for x in vals]
+
+#----------------------------------------------------------------------------
+
+_examples = '''examples:
+
+  # Project generated images
+  python %(prog)s project-generated-images --network=gdrive:networks/stylegan2-car-config-f.pkl --seeds=0,1,5
+
+  # Project real images
+  python %(prog)s project-real-images --network=gdrive:networks/stylegan2-car-config-f.pkl --dataset=car --data-dir=~/datasets
+
+'''
+
+#----------------------------------------------------------------------------
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='''StyleGAN2 projector.
+
+Run 'python %(prog)s <subcommand> --help' for subcommand help.''',
+        epilog=_examples,
+        formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+
+    subparsers = parser.add_subparsers(help='Sub-commands', dest='command')
+
+    project_generated_images_parser = subparsers.add_parser('project-generated-images', help='Project generated images')
+    project_generated_images_parser.add_argument('--network', help='Network pickle filename', dest='network_pkl', required=True)
+    project_generated_images_parser.add_argument('--seeds', type=_parse_num_range, help='List of random seeds', default=range(3))
+    project_generated_images_parser.add_argument('--num-snapshots', type=int, help='Number of snapshots (default: %(default)s)', default=5)
+    project_generated_images_parser.add_argument('--truncation-psi', type=float, help='Truncation psi (default: %(default)s)', default=1.0)
+    project_generated_images_parser.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+
+    project_real_images_parser = subparsers.add_parser('project-real-images', help='Project real images')
+    project_real_images_parser.add_argument('--network', help='Network pickle filename', dest='network_pkl', required=True)
+    project_real_images_parser.add_argument('--data-dir', help='Dataset root directory', required=True)
+    project_real_images_parser.add_argument('--dataset', help='Training dataset', dest='dataset_name', required=True)
+    project_real_images_parser.add_argument('--num-snapshots', type=int, help='Number of snapshots (default: %(default)s)', default=5)
+    project_real_images_parser.add_argument('--num-images', type=int, help='Number of images to project (default: %(default)s)', default=3)
+    project_real_images_parser.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+
+    args = parser.parse_args()
+    subcmd = args.command
+    if subcmd is None:
+        print ('Error: missing subcommand.  Re-run with --help for usage.')
+        sys.exit(1)
+
+    kwargs = vars(args)
+    sc = dnnlib.SubmitConfig()
+    sc.num_gpus = 1
+    sc.submit_target = dnnlib.SubmitTarget.LOCAL
+    sc.local.do_not_copy_source_files = True
+    sc.run_dir_root = kwargs.pop('result_dir')
+    sc.run_desc = kwargs.pop('command')
+
+    func_name_map = {
+        'project-generated-images': 'run_projector.project_generated_images',
+        'project-real-images': 'run_projector.project_real_images'
+    }
+    dnnlib.submit_run(sc, func_name_map[subcmd], **kwargs)
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    main()
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/run_training.py b/reconstruction/ostec/external/stylegan2/run_training.py
new file mode 100644
index 000000000..bc4c0a2bd
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/run_training.py
@@ -0,0 +1,195 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+import argparse
+import copy
+import os
+import sys
+
+import dnnlib
+from dnnlib import EasyDict
+
+from metrics.metric_defaults import metric_defaults
+
+#----------------------------------------------------------------------------
+
+_valid_configs = [
+    # Table 1
+    'config-a', # Baseline StyleGAN
+    'config-b', # + Weight demodulation
+    'config-c', # + Lazy regularization
+    'config-d', # + Path length regularization
+    'config-e', # + No growing, new G & D arch.
+    'config-f', # + Large networks (default)
+
+    # Table 2
+    'config-e-Gorig-Dorig',   'config-e-Gorig-Dresnet',   'config-e-Gorig-Dskip',
+    'config-e-Gresnet-Dorig', 'config-e-Gresnet-Dresnet', 'config-e-Gresnet-Dskip',
+    'config-e-Gskip-Dorig',   'config-e-Gskip-Dresnet',   'config-e-Gskip-Dskip',
+]
+
+#----------------------------------------------------------------------------
+
+def run(dataset, data_dir, result_dir, config_id, num_gpus, total_kimg, gamma, mirror_augment, metrics):
+    train     = EasyDict(run_func_name='training.training_loop.training_loop') # Options for training loop.
+    G         = EasyDict(func_name='training.networks_stylegan2.G_main')       # Options for generator network.
+    D         = EasyDict(func_name='training.networks_stylegan2.D_stylegan2')  # Options for discriminator network.
+    G_opt     = EasyDict(beta1=0.0, beta2=0.99, epsilon=1e-8)                  # Options for generator optimizer.
+    D_opt     = EasyDict(beta1=0.0, beta2=0.99, epsilon=1e-8)                  # Options for discriminator optimizer.
+    G_loss    = EasyDict(func_name='training.loss.G_logistic_ns_pathreg')      # Options for generator loss.
+    D_loss    = EasyDict(func_name='training.loss.D_logistic_r1')              # Options for discriminator loss.
+    sched     = EasyDict()                                                     # Options for TrainingSchedule.
+    grid      = EasyDict(size='8k', layout='random')                           # Options for setup_snapshot_image_grid().
+    sc        = dnnlib.SubmitConfig()                                          # Options for dnnlib.submit_run().
+    tf_config = {'rnd.np_random_seed': 1000}                                   # Options for tflib.init_tf().
+
+    train.data_dir = data_dir
+    train.total_kimg = total_kimg
+    train.mirror_augment = mirror_augment
+    train.image_snapshot_ticks = train.network_snapshot_ticks = 10
+    sched.G_lrate_base = sched.D_lrate_base = 0.002
+    sched.minibatch_size_base = 32
+    sched.minibatch_gpu_base = 4
+    D_loss.gamma = 10
+    metrics = [metric_defaults[x] for x in metrics]
+    desc = 'stylegan2'
+
+    desc += '-' + dataset
+    dataset_args = EasyDict(tfrecord_dir=dataset)
+
+    assert num_gpus in [1, 2, 4, 8]
+    sc.num_gpus = num_gpus
+    desc += '-%dgpu' % num_gpus
+
+    assert config_id in _valid_configs
+    desc += '-' + config_id
+
+    # Configs A-E: Shrink networks to match original StyleGAN.
+    if config_id != 'config-f':
+        G.fmap_base = D.fmap_base = 8 << 10
+
+    # Config E: Set gamma to 100 and override G & D architecture.
+    if config_id.startswith('config-e'):
+        D_loss.gamma = 100
+        if 'Gorig'   in config_id: G.architecture = 'orig'
+        if 'Gskip'   in config_id: G.architecture = 'skip' # (default)
+        if 'Gresnet' in config_id: G.architecture = 'resnet'
+        if 'Dorig'   in config_id: D.architecture = 'orig'
+        if 'Dskip'   in config_id: D.architecture = 'skip'
+        if 'Dresnet' in config_id: D.architecture = 'resnet' # (default)
+
+    # Configs A-D: Enable progressive growing and switch to networks that support it.
+    if config_id in ['config-a', 'config-b', 'config-c', 'config-d']:
+        sched.lod_initial_resolution = 8
+        sched.G_lrate_base = sched.D_lrate_base = 0.001
+        sched.G_lrate_dict = sched.D_lrate_dict = {128: 0.0015, 256: 0.002, 512: 0.003, 1024: 0.003}
+        sched.minibatch_size_base = 32 # (default)
+        sched.minibatch_size_dict = {8: 256, 16: 128, 32: 64, 64: 32}
+        sched.minibatch_gpu_base = 4 # (default)
+        sched.minibatch_gpu_dict = {8: 32, 16: 16, 32: 8, 64: 4}
+        G.synthesis_func = 'G_synthesis_stylegan_revised'
+        D.func_name = 'training.networks_stylegan2.D_stylegan'
+
+    # Configs A-C: Disable path length regularization.
+    if config_id in ['config-a', 'config-b', 'config-c']:
+        G_loss = EasyDict(func_name='training.loss.G_logistic_ns')
+
+    # Configs A-B: Disable lazy regularization.
+    if config_id in ['config-a', 'config-b']:
+        train.lazy_regularization = False
+
+    # Config A: Switch to original StyleGAN networks.
+    if config_id == 'config-a':
+        G = EasyDict(func_name='training.networks_stylegan.G_style')
+        D = EasyDict(func_name='training.networks_stylegan.D_basic')
+
+    if gamma is not None:
+        D_loss.gamma = gamma
+
+    sc.submit_target = dnnlib.SubmitTarget.LOCAL
+    sc.local.do_not_copy_source_files = True
+    kwargs = EasyDict(train)
+    kwargs.update(G_args=G, D_args=D, G_opt_args=G_opt, D_opt_args=D_opt, G_loss_args=G_loss, D_loss_args=D_loss)
+    kwargs.update(dataset_args=dataset_args, sched_args=sched, grid_args=grid, metric_arg_list=metrics, tf_config=tf_config)
+    kwargs.submit_config = copy.deepcopy(sc)
+    kwargs.submit_config.run_dir_root = result_dir
+    kwargs.submit_config.run_desc = desc
+    dnnlib.submit_run(**kwargs)
+
+#----------------------------------------------------------------------------
+
+def _str_to_bool(v):
+    if isinstance(v, bool):
+        return v
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+def _parse_comma_sep(s):
+    if s is None or s.lower() == 'none' or s == '':
+        return []
+    return s.split(',')
+
+#----------------------------------------------------------------------------
+
+_examples = '''examples:
+
+  # Train StyleGAN2 using the FFHQ dataset
+  python %(prog)s --num-gpus=8 --data-dir=~/datasets --config=config-f --dataset=ffhq --mirror-augment=true
+
+valid configs:
+
+  ''' + ', '.join(_valid_configs) + '''
+
+valid metrics:
+
+  ''' + ', '.join(sorted([x for x in metric_defaults.keys()])) + '''
+
+'''
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='Train StyleGAN2.',
+        epilog=_examples,
+        formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument('--result-dir', help='Root directory for run results (default: %(default)s)', default='results', metavar='DIR')
+    parser.add_argument('--data-dir', help='Dataset root directory', required=True)
+    parser.add_argument('--dataset', help='Training dataset', required=True)
+    parser.add_argument('--config', help='Training config (default: %(default)s)', default='config-f', required=True, dest='config_id', metavar='CONFIG')
+    parser.add_argument('--num-gpus', help='Number of GPUs (default: %(default)s)', default=1, type=int, metavar='N')
+    parser.add_argument('--total-kimg', help='Training length in thousands of images (default: %(default)s)', metavar='KIMG', default=25000, type=int)
+    parser.add_argument('--gamma', help='R1 regularization weight (default is config dependent)', default=None, type=float)
+    parser.add_argument('--mirror-augment', help='Mirror augment (default: %(default)s)', default=False, metavar='BOOL', type=_str_to_bool)
+    parser.add_argument('--metrics', help='Comma-separated list of metrics or "none" (default: %(default)s)', default='fid50k', type=_parse_comma_sep)
+
+    args = parser.parse_args()
+
+    if not os.path.exists(args.data_dir):
+        print ('Error: dataset root directory does not exist.')
+        sys.exit(1)
+
+    if args.config_id not in _valid_configs:
+        print ('Error: --config value must be one of: ', ', '.join(_valid_configs))
+        sys.exit(1)
+
+    for metric in args.metrics:
+        if metric not in metric_defaults:
+            print ('Error: unknown metric \'%s\'' % metric)
+            sys.exit(1)
+
+    run(**vars(args))
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    main()
+
+#----------------------------------------------------------------------------
+
diff --git a/reconstruction/ostec/external/stylegan2/test_nvcc.cu b/reconstruction/ostec/external/stylegan2/test_nvcc.cu
new file mode 100644
index 000000000..8b09bbfe0
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/test_nvcc.cu
@@ -0,0 +1,29 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#include <cstdio>
+
+void checkCudaError(cudaError_t err)
+{
+    if (err != cudaSuccess)
+    {
+        printf("%s: %s\n", cudaGetErrorName(err), cudaGetErrorString(err));
+        exit(1);
+    }
+}
+
+__global__ void cudaKernel(void)
+{
+    printf("GPU says hello.\n");
+}
+
+int main(void)
+{
+    printf("CPU says hello.\n");
+    checkCudaError(cudaLaunchKernel((void*)cudaKernel, 1, 1, NULL, 0, NULL));
+    checkCudaError(cudaDeviceSynchronize());
+    return 0;
+}
diff --git a/reconstruction/ostec/external/stylegan2/training/__init__.py b/reconstruction/ostec/external/stylegan2/training/__init__.py
new file mode 100644
index 000000000..9ab9908ef
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/__init__.py
@@ -0,0 +1,7 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+# empty
diff --git a/reconstruction/ostec/external/stylegan2/training/dataset.py b/reconstruction/ostec/external/stylegan2/training/dataset.py
new file mode 100644
index 000000000..2d1059838
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/dataset.py
@@ -0,0 +1,199 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Multi-resolution input data pipeline."""
+
+import os
+import glob
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+
+#----------------------------------------------------------------------------
+# Dataset class that loads data from tfrecords files.
+
+class TFRecordDataset:
+    def __init__(self,
+        tfrecord_dir,               # Directory containing a collection of tfrecords files.
+        resolution      = None,     # Dataset resolution, None = autodetect.
+        label_file      = None,     # Relative path of the labels file, None = autodetect.
+        max_label_size  = 0,        # 0 = no labels, 'full' = full labels, <int> = N first label components.
+        max_images      = None,     # Maximum number of images to use, None = use all images.
+        repeat          = True,     # Repeat dataset indefinitely?
+        shuffle_mb      = 4096,     # Shuffle data within specified window (megabytes), 0 = disable shuffling.
+        prefetch_mb     = 2048,     # Amount of data to prefetch (megabytes), 0 = disable prefetching.
+        buffer_mb       = 256,      # Read buffer size (megabytes).
+        num_threads     = 2):       # Number of concurrent threads.
+
+        self.tfrecord_dir       = tfrecord_dir
+        self.resolution         = None
+        self.resolution_log2    = None
+        self.shape              = []        # [channels, height, width]
+        self.dtype              = 'uint8'
+        self.dynamic_range      = [0, 255]
+        self.label_file         = label_file
+        self.label_size         = None      # components
+        self.label_dtype        = None
+        self._np_labels         = None
+        self._tf_minibatch_in   = None
+        self._tf_labels_var     = None
+        self._tf_labels_dataset = None
+        self._tf_datasets       = dict()
+        self._tf_iterator       = None
+        self._tf_init_ops       = dict()
+        self._tf_minibatch_np   = None
+        self._cur_minibatch     = -1
+        self._cur_lod           = -1
+
+        # List tfrecords files and inspect their shapes.
+        assert os.path.isdir(self.tfrecord_dir)
+        tfr_files = sorted(glob.glob(os.path.join(self.tfrecord_dir, '*.tfrecords')))
+        assert len(tfr_files) >= 1
+        tfr_shapes = []
+        for tfr_file in tfr_files:
+            tfr_opt = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.NONE)
+            for record in tf.python_io.tf_record_iterator(tfr_file, tfr_opt):
+                tfr_shapes.append(self.parse_tfrecord_np(record).shape)
+                break
+
+        # Autodetect label filename.
+        if self.label_file is None:
+            guess = sorted(glob.glob(os.path.join(self.tfrecord_dir, '*.labels')))
+            if len(guess):
+                self.label_file = guess[0]
+        elif not os.path.isfile(self.label_file):
+            guess = os.path.join(self.tfrecord_dir, self.label_file)
+            if os.path.isfile(guess):
+                self.label_file = guess
+
+        # Determine shape and resolution.
+        max_shape = max(tfr_shapes, key=np.prod)
+        self.resolution = resolution if resolution is not None else max_shape[1]
+        self.resolution_log2 = int(np.log2(self.resolution))
+        self.shape = [max_shape[0], self.resolution, self.resolution]
+        tfr_lods = [self.resolution_log2 - int(np.log2(shape[1])) for shape in tfr_shapes]
+        assert all(shape[0] == max_shape[0] for shape in tfr_shapes)
+        assert all(shape[1] == shape[2] for shape in tfr_shapes)
+        assert all(shape[1] == self.resolution // (2**lod) for shape, lod in zip(tfr_shapes, tfr_lods))
+        assert all(lod in tfr_lods for lod in range(self.resolution_log2 - 1))
+
+        # Load labels.
+        assert max_label_size == 'full' or max_label_size >= 0
+        self._np_labels = np.zeros([1<<30, 0], dtype=np.float32)
+        if self.label_file is not None and max_label_size != 0:
+            self._np_labels = np.load(self.label_file)
+            assert self._np_labels.ndim == 2
+        if max_label_size != 'full' and self._np_labels.shape[1] > max_label_size:
+            self._np_labels = self._np_labels[:, :max_label_size]
+        if max_images is not None and self._np_labels.shape[0] > max_images:
+            self._np_labels = self._np_labels[:max_images]
+        self.label_size = self._np_labels.shape[1]
+        self.label_dtype = self._np_labels.dtype.name
+
+        # Build TF expressions.
+        with tf.name_scope('Dataset'), tf.device('/cpu:0'):
+            self._tf_minibatch_in = tf.placeholder(tf.int64, name='minibatch_in', shape=[])
+            self._tf_labels_var = tflib.create_var_with_large_initial_value(self._np_labels, name='labels_var')
+            self._tf_labels_dataset = tf.data.Dataset.from_tensor_slices(self._tf_labels_var)
+            for tfr_file, tfr_shape, tfr_lod in zip(tfr_files, tfr_shapes, tfr_lods):
+                if tfr_lod < 0:
+                    continue
+                dset = tf.data.TFRecordDataset(tfr_file, compression_type='', buffer_size=buffer_mb<<20)
+                if max_images is not None:
+                    dset = dset.take(max_images)
+                dset = dset.map(self.parse_tfrecord_tf, num_parallel_calls=num_threads)
+                dset = tf.data.Dataset.zip((dset, self._tf_labels_dataset))
+                bytes_per_item = np.prod(tfr_shape) * np.dtype(self.dtype).itemsize
+                if shuffle_mb > 0:
+                    dset = dset.shuffle(((shuffle_mb << 20) - 1) // bytes_per_item + 1)
+                if repeat:
+                    dset = dset.repeat()
+                if prefetch_mb > 0:
+                    dset = dset.prefetch(((prefetch_mb << 20) - 1) // bytes_per_item + 1)
+                dset = dset.batch(self._tf_minibatch_in)
+                self._tf_datasets[tfr_lod] = dset
+            self._tf_iterator = tf.data.Iterator.from_structure(self._tf_datasets[0].output_types, self._tf_datasets[0].output_shapes)
+            self._tf_init_ops = {lod: self._tf_iterator.make_initializer(dset) for lod, dset in self._tf_datasets.items()}
+
+    def close(self):
+        pass
+
+    # Use the given minibatch size and level-of-detail for the data returned by get_minibatch_tf().
+    def configure(self, minibatch_size, lod=0):
+        lod = int(np.floor(lod))
+        assert minibatch_size >= 1 and lod in self._tf_datasets
+        if self._cur_minibatch != minibatch_size or self._cur_lod != lod:
+            self._tf_init_ops[lod].run({self._tf_minibatch_in: minibatch_size})
+            self._cur_minibatch = minibatch_size
+            self._cur_lod = lod
+
+    # Get next minibatch as TensorFlow expressions.
+    def get_minibatch_tf(self): # => images, labels
+        return self._tf_iterator.get_next()
+
+    # Get next minibatch as NumPy arrays.
+    def get_minibatch_np(self, minibatch_size, lod=0): # => images, labels
+        self.configure(minibatch_size, lod)
+        with tf.name_scope('Dataset'):
+            if self._tf_minibatch_np is None:
+                self._tf_minibatch_np = self.get_minibatch_tf()
+            return tflib.run(self._tf_minibatch_np)
+
+    # Get random labels as TensorFlow expression.
+    def get_random_labels_tf(self, minibatch_size): # => labels
+        with tf.name_scope('Dataset'):
+            if self.label_size > 0:
+                with tf.device('/cpu:0'):
+                    return tf.gather(self._tf_labels_var, tf.random_uniform([minibatch_size], 0, self._np_labels.shape[0], dtype=tf.int32))
+            return tf.zeros([minibatch_size, 0], self.label_dtype)
+
+    # Get random labels as NumPy array.
+    def get_random_labels_np(self, minibatch_size): # => labels
+        if self.label_size > 0:
+            return self._np_labels[np.random.randint(self._np_labels.shape[0], size=[minibatch_size])]
+        return np.zeros([minibatch_size, 0], self.label_dtype)
+
+    # Parse individual image from a tfrecords file into TensorFlow expression.
+    @staticmethod
+    def parse_tfrecord_tf(record):
+        features = tf.parse_single_example(record, features={
+            'shape': tf.FixedLenFeature([3], tf.int64),
+            'data': tf.FixedLenFeature([], tf.string)})
+        data = tf.decode_raw(features['data'], tf.uint8)
+        return tf.reshape(data, features['shape'])
+
+    # Parse individual image from a tfrecords file into NumPy array.
+    @staticmethod
+    def parse_tfrecord_np(record):
+        ex = tf.train.Example()
+        ex.ParseFromString(record)
+        shape = ex.features.feature['shape'].int64_list.value # pylint: disable=no-member
+        data = ex.features.feature['data'].bytes_list.value[0] # pylint: disable=no-member
+        return np.fromstring(data, np.uint8).reshape(shape)
+
+#----------------------------------------------------------------------------
+# Helper func for constructing a dataset object using the given options.
+
+def load_dataset(class_name=None, data_dir=None, verbose=False, **kwargs):
+    kwargs = dict(kwargs)
+    if 'tfrecord_dir' in kwargs:
+        if class_name is None:
+            class_name = __name__ + '.TFRecordDataset'
+        if data_dir is not None:
+            kwargs['tfrecord_dir'] = os.path.join(data_dir, kwargs['tfrecord_dir'])
+
+    assert class_name is not None
+    if verbose:
+        print('Streaming data using %s...' % class_name)
+    dataset = dnnlib.util.get_obj_by_name(class_name)(**kwargs)
+    if verbose:
+        print('Dataset shape =', np.int32(dataset.shape).tolist())
+        print('Dynamic range =', dataset.dynamic_range)
+        print('Label size    =', dataset.label_size)
+    return dataset
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/training/loss.py b/reconstruction/ostec/external/stylegan2/training/loss.py
new file mode 100644
index 000000000..7ad2fe16f
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/loss.py
@@ -0,0 +1,197 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Loss functions."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib.tflib as tflib
+from dnnlib.tflib.autosummary import autosummary
+
+#----------------------------------------------------------------------------
+# Logistic loss from the paper
+# "Generative Adversarial Nets", Goodfellow et al. 2014
+
+def G_logistic(G, D, opt, training_set, minibatch_size):
+    _ = opt
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    labels = training_set.get_random_labels_tf(minibatch_size)
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    loss = -tf.nn.softplus(fake_scores_out) # log(1-sigmoid(fake_scores_out)) # pylint: disable=invalid-unary-operand-type
+    return loss, None
+
+def G_logistic_ns(G, D, opt, training_set, minibatch_size):
+    _ = opt
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    labels = training_set.get_random_labels_tf(minibatch_size)
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    loss = tf.nn.softplus(-fake_scores_out) # -log(sigmoid(fake_scores_out))
+    return loss, None
+
+def D_logistic(G, D, opt, training_set, minibatch_size, reals, labels):
+    _ = opt, training_set
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    real_scores_out = D.get_output_for(reals, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    real_scores_out = autosummary('Loss/scores/real', real_scores_out)
+    fake_scores_out = autosummary('Loss/scores/fake', fake_scores_out)
+    loss = tf.nn.softplus(fake_scores_out) # -log(1-sigmoid(fake_scores_out))
+    loss += tf.nn.softplus(-real_scores_out) # -log(sigmoid(real_scores_out)) # pylint: disable=invalid-unary-operand-type
+    return loss, None
+
+#----------------------------------------------------------------------------
+# R1 and R2 regularizers from the paper
+# "Which Training Methods for GANs do actually Converge?", Mescheder et al. 2018
+
+def D_logistic_r1(G, D, opt, training_set, minibatch_size, reals, labels, gamma=10.0):
+    _ = opt, training_set
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    real_scores_out = D.get_output_for(reals, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    real_scores_out = autosummary('Loss/scores/real', real_scores_out)
+    fake_scores_out = autosummary('Loss/scores/fake', fake_scores_out)
+    loss = tf.nn.softplus(fake_scores_out) # -log(1-sigmoid(fake_scores_out))
+    loss += tf.nn.softplus(-real_scores_out) # -log(sigmoid(real_scores_out)) # pylint: disable=invalid-unary-operand-type
+
+    with tf.name_scope('GradientPenalty'):
+        real_grads = tf.gradients(tf.reduce_sum(real_scores_out), [reals])[0]
+        gradient_penalty = tf.reduce_sum(tf.square(real_grads), axis=[1,2,3])
+        gradient_penalty = autosummary('Loss/gradient_penalty', gradient_penalty)
+        reg = gradient_penalty * (gamma * 0.5)
+    return loss, reg
+
+def D_logistic_r2(G, D, opt, training_set, minibatch_size, reals, labels, gamma=10.0):
+    _ = opt, training_set
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    real_scores_out = D.get_output_for(reals, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    real_scores_out = autosummary('Loss/scores/real', real_scores_out)
+    fake_scores_out = autosummary('Loss/scores/fake', fake_scores_out)
+    loss = tf.nn.softplus(fake_scores_out) # -log(1-sigmoid(fake_scores_out))
+    loss += tf.nn.softplus(-real_scores_out) # -log(sigmoid(real_scores_out)) # pylint: disable=invalid-unary-operand-type
+
+    with tf.name_scope('GradientPenalty'):
+        fake_grads = tf.gradients(tf.reduce_sum(fake_scores_out), [fake_images_out])[0]
+        gradient_penalty = tf.reduce_sum(tf.square(fake_grads), axis=[1,2,3])
+        gradient_penalty = autosummary('Loss/gradient_penalty', gradient_penalty)
+        reg = gradient_penalty * (gamma * 0.5)
+    return loss, reg
+
+#----------------------------------------------------------------------------
+# WGAN loss from the paper
+# "Wasserstein Generative Adversarial Networks", Arjovsky et al. 2017
+
+def G_wgan(G, D, opt, training_set, minibatch_size):
+    _ = opt
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    labels = training_set.get_random_labels_tf(minibatch_size)
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    loss = -fake_scores_out
+    return loss, None
+
+def D_wgan(G, D, opt, training_set, minibatch_size, reals, labels, wgan_epsilon=0.001):
+    _ = opt, training_set
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    real_scores_out = D.get_output_for(reals, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    real_scores_out = autosummary('Loss/scores/real', real_scores_out)
+    fake_scores_out = autosummary('Loss/scores/fake', fake_scores_out)
+    loss = fake_scores_out - real_scores_out
+    with tf.name_scope('EpsilonPenalty'):
+        epsilon_penalty = autosummary('Loss/epsilon_penalty', tf.square(real_scores_out))
+        loss += epsilon_penalty * wgan_epsilon
+    return loss, None
+
+#----------------------------------------------------------------------------
+# WGAN-GP loss from the paper
+# "Improved Training of Wasserstein GANs", Gulrajani et al. 2017
+
+def D_wgan_gp(G, D, opt, training_set, minibatch_size, reals, labels, wgan_lambda=10.0, wgan_epsilon=0.001, wgan_target=1.0):
+    _ = opt, training_set
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    fake_images_out = G.get_output_for(latents, labels, is_training=True)
+    real_scores_out = D.get_output_for(reals, labels, is_training=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    real_scores_out = autosummary('Loss/scores/real', real_scores_out)
+    fake_scores_out = autosummary('Loss/scores/fake', fake_scores_out)
+    loss = fake_scores_out - real_scores_out
+    with tf.name_scope('EpsilonPenalty'):
+        epsilon_penalty = autosummary('Loss/epsilon_penalty', tf.square(real_scores_out))
+    loss += epsilon_penalty * wgan_epsilon
+
+    with tf.name_scope('GradientPenalty'):
+        mixing_factors = tf.random_uniform([minibatch_size, 1, 1, 1], 0.0, 1.0, dtype=fake_images_out.dtype)
+        mixed_images_out = tflib.lerp(tf.cast(reals, fake_images_out.dtype), fake_images_out, mixing_factors)
+        mixed_scores_out = D.get_output_for(mixed_images_out, labels, is_training=True)
+        mixed_scores_out = autosummary('Loss/scores/mixed', mixed_scores_out)
+        mixed_grads = tf.gradients(tf.reduce_sum(mixed_scores_out), [mixed_images_out])[0]
+        mixed_norms = tf.sqrt(tf.reduce_sum(tf.square(mixed_grads), axis=[1,2,3]))
+        mixed_norms = autosummary('Loss/mixed_norms', mixed_norms)
+        gradient_penalty = tf.square(mixed_norms - wgan_target)
+        reg = gradient_penalty * (wgan_lambda / (wgan_target**2))
+    return loss, reg
+
+#----------------------------------------------------------------------------
+# Non-saturating logistic loss with path length regularizer from the paper
+# "Analyzing and Improving the Image Quality of StyleGAN", Karras et al. 2019
+
+def G_logistic_ns_pathreg(G, D, opt, training_set, minibatch_size, pl_minibatch_shrink=2, pl_decay=0.01, pl_weight=2.0):
+    _ = opt
+    latents = tf.random_normal([minibatch_size] + G.input_shapes[0][1:])
+    labels = training_set.get_random_labels_tf(minibatch_size)
+    fake_images_out, fake_dlatents_out = G.get_output_for(latents, labels, is_training=True, return_dlatents=True)
+    fake_scores_out = D.get_output_for(fake_images_out, labels, is_training=True)
+    loss = tf.nn.softplus(-fake_scores_out) # -log(sigmoid(fake_scores_out))
+
+    # Path length regularization.
+    with tf.name_scope('PathReg'):
+
+        # Evaluate the regularization term using a smaller minibatch to conserve memory.
+        if pl_minibatch_shrink > 1:
+            pl_minibatch = minibatch_size // pl_minibatch_shrink
+            pl_latents = tf.random_normal([pl_minibatch] + G.input_shapes[0][1:])
+            pl_labels = training_set.get_random_labels_tf(pl_minibatch)
+            fake_images_out, fake_dlatents_out = G.get_output_for(pl_latents, pl_labels, is_training=True, return_dlatents=True)
+
+        # Compute |J*y|.
+        pl_noise = tf.random_normal(tf.shape(fake_images_out)) / np.sqrt(np.prod(G.output_shape[2:]))
+        pl_grads = tf.gradients(tf.reduce_sum(fake_images_out * pl_noise), [fake_dlatents_out])[0]
+        pl_lengths = tf.sqrt(tf.reduce_mean(tf.reduce_sum(tf.square(pl_grads), axis=2), axis=1))
+        pl_lengths = autosummary('Loss/pl_lengths', pl_lengths)
+
+        # Track exponential moving average of |J*y|.
+        with tf.control_dependencies(None):
+            pl_mean_var = tf.Variable(name='pl_mean', trainable=False, initial_value=0.0, dtype=tf.float32)
+        pl_mean = pl_mean_var + pl_decay * (tf.reduce_mean(pl_lengths) - pl_mean_var)
+        pl_update = tf.assign(pl_mean_var, pl_mean)
+
+        # Calculate (|J*y|-a)^2.
+        with tf.control_dependencies([pl_update]):
+            pl_penalty = tf.square(pl_lengths - pl_mean)
+            pl_penalty = autosummary('Loss/pl_penalty', pl_penalty)
+
+        # Apply weight.
+        #
+        # Note: The division in pl_noise decreases the weight by num_pixels, and the reduce_mean
+        # in pl_lengths decreases it by num_affine_layers. The effective weight then becomes:
+        #
+        # gamma_pl = pl_weight / num_pixels / num_affine_layers
+        # = 2 / (r^2) / (log2(r) * 2 - 2)
+        # = 1 / (r^2 * (log2(r) - 1))
+        # = ln(2) / (r^2 * (ln(r) - ln(2))
+        #
+        reg = pl_penalty * pl_weight
+
+    return loss, reg
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/training/misc.py b/reconstruction/ostec/external/stylegan2/training/misc.py
new file mode 100644
index 000000000..9b3444e85
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/misc.py
@@ -0,0 +1,145 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Miscellaneous utility functions."""
+
+import os
+import pickle
+import numpy as np
+import PIL.Image
+import PIL.ImageFont
+import dnnlib
+
+#----------------------------------------------------------------------------
+# Convenience wrappers for pickle that are able to load data produced by
+# older versions of the code, and from external URLs.
+
+def open_file_or_url(file_or_url):
+    if dnnlib.util.is_url(file_or_url):
+        return dnnlib.util.open_url(file_or_url, cache_dir='.stylegan2-cache')
+    return open(file_or_url, 'rb')
+
+def load_pkl(file_or_url):
+    with open_file_or_url(file_or_url) as file:
+        return pickle.load(file, encoding='latin1')
+
+def save_pkl(obj, filename):
+    with open(filename, 'wb') as file:
+        pickle.dump(obj, file, protocol=pickle.HIGHEST_PROTOCOL)
+
+#----------------------------------------------------------------------------
+# Image utils.
+
+def adjust_dynamic_range(data, drange_in, drange_out):
+    if drange_in != drange_out:
+        scale = (np.float32(drange_out[1]) - np.float32(drange_out[0])) / (np.float32(drange_in[1]) - np.float32(drange_in[0]))
+        bias = (np.float32(drange_out[0]) - np.float32(drange_in[0]) * scale)
+        data = data * scale + bias
+    return data
+
+def create_image_grid(images, grid_size=None):
+    assert images.ndim == 3 or images.ndim == 4
+    num, img_w, img_h = images.shape[0], images.shape[-1], images.shape[-2]
+
+    if grid_size is not None:
+        grid_w, grid_h = tuple(grid_size)
+    else:
+        grid_w = max(int(np.ceil(np.sqrt(num))), 1)
+        grid_h = max((num - 1) // grid_w + 1, 1)
+
+    grid = np.zeros(list(images.shape[1:-2]) + [grid_h * img_h, grid_w * img_w], dtype=images.dtype)
+    for idx in range(num):
+        x = (idx % grid_w) * img_w
+        y = (idx // grid_w) * img_h
+        grid[..., y : y + img_h, x : x + img_w] = images[idx]
+    return grid
+
+def convert_to_pil_image(image, drange=[0,1]):
+    assert image.ndim == 2 or image.ndim == 3
+    if image.ndim == 3:
+        if image.shape[0] == 1:
+            image = image[0] # grayscale CHW => HW
+        else:
+            image = image.transpose(1, 2, 0) # CHW -> HWC
+
+    image = adjust_dynamic_range(image, drange, [0,255])
+    image = np.rint(image).clip(0, 255).astype(np.uint8)
+    fmt = 'RGB' if image.ndim == 3 else 'L'
+    return PIL.Image.fromarray(image, fmt)
+
+def save_image_grid(images, filename, drange=[0,1], grid_size=None):
+    convert_to_pil_image(create_image_grid(images, grid_size), drange).save(filename)
+
+def apply_mirror_augment(minibatch):
+    mask = np.random.rand(minibatch.shape[0]) < 0.5
+    minibatch = np.array(minibatch)
+    minibatch[mask] = minibatch[mask, :, :, ::-1]
+    return minibatch
+
+#----------------------------------------------------------------------------
+# Loading data from previous training runs.
+
+def parse_config_for_previous_run(run_dir):
+    with open(os.path.join(run_dir, 'submit_config.pkl'), 'rb') as f:
+        data = pickle.load(f)
+    data = data.get('run_func_kwargs', {})
+    return dict(train=data, dataset=data.get('dataset_args', {}))
+
+#----------------------------------------------------------------------------
+# Size and contents of the image snapshot grids that are exported
+# periodically during training.
+
+def setup_snapshot_image_grid(training_set,
+    size    = '1080p',      # '1080p' = to be viewed on 1080p display, '4k' = to be viewed on 4k display.
+    layout  = 'random'):    # 'random' = grid contents are selected randomly, 'row_per_class' = each row corresponds to one class label.
+
+    # Select size.
+    gw = 1; gh = 1
+    if size == '1080p':
+        gw = np.clip(1920 // training_set.shape[2], 3, 32)
+        gh = np.clip(1080 // training_set.shape[1], 2, 32)
+    if size == '4k':
+        gw = np.clip(3840 // training_set.shape[2], 7, 32)
+        gh = np.clip(2160 // training_set.shape[1], 4, 32)
+    if size == '8k':
+        gw = np.clip(7680 // training_set.shape[2], 7, 32)
+        gh = np.clip(4320 // training_set.shape[1], 4, 32)
+
+    # Initialize data arrays.
+    reals = np.zeros([gw * gh] + training_set.shape, dtype=training_set.dtype)
+    labels = np.zeros([gw * gh, training_set.label_size], dtype=training_set.label_dtype)
+
+    # Random layout.
+    if layout == 'random':
+        reals[:], labels[:] = training_set.get_minibatch_np(gw * gh)
+
+    # Class-conditional layouts.
+    class_layouts = dict(row_per_class=[gw,1], col_per_class=[1,gh], class4x4=[4,4])
+    if layout in class_layouts:
+        bw, bh = class_layouts[layout]
+        nw = (gw - 1) // bw + 1
+        nh = (gh - 1) // bh + 1
+        blocks = [[] for _i in range(nw * nh)]
+        for _iter in range(1000000):
+            real, label = training_set.get_minibatch_np(1)
+            idx = np.argmax(label[0])
+            while idx < len(blocks) and len(blocks[idx]) >= bw * bh:
+                idx += training_set.label_size
+            if idx < len(blocks):
+                blocks[idx].append((real, label))
+                if all(len(block) >= bw * bh for block in blocks):
+                    break
+        for i, block in enumerate(blocks):
+            for j, (real, label) in enumerate(block):
+                x = (i %  nw) * bw + j %  bw
+                y = (i // nw) * bh + j // bw
+                if x < gw and y < gh:
+                    reals[x + y * gw] = real[0]
+                    labels[x + y * gw] = label[0]
+
+    return (gw, gh), reals, labels
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/training/networks_stylegan.py b/reconstruction/ostec/external/stylegan2/training/networks_stylegan.py
new file mode 100644
index 000000000..76ce31caa
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/networks_stylegan.py
@@ -0,0 +1,660 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Network architectures used in the StyleGAN paper."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+
+# NOTE: Do not import any application-specific modules here!
+# Specify all network parameters as kwargs.
+
+#----------------------------------------------------------------------------
+# Primitive ops for manipulating 4D activation tensors.
+# The gradients of these are not necessary efficient or even meaningful.
+
+def _blur2d(x, f=[1,2,1], normalize=True, flip=False, stride=1):
+    assert x.shape.ndims == 4 and all(dim.value is not None for dim in x.shape[1:])
+    assert isinstance(stride, int) and stride >= 1
+
+    # Finalize filter kernel.
+    f = np.array(f, dtype=np.float32)
+    if f.ndim == 1:
+        f = f[:, np.newaxis] * f[np.newaxis, :]
+    assert f.ndim == 2
+    if normalize:
+        f /= np.sum(f)
+    if flip:
+        f = f[::-1, ::-1]
+    f = f[:, :, np.newaxis, np.newaxis]
+    f = np.tile(f, [1, 1, int(x.shape[1]), 1])
+
+    # No-op => early exit.
+    if f.shape == (1, 1) and f[0,0] == 1:
+        return x
+
+    # Convolve using depthwise_conv2d.
+    orig_dtype = x.dtype
+    x = tf.cast(x, tf.float32)  # tf.nn.depthwise_conv2d() doesn't support fp16
+    f = tf.constant(f, dtype=x.dtype, name='filter')
+    strides = [1, 1, stride, stride]
+    x = tf.nn.depthwise_conv2d(x, f, strides=strides, padding='SAME', data_format='NCHW')
+    x = tf.cast(x, orig_dtype)
+    return x
+
+def _upscale2d(x, factor=2, gain=1):
+    assert x.shape.ndims == 4 and all(dim.value is not None for dim in x.shape[1:])
+    assert isinstance(factor, int) and factor >= 1
+
+    # Apply gain.
+    if gain != 1:
+        x *= gain
+
+    # No-op => early exit.
+    if factor == 1:
+        return x
+
+    # Upscale using tf.tile().
+    s = x.shape
+    x = tf.reshape(x, [-1, s[1], s[2], 1, s[3], 1])
+    x = tf.tile(x, [1, 1, 1, factor, 1, factor])
+    x = tf.reshape(x, [-1, s[1], s[2] * factor, s[3] * factor])
+    return x
+
+def _downscale2d(x, factor=2, gain=1):
+    assert x.shape.ndims == 4 and all(dim.value is not None for dim in x.shape[1:])
+    assert isinstance(factor, int) and factor >= 1
+
+    # 2x2, float32 => downscale using _blur2d().
+    if factor == 2 and x.dtype == tf.float32:
+        f = [np.sqrt(gain) / factor] * factor
+        return _blur2d(x, f=f, normalize=False, stride=factor)
+
+    # Apply gain.
+    if gain != 1:
+        x *= gain
+
+    # No-op => early exit.
+    if factor == 1:
+        return x
+
+    # Large factor => downscale using tf.nn.avg_pool().
+    # NOTE: Requires tf_config['graph_options.place_pruned_graph']=True to work.
+    ksize = [1, 1, factor, factor]
+    return tf.nn.avg_pool(x, ksize=ksize, strides=ksize, padding='VALID', data_format='NCHW')
+
+#----------------------------------------------------------------------------
+# High-level ops for manipulating 4D activation tensors.
+# The gradients of these are meant to be as efficient as possible.
+
+def blur2d(x, f=[1,2,1], normalize=True):
+    with tf.variable_scope('Blur2D'):
+        @tf.custom_gradient
+        def func(x):
+            y = _blur2d(x, f, normalize)
+            @tf.custom_gradient
+            def grad(dy):
+                dx = _blur2d(dy, f, normalize, flip=True)
+                return dx, lambda ddx: _blur2d(ddx, f, normalize)
+            return y, grad
+        return func(x)
+
+def upscale2d(x, factor=2):
+    with tf.variable_scope('Upscale2D'):
+        @tf.custom_gradient
+        def func(x):
+            y = _upscale2d(x, factor)
+            @tf.custom_gradient
+            def grad(dy):
+                dx = _downscale2d(dy, factor, gain=factor**2)
+                return dx, lambda ddx: _upscale2d(ddx, factor)
+            return y, grad
+        return func(x)
+
+def downscale2d(x, factor=2):
+    with tf.variable_scope('Downscale2D'):
+        @tf.custom_gradient
+        def func(x):
+            y = _downscale2d(x, factor)
+            @tf.custom_gradient
+            def grad(dy):
+                dx = _upscale2d(dy, factor, gain=1/factor**2)
+                return dx, lambda ddx: _downscale2d(ddx, factor)
+            return y, grad
+        return func(x)
+
+#----------------------------------------------------------------------------
+# Get/create weight tensor for a convolutional or fully-connected layer.
+
+def get_weight(shape, gain=np.sqrt(2), use_wscale=False, lrmul=1):
+    fan_in = np.prod(shape[:-1]) # [kernel, kernel, fmaps_in, fmaps_out] or [in, out]
+    he_std = gain / np.sqrt(fan_in) # He init
+
+    # Equalized learning rate and custom learning rate multiplier.
+    if use_wscale:
+        init_std = 1.0 / lrmul
+        runtime_coef = he_std * lrmul
+    else:
+        init_std = he_std / lrmul
+        runtime_coef = lrmul
+
+    # Create variable.
+    init = tf.initializers.random_normal(0, init_std)
+    return tf.get_variable('weight', shape=shape, initializer=init) * runtime_coef
+
+#----------------------------------------------------------------------------
+# Fully-connected layer.
+
+def dense(x, fmaps, **kwargs):
+    if len(x.shape) > 2:
+        x = tf.reshape(x, [-1, np.prod([d.value for d in x.shape[1:]])])
+    w = get_weight([x.shape[1].value, fmaps], **kwargs)
+    w = tf.cast(w, x.dtype)
+    return tf.matmul(x, w)
+
+#----------------------------------------------------------------------------
+# Convolutional layer.
+
+def conv2d(x, fmaps, kernel, **kwargs):
+    assert kernel >= 1 and kernel % 2 == 1
+    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], **kwargs)
+    w = tf.cast(w, x.dtype)
+    return tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='SAME', data_format='NCHW')
+
+#----------------------------------------------------------------------------
+# Fused convolution + scaling.
+# Faster and uses less memory than performing the operations separately.
+
+def upscale2d_conv2d(x, fmaps, kernel, fused_scale='auto', **kwargs):
+    assert kernel >= 1 and kernel % 2 == 1
+    assert fused_scale in [True, False, 'auto']
+    if fused_scale == 'auto':
+        fused_scale = min(x.shape[2:]) * 2 >= 128
+
+    # Not fused => call the individual ops directly.
+    if not fused_scale:
+        return conv2d(upscale2d(x), fmaps, kernel, **kwargs)
+
+    # Fused => perform both ops simultaneously using tf.nn.conv2d_transpose().
+    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], **kwargs)
+    w = tf.transpose(w, [0, 1, 3, 2]) # [kernel, kernel, fmaps_out, fmaps_in]
+    w = tf.pad(w, [[1,1], [1,1], [0,0], [0,0]], mode='CONSTANT')
+    w = tf.add_n([w[1:, 1:], w[:-1, 1:], w[1:, :-1], w[:-1, :-1]])
+    w = tf.cast(w, x.dtype)
+    os = [tf.shape(x)[0], fmaps, x.shape[2] * 2, x.shape[3] * 2]
+    return tf.nn.conv2d_transpose(x, w, os, strides=[1,1,2,2], padding='SAME', data_format='NCHW')
+
+def conv2d_downscale2d(x, fmaps, kernel, fused_scale='auto', **kwargs):
+    assert kernel >= 1 and kernel % 2 == 1
+    assert fused_scale in [True, False, 'auto']
+    if fused_scale == 'auto':
+        fused_scale = min(x.shape[2:]) >= 128
+
+    # Not fused => call the individual ops directly.
+    if not fused_scale:
+        return downscale2d(conv2d(x, fmaps, kernel, **kwargs))
+
+    # Fused => perform both ops simultaneously using tf.nn.conv2d().
+    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], **kwargs)
+    w = tf.pad(w, [[1,1], [1,1], [0,0], [0,0]], mode='CONSTANT')
+    w = tf.add_n([w[1:, 1:], w[:-1, 1:], w[1:, :-1], w[:-1, :-1]]) * 0.25
+    w = tf.cast(w, x.dtype)
+    return tf.nn.conv2d(x, w, strides=[1,1,2,2], padding='SAME', data_format='NCHW')
+
+#----------------------------------------------------------------------------
+# Apply bias to the given activation tensor.
+
+def apply_bias(x, lrmul=1):
+    b = tf.get_variable('bias', shape=[x.shape[1]], initializer=tf.initializers.zeros()) * lrmul
+    b = tf.cast(b, x.dtype)
+    if len(x.shape) == 2:
+        return x + b
+    return x + tf.reshape(b, [1, -1, 1, 1])
+
+#----------------------------------------------------------------------------
+# Leaky ReLU activation. More efficient than tf.nn.leaky_relu() and supports FP16.
+
+def leaky_relu(x, alpha=0.2):
+    with tf.variable_scope('LeakyReLU'):
+        alpha = tf.constant(alpha, dtype=x.dtype, name='alpha')
+        @tf.custom_gradient
+        def func(x):
+            y = tf.maximum(x, x * alpha)
+            @tf.custom_gradient
+            def grad(dy):
+                dx = tf.where(y >= 0, dy, dy * alpha)
+                return dx, lambda ddx: tf.where(y >= 0, ddx, ddx * alpha)
+            return y, grad
+        return func(x)
+
+#----------------------------------------------------------------------------
+# Pixelwise feature vector normalization.
+
+def pixel_norm(x, epsilon=1e-8):
+    with tf.variable_scope('PixelNorm'):
+        epsilon = tf.constant(epsilon, dtype=x.dtype, name='epsilon')
+        return x * tf.rsqrt(tf.reduce_mean(tf.square(x), axis=1, keepdims=True) + epsilon)
+
+#----------------------------------------------------------------------------
+# Instance normalization.
+
+def instance_norm(x, epsilon=1e-8):
+    assert len(x.shape) == 4 # NCHW
+    with tf.variable_scope('InstanceNorm'):
+        orig_dtype = x.dtype
+        x = tf.cast(x, tf.float32)
+        x -= tf.reduce_mean(x, axis=[2,3], keepdims=True)
+        epsilon = tf.constant(epsilon, dtype=x.dtype, name='epsilon')
+        x *= tf.rsqrt(tf.reduce_mean(tf.square(x), axis=[2,3], keepdims=True) + epsilon)
+        x = tf.cast(x, orig_dtype)
+        return x
+
+#----------------------------------------------------------------------------
+# Style modulation.
+
+def style_mod(x, dlatent, **kwargs):
+    with tf.variable_scope('StyleMod'):
+        style = apply_bias(dense(dlatent, fmaps=x.shape[1]*2, gain=1, **kwargs))
+        style = tf.reshape(style, [-1, 2, x.shape[1]] + [1] * (len(x.shape) - 2))
+        return x * (style[:,0] + 1) + style[:,1]
+
+#----------------------------------------------------------------------------
+# Noise input.
+
+def apply_noise(x, noise_var=None, randomize_noise=True):
+    assert len(x.shape) == 4 # NCHW
+    with tf.variable_scope('Noise'):
+        if noise_var is None or randomize_noise:
+            noise = tf.random_normal([tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
+        else:
+            noise = tf.cast(noise_var, x.dtype)
+        weight = tf.get_variable('weight', shape=[x.shape[1].value], initializer=tf.initializers.zeros())
+        return x + noise * tf.reshape(tf.cast(weight, x.dtype), [1, -1, 1, 1])
+
+#----------------------------------------------------------------------------
+# Minibatch standard deviation.
+
+def minibatch_stddev_layer(x, group_size=4, num_new_features=1):
+    with tf.variable_scope('MinibatchStddev'):
+        group_size = tf.minimum(group_size, tf.shape(x)[0])     # Minibatch must be divisible by (or smaller than) group_size.
+        s = x.shape                                             # [NCHW]  Input shape.
+        y = tf.reshape(x, [group_size, -1, num_new_features, s[1]//num_new_features, s[2], s[3]])   # [GMncHW] Split minibatch into M groups of size G. Split channels into n channel groups c.
+        y = tf.cast(y, tf.float32)                              # [GMncHW] Cast to FP32.
+        y -= tf.reduce_mean(y, axis=0, keepdims=True)           # [GMncHW] Subtract mean over group.
+        y = tf.reduce_mean(tf.square(y), axis=0)                # [MncHW]  Calc variance over group.
+        y = tf.sqrt(y + 1e-8)                                   # [MncHW]  Calc stddev over group.
+        y = tf.reduce_mean(y, axis=[2,3,4], keepdims=True)      # [Mn111]  Take average over fmaps and pixels.
+        y = tf.reduce_mean(y, axis=[2])                         # [Mn11] Split channels into c channel groups
+        y = tf.cast(y, x.dtype)                                 # [Mn11]  Cast back to original data type.
+        y = tf.tile(y, [group_size, 1, s[2], s[3]])             # [NnHW]  Replicate over group and pixels.
+        return tf.concat([x, y], axis=1)                        # [NCHW]  Append as new fmap.
+
+#----------------------------------------------------------------------------
+# Style-based generator used in the StyleGAN paper.
+# Composed of two sub-networks (G_mapping and G_synthesis) that are defined below.
+
+def G_style(
+    latents_in,                                     # First input: Latent vectors (Z) [minibatch, latent_size].
+    labels_in,                                      # Second input: Conditioning labels [minibatch, label_size].
+    truncation_psi          = 0.7,                  # Style strength multiplier for the truncation trick. None = disable.
+    truncation_cutoff       = 8,                    # Number of layers for which to apply the truncation trick. None = disable.
+    truncation_psi_val      = None,                 # Value for truncation_psi to use during validation.
+    truncation_cutoff_val   = None,                 # Value for truncation_cutoff to use during validation.
+    dlatent_avg_beta        = 0.995,                # Decay for tracking the moving average of W during training. None = disable.
+    style_mixing_prob       = 0.9,                  # Probability of mixing styles during training. None = disable.
+    is_training             = False,                # Network is under training? Enables and disables specific features.
+    is_validation           = False,                # Network is under validation? Chooses which value to use for truncation_psi.
+    is_template_graph       = False,                # True = template graph constructed by the Network class, False = actual evaluation.
+    components              = dnnlib.EasyDict(),    # Container for sub-networks. Retained between calls.
+    **kwargs):                                      # Arguments for sub-networks (G_mapping and G_synthesis).
+
+    # Validate arguments.
+    assert not is_training or not is_validation
+    assert isinstance(components, dnnlib.EasyDict)
+    if is_validation:
+        truncation_psi = truncation_psi_val
+        truncation_cutoff = truncation_cutoff_val
+    if is_training or (truncation_psi is not None and not tflib.is_tf_expression(truncation_psi) and truncation_psi == 1):
+        truncation_psi = None
+    if is_training or (truncation_cutoff is not None and not tflib.is_tf_expression(truncation_cutoff) and truncation_cutoff <= 0):
+        truncation_cutoff = None
+    if not is_training or (dlatent_avg_beta is not None and not tflib.is_tf_expression(dlatent_avg_beta) and dlatent_avg_beta == 1):
+        dlatent_avg_beta = None
+    if not is_training or (style_mixing_prob is not None and not tflib.is_tf_expression(style_mixing_prob) and style_mixing_prob <= 0):
+        style_mixing_prob = None
+
+    # Setup components.
+    if 'synthesis' not in components:
+        components.synthesis = tflib.Network('G_synthesis', func_name=G_synthesis, **kwargs)
+    num_layers = components.synthesis.input_shape[1]
+    dlatent_size = components.synthesis.input_shape[2]
+    if 'mapping' not in components:
+        components.mapping = tflib.Network('G_mapping', func_name=G_mapping, dlatent_broadcast=num_layers, **kwargs)
+
+    # Setup variables.
+    lod_in = tf.get_variable('lod', initializer=np.float32(0), trainable=False)
+    dlatent_avg = tf.get_variable('dlatent_avg', shape=[dlatent_size], initializer=tf.initializers.zeros(), trainable=False)
+
+    # Evaluate mapping network.
+    dlatents = components.mapping.get_output_for(latents_in, labels_in, **kwargs)
+
+    # Update moving average of W.
+    if dlatent_avg_beta is not None:
+        with tf.variable_scope('DlatentAvg'):
+            batch_avg = tf.reduce_mean(dlatents[:, 0], axis=0)
+            update_op = tf.assign(dlatent_avg, tflib.lerp(batch_avg, dlatent_avg, dlatent_avg_beta))
+            with tf.control_dependencies([update_op]):
+                dlatents = tf.identity(dlatents)
+
+    # Perform style mixing regularization.
+    if style_mixing_prob is not None:
+        with tf.name_scope('StyleMix'):
+            latents2 = tf.random_normal(tf.shape(latents_in))
+            dlatents2 = components.mapping.get_output_for(latents2, labels_in, **kwargs)
+            layer_idx = np.arange(num_layers)[np.newaxis, :, np.newaxis]
+            cur_layers = num_layers - tf.cast(lod_in, tf.int32) * 2
+            mixing_cutoff = tf.cond(
+                tf.random_uniform([], 0.0, 1.0) < style_mixing_prob,
+                lambda: tf.random_uniform([], 1, cur_layers, dtype=tf.int32),
+                lambda: cur_layers)
+            dlatents = tf.where(tf.broadcast_to(layer_idx < mixing_cutoff, tf.shape(dlatents)), dlatents, dlatents2)
+
+    # Apply truncation trick.
+    if truncation_psi is not None and truncation_cutoff is not None:
+        with tf.variable_scope('Truncation'):
+            layer_idx = np.arange(num_layers)[np.newaxis, :, np.newaxis]
+            ones = np.ones(layer_idx.shape, dtype=np.float32)
+            coefs = tf.where(layer_idx < truncation_cutoff, truncation_psi * ones, ones)
+            dlatents = tflib.lerp(dlatent_avg, dlatents, coefs)
+
+    # Evaluate synthesis network.
+    with tf.control_dependencies([tf.assign(components.synthesis.find_var('lod'), lod_in)]):
+        images_out = components.synthesis.get_output_for(dlatents, force_clean_graph=is_template_graph, **kwargs)
+    return tf.identity(images_out, name='images_out')
+
+#----------------------------------------------------------------------------
+# Mapping network used in the StyleGAN paper.
+
+def G_mapping(
+    latents_in,                             # First input: Latent vectors (Z) [minibatch, latent_size].
+    labels_in,                              # Second input: Conditioning labels [minibatch, label_size].
+    latent_size             = 512,          # Latent vector (Z) dimensionality.
+    label_size              = 0,            # Label dimensionality, 0 if no labels.
+    dlatent_size            = 512,          # Disentangled latent (W) dimensionality.
+    dlatent_broadcast       = None,         # Output disentangled latent (W) as [minibatch, dlatent_size] or [minibatch, dlatent_broadcast, dlatent_size].
+    mapping_layers          = 8,            # Number of mapping layers.
+    mapping_fmaps           = 512,          # Number of activations in the mapping layers.
+    mapping_lrmul           = 0.01,         # Learning rate multiplier for the mapping layers.
+    mapping_nonlinearity    = 'lrelu',      # Activation function: 'relu', 'lrelu'.
+    use_wscale              = True,         # Enable equalized learning rate?
+    normalize_latents       = True,         # Normalize latent vectors (Z) before feeding them to the mapping layers?
+    dtype                   = 'float32',    # Data type to use for activations and outputs.
+    **_kwargs):                             # Ignore unrecognized keyword args.
+
+    act, gain = {'relu': (tf.nn.relu, np.sqrt(2)), 'lrelu': (leaky_relu, np.sqrt(2))}[mapping_nonlinearity]
+
+    # Inputs.
+    latents_in.set_shape([None, latent_size])
+    labels_in.set_shape([None, label_size])
+    latents_in = tf.cast(latents_in, dtype)
+    labels_in = tf.cast(labels_in, dtype)
+    x = latents_in
+
+    # Embed labels and concatenate them with latents.
+    if label_size:
+        with tf.variable_scope('LabelConcat'):
+            w = tf.get_variable('weight', shape=[label_size, latent_size], initializer=tf.initializers.random_normal())
+            y = tf.matmul(labels_in, tf.cast(w, dtype))
+            x = tf.concat([x, y], axis=1)
+
+    # Normalize latents.
+    if normalize_latents:
+        x = pixel_norm(x)
+
+    # Mapping layers.
+    for layer_idx in range(mapping_layers):
+        with tf.variable_scope('Dense%d' % layer_idx):
+            fmaps = dlatent_size if layer_idx == mapping_layers - 1 else mapping_fmaps
+            x = dense(x, fmaps=fmaps, gain=gain, use_wscale=use_wscale, lrmul=mapping_lrmul)
+            x = apply_bias(x, lrmul=mapping_lrmul)
+            x = act(x)
+
+    # Broadcast.
+    if dlatent_broadcast is not None:
+        with tf.variable_scope('Broadcast'):
+            x = tf.tile(x[:, np.newaxis], [1, dlatent_broadcast, 1])
+
+    # Output.
+    assert x.dtype == tf.as_dtype(dtype)
+    return tf.identity(x, name='dlatents_out')
+
+#----------------------------------------------------------------------------
+# Synthesis network used in the StyleGAN paper.
+
+def G_synthesis(
+    dlatents_in,                        # Input: Disentangled latents (W) [minibatch, num_layers, dlatent_size].
+    dlatent_size        = 512,          # Disentangled latent (W) dimensionality.
+    num_channels        = 3,            # Number of output color channels.
+    resolution          = 1024,         # Output resolution.
+    fmap_base           = 8192,         # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    use_styles          = True,         # Enable style inputs?
+    const_input_layer   = True,         # First layer is a learned constant?
+    use_noise           = True,         # Enable noise inputs?
+    randomize_noise     = True,         # True = randomize noise inputs every time (non-deterministic), False = read noise inputs from variables.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu'
+    use_wscale          = True,         # Enable equalized learning rate?
+    use_pixel_norm      = False,        # Enable pixelwise feature vector normalization?
+    use_instance_norm   = True,         # Enable instance normalization?
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    fused_scale         = 'auto',       # True = fused convolution + scaling, False = separate ops, 'auto' = decide automatically.
+    blur_filter         = [1,2,1],      # Low-pass filter to apply when resampling activations. None = no filtering.
+    structure           = 'auto',       # 'fixed' = no progressive growing, 'linear' = human-readable, 'recursive' = efficient, 'auto' = select automatically.
+    is_template_graph   = False,        # True = template graph constructed by the Network class, False = actual evaluation.
+    force_clean_graph   = False,        # True = construct a clean graph that looks nice in TensorBoard, False = default behavior.
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return min(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_max)
+    def blur(x): return blur2d(x, blur_filter) if blur_filter else x
+    if is_template_graph: force_clean_graph = True
+    if force_clean_graph: randomize_noise = False
+    if structure == 'auto': structure = 'linear' if force_clean_graph else 'recursive'
+    act, gain = {'relu': (tf.nn.relu, np.sqrt(2)), 'lrelu': (leaky_relu, np.sqrt(2))}[nonlinearity]
+    num_layers = resolution_log2 * 2 - 2
+    num_styles = num_layers if use_styles else 1
+    images_out = None
+
+    # Primary inputs.
+    dlatents_in.set_shape([None, num_styles, dlatent_size])
+    dlatents_in = tf.cast(dlatents_in, dtype)
+    lod_in = tf.cast(tf.get_variable('lod', initializer=np.float32(0), trainable=False), dtype)
+
+    # Noise inputs.
+    noise_inputs = []
+    if use_noise:
+        for layer_idx in range(num_layers):
+            res = layer_idx // 2 + 2
+            shape = [1, use_noise, 2**res, 2**res]
+            noise_inputs.append(tf.get_variable('noise%d' % layer_idx, shape=shape, initializer=tf.initializers.random_normal(), trainable=False))
+
+    # Things to do at the end of each layer.
+    def layer_epilogue(x, layer_idx):
+        if use_noise:
+            x = apply_noise(x, noise_inputs[layer_idx], randomize_noise=randomize_noise)
+        x = apply_bias(x)
+        x = act(x)
+        if use_pixel_norm:
+            x = pixel_norm(x)
+        if use_instance_norm:
+            x = instance_norm(x)
+        if use_styles:
+            x = style_mod(x, dlatents_in[:, layer_idx], use_wscale=use_wscale)
+        return x
+
+    # Early layers.
+    with tf.variable_scope('4x4'):
+        if const_input_layer:
+            with tf.variable_scope('Const'):
+                x = tf.get_variable('const', shape=[1, nf(1), 4, 4], initializer=tf.initializers.ones())
+                x = layer_epilogue(tf.tile(tf.cast(x, dtype), [tf.shape(dlatents_in)[0], 1, 1, 1]), 0)
+        else:
+            with tf.variable_scope('Dense'):
+                x = dense(dlatents_in[:, 0], fmaps=nf(1)*16, gain=gain/4, use_wscale=use_wscale) # tweak gain to match the official implementation of Progressing GAN
+                x = layer_epilogue(tf.reshape(x, [-1, nf(1), 4, 4]), 0)
+        with tf.variable_scope('Conv'):
+            x = layer_epilogue(conv2d(x, fmaps=nf(1), kernel=3, gain=gain, use_wscale=use_wscale), 1)
+
+    # Building blocks for remaining layers.
+    def block(res, x): # res = 3..resolution_log2
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            with tf.variable_scope('Conv0_up'):
+                x = layer_epilogue(blur(upscale2d_conv2d(x, fmaps=nf(res-1), kernel=3, gain=gain, use_wscale=use_wscale, fused_scale=fused_scale)), res*2-4)
+            with tf.variable_scope('Conv1'):
+                x = layer_epilogue(conv2d(x, fmaps=nf(res-1), kernel=3, gain=gain, use_wscale=use_wscale), res*2-3)
+            return x
+    def torgb(res, x): # res = 2..resolution_log2
+        lod = resolution_log2 - res
+        with tf.variable_scope('ToRGB_lod%d' % lod):
+            return apply_bias(conv2d(x, fmaps=num_channels, kernel=1, gain=1, use_wscale=use_wscale))
+
+    # Fixed structure: simple and efficient, but does not support progressive growing.
+    if structure == 'fixed':
+        for res in range(3, resolution_log2 + 1):
+            x = block(res, x)
+        images_out = torgb(resolution_log2, x)
+
+    # Linear structure: simple but inefficient.
+    if structure == 'linear':
+        images_out = torgb(2, x)
+        for res in range(3, resolution_log2 + 1):
+            lod = resolution_log2 - res
+            x = block(res, x)
+            img = torgb(res, x)
+            images_out = upscale2d(images_out)
+            with tf.variable_scope('Grow_lod%d' % lod):
+                images_out = tflib.lerp_clip(img, images_out, lod_in - lod)
+
+    # Recursive structure: complex but efficient.
+    if structure == 'recursive':
+        def cset(cur_lambda, new_cond, new_lambda):
+            return lambda: tf.cond(new_cond, new_lambda, cur_lambda)
+        def grow(x, res, lod):
+            y = block(res, x)
+            img = lambda: upscale2d(torgb(res, y), 2**lod)
+            img = cset(img, (lod_in > lod), lambda: upscale2d(tflib.lerp(torgb(res, y), upscale2d(torgb(res - 1, x)), lod_in - lod), 2**lod))
+            if lod > 0: img = cset(img, (lod_in < lod), lambda: grow(y, res + 1, lod - 1))
+            return img()
+        images_out = grow(x, 3, resolution_log2 - 3)
+
+    assert images_out.dtype == tf.as_dtype(dtype)
+    return tf.identity(images_out, name='images_out')
+
+#----------------------------------------------------------------------------
+# Discriminator used in the StyleGAN paper.
+
+def D_basic(
+    images_in,                          # First input: Images [minibatch, channel, height, width].
+    labels_in,                          # Second input: Labels [minibatch, label_size].
+    num_channels        = 1,            # Number of input color channels. Overridden based on dataset.
+    resolution          = 32,           # Input resolution. Overridden based on dataset.
+    label_size          = 0,            # Dimensionality of the labels, 0 if no labels. Overridden based on dataset.
+    fmap_base           = 8192,         # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu',
+    use_wscale          = True,         # Enable equalized learning rate?
+    mbstd_group_size    = 4,            # Group size for the minibatch standard deviation layer, 0 = disable.
+    mbstd_num_features  = 1,            # Number of features for the minibatch standard deviation layer.
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    fused_scale         = 'auto',       # True = fused convolution + scaling, False = separate ops, 'auto' = decide automatically.
+    blur_filter         = [1,2,1],      # Low-pass filter to apply when resampling activations. None = no filtering.
+    structure           = 'auto',       # 'fixed' = no progressive growing, 'linear' = human-readable, 'recursive' = efficient, 'auto' = select automatically.
+    is_template_graph   = False,        # True = template graph constructed by the Network class, False = actual evaluation.
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return min(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_max)
+    def blur(x): return blur2d(x, blur_filter) if blur_filter else x
+    if structure == 'auto': structure = 'linear' if is_template_graph else 'recursive'
+    act, gain = {'relu': (tf.nn.relu, np.sqrt(2)), 'lrelu': (leaky_relu, np.sqrt(2))}[nonlinearity]
+
+    images_in.set_shape([None, num_channels, resolution, resolution])
+    labels_in.set_shape([None, label_size])
+    images_in = tf.cast(images_in, dtype)
+    labels_in = tf.cast(labels_in, dtype)
+    lod_in = tf.cast(tf.get_variable('lod', initializer=np.float32(0.0), trainable=False), dtype)
+    scores_out = None
+
+    # Building blocks.
+    def fromrgb(x, res): # res = 2..resolution_log2
+        with tf.variable_scope('FromRGB_lod%d' % (resolution_log2 - res)):
+            return act(apply_bias(conv2d(x, fmaps=nf(res-1), kernel=1, gain=gain, use_wscale=use_wscale)))
+    def block(x, res): # res = 2..resolution_log2
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            if res >= 3: # 8x8 and up
+                with tf.variable_scope('Conv0'):
+                    x = act(apply_bias(conv2d(x, fmaps=nf(res-1), kernel=3, gain=gain, use_wscale=use_wscale)))
+                with tf.variable_scope('Conv1_down'):
+                    x = act(apply_bias(conv2d_downscale2d(blur(x), fmaps=nf(res-2), kernel=3, gain=gain, use_wscale=use_wscale, fused_scale=fused_scale)))
+            else: # 4x4
+                if mbstd_group_size > 1:
+                    x = minibatch_stddev_layer(x, mbstd_group_size, mbstd_num_features)
+                with tf.variable_scope('Conv'):
+                    x = act(apply_bias(conv2d(x, fmaps=nf(res-1), kernel=3, gain=gain, use_wscale=use_wscale)))
+                with tf.variable_scope('Dense0'):
+                    x = act(apply_bias(dense(x, fmaps=nf(res-2), gain=gain, use_wscale=use_wscale)))
+                with tf.variable_scope('Dense1'):
+                    x = apply_bias(dense(x, fmaps=max(label_size, 1), gain=1, use_wscale=use_wscale))
+            return x
+
+    # Fixed structure: simple and efficient, but does not support progressive growing.
+    if structure == 'fixed':
+        x = fromrgb(images_in, resolution_log2)
+        for res in range(resolution_log2, 2, -1):
+            x = block(x, res)
+        scores_out = block(x, 2)
+
+    # Linear structure: simple but inefficient.
+    if structure == 'linear':
+        img = images_in
+        x = fromrgb(img, resolution_log2)
+        for res in range(resolution_log2, 2, -1):
+            lod = resolution_log2 - res
+            x = block(x, res)
+            img = downscale2d(img)
+            y = fromrgb(img, res - 1)
+            with tf.variable_scope('Grow_lod%d' % lod):
+                x = tflib.lerp_clip(x, y, lod_in - lod)
+        scores_out = block(x, 2)
+
+    # Recursive structure: complex but efficient.
+    if structure == 'recursive':
+        def cset(cur_lambda, new_cond, new_lambda):
+            return lambda: tf.cond(new_cond, new_lambda, cur_lambda)
+        def grow(res, lod):
+            x = lambda: fromrgb(downscale2d(images_in, 2**lod), res)
+            if lod > 0: x = cset(x, (lod_in < lod), lambda: grow(res + 1, lod - 1))
+            x = block(x(), res); y = lambda: x
+            if res > 2: y = cset(y, (lod_in > lod), lambda: tflib.lerp(x, fromrgb(downscale2d(images_in, 2**(lod+1)), res - 1), lod_in - lod))
+            return y()
+        scores_out = grow(2, resolution_log2 - 2)
+
+    # Label conditioning from "Which Training Methods for GANs do actually Converge?"
+    if label_size:
+        with tf.variable_scope('LabelSwitch'):
+            scores_out = tf.reduce_sum(scores_out * labels_in, axis=1, keepdims=True)
+
+    assert scores_out.dtype == tf.as_dtype(dtype)
+    scores_out = tf.identity(scores_out, name='scores_out')
+    return scores_out
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/training/networks_stylegan2.py b/reconstruction/ostec/external/stylegan2/training/networks_stylegan2.py
new file mode 100644
index 000000000..6c96fc192
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/networks_stylegan2.py
@@ -0,0 +1,697 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Network architectures used in the StyleGAN2 paper."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+from dnnlib.tflib.ops.upfirdn_2d import upsample_2d, downsample_2d, upsample_conv_2d, conv_downsample_2d
+from dnnlib.tflib.ops.fused_bias_act import fused_bias_act
+
+# NOTE: Do not import any application-specific modules here!
+# Specify all network parameters as kwargs.
+
+#----------------------------------------------------------------------------
+# Get/create weight tensor for a convolution or fully-connected layer.
+
+def get_weight(shape, gain=1, use_wscale=True, lrmul=1, weight_var='weight'):
+    fan_in = np.prod(shape[:-1]) # [kernel, kernel, fmaps_in, fmaps_out] or [in, out]
+    he_std = gain / np.sqrt(fan_in) # He init
+
+    # Equalized learning rate and custom learning rate multiplier.
+    if use_wscale:
+        init_std = 1.0 / lrmul
+        runtime_coef = he_std * lrmul
+    else:
+        init_std = he_std / lrmul
+        runtime_coef = lrmul
+
+    # Create variable.
+    init = tf.initializers.random_normal(0, init_std)
+    return tf.get_variable(weight_var, shape=shape, initializer=init) * runtime_coef
+
+#----------------------------------------------------------------------------
+# Fully-connected layer.
+
+def dense_layer(x, fmaps, gain=1, use_wscale=True, lrmul=1, weight_var='weight'):
+    if len(x.shape) > 2:
+        x = tf.reshape(x, [-1, np.prod([d.value for d in x.shape[1:]])])
+    w = get_weight([x.shape[1].value, fmaps], gain=gain, use_wscale=use_wscale, lrmul=lrmul, weight_var=weight_var)
+    w = tf.cast(w, x.dtype)
+    return tf.matmul(x, w)
+
+#----------------------------------------------------------------------------
+# Convolution layer with optional upsampling or downsampling.
+
+def conv2d_layer(x, fmaps, kernel, up=False, down=False, resample_kernel=None, gain=1, use_wscale=True, lrmul=1, weight_var='weight'):
+    assert not (up and down)
+    assert kernel >= 1 and kernel % 2 == 1
+    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], gain=gain, use_wscale=use_wscale, lrmul=lrmul, weight_var=weight_var)
+    if up:
+        x = upsample_conv_2d(x, tf.cast(w, x.dtype), data_format='NCHW', k=resample_kernel)
+    elif down:
+        x = conv_downsample_2d(x, tf.cast(w, x.dtype), data_format='NCHW', k=resample_kernel)
+    else:
+        x = tf.nn.conv2d(x, tf.cast(w, x.dtype), data_format='NCHW', strides=[1,1,1,1], padding='SAME')
+    return x
+
+#----------------------------------------------------------------------------
+# Apply bias and activation func.
+
+def apply_bias_act(x, act='linear', alpha=None, gain=None, lrmul=1, bias_var='bias'):
+    b = tf.get_variable(bias_var, shape=[x.shape[1]], initializer=tf.initializers.zeros()) * lrmul
+    return fused_bias_act(x, b=tf.cast(b, x.dtype), act=act, alpha=alpha, gain=gain)
+
+#----------------------------------------------------------------------------
+# Naive upsampling (nearest neighbor) and downsampling (average pooling).
+
+def naive_upsample_2d(x, factor=2):
+    with tf.variable_scope('NaiveUpsample'):
+        _N, C, H, W = x.shape.as_list()
+        x = tf.reshape(x, [-1, C, H, 1, W, 1])
+        x = tf.tile(x, [1, 1, 1, factor, 1, factor])
+        return tf.reshape(x, [-1, C, H * factor, W * factor])
+
+def naive_downsample_2d(x, factor=2):
+    with tf.variable_scope('NaiveDownsample'):
+        _N, C, H, W = x.shape.as_list()
+        x = tf.reshape(x, [-1, C, H // factor, factor, W // factor, factor])
+        return tf.reduce_mean(x, axis=[3,5])
+
+#----------------------------------------------------------------------------
+# Modulated convolution layer.
+
+def modulated_conv2d_layer(x, y, fmaps, kernel, up=False, down=False, demodulate=True, resample_kernel=None, gain=1, use_wscale=True, lrmul=1, fused_modconv=True, weight_var='weight', mod_weight_var='mod_weight', mod_bias_var='mod_bias'):
+    assert not (up and down)
+    assert kernel >= 1 and kernel % 2 == 1
+
+    # Get weight.
+    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], gain=gain, use_wscale=use_wscale, lrmul=lrmul, weight_var=weight_var)
+    ww = w[np.newaxis] # [BkkIO] Introduce minibatch dimension.
+
+    # Modulate.
+    s = dense_layer(y, fmaps=x.shape[1].value, weight_var=mod_weight_var) # [BI] Transform incoming W to style.
+    s = apply_bias_act(s, bias_var=mod_bias_var) + 1 # [BI] Add bias (initially 1).
+    ww *= tf.cast(s[:, np.newaxis, np.newaxis, :, np.newaxis], w.dtype) # [BkkIO] Scale input feature maps.
+
+    # Demodulate.
+    if demodulate:
+        d = tf.rsqrt(tf.reduce_sum(tf.square(ww), axis=[1,2,3]) + 1e-8) # [BO] Scaling factor.
+        ww *= d[:, np.newaxis, np.newaxis, np.newaxis, :] # [BkkIO] Scale output feature maps.
+
+    # Reshape/scale input.
+    if fused_modconv:
+        x = tf.reshape(x, [1, -1, x.shape[2], x.shape[3]]) # Fused => reshape minibatch to convolution groups.
+        w = tf.reshape(tf.transpose(ww, [1, 2, 3, 0, 4]), [ww.shape[1], ww.shape[2], ww.shape[3], -1])
+    else:
+        x *= tf.cast(s[:, :, np.newaxis, np.newaxis], x.dtype) # [BIhw] Not fused => scale input activations.
+
+    # Convolution with optional up/downsampling.
+    if up:
+        x = upsample_conv_2d(x, tf.cast(w, x.dtype), data_format='NCHW', k=resample_kernel)
+    elif down:
+        x = conv_downsample_2d(x, tf.cast(w, x.dtype), data_format='NCHW', k=resample_kernel)
+    else:
+        x = tf.nn.conv2d(x, tf.cast(w, x.dtype), data_format='NCHW', strides=[1,1,1,1], padding='SAME')
+
+    # Reshape/scale output.
+    if fused_modconv:
+        x = tf.reshape(x, [-1, fmaps, x.shape[2], x.shape[3]]) # Fused => reshape convolution groups back to minibatch.
+    elif demodulate:
+        x *= tf.cast(d[:, :, np.newaxis, np.newaxis], x.dtype) # [BOhw] Not fused => scale output activations.
+    return x
+
+#----------------------------------------------------------------------------
+# Minibatch standard deviation layer.
+
+def minibatch_stddev_layer(x, group_size=4, num_new_features=1):
+    group_size = tf.minimum(group_size, tf.shape(x)[0])     # Minibatch must be divisible by (or smaller than) group_size.
+    s = x.shape                                             # [NCHW]  Input shape.
+    y = tf.reshape(x, [group_size, -1, num_new_features, s[1]//num_new_features, s[2], s[3]])   # [GMncHW] Split minibatch into M groups of size G. Split channels into n channel groups c.
+    y = tf.cast(y, tf.float32)                              # [GMncHW] Cast to FP32.
+    y -= tf.reduce_mean(y, axis=0, keepdims=True)           # [GMncHW] Subtract mean over group.
+    y = tf.reduce_mean(tf.square(y), axis=0)                # [MncHW]  Calc variance over group.
+    y = tf.sqrt(y + 1e-8)                                   # [MncHW]  Calc stddev over group.
+    y = tf.reduce_mean(y, axis=[2,3,4], keepdims=True)      # [Mn111]  Take average over fmaps and pixels.
+    y = tf.reduce_mean(y, axis=[2])                         # [Mn11] Split channels into c channel groups
+    y = tf.cast(y, x.dtype)                                 # [Mn11]  Cast back to original data type.
+    y = tf.tile(y, [group_size, 1, s[2], s[3]])             # [NnHW]  Replicate over group and pixels.
+    return tf.concat([x, y], axis=1)                        # [NCHW]  Append as new fmap.
+
+#----------------------------------------------------------------------------
+# Main generator network.
+# Composed of two sub-networks (mapping and synthesis) that are defined below.
+# Used in configs B-F (Table 1).
+
+def G_main(
+    latents_in,                                         # First input: Latent vectors (Z) [minibatch, latent_size].
+    labels_in,                                          # Second input: Conditioning labels [minibatch, label_size].
+    truncation_psi          = 0.5,                      # Style strength multiplier for the truncation trick. None = disable.
+    truncation_cutoff       = None,                     # Number of layers for which to apply the truncation trick. None = disable.
+    truncation_psi_val      = None,                     # Value for truncation_psi to use during validation.
+    truncation_cutoff_val   = None,                     # Value for truncation_cutoff to use during validation.
+    dlatent_avg_beta        = 0.995,                    # Decay for tracking the moving average of W during training. None = disable.
+    style_mixing_prob       = 0.9,                      # Probability of mixing styles during training. None = disable.
+    is_training             = False,                    # Network is under training? Enables and disables specific features.
+    is_validation           = False,                    # Network is under validation? Chooses which value to use for truncation_psi.
+    return_dlatents         = False,                    # Return dlatents in addition to the images?
+    is_template_graph       = False,                    # True = template graph constructed by the Network class, False = actual evaluation.
+    components              = dnnlib.EasyDict(),        # Container for sub-networks. Retained between calls.
+    mapping_func            = 'G_mapping',              # Build func name for the mapping network.
+    synthesis_func          = 'G_synthesis_stylegan2',  # Build func name for the synthesis network.
+    **kwargs):                                          # Arguments for sub-networks (mapping and synthesis).
+
+    # Validate arguments.
+    assert not is_training or not is_validation
+    assert isinstance(components, dnnlib.EasyDict)
+    if is_validation:
+        truncation_psi = truncation_psi_val
+        truncation_cutoff = truncation_cutoff_val
+    if is_training or (truncation_psi is not None and not tflib.is_tf_expression(truncation_psi) and truncation_psi == 1):
+        truncation_psi = None
+    if is_training:
+        truncation_cutoff = None
+    if not is_training or (dlatent_avg_beta is not None and not tflib.is_tf_expression(dlatent_avg_beta) and dlatent_avg_beta == 1):
+        dlatent_avg_beta = None
+    if not is_training or (style_mixing_prob is not None and not tflib.is_tf_expression(style_mixing_prob) and style_mixing_prob <= 0):
+        style_mixing_prob = None
+
+    # Setup components.
+    if 'synthesis' not in components:
+        components.synthesis = tflib.Network('G_synthesis', func_name=globals()[synthesis_func], **kwargs)
+    num_layers = components.synthesis.input_shape[1]
+    dlatent_size = components.synthesis.input_shape[2]
+    if 'mapping' not in components:
+        components.mapping = tflib.Network('G_mapping', func_name=globals()[mapping_func], dlatent_broadcast=num_layers, **kwargs)
+
+    # Setup variables.
+    lod_in = tf.get_variable('lod', initializer=np.float32(0), trainable=False)
+    dlatent_avg = tf.get_variable('dlatent_avg', shape=[dlatent_size], initializer=tf.initializers.zeros(), trainable=False)
+
+    # Evaluate mapping network.
+    dlatents = components.mapping.get_output_for(latents_in, labels_in, is_training=is_training, **kwargs)
+    dlatents = tf.cast(dlatents, tf.float32)
+
+    # Update moving average of W.
+    if dlatent_avg_beta is not None:
+        with tf.variable_scope('DlatentAvg'):
+            batch_avg = tf.reduce_mean(dlatents[:, 0], axis=0)
+            update_op = tf.assign(dlatent_avg, tflib.lerp(batch_avg, dlatent_avg, dlatent_avg_beta))
+            with tf.control_dependencies([update_op]):
+                dlatents = tf.identity(dlatents)
+
+    # Perform style mixing regularization.
+    if style_mixing_prob is not None:
+        with tf.variable_scope('StyleMix'):
+            latents2 = tf.random_normal(tf.shape(latents_in))
+            dlatents2 = components.mapping.get_output_for(latents2, labels_in, is_training=is_training, **kwargs)
+            dlatents2 = tf.cast(dlatents2, tf.float32)
+            layer_idx = np.arange(num_layers)[np.newaxis, :, np.newaxis]
+            cur_layers = num_layers - tf.cast(lod_in, tf.int32) * 2
+            mixing_cutoff = tf.cond(
+                tf.random_uniform([], 0.0, 1.0) < style_mixing_prob,
+                lambda: tf.random_uniform([], 1, cur_layers, dtype=tf.int32),
+                lambda: cur_layers)
+            dlatents = tf.where(tf.broadcast_to(layer_idx < mixing_cutoff, tf.shape(dlatents)), dlatents, dlatents2)
+
+    # Apply truncation trick.
+    if truncation_psi is not None:
+        with tf.variable_scope('Truncation'):
+            layer_idx = np.arange(num_layers)[np.newaxis, :, np.newaxis]
+            layer_psi = np.ones(layer_idx.shape, dtype=np.float32)
+            if truncation_cutoff is None:
+                layer_psi *= truncation_psi
+            else:
+                layer_psi = tf.where(layer_idx < truncation_cutoff, layer_psi * truncation_psi, layer_psi)
+            dlatents = tflib.lerp(dlatent_avg, dlatents, layer_psi)
+
+    # Evaluate synthesis network.
+    deps = []
+    if 'lod' in components.synthesis.vars:
+        deps.append(tf.assign(components.synthesis.vars['lod'], lod_in))
+    with tf.control_dependencies(deps):
+        images_out = components.synthesis.get_output_for(dlatents, is_training=is_training, force_clean_graph=is_template_graph, **kwargs)
+
+    # Return requested outputs.
+    images_out = tf.identity(images_out, name='images_out')
+    if return_dlatents:
+        return images_out, dlatents
+    return images_out
+
+#----------------------------------------------------------------------------
+# Mapping network.
+# Transforms the input latent code (z) to the disentangled latent code (w).
+# Used in configs B-F (Table 1).
+
+def G_mapping(
+    latents_in,                             # First input: Latent vectors (Z) [minibatch, latent_size].
+    labels_in,                              # Second input: Conditioning labels [minibatch, label_size].
+    latent_size             = 512,          # Latent vector (Z) dimensionality.
+    label_size              = 0,            # Label dimensionality, 0 if no labels.
+    dlatent_size            = 512,          # Disentangled latent (W) dimensionality.
+    dlatent_broadcast       = None,         # Output disentangled latent (W) as [minibatch, dlatent_size] or [minibatch, dlatent_broadcast, dlatent_size].
+    mapping_layers          = 8,            # Number of mapping layers.
+    mapping_fmaps           = 512,          # Number of activations in the mapping layers.
+    mapping_lrmul           = 0.01,         # Learning rate multiplier for the mapping layers.
+    mapping_nonlinearity    = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+    normalize_latents       = True,         # Normalize latent vectors (Z) before feeding them to the mapping layers?
+    dtype                   = 'float32',    # Data type to use for activations and outputs.
+    **_kwargs):                             # Ignore unrecognized keyword args.
+
+    act = mapping_nonlinearity
+
+    # Inputs.
+    latents_in.set_shape([None, latent_size])
+    labels_in.set_shape([None, label_size])
+    latents_in = tf.cast(latents_in, dtype)
+    labels_in = tf.cast(labels_in, dtype)
+    x = latents_in
+
+    # Embed labels and concatenate them with latents.
+    if label_size:
+        with tf.variable_scope('LabelConcat'):
+            w = tf.get_variable('weight', shape=[label_size, latent_size], initializer=tf.initializers.random_normal())
+            y = tf.matmul(labels_in, tf.cast(w, dtype))
+            x = tf.concat([x, y], axis=1)
+
+    # Normalize latents.
+    if normalize_latents:
+        with tf.variable_scope('Normalize'):
+            x *= tf.rsqrt(tf.reduce_mean(tf.square(x), axis=1, keepdims=True) + 1e-8)
+
+    # Mapping layers.
+    for layer_idx in range(mapping_layers):
+        with tf.variable_scope('Dense%d' % layer_idx):
+            fmaps = dlatent_size if layer_idx == mapping_layers - 1 else mapping_fmaps
+            x = apply_bias_act(dense_layer(x, fmaps=fmaps, lrmul=mapping_lrmul), act=act, lrmul=mapping_lrmul)
+
+    # Broadcast.
+    if dlatent_broadcast is not None:
+        with tf.variable_scope('Broadcast'):
+            x = tf.tile(x[:, np.newaxis], [1, dlatent_broadcast, 1])
+
+    # Output.
+    assert x.dtype == tf.as_dtype(dtype)
+    return tf.identity(x, name='dlatents_out')
+
+#----------------------------------------------------------------------------
+# StyleGAN synthesis network with revised architecture (Figure 2d).
+# Implements progressive growing, but no skip connections or residual nets (Figure 7).
+# Used in configs B-D (Table 1).
+
+def G_synthesis_stylegan_revised(
+    dlatents_in,                        # Input: Disentangled latents (W) [minibatch, num_layers, dlatent_size].
+    dlatent_size        = 512,          # Disentangled latent (W) dimensionality.
+    num_channels        = 3,            # Number of output color channels.
+    resolution          = 1024,         # Output resolution.
+    fmap_base           = 16 << 10,     # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_min            = 1,            # Minimum number of feature maps in any layer.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    randomize_noise     = True,         # True = randomize noise inputs every time (non-deterministic), False = read noise inputs from variables.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    resample_kernel     = [1,3,3,1],    # Low-pass filter to apply when resampling activations. None = no filtering.
+    fused_modconv       = True,         # Implement modulated_conv2d_layer() as a single fused op?
+    structure           = 'auto',       # 'fixed' = no progressive growing, 'linear' = human-readable, 'recursive' = efficient, 'auto' = select automatically.
+    is_template_graph   = False,        # True = template graph constructed by the Network class, False = actual evaluation.
+    force_clean_graph   = False,        # True = construct a clean graph that looks nice in TensorBoard, False = default behavior.
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return np.clip(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_min, fmap_max)
+    if is_template_graph: force_clean_graph = True
+    if force_clean_graph: randomize_noise = False
+    if structure == 'auto': structure = 'linear' if force_clean_graph else 'recursive'
+    act = nonlinearity
+    num_layers = resolution_log2 * 2 - 2
+    images_out = None
+
+    # Primary inputs.
+    dlatents_in.set_shape([None, num_layers, dlatent_size])
+    dlatents_in = tf.cast(dlatents_in, dtype)
+    lod_in = tf.cast(tf.get_variable('lod', initializer=np.float32(0), trainable=False), dtype)
+
+    # Noise inputs.
+    noise_inputs = []
+    for layer_idx in range(num_layers - 1):
+        res = (layer_idx + 5) // 2
+        shape = [1, 1, 2**res, 2**res]
+        noise_inputs.append(tf.get_variable('noise%d' % layer_idx, shape=shape, initializer=tf.initializers.random_normal(), trainable=False))
+
+    # Single convolution layer with all the bells and whistles.
+    def layer(x, layer_idx, fmaps, kernel, up=False):
+        x = modulated_conv2d_layer(x, dlatents_in[:, layer_idx], fmaps=fmaps, kernel=kernel, up=up, resample_kernel=resample_kernel, fused_modconv=fused_modconv)
+        if randomize_noise:
+            noise = tf.random_normal([tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
+        else:
+            noise = tf.cast(noise_inputs[layer_idx], x.dtype)
+        noise_strength = tf.get_variable('noise_strength', shape=[], initializer=tf.initializers.zeros())
+        x += noise * tf.cast(noise_strength, x.dtype)
+        return apply_bias_act(x, act=act)
+
+    # Early layers.
+    with tf.variable_scope('4x4'):
+        with tf.variable_scope('Const'):
+            x = tf.get_variable('const', shape=[1, nf(1), 4, 4], initializer=tf.initializers.random_normal())
+            x = tf.tile(tf.cast(x, dtype), [tf.shape(dlatents_in)[0], 1, 1, 1])
+        with tf.variable_scope('Conv'):
+            x = layer(x, layer_idx=0, fmaps=nf(1), kernel=3)
+
+    # Building blocks for remaining layers.
+    def block(res, x): # res = 3..resolution_log2
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            with tf.variable_scope('Conv0_up'):
+                x = layer(x, layer_idx=res*2-5, fmaps=nf(res-1), kernel=3, up=True)
+            with tf.variable_scope('Conv1'):
+                x = layer(x, layer_idx=res*2-4, fmaps=nf(res-1), kernel=3)
+            return x
+    def torgb(res, x): # res = 2..resolution_log2
+        with tf.variable_scope('ToRGB_lod%d' % (resolution_log2 - res)):
+            return apply_bias_act(modulated_conv2d_layer(x, dlatents_in[:, res*2-3], fmaps=num_channels, kernel=1, demodulate=False, fused_modconv=fused_modconv))
+
+    # Fixed structure: simple and efficient, but does not support progressive growing.
+    if structure == 'fixed':
+        for res in range(3, resolution_log2 + 1):
+            x = block(res, x)
+        images_out = torgb(resolution_log2, x)
+
+    # Linear structure: simple but inefficient.
+    if structure == 'linear':
+        images_out = torgb(2, x)
+        for res in range(3, resolution_log2 + 1):
+            lod = resolution_log2 - res
+            x = block(res, x)
+            img = torgb(res, x)
+            with tf.variable_scope('Upsample_lod%d' % lod):
+                images_out = upsample_2d(images_out)
+            with tf.variable_scope('Grow_lod%d' % lod):
+                images_out = tflib.lerp_clip(img, images_out, lod_in - lod)
+
+    # Recursive structure: complex but efficient.
+    if structure == 'recursive':
+        def cset(cur_lambda, new_cond, new_lambda):
+            return lambda: tf.cond(new_cond, new_lambda, cur_lambda)
+        def grow(x, res, lod):
+            y = block(res, x)
+            img = lambda: naive_upsample_2d(torgb(res, y), factor=2**lod)
+            img = cset(img, (lod_in > lod), lambda: naive_upsample_2d(tflib.lerp(torgb(res, y), upsample_2d(torgb(res - 1, x)), lod_in - lod), factor=2**lod))
+            if lod > 0: img = cset(img, (lod_in < lod), lambda: grow(y, res + 1, lod - 1))
+            return img()
+        images_out = grow(x, 3, resolution_log2 - 3)
+
+    assert images_out.dtype == tf.as_dtype(dtype)
+    return tf.identity(images_out, name='images_out')
+
+#----------------------------------------------------------------------------
+# StyleGAN2 synthesis network (Figure 7).
+# Implements skip connections and residual nets (Figure 7), but no progressive growing.
+# Used in configs E-F (Table 1).
+
+def G_synthesis_stylegan2(
+    dlatents_in,                        # Input: Disentangled latents (W) [minibatch, num_layers, dlatent_size].
+    dlatent_size        = 512,          # Disentangled latent (W) dimensionality.
+    num_channels        = 3,            # Number of output color channels.
+    resolution          = 1024,         # Output resolution.
+    fmap_base           = 16 << 10,     # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_min            = 1,            # Minimum number of feature maps in any layer.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    randomize_noise     = True,         # True = randomize noise inputs every time (non-deterministic), False = read noise inputs from variables.
+    architecture        = 'skip',       # Architecture: 'orig', 'skip', 'resnet'.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    resample_kernel     = [1,3,3,1],    # Low-pass filter to apply when resampling activations. None = no filtering.
+    fused_modconv       = True,         # Implement modulated_conv2d_layer() as a single fused op?
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return np.clip(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_min, fmap_max)
+    assert architecture in ['orig', 'skip', 'resnet']
+    act = nonlinearity
+    num_layers = resolution_log2 * 2 - 2
+    images_out = None
+
+    # Primary inputs.
+    dlatents_in.set_shape([None, num_layers, dlatent_size])
+    dlatents_in = tf.cast(dlatents_in, dtype)
+
+    # Noise inputs.
+    noise_inputs = []
+    for layer_idx in range(num_layers - 1):
+        res = (layer_idx + 5) // 2
+        shape = [1, 1, 2**res, 2**res]
+        noise_inputs.append(tf.get_variable('noise%d' % layer_idx, shape=shape, initializer=tf.initializers.random_normal(), trainable=False))
+
+    # Single convolution layer with all the bells and whistles.
+    def layer(x, layer_idx, fmaps, kernel, up=False):
+        x = modulated_conv2d_layer(x, dlatents_in[:, layer_idx], fmaps=fmaps, kernel=kernel, up=up, resample_kernel=resample_kernel, fused_modconv=fused_modconv)
+        if randomize_noise:
+            noise = tf.random_normal([tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
+        else:
+            noise = tf.cast(noise_inputs[layer_idx], x.dtype)
+        noise_strength = tf.get_variable('noise_strength', shape=[], initializer=tf.initializers.zeros())
+        x += noise * tf.cast(noise_strength, x.dtype)
+        return apply_bias_act(x, act=act)
+
+    # Building blocks for main layers.
+    def block(x, res): # res = 3..resolution_log2
+        t = x
+        with tf.variable_scope('Conv0_up'):
+            x = layer(x, layer_idx=res*2-5, fmaps=nf(res-1), kernel=3, up=True)
+        with tf.variable_scope('Conv1'):
+            x = layer(x, layer_idx=res*2-4, fmaps=nf(res-1), kernel=3)
+        if architecture == 'resnet':
+            with tf.variable_scope('Skip'):
+                t = conv2d_layer(t, fmaps=nf(res-1), kernel=1, up=True, resample_kernel=resample_kernel)
+                x = (x + t) * (1 / np.sqrt(2))
+        return x
+    def upsample(y):
+        with tf.variable_scope('Upsample'):
+            return upsample_2d(y, k=resample_kernel)
+    def torgb(x, y, res): # res = 2..resolution_log2
+        with tf.variable_scope('ToRGB'):
+            t = apply_bias_act(modulated_conv2d_layer(x, dlatents_in[:, res*2-3], fmaps=num_channels, kernel=1, demodulate=False, fused_modconv=fused_modconv))
+            return t if y is None else y + t
+
+    # Early layers.
+    y = None
+    with tf.variable_scope('4x4'):
+        with tf.variable_scope('Const'):
+            x = tf.get_variable('const', shape=[1, nf(1), 4, 4], initializer=tf.initializers.random_normal())
+            x = tf.tile(tf.cast(x, dtype), [tf.shape(dlatents_in)[0], 1, 1, 1])
+        with tf.variable_scope('Conv'):
+            x = layer(x, layer_idx=0, fmaps=nf(1), kernel=3)
+        if architecture == 'skip':
+            y = torgb(x, y, 2)
+
+    # Main layers.
+    for res in range(3, resolution_log2 + 1):
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            x = block(x, res)
+            if architecture == 'skip':
+                y = upsample(y)
+            if architecture == 'skip' or res == resolution_log2:
+                y = torgb(x, y, res)
+    images_out = y
+
+    assert images_out.dtype == tf.as_dtype(dtype)
+    return tf.identity(images_out, name='images_out')
+
+#----------------------------------------------------------------------------
+# Original StyleGAN discriminator.
+# Used in configs B-D (Table 1).
+
+def D_stylegan(
+    images_in,                          # First input: Images [minibatch, channel, height, width].
+    labels_in,                          # Second input: Labels [minibatch, label_size].
+    num_channels        = 3,            # Number of input color channels. Overridden based on dataset.
+    resolution          = 1024,         # Input resolution. Overridden based on dataset.
+    label_size          = 0,            # Dimensionality of the labels, 0 if no labels. Overridden based on dataset.
+    fmap_base           = 16 << 10,     # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_min            = 1,            # Minimum number of feature maps in any layer.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+    mbstd_group_size    = 4,            # Group size for the minibatch standard deviation layer, 0 = disable.
+    mbstd_num_features  = 1,            # Number of features for the minibatch standard deviation layer.
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    resample_kernel     = [1,3,3,1],    # Low-pass filter to apply when resampling activations. None = no filtering.
+    structure           = 'auto',       # 'fixed' = no progressive growing, 'linear' = human-readable, 'recursive' = efficient, 'auto' = select automatically.
+    is_template_graph   = False,        # True = template graph constructed by the Network class, False = actual evaluation.
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return np.clip(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_min, fmap_max)
+    if structure == 'auto': structure = 'linear' if is_template_graph else 'recursive'
+    act = nonlinearity
+
+    images_in.set_shape([None, num_channels, resolution, resolution])
+    labels_in.set_shape([None, label_size])
+    images_in = tf.cast(images_in, dtype)
+    labels_in = tf.cast(labels_in, dtype)
+    lod_in = tf.cast(tf.get_variable('lod', initializer=np.float32(0.0), trainable=False), dtype)
+
+    # Building blocks for spatial layers.
+    def fromrgb(x, res): # res = 2..resolution_log2
+        with tf.variable_scope('FromRGB_lod%d' % (resolution_log2 - res)):
+            return apply_bias_act(conv2d_layer(x, fmaps=nf(res-1), kernel=1), act=act)
+    def block(x, res): # res = 2..resolution_log2
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            with tf.variable_scope('Conv0'):
+                x = apply_bias_act(conv2d_layer(x, fmaps=nf(res-1), kernel=3), act=act)
+            with tf.variable_scope('Conv1_down'):
+                x = apply_bias_act(conv2d_layer(x, fmaps=nf(res-2), kernel=3, down=True, resample_kernel=resample_kernel), act=act)
+            return x
+
+    # Fixed structure: simple and efficient, but does not support progressive growing.
+    if structure == 'fixed':
+        x = fromrgb(images_in, resolution_log2)
+        for res in range(resolution_log2, 2, -1):
+            x = block(x, res)
+
+    # Linear structure: simple but inefficient.
+    if structure == 'linear':
+        img = images_in
+        x = fromrgb(img, resolution_log2)
+        for res in range(resolution_log2, 2, -1):
+            lod = resolution_log2 - res
+            x = block(x, res)
+            with tf.variable_scope('Downsample_lod%d' % lod):
+                img = downsample_2d(img)
+            y = fromrgb(img, res - 1)
+            with tf.variable_scope('Grow_lod%d' % lod):
+                x = tflib.lerp_clip(x, y, lod_in - lod)
+
+    # Recursive structure: complex but efficient.
+    if structure == 'recursive':
+        def cset(cur_lambda, new_cond, new_lambda):
+            return lambda: tf.cond(new_cond, new_lambda, cur_lambda)
+        def grow(res, lod):
+            x = lambda: fromrgb(naive_downsample_2d(images_in, factor=2**lod), res)
+            if lod > 0: x = cset(x, (lod_in < lod), lambda: grow(res + 1, lod - 1))
+            x = block(x(), res); y = lambda: x
+            y = cset(y, (lod_in > lod), lambda: tflib.lerp(x, fromrgb(naive_downsample_2d(images_in, factor=2**(lod+1)), res - 1), lod_in - lod))
+            return y()
+        x = grow(3, resolution_log2 - 3)
+
+    # Final layers at 4x4 resolution.
+    with tf.variable_scope('4x4'):
+        if mbstd_group_size > 1:
+            with tf.variable_scope('MinibatchStddev'):
+                x = minibatch_stddev_layer(x, mbstd_group_size, mbstd_num_features)
+        with tf.variable_scope('Conv'):
+            x = apply_bias_act(conv2d_layer(x, fmaps=nf(1), kernel=3), act=act)
+        with tf.variable_scope('Dense0'):
+            x = apply_bias_act(dense_layer(x, fmaps=nf(0)), act=act)
+
+    # Output layer with label conditioning from "Which Training Methods for GANs do actually Converge?"
+    with tf.variable_scope('Output'):
+        x = apply_bias_act(dense_layer(x, fmaps=max(labels_in.shape[1], 1)))
+        if labels_in.shape[1] > 0:
+            x = tf.reduce_sum(x * labels_in, axis=1, keepdims=True)
+    scores_out = x
+
+    # Output.
+    assert scores_out.dtype == tf.as_dtype(dtype)
+    scores_out = tf.identity(scores_out, name='scores_out')
+    return scores_out
+
+#----------------------------------------------------------------------------
+# StyleGAN2 discriminator (Figure 7).
+# Implements skip connections and residual nets (Figure 7), but no progressive growing.
+# Used in configs E-F (Table 1).
+
+def D_stylegan2(
+    images_in,                          # First input: Images [minibatch, channel, height, width].
+    labels_in,                          # Second input: Labels [minibatch, label_size].
+    num_channels        = 3,            # Number of input color channels. Overridden based on dataset.
+    resolution          = 1024,         # Input resolution. Overridden based on dataset.
+    label_size          = 0,            # Dimensionality of the labels, 0 if no labels. Overridden based on dataset.
+    fmap_base           = 16 << 10,     # Overall multiplier for the number of feature maps.
+    fmap_decay          = 1.0,          # log2 feature map reduction when doubling the resolution.
+    fmap_min            = 1,            # Minimum number of feature maps in any layer.
+    fmap_max            = 512,          # Maximum number of feature maps in any layer.
+    architecture        = 'resnet',     # Architecture: 'orig', 'skip', 'resnet'.
+    nonlinearity        = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+    mbstd_group_size    = 4,            # Group size for the minibatch standard deviation layer, 0 = disable.
+    mbstd_num_features  = 1,            # Number of features for the minibatch standard deviation layer.
+    dtype               = 'float32',    # Data type to use for activations and outputs.
+    resample_kernel     = [1,3,3,1],    # Low-pass filter to apply when resampling activations. None = no filtering.
+    **_kwargs):                         # Ignore unrecognized keyword args.
+
+    resolution_log2 = int(np.log2(resolution))
+    assert resolution == 2**resolution_log2 and resolution >= 4
+    def nf(stage): return np.clip(int(fmap_base / (2.0 ** (stage * fmap_decay))), fmap_min, fmap_max)
+    assert architecture in ['orig', 'skip', 'resnet']
+    act = nonlinearity
+
+    images_in.set_shape([None, num_channels, resolution, resolution])
+    labels_in.set_shape([None, label_size])
+    images_in = tf.cast(images_in, dtype)
+    labels_in = tf.cast(labels_in, dtype)
+
+    # Building blocks for main layers.
+    def fromrgb(x, y, res): # res = 2..resolution_log2
+        with tf.variable_scope('FromRGB'):
+            t = apply_bias_act(conv2d_layer(y, fmaps=nf(res-1), kernel=1), act=act)
+            return t if x is None else x + t
+    def block(x, res): # res = 2..resolution_log2
+        t = x
+        with tf.variable_scope('Conv0'):
+            x = apply_bias_act(conv2d_layer(x, fmaps=nf(res-1), kernel=3), act=act)
+        with tf.variable_scope('Conv1_down'):
+            x = apply_bias_act(conv2d_layer(x, fmaps=nf(res-2), kernel=3, down=True, resample_kernel=resample_kernel), act=act)
+        if architecture == 'resnet':
+            with tf.variable_scope('Skip'):
+                t = conv2d_layer(t, fmaps=nf(res-2), kernel=1, down=True, resample_kernel=resample_kernel)
+                x = (x + t) * (1 / np.sqrt(2))
+        return x
+    def downsample(y):
+        with tf.variable_scope('Downsample'):
+            return downsample_2d(y, k=resample_kernel)
+
+    # Main layers.
+    x = None
+    y = images_in
+    for res in range(resolution_log2, 2, -1):
+        with tf.variable_scope('%dx%d' % (2**res, 2**res)):
+            if architecture == 'skip' or res == resolution_log2:
+                x = fromrgb(x, y, res)
+            x = block(x, res)
+            if architecture == 'skip':
+                y = downsample(y)
+
+    # Final layers.
+    with tf.variable_scope('4x4'):
+        if architecture == 'skip':
+            x = fromrgb(x, y, 2)
+        if mbstd_group_size > 1:
+            with tf.variable_scope('MinibatchStddev'):
+                x = minibatch_stddev_layer(x, mbstd_group_size, mbstd_num_features)
+        with tf.variable_scope('Conv'):
+            x = apply_bias_act(conv2d_layer(x, fmaps=nf(1), kernel=3), act=act)
+        with tf.variable_scope('Dense0'):
+            x = apply_bias_act(dense_layer(x, fmaps=nf(0)), act=act)
+
+    # Output layer with label conditioning from "Which Training Methods for GANs do actually Converge?"
+    with tf.variable_scope('Output'):
+        x = apply_bias_act(dense_layer(x, fmaps=max(labels_in.shape[1], 1)))
+        if labels_in.shape[1] > 0:
+            x = tf.reduce_sum(x * labels_in, axis=1, keepdims=True)
+    scores_out = x
+
+    # Output.
+    assert scores_out.dtype == tf.as_dtype(dtype)
+    scores_out = tf.identity(scores_out, name='scores_out')
+    return scores_out
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/external/stylegan2/training/training_loop.py b/reconstruction/ostec/external/stylegan2/training/training_loop.py
new file mode 100644
index 000000000..c2d88cf03
--- /dev/null
+++ b/reconstruction/ostec/external/stylegan2/training/training_loop.py
@@ -0,0 +1,356 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Main training script."""
+
+import numpy as np
+import tensorflow as tf
+import dnnlib
+import dnnlib.tflib as tflib
+from dnnlib.tflib.autosummary import autosummary
+
+from training import dataset
+from training import misc
+from metrics import metric_base
+
+#----------------------------------------------------------------------------
+# Just-in-time processing of training images before feeding them to the networks.
+
+def process_reals(x, labels, lod, mirror_augment, drange_data, drange_net):
+    with tf.name_scope('DynamicRange'):
+        x = tf.cast(x, tf.float32)
+        x = misc.adjust_dynamic_range(x, drange_data, drange_net)
+    if mirror_augment:
+        with tf.name_scope('MirrorAugment'):
+            x = tf.where(tf.random_uniform([tf.shape(x)[0]]) < 0.5, x, tf.reverse(x, [3]))
+    with tf.name_scope('FadeLOD'): # Smooth crossfade between consecutive levels-of-detail.
+        s = tf.shape(x)
+        y = tf.reshape(x, [-1, s[1], s[2]//2, 2, s[3]//2, 2])
+        y = tf.reduce_mean(y, axis=[3, 5], keepdims=True)
+        y = tf.tile(y, [1, 1, 1, 2, 1, 2])
+        y = tf.reshape(y, [-1, s[1], s[2], s[3]])
+        x = tflib.lerp(x, y, lod - tf.floor(lod))
+    with tf.name_scope('UpscaleLOD'): # Upscale to match the expected input/output size of the networks.
+        s = tf.shape(x)
+        factor = tf.cast(2 ** tf.floor(lod), tf.int32)
+        x = tf.reshape(x, [-1, s[1], s[2], 1, s[3], 1])
+        x = tf.tile(x, [1, 1, 1, factor, 1, factor])
+        x = tf.reshape(x, [-1, s[1], s[2] * factor, s[3] * factor])
+    return x, labels
+
+#----------------------------------------------------------------------------
+# Evaluate time-varying training parameters.
+
+def training_schedule(
+    cur_nimg,
+    training_set,
+    lod_initial_resolution  = None,     # Image resolution used at the beginning.
+    lod_training_kimg       = 600,      # Thousands of real images to show before doubling the resolution.
+    lod_transition_kimg     = 600,      # Thousands of real images to show when fading in new layers.
+    minibatch_size_base     = 32,       # Global minibatch size.
+    minibatch_size_dict     = {},       # Resolution-specific overrides.
+    minibatch_gpu_base      = 4,        # Number of samples processed at a time by one GPU.
+    minibatch_gpu_dict      = {},       # Resolution-specific overrides.
+    G_lrate_base            = 0.002,    # Learning rate for the generator.
+    G_lrate_dict            = {},       # Resolution-specific overrides.
+    D_lrate_base            = 0.002,    # Learning rate for the discriminator.
+    D_lrate_dict            = {},       # Resolution-specific overrides.
+    lrate_rampup_kimg       = 0,        # Duration of learning rate ramp-up.
+    tick_kimg_base          = 4,        # Default interval of progress snapshots.
+    tick_kimg_dict          = {8:28, 16:24, 32:20, 64:16, 128:12, 256:8, 512:6, 1024:4}): # Resolution-specific overrides.
+
+    # Initialize result dict.
+    s = dnnlib.EasyDict()
+    s.kimg = cur_nimg / 1000.0
+
+    # Training phase.
+    phase_dur = lod_training_kimg + lod_transition_kimg
+    phase_idx = int(np.floor(s.kimg / phase_dur)) if phase_dur > 0 else 0
+    phase_kimg = s.kimg - phase_idx * phase_dur
+
+    # Level-of-detail and resolution.
+    if lod_initial_resolution is None:
+        s.lod = 0.0
+    else:
+        s.lod = training_set.resolution_log2
+        s.lod -= np.floor(np.log2(lod_initial_resolution))
+        s.lod -= phase_idx
+        if lod_transition_kimg > 0:
+            s.lod -= max(phase_kimg - lod_training_kimg, 0.0) / lod_transition_kimg
+        s.lod = max(s.lod, 0.0)
+    s.resolution = 2 ** (training_set.resolution_log2 - int(np.floor(s.lod)))
+
+    # Minibatch size.
+    s.minibatch_size = minibatch_size_dict.get(s.resolution, minibatch_size_base)
+    s.minibatch_gpu = minibatch_gpu_dict.get(s.resolution, minibatch_gpu_base)
+
+    # Learning rate.
+    s.G_lrate = G_lrate_dict.get(s.resolution, G_lrate_base)
+    s.D_lrate = D_lrate_dict.get(s.resolution, D_lrate_base)
+    if lrate_rampup_kimg > 0:
+        rampup = min(s.kimg / lrate_rampup_kimg, 1.0)
+        s.G_lrate *= rampup
+        s.D_lrate *= rampup
+
+    # Other parameters.
+    s.tick_kimg = tick_kimg_dict.get(s.resolution, tick_kimg_base)
+    return s
+
+#----------------------------------------------------------------------------
+# Main training script.
+
+def training_loop(
+    G_args                  = {},       # Options for generator network.
+    D_args                  = {},       # Options for discriminator network.
+    G_opt_args              = {},       # Options for generator optimizer.
+    D_opt_args              = {},       # Options for discriminator optimizer.
+    G_loss_args             = {},       # Options for generator loss.
+    D_loss_args             = {},       # Options for discriminator loss.
+    dataset_args            = {},       # Options for dataset.load_dataset().
+    sched_args              = {},       # Options for train.TrainingSchedule.
+    grid_args               = {},       # Options for train.setup_snapshot_image_grid().
+    metric_arg_list         = [],       # Options for MetricGroup.
+    tf_config               = {},       # Options for tflib.init_tf().
+    data_dir                = None,     # Directory to load datasets from.
+    G_smoothing_kimg        = 10.0,     # Half-life of the running average of generator weights.
+    minibatch_repeats       = 4,        # Number of minibatches to run before adjusting training parameters.
+    lazy_regularization     = True,     # Perform regularization as a separate training step?
+    G_reg_interval          = 4,        # How often the perform regularization for G? Ignored if lazy_regularization=False.
+    D_reg_interval          = 16,       # How often the perform regularization for D? Ignored if lazy_regularization=False.
+    reset_opt_for_new_lod   = True,     # Reset optimizer internal state (e.g. Adam moments) when new layers are introduced?
+    total_kimg              = 25000,    # Total length of the training, measured in thousands of real images.
+    mirror_augment          = False,    # Enable mirror augment?
+    drange_net              = [-1,1],   # Dynamic range used when feeding image data to the networks.
+    image_snapshot_ticks    = 50,       # How often to save image snapshots? None = only save 'reals.png' and 'fakes-init.png'.
+    network_snapshot_ticks  = 50,       # How often to save network snapshots? None = only save 'networks-final.pkl'.
+    save_tf_graph           = False,    # Include full TensorFlow computation graph in the tfevents file?
+    save_weight_histograms  = False,    # Include weight histograms in the tfevents file?
+    resume_pkl              = None,     # Network pickle to resume training from, None = train from scratch.
+    resume_kimg             = 0.0,      # Assumed training progress at the beginning. Affects reporting and training schedule.
+    resume_time             = 0.0,      # Assumed wallclock time at the beginning. Affects reporting.
+    resume_with_new_nets    = False):   # Construct new networks according to G_args and D_args before resuming training?
+
+    # Initialize dnnlib and TensorFlow.
+    tflib.init_tf(tf_config)
+    num_gpus = dnnlib.submit_config.num_gpus
+
+    # Load training set.
+    training_set = dataset.load_dataset(data_dir=dnnlib.convert_path(data_dir), verbose=True, **dataset_args)
+    grid_size, grid_reals, grid_labels = misc.setup_snapshot_image_grid(training_set, **grid_args)
+    misc.save_image_grid(grid_reals, dnnlib.make_run_dir_path('reals.png'), drange=training_set.dynamic_range, grid_size=grid_size)
+
+    # Construct or load networks.
+    with tf.device('/gpu:0'):
+        if resume_pkl is None or resume_with_new_nets:
+            print('Constructing networks...')
+            G = tflib.Network('G', num_channels=training_set.shape[0], resolution=training_set.shape[1], label_size=training_set.label_size, **G_args)
+            D = tflib.Network('D', num_channels=training_set.shape[0], resolution=training_set.shape[1], label_size=training_set.label_size, **D_args)
+            Gs = G.clone('Gs')
+        if resume_pkl is not None:
+            print('Loading networks from "%s"...' % resume_pkl)
+            rG, rD, rGs = misc.load_pkl(resume_pkl)
+            if resume_with_new_nets: G.copy_vars_from(rG); D.copy_vars_from(rD); Gs.copy_vars_from(rGs)
+            else: G = rG; D = rD; Gs = rGs
+
+    # Print layers and generate initial image snapshot.
+    G.print_layers(); D.print_layers()
+    sched = training_schedule(cur_nimg=total_kimg*1000, training_set=training_set, **sched_args)
+    grid_latents = np.random.randn(np.prod(grid_size), *G.input_shape[1:])
+    grid_fakes = Gs.run(grid_latents, grid_labels, is_validation=True, minibatch_size=sched.minibatch_gpu)
+    misc.save_image_grid(grid_fakes, dnnlib.make_run_dir_path('fakes_init.png'), drange=drange_net, grid_size=grid_size)
+
+    # Setup training inputs.
+    print('Building TensorFlow graph...')
+    with tf.name_scope('Inputs'), tf.device('/cpu:0'):
+        lod_in               = tf.placeholder(tf.float32, name='lod_in', shape=[])
+        lrate_in             = tf.placeholder(tf.float32, name='lrate_in', shape=[])
+        minibatch_size_in    = tf.placeholder(tf.int32, name='minibatch_size_in', shape=[])
+        minibatch_gpu_in     = tf.placeholder(tf.int32, name='minibatch_gpu_in', shape=[])
+        minibatch_multiplier = minibatch_size_in // (minibatch_gpu_in * num_gpus)
+        Gs_beta              = 0.5 ** tf.div(tf.cast(minibatch_size_in, tf.float32), G_smoothing_kimg * 1000.0) if G_smoothing_kimg > 0.0 else 0.0
+
+    # Setup optimizers.
+    G_opt_args = dict(G_opt_args)
+    D_opt_args = dict(D_opt_args)
+    for args, reg_interval in [(G_opt_args, G_reg_interval), (D_opt_args, D_reg_interval)]:
+        args['minibatch_multiplier'] = minibatch_multiplier
+        args['learning_rate'] = lrate_in
+        if lazy_regularization:
+            mb_ratio = reg_interval / (reg_interval + 1)
+            args['learning_rate'] *= mb_ratio
+            if 'beta1' in args: args['beta1'] **= mb_ratio
+            if 'beta2' in args: args['beta2'] **= mb_ratio
+    G_opt = tflib.Optimizer(name='TrainG', **G_opt_args)
+    D_opt = tflib.Optimizer(name='TrainD', **D_opt_args)
+    G_reg_opt = tflib.Optimizer(name='RegG', share=G_opt, **G_opt_args)
+    D_reg_opt = tflib.Optimizer(name='RegD', share=D_opt, **D_opt_args)
+
+    # Build training graph for each GPU.
+    data_fetch_ops = []
+    for gpu in range(num_gpus):
+        with tf.name_scope('GPU%d' % gpu), tf.device('/gpu:%d' % gpu):
+
+            # Create GPU-specific shadow copies of G and D.
+            G_gpu = G if gpu == 0 else G.clone(G.name + '_shadow')
+            D_gpu = D if gpu == 0 else D.clone(D.name + '_shadow')
+
+            # Fetch training data via temporary variables.
+            with tf.name_scope('DataFetch'):
+                sched = training_schedule(cur_nimg=int(resume_kimg*1000), training_set=training_set, **sched_args)
+                reals_var = tf.Variable(name='reals', trainable=False, initial_value=tf.zeros([sched.minibatch_gpu] + training_set.shape))
+                labels_var = tf.Variable(name='labels', trainable=False, initial_value=tf.zeros([sched.minibatch_gpu, training_set.label_size]))
+                reals_write, labels_write = training_set.get_minibatch_tf()
+                reals_write, labels_write = process_reals(reals_write, labels_write, lod_in, mirror_augment, training_set.dynamic_range, drange_net)
+                reals_write = tf.concat([reals_write, reals_var[minibatch_gpu_in:]], axis=0)
+                labels_write = tf.concat([labels_write, labels_var[minibatch_gpu_in:]], axis=0)
+                data_fetch_ops += [tf.assign(reals_var, reals_write)]
+                data_fetch_ops += [tf.assign(labels_var, labels_write)]
+                reals_read = reals_var[:minibatch_gpu_in]
+                labels_read = labels_var[:minibatch_gpu_in]
+
+            # Evaluate loss functions.
+            lod_assign_ops = []
+            if 'lod' in G_gpu.vars: lod_assign_ops += [tf.assign(G_gpu.vars['lod'], lod_in)]
+            if 'lod' in D_gpu.vars: lod_assign_ops += [tf.assign(D_gpu.vars['lod'], lod_in)]
+            with tf.control_dependencies(lod_assign_ops):
+                with tf.name_scope('G_loss'):
+                    G_loss, G_reg = dnnlib.util.call_func_by_name(G=G_gpu, D=D_gpu, opt=G_opt, training_set=training_set, minibatch_size=minibatch_gpu_in, **G_loss_args)
+                with tf.name_scope('D_loss'):
+                    D_loss, D_reg = dnnlib.util.call_func_by_name(G=G_gpu, D=D_gpu, opt=D_opt, training_set=training_set, minibatch_size=minibatch_gpu_in, reals=reals_read, labels=labels_read, **D_loss_args)
+
+            # Register gradients.
+            if not lazy_regularization:
+                if G_reg is not None: G_loss += G_reg
+                if D_reg is not None: D_loss += D_reg
+            else:
+                if G_reg is not None: G_reg_opt.register_gradients(tf.reduce_mean(G_reg * G_reg_interval), G_gpu.trainables)
+                if D_reg is not None: D_reg_opt.register_gradients(tf.reduce_mean(D_reg * D_reg_interval), D_gpu.trainables)
+            G_opt.register_gradients(tf.reduce_mean(G_loss), G_gpu.trainables)
+            D_opt.register_gradients(tf.reduce_mean(D_loss), D_gpu.trainables)
+
+    # Setup training ops.
+    data_fetch_op = tf.group(*data_fetch_ops)
+    G_train_op = G_opt.apply_updates()
+    D_train_op = D_opt.apply_updates()
+    G_reg_op = G_reg_opt.apply_updates(allow_no_op=True)
+    D_reg_op = D_reg_opt.apply_updates(allow_no_op=True)
+    Gs_update_op = Gs.setup_as_moving_average_of(G, beta=Gs_beta)
+
+    # Finalize graph.
+    with tf.device('/gpu:0'):
+        try:
+            peak_gpu_mem_op = tf.contrib.memory_stats.MaxBytesInUse()
+        except tf.errors.NotFoundError:
+            peak_gpu_mem_op = tf.constant(0)
+    tflib.init_uninitialized_vars()
+
+    print('Initializing logs...')
+    summary_log = tf.summary.FileWriter(dnnlib.make_run_dir_path())
+    if save_tf_graph:
+        summary_log.add_graph(tf.get_default_graph())
+    if save_weight_histograms:
+        G.setup_weight_histograms(); D.setup_weight_histograms()
+    metrics = metric_base.MetricGroup(metric_arg_list)
+
+    print('Training for %d kimg...\n' % total_kimg)
+    dnnlib.RunContext.get().update('', cur_epoch=resume_kimg, max_epoch=total_kimg)
+    maintenance_time = dnnlib.RunContext.get().get_last_update_interval()
+    cur_nimg = int(resume_kimg * 1000)
+    cur_tick = -1
+    tick_start_nimg = cur_nimg
+    prev_lod = -1.0
+    running_mb_counter = 0
+    while cur_nimg < total_kimg * 1000:
+        if dnnlib.RunContext.get().should_stop(): break
+
+        # Choose training parameters and configure training ops.
+        sched = training_schedule(cur_nimg=cur_nimg, training_set=training_set, **sched_args)
+        assert sched.minibatch_size % (sched.minibatch_gpu * num_gpus) == 0
+        training_set.configure(sched.minibatch_gpu, sched.lod)
+        if reset_opt_for_new_lod:
+            if np.floor(sched.lod) != np.floor(prev_lod) or np.ceil(sched.lod) != np.ceil(prev_lod):
+                G_opt.reset_optimizer_state(); D_opt.reset_optimizer_state()
+        prev_lod = sched.lod
+
+        # Run training ops.
+        feed_dict = {lod_in: sched.lod, lrate_in: sched.G_lrate, minibatch_size_in: sched.minibatch_size, minibatch_gpu_in: sched.minibatch_gpu}
+        for _repeat in range(minibatch_repeats):
+            rounds = range(0, sched.minibatch_size, sched.minibatch_gpu * num_gpus)
+            run_G_reg = (lazy_regularization and running_mb_counter % G_reg_interval == 0)
+            run_D_reg = (lazy_regularization and running_mb_counter % D_reg_interval == 0)
+            cur_nimg += sched.minibatch_size
+            running_mb_counter += 1
+
+            # Fast path without gradient accumulation.
+            if len(rounds) == 1:
+                tflib.run([G_train_op, data_fetch_op], feed_dict)
+                if run_G_reg:
+                    tflib.run(G_reg_op, feed_dict)
+                tflib.run([D_train_op, Gs_update_op], feed_dict)
+                if run_D_reg:
+                    tflib.run(D_reg_op, feed_dict)
+
+            # Slow path with gradient accumulation.
+            else:
+                for _round in rounds:
+                    tflib.run(G_train_op, feed_dict)
+                if run_G_reg:
+                    for _round in rounds:
+                        tflib.run(G_reg_op, feed_dict)
+                tflib.run(Gs_update_op, feed_dict)
+                for _round in rounds:
+                    tflib.run(data_fetch_op, feed_dict)
+                    tflib.run(D_train_op, feed_dict)
+                if run_D_reg:
+                    for _round in rounds:
+                        tflib.run(D_reg_op, feed_dict)
+
+        # Perform maintenance tasks once per tick.
+        done = (cur_nimg >= total_kimg * 1000)
+        if cur_tick < 0 or cur_nimg >= tick_start_nimg + sched.tick_kimg * 1000 or done:
+            cur_tick += 1
+            tick_kimg = (cur_nimg - tick_start_nimg) / 1000.0
+            tick_start_nimg = cur_nimg
+            tick_time = dnnlib.RunContext.get().get_time_since_last_update()
+            total_time = dnnlib.RunContext.get().get_time_since_start() + resume_time
+
+            # Report progress.
+            print('tick %-5d kimg %-8.1f lod %-5.2f minibatch %-4d time %-12s sec/tick %-7.1f sec/kimg %-7.2f maintenance %-6.1f gpumem %.1f' % (
+                autosummary('Progress/tick', cur_tick),
+                autosummary('Progress/kimg', cur_nimg / 1000.0),
+                autosummary('Progress/lod', sched.lod),
+                autosummary('Progress/minibatch', sched.minibatch_size),
+                dnnlib.util.format_time(autosummary('Timing/total_sec', total_time)),
+                autosummary('Timing/sec_per_tick', tick_time),
+                autosummary('Timing/sec_per_kimg', tick_time / tick_kimg),
+                autosummary('Timing/maintenance_sec', maintenance_time),
+                autosummary('Resources/peak_gpu_mem_gb', peak_gpu_mem_op.eval() / 2**30)))
+            autosummary('Timing/total_hours', total_time / (60.0 * 60.0))
+            autosummary('Timing/total_days', total_time / (24.0 * 60.0 * 60.0))
+
+            # Save snapshots.
+            if image_snapshot_ticks is not None and (cur_tick % image_snapshot_ticks == 0 or done):
+                grid_fakes = Gs.run(grid_latents, grid_labels, is_validation=True, minibatch_size=sched.minibatch_gpu)
+                misc.save_image_grid(grid_fakes, dnnlib.make_run_dir_path('fakes%06d.png' % (cur_nimg // 1000)), drange=drange_net, grid_size=grid_size)
+            if network_snapshot_ticks is not None and (cur_tick % network_snapshot_ticks == 0 or done):
+                pkl = dnnlib.make_run_dir_path('network-snapshot-%06d.pkl' % (cur_nimg // 1000))
+                misc.save_pkl((G, D, Gs), pkl)
+                metrics.run(pkl, run_dir=dnnlib.make_run_dir_path(), data_dir=dnnlib.convert_path(data_dir), num_gpus=num_gpus, tf_config=tf_config)
+
+            # Update summaries and RunContext.
+            metrics.update_autosummaries()
+            tflib.autosummary.save_summaries(summary_log, cur_nimg)
+            dnnlib.RunContext.get().update('%.2f' % sched.lod, cur_epoch=cur_nimg // 1000, max_epoch=total_kimg)
+            maintenance_time = dnnlib.RunContext.get().get_last_update_interval() - tick_time
+
+    # Save final snapshot.
+    misc.save_pkl((G, D, Gs), dnnlib.make_run_dir_path('network-final.pkl'))
+
+    # All done.
+    summary_log.close()
+    training_set.close()
+
+#----------------------------------------------------------------------------
diff --git a/reconstruction/ostec/run_ostec.py b/reconstruction/ostec/run_ostec.py
new file mode 100644
index 000000000..e59bb8c13
--- /dev/null
+++ b/reconstruction/ostec/run_ostec.py
@@ -0,0 +1,93 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import time
+import os
+import glob
+from random import shuffle
+import argparse
+from argparse import Namespace
+import menpo.io as mio
+import menpo.image
+import cv2
+import sys
+sys.path.append("external/stylegan2")
+sys.path.append("external/deep3dfacerecon")
+sys.path.append("external/graphonomy")
+from core.operator import Operator
+from core.config import get_config
+import numpy as np
+from utils.utils import im_menpo2PIL, fix_obj
+from external.face_detector.detect_face import Face_Detector
+from FaceHairMask.MaskExtractor import MaskExtractor
+from menpo.shape import TexturedTriMesh
+import menpo3d.io as m3io
+
+def main(args):
+    source_dir = args.source_dir
+    save_dir = args.save_dir
+    os.makedirs(save_dir,exist_ok=True)
+    operator = Operator(args)
+    detector = Face_Detector()
+    if not args.ganfit:
+        from external.deep3dfacerecon.ostec_api import Deep3dModel
+        deep3dmodel = Deep3dModel()
+    maskExtractor = MaskExtractor()
+
+
+    # while True:
+    for ext in ['.png', '.jpg']:
+        print('Scanning paths...')
+        paths = glob.glob(source_dir + '/*' + ext)
+        shuffle(paths)
+        for path in paths:
+            # try: # To avoid detection errors on large datasets
+            save_path = path.replace(source_dir, save_dir)
+            pkl_path = path.replace(ext,'.pkl')
+            if not os.path.isfile(save_path.replace(ext, '.png')):
+                print('Started: ' + path)
+                start = time.time()
+
+                img = menpo.image.Image(np.transpose(cv2.imread(path)[:,:,::-1],[2,0,1])/255.0)
+
+                if args.ganfit and not os.path.isfile(pkl_path):
+                    raise Exception('Reconstruction from GANfit mode is activated and no GANFit reconstruction pickle file has been found! Either Remove --ganfit flag or Run GANFit first.')
+
+                if os.path.isfile(pkl_path): # GANFit mode
+                    fitting = mio.import_pickle(pkl_path)
+
+                else: # Deep3dReconstruction mode
+                    _, lms = detector.face_detection((img.pixels_with_channels_at_back() * 255).astype(np.uint8))
+                    fitting = deep3dmodel.recontruct(im_menpo2PIL(img), lms)
+                    img = menpo.image.Image(fitting['input'])
+
+                _, face_mask = maskExtractor.main(img)
+
+                final_uv, results_dict = operator.run(img, fitting, face_mask)
+                tmesh = TexturedTriMesh(fitting['vertices'],operator.tcoords.points,final_uv,operator.uv_trilist)
+                m3io.export_textured_mesh(tmesh,save_path.replace(ext, '.obj'),texture_extension='.png')
+                fix_obj(save_path.replace(ext, '.obj'))
+                # mio.export_image(final_uv, save_path.replace(ext, '.png'))
+
+                if args.frontalize:
+                    mio.export_image(results_dict['frontal'], save_path.replace(ext,'_frontal.png'))
+                if args.pickle:
+                    mio.export_pickle(results_dict, save_path.replace(ext,'.pkl'))
+
+                print('Total Processing Time : %.2f secs' % (time.time() - start))
+
+            # except Exception as inst:
+            #     print(type(inst))  # the exception instance
+            #     print(inst.args)  # arguments stored in .args
+            #     print(inst)  # __str__ allows args to be printed directly,
+
+if __name__ == "__main__":
+    args, unparsed = get_config()
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--source_dir', help='Directory of input 2D images')
+    parser.add_argument('--save_dir', help='Directory to save synthesized UVs')
+    args2 = parser.parse_args(unparsed)
+    args = Namespace(**vars(args), **vars(args2))
+    main(args)
diff --git a/reconstruction/ostec/utils/align2stylegan.py b/reconstruction/ostec/utils/align2stylegan.py
new file mode 100644
index 000000000..9c45953ea
--- /dev/null
+++ b/reconstruction/ostec/utils/align2stylegan.py
@@ -0,0 +1,234 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import numpy as np
+import scipy.ndimage
+import PIL.Image
+
+def create_perspective_transform_matrix(src, dst):
+    """ Creates a perspective transformation matrix which transforms points
+        in quadrilateral ``src`` to the corresponding points on quadrilateral
+        ``dst``.
+
+        Will raise a ``np.linalg.LinAlgError`` on invalid input.
+        """
+    # See:
+    # * http://xenia.media.mit.edu/~cwren/interpolator/
+    # * http://stackoverflow.com/a/14178717/71522
+    in_matrix = []
+    for (x, y), (X, Y) in zip(src, dst):
+        in_matrix.extend([
+            [x, y, 1, 0, 0, 0, -X * x, -X * y],
+            [0, 0, 0, x, y, 1, -Y * x, -Y * y],
+        ])
+
+    A = np.matrix(in_matrix, dtype=np.float)
+    B = np.array(dst).reshape(8)
+    af = np.dot(np.linalg.inv(A.T * A) * A.T, B)
+    return np.append(np.array(af).reshape(8), 1).reshape((3, 3))
+
+
+def create_perspective_transform(src, dst, round=False, splat_args=False):
+    """ Returns a function which will transform points in quadrilateral
+        ``src`` to the corresponding points on quadrilateral ``dst``::
+
+            >>> transform = create_perspective_transform(
+            ...     [(0, 0), (10, 0), (10, 10), (0, 10)],
+            ...     [(50, 50), (100, 50), (100, 100), (50, 100)],
+            ... )
+            >>> transform((5, 5))
+            (74.99999999999639, 74.999999999999957)
+
+        If ``round`` is ``True`` then points will be rounded to the nearest
+        integer and integer values will be returned.
+
+            >>> transform = create_perspective_transform(
+            ...     [(0, 0), (10, 0), (10, 10), (0, 10)],
+            ...     [(50, 50), (100, 50), (100, 100), (50, 100)],
+            ...     round=True,
+            ... )
+            >>> transform((5, 5))
+            (75, 75)
+
+        If ``splat_args`` is ``True`` the function will accept two arguments
+        instead of a tuple.
+
+            >>> transform = create_perspective_transform(
+            ...     [(0, 0), (10, 0), (10, 10), (0, 10)],
+            ...     [(50, 50), (100, 50), (100, 100), (50, 100)],
+            ...     splat_args=True,
+            ... )
+            >>> transform(5, 5)
+            (74.99999999999639, 74.999999999999957)
+
+        If the input values yield an invalid transformation matrix an identity
+        function will be returned and the ``error`` attribute will be set to a
+        description of the error::
+
+            >>> tranform = create_perspective_transform(
+            ...     np.zeros((4, 2)),
+            ...     np.zeros((4, 2)),
+            ... )
+            >>> transform((5, 5))
+            (5.0, 5.0)
+            >>> transform.error
+            'invalid input quads (...): Singular matrix
+        """
+    try:
+        transform_matrix = create_perspective_transform_matrix(src, dst)
+        error = None
+    except np.linalg.LinAlgError as e:
+        transform_matrix = np.identity(3, dtype=np.float)
+        error = "invalid input quads (%s and %s): %s" %(src, dst, e)
+        error = error.replace("\n", "")
+
+    to_eval = "def perspective_transform(%s):\n" %(
+        splat_args and "*pt" or "pt",
+    )
+    to_eval += "  res = np.dot(transform_matrix, ((pt[0], ), (pt[1], ), (1, )))\n"
+    to_eval += "  res = res / res[2]\n"
+    if round:
+        to_eval += "  return (int(round(res[0][0])), int(round(res[1][0])))\n"
+    else:
+        to_eval += "  return (res[0][0], res[1][0])\n"
+    locals = {
+        "transform_matrix": transform_matrix,
+    }
+    locals.update(globals())
+    exec(to_eval,locals,locals)
+    res = locals["perspective_transform"]
+    res.matrix = transform_matrix
+    res.error = error
+    return res
+
+
+def align_mesh2stylegan(temp_tcoords, transformation_params):
+    temp_tcoords = temp_tcoords.copy()
+    temp_tcoords[:, 0] = temp_tcoords[:, 0] - transformation_params['crop'][1]
+    temp_tcoords[:, 1] = temp_tcoords[:, 1] - transformation_params['crop'][0]
+
+    temp_tcoords[:, 0] = temp_tcoords[:, 0] + transformation_params['pad'][1]
+    temp_tcoords[:, 1] = temp_tcoords[:, 1] + transformation_params['pad'][0]
+
+    h, w = (4096, 4096)  # transformation_params['new_size']
+    transform = create_perspective_transform(
+        transformation_params['quad'],
+        [(0, 0), (0, h), (h, w), (w, 0)],
+        splat_args=True,
+    )
+    for i in range(len(temp_tcoords)):
+        temp_tcoords[i, 1], temp_tcoords[i, 0] = transform(temp_tcoords[i, 1], temp_tcoords[i, 0])
+
+    new_tcoords = temp_tcoords[:, ::-1] / (h, w)  # transformation_params['new_size']
+    new_tcoords[:, 1] = 1 - new_tcoords[:, 1]
+    return new_tcoords
+
+def align_im2stylegan(src_im, src_mask, face_landmarks, output_size=1024, transform_size=4096,
+                    enable_padding=True, x_scale=1, y_scale=1, em_scale=0.1, alpha=False):
+        # Align function from FFHQ dataset pre-processing step
+        # https://github.com/NVlabs/ffhq-dataset/blob/master/download_ffhq.py
+
+        lm = np.array(face_landmarks)
+        lm_chin = lm[0: 17]  # left-right
+        lm_eyebrow_left = lm[17: 22]  # left-right
+        lm_eyebrow_right = lm[22: 27]  # left-right
+        lm_nose = lm[27: 31]  # top-down
+        lm_nostrils = lm[31: 36]  # top-down
+        lm_eye_left = lm[36: 42]  # left-clockwise
+        lm_eye_right = lm[42: 48]  # left-clockwise
+        lm_mouth_outer = lm[48: 60]  # left-clockwise
+        lm_mouth_inner = lm[60: 68]  # left-clockwise
+
+        # Calculate auxiliary vectors.
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        eye_avg = (eye_left + eye_right) * 0.5
+        eye_to_eye = eye_right - eye_left
+        mouth_left = lm_mouth_outer[0]
+        mouth_right = lm_mouth_outer[6]
+        mouth_avg = (mouth_left + mouth_right) * 0.5
+        eye_to_mouth = mouth_avg - eye_avg
+
+        # Choose oriented crop rectangle.
+        x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+        x /= np.hypot(*x)
+        x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
+        x *= x_scale
+        y = np.flipud(x) * [-y_scale, y_scale]
+        c = eye_avg + eye_to_mouth * em_scale
+        quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+        qsize = np.hypot(*x) * 2
+        rsize = None
+
+        img = src_im.convert('RGBA').convert('RGB')
+
+        img_mask = src_mask.convert('L')
+
+        img.putalpha(img_mask)
+
+        # Shrink.
+        shrink = int(np.floor(qsize / output_size * 0.5))
+        if shrink > 1:
+            rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
+            img = img.resize(rsize, PIL.Image.ANTIALIAS)
+            quad /= shrink
+            qsize /= shrink
+
+        # Crop.
+        border = max(int(np.rint(qsize * 0.1)), 3)
+        crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+                int(np.ceil(max(quad[:, 1]))))
+        crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
+                min(crop[3] + border, img.size[1]))
+        if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
+            img = img.crop(crop)
+            quad -= crop[0:2]
+
+        # Pad.
+        pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+               int(np.ceil(max(quad[:, 1]))))
+        pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
+               max(pad[3] - img.size[1] + border, 0))
+        if enable_padding and max(pad) > border - 4:
+            pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
+            img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'constant')
+            h, w, _ = img.shape
+            y, x, _ = np.ogrid[:h, :w, :1]
+            mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
+                              1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
+            blur = qsize * 0.02
+            img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+            img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
+            img = np.uint8(np.clip(np.rint(img), 0, 255))
+            if alpha:
+                mask = 1 - np.clip(3.0 * mask, 0.0, 1.0)
+                mask = np.uint8(np.clip(np.rint(mask * 255), 0, 255))
+                img = np.concatenate((img, mask), axis=2)
+                img = PIL.Image.fromarray(img, 'RGBA')
+            else:
+                img = PIL.Image.fromarray(img, 'RGBA')
+            quad += pad[:2]
+
+        # Transform.
+        aligned_mask = PIL.Image.fromarray(np.uint8(img)[:, :, 3])
+        img = PIL.Image.fromarray(np.uint8(img)[:, :, :3])
+
+        img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(),
+                            PIL.Image.BILINEAR)
+        aligned_mask = aligned_mask.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(),
+                                              PIL.Image.BILINEAR)
+        if output_size < transform_size:
+            img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
+            aligned_mask = aligned_mask.resize((output_size, output_size), PIL.Image.ANTIALIAS)
+
+        transformation_params = {
+            'rsize': rsize,
+            'crop': crop,
+            'pad': pad,
+            'quad': quad + 0.5,
+            'new_size': (output_size, output_size)
+        }
+        # Save aligned image.
+        return img, aligned_mask, transformation_params
diff --git a/reconstruction/ostec/utils/ganfit_camera.py b/reconstruction/ostec/utils/ganfit_camera.py
new file mode 100644
index 000000000..40bbe09d6
--- /dev/null
+++ b/reconstruction/ostec/utils/ganfit_camera.py
@@ -0,0 +1,259 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import numpy as np
+import math
+
+"""Collection of functions to adapt GANFit camera parameters"""
+
+GANFIT_CAMERA_CONSTANTS = [np.array([[0.0, 0.0, 6.0]], dtype=np.float32),
+                    np.array([[0.0, 0.0, 0.0]], dtype=np.float32),
+                    np.array([[0.0, 1.0, 0.0]], dtype=np.float32),
+                    np.array([20.], dtype=np.float32)]
+
+def look_at(eye, center, world_up):
+    """Computes camera viewing matrices.
+    Functionality mimes gluLookAt (third_party/GL/glu/include/GLU/glu.h).
+    Args:
+    eye: 2-D float32 tensor with shape [batch_size, 3] containing the XYZ world
+        space position of the camera.
+    center: 2-D float32 tensor with shape [batch_size, 3] containing a position
+        along the center of the camera's gaze.
+    world_up: 2-D float32 tensor with shape [batch_size, 3] specifying the
+        world's up direction; the output camera will have no tilt with respect
+        to this direction.
+    Returns:
+    A [batch_size, 4, 4] float tensor containing a right-handed camera
+    extrinsics matrix that maps points from world space to points in eye space.
+    """
+    batch_size = center.shape[0]
+    vector_degeneracy_cutoff = 1e-6
+    forward = center - eye
+    forward_norm = np.linalg.norm(forward, axis=1, keepdims=True)
+    forward = np.divide(forward, forward_norm)
+
+    to_side = np.cross(forward, world_up)
+    to_side_norm = np.linalg.norm(to_side, axis=1, keepdims=True)
+    to_side = np.divide(to_side, to_side_norm)
+    cam_up = np.cross(to_side, forward)
+
+    w_column = np.array(
+        batch_size * [[0., 0., 0., 1.]], dtype=np.float32)  # [batch_size, 4]
+    w_column = np.reshape(w_column, [batch_size, 4, 1])
+    view_rotation = np.stack(
+        [to_side, cam_up, -forward,
+         np.zeros_like(to_side, dtype=np.float32)],
+        axis=1)  # [batch_size, 4, 3] matrix
+    view_rotation = np.concatenate(
+        [view_rotation, w_column], axis=2)  # [batch_size, 4, 4]
+
+    identity_batch = np.tile(np.expand_dims(np.eye(3), 0), [batch_size, 1, 1])
+    view_translation = np.concatenate([identity_batch, np.expand_dims(-eye, 2)], 2)
+    view_translation = np.concatenate(
+        [view_translation,
+         np.reshape(w_column, [batch_size, 1, 4])], 1)
+    camera_matrices = np.matmul(view_rotation, view_translation)
+    return camera_matrices
+
+def perspective(aspect_ratio, fov_y, near_clip, far_clip):
+    """Computes perspective transformation matrices.
+    Functionality mimes gluPerspective (third_party/GL/glu/include/GLU/glu.h).
+    Args:
+      aspect_ratio: float value specifying the image aspect ratio (width/height).
+      fov_y: 1-D float32 Tensor with shape [batch_size] specifying output vertical
+          field of views in degrees.
+      near_clip: 1-D float32 Tensor with shape [batch_size] specifying near
+          clipping plane distance.
+      far_clip: 1-D float32 Tensor with shape [batch_size] specifying far clipping
+          plane distance.
+    Returns:
+      A [batch_size, 4, 4] float tensor that maps from right-handed points in eye
+      space to left-handed points in clip space.
+    """
+    focal_lengths_y = 1.0 / np.tan(fov_y * (math.pi / 360.0))
+    depth_range = far_clip - near_clip
+    p_22 = -(far_clip + near_clip) / depth_range
+    p_23 = -2.0 * (far_clip * near_clip / depth_range)
+
+    zeros = np.zeros_like(p_23, dtype=np.float32)
+    # pyformat: disable
+    perspective_transform = np.concatenate(
+      [
+          focal_lengths_y / aspect_ratio, zeros, zeros, zeros,
+          zeros, focal_lengths_y, zeros, zeros,
+          zeros, zeros, p_22, p_23,
+          zeros, zeros, -np.ones_like(p_23, dtype=np.float32), zeros
+      ], axis=0)
+    # pyformat: enable
+    perspective_transform = np.reshape(perspective_transform, [4, 4, -1])
+    return np.transpose(perspective_transform, [2, 0, 1])
+
+def euler_matrices(angles):
+  """Computes a XYZ Tait-Bryan (improper Euler angle) rotation.
+  Returns 4x4 matrices for convenient multiplication with other transformations.
+  Args:
+    angles: a [batch_size, 3] tensor containing X, Y, and Z angles in radians.
+  Returns:
+    a [batch_size, 4, 4] tensor of matrices.
+  """
+  s = np.sin(angles)
+  c = np.cos(angles)
+  # Rename variables for readability in the matrix definition below.
+  c0, c1, c2 = (c[:, 0], c[:, 1], c[:, 2])
+  s0, s1, s2 = (s[:, 0], s[:, 1], s[:, 2])
+
+  zeros = np.zeros_like(s[:, 0])
+  ones = np.ones_like(s[:, 0])
+
+  # pyformat: disable
+  flattened = np.concatenate(
+      [
+          c2 * c1, c2 * s1 * s0 - c0 * s2, s2 * s0 + c2 * c0 * s1, zeros,
+          c1 * s2, c2 * c0 + s2 * s1 * s0, c0 * s2 * s1 - c2 * s0, zeros,
+          -s1, c1 * s0, c1 * c0, zeros,
+          zeros, zeros, zeros, ones
+      ],
+      axis=0)
+  # pyformat: enable
+  reshaped = np.reshape(flattened, [4, 4, -1])
+  return np.transpose(reshaped, [2, 0, 1])
+
+def transform_homogeneous(matrices, vertices):
+  """Applies batched 4x4 homogenous matrix transformations to 3-D vertices.
+  The vertices are input and output as as row-major, but are interpreted as
+  column vectors multiplied on the right-hand side of the matrices. More
+  explicitly, this function computes (MV^T)^T.
+  Vertices are assumed to be xyz, and are extended to xyzw with w=1.
+  Args:
+    matrices: a [batch_size, 4, 4] tensor of matrices.
+    vertices: a [batch_size, N, 3] tensor of xyz vertices.
+  Returns:
+    a [batch_size, N, 4] tensor of xyzw vertices.
+  Raises:
+    ValueError: if matrices or vertices have the wrong number of dimensions.
+  """
+  homogeneous_coord = np.ones(
+      [np.shape(vertices)[0], np.shape(vertices)[1], 1], dtype=np.float32)
+  vertices_homogeneous = np.concatenate([vertices, homogeneous_coord], 2)
+
+  return np.matmul(vertices_homogeneous[0], matrices[0].T)
+
+
+def P2sRt(P):
+    ''' decompositing camera matrix P
+    Args:
+        P: (3, 4). Affine Camera Matrix.
+    Returns:
+        s: scale factor.
+        R: (3, 3). rotation matrix.
+        t: (3,). translation.
+    '''
+    t = P[:, 3]
+    R1 = P[0:1, :3]
+    R2 = P[1:2, :3]
+    s = (np.linalg.norm(R1) + np.linalg.norm(R2)) / 2.0
+    r1 = R1 / np.linalg.norm(R1)
+    r2 = R2 / np.linalg.norm(R2)
+    r3 = np.cross(r1, r2)
+
+    R = np.concatenate((r1, r2, r3), 0)
+    return s, R, t
+
+
+# Ref: https://www.learnopencv.com/rotation-matrix-to-euler-angles/
+def isRotationMatrix(R):
+    ''' checks if a matrix is a valid rotation matrix(whether orthogonal or not)
+    '''
+    Rt = np.transpose(R)
+    shouldBeIdentity = np.dot(Rt, R)
+    I = np.identity(3, dtype=R.dtype)
+    n = np.linalg.norm(I - shouldBeIdentity)
+    return n < 1e-6
+
+
+def matrix2angle(R):
+    ''' get three Euler angles from Rotation Matrix
+    Args:
+        R: (3,3). rotation matrix
+    Returns:
+        x: pitch
+        y: yaw
+        z: roll
+    '''
+    assert (isRotationMatrix)
+    sy = math.sqrt(R[0, 0] * R[0, 0] + R[1, 0] * R[1, 0])
+
+    singular = sy < 1e-6
+
+    if not singular:
+        x = math.atan2(R[2, 1], R[2, 2])
+        y = math.atan2(-R[2, 0], sy)
+        z = math.atan2(R[1, 0], R[0, 0])
+    else:
+        x = math.atan2(-R[1, 2], R[1, 1])
+        y = math.atan2(-R[2, 0], sy)
+        z = 0
+
+    # rx, ry, rz = np.rad2deg(x), np.rad2deg(y), np.rad2deg(z)
+    # rx, ry, rz = x * 180 / np.pi, y * 180 / np.pi, z * 180 / np.pi
+    # return rx, ry, rz
+    return x, y, z
+
+def get_camera_matrices(camera_params):
+    # defaults from ganfit
+
+    camera_position = camera_params[0] + GANFIT_CAMERA_CONSTANTS[0]
+    camera_lookat = camera_params[1] + GANFIT_CAMERA_CONSTANTS[1]
+    camera_up =  camera_params[2] + GANFIT_CAMERA_CONSTANTS[2]
+
+    return look_at(camera_position, camera_lookat, camera_up)
+
+def get_pose(camera_params):
+    s, R, t = P2sRt(get_camera_matrices(camera_params)[0])
+    return matrix2angle(R)  # pitch:%.2f,\n yaw:%.2f \n , roll:%.2f \n
+
+def apply_camera_only3d(vertices, camera_params):
+
+    camera_matrices = get_camera_matrices(camera_params)
+
+    clip_space_vertices = transform_homogeneous(camera_matrices, np.array([vertices]))[:, 0:3]
+
+    return clip_space_vertices
+
+
+def apply_camera(camera_params, tmesh, fitting, ganfit_image_size = 512):
+    vertices = tmesh.points
+
+    # defaults from ganfit
+    image_width = ganfit_image_size
+    image_height = ganfit_image_size
+    near_clip = np.array([0.01])
+    far_clip = np.array([10.0])
+    fov_y =  camera_params[3] + GANFIT_CAMERA_CONSTANTS[3]
+
+    camera_matrices = get_camera_matrices(camera_params)
+    perspective_transforms = perspective(image_width / image_height, fov_y, near_clip, far_clip)
+    clip_space_transforms = np.matmul(perspective_transforms, camera_matrices)
+
+    clip_space_vertices = transform_homogeneous(clip_space_transforms, np.array([vertices]))
+
+    _MINIMUM_PERSPECTIVE_DIVIDE_THRESHOLD = 1e-6
+    clip_space_points_w = np.maximum(
+        np.abs(clip_space_vertices[:, 3:4]),
+        _MINIMUM_PERSPECTIVE_DIVIDE_THRESHOLD) * np.sign(
+        clip_space_vertices[:, 3:4])
+    normalized_device_coordinates = (
+            clip_space_vertices[:, 0:3] / clip_space_points_w)
+
+    normalized_device_coordinates = ((normalized_device_coordinates[:, 0:2] * [1, -1]) / 2 + 0.5) * [image_width,
+                                                                                                     image_height]
+
+    dense_lms_org = fitting['transformation'].apply(normalized_device_coordinates[:, ::-1])[:, ::-1]
+
+    dense_lms_org /= [1024, 1024]
+    dense_lms_org[:, 1] = 1 - dense_lms_org[:, 1]
+
+    return dense_lms_org
+
diff --git a/reconstruction/ostec/utils/generate_heatmap.py b/reconstruction/ostec/utils/generate_heatmap.py
new file mode 100644
index 000000000..63b5533c4
--- /dev/null
+++ b/reconstruction/ostec/utils/generate_heatmap.py
@@ -0,0 +1,129 @@
+import numpy as np
+import math
+import cv2
+
+# Adapted from: https://github.com/1adrianb/face-alignment/blob/master/face_alignment/utils.py
+def _gaussian(size=3, sigma=0.25, amplitude=1, normalize=False, width=None, height=None, sigma_horz=None,
+              sigma_vert=None, mean_horz=0.5, mean_vert=0.5):
+    """ Generate a guassian kernel.
+
+    Args:
+        size (int): The size of the kernel if the width or height are not specified
+        sigma (float): Standard deviation of the kernel if sigma_horz or sigma_vert are not specified
+        amplitude: The scale of the kernel
+        normalize: If True, the kernel will be normalized such as values will sum to one
+        width (int, optional): The width of the kernel
+        height (int, optional): The height of the kernel
+        sigma_horz (float, optional): Horizontal standard deviation of the kernel
+        sigma_vert (float, optional): Vertical standard deviation of the kernel
+        mean_horz (float): Horizontal mean of the kernel
+        mean_vert (float): Vertical mean of the kernel
+
+    Returns:
+        np.array: The computed gaussian kernel
+    """
+    # handle some defaults
+    if width is None:
+        width = size
+    if height is None:
+        height = size
+    if sigma_horz is None:
+        sigma_horz = sigma
+    if sigma_vert is None:
+        sigma_vert = sigma
+    center_x = mean_horz * width + 0.5
+    center_y = mean_vert * height + 0.5
+    gauss = np.empty((height, width), dtype=np.float32)
+    # generate kernel
+    for i in range(height):
+        for j in range(width):
+            gauss[i][j] = amplitude * math.exp(-(math.pow((j + 1 - center_x) / (
+                    sigma_horz * width), 2) / 2.0 + math.pow((i + 1 - center_y) / (sigma_vert * height), 2) / 2.0))
+    if normalize:
+        gauss = gauss / np.sum(gauss)
+
+    return gauss
+
+
+# Adapted from: https://github.com/1adrianb/face-alignment/blob/master/face_alignment/utils.py
+def draw_gaussian(image, point, sigma):
+    """ Draw gaussian circle at a point in an image.
+
+    Args:
+        image (np.array): An image of shape (H, W)
+        point (np.array): The center point of the guassian circle
+        sigma (float): Standard deviation of the gaussian kernel
+
+    Returns:
+        np.array: The image with the drawn gaussian.
+    """
+    # Check if the gaussian is inside
+    point[0] = round(point[0], 2)
+    point[1] = round(point[1], 2)
+
+    ul = [math.floor(point[0] - 7.5 * sigma), math.floor(point[1] - 7.5 * sigma)]
+    br = [math.floor(point[0] + 7.5 * sigma), math.floor(point[1] + 7.5 * sigma)]
+    if (ul[0] > image.shape[1] or ul[1] >
+            image.shape[0] or br[0] < 1 or br[1] < 1):
+        return image
+    size = 15 * sigma + 1
+    g = _gaussian(size, sigma=0.1)
+    g_x = [int(max(1, -ul[0])), int(min(br[0], image.shape[1])) -
+           int(max(1, ul[0])) + int(max(1, -ul[0]))]
+    g_y = [int(max(1, -ul[1])), int(min(br[1], image.shape[0])) -
+           int(max(1, ul[1])) + int(max(1, -ul[1]))]
+    img_x = [int(max(1, ul[0])), int(min(br[0], image.shape[1]))]
+    img_y = [int(max(1, ul[1])), int(min(br[1], image.shape[0]))]
+    assert (g_x[0] > 0 and g_y[1] > 0)
+    image[img_y[0] - 1:img_y[1], img_x[0] - 1:img_x[1]] = \
+        image[img_y[0] - 1:img_y[1], img_x[0] - 1:img_x[1]] + g[g_y[0] - 1:g_y[1], g_x[0] - 1:g_x[1]]
+    image[image > 1] = 1
+
+    return image
+
+
+# Adapted from: https://github.com/1adrianb/face-alignment/blob/master/face_alignment/api.py
+def generate_heatmaps(height, width, points, sigma=None):
+    """ Generate heatmaps corresponding to a set of points.
+
+    Args:
+        height (int): Heatmap height
+        width (int): Heatmap width
+        points (np.array): An array of points of shape (N, 2)
+        sigma (float, optional): Standard deviation of the gaussian kernel. If not specified it will be determined
+            from the width of the heatmap
+
+    Returns:
+        np.array: The generated heatmaps.
+    """
+    sigma = max(1, int(np.round(width / 128.))) if sigma is None else sigma
+    heatmaps = np.zeros((points.shape[0], height, width), dtype=np.float32)
+    for i in range(points.shape[0]):
+        if points[i, 0] > 0:
+            heatmaps[i] = draw_gaussian(
+                heatmaps[i], points[i], sigma)
+
+    return heatmaps
+
+
+
+if __name__ == "__main__":
+
+    #you can use [X,2] matrix
+    points = np.array([
+        [(30.2946)+8, 51.6963],
+        [(65.5318)+8, 51.5014],
+        [(48.0252)+8, 71.7366],
+        [(33.5493)+8, 92.3655],
+        [(62.7299)+8, 92.2041]], dtype=np.float32)
+
+
+    heatmaps = generate_heatmaps(width = 112,
+                                 height = 112,
+                                 points = points,
+                                 sigma = 3)
+    print(heatmaps.shape)
+
+    final_heatmap = np.sum(heatmaps, axis=0)
+    cv2.imwrite("final_heatmap.png", final_heatmap*255)
+    print("end")
\ No newline at end of file
diff --git a/reconstruction/ostec/utils/image_rasterization.py b/reconstruction/ostec/utils/image_rasterization.py
new file mode 100644
index 000000000..3b46126b7
--- /dev/null
+++ b/reconstruction/ostec/utils/image_rasterization.py
@@ -0,0 +1,70 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+from menpo3d.rasterize import (
+    rasterize_barycentric_coordinate_images,
+    rasterize_mesh_from_barycentric_coordinate_images)
+
+from menpo.transform import Rotation, Translation, Scale
+from menpo3d.camera import PerspectiveProjection, PerspectiveCamera
+import numpy as np
+
+def rotation_z(theta, degrees=True):
+   if degrees:
+       # convert to radians
+       theta = theta * np.pi / 180.0
+   return Rotation(np.array([[np.cos(theta), -np.sin(theta), 0],
+                             [np.sin(theta), np.cos(theta), 0],
+                             [0, 0, 1]]),
+                   skip_checks=True)
+
+def rotation_y(theta, degrees=True):
+   if degrees:
+       # convert to radians
+       theta = theta * np.pi / 180.0
+   return Rotation(np.array([[np.cos(theta), 0, np.sin(theta)],
+                             [0, 1, 0],
+                             [-np.sin(theta), 0, np.cos(theta)]]),
+                   skip_checks=True)
+
+def rotation_x(theta, degrees=True):
+   if degrees:
+       theta = np.deg2rad(theta)
+   return Rotation(np.array([[ 1,             0,              0],
+                             [ 0, np.cos(theta), -np.sin(theta)],
+                             [ 0, np.sin(theta), np.cos(theta)]]),
+                   skip_checks=True)
+
+def perspective_camera_for_template(img_shape, focal_length_mult=2,
+                                     pose_angle_deg=[0,0,0], cam_dist = 7):
+    f = np.array(img_shape).max() * focal_length_mult
+    rot_z = rotation_z(180 + pose_angle_deg[2])
+    rot_y = rotation_y(180 + pose_angle_deg[1])
+    rot_x = rotation_x(pose_angle_deg[0])
+    rotation = rot_z.compose_before(rot_y).compose_before(rot_x)
+
+    translation = Translation([0, 0, +cam_dist])
+    projection = PerspectiveProjection(f, img_shape)
+    return PerspectiveCamera(rotation, translation, projection)
+
+def align_mesh_to_template(source, target, scale_corrective=1.2):
+    scale = Scale((target.norm() / source.norm()) * scale_corrective,
+                  n_dims=target.n_dims)
+    translation = Translation(target.centre() - source.centre())
+    return translation.compose_before(scale)
+
+def rasterize_image( mesh, img_shape, pose_angle_deg=[0,0,0], cam_dist = 7, ):
+    camera = perspective_camera_for_template(img_shape, pose_angle_deg=pose_angle_deg, cam_dist= cam_dist)
+
+    # Pre-process - align the mesh roughly with the template
+    # aligned_mesh = align_mesh_to_template(mesh, template).apply(mesh)
+
+    mesh_in_img = camera.apply(mesh)
+
+    bcs = rasterize_barycentric_coordinate_images(mesh_in_img, img_shape)
+    img = rasterize_mesh_from_barycentric_coordinate_images(mesh_in_img, *bcs)
+#   shape_img = rasterize_shape_image_from_barycentric_coordinate_images(mesh, *bcs)
+    img.pixels = np.clip(img.pixels,0.0,1.0)
+    return img, mesh_in_img.points
\ No newline at end of file
diff --git a/reconstruction/ostec/utils/shading.py b/reconstruction/ostec/utils/shading.py
new file mode 100644
index 000000000..9b5d436ef
--- /dev/null
+++ b/reconstruction/ostec/utils/shading.py
@@ -0,0 +1,66 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+""" Renders 3D faces in python with lambertian shading
+    Author: Stylianos Ploumpis """
+
+import numpy as np
+from menpo.transform import UniformScale, Translation
+
+
+def l2_normalize(x, axis=0, epsilon=1e-12):
+    """
+    Transforms an `ndarray` to have a unit l2 norm along
+    a given direction.
+    ----------
+    x : `ndarray`
+        The array to be transformed.
+    axis : `int`
+        The axis that will be l2 unit normed.
+    epsilon: `float`
+        A small value such as to avoid division by zero.
+
+    Returns
+    -------
+    x : (D,) `ndarray`
+        The transformed array.
+    """
+    return x / np.maximum(np.linalg.norm(x, axis=axis), epsilon)
+
+
+def mesh_in_unit_sphere(mesh):
+    scale = UniformScale(1 / mesh.norm(), mesh.n_dims)
+    translation = Translation(-scale.apply(mesh).centre())
+    return translation.compose_after(scale)
+
+
+def lambertian_shading(mesh, diffuse_colour=0.4,
+                       albedo_weighting=0.6, ambient_colour=0.2,
+                       light_positions=((45, 45, 1), (30, 60, 10),(30, 60, 0.5),(0, 0, 1))):
+
+    diffuse_colour = np.asarray(diffuse_colour)
+    light_positions = l2_normalize(np.asarray(light_positions).reshape(-1, 3),
+                                   axis=0)
+
+    unit_transform = mesh_in_unit_sphere(mesh)
+    mesh = unit_transform.apply(mesh)
+
+    light_directions = l2_normalize(light_positions.reshape(-1, 1, 3) -
+                                    mesh.points[None, ...], axis=0)
+
+    # Calculate the lambertian reflectance for each light source.
+    # This will be an `ndarray` of shape(num_light_sources, num_vertices)
+    lambertian = np.sum(light_directions *
+                        mesh.vertex_normals()[None, ...], 2)[..., None]
+
+    # Sum up the contribution of all the light sources and multiply by the
+    # diffusion colour.
+    lambertian = lambertian.sum(0) * diffuse_colour + ambient_colour
+
+    mesh.colours[...] = np.clip(mesh.colours * albedo_weighting +
+                                lambertian * (1 - albedo_weighting),
+                                0, 1)
+
+    return unit_transform.pseudoinverse().apply(mesh)
\ No newline at end of file
diff --git a/reconstruction/ostec/utils/utils.py b/reconstruction/ostec/utils/utils.py
new file mode 100644
index 000000000..6eb00eb33
--- /dev/null
+++ b/reconstruction/ostec/utils/utils.py
@@ -0,0 +1,170 @@
+# Copyright (c) 2020, Baris Gecer. All rights reserved.
+#
+# This work is made available under the CC BY-NC-SA 4.0.
+# To view a copy of this license, see LICENSE
+
+import PIL.Image
+from skimage.morphology import binary_dilation,  disk
+from skimage.filters import gaussian
+from scipy.interpolate import NearestNDInterpolator
+import cv2 as cv
+from utils.shading import lambertian_shading
+from menpo3d.rasterize import rasterize_mesh
+from menpo.shape import TriMesh, TexturedTriMesh, ColouredTriMesh
+from menpo.image import Image
+from utils.image_rasterization import *
+import menpo.io as mio
+import os
+
+gold = (244/256,178/256,88/256)
+green = (26/256,140/256,57/256)
+blue = (24/256,148/256,187/256)
+red = (255/256,51/256,51/256)
+purple = (102/256,0/256,204/256)
+pink = (255/256,21/256,255/256)
+
+def perspective_camera(img_shape, focal_length_mult=2,
+                                     pose_angle_deg=[0,0,0], translation=4):
+    f = np.array(img_shape).max() * focal_length_mult
+    rot_z = rotation_z(180 + pose_angle_deg[2])
+    rot_y = rotation_y(180 + pose_angle_deg[1])
+    rot_x = rotation_x(pose_angle_deg[0])
+    rotation = rot_z.compose_before(rot_y).compose_before(rot_x)
+
+    translation = Translation([0, 0, translation])
+    projection = PerspectiveProjection(f, img_shape)
+    return PerspectiveCamera(rotation, translation, projection)
+
+def rasterize_mesh_at_template(mesh, img_shape=(1024,1024), pose_angle_deg=[0, 0, 0], translation=4, shaded=True, **kwargs):
+    camera = perspective_camera(img_shape, pose_angle_deg=pose_angle_deg, translation=translation)
+    if shaded:
+        mesh = lambertian_shading(mesh, **kwargs)
+    mesh.points[mesh.points == 0] = None
+    return rasterize_mesh(camera.apply(mesh), img_shape)
+
+def export_tmesh(tmesh, path, pose_angle_deg=[0, 30, 0], translation=4):
+    render = rasterize_mesh_at_template(tmesh, pose_angle_deg=pose_angle_deg, shaded=False, translation=translation)
+    mio.export_image(render, path, overwrite=True)
+
+def export_shape(tmesh, path, pose_angle_deg=[0, 30, 0]):
+    mesh = ColouredTriMesh(tmesh.points, tmesh.trilist,
+                           np.tile(np.array(blue), [len(tmesh.points), 1]))
+    render = rasterize_mesh_at_template(mesh, pose_angle_deg=pose_angle_deg, shaded=True)
+    mio.export_image(render, path, overwrite=True)
+
+def uv_color_normalize(img_uv_src, angle_uv_src, img_uv, angle_uv):
+    intersection = (angle_uv_src.pixels[0] < 0.6) & (angle_uv_src.pixels[0] > 0.3) & (angle_uv.pixels[0] > 0.7)
+    if np.any(intersection):
+        target_dist = img_uv_src.pixels[:, intersection]
+        target_mean = np.mean(target_dist, 1).reshape([3, -1])
+        targed_std = np.std(target_dist, 1).reshape([3, -1])
+
+        source_dist = img_uv.pixels[:, intersection]
+        source_mean = np.mean(source_dist, 1).reshape([3, -1])
+        source_std = np.std(source_dist, 1).reshape([3, -1])
+
+        temp = ((img_uv.pixels.reshape([3, -1]) - source_mean) / source_std) * targed_std + target_mean
+        return Image(temp.reshape(img_uv.pixels.shape))
+    else:
+        return img_uv
+
+def uv_stiching(img_uv_list, angle_uv_list, smoothing_sigma = 10, seamless_clone=False):
+    max_ind = np.argmax(np.array(angle_uv_list).mean(axis=1), axis=0)
+    mask_out_all = np.max(np.array(angle_uv_list).mean(axis=1),axis=0) ==-1
+    max_ind[mask_out_all] = len(img_uv_list) -1
+    all_uvs = np.clip(np.array(img_uv_list), 0, 1)
+
+    max_ind_one_hot = np.zeros((max_ind.size, len(img_uv_list)))
+    max_ind_one_hot[np.arange(max_ind.size), max_ind.flatten()] = 1
+    max_ind_one_hot = max_ind_one_hot.reshape(max_ind.shape + (-1,))
+
+    if smoothing_sigma>0:
+        max_ind_one_hot_g = gaussian(max_ind_one_hot, sigma=smoothing_sigma, multichannel=True, mode='reflect')
+    else:
+        max_ind_one_hot_g = max_ind_one_hot
+
+    max_ind_one_hot = np.tile(max_ind_one_hot, [3, 1, 1, 1])
+    max_ind_one_hot = np.transpose(max_ind_one_hot, [3, 0, 1, 2])
+
+    max_ind_one_hot_g = np.tile(max_ind_one_hot_g, [3, 1, 1, 1])
+    max_ind_one_hot_g = np.transpose(max_ind_one_hot_g, [3, 0, 1, 2])
+
+    if seamless_clone:
+        all_uvs_uint8 = (np.transpose(all_uvs,[0,2,3,1])*255).astype(np.uint8)
+        max_ind_one_uint8 = (np.transpose(1 - max_ind_one_hot,[0,2,3,1])*255).astype(np.uint8)
+        max_ind_one_uint8_g = (np.transpose(1 - max_ind_one_hot_g,[0,2,3,1])*255).astype(np.uint8)
+
+        dst = (Image(np.sum(all_uvs * max_ind_one_hot_g, axis=0)).pixels_with_channels_at_back()*255).astype(np.uint8)
+        for i, (src, mask, mask_g) in enumerate(zip(all_uvs_uint8[1:], max_ind_one_uint8[1:], max_ind_one_uint8_g[1:])):
+            mask_inv = 255-mask.copy()
+            mask_inv[:,0:5] = 255
+            mask_inv[:, -6:-1] = 255
+            mask_inv[0:5,:] = 255
+            mask_inv[-6:-1,:] = 255
+            mask_inv_dilated = binary_dilation(mask_inv.astype(np.bool)[:,:,0], disk(50)).astype(np.uint8)*255
+            mask_inv_dilated = np.transpose(np.tile(mask_inv_dilated,[3,1,1]),[1,2,0])
+
+            mask_inv_dilated_g = gaussian(mask_inv_dilated/255, sigma=100, multichannel=True, mode='reflect')
+            mixed = (src * mask_inv_dilated_g + dst * (1-mask_inv_dilated_g)).astype(np.uint8)
+
+            im_clone = cv.seamlessClone(dst, mixed, mask_inv, (512,512), cv.NORMAL_CLONE)
+            dst = (im_clone * (1-mask_g/255) +  dst* (mask_g/255)).astype(np.uint8)
+        final_uv = Image(np.transpose(dst.astype(np.float)/255,[2,0,1]))
+    else:
+        final_uv = Image(np.sum(all_uvs * max_ind_one_hot_g, axis=0))
+
+    final_uv.pixels = np.clip(final_uv.pixels,0.0, 1.0)
+    return final_uv, max_ind_one_hot
+
+def im_menpo2PIL(menpo_im):
+    return PIL.Image.fromarray((menpo_im.pixels_with_channels_at_back() * 255).astype(np.uint8))
+
+def im_PIL2menpo(pil_im):
+    return Image.init_from_channels_at_back(np.array(pil_im).astype(np.float) / 255)
+
+def fill_UV(UV):
+    mask = np.sum(UV.pixels, 0) == 0
+    xx, yy = np.meshgrid(np.arange(UV.shape[1]), np.arange(UV.shape[0]))
+    xym = np.vstack((np.ravel(xx[~mask]), np.ravel(yy[~mask]))).T
+    data = UV.pixels[:, ~mask]
+    for i in range(3):
+        interp = NearestNDInterpolator(xym, data[i])
+        result = interp(np.ravel(xx[mask]), np.ravel(yy[mask]))
+        UV.pixels[i, mask] = result
+    return UV
+
+from keras.utils import get_file
+import bz2
+def unpack_bz2(src_path):
+    data = bz2.BZ2File(src_path).read()
+    dst_path = src_path[:-4]
+    with open(dst_path, 'wb') as fp:
+        fp.write(data)
+    return dst_path
+
+
+def fix_obj(fp):
+    os.path.dirname(fp)
+    template = """# Produced by Dimensional Imaging OBJ exporter
+# http://www.di3d.com
+#
+#
+newmtl merged_material
+Ka  0.5 0.5 0.5
+Kd  0.5 0.5 0.5
+Ks  0.47 0.47 0.47
+d 1
+Ns 0
+illum 2
+map_Kd {}.png
+#
+#
+# EOF""".format(os.path.splitext(os.path.basename(fp))[0])
+    with open(os.path.join(os.path.dirname(fp), os.path.splitext(os.path.basename(fp))[0] + '.mtl'), 'w') as f:
+        f.write(template)
+
+    with open(fp, 'r+')  as f:
+        content = f.read()
+        f.seek(0, 0)
+        f.write('mtllib ' + os.path.splitext(os.path.basename(fp))[0] + '.mtl' + '\n' + content)
+
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 000000000..b8cdc0177
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,3 @@
+Cython>=0.29.28
+cmake>=3.22.3    
+numpy>=1.22.3
diff --git a/resources/11513D05.jpg b/resources/11513D05.jpg
deleted file mode 100644
index c38bb0d08..000000000
Binary files a/resources/11513D05.jpg and /dev/null differ
diff --git a/resources/arcface.png b/resources/arcface.png
deleted file mode 100644
index fa43f9e82..000000000
Binary files a/resources/arcface.png and /dev/null differ
diff --git a/resources/cov_test.jpg b/resources/cov_test.jpg
deleted file mode 100644
index 8d5bbe5a4..000000000
Binary files a/resources/cov_test.jpg and /dev/null differ
diff --git a/resources/facerecognitionfromvideo.PNG b/resources/facerecognitionfromvideo.PNG
deleted file mode 100644
index 96f7724f0..000000000
Binary files a/resources/facerecognitionfromvideo.PNG and /dev/null differ
diff --git a/resources/lfr19_wechat1.jpg b/resources/lfr19_wechat1.jpg
deleted file mode 100644
index 3ba0dadfd..000000000
Binary files a/resources/lfr19_wechat1.jpg and /dev/null differ
diff --git a/resources/mainsteps.png b/resources/mainsteps.png
deleted file mode 100644
index 1d3d1ff61..000000000
Binary files a/resources/mainsteps.png and /dev/null differ
diff --git a/resources/memoryspeed.png b/resources/memoryspeed.png
deleted file mode 100644
index d4e1a44ef..000000000
Binary files a/resources/memoryspeed.png and /dev/null differ
diff --git a/resources/retina_R50_ex1.jpg b/resources/retina_R50_ex1.jpg
deleted file mode 100644
index 603cda3a0..000000000
Binary files a/resources/retina_R50_ex1.jpg and /dev/null differ
diff --git a/resources/retina_R50_ex2.jpg b/resources/retina_R50_ex2.jpg
deleted file mode 100644
index ec2a606b8..000000000
Binary files a/resources/retina_R50_ex2.jpg and /dev/null differ
diff --git a/resources/subcenterarcfacediff.png b/resources/subcenterarcfacediff.png
deleted file mode 100644
index 541fa7663..000000000
Binary files a/resources/subcenterarcfacediff.png and /dev/null differ
diff --git a/resources/subcenterarcfaceframework.png b/resources/subcenterarcfaceframework.png
deleted file mode 100644
index b7be824b6..000000000
Binary files a/resources/subcenterarcfaceframework.png and /dev/null differ
diff --git a/resources/widerfacevaltest.png b/resources/widerfacevaltest.png
deleted file mode 100644
index 65807decf..000000000
Binary files a/resources/widerfacevaltest.png and /dev/null differ
diff --git a/sample-images/t2.jpg b/sample-images/t2.jpg
deleted file mode 100644
index dcca9303f..000000000
Binary files a/sample-images/t2.jpg and /dev/null differ
diff --git a/server/.models/.gitkeep b/server/.models/.gitkeep
new file mode 100644
index 000000000..8b1378917
--- /dev/null
+++ b/server/.models/.gitkeep
@@ -0,0 +1 @@
+
diff --git a/server/LICENSING.md b/server/LICENSING.md
new file mode 100644
index 000000000..9fba71cda
--- /dev/null
+++ b/server/LICENSING.md
@@ -0,0 +1,8 @@
+# Licensing
+
+- InsightFace Server source code and its Python SDK are licensed under MIT.
+- InsightFace open-source model packages are licensed for non-commercial
+  academic research only. For commercial licensing, visit
+  <https://www.insightface.ai>.
+- Every model directory loaded by the Server includes a `MODEL.LICENSE` file
+  describing that model's license.
diff --git a/server/Makefile b/server/Makefile
new file mode 100644
index 000000000..4f738c4d0
--- /dev/null
+++ b/server/Makefile
@@ -0,0 +1,72 @@
+SHELL := /bin/sh
+PYTHON ?= python3.11
+PYTHONPATH_VALUE := backend:../python-package
+CPU_IMAGE ?= ghcr.io/deepinsight/insightface-server:0.2.0-cpu
+CUDA_IMAGE ?= ghcr.io/deepinsight/insightface-server:0.2.0-cuda12
+SERVER_VERSION ?= 0.2.0
+NATIVE_CPU_BUILD_DIR ?= build/native-search-cpu
+DOCKER_BUILD_ARGS ?= --build-arg HTTP_PROXY --build-arg HTTPS_PROXY \
+	--build-arg http_proxy --build-arg https_proxy \
+	--build-arg INSIGHTFACE_SERVER_VERSION=$(SERVER_VERSION)
+
+.PHONY: lint test test-api test-sdk test-frontend test-native-cpu update-api-docs build-cpu build-cuda12 \
+	run-cpu run-cuda12 test-cpu test-cuda12 test-consistency smoke-test release-preflight \
+	release-precheck
+
+lint:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m ruff check backend tests
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m mypy backend/insightface_server
+
+test:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m pytest tests -m "not cpu and not cuda and not consistency and not docker and not e2e"
+
+test-api:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m pytest tests/api tests/integration
+
+test-sdk:
+	$(PYTHON) -m pytest tests/sdk
+
+test-frontend:
+	node --test tests/frontend/*.test.mjs
+
+update-api-docs:
+	$(PYTHON) scripts/update_openapi_snapshot.py
+
+test-native-cpu:
+	cmake -S native/search -B $(NATIVE_CPU_BUILD_DIR) \
+		-DIFS_SEARCH_BUILD_CPU=ON \
+		-DIFS_SEARCH_BUILD_CUDA=OFF \
+		-DIFS_SEARCH_BUILD_TESTS=ON
+	cmake --build $(NATIVE_CPU_BUILD_DIR) --parallel
+	ctest --test-dir $(NATIVE_CPU_BUILD_DIR) --output-on-failure
+
+build-cpu:
+	docker build $(DOCKER_BUILD_ARGS) --platform linux/amd64 -f docker/Dockerfile.cpu -t $(CPU_IMAGE) ..
+
+build-cuda12:
+	docker build $(DOCKER_BUILD_ARGS) --platform linux/amd64 -f docker/Dockerfile.cuda12 -t $(CUDA_IMAGE) ..
+
+run-cpu:
+	docker compose -f deploy/compose.cpu.yml up -d --no-build
+
+run-cuda12:
+	docker compose -f deploy/compose.cuda12.yml up -d --no-build
+
+test-cpu:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m pytest tests/cpu -m cpu
+
+test-cuda12:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m pytest tests/cuda -m cuda
+
+test-consistency:
+	PYTHONPATH=$(PYTHONPATH_VALUE) $(PYTHON) -m pytest tests/integration/test_consistency.py -m consistency
+
+smoke-test:
+	$(PYTHON) scripts/smoke_test.py --base-url http://127.0.0.1:8080
+
+release-preflight:
+	$(PYTHON) scripts/release_preflight.py --version $(SERVER_VERSION)
+
+release-precheck:
+	$(PYTHON) scripts/release_preflight.py --version $(SERVER_VERSION) --relaxed \
+		--output test-results/release-preflight.json
diff --git a/server/README.de.md b/server/README.de.md
new file mode 100644
index 000000000..27a65a689
--- /dev/null
+++ b/server/README.de.md
@@ -0,0 +1,288 @@
+# InsightFace Server
+
+**Sprachen:** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · Deutsch · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **Eine GPU. 50M+ Gesichtsvektoren. Rasante Suche mit INT8-Feature-Quantisierung ohne wesentlichen Genauigkeitsverlust.**
+
+**Ein selbst gehosteter Gesichtserkennungsserver mit Web UI, verständlicher
+REST API, SQLite und lokaler CPU- oder NVIDIA-GPU-Inferenz in einem Container.**
+
+```text
+Bild hochladen -> erkennen, vergleichen, registrieren oder suchen
+```
+
+> **Modelllizenz:** Öffentliche vortrainierte InsightFace-Modelle sind in der
+> Regel nur für nichtkommerzielle Forschung freigegeben. Kommerzielle Nutzung
+> erfordert eine separate Genehmigung von
+> [InsightFace](https://www.insightface.ai).
+
+InsightFace Server ist für übliche Gesichtserkennungsabläufe eine einfachere,
+datenschutzorientierte Alternative zu AWS Rekognition auf eigener
+Infrastruktur. Bilder, Embeddings, Modelle und Indizes können im eigenen Netz
+bleiben. Er ist **kein** AWS-kompatibler Ersatz und implementiert weder SigV4,
+IAM, Region noch AWS-Ressourcensemantik.
+
+Aktuelle Version: **0.2.0**, Linux x86_64.
+
+| Laufzeit | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+Die beweglichen Tags `cpu` und `cuda12` bezeichnen die neueste stabile Version
+der jeweiligen Laufzeitfamilie. Ein mehrdeutiges `latest` gibt es nicht. Siehe
+[Maintainer Guide — English](docs/maintainer-guide.md) für die Release-Regeln.
+
+![InsightFace Server Dashboard auf Englisch](docs/images/customer/dashboard-en.jpg)
+
+## Funktionsübersicht
+
+- SCRFD-Erkennung, fünf Landmarken, Ausrichtung, ArcFace-Embeddings, L2
+  normalization, rohe cosine similarity und exakte 1:N-Personensuche.
+- Mehrere Erkennungsauflösungen mit gemeinsamem NMS sowie `largest` und
+  `center_largest` zur Ein-Gesicht-Auswahl.
+- `Collection -> Person -> FaceSample`, modellgebundene Collections,
+  Mehrbildregistrierung mit Teilerfolg, metadata und klaren Ablehnungsgründen.
+- Registrierung mit `review_mode` `off`, `standard` oder `strict`; optional
+  vorab berechnete `external_trusted` Embeddings.
+- Exakte GPU-Suche mit FP32-, FP16-, BF16- und INT8-Vektorspeicherung.
+- Mehrsprachige Web UI für Dashboard, Collections, Personen, Detect, Compare,
+  Search, RTSP-Monitoring, Systemdiagnose und Hilfe.
+- 29 snake_case-REST-Operationen unter `/v1`, einschließlich geschütztem
+  `/v1/embeddings`, plus schlankem typisierten Python SDK.
+- Serverseitige RTSP Monitors mit begrenzten In-Memory-Ereignissen, mehreren
+  Clients und optionalem `preview.mjpeg`; das Schließen des Browsers beendet
+  die Überwachung nicht.
+- SQLite als dauerhafte Quelle, rekonstruierbare exakte In-Memory-Indizes,
+  schreibgeschütztes `/models`, persistentes `/data`, Migrationen,
+  Healthchecks und strikte CUDA-Prüfung ohne stillen CPU-Fallback.
+- JPEG-, PNG- und WebP-Eingaben; Originaluploads werden standardmäßig nicht
+  gespeichert.
+
+### GPU-Suchleistung auf der RTX 5090
+
+Auf einer NVIDIA GeForce RTX 5090 (32.607 MiB) speicherte der native exakte
+CUDA-Flat-Index mit INT8 bis zu **58,9M 512-dimensionale Bildvektoren**.
+
+| GPU-Datentyp | Maximale Bildvektoren | 10M Top-5 p50 | 10M serielle QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15,8M | 12,84 ms | 77,85 |
+| FP16 | 30,7M | 6,83 ms | 146,32 |
+| BF16 | 30,7M | 6,83 ms | 146,33 |
+| INT8 | **58,9M** | **3,84 ms** | **260,81** |
+
+INT8 erreichte gegenüber FP32 die 3,73-fache gemessene Kapazität und den
+3,35-fachen 10M-Top-5-Durchsatz. Dies sind reine GPU-Messungen auf derselben
+RTX 5090 mit Driver 580.105.08 und CUDA 12.9. Die Kapazität ist die isolierte
+Grenze des nativen Index ohne geladene ONNX-Modelle oder Serverlast. Der
+Geschwindigkeitstest nutzt exakt 10M Bildvektoren, eine vollständige exakte
+GPU-residente Top-5-Suche, eine Anfrage gleichzeitig, 10 Warm-ups und 100
+Messungen. Innerhalb der jeweiligen Speicherung ist die Suche exakt;
+Quantisierung kann dennoch Scores gegenüber FP32 verändern. Produktion
+benötigt VRAM-Reserve für Modelle, Requests, Parallelität, Index-Neuaufbau und
+den Allocator.
+
+### Multiethnische ICCV21-MFR-MR-ALL-Genauigkeit
+
+Wir haben die nativen Suchprofile auf dem multiethnischen MR-Testdatensatz von
+[ICCV21-MFR](../challenges/iccv21-mfr/) mit dem MR-ALL-All-Pairs-1:1-Protokoll
+bei FAR `1e-6` evaluiert. Alle Profile verwenden dieselben einmalig über die
+Server API extrahierten, L2-normalisierten 512-dimensionalen
+`buffalo_l`-Embeddings; nur Vektorspeicherung und Rechendarstellung der Suche
+ändern sich.
+
+| Suchprofil | MR-ALL bei FAR 1e-6 | Cosine-Schwelle | Differenz zu FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91,249107 % | 0,407787 | — |
+| FP16 | 91,249197 % | 0,407787 | +0,000090 Prozentpunkte |
+| BF16 | 91,248502 % | 0,407787 | -0,000605 Prozentpunkte |
+| **INT8** | **91,248005 %** | **0,407739** | **-0,001102 Prozentpunkte** |
+
+**INT8 zeigt in diesem Benchmark keinen wesentlichen Genauigkeitsverlust:**
+Mit der challenge-üblichen Darstellung auf zwei Dezimalstellen erreichen FP32
+und INT8 jeweils **91,25 % MR-ALL**; die ungerundete Differenz beträgt nur
+0,0011 Prozentpunkte. Gleichzeitig bleiben die oben gezeigten Vorteile von
+3,73-facher gemessener Kapazität und 3,35-fachem 10M-Top-5-Durchsatz erhalten.
+Verglichen wird die Präzision von Vektorspeicherung und Suche, nicht die
+INT8-Modellinferenz.
+
+![Collection-Verwaltung auf Englisch](docs/images/customer/collections-en.jpg)
+
+![RTSP Monitor auf Englisch; private Adresse unkenntlich](docs/images/customer/monitoring-en.jpg)
+
+## Schnellstart
+
+Voraussetzungen:
+
+- Linux x86_64 mit Docker Engine und Docker Compose;
+- für CUDA eine unterstützte NVIDIA GPU, NVIDIA Driver und NVIDIA Container
+  Toolkit.
+
+Auf dem Host sind Python, OpenCV, ONNX Runtime, CUDA Toolkit und cuDNN nicht
+erforderlich. Öffentliche Images enthalten keine Modelle, Kundendaten, API Keys
+oder Produktionskonfiguration.
+
+Installieren Sie in einem vollständigen InsightFace-Checkout ein Modell nach
+`server/.models`:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+Das Modellwerkzeug unterstützt außerdem `buffalo_m`, `buffalo_sc` und
+`antelopev2`. Es schreibt `manifest.json` und die signierte `MODEL.LICENSE`;
+`models verify` prüft das installierte Paket. Die Modellbedingungen sind von
+der Server-Quelllizenz getrennt.
+
+CPU starten:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Stattdessen CUDA 12 starten:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Öffnen Sie für CPU `http://SERVER:18097/` oder für CUDA
+`http://SERVER:18098/`. Erstellen Sie eine Collection, registrieren Sie eine
+Person mit einem oder mehreren Bildern und suchen Sie mit einem anderen Bild.
+`docker compose ... down` ohne `-v` erhält das Datenvolume.
+
+Die mitgelieferten Compose-Dateien deaktivieren Authentifizierung für isolierte
+Tests standardmäßig. Vor der Freigabe für andere Benutzer oder Netze:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='durch-ein-langes-zufaelliges-geheimnis-ersetzen'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Den vollständigen ersten Ablauf erklärt das
+[Benutzerhandbuch](docs/user-guide.de.md).
+
+## Aus Quellcode bauen
+
+Die Dockerfiles kopieren `server/` und ausgewählte Inferenzmodule aus
+`python-package/insightface/`; daher ist das vollständige Repository der
+Build-Kontext.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` erzwingt das lokal gebaute Image. Der Build lädt weiterhin
+festgeschriebene Basis-Images und Abhängigkeiten; die Modellinstallation lädt
+das separat akzeptierte Modellpaket.
+
+## Kernverhalten
+
+- Similarity ist der rohe Cosinuswert, keine Wahrscheinlichkeit. Thresholds
+  liegen in `0.0..1.0`, Standard ist `0.4`.
+- Eine Collection bindet Modell und embedding contract. Bei Abweichung bleibt
+  sie sichtbar, Registrierung/Suche liefert `collection_model_mismatch`.
+- Das Detection Profile beim Start wird in neue Collections kopiert; danach
+  kann ihr Profil unabhängig für folgende Requests geändert werden.
+- Optionale Gesichtsspeicherung enthält einen auf 112x112 skalierten
+  bounding-box JPEG crop, nicht das Original und nicht den ausgerichteten
+  Erkennungseingang; standardmäßig aus.
+- SQLite-Commits sind maßgeblich. Indexänderungen sind vor einer erfolgreichen
+  Registrierungs-/Löschantwort abgeschlossen; nach Neustart erfolgt der Aufbau
+  aus SQLite.
+- Responses enthalten `x-request-id`; Listen verwenden undurchsichtige
+  signierte cursor.
+
+Exakte Felder, Defaults, Lebenszyklen und Fehlerverhalten stehen ausschließlich
+in den unten verlinkten Detaildokumenten.
+
+## API und SDK
+
+Hauptgruppen:
+
+- System: `/v1/health`, `/v1/system`, `/v1/models`;
+- zustandslose Gesichtsoperationen: `/v1/detect`, `/v1/compare`,
+  `/v1/embeddings`;
+- CRUD für Collection, Person und FaceSample;
+- Personensuche in einer Collection;
+- Konfiguration, Status, Ereignisse und Vorschau von RTSP Monitors.
+
+Alle Parameter, Responses, Fehler und Beispiele enthält der
+[REST-API-Leitfaden](docs/api.de.md). Interaktives OpenAPI bleibt unter `/docs`.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+Installation, Eingaben, Methoden und vollständige SDK-Abläufe stehen im
+[Benutzerhandbuch](docs/user-guide.de.md).
+
+## Sicherheit
+
+Gesichtsbilder und Embeddings sind biometrische Daten. Aktivieren Sie bei
+Netzbetrieb die Authentifizierung, terminieren Sie HTTPS an einem vertrauenswürdigen
+Reverse Proxy, beschränken Sie Docker- und Volume-Zugriff, lassen Sie breites
+CORS aus und definieren Sie Backup, Aufbewahrung, Löschung, Einwilligung und
+Incident Response. Bilder, Embeddings, RTSP-Zugangsdaten und API Keys gehören
+nicht in Logs.
+
+Der Server enthält kein TLS, Benutzerkonten, RBAC, Cloud-IAM oder
+Rechtskonformitätsmodul. Betrieb und Sicherheit stehen im
+[Benutzerhandbuch](docs/user-guide.de.md).
+
+## Umfang der ersten Phase
+
+Nicht enthalten sind AWS-/CompreFace-Kompatibilität, CUDA 11, Jetson, ARM64,
+Windows Container, TensorRT, Kubernetes, verteilte Worker, persistente
+Monitor-Ereignisse oder Aufnahme/NVR sowie Liveness-, Deepfake- oder
+demografische Analysen.
+
+## Dokumentation
+
+- [Benutzerhandbuch](docs/user-guide.de.md) — Installation, Konfiguration,
+  Modelle, Web UI, SDK, GPU, Sicherheit, Backup und Fehlerbehebung.
+- [REST-API-Leitfaden](docs/api.de.md) — alle öffentlichen Endpunkte, Felder,
+  Verhalten, Ergebnisse, Fehler, Pagination und Beispiele.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — Architektur,
+  Suchinternes, Tests, Beiträge und Container-Releases.
+
+GitHub und die Web-UI-Hilfe verwenden dieselben lokalisierten Markdown-Dateien;
+nur die Darstellung unterscheidet sich.
+
+## Lizenz
+
+Der einzige Einstieg für Lizenzinformationen ist
+[LICENSING.md](LICENSING.md). Server-Quellcode und Python SDK stehen unter MIT
+License; diese Erklärung gilt nicht für Modelldateien, Modellgewichte,
+Datensätze oder Drittkomponenten. Öffentliche InsightFace-Modelle sind ohne
+separate Genehmigung in der Regel nur für nichtkommerzielle Forschung bestimmt.
+Kommerzielle Lizenzierung: <https://www.insightface.ai>.
diff --git a/server/README.es.md b/server/README.es.md
new file mode 100644
index 000000000..cf1cfc314
--- /dev/null
+++ b/server/README.es.md
@@ -0,0 +1,279 @@
+# InsightFace Server
+
+**Idiomas:** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · Español · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **Una GPU. 50M+ de vectores faciales. Búsqueda ultrarrápida con cuantización INT8 de características y sin pérdida material de precisión.**
+
+**Servidor de reconocimiento facial autohospedado con Web UI, una REST API
+directa, SQLite e inferencia local por CPU o GPU NVIDIA en un solo contenedor.**
+
+```text
+subir una imagen -> detectar, comparar, registrar o buscar
+```
+
+> **Licencia del modelo:** los modelos públicos preentrenados de InsightFace
+> suelen estar disponibles solo para investigación no comercial. El uso
+> comercial requiere una autorización independiente de
+> [InsightFace](https://www.insightface.ai).
+
+InsightFace Server es una alternativa más sencilla y centrada en la privacidad
+a AWS Rekognition para flujos habituales ejecutados en infraestructura propia.
+Imágenes, embeddings, modelos e índices pueden permanecer dentro de su red. No
+es un reemplazo compatible con AWS y no implementa SigV4, IAM, Region ni
+semántica de recursos AWS.
+
+Versión actual: **0.2.0**, Linux x86_64.
+
+| Entorno | Imagen |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| GPU NVIDIA | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+Las etiquetas móviles `cpu` y `cuda12` identifican la última versión estable de
+cada familia. No se usa una etiqueta ambigua `latest`. Consulte
+[Maintainer Guide — English](docs/maintainer-guide.md) para la política.
+
+![Panel de InsightFace Server en inglés](docs/images/customer/dashboard-en.jpg)
+
+## Funciones principales
+
+- Detección SCRFD, cinco landmarks, alineación, embeddings ArcFace, L2
+  normalization, cosine similarity original y búsqueda exacta 1:N de Person.
+- Detección multirresolución con un único NMS combinado y selección
+  `largest` o `center_largest`.
+- `Collection -> Person -> FaceSample`, Collections vinculadas al modelo,
+  registro de varias imágenes con éxito parcial, metadata y razones claras.
+- `review_mode` de registro: `off`, `standard` o `strict`; embeddings
+  precalculados opcionales mediante `external_trusted`.
+- Búsqueda GPU exacta con almacenamiento vectorial FP32, FP16, BF16 e INT8.
+- Web UI multilingüe para Panel, Collections, Personas, Detect, Compare,
+  Search, monitorización RTSP, System y Help.
+- 29 operaciones REST snake_case bajo `/v1`, incluido `/v1/embeddings`
+  protegido, más un SDK Python ligero y tipado.
+- Monitors RTSP persistentes del servidor, eventos acotados en memoria,
+  múltiples clientes y `preview.mjpeg` opcional; cerrar el navegador no detiene
+  el monitor.
+- SQLite como fuente persistente, índices exactos reconstruibles en memoria,
+  `/models` de solo lectura, `/data` persistente, migraciones, health checks y
+  validación CUDA estricta sin fallback silencioso a CPU.
+- Entrada JPEG, PNG y WebP; los originales no se conservan por defecto.
+
+### Rendimiento de búsqueda GPU en RTX 5090
+
+En una NVIDIA GeForce RTX 5090 (32.607 MiB), el índice plano exacto CUDA nativo
+almacenó hasta **58,9M vectores de imagen de 512 dimensiones con INT8**.
+
+| Tipo de datos GPU | Máximo de vectores de imagen | 10M Top-5 p50 | 10M QPS serial |
+| --- | ---: | ---: | ---: |
+| FP32 | 15,8M | 12,84 ms | 77,85 |
+| FP16 | 30,7M | 6,83 ms | 146,32 |
+| BF16 | 30,7M | 6,83 ms | 146,33 |
+| INT8 | **58,9M** | **3,84 ms** | **260,81** |
+
+INT8 obtuvo 3,73 veces la capacidad medida y 3,35 veces el rendimiento Top-5
+de FP32 sobre 10M. Son mediciones solo de GPU en la misma RTX 5090 con Driver
+580.105.08 y CUDA 12.9. La capacidad es el límite aislado del índice nativo sin
+modelos ONNX cargados ni carga del Server. La velocidad usa exactamente 10M
+vectores de imagen, barrido exacto Top-5 residente en GPU, una consulta en
+curso, 10 calentamientos y 100 mediciones. La búsqueda es exacta dentro de cada
+representación almacenada; la cuantización puede cambiar scores frente a FP32.
+Producción debe reservar VRAM para modelos, requests, concurrencia,
+reconstrucciones del índice y el allocator.
+
+### Precisión MR-ALL multirracial de ICCV21-MFR
+
+Evaluamos los perfiles de búsqueda nativos sobre el conjunto multirracial (MR)
+de [ICCV21-MFR](../challenges/iccv21-mfr/) con su protocolo MR-ALL 1:1 de todos
+los pares y FAR `1e-6`. Todos los perfiles usan los mismos embeddings
+`buffalo_l` de 512 dimensiones, normalizados con L2 y extraídos una sola vez
+mediante la Server API; solo cambia la representación de almacenamiento y
+cálculo de la búsqueda.
+
+| Perfil de búsqueda | MR-ALL con FAR 1e-6 | Umbral cosine | Diferencia frente a FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91,249107 % | 0,407787 | — |
+| FP16 | 91,249197 % | 0,407787 | +0,000090 puntos porcentuales |
+| BF16 | 91,248502 % | 0,407787 | -0,000605 puntos porcentuales |
+| **INT8** | **91,248005 %** | **0,407739** | **-0,001102 puntos porcentuales** |
+
+**INT8 no presenta una pérdida material de precisión en este benchmark:**
+con el redondeo a dos decimales habitual del challenge, FP32 e INT8 obtienen
+**91,25 % de MR-ALL**; la diferencia sin redondear es de solo 0,0011 puntos
+porcentuales. A la vez, conserva las ventajas mostradas arriba de 3,73 veces la
+capacidad medida y 3,35 veces el rendimiento Top-5 sobre 10M. Esta comparación
+mide la precisión del almacenamiento y la búsqueda vectorial, no la inferencia
+de modelos INT8.
+
+![Gestión de Collections en inglés](docs/images/customer/collections-en.jpg)
+
+![Monitor RTSP en inglés; dirección privada ocultada](docs/images/customer/monitoring-en.jpg)
+
+## Inicio rápido
+
+Requisitos:
+
+- Linux x86_64 con Docker Engine y Docker Compose;
+- para CUDA, GPU NVIDIA compatible, NVIDIA Driver y NVIDIA Container Toolkit.
+
+El host no necesita Python, OpenCV, ONNX Runtime, CUDA Toolkit ni cuDNN. Las
+imágenes públicas no incluyen modelos, datos de clientes, API Keys ni
+configuración de producción.
+
+Desde un checkout completo de InsightFace, instale un modelo en
+`server/.models`:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+La herramienta también admite `buffalo_m`, `buffalo_sc` y `antelopev2`.
+Escribe `manifest.json` y el `MODEL.LICENSE` firmado; `models verify` comprueba
+el paquete. Los términos del modelo son independientes de la licencia del
+código Server.
+
+Iniciar CPU:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Iniciar CUDA 12 en su lugar:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Abra `http://SERVIDOR:18097/` para CPU o `http://SERVIDOR:18098/` para CUDA.
+Cree una Collection, registre una Person con una o varias fotos y busque con
+otra. Use `docker compose ... down` sin `-v` para conservar el volumen.
+
+Los Compose incluidos desactivan la autenticación por defecto para evaluación
+aislada. Antes de exponer el servicio a otras personas o redes:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='sustituya-por-un-secreto-aleatorio-largo'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Consulte la [guía para principiantes](docs/user-guide.es.md) para el flujo
+completo.
+
+## Compilar desde el código fuente
+
+Los Dockerfiles copian `server/` y módulos de inferencia seleccionados de
+`python-package/insightface/`, por lo que el repositorio completo es el contexto
+de compilación.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` obliga a usar la imagen local. La compilación aún descarga las
+imágenes base y dependencias fijadas; la instalación descarga por separado el
+paquete de modelo cuya licencia se aceptó.
+
+## Comportamiento esencial
+
+- Similarity es el coseno original, no una probabilidad. Los thresholds usan
+  `0.0..1.0` y el valor predeterminado es `0.4`.
+- Una Collection fija el modelo y el embedding contract. Si no coinciden, sigue
+  visible pero registro/búsqueda devuelve `collection_model_mismatch`.
+- El Detection Profile de inicio se copia a nuevas Collections; después puede
+  modificarse de forma independiente para las siguientes peticiones.
+- El guardado opcional conserva un bounding-box JPEG crop redimensionado a
+  112x112, no el original ni la entrada alineada de reconocimiento; está
+  desactivado por defecto.
+- Los commits SQLite son autoritativos. El índice se sincroniza antes de una
+  respuesta correcta de registro/borrado y se reconstruye tras reiniciar.
+- Las respuestas incluyen `x-request-id`; las listas usan cursor firmados y
+  opacos.
+
+Los campos exactos, defaults, ciclos de vida y errores están únicamente en la
+documentación detallada enlazada a continuación.
+
+## API y SDK
+
+Grupos principales:
+
+- sistema: `/v1/health`, `/v1/system`, `/v1/models`;
+- operaciones sin estado: `/v1/detect`, `/v1/compare`, `/v1/embeddings`;
+- CRUD de Collection, Person y FaceSample;
+- búsqueda de Person en Collection;
+- configuración, estado, eventos y vista previa de RTSP Monitor.
+
+La [guía REST API](docs/api.es.md) contiene parámetros, respuestas, errores y
+ejemplos. OpenAPI interactivo sigue disponible en `/docs`.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+La instalación, entradas, métodos y flujos completos del SDK están en la
+[guía de usuario](docs/user-guide.es.md).
+
+## Seguridad
+
+Las imágenes faciales y embeddings son datos biométricos. En red, active la
+autenticación, termine HTTPS en un proxy inverso fiable, restrinja Docker y los
+volúmenes, mantenga deshabilitado CORS amplio y defina copias de seguridad,
+retención, borrado, consentimiento y respuesta a incidentes. No registre
+imágenes, embeddings, credenciales RTSP ni API Keys.
+
+El Server no incluye TLS, cuentas, RBAC, cloud IAM ni una capa de cumplimiento
+legal. La operación y seguridad están en la
+[guía de usuario](docs/user-guide.es.md).
+
+## Alcance de la primera fase
+
+No incluye compatibilidad AWS/CompreFace, CUDA 11, Jetson, ARM64, Windows
+Container, TensorRT, Kubernetes, Workers distribuidos, eventos Monitor
+persistentes o grabación/NVR, liveness, deepfake ni análisis demográfico.
+
+## Documentación
+
+- [Guía de usuario](docs/user-guide.es.md) — instalación, configuración,
+  modelos, Web UI, SDK, GPU, seguridad, backup y resolución de problemas.
+- [Guía REST API](docs/api.es.md) — todos los endpoints, campos, comportamientos,
+  resultados, errores, paginación y ejemplos.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — arquitectura,
+  búsqueda interna, pruebas, contribuciones y releases de contenedor.
+
+GitHub y la ayuda Web UI leen los mismos Markdown localizados; solo cambia la
+presentación.
+
+## Licencia
+
+El único punto de entrada de licencias es [LICENSING.md](LICENSING.md). El
+código Server y el SDK Python usan MIT License; esta declaración no cubre
+archivos ni pesos de modelos, datasets o componentes de terceros. Los modelos
+públicos de InsightFace suelen limitarse a investigación no comercial salvo
+autorización separada. Licencia comercial: <https://www.insightface.ai>.
diff --git a/server/README.fr.md b/server/README.fr.md
new file mode 100644
index 000000000..c191cdedf
--- /dev/null
+++ b/server/README.fr.md
@@ -0,0 +1,280 @@
+# InsightFace Server
+
+**Langues :** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · Français · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **Un seul GPU. 50M+ de vecteurs faciaux. Recherche ultrarapide par quantification INT8 des caractéristiques, sans perte de précision significative.**
+
+**Un serveur de reconnaissance faciale autohébergé réunissant Web UI, REST API
+simple, SQLite et inférence locale CPU ou GPU NVIDIA dans un seul conteneur.**
+
+```text
+téléverser une image -> détecter, comparer, inscrire ou rechercher
+```
+
+> **Licence des modèles :** les modèles publics préentraînés InsightFace sont
+> généralement réservés à la recherche non commerciale. L’usage commercial
+> exige une autorisation distincte d’[InsightFace](https://www.insightface.ai).
+
+InsightFace Server est une alternative plus simple et orientée confidentialité
+à AWS Rekognition pour les usages courants sur votre propre infrastructure. Les
+images, embeddings, modèles et index peuvent rester dans votre réseau. Ce n’est
+pas un remplacement compatible AWS et il n’implémente ni SigV4, IAM, Region, ni
+la sémantique des ressources AWS.
+
+Version actuelle : **0.2.0**, Linux x86_64.
+
+| Environnement | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| GPU NVIDIA | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+Les tags mobiles `cpu` et `cuda12` désignent la dernière version stable de
+chaque famille. Aucun tag ambigu `latest` n’est fourni. Voir le
+[Maintainer Guide — English](docs/maintainer-guide.md) pour la politique.
+
+![Dashboard InsightFace Server en anglais](docs/images/customer/dashboard-en.jpg)
+
+## Fonctionnalités principales
+
+- Détection SCRFD, cinq landmarks, alignement, embeddings ArcFace, L2
+  normalization, cosine similarity brut et recherche Person 1:N exacte.
+- Détection multirésolution avec un NMS global après fusion, et sélection
+  `largest` ou `center_largest`.
+- `Collection -> Person -> FaceSample`, Collections liées au modèle,
+  inscription multi-image avec succès partiel, metadata et motifs explicites.
+- `review_mode` d’inscription : `off`, `standard` ou `strict`; embeddings
+  pré-calculés facultatifs via `external_trusted`.
+- Recherche GPU exacte avec stockage vectoriel FP32, FP16, BF16 et INT8.
+- Web UI multilingue pour Dashboard, Collections, People, Detect, Compare,
+  Search, surveillance RTSP, System et Help.
+- 29 opérations REST snake_case sous `/v1`, dont `/v1/embeddings` protégé, et
+  un SDK Python léger et typé.
+- RTSP Monitors persistants côté serveur, événements bornés en mémoire,
+  plusieurs clients et `preview.mjpeg` facultatif ; fermer le navigateur
+  n’arrête pas la surveillance.
+- SQLite comme source durable, index exacts mémoire reconstruisibles,
+  `/models` en lecture seule, `/data` persistant, migrations, health checks et
+  validation CUDA stricte sans fallback CPU silencieux.
+- Entrées JPEG, PNG et WebP ; les originaux ne sont pas conservés par défaut.
+
+### Performances de recherche GPU sur RTX 5090
+
+Sur une NVIDIA GeForce RTX 5090 (32 607 MiB), l’index CUDA flat exact natif a
+stocké jusqu’à **58,9M vecteurs d’image de 512 dimensions en INT8**.
+
+| Type de données GPU | Vecteurs d’image maximum | 10M Top-5 p50 | 10M QPS série |
+| --- | ---: | ---: | ---: |
+| FP32 | 15,8M | 12,84 ms | 77,85 |
+| FP16 | 30,7M | 6,83 ms | 146,32 |
+| BF16 | 30,7M | 6,83 ms | 146,33 |
+| INT8 | **58,9M** | **3,84 ms** | **260,81** |
+
+INT8 atteint 3,73 fois la capacité mesurée et 3,35 fois le débit Top-5 de FP32
+sur 10M. Ces mesures GPU proviennent de la même RTX 5090 avec Driver 580.105.08
+et CUDA 12.9. La capacité est la limite isolée de l’index natif, sans modèles
+ONNX chargés ni charge Server. La vitesse utilise exactement 10M vecteurs
+d’image, un balayage exact Top-5 résident GPU, une seule requête en vol, 10
+warm-ups et 100 mesures. La recherche est exacte dans chaque représentation
+stockée ; la quantification peut néanmoins modifier les scores par rapport à
+FP32. La production doit réserver de la VRAM pour modèles, requêtes,
+concurrence, reconstruction d’index et allocator.
+
+### Précision MR-ALL multiethnique ICCV21-MFR
+
+Nous avons évalué les profils de recherche natifs sur le jeu de test
+multiethnique (MR) d’[ICCV21-MFR](../challenges/iccv21-mfr/) avec son protocole
+MR-ALL 1:1 toutes paires à FAR `1e-6`. Tous les profils utilisent les mêmes
+embeddings `buffalo_l` de 512 dimensions, normalisés L2 et extraits une seule
+fois via la Server API ; seules changent les représentations de stockage et de
+calcul de la recherche.
+
+| Profil de recherche | MR-ALL à FAR 1e-6 | Seuil cosine | Écart par rapport à FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91,249107 % | 0,407787 | — |
+| FP16 | 91,249197 % | 0,407787 | +0,000090 point |
+| BF16 | 91,248502 % | 0,407787 | -0,000605 point |
+| **INT8** | **91,248005 %** | **0,407739** | **-0,001102 point** |
+
+**INT8 ne présente aucune perte de précision significative sur ce
+benchmark :** avec l’affichage à deux décimales utilisé par le challenge, FP32
+et INT8 obtiennent tous deux **91,25 % de MR-ALL** ; l’écart non arrondi n’est
+que de 0,0011 point. Les avantages indiqués plus haut — capacité mesurée
+3,73 fois supérieure et débit Top-5 sur 10M 3,35 fois supérieur — sont
+conservés. Cette comparaison mesure la précision du stockage et de la recherche
+vectorielle, pas l’inférence de modèle INT8.
+
+![Gestion des Collections en anglais](docs/images/customer/collections-en.jpg)
+
+![RTSP Monitor en anglais ; adresse privée masquée](docs/images/customer/monitoring-en.jpg)
+
+## Démarrage rapide
+
+Prérequis :
+
+- Linux x86_64 avec Docker Engine et Docker Compose ;
+- pour CUDA, GPU NVIDIA compatible, NVIDIA Driver et NVIDIA Container Toolkit.
+
+L’hôte n’a pas besoin de Python, OpenCV, ONNX Runtime, CUDA Toolkit ou cuDNN.
+Les images publiques n’incluent ni modèles, données client, API Keys, ni
+configuration de production.
+
+Depuis un checkout complet d’InsightFace, installez un modèle dans
+`server/.models` :
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+L’outil accepte aussi `buffalo_m`, `buffalo_sc` et `antelopev2`. Il écrit
+`manifest.json` et le `MODEL.LICENSE` signé ; `models verify` contrôle le
+paquet. Les conditions du modèle sont distinctes de la licence du code Server.
+
+Démarrer CPU :
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Démarrer CUDA 12 à la place :
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Ouvrez `http://SERVEUR:18097/` pour CPU ou `http://SERVEUR:18098/` pour CUDA.
+Créez une Collection, inscrivez une Person avec une ou plusieurs photos, puis
+recherchez avec une autre photo. `docker compose ... down` sans `-v` conserve
+le volume.
+
+Les fichiers Compose fournis désactivent l’authentification par défaut pour une
+évaluation isolée. Avant d’exposer le service à d’autres utilisateurs ou
+réseaux :
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='remplacer-par-un-long-secret-aleatoire'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Le [guide utilisateur](docs/user-guide.fr.md) décrit tout le premier parcours.
+
+## Construire depuis les sources
+
+Les Dockerfiles copient `server/` et certains modules d’inférence de
+`python-package/insightface/` ; le dépôt complet est donc le contexte de build.
+
+CPU :
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12 :
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` garantit l’utilisation de l’image locale. Le build télécharge
+toujours les images de base et dépendances verrouillées ; l’installation du
+modèle télécharge séparément le paquet dont la licence a été acceptée.
+
+## Comportement essentiel
+
+- Similarity est le cosinus brut, pas une probabilité. Les thresholds utilisent
+  `0.0..1.0`, valeur par défaut `0.4`.
+- Une Collection fixe le modèle et l’embedding contract. En cas d’écart elle
+  reste visible, mais inscription/recherche renvoie
+  `collection_model_mismatch`.
+- Le Detection Profile de démarrage est copié dans les nouvelles Collections ;
+  leur profil peut ensuite changer indépendamment pour les requêtes suivantes.
+- Le stockage facultatif conserve un bounding-box JPEG crop redimensionné en
+  112x112, pas l’original ni l’entrée alignée de reconnaissance ; désactivé par
+  défaut.
+- Les commits SQLite font foi. L’index est synchronisé avant toute réponse
+  d’inscription/suppression réussie et reconstruit depuis SQLite au redémarrage.
+- Les réponses contiennent `x-request-id` ; les listes utilisent des cursor
+  opaques et signés.
+
+Les champs, defaults, cycles de vie et erreurs exacts sont maintenus uniquement
+dans les documents détaillés ci-dessous.
+
+## API et SDK
+
+Groupes principaux :
+
+- système : `/v1/health`, `/v1/system`, `/v1/models` ;
+- visage sans état : `/v1/detect`, `/v1/compare`, `/v1/embeddings` ;
+- CRUD Collection, Person et FaceSample ;
+- recherche de Person dans une Collection ;
+- configuration, état, événements et aperçu des RTSP Monitors.
+
+Le [guide REST API](docs/api.fr.md) contient tous les paramètres, réponses,
+erreurs et exemples. OpenAPI interactif reste disponible sous `/docs`.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+L’installation, les entrées, les méthodes et les parcours complets du SDK sont
+dans le [guide utilisateur](docs/user-guide.fr.md).
+
+## Sécurité
+
+Les images faciales et embeddings sont des données biométriques. En réseau,
+activez l’authentification, terminez HTTPS sur un reverse proxy fiable,
+restreignez Docker et les volumes, laissez le CORS large désactivé et définissez
+backup, rétention, suppression, consentement et réponse aux incidents. Ne
+journalisez ni images, embeddings, identifiants RTSP, ni API Keys.
+
+Le Server n’inclut ni TLS, comptes utilisateurs, RBAC, cloud IAM, ni couche de
+conformité légale. L’exploitation et la sécurité sont dans le
+[guide utilisateur](docs/user-guide.fr.md).
+
+## Périmètre de la première phase
+
+Sont exclus : compatibilité AWS/CompreFace, CUDA 11, Jetson, ARM64, Windows
+Container, TensorRT, Kubernetes, Workers distribués, événements Monitor
+persistants ou enregistrement/NVR, liveness, deepfake et attributs
+démographiques.
+
+## Documentation
+
+- [Guide utilisateur](docs/user-guide.fr.md) — installation, configuration,
+  modèles, Web UI, SDK, GPU, sécurité, backup et dépannage.
+- [Guide REST API](docs/api.fr.md) — tous les endpoints, champs, comportements,
+  résultats, erreurs, pagination et exemples.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — architecture,
+  recherche interne, tests, contribution et releases des conteneurs.
+
+GitHub et l’aide Web UI lisent les mêmes Markdown localisés ; seule la
+présentation diffère.
+
+## Licence
+
+Le point d’entrée unique des licences est [LICENSING.md](LICENSING.md). Le code
+Server et le SDK Python sont sous MIT License ; cette déclaration ne couvre ni
+fichiers ou poids de modèles, ni datasets, ni composants tiers. Les modèles
+publics InsightFace sont généralement limités à la recherche non commerciale
+sans autorisation séparée. Licence commerciale :
+<https://www.insightface.ai>.
diff --git a/server/README.ja.md b/server/README.ja.md
new file mode 100644
index 000000000..f2b4f52ca
--- /dev/null
+++ b/server/README.ja.md
@@ -0,0 +1,267 @@
+# InsightFace Server
+
+**言語:** [English](README.md) · [中文](README.zh-CN.md) · 日本語 · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **GPU 1基で50M+の顔ベクトル。INT8特徴量子化による高速検索を、実質的な精度低下なく実現。**
+
+**Web UI、分かりやすい REST API、SQLite、ローカル CPU または NVIDIA GPU
+推論を1つのコンテナで提供するセルフホスト型顔認識サーバーです。**
+
+```text
+画像をアップロード -> 検出、比較、登録、検索
+```
+
+> **モデルライセンス:** 公開 InsightFace 学習済みモデルは通常、非商用研究用途に
+> 限定されます。商用利用には個別の許諾が必要です。
+> [InsightFace 公式サイト](https://www.insightface.ai)からお問い合わせください。
+
+InsightFace Server は、管理下のインフラで一般的な顔認識を行うための、AWS
+Rekognition よりシンプルでプライバシー重視の選択肢です。画像、Embedding、
+モデル、索引をネットワーク内に保持できます。AWS 互換製品ではなく、SigV4、
+IAM、Region、AWS リソース意味論は実装しません。
+
+現在のリリース: **0.2.0**、Linux x86_64。
+
+| Runtime | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+移動タグ `cpu` と `cuda12` は各 Runtime の最新安定版を示します。曖昧な
+`latest` は提供しません。リリース方針は
+[Maintainer Guide — English](docs/maintainer-guide.md)を参照してください。
+
+![英語版 InsightFace Server Dashboard](docs/images/customer/dashboard-en.jpg)
+
+## 主な機能
+
+- SCRFD 検出、5点ランドマーク、アラインメント、ArcFace Embedding、L2
+  normalization、元の cosine similarity、厳密な 1:N Person 検索。
+- 複数解像度の候補を統合して1回の NMS を実行し、`largest` と
+  `center_largest` の単一顔選択をサポート。
+- `Collection -> Person -> FaceSample`、モデルに固定された Collection、
+  複数画像登録、部分成功、metadata、明確な拒否理由。
+- 登録 `review_mode` は `off`、`standard`、`strict`。`external_trusted`
+  の事前計算済み Embedding も利用可能。
+- GPU 厳密検索は FP32、FP16、BF16、INT8 のベクトル保存に対応。
+- Dashboard、Collections、People、Detect、Compare、Search、RTSP Monitor、
+  System、Help を備えた多言語 Web UI。
+- `/v1` 配下の29個の snake_case REST operation。保護された
+  `/v1/embeddings` と軽量 Python SDK を含みます。
+- 有界メモリイベント、複数クライアント、任意の `preview.mjpeg` を備えた
+  サーバー側 RTSP Monitor。ブラウザを閉じても監視は停止しません。
+- SQLite を永続的な正本とし、再構築可能なメモリ内厳密索引、読み取り専用
+  `/models`、永続 `/data`、migration、health check、CPU fallback を許さない
+  CUDA 起動検証を提供。
+- JPEG、PNG、WebP をサポートし、元のアップロード画像は既定で保持しません。
+
+### RTX 5090 GPU 検索性能
+
+1基の NVIDIA GeForce RTX 5090（32,607 MiB）で、ネイティブ CUDA
+exact-flat index は INT8 使用時に **58.9M 個の512次元画像ベクトル**を
+格納できました。
+
+| GPU データ型 | 最大画像ベクトル数 | 10M Top-5 p50 | 10M serial QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15.8M | 12.84 ms | 77.85 |
+| FP16 | 30.7M | 6.83 ms | 146.32 |
+| BF16 | 30.7M | 6.83 ms | 146.33 |
+| INT8 | **58.9M** | **3.84 ms** | **260.81** |
+
+INT8 は FP32 に対し、実測容量 3.73 倍、10M Top-5 throughput 3.35 倍でした。
+数値は同一の RTX 5090、Driver 580.105.08、CUDA 12.9 による GPU
+測定です。容量は ONNX モデルや Server 負荷を含まない独立ネイティブ索引の
+上限です。速度は正確に 10M 個の画像ベクトルを対象に、GPU 常駐 Top-5
+全件厳密走査、同時1 query、warm-up 10回と測定100回で計測しました。
+索引は各保存表現内では厳密ですが、量子化により FP32 から score が変わる
+場合があります。本番ではモデル、request、並行処理、索引再構築、allocator
+のための VRAM 余裕が必要です。
+
+### ICCV21-MFR 多人種 MR-ALL 精度
+
+[ICCV21-MFR](../challenges/iccv21-mfr/) の多人種（MR）テストセットで、
+MR-ALL の全組合せ 1:1 プロトコルと FAR `1e-6` を用いてネイティブ検索
+profile を評価しました。すべての profile は Server API で一度だけ抽出し
+L2 normalization を施した同一の512次元 `buffalo_l` embedding を使用し、
+ベクトルの保存形式と検索時の計算表現だけを変更しています。
+
+| 検索 profile | FAR 1e-6 での MR-ALL | Cosine threshold | FP32 との差 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91.249107% | 0.407787 | — |
+| FP16 | 91.249197% | 0.407787 | +0.000090ポイント |
+| BF16 | 91.248502% | 0.407787 | -0.000605ポイント |
+| **INT8** | **91.248005%** | **0.407739** | **-0.001102ポイント** |
+
+**このベンチマークでは INT8 に実質的な精度低下はありません。**
+challenge と同じ小数第2位までの表示では FP32 と INT8 の MR-ALL はともに
+**91.25%** で、丸め前の差もわずか0.0011ポイントです。同時に、上記の実測
+容量3.73倍と 10M Top-5 throughput 3.35倍の利点を維持します。この比較は
+ベクトル保存・検索精度の評価であり、INT8 モデル推論の評価ではありません。
+
+![英語版 Collection 管理](docs/images/customer/collections-en.jpg)
+
+![英語版 RTSP Monitor。プライベートアドレスはマスク済み](docs/images/customer/monitoring-en.jpg)
+
+## クイックスタート
+
+必要条件:
+
+- Docker Engine と Docker Compose を備えた Linux x86_64。
+- CUDA 版は、対応 NVIDIA GPU、NVIDIA Driver、NVIDIA Container Toolkit。
+
+ホストに Python、OpenCV、ONNX Runtime、CUDA Toolkit、cuDNN は不要です。
+公開 Image にモデル、顧客データ、API Key、本番設定は含まれません。
+
+InsightFace リポジトリ全体の checkout で、`server/.models` にモデルを
+インストールします。
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+モデルツールは `buffalo_m`、`buffalo_sc`、`antelopev2` にも対応します。
+`manifest.json` と署名済み `MODEL.LICENSE` を作成し、`models verify` で
+検証できます。モデル条件は Server ソースライセンスとは別です。
+
+CPU を起動:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+代わりに CUDA 12 を起動:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+CPU は `http://SERVER:18097/`、CUDA は `http://SERVER:18098/` を開きます。
+Collection を作成し、Person に1枚以上の画像を登録して、別の画像で検索します。
+データを保持するには `-v` を付けずに `docker compose ... down` します。
+
+同梱 Compose は隔離評価向けに認証を既定で無効にします。他のユーザーや
+ネットワークへ公開する前に次を設定します。
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='十分に長いランダムな秘密へ置換'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+完全な初回手順は[初心者向けユーザーガイド](docs/user-guide.ja.md)を参照してください。
+
+## ソースからビルド
+
+Dockerfile は `server/` と `python-package/insightface/` の選択された推論
+モジュールをコピーするため、リポジトリ全体が build context です。
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` によりローカル build の Image を使用します。build 時には固定済み
+base image と依存を、モデル install 時には許諾したモデル package を別途
+download します。
+
+## コア動作
+
+- Similarity は確率ではなく元の cosine 値です。Threshold は `0.0..1.0`、
+  既定は `0.4`。
+- Collection はモデルと embedding contract に固定されます。不一致でも表示は
+  できますが、登録/検索は `collection_model_mismatch` を返します。
+- 起動時 Detection Profile は新しい Collection にコピーされ、その後の
+  Collection Profile は独立して変更でき、次の Request から有効になります。
+- 任意の顔保存は 112x112 にリサイズした bounding-box JPEG crop であり、
+  元画像や認識用のアライン済み入力ではありません。既定は無効です。
+- SQLite commit が正本です。登録/削除の成功応答前に索引を同期し、再起動後は
+  SQLite から再構築します。
+- Response は `x-request-id` を持ち、一覧 API は不透明な署名 cursor を使います。
+
+正確な field、default、lifecycle、failure behavior は以下の詳細文書が正本です。
+
+## API と SDK
+
+主な API group:
+
+- System: `/v1/health`、`/v1/system`、`/v1/models`。
+- Stateless face: `/v1/detect`、`/v1/compare`、`/v1/embeddings`。
+- Collection、Person、FaceSample CRUD。
+- Collection の Person search。
+- RTSP Monitor の設定、state、events、preview。
+
+全 parameter、response、error、example は
+[REST API 利用ガイド](docs/api.ja.md)を参照してください。Interactive OpenAPI
+は `/docs` に残ります。
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+SDK のインストール、入力形式、メソッド、完全な手順は
+[ユーザーガイド](docs/user-guide.ja.md)を参照してください。
+
+## セキュリティ
+
+顔画像と Embedding は生体情報です。ネットワーク公開時は認証を有効にし、信頼する
+reverse proxy で HTTPS を終端し、Docker と volume へのアクセスを制限し、広い
+CORS を無効のまま保ち、backup、retention、deletion、consent、incident response
+を定義してください。画像、Embedding、RTSP credential、API Key を log に
+記録しないでください。
+
+Server は TLS、ユーザーアカウント、RBAC、cloud IAM、法令遵守レイヤーを内蔵
+しません。運用とセキュリティは[ユーザーガイド](docs/user-guide.ja.md)を
+参照してください。
+
+## フェーズ1の範囲
+
+AWS/CompreFace 互換、CUDA 11、Jetson、ARM64、Windows Container、
+TensorRT、Kubernetes、分散 Worker、Monitor event の永続化、録画/NVR、
+liveness、deepfake、属性分析は実装しません。
+
+## ドキュメント
+
+- [ユーザーガイド](docs/user-guide.ja.md) — インストール、設定、モデル、
+  Web UI、SDK、GPU、セキュリティ、バックアップ、トラブルシュート。
+- [REST API 利用ガイド](docs/api.ja.md) — 全公開 Endpoint、Field、挙動、
+  Result、Error、Pagination、Example。
+- [Maintainer Guide — English](docs/maintainer-guide.md) — Architecture、
+  Search 内部、Test、Contribution、Container Release。
+
+GitHub と Web UI Help は同じローカライズ済み User Guide/API Guide Markdown
+を読み込み、表示方法だけが異なります。
+
+## ライセンス
+
+ライセンスの唯一の入口は [LICENSING.md](LICENSING.md) です。Server
+ソースと Python SDK は MIT License ですが、この宣言はモデルファイル、
+モデル重み、データセット、第三者コンポーネントには適用されません。公開
+InsightFace 学習済みモデルは、別途許諾がない限り通常は非商用研究用途に
+限定されます。商用ライセンス: <https://www.insightface.ai>。
diff --git a/server/README.ko.md b/server/README.ko.md
new file mode 100644
index 000000000..c42be2601
--- /dev/null
+++ b/server/README.ko.md
@@ -0,0 +1,267 @@
+# InsightFace Server
+
+**언어:** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · 한국어
+
+> **GPU 한 장으로 50M+ 얼굴 벡터. INT8 특징 양자화 기반 초고속 검색을 실질적인 정확도 손실 없이 제공합니다.**
+
+**Web UI, 직관적인 REST API, SQLite, 로컬 CPU 또는 NVIDIA GPU 추론을
+하나의 컨테이너로 제공하는 셀프 호스팅 얼굴 인식 서버입니다.**
+
+```text
+이미지 업로드 -> 검출, 비교, 등록 또는 검색
+```
+
+> **모델 라이선스:** 공개 InsightFace 사전 학습 모델은 일반적으로 비상업적 연구
+> 용도로만 제공됩니다. 상업적 사용에는
+> [InsightFace](https://www.insightface.ai)의 별도 허가가 필요합니다.
+
+InsightFace Server는 직접 관리하는 인프라에서 일반적인 얼굴 인식 워크플로를
+실행하기 위한, AWS Rekognition보다 단순하고 개인정보 보호에 초점을 둔
+대안입니다. 이미지, embedding, 모델, 인덱스를 자체 네트워크 안에 둘 수 있습니다.
+AWS 호환 대체품은 아니며 SigV4, IAM, Region 또는 AWS 리소스 의미 체계를
+구현하지 않습니다.
+
+현재 릴리스: **0.2.0**, Linux x86_64.
+
+| Runtime | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+이동 태그 `cpu`와 `cuda12`는 각 Runtime 계열의 최신 안정 버전을 가리킵니다.
+모호한 `latest` 태그는 사용하지 않습니다.
+[Maintainer Guide — English](docs/maintainer-guide.md)의 릴리스 정책을
+참고하세요.
+
+![영문 InsightFace Server Dashboard](docs/images/customer/dashboard-en.jpg)
+
+## 주요 기능
+
+- SCRFD 얼굴 검출, 5점 landmark, 정렬, ArcFace embedding, L2
+  normalization, 원본 cosine similarity, 정확한 1:N Person 검색.
+- 여러 해상도의 후보를 합친 뒤 한 번의 NMS를 수행하며 `largest`와
+  `center_largest` 단일 얼굴 선택을 지원.
+- `Collection -> Person -> FaceSample`, 모델에 고정된 Collection,
+  다중 이미지 등록의 부분 성공, metadata와 명확한 거절 사유.
+- 등록 `review_mode`: `off`, `standard`, `strict`; 선택적인
+  `external_trusted` 사전 계산 embedding.
+- FP32, FP16, BF16, INT8 벡터 저장을 사용하는 정확한 GPU 검색.
+- Dashboard, Collections, People, Detect, Compare, Search, RTSP 모니터링,
+  System, Help를 제공하는 다국어 Web UI.
+- `/v1` 아래 29개 snake_case REST operation, 보호되는
+  `/v1/embeddings`, 가볍고 타입이 지정된 Python SDK.
+- 제한된 메모리 이벤트, 여러 클라이언트, 선택적 `preview.mjpeg`를 지원하는
+  서버 측 RTSP Monitor. 브라우저를 닫아도 모니터링은 중지되지 않습니다.
+- SQLite를 영구 원본으로 사용하고, 재구축 가능한 메모리 정확 인덱스,
+  읽기 전용 `/models`, 영구 `/data`, migration, health check, 조용한 CPU
+  fallback을 금지하는 엄격한 CUDA 시작 검증 제공.
+- JPEG, PNG, WebP 입력을 지원하며 원본 업로드는 기본적으로 보관하지 않음.
+
+### RTX 5090 GPU 검색 성능
+
+NVIDIA GeForce RTX 5090 한 장(32,607 MiB)에서 네이티브 CUDA exact-flat
+인덱스는 INT8로 최대 **58.9M개의 512차원 이미지 벡터**를 저장했습니다.
+
+| GPU 데이터 유형 | 최대 이미지 벡터 수 | 10M Top-5 p50 | 10M 직렬 QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15.8M | 12.84 ms | 77.85 |
+| FP16 | 30.7M | 6.83 ms | 146.32 |
+| BF16 | 30.7M | 6.83 ms | 146.33 |
+| INT8 | **58.9M** | **3.84 ms** | **260.81** |
+
+INT8은 FP32 대비 실측 용량 3.73배, 10M Top-5 처리량 3.35배를 기록했습니다.
+모든 값은 같은 RTX 5090, Driver 580.105.08, CUDA 12.9에서 측정한 GPU
+결과입니다. 용량은 ONNX 모델과 Server 부하를 제외한 독립 네이티브 인덱스의
+한계입니다. 속도는 정확히 10M 이미지 벡터, GPU 상주 Top-5 전체 정확 스캔,
+동시 query 1개, warm-up 10회와 측정 100회 조건입니다. 검색은 각 저장 표현
+내에서는 정확하지만 양자화로 인해 FP32 대비 score가 달라질 수 있습니다.
+운영 환경에서는 모델, request, 동시성, 인덱스 재구축, allocator를 위한 VRAM
+여유가 필요합니다.
+
+### ICCV21-MFR 다인종 MR-ALL 정확도
+
+[ICCV21-MFR](../challenges/iccv21-mfr/)의 다인종(MR) 테스트 세트에서 MR-ALL
+전체 쌍 1:1 프로토콜과 FAR `1e-6` 조건으로 네이티브 검색 profile을
+평가했습니다. 모든 profile은 Server API로 한 번만 추출하고 L2 normalization한
+동일한 512차원 `buffalo_l` embedding을 사용했으며, 벡터 저장 및 검색 계산
+표현만 변경했습니다.
+
+| 검색 profile | FAR 1e-6의 MR-ALL | Cosine 임계값 | FP32 대비 차이 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91.249107% | 0.407787 | — |
+| FP16 | 91.249197% | 0.407787 | +0.000090%p |
+| BF16 | 91.248502% | 0.407787 | -0.000605%p |
+| **INT8** | **91.248005%** | **0.407739** | **-0.001102%p** |
+
+**이 benchmark에서 INT8은 실질적인 정확도 손실이 없습니다.** challenge와
+같이 소수점 둘째 자리까지 표시하면 FP32와 INT8 모두 **91.25% MR-ALL**이며,
+반올림 전 차이도 0.0011%p에 불과합니다. 동시에 위에 제시한 실측 용량 3.73배와
+10M Top-5 처리량 3.35배의 이점을 유지합니다. 이 비교는 벡터 저장 및 검색
+정밀도를 측정한 것이며 INT8 모델 추론을 측정한 것이 아닙니다.
+
+![영문 Collection 관리](docs/images/customer/collections-en.jpg)
+
+![영문 RTSP Monitor, 비공개 주소 마스킹](docs/images/customer/monitoring-en.jpg)
+
+## 빠른 시작
+
+요구 사항:
+
+- Docker Engine과 Docker Compose가 설치된 Linux x86_64;
+- CUDA는 지원되는 NVIDIA GPU, NVIDIA Driver, NVIDIA Container Toolkit.
+
+호스트에 Python, OpenCV, ONNX Runtime, CUDA Toolkit 또는 cuDNN을 설치할
+필요가 없습니다. 공개 Image에는 모델, 고객 데이터, API Key, 운영 설정이
+포함되지 않습니다.
+
+전체 InsightFace 저장소 checkout에서 모델을 `server/.models`에 설치합니다.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+모델 도구는 `buffalo_m`, `buffalo_sc`, `antelopev2`도 지원합니다.
+`manifest.json`과 서명된 `MODEL.LICENSE`를 기록하며 `models verify`로
+검증할 수 있습니다. 모델 조건은 Server 소스 라이선스와 별개입니다.
+
+CPU 시작:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+대신 CUDA 12 시작:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+CPU는 `http://SERVER:18097/`, CUDA는 `http://SERVER:18098/`를 엽니다.
+Collection을 만들고 한 장 이상의 사진으로 Person을 등록한 뒤 다른 사진으로
+검색합니다. 데이터 volume을 유지하려면 `-v` 없이
+`docker compose ... down`을 사용합니다.
+
+제공되는 Compose는 격리 평가를 위해 인증을 기본적으로 끕니다. 다른 사용자나
+네트워크에 공개하기 전에:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='충분히-긴-무작위-비밀로-교체'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+전체 첫 실행 절차는 [초보자 사용자 가이드](docs/user-guide.ko.md)를 참고하세요.
+
+## 소스에서 빌드
+
+Dockerfile은 `server/`와 `python-package/insightface/`의 선택된 추론 모듈을
+복사하므로 전체 저장소가 build context입니다.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never`는 로컬 빌드 Image 사용을 보장합니다. 빌드 중에는 고정된 base
+Image와 의존성을, 모델 설치 중에는 라이선스를 수락한 모델 package를 별도로
+다운로드합니다.
+
+## 핵심 동작
+
+- Similarity는 확률이 아닌 원본 cosine 값입니다. Threshold는 `0.0..1.0`,
+  기본값은 `0.4`입니다.
+- Collection은 모델과 embedding contract에 고정됩니다. 불일치 시에도 보이지만
+  등록/검색은 `collection_model_mismatch`를 반환합니다.
+- 시작 Detection Profile은 새 Collection에 복사되고, 이후 Profile은 다음
+  요청부터 독립적으로 변경할 수 있습니다.
+- 선택적 얼굴 저장은 112x112로 조정된 bounding-box JPEG crop이며 원본이나
+  인식용 정렬 입력이 아닙니다. 기본적으로 꺼져 있습니다.
+- SQLite commit이 원본입니다. 등록/삭제 성공 응답 전에 인덱스를 동기화하고
+  재시작 후 SQLite에서 다시 구축합니다.
+- 응답은 `x-request-id`를 포함하며 목록은 불투명한 서명 cursor를 사용합니다.
+
+정확한 field, default, lifecycle, 오류 동작은 아래 상세 문서에서만 관리합니다.
+
+## API와 SDK
+
+주요 그룹:
+
+- 시스템: `/v1/health`, `/v1/system`, `/v1/models`;
+- 무상태 얼굴: `/v1/detect`, `/v1/compare`, `/v1/embeddings`;
+- Collection, Person, FaceSample CRUD;
+- Collection Person 검색;
+- RTSP Monitor 설정, 상태, 이벤트, 미리보기.
+
+모든 parameter, response, error, example은
+[REST API 가이드](docs/api.ko.md)를 참고하세요. 대화형 OpenAPI는 `/docs`에
+유지됩니다.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+SDK 설치, 입력 형식, 메서드와 전체 흐름은
+[사용자 가이드](docs/user-guide.ko.md)를 참고하세요.
+
+## 보안
+
+얼굴 이미지와 embedding은 생체 데이터입니다. 네트워크 배포에서는 인증을 켜고,
+신뢰하는 reverse proxy에서 HTTPS를 종료하며, Docker와 volume 접근을 제한하고,
+광범위한 CORS를 끈 상태로 유지하고, backup, 보존, 삭제, 동의, incident response
+정책을 정의하세요. 이미지, embedding, RTSP credential, API Key를 로그에
+남기지 마세요.
+
+Server는 TLS, 사용자 계정, RBAC, cloud IAM 또는 법적 준수 계층을 내장하지
+않습니다. 운영과 보안은 [사용자 가이드](docs/user-guide.ko.md)를 참고하세요.
+
+## 1단계 범위
+
+AWS/CompreFace 호환, CUDA 11, Jetson, ARM64, Windows Container, TensorRT,
+Kubernetes, 분산 Worker, 영구 Monitor 이벤트 또는 녹화/NVR, liveness,
+deepfake, 인구통계 속성은 구현하지 않습니다.
+
+## 문서
+
+- [사용자 가이드](docs/user-guide.ko.md) — 설치, 설정, 모델, Web UI, SDK,
+  GPU, 보안, 백업과 문제 해결.
+- [REST API 가이드](docs/api.ko.md) — 모든 endpoint, 필드, 동작, 결과, 오류,
+  pagination과 예시.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — 아키텍처, 검색 내부,
+  테스트, 기여 규칙과 컨테이너 릴리스.
+
+GitHub와 Web UI 도움말은 동일한 현지화 Markdown을 읽으며 표시 방식만 다릅니다.
+
+## 라이선스
+
+단일 라이선스 진입점은 [LICENSING.md](LICENSING.md)입니다. Server 소스와
+Python SDK는 MIT License이지만 이 선언은 모델 파일, 모델 가중치, dataset,
+타사 컴포넌트에는 적용되지 않습니다. 공개 InsightFace 모델은 별도 허가가
+없다면 일반적으로 비상업적 연구로 제한됩니다. 상업 라이선스:
+<https://www.insightface.ai>.
diff --git a/server/README.md b/server/README.md
new file mode 100644
index 000000000..722e978e4
--- /dev/null
+++ b/server/README.md
@@ -0,0 +1,281 @@
+# InsightFace Server
+
+**Languages:** English · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **One GPU. 50M+ face vectors. High-speed INT8-quantized search with no material accuracy loss.**
+
+**A self-hosted face-recognition server with a Web UI, a straightforward REST
+API, SQLite, and local CPU or NVIDIA GPU inference in one container.**
+
+```text
+upload an image -> detect, compare, enroll, or search
+```
+
+> **Model license:** Public InsightFace pretrained models are generally
+> available for non-commercial research only. Commercial use requires separate
+> authorization from [InsightFace](https://www.insightface.ai).
+
+InsightFace Server is a simpler, privacy-focused alternative to AWS Rekognition
+for common face-recognition workflows on infrastructure you control. Images,
+embeddings, models, and indexes can remain inside your network. It is **not** an
+AWS-compatible replacement and does not implement SigV4, IAM, Region, or AWS
+resource semantics.
+
+Current release: **0.2.0**, Linux x86_64.
+
+| Runtime | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+The moving `cpu` and `cuda12` tags identify the latest stable release in each
+runtime family. There is no ambiguous `latest` tag. See
+[Maintainer Guide — English](docs/maintainer-guide.md) for the release policy.
+
+![InsightFace Server dashboard in English](docs/images/customer/dashboard-en.jpg)
+
+## Highlights
+
+- SCRFD face detection, five landmarks, alignment, ArcFace embeddings, L2
+  normalization, raw cosine similarity, and exact 1:N Person search.
+- Multi-resolution detection with one merged NMS and `largest` or
+  `center_largest` single-face selection.
+- `Collection -> Person -> FaceSample` storage with model-bound Collections,
+  multi-image enrollment, partial success, metadata, and explicit rejection
+  reasons.
+- Enrollment `review_mode`: `off`, `standard`, or `strict`; optional
+  `external_trusted` precomputed embeddings.
+- Exact GPU search with FP32, FP16, BF16, and INT8 vector storage.
+- Multilingual Web UI for Dashboard, Collections, People, Detect, Compare,
+  Search, RTSP monitoring, System diagnostics, and Help.
+- 29 snake_case REST operations under `/v1`, including authenticated
+  `/v1/embeddings`, plus a typed lightweight Python SDK.
+- Persistent server-side RTSP Monitors with bounded in-memory events,
+  independent clients, and optional `preview.mjpeg`; closing the browser does
+  not stop monitoring.
+- SQLite as the durable source of truth, disposable in-memory exact indexes,
+  read-only `/models`, persistent `/data`, migrations, health checks, and
+  strict CUDA startup validation without silent CPU fallback.
+- JPEG, PNG, and WebP input; original uploads are not retained by default.
+
+### RTX 5090 GPU search performance
+
+On one NVIDIA GeForce RTX 5090 (32,607 MiB), the native CUDA exact-flat index
+stored up to **58.9M 512-dimensional image vectors in INT8**.
+
+| GPU data type | Maximum image vectors | 10M Top-5 p50 | 10M serial QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15.8M | 12.84 ms | 77.85 |
+| FP16 | 30.7M | 6.83 ms | 146.32 |
+| BF16 | 30.7M | 6.83 ms | 146.33 |
+| INT8 | **58.9M** | **3.84 ms** | **260.81** |
+
+INT8 delivered 3.73x the measured capacity and 3.35x the 10M Top-5 throughput
+of FP32. These are GPU-only measurements on the same RTX 5090 with Driver
+580.105.08 and CUDA 12.9. Capacity is an isolated native-index ceiling without
+loaded ONNX models or Server workload. Speed uses exactly 10M image vectors,
+exhaustive GPU-resident Top-5, one query in flight, 10 warm-ups, and 100
+measured queries. The index is exact within each stored representation;
+quantization can still change scores relative to FP32. A production deployment
+must reserve VRAM for models, requests, concurrency, index rebuilds, and
+allocator headroom.
+
+### ICCV21-MFR multi-racial MR-ALL accuracy
+
+We evaluated the native search profiles on the multi-racial (MR) test set from
+[ICCV21-MFR](../challenges/iccv21-mfr/) using its MR-ALL all-pairs 1:1 protocol
+at FAR `1e-6`. Every profile used the same L2-normalized 512-dimensional
+`buffalo_l` embeddings extracted once through the Server API; only the stored
+vector and search-compute representation changed.
+
+| Search profile | MR-ALL at FAR 1e-6 | Cosine threshold | Difference from FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91.249107% | 0.407787 | — |
+| FP16 | 91.249197% | 0.407787 | +0.000090 pp |
+| BF16 | 91.248502% | 0.407787 | -0.000605 pp |
+| **INT8** | **91.248005%** | **0.407739** | **-0.001102 pp** |
+
+**INT8 has no material accuracy loss on this benchmark:** at the
+challenge-style two-decimal reporting precision, FP32 and INT8 both score
+**91.25% MR-ALL**, while the unrounded difference is only 0.0011 percentage
+points. This preserves the 3.73x measured capacity and 3.35x 10M Top-5
+throughput advantages shown above. The comparison measures vector
+storage/search precision, not INT8 model inference.
+
+![Collection management in English](docs/images/customer/collections-en.jpg)
+
+![RTSP Monitor in English; private address redacted](docs/images/customer/monitoring-en.jpg)
+
+## Quick start
+
+Requirements:
+
+- Linux x86_64 with Docker Engine and Docker Compose;
+- for CUDA: a supported NVIDIA GPU, NVIDIA Driver, and NVIDIA Container
+  Toolkit.
+
+The host does not need Python, OpenCV, ONNX Runtime, CUDA Toolkit, or cuDNN.
+Public images do not contain models, customer data, API keys, or production
+configuration.
+
+From a complete InsightFace repository checkout, install a model into
+`server/.models`:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+The model tool also supports `buffalo_m`, `buffalo_sc`, and `antelopev2`. It
+writes `manifest.json` and the signed `MODEL.LICENSE`; use `models verify` to
+check the installed package. Model terms remain separate from the Server source
+license.
+
+Start CPU:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Start CUDA 12 instead:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Open `http://SERVER:18097/` for CPU or `http://SERVER:18098/` for CUDA. Create a
+Collection, register a Person with one or more photos, and search with another
+photo. Stop with `docker compose ... down` without `-v` to retain the database
+volume.
+
+The supplied Compose files set authentication to `false` for isolated
+evaluation. Before exposing the service to other users or networks:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='replace-with-a-long-random-secret'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+See the [beginner user guide](docs/user-guide.md) for the complete first-run
+workflow.
+
+## Build from source
+
+The Dockerfiles copy `server/` and selected inference modules from
+`python-package/insightface/`, so the complete repository is the build context.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` ensures Compose uses the locally built image. The build still
+downloads pinned base images and dependencies; model installation separately
+downloads the accepted model package.
+
+## Core behavior
+
+- Similarity is raw cosine, not probability. Thresholds use `0.0..1.0` and
+  default to `0.4`.
+- A Collection pins the model and embedding contract. A mismatch remains
+  visible but enrollment/search returns `collection_model_mismatch`.
+- The startup detection profile is copied into new Collections; Collection
+  profiles can then be updated independently for subsequent requests.
+- Optional face storage saves a 112x112 bounding-box JPEG crop, not the original
+  upload or the aligned recognition input, and is off by default.
+- SQLite commits are authoritative. Index mutations complete before successful
+  enrollment/deletion responses return, and indexes rebuild from SQLite after
+  restart.
+- Responses carry `x-request-id`; list APIs use opaque signed cursors.
+
+Exact fields, defaults, lifecycle rules, and failure behavior live in the
+linked documentation below.
+
+## API and SDK
+
+Main API groups:
+
+- system: `/v1/health`, `/v1/system`, `/v1/models`;
+- stateless faces: `/v1/detect`, `/v1/compare`, `/v1/embeddings`;
+- Collection, Person, and FaceSample CRUD;
+- Collection Person search;
+- RTSP Monitor configuration, state, events, and preview.
+
+Use the [complete REST API guide](docs/api.md) for every parameter, response,
+error, and example. Interactive OpenAPI remains available at `/docs`.
+
+Python:
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+See the [User Guide](docs/user-guide.md) for installation, inputs, methods, and
+complete SDK workflows.
+
+## Security
+
+Face images and embeddings are biometric data. For network deployment, enable
+authentication, terminate HTTPS at a trusted reverse proxy, restrict Docker and
+volume access, keep broad CORS disabled, and define backup, retention, deletion,
+consent, and incident-response policies. Do not log images, embeddings, RTSP
+credentials, or API keys.
+
+The Server does not provide built-in TLS, user accounts, RBAC, cloud IAM, or a
+legal-compliance layer. Deployment and security guidance is in the
+[User Guide](docs/user-guide.md).
+
+## Phase-one scope
+
+This release does not implement AWS/CompreFace compatibility, CUDA 11, Jetson,
+ARM64, Windows containers, TensorRT, Kubernetes, distributed Workers, persistent
+Monitor events or recording/NVR, liveness, deepfake detection, or demographic
+attributes.
+
+## Documentation
+
+- [User Guide](docs/user-guide.md) — complete installation, configuration,
+  model, Web UI, SDK, GPU, security, backup, and troubleshooting guidance.
+- [REST API Guide](docs/api.md) — every public endpoint, field, behavior,
+  result, error, pagination rule, and example.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — architecture,
+  search internals, testing, contribution rules, and container releases.
+
+The localized User Guide and API Guide are the exact Markdown sources rendered
+by the Web UI Help page; only their presentation differs.
+
+## License
+
+See the single [LICENSING.md](LICENSING.md) entry point. In short, the Server
+source and Python SDK are MIT licensed; that declaration does not cover model
+files, model weights, datasets, or third-party components. Public InsightFace
+pretrained models are generally limited to non-commercial research unless
+separately authorized. Commercial licensing information:
+<https://www.insightface.ai>.
diff --git a/server/README.pt.md b/server/README.pt.md
new file mode 100644
index 000000000..12a74b09d
--- /dev/null
+++ b/server/README.pt.md
@@ -0,0 +1,272 @@
+# InsightFace Server
+
+**Idiomas:** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · Português · [한국어](README.ko.md)
+
+> **Uma GPU. 50M+ de vetores faciais. Pesquisa ultrarrápida com quantização INT8 de características e sem perda material de precisão.**
+
+**Servidor de reconhecimento facial autoalojado com Web UI, REST API simples,
+SQLite e inferência local por CPU ou GPU NVIDIA num único contentor.**
+
+```text
+carregar uma imagem -> detetar, comparar, registar ou pesquisar
+```
+
+> **Licença do modelo:** os modelos públicos pré-treinados InsightFace estão
+> normalmente disponíveis apenas para investigação não comercial. O uso
+> comercial exige autorização separada da
+> [InsightFace](https://www.insightface.ai).
+
+InsightFace Server é uma alternativa mais simples e centrada na privacidade ao
+AWS Rekognition para fluxos comuns em infraestrutura própria. Imagens,
+embeddings, modelos e índices podem permanecer na sua rede. Não é um substituto
+compatível com AWS e não implementa SigV4, IAM, Region ou semântica de recursos
+AWS.
+
+Versão atual: **0.2.0**, Linux x86_64.
+
+| Ambiente | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| GPU NVIDIA | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+As tags móveis `cpu` e `cuda12` indicam a versão estável mais recente de cada
+família. Não existe a tag ambígua `latest`. Consulte
+[Maintainer Guide — English](docs/maintainer-guide.md) para a política.
+
+![Dashboard InsightFace Server em inglês](docs/images/customer/dashboard-en.jpg)
+
+## Principais funcionalidades
+
+- Deteção SCRFD, cinco landmarks, alinhamento, embeddings ArcFace, L2
+  normalization, cosine similarity original e pesquisa Person 1:N exata.
+- Deteção multirresolução com um único NMS após fusão e seleção
+  `largest` ou `center_largest`.
+- `Collection -> Person -> FaceSample`, Collections ligadas ao modelo,
+  registo multi-imagem com sucesso parcial, metadata e motivos explícitos.
+- `review_mode` de registo: `off`, `standard` ou `strict`; embeddings
+  pré-calculados opcionais por `external_trusted`.
+- Pesquisa GPU exata com armazenamento vetorial FP32, FP16, BF16 e INT8.
+- Web UI multilingue para Dashboard, Collections, People, Detect, Compare,
+  Search, monitorização RTSP, System e Help.
+- 29 operações REST snake_case em `/v1`, incluindo `/v1/embeddings` protegido,
+  mais SDK Python leve e tipado.
+- RTSP Monitors persistentes no servidor, eventos limitados em memória, vários
+  clientes e `preview.mjpeg` opcional; fechar o browser não para a monitorização.
+- SQLite como fonte persistente, índices exatos reconstruíveis em memória,
+  `/models` só de leitura, `/data` persistente, migrations, health checks e
+  validação CUDA estrita sem fallback CPU silencioso.
+- Entrada JPEG, PNG e WebP; originais não são guardados por predefinição.
+
+### Desempenho de pesquisa GPU na RTX 5090
+
+Numa NVIDIA GeForce RTX 5090 (32.607 MiB), o índice flat exato CUDA nativo
+armazenou até **58,9M vetores de imagem de 512 dimensões em INT8**.
+
+| Tipo de dados GPU | Máximo de vetores de imagem | 10M Top-5 p50 | 10M QPS em série |
+| --- | ---: | ---: | ---: |
+| FP32 | 15,8M | 12,84 ms | 77,85 |
+| FP16 | 30,7M | 6,83 ms | 146,32 |
+| BF16 | 30,7M | 6,83 ms | 146,33 |
+| INT8 | **58,9M** | **3,84 ms** | **260,81** |
+
+INT8 atingiu 3,73 vezes a capacidade medida e 3,35 vezes o throughput Top-5 de
+FP32 sobre 10M. São medições apenas de GPU na mesma RTX 5090 com Driver
+580.105.08 e CUDA 12.9. A capacidade é o limite isolado do índice nativo sem
+modelos ONNX carregados nem carga do Server. A velocidade usa exatamente 10M
+vetores de imagem, varrimento exato Top-5 residente na GPU, uma pesquisa em
+curso, 10 warm-ups e 100 medições. A pesquisa é exata dentro de cada
+representação armazenada; a quantização pode alterar scores face a FP32.
+Produção deve reservar VRAM para modelos, requests, concorrência, reconstrução
+do índice e allocator.
+
+### Precisão MR-ALL multirracial do ICCV21-MFR
+
+Avaliámos os perfis de pesquisa nativos no conjunto de teste multirracial (MR)
+do [ICCV21-MFR](../challenges/iccv21-mfr/) com o protocolo MR-ALL 1:1 de todos
+os pares e FAR `1e-6`. Todos os perfis usam os mesmos embeddings `buffalo_l` de
+512 dimensões, normalizados com L2 e extraídos uma única vez pela Server API;
+apenas mudam as representações de armazenamento e de cálculo da pesquisa.
+
+| Perfil de pesquisa | MR-ALL com FAR 1e-6 | Limiar cosine | Diferença para FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91,249107 % | 0,407787 | — |
+| FP16 | 91,249197 % | 0,407787 | +0,000090 pontos percentuais |
+| BF16 | 91,248502 % | 0,407787 | -0,000605 pontos percentuais |
+| **INT8** | **91,248005 %** | **0,407739** | **-0,001102 pontos percentuais** |
+
+**O INT8 não apresenta perda material de precisão neste benchmark:**
+com a apresentação de duas casas decimais usada pelo challenge, FP32 e INT8
+obtêm ambos **91,25 % de MR-ALL**; a diferença sem arredondamento é de apenas
+0,0011 pontos percentuais. Mantêm-se também as vantagens acima de 3,73 vezes a
+capacidade medida e 3,35 vezes o throughput Top-5 sobre 10M. Esta comparação
+mede a precisão do armazenamento e da pesquisa vetorial, não a inferência de
+modelos INT8.
+
+![Gestão de Collections em inglês](docs/images/customer/collections-en.jpg)
+
+![RTSP Monitor em inglês; endereço privado ocultado](docs/images/customer/monitoring-en.jpg)
+
+## Início rápido
+
+Requisitos:
+
+- Linux x86_64 com Docker Engine e Docker Compose;
+- para CUDA, GPU NVIDIA suportada, NVIDIA Driver e NVIDIA Container Toolkit.
+
+O host não precisa de Python, OpenCV, ONNX Runtime, CUDA Toolkit ou cuDNN.
+As Images públicas não incluem modelos, dados de clientes, API Keys ou
+configuração de produção.
+
+Num checkout completo do InsightFace, instale um modelo em `server/.models`:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+A ferramenta também suporta `buffalo_m`, `buffalo_sc` e `antelopev2`. Escreve
+`manifest.json` e o `MODEL.LICENSE` assinado; `models verify` valida o pacote.
+Os termos do modelo são separados da licença do código Server.
+
+Iniciar CPU:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Iniciar CUDA 12 em alternativa:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Abra `http://SERVIDOR:18097/` para CPU ou `http://SERVIDOR:18098/` para CUDA.
+Crie uma Collection, registe uma Person com uma ou várias fotografias e
+pesquise com outra. `docker compose ... down` sem `-v` preserva o volume.
+
+Os Compose fornecidos desativam autenticação por predefinição para avaliação
+isolada. Antes de expor o serviço a outros utilizadores ou redes:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='substitua-por-um-segredo-aleatorio-longo'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Consulte o [guia para iniciantes](docs/user-guide.pt.md) para o fluxo completo.
+
+## Compilar a partir do código
+
+Os Dockerfiles copiam `server/` e módulos de inferência selecionados de
+`python-package/insightface/`; por isso o repositório completo é o contexto de
+build.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` garante o uso da Image local. O build ainda descarrega Images
+base e dependências fixadas; a instalação descarrega separadamente o pacote de
+modelo cuja licença foi aceite.
+
+## Comportamento essencial
+
+- Similarity é o coseno original, não probabilidade. Thresholds usam
+  `0.0..1.0`, com predefinição `0.4`.
+- Uma Collection fixa modelo e embedding contract. Em caso de divergência
+  continua visível, mas registo/pesquisa devolve `collection_model_mismatch`.
+- O Detection Profile de arranque é copiado para novas Collections; depois pode
+  ser alterado independentemente para pedidos seguintes.
+- O armazenamento opcional guarda um bounding-box JPEG crop redimensionado
+  para 112x112, não o original nem a entrada alinhada de reconhecimento; vem
+  desligado por predefinição.
+- Commits SQLite são autoritativos. O índice sincroniza antes da resposta de
+  registo/remoção e é reconstruído de SQLite após reinício.
+- Respostas incluem `x-request-id`; listas usam cursor opacos e assinados.
+
+Campos, defaults, ciclos de vida e erros exatos são mantidos apenas nos
+documentos detalhados abaixo.
+
+## API e SDK
+
+Grupos principais:
+
+- sistema: `/v1/health`, `/v1/system`, `/v1/models`;
+- faces sem estado: `/v1/detect`, `/v1/compare`, `/v1/embeddings`;
+- CRUD de Collection, Person e FaceSample;
+- pesquisa Person numa Collection;
+- configuração, estado, eventos e preview de RTSP Monitor.
+
+O [guia REST API](docs/api.pt.md) contém todos os parâmetros, respostas, erros e
+exemplos. OpenAPI interativo continua em `/docs`.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+A instalação, entradas, métodos e fluxos completos do SDK estão no
+[guia do utilizador](docs/user-guide.pt.md).
+
+## Segurança
+
+Imagens faciais e embeddings são dados biométricos. Em rede, ative autenticação,
+termine HTTPS num reverse proxy de confiança, restrinja Docker e volumes,
+mantenha CORS amplo desligado e defina backup, retenção, eliminação,
+consentimento e resposta a incidentes. Não registe imagens, embeddings,
+credenciais RTSP ou API Keys.
+
+O Server não inclui TLS, contas, RBAC, cloud IAM ou camada jurídica. Operação e
+segurança estão no [guia do utilizador](docs/user-guide.pt.md).
+
+## Âmbito da primeira fase
+
+Não inclui compatibilidade AWS/CompreFace, CUDA 11, Jetson, ARM64, Windows
+Container, TensorRT, Kubernetes, Workers distribuídos, eventos Monitor
+persistentes ou gravação/NVR, liveness, deepfake ou atributos demográficos.
+
+## Documentação
+
+- [Guia do utilizador](docs/user-guide.pt.md) — instalação, configuração,
+  modelos, Web UI, SDK, GPU, segurança, backup e resolução de problemas.
+- [Guia REST API](docs/api.pt.md) — todos os endpoints, campos, comportamentos,
+  resultados, erros, paginação e exemplos.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — arquitetura,
+  pesquisa interna, testes, contribuições e releases de contentores.
+
+GitHub e a ajuda Web UI leem os mesmos Markdown localizados; só muda a
+apresentação.
+
+## Licença
+
+O ponto único de licenciamento é [LICENSING.md](LICENSING.md). O código Server
+e o SDK Python usam MIT License; esta declaração não cobre ficheiros ou pesos
+de modelos, datasets ou componentes de terceiros. Modelos públicos InsightFace
+estão normalmente limitados a investigação não comercial sem autorização
+separada. Licenciamento comercial: <https://www.insightface.ai>.
diff --git a/server/README.ru.md b/server/README.ru.md
new file mode 100644
index 000000000..4b4f7622e
--- /dev/null
+++ b/server/README.ru.md
@@ -0,0 +1,274 @@
+# InsightFace Server
+
+**Языки:** [English](README.md) · [中文](README.zh-CN.md) · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · Русский · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **Один GPU. 50M+ векторов лиц. Сверхбыстрый поиск с INT8-квантованием признаков без существенной потери точности.**
+
+**Самостоятельно размещаемый сервер распознавания лиц с Web UI, понятным REST
+API, SQLite и локальным CPU- или NVIDIA GPU-инференсом в одном контейнере.**
+
+```text
+загрузить изображение -> обнаружить, сравнить, зарегистрировать или найти
+```
+
+> **Лицензия модели:** публичные предобученные модели InsightFace обычно
+> доступны только для некоммерческих исследований. Для коммерческого применения
+> требуется отдельное разрешение [InsightFace](https://www.insightface.ai).
+
+InsightFace Server — более простая и ориентированная на приватность альтернатива
+AWS Rekognition для типовых сценариев на собственной инфраструктуре. Изображения,
+embeddings, модели и индексы могут оставаться внутри вашей сети. Это **не**
+AWS-совместимая замена: SigV4, IAM, Region и семантика ресурсов AWS не
+реализованы.
+
+Текущая версия: **0.2.0**, Linux x86_64.
+
+| Среда | Image |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+Подвижные теги `cpu` и `cuda12` указывают на последнюю стабильную версию своей
+семьи. Неоднозначный `latest` не используется. См.
+[Maintainer Guide — English](docs/maintainer-guide.md) о правилах релиза.
+
+![Английский Dashboard InsightFace Server](docs/images/customer/dashboard-en.jpg)
+
+## Основные возможности
+
+- Детекция SCRFD, пять landmarks, выравнивание, ArcFace embeddings, L2
+  normalization, исходный cosine similarity и точный поиск Person 1:N.
+- Многоразмерная детекция с единым NMS после объединения и выбором
+  `largest` либо `center_largest`.
+- `Collection -> Person -> FaceSample`, привязка Collection к модели,
+  регистрация нескольких изображений с частичным успехом, metadata и понятными
+  причинами отказа.
+- `review_mode` регистрации: `off`, `standard`, `strict`; дополнительно
+  предвычисленные `external_trusted` embeddings.
+- Точный GPU-поиск с хранением векторов FP32, FP16, BF16 и INT8.
+- Многоязычный Web UI: Dashboard, Collections, People, Detect, Compare, Search,
+  RTSP-мониторинг, System и Help.
+- 29 snake_case REST-операций под `/v1`, включая защищённый
+  `/v1/embeddings`, и лёгкий типизированный Python SDK.
+- Серверные RTSP Monitors с ограниченными событиями в памяти, несколькими
+  клиентами и необязательным `preview.mjpeg`; закрытие браузера не останавливает
+  мониторинг.
+- SQLite как постоянный источник, перестраиваемые точные индексы в памяти,
+  `/models` только для чтения, постоянный `/data`, миграции, health checks и
+  строгая проверка CUDA без скрытого CPU fallback.
+- JPEG, PNG и WebP; исходные загрузки по умолчанию не сохраняются.
+
+### Производительность GPU-поиска на RTX 5090
+
+На одной NVIDIA GeForce RTX 5090 (32 607 MiB) нативный точный CUDA flat-индекс
+с INT8 вместил до **58,9M 512-мерных векторов изображений**.
+
+| Тип данных GPU | Максимум векторов изображений | 10M Top-5 p50 | 10M последовательных QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15,8M | 12,84 ms | 77,85 |
+| FP16 | 30,7M | 6,83 ms | 146,32 |
+| BF16 | 30,7M | 6,83 ms | 146,33 |
+| INT8 | **58,9M** | **3,84 ms** | **260,81** |
+
+INT8 дал в 3,73 раза большую измеренную ёмкость и в 3,35 раза больший
+10M Top-5 throughput, чем FP32. Это только GPU-измерения на одной RTX 5090 с
+Driver 580.105.08 и CUDA 12.9. Ёмкость — изолированный предел нативного индекса
+без загруженных ONNX-моделей и нагрузки Server. Скорость измерена ровно на 10M
+векторах изображений, полным точным GPU-resident Top-5 сканированием, с одним
+запросом одновременно, 10 прогревами и 100 измерениями. Поиск точен внутри
+каждого сохранённого представления, однако квантование может изменить score
+относительно FP32. В production нужен запас VRAM для моделей, запросов,
+параллельности, перестроения индекса и allocator.
+
+### Точность MR-ALL на многорасовом наборе ICCV21-MFR
+
+Нативные профили поиска оценены на многорасовом (MR) тестовом наборе
+[ICCV21-MFR](../challenges/iccv21-mfr/) по протоколу MR-ALL 1:1 для всех пар
+при FAR `1e-6`. Во всех профилях используются одни и те же 512-мерные
+L2-нормализованные embeddings `buffalo_l`, однократно извлечённые через Server
+API; меняются только представление хранения векторов и вычисления при поиске.
+
+| Профиль поиска | MR-ALL при FAR 1e-6 | Порог cosine | Разница с FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91,249107 % | 0,407787 | — |
+| FP16 | 91,249197 % | 0,407787 | +0,000090 п. п. |
+| BF16 | 91,248502 % | 0,407787 | -0,000605 п. п. |
+| **INT8** | **91,248005 %** | **0,407739** | **-0,001102 п. п.** |
+
+**INT8 не показывает существенной потери точности в этом benchmark:**
+при принятом в challenge округлении до двух знаков FP32 и INT8 дают одинаковые
+**91,25 % MR-ALL**, а разница без округления составляет лишь 0,0011 процентного
+пункта. При этом сохраняются показанные выше преимущества: в 3,73 раза большая
+измеренная ёмкость и в 3,35 раза больший 10M Top-5 throughput. Сравнение
+относится к точности хранения векторов и поиска, а не к INT8-инференсу модели.
+
+![Управление Collections на английском](docs/images/customer/collections-en.jpg)
+
+![RTSP Monitor на английском; частный адрес скрыт](docs/images/customer/monitoring-en.jpg)
+
+## Быстрый старт
+
+Требования:
+
+- Linux x86_64 с Docker Engine и Docker Compose;
+- для CUDA — совместимая NVIDIA GPU, NVIDIA Driver и NVIDIA Container Toolkit.
+
+На хосте не нужны Python, OpenCV, ONNX Runtime, CUDA Toolkit или cuDNN.
+Публичные Images не содержат модели, клиентские данные, API Keys или
+production-конфигурацию.
+
+В полном checkout InsightFace установите модель в `server/.models`:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+Инструмент также поддерживает `buffalo_m`, `buffalo_sc` и `antelopev2`. Он
+создаёт `manifest.json` и подписанную `MODEL.LICENSE`; `models verify` проверяет
+пакет. Условия модели отделены от лицензии исходного кода Server.
+
+Запустить CPU:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Вместо этого запустить CUDA 12:
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+Для CPU откройте `http://СЕРВЕР:18097/`, для CUDA —
+`http://СЕРВЕР:18098/`. Создайте Collection, зарегистрируйте Person по одной
+или нескольким фотографиям и найдите её по другой. `docker compose ... down`
+без `-v` сохраняет volume.
+
+Поставляемые Compose по умолчанию отключают аутентификацию для изолированной
+оценки. Перед доступом других пользователей или сетей:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='замените-на-длинный-случайный-секрет'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Полный первый сценарий описан в [руководстве пользователя](docs/user-guide.ru.md).
+
+## Сборка из исходного кода
+
+Dockerfiles копируют `server/` и выбранные inference-модули из
+`python-package/insightface/`, поэтому контекстом сборки служит весь репозиторий.
+
+CPU:
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12:
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` гарантирует использование локального Image. Сборка всё равно
+загружает зафиксированные базовые Images и зависимости, а установка модели —
+отдельно принятый пакет модели.
+
+## Основные правила
+
+- Similarity — исходный cosine, не вероятность. Threshold находится в
+  `0.0..1.0`, значение по умолчанию `0.4`.
+- Collection фиксирует модель и embedding contract. При несовпадении она видна,
+  но регистрация/поиск возвращает `collection_model_mismatch`.
+- Стартовый Detection Profile копируется в новые Collections; затем профиль
+  можно независимо менять для последующих запросов.
+- Необязательное сохранение лица записывает bounding-box JPEG crop,
+  масштабированный до 112x112, а не оригинал и не выровненный вход распознавания;
+  по умолчанию выключено.
+- Commit SQLite является источником истины. Индекс синхронизируется до успешного
+  ответа регистрации/удаления и строится из SQLite после перезапуска.
+- Ответы содержат `x-request-id`; списки используют непрозрачный подписанный
+  cursor.
+
+Точные поля, defaults, жизненный цикл и ошибки поддерживаются только в
+подробной документации ниже.
+
+## API и SDK
+
+Основные группы:
+
+- система: `/v1/health`, `/v1/system`, `/v1/models`;
+- операции без состояния: `/v1/detect`, `/v1/compare`, `/v1/embeddings`;
+- CRUD Collection, Person и FaceSample;
+- поиск Person в Collection;
+- конфигурация, состояние, события и preview RTSP Monitor.
+
+Все параметры, ответы, ошибки и примеры содержит
+[руководство REST API](docs/api.ru.md). Интерактивный OpenAPI доступен в `/docs`.
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+Установка, входные типы, методы и полные сценарии SDK приведены в
+[руководстве пользователя](docs/user-guide.ru.md).
+
+## Безопасность
+
+Изображения лиц и embeddings — биометрические данные. При сетевом доступе
+включите аутентификацию, завершайте HTTPS на доверенном reverse proxy,
+ограничьте Docker и volumes, не включайте широкий CORS и определите backup,
+retention, удаление, согласие и реакцию на инциденты. Не записывайте в логи
+изображения, embeddings, RTSP credentials и API Keys.
+
+Server не включает TLS, учётные записи, RBAC, cloud IAM или правовой слой.
+Эксплуатация и безопасность описаны в
+[руководстве пользователя](docs/user-guide.ru.md).
+
+## Границы первой фазы
+
+Не реализованы AWS/CompreFace-совместимость, CUDA 11, Jetson, ARM64, Windows
+Container, TensorRT, Kubernetes, распределённые Workers, постоянные события
+Monitor или запись/NVR, liveness, deepfake и демографические атрибуты.
+
+## Документация
+
+- [Руководство пользователя](docs/user-guide.ru.md) — установка, настройки,
+  модели, Web UI, SDK, GPU, безопасность, backup и диагностика.
+- [Руководство REST API](docs/api.ru.md) — все endpoints, поля, поведение,
+  результаты, ошибки, pagination и примеры.
+- [Maintainer Guide — English](docs/maintainer-guide.md) — архитектура,
+  внутренний поиск, тесты, вклад и релизы контейнеров.
+
+GitHub и справка Web UI читают одни и те же локализованные Markdown; отличается
+только отображение.
+
+## Лицензия
+
+Единая точка входа для лицензий — [LICENSING.md](LICENSING.md). Код Server и
+Python SDK распространяются по MIT License; это заявление не охватывает файлы
+и веса моделей, datasets или сторонние компоненты. Публичные модели InsightFace
+обычно ограничены некоммерческими исследованиями без отдельного разрешения.
+Коммерческое лицензирование: <https://www.insightface.ai>.
diff --git a/server/README.zh-CN.md b/server/README.zh-CN.md
new file mode 100644
index 000000000..88c0b388e
--- /dev/null
+++ b/server/README.zh-CN.md
@@ -0,0 +1,254 @@
+# InsightFace Server
+
+**语言：** [English](README.md) · 中文 · [日本語](README.ja.md) · [Deutsch](README.de.md) · [Español](README.es.md) · [Français](README.fr.md) · [Русский](README.ru.md) · [Português](README.pt.md) · [한국어](README.ko.md)
+
+> **单张 GPU，承载 50M+ 人脸向量，提供基于 INT8 特征量化的极速检索，精度无实质损失。**
+
+**一个容器同时提供 Web UI、直观的 REST API、SQLite，以及本地 CPU 或
+NVIDIA GPU 人脸识别推理。**
+
+```text
+上传图片 -> 检测、比对、注册或搜索
+```
+
+> **模型许可：** InsightFace 公开预训练模型通常仅限非商业研究用途。商业使用
+> 需要单独授权，请前往 [InsightFace 官网](https://www.insightface.ai) 获取授权信息。
+
+InsightFace Server 面向常见的人脸识别流程，是比 AWS Rekognition 更简单、
+更注重数据隐私的自托管方案。图片、特征、模型和索引都可以留在自己的网络内。
+它**不是** AWS 兼容替代品，不实现 SigV4、IAM、Region 或 AWS 资源语义。
+
+当前版本：**0.2.0**，Linux x86_64。
+
+| 运行环境 | 镜像 |
+| --- | --- |
+| CPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cpu` |
+| NVIDIA GPU | `ghcr.io/deepinsight/insightface-server:0.2.0-cuda12` |
+
+滚动标签 `cpu` 和 `cuda12` 分别指向对应运行环境的最新稳定版本，不提供含义模糊的
+`latest`。发布规则见[维护者指南（仅英文）](docs/maintainer-guide.md)。
+
+![InsightFace Server 英文仪表盘](docs/images/customer/dashboard-en.jpg)
+
+## 功能概览
+
+- SCRFD 人脸检测、五点关键点、对齐、ArcFace 特征、L2 normalization、原始
+  cosine similarity 和精确 1:N Person 搜索。
+- 多分辨率检测后合并执行一次 NMS，单脸策略支持 `largest` 和
+  `center_largest`。
+- `Collection -> Person -> FaceSample` 数据模型；Collection 绑定模型，多图片
+  注册支持部分成功、metadata 和明确的拒绝原因。
+- 注册 `review_mode` 支持 `off`、`standard`、`strict`，也支持
+  `external_trusted` 外部可信特征。
+- GPU 精确检索支持 FP32、FP16、BF16 和 INT8 向量存储。
+- 多语言 Web UI 覆盖仪表盘、人员库、人员、人脸检测、比对、搜索、RTSP 监控、
+  系统诊断和帮助。
+- `/v1` 下提供 29 个 snake_case REST 接口，包括受保护的
+  `/v1/embeddings`，并附带轻量 Python SDK。
+- 服务端 RTSP Monitor 独立运行、保存有限的内存事件、支持多个客户端，并可选
+  `preview.mjpeg`；关闭浏览器不会停止监控。
+- SQLite 是持久化事实来源；内存精确索引可重建；`/models` 只读、`/data`
+  持久化，并提供 migration、健康检查和禁止静默 CPU 回退的严格 CUDA 启动验证。
+- 支持 JPEG、PNG 和 WebP；默认不保留原始上传图片。
+
+### RTX 5090 GPU 检索性能
+
+在单张 NVIDIA GeForce RTX 5090（32,607 MiB）上，原生 CUDA 精确全量扫描索引
+使用 INT8 时，实测最多可保存 **58.9M 个 512 维图片特征向量**。
+
+| GPU 数据类型 | 最大图片向量数 | 10M Top-5 p50 | 10M 串行 QPS |
+| --- | ---: | ---: | ---: |
+| FP32 | 15.8M | 12.84 ms | 77.85 |
+| FP16 | 30.7M | 6.83 ms | 146.32 |
+| BF16 | 30.7M | 6.83 ms | 146.33 |
+| INT8 | **58.9M** | **3.84 ms** | **260.81** |
+
+与 FP32 相比，INT8 的实测容量为 3.73 倍，10M Top-5 吞吐为 3.35 倍。以上
+仅为同一张 RTX 5090、Driver 580.105.08、CUDA 12.9 上的 GPU 实测。容量是
+未加载 ONNX 模型和 Server 工作负载时的独立原生索引极限；速度测试固定为
+10M 个图片特征向量，执行 GPU 驻留的 Top-5 全量精确扫描，单请求串行，预热
+10 次后测量 100 次。索引在各自存储表示内是精确搜索，但量化仍可能使分数
+相对 FP32 发生变化。生产部署还必须为模型、请求、并发、索引重建和显存
+分配器预留空间。
+
+### ICCV21-MFR 多人种 MR-ALL 精度
+
+我们在 [ICCV21-MFR](../challenges/iccv21-mfr/) 的多人种（MR）测试集上，按照
+MR-ALL 全组队 1:1 协议和 FAR `1e-6` 测试了原生检索 profile。所有 profile
+复用同一批由 Server API 一次性提取并完成 L2 normalization 的 512 维
+`buffalo_l` 特征，仅改变向量存储和检索计算表示。
+
+| 检索 profile | FAR 1e-6 下的 MR-ALL | Cosine 阈值 | 相对 FP32 |
+| --- | ---: | ---: | ---: |
+| FP32 | 91.249107% | 0.407787 | — |
+| FP16 | 91.249197% | 0.407787 | +0.000090 个百分点 |
+| BF16 | 91.248502% | 0.407787 | -0.000605 个百分点 |
+| **INT8** | **91.248005%** | **0.407739** | **-0.001102 个百分点** |
+
+**INT8 在该测评中没有实质精度损失：**按照挑战常用的两位小数展示，
+FP32 和 INT8 的 MR-ALL 均为 **91.25%**，未四舍五入的差异也只有 0.0011
+个百分点，同时保留上文 3.73 倍实测容量和 3.35 倍 10M Top-5 吞吐优势。
+这里对比的是向量存储与检索精度，并非 INT8 模型推理。
+
+![英文 Collection 管理页面](docs/images/customer/collections-en.jpg)
+
+![英文 RTSP Monitor 页面，私有地址已遮挡](docs/images/customer/monitoring-en.jpg)
+
+## 快速开始
+
+环境要求：
+
+- 安装 Docker Engine 和 Docker Compose 的 Linux x86_64；
+- CUDA 版本还需要支持的 NVIDIA GPU、NVIDIA Driver 和 NVIDIA Container
+  Toolkit。
+
+宿主机不需要安装 Python、OpenCV、ONNX Runtime、CUDA Toolkit 或 cuDNN。
+公开镜像不包含模型、客户数据、API Key 或生产配置。
+
+在完整 InsightFace 仓库中，将模型安装到 `server/.models`：
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+```
+
+模型工具还支持 `buffalo_m`、`buffalo_sc` 和 `antelopev2`。安装会生成
+`manifest.json` 和签名的 `MODEL.LICENSE`，可以使用 `models verify` 核验。
+模型许可独立于 Server 源码许可。
+
+启动 CPU：
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+改为启动 CUDA 12：
+
+```bash
+docker compose -f server/deploy/compose.cuda12.yml pull
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml up -d
+curl -fsS http://127.0.0.1:18098/v1/health
+```
+
+CPU 打开 `http://服务器地址:18097/`，CUDA 打开
+`http://服务器地址:18098/`。创建 Collection、为 Person 上传一张或多张注册照，
+再用另一张照片搜索。停止时使用不带 `-v` 的 `docker compose ... down`，即可保留
+数据库卷。
+
+项目提供的 Compose 配置在隔离评估环境中默认关闭认证。对其他用户或网络开放前：
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='请替换为足够长的随机密钥'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+完整的首次使用流程参见[新手用户指南](docs/user-guide.zh-CN.md)。
+
+## 从源码构建
+
+Dockerfile 会复制 `server/` 和 `python-package/insightface/` 中选定的推理模块，
+所以必须使用完整仓库作为构建上下文。
+
+CPU：
+
+```bash
+make -C server build-cpu
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  up -d --no-build --pull never
+```
+
+CUDA 12：
+
+```bash
+make -C server build-cuda12
+docker compose -f server/deploy/compose.cuda12.yml \
+  run --rm --pull never models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cuda12.yml \
+  up -d --no-build --pull never
+```
+
+`--pull never` 确保 Compose 使用本地构建的镜像。构建过程仍会下载锁定的基础镜像
+和依赖；模型安装会单独下载已接受许可的模型包。
+
+## 核心行为
+
+- Similarity 是原始 cosine 值，不是概率；阈值使用 `0.0..1.0`，默认 `0.4`。
+- Collection 固定绑定模型和 embedding contract。模型不匹配时仍可查看，但注册和
+  搜索返回 `collection_model_mismatch`。
+- 启动时的检测 Profile 会复制给新 Collection，之后 Collection Profile 可以独立
+  修改，并从下一次请求生效。
+- 可选人脸保存内容是缩放为 112x112 的 bounding-box JPEG crop，不是原始上传图片，
+  也不是识别模型使用的对齐输入；默认关闭。
+- SQLite 提交是事实来源。注册或删除成功返回前会同步索引；重启后索引从 SQLite
+  重建。
+- 响应包含 `x-request-id`，列表接口使用不透明的签名 cursor。
+
+准确字段、默认值、生命周期和失败行为以末尾延伸文档为准。
+
+## API 和 SDK
+
+主要 API 分组：
+
+- 系统：`/v1/health`、`/v1/system`、`/v1/models`；
+- 无状态人脸接口：`/v1/detect`、`/v1/compare`、`/v1/embeddings`；
+- Collection、Person、FaceSample CRUD；
+- Collection Person 搜索；
+- RTSP Monitor 配置、状态、事件和预览。
+
+所有参数、响应、错误和示例见[完整 REST API 使用指南](docs/api.zh-CN.md)。
+交互式 OpenAPI 仍保留在 `/docs`。
+
+Python：
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:18097", api_key=None) as client:
+    faces = client.detect("photo.jpg")
+    matches = client.search("employees", "unknown.jpg", limit=5)
+```
+
+SDK 安装、图片输入、方法和完整流程见[用户指南](docs/user-guide.zh-CN.md)。
+
+## 安全提示
+
+人脸图片和 embedding 属于生物特征数据。网络部署时应开启认证，通过可信反向代理
+终止 HTTPS，限制 Docker 和数据卷访问，保持宽泛 CORS 关闭，并制定备份、留存、
+删除、同意和安全事件处理策略。日志中不得记录图片、embedding、RTSP 凭据或
+API Key。
+
+Server 不内置 TLS、用户账户、RBAC、云 IAM 或法律合规层。部署和安全操作见
+[用户指南](docs/user-guide.zh-CN.md)。
+
+## 第一阶段范围
+
+当前版本不实现 AWS/CompreFace 兼容、CUDA 11、Jetson、ARM64、Windows
+Container、TensorRT、Kubernetes、分布式 Worker、持久化 Monitor 事件或录像/NVR，
+也不实现活体、Deepfake Detection 和人口属性分析。
+
+## 文档
+
+- [用户指南](docs/user-guide.zh-CN.md)：完整覆盖安装、配置、模型、Web UI、
+  SDK、GPU、安全、备份和故障定位。
+- [REST API 使用指南](docs/api.zh-CN.md)：覆盖每个公开接口、字段、行为、结果、
+  错误、分页规则和示例。
+- [维护者指南（仅英文）](docs/maintainer-guide.md)：架构、检索内部实现、测试、
+  贡献规则和容器发布。
+
+GitHub 与 Web UI 帮助页读取完全相同的本地化 User Guide 和 API Guide Markdown，
+区别只在渲染方式。
+
+## 许可证
+
+统一许可入口见 [LICENSING.md](LICENSING.md)。简而言之，Server 源码和 Python
+SDK 采用 MIT License；该声明不覆盖模型文件、模型权重、数据集或第三方组件。
+InsightFace 公开预训练模型通常仅限非商业研究用途，除非另行获得授权。商业
+授权信息：<https://www.insightface.ai>。
diff --git a/server/backend/insightface_server/__init__.py b/server/backend/insightface_server/__init__.py
new file mode 100644
index 000000000..cb2235b96
--- /dev/null
+++ b/server/backend/insightface_server/__init__.py
@@ -0,0 +1,3 @@
+"""InsightFace Server."""
+
+__version__ = "0.2.0"
diff --git a/server/backend/insightface_server/api/__init__.py b/server/backend/insightface_server/api/__init__.py
new file mode 100644
index 000000000..3c92c360c
--- /dev/null
+++ b/server/backend/insightface_server/api/__init__.py
@@ -0,0 +1 @@
+"""HTTP API boundary."""
diff --git a/server/backend/insightface_server/api/auth.py b/server/backend/insightface_server/api/auth.py
new file mode 100644
index 000000000..c0f56fd3b
--- /dev/null
+++ b/server/backend/insightface_server/api/auth.py
@@ -0,0 +1,40 @@
+from __future__ import annotations
+
+import logging
+
+from fastapi import Request
+
+from ..config import Settings
+from ..errors import ApiError
+from ..storage import Repository
+
+LOGGER = logging.getLogger("insightface_server")
+
+
+class ApiKeyAuthenticator:
+    def __init__(self, settings: Settings, repository: Repository):
+        self.settings = settings
+        self.repository = repository
+
+    def initialize(self) -> None:
+        if not self.settings.auth_enabled:
+            return
+        if self.settings.startup_api_key:
+            result = self.repository.sync_api_key(self.settings.startup_api_key)
+            if result != "unchanged":
+                LOGGER.info("api_key_%s", result)
+        if not self.repository.has_api_keys():
+            raise RuntimeError(
+                "Authentication is enabled but no API key exists; set INSIGHTFACE_API_KEY "
+                "when starting the server"
+            )
+
+    def require(self, request: Request) -> None:
+        if not self.settings.auth_enabled:
+            return
+        authorization = request.headers.get("authorization", "")
+        scheme, separator, key = authorization.partition(" ")
+        if separator != " " or scheme.lower() != "bearer" or not key:
+            raise ApiError("unauthorized", "A valid Bearer API key is required.", 401)
+        if not self.repository.verify_api_key(key):
+            raise ApiError("unauthorized", "A valid Bearer API key is required.", 401)
diff --git a/server/backend/insightface_server/api/responses.py b/server/backend/insightface_server/api/responses.py
new file mode 100644
index 000000000..ef7fab7e8
--- /dev/null
+++ b/server/backend/insightface_server/api/responses.py
@@ -0,0 +1,373 @@
+"""Public response schemas for the versioned REST API.
+
+The route implementations intentionally return ``JSONResponse`` objects so the
+wire format remains explicit. These models document and test that format
+without asking FastAPI to filter successful response payloads.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Literal
+from uuid import UUID
+
+from pydantic import BaseModel, ConfigDict
+
+
+class PublicResponseModel(BaseModel):
+    model_config = ConfigDict(extra="allow")
+
+
+class ResponseEnvelope(PublicResponseModel):
+    request_id: UUID
+
+
+class ErrorBody(PublicResponseModel):
+    code: str
+    message: str
+    details: dict[str, Any]
+
+
+class ErrorEnvelope(ResponseEnvelope):
+    error: ErrorBody
+
+
+class PixelBox(PublicResponseModel):
+    x: int
+    y: int
+    width: int
+    height: int
+
+
+class NormalizedBox(PublicResponseModel):
+    left: float
+    top: float
+    width: float
+    height: float
+
+
+class BoundingBox(PublicResponseModel):
+    pixels: PixelBox
+    normalized: NormalizedBox
+
+
+class FaceQuality(PublicResponseModel):
+    score: float
+    sharpness: float
+    brightness: float
+    pose: float
+
+
+class FaceObservation(PublicResponseModel):
+    bbox: BoundingBox
+    landmarks: list[list[float]]
+    detection_score: float
+    quality: FaceQuality
+    embedding: list[float] | None = None
+
+
+class ModelComponent(PublicResponseModel):
+    model_id: str
+    model_version: str
+    task: str
+    sha256: str
+    file: str | None = None
+    input_size: list[int] | None = None
+    embedding_dimension: int | None = None
+    preprocessing_version: str | None = None
+
+
+class ModelSummary(PublicResponseModel):
+    model_id: str
+    model_version: str
+    model_digest: str
+    embedding_dimension: int
+    preprocessing_version: str
+    provider: str
+    models: list[ModelComponent]
+    license: dict[str, Any] | None
+
+
+class DetectionProfile(PublicResponseModel):
+    input_sizes: list[list[int]]
+    threshold: float
+    nms_threshold: float
+    single_face_selection: Literal["largest", "center_largest"]
+
+
+class Collection(PublicResponseModel):
+    id: str
+    name: str
+    description: str
+    default_threshold: float
+    model_id: str
+    model_version: str
+    model_digest: str
+    embedding_dimension: int
+    preprocessing_version: str
+    embedding_contract_id: str
+    search_profile: str
+    capacity_rows: int
+    max_faces_per_person: int
+    load_policy: Literal["eager", "lazy"]
+    search_revision: int
+    save_face_crops: bool
+    metadata: dict[str, Any]
+    person_count: int
+    face_count: int
+    created_at: str
+    updated_at: str
+    detection: DetectionProfile
+    detection_revision: int
+
+
+class Person(PublicResponseModel):
+    id: str
+    name: str | None
+    external_id: str | None
+    metadata: dict[str, Any]
+    face_count: int
+    created_at: str
+    updated_at: str
+
+
+class FaceSample(PublicResponseModel):
+    id: str
+    person_id: str | None = None
+    bounding_box: BoundingBox
+    landmarks: list[list[float]] | None
+    detection_score: float
+    quality: FaceQuality
+    model_id: str
+    model_version: str
+    model_digest: str
+    preprocessing_version: str
+    embedding_source: Literal["server", "external_trusted"]
+    embedding_contract_id: str | None
+    has_crop: bool
+    created_at: str
+
+
+class RejectedImage(PublicResponseModel):
+    index: int
+    filename: str
+    reason: str
+
+
+class Match(PublicResponseModel):
+    person: Person
+    similarity: float
+    matched_face_id: str
+
+
+class MonitorSource(PublicResponseModel):
+    type: Literal["rtsp"]
+    url: str
+
+
+class MonitorEventPolicy(PublicResponseModel):
+    confirm_frames: int
+    absence_timeout_seconds: float
+    cooldown_seconds: float
+    emit_unknown: bool
+
+
+class MonitorRuntime(PublicResponseModel):
+    status: str
+    connected: bool
+    stream_epoch: str | None
+    last_frame_at: str | None
+    last_error: dict[str, Any] | str | None
+    preview_active: bool
+    preview_viewers: int
+
+
+class Monitor(PublicResponseModel):
+    id: str
+    name: str
+    description: str
+    enabled: bool
+    source: MonitorSource
+    collection_id: str
+    inference_fps: float
+    match_threshold: float | None
+    event_buffer_size: int
+    event_policy: MonitorEventPolicy
+    preview_enabled: bool
+    runtime: MonitorRuntime
+    created_at: str
+    updated_at: str
+
+
+class MonitorStateSource(PublicResponseModel):
+    width: int | None
+    height: int | None
+    source_fps: float | None
+
+
+class MonitorInferenceState(PublicResponseModel):
+    configured_fps: float
+    actual_fps: float
+    processing_ms: float | None
+    last_frame_at: str | None
+    last_inference_at: str | None
+    frame_sequence: int
+    decoded_frames: int
+    processed_frames: int
+    dropped_frames: int
+    capacity_limited: bool
+
+
+class MonitorPreviewState(PublicResponseModel):
+    enabled: bool
+    active: bool
+    viewers: int
+    frames: int
+
+
+class MonitorState(PublicResponseModel):
+    monitor_id: str
+    stream_epoch: str | None
+    status: str
+    connected: bool
+    source: MonitorStateSource
+    inference: MonitorInferenceState
+    threshold: float | None
+    faces: list[dict[str, Any]]
+    matched_faces: int
+    unknown_faces: int
+    preview: MonitorPreviewState
+    reconnects: int
+    inference_errors: int
+    last_error: dict[str, Any] | str | None
+    started_at: str | None
+
+
+class MonitorEvent(PublicResponseModel):
+    id: UUID
+    sequence: int
+    type: str
+    monitor_id: str
+    stream_epoch: str
+    occurred_at: str
+    track_id: str | None = None
+    person: dict[str, Any] | None = None
+    similarity: float | None = None
+    threshold: float | None = None
+    matched_face_id: str | None = None
+    face: dict[str, Any] | None = None
+    error: dict[str, Any] | None = None
+
+
+class HealthResponse(ResponseEnvelope):
+    status: Literal["ready", "not_ready"]
+    version: str
+    auth_enabled: bool
+
+
+class SystemResponse(ResponseEnvelope):
+    server_version: str
+    os: str
+    architecture: str
+    cpu: dict[str, Any]
+    runtime: dict[str, Any]
+    execution_provider: str
+    model: ModelSummary
+    database: dict[str, Any]
+    data: dict[str, Any]
+    models: dict[str, Any]
+    stats: dict[str, int]
+    search: dict[str, Any]
+    api_key: dict[str, bool]
+    safe_config: dict[str, Any]
+    recent_errors: list[dict[str, Any]]
+
+
+class ModelsResponse(ResponseEnvelope):
+    models: list[ModelComponent]
+    execution_provider: str
+    license: dict[str, Any] | None
+
+
+class DetectResponse(ResponseEnvelope):
+    faces: list[FaceObservation]
+    processing_ms: float
+
+
+class CompareResponse(ResponseEnvelope):
+    matched: bool
+    similarity: float
+    threshold: float
+    source_face: FaceObservation
+    target_face: FaceObservation
+    processing_ms: float
+
+
+class EmbeddingsResponse(ResponseEnvelope):
+    faces: list[FaceObservation]
+    model: ModelSummary
+    processing_ms: float
+
+
+class CollectionResponse(ResponseEnvelope):
+    collection: Collection
+
+
+class CollectionPageResponse(ResponseEnvelope):
+    collections: list[Collection]
+    next_cursor: str | None
+
+
+class PersonResponse(ResponseEnvelope):
+    person: Person
+
+
+class PersonRegistrationResponse(PersonResponse):
+    faces: list[FaceSample]
+    rejected_images: list[RejectedImage]
+
+
+class PersonPageResponse(ResponseEnvelope):
+    persons: list[Person]
+    next_cursor: str | None
+
+
+class FaceRegistrationResponse(ResponseEnvelope):
+    faces: list[FaceSample]
+    rejected_images: list[RejectedImage]
+
+
+class FacePageResponse(ResponseEnvelope):
+    faces: list[FaceSample]
+    next_cursor: str | None
+
+
+class SearchResponse(ResponseEnvelope):
+    searched_face: FaceObservation
+    matches: list[Match]
+    threshold: float
+    processing_ms: float
+
+
+class MonitorResponse(ResponseEnvelope):
+    monitor: Monitor
+
+
+class MonitorPageResponse(ResponseEnvelope):
+    monitors: list[Monitor]
+    next_cursor: str | None
+
+
+class MonitorStateResponse(ResponseEnvelope):
+    state: MonitorState
+
+
+class MonitorEventPageResponse(ResponseEnvelope):
+    monitor_id: str
+    stream_epoch: str | None
+    events: list[MonitorEvent]
+    oldest_sequence: int | None
+    latest_sequence: int
+    has_more: bool
+    truncated: bool
+    stream_reset: bool
+    next_cursor: str
diff --git a/server/backend/insightface_server/api/schemas.py b/server/backend/insightface_server/api/schemas.py
new file mode 100644
index 000000000..eec899f3f
--- /dev/null
+++ b/server/backend/insightface_server/api/schemas.py
@@ -0,0 +1,231 @@
+from __future__ import annotations
+
+import re
+from typing import Any, Literal
+from urllib.parse import urlsplit
+
+from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
+
+from ..config import normalize_detector_input_sizes
+
+ID_PATTERN = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$")
+
+
+def validate_id(value: str) -> str:
+    if ID_PATTERN.fullmatch(value) is None:
+        raise ValueError(
+            "must be 1-64 characters starting with a letter or digit and containing "
+            "only letters, digits, dot, underscore, or hyphen"
+        )
+    return value
+
+
+def validate_collection_id(value: str) -> str:
+    if value == "_default":
+        return value
+    return validate_id(value)
+
+
+class StrictModel(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+
+SearchProfile = Literal["fp32_v1", "fp16_v1", "bf16_v1", "int8_x736_v1", "int8_x1000_v1"]
+SearchLoadPolicy = Literal["eager", "lazy"]
+ReviewMode = Literal["off", "standard", "strict"]
+EmbeddingMode = Literal["server", "external_trusted"]
+SingleFaceSelection = Literal["largest", "center_largest"]
+
+
+class CollectionDetectionCreate(StrictModel):
+    input_sizes: list[tuple[int, int]] | None = None
+    threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    nms_threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    single_face_selection: SingleFaceSelection | None = None
+
+    @field_validator("input_sizes")
+    @classmethod
+    def validate_input_sizes(
+        cls, value: list[tuple[int, int]] | None
+    ) -> list[tuple[int, int]] | None:
+        if value is None:
+            return None
+        return list(normalize_detector_input_sizes(value))
+
+
+class CollectionDetectionPatch(StrictModel):
+    input_sizes: list[tuple[int, int]] | None = None
+    threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    nms_threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    single_face_selection: SingleFaceSelection | None = None
+
+    @field_validator("input_sizes")
+    @classmethod
+    def validate_input_sizes(
+        cls, value: list[tuple[int, int]] | None
+    ) -> list[tuple[int, int]] | None:
+        if value is None:
+            return None
+        return list(normalize_detector_input_sizes(value))
+
+    @model_validator(mode="after")
+    def reject_nulls(self) -> CollectionDetectionPatch:
+        for name in self.model_fields_set:
+            if getattr(self, name) is None:
+                raise ValueError(f"{name} cannot be null")
+        return self
+
+
+class CollectionSearchCreate(StrictModel):
+    profile: SearchProfile | None = None
+    capacity_rows: int | None = Field(default=None, ge=1, le=2**63 - 1)
+    max_faces_per_person: int | None = Field(default=None, ge=1, le=1_000_000)
+    load_policy: SearchLoadPolicy | None = None
+
+
+class CollectionSearchPatch(StrictModel):
+    capacity_rows: int | None = Field(default=None, ge=1, le=2**63 - 1)
+    max_faces_per_person: int | None = Field(default=None, ge=1, le=1_000_000)
+    load_policy: SearchLoadPolicy | None = None
+
+    @model_validator(mode="after")
+    def reject_nulls(self) -> CollectionSearchPatch:
+        for name in self.model_fields_set:
+            if getattr(self, name) is None:
+                raise ValueError(f"{name} cannot be null")
+        return self
+
+
+class CollectionCreate(StrictModel):
+    id: str
+    name: str = Field(min_length=1, max_length=200)
+    description: str = Field(default="", max_length=2000)
+    threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    save_face_crops: bool | None = None
+    metadata: dict[str, Any] = Field(default_factory=dict)
+    detection: CollectionDetectionCreate = Field(
+        default_factory=CollectionDetectionCreate
+    )
+    search: CollectionSearchCreate = Field(default_factory=CollectionSearchCreate)
+
+    _id = field_validator("id")(validate_collection_id)
+
+
+class CollectionPatch(StrictModel):
+    name: str | None = Field(default=None, min_length=1, max_length=200)
+    description: str | None = Field(default=None, max_length=2000)
+    threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    save_face_crops: bool | None = None
+    metadata: dict[str, Any] | None = None
+    detection: CollectionDetectionPatch | None = None
+    search: CollectionSearchPatch | None = None
+
+    @model_validator(mode="after")
+    def reject_nulls(self) -> CollectionPatch:
+        for name in self.model_fields_set:
+            if getattr(self, name) is None:
+                raise ValueError(f"{name} cannot be null")
+        return self
+
+
+class PersonPatch(StrictModel):
+    name: str | None = Field(default=None, max_length=200)
+    external_id: str | None = Field(default=None, max_length=200)
+    metadata: dict[str, Any] | None = None
+
+    @model_validator(mode="after")
+    def reject_null_metadata(self) -> PersonPatch:
+        if "metadata" in self.model_fields_set and self.metadata is None:
+            raise ValueError("metadata cannot be null")
+        return self
+
+
+class MonitorSource(StrictModel):
+    type: Literal["rtsp"] = "rtsp"
+    url: str = Field(min_length=8, max_length=2048)
+
+    @field_validator("url")
+    @classmethod
+    def validate_rtsp_url(cls, value: str) -> str:
+        normalized = value.strip()
+        parsed = urlsplit(normalized)
+        if parsed.scheme.lower() not in {"rtsp", "rtsps"}:
+            raise ValueError("url must use the rtsp:// or rtsps:// scheme")
+        if not parsed.hostname:
+            raise ValueError("url must include a host")
+        if parsed.fragment:
+            raise ValueError("url must not include a fragment")
+        try:
+            _ = parsed.port
+        except ValueError as exc:
+            raise ValueError("url contains an invalid port") from exc
+        return normalized
+
+
+class MonitorEventPolicy(StrictModel):
+    confirm_frames: int = Field(default=3, ge=1, le=100)
+    absence_timeout_seconds: float = Field(default=3.0, ge=0.1, le=300.0)
+    cooldown_seconds: float = Field(default=10.0, ge=0.0, le=3600.0)
+    emit_unknown: bool = True
+
+
+class MonitorEventPolicyPatch(StrictModel):
+    confirm_frames: int | None = Field(default=None, ge=1, le=100)
+    absence_timeout_seconds: float | None = Field(default=None, ge=0.1, le=300.0)
+    cooldown_seconds: float | None = Field(default=None, ge=0.0, le=3600.0)
+    emit_unknown: bool | None = None
+
+    @model_validator(mode="after")
+    def validate_patch(self) -> MonitorEventPolicyPatch:
+        if not self.model_fields_set:
+            raise ValueError("at least one event policy field must be supplied")
+        for name in self.model_fields_set:
+            if getattr(self, name) is None:
+                raise ValueError(f"{name} cannot be null")
+        return self
+
+
+class MonitorCreate(StrictModel):
+    id: str
+    name: str = Field(min_length=1, max_length=200)
+    description: str = Field(default="", max_length=2000)
+    enabled: bool = True
+    source: MonitorSource
+    collection_id: str
+    inference_fps: float = Field(default=2.0, ge=0.1, le=30.0)
+    match_threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    event_buffer_size: int = Field(default=1000, ge=10, le=10000)
+    event_policy: MonitorEventPolicy = Field(default_factory=MonitorEventPolicy)
+    preview_enabled: bool = False
+
+    _id = field_validator("id")(validate_id)
+    _collection_id = field_validator("collection_id")(validate_collection_id)
+
+
+class MonitorPatch(StrictModel):
+    name: str | None = Field(default=None, min_length=1, max_length=200)
+    description: str | None = Field(default=None, max_length=2000)
+    enabled: bool | None = None
+    source: MonitorSource | None = None
+    collection_id: str | None = None
+    inference_fps: float | None = Field(default=None, ge=0.1, le=30.0)
+    match_threshold: float | None = Field(default=None, ge=0.0, le=1.0)
+    event_buffer_size: int | None = Field(default=None, ge=10, le=10000)
+    event_policy: MonitorEventPolicyPatch | None = None
+    preview_enabled: bool | None = None
+
+    @field_validator("collection_id")
+    @classmethod
+    def validate_optional_collection_id(cls, value: str | None) -> str | None:
+        return validate_collection_id(value) if value is not None else None
+
+    @model_validator(mode="after")
+    def validate_patch(self) -> MonitorPatch:
+        if not self.model_fields_set:
+            raise ValueError("at least one field must be supplied")
+        for name in self.model_fields_set:
+            if name == "match_threshold":
+                continue
+            if getattr(self, name) is None:
+                raise ValueError(f"{name} cannot be null")
+        return self
diff --git a/server/backend/insightface_server/app.py b/server/backend/insightface_server/app.py
new file mode 100644
index 000000000..eb1bc3d58
--- /dev/null
+++ b/server/backend/insightface_server/app.py
@@ -0,0 +1,1501 @@
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+import platform
+import sqlite3
+import time
+import uuid
+from collections import deque
+from contextlib import asynccontextmanager
+from functools import partial
+from pathlib import Path
+from typing import Any
+
+from fastapi import FastAPI, File, Form, Query, Request, UploadFile
+from fastapi.exceptions import RequestValidationError
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.openapi.utils import get_openapi
+from fastapi.responses import FileResponse, JSONResponse, Response, StreamingResponse
+from fastapi.staticfiles import StaticFiles
+from starlette.exceptions import HTTPException as StarletteHTTPException
+
+from . import __version__
+from .api.auth import ApiKeyAuthenticator
+from .api.responses import (
+    CollectionPageResponse,
+    CollectionResponse,
+    CompareResponse,
+    DetectResponse,
+    EmbeddingsResponse,
+    ErrorEnvelope,
+    FacePageResponse,
+    FaceRegistrationResponse,
+    HealthResponse,
+    ModelsResponse,
+    MonitorEventPageResponse,
+    MonitorPageResponse,
+    MonitorResponse,
+    MonitorStateResponse,
+    PersonPageResponse,
+    PersonRegistrationResponse,
+    PersonResponse,
+    SearchResponse,
+    SystemResponse,
+)
+from .api.schemas import (
+    CollectionCreate,
+    CollectionPatch,
+    EmbeddingMode,
+    MonitorCreate,
+    MonitorPatch,
+    PersonPatch,
+    ReviewMode,
+    validate_id,
+)
+from .config import DetectionProfile, Settings
+from .errors import ApiError, bad_request, conflict, not_found, unprocessable
+from .inference import create_engine
+from .request_context import REQUEST_DEADLINE, remaining_seconds
+from .search import SearchIndexError, SearchIndexManager, create_search_backend
+from .services import (
+    FaceService,
+    ImageData,
+    ImageLoader,
+    MonitorAlreadyExistsError,
+    MonitorLimitError,
+    MonitorManager,
+    MonitorNotFoundError,
+    MonitorPreviewDisabledError,
+    MonitorUnavailableError,
+)
+from .services.presentation import face_result
+from .storage import (
+    CursorCodec,
+    Database,
+    FaceCropStore,
+    Repository,
+    SecretCodec,
+)
+
+LOGGER = logging.getLogger("insightface_server")
+PACKAGE_DIR = Path(__file__).resolve().parent
+SERVER_DIR = PACKAGE_DIR.parents[1]
+MAINTAINER_GUIDE = SERVER_DIR / "docs" / "maintainer-guide.md"
+DOCUMENTS: dict[str, dict[str, Path]] = {
+    "en": {
+        "api": SERVER_DIR / "docs" / "api.md",
+        "user-guide": SERVER_DIR / "docs" / "user-guide.md",
+        "maintainer": MAINTAINER_GUIDE,
+    },
+    "zh": {
+        "api": SERVER_DIR / "docs" / "api.zh-CN.md",
+        "user-guide": SERVER_DIR / "docs" / "user-guide.zh-CN.md",
+        "maintainer": MAINTAINER_GUIDE,
+    },
+    **{
+        language: {
+            "api": SERVER_DIR / "docs" / f"api.{language}.md",
+            "user-guide": SERVER_DIR / "docs" / f"user-guide.{language}.md",
+            "maintainer": MAINTAINER_GUIDE,
+        }
+        for language in ("ja", "de", "es", "fr", "ru", "pt", "ko")
+    },
+}
+
+
+def _request_id(request: Request) -> str:
+    return str(getattr(request.state, "request_id", "unknown"))
+
+
+def _json(request: Request, content: dict[str, Any], *, status_code: int = 200) -> JSONResponse:
+    return JSONResponse(
+        status_code=status_code,
+        content={**content, "request_id": _request_id(request)},
+    )
+
+
+def _metadata(raw: str) -> dict[str, Any]:
+    try:
+        value = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        raise bad_request("metadata must be valid JSON.", code="invalid_metadata") from exc
+    if not isinstance(value, dict):
+        raise bad_request("metadata must be a JSON object.", code="invalid_metadata")
+    return value
+
+
+def _enrollment_embeddings(
+    *,
+    embedding_mode: EmbeddingMode,
+    external_embeddings: str | None,
+    embedding_contract_id: str | None,
+    image_count: int,
+) -> list[Any] | None:
+    """Validate multipart-level trusted embedding fields without logging values."""
+
+    if embedding_mode == "server":
+        if external_embeddings is not None or embedding_contract_id is not None:
+            raise bad_request(
+                "External embedding fields require embedding_mode=external_trusted.",
+                code="unexpected_external_embedding",
+            )
+        return None
+    if external_embeddings is None:
+        raise bad_request(
+            "external_embeddings is required for external_trusted enrollment.",
+            code="missing_external_embeddings",
+        )
+    if embedding_contract_id is None or not embedding_contract_id.strip():
+        raise bad_request(
+            "embedding_contract_id is required for external_trusted enrollment.",
+            code="missing_embedding_contract_id",
+        )
+    try:
+        values = json.loads(external_embeddings)
+    except (json.JSONDecodeError, TypeError) as exc:
+        raise bad_request(
+            "external_embeddings must be a JSON array of vectors.",
+            code="invalid_external_embeddings",
+        ) from exc
+    if not isinstance(values, list):
+        raise bad_request(
+            "external_embeddings must be a JSON array of vectors.",
+            code="invalid_external_embeddings",
+        )
+    if len(values) != image_count:
+        raise bad_request(
+            "external_embeddings must contain exactly one vector for each image.",
+            code="external_embedding_count_mismatch",
+            image_count=image_count,
+            embedding_count=len(values),
+        )
+    return values
+
+
+def _path_status(path: Path) -> dict[str, Any]:
+    return {
+        "path": str(path),
+        "exists": path.exists(),
+        "readable": os.access(path, os.R_OK),
+        "writable": os.access(path, os.W_OK),
+    }
+
+
+def _cpu_name() -> str:
+    value = platform.processor()
+    if value:
+        return value
+    try:
+        for line in Path("/proc/cpuinfo").read_text(encoding="utf-8").splitlines():
+            if line.lower().startswith("model name"):
+                return line.split(":", 1)[1].strip()
+    except OSError:
+        pass
+    return platform.machine()
+
+
+async def _blocking(function, *args, **kwargs):
+    """Run blocking model/storage work without delaying timeout responses."""
+
+    work = asyncio.to_thread(partial(function, *args, **kwargs))
+    remaining = remaining_seconds()
+    if remaining is None:
+        return await work
+    try:
+        return await asyncio.wait_for(work, timeout=max(0.001, remaining - 0.05))
+    except TimeoutError as exc:
+        raise ApiError("request_timeout", "The request timed out.", 503) from exc
+
+
+def create_app(settings: Settings | None = None) -> FastAPI:
+    configured = settings or Settings.from_env()
+    recent_errors: deque[dict[str, Any]] = deque(maxlen=20)
+
+    @asynccontextmanager
+    async def lifespan(app: FastAPI):
+        logging.basicConfig(
+            level=getattr(logging, configured.log_level, logging.INFO),
+            format="%(asctime)s %(levelname)s %(name)s %(message)s",
+        )
+        database = Database(
+            configured.data_dir / "insightface-server.db", SERVER_DIR / "migrations"
+        )
+        search_indexes: SearchIndexManager | None = None
+        monitors: MonitorManager | None = None
+        engine = None
+        app.state.ready = False
+        database.acquire_process_lock()
+        try:
+            database.initialize()
+            repository = Repository(database)
+            authenticator = ApiKeyAuthenticator(configured, repository)
+            authenticator.initialize()
+            engine = create_engine(configured)
+            engine.startup()
+            search_backend = create_search_backend(configured)
+            search_indexes = SearchIndexManager(
+                repository,
+                search_backend,
+                build_batch_rows=configured.search_build_batch_rows,
+            )
+            search_indexes.startup()
+            image_loader = ImageLoader(configured)
+            cursors = CursorCodec(configured.data_dir / "cursor.key")
+            service = FaceService(
+                configured,
+                repository,
+                cursors,
+                FaceCropStore(configured.data_dir, enabled=configured.save_face_crops),
+                engine,
+                search_indexes,
+            )
+            monitors = MonitorManager(
+                service,
+                repository,
+                cursors,
+                SecretCodec(configured.data_dir / "monitor-credentials.key"),
+                max_monitors=configured.rtsp_max_streams,
+                max_faces=configured.max_detected_faces,
+                preview_fps=configured.rtsp_preview_fps,
+                jpeg_quality=configured.rtsp_jpeg_quality,
+                open_timeout_seconds=configured.rtsp_open_timeout_seconds,
+                read_timeout_seconds=configured.rtsp_read_timeout_seconds,
+                reconnect_delay_seconds=configured.rtsp_reconnect_delay_seconds,
+            )
+            monitors.startup()
+            app.state.settings = configured
+            app.state.database = database
+            app.state.repository = repository
+            app.state.authenticator = authenticator
+            app.state.engine = engine
+            app.state.search_indexes = search_indexes
+            app.state.image_loader = image_loader
+            app.state.service = service
+            app.state.monitors = monitors
+            app.state.ready = True
+            LOGGER.info(
+                "server_ready version=%s provider=%s inference_max_concurrency=%s "
+                "model=%s runtime=%s search=%s",
+                __version__,
+                engine.summary.provider,
+                configured.effective_inference_max_concurrency,
+                json.dumps(engine.summary.as_dict(), sort_keys=True),
+                json.dumps(engine.runtime_summary(), sort_keys=True),
+                json.dumps(search_indexes.runtime_summary(), sort_keys=True),
+            )
+            LOGGER.info(
+                "web_ui_%s",
+                "disabled" if configured.web_ui_disabled else "enabled",
+            )
+            yield
+        finally:
+            app.state.ready = False
+            if monitors is not None:
+                monitors.close()
+            if search_indexes is not None:
+                search_indexes.close()
+            if engine is not None:
+                engine.close()
+            database.release_process_lock()
+
+    app = FastAPI(
+        title="InsightFace Server",
+        version=__version__,
+        description="Simple self-hosted face detection, comparison, registration, and search.",
+        lifespan=lifespan,
+        docs_url=None,
+        redoc_url=None,
+    )
+    if configured.cors_origins:
+        app.add_middleware(
+            CORSMiddleware,
+            allow_origins=list(configured.cors_origins),
+            allow_credentials=False,
+            allow_methods=["GET", "POST", "PATCH", "DELETE"],
+            allow_headers=["Authorization", "Content-Type"],
+        )
+
+    @app.middleware("http")
+    async def request_context(request: Request, call_next):
+        request_id = str(uuid.uuid4())
+        request.state.request_id = request_id
+        started = time.monotonic()
+        deadline_token = REQUEST_DEADLINE.set(started + configured.request_timeout_seconds)
+        try:
+            content_length = int(request.headers.get("content-length", "0"))
+        except ValueError:
+            content_length = 0
+        if content_length > configured.max_request_bytes:
+            response: Response = JSONResponse(
+                status_code=413,
+                content=ApiError("request_too_large", "The request body is too large.", 413).body(
+                    request_id
+                ),
+            )
+        else:
+            try:
+                chunks: list[bytes] = []
+                received = 0
+                async for chunk in request.stream():
+                    received += len(chunk)
+                    if received > configured.max_request_bytes:
+                        raise ApiError("request_too_large", "The request body is too large.", 413)
+                    chunks.append(chunk)
+                # Function middleware receives Starlette's cached Request. Setting
+                # its bounded body lets the downstream multipart parser replay it.
+                request._body = b"".join(chunks)  # noqa: SLF001
+                remaining = configured.request_timeout_seconds - (time.monotonic() - started)
+                if remaining <= 0:
+                    raise TimeoutError
+                response = await asyncio.wait_for(call_next(request), timeout=remaining)
+            except ApiError as request_error:
+                recent_errors.append(
+                    {
+                        "code": request_error.code,
+                        "path": request.url.path,
+                        "request_id": request_id,
+                    }
+                )
+                response = JSONResponse(
+                    status_code=request_error.status_code,
+                    content=request_error.body(request_id),
+                )
+            except TimeoutError:
+                timeout_error = ApiError("request_timeout", "The request timed out.", 503)
+                recent_errors.append(
+                    {
+                        "code": timeout_error.code,
+                        "path": request.url.path,
+                        "request_id": request_id,
+                    }
+                )
+                response = JSONResponse(status_code=503, content=timeout_error.body(request_id))
+        response.headers["x-request-id"] = request_id
+        response.headers["X-Content-Type-Options"] = "nosniff"
+        response.headers["X-Frame-Options"] = "DENY"
+        response.headers["Referrer-Policy"] = "no-referrer"
+        response.headers["Permissions-Policy"] = "camera=(), microphone=(), geolocation=()"
+        response.headers["Content-Security-Policy"] = (
+            "default-src 'self'; img-src 'self' data: blob:; media-src 'self' blob:; "
+            "style-src 'self'; script-src 'self'; connect-src 'self'; object-src 'none'; "
+            "base-uri 'none'; frame-ancestors 'none'; form-action 'self'"
+        )
+        REQUEST_DEADLINE.reset(deadline_token)
+        return response
+
+    @app.exception_handler(ApiError)
+    async def api_error_handler(request: Request, exc: ApiError):
+        recent_errors.append(
+            {"code": exc.code, "path": request.url.path, "request_id": _request_id(request)}
+        )
+        response = JSONResponse(status_code=exc.status_code, content=exc.body(_request_id(request)))
+        if exc.status_code == 401:
+            response.headers["WWW-Authenticate"] = "Bearer"
+        return response
+
+    @app.exception_handler(RequestValidationError)
+    async def validation_error_handler(request: Request, exc: RequestValidationError):
+        first = exc.errors()[0] if exc.errors() else {"msg": "invalid request"}
+        error = bad_request(f"Request validation failed: {first.get('msg', 'invalid request')}")
+        return JSONResponse(status_code=400, content=error.body(_request_id(request)))
+
+    @app.exception_handler(StarletteHTTPException)
+    async def http_error_handler(request: Request, exc: StarletteHTTPException):
+        code = {
+            404: "route_not_found",
+            405: "method_not_allowed",
+            413: "request_too_large",
+        }.get(exc.status_code, "http_error")
+        error = ApiError(code, str(exc.detail), exc.status_code)
+        return JSONResponse(status_code=exc.status_code, content=error.body(_request_id(request)))
+
+    @app.exception_handler(Exception)
+    async def unexpected_error_handler(request: Request, exc: Exception):
+        LOGGER.exception(
+            "unhandled_request_error request_id=%s path=%s type=%s",
+            _request_id(request),
+            request.url.path,
+            type(exc).__name__,
+        )
+        recent_errors.append(
+            {"code": "internal_error", "path": request.url.path, "request_id": _request_id(request)}
+        )
+        error = ApiError("internal_error", "The server could not complete the request.", 500)
+        return JSONResponse(status_code=500, content=error.body(_request_id(request)))
+
+    async def authorize(request: Request) -> None:
+        await _blocking(app.state.authenticator.require, request)
+
+    async def load_images(
+        uploads: list[UploadFile],
+    ) -> tuple[list[ImageData], list[int], list[dict[str, Any]]]:
+        loaded: list[ImageData] = []
+        original_indices: list[int] = []
+        rejected: list[dict[str, Any]] = []
+        for index, upload in enumerate(uploads):
+            try:
+                loaded.append(await app.state.image_loader.from_upload(upload))
+                original_indices.append(index)
+            except ApiError as exc:
+                rejected.append(
+                    {
+                        "index": index,
+                        "filename": upload.filename or "upload",
+                        "reason": exc.code,
+                    }
+                )
+        return loaded, original_indices, rejected
+
+    def remap_rejections(
+        rejected: list[dict[str, Any]], original_indices: list[int]
+    ) -> list[dict[str, Any]]:
+        result: list[dict[str, Any]] = []
+        for item in rejected:
+            value = dict(item)
+            relative = int(value["index"])
+            if 0 <= relative < len(original_indices):
+                value["index"] = original_indices[relative]
+            result.append(value)
+        return result
+
+    @app.get(
+        "/v1/health",
+        tags=["system"],
+        response_model=HealthResponse,
+        responses={
+            503: {
+                "model": HealthResponse,
+                "description": "The process is running but is not ready.",
+            }
+        },
+    )
+    async def health(request: Request):
+        ready = bool(getattr(app.state, "ready", False))
+        if ready:
+            try:
+                status = await _blocking(app.state.database.status)
+                ready = status.get("quick_check") == "ok"
+            except Exception:
+                ready = False
+        return _json(
+            request,
+            {
+                "status": "ready" if ready else "not_ready",
+                "version": __version__,
+                "auth_enabled": configured.auth_enabled,
+            },
+            status_code=200 if ready else 503,
+        )
+
+    @app.get("/v1/system", tags=["system"], response_model=SystemResponse)
+    async def system(request: Request):
+        await authorize(request)
+        runtime = app.state.engine.runtime_summary()
+        return _json(
+            request,
+            {
+                "server_version": __version__,
+                "os": platform.platform(),
+                "architecture": platform.machine(),
+                "cpu": {"model": _cpu_name(), "logical_cores": os.cpu_count()},
+                "runtime": runtime,
+                "execution_provider": app.state.engine.summary.provider,
+                "model": app.state.engine.summary.as_dict(),
+                "database": app.state.database.status(),
+                "data": _path_status(configured.data_dir),
+                "models": _path_status(configured.models_dir),
+                "stats": app.state.repository.stats(),
+                "search": app.state.search_indexes.runtime_summary(),
+                "api_key": {
+                    "authentication_enabled": configured.auth_enabled,
+                    "configured": app.state.repository.has_api_keys(),
+                },
+                "safe_config": {
+                    "max_image_bytes": configured.max_image_bytes,
+                    "max_image_pixels": configured.max_image_pixels,
+                    "max_registration_images": configured.max_registration_images,
+                    "max_request_bytes": configured.max_request_bytes,
+                    "request_timeout_seconds": configured.request_timeout_seconds,
+                    "inference_max_concurrency": (
+                        configured.effective_inference_max_concurrency
+                    ),
+                    "save_face_crops": configured.save_face_crops,
+                    "default_similarity_threshold": configured.default_threshold,
+                    "default_search_profile": configured.default_search_profile,
+                    "default_search_capacity_rows": configured.default_search_capacity_rows,
+                    "max_search_capacity_rows": configured.max_search_capacity_rows,
+                    "default_max_faces_per_person": configured.default_max_faces_per_person,
+                    "default_search_load_policy": configured.default_search_load_policy,
+                    "search_backend": configured.search_backend,
+                    "search_device_id": configured.search_device_id,
+                    "search_topk_mode": configured.search_topk_mode,
+                    "detector_input_sizes": [
+                        list(size) for size in configured.detector_input_sizes
+                    ],
+                    "detection": configured.detection_profile.as_dict(),
+                    "max_detected_faces": configured.max_detected_faces,
+                    "rtsp_max_streams": configured.rtsp_max_streams,
+                    "rtsp_preview_fps": configured.rtsp_preview_fps,
+                    "web_ui_disabled": configured.web_ui_disabled,
+                    "config_file": (
+                        str(configured.config_file) if configured.config_file else None
+                    ),
+                },
+                "recent_errors": list(recent_errors)[-10:],
+            },
+        )
+
+    @app.get("/v1/models", tags=["system"], response_model=ModelsResponse)
+    async def models(request: Request):
+        await authorize(request)
+        return _json(
+            request,
+            {
+                "models": [dict(model) for model in app.state.engine.summary.models],
+                "execution_provider": app.state.engine.summary.provider,
+                "license": app.state.engine.summary.license,
+            },
+        )
+
+    @app.post("/v1/detect", tags=["faces"], response_model=DetectResponse)
+    async def detect(
+        request: Request,
+        image: UploadFile = File(...),
+        max_faces: int | None = Form(default=None, ge=1, le=100),
+        collection_id: str | None = Form(default=None),
+        min_score: float | None = Form(default=None, ge=0.0, le=1.0, deprecated=True),
+    ):
+        await authorize(request)
+        if min_score is not None:
+            raise bad_request(
+                "min_score is configured by the system or Collection detection profile.",
+                code="request_detection_override_not_supported",
+            )
+        started = time.perf_counter()
+        loaded = await app.state.image_loader.from_upload(image)
+        profile = await _blocking(app.state.service.detection_profile, collection_id)
+        faces = await _blocking(
+            app.state.service.detect,
+            loaded,
+            max_faces=max_faces,
+            detection_profile=profile,
+        )
+        return _json(
+            request,
+            {"faces": faces, "processing_ms": round((time.perf_counter() - started) * 1000, 3)},
+        )
+
+    @app.post("/v1/compare", tags=["faces"], response_model=CompareResponse)
+    async def compare(
+        request: Request,
+        source: UploadFile = File(...),
+        target: UploadFile = File(...),
+        threshold: float | None = Form(default=None, ge=0.0, le=1.0),
+        collection_id: str | None = Form(default=None),
+    ):
+        await authorize(request)
+        started = time.perf_counter()
+        source_image, target_image = await asyncio.gather(
+            app.state.image_loader.from_upload(source),
+            app.state.image_loader.from_upload(target),
+        )
+        profile = await _blocking(app.state.service.detection_profile, collection_id)
+        source_face = await _blocking(
+            app.state.service.selected_face,
+            source_image,
+            detection_profile=profile,
+        )
+        target_face = await _blocking(
+            app.state.service.selected_face,
+            target_image,
+            detection_profile=profile,
+        )
+        assert source_face.embedding is not None and target_face.embedding is not None
+        from .services.core import similarity
+
+        score = similarity(source_face.embedding, target_face.embedding)
+        selected_threshold = configured.default_threshold if threshold is None else threshold
+        source_height, source_width = source_image.pixels.shape[:2]
+        target_height, target_width = target_image.pixels.shape[:2]
+        return _json(
+            request,
+            {
+                "matched": score >= selected_threshold,
+                "similarity": score,
+                "threshold": selected_threshold,
+                "source_face": face_result(source_face, source_width, source_height),
+                "target_face": face_result(target_face, target_width, target_height),
+                "processing_ms": round((time.perf_counter() - started) * 1000, 3),
+            },
+        )
+
+    @app.post("/v1/embeddings", tags=["faces"], response_model=EmbeddingsResponse)
+    async def embeddings(
+        request: Request,
+        image: UploadFile = File(...),
+        collection_id: str | None = Form(default=None),
+        face_selection: str | None = Form(default=None, deprecated=True),
+    ):
+        await authorize(request)
+        started = time.perf_counter()
+        if face_selection is not None:
+            raise bad_request(
+                "face_selection is configured by the system or Collection detection profile.",
+                code="request_detection_override_not_supported",
+            )
+        loaded = await app.state.image_loader.from_upload(image)
+        profile = await _blocking(app.state.service.detection_profile, collection_id)
+        selected = await _blocking(
+            app.state.service.selected_face,
+            loaded,
+            detection_profile=profile,
+        )
+        faces = [selected]
+        height, width = loaded.pixels.shape[:2]
+        values = []
+        for face in faces:
+            if face.embedding is None:
+                continue
+            values.append(
+                {
+                    **face_result(face, width, height),
+                    "embedding": [round(float(value), 8) for value in face.embedding],
+                }
+            )
+        return _json(
+            request,
+            {
+                "faces": values,
+                "model": app.state.engine.summary.as_dict(),
+                "processing_ms": round((time.perf_counter() - started) * 1000, 3),
+            },
+        )
+
+    @app.post(
+        "/v1/collections",
+        tags=["collections"],
+        status_code=201,
+        response_model=CollectionResponse,
+    )
+    async def create_collection(request: Request, payload: CollectionCreate):
+        await authorize(request)
+        summary = app.state.engine.summary
+        profile = (
+            configured.default_search_profile
+            if payload.search.profile is None
+            else payload.search.profile
+        )
+        try:
+            app.state.search_indexes.validate_configuration(
+                profile=profile, dimension=summary.embedding_dimension
+            )
+        except SearchIndexError as exc:
+            raise bad_request(str(exc), code="unsupported_search_profile") from exc
+        try:
+            detection_profile = DetectionProfile.from_mapping(
+                payload.detection.model_dump(exclude_none=True),
+                defaults=configured.detection_profile,
+            )
+            app.state.engine.validate_detection_profile(detection_profile)
+            capacity_rows = (
+                configured.default_search_capacity_rows
+                if payload.search.capacity_rows is None
+                else payload.search.capacity_rows
+            )
+            if capacity_rows > configured.max_search_capacity_rows:
+                raise bad_request(
+                    "capacity_rows exceeds this deployment's configured maximum.",
+                    code="search_capacity_too_large",
+                    max_capacity_rows=configured.max_search_capacity_rows,
+                )
+            values = {
+                "id": payload.id,
+                "name": payload.name,
+                "description": payload.description,
+                "default_threshold": (
+                    configured.default_threshold if payload.threshold is None else payload.threshold
+                ),
+                "metadata": payload.metadata,
+                "save_face_crops": (
+                    configured.save_face_crops
+                    if payload.save_face_crops is None
+                    else payload.save_face_crops
+                ),
+                "model_id": summary.model_id,
+                "model_version": summary.model_version,
+                "model_digest": summary.model_digest,
+                "embedding_dimension": summary.embedding_dimension,
+                "preprocessing_version": summary.preprocessing_version,
+                "search_profile": profile,
+                "capacity_rows": capacity_rows,
+                "max_faces_per_person": (
+                    configured.default_max_faces_per_person
+                    if payload.search.max_faces_per_person is None
+                    else payload.search.max_faces_per_person
+                ),
+                "load_policy": (
+                    "eager"
+                    if payload.id == "_default" and payload.search.load_policy is None
+                    else (
+                        configured.default_search_load_policy
+                        if payload.search.load_policy is None
+                        else payload.search.load_policy
+                    )
+                ),
+                "detector_input_sizes": [
+                    list(size) for size in detection_profile.input_sizes
+                ],
+                "detector_threshold": detection_profile.threshold,
+                "detector_nms_threshold": detection_profile.nms_threshold,
+                "single_face_selection": detection_profile.single_face_selection,
+            }
+            item = await _blocking(
+                app.state.search_indexes.run_create,
+                values,
+                lambda: app.state.repository.create_collection(values),
+            )
+        except sqlite3.IntegrityError as exc:
+            raise conflict("The collection already exists.", code="collection_exists") from exc
+        except ValueError as exc:
+            raise bad_request(str(exc), code="invalid_detection_profile") from exc
+        except SearchIndexError as exc:
+            raise ApiError(
+                "search_index_unavailable",
+                "The Collection search index could not be allocated.",
+                503,
+                {"reason": str(exc)},
+            ) from exc
+        return _json(request, {"collection": item}, status_code=201)
+
+    @app.get(
+        "/v1/collections",
+        tags=["collections"],
+        response_model=CollectionPageResponse,
+    )
+    async def list_collections(
+        request: Request,
+        limit: int = Query(default=50, ge=1, le=100),
+        cursor: str | None = Query(default=None),
+    ):
+        await authorize(request)
+        after = app.state.service.cursors.decode(cursor, "collections")
+        rows = app.state.repository.list_collections(after, limit + 1)
+        page, more = rows[:limit], len(rows) > limit
+        next_cursor = (
+            app.state.service.cursors.encode("collections", page[-1]["id"])
+            if more and page
+            else None
+        )
+        return _json(request, {"collections": page, "next_cursor": next_cursor})
+
+    @app.get(
+        "/v1/collections/{collection_id}",
+        tags=["collections"],
+        response_model=CollectionResponse,
+    )
+    async def get_collection(request: Request, collection_id: str):
+        await authorize(request)
+        return _json(request, {"collection": app.state.service.collection(collection_id)})
+
+    @app.patch(
+        "/v1/collections/{collection_id}",
+        tags=["collections"],
+        response_model=CollectionResponse,
+    )
+    async def patch_collection(request: Request, collection_id: str, payload: CollectionPatch):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        changes = payload.model_dump(exclude_unset=True)
+        if "threshold" in changes:
+            changes["default_threshold"] = changes.pop("threshold")
+        search_changes = changes.pop("search", None)
+        if search_changes:
+            changes.update(search_changes)
+        detection_changes = changes.pop("detection", None)
+        if detection_changes:
+            mapping = {
+                "input_sizes": "detector_input_sizes",
+                "threshold": "detector_threshold",
+                "nms_threshold": "detector_nms_threshold",
+                "single_face_selection": "single_face_selection",
+            }
+            for name, value in detection_changes.items():
+                changes[mapping[name]] = value
+        if (
+            "capacity_rows" in changes
+            and int(changes["capacity_rows"]) > configured.max_search_capacity_rows
+        ):
+            raise bad_request(
+                "capacity_rows exceeds this deployment's configured maximum.",
+                code="search_capacity_too_large",
+                max_capacity_rows=configured.max_search_capacity_rows,
+            )
+        item = await _blocking(app.state.service.update_collection, collection_id, changes)
+        return _json(request, {"collection": item})
+
+    @app.delete("/v1/collections/{collection_id}", tags=["collections"], status_code=204)
+    async def delete_collection(
+        request: Request, collection_id: str, force: bool = Query(default=False)
+    ):
+        await authorize(request)
+        await _blocking(app.state.service.delete_collection, collection_id, force=force)
+        return Response(status_code=204)
+
+    @app.post(
+        "/v1/collections/{collection_id}/persons",
+        tags=["persons"],
+        status_code=201,
+        response_model=PersonRegistrationResponse,
+    )
+    async def create_person(
+        request: Request,
+        collection_id: str,
+        images: list[UploadFile] = File(...),
+        id: str | None = Form(default=None),
+        name: str | None = Form(default=None, max_length=200),
+        external_id: str | None = Form(default=None, max_length=200),
+        metadata: str = Form(default="{}"),
+        review_mode: ReviewMode = Form(default="off"),
+        embedding_mode: EmbeddingMode = Form(default="server"),
+        external_embeddings: str | None = Form(default=None),
+        embedding_contract_id: str | None = Form(default=None, max_length=256),
+    ):
+        await authorize(request)
+        try:
+            person_id = validate_id(id) if id else str(uuid.uuid4())
+        except ValueError as exc:
+            raise bad_request(str(exc), code="invalid_person_id") from exc
+        if not images:
+            raise bad_request("At least one image is required.")
+        if len(images) > configured.max_registration_images:
+            raise bad_request(
+                f"At most {configured.max_registration_images} images may be registered at once.",
+                code="too_many_images",
+            )
+        supplied_embeddings = _enrollment_embeddings(
+            embedding_mode=embedding_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+            image_count=len(images),
+        )
+        if embedding_mode == "external_trusted":
+            collection = await _blocking(app.state.service.collection, collection_id)
+            expected_contract = str(collection["embedding_contract_id"])
+            if embedding_contract_id != expected_contract:
+                raise ApiError(
+                    "embedding_contract_mismatch",
+                    "The external embedding contract does not match this Collection.",
+                    409,
+                    {
+                        "expected_embedding_contract_id": expected_contract,
+                        "provided_embedding_contract_id": embedding_contract_id,
+                    },
+                )
+        loaded, indices, invalid = await load_images(images)
+        if not loaded:
+            raise unprocessable(
+                "None of the uploaded images were valid.",
+                code="registration_failed",
+                rejected_images=invalid,
+            )
+        try:
+            loaded_embeddings = (
+                [supplied_embeddings[index] for index in indices]
+                if supplied_embeddings is not None
+                else None
+            )
+            person, faces, rejected = await _blocking(
+                app.state.service.create_person,
+                collection_id,
+                {
+                    "id": person_id,
+                    "name": name,
+                    "external_id": external_id,
+                    "metadata": _metadata(metadata),
+                },
+                loaded,
+                review_mode=review_mode,
+                embedding_mode=embedding_mode,
+                external_embeddings=loaded_embeddings,
+                embedding_contract_id=embedding_contract_id,
+            )
+        except ApiError as exc:
+            if exc.code == "registration_failed" and invalid:
+                exc.details["rejected_images"] = invalid + remap_rejections(
+                    list(exc.details.get("rejected_images", [])), indices
+                )
+            raise
+        return _json(
+            request,
+            {
+                "person": person,
+                "faces": faces,
+                "rejected_images": invalid + remap_rejections(rejected, indices),
+            },
+            status_code=201,
+        )
+
+    @app.get(
+        "/v1/collections/{collection_id}/persons",
+        tags=["persons"],
+        response_model=PersonPageResponse,
+    )
+    async def list_persons(
+        request: Request,
+        collection_id: str,
+        limit: int = Query(default=50, ge=1, le=100),
+        cursor: str | None = Query(default=None),
+        search: str | None = Query(default=None, max_length=200),
+    ):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        scope = f"persons:{collection_id}:{search or ''}"
+        after = app.state.service.cursors.decode(cursor, scope)
+        rows = app.state.repository.list_persons(collection_id, after, limit + 1, search)
+        page, more = rows[:limit], len(rows) > limit
+        next_cursor = (
+            app.state.service.cursors.encode(scope, page[-1]["id"]) if more and page else None
+        )
+        return _json(request, {"persons": page, "next_cursor": next_cursor})
+
+    @app.get(
+        "/v1/collections/{collection_id}/persons/{person_id}",
+        tags=["persons"],
+        response_model=PersonResponse,
+    )
+    async def get_person(request: Request, collection_id: str, person_id: str):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        person = app.state.repository.get_person(collection_id, person_id)
+        if person is None:
+            raise not_found("person", person_id)
+        return _json(request, {"person": person})
+
+    @app.patch(
+        "/v1/collections/{collection_id}/persons/{person_id}",
+        tags=["persons"],
+        response_model=PersonResponse,
+    )
+    async def patch_person(
+        request: Request, collection_id: str, person_id: str, payload: PersonPatch
+    ):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        try:
+            person = app.state.repository.update_person(
+                collection_id, person_id, payload.model_dump(exclude_unset=True)
+            )
+        except sqlite3.IntegrityError as exc:
+            raise conflict("The external ID already exists.", code="external_id_exists") from exc
+        if person is None:
+            raise not_found("person", person_id)
+        return _json(request, {"person": person})
+
+    @app.delete(
+        "/v1/collections/{collection_id}/persons/{person_id}",
+        tags=["persons"],
+        status_code=204,
+    )
+    async def delete_person(request: Request, collection_id: str, person_id: str):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        crops = await _blocking(app.state.service.delete_person, collection_id, person_id)
+        for crop in crops:
+            app.state.service.crops.delete(crop)
+        return Response(status_code=204)
+
+    @app.post(
+        "/v1/collections/{collection_id}/persons/{person_id}/faces",
+        tags=["persons"],
+        status_code=201,
+        response_model=FaceRegistrationResponse,
+    )
+    async def add_faces(
+        request: Request,
+        collection_id: str,
+        person_id: str,
+        images: list[UploadFile] = File(...),
+        review_mode: ReviewMode = Form(default="off"),
+        embedding_mode: EmbeddingMode = Form(default="server"),
+        external_embeddings: str | None = Form(default=None),
+        embedding_contract_id: str | None = Form(default=None, max_length=256),
+    ):
+        await authorize(request)
+        if len(images) > configured.max_registration_images:
+            raise bad_request(
+                f"At most {configured.max_registration_images} images may be registered at once.",
+                code="too_many_images",
+            )
+        supplied_embeddings = _enrollment_embeddings(
+            embedding_mode=embedding_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+            image_count=len(images),
+        )
+        if embedding_mode == "external_trusted":
+            collection = await _blocking(app.state.service.collection, collection_id)
+            expected_contract = str(collection["embedding_contract_id"])
+            if embedding_contract_id != expected_contract:
+                raise ApiError(
+                    "embedding_contract_mismatch",
+                    "The external embedding contract does not match this Collection.",
+                    409,
+                    {
+                        "expected_embedding_contract_id": expected_contract,
+                        "provided_embedding_contract_id": embedding_contract_id,
+                    },
+                )
+        loaded, indices, invalid = await load_images(images)
+        loaded_embeddings = (
+            [supplied_embeddings[index] for index in indices]
+            if supplied_embeddings is not None
+            else None
+        )
+        faces, rejected = await _blocking(
+            app.state.service.add_faces,
+            collection_id,
+            person_id,
+            loaded,
+            review_mode=review_mode,
+            embedding_mode=embedding_mode,
+            external_embeddings=loaded_embeddings,
+            embedding_contract_id=embedding_contract_id,
+        )
+        return _json(
+            request,
+            {
+                "faces": faces,
+                "rejected_images": invalid + remap_rejections(rejected, indices),
+            },
+            status_code=201,
+        )
+
+    @app.get(
+        "/v1/collections/{collection_id}/persons/{person_id}/faces",
+        tags=["persons"],
+        response_model=FacePageResponse,
+    )
+    async def list_faces(
+        request: Request,
+        collection_id: str,
+        person_id: str,
+        limit: int = Query(default=50, ge=1, le=100),
+        cursor: str | None = Query(default=None),
+    ):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        if app.state.repository.get_person(collection_id, person_id) is None:
+            raise not_found("person", person_id)
+        scope = f"faces:{collection_id}:{person_id}"
+        after = app.state.service.cursors.decode(cursor, scope)
+        rows = app.state.repository.list_faces(collection_id, person_id, after, limit + 1)
+        page, more = rows[:limit], len(rows) > limit
+        next_cursor = (
+            app.state.service.cursors.encode(scope, page[-1]["id"]) if more and page else None
+        )
+        return _json(
+            request,
+            {
+                "faces": [app.state.service.public_face(item) for item in page],
+                "next_cursor": next_cursor,
+            },
+        )
+
+    @app.get(
+        "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image",
+        tags=["persons"],
+        response_class=Response,
+        responses={
+            200: {
+                "description": "Saved 112 x 112 face crop",
+                "content": {"image/jpeg": {"schema": {"type": "string", "format": "binary"}}},
+            }
+        },
+    )
+    async def get_face_image(request: Request, collection_id: str, person_id: str, face_id: str):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        face = await _blocking(
+            app.state.repository.get_face_crop, collection_id, person_id, face_id
+        )
+        if face is None:
+            existing = await _blocking(
+                app.state.repository.get_face, collection_id, person_id, face_id
+            )
+            if existing is None:
+                raise not_found("face", face_id)
+            raise ApiError(
+                "face_image_not_found",
+                "The FaceSample has no saved image.",
+                404,
+            )
+
+        image = face.get("bytes")
+        if image is None and face.get("crop_path"):
+            try:
+                image = await _blocking(app.state.service.crops.read, str(face["crop_path"]))
+            except (OSError, RuntimeError):
+                image = None
+        if image is None:
+            raise ApiError(
+                "face_image_not_found",
+                "The FaceSample has no saved image.",
+                404,
+            )
+        return Response(
+            content=bytes(image),
+            media_type="image/jpeg",
+            headers={"Cache-Control": "no-store"},
+        )
+
+    @app.delete(
+        "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}",
+        tags=["persons"],
+        status_code=204,
+    )
+    async def delete_face(request: Request, collection_id: str, person_id: str, face_id: str):
+        await authorize(request)
+        app.state.service.collection(collection_id)
+        crop = await _blocking(app.state.service.delete_face, collection_id, person_id, face_id)
+        app.state.service.crops.delete(crop)
+        return Response(status_code=204)
+
+    @app.post(
+        "/v1/collections/{collection_id}/search",
+        tags=["search"],
+        response_model=SearchResponse,
+    )
+    async def search(
+        request: Request,
+        collection_id: str,
+        image: UploadFile = File(...),
+        limit: int = Form(default=5, ge=1, le=100),
+        threshold: float | None = Form(default=None, ge=0.0, le=1.0),
+        face_selection: str | None = Form(default=None, deprecated=True),
+    ):
+        await authorize(request)
+        if face_selection is not None:
+            raise bad_request(
+                "face_selection is configured by the Collection detection profile.",
+                code="request_detection_override_not_supported",
+            )
+        started = time.perf_counter()
+        loaded = await app.state.image_loader.from_upload(image)
+        searched_face, matches, selected_threshold = await _blocking(
+            app.state.service.search,
+            collection_id,
+            loaded,
+            limit=limit,
+            threshold=threshold,
+        )
+        return _json(
+            request,
+            {
+                "searched_face": searched_face,
+                "matches": matches,
+                "threshold": selected_threshold,
+                "processing_ms": round((time.perf_counter() - started) * 1000, 3),
+            },
+        )
+
+    def monitor_values(payload: MonitorCreate | MonitorPatch) -> dict[str, Any]:
+        values = payload.model_dump(exclude_unset=True)
+        source = values.pop("source", None)
+        if source is not None:
+            values["source_type"] = source["type"]
+            values["url"] = source["url"]
+        policy = values.pop("event_policy", None)
+        if policy is not None:
+            values.update(policy)
+        return values
+
+    @app.post(
+        "/v1/monitors",
+        tags=["monitors"],
+        status_code=201,
+        response_model=MonitorResponse,
+        responses={
+            429: {
+                "model": ErrorEnvelope,
+                "description": "The enabled Monitor capacity is exhausted.",
+            }
+        },
+    )
+    async def create_monitor(request: Request, payload: MonitorCreate):
+        await authorize(request)
+        try:
+            monitor = await _blocking(
+                app.state.monitors.create,
+                monitor_values(payload),
+            )
+        except MonitorAlreadyExistsError as exc:
+            raise conflict(
+                "The Monitor already exists.",
+                code="monitor_exists",
+            ) from exc
+        except MonitorLimitError as exc:
+            raise ApiError(
+                "monitor_limit_exceeded",
+                "The maximum number of enabled Monitors is already running.",
+                429,
+                {"max_monitors": configured.rtsp_max_streams},
+            ) from exc
+        return _json(request, {"monitor": monitor}, status_code=201)
+
+    @app.get(
+        "/v1/monitors",
+        tags=["monitors"],
+        response_model=MonitorPageResponse,
+    )
+    async def list_monitors(
+        request: Request,
+        limit: int = Query(default=50, ge=1, le=100),
+        cursor: str | None = Query(default=None),
+    ):
+        await authorize(request)
+        after = app.state.service.cursors.decode(cursor, "monitors")
+        rows = await _blocking(app.state.monitors.list, after, limit + 1)
+        page, more = rows[:limit], len(rows) > limit
+        next_cursor = (
+            app.state.service.cursors.encode("monitors", page[-1]["id"])
+            if more and page
+            else None
+        )
+        return _json(
+            request,
+            {"monitors": page, "next_cursor": next_cursor},
+        )
+
+    @app.get(
+        "/v1/monitors/{monitor_id}",
+        tags=["monitors"],
+        response_model=MonitorResponse,
+    )
+    async def get_monitor(request: Request, monitor_id: str):
+        await authorize(request)
+        try:
+            monitor = await _blocking(app.state.monitors.get, monitor_id)
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        return _json(request, {"monitor": monitor})
+
+    @app.patch(
+        "/v1/monitors/{monitor_id}",
+        tags=["monitors"],
+        response_model=MonitorResponse,
+        responses={
+            429: {
+                "model": ErrorEnvelope,
+                "description": "The enabled Monitor capacity is exhausted.",
+            }
+        },
+    )
+    async def patch_monitor(
+        request: Request,
+        monitor_id: str,
+        payload: MonitorPatch,
+    ):
+        await authorize(request)
+        try:
+            monitor = await _blocking(
+                app.state.monitors.update,
+                monitor_id,
+                monitor_values(payload),
+            )
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        except MonitorLimitError as exc:
+            raise ApiError(
+                "monitor_limit_exceeded",
+                "The maximum number of enabled Monitors is already running.",
+                429,
+                {"max_monitors": configured.rtsp_max_streams},
+            ) from exc
+        return _json(request, {"monitor": monitor})
+
+    @app.delete(
+        "/v1/monitors/{monitor_id}",
+        tags=["monitors"],
+        status_code=204,
+    )
+    async def delete_monitor(request: Request, monitor_id: str):
+        await authorize(request)
+        try:
+            await _blocking(app.state.monitors.delete, monitor_id)
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        return Response(status_code=204)
+
+    @app.get(
+        "/v1/monitors/{monitor_id}/state",
+        tags=["monitors"],
+        response_model=MonitorStateResponse,
+    )
+    async def monitor_state(request: Request, monitor_id: str):
+        await authorize(request)
+        try:
+            state = await _blocking(app.state.monitors.state, monitor_id)
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        return _json(request, {"state": state})
+
+    @app.get(
+        "/v1/monitors/{monitor_id}/events",
+        tags=["monitors"],
+        response_model=MonitorEventPageResponse,
+    )
+    async def monitor_events(
+        request: Request,
+        monitor_id: str,
+        cursor: str | None = Query(default=None),
+        limit: int = Query(default=100, ge=1, le=1000),
+    ):
+        await authorize(request)
+        try:
+            page = await _blocking(
+                app.state.monitors.events,
+                monitor_id,
+                cursor=cursor,
+                limit=limit,
+            )
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        return _json(request, page)
+
+    @app.get(
+        "/v1/monitors/{monitor_id}/preview.mjpeg",
+        tags=["monitors"],
+        response_class=StreamingResponse,
+        responses={
+            200: {
+                "description": "Live Motion JPEG preview stream.",
+                "content": {
+                    "multipart/x-mixed-replace": {
+                        "schema": {"type": "string", "format": "binary"}
+                    }
+                },
+            }
+        },
+    )
+    async def monitor_preview(request: Request, monitor_id: str):
+        await authorize(request)
+        try:
+            stream = app.state.monitors.preview(monitor_id)
+        except MonitorNotFoundError as exc:
+            raise not_found("monitor", monitor_id) from exc
+        except MonitorPreviewDisabledError as exc:
+            raise conflict(
+                "Web preview is disabled for this Monitor.",
+                code="preview_disabled",
+            ) from exc
+        except MonitorUnavailableError as exc:
+            raise ApiError(
+                "stream_unavailable",
+                "The Monitor does not currently have an RTSP frame to preview.",
+                503,
+            ) from exc
+        return StreamingResponse(
+            stream,
+            media_type="multipart/x-mixed-replace; boundary=frame",
+            headers={
+                "Cache-Control": "no-store, no-cache, must-revalidate",
+                "Pragma": "no-cache",
+                "X-Accel-Buffering": "no",
+            },
+        )
+
+    if not configured.web_ui_disabled:
+        frontend = SERVER_DIR / "frontend"
+        app.mount("/assets", StaticFiles(directory=frontend), name="assets")
+        app.mount(
+            "/guide-images",
+            StaticFiles(directory=SERVER_DIR / "docs" / "images"),
+            name="guide-images",
+        )
+
+        @app.get("/", include_in_schema=False)
+        async def index():
+            return FileResponse(frontend / "index.html")
+
+        @app.get("/docs", include_in_schema=False)
+        async def docs():
+            return FileResponse(frontend / "openapi.html")
+
+        @app.get(
+            "/guide-content/{locale_code}/{document}.md",
+            include_in_schema=False,
+        )
+        async def guide_content(locale_code: str, document: str):
+            """Serve only versioned Markdown bundled in this image."""
+
+            path = DOCUMENTS.get(locale_code, {}).get(document)
+            if path is None or not path.is_file():
+                raise StarletteHTTPException(
+                    status_code=404, detail="Documentation not found"
+                )
+            return FileResponse(
+                path,
+                media_type="text/markdown; charset=utf-8",
+                headers={
+                    "Cache-Control": "public, max-age=300",
+                    "X-Content-Type-Options": "nosniff",
+                },
+            )
+
+    def custom_openapi() -> dict[str, Any]:
+        if app.openapi_schema:
+            return app.openapi_schema
+        schema = get_openapi(
+            title=app.title,
+            version=app.version,
+            description=app.description,
+            routes=app.routes,
+        )
+        components = schema.setdefault("components", {})
+        components.setdefault("securitySchemes", {})["bearerAuth"] = {
+            "type": "http",
+            "scheme": "bearer",
+        }
+        components.setdefault("schemas", {})["ErrorEnvelope"] = {
+            "type": "object",
+            "required": ["error", "request_id"],
+            "properties": {
+                "error": {
+                    "type": "object",
+                    "required": ["code", "message", "details"],
+                    "properties": {
+                        "code": {"type": "string"},
+                        "message": {"type": "string"},
+                        "details": {"type": "object"},
+                    },
+                },
+                "request_id": {"type": "string", "format": "uuid"},
+            },
+        }
+        error_response = {
+            "description": "Error",
+            "content": {
+                "application/json": {"schema": {"$ref": "#/components/schemas/ErrorEnvelope"}}
+            },
+        }
+        for path, operations in schema.get("paths", {}).items():
+            for operation in operations.values():
+                if not isinstance(operation, dict) or "responses" not in operation:
+                    continue
+                if path.startswith("/v1/") and path != "/v1/health":
+                    operation["security"] = [{"bearerAuth": []}]
+                operation["responses"].pop("422", None)
+                for status in ("400", "401", "404", "409", "413", "422", "500", "503"):
+                    operation["responses"].setdefault(status, error_response)
+        app.openapi_schema = schema
+        return schema
+
+    app.openapi = custom_openapi  # type: ignore[method-assign]
+
+    return app
+
+
+app = create_app()
diff --git a/server/backend/insightface_server/config.py b/server/backend/insightface_server/config.py
new file mode 100644
index 000000000..998935636
--- /dev/null
+++ b/server/backend/insightface_server/config.py
@@ -0,0 +1,516 @@
+from __future__ import annotations
+
+import math
+import os
+import tomllib
+from collections.abc import Mapping
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Literal, cast
+
+SingleFaceSelection = Literal["largest", "center_largest"]
+
+SUPPORTED_SEARCH_PROFILES = (
+    "fp32_v1",
+    "fp16_v1",
+    "bf16_v1",
+    "int8_x736_v1",
+    "int8_x1000_v1",
+)
+SUPPORTED_SEARCH_LOAD_POLICIES = ("eager", "lazy")
+SUPPORTED_SEARCH_BACKENDS = ("auto", "reference", "native_cpu", "native_cuda")
+SUPPORTED_SEARCH_TOPK_MODES = ("auto", "host", "device")
+DEFAULT_DETECTOR_INPUT_SIZES = ((96, 96), (512, 512))
+DEFAULT_DETECTOR_THRESHOLD = 0.50
+DEFAULT_DETECTOR_NMS_THRESHOLD = 0.40
+DEFAULT_SINGLE_FACE_SELECTION: SingleFaceSelection = "largest"
+DEFAULT_MAX_DETECTED_FACES = 100
+DEFAULT_CPU_INFERENCE_MAX_CONCURRENCY = 4
+DEFAULT_CUDA_INFERENCE_MAX_CONCURRENCY = 8
+MAX_INFERENCE_MAX_CONCURRENCY = 256
+SUPPORTED_SINGLE_FACE_SELECTIONS = ("largest", "center_largest")
+MAX_DETECTOR_INPUT_SIZE_COUNT = 4
+MIN_DETECTOR_INPUT_SIDE = 32
+MAX_DETECTOR_INPUT_SIDE = 2048
+MAX_DETECTOR_INPUT_PIXELS = 4 * 1024 * 1024
+
+
+def _finite_unit_interval(value: object, *, name: str) -> float:
+    if isinstance(value, bool) or not isinstance(value, int | float):
+        raise ValueError(f"{name} must be a number")
+    normalized = float(value)
+    if not math.isfinite(normalized) or not 0.0 <= normalized <= 1.0:
+        raise ValueError(f"{name} must be finite and within 0.0..1.0")
+    return normalized
+
+
+def normalize_single_face_selection(value: object) -> SingleFaceSelection:
+    if not isinstance(value, str):
+        raise ValueError("single_face_selection must be a string")
+    normalized = value.strip().lower()
+    if normalized not in SUPPORTED_SINGLE_FACE_SELECTIONS:
+        allowed = ", ".join(SUPPORTED_SINGLE_FACE_SELECTIONS)
+        raise ValueError(f"single_face_selection must be one of: {allowed}")
+    return cast(SingleFaceSelection, normalized)
+
+
+def normalize_detector_input_sizes(value: object) -> tuple[tuple[int, int], ...]:
+    """Validate and canonicalize the startup-only detector-size setting.
+
+    SCRFD's largest feature-map stride is 32. Requiring aligned dimensions
+    prevents ambiguous partial feature-map behavior, while the count, side and
+    total-pixel limits bound a configuration mistake before it
+    can exhaust CPU RAM or GPU VRAM.
+    """
+
+    if not isinstance(value, list | tuple) or not value:
+        raise ValueError("detector_input_sizes must be a non-empty list of [width, height]")
+    if len(value) > MAX_DETECTOR_INPUT_SIZE_COUNT:
+        raise ValueError(
+            f"detector_input_sizes supports at most {MAX_DETECTOR_INPUT_SIZE_COUNT} sizes"
+        )
+    normalized: list[tuple[int, int]] = []
+    total_pixels = 0
+    for item in value:
+        if not isinstance(item, list | tuple) or len(item) != 2:
+            raise ValueError("each detector input size must be [width, height]")
+        width, height = item
+        if (
+            isinstance(width, bool)
+            or isinstance(height, bool)
+            or not isinstance(width, int)
+            or not isinstance(height, int)
+        ):
+            raise ValueError("detector input width and height must be integers")
+        if not (
+            MIN_DETECTOR_INPUT_SIDE <= width <= MAX_DETECTOR_INPUT_SIDE
+            and MIN_DETECTOR_INPUT_SIDE <= height <= MAX_DETECTOR_INPUT_SIDE
+        ):
+            raise ValueError(
+                "detector input width and height must be within "
+                f"{MIN_DETECTOR_INPUT_SIDE}..{MAX_DETECTOR_INPUT_SIDE}"
+            )
+        if width % 32 or height % 32:
+            raise ValueError("detector input width and height must be multiples of 32")
+        size = (width, height)
+        if size in normalized:
+            raise ValueError("detector input sizes must not contain duplicates")
+        normalized.append(size)
+        total_pixels += width * height
+    if total_pixels > MAX_DETECTOR_INPUT_PIXELS:
+        raise ValueError(
+            "combined detector input pixels exceed the supported system limit "
+            f"of {MAX_DETECTOR_INPUT_PIXELS}"
+        )
+    return tuple(normalized)
+
+
+@dataclass(frozen=True, slots=True)
+class DetectionProfile:
+    """One immutable detector policy used for the whole request."""
+
+    input_sizes: tuple[tuple[int, int], ...] = DEFAULT_DETECTOR_INPUT_SIZES
+    threshold: float = DEFAULT_DETECTOR_THRESHOLD
+    nms_threshold: float = DEFAULT_DETECTOR_NMS_THRESHOLD
+    single_face_selection: SingleFaceSelection = DEFAULT_SINGLE_FACE_SELECTION
+
+    @classmethod
+    def from_mapping(
+        cls,
+        value: Mapping[str, Any] | None,
+        *,
+        defaults: DetectionProfile | None = None,
+    ) -> DetectionProfile:
+        base = defaults or cls()
+        document = value or {}
+        unknown = set(document) - {
+            "input_sizes",
+            "threshold",
+            "nms_threshold",
+            "single_face_selection",
+        }
+        if unknown:
+            names = ", ".join(sorted(unknown))
+            raise ValueError(f"Unsupported detection setting(s): {names}")
+        return cls(
+            input_sizes=normalize_detector_input_sizes(
+                document.get("input_sizes", base.input_sizes)
+            ),
+            threshold=_finite_unit_interval(
+                document.get("threshold", base.threshold), name="detection threshold"
+            ),
+            nms_threshold=_finite_unit_interval(
+                document.get("nms_threshold", base.nms_threshold),
+                name="detection nms_threshold",
+            ),
+            single_face_selection=normalize_single_face_selection(
+                document.get("single_face_selection", base.single_face_selection)
+            ),
+        )
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "input_sizes": [list(size) for size in self.input_sizes],
+            "threshold": self.threshold,
+            "nms_threshold": self.nms_threshold,
+            "single_face_selection": self.single_face_selection,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class ServerFileConfig:
+    detection: DetectionProfile = DetectionProfile()
+    max_detected_faces: int = DEFAULT_MAX_DETECTED_FACES
+    inference_max_concurrency: int | None = None
+    web_ui_disabled: bool = False
+
+
+def default_inference_max_concurrency(execution_provider: str) -> int:
+    return (
+        DEFAULT_CUDA_INFERENCE_MAX_CONCURRENCY
+        if execution_provider == "CUDAExecutionProvider"
+        else DEFAULT_CPU_INFERENCE_MAX_CONCURRENCY
+    )
+
+
+def normalize_inference_max_concurrency(
+    value: object, *, name: str = "[inference].max_concurrency"
+) -> int | None:
+    if isinstance(value, str) and value.strip().lower() == "auto":
+        return None
+    if isinstance(value, bool) or not isinstance(value, int):
+        raise ValueError(f"{name} must be 'auto' or an integer")
+    if not 1 <= value <= MAX_INFERENCE_MAX_CONCURRENCY:
+        raise ValueError(
+            f"{name} must be within 1..{MAX_INFERENCE_MAX_CONCURRENCY}"
+        )
+    return value
+
+
+def load_server_config(
+    path: Path | None,
+) -> ServerFileConfig:
+    """Read and validate the immutable startup configuration from TOML."""
+
+    if path is None:
+        return ServerFileConfig()
+    try:
+        with path.open("rb") as stream:
+            document = tomllib.load(stream)
+    except OSError as exc:
+        raise ValueError(f"Unable to read INSIGHTFACE_CONFIG_FILE '{path}': {exc}") from exc
+    except tomllib.TOMLDecodeError as exc:
+        raise ValueError(f"INSIGHTFACE_CONFIG_FILE '{path}' is not valid TOML: {exc}") from exc
+
+    unknown_sections = set(document) - {"detection", "inference", "web"}
+    if unknown_sections:
+        names = ", ".join(sorted(unknown_sections))
+        raise ValueError(f"Unsupported server configuration section(s): {names}")
+    detection = document.get("detection", {})
+    if not isinstance(detection, dict):
+        raise ValueError("[detection] must be a TOML table")
+    unknown_detection = set(detection) - {
+        "input_sizes",
+        "threshold",
+        "nms_threshold",
+        "single_face_selection",
+        "max_detected_faces",
+    }
+    if unknown_detection:
+        names = ", ".join(sorted(unknown_detection))
+        raise ValueError(f"Unsupported [detection] setting(s): {names}")
+    web = document.get("web", {})
+    if not isinstance(web, dict):
+        raise ValueError("[web] must be a TOML table")
+    unknown_web = set(web) - {"disabled"}
+    if unknown_web:
+        names = ", ".join(sorted(unknown_web))
+        raise ValueError(f"Unsupported [web] setting(s): {names}")
+    web_ui_disabled = web.get("disabled", False)
+    if not isinstance(web_ui_disabled, bool):
+        raise ValueError("[web].disabled must be a boolean")
+    inference = document.get("inference", {})
+    if not isinstance(inference, dict):
+        raise ValueError("[inference] must be a TOML table")
+    unknown_inference = set(inference) - {"max_concurrency"}
+    if unknown_inference:
+        names = ", ".join(sorted(unknown_inference))
+        raise ValueError(f"Unsupported [inference] setting(s): {names}")
+    inference_max_concurrency = normalize_inference_max_concurrency(
+        inference.get("max_concurrency", "auto")
+    )
+    max_detected_faces = detection.get(
+        "max_detected_faces", DEFAULT_MAX_DETECTED_FACES
+    )
+    if (
+        isinstance(max_detected_faces, bool)
+        or not isinstance(max_detected_faces, int)
+        or not 1 <= max_detected_faces <= 100
+    ):
+        raise ValueError("[detection].max_detected_faces must be an integer within 1..100")
+    profile_values = {
+        key: value
+        for key, value in detection.items()
+        if key != "max_detected_faces"
+    }
+    return ServerFileConfig(
+        detection=DetectionProfile.from_mapping(profile_values),
+        max_detected_faces=max_detected_faces,
+        inference_max_concurrency=inference_max_concurrency,
+        web_ui_disabled=web_ui_disabled,
+    )
+
+
+def load_detector_input_sizes(path: Path | None) -> tuple[tuple[int, int], ...]:
+    """Backward-compatible helper for the detector portion of server TOML."""
+
+    return load_server_config(path).detection.input_sizes
+
+
+def _bool(name: str, default: bool) -> bool:
+    raw = os.getenv(name)
+    if raw is None:
+        return default
+    value = raw.strip().lower()
+    if value in {"1", "true", "yes", "on"}:
+        return True
+    if value in {"0", "false", "no", "off"}:
+        return False
+    raise ValueError(f"{name} must be an explicit boolean")
+
+
+def _int(name: str, default: int, *, minimum: int = 1) -> int:
+    value = int(os.getenv(name, str(default)))
+    if value < minimum:
+        raise ValueError(f"{name} must be >= {minimum}")
+    return value
+
+
+def _float(name: str, default: float, *, minimum: float, maximum: float) -> float:
+    value = float(os.getenv(name, str(default)))
+    if not math.isfinite(value) or not minimum <= value <= maximum:
+        raise ValueError(f"{name} must be finite and within {minimum}..{maximum}")
+    return value
+
+
+def _choice(name: str, default: str, choices: tuple[str, ...]) -> str:
+    value = os.getenv(name, default).strip().lower()
+    if value not in choices:
+        allowed = ", ".join(choices)
+        raise ValueError(f"{name} must be one of: {allowed}")
+    return value
+
+
+@dataclass(frozen=True, slots=True)
+class Settings:
+    data_dir: Path
+    models_dir: Path
+    inference_mode: str
+    execution_provider: str
+    auth_enabled: bool
+    startup_api_key: str | None
+    max_image_bytes: int
+    max_image_pixels: int
+    max_request_bytes: int
+    max_registration_images: int
+    request_timeout_seconds: int
+    detector_threshold: float
+    registration_min_score: float
+    registration_min_quality: float
+    registration_min_face_size: int
+    default_threshold: float
+    cors_origins: tuple[str, ...]
+    log_level: str
+    save_face_crops: bool
+    device_id: int
+    inference_max_concurrency: int | None = None
+    detector_input_sizes: tuple[tuple[int, int], ...] = DEFAULT_DETECTOR_INPUT_SIZES
+    detector_nms_threshold: float = DEFAULT_DETECTOR_NMS_THRESHOLD
+    detector_single_face_selection: SingleFaceSelection = DEFAULT_SINGLE_FACE_SELECTION
+    max_detected_faces: int = DEFAULT_MAX_DETECTED_FACES
+    config_file: Path | None = None
+    default_search_profile: str = "fp32_v1"
+    default_search_capacity_rows: int = 100_000
+    max_search_capacity_rows: int = 10_000_000
+    default_max_faces_per_person: int = 20
+    default_search_load_policy: str = "lazy"
+    search_backend: str = "auto"
+    search_library_path: Path | None = None
+    search_device_id: int = 0
+    search_topk_mode: str = "auto"
+    search_build_batch_rows: int = 4096
+    web_ui_disabled: bool = False
+    rtsp_max_streams: int = 4
+    rtsp_preview_fps: float = 5.0
+    rtsp_jpeg_quality: int = 82
+    rtsp_open_timeout_seconds: float = 5.0
+    rtsp_read_timeout_seconds: float = 5.0
+    rtsp_reconnect_delay_seconds: float = 1.0
+
+    @property
+    def detection_profile(self) -> DetectionProfile:
+        return DetectionProfile(
+            input_sizes=self.detector_input_sizes,
+            threshold=self.detector_threshold,
+            nms_threshold=self.detector_nms_threshold,
+            single_face_selection=self.detector_single_face_selection,
+        )
+
+    @property
+    def effective_inference_max_concurrency(self) -> int:
+        return self.inference_max_concurrency or default_inference_max_concurrency(
+            self.execution_provider
+        )
+
+    @classmethod
+    def from_env(cls) -> Settings:
+        mode = os.getenv("INSIGHTFACE_INFERENCE_MODE", "onnx").strip().lower()
+        if mode not in {"onnx", "mock"}:
+            raise ValueError("INSIGHTFACE_INFERENCE_MODE must be onnx or mock")
+        provider = os.getenv(
+            "INSIGHTFACE_EXECUTION_PROVIDER", "CPUExecutionProvider"
+        ).strip()
+        if provider not in {"CPUExecutionProvider", "CUDAExecutionProvider"}:
+            raise ValueError("Unsupported INSIGHTFACE_EXECUTION_PROVIDER")
+        origins = tuple(
+            part.strip()
+            for part in os.getenv("INSIGHTFACE_CORS_ORIGINS", "").split(",")
+            if part.strip()
+        )
+        device_id = _int("INSIGHTFACE_DEVICE_ID", 0, minimum=0)
+        search_library = os.getenv("INSIGHTFACE_SEARCH_LIBRARY")
+        config_file_value = os.getenv("INSIGHTFACE_CONFIG_FILE")
+        config_file = (
+            Path(config_file_value).expanduser() if config_file_value else None
+        )
+        file_config = load_server_config(config_file)
+        inference_concurrency_value = os.getenv(
+            "INSIGHTFACE_INFERENCE_MAX_CONCURRENCY"
+        )
+        if inference_concurrency_value is None:
+            inference_max_concurrency = file_config.inference_max_concurrency
+        else:
+            raw_value = inference_concurrency_value.strip().lower()
+            try:
+                raw_concurrency: object = (
+                    int(inference_concurrency_value) if raw_value != "auto" else "auto"
+                )
+            except ValueError as exc:
+                raise ValueError(
+                    "INSIGHTFACE_INFERENCE_MAX_CONCURRENCY must be 'auto' or an integer"
+                ) from exc
+            inference_max_concurrency = normalize_inference_max_concurrency(
+                raw_concurrency, name="INSIGHTFACE_INFERENCE_MAX_CONCURRENCY"
+            )
+        default_capacity_rows = _int(
+            "INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS", 100_000
+        )
+        max_capacity_rows = _int(
+            "INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS", 10_000_000
+        )
+        if default_capacity_rows > max_capacity_rows:
+            raise ValueError(
+                "INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS must not exceed "
+                "INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS"
+            )
+        rtsp_jpeg_quality = _int("INSIGHTFACE_RTSP_JPEG_QUALITY", 82)
+        if rtsp_jpeg_quality > 100:
+            raise ValueError("INSIGHTFACE_RTSP_JPEG_QUALITY must be within 1..100")
+        return cls(
+            data_dir=Path(os.getenv("INSIGHTFACE_DATA_DIR", "/data")).expanduser(),
+            models_dir=Path(os.getenv("INSIGHTFACE_MODELS_DIR", "/models")).expanduser(),
+            inference_mode=mode,
+            execution_provider=provider,
+            auth_enabled=_bool("INSIGHTFACE_AUTH_ENABLED", True),
+            startup_api_key=os.getenv("INSIGHTFACE_API_KEY") or None,
+            max_image_bytes=_int("INSIGHTFACE_MAX_IMAGE_BYTES", 10 * 1024 * 1024),
+            max_image_pixels=_int("INSIGHTFACE_MAX_IMAGE_PIXELS", 40_000_000),
+            max_request_bytes=_int("INSIGHTFACE_MAX_REQUEST_BYTES", 64 * 1024 * 1024),
+            max_registration_images=_int("INSIGHTFACE_MAX_REGISTRATION_IMAGES", 20),
+            request_timeout_seconds=_int("INSIGHTFACE_REQUEST_TIMEOUT_SECONDS", 60),
+            detector_threshold=_float(
+                "INSIGHTFACE_DETECTOR_THRESHOLD",
+                file_config.detection.threshold,
+                minimum=0.0,
+                maximum=1.0,
+            ),
+            registration_min_score=_float(
+                "INSIGHTFACE_REGISTRATION_MIN_SCORE", 0.6, minimum=0.0, maximum=1.0
+            ),
+            registration_min_quality=_float(
+                "INSIGHTFACE_REGISTRATION_MIN_QUALITY", 0.35, minimum=0.0, maximum=1.0
+            ),
+            registration_min_face_size=_int(
+                "INSIGHTFACE_REGISTRATION_MIN_FACE_SIZE", 40
+            ),
+            default_threshold=_float(
+                "INSIGHTFACE_DEFAULT_THRESHOLD", 0.4, minimum=0.0, maximum=1.0
+            ),
+            cors_origins=origins,
+            log_level=os.getenv("INSIGHTFACE_LOG_LEVEL", "INFO").upper(),
+            save_face_crops=_bool("INSIGHTFACE_SAVE_FACE_CROPS", False),
+            device_id=device_id,
+            inference_max_concurrency=(
+                inference_max_concurrency
+                or default_inference_max_concurrency(provider)
+            ),
+            detector_input_sizes=file_config.detection.input_sizes,
+            detector_nms_threshold=file_config.detection.nms_threshold,
+            detector_single_face_selection=file_config.detection.single_face_selection,
+            max_detected_faces=file_config.max_detected_faces,
+            config_file=config_file,
+            default_search_profile=_choice(
+                "INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE",
+                "fp32_v1",
+                SUPPORTED_SEARCH_PROFILES,
+            ),
+            default_search_capacity_rows=default_capacity_rows,
+            max_search_capacity_rows=max_capacity_rows,
+            default_max_faces_per_person=_int(
+                "INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON", 20
+            ),
+            default_search_load_policy=_choice(
+                "INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY",
+                "lazy",
+                SUPPORTED_SEARCH_LOAD_POLICIES,
+            ),
+            search_backend=_choice(
+                "INSIGHTFACE_SEARCH_BACKEND", "auto", SUPPORTED_SEARCH_BACKENDS
+            ),
+            search_library_path=(
+                Path(search_library).expanduser() if search_library else None
+            ),
+            search_device_id=_int(
+                "INSIGHTFACE_SEARCH_DEVICE_ID", device_id, minimum=0
+            ),
+            search_topk_mode=_choice(
+                "INSIGHTFACE_SEARCH_TOPK_MODE", "auto", SUPPORTED_SEARCH_TOPK_MODES
+            ),
+            search_build_batch_rows=_int(
+                "INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS", 4096
+            ),
+            web_ui_disabled=file_config.web_ui_disabled,
+            rtsp_max_streams=_int("INSIGHTFACE_RTSP_MAX_STREAMS", 4),
+            rtsp_preview_fps=_float(
+                "INSIGHTFACE_RTSP_PREVIEW_FPS", 5.0, minimum=1.0, maximum=30.0
+            ),
+            rtsp_jpeg_quality=rtsp_jpeg_quality,
+            rtsp_open_timeout_seconds=_float(
+                "INSIGHTFACE_RTSP_OPEN_TIMEOUT_SECONDS",
+                5.0,
+                minimum=0.5,
+                maximum=60.0,
+            ),
+            rtsp_read_timeout_seconds=_float(
+                "INSIGHTFACE_RTSP_READ_TIMEOUT_SECONDS",
+                5.0,
+                minimum=0.5,
+                maximum=60.0,
+            ),
+            rtsp_reconnect_delay_seconds=_float(
+                "INSIGHTFACE_RTSP_RECONNECT_DELAY_SECONDS",
+                1.0,
+                minimum=0.1,
+                maximum=30.0,
+            ),
+        )
diff --git a/server/backend/insightface_server/errors.py b/server/backend/insightface_server/errors.py
new file mode 100644
index 000000000..765352f0d
--- /dev/null
+++ b/server/backend/insightface_server/errors.py
@@ -0,0 +1,45 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass(slots=True)
+class ApiError(Exception):
+    code: str
+    message: str
+    status_code: int = 400
+    details: dict[str, Any] = field(default_factory=dict)
+
+    def __str__(self) -> str:
+        return self.message
+
+    def body(self, request_id: str) -> dict[str, Any]:
+        return {
+            "error": {
+                "code": self.code,
+                "message": self.message,
+                "details": self.details,
+            },
+            "request_id": request_id,
+        }
+
+
+def bad_request(message: str, *, code: str = "invalid_request", **details: Any) -> ApiError:
+    return ApiError(code, message, 400, details)
+
+
+def not_found(resource: str, identifier: str) -> ApiError:
+    return ApiError(
+        f"{resource}_not_found",
+        f"{resource.replace('_', ' ').title()} '{identifier}' was not found.",
+        404,
+    )
+
+
+def conflict(message: str, *, code: str = "resource_conflict") -> ApiError:
+    return ApiError(code, message, 409)
+
+
+def unprocessable(message: str, *, code: str, **details: Any) -> ApiError:
+    return ApiError(code, message, 422, details)
diff --git a/server/backend/insightface_server/inference/__init__.py b/server/backend/insightface_server/inference/__init__.py
new file mode 100644
index 000000000..6df391334
--- /dev/null
+++ b/server/backend/insightface_server/inference/__init__.py
@@ -0,0 +1,22 @@
+from .base import (
+    EngineSummary,
+    FaceObservation,
+    InferenceEngine,
+    cosine_similarity,
+    l2_normalize,
+    raw_cosine_similarity,
+)
+from .factory import create_engine
+from .quality import RegistrationQualityPolicy, registration_rejection_reasons
+
+__all__ = [
+    "EngineSummary",
+    "FaceObservation",
+    "InferenceEngine",
+    "RegistrationQualityPolicy",
+    "cosine_similarity",
+    "create_engine",
+    "l2_normalize",
+    "raw_cosine_similarity",
+    "registration_rejection_reasons",
+]
diff --git a/server/backend/insightface_server/inference/base.py b/server/backend/insightface_server/inference/base.py
new file mode 100644
index 000000000..66b74eaf3
--- /dev/null
+++ b/server/backend/insightface_server/inference/base.py
@@ -0,0 +1,157 @@
+from __future__ import annotations
+
+import os
+import platform
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Protocol, runtime_checkable
+
+import numpy as np
+
+from ..config import DetectionProfile
+
+
+@dataclass(frozen=True, slots=True)
+class EngineSummary:
+    model_id: str
+    model_version: str
+    model_digest: str
+    embedding_dimension: int
+    preprocessing_version: str
+    provider: str
+    models: tuple[dict[str, object], ...] = ()
+    license: dict[str, object] | None = None
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "model_id": self.model_id,
+            "model_version": self.model_version,
+            "model_digest": self.model_digest,
+            "embedding_dimension": self.embedding_dimension,
+            "preprocessing_version": self.preprocessing_version,
+            "provider": self.provider,
+            "models": [dict(model) for model in self.models],
+            "license": dict(self.license) if self.license is not None else None,
+        }
+
+
+@dataclass(slots=True)
+class FaceObservation:
+    """Internal face result. Coordinates are pixel-space ``(x1, y1, x2, y2)``."""
+
+    bbox: tuple[float, float, float, float]
+    detection_score: float
+    landmarks: np.ndarray | None
+    embedding: np.ndarray | None
+    quality_score: float = 0.0
+    sharpness: float = 0.0
+    brightness: float = 0.0
+    yaw: float = 0.0
+    pitch: float = 0.0
+    roll: float = 0.0
+    rejection_reasons: list[str] = field(default_factory=list)
+
+    @property
+    def confidence(self) -> float:
+        """Compatibility alias; new API responses call this ``detection_score``."""
+
+        return self.detection_score
+
+    @property
+    def area(self) -> float:
+        return max(0.0, self.bbox[2] - self.bbox[0]) * max(
+            0.0, self.bbox[3] - self.bbox[1]
+        )
+
+    @property
+    def bbox_xywh(self) -> tuple[float, float, float, float]:
+        left, top, right, bottom = self.bbox
+        return left, top, max(0.0, right - left), max(0.0, bottom - top)
+
+    def normalized_bbox(
+        self, image_width: int, image_height: int
+    ) -> tuple[float, float, float, float]:
+        if image_width <= 0 or image_height <= 0:
+            raise ValueError("Image dimensions must be positive")
+        left, top, right, bottom = self.bbox
+        clipped_left = float(np.clip(left, 0.0, image_width))
+        clipped_top = float(np.clip(top, 0.0, image_height))
+        clipped_right = float(np.clip(right, clipped_left, image_width))
+        clipped_bottom = float(np.clip(bottom, clipped_top, image_height))
+        return (
+            clipped_left / image_width,
+            clipped_top / image_height,
+            (clipped_right - clipped_left) / image_width,
+            (clipped_bottom - clipped_top) / image_height,
+        )
+
+
+@runtime_checkable
+class InferenceEngine(Protocol):
+    summary: EngineSummary
+
+    def startup(self) -> None: ...
+
+    def analyze(
+        self,
+        image: np.ndarray,
+        *,
+        require_embeddings: bool = True,
+        max_faces: int | None = None,
+        min_score: float | None = None,
+        detection_profile: DetectionProfile | None = None,
+    ) -> list[FaceObservation]: ...
+
+    def validate_detection_profile(self, profile: DetectionProfile) -> None: ...
+
+    def runtime_summary(self) -> dict[str, object]: ...
+
+    def close(self) -> None: ...
+
+
+def validate_image(image: np.ndarray) -> None:
+    if not isinstance(image, np.ndarray):
+        raise TypeError("image must be a numpy array")
+    if image.ndim != 3 or image.shape[2] != 3 or image.size == 0:
+        raise ValueError("image must be a non-empty BGR array with three channels")
+    if image.dtype != np.uint8:
+        raise ValueError("image must use uint8 pixels")
+
+
+def cpu_summary() -> dict[str, object]:
+    model = platform.processor().strip()
+    cpuinfo = Path("/proc/cpuinfo")
+    if not model and cpuinfo.is_file():
+        try:
+            for line in cpuinfo.read_text(encoding="utf-8", errors="replace").splitlines():
+                if line.lower().startswith("model name") and ":" in line:
+                    model = line.split(":", 1)[1].strip()
+                    break
+        except OSError:
+            pass
+    return {"model": model or "unknown", "logical_cores": os.cpu_count()}
+
+
+def l2_normalize(value: np.ndarray) -> np.ndarray:
+    vector = np.asarray(value, dtype=np.float32).reshape(-1)
+    norm = float(np.linalg.norm(vector))
+    if not np.isfinite(norm) or norm <= 1e-12:
+        raise ValueError("Model returned a zero or non-finite embedding")
+    result = vector / norm
+    if not np.all(np.isfinite(result)):
+        raise ValueError("Model returned a non-finite embedding")
+    return result.astype(np.float32, copy=False)
+
+
+def raw_cosine_similarity(left: np.ndarray, right: np.ndarray) -> float:
+    left_normalized = l2_normalize(left)
+    right_normalized = l2_normalize(right)
+    if left_normalized.shape != right_normalized.shape:
+        raise ValueError("Embedding dimensions do not match")
+    return float(np.clip(np.dot(left_normalized, right_normalized), -1.0, 1.0))
+
+
+def cosine_similarity(left: np.ndarray, right: np.ndarray) -> float:
+    """Return the raw cosine similarity in the inclusive ``[-1, 1]`` range."""
+
+    return raw_cosine_similarity(left, right)
diff --git a/server/backend/insightface_server/inference/concurrency.py b/server/backend/insightface_server/inference/concurrency.py
new file mode 100644
index 000000000..bfbfc05d5
--- /dev/null
+++ b/server/backend/insightface_server/inference/concurrency.py
@@ -0,0 +1,72 @@
+from __future__ import annotations
+
+import threading
+from collections.abc import Iterator
+from contextlib import contextmanager
+
+from ..request_context import check_request_deadline, remaining_seconds
+
+
+class InferenceConcurrencyLimiter:
+    """Bound all process-wide model work while allowing queued requests.
+
+    The limiter is deliberately owned by the inference engine, so HTTP calls,
+    enrollment work and RTSP frames all consume the same finite model budget.
+    Closing rejects queued work and waits for active Session.run calls to drain
+    before the model objects are released.
+    """
+
+    def __init__(self, max_concurrency: int) -> None:
+        if isinstance(max_concurrency, bool) or not isinstance(max_concurrency, int):
+            raise ValueError("inference max_concurrency must be an integer")
+        if max_concurrency <= 0:
+            raise ValueError("inference max_concurrency must be positive")
+        self.max_concurrency = max_concurrency
+        self._condition = threading.Condition()
+        self._active = 0
+        self._waiting = 0
+        self._peak_active = 0
+        self._closed = False
+
+    @contextmanager
+    def slot(self) -> Iterator[None]:
+        with self._condition:
+            if self._closed:
+                raise RuntimeError("Inference engine is closed")
+            self._waiting += 1
+            try:
+                while self._active >= self.max_concurrency:
+                    check_request_deadline()
+                    remaining = remaining_seconds()
+                    timeout = None if remaining is None else min(0.1, max(0.001, remaining))
+                    self._condition.wait(timeout=timeout)
+                    if self._closed:
+                        raise RuntimeError("Inference engine is closed")
+                check_request_deadline()
+                self._active += 1
+                self._peak_active = max(self._peak_active, self._active)
+            finally:
+                self._waiting -= 1
+
+        try:
+            yield
+        finally:
+            with self._condition:
+                self._active -= 1
+                self._condition.notify_all()
+
+    def close_and_wait(self) -> None:
+        with self._condition:
+            self._closed = True
+            self._condition.notify_all()
+            while self._active:
+                self._condition.wait()
+
+    def summary(self) -> dict[str, int]:
+        with self._condition:
+            return {
+                "max_concurrency": self.max_concurrency,
+                "active": self._active,
+                "waiting": self._waiting,
+                "peak_active": self._peak_active,
+            }
diff --git a/server/backend/insightface_server/inference/factory.py b/server/backend/insightface_server/inference/factory.py
new file mode 100644
index 000000000..7e01952b8
--- /dev/null
+++ b/server/backend/insightface_server/inference/factory.py
@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+from typing import Any
+
+from ..config import (
+    DEFAULT_DETECTOR_INPUT_SIZES,
+    DetectionProfile,
+    default_inference_max_concurrency,
+)
+from .base import InferenceEngine
+from .mock import MockInferenceEngine
+
+
+def _setting(settings: object, name: str, default: Any = None) -> Any:
+    if isinstance(settings, dict):
+        return settings.get(name, default)
+    return getattr(settings, name, default)
+
+
+def create_engine(settings: object) -> InferenceEngine:
+    """Create, but do not start, the configured process-wide inference engine."""
+
+    mode = str(_setting(settings, "inference_mode", "onnx")).lower()
+    provider = str(_setting(settings, "execution_provider", "CPUExecutionProvider"))
+    configured_concurrency = _setting(settings, "inference_max_concurrency", None)
+    max_concurrency = (
+        default_inference_max_concurrency(provider)
+        if configured_concurrency is None
+        else int(configured_concurrency)
+    )
+    if mode == "mock":
+        default_profile = getattr(settings, "detection_profile", None)
+        if not isinstance(default_profile, DetectionProfile):
+            default_profile = None
+        return MockInferenceEngine(
+            provider=provider,
+            embedding_dimension=int(_setting(settings, "embedding_dimension", 512)),
+            detector_input_sizes=_setting(
+                settings, "detector_input_sizes", DEFAULT_DETECTOR_INPUT_SIZES
+            ),
+            detection_profile=default_profile,
+            max_concurrency=max_concurrency,
+        )
+    if mode != "onnx":
+        raise ValueError(f"Unsupported inference_mode: {mode}")
+
+    from .onnx_engine import OnnxInsightFaceEngine
+
+    return OnnxInsightFaceEngine(settings)
diff --git a/server/backend/insightface_server/inference/mock.py b/server/backend/insightface_server/inference/mock.py
new file mode 100644
index 000000000..bc1242297
--- /dev/null
+++ b/server/backend/insightface_server/inference/mock.py
@@ -0,0 +1,178 @@
+from __future__ import annotations
+
+import hashlib
+import platform
+import threading
+
+import cv2
+import numpy as np
+
+from ..config import (
+    DEFAULT_DETECTOR_INPUT_SIZES,
+    DetectionProfile,
+    default_inference_max_concurrency,
+    normalize_detector_input_sizes,
+    normalize_single_face_selection,
+)
+from .base import EngineSummary, FaceObservation, cpu_summary, l2_normalize, validate_image
+from .concurrency import InferenceConcurrencyLimiter
+from .quality import enrich_quality
+
+
+class MockInferenceEngine:
+    """Deterministic synthetic engine for public CI and explicit development mode."""
+
+    def __init__(
+        self,
+        provider: str = "CPUExecutionProvider",
+        embedding_dimension: int = 512,
+        detector_input_sizes: object = DEFAULT_DETECTOR_INPUT_SIZES,
+        detection_profile: DetectionProfile | None = None,
+        max_concurrency: int | None = None,
+    ) -> None:
+        self.summary = EngineSummary(
+            model_id="synthetic-recognition",
+            model_version="1",
+            model_digest=hashlib.sha256(b"simple-insightface-server-mock-v1").hexdigest(),
+            embedding_dimension=embedding_dimension,
+            preprocessing_version="mock-1",
+            provider=provider,
+            models=(
+                {
+                    "model_id": "synthetic-detection",
+                    "model_version": "1",
+                    "task": "face_detection",
+                    "sha256": hashlib.sha256(b"synthetic-detection").hexdigest(),
+                },
+                {
+                    "model_id": "synthetic-recognition",
+                    "model_version": "1",
+                    "task": "face_recognition",
+                    "sha256": hashlib.sha256(b"synthetic-recognition").hexdigest(),
+                },
+            ),
+        )
+        self._embedding_dimension = embedding_dimension
+        self._default_detection_profile = detection_profile or DetectionProfile(
+            input_sizes=normalize_detector_input_sizes(detector_input_sizes)
+        )
+        self._detector_input_sizes = self._default_detection_profile.input_sizes
+        self._started = False
+        self._closed = False
+        self._lifecycle_lock = threading.RLock()
+        self._concurrency = InferenceConcurrencyLimiter(
+            max_concurrency or default_inference_max_concurrency(provider)
+        )
+
+    def startup(self) -> None:
+        with self._lifecycle_lock:
+            if self._closed:
+                raise RuntimeError("Inference engine is closed")
+            self._started = True
+
+    def analyze(
+        self,
+        image: np.ndarray,
+        *,
+        require_embeddings: bool = True,
+        max_faces: int | None = None,
+        min_score: float | None = None,
+        detection_profile: DetectionProfile | None = None,
+    ) -> list[FaceObservation]:
+        validate_image(image)
+        if max_faces is not None and max_faces <= 0:
+            raise ValueError("max_faces must be positive")
+        if min_score is not None and not 0.0 <= min_score <= 1.0:
+            raise ValueError("min_score must be between 0 and 1")
+        with self._concurrency.slot():
+            with self._lifecycle_lock:
+                if not self._started or self._closed:
+                    raise RuntimeError("Inference engine has not been started")
+            gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+            if min(image.shape[:2]) < 16 or float(np.std(gray)) < 1.0:
+                return []
+
+            height, width = image.shape[:2]
+            face_count = 2 if width >= height * 1.8 else 1
+            observations = [
+                self._observation(image, index=index, count=face_count, embeddings=require_embeddings)
+                for index in range(face_count)
+            ]
+            profile = detection_profile or self._default_detection_profile
+            self.validate_detection_profile(profile)
+            threshold = profile.threshold if min_score is None else max(profile.threshold, min_score)
+            observations = [face for face in observations if face.detection_score >= threshold]
+            observations.sort(key=lambda face: face.area, reverse=True)
+            return observations if max_faces is None else observations[:max_faces]
+
+    def validate_detection_profile(self, profile: DetectionProfile) -> None:
+        normalize_detector_input_sizes(profile.input_sizes)
+        if not 0.0 <= profile.threshold <= 1.0:
+            raise ValueError("detection threshold must be between 0 and 1")
+        if not 0.0 <= profile.nms_threshold <= 1.0:
+            raise ValueError("detection nms_threshold must be between 0 and 1")
+        normalize_single_face_selection(profile.single_face_selection)
+
+    def _observation(
+        self, image: np.ndarray, *, index: int, count: int, embeddings: bool
+    ) -> FaceObservation:
+        height, width = image.shape[:2]
+        side = float(min(height, width) * (0.62 if index == 0 else 0.46))
+        center_x = width * ((index + 1) / (count + 1))
+        center_y = height / 2.0
+        left, top = center_x - side / 2.0, center_y - side / 2.0
+        right, bottom = center_x + side / 2.0, center_y + side / 2.0
+        landmarks = np.asarray(
+            [
+                [center_x - side * 0.18, center_y - side * 0.12],
+                [center_x + side * 0.18, center_y - side * 0.12],
+                [center_x, center_y],
+                [center_x - side * 0.13, center_y + side * 0.20],
+                [center_x + side * 0.13, center_y + side * 0.20],
+            ],
+            dtype=np.float32,
+        )
+        embedding = None
+        if embeddings:
+            resized = cv2.resize(image, (32, 32), interpolation=cv2.INTER_AREA)
+            seed = hashlib.sha256(resized.tobytes() + index.to_bytes(2, "big")).digest()
+            generator = np.random.default_rng(int.from_bytes(seed[:8], "big"))
+            embedding = l2_normalize(
+                generator.standard_normal(self._embedding_dimension, dtype=np.float32)
+            )
+        face = FaceObservation(
+            bbox=(left, top, right, bottom),
+            detection_score=0.99 - index * 0.02,
+            landmarks=landmarks,
+            embedding=embedding,
+        )
+        enrich_quality(image, face)
+        return face
+
+    def runtime_summary(self) -> dict[str, object]:
+        with self._lifecycle_lock:
+            summary: dict[str, object] = {
+                "mode": "mock",
+                "os": platform.platform(),
+                "architecture": platform.machine(),
+                "cpu": cpu_summary(),
+                "execution_provider": self.summary.provider,
+                "detector_input_sizes": [list(size) for size in self._detector_input_sizes],
+                "detection_profile": self._default_detection_profile.as_dict(),
+                "strict_provider_audit": {},
+            }
+        summary["inference_concurrency"] = self._concurrency.summary()
+        return summary
+
+    def runtime_details(self) -> dict[str, object]:
+        """Compatibility alias for internal callers during the server transition."""
+
+        return self.runtime_summary()
+
+    def close(self) -> None:
+        with self._lifecycle_lock:
+            if self._closed:
+                return
+            self._closed = True
+            self._started = False
+        self._concurrency.close_and_wait()
diff --git a/server/backend/insightface_server/inference/onnx_engine.py b/server/backend/insightface_server/inference/onnx_engine.py
new file mode 100644
index 000000000..ce314a235
--- /dev/null
+++ b/server/backend/insightface_server/inference/onnx_engine.py
@@ -0,0 +1,600 @@
+from __future__ import annotations
+
+import copy
+import ctypes
+import hashlib
+import json
+import os
+import platform
+import subprocess
+import threading
+from collections import Counter
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+from ..config import (
+    DEFAULT_DETECTOR_INPUT_SIZES,
+    DEFAULT_DETECTOR_NMS_THRESHOLD,
+    DetectionProfile,
+    default_inference_max_concurrency,
+    normalize_detector_input_sizes,
+    normalize_single_face_selection,
+)
+from ..licensing import verify_model_license
+from ..models import ModelBundle, load_manifest
+from .base import EngineSummary, FaceObservation, cpu_summary, l2_normalize, validate_image
+from .concurrency import InferenceConcurrencyLimiter
+from .quality import enrich_quality
+
+_CPU_METADATA_OPERATORS = frozenset(
+    {"Cast", "Concat", "Gather", "Shape", "Size", "Slice", "Squeeze", "Unsqueeze"}
+)
+_CPU_METADATA_DTYPES = frozenset({"int32", "int64"})
+_MAX_CPU_METADATA_OUTPUT_BYTES = 4096
+_SUPPORTED_COMPUTE_CAPABILITIES = frozenset(
+    {(7, 5), (8, 0), (8, 6), (8, 9), (9, 0), (10, 0), (10, 3), (12, 0)}
+)
+_BLACKWELL_COMPUTE_CAPABILITIES = frozenset({(10, 0), (10, 3), (12, 0)})
+
+
+def _setting(settings: object, name: str, default: Any = None) -> Any:
+    if isinstance(settings, dict):
+        return settings.get(name, default)
+    return getattr(settings, name, default)
+
+
+def _integer_metadata_shapes(value: object) -> bool:
+    if not isinstance(value, list) or not value:
+        return False
+    for tensor in value:
+        if not isinstance(tensor, dict) or len(tensor) != 1:
+            return False
+        dtype, shape = next(iter(tensor.items()))
+        if dtype not in _CPU_METADATA_DTYPES or not isinstance(shape, list):
+            return False
+        if any(not isinstance(dimension, int) or dimension < 0 for dimension in shape):
+            return False
+    return True
+
+
+def audit_cuda_profile(events: object, *, model_name: str) -> dict[str, object]:
+    """Reject silent CPU model compute while permitting tiny shape bookkeeping nodes."""
+
+    if not isinstance(events, list):
+        raise RuntimeError(f"Invalid ONNX Runtime profile for {model_name}")
+    cuda_kernels = 0
+    cpu_metadata_operators: Counter[str] = Counter()
+    rejected: list[dict[str, object]] = []
+    for event in events:
+        if not isinstance(event, dict) or event.get("cat") != "Node":
+            continue
+        args = event.get("args")
+        if not isinstance(args, dict):
+            continue
+        provider = args.get("provider")
+        operator = args.get("op_name")
+        if provider == "CUDAExecutionProvider":
+            cuda_kernels += 1
+            continue
+        if not provider:
+            continue
+        if provider != "CPUExecutionProvider":
+            rejected.append(
+                {
+                    "name": event.get("name"),
+                    "operator": operator,
+                    "provider": provider,
+                }
+            )
+            continue
+        try:
+            output_bytes = int(args.get("output_size", -1))
+        except (TypeError, ValueError):
+            output_bytes = -1
+        metadata_only = (
+            operator in _CPU_METADATA_OPERATORS
+            and 0 <= output_bytes <= _MAX_CPU_METADATA_OUTPUT_BYTES
+            and _integer_metadata_shapes(args.get("output_type_shape"))
+        )
+        if metadata_only:
+            cpu_metadata_operators[str(operator)] += 1
+        else:
+            rejected.append(
+                {
+                    "name": event.get("name"),
+                    "operator": operator,
+                    "provider": provider,
+                    "output_bytes": output_bytes,
+                    "output_type_shape": args.get("output_type_shape"),
+                }
+            )
+    if cuda_kernels == 0:
+        raise RuntimeError(f"GPU strict audit found no CUDA kernels for {model_name}")
+    if rejected:
+        compact = json.dumps(rejected, separators=(",", ":"), sort_keys=True)
+        raise RuntimeError(
+            f"GPU strict audit rejected CPU/non-CUDA compute for {model_name}: {compact}"
+        )
+    return {
+        "accepted": True,
+        "cuda_kernel_count": cuda_kernels,
+        "cpu_shape_kernel_count": sum(cpu_metadata_operators.values()),
+        "cpu_shape_operators": dict(sorted(cpu_metadata_operators.items())),
+        "policy": "CUDA compute required; CPU limited to small integer shape metadata",
+    }
+
+
+def _cuda_runtime_version() -> str | None:
+    try:
+        library = ctypes.CDLL("libcudart.so.12")
+        value = ctypes.c_int()
+        if library.cudaRuntimeGetVersion(ctypes.byref(value)) != 0:
+            return None
+    except (OSError, AttributeError, ValueError):
+        return None
+    raw = int(value.value)
+    return f"{raw // 1000}.{(raw % 1000) // 10}"
+
+
+def _cudnn_version() -> str | None:
+    try:
+        library = ctypes.CDLL("libcudnn.so.9")
+        getter = library.cudnnGetVersion
+        getter.restype = ctypes.c_size_t
+        raw = int(getter())
+    except (OSError, AttributeError, ValueError):
+        return None
+    if raw <= 0:
+        return None
+    return f"{raw // 10000}.{(raw % 10000) // 100}.{raw % 100}"
+
+
+def _gpu_details() -> list[dict[str, str]]:
+    command = [
+        "nvidia-smi",
+        "--query-gpu=name,compute_cap,driver_version",
+        "--format=csv,noheader,nounits",
+    ]
+    try:
+        result = subprocess.run(
+            command,
+            check=True,
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except (FileNotFoundError, subprocess.SubprocessError):
+        return []
+    gpus: list[dict[str, str]] = []
+    for line in result.stdout.splitlines():
+        values = [item.strip() for item in line.split(",")]
+        if len(values) == 3:
+            gpus.append(
+                {
+                    "name": values[0],
+                    "compute_capability": values[1],
+                    "driver_version": values[2],
+                }
+            )
+    return gpus
+
+
+def _version_tuple(value: str) -> tuple[int, ...]:
+    parts: list[int] = []
+    for item in value.split("."):
+        digits = "".join(character for character in item if character.isdigit())
+        if not digits:
+            break
+        parts.append(int(digits))
+    return tuple(parts)
+
+
+def validate_gpu_requirements(gpus: list[dict[str, str]], cuda: str | None) -> None:
+    if not gpus:
+        raise RuntimeError("CUDA Session exists but NVIDIA GPU diagnostics failed")
+    for gpu in gpus:
+        capability = _version_tuple(gpu["compute_capability"])
+        driver = _version_tuple(gpu["driver_version"])
+        architecture = capability[:2]
+        if architecture not in _SUPPORTED_COMPUTE_CAPABILITIES:
+            raise RuntimeError(
+                f"Unsupported CUDA Compute Capability {gpu['compute_capability']} "
+                f"on {gpu['name']}"
+            )
+        if architecture in _BLACKWELL_COMPUTE_CAPABILITIES:
+            if driver < (570, 26):
+                raise RuntimeError(
+                    "Blackwell requires NVIDIA Driver 570.26 or newer; "
+                    f"found {gpu['driver_version']}"
+                )
+            if cuda is None or _version_tuple(cuda) < (12, 8):
+                raise RuntimeError(
+                    f"Blackwell requires CUDA Runtime 12.8 or newer; found {cuda}"
+                )
+        elif driver < (535,):
+            raise RuntimeError(
+                "Turing through Hopper require NVIDIA Driver R535 or newer; "
+                f"found {gpu['driver_version']}"
+            )
+
+
+class OnnxInsightFaceEngine:
+    """Shared SCRFD + ArcFace implementation with exactly one Session per model."""
+
+    def __init__(self, settings: object) -> None:
+        self._settings = settings
+        self._provider = str(
+            _setting(settings, "execution_provider", "CPUExecutionProvider")
+        )
+        if self._provider not in {"CPUExecutionProvider", "CUDAExecutionProvider"}:
+            raise ValueError(f"Unsupported execution_provider: {self._provider}")
+        models_dir = Path(str(_setting(settings, "models_dir", "/models")))
+        self.bundle: ModelBundle = load_manifest(models_dir)
+        model_license = verify_model_license(
+            self.bundle.license_path, expected_model_id=self.bundle.model_id
+        )
+        recognizer = self.bundle.recognizer
+        bundle_identity = [
+            {
+                "sha256": model.sha256,
+                "input_size": model.input_size,
+                "input_mean": model.input_mean,
+                "input_std": model.input_std,
+                "preprocessing_version": model.preprocessing_version,
+            }
+            for model in self.bundle.models
+        ]
+        bundle_digest = hashlib.sha256(
+            json.dumps(bundle_identity, separators=(",", ":"), sort_keys=True).encode()
+        ).hexdigest()
+        self.summary = EngineSummary(
+            model_id=self.bundle.model_id,
+            model_version=self.bundle.model_version,
+            # Collection semantics include detection/alignment as well as recognition.
+            model_digest=bundle_digest,
+            embedding_dimension=int(recognizer.embedding_dimension or 0),
+            preprocessing_version=recognizer.preprocessing_version,
+            provider=self._provider,
+            models=tuple(model.public_summary() for model in self.bundle.models),
+            license=model_license.public_summary(),
+        )
+        self._device_id = int(_setting(settings, "device_id", 0))
+        self._detector_threshold = float(_setting(settings, "detector_threshold", 0.50))
+        self._detector_input_sizes = normalize_detector_input_sizes(
+            _setting(settings, "detector_input_sizes", DEFAULT_DETECTOR_INPUT_SIZES)
+        )
+        self._detector_nms_threshold = float(
+            _setting(settings, "detector_nms_threshold", DEFAULT_DETECTOR_NMS_THRESHOLD)
+        )
+        if not 0.0 <= self._detector_threshold <= 1.0:
+            raise ValueError("detector_threshold must be between 0 and 1")
+        if not 0.0 <= self._detector_nms_threshold <= 1.0:
+            raise ValueError("detector_nms_threshold must be between 0 and 1")
+        self._default_detection_profile = DetectionProfile.from_mapping(
+            {
+                "input_sizes": self._detector_input_sizes,
+                "threshold": self._detector_threshold,
+                "nms_threshold": self._detector_nms_threshold,
+                "single_face_selection": _setting(
+                    settings, "detector_single_face_selection", "largest"
+                ),
+            }
+        )
+        configured_concurrency = _setting(settings, "inference_max_concurrency", None)
+        self._max_concurrency = (
+            default_inference_max_concurrency(self._provider)
+            if configured_concurrency is None
+            else int(configured_concurrency)
+        )
+        self._concurrency = InferenceConcurrencyLimiter(self._max_concurrency)
+        self._lifecycle_lock = threading.RLock()
+        self._startup_attempted = False
+        self._started = False
+        self._closed = False
+        self._startup_error: BaseException | None = None
+        self._detector: Any = None
+        self._recognizer: Any = None
+        self._detector_session: Any = None
+        self._recognizer_session: Any = None
+        self._runtime: dict[str, object] = {}
+
+    def _new_session(self, path: Path) -> Any:
+        import onnxruntime as ort
+
+        options = ort.SessionOptions()
+        options.log_severity_level = 3
+        options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+        if self._provider == "CUDAExecutionProvider":
+            options.enable_profiling = True
+            profile_id = hashlib.sha256(str(path).encode()).hexdigest()[:12]
+            options.profile_file_prefix = f"/tmp/insightface-ort-{profile_id}"
+            providers: list[Any] = [
+                (
+                    "CUDAExecutionProvider",
+                    {
+                        "device_id": self._device_id,
+                        "do_copy_in_default_stream": True,
+                    },
+                ),
+                # Dynamic SCRFD graphs contain a few shape-only nodes which ORT assigns
+                # to CPU. The warm-up profile below audits every such assignment.
+                "CPUExecutionProvider",
+            ]
+        else:
+            providers = ["CPUExecutionProvider"]
+        session = ort.InferenceSession(
+            str(path), sess_options=options, providers=providers
+        )
+        actual = session.get_providers()
+        if not actual or actual[0] != self._provider:
+            raise RuntimeError(
+                f"Required primary {self._provider} is unavailable; "
+                f"Session providers are {actual}"
+            )
+        return session
+
+    @staticmethod
+    def _finish_cuda_profile(session: Any, *, model_name: str) -> dict[str, object]:
+        profile_value = session.end_profiling()
+        if not profile_value:
+            raise RuntimeError(f"ONNX Runtime did not produce a profile for {model_name}")
+        profile_path = Path(profile_value)
+        try:
+            events = json.loads(profile_path.read_text(encoding="utf-8"))
+            return audit_cuda_profile(events, model_name=model_name)
+        except (OSError, json.JSONDecodeError) as exc:
+            raise RuntimeError(
+                f"Unable to audit ONNX Runtime profile for {model_name}"
+            ) from exc
+        finally:
+            profile_path.unlink(missing_ok=True)
+
+    def startup(self) -> None:
+        with self._lifecycle_lock:
+            if self._closed:
+                raise RuntimeError("Inference engine is closed")
+            if self._started:
+                return
+            if self._startup_attempted:
+                message = "Previous inference engine startup failed"
+                raise RuntimeError(message) from self._startup_error
+            self._startup_attempted = True
+            try:
+                self._startup_once()
+            except BaseException as exc:
+                self._startup_error = exc
+                self._detector = None
+                self._recognizer = None
+                self._detector_session = None
+                self._recognizer_session = None
+                raise
+            self._started = True
+
+    def _startup_once(self) -> None:
+        import onnxruntime as ort
+        from insightface.model_zoo.arcface_onnx import ArcFaceONNX
+        from insightface.model_zoo.scrfd import SCRFD
+
+        available = list(ort.get_available_providers())
+        if self._provider not in available:
+            raise RuntimeError(
+                f"Required {self._provider} is unavailable; ONNX Runtime reports {available}"
+            )
+
+        # These are the only two Session constructions in the engine lifecycle.
+        self._detector_session = self._new_session(self.bundle.detector.path)
+        self._recognizer_session = self._new_session(self.bundle.recognizer.path)
+        self._detector = SCRFD(
+            model_file=str(self.bundle.detector.path), session=self._detector_session
+        )
+        self._detector.input_mean = self.bundle.detector.input_mean
+        self._detector.input_std = self.bundle.detector.input_std
+        self._assert_detector_sizes_supported(self._detector_input_sizes)
+        self._detector.prepare(
+            self._device_id if self._provider == "CUDAExecutionProvider" else -1,
+            input_size=self._detector_input_sizes,
+            det_thresh=self._detector_threshold,
+            nms_thresh=self._detector_nms_threshold,
+        )
+        self._recognizer = ArcFaceONNX(
+            model_file=str(self.bundle.recognizer.path),
+            session=self._recognizer_session,
+        )
+        self._recognizer.input_mean = self.bundle.recognizer.input_mean
+        self._recognizer.input_std = self.bundle.recognizer.input_std
+
+        self._warm_up()
+        provider_audit: dict[str, object] = {}
+        if self._provider == "CUDAExecutionProvider":
+            provider_audit = {
+                "detector": self._finish_cuda_profile(
+                    self._detector_session, model_name="detector"
+                ),
+                "recognizer": self._finish_cuda_profile(
+                    self._recognizer_session, model_name="recognizer"
+                ),
+            }
+
+        cuda = _cuda_runtime_version()
+        cudnn = _cudnn_version()
+        gpus = _gpu_details()
+        if self._provider == "CUDAExecutionProvider":
+            if cuda is None or cudnn is None:
+                raise RuntimeError(
+                    "CUDA provider initialized but actual CUDA/cuDNN runtime versions "
+                    f"could not be read (cuda={cuda}, cudnn={cudnn})"
+                )
+            validate_gpu_requirements(gpus, cuda)
+
+        self._runtime = {
+            "mode": "onnx",
+            "os": platform.platform(),
+            "architecture": platform.machine(),
+            "cpu": cpu_summary(),
+            "onnx_runtime_version": ort.__version__,
+            "available_execution_providers": available,
+            "execution_provider": self._provider,
+            "detector_input_sizes": [list(size) for size in self._detector_input_sizes],
+            "detector_session_providers": self._detector_session.get_providers(),
+            "recognizer_session_providers": self._recognizer_session.get_providers(),
+            "strict_provider_audit": provider_audit,
+            "cuda_runtime_version": cuda,
+            "cuda_image_version": os.getenv("CUDA_VERSION"),
+            "cudnn_version": cudnn,
+            "gpus": gpus,
+            "models": [dict(model) for model in self.summary.models],
+        }
+
+    def _warm_up(self) -> None:
+        self._warm_up_detector_sizes(self._detector_input_sizes)
+
+        recognizer_input = self._recognizer_session.get_inputs()[0]
+        width, height = self.bundle.recognizer.input_size
+        recognition_data = np.zeros((1, 3, height, width), dtype=np.float32)
+        recognition_outputs = self._recognizer_session.run(
+            None, {recognizer_input.name: recognition_data}
+        )
+        if not recognition_outputs:
+            raise RuntimeError("Recognition model warm-up returned no outputs")
+        actual_dimension = np.asarray(recognition_outputs[0]).reshape(1, -1).shape[1]
+        if actual_dimension != self.summary.embedding_dimension:
+            raise RuntimeError(
+                "Recognition warm-up output dimension does not match manifest: "
+                f"expected {self.summary.embedding_dimension}, got {actual_dimension}"
+            )
+
+    def _assert_detector_sizes_supported(
+        self, sizes: tuple[tuple[int, int], ...]
+    ) -> None:
+        static_size = getattr(self._detector, "static_input_size", None)
+        if static_size is not None and sizes != (tuple(static_size),):
+            raise ValueError(
+                "The detection model has a static input and only supports "
+                f"{tuple(static_size)}"
+            )
+
+    def _warm_up_detector_sizes(self, sizes: tuple[tuple[int, int], ...]) -> None:
+        detector_input = self._detector_session.get_inputs()[0]
+        for detector_width, detector_height in sizes:
+            detector_data = np.zeros(
+                (1, 3, detector_height, detector_width), dtype=np.float32
+            )
+            detector_outputs = self._detector_session.run(
+                None, {detector_input.name: detector_data}
+            )
+            if not detector_outputs:
+                raise RuntimeError(
+                    "Detection model warm-up returned no outputs for input size "
+                    f"{detector_width}x{detector_height}"
+                )
+
+    def analyze(
+        self,
+        image: np.ndarray,
+        *,
+        require_embeddings: bool = True,
+        max_faces: int | None = None,
+        min_score: float | None = None,
+        detection_profile: DetectionProfile | None = None,
+    ) -> list[FaceObservation]:
+        from insightface.utils import face_align
+
+        validate_image(image)
+        if max_faces is not None and max_faces <= 0:
+            raise ValueError("max_faces must be positive")
+        if min_score is not None and not 0.0 <= min_score <= 1.0:
+            raise ValueError("min_score must be between 0 and 1")
+
+        with self._concurrency.slot():
+            with self._lifecycle_lock:
+                if not self._started or self._closed:
+                    raise RuntimeError("Inference engine has not been started")
+                detector = self._detector
+                recognizer = self._recognizer
+                profile = detection_profile or self._default_detection_profile
+                self.validate_detection_profile(profile)
+
+            # SCRFD's policy and anchor cache live on its Python wrapper. Each
+            # request gets independent mutable wrapper state while all copies
+            # share the one process-wide, thread-safe ORT Session.
+            request_detector = copy.copy(detector)
+            request_detector.center_cache = dict(detector.center_cache)
+            request_detector.det_thresh = profile.threshold
+            request_detector.nms_thresh = profile.nms_threshold
+            boxes, keypoints = request_detector.detect(
+                image, input_size=profile.input_sizes, max_num=0
+            )
+            image_height, image_width = image.shape[:2]
+            observations: list[FaceObservation] = []
+            threshold = profile.threshold if min_score is None else max(profile.threshold, min_score)
+            for index, row in enumerate(boxes):
+                detection_score = float(np.clip(row[4], 0.0, 1.0))
+                if detection_score < threshold:
+                    continue
+                left = float(np.clip(row[0], 0, image_width))
+                top = float(np.clip(row[1], 0, image_height))
+                right = float(np.clip(row[2], 0, image_width))
+                bottom = float(np.clip(row[3], 0, image_height))
+                landmarks = (
+                    None if keypoints is None else np.asarray(keypoints[index], dtype=np.float32)
+                )
+                embedding = None
+                if require_embeddings:
+                    if landmarks is None or landmarks.shape != (5, 2):
+                        continue
+                    aligned = face_align.norm_crop(
+                        image,
+                        landmark=landmarks,
+                        image_size=self.bundle.recognizer.input_size[0],
+                    )
+                    embedding = l2_normalize(recognizer.get_feat(aligned)[0])
+                observation = FaceObservation(
+                    bbox=(left, top, right, bottom),
+                    detection_score=detection_score,
+                    landmarks=landmarks,
+                    embedding=embedding,
+                )
+                enrich_quality(image, observation)
+                observations.append(observation)
+
+            observations.sort(key=lambda face: face.area, reverse=True)
+            return observations if max_faces is None else observations[:max_faces]
+
+    def validate_detection_profile(self, profile: DetectionProfile) -> None:
+        sizes = normalize_detector_input_sizes(profile.input_sizes)
+        if not 0.0 <= profile.threshold <= 1.0:
+            raise ValueError("detection threshold must be between 0 and 1")
+        if not 0.0 <= profile.nms_threshold <= 1.0:
+            raise ValueError("detection nms_threshold must be between 0 and 1")
+        normalize_single_face_selection(profile.single_face_selection)
+        self._assert_detector_sizes_supported(sizes)
+
+    def runtime_summary(self) -> dict[str, object]:
+        with self._lifecycle_lock:
+            summary = dict(self._runtime)
+            summary["detector_input_sizes"] = [
+                list(size) for size in self._detector_input_sizes
+            ]
+            summary["detection_profile"] = self._default_detection_profile.as_dict()
+        summary["inference_concurrency"] = self._concurrency.summary()
+        return summary
+
+    def runtime_details(self) -> dict[str, object]:
+        """Compatibility alias for internal callers during the server transition."""
+
+        return self.runtime_summary()
+
+    def close(self) -> None:
+        with self._lifecycle_lock:
+            if self._closed:
+                return
+            self._closed = True
+            self._started = False
+        self._concurrency.close_and_wait()
+        with self._lifecycle_lock:
+            self._detector = None
+            self._recognizer = None
+            self._detector_session = None
+            self._recognizer_session = None
diff --git a/server/backend/insightface_server/inference/quality.py b/server/backend/insightface_server/inference/quality.py
new file mode 100644
index 000000000..e997094db
--- /dev/null
+++ b/server/backend/insightface_server/inference/quality.py
@@ -0,0 +1,98 @@
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+
+import cv2
+import numpy as np
+
+from .base import FaceObservation
+
+
+@dataclass(frozen=True, slots=True)
+class RegistrationQualityPolicy:
+    min_face_size: int = 40
+    min_detection_score: float = 0.50
+    min_quality_score: float = 0.25
+    min_sharpness: float = 0.08
+    min_brightness: float = 0.06
+    max_brightness: float = 0.94
+    max_abs_pose_degrees: float = 45.0
+
+
+def enrich_quality(image: np.ndarray, face: FaceObservation) -> None:
+    """Populate deterministic, documented quality heuristics on a face result."""
+
+    image_height, image_width = image.shape[:2]
+    left, top, right, bottom = face.bbox
+    x1, y1 = max(0, int(math.floor(left))), max(0, int(math.floor(top)))
+    x2 = min(image_width, int(math.ceil(right)))
+    y2 = min(image_height, int(math.ceil(bottom)))
+    crop = image[y1:y2, x1:x2]
+    if crop.size:
+        gray = cv2.cvtColor(crop, cv2.COLOR_BGR2GRAY)
+        face.brightness = float(np.clip(np.mean(gray) / 255.0, 0.0, 1.0))
+        laplacian_variance = float(cv2.Laplacian(gray, cv2.CV_64F).var())
+        face.sharpness = float(np.clip(1.0 - math.exp(-laplacian_variance / 100.0), 0, 1))
+
+    if face.landmarks is not None and face.landmarks.shape == (5, 2):
+        points = face.landmarks.astype(np.float64)
+        eyes_midpoint = (points[0] + points[1]) / 2.0
+        mouth_midpoint = (points[3] + points[4]) / 2.0
+        eye_delta = points[1] - points[0]
+        eye_distance = max(float(np.linalg.norm(eye_delta)), 1e-6)
+        face.roll = float(np.degrees(np.arctan2(eye_delta[1], eye_delta[0])))
+        face.yaw = float((points[2, 0] - eyes_midpoint[0]) / eye_distance * 90.0)
+        vertical = max(float(mouth_midpoint[1] - eyes_midpoint[1]), 1e-6)
+        face.pitch = float(
+            ((points[2, 1] - eyes_midpoint[1]) / vertical - 0.45) * 90.0
+        )
+
+    face_width = max(0.0, right - left)
+    face_height = max(0.0, bottom - top)
+    size_score = float(np.clip(min(face_width, face_height) / 80.0, 0.0, 1.0))
+    brightness_score = float(np.clip(1.0 - abs(face.brightness - 0.5) / 0.5, 0.0, 1.0))
+    pose_score = float(
+        np.clip(
+            1.0 - max(abs(face.yaw), abs(face.pitch), abs(face.roll)) / 90.0,
+            0.0,
+            1.0,
+        )
+    )
+    face.quality_score = float(
+        np.clip(
+            0.35 * face.detection_score
+            + 0.25 * face.sharpness
+            + 0.15 * brightness_score
+            + 0.15 * size_score
+            + 0.10 * pose_score,
+            0.0,
+            1.0,
+        )
+    )
+
+
+def registration_rejection_reasons(
+    face: FaceObservation,
+    policy: RegistrationQualityPolicy | None = None,
+) -> list[str]:
+    """Return stable public reason codes for a candidate registration face."""
+
+    selected = policy or RegistrationQualityPolicy()
+    width = max(0.0, face.bbox[2] - face.bbox[0])
+    height = max(0.0, face.bbox[3] - face.bbox[1])
+    reasons: list[str] = []
+    if min(width, height) < selected.min_face_size:
+        reasons.append("face_too_small")
+    if face.detection_score < selected.min_detection_score:
+        reasons.append("low_detection_score")
+    if face.sharpness < selected.min_sharpness:
+        reasons.append("low_quality")
+    if not selected.min_brightness <= face.brightness <= selected.max_brightness:
+        if "low_quality" not in reasons:
+            reasons.append("low_quality")
+    if face.quality_score < selected.min_quality_score and "low_quality" not in reasons:
+        reasons.append("low_quality")
+    if max(abs(face.yaw), abs(face.pitch), abs(face.roll)) > selected.max_abs_pose_degrees:
+        reasons.append("extreme_pose")
+    return reasons
diff --git a/server/backend/insightface_server/licensing/__init__.py b/server/backend/insightface_server/licensing/__init__.py
new file mode 100644
index 000000000..a2a2e2165
--- /dev/null
+++ b/server/backend/insightface_server/licensing/__init__.py
@@ -0,0 +1,17 @@
+"""Verification for InsightFace-issued model licenses."""
+
+from .model_license import (
+    LICENSE_FILENAME,
+    ModelLicense,
+    ModelLicenseError,
+    canonical_license_bytes,
+    verify_model_license,
+)
+
+__all__ = [
+    "LICENSE_FILENAME",
+    "ModelLicense",
+    "ModelLicenseError",
+    "canonical_license_bytes",
+    "verify_model_license",
+]
diff --git a/server/backend/insightface_server/licensing/defaults/antelopev2/MODEL.LICENSE b/server/backend/insightface_server/licensing/defaults/antelopev2/MODEL.LICENSE
new file mode 100644
index 000000000..c88e6dd17
--- /dev/null
+++ b/server/backend/insightface_server/licensing/defaults/antelopev2/MODEL.LICENSE
@@ -0,0 +1,9 @@
+{
+  "license_version": 1,
+  "license_id": "antelopev2-public-v1",
+  "issuer": "InsightFace",
+  "model_id": "antelopev2",
+  "grant": "non-commercial",
+  "valid_from": "2021-09-22T00:00:00Z",
+  "signature": "d_3EK5bUdYWMXESwLMO1tUmUW_pXO3BmFlBQtq9gLx7ruHQA7SFi21JVfCjvecPiUf21SVeRNqpIP99m4OEZCg"
+}
diff --git a/server/backend/insightface_server/licensing/defaults/buffalo_l/MODEL.LICENSE b/server/backend/insightface_server/licensing/defaults/buffalo_l/MODEL.LICENSE
new file mode 100644
index 000000000..bc7411882
--- /dev/null
+++ b/server/backend/insightface_server/licensing/defaults/buffalo_l/MODEL.LICENSE
@@ -0,0 +1,9 @@
+{
+  "license_version": 1,
+  "license_id": "buffalo_l-public-v1",
+  "issuer": "InsightFace",
+  "model_id": "buffalo_l",
+  "grant": "non-commercial",
+  "valid_from": "2021-09-22T00:00:00Z",
+  "signature": "mkF_zjs_gw5lzWlN6DXlBWPY6ZK7diTGoSORdp33ISLegqxSrL930a--bHEGOOiUO4n-W9kX5qQc6kbEkGnUBQ"
+}
diff --git a/server/backend/insightface_server/licensing/defaults/buffalo_m/MODEL.LICENSE b/server/backend/insightface_server/licensing/defaults/buffalo_m/MODEL.LICENSE
new file mode 100644
index 000000000..4012e5aa9
--- /dev/null
+++ b/server/backend/insightface_server/licensing/defaults/buffalo_m/MODEL.LICENSE
@@ -0,0 +1,9 @@
+{
+  "license_version": 1,
+  "license_id": "buffalo_m-public-v1",
+  "issuer": "InsightFace",
+  "model_id": "buffalo_m",
+  "grant": "non-commercial",
+  "valid_from": "2021-09-22T00:00:00Z",
+  "signature": "_Tu01B-HHWeAWr8jwxvl1SmEaupMbUNWb9SfnyLUDa5x2se8PQg-BE6M-MM5em2VEglwgW_K5Que30CzYFbVDA"
+}
diff --git a/server/backend/insightface_server/licensing/defaults/buffalo_sc/MODEL.LICENSE b/server/backend/insightface_server/licensing/defaults/buffalo_sc/MODEL.LICENSE
new file mode 100644
index 000000000..404069103
--- /dev/null
+++ b/server/backend/insightface_server/licensing/defaults/buffalo_sc/MODEL.LICENSE
@@ -0,0 +1,9 @@
+{
+  "license_version": 1,
+  "license_id": "buffalo_sc-public-v1",
+  "issuer": "InsightFace",
+  "model_id": "buffalo_sc",
+  "grant": "non-commercial",
+  "valid_from": "2021-09-22T00:00:00Z",
+  "signature": "hyj_imJNDac01OTmlPYVfWkrrcQzx2MSMZXx0hu1Z1EniCN7dVXx-t3zKHOAtzvNdPqCioobExvdcF5w-I2nDw"
+}
diff --git a/server/backend/insightface_server/licensing/model_license.py b/server/backend/insightface_server/licensing/model_license.py
new file mode 100644
index 000000000..321d97d5e
--- /dev/null
+++ b/server/backend/insightface_server/licensing/model_license.py
@@ -0,0 +1,281 @@
+"""Small, offline Ed25519 model-license verifier.
+
+The license is a compliance credential for a logical ``model_id``.  It is
+deliberately not bound to an ONNX digest so an operator may create an FP16,
+INT8, optimized ONNX, or TensorRT derivative without asking for a new license.
+"""
+
+from __future__ import annotations
+
+import base64
+import binascii
+import json
+import re
+from collections.abc import Iterable
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+import rfc8785
+from cryptography.exceptions import InvalidSignature
+from cryptography.hazmat.primitives import serialization
+from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PublicKey
+
+LICENSE_FILENAME = "MODEL.LICENSE"
+LICENSE_VERSION = 1
+LICENSE_ISSUER = "InsightFace"
+GRANTS = frozenset({"non-commercial", "commercial"})
+MAX_LICENSE_BYTES = 64 * 1024
+_MODEL_ID = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$")
+_SIGNATURE = re.compile(r"^[A-Za-z0-9_-]+$")
+_REQUIRED_FIELDS = frozenset(
+    {
+        "license_version",
+        "license_id",
+        "issuer",
+        "model_id",
+        "grant",
+        "valid_from",
+        "signature",
+    }
+)
+_OPTIONAL_FIELDS = frozenset({"customer", "reference", "valid_until"})
+
+
+class ModelLicenseError(RuntimeError):
+    """A model license is missing, invalid, not active, or expired."""
+
+
+@dataclass(frozen=True, slots=True)
+class ModelLicense:
+    license_id: str
+    issuer: str
+    model_id: str
+    grant: str
+    valid_from: datetime
+    valid_until: datetime | None
+    customer: str | None = None
+    reference: str | None = None
+
+    @property
+    def commercial_use_permitted(self) -> bool:
+        return self.grant == "commercial"
+
+    def public_summary(self) -> dict[str, object]:
+        return {
+            "license_id": self.license_id,
+            "issuer": self.issuer,
+            "model_id": self.model_id,
+            "grant": self.grant,
+            "customer": self.customer,
+            "reference": self.reference,
+            "valid_from": _format_time(self.valid_from),
+            "valid_until": (
+                _format_time(self.valid_until) if self.valid_until is not None else None
+            ),
+            "signature_valid": True,
+            "commercial_use_permitted": self.commercial_use_permitted,
+        }
+
+
+def _format_time(value: datetime) -> str:
+    return value.astimezone(UTC).isoformat().replace("+00:00", "Z")
+
+
+def _reject_duplicate_keys(items: list[tuple[str, Any]]) -> dict[str, Any]:
+    result: dict[str, Any] = {}
+    for key, value in items:
+        if key in result:
+            raise ModelLicenseError(f"Duplicate field in model license: {key}")
+        result[key] = value
+    return result
+
+
+def _read_document(path: Path) -> dict[str, Any]:
+    try:
+        size = path.stat().st_size
+        if size <= 0 or size > MAX_LICENSE_BYTES:
+            raise ModelLicenseError(
+                f"Model license must be between 1 and {MAX_LICENSE_BYTES} bytes: {path}"
+            )
+        raw = json.loads(
+            path.read_text(encoding="utf-8"),
+            object_pairs_hook=_reject_duplicate_keys,
+            parse_constant=lambda value: (_ for _ in ()).throw(
+                ValueError(f"Non-standard JSON number: {value}")
+            ),
+        )
+    except ModelLicenseError:
+        raise
+    except (OSError, UnicodeError, json.JSONDecodeError, ValueError) as exc:
+        raise ModelLicenseError(f"Unable to read model license {path}: {exc}") from exc
+    if not isinstance(raw, dict):
+        raise ModelLicenseError("Model license root must be a JSON object")
+    return raw
+
+
+def _required_string(document: dict[str, Any], field: str, *, maximum: int = 256) -> str:
+    value = document.get(field)
+    if not isinstance(value, str) or not value.strip() or len(value) > maximum:
+        raise ModelLicenseError(
+            f"Model license field {field!r} must be a non-empty string of at most "
+            f"{maximum} characters"
+        )
+    return value.strip()
+
+
+def _optional_string(document: dict[str, Any], field: str) -> str | None:
+    if field not in document:
+        return None
+    return _required_string(document, field)
+
+
+def _parse_time(document: dict[str, Any], field: str, *, required: bool) -> datetime | None:
+    if field not in document:
+        if required:
+            raise ModelLicenseError(f"Model license is missing required field: {field}")
+        return None
+    value = _required_string(document, field, maximum=32)
+    if not value.endswith("Z"):
+        raise ModelLicenseError(f"Model license field {field!r} must use UTC and end in Z")
+    try:
+        parsed = datetime.fromisoformat(value[:-1] + "+00:00")
+    except ValueError as exc:
+        raise ModelLicenseError(f"Invalid UTC timestamp in {field!r}: {value}") from exc
+    if parsed.utcoffset() != UTC.utcoffset(parsed):
+        raise ModelLicenseError(f"Model license field {field!r} must use UTC")
+    return parsed
+
+
+def _signature_bytes(value: str) -> bytes:
+    if not _SIGNATURE.fullmatch(value):
+        raise ModelLicenseError("Model license signature must use unpadded base64url")
+    try:
+        padding = "=" * (-len(value) % 4)
+        decoded = base64.b64decode(value + padding, altchars=b"-_", validate=True)
+    except (binascii.Error, ValueError) as exc:
+        raise ModelLicenseError("Invalid model license signature encoding") from exc
+    if len(decoded) != 64:
+        raise ModelLicenseError("An Ed25519 model license signature must be 64 bytes")
+    return decoded
+
+
+def canonical_license_bytes(document: dict[str, Any]) -> bytes:
+    """Canonicalize all signed fields, excluding only ``signature``."""
+
+    unsigned = {key: value for key, value in document.items() if key != "signature"}
+    try:
+        return rfc8785.dumps(unsigned)
+    except rfc8785.CanonicalizationError as exc:
+        raise ModelLicenseError(f"Model license cannot be canonicalized: {exc}") from exc
+
+
+def _trusted_keys(directory: Path | None = None) -> tuple[Ed25519PublicKey, ...]:
+    key_dir = directory or Path(__file__).with_name("trusted_keys")
+    paths = sorted(
+        path for path in key_dir.glob("*.pem") if not path.name.startswith(".")
+    )
+    if not paths:
+        raise ModelLicenseError(f"No trusted InsightFace model-license keys found in {key_dir}")
+    keys: list[Ed25519PublicKey] = []
+    for path in paths:
+        try:
+            loaded = serialization.load_pem_public_key(path.read_bytes())
+        except (OSError, ValueError, TypeError) as exc:
+            raise ModelLicenseError(f"Invalid trusted public key {path.name}: {exc}") from exc
+        if not isinstance(loaded, Ed25519PublicKey):
+            raise ModelLicenseError(f"Trusted key {path.name} is not an Ed25519 public key")
+        keys.append(loaded)
+    return tuple(keys)
+
+
+def verify_model_license(
+    path: str | Path,
+    *,
+    expected_model_id: str,
+    now: datetime | None = None,
+    public_keys: Iterable[Ed25519PublicKey] | None = None,
+) -> ModelLicense:
+    """Verify signature, model scope, issuer, grant, and validity period."""
+
+    license_path = Path(path)
+    if not license_path.is_file():
+        raise ModelLicenseError(f"Required model license file is missing: {license_path}")
+    document = _read_document(license_path)
+    fields = frozenset(document)
+    missing = sorted(_REQUIRED_FIELDS - fields)
+    unknown = sorted(fields - _REQUIRED_FIELDS - _OPTIONAL_FIELDS)
+    if missing:
+        raise ModelLicenseError(f"Model license is missing required fields: {', '.join(missing)}")
+    if unknown:
+        raise ModelLicenseError(f"Model license contains unsupported fields: {', '.join(unknown)}")
+    if document.get("license_version") != LICENSE_VERSION:
+        raise ModelLicenseError(
+            f"Unsupported model license version: {document.get('license_version')!r}"
+        )
+
+    license_id = _required_string(document, "license_id")
+    issuer = _required_string(document, "issuer")
+    if issuer != LICENSE_ISSUER:
+        raise ModelLicenseError(
+            f"Model license issuer must be {LICENSE_ISSUER!r}; found {issuer!r}"
+        )
+    model_id = _required_string(document, "model_id", maximum=128)
+    if not _MODEL_ID.fullmatch(model_id):
+        raise ModelLicenseError("Model license model_id has an invalid format")
+    if model_id != expected_model_id:
+        raise ModelLicenseError(
+            f"Model license is for {model_id!r}, not the active model {expected_model_id!r}"
+        )
+    grant = _required_string(document, "grant", maximum=32)
+    if grant not in GRANTS:
+        raise ModelLicenseError(
+            f"Model license grant must be one of: {', '.join(sorted(GRANTS))}"
+        )
+    customer = _optional_string(document, "customer")
+    reference = _optional_string(document, "reference")
+    if grant == "commercial" and customer is None:
+        raise ModelLicenseError("A commercial model license must identify the customer")
+
+    valid_from = _parse_time(document, "valid_from", required=True)
+    valid_until = _parse_time(document, "valid_until", required=False)
+    assert valid_from is not None
+    if valid_until is not None and valid_until <= valid_from:
+        raise ModelLicenseError("Model license valid_until must be later than valid_from")
+
+    signature = _signature_bytes(_required_string(document, "signature", maximum=128))
+    signed = canonical_license_bytes(document)
+    keys = tuple(public_keys) if public_keys is not None else _trusted_keys()
+    if not keys:
+        raise ModelLicenseError("No trusted InsightFace model-license keys are configured")
+    for key in keys:
+        try:
+            key.verify(signature, signed)
+            break
+        except InvalidSignature:
+            continue
+    else:
+        raise ModelLicenseError("Model license signature verification failed")
+
+    current = now or datetime.now(UTC)
+    if current.tzinfo is None:
+        raise ModelLicenseError("License verification time must be timezone-aware")
+    current = current.astimezone(UTC)
+    if current < valid_from:
+        raise ModelLicenseError(
+            f"Model license is not active until {_format_time(valid_from)}"
+        )
+    if valid_until is not None and current >= valid_until:
+        raise ModelLicenseError(f"Model license expired at {_format_time(valid_until)}")
+
+    return ModelLicense(
+        license_id=license_id,
+        issuer=issuer,
+        model_id=model_id,
+        grant=grant,
+        valid_from=valid_from,
+        valid_until=valid_until,
+        customer=customer,
+        reference=reference,
+    )
diff --git a/server/backend/insightface_server/licensing/trusted_keys/README.md b/server/backend/insightface_server/licensing/trusted_keys/README.md
new file mode 100644
index 000000000..6873a5ea0
--- /dev/null
+++ b/server/backend/insightface_server/licensing/trusted_keys/README.md
@@ -0,0 +1,12 @@
+# Trusted model-license keys
+
+`insightface-model-license-public-ed25519.pem` is the active InsightFace Ed25519
+public key used to verify `MODEL.LICENSE` files offline.
+
+- Status: active
+- Activated: 2026-07-22
+- Public-key DER SHA-256: `afa29e9508d9ae7a1308974a8f044a49e1d2c8ed7dba6f2a02f66ce0c450181d`
+
+Only public keys belong here. Private issuer keys must remain under the
+Git-ignored repository-root `.private/license-issuer/` directory and must never
+enter a source archive or container image.
diff --git a/server/backend/insightface_server/licensing/trusted_keys/insightface-model-license-public-ed25519.pem b/server/backend/insightface_server/licensing/trusted_keys/insightface-model-license-public-ed25519.pem
new file mode 100644
index 000000000..bf1b32768
--- /dev/null
+++ b/server/backend/insightface_server/licensing/trusted_keys/insightface-model-license-public-ed25519.pem
@@ -0,0 +1,3 @@
+-----BEGIN PUBLIC KEY-----
+MCowBQYDK2VwAyEAJyP2ZksitIXEOZX1cOauRUbsnmNvNmEAA5aXvcNUsA8=
+-----END PUBLIC KEY-----
diff --git a/server/backend/insightface_server/models/__init__.py b/server/backend/insightface_server/models/__init__.py
new file mode 100644
index 000000000..2a22cfe71
--- /dev/null
+++ b/server/backend/insightface_server/models/__init__.py
@@ -0,0 +1,25 @@
+from .embedding_contract import (
+    EMBEDDING_CONTRACT_PREFIX,
+    embedding_contract_id,
+    embedding_contract_id_for_collection,
+)
+from .manifest import (
+    DETECTION_TASK,
+    RECOGNITION_TASK,
+    ModelBundle,
+    ModelSpec,
+    load_manifest,
+    sha256_file,
+)
+
+__all__ = [
+    "DETECTION_TASK",
+    "EMBEDDING_CONTRACT_PREFIX",
+    "RECOGNITION_TASK",
+    "ModelBundle",
+    "ModelSpec",
+    "embedding_contract_id",
+    "embedding_contract_id_for_collection",
+    "load_manifest",
+    "sha256_file",
+]
diff --git a/server/backend/insightface_server/models/embedding_contract.py b/server/backend/insightface_server/models/embedding_contract.py
new file mode 100644
index 000000000..8b099ff23
--- /dev/null
+++ b/server/backend/insightface_server/models/embedding_contract.py
@@ -0,0 +1,51 @@
+from __future__ import annotations
+
+import hashlib
+import json
+from collections.abc import Mapping
+from typing import Any
+
+EMBEDDING_CONTRACT_PREFIX = "ifsemb-v1-sha256:"
+
+
+def embedding_contract_id(
+    *,
+    model_id: str,
+    model_version: str,
+    model_digest: str,
+    embedding_dimension: int,
+    preprocessing_version: str,
+) -> str:
+    """Return the stable identity of one Collection embedding contract.
+
+    The ordered JSON tuple and prefix are deliberately versioned. A future change
+    to alignment, preprocessing identity, or canonicalization must use a new
+    prefix instead of changing the meaning of existing identifiers.
+    """
+
+    canonical_tuple = (
+        model_id,
+        model_version,
+        model_digest.lower(),
+        int(embedding_dimension),
+        preprocessing_version,
+    )
+    canonical_bytes = json.dumps(
+        canonical_tuple,
+        ensure_ascii=False,
+        separators=(",", ":"),
+    ).encode("utf-8")
+    digest = hashlib.sha256(canonical_bytes).hexdigest()
+    return EMBEDDING_CONTRACT_PREFIX + digest
+
+
+def embedding_contract_id_for_collection(collection: Mapping[str, Any]) -> str:
+    """Derive an embedding contract identifier from Collection model fields."""
+
+    return embedding_contract_id(
+        model_id=str(collection["model_id"]),
+        model_version=str(collection["model_version"]),
+        model_digest=str(collection["model_digest"]),
+        embedding_dimension=int(collection["embedding_dimension"]),
+        preprocessing_version=str(collection["preprocessing_version"]),
+    )
diff --git a/server/backend/insightface_server/models/manifest.py b/server/backend/insightface_server/models/manifest.py
new file mode 100644
index 000000000..af533b261
--- /dev/null
+++ b/server/backend/insightface_server/models/manifest.py
@@ -0,0 +1,300 @@
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+from ..licensing import LICENSE_FILENAME
+
+DETECTION_TASK = "face_detection"
+RECOGNITION_TASK = "face_recognition"
+MANIFEST_VERSION = 1
+_MODEL_ID = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$")
+
+
+@dataclass(frozen=True, slots=True)
+class ModelSpec:
+    """One ONNX component described by ``/models/manifest.json``.
+
+    ``sha256`` is calculated from the active artifact for diagnostics and the
+    Collection embedding contract. It is not compared with the manifest and is
+    never part of model-license authorization.
+    """
+
+    model_id: str
+    model_version: str
+    task: str
+    path: Path
+    input_size: tuple[int, int]
+    embedding_dimension: int | None
+    preprocessing_version: str
+    sha256: str
+    input_mean: float
+    input_std: float
+
+    def public_summary(self) -> dict[str, object]:
+        return {
+            "model_id": self.model_id,
+            "model_version": self.model_version,
+            "task": self.task,
+            "file": self.path.name,
+            "input_size": list(self.input_size),
+            "embedding_dimension": self.embedding_dimension,
+            "preprocessing_version": self.preprocessing_version,
+            "sha256": self.sha256,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class ModelBundle:
+    model_id: str
+    model_version: str
+    display_name: str
+    license_path: Path
+    detector: ModelSpec
+    recognizer: ModelSpec
+    legacy_manifest: bool = False
+
+    @property
+    def models(self) -> tuple[ModelSpec, ModelSpec]:
+        return (self.detector, self.recognizer)
+
+
+def sha256_file(path: Path) -> str:
+    hasher = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            hasher.update(chunk)
+    return hasher.hexdigest()
+
+
+def _required_string(raw: dict[str, Any], key: str, *, context: str) -> str:
+    value = raw.get(key)
+    if not isinstance(value, str) or not value.strip():
+        raise RuntimeError(f"Manifest field {key!r} must be a non-empty string in {context}")
+    return value.strip()
+
+
+def _safe_file(
+    models_dir: Path, filename: str, *, suffix: str, must_exist: bool = True
+) -> Path:
+    root = models_dir.resolve()
+    path = (root / filename).resolve()
+    if root not in path.parents or path.suffix.lower() != suffix:
+        raise RuntimeError(f"Unsafe {suffix} path in manifest: {filename}")
+    if must_exist and not path.is_file():
+        raise RuntimeError(f"Model package file does not exist: {path}")
+    return path
+
+
+def _positive_pair(value: object, *, field: str) -> tuple[int, int]:
+    if (
+        not isinstance(value, list)
+        or len(value) != 2
+        or any(isinstance(item, bool) or not isinstance(item, int) or item <= 0 for item in value)
+    ):
+        raise RuntimeError(f"{field} must contain two positive integers")
+    return value[0], value[1]
+
+
+def _positive_integer(value: object, *, field: str) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or value <= 0:
+        raise RuntimeError(f"{field} must be a positive integer")
+    return value
+
+
+def _new_manifest(raw: dict[str, Any], models_dir: Path) -> ModelBundle:
+    allowed = {
+        "manifest_version",
+        "model_id",
+        "model_version",
+        "display_name",
+        "files",
+        "recognition",
+        "license",
+    }
+    unknown = sorted(set(raw) - allowed)
+    if unknown:
+        raise RuntimeError(f"manifest.json contains unsupported fields: {', '.join(unknown)}")
+    if raw.get("manifest_version") != MANIFEST_VERSION:
+        raise RuntimeError(
+            f"manifest_version must be {MANIFEST_VERSION}; found {raw.get('manifest_version')!r}"
+        )
+    model_id = _required_string(raw, "model_id", context="manifest.json")
+    if not _MODEL_ID.fullmatch(model_id):
+        raise RuntimeError("manifest model_id has an invalid format")
+    model_version = _required_string(raw, "model_version", context="manifest.json")
+    display = raw.get("display_name", model_id)
+    if not isinstance(display, str) or not display.strip():
+        raise RuntimeError("display_name must be a non-empty string when provided")
+
+    files = raw.get("files")
+    if not isinstance(files, dict) or set(files) != {"detector", "recognizer"}:
+        raise RuntimeError("files must contain exactly detector and recognizer")
+    detector_name = _required_string(files, "detector", context="files")
+    recognizer_name = _required_string(files, "recognizer", context="files")
+    detector_path = _safe_file(models_dir, detector_name, suffix=".onnx")
+    recognizer_path = _safe_file(models_dir, recognizer_name, suffix=".onnx")
+
+    recognition = raw.get("recognition")
+    if not isinstance(recognition, dict):
+        raise RuntimeError("recognition must be an object")
+    if set(recognition) != {"input_size", "embedding_dimension", "preprocessing"}:
+        raise RuntimeError(
+            "recognition must contain exactly input_size, embedding_dimension, and preprocessing"
+        )
+    recognition_size = _positive_pair(
+        recognition.get("input_size"), field="recognition.input_size"
+    )
+    if recognition_size[0] != recognition_size[1]:
+        raise RuntimeError("Phase-one recognition input_size must be square")
+    dimension = _positive_integer(
+        recognition.get("embedding_dimension"), field="recognition.embedding_dimension"
+    )
+    preprocessing = _required_string(
+        recognition, "preprocessing", context="recognition"
+    )
+
+    license_name = _required_string(raw, "license", context="manifest.json")
+    if Path(license_name).name != LICENSE_FILENAME:
+        raise RuntimeError(f"The model license filename must be {LICENSE_FILENAME}")
+    license_path = _safe_file(
+        models_dir, license_name, suffix=".license", must_exist=False
+    )
+
+    detector = ModelSpec(
+        model_id=model_id,
+        model_version=model_version,
+        task=DETECTION_TASK,
+        path=detector_path,
+        # Active SCRFD resolutions are startup configuration, not manifest data.
+        input_size=(640, 640),
+        embedding_dimension=None,
+        preprocessing_version="insightface-scrfd-1",
+        sha256=sha256_file(detector_path),
+        input_mean=127.5,
+        input_std=128.0,
+    )
+    recognizer = ModelSpec(
+        model_id=model_id,
+        model_version=model_version,
+        task=RECOGNITION_TASK,
+        path=recognizer_path,
+        input_size=recognition_size,
+        embedding_dimension=dimension,
+        preprocessing_version=preprocessing,
+        sha256=sha256_file(recognizer_path),
+        input_mean=127.5,
+        input_std=127.5,
+    )
+    return ModelBundle(
+        model_id=model_id,
+        model_version=model_version,
+        display_name=display.strip(),
+        license_path=license_path,
+        detector=detector,
+        recognizer=recognizer,
+    )
+
+
+def _legacy_spec(raw: object, models_dir: Path) -> ModelSpec:
+    if not isinstance(raw, dict):
+        raise RuntimeError("Every models entry in legacy manifest.json must be an object")
+    filename = _required_string(raw, "file", context="legacy models entry")
+    task = _required_string(raw, "task", context=filename)
+    if task not in {DETECTION_TASK, RECOGNITION_TASK}:
+        raise RuntimeError(f"Unsupported model task {task!r}")
+    path = _safe_file(models_dir, filename, suffix=".onnx")
+    size = _positive_pair(raw.get("input_size"), field=f"input_size for {filename}")
+    dimension_value = raw.get("embedding_dimension")
+    dimension = (
+        None
+        if dimension_value is None
+        else _positive_integer(dimension_value, field=f"embedding_dimension for {filename}")
+    )
+    default_std = 128.0 if task == DETECTION_TASK else 127.5
+    try:
+        input_mean = float(raw.get("input_mean", 127.5))
+        input_std = float(raw.get("input_std", default_std))
+    except (TypeError, ValueError) as exc:
+        raise RuntimeError(f"Invalid input_mean/input_std for {filename}") from exc
+    if not math.isfinite(input_mean) or not math.isfinite(input_std) or input_std <= 0:
+        raise RuntimeError(f"Invalid input_mean/input_std for {filename}")
+    return ModelSpec(
+        model_id=_required_string(raw, "model_id", context=filename),
+        model_version=_required_string(raw, "model_version", context=filename),
+        task=task,
+        path=path,
+        input_size=size,
+        embedding_dimension=dimension,
+        preprocessing_version=_required_string(
+            raw, "preprocessing_version", context=filename
+        ),
+        # A legacy declared digest is intentionally not enforced. It is only
+        # recalculated for diagnostics and Collection compatibility.
+        sha256=sha256_file(path),
+        input_mean=input_mean,
+        input_std=input_std,
+    )
+
+
+def _legacy_manifest(raw: dict[str, Any], models_dir: Path) -> ModelBundle:
+    entries = raw.get("models")
+    if not isinstance(entries, list) or not entries:
+        raise RuntimeError("manifest.json must use manifest_version 1")
+    specs = [_legacy_spec(entry, models_dir) for entry in entries]
+    by_task = {spec.task: spec for spec in specs}
+    if len(specs) != 2 or len(by_task) != 2 or set(by_task) != {DETECTION_TASK, RECOGNITION_TASK}:
+        raise RuntimeError(
+            "Phase one requires exactly one face_detection and one face_recognition model"
+        )
+    recognizer = by_task[RECOGNITION_TASK]
+    if recognizer.embedding_dimension is None:
+        raise RuntimeError("face_recognition must declare embedding_dimension")
+    package = raw.get("package")
+    package_name = package.get("name") if isinstance(package, dict) else None
+    model_id = package_name if isinstance(package_name, str) and package_name else recognizer.model_id
+    parents = {spec.path.parent for spec in specs}
+    if len(parents) != 1:
+        raise RuntimeError("Legacy model files must share one package directory")
+    license_path = parents.pop() / LICENSE_FILENAME
+    return ModelBundle(
+        model_id=model_id,
+        model_version=(
+            str(package.get("release"))
+            if isinstance(package, dict) and package.get("release")
+            else recognizer.model_version
+        ),
+        display_name=model_id,
+        license_path=license_path,
+        detector=by_task[DETECTION_TASK],
+        recognizer=recognizer,
+        legacy_manifest=True,
+    )
+
+
+def load_manifest(models_dir: str | Path) -> ModelBundle:
+    """Load the current compact manifest, with read-only legacy migration support."""
+
+    root = Path(models_dir)
+    manifest_path = root / "manifest.json"
+    if not manifest_path.is_file():
+        raise RuntimeError(f"Required model manifest is missing: {manifest_path}")
+    try:
+        raw = json.loads(
+            manifest_path.read_text(encoding="utf-8"),
+            parse_constant=lambda value: (_ for _ in ()).throw(
+                ValueError(f"Non-standard JSON number: {value}")
+            ),
+        )
+    except (OSError, json.JSONDecodeError, ValueError) as exc:
+        raise RuntimeError(f"Invalid model manifest: {exc}") from exc
+    if not isinstance(raw, dict):
+        raise RuntimeError("Model manifest root must be an object")
+    if "manifest_version" in raw:
+        return _new_manifest(raw, root)
+    return _legacy_manifest(raw, root)
diff --git a/server/backend/insightface_server/models/packages.py b/server/backend/insightface_server/models/packages.py
new file mode 100644
index 000000000..6ae0d1f31
--- /dev/null
+++ b/server/backend/insightface_server/models/packages.py
@@ -0,0 +1,587 @@
+"""Explicitly installed model packages for InsightFace Server.
+
+Official downloads are checksum-verified before installation. Runtime model
+licenses intentionally bind only the logical ``model_id`` so customers may use
+converted FP16, INT8, optimized ONNX, or TensorRT artifacts.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import stat
+import tempfile
+import uuid
+import zipfile
+from collections.abc import Callable
+from dataclasses import dataclass
+from pathlib import Path, PurePosixPath
+from typing import IO
+
+import requests
+from requests.adapters import HTTPAdapter
+from urllib3.util.retry import Retry
+
+from ..licensing import LICENSE_FILENAME, ModelLicense, verify_model_license
+from .manifest import load_manifest, sha256_file
+
+MODEL_LICENSE_SUMMARY = "Non-commercial research use only."
+MODEL_LICENSE_ID = "non-commercial-research-only"
+MODEL_LICENSE_FILENAME = LICENSE_FILENAME
+MODEL_LICENSE_URL = "https://github.com/deepinsight/insightface#license"
+COMMERCIAL_LICENSE_URL = "https://www.insightface.ai"
+DEFAULT_LICENSES_DIR = (
+    Path(__file__).resolve().parents[1] / "licensing" / "defaults"
+)
+MAX_ARCHIVE_BYTES = 400 * 1024 * 1024
+MAX_MODEL_BYTES = 256 * 1024 * 1024
+
+
+class ModelPackageError(RuntimeError):
+    """A safe, user-facing model package operation failed."""
+
+
+@dataclass(frozen=True, slots=True)
+class PackageFile:
+    filename: str
+    sha256: str
+    model_id: str
+    model_version: str
+    task: str
+    input_size: tuple[int, int]
+    preprocessing_version: str
+    input_mean: float
+    input_std: float
+    embedding_dimension: int | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class ModelPackage:
+    name: str
+    release: str
+    display_name: str
+    url: str
+    archive_sha256: str
+    files: tuple[PackageFile, ...]
+    license_summary: str = MODEL_LICENSE_SUMMARY
+    license_id: str = MODEL_LICENSE_ID
+    license_url: str = MODEL_LICENSE_URL
+    commercial_license_url: str = COMMERCIAL_LICENSE_URL
+
+
+BUFFALO_L = ModelPackage(
+    name="buffalo_l",
+    release="v0.7",
+    display_name="Buffalo_L",
+    url="https://github.com/deepinsight/insightface/releases/download/v0.7/buffalo_l.zip",
+    archive_sha256="80ffe37d8a5940d59a7384c201a2a38d4741f2f3c51eef46ebb28218a7b0ca2f",
+    files=(
+        PackageFile(
+            filename="det_10g.onnx",
+            sha256="5838f7fe053675b1c7a08b633df49e7af5495cee0493c7dcf6697200b85b5b91",
+            model_id="scrfd-detection",
+            model_version="1",
+            task="face_detection",
+            input_size=(640, 640),
+            preprocessing_version="insightface-scrfd-1",
+            input_mean=127.5,
+            input_std=128.0,
+        ),
+        PackageFile(
+            filename="w600k_r50.onnx",
+            sha256="4c06341c33c2ca1f86781dab0e829f88ad5b64be9fba56e56bc9ebdefc619e43",
+            model_id="buffalo_l-recognition",
+            model_version="1",
+            task="face_recognition",
+            input_size=(112, 112),
+            preprocessing_version="insightface-arcface-1",
+            input_mean=127.5,
+            input_std=127.5,
+            embedding_dimension=512,
+        ),
+    ),
+)
+
+BUFFALO_M = ModelPackage(
+    name="buffalo_m",
+    release="v0.7",
+    display_name="Buffalo_M",
+    url="https://github.com/deepinsight/insightface/releases/download/v0.7/buffalo_m.zip",
+    archive_sha256="d98264bd8f2dc75cbc2ddce2a14e636e02bb857b3051c234b737bf3b614edca9",
+    files=(
+        PackageFile(
+            filename="det_2.5g.onnx",
+            sha256="041f73f47371333d1d17a6fee6c8ab4e6aecabefe398ff32cca4e2d5eaee0af9",
+            model_id="buffalo_m-detection",
+            model_version="1",
+            task="face_detection",
+            input_size=(640, 640),
+            preprocessing_version="insightface-scrfd-1",
+            input_mean=127.5,
+            input_std=128.0,
+        ),
+        PackageFile(
+            filename="w600k_r50.onnx",
+            sha256="4c06341c33c2ca1f86781dab0e829f88ad5b64be9fba56e56bc9ebdefc619e43",
+            model_id="buffalo_m-recognition",
+            model_version="1",
+            task="face_recognition",
+            input_size=(112, 112),
+            preprocessing_version="insightface-arcface-1",
+            input_mean=127.5,
+            input_std=127.5,
+            embedding_dimension=512,
+        ),
+    ),
+)
+
+BUFFALO_SC = ModelPackage(
+    name="buffalo_sc",
+    release="v0.7",
+    display_name="Buffalo_SC",
+    url="https://github.com/deepinsight/insightface/releases/download/v0.7/buffalo_sc.zip",
+    archive_sha256="57d31b56b6ffa911c8a73cfc1707c73cab76efe7f13b675a05223bf42de47c72",
+    files=(
+        PackageFile(
+            filename="det_500m.onnx",
+            sha256="5e4447f50245bbd7966bd6c0fa52938c61474a04ec7def48753668a9d8b4ea3a",
+            model_id="buffalo_sc-detection",
+            model_version="1",
+            task="face_detection",
+            input_size=(640, 640),
+            preprocessing_version="insightface-scrfd-1",
+            input_mean=127.5,
+            input_std=128.0,
+        ),
+        PackageFile(
+            filename="w600k_mbf.onnx",
+            sha256="9cc6e4a75f0e2bf0b1aed94578f144d15175f357bdc05e815e5c4a02b319eb4f",
+            model_id="buffalo_sc-recognition",
+            model_version="1",
+            task="face_recognition",
+            input_size=(112, 112),
+            preprocessing_version="insightface-arcface-1",
+            input_mean=127.5,
+            input_std=127.5,
+            embedding_dimension=512,
+        ),
+    ),
+)
+
+ANTELOPEV2 = ModelPackage(
+    name="antelopev2",
+    release="v0.7",
+    display_name="AntelopeV2",
+    url="https://github.com/deepinsight/insightface/releases/download/v0.7/antelopev2.zip",
+    archive_sha256="8e182f14fc6e80b3bfa375b33eb6cff7ee05d8ef7633e738d1c89021dcf0c5c5",
+    files=(
+        PackageFile(
+            filename="scrfd_10g_bnkps.onnx",
+            sha256="5838f7fe053675b1c7a08b633df49e7af5495cee0493c7dcf6697200b85b5b91",
+            model_id="antelopev2-detection",
+            model_version="1",
+            task="face_detection",
+            input_size=(640, 640),
+            preprocessing_version="insightface-scrfd-1",
+            input_mean=127.5,
+            input_std=128.0,
+        ),
+        PackageFile(
+            filename="glintr100.onnx",
+            sha256="4ab1d6435d639628a6f3e5008dd4f929edf4c4124b1a7169e1048f9fef534cdf",
+            model_id="antelopev2-recognition",
+            model_version="1",
+            task="face_recognition",
+            input_size=(112, 112),
+            preprocessing_version="insightface-arcface-1",
+            input_mean=127.5,
+            input_std=127.5,
+            embedding_dimension=512,
+        ),
+    ),
+)
+
+PACKAGES: dict[str, ModelPackage] = {
+    package.name: package
+    for package in (BUFFALO_L, BUFFALO_M, BUFFALO_SC, ANTELOPEV2)
+}
+
+
+def package_for_name(name: str) -> ModelPackage:
+    try:
+        return PACKAGES[name]
+    except KeyError as exc:
+        supported = ", ".join(sorted(PACKAGES))
+        raise ModelPackageError(
+            f"Unknown model package {name!r}. Supported packages: {supported}"
+        ) from exc
+
+
+def license_notice(package: ModelPackage) -> str:
+    return "\n".join(
+        (
+            "LICENSE NOTICE",
+            f"{package.name} is provided for non-commercial research use only.",
+            "Commercial use requires a separate license.",
+            f"Licensing information: {package.commercial_license_url}",
+        )
+    )
+
+
+def model_license_document(package: ModelPackage) -> str:
+    """Return the InsightFace-signed default license for a catalog package."""
+
+    path = DEFAULT_LICENSES_DIR / package.name / MODEL_LICENSE_FILENAME
+    try:
+        document = path.read_text(encoding="utf-8")
+    except (OSError, UnicodeError) as exc:
+        raise ModelPackageError(
+            f"Bundled default license is unavailable for {package.name}: {exc}"
+        ) from exc
+    return document
+
+
+def _package_matches_installed(package: ModelPackage, models_dir: Path) -> bool:
+    try:
+        bundle = load_manifest(models_dir)
+    except RuntimeError:
+        return False
+    return bundle.model_id == package.name
+
+
+def installed_package_name(models_dir: Path) -> str | None:
+    for package in PACKAGES.values():
+        if _package_matches_installed(package, models_dir):
+            return package.name
+    return None
+
+
+def _requests_session() -> requests.Session:
+    retry = Retry(
+        total=5,
+        connect=5,
+        read=5,
+        status=5,
+        backoff_factor=1.0,
+        status_forcelist=(429, 500, 502, 503, 504),
+        allowed_methods=frozenset({"GET"}),
+        raise_on_status=False,
+    )
+    session = requests.Session()
+    session.mount("https://", HTTPAdapter(max_retries=retry))
+    session.mount("http://", HTTPAdapter(max_retries=retry))
+    return session
+
+
+def download_archive(package: ModelPackage, destination: Path) -> None:
+    """Download one pinned archive with bounds, retry, and SHA-256 validation."""
+
+    hasher = hashlib.sha256()
+    downloaded = 0
+    next_report = 10
+    try:
+        with _requests_session() as session:
+            response = session.get(
+                package.url,
+                stream=True,
+                timeout=(15, 300),
+                headers={"User-Agent": "InsightFace-Server-Model-Installer/0.2.0"},
+            )
+            response.raise_for_status()
+            declared = response.headers.get("Content-Length")
+            if declared and int(declared) > MAX_ARCHIVE_BYTES:
+                raise ModelPackageError(
+                    f"Model archive exceeds the {MAX_ARCHIVE_BYTES // 1024 // 1024} MiB limit"
+                )
+            total = int(declared) if declared and declared.isdigit() else 0
+            with destination.open("xb") as output:
+                for chunk in response.iter_content(chunk_size=1024 * 1024):
+                    if not chunk:
+                        continue
+                    downloaded += len(chunk)
+                    if downloaded > MAX_ARCHIVE_BYTES:
+                        raise ModelPackageError(
+                            f"Model archive exceeds the {MAX_ARCHIVE_BYTES // 1024 // 1024} MiB limit"
+                        )
+                    output.write(chunk)
+                    hasher.update(chunk)
+                    if total:
+                        progress = min(100, int(downloaded * 100 / total))
+                        if progress >= next_report:
+                            print(f"Download progress: {progress}%")
+                            next_report += 10
+                output.flush()
+                os.fsync(output.fileno())
+    except (OSError, requests.RequestException, ValueError) as exc:
+        if isinstance(exc, ModelPackageError):
+            raise
+        raise ModelPackageError(f"Unable to download {package.name}: {exc}") from exc
+
+    actual = hasher.hexdigest()
+    if actual != package.archive_sha256:
+        raise ModelPackageError(
+            f"Archive SHA-256 mismatch: expected {package.archive_sha256}, actual {actual}"
+        )
+
+
+def _safe_member_for_filename(archive: zipfile.ZipFile, filename: str) -> zipfile.ZipInfo:
+    matches = []
+    for member in archive.infolist():
+        path = PurePosixPath(member.filename)
+        mode = member.external_attr >> 16
+        if path.name == filename and not member.is_dir():
+            if stat.S_ISLNK(mode):
+                raise ModelPackageError(f"Archive entry must not be a symlink: {member.filename}")
+            matches.append(member)
+    if len(matches) != 1:
+        raise ModelPackageError(
+            f"Archive must contain exactly one {filename}; found {len(matches)}"
+        )
+    member = matches[0]
+    if member.file_size > MAX_MODEL_BYTES:
+        raise ModelPackageError(f"Archive entry is too large: {filename}")
+    return member
+
+
+def _copy_checked(source: IO[bytes], destination: Path, expected_sha256: str) -> None:
+    hasher = hashlib.sha256()
+    written = 0
+    with destination.open("xb") as output:
+        while chunk := source.read(1024 * 1024):
+            written += len(chunk)
+            if written > MAX_MODEL_BYTES:
+                raise ModelPackageError(f"Extracted model is too large: {destination.name}")
+            output.write(chunk)
+            hasher.update(chunk)
+        output.flush()
+        os.fsync(output.fileno())
+    actual = hasher.hexdigest()
+    if actual != expected_sha256:
+        raise ModelPackageError(
+            f"SHA-256 mismatch for {destination.name}: expected {expected_sha256}, actual {actual}"
+        )
+
+
+def extract_required_models(package: ModelPackage, archive_path: Path, output_dir: Path) -> None:
+    try:
+        with zipfile.ZipFile(archive_path) as archive:
+            for item in package.files:
+                member = _safe_member_for_filename(archive, item.filename)
+                with archive.open(member) as source:
+                    _copy_checked(source, output_dir / item.filename, item.sha256)
+    except (OSError, zipfile.BadZipFile, RuntimeError) as exc:
+        if isinstance(exc, ModelPackageError):
+            raise
+        raise ModelPackageError(f"Invalid model archive: {exc}") from exc
+
+
+def _manifest(package: ModelPackage, relative_bundle: Path) -> dict[str, object]:
+    by_task = {item.task: item for item in package.files}
+    detector = by_task.get("face_detection")
+    recognizer = by_task.get("face_recognition")
+    if detector is None or recognizer is None or recognizer.embedding_dimension is None:
+        raise ModelPackageError(
+            f"Package {package.name} must contain one detector and one recognizer"
+        )
+    return {
+        "manifest_version": 1,
+        "model_id": package.name,
+        "model_version": package.release,
+        "display_name": package.display_name,
+        "files": {
+            "detector": (relative_bundle / detector.filename).as_posix(),
+            "recognizer": (relative_bundle / recognizer.filename).as_posix(),
+        },
+        "recognition": {
+            "input_size": list(recognizer.input_size),
+            "embedding_dimension": recognizer.embedding_dimension,
+            "preprocessing": recognizer.preprocessing_version,
+        },
+        "license": (relative_bundle / MODEL_LICENSE_FILENAME).as_posix(),
+    }
+
+
+def _write_json_fsync(path: Path, payload: dict[str, object]) -> None:
+    with path.open("x", encoding="utf-8") as output:
+        json.dump(payload, output, indent=2)
+        output.write("\n")
+        output.flush()
+        os.fsync(output.fileno())
+
+
+def _write_text_fsync(path: Path, payload: str) -> None:
+    with path.open("x", encoding="utf-8") as output:
+        output.write(payload)
+        output.flush()
+        os.fsync(output.fileno())
+
+
+def _managed_license_location(
+    package: ModelPackage, models_dir: Path
+) -> tuple[Path, Path]:
+    root = models_dir.expanduser().resolve()
+    bundle = load_manifest(root)
+    if bundle.model_id != package.name:
+        raise ModelPackageError(
+            f"Installed model_id is {bundle.model_id!r}, not {package.name!r}"
+        )
+    try:
+        relative_license = bundle.license_path.relative_to(root)
+    except ValueError as exc:  # pragma: no cover - load_manifest already confines paths
+        raise ModelPackageError("Installed model license escapes the model directory") from exc
+    return root / relative_license, relative_license
+
+
+LicenseProvider = Callable[[ModelPackage], str]
+
+
+def _ensure_model_license(
+    package: ModelPackage,
+    models_dir: Path,
+    *,
+    license_provider: LicenseProvider,
+) -> None:
+    """Install a default license and migrate a recognized legacy manifest."""
+
+    root = models_dir.expanduser().resolve()
+    owner = root.stat()
+    bundle = load_manifest(root)
+    license_path, relative_license = _managed_license_location(package, root)
+    if license_path.exists():
+        verify_model_license(license_path, expected_model_id=package.name)
+    else:
+        _write_text_fsync(license_path, license_provider(package))
+        verify_model_license(license_path, expected_model_id=package.name)
+
+    if bundle.legacy_manifest:
+        parents = {model.path.parent for model in bundle.models}
+        if len(parents) != 1:
+            raise ModelPackageError(
+                f"Installed {package.name} model files must share one package directory"
+            )
+        relative_bundle = parents.pop().relative_to(root)
+        payload = _manifest(package, relative_bundle)
+        if payload["license"] != relative_license.as_posix():
+            raise ModelPackageError("Legacy model license location is inconsistent")
+        temporary_manifest = root / f".manifest-{uuid.uuid4().hex}.json"
+        try:
+            _write_json_fsync(temporary_manifest, payload)
+            os.replace(temporary_manifest, root / "manifest.json")
+        finally:
+            temporary_manifest.unlink(missing_ok=True)
+
+    _restore_owner(license_path, owner.st_uid, owner.st_gid)
+    _restore_owner(root / "manifest.json", owner.st_uid, owner.st_gid)
+    license_path.chmod(0o644)
+    (root / "manifest.json").chmod(0o644)
+
+
+def _verify_model_license(package: ModelPackage, models_dir: Path) -> ModelLicense:
+    root = models_dir.expanduser().resolve()
+    license_path, _relative_license = _managed_license_location(package, root)
+    return verify_model_license(license_path, expected_model_id=package.name)
+
+
+def _restore_owner(root: Path, uid: int, gid: int) -> None:
+    if not hasattr(os, "chown") or os.geteuid() != 0:
+        return
+    paths = (root, *root.rglob("*")) if root.is_dir() else (root,)
+    for path in paths:
+        os.chown(path, uid, gid, follow_symlinks=False)
+
+
+Downloader = Callable[[ModelPackage, Path], None]
+
+
+def install_package(
+    package: ModelPackage,
+    models_dir: Path,
+    *,
+    downloader: Downloader = download_archive,
+    license_provider: LicenseProvider = model_license_document,
+) -> str:
+    """Install a verified package and publish its manifest last."""
+
+    models_dir = models_dir.expanduser().resolve()
+    if not models_dir.is_dir():
+        raise ModelPackageError(f"Configured model directory does not exist: {models_dir}")
+    if _package_matches_installed(package, models_dir):
+        _ensure_model_license(
+            package, models_dir, license_provider=license_provider
+        )
+        return "already_installed"
+    manifest_path = models_dir / "manifest.json"
+    if manifest_path.exists():
+        raise ModelPackageError(
+            "A different or invalid manifest.json already exists. Back up the model directory "
+            "and remove the conflicting bundle before installing."
+        )
+
+    owner = models_dir.stat()
+    bundle_name = f"{package.name}-{package.release}-{package.archive_sha256[:12]}"
+    bundles_dir = models_dir / ".bundles"
+    bundles_dir.mkdir(mode=0o755, exist_ok=True)
+    destination = bundles_dir / bundle_name
+
+    with tempfile.TemporaryDirectory(prefix=".install-", dir=models_dir) as temporary:
+        temporary_dir = Path(temporary)
+        archive_path = temporary_dir / f"{package.name}.zip"
+        payload_dir = temporary_dir / "payload"
+        payload_dir.mkdir()
+        print(f"Downloading {package.name} {package.release} from {package.url}")
+        downloader(package, archive_path)
+        print("Archive SHA-256 verified.")
+        extract_required_models(package, archive_path, payload_dir)
+
+        _write_text_fsync(
+            payload_dir / MODEL_LICENSE_FILENAME, license_provider(package)
+        )
+        local_manifest = _manifest(package, Path("."))
+        _write_json_fsync(payload_dir / "manifest.json", local_manifest)
+        load_manifest(payload_dir)
+
+        if destination.exists():
+            destination_valid = destination.is_dir() and all(
+                (destination / item.filename).is_file()
+                and sha256_file(destination / item.filename) == item.sha256
+                for item in package.files
+            )
+            if not destination_valid:
+                raise ModelPackageError(f"Conflicting package directory already exists: {destination}")
+        else:
+            os.replace(payload_dir, destination)
+        load_manifest(destination)
+
+        relative_bundle = Path(".bundles") / bundle_name
+        manifest_payload = _manifest(package, relative_bundle)
+        temporary_manifest = models_dir / f".manifest-{uuid.uuid4().hex}.json"
+        try:
+            _write_json_fsync(temporary_manifest, manifest_payload)
+            os.replace(temporary_manifest, manifest_path)
+        finally:
+            temporary_manifest.unlink(missing_ok=True)
+
+    _restore_owner(bundles_dir, owner.st_uid, owner.st_gid)
+    _restore_owner(manifest_path, owner.st_uid, owner.st_gid)
+    manifest_path.chmod(0o644)
+    for item in package.files:
+        (destination / item.filename).chmod(0o644)
+    (destination / MODEL_LICENSE_FILENAME).chmod(0o644)
+    load_manifest(models_dir)
+    return "installed"
+
+
+def verify_installed(
+    models_dir: Path,
+) -> tuple[str | None, tuple[dict[str, object], ...], ModelLicense]:
+    bundle = load_manifest(models_dir)
+    installed = installed_package_name(models_dir)
+    license_info = verify_model_license(
+        bundle.license_path, expected_model_id=bundle.model_id
+    )
+    return (
+        installed,
+        tuple(model.public_summary() for model in bundle.models),
+        license_info,
+    )
diff --git a/server/backend/insightface_server/models_cli.py b/server/backend/insightface_server/models_cli.py
new file mode 100644
index 000000000..33a999b69
--- /dev/null
+++ b/server/backend/insightface_server/models_cli.py
@@ -0,0 +1,145 @@
+"""Manage explicitly installed model packages."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+from .licensing import ModelLicense
+from .models.packages import (
+    PACKAGES,
+    ModelPackage,
+    ModelPackageError,
+    install_package,
+    installed_package_name,
+    license_notice,
+    package_for_name,
+    verify_installed,
+)
+
+
+def _models_dir(value: str | None) -> Path:
+    configured = value if value is not None else os.environ.get("INSIGHTFACE_MODELS_DIR", "/models")
+    return Path(configured)
+
+
+def _confirm_license(package: ModelPackage, accepted: bool) -> None:
+    print(license_notice(package))
+    if accepted:
+        return
+    if not sys.stdin.isatty():
+        raise ModelPackageError(
+            "License acceptance is required in non-interactive mode; add --accept-license."
+        )
+    answer = input("Accept this model license and continue? [y/N] ").strip().lower()
+    if answer not in {"y", "yes"}:
+        raise ModelPackageError("Model installation cancelled; license was not accepted.")
+
+
+def _print_package(package: ModelPackage, models_dir: Path) -> None:
+    status = "installed" if installed_package_name(models_dir) == package.name else "not installed"
+    print(f"{package.name}\t{package.release}\t{status}")
+
+
+def _print_verified_license(license_info: ModelLicense) -> None:
+    summary = license_info.public_summary()
+    print("LICENSE VERIFIED")
+    print(f"Issuer: {summary['issuer']}")
+    print(f"License ID: {summary['license_id']}")
+    print(f"Model ID: {summary['model_id']}")
+    print(f"Grant: {summary['grant']}")
+    if summary["customer"]:
+        print(f"Customer: {summary['customer']}")
+    if summary["reference"]:
+        print(f"Reference: {summary['reference']}")
+    print(f"Valid from: {summary['valid_from']}")
+    print(f"Valid until: {summary['valid_until'] or 'perpetual'}")
+    print("Signature: VALID")
+    print(
+        "Commercial use: "
+        + ("PERMITTED" if summary["commercial_use_permitted"] else "NOT PERMITTED")
+    )
+    if not summary["commercial_use_permitted"]:
+        print("Commercial licensing: https://www.insightface.ai")
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(prog="models", description=__doc__)
+    parser.add_argument("--models-dir", help="Override INSIGHTFACE_MODELS_DIR")
+    commands = parser.add_subparsers(dest="command", required=True)
+
+    commands.add_parser("list", help="List supported packages and installation status")
+    install = commands.add_parser("install", help="Download and install a verified package")
+    install.add_argument(
+        "model",
+        help="Package name: buffalo_l, buffalo_m, buffalo_sc, or antelopev2",
+    )
+    install.add_argument(
+        "--accept-license",
+        action="store_true",
+        help="Accept the displayed model license in non-interactive environments",
+    )
+    verify = commands.add_parser(
+        "verify", help="Verify model files, manifest, and the signed model license"
+    )
+    verify.add_argument("model", nargs="?", help="Optionally require this installed package")
+    info = commands.add_parser("info", help="Show package source, hashes, and license")
+    info.add_argument("model", help="Package name")
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = _build_parser()
+    args = parser.parse_args(argv)
+    models_dir = _models_dir(args.models_dir)
+    try:
+        if args.command == "list":
+            print("NAME\tRELEASE\tSTATUS")
+            for package in PACKAGES.values():
+                _print_package(package, models_dir)
+            return 0
+        if args.command == "info":
+            package = package_for_name(args.model)
+            print(f"Name: {package.name}")
+            print(f"Release: {package.release}")
+            print(f"Source: {package.url}")
+            print(f"Archive SHA-256: {package.archive_sha256}")
+            for item in package.files:
+                print(f"{item.filename} SHA-256: {item.sha256}")
+            print(license_notice(package))
+            return 0
+        if args.command == "install":
+            package = package_for_name(args.model)
+            if installed_package_name(models_dir) != package.name:
+                _confirm_license(package, args.accept_license)
+            status = install_package(package, models_dir)
+            print(
+                f"Model package {package.name} is installed and verified in "
+                f"{models_dir.expanduser().resolve()}."
+                if status == "installed"
+                else f"Model package {package.name} is already installed and verified."
+            )
+            print(license_notice(package))
+            return 0
+        if args.command == "verify":
+            installed, models, license_info = verify_installed(models_dir)
+            if args.model and installed != args.model:
+                raise ModelPackageError(
+                    f"Expected {args.model!r}, but the installed package is {installed or 'custom/unknown'}."
+                )
+            print(f"Installed package: {installed or 'custom/unknown'}")
+            for model in models:
+                print(f"Model file: OK ({model['file']})")
+            _print_verified_license(license_info)
+            return 0
+        parser.error("unknown command")
+    except (ModelPackageError, RuntimeError, OSError) as exc:
+        print(f"models: error: {exc}", file=sys.stderr)
+        return 2
+    return 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/server/backend/insightface_server/request_context.py b/server/backend/insightface_server/request_context.py
new file mode 100644
index 000000000..84cddaf74
--- /dev/null
+++ b/server/backend/insightface_server/request_context.py
@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+import contextvars
+import time
+from collections.abc import Iterator
+from contextlib import contextmanager
+
+
+class RequestDeadlineExceeded(RuntimeError):
+    """Raised in a worker before it commits work for an expired request."""
+
+
+REQUEST_DEADLINE: contextvars.ContextVar[float | None] = contextvars.ContextVar(
+    "insightface_request_deadline", default=None
+)
+
+
+@contextmanager
+def without_request_deadline() -> Iterator[None]:
+    """Let an internal compensation restore an already-modified invariant."""
+
+    token = REQUEST_DEADLINE.set(None)
+    try:
+        yield
+    finally:
+        REQUEST_DEADLINE.reset(token)
+
+
+def remaining_seconds() -> float | None:
+    deadline = REQUEST_DEADLINE.get()
+    return None if deadline is None else deadline - time.monotonic()
+
+
+def check_request_deadline() -> None:
+    remaining = remaining_seconds()
+    if remaining is not None and remaining <= 0:
+        raise RequestDeadlineExceeded("The originating request deadline has expired")
diff --git a/server/backend/insightface_server/search/__init__.py b/server/backend/insightface_server/search/__init__.py
new file mode 100644
index 000000000..4beaadc0f
--- /dev/null
+++ b/server/backend/insightface_server/search/__init__.py
@@ -0,0 +1,42 @@
+from .base import (
+    SEARCH_PROFILES,
+    IndexRecord,
+    IndexStats,
+    MutableSearchIndex,
+    PersonHit,
+    SearchBackend,
+    SearchIndexCapacityError,
+    SearchIndexError,
+    SearchIndexStateError,
+)
+from .factory import NativeSearchBackend, ReferenceSearchBackend, create_search_backend
+from .manager import (
+    CollectionIndexState,
+    SearchIndexManager,
+    SearchMutationCommittedError,
+)
+from .native import NativeCapabilities, NativeSearchIndex, NativeSearchLibrary
+from .reference import ReferenceSearchIndex, profile_similarity
+
+__all__ = [
+    "SEARCH_PROFILES",
+    "IndexRecord",
+    "IndexStats",
+    "MutableSearchIndex",
+    "NativeSearchBackend",
+    "NativeCapabilities",
+    "NativeSearchIndex",
+    "NativeSearchLibrary",
+    "PersonHit",
+    "ReferenceSearchIndex",
+    "profile_similarity",
+    "ReferenceSearchBackend",
+    "SearchBackend",
+    "SearchIndexManager",
+    "SearchIndexCapacityError",
+    "SearchIndexError",
+    "SearchIndexStateError",
+    "SearchMutationCommittedError",
+    "CollectionIndexState",
+    "create_search_backend",
+]
diff --git a/server/backend/insightface_server/search/base.py b/server/backend/insightface_server/search/base.py
new file mode 100644
index 000000000..06c25af2d
--- /dev/null
+++ b/server/backend/insightface_server/search/base.py
@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Protocol, runtime_checkable
+
+import numpy as np
+
+SEARCH_PROFILES = frozenset(
+    {"fp32_v1", "fp16_v1", "bf16_v1", "int8_x1000_v1", "int8_x736_v1"}
+)
+
+
+class SearchIndexError(RuntimeError):
+    """Base error raised by an in-process vector index."""
+
+
+class SearchIndexCapacityError(SearchIndexError):
+    """The configured hard row capacity would be exceeded."""
+
+
+class SearchIndexStateError(SearchIndexError):
+    """The index cannot serve the requested operation in its current state."""
+
+
+@dataclass(frozen=True, slots=True)
+class IndexRecord:
+    vector_id: int
+    person_numeric_id: int
+    embedding: np.ndarray
+
+
+@dataclass(frozen=True, slots=True)
+class PersonHit:
+    person_numeric_id: int
+    vector_id: int
+    cosine: float
+
+
+@dataclass(frozen=True, slots=True)
+class IndexStats:
+    profile: str
+    dimension: int
+    live_rows: int
+    physical_rows: int
+    capacity_rows: int
+    tombstone_rows: int
+    reallocations: int = 0
+
+
+@runtime_checkable
+class MutableSearchIndex(Protocol):
+    profile: str
+    dimension: int
+
+    def add_batch(self, records: list[IndexRecord]) -> None: ...
+
+    def remove_batch(self, vector_ids: list[int]) -> set[int]: ...
+
+    def search_persons(self, query: np.ndarray, limit: int) -> list[PersonHit]: ...
+
+    def stats(self) -> IndexStats: ...
+
+    def close(self) -> None: ...
+
+
+@runtime_checkable
+class SearchBackend(Protocol):
+    name: str
+
+    def create_index(
+        self, *, profile: str, dimension: int, capacity_rows: int
+    ) -> MutableSearchIndex: ...
+
+    def runtime_summary(self) -> dict[str, Any]: ...
+
+    def close(self) -> None: ...
diff --git a/server/backend/insightface_server/search/factory.py b/server/backend/insightface_server/search/factory.py
new file mode 100644
index 000000000..e6f477720
--- /dev/null
+++ b/server/backend/insightface_server/search/factory.py
@@ -0,0 +1,127 @@
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from .base import SEARCH_PROFILES, MutableSearchIndex
+from .native import DIMENSION, NativeSearchLibrary
+from .reference import ReferenceSearchIndex
+
+
+class ReferenceSearchBackend:
+    name = "reference"
+
+    def create_index(
+        self, *, profile: str, dimension: int, capacity_rows: int
+    ) -> MutableSearchIndex:
+        return ReferenceSearchIndex(
+            profile=profile, dimension=dimension, capacity_rows=capacity_rows
+        )
+
+    def runtime_summary(self) -> dict[str, Any]:
+        return {
+            "backend": self.name,
+            "implementation": "NumPy exact grouped Flat-IP reference",
+            "production": False,
+            "profiles": sorted(SEARCH_PROFILES),
+            "grouped_person_topk": True,
+            "grouped_person_topk_mode": "host_reference",
+        }
+
+    def close(self) -> None:
+        return None
+
+
+class NativeSearchBackend:
+    def __init__(
+        self,
+        path: Path,
+        *,
+        backend: str,
+        device: int,
+        topk_mode: str,
+    ) -> None:
+        self.name = backend
+        self.topk_mode = topk_mode
+        self.library = NativeSearchLibrary(path, backend=backend, device=device)
+        self._self_test()
+
+    def _self_test(self) -> None:
+        import numpy as np
+
+        index = self.library.create_index(
+            profile="fp32_v1", capacity_rows=1, topk_mode=self.topk_mode
+        )
+        try:
+            from .base import IndexRecord
+
+            vector = np.full(DIMENSION, 1.0 / np.sqrt(DIMENSION), dtype=np.float32)
+            index.add_batch([IndexRecord(1, 1, vector)])
+            hits = index.search_persons(vector, 1)
+            if len(hits) != 1 or hits[0].vector_id != 1 or abs(hits[0].cosine - 1.0) > 1e-4:
+                raise RuntimeError("native grouped Top-K self-test returned an invalid result")
+            if index.remove_batch([1]) != {1} or index.stats().live_rows != 0:
+                raise RuntimeError("native mutation self-test failed")
+        finally:
+            index.close()
+
+    def create_index(
+        self, *, profile: str, dimension: int, capacity_rows: int
+    ) -> MutableSearchIndex:
+        if dimension != DIMENSION:
+            raise ValueError(
+                f"native search requires dimension {DIMENSION}; received {dimension}"
+            )
+        return self.library.create_index(
+            profile=profile,
+            capacity_rows=capacity_rows,
+            topk_mode=self.topk_mode,
+        )
+
+    def runtime_summary(self) -> dict[str, Any]:
+        capabilities = self.library.capabilities
+        return {
+            "backend": self.name,
+            "profiles": list(capabilities.profiles),
+            "device": capabilities.device,
+            "compute_capability": capabilities.compute_capability,
+            "cuda_runtime_version": capabilities.cuda_runtime_version,
+            "cuda_driver_version": capabilities.cuda_driver_version,
+            "build_info": capabilities.build_info,
+            "topk_mode": self.topk_mode,
+            "grouped_person_topk": True,
+            "grouped_person_topk_mode": capabilities.grouped_person_topk_mode,
+        }
+
+    def close(self) -> None:
+        return None
+
+
+def create_search_backend(settings: Any) -> ReferenceSearchBackend | NativeSearchBackend:
+    requested = str(getattr(settings, "search_backend", "auto"))
+    inference_mode = str(getattr(settings, "inference_mode", "onnx"))
+    provider = str(getattr(settings, "execution_provider", "CPUExecutionProvider"))
+    if requested == "auto":
+        if inference_mode == "mock":
+            requested = "reference"
+        elif provider == "CUDAExecutionProvider":
+            requested = "native_cuda"
+        else:
+            requested = "native_cpu"
+    if requested == "reference":
+        return ReferenceSearchBackend()
+    if requested not in {"native_cpu", "native_cuda"}:
+        raise ValueError(f"unsupported search backend: {requested}")
+    configured_path = getattr(settings, "search_library_path", None)
+    filename = (
+        "libifs_search_cuda.so" if requested == "native_cuda" else "libifs_search_cpu.so"
+    )
+    path = Path(configured_path) if configured_path else Path(
+        "/opt/insightface/server/native/lib"
+    ) / filename
+    return NativeSearchBackend(
+        path,
+        backend=requested,
+        device=int(getattr(settings, "search_device_id", 0)),
+        topk_mode=str(getattr(settings, "search_topk_mode", "auto")),
+    )
diff --git a/server/backend/insightface_server/search/manager.py b/server/backend/insightface_server/search/manager.py
new file mode 100644
index 000000000..6a3b40a47
--- /dev/null
+++ b/server/backend/insightface_server/search/manager.py
@@ -0,0 +1,739 @@
+from __future__ import annotations
+
+import logging
+import threading
+from collections.abc import Callable, Iterator
+from contextlib import contextmanager
+from dataclasses import dataclass, field
+from enum import StrEnum
+from typing import Any, TypeVar
+
+import numpy as np
+
+from ..storage import Repository, SearchMutation
+from .base import (
+    IndexRecord,
+    MutableSearchIndex,
+    SearchBackend,
+    SearchIndexCapacityError,
+    SearchIndexError,
+)
+from .reference import profile_similarity
+from .synchronization import ReadWriteLock
+
+LOGGER = logging.getLogger("insightface_server.search")
+T = TypeVar("T")
+
+
+class CollectionIndexState(StrEnum):
+    UNLOADED = "unloaded"
+    BUILDING = "building"
+    READY = "ready"
+    DIRTY = "dirty"
+    REBUILDING = "rebuilding"
+    FAILED = "failed"
+    DRAINING = "draining"
+    CLOSED = "closed"
+
+
+class SearchMutationCommittedError(SearchIndexError):
+    """SQLite committed, but its derived search generation could not catch up."""
+
+    def __init__(
+        self,
+        collection_id: str,
+        revision: int,
+        detail: str,
+        *,
+        committed_result: Any = None,
+    ):
+        self.collection_id = collection_id
+        self.revision = revision
+        self.committed_result = committed_result
+        super().__init__(
+            f"search mutation committed at revision {revision}, but indexing failed: {detail}"
+        )
+
+
+@dataclass(slots=True)
+class _CollectionSlot:
+    collection_id: str
+    lock: ReadWriteLock = field(default_factory=ReadWriteLock)
+    state: CollectionIndexState = CollectionIndexState.UNLOADED
+    index: MutableSearchIndex | None = None
+    applied_revision: int = 0
+    generation: int = 0
+    last_error: str | None = None
+
+
+class SearchIndexManager:
+    """Own one disposable exact-search generation per active Collection."""
+
+    def __init__(
+        self,
+        repository: Repository,
+        backend: SearchBackend,
+        *,
+        build_batch_rows: int = 4096,
+    ) -> None:
+        if build_batch_rows <= 0:
+            raise ValueError("build_batch_rows must be positive")
+        self.repository = repository
+        self.backend = backend
+        self.build_batch_rows = build_batch_rows
+        self._slots: dict[str, _CollectionSlot] = {}
+        self._lock = threading.RLock()
+        self._closed = False
+
+    def _slot(self, collection_id: str) -> _CollectionSlot:
+        with self._lock:
+            if self._closed:
+                raise SearchIndexError("search index manager is closed")
+            return self._slots.setdefault(collection_id, _CollectionSlot(collection_id))
+
+    @contextmanager
+    def _locked_slot(self, collection_id: str) -> Iterator[_CollectionSlot]:
+        """Lock the currently registered slot, retrying across delete/recreate."""
+
+        while True:
+            slot = self._slot(collection_id)
+            slot.lock.acquire_write()
+            with self._lock:
+                registered = self._slots.get(collection_id) is slot
+            if registered:
+                break
+            slot.lock.release_write()
+        try:
+            yield slot
+        finally:
+            slot.lock.release_write()
+
+    @contextmanager
+    def _ready_read_slot(
+        self, collection_id: str
+    ) -> Iterator[tuple[_CollectionSlot, MutableSearchIndex]]:
+        """Yield one stable, current generation while permitting peer readers."""
+
+        while True:
+            with self._locked_slot(collection_id) as slot:
+                index = self._ensure_ready_locked(slot)
+                state = self.repository.get_search_state(collection_id)
+                if state is None:
+                    raise KeyError(collection_id)
+                if state.revision != slot.applied_revision:
+                    slot.state = CollectionIndexState.DIRTY
+                    index = self._ensure_ready_locked(slot)
+                    state = self.repository.get_search_state(collection_id)
+                    if state is None:
+                        raise KeyError(collection_id)
+                if state.max_seq:
+                    self._acknowledge(collection_id, state.max_seq)
+                expected_revision = slot.applied_revision
+                expected_generation = slot.generation
+
+            slot.lock.acquire_read()
+            try:
+                with self._lock:
+                    registered = self._slots.get(collection_id) is slot
+                stable = (
+                    registered
+                    and slot.index is index
+                    and slot.state == CollectionIndexState.READY
+                    and slot.applied_revision == expected_revision
+                    and slot.generation == expected_generation
+                )
+                if not stable:
+                    continue
+                yield slot, index
+                return
+            finally:
+                slot.lock.release_read()
+
+    def _mark_generation_dirty(
+        self, collection_id: str, expected_index: MutableSearchIndex
+    ) -> None:
+        with self._locked_slot(collection_id) as slot:
+            if slot.index is expected_index:
+                slot.state = CollectionIndexState.DIRTY
+
+    def startup(self) -> None:
+        # Search generations are process-local. After a restart every eager
+        # Collection is rebuilt from an authoritative SQLite snapshot and lazy
+        # Collections do the same on first use, so pre-restart outbox rows are
+        # no longer needed for recovery.
+        reset = self.repository.reset_search_changes_before_full_rebuild()
+        if reset:
+            LOGGER.info("discarded_stale_search_changes count=%s", reset)
+        after = ""
+        while True:
+            collections = self.repository.list_collections(after, 100)
+            if not collections:
+                break
+            for collection in collections:
+                if collection.get("load_policy") == "eager":
+                    self.ensure_ready(collection["id"])
+            after = collections[-1]["id"]
+            if len(collections) < 100:
+                break
+
+    def validate_configuration(self, *, profile: str, dimension: int) -> None:
+        summary = self.backend.runtime_summary()
+        supported = summary.get("profiles")
+        if isinstance(supported, list) and profile not in supported:
+            raise SearchIndexError(
+                f"search backend {self.backend.name} does not support profile {profile}"
+            )
+        if self.backend.name.startswith("native_") and dimension != 512:
+            raise SearchIndexError("native search requires 512-dimensional embeddings")
+
+    def run_create(
+        self,
+        collection: dict[str, Any],
+        operation: Callable[[], dict[str, Any]],
+    ) -> dict[str, Any]:
+        """Preallocate an eager empty generation before committing its row."""
+
+        collection_id = str(collection["id"])
+        with self._locked_slot(collection_id) as slot:
+            # Let SQLite produce the authoritative duplicate error without
+            # disturbing an already-active generation.
+            if self.repository.get_collection(collection_id) is not None:
+                return operation()
+            self.validate_configuration(
+                profile=str(collection["search_profile"]),
+                dimension=int(collection["embedding_dimension"]),
+            )
+            staging: MutableSearchIndex | None = None
+            try:
+                if collection.get("load_policy") == "eager":
+                    slot.state = CollectionIndexState.BUILDING
+                    staging = self._new_index(collection)
+                    stats = staging.stats()
+                    if stats.live_rows != 0:
+                        raise SearchIndexError("a new Collection index is not empty")
+                result = operation()
+            except Exception:
+                if staging is not None:
+                    staging.close()
+                slot.state = CollectionIndexState.UNLOADED
+                with self._lock:
+                    if self._slots.get(collection_id) is slot:
+                        self._slots.pop(collection_id, None)
+                raise
+
+            if staging is not None:
+                slot.index = staging
+                slot.applied_revision = int(result["search_revision"])
+                slot.generation += 1
+                slot.state = CollectionIndexState.READY
+                slot.last_error = None
+            else:
+                slot.state = CollectionIndexState.UNLOADED
+            return result
+
+    def collection_created(self, collection: dict[str, Any]) -> None:
+        self.validate_configuration(
+            profile=str(collection["search_profile"]),
+            dimension=int(collection["embedding_dimension"]),
+        )
+        if collection.get("load_policy") == "eager":
+            self.ensure_ready(str(collection["id"]))
+
+    @staticmethod
+    def _record(value: Any) -> IndexRecord:
+        return IndexRecord(
+            vector_id=int(value.vector_id),
+            person_numeric_id=int(value.person_numeric_id),
+            embedding=value.embedding,
+        )
+
+    def _new_index(self, collection: dict[str, Any]) -> MutableSearchIndex:
+        try:
+            return self.backend.create_index(
+                profile=str(collection["search_profile"]),
+                dimension=int(collection["embedding_dimension"]),
+                capacity_rows=int(collection["capacity_rows"]),
+            )
+        except SearchIndexError:
+            raise
+        except Exception as exc:
+            raise SearchIndexError(
+                f"search backend could not allocate the configured index: {exc}"
+            ) from exc
+
+    def _acknowledge(self, collection_id: str, through_seq: int) -> None:
+        """Best-effort outbox cleanup after the index reaches a durable barrier."""
+
+        try:
+            self.repository.acknowledge_search_changes(collection_id, through_seq)
+        except Exception:
+            # The authoritative rows and current generation are already
+            # consistent. Retaining an event is safe and a later successful
+            # acknowledgement (or process restart) will remove it.
+            LOGGER.warning(
+                "unable_to_acknowledge_search_changes collection=%s through_seq=%s",
+                collection_id,
+                through_seq,
+                exc_info=True,
+            )
+
+    def _create_staging(
+        self, collection: dict[str, Any]
+    ) -> tuple[MutableSearchIndex, int, int, int]:
+        """Build a disposable generation without changing the active slot."""
+
+        collection_id = str(collection["id"])
+        staging = self._new_index(collection)
+        try:
+            source_revision: int | None = None
+            source_max_seq: int | None = None
+            loaded_rows = 0
+            for batch in self.repository.iter_index_faces(
+                collection_id, self.build_batch_rows
+            ):
+                if source_revision is None:
+                    source_revision = int(batch.revision)
+                    source_max_seq = int(batch.max_seq)
+                elif source_revision != int(batch.revision):
+                    raise SearchIndexError("SQLite index snapshot revision changed")
+                elif source_max_seq != int(batch.max_seq):
+                    raise SearchIndexError("SQLite index snapshot sequence changed")
+                records = [self._record(face) for face in batch.faces]
+                staging.add_batch(records)
+                loaded_rows += len(records)
+            if source_revision is None:
+                raise SearchIndexError("repository returned no index snapshot")
+            assert source_max_seq is not None
+            committed_revision = self.repository.get_search_revision(collection_id)
+            if committed_revision is None:
+                raise KeyError(collection_id)
+            if committed_revision != source_revision:
+                raise SearchIndexError(
+                    "Collection changed while its index was building; retry is required"
+                )
+            stats = staging.stats()
+            if stats.live_rows != loaded_rows:
+                raise SearchIndexError(
+                    f"index build count mismatch: loaded={loaded_rows} "
+                    f"native={stats.live_rows}"
+                )
+            return staging, source_revision, source_max_seq, loaded_rows
+        except (KeyError, SearchIndexError):
+            staging.close()
+            raise
+        except Exception as exc:
+            staging.close()
+            raise SearchIndexError(f"search index build failed: {exc}") from exc
+
+    def _build_locked(
+        self, slot: _CollectionSlot, collection: dict[str, Any]
+    ) -> MutableSearchIndex:
+        slot.state = (
+            CollectionIndexState.REBUILDING
+            if slot.index is not None
+            else CollectionIndexState.BUILDING
+        )
+        staging: MutableSearchIndex | None = None
+        try:
+            staging, source_revision, source_max_seq, loaded_rows = self._create_staging(
+                collection
+            )
+            stats = staging.stats()
+            old = slot.index
+            slot.index = staging
+            slot.applied_revision = source_revision
+            slot.generation += 1
+            slot.state = CollectionIndexState.READY
+            slot.last_error = None
+            self._acknowledge(slot.collection_id, source_max_seq)
+            if old is not None:
+                try:
+                    old.close()
+                except Exception:
+                    # The new generation is already authoritative. A best-effort
+                    # close failure must not roll the slot back to FAILED.
+                    LOGGER.warning(
+                        "unable_to_close_retired_search_generation collection=%s",
+                        slot.collection_id,
+                        exc_info=True,
+                    )
+            LOGGER.info(
+                "search_index_ready collection=%s generation=%s revision=%s rows=%s "
+                "capacity=%s profile=%s backend=%s",
+                slot.collection_id,
+                slot.generation,
+                slot.applied_revision,
+                stats.live_rows,
+                stats.capacity_rows,
+                stats.profile,
+                self.backend.name,
+            )
+            return slot.index
+        except Exception as exc:
+            if staging is not None and slot.index is not staging:
+                try:
+                    staging.close()
+                except Exception:
+                    LOGGER.warning(
+                        "unable_to_close_failed_search_generation collection=%s",
+                        slot.collection_id,
+                        exc_info=True,
+                    )
+            slot.last_error = f"{type(exc).__name__}: {exc}"
+            slot.state = CollectionIndexState.FAILED
+            raise
+
+    def _ensure_ready_locked(
+        self, slot: _CollectionSlot, collection: dict[str, Any] | None = None
+    ) -> MutableSearchIndex:
+        item = collection or self.repository.get_collection(slot.collection_id)
+        if item is None:
+            raise KeyError(slot.collection_id)
+        committed = int(item["search_revision"])
+        if (
+            slot.state == CollectionIndexState.READY
+            and slot.index is not None
+            and slot.applied_revision == committed
+        ):
+            return slot.index
+        if slot.index is not None:
+            slot.state = CollectionIndexState.DIRTY
+        return self._build_locked(slot, item)
+
+    def ensure_ready(self, collection_id: str) -> None:
+        with self._locked_slot(collection_id) as slot:
+            self._ensure_ready_locked(slot)
+
+    def _apply_mutation_locked(
+        self, slot: _CollectionSlot, mutation: SearchMutation
+    ) -> None:
+        if mutation.collection_id != slot.collection_id:
+            raise SearchIndexError("search mutation Collection does not match its index")
+        if slot.index is None or slot.state != CollectionIndexState.READY:
+            self._ensure_ready_locked(slot)
+            return
+        if mutation.revision != slot.applied_revision + 1:
+            slot.state = CollectionIndexState.DIRTY
+            self._ensure_ready_locked(slot)
+            return
+        additions = [
+            int(change.vector_id)
+            for change in mutation.changes
+            if change.operation == "add" and change.vector_id is not None
+        ]
+        deletions = [
+            int(change.vector_id)
+            for change in mutation.changes
+            if change.operation == "delete" and change.vector_id is not None
+        ]
+        try:
+            if additions:
+                faces = self.repository.get_index_faces_by_vector_ids(
+                    slot.collection_id, additions
+                )
+                if set(faces) != set(additions):
+                    raise SearchIndexError("an added FaceSample is missing from SQLite")
+                slot.index.add_batch([self._record(faces[value]) for value in additions])
+            if deletions:
+                slot.index.remove_batch(deletions)
+            slot.applied_revision = mutation.revision
+            current = self.repository.get_search_revision(slot.collection_id)
+            if current != slot.applied_revision:
+                raise SearchIndexError("index revision did not reach the SQLite barrier")
+            self._acknowledge(slot.collection_id, mutation.max_seq)
+        except Exception:
+            slot.state = CollectionIndexState.DIRTY
+            # A failed native mutation may have partially changed the handle.
+            # Never reuse it; rebuild deterministically from SQLite.
+            if slot.index is not None:
+                slot.index.close()
+                slot.index = None
+            self._ensure_ready_locked(slot)
+
+    def run_mutation(
+        self,
+        collection_id: str,
+        operation: Callable[[], tuple[T, SearchMutation | None]],
+        *,
+        expected_additions: int = 0,
+    ) -> T:
+        with self._locked_slot(collection_id) as slot:
+            index = self._ensure_ready_locked(slot)
+            if expected_additions:
+                stats = index.stats()
+                if stats.live_rows + expected_additions > stats.capacity_rows:
+                    raise SearchIndexCapacityError(
+                        f"Collection {collection_id} capacity {stats.capacity_rows} "
+                        "would be exceeded"
+                    )
+            result, mutation = operation()
+            if mutation is None:
+                return result
+            try:
+                self._apply_mutation_locked(slot, mutation)
+            except Exception as exc:
+                raise SearchMutationCommittedError(
+                    collection_id,
+                    mutation.revision,
+                    str(exc),
+                    committed_result=result,
+                ) from exc
+            return result
+
+    def run_drop(
+        self,
+        collection_id: str,
+        operation: Callable[[], tuple[T, SearchMutation | None]],
+    ) -> T:
+        with self._locked_slot(collection_id) as slot:
+            result, mutation = operation()
+            if mutation is None:
+                return result
+            self._acknowledge(collection_id, mutation.max_seq)
+            if slot.index is not None:
+                slot.state = CollectionIndexState.DRAINING
+                try:
+                    slot.index.close()
+                except Exception:
+                    # SQLite deletion is already committed. Native handles are
+                    # derived state, so close failure must not block slot/crop
+                    # cleanup or turn a durable delete into an ambiguous error.
+                    LOGGER.warning(
+                        "unable_to_close_dropped_search_generation collection=%s",
+                        collection_id,
+                        exc_info=True,
+                    )
+                slot.index = None
+            slot.state = CollectionIndexState.CLOSED
+            slot.applied_revision = 0
+            # Remove while still holding slot.lock. A waiter that captured this
+            # slot will fail _locked_slot's identity check and retry with the
+            # replacement slot before recreating the same Collection ID.
+            with self._lock:
+                if self._slots.get(collection_id) is slot:
+                    self._slots.pop(collection_id, None)
+        return result
+
+    def run_configuration_update(
+        self,
+        collection_id: str,
+        operation: Callable[[], T],
+        *,
+        changes: dict[str, Any],
+    ) -> T:
+        with self._locked_slot(collection_id) as slot:
+            current = self.repository.get_collection(collection_id)
+            if current is None:
+                raise KeyError(collection_id)
+            effective_changes = {
+                key: value
+                for key, value in changes.items()
+                if current.get(key) != value
+            }
+            target = {**current, **effective_changes}
+            target_policy = str(target["load_policy"])
+            explicitly_unload = (
+                "load_policy" in effective_changes and target_policy == "lazy"
+            )
+            needs_staging = not explicitly_unload and (
+                (target_policy == "eager" and slot.index is None)
+                or ("capacity_rows" in effective_changes and slot.index is not None)
+            )
+            previous_state = slot.state
+            staging: MutableSearchIndex | None = None
+            staging_revision = 0
+            staging_max_seq = 0
+            if needs_staging:
+                slot.state = (
+                    CollectionIndexState.REBUILDING
+                    if slot.index is not None
+                    else CollectionIndexState.BUILDING
+                )
+                try:
+                    staging, staging_revision, staging_max_seq, _ = self._create_staging(
+                        target
+                    )
+                except Exception as exc:
+                    slot.state = previous_state
+                    slot.last_error = f"{type(exc).__name__}: {exc}"
+                    raise
+
+            try:
+                result = operation()
+            except Exception:
+                if staging is not None:
+                    staging.close()
+                slot.state = previous_state
+                raise
+
+            if explicitly_unload:
+                if slot.index is not None:
+                    slot.index.close()
+                    slot.index = None
+                slot.state = CollectionIndexState.UNLOADED
+                slot.applied_revision = 0
+                slot.last_error = None
+            elif staging is not None:
+                old = slot.index
+                slot.index = staging
+                slot.applied_revision = staging_revision
+                slot.generation += 1
+                slot.state = CollectionIndexState.READY
+                slot.last_error = None
+                self._acknowledge(collection_id, staging_max_seq)
+                if old is not None:
+                    try:
+                        old.close()
+                    except Exception:
+                        LOGGER.warning(
+                            "unable_to_close_retired_search_generation collection=%s",
+                            collection_id,
+                            exc_info=True,
+                        )
+            return result
+
+    def search(
+        self,
+        collection_id: str,
+        query: np.ndarray,
+        *,
+        limit: int,
+        threshold: float,
+    ) -> list[dict[str, Any]]:
+        stale_index: MutableSearchIndex | None = None
+        stale_message: str | None = None
+        with self._ready_read_slot(collection_id) as (slot, index):
+            hits = index.search_persons(query, limit)
+            mappings = self.repository.get_index_face_mappings(
+                collection_id, [hit.vector_id for hit in hits]
+            )
+            if len(mappings) != len(hits):
+                stale_index = index
+                stale_message = "native result is missing from SQLite hydration"
+            matches: list[dict[str, Any]] = []
+            for hit in hits if stale_message is None else []:
+                item = mappings[hit.vector_id]
+                if int(item["person_numeric_id"]) != hit.person_numeric_id:
+                    stale_index = index
+                    stale_message = "native Person mapping disagrees with SQLite"
+                    break
+                # Match on the unrounded raw cosine. Rounding before this
+                # comparison can flip decisions immediately below a threshold.
+                cosine = float(np.clip(hit.cosine, -1.0, 1.0))
+                if cosine < threshold:
+                    continue
+                matches.append(
+                    {
+                        "person": item["person"],
+                        "similarity": cosine,
+                        "matched_face_id": item["face_id"],
+                    }
+                )
+        if stale_index is not None:
+            self._mark_generation_dirty(collection_id, stale_index)
+            raise SearchIndexError(stale_message or "search generation is inconsistent")
+        matches.sort(key=lambda value: (-value["similarity"], value["person"]["id"]))
+        return matches[:limit]
+
+    def best_other_person(
+        self,
+        collection_id: str,
+        query: np.ndarray,
+        *,
+        exclude_person_id: str,
+    ) -> dict[str, Any] | None:
+        """Return the strongest exact match outside ``exclude_person_id``.
+
+        Strict enrollment review uses the Collection's configured in-memory
+        profile, so its class-out score has exactly the same quantization and
+        accumulation semantics as normal search. ``search_persons`` already
+        groups FaceSamples by Person; requesting two hits is sufficient because
+        at most one grouped hit can belong to the excluded Person.
+        """
+
+        stale_index: MutableSearchIndex | None = None
+        stale_message: str | None = None
+        result: dict[str, Any] | None = None
+        with self._ready_read_slot(collection_id) as (slot, index):
+            hits = index.search_persons(query, 2)
+            mappings = self.repository.get_index_face_mappings(
+                collection_id,
+                [hit.vector_id for hit in hits],
+                include_embedding=True,
+            )
+            if len(mappings) != len(hits):
+                stale_index = index
+                stale_message = "native result is missing from SQLite hydration"
+
+            for hit in hits if stale_message is None else []:
+                item = mappings[hit.vector_id]
+                if int(item["person_numeric_id"]) != hit.person_numeric_id:
+                    stale_index = index
+                    stale_message = "native Person mapping disagrees with SQLite"
+                    break
+                person = item["person"]
+                if str(person["id"]) == exclude_person_id:
+                    continue
+                result = {
+                    "person_id": str(person["id"]),
+                    "face_id": str(item["face_id"]),
+                    # Public search results are clipped to cosine's range. The
+                    # strict reviewer needs the profile's raw accumulator score
+                    # on both sides of `same > other`, including rare INT8
+                    # values slightly above 1 after quantization.
+                    "similarity": profile_similarity(
+                        index.profile, query, item["embedding"]
+                    ),
+                }
+                break
+        if stale_index is not None:
+            self._mark_generation_dirty(collection_id, stale_index)
+            raise SearchIndexError(stale_message or "search generation is inconsistent")
+        return result
+
+    def runtime_summary(self) -> dict[str, Any]:
+        with self._lock:
+            slots = list(self._slots.values())
+        collections: list[dict[str, Any]] = []
+        for slot in slots:
+            with slot.lock.read():
+                value: dict[str, Any] = {
+                    "collection_id": slot.collection_id,
+                    "state": slot.state.value,
+                    "generation": slot.generation,
+                    "applied_revision": slot.applied_revision,
+                }
+                if slot.last_error:
+                    value["last_error"] = slot.last_error
+                if slot.index is not None:
+                    stats = slot.index.stats()
+                    value.update(
+                        {
+                            "profile": stats.profile,
+                            "live_rows": stats.live_rows,
+                            "physical_rows": stats.physical_rows,
+                            "capacity_rows": stats.capacity_rows,
+                            "tombstone_rows": stats.tombstone_rows,
+                            "reallocations": stats.reallocations,
+                        }
+                    )
+                collections.append(value)
+        return {
+            **self.backend.runtime_summary(),
+            "collections": sorted(collections, key=lambda value: value["collection_id"]),
+        }
+
+    def close(self) -> None:
+        with self._lock:
+            if self._closed:
+                return
+            self._closed = True
+            slots = list(self._slots.values())
+        for slot in slots:
+            with slot.lock.write():
+                slot.state = CollectionIndexState.DRAINING
+                if slot.index is not None:
+                    slot.index.close()
+                    slot.index = None
+                slot.state = CollectionIndexState.CLOSED
+        self.backend.close()
diff --git a/server/backend/insightface_server/search/native.py b/server/backend/insightface_server/search/native.py
new file mode 100644
index 000000000..01a978700
--- /dev/null
+++ b/server/backend/insightface_server/search/native.py
@@ -0,0 +1,417 @@
+from __future__ import annotations
+
+import ctypes
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+
+from .base import (
+    IndexRecord,
+    IndexStats,
+    PersonHit,
+    SearchIndexCapacityError,
+    SearchIndexError,
+    SearchIndexStateError,
+)
+from .synchronization import ReadWriteLock
+
+ABI_VERSION = 2
+DIMENSION = 512
+
+_PROFILE_CODES = {
+    "fp32_v1": 0,
+    "fp16_v1": 1,
+    "bf16_v1": 2,
+    "int8_x1000_v1": 3,
+    "int8_x736_v1": 4,
+}
+_BACKEND_CODES = {"native_cpu": 1, "native_cuda": 2}
+_TOPK_CODES = {"auto": 0, "host": 1, "device": 2}
+
+_OK = 0
+_OUT_OF_MEMORY = 2
+_UNSUPPORTED = 3
+_CAPACITY_EXCEEDED = 6
+_CAP_GROUPED_PERSON_TOPK = 1 << 5
+_CAP_GROUPED_HOST_REFERENCE = 1 << 8
+_CAP_GROUPED_DEVICE_RESIDENT = 1 << 9
+
+
+class _Capabilities(ctypes.Structure):
+    _fields_ = [
+        ("struct_size", ctypes.c_uint32),
+        ("abi_version", ctypes.c_uint32),
+        ("dimension", ctypes.c_uint32),
+        ("backend", ctypes.c_uint32),
+        ("profile_mask", ctypes.c_uint64),
+        ("flags", ctypes.c_uint64),
+        ("device_topk_limit", ctypes.c_uint64),
+        ("device", ctypes.c_int32),
+        ("compute_capability_major", ctypes.c_int32),
+        ("compute_capability_minor", ctypes.c_int32),
+        ("cuda_runtime_version", ctypes.c_int32),
+        ("cuda_driver_version", ctypes.c_int32),
+    ]
+
+
+class _CreateOptions(ctypes.Structure):
+    _fields_ = [
+        ("struct_size", ctypes.c_uint32),
+        ("profile", ctypes.c_uint32),
+        ("reserve_rows", ctypes.c_uint64),
+        ("max_rows", ctypes.c_uint64),
+        ("device", ctypes.c_int32),
+        ("topk_mode", ctypes.c_uint32),
+        ("growth_factor", ctypes.c_double),
+        ("reserved", ctypes.c_uint64 * 4),
+    ]
+
+
+class _Stats(ctypes.Structure):
+    _fields_ = [
+        ("struct_size", ctypes.c_uint32),
+        ("backend", ctypes.c_uint32),
+        ("profile", ctypes.c_uint32),
+        ("device", ctypes.c_int32),
+        ("physical_rows", ctypes.c_uint64),
+        ("live_rows", ctypes.c_uint64),
+        ("capacity_rows", ctypes.c_uint64),
+        ("max_rows", ctypes.c_uint64),
+        ("tombstone_rows", ctypes.c_uint64),
+        ("reallocations", ctypes.c_uint64),
+        ("bytes_per_vector", ctypes.c_uint64),
+    ]
+
+
+class _Timings(ctypes.Structure):
+    _fields_ = [
+        ("struct_size", ctypes.c_uint32),
+        ("reserved", ctypes.c_uint32),
+        ("kernel_ms", ctypes.c_double),
+        ("topk_ms", ctypes.c_double),
+        ("total_ms", ctypes.c_double),
+    ]
+
+
+@dataclass(frozen=True, slots=True)
+class NativeCapabilities:
+    backend: str
+    profiles: tuple[str, ...]
+    flags: int
+    device: int
+    compute_capability: str | None
+    cuda_runtime_version: int
+    cuda_driver_version: int
+    build_info: str
+
+    @property
+    def grouped_person_topk_mode(self) -> str:
+        if self.flags & _CAP_GROUPED_DEVICE_RESIDENT:
+            return "device_exact"
+        if self.flags & _CAP_GROUPED_HOST_REFERENCE:
+            return "host_reference"
+        return "unavailable"
+
+
+def _decode(value: bytes | None) -> str:
+    return value.decode("utf-8", errors="replace") if value else ""
+
+
+def _validated_vectors(records: list[IndexRecord]) -> np.ndarray:
+    values = np.asarray([record.embedding for record in records], dtype=np.float32)
+    if values.shape != (len(records), DIMENSION):
+        raise ValueError(f"native search requires vectors shaped (N, {DIMENSION})")
+    if not np.all(np.isfinite(values)):
+        raise ValueError("embedding contains non-finite values")
+    norms = np.linalg.norm(values, axis=1)
+    if np.any(np.abs(norms - 1.0) > 2e-4):
+        raise ValueError("all embeddings must be L2-normalized")
+    return np.ascontiguousarray(values)
+
+
+def _validated_query(query: np.ndarray) -> np.ndarray:
+    value = np.ascontiguousarray(query, dtype=np.float32).reshape(-1)
+    if value.shape != (DIMENSION,):
+        raise ValueError(f"native search query must contain {DIMENSION} values")
+    if not np.all(np.isfinite(value)):
+        raise ValueError("query contains non-finite values")
+    norm = float(np.linalg.norm(value))
+    if abs(norm - 1.0) > 2e-4:
+        raise ValueError("query must be L2-normalized")
+    return value
+
+
+class NativeSearchLibrary:
+    """Validated process-wide handle for one CPU or CUDA ABI-v2 library."""
+
+    def __init__(self, path: Path, *, backend: str, device: int) -> None:
+        if backend not in _BACKEND_CODES:
+            raise ValueError(f"unsupported native backend: {backend}")
+        self.path = Path(path)
+        self.backend = backend
+        self.device = -1 if backend == "native_cpu" else int(device)
+        try:
+            self._library = ctypes.CDLL(str(self.path))
+        except OSError as exc:
+            raise SearchIndexError(f"unable to load native search library: {exc}") from exc
+        self._bind()
+        actual_abi = int(self._library.ifs_search_abi_version())
+        if actual_abi != ABI_VERSION:
+            raise SearchIndexError(
+                f"native search ABI mismatch: expected {ABI_VERSION}, got {actual_abi}"
+            )
+        dimension = int(self._library.ifs_search_dimension())
+        if dimension != DIMENSION:
+            raise SearchIndexError(
+                f"native search dimension mismatch: expected {DIMENSION}, got {dimension}"
+            )
+        raw = _Capabilities()
+        raw.struct_size = ctypes.sizeof(_Capabilities)
+        self._check(self._library.ifs_search_get_capabilities(self.device, ctypes.byref(raw)))
+        expected_backend = _BACKEND_CODES[backend]
+        if raw.backend != expected_backend:
+            raise SearchIndexError(
+                f"native library backend mismatch: expected {expected_backend}, got {raw.backend}"
+            )
+        if raw.abi_version != ABI_VERSION or raw.dimension != DIMENSION:
+            raise SearchIndexError("native capability metadata is inconsistent")
+        if not raw.flags & _CAP_GROUPED_PERSON_TOPK:
+            raise SearchIndexError("native library lacks exact grouped Person Top-K")
+        supported = tuple(
+            name for name, code in _PROFILE_CODES.items() if raw.profile_mask & (1 << code)
+        )
+        capability = None
+        if backend == "native_cuda" and raw.compute_capability_major >= 0:
+            capability = f"{raw.compute_capability_major}.{raw.compute_capability_minor}"
+        self.capabilities = NativeCapabilities(
+            backend=backend,
+            profiles=supported,
+            flags=int(raw.flags),
+            device=int(raw.device),
+            compute_capability=capability,
+            cuda_runtime_version=int(raw.cuda_runtime_version),
+            cuda_driver_version=int(raw.cuda_driver_version),
+            build_info=_decode(self._library.ifs_search_build_info()),
+        )
+
+    def _bind(self) -> None:
+        library = self._library
+        library.ifs_search_abi_version.restype = ctypes.c_uint32
+        library.ifs_search_dimension.restype = ctypes.c_uint32
+        library.ifs_search_build_info.restype = ctypes.c_char_p
+        library.ifs_search_last_error.restype = ctypes.c_char_p
+        library.ifs_search_status_string.argtypes = [ctypes.c_int]
+        library.ifs_search_status_string.restype = ctypes.c_char_p
+        library.ifs_search_get_capabilities.argtypes = [
+            ctypes.c_int32,
+            ctypes.POINTER(_Capabilities),
+        ]
+        library.ifs_search_get_capabilities.restype = ctypes.c_int
+        library.ifs_search_create.argtypes = [
+            ctypes.POINTER(_CreateOptions),
+            ctypes.POINTER(ctypes.c_void_p),
+        ]
+        library.ifs_search_create.restype = ctypes.c_int
+        library.ifs_search_destroy.argtypes = [ctypes.c_void_p]
+        library.ifs_search_destroy.restype = None
+        library.ifs_search_add_batch.argtypes = [
+            ctypes.c_void_p,
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.POINTER(ctypes.c_float),
+            ctypes.c_uint64,
+        ]
+        library.ifs_search_add_batch.restype = ctypes.c_int
+        library.ifs_search_delete_batch.argtypes = [
+            ctypes.c_void_p,
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.c_uint64,
+            ctypes.POINTER(ctypes.c_uint64),
+        ]
+        library.ifs_search_delete_batch.restype = ctypes.c_int
+        library.ifs_search_grouped_topk.argtypes = [
+            ctypes.c_void_p,
+            ctypes.POINTER(ctypes.c_float),
+            ctypes.c_uint64,
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.POINTER(ctypes.c_float),
+            ctypes.POINTER(ctypes.c_uint64),
+            ctypes.POINTER(_Timings),
+        ]
+        library.ifs_search_grouped_topk.restype = ctypes.c_int
+        library.ifs_search_get_stats.argtypes = [ctypes.c_void_p, ctypes.POINTER(_Stats)]
+        library.ifs_search_get_stats.restype = ctypes.c_int
+
+    def _check(self, status: int) -> None:
+        if status == _OK:
+            return
+        detail = _decode(self._library.ifs_search_last_error())
+        label = _decode(self._library.ifs_search_status_string(status)) or str(status)
+        message = f"native search {label}: {detail or 'no detail'}"
+        if status == _CAPACITY_EXCEEDED:
+            raise SearchIndexCapacityError(message)
+        if status in {_OUT_OF_MEMORY, _UNSUPPORTED}:
+            raise SearchIndexError(message)
+        raise SearchIndexError(message)
+
+    def create_index(
+        self,
+        *,
+        profile: str,
+        capacity_rows: int,
+        topk_mode: str = "auto",
+    ) -> NativeSearchIndex:
+        if profile not in _PROFILE_CODES:
+            raise ValueError(f"unsupported search profile: {profile}")
+        if profile not in self.capabilities.profiles:
+            raise SearchIndexError(
+                f"{self.backend} does not support search profile {profile}"
+            )
+        if topk_mode not in _TOPK_CODES:
+            raise ValueError(f"unsupported Top-K mode: {topk_mode}")
+        if capacity_rows <= 0:
+            raise ValueError("capacity_rows must be positive")
+        options = _CreateOptions()
+        options.struct_size = ctypes.sizeof(_CreateOptions)
+        options.profile = _PROFILE_CODES[profile]
+        options.reserve_rows = capacity_rows
+        options.max_rows = capacity_rows
+        options.device = self.device
+        options.topk_mode = _TOPK_CODES[topk_mode]
+        options.growth_factor = 1.5
+        handle = ctypes.c_void_p()
+        self._check(self._library.ifs_search_create(ctypes.byref(options), ctypes.byref(handle)))
+        if not handle.value:
+            raise SearchIndexError("native search returned a null index handle")
+        return NativeSearchIndex(self, handle, profile=profile)
+
+
+class NativeSearchIndex:
+    dimension = DIMENSION
+
+    def __init__(
+        self,
+        library: NativeSearchLibrary,
+        handle: ctypes.c_void_p,
+        *,
+        profile: str,
+    ) -> None:
+        self.library = library
+        self.profile = profile
+        self._handle = handle
+        self._lock = ReadWriteLock()
+        self._closed = False
+
+    def _ensure_open(self) -> None:
+        if self._closed or not self._handle.value:
+            raise SearchIndexStateError("native search index is closed")
+
+    def add_batch(self, records: list[IndexRecord]) -> None:
+        if not records:
+            return
+        values = _validated_vectors(records)
+        vector_ids = np.ascontiguousarray(
+            [int(record.vector_id) for record in records], dtype=np.uint64
+        )
+        group_ids = np.ascontiguousarray(
+            [int(record.person_numeric_id) for record in records], dtype=np.uint64
+        )
+        if np.any(vector_ids == 0) or np.any(group_ids == 0):
+            raise ValueError("native vector and Person IDs must be positive")
+        with self._lock.write():
+            self._ensure_open()
+            self.library._check(
+                self.library._library.ifs_search_add_batch(
+                    self._handle,
+                    vector_ids.ctypes.data_as(ctypes.POINTER(ctypes.c_uint64)),
+                    group_ids.ctypes.data_as(ctypes.POINTER(ctypes.c_uint64)),
+                    values.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
+                    len(records),
+                )
+            )
+
+    def remove_batch(self, vector_ids: list[int]) -> set[int]:
+        if not vector_ids:
+            return set()
+        values = np.ascontiguousarray(vector_ids, dtype=np.uint64)
+        if np.any(values == 0):
+            raise ValueError("native vector IDs must be positive")
+        removed = ctypes.c_uint64()
+        with self._lock.write():
+            self._ensure_open()
+            self.library._check(
+                self.library._library.ifs_search_delete_batch(
+                    self._handle,
+                    values.ctypes.data_as(ctypes.POINTER(ctypes.c_uint64)),
+                    len(vector_ids),
+                    ctypes.byref(removed),
+                )
+            )
+        if removed.value != len(vector_ids):
+            raise SearchIndexError(
+                "native delete count mismatch; the collection index must be rebuilt"
+            )
+        return {int(value) for value in vector_ids}
+
+    def search_persons(self, query: np.ndarray, limit: int) -> list[PersonHit]:
+        if limit <= 0:
+            raise ValueError("limit must be positive")
+        value = _validated_query(query)
+        group_ids = np.empty(limit, dtype=np.uint64)
+        vector_ids = np.empty(limit, dtype=np.uint64)
+        scores = np.empty(limit, dtype=np.float32)
+        count = ctypes.c_uint64()
+        timings = _Timings()
+        timings.struct_size = ctypes.sizeof(_Timings)
+        with self._lock.read():
+            self._ensure_open()
+            self.library._check(
+                self.library._library.ifs_search_grouped_topk(
+                    self._handle,
+                    value.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
+                    limit,
+                    group_ids.ctypes.data_as(ctypes.POINTER(ctypes.c_uint64)),
+                    vector_ids.ctypes.data_as(ctypes.POINTER(ctypes.c_uint64)),
+                    scores.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
+                    ctypes.byref(count),
+                    ctypes.byref(timings),
+                )
+            )
+        return [
+            PersonHit(
+                person_numeric_id=int(group_ids[offset]),
+                vector_id=int(vector_ids[offset]),
+                cosine=float(np.clip(scores[offset], -1.0, 1.0)),
+            )
+            for offset in range(int(count.value))
+        ]
+
+    def stats(self) -> IndexStats:
+        raw = _Stats()
+        raw.struct_size = ctypes.sizeof(_Stats)
+        with self._lock.read():
+            self._ensure_open()
+            self.library._check(
+                self.library._library.ifs_search_get_stats(self._handle, ctypes.byref(raw))
+            )
+        return IndexStats(
+            profile=self.profile,
+            dimension=DIMENSION,
+            live_rows=int(raw.live_rows),
+            physical_rows=int(raw.physical_rows),
+            capacity_rows=int(raw.max_rows or raw.capacity_rows),
+            tombstone_rows=int(raw.tombstone_rows),
+            reallocations=int(raw.reallocations),
+        )
+
+    def close(self) -> None:
+        with self._lock.write():
+            if self._closed:
+                return
+            handle = self._handle
+            self._handle = ctypes.c_void_p()
+            self._closed = True
+            if handle.value:
+                self.library._library.ifs_search_destroy(handle)
diff --git a/server/backend/insightface_server/search/reference.py b/server/backend/insightface_server/search/reference.py
new file mode 100644
index 000000000..9ea539d76
--- /dev/null
+++ b/server/backend/insightface_server/search/reference.py
@@ -0,0 +1,262 @@
+from __future__ import annotations
+
+import numpy as np
+
+from .base import (
+    SEARCH_PROFILES,
+    IndexRecord,
+    IndexStats,
+    PersonHit,
+    SearchIndexCapacityError,
+    SearchIndexStateError,
+)
+from .synchronization import ReadWriteLock
+
+INT8_PROFILE_SCALES = {
+    "int8_x1000_v1": 1000,
+    "int8_x736_v1": 736,
+}
+
+
+def _validated_embedding(value: np.ndarray, dimension: int) -> np.ndarray:
+    vector = np.asarray(value, dtype=np.float32).reshape(-1)
+    if vector.shape != (dimension,):
+        raise ValueError(f"embedding must have shape ({dimension},)")
+    if not np.all(np.isfinite(vector)):
+        raise ValueError("embedding contains non-finite values")
+    norm = float(np.linalg.norm(vector))
+    if abs(norm - 1.0) > 2e-4:
+        raise ValueError(f"embedding must be L2-normalized; norm={norm:.8f}")
+    return np.ascontiguousarray(vector)
+
+
+def _round_half_away_from_zero(value: np.ndarray) -> np.ndarray:
+    return np.where(value >= 0, np.floor(value + 0.5), np.ceil(value - 0.5))
+
+
+def quantize_int8(value: np.ndarray, scale: int) -> np.ndarray:
+    if scale <= 0:
+        raise ValueError("INT8 scale must be positive")
+    scaled = _round_half_away_from_zero(
+        np.asarray(value, dtype=np.float32) * np.float32(scale)
+    )
+    return np.clip(scaled, -128, 127).astype(np.int8)
+
+
+def quantize_int8_x1000(value: np.ndarray) -> np.ndarray:
+    return quantize_int8(value, 1000)
+
+
+def quantize_int8_x736(value: np.ndarray) -> np.ndarray:
+    return quantize_int8(value, 736)
+
+
+def quantize_bf16_to_fp32(value: np.ndarray) -> np.ndarray:
+    """Round FP32 to BF16 (nearest-even) while retaining a NumPy FP32 container."""
+
+    contiguous = np.ascontiguousarray(value, dtype=np.float32)
+    bits = contiguous.view(np.uint32).copy()
+    rounding = np.uint32(0x7FFF) + ((bits >> np.uint32(16)) & np.uint32(1))
+    bits = (bits + rounding) & np.uint32(0xFFFF0000)
+    return bits.view(np.float32)
+
+
+def profile_similarity(profile: str, left: np.ndarray, right: np.ndarray) -> float:
+    """Return the profile's unclipped raw similarity for two normalized vectors.
+
+    This is the scalar semantic reference used by enrollment review. It applies
+    the same storage/query conversion and accumulator scale as the selected
+    search profile, but deliberately does not clamp or round the returned score.
+    """
+
+    if profile not in SEARCH_PROFILES:
+        raise ValueError(f"unsupported search profile: {profile}")
+    left_value = np.asarray(left, dtype=np.float32).reshape(-1)
+    right_value = np.asarray(right, dtype=np.float32).reshape(-1)
+    if left_value.size == 0 or left_value.shape != right_value.shape:
+        raise ValueError("embeddings must have the same non-empty shape")
+    left_value = _validated_embedding(left_value, left_value.size)
+    right_value = _validated_embedding(right_value, right_value.size)
+
+    scale = INT8_PROFILE_SCALES.get(profile)
+    if scale is not None:
+        left_int = quantize_int8(left_value, scale).astype(np.int32)
+        right_int = quantize_int8(right_value, scale).astype(np.int32)
+        accumulator = int(np.dot(left_int, right_int))
+        return accumulator / float(scale * scale)
+    if profile == "fp16_v1":
+        left_value = left_value.astype(np.float16).astype(np.float32)
+        right_value = right_value.astype(np.float16).astype(np.float32)
+    elif profile == "bf16_v1":
+        left_value = quantize_bf16_to_fp32(left_value)
+        right_value = quantize_bf16_to_fp32(right_value)
+    return float(np.dot(left_value, right_value))
+
+
+class ReferenceSearchIndex:
+    """Deterministic exact grouped Flat-IP reference used by tests and shadowing.
+
+    This implementation deliberately favors a small, auditable contract over
+    throughput. Production images use the native CPU/CUDA implementations.
+    """
+
+    def __init__(self, *, profile: str, dimension: int, capacity_rows: int) -> None:
+        if profile not in SEARCH_PROFILES:
+            raise ValueError(f"unsupported search profile: {profile}")
+        if dimension <= 0:
+            raise ValueError("dimension must be positive")
+        if capacity_rows <= 0:
+            raise ValueError("capacity_rows must be positive")
+        self.profile = profile
+        self.dimension = dimension
+        self._capacity = capacity_rows
+        self._lock = ReadWriteLock()
+        self._closed = False
+        self._vector_ids = np.zeros(capacity_rows, dtype=np.uint64)
+        self._person_ids = np.zeros(capacity_rows, dtype=np.uint64)
+        self._alive = np.zeros(capacity_rows, dtype=np.bool_)
+        if profile in INT8_PROFILE_SCALES:
+            self._vectors = np.zeros((capacity_rows, dimension), dtype=np.int8)
+        elif profile == "fp16_v1":
+            self._vectors = np.zeros((capacity_rows, dimension), dtype=np.float16)
+        else:
+            self._vectors = np.zeros((capacity_rows, dimension), dtype=np.float32)
+        self._id_to_slot: dict[int, int] = {}
+        self._free_slots: list[int] = []
+        self._physical_rows = 0
+
+    def _ensure_open(self) -> None:
+        if self._closed:
+            raise SearchIndexStateError("search index is closed")
+
+    def _encode(self, value: np.ndarray) -> np.ndarray:
+        int8_scale = INT8_PROFILE_SCALES.get(self.profile)
+        if int8_scale is not None:
+            return quantize_int8(value, int8_scale)
+        if self.profile == "fp16_v1":
+            return value.astype(np.float16)
+        if self.profile == "bf16_v1":
+            return quantize_bf16_to_fp32(value)
+        return value.astype(np.float32, copy=False)
+
+    def add_batch(self, records: list[IndexRecord]) -> None:
+        if not records:
+            return
+        with self._lock.write():
+            self._ensure_open()
+            ids = [int(record.vector_id) for record in records]
+            if any(value <= 0 or value > np.iinfo(np.int64).max for value in ids):
+                raise ValueError("vector_id must be within 1..2^63-1")
+            if any(int(record.person_numeric_id) <= 0 for record in records):
+                raise ValueError("person_numeric_id must be positive")
+            if len(ids) != len(set(ids)) or any(value in self._id_to_slot for value in ids):
+                raise ValueError("active vector_id values must be unique")
+            live_rows = len(self._id_to_slot)
+            if live_rows + len(records) > self._capacity:
+                raise SearchIndexCapacityError(
+                    f"index capacity {self._capacity} would be exceeded"
+                )
+            encoded = [
+                self._encode(_validated_embedding(record.embedding, self.dimension))
+                for record in records
+            ]
+            available_new = self._capacity - self._physical_rows
+            required_new = max(0, len(records) - len(self._free_slots))
+            if required_new > available_new:
+                raise SearchIndexCapacityError("no physical slots are available")
+            for record, vector in zip(records, encoded, strict=True):
+                if self._free_slots:
+                    slot = self._free_slots.pop()
+                else:
+                    slot = self._physical_rows
+                    self._physical_rows += 1
+                vector_id = int(record.vector_id)
+                self._vectors[slot] = vector
+                self._vector_ids[slot] = vector_id
+                self._person_ids[slot] = int(record.person_numeric_id)
+                self._alive[slot] = True
+                self._id_to_slot[vector_id] = slot
+
+    def remove_batch(self, vector_ids: list[int]) -> set[int]:
+        removed: set[int] = set()
+        with self._lock.write():
+            self._ensure_open()
+            for value in vector_ids:
+                vector_id = int(value)
+                slot = self._id_to_slot.pop(vector_id, None)
+                if slot is None:
+                    continue
+                self._alive[slot] = False
+                self._free_slots.append(slot)
+                removed.add(vector_id)
+        return removed
+
+    def _scores(self, query: np.ndarray, slots: np.ndarray) -> np.ndarray:
+        int8_scale = INT8_PROFILE_SCALES.get(self.profile)
+        if int8_scale is not None:
+            encoded_query = quantize_int8(query, int8_scale).astype(np.int32)
+            matrix = self._vectors[slots].astype(np.int32)
+            values = matrix @ encoded_query
+            return values.astype(np.float32) / np.float32(int8_scale * int8_scale)
+        if self.profile == "fp16_v1":
+            encoded_query = query.astype(np.float16).astype(np.float32)
+            matrix = self._vectors[slots].astype(np.float32)
+        elif self.profile == "bf16_v1":
+            encoded_query = quantize_bf16_to_fp32(query)
+            matrix = self._vectors[slots]
+        else:
+            encoded_query = query
+            matrix = self._vectors[slots]
+        return np.asarray(matrix @ encoded_query, dtype=np.float32)
+
+    def search_persons(self, query: np.ndarray, limit: int) -> list[PersonHit]:
+        if limit <= 0:
+            raise ValueError("limit must be positive")
+        normalized = _validated_embedding(query, self.dimension)
+        with self._lock.read():
+            self._ensure_open()
+            slots = np.flatnonzero(self._alive[: self._physical_rows])
+            if slots.size == 0:
+                return []
+            scores = self._scores(normalized, slots)
+            best: dict[int, tuple[float, int]] = {}
+            for slot, raw_score in zip(slots.tolist(), scores.tolist(), strict=True):
+                person_id = int(self._person_ids[slot])
+                vector_id = int(self._vector_ids[slot])
+                score = float(raw_score)
+                current = best.get(person_id)
+                if current is None or score > current[0] or (
+                    score == current[0] and vector_id < current[1]
+                ):
+                    best[person_id] = (score, vector_id)
+            ordered = [
+                (person_id, value[1], value[0]) for person_id, value in best.items()
+            ]
+            ordered.sort(key=lambda hit: (-hit[2], hit[0], hit[1]))
+            return [
+                PersonHit(
+                    person_numeric_id=person_id,
+                    vector_id=vector_id,
+                    cosine=float(np.clip(score, -1.0, 1.0)),
+                )
+                for person_id, vector_id, score in ordered[:limit]
+            ]
+
+    def stats(self) -> IndexStats:
+        with self._lock.read():
+            self._ensure_open()
+            live = len(self._id_to_slot)
+            return IndexStats(
+                profile=self.profile,
+                dimension=self.dimension,
+                live_rows=live,
+                physical_rows=self._physical_rows,
+                capacity_rows=self._capacity,
+                tombstone_rows=self._physical_rows - live,
+            )
+
+    def close(self) -> None:
+        with self._lock.write():
+            self._closed = True
+            self._id_to_slot.clear()
+            self._free_slots.clear()
diff --git a/server/backend/insightface_server/search/synchronization.py b/server/backend/insightface_server/search/synchronization.py
new file mode 100644
index 000000000..4982f52d2
--- /dev/null
+++ b/server/backend/insightface_server/search/synchronization.py
@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+import threading
+from collections.abc import Iterator
+from contextlib import contextmanager
+
+
+class ReadWriteLock:
+    """Small writer-preferring lock for immutable index-generation reads."""
+
+    def __init__(self) -> None:
+        self._condition = threading.Condition()
+        self._readers = 0
+        self._reader_depths: dict[int, int] = {}
+        self._writer = False
+        self._writer_owner: int | None = None
+        self._writer_depth = 0
+        self._waiting_writers = 0
+
+    def acquire_read(self) -> None:
+        owner = threading.get_ident()
+        with self._condition:
+            existing_depth = self._reader_depths.get(owner, 0)
+            while (
+                self._writer_owner not in {None, owner}
+                or (self._waiting_writers and existing_depth == 0 and not self._writer)
+            ):
+                self._condition.wait()
+            self._readers += 1
+            self._reader_depths[owner] = existing_depth + 1
+
+    def release_read(self) -> None:
+        owner = threading.get_ident()
+        with self._condition:
+            depth = self._reader_depths.get(owner, 0)
+            if self._readers <= 0 or depth <= 0:
+                raise RuntimeError("read lock released without being acquired")
+            self._readers -= 1
+            if depth == 1:
+                self._reader_depths.pop(owner)
+            else:
+                self._reader_depths[owner] = depth - 1
+            if self._readers == 0:
+                self._condition.notify_all()
+
+    def acquire_write(self) -> None:
+        owner = threading.get_ident()
+        with self._condition:
+            if self._writer_owner == owner:
+                self._writer_depth += 1
+                return
+            self._waiting_writers += 1
+            try:
+                while self._writer or self._readers:
+                    self._condition.wait()
+                self._writer = True
+                self._writer_owner = owner
+                self._writer_depth = 1
+            finally:
+                self._waiting_writers -= 1
+
+    def release_write(self) -> None:
+        owner = threading.get_ident()
+        with self._condition:
+            if not self._writer or self._writer_owner != owner:
+                raise RuntimeError("write lock released without being acquired")
+            self._writer_depth -= 1
+            if self._writer_depth:
+                return
+            self._writer = False
+            self._writer_owner = None
+            self._condition.notify_all()
+
+    @contextmanager
+    def read(self) -> Iterator[None]:
+        self.acquire_read()
+        try:
+            yield
+        finally:
+            self.release_read()
+
+    @contextmanager
+    def write(self) -> Iterator[None]:
+        self.acquire_write()
+        try:
+            yield
+        finally:
+            self.release_write()
diff --git a/server/backend/insightface_server/services/__init__.py b/server/backend/insightface_server/services/__init__.py
new file mode 100644
index 000000000..75151d341
--- /dev/null
+++ b/server/backend/insightface_server/services/__init__.py
@@ -0,0 +1,26 @@
+from .core import FaceService
+from .images import ImageData, ImageLoader
+from .rtsp import (
+    MonitorAlreadyExistsError,
+    MonitorLimitError,
+    MonitorManager,
+    MonitorNotFoundError,
+    MonitorOptions,
+    MonitorPreviewDisabledError,
+    MonitorSession,
+    MonitorUnavailableError,
+)
+
+__all__ = [
+    "FaceService",
+    "ImageData",
+    "ImageLoader",
+    "MonitorAlreadyExistsError",
+    "MonitorLimitError",
+    "MonitorManager",
+    "MonitorNotFoundError",
+    "MonitorOptions",
+    "MonitorPreviewDisabledError",
+    "MonitorSession",
+    "MonitorUnavailableError",
+]
diff --git a/server/backend/insightface_server/services/core.py b/server/backend/insightface_server/services/core.py
new file mode 100644
index 000000000..bca115213
--- /dev/null
+++ b/server/backend/insightface_server/services/core.py
@@ -0,0 +1,944 @@
+from __future__ import annotations
+
+import logging
+import sqlite3
+import uuid
+from typing import Any
+
+import numpy as np
+
+from ..api.schemas import EmbeddingMode, ReviewMode
+from ..config import DetectionProfile, Settings
+from ..errors import ApiError, bad_request, conflict, not_found, unprocessable
+from ..inference import FaceObservation, InferenceEngine
+from ..search import (
+    SearchIndexCapacityError,
+    SearchIndexError,
+    SearchIndexManager,
+    SearchMutationCommittedError,
+    profile_similarity,
+)
+from ..storage import (
+    CollectionCapacityExceeded,
+    CursorCodec,
+    FaceCropStore,
+    MaxFacesPerPersonExceeded,
+    Repository,
+    utc_now,
+)
+from .images import ImageData
+from .presentation import bounding_box, clamp01, face_result, quality
+
+LOGGER = logging.getLogger("insightface_server.service")
+CENTER_LARGEST_DISTANCE_WEIGHT = 2.0
+
+
+def similarity(left: np.ndarray, right: np.ndarray) -> float:
+    return float(np.clip(np.dot(left, right), -1.0, 1.0))
+
+
+class FaceService:
+    def __init__(
+        self,
+        settings: Settings,
+        repository: Repository,
+        cursors: CursorCodec,
+        crops: FaceCropStore,
+        engine: InferenceEngine,
+        search_indexes: SearchIndexManager,
+    ):
+        self.settings = settings
+        self.repository = repository
+        self.cursors = cursors
+        self.crops = crops
+        self.engine = engine
+        self.search_indexes = search_indexes
+
+    @staticmethod
+    def _search_unavailable(exc: Exception) -> ApiError:
+        details: dict[str, Any] = {}
+        if isinstance(exc, SearchMutationCommittedError):
+            details = {
+                "write_committed": True,
+                "collection_id": exc.collection_id,
+                "revision": exc.revision,
+            }
+        return ApiError(
+            "search_index_unavailable",
+            "The Collection search index is unavailable and is being rebuilt.",
+            503,
+            details,
+        )
+
+    @staticmethod
+    def _quota_error(exc: Exception) -> ApiError:
+        if isinstance(exc, MaxFacesPerPersonExceeded):
+            return ApiError(
+                "person_face_limit_exceeded",
+                "The Person has reached the configured FaceSample limit.",
+                409,
+                {
+                    "max_faces": exc.max_faces,
+                    "requested_faces": exc.requested_faces,
+                },
+            )
+        if isinstance(exc, CollectionCapacityExceeded):
+            return ApiError(
+                "collection_capacity_exceeded",
+                "The Collection has reached its configured search capacity.",
+                409,
+                {
+                    "capacity_rows": exc.capacity_rows,
+                    "requested_rows": exc.requested_rows,
+                },
+            )
+        return conflict(
+            "The Collection has reached its configured search capacity.",
+            code="collection_capacity_exceeded",
+        )
+
+    def _delete_crop_paths(self, paths: list[str | None]) -> None:
+        for path in paths:
+            try:
+                self.crops.delete(path)
+            except OSError:
+                LOGGER.warning("unable_to_remove_orphan_face_crop", exc_info=True)
+
+    def raw_collection(self, collection_id: str) -> dict[str, Any]:
+        """Return Collection metadata without enforcing the active model contract."""
+
+        item = self.repository.get_collection(collection_id)
+        if item is None:
+            raise not_found("collection", collection_id)
+        return item
+
+    def collection(self, collection_id: str) -> dict[str, Any]:
+        item = self.raw_collection(collection_id)
+        summary = self.engine.summary
+        expected = (
+            summary.model_id,
+            summary.model_version,
+            summary.model_digest,
+            summary.embedding_dimension,
+            summary.preprocessing_version,
+        )
+        actual = (
+            item["model_id"],
+            item["model_version"],
+            item["model_digest"],
+            item["embedding_dimension"],
+            item["preprocessing_version"],
+        )
+        if expected != actual:
+            raise ApiError(
+                "collection_model_mismatch",
+                "The collection is bound to a different model; rebuild or migrate it.",
+                409,
+            )
+        return item
+
+    @staticmethod
+    def collection_detection_profile(collection: dict[str, Any]) -> DetectionProfile:
+        return DetectionProfile.from_mapping(collection.get("detection"))
+
+    def detection_profile(self, collection_id: str | None = None) -> DetectionProfile:
+        if collection_id is None:
+            return self.settings.detection_profile
+        return self.collection_detection_profile(self.collection(collection_id))
+
+    def analyze(
+        self,
+        image: ImageData,
+        *,
+        embeddings: bool = True,
+        detection_profile: DetectionProfile | None = None,
+    ) -> list[FaceObservation]:
+        profile = detection_profile or self.settings.detection_profile
+        faces = self.engine.analyze(
+            image.pixels,
+            require_embeddings=embeddings,
+            detection_profile=profile,
+        )
+        return sorted(faces, key=lambda face: -face.area)
+
+    def detect(
+        self,
+        image: ImageData,
+        *,
+        max_faces: int | None,
+        detection_profile: DetectionProfile | None = None,
+    ) -> list[dict[str, Any]]:
+        faces = self.analyze(
+            image, embeddings=False, detection_profile=detection_profile
+        )
+        limit = self.settings.max_detected_faces
+        if max_faces is not None:
+            limit = min(limit, max_faces)
+        faces = faces[:limit]
+        height, width = image.pixels.shape[:2]
+        return [face_result(face, width, height) for face in faces]
+
+    @staticmethod
+    def _selection_score(
+        face: FaceObservation,
+        *,
+        image_width: int,
+        image_height: int,
+        strategy: str,
+    ) -> float:
+        left, top, right, bottom = face.bbox
+        if strategy == "largest":
+            return face.area
+        if strategy != "center_largest":
+            raise ValueError(f"Unsupported single-face selection strategy: {strategy}")
+        center_x = (left + right) * 0.5
+        center_y = (top + bottom) * 0.5
+        distance_squared = (
+            (center_x - float(image_width) * 0.5) ** 2
+            + (center_y - float(image_height) * 0.5) ** 2
+        )
+        # This follows the long-standing InsightFace/FaceAnalysis selection
+        # metric. Detection confidence is intentionally not part of the score.
+        return face.area - CENTER_LARGEST_DISTANCE_WEIGHT * distance_squared
+
+    def select_from_observations(
+        self,
+        faces: list[FaceObservation],
+        image: ImageData,
+        profile: DetectionProfile,
+    ) -> FaceObservation:
+        if not faces:
+            raise ValueError("faces must not be empty")
+        height, width = image.pixels.shape[:2]
+        # max() performs one linear scan and, like numpy.argmax(), keeps the
+        # first observation when scores are exactly equal.
+        return max(
+            faces,
+            key=lambda face: self._selection_score(
+                face,
+                image_width=width,
+                image_height=height,
+                strategy=profile.single_face_selection,
+            ),
+        )
+
+    def selected_face(
+        self, image: ImageData, *, detection_profile: DetectionProfile | None = None
+    ) -> FaceObservation:
+        profile = detection_profile or self.settings.detection_profile
+        faces = self.analyze(
+            image, embeddings=True, detection_profile=profile
+        )
+        if not faces:
+            raise unprocessable("No usable face was detected.", code="face_not_found")
+        return self.select_from_observations(faces, image, profile)
+
+    def largest_face(self, image: ImageData) -> FaceObservation:
+        """Backward-compatible internal alias using the system profile."""
+
+        return self.selected_face(image, detection_profile=self.settings.detection_profile)
+
+    def _registration_reason(
+        self,
+        image: ImageData,
+        faces: list[FaceObservation],
+        *,
+        review_mode: ReviewMode,
+        embedding_dimension: int,
+        embedding: np.ndarray | None = None,
+        use_supplied_embedding: bool = False,
+        require_landmarks: bool = False,
+        invalid_embedding_reason: str = "invalid_embedding",
+    ) -> str | None:
+        if not faces:
+            return "face_not_found"
+        # Off is the low-friction enrollment mode: when several faces are
+        # present, analyze() has already sorted them by area and we register the
+        # largest one. Standard and strict retain the unambiguous single-face
+        # requirement before applying their additional review rules.
+        if review_mode != "off" and len(faces) != 1:
+            return "multiple_faces"
+        face = faces[0]
+        if require_landmarks:
+            if face.landmarks is None:
+                return "invalid_landmarks"
+            landmarks = np.asarray(face.landmarks, dtype=np.float32)
+            if landmarks.shape != (5, 2) or not np.all(np.isfinite(landmarks)):
+                return "invalid_landmarks"
+        # In trusted mode ``None`` means the supplied vector was invalid. Never
+        # fall back to an embedding a misbehaving engine may have returned while
+        # require_embeddings=False.
+        candidate = embedding if use_supplied_embedding else face.embedding
+        if candidate is None:
+            return invalid_embedding_reason
+        candidate = np.asarray(candidate, dtype=np.float32).reshape(-1)
+        if (
+            candidate.shape != (embedding_dimension,)
+            or not np.all(np.isfinite(candidate))
+            or abs(float(np.linalg.norm(candidate)) - 1.0) > 2e-4
+        ):
+            return invalid_embedding_reason
+        if review_mode == "off":
+            return None
+        width = face.bbox[2] - face.bbox[0]
+        height = face.bbox[3] - face.bbox[1]
+        if min(width, height) < self.settings.registration_min_face_size:
+            return "face_too_small"
+        if face.confidence < self.settings.registration_min_score:
+            return "low_detection_score"
+        if any(abs(float(getattr(face, field, 0.0))) > 45.0 for field in ("yaw", "pitch", "roll")):
+            return "extreme_pose"
+        if quality(face)["score"] < self.settings.registration_min_quality:
+            return "low_quality"
+        return None
+
+    @staticmethod
+    def _trusted_embedding(value: Any, embedding_dimension: int) -> np.ndarray | None:
+        """Validate and canonically normalize one trusted JSON embedding.
+
+        JSON has no native float32 scalar type, so the public contract means the
+        values must be scalar JSON numbers that are representable as a finite
+        float32 vector.  The caller already trusts image/embedding identity; this
+        function deliberately performs no recognition inference or identity check.
+        """
+
+        if not isinstance(value, list) or len(value) != embedding_dimension:
+            return None
+        if any(isinstance(item, bool) or not isinstance(item, int | float) for item in value):
+            return None
+        try:
+            with np.errstate(over="ignore", invalid="ignore"):
+                embedding = np.asarray(value, dtype=np.float32)
+        except (TypeError, ValueError, OverflowError):
+            return None
+        if embedding.shape != (embedding_dimension,) or not np.all(np.isfinite(embedding)):
+            return None
+        with np.errstate(over="ignore", invalid="ignore"):
+            norm = float(np.linalg.norm(embedding))
+        if not np.isfinite(norm) or norm <= 1e-12 or abs(norm - 1.0) > 2e-4:
+            return None
+        # Normalize even an already-valid input to remove harmless serialization
+        # error before SQLite persistence and search-profile quantization.
+        return (embedding / norm).astype(np.float32, copy=False)
+
+    @staticmethod
+    def _validate_embedding_mode(
+        collection: dict[str, Any],
+        *,
+        embedding_mode: EmbeddingMode,
+        external_embeddings: list[Any] | None,
+        embedding_contract_id: str | None,
+        image_count: int,
+    ) -> None:
+        if embedding_mode == "server":
+            if external_embeddings is not None or embedding_contract_id is not None:
+                raise bad_request(
+                    "External embedding fields require embedding_mode=external_trusted.",
+                    code="unexpected_external_embedding",
+                )
+            return
+        if external_embeddings is None:
+            raise bad_request(
+                "external_embeddings is required for external_trusted enrollment.",
+                code="missing_external_embeddings",
+            )
+        if len(external_embeddings) != image_count:
+            raise bad_request(
+                "external_embeddings must contain exactly one vector for each image.",
+                code="external_embedding_count_mismatch",
+                image_count=image_count,
+                embedding_count=len(external_embeddings),
+            )
+        if not embedding_contract_id:
+            raise bad_request(
+                "embedding_contract_id is required for external_trusted enrollment.",
+                code="missing_embedding_contract_id",
+            )
+        expected = str(collection["embedding_contract_id"])
+        if embedding_contract_id != expected:
+            raise ApiError(
+                "embedding_contract_mismatch",
+                "The external embedding contract does not match this Collection.",
+                409,
+                {
+                    "expected_embedding_contract_id": expected,
+                    "provided_embedding_contract_id": embedding_contract_id,
+                },
+            )
+
+    def prepare_registration(
+        self,
+        collection: dict[str, Any],
+        images: list[ImageData],
+        *,
+        review_mode: ReviewMode,
+        embedding_mode: EmbeddingMode = "server",
+        external_embeddings: list[Any] | None = None,
+        embedding_contract_id: str | None = None,
+    ) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+        self._validate_embedding_mode(
+            collection,
+            embedding_mode=embedding_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+            image_count=len(images),
+        )
+        accepted: list[dict[str, Any]] = []
+        rejected: list[dict[str, Any]] = []
+        detection_profile = self.collection_detection_profile(collection)
+        try:
+            for index, image in enumerate(images):
+                trusted_embedding = None
+                if embedding_mode == "external_trusted":
+                    assert external_embeddings is not None
+                    trusted_embedding = self._trusted_embedding(
+                        external_embeddings[index],
+                        int(collection["embedding_dimension"]),
+                    )
+                faces = self.analyze(
+                    image,
+                    # Trusted enrollment still runs detector, landmarks and
+                    # quality but never invokes the recognizer.
+                    embeddings=embedding_mode == "server",
+                    detection_profile=detection_profile,
+                )
+                if review_mode == "off":
+                    faces = (
+                        [
+                            self.select_from_observations(
+                                faces,
+                                image,
+                                detection_profile,
+                            )
+                        ]
+                        if faces
+                        else []
+                    )
+                reason = self._registration_reason(
+                    image,
+                    faces,
+                    review_mode=review_mode,
+                    embedding_dimension=int(collection["embedding_dimension"]),
+                    embedding=trusted_embedding,
+                    use_supplied_embedding=embedding_mode == "external_trusted",
+                    require_landmarks=embedding_mode == "external_trusted",
+                    invalid_embedding_reason=(
+                        "invalid_external_embedding"
+                        if embedding_mode == "external_trusted"
+                        else "invalid_embedding"
+                    ),
+                )
+                if reason:
+                    rejected.append({"index": index, "filename": image.filename, "reason": reason})
+                    continue
+                face = faces[0]
+                final_embedding = (
+                    trusted_embedding if embedding_mode == "external_trusted" else face.embedding
+                )
+                assert final_embedding is not None
+                face_id = str(uuid.uuid4())
+                height, width = image.pixels.shape[:2]
+                crop_image = (
+                    self.crops.encode(image.pixels, face.bbox)
+                    if bool(collection.get("save_face_crops", False))
+                    else None
+                )
+                accepted.append(
+                    {
+                        "id": face_id,
+                        "embedding": final_embedding,
+                        "embedding_dimension": collection["embedding_dimension"],
+                        "bounding_box": bounding_box(face, width, height),
+                        "landmarks": (
+                            face.landmarks.tolist() if face.landmarks is not None else None
+                        ),
+                        "detection_score": clamp01(face.confidence),
+                        "quality": quality(face),
+                        "model_id": collection["model_id"],
+                        "model_version": collection["model_version"],
+                        "model_digest": collection["model_digest"],
+                        "preprocessing_version": collection["preprocessing_version"],
+                        "embedding_source": embedding_mode,
+                        "embedding_contract_id": collection["embedding_contract_id"],
+                        # New FaceSamples keep their optional crop in SQLite. The
+                        # path field remains null while legacy path-backed crops
+                        # continue to be readable and are not migrated implicitly.
+                        "crop_path": None,
+                        "crop_image": crop_image,
+                        "crop_media_type": "image/jpeg" if crop_image is not None else None,
+                        "created_at": utc_now(),
+                        "_registration_index": index,
+                        "_registration_filename": image.filename,
+                    }
+                )
+        except Exception:
+            for item in accepted:
+                try:
+                    self.crops.delete(item.get("crop_path"))
+                except OSError:
+                    LOGGER.warning("unable_to_remove_orphan_face_crop", exc_info=True)
+            raise
+        return accepted, rejected
+
+    def _strict_registration_review(
+        self,
+        collection: dict[str, Any],
+        person_id: str,
+        faces: list[dict[str, Any]],
+        existing_embeddings: list[np.ndarray],
+    ) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+        """Apply deterministic class-in versus class-out enrollment review."""
+
+        accepted: list[dict[str, Any]] = []
+        rejected: list[dict[str, Any]] = []
+        same_embeddings = list(existing_embeddings)
+        profile = str(collection["search_profile"])
+        for face in faces:
+            embedding = np.asarray(face["embedding"], dtype=np.float32)
+            # The first usable sample bootstraps an empty Person. Samples later in
+            # this same multipart request can use it as a class-in reference.
+            if not same_embeddings:
+                accepted.append(face)
+                same_embeddings.append(embedding)
+                continue
+
+            max_same = max(
+                profile_similarity(profile, embedding, reference) for reference in same_embeddings
+            )
+            try:
+                other = self.search_indexes.best_other_person(
+                    str(collection["id"]),
+                    embedding,
+                    exclude_person_id=person_id,
+                )
+            except SearchIndexError as exc:
+                raise self._search_unavailable(exc) from exc
+
+            if other is not None and max_same <= float(other["similarity"]):
+                self._delete_crop_paths([face.get("crop_path")])
+                rejected.append(
+                    {
+                        "index": int(face["_registration_index"]),
+                        "filename": str(face["_registration_filename"]),
+                        "reason": "identity_similarity_conflict",
+                        "same_person_similarity": float(max_same),
+                        "other_person_similarity": float(other["similarity"]),
+                        "other_person_id": str(other["person_id"]),
+                        "matched_face_id": str(other["face_id"]),
+                    }
+                )
+                continue
+
+            accepted.append(face)
+            same_embeddings.append(embedding)
+        return accepted, rejected
+
+    @staticmethod
+    def public_face(face: dict[str, Any]) -> dict[str, Any]:
+        result = {
+            key: face[key]
+            for key in (
+                "id",
+                "person_id",
+                "bounding_box",
+                "landmarks",
+                "detection_score",
+                "quality",
+                "model_id",
+                "model_version",
+                "model_digest",
+                "preprocessing_version",
+                "embedding_source",
+                "embedding_contract_id",
+                "created_at",
+            )
+            if key in face
+        }
+        result["has_crop"] = bool(
+            face.get(
+                "has_crop",
+                face.get("crop_image") is not None or face.get("crop_path"),
+            )
+        )
+        return result
+
+    def create_person(
+        self,
+        collection_id: str,
+        person: dict[str, Any],
+        images: list[ImageData],
+        *,
+        review_mode: ReviewMode = "off",
+        embedding_mode: EmbeddingMode = "server",
+        external_embeddings: list[Any] | None = None,
+        embedding_contract_id: str | None = None,
+    ) -> tuple[dict[str, Any], list[dict[str, Any]], list[dict[str, Any]]]:
+        collection = self.collection(collection_id)
+        faces, rejected = self.prepare_registration(
+            collection,
+            images,
+            review_mode=review_mode,
+            embedding_mode=embedding_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+        )
+        if not faces:
+            raise unprocessable(
+                "None of the uploaded images contained one registrable face.",
+                code="registration_failed",
+                rejected_images=rejected,
+            )
+        try:
+            if review_mode == "strict":
+                # Keep the identity decision and write under one Collection
+                # lock, so another enrollment cannot change the class-out
+                # maximum between review and commit.
+                def strict_create():
+                    reviewed, strict_rejected = self._strict_registration_review(
+                        collection, str(person["id"]), faces, []
+                    )
+                    created_item, mutation = self.repository.create_person_with_faces_with_mutation(
+                        collection_id, person, reviewed
+                    )
+                    return (created_item, reviewed, strict_rejected), mutation
+
+                created, faces, strict_rejected = self.search_indexes.run_mutation(
+                    collection_id,
+                    strict_create,
+                    # The Repository checks capacity using only rows that pass
+                    # strict review; the native index has the same hard limit.
+                    expected_additions=0,
+                )
+                rejected.extend(strict_rejected)
+            else:
+                created = self.search_indexes.run_mutation(
+                    collection_id,
+                    lambda: self.repository.create_person_with_faces_with_mutation(
+                        collection_id, person, faces
+                    ),
+                    expected_additions=len(faces),
+                )
+        except SearchMutationCommittedError as exc:
+            raise self._search_unavailable(exc) from exc
+        except KeyError as exc:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise not_found("collection", collection_id) from exc
+        except (CollectionCapacityExceeded, MaxFacesPerPersonExceeded) as exc:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise self._quota_error(exc) from exc
+        except SearchIndexCapacityError as exc:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise self._quota_error(exc) from exc
+        except SearchIndexError as exc:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise self._search_unavailable(exc) from exc
+        except sqlite3.IntegrityError as exc:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise conflict(
+                "The person ID or external ID already exists.", code="person_exists"
+            ) from exc
+        except Exception:
+            self._delete_crop_paths([face.get("crop_path") for face in faces])
+            raise
+        return created, [self.public_face(face) for face in faces], rejected
+
+    def add_faces(
+        self,
+        collection_id: str,
+        person_id: str,
+        images: list[ImageData],
+        *,
+        review_mode: ReviewMode = "off",
+        embedding_mode: EmbeddingMode = "server",
+        external_embeddings: list[Any] | None = None,
+        embedding_contract_id: str | None = None,
+    ) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+        collection = self.collection(collection_id)
+        if self.repository.get_person(collection_id, person_id) is None:
+            raise not_found("person", person_id)
+        faces, rejected = self.prepare_registration(
+            collection,
+            images,
+            review_mode=review_mode,
+            embedding_mode=embedding_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+        )
+        if faces:
+            try:
+                if review_mode == "strict":
+
+                    def strict_add():
+                        if self.repository.get_person(collection_id, person_id) is None:
+                            raise KeyError(person_id)
+                        reviewed, strict_rejected = self._strict_registration_review(
+                            collection,
+                            person_id,
+                            faces,
+                            self.repository.list_face_embeddings(collection_id, person_id),
+                        )
+                        mutation = (
+                            self.repository.add_faces_with_mutation(
+                                collection_id, person_id, reviewed
+                            )
+                            if reviewed
+                            else None
+                        )
+                        return (reviewed, strict_rejected), mutation
+
+                    faces, strict_rejected = self.search_indexes.run_mutation(
+                        collection_id,
+                        strict_add,
+                        expected_additions=0,
+                    )
+                    rejected.extend(strict_rejected)
+                else:
+                    self.search_indexes.run_mutation(
+                        collection_id,
+                        lambda: (
+                            None,
+                            self.repository.add_faces_with_mutation(
+                                collection_id, person_id, faces
+                            ),
+                        ),
+                        expected_additions=len(faces),
+                    )
+            except SearchMutationCommittedError as exc:
+                raise self._search_unavailable(exc) from exc
+            except KeyError as exc:
+                self._delete_crop_paths([face.get("crop_path") for face in faces])
+                raise not_found("person", person_id) from exc
+            except (CollectionCapacityExceeded, MaxFacesPerPersonExceeded) as exc:
+                self._delete_crop_paths([face.get("crop_path") for face in faces])
+                raise self._quota_error(exc) from exc
+            except SearchIndexCapacityError as exc:
+                self._delete_crop_paths([face.get("crop_path") for face in faces])
+                raise self._quota_error(exc) from exc
+            except SearchIndexError as exc:
+                self._delete_crop_paths([face.get("crop_path") for face in faces])
+                raise self._search_unavailable(exc) from exc
+            except Exception:
+                self._delete_crop_paths([face.get("crop_path") for face in faces])
+                raise
+        return [self.public_face(face) for face in faces], rejected
+
+    def search(
+        self,
+        collection_id: str,
+        image: ImageData,
+        *,
+        limit: int,
+        threshold: float | None,
+    ) -> tuple[dict[str, Any], list[dict[str, Any]], float]:
+        collection = self.collection(collection_id)
+        profile = self.collection_detection_profile(collection)
+        face = self.selected_face(image, detection_profile=profile)
+        assert face.embedding is not None
+        selected_threshold = (
+            float(collection["default_threshold"]) if threshold is None else threshold
+        )
+        try:
+            matches = self.search_indexes.search(
+                collection_id,
+                face.embedding,
+                limit=limit,
+                threshold=selected_threshold,
+            )
+        except KeyError as exc:
+            raise not_found("collection", collection_id) from exc
+        except SearchIndexError as exc:
+            raise self._search_unavailable(exc) from exc
+        height, width = image.pixels.shape[:2]
+        return face_result(face, width, height), matches[:limit], selected_threshold
+
+    def search_all_faces(
+        self,
+        collection_id: str,
+        image: ImageData,
+        *,
+        max_faces: int,
+        threshold: float | None,
+    ) -> tuple[list[dict[str, Any]], float]:
+        """Detect once and search each retained face against one Collection.
+
+        This is the shared server-side path used by RTSP recognition. It avoids
+        re-detecting cropped faces and never exposes embeddings in stream state.
+        """
+
+        collection = self.collection(collection_id)
+        profile = self.collection_detection_profile(collection)
+        faces = self.analyze(image, embeddings=True, detection_profile=profile)
+        faces = faces[: min(max_faces, self.settings.max_detected_faces)]
+        selected_threshold = (
+            float(collection["default_threshold"]) if threshold is None else threshold
+        )
+        height, width = image.pixels.shape[:2]
+        results: list[dict[str, Any]] = []
+        try:
+            for face in faces:
+                assert face.embedding is not None
+                matches = self.search_indexes.search(
+                    collection_id,
+                    face.embedding,
+                    limit=1,
+                    threshold=selected_threshold,
+                )
+                results.append(
+                    {
+                        "face": face_result(face, width, height),
+                        "match": matches[0] if matches else None,
+                    }
+                )
+        except KeyError as exc:
+            raise not_found("collection", collection_id) from exc
+        except SearchIndexError as exc:
+            raise self._search_unavailable(exc) from exc
+        return results, selected_threshold
+
+    def delete_collection(self, collection_id: str, *, force: bool) -> None:
+        # Deletion is an administrative cleanup operation. It intentionally
+        # permits a Collection bound to an older model while enrollment,
+        # inference, and search continue to enforce the active contract.
+        self.raw_collection(collection_id)
+        status, crop_paths = self.search_indexes.run_drop(
+            collection_id,
+            lambda: self._delete_collection_mutation(collection_id, force=force),
+        )
+        if status == "not_empty":
+            raise conflict(
+                "The collection is not empty; repeat with force=true.",
+                code="collection_not_empty",
+            )
+        if status == "in_use":
+            raise conflict(
+                "The Collection is used by one or more Monitors; update or delete "
+                "those Monitors first.",
+                code="collection_in_use",
+            )
+        if status == "missing":
+            raise not_found("collection", collection_id)
+        for path in crop_paths:
+            self.crops.delete(path)
+        self.crops.delete_collection(collection_id)
+
+    def _delete_collection_mutation(
+        self, collection_id: str, *, force: bool
+    ) -> tuple[tuple[str, list[str]], Any]:
+        status, crops, mutation = self.repository.delete_collection_with_mutation(
+            collection_id, force=force
+        )
+        return (status, crops), mutation
+
+    def delete_person(self, collection_id: str, person_id: str) -> list[str]:
+        self.collection(collection_id)
+
+        def operation():
+            deleted, crops, mutation = self.repository.delete_person_with_mutation(
+                collection_id, person_id
+            )
+            return (deleted, crops), mutation
+
+        try:
+            deleted, crops = self.search_indexes.run_mutation(collection_id, operation)
+        except SearchMutationCommittedError as exc:
+            committed = exc.committed_result
+            if isinstance(committed, tuple) and len(committed) == 2:
+                committed_deleted, committed_crops = committed
+                if committed_deleted and isinstance(committed_crops, list):
+                    self._delete_crop_paths(committed_crops)
+            raise self._search_unavailable(exc) from exc
+        except KeyError as exc:
+            raise not_found("collection", collection_id) from exc
+        except SearchIndexError as exc:
+            raise self._search_unavailable(exc) from exc
+        if not deleted:
+            raise not_found("person", person_id)
+        return crops
+
+    def delete_face(self, collection_id: str, person_id: str, face_id: str) -> str | None:
+        self.collection(collection_id)
+
+        def operation():
+            deleted, crop, mutation = self.repository.delete_face_with_mutation(
+                collection_id, person_id, face_id
+            )
+            return (deleted, crop), mutation
+
+        try:
+            deleted, crop = self.search_indexes.run_mutation(collection_id, operation)
+        except SearchMutationCommittedError as exc:
+            committed = exc.committed_result
+            if isinstance(committed, tuple) and len(committed) == 2:
+                committed_deleted, committed_crop = committed
+                if committed_deleted:
+                    self._delete_crop_paths([committed_crop])
+            raise self._search_unavailable(exc) from exc
+        except KeyError as exc:
+            raise not_found("collection", collection_id) from exc
+        except SearchIndexError as exc:
+            raise self._search_unavailable(exc) from exc
+        if not deleted:
+            raise not_found("face", face_id)
+        return crop
+
+    def update_collection(self, collection_id: str, changes: dict[str, Any]) -> dict[str, Any]:
+        current = self.collection(collection_id)
+        detection_columns = {
+            "detector_input_sizes": "input_sizes",
+            "detector_threshold": "threshold",
+            "detector_nms_threshold": "nms_threshold",
+            "single_face_selection": "single_face_selection",
+        }
+
+        def current_value(key: str) -> Any:
+            detection_key = detection_columns.get(key)
+            if detection_key is not None:
+                value = current["detection"][detection_key]
+                if key == "detector_input_sizes":
+                    return [list(size) for size in value]
+                return value
+            return current.get(key)
+
+        # Browsers and SDKs may submit a complete edit form. Only actual changes
+        # should update SQLite or allocate a second index generation.
+        effective_changes = {
+            key: value for key, value in changes.items() if current_value(key) != value
+        }
+        if not effective_changes:
+            return current
+        if any(key in effective_changes for key in detection_columns):
+            target_detection = dict(current["detection"])
+            for column, detection_key in detection_columns.items():
+                if column in effective_changes:
+                    target_detection[detection_key] = effective_changes[column]
+            profile = DetectionProfile.from_mapping(target_detection)
+            try:
+                self.engine.validate_detection_profile(profile)
+            except ValueError as exc:
+                raise bad_request(
+                    str(exc), code="invalid_detection_profile"
+                ) from exc
+        if "capacity_rows" in effective_changes and int(effective_changes["capacity_rows"]) < int(
+            current["face_count"]
+        ):
+            raise self._quota_error(
+                CollectionCapacityExceeded(
+                    capacity_rows=int(effective_changes["capacity_rows"]),
+                    requested_rows=int(current["face_count"]),
+                )
+            )
+        try:
+            item = self.search_indexes.run_configuration_update(
+                collection_id,
+                lambda: self.repository.update_collection(collection_id, effective_changes),
+                changes=effective_changes,
+            )
+        except (CollectionCapacityExceeded, MaxFacesPerPersonExceeded) as exc:
+            raise self._quota_error(exc) from exc
+        except SearchIndexError as exc:
+            raise self._search_unavailable(exc) from exc
+        except KeyError as exc:
+            raise not_found("collection", collection_id) from exc
+        if item is None:
+            raise not_found("collection", collection_id)
+        return item
diff --git a/server/backend/insightface_server/services/images.py b/server/backend/insightface_server/services/images.py
new file mode 100644
index 000000000..d8717ee0b
--- /dev/null
+++ b/server/backend/insightface_server/services/images.py
@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+import hashlib
+from dataclasses import dataclass
+from functools import partial
+from io import BytesIO
+
+import cv2
+import numpy as np
+from anyio import to_thread
+from fastapi import UploadFile
+from PIL import Image, ImageOps, UnidentifiedImageError
+
+from ..config import Settings
+from ..errors import ApiError
+
+
+@dataclass(frozen=True, slots=True)
+class ImageData:
+    content: bytes
+    sha256: str
+    pixels: np.ndarray
+    filename: str
+
+
+class ImageLoader:
+    def __init__(self, settings: Settings):
+        self.settings = settings
+
+    async def from_upload(self, upload: UploadFile) -> ImageData:
+        chunks = bytearray()
+        while chunk := await upload.read(1024 * 1024):
+            if len(chunks) + len(chunk) > self.settings.max_image_bytes:
+                raise ApiError(
+                    "image_too_large",
+                    "The image exceeds the configured byte limit.",
+                    413,
+                    {"filename": upload.filename},
+                )
+            chunks.extend(chunk)
+        return await to_thread.run_sync(
+            partial(self.from_bytes, bytes(chunks), filename=upload.filename or "upload"),
+            abandon_on_cancel=True,
+        )
+
+    def from_bytes(self, content: bytes, *, filename: str = "image") -> ImageData:
+        if not content:
+            raise ApiError("invalid_image", "The image is empty.", 422)
+        if len(content) > self.settings.max_image_bytes:
+            raise ApiError("image_too_large", "The image exceeds the byte limit.", 413)
+        try:
+            with Image.open(BytesIO(content)) as source:
+                width, height = source.size
+                image_format = source.format
+                if width <= 0 or height <= 0 or width * height > self.settings.max_image_pixels:
+                    raise ApiError("image_too_large", "The image pixel limit was exceeded.", 413)
+                oriented = ImageOps.exif_transpose(source).convert("RGB")
+                pixels = cv2.cvtColor(np.asarray(oriented), cv2.COLOR_RGB2BGR)
+        except (UnidentifiedImageError, OSError, ValueError, Image.DecompressionBombError) as exc:
+            raise ApiError("invalid_image", "The image could not be decoded.", 422) from exc
+        if image_format not in {"JPEG", "PNG", "WEBP"}:
+            raise ApiError(
+                "invalid_image",
+                "Only JPEG, PNG, and WebP images are supported.",
+                422,
+            )
+        if pixels is None or pixels.ndim != 3 or pixels.shape[2] != 3:
+            raise ApiError("invalid_image", "The image could not be decoded.", 422)
+        if int(pixels.shape[0]) * int(pixels.shape[1]) > self.settings.max_image_pixels:
+            raise ApiError("image_too_large", "The image pixel limit was exceeded.", 413)
+        return ImageData(
+            content=content,
+            sha256=hashlib.sha256(content).hexdigest(),
+            pixels=pixels,
+            filename=filename,
+        )
diff --git a/server/backend/insightface_server/services/presentation.py b/server/backend/insightface_server/services/presentation.py
new file mode 100644
index 000000000..73d182244
--- /dev/null
+++ b/server/backend/insightface_server/services/presentation.py
@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+
+from ..inference import FaceObservation
+
+
+def clamp01(value: float) -> float:
+    return round(float(np.clip(value, 0.0, 1.0)), 6)
+
+
+def quality(face: FaceObservation) -> dict[str, float]:
+    sharpness = clamp01(float(getattr(face, "sharpness", 0.0)))
+    brightness = clamp01(float(getattr(face, "brightness", 0.0)))
+    yaw = float(getattr(face, "yaw", 0.0))
+    pitch = float(getattr(face, "pitch", 0.0))
+    roll = float(getattr(face, "roll", 0.0))
+    pose = clamp01(1.0 - min(1.0, (abs(yaw) + abs(pitch) + abs(roll)) / 135.0))
+    provided = getattr(face, "quality_score", None)
+    score = clamp01(
+        float(provided)
+        if provided is not None
+        else 0.4 * sharpness + 0.35 * brightness + 0.25 * pose
+    )
+    return {
+        "score": score,
+        "sharpness": sharpness,
+        "brightness": brightness,
+        "pose": pose,
+    }
+
+
+def bounding_box(
+    face: FaceObservation, image_width: int, image_height: int
+) -> dict[str, Any]:
+    left, top, right, bottom = face.bbox
+    x = max(0, min(image_width, int(round(left))))
+    y = max(0, min(image_height, int(round(top))))
+    right_px = max(x, min(image_width, int(round(right))))
+    bottom_px = max(y, min(image_height, int(round(bottom))))
+    width = right_px - x
+    height = bottom_px - y
+    return {
+        "pixels": {"x": x, "y": y, "width": width, "height": height},
+        "normalized": {
+            "left": round(x / image_width, 8),
+            "top": round(y / image_height, 8),
+            "width": round(width / image_width, 8),
+            "height": round(height / image_height, 8),
+        },
+    }
+
+
+def face_result(
+    face: FaceObservation, image_width: int, image_height: int
+) -> dict[str, Any]:
+    landmarks = []
+    if face.landmarks is not None:
+        landmarks = [
+            [round(float(point[0]), 3), round(float(point[1]), 3)]
+            for point in face.landmarks
+        ]
+    return {
+        "bbox": bounding_box(face, image_width, image_height),
+        "landmarks": landmarks,
+        "detection_score": clamp01(face.confidence),
+        "quality": quality(face),
+    }
diff --git a/server/backend/insightface_server/services/rtsp.py b/server/backend/insightface_server/services/rtsp.py
new file mode 100644
index 000000000..2a0cbc15a
--- /dev/null
+++ b/server/backend/insightface_server/services/rtsp.py
@@ -0,0 +1,1299 @@
+from __future__ import annotations
+
+import copy
+import logging
+import math
+import sqlite3
+import threading
+import time
+import uuid
+from collections import deque
+from collections.abc import Callable, Iterator
+from dataclasses import dataclass, replace
+from datetime import UTC, datetime
+from typing import Any
+from urllib.parse import urlsplit
+
+import cv2
+import numpy as np
+
+from ..errors import ApiError
+from ..request_context import without_request_deadline
+from ..storage import CursorCodec, Repository, SecretCodec
+from .core import FaceService
+from .images import ImageData
+
+LOGGER = logging.getLogger("insightface_server.monitors")
+
+
+class MonitorNotFoundError(KeyError):
+    pass
+
+
+class MonitorAlreadyExistsError(ValueError):
+    pass
+
+
+class MonitorLimitError(RuntimeError):
+    pass
+
+
+class MonitorPreviewDisabledError(RuntimeError):
+    pass
+
+
+class MonitorUnavailableError(RuntimeError):
+    pass
+
+
+@dataclass(frozen=True, slots=True)
+class MonitorOptions:
+    id: str
+    name: str
+    description: str
+    enabled: bool
+    url: str
+    collection_id: str
+    inference_fps: float
+    match_threshold: float | None
+    event_buffer_size: int
+    confirm_frames: int
+    absence_timeout_seconds: float
+    cooldown_seconds: float
+    emit_unknown: bool
+    preview_enabled: bool
+    created_at: str
+    updated_at: str
+
+
+@dataclass(slots=True)
+class _Track:
+    id: str
+    kind: str
+    person: dict[str, Any] | None
+    matched_face_id: str | None
+    similarity: float | None
+    face: dict[str, Any]
+    first_seen: float
+    last_seen: float
+    consecutive: int = 1
+    confirmed: bool = False
+
+
+def next_inference_time(*, started: float, finished: float, inference_fps: float) -> float:
+    """Return the next eligible start time without ever scheduling catch-up work."""
+
+    interval = 1.0 / inference_fps
+    return max(started + interval, finished)
+
+
+def redacted_rtsp_source(url: str) -> str:
+    """Return a display-safe source that never includes credentials or query values."""
+
+    parsed = urlsplit(url)
+    host = parsed.hostname or "unknown"
+    if ":" in host and not host.startswith("["):
+        host = f"[{host}]"
+    port = f":{parsed.port}" if parsed.port is not None else ""
+    path = parsed.path or "/"
+    return f"{parsed.scheme.lower()}://{host}{port}{path}"
+
+
+def _utc_now() -> str:
+    return datetime.now(UTC).isoformat(timespec="milliseconds").replace("+00:00", "Z")
+
+
+def _open_capture(url: str, open_timeout_ms: int, read_timeout_ms: int):
+    parameters: list[int] = []
+    if hasattr(cv2, "CAP_PROP_OPEN_TIMEOUT_MSEC"):
+        parameters.extend([cv2.CAP_PROP_OPEN_TIMEOUT_MSEC, open_timeout_ms])
+    if hasattr(cv2, "CAP_PROP_READ_TIMEOUT_MSEC"):
+        parameters.extend([cv2.CAP_PROP_READ_TIMEOUT_MSEC, read_timeout_ms])
+    try:
+        return cv2.VideoCapture(url, cv2.CAP_FFMPEG, parameters)
+    except (TypeError, cv2.error):
+        capture = cv2.VideoCapture()
+        if hasattr(cv2, "CAP_PROP_OPEN_TIMEOUT_MSEC"):
+            capture.set(cv2.CAP_PROP_OPEN_TIMEOUT_MSEC, open_timeout_ms)
+        if hasattr(cv2, "CAP_PROP_READ_TIMEOUT_MSEC"):
+            capture.set(cv2.CAP_PROP_READ_TIMEOUT_MSEC, read_timeout_ms)
+        capture.open(url, cv2.CAP_FFMPEG)
+        return capture
+
+
+def _monitor_result(item: dict[str, Any]) -> dict[str, Any]:
+    match = item.get("match")
+    face = copy.deepcopy(item.get("face") or {})
+    if not isinstance(match, dict):
+        return {
+            "status": "unknown",
+            "face": face,
+            "person": None,
+            "similarity": None,
+            "matched_face_id": None,
+        }
+    source_person = match.get("person") or {}
+    person = {
+        "id": str(source_person.get("id") or match.get("person_id") or ""),
+        "name": str(source_person.get("name") or ""),
+        "external_id": str(source_person.get("external_id") or ""),
+    }
+    return {
+        "status": "matched",
+        "face": face,
+        "person": person,
+        "similarity": round(float(match.get("similarity", 0.0)), 6),
+        "matched_face_id": match.get("matched_face_id"),
+    }
+
+
+def _normalized_box(result: dict[str, Any]) -> dict[str, float] | None:
+    value = result.get("face", {}).get("bbox", {}).get("normalized", {})
+    try:
+        return {
+            "left": float(value["left"]),
+            "top": float(value["top"]),
+            "width": float(value["width"]),
+            "height": float(value["height"]),
+        }
+    except (KeyError, TypeError, ValueError):
+        return None
+
+
+def _intersection_over_union(
+    left: dict[str, Any] | _Track,
+    right: dict[str, Any] | _Track,
+) -> float:
+    left_result = (
+        {"face": left.face} if isinstance(left, _Track) else left
+    )
+    right_result = (
+        {"face": right.face} if isinstance(right, _Track) else right
+    )
+    a = _normalized_box(left_result)
+    b = _normalized_box(right_result)
+    if a is None or b is None:
+        return 0.0
+    ax2 = a["left"] + a["width"]
+    ay2 = a["top"] + a["height"]
+    bx2 = b["left"] + b["width"]
+    by2 = b["top"] + b["height"]
+    intersection_width = max(0.0, min(ax2, bx2) - max(a["left"], b["left"]))
+    intersection_height = max(0.0, min(ay2, by2) - max(a["top"], b["top"]))
+    intersection = intersection_width * intersection_height
+    union = a["width"] * a["height"] + b["width"] * b["height"] - intersection
+    return intersection / union if union > 0.0 else 0.0
+
+
+class MonitorSession:
+    """One long-running RTSP reader, inference loop, event buffer, and preview."""
+
+    def __init__(
+        self,
+        service: FaceService,
+        options: MonitorOptions,
+        *,
+        max_faces: int,
+        preview_fps: float,
+        jpeg_quality: int,
+        open_timeout_seconds: float,
+        read_timeout_seconds: float,
+        reconnect_delay_seconds: float,
+        capture_factory: Callable[[str, int, int], Any] = _open_capture,
+    ) -> None:
+        self.service = service
+        self.options = options
+        self.max_faces = max_faces
+        self.preview_fps = preview_fps
+        self.jpeg_quality = jpeg_quality
+        self.open_timeout_ms = int(open_timeout_seconds * 1000)
+        self.read_timeout_ms = int(read_timeout_seconds * 1000)
+        self.reconnect_delay_seconds = reconnect_delay_seconds
+        self.capture_factory = capture_factory
+        self.stream_epoch = str(uuid.uuid4())
+        self.started_at = _utc_now()
+
+        self._condition = threading.Condition(threading.RLock())
+        self._stop_event = threading.Event()
+        self._reader: threading.Thread | None = None
+        self._worker: threading.Thread | None = None
+        self._state = "starting"
+        self._connected = False
+        self._connection_error: str | None = None
+        self._inference_error: str | None = None
+        self._latest_frame: np.ndarray | None = None
+        self._latest_frame_sequence = 0
+        self._processed_frame_sequence = 0
+        self._latest_jpeg: bytes | None = None
+        self._preview_sequence = 0
+        self._results: list[dict[str, Any]] = []
+        self._threshold = options.match_threshold
+        self._decoded_frames = 0
+        self._processed_frames = 0
+        self._skipped_frames = 0
+        self._preview_frames = 0
+        self._reconnects = 0
+        self._inference_errors = 0
+        self._last_processing_ms: float | None = None
+        self._last_frame_at: str | None = None
+        self._last_inference_at: str | None = None
+        self._source_fps: float | None = None
+        self._width: int | None = None
+        self._height: int | None = None
+        self._viewer_count = 0
+        self._inference_times: deque[float] = deque(maxlen=120)
+        self._events: deque[dict[str, Any]] = deque(maxlen=options.event_buffer_size)
+        self._event_sequence = 0
+        self._tracks: dict[str, _Track] = {}
+        self._last_emitted: dict[tuple[str, str], float] = {}
+        self._active_errors: set[str] = set()
+
+    def start(self) -> None:
+        with self._condition:
+            if self._reader is not None:
+                return
+            self._reader = threading.Thread(
+                target=self._reader_loop,
+                name=f"monitor-reader-{self.options.id}",
+                daemon=True,
+            )
+            self._worker = threading.Thread(
+                target=self._inference_loop,
+                name=f"monitor-inference-{self.options.id}",
+                daemon=True,
+            )
+            self._reader.start()
+            self._worker.start()
+
+    def stop(self) -> None:
+        self._stop_event.set()
+        with self._condition:
+            self._condition.notify_all()
+        current = threading.current_thread()
+        for thread in (self._reader, self._worker):
+            if thread is not None and thread is not current:
+                thread.join(timeout=max(2.0, self.read_timeout_ms / 1000.0 + 1.0))
+        with self._condition:
+            self._connected = False
+            self._state = "stopped"
+            self._latest_jpeg = None
+            self._condition.notify_all()
+
+    @property
+    def stopped(self) -> bool:
+        return self._stop_event.is_set()
+
+    @property
+    def preview_enabled(self) -> bool:
+        with self._condition:
+            return self.options.preview_enabled
+
+    @property
+    def connected(self) -> bool:
+        with self._condition:
+            return self._connected and self._latest_frame is not None
+
+    def apply_live_options(
+        self,
+        *,
+        preview_enabled: bool | None = None,
+        event_buffer_size: int | None = None,
+        name: str | None = None,
+        description: str | None = None,
+        updated_at: str | None = None,
+    ) -> None:
+        with self._condition:
+            changes: dict[str, Any] = {}
+            if preview_enabled is not None:
+                changes["preview_enabled"] = preview_enabled
+            if name is not None:
+                changes["name"] = name
+            if description is not None:
+                changes["description"] = description
+            if updated_at is not None:
+                changes["updated_at"] = updated_at
+            if event_buffer_size is not None:
+                changes["event_buffer_size"] = event_buffer_size
+                self._events = deque(
+                    list(self._events)[-event_buffer_size:],
+                    maxlen=event_buffer_size,
+                )
+            if changes:
+                self.options = replace(self.options, **changes)
+            if not self.options.preview_enabled:
+                self._latest_jpeg = None
+            self._condition.notify_all()
+
+    def summary(self) -> dict[str, Any]:
+        state = self.state()
+        return {
+            "status": state["status"],
+            "connected": state["connected"],
+            "stream_epoch": state["stream_epoch"],
+            "last_frame_at": state["inference"]["last_frame_at"],
+            "last_error": state["last_error"],
+            "preview_active": state["preview"]["active"],
+            "preview_viewers": state["preview"]["viewers"],
+        }
+
+    def state(self) -> dict[str, Any]:
+        with self._condition:
+            times = tuple(self._inference_times)
+            effective_fps = 0.0
+            if len(times) > 1 and times[-1] > times[0]:
+                effective_fps = (len(times) - 1) / (times[-1] - times[0])
+            status = self._state
+            if self._connected and self._inference_error:
+                status = "degraded"
+            faces = copy.deepcopy(self._results)
+            matched = sum(item.get("status") == "matched" for item in faces)
+            return {
+                "monitor_id": self.options.id,
+                "stream_epoch": self.stream_epoch,
+                "status": status,
+                "connected": self._connected,
+                "source": {
+                    "width": self._width,
+                    "height": self._height,
+                    "source_fps": self._source_fps,
+                },
+                "inference": {
+                    "configured_fps": self.options.inference_fps,
+                    "actual_fps": round(effective_fps, 3),
+                    "processing_ms": (
+                        round(self._last_processing_ms, 3)
+                        if self._last_processing_ms is not None
+                        else None
+                    ),
+                    "last_frame_at": self._last_frame_at,
+                    "last_inference_at": self._last_inference_at,
+                    "frame_sequence": self._latest_frame_sequence,
+                    "decoded_frames": self._decoded_frames,
+                    "processed_frames": self._processed_frames,
+                    "dropped_frames": self._skipped_frames,
+                    "capacity_limited": bool(
+                        self._last_processing_ms is not None
+                        and self._last_processing_ms
+                        > 1000.0 / self.options.inference_fps
+                    ),
+                },
+                "threshold": self._threshold,
+                "faces": faces,
+                "matched_faces": matched,
+                "unknown_faces": len(faces) - matched,
+                "preview": {
+                    "enabled": self.options.preview_enabled,
+                    "active": bool(
+                        self.options.preview_enabled and self._viewer_count > 0
+                    ),
+                    "viewers": self._viewer_count,
+                    "frames": self._preview_frames,
+                },
+                "reconnects": self._reconnects,
+                "inference_errors": self._inference_errors,
+                "last_error": self._inference_error or self._connection_error,
+                "started_at": self.started_at,
+            }
+
+    def event_page(
+        self,
+        *,
+        cursor_epoch: str | None,
+        after_sequence: int | None,
+        limit: int,
+    ) -> dict[str, Any]:
+        with self._condition:
+            events = list(self._events)
+            oldest = int(events[0]["sequence"]) if events else self._event_sequence + 1
+            latest = self._event_sequence
+            stream_reset = (
+                cursor_epoch is not None and cursor_epoch != self.stream_epoch
+            )
+            truncated = False
+            if cursor_epoch is None or stream_reset:
+                selected = events[-limit:]
+            else:
+                after = int(after_sequence or 0)
+                truncated = bool(events and after < oldest - 1)
+                selected = [
+                    item
+                    for item in events
+                    if int(item["sequence"]) > after
+                ][:limit]
+            if selected:
+                next_sequence = int(selected[-1]["sequence"])
+            elif cursor_epoch == self.stream_epoch and after_sequence is not None:
+                next_sequence = min(int(after_sequence), latest)
+            else:
+                next_sequence = latest
+            has_more = any(
+                int(item["sequence"]) > next_sequence for item in events
+            )
+            return {
+                "monitor_id": self.options.id,
+                "stream_epoch": self.stream_epoch,
+                "events": copy.deepcopy(selected),
+                "oldest_sequence": oldest if events else None,
+                "latest_sequence": latest,
+                "next_sequence": next_sequence,
+                "has_more": has_more,
+                "truncated": truncated,
+                "stream_reset": stream_reset,
+            }
+
+    def iter_mjpeg(self) -> Iterator[bytes]:
+        last_sequence = -1
+        with self._condition:
+            if not self.options.preview_enabled:
+                raise MonitorPreviewDisabledError
+            if not self._connected or self._latest_frame is None:
+                raise MonitorUnavailableError
+            self._viewer_count += 1
+            self._condition.notify_all()
+        try:
+            while not self._stop_event.is_set():
+                with self._condition:
+                    if not self.options.preview_enabled:
+                        break
+                    if (
+                        self._latest_jpeg is None
+                        or self._preview_sequence == last_sequence
+                    ):
+                        self._condition.wait(timeout=5.0)
+                    if self._stop_event.is_set() or not self.options.preview_enabled:
+                        break
+                    jpeg = self._latest_jpeg
+                    sequence = self._preview_sequence
+                if jpeg is None or sequence == last_sequence:
+                    continue
+                last_sequence = sequence
+                yield (
+                    b"--frame\r\n"
+                    b"Content-Type: image/jpeg\r\n"
+                    + f"Content-Length: {len(jpeg)}\r\n\r\n".encode("ascii")
+                    + jpeg
+                    + b"\r\n"
+                )
+        finally:
+            with self._condition:
+                self._viewer_count = max(0, self._viewer_count - 1)
+                if self._viewer_count == 0:
+                    self._latest_jpeg = None
+
+    def _append_event(
+        self,
+        event_type: str,
+        *,
+        now: float,
+        track: _Track | None = None,
+        error: dict[str, Any] | None = None,
+    ) -> None:
+        identity = (
+            str(track.person.get("id"))
+            if track is not None and track.person
+            else (track.id if track is not None else event_type)
+        )
+        cooldown_key = (event_type, identity)
+        last = self._last_emitted.get(cooldown_key)
+        if (
+            last is not None
+            and self.options.cooldown_seconds > 0
+            and now - last < self.options.cooldown_seconds
+        ):
+            return
+        self._last_emitted[cooldown_key] = now
+        self._event_sequence += 1
+        event: dict[str, Any] = {
+            "id": str(uuid.uuid4()),
+            "sequence": self._event_sequence,
+            "type": event_type,
+            "monitor_id": self.options.id,
+            "stream_epoch": self.stream_epoch,
+            "occurred_at": _utc_now(),
+        }
+        if track is not None:
+            event.update(
+                {
+                    "track_id": track.id,
+                    "person": copy.deepcopy(track.person),
+                    "similarity": track.similarity,
+                    "threshold": self._threshold,
+                    "matched_face_id": track.matched_face_id,
+                    "face": copy.deepcopy(track.face),
+                }
+            )
+        if error is not None:
+            event["error"] = copy.deepcopy(error)
+        self._events.append(event)
+
+    def _mark_error(
+        self,
+        *,
+        source: str,
+        code: str,
+        message: str,
+        now: float,
+    ) -> None:
+        with self._condition:
+            if source not in self._active_errors:
+                self._append_event(
+                    "monitor_error",
+                    now=now,
+                    error={"code": code, "message": message},
+                )
+            self._active_errors.add(source)
+
+    def _mark_recovered(self, *, source: str, now: float) -> None:
+        with self._condition:
+            was_active = source in self._active_errors
+            self._active_errors.discard(source)
+            if was_active and not self._active_errors:
+                self._append_event("monitor_recovered", now=now)
+
+    def _reader_loop(self) -> None:
+        retry_delay = self.reconnect_delay_seconds
+        next_preview = time.monotonic()
+        while not self._stop_event.is_set():
+            with self._condition:
+                self._state = (
+                    "starting" if self._decoded_frames == 0 else "reconnecting"
+                )
+                self._connected = False
+            capture = None
+            try:
+                capture = self.capture_factory(
+                    self.options.url,
+                    self.open_timeout_ms,
+                    self.read_timeout_ms,
+                )
+                if capture is None or not capture.isOpened():
+                    raise RuntimeError("capture_open_failed")
+                source_fps = float(capture.get(cv2.CAP_PROP_FPS) or 0.0)
+                with self._condition:
+                    self._source_fps = (
+                        round(source_fps, 3)
+                        if math.isfinite(source_fps) and source_fps > 0
+                        else None
+                    )
+                retry_delay = self.reconnect_delay_seconds
+                while not self._stop_event.is_set():
+                    ok, frame = capture.read()
+                    if not ok or frame is None or frame.ndim != 3:
+                        raise RuntimeError("capture_read_failed")
+                    now = time.monotonic()
+                    with self._condition:
+                        self._decoded_frames += 1
+                        self._latest_frame_sequence += 1
+                        self._latest_frame = frame
+                        self._width = int(frame.shape[1])
+                        self._height = int(frame.shape[0])
+                        self._last_frame_at = _utc_now()
+                        self._connected = True
+                        self._state = "running"
+                        self._connection_error = None
+                        viewers = self._viewer_count
+                        preview_enabled = self.options.preview_enabled
+                        self._condition.notify_all()
+                    self._mark_recovered(source="source", now=now)
+                    if preview_enabled and viewers > 0 and now >= next_preview:
+                        self._publish_preview(frame)
+                        next_preview = max(
+                            next_preview + 1.0 / self.preview_fps,
+                            now,
+                        )
+            except Exception as exc:
+                if not self._stop_event.is_set():
+                    LOGGER.warning(
+                        "monitor_source_unavailable monitor_id=%s error_type=%s",
+                        self.options.id,
+                        type(exc).__name__,
+                    )
+                    message = "The RTSP source is unavailable; reconnecting."
+                    with self._condition:
+                        self._connected = False
+                        self._connection_error = message
+                        self._state = "reconnecting"
+                        self._reconnects += 1
+                    self._mark_error(
+                        source="source",
+                        code="rtsp_source_unavailable",
+                        message=message,
+                        now=time.monotonic(),
+                    )
+            finally:
+                if capture is not None:
+                    try:
+                        capture.release()
+                    except Exception:
+                        pass
+            if not self._stop_event.wait(retry_delay):
+                retry_delay = min(
+                    30.0,
+                    max(self.reconnect_delay_seconds, retry_delay * 2.0),
+                )
+
+    def _inference_loop(self) -> None:
+        next_allowed = time.monotonic()
+        while not self._stop_event.is_set():
+            frame: np.ndarray | None = None
+            sequence = 0
+            with self._condition:
+                now = time.monotonic()
+                fresh_frame = (
+                    self._latest_frame is not None
+                    and self._latest_frame_sequence != self._processed_frame_sequence
+                )
+                if fresh_frame and now >= next_allowed:
+                    assert self._latest_frame is not None
+                    frame = self._latest_frame.copy()
+                    sequence = self._latest_frame_sequence
+                else:
+                    wait_time = 0.5
+                    if fresh_frame:
+                        wait_time = max(0.001, min(wait_time, next_allowed - now))
+                    self._condition.wait(timeout=wait_time)
+                    self._expire_tracks(time.monotonic())
+                    continue
+            started = time.monotonic()
+            results: list[dict[str, Any]] | None = None
+            selected_threshold: float | None = None
+            error_message: str | None = None
+            try:
+                assert frame is not None
+                image = ImageData(
+                    content=b"",
+                    sha256="",
+                    pixels=frame,
+                    filename="rtsp-frame",
+                )
+                raw_results, selected_threshold = self.service.search_all_faces(
+                    self.options.collection_id,
+                    image,
+                    max_faces=self.max_faces,
+                    threshold=self.options.match_threshold,
+                )
+                results = [_monitor_result(item) for item in raw_results]
+            except ApiError as exc:
+                error_message = f"{exc.code}: {exc.message}"
+            except Exception as exc:
+                LOGGER.exception(
+                    "monitor_inference_failed monitor_id=%s error_type=%s",
+                    self.options.id,
+                    type(exc).__name__,
+                )
+                error_message = "internal_error: The frame could not be processed."
+            finished = time.monotonic()
+            with self._condition:
+                previous_sequence = self._processed_frame_sequence
+                self._processed_frame_sequence = sequence
+                self._skipped_frames += max(0, sequence - previous_sequence - 1)
+                self._processed_frames += 1
+                self._last_processing_ms = (finished - started) * 1000.0
+                self._last_inference_at = _utc_now()
+                self._inference_times.append(started)
+                if results is not None:
+                    self._threshold = selected_threshold
+                    self._results = self._update_tracks(results, finished)
+                    self._inference_error = None
+                else:
+                    self._inference_errors += 1
+                    self._inference_error = error_message
+            if results is None:
+                self._mark_error(
+                    source="inference",
+                    code="monitor_inference_failed",
+                    message=error_message or "The frame could not be processed.",
+                    now=finished,
+                )
+            elif self._connected:
+                self._mark_recovered(source="inference", now=finished)
+            next_allowed = next_inference_time(
+                started=started,
+                finished=finished,
+                inference_fps=self.options.inference_fps,
+            )
+
+    def _update_tracks(
+        self,
+        results: list[dict[str, Any]],
+        now: float,
+    ) -> list[dict[str, Any]]:
+        unmatched_tracks = set(self._tracks)
+        rendered: list[dict[str, Any]] = []
+        for result in results:
+            best_id: str | None = None
+            best_iou = 0.15
+            incoming_person = (result.get("person") or {}).get("id")
+            for track_id in unmatched_tracks:
+                track = self._tracks[track_id]
+                track_person = (track.person or {}).get("id")
+                if incoming_person and track_person and incoming_person != track_person:
+                    continue
+                overlap = _intersection_over_union(track, result)
+                if overlap > best_iou:
+                    best_iou = overlap
+                    best_id = track_id
+            if best_id is None:
+                track = _Track(
+                    id=str(uuid.uuid4()),
+                    kind=str(result["status"]),
+                    person=copy.deepcopy(result.get("person")),
+                    matched_face_id=result.get("matched_face_id"),
+                    similarity=result.get("similarity"),
+                    face=copy.deepcopy(result.get("face") or {}),
+                    first_seen=now,
+                    last_seen=now,
+                )
+                self._tracks[track.id] = track
+            else:
+                unmatched_tracks.remove(best_id)
+                track = self._tracks[best_id]
+                incoming_kind = str(result["status"])
+                if (
+                    track.kind != incoming_kind
+                    or (track.person or {}).get("id") != incoming_person
+                ):
+                    if track.confirmed:
+                        self._emit_track_event(track, entering=False, now=now)
+                    track.kind = incoming_kind
+                    track.first_seen = now
+                    track.consecutive = 1
+                    track.confirmed = False
+                else:
+                    track.consecutive += 1
+                track.last_seen = now
+                track.person = copy.deepcopy(result.get("person"))
+                track.matched_face_id = result.get("matched_face_id")
+                track.similarity = result.get("similarity")
+                track.face = copy.deepcopy(result.get("face") or {})
+            if not track.confirmed and track.consecutive >= self.options.confirm_frames:
+                track.confirmed = True
+                self._emit_track_event(track, entering=True, now=now)
+            rendered.append(
+                {
+                    **copy.deepcopy(result),
+                    "track_id": track.id,
+                }
+            )
+        self._expire_tracks(now)
+        return rendered
+
+    def _expire_tracks(self, now: float) -> None:
+        expired = [
+            track_id
+            for track_id, track in self._tracks.items()
+            if now - track.last_seen >= self.options.absence_timeout_seconds
+        ]
+        for track_id in expired:
+            track = self._tracks.pop(track_id)
+            if track.confirmed:
+                self._emit_track_event(track, entering=False, now=now)
+
+    def _emit_track_event(self, track: _Track, *, entering: bool, now: float) -> None:
+        if track.kind == "unknown" and not self.options.emit_unknown:
+            return
+        prefix = "person" if track.kind == "matched" else "unknown"
+        suffix = "enter" if entering else "exit"
+        self._append_event(f"{prefix}_{suffix}", now=now, track=track)
+
+    def _publish_preview(self, frame: np.ndarray) -> None:
+        ok, encoded = cv2.imencode(
+            ".jpg",
+            frame,
+            [cv2.IMWRITE_JPEG_QUALITY, self.jpeg_quality],
+        )
+        if not ok:
+            return
+        with self._condition:
+            if not self.options.preview_enabled or self._viewer_count == 0:
+                return
+            self._latest_jpeg = encoded.tobytes()
+            self._preview_frames += 1
+            self._preview_sequence += 1
+            self._condition.notify_all()
+
+
+class MonitorManager:
+    def __init__(
+        self,
+        service: FaceService,
+        repository: Repository,
+        cursors: CursorCodec,
+        credentials: SecretCodec,
+        *,
+        max_monitors: int = 4,
+        max_faces: int = 100,
+        preview_fps: float = 5.0,
+        jpeg_quality: int = 82,
+        open_timeout_seconds: float = 5.0,
+        read_timeout_seconds: float = 5.0,
+        reconnect_delay_seconds: float = 1.0,
+        capture_factory: Callable[[str, int, int], Any] = _open_capture,
+        session_factory: Callable[..., MonitorSession] = MonitorSession,
+    ) -> None:
+        self.service = service
+        self.repository = repository
+        self.cursors = cursors
+        self.credentials = credentials
+        self.max_monitors = max_monitors
+        self.max_faces = max_faces
+        self.preview_fps = preview_fps
+        self.jpeg_quality = jpeg_quality
+        self.open_timeout_seconds = open_timeout_seconds
+        self.read_timeout_seconds = read_timeout_seconds
+        self.reconnect_delay_seconds = reconnect_delay_seconds
+        self.capture_factory = capture_factory
+        self.session_factory = session_factory
+        # Lifecycle mutations are deliberately serialized.  Capacity checks,
+        # durable configuration changes, and session publication must be one
+        # operation; protecting only individual _sessions reads allows two
+        # concurrent callers to both reserve the final slot.
+        self._mutation_lock = threading.RLock()
+        self._lock = threading.RLock()
+        self._sessions: dict[str, MonitorSession] = {}
+        self._startup_errors: dict[str, str] = {}
+        self._closed = False
+
+    def startup(self) -> None:
+        with self._mutation_lock:
+            self._require_open()
+            for row in self.repository.list_enabled_monitors():
+                monitor_id = str(row["id"])
+                with self._lock:
+                    current = self._sessions.get(monitor_id)
+                if current is not None and not current.stopped:
+                    continue
+                try:
+                    self._start_row(row)
+                except Exception as exc:
+                    LOGGER.error(
+                        "monitor_restore_failed monitor_id=%s error_type=%s",
+                        monitor_id,
+                        type(exc).__name__,
+                    )
+                    with self._lock:
+                        self._startup_errors[monitor_id] = (
+                            "The Monitor could not be restored at startup."
+                        )
+
+    def close(self) -> None:
+        with self._mutation_lock:
+            self._closed = True
+            with self._lock:
+                sessions = list(self._sessions.items())
+                self._sessions.clear()
+            for monitor_id, session in sessions:
+                self._stop_session(session, monitor_id=monitor_id)
+
+    def create(self, item: dict[str, Any]) -> dict[str, Any]:
+        with self._mutation_lock:
+            self._require_open()
+            collection_id = str(item["collection_id"])
+            self.service.collection(collection_id)
+            monitor_id = str(item["id"])
+            if self.repository.get_monitor(monitor_id) is not None:
+                raise MonitorAlreadyExistsError(monitor_id)
+            if item.get("enabled", True):
+                self._ensure_capacity(excluding=None)
+            values = dict(item)
+            values["source_url_ciphertext"] = self.credentials.encrypt(
+                str(values.pop("url")),
+                scope=f"monitor:{monitor_id}",
+            )
+            try:
+                row = self.repository.create_monitor(values)
+            except sqlite3.IntegrityError as exc:
+                if self.repository.get_monitor(monitor_id) is not None:
+                    raise MonitorAlreadyExistsError(monitor_id) from exc
+                # A Collection delete can win after the validation read but
+                # before this insert. Re-run the public lookup so that race is
+                # reported as the existing collection_not_found API error.
+                self.service.collection(collection_id)
+                raise
+            if row["enabled"]:
+                try:
+                    candidate = self._start_candidate(row)
+                except Exception:
+                    try:
+                        # The row was already committed. Its cleanup must not be
+                        # rejected solely because session startup consumed the
+                        # originating request's remaining deadline.
+                        with without_request_deadline():
+                            self.repository.delete_monitor(monitor_id)
+                    except Exception:
+                        LOGGER.exception(
+                            "monitor_create_compensation_failed monitor_id=%s",
+                            monitor_id,
+                        )
+                    raise
+                previous = self._publish_session(monitor_id, candidate)
+                if previous is not None:
+                    self._stop_session(previous, monitor_id=monitor_id)
+            return self._public_monitor(row)
+
+    def list(self, after: str, limit: int) -> list[dict[str, Any]]:
+        return [
+            self._public_monitor(row)
+            for row in self.repository.list_monitors(after, limit)
+        ]
+
+    def get(self, monitor_id: str) -> dict[str, Any]:
+        row = self.repository.get_monitor(monitor_id)
+        if row is None:
+            raise MonitorNotFoundError(monitor_id)
+        return self._public_monitor(row)
+
+    def update(self, monitor_id: str, changes: dict[str, Any]) -> dict[str, Any]:
+        with self._mutation_lock:
+            self._require_open()
+            existing = self.repository.get_monitor(monitor_id)
+            if existing is None:
+                raise MonitorNotFoundError(monitor_id)
+            collection_changed = "collection_id" in changes
+            if collection_changed:
+                self.service.collection(str(changes["collection_id"]))
+            desired_enabled = bool(changes.get("enabled", existing["enabled"]))
+            persisted = dict(changes)
+            if "url" in persisted:
+                persisted["source_url_ciphertext"] = self.credentials.encrypt(
+                    str(persisted.pop("url")),
+                    scope=f"monitor:{monitor_id}",
+                )
+
+            restart_fields = {
+                "url",
+                "collection_id",
+                "inference_fps",
+                "match_threshold",
+                "confirm_frames",
+                "absence_timeout_seconds",
+                "cooldown_seconds",
+                "emit_unknown",
+            }
+            with self._lock:
+                session = self._sessions.get(monitor_id)
+            should_restart = bool(set(changes) & restart_fields)
+            needs_candidate = desired_enabled and (
+                session is None or session.stopped or should_restart
+            )
+            candidate: MonitorSession | None = None
+            if needs_candidate:
+                self._ensure_capacity(excluding=monitor_id)
+                proposed = dict(existing)
+                proposed.update(persisted)
+                proposed["enabled"] = True
+                candidate = self._start_candidate(proposed)
+
+            try:
+                updated = self.repository.update_monitor(monitor_id, persisted)
+            except Exception:
+                if candidate is not None:
+                    self._stop_session(candidate, monitor_id=monitor_id)
+                if collection_changed:
+                    # As in create(), turn a foreign-key race into the stable
+                    # collection_not_found API error when appropriate.
+                    self.service.collection(str(changes["collection_id"]))
+                raise
+            if updated is None:
+                if candidate is not None:
+                    self._stop_session(candidate, monitor_id=monitor_id)
+                raise MonitorNotFoundError(monitor_id)
+
+            if not updated["enabled"]:
+                with self._lock:
+                    previous = self._sessions.pop(monitor_id, None)
+                if previous is not None:
+                    self._stop_session(previous, monitor_id=monitor_id)
+            elif candidate is not None:
+                previous = self._publish_session(monitor_id, candidate)
+                if previous is not None and previous is not candidate:
+                    self._stop_session(previous, monitor_id=monitor_id)
+            elif session is not None:
+                session.apply_live_options(
+                    preview_enabled=(
+                        bool(updated["preview_enabled"])
+                        if "preview_enabled" in changes
+                        else None
+                    ),
+                    event_buffer_size=(
+                        int(updated["event_buffer_size"])
+                        if "event_buffer_size" in changes
+                        else None
+                    ),
+                    name=str(updated["name"]) if "name" in changes else None,
+                    description=(
+                        str(updated["description"])
+                        if "description" in changes
+                        else None
+                    ),
+                    updated_at=str(updated["updated_at"]),
+                )
+            with self._lock:
+                self._startup_errors.pop(monitor_id, None)
+            return self._public_monitor(updated)
+
+    def delete(self, monitor_id: str) -> None:
+        with self._mutation_lock:
+            self._require_open()
+            if self.repository.get_monitor(monitor_id) is None:
+                raise MonitorNotFoundError(monitor_id)
+            # Keep the live session intact if the durable delete fails. Once
+            # the row is gone, detach before stopping so readers cannot obtain
+            # a session for a deleted Monitor.
+            if not self.repository.delete_monitor(monitor_id):
+                raise MonitorNotFoundError(monitor_id)
+            with self._lock:
+                session = self._sessions.pop(monitor_id, None)
+                self._startup_errors.pop(monitor_id, None)
+            if session is not None:
+                self._stop_session(session, monitor_id=monitor_id)
+
+    def state(self, monitor_id: str) -> dict[str, Any]:
+        row = self.repository.get_monitor(monitor_id)
+        if row is None:
+            raise MonitorNotFoundError(monitor_id)
+        with self._lock:
+            session = self._sessions.get(monitor_id)
+            error = self._startup_errors.get(monitor_id)
+        if session is not None:
+            return session.state()
+        return {
+            "monitor_id": monitor_id,
+            "stream_epoch": None,
+            "status": "error" if error else "stopped",
+            "connected": False,
+            "source": {"width": None, "height": None, "source_fps": None},
+            "inference": {
+                "configured_fps": float(row["inference_fps"]),
+                "actual_fps": 0.0,
+                "processing_ms": None,
+                "last_frame_at": None,
+                "last_inference_at": None,
+                "frame_sequence": 0,
+                "decoded_frames": 0,
+                "processed_frames": 0,
+                "dropped_frames": 0,
+                "capacity_limited": False,
+            },
+            "threshold": row["match_threshold"],
+            "faces": [],
+            "matched_faces": 0,
+            "unknown_faces": 0,
+            "preview": {
+                "enabled": bool(row["preview_enabled"]),
+                "active": False,
+                "viewers": 0,
+                "frames": 0,
+            },
+            "reconnects": 0,
+            "inference_errors": 0,
+            "last_error": error,
+            "started_at": None,
+        }
+
+    def events(
+        self,
+        monitor_id: str,
+        *,
+        cursor: str | None,
+        limit: int,
+    ) -> dict[str, Any]:
+        if self.repository.get_monitor(monitor_id) is None:
+            raise MonitorNotFoundError(monitor_id)
+        scope = f"monitor-events:{monitor_id}"
+        cursor_epoch: str | None = None
+        after_sequence: int | None = None
+        if cursor:
+            raw = self.cursors.decode(cursor, scope)
+            try:
+                cursor_epoch, sequence_text = raw.rsplit(":", 1)
+                after_sequence = int(sequence_text)
+                if after_sequence < 0 or not cursor_epoch:
+                    raise ValueError
+            except ValueError as exc:
+                from ..errors import bad_request
+
+                raise bad_request(
+                    "The cursor is invalid.",
+                    code="invalid_cursor",
+                ) from exc
+        with self._lock:
+            session = self._sessions.get(monitor_id)
+        if session is None:
+            page: dict[str, Any] = {
+                "monitor_id": monitor_id,
+                "stream_epoch": None,
+                "events": [],
+                "oldest_sequence": None,
+                "latest_sequence": 0,
+                "next_sequence": 0,
+                "has_more": False,
+                "truncated": False,
+                "stream_reset": cursor_epoch is not None,
+            }
+            next_epoch = "stopped"
+        else:
+            page = session.event_page(
+                cursor_epoch=cursor_epoch,
+                after_sequence=after_sequence,
+                limit=limit,
+            )
+            next_epoch = session.stream_epoch
+        page["next_cursor"] = self.cursors.encode(
+            scope,
+            f"{next_epoch}:{page.pop('next_sequence')}",
+        )
+        return page
+
+    def preview(self, monitor_id: str) -> Iterator[bytes]:
+        row = self.repository.get_monitor(monitor_id)
+        if row is None:
+            raise MonitorNotFoundError(monitor_id)
+        if not bool(row["preview_enabled"]):
+            raise MonitorPreviewDisabledError
+        with self._lock:
+            session = self._sessions.get(monitor_id)
+        if session is None or session.stopped:
+            raise MonitorUnavailableError
+        if not session.connected:
+            raise MonitorUnavailableError
+        return session.iter_mjpeg()
+
+    def _ensure_capacity(self, *, excluding: str | None) -> None:
+        with self._lock:
+            active = sum(
+                not session.stopped and monitor_id != excluding
+                for monitor_id, session in self._sessions.items()
+            )
+        if active >= self.max_monitors:
+            raise MonitorLimitError
+
+    def _start_row(self, row: dict[str, Any]) -> MonitorSession:
+        with self._mutation_lock:
+            self._require_open()
+            monitor_id = str(row["id"])
+            self._ensure_capacity(excluding=monitor_id)
+            session = self._start_candidate(row)
+            previous = self._publish_session(monitor_id, session)
+            if previous is not None and previous is not session:
+                self._stop_session(previous, monitor_id=monitor_id)
+            return session
+
+    def _start_candidate(self, row: dict[str, Any]) -> MonitorSession:
+        options = self._options(row)
+        session = self.session_factory(
+            self.service,
+            options,
+            max_faces=self.max_faces,
+            preview_fps=self.preview_fps,
+            jpeg_quality=self.jpeg_quality,
+            open_timeout_seconds=self.open_timeout_seconds,
+            read_timeout_seconds=self.read_timeout_seconds,
+            reconnect_delay_seconds=self.reconnect_delay_seconds,
+            capture_factory=self.capture_factory,
+        )
+        try:
+            session.start()
+        except Exception:
+            self._stop_session(session, monitor_id=options.id)
+            raise
+        return session
+
+    def _publish_session(
+        self,
+        monitor_id: str,
+        session: MonitorSession,
+    ) -> MonitorSession | None:
+        with self._lock:
+            previous = self._sessions.get(monitor_id)
+            self._sessions[monitor_id] = session
+            self._startup_errors.pop(monitor_id, None)
+        return previous
+
+    @staticmethod
+    def _stop_session(session: MonitorSession, *, monitor_id: str) -> None:
+        try:
+            session.stop()
+        except Exception:
+            LOGGER.exception(
+                "monitor_stop_failed monitor_id=%s",
+                monitor_id,
+            )
+
+    def _require_open(self) -> None:
+        if self._closed:
+            raise RuntimeError("Monitor manager is closed.")
+
+    def _options(self, row: dict[str, Any]) -> MonitorOptions:
+        monitor_id = str(row["id"])
+        url = self.credentials.decrypt(
+            str(row["source_url_ciphertext"]),
+            scope=f"monitor:{monitor_id}",
+        )
+        return MonitorOptions(
+            id=monitor_id,
+            name=str(row["name"]),
+            description=str(row["description"]),
+            enabled=bool(row["enabled"]),
+            url=url,
+            collection_id=str(row["collection_id"]),
+            inference_fps=float(row["inference_fps"]),
+            match_threshold=(
+                float(row["match_threshold"])
+                if row["match_threshold"] is not None
+                else None
+            ),
+            event_buffer_size=int(row["event_buffer_size"]),
+            confirm_frames=int(row["confirm_frames"]),
+            absence_timeout_seconds=float(row["absence_timeout_seconds"]),
+            cooldown_seconds=float(row["cooldown_seconds"]),
+            emit_unknown=bool(row["emit_unknown"]),
+            preview_enabled=bool(row["preview_enabled"]),
+            created_at=str(row["created_at"]),
+            updated_at=str(row["updated_at"]),
+        )
+
+    def _public_monitor(self, row: dict[str, Any]) -> dict[str, Any]:
+        options = self._options(row)
+        with self._lock:
+            session = self._sessions.get(options.id)
+            startup_error = self._startup_errors.get(options.id)
+        runtime = (
+            session.summary()
+            if session is not None
+            else {
+                "status": (
+                    "error" if startup_error else "stopped"
+                ),
+                "connected": False,
+                "stream_epoch": None,
+                "last_frame_at": None,
+                "last_error": startup_error,
+                "preview_active": False,
+                "preview_viewers": 0,
+            }
+        )
+        return {
+            "id": options.id,
+            "name": options.name,
+            "description": options.description,
+            "enabled": options.enabled,
+            "source": {
+                "type": "rtsp",
+                "url": redacted_rtsp_source(options.url),
+            },
+            "collection_id": options.collection_id,
+            "inference_fps": options.inference_fps,
+            "match_threshold": options.match_threshold,
+            "event_buffer_size": options.event_buffer_size,
+            "event_policy": {
+                "confirm_frames": options.confirm_frames,
+                "absence_timeout_seconds": options.absence_timeout_seconds,
+                "cooldown_seconds": options.cooldown_seconds,
+                "emit_unknown": options.emit_unknown,
+            },
+            "preview_enabled": options.preview_enabled,
+            "runtime": runtime,
+            "created_at": options.created_at,
+            "updated_at": options.updated_at,
+        }
diff --git a/server/backend/insightface_server/storage/__init__.py b/server/backend/insightface_server/storage/__init__.py
new file mode 100644
index 000000000..23d56d7b6
--- /dev/null
+++ b/server/backend/insightface_server/storage/__init__.py
@@ -0,0 +1,35 @@
+from .crops import FaceCropStore
+from .database import Database
+from .pagination import CursorCodec
+from .repository import (
+    SEARCH_LOAD_POLICIES,
+    SEARCH_PROFILES,
+    CollectionCapacityExceeded,
+    IndexFace,
+    IndexFaceBatch,
+    MaxFacesPerPersonExceeded,
+    Repository,
+    SearchChange,
+    SearchMutation,
+    SearchState,
+    utc_now,
+)
+from .secrets import SecretCodec
+
+__all__ = [
+    "SEARCH_LOAD_POLICIES",
+    "SEARCH_PROFILES",
+    "CollectionCapacityExceeded",
+    "CursorCodec",
+    "Database",
+    "FaceCropStore",
+    "IndexFace",
+    "IndexFaceBatch",
+    "MaxFacesPerPersonExceeded",
+    "Repository",
+    "SearchChange",
+    "SearchMutation",
+    "SearchState",
+    "SecretCodec",
+    "utc_now",
+]
diff --git a/server/backend/insightface_server/storage/crops.py b/server/backend/insightface_server/storage/crops.py
new file mode 100644
index 000000000..9f151e3fb
--- /dev/null
+++ b/server/backend/insightface_server/storage/crops.py
@@ -0,0 +1,101 @@
+from __future__ import annotations
+
+import hashlib
+import os
+import shutil
+import uuid
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+
+class FaceCropStore:
+    MEDIA_TYPE = "image/jpeg"
+
+    def __init__(self, data_dir: Path, *, enabled: bool):
+        self.data_dir = data_dir
+        self.root = data_dir / "face-crops"
+        self.enabled = enabled
+        if enabled:
+            self.root.mkdir(parents=True, exist_ok=True, mode=0o700)
+            self.root.chmod(0o700)
+
+    @staticmethod
+    def _directory(collection_id: str) -> str:
+        return hashlib.sha256(collection_id.encode()).hexdigest()
+
+    @staticmethod
+    def encode(
+        image: np.ndarray,
+        bbox: tuple[float, float, float, float],
+    ) -> bytes:
+        """Return a 112x112 JPEG crop without writing it to disk."""
+
+        height, width = image.shape[:2]
+        left = max(0, min(width, int(bbox[0])))
+        top = max(0, min(height, int(bbox[1])))
+        right = max(0, min(width, int(np.ceil(bbox[2]))))
+        bottom = max(0, min(height, int(np.ceil(bbox[3]))))
+        crop = image[top:bottom, left:right]
+        if crop.size == 0:
+            raise RuntimeError("Cannot persist an empty face crop")
+        crop = cv2.resize(crop, (112, 112), interpolation=cv2.INTER_AREA)
+        ok, encoded = cv2.imencode(".jpg", crop, [cv2.IMWRITE_JPEG_QUALITY, 92])
+        if not ok:
+            raise RuntimeError("Cannot encode face crop")
+        return encoded.tobytes()
+
+    def _managed_path(self, relative_path: str) -> Path:
+        relative = Path(relative_path)
+        root = self.root.resolve()
+        target = (self.data_dir / relative).resolve()
+        if relative.is_absolute() or root not in target.parents:
+            raise RuntimeError("Face crop path is outside managed storage")
+        return target
+
+    def save(
+        self,
+        collection_id: str,
+        face_id: str,
+        image: np.ndarray,
+        bbox: tuple[float, float, float, float],
+    ) -> str | None:
+        if not self.enabled:
+            return None
+        encoded = self.encode(image, bbox)
+        directory = self.root / self._directory(collection_id)
+        directory.mkdir(parents=True, exist_ok=True, mode=0o700)
+        target = directory / f"{face_id}.jpg"
+        temporary = directory / f".{face_id}.{uuid.uuid4().hex}.tmp"
+        try:
+            with temporary.open("xb") as stream:
+                stream.write(encoded)
+            temporary.chmod(0o600)
+            os.replace(temporary, target)
+            target.chmod(0o600)
+        finally:
+            temporary.unlink(missing_ok=True)
+        return str(target.relative_to(self.data_dir))
+
+    def read(self, relative_path: str | None) -> bytes | None:
+        """Read a legacy on-disk crop while rejecting paths outside the store."""
+
+        if not relative_path:
+            return None
+        target = self._managed_path(relative_path)
+        try:
+            return target.read_bytes()
+        except FileNotFoundError:
+            return None
+
+    def delete(self, relative_path: str | None) -> None:
+        if not relative_path:
+            return
+        target = self._managed_path(relative_path)
+        target.unlink(missing_ok=True)
+
+    def delete_collection(self, collection_id: str) -> None:
+        directory = self.root / self._directory(collection_id)
+        if directory.is_dir():
+            shutil.rmtree(directory)
diff --git a/server/backend/insightface_server/storage/database.py b/server/backend/insightface_server/storage/database.py
new file mode 100644
index 000000000..7aa71211c
--- /dev/null
+++ b/server/backend/insightface_server/storage/database.py
@@ -0,0 +1,140 @@
+from __future__ import annotations
+
+import contextlib
+import fcntl
+import re
+import sqlite3
+import threading
+from collections.abc import Iterator
+from pathlib import Path
+from typing import BinaryIO
+
+from ..request_context import check_request_deadline
+
+
+class Database:
+    _MIGRATION_NAME = re.compile(r"\d{4}_[A-Za-z0-9_.-]+\.sql")
+
+    def __init__(self, path: Path, migrations_dir: Path):
+        self.path = path
+        self.migrations_dir = migrations_dir
+        self._write_lock = threading.RLock()
+        self._process_lock: BinaryIO | None = None
+
+    def acquire_process_lock(self) -> None:
+        """Reject a second Server process using the same durable data store."""
+
+        with self._write_lock:
+            if self._process_lock is not None:
+                return
+            self.path.parent.mkdir(parents=True, exist_ok=True)
+            lock_path = self.path.with_suffix(self.path.suffix + ".server.lock")
+            lock = lock_path.open("a+b")
+            lock_path.chmod(0o600)
+            try:
+                fcntl.flock(lock.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+            except BlockingIOError as exc:
+                lock.close()
+                raise RuntimeError(
+                    "the data directory is already in use by another InsightFace "
+                    "Server process; phase one requires exactly one worker"
+                ) from exc
+            self._process_lock = lock
+
+    def release_process_lock(self) -> None:
+        with self._write_lock:
+            lock = self._process_lock
+            self._process_lock = None
+            if lock is not None:
+                fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
+                lock.close()
+
+    def initialize(self) -> None:
+        with self._write_lock:
+            self.path.parent.mkdir(parents=True, exist_ok=True)
+            lock_path = self.path.with_suffix(self.path.suffix + ".migration.lock")
+            with lock_path.open("a+b") as lock:
+                lock_path.chmod(0o600)
+                fcntl.flock(lock.fileno(), fcntl.LOCK_EX)
+                with self._connect() as connection:
+                    connection.execute(
+                        "CREATE TABLE IF NOT EXISTS schema_migrations "
+                        "(version TEXT PRIMARY KEY, "
+                        "applied_at TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP)"
+                    )
+                    connection.commit()
+                    applied = {
+                        row[0]
+                        for row in connection.execute("SELECT version FROM schema_migrations")
+                    }
+                    candidates = (
+                        path
+                        for path in self.migrations_dir.glob("*.sql")
+                        if self._MIGRATION_NAME.fullmatch(path.name)
+                    )
+                    for migration in sorted(candidates):
+                        if migration.name in applied:
+                            continue
+                        # Migration names are constrained by _MIGRATION_NAME, so
+                        # the quoted version cannot contain SQL metacharacters.
+                        script = migration.read_text(encoding="utf-8")
+                        transaction = (
+                            "BEGIN IMMEDIATE;\n"
+                            + script
+                            + "\nINSERT INTO schema_migrations(version) VALUES "
+                            + f"('{migration.name}');\nCOMMIT;"
+                        )
+                        try:
+                            connection.executescript(transaction)
+                        except Exception:
+                            connection.rollback()
+                            raise
+                fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
+            self.path.chmod(0o600)
+
+    def _connect(self) -> sqlite3.Connection:
+        connection = sqlite3.connect(self.path, timeout=15, check_same_thread=False)
+        self.path.chmod(0o600)
+        connection.row_factory = sqlite3.Row
+        connection.execute("PRAGMA foreign_keys=ON")
+        connection.execute("PRAGMA journal_mode=WAL")
+        connection.execute("PRAGMA synchronous=NORMAL")
+        connection.execute("PRAGMA busy_timeout=15000")
+        return connection
+
+    @contextlib.contextmanager
+    def read(self) -> Iterator[sqlite3.Connection]:
+        connection = self._connect()
+        try:
+            yield connection
+        finally:
+            connection.close()
+
+    @contextlib.contextmanager
+    def write(self) -> Iterator[sqlite3.Connection]:
+        with self._write_lock:
+            check_request_deadline()
+            connection = self._connect()
+            try:
+                connection.execute("BEGIN IMMEDIATE")
+                yield connection
+                check_request_deadline()
+                connection.commit()
+            except Exception:
+                connection.rollback()
+                raise
+            finally:
+                connection.close()
+
+    def status(self) -> dict[str, object]:
+        with self.read() as connection:
+            quick = connection.execute("PRAGMA quick_check").fetchone()[0]
+            count = connection.execute(
+                "SELECT count(*) FROM schema_migrations"
+            ).fetchone()[0]
+        return {
+            "path": str(self.path),
+            "quick_check": quick,
+            "migration_count": int(count),
+            "wal": True,
+        }
diff --git a/server/backend/insightface_server/storage/pagination.py b/server/backend/insightface_server/storage/pagination.py
new file mode 100644
index 000000000..126065de8
--- /dev/null
+++ b/server/backend/insightface_server/storage/pagination.py
@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+import base64
+import hashlib
+import hmac
+import json
+import os
+from pathlib import Path
+
+from ..errors import bad_request
+
+
+def _encode(raw: bytes) -> str:
+    return base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii")
+
+
+def _decode(value: str) -> bytes:
+    decoded = base64.urlsafe_b64decode(value + "=" * (-len(value) % 4))
+    if _encode(decoded) != value:
+        raise ValueError("non-canonical base64")
+    return decoded
+
+
+class CursorCodec:
+    def __init__(self, secret_path: Path):
+        secret_path.parent.mkdir(parents=True, exist_ok=True)
+        created = os.urandom(32)
+        try:
+            descriptor = os.open(secret_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
+            with os.fdopen(descriptor, "wb") as stream:
+                stream.write(created)
+            self._secret = created
+        except FileExistsError:
+            flags = os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0)
+            descriptor = os.open(secret_path, flags)
+            with os.fdopen(descriptor, "rb") as stream:
+                self._secret = stream.read()
+        secret_path.chmod(0o600)
+        if len(self._secret) < 32:
+            raise RuntimeError("Cursor secret must contain at least 32 bytes")
+
+    def encode(self, scope: str, after: str) -> str:
+        payload = json.dumps(
+            {"v": 1, "scope": scope, "after": after},
+            sort_keys=True,
+            separators=(",", ":"),
+        ).encode()
+        signature = hmac.new(self._secret, payload, hashlib.sha256).digest()[:16]
+        return f"{_encode(payload)}.{_encode(signature)}"
+
+    def decode(self, token: str | None, scope: str) -> str:
+        if not token:
+            return ""
+        try:
+            if len(token) > 2048:
+                raise ValueError("token too long")
+            encoded_payload, encoded_signature = token.split(".", 1)
+            payload = _decode(encoded_payload)
+            signature = _decode(encoded_signature)
+            expected = hmac.new(self._secret, payload, hashlib.sha256).digest()[:16]
+            value = json.loads(payload)
+            if not hmac.compare_digest(signature, expected):
+                raise ValueError("signature")
+            if value != {"v": 1, "scope": scope, "after": value.get("after")}:
+                raise ValueError("scope")
+            after = value["after"]
+            if not isinstance(after, str):
+                raise ValueError("after")
+            return after
+        except (ValueError, TypeError, KeyError, json.JSONDecodeError) as exc:
+            raise bad_request("The cursor is invalid.", code="invalid_cursor") from exc
diff --git a/server/backend/insightface_server/storage/repository.py b/server/backend/insightface_server/storage/repository.py
new file mode 100644
index 000000000..3ca6ad388
--- /dev/null
+++ b/server/backend/insightface_server/storage/repository.py
@@ -0,0 +1,1309 @@
+from __future__ import annotations
+
+import hashlib
+import hmac
+import json
+import os
+import sqlite3
+import uuid
+from collections.abc import Iterator, Sequence
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from typing import Any
+
+import numpy as np
+
+from ..config import DetectionProfile
+from ..models.embedding_contract import embedding_contract_id_for_collection
+from .database import Database
+
+SEARCH_PROFILES = frozenset(
+    {"fp32_v1", "fp16_v1", "bf16_v1", "int8_x1000_v1", "int8_x736_v1"}
+)
+SEARCH_LOAD_POLICIES = frozenset({"lazy", "eager"})
+
+
+class CollectionCapacityExceeded(ValueError):
+    def __init__(self, *, capacity_rows: int, requested_rows: int):
+        self.capacity_rows = capacity_rows
+        self.requested_rows = requested_rows
+        super().__init__(
+            f"collection capacity {capacity_rows} would be exceeded by {requested_rows} rows"
+        )
+
+
+class MaxFacesPerPersonExceeded(ValueError):
+    def __init__(self, *, max_faces: int, requested_faces: int):
+        self.max_faces = max_faces
+        self.requested_faces = requested_faces
+        super().__init__(
+            f"person face limit {max_faces} would be exceeded by {requested_faces} faces"
+        )
+
+
+@dataclass(frozen=True, slots=True)
+class SearchChange:
+    seq: int
+    collection_id: str
+    revision: int
+    operation: str
+    vector_id: int | None
+    person_numeric_id: int | None
+    face_id: str | None
+    created_at: str
+
+
+@dataclass(frozen=True, slots=True)
+class SearchMutation:
+    collection_id: str
+    revision: int
+    changes: tuple[SearchChange, ...]
+
+    @property
+    def max_seq(self) -> int:
+        return max((change.seq for change in self.changes), default=0)
+
+
+@dataclass(frozen=True, slots=True)
+class IndexFace:
+    vector_id: int
+    person_numeric_id: int
+    face_id: str
+    person_id: str
+    embedding: np.ndarray
+
+
+@dataclass(frozen=True, slots=True)
+class IndexFaceBatch:
+    """One batch from a single consistent SQLite read snapshot."""
+
+    revision: int
+    faces: tuple[IndexFace, ...]
+    max_seq: int
+
+
+@dataclass(frozen=True, slots=True)
+class SearchState:
+    collection_id: str
+    revision: int
+    max_seq: int
+
+
+def utc_now() -> str:
+    return datetime.now(UTC).isoformat(timespec="milliseconds").replace("+00:00", "Z")
+
+
+class Repository:
+    def __init__(self, database: Database):
+        self.database = database
+
+    @staticmethod
+    def _collection_row(row: sqlite3.Row) -> dict[str, Any]:
+        value = dict(row)
+        value["metadata"] = json.loads(value.pop("metadata_json"))
+        defaults = DetectionProfile()
+        value["detection"] = {
+            "input_sizes": json.loads(
+                value.pop(
+                    "detector_input_sizes_json",
+                    json.dumps(defaults.as_dict()["input_sizes"]),
+                )
+            ),
+            "threshold": float(value.pop("detector_threshold", defaults.threshold)),
+            "nms_threshold": float(
+                value.pop("detector_nms_threshold", defaults.nms_threshold)
+            ),
+            "single_face_selection": value.pop(
+                "single_face_selection", defaults.single_face_selection
+            ),
+        }
+        value.setdefault("detection_revision", 1)
+        value["person_count"] = int(value.get("person_count", 0))
+        value["face_count"] = int(value.get("face_count", 0))
+        value["save_face_crops"] = bool(value.get("save_face_crops", False))
+        value["embedding_contract_id"] = embedding_contract_id_for_collection(value)
+        return value
+
+    @staticmethod
+    def _person_row(row: sqlite3.Row) -> dict[str, Any]:
+        value = dict(row)
+        value["metadata"] = json.loads(value.pop("metadata_json"))
+        value["face_count"] = int(value.get("face_count", 0))
+        value.pop("collection_id", None)
+        return value
+
+    @staticmethod
+    def _monitor_row(row: sqlite3.Row) -> dict[str, Any]:
+        value = dict(row)
+        value["enabled"] = bool(value["enabled"])
+        value["emit_unknown"] = bool(value.pop("emit_unknown"))
+        value["preview_enabled"] = bool(value["preview_enabled"])
+        return value
+
+    @staticmethod
+    def _face_row(row: sqlite3.Row, *, include_embedding: bool = True) -> dict[str, Any]:
+        value = dict(row)
+        value["bounding_box"] = json.loads(value.pop("bounding_box_json"))
+        landmarks_json = value.pop("landmarks_json")
+        value["landmarks"] = json.loads(landmarks_json) if landmarks_json else None
+        value["quality"] = json.loads(value.pop("quality_json"))
+        if include_embedding:
+            value["embedding"] = np.frombuffer(
+                value["embedding"], dtype=np.float32
+            ).copy()
+        else:
+            value.pop("embedding", None)
+        # Crop bytes are intentionally only exposed by get_face_crop(). Keep this
+        # defensive pop for callers constructing custom SELECTs in tests.
+        crop_image = value.pop("crop_image", None)
+        value["has_crop"] = bool(
+            value.get("has_crop", crop_image is not None or value.get("crop_path"))
+        )
+        value.pop("crop_path", None)
+        value.pop("collection_id", None)
+        return value
+
+    @staticmethod
+    def _face_columns(*, include_embedding: bool) -> str:
+        columns = """
+            id,collection_id,person_id,embedding_dimension,
+            bounding_box_json,landmarks_json,detection_score,quality_json,
+            model_id,model_version,model_digest,preprocessing_version,
+            embedding_source,embedding_contract_id,crop_path,created_at,
+            CASE WHEN crop_image IS NOT NULL OR crop_path IS NOT NULL
+                 THEN 1 ELSE 0 END AS has_crop
+        """
+        if include_embedding:
+            return "embedding," + columns
+        return columns
+
+    @staticmethod
+    def _collection_query() -> str:
+        return """
+            SELECT c.*,
+              (SELECT count(*) FROM persons p WHERE p.collection_id=c.id) AS person_count,
+              (SELECT count(*) FROM face_samples f WHERE f.collection_id=c.id) AS face_count
+            FROM collections c
+        """
+
+    def create_collection(self, item: dict[str, Any]) -> dict[str, Any]:
+        now = utc_now()
+        detection = DetectionProfile()
+        with self.database.write() as connection:
+            connection.execute(
+                """INSERT INTO collections(
+                    id,name,description,default_threshold,model_id,model_version,
+                    model_digest,embedding_dimension,preprocessing_version,
+                    metadata_json,created_at,updated_at,search_profile,capacity_rows,
+                    max_faces_per_person,load_policy,save_face_crops,
+                    detector_input_sizes_json,detector_threshold,
+                    detector_nms_threshold,single_face_selection,detection_revision
+                ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
+                (
+                    item["id"],
+                    item["name"],
+                    item.get("description", ""),
+                    item["default_threshold"],
+                    item["model_id"],
+                    item["model_version"],
+                    item["model_digest"],
+                    item["embedding_dimension"],
+                    item["preprocessing_version"],
+                    json.dumps(item.get("metadata", {}), sort_keys=True),
+                    now,
+                    now,
+                    item.get("search_profile", "fp32_v1"),
+                    int(item.get("capacity_rows", 100000)),
+                    int(item.get("max_faces_per_person", 20)),
+                    item.get("load_policy", "lazy"),
+                    int(bool(item.get("save_face_crops", False))),
+                    json.dumps(
+                        item.get(
+                            "detector_input_sizes", detection.as_dict()["input_sizes"]
+                        ),
+                        separators=(",", ":"),
+                    ),
+                    float(item.get("detector_threshold", detection.threshold)),
+                    float(
+                        item.get("detector_nms_threshold", detection.nms_threshold)
+                    ),
+                    item.get(
+                        "single_face_selection", detection.single_face_selection
+                    ),
+                    1,
+                ),
+            )
+            row = connection.execute(
+                self._collection_query() + " WHERE c.id=?", (item["id"],)
+            ).fetchone()
+            if row is None:  # pragma: no cover - guarded by the INSERT above
+                raise RuntimeError("new Collection could not be read in its transaction")
+            result = self._collection_row(row)
+        return result
+
+    def get_collection(self, collection_id: str) -> dict[str, Any] | None:
+        with self.database.read() as connection:
+            row = connection.execute(
+                self._collection_query() + " WHERE c.id=?", (collection_id,)
+            ).fetchone()
+        return self._collection_row(row) if row else None
+
+    def list_collections(self, after: str, limit: int) -> list[dict[str, Any]]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                self._collection_query()
+                + " WHERE c.id>? ORDER BY c.id LIMIT ?",
+                (after, limit),
+            ).fetchall()
+        return [self._collection_row(row) for row in rows]
+
+    def create_monitor(self, item: dict[str, Any]) -> dict[str, Any]:
+        now = utc_now()
+        with self.database.write() as connection:
+            connection.execute(
+                """INSERT INTO monitors(
+                    id,name,description,enabled,source_type,source_url_ciphertext,
+                    collection_id,inference_fps,match_threshold,event_buffer_size,
+                    confirm_frames,absence_timeout_seconds,cooldown_seconds,
+                    emit_unknown,preview_enabled,created_at,updated_at
+                ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
+                (
+                    item["id"],
+                    item["name"],
+                    item.get("description", ""),
+                    int(bool(item.get("enabled", True))),
+                    item.get("source_type", "rtsp"),
+                    item["source_url_ciphertext"],
+                    item["collection_id"],
+                    float(item.get("inference_fps", 2.0)),
+                    item.get("match_threshold"),
+                    int(item.get("event_buffer_size", 1000)),
+                    int(item.get("confirm_frames", 3)),
+                    float(item.get("absence_timeout_seconds", 3.0)),
+                    float(item.get("cooldown_seconds", 10.0)),
+                    int(bool(item.get("emit_unknown", True))),
+                    int(bool(item.get("preview_enabled", False))),
+                    now,
+                    now,
+                ),
+            )
+            row = connection.execute(
+                "SELECT * FROM monitors WHERE id=?", (item["id"],)
+            ).fetchone()
+        if row is None:  # pragma: no cover - guarded by INSERT
+            raise RuntimeError("new Monitor could not be read in its transaction")
+        return self._monitor_row(row)
+
+    def get_monitor(self, monitor_id: str) -> dict[str, Any] | None:
+        with self.database.read() as connection:
+            row = connection.execute(
+                "SELECT * FROM monitors WHERE id=?", (monitor_id,)
+            ).fetchone()
+        return self._monitor_row(row) if row else None
+
+    def list_monitors(self, after: str, limit: int) -> list[dict[str, Any]]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                "SELECT * FROM monitors WHERE id>? ORDER BY id LIMIT ?",
+                (after, limit),
+            ).fetchall()
+        return [self._monitor_row(row) for row in rows]
+
+    def list_enabled_monitors(self) -> list[dict[str, Any]]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                "SELECT * FROM monitors WHERE enabled=1 ORDER BY id"
+            ).fetchall()
+        return [self._monitor_row(row) for row in rows]
+
+    def update_monitor(
+        self, monitor_id: str, changes: dict[str, Any]
+    ) -> dict[str, Any] | None:
+        allowed = {
+            "name",
+            "description",
+            "enabled",
+            "source_type",
+            "source_url_ciphertext",
+            "collection_id",
+            "inference_fps",
+            "match_threshold",
+            "event_buffer_size",
+            "confirm_frames",
+            "absence_timeout_seconds",
+            "cooldown_seconds",
+            "emit_unknown",
+            "preview_enabled",
+        }
+        assignments: list[str] = []
+        values: list[Any] = []
+        for key, value in changes.items():
+            if key not in allowed:
+                continue
+            assignments.append(f"{key}=?")
+            if key in {"enabled", "emit_unknown", "preview_enabled"}:
+                value = int(bool(value))
+            values.append(value)
+        if not assignments:
+            return self.get_monitor(monitor_id)
+        assignments.append("updated_at=?")
+        values.extend([utc_now(), monitor_id])
+        with self.database.write() as connection:
+            cursor = connection.execute(
+                f"UPDATE monitors SET {', '.join(assignments)} WHERE id=?",
+                values,
+            )
+            if not cursor.rowcount:
+                return None
+            row = connection.execute(
+                "SELECT * FROM monitors WHERE id=?", (monitor_id,)
+            ).fetchone()
+        return self._monitor_row(row) if row else None
+
+    def delete_monitor(self, monitor_id: str) -> bool:
+        with self.database.write() as connection:
+            cursor = connection.execute(
+                "DELETE FROM monitors WHERE id=?", (monitor_id,)
+            )
+        return bool(cursor.rowcount)
+
+    def monitor_ids_for_collection(self, collection_id: str) -> list[str]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                "SELECT id FROM monitors WHERE collection_id=? ORDER BY id",
+                (collection_id,),
+            ).fetchall()
+        return [str(row[0]) for row in rows]
+
+    def update_collection(
+        self, collection_id: str, changes: dict[str, Any]
+    ) -> dict[str, Any] | None:
+        allowed = {
+            "name": "name",
+            "description": "description",
+            "default_threshold": "default_threshold",
+            "metadata": "metadata_json",
+            "capacity_rows": "capacity_rows",
+            "max_faces_per_person": "max_faces_per_person",
+            "load_policy": "load_policy",
+            "save_face_crops": "save_face_crops",
+            "detector_input_sizes": "detector_input_sizes_json",
+            "detector_threshold": "detector_threshold",
+            "detector_nms_threshold": "detector_nms_threshold",
+            "single_face_selection": "single_face_selection",
+        }
+        assignments: list[str] = []
+        values: list[Any] = []
+        for key, column in allowed.items():
+            if key not in changes:
+                continue
+            assignments.append(f"{column}=?")
+            value = changes[key]
+            if key == "metadata":
+                value = json.dumps(value, sort_keys=True)
+            elif key == "detector_input_sizes":
+                value = json.dumps(value, separators=(",", ":"))
+            elif key == "save_face_crops":
+                value = int(bool(value))
+            values.append(value)
+        if not assignments:
+            return self.get_collection(collection_id)
+        if any(key.startswith("detector_") or key == "single_face_selection" for key in changes):
+            assignments.append("detection_revision=detection_revision+1")
+        assignments.append("updated_at=?")
+        values.extend([utc_now(), collection_id])
+        result: dict[str, Any] | None = None
+        with self.database.write() as connection:
+            if "capacity_rows" in changes:
+                face_count = int(
+                    connection.execute(
+                        "SELECT count(*) FROM face_samples WHERE collection_id=?",
+                        (collection_id,),
+                    ).fetchone()[0]
+                )
+                if int(changes["capacity_rows"]) < face_count:
+                    raise CollectionCapacityExceeded(
+                        capacity_rows=int(changes["capacity_rows"]),
+                        requested_rows=face_count,
+                    )
+            if "max_faces_per_person" in changes:
+                largest = int(
+                    connection.execute(
+                        """SELECT coalesce(max(face_count), 0)
+                           FROM (
+                             SELECT count(*) AS face_count FROM face_samples
+                             WHERE collection_id=? GROUP BY person_id
+                           )""",
+                        (collection_id,),
+                    ).fetchone()[0]
+                )
+                if int(changes["max_faces_per_person"]) < largest:
+                    raise MaxFacesPerPersonExceeded(
+                        max_faces=int(changes["max_faces_per_person"]),
+                        requested_faces=largest,
+                    )
+            cursor = connection.execute(
+                f"UPDATE collections SET {', '.join(assignments)} WHERE id=?", values
+            )
+            if cursor.rowcount:
+                row = connection.execute(
+                    self._collection_query() + " WHERE c.id=?", (collection_id,)
+                ).fetchone()
+                if row is None:  # pragma: no cover - guarded by the UPDATE above
+                    raise RuntimeError("updated Collection disappeared in its transaction")
+                result = self._collection_row(row)
+        return result
+
+    def delete_collection(
+        self, collection_id: str, *, force: bool
+    ) -> tuple[str, list[str]]:
+        status, crops, _ = self.delete_collection_with_mutation(
+            collection_id, force=force
+        )
+        return status, crops
+
+    def delete_collection_with_mutation(
+        self, collection_id: str, *, force: bool
+    ) -> tuple[str, list[str], SearchMutation | None]:
+        mutation: SearchMutation | None = None
+        with self.database.write() as connection:
+            exists = connection.execute(
+                "SELECT 1 FROM collections WHERE id=?", (collection_id,)
+            ).fetchone()
+            if not exists:
+                return "missing", [], None
+            if connection.execute(
+                "SELECT 1 FROM monitors WHERE collection_id=? LIMIT 1",
+                (collection_id,),
+            ).fetchone():
+                return "in_use", [], None
+            if not force:
+                counts = connection.execute(
+                    "SELECT (SELECT count(*) FROM persons WHERE collection_id=?), "
+                    "(SELECT count(*) FROM face_samples WHERE collection_id=?)",
+                    (collection_id, collection_id),
+                ).fetchone()
+                if int(counts[0]) or int(counts[1]):
+                    return "not_empty", [], None
+            crops = [
+                row[0]
+                for row in connection.execute(
+                    "SELECT crop_path FROM face_samples WHERE collection_id=? "
+                    "AND crop_path IS NOT NULL",
+                    (collection_id,),
+                )
+            ]
+            revision = self._next_search_revision(connection, collection_id)
+            mutation = self._record_search_changes(
+                connection,
+                collection_id,
+                revision,
+                [("clear", None, None, None)],
+            )
+            cursor = connection.execute("DELETE FROM collections WHERE id=?", (collection_id,))
+        return ("deleted" if cursor.rowcount else "missing"), crops, mutation
+
+    def create_person_with_faces(
+        self, collection_id: str, person: dict[str, Any], faces: list[dict[str, Any]]
+    ) -> dict[str, Any]:
+        created, _ = self.create_person_with_faces_with_mutation(
+            collection_id, person, faces
+        )
+        return created
+
+    def create_person_with_faces_with_mutation(
+        self, collection_id: str, person: dict[str, Any], faces: list[dict[str, Any]]
+    ) -> tuple[dict[str, Any], SearchMutation | None]:
+        now = utc_now()
+        mutation: SearchMutation | None = None
+        created: dict[str, Any] | None = None
+        with self.database.write() as connection:
+            self._check_face_limits(
+                connection, collection_id, person_id=None, add_count=len(faces)
+            )
+            connection.execute(
+                """INSERT INTO persons(
+                    collection_id,id,name,external_id,metadata_json,created_at,updated_at
+                ) VALUES(?,?,?,?,?,?,?)""",
+                (
+                    collection_id,
+                    person["id"],
+                    person.get("name"),
+                    person.get("external_id"),
+                    json.dumps(person.get("metadata", {}), sort_keys=True),
+                    now,
+                    now,
+                ),
+            )
+            person_numeric_id = self._ensure_person_numeric_id(
+                connection, collection_id, person["id"]
+            )
+            added: list[tuple[int, int, str]] = []
+            for face in faces:
+                vector_id = self._insert_face(
+                    connection,
+                    collection_id,
+                    person["id"],
+                    person_numeric_id,
+                    face,
+                )
+                added.append((vector_id, person_numeric_id, face["id"]))
+            if added:
+                revision = self._next_search_revision(connection, collection_id)
+                mutation = self._record_search_changes(
+                    connection,
+                    collection_id,
+                    revision,
+                    [("add", *entry) for entry in added],
+                )
+            created = self._get_person_from_connection(
+                connection, collection_id, person["id"]
+            )
+        assert created is not None
+        return created, mutation
+
+    @staticmethod
+    def _insert_face(
+        connection: sqlite3.Connection,
+        collection_id: str,
+        person_id: str,
+        person_numeric_id: int,
+        face: dict[str, Any],
+    ) -> int:
+        connection.execute(
+            """INSERT INTO face_samples(
+                id,collection_id,person_id,embedding,embedding_dimension,
+                bounding_box_json,landmarks_json,detection_score,quality_json,
+                model_id,model_version,model_digest,preprocessing_version,
+                embedding_source,embedding_contract_id,
+                crop_path,crop_image,crop_media_type,created_at
+            ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
+            (
+                face["id"],
+                collection_id,
+                person_id,
+                np.asarray(face["embedding"], dtype=np.float32).tobytes(),
+                int(face["embedding_dimension"]),
+                json.dumps(face["bounding_box"], sort_keys=True),
+                json.dumps(face.get("landmarks")),
+                face["detection_score"],
+                json.dumps(face["quality"], sort_keys=True),
+                face["model_id"],
+                face["model_version"],
+                face["model_digest"],
+                face["preprocessing_version"],
+                face.get("embedding_source", "server"),
+                face.get("embedding_contract_id"),
+                face.get("crop_path"),
+                face.get("crop_image"),
+                face.get("crop_media_type"),
+                face.get("created_at", utc_now()),
+            ),
+        )
+        cursor = connection.execute(
+            """INSERT INTO search_vector_ids(
+                   collection_id,person_id,person_numeric_id,face_id
+               ) VALUES(?,?,?,?)""",
+            (collection_id, person_id, person_numeric_id, face["id"]),
+        )
+        if cursor.lastrowid is None:
+            raise RuntimeError("failed to allocate stable face search identifier")
+        return int(cursor.lastrowid)
+
+    def add_faces(
+        self, collection_id: str, person_id: str, faces: list[dict[str, Any]]
+    ) -> None:
+        self.add_faces_with_mutation(collection_id, person_id, faces)
+
+    def add_faces_with_mutation(
+        self, collection_id: str, person_id: str, faces: list[dict[str, Any]]
+    ) -> SearchMutation | None:
+        mutation: SearchMutation | None = None
+        with self.database.write() as connection:
+            exists = connection.execute(
+                "SELECT 1 FROM persons WHERE collection_id=? AND id=?",
+                (collection_id, person_id),
+            ).fetchone()
+            if not exists:
+                raise KeyError(person_id)
+            self._check_face_limits(
+                connection, collection_id, person_id=person_id, add_count=len(faces)
+            )
+            person_numeric_id = self._ensure_person_numeric_id(
+                connection, collection_id, person_id
+            )
+            added: list[tuple[int, int, str]] = []
+            for face in faces:
+                vector_id = self._insert_face(
+                    connection,
+                    collection_id,
+                    person_id,
+                    person_numeric_id,
+                    face,
+                )
+                added.append((vector_id, person_numeric_id, face["id"]))
+            connection.execute(
+                "UPDATE persons SET updated_at=? WHERE collection_id=? AND id=?",
+                (utc_now(), collection_id, person_id),
+            )
+            if added:
+                revision = self._next_search_revision(connection, collection_id)
+                mutation = self._record_search_changes(
+                    connection,
+                    collection_id,
+                    revision,
+                    [("add", *entry) for entry in added],
+                )
+        return mutation
+
+    def get_person(self, collection_id: str, person_id: str) -> dict[str, Any] | None:
+        with self.database.read() as connection:
+            return self._get_person_from_connection(connection, collection_id, person_id)
+
+    @classmethod
+    def _get_person_from_connection(
+        cls, connection: sqlite3.Connection, collection_id: str, person_id: str
+    ) -> dict[str, Any] | None:
+        row = connection.execute(
+            """SELECT p.*, count(f.id) AS face_count
+               FROM persons p LEFT JOIN face_samples f
+                 ON f.collection_id=p.collection_id AND f.person_id=p.id
+               WHERE p.collection_id=? AND p.id=?
+               GROUP BY p.collection_id,p.id""",
+            (collection_id, person_id),
+        ).fetchone()
+        return cls._person_row(row) if row else None
+
+    def list_persons(
+        self, collection_id: str, after: str, limit: int, query: str | None = None
+    ) -> list[dict[str, Any]]:
+        filters = ["p.collection_id=?", "p.id>?"]
+        values: list[Any] = [collection_id, after]
+        if query:
+            filters.append(
+                "(lower(p.id) LIKE ? OR lower(coalesce(p.name,'')) LIKE ? "
+                "OR lower(coalesce(p.external_id,'')) LIKE ?)"
+            )
+            needle = f"%{query.lower()}%"
+            values.extend([needle, needle, needle])
+        values.append(limit)
+        with self.database.read() as connection:
+            rows = connection.execute(
+                f"""SELECT p.*, count(f.id) AS face_count
+                    FROM persons p LEFT JOIN face_samples f
+                      ON f.collection_id=p.collection_id AND f.person_id=p.id
+                    WHERE {' AND '.join(filters)}
+                    GROUP BY p.collection_id,p.id ORDER BY p.id LIMIT ?""",
+                values,
+            ).fetchall()
+        return [self._person_row(row) for row in rows]
+
+    def update_person(
+        self, collection_id: str, person_id: str, changes: dict[str, Any]
+    ) -> dict[str, Any] | None:
+        allowed = {"name": "name", "external_id": "external_id", "metadata": "metadata_json"}
+        assignments: list[str] = []
+        values: list[Any] = []
+        for key, column in allowed.items():
+            if key not in changes:
+                continue
+            assignments.append(f"{column}=?")
+            values.append(
+                json.dumps(changes[key], sort_keys=True) if key == "metadata" else changes[key]
+            )
+        if not assignments:
+            return self.get_person(collection_id, person_id)
+        assignments.append("updated_at=?")
+        values.extend([utc_now(), collection_id, person_id])
+        with self.database.write() as connection:
+            cursor = connection.execute(
+                f"UPDATE persons SET {', '.join(assignments)} "
+                "WHERE collection_id=? AND id=?",
+                values,
+            )
+        return self.get_person(collection_id, person_id) if cursor.rowcount else None
+
+    def delete_person(self, collection_id: str, person_id: str) -> tuple[bool, list[str]]:
+        deleted, crops, _ = self.delete_person_with_mutation(collection_id, person_id)
+        return deleted, crops
+
+    def delete_person_with_mutation(
+        self, collection_id: str, person_id: str
+    ) -> tuple[bool, list[str], SearchMutation | None]:
+        mutation: SearchMutation | None = None
+        with self.database.write() as connection:
+            crops = [
+                row[0]
+                for row in connection.execute(
+                    "SELECT crop_path FROM face_samples WHERE collection_id=? "
+                    "AND person_id=? AND crop_path IS NOT NULL",
+                    (collection_id, person_id),
+                )
+            ]
+            exists = connection.execute(
+                "SELECT 1 FROM persons WHERE collection_id=? AND id=?",
+                (collection_id, person_id),
+            ).fetchone()
+            if not exists:
+                return False, [], None
+            mappings = self._face_mappings_for_person(
+                connection, collection_id, person_id
+            )
+            if mappings:
+                revision = self._next_search_revision(connection, collection_id)
+                mutation = self._record_search_changes(
+                    connection,
+                    collection_id,
+                    revision,
+                    [
+                        ("delete", vector_id, person_numeric_id, face_id)
+                        for vector_id, person_numeric_id, face_id in mappings
+                    ],
+                )
+            cursor = connection.execute(
+                "DELETE FROM persons WHERE collection_id=? AND id=?",
+                (collection_id, person_id),
+            )
+        return cursor.rowcount > 0, crops, mutation
+
+    def list_faces(
+        self, collection_id: str, person_id: str, after: str, limit: int
+    ) -> list[dict[str, Any]]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                f"""SELECT {self._face_columns(include_embedding=False)}
+                   FROM face_samples
+                   WHERE collection_id=? AND person_id=? AND id>?
+                   ORDER BY id LIMIT ?""",
+                (collection_id, person_id, after, limit),
+            ).fetchall()
+        return [self._face_row(row, include_embedding=False) for row in rows]
+
+    def list_face_embeddings(
+        self, collection_id: str, person_id: str
+    ) -> list[np.ndarray]:
+        """Return private FP32 embeddings for one Person's enrollment review."""
+
+        with self.database.read() as connection:
+            rows = connection.execute(
+                """SELECT embedding FROM face_samples
+                   WHERE collection_id=? AND person_id=? ORDER BY id""",
+                (collection_id, person_id),
+            ).fetchall()
+        return [
+            np.frombuffer(row["embedding"], dtype=np.float32).copy() for row in rows
+        ]
+
+    def get_face(
+        self, collection_id: str, person_id: str, face_id: str
+    ) -> dict[str, Any] | None:
+        with self.database.read() as connection:
+            row = connection.execute(
+                f"""SELECT {self._face_columns(include_embedding=True)}
+                    FROM face_samples
+                    WHERE collection_id=? AND person_id=? AND id=?""",
+                (collection_id, person_id, face_id),
+            ).fetchone()
+        return self._face_row(row) if row else None
+
+    def get_face_crop(
+        self, collection_id: str, person_id: str, face_id: str
+    ) -> dict[str, Any] | None:
+        """Return crop payload metadata without including it in normal face reads."""
+
+        with self.database.read() as connection:
+            row = connection.execute(
+                """SELECT crop_image,crop_media_type,crop_path
+                   FROM face_samples
+                   WHERE collection_id=? AND person_id=? AND id=?""",
+                (collection_id, person_id, face_id),
+            ).fetchone()
+        if row is None or (row["crop_image"] is None and row["crop_path"] is None):
+            return None
+        return {
+            "bytes": (
+                bytes(row["crop_image"]) if row["crop_image"] is not None else None
+            ),
+            "media_type": row["crop_media_type"] or "image/jpeg",
+            "crop_path": row["crop_path"],
+        }
+
+    def delete_face(
+        self, collection_id: str, person_id: str, face_id: str
+    ) -> tuple[bool, str | None]:
+        deleted, crop, _ = self.delete_face_with_mutation(
+            collection_id, person_id, face_id
+        )
+        return deleted, crop
+
+    def delete_face_with_mutation(
+        self, collection_id: str, person_id: str, face_id: str
+    ) -> tuple[bool, str | None, SearchMutation | None]:
+        mutation: SearchMutation | None = None
+        with self.database.write() as connection:
+            row = connection.execute(
+                """SELECT f.crop_path, v.numeric_id AS vector_id,
+                          v.person_numeric_id
+                   FROM face_samples f
+                   JOIN search_vector_ids v ON v.face_id=f.id
+                   WHERE f.collection_id=? AND f.person_id=? AND f.id=?""",
+                (collection_id, person_id, face_id),
+            ).fetchone()
+            if not row:
+                return False, None, None
+            revision = self._next_search_revision(connection, collection_id)
+            mutation = self._record_search_changes(
+                connection,
+                collection_id,
+                revision,
+                [
+                    (
+                        "delete",
+                        int(row["vector_id"]),
+                        int(row["person_numeric_id"]),
+                        face_id,
+                    )
+                ],
+            )
+            connection.execute("DELETE FROM face_samples WHERE id=?", (face_id,))
+            connection.execute(
+                "UPDATE persons SET updated_at=? WHERE collection_id=? AND id=?",
+                (utc_now(), collection_id, person_id),
+            )
+        return True, row["crop_path"], mutation
+
+    def all_faces(self, collection_id: str) -> list[dict[str, Any]]:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                f"""SELECT {self._face_columns(include_embedding=True)}
+                    FROM face_samples WHERE collection_id=? ORDER BY person_id,id""",
+                (collection_id,),
+            ).fetchall()
+        return [self._face_row(row) for row in rows]
+
+    @staticmethod
+    def _ensure_person_numeric_id(
+        connection: sqlite3.Connection, collection_id: str, person_id: str
+    ) -> int:
+        connection.execute(
+            """INSERT OR IGNORE INTO search_person_ids(collection_id,person_id)
+               VALUES(?,?)""",
+            (collection_id, person_id),
+        )
+        row = connection.execute(
+            """SELECT numeric_id FROM search_person_ids
+               WHERE collection_id=? AND person_id=?""",
+            (collection_id, person_id),
+        ).fetchone()
+        if row is None:
+            raise RuntimeError("failed to allocate stable person search identifier")
+        return int(row["numeric_id"])
+
+    @staticmethod
+    def _next_search_revision(
+        connection: sqlite3.Connection, collection_id: str
+    ) -> int:
+        cursor = connection.execute(
+            """UPDATE collections
+               SET search_revision=search_revision+1, updated_at=?
+               WHERE id=?""",
+            (utc_now(), collection_id),
+        )
+        if cursor.rowcount != 1:
+            raise KeyError(collection_id)
+        row = connection.execute(
+            "SELECT search_revision FROM collections WHERE id=?", (collection_id,)
+        ).fetchone()
+        assert row is not None
+        return int(row["search_revision"])
+
+    @staticmethod
+    def _record_search_changes(
+        connection: sqlite3.Connection,
+        collection_id: str,
+        revision: int,
+        changes: Sequence[tuple[str, int | None, int | None, str | None]],
+    ) -> SearchMutation:
+        recorded: list[SearchChange] = []
+        created_at = utc_now()
+        for operation, vector_id, person_numeric_id, face_id in changes:
+            cursor = connection.execute(
+                """INSERT INTO search_changes(
+                       collection_id,revision,operation,vector_id,
+                       person_numeric_id,face_id,created_at
+                   ) VALUES(?,?,?,?,?,?,?)""",
+                (
+                    collection_id,
+                    revision,
+                    operation,
+                    vector_id,
+                    person_numeric_id,
+                    face_id,
+                    created_at,
+                ),
+            )
+            if cursor.lastrowid is None:
+                raise RuntimeError("failed to allocate search change sequence")
+            recorded.append(
+                SearchChange(
+                    seq=int(cursor.lastrowid),
+                    collection_id=collection_id,
+                    revision=revision,
+                    operation=operation,
+                    vector_id=vector_id,
+                    person_numeric_id=person_numeric_id,
+                    face_id=face_id,
+                    created_at=created_at,
+                )
+            )
+        return SearchMutation(collection_id, revision, tuple(recorded))
+
+    @staticmethod
+    def _face_mappings_for_person(
+        connection: sqlite3.Connection, collection_id: str, person_id: str
+    ) -> list[tuple[int, int, str]]:
+        rows = connection.execute(
+            """SELECT numeric_id,person_numeric_id,face_id
+               FROM search_vector_ids
+               WHERE collection_id=? AND person_id=? ORDER BY numeric_id""",
+            (collection_id, person_id),
+        ).fetchall()
+        return [
+            (int(row["numeric_id"]), int(row["person_numeric_id"]), row["face_id"])
+            for row in rows
+        ]
+
+    @staticmethod
+    def _check_face_limits(
+        connection: sqlite3.Connection,
+        collection_id: str,
+        *,
+        person_id: str | None,
+        add_count: int,
+    ) -> None:
+        if add_count < 0:
+            raise ValueError("add_count must not be negative")
+        row = connection.execute(
+            """SELECT capacity_rows,max_faces_per_person
+               FROM collections WHERE id=?""",
+            (collection_id,),
+        ).fetchone()
+        if row is None:
+            raise KeyError(collection_id)
+        collection_count = int(
+            connection.execute(
+                "SELECT count(*) FROM face_samples WHERE collection_id=?",
+                (collection_id,),
+            ).fetchone()[0]
+        )
+        desired_collection_count = collection_count + add_count
+        capacity_rows = int(row["capacity_rows"])
+        if desired_collection_count > capacity_rows:
+            raise CollectionCapacityExceeded(
+                capacity_rows=capacity_rows,
+                requested_rows=desired_collection_count,
+            )
+        current_person_count = 0
+        if person_id is not None:
+            current_person_count = int(
+                connection.execute(
+                    """SELECT count(*) FROM face_samples
+                       WHERE collection_id=? AND person_id=?""",
+                    (collection_id, person_id),
+                ).fetchone()[0]
+            )
+        desired_person_count = current_person_count + add_count
+        max_faces = int(row["max_faces_per_person"])
+        if desired_person_count > max_faces:
+            raise MaxFacesPerPersonExceeded(
+                max_faces=max_faces,
+                requested_faces=desired_person_count,
+            )
+
+    @staticmethod
+    def _search_change_row(row: sqlite3.Row) -> SearchChange:
+        return SearchChange(
+            seq=int(row["seq"]),
+            collection_id=row["collection_id"],
+            revision=int(row["revision"]),
+            operation=row["operation"],
+            vector_id=int(row["vector_id"]) if row["vector_id"] is not None else None,
+            person_numeric_id=(
+                int(row["person_numeric_id"])
+                if row["person_numeric_id"] is not None
+                else None
+            ),
+            face_id=row["face_id"],
+            created_at=row["created_at"],
+        )
+
+    def get_search_revision(self, collection_id: str) -> int | None:
+        with self.database.read() as connection:
+            row = connection.execute(
+                "SELECT search_revision FROM collections WHERE id=?", (collection_id,)
+            ).fetchone()
+        return int(row["search_revision"]) if row else None
+
+    def get_search_state(self, collection_id: str) -> SearchState | None:
+        with self.database.read() as connection:
+            row = connection.execute(
+                """SELECT c.search_revision,
+                          coalesce((SELECT max(s.seq) FROM search_changes s
+                                    WHERE s.collection_id=c.id), 0) AS max_seq
+                   FROM collections c WHERE c.id=?""",
+                (collection_id,),
+            ).fetchone()
+        if row is None:
+            return None
+        return SearchState(collection_id, int(row["search_revision"]), int(row["max_seq"]))
+
+    def get_search_changes(
+        self, collection_id: str, after_seq: int = 0, limit: int = 1000
+    ) -> list[SearchChange]:
+        if after_seq < 0:
+            raise ValueError("after_seq must not be negative")
+        if limit < 1:
+            raise ValueError("limit must be positive")
+        with self.database.read() as connection:
+            rows = connection.execute(
+                """SELECT * FROM search_changes
+                   WHERE collection_id=? AND seq>?
+                   ORDER BY seq LIMIT ?""",
+                (collection_id, after_seq, limit),
+            ).fetchall()
+        return [self._search_change_row(row) for row in rows]
+
+    def acknowledge_search_changes(
+        self, collection_id: str, through_seq: int
+    ) -> int:
+        """Delete durable index events through a globally monotonic sequence."""
+
+        if through_seq < 0:
+            raise ValueError("through_seq must not be negative")
+        with self.database.write() as connection:
+            cursor = connection.execute(
+                """DELETE FROM search_changes
+                   WHERE collection_id=? AND seq<=?""",
+                (collection_id, through_seq),
+            )
+        return int(cursor.rowcount)
+
+    def reset_search_changes_before_full_rebuild(self) -> int:
+        """Discard stale events when no in-memory generation survives startup."""
+
+        with self.database.write() as connection:
+            cursor = connection.execute("DELETE FROM search_changes")
+        return int(cursor.rowcount)
+
+    def iter_index_faces(
+        self, collection_id: str, batch_size: int = 1000
+    ) -> Iterator[IndexFaceBatch]:
+        """Stream an index build from one WAL snapshot without materializing it."""
+
+        if batch_size < 1:
+            raise ValueError("batch_size must be positive")
+        with self.database.read() as connection:
+            # Explicit BEGIN holds one stable read snapshot while WAL writers continue.
+            connection.execute("BEGIN")
+            collection = connection.execute(
+                """SELECT c.search_revision,
+                          coalesce((SELECT max(s.seq) FROM search_changes s
+                                    WHERE s.collection_id=c.id), 0) AS max_seq
+                   FROM collections c WHERE c.id=?""",
+                (collection_id,),
+            ).fetchone()
+            if collection is None:
+                raise KeyError(collection_id)
+            revision = int(collection["search_revision"])
+            max_seq = int(collection["max_seq"])
+            cursor = connection.execute(
+                """SELECT v.numeric_id AS vector_id,v.person_numeric_id,
+                          v.face_id,v.person_id,f.embedding
+                   FROM search_vector_ids v
+                   JOIN face_samples f ON f.id=v.face_id
+                   WHERE v.collection_id=? ORDER BY v.numeric_id""",
+                (collection_id,),
+            )
+            emitted = False
+            while rows := cursor.fetchmany(batch_size):
+                emitted = True
+                yield IndexFaceBatch(
+                    revision,
+                    tuple(
+                        IndexFace(
+                            vector_id=int(row["vector_id"]),
+                            person_numeric_id=int(row["person_numeric_id"]),
+                            face_id=row["face_id"],
+                            person_id=row["person_id"],
+                            embedding=np.frombuffer(row["embedding"], dtype=np.float32).copy(),
+                        )
+                        for row in rows
+                    ),
+                    max_seq,
+                )
+            if not emitted:
+                yield IndexFaceBatch(revision, (), max_seq)
+
+    def get_index_faces_by_vector_ids(
+        self, collection_id: str, vector_ids: Sequence[int]
+    ) -> dict[int, IndexFace]:
+        result: dict[int, IndexFace] = {}
+        unique_ids = list(dict.fromkeys(int(value) for value in vector_ids))
+        with self.database.read() as connection:
+            for offset in range(0, len(unique_ids), 500):
+                chunk = unique_ids[offset : offset + 500]
+                if not chunk:
+                    continue
+                placeholders = ",".join("?" for _ in chunk)
+                rows = connection.execute(
+                    f"""SELECT v.numeric_id AS vector_id,v.person_numeric_id,
+                               v.face_id,v.person_id,f.embedding
+                        FROM search_vector_ids v
+                        JOIN face_samples f ON f.id=v.face_id
+                        WHERE v.collection_id=?
+                          AND v.numeric_id IN ({placeholders})""",
+                    [collection_id, *chunk],
+                ).fetchall()
+                for row in rows:
+                    vector_id = int(row["vector_id"])
+                    result[vector_id] = IndexFace(
+                        vector_id=vector_id,
+                        person_numeric_id=int(row["person_numeric_id"]),
+                        face_id=row["face_id"],
+                        person_id=row["person_id"],
+                        embedding=np.frombuffer(row["embedding"], dtype=np.float32).copy(),
+                    )
+        return result
+
+    def get_index_face_mappings(
+        self,
+        collection_id: str,
+        vector_ids: Sequence[int],
+        *,
+        include_embedding: bool = False,
+    ) -> dict[int, dict[str, Any]]:
+        """Hydrate native results in batches, avoiding one query per Person."""
+
+        result: dict[int, dict[str, Any]] = {}
+        unique_ids = list(dict.fromkeys(int(value) for value in vector_ids))
+        with self.database.read() as connection:
+            for offset in range(0, len(unique_ids), 400):
+                chunk = unique_ids[offset : offset + 400]
+                if not chunk:
+                    continue
+                placeholders = ",".join("?" for _ in chunk)
+                rows = connection.execute(
+                    f"""SELECT v.numeric_id AS vector_id,v.person_numeric_id,
+                               v.face_id,v.person_id,
+                               p.name,p.external_id,p.metadata_json,
+                               p.created_at AS person_created_at,
+                               p.updated_at AS person_updated_at,
+                               (SELECT count(*) FROM face_samples pf
+                                WHERE pf.collection_id=p.collection_id
+                                  AND pf.person_id=p.id) AS face_count,
+                               f.embedding
+                        FROM search_vector_ids v
+                        JOIN persons p ON p.collection_id=v.collection_id
+                                      AND p.id=v.person_id
+                        JOIN face_samples f ON f.id=v.face_id
+                        WHERE v.collection_id=?
+                          AND v.numeric_id IN ({placeholders})""",
+                    [collection_id, *chunk],
+                ).fetchall()
+                for row in rows:
+                    vector_id = int(row["vector_id"])
+                    value: dict[str, Any] = {
+                        "vector_id": vector_id,
+                        "person_numeric_id": int(row["person_numeric_id"]),
+                        "face_id": row["face_id"],
+                        "person_id": row["person_id"],
+                        "person": {
+                            "id": row["person_id"],
+                            "name": row["name"],
+                            "external_id": row["external_id"],
+                            "metadata": json.loads(row["metadata_json"]),
+                            "face_count": int(row["face_count"]),
+                            "created_at": row["person_created_at"],
+                            "updated_at": row["person_updated_at"],
+                        },
+                    }
+                    if include_embedding:
+                        value["embedding"] = np.frombuffer(
+                            row["embedding"], dtype=np.float32
+                        ).copy()
+                    result[vector_id] = value
+        return result
+
+    def stats(self) -> dict[str, int]:
+        with self.database.read() as connection:
+            collection_count = connection.execute("SELECT count(*) FROM collections").fetchone()[0]
+            person_count = connection.execute("SELECT count(*) FROM persons").fetchone()[0]
+            face_count = connection.execute("SELECT count(*) FROM face_samples").fetchone()[0]
+        return {
+            "collection_count": int(collection_count),
+            "person_count": int(person_count),
+            "face_count": int(face_count),
+        }
+
+    def sync_api_key(self, raw_key: str) -> str:
+        """Make ``raw_key`` the only active key, retaining inactive history.
+
+        The comparison and any replacement happen inside one SQLite write
+        transaction. Repeating the same startup key is a true no-op: its row,
+        salt, digest, and creation timestamp remain unchanged.
+        """
+
+        if not raw_key:
+            raise ValueError("API key must not be empty")
+        with self.database.write() as connection:
+            rows = connection.execute(
+                "SELECT id,salt,digest FROM api_keys WHERE active=1 ORDER BY created_at,id"
+            ).fetchall()
+            matching_id: str | None = None
+            for row in rows:
+                candidate = hashlib.scrypt(
+                    raw_key.encode(), salt=row["salt"], n=2**14, r=8, p=1, dklen=32
+                )
+                if hmac.compare_digest(candidate, row["digest"]):
+                    matching_id = str(row["id"])
+                    break
+            if matching_id is not None and len(rows) == 1:
+                return "unchanged"
+            if matching_id is not None:
+                connection.execute(
+                    "UPDATE api_keys SET active=0 WHERE active=1 AND id<>?",
+                    (matching_id,),
+                )
+                return "rotated"
+
+            salt = os.urandom(16)
+            digest = hashlib.scrypt(
+                raw_key.encode(), salt=salt, n=2**14, r=8, p=1, dklen=32
+            )
+            result = "created" if not rows else "rotated"
+            connection.execute("UPDATE api_keys SET active=0 WHERE active=1")
+            connection.execute(
+                "INSERT INTO api_keys(id,label,salt,digest,active,created_at) "
+                "VALUES(?,?,?,?,1,?)",
+                (str(uuid.uuid4()), "startup", salt, digest, utc_now()),
+            )
+            return result
+
+    def has_api_keys(self) -> bool:
+        with self.database.read() as connection:
+            return bool(
+                connection.execute("SELECT 1 FROM api_keys WHERE active=1 LIMIT 1").fetchone()
+            )
+
+    def verify_api_key(self, raw_key: str) -> bool:
+        with self.database.read() as connection:
+            rows = connection.execute(
+                "SELECT salt,digest FROM api_keys WHERE active=1"
+            ).fetchall()
+        for row in rows:
+            candidate = hashlib.scrypt(
+                raw_key.encode(), salt=row["salt"], n=2**14, r=8, p=1, dklen=32
+            )
+            if hmac.compare_digest(candidate, row["digest"]):
+                return True
+        return False
diff --git a/server/backend/insightface_server/storage/secrets.py b/server/backend/insightface_server/storage/secrets.py
new file mode 100644
index 000000000..17357caf6
--- /dev/null
+++ b/server/backend/insightface_server/storage/secrets.py
@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import base64
+import os
+from pathlib import Path
+
+from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+
+def _encode(value: bytes) -> str:
+    return base64.urlsafe_b64encode(value).rstrip(b"=").decode("ascii")
+
+
+def _decode(value: str) -> bytes:
+    decoded = base64.urlsafe_b64decode(value + "=" * (-len(value) % 4))
+    if _encode(decoded) != value:
+        raise ValueError("non-canonical base64")
+    return decoded
+
+
+class SecretCodec:
+    """Encrypt recoverable local secrets with a key stored beside the database."""
+
+    def __init__(self, secret_path: Path):
+        secret_path.parent.mkdir(parents=True, exist_ok=True)
+        created = os.urandom(32)
+        try:
+            descriptor = os.open(
+                secret_path,
+                os.O_WRONLY | os.O_CREAT | os.O_EXCL,
+                0o600,
+            )
+            with os.fdopen(descriptor, "wb") as stream:
+                stream.write(created)
+            self._secret = created
+        except FileExistsError:
+            flags = os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0)
+            descriptor = os.open(secret_path, flags)
+            with os.fdopen(descriptor, "rb") as stream:
+                self._secret = stream.read()
+        secret_path.chmod(0o600)
+        if len(self._secret) != 32:
+            raise RuntimeError("Monitor credential key must contain exactly 32 bytes")
+        self._cipher = AESGCM(self._secret)
+
+    def encrypt(self, value: str, *, scope: str) -> str:
+        nonce = os.urandom(12)
+        ciphertext = self._cipher.encrypt(
+            nonce,
+            value.encode("utf-8"),
+            scope.encode("utf-8"),
+        )
+        return f"v1.{_encode(nonce + ciphertext)}"
+
+    def decrypt(self, value: str, *, scope: str) -> str:
+        try:
+            version, encoded = value.split(".", 1)
+            if version != "v1":
+                raise ValueError("unsupported version")
+            payload = _decode(encoded)
+            if len(payload) < 13:
+                raise ValueError("truncated value")
+            plaintext = self._cipher.decrypt(
+                payload[:12],
+                payload[12:],
+                scope.encode("utf-8"),
+            )
+            return plaintext.decode("utf-8")
+        except Exception as exc:
+            raise RuntimeError("Stored Monitor credentials could not be decrypted") from exc
diff --git a/server/config/server.toml b/server/config/server.toml
new file mode 100644
index 000000000..be7243c82
--- /dev/null
+++ b/server/config/server.toml
@@ -0,0 +1,34 @@
+# InsightFace Server settings loaded once during process startup.
+# Restart the container after changing this file.
+
+[inference]
+# "auto" resolves to 4 concurrent model pipelines on CPU and 8 on CUDA.
+# A positive integer overrides the provider-specific default. API calls,
+# enrollment and RTSP frames share this one process-wide budget.
+max_concurrency = "auto"
+
+[detection]
+# Each entry is [width, height]. Dynamic SCRFD models run every configured
+# resolution, map all candidates to source-image coordinates, then apply one
+# global NMS pass to the merged candidate set.
+input_sizes = [[96, 96], [512, 512]]
+
+# Minimum detector confidence. It is applied while SCRFD candidates are
+# generated, before the merged NMS pass.
+threshold = 0.50
+
+# IoU threshold used by the single global NMS pass.
+nms_threshold = 0.40
+
+# Used by operations that require one face. Supported values:
+# "largest" and "center_largest". The latter maximizes the pixel-space score
+# area - 2.0 * squared_distance(face_box_center, image_center).
+single_face_selection = "largest"
+
+# Deployment-wide safety limit. A request may ask for fewer results, never more.
+max_detected_faces = 100
+
+[web]
+# false (default): serve the Web UI, interactive API reference and guides.
+# true: API-only mode; keep /v1 and /openapi.json, but do not register UI routes.
+disabled = false
diff --git a/server/deploy/compose.cpu.yml b/server/deploy/compose.cpu.yml
new file mode 100644
index 000000000..1c7c1d0bc
--- /dev/null
+++ b/server/deploy/compose.cpu.yml
@@ -0,0 +1,90 @@
+name: insightface-simple-cpu
+
+x-models-path: &models-path ../.models
+
+services:
+  server:
+    image: ghcr.io/deepinsight/insightface-server:0.2.0-cpu
+    platform: linux/amd64
+    init: true
+    read_only: true
+    user: "10001:10001"
+    restart: unless-stopped
+    stop_grace_period: 30s
+    ports:
+      - "18097:8080"
+    environment:
+      INSIGHTFACE_AUTH_ENABLED: "${INSIGHTFACE_AUTH_ENABLED:-false}"
+      INSIGHTFACE_CONFIG_FILE: /etc/insightface/server.toml
+      INSIGHTFACE_API_KEY: "${INSIGHTFACE_API_KEY:-}"
+      INSIGHTFACE_CORS_ORIGINS: "${INSIGHTFACE_CORS_ORIGINS:-}"
+      INSIGHTFACE_LOG_LEVEL: "${INSIGHTFACE_LOG_LEVEL:-INFO}"
+      INSIGHTFACE_SAVE_FACE_CROPS: "${INSIGHTFACE_SAVE_FACE_CROPS:-false}"
+      INSIGHTFACE_DEFAULT_THRESHOLD: "${INSIGHTFACE_DEFAULT_THRESHOLD:-0.4}"
+      INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE: "${INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE:-fp32_v1}"
+      INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS: "${INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS:-100000}"
+      INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS: "${INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS:-10000000}"
+      INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON: "${INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON:-20}"
+      INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY: "${INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY:-lazy}"
+      INSIGHTFACE_SEARCH_DEVICE_ID: "${INSIGHTFACE_SEARCH_DEVICE_ID:-0}"
+      INSIGHTFACE_SEARCH_TOPK_MODE: "${INSIGHTFACE_SEARCH_TOPK_MODE:-auto}"
+      INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS: "${INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS:-4096}"
+      INSIGHTFACE_INFERENCE_MODE: onnx
+      INSIGHTFACE_EXECUTION_PROVIDER: CPUExecutionProvider
+    volumes:
+      - type: bind
+        source: ../config/server.toml
+        target: /etc/insightface/server.toml
+        read_only: true
+        bind:
+          create_host_path: false
+      - type: volume
+        source: data
+        target: /data
+      - type: bind
+        source: *models-path
+        target: /models
+        read_only: true
+        bind:
+          create_host_path: false
+    tmpfs:
+      - /tmp:rw,noexec,nosuid,nodev,size=128m
+    cap_drop: [ALL]
+    security_opt: [no-new-privileges:true]
+    pids_limit: 512
+    shm_size: 256m
+
+  models:
+    profiles: [tools]
+    image: ghcr.io/deepinsight/insightface-server:0.2.0-cpu
+    platform: linux/amd64
+    init: true
+    read_only: true
+    user: "${INSIGHTFACE_MODELS_UID:-1000}:${INSIGHTFACE_MODELS_GID:-1000}"
+    group_add: ["10001"]
+    entrypoint: ["python", "-m", "insightface_server.models_cli"]
+    environment:
+      INSIGHTFACE_MODELS_DIR: /models
+      HTTP_PROXY: "${HTTP_PROXY:-}"
+      HTTPS_PROXY: "${HTTPS_PROXY:-}"
+      NO_PROXY: "${NO_PROXY:-}"
+      http_proxy: "${http_proxy:-}"
+      https_proxy: "${https_proxy:-}"
+      no_proxy: "${no_proxy:-}"
+    volumes:
+      - type: bind
+        source: *models-path
+        target: /models
+        bind:
+          create_host_path: false
+    tmpfs:
+      - /tmp:rw,noexec,nosuid,nodev,size=64m
+    cap_drop: [ALL]
+    security_opt: [no-new-privileges:true]
+    pids_limit: 128
+    healthcheck:
+      disable: true
+
+volumes:
+  data:
+    name: insightface-simple-cpu-data
diff --git a/server/deploy/compose.cuda12.yml b/server/deploy/compose.cuda12.yml
new file mode 100644
index 000000000..8c71cf8aa
--- /dev/null
+++ b/server/deploy/compose.cuda12.yml
@@ -0,0 +1,95 @@
+name: insightface-simple-cuda12
+
+x-models-path: &models-path ../.models
+
+services:
+  server:
+    image: ghcr.io/deepinsight/insightface-server:0.2.0-cuda12
+    platform: linux/amd64
+    init: true
+    read_only: true
+    user: "10001:10001"
+    restart: unless-stopped
+    stop_grace_period: 30s
+    ports:
+      - "18098:8080"
+    environment:
+      INSIGHTFACE_AUTH_ENABLED: "${INSIGHTFACE_AUTH_ENABLED:-false}"
+      INSIGHTFACE_CONFIG_FILE: /etc/insightface/server.toml
+      INSIGHTFACE_API_KEY: "${INSIGHTFACE_API_KEY:-}"
+      INSIGHTFACE_CORS_ORIGINS: "${INSIGHTFACE_CORS_ORIGINS:-}"
+      INSIGHTFACE_LOG_LEVEL: "${INSIGHTFACE_LOG_LEVEL:-INFO}"
+      INSIGHTFACE_SAVE_FACE_CROPS: "${INSIGHTFACE_SAVE_FACE_CROPS:-false}"
+      INSIGHTFACE_DEFAULT_THRESHOLD: "${INSIGHTFACE_DEFAULT_THRESHOLD:-0.4}"
+      INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE: "${INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE:-fp32_v1}"
+      INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS: "${INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS:-100000}"
+      INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS: "${INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS:-10000000}"
+      INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON: "${INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON:-20}"
+      INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY: "${INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY:-lazy}"
+      INSIGHTFACE_SEARCH_DEVICE_ID: "${INSIGHTFACE_SEARCH_DEVICE_ID:-0}"
+      INSIGHTFACE_SEARCH_TOPK_MODE: "${INSIGHTFACE_SEARCH_TOPK_MODE:-auto}"
+      INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS: "${INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS:-4096}"
+      INSIGHTFACE_INFERENCE_MODE: onnx
+      INSIGHTFACE_EXECUTION_PROVIDER: CUDAExecutionProvider
+      INSIGHTFACE_STRICT_CUDA: "1"
+      NVIDIA_VISIBLE_DEVICES: all
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+      CUDA_MODULE_LOADING: LAZY
+    gpus: all
+    volumes:
+      - type: bind
+        source: ../config/server.toml
+        target: /etc/insightface/server.toml
+        read_only: true
+        bind:
+          create_host_path: false
+      - type: volume
+        source: data
+        target: /data
+      - type: bind
+        source: *models-path
+        target: /models
+        read_only: true
+        bind:
+          create_host_path: false
+    tmpfs:
+      - /tmp:rw,noexec,nosuid,nodev,size=128m
+    cap_drop: [ALL]
+    security_opt: [no-new-privileges:true]
+    pids_limit: 512
+    shm_size: 256m
+
+  models:
+    profiles: [tools]
+    image: ghcr.io/deepinsight/insightface-server:0.2.0-cuda12
+    platform: linux/amd64
+    init: true
+    read_only: true
+    user: "${INSIGHTFACE_MODELS_UID:-1000}:${INSIGHTFACE_MODELS_GID:-1000}"
+    group_add: ["10001"]
+    entrypoint: ["python", "-m", "insightface_server.models_cli"]
+    environment:
+      INSIGHTFACE_MODELS_DIR: /models
+      HTTP_PROXY: "${HTTP_PROXY:-}"
+      HTTPS_PROXY: "${HTTPS_PROXY:-}"
+      NO_PROXY: "${NO_PROXY:-}"
+      http_proxy: "${http_proxy:-}"
+      https_proxy: "${https_proxy:-}"
+      no_proxy: "${no_proxy:-}"
+    volumes:
+      - type: bind
+        source: *models-path
+        target: /models
+        bind:
+          create_host_path: false
+    tmpfs:
+      - /tmp:rw,noexec,nosuid,nodev,size=64m
+    cap_drop: [ALL]
+    security_opt: [no-new-privileges:true]
+    pids_limit: 128
+    healthcheck:
+      disable: true
+
+volumes:
+  data:
+    name: insightface-simple-cuda12-data
diff --git a/server/docker/Dockerfile.cpu b/server/docker/Dockerfile.cpu
new file mode 100644
index 000000000..1485e5432
--- /dev/null
+++ b/server/docker/Dockerfile.cpu
@@ -0,0 +1,106 @@
+# syntax=docker/dockerfile:1.7@sha256:a57df69d0ea827fb7266491f2813635de6f17269be881f696fbfdf2d83dda33e
+
+FROM python:3.11-slim-bookworm@sha256:28255a3ace7eb4c48bc1b57b90af29e1bc82b4fd6c60614a8e3dce61b87ff941 AS native-search-builder
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+RUN set -eux; \
+    for attempt in 1 2 3 4 5; do \
+      if apt-get -o Acquire::Retries=5 update \
+        && apt-get -o Acquire::Retries=5 install --yes --no-install-recommends \
+          build-essential cmake; then \
+        break; \
+      fi; \
+      test "$attempt" -lt 5; \
+      sleep 2; \
+    done; \
+    rm -rf /var/lib/apt/lists/*
+
+COPY server/native/search /src/native-search
+RUN cmake -S /src/native-search -B /build/native-search \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DIFS_SEARCH_BUILD_CPU=ON \
+      -DIFS_SEARCH_BUILD_CUDA=OFF \
+      -DIFS_SEARCH_BUILD_TESTS=ON \
+    && cmake --build /build/native-search --parallel \
+    && ctest --test-dir /build/native-search --output-on-failure
+
+FROM python:3.11-slim-bookworm@sha256:28255a3ace7eb4c48bc1b57b90af29e1bc82b4fd6c60614a8e3dce61b87ff941
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+LABEL org.opencontainers.image.title="InsightFace Server CPU" \
+      org.opencontainers.image.description="Simple InsightFace Server for linux/amd64 CPU inference" \
+      org.opencontainers.image.source="https://github.com/deepinsight/insightface"
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PYTHONPATH=/opt/insightface/server/backend:/opt/insightface/python-package \
+    INSIGHTFACE_DATA_DIR=/data \
+    INSIGHTFACE_MODELS_DIR=/models \
+    INSIGHTFACE_CONFIG_FILE=/etc/insightface/server.toml \
+    INSIGHTFACE_INFERENCE_MODE=onnx \
+    INSIGHTFACE_EXECUTION_PROVIDER=CPUExecutionProvider \
+    INSIGHTFACE_SEARCH_BACKEND=native_cpu \
+    INSIGHTFACE_SEARCH_LIBRARY=/opt/insightface/server/native/lib/libifs_search_cpu.so
+
+RUN apt-get -o Acquire::Retries=5 update \
+    && apt-get -o Acquire::Retries=5 install --yes --no-install-recommends ca-certificates libgomp1 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY server/requirements.common.lock server/requirements.cpu.lock /tmp/requirements/
+RUN python -m pip install --no-cache-dir -r /tmp/requirements/requirements.cpu.lock
+
+COPY server/scripts/verify_runtime.py /opt/insightface/server/scripts/verify_runtime.py
+RUN python /opt/insightface/server/scripts/verify_runtime.py \
+      --mode build --provider CPUExecutionProvider
+
+# Only inference modules required by the server are copied. Bundled sample images
+# and training/evaluation code are intentionally excluded from the final image.
+COPY python-package/insightface/__init__.py /opt/insightface/python-package/insightface/__init__.py
+COPY python-package/insightface/app /opt/insightface/python-package/insightface/app
+COPY python-package/insightface/model_zoo /opt/insightface/python-package/insightface/model_zoo
+COPY python-package/insightface/utils /opt/insightface/python-package/insightface/utils
+COPY python-package/insightface/data/__init__.py \
+     python-package/insightface/data/image.py \
+     python-package/insightface/data/pickle_object.py \
+     /opt/insightface/python-package/insightface/data/
+COPY python-package/insightface/thirdparty/__init__.py /opt/insightface/python-package/insightface/thirdparty/__init__.py
+
+COPY server/backend /opt/insightface/server/backend
+COPY server/frontend /opt/insightface/server/frontend
+COPY server/migrations /opt/insightface/server/migrations
+COPY server/config/server.toml /etc/insightface/server.toml
+COPY server/LICENSING.md /opt/insightface/server/LICENSING.md
+COPY server/docs/api*.md /opt/insightface/server/docs/
+COPY server/docs/user-guide*.md /opt/insightface/server/docs/
+COPY server/docs/maintainer-guide.md /opt/insightface/server/docs/maintainer-guide.md
+COPY server/docs/images /opt/insightface/server/docs/images
+COPY --from=native-search-builder \
+     /build/native-search/cpu/libifs_search_cpu.so.2.0.0 \
+     /opt/insightface/server/native/lib/libifs_search_cpu.so
+
+RUN groupadd --gid 10001 insightface \
+    && useradd --uid 10001 --gid 10001 --no-create-home --home-dir /nonexistent \
+       --shell /usr/sbin/nologin insightface \
+    && install -d -o insightface -g insightface -m 0750 /data \
+    && install -d -o root -g insightface -m 0550 /models \
+    && chown -R root:insightface /opt/insightface \
+    && chmod -R u=rwX,g=rX,o= /opt/insightface
+
+ARG INSIGHTFACE_SERVER_VERSION=0.2.0
+LABEL org.opencontainers.image.version="${INSIGHTFACE_SERVER_VERSION}"
+
+WORKDIR /opt/insightface/server
+USER 10001:10001
+
+VOLUME ["/data"]
+EXPOSE 8080
+STOPSIGNAL SIGTERM
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=90s --retries=3 \
+  CMD ["python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8080/v1/health', timeout=3).read()"]
+
+ENTRYPOINT ["python", "/opt/insightface/server/scripts/verify_runtime.py", "--mode", "startup", "--provider", "CPUExecutionProvider", "--"]
+CMD ["python", "-m", "uvicorn", "insightface_server.app:app", "--host", "0.0.0.0", "--port", "8080", "--no-access-log"]
diff --git a/server/docker/Dockerfile.cuda12 b/server/docker/Dockerfile.cuda12
new file mode 100644
index 000000000..99e157829
--- /dev/null
+++ b/server/docker/Dockerfile.cuda12
@@ -0,0 +1,128 @@
+# syntax=docker/dockerfile:1.7@sha256:a57df69d0ea827fb7266491f2813635de6f17269be881f696fbfdf2d83dda33e
+
+FROM nvidia/cuda:12.9.1-devel-ubuntu22.04@sha256:bd4e2680a261c212f1e2fea241606f71497dc67a417f73175d794ec8212b5ba8 AS native-search-builder
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get -o Acquire::Retries=5 update \
+    && apt-get -o Acquire::Retries=5 install --yes --no-install-recommends cmake g++ \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY server/native/search /src/native-search
+RUN cmake -S /src/native-search -B /build/native-search \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DIFS_SEARCH_BUILD_CPU=OFF \
+      -DIFS_SEARCH_BUILD_CUDA=ON \
+      -DIFS_SEARCH_BUILD_TESTS=ON \
+    && cmake --build /build/native-search --parallel \
+    && ctest --test-dir /build/native-search --output-on-failure
+
+FROM nvcr.io/nvidia/cuda:12.9.1-runtime-ubuntu22.04@sha256:6553b9635f35d992cf0473f55d1e998935a2dd1e2e604d3cbfb2bf295a8faa79
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+LABEL org.opencontainers.image.title="InsightFace Server CUDA 12" \
+      org.opencontainers.image.description="Simple InsightFace Server for linux/amd64 NVIDIA GPU inference" \
+      org.opencontainers.image.source="https://github.com/deepinsight/insightface"
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PATH=/opt/venv/bin:$PATH \
+    PYTHONPATH=/opt/insightface/server/backend:/opt/insightface/python-package \
+    INSIGHTFACE_DATA_DIR=/data \
+    INSIGHTFACE_MODELS_DIR=/models \
+    INSIGHTFACE_CONFIG_FILE=/etc/insightface/server.toml \
+    INSIGHTFACE_INFERENCE_MODE=onnx \
+    INSIGHTFACE_EXECUTION_PROVIDER=CUDAExecutionProvider \
+    INSIGHTFACE_STRICT_CUDA=1 \
+    INSIGHTFACE_SEARCH_BACKEND=native_cuda \
+    INSIGHTFACE_SEARCH_LIBRARY=/opt/insightface/server/native/lib/libifs_search_cuda.so \
+    NVIDIA_DRIVER_CAPABILITIES=compute,utility \
+    CUDA_MODULE_LOADING=LAZY
+
+# Ubuntu 22.04 does not include Python 3.11. The deadsnakes repository is used
+# inside the image only; this never changes the customer's host packages.
+RUN set -eux; \
+    for attempt in 1 2 3 4 5; do \
+      if apt-get -o Acquire::Retries=5 update \
+        && apt-get -o Acquire::Retries=5 install --yes --no-install-recommends \
+          ca-certificates libgomp1 software-properties-common; then \
+        break; \
+      fi; \
+      test "$attempt" -lt 5; \
+      sleep 2; \
+    done; \
+    for attempt in 1 2 3 4 5; do \
+      if add-apt-repository --yes ppa:deadsnakes/ppa; then \
+        break; \
+      fi; \
+      test "$attempt" -lt 5; \
+      sleep 2; \
+    done; \
+    for attempt in 1 2 3 4 5; do \
+      if apt-get -o Acquire::Retries=5 update \
+        && apt-get -o Acquire::Retries=5 install --fix-missing --yes \
+          --no-install-recommends python3.11 python3.11-venv \
+          libcudnn9-cuda-12=9.24.0.43-1; then \
+        break; \
+      fi; \
+      test "$attempt" -lt 5; \
+      sleep 2; \
+    done; \
+    python3.11 -m venv /opt/venv; \
+    rm -rf /var/lib/apt/lists/*
+
+COPY server/requirements.common.lock server/requirements.cuda12.lock /tmp/requirements/
+RUN python -m pip install --no-cache-dir -r /tmp/requirements/requirements.cuda12.lock
+
+COPY server/scripts/verify_runtime.py /opt/insightface/server/scripts/verify_runtime.py
+RUN python /opt/insightface/server/scripts/verify_runtime.py \
+      --mode build --provider CUDAExecutionProvider
+
+COPY python-package/insightface/__init__.py /opt/insightface/python-package/insightface/__init__.py
+COPY python-package/insightface/app /opt/insightface/python-package/insightface/app
+COPY python-package/insightface/model_zoo /opt/insightface/python-package/insightface/model_zoo
+COPY python-package/insightface/utils /opt/insightface/python-package/insightface/utils
+COPY python-package/insightface/data/__init__.py \
+     python-package/insightface/data/image.py \
+     python-package/insightface/data/pickle_object.py \
+     /opt/insightface/python-package/insightface/data/
+COPY python-package/insightface/thirdparty/__init__.py /opt/insightface/python-package/insightface/thirdparty/__init__.py
+
+COPY server/backend /opt/insightface/server/backend
+COPY server/frontend /opt/insightface/server/frontend
+COPY server/migrations /opt/insightface/server/migrations
+COPY server/config/server.toml /etc/insightface/server.toml
+COPY server/LICENSING.md /opt/insightface/server/LICENSING.md
+COPY server/docs/api*.md /opt/insightface/server/docs/
+COPY server/docs/user-guide*.md /opt/insightface/server/docs/
+COPY server/docs/maintainer-guide.md /opt/insightface/server/docs/maintainer-guide.md
+COPY server/docs/images /opt/insightface/server/docs/images
+COPY --from=native-search-builder \
+     /build/native-search/cuda/libifs_search_cuda.so.2.0.0 \
+     /opt/insightface/server/native/lib/libifs_search_cuda.so
+
+RUN groupadd --gid 10001 insightface \
+    && useradd --uid 10001 --gid 10001 --no-create-home --home-dir /nonexistent \
+       --shell /usr/sbin/nologin insightface \
+    && install -d -o insightface -g insightface -m 0750 /data \
+    && install -d -o root -g insightface -m 0550 /models \
+    && chown -R root:insightface /opt/insightface \
+    && chmod -R u=rwX,g=rX,o= /opt/insightface
+
+ARG INSIGHTFACE_SERVER_VERSION=0.2.0
+LABEL org.opencontainers.image.version="${INSIGHTFACE_SERVER_VERSION}"
+
+WORKDIR /opt/insightface/server
+USER 10001:10001
+
+VOLUME ["/data"]
+EXPOSE 8080
+STOPSIGNAL SIGTERM
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=120s --retries=3 \
+  CMD ["python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8080/v1/health', timeout=3).read()"]
+
+ENTRYPOINT ["python", "/opt/insightface/server/scripts/verify_runtime.py", "--mode", "startup", "--provider", "CUDAExecutionProvider", "--"]
+CMD ["python", "-m", "uvicorn", "insightface_server.app:app", "--host", "0.0.0.0", "--port", "8080", "--no-access-log"]
diff --git a/server/docs/api.de.md b/server/docs/api.de.md
new file mode 100644
index 000000000..a494b9a54
--- /dev/null
+++ b/server/docs/api.de.md
@@ -0,0 +1,170 @@
+# InsightFace Server REST-API-Leitfaden
+
+**Sprachen:** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · Deutsch · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+Dieses Dokument erklärt Zweck, Eingaben, Serververhalten, Erfolg und typische Fehler aller öffentlichen Endpunkte. Start und Modellinstallation stehen im [Benutzerhandbuch](user-guide.de.md); das exakte laufende Schema finden Sie unter `/docs` und `/openapi.json`.
+
+## Gemeinsame Regeln
+
+- Basispfad `/v1`, JSON in `snake_case`, Bilder als JPEG/PNG/WebP multipart.
+- Die mitgelieferte Compose-Konfiguration deaktiviert Authentifizierung für isolierte Tests. Ist sie aktiv, benötigen alle Endpunkte außer health `Authorization: Bearer <api_key>`. Bei deaktivierter Authentifizierung den Header ganz weglassen.
+- Jede Antwort trägt `x-request-id`; JSON wiederholt ihn als `request_id`.
+- Confidence/quality/threshold liegen in `0..1`. Similarity ist keine Wahrscheinlichkeit, sondern roher Cosinus in `[-1,1]`; Standard `0.4`, Treffer bei `similarity >= threshold`.
+- Cursor sind undurchsichtig und nur unverändert mit demselben Pfad, Collection-, Person- und Filterkontext wiederzuverwenden.
+- Übliche Statuscodes: 400 Eingabe, 401 Auth, 404 nicht gefunden, 409 Konflikt, 413 Größe, 422 Bild/Gesicht, 429 Limit, 503 Timeout/Modell/Index.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## System
+
+### `GET /v1/health`
+
+**Zweck/Eingabe:** Öffentliche Readiness, keine Parameter und keine Authentifizierung. **Ergebnis:** Prüft Start und SQLite quick_check; 200 mit `status`, `auth_enabled`, `request_id`. **Fehler:** `503 not_ready`.
+
+### `GET /v1/system`
+
+**Zweck/Eingabe:** Sichere Betriebsdiagnose, keine Parameter. **Ergebnis:** 200 mit OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, Modell, DB, Mounts, Zählern, Suchbackend, sicherer Konfiguration und Inferenz-Parallelität; keine Secrets/Bilder/Embeddings. **Fehler:** 401, 503.
+
+### `GET /v1/models`
+
+**Zweck/Eingabe:** Verifizierte Detector-/Recognizer-Modelle, Provider und Lizenz lesen; keine Parameter. **Ergebnis:** 200 `models`, `execution_provider`, `license`. **Fehler:** 401.
+
+## Zustandslose Bildoperationen
+
+### `POST /v1/detect`
+
+**Eingabe:** multipart `image` erforderlich, `max_faces` 1–100, optional `collection_id`. **Verhalten/Ergebnis:** Mehrfachauflösung, gemeinsame NMS, Flächensortierung; 200 `faces` mit Boxen/5 Punkten/Score/Qualität und `processing_ms`. Kein Gesicht ist eine leere Erfolgsliste. **Fehler:** 400 alter min_score, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**Eingabe:** multipart `source`, `target`, optional `threshold` 0–1 und `collection_id`. **Ergebnis:** Wählt je ein Gesicht und liefert 200 `matched`, Cosinus-`similarity`, effektiven threshold, beide Faces und Laufzeit. **Fehler:** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**Eingabe:** multipart `image`, optional `collection_id`. **Ergebnis:** 200 mit ausgewähltem Face, L2-normalisiertem Embedding, Modell und Laufzeit. Für normale Registrierung nicht nötig; Embeddings werden nicht geloggt. **Fehler:** 400 alter face_selection, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Eingabe:** JSON `id` und `name`; optional description, threshold (Standard 0.4), metadata, save_face_crops, `detection` und `search` mit profile/capacity/max_faces_per_person/load_policy. **Verhalten/Ergebnis:** Bindet Modell, Vorverarbeitung und Suchvertrag; 201 mit vollständig aufgelöster `collection`. **Fehler:** 400 Konfiguration, 409 exists, 503 Index.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**Eingabe:** query `limit` 1–100 (Standard 50), optionaler Cursor. **Ergebnis:** 200 `collections`, nullable `next_cursor`. **Fehler:** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**Eingabe:** Collection-ID im Pfad. **Ergebnis:** 200 `collection`, Person-/Face-Zähler, `embedding_contract_id`. **Fehler:** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Eingabe:** Pfad-ID; JSON name/description/threshold/metadata/save_face_crops, Suchkapazität/max/load und Detection. Null, unbekannte Felder, Modellbindung und search profile sind nicht änderbar. **Ergebnis:** 200 vollständige Collection; neue Detection gilt ab dem nächsten Request. **Fehler:** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Eingabe:** Pfad-ID, query `force=false`; für nicht leere Collection true. **Ergebnis:** 204 ohne Body. **Fehler:** 404, 409 collection_not_empty, 503.
+
+## Personen und FaceSamples
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Eingabe:** Pfad Collection; multipart wiederholtes `images`, optional id/name/external_id, JSON-String metadata, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; extern benötigt Vektoren und contract ID. **Verhalten/Ergebnis:** Prüft jedes Bild; 201 `person`, akzeptierte `faces`, `rejected_images`; Teilerfolg erlaubt, alle abgelehnt ergibt 422 ohne Person. **Fehler:** 400, 404, 409 ID/Vertrag/Kapazität, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Eingabe:** Collection-ID, query limit/cursor/`search` über ID, Name oder externe ID. **Ergebnis:** 200 `persons`, `next_cursor`. **Fehler:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Eingabe:** Collection- und Person-ID. **Ergebnis:** 200 `person` mit face_count. **Fehler:** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Eingabe:** Pfad-IDs; JSON name/external_id/object metadata. **Ergebnis:** 200 aktualisierte Person. **Fehler:** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Eingabe:** Pfad-IDs. **Ergebnis:** Löscht Person, alle FaceSamples, Embeddings und Crops, synchronisiert Index, 204. **Fehler:** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Eingabe:** Pfad-IDs; wiederholte images und dieselben Review-/Embedding-Felder wie Person-Erstellung. **Ergebnis:** 201 `faces`, `rejected_images`, Teilerfolg möglich. **Fehler:** Registrierungsfehler plus 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Eingabe:** Pfad-IDs; query limit 1–100 und Cursor. **Ergebnis:** 200 Face-Metadaten, `has_crop`, `next_cursor`, ohne Embedding/Bildbytes. **Fehler:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Eingabe:** drei Pfad-IDs. **Ergebnis:** Falls gespeichert 200 `image/jpeg`, 112×112 Crop, `Cache-Control:no-store`; Request-ID nur im Header. **Fehler:** 401, 404 Face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Eingabe:** drei Pfad-IDs. **Ergebnis:** Löscht Embedding/Crop/Indexzeile, 204. **Fehler:** 404, 503.
+
+## Suche
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Eingabe:** Collection-ID; multipart `image`, `limit` 1–100 (Standard 5), optional threshold, sonst Collection-Wert. **Verhalten/Ergebnis:** Sucht das gewählte Face über alle Samples, pro Person gilt der höchste Wert; 200 `searched_face`, sortierte `matches`, threshold und Laufzeit. Kein Treffer ist leere Liste. **Fehler:** 404, 409 Modell, 413, 422 Bild/Gesicht, 503 Index/Timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## RTSP-Monitore
+
+Monitor-Konfigurationen liegen dauerhaft in SQLite; aktivierte Aufgaben werden nach einem Server-Neustart wiederhergestellt. Videobilder werden nicht gespeichert, Ereignisse nur in einem begrenzten RAM-Ringpuffer.
+
+### `POST /v1/monitors`
+
+**Zweck:** Einen dauerhaften RTSP-Erkennungsmonitor anlegen. **Eingabe:** JSON mit ID, Name, `source`, Collection, `inference_fps` (Standard 2), optionalem Schwellwert, Puffer/Event-Regeln und `preview_enabled` (Standard false). **Ergebnis:** 201 mit redigiertem `monitor`; Zugangsdaten werden verschlüsselt gespeichert. **Fehler:** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**Zweck:** Monitore seitenweise auflisten. **Eingabe:** `limit` 1–100 (Standard 50) und der unveränderte, undurchsichtige `cursor` der vorigen Antwort. **Ergebnis:** 200 mit `monitors` und `next_cursor`, niemals mit Zugangsdaten. **Fehler:** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Zweck:** Konfiguration und Laufzeitübersicht eines Monitors lesen. **Eingabe:** `monitor_id` im Pfad. **Ergebnis:** 200 mit Event-Regeln, redigierter Quelle, Preview-Einstellung und Status. **Fehler:** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Zweck:** Alle veränderbaren Felder teilweise ändern und über `enabled` starten/stoppen. **Eingabe:** JSON-Teilobjekt; auch `event_policy` ist partiell, ein null-Schwellwert erbt den Collection-Wert. **Ergebnis:** 200 mit vollständigem Monitor; Quelle/Collection/Rate/Regeln starten die Aufgabe neu. **Fehler:** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Zweck:** Einen Monitor dauerhaft entfernen. **Eingabe:** `monitor_id` im Pfad. **Ergebnis:** Decoder, Inferenz und RTSP-Verbindung werden gestoppt, RAM-Ereignisse verworfen, HTTP 204; die Collection bleibt bestehen. **Fehler:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Zweck:** Den aktuellen Zustand für Headless-Clients abfragen. **Eingabe:** Monitor-ID im Pfad. **Ergebnis:** 200 mit Verbindung, effektiver FPS, Laufzeit, übersprungenen Frames, aktuellen Treffern/unbekannten Gesichtern, Preview, Reconnects und Fehlern, ohne Embeddings. **Fehler:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Zweck:** Flüchtige Enter/Exit/Error/Recovery-Ereignisse abrufen. **Eingabe:** `limit` 1–1000 und der letzte undurchsichtige `cursor`. **Ergebnis:** 200 mit `events`, `next_cursor`, `truncated` und `stream_reset`; Neustarts löschen Ereignisse. **Fehler:** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Zweck:** Die standardmäßig deaktivierte rohe MJPEG-Vorschau öffnen. **Eingabe:** Pfad-ID und normaler Bearer-Header; kein API-Key in der URL. **Ergebnis:** Lang laufendes `multipart/x-mixed-replace`, nur bei Zuschauern codiert; Boxen zeichnet der Client aus `/state`. **Fehler:** 401, 404, 409 `preview_disabled`, 503.
+
+## Retry-Regel
+
+GET darf wiederholt werden. DELETE erst nach Statusprüfung. Bei unklarem Netzwerkergebnis einer Person-/Face-Erstellung vor erneutem POST per ID lesen. Nur 429 und temporäre 503 mit begrenztem exponentiellem Backoff plus Jitter wiederholen; 4xx-Eingaben korrigieren.
diff --git a/server/docs/api.es.md b/server/docs/api.es.md
new file mode 100644
index 000000000..8b59a8f09
--- /dev/null
+++ b/server/docs/api.es.md
@@ -0,0 +1,170 @@
+# Guía de uso de la API REST de InsightFace Server
+
+**Idiomas:** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · Español · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+Esta guía explica el propósito, la entrada, el trabajo del servidor, el resultado y los errores de todas las rutas públicas. Consulte instalación y primer uso en la [guía de usuario](user-guide.es.md) y el esquema exacto en ejecución en `/docs` o `/openapi.json`.
+
+## Reglas comunes
+
+- Ruta base `/v1`, JSON `snake_case`, imágenes JPEG/PNG/WebP como multipart.
+- El Compose incluido desactiva auth para evaluación aislada. Si se activa, todo salvo health requiere `Authorization: Bearer <api_key>`; si está desactivada, omita el header, no envíe uno vacío.
+- Toda respuesta tiene `x-request-id`; JSON repite `request_id`.
+- confidence/quality/threshold usan `0..1`. Similarity no es probabilidad: es coseno `[-1,1]`; valor predeterminado `0.4`, coincide si `similarity >= threshold`.
+- Un cursor es opaco y solo se reutiliza sin cambios con la misma ruta, Collection, Person y filtro.
+- Estados habituales: 400 entrada, 401 auth, 404 ausente, 409 conflicto, 413 tamaño, 422 imagen/rostro, 429 límite, 503 timeout/modelo/índice.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## Sistema
+
+### `GET /v1/health`
+
+**Uso/entrada:** readiness pública, sin parámetros ni autenticación. **Resultado:** comprueba inicio y SQLite quick_check; 200 con `status`, `auth_enabled`, `request_id`. **Error:** `503 not_ready`.
+
+### `GET /v1/system`
+
+**Uso/entrada:** diagnóstico seguro, sin parámetros. **Resultado:** 200 con OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, modelo, DB, montajes, totales, búsqueda, configuración segura y concurrencia; nunca secretos, imágenes ni embeddings. **Errores:** 401, 503.
+
+### `GET /v1/models`
+
+**Uso/entrada:** modelos detector/recognizer verificados, Provider y licencia; sin parámetros. **Resultado:** 200 `models`, `execution_provider`, `license`. **Error:** 401.
+
+## Operaciones faciales sin estado
+
+### `POST /v1/detect`
+
+**Entrada:** multipart `image` obligatorio, `max_faces` 1–100, `collection_id` opcional. **Proceso/resultado:** fusiona resoluciones, NMS global y orden por área; 200 `faces` con cajas/5 puntos/score/calidad y `processing_ms`. Sin rostro es lista vacía correcta. **Errores:** 400 min_score antiguo, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**Entrada:** multipart `source`, `target`, `threshold` opcional 0–1 y `collection_id`. **Resultado:** elige un rostro por imagen; 200 `matched`, coseno `similarity`, threshold efectivo, ambos rostros y tiempo. **Errores:** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**Entrada:** multipart `image`, `collection_id` opcional. **Resultado:** 200 con rostro seleccionado, embedding L2, modelo y tiempo. No hace falta para registro normal y el vector no se registra en logs. **Errores:** 400 face_selection antiguo, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Entrada:** JSON `id`, `name`; opcionales description, threshold (0.4), metadata, save_face_crops, `detection` y `search` con profile/capacity/max_faces_per_person/load_policy. **Proceso/resultado:** fija modelo, preprocesamiento y contrato de búsqueda; 201 con `collection` resuelta. **Errores:** 400 configuración, 409 exists, 503 índice.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**Entrada:** query `limit` 1–100 (50), cursor opcional. **Resultado:** 200 `collections`, `next_cursor` nullable. **Errores:** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**Entrada:** ID de Collection en path. **Resultado:** 200 `collection`, conteos Person/Face y `embedding_contract_id`. **Error:** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Entrada:** path ID; JSON name/description/threshold/metadata/save_face_crops, capacidad/max/load de búsqueda y detection. No admite null, campos desconocidos, cambiar modelo ni search profile. **Resultado:** 200 Collection completa; detection rige desde la siguiente petición. **Errores:** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Entrada:** path ID; query `force=false`, true para no vacía. **Resultado:** 204 sin body. **Errores:** 404, 409 collection_not_empty, 503.
+
+## Personas y FaceSamples
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Entrada:** path Collection; multipart `images` repetible, id/name/external_id opcionales, metadata como JSON string, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; modo externo añade vectores y contract ID. **Proceso/resultado:** revisa cada imagen; 201 `person`, `faces` aceptadas y `rejected_images`; admite éxito parcial, todo rechazado devuelve 422 sin crear Person. **Errores:** 400, 404, 409 ID/contrato/capacidad, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Entrada:** Collection; query limit/cursor/`search` por ID, nombre o external ID. **Resultado:** 200 `persons`, `next_cursor`. **Errores:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs Collection y Person. **Resultado:** 200 `person` con face_count. **Error:** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs; JSON name/external_id/metadata objeto. **Resultado:** 200 Person actualizada. **Errores:** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs. **Resultado:** elimina Person, FaceSamples, embeddings y crops, sincroniza índice, 204. **Errores:** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrada:** IDs; images repetibles y mismos campos review/embedding que crear Person. **Resultado:** 201 `faces`, `rejected_images`, éxito parcial. **Errores:** registro más 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrada:** IDs; query limit 1–100 y cursor. **Resultado:** 200 metadata de `faces`, `has_crop`, `next_cursor`, sin embedding ni bytes. **Errores:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Entrada:** tres IDs. **Resultado:** si existe, 200 `image/jpeg`, crop 112×112, `Cache-Control:no-store`; request ID solo en header. **Errores:** 401, 404 face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Entrada:** tres IDs. **Resultado:** elimina embedding/crop/fila del índice, 204. **Errores:** 404, 503.
+
+## Búsqueda
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Entrada:** Collection; multipart `image`, `limit` 1–100 (5), threshold opcional o valor de Collection. **Proceso/resultado:** compara el rostro elegido con todas las muestras y usa el máximo por Person; 200 `searched_face`, `matches` ordenados, threshold y tiempo. Sin match es lista vacía. **Errores:** 404, 409 modelo, 413, 422 imagen/rostro, 503 índice/timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## Monitores RTSP
+
+La configuración del Monitor persiste en SQLite y una tarea habilitada se restaura al reiniciar el servidor. No se guardan fotogramas; los eventos viven solo en un búfer circular limitado en memoria.
+
+### `POST /v1/monitors`
+
+**Uso:** Crear un Monitor RTSP persistente. **Entrada:** JSON con ID, nombre, `source`, Collection, `inference_fps` (2), umbral opcional, búfer/política de eventos y `preview_enabled` (false). **Resultado:** 201 con `monitor` censurado; las credenciales se cifran al almacenarse. **Errores:** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**Uso:** Enumerar Monitores con paginación. **Entrada:** `limit` 1–100 (50) y el `cursor` opaco de la respuesta anterior, sin modificarlo. **Resultado:** 200 con `monitors` y `next_cursor`, nunca credenciales. **Errores:** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Uso:** Leer la configuración y el resumen de ejecución de un Monitor. **Entrada:** `monitor_id` en la ruta. **Resultado:** 200 con política de eventos, fuente censurada, vista previa y estado. **Errores:** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Uso:** Actualizar parcialmente campos salvo el ID y arrancar/parar con `enabled`. **Entrada:** JSON parcial; `event_policy` también es parcial y umbral null hereda la Collection. **Resultado:** 200 con el Monitor completo; cambiar fuente/Collection/frecuencia/política reinicia la tarea. **Errores:** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Uso:** Eliminar permanentemente un Monitor. **Entrada:** `monitor_id` en la ruta. **Resultado:** detiene decodificador, inferencia y RTSP, descarta eventos de memoria y devuelve 204; no elimina la Collection. **Errores:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Uso:** Consultar el estado actual desde clientes sin interfaz. **Entrada:** ID del Monitor. **Resultado:** 200 con conexión, FPS efectivo, latencia, saltos, rostros reconocidos/desconocidos, vista previa, reconexiones y error seguro, sin embeddings. **Errores:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Uso:** Obtener eventos volátiles de entrada/salida/error/recuperación. **Entrada:** `limit` 1–1000 y el último `cursor` opaco. **Resultado:** 200 con `events`, `next_cursor`, `truncated` y `stream_reset`; el reinicio pierde los eventos. **Errores:** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Uso:** Abrir la vista previa MJPEG cruda, desactivada por defecto. **Entrada:** ID y cabecera Bearer normal; nunca la clave en la URL. **Resultado:** `multipart/x-mixed-replace` largo, codificado solo con espectadores; el cliente dibuja cajas desde `/state`. **Errores:** 401, 404, 409 `preview_disabled`, 503.
+
+## Reintentos
+
+GET se puede reintentar. Verifique estado antes de repetir DELETE. Si el resultado de crear Person/Face es incierto por red, consulte el ID antes de repetir POST. Reintente solo 429 y 503 transitorios con backoff exponencial limitado y jitter; corrija los 4xx.
diff --git a/server/docs/api.fr.md b/server/docs/api.fr.md
new file mode 100644
index 000000000..088bd09f8
--- /dev/null
+++ b/server/docs/api.fr.md
@@ -0,0 +1,170 @@
+# Guide d’utilisation de l’API REST InsightFace Server
+
+**Langues :** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · Français · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+Ce guide décrit l’objectif, les entrées, le traitement serveur, le résultat et les erreurs de chaque route publique. Pour installer et effectuer une première recherche, consultez le [guide utilisateur](user-guide.fr.md). Le schéma exact de l’instance est disponible sous `/docs` et `/openapi.json`.
+
+## Règles communes
+
+- Base `/v1`, JSON en `snake_case`, images JPEG/PNG/WebP en multipart.
+- Le Compose fourni désactive l’authentification pour une évaluation isolée. Lorsqu’elle est active, tout sauf health exige `Authorization: Bearer <api_key>` ; sinon omettez totalement cet en-tête.
+- Chaque réponse porte `x-request-id`, répété dans le JSON par `request_id`.
+- confidence/quality/threshold utilisent `0..1`. Similarity n’est pas une probabilité : cosine brut `[-1,1]`, seuil par défaut `0.4`, correspondance si `similarity >= threshold`.
+- Un cursor est opaque et ne se réutilise, inchangé, qu’avec la même route, Collection, Person et filtre.
+- Codes usuels : 400 entrée, 401 auth, 404 absent, 409 conflit, 413 taille, 422 image/visage, 429 limite, 503 timeout/modèle/index.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## Système
+
+### `GET /v1/health`
+
+**Usage/entrée :** readiness publique, sans paramètre ni auth. **Résultat :** vérifie démarrage et SQLite quick_check ; 200 avec `status`, `auth_enabled`, `request_id`. **Erreur :** `503 not_ready`.
+
+### `GET /v1/system`
+
+**Usage/entrée :** diagnostic sûr, sans paramètre. **Résultat :** 200 avec OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, modèle, DB, montages, compteurs, recherche, configuration sûre et concurrence ; jamais de secrets, images ou embeddings. **Erreurs :** 401, 503.
+
+### `GET /v1/models`
+
+**Usage/entrée :** modèles detector/recognizer vérifiés, Provider et licence ; sans paramètre. **Résultat :** 200 `models`, `execution_provider`, `license`. **Erreur :** 401.
+
+## Opérations faciales sans état
+
+### `POST /v1/detect`
+
+**Entrée :** multipart `image` requis, `max_faces` 1–100, `collection_id` facultatif. **Traitement/résultat :** fusion multi-résolution, NMS globale, tri par aire ; 200 `faces` avec boîtes/5 points/score/qualité et `processing_ms`. Aucun visage donne une liste vide valide. **Erreurs :** 400 ancien min_score, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**Entrée :** multipart `source`, `target`, `threshold` facultatif 0–1 et `collection_id`. **Résultat :** choisit un visage par image ; 200 `matched`, cosine `similarity`, seuil effectif, deux visages et durée. **Erreurs :** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**Entrée :** multipart `image`, `collection_id` facultatif. **Résultat :** 200 avec visage choisi, embedding L2, modèle et durée. Inutile à l’inscription normale ; le vecteur n’est pas journalisé. **Erreurs :** 400 ancien face_selection, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Entrée :** JSON `id`, `name`; facultatifs description, threshold (0.4), metadata, save_face_crops, `detection`, `search` avec profile/capacity/max_faces_per_person/load_policy. **Traitement/résultat :** fixe modèle, prétraitement et contrat de recherche ; 201 avec `collection` résolue. **Erreurs :** 400 configuration, 409 exists, 503 index.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**Entrée :** query `limit` 1–100 (50), cursor facultatif. **Résultat :** 200 `collections`, `next_cursor` nullable. **Erreurs :** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**Entrée :** ID Collection dans le chemin. **Résultat :** 200 `collection`, compteurs Person/Face et `embedding_contract_id`. **Erreur :** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Entrée :** ID ; JSON name/description/threshold/metadata/save_face_crops, capacité/max/load de recherche et detection. Null, champs inconnus, modèle et search profile sont immuables. **Résultat :** 200 Collection complète ; detection s’applique à la requête suivante. **Erreurs :** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Entrée :** ID ; query `force=false`, true si non vide. **Résultat :** 204 sans corps. **Erreurs :** 404, 409 collection_not_empty, 503.
+
+## Personnes et FaceSamples
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Entrée :** Collection ; multipart `images` répétable, id/name/external_id facultatifs, metadata JSON texte, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; le mode externe ajoute vecteurs et contract ID. **Traitement/résultat :** contrôle chaque image ; 201 `person`, `faces` acceptées et `rejected_images`, succès partiel autorisé ; tout rejeté donne 422 sans Person. **Erreurs :** 400, 404, 409 ID/contrat/capacité, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Entrée :** Collection ; query limit/cursor/`search` sur ID, nom ou ID externe. **Résultat :** 200 `persons`, `next_cursor`. **Erreurs :** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrée :** IDs Collection et Person. **Résultat :** 200 `person` avec face_count. **Erreur :** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrée :** IDs ; JSON name/external_id/metadata objet. **Résultat :** 200 Person mise à jour. **Erreurs :** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrée :** IDs. **Résultat :** supprime Person, FaceSamples, embeddings et crops, synchronise l’index, 204. **Erreurs :** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrée :** IDs ; images répétables et mêmes champs review/embedding que la création. **Résultat :** 201 `faces`, `rejected_images`, succès partiel. **Erreurs :** inscription plus 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrée :** IDs ; query limit 1–100 et cursor. **Résultat :** 200 métadonnées `faces`, `has_crop`, `next_cursor`, sans embedding ni octets. **Erreurs :** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Entrée :** trois IDs. **Résultat :** si conservé, 200 `image/jpeg`, crop 112×112, `Cache-Control:no-store`; request ID dans l’en-tête seulement. **Erreurs :** 401, 404 face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Entrée :** trois IDs. **Résultat :** supprime embedding/crop/ligne d’index, 204. **Erreurs :** 404, 503.
+
+## Recherche
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Entrée :** Collection ; multipart `image`, `limit` 1–100 (5), threshold facultatif ou valeur Collection. **Traitement/résultat :** compare le visage choisi à tous les samples et garde le maximum par Person ; 200 `searched_face`, `matches` triés, seuil et durée. Aucun match est une liste vide. **Erreurs :** 404, 409 modèle, 413, 422 image/visage, 503 index/timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## Moniteurs RTSP
+
+La configuration d’un Monitor persiste dans SQLite et toute tâche activée est restaurée au redémarrage du serveur. Les images ne sont pas enregistrées ; les événements restent dans un tampon circulaire borné en mémoire.
+
+### `POST /v1/monitors`
+
+**Usage :** Créer un Monitor RTSP persistant. **Entrée :** JSON avec ID, nom, `source`, Collection, `inference_fps` (2), seuil facultatif, tampon/politique d’événements et `preview_enabled` (false). **Résultat :** 201 avec `monitor` masqué ; les identifiants sont chiffrés au repos. **Erreurs :** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**Usage :** Lister les Monitors avec pagination. **Entrée :** `limit` 1–100 (50) et le `cursor` opaque de la réponse précédente, inchangé. **Résultat :** 200 avec `monitors` et `next_cursor`, jamais les identifiants. **Erreurs :** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Usage :** Lire la configuration et le résumé d’exécution d’un Monitor. **Entrée :** `monitor_id` dans le chemin. **Résultat :** 200 avec politique d’événements, source masquée, aperçu et état. **Erreurs :** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Usage :** Modifier partiellement sauf l’ID et démarrer/arrêter via `enabled`. **Entrée :** JSON partiel ; `event_policy` est aussi partiel et un seuil null hérite de la Collection. **Résultat :** 200 avec le Monitor complet ; source/Collection/fréquence/politique relancent la tâche. **Erreurs :** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Usage :** Supprimer définitivement un Monitor. **Entrée :** `monitor_id` dans le chemin. **Résultat :** arrête décodage, inférence et RTSP, efface les événements mémoire puis renvoie 204 ; la Collection reste. **Erreurs :** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Usage :** Interroger l’état courant depuis un client sans interface. **Entrée :** ID du Monitor. **Résultat :** 200 avec connexion, FPS effectif, latence, images sautées, visages reconnus/inconnus, aperçu, reconnexions et erreur sûre, sans embeddings. **Erreurs :** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Usage :** Lire les événements volatils entrée/sortie/erreur/rétablissement. **Entrée :** `limit` 1–1000 et le dernier `cursor` opaque. **Résultat :** 200 avec `events`, `next_cursor`, `truncated` et `stream_reset` ; un redémarrage perd les événements. **Erreurs :** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Usage :** Ouvrir l’aperçu MJPEG brut, désactivé par défaut. **Entrée :** ID et en-tête Bearer normal ; jamais la clé dans l’URL. **Résultat :** long `multipart/x-mixed-replace`, encodé seulement avec spectateurs ; le client trace les cadres via `/state`. **Erreurs :** 401, 404, 409 `preview_disabled`, 503.
+
+## Réessais
+
+GET peut être réessayé. Vérifiez l’état avant de répéter DELETE. Si une création Person/Face a un résultat réseau incertain, lisez l’ID avant un nouveau POST. Réessayez seulement 429 et 503 transitoires avec backoff exponentiel borné et jitter ; corrigez les 4xx.
diff --git a/server/docs/api.ja.md b/server/docs/api.ja.md
new file mode 100644
index 000000000..6bf8d87c8
--- /dev/null
+++ b/server/docs/api.ja.md
@@ -0,0 +1,170 @@
+# InsightFace Server REST API 利用ガイド
+
+**言語:** [English](api.md) · [中文](api.zh-CN.md) · 日本語 · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+この文書は全公開 API の用途、入力、サーバー処理、正常結果、主なエラーを説明します。コンテナとモデルの準備は [ユーザーガイド](user-guide.ja.md)、実行中の厳密な Schema は `/docs` または `/openapi.json` を参照してください。
+
+## 共通規則と最初の呼び出し
+
+- ベースパスは `/v1`、JSON は `snake_case`、画像は JPEG/PNG/WebP の multipart です。
+- 同梱 Compose は隔離評価向けに認証を既定で無効にします。有効時は health 以外へ `Authorization: Bearer <api_key>` を送ります。無効時は空の Authorization を送らず省略します。
+- 全レスポンスに `x-request-id`、JSON には同じ `request_id` があります。
+- confidence/quality/threshold は `0..1`。similarity は確率ではなく `[-1,1]` の生 cosine で、既定しきい値は `0.4`、一致条件は `similarity >= threshold` です。
+- `cursor` は不透明です。同じパス、Collection、Person、filter にそのまま返してください。
+- 主な状態は 400 入力、401 認証、404 不在、409 競合、413 サイズ、422 画像/顔、429 上限、503 timeout/model/index です。
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## システム
+
+### `GET /v1/health`
+
+**用途/入力:** 公開 readiness。パラメーターなし、認証不要。**処理/結果:** DB quick check と起動状態を確認し、200 で `status`、`auth_enabled`、`request_id`。**エラー:** 未準備は `503 not_ready`。
+
+### `GET /v1/system`
+
+**用途/入力:** 安全な運用診断。パラメーターなし。**結果:** 200 で OS/CPU/GPU、Driver、CUDA/cuDNN/ORT、Provider、モデル、DB、mount、件数、検索 backend、安全設定、推論並行数。秘密・画像・embedding は含みません。**エラー:** 401、503。
+
+### `GET /v1/models`
+
+**用途/入力:** 検証済み detector/recognizer、実 Provider、License を確認。パラメーターなし。**結果:** 200 の `models`、`execution_provider`、`license`。**エラー:** 401。
+
+## ステートレス顔処理
+
+### `POST /v1/detect`
+
+**入力:** multipart `image` 必須、`max_faces` 1～100、任意 `collection_id`。**処理/結果:** 複数解像度候補を統合して一度 NMS、面積順の `faces`、box/5点/score/quality、`processing_ms`。顔なしは 200 の空配列。**エラー:** 400 旧 min_score、404 Collection、413、422 invalid_image、503。
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**入力:** multipart `source`、`target` 必須、`threshold` 0～1、任意 `collection_id`。**処理/結果:** 各画像から設定戦略で1顔を選び、200 で `matched`、cosine `similarity`、実 threshold、両 face、処理時間。**エラー:** 404、413、422 invalid_image/face_not_found、503。
+
+### `POST /v1/embeddings`
+
+**入力:** multipart `image` 必須、任意 `collection_id`。**結果:** 200 で選択 face、L2 正規化 embedding、model、処理時間。通常登録には不要で、値はログされません。**エラー:** 400 旧 face_selection、404、413、422、503。
+
+## Collection
+
+### `POST /v1/collections`
+
+**入力:** JSON `id`、`name` 必須。任意 `description`、`threshold`(既定0.4)、`metadata`、`save_face_crops`、`detection`、`search`。search は profile/capacity_rows/max_faces_per_person/load_policy。**処理/結果:** モデル・前処理・検索契約を固定し、201 で解決済み `collection`。**エラー:** 400 profile/detection/capacity、409 exists、503 index。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**入力:** query `limit` 1～100(既定50)、任意不透明 `cursor`。**結果:** 200 の `collections` と nullable `next_cursor`。**エラー:** 400 invalid_cursor、401。
+
+### `GET /v1/collections/{collection_id}`
+
+**入力:** path `collection_id`。**結果:** 200 の `collection`、Person/Face件数、`embedding_contract_id`。**エラー:** 404。
+
+### `PATCH /v1/collections/{collection_id}`
+
+**入力:** path ID。JSON で name/description/threshold/metadata/save_face_crops、search の capacity/max/load、detection を変更。null、未知フィールド、モデルと search profile の変更は不可。**結果:** 200 の完全な更新 Collection、次リクエストから反映。**エラー:** 400、404、409、503。
+
+### `DELETE /v1/collections/{collection_id}`
+
+**入力:** path ID、query `force=false`。非空を消す時のみ true。**結果:** 204 本文なし。**エラー:** 404、409 collection_not_empty、503。
+
+## Person と FaceSample
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**入力:** path Collection。multipart の repeatable `images` 必須、任意 id/name/external_id、JSON文字列 metadata、`review_mode=off|standard|strict`、`embedding_mode=server|external_trusted`。外部特徴では vector 配列と contract ID も必須。**処理/結果:** 画像ごとに審査し、201 で `person`、受理 `faces`、`rejected_images`。部分成功可。全失敗は 422 で Person を作りません。**エラー:** 400、404、409 ID/contract/capacity、413、422、503。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**入力:** path Collection、query limit/cursor/`search`(ID・名前・外部ID)。**結果:** 200 の `persons` と `next_cursor`。**エラー:** 400 cursor、404。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**入力:** Collection ID と Person ID。**結果:** 200 の `person` と face_count。**エラー:** 404。
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**入力:** path IDs、JSON name/external_id/object metadata。**結果:** 200 の更新 `person`。**エラー:** 400、404、409 external_id_exists。
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**入力:** path IDs。**処理/結果:** Person、全 FaceSample、embedding、任意 crop を消し索引へ同期、204。**エラー:** 404、503。
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**入力:** path IDs、repeatable images と review/embedding fields（Person作成と同じ）。**結果:** 201 の `faces` と `rejected_images`、部分成功可。**エラー:** 登録エラーと 404 Person。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**入力:** path IDs、query limit 1～100 と cursor。**結果:** 200 の metadata `faces`、`has_crop`、`next_cursor`。embedding/画像byteは返しません。**エラー:** 400 cursor、404。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**入力:** 3 path IDs。**結果:** 保存済みの場合 200 `image/jpeg` 112×112 crop、`Cache-Control:no-store`。request ID はheaderのみ。**エラー:** 401、404 face/face_image_not_found。
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**入力:** 3 path IDs。**結果:** embedding/crop/索引行を削除し204。**エラー:** 404、503。
+
+## 検索
+
+### `POST /v1/collections/{collection_id}/search`
+
+**入力:** path Collection、multipart `image`、`limit` 1～100(既定5)、任意 threshold(省略時Collection値)。**処理/結果:** 選択顔を全 FaceSample と比較し、Personごとの最高値を降順で返します。200 の `searched_face`、`matches`、threshold、処理時間。該当なしは空配列。**エラー:** 404、409 model、413、422 image/face、503 index/timeout。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## RTSP Monitor
+
+Monitor設定はSQLiteに永続化され、有効なタスクはServer再起動後に復元されます。動画frameは保存せず、eventは上限付きmemory bufferだけに保持します。
+
+### `POST /v1/monitors`
+
+**用途:** 永続的なRTSP認識Monitorを作成します。**入力:** JSONのID、名前、`source`、Collection、`inference_fps`（既定2）、任意threshold、buffer/event policy、`preview_enabled`（既定false）。**結果:** 201で認証情報を除いた`monitor`。URL資格情報は暗号化保存されます。**エラー:** 400、404、409、429。
+
+### `GET /v1/monitors`
+
+**用途:** Monitor一覧をページングします。**入力:** `limit` 1～100（既定50）と、前回の不透明な`cursor`。**結果:** 200の`monitors`と`next_cursor`；認証情報は返しません。**エラー:** 400 `invalid_cursor`、401。
+
+### `GET /v1/monitors/{monitor_id}`
+
+**用途:** 1件の設定とruntime要約を取得します。**入力:** pathの`monitor_id`。**結果:** 200の`monitor`にevent policy、脱敏source、preview設定、状態を含みます。**エラー:** 401、404 `monitor_not_found`。
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**用途:** ID以外を部分更新し、`enabled`で開始/停止します。**入力:** JSONの変更フィールド；`event_policy`も部分更新でき、thresholdのnullはCollection値を継承します。**結果:** 200の完全な`monitor`。source/Collection/rate/policy変更時はtaskを再起動します。**エラー:** 400、404、429。
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**用途:** Monitorを恒久削除します。**入力:** pathの`monitor_id`。**結果:** decoder/inference/RTSP接続を停止し、memory eventを破棄して204；Collectionは削除しません。**エラー:** 401、404。
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**用途:** headless clientが現在状態をpollします。**入力:** pathのMonitor ID。**結果:** 200で接続、実効FPS、処理時間、skip、現在の認識/未登録face、preview、再接続、エラーを返し、embeddingは返しません。**エラー:** 401、404。
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**用途:** 保存しない最近のenter/exit/error/recovery eventを取得します。**入力:** `limit` 1～1000と前回の不透明な`cursor`。**結果:** 200で`events`、`next_cursor`、`truncated`、`stream_reset`；再起動でeventは失われます。**エラー:** 400 `invalid_cursor`、401、404。
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**用途:** 既定で無効のraw MJPEG previewを開きます。**入力:** path IDと通常のBearer header；API keyをURLに入れません。**結果:** viewerがいる間だけencodeする長時間`multipart/x-mixed-replace`で、boxはclientが`/state`から描画します。**エラー:** 401、404、409 `preview_disabled`、503。
+
+## クライアント実装チェック
+
+GET は再試行可能です。DELETE は状態確認後に再試行してください。Person/Face作成で通信結果が不明な場合は、同じIDを再送する前にGETで確認します。429/一時503だけを上限付き指数バックオフとjitterで再試行し、4xx入力エラーは修正してください。
diff --git a/server/docs/api.ko.md b/server/docs/api.ko.md
new file mode 100644
index 000000000..fbf790236
--- /dev/null
+++ b/server/docs/api.ko.md
@@ -0,0 +1,170 @@
+# InsightFace Server REST API 사용 가이드
+
+**언어:** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · 한국어
+
+이 문서는 모든 공개 API의 목적, 입력, 서버 처리, 성공 결과와 오류를 설명합니다. 설치와 첫 검색은 [사용자 가이드](user-guide.ko.md)를, 현재 실행 버전의 정확한 스키마는 `/docs`와 `/openapi.json`을 확인하세요.
+
+## 공통 규칙
+
+- 기본 경로 `/v1`, JSON은 `snake_case`, 이미지는 JPEG/PNG/WebP multipart입니다.
+- 제공되는 Compose는 격리 평가용으로 인증이 기본 비활성화됩니다. 활성화 시 health 외에는 `Authorization: Bearer <api_key>`가 필요하고, 비활성화 시 빈 헤더를 보내지 말고 완전히 생략합니다.
+- 모든 응답에 `x-request-id`, JSON에는 같은 `request_id`가 있습니다.
+- confidence/quality/threshold는 `0..1`입니다. Similarity는 확률이 아닌 원본 cosine `[-1,1]`이며 기본 threshold는 `0.4`, `similarity >= threshold`일 때 일치합니다.
+- cursor는 불투명하며 같은 경로, Collection, Person, filter에 변경 없이 다시 보냅니다.
+- 일반 상태: 400 입력, 401 인증, 404 없음, 409 충돌, 413 크기, 422 이미지/얼굴, 429 제한, 503 timeout/model/index.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## 시스템
+
+### `GET /v1/health`
+
+**용도/입력:** 공개 readiness, 파라미터와 인증 없음. **결과:** 시작 상태와 SQLite quick_check를 확인하고 200 `status`, `auth_enabled`, `request_id`. **오류:** `503 not_ready`.
+
+### `GET /v1/system`
+
+**용도/입력:** 안전한 운영 진단, 파라미터 없음. **결과:** 200 OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, 모델, DB, mount, 수량, 검색, 안전 설정, 추론 동시성. 비밀, 이미지, embedding은 제외됩니다. **오류:** 401, 503.
+
+### `GET /v1/models`
+
+**용도/입력:** 검증된 detector/recognizer, Provider, License 확인. 파라미터 없음. **결과:** 200 `models`, `execution_provider`, `license`. **오류:** 401.
+
+## 무상태 얼굴 처리
+
+### `POST /v1/detect`
+
+**입력:** multipart `image` 필수, `max_faces` 1–100, 선택 `collection_id`. **처리/결과:** 여러 해상도를 합쳐 전역 NMS, 면적순 정렬; 200 `faces`의 box/5점/score/quality와 `processing_ms`. 얼굴 없음은 정상 빈 목록입니다. **오류:** 400 구 min_score, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**입력:** multipart `source`, `target`, 선택 `threshold` 0–1과 `collection_id`. **결과:** 각 이미지에서 한 얼굴을 골라 200 `matched`, cosine `similarity`, 실제 threshold, 두 face, 처리 시간. **오류:** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**입력:** multipart `image`, 선택 `collection_id`. **결과:** 200 선택 face, L2 embedding, 모델, 시간. 일반 등록에는 필요 없고 vector는 로그에 남지 않습니다. **오류:** 400 구 face_selection, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**입력:** JSON `id`, `name`; 선택 description, threshold(0.4), metadata, save_face_crops, `detection`, `search`의 profile/capacity/max_faces_per_person/load_policy. **처리/결과:** 모델, 전처리, 검색 계약을 고정하고 201 완전한 `collection`. **오류:** 400 설정, 409 exists, 503 index.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**입력:** query `limit` 1–100(50), 선택 cursor. **결과:** 200 `collections`, nullable `next_cursor`. **오류:** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**입력:** 경로 Collection ID. **결과:** 200 `collection`, Person/Face 수, `embedding_contract_id`. **오류:** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**입력:** ID; JSON name/description/threshold/metadata/save_face_crops, 검색 capacity/max/load와 detection. null, 알 수 없는 필드, 모델과 search profile 변경은 불가합니다. **결과:** 200 전체 Collection; detection은 다음 요청부터 적용됩니다. **오류:** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**입력:** ID; query `force=false`, 비어 있지 않으면 true. **결과:** 204 본문 없음. **오류:** 404, 409 collection_not_empty, 503.
+
+## Person과 FaceSample
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**입력:** Collection; multipart 반복 `images`, 선택 id/name/external_id, JSON 문자열 metadata, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; 외부 모드는 벡터와 contract ID도 필요합니다. **처리/결과:** 이미지별 심사; 201 `person`, 승인 `faces`, `rejected_images`, 부분 성공 가능. 전부 거절되면 Person 없이 422. **오류:** 400, 404, 409 ID/계약/용량, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**입력:** Collection; query limit/cursor/`search`로 ID, 이름, external ID 검색. **결과:** 200 `persons`, `next_cursor`. **오류:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**입력:** Collection과 Person ID. **결과:** 200 `person`, face_count. **오류:** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**입력:** IDs; JSON name/external_id/object metadata. **결과:** 200 수정된 Person. **오류:** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**입력:** IDs. **결과:** Person, FaceSamples, embeddings, crops를 삭제하고 인덱스 동기화, 204. **오류:** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**입력:** IDs; 반복 images와 Person 생성과 같은 review/embedding 필드. **결과:** 201 `faces`, `rejected_images`, 부분 성공. **오류:** 등록 오류와 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**입력:** IDs; query limit 1–100, cursor. **결과:** 200 `faces` metadata, `has_crop`, `next_cursor`, embedding/bytes 제외. **오류:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**입력:** 세 ID. **결과:** 저장된 경우 200 `image/jpeg`, 112×112 crop, `Cache-Control:no-store`; request ID는 헤더에만 있습니다. **오류:** 401, 404 face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**입력:** 세 ID. **결과:** embedding/crop/index row 삭제, 204. **오류:** 404, 503.
+
+## 검색
+
+### `POST /v1/collections/{collection_id}/search`
+
+**입력:** Collection; multipart `image`, `limit` 1–100(5), 선택 threshold 또는 Collection 값. **처리/결과:** 선택 얼굴을 모든 sample과 비교하고 Person별 최고값 사용; 200 `searched_face`, 정렬 `matches`, threshold, 시간. 일치 없음은 빈 목록. **오류:** 404, 409 모델, 413, 422 이미지/얼굴, 503 index/timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## RTSP Monitor
+
+Monitor 설정은 SQLite에 영구 저장되고 활성화된 작업은 서버 재시작 후 복원됩니다. 영상 프레임은 저장하지 않으며 이벤트는 용량이 제한된 메모리 링 버퍼에만 유지됩니다.
+
+### `POST /v1/monitors`
+
+**용도:** 영구 RTSP 인식 Monitor를 만듭니다. **입력:** ID, 이름, `source`, Collection, `inference_fps`(2), 선택 임계값, 이벤트 버퍼/정책, `preview_enabled`(false)를 담은 JSON입니다. **결과:** 자격증명을 제거한 `monitor`와 201; 자격증명은 암호화 저장됩니다. **오류:** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**용도:** Monitor 목록을 페이지 단위로 조회합니다. **입력:** `limit` 1~100(50)과 이전 응답의 불투명한 `cursor`를 변경 없이 사용합니다. **결과:** 200의 `monitors`, `next_cursor`이며 자격증명은 반환하지 않습니다. **오류:** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**용도:** 한 Monitor의 설정과 실행 요약을 읽습니다. **입력:** 경로의 `monitor_id`입니다. **결과:** 200으로 이벤트 정책, 마스킹된 source, preview 설정과 상태를 반환합니다. **오류:** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**용도:** ID 외 필드를 일부 수정하고 `enabled`로 시작/중지합니다. **입력:** 부분 JSON이며 `event_policy`도 일부만 보낼 수 있고 null 임계값은 Collection 값을 상속합니다. **결과:** 200의 전체 Monitor; source/Collection/속도/정책 변경은 작업을 재시작합니다. **오류:** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**용도:** Monitor를 영구 삭제합니다. **입력:** 경로의 `monitor_id`입니다. **결과:** 디코더, 추론, RTSP 연결을 중지하고 메모리 이벤트를 버린 뒤 204를 반환하며 Collection은 유지합니다. **오류:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**용도:** 화면 없는 클라이언트에서 현재 상태를 폴링합니다. **입력:** Monitor ID입니다. **결과:** 200으로 연결, 실제 FPS, 지연, 건너뛴 프레임, 현재 인식/미등록 얼굴, preview, 재연결과 안전한 오류를 반환하며 embedding은 제외합니다. **오류:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**용도:** 휘발성 입장/퇴장/오류/복구 이벤트를 조회합니다. **입력:** `limit` 1~1000과 마지막 불투명 `cursor`입니다. **결과:** 200의 `events`, `next_cursor`, `truncated`, `stream_reset`; 재시작 시 이벤트가 사라집니다. **오류:** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**용도:** 기본 비활성인 원본 MJPEG preview를 엽니다. **입력:** ID와 일반 Bearer 헤더이며 API 키를 URL에 넣지 않습니다. **결과:** 시청자가 있을 때만 인코딩하는 장기 `multipart/x-mixed-replace`; 상자는 클라이언트가 `/state`로 그립니다. **오류:** 401, 404, 409 `preview_disabled`, 503.
+
+## 재시도
+
+GET은 재시도할 수 있습니다. DELETE 재시도 전 상태를 확인하세요. Person/Face 생성의 네트워크 결과가 불확실하면 POST 전 ID를 조회합니다. 429와 일시적 503만 제한된 exponential backoff와 jitter로 재시도하고 4xx는 요청을 수정하세요.
diff --git a/server/docs/api.md b/server/docs/api.md
new file mode 100644
index 000000000..0203398d0
--- /dev/null
+++ b/server/docs/api.md
@@ -0,0 +1,857 @@
+# InsightFace Server REST API usage guide
+
+**Languages:** English · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+This document explains how to call every public endpoint, what each parameter
+means, what work the server performs, and what a successful result looks like.
+Start with the [step-by-step user guide](user-guide.md) if the
+container and model are not running yet. The API is rooted at `/v1`, uses
+snake_case JSON fields, and accepts images as `multipart/form-data`. It is not
+an AWS Rekognition or CompreFace compatibility contract.
+
+## Common behavior
+
+The supplied Compose files disable authentication by default for isolated
+evaluation. When an operator enables authentication, every endpoint except
+`GET /v1/health` requires:
+
+```http
+Authorization: Bearer <api_key>
+```
+
+`GET /v1/health` is public so container orchestrators and the Web UI can check
+readiness and whether API authentication is enabled. Other `/v1` endpoints
+require authentication only when `auth_enabled` is true.
+
+Every response has an `x-request-id` UUID header. JSON response bodies repeat it
+as `request_id`. Successful DELETE requests return HTTP 204 with no body.
+
+Detection and quality scores use `0.0..1.0`:
+
+- `detection_score` is detector confidence;
+- `quality.score`, `sharpness`, `brightness`, and `pose` are documented local
+  quality signals, not AWS metrics;
+- `similarity` is the raw cosine value in `[-1.0, 1.0]`, not a probability;
+- recognition `threshold` accepts `[0.0, 1.0]` and defaults to `0.4`;
+- `threshold` is inclusive: `similarity >= threshold` matches.
+
+Bounding boxes include both forms:
+
+```json
+{
+  "pixels": {"x": 120, "y": 80, "width": 240, "height": 280},
+  "normalized": {"left": 0.12, "top": 0.08, "width": 0.24, "height": 0.28}
+}
+```
+
+JPEG, PNG, and WebP are accepted. EXIF orientation is applied before inference.
+The default compressed-image limit is 10 MiB, decoded limit is 40 million
+pixels, and whole-request limit is 64 MiB.
+
+## Errors
+
+Errors use standard HTTP status codes and one envelope:
+
+```json
+{
+  "error": {
+    "code": "face_not_found",
+    "message": "No usable face was detected.",
+    "details": {}
+  },
+  "request_id": "3ed21e89-4595-4eed-a699-1df42ca62032"
+}
+```
+
+Common status mappings include 400 invalid parameters, 401 missing/invalid API
+key, 404 missing resource, 409 resource/model conflict, 413 request or image too
+large, 422 invalid image or unusable face, 500 unexpected error, and 503 timeout
+or unavailable runtime. Phase one does not include a built-in rate limiter.
+
+OpenAPI is available at `/openapi.json`, with a same-origin interactive viewer at
+`/docs`.
+
+## First API workflow
+
+Set the server address once. Leave `AUTH_HEADER` empty only when health reports
+`auth_enabled: false`:
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+Then create a Collection, enroll one Person, and search with a different image:
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" \
+  -H 'Content-Type: application/json' \
+  -d '{"id":"employees","name":"Employees","threshold":0.4}'
+
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" \
+  -F 'id=alice' -F 'name=Alice' -F 'review_mode=off' \
+  -F 'images=@alice-enroll.jpg'
+
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" \
+  -F 'image=@alice-query.jpg' -F 'limit=5'
+```
+
+Do not send an empty `Authorization` header when authentication is disabled;
+omit the header completely. Shell examples below show the header so they work
+unchanged in authenticated deployments.
+
+## Client rules and retry safety
+
+- Use a client timeout longer than the configured server request timeout.
+- Treat `x-request-id` as the correlation ID and log it without logging images,
+  embeddings, RTSP credentials, or API keys.
+- Reuse an opaque `next_cursor` only with the same endpoint, Collection, Person,
+  and filter values. Never parse or construct it.
+- GET requests are safe to retry. Retry DELETE only after checking current
+  state. Do not automatically retry Person/FaceSample creation after an
+  ambiguous network failure; first query the resource ID supplied by the client.
+- `429` and transient `503` may be retried with bounded exponential backoff and
+  jitter. Validation `4xx` responses require changing the request.
+- Content type matters: Collection/Person patches use JSON; image operations and
+  enrollment use multipart; the saved face endpoint returns JPEG; MJPEG is a
+  streaming response.
+
+## System
+
+### `GET /v1/health`
+
+**Use:** Public readiness and container health check. **Parameters:** none.
+
+**Expected work and result:** It returns 200 with `status: "ready"` when startup has
+completed and SQLite `quick_check` succeeds, otherwise 503. The public
+`auth_enabled` boolean lets clients decide whether to present API-key controls;
+it does not expose the configured key or its hash.
+
+```bash
+curl -sS "${BASE_URL}/v1/health"
+```
+
+```json
+{"status":"ready","auth_enabled":false,"request_id":"..."}
+```
+
+**Common errors:** `503 not_ready` while models, database, or search indexes are
+not ready. Health deliberately does not require authentication.
+
+### `GET /v1/system`
+
+**Use:** Operator diagnostics. **Parameters:** none.
+
+**Expected work and result:** Returns safe diagnostics: server/OS/architecture/CPU, GPU and Compute Capability
+when present, NVIDIA driver, CUDA/cuDNN/ONNX Runtime, actual providers, model
+summary, database and path status, aggregate counts, API-key state, safe limits,
+and recent structured error summaries. It does not return API keys, images, or
+embeddings. `safe_config.detection` reports the immutable system detection
+profile and `safe_config.max_detected_faces` reports its safety cap. The API has
+no runtime mutation endpoint for the system profile.
+`safe_config.inference_max_concurrency` reports the resolved process-wide model
+budget (CPU default 4, CUDA default 8). Runtime diagnostics additionally expose
+active, waiting and peak model work. Detect, Compare, Embeddings, enrollment,
+Search query extraction and RTSP recognition share this budget.
+
+```bash
+curl -sS "${BASE_URL}/v1/system" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with runtime, model, database, path, aggregate count,
+search backend, safe configuration, authentication-state, and recent-error
+objects. Secrets, images, crops, and embeddings are excluded.
+
+**Common errors:** `401 unauthorized`; `503 request_timeout` if diagnostics
+cannot finish before the request deadline.
+
+Collection creation copies this system profile unless the request supplies
+overrides. Collection profiles are stored in SQLite, may be patched, and are
+used by Collection-bound enrollment and Search operations. Existing embeddings
+are not automatically re-extracted after a profile change.
+
+### `GET /v1/models`
+
+**Use:** Read the verified model bundle and actual provider. **Parameters:** none.
+
+**Expected work and result:** HTTP 200 returns `models`,
+`execution_provider`, and the verified license summary. It does not return model
+bytes or the private signing key.
+
+```bash
+curl -sS "${BASE_URL}/v1/models" -H "${AUTH_HEADER}"
+```
+
+**Common errors:** `401 unauthorized`.
+
+## Stateless face operations
+
+### `POST /v1/detect`
+
+**Use:** Detect all usable faces without writing data. Send
+`multipart/form-data` with:
+
+- `image` (required);
+- `max_faces` (optional, 1–100);
+- `collection_id` (optional; uses that Collection's detection profile instead
+  of the system profile).
+
+Faces are ordered by descending area. No face is a successful response with
+`faces: []`.
+
+```bash
+curl -sS http://localhost:18097/v1/detect \
+  -H "Authorization: Bearer ${INSIGHTFACE_API_KEY}" \
+  -F 'image=@group.jpg' \
+  -F 'max_faces=10' \
+  -F 'collection_id=employees'
+```
+
+**Success:** HTTP 200 with `faces`, `processing_ms`, and `request_id`. Each face
+contains pixel/normalized bounding boxes, five landmarks, detector confidence,
+and quality signals. Embeddings are not returned or persisted.
+
+**Common errors:** `400 request_detection_override_not_supported` for the
+deprecated `min_score`; `404` for an unknown `collection_id`; `413` for size
+limits; `422 invalid_image`; `503 request_timeout`.
+
+### `POST /v1/compare`
+
+**Use:** Compare one selected face from each of two images without persistence.
+Send `multipart/form-data` with:
+
+- `source` and `target` (required);
+- `threshold` (optional, `0.0..1.0`; server default `0.4`);
+- `collection_id` (optional detection-profile source).
+
+The active profile's single-face strategy selects one face in each image. A
+missing face in either image returns `422 face_not_found`.
+
+```bash
+curl -sS http://localhost:18097/v1/compare \
+  -H "Authorization: Bearer ${INSIGHTFACE_API_KEY}" \
+  -F 'source=@source.jpg' \
+  -F 'target=@target.jpg' \
+  -F 'threshold=0.4'
+```
+
+**Success:** HTTP 200 with `matched`, raw cosine `similarity`, the effective
+`threshold`, selected source/target face summaries, `processing_ms`, and
+`request_id`.
+
+**Common errors:** `404` for an unknown Collection; `413` for size limits;
+`422 invalid_image` or `face_not_found`; `503 request_timeout`.
+
+### `POST /v1/embeddings`
+
+**Use:** Extract the selected face embedding for a trusted integration. Send
+`multipart/form-data` with:
+
+- `image` (required);
+- `collection_id` (optional detection-profile source).
+
+Returns the profile-selected face with its embedding and model information. This
+authenticated endpoint is intentionally not used by normal registration/search
+flows. Embeddings are sensitive biometric templates and are not logged.
+
+```bash
+curl -sS "${BASE_URL}/v1/embeddings" -H "${AUTH_HEADER}" \
+  -F 'image=@portrait.jpg' -F 'collection_id=employees'
+```
+
+**Success:** HTTP 200 with one item in `faces`, its L2-normalized embedding,
+`model`, `processing_ms`, and `request_id`.
+
+**Common errors:** `400 request_detection_override_not_supported` for the
+deprecated `face_selection`; `404` unknown Collection; `413`; `422
+invalid_image` or `face_not_found`; `503 request_timeout`.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Use:** Create an isolated identity database and pin its model, detection, and
+search contract. Send `application/json`:
+
+```json
+{
+  "id": "employees",
+  "name": "Company Employees",
+  "description": "Employee face collection",
+  "threshold": 0.4,
+  "save_face_crops": false,
+  "detection": {
+    "input_sizes": [[96, 96], [512, 512]],
+    "threshold": 0.5,
+    "nms_threshold": 0.4,
+    "single_face_selection": "largest"
+  },
+  "search": {
+    "profile": "fp32_v1",
+    "capacity_rows": 100000,
+    "max_faces_per_person": 20,
+    "load_policy": "lazy"
+  },
+  "metadata": {"site": "shanghai"}
+}
+```
+
+`id` is `_default`, or 1–64 characters beginning with a letter or digit and
+otherwise containing letters, digits, `.`, `_`, and `-`. `name` is required. If
+omitted, `threshold` uses `INSIGHTFACE_DEFAULT_THRESHOLD`.
+
+`search.profile` accepts exactly `fp32_v1`, `fp16_v1`, `bf16_v1`,
+`int8_x736_v1`, or `int8_x1000_v1`; there is no implicit rerank profile. The
+recommended/default INT8 scale is 736, while the overall Collection default
+remains FP32. The other defaults are 100,000 rows, 20 FaceSamples per Person,
+and lazy loading. `_default` uses eager
+loading when no load policy is provided. The resolved values are persisted and
+returned as `search_profile`, `capacity_rows`, `max_faces_per_person`, and
+`load_policy` on Collection resources.
+
+The CPU native backend supports FP32, BF16, and INT8; FP16 is CUDA-only. The
+CUDA backend supports all five profiles. A persisted profile unsupported by the
+active backend fails explicitly and is never silently converted to another
+profile or execution provider. CUDA BF16 additionally requires an SM80-or-newer
+device. See the [User Guide](user-guide.md) for profile and capacity semantics
+and the complete support matrix.
+
+Creation binds model ID, version, bundle digest, embedding dimension, and
+preprocessing version. Those fields cannot be patched. Every Collection
+response also exposes a stable opaque `embedding_contract_id` derived from
+those pinned fields. Clients must copy, not construct, that ID when using
+trusted external enrollment.
+
+`detection` is optional and may partially override the system profile copied at
+creation. `single_face_selection` accepts `largest` and `center_largest`.
+`center_largest` maximizes one pixel-space score:
+`area - 2.0 * ((face_cx - image_cx)^2 + (face_cy - image_cy)^2)`.
+Detection confidence does not participate in this selection score.
+Collection responses return the resolved `detection` object and a monotonically
+increasing `detection_revision`.
+
+`save_face_crops` is optional and defaults to the deployment's
+`INSIGHTFACE_SAVE_FACE_CROPS` value, which is `false` unless explicitly changed.
+The resolved boolean is persisted on the Collection and does not follow later
+environment changes. When enabled, accepted 112×112 bounding-box crops—not original
+uploads—are encoded as JPEG and stored as BLOBs in SQLite. This can materially
+increase database and backup size.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" \
+  -H 'Content-Type: application/json' \
+  -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+**Success:** HTTP 201 with the resolved `collection`, including immutable model
+binding, detection profile, search settings, counts, and timestamps.
+
+**Common errors:** `400 invalid_detection_profile`,
+`unsupported_search_profile`, or `search_capacity_too_large`; `409
+collection_exists`; `503 search_index_unavailable`.
+
+### `GET /v1/collections`
+
+**Use:** List Collections. **Query parameters:** `limit` integer 1–100, default
+50; `cursor` optional opaque continuation token.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections?limit=50" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `collections` and nullable `next_cursor`. Pass the
+returned cursor unchanged to the same endpoint. **Common errors:** `400
+invalid_cursor`; `401 unauthorized`.
+
+### `GET /v1/collections/{collection_id}`
+
+**Use:** Read one Collection. **Path parameter:** `collection_id`, the stable
+Collection ID.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `collection`, current `person_count`, `face_count`,
+and `embedding_contract_id`. **Common errors:** `404 resource_not_found`; `409
+collection_model_mismatch` when model-bound use is attempted under an
+incompatible active bundle.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Use:** Update mutable Collection policy. **Path parameter:** `collection_id`.
+The `application/json` body may update `name`, `description`, `threshold`, `metadata`, and
+`save_face_crops`. Changing crop storage affects later registration requests:
+existing crops are neither backfilled nor deleted, and an in-flight registration
+may finish using the value it already read. A nested
+`search` object may update `capacity_rows`, `max_faces_per_person`, and
+`load_policy`; incompatible reductions are rejected. `search_profile` is
+immutable through this endpoint because changing it requires an index rebuild.
+The nested `detection` object may update any detection-profile field. An
+in-flight request keeps its original immutable snapshot; later requests see the
+new revision. No existing FaceSample is reprocessed.
+Unknown fields and explicit nulls are rejected.
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/collections/employees" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"threshold":0.45,"detection":{"single_face_selection":"center_largest"}}'
+```
+
+**Success:** HTTP 200 with the complete updated `collection`; later requests see
+the new policy. **Common errors:** `400` for null/unknown/invalid values; `404`;
+`409` for incompatible capacity reductions or model contract; `503` if the
+index cannot apply the mutation.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Use:** Delete a Collection. **Path parameter:** `collection_id`. **Query
+parameter:** `force` boolean, default `false`.
+
+An empty Collection is deleted. A non-empty Collection returns
+`409 collection_not_empty`; repeat with `force=true` only when deleting all
+People and FaceSamples is intended.
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees?force=true" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 204 with no body. **Common errors:** `404`; `409
+collection_not_empty`; `503` if the index cannot complete deletion.
+
+Registration that would exceed `capacity_rows` returns
+`409 collection_capacity_exceeded` without committing the extra FaceSample.
+Registration that would exceed `max_faces_per_person` returns
+`409 person_face_limit_exceeded`.
+
+## People and FaceSamples
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Use:** Create a Person and enroll one or more FaceSamples in one request.
+**Path parameter:** `collection_id`. Send `multipart/form-data` fields:
+
+- `images` (required and repeatable; default maximum 20);
+- `id` (optional; a UUID is generated when omitted);
+- `name`, `external_id` (optional);
+- `metadata` (optional JSON object encoded as a multipart string, default `{}`);
+- `review_mode` (`off`, `standard`, or `strict`; default `off`).
+- `embedding_mode` (`server` or `external_trusted`; default `server`);
+- `external_embeddings` (required only for `external_trusted`; JSON array with
+  exactly one feature vector per `images` part);
+- `embedding_contract_id` (required only for `external_trusted`; copy the
+  current Collection value exactly).
+
+```bash
+curl -sS http://localhost:18097/v1/collections/employees/persons \
+  -H "Authorization: Bearer ${INSIGHTFACE_API_KEY}" \
+  -F 'id=employee-001' \
+  -F 'name=Alice' \
+  -F 'external_id=HR-1001' \
+  -F 'metadata={"department":"sales"}' \
+  -F 'review_mode=standard' \
+  -F 'images=@alice1.jpg' \
+  -F 'images=@alice2.jpg'
+```
+
+Every mode requires a valid image containing at least one detected face and a
+finite, correctly sized, L2-normalized embedding. `off` uses the Collection's
+single-face strategy and skips configurable quality thresholds. `standard` and `strict`
+require exactly one face; `standard` additionally applies the configured
+minimum face size, detector score, quality, and pose rules. `strict` applies
+`standard`, then requires the candidate's maximum similarity to its Person's
+existing samples to be strictly greater than its maximum similarity to every
+other Person. Similarity is evaluated with the Collection's pinned search
+profile; a tie is rejected.
+
+When a Person has no existing FaceSample, the first standard-quality candidate
+bootstraps that Person and skips the similarity comparison. Later candidates in
+the same multipart request use earlier accepted candidates as class-in
+references. A batch can partially succeed:
+
+With `embedding_mode=server`, the service aligns each accepted face, extracts
+its recognition feature, and L2-normalizes it. With
+`embedding_mode=external_trusted`, the image is still decoded, detected, and
+subjected to the same `review_mode` rules, but the recognition model is not
+run. In `off`, the trusted caller asserts that vector `i` belongs to the largest
+face in image part `i`. There is no automatic fallback to server extraction;
+image and vector counts must match. Invalid or rejected images do not cause
+their supplied vectors to be enrolled.
+
+The external vector must contain finite numeric values, be nonzero, match the
+Collection dimension and `embedding_contract_id`, and have L2 norm within
+`1.0 ± 0.0002`. A vector outside that tolerance is rejected for that image as
+`invalid_external_embedding`; it is not silently repaired or replaced. A
+passing vector is normalized once more after FP32 conversion to remove small
+floating-point drift before SQLite commit.
+`strict` review performs class-in/class-out comparisons using that final
+external vector. A trusted caller is solely responsible for ensuring that the
+vector was extracted from the paired image with the declared pipeline; the
+server deliberately does not re-extract a feature to prove the association.
+
+```json
+{
+  "person": {"id": "employee-001", "face_count": 1},
+  "faces": [{"id": "a-face-uuid", "quality": {"score": 0.91}}],
+  "rejected_images": [
+    {"index": 1, "filename": "alice2.jpg", "reason": "multiple_faces"}
+  ],
+  "request_id": "a-uuid"
+}
+```
+
+Current rejection reasons include `invalid_image`, `image_too_large`,
+`face_not_found`, `multiple_faces`, `face_too_small`, `low_detection_score`,
+`low_quality`, `extreme_pose`, `invalid_embedding`, and
+`identity_similarity_conflict`. A strict similarity rejection also reports
+`same_person_similarity`, `other_person_similarity`, `other_person_id`, and
+`matched_face_id`. If no image is accepted, the request returns
+`422 registration_failed` and no Person is created.
+
+**Success:** HTTP 201 with `person`, accepted `faces`, `rejected_images`, and
+`request_id`. Partial success is still HTTP 201.
+
+**Common errors:** `400` invalid ID/metadata or too many images; `404`
+Collection; `409` Person/external-ID, embedding-contract, capacity, or
+per-Person limit conflict; `413`; `422 registration_failed`; `503
+search_index_unavailable`. If a 503 contains `write_committed: true`, do not
+blindly retry—read the Person first.
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Use:** List or filter People. **Path parameter:** `collection_id`. **Query
+parameters:** `limit` 1–100, default 50; opaque `cursor`; optional `search` up to
+200 characters matching Person ID, name, or external ID.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons?limit=50&search=alice" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `persons` and nullable `next_cursor`. **Common
+errors:** `400 invalid_cursor`; `404` Collection.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Use:** Read one Person. **Path parameters:** `collection_id` and `person_id`.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/alice" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `person`, including current `face_count` and
+timestamps. **Common errors:** `404` Collection or Person.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Use:** Update Person display data. **Path parameters:** `collection_id` and
+`person_id`. The JSON body accepts `name`, `external_id`, and object `metadata`;
+unknown fields are rejected and `metadata` cannot be null.
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/collections/employees/persons/alice" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"name":"Alice Chen","metadata":{"department":"sales"}}'
+```
+
+**Success:** HTTP 200 with the complete updated `person`. **Common errors:**
+`400` invalid body; `404`; `409 external_id_exists`.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Use:** Delete one Person, all FaceSamples, embeddings, and optional crops.
+**Path parameters:** `collection_id` and `person_id`.
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees/persons/alice" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 204 with no body; later searches cannot return the deleted
+Person. **Common errors:** `404`; `503 search_index_unavailable`.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Use:** Add FaceSamples to an existing Person. **Path parameters:**
+`collection_id` and `person_id`. Add repeatable multipart `images`.
+`review_mode` has the
+same values, default, registration rules, and partial-success response described
+above. `embedding_mode`, `external_embeddings`, and `embedding_contract_id`
+also have exactly the same semantics as Person creation.
+
+See the [User Guide](user-guide.md) for the complete
+request example and trust-boundary guidance.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/alice/faces" \
+  -H "${AUTH_HEADER}" -F 'review_mode=standard' \
+  -F 'images=@alice-2.jpg' -F 'images=@alice-3.webp'
+```
+
+**Success:** HTTP 201 with accepted `faces` and `rejected_images`; partial
+success is allowed. **Common errors:** the same registration, capacity,
+contract, size, quality, and index errors as Person creation, plus `404` Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Use:** Page through FaceSample metadata. **Path parameters:** `collection_id`
+and `person_id`. **Query parameters:** `limit` 1–100, default 50, and opaque
+`cursor`. Stored embeddings
+and crop bytes are not returned. Each item has `has_crop: true` only when a
+stored crop is available.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/alice/faces?limit=50" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `faces` and nullable `next_cursor`. **Common errors:**
+`400 invalid_cursor`; `404` Collection or Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Use:** Download an optional saved crop for administration. **Path parameters:**
+`collection_id`, `person_id`, and `face_id`. Returns the stored 112×112
+bounding-box crop as `image/jpeg`. The endpoint requires the same
+Bearer authentication as other non-health API calls and returns
+`Cache-Control: no-store`. Clients should not treat it as an original image. If
+the FaceSample exists but has no stored crop, the server returns a not-found
+error rather than synthesizing or reconstructing an image.
+
+```bash
+curl -sS http://localhost:18097/v1/collections/employees/persons/employee-001/faces/face-uuid/image \
+  -H "Authorization: Bearer ${INSIGHTFACE_API_KEY}" \
+  -o face-crop.jpg
+```
+
+**Success:** HTTP 200 JPEG with `Cache-Control: no-store`; this response has no
+JSON `request_id`, so use the `x-request-id` header. **Common errors:** `404`
+FaceSample or `face_image_not_found`; `401 unauthorized`.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Use:** Delete one FaceSample, embedding, and optional crop. **Path parameters:**
+`collection_id`, `person_id`, and `face_id`.
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees/persons/alice/faces/face-uuid" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 204 with no body; the row is removed from the active index
+before success. **Common errors:** `404`; `503 search_index_unavailable`.
+
+Successful face and Person deletions synchronously update the active search
+generation. A later search in the same process cannot return the deleted row.
+
+## Search
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Use:** Search a Collection using the selected face in one query image.
+**Path parameter:** `collection_id`. Multipart fields:
+
+- `image` (required);
+- `limit` (optional, 1–100, default 5);
+- `threshold` (optional `0.0..1.0`; defaults to the Collection threshold).
+
+The Collection profile selects the input face, which is searched against every FaceSample. Each
+Person receives its highest FaceSample score. Only People at or above threshold
+are returned, in descending score order. No match is `matches: []`; no usable
+query face is `422 face_not_found`.
+
+```bash
+curl -sS http://localhost:18097/v1/collections/employees/search \
+  -H "Authorization: Bearer ${INSIGHTFACE_API_KEY}" \
+  -F 'image=@unknown.jpg' \
+  -F 'limit=5'
+```
+
+Example match:
+
+```json
+{
+  "person": {
+    "id": "employee-001",
+    "name": "Alice",
+    "external_id": "HR-1001",
+    "metadata": {"department": "sales"}
+  },
+  "similarity": 0.8642,
+  "matched_face_id": "a-face-uuid"
+}
+```
+
+**Success:** HTTP 200 with `searched_face`, ordered `matches`, effective
+`threshold`, `processing_ms`, and `request_id`. No match is `matches: []`.
+
+**Common errors:** `404` Collection; `409 collection_model_mismatch`; `413`;
+`422 invalid_image` or `face_not_found`; `503 search_index_unavailable` or
+`request_timeout`.
+
+## RTSP Monitors
+
+A Monitor is a persistent server-side RTSP recognition task. Its configuration is
+stored in SQLite and an enabled Monitor resumes when the Server restarts. Video
+frames are never saved. Recent events live only in a bounded memory buffer and
+are lost on restart. The decoder retains only the newest frame, so slow inference
+reduces the effective rate instead of building a delayed frame queue.
+
+### `POST /v1/monitors`
+
+**Use:** Create and optionally start a persistent Monitor. **Body:** Send
+`application/json` with the configuration below. `source.url` accepts only
+`rtsp://` or `rtsps://`; credentials are AES-GCM encrypted under `/data` and the
+API only returns a redacted source.
+
+```json
+{
+  "id": "front-gate",
+  "name": "Front gate",
+  "description": "Main entrance",
+  "enabled": true,
+  "source": {"type": "rtsp", "url": "rtsp://viewer:secret@camera.example/live"},
+  "collection_id": "employees",
+  "inference_fps": 2.0,
+  "match_threshold": null,
+  "event_buffer_size": 1000,
+  "event_policy": {
+    "confirm_frames": 3,
+    "absence_timeout_seconds": 3.0,
+    "cooldown_seconds": 10.0,
+    "emit_unknown": true
+  },
+  "preview_enabled": false
+}
+```
+
+`match_threshold: null` inherits the Collection threshold.
+`event_buffer_size` is 10–10000. Web preview is deliberately off by default;
+recognition and event collection do not require a viewer.
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors" -H "${AUTH_HEADER}" \
+  -H 'Content-Type: application/json' -d @monitor.json
+```
+
+**Success:** HTTP 201 with `monitor`, its redacted source, effective defaults, and
+runtime summary. **Common errors:** `400 invalid_request`; `404` Collection;
+`409 monitor_exists`; `429 monitor_limit_exceeded`.
+
+### `GET /v1/monitors`
+
+**Use:** List persistent Monitor configurations and compact runtime summaries.
+**Query:** `limit` is 1–100 (default 50); `cursor` is an opaque value returned as
+`next_cursor` and must not be parsed or modified by clients.
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors?limit=50" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with ordered `monitors` and nullable `next_cursor`.
+**Common errors:** `400 invalid_cursor` for an invalid, altered, or wrong-scope
+cursor; `401 unauthorized` when authentication is enabled.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Use:** Read one persisted Monitor configuration and its latest runtime summary.
+**Path:** `monitor_id` is the caller-selected ID used during creation. The
+returned RTSP URL omits user information and query values.
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors/front-gate" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `monitor`, including `event_policy`,
+`preview_enabled`, timestamps, and `runtime`. **Common errors:** `404
+monitor_not_found`; `401 unauthorized`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Use:** Partially update a Monitor; its `id` cannot change. **Body:** Supply one
+or more mutable fields from create. `event_policy` itself is partial. Send a new
+`source` only when rotating the RTSP URL or credentials. Set
+`match_threshold` to `null` to return to the Collection default.
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/monitors/front-gate" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"inference_fps":1.5,"event_policy":{"confirm_frames":5}}'
+```
+
+Changing source, Collection, cadence, threshold, or event policy restarts that
+Monitor task. Set `enabled` to `false` or `true` to stop or start it. Name,
+description, preview, and buffer-size changes apply without a task restart.
+
+**Success:** HTTP 200 with the complete updated `monitor`. **Common errors:**
+`400 invalid_request`; `404` Monitor or Collection; `429
+monitor_limit_exceeded`.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Use:** Permanently remove one Monitor configuration. **Path:** `monitor_id`.
+This stops its decoder and inference threads, releases the RTSP connection, and
+discards its in-memory state and events. It does not delete its Collection.
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/monitors/front-gate" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 204 with no body. **Common errors:** `404 monitor_not_found`;
+`401 unauthorized`.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Use:** Poll the live state needed by headless clients or the Web UI.
+**Result fields:** `status`, `connected`, source dimensions/FPS, configured and
+actual inference rate, processing time, skipped frames, current recognized and
+unknown faces, preview viewers, reconnect count, and last safe error. Embeddings
+and source credentials are never included.
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors/front-gate/state" -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `state`; a disabled Monitor normally reports
+`stopped`. **Common errors:** `404 monitor_not_found`; `401 unauthorized`.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Use:** Pull recent enter, exit, error, and recovery events without maintaining
+a long-lived connection. **Query:** `limit` is 1–1000 (default 100); pass the
+previous `next_cursor` on the next poll. The cursor is an opaque signed string
+containing an internal stream epoch and sequence.
+
+The first call without a cursor returns the newest events up to `limit`.
+Subsequent calls return later events. `truncated: true` means the client fell
+behind the bounded ring; `stream_reset: true` means the task restarted and the
+old cursor belongs to another epoch. Events are not durable and are lost when
+the process restarts.
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors/front-gate/events?limit=100" \
+  -H "${AUTH_HEADER}"
+```
+
+**Success:** HTTP 200 with `events`, `next_cursor`, `has_more`, `truncated`, and
+`stream_reset`. **Common errors:** `400 invalid_cursor`; `404
+monitor_not_found`; `401 unauthorized`.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Use:** Open an optional raw MJPEG preview. **Authentication:** Send the same
+Bearer header as every other API call; never put the API key in a URL. The
+endpoint returns unannotated `multipart/x-mixed-replace` JPEG frames. A client
+uses `/state` to draw boxes and labels.
+
+JPEG encoding runs lazily only while `preview_enabled` is true and at least one
+viewer is connected. Closing the preview does not stop recognition. Clients
+should reconnect with bounded backoff after transport loss.
+
+**Success:** HTTP 200 as a long-lived binary stream, not JSON. **Common errors:**
+`409 preview_disabled`; `503 stream_unavailable`; `404 monitor_not_found`; `401
+unauthorized`.
diff --git a/server/docs/api.pt.md b/server/docs/api.pt.md
new file mode 100644
index 000000000..637c76000
--- /dev/null
+++ b/server/docs/api.pt.md
@@ -0,0 +1,170 @@
+# Guia de utilização da API REST do InsightFace Server
+
+**Idiomas:** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · Português · [한국어](api.ko.md)
+
+Este guia descreve o objetivo, entrada, trabalho do servidor, resultado e erros de cada API pública. Para instalação e primeira pesquisa consulte o [guia do utilizador](user-guide.pt.md); o esquema exato da instância está em `/docs` e `/openapi.json`.
+
+## Regras comuns
+
+- Base `/v1`, JSON em `snake_case`, imagens JPEG/PNG/WebP em multipart.
+- O Compose fornecido desativa autenticação para avaliação isolada. Quando ativa, tudo exceto health requer `Authorization: Bearer <api_key>`; quando inativa omita completamente o cabeçalho.
+- Cada resposta tem `x-request-id`; JSON repete-o como `request_id`.
+- confidence/quality/threshold usam `0..1`. Similarity não é probabilidade: é cosine bruto `[-1,1]`; predefinição `0.4`, match quando `similarity >= threshold`.
+- Cursor é opaco e só deve ser devolvido sem alterações à mesma rota, Collection, Person e filtro.
+- Estados comuns: 400 entrada, 401 auth, 404 ausente, 409 conflito, 413 tamanho, 422 imagem/rosto, 429 limite, 503 timeout/modelo/índice.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## Sistema
+
+### `GET /v1/health`
+
+**Uso/entrada:** readiness pública, sem parâmetros nem auth. **Resultado:** verifica arranque e SQLite quick_check; 200 com `status`, `auth_enabled`, `request_id`. **Erro:** `503 not_ready`.
+
+### `GET /v1/system`
+
+**Uso/entrada:** diagnóstico seguro, sem parâmetros. **Resultado:** 200 com OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, modelo, DB, mounts, contagens, pesquisa, configuração segura e concorrência; nunca segredos, imagens ou embeddings. **Erros:** 401, 503.
+
+### `GET /v1/models`
+
+**Uso/entrada:** detector/recognizer verificados, Provider e licença; sem parâmetros. **Resultado:** 200 `models`, `execution_provider`, `license`. **Erro:** 401.
+
+## Operações faciais sem estado
+
+### `POST /v1/detect`
+
+**Entrada:** multipart `image` obrigatório, `max_faces` 1–100, `collection_id` opcional. **Processo/resultado:** combina resoluções, NMS global, ordena por área; 200 `faces` com caixas/5 pontos/score/qualidade e `processing_ms`. Sem rosto é lista vazia válida. **Erros:** 400 min_score antigo, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**Entrada:** multipart `source`, `target`, `threshold` opcional 0–1 e `collection_id`. **Resultado:** escolhe um rosto por imagem; 200 `matched`, cosine `similarity`, threshold efetivo, ambos os rostos e tempo. **Erros:** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**Entrada:** multipart `image`, `collection_id` opcional. **Resultado:** 200 com rosto escolhido, embedding L2, modelo e tempo. Desnecessário para registo normal; vetor não é registado em logs. **Erros:** 400 face_selection antigo, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Entrada:** JSON `id`, `name`; opcionais description, threshold (0.4), metadata, save_face_crops, `detection`, `search` com profile/capacity/max_faces_per_person/load_policy. **Processo/resultado:** fixa modelo, pré-processamento e contrato de pesquisa; 201 com `collection` resolvida. **Erros:** 400 configuração, 409 exists, 503 índice.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**Entrada:** query `limit` 1–100 (50), cursor opcional. **Resultado:** 200 `collections`, `next_cursor` nullable. **Erros:** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**Entrada:** ID da Collection no path. **Resultado:** 200 `collection`, contagens Person/Face e `embedding_contract_id`. **Erro:** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Entrada:** ID; JSON name/description/threshold/metadata/save_face_crops, capacidade/max/load de pesquisa e detection. Null, campos desconhecidos, modelo e search profile não podem mudar. **Resultado:** 200 Collection completa; detection vale no próximo pedido. **Erros:** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Entrada:** ID; query `force=false`, true se não vazia. **Resultado:** 204 sem corpo. **Erros:** 404, 409 collection_not_empty, 503.
+
+## Pessoas e FaceSamples
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Entrada:** Collection; multipart `images` repetível, id/name/external_id opcionais, metadata como JSON texto, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; modo externo acrescenta vetores e contract ID. **Processo/resultado:** revê cada imagem; 201 `person`, `faces` aceites e `rejected_images`, sucesso parcial; tudo rejeitado dá 422 sem Person. **Erros:** 400, 404, 409 ID/contrato/capacidade, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Entrada:** Collection; query limit/cursor/`search` por ID, nome ou external ID. **Resultado:** 200 `persons`, `next_cursor`. **Erros:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs Collection e Person. **Resultado:** 200 `person` com face_count. **Erro:** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs; JSON name/external_id/metadata objeto. **Resultado:** 200 Person atualizada. **Erros:** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Entrada:** IDs. **Resultado:** elimina Person, FaceSamples, embeddings e crops, sincroniza índice, 204. **Erros:** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrada:** IDs; images repetíveis e os mesmos campos review/embedding da criação. **Resultado:** 201 `faces`, `rejected_images`, sucesso parcial. **Erros:** registo mais 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Entrada:** IDs; query limit 1–100 e cursor. **Resultado:** 200 metadata `faces`, `has_crop`, `next_cursor`, sem embedding nem bytes. **Erros:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Entrada:** três IDs. **Resultado:** se guardado, 200 `image/jpeg`, crop 112×112, `Cache-Control:no-store`; request ID apenas no header. **Erros:** 401, 404 face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Entrada:** três IDs. **Resultado:** elimina embedding/crop/linha do índice, 204. **Erros:** 404, 503.
+
+## Pesquisa
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Entrada:** Collection; multipart `image`, `limit` 1–100 (5), threshold opcional ou valor da Collection. **Processo/resultado:** compara o rosto escolhido com todas as samples e usa o máximo por Person; 200 `searched_face`, `matches` ordenados, threshold e tempo. Sem match é lista vazia. **Erros:** 404, 409 modelo, 413, 422 imagem/rosto, 503 índice/timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## Monitores RTSP
+
+A configuração do Monitor persiste no SQLite e uma tarefa ativada é restaurada após reiniciar o servidor. Os fotogramas não são guardados; os eventos vivem apenas num buffer circular limitado em memória.
+
+### `POST /v1/monitors`
+
+**Uso:** Criar um Monitor RTSP persistente. **Entrada:** JSON com ID, nome, `source`, Collection, `inference_fps` (2), limiar opcional, buffer/política de eventos e `preview_enabled` (false). **Resultado:** 201 com `monitor` ocultado; credenciais são encriptadas em repouso. **Erros:** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**Uso:** Listar Monitores com paginação. **Entrada:** `limit` 1–100 (50) e o `cursor` opaco da resposta anterior, sem alterações. **Resultado:** 200 com `monitors` e `next_cursor`, nunca as credenciais. **Erros:** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Uso:** Ler configuração e resumo de execução de um Monitor. **Entrada:** `monitor_id` no caminho. **Resultado:** 200 com política de eventos, origem ocultada, preview e estado. **Erros:** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Uso:** Atualizar parcialmente exceto o ID e iniciar/parar com `enabled`. **Entrada:** JSON parcial; `event_policy` também é parcial e limiar null herda a Collection. **Resultado:** 200 com Monitor completo; origem/Collection/frequência/política reiniciam a tarefa. **Erros:** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Uso:** Remover permanentemente um Monitor. **Entrada:** `monitor_id` no caminho. **Resultado:** para descodificação, inferência e RTSP, descarta eventos de memória e devolve 204; não remove a Collection. **Erros:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Uso:** Consultar o estado atual em clientes sem interface. **Entrada:** ID do Monitor. **Resultado:** 200 com ligação, FPS efetivo, latência, saltos, rostos reconhecidos/desconhecidos, preview, reconexões e erro seguro, sem embeddings. **Erros:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Uso:** Obter eventos voláteis de entrada/saída/erro/recuperação. **Entrada:** `limit` 1–1000 e último `cursor` opaco. **Resultado:** 200 com `events`, `next_cursor`, `truncated` e `stream_reset`; reiniciar perde eventos. **Erros:** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Uso:** Abrir o preview MJPEG bruto, desativado por predefinição. **Entrada:** ID e cabeçalho Bearer normal; nunca a chave no URL. **Resultado:** `multipart/x-mixed-replace` longo, codificado só com observadores; o cliente desenha caixas via `/state`. **Erros:** 401, 404, 409 `preview_disabled`, 503.
+
+## Repetição de pedidos
+
+GET pode ser repetido. Verifique o estado antes de repetir DELETE. Se o resultado de criar Person/Face for incerto pela rede, consulte o ID antes de novo POST. Repita apenas 429 e 503 transitórios com backoff exponencial limitado e jitter; corrija os 4xx.
diff --git a/server/docs/api.ru.md b/server/docs/api.ru.md
new file mode 100644
index 000000000..7515def1b
--- /dev/null
+++ b/server/docs/api.ru.md
@@ -0,0 +1,170 @@
+# Руководство по REST API InsightFace Server
+
+**Языки:** [English](api.md) · [中文](api.zh-CN.md) · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · Русский · [Português](api.pt.md) · [한국어](api.ko.md)
+
+Здесь описаны назначение, входные данные, работа сервера, успешный результат и ошибки каждого публичного API. Установка и первый поиск приведены в [руководстве пользователя](user-guide.ru.md), точная схема запущенной версии — в `/docs` и `/openapi.json`.
+
+## Общие правила
+
+- Базовый путь `/v1`, JSON в `snake_case`, изображения JPEG/PNG/WebP как multipart.
+- Поставляемый Compose отключает авторизацию для изолированной оценки. Если она включена, всё кроме health требует `Authorization: Bearer <api_key>`; если выключена, заголовок нужно полностью убрать.
+- В каждом ответе есть `x-request-id`, а JSON повторяет его как `request_id`.
+- confidence/quality/threshold находятся в `0..1`. Similarity — не вероятность, а исходный cosine `[-1,1]`; порог по умолчанию `0.4`, совпадение при `similarity >= threshold`.
+- Cursor непрозрачен и возвращается без изменений только в тот же путь, Collection, Person и фильтр.
+- Частые статусы: 400 ввод, 401 auth, 404 нет ресурса, 409 конфликт, 413 размер, 422 изображение/лицо, 429 лимит, 503 timeout/модель/индекс.
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+curl -fsS "${BASE_URL}/v1/health"
+```
+
+## Система
+
+### `GET /v1/health`
+
+**Назначение/ввод:** публичная readiness-проверка, без параметров и auth. **Результат:** проверяет запуск и SQLite quick_check; 200 с `status`, `auth_enabled`, `request_id`. **Ошибка:** `503 not_ready`.
+
+### `GET /v1/system`
+
+**Назначение/ввод:** безопасная диагностика, без параметров. **Результат:** 200 с OS/CPU/GPU, Driver, CUDA/cuDNN/ORT, Provider, моделью, DB, mount, счётчиками, поиском, безопасной конфигурацией и параллелизмом; без секретов, изображений и embeddings. **Ошибки:** 401, 503.
+
+### `GET /v1/models`
+
+**Назначение/ввод:** проверенные detector/recognizer, Provider и лицензия; без параметров. **Результат:** 200 `models`, `execution_provider`, `license`. **Ошибка:** 401.
+
+## Stateless-операции
+
+### `POST /v1/detect`
+
+**Ввод:** multipart `image` обязателен, `max_faces` 1–100, необязательный `collection_id`. **Работа/результат:** объединяет разрешения, делает общую NMS и сортирует по площади; 200 `faces` с рамками/5 точками/score/quality и `processing_ms`. Нет лица — корректный пустой список. **Ошибки:** 400 старый min_score, 404 Collection, 413, 422 invalid_image, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" -F 'image=@group.jpg' -F 'max_faces=10'
+```
+
+### `POST /v1/compare`
+
+**Ввод:** multipart `source`, `target`, необязательные `threshold` 0–1 и `collection_id`. **Результат:** выбирает одно лицо из каждого изображения; 200 `matched`, cosine `similarity`, фактический threshold, оба face и время. **Ошибки:** 404, 413, 422 invalid_image/face_not_found, 503.
+
+### `POST /v1/embeddings`
+
+**Ввод:** multipart `image`, необязательный `collection_id`. **Результат:** 200 с выбранным face, L2-нормированным embedding, моделью и временем. Для обычной регистрации не нужен; в лог не записывается. **Ошибки:** 400 старый face_selection, 404, 413, 422, 503.
+
+## Collections
+
+### `POST /v1/collections`
+
+**Ввод:** JSON `id`, `name`; необязательные description, threshold (0.4), metadata, save_face_crops, `detection`, `search` с profile/capacity/max_faces_per_person/load_policy. **Работа/результат:** фиксирует модель, предобработку и поисковый контракт; 201 с полной `collection`. **Ошибки:** 400 конфигурация, 409 exists, 503 индекс.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" -H 'Content-Type: application/json' -d '{"id":"employees","name":"Employees","threshold":0.4}'
+```
+
+### `GET /v1/collections`
+
+**Ввод:** query `limit` 1–100 (50), необязательный cursor. **Результат:** 200 `collections`, nullable `next_cursor`. **Ошибки:** 400 invalid_cursor, 401.
+
+### `GET /v1/collections/{collection_id}`
+
+**Ввод:** ID Collection в пути. **Результат:** 200 `collection`, количества Person/Face и `embedding_contract_id`. **Ошибка:** 404.
+
+### `PATCH /v1/collections/{collection_id}`
+
+**Ввод:** ID; JSON name/description/threshold/metadata/save_face_crops, capacity/max/load поиска и detection. Null, неизвестные поля, модель и search profile менять нельзя. **Результат:** 200 полная Collection; detection действует со следующего запроса. **Ошибки:** 400, 404, 409, 503.
+
+### `DELETE /v1/collections/{collection_id}`
+
+**Ввод:** ID; query `force=false`, true для непустой Collection. **Результат:** 204 без тела. **Ошибки:** 404, 409 collection_not_empty, 503.
+
+## Person и FaceSample
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**Ввод:** Collection; multipart повторяемые `images`, необязательные id/name/external_id, metadata как JSON-строка, `review_mode=off|standard|strict`, `embedding_mode=server|external_trusted`; внешнему режиму нужны векторы и contract ID. **Работа/результат:** проверяет каждое изображение; 201 `person`, принятые `faces`, `rejected_images`; частичный успех разрешён, все отклонены — 422 без Person. **Ошибки:** 400, 404, 409 ID/контракт/ёмкость, 413, 422, 503.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" -F 'id=alice' -F 'review_mode=off' -F 'images=@alice.jpg'
+```
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**Ввод:** Collection; query limit/cursor/`search` по ID, имени или external ID. **Результат:** 200 `persons`, `next_cursor`. **Ошибки:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**Ввод:** ID Collection и Person. **Результат:** 200 `person` с face_count. **Ошибка:** 404.
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**Ввод:** IDs; JSON name/external_id/metadata-объект. **Результат:** 200 обновлённый Person. **Ошибки:** 400, 404, 409 external_id_exists.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**Ввод:** IDs. **Результат:** удаляет Person, FaceSamples, embeddings и crops, синхронизирует индекс, 204. **Ошибки:** 404, 503.
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Ввод:** IDs; повторяемые images и те же review/embedding-поля, что при создании Person. **Результат:** 201 `faces`, `rejected_images`, возможен частичный успех. **Ошибки:** ошибки регистрации плюс 404 Person.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**Ввод:** IDs; query limit 1–100 и cursor. **Результат:** 200 metadata `faces`, `has_crop`, `next_cursor`, без embedding и bytes. **Ошибки:** 400 cursor, 404.
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**Ввод:** три ID. **Результат:** если сохранено, 200 `image/jpeg`, crop 112×112, `Cache-Control:no-store`; request ID только в header. **Ошибки:** 401, 404 face/face_image_not_found.
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**Ввод:** три ID. **Результат:** удаляет embedding/crop/строку индекса, 204. **Ошибки:** 404, 503.
+
+## Поиск
+
+### `POST /v1/collections/{collection_id}/search`
+
+**Ввод:** Collection; multipart `image`, `limit` 1–100 (5), необязательный threshold или значение Collection. **Работа/результат:** сравнивает выбранное лицо со всеми samples, берёт максимум по Person; 200 `searched_face`, отсортированные `matches`, threshold и время. Нет совпадений — пустой список. **Ошибки:** 404, 409 модель, 413, 422 изображение/лицо, 503 индекс/timeout.
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" -F 'image=@query.jpg' -F 'limit=5'
+```
+
+## RTSP-мониторы
+
+Конфигурация Monitor сохраняется в SQLite, а включённая задача восстанавливается после перезапуска сервера. Видеокадры не сохраняются; события находятся только в ограниченном кольцевом буфере памяти.
+
+### `POST /v1/monitors`
+
+**Назначение:** Создать постоянный RTSP Monitor. **Ввод:** JSON с ID, именем, `source`, Collection, `inference_fps` (2), необязательным порогом, буфером/политикой событий и `preview_enabled` (false). **Результат:** 201 с очищенным `monitor`; реквизиты шифруются при хранении. **Ошибки:** 400, 404, 409, 429.
+
+### `GET /v1/monitors`
+
+**Назначение:** Постранично перечислить Monitors. **Ввод:** `limit` 1–100 (50) и неизменённый непрозрачный `cursor` прошлого ответа. **Результат:** 200 с `monitors` и `next_cursor`, без реквизитов доступа. **Ошибки:** 400 `invalid_cursor`, 401.
+
+### `GET /v1/monitors/{monitor_id}`
+
+**Назначение:** Прочитать конфигурацию и сводку выполнения Monitor. **Ввод:** `monitor_id` в пути. **Результат:** 200 с политикой событий, очищенным источником, preview и состоянием. **Ошибки:** 401, 404 `monitor_not_found`.
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**Назначение:** Частично изменить поля кроме ID и включать/выключать через `enabled`. **Ввод:** частичный JSON; `event_policy` также частичная, null-порог наследует Collection. **Результат:** 200 с полным Monitor; источник/Collection/частота/политика перезапускают задачу. **Ошибки:** 400, 404, 429.
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**Назначение:** Навсегда удалить Monitor. **Ввод:** `monitor_id` в пути. **Результат:** останавливает декодер, инференс и RTSP, удаляет события из памяти и возвращает 204; Collection остаётся. **Ошибки:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**Назначение:** Опросить текущее состояние из headless-клиента. **Ввод:** ID Monitor. **Результат:** 200 со связью, фактическим FPS, задержкой, пропусками, текущими известными/неизвестными лицами, preview, переподключениями и безопасной ошибкой, без embeddings. **Ошибки:** 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**Назначение:** Получить временные события входа/выхода/ошибки/восстановления. **Ввод:** `limit` 1–1000 и последний непрозрачный `cursor`. **Результат:** 200 с `events`, `next_cursor`, `truncated`, `stream_reset`; перезапуск удаляет события. **Ошибки:** 400 `invalid_cursor`, 401, 404.
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**Назначение:** Открыть необязательный сырой MJPEG preview, по умолчанию выключенный. **Ввод:** ID и обычный Bearer-заголовок; ключ нельзя помещать в URL. **Результат:** долгий `multipart/x-mixed-replace`, кодируемый только при зрителях; рамки клиент берёт из `/state`. **Ошибки:** 401, 404, 409 `preview_disabled`, 503.
+
+## Повтор запросов
+
+GET можно повторять. Перед повтором DELETE проверьте состояние. Если результат создания Person/Face неизвестен из-за сети, прочитайте ID до нового POST. Повторяйте только 429 и временные 503 с ограниченным exponential backoff и jitter; ошибки 4xx требуют исправления запроса.
diff --git a/server/docs/api.zh-CN.md b/server/docs/api.zh-CN.md
new file mode 100644
index 000000000..23090dff8
--- /dev/null
+++ b/server/docs/api.zh-CN.md
@@ -0,0 +1,567 @@
+# InsightFace Server REST API使用手册
+
+**语言：** [English](api.md) · 中文 · [日本語](api.ja.md) · [Deutsch](api.de.md) · [Español](api.es.md) · [Français](api.fr.md) · [Русский](api.ru.md) · [Português](api.pt.md) · [한국어](api.ko.md)
+
+本文覆盖每个公开接口的调用方式、参数含义、服务端执行过程、成功结果和常见错误。
+如果容器和模型还没有启动，请先阅读
+[分步用户指南](user-guide.zh-CN.md)。实时OpenAPI Schema位于`/docs`和
+`/openapi.json`；本文负责解释“怎么用”和“结果代表什么”。
+
+## 通用约定
+
+- API基础路径为`/v1`，JSON字段统一使用`snake_case`。
+- Collection/PATCH请求使用`application/json`；图片和注册使用
+  `multipart/form-data`；裁剪图返回`image/jpeg`；摄像头预览返回MJPEG流。
+- JPEG、PNG和WebP均支持。默认压缩图片上限10 MiB、解码像素上限4000万、整个请求
+  上限64 MiB；实际值以`GET /v1/system`为准。
+- 每个响应都有`x-request-id`响应头；JSON响应还包含同一个`request_id`。排错时记录
+  这个ID，但不要记录图片、embedding、API Key或RTSP凭据。
+- `detection_score`和质量分数范围为`0.0..1.0`。`similarity`是原始cosine值，范围
+  `[-1.0,1.0]`，不是概率。公开匹配阈值范围为`[0.0,1.0]`，判定为
+  `similarity >= threshold`，默认阈值`0.4`。
+- 列表接口的`cursor`是不透明令牌，只能原样交回同一个接口、同一Collection、同一
+  Person和同一筛选条件，不要解析或自行构造。
+
+项目附带的Compose配置在隔离评估环境中默认关闭认证。`GET /v1/health`始终公开；
+管理员启用认证后，其他接口必须发送：
+
+```http
+Authorization: Bearer <api_key>
+```
+
+认证关闭时不要发送空的`Authorization`头，直接省略该字段。
+
+统一错误格式：
+
+```json
+{
+  "error": {
+    "code": "face_not_found",
+    "message": "No usable face was detected.",
+    "details": {}
+  },
+  "request_id": "3ed21e89-4595-4eed-a699-1df42ca62032"
+}
+```
+
+常用状态码：`400`参数错误、`401`未认证、`404`资源不存在、`409`状态或契约冲突、
+`413`请求/图片过大、`422`图片或人脸不符合处理要求、`429`达到流数量限制、`500`
+内部错误、`503`超时或模型/索引不可用。
+
+## 第一次API调用
+
+```bash
+BASE_URL=http://127.0.0.1:18097
+AUTH_HEADER="Authorization: Bearer ${INSIGHTFACE_API_KEY}"
+
+curl -fsS "${BASE_URL}/v1/health"
+
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" \
+  -H 'Content-Type: application/json' \
+  -d '{"id":"employees","name":"员工库","threshold":0.4}'
+
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" \
+  -F 'id=alice' -F 'name=Alice' -F 'review_mode=off' \
+  -F 'images=@alice-enroll.jpg'
+
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" \
+  -F 'image=@alice-query.jpg' -F 'limit=5'
+```
+
+认证关闭时从后三条命令中删除`-H "${AUTH_HEADER}"`。
+
+## 系统接口
+
+### `GET /v1/health`
+
+**用途：** 容器健康检查和就绪探测，公开且无需认证。
+
+**参数：** 无。
+
+**执行与结果：** 检查启动状态和SQLite `quick_check`。就绪时HTTP 200：
+
+```json
+{"status":"ready","auth_enabled":false,"request_id":"..."}
+```
+
+```bash
+curl -sS "${BASE_URL}/v1/health"
+```
+
+**常见错误：** 模型、数据库或索引尚未就绪时返回`503 not_ready`。
+
+### `GET /v1/system`
+
+**用途：** 管理员查看安全的运行诊断。
+
+**参数：** 无。
+
+**执行与结果：** HTTP 200返回Server/OS/CPU/GPU、Compute Capability、Driver、
+CUDA、cuDNN、ORT、实际Provider、模型与License、数据库、挂载目录、Collection/
+Person/Face数量、检索后端、安全配置、推理并发和最近错误。不会返回密钥、图片或特征。
+
+```bash
+curl -sS "${BASE_URL}/v1/system" -H "${AUTH_HEADER}"
+```
+
+**常见错误：** `401 unauthorized`、`503 request_timeout`。
+
+### `GET /v1/models`
+
+**用途：** 查看当前已验证的检测/识别模型、实际Provider和模型授权摘要。
+
+**参数：** 无。
+
+**执行与结果：** HTTP 200返回`models`、`execution_provider`和`license`，不返回
+ONNX文件内容或签名私钥。
+
+```bash
+curl -sS "${BASE_URL}/v1/models" -H "${AUTH_HEADER}"
+```
+
+**常见错误：** `401 unauthorized`。
+
+## 无状态人脸接口
+
+### `POST /v1/detect`
+
+**用途：** 检测一张图片中的所有可用人脸，不写数据库。
+
+**表单参数：** `image`必填；`max_faces`可选，1～100；`collection_id`可选，指定后
+使用该Collection的检测配置，否则使用系统配置。旧参数`min_score`不再支持。
+
+```bash
+curl -sS "${BASE_URL}/v1/detect" -H "${AUTH_HEADER}" \
+  -F 'image=@group.webp' -F 'max_faces=10' -F 'collection_id=employees'
+```
+
+**执行与结果：** 对配置中的每个输入尺寸检测，合并候选后做一次全局NMS，按人脸面积
+降序返回。HTTP 200包含`faces`、`processing_ms`和`request_id`；每张脸包含像素/
+归一化框、五点关键点、检测分数和质量信息。无人脸是成功的`faces: []`。
+
+**常见错误：** `400 request_detection_override_not_supported`、`404` Collection、
+`413`、`422 invalid_image`、`503 request_timeout`。
+
+### `POST /v1/compare`
+
+**用途：** 比对两张图片中按策略选中的单张脸，不持久化。
+
+**表单参数：** `source`和`target`必填；`threshold`可选0～1，默认0.4；
+`collection_id`可选，用于选择Collection检测配置。
+
+```bash
+curl -sS "${BASE_URL}/v1/compare" -H "${AUTH_HEADER}" \
+  -F 'source=@source.jpg' -F 'target=@target.png' -F 'threshold=0.4'
+```
+
+**执行与结果：** 按`largest`或`center_largest`分别选脸、对齐、抽取并L2归一化特征，
+计算原始cosine。HTTP 200返回`matched`、`similarity`、实际`threshold`、两张选中脸、
+`processing_ms`和`request_id`。
+
+**常见错误：** `404` Collection、`413`、`422 invalid_image`或`face_not_found`、
+`503 request_timeout`。
+
+### `POST /v1/embeddings`
+
+**用途：** 为可信集成方抽取一张选中脸的特征；普通注册/搜索不需要调用它。
+
+**表单参数：** `image`必填；`collection_id`可选。旧`face_selection`请求参数不再支持。
+
+```bash
+curl -sS "${BASE_URL}/v1/embeddings" -H "${AUTH_HEADER}" \
+  -F 'image=@portrait.jpg' -F 'collection_id=employees'
+```
+
+**执行与结果：** HTTP 200返回一个`faces`项、L2归一化embedding、`model`、
+`processing_ms`和`request_id`。embedding属于敏感生物特征，服务不会记录其内容。
+
+**常见错误：** `400 request_detection_override_not_supported`、`404`、`413`、
+`422 invalid_image`或`face_not_found`、`503`。
+
+## Collection接口
+
+### `POST /v1/collections`
+
+**用途：** 创建独立人员库，并固定模型、检测和搜索契约。
+
+**JSON参数：** `id`、`name`必填；`description`默认空字符串；`threshold`默认0.4；
+`metadata`默认`{}`；`save_face_crops`默认false。可选`detection`包含`input_sizes`、
+`threshold`、`nms_threshold`、`single_face_selection`；可选`search`包含`profile`、
+`capacity_rows`、`max_faces_per_person`和`load_policy`。ID为`_default`或1～64位，
+首位是字母/数字，其余允许字母、数字、点、下划线和短横线。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections" -H "${AUTH_HEADER}" \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "id":"employees",
+    "name":"公司员工",
+    "threshold":0.4,
+    "search":{"profile":"fp32_v1","capacity_rows":100000,"max_faces_per_person":20,"load_policy":"lazy"},
+    "detection":{"input_sizes":[[96,96],[512,512]],"threshold":0.5,"nms_threshold":0.4,"single_face_selection":"largest"}
+  }'
+```
+
+**执行与结果：** 分配索引并固定当前模型ID、版本、digest、512维特征与预处理版本。
+HTTP 201返回完整`collection`、解析后的默认值、计数和时间戳。
+
+**常见错误：** `400 invalid_detection_profile`、`unsupported_search_profile`或
+`search_capacity_too_large`；`409 collection_exists`；`503 search_index_unavailable`。
+
+### `GET /v1/collections`
+
+**用途：** 分页列出人员库。
+
+**查询参数：** `limit` 1～100，默认50；`cursor`可选不透明令牌。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections?limit=50" -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`collections`和可空`next_cursor`。**常见错误：**
+`400 invalid_cursor`、`401 unauthorized`。
+
+### `GET /v1/collections/{collection_id}`
+
+**用途：** 获取一个人员库。**路径参数：** `collection_id`。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees" -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`collection`、实时`person_count`、`face_count`和
+`embedding_contract_id`。**常见错误：** `404 resource_not_found`。
+
+### `PATCH /v1/collections/{collection_id}`
+
+**用途：** 修改Collection可变策略。**路径参数：** `collection_id`。
+
+**JSON参数：** 可提交`name`、`description`、`threshold`、`metadata`、
+`save_face_crops`；`search`只能修改`capacity_rows`、`max_faces_per_person`、
+`load_policy`；`detection`可修改检测配置。模型绑定和`search.profile`不可修改，未知字段
+及显式null会被拒绝。检测修改从下一次请求生效，不重算已有特征。
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/collections/employees" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"threshold":0.45,"detection":{"single_face_selection":"center_largest"}}'
+```
+
+**结果：** HTTP 200返回完整更新后的`collection`。**常见错误：** `400`、`404`、
+`409`容量/模型契约冲突、`503`索引更新失败。
+
+### `DELETE /v1/collections/{collection_id}`
+
+**用途：** 删除人员库。**路径参数：** `collection_id`；**查询参数：** `force`
+布尔值，默认false。非空Collection必须明确`force=true`。
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees?force=true" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 204，无响应体。**常见错误：** `404`、`409 collection_not_empty`、
+`503 search_index_unavailable`。
+
+## Person与FaceSample接口
+
+### `POST /v1/collections/{collection_id}/persons`
+
+**用途：** 一次创建Person并注册一张或多张FaceSample。
+
+**路径参数：** `collection_id`。**表单参数：** `images`必填且可重复，默认最多20张；
+`id`可选，省略后生成UUID；`name`、`external_id`可选；`metadata`是JSON对象字符串，
+默认`{}`；`review_mode`为`off|standard|strict`，默认`off`；`embedding_mode`为
+`server|external_trusted`，默认`server`。外部模式还必须提交与图片一一对应的
+`external_embeddings` JSON数组以及Collection返回的`embedding_contract_id`。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons" -H "${AUTH_HEADER}" \
+  -F 'id=employee-001' -F 'name=Alice' -F 'external_id=HR-1001' \
+  -F 'metadata={"department":"sales"}' -F 'review_mode=standard' \
+  -F 'images=@alice1.jpg' -F 'images=@alice2.webp'
+```
+
+**执行与结果：** `off`按Collection策略选脸并允许多人脸；`standard`要求恰好一张脸并
+执行尺寸、检测分数、清晰度、亮度和姿态审查；`strict`还要求最佳类内相似度严格大于
+最佳类外相似度。HTTP 201返回`person`、成功`faces`和逐图片`rejected_images`，允许
+部分成功。所有图片失败时返回`422 registration_failed`且不创建Person。
+
+**常见错误：** `400` ID/metadata/图片数量；`404` Collection；`409` Person、外部ID、
+embedding契约、容量或每人样本上限冲突；`413`；`422 registration_failed`；`503
+search_index_unavailable`。若503详情含`write_committed:true`，先查询Person再决定
+是否重试。
+
+### `GET /v1/collections/{collection_id}/persons`
+
+**用途：** 分页列出或筛选人员。**路径参数：** `collection_id`。**查询参数：**
+`limit` 1～100，默认50；`cursor`；`search`可选，匹配Person ID、姓名或外部ID。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons?limit=50&search=alice" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`persons`和`next_cursor`。**常见错误：**
+`400 invalid_cursor`、`404` Collection。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}`
+
+**用途：** 获取一个Person。**路径参数：** `collection_id`、`person_id`。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/employee-001" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`person`、当前`face_count`和时间戳。**常见错误：** `404`。
+
+### `PATCH /v1/collections/{collection_id}/persons/{person_id}`
+
+**用途：** 修改Person展示信息。**路径参数：** `collection_id`、`person_id`。
+**JSON参数：** `name`、`external_id`、对象`metadata`；未知字段拒绝，metadata不可null。
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/collections/employees/persons/employee-001" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"name":"Alice Chen","metadata":{"department":"sales"}}'
+```
+
+**结果：** HTTP 200返回完整`person`。**常见错误：** `400`、`404`、
+`409 external_id_exists`。
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}`
+
+**用途：** 删除Person及其全部FaceSample、embedding和可选裁剪图。
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees/persons/employee-001" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 204，无响应体；成功后搜索不会再返回该Person。**常见错误：**
+`404`、`503 search_index_unavailable`。
+
+### `POST /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**用途：** 给已有Person增量加入FaceSample。
+
+**路径参数：** `collection_id`、`person_id`。**表单参数：** 可重复`images`、
+`review_mode`、`embedding_mode`、`external_embeddings`、`embedding_contract_id`，
+含义与创建Person完全相同。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/employee-001/faces" \
+  -H "${AUTH_HEADER}" -F 'review_mode=standard' \
+  -F 'images=@alice3.jpg' -F 'images=@alice4.png'
+```
+
+**结果：** HTTP 201返回成功`faces`和逐图片`rejected_images`，允许部分成功。
+**常见错误：** 与注册Person相同，另有`404` Person。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces`
+
+**用途：** 分页列出FaceSample元数据，不返回embedding或图片字节。
+
+**路径参数：** `collection_id`、`person_id`；**查询参数：** `limit` 1～100，默认50；
+`cursor`可选。`has_crop`表示是否存在已保存裁剪图。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/employee-001/faces?limit=50" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`faces`和`next_cursor`。**常见错误：**
+`400 invalid_cursor`、`404` Collection或Person。
+
+### `GET /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image`
+
+**用途：** 下载启用保存后存在的112×112管理用人脸裁剪图。它不是原始上传图片。
+
+**路径参数：** `collection_id`、`person_id`、`face_id`。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/persons/employee-001/faces/FACE_ID/image" \
+  -H "${AUTH_HEADER}" -o face-crop.jpg
+```
+
+**结果：** HTTP 200 `image/jpeg`，带`Cache-Control: no-store`。非JSON响应的请求ID只在
+`x-request-id`头中。**常见错误：** `401`、`404` FaceSample或
+`face_image_not_found`。
+
+### `DELETE /v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}`
+
+**用途：** 删除一个FaceSample、embedding和可选裁剪图。
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/collections/employees/persons/employee-001/faces/FACE_ID" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 204，无响应体；返回成功前同步从活动索引移除。**常见错误：**
+`404`、`503 search_index_unavailable`。
+
+## 搜索接口
+
+### `POST /v1/collections/{collection_id}/search`
+
+**用途：** 用一张查询图片在指定人员库中执行1:N Person搜索。
+
+**路径参数：** `collection_id`。**表单参数：** `image`必填；`limit` 1～100，默认5；
+`threshold`可选0～1，省略后使用Collection阈值。旧`face_selection`参数不再支持。
+
+```bash
+curl -sS "${BASE_URL}/v1/collections/employees/search" -H "${AUTH_HEADER}" \
+  -F 'image=@unknown.webp' -F 'limit=5' -F 'threshold=0.4'
+```
+
+**执行与结果：** 按Collection配置选择查询脸，扫描所有有效FaceSample，每个Person取
+最高FaceSample相似度，按相似度降序且只返回达到阈值的结果。HTTP 200返回
+`searched_face`、`matches`、实际`threshold`、`processing_ms`和`request_id`。
+无匹配是成功的`matches: []`。
+
+**常见错误：** `404` Collection、`409 collection_model_mismatch`、`413`、
+`422 invalid_image`或`face_not_found`、`503 search_index_unavailable`或超时。
+
+## RTSP Monitor监控任务
+
+Monitor是持久化的服务端RTSP识别任务。配置保存在SQLite中，处于启用状态的Monitor会
+在Server重启后自动恢复。系统不保存视频帧；事件只存在于有容量上限的内存环形缓冲区，
+进程重启后丢失。解码器只保留最新帧，因此推理变慢时会降低实际执行频率，而不会积压
+已经过时的视频帧。
+
+### `POST /v1/monitors`
+
+**用途：** 创建并可选择立即启动一个Monitor。**请求体：** 使用JSON；`source.url`
+只允许`rtsp://`或`rtsps://`。凭据使用AES-GCM加密保存在`/data`，API只返回脱敏地址。
+
+```json
+{
+  "id": "front-gate",
+  "name": "公司前门",
+  "description": "主入口",
+  "enabled": true,
+  "source": {"type": "rtsp", "url": "rtsp://viewer:secret@camera.example/live"},
+  "collection_id": "employees",
+  "inference_fps": 2.0,
+  "match_threshold": null,
+  "event_buffer_size": 1000,
+  "event_policy": {
+    "confirm_frames": 3,
+    "absence_timeout_seconds": 3.0,
+    "cooldown_seconds": 10.0,
+    "emit_unknown": true
+  },
+  "preview_enabled": false
+}
+```
+
+`match_threshold: null`表示继承Collection阈值；`event_buffer_size`范围10～10000。
+Web预览默认关闭，不打开预览也会持续识别并产生事件。
+
+**结果：** HTTP 201返回完整`monitor`、脱敏源、实际默认值和运行摘要。
+**常见错误：** `400 invalid_request`、`404` Collection、`409 monitor_exists`、
+`429 monitor_limit_exceeded`。
+
+### `GET /v1/monitors`
+
+**用途：** 分页列出持久化Monitor配置和简要运行状态。**查询参数：** `limit`范围
+1～100，默认50；`cursor`必须原样使用上次响应中的`next_cursor`，客户端不应解析。
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors?limit=50" -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回有序`monitors`和可空`next_cursor`。**常见错误：**
+`400 invalid_cursor`表示令牌无效、被修改或作用域不匹配；启用认证时也可能返回401。
+
+### `GET /v1/monitors/{monitor_id}`
+
+**用途：** 读取一个Monitor的持久化配置和最新运行摘要。**路径参数：**
+`monitor_id`是创建时由调用方指定的ID；响应中的RTSP地址不包含用户名、密码和查询值。
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors/front-gate" -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`monitor`，包括事件策略、预览开关、时间戳和`runtime`。
+**常见错误：** `404 monitor_not_found`、`401 unauthorized`。
+
+### `PATCH /v1/monitors/{monitor_id}`
+
+**用途：** 局部修改Monitor，`id`不可修改。**请求体：** 至少提供一个创建接口中的
+可变字段；`event_policy`也支持局部字段。只有更换RTSP地址或凭据时才发送`source`；
+将`match_threshold`设为`null`可恢复继承Collection阈值。
+
+```bash
+curl -sS -X PATCH "${BASE_URL}/v1/monitors/front-gate" \
+  -H "${AUTH_HEADER}" -H 'Content-Type: application/json' \
+  -d '{"inference_fps":1.5,"event_policy":{"confirm_frames":5}}'
+```
+
+修改源、Collection、执行频率、阈值或事件策略会重启该任务；`enabled`控制启停。
+名称、描述、预览和缓冲容量可以在线生效。
+
+**结果：** HTTP 200返回更新后的完整`monitor`。**常见错误：** `400
+invalid_request`、`404` Monitor或Collection、`429 monitor_limit_exceeded`。
+
+### `DELETE /v1/monitors/{monitor_id}`
+
+**用途：** 永久删除一个Monitor配置。**路径参数：** `monitor_id`。操作会停止解码与
+推理线程、释放RTSP连接并丢弃内存状态和事件，但不会删除其Collection。
+
+```bash
+curl -sS -X DELETE "${BASE_URL}/v1/monitors/front-gate" \
+  -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 204，无响应体。**常见错误：** `404 monitor_not_found`、
+`401 unauthorized`。
+
+### `GET /v1/monitors/{monitor_id}/state`
+
+**用途：** 供无界面客户端或Web UI轮询当前运行状态。**返回字段：** 包含连接状态、
+源分辨率/FPS、配置与实际推理频率、耗时、跳帧、当前已识别与陌生人脸、预览查看者、
+重连次数和安全的最近错误；不会包含embedding或源凭据。
+
+```bash
+curl -sS "${BASE_URL}/v1/monitors/front-gate/state" -H "${AUTH_HEADER}"
+```
+
+**结果：** HTTP 200返回`state`，停用的Monitor通常为`stopped`。**常见错误：**
+`404 monitor_not_found`、`401 unauthorized`。
+
+### `GET /v1/monitors/{monitor_id}/events`
+
+**用途：** 通过短轮询获取最近的进入、离开、错误和恢复事件，无需保持长连接。
+**查询参数：** `limit`范围1～1000，默认100；下一次请求原样携带上次的
+`next_cursor`。cursor是包含内部任务epoch和序号的签名不透明字符串。
+
+第一次不带cursor时返回最新的若干事件；后续只返回更新事件。`truncated: true`表示
+客户端落后于环形缓冲区，`stream_reset: true`表示任务已重启，旧cursor属于上一个
+epoch。事件不落盘，Server进程重启后会丢失。
+
+**结果：** HTTP 200返回`events`、`next_cursor`、`has_more`、`truncated`和
+`stream_reset`。**常见错误：** `400 invalid_cursor`、`404 monitor_not_found`、
+`401 unauthorized`。
+
+### `GET /v1/monitors/{monitor_id}/preview.mjpeg`
+
+**用途：** 打开可选的原始MJPEG预览。**认证：** 与其他API一样使用Bearer请求头，
+不要把API Key放进URL。接口返回未画框的`multipart/x-mixed-replace` JPEG流，客户端
+结合`/state`自行绘制人脸框、ID和相似度。
+
+只有`preview_enabled=true`且至少有一个查看者时才进行JPEG编码；关闭预览不会停止
+识别，传输中断后客户端应采用有上限的退避方式重连。
+
+**结果：** HTTP 200长连接二进制流，不是JSON。**常见错误：** `409
+preview_disabled`、`503 stream_unavailable`、`404 monitor_not_found`、401。
+
+## 生产客户端检查表
+
+- 先调用`/v1/health`，再读取`/v1/system`确认Provider、模型和阈值配置。
+- GET可以安全重试；DELETE重试前先读取资源状态。创建Person/FaceSample遇到网络结果
+  不确定时，先按调用方指定ID查询，不要直接重复注册。
+- `429`和临时`503`可使用带抖动的有界指数退避；其他4xx应修正请求而不是重试。
+- 升级前保存当前镜像digest、模型ID/digest、数据库备份和API版本。不要让两个Server
+  进程同时写同一个`/data`目录。
diff --git a/server/docs/images/customer/collections-en.jpg b/server/docs/images/customer/collections-en.jpg
new file mode 100644
index 000000000..1c4b5c8d9
Binary files /dev/null and b/server/docs/images/customer/collections-en.jpg differ
diff --git a/server/docs/images/customer/dashboard-en.jpg b/server/docs/images/customer/dashboard-en.jpg
new file mode 100644
index 000000000..c0d5d5959
Binary files /dev/null and b/server/docs/images/customer/dashboard-en.jpg differ
diff --git a/server/docs/images/customer/monitoring-en.jpg b/server/docs/images/customer/monitoring-en.jpg
new file mode 100644
index 000000000..83d172863
Binary files /dev/null and b/server/docs/images/customer/monitoring-en.jpg differ
diff --git a/server/docs/maintainer-guide.md b/server/docs/maintainer-guide.md
new file mode 100644
index 000000000..2272cdfb8
--- /dev/null
+++ b/server/docs/maintainer-guide.md
@@ -0,0 +1,674 @@
+# InsightFace Server maintainer guide
+
+> **Maintainer reference — English only.**
+>
+> This document covers source architecture, implementation contracts, testing,
+> and release work. Operators and API consumers should use the localized
+> [User Guide](user-guide.md) and [REST API Guide](api.md). No information
+> required for normal installation or use is intentionally kept only here.
+
+## 1. Scope and repository boundaries
+
+The phase-one product is one process and one container:
+
+```text
+Web UI
+REST API
+SQLite
+local ONNX Runtime inference
+mutable in-memory exact search indexes
+server-side RTSP Monitor tasks
+```
+
+Primary code lives under `server/`. It may import selected existing modules from
+`python-package/insightface/`, but Server work must not modify algorithm,
+training, or package behavior unless a separate upstream change explicitly
+requires it.
+
+Do not commit model binaries, signed customer licenses, issuer private keys,
+real face images, customer data, generated databases, or production
+configuration. `/models` is read-only runtime input; `/data` is mutable,
+persistent runtime state.
+
+## 2. Source architecture
+
+```text
+server/
+├── backend/insightface_server/
+│   ├── api/          request and response schemas, authentication
+│   ├── inference/    provider-independent pipeline and ONNX implementation
+│   ├── licensing/    offline model-license verification and trusted keys
+│   ├── models/       manifest, packages, embedding contract
+│   ├── search/       Python facade, native ABI, reference implementation
+│   ├── services/     application workflows and RTSP tasks
+│   └── storage/      migrations, repository, secrets, crop storage
+├── native/           C/C++ CPU and CUDA exact-search implementations
+├── frontend/         dependency-free Web UI and OpenAPI viewer
+├── sdk/python/       lightweight typed synchronous client
+├── migrations/       ordered SQLite migrations
+├── docker/           pinned CPU and CUDA image definitions
+├── deploy/           public Compose definitions
+├── config/           startup-only TOML
+├── docs/             localized user/API guides and this guide
+├── scripts/          manifest, snapshot, smoke, and opt-in validation tools
+└── tests/            unit, API, SDK, UI, native, real-model, Docker tests
+```
+
+FastAPI owns process lifecycle. Storage and model readiness complete before
+readiness becomes healthy. The Web UI calls only public `/v1` operations.
+SQLite is authoritative; native indexes are disposable projections.
+
+## 3. Inference lifecycle and concurrency
+
+Each process creates exactly one detector ONNX Runtime Session and one
+recognizer Session. Sessions are reused concurrently; a request must never
+construct a model Session. Request-specific detector policy and dynamic input
+state must remain local to the call.
+
+The process-wide inference limiter is shared by Detect, Compare, Embeddings,
+enrollment, Search query extraction, and RTSP frames. `max_concurrency="auto"`
+resolves to 4 on CPU and 8 on CUDA. FastAPI requests are asynchronous, while
+blocking decode and inference work runs in worker threads. The limiter bounds
+model pipelines rather than serializing the whole HTTP request.
+
+ONNX Runtime Sessions support concurrent `run` calls. Do not add a global
+inference lock to work around request-local mutable state; eliminate or isolate
+that state. Locks remain appropriate for:
+
+- SQLite writes and migrations;
+- one Collection's index mutation/rebuild/revision barrier;
+- Monitor lifecycle and bounded event-ring mutation;
+- cache publication where duplicate construction would violate an invariant.
+
+Dynamic SCRFD evaluates every configured input resolution, maps candidates back
+to source-image coordinates, concatenates them, and applies exactly one global
+NMS. It does not NMS each resolution independently and merge the winners.
+Single-face selection supports:
+
+```text
+largest
+center_largest
+```
+
+`center_largest` maximizes:
+
+```text
+area - 2.0 * squared_distance(face_box_center, image_center)
+```
+
+Confidence is not part of that selection score.
+
+Startup performs real detector warm-up at every configured resolution and a
+real recognizer warm-up, validates the embedding dimension, and publishes the
+actual runtime summary through `/v1/system` and `/v1/models`.
+
+## 4. CUDA strict-provider contract
+
+The CUDA image fixes Ubuntu 22.04, Python 3.11, CUDA Runtime 12.9.1, cuDNN
+9.24.0, and the Microsoft CUDA 12 `onnxruntime-gpu==1.27.0` wheel. Never infer
+an actual runtime version from a tag alone.
+
+Startup must fail unless all of these succeed:
+
+1. Linux x86_64 and version-pin validation;
+2. GPU, Compute Capability, and Driver discovery;
+3. `CUDAExecutionProvider` availability and primary Session placement;
+4. actual loaded CUDA and cuDNN library version inspection;
+5. manifest and signed model-license verification;
+6. real detector and recognizer Session creation;
+7. real warm-up through both graphs;
+8. recognition output-dimension verification;
+9. ORT profile evidence for CUDA kernels;
+10. strict rejection of unexpected CPU/non-CUDA model compute.
+
+Dynamic graphs may place bounded integer shape-metadata operations on
+`CPUExecutionProvider`. The strict audit may allow only explicit reviewed
+operators with bounded metadata output. Never add convolution, recognition, or
+a broad operator class to silence an audit failure.
+
+Architecture compatibility, actual validation, and Community Tested reports
+are distinct labels. A hardware claim requires a dated record containing the
+source revision, image digest, GPU/Driver, actual CUDA/cuDNN/ORT, model
+identities/digests, provider lists, strict audit, functional flow, and
+consistency result.
+
+## 5. Model bundle and offline license design
+
+A model bundle contains detector and recognizer ONNX files, `manifest.json`,
+and `MODEL.LICENSE`. Normal startup never downloads models. The `models`
+Compose tool supports install and verify for public packages; the manifest
+helper supports controlled private bundles.
+
+The manifest is the runtime truth for:
+
+- detector and recognizer file names;
+- task and model identities;
+- model version;
+- input size and dynamic-input behavior;
+- embedding dimension;
+- normalization and preprocessing versions;
+- artifact SHA-256 used for diagnostics and Collection compatibility.
+
+The signed model license is an offline compliance credential, not DRM. It binds
+authorization to `model_id`, issuer, use, and optional validity interval. It
+does not bind the exact artifact digest so an authorized customer may perform
+approved FP16 or other format conversion. Verification uses the bundled
+InsightFace Ed25519 public key. The issuer private key stays under an ignored
+private issuer directory and never enters a container or commit.
+
+The four public package licenses start at `2021-09-22T00:00:00Z`, have no end
+date, and state non-commercial use. A future private model can share a detector
+with a public bundle but needs a license whose `model_id` matches the private
+recognizer identity. A not-yet-effective or expired commercial license must
+fail startup clearly.
+
+Changing detector or recognizer can change detection, alignment, preprocessing,
+or embedding semantics. The computed bundle contract therefore pins every
+Collection. Do not bypass `collection_model_mismatch`; phase one intentionally
+requires explicit rebuild or migration.
+
+## 6. Embedding and score contract
+
+The recognizer consumes the five-point aligned face. Optional persisted crops
+are a separate 112×112 resized bounding-box crop; they are not the aligned
+recognition tensor.
+
+Canonical durable embeddings are FP32 and L2-normalized. Query embeddings are
+also normalized. Public search and threshold decisions use:
+
+```text
+similarity = dot(normalized_database_embedding, normalized_query_embedding)
+```
+
+Similarity is raw cosine in `[-1, 1]`, not probability. Public thresholds are
+restricted to `[0, 1]`, default to `0.4`, and match inclusively with
+`similarity >= threshold`.
+
+`external_trusted` intentionally skips recognition extraction. Every feature
+still requires a same-index image for decode, detection, selection, quality,
+metadata, and optional crop behavior. Validate image/vector count, exact
+Collection `embedding_contract_id`, dimension, finite values, nonzero norm, and
+L2 norm within `1.0 ± 0.0002`. Passing values are converted to canonical FP32
+and normalized once for floating-point drift.
+
+The service cannot prove that an external vector belongs to its supplied image.
+This mode places the caller inside the biometric trust boundary. Never log the
+vector, silently replace it with a server feature, or invoke the recognizer
+after an `external_trusted` validation failure.
+
+Review modes:
+
+- `off`: select using the Collection strategy; multiple faces are allowed;
+- `standard`: exactly one usable face plus configured size, detector,
+  sharpness, brightness, aggregate quality, and pose checks;
+- `strict`: standard plus best within-Person similarity greater than best
+  outside-Person similarity.
+
+Quality rules are documented heuristics, not AWS-equivalent quality scores.
+
+## 7. Exact search implementation
+
+Phase one uses one exact mutable in-process index per active Collection. It does
+not use FAISS and does not perform approximate candidate selection. A Person's
+score is the maximum similarity among all of that Person's live FaceSamples.
+Ties are deterministic.
+
+Profiles:
+
+| Profile | Stored vector | CPU | CUDA | Notes |
+| --- | --- | --- | --- | --- |
+| `fp32_v1` | FP32 | yes | yes | flat FP32 inner product |
+| `fp16_v1` | FP16 | no | yes | low-precision approximation |
+| `bf16_v1` | BF16 | supported CPU | SM80+ | low-precision approximation |
+| `int8_x736_v1` | INT8 | yes | yes | recommended INT8 |
+| `int8_x1000_v1` | INT8 | yes | yes | compatibility contract |
+
+Production images load the native CPU or CUDA shared library and expose only
+its capability mask. Python then rejects an unavailable profile before
+Collection creation/load; it does not attempt to load an incompatible symbol
+and continue.
+
+INT8 encoding is deterministic:
+
+```text
+q = clamp(round_half_away_from_zero(x * S), -128, 127)
+score_internal = int32_dot(q_database, q_query) / S²
+```
+
+`S` is 736 or 1000 as encoded in the immutable profile name. Accumulation is
+INT32. Exact row and Person ordering uses the unclipped internal score; only the
+public response is clamped to cosine range. There is no rerank stage.
+
+All profiles exhaustively score every live FaceSample, so “exact” describes
+candidate coverage. Low-precision arithmetic can still produce ordering
+differences relative to FP32.
+
+## 8. Capacity, mutation, and restart invariants
+
+`capacity_rows` is both initial reservation and the Server-level live-row limit.
+The default is 100,000, guarded by a deployment maximum of 10,000,000. Vector
+bytes alone for 512 dimensions are:
+
+```text
+FP32: 2,048 bytes/row
+FP16 or BF16: 1,024 bytes/row
+INT8: 512 bytes/row
+```
+
+IDs, liveness metadata, group mappings, score buffers, Top-K workspaces, and
+allocator overhead are additional. CUDA reserves row/group metadata and
+grouped Top-K workspaces with Collection capacity to avoid first-query growth.
+
+SQLite is authoritative. An accepted add/delete holds the Collection index
+lock and follows this durable barrier:
+
+1. write FaceSample change, pending search change, and next
+   `search_revision` in one SQLite transaction;
+2. commit SQLite;
+3. apply the mutation to the active native generation;
+4. verify native applied revision equals SQLite revision;
+5. acknowledge covered pending changes;
+6. return success.
+
+A successful add is therefore immediately searchable; a successful deletion is
+immediately absent. If native mutation fails after the commit, discard and
+rebuild the generation. If rebuild also fails, return
+`503 search_index_unavailable` with `write_committed=true` and committed
+revision.
+
+No index binary is persisted. At startup, eager Collections rebuild before
+readiness; lazy Collections rebuild on first search. Rebuild streams SQLite in
+batches, verifies the revision did not change, validates row counts, and
+atomically publishes the generation.
+
+CPU deletion reuses a slot. CUDA deletion creates a tombstone; a later add may
+trigger a deterministic rebuild from live SQLite rows to reuse physical
+capacity. The process lock beside SQLite prevents two application workers from
+serving unsafe independent indexes over one data directory.
+
+CUDA grouped Top-K stays device-resident and evaluates every live row, reduces
+to the best FaceSample per Person, and returns only final records over PCIe.
+CPU uses the same score/order contract with host result handling.
+
+## 9. SQLite, cursors, secrets, and crops
+
+Add a new ordered migration instead of modifying an applied schema. Migrations
+run transactionally. SQLite uses foreign keys, WAL, a busy timeout, and an
+in-process write lock.
+
+Opaque list and event cursors must remain signed and scoped. Resource list
+cursors use `/data/cursor.key` and must not reveal offsets or SQL. Monitor event
+cursors contain an opaque epoch/sequence contract and report truncation or
+stream reset without promising durability.
+
+The startup API key is stored only as a random-salted scrypt digest. A different
+`INSIGHTFACE_API_KEY` on a later start atomically deactivates the previous
+credential and activates the new digest. Phase one deliberately has no runtime
+multi-key, scope, role, or revocation API.
+
+RTSP credentials are encrypted under a data-volume key. API responses always
+redact user information and query secrets. Never include credentials in preview
+URLs, logs, errors, or events.
+
+`save_face_crops` is resolved per Collection and defaults false. Accepted
+registrations may store one 112×112 bounding-box JPEG BLOB. Original uploads,
+aligned recognition inputs, rejected images, and RTSP frames are never stored.
+Logical deletion is not forensic erasure from WAL, snapshots, or backups.
+
+## 10. RTSP Monitor runtime
+
+Monitor configuration persists in SQLite. Each enabled Monitor owns a server
+task, decoder lifecycle, latest-frame slot, schedule, state snapshot, and
+bounded in-memory event ring. It shares the process inference budget and model
+Sessions; it does not create separate ONNX Sessions.
+
+The decoder keeps only the newest frame. If inference takes longer than the
+requested interval, stale frames are counted as skipped instead of queued.
+Multiple clients can independently poll one Monitor's state and recent events;
+they see the same server state but maintain their own cursors.
+
+Events are intentionally non-durable. If no client polls, old events fall out
+of the bounded ring. A process/task restart creates a new stream epoch.
+
+`preview_enabled` defaults false. Recognition and events do not depend on a
+preview client. JPEG encoding begins lazily only while preview is enabled and
+at least one viewer is connected. `/preview.mjpeg` carries raw frames; clients
+draw labels from `/state`. Closing every browser must not stop the Monitor.
+
+Lifecycle updates:
+
+- source, Collection, cadence, threshold, or event-policy changes restart the
+  task;
+- name, description, preview, and buffer-size changes do not require restart;
+- `enabled=false` stops runtime but preserves configuration;
+- DELETE stops runtime and removes configuration and volatile state.
+
+## 11. Security invariants
+
+The Server processes sensitive biometric data. Preserve these design rules:
+
+- no image, embedding, API key, RTSP credential, or multipart-body logging;
+- default-deny CORS, exact trusted origins only;
+- image-byte, decoded-pixel, request-body, image-count, and request-time limits;
+- `/models` read-only and `/data` the only durable writable area;
+- non-root UID/GID 10001, read-only root filesystem, dropped capabilities,
+  `no-new-privileges`, bounded `/tmp`;
+- no remote model download during normal startup;
+- authenticated and non-cacheable crop/embedding access;
+- self-hosted UI with CSP and no CDN, analytics, remote font, or third-party JS;
+- plain HTTP only inside the container; deployment owns TLS termination,
+  ingress/egress restriction, and rate limiting.
+
+Do not present face recognition as the sole control for a high-impact decision.
+Thresholds and quality settings require deployment-specific validation.
+
+## 12. Local development
+
+Use Python 3.11:
+
+```bash
+python3.11 -m venv server/.venv
+. server/.venv/bin/activate
+python -m pip install -r server/requirements.dev.lock
+```
+
+CPU and CUDA lock files remain separate because only one ONNX Runtime package
+may be installed in each image. Frontend tests use Node's built-in test runner
+and have no package-install step.
+
+Unified commands from the repository root:
+
+```bash
+make -C server lint
+make -C server test
+make -C server test-api
+make -C server test-sdk
+make -C server test-frontend
+make -C server test-native-cpu
+make -C server build-cpu
+make -C server build-cuda12
+make -C server run-cpu
+make -C server run-cuda12
+make -C server test-cpu
+make -C server test-cuda12
+make -C server test-consistency
+make -C server smoke-test
+```
+
+Public tests use mock inference and synthetic data; they must not require
+models, commercial assets, real faces, GPUs, cameras, or external network
+services. Published images force ONNX mode and reject mock inference.
+
+## 13. Native and real-model validation
+
+`make -C server test-native-cpu` builds the native library and checks the C ABI,
+score semantics, grouped Person Top-K, capacity add/delete/reuse, CPU
+FP32/BF16/INT8, and explicit unsupported FP16 behavior. The library avoids
+`-march=native`; runtime dispatch selects a compatible optimized kernel.
+
+Real-model tests are opt-in and use ignored `/models` plus private authorized
+images. Record at least:
+
+- detection count and area ordering;
+- embedding dimension and L2 norm;
+- CPU/GPU cosine tolerance;
+- 1:N Person Top-K order;
+- decisions near thresholds;
+- actual Session providers and CUDA strict audit;
+- model/runtime/hardware identity.
+
+Docker validation records build result/digest, build-time verifier, startup
+output, actual `/v1/system`, CRUD/enrollment/search, restart persistence,
+strict CUDA audit, and CPU/GPU consistency. Never derive a validation claim
+only from Dockerfile text or a base-image tag.
+
+RTSP E2E additionally checks credential encryption/redaction, task restoration,
+bounded event cursor behavior, no-backlog scheduling, viewer-independent
+execution, reconnects, optional preview, and client-side overlays.
+
+## 14. Public API documentation change gate
+
+A public API addition, modification, rename, deprecation, or removal is not
+complete until the same change includes:
+
+1. FastAPI/Pydantic behavior and OpenAPI metadata;
+2. all nine localized `docs/api*.md` operation sections, including purpose,
+   authentication, parameters, defaults/ranges/enums, server behavior, success,
+   errors, side effects, pagination/retry guidance, and an example;
+3. all affected localized `docs/user-guide*.md` workflows;
+4. Python SDK and Web UI behavior when exposed there;
+5. API, SDK, UI, and documentation contract tests;
+6. reviewed `make -C server update-api-docs` snapshot diff;
+7. compatibility or migration notes in the affected guides.
+
+`tests/api/test_documentation_contract.py` compares the runtime public
+method/path set with every localized API Guide and compares live OpenAPI with
+`docs/openapi.snapshot.json`. Never resolve a failure by updating only the
+snapshot.
+
+README files are GitHub overview/quick-start pages and are not copied into
+images. User Guides and API Guides are the single user-facing Markdown sources
+for both GitHub and Web UI. This file is the single English-only maintainer
+reference. `/docs` and `/openapi.json` remain the live machine-oriented schema.
+
+## 15. Container versioning and release
+
+InsightFace Server has no repository-hosted CI or release pipeline. The
+repository owner performs validation, image publication, and stable-channel
+promotion manually from trusted Linux hosts. This keeps registry credentials
+and private model/test assets outside the repository.
+
+All public variants share:
+
+```text
+ghcr.io/deepinsight/insightface-server
+```
+
+Each stable version has two immutable tags built from one source revision:
+
+```text
+<major>.<minor>.<patch>-cpu
+<major>.<minor>.<patch>-cuda12
+```
+
+`cpu` and `cuda12` are moving stable channels. Never publish `latest`, never
+overwrite an immutable version tag, and never move either stable channel until
+both versioned variants have passed validation.
+
+### 15.1 Prepare and validate the source
+
+Use the relaxed precheck while the tree contains local edits:
+
+```bash
+make -C server release-precheck
+```
+
+Before publication, commit all release source and use a clean worktree. The
+strict check verifies synchronized backend, SDK, documentation, Compose, Docker,
+license metadata, Git revision, and release-tag state:
+
+```bash
+make -C server release-preflight
+make -C server lint
+make -C server test
+make -C server test-api
+make -C server test-sdk
+make -C server test-frontend
+make -C server test-native-cpu
+```
+
+Build both Python distributions in isolated output directories and run metadata
+validation with the maintainer's packaging environment:
+
+```bash
+python3.11 -m build --sdist --wheel --outdir /tmp/ifs-server-dist server
+python3.11 -m build --sdist --wheel --outdir /tmp/ifs-sdk-dist server/sdk/python
+python3.11 -m twine check /tmp/ifs-server-dist/* /tmp/ifs-sdk-dist/*
+```
+
+### 15.2 Build and qualify both images
+
+Build both immutable version tags from the exact committed revision:
+
+```bash
+export RELEASE_VERSION=0.2.0
+export RELEASE_IMAGE=ghcr.io/deepinsight/insightface-server
+export RELEASE_SHA="$(git rev-parse HEAD)"
+
+make -C server build-cpu \
+  SERVER_VERSION="$RELEASE_VERSION" \
+  CPU_IMAGE="$RELEASE_IMAGE:$RELEASE_VERSION-cpu"
+make -C server build-cuda12 \
+  SERVER_VERSION="$RELEASE_VERSION" \
+  CUDA_IMAGE="$RELEASE_IMAGE:$RELEASE_VERSION-cuda12"
+```
+
+Run the CPU image on a Linux x86-64 host and the CUDA image on a compatible
+NVIDIA host with the verified model directory mounted read-only. For both
+variants, check startup, `/v1/health`, `/v1/system`, CRUD/enrollment/search,
+restart persistence, and a private authorized image containing exactly one
+usable face:
+
+```bash
+python3.11 server/scripts/smoke_test.py \
+  --base-url http://127.0.0.1:8080 \
+  --image /absolute/path/to/release-test-image.jpg
+```
+
+The CUDA result must report `CUDAExecutionProvider` and pass strict-provider
+startup checks. Compare CPU and CUDA results for the same image and retain only
+redacted validation notes; never commit the image, embedding, credentials, RTSP
+URLs, or raw biometric output.
+
+### 15.3 Lock the source identity
+
+Fetch tags before any registry write and stop on a tag conflict. Reuse an
+existing release tag only when it is an annotated, validly signed tag for the
+exact release commit; otherwise create and verify it. Pushing the same
+source-bound tag is safe to retry, while a conflicting remote tag is rejected:
+
+```bash
+git fetch origin --tags || exit 1
+if git show-ref --verify --quiet "refs/tags/server-v$RELEASE_VERSION"; then
+  if test "$(git rev-list -n 1 "server-v$RELEASE_VERSION")" != "$RELEASE_SHA"; then
+    echo "release tag points to another revision" >&2
+    exit 1
+  fi
+else
+  git tag -s "server-v$RELEASE_VERSION" "$RELEASE_SHA" \
+    -m "InsightFace Server $RELEASE_VERSION" || exit 1
+fi
+git tag -v "server-v$RELEASE_VERSION" || exit 1
+git push origin \
+  "refs/tags/server-v$RELEASE_VERSION:refs/tags/server-v$RELEASE_VERSION" ||
+  exit 1
+```
+
+### 15.4 Publish manually to GHCR
+
+Authenticate using a maintainer token with package-write permission, supplied
+through standard input:
+
+```bash
+read -r -p "GHCR user: " GHCR_USER
+read -r -s -p "GHCR token: " GHCR_TOKEN
+printf '%s' "$GHCR_TOKEN" |
+  docker login ghcr.io --username "$GHCR_USER" --password-stdin
+unset GHCR_TOKEN
+```
+
+Before pushing, run both commands below. Each must fail specifically because
+the manifest is not found. A returned manifest means the immutable tag already
+exists; an authentication, DNS, timeout, or other operational error is also a
+hard stop. Never infer “not found” from an arbitrary command failure.
+
+```bash
+for VARIANT in cpu cuda12; do
+  if docker buildx imagetools inspect \
+    "$RELEASE_IMAGE:$RELEASE_VERSION-$VARIANT"; then
+    echo "immutable tag already exists; stop" >&2
+    exit 1
+  fi
+  read -r -p "Confirm $VARIANT failed only with manifest-not-found [yes]: " CONFIRM
+  test "$CONFIRM" = yes || exit 1
+done
+```
+
+Only after confirming both immutable tags are absent, push both validated
+images:
+
+```bash
+docker push "$RELEASE_IMAGE:$RELEASE_VERSION-cpu"
+docker push "$RELEASE_IMAGE:$RELEASE_VERSION-cuda12"
+```
+
+Resolve and validate the two remote digests. These digest references, not the
+tag names, are the promotion inputs:
+
+```bash
+export CPU_DIGEST="$(
+  docker buildx imagetools inspect \
+    "$RELEASE_IMAGE:$RELEASE_VERSION-cpu" \
+    --format '{{json .Manifest}}' |
+  python3.11 -c 'import json,sys; print(json.load(sys.stdin)["digest"])'
+)"
+export CUDA_DIGEST="$(
+  docker buildx imagetools inspect \
+    "$RELEASE_IMAGE:$RELEASE_VERSION-cuda12" \
+    --format '{{json .Manifest}}' |
+  python3.11 -c 'import json,sys; print(json.load(sys.stdin)["digest"])'
+)"
+case "$CPU_DIGEST:$CUDA_DIGEST" in
+  sha256:*:sha256:*) ;;
+  *) echo "invalid release digest" >&2; exit 1 ;;
+esac
+```
+
+Record any existing `cpu` and `cuda12` stable-channel digests before changing
+them. Only after both versioned digests are readable and verified, move the two
+stable channels to those exact digest references:
+
+```bash
+docker buildx imagetools create \
+  --prefer-index=false \
+  --tag "$RELEASE_IMAGE:cpu" "$RELEASE_IMAGE@$CPU_DIGEST"
+docker buildx imagetools create \
+  --prefer-index=false \
+  --tag "$RELEASE_IMAGE:cuda12" "$RELEASE_IMAGE@$CUDA_DIGEST"
+```
+
+Read both stable tags back and require their digests to match the recorded
+versioned digests:
+
+```bash
+test "$(
+  docker buildx imagetools inspect "$RELEASE_IMAGE:cpu" \
+    --format '{{json .Manifest}}' |
+  python3.11 -c 'import json,sys; print(json.load(sys.stdin)["digest"])'
+)" = "$CPU_DIGEST"
+test "$(
+  docker buildx imagetools inspect "$RELEASE_IMAGE:cuda12" \
+    --format '{{json .Manifest}}' |
+  python3.11 -c 'import json,sys; print(json.load(sys.stdin)["digest"])'
+)" = "$CUDA_DIGEST"
+docker logout ghcr.io
+```
+
+Registry updates to `cpu` and `cuda12` are not atomic. Record their previous
+digests before changing them. If promotion is interrupted, stop, inspect all
+four tags, and manually restore both stable channels to the recorded digests or
+finish the matching pair; never rebuild or overwrite a versioned tag.
+
+## 16. Contribution checklist
+
+- Preserve unrelated work in a dirty tree.
+- Add migrations; never rewrite an applied schema.
+- Keep model Sessions singleton and request policy isolated.
+- Keep SQLite authoritative and index mutations behind the revision barrier.
+- Keep API, SDK, Web UI, all localized guides, and OpenAPI synchronized.
+- Test error and retry behavior, not only the success path.
+- Retain actual hardware/model evidence for compatibility claims.
+- Do not weaken provider, license, digest, capacity, or input checks to make an
+  incompatible environment appear healthy.
+- Never push images, code, releases, or customer artifacts unless explicitly
+  authorized.
diff --git a/server/docs/openapi.snapshot.json b/server/docs/openapi.snapshot.json
new file mode 100644
index 000000000..f76384596
--- /dev/null
+++ b/server/docs/openapi.snapshot.json
@@ -0,0 +1,6783 @@
+{
+  "components": {
+    "schemas": {
+      "Body_add_faces_v1_collections__collection_id__persons__person_id__faces_post": {
+        "properties": {
+          "embedding_contract_id": {
+            "anyOf": [
+              {
+                "maxLength": 256,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Embedding Contract Id"
+          },
+          "embedding_mode": {
+            "default": "server",
+            "enum": [
+              "server",
+              "external_trusted"
+            ],
+            "title": "Embedding Mode",
+            "type": "string"
+          },
+          "external_embeddings": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "External Embeddings"
+          },
+          "images": {
+            "items": {
+              "contentMediaType": "application/octet-stream",
+              "type": "string"
+            },
+            "title": "Images",
+            "type": "array"
+          },
+          "review_mode": {
+            "default": "off",
+            "enum": [
+              "off",
+              "standard",
+              "strict"
+            ],
+            "title": "Review Mode",
+            "type": "string"
+          }
+        },
+        "required": [
+          "images"
+        ],
+        "title": "Body_add_faces_v1_collections__collection_id__persons__person_id__faces_post",
+        "type": "object"
+      },
+      "Body_compare_v1_compare_post": {
+        "properties": {
+          "collection_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Collection Id"
+          },
+          "source": {
+            "contentMediaType": "application/octet-stream",
+            "title": "Source",
+            "type": "string"
+          },
+          "target": {
+            "contentMediaType": "application/octet-stream",
+            "title": "Target",
+            "type": "string"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "required": [
+          "source",
+          "target"
+        ],
+        "title": "Body_compare_v1_compare_post",
+        "type": "object"
+      },
+      "Body_create_person_v1_collections__collection_id__persons_post": {
+        "properties": {
+          "embedding_contract_id": {
+            "anyOf": [
+              {
+                "maxLength": 256,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Embedding Contract Id"
+          },
+          "embedding_mode": {
+            "default": "server",
+            "enum": [
+              "server",
+              "external_trusted"
+            ],
+            "title": "Embedding Mode",
+            "type": "string"
+          },
+          "external_embeddings": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "External Embeddings"
+          },
+          "external_id": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "External Id"
+          },
+          "id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Id"
+          },
+          "images": {
+            "items": {
+              "contentMediaType": "application/octet-stream",
+              "type": "string"
+            },
+            "title": "Images",
+            "type": "array"
+          },
+          "metadata": {
+            "default": "{}",
+            "title": "Metadata",
+            "type": "string"
+          },
+          "name": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Name"
+          },
+          "review_mode": {
+            "default": "off",
+            "enum": [
+              "off",
+              "standard",
+              "strict"
+            ],
+            "title": "Review Mode",
+            "type": "string"
+          }
+        },
+        "required": [
+          "images"
+        ],
+        "title": "Body_create_person_v1_collections__collection_id__persons_post",
+        "type": "object"
+      },
+      "Body_detect_v1_detect_post": {
+        "properties": {
+          "collection_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Collection Id"
+          },
+          "image": {
+            "contentMediaType": "application/octet-stream",
+            "title": "Image",
+            "type": "string"
+          },
+          "max_faces": {
+            "anyOf": [
+              {
+                "maximum": 100.0,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Max Faces"
+          },
+          "min_score": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "deprecated": true,
+            "title": "Min Score"
+          }
+        },
+        "required": [
+          "image"
+        ],
+        "title": "Body_detect_v1_detect_post",
+        "type": "object"
+      },
+      "Body_embeddings_v1_embeddings_post": {
+        "properties": {
+          "collection_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Collection Id"
+          },
+          "face_selection": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "deprecated": true,
+            "title": "Face Selection"
+          },
+          "image": {
+            "contentMediaType": "application/octet-stream",
+            "title": "Image",
+            "type": "string"
+          }
+        },
+        "required": [
+          "image"
+        ],
+        "title": "Body_embeddings_v1_embeddings_post",
+        "type": "object"
+      },
+      "Body_search_v1_collections__collection_id__search_post": {
+        "properties": {
+          "face_selection": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "deprecated": true,
+            "title": "Face Selection"
+          },
+          "image": {
+            "contentMediaType": "application/octet-stream",
+            "title": "Image",
+            "type": "string"
+          },
+          "limit": {
+            "default": 5,
+            "maximum": 100.0,
+            "minimum": 1.0,
+            "title": "Limit",
+            "type": "integer"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "required": [
+          "image"
+        ],
+        "title": "Body_search_v1_collections__collection_id__search_post",
+        "type": "object"
+      },
+      "BoundingBox": {
+        "additionalProperties": true,
+        "properties": {
+          "normalized": {
+            "$ref": "#/components/schemas/NormalizedBox"
+          },
+          "pixels": {
+            "$ref": "#/components/schemas/PixelBox"
+          }
+        },
+        "required": [
+          "pixels",
+          "normalized"
+        ],
+        "title": "BoundingBox",
+        "type": "object"
+      },
+      "Collection": {
+        "additionalProperties": true,
+        "properties": {
+          "capacity_rows": {
+            "title": "Capacity Rows",
+            "type": "integer"
+          },
+          "created_at": {
+            "title": "Created At",
+            "type": "string"
+          },
+          "default_threshold": {
+            "title": "Default Threshold",
+            "type": "number"
+          },
+          "description": {
+            "title": "Description",
+            "type": "string"
+          },
+          "detection": {
+            "$ref": "#/components/schemas/DetectionProfile"
+          },
+          "detection_revision": {
+            "title": "Detection Revision",
+            "type": "integer"
+          },
+          "embedding_contract_id": {
+            "title": "Embedding Contract Id",
+            "type": "string"
+          },
+          "embedding_dimension": {
+            "title": "Embedding Dimension",
+            "type": "integer"
+          },
+          "face_count": {
+            "title": "Face Count",
+            "type": "integer"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "load_policy": {
+            "enum": [
+              "eager",
+              "lazy"
+            ],
+            "title": "Load Policy",
+            "type": "string"
+          },
+          "max_faces_per_person": {
+            "title": "Max Faces Per Person",
+            "type": "integer"
+          },
+          "metadata": {
+            "additionalProperties": true,
+            "title": "Metadata",
+            "type": "object"
+          },
+          "model_digest": {
+            "title": "Model Digest",
+            "type": "string"
+          },
+          "model_id": {
+            "title": "Model Id",
+            "type": "string"
+          },
+          "model_version": {
+            "title": "Model Version",
+            "type": "string"
+          },
+          "name": {
+            "title": "Name",
+            "type": "string"
+          },
+          "person_count": {
+            "title": "Person Count",
+            "type": "integer"
+          },
+          "preprocessing_version": {
+            "title": "Preprocessing Version",
+            "type": "string"
+          },
+          "save_face_crops": {
+            "title": "Save Face Crops",
+            "type": "boolean"
+          },
+          "search_profile": {
+            "title": "Search Profile",
+            "type": "string"
+          },
+          "search_revision": {
+            "title": "Search Revision",
+            "type": "integer"
+          },
+          "updated_at": {
+            "title": "Updated At",
+            "type": "string"
+          }
+        },
+        "required": [
+          "id",
+          "name",
+          "description",
+          "default_threshold",
+          "model_id",
+          "model_version",
+          "model_digest",
+          "embedding_dimension",
+          "preprocessing_version",
+          "embedding_contract_id",
+          "search_profile",
+          "capacity_rows",
+          "max_faces_per_person",
+          "load_policy",
+          "search_revision",
+          "save_face_crops",
+          "metadata",
+          "person_count",
+          "face_count",
+          "created_at",
+          "updated_at",
+          "detection",
+          "detection_revision"
+        ],
+        "title": "Collection",
+        "type": "object"
+      },
+      "CollectionCreate": {
+        "additionalProperties": false,
+        "properties": {
+          "description": {
+            "default": "",
+            "maxLength": 2000,
+            "title": "Description",
+            "type": "string"
+          },
+          "detection": {
+            "$ref": "#/components/schemas/CollectionDetectionCreate"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "metadata": {
+            "additionalProperties": true,
+            "title": "Metadata",
+            "type": "object"
+          },
+          "name": {
+            "maxLength": 200,
+            "minLength": 1,
+            "title": "Name",
+            "type": "string"
+          },
+          "save_face_crops": {
+            "anyOf": [
+              {
+                "type": "boolean"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Save Face Crops"
+          },
+          "search": {
+            "$ref": "#/components/schemas/CollectionSearchCreate"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "required": [
+          "id",
+          "name"
+        ],
+        "title": "CollectionCreate",
+        "type": "object"
+      },
+      "CollectionDetectionCreate": {
+        "additionalProperties": false,
+        "properties": {
+          "input_sizes": {
+            "anyOf": [
+              {
+                "items": {
+                  "maxItems": 2,
+                  "minItems": 2,
+                  "prefixItems": [
+                    {
+                      "type": "integer"
+                    },
+                    {
+                      "type": "integer"
+                    }
+                  ],
+                  "type": "array"
+                },
+                "type": "array"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Input Sizes"
+          },
+          "nms_threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Nms Threshold"
+          },
+          "single_face_selection": {
+            "anyOf": [
+              {
+                "enum": [
+                  "largest",
+                  "center_largest"
+                ],
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Single Face Selection"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "title": "CollectionDetectionCreate",
+        "type": "object"
+      },
+      "CollectionDetectionPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "input_sizes": {
+            "anyOf": [
+              {
+                "items": {
+                  "maxItems": 2,
+                  "minItems": 2,
+                  "prefixItems": [
+                    {
+                      "type": "integer"
+                    },
+                    {
+                      "type": "integer"
+                    }
+                  ],
+                  "type": "array"
+                },
+                "type": "array"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Input Sizes"
+          },
+          "nms_threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Nms Threshold"
+          },
+          "single_face_selection": {
+            "anyOf": [
+              {
+                "enum": [
+                  "largest",
+                  "center_largest"
+                ],
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Single Face Selection"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "title": "CollectionDetectionPatch",
+        "type": "object"
+      },
+      "CollectionPageResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "collections": {
+            "items": {
+              "$ref": "#/components/schemas/Collection"
+            },
+            "title": "Collections",
+            "type": "array"
+          },
+          "next_cursor": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Next Cursor"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "collections",
+          "next_cursor"
+        ],
+        "title": "CollectionPageResponse",
+        "type": "object"
+      },
+      "CollectionPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "description": {
+            "anyOf": [
+              {
+                "maxLength": 2000,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Description"
+          },
+          "detection": {
+            "anyOf": [
+              {
+                "$ref": "#/components/schemas/CollectionDetectionPatch"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "metadata": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Metadata"
+          },
+          "name": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "minLength": 1,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Name"
+          },
+          "save_face_crops": {
+            "anyOf": [
+              {
+                "type": "boolean"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Save Face Crops"
+          },
+          "search": {
+            "anyOf": [
+              {
+                "$ref": "#/components/schemas/CollectionSearchPatch"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          }
+        },
+        "title": "CollectionPatch",
+        "type": "object"
+      },
+      "CollectionResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "collection": {
+            "$ref": "#/components/schemas/Collection"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "collection"
+        ],
+        "title": "CollectionResponse",
+        "type": "object"
+      },
+      "CollectionSearchCreate": {
+        "additionalProperties": false,
+        "properties": {
+          "capacity_rows": {
+            "anyOf": [
+              {
+                "maximum": 9.223372036854776e+18,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Capacity Rows"
+          },
+          "load_policy": {
+            "anyOf": [
+              {
+                "enum": [
+                  "eager",
+                  "lazy"
+                ],
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Load Policy"
+          },
+          "max_faces_per_person": {
+            "anyOf": [
+              {
+                "maximum": 1000000.0,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Max Faces Per Person"
+          },
+          "profile": {
+            "anyOf": [
+              {
+                "enum": [
+                  "fp32_v1",
+                  "fp16_v1",
+                  "bf16_v1",
+                  "int8_x736_v1",
+                  "int8_x1000_v1"
+                ],
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Profile"
+          }
+        },
+        "title": "CollectionSearchCreate",
+        "type": "object"
+      },
+      "CollectionSearchPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "capacity_rows": {
+            "anyOf": [
+              {
+                "maximum": 9.223372036854776e+18,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Capacity Rows"
+          },
+          "load_policy": {
+            "anyOf": [
+              {
+                "enum": [
+                  "eager",
+                  "lazy"
+                ],
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Load Policy"
+          },
+          "max_faces_per_person": {
+            "anyOf": [
+              {
+                "maximum": 1000000.0,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Max Faces Per Person"
+          }
+        },
+        "title": "CollectionSearchPatch",
+        "type": "object"
+      },
+      "CompareResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "matched": {
+            "title": "Matched",
+            "type": "boolean"
+          },
+          "processing_ms": {
+            "title": "Processing Ms",
+            "type": "number"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "similarity": {
+            "title": "Similarity",
+            "type": "number"
+          },
+          "source_face": {
+            "$ref": "#/components/schemas/FaceObservation"
+          },
+          "target_face": {
+            "$ref": "#/components/schemas/FaceObservation"
+          },
+          "threshold": {
+            "title": "Threshold",
+            "type": "number"
+          }
+        },
+        "required": [
+          "request_id",
+          "matched",
+          "similarity",
+          "threshold",
+          "source_face",
+          "target_face",
+          "processing_ms"
+        ],
+        "title": "CompareResponse",
+        "type": "object"
+      },
+      "DetectResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "faces": {
+            "items": {
+              "$ref": "#/components/schemas/FaceObservation"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "processing_ms": {
+            "title": "Processing Ms",
+            "type": "number"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "faces",
+          "processing_ms"
+        ],
+        "title": "DetectResponse",
+        "type": "object"
+      },
+      "DetectionProfile": {
+        "additionalProperties": true,
+        "properties": {
+          "input_sizes": {
+            "items": {
+              "items": {
+                "type": "integer"
+              },
+              "type": "array"
+            },
+            "title": "Input Sizes",
+            "type": "array"
+          },
+          "nms_threshold": {
+            "title": "Nms Threshold",
+            "type": "number"
+          },
+          "single_face_selection": {
+            "enum": [
+              "largest",
+              "center_largest"
+            ],
+            "title": "Single Face Selection",
+            "type": "string"
+          },
+          "threshold": {
+            "title": "Threshold",
+            "type": "number"
+          }
+        },
+        "required": [
+          "input_sizes",
+          "threshold",
+          "nms_threshold",
+          "single_face_selection"
+        ],
+        "title": "DetectionProfile",
+        "type": "object"
+      },
+      "EmbeddingsResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "faces": {
+            "items": {
+              "$ref": "#/components/schemas/FaceObservation"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "model": {
+            "$ref": "#/components/schemas/ModelSummary"
+          },
+          "processing_ms": {
+            "title": "Processing Ms",
+            "type": "number"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "faces",
+          "model",
+          "processing_ms"
+        ],
+        "title": "EmbeddingsResponse",
+        "type": "object"
+      },
+      "ErrorBody": {
+        "additionalProperties": true,
+        "properties": {
+          "code": {
+            "title": "Code",
+            "type": "string"
+          },
+          "details": {
+            "additionalProperties": true,
+            "title": "Details",
+            "type": "object"
+          },
+          "message": {
+            "title": "Message",
+            "type": "string"
+          }
+        },
+        "required": [
+          "code",
+          "message",
+          "details"
+        ],
+        "title": "ErrorBody",
+        "type": "object"
+      },
+      "ErrorEnvelope": {
+        "properties": {
+          "error": {
+            "properties": {
+              "code": {
+                "type": "string"
+              },
+              "details": {
+                "type": "object"
+              },
+              "message": {
+                "type": "string"
+              }
+            },
+            "required": [
+              "code",
+              "message",
+              "details"
+            ],
+            "type": "object"
+          },
+          "request_id": {
+            "format": "uuid",
+            "type": "string"
+          }
+        },
+        "required": [
+          "error",
+          "request_id"
+        ],
+        "type": "object"
+      },
+      "FaceObservation": {
+        "additionalProperties": true,
+        "properties": {
+          "bbox": {
+            "$ref": "#/components/schemas/BoundingBox"
+          },
+          "detection_score": {
+            "title": "Detection Score",
+            "type": "number"
+          },
+          "embedding": {
+            "anyOf": [
+              {
+                "items": {
+                  "type": "number"
+                },
+                "type": "array"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Embedding"
+          },
+          "landmarks": {
+            "items": {
+              "items": {
+                "type": "number"
+              },
+              "type": "array"
+            },
+            "title": "Landmarks",
+            "type": "array"
+          },
+          "quality": {
+            "$ref": "#/components/schemas/FaceQuality"
+          }
+        },
+        "required": [
+          "bbox",
+          "landmarks",
+          "detection_score",
+          "quality"
+        ],
+        "title": "FaceObservation",
+        "type": "object"
+      },
+      "FacePageResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "faces": {
+            "items": {
+              "$ref": "#/components/schemas/FaceSample"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "next_cursor": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Next Cursor"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "faces",
+          "next_cursor"
+        ],
+        "title": "FacePageResponse",
+        "type": "object"
+      },
+      "FaceQuality": {
+        "additionalProperties": true,
+        "properties": {
+          "brightness": {
+            "title": "Brightness",
+            "type": "number"
+          },
+          "pose": {
+            "title": "Pose",
+            "type": "number"
+          },
+          "score": {
+            "title": "Score",
+            "type": "number"
+          },
+          "sharpness": {
+            "title": "Sharpness",
+            "type": "number"
+          }
+        },
+        "required": [
+          "score",
+          "sharpness",
+          "brightness",
+          "pose"
+        ],
+        "title": "FaceQuality",
+        "type": "object"
+      },
+      "FaceRegistrationResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "faces": {
+            "items": {
+              "$ref": "#/components/schemas/FaceSample"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "rejected_images": {
+            "items": {
+              "$ref": "#/components/schemas/RejectedImage"
+            },
+            "title": "Rejected Images",
+            "type": "array"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "faces",
+          "rejected_images"
+        ],
+        "title": "FaceRegistrationResponse",
+        "type": "object"
+      },
+      "FaceSample": {
+        "additionalProperties": true,
+        "properties": {
+          "bounding_box": {
+            "$ref": "#/components/schemas/BoundingBox"
+          },
+          "created_at": {
+            "title": "Created At",
+            "type": "string"
+          },
+          "detection_score": {
+            "title": "Detection Score",
+            "type": "number"
+          },
+          "embedding_contract_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Embedding Contract Id"
+          },
+          "embedding_source": {
+            "enum": [
+              "server",
+              "external_trusted"
+            ],
+            "title": "Embedding Source",
+            "type": "string"
+          },
+          "has_crop": {
+            "title": "Has Crop",
+            "type": "boolean"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "landmarks": {
+            "anyOf": [
+              {
+                "items": {
+                  "items": {
+                    "type": "number"
+                  },
+                  "type": "array"
+                },
+                "type": "array"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Landmarks"
+          },
+          "model_digest": {
+            "title": "Model Digest",
+            "type": "string"
+          },
+          "model_id": {
+            "title": "Model Id",
+            "type": "string"
+          },
+          "model_version": {
+            "title": "Model Version",
+            "type": "string"
+          },
+          "person_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Person Id"
+          },
+          "preprocessing_version": {
+            "title": "Preprocessing Version",
+            "type": "string"
+          },
+          "quality": {
+            "$ref": "#/components/schemas/FaceQuality"
+          }
+        },
+        "required": [
+          "id",
+          "bounding_box",
+          "landmarks",
+          "detection_score",
+          "quality",
+          "model_id",
+          "model_version",
+          "model_digest",
+          "preprocessing_version",
+          "embedding_source",
+          "embedding_contract_id",
+          "has_crop",
+          "created_at"
+        ],
+        "title": "FaceSample",
+        "type": "object"
+      },
+      "HTTPValidationError": {
+        "properties": {
+          "detail": {
+            "items": {
+              "$ref": "#/components/schemas/ValidationError"
+            },
+            "title": "Detail",
+            "type": "array"
+          }
+        },
+        "title": "HTTPValidationError",
+        "type": "object"
+      },
+      "HealthResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "auth_enabled": {
+            "title": "Auth Enabled",
+            "type": "boolean"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "status": {
+            "enum": [
+              "ready",
+              "not_ready"
+            ],
+            "title": "Status",
+            "type": "string"
+          },
+          "version": {
+            "title": "Version",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "status",
+          "version",
+          "auth_enabled"
+        ],
+        "title": "HealthResponse",
+        "type": "object"
+      },
+      "Match": {
+        "additionalProperties": true,
+        "properties": {
+          "matched_face_id": {
+            "title": "Matched Face Id",
+            "type": "string"
+          },
+          "person": {
+            "$ref": "#/components/schemas/Person"
+          },
+          "similarity": {
+            "title": "Similarity",
+            "type": "number"
+          }
+        },
+        "required": [
+          "person",
+          "similarity",
+          "matched_face_id"
+        ],
+        "title": "Match",
+        "type": "object"
+      },
+      "ModelComponent": {
+        "additionalProperties": true,
+        "properties": {
+          "embedding_dimension": {
+            "anyOf": [
+              {
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Embedding Dimension"
+          },
+          "file": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "File"
+          },
+          "input_size": {
+            "anyOf": [
+              {
+                "items": {
+                  "type": "integer"
+                },
+                "type": "array"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Input Size"
+          },
+          "model_id": {
+            "title": "Model Id",
+            "type": "string"
+          },
+          "model_version": {
+            "title": "Model Version",
+            "type": "string"
+          },
+          "preprocessing_version": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Preprocessing Version"
+          },
+          "sha256": {
+            "title": "Sha256",
+            "type": "string"
+          },
+          "task": {
+            "title": "Task",
+            "type": "string"
+          }
+        },
+        "required": [
+          "model_id",
+          "model_version",
+          "task",
+          "sha256"
+        ],
+        "title": "ModelComponent",
+        "type": "object"
+      },
+      "ModelSummary": {
+        "additionalProperties": true,
+        "properties": {
+          "embedding_dimension": {
+            "title": "Embedding Dimension",
+            "type": "integer"
+          },
+          "license": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "License"
+          },
+          "model_digest": {
+            "title": "Model Digest",
+            "type": "string"
+          },
+          "model_id": {
+            "title": "Model Id",
+            "type": "string"
+          },
+          "model_version": {
+            "title": "Model Version",
+            "type": "string"
+          },
+          "models": {
+            "items": {
+              "$ref": "#/components/schemas/ModelComponent"
+            },
+            "title": "Models",
+            "type": "array"
+          },
+          "preprocessing_version": {
+            "title": "Preprocessing Version",
+            "type": "string"
+          },
+          "provider": {
+            "title": "Provider",
+            "type": "string"
+          }
+        },
+        "required": [
+          "model_id",
+          "model_version",
+          "model_digest",
+          "embedding_dimension",
+          "preprocessing_version",
+          "provider",
+          "models",
+          "license"
+        ],
+        "title": "ModelSummary",
+        "type": "object"
+      },
+      "ModelsResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "execution_provider": {
+            "title": "Execution Provider",
+            "type": "string"
+          },
+          "license": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "License"
+          },
+          "models": {
+            "items": {
+              "$ref": "#/components/schemas/ModelComponent"
+            },
+            "title": "Models",
+            "type": "array"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "models",
+          "execution_provider",
+          "license"
+        ],
+        "title": "ModelsResponse",
+        "type": "object"
+      },
+      "Monitor": {
+        "additionalProperties": true,
+        "properties": {
+          "collection_id": {
+            "title": "Collection Id",
+            "type": "string"
+          },
+          "created_at": {
+            "title": "Created At",
+            "type": "string"
+          },
+          "description": {
+            "title": "Description",
+            "type": "string"
+          },
+          "enabled": {
+            "title": "Enabled",
+            "type": "boolean"
+          },
+          "event_buffer_size": {
+            "title": "Event Buffer Size",
+            "type": "integer"
+          },
+          "event_policy": {
+            "$ref": "#/components/schemas/MonitorEventPolicy-Output"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "inference_fps": {
+            "title": "Inference Fps",
+            "type": "number"
+          },
+          "match_threshold": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Match Threshold"
+          },
+          "name": {
+            "title": "Name",
+            "type": "string"
+          },
+          "preview_enabled": {
+            "title": "Preview Enabled",
+            "type": "boolean"
+          },
+          "runtime": {
+            "$ref": "#/components/schemas/MonitorRuntime"
+          },
+          "source": {
+            "$ref": "#/components/schemas/MonitorSource-Output"
+          },
+          "updated_at": {
+            "title": "Updated At",
+            "type": "string"
+          }
+        },
+        "required": [
+          "id",
+          "name",
+          "description",
+          "enabled",
+          "source",
+          "collection_id",
+          "inference_fps",
+          "match_threshold",
+          "event_buffer_size",
+          "event_policy",
+          "preview_enabled",
+          "runtime",
+          "created_at",
+          "updated_at"
+        ],
+        "title": "Monitor",
+        "type": "object"
+      },
+      "MonitorCreate": {
+        "additionalProperties": false,
+        "properties": {
+          "collection_id": {
+            "title": "Collection Id",
+            "type": "string"
+          },
+          "description": {
+            "default": "",
+            "maxLength": 2000,
+            "title": "Description",
+            "type": "string"
+          },
+          "enabled": {
+            "default": true,
+            "title": "Enabled",
+            "type": "boolean"
+          },
+          "event_buffer_size": {
+            "default": 1000,
+            "maximum": 10000.0,
+            "minimum": 10.0,
+            "title": "Event Buffer Size",
+            "type": "integer"
+          },
+          "event_policy": {
+            "$ref": "#/components/schemas/MonitorEventPolicy-Input"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "inference_fps": {
+            "default": 2.0,
+            "maximum": 30.0,
+            "minimum": 0.1,
+            "title": "Inference Fps",
+            "type": "number"
+          },
+          "match_threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Match Threshold"
+          },
+          "name": {
+            "maxLength": 200,
+            "minLength": 1,
+            "title": "Name",
+            "type": "string"
+          },
+          "preview_enabled": {
+            "default": false,
+            "title": "Preview Enabled",
+            "type": "boolean"
+          },
+          "source": {
+            "$ref": "#/components/schemas/MonitorSource-Input"
+          }
+        },
+        "required": [
+          "id",
+          "name",
+          "source",
+          "collection_id"
+        ],
+        "title": "MonitorCreate",
+        "type": "object"
+      },
+      "MonitorEvent": {
+        "additionalProperties": true,
+        "properties": {
+          "error": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Error"
+          },
+          "face": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Face"
+          },
+          "id": {
+            "format": "uuid",
+            "title": "Id",
+            "type": "string"
+          },
+          "matched_face_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Matched Face Id"
+          },
+          "monitor_id": {
+            "title": "Monitor Id",
+            "type": "string"
+          },
+          "occurred_at": {
+            "title": "Occurred At",
+            "type": "string"
+          },
+          "person": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Person"
+          },
+          "sequence": {
+            "title": "Sequence",
+            "type": "integer"
+          },
+          "similarity": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Similarity"
+          },
+          "stream_epoch": {
+            "title": "Stream Epoch",
+            "type": "string"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          },
+          "track_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Track Id"
+          },
+          "type": {
+            "title": "Type",
+            "type": "string"
+          }
+        },
+        "required": [
+          "id",
+          "sequence",
+          "type",
+          "monitor_id",
+          "stream_epoch",
+          "occurred_at"
+        ],
+        "title": "MonitorEvent",
+        "type": "object"
+      },
+      "MonitorEventPageResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "events": {
+            "items": {
+              "$ref": "#/components/schemas/MonitorEvent"
+            },
+            "title": "Events",
+            "type": "array"
+          },
+          "has_more": {
+            "title": "Has More",
+            "type": "boolean"
+          },
+          "latest_sequence": {
+            "title": "Latest Sequence",
+            "type": "integer"
+          },
+          "monitor_id": {
+            "title": "Monitor Id",
+            "type": "string"
+          },
+          "next_cursor": {
+            "title": "Next Cursor",
+            "type": "string"
+          },
+          "oldest_sequence": {
+            "anyOf": [
+              {
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Oldest Sequence"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "stream_epoch": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Stream Epoch"
+          },
+          "stream_reset": {
+            "title": "Stream Reset",
+            "type": "boolean"
+          },
+          "truncated": {
+            "title": "Truncated",
+            "type": "boolean"
+          }
+        },
+        "required": [
+          "request_id",
+          "monitor_id",
+          "stream_epoch",
+          "events",
+          "oldest_sequence",
+          "latest_sequence",
+          "has_more",
+          "truncated",
+          "stream_reset",
+          "next_cursor"
+        ],
+        "title": "MonitorEventPageResponse",
+        "type": "object"
+      },
+      "MonitorEventPolicy-Input": {
+        "additionalProperties": false,
+        "properties": {
+          "absence_timeout_seconds": {
+            "default": 3.0,
+            "maximum": 300.0,
+            "minimum": 0.1,
+            "title": "Absence Timeout Seconds",
+            "type": "number"
+          },
+          "confirm_frames": {
+            "default": 3,
+            "maximum": 100.0,
+            "minimum": 1.0,
+            "title": "Confirm Frames",
+            "type": "integer"
+          },
+          "cooldown_seconds": {
+            "default": 10.0,
+            "maximum": 3600.0,
+            "minimum": 0.0,
+            "title": "Cooldown Seconds",
+            "type": "number"
+          },
+          "emit_unknown": {
+            "default": true,
+            "title": "Emit Unknown",
+            "type": "boolean"
+          }
+        },
+        "title": "MonitorEventPolicy",
+        "type": "object"
+      },
+      "MonitorEventPolicy-Output": {
+        "additionalProperties": true,
+        "properties": {
+          "absence_timeout_seconds": {
+            "title": "Absence Timeout Seconds",
+            "type": "number"
+          },
+          "confirm_frames": {
+            "title": "Confirm Frames",
+            "type": "integer"
+          },
+          "cooldown_seconds": {
+            "title": "Cooldown Seconds",
+            "type": "number"
+          },
+          "emit_unknown": {
+            "title": "Emit Unknown",
+            "type": "boolean"
+          }
+        },
+        "required": [
+          "confirm_frames",
+          "absence_timeout_seconds",
+          "cooldown_seconds",
+          "emit_unknown"
+        ],
+        "title": "MonitorEventPolicy",
+        "type": "object"
+      },
+      "MonitorEventPolicyPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "absence_timeout_seconds": {
+            "anyOf": [
+              {
+                "maximum": 300.0,
+                "minimum": 0.1,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Absence Timeout Seconds"
+          },
+          "confirm_frames": {
+            "anyOf": [
+              {
+                "maximum": 100.0,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Confirm Frames"
+          },
+          "cooldown_seconds": {
+            "anyOf": [
+              {
+                "maximum": 3600.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Cooldown Seconds"
+          },
+          "emit_unknown": {
+            "anyOf": [
+              {
+                "type": "boolean"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Emit Unknown"
+          }
+        },
+        "title": "MonitorEventPolicyPatch",
+        "type": "object"
+      },
+      "MonitorInferenceState": {
+        "additionalProperties": true,
+        "properties": {
+          "actual_fps": {
+            "title": "Actual Fps",
+            "type": "number"
+          },
+          "capacity_limited": {
+            "title": "Capacity Limited",
+            "type": "boolean"
+          },
+          "configured_fps": {
+            "title": "Configured Fps",
+            "type": "number"
+          },
+          "decoded_frames": {
+            "title": "Decoded Frames",
+            "type": "integer"
+          },
+          "dropped_frames": {
+            "title": "Dropped Frames",
+            "type": "integer"
+          },
+          "frame_sequence": {
+            "title": "Frame Sequence",
+            "type": "integer"
+          },
+          "last_frame_at": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Last Frame At"
+          },
+          "last_inference_at": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Last Inference At"
+          },
+          "processed_frames": {
+            "title": "Processed Frames",
+            "type": "integer"
+          },
+          "processing_ms": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Processing Ms"
+          }
+        },
+        "required": [
+          "configured_fps",
+          "actual_fps",
+          "processing_ms",
+          "last_frame_at",
+          "last_inference_at",
+          "frame_sequence",
+          "decoded_frames",
+          "processed_frames",
+          "dropped_frames",
+          "capacity_limited"
+        ],
+        "title": "MonitorInferenceState",
+        "type": "object"
+      },
+      "MonitorPageResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "monitors": {
+            "items": {
+              "$ref": "#/components/schemas/Monitor"
+            },
+            "title": "Monitors",
+            "type": "array"
+          },
+          "next_cursor": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Next Cursor"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "monitors",
+          "next_cursor"
+        ],
+        "title": "MonitorPageResponse",
+        "type": "object"
+      },
+      "MonitorPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "collection_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Collection Id"
+          },
+          "description": {
+            "anyOf": [
+              {
+                "maxLength": 2000,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Description"
+          },
+          "enabled": {
+            "anyOf": [
+              {
+                "type": "boolean"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Enabled"
+          },
+          "event_buffer_size": {
+            "anyOf": [
+              {
+                "maximum": 10000.0,
+                "minimum": 10.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Event Buffer Size"
+          },
+          "event_policy": {
+            "anyOf": [
+              {
+                "$ref": "#/components/schemas/MonitorEventPolicyPatch"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "inference_fps": {
+            "anyOf": [
+              {
+                "maximum": 30.0,
+                "minimum": 0.1,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Inference Fps"
+          },
+          "match_threshold": {
+            "anyOf": [
+              {
+                "maximum": 1.0,
+                "minimum": 0.0,
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Match Threshold"
+          },
+          "name": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "minLength": 1,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Name"
+          },
+          "preview_enabled": {
+            "anyOf": [
+              {
+                "type": "boolean"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Preview Enabled"
+          },
+          "source": {
+            "anyOf": [
+              {
+                "$ref": "#/components/schemas/MonitorSource-Input"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          }
+        },
+        "title": "MonitorPatch",
+        "type": "object"
+      },
+      "MonitorPreviewState": {
+        "additionalProperties": true,
+        "properties": {
+          "active": {
+            "title": "Active",
+            "type": "boolean"
+          },
+          "enabled": {
+            "title": "Enabled",
+            "type": "boolean"
+          },
+          "frames": {
+            "title": "Frames",
+            "type": "integer"
+          },
+          "viewers": {
+            "title": "Viewers",
+            "type": "integer"
+          }
+        },
+        "required": [
+          "enabled",
+          "active",
+          "viewers",
+          "frames"
+        ],
+        "title": "MonitorPreviewState",
+        "type": "object"
+      },
+      "MonitorResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "monitor": {
+            "$ref": "#/components/schemas/Monitor"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "monitor"
+        ],
+        "title": "MonitorResponse",
+        "type": "object"
+      },
+      "MonitorRuntime": {
+        "additionalProperties": true,
+        "properties": {
+          "connected": {
+            "title": "Connected",
+            "type": "boolean"
+          },
+          "last_error": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Last Error"
+          },
+          "last_frame_at": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Last Frame At"
+          },
+          "preview_active": {
+            "title": "Preview Active",
+            "type": "boolean"
+          },
+          "preview_viewers": {
+            "title": "Preview Viewers",
+            "type": "integer"
+          },
+          "status": {
+            "title": "Status",
+            "type": "string"
+          },
+          "stream_epoch": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Stream Epoch"
+          }
+        },
+        "required": [
+          "status",
+          "connected",
+          "stream_epoch",
+          "last_frame_at",
+          "last_error",
+          "preview_active",
+          "preview_viewers"
+        ],
+        "title": "MonitorRuntime",
+        "type": "object"
+      },
+      "MonitorSource-Input": {
+        "additionalProperties": false,
+        "properties": {
+          "type": {
+            "const": "rtsp",
+            "default": "rtsp",
+            "title": "Type",
+            "type": "string"
+          },
+          "url": {
+            "maxLength": 2048,
+            "minLength": 8,
+            "title": "Url",
+            "type": "string"
+          }
+        },
+        "required": [
+          "url"
+        ],
+        "title": "MonitorSource",
+        "type": "object"
+      },
+      "MonitorSource-Output": {
+        "additionalProperties": true,
+        "properties": {
+          "type": {
+            "const": "rtsp",
+            "title": "Type",
+            "type": "string"
+          },
+          "url": {
+            "title": "Url",
+            "type": "string"
+          }
+        },
+        "required": [
+          "type",
+          "url"
+        ],
+        "title": "MonitorSource",
+        "type": "object"
+      },
+      "MonitorState": {
+        "additionalProperties": true,
+        "properties": {
+          "connected": {
+            "title": "Connected",
+            "type": "boolean"
+          },
+          "faces": {
+            "items": {
+              "additionalProperties": true,
+              "type": "object"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "inference": {
+            "$ref": "#/components/schemas/MonitorInferenceState"
+          },
+          "inference_errors": {
+            "title": "Inference Errors",
+            "type": "integer"
+          },
+          "last_error": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Last Error"
+          },
+          "matched_faces": {
+            "title": "Matched Faces",
+            "type": "integer"
+          },
+          "monitor_id": {
+            "title": "Monitor Id",
+            "type": "string"
+          },
+          "preview": {
+            "$ref": "#/components/schemas/MonitorPreviewState"
+          },
+          "reconnects": {
+            "title": "Reconnects",
+            "type": "integer"
+          },
+          "source": {
+            "$ref": "#/components/schemas/MonitorStateSource"
+          },
+          "started_at": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Started At"
+          },
+          "status": {
+            "title": "Status",
+            "type": "string"
+          },
+          "stream_epoch": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Stream Epoch"
+          },
+          "threshold": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Threshold"
+          },
+          "unknown_faces": {
+            "title": "Unknown Faces",
+            "type": "integer"
+          }
+        },
+        "required": [
+          "monitor_id",
+          "stream_epoch",
+          "status",
+          "connected",
+          "source",
+          "inference",
+          "threshold",
+          "faces",
+          "matched_faces",
+          "unknown_faces",
+          "preview",
+          "reconnects",
+          "inference_errors",
+          "last_error",
+          "started_at"
+        ],
+        "title": "MonitorState",
+        "type": "object"
+      },
+      "MonitorStateResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "state": {
+            "$ref": "#/components/schemas/MonitorState"
+          }
+        },
+        "required": [
+          "request_id",
+          "state"
+        ],
+        "title": "MonitorStateResponse",
+        "type": "object"
+      },
+      "MonitorStateSource": {
+        "additionalProperties": true,
+        "properties": {
+          "height": {
+            "anyOf": [
+              {
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Height"
+          },
+          "source_fps": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Source Fps"
+          },
+          "width": {
+            "anyOf": [
+              {
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Width"
+          }
+        },
+        "required": [
+          "width",
+          "height",
+          "source_fps"
+        ],
+        "title": "MonitorStateSource",
+        "type": "object"
+      },
+      "NormalizedBox": {
+        "additionalProperties": true,
+        "properties": {
+          "height": {
+            "title": "Height",
+            "type": "number"
+          },
+          "left": {
+            "title": "Left",
+            "type": "number"
+          },
+          "top": {
+            "title": "Top",
+            "type": "number"
+          },
+          "width": {
+            "title": "Width",
+            "type": "number"
+          }
+        },
+        "required": [
+          "left",
+          "top",
+          "width",
+          "height"
+        ],
+        "title": "NormalizedBox",
+        "type": "object"
+      },
+      "Person": {
+        "additionalProperties": true,
+        "properties": {
+          "created_at": {
+            "title": "Created At",
+            "type": "string"
+          },
+          "external_id": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "External Id"
+          },
+          "face_count": {
+            "title": "Face Count",
+            "type": "integer"
+          },
+          "id": {
+            "title": "Id",
+            "type": "string"
+          },
+          "metadata": {
+            "additionalProperties": true,
+            "title": "Metadata",
+            "type": "object"
+          },
+          "name": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Name"
+          },
+          "updated_at": {
+            "title": "Updated At",
+            "type": "string"
+          }
+        },
+        "required": [
+          "id",
+          "name",
+          "external_id",
+          "metadata",
+          "face_count",
+          "created_at",
+          "updated_at"
+        ],
+        "title": "Person",
+        "type": "object"
+      },
+      "PersonPageResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "next_cursor": {
+            "anyOf": [
+              {
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Next Cursor"
+          },
+          "persons": {
+            "items": {
+              "$ref": "#/components/schemas/Person"
+            },
+            "title": "Persons",
+            "type": "array"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "persons",
+          "next_cursor"
+        ],
+        "title": "PersonPageResponse",
+        "type": "object"
+      },
+      "PersonPatch": {
+        "additionalProperties": false,
+        "properties": {
+          "external_id": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "External Id"
+          },
+          "metadata": {
+            "anyOf": [
+              {
+                "additionalProperties": true,
+                "type": "object"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Metadata"
+          },
+          "name": {
+            "anyOf": [
+              {
+                "maxLength": 200,
+                "type": "string"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Name"
+          }
+        },
+        "title": "PersonPatch",
+        "type": "object"
+      },
+      "PersonRegistrationResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "faces": {
+            "items": {
+              "$ref": "#/components/schemas/FaceSample"
+            },
+            "title": "Faces",
+            "type": "array"
+          },
+          "person": {
+            "$ref": "#/components/schemas/Person"
+          },
+          "rejected_images": {
+            "items": {
+              "$ref": "#/components/schemas/RejectedImage"
+            },
+            "title": "Rejected Images",
+            "type": "array"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "person",
+          "faces",
+          "rejected_images"
+        ],
+        "title": "PersonRegistrationResponse",
+        "type": "object"
+      },
+      "PersonResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "person": {
+            "$ref": "#/components/schemas/Person"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          }
+        },
+        "required": [
+          "request_id",
+          "person"
+        ],
+        "title": "PersonResponse",
+        "type": "object"
+      },
+      "PixelBox": {
+        "additionalProperties": true,
+        "properties": {
+          "height": {
+            "title": "Height",
+            "type": "integer"
+          },
+          "width": {
+            "title": "Width",
+            "type": "integer"
+          },
+          "x": {
+            "title": "X",
+            "type": "integer"
+          },
+          "y": {
+            "title": "Y",
+            "type": "integer"
+          }
+        },
+        "required": [
+          "x",
+          "y",
+          "width",
+          "height"
+        ],
+        "title": "PixelBox",
+        "type": "object"
+      },
+      "RejectedImage": {
+        "additionalProperties": true,
+        "properties": {
+          "filename": {
+            "title": "Filename",
+            "type": "string"
+          },
+          "index": {
+            "title": "Index",
+            "type": "integer"
+          },
+          "reason": {
+            "title": "Reason",
+            "type": "string"
+          }
+        },
+        "required": [
+          "index",
+          "filename",
+          "reason"
+        ],
+        "title": "RejectedImage",
+        "type": "object"
+      },
+      "SearchResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "matches": {
+            "items": {
+              "$ref": "#/components/schemas/Match"
+            },
+            "title": "Matches",
+            "type": "array"
+          },
+          "processing_ms": {
+            "title": "Processing Ms",
+            "type": "number"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "searched_face": {
+            "$ref": "#/components/schemas/FaceObservation"
+          },
+          "threshold": {
+            "title": "Threshold",
+            "type": "number"
+          }
+        },
+        "required": [
+          "request_id",
+          "searched_face",
+          "matches",
+          "threshold",
+          "processing_ms"
+        ],
+        "title": "SearchResponse",
+        "type": "object"
+      },
+      "SystemResponse": {
+        "additionalProperties": true,
+        "properties": {
+          "api_key": {
+            "additionalProperties": {
+              "type": "boolean"
+            },
+            "title": "Api Key",
+            "type": "object"
+          },
+          "architecture": {
+            "title": "Architecture",
+            "type": "string"
+          },
+          "cpu": {
+            "additionalProperties": true,
+            "title": "Cpu",
+            "type": "object"
+          },
+          "data": {
+            "additionalProperties": true,
+            "title": "Data",
+            "type": "object"
+          },
+          "database": {
+            "additionalProperties": true,
+            "title": "Database",
+            "type": "object"
+          },
+          "execution_provider": {
+            "title": "Execution Provider",
+            "type": "string"
+          },
+          "model": {
+            "$ref": "#/components/schemas/ModelSummary"
+          },
+          "models": {
+            "additionalProperties": true,
+            "title": "Models",
+            "type": "object"
+          },
+          "os": {
+            "title": "Os",
+            "type": "string"
+          },
+          "recent_errors": {
+            "items": {
+              "additionalProperties": true,
+              "type": "object"
+            },
+            "title": "Recent Errors",
+            "type": "array"
+          },
+          "request_id": {
+            "format": "uuid",
+            "title": "Request Id",
+            "type": "string"
+          },
+          "runtime": {
+            "additionalProperties": true,
+            "title": "Runtime",
+            "type": "object"
+          },
+          "safe_config": {
+            "additionalProperties": true,
+            "title": "Safe Config",
+            "type": "object"
+          },
+          "search": {
+            "additionalProperties": true,
+            "title": "Search",
+            "type": "object"
+          },
+          "server_version": {
+            "title": "Server Version",
+            "type": "string"
+          },
+          "stats": {
+            "additionalProperties": {
+              "type": "integer"
+            },
+            "title": "Stats",
+            "type": "object"
+          }
+        },
+        "required": [
+          "request_id",
+          "server_version",
+          "os",
+          "architecture",
+          "cpu",
+          "runtime",
+          "execution_provider",
+          "model",
+          "database",
+          "data",
+          "models",
+          "stats",
+          "search",
+          "api_key",
+          "safe_config",
+          "recent_errors"
+        ],
+        "title": "SystemResponse",
+        "type": "object"
+      },
+      "ValidationError": {
+        "properties": {
+          "ctx": {
+            "title": "Context",
+            "type": "object"
+          },
+          "input": {
+            "title": "Input"
+          },
+          "loc": {
+            "items": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "integer"
+                }
+              ]
+            },
+            "title": "Location",
+            "type": "array"
+          },
+          "msg": {
+            "title": "Message",
+            "type": "string"
+          },
+          "type": {
+            "title": "Error Type",
+            "type": "string"
+          }
+        },
+        "required": [
+          "loc",
+          "msg",
+          "type"
+        ],
+        "title": "ValidationError",
+        "type": "object"
+      }
+    },
+    "securitySchemes": {
+      "bearerAuth": {
+        "scheme": "bearer",
+        "type": "http"
+      }
+    }
+  },
+  "info": {
+    "description": "Simple self-hosted face detection, comparison, registration, and search.",
+    "title": "InsightFace Server",
+    "version": "0.2.0"
+  },
+  "openapi": "3.1.0",
+  "paths": {
+    "/v1/collections": {
+      "get": {
+        "operationId": "list_collections_v1_collections_get",
+        "parameters": [
+          {
+            "in": "query",
+            "name": "limit",
+            "required": false,
+            "schema": {
+              "default": 50,
+              "maximum": 100,
+              "minimum": 1,
+              "title": "Limit",
+              "type": "integer"
+            }
+          },
+          {
+            "in": "query",
+            "name": "cursor",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Cursor"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/CollectionPageResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "List Collections",
+        "tags": [
+          "collections"
+        ]
+      },
+      "post": {
+        "operationId": "create_collection_v1_collections_post",
+        "requestBody": {
+          "content": {
+            "application/json": {
+              "schema": {
+                "$ref": "#/components/schemas/CollectionCreate"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "201": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/CollectionResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Create Collection",
+        "tags": [
+          "collections"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}": {
+      "delete": {
+        "operationId": "delete_collection_v1_collections__collection_id__delete",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "query",
+            "name": "force",
+            "required": false,
+            "schema": {
+              "default": false,
+              "title": "Force",
+              "type": "boolean"
+            }
+          }
+        ],
+        "responses": {
+          "204": {
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Delete Collection",
+        "tags": [
+          "collections"
+        ]
+      },
+      "get": {
+        "operationId": "get_collection_v1_collections__collection_id__get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/CollectionResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Get Collection",
+        "tags": [
+          "collections"
+        ]
+      },
+      "patch": {
+        "operationId": "patch_collection_v1_collections__collection_id__patch",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "application/json": {
+              "schema": {
+                "$ref": "#/components/schemas/CollectionPatch"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/CollectionResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Patch Collection",
+        "tags": [
+          "collections"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/persons": {
+      "get": {
+        "operationId": "list_persons_v1_collections__collection_id__persons_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "query",
+            "name": "limit",
+            "required": false,
+            "schema": {
+              "default": 50,
+              "maximum": 100,
+              "minimum": 1,
+              "title": "Limit",
+              "type": "integer"
+            }
+          },
+          {
+            "in": "query",
+            "name": "cursor",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Cursor"
+            }
+          },
+          {
+            "in": "query",
+            "name": "search",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "maxLength": 200,
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Search"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/PersonPageResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "List Persons",
+        "tags": [
+          "persons"
+        ]
+      },
+      "post": {
+        "operationId": "create_person_v1_collections__collection_id__persons_post",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_create_person_v1_collections__collection_id__persons_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "201": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/PersonRegistrationResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Create Person",
+        "tags": [
+          "persons"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/persons/{person_id}": {
+      "delete": {
+        "operationId": "delete_person_v1_collections__collection_id__persons__person_id__delete",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "204": {
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Delete Person",
+        "tags": [
+          "persons"
+        ]
+      },
+      "get": {
+        "operationId": "get_person_v1_collections__collection_id__persons__person_id__get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/PersonResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Get Person",
+        "tags": [
+          "persons"
+        ]
+      },
+      "patch": {
+        "operationId": "patch_person_v1_collections__collection_id__persons__person_id__patch",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "application/json": {
+              "schema": {
+                "$ref": "#/components/schemas/PersonPatch"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/PersonResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Patch Person",
+        "tags": [
+          "persons"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/persons/{person_id}/faces": {
+      "get": {
+        "operationId": "list_faces_v1_collections__collection_id__persons__person_id__faces_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "query",
+            "name": "limit",
+            "required": false,
+            "schema": {
+              "default": 50,
+              "maximum": 100,
+              "minimum": 1,
+              "title": "Limit",
+              "type": "integer"
+            }
+          },
+          {
+            "in": "query",
+            "name": "cursor",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Cursor"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/FacePageResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "List Faces",
+        "tags": [
+          "persons"
+        ]
+      },
+      "post": {
+        "operationId": "add_faces_v1_collections__collection_id__persons__person_id__faces_post",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_add_faces_v1_collections__collection_id__persons__person_id__faces_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "201": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/FaceRegistrationResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Add Faces",
+        "tags": [
+          "persons"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}": {
+      "delete": {
+        "operationId": "delete_face_v1_collections__collection_id__persons__person_id__faces__face_id__delete",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "face_id",
+            "required": true,
+            "schema": {
+              "title": "Face Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "204": {
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Delete Face",
+        "tags": [
+          "persons"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image": {
+      "get": {
+        "operationId": "get_face_image_v1_collections__collection_id__persons__person_id__faces__face_id__image_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "person_id",
+            "required": true,
+            "schema": {
+              "title": "Person Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "path",
+            "name": "face_id",
+            "required": true,
+            "schema": {
+              "title": "Face Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "image/jpeg": {
+                "schema": {
+                  "format": "binary",
+                  "type": "string"
+                }
+              }
+            },
+            "description": "Saved 112 x 112 face crop"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Get Face Image",
+        "tags": [
+          "persons"
+        ]
+      }
+    },
+    "/v1/collections/{collection_id}/search": {
+      "post": {
+        "operationId": "search_v1_collections__collection_id__search_post",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "collection_id",
+            "required": true,
+            "schema": {
+              "title": "Collection Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_search_v1_collections__collection_id__search_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/SearchResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Search",
+        "tags": [
+          "search"
+        ]
+      }
+    },
+    "/v1/compare": {
+      "post": {
+        "operationId": "compare_v1_compare_post",
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_compare_v1_compare_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/CompareResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Compare",
+        "tags": [
+          "faces"
+        ]
+      }
+    },
+    "/v1/detect": {
+      "post": {
+        "operationId": "detect_v1_detect_post",
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_detect_v1_detect_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/DetectResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Detect",
+        "tags": [
+          "faces"
+        ]
+      }
+    },
+    "/v1/embeddings": {
+      "post": {
+        "operationId": "embeddings_v1_embeddings_post",
+        "requestBody": {
+          "content": {
+            "multipart/form-data": {
+              "schema": {
+                "$ref": "#/components/schemas/Body_embeddings_v1_embeddings_post"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/EmbeddingsResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Embeddings",
+        "tags": [
+          "faces"
+        ]
+      }
+    },
+    "/v1/health": {
+      "get": {
+        "operationId": "health_v1_health_get",
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/HealthResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/HealthResponse"
+                }
+              }
+            },
+            "description": "The process is running but is not ready."
+          }
+        },
+        "summary": "Health",
+        "tags": [
+          "system"
+        ]
+      }
+    },
+    "/v1/models": {
+      "get": {
+        "operationId": "models_v1_models_get",
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ModelsResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Models",
+        "tags": [
+          "system"
+        ]
+      }
+    },
+    "/v1/monitors": {
+      "get": {
+        "operationId": "list_monitors_v1_monitors_get",
+        "parameters": [
+          {
+            "in": "query",
+            "name": "limit",
+            "required": false,
+            "schema": {
+              "default": 50,
+              "maximum": 100,
+              "minimum": 1,
+              "title": "Limit",
+              "type": "integer"
+            }
+          },
+          {
+            "in": "query",
+            "name": "cursor",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Cursor"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorPageResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "List Monitors",
+        "tags": [
+          "monitors"
+        ]
+      },
+      "post": {
+        "operationId": "create_monitor_v1_monitors_post",
+        "requestBody": {
+          "content": {
+            "application/json": {
+              "schema": {
+                "$ref": "#/components/schemas/MonitorCreate"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "201": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "429": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "The enabled Monitor capacity is exhausted."
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Create Monitor",
+        "tags": [
+          "monitors"
+        ]
+      }
+    },
+    "/v1/monitors/{monitor_id}": {
+      "delete": {
+        "operationId": "delete_monitor_v1_monitors__monitor_id__delete",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "204": {
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Delete Monitor",
+        "tags": [
+          "monitors"
+        ]
+      },
+      "get": {
+        "operationId": "get_monitor_v1_monitors__monitor_id__get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Get Monitor",
+        "tags": [
+          "monitors"
+        ]
+      },
+      "patch": {
+        "operationId": "patch_monitor_v1_monitors__monitor_id__patch",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          }
+        ],
+        "requestBody": {
+          "content": {
+            "application/json": {
+              "schema": {
+                "$ref": "#/components/schemas/MonitorPatch"
+              }
+            }
+          },
+          "required": true
+        },
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "429": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "The enabled Monitor capacity is exhausted."
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Patch Monitor",
+        "tags": [
+          "monitors"
+        ]
+      }
+    },
+    "/v1/monitors/{monitor_id}/events": {
+      "get": {
+        "operationId": "monitor_events_v1_monitors__monitor_id__events_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          },
+          {
+            "in": "query",
+            "name": "cursor",
+            "required": false,
+            "schema": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Cursor"
+            }
+          },
+          {
+            "in": "query",
+            "name": "limit",
+            "required": false,
+            "schema": {
+              "default": 100,
+              "maximum": 1000,
+              "minimum": 1,
+              "title": "Limit",
+              "type": "integer"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorEventPageResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Monitor Events",
+        "tags": [
+          "monitors"
+        ]
+      }
+    },
+    "/v1/monitors/{monitor_id}/preview.mjpeg": {
+      "get": {
+        "operationId": "monitor_preview_v1_monitors__monitor_id__preview_mjpeg_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "multipart/x-mixed-replace": {
+                "schema": {
+                  "format": "binary",
+                  "type": "string"
+                }
+              }
+            },
+            "description": "Live Motion JPEG preview stream."
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Monitor Preview",
+        "tags": [
+          "monitors"
+        ]
+      }
+    },
+    "/v1/monitors/{monitor_id}/state": {
+      "get": {
+        "operationId": "monitor_state_v1_monitors__monitor_id__state_get",
+        "parameters": [
+          {
+            "in": "path",
+            "name": "monitor_id",
+            "required": true,
+            "schema": {
+              "title": "Monitor Id",
+              "type": "string"
+            }
+          }
+        ],
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/MonitorStateResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "Monitor State",
+        "tags": [
+          "monitors"
+        ]
+      }
+    },
+    "/v1/system": {
+      "get": {
+        "operationId": "system_v1_system_get",
+        "responses": {
+          "200": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/SystemResponse"
+                }
+              }
+            },
+            "description": "Successful Response"
+          },
+          "400": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "401": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "404": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "409": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "413": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "422": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "500": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          },
+          "503": {
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ErrorEnvelope"
+                }
+              }
+            },
+            "description": "Error"
+          }
+        },
+        "security": [
+          {
+            "bearerAuth": []
+          }
+        ],
+        "summary": "System",
+        "tags": [
+          "system"
+        ]
+      }
+    }
+  }
+}
diff --git a/server/docs/user-guide.de.md b/server/docs/user-guide.de.md
new file mode 100644
index 000000000..f9fbb0812
--- /dev/null
+++ b/server/docs/user-guide.de.md
@@ -0,0 +1,146 @@
+# InsightFace Server Benutzerhandbuch
+
+**Sprachen:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · Deutsch · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+Dieses Handbuch führt Erstanwender vom leeren Checkout bis zur ersten erfolgreichen Personensuche. Dieselben Funktionen stehen über Web UI, `/v1` und Python SDK bereit. Alle HTTP-Felder und Ergebnisse beschreibt der [API-Leitfaden](api.de.md).
+
+## Vom Start bis zur ersten Suche
+
+CPU benötigt Linux x86_64, Docker Engine und Docker Compose. CUDA benötigt zusätzlich einen kompatiblen NVIDIA-Treiber und NVIDIA Container Toolkit; CUDA, cuDNN, ORT, Python und OpenCV müssen nicht auf dem Host installiert sein.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Für GPU verwenden Sie `compose.cuda12.yml` und Port `18098`. Vor dem Download erscheint die Modelllizenz. Öffentliche InsightFace-Modelle sind ohne separate kommerzielle Lizenz nur für nichtkommerzielle Forschung bestimmt.
+
+Die mitgelieferte Compose-Konfiguration deaktiviert Authentifizierung standardmäßig für isolierte Tests. Für andere Benutzer oder Netze setzen Sie vor dem Start `INSIGHTFACE_AUTH_ENABLED=true` und einen langen `INSIGHTFACE_API_KEY`. Danach: Dashboard prüfen, Collection erstellen, Person registrieren und mit einem anderen Bild suchen. Stoppen Sie mit `docker compose ... down` ohne `-v`, damit das Datenvolume erhalten bleibt.
+
+## 1. Anmelden und Bereitschaft prüfen
+
+Öffnen Sie für CPU `http://SERVER:18097/` oder für CUDA 12 `http://SERVER:18098/`. Wenn Authentifizierung aktiv ist, tragen Sie unter **API-Schlüssel konfigurieren** den vom Betreiber erhaltenen Key ein. Er verbleibt nur im Speicher des Tabs und wird beim Neuladen oder Schließen gelöscht.
+
+Prüfen Sie unter **Übersicht** oder **System**, dass Dienst, Datenbank, Modelle und Provider bereit sind. Eine CUDA-Instanz muss `CUDAExecutionProvider` melden und fällt nicht still auf CPU zurück.
+
+## 2. Collection anlegen
+
+Unter **Sammlungen** → **Neue Sammlung** setzen Sie eine stabile ID, Name,
+Standard-Cosinus-Schwelle (`0.4`), ein verfügbares Suchprofil, Kapazität und
+maximale FaceSamples pro Person. Die Speicherung eines auf 112×112 skalierten
+`bounding-box crop` als JPEG ist standardmäßig aus; es ist nicht der
+ausgerichtete Erkennungseingang.
+
+Eine Collection ist an Modell-ID, Version, Digest, Dimension und Vorverarbeitung gebunden. Nach einem Modellwechsel bleibt sie sichtbar, aber Registrierung und Suche werden bei abweichendem Vertrag ausdrücklich abgelehnt.
+
+Das Erkennungsprofil kopiert beim Anlegen die Systemwerte und kann später für Eingabegrößen, Erkennungs-/NMS-Schwelle und Ein-Gesicht-Strategie geändert werden. `largest` priorisiert die Fläche; `center_largest` maximiert `Fläche - 2,0 × quadrierter Pixelabstand zwischen Box- und Bildmitte`. Die Erkennungskonfidenz gehört nicht zu diesem Wert.
+
+## 3. Person registrieren
+
+Wählen Sie unter **Personen** eine Collection und **Person registrieren**. Geben Sie optional ID, Name, externe ID, JSON-Metadaten und ein oder mehrere JPEG-, PNG- oder WebP-Bilder an.
+
+- `off`: verwendet die Ein-Gesicht-Strategie der Collection; mehrere Gesichter sind erlaubt.
+- `standard`: genau ein nutzbares Gesicht sowie Prüfungen von Größe, Erkennungswert, Schärfe, Helligkeit und Pose.
+- `strict`: zusätzlich muss die beste Ähnlichkeit innerhalb der Person höher sein als die beste Ähnlichkeit zu anderen Personen.
+
+Stapelregistrierung kann teilweise erfolgreich sein und meldet den Ablehnungsgrund pro Bild. Originale werden nicht gespeichert. `external_trusted` akzeptiert ein L2-normalisiertes Embedding; das Bild bleibt für Erkennung und Qualitätsprüfung erforderlich, das Embedding wird aber nicht erneut extrahiert.
+
+## 4. Erkennen, vergleichen und suchen
+
+**Erkennen** zeigt Boxen, fünf Landmarken, Erkennungswert und Qualität; kein Gesicht ist eine erfolgreiche leere Liste. **Vergleichen** wählt mit dem System- oder Collection-Profil je ein Gesicht und liefert Cosinus-`similarity`, `threshold` und `matched`. Ähnlichkeit ist keine Wahrscheinlichkeit.
+
+Unter **Suchen** wählen Sie Collection und Bild. Der Person-Score ist die höchste Ähnlichkeit aller FaceSamples dieser Person. Ergebnisse sind absteigend sortiert; kein Treffer ist eine leere Liste. Neue Samples werden zuerst in SQLite bestätigt und vor der Erfolgsantwort in den Speicherindex eingefügt. Beim Neustart wird der Index aus SQLite neu aufgebaut.
+
+## 5. RTSP-Kameraüberwachung
+
+Erstellen Sie unter **Kameraüberwachung** einen dauerhaften Monitor und konfigurieren Sie RTSP-Quelle, Collection, Inferenzrate, optionalen Schwellwert und Ereignisregeln. Die Vorschau ist standardmäßig aus; Erkennung und Ereignisse laufen trotzdem. Bei aktivierter Vorschau zeichnet die Web UI grüne registrierte und orange unbekannte Gesichter aus den `/state`-Daten über rohe Bilder.
+
+Der Monitor läuft unabhängig vom Browser; aktivierte Aufgaben werden nach Server-Neustart wiederhergestellt. Einstellungen liegen in SQLite und RTSP-Zugangsdaten verschlüsselt unter `/data`, Videobilder und Ereignisse werden jedoch nicht gespeichert. Ereignisse bleiben nur im begrenzten RAM-Puffer. Der Decoder behält den neuesten Frame und überspringt veraltete Frames statt sie aufzustauen.
+
+## 6. Daten, Backup und Sicherheit
+
+Persistieren Sie `/data` und mounten Sie `/models` schreibgeschützt. Sichern Sie SQLite und optional gespeicherte Gesichtsbilder gemeinsam vor Massenänderungen. API Keys werden gehasht; ein neuer `INSIGHTFACE_API_KEY` beim Start mit demselben Volume rotiert den aktiven Key. Bilder, Embeddings und Keys gehören nicht in Logs.
+
+Der OpenAPI-Schema-Explorer für Entwickler liegt unter `/docs`; aufgabenbezogene API-Anleitungen stehen in dieser Hilfe. Nennen Sie bei Fehlern `x-request-id`. Prüfen Sie bei `401` den Key, bei `409 collection_model_mismatch` den Modellvertrag und bei `422 face_not_found` das Eingabebild.
+
+## 7. Modelle und Lizenzen
+
+Die Images enthalten keine Modelle. Der normale Start bleibt offline; der
+einmalige Dienst `models` installiert nach `server/.models`:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+Unterstützt sind `buffalo_l` (`det_10g.onnx` + `w600k_r50.onnx`),
+`buffalo_m`, `buffalo_sc` und `antelopev2`. Die Installation erzeugt
+`manifest.json` und die signierte `MODEL.LICENSE`. Ohne `--accept-license`
+zeigt das Werkzeug die Bedingungen und lädt nichts herunter. Öffentliche
+vortrainierte InsightFace-Modelle sind ohne separate kommerzielle Lizenz nur
+für nichtkommerzielle Forschung bestimmt.
+
+## 8. Startkonfiguration und Suche
+
+`server/config/server.toml` wird einmal beim Start gelesen; Änderungen benötigen
+einen Neustart. Standardwerte sind `input_sizes=[[96,96],[512,512]]`,
+Detektionsschwelle `0.50`, NMS `0.40`, `single_face_selection="largest"` und
+höchstens 100 Gesichter. SCRFD führt jede Auflösung aus, projiziert alle
+Kandidaten in das Originalbild und wendet einmal globales NMS an.
+`max_concurrency="auto"` bedeutet CPU 4 und CUDA 8.
+`[web].disabled=true` lässt nur `/v1` und `/openapi.json` aktiv.
+
+System zeigt nur verfügbare Suchprofile. Ein Profil ist nach Erstellung einer
+Collection fest und kann nicht pro Search gewechselt werden:
+
+- `fp32_v1`: Standard für CPU/CUDA;
+- `fp16_v1`: CUDA;
+- `bf16_v1`: unterstützte CPU oder SM80+ CUDA;
+- `int8_x736_v1`: empfohlenes INT8 für CPU/CUDA, INT32-Akkumulation;
+- `int8_x1000_v1`: Kompatibilität bestehender Collections.
+
+Alle Profile durchsuchen jeden FaceSample vollständig und sind keine
+ANN-Indizes; öffentliche Werte bleiben raw cosine. `capacity_rows` ist
+standardmäßig `100000`, der Guardrail `10000000`,
+`max_faces_per_person=20`. Bei 512 Dimensionen benötigt nur der Vektor ungefähr
+2.048 Byte FP32, 1.024 Byte FP16/BF16 oder 512 Byte INT8 pro Zeile.
+
+## 9. SDK, eigener Build und Datenbetrieb
+
+Das Python SDK akzeptiert Pfade, bytes und file-like objects und bietet typisierte
+Methoden für Detect, Compare, Collection, Registrierung, Search und Monitor.
+Den vollständigen HTTP-Vertrag beschreibt der [API-Leitfaden](api.de.md).
+
+Aus einem vollständigen Repository kann jeder die Images bauen:
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Für lokale Images verwenden Compose-Befehle `--pull never`. Unveränderliche
+Tags sind `0.2.0-cpu` und `0.2.0-cuda12`; `cpu` und `cuda12` zeigen auf die
+jeweils neueste stabile Variante, ein `latest` gibt es absichtlich nicht.
+Vor einem Upgrade Schreibzugriffe stoppen und `/data` samt Crop-Speicher
+SQLite-sicher sichern. `docker compose down -v` nicht verwenden, weil es das
+Datenvolume löscht.
+
+## 10. GPU, Netzwerk und Fehlerbehebung
+
+Das CUDA-Image enthält CUDA Runtime 12.9.1, cuDNN 9.24.0 und
+`onnxruntime-gpu==1.27.0`. Turing/Ampere/Ada/Hopper benötigen mindestens R535,
+Blackwell/RTX 50 mindestens 570.26; für neue Installationen wird ein stabiler
+R580 oder neuer empfohlen. Der Start prüft GPU, Compute Capability, Driver,
+CUDA/cuDNN/ORT, Provider, echte Modell-Sessions und Warm-up und verweigert
+stillen CPU-Fallback.
+
+Bei Netzzugriff HTTPS an einem vertrauenswürdigen Reverse Proxy terminieren,
+CORS-Ursprünge sowie Rate/Body/Timeout begrenzen und `/data` und Backups als
+biometrische Daten schützen. Bilder, Embeddings und Keys nie protokollieren.
+Phase eins besitzt nur einen undifferenzierten API Key und ist kein
+Mandanten-Berechtigungssystem.
diff --git a/server/docs/user-guide.es.md b/server/docs/user-guide.es.md
new file mode 100644
index 000000000..e8be41886
--- /dev/null
+++ b/server/docs/user-guide.es.md
@@ -0,0 +1,142 @@
+# Guía de usuario de InsightFace Server
+
+**Idiomas:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · Español · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+Esta guía lleva a un usuario nuevo desde un directorio vacío hasta la primera búsqueda correcta. Las mismas funciones están disponibles en la Web UI, `/v1` y el SDK Python. Consulte todos los campos y resultados HTTP en la [guía de API](api.es.md).
+
+## Desde cero hasta la primera búsqueda
+
+CPU requiere Linux x86_64, Docker Engine y Docker Compose. CUDA añade un Driver NVIDIA compatible y NVIDIA Container Toolkit; no instale CUDA, cuDNN, ORT, Python ni OpenCV en el host.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Para GPU use `compose.cuda12.yml` y el puerto `18098`. El instalador muestra la licencia antes de descargar; los modelos públicos de InsightFace son solo para investigación no comercial salvo licencia comercial independiente.
+
+El Compose incluido desactiva la autenticación por defecto para evaluación aislada. Antes de exponer el servicio, defina `INSIGHTFACE_AUTH_ENABLED=true` y un `INSIGHTFACE_API_KEY` largo. Después compruebe el Panel, cree una Collection, registre una Person y busque con otra foto. Detenga con `docker compose ... down` sin `-v` para conservar el volumen.
+
+## 1. Acceso y estado
+
+Abra `http://SERVIDOR:18097/` para CPU o `http://SERVIDOR:18098/` para CUDA 12. Si hay autenticación, use **Configurar clave API**, pegue la clave del operador y aplíquela a la pestaña. Solo permanece en memoria y se elimina al recargar o cerrar.
+
+Compruebe en **Panel** o **Sistema** que servicio, base de datos, modelos y Provider estén listos. CUDA debe mostrar `CUDAExecutionProvider` y nunca vuelve silenciosamente a CPU.
+
+## 2. Crear una Collection
+
+En **Colecciones** → **Nueva colección**, indique un ID estable, nombre, umbral
+coseno (`0.4` inicialmente), perfil disponible, capacidad y máximo de
+FaceSamples por persona. Guardar como JPEG un `bounding-box crop` redimensionado
+a 112×112 está desactivado por defecto; no es la entrada alineada de
+reconocimiento.
+
+La Collection queda fijada al ID, versión, digest, dimensión y preprocesamiento del modelo. Tras cambiar el modelo, una colección antigua sigue visible, pero su registro y búsqueda se rechazan si el contrato no coincide.
+
+El perfil de detección copia los valores del sistema al crear la Collection y después permite cambiar tamaños de entrada, umbrales de detección/NMS y estrategia de un rostro. `largest` prioriza el área; `center_largest` maximiza `área - 2,0 × distancia en píxeles al cuadrado entre el centro del cuadro y el de la imagen`. La confianza de detección no participa en esta puntuación.
+
+## 3. Registrar una Person
+
+En **Personas**, seleccione la Collection y **Registrar persona**. Puede indicar ID, nombre, ID externo, metadata JSON y una o varias imágenes JPEG, PNG o WebP.
+
+- `off`: usa la estrategia de un rostro de la Collection y permite varios rostros.
+- `standard`: exige un rostro utilizable y valida tamaño, detección, nitidez, iluminación y pose.
+- `strict`: además exige que la mejor similitud interna sea mayor que la mejor similitud con otra persona.
+
+El lote permite éxito parcial y explica cada rechazo. No se guardan originales. `external_trusted` acepta un embedding normalizado L2; la imagen sigue siendo obligatoria para detección y calidad, pero no se vuelve a extraer el vector.
+
+## 4. Detectar, comparar y buscar
+
+**Detectar** muestra cajas, cinco puntos, detección y calidad; sin rostros devuelve una lista vacía correcta. **Comparar** usa el perfil del sistema o de una Collection para elegir un rostro por imagen y devuelve `similarity` coseno, `threshold` y `matched`. La similitud no es probabilidad.
+
+En **Buscar**, seleccione Collection e imagen. La puntuación de una persona es la mayor similitud entre sus FaceSamples. Los resultados se ordenan de mayor a menor; sin coincidencia es una lista vacía. Cada muestra se confirma primero en SQLite y se añade al índice antes de responder. Al reiniciar, el índice se reconstruye desde SQLite.
+
+## 5. Monitorización de cámara RTSP
+
+En **Monitorización de cámaras**, cree un Monitor persistente y configure fuente RTSP, Collection, frecuencia, umbral opcional y política de eventos. La vista previa está desactivada por defecto; reconocimiento y eventos continúan sin ella. Al activarla, la Web UI dibuja cajas verdes para personas registradas y naranjas para rostros desconocidos usando `/state` sobre imágenes crudas.
+
+El Monitor funciona independientemente del navegador y las tareas activas se restauran tras reiniciar el servidor. La configuración vive en SQLite y las credenciales RTSP cifradas en `/data`; no se guardan fotogramas ni eventos. Los eventos solo permanecen en un búfer de memoria limitado. El decodificador conserva el último fotograma y omite los obsoletos en lugar de acumularlos.
+
+## 6. Datos y seguridad
+
+Conserve `/data` y monte `/models` como solo lectura. Antes de operaciones masivas, copie SQLite y los recortes juntos. Las claves se guardan como hash; iniciar el mismo volumen con otro `INSIGHTFACE_API_KEY` rota la clave activa. No registre imágenes, embeddings ni claves.
+
+El explorador de esquemas OpenAPI para desarrolladores está en `/docs`; las instrucciones prácticas de la API están en esta ayuda. Incluya `x-request-id` al comunicar incidencias. `401` indica clave, `409 collection_model_mismatch` contrato de modelo y `422 face_not_found` ausencia de rostro válido.
+
+## 7. Modelos y licencias
+
+Las imágenes no incluyen modelos. El inicio normal permanece sin conexión; el
+servicio puntual `models` instala en `server/.models`:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+Se admiten `buffalo_l` (`det_10g.onnx` + `w600k_r50.onnx`), `buffalo_m`,
+`buffalo_sc` y `antelopev2`. La instalación crea `manifest.json` y
+`MODEL.LICENSE` firmada. Sin `--accept-license`, la herramienta muestra los
+términos y termina sin descargar. Los modelos públicos preentrenados de
+InsightFace son solo para investigación no comercial salvo licencia comercial
+independiente.
+
+## 8. Configuración de inicio y búsqueda
+
+`server/config/server.toml` se lee una vez al arrancar; los cambios requieren
+reiniciar. Los valores son `input_sizes=[[96,96],[512,512]]`, umbral de
+detección `0.50`, NMS `0.40`, `single_face_selection="largest"` y máximo 100
+rostros. SCRFD ejecuta cada resolución, lleva los candidatos a la imagen
+original y hace un único NMS global. `max_concurrency="auto"` equivale a CPU 4
+y CUDA 8. `[web].disabled=true` conserva solo `/v1` y `/openapi.json`.
+
+System anuncia únicamente los perfiles disponibles. El perfil se fija al crear
+la Collection y no puede cambiarse por petición:
+
+- `fp32_v1`: CPU/CUDA estándar;
+- `fp16_v1`: CUDA;
+- `bf16_v1`: CPU compatible o CUDA SM80+;
+- `int8_x736_v1`: INT8 recomendado en CPU/CUDA, acumulación INT32;
+- `int8_x1000_v1`: compatibilidad de Collections existentes.
+
+Todos recorren cada FaceSample y no son índices ANN; la salida sigue siendo raw
+cosine. `capacity_rows` vale `100000`, el límite global `10000000` y
+`max_faces_per_person=20`. Para 512 dimensiones, solo el vector ocupa unos
+2.048 bytes FP32, 1.024 FP16/BF16 o 512 INT8 por fila.
+
+## 9. SDK, compilación y operación de datos
+
+El SDK Python admite ruta, bytes y objeto tipo archivo, con métodos tipados para
+Detect, Compare, Collections, registro, Search y Monitors. Consulte el contrato
+HTTP en la [guía de API](api.es.md).
+
+El usuario puede compilar desde un checkout completo:
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Use `--pull never` con Compose para usar la imagen local. Los tags inmutables
+son `0.2.0-cpu` y `0.2.0-cuda12`; `cpu` y `cuda12` apuntan a la última estable
+y no existe `latest`. Antes de actualizar, pare las escrituras y haga una copia
+SQLite segura de `/data` y los crops. No use `docker compose down -v`: elimina
+el volumen de datos.
+
+## 10. GPU, red y resolución de problemas
+
+La imagen CUDA contiene CUDA Runtime 12.9.1, cuDNN 9.24.0 y
+`onnxruntime-gpu==1.27.0`. Turing/Ampere/Ada/Hopper requieren R535 o posterior,
+Blackwell/RTX 50 requieren 570.26 o posterior; para nuevas instalaciones se
+recomienda una R580 estable o posterior. El inicio valida GPU, Compute
+Capability, Driver, CUDA/cuDNN/ORT, Provider, Sessions reales y warm-up, y
+rechaza el fallback silencioso a CPU.
+
+Al exponer la red, termine HTTPS en un proxy inverso de confianza, limite
+orígenes CORS, tasa, cuerpo y tiempo, y proteja `/data` y copias como datos
+biométricos. No registre imágenes, embeddings ni claves. La fase uno tiene una
+única API Key sin roles; no es autorización multi-tenant.
diff --git a/server/docs/user-guide.fr.md b/server/docs/user-guide.fr.md
new file mode 100644
index 000000000..9125aa674
--- /dev/null
+++ b/server/docs/user-guide.fr.md
@@ -0,0 +1,144 @@
+# Guide utilisateur InsightFace Server
+
+**Langues :** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · Français · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+Ce guide accompagne un nouvel utilisateur depuis un répertoire vide jusqu’à la première recherche réussie. Les mêmes fonctions existent dans l’interface Web, `/v1` et le SDK Python. Tous les champs et résultats HTTP sont décrits dans le [guide API](api.fr.md).
+
+## De zéro à la première recherche
+
+La version CPU nécessite Linux x86_64, Docker Engine et Docker Compose. CUDA exige en plus un pilote NVIDIA compatible et NVIDIA Container Toolkit ; il n’est pas nécessaire d’installer CUDA, cuDNN, ORT, Python ou OpenCV sur l’hôte.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Pour le GPU, utilisez `compose.cuda12.yml` et le port `18098`. L’installateur affiche la licence avant téléchargement ; les modèles publics InsightFace sont réservés à la recherche non commerciale sans licence commerciale distincte.
+
+Le Compose fourni désactive l’authentification par défaut pour une évaluation isolée. Avant toute exposition réseau, définissez `INSIGHTFACE_AUTH_ENABLED=true` et un long `INSIGHTFACE_API_KEY`. Vérifiez ensuite le Dashboard, créez une Collection, inscrivez une Person et recherchez-la avec une autre image. Arrêtez avec `docker compose ... down` sans `-v` pour conserver le volume.
+
+## 1. Connexion et état
+
+Ouvrez `http://SERVEUR:18097/` pour le CPU ou `http://SERVEUR:18098/` pour CUDA 12. Si l’authentification est active, choisissez **Configurer la clé API**, collez la clé fournie et appliquez-la à l’onglet. Elle reste uniquement en mémoire et disparaît au rechargement ou à la fermeture.
+
+Dans **Tableau de bord** ou **Système**, vérifiez que service, base, modèles et Provider sont prêts. CUDA doit afficher `CUDAExecutionProvider` et ne bascule jamais silencieusement sur CPU.
+
+## 2. Créer une Collection
+
+Dans **Collections** → **Nouvelle collection**, définissez un ID stable, un nom,
+le seuil cosinus (`0.4` au départ), un profil disponible, la capacité et le
+nombre maximal de FaceSamples par personne. La conservation JPEG d’un
+`bounding-box crop` redimensionné en 112×112 est désactivée par défaut ; ce
+n’est pas l’entrée alignée du modèle de reconnaissance.
+
+La Collection est liée à l’ID, la version, le digest, la dimension et le prétraitement du modèle. Après un changement de modèle, elle reste visible mais inscription et recherche sont refusées si le contrat diffère.
+
+Le profil de détection copie les valeurs système à la création, puis permet de modifier tailles d’entrée, seuils détection/NMS et stratégie mono-visage. `largest` privilégie la surface ; `center_largest` maximise `surface - 2,0 × distance en pixels au carré entre le centre du cadre et celui de l’image`. La confiance de détection ne participe pas à ce score.
+
+## 3. Inscrire une Person
+
+Dans **Personnes**, sélectionnez la Collection puis **Inscrire une personne**. Saisissez éventuellement ID, nom, ID externe, metadata JSON et une ou plusieurs images JPEG, PNG ou WebP.
+
+- `off` : utilise la stratégie mono-visage de la Collection et autorise plusieurs visages ;
+- `standard` : impose un visage exploitable et contrôle taille, détection, netteté, luminosité et pose ;
+- `strict` : impose aussi que la meilleure similarité interne soit supérieure à la meilleure similarité externe.
+
+Un lot accepte un succès partiel et détaille chaque rejet. Les originaux ne sont pas stockés. `external_trusted` accepte un embedding normalisé L2 ; l’image reste obligatoire pour détection et qualité, mais le vecteur n’est pas réextrait.
+
+## 4. Détecter, comparer et rechercher
+
+**Détecter** affiche boîtes, cinq points, score et qualité ; aucun visage renvoie une liste vide valide. **Comparer** utilise le profil système ou Collection pour choisir un visage par image et renvoie `similarity` cosinus, `threshold` et `matched`. La similarité n’est pas une probabilité.
+
+Dans **Rechercher**, choisissez Collection et image. Le score d’une personne est la meilleure similarité de ses FaceSamples. Les résultats sont triés par ordre décroissant ; aucun résultat donne une liste vide. Chaque échantillon est d’abord validé dans SQLite puis ajouté à l’index avant la réponse. Au redémarrage, l’index est reconstruit depuis SQLite.
+
+## 5. Surveillance de caméra RTSP
+
+Dans **Surveillance caméra**, créez un Monitor persistant et configurez source RTSP, Collection, fréquence, seuil facultatif et politique d’événements. L’aperçu est désactivé par défaut ; reconnaissance et événements continuent sans lui. Lorsqu’il est actif, la Web UI trace les inscrits en vert et les inconnus en orange depuis `/state` sur des images brutes.
+
+Le Monitor fonctionne indépendamment du navigateur et les tâches actives sont restaurées après redémarrage. La configuration est dans SQLite et les identifiants RTSP sont chiffrés dans `/data`, mais images et événements ne sont pas enregistrés. Les événements restent seulement dans un tampon mémoire borné. Le décodeur garde l’image la plus récente et ignore les anciennes au lieu de les empiler.
+
+## 6. Données et sécurité
+
+Persistez `/data` et montez `/models` en lecture seule. Sauvegardez ensemble SQLite et les recadrages avant une opération massive. Les clés sont hashées ; redémarrer le même volume avec un autre `INSIGHTFACE_API_KEY` fait tourner la clé active. Ne journalisez ni images, ni embeddings, ni clés.
+
+L’explorateur de schéma OpenAPI destiné aux développeurs se trouve sous `/docs` ; les instructions API orientées tâches sont dans cette aide. Fournissez `x-request-id` lors d’un incident. `401` concerne la clé, `409 collection_model_mismatch` le contrat modèle, `422 face_not_found` l’absence de visage exploitable.
+
+## 7. Modèles et licences
+
+Les images ne contiennent aucun modèle. Le démarrage normal reste hors ligne ;
+le service ponctuel `models` installe dans `server/.models` :
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+Les paquets pris en charge sont `buffalo_l` (`det_10g.onnx` +
+`w600k_r50.onnx`), `buffalo_m`, `buffalo_sc` et `antelopev2`. L’installation
+crée `manifest.json` et le fichier signé `MODEL.LICENSE`. Sans
+`--accept-license`, l’outil affiche les conditions puis s’arrête sans
+télécharger. Les modèles préentraînés publics InsightFace sont réservés à la
+recherche non commerciale sans licence commerciale séparée.
+
+## 8. Configuration de démarrage et recherche
+
+`server/config/server.toml` est lu une seule fois au démarrage ; toute
+modification exige un redémarrage. Valeurs initiales :
+`input_sizes=[[96,96],[512,512]]`, seuil de détection `0.50`, NMS `0.40`,
+`single_face_selection="largest"` et 100 visages au maximum. SCRFD exécute
+chaque résolution, reprojette tous les candidats sur l’image source et effectue
+un seul NMS global. `max_concurrency="auto"` signifie CPU 4 et CUDA 8.
+`[web].disabled=true` ne conserve que `/v1` et `/openapi.json`.
+
+System n’annonce que les profils réellement disponibles. Le profil est fixé à
+la création de la Collection et n’est pas sélectionnable par requête :
+
+- `fp32_v1` : CPU/CUDA standard ;
+- `fp16_v1` : CUDA ;
+- `bf16_v1` : CPU compatible ou CUDA SM80+ ;
+- `int8_x736_v1` : INT8 recommandé CPU/CUDA, accumulation INT32 ;
+- `int8_x1000_v1` : compatibilité des Collections existantes.
+
+Tous parcourent chaque FaceSample et ne sont pas des index ANN ; le score public
+reste raw cosine. `capacity_rows=100000`, garde-fou `10000000` et
+`max_faces_per_person=20`. Pour 512 dimensions, le vecteur seul représente
+environ 2 048 octets FP32, 1 024 FP16/BF16 ou 512 INT8 par ligne.
+
+## 9. SDK, construction et exploitation des données
+
+Le SDK Python accepte chemin, bytes et objet fichier et fournit des méthodes
+typées pour Detect, Compare, Collections, inscription, Search et Monitors.
+Consultez le contrat HTTP dans le [guide API](api.fr.md).
+
+Tout utilisateur peut construire depuis le dépôt complet :
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Ajoutez `--pull never` aux commandes Compose pour employer l’image locale. Les
+tags immuables sont `0.2.0-cpu` et `0.2.0-cuda12`; `cpu` et `cuda12` suivent la
+dernière version stable et aucun `latest` n’est publié. Avant mise à niveau,
+arrêtez les écritures et sauvegardez `/data` et les crops avec une méthode sûre
+pour SQLite. N’utilisez pas `docker compose down -v`, qui supprime le volume.
+
+## 10. GPU, réseau et dépannage
+
+L’image CUDA contient CUDA Runtime 12.9.1, cuDNN 9.24.0 et
+`onnxruntime-gpu==1.27.0`. Turing/Ampere/Ada/Hopper demandent R535 ou plus,
+Blackwell/RTX 50 demandent 570.26 ou plus ; une R580 stable ou plus récente est
+recommandée. Au démarrage, GPU, Compute Capability, Driver, CUDA/cuDNN/ORT,
+Provider, Sessions réelles et warm-up sont vérifiés ; aucun repli CPU silencieux
+n’est permis.
+
+Pour une exposition réseau, terminez HTTPS sur un reverse proxy fiable,
+restreignez CORS, débit, taille et délais, puis protégez `/data` et les backups
+comme données biométriques. Ne journalisez jamais images, embeddings ou clés.
+La phase un ne possède qu’une API Key sans rôles et n’est pas un système
+d’autorisation multi-tenant.
diff --git a/server/docs/user-guide.ja.md b/server/docs/user-guide.ja.md
new file mode 100644
index 000000000..68ac739fa
--- /dev/null
+++ b/server/docs/user-guide.ja.md
@@ -0,0 +1,139 @@
+# InsightFace Server ユーザーガイド
+
+**言語:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · 日本語 · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+このガイドは初めて利用する方のために、空の作業ディレクトリから最初の検索成功までを順番に説明します。同じ機能は Web UI、`/v1` API、Python SDK から利用できます。全 HTTP 項目とレスポンスは [API 利用ガイド](api.ja.md) を参照してください。
+
+## ゼロから起動して最初の検索を行う
+
+CPU 版には Linux x86_64、Docker Engine、Docker Compose が必要です。CUDA 版には対応 NVIDIA Driver と NVIDIA Container Toolkit も必要ですが、ホスト側 CUDA、cuDNN、ORT、Python、OpenCV は不要です。
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+GPU では `compose.cuda12.yml` とポート `18098` を使用します。モデルのダウンロード前にライセンスが表示されます。公開済み InsightFace 事前学習モデルは、別途商用ライセンスがない限り非商用研究用途に限定されます。
+
+同梱 Compose は隔離評価向けに認証を既定で無効にしています。有効化する場合は起動前に `INSIGHTFACE_AUTH_ENABLED=true` と長い `INSIGHTFACE_API_KEY` を設定します。UI を開き、Dashboard確認 → Collection作成 → Person登録 → 別画像でSearchの順に進めます。停止は `docker compose ... down` を使用し、データを保持する場合は `-v` を付けないでください。
+
+## 1. ログインと準備確認
+
+CPU は `http://SERVER:18097/`、CUDA 12 は `http://SERVER:18098/` を開きます。認証が有効な場合は **API キーを設定** から管理者の Key を入力します。Key は現在のタブのメモリだけに保持され、再読み込みまたはタブを閉じると消えます。
+
+**ダッシュボード** または **システム** で、サービス、データベース、モデル、Provider が ready であることを確認します。CUDA 版は `CUDAExecutionProvider` を表示しなければならず、CPU へ自動フォールバックしません。
+
+## 2. Collection を作成
+
+**コレクション** → **新規コレクション** で、安定した ID、名前、既定 cosine
+しきい値（初期値 `0.4`）、利用可能な検索プロファイル、容量、人物ごとの最大
+FaceSample 数を設定します。112×112 にリサイズした `bounding-box crop` JPEG
+の保存は既定でオフです。これは認識モデル用のアライン済み入力ではありません。
+
+Collection はモデル ID、バージョン、digest、次元、前処理に固定されます。モデル変更後も古い Collection は表示されますが、契約が異なる登録・検索は明示的に拒否されます。
+
+検出設定は作成時にシステム既定値をコピーし、入力サイズ、検出/NMS しきい値、単一顔戦略を後から変更できます。`largest` は面積優先、`center_largest` は `面積 - 2.0 × 顔枠中心と画像中心のピクセル距離の二乗` を最大化します。検出信頼度はこのスコアに含みません。
+
+## 3. Person を登録
+
+**人物** で Collection を選び、**人物を登録** を開きます。ID、名前、外部 ID、JSON metadata と 1 枚以上の JPEG、PNG、または WebP を指定します。
+
+- `off`: Collection の単一顔戦略を使用し、複数顔を許可します。
+- `standard`: 1 つの有効顔を要求し、サイズ、検出値、鮮明度、明るさ、姿勢を確認します。
+- `strict`: standard に加え、最良の人物内 similarity が最良の人物外 similarity より高いことを要求します。
+
+一括登録は部分成功を返します。拒否理由を確認してから再試行してください。元画像は保存されません。`external_trusted` では L2 正規化済み embedding を利用でき、画像は品質確認に必要ですが特徴量の再抽出は行いません。
+
+## 4. 検出・比較・検索
+
+**検出** は顔矩形、5 点、検出値、品質を表示し、顔なしは空リストで成功します。**比較** は選択したシステムまたは Collection の戦略で各画像から 1 顔を選び、cosine `similarity`、`threshold`、`matched` を返します。Similarity は確率ではありません。
+
+**検索** で Collection と画像を選択します。Collection の戦略で顔を選び、人物の全 FaceSample 中の最高 similarity を人物スコアとして降順に返します。一致なしは空リストです。新規 FaceSample は SQLite へ commit 後、応答前にメモリ索引へ追加されます。再起動時は SQLite から再構築します。
+
+## 5. RTSP カメラ監視
+
+**カメラ監視** で永続的な Monitor を作成し、RTSP 接続先、Collection、推論頻度、任意のしきい値、event policy を設定します。Preview は既定で無効で、無効でも認識と event は継続します。有効時は Web UI が raw frame と `/state` から緑の登録人物枠、オレンジの未登録顔枠を描画します。
+
+Monitor はブラウザーと独立して動作し、有効な task は Server 再起動後に復元されます。設定は SQLite、RTSP 認証情報は `/data` に暗号化保存されますが、動画 frame と event は保存しません。Event は上限付き memory buffer だけに残り、再起動で失われます。Decoder は最新 frame だけを保持し、遅い処理では古い frame を queue に積まず skip します。
+
+## 6. データと安全性
+
+`/data` を永続化し、`/models` は読み取り専用にします。大量削除前に SQLite と顔画像領域を一緒にバックアップしてください。API Key は hash 保存され、同じデータ volume で異なる `INSIGHTFACE_API_KEY` を指定して再起動すると Key がローテーションされます。画像、embedding、Key をログへ出力しないでください。
+
+開発者向け OpenAPI スキーマエクスプローラーは `/docs`、操作別の API 説明はこのヘルプ内にあります。障害報告には応答ヘッダーの `x-request-id` を含めてください。`401` は Key、`409 collection_model_mismatch` はモデル契約、`422 face_not_found` は有効顔を確認します。
+
+## 7. モデルとライセンス
+
+イメージにはモデルを含みません。通常起動はオフラインで、1 回限りの `models`
+サービスが `server/.models` へインストールします。
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+対応パッケージは `buffalo_l`（`det_10g.onnx` + `w600k_r50.onnx`）、
+`buffalo_m`、`buffalo_sc`、`antelopev2` です。インストール後は
+`manifest.json` と署名済み `MODEL.LICENSE` が残ります。
+`--accept-license` がない場合は条項を表示してダウンロードせず終了します。
+公開 InsightFace 学習済みモデルは、別途商用ライセンスがない限り非商用研究用です。
+
+## 8. 起動設定と検索
+
+`server/config/server.toml` は起動時に一度だけ読み込まれ、変更にはコンテナ再起動が
+必要です。既定値は `input_sizes=[[96,96],[512,512]]`、検出しきい値 `0.50`、
+NMS `0.40`、`single_face_selection="largest"`、最大 100 顔です。SCRFD は各解像度
+を実行し、元画像座標へ戻した全候補に 1 回だけグローバル NMS を行います。
+`max_concurrency="auto"` は CPU 4、CUDA 8 です。`[web].disabled=true` では
+`/v1` と `/openapi.json` だけを提供します。
+
+利用可能な検索 Profile は System に表示されます。Collection 作成後は変更できず、
+リクエスト単位でも指定できません。
+
+- `fp32_v1`: CPU/CUDA の標準。
+- `fp16_v1`: CUDA。
+- `bf16_v1`: 対応 CPU または SM80+ CUDA。
+- `int8_x736_v1`: CPU/CUDA の推奨 INT8。INT32 で累積。
+- `int8_x1000_v1`: 既存 Collection 互換用。
+
+すべての Profile は全 FaceSample を走査する Flat 検索で ANN ではありません。
+公開スコアは常に raw cosine です。`capacity_rows` の既定は `100000`、上限ガードは
+`10000000`、`max_faces_per_person` は `20` です。512 次元の純ベクトル容量は 1 行
+あたり FP32 2,048 byte、FP16/BF16 1,024 byte、INT8 512 byte が目安です。
+
+## 9. SDK、ビルド、データ運用
+
+SDK は path、bytes、file-like object に対応し、`detect`、`compare`、
+`create_collection`、`add_person`、`search`、Monitor 操作を型付きで提供します。
+詳細な HTTP 契約は [API 利用ガイド](api.ja.md)を参照してください。
+
+完全なリポジトリからユーザー自身でビルドできます。
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+ローカルイメージを使う Compose 操作には `--pull never` を付けます。公開固定 Tag は
+`0.2.0-cpu` と `0.2.0-cuda12`、移動 Tag は `cpu` と `cuda12` で、`latest` は
+ありません。アップグレード前に書き込みを止め、SQLite-safe な方法で `/data` と
+crop を一緒にバックアップしてください。`docker compose down -v` は Volume を
+削除するため使わないでください。
+
+## 10. GPU、ネットワーク、トラブルシュート
+
+CUDA イメージは CUDA Runtime 12.9.1、cuDNN 9.24.0、
+`onnxruntime-gpu==1.27.0` を含みます。Turing/Ampere/Ada/Hopper は Driver R535
+以上、Blackwell/RTX 50 は 570.26 以上、新規導入は安定版 R580 以上を推奨します。
+起動時に GPU、Compute Capability、Driver、CUDA/cuDNN/ORT、Provider、実モデル
+Session と warm-up を検証し、CPU への暗黙 fallback は拒否します。
+
+ネットワーク公開時は信頼できる Reverse Proxy で HTTPS を終端し、CORS origin、
+rate/body/timeout を制限してください。画像、embedding、Key をログに残さず、
+`/data` とバックアップを生体情報として保護します。Phase 1 は権限区別のない単一
+API Key であり、マルチテナント認可機能ではありません。
diff --git a/server/docs/user-guide.ko.md b/server/docs/user-guide.ko.md
new file mode 100644
index 000000000..f698e5caa
--- /dev/null
+++ b/server/docs/user-guide.ko.md
@@ -0,0 +1,139 @@
+# InsightFace Server 사용자 가이드
+
+**언어:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · 한국어
+
+이 가이드는 처음 사용하는 사용자가 빈 작업 폴더에서 시작해 첫 검색에 성공할 때까지 단계별로 설명합니다. 같은 기능을 Web UI, `/v1` API, Python SDK에서 사용할 수 있습니다. 모든 HTTP 필드와 결과는 [API 사용 가이드](api.ko.md)를 확인하세요.
+
+## 처음부터 첫 검색까지
+
+CPU 버전에는 Linux x86_64, Docker Engine, Docker Compose가 필요합니다. CUDA 버전에는 호환 NVIDIA Driver와 NVIDIA Container Toolkit도 필요하지만 호스트 CUDA, cuDNN, ORT, Python, OpenCV는 설치하지 않아도 됩니다.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+GPU는 `compose.cuda12.yml`과 포트 `18098`을 사용합니다. 설치 전 모델 라이선스가 표시되며, 별도 상용 라이선스가 없으면 공개 InsightFace 모델은 비상업 연구용으로만 사용할 수 있습니다.
+
+제공되는 Compose는 격리된 평가 환경을 위해 인증이 기본적으로 꺼져 있습니다. 네트워크에 공개하기 전 `INSIGHTFACE_AUTH_ENABLED=true`와 긴 `INSIGHTFACE_API_KEY`를 설정하세요. Dashboard 확인, Collection 생성, Person 등록, 다른 이미지로 Search 순서로 진행합니다. 데이터 볼륨을 보존하려면 `docker compose ... down`에 `-v`를 붙이지 마세요.
+
+## 1. 로그인과 준비 상태
+
+CPU는 `http://SERVER:18097/`, CUDA 12는 `http://SERVER:18098/`을 엽니다. 인증이 켜져 있으면 **API 키 설정**에서 운영자가 제공한 Key를 입력하고 현재 탭에 적용합니다. Key는 탭 메모리에만 있으며 새로고침하거나 닫으면 삭제됩니다.
+
+**대시보드** 또는 **시스템**에서 서비스, 데이터베이스, 모델, Provider가 준비되었는지 확인합니다. CUDA는 `CUDAExecutionProvider`를 표시해야 하며 CPU로 조용히 전환하지 않습니다.
+
+## 2. Collection 만들기
+
+**컬렉션** → **새 컬렉션**에서 안정적인 ID, 이름, cosine 임계값(초기 `0.4`),
+사용 가능한 검색 프로필, 용량, 사람별 최대 FaceSample 수를 설정합니다. 112×112로
+조정한 `bounding-box crop` JPEG 저장은 기본적으로 꺼져 있으며, 인식 모델의 정렬
+입력이 아닙니다.
+
+Collection은 모델 ID, 버전, digest, 차원, 전처리에 고정됩니다. 모델을 바꿔도 이전 Collection은 보이지만 계약이 다르면 등록과 검색이 명시적으로 거부됩니다.
+
+검출 프로필은 Collection 생성 시 시스템 값을 복사하며 이후 입력 크기, 검출/NMS 임계값, 단일 얼굴 전략을 변경할 수 있습니다. `largest`는 면적을 우선하고, `center_largest`는 `면적 - 2.0 × 얼굴 상자 중심과 이미지 중심 사이의 픽셀 거리 제곱`을 최대화합니다. 검출 신뢰도는 이 점수에 포함되지 않습니다.
+
+## 3. Person 등록
+
+**사람**에서 Collection을 선택하고 **사람 등록**을 엽니다. ID, 이름, 외부 ID, JSON metadata와 여러 JPEG, PNG 또는 WebP를 지정할 수 있습니다.
+
+- `off`: Collection의 단일 얼굴 전략을 사용하며 여러 얼굴을 허용합니다.
+- `standard`: 하나의 사용 가능한 얼굴과 크기, 검출, 선명도, 밝기, 자세 검사를 요구합니다.
+- `strict`: standard에 더해 최상의 클래스 내 similarity가 최상의 클래스 외 similarity보다 커야 합니다.
+
+일괄 등록은 부분 성공과 각 거부 이유를 반환합니다. 원본은 저장하지 않습니다. `external_trusted`는 L2 정규화된 embedding을 받으며 이미지로 검출과 품질은 검사하지만 특징을 다시 추출하지 않습니다.
+
+## 4. 검출, 비교, 검색
+
+**검출**은 상자, 5개 점, 검출 점수, 품질을 표시하며 얼굴 없음은 정상적인 빈 목록입니다. **비교**는 시스템 또는 Collection 프로필로 각 이미지에서 한 얼굴을 선택하고 cosine `similarity`, `threshold`, `matched`를 반환합니다. 유사도는 확률이 아닙니다.
+
+**검색**에서 Collection과 이미지를 선택합니다. 한 사람의 점수는 모든 FaceSample 중 최고 similarity입니다. 결과는 내림차순이며 일치 없음은 빈 목록입니다. 새 샘플은 SQLite에 commit된 다음 성공 응답 전에 메모리 인덱스에 추가됩니다. 재시작 시 SQLite에서 재구축합니다.
+
+## 5. RTSP 카메라 모니터링
+
+**카메라 모니터링**에서 영구 Monitor를 만들고 RTSP source, Collection, 추론 속도, 선택 임계값과 이벤트 정책을 설정합니다. 미리보기는 기본으로 꺼져 있으며 인식과 이벤트는 계속 작동합니다. 켜면 Web UI가 원본 프레임 위에 `/state` 결과로 등록 인물은 초록색, 미등록 얼굴은 주황색 상자를 그립니다.
+
+Monitor는 브라우저와 독립적으로 실행되고 활성 작업은 서버 재시작 후 복원됩니다. 설정은 SQLite에, RTSP 자격증명은 `/data`에 암호화 저장되지만 영상 프레임과 이벤트는 저장하지 않습니다. 이벤트는 제한된 메모리 버퍼에만 남습니다. 디코더는 최신 프레임만 보관하고 오래된 프레임을 쌓지 않고 건너뜁니다.
+
+## 6. 데이터와 보안
+
+`/data`를 영속화하고 `/models`는 읽기 전용으로 마운트합니다. 대량 작업 전 SQLite와 face crop 영역을 함께 백업하세요. Key는 hash로 저장되며 같은 volume을 다른 `INSIGHTFACE_API_KEY`로 시작하면 활성 Key가 교체됩니다. 이미지, embedding, Key를 로그에 남기지 마세요.
+
+개발자용 OpenAPI 스키마 탐색기는 `/docs`에 있으며 작업 중심 API 안내는 이 도움말에 있습니다. 문제 보고 시 `x-request-id`를 포함하세요. `401`은 Key, `409 collection_model_mismatch`는 모델 계약, `422 face_not_found`는 사용 가능한 얼굴을 확인합니다.
+
+## 7. 모델과 라이선스
+
+이미지에는 모델이 포함되지 않습니다. 일반 Server 시작은 오프라인이며, 일회성
+`models` 서비스가 `server/.models`에 설치합니다.
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+지원 패키지는 `buffalo_l`(`det_10g.onnx` + `w600k_r50.onnx`),
+`buffalo_m`, `buffalo_sc`, `antelopev2`입니다. 설치하면 `manifest.json`과
+서명된 `MODEL.LICENSE`가 생성됩니다. `--accept-license`가 없으면 조건만
+표시하고 다운로드하지 않습니다. 공개 InsightFace 사전 학습 모델은 별도 상업용
+라이선스가 없는 경우 비상업 연구용입니다.
+
+## 8. 시작 설정과 검색
+
+`server/config/server.toml`은 시작 시 한 번만 읽으며 변경 후 재시작해야 합니다.
+기본값은 `input_sizes=[[96,96],[512,512]]`, 검출 threshold `0.50`, NMS `0.40`,
+`single_face_selection="largest"`, 최대 100개 얼굴입니다. SCRFD는 각 해상도를
+실행하고 원본 좌표로 합친 모든 후보에 전역 NMS를 한 번 수행합니다.
+`max_concurrency="auto"`는 CPU 4, CUDA 8입니다.
+`[web].disabled=true`이면 `/v1`과 `/openapi.json`만 제공합니다.
+
+System은 현재 환경에서 사용 가능한 Profile만 표시합니다. Collection 생성 후 고정되며
+Search 요청별로 바꿀 수 없습니다.
+
+- `fp32_v1`: CPU/CUDA 표준
+- `fp16_v1`: CUDA
+- `bf16_v1`: 지원 CPU 또는 SM80+ CUDA
+- `int8_x736_v1`: CPU/CUDA 권장 INT8, INT32 누적
+- `int8_x1000_v1`: 기존 Collection 호환
+
+모든 Profile은 전체 FaceSample을 순회하는 Flat 검색이며 ANN이 아닙니다. 공개 score는
+raw cosine입니다. `capacity_rows=100000`, 배포 한도 `10000000`,
+`max_faces_per_person=20`이 기본입니다. 512차원 순수 벡터는 행당 FP32 2,048 byte,
+FP16/BF16 1,024 byte, INT8 512 byte 정도입니다.
+
+## 9. SDK, 빌드와 데이터 운영
+
+Python SDK는 경로, bytes, file-like object를 지원하고 Detect, Compare,
+Collections, 등록, Search, Monitors의 타입 지정 메서드를 제공합니다. 전체 HTTP
+계약은 [API 사용 가이드](api.ko.md)를 확인하세요.
+
+전체 저장소에서 사용자가 직접 이미지를 빌드할 수 있습니다.
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+로컬 이미지를 쓸 때 Compose에 `--pull never`를 추가합니다. 고정 Tag는
+`0.2.0-cpu`, `0.2.0-cuda12`이고 이동 Tag `cpu`, `cuda12`는 최신 안정 버전을
+가리키며 `latest`는 없습니다. 업그레이드 전 쓰기를 중지하고 `/data`와 crop을
+SQLite-safe 방식으로 백업하세요. `docker compose down -v`는 데이터 Volume을
+삭제하므로 사용하지 마세요.
+
+## 10. GPU, 네트워크와 문제 해결
+
+CUDA 이미지는 CUDA Runtime 12.9.1, cuDNN 9.24.0,
+`onnxruntime-gpu==1.27.0`을 포함합니다. Turing/Ampere/Ada/Hopper는 R535+,
+Blackwell/RTX 50은 570.26+가 필요하며 신규 배포는 안정적인 R580+를 권장합니다.
+시작 시 GPU, Compute Capability, Driver, CUDA/cuDNN/ORT, Provider, 실제 모델
+Session과 warm-up을 검사하며 CPU로 조용히 fallback하지 않습니다.
+
+네트워크에 공개할 때 신뢰할 수 있는 Reverse Proxy에서 HTTPS를 종료하고 CORS origin,
+rate, body, timeout을 제한하며 `/data`와 백업을 생체 데이터로 보호하세요. 이미지,
+embedding, Key를 로그에 남기지 마세요. 1단계는 역할 없는 단일 API Key만 지원하며
+multi-tenant 권한 시스템이 아닙니다.
diff --git a/server/docs/user-guide.md b/server/docs/user-guide.md
new file mode 100644
index 000000000..4f280c7fe
--- /dev/null
+++ b/server/docs/user-guide.md
@@ -0,0 +1,283 @@
+# InsightFace Server user guide
+
+**Languages:** English · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+This is the step-by-step operating guide for first-time users. It starts with an
+empty checkout and ends with a working Collection, enrolled Person, and search
+result. The same operations are available through the Web UI, `/v1` API, and
+Python SDK. For every HTTP field and response, open the
+[API usage guide](api.md).
+
+## Start here: from zero to a working server
+
+You need a Linux x86_64 host with Docker Engine and Docker Compose. A CUDA
+deployment additionally needs a supported NVIDIA driver and NVIDIA Container
+Toolkit. Do not install host CUDA, cuDNN, ONNX Runtime, Python, or OpenCV.
+
+CPU example:
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+For NVIDIA GPU, replace `compose.cpu.yml` with `compose.cuda12.yml` and use port
+`18098`. The model installer shows the model license before download. Public
+InsightFace pretrained models are restricted to non-commercial research unless
+you have a separate commercial license.
+
+The bundled Compose files default to `auth_enabled=false` for isolated evaluation.
+No API key field is required in that mode, and the Web UI hides its key control.
+Before exposing the service to other users or networks, enable authentication
+before startup:
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='replace-with-a-long-random-secret'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+Open `http://SERVER:18097/` for CPU or `http://SERVER:18098/` for CUDA. Complete
+the first workflow in this order: check **Dashboard**, create a Collection,
+register one Person with at least one clear image, then use **Search** with a
+different image of that Person. A successful no-match is an empty list; it is
+not a server failure. Stop with `docker compose ... down` without `-v`; adding
+`-v` permanently removes the named data volume.
+
+## 1. Sign in and check readiness
+
+Open `http://SERVER:18097/` for CPU or `http://SERVER:18098/` for CUDA 12. If authentication is enabled, choose **Configure API key**, paste the key supplied by the operator, and select **Use for this tab**. The browser keeps it only in memory; reloading or closing the tab clears it.
+
+Check **Dashboard** or **System** before enrolling data. The service, database, model and provider must be ready. A CUDA deployment must report `CUDAExecutionProvider`; it never silently falls back to CPU.
+
+## 2. Create a Collection
+
+Open **Collections**, choose **New collection**, and set:
+
+- a stable ID such as `employees`;
+- a display name and optional metadata;
+- the default cosine threshold, initially `0.4`;
+- a search profile supported by the current host;
+- capacity and maximum FaceSamples per Person;
+- detector input sizes, detector/NMS thresholds, and a single-face strategy;
+- optional 112×112 `bounding-box crop` JPEG storage—not an aligned recognition
+  input—disabled by default.
+
+A Collection is pinned to the active model identity, digest, embedding dimension and preprocessing version. Its detection profile starts as a copy of the system profile and may be changed later; each update affects the next request and increments `detection_revision`, but does not reprocess existing FaceSamples. `largest` prefers area. `center_largest` maximizes `area - 2.0 × squared pixel distance from the face-box center to the image center`; detection confidence is not part of this score.
+
+## 3. Register a Person
+
+Open **People**, select a Collection, then **Register person**. Provide an optional stable person ID, name, external ID and JSON metadata. Drop one or more JPEG, PNG, or WebP images.
+
+Enrollment review modes are:
+
+- `off`: use the Collection single-face strategy; multiple faces are allowed;
+- `standard`: require one usable face and apply size, detection, sharpness, brightness and pose checks;
+- `strict`: apply standard checks and require the sample's best within-person similarity to exceed its best outside-person similarity.
+
+Batch enrollment supports partial success. Review each rejected image and its
+reason before retrying; the service does not retain rejected originals. When
+crop storage is enabled, only a `bounding-box crop` resized to 112×112 is
+stored—not the original upload or the aligned recognition input.
+
+Trusted systems may send a precomputed, L2-normalized embedding using `external_trusted`. An image is still required for detection and quality review, but the server does not re-extract the embedding. The embedding contract must exactly match the Collection.
+
+## 4. Detect and compare
+
+Use **Detect** to upload one image and inspect boxes, five landmarks, confidence and heuristic quality. No face is a successful result with an empty list.
+
+Use **Compare** to upload source and target images. Select the system or a Collection detection profile; its strategy chooses one usable face in each image. The result contains raw cosine `similarity`, the selected `threshold`, and `matched`. Similarity is not a probability. If either image has no usable face, the API returns `422 face_not_found`.
+
+## 5. Search a Collection
+
+Open **Search**, select the Collection, upload a query image, set a result limit and optionally override the threshold. The Collection detection profile chooses the query face. Results are sorted by similarity; a Person's score is the maximum score among that Person's FaceSamples. No match is a successful empty list.
+
+Newly accepted FaceSamples are committed to SQLite and then added to the in-memory index before the successful response is returned. Deletions update both stores. On restart the index is rebuilt from SQLite, which remains authoritative.
+
+## 6. RTSP camera monitoring
+
+Open **Camera monitoring** and choose **New Monitor**. Give the task an ID and
+name, enter an `rtsp://` or `rtsps://` source, select a Collection, and choose the
+inference rate and optional match threshold. The event settings control how many
+consecutive observations confirm a face, when absence creates an exit event, the
+duplicate-event cooldown, and how many recent events stay in memory.
+
+**Web video preview is off by default.** Enable it only when an operator needs a
+visual check. Recognition and events continue without a preview. When enabled,
+the server sends raw JPEG frames and the Web UI draws green boxes for enrolled
+people and amber boxes for detected but unenrolled faces using `/state` results.
+
+The Monitor runs on the server independently of the browser. Closing the page
+does not stop it, and enabled Monitors resume after a Server restart. Use
+**Start/Stop** to change `enabled`, **Edit** to rotate the RTSP source or tune its
+settings, and **Delete** to remove the task. The decoder keeps only the newest
+frame; if processing exceeds the requested interval, stale frames are skipped
+instead of queued.
+
+Monitor configuration is stored in SQLite. RTSP credentials are encrypted in
+`/data` and are never returned by the API. Video frames are not saved. Recent
+enter/exit/error/recovery events exist only in a bounded in-memory ring and are
+lost on restart. Use HTTPS when the UI/API crosses an untrusted network and
+restrict Monitor administration to trusted operators.
+
+## 7. Update and delete data
+
+Collections and Persons can be edited from their lists. Deleting a FaceSample removes its embedding and optional crop. Deleting a non-empty Collection requires explicit force confirmation. Back up `/data` before bulk or destructive maintenance.
+
+## 8. API and Python SDK
+
+The developer OpenAPI schema explorer is at `/docs`; task-oriented API instructions are in this Help manual. Every API response carries `x-request-id`; include it when reporting a problem.
+
+```python
+from insightface_server import Client
+
+client = Client("http://localhost:18097", api_key="your-key")
+client.create_collection(collection_id="employees", name="Employees", threshold=0.4)
+client.add_person("employees", person_id="alice", images=["alice-1.jpg", "alice-2.jpg"])
+matches = client.search("employees", "query.jpg", limit=5)
+```
+
+## 9. Data, backup and security
+
+- Persist `/data`; mount `/models` read-only.
+- Back up the SQLite database and configured crop storage together while writes are stopped or by using a SQLite-safe snapshot method.
+- API keys are stored as hashes. Supplying a different `INSIGHTFACE_API_KEY` on a later start intentionally rotates the active key for that data volume.
+- Do not log images, embeddings or keys. Keep broad CORS disabled unless required.
+- Model files are not included in the image. InsightFace-provided open-source
+  pretrained models, including `buffalo_l`, are for non-commercial research use
+  only. Commercial use requires a separate license; visit
+  <https://www.insightface.ai>. The **System** page displays the same notice.
+
+## 10. Troubleshooting
+
+`401 unauthorized` means the tab has no current key or the key was rotated. `409 collection_model_mismatch` means the Collection was created with a different model contract. `422 face_not_found` means no usable face was selected. A CUDA startup failure is intentional when the Driver, GPU, model session, provider or warm-up validation fails. Check **System**, container logs and the response `request_id`.
+
+## 11. Models and model licenses
+
+The images do not contain models. The one-shot `models` service installs a
+package into `server/.models`, while normal Server startup remains offline:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+Supported public packages are `buffalo_l` (`det_10g.onnx` +
+`w600k_r50.onnx`), `buffalo_m` (`det_2.5g.onnx` + `w600k_r50.onnx`),
+`buffalo_sc` (`det_500m.onnx` + `w600k_mbf.onnx`), and `antelopev2`
+(`scrfd_10g_bnkps.onnx` + `glintr100.onnx`). Installation creates
+`manifest.json` and signed `MODEL.LICENSE`. Without `--accept-license`, the
+tool prints the terms and exits without downloading. `models verify` validates
+the package identity, signed license, validity dates, and current authorization.
+
+Public InsightFace pretrained models are for non-commercial research unless a
+separate commercial license has been issued. A private model can use the same
+manifest and offline signed license format. The license identifies `model_id`;
+it is a compliance credential, not DRM or a model-file checksum.
+
+## 12. Startup-only configuration
+
+The common startup file is `server/config/server.toml`. Compose mounts it
+read-only at `/etc/insightface/server.toml`; edit it before startup and restart
+the container to apply a change. Defaults are:
+
+```toml
+[inference]
+max_concurrency = "auto" # CPU 4, CUDA 8
+
+[detection]
+input_sizes = [[96, 96], [512, 512]]
+threshold = 0.50
+nms_threshold = 0.40
+single_face_selection = "largest"
+max_detected_faces = 100
+
+[web]
+disabled = false
+```
+
+Dynamic SCRFD runs every configured resolution, maps candidates back to the
+source image, merges all candidates, and performs one global NMS. Settings are
+read once; there is no runtime settings API. New Collections copy the system
+detection profile, after which their profile can be updated independently for
+the next request. Stateless Detect and Embeddings use the system profile;
+Compare can use the system profile or a selected Collection; enrollment and
+search use their Collection.
+
+Set `[web].disabled=true` for API-only operation. `/v1` and `/openapi.json`
+remain available, while `/`, `/docs`, guides, and frontend assets are not
+registered.
+
+## 13. Exact-search profiles and capacity
+
+The System response advertises only profiles available on the current CPU/GPU.
+A Collection fixes one profile when created; a search request cannot change it.
+
+| Profile | Stored representation | Typical availability |
+| --- | --- | --- |
+| `fp32_v1` | FP32 | CPU and CUDA |
+| `fp16_v1` | FP16 | CUDA |
+| `bf16_v1` | BF16 | supported CPU or SM80+ CUDA |
+| `int8_x736_v1` | INT8, scale 736 | CPU and CUDA; recommended INT8 |
+| `int8_x1000_v1` | INT8, scale 1000 | compatibility profile |
+
+All are flat exhaustive searches over every live FaceSample; low-precision
+profiles approximate FP32 scores but are not ANN indexes. INT8 dot products
+accumulate into INT32. Public similarities and thresholds remain raw cosine.
+
+`capacity_rows` reserves the maximum live rows for that Collection and avoids
+routine growth pauses. Approximate vector storage for 512 dimensions is
+2,048 bytes per FP32 row, 1,024 per FP16/BF16 row, and 512 per INT8 row,
+before IDs and workspaces. Set capacity from an actual memory budget. The
+default is `100000`; the deployment guardrail defaults to `10000000`.
+`max_faces_per_person` defaults to `20` and limits sample count, not the number
+of people.
+
+## 14. CUDA support and fail-fast verification
+
+The CUDA image contains CUDA Runtime 12.9.1, cuDNN 9.24.0, Python 3.11, and
+`onnxruntime-gpu==1.27.0`. The host needs only Driver, Docker Engine, NVIDIA
+Container Toolkit, and a compatible GPU.
+
+- Turing, Ampere, Ada, and Hopper: Driver R535 or newer.
+- Blackwell and RTX 50 series: Driver 570.26 or newer.
+- New deployments should prefer a stable R580 or newer driver.
+
+Architecture compatibility is not a claim that every GPU SKU is formally
+certified. On every CUDA start the Server checks GPU model, Compute Capability,
+Driver, actual CUDA/cuDNN/ORT versions, the presence of
+`CUDAExecutionProvider`, real detector and recognizer Sessions, and real
+warm-up inference. It audits provider placement and terminates instead of
+silently falling back to CPU. Confirm the result on **System** before use.
+
+## 15. Build, upgrade, backup, and recovery
+
+Users may build both images from the complete repository checkout:
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Then add `--pull never` to Compose model/install and `up` commands to use the
+local image. Builds use pinned base images and locked dependencies, but require
+network access for those inputs. The public tags are
+`0.2.0-cpu`/`0.2.0-cuda12`; moving `cpu`/`cuda12` tags point to the latest
+stable variant, and there is deliberately no `latest` tag.
+
+Before upgrading, stop writes and create a SQLite-safe snapshot of `/data`
+plus any crop storage. Keep `/models` and its license files. Start the new
+container against a copy first, check migrations and `/v1/health`, then verify
+the model contract and a known search. Use `docker compose down` without `-v`;
+`docker compose down -v` deletes the named data volume.
+
+For network exposure, terminate HTTPS at a trusted reverse proxy, allow only
+required origins rather than broad CORS, apply edge rate/body/time limits, and
+protect the data volume and backups as biometric data. The Server has one
+undifferentiated API key in phase one and is not a multi-tenant authorization
+system.
diff --git a/server/docs/user-guide.pt.md b/server/docs/user-guide.pt.md
new file mode 100644
index 000000000..b49de14ab
--- /dev/null
+++ b/server/docs/user-guide.pt.md
@@ -0,0 +1,138 @@
+# Guia do utilizador do InsightFace Server
+
+**Idiomas:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · Português · [한국어](user-guide.ko.md)
+
+Este guia conduz um novo utilizador desde uma pasta vazia até à primeira pesquisa bem-sucedida. As mesmas funções estão disponíveis na Web UI, em `/v1` e no SDK Python. Consulte todos os campos e resultados HTTP no [guia da API](api.pt.md).
+
+## Do zero à primeira pesquisa
+
+CPU requer Linux x86_64, Docker Engine e Docker Compose. CUDA requer ainda um Driver NVIDIA compatível e NVIDIA Container Toolkit; não instale CUDA, cuDNN, ORT, Python ou OpenCV no anfitrião.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Para GPU use `compose.cuda12.yml` e a porta `18098`. O instalador mostra a licença antes do download; os modelos públicos InsightFace destinam-se apenas a investigação não comercial sem uma licença comercial separada.
+
+O Compose fornecido desativa a autenticação por predefinição para avaliação isolada. Antes de expor o serviço, defina `INSIGHTFACE_AUTH_ENABLED=true` e uma `INSIGHTFACE_API_KEY` longa. Verifique depois o Dashboard, crie uma Collection, registe uma Person e pesquise com outra imagem. Pare com `docker compose ... down` sem `-v` para preservar o volume.
+
+## 1. Entrada e estado
+
+Abra `http://SERVIDOR:18097/` para CPU ou `http://SERVIDOR:18098/` para CUDA 12. Se a autenticação estiver ativa, escolha **Configurar chave API**, introduza a chave do operador e use-a neste separador. Fica apenas em memória e desaparece ao atualizar ou fechar.
+
+Em **Painel** ou **Sistema**, confirme que serviço, base de dados, modelos e Provider estão prontos. CUDA deve indicar `CUDAExecutionProvider` e nunca recua silenciosamente para CPU.
+
+## 2. Criar uma Collection
+
+Em **Coleções** → **Nova coleção**, defina ID estável, nome, limiar cosine
+(`0.4` inicialmente), perfil disponível, capacidade e máximo de FaceSamples por
+pessoa. Guardar em JPEG um `bounding-box crop` redimensionado para 112×112 está
+desligado por predefinição; não é a entrada alinhada de reconhecimento.
+
+A Collection fica ligada ao ID, versão, digest, dimensão e pré-processamento do modelo. Após mudar o modelo, continua visível, mas registo e pesquisa são recusados quando o contrato não coincide.
+
+O perfil de deteção copia os valores do sistema ao criar a Collection e depois permite alterar tamanhos de entrada, limiares de deteção/NMS e estratégia de um rosto. `largest` prioriza a área; `center_largest` maximiza `área - 2,0 × distância em píxeis ao quadrado entre o centro da caixa e o da imagem`. A confiança de deteção não participa nesta pontuação.
+
+## 3. Registar uma Person
+
+Em **Pessoas**, selecione a Collection e **Registar pessoa**. Pode indicar ID, nome, ID externo, metadata JSON e várias imagens JPEG, PNG ou WebP.
+
+- `off`: usa a estratégia de um rosto da Collection e permite vários rostos;
+- `standard`: exige um rosto utilizável e verifica tamanho, deteção, nitidez, brilho e pose;
+- `strict`: exige também que a melhor similaridade interna seja superior à melhor similaridade com outra pessoa.
+
+O lote aceita sucesso parcial e explica cada rejeição. Os originais não são guardados. `external_trusted` aceita um embedding normalizado L2; a imagem continua obrigatória para deteção e qualidade, mas o vetor não é extraído novamente.
+
+## 4. Detetar, comparar e pesquisar
+
+**Detetar** mostra caixas, cinco pontos, pontuação e qualidade; sem rostos devolve uma lista vazia válida. **Comparar** usa o perfil do sistema ou da Collection para escolher um rosto por imagem e devolve `similarity` cosine, `threshold` e `matched`. Similaridade não é probabilidade.
+
+Em **Pesquisar**, escolha Collection e imagem. A pontuação da pessoa é a maior similaridade dos seus FaceSamples. Os resultados são decrescentes; sem correspondência é uma lista vazia. Cada amostra é confirmada no SQLite e adicionada ao índice antes da resposta. No reinício, o índice é reconstruído a partir do SQLite.
+
+## 5. Monitorização de câmara RTSP
+
+Em **Monitorização de câmaras**, crie um Monitor persistente e configure origem RTSP, Collection, frequência, limiar opcional e política de eventos. A pré-visualização está desligada por predefinição; reconhecimento e eventos continuam sem ela. Quando ativa, a Web UI desenha pessoas registadas a verde e desconhecidas a laranja a partir de `/state` sobre imagens brutas.
+
+O Monitor funciona independentemente do navegador e tarefas ativas são restauradas ao reiniciar o servidor. A configuração fica no SQLite e credenciais RTSP encriptadas em `/data`, mas fotogramas e eventos não são guardados. Eventos ficam apenas num buffer de memória limitado. O descodificador mantém o último fotograma e ignora os antigos em vez de os acumular.
+
+## 6. Dados e segurança
+
+Mantenha `/data` persistente e `/models` só de leitura. Antes de operações em massa, copie SQLite e face crops juntos. As chaves são guardadas como hash; iniciar o mesmo volume com outro `INSIGHTFACE_API_KEY` roda a chave ativa. Não registe imagens, embeddings nem chaves.
+
+O explorador de esquema OpenAPI para programadores está em `/docs`; as instruções práticas da API estão nesta ajuda. Inclua `x-request-id` ao comunicar problemas. `401` indica chave, `409 collection_model_mismatch` contrato do modelo e `422 face_not_found` ausência de rosto utilizável.
+
+## 7. Modelos e licenças
+
+As imagens não incluem modelos. O arranque normal fica offline; o serviço
+pontual `models` instala em `server/.models`:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+São suportados `buffalo_l` (`det_10g.onnx` + `w600k_r50.onnx`), `buffalo_m`,
+`buffalo_sc` e `antelopev2`. A instalação cria `manifest.json` e
+`MODEL.LICENSE` assinada. Sem `--accept-license`, a ferramenta mostra os termos
+e termina sem descarregar. Os modelos públicos pré-treinados do InsightFace são
+apenas para investigação não comercial sem licença comercial separada.
+
+## 8. Configuração de arranque e pesquisa
+
+`server/config/server.toml` é lido uma vez ao iniciar; alterações requerem
+reinício. Predefinições: `input_sizes=[[96,96],[512,512]]`, limiar de deteção
+`0.50`, NMS `0.40`, `single_face_selection="largest"` e máximo 100 rostos.
+SCRFD executa cada resolução, converte candidatos para a imagem original e faz
+um único NMS global. `max_concurrency="auto"` significa CPU 4 e CUDA 8.
+`[web].disabled=true` mantém apenas `/v1` e `/openapi.json`.
+
+System anuncia somente perfis disponíveis. O perfil fica fixo quando se cria a
+Collection e não pode mudar por pedido:
+
+- `fp32_v1`: CPU/CUDA padrão;
+- `fp16_v1`: CUDA;
+- `bf16_v1`: CPU compatível ou CUDA SM80+;
+- `int8_x736_v1`: INT8 recomendado CPU/CUDA, acumulação INT32;
+- `int8_x1000_v1`: compatibilidade com Collections existentes.
+
+Todos percorrem cada FaceSample, não são índices ANN e devolvem raw cosine.
+`capacity_rows=100000`, limite global `10000000` e
+`max_faces_per_person=20`. Em 512 dimensões, só o vetor ocupa aproximadamente
+2 048 bytes FP32, 1 024 FP16/BF16 ou 512 INT8 por linha.
+
+## 9. SDK, compilação e operação de dados
+
+O SDK Python aceita caminho, bytes e file-like object e fornece métodos tipados
+para Detect, Compare, Collections, registo, Search e Monitors. Consulte o
+contrato HTTP no [guia da API](api.pt.md).
+
+Qualquer utilizador pode compilar a partir do repositório completo:
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Use `--pull never` no Compose para a imagem local. Os tags imutáveis são
+`0.2.0-cpu` e `0.2.0-cuda12`; `cpu` e `cuda12` apontam à última versão estável,
+sem tag `latest`. Antes de atualizar, pare escritas e faça backup SQLite-safe de
+`/data` e crops. Não use `docker compose down -v`, pois elimina o volume.
+
+## 10. GPU, rede e resolução de problemas
+
+A imagem CUDA contém CUDA Runtime 12.9.1, cuDNN 9.24.0 e
+`onnxruntime-gpu==1.27.0`. Turing/Ampere/Ada/Hopper exigem R535+, Blackwell e
+RTX 50 exigem 570.26+; recomenda-se R580 estável ou superior. No arranque são
+verificados GPU, Compute Capability, Driver, CUDA/cuDNN/ORT, Provider, Sessions
+reais e warm-up; fallback silencioso para CPU é recusado.
+
+Ao expor a rede, termine HTTPS num reverse proxy de confiança, limite origins
+CORS, rate, body e timeout, e proteja `/data` e backups como dados biométricos.
+Não registe imagens, embeddings ou chaves. A fase um tem uma única API Key sem
+roles e não é autorização multi-tenant.
diff --git a/server/docs/user-guide.ru.md b/server/docs/user-guide.ru.md
new file mode 100644
index 000000000..fa6906dec
--- /dev/null
+++ b/server/docs/user-guide.ru.md
@@ -0,0 +1,141 @@
+# Руководство пользователя InsightFace Server
+
+**Языки:** [English](user-guide.md) · [中文](user-guide.zh-CN.md) · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · Русский · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+Это пошаговое руководство ведёт нового пользователя от пустого каталога до первого успешного поиска. Те же операции доступны через Web UI, `/v1` и Python SDK. Все HTTP-поля и результаты приведены в [руководстве по API](api.ru.md).
+
+## От нуля до первого поиска
+
+Для CPU нужны Linux x86_64, Docker Engine и Docker Compose. Для CUDA также нужны совместимый драйвер NVIDIA и NVIDIA Container Toolkit; устанавливать CUDA, cuDNN, ORT, Python или OpenCV на хост не требуется.
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+Для GPU используйте `compose.cuda12.yml` и порт `18098`. Перед загрузкой установщик показывает лицензию; публичные модели InsightFace разрешены только для некоммерческих исследований без отдельной коммерческой лицензии.
+
+В поставляемом Compose аутентификация по умолчанию отключена для изолированной оценки. Перед сетевым доступом задайте `INSIGHTFACE_AUTH_ENABLED=true` и длинный `INSIGHTFACE_API_KEY`. Затем проверьте Dashboard, создайте Collection, зарегистрируйте Person и выполните поиск по другому фото. Останавливайте через `docker compose ... down` без `-v`, чтобы сохранить volume.
+
+## 1. Вход и готовность
+
+Для CPU откройте `http://СЕРВЕР:18097/`, для CUDA 12 — `http://СЕРВЕР:18098/`. Если включена аутентификация, выберите **Настроить API-ключ**, вставьте ключ оператора и примените его к вкладке. Он хранится только в памяти и удаляется после обновления или закрытия.
+
+В разделах **Панель** или **Система** убедитесь, что сервис, база, модели и Provider готовы. CUDA должен показывать `CUDAExecutionProvider` и не переключается на CPU без ошибки.
+
+## 2. Создание Collection
+
+В **Коллекции** → **Новая коллекция** задайте стабильный ID, имя, cosine-порог
+(`0.4` сначала), доступный профиль, ёмкость и максимум FaceSamples на человека.
+Сохранение JPEG `bounding-box crop`, масштабированного до 112×112, по умолчанию
+выключено; это не выровненный вход модели распознавания.
+
+Collection фиксирует ID, версию, digest, размерность и предобработку модели. После смены модели старая Collection видима, но регистрация и поиск отклоняются, если контракт не совпадает.
+
+Профиль детекции при создании копирует системные значения, после чего можно менять входные размеры, пороги детекции/NMS и стратегию одного лица. `largest` выбирает по площади; `center_largest` максимизирует `площадь - 2,0 × квадрат расстояния в пикселях между центром рамки и центром изображения`. Уверенность детектора в этот балл не входит.
+
+## 3. Регистрация Person
+
+В **Люди** выберите Collection и **Зарегистрировать человека**. Можно указать ID, имя, внешний ID, JSON metadata и несколько JPEG, PNG или WebP.
+
+- `off`: использует стратегию одного лица Collection; несколько лиц разрешены;
+- `standard`: требует одно пригодное лицо и проверяет размер, детекцию, резкость, яркость и позу;
+- `strict`: дополнительно требует, чтобы лучшая внутриклассовая similarity была выше лучшей similarity с другим человеком.
+
+Пакет может завершиться частично и сообщает причину каждого отказа. Оригиналы не сохраняются. `external_trusted` принимает L2-нормализованный embedding; изображение всё равно нужно для детекции и качества, но вектор повторно не извлекается.
+
+## 4. Детекция, сравнение и поиск
+
+**Детекция** показывает рамки, пять точек, score и качество; отсутствие лиц — успешный пустой список. **Сравнение** выбирает по одному лицу с системным профилем или профилем Collection и возвращает cosine `similarity`, `threshold`, `matched`. Similarity не является вероятностью.
+
+В **Поиск** выберите Collection и изображение. Оценка человека — максимум по его FaceSamples. Результаты отсортированы по убыванию; нет совпадений — пустой список. Образец сначала фиксируется в SQLite, затем до ответа добавляется в индекс. После перезапуска индекс строится из SQLite.
+
+## 5. Мониторинг RTSP-камеры
+
+В **Мониторинге камер** создайте постоянный Monitor и настройте RTSP-источник, Collection, частоту, необязательный порог и правила событий. Просмотр по умолчанию выключен; распознавание и события работают без него. При включении Web UI рисует зарегистрированных зелёным, неизвестных оранжевым по `/state` поверх сырых кадров.
+
+Monitor работает независимо от браузера, включённые задачи восстанавливаются после перезапуска сервера. Конфигурация хранится в SQLite, реквизиты RTSP зашифрованы в `/data`, но кадры и события не сохраняются. События живут только в ограниченном буфере памяти. Декодер держит последний кадр и пропускает устаревшие вместо накопления очереди.
+
+## 6. Данные и безопасность
+
+Сохраняйте `/data`, подключайте `/models` только для чтения. Перед массовыми изменениями копируйте SQLite и область face crops вместе. Ключи хранятся как hash; запуск того же volume с новым `INSIGHTFACE_API_KEY` меняет активный ключ. Не пишите изображения, embeddings и ключи в журнал.
+
+Обозреватель схемы OpenAPI для разработчиков доступен по `/docs`; практические инструкции API находятся в этой справке. Для диагностики сообщайте `x-request-id`. `401` означает проблему ключа, `409 collection_model_mismatch` — контракта модели, `422 face_not_found` — отсутствия пригодного лица.
+
+## 7. Модели и лицензии
+
+Образы не содержат моделей. Обычный запуск не требует сети; одноразовый сервис
+`models` устанавливает пакет в `server/.models`:
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+Поддерживаются `buffalo_l` (`det_10g.onnx` + `w600k_r50.onnx`), `buffalo_m`,
+`buffalo_sc` и `antelopev2`. Установка создаёт `manifest.json` и подписанный
+`MODEL.LICENSE`. Без `--accept-license` инструмент показывает условия и
+завершается без загрузки. Публичные предобученные модели InsightFace
+разрешены только для некоммерческих исследований без отдельной коммерческой
+лицензии.
+
+## 8. Настройки запуска и поиск
+
+`server/config/server.toml` читается один раз при старте; изменения требуют
+перезапуска. По умолчанию: `input_sizes=[[96,96],[512,512]]`, порог детектора
+`0.50`, NMS `0.40`, `single_face_selection="largest"` и максимум 100 лиц.
+SCRFD запускает каждое разрешение, переносит кандидатов в координаты оригинала
+и выполняет один общий NMS. `max_concurrency="auto"` означает CPU 4 и CUDA 8.
+`[web].disabled=true` оставляет только `/v1` и `/openapi.json`.
+
+System показывает лишь доступные профили. Профиль фиксируется при создании
+Collection и не выбирается в отдельном Search:
+
+- `fp32_v1`: стандарт CPU/CUDA;
+- `fp16_v1`: CUDA;
+- `bf16_v1`: поддерживаемый CPU или CUDA SM80+;
+- `int8_x736_v1`: рекомендуемый INT8 CPU/CUDA, накопление INT32;
+- `int8_x1000_v1`: совместимость старых Collections.
+
+Все профили просматривают каждый FaceSample и не являются ANN-индексами;
+публичный score остаётся raw cosine. Значения по умолчанию:
+`capacity_rows=100000`, общий предел `10000000`,
+`max_faces_per_person=20`. Для 512 измерений только вектор занимает примерно
+2048 байт FP32, 1024 FP16/BF16 или 512 INT8 на строку.
+
+## 9. SDK, сборка и работа с данными
+
+Python SDK принимает путь, bytes и file-like object и предоставляет
+типизированные Detect, Compare, Collections, регистрацию, Search и Monitors.
+Полный HTTP-контракт находится в [руководстве по API](api.ru.md).
+
+Пользователь может собрать образы из полного репозитория:
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+Для локального образа добавляйте `--pull never` к Compose. Неизменяемые теги:
+`0.2.0-cpu` и `0.2.0-cuda12`; движущиеся `cpu` и `cuda12` указывают на
+последнюю стабильную версию, а `latest` не публикуется. Перед обновлением
+остановите запись и сделайте SQLite-safe копию `/data` вместе с crops.
+Не используйте `docker compose down -v`: команда удаляет volume данных.
+
+## 10. GPU, сеть и устранение неполадок
+
+CUDA-образ содержит CUDA Runtime 12.9.1, cuDNN 9.24.0 и
+`onnxruntime-gpu==1.27.0`. Turing/Ampere/Ada/Hopper требуют R535+, Blackwell и
+RTX 50 — 570.26+; для новых установок рекомендуется стабильный R580+.
+При старте проверяются GPU, Compute Capability, Driver, CUDA/cuDNN/ORT,
+Provider, настоящие model Sessions и warm-up; скрытый переход на CPU запрещён.
+
+При сетевой публикации завершайте HTTPS на доверенном reverse proxy,
+ограничивайте CORS origins, rate, body и timeout и защищайте `/data` и backup
+как биометрические данные. Не журналируйте изображения, embeddings и ключи.
+В первой фазе есть один API Key без ролей — это не multi-tenant авторизация.
diff --git a/server/docs/user-guide.zh-CN.md b/server/docs/user-guide.zh-CN.md
new file mode 100644
index 000000000..66cc64afd
--- /dev/null
+++ b/server/docs/user-guide.zh-CN.md
@@ -0,0 +1,251 @@
+# InsightFace Server 用户指南
+
+**语言：** [English](user-guide.md) · 中文 · [日本語](user-guide.ja.md) · [Deutsch](user-guide.de.md) · [Español](user-guide.es.md) · [Français](user-guide.fr.md) · [Русский](user-guide.ru.md) · [Português](user-guide.pt.md) · [한국어](user-guide.ko.md)
+
+这是面向第一次使用者的分步操作指南：从一个空的项目目录开始，直到创建人员库、
+注册人员并得到第一次搜索结果。相同能力可以通过 Web UI、`/v1` API 和 Python SDK
+使用；每个HTTP字段和响应的完整说明请查看
+[API使用手册](api.zh-CN.md)。
+
+## 从这里开始：从零启动到第一次成功搜索
+
+CPU版需要Linux x86_64、Docker Engine和Docker Compose。CUDA版还需要兼容的
+NVIDIA Driver与NVIDIA Container Toolkit；宿主机不需要安装CUDA、cuDNN、
+ONNX Runtime、Python或OpenCV。
+
+CPU启动示例：
+
+```bash
+mkdir -p server/.models
+docker compose -f server/deploy/compose.cpu.yml pull
+docker compose -f server/deploy/compose.cpu.yml run --rm models install buffalo_l
+docker compose -f server/deploy/compose.cpu.yml up -d
+curl -fsS http://127.0.0.1:18097/v1/health
+```
+
+GPU版把`compose.cpu.yml`替换成`compose.cuda12.yml`，健康检查端口改为`18098`。
+模型安装器会在下载前展示许可。InsightFace公开预训练模型默认仅允许非商业研究，
+商业使用需要单独授权。
+
+随项目提供的Compose配置在隔离评估环境中默认`auth_enabled=false`，此时API不用传
+认证字段，Web UI也会隐藏API Key输入。对其他用户或网络开放前，应在首次启动前启用：
+
+```bash
+export INSIGHTFACE_AUTH_ENABLED=true
+export INSIGHTFACE_API_KEY='replace-with-a-long-random-secret'
+docker compose -f server/deploy/compose.cpu.yml up -d
+```
+
+CPU访问`http://服务器地址:18097/`，GPU访问`http://服务器地址:18098/`。第一次操作
+按以下顺序完成：确认仪表盘全部就绪、创建Collection、用至少一张清晰图片注册Person、
+再用该人员的另一张图片执行Search。没有匹配时返回空列表，这是正常成功结果。
+停止服务使用`docker compose ... down`且不要加`-v`；`-v`会永久删除命名数据卷。
+
+## 1. 登录并检查就绪状态
+
+CPU 打开 `http://服务器地址:18097/`，CUDA 12 打开 `http://服务器地址:18098/`。如果启用了认证，点击 **配置 API Key**，粘贴管理员提供的 Key，再选择 **在此标签页使用**。Key 只保留在当前标签页内存中，刷新或关闭页面后即清除。
+
+注册数据前请查看 **仪表盘** 或 **系统**。服务、数据库、模型和 Provider 均应为就绪。CUDA 部署必须显示 `CUDAExecutionProvider`，不会静默回退到 CPU。
+
+## 2. 创建 Collection
+
+打开 **人员库**，选择 **新建人员库**，设置：
+
+- 稳定 ID，例如 `employees`；
+- 展示名称、描述和可选 metadata；
+- 默认 cosine 阈值，初始建议 `0.4`；
+- 当前主机支持的 search profile；
+- 容量和每个 Person 最多 FaceSample 数；
+- 检测输入尺寸、检测/NMS 阈值以及单脸挑选策略；
+- 是否保存缩放为 112×112 的 `bounding-box crop` JPEG；它不是识别模型使用的
+  对齐输入，默认关闭。
+
+Collection 会固定绑定模型 ID、版本、digest、特征维度和预处理版本。检测配置在创建时复制系统默认值，之后可以单独修改；修改从下一次请求生效并递增 `detection_revision`，但不会重新处理已有 FaceSample。`largest` 优先面积；`center_largest` 最大化 `人脸面积 - 2.0 × 人脸框中心到图像中心的像素距离平方`，检测置信度不参与该分数。
+
+## 3. 注册 Person
+
+打开 **人员**，选择 Collection，再点击 **注册人员**。可填写稳定的 Person ID、姓名、外部 ID 和 JSON metadata，然后拖入一张或多张 JPEG、PNG 或 WebP 图片。
+
+入库审查模式：
+
+- `off`：使用 Collection 的单脸挑选策略，允许图片中存在多张脸；
+- `standard`：要求一张可用脸，并检查尺寸、检测分数、清晰度、亮度和姿态；
+- `strict`：在 standard 基础上，要求样本的最佳类内相似度高于最佳类外相似度。
+
+批量注册支持部分成功。请根据每张失败图片的原因处理后重试；系统不保存被拒绝的
+原图。启用人脸图保存时，只保存缩放为 112×112 的 `bounding-box crop`，不保存
+原始上传图片或识别模型使用的对齐输入。
+
+可信系统可以用 `external_trusted` 提交预先抽取并 L2 归一化的 embedding。仍须同时提供图片完成检测和质量审查，但服务不会再次抽取特征；embedding contract 必须与 Collection 完全一致。
+
+## 4. 检测与比对
+
+在 **检测** 中上传单图，可查看人脸框、五点关键点、检测分数和启发式质量信息。无人脸是成功的空列表。
+
+在 **比对** 中分别上传 source 和 target，并可选择系统或 Collection 检测配置。配置中的策略从两张图各挑选一张可用脸，返回原始 cosine `similarity`、`threshold` 和 `matched`。Similarity 不是概率；任一图片没有可用脸时返回 `422 face_not_found`。
+
+## 5. 搜索人员库
+
+打开 **搜索**，选择 Collection，上传查询图片并设置返回数量；也可以临时覆盖阈值。系统按 Collection 检测配置挑选查询脸，按相似度降序返回。Person 得分取其所有 FaceSample 的最高相似度。无匹配是成功的空列表。
+
+新 FaceSample 会先提交到 SQLite，再加入内存索引，然后才返回成功；删除同时更新两处。重启时从 SQLite 重建索引，SQLite 始终是权威数据源。
+
+## 6. RTSP 摄像头监控
+
+打开 **摄像头监控**，点击 **新建监控任务**。填写任务ID和名称，输入`rtsp://`或
+`rtsps://`地址，选择Collection，并设置每秒推理次数和可选匹配阈值。事件策略可以
+设置连续多少帧后确认、离开超时、重复事件冷却时间，以及内存中保留的最近事件数量。
+
+**Web视频预览默认关闭。** 只有管理员需要查看画面时才开启；不开预览也会持续识别
+和生成事件。开启后服务器传输原始JPEG帧，Web UI依据`/state`结果绘制标注：绿色框
+表示已入库人员，橙色框表示检测到但未入库的人脸。
+
+Monitor独立运行在服务器端，关闭浏览器不会停止；处于启用状态的任务会在Server重启
+后自动恢复。使用 **启动/停止** 修改`enabled`，使用 **编辑** 更换RTSP源或调整参数，
+使用 **删除** 永久移除任务。解码器只保留最新帧；推理耗时超过设定周期时直接跳过
+过时帧，不会排队补跑。
+
+Monitor配置保存在SQLite中，RTSP凭据加密保存在`/data`且API不会回传。视频帧不会
+保存；进入、离开、错误和恢复事件只保留在有上限的内存环形缓冲区，进程重启后丢失。
+Web UI/API跨越不可信网络时应使用HTTPS，并只允许可信管理员管理Monitor。
+
+## 7. 修改与删除
+
+可在列表中修改 Collection 和 Person。删除 FaceSample 会同时删除 embedding 和可选裁剪图。删除非空 Collection 需要明确确认 `force`。批量或破坏性操作前先备份 `/data`。
+
+## 8. API 与 Python SDK
+
+面向开发者的 OpenAPI Schema 浏览器位于 `/docs`；任务式 API 使用说明就在本帮助中。每个响应都带 `x-request-id`，报告问题时请一并提供。
+
+```python
+from insightface_server import Client
+
+client = Client("http://localhost:18097", api_key="your-key")
+client.create_collection(collection_id="employees", name="员工库", threshold=0.4)
+client.add_person("employees", person_id="alice", images=["alice-1.jpg", "alice-2.jpg"])
+matches = client.search("employees", "query.jpg", limit=5)
+```
+
+## 9. 数据、备份与安全
+
+- 持久化挂载 `/data`，`/models` 只读挂载。
+- 停止写入后备份 SQLite 和裁剪图目录，或使用 SQLite 安全快照方式。
+- API Key 只以 hash 保存。后续启动同一数据卷时传入不同 `INSIGHTFACE_API_KEY`，会主动轮换当前 Key。
+- 不要记录图片、embedding 或 Key；除非确有需要，不要开启宽泛 CORS。
+- 公开镜像不包含模型。InsightFace提供的开源预训练模型（包括 `buffalo_l`）仅限
+  非商业研究使用；商业使用需要单独许可，请访问 <https://www.insightface.ai>。
+  **系统** 页面也会显示相同提示。
+
+## 10. 故障定位
+
+`401 unauthorized` 表示当前标签页未配置 Key 或 Key 已轮换。`409 collection_model_mismatch` 表示 Collection 与当前模型契约不同。`422 face_not_found` 表示没有选出可用脸。CUDA 模式在 Driver、GPU、模型 Session、Provider 或 warm-up 检查失败时会主动终止。请查看 **系统**、容器日志和响应中的 `request_id`。
+
+## 11. 模型与模型许可
+
+镜像不包含模型。一次性的`models`工具把模型安装到`server/.models`，正常Server
+启动无需联网：
+
+```bash
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models install buffalo_l --accept-license
+docker compose -f server/deploy/compose.cpu.yml \
+  run --rm models verify buffalo_l
+```
+
+公开包包括：`buffalo_l`（`det_10g.onnx` + `w600k_r50.onnx`）、
+`buffalo_m`（`det_2.5g.onnx` + `w600k_r50.onnx`）、
+`buffalo_sc`（`det_500m.onnx` + `w600k_mbf.onnx`）和`antelopev2`
+（`scrfd_10g_bnkps.onnx` + `glintr100.onnx`）。安装会生成`manifest.json`
+与签名的`MODEL.LICENSE`。不带`--accept-license`时，工具只显示许可并退出，不会
+下载。`models verify`会核验包身份、签名、有效期和当前授权状态。
+
+InsightFace公开预训练模型默认仅限非商业研究；商业使用需另行获取授权。私有模型也
+可以使用同样的manifest和离线签名许可。许可按`model_id`表达授权，是合规凭证，
+不是DRM，也不要求模型文件SHA-256保持不变。
+
+## 12. 仅启动时生效的配置
+
+通用配置文件为`server/config/server.toml`，Compose将其只读挂载到
+`/etc/insightface/server.toml`。修改后必须重启容器，默认值如下：
+
+```toml
+[inference]
+max_concurrency = "auto" # CPU为4，CUDA为8
+
+[detection]
+input_sizes = [[96, 96], [512, 512]]
+threshold = 0.50
+nms_threshold = 0.40
+single_face_selection = "largest"
+max_detected_faces = 100
+
+[web]
+disabled = false
+```
+
+动态SCRFD会分别运行所有分辨率，把候选框映射回原图后合并，并只执行一次全局NMS。
+系统配置只在启动时读取，不提供运行时修改API。新Collection会复制系统检测配置，
+之后可独立修改并从下一次请求生效。无状态Detect和Embeddings使用系统配置；Compare
+可使用系统配置或指定Collection；注册与Search始终使用Collection配置。
+
+将`[web].disabled=true`可只启动API。此时`/v1`和`/openapi.json`仍可用，但不会
+注册`/`、`/docs`、帮助文档和前端静态资源。
+
+## 13. 精确检索Profile与容量
+
+**系统**接口只公布当前CPU/GPU真正可用的Profile。Collection在创建时固定Profile，
+不能在单次Search请求中临时切换。
+
+| Profile | 存储类型 | 常见可用环境 |
+| --- | --- | --- |
+| `fp32_v1` | FP32 | CPU与CUDA |
+| `fp16_v1` | FP16 | CUDA |
+| `bf16_v1` | BF16 | 支持的CPU或SM80+ CUDA |
+| `int8_x736_v1` | INT8，scale 736 | CPU与CUDA；推荐INT8 |
+| `int8_x1000_v1` | INT8，scale 1000 | 兼容已有Collection |
+
+这些实现都会遍历全部有效FaceSample，属于Flat精确全量搜索，不是ANN索引。低精度
+Profile会近似FP32分数；INT8点积使用INT32累加。对外相似度和阈值始终是原始cosine。
+
+`capacity_rows`预留该Collection的最大有效行数，避免常规扩容停顿。512维向量的
+大致纯特征占用为：FP32每行2,048字节，FP16/BF16每行1,024字节，INT8每行512字节，
+还需额外计算ID与工作区。默认容量`100000`，部署级上限默认`10000000`。
+`max_faces_per_person`默认`20`，限制单人样本数，不限制Person数量。
+
+## 14. CUDA支持与严格启动检查
+
+CUDA镜像包含CUDA Runtime 12.9.1、cuDNN 9.24.0、Python 3.11和
+`onnxruntime-gpu==1.27.0`。宿主机只需Driver、Docker Engine、NVIDIA Container
+Toolkit和兼容GPU。
+
+- Turing、Ampere、Ada、Hopper：Driver R535或更高；
+- Blackwell与RTX 50系列：Driver 570.26或更高；
+- 新部署建议使用稳定的R580或更高版本。
+
+架构兼容不等于所有GPU型号都已经正式认证。每次CUDA启动都会核验GPU型号、
+Compute Capability、Driver、实际CUDA/cuDNN/ORT版本、`CUDAExecutionProvider`、
+真实检测与识别Session以及真实warm-up推理，并审计Provider分配。任何关键检查失败
+都会终止启动，不会静默回退CPU。使用前请在 **系统** 页面确认结果。
+
+## 15. 构建、升级、备份与恢复
+
+用户可从完整仓库自行构建：
+
+```bash
+make -C server build-cpu
+make -C server build-cuda12
+```
+
+随后在Compose的模型安装与`up`命令中加入`--pull never`，即可使用本地镜像。构建
+使用固定基础镜像和锁定依赖，但仍需联网获取这些输入。公开版本Tag为
+`0.2.0-cpu`和`0.2.0-cuda12`；移动Tag `cpu`/`cuda12`分别指向最新稳定版本，
+明确不发布含义模糊的`latest`。
+
+升级前停止写入，使用SQLite安全方式备份`/data`以及可选裁剪图，并保留`/models`
+和许可文件。先用数据副本启动新镜像，检查migration、`/v1/health`、模型契约和一条
+已知Search，再切换正式数据。停止使用`docker compose down`且不要带`-v`；
+`docker compose down -v`会删除命名数据卷。
+
+跨网络使用时，应在可信反向代理终止HTTPS，只开放必要的CORS origin，并在边缘限制
+速率、请求体和超时。数据卷及备份应按生物识别数据保护。第一阶段只有一个不区分权限
+的API Key，不应把它当作多租户授权系统。
diff --git a/server/frontend/README.md b/server/frontend/README.md
new file mode 100644
index 000000000..e08a56123
--- /dev/null
+++ b/server/frontend/README.md
@@ -0,0 +1,28 @@
+# InsightFace Server Web UI
+
+This directory is a build-free, same-origin browser client for the public
+`/v1` REST API. The server mounts these files at `/assets`, serves `index.html`
+at `/`, and serves `openapi.html` at `/docs`. The Help area renders the bundled
+User Guide or complete API Guide in all nine supported languages, plus the
+English-only Maintainer Guide. A document switcher displays one source at a
+time with its own searchable table of contents. These are the same Markdown
+files linked from GitHub; `/docs` remains the live, schema-derived API
+reference. Repository READMEs are not served by the application.
+
+The UI uses only browser-native HTML, CSS, JavaScript modules, images, and canvas.
+It has no package-manager, CDN, analytics, or third-party font
+dependency. The API key is held in a private JavaScript field for the lifetime
+of the current tab; it is never written to browser storage, cookies, or URLs.
+
+Camera Monitoring administers persistent server-side RTSP Monitors through the
+public `/v1` API. Closing the page never stops a task. The UI polls signed-cursor
+in-memory events and live state; optional raw MJPEG preview is off by default
+and encoded only while viewed. The browser draws green enrolled and amber
+unknown-face overlays from `/state`. It never requests local camera permission
+or stores frames, events, RTSP URLs, credentials, or API keys.
+
+Run the browser-independent checks from the repository root:
+
+```bash
+node --test server/tests/frontend/*.test.mjs
+```
diff --git a/server/frontend/api.mjs b/server/frontend/api.mjs
new file mode 100644
index 000000000..f0cd6a774
--- /dev/null
+++ b/server/frontend/api.mjs
@@ -0,0 +1,347 @@
+export class ApiError extends Error {
+  constructor({ code = "request_failed", message = "Request failed.", details = {}, status = 0, requestId = "" } = {}) {
+    super(message);
+    this.name = "ApiError";
+    this.code = code;
+    this.details = details;
+    this.status = status;
+    this.requestId = requestId;
+  }
+}
+
+export class ApiClient {
+  #apiKey = "";
+
+  constructor(baseUrl = "", { fetchFn = globalThis.fetch?.bind(globalThis) } = {}) {
+    if (!fetchFn) throw new TypeError("A fetch implementation is required");
+    this.baseUrl = String(baseUrl).replace(/\/$/, "");
+    this.fetchFn = fetchFn;
+  }
+
+  setApiKey(value) {
+    this.#apiKey = String(value ?? "").trim();
+  }
+
+  clearApiKey() {
+    this.#apiKey = "";
+  }
+
+  get hasApiKey() {
+    return Boolean(this.#apiKey);
+  }
+
+  async request(path, { method = "GET", body, query, signal, keepalive = false } = {}) {
+    const url = new URL(`${this.baseUrl}${path}`, globalThis.location?.origin ?? "http://localhost");
+    for (const [key, value] of Object.entries(query ?? {})) {
+      if (value !== undefined && value !== null && value !== "") url.searchParams.set(key, String(value));
+    }
+    const headers = new Headers({ Accept: "application/json" });
+    if (this.#apiKey) headers.set("Authorization", `Bearer ${this.#apiKey}`);
+
+    let requestBody = body;
+    if (body && !(body instanceof FormData)) {
+      headers.set("Content-Type", "application/json");
+      requestBody = JSON.stringify(body);
+    }
+
+    let response;
+    try {
+      response = await this.fetchFn(url, { method, headers, body: requestBody, signal, keepalive });
+    } catch (error) {
+      if (error?.name === "AbortError") {
+        throw new ApiError({ code: "request_cancelled", message: "The request was cancelled." });
+      }
+      throw new ApiError({ code: "network_error", message: "Could not reach the InsightFace Server." });
+    }
+
+    const requestId = response.headers.get("x-request-id") ?? "";
+    if (response.status === 204) return null;
+    const text = await response.text();
+    let payload = null;
+    if (text) {
+      try {
+        payload = JSON.parse(text);
+      } catch {
+        if (!response.ok) {
+          throw new ApiError({ status: response.status, requestId, message: "The server returned an invalid error response." });
+        }
+        throw new ApiError({ status: response.status, requestId, code: "invalid_response", message: "The server returned invalid JSON." });
+      }
+    }
+
+    if (!response.ok) {
+      const error = payload?.error ?? payload ?? {};
+      throw new ApiError({
+        code: error.code ?? "request_failed",
+        message: error.message ?? `Request failed with HTTP ${response.status}.`,
+        details: error.details ?? {},
+        status: response.status,
+        requestId: payload?.request_id ?? requestId,
+      });
+    }
+    if (payload && typeof payload === "object" && !payload.request_id && requestId) payload.request_id = requestId;
+    return payload;
+  }
+
+  async requestBlob(path, { signal } = {}) {
+    const url = new URL(`${this.baseUrl}${path}`, globalThis.location?.origin ?? "http://localhost");
+    const headers = new Headers({ Accept: "image/jpeg" });
+    if (this.#apiKey) headers.set("Authorization", `Bearer ${this.#apiKey}`);
+
+    let response;
+    try {
+      response = await this.fetchFn(url, { method: "GET", headers, signal });
+    } catch (error) {
+      if (error?.name === "AbortError") {
+        throw new ApiError({ code: "request_cancelled", message: "The request was cancelled." });
+      }
+      throw new ApiError({ code: "network_error", message: "Could not reach the InsightFace Server." });
+    }
+
+    const requestId = response.headers.get("x-request-id") ?? "";
+    if (!response.ok) {
+      let payload = null;
+      try {
+        const text = await response.text();
+        payload = text ? JSON.parse(text) : null;
+      } catch {
+        payload = null;
+      }
+      const error = payload?.error ?? payload ?? {};
+      throw new ApiError({
+        code: error.code ?? "request_failed",
+        message: error.message ?? `Request failed with HTTP ${response.status}.`,
+        details: error.details ?? {},
+        status: response.status,
+        requestId: payload?.request_id ?? requestId,
+      });
+    }
+
+    const mediaType = (response.headers.get("content-type") ?? "").split(";", 1)[0].trim().toLowerCase();
+    const blob = await response.blob();
+    if (mediaType !== "image/jpeg" || !blob.size) {
+      throw new ApiError({
+        code: "invalid_response",
+        message: "The server returned an invalid face crop.",
+        status: response.status,
+        requestId,
+      });
+    }
+    return blob;
+  }
+
+  async requestStream(path, { signal } = {}) {
+    const url = new URL(`${this.baseUrl}${path}`, globalThis.location?.origin ?? "http://localhost");
+    const headers = new Headers({ Accept: "multipart/x-mixed-replace" });
+    if (this.#apiKey) headers.set("Authorization", `Bearer ${this.#apiKey}`);
+    let response;
+    try {
+      response = await this.fetchFn(url, { method: "GET", headers, signal });
+    } catch (error) {
+      if (error?.name === "AbortError") throw error;
+      throw new ApiError({ code: "network_error", message: "Could not reach the InsightFace Server." });
+    }
+    if (!response.ok) {
+      const requestId = response.headers.get("x-request-id") ?? "";
+      let payload = null;
+      try {
+        const text = await response.text();
+        payload = text ? JSON.parse(text) : null;
+      } catch {
+        payload = null;
+      }
+      const error = payload?.error ?? payload ?? {};
+      throw new ApiError({
+        code: error.code ?? "request_failed",
+        message: error.message ?? `Request failed with HTTP ${response.status}.`,
+        details: error.details ?? {},
+        status: response.status,
+        requestId: payload?.request_id ?? requestId,
+      });
+    }
+    return response;
+  }
+
+  health() { return this.request("/v1/health"); }
+  system() { return this.request("/v1/system"); }
+  models() { return this.request("/v1/models"); }
+
+  detect(image, { maxFaces, collectionId } = {}) {
+    const form = new FormData();
+    form.append("image", image);
+    if (maxFaces !== undefined && maxFaces !== "") form.append("max_faces", String(maxFaces));
+    if (collectionId) form.append("collection_id", collectionId);
+    return this.request("/v1/detect", { method: "POST", body: form });
+  }
+
+  compare(source, target, { threshold, collectionId } = {}) {
+    const form = new FormData();
+    form.append("source", source);
+    form.append("target", target);
+    if (threshold !== undefined && threshold !== "") form.append("threshold", String(threshold));
+    if (collectionId) form.append("collection_id", collectionId);
+    return this.request("/v1/compare", { method: "POST", body: form });
+  }
+
+  embeddings(image, { collectionId } = {}) {
+    const form = new FormData();
+    form.append("image", image);
+    if (collectionId) form.append("collection_id", collectionId);
+    return this.request("/v1/embeddings", { method: "POST", body: form });
+  }
+
+  createCollection(input) { return this.request("/v1/collections", { method: "POST", body: input }); }
+  listCollections(query) { return this.request("/v1/collections", { query }); }
+  getCollection(id) { return this.request(`/v1/collections/${encodeURIComponent(id)}`); }
+  updateCollection(id, input) { return this.request(`/v1/collections/${encodeURIComponent(id)}`, { method: "PATCH", body: input }); }
+  deleteCollection(id, force = false) {
+    return this.request(`/v1/collections/${encodeURIComponent(id)}`, { method: "DELETE", query: { force } });
+  }
+
+  createPerson(collectionId, {
+    id,
+    name,
+    externalId,
+    metadata,
+    images,
+    reviewMode = "off",
+    embeddingMode = "server",
+    externalEmbeddings,
+    embeddingContractId,
+  }) {
+    const form = new FormData();
+    if (id) form.append("id", id);
+    if (name) form.append("name", name);
+    if (externalId) form.append("external_id", externalId);
+    form.append("metadata", JSON.stringify(metadata ?? {}));
+    form.append("review_mode", reviewMode);
+    form.append("embedding_mode", embeddingMode);
+    if (embeddingMode === "external_trusted") {
+      form.append("external_embeddings", JSON.stringify(externalEmbeddings));
+      form.append("embedding_contract_id", embeddingContractId ?? "");
+    }
+    for (const image of images ?? []) form.append("images", image);
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons`, { method: "POST", body: form });
+  }
+
+  listPeople(collectionId, query) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons`, { query });
+  }
+  getPerson(collectionId, personId) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}`);
+  }
+  updatePerson(collectionId, personId, input) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}`, { method: "PATCH", body: input });
+  }
+  deletePerson(collectionId, personId) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}`, { method: "DELETE" });
+  }
+  addFaces(collectionId, personId, images, {
+    reviewMode = "off",
+    embeddingMode = "server",
+    externalEmbeddings,
+    embeddingContractId,
+  } = {}) {
+    const form = new FormData();
+    form.append("review_mode", reviewMode);
+    form.append("embedding_mode", embeddingMode);
+    if (embeddingMode === "external_trusted") {
+      form.append("external_embeddings", JSON.stringify(externalEmbeddings));
+      form.append("embedding_contract_id", embeddingContractId ?? "");
+    }
+    for (const image of images ?? []) form.append("images", image);
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}/faces`, { method: "POST", body: form });
+  }
+  listFaces(collectionId, personId, query) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}/faces`, { query });
+  }
+  deleteFace(collectionId, personId, faceId) {
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}/faces/${encodeURIComponent(faceId)}`, { method: "DELETE" });
+  }
+  getFaceCrop(collectionId, personId, faceId) {
+    return this.requestBlob(`/v1/collections/${encodeURIComponent(collectionId)}/persons/${encodeURIComponent(personId)}/faces/${encodeURIComponent(faceId)}/image`);
+  }
+
+  search(collectionId, image, { limit = 5, threshold } = {}) {
+    const form = new FormData();
+    form.append("image", image);
+    form.append("limit", String(limit));
+    if (threshold !== undefined && threshold !== "") form.append("threshold", String(threshold));
+    return this.request(`/v1/collections/${encodeURIComponent(collectionId)}/search`, { method: "POST", body: form });
+  }
+
+  createMonitor({
+    id,
+    name,
+    description = "",
+    enabled = true,
+    url,
+    collectionId,
+    inferenceFps = 2,
+    matchThreshold,
+    eventBufferSize = 1000,
+    confirmFrames = 3,
+    absenceTimeoutSeconds = 3,
+    cooldownSeconds = 10,
+    emitUnknown = true,
+    previewEnabled = false,
+  } = {}) {
+    const body = {
+      id,
+      name,
+      description,
+      enabled,
+      source: { type: "rtsp", url },
+      collection_id: collectionId,
+      inference_fps: Number(inferenceFps),
+      match_threshold: matchThreshold === undefined || matchThreshold === null || matchThreshold === ""
+        ? null
+        : Number(matchThreshold),
+      event_buffer_size: Number(eventBufferSize),
+      event_policy: {
+        confirm_frames: Number(confirmFrames),
+        absence_timeout_seconds: Number(absenceTimeoutSeconds),
+        cooldown_seconds: Number(cooldownSeconds),
+        emit_unknown: Boolean(emitUnknown),
+      },
+      preview_enabled: Boolean(previewEnabled),
+    };
+    return this.request("/v1/monitors", { method: "POST", body });
+  }
+
+  listMonitors(query) {
+    return this.request("/v1/monitors", { query });
+  }
+
+  getMonitor(monitorId) {
+    return this.request(`/v1/monitors/${encodeURIComponent(monitorId)}`);
+  }
+
+  updateMonitor(monitorId, input) {
+    return this.request(`/v1/monitors/${encodeURIComponent(monitorId)}`, {
+      method: "PATCH",
+      body: input,
+    });
+  }
+
+  deleteMonitor(monitorId) {
+    return this.request(`/v1/monitors/${encodeURIComponent(monitorId)}`, {
+      method: "DELETE",
+    });
+  }
+
+  monitorState(monitorId) {
+    return this.request(`/v1/monitors/${encodeURIComponent(monitorId)}/state`);
+  }
+
+  monitorEvents(monitorId, query) {
+    return this.request(`/v1/monitors/${encodeURIComponent(monitorId)}/events`, { query });
+  }
+
+  monitorPreview(monitorId, { signal } = {}) {
+    return this.requestStream(
+      `/v1/monitors/${encodeURIComponent(monitorId)}/preview.mjpeg`,
+      { signal },
+    );
+  }
+}
diff --git a/server/frontend/app.mjs b/server/frontend/app.mjs
new file mode 100644
index 000000000..3dd02624c
--- /dev/null
+++ b/server/frontend/app.mjs
@@ -0,0 +1,2033 @@
+import { ApiClient, ApiError } from "./api.mjs?v=0.2.0-r13";
+import { initializeI18n, locale, t, translateTree } from "./i18n.mjs?v=0.2.0-r13";
+import { renderMarkdown } from "./markdown.mjs?v=0.2.0-r13";
+import {
+  applySearchProfileAvailability,
+  authenticationEnabledFromHealth,
+  faceLandmarks,
+  facePixels,
+  filterPeople,
+  formatCosine,
+  formatDuration,
+  formatScore,
+  listItems,
+  parseExternalEmbeddings,
+  parseMetadata,
+  searchProfilesFromSystem,
+} from "./core.mjs?v=0.2.0-r13";
+
+const client = new ApiClient(window.location.origin);
+const state = {
+  route: "dashboard",
+  collections: [],
+  people: [],
+  selectedCollection: "",
+  selectedPerson: null,
+  system: null,
+  models: [],
+  errors: [],
+  serverErrors: [],
+  monitors: [],
+  selectedMonitorId: "",
+  editingMonitorId: "",
+  monitorPollTimer: null,
+  monitorEventCursor: "",
+  monitorEvents: [],
+  previewAbortController: null,
+  previewObjectUrl: "",
+  videoStatus: null,
+  cropObjectUrls: new Set(),
+  authEnabled: null,
+  helpDocument: "user-guide",
+  helpLoadRevision: 0,
+};
+
+const helpDocuments = [
+  { name: "user-guide", label: "User guide", localized: true },
+  { name: "api", label: "API guide", localized: true },
+  { name: "maintainer", label: "Maintainer guide", localized: false },
+];
+
+const pageTitles = {
+  dashboard: "Dashboard",
+  collections: "Collections",
+  people: "People",
+  detect: "Detect",
+  compare: "Compare",
+  search: "Search",
+  video: "Camera monitoring",
+  system: "System",
+  help: "Help",
+};
+
+const $ = (selector, root = document) => root.querySelector(selector);
+const $$ = (selector, root = document) => [...root.querySelectorAll(selector)];
+
+function element(tag, { className = "", text = "", title = "" } = {}) {
+  const node = document.createElement(tag);
+  if (className) node.className = className;
+  if (text !== "") node.textContent = t(text);
+  if (title) node.title = t(title);
+  return node;
+}
+
+function firstValue(...values) {
+  return values.find((value) => value !== undefined && value !== null && value !== "");
+}
+
+function valueAt(object, path) {
+  return path.split(".").reduce((value, key) => value?.[key], object);
+}
+
+function readAny(object, ...paths) {
+  return firstValue(...paths.map((path) => valueAt(object, path)));
+}
+
+function unwrap(payload, key) {
+  return payload?.[key] ?? payload;
+}
+
+function selectedCollection() {
+  return state.collections.find((collection) => collection.id === state.selectedCollection);
+}
+
+function configureEmbeddingFields(form) {
+  const mode = form?.elements?.embedding_mode?.value ?? "server";
+  const fields = $(".external-embedding-fields", form);
+  if (!fields) return;
+  fields.hidden = mode !== "external_trusted";
+  const collection = selectedCollection();
+  const contract = collection?.embedding_contract_id || "Unavailable";
+  $$(".embedding-contract", form).forEach((target) => { target.textContent = contract; });
+}
+
+function enrollmentEmbeddingInput(form, images) {
+  const embeddingMode = form.elements.embedding_mode?.value ?? "server";
+  if (embeddingMode !== "external_trusted") {
+    return { embeddingMode: "server" };
+  }
+  const collection = selectedCollection();
+  const embeddingContractId = collection?.embedding_contract_id;
+  if (!embeddingContractId) {
+    throw new Error("The selected Collection does not report an embedding contract. Refresh it or use server extraction.");
+  }
+  return {
+    embeddingMode,
+    embeddingContractId,
+    externalEmbeddings: parseExternalEmbeddings(
+      form.elements.external_embeddings.value,
+      images.length,
+      collection.embedding_dimension,
+    ),
+  };
+}
+
+function setText(selector, text) {
+  const target = $(selector);
+  if (target) target.textContent = text ?? "—";
+}
+
+function releaseCropObjectUrl(url) {
+  if (!url || !state.cropObjectUrls.delete(url)) return;
+  URL.revokeObjectURL(url);
+}
+
+function releaseAllCropObjectUrls() {
+  for (const url of state.cropObjectUrls) URL.revokeObjectURL(url);
+  state.cropObjectUrls.clear();
+}
+
+function setBusy(control, busy, busyText = "Working…") {
+  if (!control) return;
+  if (busy) {
+    control.dataset.originalText = control.textContent;
+    control.textContent = t(busyText);
+    control.disabled = true;
+  } else {
+    control.textContent = control.dataset.originalText || control.textContent;
+    control.disabled = false;
+    delete control.dataset.originalText;
+  }
+}
+
+function toast(title, message = "", kind = "success") {
+  const region = $("#toast-region");
+  const item = element("div", { className: `toast ${kind === "error" ? "error" : ""}` });
+  item.append(element("strong", { text: title }));
+  if (message) item.append(element("span", { text: message }));
+  region.append(item);
+  window.setTimeout(() => item.remove(), 5000);
+}
+
+function describeError(error) {
+  if (error instanceof ApiError) return `${error.code}: ${t(error.message)}`;
+  return t(error?.message || "Unexpected error.");
+}
+
+function handleError(error, context = "Request failed") {
+  const message = describeError(error);
+  const requestId = error?.requestId ? ` ${t("Request: {id}", { id: error.requestId })}` : "";
+  state.errors.unshift({ context, message, requestId, at: new Date() });
+  state.errors = state.errors.slice(0, 8);
+  renderRecentErrors();
+  toast(context, `${message}${requestId}`, "error");
+}
+
+async function guarded(control, operation, context, busyText) {
+  setBusy(control, true, busyText);
+  try {
+    return await operation();
+  } catch (error) {
+    handleError(error, context);
+    return undefined;
+  } finally {
+    setBusy(control, false);
+  }
+}
+
+function renderRecentErrors() {
+  const container = $("#recent-errors");
+  container.replaceChildren();
+  const serverErrors = (state.serverErrors ?? []).map((item) => ({
+    context: item.code || item.error_code || "Server error",
+    message: item.message || item.detail || item.path || "An error was recorded by the server.",
+    requestId: item.request_id ? ` ${t("Request: {id}", { id: item.request_id })}` : "",
+    at: item.timestamp ? new Date(item.timestamp) : null,
+  }));
+  const errors = [...state.errors, ...serverErrors].slice(0, 8);
+  if (!errors.length) {
+    container.className = "empty-state compact";
+    container.append(element("strong", { text: "No recent errors" }), element("span", { text: "API failures from this browser tab will appear here." }));
+    return;
+  }
+  container.className = "";
+  for (const error of errors) {
+    const entry = element("div", { className: "error-entry" });
+    entry.append(
+      element("strong", { text: error.context }),
+      element("span", { text: `${error.message}${error.requestId}` }),
+      element("time", { text: error.at && !Number.isNaN(error.at.getTime()) ? error.at.toLocaleTimeString() : t("Server") }),
+    );
+    container.append(entry);
+  }
+}
+
+async function refreshHealth() {
+  const chip = $("#sidebar-health");
+  try {
+    const health = await client.health();
+    applyAuthenticationMode(health);
+    const status = String(firstValue(health?.status, health?.state, "healthy"));
+    chip.className = `server-chip ${/ok|healthy|ready/i.test(status) ? "healthy" : "error"}`;
+    chip.lastElementChild.textContent = t(status);
+    return health;
+  } catch (error) {
+    chip.className = "server-chip error";
+    chip.lastElementChild.textContent = t("Unavailable");
+    throw error;
+  }
+}
+
+function applyAuthenticationMode(health) {
+  const enabled = authenticationEnabledFromHealth(health);
+  state.authEnabled = enabled;
+  if (!enabled) {
+    client.clearApiKey();
+    const dialog = $("#api-key-dialog");
+    if (dialog?.open) dialog.close();
+  }
+  $$('[data-auth-ui]').forEach((node) => { node.hidden = !enabled; });
+  updateAuthState();
+}
+
+function navigate(route) {
+  if (!pageTitles[route]) return;
+  if (state.route === "video" && route !== "video") stopMonitorView();
+  state.route = route;
+  $$("[data-page]").forEach((page) => { page.hidden = page.dataset.page !== route; });
+  $$(".nav-item[data-route]").forEach((item) => {
+    const selected = item.dataset.route === route;
+    item.classList.toggle("active", selected);
+    if (selected) item.setAttribute("aria-current", "page");
+    else item.removeAttribute("aria-current");
+  });
+  setText("#page-title", t(pageTitles[route]));
+  document.body.classList.remove("menu-open");
+  $("#menu-toggle").setAttribute("aria-expanded", "false");
+  document.title = `${t(pageTitles[route])} · InsightFace Server`;
+  window.history.replaceState(null, "", `#${route}`);
+  void refreshRoute(route);
+}
+
+async function refreshRoute(route = state.route) {
+  if (route === "dashboard") await loadDashboard();
+  if (route === "collections") await loadCollections();
+  if (route === "people") {
+    await loadCollections();
+    if (state.selectedCollection) await loadPeople();
+  }
+  if (["detect", "compare", "search", "video"].includes(route)) await loadCollections();
+  if (route === "video") await loadMonitors();
+  if (route === "system") await loadSystem();
+  if (route === "help") await loadDocumentation();
+}
+
+function documentationTarget(href) {
+  const value = String(href || "").trim();
+  if (/^(?:\.\/)?maintainer-guide\.md(?:#[A-Za-z0-9_.-]+)?$/.test(value)) {
+    return "maintainer";
+  }
+  const local = /^(?:\.\/)?(user-guide|api)(?:\.(?:zh-CN|ja|de|es|fr|ru|pt|ko))?\.md(?:#[A-Za-z0-9_.-]+)?$/.exec(value);
+  if (local) return local[1];
+  const bundled = /^\/guide-content\/(?:en|zh|ja|de|es|fr|ru|pt|ko)\/(user-guide|api|maintainer)\.md(?:#[A-Za-z0-9_.-]+)?$/.exec(value);
+  return bundled?.[1] ?? "";
+}
+
+function renderHelpDocumentSwitcher() {
+  const switcher = $("#documentation-switcher");
+  switcher.replaceChildren();
+  for (const documentDefinition of helpDocuments) {
+    const button = element("button", { className: "documentation-switch" });
+    button.type = "button";
+    button.dataset.helpDocument = documentDefinition.name;
+    button.setAttribute("aria-pressed", String(state.helpDocument === documentDefinition.name));
+    button.classList.toggle("active", state.helpDocument === documentDefinition.name);
+    button.append(element("span", { text: documentDefinition.label }));
+    if (!documentDefinition.localized) {
+      button.append(element("small", { text: "English only" }));
+    }
+    button.addEventListener("click", () => {
+      if (state.helpDocument === documentDefinition.name) {
+        $("#markdown-document")?.scrollIntoView({ behavior: "smooth", block: "start" });
+        return;
+      }
+      state.helpDocument = documentDefinition.name;
+      renderHelpDocumentSwitcher();
+      void loadDocumentation();
+    });
+    switcher.append(button);
+  }
+}
+
+function embeddedDocumentationChapter(documentName, markdown) {
+  const chapter = document.createElement("section");
+  chapter.className = "documentation-chapter";
+  chapter.dataset.document = documentName;
+  chapter.innerHTML = renderMarkdown(markdown);
+
+  const languageNavigation = chapter.querySelector("p");
+  const languageLinks = [...(languageNavigation?.querySelectorAll("a[href]") ?? [])]
+    .filter((link) => documentationTarget(link.getAttribute("href")));
+  if (languageLinks.length >= 4) {
+    languageNavigation.remove();
+  }
+
+  [...chapter.querySelectorAll("h1, h2, h3, h4, h5")].forEach((heading, index) => {
+    const currentLevel = Number(heading.tagName.slice(1));
+    const replacement = document.createElement(`h${Math.min(currentLevel + 1, 6)}`);
+    replacement.id = index === 0 ? `help-${documentName}` : `help-${documentName}-${index}`;
+    replacement.replaceChildren(...heading.childNodes);
+    heading.replaceWith(replacement);
+  });
+  for (const link of chapter.querySelectorAll("a[href]")) {
+    const targetDocument = documentationTarget(link.getAttribute("href"));
+    if (!targetDocument) continue;
+    link.setAttribute("href", "#help");
+    link.removeAttribute("target");
+    link.removeAttribute("rel");
+    link.addEventListener("click", (event) => {
+      event.preventDefault();
+      state.helpDocument = targetDocument;
+      renderHelpDocumentSwitcher();
+      void loadDocumentation();
+    });
+  }
+  return chapter;
+}
+
+function filterDocumentationToc() {
+  const query = $("#documentation-filter").value.trim().toLocaleLowerCase();
+  $$(".documentation-toc-link", $("#documentation-toc")).forEach((button) => {
+    button.hidden = Boolean(query) && !button.textContent.toLocaleLowerCase().includes(query);
+  });
+  $$(".documentation-toc-group", $("#documentation-toc")).forEach((group) => {
+    group.hidden = $$(".documentation-toc-link", group).every((button) => button.hidden);
+  });
+}
+
+function renderDocumentationToc(chapters) {
+  const navigation = $("#documentation-toc");
+  navigation.replaceChildren();
+  for (const chapter of chapters) {
+    const group = element("div", { className: "documentation-toc-group" });
+    for (const heading of chapter.querySelectorAll("h2, h3, h4")) {
+      const button = document.createElement("button");
+      button.type = "button";
+      button.className = `documentation-toc-link level-${heading.tagName.slice(1)}`;
+      button.textContent = heading.textContent;
+      button.addEventListener("click", () => {
+        heading.scrollIntoView({ behavior: "smooth", block: "start" });
+      });
+      group.append(button);
+    }
+    navigation.append(group);
+  }
+  filterDocumentationToc();
+}
+
+async function loadDocumentation() {
+  const revision = ++state.helpLoadRevision;
+  const documentDefinition = helpDocuments.find(({ name }) => name === state.helpDocument)
+    ?? helpDocuments[0];
+  renderHelpDocumentSwitcher();
+  const root = $("#markdown-document");
+  root.replaceChildren();
+  root.setAttribute("aria-busy", "true");
+  const loading = element("div", { className: "loading-document" });
+  loading.append(element("span"), element("strong", { text: "Loading documentation…" }));
+  root.append(loading);
+  $("#documentation-toc").replaceChildren();
+  $("#documentation-filter").value = "";
+  try {
+    const documentLocale = documentDefinition.localized ? locale() : "en";
+    const response = await fetch(`/guide-content/${documentLocale}/${documentDefinition.name}.md`, {
+      headers: { Accept: "text/markdown" },
+    });
+    if (!response.ok) throw new Error(`${documentDefinition.name}: HTTP ${response.status}`);
+    const markdown = await response.text();
+    if (revision !== state.helpLoadRevision) return;
+    const chapter = embeddedDocumentationChapter(documentDefinition.name, markdown);
+    root.replaceChildren(chapter);
+    root.setAttribute("aria-busy", "false");
+    renderDocumentationToc([chapter]);
+  } catch (error) {
+    if (revision !== state.helpLoadRevision) return;
+    root.setAttribute("aria-busy", "false");
+    root.replaceChildren(
+      element("div", { className: "empty-state" }),
+    );
+    root.firstElementChild.append(
+      element("strong", { text: "Could not load documentation." }),
+      element("span", { text: error.message }),
+    );
+  }
+}
+
+function statusRow(label, value, className = "") {
+  const row = element("div");
+  row.append(element("dt", { text: label }));
+  const description = element("dd", { text: String(value ?? "—"), className });
+  row.append(description);
+  return row;
+}
+
+function systemProvider(system) {
+  return readAny(system, "execution_provider", "provider", "onnx_runtime.provider", "runtime.execution_provider") ?? "Unknown";
+}
+
+function systemModels(system) {
+  const fromSystem = listItems(system, ["models"]);
+  return fromSystem.length ? fromSystem : state.models;
+}
+
+function updateCreateCollectionProfiles() {
+  const select = $("#collection-create-form select[name=search_profile]");
+  return applySearchProfileAvailability(select, searchProfilesFromSystem(state.system));
+}
+
+function newCollectionCropDefault() {
+  return Boolean(state.system?.safe_config?.save_face_crops);
+}
+
+function systemDetectionDefaults() {
+  return state.system?.safe_config?.detection ?? {
+    input_sizes: [[96, 96], [512, 512]],
+    threshold: 0.5,
+    nms_threshold: 0.4,
+    single_face_selection: "largest",
+  };
+}
+
+function formatDetectorInputSizes(value) {
+  return (value ?? []).map((size) => `${size[0]}x${size[1]}`).join(", ");
+}
+
+function parseDetectorInputSizes(value) {
+  const parts = String(value ?? "").split(",").map((part) => part.trim()).filter(Boolean);
+  if (!parts.length) throw new Error("At least one detector input size is required.");
+  return parts.map((part) => {
+    const match = part.match(/^(\d+)\s*[x×]\s*(\d+)$/i);
+    if (!match) throw new Error(`Invalid detector input size: ${part}`);
+    return [Number(match[1]), Number(match[2])];
+  });
+}
+
+async function loadDashboard() {
+  const [healthResult, systemResult, collectionResult, modelResult] = await Promise.allSettled([
+    refreshHealth(), client.system(), client.listCollections({ limit: 100 }), client.models(),
+  ]);
+  if (systemResult.status === "fulfilled") {
+    state.system = unwrap(systemResult.value, "system");
+    updateCreateCollectionProfiles();
+  }
+  if (collectionResult.status === "fulfilled") {
+    state.collections = listItems(collectionResult.value, ["collections"]);
+    updateCollectionSelects();
+  }
+  if (modelResult.status === "fulfilled") state.models = listItems(modelResult.value, ["models"]);
+
+  const system = state.system ?? {};
+  state.serverErrors = Array.isArray(system.recent_errors) ? system.recent_errors : [];
+  renderRecentErrors();
+  const collections = state.collections;
+  const personCount = Number(system.stats?.person_count ?? collections.reduce((sum, item) => sum + Number(item.person_count ?? 0), 0));
+  const faceCount = Number(system.stats?.face_count ?? collections.reduce((sum, item) => sum + Number(item.face_count ?? 0), 0));
+  const models = systemModels(system);
+  const model = models[0] ?? {};
+  const provider = systemProvider(system);
+  setText("#dashboard-provider", provider);
+  setText("#metric-collections", String(system.stats?.collection_count ?? collections.length));
+  setText("#metric-people", String(personCount));
+  setText("#metric-faces", String(faceCount));
+  setText("#metric-model", firstValue(model.model_id, model.id, readAny(system, "model.model_id"), "Not loaded"));
+  setText("#metric-model-version", firstValue(model.model_version, model.version, readAny(system, "model.model_version"), "—"));
+
+  const status = $("#dashboard-status");
+  status.replaceChildren(
+    statusRow("API readiness", healthResult.status === "fulfilled" ? firstValue(healthResult.value?.status, "ready") : "unavailable"),
+    statusRow("Database", readAny(system, "database.quick_check", "database.status", "database_state") ?? "unknown"),
+    statusRow("Execution provider", provider),
+    statusRow("ONNX Runtime", readAny(system, "onnx_runtime_version", "onnx_runtime.version", "runtime.onnx_runtime_version") ?? "—"),
+  );
+}
+
+async function loadCollections() {
+  const capabilityRequest = state.system === null
+    ? client.system().catch((error) => {
+      handleError(error, "Could not load search capabilities");
+      return null;
+    })
+    : Promise.resolve(null);
+  const [payload, systemPayload] = await Promise.all([
+    guarded(null, () => client.listCollections({ limit: 100 }), "Could not load collections"),
+    capabilityRequest,
+  ]);
+  if (systemPayload !== null) {
+    state.system = unwrap(systemPayload, "system");
+    updateCreateCollectionProfiles();
+  }
+  if (payload === undefined) return;
+  state.collections = listItems(payload, ["collections"]);
+  renderCollections();
+  updateCollectionSelects();
+  configureEmbeddingFields($("#person-create-form"));
+}
+
+function collectionLabel(collection) {
+  return collection.name || collection.id || "Unnamed collection";
+}
+
+function updateCollectionSelects() {
+  for (const select of [$("#people-collection"), $("#search-collection"), $("#video-collection"), $("#detect-collection"), $("#compare-collection")]) {
+    if (!select) continue;
+    const current = select.value;
+    const prompt = select.options[0]?.textContent ?? "Select a collection";
+    select.replaceChildren(new Option(prompt, ""));
+    for (const collection of state.collections) select.add(new Option(`${collectionLabel(collection)} (${collection.id})`, collection.id));
+    if (state.collections.some((collection) => collection.id === current)) select.value = current;
+  }
+  const peopleSelect = $("#people-collection");
+  if (state.selectedCollection && state.collections.some((item) => item.id === state.selectedCollection)) peopleSelect.value = state.selectedCollection;
+}
+
+function renderCollections() {
+  const table = $("#collection-table");
+  const query = $("#collection-filter").value.trim().toLocaleLowerCase();
+  const collections = state.collections.filter((item) => !query || [item.id, item.name, item.description].some((value) => String(value ?? "").toLocaleLowerCase().includes(query)));
+  table.replaceChildren();
+  $("#collection-empty").hidden = collections.length > 0;
+  $("#collection-count").textContent = t(
+    collections.length === 1 ? "{count} collection" : "{count} collections",
+    { count: collections.length },
+  );
+  for (const collection of collections) {
+    const row = document.createElement("tr");
+    const identity = document.createElement("td");
+    identity.append(element("span", { className: "table-title", text: collectionLabel(collection) }), element("span", { className: "table-subtitle", text: collection.id }));
+    const threshold = document.createElement("td");
+    threshold.append(element("span", { className: "badge", text: formatCosine(firstValue(collection.default_threshold, collection.threshold, 0)) }));
+    const profile = document.createElement("td");
+    profile.append(element("span", { className: "badge", text: firstValue(collection.search_profile, "fp32_v1") }));
+    const crops = document.createElement("td");
+    crops.append(element("span", { className: "badge", text: collection.save_face_crops ? "On" : "Off" }));
+    const model = document.createElement("td");
+    model.append(element("span", { className: "table-title", text: firstValue(collection.model_id, "—") }), element("span", { className: "table-subtitle", text: `v${firstValue(collection.model_version, "—")}` }));
+    const people = element("td", { text: String(collection.person_count ?? 0) });
+    const faces = element("td", { text: String(collection.face_count ?? 0) });
+    const actions = element("td", { className: "table-actions" });
+    const manage = element("button", { className: "button secondary small", text: "People" });
+    manage.type = "button";
+    manage.addEventListener("click", () => {
+      state.selectedCollection = collection.id;
+      navigate("people");
+    });
+    const edit = element("button", { className: "button secondary small", text: "Edit" });
+    edit.type = "button";
+    edit.addEventListener("click", () => openCollectionEdit(collection));
+    const remove = element("button", { className: "button danger small", text: "Delete" });
+    remove.type = "button";
+    remove.addEventListener("click", () => void deleteCollection(collection));
+    actions.append(manage, edit, remove);
+    row.append(identity, threshold, profile, crops, model, people, faces, actions);
+    table.append(row);
+  }
+}
+
+function openCollectionEdit(collection) {
+  const form = $("#collection-edit-form");
+  form.elements.id.value = collection.id;
+  form.elements.name.value = collection.name ?? "";
+  form.elements.description.value = collection.description ?? "";
+  form.elements.threshold.value = firstValue(collection.default_threshold, collection.threshold, 0.4);
+  form.elements.search_profile.value = firstValue(collection.search_profile, "fp32_v1");
+  form.elements.capacity_rows.value = firstValue(collection.capacity_rows, 100000);
+  form.elements.max_faces_per_person.value = firstValue(collection.max_faces_per_person, 20);
+  form.elements.load_policy.value = firstValue(collection.load_policy, "lazy");
+  form.elements.save_face_crops.checked = Boolean(collection.save_face_crops);
+  const detection = collection.detection ?? systemDetectionDefaults();
+  form.elements.detector_input_sizes.value = formatDetectorInputSizes(detection.input_sizes);
+  form.elements.detector_threshold.value = firstValue(detection.threshold, 0.5);
+  form.elements.detector_nms_threshold.value = firstValue(detection.nms_threshold, 0.4);
+  form.elements.single_face_selection.value = firstValue(detection.single_face_selection, "largest");
+  form.elements.metadata.value = JSON.stringify(collection.metadata ?? {}, null, 2);
+  $("#collection-edit-dialog").showModal();
+}
+
+async function deleteCollection(collection) {
+  const nonEmpty = Number(collection.person_count ?? 0) > 0 || Number(collection.face_count ?? 0) > 0;
+  const prompt = nonEmpty
+    ? t("Collection “{name}” is not empty. Delete it and all people and face samples?", { name: collectionLabel(collection) })
+    : t("Delete collection “{name}”?", { name: collectionLabel(collection) });
+  if (!window.confirm(prompt)) return;
+  const result = await guarded(null, () => client.deleteCollection(collection.id, nonEmpty), "Collection could not be deleted");
+  if (result === undefined) return;
+  toast("Collection deleted", collection.id);
+  if (state.selectedCollection === collection.id) {
+    state.selectedCollection = "";
+    state.selectedPerson = null;
+  }
+  await loadCollections();
+}
+
+async function loadPeople() {
+  const collectionId = state.selectedCollection || $("#people-collection").value;
+  state.selectedCollection = collectionId;
+  if (!collectionId) {
+    state.people = [];
+    state.selectedPerson = null;
+    renderPeople();
+    return;
+  }
+  const payload = await guarded(null, () => client.listPeople(collectionId, { limit: 100 }), "Could not load people");
+  if (payload === undefined) return;
+  state.people = listItems(payload, ["persons", "people"]);
+  if (state.selectedPerson && !state.people.some((person) => person.id === state.selectedPerson.id)) state.selectedPerson = null;
+  renderPeople();
+}
+
+function initials(person) {
+  const text = String(person.name || person.id || "?").trim();
+  const words = text.split(/\s+/).filter(Boolean);
+  return (words.length > 1 ? `${words[0][0]}${words.at(-1)[0]}` : text.slice(0, 2)).toUpperCase();
+}
+
+function renderPeople() {
+  const list = $("#person-list");
+  const people = filterPeople(state.people, $("#people-filter").value);
+  list.replaceChildren();
+  setText("#people-count", t(people.length === 1 ? "{count} person" : "{count} people", { count: people.length }));
+  setText("#people-list-title", state.selectedCollection ? t("People in {collection}", { collection: state.selectedCollection }) : t("People"));
+  $("#people-empty").hidden = people.length > 0;
+  if (!people.length) {
+    const title = state.selectedCollection ? "No people found" : "Select a collection";
+    const message = state.selectedCollection ? "Register a person or change the search." : "Its enrolled people will appear here.";
+    $("#people-empty").replaceChildren(element("strong", { text: title }), element("span", { text: message }));
+  }
+  for (const person of people) {
+    const button = element("button", { className: `person-row ${state.selectedPerson?.id === person.id ? "selected" : ""}` });
+    button.type = "button";
+    const identity = element("span");
+    identity.append(element("strong", { text: person.name || person.id }), element("small", { text: person.external_id || person.id }));
+    const faceCount = Number(person.face_count ?? 0);
+    button.append(
+      element("span", { className: "avatar", text: initials(person) }),
+      identity,
+      element("span", { className: "face-count", text: t(faceCount === 1 ? "{count} face" : "{count} faces", { count: faceCount }) }),
+    );
+    button.addEventListener("click", () => void selectPerson(person));
+    list.append(button);
+  }
+  if (!state.selectedPerson) renderPersonDetail();
+}
+
+async function selectPerson(person) {
+  const payload = await guarded(null, () => client.getPerson(state.selectedCollection, person.id), "Could not load person");
+  if (payload === undefined) return;
+  state.selectedPerson = unwrap(payload, "person");
+  renderPeople();
+  await renderPersonDetail();
+}
+
+function renderRejections(target, rejected, files = []) {
+  const items = rejected ?? [];
+  target.replaceChildren();
+  target.hidden = items.length === 0;
+  if (!items.length) return;
+  target.append(element("strong", { text: t(items.length === 1 ? "{count} image was rejected" : "{count} images were rejected", { count: items.length }) }));
+  items.forEach((item, index) => {
+    const row = element("div", { className: "rejection-item" });
+    const file = files[Number(item.index)];
+    row.append(
+      element("span", { text: item.filename || item.file_name || file?.name || t("Image {index}", { index: index + 1 }) }),
+      element("strong", { text: item.reason || item.code || "rejected" }),
+    );
+    target.append(row);
+  });
+}
+
+async function renderPersonDetail() {
+  const panel = $("#person-detail");
+  const person = state.selectedPerson;
+  if (!person) {
+    panel.replaceChildren();
+    const empty = element("div", { className: "empty-state" });
+    empty.append(element("strong", { text: "No person selected" }), element("span", { text: "Choose a person to view details and face samples." }));
+    panel.append(empty);
+    return;
+  }
+
+  panel.innerHTML = `
+    <div class="person-detail-header">
+      <div><p class="eyebrow">PERSON</p><h3 class="person-name"></h3><p class="person-identity"></p></div>
+      <div class="table-actions"><button class="button secondary small edit-person" type="button">Edit</button><button class="button danger small delete-person" type="button">Delete</button></div>
+    </div>
+    <div class="person-detail-body">
+      <form class="person-edit-form" hidden>
+        <div class="form-grid two"><label><span class="field-caption">Name</span><input name="name" maxlength="200"></label><label><span class="field-caption">External ID</span><input name="external_id" maxlength="200"></label></div>
+        <label>Metadata (JSON)<textarea name="metadata" rows="3" spellcheck="false"></textarea></label>
+        <div class="form-actions"><button class="button primary small" type="submit">Save</button><button class="button secondary small cancel-edit" type="button">Cancel</button></div>
+      </form>
+      <div class="person-summary"><p class="eyebrow">METADATA</p><pre class="person-metadata"></pre></div>
+      <div><div class="panel-heading"><div><p class="eyebrow">FACE SAMPLES</p><h3 class="faces-heading">Registered faces</h3></div></div>
+        <form class="add-faces-form"><label class="drop-zone compact-drop"><input name="images" type="file" accept="image/jpeg,image/png,image/webp" multiple required><strong>Add registration photos</strong><span>JPEG, PNG, or WebP</span><span class="file-summary">No files selected</span></label><div class="form-grid two"><label><span class="field-caption">Enrollment review</span><select name="review_mode"><option value="off" selected>Off · collection selection</option><option value="standard">Standard quality</option><option value="strict">Strict identity</option></select></label><label><span class="field-caption">Embedding source</span><select name="embedding_mode"><option value="server" selected>Extract from image</option><option value="external_trusted">Trusted external feature</option></select></label></div><p class="hint">Off uses the Collection single-face selection strategy; standard and strict require exactly one.</p><p class="hint embedding-contract-summary">Collection embedding contract: <code class="embedding-contract"></code></p><div class="external-embedding-fields" hidden><p class="hint">The server trusts the image/vector pairing and does not re-extract the feature.</p><label>External embeddings (JSON, one vector per image)<textarea name="external_embeddings" rows="5" spellcheck="false" placeholder="[[0.0123, -0.0456, ...]]"></textarea></label></div><div class="form-actions"><button class="button secondary small" type="submit">Add face samples</button><span class="form-status" role="status"></span></div></form>
+        <div class="rejection-list face-rejections" hidden></div><div class="face-sample-grid"></div>
+      </div>
+    </div>`;
+
+  translateTree(panel);
+  $(".person-name", panel).textContent = person.name || person.id;
+  $(".person-identity", panel).textContent = [person.id, person.external_id].filter(Boolean).join(" · ");
+  $(".person-metadata", panel).textContent = JSON.stringify(person.metadata ?? {}, null, 2);
+  const editForm = $(".person-edit-form", panel);
+  editForm.elements.name.value = person.name ?? "";
+  editForm.elements.external_id.value = person.external_id ?? "";
+  editForm.elements.metadata.value = JSON.stringify(person.metadata ?? {}, null, 2);
+  $(".edit-person", panel).addEventListener("click", () => { editForm.hidden = false; });
+  $(".cancel-edit", panel).addEventListener("click", () => { editForm.hidden = true; });
+  $(".delete-person", panel).addEventListener("click", () => void deleteSelectedPerson());
+  editForm.addEventListener("submit", (event) => void updateSelectedPerson(event));
+  const addForm = $(".add-faces-form", panel);
+  setupFileInput($("input[type=file]", addForm));
+  addForm.elements.embedding_mode.addEventListener("change", () => configureEmbeddingFields(addForm));
+  configureEmbeddingFields(addForm);
+  addForm.addEventListener("submit", (event) => void addPersonFaces(event));
+
+  const facePayload = await guarded(null, () => client.listFaces(state.selectedCollection, person.id, { limit: 100 }), "Could not load face samples");
+  if (facePayload === undefined || state.selectedPerson?.id !== person.id) return;
+  renderFaceSamples(listItems(facePayload, ["faces", "face_samples"]));
+}
+
+function renderFaceSamples(faces) {
+  const panel = $("#person-detail");
+  const grid = $(".face-sample-grid", panel);
+  if (!grid) return;
+  releaseAllCropObjectUrls();
+  grid.replaceChildren();
+  $(".faces-heading", panel).textContent = t("Registered faces ({count})", { count: faces.length });
+  if (!faces.length) {
+    const empty = element("div", { className: "empty-state compact" });
+    empty.append(element("strong", { text: "No face samples" }), element("span", { text: "Add one or more clear photos." }));
+    grid.append(empty);
+    return;
+  }
+  for (const face of faces) {
+    const card = element("article", { className: "face-card" });
+    card.append(element("strong", { text: face.id }));
+    const details = document.createElement("dl");
+    details.append(
+      statusRow("Detection", formatScore(firstValue(face.detection_score, face.confidence))),
+      statusRow("Quality", formatScore(firstValue(face.quality?.score, face.quality_score))),
+      statusRow("Embedding", face.embedding_source === "external_trusted" ? "External trusted" : "Server extracted"),
+      statusRow("Stored crop", face.has_crop ? "Available" : "Not saved"),
+      statusRow("Created", formatTimestamp(face.created_at)),
+    );
+    const crop = document.createElement("img");
+    crop.alt = t("Saved crop for face sample {id}", { id: face.id });
+    crop.hidden = true;
+    if (face.has_crop) {
+      const view = element("button", { className: "button secondary small full-width", text: "View crop" });
+      view.type = "button";
+      view.addEventListener("click", () => void toggleFaceCrop(face, view, crop));
+      card.append(view, crop);
+    }
+    const remove = element("button", { className: "button danger small full-width", text: "Delete sample" });
+    remove.type = "button";
+    remove.addEventListener("click", () => void deleteFaceSample(face));
+    card.append(details, remove);
+    grid.append(card);
+  }
+}
+
+async function toggleFaceCrop(face, button, image) {
+  if (!image.hidden) {
+    releaseCropObjectUrl(image.dataset.cropUrl);
+    delete image.dataset.cropUrl;
+    image.hidden = true;
+    image.removeAttribute("src");
+    button.textContent = t("View crop");
+    return;
+  }
+  const blob = await guarded(
+    button,
+    () => client.getFaceCrop(state.selectedCollection, state.selectedPerson.id, face.id),
+    "Face crop could not be loaded",
+    "Loading…",
+  );
+  if (blob === undefined || !button.isConnected) return;
+  const url = URL.createObjectURL(blob);
+  state.cropObjectUrls.add(url);
+  image.dataset.cropUrl = url;
+  const release = () => {
+    releaseCropObjectUrl(url);
+    if (image.dataset.cropUrl === url) delete image.dataset.cropUrl;
+  };
+  image.addEventListener("load", release, { once: true });
+  image.addEventListener("error", () => {
+    release();
+    image.hidden = true;
+    image.removeAttribute("src");
+    button.textContent = t("View crop");
+    handleError(new Error("The stored face crop could not be decoded."), "Face crop could not be displayed");
+  }, { once: true });
+  image.src = url;
+  image.hidden = false;
+  button.textContent = t("Hide crop");
+}
+
+function formatTimestamp(value) {
+  if (!value) return "—";
+  const date = new Date(value);
+  return Number.isNaN(date.getTime()) ? String(value) : date.toLocaleString();
+}
+
+async function updateSelectedPerson(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const submit = $("button[type=submit]", form);
+  let metadata;
+  try { metadata = parseMetadata(form.elements.metadata.value); } catch (error) { handleError(error, "Invalid person metadata"); return; }
+  const payload = await guarded(submit, () => client.updatePerson(state.selectedCollection, state.selectedPerson.id, {
+    name: form.elements.name.value.trim() || null,
+    external_id: form.elements.external_id.value.trim() || null,
+    metadata,
+  }), "Person could not be updated", "Saving…");
+  if (payload === undefined) return;
+  state.selectedPerson = unwrap(payload, "person");
+  toast("Person updated", state.selectedPerson.id);
+  await loadPeople();
+  const refreshed = state.people.find((item) => item.id === state.selectedPerson?.id);
+  if (refreshed) await selectPerson(refreshed);
+}
+
+async function deleteSelectedPerson() {
+  const person = state.selectedPerson;
+  if (!person || !window.confirm(t("Delete “{name}” and all registered face samples?", { name: person.name || person.id }))) return;
+  const result = await guarded(null, () => client.deletePerson(state.selectedCollection, person.id), "Person could not be deleted");
+  if (result === undefined) return;
+  toast("Person deleted", person.id);
+  state.selectedPerson = null;
+  await loadPeople();
+}
+
+async function addPersonFaces(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const button = $("button[type=submit]", form);
+  const images = [...form.elements.images.files];
+  if (!images.length) return;
+  let embeddingInput;
+  try { embeddingInput = enrollmentEmbeddingInput(form, images); } catch (error) { handleError(error, "Invalid external embeddings"); return; }
+  const payload = await guarded(button, () => client.addFaces(state.selectedCollection, state.selectedPerson.id, images, { reviewMode: form.elements.review_mode.value, ...embeddingInput }), "Face samples could not be added", "Uploading…");
+  if (payload === undefined) return;
+  renderRejections($(".face-rejections", $("#person-detail")), payload.rejected_images ?? payload.rejected ?? [], images);
+  const addedCount = listItems(payload, ["faces", "face_samples"]).length;
+  toast("Registration complete", t(addedCount === 1 ? "{count} face sample added." : "{count} face samples added.", { count: addedCount }));
+  form.elements.external_embeddings.value = "";
+  const id = state.selectedPerson.id;
+  await loadPeople();
+  const person = state.people.find((item) => item.id === id);
+  if (person) await selectPerson(person);
+}
+
+async function deleteFaceSample(face) {
+  if (!window.confirm(t("Delete face sample {id}?", { id: face.id }))) return;
+  const result = await guarded(null, () => client.deleteFace(state.selectedCollection, state.selectedPerson.id, face.id), "Face sample could not be deleted");
+  if (result === undefined) return;
+  toast("Face sample deleted", face.id);
+  const person = state.selectedPerson;
+  await selectPerson(person);
+}
+
+function fileSummary(files) {
+  if (!files?.length) return t("No files selected");
+  if (files.length === 1) return files[0].name;
+  return t("{count} files selected", { count: files.length });
+}
+
+function setupFileInput(input) {
+  if (!input || input.dataset.ready) return;
+  input.dataset.ready = "true";
+  const zone = input.closest(".drop-zone");
+  const summary = $(".file-summary", zone);
+  const update = () => { if (summary) summary.textContent = fileSummary(input.files); };
+  input.addEventListener("change", update);
+  for (const eventName of ["dragenter", "dragover"]) zone.addEventListener(eventName, () => zone.classList.add("dragging"));
+  for (const eventName of ["dragleave", "drop"]) zone.addEventListener(eventName, () => zone.classList.remove("dragging"));
+}
+
+function setupImagePreview(input, canvas, onReset = () => {}) {
+  input.addEventListener("change", () => {
+    const file = input.files[0];
+    onReset();
+    if (!file) {
+      canvas.closest(".canvas-stage")?.classList.add("empty");
+      return;
+    }
+    void renderImageFaces(file, canvas).catch((error) => handleError(error, "Image preview failed"));
+  });
+}
+
+function imageFromFile(file) {
+  return new Promise((resolve, reject) => {
+    const url = URL.createObjectURL(file);
+    const image = new Image();
+    image.onload = () => { URL.revokeObjectURL(url); resolve(image); };
+    image.onerror = () => { URL.revokeObjectURL(url); reject(new Error("The selected image could not be decoded.")); };
+    image.src = url;
+  });
+}
+
+function drawFaceOverlays(canvas, faces, { labels = [] } = {}) {
+  const context = canvas.getContext("2d");
+  const lineWidth = Math.max(2, Math.round(canvas.width / 600 * 2));
+  context.lineWidth = lineWidth;
+  context.font = `${Math.max(12, Math.round(canvas.width / 60))}px ui-sans-serif, sans-serif`;
+  context.textBaseline = "bottom";
+  faces.forEach((face, index) => {
+    const box = facePixels(face, canvas.width, canvas.height);
+    if (!box) return;
+    const color = index === 0 ? "#818cf8" : "#22d3ee";
+    context.strokeStyle = color;
+    context.fillStyle = color;
+    context.strokeRect(box.x, box.y, box.width, box.height);
+    for (const point of faceLandmarks(face)) {
+      const x = Math.abs(point.x) <= 1 ? point.x * canvas.width : point.x;
+      const y = Math.abs(point.y) <= 1 ? point.y * canvas.height : point.y;
+      context.beginPath();
+      context.arc(x, y, Math.max(2, lineWidth * 1.2), 0, Math.PI * 2);
+      context.fill();
+    }
+    const label = labels[index];
+    if (label) {
+      const metrics = context.measureText(label);
+      const textHeight = Math.max(17, Math.round(canvas.width / 45));
+      const top = Math.max(0, box.y - textHeight);
+      context.fillStyle = "rgba(9, 9, 11, .9)";
+      context.fillRect(box.x, top, metrics.width + 12, textHeight);
+      context.fillStyle = color;
+      context.fillText(label, box.x + 6, top + textHeight - 2);
+    }
+  });
+}
+
+async function renderImageFaces(file, canvas, faces = [], options = {}) {
+  const image = await imageFromFile(file);
+  canvas.width = image.naturalWidth;
+  canvas.height = image.naturalHeight;
+  const context = canvas.getContext("2d");
+  context.drawImage(image, 0, 0);
+  drawFaceOverlays(canvas, faces, options);
+  canvas.closest(".canvas-stage")?.classList.remove("empty");
+}
+
+function resultFaces(payload) {
+  return listItems(payload, ["faces", "detections"]);
+}
+
+function renderFaceResults(target, faces) {
+  target.replaceChildren();
+  faces.forEach((face, index) => {
+    const quality = face.quality ?? {};
+    const row = element("div", { className: "face-result" });
+    const details = element("div", { className: "quality-bars" });
+    details.append(
+      element("span", { text: t("Detection {score}", { score: formatScore(firstValue(face.detection_score, face.confidence)) }) }),
+      element("span", { text: t("Quality {score}", { score: formatScore(quality.score) }) }),
+      element("span", { text: t("Sharpness {score}", { score: formatScore(quality.sharpness) }) }),
+    );
+    row.append(element("span", { className: "face-index", text: String(index + 1) }), details, element("strong", { text: t("Brightness {score}", { score: formatScore(quality.brightness) }) }));
+    target.append(row);
+  });
+}
+
+async function submitDetect(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const button = $("button[type=submit]", form);
+  const file = form.elements.image.files[0];
+  const payload = await guarded(button, () => client.detect(file, { maxFaces: form.elements.max_faces.value, collectionId: form.elements.collection.value }), "Detection failed", "Detecting…");
+  if (payload === undefined) return;
+  const faces = resultFaces(payload);
+  await renderImageFaces(file, $("#detect-canvas"), faces, { labels: faces.map((face, index) => `#${index + 1} · ${formatScore(firstValue(face.detection_score, face.confidence))}`) });
+  renderFaceResults($("#detect-results"), faces);
+  $("#detect-summary").textContent = t(
+    faces.length === 1 ? "{count} face detected · {duration}" : "{count} faces detected · {duration}",
+    { count: faces.length, duration: formatDuration(payload.processing_ms) },
+  );
+}
+
+async function submitCompare(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const button = $("button[type=submit]", form);
+  const source = form.elements.source.files[0];
+  const target = form.elements.target.files[0];
+  const threshold = form.elements.threshold.value;
+  const payload = await guarded(button, () => client.compare(source, target, { threshold, collectionId: form.elements.collection.value }), "Comparison failed", "Comparing…");
+  if (payload === undefined) return;
+  await Promise.all([
+    renderImageFaces(source, $("#compare-source-canvas"), [payload.source_face].filter(Boolean)),
+    renderImageFaces(target, $("#compare-target-canvas"), [payload.target_face].filter(Boolean)),
+  ]);
+  const verdict = $("#compare-verdict");
+  verdict.className = `compare-verdict ${payload.matched ? "match" : "no-match"}`;
+  $(".verdict-mark", verdict).textContent = payload.matched ? "✓" : "×";
+  $("strong", verdict).textContent = t(payload.matched ? "Match" : "No match");
+  $("small", verdict).textContent = t("{similarity} similarity · threshold {threshold} · {duration}", {
+    similarity: formatCosine(payload.similarity),
+    threshold: formatCosine(payload.threshold),
+    duration: formatDuration(payload.processing_ms),
+  });
+}
+
+function renderMatches(target, matches) {
+  target.replaceChildren();
+  if (!matches.length) {
+    const empty = element("div", { className: "empty-state compact" });
+    empty.append(element("strong", { text: "No matches above threshold" }), element("span", { text: "Try a clearer image or review the collection threshold." }));
+    target.append(empty);
+    return;
+  }
+  matches.forEach((match, index) => {
+    const person = match.person ?? {};
+    const card = element("div", { className: "match-card" });
+    const identity = element("div", { className: "match-person" });
+    const metadata = person.metadata && Object.keys(person.metadata).length ? JSON.stringify(person.metadata) : "";
+    identity.append(
+      element("strong", { text: person.name || person.id || match.person_id || "Unknown" }),
+      element("small", { text: [person.external_id || person.id, t("Face {id}", { id: match.matched_face_id ?? "—" }), metadata].filter(Boolean).join(" · "), title: metadata }),
+    );
+    card.append(element("span", { className: "match-rank", text: `#${index + 1}` }), identity, element("span", { className: "match-score", text: formatCosine(match.similarity) }));
+    target.append(card);
+  });
+}
+
+async function submitSearch(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const button = $("button[type=submit]", form);
+  const file = form.elements.image.files[0];
+  const payload = await guarded(button, () => client.search(form.elements.collection.value, file, {
+    limit: form.elements.limit.value,
+    threshold: form.elements.threshold.value,
+  }), "Search failed", "Searching…");
+  if (payload === undefined) return;
+  const searchedFace = payload.searched_face;
+  await renderImageFaces(file, $("#search-canvas"), [searchedFace].filter(Boolean), { labels: [t("Search face")] });
+  const matches = listItems(payload, ["matches"]);
+  renderMatches($("#search-results"), matches);
+  $("#search-summary").textContent = t(
+    matches.length === 1 ? "{count} match · threshold {threshold} · {duration}" : "{count} matches · threshold {threshold} · {duration}",
+    { count: matches.length, threshold: formatCosine(payload.threshold), duration: formatDuration(payload.processing_ms) },
+  );
+}
+
+async function loadSystem() {
+  const button = $("#system-refresh");
+  const result = await guarded(button, async () => {
+    const [systemPayload, modelsPayload] = await Promise.all([
+      client.system(),
+      client.models(),
+    ]);
+    return {
+      system: unwrap(systemPayload, "system"),
+      models: listItems(modelsPayload, ["models"]),
+    };
+  }, "Could not load system diagnostics", "Refreshing…");
+  if (!result) return;
+  state.system = result.system;
+  state.models = result.models;
+  updateCreateCollectionProfiles();
+  renderSystem();
+}
+
+function renderSystem() {
+  const system = state.system ?? {};
+  const osName = firstValue(readAny(system, "os.name"), readAny(system, "platform.os"), system.os, "—");
+  const architecture = firstValue(readAny(system, "os.architecture"), readAny(system, "platform.architecture"), system.architecture, "—");
+  const provider = systemProvider(system);
+  const gpuEntries = listItems(readAny(system, "runtime.gpus", "gpu", "gpus", "hardware.gpus"), ["devices"]);
+  const gpu = gpuEntries[0] ?? (typeof system.gpu === "object" ? system.gpu : {});
+  setText("#system-server", firstValue(system.server_version, system.version, "—"));
+  setText("#system-platform", `${osName} · ${architecture}`);
+  setText("#system-provider", provider);
+  setText("#system-ort", `ORT ${readAny(system, "runtime.onnx_runtime_version", "onnx_runtime_version", "onnx_runtime.version") ?? "—"}`);
+  setText("#system-gpu", firstValue(gpu.name, gpu.model, system.gpu_model, "CPU"));
+  setText("#system-driver", `Driver ${firstValue(gpu.driver_version, system.nvidia_driver, system.driver_version, "—")}`);
+  setText("#system-auth", t(client.hasApiKey ? "Configured" : "Not configured"));
+  const detectorInputSizes = readAny(system, "runtime.detector_input_sizes", "safe_config.detector_input_sizes");
+  const detectorInputLabel = Array.isArray(detectorInputSizes)
+    ? detectorInputSizes.map((size) => Array.isArray(size) ? size.join(" × ") : String(size)).join(" + ")
+    : "—";
+
+  $("#system-runtime").replaceChildren(
+    statusRow("OS", osName),
+    statusRow("Architecture", architecture),
+    statusRow("CPU", firstValue(readAny(system, "cpu.model"), system.cpu_model, system.cpu, "—")),
+    statusRow("GPU", firstValue(gpu.name, gpu.model, system.gpu_model, "Not reported")),
+    statusRow("Compute capability", firstValue(gpu.compute_capability, system.compute_capability, "—")),
+    statusRow("NVIDIA driver", firstValue(gpu.driver_version, system.nvidia_driver, system.driver_version, "—")),
+    statusRow("CUDA runtime", firstValue(readAny(system, "runtime.cuda_runtime_version"), system.cuda_runtime_version, readAny(system, "cuda.runtime_version"), "—")),
+    statusRow("cuDNN", firstValue(readAny(system, "runtime.cudnn_version"), system.cudnn_version, readAny(system, "cudnn.version"), "—")),
+    statusRow("ONNX Runtime", firstValue(readAny(system, "runtime.onnx_runtime_version"), system.onnx_runtime_version, readAny(system, "onnx_runtime.version"), "—")),
+    statusRow("Execution provider", provider),
+    statusRow("Inference concurrency", firstValue(readAny(system, "runtime.inference_concurrency.max_concurrency"), system.safe_config?.inference_max_concurrency, "—")),
+    statusRow("Detector input sizes", detectorInputLabel),
+    statusRow("Detection threshold", firstValue(system.safe_config?.detection?.threshold, "—")),
+    statusRow("Detector NMS threshold", firstValue(system.safe_config?.detection?.nms_threshold, "—")),
+    statusRow("Single-face selection", firstValue(system.safe_config?.detection?.single_face_selection, "—")),
+    statusRow("Startup config", firstValue(system.safe_config?.config_file, "built-in defaults")),
+  );
+  const databaseStatus = firstValue(readAny(system, "database.quick_check"), readAny(system, "database.status"), system.database_status, "unknown");
+  const dataStatus = system.data && typeof system.data === "object" ? (system.data.exists && system.data.writable ? "ready, writable" : "unavailable") : firstValue(readAny(system, "paths.data.status"), system.data_status, "unknown");
+  const modelsStatus = system.models && typeof system.models === "object" ? (system.models.exists && system.models.readable ? "ready, readable" : "unavailable") : firstValue(readAny(system, "models_path.status"), readAny(system, "paths.models.status"), system.models_status, "unknown");
+  const storageRows = [
+    statusRow("Database", databaseStatus, /ok|healthy|ready/i.test(String(databaseStatus)) ? "good" : ""),
+    statusRow("Database path", firstValue(readAny(system, "database.path"), system.database_path, "—")),
+    statusRow("/data", dataStatus, /ok|ready|writable/i.test(String(dataStatus)) ? "good" : ""),
+    statusRow("/data path", firstValue(readAny(system, "data.path"), readAny(system, "paths.data.path"), "/data")),
+    statusRow("/models", modelsStatus, /ok|ready|read.?only/i.test(String(modelsStatus)) ? "good" : ""),
+    statusRow("/models path", firstValue(readAny(system, "models.path"), readAny(system, "models_path.path"), readAny(system, "paths.models.path"), "/models")),
+    statusRow("Authentication", system.api_key?.authentication_enabled === false ? "development mode (disabled)" : system.api_key?.authentication_enabled === true ? "enabled" : "unknown"),
+    statusRow("Maximum image size", system.safe_config?.max_image_bytes ? `${(Number(system.safe_config.max_image_bytes) / 1024 / 1024).toFixed(1)} MiB` : "—"),
+    statusRow("Maximum pixels", firstValue(system.safe_config?.max_image_pixels, "—")),
+    statusRow("Request timeout", system.safe_config?.request_timeout_seconds ? t("{count} seconds", { count: system.safe_config.request_timeout_seconds }) : "—"),
+    statusRow("New-collection crop default", system.safe_config?.save_face_crops ? "enabled" : "disabled"),
+  ];
+  if (system.api_key?.authentication_enabled !== false) {
+    storageRows.splice(
+      7,
+      0,
+      statusRow("Server API key", system.api_key?.configured === true ? "configured" : system.api_key?.configured === false ? "not configured" : "unknown"),
+    );
+  }
+  $("#system-storage").replaceChildren(...storageRows);
+
+  const modelGrid = $("#system-models");
+  modelGrid.replaceChildren();
+  for (const model of systemModels(system)) {
+    const card = element("article", { className: "model-card" });
+    card.append(element("h4", { text: firstValue(model.model_id, model.id, "Unnamed model") }));
+    const details = document.createElement("dl");
+    details.append(
+      statusRow("Task", firstValue(model.task, "—")),
+      statusRow("Version", firstValue(model.model_version, model.version, "—")),
+      statusRow("Digest", firstValue(model.sha256, model.model_digest, "—")),
+      statusRow("Input", Array.isArray(model.input_size) ? model.input_size.join(" × ") : "—"),
+      statusRow("Embedding", firstValue(model.embedding_dimension, "—")),
+    );
+    card.append(details);
+    modelGrid.append(card);
+  }
+  if (!modelGrid.children.length) modelGrid.append(element("p", { className: "muted", text: "No loaded models were reported." }));
+}
+
+function selectedMonitor() {
+  return state.monitors.find((monitor) => monitor.id === state.selectedMonitorId) ?? null;
+}
+
+function videoIdentity(result) {
+  const person = result?.person ?? {};
+  return {
+    id: String(person.id ?? ""),
+    name: String(person.name ?? ""),
+    similarity: Number(result?.similarity),
+  };
+}
+
+function monitorStatusLabel(status) {
+  return {
+    running: "Running",
+    starting: "Starting",
+    reconnecting: "Reconnecting",
+    degraded: "Degraded",
+    stopped: "Stopped",
+    error: "Error",
+  }[status] || status || "Unavailable";
+}
+
+function monitorEventLabel(type) {
+  return {
+    person_enter: "Known person entered",
+    person_exit: "Known person left",
+    unknown_enter: "Unknown person entered",
+    unknown_exit: "Unknown person left",
+    monitor_error: "Monitor error",
+    monitor_recovered: "Monitor recovered",
+  }[type] || type || "Event";
+}
+
+function renderMonitorList() {
+  const target = $("#monitor-list");
+  target.replaceChildren();
+  $("#monitor-empty").hidden = state.monitors.length > 0;
+  for (const monitor of state.monitors) {
+    const button = element("button", {
+      className: `monitor-list-item ${monitor.id === state.selectedMonitorId ? "selected" : ""}`,
+    });
+    button.type = "button";
+    const identity = element("span");
+    identity.append(
+      element("strong", { text: monitor.name || monitor.id }),
+      element("small", { text: `${monitor.id} · ${monitor.collection_id}` }),
+    );
+    const runtimeStatus = monitor.runtime?.status || (monitor.enabled ? "starting" : "stopped");
+    const status = element("span", {
+      className: `monitor-list-state ${runtimeStatus}`,
+      text: monitorStatusLabel(runtimeStatus),
+    });
+    button.append(identity, status);
+    button.addEventListener("click", () => selectMonitor(monitor.id));
+    target.append(button);
+  }
+}
+
+function renderSelectedMonitor() {
+  const monitor = selectedMonitor();
+  $("#selected-monitor-name").textContent = monitor?.name || t("Select a Monitor");
+  $("#selected-monitor-id").textContent = monitor?.id || "—";
+  const toggle = $("#monitor-toggle");
+  const previewToggle = $("#monitor-preview-toggle");
+  const edit = $("#monitor-edit");
+  const remove = $("#monitor-delete");
+  for (const control of [toggle, previewToggle, edit, remove]) control.disabled = !monitor;
+  toggle.textContent = t(monitor?.enabled ? "Stop" : "Start");
+  previewToggle.textContent = t(monitor?.preview_enabled ? "Disable preview" : "Enable preview");
+  $("#monitor-config").hidden = !monitor;
+  $("#monitor-source").textContent = monitor?.source?.url || "—";
+  $("#monitor-collection").textContent = monitor?.collection_id || "—";
+  $("#monitor-threshold").textContent = monitor
+    ? monitor.match_threshold === null || monitor.match_threshold === undefined
+      ? t("Collection default")
+      : formatCosine(monitor.match_threshold)
+    : "—";
+  $("#monitor-event-capacity").textContent = monitor
+    ? String(monitor.event_buffer_size ?? "—")
+    : "—";
+  if (!monitor) {
+    $("#video-idle-title").textContent = t("No Monitor selected");
+    $("#video-idle-message").textContent = t("Select a Monitor to inspect its live recognition state.");
+    $("#video-status").textContent = t("Select a Monitor to inspect it.");
+    $("#video-metrics").hidden = true;
+    renderVideoMatches([]);
+    renderMonitorEvents();
+  } else if (!monitor.preview_enabled) {
+    $("#video-idle-title").textContent = t("Preview is disabled");
+    $("#video-idle-message").textContent = t("Recognition and events continue without video preview.");
+  } else if (!monitor.enabled) {
+    $("#video-idle-title").textContent = t("Monitor is stopped.");
+    $("#video-idle-message").textContent = t("Start the Monitor to receive frames.");
+  } else {
+    $("#video-idle-title").textContent = t("Waiting for preview");
+    $("#video-idle-message").textContent = t("The preview will appear after the RTSP source connects.");
+  }
+}
+
+async function loadMonitors({ preserveSelection = true } = {}) {
+  const payload = await guarded(
+    null,
+    () => client.listMonitors({ limit: 100 }),
+    "Could not load Monitors",
+  );
+  if (payload === undefined) return;
+  const previous = preserveSelection ? state.selectedMonitorId : "";
+  state.monitors = listItems(payload, ["monitors"]);
+  state.selectedMonitorId = state.monitors.some((item) => item.id === previous)
+    ? previous
+    : state.monitors[0]?.id || "";
+  renderMonitorList();
+  renderSelectedMonitor();
+  if (state.selectedMonitorId) {
+    state.monitorEventCursor = "";
+    state.monitorEvents = [];
+    stopMonitorPreview();
+    void pollMonitor(state.selectedMonitorId);
+  }
+}
+
+function selectMonitor(monitorId) {
+  if (monitorId === state.selectedMonitorId) return;
+  if (state.editingMonitorId) closeMonitorForm();
+  state.selectedMonitorId = monitorId;
+  state.monitorEventCursor = "";
+  state.monitorEvents = [];
+  state.videoStatus = null;
+  if (state.monitorPollTimer !== null) window.clearTimeout(state.monitorPollTimer);
+  state.monitorPollTimer = null;
+  stopMonitorPreview();
+  renderMonitorList();
+  renderSelectedMonitor();
+  renderMonitorEvents();
+  void pollMonitor(monitorId);
+}
+
+function renderVideoMatches(recognitions = []) {
+  const target = $("#video-matches");
+  target.replaceChildren();
+  recognitions.forEach((result, index) => {
+    const recognized = result?.status === "matched";
+    const row = element("div", { className: `video-match ${recognized ? "recognized" : "not-enrolled"}` });
+    const number = element("span", { className: "face-index", text: String(index + 1) });
+    const summary = element("div", { className: "video-match-person" });
+    if (recognized) {
+      const identity = videoIdentity(result);
+      summary.append(
+        element("strong", { text: identity.name || identity.id }),
+        element("small", { text: t("ID {id}", { id: identity.id }) }),
+      );
+      row.append(number, summary, element("strong", { text: formatCosine(identity.similarity) }));
+    } else {
+      const score = formatScore(result?.face?.detection_score);
+      summary.append(
+        element("strong", { text: t("Detected, not enrolled") }),
+        element("small", { text: t("Detection score {score}", { score }) }),
+      );
+      row.append(number, summary, element("strong", { text: "—" }));
+    }
+    target.append(row);
+  });
+}
+
+function drawMonitorOverlay(stream) {
+  const canvas = $("#rtsp-overlay");
+  const stage = canvas.parentElement;
+  const width = stage.clientWidth;
+  const height = stage.clientHeight;
+  if (!width || !height) return;
+  const ratio = window.devicePixelRatio || 1;
+  canvas.width = Math.round(width * ratio);
+  canvas.height = Math.round(height * ratio);
+  const context = canvas.getContext("2d");
+  context.scale(ratio, ratio);
+  context.clearRect(0, 0, width, height);
+  const sourceWidth = Number(stream?.source?.width);
+  const sourceHeight = Number(stream?.source?.height);
+  if (!sourceWidth || !sourceHeight || !stream?.connected) return;
+  const scale = Math.min(width / sourceWidth, height / sourceHeight);
+  const renderedWidth = sourceWidth * scale;
+  const renderedHeight = sourceHeight * scale;
+  const offsetX = (width - renderedWidth) / 2;
+  const offsetY = (height - renderedHeight) / 2;
+  context.font = "700 13px system-ui, sans-serif";
+  context.lineWidth = 2;
+  for (const result of stream.faces ?? []) {
+    const box = result?.face?.bbox?.normalized;
+    if (!box) continue;
+    const x = offsetX + Number(box.left) * renderedWidth;
+    const y = offsetY + Number(box.top) * renderedHeight;
+    const boxWidth = Number(box.width) * renderedWidth;
+    const boxHeight = Number(box.height) * renderedHeight;
+    const recognized = result.status === "matched";
+    const color = recognized ? "#57d368" : "#ff9e20";
+    const identity = videoIdentity(result);
+    const label = recognized
+      ? `${identity.name || identity.id} ${formatCosine(identity.similarity)}`
+      : t("Not enrolled");
+    context.strokeStyle = color;
+    context.strokeRect(x, y, boxWidth, boxHeight);
+    const labelWidth = context.measureText(label).width + 12;
+    const labelY = Math.max(0, y - 23);
+    context.fillStyle = color;
+    context.fillRect(x, labelY, labelWidth, 23);
+    context.fillStyle = "#09090b";
+    context.fillText(label, x + 6, labelY + 16);
+  }
+}
+
+function renderVideoStatus(stream) {
+  state.videoStatus = stream;
+  $("#video-metrics").hidden = false;
+  const connected = stream?.connected === true;
+  const statusLabel = connected
+    ? "Connected"
+    : ["starting", "connecting"].includes(stream?.status)
+      ? "Connecting"
+      : stream?.status === "reconnecting"
+        ? "Reconnecting"
+        : stream?.status || "Unavailable";
+  $("#video-connection").textContent = t(statusLabel);
+  const effective = Number(stream?.inference?.actual_fps);
+  const requested = Number(stream?.inference?.configured_fps);
+  $("#video-effective-fps").textContent = Number.isFinite(effective) && Number.isFinite(requested)
+    ? t("{effective} / {requested} per second", { effective: effective.toFixed(2), requested: requested.toFixed(2) })
+    : "—";
+  $("#video-latency").textContent = formatDuration(stream?.inference?.processing_ms);
+  $("#video-skipped").textContent = String(stream?.inference?.dropped_frames ?? 0);
+  renderVideoMatches(stream?.faces ?? []);
+  drawMonitorOverlay(stream);
+
+  const monitor = selectedMonitor();
+  if (!monitor?.enabled) {
+    $("#video-status").textContent = t("Monitor is stopped.");
+  } else if (stream?.last_error && !connected) {
+    $("#video-status").textContent = t("RTSP stream unavailable. Reconnecting…");
+  } else if (!connected) {
+    $("#video-status").textContent = t("Connecting to RTSP camera…");
+  } else {
+    const timing = stream?.inference?.capacity_limited
+      ? t("Inference-limited; stale frames are being skipped")
+      : t("Running at the configured cadence");
+    $("#video-status").textContent = t("{recognized} recognized · {unknown} not enrolled · {timing}", {
+      recognized: Number(stream?.matched_faces ?? 0),
+      unknown: Number(stream?.unknown_faces ?? 0),
+      timing,
+    });
+  }
+  if (monitor?.preview_enabled && connected) startMonitorPreview(monitor.id);
+  else stopMonitorPreview();
+}
+
+function renderMonitorEvents() {
+  const target = $("#monitor-events");
+  target.replaceChildren();
+  $("#monitor-event-count").textContent = String(state.monitorEvents.length);
+  for (const event of [...state.monitorEvents].reverse()) {
+    const row = element("div", { className: "monitor-event" });
+    const eventLabel = monitorEventLabel(event.type);
+    const type = element("span", { className: "badge", text: eventLabel });
+    const summary = element("span");
+    const identity = event.person?.name
+      || event.person?.id
+      || (event.type === "monitor_error" || event.type === "monitor_recovered"
+        ? t(eventLabel)
+        : t("Unknown person"));
+    summary.append(
+      element("strong", { text: identity }),
+      element("small", {
+        text: event.similarity === null || event.similarity === undefined
+          ? event.track_id || event.error?.message || "—"
+          : t("Similarity {score}", { score: formatCosine(event.similarity) }),
+      }),
+    );
+    const occurred = event.occurred_at ? new Date(event.occurred_at) : null;
+    const timestamp = element("time", {
+      text: occurred && !Number.isNaN(occurred.getTime()) ? occurred.toLocaleTimeString() : "—",
+    });
+    row.append(type, summary, timestamp);
+    target.append(row);
+  }
+}
+
+async function pollMonitorEvents(monitorId) {
+  const query = { limit: 100 };
+  if (state.monitorEventCursor) query.cursor = state.monitorEventCursor;
+  const payload = await client.monitorEvents(monitorId, query);
+  if (monitorId !== state.selectedMonitorId) return;
+  if (payload.stream_reset || payload.truncated) state.monitorEvents = [];
+  state.monitorEvents.push(...(payload.events ?? []));
+  state.monitorEvents = state.monitorEvents.slice(-1000);
+  state.monitorEventCursor = payload.next_cursor || "";
+  renderMonitorEvents();
+}
+
+async function pollMonitor(monitorId) {
+  if (!monitorId || monitorId !== state.selectedMonitorId || state.route !== "video") return;
+  try {
+    const [statusPayload, monitorPayload] = await Promise.all([
+      client.monitorState(monitorId),
+      client.getMonitor(monitorId),
+      pollMonitorEvents(monitorId),
+    ]);
+    if (monitorId !== state.selectedMonitorId) return;
+    const refreshed = unwrap(monitorPayload, "monitor");
+    state.monitors = state.monitors.map((item) => item.id === monitorId ? refreshed : item);
+    renderMonitorList();
+    renderSelectedMonitor();
+    renderVideoStatus(unwrap(statusPayload, "state"));
+  } catch (error) {
+    if (monitorId !== state.selectedMonitorId) return;
+    $("#video-status").textContent = describeError(error);
+    if (error instanceof ApiError && error.status === 404) {
+      await loadMonitors({ preserveSelection: false });
+      return;
+    }
+  }
+  if (monitorId === state.selectedMonitorId && state.route === "video") {
+    state.monitorPollTimer = window.setTimeout(() => void pollMonitor(monitorId), 1000);
+  }
+}
+
+function concatBytes(left, right) {
+  const combined = new Uint8Array(left.length + right.length);
+  combined.set(left);
+  combined.set(right, left.length);
+  return combined;
+}
+
+function headerEndIndex(buffer) {
+  for (let index = 0; index <= buffer.length - 4; index += 1) {
+    if (buffer[index] === 13 && buffer[index + 1] === 10 && buffer[index + 2] === 13 && buffer[index + 3] === 10) return index;
+  }
+  return -1;
+}
+
+async function consumeMjpeg(response, signal) {
+  const reader = response.body?.getReader();
+  if (!reader) throw new Error("The browser cannot read this preview stream.");
+  let buffer = new Uint8Array();
+  while (!signal.aborted) {
+    const { value, done } = await reader.read();
+    if (done) break;
+    buffer = concatBytes(buffer, value);
+    while (buffer.length) {
+      const headerEnd = headerEndIndex(buffer);
+      if (headerEnd < 0) break;
+      const header = new TextDecoder("ascii").decode(buffer.slice(0, headerEnd));
+      const lengthMatch = header.match(/Content-Length:\s*(\d+)/i);
+      if (!lengthMatch) {
+        buffer = buffer.slice(headerEnd + 4);
+        continue;
+      }
+      const contentLength = Number(lengthMatch[1]);
+      const contentStart = headerEnd + 4;
+      if (buffer.length < contentStart + contentLength) break;
+      const jpeg = buffer.slice(contentStart, contentStart + contentLength);
+      buffer = buffer.slice(Math.min(buffer.length, contentStart + contentLength + 2));
+      const url = URL.createObjectURL(new Blob([jpeg], { type: "image/jpeg" }));
+      const preview = $("#rtsp-preview");
+      const previous = state.previewObjectUrl;
+      state.previewObjectUrl = url;
+      preview.src = url;
+      preview.hidden = false;
+      $("#rtsp-overlay").hidden = false;
+      $("#video-idle").hidden = true;
+      if (previous) URL.revokeObjectURL(previous);
+    }
+  }
+}
+
+function startMonitorPreview(monitorId) {
+  if (state.previewAbortController || monitorId !== state.selectedMonitorId) return;
+  const controller = new AbortController();
+  state.previewAbortController = controller;
+  void (async () => {
+    try {
+      const response = await client.monitorPreview(monitorId, { signal: controller.signal });
+      await consumeMjpeg(response, controller.signal);
+    } catch (error) {
+      if (!controller.signal.aborted && monitorId === state.selectedMonitorId) {
+        $("#video-status").textContent = describeError(error);
+      }
+    } finally {
+      if (state.previewAbortController === controller) state.previewAbortController = null;
+    }
+  })();
+}
+
+function stopMonitorPreview() {
+  state.previewAbortController?.abort();
+  state.previewAbortController = null;
+  const preview = $("#rtsp-preview");
+  preview.removeAttribute("src");
+  preview.hidden = true;
+  $("#rtsp-overlay").hidden = true;
+  const context = $("#rtsp-overlay").getContext("2d");
+  context.clearRect(0, 0, $("#rtsp-overlay").width, $("#rtsp-overlay").height);
+  if (state.previewObjectUrl) URL.revokeObjectURL(state.previewObjectUrl);
+  state.previewObjectUrl = "";
+  $("#video-idle").hidden = false;
+}
+
+function stopMonitorView() {
+  if (state.monitorPollTimer !== null) window.clearTimeout(state.monitorPollTimer);
+  state.monitorPollTimer = null;
+  stopMonitorPreview();
+}
+
+function resetMonitorFormDefaults() {
+  const form = $("#monitor-create-form");
+  form.reset();
+  form.elements.inference_fps.value = "2";
+  form.elements.event_buffer_size.value = "1000";
+  form.elements.confirm_frames.value = "3";
+  form.elements.absence_timeout_seconds.value = "3";
+  form.elements.cooldown_seconds.value = "10";
+  form.elements.enabled.checked = true;
+  form.elements.emit_unknown.checked = true;
+  form.elements.preview_enabled.checked = false;
+}
+
+function renderMonitorFormMode() {
+  const editing = Boolean(state.editingMonitorId);
+  const form = $("#monitor-create-form");
+  $("#monitor-form-eyebrow").textContent = t(editing ? "EDIT MONITOR" : "NEW MONITOR");
+  $("#monitor-form-title").textContent = t(editing ? "Edit Monitor" : "Create Monitor");
+  $("#monitor-form-submit").textContent = t(editing ? "Save changes" : "Create Monitor");
+  $("#monitor-enabled-label").textContent = t(editing ? "Enabled" : "Start after creation");
+  $("#monitor-enabled-help").textContent = t(
+    editing ? "Run this Monitor on the server." : "Run independently of this browser.",
+  );
+  form.elements.id.readOnly = editing;
+  form.elements.url.required = !editing;
+  form.elements.url.placeholder = editing
+    ? t("Leave blank to keep the current RTSP URL")
+    : "rtsp://camera.example:8554/stream";
+}
+
+function openCreateMonitorForm() {
+  state.editingMonitorId = "";
+  resetMonitorFormDefaults();
+  renderMonitorFormMode();
+  const form = $("#monitor-create-form");
+  form.hidden = false;
+  form.elements.id.focus();
+}
+
+function openEditMonitorForm() {
+  const monitor = selectedMonitor();
+  if (!monitor) return;
+  state.editingMonitorId = monitor.id;
+  resetMonitorFormDefaults();
+  const form = $("#monitor-create-form");
+  form.elements.id.value = monitor.id;
+  form.elements.name.value = monitor.name ?? "";
+  form.elements.description.value = monitor.description ?? "";
+  form.elements.url.value = "";
+  form.elements.collection_id.value = monitor.collection_id;
+  form.elements.inference_fps.value = String(monitor.inference_fps ?? 2);
+  form.elements.match_threshold.value = monitor.match_threshold ?? "";
+  form.elements.event_buffer_size.value = String(monitor.event_buffer_size ?? 1000);
+  form.elements.confirm_frames.value = String(monitor.event_policy?.confirm_frames ?? 3);
+  form.elements.absence_timeout_seconds.value = String(monitor.event_policy?.absence_timeout_seconds ?? 3);
+  form.elements.cooldown_seconds.value = String(monitor.event_policy?.cooldown_seconds ?? 10);
+  form.elements.enabled.checked = monitor.enabled === true;
+  form.elements.emit_unknown.checked = monitor.event_policy?.emit_unknown !== false;
+  form.elements.preview_enabled.checked = monitor.preview_enabled === true;
+  renderMonitorFormMode();
+  form.hidden = false;
+  form.elements.name.focus();
+}
+
+function closeMonitorForm() {
+  state.editingMonitorId = "";
+  $("#monitor-create-form").hidden = true;
+}
+
+async function saveMonitor(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const editingMonitorId = state.editingMonitorId;
+  const common = {
+    name: form.elements.name.value.trim(),
+    description: form.elements.description.value.trim(),
+    enabled: form.elements.enabled.checked,
+    collection_id: form.elements.collection_id.value,
+    inference_fps: Number(form.elements.inference_fps.value),
+    match_threshold: form.elements.match_threshold.value === ""
+      ? null
+      : Number(form.elements.match_threshold.value),
+    event_buffer_size: Number(form.elements.event_buffer_size.value),
+    event_policy: {
+      confirm_frames: Number(form.elements.confirm_frames.value),
+      absence_timeout_seconds: Number(form.elements.absence_timeout_seconds.value),
+      cooldown_seconds: Number(form.elements.cooldown_seconds.value),
+      emit_unknown: form.elements.emit_unknown.checked,
+    },
+    preview_enabled: form.elements.preview_enabled.checked,
+  };
+  const rtspUrl = form.elements.url.value.trim();
+  if (editingMonitorId && rtspUrl) common.source = { type: "rtsp", url: rtspUrl };
+  const payload = await guarded(
+    $("button[type=submit]", form),
+    () => editingMonitorId
+      ? client.updateMonitor(editingMonitorId, common)
+      : client.createMonitor({
+          id: form.elements.id.value.trim(),
+          name: common.name,
+          description: common.description,
+          enabled: common.enabled,
+          url: rtspUrl,
+          collectionId: common.collection_id,
+          inferenceFps: common.inference_fps,
+          matchThreshold: common.match_threshold,
+          eventBufferSize: common.event_buffer_size,
+          confirmFrames: common.event_policy.confirm_frames,
+          absenceTimeoutSeconds: common.event_policy.absence_timeout_seconds,
+          cooldownSeconds: common.event_policy.cooldown_seconds,
+          emitUnknown: common.event_policy.emit_unknown,
+          previewEnabled: common.preview_enabled,
+        }),
+    editingMonitorId ? "Monitor could not be updated" : "Monitor could not be created",
+    editingMonitorId ? "Saving…" : "Creating…",
+  );
+  if (payload === undefined) return;
+  const monitor = unwrap(payload, "monitor");
+  toast(editingMonitorId ? "Monitor updated" : "Monitor created", monitor.id);
+  closeMonitorForm();
+  if (editingMonitorId) {
+    state.monitors = state.monitors.map((item) => item.id === monitor.id ? monitor : item);
+    if (!monitor.preview_enabled || !monitor.enabled) stopMonitorPreview();
+    renderMonitorList();
+    renderSelectedMonitor();
+    void pollMonitor(monitor.id);
+    return;
+  }
+  await loadMonitors({ preserveSelection: false });
+  selectMonitor(monitor.id);
+}
+
+async function toggleMonitorEnabled() {
+  const monitor = selectedMonitor();
+  if (!monitor) return;
+  const payload = await guarded(
+    $("#monitor-toggle"),
+    () => client.updateMonitor(monitor.id, { enabled: !monitor.enabled }),
+    "Monitor could not be updated",
+    monitor.enabled ? "Stopping…" : "Starting…",
+  );
+  if (payload === undefined) return;
+  const updated = unwrap(payload, "monitor");
+  state.monitors = state.monitors.map((item) => item.id === updated.id ? updated : item);
+  if (!updated.enabled) stopMonitorPreview();
+  renderMonitorList();
+  renderSelectedMonitor();
+  void pollMonitor(updated.id);
+}
+
+async function toggleMonitorPreview() {
+  const monitor = selectedMonitor();
+  if (!monitor) return;
+  const payload = await guarded(
+    $("#monitor-preview-toggle"),
+    () => client.updateMonitor(monitor.id, { preview_enabled: !monitor.preview_enabled }),
+    "Monitor preview could not be updated",
+    "Saving…",
+  );
+  if (payload === undefined) return;
+  const updated = unwrap(payload, "monitor");
+  state.monitors = state.monitors.map((item) => item.id === updated.id ? updated : item);
+  if (!updated.preview_enabled) stopMonitorPreview();
+  renderMonitorList();
+  renderSelectedMonitor();
+}
+
+async function deleteSelectedMonitor() {
+  const monitor = selectedMonitor();
+  if (!monitor || !window.confirm(t("Delete Monitor {id}?", { id: monitor.id }))) return;
+  const result = await guarded(
+    $("#monitor-delete"),
+    () => client.deleteMonitor(monitor.id),
+    "Monitor could not be deleted",
+    "Deleting…",
+  );
+  if (result === undefined) return;
+  toast("Monitor deleted", monitor.id);
+  stopMonitorView();
+  state.selectedMonitorId = "";
+  state.monitorEvents = [];
+  state.monitorEventCursor = "";
+  await loadMonitors({ preserveSelection: false });
+}
+
+function bindEvents() {
+  $$('[data-route]').forEach((control) => control.addEventListener("click", (event) => {
+    if (control.tagName === "A") return;
+    event.preventDefault();
+    navigate(control.dataset.route);
+  }));
+  $("#menu-toggle").addEventListener("click", () => {
+    const open = document.body.classList.toggle("menu-open");
+    $("#menu-toggle").setAttribute("aria-expanded", String(open));
+  });
+  $("#refresh-page").addEventListener("click", () => void refreshRoute());
+  $("#clear-errors").addEventListener("click", () => { state.errors = []; state.serverErrors = []; renderRecentErrors(); });
+
+  const keyDialog = $("#api-key-dialog");
+  $("#open-api-key").addEventListener("click", () => keyDialog.showModal());
+  $("#api-key-form").addEventListener("submit", (event) => {
+    event.preventDefault();
+    client.setApiKey(event.currentTarget.elements.api_key.value);
+    event.currentTarget.elements.api_key.value = "";
+    keyDialog.close();
+    updateAuthState();
+    toast("API key configured", "It will be forgotten when this tab closes.");
+    void refreshRoute();
+  });
+  $("#clear-api-key").addEventListener("click", () => {
+    client.clearApiKey();
+    $("#api-key-form").elements.api_key.value = "";
+    keyDialog.close();
+    updateAuthState();
+    toast("API key cleared");
+  });
+
+  $("#show-create-collection").addEventListener("click", () => {
+    updateCreateCollectionProfiles();
+    const form = $("#collection-create-form");
+    form.elements.save_face_crops.checked = newCollectionCropDefault();
+    const detection = systemDetectionDefaults();
+    form.elements.detector_input_sizes.value = formatDetectorInputSizes(detection.input_sizes);
+    form.elements.detector_threshold.value = String(detection.threshold);
+    form.elements.detector_nms_threshold.value = String(detection.nms_threshold);
+    form.elements.single_face_selection.value = detection.single_face_selection;
+    form.hidden = false;
+    $("input", form).focus();
+  });
+  $("#collection-create-form .close-form").addEventListener("click", () => { $("#collection-create-form").hidden = true; });
+  $("#collection-create-form").addEventListener("submit", (event) => void createCollection(event));
+  $("#collection-filter").addEventListener("input", renderCollections);
+  $("#collection-edit-form").addEventListener("submit", (event) => void updateCollection(event));
+
+  $("#people-collection").addEventListener("change", (event) => { state.selectedCollection = event.target.value; state.selectedPerson = null; configureEmbeddingFields($("#person-create-form")); void loadPeople(); });
+  $("#people-filter").addEventListener("input", renderPeople);
+  $("#show-create-person").addEventListener("click", () => {
+    if (!state.selectedCollection) { handleError(new Error("Select a collection first."), "Cannot register person"); return; }
+    configureEmbeddingFields($("#person-create-form"));
+    $("#person-create-form").hidden = false;
+  });
+  $("#person-create-form .close-form").addEventListener("click", () => { $("#person-create-form").hidden = true; });
+  $("#person-create-form").elements.embedding_mode.addEventListener("change", (event) => configureEmbeddingFields(event.currentTarget.form));
+  $("#person-create-form").addEventListener("submit", (event) => void createPerson(event));
+
+  $("#detect-form").addEventListener("submit", (event) => void submitDetect(event));
+  $("#compare-form").addEventListener("submit", (event) => void submitCompare(event));
+  $("#search-form").addEventListener("submit", (event) => void submitSearch(event));
+  $("#compare-threshold").addEventListener("input", (event) => { $("#compare-threshold-output").value = Number(event.target.value).toFixed(2); });
+
+  $("#system-refresh").addEventListener("click", () => void loadSystem());
+  $("#documentation-filter").addEventListener("input", filterDocumentationToc);
+  window.addEventListener("insightface:localechange", () => {
+    updateAuthState();
+    renderRecentErrors();
+    if (state.route === "collections") renderCollections();
+    if (state.route === "people") renderPeople();
+    if (state.route === "video") {
+      renderMonitorList();
+      renderSelectedMonitor();
+      renderMonitorEvents();
+      if (state.videoStatus) renderVideoStatus(state.videoStatus);
+      if (!$("#monitor-create-form").hidden) renderMonitorFormMode();
+    }
+    setText("#page-title", t(pageTitles[state.route]));
+    document.title = `${t(pageTitles[state.route])} · InsightFace Server`;
+    if (state.route === "help") void loadDocumentation();
+  });
+  $("#show-create-monitor").addEventListener("click", openCreateMonitorForm);
+  $("#monitor-create-form .close-form").addEventListener("click", closeMonitorForm);
+  $("#monitor-create-form").addEventListener("submit", (event) => void saveMonitor(event));
+  $("#refresh-monitors").addEventListener("click", () => void loadMonitors());
+  $("#monitor-toggle").addEventListener("click", () => void toggleMonitorEnabled());
+  $("#monitor-preview-toggle").addEventListener("click", () => void toggleMonitorPreview());
+  $("#monitor-edit").addEventListener("click", openEditMonitorForm);
+  $("#monitor-delete").addEventListener("click", () => void deleteSelectedMonitor());
+  $("#rtsp-preview").addEventListener("load", () => { $("#video-idle").hidden = true; });
+  $("#rtsp-preview").addEventListener("error", () => {
+    if (state.selectedMonitorId) $("#video-status").textContent = t("Waiting for the first RTSP frame…");
+  });
+  window.addEventListener("beforeunload", stopMonitorView);
+  window.addEventListener("beforeunload", releaseAllCropObjectUrls);
+
+  $$('input[type="file"]').forEach(setupFileInput);
+  setupImagePreview($("#detect-form input[name=image]"), $("#detect-canvas"), () => {
+    $("#detect-results").replaceChildren();
+    $("#detect-summary").textContent = t("Image ready. Press Detect faces to process it.");
+  });
+  setupImagePreview($("#compare-form input[name=source]"), $("#compare-source-canvas"), resetCompareVerdict);
+  setupImagePreview($("#compare-form input[name=target]"), $("#compare-target-canvas"), resetCompareVerdict);
+  setupImagePreview($("#search-form input[name=image]"), $("#search-canvas"), () => {
+    $("#search-results").replaceChildren();
+    $("#search-summary").textContent = t("Image ready. Press Search people to process it.");
+  });
+}
+
+function resetCompareVerdict() {
+  const verdict = $("#compare-verdict");
+  verdict.className = "compare-verdict";
+  $(".verdict-mark", verdict).textContent = "—";
+  $("strong", verdict).textContent = t("Ready to compare");
+  $("small", verdict).textContent = t("Similarity is not a probability.");
+}
+
+function updateAuthState() {
+  const stateNode = $("#auth-state");
+  stateNode.classList.toggle("configured", client.hasApiKey);
+  stateNode.innerHTML = "";
+  stateNode.append(element("span", { text: client.hasApiKey ? "●" : "○" }), document.createTextNode(` ${t(client.hasApiKey ? "API key configured" : "No API key")}`));
+  setText("#system-auth", t(client.hasApiKey ? "Configured" : "Not configured"));
+}
+
+async function createCollection(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const button = $("button[type=submit]", form);
+  let metadata;
+  try { metadata = parseMetadata(form.elements.metadata.value); } catch (error) { handleError(error, "Invalid collection metadata"); return; }
+  let detectorInputSizes;
+  try { detectorInputSizes = parseDetectorInputSizes(form.elements.detector_input_sizes.value); } catch (error) { handleError(error, "Invalid detection profile"); return; }
+  const payload = await guarded(button, () => client.createCollection({
+    id: form.elements.id.value.trim(),
+    name: form.elements.name.value.trim(),
+    description: form.elements.description.value.trim(),
+    threshold: Number(form.elements.threshold.value),
+    metadata,
+    save_face_crops: form.elements.save_face_crops.checked,
+    detection: {
+      input_sizes: detectorInputSizes,
+      threshold: Number(form.elements.detector_threshold.value),
+      nms_threshold: Number(form.elements.detector_nms_threshold.value),
+      single_face_selection: form.elements.single_face_selection.value,
+    },
+    search: {
+      profile: form.elements.search_profile.value,
+      capacity_rows: Number(form.elements.capacity_rows.value),
+      max_faces_per_person: Number(form.elements.max_faces_per_person.value),
+      ...(form.elements.load_policy.value ? { load_policy: form.elements.load_policy.value } : {}),
+    },
+  }), "Collection could not be created", "Creating…");
+  if (payload === undefined) return;
+  const collection = unwrap(payload, "collection");
+  toast("Collection created", collection.id ?? form.elements.id.value);
+  form.reset();
+  form.elements.threshold.value = "0.4";
+  form.elements.search_profile.value = "fp32_v1";
+  form.elements.capacity_rows.value = "100000";
+  form.elements.max_faces_per_person.value = "20";
+  form.elements.load_policy.value = "";
+  form.elements.save_face_crops.checked = newCollectionCropDefault();
+  const detection = systemDetectionDefaults();
+  form.elements.detector_input_sizes.value = formatDetectorInputSizes(detection.input_sizes);
+  form.elements.detector_threshold.value = String(detection.threshold);
+  form.elements.detector_nms_threshold.value = String(detection.nms_threshold);
+  form.elements.single_face_selection.value = detection.single_face_selection;
+  form.elements.metadata.value = "{}";
+  updateCreateCollectionProfiles();
+  form.hidden = true;
+  await loadCollections();
+}
+
+async function updateCollection(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  let metadata;
+  try { metadata = parseMetadata(form.elements.metadata.value); } catch (error) { handleError(error, "Invalid collection metadata"); return; }
+  let detectorInputSizes;
+  try { detectorInputSizes = parseDetectorInputSizes(form.elements.detector_input_sizes.value); } catch (error) { handleError(error, "Invalid detection profile"); return; }
+  const payload = await guarded($("button[type=submit]", form), () => client.updateCollection(form.elements.id.value, {
+    name: form.elements.name.value.trim(),
+    description: form.elements.description.value.trim(),
+    threshold: Number(form.elements.threshold.value),
+    metadata,
+    save_face_crops: form.elements.save_face_crops.checked,
+    detection: {
+      input_sizes: detectorInputSizes,
+      threshold: Number(form.elements.detector_threshold.value),
+      nms_threshold: Number(form.elements.detector_nms_threshold.value),
+      single_face_selection: form.elements.single_face_selection.value,
+    },
+    search: {
+      capacity_rows: Number(form.elements.capacity_rows.value),
+      max_faces_per_person: Number(form.elements.max_faces_per_person.value),
+      load_policy: form.elements.load_policy.value,
+    },
+  }), "Collection could not be updated", "Saving…");
+  if (payload === undefined) return;
+  $("#collection-edit-dialog").close();
+  toast("Collection updated", form.elements.id.value);
+  await loadCollections();
+}
+
+async function createPerson(event) {
+  event.preventDefault();
+  const form = event.currentTarget;
+  const images = [...form.elements.images.files];
+  let metadata;
+  try { metadata = parseMetadata(form.elements.metadata.value); } catch (error) { handleError(error, "Invalid person metadata"); return; }
+  let embeddingInput;
+  try { embeddingInput = enrollmentEmbeddingInput(form, images); } catch (error) { handleError(error, "Invalid external embeddings"); return; }
+  const button = $("button[type=submit]", form);
+  setBusy(button, true, "Registering…");
+  let payload;
+  try {
+    payload = await client.createPerson(state.selectedCollection, {
+      id: form.elements.id.value.trim(),
+      name: form.elements.name.value.trim(),
+      externalId: form.elements.external_id.value.trim(),
+      metadata,
+      images,
+      reviewMode: form.elements.review_mode.value,
+      ...embeddingInput,
+    });
+  } catch (error) {
+    const rejected = error instanceof ApiError ? error.details?.rejected_images : null;
+    if (Array.isArray(rejected)) renderRejections($("#person-rejections"), rejected, images);
+    handleError(error, "Person could not be registered");
+    return;
+  } finally {
+    setBusy(button, false);
+  }
+  const faces = listItems(payload, ["faces", "face_samples"]);
+  const rejected = payload.rejected_images ?? payload.rejected ?? [];
+  renderRejections($("#person-rejections"), rejected, images);
+  toast("Registration complete", t("{accepted} accepted, {rejected} rejected.", { accepted: faces.length, rejected: rejected.length }));
+  form.elements.images.value = "";
+  form.elements.external_embeddings.value = "";
+  $(".file-summary", form).textContent = t("No files selected");
+  await loadPeople();
+  const person = unwrap(payload, "person");
+  const listed = state.people.find((item) => item.id === person?.id);
+  if (listed) await selectPerson(listed);
+}
+
+async function initialize() {
+  initializeI18n();
+  bindEvents();
+  translateTree(document);
+  updateAuthState();
+  renderRecentErrors();
+  try { await refreshHealth(); } catch { /* Health UI already shows unavailable. */ }
+  const requestedRoute = window.location.hash.slice(1);
+  const initialRoute = requestedRoute === "camera" ? "video" : requestedRoute;
+  navigate(pageTitles[initialRoute] ? initialRoute : "dashboard");
+}
+
+void initialize();
diff --git a/server/frontend/core.mjs b/server/frontend/core.mjs
new file mode 100644
index 000000000..215222ae3
--- /dev/null
+++ b/server/frontend/core.mjs
@@ -0,0 +1,211 @@
+/**
+ * Browser-independent helpers shared by the UI and its Node tests.
+ * Keep this module free of DOM globals so it can be imported by `node --test`.
+ */
+
+export const DEFAULT_THRESHOLD = 0.4;
+export const SEARCH_PROFILES = Object.freeze([
+  "fp32_v1",
+  "fp16_v1",
+  "bf16_v1",
+  "int8_x736_v1",
+  "int8_x1000_v1",
+]);
+
+/**
+ * Older servers did not publish auth_enabled in their health response and
+ * required authentication by default. Keep their API-key controls visible.
+ */
+export function authenticationEnabledFromHealth(health) {
+  return health?.auth_enabled !== false;
+}
+
+/**
+ * Return the public profiles reported by GET /v1/system. `null` means the
+ * server did not expose capability information (for compatibility with an
+ * older server); an empty array means it explicitly supports none of them.
+ */
+export function searchProfilesFromSystem(system) {
+  const profiles = system?.search?.profiles;
+  if (!Array.isArray(profiles)) return null;
+  const reported = new Set(profiles.filter((profile) => typeof profile === "string"));
+  return SEARCH_PROFILES.filter((profile) => reported.has(profile));
+}
+
+/**
+ * Apply runtime capabilities to the create-Collection profile control.
+ * Existing Collection profile fields are deliberately not passed here: a
+ * Collection is pinned to its profile and its edit form only displays it.
+ */
+export function applySearchProfileAvailability(select, supportedProfiles) {
+  if (!select) return "";
+  const options = [...(select.options ?? [])];
+  const reported = Array.isArray(supportedProfiles) ? new Set(supportedProfiles) : null;
+
+  for (const option of options) {
+    option.disabled = reported !== null && !reported.has(option.value);
+  }
+
+  const enabled = options.filter((option) => !option.disabled);
+  if (!enabled.some((option) => option.value === select.value)) {
+    select.value = enabled[0]?.value ?? "";
+  }
+  select.disabled = enabled.length === 0;
+  return select.value;
+}
+
+export function clamp(value, minimum, maximum) {
+  const number = Number(value);
+  if (!Number.isFinite(number)) return minimum;
+  return Math.min(maximum, Math.max(minimum, number));
+}
+
+export function normalizeThreshold(value, fallback = DEFAULT_THRESHOLD) {
+  if (value === "" || value === null || value === undefined) return fallback;
+  return clamp(value, 0, 1);
+}
+
+export function formatScore(value, digits = 3) {
+  const score = Number(value);
+  if (!Number.isFinite(score)) return "—";
+  return score.toFixed(digits);
+}
+
+export function formatCosine(value, digits = 4) {
+  const cosine = Number(value);
+  if (!Number.isFinite(cosine)) return "—";
+  return cosine.toFixed(digits);
+}
+
+export function formatDuration(value) {
+  const duration = Number(value);
+  if (!Number.isFinite(duration)) return "—";
+  return `${duration.toFixed(duration < 10 ? 1 : 0)} ms`;
+}
+
+export function parseMetadata(value) {
+  const text = String(value ?? "").trim();
+  if (!text) return {};
+  let parsed;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    throw new Error("Metadata must be valid JSON.");
+  }
+  if (!parsed || Array.isArray(parsed) || typeof parsed !== "object") {
+    throw new Error("Metadata must be a JSON object.");
+  }
+  return parsed;
+}
+
+export function parseExternalEmbeddings(value, expectedCount, expectedDimension) {
+  const text = String(value ?? "").trim();
+  let parsed;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    throw new Error("External embeddings must be valid JSON.");
+  }
+  if (!Array.isArray(parsed)) {
+    throw new Error("External embeddings must be a JSON array of vectors.");
+  }
+  if (parsed.length !== Number(expectedCount)) {
+    throw new Error(`Provide exactly one external embedding for each image (${parsed.length} != ${expectedCount}).`);
+  }
+  return parsed.map((vector, index) => {
+    if (!Array.isArray(vector) || vector.length === 0) {
+      throw new Error(`External embedding ${index + 1} must be a non-empty array.`);
+    }
+    if (expectedDimension && vector.length !== Number(expectedDimension)) {
+      throw new Error(`External embedding ${index + 1} must contain ${expectedDimension} values.`);
+    }
+    const values = vector.map((value) => {
+      if (typeof value !== "number" || !Number.isFinite(value)) {
+        throw new Error(`External embedding ${index + 1} contains a non-finite or non-numeric value.`);
+      }
+      return value;
+    });
+    const norm = Math.sqrt(values.reduce((sum, value) => sum + value * value, 0));
+    if (!Number.isFinite(norm) || Math.abs(norm - 1) > 2e-4) {
+      throw new Error(`External embedding ${index + 1} must be L2-normalized within 0.0002 of unit norm.`);
+    }
+    return values;
+  });
+}
+
+export function listItems(payload, preferredKeys = []) {
+  if (Array.isArray(payload)) return payload;
+  if (!payload || typeof payload !== "object") return [];
+  for (const key of [...preferredKeys, "items", "results"]) {
+    if (Array.isArray(payload[key])) return payload[key];
+  }
+  return [];
+}
+
+export function facePixels(face, imageWidth, imageHeight) {
+  const bbox = face?.bbox ?? face?.bounding_box ?? {};
+  const pixels = bbox.pixels ?? bbox.pixel ?? {};
+  if ([pixels.x, pixels.y, pixels.width, pixels.height].every(Number.isFinite)) {
+    return {
+      x: Number(pixels.x),
+      y: Number(pixels.y),
+      width: Number(pixels.width),
+      height: Number(pixels.height),
+    };
+  }
+
+  const normalized = bbox.normalized ?? bbox;
+  const left = Number(normalized.left ?? normalized.x);
+  const top = Number(normalized.top ?? normalized.y);
+  const width = Number(normalized.width);
+  const height = Number(normalized.height);
+  if (![left, top, width, height].every(Number.isFinite)) return null;
+  return {
+    x: left * imageWidth,
+    y: top * imageHeight,
+    width: width * imageWidth,
+    height: height * imageHeight,
+  };
+}
+
+export function faceLandmarks(face) {
+  const points = face?.landmarks;
+  if (!Array.isArray(points)) return [];
+  return points
+    .map((point) => {
+      if (Array.isArray(point) && point.length >= 2) {
+        return { x: Number(point[0]), y: Number(point[1]) };
+      }
+      if (point && typeof point === "object") {
+        return { x: Number(point.x), y: Number(point.y) };
+      }
+      return null;
+    })
+    .filter((point) => point && Number.isFinite(point.x) && Number.isFinite(point.y));
+}
+
+export function bestPersonMatches(samples, threshold, limit) {
+  const floor = normalizeThreshold(threshold, 0);
+  const count = Math.max(1, Math.floor(Number(limit) || 5));
+  const byPerson = new Map();
+  for (const sample of samples ?? []) {
+    const personId = String(sample?.person?.id ?? sample?.person_id ?? "");
+    const similarity = Number(sample?.similarity);
+    if (!personId || !Number.isFinite(similarity) || similarity < floor) continue;
+    const previous = byPerson.get(personId);
+    if (!previous || similarity > Number(previous.similarity)) byPerson.set(personId, sample);
+  }
+  return [...byPerson.values()]
+    .sort((left, right) => Number(right.similarity) - Number(left.similarity))
+    .slice(0, count);
+}
+
+export function filterPeople(people, query) {
+  const needle = String(query ?? "").trim().toLocaleLowerCase();
+  if (!needle) return [...(people ?? [])];
+  return (people ?? []).filter((person) =>
+    [person.id, person.name, person.external_id]
+      .filter(Boolean)
+      .some((value) => String(value).toLocaleLowerCase().includes(needle)),
+  );
+}
diff --git a/server/frontend/i18n.mjs b/server/frontend/i18n.mjs
new file mode 100644
index 000000000..c5b43ac9b
--- /dev/null
+++ b/server/frontend/i18n.mjs
@@ -0,0 +1,566 @@
+import { SHARED_UI_MESSAGES, UI_CATALOGS } from "./ui-translations.mjs?v=0.2.0-r13";
+
+export const LANGUAGES = Object.freeze([
+  { code: "en", label: "English", htmlLang: "en" },
+  { code: "zh", label: "中文", htmlLang: "zh-CN" },
+  { code: "ja", label: "日本語", htmlLang: "ja" },
+  { code: "de", label: "Deutsch", htmlLang: "de" },
+  { code: "es", label: "Español", htmlLang: "es" },
+  { code: "fr", label: "Français", htmlLang: "fr" },
+  { code: "ru", label: "Русский", htmlLang: "ru" },
+  { code: "pt", label: "Português", htmlLang: "pt" },
+  { code: "ko", label: "한국어", htmlLang: "ko" },
+]);
+
+const catalogs = {
+  zh: {
+    "Skip to content": "跳到主要内容", "Primary navigation": "主导航", "InsightFace Server dashboard": "InsightFace Server 仪表盘",
+    "Server console": "服务器控制台", Connecting: "正在连接", Dashboard: "仪表盘", Collections: "人员库", People: "人员",
+    Detect: "检测", Compare: "比对", Search: "搜索", "Live camera": "实时摄像头", System: "系统", Help: "帮助",
+    "API docs": "API 文档", "Configure API key": "配置 API Key", "Toggle navigation": "切换导航", "No API key": "未设置 API Key",
+    "Refresh current page": "刷新当前页面", Refresh: "刷新", "LOCAL FACE INTELLIGENCE": "本地人脸智能",
+    "Your face recognition service,": "您的人脸识别服务，", "at a glance.": "一目了然。", "Execution provider": "执行 Provider",
+    "Database totals": "数据库统计", "Face samples": "人脸样本", Model: "模型", SERVICE: "服务", "Runtime status": "运行状态",
+    "View diagnostics →": "查看诊断 →", "THIS TAB": "当前标签页", "Recent errors": "最近错误", Clear: "清除",
+    "No recent errors": "没有最近错误", "IDENTITY SPACES": "身份空间", "＋ New collection": "＋ 新建人员库",
+    CREATE: "创建", "New collection": "新建人员库", "Collection ID": "人员库 ID", Name: "名称", Description: "描述",
+    "Default cosine threshold": "默认余弦阈值", "Search profile": "搜索配置", "Capacity (faces)": "容量（人脸）",
+    "Maximum faces per person": "每人最多人脸数", "Load policy": "加载策略", "Server default": "服务器默认", Lazy: "延迟加载", Eager: "启动加载",
+    "Save 112×112 face crops in SQLite": "在 SQLite 中保存 112×112 人脸裁剪图", "Metadata (JSON)": "元数据（JSON）",
+    "Create collection": "创建人员库", "Filter collections": "筛选人员库", Collection: "人员库", Threshold: "阈值", Profile: "配置",
+    Crops: "裁剪图", Faces: "人脸", Actions: "操作", "No collections yet": "还没有人员库", ENROLLMENT: "人员注册",
+    "People & face samples": "人员与人脸样本", "＋ Register person": "＋ 注册人员", "Select a collection": "选择人员库",
+    "Search people": "搜索人员", REGISTER: "注册", "New person": "新人员", "Person ID": "人员 ID", optional: "可选",
+    "External ID": "外部 ID", "Enrollment review": "入库审查", "Off · largest face": "关闭 · 最大人脸", "Standard quality": "标准质量",
+    "Strict identity": "严格身份", "Embedding source": "特征来源", "Extract from image": "从图片抽取", "Trusted external feature": "可信外部特征",
+    "Drop registration photos here": "将注册照片拖到这里", "Register person": "注册人员", DIRECTORY: "人员目录", "No person selected": "未选择人员",
+    STATELESS: "无状态", "Detect faces": "检测人脸", "Drop an image": "拖入图片", "Maximum faces": "最大人脸数", "Minimum score": "最低分数",
+    "Image preview": "图片预览", "ONE-TO-ONE": "一对一", "Compare two faces": "比对两张人脸", "Source image": "源图片", "Target image": "目标图片",
+    "Choose source": "选择源图片", "Choose target": "选择目标图片", "Match threshold": "匹配阈值", "Compare faces": "比对人脸",
+    "Waiting for images": "等待图片", "ONE-TO-MANY": "一对多", "Search a collection": "搜索人员库", "Drop a query image": "拖入查询图片",
+    "Top results": "返回数量", Default: "默认", "Query preview": "查询预览", "BROWSER DEMO": "浏览器演示", "Camera is off": "摄像头已关闭",
+    Mode: "模式", "Detect only": "仅检测", "Frame rate": "帧率", "Start camera": "启动摄像头", "Stop camera": "停止摄像头",
+    ADMINISTRATION: "管理", "Refresh system": "刷新系统", Server: "服务器", Provider: "Provider", Accelerator: "加速器",
+    "API key in UI": "UI 中的 API Key", "Not configured": "未配置", RUNTIME: "运行时", "Hardware & software": "硬件与软件",
+    STORAGE: "存储", "Database & mounts": "数据库与挂载", MODELS: "模型", "Loaded models": "已加载模型", AUTHENTICATION: "认证",
+    "API key": "API Key", "Paste API key": "粘贴 API Key", "Clear key": "清除 Key", "Use for this tab": "在此标签页使用",
+    "Edit collection": "编辑人员库", Cancel: "取消", "Save changes": "保存修改", README: "README", "User guide": "用户指南",
+    "Open document": "打开文档", "Documentation": "文档", "Could not load documentation.": "无法加载文档。",
+    "The key stays in this tab's memory and is never saved by the UI.": "Key 仅保留在当前标签页内存中，UI 不会保存它。",
+    "Similarity is not a probability.": "相似度不是概率。", "No files selected": "未选择文件", "JPEG, PNG, or WebP": "JPEG、PNG 或 WebP",
+    "Detector input sizes": "检测输入尺寸", "Detection threshold": "检测阈值", "NMS threshold": "NMS 阈值", "Single-face selection": "单脸挑选策略",
+    "Largest face": "最大脸", "Center-weighted largest": "中心加权最大脸", "Detection profile": "检测配置", "Off · collection selection": "关闭 · 使用人员库策略",
+  },
+  ja: {
+    "Skip to content": "メインコンテンツへ", "Primary navigation": "メインナビゲーション", "Server console": "サーバーコンソール", Connecting: "接続中",
+    Dashboard: "ダッシュボード", Collections: "コレクション", People: "人物", Detect: "検出", Compare: "比較", Search: "検索",
+    "Live camera": "ライブカメラ", System: "システム", Help: "ヘルプ", "API docs": "API ドキュメント", "Configure API key": "API キーを設定",
+    "No API key": "API キーなし", "Refresh current page": "現在のページを更新", Refresh: "更新", "Execution provider": "実行プロバイダー",
+    "Face samples": "顔サンプル", Model: "モデル", "Runtime status": "実行状態", "Recent errors": "最近のエラー", Clear: "クリア",
+    "No recent errors": "最近のエラーはありません", "＋ New collection": "＋ 新規コレクション", "New collection": "新規コレクション",
+    "Collection ID": "コレクション ID", Name: "名前", Description: "説明", "Default cosine threshold": "既定のコサインしきい値",
+    "Search profile": "検索プロファイル", "Capacity (faces)": "容量（顔）", "Maximum faces per person": "人物ごとの最大顔数",
+    "Load policy": "ロードポリシー", "Server default": "サーバー既定", Lazy: "遅延", Eager: "起動時", "Metadata (JSON)": "メタデータ（JSON）",
+    "Create collection": "コレクションを作成", Collection: "コレクション", Threshold: "しきい値", Profile: "プロファイル", Faces: "顔",
+    ENROLLMENT: "登録", "People & face samples": "人物と顔サンプル", "＋ Register person": "＋ 人物を登録", "Select a collection": "コレクションを選択",
+    "Search people": "人物を検索", "New person": "新規人物", "Person ID": "人物 ID", optional: "任意", "External ID": "外部 ID",
+    "Enrollment review": "登録レビュー", "Off · largest face": "オフ · 最大の顔", "Standard quality": "標準品質", "Strict identity": "厳格な本人確認",
+    "Embedding source": "特徴量ソース", "Extract from image": "画像から抽出", "Trusted external feature": "信頼済み外部特徴量",
+    "Drop registration photos here": "登録写真をここにドロップ", "Register person": "人物を登録", "No person selected": "人物が選択されていません",
+    "Detect faces": "顔を検出", "Drop an image": "画像をドロップ", "Maximum faces": "最大顔数", "Minimum score": "最小スコア",
+    "Compare two faces": "2つの顔を比較", "Source image": "ソース画像", "Target image": "ターゲット画像", "Choose source": "ソースを選択",
+    "Choose target": "ターゲットを選択", "Match threshold": "一致しきい値", "Compare faces": "顔を比較", "Waiting for images": "画像待機中",
+    "Search a collection": "コレクションを検索", "Drop a query image": "検索画像をドロップ", "Top results": "上位結果", Default: "既定",
+    "Camera is off": "カメラはオフです", Mode: "モード", "Detect only": "検出のみ", "Frame rate": "フレームレート",
+    "Start camera": "カメラを開始", "Stop camera": "カメラを停止", ADMINISTRATION: "管理", "Refresh system": "システムを更新",
+    Server: "サーバー", Provider: "プロバイダー", Accelerator: "アクセラレーター", "Not configured": "未設定", RUNTIME: "ランタイム",
+    "Hardware & software": "ハードウェアとソフトウェア", STORAGE: "ストレージ", "Database & mounts": "データベースとマウント",
+    "Loaded models": "ロード済みモデル", AUTHENTICATION: "認証", "API key": "API キー", "Paste API key": "API キーを貼り付け",
+    "Clear key": "キーをクリア", "Use for this tab": "このタブで使用", "Edit collection": "コレクションを編集", Cancel: "キャンセル",
+    "Save changes": "変更を保存", README: "README", "User guide": "ユーザーガイド", Documentation: "ドキュメント", "Open document": "ドキュメントを開く",
+    "Could not load documentation.": "ドキュメントを読み込めませんでした。", "Similarity is not a probability.": "類似度は確率ではありません。",
+    "No files selected": "ファイル未選択", "JPEG, PNG, or WebP": "JPEG、PNG、または WebP",
+    "Detector input sizes": "検出入力サイズ", "Detection threshold": "検出しきい値", "NMS threshold": "NMS しきい値", "Single-face selection": "単一顔の選択",
+    "Largest face": "最大の顔", "Center-weighted largest": "中心重み付き最大", "Detection profile": "検出プロファイル", "Off · collection selection": "オフ · Collection の選択",
+  },
+  de: {
+    "Skip to content": "Zum Inhalt springen", "Primary navigation": "Hauptnavigation", "Server console": "Serverkonsole", Connecting: "Verbindung wird hergestellt",
+    Dashboard: "Übersicht", Collections: "Sammlungen", People: "Personen", Detect: "Erkennen", Compare: "Vergleichen", Search: "Suchen",
+    "Live camera": "Live-Kamera", System: "System", Help: "Hilfe", "API docs": "API-Dokumentation", "Configure API key": "API-Schlüssel konfigurieren",
+    "No API key": "Kein API-Schlüssel", "Refresh current page": "Aktuelle Seite neu laden", Refresh: "Aktualisieren", "Execution provider": "Ausführungs-Provider",
+    "Face samples": "Gesichtsproben", Model: "Modell", "Runtime status": "Laufzeitstatus", "Recent errors": "Letzte Fehler", Clear: "Leeren",
+    "No recent errors": "Keine aktuellen Fehler", "＋ New collection": "＋ Neue Sammlung", "New collection": "Neue Sammlung", "Collection ID": "Sammlungs-ID",
+    Name: "Name", Description: "Beschreibung", "Default cosine threshold": "Standard-Cosinus-Schwelle", "Search profile": "Suchprofil",
+    "Capacity (faces)": "Kapazität (Gesichter)", "Maximum faces per person": "Maximale Gesichter pro Person", "Load policy": "Ladestrategie",
+    "Server default": "Serverstandard", Lazy: "Verzögert", Eager: "Beim Start", "Metadata (JSON)": "Metadaten (JSON)",
+    "Create collection": "Sammlung erstellen", Collection: "Sammlung", Threshold: "Schwelle", Profile: "Profil", Faces: "Gesichter",
+    "People & face samples": "Personen und Gesichtsproben", "＋ Register person": "＋ Person registrieren", "Select a collection": "Sammlung auswählen",
+    "Search people": "Personen suchen", "New person": "Neue Person", "Person ID": "Personen-ID", optional: "optional", "External ID": "Externe ID",
+    "Enrollment review": "Registrierungsprüfung", "Off · largest face": "Aus · größtes Gesicht", "Standard quality": "Standardqualität", "Strict identity": "Strikte Identität",
+    "Embedding source": "Embedding-Quelle", "Extract from image": "Aus Bild extrahieren", "Trusted external feature": "Vertrauenswürdiges externes Merkmal",
+    "Drop registration photos here": "Registrierungsfotos hier ablegen", "Register person": "Person registrieren", "No person selected": "Keine Person ausgewählt",
+    "Detect faces": "Gesichter erkennen", "Drop an image": "Bild hier ablegen", "Maximum faces": "Maximale Gesichter", "Minimum score": "Mindestwert",
+    "Compare two faces": "Zwei Gesichter vergleichen", "Source image": "Quellbild", "Target image": "Zielbild", "Choose source": "Quelle wählen",
+    "Choose target": "Ziel wählen", "Match threshold": "Übereinstimmungsschwelle", "Compare faces": "Gesichter vergleichen", "Waiting for images": "Warte auf Bilder",
+    "Search a collection": "Sammlung durchsuchen", "Drop a query image": "Suchbild hier ablegen", "Top results": "Top-Ergebnisse", Default: "Standard",
+    "Camera is off": "Kamera ist aus", Mode: "Modus", "Detect only": "Nur erkennen", "Frame rate": "Bildrate", "Start camera": "Kamera starten",
+    "Stop camera": "Kamera stoppen", ADMINISTRATION: "VERWALTUNG", "Refresh system": "System aktualisieren", Server: "Server", Provider: "Provider",
+    Accelerator: "Beschleuniger", "Not configured": "Nicht konfiguriert", RUNTIME: "LAUFZEIT", "Hardware & software": "Hardware und Software",
+    STORAGE: "SPEICHER", "Database & mounts": "Datenbank und Einbindungen", "Loaded models": "Geladene Modelle", AUTHENTICATION: "AUTHENTIFIZIERUNG",
+    "API key": "API-Schlüssel", "Paste API key": "API-Schlüssel einfügen", "Clear key": "Schlüssel löschen", "Use for this tab": "Für diesen Tab verwenden",
+    "Edit collection": "Sammlung bearbeiten", Cancel: "Abbrechen", "Save changes": "Änderungen speichern", README: "README", "User guide": "Benutzerhandbuch",
+    Documentation: "Dokumentation", "Open document": "Dokument öffnen", "Could not load documentation.": "Dokumentation konnte nicht geladen werden.",
+    "Similarity is not a probability.": "Ähnlichkeit ist keine Wahrscheinlichkeit.", "No files selected": "Keine Dateien ausgewählt", "JPEG, PNG, or WebP": "JPEG, PNG oder WebP",
+    "Detector input sizes": "Detektor-Eingabegrößen", "Detection threshold": "Erkennungsschwelle", "NMS threshold": "NMS-Schwelle", "Single-face selection": "Ein-Gesicht-Auswahl",
+    "Largest face": "Größtes Gesicht", "Center-weighted largest": "Zentrumsgewichtet größtes", "Detection profile": "Erkennungsprofil", "Off · collection selection": "Aus · Collection-Auswahl",
+  },
+  es: {
+    "Skip to content": "Ir al contenido", "Primary navigation": "Navegación principal", "Server console": "Consola del servidor", Connecting: "Conectando",
+    Dashboard: "Panel", Collections: "Colecciones", People: "Personas", Detect: "Detectar", Compare: "Comparar", Search: "Buscar",
+    "Live camera": "Cámara en vivo", System: "Sistema", Help: "Ayuda", "API docs": "Documentación API", "Configure API key": "Configurar clave API",
+    "No API key": "Sin clave API", "Refresh current page": "Actualizar página actual", Refresh: "Actualizar", "Execution provider": "Proveedor de ejecución",
+    "Face samples": "Muestras faciales", Model: "Modelo", "Runtime status": "Estado de ejecución", "Recent errors": "Errores recientes", Clear: "Limpiar",
+    "No recent errors": "No hay errores recientes", "＋ New collection": "＋ Nueva colección", "New collection": "Nueva colección", "Collection ID": "ID de colección",
+    Name: "Nombre", Description: "Descripción", "Default cosine threshold": "Umbral coseno predeterminado", "Search profile": "Perfil de búsqueda",
+    "Capacity (faces)": "Capacidad (rostros)", "Maximum faces per person": "Máximo de rostros por persona", "Load policy": "Política de carga",
+    "Server default": "Predeterminado del servidor", Lazy: "Diferida", Eager: "Al iniciar", "Metadata (JSON)": "Metadatos (JSON)",
+    "Create collection": "Crear colección", Collection: "Colección", Threshold: "Umbral", Profile: "Perfil", Faces: "Rostros",
+    "People & face samples": "Personas y muestras faciales", "＋ Register person": "＋ Registrar persona", "Select a collection": "Seleccionar colección",
+    "Search people": "Buscar personas", "New person": "Nueva persona", "Person ID": "ID de persona", optional: "opcional", "External ID": "ID externo",
+    "Enrollment review": "Revisión de registro", "Off · largest face": "Desactivada · rostro mayor", "Standard quality": "Calidad estándar", "Strict identity": "Identidad estricta",
+    "Embedding source": "Origen del embedding", "Extract from image": "Extraer de la imagen", "Trusted external feature": "Vector externo confiable",
+    "Drop registration photos here": "Suelta aquí las fotos de registro", "Register person": "Registrar persona", "No person selected": "Ninguna persona seleccionada",
+    "Detect faces": "Detectar rostros", "Drop an image": "Suelta una imagen", "Maximum faces": "Máximo de rostros", "Minimum score": "Puntuación mínima",
+    "Compare two faces": "Comparar dos rostros", "Source image": "Imagen de origen", "Target image": "Imagen de destino", "Choose source": "Elegir origen",
+    "Choose target": "Elegir destino", "Match threshold": "Umbral de coincidencia", "Compare faces": "Comparar rostros", "Waiting for images": "Esperando imágenes",
+    "Search a collection": "Buscar en una colección", "Drop a query image": "Suelta una imagen de consulta", "Top results": "Mejores resultados", Default: "Predeterminado",
+    "Camera is off": "La cámara está apagada", Mode: "Modo", "Detect only": "Solo detectar", "Frame rate": "Fotogramas por segundo",
+    "Start camera": "Iniciar cámara", "Stop camera": "Detener cámara", ADMINISTRATION: "ADMINISTRACIÓN", "Refresh system": "Actualizar sistema",
+    Server: "Servidor", Provider: "Proveedor", Accelerator: "Acelerador", "Not configured": "Sin configurar", RUNTIME: "EJECUCIÓN",
+    "Hardware & software": "Hardware y software", STORAGE: "ALMACENAMIENTO", "Database & mounts": "Base de datos y montajes", "Loaded models": "Modelos cargados",
+    AUTHENTICATION: "AUTENTICACIÓN", "API key": "Clave API", "Paste API key": "Pegar clave API", "Clear key": "Borrar clave",
+    "Use for this tab": "Usar en esta pestaña", "Edit collection": "Editar colección", Cancel: "Cancelar", "Save changes": "Guardar cambios",
+    README: "README", "User guide": "Guía de usuario", Documentation: "Documentación", "Open document": "Abrir documento",
+    "Could not load documentation.": "No se pudo cargar la documentación.", "Similarity is not a probability.": "La similitud no es una probabilidad.",
+    "No files selected": "No hay archivos seleccionados", "JPEG, PNG, or WebP": "JPEG, PNG o WebP",
+    "Detector input sizes": "Tamaños de entrada del detector", "Detection threshold": "Umbral de detección", "NMS threshold": "Umbral NMS", "Single-face selection": "Selección de un rostro",
+    "Largest face": "Rostro mayor", "Center-weighted largest": "Mayor ponderado por centro", "Detection profile": "Perfil de detección", "Off · collection selection": "Desactivada · selección de Collection",
+  },
+  fr: {
+    "Skip to content": "Aller au contenu", "Primary navigation": "Navigation principale", "Server console": "Console serveur", Connecting: "Connexion",
+    Dashboard: "Tableau de bord", Collections: "Collections", People: "Personnes", Detect: "Détecter", Compare: "Comparer", Search: "Rechercher",
+    "Live camera": "Caméra en direct", System: "Système", Help: "Aide", "API docs": "Documentation API", "Configure API key": "Configurer la clé API",
+    "No API key": "Aucune clé API", "Refresh current page": "Actualiser la page", Refresh: "Actualiser", "Execution provider": "Provider d’exécution",
+    "Face samples": "Échantillons faciaux", Model: "Modèle", "Runtime status": "État d’exécution", "Recent errors": "Erreurs récentes", Clear: "Effacer",
+    "No recent errors": "Aucune erreur récente", "＋ New collection": "＋ Nouvelle collection", "New collection": "Nouvelle collection", "Collection ID": "ID de collection",
+    Name: "Nom", Description: "Description", "Default cosine threshold": "Seuil cosinus par défaut", "Search profile": "Profil de recherche",
+    "Capacity (faces)": "Capacité (visages)", "Maximum faces per person": "Nombre maximal de visages par personne", "Load policy": "Politique de chargement",
+    "Server default": "Valeur serveur", Lazy: "Différé", Eager: "Au démarrage", "Metadata (JSON)": "Métadonnées (JSON)",
+    "Create collection": "Créer la collection", Collection: "Collection", Threshold: "Seuil", Profile: "Profil", Faces: "Visages",
+    "People & face samples": "Personnes et échantillons faciaux", "＋ Register person": "＋ Inscrire une personne", "Select a collection": "Choisir une collection",
+    "Search people": "Rechercher des personnes", "New person": "Nouvelle personne", "Person ID": "ID de personne", optional: "facultatif", "External ID": "ID externe",
+    "Enrollment review": "Contrôle d’inscription", "Off · largest face": "Désactivé · plus grand visage", "Standard quality": "Qualité standard", "Strict identity": "Identité stricte",
+    "Embedding source": "Source de l’embedding", "Extract from image": "Extraire de l’image", "Trusted external feature": "Vecteur externe approuvé",
+    "Drop registration photos here": "Déposez les photos d’inscription ici", "Register person": "Inscrire la personne", "No person selected": "Aucune personne sélectionnée",
+    "Detect faces": "Détecter les visages", "Drop an image": "Déposez une image", "Maximum faces": "Nombre maximal de visages", "Minimum score": "Score minimal",
+    "Compare two faces": "Comparer deux visages", "Source image": "Image source", "Target image": "Image cible", "Choose source": "Choisir la source",
+    "Choose target": "Choisir la cible", "Match threshold": "Seuil de correspondance", "Compare faces": "Comparer les visages", "Waiting for images": "En attente d’images",
+    "Search a collection": "Rechercher dans une collection", "Drop a query image": "Déposez une image de recherche", "Top results": "Meilleurs résultats", Default: "Par défaut",
+    "Camera is off": "La caméra est arrêtée", Mode: "Mode", "Detect only": "Détection uniquement", "Frame rate": "Fréquence d’images",
+    "Start camera": "Démarrer la caméra", "Stop camera": "Arrêter la caméra", ADMINISTRATION: "ADMINISTRATION", "Refresh system": "Actualiser le système",
+    Server: "Serveur", Provider: "Provider", Accelerator: "Accélérateur", "Not configured": "Non configuré", RUNTIME: "EXÉCUTION",
+    "Hardware & software": "Matériel et logiciel", STORAGE: "STOCKAGE", "Database & mounts": "Base de données et montages", "Loaded models": "Modèles chargés",
+    AUTHENTICATION: "AUTHENTIFICATION", "API key": "Clé API", "Paste API key": "Coller la clé API", "Clear key": "Effacer la clé",
+    "Use for this tab": "Utiliser dans cet onglet", "Edit collection": "Modifier la collection", Cancel: "Annuler", "Save changes": "Enregistrer",
+    README: "README", "User guide": "Guide utilisateur", Documentation: "Documentation", "Open document": "Ouvrir le document",
+    "Could not load documentation.": "Impossible de charger la documentation.", "Similarity is not a probability.": "La similarité n’est pas une probabilité.",
+    "No files selected": "Aucun fichier sélectionné", "JPEG, PNG, or WebP": "JPEG, PNG ou WebP",
+    "Detector input sizes": "Tailles d’entrée du détecteur", "Detection threshold": "Seuil de détection", "NMS threshold": "Seuil NMS", "Single-face selection": "Sélection mono-visage",
+    "Largest face": "Plus grand visage", "Center-weighted largest": "Plus grand pondéré par le centre", "Detection profile": "Profil de détection", "Off · collection selection": "Désactivé · sélection Collection",
+  },
+  ru: {
+    "Skip to content": "Перейти к содержимому", "Primary navigation": "Основная навигация", "Server console": "Консоль сервера", Connecting: "Подключение",
+    Dashboard: "Панель", Collections: "Коллекции", People: "Люди", Detect: "Детекция", Compare: "Сравнение", Search: "Поиск",
+    "Live camera": "Камера", System: "Система", Help: "Помощь", "API docs": "Документация API", "Configure API key": "Настроить API-ключ",
+    "No API key": "Нет API-ключа", "Refresh current page": "Обновить страницу", Refresh: "Обновить", "Execution provider": "Provider выполнения",
+    "Face samples": "Образцы лиц", Model: "Модель", "Runtime status": "Состояние", "Recent errors": "Последние ошибки", Clear: "Очистить",
+    "No recent errors": "Недавних ошибок нет", "＋ New collection": "＋ Новая коллекция", "New collection": "Новая коллекция", "Collection ID": "ID коллекции",
+    Name: "Имя", Description: "Описание", "Default cosine threshold": "Порог cosine по умолчанию", "Search profile": "Профиль поиска",
+    "Capacity (faces)": "Ёмкость (лица)", "Maximum faces per person": "Максимум лиц на человека", "Load policy": "Политика загрузки",
+    "Server default": "Настройка сервера", Lazy: "Ленивая", Eager: "При запуске", "Metadata (JSON)": "Метаданные (JSON)",
+    "Create collection": "Создать коллекцию", Collection: "Коллекция", Threshold: "Порог", Profile: "Профиль", Faces: "Лица",
+    "People & face samples": "Люди и образцы лиц", "＋ Register person": "＋ Зарегистрировать человека", "Select a collection": "Выберите коллекцию",
+    "Search people": "Поиск людей", "New person": "Новый человек", "Person ID": "ID человека", optional: "необязательно", "External ID": "Внешний ID",
+    "Enrollment review": "Проверка регистрации", "Off · largest face": "Выкл. · крупнейшее лицо", "Standard quality": "Стандартное качество", "Strict identity": "Строгая идентичность",
+    "Embedding source": "Источник embedding", "Extract from image": "Извлечь из изображения", "Trusted external feature": "Доверенный внешний вектор",
+    "Drop registration photos here": "Перетащите сюда фотографии", "Register person": "Зарегистрировать", "No person selected": "Человек не выбран",
+    "Detect faces": "Обнаружить лица", "Drop an image": "Перетащите изображение", "Maximum faces": "Максимум лиц", "Minimum score": "Минимальный score",
+    "Compare two faces": "Сравнить два лица", "Source image": "Исходное изображение", "Target image": "Целевое изображение", "Choose source": "Выбрать исходное",
+    "Choose target": "Выбрать целевое", "Match threshold": "Порог совпадения", "Compare faces": "Сравнить лица", "Waiting for images": "Ожидание изображений",
+    "Search a collection": "Поиск по коллекции", "Drop a query image": "Перетащите изображение запроса", "Top results": "Лучшие результаты", Default: "По умолчанию",
+    "Camera is off": "Камера выключена", Mode: "Режим", "Detect only": "Только детекция", "Frame rate": "Частота кадров",
+    "Start camera": "Включить камеру", "Stop camera": "Остановить камеру", ADMINISTRATION: "АДМИНИСТРИРОВАНИЕ", "Refresh system": "Обновить систему",
+    Server: "Сервер", Provider: "Provider", Accelerator: "Ускоритель", "Not configured": "Не настроено", RUNTIME: "СРЕДА",
+    "Hardware & software": "Аппаратное и программное обеспечение", STORAGE: "ХРАНИЛИЩЕ", "Database & mounts": "База данных и точки монтирования",
+    "Loaded models": "Загруженные модели", AUTHENTICATION: "АУТЕНТИФИКАЦИЯ", "API key": "API-ключ", "Paste API key": "Вставьте API-ключ",
+    "Clear key": "Очистить ключ", "Use for this tab": "Использовать в этой вкладке", "Edit collection": "Изменить коллекцию", Cancel: "Отмена",
+    "Save changes": "Сохранить", README: "README", "User guide": "Руководство", Documentation: "Документация", "Open document": "Открыть документ",
+    "Could not load documentation.": "Не удалось загрузить документацию.", "Similarity is not a probability.": "Сходство не является вероятностью.",
+    "No files selected": "Файлы не выбраны", "JPEG, PNG, or WebP": "JPEG, PNG или WebP",
+    "Detector input sizes": "Входные размеры детектора", "Detection threshold": "Порог детекции", "NMS threshold": "Порог NMS", "Single-face selection": "Выбор одного лица",
+    "Largest face": "Крупнейшее лицо", "Center-weighted largest": "Наибольшее с весом центра", "Detection profile": "Профиль детекции", "Off · collection selection": "Выкл. · выбор Collection",
+  },
+  pt: {
+    "Skip to content": "Ir para o conteúdo", "Primary navigation": "Navegação principal", "Server console": "Consola do servidor", Connecting: "A ligar",
+    Dashboard: "Painel", Collections: "Coleções", People: "Pessoas", Detect: "Detetar", Compare: "Comparar", Search: "Pesquisar",
+    "Live camera": "Câmara ao vivo", System: "Sistema", Help: "Ajuda", "API docs": "Documentação API", "Configure API key": "Configurar chave API",
+    "No API key": "Sem chave API", "Refresh current page": "Atualizar página", Refresh: "Atualizar", "Execution provider": "Provider de execução",
+    "Face samples": "Amostras faciais", Model: "Modelo", "Runtime status": "Estado de execução", "Recent errors": "Erros recentes", Clear: "Limpar",
+    "No recent errors": "Sem erros recentes", "＋ New collection": "＋ Nova coleção", "New collection": "Nova coleção", "Collection ID": "ID da coleção",
+    Name: "Nome", Description: "Descrição", "Default cosine threshold": "Limiar cosine predefinido", "Search profile": "Perfil de pesquisa",
+    "Capacity (faces)": "Capacidade (rostos)", "Maximum faces per person": "Máximo de rostos por pessoa", "Load policy": "Política de carregamento",
+    "Server default": "Predefinição do servidor", Lazy: "Adiado", Eager: "No arranque", "Metadata (JSON)": "Metadados (JSON)",
+    "Create collection": "Criar coleção", Collection: "Coleção", Threshold: "Limiar", Profile: "Perfil", Faces: "Rostos",
+    "People & face samples": "Pessoas e amostras faciais", "＋ Register person": "＋ Registar pessoa", "Select a collection": "Selecionar coleção",
+    "Search people": "Pesquisar pessoas", "New person": "Nova pessoa", "Person ID": "ID da pessoa", optional: "opcional", "External ID": "ID externo",
+    "Enrollment review": "Revisão de registo", "Off · largest face": "Desativada · maior rosto", "Standard quality": "Qualidade padrão", "Strict identity": "Identidade rigorosa",
+    "Embedding source": "Origem do embedding", "Extract from image": "Extrair da imagem", "Trusted external feature": "Vetor externo confiável",
+    "Drop registration photos here": "Largue aqui as fotos de registo", "Register person": "Registar pessoa", "No person selected": "Nenhuma pessoa selecionada",
+    "Detect faces": "Detetar rostos", "Drop an image": "Largue uma imagem", "Maximum faces": "Máximo de rostos", "Minimum score": "Pontuação mínima",
+    "Compare two faces": "Comparar dois rostos", "Source image": "Imagem de origem", "Target image": "Imagem de destino", "Choose source": "Escolher origem",
+    "Choose target": "Escolher destino", "Match threshold": "Limiar de correspondência", "Compare faces": "Comparar rostos", "Waiting for images": "À espera de imagens",
+    "Search a collection": "Pesquisar numa coleção", "Drop a query image": "Largue uma imagem de consulta", "Top results": "Melhores resultados", Default: "Predefinido",
+    "Camera is off": "A câmara está desligada", Mode: "Modo", "Detect only": "Apenas detetar", "Frame rate": "Taxa de fotogramas",
+    "Start camera": "Iniciar câmara", "Stop camera": "Parar câmara", ADMINISTRATION: "ADMINISTRAÇÃO", "Refresh system": "Atualizar sistema",
+    Server: "Servidor", Provider: "Provider", Accelerator: "Acelerador", "Not configured": "Não configurado", RUNTIME: "EXECUÇÃO",
+    "Hardware & software": "Hardware e software", STORAGE: "ARMAZENAMENTO", "Database & mounts": "Base de dados e montagens", "Loaded models": "Modelos carregados",
+    AUTHENTICATION: "AUTENTICAÇÃO", "API key": "Chave API", "Paste API key": "Colar chave API", "Clear key": "Limpar chave",
+    "Use for this tab": "Usar neste separador", "Edit collection": "Editar coleção", Cancel: "Cancelar", "Save changes": "Guardar alterações",
+    README: "README", "User guide": "Guia do utilizador", Documentation: "Documentação", "Open document": "Abrir documento",
+    "Could not load documentation.": "Não foi possível carregar a documentação.", "Similarity is not a probability.": "A similaridade não é uma probabilidade.",
+    "No files selected": "Nenhum ficheiro selecionado", "JPEG, PNG, or WebP": "JPEG, PNG ou WebP",
+    "Detector input sizes": "Tamanhos de entrada do detetor", "Detection threshold": "Limiar de deteção", "NMS threshold": "Limiar NMS", "Single-face selection": "Seleção de um rosto",
+    "Largest face": "Maior rosto", "Center-weighted largest": "Maior com peso central", "Detection profile": "Perfil de deteção", "Off · collection selection": "Desativado · seleção da Collection",
+  },
+  ko: {
+    "Skip to content": "본문으로 이동", "Primary navigation": "주요 탐색", "Server console": "서버 콘솔", Connecting: "연결 중",
+    Dashboard: "대시보드", Collections: "컬렉션", People: "사람", Detect: "검출", Compare: "비교", Search: "검색",
+    "Live camera": "라이브 카메라", System: "시스템", Help: "도움말", "API docs": "API 문서", "Configure API key": "API 키 설정",
+    "No API key": "API 키 없음", "Refresh current page": "현재 페이지 새로고침", Refresh: "새로고침", "Execution provider": "실행 Provider",
+    "Face samples": "얼굴 샘플", Model: "모델", "Runtime status": "실행 상태", "Recent errors": "최근 오류", Clear: "지우기",
+    "No recent errors": "최근 오류 없음", "＋ New collection": "＋ 새 컬렉션", "New collection": "새 컬렉션", "Collection ID": "컬렉션 ID",
+    Name: "이름", Description: "설명", "Default cosine threshold": "기본 cosine 임계값", "Search profile": "검색 프로필",
+    "Capacity (faces)": "용량(얼굴)", "Maximum faces per person": "사람별 최대 얼굴 수", "Load policy": "로드 정책",
+    "Server default": "서버 기본값", Lazy: "지연", Eager: "시작 시", "Metadata (JSON)": "메타데이터(JSON)",
+    "Create collection": "컬렉션 만들기", Collection: "컬렉션", Threshold: "임계값", Profile: "프로필", Faces: "얼굴",
+    "People & face samples": "사람 및 얼굴 샘플", "＋ Register person": "＋ 사람 등록", "Select a collection": "컬렉션 선택",
+    "Search people": "사람 검색", "New person": "새 사람", "Person ID": "사람 ID", optional: "선택", "External ID": "외부 ID",
+    "Enrollment review": "등록 검토", "Off · largest face": "끄기 · 가장 큰 얼굴", "Standard quality": "표준 품질", "Strict identity": "엄격한 신원",
+    "Embedding source": "임베딩 소스", "Extract from image": "이미지에서 추출", "Trusted external feature": "신뢰된 외부 특징",
+    "Drop registration photos here": "등록 사진을 여기에 놓으세요", "Register person": "사람 등록", "No person selected": "선택된 사람이 없습니다",
+    "Detect faces": "얼굴 검출", "Drop an image": "이미지를 놓으세요", "Maximum faces": "최대 얼굴 수", "Minimum score": "최소 점수",
+    "Compare two faces": "두 얼굴 비교", "Source image": "원본 이미지", "Target image": "대상 이미지", "Choose source": "원본 선택",
+    "Choose target": "대상 선택", "Match threshold": "일치 임계값", "Compare faces": "얼굴 비교", "Waiting for images": "이미지 대기 중",
+    "Search a collection": "컬렉션 검색", "Drop a query image": "검색 이미지를 놓으세요", "Top results": "상위 결과", Default: "기본값",
+    "Camera is off": "카메라가 꺼져 있습니다", Mode: "모드", "Detect only": "검출만", "Frame rate": "프레임 속도",
+    "Start camera": "카메라 시작", "Stop camera": "카메라 중지", ADMINISTRATION: "관리", "Refresh system": "시스템 새로고침",
+    Server: "서버", Provider: "Provider", Accelerator: "가속기", "Not configured": "설정되지 않음", RUNTIME: "런타임",
+    "Hardware & software": "하드웨어 및 소프트웨어", STORAGE: "스토리지", "Database & mounts": "데이터베이스 및 마운트", "Loaded models": "로드된 모델",
+    AUTHENTICATION: "인증", "API key": "API 키", "Paste API key": "API 키 붙여넣기", "Clear key": "키 지우기",
+    "Use for this tab": "이 탭에서 사용", "Edit collection": "컬렉션 편집", Cancel: "취소", "Save changes": "변경 저장",
+    README: "README", "User guide": "사용자 가이드", Documentation: "문서", "Open document": "문서 열기",
+    "Could not load documentation.": "문서를 불러올 수 없습니다.", "Similarity is not a probability.": "유사도는 확률이 아닙니다.",
+    "No files selected": "선택된 파일 없음", "JPEG, PNG, or WebP": "JPEG, PNG 또는 WebP",
+    "Detector input sizes": "검출기 입력 크기", "Detection threshold": "검출 임계값", "NMS threshold": "NMS 임계값", "Single-face selection": "단일 얼굴 선택",
+    "Largest face": "가장 큰 얼굴", "Center-weighted largest": "중심 가중 최대 얼굴", "Detection profile": "검출 프로필", "Off · collection selection": "끄기 · Collection 선택",
+  },
+};
+
+const supplementalCatalogs = {
+  zh: {
+    Language: "语言", "Version-matched documentation is bundled with this server and works without internet access.": "与当前版本匹配的文档已内置，无需联网即可阅读。",
+    "InsightFace Server dashboard": "InsightFace Server 仪表盘", "Detect, compare, enroll, and search without sending images outside your server.": "无需将图片发送到服务器之外，即可完成检测、比对、注册和搜索。",
+    "separate identity spaces": "独立身份空间", "enrolled identities": "已注册身份", "stored embeddings": "已存特征", "API readiness": "API 就绪状态", Database: "数据库",
+    "API failures from this browser tab will appear here.": "此浏览器标签页中的 API 错误会显示在这里。", "tab memory only": "仅保留在标签页内存",
+    "The displayed language follows the console language selector.": "文档语言跟随控制台语言选择。", "Loading documentation…": "正在加载文档…",
+    "API key configured": "API Key 已配置", Configured: "已配置", "Ready to compare": "可以开始比对", Close: "关闭",
+    "Open console": "打开控制台", "OpenAPI JSON": "OpenAPI JSON", "API Reference": "API 参考", "Simple face recognition APIs.": "简单的人脸识别 API。",
+    "Base path": "基础路径", "API operations": "API 操作", "Filter operations": "筛选操作", "Loading API schema": "正在加载 API Schema",
+    COMPONENTS: "组件", "Data schemas": "数据 Schema", Parameters: "参数", required: "必填", "Request body": "请求体", Responses: "响应",
+    "Open-source model license": "开源模型许可", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "InsightFace 提供的开源预训练模型（包括 buffalo_l）仅限非商业研究使用。", "Commercial use requires a separate license.": "商业使用需要单独许可。", "Commercial licensing": "商业许可信息",
+  },
+  ja: {
+    Language: "言語", "Version-matched documentation is bundled with this server and works without internet access.": "このサーバーと同じバージョンのドキュメントが内蔵され、オフラインで利用できます。",
+    "InsightFace Server dashboard": "InsightFace Server ダッシュボード", "The key stays in this tab's memory and is never saved by the UI.": "Key はこのタブのメモリだけに保持され、UI には保存されません。",
+    "Detect, compare, enroll, and search without sending images outside your server.": "画像をサーバー外へ送信せずに、検出、比較、登録、検索を実行できます。", "separate identity spaces": "独立した ID 空間", "enrolled identities": "登録済み ID",
+    "stored embeddings": "保存済み特徴量", "API readiness": "API 準備状態", Database: "データベース", "API failures from this browser tab will appear here.": "このタブの API エラーがここに表示されます。", "tab memory only": "タブのメモリのみ",
+    "The displayed language follows the console language selector.": "文書の言語はコンソールの言語設定に従います。", "Loading documentation…": "ドキュメントを読み込み中…",
+    "API key configured": "API キー設定済み", Configured: "設定済み", "Ready to compare": "比較できます", Close: "閉じる",
+    "Open console": "コンソールを開く", "API Reference": "API リファレンス", "Simple face recognition APIs.": "シンプルな顔認識 API。", "Base path": "ベースパス",
+    "API operations": "API 操作", "Filter operations": "操作を絞り込み", "Loading API schema": "API スキーマを読み込み中", COMPONENTS: "コンポーネント",
+    "Data schemas": "データスキーマ", Parameters: "パラメーター", required: "必須", "Request body": "リクエスト本文", Responses: "レスポンス",
+    "Open-source model license": "オープンソースモデルのライセンス", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "buffalo_l を含む InsightFace 提供のオープンソース事前学習モデルは、非商用の研究目的に限り利用できます。", "Commercial use requires a separate license.": "商用利用には別途ライセンスが必要です。", "Commercial licensing": "商用ライセンス",
+  },
+  de: {
+    Language: "Sprache", "Version-matched documentation is bundled with this server and works without internet access.": "Zur Serverversion passende Dokumentation ist enthalten und offline verfügbar.",
+    "InsightFace Server dashboard": "InsightFace Server Übersicht", "The key stays in this tab's memory and is never saved by the UI.": "Der Key bleibt nur im Speicher dieses Tabs und wird von der UI nicht gespeichert.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "Erkennen, vergleichen, registrieren und suchen, ohne Bilder vom Server zu übertragen.", "separate identity spaces": "getrennte Identitätsräume", "enrolled identities": "registrierte Identitäten",
+    "stored embeddings": "gespeicherte Embeddings", "API readiness": "API-Bereitschaft", Database: "Datenbank", "API failures from this browser tab will appear here.": "API-Fehler dieses Browser-Tabs erscheinen hier.", "tab memory only": "nur Tab-Speicher",
+    "The displayed language follows the console language selector.": "Die Dokumentsprache folgt der Sprachauswahl der Konsole.", "Loading documentation…": "Dokumentation wird geladen…",
+    "API key configured": "API-Schlüssel konfiguriert", Configured: "Konfiguriert", "Ready to compare": "Bereit zum Vergleichen", Close: "Schließen",
+    "Open console": "Konsole öffnen", "API Reference": "API-Referenz", "Simple face recognition APIs.": "Einfache Gesichtserkennungs-APIs.", "Base path": "Basispfad",
+    "API operations": "API-Operationen", "Filter operations": "Operationen filtern", "Loading API schema": "API-Schema wird geladen", COMPONENTS: "KOMPONENTEN",
+    "Data schemas": "Datenschemas", Parameters: "Parameter", required: "erforderlich", "Request body": "Request-Body", Responses: "Antworten",
+    "Open-source model license": "Lizenz für Open-Source-Modelle", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "Von InsightFace bereitgestellte Open-Source-Modelle, einschließlich buffalo_l, dürfen nur für nichtkommerzielle Forschung verwendet werden.", "Commercial use requires a separate license.": "Die kommerzielle Nutzung erfordert eine separate Lizenz.", "Commercial licensing": "Kommerzielle Lizenzierung",
+  },
+  es: {
+    Language: "Idioma", "Version-matched documentation is bundled with this server and works without internet access.": "La documentación correspondiente a esta versión está integrada y funciona sin internet.",
+    "InsightFace Server dashboard": "Panel de InsightFace Server", "The key stays in this tab's memory and is never saved by the UI.": "La clave solo permanece en la memoria de esta pestaña y la UI no la guarda.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "Detecte, compare, registre y busque sin enviar imágenes fuera de su servidor.", "separate identity spaces": "espacios de identidad separados", "enrolled identities": "identidades registradas",
+    "stored embeddings": "vectores guardados", "API readiness": "disponibilidad API", Database: "Base de datos", "API failures from this browser tab will appear here.": "Los errores API de esta pestaña aparecerán aquí.", "tab memory only": "solo memoria de pestaña",
+    "The displayed language follows the console language selector.": "El idioma del documento sigue el selector de la consola.", "Loading documentation…": "Cargando documentación…",
+    "API key configured": "Clave API configurada", Configured: "Configurada", "Ready to compare": "Listo para comparar", Close: "Cerrar",
+    "Open console": "Abrir consola", "API Reference": "Referencia API", "Simple face recognition APIs.": "API sencillas de reconocimiento facial.", "Base path": "Ruta base",
+    "API operations": "Operaciones API", "Filter operations": "Filtrar operaciones", "Loading API schema": "Cargando esquema API", COMPONENTS: "COMPONENTES",
+    "Data schemas": "Esquemas de datos", Parameters: "Parámetros", required: "obligatorio", "Request body": "Cuerpo de solicitud", Responses: "Respuestas",
+    "Open-source model license": "Licencia del modelo de código abierto", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "Los modelos preentrenados de código abierto proporcionados por InsightFace, incluido buffalo_l, son solo para investigación no comercial.", "Commercial use requires a separate license.": "El uso comercial requiere una licencia independiente.", "Commercial licensing": "Licencia comercial",
+  },
+  fr: {
+    Language: "Langue", "Version-matched documentation is bundled with this server and works without internet access.": "La documentation correspondant à cette version est intégrée et disponible hors ligne.",
+    "InsightFace Server dashboard": "Tableau de bord InsightFace Server", "The key stays in this tab's memory and is never saved by the UI.": "La clé reste uniquement en mémoire dans cet onglet et n'est pas enregistrée par l'interface.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "Détectez, comparez, inscrivez et recherchez sans envoyer les images hors de votre serveur.", "separate identity spaces": "espaces d'identité séparés", "enrolled identities": "identités inscrites",
+    "stored embeddings": "embeddings enregistrés", "API readiness": "disponibilité API", Database: "Base de données", "API failures from this browser tab will appear here.": "Les erreurs API de cet onglet apparaîtront ici.", "tab memory only": "mémoire de l'onglet uniquement",
+    "The displayed language follows the console language selector.": "La langue du document suit le sélecteur de la console.", "Loading documentation…": "Chargement de la documentation…",
+    "API key configured": "Clé API configurée", Configured: "Configurée", "Ready to compare": "Prêt à comparer", Close: "Fermer",
+    "Open console": "Ouvrir la console", "API Reference": "Référence API", "Simple face recognition APIs.": "API simples de reconnaissance faciale.", "Base path": "Chemin de base",
+    "API operations": "Opérations API", "Filter operations": "Filtrer les opérations", "Loading API schema": "Chargement du schéma API", COMPONENTS: "COMPOSANTS",
+    "Data schemas": "Schémas de données", Parameters: "Paramètres", required: "obligatoire", "Request body": "Corps de requête", Responses: "Réponses",
+    "Open-source model license": "Licence du modèle open source", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "Les modèles préentraînés open source fournis par InsightFace, dont buffalo_l, sont réservés à la recherche non commerciale.", "Commercial use requires a separate license.": "L’utilisation commerciale nécessite une licence distincte.", "Commercial licensing": "Licence commerciale",
+  },
+  ru: {
+    Language: "Язык", "Version-matched documentation is bundled with this server and works without internet access.": "Документация этой версии встроена в сервер и доступна без интернета.",
+    "InsightFace Server dashboard": "Панель InsightFace Server", "The key stays in this tab's memory and is never saved by the UI.": "Ключ хранится только в памяти вкладки и не сохраняется интерфейсом.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "Детектируйте, сравнивайте, регистрируйте и ищите, не отправляя изображения за пределы сервера.", "separate identity spaces": "отдельные пространства идентичностей", "enrolled identities": "зарегистрированные личности",
+    "stored embeddings": "сохранённые векторы", "API readiness": "готовность API", Database: "База данных", "API failures from this browser tab will appear here.": "Ошибки API этой вкладки появятся здесь.", "tab memory only": "только память вкладки",
+    "The displayed language follows the console language selector.": "Язык документа соответствует выбору в консоли.", "Loading documentation…": "Загрузка документации…",
+    "API key configured": "API-ключ настроен", Configured: "Настроено", "Ready to compare": "Готово к сравнению", Close: "Закрыть",
+    "Open console": "Открыть консоль", "API Reference": "Справочник API", "Simple face recognition APIs.": "Простые API распознавания лиц.", "Base path": "Базовый путь",
+    "API operations": "Операции API", "Filter operations": "Фильтр операций", "Loading API schema": "Загрузка схемы API", COMPONENTS: "КОМПОНЕНТЫ",
+    "Data schemas": "Схемы данных", Parameters: "Параметры", required: "обязательно", "Request body": "Тело запроса", Responses: "Ответы",
+    "Open-source model license": "Лицензия модели с открытым исходным кодом", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "Предоставленные InsightFace открытые предобученные модели, включая buffalo_l, предназначены только для некоммерческих исследований.", "Commercial use requires a separate license.": "Для коммерческого использования требуется отдельная лицензия.", "Commercial licensing": "Коммерческое лицензирование",
+  },
+  pt: {
+    Language: "Idioma", "Version-matched documentation is bundled with this server and works without internet access.": "A documentação desta versão está incluída e funciona sem internet.",
+    "InsightFace Server dashboard": "Painel do InsightFace Server", "The key stays in this tab's memory and is never saved by the UI.": "A chave fica apenas na memória deste separador e não é guardada pela interface.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "Detete, compare, registe e pesquise sem enviar imagens para fora do seu servidor.", "separate identity spaces": "espaços de identidade separados", "enrolled identities": "identidades registadas",
+    "stored embeddings": "vetores guardados", "API readiness": "disponibilidade da API", Database: "Base de dados", "API failures from this browser tab will appear here.": "Os erros API deste separador aparecerão aqui.", "tab memory only": "apenas memória do separador",
+    "The displayed language follows the console language selector.": "O idioma do documento segue o seletor da consola.", "Loading documentation…": "A carregar documentação…",
+    "API key configured": "Chave API configurada", Configured: "Configurada", "Ready to compare": "Pronto para comparar", Close: "Fechar",
+    "Open console": "Abrir consola", "API Reference": "Referência API", "Simple face recognition APIs.": "APIs simples de reconhecimento facial.", "Base path": "Caminho base",
+    "API operations": "Operações API", "Filter operations": "Filtrar operações", "Loading API schema": "A carregar esquema API", COMPONENTS: "COMPONENTES",
+    "Data schemas": "Esquemas de dados", Parameters: "Parâmetros", required: "obrigatório", "Request body": "Corpo do pedido", Responses: "Respostas",
+    "Open-source model license": "Licença do modelo open source", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "Os modelos pré-treinados open source fornecidos pela InsightFace, incluindo buffalo_l, destinam-se apenas a investigação não comercial.", "Commercial use requires a separate license.": "A utilização comercial requer uma licença separada.", "Commercial licensing": "Licenciamento comercial",
+  },
+  ko: {
+    Language: "언어", "Version-matched documentation is bundled with this server and works without internet access.": "현재 서버 버전에 맞는 문서가 내장되어 오프라인에서도 작동합니다.",
+    "InsightFace Server dashboard": "InsightFace Server 대시보드", "The key stays in this tab's memory and is never saved by the UI.": "Key는 현재 탭의 메모리에만 있으며 UI에 저장되지 않습니다.",
+    "Detect, compare, enroll, and search without sending images outside your server.": "이미지를 서버 밖으로 보내지 않고 검출, 비교, 등록, 검색할 수 있습니다.", "separate identity spaces": "분리된 신원 공간", "enrolled identities": "등록된 신원",
+    "stored embeddings": "저장된 임베딩", "API readiness": "API 준비 상태", Database: "데이터베이스", "API failures from this browser tab will appear here.": "이 브라우저 탭의 API 오류가 여기에 표시됩니다.", "tab memory only": "탭 메모리에만 저장",
+    "The displayed language follows the console language selector.": "문서 언어는 콘솔 언어 선택을 따릅니다.", "Loading documentation…": "문서를 불러오는 중…",
+    "API key configured": "API 키 설정됨", Configured: "설정됨", "Ready to compare": "비교 준비 완료", Close: "닫기",
+    "Open console": "콘솔 열기", "API Reference": "API 참조", "Simple face recognition APIs.": "간단한 얼굴 인식 API.", "Base path": "기본 경로",
+    "API operations": "API 작업", "Filter operations": "작업 필터", "Loading API schema": "API 스키마 불러오는 중", COMPONENTS: "컴포넌트",
+    "Data schemas": "데이터 스키마", Parameters: "매개변수", required: "필수", "Request body": "요청 본문", Responses: "응답",
+    "Open-source model license": "오픈 소스 모델 라이선스", "InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.": "buffalo_l을 포함하여 InsightFace가 제공하는 오픈 소스 사전 학습 모델은 비상업적 연구 용도로만 사용할 수 있습니다.", "Commercial use requires a separate license.": "상업적 사용에는 별도의 라이선스가 필요합니다.", "Commercial licensing": "상업용 라이선스",
+  },
+};
+
+const videoCatalogs = {
+  zh: {
+    "Video recognition": "视频识别", "LOCAL VIDEO": "本地视频", "Play a local video and identify every detected face. The video stays in your browser; only sampled JPEG frames are sent to this server.": "播放本地视频并识别每张检测到的人脸。视频保留在浏览器中，仅向服务器发送采样的 JPEG 帧。",
+    "Video recognition face overlays": "视频识别人脸框", "No video selected": "未选择视频", "Choose a local video. It stays in this browser; only sampled JPEG frames are sent for recognition.": "选择本地视频。视频保留在浏览器中，仅发送采样的 JPEG 帧进行识别。",
+    "Choose a video": "选择视频", "MP4, WebM, MOV, or MKV supported by this browser": "浏览器支持的 MP4、WebM、MOV 或 MKV", "No file selected": "未选择文件", "blank = collection default": "留空 = 人员库默认值", "Sampling rate": "采样频率", "Maximum faces per frame": "每帧最多人脸数",
+    "Start recognition": "开始识别", "Stop recognition": "停止识别", "Select a video, then start recognition.": "选择视频，然后开始识别。", Unrecognized: "未识别", "Below threshold": "低于阈值",
+    "Recognition running": "正在识别", "Video is playing. Sampled frames are processed without saving the video.": "视频正在播放。系统处理采样帧，但不会保存视频。", "Recognition stopped. The video remains in this browser.": "识别已停止。视频仍仅保留在浏览器中。",
+    "Video finished. The last recognition result remains visible.": "视频播放结束，最后一次识别结果仍会保留。", "Video selected. Press Start recognition to begin.": "视频已选择，点击“开始识别”即可运行。", "ID {id}": "ID {id}", "ID {id} · {similarity}": "ID {id} · {similarity}", "{faces} faces · {recognized} recognized · {duration}": "检测 {faces} 张人脸 · 识别 {recognized} 张 · {duration}",
+  },
+  ja: {
+    "Video recognition": "動画認識", "LOCAL VIDEO": "ローカル動画", "No video selected": "動画が選択されていません", "Choose a video": "動画を選択", "Sampling rate": "サンプリング頻度", "Maximum faces per frame": "フレームごとの最大顔数", "Start recognition": "認識を開始", "Stop recognition": "認識を停止", "Select a video, then start recognition.": "動画を選択して認識を開始してください。", Unrecognized: "未認識", "Below threshold": "しきい値未満", "Recognition running": "認識中",
+  },
+  de: {
+    "Video recognition": "Videoerkennung", "LOCAL VIDEO": "LOKALES VIDEO", "No video selected": "Kein Video ausgewählt", "Choose a video": "Video auswählen", "Sampling rate": "Abtastrate", "Maximum faces per frame": "Maximale Gesichter pro Bild", "Start recognition": "Erkennung starten", "Stop recognition": "Erkennung stoppen", "Select a video, then start recognition.": "Video auswählen und dann die Erkennung starten.", Unrecognized: "Nicht erkannt", "Below threshold": "Unter dem Schwellenwert", "Recognition running": "Erkennung läuft",
+  },
+  es: {
+    "Video recognition": "Reconocimiento de vídeo", "LOCAL VIDEO": "VÍDEO LOCAL", "No video selected": "Ningún vídeo seleccionado", "Choose a video": "Elegir vídeo", "Sampling rate": "Frecuencia de muestreo", "Maximum faces per frame": "Máximo de rostros por fotograma", "Start recognition": "Iniciar reconocimiento", "Stop recognition": "Detener reconocimiento", "Select a video, then start recognition.": "Seleccione un vídeo y luego inicie el reconocimiento.", Unrecognized: "Sin reconocer", "Below threshold": "Por debajo del umbral", "Recognition running": "Reconocimiento en curso",
+  },
+  fr: {
+    "Video recognition": "Reconnaissance vidéo", "LOCAL VIDEO": "VIDÉO LOCALE", "No video selected": "Aucune vidéo sélectionnée", "Choose a video": "Choisir une vidéo", "Sampling rate": "Fréquence d’échantillonnage", "Maximum faces per frame": "Visages maximum par image", "Start recognition": "Démarrer la reconnaissance", "Stop recognition": "Arrêter la reconnaissance", "Select a video, then start recognition.": "Choisissez une vidéo, puis démarrez la reconnaissance.", Unrecognized: "Non reconnu", "Below threshold": "Sous le seuil", "Recognition running": "Reconnaissance en cours",
+  },
+  ru: {
+    "Video recognition": "Распознавание видео", "LOCAL VIDEO": "ЛОКАЛЬНОЕ ВИДЕО", "No video selected": "Видео не выбрано", "Choose a video": "Выбрать видео", "Sampling rate": "Частота выборки", "Maximum faces per frame": "Максимум лиц в кадре", "Start recognition": "Начать распознавание", "Stop recognition": "Остановить распознавание", "Select a video, then start recognition.": "Выберите видео и запустите распознавание.", Unrecognized: "Не распознано", "Below threshold": "Ниже порога", "Recognition running": "Идёт распознавание",
+  },
+  pt: {
+    "Video recognition": "Reconhecimento de vídeo", "LOCAL VIDEO": "VÍDEO LOCAL", "No video selected": "Nenhum vídeo selecionado", "Choose a video": "Escolher vídeo", "Sampling rate": "Taxa de amostragem", "Maximum faces per frame": "Máximo de rostos por fotograma", "Start recognition": "Iniciar reconhecimento", "Stop recognition": "Parar reconhecimento", "Select a video, then start recognition.": "Selecione um vídeo e inicie o reconhecimento.", Unrecognized: "Não reconhecido", "Below threshold": "Abaixo do limiar", "Recognition running": "Reconhecimento em curso",
+  },
+  ko: {
+    "Video recognition": "동영상 인식", "LOCAL VIDEO": "로컬 동영상", "No video selected": "선택된 동영상 없음", "Choose a video": "동영상 선택", "Sampling rate": "샘플링 속도", "Maximum faces per frame": "프레임당 최대 얼굴 수", "Start recognition": "인식 시작", "Stop recognition": "인식 중지", "Select a video, then start recognition.": "동영상을 선택한 다음 인식을 시작하세요.", Unrecognized: "인식되지 않음", "Below threshold": "임계값 미만", "Recognition running": "인식 중",
+  },
+};
+
+let activeLocale = "en";
+const originalText = new WeakMap();
+const originalAttributes = new WeakMap();
+const catalogLayers = [UI_CATALOGS, videoCatalogs, supplementalCatalogs, catalogs];
+
+function catalogTemplate(language, source) {
+  for (const catalog of catalogLayers) {
+    if (Object.prototype.hasOwnProperty.call(catalog[language] ?? {}, source)) return catalog[language][source];
+  }
+  return undefined;
+}
+
+function escapePattern(value) {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+
+function compileLocalizedTemplate(source, localized) {
+  const variables = [];
+  let cursor = 0;
+  let pattern = "";
+  for (const match of localized.matchAll(/\{([A-Za-z0-9_]+)\}/g)) {
+    pattern += escapePattern(localized.slice(cursor, match.index));
+    pattern += "(.+?)";
+    variables.push(match[1]);
+    cursor = match.index + match[0].length;
+  }
+  if (!variables.length) return null;
+  pattern += escapePattern(localized.slice(cursor));
+  return { source, variables, pattern: new RegExp(`^${pattern}$`) };
+}
+
+const reverseExactTranslations = new Map();
+const reverseTemplateTranslations = [];
+for (const language of LANGUAGES.filter(({ code }) => code !== "en")) {
+  for (const catalog of catalogLayers) {
+    for (const [source, localized] of Object.entries(catalog[language.code] ?? {})) {
+      if (localized === source) continue;
+      if (!reverseExactTranslations.has(localized)) reverseExactTranslations.set(localized, source);
+      const compiled = compileLocalizedTemplate(source, localized);
+      if (compiled) reverseTemplateTranslations.push(compiled);
+    }
+  }
+}
+
+function canonicalMessage(value) {
+  const source = String(value);
+  const exact = reverseExactTranslations.get(source);
+  if (exact) return { source: exact, variables: {} };
+  for (const candidate of reverseTemplateTranslations) {
+    const match = candidate.pattern.exec(source);
+    if (!match) continue;
+    return {
+      source: candidate.source,
+      variables: Object.fromEntries(candidate.variables.map((name, index) => [name, match[index + 1]])),
+    };
+  }
+  return { source, variables: {} };
+}
+
+export function normalizeLocale(value) {
+  const candidate = String(value || "").trim().toLowerCase().replace("_", "-");
+  const base = candidate.split("-")[0];
+  return LANGUAGES.some((language) => language.code === base) ? base : "en";
+}
+
+export function detectLocale({ stored, languages = [] } = {}) {
+  if (stored && LANGUAGES.some((language) => language.code === stored)) return stored;
+  for (const language of languages) {
+    const normalized = normalizeLocale(language);
+    if (normalized !== "en" || String(language).toLowerCase().startsWith("en")) return normalized;
+  }
+  return "en";
+}
+
+export function locale() {
+  return activeLocale;
+}
+
+export function t(source, variables = {}, language = activeLocale) {
+  const direct = catalogTemplate(language, source);
+  const canonical = direct === undefined ? canonicalMessage(source) : { source, variables: {} };
+  const template = direct ?? catalogTemplate(language, canonical.source) ?? source;
+  const values = { ...canonical.variables, ...variables };
+  return String(template).replace(/\{([A-Za-z0-9_]+)\}/g, (_, name) => String(values[name] ?? `{${name}}`));
+}
+
+export function hasTranslation(source, language) {
+  if (language === "en" || SHARED_UI_MESSAGES.has(source)) return true;
+  return catalogLayers
+    .some((catalog) => Object.prototype.hasOwnProperty.call(catalog[language] ?? {}, source));
+}
+
+function translateTextNode(node) {
+  if (!originalText.has(node)) originalText.set(node, node.nodeValue || "");
+  const source = originalText.get(node);
+  const trimmed = source.trim();
+  if (!trimmed) return;
+  const translated = t(trimmed);
+  node.nodeValue = source.replace(trimmed, translated);
+}
+
+function translateAttributes(element) {
+  const names = ["aria-label", "placeholder", "title"];
+  if (!originalAttributes.has(element)) originalAttributes.set(element, {});
+  const saved = originalAttributes.get(element);
+  for (const name of names) {
+    if (!element.hasAttribute(name)) continue;
+    if (!(name in saved)) saved[name] = element.getAttribute(name);
+    element.setAttribute(name, t(saved[name]));
+  }
+}
+
+export function translateTree(root = document) {
+  if (root.nodeType === Node.TEXT_NODE) {
+    translateTextNode(root);
+    return;
+  }
+  if (!(root instanceof Element || root instanceof Document)) return;
+  if (root instanceof Element) translateAttributes(root);
+  const walker = document.createTreeWalker(root, NodeFilter.SHOW_ELEMENT | NodeFilter.SHOW_TEXT);
+  let node = walker.nextNode();
+  while (node) {
+    if (node.nodeType === Node.TEXT_NODE) {
+      const parent = node.parentElement;
+      if (parent && !["SCRIPT", "STYLE", "CODE", "PRE", "TEXTAREA"].includes(parent.tagName)) translateTextNode(node);
+    } else {
+      translateAttributes(node);
+    }
+    node = walker.nextNode();
+  }
+}
+
+export function setLocale(language, { persist = true, announce = true } = {}) {
+  activeLocale = normalizeLocale(language);
+  const metadata = LANGUAGES.find((item) => item.code === activeLocale) ?? LANGUAGES[0];
+  document.documentElement.lang = metadata.htmlLang;
+  if (persist) window.localStorage.setItem("insightface.locale", activeLocale);
+  translateTree(document);
+  const selector = document.querySelector("#language-select");
+  if (selector) selector.value = activeLocale;
+  if (announce) window.dispatchEvent(new CustomEvent("insightface:localechange", { detail: { locale: activeLocale } }));
+  return activeLocale;
+}
+
+export function initializeI18n() {
+  let stored = null;
+  try {
+    stored = window.localStorage.getItem("insightface.locale");
+  } catch {
+    stored = null;
+  }
+  const selected = detectLocale({ stored, languages: navigator.languages || [navigator.language] });
+  setLocale(selected, { persist: false, announce: false });
+  const selector = document.querySelector("#language-select");
+  selector?.addEventListener("change", () => setLocale(selector.value));
+  return selected;
+}
diff --git a/server/frontend/index.html b/server/frontend/index.html
new file mode 100644
index 000000000..e9b03ca27
--- /dev/null
+++ b/server/frontend/index.html
@@ -0,0 +1,460 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="utf-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1">
+    <meta name="color-scheme" content="dark">
+    <meta name="theme-color" content="#09090b">
+    <meta
+      http-equiv="Content-Security-Policy"
+      content="default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' blob: data:; connect-src 'self'; media-src 'self' blob:; object-src 'none'; base-uri 'none'; form-action 'self'"
+    >
+    <meta name="description" content="Self-hosted InsightFace detection, comparison, enrollment, and search.">
+    <title>InsightFace Server</title>
+    <link rel="icon" href="/assets/insightface-mark.svg" type="image/svg+xml">
+    <link rel="stylesheet" href="/assets/styles.css?v=0.2.0-r13">
+  </head>
+  <body>
+    <a class="skip-link" href="#main-content">Skip to content</a>
+    <div class="app-shell">
+      <aside class="sidebar" aria-label="Primary navigation">
+        <button class="brand" type="button" data-route="dashboard" aria-label="InsightFace Server dashboard">
+          <span class="brand-mark" aria-hidden="true">
+            <svg viewBox="0 0 24 24" fill="none">
+              <circle cx="12" cy="8" r="4"></circle>
+              <path d="M5.5 21a7.5 7.5 0 0 1 13 0"></path>
+            </svg>
+          </span>
+          <div>
+            <strong>InsightFace</strong>
+            <small>Server console</small>
+          </div>
+        </button>
+
+        <div class="server-chip" id="sidebar-health">
+          <span class="status-dot" aria-hidden="true"></span>
+          <span>Connecting</span>
+        </div>
+
+        <nav class="nav-list">
+          <button class="nav-item active" type="button" data-route="dashboard" aria-current="page">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><path d="m3 11 9-8 9 8"></path><path d="M5 10v10h14V10"></path><path d="M9 20v-6h6v6"></path></svg></span>Dashboard
+          </button>
+          <button class="nav-item" type="button" data-route="collections">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><ellipse cx="12" cy="5" rx="8" ry="3"></ellipse><path d="M4 5v7c0 1.7 3.6 3 8 3s8-1.3 8-3V5"></path><path d="M4 12v7c0 1.7 3.6 3 8 3s8-1.3 8-3v-7"></path></svg></span>Collections
+          </button>
+          <button class="nav-item" type="button" data-route="people">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><path d="M16 21v-2a4 4 0 0 0-4-4H6a4 4 0 0 0-4 4v2"></path><circle cx="9" cy="7" r="4"></circle><path d="M22 21v-2a4 4 0 0 0-3-3.87"></path><path d="M16 3.13a4 4 0 0 1 0 7.75"></path></svg></span>People
+          </button>
+          <button class="nav-item" type="button" data-route="detect">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><path d="M3 9V5a2 2 0 0 1 2-2h4"></path><path d="M15 3h4a2 2 0 0 1 2 2v4"></path><path d="M21 15v4a2 2 0 0 1-2 2h-4"></path><path d="M9 21H5a2 2 0 0 1-2-2v-4"></path><circle cx="12" cy="10" r="3"></circle><path d="M7 18a5 5 0 0 1 10 0"></path></svg></span>Detect
+          </button>
+          <button class="nav-item" type="button" data-route="compare">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><path d="m16 3 4 4-4 4"></path><path d="M20 7H4"></path><path d="m8 21-4-4 4-4"></path><path d="M4 17h16"></path></svg></span>Compare
+          </button>
+          <button class="nav-item" type="button" data-route="search">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><circle cx="11" cy="11" r="7"></circle><path d="m20 20-4-4"></path></svg></span>Search
+          </button>
+          <button class="nav-item" type="button" data-route="video">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><path d="m16 13 5 3V8l-5 3"></path><rect x="3" y="6" width="13" height="12" rx="2"></rect></svg></span>Camera monitoring
+          </button>
+          <button class="nav-item" type="button" data-route="system">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><rect x="4" y="4" width="16" height="16" rx="2"></rect><rect x="9" y="9" width="6" height="6"></rect><path d="M9 1v3M15 1v3M9 20v3M15 20v3M20 9h3M20 14h3M1 9h3M1 14h3"></path></svg></span>System
+          </button>
+          <button class="nav-item" type="button" data-route="help">
+            <span class="nav-icon" aria-hidden="true"><svg viewBox="0 0 24 24"><circle cx="12" cy="12" r="9"></circle><path d="M9.7 9a2.5 2.5 0 1 1 3.8 2.1c-.9.6-1.5 1.1-1.5 2.4"></path><path d="M12 17h.01"></path></svg></span>Help
+          </button>
+        </nav>
+
+        <div class="sidebar-footer" data-auth-ui hidden>
+          <button class="button secondary full-width" id="open-api-key" type="button">Configure API key</button>
+          <p class="muted tiny">The key stays in this tab's memory and is never saved by the UI.</p>
+        </div>
+      </aside>
+
+      <main id="main-content" class="main-content" tabindex="-1">
+        <header class="topbar">
+          <button class="menu-button" id="menu-toggle" type="button" aria-label="Toggle navigation" aria-expanded="false">☰</button>
+          <div>
+            <p class="eyebrow">INSIGHTFACE SERVER</p>
+            <h1 id="page-title">Dashboard</h1>
+          </div>
+          <div class="topbar-actions">
+            <label class="language-control" title="Language">
+              <span class="sr-only">Language</span>
+              <select id="language-select" aria-label="Language">
+                <option value="en">English</option>
+                <option value="zh">中文</option>
+                <option value="ja">日本語</option>
+                <option value="de">Deutsch</option>
+                <option value="es">Español</option>
+                <option value="fr">Français</option>
+                <option value="ru">Русский</option>
+                <option value="pt">Português</option>
+                <option value="ko">한국어</option>
+              </select>
+            </label>
+            <span class="auth-state" id="auth-state" data-auth-ui hidden><span aria-hidden="true">○</span> No API key</span>
+            <button class="icon-button" id="refresh-page" type="button" aria-label="Refresh current page" title="Refresh">↻</button>
+          </div>
+        </header>
+
+        <div class="page-stack">
+          <section class="page" data-page="dashboard" aria-labelledby="dashboard-heading">
+            <div class="hero-panel">
+              <div>
+                <p class="eyebrow">LOCAL FACE INTELLIGENCE</p>
+                <h2 id="dashboard-heading">Your face recognition service, <span class="gradient-text">at a glance.</span></h2>
+                <p>Detect, compare, enroll, and search without sending images outside your server.</p>
+              </div>
+              <div class="hero-provider">
+                <span>Execution provider</span>
+                <strong id="dashboard-provider">Checking…</strong>
+              </div>
+            </div>
+
+            <div class="metric-grid" aria-label="Database totals">
+              <article class="metric-card"><span>Collections</span><strong id="metric-collections">—</strong><small>separate identity spaces</small></article>
+              <article class="metric-card"><span>People</span><strong id="metric-people">—</strong><small>enrolled identities</small></article>
+              <article class="metric-card"><span>Face samples</span><strong id="metric-faces">—</strong><small>stored embeddings</small></article>
+              <article class="metric-card"><span>Model</span><strong class="metric-text" id="metric-model">—</strong><small id="metric-model-version">not loaded</small></article>
+            </div>
+
+            <div class="two-column">
+              <article class="panel">
+                <div class="panel-heading"><div><p class="eyebrow">SERVICE</p><h3>Runtime status</h3></div><button class="text-button" type="button" data-route="system">View diagnostics →</button></div>
+                <dl class="status-list" id="dashboard-status"></dl>
+              </article>
+              <article class="panel">
+                <div class="panel-heading"><div><p class="eyebrow">THIS TAB</p><h3>Recent errors</h3></div><button class="text-button" id="clear-errors" type="button">Clear</button></div>
+                <div id="recent-errors" class="empty-state compact"><strong>No recent errors</strong><span>API failures from this browser tab will appear here.</span></div>
+              </article>
+            </div>
+          </section>
+
+          <section class="page" data-page="collections" aria-labelledby="collections-heading" hidden>
+            <div class="section-heading">
+              <div><p class="eyebrow">IDENTITY SPACES</p><h2 id="collections-heading">Collections</h2><p>Each collection is pinned to the model and preprocessing version active when it is created.</p></div>
+              <button class="button primary" id="show-create-collection" type="button">＋ New collection</button>
+            </div>
+
+            <form class="panel form-panel collapsible" id="collection-create-form" hidden>
+              <div class="panel-heading"><div><p class="eyebrow">CREATE</p><h3>New collection</h3></div><button class="icon-button close-form" type="button" aria-label="Close">×</button></div>
+              <div class="form-grid three">
+                <label><span class="field-caption">Collection ID</span><input name="id" required pattern="(_default|[A-Za-z0-9][A-Za-z0-9._-]{0,63})" maxlength="64" placeholder="employees" autocomplete="off"><small>Letters, numbers, dot, dash, underscore; max 64. `_default` is reserved for the default collection.</small></label>
+                <label><span class="field-caption">Name</span><input name="name" required maxlength="128" placeholder="Company Employees"></label>
+                <label><span class="field-caption">Default cosine threshold</span><input name="threshold" type="number" min="0" max="1" step="0.01" value="0.4" required></label>
+                <label><span class="field-caption">Search profile</span><select name="search_profile" required><option value="fp32_v1">FP32</option><option value="fp16_v1">FP16</option><option value="bf16_v1">BF16</option><option value="int8_x736_v1">INT8 ×736</option><option value="int8_x1000_v1">INT8 ×1000</option></select></label>
+                <label><span class="field-caption">Capacity (faces)</span><input name="capacity_rows" type="number" min="1" step="1" value="100000" required></label>
+                <label><span class="field-caption">Maximum faces per person</span><input name="max_faces_per_person" type="number" min="1" max="1000000" step="1" value="20" required></label>
+                <label><span class="field-caption">Load policy</span><select name="load_policy"><option value="">Server default</option><option value="lazy">Lazy</option><option value="eager">Eager</option></select></label>
+                <label><span class="field-caption">Detector input sizes</span><input name="detector_input_sizes" value="96x96, 512x512" required><small>Comma-separated width×height pairs; each side must be a multiple of 32.</small></label>
+                <label><span class="field-caption">Detection threshold</span><input name="detector_threshold" type="number" min="0" max="1" step="0.01" value="0.5" required></label>
+                <label><span class="field-caption">NMS threshold</span><input name="detector_nms_threshold" type="number" min="0" max="1" step="0.01" value="0.4" required></label>
+                <label><span class="field-caption">Single-face selection</span><select name="single_face_selection" required><option value="largest">Largest face</option><option value="center_largest">Center-weighted largest</option></select></label>
+              </div>
+              <label class="checkbox-field"><input name="save_face_crops" type="checkbox"><span>Save 112×112 face crops in SQLite</span><small>Off by default. This never stores the original upload and may substantially increase database size.</small></label>
+              <label>Description<textarea name="description" rows="2" maxlength="1000" placeholder="What belongs in this collection?"></textarea></label>
+              <label>Metadata (JSON)<textarea name="metadata" rows="3" spellcheck="false">{}</textarea></label>
+              <div class="form-actions"><button class="button primary" type="submit">Create collection</button><span class="form-status" role="status"></span></div>
+            </form>
+
+            <article class="panel">
+              <div class="table-toolbar">
+                <label class="search-field"><span class="sr-only">Filter collections</span><input id="collection-filter" type="search" placeholder="Filter collections…"></label>
+                <span class="muted" id="collection-count">0 collections</span>
+              </div>
+              <div class="table-wrap">
+                <table>
+                  <thead><tr><th>Collection</th><th>Threshold</th><th>Profile</th><th>Crops</th><th>Model</th><th>People</th><th>Faces</th><th><span class="sr-only">Actions</span></th></tr></thead>
+                  <tbody id="collection-table"></tbody>
+                </table>
+              </div>
+              <div class="empty-state" id="collection-empty"><strong>No collections yet</strong><span>Create one to start enrolling people.</span></div>
+            </article>
+          </section>
+
+          <section class="page" data-page="people" aria-labelledby="people-heading" hidden>
+            <div class="section-heading">
+              <div><p class="eyebrow">ENROLLMENT</p><h2 id="people-heading">People &amp; face samples</h2><p>Register several clear photos for each real person to improve search robustness.</p></div>
+              <button class="button primary" id="show-create-person" type="button">＋ Register person</button>
+            </div>
+
+            <div class="toolbar panel slim">
+              <label>Collection<select id="people-collection"><option value="">Select a collection</option></select></label>
+              <label class="grow">Search people<input id="people-filter" type="search" placeholder="Name, ID, or external ID"></label>
+            </div>
+
+            <form class="panel form-panel collapsible" id="person-create-form" hidden>
+              <div class="panel-heading"><div><p class="eyebrow">REGISTER</p><h3>New person</h3></div><button class="icon-button close-form" type="button" aria-label="Close">×</button></div>
+              <div class="form-grid three">
+                <label><span class="field-caption">Person ID <span class="optional">optional</span></span><input name="id" maxlength="64" pattern="[A-Za-z0-9][A-Za-z0-9._-]{0,63}" placeholder="employee-001"></label>
+                <label><span class="field-caption">Name</span><input name="name" maxlength="200" placeholder="Alice"></label>
+                <label><span class="field-caption">External ID</span><input name="external_id" maxlength="200" placeholder="HR-1001"></label>
+                <label><span class="field-caption">Enrollment review</span><select name="review_mode"><option value="off" selected>Off · collection selection</option><option value="standard">Standard quality</option><option value="strict">Strict identity</option></select></label>
+                <label><span class="field-caption">Embedding source</span><select name="embedding_mode"><option value="server" selected>Extract from image</option><option value="external_trusted">Trusted external feature</option></select></label>
+              </div>
+              <p class="hint embedding-contract-summary">Collection embedding contract: <code class="embedding-contract">Select a Collection</code></p>
+              <div class="external-embedding-fields" hidden>
+                <p class="hint">Images are still required for detection and quality review; the server will not extract or verify these features.</p>
+                <label>External embeddings (JSON, one vector per image)<textarea name="external_embeddings" rows="6" spellcheck="false" placeholder="[[0.0123, -0.0456, ...]]"></textarea></label>
+              </div>
+              <label>Metadata (JSON)<textarea name="metadata" rows="3" spellcheck="false">{}</textarea></label>
+              <label class="drop-zone" id="person-drop-zone">
+                <input name="images" type="file" accept="image/jpeg,image/png,image/webp" multiple required>
+                <span class="drop-icon" aria-hidden="true">＋</span>
+                <strong>Drop registration photos here</strong>
+                <span>or choose one or more JPEG, PNG, or WebP images</span>
+                <small class="file-summary">No files selected</small>
+              </label>
+              <p class="hint">Off uses the Collection single-face selection strategy. Standard and strict require exactly one clear, usable face.</p>
+              <div class="form-actions"><button class="button primary" type="submit">Register person</button><span class="form-status" role="status"></span></div>
+              <div class="rejection-list" id="person-rejections" hidden></div>
+            </form>
+
+            <div class="people-layout">
+              <article class="panel people-list-panel">
+                <div class="panel-heading"><div><p class="eyebrow">DIRECTORY</p><h3 id="people-list-title">People</h3></div><span class="muted" id="people-count">0 people</span></div>
+                <div class="person-list" id="person-list"></div>
+                <div class="empty-state" id="people-empty"><strong>Select a collection</strong><span>Its enrolled people will appear here.</span></div>
+              </article>
+              <article class="panel detail-panel" id="person-detail">
+                <div class="empty-state"><strong>No person selected</strong><span>Choose a person to view details and face samples.</span></div>
+              </article>
+            </div>
+          </section>
+
+          <section class="page" data-page="detect" aria-labelledby="detect-heading" hidden>
+            <div class="section-heading"><div><p class="eyebrow">STATELESS</p><h2 id="detect-heading">Detect faces</h2><p>Find faces, five-point landmarks, confidence, and quality in one image.</p></div></div>
+            <div class="tool-layout">
+              <form class="panel tool-controls" id="detect-form">
+                <label class="drop-zone compact-drop" id="detect-drop-zone">
+                  <input name="image" type="file" accept="image/jpeg,image/png,image/webp" required>
+                  <span class="drop-icon" aria-hidden="true">⌗</span><strong>Drop an image</strong><span>JPEG, PNG, or WebP</span><small class="file-summary">No file selected</small>
+                </label>
+                <div class="form-grid two">
+                  <label><span class="field-caption">Maximum faces</span><input name="max_faces" type="number" min="1" max="100" placeholder="All"></label>
+                  <label><span class="field-caption">Detection profile</span><select name="collection" id="detect-collection"><option value="">System default</option></select></label>
+                </div>
+                <button class="button primary full-width" type="submit">Detect faces</button>
+                <div class="result-summary" id="detect-summary" aria-live="polite">Choose an image to begin.</div>
+              </form>
+              <article class="panel media-panel">
+                <div class="canvas-stage empty" id="detect-stage"><canvas id="detect-canvas" aria-label="Uploaded image with detected face overlays"></canvas><span class="canvas-placeholder">Image preview</span></div>
+                <div class="face-result-list" id="detect-results"></div>
+              </article>
+            </div>
+          </section>
+
+          <section class="page" data-page="compare" aria-labelledby="compare-heading" hidden>
+            <div class="section-heading"><div><p class="eyebrow">ONE-TO-ONE</p><h2 id="compare-heading">Compare two faces</h2><p>The selected detection profile chooses one face from each image for cosine comparison.</p></div></div>
+            <form id="compare-form">
+              <div class="compare-grid">
+                <article class="panel media-panel">
+                  <div class="panel-heading"><h3>Source image</h3></div>
+                  <label class="drop-zone compact-drop" id="compare-source-drop"><input name="source" type="file" accept="image/jpeg,image/png,image/webp" required><strong>Choose source</strong><span>JPEG, PNG, or WebP</span><span class="file-summary">No file selected</span></label>
+                  <div class="canvas-stage empty"><canvas id="compare-source-canvas" aria-label="Source image face overlay"></canvas><span class="canvas-placeholder">Source preview</span></div>
+                </article>
+                <article class="panel media-panel">
+                  <div class="panel-heading"><h3>Target image</h3></div>
+                  <label class="drop-zone compact-drop" id="compare-target-drop"><input name="target" type="file" accept="image/jpeg,image/png,image/webp" required><strong>Choose target</strong><span>JPEG, PNG, or WebP</span><span class="file-summary">No file selected</span></label>
+                  <div class="canvas-stage empty"><canvas id="compare-target-canvas" aria-label="Target image face overlay"></canvas><span class="canvas-placeholder">Target preview</span></div>
+                </article>
+              </div>
+              <div class="panel compare-action-bar">
+                <label><span class="field-caption">Detection profile</span><select name="collection" id="compare-collection"><option value="">System default</option></select></label>
+                <label><span class="field-caption">Match threshold <output id="compare-threshold-output">0.40</output></span><input name="threshold" id="compare-threshold" type="range" min="0" max="1" step="0.01" value="0.4"></label>
+                <button class="button primary" type="submit">Compare faces</button>
+                <div class="compare-verdict" id="compare-verdict" aria-live="polite"><span class="verdict-mark">—</span><div><strong>Waiting for images</strong><small>Similarity is not a probability.</small></div></div>
+              </div>
+            </form>
+          </section>
+
+          <section class="page" data-page="search" aria-labelledby="search-heading" hidden>
+            <div class="section-heading"><div><p class="eyebrow">ONE-TO-MANY</p><h2 id="search-heading">Search a collection</h2><p>Find the closest enrolled people using the Collection single-face selection strategy.</p></div></div>
+            <div class="tool-layout">
+              <form class="panel tool-controls" id="search-form">
+                <label>Collection<select name="collection" id="search-collection" required><option value="">Select a collection</option></select></label>
+                <label class="drop-zone compact-drop" id="search-drop-zone"><input name="image" type="file" accept="image/jpeg,image/png,image/webp" required><span class="drop-icon" aria-hidden="true">⌕</span><strong>Drop a query image</strong><span>JPEG, PNG, or WebP</span><small class="file-summary">No file selected</small></label>
+                <div class="form-grid two search-parameters">
+                  <label><span class="field-caption">Top results</span><input name="limit" type="number" min="1" max="100" value="5" required></label>
+                  <label><span class="field-caption">Threshold <span class="optional">blank = collection default</span></span><input name="threshold" type="number" min="0" max="1" step="0.01" placeholder="Default"></label>
+                </div>
+                <button class="button primary full-width" type="submit">Search people</button>
+                <div class="result-summary" id="search-summary" aria-live="polite">Choose a collection and image.</div>
+              </form>
+              <article class="panel media-panel">
+                <div class="canvas-stage empty" id="search-stage"><canvas id="search-canvas" aria-label="Search image face overlay"></canvas><span class="canvas-placeholder">Query preview</span></div>
+                <div class="match-list" id="search-results"></div>
+              </article>
+            </div>
+          </section>
+
+          <section class="page" data-page="video" aria-labelledby="video-heading" hidden>
+            <div class="section-heading"><div><p class="eyebrow">RTSP MONITORS</p><h2 id="video-heading">Camera monitoring</h2><p>Create persistent server-side Monitors, inspect their current state, and pull recent in-memory events. Video frames and events are not saved.</p></div><button class="button primary" id="show-create-monitor" type="button">New Monitor</button></div>
+            <div class="two-column monitor-admin">
+              <article class="panel">
+                <div class="panel-heading"><div><p class="eyebrow">CONFIGURATION</p><h3>Monitors</h3></div><button class="text-button" id="refresh-monitors" type="button">Refresh</button></div>
+                <div class="monitor-list" id="monitor-list"></div>
+                <div class="empty-state compact" id="monitor-empty"><strong>No Monitors</strong><span>Create a Monitor to connect an RTSP camera.</span></div>
+              </article>
+              <form class="panel form-panel" id="monitor-create-form" hidden>
+                <div class="panel-heading"><div><p class="eyebrow" id="monitor-form-eyebrow">NEW MONITOR</p><h3 id="monitor-form-title">Create Monitor</h3></div><button class="icon-button close-form" type="button" aria-label="Close">×</button></div>
+                <div class="form-grid two">
+                  <label><span class="field-caption">Monitor ID</span><input name="id" maxlength="64" pattern="[A-Za-z0-9][A-Za-z0-9._-]{0,63}" placeholder="front-gate" required></label>
+                  <label><span class="field-caption">Name</span><input name="name" maxlength="200" placeholder="Company front gate" required></label>
+                </div>
+                <label><span class="field-caption">Description</span><input name="description" maxlength="2000" placeholder="Optional location or purpose"></label>
+                <label><span class="field-caption">RTSP URL</span><input name="url" id="rtsp-url" type="url" inputmode="url" autocomplete="off" spellcheck="false" placeholder="rtsp://camera.example:8554/stream" required><small>The credential is encrypted in /data and is never returned by the API.</small></label>
+                <label><span class="field-caption">Collection</span><select name="collection_id" id="video-collection" required><option value="">Select a collection</option></select></label>
+                <div class="form-grid two">
+                  <label><span class="field-caption">Inference rate</span><input name="inference_fps" type="number" min="0.1" max="30" step="0.1" value="2" required><small>Executions per second</small></label>
+                  <label><span class="field-caption">Match threshold <span class="optional">blank = collection default</span></span><input name="match_threshold" type="number" min="0" max="1" step="0.01" placeholder="Default"></label>
+                </div>
+                <div class="form-grid two">
+                  <label><span class="field-caption">Recent event capacity</span><input name="event_buffer_size" type="number" min="10" max="10000" value="1000" required></label>
+                  <label><span class="field-caption">Confirmation frames</span><input name="confirm_frames" type="number" min="1" max="100" value="3" required></label>
+                </div>
+                <div class="form-grid two">
+                  <label><span class="field-caption">Absence timeout</span><input name="absence_timeout_seconds" type="number" min="0.1" max="300" step="0.1" value="3" required><small>Seconds before an exit event</small></label>
+                  <label><span class="field-caption">Event cooldown</span><input name="cooldown_seconds" type="number" min="0" max="3600" step="0.1" value="10" required><small>Seconds between repeated events</small></label>
+                </div>
+                <label class="checkbox-line"><input name="enabled" type="checkbox" checked><span><strong id="monitor-enabled-label">Start after creation</strong><small id="monitor-enabled-help">Run independently of this browser.</small></span></label>
+                <label class="checkbox-line"><input name="emit_unknown" type="checkbox" checked><span><strong>Emit unknown-person events</strong><small>Include confirmed faces that are not enrolled.</small></span></label>
+                <label class="checkbox-line"><input name="preview_enabled" type="checkbox"><span><strong>Enable Web video preview</strong><small>Off by default. Recognition and events work without preview.</small></span></label>
+                <button class="button primary full-width" id="monitor-form-submit" type="submit">Create Monitor</button>
+              </form>
+            </div>
+            <div class="tool-layout video-layout">
+              <article class="panel video-view">
+                <div class="video-stage">
+                  <img id="rtsp-preview" alt="RTSP camera preview" hidden>
+                  <canvas id="rtsp-overlay" aria-label="Recognition overlay" hidden></canvas>
+                  <div class="video-idle" id="video-idle"><span class="video-symbol" aria-hidden="true">●</span><strong id="video-idle-title">No Monitor selected</strong><span id="video-idle-message">Select a Monitor to inspect its live recognition state.</span></div>
+                </div>
+                <div class="stream-legend" aria-label="Recognition color legend">
+                  <span><i class="legend-swatch recognized" aria-hidden="true"></i>Recognized person</span>
+                  <span><i class="legend-swatch not-enrolled" aria-hidden="true"></i>Detected, not enrolled</span>
+                </div>
+              </article>
+              <article class="panel tool-controls" id="monitor-runtime">
+                <div class="panel-heading"><div><p class="eyebrow">CURRENT STATE</p><h3 id="selected-monitor-name">Select a Monitor</h3></div><span class="badge" id="selected-monitor-id">—</span></div>
+                <div class="monitor-actions">
+                  <button class="button secondary small" id="monitor-toggle" type="button" disabled>Start</button>
+                  <button class="button secondary small" id="monitor-preview-toggle" type="button" disabled>Enable preview</button>
+                  <button class="button secondary small" id="monitor-edit" type="button" disabled>Edit</button>
+                  <button class="button danger small" id="monitor-delete" type="button" disabled>Delete</button>
+                </div>
+                <dl class="stream-metrics" id="monitor-config" hidden>
+                  <div><dt>Source</dt><dd id="monitor-source">—</dd></div>
+                  <div><dt>Collection</dt><dd id="monitor-collection">—</dd></div>
+                  <div><dt>Match threshold</dt><dd id="monitor-threshold">—</dd></div>
+                  <div><dt>Recent event capacity</dt><dd id="monitor-event-capacity">—</dd></div>
+                </dl>
+                <div class="result-summary" id="video-status" aria-live="polite">Select a Monitor to inspect it.</div>
+                <dl class="stream-metrics" id="video-metrics" hidden>
+                  <div><dt>Connection</dt><dd id="video-connection">—</dd></div>
+                  <div><dt>Effective rate</dt><dd id="video-effective-fps">—</dd></div>
+                  <div><dt>Inference time</dt><dd id="video-latency">—</dd></div>
+                  <div><dt>Skipped frames</dt><dd id="video-skipped">—</dd></div>
+                </dl>
+                <div class="video-matches" id="video-matches"></div>
+                <div class="panel-heading monitor-events-heading"><div><p class="eyebrow">IN-MEMORY BUFFER</p><h3>Recent events</h3></div><span class="badge" id="monitor-event-count">0</span></div>
+                <div class="monitor-events" id="monitor-events"></div>
+              </article>
+            </div>
+          </section>
+
+          <section class="page" data-page="system" aria-labelledby="system-heading" hidden>
+            <div class="section-heading"><div><p class="eyebrow">ADMINISTRATION</p><h2 id="system-heading">System</h2><p>Runtime diagnostics and safe system configuration. Secrets and embeddings are never displayed.</p></div><button class="button secondary" id="system-refresh" type="button">Refresh system</button></div>
+            <div class="metric-grid system-metrics">
+              <article class="metric-card"><span>Server</span><strong class="metric-text" id="system-server">—</strong><small id="system-platform">—</small></article>
+              <article class="metric-card"><span>Provider</span><strong class="metric-text" id="system-provider">—</strong><small id="system-ort">—</small></article>
+              <article class="metric-card"><span>Accelerator</span><strong class="metric-text" id="system-gpu">—</strong><small id="system-driver">—</small></article>
+              <article class="metric-card" data-auth-ui hidden><span>API key in UI</span><strong class="metric-text" id="system-auth">Not configured</strong><small>tab memory only</small></article>
+            </div>
+            <div class="two-column">
+              <article class="panel"><div class="panel-heading"><div><p class="eyebrow">RUNTIME</p><h3>Hardware &amp; software</h3></div></div><dl class="diagnostic-list" id="system-runtime"></dl></article>
+              <article class="panel"><div class="panel-heading"><div><p class="eyebrow">STORAGE</p><h3>Database &amp; mounts</h3></div></div><dl class="diagnostic-list" id="system-storage"></dl></article>
+            </div>
+            <article class="panel">
+              <div class="panel-heading"><div><p class="eyebrow">MODELS</p><h3>Loaded models</h3></div></div>
+              <aside class="model-license-notice" aria-label="Open-source model license">
+                <strong>Open-source model license</strong>
+                <p>InsightFace-provided open-source pretrained models, including buffalo_l, are for non-commercial research use only.</p>
+                <p>Commercial use requires a separate license. <a href="https://www.insightface.ai" target="_blank" rel="noopener noreferrer">Commercial licensing</a></p>
+              </aside>
+              <div class="model-grid" id="system-models"></div>
+            </article>
+            <article class="panel developer-tools-panel">
+              <div>
+                <p class="eyebrow">DEVELOPER TOOLS</p>
+                <h3>OpenAPI Schema</h3>
+                <p>Inspect the exact paths, parameters, and response schemas exposed by this running server.</p>
+              </div>
+              <div class="developer-tools-actions">
+                <a class="button secondary" href="/docs" target="_blank" rel="noopener">Open schema explorer</a>
+                <a class="text-link" href="/openapi.json" target="_blank" rel="noopener">OpenAPI JSON</a>
+              </div>
+            </article>
+          </section>
+
+          <section class="page" data-page="help" aria-labelledby="help-heading" hidden>
+            <div class="section-heading">
+              <div><p class="eyebrow">DOCUMENTATION</p><h2 id="help-heading">Help</h2><p>Beginner guidance and complete API documentation are bundled with this server and work without internet access.</p></div>
+            </div>
+            <div class="documentation-switcher" id="documentation-switcher" role="group" aria-label="Documentation"></div>
+            <div class="documentation-layout">
+              <aside class="panel documentation-nav" aria-label="Documentation">
+                <strong>On this page</strong>
+                <label class="documentation-filter"><span>Find a topic</span><input id="documentation-filter" type="search" placeholder="Search help topics…"></label>
+                <nav class="documentation-toc" id="documentation-toc" aria-label="Help topics"></nav>
+                <p class="muted tiny">The displayed language follows the console language selector.</p>
+              </aside>
+              <article class="panel markdown-document" id="markdown-document" aria-live="polite">
+                <div class="loading-document"><span></span><strong>Loading documentation…</strong></div>
+              </article>
+            </div>
+            <aside class="panel documentation-developer-note">
+              <div><strong>Developer schema</strong><p>Use the OpenAPI schema when generating clients or checking exact runtime types.</p></div>
+              <a class="button secondary" href="/docs" target="_blank" rel="noopener">Open schema explorer</a>
+            </aside>
+          </section>
+        </div>
+      </main>
+    </div>
+
+    <dialog id="api-key-dialog" data-auth-ui hidden>
+      <form method="dialog" id="api-key-form">
+        <div class="dialog-heading"><div><p class="eyebrow">AUTHENTICATION</p><h2>API key</h2></div><button class="icon-button" value="cancel" aria-label="Close">×</button></div>
+        <p>Requests from this tab will send the key as a Bearer token. It is not written to cookies, local storage, logs, or the URL.</p>
+        <label>API key<input name="api_key" type="password" autocomplete="off" spellcheck="false" placeholder="Paste API key"></label>
+        <div class="dialog-actions"><button class="button secondary" id="clear-api-key" type="button">Clear key</button><button class="button primary" value="default" type="submit">Use for this tab</button></div>
+      </form>
+    </dialog>
+
+    <dialog id="collection-edit-dialog">
+      <form method="dialog" id="collection-edit-form">
+        <div class="dialog-heading"><div><p class="eyebrow">COLLECTION</p><h2>Edit collection</h2></div><button class="icon-button" value="cancel" aria-label="Close">×</button></div>
+        <input name="id" type="hidden">
+        <label>Name<input name="name" required maxlength="128"></label>
+        <label>Description<textarea name="description" rows="3" maxlength="1000"></textarea></label>
+        <label>Default cosine threshold<input name="threshold" required type="number" min="0" max="1" step="0.01"></label>
+        <label>Search profile<input name="search_profile" readonly></label>
+        <label>Capacity (faces)<input name="capacity_rows" required type="number" min="1" step="1"></label>
+        <label>Maximum faces per person<input name="max_faces_per_person" required type="number" min="1" max="1000000" step="1"></label>
+        <label>Load policy<select name="load_policy"><option value="lazy">Lazy</option><option value="eager">Eager</option></select></label>
+        <label>Detector input sizes<input name="detector_input_sizes" required><small>Comma-separated width×height pairs.</small></label>
+        <label>Detection threshold<input name="detector_threshold" required type="number" min="0" max="1" step="0.01"></label>
+        <label>NMS threshold<input name="detector_nms_threshold" required type="number" min="0" max="1" step="0.01"></label>
+        <label>Single-face selection<select name="single_face_selection"><option value="largest">Largest face</option><option value="center_largest">Center-weighted largest</option></select></label>
+        <label class="checkbox-field"><input name="save_face_crops" type="checkbox"><span>Save 112×112 face crops in SQLite</span><small>Applies to later registrations; an already-running request may use the previous setting.</small></label>
+        <label>Metadata (JSON)<textarea name="metadata" rows="4" spellcheck="false"></textarea></label>
+        <div class="dialog-actions"><button class="button secondary" value="cancel">Cancel</button><button class="button primary" value="default" type="submit">Save changes</button></div>
+      </form>
+    </dialog>
+
+    <div class="toast-region" id="toast-region" role="status" aria-live="polite" aria-atomic="true"></div>
+    <script type="module" src="/assets/app.mjs?v=0.2.0-r13"></script>
+  </body>
+</html>
diff --git a/server/frontend/insightface-mark.svg b/server/frontend/insightface-mark.svg
new file mode 100644
index 000000000..797afffc0
--- /dev/null
+++ b/server/frontend/insightface-mark.svg
@@ -0,0 +1,7 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" role="img" aria-label="InsightFace">
+  <rect width="24" height="24" rx="6" fill="#6366f1"/>
+  <g fill="none" stroke="#fff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <circle cx="12" cy="8" r="4"/>
+    <path d="M5.5 21a7.5 7.5 0 0 1 13 0"/>
+  </g>
+</svg>
diff --git a/server/frontend/markdown.mjs b/server/frontend/markdown.mjs
new file mode 100644
index 000000000..d98487058
--- /dev/null
+++ b/server/frontend/markdown.mjs
@@ -0,0 +1,124 @@
+function escapeHtml(value) {
+  return String(value)
+    .replaceAll("&", "&amp;")
+    .replaceAll("<", "&lt;")
+    .replaceAll(">", "&gt;")
+    .replaceAll('"', "&quot;")
+    .replaceAll("'", "&#39;");
+}
+
+function safeHref(value) {
+  const href = String(value).trim();
+  if (/^(?:https?:\/\/|\/|#|\.\.?\/)/i.test(href)) return escapeHtml(href);
+  if (/^(?:(?:user-guide|api)(?:\.(?:zh-CN|ja|de|es|fr|ru|pt|ko))?|maintainer-guide)\.md(?:#[A-Za-z0-9_.-]+)?$/.test(href)) {
+    return escapeHtml(href);
+  }
+  return "#";
+}
+
+function safeImageHref(value) {
+  const href = String(value).trim();
+  if (href.startsWith("docs/images/")) return escapeHtml(`/guide-images/${href.slice("docs/images/".length)}`);
+  return safeHref(href);
+}
+
+function inline(value) {
+  let output = escapeHtml(value);
+  output = output.replace(/!\[([^\]]*)\]\(([^)]+)\)/g, (_, label, href) => `<img src="${safeImageHref(href)}" alt="${label}" loading="lazy">`);
+  output = output.replace(/`([^`]+)`/g, "<code>$1</code>");
+  output = output.replace(/\*\*([^*]+)\*\*/g, "<strong>$1</strong>");
+  output = output.replace(/\*([^*]+)\*/g, "<em>$1</em>");
+  output = output.replace(/\[([^\]]+)\]\(([^)]+)\)/g, (_, label, href) => `<a href="${safeHref(href)}" target="_blank" rel="noopener">${label}</a>`);
+  return output;
+}
+
+function isTableSeparator(value) {
+  return /^\s*\|?(?:\s*:?-{3,}:?\s*\|)+\s*:?-{3,}:?\s*\|?\s*$/.test(value);
+}
+
+function tableCells(value) {
+  return value.trim().replace(/^\||\|$/g, "").split("|").map((cell) => cell.trim());
+}
+
+export function renderMarkdown(markdown) {
+  const lines = String(markdown || "").replaceAll("\r\n", "\n").split("\n");
+  const output = [];
+  let index = 0;
+
+  while (index < lines.length) {
+    const line = lines[index];
+    if (!line.trim()) {
+      index += 1;
+      continue;
+    }
+
+    if (line.startsWith("```")) {
+      const language = line.slice(3).trim().replace(/[^A-Za-z0-9_-]/g, "");
+      const code = [];
+      index += 1;
+      while (index < lines.length && !lines[index].startsWith("```")) {
+        code.push(lines[index]);
+        index += 1;
+      }
+      if (index < lines.length) index += 1;
+      output.push(`<pre><code${language ? ` class="language-${language}"` : ""}>${escapeHtml(code.join("\n"))}</code></pre>`);
+      continue;
+    }
+
+    const heading = /^(#{1,6})\s+(.+)$/.exec(line);
+    if (heading) {
+      const level = heading[1].length;
+      output.push(`<h${level}>${inline(heading[2])}</h${level}>`);
+      index += 1;
+      continue;
+    }
+
+    if (index + 1 < lines.length && line.includes("|") && isTableSeparator(lines[index + 1])) {
+      const headers = tableCells(line);
+      index += 2;
+      const rows = [];
+      while (index < lines.length && lines[index].includes("|") && lines[index].trim()) {
+        rows.push(tableCells(lines[index]));
+        index += 1;
+      }
+      output.push(`<div class="markdown-table-wrap"><table><thead><tr>${headers.map((cell) => `<th>${inline(cell)}</th>`).join("")}</tr></thead><tbody>${rows.map((row) => `<tr>${row.map((cell) => `<td>${inline(cell)}</td>`).join("")}</tr>`).join("")}</tbody></table></div>`);
+      continue;
+    }
+
+    const unordered = /^\s*[-*+]\s+(.+)$/.exec(line);
+    const ordered = /^\s*\d+[.)]\s+(.+)$/.exec(line);
+    if (unordered || ordered) {
+      const tag = unordered ? "ul" : "ol";
+      const items = [];
+      while (index < lines.length) {
+        const match = tag === "ul" ? /^\s*[-*+]\s+(.+)$/.exec(lines[index]) : /^\s*\d+[.)]\s+(.+)$/.exec(lines[index]);
+        if (!match) break;
+        items.push(`<li>${inline(match[1])}</li>`);
+        index += 1;
+      }
+      output.push(`<${tag}>${items.join("")}</${tag}>`);
+      continue;
+    }
+
+    if (line.startsWith("> ")) {
+      const quotes = [];
+      while (index < lines.length && lines[index].startsWith("> ")) {
+        quotes.push(lines[index].slice(2));
+        index += 1;
+      }
+      output.push(`<blockquote>${inline(quotes.join(" "))}</blockquote>`);
+      continue;
+    }
+
+    const paragraph = [line.trim()];
+    index += 1;
+    while (index < lines.length && lines[index].trim() && !/^(?:#{1,6}\s|```|\s*[-*+]\s+|\s*\d+[.)]\s+|> )/.test(lines[index])) {
+      if (index + 1 < lines.length && isTableSeparator(lines[index + 1])) break;
+      paragraph.push(lines[index].trim());
+      index += 1;
+    }
+    output.push(`<p>${inline(paragraph.join(" "))}</p>`);
+  }
+
+  return output.join("\n");
+}
diff --git a/server/frontend/openapi.css b/server/frontend/openapi.css
new file mode 100644
index 000000000..eb4af5251
--- /dev/null
+++ b/server/frontend/openapi.css
@@ -0,0 +1,158 @@
+:root {
+  --ink: #172029;
+  --muted: #69767f;
+  --paper: #f4f6f4;
+  --surface: #fff;
+  --line: #dce4e0;
+  --nav: #12251f;
+  --accent: #087f66;
+  --get: #1769aa;
+  --post: #087f66;
+  --patch: #9a6307;
+  --delete: #b43b3b;
+  font-family: Inter, ui-sans-serif, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
+  color: var(--ink);
+  background: var(--paper);
+}
+* { box-sizing: border-box; }
+body { margin: 0; line-height: 1.5; }
+a { color: var(--accent); }
+code, pre { font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; }
+.docs-header { position: sticky; z-index: 10; top: 0; display: flex; align-items: center; justify-content: space-between; min-height: 65px; padding: 0 clamp(18px, 5vw, 72px); color: #ecf7f3; background: rgb(18 37 31 / 96%); border-bottom: 1px solid rgb(255 255 255 / 9%); backdrop-filter: blur(12px); }
+.brand-link { display: flex; align-items: center; gap: 10px; color: inherit; text-decoration: none; }
+.brand-link strong, .brand-link small { display: block; }
+.brand-link strong { font-size: 14px; }
+.brand-link small { color: #97afa6; font-size: 9px; letter-spacing: .1em; text-transform: uppercase; }
+.brand-mark { display: grid; place-items: center; width: 32px; height: 32px; color: #70d9b6; background: rgb(255 255 255 / 7%); border: 1px solid rgb(255 255 255 / 12%); border-radius: 8px; font-size: 10px; font-weight: 900; }
+.header-links { display: flex; align-items: center; gap: 18px; }
+.docs-language select { min-width: 112px; padding: 7px 28px 7px 9px; color: inherit; background: rgb(255 255 255 / 5%); border: 1px solid rgb(255 255 255 / 14%); border-radius: 7px; font: inherit; }
+.docs-language option { color: #172029; background: #fff; }
+.sr-only { position: absolute; width: 1px; height: 1px; padding: 0; margin: -1px; overflow: hidden; clip: rect(0, 0, 0, 0); white-space: nowrap; border: 0; }
+.header-links a { color: #b9cbc4; font-size: 11px; font-weight: 700; text-decoration: none; }
+.header-links a:hover { color: #fff; }
+main { width: min(1450px, 100%); margin: auto; padding: 52px clamp(18px, 5vw, 72px) 72px; }
+.docs-intro { position: relative; overflow: hidden; padding: clamp(28px, 5vw, 52px); color: #eaf8f3; background: linear-gradient(125deg, #163e32, #142923); border-radius: 17px; }
+.docs-intro::after { content: ""; position: absolute; right: -100px; bottom: -210px; width: 390px; height: 390px; border: 1px solid rgb(117 224 190 / 14%); border-radius: 50%; box-shadow: 0 0 0 55px rgb(117 224 190 / 4%), 0 0 0 110px rgb(117 224 190 / 3%); }
+.eyebrow { margin: 0 0 5px; color: #2b9a7e; font-size: 10px; font-weight: 850; letter-spacing: .15em; text-transform: uppercase; }
+.docs-intro .eyebrow { color: #75e0be; }
+h1 { position: relative; z-index: 1; max-width: 720px; margin: 0; font-size: clamp(31px, 5vw, 52px); line-height: 1.05; letter-spacing: -.045em; }
+.docs-intro > p:not(.eyebrow) { position: relative; z-index: 1; max-width: 700px; color: #aac1b8; }
+.docs-intro .guide-callout { margin-top: 18px; padding: 12px 14px; color: #cbd5e1; background: rgb(255 255 255 / 5%); border: 1px solid rgb(255 255 255 / 10%); border-radius: 9px; }
+.docs-intro .guide-callout a { color: #a5b4fc; font-weight: 700; }
+.base-url { position: relative; z-index: 1; display: inline-flex; align-items: center; gap: 12px; margin-top: 12px; padding: 8px 12px; background: rgb(255 255 255 / 7%); border: 1px solid rgb(255 255 255 / 11%); border-radius: 8px; font-size: 11px; }
+.base-url span { color: #9bb3a9; }
+.base-url code { color: #82e8c5; }
+.docs-layout { display: grid; grid-template-columns: 250px minmax(0, 1fr); gap: 24px; margin-top: 34px; align-items: start; }
+.docs-sidebar { position: sticky; top: 88px; max-height: calc(100vh - 110px); padding: 17px; overflow-y: auto; background: var(--surface); border: 1px solid var(--line); border-radius: 12px; }
+.docs-sidebar label { display: grid; gap: 5px; color: var(--muted); font-size: 10px; font-weight: 750; }
+.docs-sidebar input { width: 100%; height: 35px; padding: 0 9px; color: var(--ink); background: #fff; border: 1px solid var(--line); border-radius: 7px; outline: 0; }
+.docs-sidebar input:focus { border-color: var(--accent); box-shadow: 0 0 0 3px rgb(8 127 102 / 10%); }
+#operation-nav { display: grid; gap: 2px; margin-top: 13px; }
+#operation-nav a { display: grid; grid-template-columns: 45px 1fr; gap: 7px; align-items: center; padding: 7px 8px; color: #435049; border-radius: 6px; font-size: 10px; text-decoration: none; }
+#operation-nav a:hover { background: #edf5f1; }
+#operation-nav b { font-size: 8px; letter-spacing: .05em; }
+#operation-nav span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.operations { display: grid; gap: 13px; min-width: 0; }
+.operation { overflow: hidden; background: var(--surface); border: 1px solid var(--line); border-radius: 12px; }
+.operation[hidden] { display: none; }
+.operation summary { display: grid; grid-template-columns: 64px minmax(180px, auto) 1fr; gap: 12px; align-items: center; padding: 16px 18px; cursor: pointer; list-style: none; }
+.operation summary::-webkit-details-marker { display: none; }
+.method { display: inline-grid; place-items: center; width: 57px; height: 25px; color: #fff; border-radius: 6px; font-size: 9px; font-weight: 900; letter-spacing: .06em; }
+.method.get { background: var(--get); }.method.post { background: var(--post); }.method.patch { background: var(--patch); }.method.delete { background: var(--delete); }
+.operation-path { color: #26322d; font-size: 12px; font-weight: 750; overflow-wrap: anywhere; }
+.operation-summary { color: var(--muted); font-size: 11px; }
+.operation-body { padding: 3px 18px 20px 94px; border-top: 1px solid #edf0ee; }
+.operation-description { color: var(--muted); font-size: 12px; }
+.operation-section { margin-top: 18px; }
+.operation-section h3 { margin: 0 0 8px; font-size: 11px; letter-spacing: .08em; text-transform: uppercase; }
+.parameter { display: grid; grid-template-columns: 170px 1fr; gap: 14px; padding: 9px 0; border-bottom: 1px solid #edf0ee; font-size: 11px; }
+.parameter:last-child { border: 0; }
+.parameter-name { font-weight: 800; overflow-wrap: anywhere; }
+.required { margin-left: 5px; color: var(--delete); font-size: 8px; text-transform: uppercase; }
+.parameter-meta { color: var(--muted); }
+.code-block { max-height: 380px; margin: 0; padding: 13px; overflow: auto; color: #d9eae3; background: #172620; border-radius: 8px; font-size: 10px; line-height: 1.55; }
+.loading { display: grid; place-items: center; min-height: 260px; color: var(--muted); font-size: 12px; }
+.loading span { width: 24px; height: 24px; border: 2px solid var(--line); border-top-color: var(--accent); border-radius: 50%; animation: spin .8s linear infinite; }
+@keyframes spin { to { transform: rotate(360deg); } }
+.error { padding: 25px; color: #812e2e; background: #fde9e8; border: 1px solid #f1c9c7; border-radius: 12px; }
+.schema-section { margin-top: 54px; }
+.section-heading { display: flex; align-items: end; justify-content: space-between; gap: 20px; margin-bottom: 17px; }
+.section-heading h2 { margin: 0; font-size: 27px; letter-spacing: -.035em; }
+.section-heading > span { color: var(--muted); font-size: 11px; }
+.schema-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(310px, 1fr)); gap: 12px; }
+.schema-card { min-width: 0; background: var(--surface); border: 1px solid var(--line); border-radius: 10px; }
+.schema-card summary { padding: 13px 15px; cursor: pointer; font: 750 11px ui-monospace, SFMono-Regular, Menlo, monospace; overflow-wrap: anywhere; }
+.schema-card pre { max-height: 340px; margin: 0 10px 10px; padding: 11px; overflow: auto; color: #d9eae3; background: #172620; border-radius: 7px; font-size: 9px; }
+footer { padding: 24px; color: var(--muted); background: #e8edea; border-top: 1px solid var(--line); text-align: center; font-size: 10px; }
+@media (max-width: 830px) { .docs-layout { grid-template-columns: 1fr; }.docs-sidebar { position: static; max-height: none; }.operation summary { grid-template-columns: 58px 1fr; }.operation-summary { grid-column: 2; }.operation-body { padding-left: 18px; } }
+@media (max-width: 520px) { .header-links a:last-child { display: none; }.docs-header { padding-inline: 14px; }.operation summary { grid-template-columns: 1fr; }.operation-summary { grid-column: auto; }.parameter { grid-template-columns: 1fr; gap: 2px; }.schema-grid { grid-template-columns: 1fr; } }
+@media (prefers-reduced-motion: reduce) { * { animation-duration: .001ms !important; } }
+
+/* InsightFace.ai visual identity. */
+:root {
+  --ink: #fafafa;
+  --muted: #a1a1aa;
+  --paper: #09090b;
+  --surface: #111113;
+  --line: #27272a;
+  --nav: #09090b;
+  --accent: #6366f1;
+  --get: #22d3ee;
+  --post: #6366f1;
+  --patch: #d97706;
+  --delete: #ef4444;
+  font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif;
+  color-scheme: dark;
+}
+html,
+body { background: var(--paper); }
+body {
+  color: var(--ink);
+  background-image:
+    radial-gradient(circle at 75% -10%, rgb(99 102 241 / 12%), transparent 34rem),
+    linear-gradient(rgb(99 102 241 / 3%) 1px, transparent 1px),
+    linear-gradient(90deg, rgb(99 102 241 / 3%) 1px, transparent 1px);
+  background-size: auto, 64px 64px, 64px 64px;
+  background-attachment: fixed;
+}
+::selection { color: #fff; background: rgb(99 102 241 / 55%); }
+.docs-header { color: var(--ink); background: rgb(9 9 11 / 80%); border-bottom-color: rgb(39 39 42 / 78%); backdrop-filter: blur(20px); }
+.brand-link strong { font-size: 15px; font-weight: 650; }
+.brand-link small { color: var(--muted); }
+.brand-mark { width: 32px; height: 32px; color: #fff; background: #6366f1; border: 0; border-radius: 8px; box-shadow: 0 0 22px rgb(99 102 241 / 24%); }
+.brand-mark svg { width: 20px; height: 20px; stroke: #fff; stroke-width: 2; }
+.header-links a { color: var(--muted); }
+.header-links a:hover { color: var(--ink); }
+.docs-intro {
+  color: var(--ink);
+  background:
+    linear-gradient(rgb(99 102 241 / 5%) 1px, transparent 1px),
+    linear-gradient(90deg, rgb(99 102 241 / 5%) 1px, transparent 1px),
+    radial-gradient(circle at 82% 10%, rgb(99 102 241 / 20%), transparent 42%),
+    linear-gradient(135deg, #111113, #0c0c10);
+  background-size: 48px 48px, 48px 48px, auto, auto;
+  border: 1px solid var(--line);
+  border-radius: 20px;
+  box-shadow: 0 0 80px -28px rgb(99 102 241 / 42%);
+}
+.docs-intro::after { background: rgb(99 102 241 / 8%); border-color: rgb(99 102 241 / 16%); box-shadow: 0 0 100px rgb(99 102 241 / 13%); }
+.eyebrow,
+.docs-intro .eyebrow { color: #818cf8; }
+.docs-intro > p:not(.eyebrow) { color: var(--muted); }
+.base-url { background: rgb(24 24 27 / 82%); border-color: var(--line); }
+.base-url span { color: var(--muted); }
+.base-url code { color: #22d3ee; }
+.docs-sidebar,
+.operation,
+.schema-card { background: rgb(17 17 19 / 95%); border-color: var(--line); box-shadow: 0 16px 42px rgb(0 0 0 / 12%); }
+.docs-sidebar input { color: var(--ink); background: #18181b; border-color: #3f3f46; }
+.docs-sidebar input:focus { border-color: #6366f1; box-shadow: 0 0 0 3px rgb(99 102 241 / 16%); }
+#operation-nav a { color: #d4d4d8; }
+#operation-nav a:hover { color: #fff; background: rgb(99 102 241 / 10%); }
+.operation-path { color: #f4f4f5; }
+.operation-body,
+.parameter { border-color: var(--line); }
+.code-block,
+.schema-card pre { color: #d4d4d8; background: #09090b; border: 1px solid var(--line); }
+.error { color: #fecaca; background: rgb(239 68 68 / 10%); border-color: rgb(239 68 68 / 25%); }
+footer { color: var(--muted); background: #0d0d0f; border-color: var(--line); }
diff --git a/server/frontend/openapi.html b/server/frontend/openapi.html
new file mode 100644
index 000000000..70d7c38da
--- /dev/null
+++ b/server/frontend/openapi.html
@@ -0,0 +1,53 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="utf-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1">
+    <meta name="color-scheme" content="dark">
+    <meta name="theme-color" content="#09090b">
+    <meta http-equiv="Content-Security-Policy" content="default-src 'self'; script-src 'self'; style-src 'self'; connect-src 'self'; img-src 'self' data:; object-src 'none'; base-uri 'none'; form-action 'none'">
+    <meta name="description" content="Live InsightFace Server OpenAPI schema explorer.">
+    <title>OpenAPI Schema · InsightFace Server</title>
+    <link rel="icon" href="/assets/insightface-mark.svg" type="image/svg+xml">
+    <link rel="stylesheet" href="/assets/openapi.css">
+  </head>
+  <body>
+    <header class="docs-header">
+      <a class="brand-link" href="/">
+        <span class="brand-mark" aria-hidden="true"><svg viewBox="0 0 24 24" fill="none"><circle cx="12" cy="8" r="4"></circle><path d="M5.5 21a7.5 7.5 0 0 1 13 0"></path></svg></span>
+        <span><strong>InsightFace Server</strong><small>OpenAPI Schema</small></span>
+      </a>
+      <div class="header-links">
+        <label class="docs-language" title="Language"><span class="sr-only">Language</span><select id="language-select" aria-label="Language"><option value="en">English</option><option value="zh">中文</option><option value="ja">日本語</option><option value="de">Deutsch</option><option value="es">Español</option><option value="fr">Français</option><option value="ru">Русский</option><option value="pt">Português</option><option value="ko">한국어</option></select></label>
+        <a href="/#help">Help</a>
+        <a href="/">Open console</a><a href="/openapi.json" target="_blank" rel="noopener">OpenAPI JSON</a>
+      </div>
+    </header>
+    <main>
+      <section class="docs-intro">
+        <p class="eyebrow">DEVELOPER TOOL · OPENAPI 3</p>
+        <h1>Live API schema.</h1>
+        <p id="api-description">Loading the server's OpenAPI document…</p>
+        <p class="guide-callout">New to the server? Follow the <a href="/#help">step-by-step user guide</a> before using this live schema reference.</p>
+        <div class="base-url"><span>Base path</span><code id="base-url">/v1</code></div>
+      </section>
+
+      <div class="docs-layout">
+        <aside class="docs-sidebar" aria-label="API operations">
+          <label><span>Filter operations</span><input id="operation-filter" type="search" placeholder="detect, collections…"></label>
+          <nav id="operation-nav"></nav>
+        </aside>
+        <section class="operations" id="operations" aria-live="polite">
+          <div class="loading"><span></span><strong>Loading API schema</strong></div>
+        </section>
+      </div>
+
+      <section class="schema-section">
+        <div class="section-heading"><div><p class="eyebrow">COMPONENTS</p><h2>Data schemas</h2></div><span id="schema-count"></span></div>
+        <div class="schema-grid" id="schema-grid"></div>
+      </section>
+    </main>
+    <footer>Generated locally from <code>/openapi.json</code>. No external scripts, fonts, or services are used.</footer>
+    <script type="module" src="/assets/openapi.js?v=0.2.0-r13"></script>
+  </body>
+</html>
diff --git a/server/frontend/openapi.js b/server/frontend/openapi.js
new file mode 100644
index 000000000..19f3abb38
--- /dev/null
+++ b/server/frontend/openapi.js
@@ -0,0 +1,155 @@
+import { initializeI18n, t, translateTree } from "./i18n.mjs?v=0.2.0-r13";
+
+const operationsRoot = document.querySelector("#operations");
+const navigationRoot = document.querySelector("#operation-nav");
+const filterInput = document.querySelector("#operation-filter");
+const schemaRoot = document.querySelector("#schema-grid");
+const methods = ["get", "post", "put", "patch", "delete"];
+let specification;
+
+function element(tag, className, text) {
+  const node = document.createElement(tag);
+  if (className) node.className = className;
+  if (text !== undefined) node.textContent = t(text);
+  return node;
+}
+
+function safeId(method, path) {
+  return `${method}-${path}`.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-|-$/g, "");
+}
+
+function compactSchema(schema) {
+  if (!schema) return {};
+  const clone = structuredClone(schema);
+  return clone;
+}
+
+function schemaLabel(schema) {
+  if (!schema) return "unknown";
+  if (schema.$ref) return schema.$ref.split("/").at(-1);
+  if (schema.type === "array") return `array<${schemaLabel(schema.items)}>`;
+  return schema.type || schema.format || "object";
+}
+
+function addParameterRows(container, parameters) {
+  if (!parameters.length) return;
+  const section = element("section", "operation-section");
+  section.append(element("h3", "", "Parameters"));
+  for (const parameter of parameters) {
+    const row = element("div", "parameter");
+    const name = element("div", "parameter-name", parameter.name);
+    if (parameter.required) name.append(element("span", "required", "required"));
+    const metadata = element("div", "parameter-meta");
+    metadata.textContent = `${parameter.in} · ${schemaLabel(parameter.schema)}`;
+    if (parameter.description) metadata.append(document.createElement("br"), document.createTextNode(t(parameter.description)));
+    row.append(name, metadata);
+    section.append(row);
+  }
+  container.append(section);
+}
+
+function addRequestBody(container, requestBody) {
+  if (!requestBody) return;
+  const section = element("section", "operation-section");
+  const heading = element("h3", "", "Request body");
+  if (requestBody.required) heading.append(element("span", "required", "required"));
+  section.append(heading);
+  for (const [contentType, media] of Object.entries(requestBody.content || {})) {
+    section.append(element("p", "parameter-meta", contentType));
+    const pre = element("pre", "code-block", JSON.stringify(compactSchema(media.schema), null, 2));
+    section.append(pre);
+  }
+  container.append(section);
+}
+
+function addResponses(container, responses) {
+  const section = element("section", "operation-section");
+  section.append(element("h3", "", "Responses"));
+  for (const [status, response] of Object.entries(responses || {})) {
+    const row = element("div", "parameter");
+    row.append(element("div", "parameter-name", status), element("div", "parameter-meta", response.description || "Response"));
+    section.append(row);
+  }
+  container.append(section);
+}
+
+function renderOperation(method, path, operation, inheritedParameters) {
+  const id = safeId(method, path);
+  const details = element("details", "operation");
+  details.id = id;
+  details.dataset.search = `${method} ${path} ${operation.summary || ""} ${(operation.tags || []).join(" ")}`.toLowerCase();
+  const summary = document.createElement("summary");
+  summary.append(element("span", `method ${method}`, method.toUpperCase()), element("code", "operation-path", path), element("span", "operation-summary", operation.summary || operation.operationId || ""));
+  details.append(summary);
+  const body = element("div", "operation-body");
+  if (operation.description) body.append(element("p", "operation-description", operation.description));
+  addParameterRows(body, [...(inheritedParameters || []), ...(operation.parameters || [])]);
+  addRequestBody(body, operation.requestBody);
+  addResponses(body, operation.responses);
+  details.append(body);
+  operationsRoot.append(details);
+
+  const link = document.createElement("a");
+  link.href = `#${id}`;
+  link.dataset.search = details.dataset.search;
+  link.append(element("b", "", method.toUpperCase()), element("span", "", operation.summary || path));
+  link.addEventListener("click", () => { details.open = true; });
+  navigationRoot.append(link);
+}
+
+function renderSchemas(schemas) {
+  schemaRoot.replaceChildren();
+  const entries = Object.entries(schemas || {}).sort(([left], [right]) => left.localeCompare(right));
+  document.querySelector("#schema-count").textContent = t("{count} schemas", { count: entries.length });
+  for (const [name, schema] of entries) {
+    const details = element("details", "schema-card");
+    const summary = document.createElement("summary");
+    summary.textContent = name;
+    details.append(summary, element("pre", "", JSON.stringify(schema, null, 2)));
+    schemaRoot.append(details);
+  }
+}
+
+function renderDocument(specification) {
+  operationsRoot.replaceChildren();
+  navigationRoot.replaceChildren();
+  document.querySelector("#api-description").textContent = specification.info?.description
+    ? t(specification.info.description)
+    : `${specification.info?.title || "InsightFace Server"} ${specification.info?.version || ""}`;
+  const server = specification.servers?.[0]?.url;
+  document.querySelector("#base-url").textContent = server || "/v1";
+  for (const [path, pathItem] of Object.entries(specification.paths || {})) {
+    for (const method of methods) {
+      if (pathItem[method]) renderOperation(method, path, pathItem[method], pathItem.parameters);
+    }
+  }
+  if (!operationsRoot.children.length) operationsRoot.append(element("div", "error", "The OpenAPI document contains no operations."));
+  renderSchemas(specification.components?.schemas);
+}
+
+function applyFilter() {
+  const query = filterInput.value.trim().toLowerCase();
+  document.querySelectorAll(".operation").forEach((node) => { node.hidden = Boolean(query) && !node.dataset.search.includes(query); });
+  navigationRoot.querySelectorAll("a").forEach((node) => { node.hidden = Boolean(query) && !node.dataset.search.includes(query); });
+}
+
+filterInput.addEventListener("input", applyFilter);
+window.addEventListener("insightface:localechange", () => {
+  translateTree(document);
+  document.title = `${t("OpenAPI Schema")} · InsightFace Server`;
+  if (specification) renderDocument(specification);
+  applyFilter();
+});
+
+initializeI18n();
+translateTree(document);
+document.title = `${t("OpenAPI Schema")} · InsightFace Server`;
+
+try {
+  const response = await fetch("/openapi.json", { headers: { Accept: "application/json" } });
+  if (!response.ok) throw new Error(`HTTP ${response.status}`);
+  specification = await response.json();
+  renderDocument(specification);
+} catch (error) {
+  operationsRoot.replaceChildren(element("div", "error", t("Could not load OpenAPI document: {message}", { message: error.message })));
+}
diff --git a/server/frontend/styles.css b/server/frontend/styles.css
new file mode 100644
index 000000000..43a41a984
--- /dev/null
+++ b/server/frontend/styles.css
@@ -0,0 +1,714 @@
+:root {
+  --ink: #172029;
+  --muted: #68747f;
+  --faint: #8a959e;
+  --paper: #f4f6f4;
+  --surface: #ffffff;
+  --line: #dfe5e2;
+  --line-strong: #cad4d0;
+  --nav: #12251f;
+  --nav-soft: #1c3930;
+  --accent: #087f66;
+  --accent-dark: #05654f;
+  --accent-soft: #ddf4ed;
+  --warning: #a25d08;
+  --warning-soft: #fff0d9;
+  --danger: #b43b3b;
+  --danger-soft: #fde8e7;
+  --success: #087f66;
+  --shadow: 0 1px 2px rgb(18 37 31 / 5%), 0 10px 32px rgb(18 37 31 / 5%);
+  --radius: 14px;
+  --radius-small: 9px;
+  font-family: Inter, ui-sans-serif, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
+  color: var(--ink);
+  background: var(--paper);
+  font-synthesis: none;
+}
+
+* { box-sizing: border-box; }
+html { min-width: 320px; background: var(--paper); }
+body { margin: 0; min-height: 100vh; background: var(--paper); line-height: 1.5; }
+button, input, textarea, select { font: inherit; }
+button, select { cursor: pointer; }
+button:disabled { cursor: not-allowed; opacity: .55; }
+img, canvas, video { max-width: 100%; }
+[hidden] { display: none !important; }
+
+:focus-visible { outline: 3px solid rgb(8 127 102 / 30%); outline-offset: 2px; }
+.skip-link { position: fixed; z-index: 1000; top: 10px; left: 10px; padding: 8px 14px; background: var(--surface); color: var(--ink); border-radius: 8px; transform: translateY(-160%); }
+.skip-link:focus { transform: none; }
+.sr-only { position: absolute; width: 1px; height: 1px; padding: 0; margin: -1px; overflow: hidden; clip: rect(0, 0, 0, 0); white-space: nowrap; border: 0; }
+.muted { color: var(--muted); }
+.tiny { font-size: 11px; }
+.optional { color: var(--faint); font-size: 12px; font-weight: 500; }
+
+.app-shell { min-height: 100vh; }
+.sidebar {
+  position: fixed;
+  inset: 0 auto 0 0;
+  width: 248px;
+  z-index: 20;
+  display: flex;
+  flex-direction: column;
+  padding: 24px 16px 18px;
+  color: #f2f8f5;
+  background:
+    radial-gradient(circle at 16px 0, rgb(43 112 90 / 28%), transparent 210px),
+    var(--nav);
+}
+.brand { display: flex; align-items: center; gap: 12px; padding: 0 8px 22px; }
+.brand strong { display: block; font-size: 17px; letter-spacing: -.02em; }
+.brand small { display: block; color: #9db5ab; font-size: 11px; letter-spacing: .07em; text-transform: uppercase; }
+.brand-mark { display: grid; place-items: center; width: 37px; height: 37px; flex: 0 0 auto; border: 1px solid rgb(255 255 255 / 18%); border-radius: 11px; background: rgb(255 255 255 / 8%); }
+.brand-mark::before, .brand-mark::after, .brand-mark span { content: ""; position: absolute; border: 1.5px solid #75e0be; border-radius: 50%; }
+.brand-mark { position: relative; }
+.brand-mark::before { width: 20px; height: 12px; }
+.brand-mark::after { width: 12px; height: 20px; }
+.brand-mark span { width: 4px; height: 4px; background: #75e0be; }
+
+.server-chip { display: flex; align-items: center; gap: 8px; margin: 0 5px 18px; padding: 8px 11px; color: #b6c9c1; background: rgb(255 255 255 / 5%); border: 1px solid rgb(255 255 255 / 8%); border-radius: 9px; font-size: 12px; }
+.status-dot { width: 7px; height: 7px; border-radius: 50%; background: #aeb6b3; box-shadow: 0 0 0 3px rgb(174 182 179 / 12%); }
+.server-chip.healthy .status-dot { background: #59d6ac; box-shadow: 0 0 0 3px rgb(89 214 172 / 14%); }
+.server-chip.error .status-dot { background: #ff7c71; box-shadow: 0 0 0 3px rgb(255 124 113 / 14%); }
+
+.nav-list { display: grid; gap: 3px; overflow-y: auto; }
+.nav-item { display: flex; align-items: center; gap: 11px; width: 100%; min-height: 42px; padding: 0 12px; color: #b7c9c2; background: transparent; border: 0; border-radius: 9px; text-align: left; text-decoration: none; font-size: 13px; font-weight: 600; transition: background .15s, color .15s; }
+.nav-item:hover { color: #fff; background: rgb(255 255 255 / 6%); }
+.nav-item.active { color: #fff; background: var(--nav-soft); box-shadow: inset 3px 0 #53d0a6; }
+.nav-icon { display: inline-grid; place-items: center; width: 19px; color: #7dd4b5; font-size: 16px; font-weight: 400; }
+.sidebar-footer { margin-top: auto; padding: 18px 5px 0; }
+.sidebar-footer p { margin: 10px 3px 0; color: #8da49b; line-height: 1.45; }
+
+.main-content { min-height: 100vh; margin-left: 248px; }
+.topbar { position: sticky; z-index: 10; top: 0; display: flex; align-items: center; gap: 16px; min-height: 78px; padding: 13px clamp(22px, 4vw, 52px); background: rgb(244 246 244 / 90%); border-bottom: 1px solid rgb(202 212 208 / 65%); backdrop-filter: blur(14px); }
+.topbar h1 { margin: 0; font-size: 23px; line-height: 1.2; letter-spacing: -.03em; }
+.eyebrow { margin: 0 0 3px; color: var(--accent); font-size: 10px; font-weight: 800; letter-spacing: .14em; text-transform: uppercase; }
+.topbar-actions { display: flex; align-items: center; gap: 10px; margin-left: auto; }
+.auth-state { display: inline-flex; align-items: center; gap: 7px; color: var(--muted); font-size: 12px; }
+.auth-state.configured { color: var(--accent); font-weight: 700; }
+.menu-button { display: none; width: 40px; height: 40px; border: 1px solid var(--line); background: var(--surface); border-radius: 9px; }
+
+.page-stack { width: min(1520px, 100%); margin: 0 auto; padding: 34px clamp(22px, 4vw, 52px) 64px; }
+.page { animation: page-in .18s ease-out; }
+@keyframes page-in { from { opacity: 0; transform: translateY(3px); } to { opacity: 1; transform: none; } }
+.page > * + * { margin-top: 20px; }
+.page h2 { margin: 0; font-size: clamp(23px, 2.2vw, 32px); line-height: 1.15; letter-spacing: -.035em; }
+.page h3 { margin: 0; font-size: 16px; letter-spacing: -.015em; }
+.section-heading { display: flex; align-items: end; justify-content: space-between; gap: 24px; margin-bottom: 24px; }
+.section-heading p:last-child { max-width: 720px; margin: 8px 0 0; color: var(--muted); font-size: 14px; }
+
+.hero-panel { position: relative; overflow: hidden; display: flex; justify-content: space-between; gap: 32px; min-height: 178px; padding: clamp(28px, 4vw, 46px); color: #effaf6; background: linear-gradient(123deg, #153d31, #132b25); border-radius: 18px; box-shadow: var(--shadow); }
+.hero-panel::after { content: ""; position: absolute; right: -70px; bottom: -125px; width: 310px; height: 310px; border: 1px solid rgb(117 224 190 / 15%); border-radius: 50%; box-shadow: 0 0 0 45px rgb(117 224 190 / 4%), 0 0 0 92px rgb(117 224 190 / 3%); }
+.hero-panel .eyebrow { color: #75e0be; }
+.hero-panel h2 { position: relative; z-index: 1; max-width: 620px; font-size: clamp(26px, 3vw, 41px); }
+.hero-panel p:last-child { position: relative; z-index: 1; max-width: 610px; margin: 13px 0 0; color: #a9c1b8; }
+.hero-provider { position: relative; z-index: 1; align-self: center; min-width: 230px; padding: 17px 20px; background: rgb(255 255 255 / 7%); border: 1px solid rgb(255 255 255 / 11%); border-radius: 12px; }
+.hero-provider span { display: block; color: #a9c1b8; font-size: 11px; letter-spacing: .08em; text-transform: uppercase; }
+.hero-provider strong { display: block; margin-top: 5px; font-size: 15px; overflow-wrap: anywhere; }
+
+.metric-grid { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 14px; }
+.metric-card { min-width: 0; padding: 20px 21px; background: var(--surface); border: 1px solid var(--line); border-radius: var(--radius); box-shadow: var(--shadow); }
+.metric-card > span { display: block; color: var(--muted); font-size: 12px; font-weight: 650; }
+.metric-card > strong { display: block; margin: 8px 0 2px; font-size: 30px; line-height: 1.1; letter-spacing: -.035em; }
+.metric-card > strong.metric-text { overflow: hidden; font-size: 16px; line-height: 2.05; text-overflow: ellipsis; white-space: nowrap; }
+.metric-card > small { display: block; overflow: hidden; color: var(--faint); font-size: 11px; text-overflow: ellipsis; white-space: nowrap; }
+.system-metrics { margin-bottom: 20px; }
+
+.two-column { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 18px; }
+.panel { min-width: 0; padding: 21px; background: var(--surface); border: 1px solid var(--line); border-radius: var(--radius); box-shadow: var(--shadow); }
+.panel.slim { padding: 15px 18px; }
+.panel-heading { display: flex; align-items: center; justify-content: space-between; gap: 16px; margin-bottom: 17px; }
+.panel-heading > .muted { font-size: 12px; white-space: nowrap; }
+.panel-heading .eyebrow { margin-bottom: 2px; }
+.status-list, .diagnostic-list { margin: 0; }
+.status-list > div, .diagnostic-list > div { display: flex; align-items: start; justify-content: space-between; gap: 18px; padding: 11px 0; border-bottom: 1px solid #edf0ee; }
+.status-list > div:last-child, .diagnostic-list > div:last-child { border-bottom: 0; }
+.status-list dt, .diagnostic-list dt { color: var(--muted); font-size: 12px; }
+.status-list dd, .diagnostic-list dd { margin: 0; max-width: 68%; font-size: 12px; font-weight: 700; text-align: right; overflow-wrap: anywhere; }
+.diagnostic-list dd.good { color: var(--success); }
+.diagnostic-list dd.bad { color: var(--danger); }
+
+.button { display: inline-flex; align-items: center; justify-content: center; gap: 7px; min-height: 39px; padding: 0 15px; border: 1px solid transparent; border-radius: var(--radius-small); font-size: 12px; font-weight: 750; text-decoration: none; transition: transform .1s, box-shadow .15s, background .15s; }
+.button:active:not(:disabled) { transform: translateY(1px); }
+.button.primary { color: #fff; background: var(--accent); box-shadow: 0 4px 12px rgb(8 127 102 / 16%); }
+.button.primary:hover { background: var(--accent-dark); }
+.button.secondary { color: var(--ink); background: var(--surface); border-color: var(--line-strong); }
+.sidebar .button.secondary { color: #e3efea; background: rgb(255 255 255 / 7%); border-color: rgb(255 255 255 / 13%); }
+.button.danger { color: var(--danger); background: var(--danger-soft); border-color: #f3cbc8; }
+.button.small { min-height: 31px; padding: 0 10px; }
+.full-width { width: 100%; }
+.icon-button { display: inline-grid; place-items: center; width: 36px; height: 36px; padding: 0; color: var(--muted); background: var(--surface); border: 1px solid var(--line); border-radius: 9px; font-size: 19px; }
+.icon-button:hover { color: var(--ink); border-color: var(--line-strong); }
+.text-button { padding: 0; color: var(--accent); background: none; border: 0; font-size: 12px; font-weight: 700; }
+
+label { display: grid; gap: 6px; color: #39434b; font-size: 12px; font-weight: 700; }
+input, textarea, select { width: 100%; color: var(--ink); background: #fff; border: 1px solid var(--line-strong); border-radius: 8px; outline: none; }
+input, select { height: 39px; padding: 0 11px; }
+textarea { min-height: 68px; padding: 9px 11px; resize: vertical; }
+input:hover, textarea:hover, select:hover { border-color: #a9bbb4; }
+input:focus, textarea:focus, select:focus { border-color: var(--accent); box-shadow: 0 0 0 3px rgb(8 127 102 / 10%); }
+label small, .hint { margin: 0; color: var(--faint); font-size: 10px; font-weight: 500; }
+.embedding-contract-summary code { overflow-wrap: anywhere; user-select: all; }
+.checkbox-field { grid-template-columns: auto 1fr; align-items: center; justify-content: start; }
+.checkbox-field input[type="checkbox"] { width: 17px; height: 17px; padding: 0; accent-color: var(--accent); }
+.checkbox-field small { grid-column: 2; }
+input[type="range"] { height: 22px; padding: 0; accent-color: var(--accent); border: 0; box-shadow: none; }
+output { float: right; color: var(--accent); font-weight: 800; }
+.form-grid { display: grid; gap: 15px; }
+.form-grid > label {
+  align-self: start;
+  grid-template-rows: minmax(2.65em, auto) auto auto;
+  align-content: start;
+}
+.form-grid.two { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+.form-grid.three { grid-template-columns: repeat(3, minmax(0, 1fr)); }
+.search-parameters { align-items: end; }
+.field-caption { display: flex; align-items: baseline; align-self: end; justify-content: space-between; gap: 6px; min-width: 0; }
+.field-caption .optional { flex: 0 1 auto; font-size: 9px; line-height: 1.15; text-align: right; }
+.form-panel { display: grid; gap: 15px; margin-bottom: 20px; border-color: #c3ded5; }
+.form-actions { display: flex; align-items: center; gap: 14px; }
+.form-status { color: var(--muted); font-size: 12px; }
+.form-status.error { color: var(--danger); }
+.toolbar { display: flex; align-items: end; gap: 14px; margin-bottom: 18px; }
+.toolbar label { min-width: 220px; }
+.toolbar .grow { flex: 1; }
+
+.drop-zone { position: relative; display: grid; place-items: center; gap: 3px; min-height: 145px; padding: 22px; color: var(--muted); background: #fafcfb; border: 1.5px dashed #b9cbc4; border-radius: 11px; text-align: center; transition: background .15s, border-color .15s; }
+.drop-zone:hover, .drop-zone.dragging { background: var(--accent-soft); border-color: var(--accent); }
+.drop-zone input[type="file"] { position: absolute; inset: 0; width: 100%; height: 100%; padding: 0; opacity: 0; cursor: pointer; }
+.drop-zone strong { color: var(--ink); font-size: 13px; }
+.drop-zone > span:not(.drop-icon) { font-size: 11px; font-weight: 500; }
+.drop-zone .file-summary { max-width: 100%; margin-top: 7px; color: var(--accent); overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.drop-icon { display: grid; place-items: center; width: 31px; height: 31px; margin-bottom: 4px; color: var(--accent); background: var(--accent-soft); border-radius: 50%; font-size: 18px; }
+.compact-drop { min-height: 106px; padding: 15px; }
+
+.table-toolbar { display: flex; align-items: center; justify-content: space-between; gap: 16px; margin-bottom: 13px; }
+.search-field { width: min(360px, 100%); }
+.table-wrap { overflow-x: auto; }
+table { width: 100%; border-collapse: collapse; }
+th { padding: 9px 11px; color: var(--faint); border-bottom: 1px solid var(--line); font-size: 10px; letter-spacing: .08em; text-align: left; text-transform: uppercase; white-space: nowrap; }
+td { padding: 13px 11px; border-bottom: 1px solid #edf0ee; font-size: 12px; vertical-align: middle; }
+tbody tr:last-child td { border-bottom: 0; }
+tbody tr:hover { background: #fafcfb; }
+.table-title { display: block; color: var(--ink); font-weight: 750; }
+.table-subtitle { display: block; max-width: 260px; color: var(--faint); font-size: 10px; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.table-actions { display: flex; justify-content: end; gap: 6px; white-space: nowrap; }
+.badge { display: inline-flex; align-items: center; padding: 3px 7px; color: var(--accent-dark); background: var(--accent-soft); border-radius: 999px; font-size: 10px; font-weight: 800; }
+
+.empty-state { display: grid; place-items: center; min-height: 160px; padding: 24px; color: var(--muted); text-align: center; }
+.empty-state strong { color: var(--ink); font-size: 14px; }
+.empty-state span { margin-top: 4px; font-size: 11px; }
+.empty-state.compact { min-height: 160px; }
+.error-entry { padding: 10px 0; border-bottom: 1px solid #edf0ee; }
+.error-entry:last-child { border: 0; }
+.error-entry strong { display: block; color: var(--danger); font-size: 11px; }
+.error-entry span { display: block; margin-top: 3px; color: var(--muted); font-size: 11px; }
+.error-entry time { color: var(--faint); font-size: 9px; }
+
+.people-layout { display: grid; grid-template-columns: minmax(300px, .85fr) minmax(390px, 1.4fr); gap: 18px; }
+.people-list-panel, .detail-panel { min-height: 440px; }
+.person-list { display: grid; gap: 4px; max-height: 650px; overflow-y: auto; }
+.person-row { display: flex; align-items: center; gap: 12px; width: 100%; padding: 11px; color: var(--ink); background: transparent; border: 1px solid transparent; border-radius: 9px; text-align: left; }
+.person-row:hover { background: #f6f9f7; }
+.person-row.selected { background: var(--accent-soft); border-color: #bee1d5; }
+.avatar { display: grid; place-items: center; width: 36px; height: 36px; flex: 0 0 auto; color: var(--accent-dark); background: #d9eee7; border-radius: 50%; font-size: 12px; font-weight: 850; text-transform: uppercase; }
+.person-row > span:nth-child(2) { min-width: 0; flex: 1; }
+.person-row strong, .person-row small { display: block; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.person-row strong { font-size: 12px; }
+.person-row small { color: var(--faint); font-size: 10px; }
+.face-count { color: var(--muted); font-size: 10px; white-space: nowrap; }
+.person-detail-header { display: flex; align-items: start; justify-content: space-between; gap: 16px; padding-bottom: 17px; border-bottom: 1px solid var(--line); }
+.person-detail-header h3 { font-size: 20px; }
+.person-detail-header p { margin: 3px 0 0; color: var(--muted); font-size: 11px; }
+.person-detail-body { display: grid; gap: 17px; padding-top: 17px; }
+.person-metadata { margin: 0; padding: 12px; overflow-x: auto; color: #43504a; background: #f6f8f7; border-radius: 8px; font: 11px/1.55 ui-monospace, SFMono-Regular, Menlo, monospace; }
+.face-sample-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(170px, 1fr)); gap: 10px; }
+.face-card { padding: 12px; background: #f8faf9; border: 1px solid var(--line); border-radius: 9px; }
+.face-card img { display: block; width: 112px; height: 112px; margin: 8px auto; object-fit: cover; border-radius: 8px; }
+.face-card img[hidden] { display: none; }
+.face-card strong { display: block; font-size: 11px; overflow: hidden; text-overflow: ellipsis; }
+.face-card dl { margin: 8px 0; font-size: 10px; }
+.face-card dl > div { display: flex; justify-content: space-between; gap: 8px; }
+.face-card dt { color: var(--faint); }
+.face-card dd { margin: 0; font-weight: 700; }
+.rejection-list { padding: 12px; color: #744716; background: var(--warning-soft); border: 1px solid #f2d4a7; border-radius: 9px; }
+.rejection-item { display: flex; justify-content: space-between; gap: 12px; padding: 5px 0; font-size: 11px; }
+
+.tool-layout { display: grid; grid-template-columns: minmax(260px, 340px) minmax(0, 1fr); gap: 18px; }
+.tool-controls { display: flex; flex-direction: column; gap: 16px; align-self: start; }
+.media-panel { min-height: 450px; }
+.canvas-stage { position: relative; display: grid; place-items: center; min-height: 330px; overflow: hidden; background-color: #17221e; background-image: linear-gradient(45deg, rgb(255 255 255 / 2%) 25%, transparent 25%), linear-gradient(-45deg, rgb(255 255 255 / 2%) 25%, transparent 25%); background-size: 24px 24px; border-radius: 10px; }
+.canvas-stage canvas { display: block; width: auto; height: auto; max-width: 100%; max-height: 600px; }
+.canvas-stage.empty canvas { display: none; }
+.canvas-placeholder { position: absolute; color: #8ba098; font-size: 12px; }
+.canvas-stage:not(.empty) .canvas-placeholder { display: none; }
+.result-summary { padding: 12px; color: var(--muted); background: #f5f8f6; border-radius: 9px; font-size: 11px; }
+.result-summary strong { color: var(--ink); }
+.face-result-list, .match-list { display: grid; gap: 8px; margin-top: 14px; }
+.face-result { display: grid; grid-template-columns: auto 1fr auto; gap: 12px; align-items: center; padding: 10px 12px; background: #f7faf8; border: 1px solid var(--line); border-radius: 9px; font-size: 11px; }
+.face-index { display: grid; place-items: center; width: 26px; height: 26px; color: var(--accent-dark); background: var(--accent-soft); border-radius: 50%; font-weight: 850; }
+.quality-bars { display: flex; gap: 9px; color: var(--muted); }
+.quality-bars strong { color: var(--ink); }
+
+.compare-grid { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 18px; }
+.compare-grid .canvas-stage { min-height: 300px; margin-top: 14px; }
+.compare-action-bar { display: grid; grid-template-columns: minmax(190px, .8fr) minmax(240px, 1fr) auto minmax(250px, 1.2fr); gap: 22px; align-items: end; margin-top: 18px; }
+.compare-action-bar > label { grid-template-rows: minmax(2.65em, auto) auto; align-content: end; }
+.compare-action-bar > .button { align-self: end; }
+.compare-verdict { display: flex; align-items: center; align-self: center; justify-self: end; gap: 10px; }
+.compare-verdict .verdict-mark { display: grid; place-items: center; width: 36px; height: 36px; color: var(--muted); background: #edf0ee; border-radius: 50%; font-weight: 900; }
+.compare-verdict strong, .compare-verdict small { display: block; }
+.compare-verdict strong { font-size: 13px; }
+.compare-verdict small { color: var(--faint); font-size: 10px; }
+.compare-verdict.match .verdict-mark { color: #fff; background: var(--success); }
+.compare-verdict.no-match .verdict-mark { color: #fff; background: var(--danger); }
+.match-card { display: grid; grid-template-columns: auto 1fr auto; align-items: center; gap: 12px; padding: 12px; background: #f7faf8; border: 1px solid var(--line); border-radius: 9px; }
+.match-rank { color: var(--faint); font-size: 11px; font-weight: 800; }
+.match-person { min-width: 0; }
+.match-person strong, .match-person small { display: block; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.match-person strong { font-size: 12px; }
+.match-person small { color: var(--faint); font-size: 10px; }
+.match-score { color: var(--accent-dark); font-size: 16px; font-weight: 850; }
+
+.video-stage { position: relative; display: grid; place-items: center; min-height: 520px; overflow: hidden; background: #111b18; border-radius: 10px; }
+.video-stage img { display: block; width: 100%; height: 100%; min-height: 520px; max-height: 76vh; object-fit: contain; }
+.video-stage canvas { position: absolute; inset: 0; width: 100%; height: 100%; pointer-events: none; }
+.video-idle { position: absolute; inset: 0; display: grid; place-content: center; justify-items: center; gap: 7px; padding: 20px; color: #91a49d; text-align: center; }
+.video-idle strong { color: #d5e2dd; font-size: 14px; }
+.video-idle span:last-child { max-width: 390px; font-size: 11px; }
+.video-symbol { display: grid; place-items: center; width: 48px; height: 48px; margin-bottom: 4px; padding-left: 3px; color: #71d6b4; background: rgb(113 214 180 / 10%); border: 1px solid rgb(113 214 180 / 20%); border-radius: 50%; font-size: 18px; }
+.video-matches { display: grid; gap: 6px; }
+.video-match { display: grid; grid-template-columns: auto minmax(0, 1fr) auto; align-items: center; gap: 10px; padding: 8px 10px; background: var(--accent-soft); border-radius: 8px; font-size: 11px; }
+.video-match-person { min-width: 0; }
+.video-match-person strong,
+.video-match-person small { display: block; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.video-match-person small { color: var(--faint); }
+.video-match > strong:last-child { color: var(--accent-dark); }
+.video-match.recognized { border-left: 3px solid #57d368; }
+.video-match.not-enrolled { border-left: 3px solid #ff9e20; }
+.stream-legend { display: flex; flex-wrap: wrap; gap: 16px; margin-top: 12px; color: var(--muted); font-size: 11px; }
+.stream-legend span { display: inline-flex; align-items: center; gap: 7px; }
+.legend-swatch { width: 10px; height: 10px; border-radius: 3px; box-shadow: 0 0 0 2px rgb(255 255 255 / 8%); }
+.legend-swatch.recognized { background: #57d368; }
+.legend-swatch.not-enrolled { background: #ff9e20; }
+.form-note { margin: 0; padding: 10px 11px; color: var(--muted); background: var(--accent-soft); border: 1px solid var(--line); border-radius: 8px; font-size: 10px; line-height: 1.5; }
+.stream-metrics { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 8px; margin: 0; }
+.stream-metrics > div { min-width: 0; padding: 9px 10px; background: var(--accent-soft); border: 1px solid var(--line); border-radius: 8px; }
+.stream-metrics dt { color: var(--muted); font-size: 9px; text-transform: uppercase; letter-spacing: .06em; }
+.stream-metrics dd { overflow: hidden; margin: 3px 0 0; font-size: 12px; font-weight: 800; text-overflow: ellipsis; white-space: nowrap; }
+.monitor-admin { align-items: start; margin-bottom: 18px; }
+.monitor-list { display: grid; gap: 8px; }
+.monitor-list-item { display: grid; grid-template-columns: minmax(0, 1fr) auto; gap: 12px; align-items: center; width: 100%; padding: 11px 12px; color: var(--ink); background: var(--accent-soft); border: 1px solid var(--line); border-radius: 9px; text-align: left; cursor: pointer; }
+.monitor-list-item:hover,
+.monitor-list-item.selected { border-color: var(--accent); background: var(--accent-soft-strong, var(--accent-soft)); }
+.monitor-list-item strong,
+.monitor-list-item small { display: block; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.monitor-list-item small { margin-top: 3px; color: var(--muted); }
+.monitor-list-state { display: inline-flex; align-items: center; gap: 6px; color: var(--muted); font-size: 10px; font-weight: 800; text-transform: uppercase; }
+.monitor-list-state::before { width: 7px; height: 7px; background: #71717a; border-radius: 50%; content: ""; }
+.monitor-list-state.running::before { background: #22c55e; box-shadow: 0 0 0 3px rgb(34 197 94 / 12%); }
+.monitor-list-state.error::before,
+.monitor-list-state.reconnecting::before { background: #f59e0b; }
+.monitor-actions { display: flex; flex-wrap: wrap; gap: 8px; }
+.monitor-events-heading { margin-top: 18px; }
+.monitor-events { display: grid; gap: 7px; max-height: 260px; overflow: auto; }
+.monitor-event { display: grid; grid-template-columns: auto minmax(0, 1fr) auto; gap: 9px; align-items: center; padding: 8px 9px; border: 1px solid var(--line); border-radius: 8px; font-size: 10px; }
+.monitor-event strong,
+.monitor-event small { display: block; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.monitor-event small { color: var(--muted); }
+.monitor-event time { color: var(--faint); font-variant-numeric: tabular-nums; }
+
+.model-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(240px, 1fr)); gap: 11px; }
+.model-license-notice { margin: 0 0 16px; padding: 14px 16px; color: #704207; background: var(--warning-soft); border: 1px solid #f0d2a6; border-radius: 10px; }
+.model-license-notice strong { display: block; margin-bottom: 4px; color: #704207; font-size: 12px; }
+.model-license-notice p { margin: 3px 0; font-size: 11px; }
+.model-license-notice a { color: #704207; font-weight: 800; }
+.model-card { padding: 14px; background: #f7faf8; border: 1px solid var(--line); border-radius: 9px; }
+.model-card h4 { margin: 0 0 8px; font-size: 13px; }
+.model-card dl { margin: 0; }
+.model-card dl > div { display: flex; justify-content: space-between; gap: 12px; padding: 3px 0; font-size: 10px; }
+.model-card dt { color: var(--faint); }
+.model-card dd { margin: 0; max-width: 65%; font-weight: 650; text-align: right; overflow-wrap: anywhere; }
+
+dialog { width: min(520px, calc(100% - 30px)); padding: 0; color: var(--ink); background: var(--surface); border: 1px solid var(--line); border-radius: 15px; box-shadow: 0 24px 80px rgb(18 37 31 / 25%); }
+dialog::backdrop { background: rgb(12 24 20 / 55%); backdrop-filter: blur(2px); }
+dialog form { display: grid; gap: 15px; padding: 24px; }
+dialog p { margin: 0; color: var(--muted); font-size: 12px; }
+.dialog-heading { display: flex; align-items: start; justify-content: space-between; gap: 15px; }
+.dialog-heading h2 { margin: 0; font-size: 22px; }
+.dialog-actions { display: flex; justify-content: end; gap: 9px; padding-top: 4px; }
+
+.toast-region { position: fixed; z-index: 100; right: 22px; bottom: 22px; display: grid; gap: 9px; width: min(370px, calc(100% - 44px)); pointer-events: none; }
+.toast { padding: 13px 15px; color: #e9f6f1; background: #17332a; border: 1px solid rgb(255 255 255 / 10%); border-radius: 10px; box-shadow: 0 16px 40px rgb(18 37 31 / 23%); font-size: 12px; animation: toast-in .18s ease-out; }
+.toast.error { color: #fff0ef; background: #6b2b2b; }
+.toast strong { display: block; margin-bottom: 2px; }
+.toast span { color: rgb(255 255 255 / 72%); }
+@keyframes toast-in { from { opacity: 0; transform: translateY(7px); } }
+
+@media (max-width: 1050px) {
+  .metric-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+  .people-layout { grid-template-columns: 1fr; }
+  .tool-layout { grid-template-columns: 300px minmax(0, 1fr); }
+  .compare-action-bar { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+  .compare-verdict { grid-column: auto; justify-self: start; }
+  .form-grid.three { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+}
+
+@media (max-width: 780px) {
+  .sidebar { transform: translateX(-100%); transition: transform .2s ease; box-shadow: 16px 0 40px rgb(0 0 0 / 18%); }
+  body.menu-open .sidebar { transform: none; }
+  .main-content { margin-left: 0; }
+  .menu-button { display: inline-grid; place-items: center; }
+  .topbar { padding-inline: 16px; }
+  .page-stack { padding: 25px 16px 48px; }
+  .hero-panel { display: block; }
+  .hero-provider { margin-top: 23px; }
+  .two-column, .tool-layout, .compare-grid { grid-template-columns: 1fr; }
+  .tool-controls { order: 0; }
+  .media-panel { order: 1; min-height: auto; }
+  .video-stage,
+  .video-stage img { min-height: 360px; }
+  .section-heading { align-items: start; }
+  .form-grid.three { grid-template-columns: 1fr; }
+  .auth-state { display: none; }
+}
+
+@media (max-width: 520px) {
+  .metric-grid { grid-template-columns: 1fr; }
+  .section-heading { display: grid; }
+  .section-heading .button { width: 100%; }
+  .form-grid.two { grid-template-columns: 1fr; }
+  .toolbar { display: grid; }
+  .toolbar label { min-width: 0; }
+  .compare-action-bar { grid-template-columns: 1fr; }
+  .compare-verdict { grid-column: auto; }
+  .table-toolbar { align-items: stretch; flex-direction: column; }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  *, *::before, *::after { scroll-behavior: auto !important; animation-duration: .001ms !important; transition-duration: .001ms !important; }
+}
+
+/* InsightFace.ai brand system: dark zinc surfaces with indigo and cyan accents. */
+:root {
+  --ink: #fafafa;
+  --muted: #a1a1aa;
+  --faint: #71717a;
+  --paper: #09090b;
+  --surface: #111113;
+  --line: #27272a;
+  --line-strong: #3f3f46;
+  --nav: #09090b;
+  --nav-soft: #18181b;
+  --accent: #6366f1;
+  --accent-dark: #818cf8;
+  --accent-soft: rgb(99 102 241 / 12%);
+  --warning: #fbbf24;
+  --warning-soft: rgb(245 158 11 / 12%);
+  --danger: #f87171;
+  --danger-soft: rgb(239 68 68 / 12%);
+  --success: #34d399;
+  --shadow: 0 1px 2px rgb(0 0 0 / 28%), 0 18px 45px rgb(0 0 0 / 16%);
+  --radius: 16px;
+  --radius-small: 9px;
+  font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif;
+  color-scheme: dark;
+}
+
+html,
+body {
+  background: var(--paper);
+}
+
+body {
+  color: var(--ink);
+  background-image:
+    radial-gradient(circle at 78% -10%, rgb(99 102 241 / 12%), transparent 33rem),
+    linear-gradient(rgb(99 102 241 / 3%) 1px, transparent 1px),
+    linear-gradient(90deg, rgb(99 102 241 / 3%) 1px, transparent 1px);
+  background-size: auto, 64px 64px, 64px 64px;
+  background-attachment: fixed;
+}
+
+::selection { color: #fff; background: rgb(99 102 241 / 55%); }
+::-webkit-scrollbar { width: 10px; height: 10px; }
+::-webkit-scrollbar-track { background: #09090b; }
+::-webkit-scrollbar-thumb { background: #3f3f46; border: 2px solid #09090b; border-radius: 999px; }
+::-webkit-scrollbar-thumb:hover { background: #52525b; }
+
+:focus-visible { outline-color: rgb(99 102 241 / 55%); }
+.skip-link { border: 1px solid var(--line); box-shadow: var(--shadow); }
+
+.sidebar {
+  padding-top: 20px;
+  color: var(--ink);
+  background:
+    radial-gradient(circle at 28px 10px, rgb(99 102 241 / 14%), transparent 220px),
+    rgb(9 9 11 / 96%);
+  border-right: 1px solid rgb(39 39 42 / 85%);
+  backdrop-filter: blur(20px);
+}
+
+.brand {
+  width: 100%;
+  color: var(--ink);
+  background: transparent;
+  border: 0;
+  text-align: left;
+}
+.brand:hover .brand-mark { background: #818cf8; box-shadow: 0 0 26px rgb(99 102 241 / 38%); }
+.brand strong { font-size: 18px; font-weight: 650; letter-spacing: -.025em; }
+.brand small { color: var(--muted); font-size: 10px; font-weight: 650; letter-spacing: .12em; }
+.brand-mark {
+  width: 32px;
+  height: 32px;
+  overflow: hidden;
+  background: var(--accent);
+  border: 0;
+  border-radius: 8px;
+  box-shadow: 0 0 20px rgb(99 102 241 / 20%);
+  transition: background .2s, box-shadow .2s;
+}
+.brand-mark::before,
+.brand-mark::after { display: none; }
+.brand-mark svg { width: 20px; height: 20px; stroke: #fff; stroke-width: 2; }
+
+.server-chip {
+  color: var(--muted);
+  background: rgb(24 24 27 / 72%);
+  border-color: var(--line);
+}
+.server-chip.healthy .status-dot { background: #34d399; box-shadow: 0 0 0 3px rgb(52 211 153 / 12%); }
+.server-chip.error .status-dot { background: #f87171; box-shadow: 0 0 0 3px rgb(248 113 113 / 12%); }
+
+.nav-list { gap: 4px; }
+.nav-item {
+  min-height: 41px;
+  color: var(--muted);
+  border: 1px solid transparent;
+  border-radius: 9px;
+  font-weight: 550;
+}
+.nav-item:hover { color: var(--ink); background: var(--nav-soft); border-color: rgb(63 63 70 / 65%); }
+.nav-item.active {
+  color: #fff;
+  background: rgb(99 102 241 / 13%);
+  border-color: rgb(99 102 241 / 28%);
+  box-shadow: none;
+}
+.nav-icon { color: #818cf8; }
+.nav-icon svg { width: 17px; height: 17px; fill: none; stroke: currentColor; stroke-width: 1.8; stroke-linecap: round; stroke-linejoin: round; }
+.sidebar-footer { border-top: 1px solid rgb(39 39 42 / 75%); }
+.sidebar-footer p { color: #71717a; }
+
+.main-content { background: rgb(9 9 11 / 34%); }
+.topbar {
+  min-height: 74px;
+  background: rgb(9 9 11 / 78%);
+  border-bottom-color: rgb(39 39 42 / 75%);
+  backdrop-filter: blur(20px);
+}
+.topbar h1 { color: var(--ink); font-size: 22px; font-weight: 650; }
+.eyebrow { color: #818cf8; }
+.auth-state.configured { color: #818cf8; }
+.menu-button { color: var(--ink); background: var(--surface); border-color: var(--line); }
+.page h2,
+.page h3 { color: var(--ink); }
+.section-heading p:last-child { color: var(--muted); }
+
+.hero-panel {
+  min-height: 190px;
+  color: var(--ink);
+  background:
+    linear-gradient(rgb(99 102 241 / 5%) 1px, transparent 1px),
+    linear-gradient(90deg, rgb(99 102 241 / 5%) 1px, transparent 1px),
+    radial-gradient(circle at 82% 15%, rgb(99 102 241 / 19%), transparent 42%),
+    linear-gradient(135deg, #111113, #0c0c10);
+  background-size: 48px 48px, 48px 48px, auto, auto;
+  border: 1px solid var(--line);
+  border-radius: 20px;
+  box-shadow: 0 0 80px -30px rgb(99 102 241 / 42%);
+}
+.hero-panel::after {
+  right: -100px;
+  bottom: -210px;
+  width: 420px;
+  height: 420px;
+  background: rgb(99 102 241 / 8%);
+  border: 1px solid rgb(99 102 241 / 16%);
+  box-shadow: 0 0 100px rgb(99 102 241 / 14%);
+}
+.hero-panel .eyebrow { color: #818cf8; }
+.hero-panel h2 { font-weight: 720; letter-spacing: -.045em; }
+.hero-panel p:last-child { color: var(--muted); }
+.gradient-text {
+  color: transparent;
+  background: linear-gradient(90deg, #6366f1, #818cf8, #22d3ee);
+  background-clip: text;
+  -webkit-background-clip: text;
+}
+.hero-provider { background: rgb(24 24 27 / 78%); border-color: var(--line); }
+.hero-provider span { color: var(--muted); }
+.hero-provider strong { color: #e4e4e7; }
+
+.metric-card,
+.panel {
+  background: rgb(17 17 19 / 94%);
+  border-color: var(--line);
+  box-shadow: var(--shadow);
+  transition: border-color .2s, box-shadow .2s;
+}
+.metric-card:hover { border-color: rgb(99 102 241 / 38%); box-shadow: 0 0 36px -18px rgb(99 102 241 / 35%); }
+.metric-card > span,
+.metric-card > small { color: var(--muted); }
+.status-list > div,
+.diagnostic-list > div,
+.error-entry,
+.person-detail-header { border-color: var(--line); }
+.diagnostic-list dd.good { color: var(--success); }
+
+.button { font-weight: 650; }
+.button.primary { background: var(--accent); box-shadow: 0 7px 24px rgb(99 102 241 / 20%); }
+.button.primary:hover { background: #818cf8; box-shadow: 0 8px 30px rgb(99 102 241 / 30%); }
+.button.secondary { color: var(--ink); background: var(--surface); border-color: var(--line-strong); }
+.button.secondary:hover { background: #18181b; border-color: #52525b; }
+.sidebar .button.secondary { color: var(--ink); background: #18181b; border-color: #3f3f46; }
+.sidebar .button.secondary:hover { border-color: rgb(99 102 241 / 55%); }
+.button.danger { color: #fca5a5; background: var(--danger-soft); border-color: rgb(239 68 68 / 22%); }
+.icon-button { color: var(--muted); background: var(--surface); border-color: var(--line); }
+.icon-button:hover { color: var(--ink); background: #18181b; border-color: #52525b; }
+.text-button,
+output { color: #818cf8; }
+
+label { color: #d4d4d8; font-weight: 600; }
+input,
+textarea,
+select {
+  color: var(--ink);
+  background: #18181b;
+  border-color: #3f3f46;
+}
+input::placeholder,
+textarea::placeholder { color: #71717a; }
+input:hover,
+textarea:hover,
+select:hover { border-color: #52525b; }
+input:focus,
+textarea:focus,
+select:focus { border-color: #6366f1; box-shadow: 0 0 0 3px rgb(99 102 241 / 16%); }
+input[type="range"],
+.checkbox-field input[type="checkbox"] { accent-color: #6366f1; }
+.form-panel { border-color: rgb(99 102 241 / 30%); }
+.form-status.error { color: #f87171; }
+
+.drop-zone {
+  color: var(--muted);
+  background: #0d0d0f;
+  border-color: #3f3f46;
+}
+.drop-zone:hover,
+.drop-zone.dragging { background: rgb(99 102 241 / 8%); border-color: #6366f1; }
+.drop-zone strong { color: var(--ink); }
+.drop-zone .file-summary { color: #818cf8; }
+.drop-icon { color: #a5b4fc; background: rgb(99 102 241 / 14%); }
+
+th { color: #71717a; border-color: var(--line); }
+td { border-color: var(--line); }
+tbody tr:hover { background: rgb(99 102 241 / 5%); }
+.table-title,
+.empty-state strong { color: var(--ink); }
+.table-subtitle { color: var(--muted); }
+.badge { color: #a5b4fc; background: rgb(99 102 241 / 14%); }
+
+.person-row { color: var(--ink); }
+.person-row:hover { background: #18181b; }
+.person-row.selected { background: rgb(99 102 241 / 12%); border-color: rgb(99 102 241 / 30%); }
+.avatar { color: #c7d2fe; background: rgb(99 102 241 / 18%); }
+.person-metadata { color: #d4d4d8; background: #0d0d0f; border: 1px solid var(--line); }
+.face-card,
+.model-card,
+.face-result,
+.match-card { background: #18181b; border-color: var(--line); }
+.rejection-list { color: #fcd34d; background: var(--warning-soft); border-color: rgb(245 158 11 / 25%); }
+
+.canvas-stage,
+.video-stage { background-color: #09090b; border: 1px solid var(--line); }
+.canvas-placeholder,
+.video-idle { color: var(--muted); }
+.video-idle strong { color: #e4e4e7; }
+.video-symbol { color: #22d3ee; background: rgb(34 211 238 / 10%); border-color: rgb(34 211 238 / 25%); }
+.result-summary { color: var(--muted); background: #18181b; border: 1px solid var(--line); }
+.result-summary strong { color: var(--ink); }
+.face-index { color: #c7d2fe; background: rgb(99 102 241 / 16%); }
+.compare-verdict .verdict-mark { color: var(--muted); background: #27272a; }
+.match-score,
+.video-match > strong:last-child { color: #818cf8; }
+.video-match { background: rgb(99 102 241 / 12%); border: 1px solid rgb(99 102 241 / 20%); }
+
+dialog { color: var(--ink); background: #111113; border-color: #3f3f46; box-shadow: 0 32px 100px rgb(0 0 0 / 55%); }
+dialog::backdrop { background: rgb(0 0 0 / 68%); backdrop-filter: blur(5px); }
+.toast { color: #f4f4f5; background: #18181b; border-color: #3f3f46; box-shadow: 0 18px 50px rgb(0 0 0 / 38%); }
+.toast.error { color: #fee2e2; background: #2a1215; border-color: rgb(248 113 113 / 25%); }
+
+.language-control { display: inline-flex; align-items: center; margin: 0; }
+.language-control select { width: auto; min-width: 118px; height: 38px; padding: 0 34px 0 12px; border-color: var(--line-strong); background-color: rgb(17 17 19 / 92%); color: var(--ink); }
+
+.developer-tools-panel,
+.documentation-developer-note { display: flex; align-items: center; justify-content: space-between; gap: 24px; }
+.developer-tools-panel h3,
+.documentation-developer-note strong { display: block; margin: 2px 0 6px; color: var(--ink); }
+.developer-tools-panel p,
+.documentation-developer-note p { margin: 0; color: var(--muted); line-height: 1.6; }
+.developer-tools-actions { display: flex; flex: 0 0 auto; align-items: center; gap: 14px; }
+.text-link { color: #818cf8; font-size: .82rem; font-weight: 700; text-decoration: none; }
+.text-link:hover { color: #c7d2fe; text-decoration: underline; }
+.documentation-developer-note { margin-top: 18px; padding: 18px 20px; }
+
+.documentation-switcher { display: flex; flex-wrap: wrap; gap: 8px; margin-bottom: 18px; padding: 6px; background: #111113; border: 1px solid var(--line); border-radius: 12px; }
+.documentation-switch { display: inline-flex; min-height: 42px; align-items: center; gap: 8px; padding: 9px 14px; color: var(--muted); background: transparent; border: 1px solid transparent; border-radius: 8px; font: inherit; font-weight: 700; cursor: pointer; }
+.documentation-switch:hover, .documentation-switch:focus-visible { color: var(--ink); background: var(--nav-soft); outline: none; }
+.documentation-switch.active { color: #eef2ff; background: rgb(99 102 241 / 16%); border-color: rgb(129 140 248 / 34%); }
+.documentation-switch small { padding: 2px 6px; color: #a5b4fc; background: rgb(99 102 241 / 14%); border-radius: 999px; font-size: .65rem; font-weight: 800; letter-spacing: .02em; }
+.documentation-layout { display: grid; grid-template-columns: 280px minmax(0, 1fr); gap: 18px; align-items: start; }
+.documentation-nav { position: sticky; top: 92px; display: flex; max-height: calc(100vh - 116px); flex-direction: column; gap: 12px; padding: 16px; overflow: hidden; }
+.documentation-nav > strong { padding: 0 4px; font-size: .92rem; }
+.documentation-nav p { margin: 2px 4px 0; }
+.documentation-filter { gap: 6px; margin: 0; padding: 0 4px; font-size: .72rem; }
+.documentation-filter input { height: 38px; padding: 8px 10px; font-size: .82rem; }
+.documentation-toc { display: grid; gap: 14px; padding-right: 4px; overflow-y: auto; overscroll-behavior: contain; scrollbar-width: thin; }
+.documentation-toc-group { display: grid; gap: 3px; padding-bottom: 12px; border-bottom: 1px solid var(--line); }
+.documentation-toc-group:last-child { padding-bottom: 0; border-bottom: 0; }
+.documentation-toc-link { width: 100%; padding: 7px 9px; color: var(--muted); background: transparent; border: 0; border-radius: 8px; font: inherit; font-size: .78rem; line-height: 1.35; text-align: left; cursor: pointer; }
+.documentation-toc-link:hover, .documentation-toc-link:focus-visible { color: var(--ink); background: var(--nav-soft); outline: none; }
+.documentation-toc-link.level-2 { color: var(--ink); font-size: .9rem; font-weight: 700; }
+.documentation-toc-link.level-3 { padding-left: 16px; font-weight: 600; }
+.documentation-toc-link.level-4 { padding-left: 25px; font-family: var(--mono); font-size: .7rem; }
+.markdown-document { min-height: 65vh; padding: clamp(22px, 4vw, 48px); overflow: hidden; }
+.markdown-document > :first-child { margin-top: 0; }
+.markdown-document h2 { margin-top: 0; font-size: clamp(2rem, 4vw, 3.2rem); letter-spacing: -.045em; scroll-margin-top: 100px; }
+.markdown-document h3 { margin-top: 2.2em; padding-top: .5em; border-top: 1px solid var(--line); font-size: 1.45rem; scroll-margin-top: 100px; }
+.markdown-document h4 { margin-top: 1.7em; font-size: 1.08rem; scroll-margin-top: 100px; }
+.markdown-document p, .markdown-document li { color: var(--muted); line-height: 1.75; }
+.markdown-document li + li { margin-top: .35em; }
+.markdown-document a { color: #818cf8; }
+.markdown-document code { padding: .15em .4em; color: #c7d2fe; background: #09090b; border: 1px solid var(--line); border-radius: 6px; }
+.markdown-document pre { padding: 18px; overflow: auto; background: #09090b; border: 1px solid var(--line); border-radius: 12px; }
+.markdown-document pre code { padding: 0; color: #e4e4e7; background: transparent; border: 0; }
+.markdown-document img { display: block; width: min(100%, 1280px); height: auto; margin: 18px auto; border: 1px solid var(--line); border-radius: 12px; box-shadow: 0 16px 40px rgb(0 0 0 / 28%); }
+.markdown-document blockquote { margin: 1.5em 0; padding: 2px 18px; color: var(--muted); border-left: 3px solid var(--accent); }
+.markdown-table-wrap { max-width: 100%; overflow-x: auto; border: 1px solid var(--line); border-radius: 12px; }
+.markdown-document table { min-width: 640px; }
+.loading-document { min-height: 240px; display: grid; place-content: center; gap: 12px; color: var(--muted); text-align: center; }
+.loading-document span { width: 28px; height: 28px; margin: auto; border: 2px solid var(--line); border-top-color: #818cf8; border-radius: 50%; animation: spin .8s linear infinite; }
+
+@keyframes spin { to { transform: rotate(360deg); } }
+
+@media (max-width: 780px) {
+  .language-control select { min-width: 96px; max-width: 118px; }
+  .developer-tools-panel,
+  .documentation-developer-note { display: grid; }
+  .developer-tools-actions { flex-wrap: wrap; }
+  .documentation-layout { grid-template-columns: 1fr; }
+  .documentation-nav { position: static; max-height: none; }
+  .documentation-toc { max-height: 280px; }
+}
diff --git a/server/frontend/ui-translations.mjs b/server/frontend/ui-translations.mjs
new file mode 100644
index 000000000..44cf2300c
--- /dev/null
+++ b/server/frontend/ui-translations.mjs
@@ -0,0 +1,452 @@
+const localeCodes = ["zh", "ja", "de", "es", "fr", "ru", "pt", "ko"];
+
+// One source string followed by Chinese, Japanese, German, Spanish, French,
+// Russian, Portuguese, and Korean. Keeping each phrase on one row makes gaps
+// visible in review and lets tests enforce identical coverage for every locale.
+const rows = [
+  ["API guide", "API 使用手册", "API 利用ガイド", "API-Leitfaden", "Guía de API", "Guide API", "Руководство по API", "Guia da API", "API 사용 가이드"],
+  ["Maintainer guide", "维护者指南", "メンテナーガイド", "Maintainer-Handbuch", "Guía de mantenimiento", "Guide de maintenance", "Руководство сопровождающего", "Guia de manutenção", "유지관리자 가이드"],
+  ["English only", "仅英文", "英語のみ", "Nur Englisch", "Solo en inglés", "Anglais uniquement", "Только на английском", "Apenas em inglês", "영어만 제공"],
+  ["Beginner guidance and complete API documentation are bundled with this server and work without internet access.", "面向新手的使用指南和完整 API 文档已内置在服务器中，无需联网即可阅读。", "初心者向けガイドと完全な API ドキュメントがサーバーに内蔵され、オフラインで利用できます。", "Einsteigerleitfaden und vollständige API-Dokumentation sind im Server enthalten und offline verfügbar.", "La guía para principiantes y la documentación completa de la API están integradas en el servidor y funcionan sin internet.", "Le guide pour débutants et la documentation API complète sont intégrés au serveur et disponibles hors ligne.", "Руководство для начинающих и полная документация API встроены в сервер и доступны без интернета.", "O guia para principiantes e a documentação completa da API estão incluídos no servidor e funcionam sem internet.", "초보자 가이드와 전체 API 문서가 서버에 포함되어 오프라인에서도 사용할 수 있습니다."],
+  ["On this page", "本页目录", "このページの目次", "Auf dieser Seite", "En esta página", "Sur cette page", "На этой странице", "Nesta página", "이 페이지의 목차"],
+  ["Find a topic", "查找主题", "トピックを検索", "Thema suchen", "Buscar un tema", "Rechercher un sujet", "Найти тему", "Procurar um tópico", "주제 찾기"],
+  ["Search help topics…", "搜索帮助主题…", "ヘルプトピックを検索…", "Hilfethemen durchsuchen…", "Buscar temas de ayuda…", "Rechercher dans l’aide…", "Поиск по темам справки…", "Pesquisar tópicos de ajuda…", "도움말 주제 검색…"],
+  ["Help topics", "帮助主题", "ヘルプトピック", "Hilfethemen", "Temas de ayuda", "Sujets d’aide", "Темы справки", "Tópicos de ajuda", "도움말 주제"],
+  ["DEVELOPER TOOLS", "开发者工具", "開発者ツール", "ENTWICKLERWERKZEUGE", "HERRAMIENTAS PARA DESARROLLADORES", "OUTILS DE DÉVELOPPEMENT", "ИНСТРУМЕНТЫ РАЗРАБОТЧИКА", "FERRAMENTAS DE DESENVOLVIMENTO", "개발자 도구"],
+  ["OpenAPI Schema", "OpenAPI Schema", "OpenAPI スキーマ", "OpenAPI-Schema", "Esquema OpenAPI", "Schéma OpenAPI", "Схема OpenAPI", "Esquema OpenAPI", "OpenAPI 스키마"],
+  ["OpenAPI Schema · InsightFace Server", "OpenAPI Schema · InsightFace Server", "OpenAPI スキーマ · InsightFace Server", "OpenAPI-Schema · InsightFace Server", "Esquema OpenAPI · InsightFace Server", "Schéma OpenAPI · InsightFace Server", "Схема OpenAPI · InsightFace Server", "Esquema OpenAPI · InsightFace Server", "OpenAPI 스키마 · InsightFace Server"],
+  ["Inspect the exact paths, parameters, and response schemas exposed by this running server.", "检查当前运行实例实际公开的路径、参数和响应 Schema。", "実行中のサーバーが公開する正確なパス、パラメーター、レスポンススキーマを確認します。", "Prüfen Sie die exakten Pfade, Parameter und Antwortschemas dieses laufenden Servers.", "Consulte las rutas, los parámetros y los esquemas de respuesta exactos que expone este servidor.", "Consultez les chemins, paramètres et schémas de réponse exacts exposés par ce serveur.", "Просматривайте точные пути, параметры и схемы ответов, доступные на этом сервере.", "Consulte os caminhos, parâmetros e esquemas de resposta exatos expostos por este servidor.", "현재 서버가 제공하는 정확한 경로, 매개변수와 응답 스키마를 확인합니다."],
+  ["Open schema explorer", "打开 Schema 浏览器", "スキーマエクスプローラーを開く", "Schema-Explorer öffnen", "Abrir explorador de esquemas", "Ouvrir l’explorateur de schéma", "Открыть обозреватель схемы", "Abrir explorador de esquema", "스키마 탐색기 열기"],
+  ["Developer schema", "开发者 Schema", "開発者向けスキーマ", "Entwicklerschema", "Esquema para desarrolladores", "Schéma développeur", "Схема для разработчиков", "Esquema para programadores", "개발자 스키마"],
+  ["Use the OpenAPI schema when generating clients or checking exact runtime types.", "生成客户端或核对运行时精确类型时，请使用 OpenAPI Schema。", "クライアント生成や実行時の正確な型確認には OpenAPI スキーマを使用してください。", "Verwenden Sie das OpenAPI-Schema zum Erzeugen von Clients oder Prüfen exakter Laufzeittypen.", "Use el esquema OpenAPI para generar clientes o comprobar los tipos exactos en ejecución.", "Utilisez le schéma OpenAPI pour générer des clients ou vérifier les types exacts à l’exécution.", "Используйте схему OpenAPI для генерации клиентов и проверки точных типов работающего сервера.", "Use o esquema OpenAPI para gerar clientes ou verificar os tipos exatos em execução.", "클라이언트 생성이나 정확한 런타임 타입 확인에는 OpenAPI 스키마를 사용하세요."],
+  ["DEVELOPER TOOL · OPENAPI 3", "开发者工具 · OPENAPI 3", "開発者ツール · OPENAPI 3", "ENTWICKLERWERKZEUG · OPENAPI 3", "HERRAMIENTA PARA DESARROLLADORES · OPENAPI 3", "OUTIL DE DÉVELOPPEMENT · OPENAPI 3", "ИНСТРУМЕНТ РАЗРАБОТЧИКА · OPENAPI 3", "FERRAMENTA DE DESENVOLVIMENTO · OPENAPI 3", "개발자 도구 · OPENAPI 3"],
+  ["Live API schema.", "实时 API Schema。", "ライブ API スキーマ。", "Live-API-Schema.", "Esquema de API en vivo.", "Schéma API dynamique.", "Актуальная схема API.", "Esquema da API em tempo real.", "실시간 API 스키마."],
+  ["New to the server? Follow the", "第一次使用？请先阅读", "初めて利用する場合は", "Neu beim Server? Lesen Sie zuerst den", "¿Es la primera vez? Siga la", "Première utilisation ? Suivez le", "Впервые используете сервер? Откройте", "É a primeira vez? Siga o", "처음 사용하시나요? 먼저"],
+  ["step-by-step user guide", "分步用户指南", "ステップ形式のユーザーガイド", "Benutzerleitfaden mit einzelnen Schritten", "guía de usuario paso a paso", "guide utilisateur pas à pas", "пошаговое руководство пользователя", "guia do utilizador passo a passo", "단계별 사용자 가이드"],
+  ["before using this live schema reference.", "，再查看实时Schema参考。", "を確認してから、このライブスキーマ参照を利用してください。", "bevor Sie diese Live-Schemareferenz verwenden.", "antes de usar esta referencia de esquema en vivo.", "avant d'utiliser cette référence de schéma dynamique.", "перед использованием этого справочника по актуальной схеме.", "antes de usar esta referência de esquema em tempo real.", "를 읽은 뒤 실시간 스키마 참조를 사용하세요."],
+  ["Toggle navigation", "切换导航", "ナビゲーションを切り替え", "Navigation umschalten", "Alternar navegación", "Afficher/masquer la navigation", "Переключить навигацию", "Alternar navegação", "탐색 메뉴 전환"],
+  ["Checking…", "正在检查…", "確認中…", "Wird geprüft…", "Comprobando…", "Vérification…", "Проверка…", "A verificar…", "확인 중…"],
+  ["not loaded", "未加载", "未ロード", "nicht geladen", "no cargado", "non chargé", "не загружено", "não carregado", "로드되지 않음"],
+  ["LOCAL FACE INTELLIGENCE", "本地人脸智能", "ローカル顔認識", "LOKALE GESICHTSINTELLIGENZ", "INTELIGENCIA FACIAL LOCAL", "INTELLIGENCE FACIALE LOCALE", "ЛОКАЛЬНОЕ РАСПОЗНАВАНИЕ ЛИЦ", "INTELIGÊNCIA FACIAL LOCAL", "로컬 얼굴 인텔리전스"],
+  ["Your face recognition service,", "您的人脸识别服务，", "顔認識サービスを", "Ihr Gesichtserkennungsdienst", "Su servicio de reconocimiento facial,", "Votre service de reconnaissance faciale,", "Ваш сервис распознавания лиц —", "O seu serviço de reconhecimento facial,", "얼굴 인식 서비스를"],
+  ["at a glance.", "一目了然。", "ひと目で確認。", "auf einen Blick.", "de un vistazo.", "en un coup d’œil.", "всё на одном экране.", "num relance.", "한눈에 확인하세요."],
+  ["Database totals", "数据库统计", "データベース集計", "Datenbank gesamt", "Totales de la base de datos", "Totaux de la base", "Итоги базы данных", "Totais da base de dados", "데이터베이스 합계"],
+  ["SERVICE", "服务", "サービス", "DIENST", "SERVICIO", "SERVICE", "СЕРВИС", "SERVIÇO", "서비스"],
+  ["View diagnostics →", "查看诊断 →", "診断を表示 →", "Diagnose anzeigen →", "Ver diagnóstico →", "Voir le diagnostic →", "Открыть диагностику →", "Ver diagnóstico →", "진단 보기 →"],
+  ["THIS TAB", "当前标签页", "このタブ", "DIESER TAB", "ESTA PESTAÑA", "CET ONGLET", "ЭТА ВКЛАДКА", "ESTE SEPARADOR", "현재 탭"],
+  ["IDENTITY SPACES", "身份空间", "ID 空間", "IDENTITÄTSRÄUME", "ESPACIOS DE IDENTIDAD", "ESPACES D’IDENTITÉ", "ПРОСТРАНСТВА ИДЕНТИЧНОСТЕЙ", "ESPAÇOS DE IDENTIDADE", "신원 공간"],
+  ["Each collection is pinned to the model and preprocessing version active when it is created.", "每个人员库都固定绑定到创建时使用的模型和预处理版本。", "各コレクションは作成時のモデルと前処理バージョンに固定されます。", "Jede Sammlung bleibt an das bei ihrer Erstellung aktive Modell und die Vorverarbeitung gebunden.", "Cada colección queda vinculada al modelo y al preprocesamiento activos al crearla.", "Chaque collection reste liée au modèle et au prétraitement actifs lors de sa création.", "Каждая коллекция привязана к модели и версии предобработки на момент создания.", "Cada coleção fica associada ao modelo e ao pré-processamento ativos na criação.", "각 컬렉션은 생성 시점의 모델과 전처리 버전에 고정됩니다."],
+  ["CREATE", "创建", "作成", "ERSTELLEN", "CREAR", "CRÉER", "СОЗДАНИЕ", "CRIAR", "생성"],
+  ["Letters, numbers, dot, dash, underscore; max 64. `_default` is reserved for the default collection.", "可使用字母、数字、点、短横线和下划线，最多 64 个字符；`_default` 保留给默认人员库。", "英数字、ピリオド、ハイフン、アンダースコアを使用でき、最大 64 文字です。`_default` は既定コレクション用に予約されています。", "Buchstaben, Ziffern, Punkt, Bindestrich und Unterstrich; maximal 64 Zeichen. `_default` ist für die Standardsammlung reserviert.", "Letras, números, punto, guion y guion bajo; máximo 64 caracteres. `_default` está reservado para la colección predeterminada.", "Lettres, chiffres, point, tiret et soulignement ; 64 caractères maximum. `_default` est réservé à la collection par défaut.", "Буквы, цифры, точка, дефис и подчёркивание; не более 64 символов. `_default` зарезервирован для коллекции по умолчанию.", "Letras, números, ponto, hífen e sublinhado; máximo de 64 caracteres. `_default` está reservado para a coleção predefinida.", "영문자, 숫자, 마침표, 하이픈, 밑줄을 사용할 수 있으며 최대 64자입니다. `_default`는 기본 컬렉션용으로 예약되어 있습니다."],
+  ["Company Employees", "公司员工", "会社従業員", "Unternehmensmitarbeiter", "Empleados de la empresa", "Employés de l’entreprise", "Сотрудники компании", "Colaboradores da empresa", "회사 직원"],
+  ["Comma-separated width×height pairs; each side must be a multiple of 32.", "使用逗号分隔的宽×高组合；每条边都必须是 32 的倍数。", "幅×高さをカンマで区切ります。各辺は 32 の倍数にしてください。", "Kommagetrennte Breite×Höhe-Paare; jede Seite muss ein Vielfaches von 32 sein.", "Pares ancho×alto separados por comas; cada lado debe ser múltiplo de 32.", "Paires largeur×hauteur séparées par des virgules ; chaque côté doit être un multiple de 32.", "Пары ширина×высота через запятую; каждая сторона должна быть кратна 32.", "Pares largura×altura separados por vírgulas; cada lado deve ser múltiplo de 32.", "쉼표로 구분한 너비×높이 쌍이며 각 변은 32의 배수여야 합니다."],
+  ["Off by default. This never stores the original upload and may substantially increase database size.", "默认关闭。不会保存原始上传图片，但启用后可能显著增大数据库。", "既定ではオフです。元のアップロード画像は保存しませんが、データベース容量が大きく増える場合があります。", "Standardmäßig aus. Das Originalbild wird nie gespeichert; die Datenbank kann jedoch deutlich größer werden.", "Desactivado por defecto. Nunca guarda la imagen original y puede aumentar mucho el tamaño de la base de datos.", "Désactivé par défaut. L’image d’origine n’est jamais conservée et la base peut fortement grossir.", "По умолчанию выключено. Исходное изображение не сохраняется, но размер базы может значительно увеличиться.", "Desligado por predefinição. Nunca guarda o original e pode aumentar bastante a base de dados.", "기본값은 꺼짐입니다. 원본 업로드는 저장하지 않지만 데이터베이스 크기가 크게 늘 수 있습니다."],
+  ["Filter collections", "筛选人员库", "コレクションを絞り込み", "Sammlungen filtern", "Filtrar colecciones", "Filtrer les collections", "Фильтр коллекций", "Filtrar coleções", "컬렉션 필터"],
+  ["Filter collections…", "筛选人员库…", "コレクションを絞り込み…", "Sammlungen filtern…", "Filtrar colecciones…", "Filtrer les collections…", "Фильтр коллекций…", "Filtrar coleções…", "컬렉션 필터…"],
+  ["0 collections", "0 个人员库", "0 コレクション", "0 Sammlungen", "0 colecciones", "0 collection", "0 коллекций", "0 coleções", "컬렉션 0개"],
+  ["Crops", "裁剪图", "クロップ", "Ausschnitte", "Recortes", "Recadrages", "Кропы", "Recortes", "크롭"],
+  ["Actions", "操作", "操作", "Aktionen", "Acciones", "Actions", "Действия", "Ações", "작업"],
+  ["No collections yet", "还没有人员库", "コレクションはまだありません", "Noch keine Sammlungen", "Aún no hay colecciones", "Aucune collection", "Коллекций пока нет", "Ainda não existem coleções", "아직 컬렉션이 없습니다"],
+  ["Create one to start enrolling people.", "创建一个人员库后即可开始注册人员。", "コレクションを作成して人物登録を開始します。", "Erstellen Sie eine Sammlung, um Personen zu registrieren.", "Cree una para empezar a registrar personas.", "Créez-en une pour commencer à inscrire des personnes.", "Создайте коллекцию, чтобы начать регистрацию людей.", "Crie uma para começar a registar pessoas.", "컬렉션을 생성한 후 인물 등록을 시작하세요."],
+  ["ENROLLMENT", "人员注册", "登録", "REGISTRIERUNG", "REGISTRO", "INSCRIPTION", "РЕГИСТРАЦИЯ", "REGISTO", "등록"],
+  ["Register several clear photos for each real person to improve search robustness.", "为每位真实人员注册多张清晰照片，可提高搜索稳定性。", "人物ごとに鮮明な写真を複数登録すると、検索が安定します。", "Mehrere klare Fotos pro Person verbessern die Zuverlässigkeit der Suche.", "Registre varias fotos nítidas de cada persona para mejorar la búsqueda.", "Inscrivez plusieurs photos nettes par personne pour fiabiliser la recherche.", "Зарегистрируйте несколько чётких фото каждого человека для более устойчивого поиска.", "Registe várias fotos nítidas de cada pessoa para tornar a pesquisa mais robusta.", "인물마다 선명한 사진을 여러 장 등록하면 검색 안정성이 향상됩니다."],
+  ["Name, ID, or external ID", "姓名、ID 或外部 ID", "名前、ID、外部 ID", "Name, ID oder externe ID", "Nombre, ID o ID externo", "Nom, ID ou ID externe", "Имя, ID или внешний ID", "Nome, ID ou ID externo", "이름, ID 또는 외부 ID"],
+  ["REGISTER", "注册", "登録", "REGISTRIEREN", "REGISTRAR", "INSCRIRE", "РЕГИСТРАЦИЯ", "REGISTAR", "등록"],
+  ["Collection embedding contract:", "人员库特征契约：", "コレクションの特徴量契約：", "Embedding-Vertrag der Sammlung:", "Contrato de embeddings de la colección:", "Contrat d’embedding de la collection :", "Контракт эмбеддингов коллекции:", "Contrato de embeddings da coleção:", "컬렉션 임베딩 계약:"],
+  ["Images are still required for detection and quality review; the server will not extract or verify these features.", "仍需提供图片用于检测和质量审查；服务器不会抽取或校验这些特征。", "検出と品質確認のため画像は必要です。サーバーは特徴量の抽出や検証を行いません。", "Bilder bleiben für Erkennung und Qualitätsprüfung erforderlich; der Server extrahiert oder prüft diese Merkmale nicht.", "Las imágenes siguen siendo necesarias para detección y calidad; el servidor no extraerá ni verificará estos vectores.", "Les images restent requises pour la détection et la qualité ; le serveur n’extrait ni ne vérifie ces vecteurs.", "Изображения всё равно нужны для детекции и проверки качества; сервер не извлекает и не проверяет эти признаки.", "As imagens continuam necessárias para deteção e qualidade; o servidor não extrai nem valida estes vetores.", "검출 및 품질 검토를 위해 이미지는 계속 필요하며 서버는 이 특징을 추출하거나 검증하지 않습니다."],
+  ["External embeddings (JSON, one vector per image)", "外部特征（JSON，每张图片一个向量）", "外部特徴量（JSON、画像ごとに 1 ベクトル）", "Externe Embeddings (JSON, ein Vektor pro Bild)", "Embeddings externos (JSON, un vector por imagen)", "Embeddings externes (JSON, un vecteur par image)", "Внешние эмбеддинги (JSON, один вектор на изображение)", "Embeddings externos (JSON, um vetor por imagem)", "외부 임베딩(JSON, 이미지당 벡터 1개)"],
+  ["or choose one or more JPEG, PNG, or WebP images", "或选择一张或多张 JPEG、PNG 或 WebP 图片", "または JPEG、PNG、WebP 画像を 1 枚以上選択", "oder ein oder mehrere JPEG-, PNG- oder WebP-Bilder auswählen", "o elija una o varias imágenes JPEG, PNG o WebP", "ou choisissez une ou plusieurs images JPEG, PNG ou WebP", "или выберите одно или несколько изображений JPEG, PNG или WebP", "ou escolha uma ou mais imagens JPEG, PNG ou WebP", "또는 JPEG, PNG, WebP 이미지를 하나 이상 선택"],
+  ["Off uses the Collection single-face selection strategy. Standard and strict require exactly one clear, usable face.", "关闭时使用人员库的单脸挑选策略；标准和严格模式要求恰好一张清晰可用的人脸。", "オフではコレクションの単一顔選択を使用します。標準と厳格では鮮明で利用可能な顔が 1 つだけ必要です。", "Aus verwendet die Ein-Gesicht-Auswahl der Sammlung. Standard und Strikt verlangen genau ein klares, nutzbares Gesicht.", "Desactivado usa la selección de rostro de la colección. Estándar y estricto requieren exactamente un rostro claro y válido.", "Désactivé utilise la sélection de la collection. Standard et strict exigent exactement un visage net et exploitable.", "В режиме «Выкл.» используется стратегия коллекции. Стандартный и строгий режимы требуют ровно одно чёткое пригодное лицо.", "Desligado usa a seleção da coleção. Normal e estrito exigem exatamente um rosto nítido e utilizável.", "꺼짐은 컬렉션의 단일 얼굴 선택 전략을 사용합니다. 표준 및 엄격 모드는 선명하고 사용 가능한 얼굴이 정확히 하나여야 합니다."],
+  ["DIRECTORY", "人员目录", "人物一覧", "VERZEICHNIS", "DIRECTORIO", "RÉPERTOIRE", "КАТАЛОГ", "DIRETÓRIO", "인물 목록"],
+  ["0 people", "0 人", "0 人", "0 Personen", "0 personas", "0 personne", "0 человек", "0 pessoas", "인물 0명"],
+  ["Its enrolled people will appear here.", "已注册人员会显示在这里。", "登録された人物がここに表示されます。", "Die registrierten Personen erscheinen hier.", "Las personas registradas aparecerán aquí.", "Les personnes inscrites apparaîtront ici.", "Зарегистрированные люди появятся здесь.", "As pessoas registadas aparecerão aqui.", "등록된 인물이 여기에 표시됩니다."],
+  ["Choose a person to view details and face samples.", "选择人员以查看详情和人脸样本。", "人物を選択して詳細と顔サンプルを表示します。", "Wählen Sie eine Person, um Details und Gesichtsproben anzuzeigen.", "Elija una persona para ver sus detalles y muestras faciales.", "Choisissez une personne pour voir ses détails et échantillons faciaux.", "Выберите человека, чтобы увидеть сведения и образцы лиц.", "Escolha uma pessoa para ver os detalhes e as amostras faciais.", "인물을 선택하여 세부 정보와 얼굴 샘플을 확인하세요."],
+  ["STATELESS", "无状态", "ステートレス", "STATUSLOS", "SIN ESTADO", "SANS ÉTAT", "БЕЗ СОСТОЯНИЯ", "SEM ESTADO", "무상태"],
+  ["Find faces, five-point landmarks, confidence, and quality in one image.", "在一张图片中检测人脸、五点关键点、置信度和质量。", "1 枚の画像から顔、5 点ランドマーク、信頼度、品質を検出します。", "Findet Gesichter, fünf Landmarken, Konfidenz und Qualität in einem Bild.", "Detecta rostros, cinco puntos, confianza y calidad en una imagen.", "Détecte les visages, cinq points, la confiance et la qualité dans une image.", "Находит лица, пять ключевых точек, уверенность и качество на одном изображении.", "Deteta rostos, cinco pontos, confiança e qualidade numa imagem.", "한 이미지에서 얼굴, 5점 랜드마크, 신뢰도와 품질을 검출합니다."],
+  ["No file selected", "未选择文件", "ファイル未選択", "Keine Datei ausgewählt", "Ningún archivo seleccionado", "Aucun fichier sélectionné", "Файл не выбран", "Nenhum ficheiro selecionado", "선택된 파일 없음"],
+  ["All", "全部", "すべて", "Alle", "Todos", "Tous", "Все", "Todos", "전체"],
+  ["System default", "系统默认", "システム既定", "Systemstandard", "Predeterminado del sistema", "Valeur système", "Системное значение", "Predefinição do sistema", "시스템 기본값"],
+  ["Choose an image to begin.", "选择图片以开始。", "画像を選択してください。", "Wählen Sie zunächst ein Bild.", "Elija una imagen para empezar.", "Choisissez une image pour commencer.", "Выберите изображение, чтобы начать.", "Escolha uma imagem para começar.", "시작하려면 이미지를 선택하세요."],
+  ["Uploaded image with detected face overlays", "带有人脸检测框的已上传图片", "検出枠を重ねたアップロード画像", "Hochgeladenes Bild mit erkannten Gesichtern", "Imagen subida con recuadros faciales", "Image importée avec cadres de visages", "Загруженное изображение с рамками лиц", "Imagem enviada com caixas de rosto", "검출 얼굴 오버레이가 있는 업로드 이미지"],
+  ["Image preview", "图片预览", "画像プレビュー", "Bildvorschau", "Vista previa", "Aperçu de l’image", "Предпросмотр изображения", "Pré-visualização da imagem", "이미지 미리보기"],
+  ["ONE-TO-ONE", "一对一", "1 対 1", "EINS-ZU-EINS", "UNO A UNO", "UN À UN", "ОДИН К ОДНОМУ", "UM PARA UM", "일대일"],
+  ["The selected detection profile chooses one face from each image for cosine comparison.", "所选检测配置会从每张图片中挑选一张人脸进行 cosine 比对。", "選択した検出プロファイルが各画像から 1 顔を選び、コサイン比較します。", "Das gewählte Erkennungsprofil wählt je Bild ein Gesicht für den Cosinusvergleich.", "El perfil de detección elige un rostro de cada imagen para la comparación coseno.", "Le profil de détection choisit un visage par image pour la comparaison cosinus.", "Выбранный профиль выбирает по одному лицу из каждого изображения для cosine-сравнения.", "O perfil de deteção escolhe um rosto de cada imagem para comparação cosine.", "선택한 검출 프로필이 각 이미지에서 얼굴 하나를 골라 코사인 비교합니다."],
+  ["Source image face overlay", "源图片人脸框", "ソース画像の顔オーバーレイ", "Gesichtsrahmen im Quellbild", "Recuadro facial de origen", "Cadre facial de l’image source", "Рамка лица на исходном изображении", "Caixa facial da imagem de origem", "소스 이미지 얼굴 오버레이"],
+  ["Source preview", "源图片预览", "ソースプレビュー", "Quellvorschau", "Vista previa de origen", "Aperçu source", "Предпросмотр исходного изображения", "Pré-visualização da origem", "소스 미리보기"],
+  ["Target image face overlay", "目标图片人脸框", "ターゲット画像の顔オーバーレイ", "Gesichtsrahmen im Zielbild", "Recuadro facial de destino", "Cadre facial de l’image cible", "Рамка лица на целевом изображении", "Caixa facial da imagem de destino", "대상 이미지 얼굴 오버레이"],
+  ["Target preview", "目标图片预览", "ターゲットプレビュー", "Zielvorschau", "Vista previa de destino", "Aperçu cible", "Предпросмотр целевого изображения", "Pré-visualização do destino", "대상 미리보기"],
+  ["ONE-TO-MANY", "一对多", "1 対多", "EINS-ZU-VIELE", "UNO A MUCHOS", "UN À PLUSIEURS", "ОДИН КО МНОГИМ", "UM PARA MUITOS", "일대다"],
+  ["Find the closest enrolled people using the Collection single-face selection strategy.", "使用人员库的单脸挑选策略查找最接近的已注册人员。", "コレクションの単一顔選択を使って最も近い登録人物を検索します。", "Sucht mit der Ein-Gesicht-Auswahl der Sammlung nach den ähnlichsten registrierten Personen.", "Busca las personas registradas más cercanas usando la selección de rostro de la colección.", "Recherche les personnes inscrites les plus proches avec la sélection de visage de la collection.", "Ищет ближайших зарегистрированных людей с помощью стратегии выбора лица коллекции.", "Procura as pessoas registadas mais próximas usando a seleção de rosto da coleção.", "컬렉션의 단일 얼굴 선택 전략으로 가장 가까운 등록 인물을 검색합니다."],
+  ["blank = collection default", "留空 = 人员库默认值", "空欄 = コレクション既定値", "leer = Sammlungsstandard", "vacío = valor de la colección", "vide = valeur de la collection", "пусто = значение коллекции", "vazio = predefinição da coleção", "비움 = 컬렉션 기본값"],
+  ["Choose a collection and image.", "请选择人员库和图片。", "コレクションと画像を選択してください。", "Wählen Sie eine Sammlung und ein Bild.", "Elija una colección y una imagen.", "Choisissez une collection et une image.", "Выберите коллекцию и изображение.", "Escolha uma coleção e uma imagem.", "컬렉션과 이미지를 선택하세요."],
+  ["Search image face overlay", "搜索图片人脸框", "検索画像の顔オーバーレイ", "Gesichtsrahmen im Suchbild", "Recuadro facial de búsqueda", "Cadre facial de l’image de recherche", "Рамка лица на поисковом изображении", "Caixa facial da imagem de pesquisa", "검색 이미지 얼굴 오버레이"],
+  ["Query preview", "查询预览", "検索プレビュー", "Suchvorschau", "Vista previa de consulta", "Aperçu de la requête", "Предпросмотр запроса", "Pré-visualização da consulta", "쿼리 미리보기"],
+  ["Play a local video and identify every detected face. The video stays in your browser; only sampled JPEG frames are sent to this server.", "播放本地视频并识别每张检测到的人脸。视频保留在浏览器中，仅向服务器发送采样的 JPEG 帧。", "ローカル動画を再生し、検出された顔を識別します。動画はブラウザー内に留まり、サンプリングした JPEG フレームだけをサーバーへ送信します。", "Spielt ein lokales Video ab und identifiziert erkannte Gesichter. Das Video bleibt im Browser; nur ausgewählte JPEG-Bilder werden gesendet.", "Reproduce un vídeo local e identifica cada rostro. El vídeo permanece en el navegador; solo se envían fotogramas JPEG muestreados.", "Lit une vidéo locale et identifie chaque visage. La vidéo reste dans le navigateur ; seuls des JPEG échantillonnés sont envoyés.", "Воспроизводит локальное видео и распознаёт найденные лица. Видео остаётся в браузере; отправляются только выбранные JPEG-кадры.", "Reproduz um vídeo local e identifica cada rosto. O vídeo fica no navegador; apenas fotogramas JPEG amostrados são enviados.", "로컬 동영상을 재생하고 검출된 얼굴을 식별합니다. 동영상은 브라우저에 남고 샘플링된 JPEG 프레임만 서버로 전송됩니다."],
+  ["Video recognition face overlays", "视频识别人脸框", "動画認識の顔オーバーレイ", "Gesichtsrahmen der Videoerkennung", "Recuadros de reconocimiento de vídeo", "Cadres de reconnaissance vidéo", "Рамки распознавания видео", "Caixas de reconhecimento de vídeo", "동영상 인식 얼굴 오버레이"],
+  ["Choose a local video. It stays in this browser; only sampled JPEG frames are sent for recognition.", "选择本地视频。视频保留在此浏览器中，仅发送采样 JPEG 帧用于识别。", "ローカル動画を選択します。動画はこのブラウザー内に留まり、認識用の JPEG フレームだけが送信されます。", "Wählen Sie ein lokales Video. Es bleibt im Browser; nur ausgewählte JPEG-Bilder werden zur Erkennung gesendet.", "Elija un vídeo local. Permanecerá en el navegador; solo se enviarán fotogramas JPEG muestreados.", "Choisissez une vidéo locale. Elle reste dans ce navigateur ; seuls des JPEG échantillonnés sont envoyés.", "Выберите локальное видео. Оно останется в браузере; для распознавания отправляются только выбранные JPEG-кадры.", "Escolha um vídeo local. Fica neste navegador; apenas fotogramas JPEG amostrados são enviados.", "로컬 동영상을 선택하세요. 동영상은 브라우저에 남고 샘플링된 JPEG 프레임만 인식용으로 전송됩니다."],
+  ["MP4, WebM, MOV, or MKV supported by this browser", "此浏览器支持的 MP4、WebM、MOV 或 MKV", "このブラウザーが対応する MP4、WebM、MOV、MKV", "Vom Browser unterstütztes MP4, WebM, MOV oder MKV", "MP4, WebM, MOV o MKV compatible con el navegador", "MP4, WebM, MOV ou MKV pris en charge par le navigateur", "Поддерживаемые браузером MP4, WebM, MOV или MKV", "MP4, WebM, MOV ou MKV suportado pelo navegador", "브라우저가 지원하는 MP4, WebM, MOV 또는 MKV"],
+  ["ADMINISTRATION", "管理", "管理", "VERWALTUNG", "ADMINISTRACIÓN", "ADMINISTRATION", "АДМИНИСТРИРОВАНИЕ", "ADMINISTRAÇÃO", "관리"],
+  ["Runtime diagnostics and safe system configuration. Secrets and embeddings are never displayed.", "显示运行时诊断和安全的系统配置，绝不展示密钥或特征。", "ランタイム診断と安全なシステム設定を表示します。秘密情報や特徴量は表示しません。", "Laufzeitdiagnose und sichere Systemkonfiguration. Geheimnisse und Embeddings werden nie angezeigt.", "Diagnóstico de ejecución y configuración segura. Nunca se muestran secretos ni embeddings.", "Diagnostic d’exécution et configuration sûre. Les secrets et embeddings ne sont jamais affichés.", "Диагностика и безопасная конфигурация. Секреты и эмбеддинги никогда не показываются.", "Diagnóstico e configuração segura. Segredos e embeddings nunca são apresentados.", "런타임 진단과 안전한 시스템 구성을 표시하며 비밀 정보와 임베딩은 노출하지 않습니다."],
+  ["Provider", "执行提供程序", "プロバイダー", "Provider", "Proveedor", "Fournisseur", "Провайдер", "Fornecedor", "실행 공급자"],
+  ["API key in UI", "UI 中的 API Key", "UI の API キー", "API-Schlüssel in der UI", "Clave API en la UI", "Clé API dans l’interface", "API-ключ в интерфейсе", "Chave API na interface", "UI의 API 키"],
+  ["MODELS", "模型", "モデル", "MODELLE", "MODELOS", "MODÈLES", "МОДЕЛИ", "MODELOS", "모델"],
+  ["DOCUMENTATION", "文档", "ドキュメント", "DOKUMENTATION", "DOCUMENTACIÓN", "DOCUMENTATION", "ДОКУМЕНТАЦИЯ", "DOCUMENTAÇÃO", "문서"],
+  ["Documentation", "文档", "ドキュメント", "Dokumentation", "Documentación", "Documentation", "Документация", "Documentação", "문서"],
+  ["Requests from this tab will send the key as a Bearer token. It is not written to cookies, local storage, logs, or the URL.", "此标签页的请求会将 Key 作为 Bearer Token 发送；不会写入 Cookie、本地存储、日志或 URL。", "このタブのリクエストはキーを Bearer トークンとして送信します。Cookie、ローカルストレージ、ログ、URL には保存しません。", "Anfragen dieses Tabs senden den Schlüssel als Bearer-Token. Er wird weder in Cookies, lokalem Speicher, Logs noch der URL gespeichert.", "Las solicitudes de esta pestaña envían la clave como token Bearer. No se guarda en cookies, almacenamiento local, registros ni URL.", "Les requêtes de cet onglet envoient la clé comme jeton Bearer. Elle n’est inscrite ni dans les cookies, le stockage local, les journaux ou l’URL.", "Запросы этой вкладки передают ключ как Bearer-токен. Он не записывается в cookie, локальное хранилище, журналы или URL.", "Os pedidos deste separador enviam a chave como token Bearer. Não é guardada em cookies, armazenamento local, registos ou URL.", "이 탭의 요청은 키를 Bearer 토큰으로 전송하며 쿠키, 로컬 저장소, 로그 또는 URL에 기록하지 않습니다."],
+  ["COLLECTION", "人员库", "コレクション", "SAMMLUNG", "COLECCIÓN", "COLLECTION", "КОЛЛЕКЦИЯ", "COLEÇÃO", "컬렉션"],
+  ["Comma-separated width×height pairs.", "使用逗号分隔的宽×高组合。", "幅×高さをカンマで区切ります。", "Kommagetrennte Breite×Höhe-Paare.", "Pares ancho×alto separados por comas.", "Paires largeur×hauteur séparées par des virgules.", "Пары ширина×высота через запятую.", "Pares largura×altura separados por vírgulas.", "쉼표로 구분한 너비×높이 쌍입니다."],
+  ["Applies to later registrations; an already-running request may use the previous setting.", "适用于后续注册；已经在运行的请求可能仍使用旧设置。", "以降の登録に適用されます。実行中のリクエストは以前の設定を使う場合があります。", "Gilt für spätere Registrierungen; eine laufende Anfrage kann noch die vorige Einstellung verwenden.", "Se aplica a registros posteriores; una solicitud en curso puede usar el ajuste anterior.", "S’applique aux inscriptions suivantes ; une requête en cours peut encore utiliser l’ancien réglage.", "Применяется к последующим регистрациям; уже выполняющийся запрос может использовать прежнюю настройку.", "Aplica-se aos registos seguintes; um pedido em curso pode usar a definição anterior.", "이후 등록부터 적용되며 이미 실행 중인 요청은 이전 설정을 사용할 수 있습니다."],
+  ["Unavailable", "不可用", "利用不可", "Nicht verfügbar", "No disponible", "Indisponible", "Недоступно", "Indisponível", "사용할 수 없음"],
+  ["Unknown", "未知", "不明", "Unbekannt", "Desconocido", "Inconnu", "Неизвестно", "Desconhecido", "알 수 없음"],
+  ["Not loaded", "未加载", "未ロード", "Nicht geladen", "No cargado", "Non chargé", "Не загружено", "Não carregado", "로드되지 않음"],
+  ["healthy", "健康", "正常", "fehlerfrei", "saludable", "sain", "исправно", "saudável", "정상"],
+  ["ready", "就绪", "準備完了", "bereit", "listo", "prêt", "готово", "pronto", "준비됨"],
+  ["unavailable", "不可用", "利用不可", "nicht verfügbar", "no disponible", "indisponible", "недоступно", "indisponível", "사용할 수 없음"],
+  ["unknown", "未知", "不明", "unbekannt", "desconocido", "inconnu", "неизвестно", "desconhecido", "알 수 없음"],
+  ["ok", "正常", "正常", "OK", "correcto", "OK", "норма", "OK", "정상"],
+  ["On", "开启", "オン", "Ein", "Activado", "Activé", "Вкл.", "Ligado", "켜짐"],
+  ["Off", "关闭", "オフ", "Aus", "Desactivado", "Désactivé", "Выкл.", "Desligado", "꺼짐"],
+  ["Edit", "编辑", "編集", "Bearbeiten", "Editar", "Modifier", "Изменить", "Editar", "편집"],
+  ["Delete", "删除", "削除", "Löschen", "Eliminar", "Supprimer", "Удалить", "Eliminar", "삭제"],
+  ["Save", "保存", "保存", "Speichern", "Guardar", "Enregistrer", "Сохранить", "Guardar", "저장"],
+  ["Request: {id}", "请求：{id}", "リクエスト：{id}", "Anfrage: {id}", "Solicitud: {id}", "Requête : {id}", "Запрос: {id}", "Pedido: {id}", "요청: {id}"],
+  ["Unexpected error.", "发生意外错误。", "予期しないエラーが発生しました。", "Unerwarteter Fehler.", "Error inesperado.", "Erreur inattendue.", "Непредвиденная ошибка.", "Erro inesperado.", "예기치 않은 오류가 발생했습니다."],
+  ["{count} collection", "{count} 个人员库", "{count} コレクション", "{count} Sammlung", "{count} colección", "{count} collection", "{count} коллекция", "{count} coleção", "컬렉션 {count}개"],
+  ["{count} collections", "{count} 个人员库", "{count} コレクション", "{count} Sammlungen", "{count} colecciones", "{count} collections", "{count} коллекций", "{count} coleções", "컬렉션 {count}개"],
+  ["{count} person", "{count} 人", "{count} 人", "{count} Person", "{count} persona", "{count} personne", "{count} человек", "{count} pessoa", "인물 {count}명"],
+  ["{count} people", "{count} 人", "{count} 人", "{count} Personen", "{count} personas", "{count} personnes", "{count} человек", "{count} pessoas", "인물 {count}명"],
+  ["{count} face", "{count} 张人脸", "{count} 顔", "{count} Gesicht", "{count} rostro", "{count} visage", "{count} лицо", "{count} rosto", "얼굴 {count}개"],
+  ["{count} faces", "{count} 张人脸", "{count} 顔", "{count} Gesichter", "{count} rostros", "{count} visages", "{count} лиц", "{count} rostos", "얼굴 {count}개"],
+  ["{count} files selected", "已选择 {count} 个文件", "{count} ファイルを選択", "{count} Dateien ausgewählt", "{count} archivos seleccionados", "{count} fichiers sélectionnés", "Выбрано файлов: {count}", "{count} ficheiros selecionados", "파일 {count}개 선택됨"],
+  ["{count} image was rejected", "{count} 张图片被拒绝", "{count} 枚の画像が拒否されました", "{count} Bild wurde abgelehnt", "Se rechazó {count} imagen", "{count} image a été rejetée", "Отклонено изображений: {count}", "{count} imagem foi rejeitada", "이미지 {count}개가 거부됨"],
+  ["{count} images were rejected", "{count} 张图片被拒绝", "{count} 枚の画像が拒否されました", "{count} Bilder wurden abgelehnt", "Se rechazaron {count} imágenes", "{count} images ont été rejetées", "Отклонено изображений: {count}", "{count} imagens foram rejeitadas", "이미지 {count}개가 거부됨"],
+  ["Image {index}", "图片 {index}", "画像 {index}", "Bild {index}", "Imagen {index}", "Image {index}", "Изображение {index}", "Imagem {index}", "이미지 {index}"],
+  ["rejected", "已拒绝", "拒否", "abgelehnt", "rechazado", "rejeté", "отклонено", "rejeitado", "거부됨"],
+  ["Unnamed collection", "未命名人员库", "名称未設定コレクション", "Unbenannte Sammlung", "Colección sin nombre", "Collection sans nom", "Коллекция без имени", "Coleção sem nome", "이름 없는 컬렉션"],
+  ["Collection “{name}” is not empty. Delete it and all people and face samples?", "人员库“{name}”不为空。是否删除它及其中所有人员和人脸样本？", "コレクション「{name}」は空ではありません。人物と顔サンプルをすべて削除しますか？", "Die Sammlung „{name}“ ist nicht leer. Sammlung, Personen und Gesichtsproben löschen?", "La colección «{name}» no está vacía. ¿Eliminarla con todas las personas y muestras?", "La collection « {name} » n’est pas vide. La supprimer avec toutes les personnes et tous les échantillons ?", "Коллекция «{name}» не пуста. Удалить её вместе со всеми людьми и образцами лиц?", "A coleção “{name}” não está vazia. Eliminá-la com todas as pessoas e amostras?", "컬렉션 ‘{name}’이 비어 있지 않습니다. 모든 인물과 얼굴 샘플을 함께 삭제할까요?"],
+  ["Delete collection “{name}”?", "删除人员库“{name}”？", "コレクション「{name}」を削除しますか？", "Sammlung „{name}“ löschen?", "¿Eliminar la colección «{name}»?", "Supprimer la collection « {name} » ?", "Удалить коллекцию «{name}»?", "Eliminar a coleção “{name}”?", "컬렉션 ‘{name}’을 삭제할까요?"],
+  ["Collection created", "人员库已创建", "コレクションを作成しました", "Sammlung erstellt", "Colección creada", "Collection créée", "Коллекция создана", "Coleção criada", "컬렉션 생성됨"],
+  ["Collection updated", "人员库已更新", "コレクションを更新しました", "Sammlung aktualisiert", "Colección actualizada", "Collection mise à jour", "Коллекция обновлена", "Coleção atualizada", "컬렉션 업데이트됨"],
+  ["Collection deleted", "人员库已删除", "コレクションを削除しました", "Sammlung gelöscht", "Colección eliminada", "Collection supprimée", "Коллекция удалена", "Coleção eliminada", "컬렉션 삭제됨"],
+  ["People in {collection}", "{collection} 中的人员", "{collection} の人物", "Personen in {collection}", "Personas en {collection}", "Personnes dans {collection}", "Люди в {collection}", "Pessoas em {collection}", "{collection}의 인물"],
+  ["No people found", "未找到人员", "人物が見つかりません", "Keine Personen gefunden", "No se encontraron personas", "Aucune personne trouvée", "Люди не найдены", "Nenhuma pessoa encontrada", "인물을 찾지 못했습니다"],
+  ["Register a person or change the search.", "请注册人员或修改搜索条件。", "人物を登録するか検索条件を変更してください。", "Registrieren Sie eine Person oder ändern Sie die Suche.", "Registre una persona o cambie la búsqueda.", "Inscrivez une personne ou modifiez la recherche.", "Зарегистрируйте человека или измените поиск.", "Registe uma pessoa ou altere a pesquisa.", "인물을 등록하거나 검색 조건을 변경하세요."],
+  ["PERSON", "人员", "人物", "PERSON", "PERSONA", "PERSONNE", "ЧЕЛОВЕК", "PESSOA", "인물"],
+  ["METADATA", "元数据", "メタデータ", "METADATEN", "METADATOS", "MÉTADONNÉES", "МЕТАДАННЫЕ", "METADADOS", "메타데이터"],
+  ["FACE SAMPLES", "人脸样本", "顔サンプル", "GESICHTSPROBEN", "MUESTRAS FACIALES", "ÉCHANTILLONS FACIAUX", "ОБРАЗЦЫ ЛИЦ", "AMOSTRAS FACIAIS", "얼굴 샘플"],
+  ["Registered faces", "已注册人脸", "登録済みの顔", "Registrierte Gesichter", "Rostros registrados", "Visages inscrits", "Зарегистрированные лица", "Rostos registados", "등록된 얼굴"],
+  ["Registered faces ({count})", "已注册人脸（{count}）", "登録済みの顔（{count}）", "Registrierte Gesichter ({count})", "Rostros registrados ({count})", "Visages inscrits ({count})", "Зарегистрированные лица ({count})", "Rostos registados ({count})", "등록된 얼굴 ({count})"],
+  ["Add registration photos", "添加注册照片", "登録写真を追加", "Registrierungsfotos hinzufügen", "Añadir fotos de registro", "Ajouter des photos d’inscription", "Добавить регистрационные фото", "Adicionar fotos de registo", "등록 사진 추가"],
+  ["Off uses the Collection single-face selection strategy; standard and strict require exactly one.", "关闭时使用人员库的单脸挑选策略；标准和严格模式要求恰好一张人脸。", "オフではコレクションの単一顔選択を使用し、標準と厳格では顔が 1 つだけ必要です。", "Aus verwendet die Ein-Gesicht-Auswahl der Sammlung; Standard und Strikt verlangen genau ein Gesicht.", "Desactivado usa la selección de la colección; estándar y estricto exigen exactamente un rostro.", "Désactivé utilise la sélection de la collection ; standard et strict exigent exactement un visage.", "«Выкл.» использует стратегию коллекции; стандартный и строгий режимы требуют ровно одно лицо.", "Desligado usa a seleção da coleção; normal e estrito exigem exatamente um rosto.", "꺼짐은 컬렉션 선택 전략을 사용하며 표준 및 엄격 모드는 얼굴이 정확히 하나여야 합니다."],
+  ["The server trusts the image/vector pairing and does not re-extract the feature.", "服务器信任图片与向量的对应关系，不会重新抽取特征。", "サーバーは画像とベクトルの対応を信頼し、特徴量を再抽出しません。", "Der Server vertraut der Bild/Vektor-Zuordnung und extrahiert das Merkmal nicht erneut.", "El servidor confía en la relación imagen/vector y no vuelve a extraer el vector.", "Le serveur fait confiance à l’association image/vecteur et ne réextrait pas le vecteur.", "Сервер доверяет соответствию изображения и вектора и не извлекает признак повторно.", "O servidor confia na associação imagem/vetor e não volta a extrair o vetor.", "서버는 이미지와 벡터의 연결을 신뢰하며 특징을 다시 추출하지 않습니다."],
+  ["Add face samples", "添加人脸样本", "顔サンプルを追加", "Gesichtsproben hinzufügen", "Añadir muestras faciales", "Ajouter des échantillons faciaux", "Добавить образцы лиц", "Adicionar amostras faciais", "얼굴 샘플 추가"],
+  ["No face samples", "没有人脸样本", "顔サンプルがありません", "Keine Gesichtsproben", "No hay muestras faciales", "Aucun échantillon facial", "Нет образцов лиц", "Sem amostras faciais", "얼굴 샘플 없음"],
+  ["Add one or more clear photos.", "请添加一张或多张清晰照片。", "鮮明な写真を 1 枚以上追加してください。", "Fügen Sie ein oder mehrere klare Fotos hinzu.", "Añada una o varias fotos nítidas.", "Ajoutez une ou plusieurs photos nettes.", "Добавьте одно или несколько чётких фото.", "Adicione uma ou mais fotos nítidas.", "선명한 사진을 하나 이상 추가하세요."],
+  ["Detection", "检测", "検出", "Erkennung", "Detección", "Détection", "Детекция", "Deteção", "검출"],
+  ["Quality", "质量", "品質", "Qualität", "Calidad", "Qualité", "Качество", "Qualidade", "품질"],
+  ["Embedding", "特征", "特徴量", "Embedding", "Embedding", "Embedding", "Эмбеддинг", "Embedding", "임베딩"],
+  ["External trusted", "可信外部特征", "信頼済み外部特徴量", "Extern vertrauenswürdig", "Externo confiable", "Externe approuvée", "Доверенный внешний", "Externo confiável", "신뢰된 외부 특징"],
+  ["Server extracted", "服务器抽取", "サーバー抽出", "Vom Server extrahiert", "Extraído por el servidor", "Extraite par le serveur", "Извлечено сервером", "Extraído pelo servidor", "서버 추출"],
+  ["Stored crop", "已存裁剪图", "保存済みクロップ", "Gespeicherter Ausschnitt", "Recorte guardado", "Recadrage enregistré", "Сохранённый кроп", "Recorte guardado", "저장된 크롭"],
+  ["Available", "可用", "利用可能", "Verfügbar", "Disponible", "Disponible", "Доступно", "Disponível", "사용 가능"],
+  ["Not saved", "未保存", "未保存", "Nicht gespeichert", "No guardado", "Non enregistré", "Не сохранено", "Não guardado", "저장되지 않음"],
+  ["Created", "创建时间", "作成日時", "Erstellt", "Creado", "Créé", "Создано", "Criado", "생성 시각"],
+  ["View crop", "查看裁剪图", "クロップを表示", "Ausschnitt anzeigen", "Ver recorte", "Voir le recadrage", "Показать кроп", "Ver recorte", "크롭 보기"],
+  ["Hide crop", "隐藏裁剪图", "クロップを隠す", "Ausschnitt ausblenden", "Ocultar recorte", "Masquer le recadrage", "Скрыть кроп", "Ocultar recorte", "크롭 숨기기"],
+  ["Delete sample", "删除样本", "サンプルを削除", "Probe löschen", "Eliminar muestra", "Supprimer l’échantillon", "Удалить образец", "Eliminar amostra", "샘플 삭제"],
+  ["Saved crop for face sample {id}", "人脸样本 {id} 的已保存裁剪图", "顔サンプル {id} の保存済みクロップ", "Gespeicherter Ausschnitt der Gesichtsprobe {id}", "Recorte guardado de la muestra {id}", "Recadrage enregistré de l’échantillon {id}", "Сохранённый кроп образца лица {id}", "Recorte guardado da amostra {id}", "얼굴 샘플 {id}의 저장된 크롭"],
+  ["Delete “{name}” and all registered face samples?", "删除“{name}”及其所有已注册人脸样本？", "「{name}」と登録済み顔サンプルをすべて削除しますか？", "„{name}“ und alle registrierten Gesichtsproben löschen?", "¿Eliminar «{name}» y todas sus muestras faciales?", "Supprimer « {name} » et tous ses échantillons faciaux ?", "Удалить «{name}» и все зарегистрированные образцы лиц?", "Eliminar “{name}” e todas as amostras faciais?", "‘{name}’과 등록된 모든 얼굴 샘플을 삭제할까요?"],
+  ["Delete face sample {id}?", "删除人脸样本 {id}？", "顔サンプル {id} を削除しますか？", "Gesichtsprobe {id} löschen?", "¿Eliminar la muestra facial {id}?", "Supprimer l’échantillon facial {id} ?", "Удалить образец лица {id}?", "Eliminar a amostra facial {id}?", "얼굴 샘플 {id}을 삭제할까요?"],
+  ["Person updated", "人员已更新", "人物を更新しました", "Person aktualisiert", "Persona actualizada", "Personne mise à jour", "Человек обновлён", "Pessoa atualizada", "인물 업데이트됨"],
+  ["Person deleted", "人员已删除", "人物を削除しました", "Person gelöscht", "Persona eliminada", "Personne supprimée", "Человек удалён", "Pessoa eliminada", "인물 삭제됨"],
+  ["Registration complete", "注册完成", "登録完了", "Registrierung abgeschlossen", "Registro completado", "Inscription terminée", "Регистрация завершена", "Registo concluído", "등록 완료"],
+  ["{count} face sample added.", "已添加 {count} 个人脸样本。", "顔サンプルを {count} 件追加しました。", "{count} Gesichtsprobe hinzugefügt.", "Se añadió {count} muestra facial.", "{count} échantillon facial ajouté.", "Добавлен образец лица: {count}.", "Foi adicionada {count} amostra facial.", "얼굴 샘플 {count}개를 추가했습니다."],
+  ["{count} face samples added.", "已添加 {count} 个人脸样本。", "顔サンプルを {count} 件追加しました。", "{count} Gesichtsproben hinzugefügt.", "Se añadieron {count} muestras faciales.", "{count} échantillons faciaux ajoutés.", "Добавлено образцов лиц: {count}.", "Foram adicionadas {count} amostras faciais.", "얼굴 샘플 {count}개를 추가했습니다."],
+  ["Face sample deleted", "人脸样本已删除", "顔サンプルを削除しました", "Gesichtsprobe gelöscht", "Muestra facial eliminada", "Échantillon facial supprimé", "Образец лица удалён", "Amostra facial eliminada", "얼굴 샘플 삭제됨"],
+  ["Detection {score}", "检测 {score}", "検出 {score}", "Erkennung {score}", "Detección {score}", "Détection {score}", "Детекция {score}", "Deteção {score}", "검출 {score}"],
+  ["Quality {score}", "质量 {score}", "品質 {score}", "Qualität {score}", "Calidad {score}", "Qualité {score}", "Качество {score}", "Qualidade {score}", "품질 {score}"],
+  ["Sharpness {score}", "清晰度 {score}", "鮮明度 {score}", "Schärfe {score}", "Nitidez {score}", "Netteté {score}", "Резкость {score}", "Nitidez {score}", "선명도 {score}"],
+  ["Brightness {score}", "亮度 {score}", "明るさ {score}", "Helligkeit {score}", "Brillo {score}", "Luminosité {score}", "Яркость {score}", "Luminosidade {score}", "밝기 {score}"],
+  ["{count} face detected · {duration}", "检测到 {count} 张人脸 · {duration}", "{count} 顔を検出 · {duration}", "{count} Gesicht erkannt · {duration}", "{count} rostro detectado · {duration}", "{count} visage détecté · {duration}", "Обнаружено лиц: {count} · {duration}", "{count} rosto detetado · {duration}", "얼굴 {count}개 검출 · {duration}"],
+  ["{count} faces detected · {duration}", "检测到 {count} 张人脸 · {duration}", "{count} 顔を検出 · {duration}", "{count} Gesichter erkannt · {duration}", "{count} rostros detectados · {duration}", "{count} visages détectés · {duration}", "Обнаружено лиц: {count} · {duration}", "{count} rostos detetados · {duration}", "얼굴 {count}개 검출 · {duration}"],
+  ["Match", "匹配", "一致", "Übereinstimmung", "Coincidencia", "Correspondance", "Совпадение", "Correspondência", "일치"],
+  ["No match", "不匹配", "不一致", "Keine Übereinstimmung", "Sin coincidencia", "Aucune correspondance", "Нет совпадения", "Sem correspondência", "불일치"],
+  ["{similarity} similarity · threshold {threshold} · {duration}", "相似度 {similarity} · 阈值 {threshold} · {duration}", "類似度 {similarity} · しきい値 {threshold} · {duration}", "Ähnlichkeit {similarity} · Schwelle {threshold} · {duration}", "Similitud {similarity} · umbral {threshold} · {duration}", "Similarité {similarity} · seuil {threshold} · {duration}", "Сходство {similarity} · порог {threshold} · {duration}", "Similaridade {similarity} · limiar {threshold} · {duration}", "유사도 {similarity} · 임계값 {threshold} · {duration}"],
+  ["No matches above threshold", "没有超过阈值的匹配", "しきい値以上の一致はありません", "Keine Treffer über der Schwelle", "No hay coincidencias por encima del umbral", "Aucune correspondance au-dessus du seuil", "Совпадений выше порога нет", "Sem correspondências acima do limiar", "임계값 이상의 일치 항목이 없습니다"],
+  ["Try a clearer image or review the collection threshold.", "请尝试更清晰的图片，或检查人员库阈值。", "より鮮明な画像を試すか、コレクションのしきい値を確認してください。", "Versuchen Sie ein klareres Bild oder prüfen Sie die Sammlungsschwelle.", "Pruebe con una imagen más nítida o revise el umbral de la colección.", "Essayez une image plus nette ou vérifiez le seuil de la collection.", "Попробуйте более чёткое изображение или проверьте порог коллекции.", "Experimente uma imagem mais nítida ou reveja o limiar da coleção.", "더 선명한 이미지를 사용하거나 컬렉션 임계값을 확인하세요."],
+  ["Face {id}", "人脸 {id}", "顔 {id}", "Gesicht {id}", "Rostro {id}", "Visage {id}", "Лицо {id}", "Rosto {id}", "얼굴 {id}"],
+  ["Search face", "查询人脸", "検索顔", "Suchgesicht", "Rostro de búsqueda", "Visage recherché", "Искомое лицо", "Rosto pesquisado", "검색 얼굴"],
+  ["{count} match · threshold {threshold} · {duration}", "{count} 个匹配 · 阈值 {threshold} · {duration}", "{count} 件一致 · しきい値 {threshold} · {duration}", "{count} Treffer · Schwelle {threshold} · {duration}", "{count} coincidencia · umbral {threshold} · {duration}", "{count} correspondance · seuil {threshold} · {duration}", "Совпадений: {count} · порог {threshold} · {duration}", "{count} correspondência · limiar {threshold} · {duration}", "일치 {count}건 · 임계값 {threshold} · {duration}"],
+  ["{count} matches · threshold {threshold} · {duration}", "{count} 个匹配 · 阈值 {threshold} · {duration}", "{count} 件一致 · しきい値 {threshold} · {duration}", "{count} Treffer · Schwelle {threshold} · {duration}", "{count} coincidencias · umbral {threshold} · {duration}", "{count} correspondances · seuil {threshold} · {duration}", "Совпадений: {count} · порог {threshold} · {duration}", "{count} correspondências · limiar {threshold} · {duration}", "일치 {count}건 · 임계값 {threshold} · {duration}"],
+  ["Image ready. Press Detect faces to process it.", "图片已就绪。点击“检测人脸”开始处理。", "画像の準備ができました。「顔を検出」を押してください。", "Bild bereit. Klicken Sie auf „Gesichter erkennen“.", "Imagen lista. Pulse «Detectar rostros» para procesarla.", "Image prête. Cliquez sur « Détecter les visages ».", "Изображение готово. Нажмите «Обнаружить лица».", "Imagem pronta. Prima «Detetar rostos» para processar.", "이미지가 준비되었습니다. ‘얼굴 검출’을 눌러 처리하세요."],
+  ["Image ready. Press Search people to process it.", "图片已就绪。点击“搜索人员”开始处理。", "画像の準備ができました。「人物を検索」を押してください。", "Bild bereit. Klicken Sie auf „Personen suchen“.", "Imagen lista. Pulse «Buscar personas» para procesarla.", "Image prête. Cliquez sur « Rechercher des personnes ».", "Изображение готово. Нажмите «Найти людей».", "Imagem pronta. Prima «Pesquisar pessoas» para processar.", "이미지가 준비되었습니다. ‘인물 검색’을 눌러 처리하세요."],
+  ["{accepted} accepted, {rejected} rejected.", "成功 {accepted} 张，拒绝 {rejected} 张。", "受理 {accepted} 件、拒否 {rejected} 件。", "{accepted} akzeptiert, {rejected} abgelehnt.", "{accepted} aceptadas, {rejected} rechazadas.", "{accepted} acceptées, {rejected} rejetées.", "Принято: {accepted}, отклонено: {rejected}.", "{accepted} aceites, {rejected} rejeitadas.", "성공 {accepted}개, 거부 {rejected}개."],
+  ["Ready to compare", "可以开始比对", "比較の準備完了", "Bereit zum Vergleichen", "Listo para comparar", "Prêt à comparer", "Готово к сравнению", "Pronto para comparar", "비교 준비 완료"],
+  ["OS", "操作系统", "OS", "Betriebssystem", "Sistema operativo", "Système d’exploitation", "ОС", "Sistema operativo", "운영체제"],
+  ["Architecture", "架构", "アーキテクチャ", "Architektur", "Arquitectura", "Architecture", "Архитектура", "Arquitetura", "아키텍처"],
+  ["CPU", "CPU", "CPU", "CPU", "CPU", "CPU", "CPU", "CPU", "CPU"],
+  ["GPU", "GPU", "GPU", "GPU", "GPU", "GPU", "GPU", "GPU", "GPU"],
+  ["Not reported", "未报告", "未報告", "Nicht gemeldet", "No informado", "Non indiqué", "Не указано", "Não indicado", "보고되지 않음"],
+  ["Compute capability", "计算能力", "Compute Capability", "Compute Capability", "Capacidad de cómputo", "Capacité de calcul", "Вычислительная способность", "Capacidade de computação", "컴퓨팅 기능"],
+  ["NVIDIA driver", "NVIDIA 驱动", "NVIDIA ドライバー", "NVIDIA-Treiber", "Controlador NVIDIA", "Pilote NVIDIA", "Драйвер NVIDIA", "Controlador NVIDIA", "NVIDIA 드라이버"],
+  ["CUDA runtime", "CUDA 运行时", "CUDA ランタイム", "CUDA-Laufzeit", "Runtime de CUDA", "Runtime CUDA", "Среда выполнения CUDA", "Runtime CUDA", "CUDA 런타임"],
+  ["ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime", "ONNX Runtime"],
+  ["Detector NMS threshold", "检测器 NMS 阈值", "検出器 NMS しきい値", "Detektor-NMS-Schwelle", "Umbral NMS del detector", "Seuil NMS du détecteur", "NMS-порог детектора", "Limiar NMS do detetor", "검출기 NMS 임계값"],
+  ["Startup config", "启动配置", "起動設定", "Startkonfiguration", "Configuración de inicio", "Configuration de démarrage", "Конфигурация запуска", "Configuração de arranque", "시작 구성"],
+  ["built-in defaults", "内置默认值", "組み込み既定値", "integrierte Standardwerte", "valores integrados", "valeurs intégrées", "встроенные значения", "predefinições integradas", "내장 기본값"],
+  ["Database", "数据库", "データベース", "Datenbank", "Base de datos", "Base de données", "База данных", "Base de dados", "데이터베이스"],
+  ["Database path", "数据库路径", "データベースパス", "Datenbankpfad", "Ruta de la base de datos", "Chemin de la base", "Путь к базе данных", "Caminho da base de dados", "데이터베이스 경로"],
+  ["ready, writable", "就绪，可写", "準備完了、書き込み可", "bereit, beschreibbar", "lista, escritura habilitada", "prêt, accessible en écriture", "готово, доступна запись", "pronto, gravável", "준비됨, 쓰기 가능"],
+  ["ready, readable", "就绪，可读", "準備完了、読み取り可", "bereit, lesbar", "listo, legible", "prêt, lisible", "готово, доступно чтение", "pronto, legível", "준비됨, 읽기 가능"],
+  ["Authentication", "认证", "認証", "Authentifizierung", "Autenticación", "Authentification", "Аутентификация", "Autenticação", "인증"],
+  ["development mode (disabled)", "开发模式（已关闭认证）", "開発モード（認証無効）", "Entwicklungsmodus (deaktiviert)", "modo de desarrollo (desactivada)", "mode développement (désactivée)", "режим разработки (отключена)", "modo de desenvolvimento (desativada)", "개발 모드(인증 꺼짐)"],
+  ["enabled", "已启用", "有効", "aktiviert", "activada", "activée", "включено", "ativado", "활성화됨"],
+  ["disabled", "已禁用", "無効", "deaktiviert", "desactivada", "désactivé", "отключено", "desativado", "비활성화됨"],
+  ["Maximum image size", "最大图片大小", "最大画像サイズ", "Maximale Bildgröße", "Tamaño máximo de imagen", "Taille maximale d’image", "Максимальный размер изображения", "Tamanho máximo da imagem", "최대 이미지 크기"],
+  ["Maximum pixels", "最大像素数", "最大ピクセル数", "Maximale Pixelzahl", "Máximo de píxeles", "Nombre maximal de pixels", "Максимальное число пикселей", "Máximo de píxeis", "최대 픽셀 수"],
+  ["Request timeout", "请求超时", "リクエストタイムアウト", "Anfrage-Timeout", "Tiempo límite de solicitud", "Délai de requête", "Тайм-аут запроса", "Tempo limite do pedido", "요청 제한 시간"],
+  ["Inference concurrency", "推理并发数", "推論同時実行数", "Inferenz-Parallelität", "Concurrencia de inferencia", "Concurrence d’inférence", "Параллелизм инференса", "Concorrência de inferência", "추론 동시 실행 수"],
+  ["New-collection crop default", "新人员库裁剪图默认值", "新規コレクションのクロップ既定値", "Ausschnitt-Standard für neue Sammlungen", "Recorte predeterminado para nuevas colecciones", "Recadrage par défaut des nouvelles collections", "Кропы по умолчанию для новых коллекций", "Recorte predefinido para novas coleções", "새 컬렉션 크롭 기본값"],
+  ["Server API key", "服务器 API Key", "サーバー API キー", "Server-API-Schlüssel", "Clave API del servidor", "Clé API du serveur", "API-ключ сервера", "Chave API do servidor", "서버 API 키"],
+  ["configured", "已配置", "設定済み", "konfiguriert", "configurada", "configurée", "настроено", "configurada", "구성됨"],
+  ["not configured", "未配置", "未設定", "nicht konfiguriert", "sin configurar", "non configurée", "не настроено", "não configurada", "구성되지 않음"],
+  ["Configured", "已配置", "設定済み", "Konfiguriert", "Configurada", "Configurée", "Настроено", "Configurada", "구성됨"],
+  ["Task", "任务", "タスク", "Aufgabe", "Tarea", "Tâche", "Задача", "Tarefa", "작업"],
+  ["Version", "版本", "バージョン", "Version", "Versión", "Version", "Версия", "Versão", "버전"],
+  ["Digest", "摘要", "ダイジェスト", "Prüfsumme", "Resumen", "Empreinte", "Дайджест", "Resumo", "다이제스트"],
+  ["Input", "输入", "入力", "Eingabe", "Entrada", "Entrée", "Вход", "Entrada", "입력"],
+  ["No loaded models were reported.", "未报告已加载的模型。", "ロード済みモデルは報告されていません。", "Es wurden keine geladenen Modelle gemeldet.", "No se informó de ningún modelo cargado.", "Aucun modèle chargé n’a été signalé.", "Загруженные модели не указаны.", "Não foram indicados modelos carregados.", "로드된 모델이 보고되지 않았습니다."],
+  ["Unnamed model", "未命名模型", "名称未設定モデル", "Unbenanntes Modell", "Modelo sin nombre", "Modèle sans nom", "Модель без имени", "Modelo sem nome", "이름 없는 모델"],
+  ["{count} seconds", "{count} 秒", "{count} 秒", "{count} Sekunden", "{count} segundos", "{count} secondes", "{count} с", "{count} segundos", "{count}초"],
+  ["Unrecognized", "未识别", "未認識", "Nicht erkannt", "No reconocido", "Non reconnu", "Не распознано", "Não reconhecido", "인식되지 않음"],
+  ["Below threshold", "低于阈值", "しきい値未満", "Unter der Schwelle", "Por debajo del umbral", "Sous le seuil", "Ниже порога", "Abaixo do limiar", "임계값 미만"],
+  ["Recognition running", "识别运行中", "認識を実行中", "Erkennung läuft", "Reconocimiento en curso", "Reconnaissance en cours", "Распознавание выполняется", "Reconhecimento em curso", "인식 실행 중"],
+  ["Video is playing. Sampled frames are processed without saving the video.", "视频正在播放。系统处理采样帧，但不会保存视频。", "動画を再生中です。サンプリングしたフレームを処理し、動画は保存しません。", "Das Video wird abgespielt. Stichproben-Frames werden verarbeitet, ohne das Video zu speichern.", "El vídeo se está reproduciendo. Los fotogramas muestreados se procesan sin guardar el vídeo.", "La vidéo est en lecture. Les images échantillonnées sont traitées sans enregistrer la vidéo.", "Видео воспроизводится. Выборочные кадры обрабатываются без сохранения видео.", "O vídeo está a ser reproduzido. Os fotogramas amostrados são processados sem guardar o vídeo.", "비디오를 재생 중입니다. 샘플 프레임을 처리하며 비디오는 저장하지 않습니다."],
+  ["Recognition stopped. The video remains in this browser.", "识别已停止。视频仍保留在当前浏览器中。", "認識を停止しました。動画はこのブラウザ内に残ります。", "Erkennung gestoppt. Das Video bleibt in diesem Browser.", "Reconocimiento detenido. El vídeo permanece en este navegador.", "Reconnaissance arrêtée. La vidéo reste dans ce navigateur.", "Распознавание остановлено. Видео остаётся в этом браузере.", "Reconhecimento parado. O vídeo permanece neste navegador.", "인식이 중지되었습니다. 비디오는 이 브라우저에 남아 있습니다."],
+  ["Video finished. The last recognition result remains visible.", "视频播放完毕。最后一次识别结果仍会显示。", "動画が終了しました。最後の認識結果は表示されたままです。", "Video beendet. Das letzte Erkennungsergebnis bleibt sichtbar.", "El vídeo ha terminado. El último resultado sigue visible.", "La vidéo est terminée. Le dernier résultat reste visible.", "Видео завершено. Последний результат остаётся на экране.", "O vídeo terminou. O último resultado permanece visível.", "비디오가 끝났습니다. 마지막 인식 결과는 계속 표시됩니다."],
+  ["Video selected. Press Start recognition to begin.", "视频已选择。点击“开始识别”即可开始。", "動画を選択しました。「認識を開始」を押してください。", "Video ausgewählt. Klicken Sie auf „Erkennung starten“.", "Vídeo seleccionado. Pulse «Iniciar reconocimiento».", "Vidéo sélectionnée. Cliquez sur « Démarrer la reconnaissance ».", "Видео выбрано. Нажмите «Начать распознавание».", "Vídeo selecionado. Prima «Iniciar reconhecimento».", "비디오가 선택되었습니다. ‘인식 시작’을 누르세요."],
+  ["Select a video, then start recognition.", "选择视频，然后开始识别。", "動画を選択してから認識を開始してください。", "Wählen Sie ein Video und starten Sie dann die Erkennung.", "Seleccione un vídeo y después inicie el reconocimiento.", "Sélectionnez une vidéo, puis démarrez la reconnaissance.", "Выберите видео, затем запустите распознавание.", "Selecione um vídeo e inicie o reconhecimento.", "비디오를 선택한 다음 인식을 시작하세요."],
+  ["ID {id}", "ID {id}", "ID {id}", "ID {id}", "ID {id}", "ID {id}", "ID {id}", "ID {id}", "ID {id}"],
+  ["ID {id} · {similarity}", "ID {id} · 相似度 {similarity}", "ID {id} · 類似度 {similarity}", "ID {id} · Ähnlichkeit {similarity}", "ID {id} · similitud {similarity}", "ID {id} · similarité {similarity}", "ID {id} · сходство {similarity}", "ID {id} · similaridade {similarity}", "ID {id} · 유사도 {similarity}"],
+  ["{faces} faces · {recognized} recognized · {duration}", "{faces} 张人脸 · 识别出 {recognized} 张 · {duration}", "{faces} 顔 · {recognized} 件認識 · {duration}", "{faces} Gesichter · {recognized} erkannt · {duration}", "{faces} rostros · {recognized} reconocidos · {duration}", "{faces} visages · {recognized} reconnus · {duration}", "Лиц: {faces} · распознано: {recognized} · {duration}", "{faces} rostos · {recognized} reconhecidos · {duration}", "얼굴 {faces}개 · 인식 {recognized}개 · {duration}"],
+  ["Start recognition", "开始识别", "認識を開始", "Erkennung starten", "Iniciar reconocimiento", "Démarrer la reconnaissance", "Начать распознавание", "Iniciar reconhecimento", "인식 시작"],
+  ["Starting…", "正在启动…", "開始中…", "Wird gestartet…", "Iniciando…", "Démarrage…", "Запуск…", "A iniciar…", "시작 중…"],
+  ["API key configured", "API Key 已配置", "API キーを設定しました", "API-Schlüssel konfiguriert", "Clave API configurada", "Clé API configurée", "API-ключ настроен", "Chave API configurada", "API 키 구성됨"],
+  ["API key cleared", "API Key 已清除", "API キーを消去しました", "API-Schlüssel gelöscht", "Clave API borrada", "Clé API effacée", "API-ключ удалён", "Chave API limpa", "API 키 지워짐"],
+  ["It will be forgotten when this tab closes.", "关闭此标签页后，Key 会被清除。", "このタブを閉じるとキーは破棄されます。", "Beim Schließen dieses Tabs wird er verworfen.", "Se olvidará al cerrar esta pestaña.", "Elle sera oubliée à la fermeture de cet onglet.", "Ключ будет забыт при закрытии вкладки.", "Será esquecida quando este separador fechar.", "이 탭을 닫으면 키가 삭제됩니다."],
+  ["Working…", "处理中…", "処理中…", "Wird verarbeitet…", "Procesando…", "Traitement…", "Обработка…", "A processar…", "처리 중…"],
+  ["Loading…", "加载中…", "読み込み中…", "Wird geladen…", "Cargando…", "Chargement…", "Загрузка…", "A carregar…", "로드 중…"],
+  ["Loading documentation…", "正在加载文档…", "ドキュメントを読み込み中…", "Dokumentation wird geladen…", "Cargando documentación…", "Chargement de la documentation…", "Загрузка документации…", "A carregar documentação…", "문서 로드 중…"],
+  ["Creating…", "正在创建…", "作成中…", "Wird erstellt…", "Creando…", "Création…", "Создание…", "A criar…", "생성 중…"],
+  ["Saving…", "正在保存…", "保存中…", "Wird gespeichert…", "Guardando…", "Enregistrement…", "Сохранение…", "A guardar…", "저장 중…"],
+  ["Uploading…", "正在上传…", "アップロード中…", "Wird hochgeladen…", "Subiendo…", "Envoi…", "Отправка…", "A enviar…", "업로드 중…"],
+  ["Detecting…", "正在检测…", "検出中…", "Erkennung läuft…", "Detectando…", "Détection…", "Детекция…", "A detetar…", "검출 중…"],
+  ["Comparing…", "正在比对…", "比較中…", "Vergleich läuft…", "Comparando…", "Comparaison…", "Сравнение…", "A comparar…", "비교 중…"],
+  ["Searching…", "正在搜索…", "検索中…", "Suche läuft…", "Buscando…", "Recherche…", "Поиск…", "A pesquisar…", "검색 중…"],
+  ["Refreshing…", "正在刷新…", "更新中…", "Wird aktualisiert…", "Actualizando…", "Actualisation…", "Обновление…", "A atualizar…", "새로 고치는 중…"],
+  ["Registering…", "正在注册…", "登録中…", "Registrierung läuft…", "Registrando…", "Inscription…", "Регистрация…", "A registar…", "등록 중…"],
+  ["Request failed", "请求失败", "リクエスト失敗", "Anfrage fehlgeschlagen", "Solicitud fallida", "Échec de la requête", "Ошибка запроса", "Pedido falhou", "요청 실패"],
+  ["Server error", "服务器错误", "サーバーエラー", "Serverfehler", "Error del servidor", "Erreur serveur", "Ошибка сервера", "Erro do servidor", "서버 오류"],
+  ["An error was recorded by the server.", "服务器记录了一个错误。", "サーバーにエラーが記録されました。", "Der Server hat einen Fehler protokolliert.", "El servidor registró un error.", "Le serveur a enregistré une erreur.", "Сервер зарегистрировал ошибку.", "O servidor registou um erro.", "서버에 오류가 기록되었습니다."],
+  ["API failures from this browser tab will appear here.", "当前浏览器标签页中的 API 失败会显示在这里。", "このブラウザタブの API エラーがここに表示されます。", "API-Fehler dieses Browser-Tabs werden hier angezeigt.", "Los fallos de API de esta pestaña aparecerán aquí.", "Les échecs d’API de cet onglet apparaîtront ici.", "Ошибки API из этой вкладки появятся здесь.", "As falhas da API deste separador aparecem aqui.", "이 브라우저 탭의 API 오류가 여기에 표시됩니다."],
+  ["Could not load search capabilities", "无法加载搜索能力", "検索機能を読み込めませんでした", "Suchfunktionen konnten nicht geladen werden", "No se pudieron cargar las capacidades de búsqueda", "Impossible de charger les capacités de recherche", "Не удалось загрузить возможности поиска", "Não foi possível carregar as capacidades de pesquisa", "검색 기능을 로드할 수 없습니다"],
+  ["Could not load collections", "无法加载人员库", "コレクションを読み込めませんでした", "Sammlungen konnten nicht geladen werden", "No se pudieron cargar las colecciones", "Impossible de charger les collections", "Не удалось загрузить коллекции", "Não foi possível carregar as coleções", "컬렉션을 로드할 수 없습니다"],
+  ["Collection could not be deleted", "无法删除人员库", "コレクションを削除できませんでした", "Sammlung konnte nicht gelöscht werden", "No se pudo eliminar la colección", "Impossible de supprimer la collection", "Не удалось удалить коллекцию", "Não foi possível eliminar a coleção", "컬렉션을 삭제할 수 없습니다"],
+  ["Could not load people", "无法加载人员", "人物を読み込めませんでした", "Personen konnten nicht geladen werden", "No se pudieron cargar las personas", "Impossible de charger les personnes", "Не удалось загрузить людей", "Não foi possível carregar as pessoas", "인물을 로드할 수 없습니다"],
+  ["Could not load person", "无法加载人员详情", "人物を読み込めませんでした", "Person konnte nicht geladen werden", "No se pudo cargar la persona", "Impossible de charger la personne", "Не удалось загрузить человека", "Não foi possível carregar a pessoa", "인물을 로드할 수 없습니다"],
+  ["Could not load face samples", "无法加载人脸样本", "顔サンプルを読み込めませんでした", "Gesichtsproben konnten nicht geladen werden", "No se pudieron cargar las muestras faciales", "Impossible de charger les échantillons faciaux", "Не удалось загрузить образцы лиц", "Não foi possível carregar as amostras faciais", "얼굴 샘플을 로드할 수 없습니다"],
+  ["Face crop could not be loaded", "无法加载人脸裁剪图", "顔クロップを読み込めませんでした", "Gesichtsausschnitt konnte nicht geladen werden", "No se pudo cargar el recorte facial", "Impossible de charger le recadrage facial", "Не удалось загрузить кроп лица", "Não foi possível carregar o recorte facial", "얼굴 크롭을 로드할 수 없습니다"],
+  ["Face crop could not be displayed", "无法显示人脸裁剪图", "顔クロップを表示できませんでした", "Gesichtsausschnitt konnte nicht angezeigt werden", "No se pudo mostrar el recorte facial", "Impossible d’afficher le recadrage facial", "Не удалось показать кроп лица", "Não foi possível apresentar o recorte facial", "얼굴 크롭을 표시할 수 없습니다"],
+  ["The stored face crop could not be decoded.", "无法解码已保存的人脸裁剪图。", "保存済み顔クロップをデコードできませんでした。", "Der gespeicherte Gesichtsausschnitt konnte nicht dekodiert werden.", "No se pudo decodificar el recorte facial guardado.", "Le recadrage facial enregistré n’a pas pu être décodé.", "Не удалось декодировать сохранённый кроп лица.", "Não foi possível descodificar o recorte facial guardado.", "저장된 얼굴 크롭을 디코딩할 수 없습니다."],
+  ["Invalid person metadata", "人员元数据无效", "人物メタデータが無効です", "Ungültige Personenmetadaten", "Metadatos de persona no válidos", "Métadonnées de personne invalides", "Некорректные метаданные человека", "Metadados da pessoa inválidos", "인물 메타데이터가 올바르지 않습니다"],
+  ["Person could not be updated", "无法更新人员", "人物を更新できませんでした", "Person konnte nicht aktualisiert werden", "No se pudo actualizar la persona", "Impossible de mettre à jour la personne", "Не удалось обновить человека", "Não foi possível atualizar a pessoa", "인물을 업데이트할 수 없습니다"],
+  ["Person could not be deleted", "无法删除人员", "人物を削除できませんでした", "Person konnte nicht gelöscht werden", "No se pudo eliminar la persona", "Impossible de supprimer la personne", "Не удалось удалить человека", "Não foi possível eliminar a pessoa", "인물을 삭제할 수 없습니다"],
+  ["Invalid external embeddings", "外部特征格式无效", "外部特徴量が無効です", "Ungültige externe Embeddings", "Embeddings externos no válidos", "Embeddings externes invalides", "Некорректные внешние эмбеддинги", "Embeddings externos inválidos", "외부 임베딩이 올바르지 않습니다"],
+  ["Face samples could not be added", "无法添加人脸样本", "顔サンプルを追加できませんでした", "Gesichtsproben konnten nicht hinzugefügt werden", "No se pudieron añadir las muestras faciales", "Impossible d’ajouter les échantillons faciaux", "Не удалось добавить образцы лиц", "Não foi possível adicionar as amostras faciais", "얼굴 샘플을 추가할 수 없습니다"],
+  ["Face sample could not be deleted", "无法删除人脸样本", "顔サンプルを削除できませんでした", "Gesichtsprobe konnte nicht gelöscht werden", "No se pudo eliminar la muestra facial", "Impossible de supprimer l’échantillon facial", "Не удалось удалить образец лица", "Não foi possível eliminar a amostra facial", "얼굴 샘플을 삭제할 수 없습니다"],
+  ["Image preview failed", "图片预览失败", "画像プレビューに失敗しました", "Bildvorschau fehlgeschlagen", "Falló la vista previa", "Échec de l’aperçu de l’image", "Ошибка предпросмотра изображения", "Falha na pré-visualização da imagem", "이미지 미리보기에 실패했습니다"],
+  ["The selected image could not be decoded.", "无法解码所选图片。", "選択した画像をデコードできませんでした。", "Das ausgewählte Bild konnte nicht dekodiert werden.", "No se pudo decodificar la imagen seleccionada.", "L’image sélectionnée n’a pas pu être décodée.", "Не удалось декодировать выбранное изображение.", "Não foi possível descodificar a imagem selecionada.", "선택한 이미지를 디코딩할 수 없습니다."],
+  ["Detection failed", "检测失败", "検出に失敗しました", "Erkennung fehlgeschlagen", "Falló la detección", "Échec de la détection", "Ошибка детекции", "Falha na deteção", "검출 실패"],
+  ["Comparison failed", "比对失败", "比較に失敗しました", "Vergleich fehlgeschlagen", "Falló la comparación", "Échec de la comparaison", "Ошибка сравнения", "Falha na comparação", "비교 실패"],
+  ["Search failed", "搜索失败", "検索に失敗しました", "Suche fehlgeschlagen", "Falló la búsqueda", "Échec de la recherche", "Ошибка поиска", "Falha na pesquisa", "검색 실패"],
+  ["Could not load system diagnostics", "无法加载系统诊断", "システム診断を読み込めませんでした", "Systemdiagnose konnte nicht geladen werden", "No se pudo cargar el diagnóstico del sistema", "Impossible de charger le diagnostic système", "Не удалось загрузить диагностику системы", "Não foi possível carregar o diagnóstico do sistema", "시스템 진단을 로드할 수 없습니다"],
+  ["Choose a playable video first.", "请先选择可播放的视频。", "再生可能な動画を選択してください。", "Wählen Sie zuerst ein abspielbares Video.", "Seleccione primero un vídeo reproducible.", "Choisissez d’abord une vidéo lisible.", "Сначала выберите воспроизводимое видео.", "Escolha primeiro um vídeo reproduzível.", "먼저 재생 가능한 비디오를 선택하세요."],
+  ["Choose a collection before starting recognition.", "开始识别前请选择人员库。", "認識を開始する前にコレクションを選択してください。", "Wählen Sie vor Beginn der Erkennung eine Sammlung.", "Seleccione una colección antes de iniciar el reconocimiento.", "Choisissez une collection avant de démarrer la reconnaissance.", "Перед распознаванием выберите коллекцию.", "Escolha uma coleção antes de iniciar o reconhecimento.", "인식을 시작하기 전에 컬렉션을 선택하세요."],
+  ["Recognition not started", "识别未启动", "認識は開始されていません", "Erkennung nicht gestartet", "Reconocimiento no iniciado", "Reconnaissance non démarrée", "Распознавание не запущено", "Reconhecimento não iniciado", "인식이 시작되지 않음"],
+  ["Video frame failed", "视频帧处理失败", "動画フレーム処理に失敗しました", "Video-Frame fehlgeschlagen", "Falló el fotograma de vídeo", "Échec du traitement de l’image vidéo", "Ошибка обработки видеокадра", "Falha no fotograma do vídeo", "비디오 프레임 처리 실패"],
+  ["Video could not start", "无法启动视频", "動画を開始できませんでした", "Video konnte nicht gestartet werden", "No se pudo iniciar el vídeo", "Impossible de démarrer la vidéo", "Не удалось запустить видео", "Não foi possível iniciar o vídeo", "비디오를 시작할 수 없습니다"],
+  ["Select a collection first.", "请先选择人员库。", "先にコレクションを選択してください。", "Wählen Sie zuerst eine Sammlung.", "Seleccione primero una colección.", "Sélectionnez d’abord une collection.", "Сначала выберите коллекцию.", "Selecione primeiro uma coleção.", "먼저 컬렉션을 선택하세요."],
+  ["Cannot register person", "无法注册人员", "人物を登録できません", "Person kann nicht registriert werden", "No se puede registrar la persona", "Impossible d’inscrire la personne", "Невозможно зарегистрировать человека", "Não é possível registar a pessoa", "인물을 등록할 수 없습니다"],
+  ["This browser could not decode the selected video.", "此浏览器无法解码所选视频。", "このブラウザは選択した動画をデコードできませんでした。", "Dieser Browser konnte das gewählte Video nicht dekodieren.", "Este navegador no pudo decodificar el vídeo seleccionado.", "Ce navigateur n’a pas pu décoder la vidéo sélectionnée.", "Браузер не смог декодировать выбранное видео.", "Este navegador não conseguiu descodificar o vídeo selecionado.", "이 브라우저에서 선택한 비디오를 디코딩할 수 없습니다."],
+  ["Video unavailable", "视频不可用", "動画を利用できません", "Video nicht verfügbar", "Vídeo no disponible", "Vidéo indisponible", "Видео недоступно", "Vídeo indisponível", "비디오를 사용할 수 없음"],
+  ["Invalid collection metadata", "人员库元数据无效", "コレクションのメタデータが無効です", "Ungültige Sammlungsmetadaten", "Metadatos de colección no válidos", "Métadonnées de collection invalides", "Некорректные метаданные коллекции", "Metadados da coleção inválidos", "컬렉션 메타데이터가 올바르지 않습니다"],
+  ["Invalid detection profile", "检测配置无效", "検出プロファイルが無効です", "Ungültiges Erkennungsprofil", "Perfil de detección no válido", "Profil de détection invalide", "Некорректный профиль детекции", "Perfil de deteção inválido", "검출 프로필이 올바르지 않습니다"],
+  ["Collection could not be created", "无法创建人员库", "コレクションを作成できませんでした", "Sammlung konnte nicht erstellt werden", "No se pudo crear la colección", "Impossible de créer la collection", "Не удалось создать коллекцию", "Não foi possível criar a coleção", "컬렉션을 생성할 수 없습니다"],
+  ["Collection could not be updated", "无法更新人员库", "コレクションを更新できませんでした", "Sammlung konnte nicht aktualisiert werden", "No se pudo actualizar la colección", "Impossible de mettre à jour la collection", "Не удалось обновить коллекцию", "Não foi possível atualizar a coleção", "컬렉션을 업데이트할 수 없습니다"],
+  ["Person could not be registered", "无法注册人员", "人物を登録できませんでした", "Person konnte nicht registriert werden", "No se pudo registrar la persona", "Impossible d’inscrire la personne", "Не удалось зарегистрировать человека", "Não foi possível registar a pessoa", "인물을 등록할 수 없습니다"],
+  ["Could not encode the sampled video frame.", "无法编码采样视频帧。", "サンプリングした動画フレームをエンコードできませんでした。", "Der abgetastete Video-Frame konnte nicht kodiert werden.", "No se pudo codificar el fotograma de vídeo muestreado.", "Impossible d’encoder l’image vidéo échantillonnée.", "Не удалось закодировать выбранный видеокадр.", "Não foi possível codificar o fotograma de vídeo amostrado.", "샘플 비디오 프레임을 인코딩할 수 없습니다."],
+  ["Could not load documentation.", "无法加载文档。", "ドキュメントを読み込めませんでした。", "Dokumentation konnte nicht geladen werden.", "No se pudo cargar la documentación.", "Impossible de charger la documentation.", "Не удалось загрузить документацию.", "Não foi possível carregar a documentação.", "문서를 로드할 수 없습니다."],
+  ["REST API · VERSION 1", "REST API · 版本 1", "REST API · バージョン 1", "REST API · VERSION 1", "API REST · VERSIÓN 1", "API REST · VERSION 1", "REST API · ВЕРСИЯ 1", "API REST · VERSÃO 1", "REST API · 버전 1"],
+  ["Loading the server's OpenAPI document…", "正在加载服务器的 OpenAPI 文档…", "サーバーの OpenAPI 文書を読み込み中…", "OpenAPI-Dokument des Servers wird geladen…", "Cargando el documento OpenAPI del servidor…", "Chargement du document OpenAPI du serveur…", "Загрузка OpenAPI-документа сервера…", "A carregar o documento OpenAPI do servidor…", "서버 OpenAPI 문서를 로드 중…"],
+  ["detect, collections…", "检测、人员库…", "検出、コレクション…", "Erkennung, Sammlungen…", "detección, colecciones…", "détection, collections…", "детекция, коллекции…", "deteção, coleções…", "검출, 컬렉션…"],
+  ["The OpenAPI document contains no operations.", "OpenAPI 文档中没有操作。", "OpenAPI 文書に操作がありません。", "Das OpenAPI-Dokument enthält keine Operationen.", "El documento OpenAPI no contiene operaciones.", "Le document OpenAPI ne contient aucune opération.", "В OpenAPI-документе нет операций.", "O documento OpenAPI não contém operações.", "OpenAPI 문서에 작업이 없습니다."],
+  ["Response", "响应", "レスポンス", "Antwort", "Respuesta", "Réponse", "Ответ", "Resposta", "응답"],
+  ["{count} schemas", "{count} 个 Schema", "{count} スキーマ", "{count} Schemas", "{count} esquemas", "{count} schémas", "Схем: {count}", "{count} esquemas", "스키마 {count}개"],
+  ["Generated locally from /openapi.json. No external scripts, fonts, or services are used.", "由 /openapi.json 在本地生成，不使用外部脚本、字体或服务。", "/openapi.json からローカル生成され、外部スクリプト、フォント、サービスは使用しません。", "Lokal aus /openapi.json erzeugt. Es werden keine externen Skripte, Schriften oder Dienste verwendet.", "Generado localmente desde /openapi.json. No usa scripts, fuentes ni servicios externos.", "Généré localement depuis /openapi.json. Aucun script, police ou service externe n’est utilisé.", "Сформировано локально из /openapi.json. Внешние скрипты, шрифты и сервисы не используются.", "Gerado localmente a partir de /openapi.json. Não usa scripts, tipos de letra ou serviços externos.", "/openapi.json에서 로컬로 생성되며 외부 스크립트, 글꼴 또는 서비스를 사용하지 않습니다."],
+  ["Generated locally from", "本地生成自", "ローカル生成元：", "Lokal erzeugt aus", "Generado localmente desde", "Généré localement depuis", "Сформировано локально из", "Gerado localmente a partir de", "로컬 생성 원본:"],
+  [". No external scripts, fonts, or services are used.", "。不使用外部脚本、字体或服务。", "。外部スクリプト、フォント、サービスは使用しません。", ". Es werden keine externen Skripte, Schriften oder Dienste verwendet.", ". No usa scripts, fuentes ni servicios externos.", ". Aucun script, police ou service externe n’est utilisé.", ". Внешние скрипты, шрифты и сервисы не используются.", ". Não usa scripts, tipos de letra ou serviços externos.", ". 외부 스크립트, 글꼴 또는 서비스를 사용하지 않습니다."],
+  ["Simple self-hosted face detection, comparison, registration, and search.", "简单的自托管人脸检测、比对、注册和搜索服务。", "シンプルなセルフホスト型の顔検出、比較、登録、検索サービスです。", "Einfache selbst gehostete Gesichtserkennung, Vergleiche, Registrierung und Suche.", "Detección, comparación, registro y búsqueda facial sencilla y autohospedada.", "Détection, comparaison, inscription et recherche faciales simples et auto-hébergées.", "Простой локальный сервис детекции, сравнения, регистрации и поиска лиц.", "Deteção, comparação, registo e pesquisa facial simples e autoalojada.", "간단한 자체 호스팅 얼굴 검출, 비교, 등록 및 검색 서비스입니다."],
+  ["Create Collection", "创建人员库", "コレクションを作成", "Sammlung erstellen", "Crear colección", "Créer une collection", "Создать коллекцию", "Criar coleção", "컬렉션 생성"],
+  ["List Collections", "列出人员库", "コレクション一覧", "Sammlungen auflisten", "Listar colecciones", "Lister les collections", "Список коллекций", "Listar coleções", "컬렉션 목록"],
+  ["Get Collection", "获取人员库", "コレクションを取得", "Sammlung abrufen", "Obtener colección", "Obtenir la collection", "Получить коллекцию", "Obter coleção", "컬렉션 조회"],
+  ["Patch Collection", "修改人员库", "コレクションを更新", "Sammlung ändern", "Modificar colección", "Modifier la collection", "Изменить коллекцию", "Alterar coleção", "컬렉션 수정"],
+  ["Delete Collection", "删除人员库", "コレクションを削除", "Sammlung löschen", "Eliminar colección", "Supprimer la collection", "Удалить коллекцию", "Eliminar coleção", "컬렉션 삭제"],
+  ["Create Person", "创建人员", "人物を作成", "Person erstellen", "Crear persona", "Créer une personne", "Создать человека", "Criar pessoa", "인물 생성"],
+  ["List Persons", "列出人员", "人物一覧", "Personen auflisten", "Listar personas", "Lister les personnes", "Список людей", "Listar pessoas", "인물 목록"],
+  ["Get Person", "获取人员", "人物を取得", "Person abrufen", "Obtener persona", "Obtenir la personne", "Получить человека", "Obter pessoa", "인물 조회"],
+  ["Patch Person", "修改人员", "人物を更新", "Person ändern", "Modificar persona", "Modifier la personne", "Изменить человека", "Alterar pessoa", "인물 수정"],
+  ["Delete Person", "删除人员", "人物を削除", "Person löschen", "Eliminar persona", "Supprimer la personne", "Удалить человека", "Eliminar pessoa", "인물 삭제"],
+  ["Add Faces", "添加人脸", "顔を追加", "Gesichter hinzufügen", "Añadir rostros", "Ajouter des visages", "Добавить лица", "Adicionar rostos", "얼굴 추가"],
+  ["List Faces", "列出人脸", "顔一覧", "Gesichter auflisten", "Listar rostros", "Lister les visages", "Список лиц", "Listar rostos", "얼굴 목록"],
+  ["Get Face Image", "获取人脸图片", "顔画像を取得", "Gesichtsbild abrufen", "Obtener imagen facial", "Obtenir l’image faciale", "Получить изображение лица", "Obter imagem facial", "얼굴 이미지 조회"],
+  ["Delete Face", "删除人脸", "顔を削除", "Gesicht löschen", "Eliminar rostro", "Supprimer le visage", "Удалить лицо", "Eliminar rosto", "얼굴 삭제"],
+  ["Embeddings", "特征向量", "特徴量", "Embeddings", "Embeddings", "Embeddings", "Эмбеддинги", "Embeddings", "임베딩"],
+  ["Could not load OpenAPI document: {message}", "无法加载 OpenAPI 文档：{message}", "OpenAPI 文書を読み込めませんでした：{message}", "OpenAPI-Dokument konnte nicht geladen werden: {message}", "No se pudo cargar el documento OpenAPI: {message}", "Impossible de charger le document OpenAPI : {message}", "Не удалось загрузить OpenAPI-документ: {message}", "Não foi possível carregar o documento OpenAPI: {message}", "OpenAPI 문서를 로드할 수 없습니다: {message}"],
+  ["API Reference · InsightFace Server", "API 参考 · InsightFace Server", "API リファレンス · InsightFace Server", "API-Referenz · InsightFace Server", "Referencia API · InsightFace Server", "Référence API · InsightFace Server", "Справочник API · InsightFace Server", "Referência API · InsightFace Server", "API 참조 · InsightFace Server"],
+  ["Save 112×112 face crops in SQLite", "在 SQLite 中保存 112×112 对齐人脸", "112×112 の顔クロップを SQLite に保存", "112×112-Gesichtsausschnitte in SQLite speichern", "Guardar recortes faciales de 112×112 en SQLite", "Enregistrer les visages recadrés 112×112 dans SQLite", "Сохранять кропы лиц 112×112 в SQLite", "Guardar recortes faciais 112×112 em SQLite", "112×112 얼굴 크롭을 SQLite에 저장"],
+  ["/data path", "/data 路径", "/data パス", "/data-Pfad", "Ruta de /data", "Chemin /data", "Путь /data", "Caminho de /data", "/data 경로"],
+  ["/models path", "/models 路径", "/models パス", "/models-Pfad", "Ruta de /models", "Chemin /models", "Путь /models", "Caminho de /models", "/models 경로"],
+  ["Camera recognition", "摄像头识别", "カメラ認識", "Kameraerkennung", "Reconocimiento de cámara", "Reconnaissance caméra", "Распознавание с камеры", "Reconhecimento por câmara", "카메라 인식"],
+  ["RTSP CAMERA", "RTSP 摄像头", "RTSP カメラ", "RTSP-KAMERA", "CÁMARA RTSP", "CAMÉRA RTSP", "RTSP-КАМЕРА", "CÂMARA RTSP", "RTSP 카메라"],
+  ["Connect the server to an RTSP camera and recognize faces at a controlled inference rate. Frames are processed transiently and are not saved.", "将服务器连接到 RTSP 摄像头，并按设定频率进行人脸识别。视频帧仅临时处理，不会保存。", "サーバーを RTSP カメラへ接続し、指定した推論頻度で顔を認識します。フレームは一時的に処理され、保存されません。", "Verbindet den Server mit einer RTSP-Kamera und erkennt Gesichter mit kontrollierter Inferenzrate. Frames werden nur vorübergehend verarbeitet und nicht gespeichert.", "Conecta el servidor a una cámara RTSP y reconoce rostros a una frecuencia controlada. Los fotogramas se procesan de forma transitoria y no se guardan.", "Connecte le serveur à une caméra RTSP et reconnaît les visages à une fréquence maîtrisée. Les images sont traitées temporairement et ne sont pas enregistrées.", "Подключает сервер к RTSP-камере и распознаёт лица с заданной частотой. Кадры обрабатываются временно и не сохраняются.", "Liga o servidor a uma câmara RTSP e reconhece rostos a uma frequência controlada. Os fotogramas são processados temporariamente e não são guardados.", "서버를 RTSP 카메라에 연결하고 설정한 추론 속도로 얼굴을 인식합니다. 프레임은 일시적으로 처리되며 저장되지 않습니다."],
+  ["Annotated RTSP camera stream", "带识别标注的 RTSP 摄像头画面", "認識結果を重ねた RTSP カメラ映像", "Annotierter RTSP-Kamerastream", "Flujo de cámara RTSP anotado", "Flux caméra RTSP annoté", "RTSP-поток камеры с разметкой", "Fluxo da câmara RTSP anotado", "인식 결과가 표시된 RTSP 카메라 스트림"],
+  ["No camera connected", "未连接摄像头", "カメラ未接続", "Keine Kamera verbunden", "Ninguna cámara conectada", "Aucune caméra connectée", "Камера не подключена", "Nenhuma câmara ligada", "연결된 카메라 없음"],
+  ["Enter an RTSP URL and start recognition. The server pulls the stream; this browser does not access a local camera.", "输入 RTSP 地址并开始识别。视频流由服务器拉取，浏览器不会访问本机摄像头。", "RTSP URL を入力して認識を開始してください。ストリームはサーバーが取得し、ブラウザーはローカルカメラへアクセスしません。", "Geben Sie eine RTSP-URL ein und starten Sie die Erkennung. Der Server ruft den Stream ab; dieser Browser greift nicht auf eine lokale Kamera zu.", "Introduzca una URL RTSP e inicie el reconocimiento. El servidor obtiene el flujo; el navegador no accede a ninguna cámara local.", "Saisissez une URL RTSP et lancez la reconnaissance. Le serveur récupère le flux ; ce navigateur n’accède pas à la caméra locale.", "Введите RTSP-адрес и запустите распознавание. Поток получает сервер; браузер не обращается к локальной камере.", "Introduza um URL RTSP e inicie o reconhecimento. O servidor obtém o fluxo; o navegador não acede a uma câmara local.", "RTSP URL을 입력하고 인식을 시작하세요. 서버가 스트림을 가져오며 브라우저는 로컬 카메라에 접근하지 않습니다."],
+  ["Recognition color legend", "识别颜色图例", "認識色の凡例", "Farblegende der Erkennung", "Leyenda de colores de reconocimiento", "Légende des couleurs de reconnaissance", "Цветовая легенда распознавания", "Legenda de cores do reconhecimento", "인식 색상 범례"],
+  ["Recognized person", "已识别人员", "登録済み人物", "Erkannte Person", "Persona reconocida", "Personne reconnue", "Распознанный человек", "Pessoa reconhecida", "인식된 사람"],
+  ["Detected, not enrolled", "检测到人脸，未入库", "顔を検出・未登録", "Erkannt, nicht registriert", "Detectado, no registrado", "Détecté, non inscrit", "Лицо обнаружено, но не зарегистрировано", "Detetado, não registado", "얼굴 검출됨, 미등록"],
+  ["RTSP URL", "RTSP 地址", "RTSP URL", "RTSP-URL", "URL RTSP", "URL RTSP", "RTSP-адрес", "URL RTSP", "RTSP URL"],
+  ["The URL is kept in memory only and is never returned with credentials.", "地址仅保留在内存中，返回结果绝不会包含认证信息。", "URL はメモリ内だけに保持され、認証情報を含めて返されることはありません。", "Die URL bleibt nur im Arbeitsspeicher und wird nie mit Zugangsdaten zurückgegeben.", "La URL solo se mantiene en memoria y nunca se devuelve con credenciales.", "L’URL reste uniquement en mémoire et n’est jamais renvoyée avec ses identifiants.", "Адрес хранится только в памяти и никогда не возвращается вместе с учётными данными.", "O URL é mantido apenas em memória e nunca é devolvido com credenciais.", "URL은 메모리에만 유지되며 인증 정보와 함께 반환되지 않습니다."],
+  ["Inference rate", "推理频率", "推論頻度", "Inferenzrate", "Frecuencia de inferencia", "Fréquence d’inférence", "Частота инференса", "Frequência de inferência", "추론 속도"],
+  ["Executions per second", "每秒执行次数", "1 秒あたりの実行回数", "Ausführungen pro Sekunde", "Ejecuciones por segundo", "Exécutions par seconde", "Запусков в секунду", "Execuções por segundo", "초당 실행 횟수"],
+  ["Only the latest frame is retained. If inference takes longer than the configured interval, the effective rate decreases and stale frames are skipped.", "系统只保留最新帧。如果一次推理超过设定周期，实际频率会自动降低，并跳过已经过时的帧。", "最新フレームだけを保持します。推論が設定間隔より長い場合、実効頻度を下げて古いフレームをスキップします。", "Nur der neueste Frame wird behalten. Dauert die Inferenz länger als das eingestellte Intervall, sinkt die effektive Rate und veraltete Frames werden übersprungen.", "Solo se conserva el fotograma más reciente. Si la inferencia tarda más que el intervalo configurado, baja la frecuencia efectiva y se omiten los fotogramas obsoletos.", "Seule l’image la plus récente est conservée. Si l’inférence dépasse l’intervalle configuré, la fréquence effective diminue et les images obsolètes sont ignorées.", "Сохраняется только самый свежий кадр. Если инференс длится дольше заданного интервала, фактическая частота снижается, а устаревшие кадры пропускаются.", "Apenas o fotograma mais recente é mantido. Se a inferência demorar mais do que o intervalo configurado, a frequência efetiva diminui e os fotogramas antigos são ignorados.", "가장 최신 프레임만 유지합니다. 추론 시간이 설정 간격보다 길면 실제 속도가 낮아지고 오래된 프레임은 건너뜁니다."],
+  ["Enter an RTSP URL and choose a Collection.", "请输入 RTSP 地址并选择人员库。", "RTSP URL を入力してコレクションを選択してください。", "Geben Sie eine RTSP-URL ein und wählen Sie eine Sammlung.", "Introduzca una URL RTSP y elija una colección.", "Saisissez une URL RTSP et choisissez une collection.", "Введите RTSP-адрес и выберите коллекцию.", "Introduza um URL RTSP e escolha uma coleção.", "RTSP URL을 입력하고 컬렉션을 선택하세요."],
+  ["Connection", "连接状态", "接続", "Verbindung", "Conexión", "Connexion", "Подключение", "Ligação", "연결"],
+  ["Effective rate", "实际频率", "実効頻度", "Effektive Rate", "Frecuencia efectiva", "Fréquence effective", "Фактическая частота", "Frequência efetiva", "실제 속도"],
+  ["Inference time", "推理耗时", "推論時間", "Inferenzzeit", "Tiempo de inferencia", "Temps d’inférence", "Время инференса", "Tempo de inferência", "추론 시간"],
+  ["Skipped frames", "跳过帧数", "スキップしたフレーム", "Übersprungene Frames", "Fotogramas omitidos", "Images ignorées", "Пропущенные кадры", "Fotogramas ignorados", "건너뛴 프레임"],
+  ["Detection score {score}", "检测分数 {score}", "検出スコア {score}", "Erkennungsscore {score}", "Puntuación de detección {score}", "Score de détection {score}", "Оценка детекции {score}", "Pontuação de deteção {score}", "검출 점수 {score}"],
+  ["Connected", "已连接", "接続済み", "Verbunden", "Conectado", "Connecté", "Подключено", "Ligado", "연결됨"],
+  ["Reconnecting", "正在重连", "再接続中", "Verbindung wird wiederhergestellt", "Reconectando", "Reconnexion", "Повторное подключение", "A restabelecer ligação", "재연결 중"],
+  ["{effective} / {requested} per second", "每秒 {effective} / {requested}", "毎秒 {effective} / {requested}", "{effective} / {requested} pro Sekunde", "{effective} / {requested} por segundo", "{effective} / {requested} par seconde", "{effective} / {requested} в секунду", "{effective} / {requested} por segundo", "초당 {effective} / {requested}"],
+  ["RTSP stream unavailable. Reconnecting…", "RTSP 视频流不可用，正在重连…", "RTSP ストリームを利用できません。再接続中…", "RTSP-Stream nicht verfügbar. Verbindung wird wiederhergestellt…", "Flujo RTSP no disponible. Reconectando…", "Flux RTSP indisponible. Reconnexion…", "RTSP-поток недоступен. Повторное подключение…", "Fluxo RTSP indisponível. A restabelecer ligação…", "RTSP 스트림을 사용할 수 없습니다. 재연결 중…"],
+  ["Connecting to RTSP camera…", "正在连接 RTSP 摄像头…", "RTSP カメラへ接続中…", "Verbindung zur RTSP-Kamera wird hergestellt…", "Conectando a la cámara RTSP…", "Connexion à la caméra RTSP…", "Подключение к RTSP-камере…", "A ligar à câmara RTSP…", "RTSP 카메라에 연결 중…"],
+  ["Inference-limited; stale frames are being skipped", "受推理速度限制，正在跳过过时帧", "推論速度が上限のため、古いフレームをスキップ中", "Durch Inferenz begrenzt; veraltete Frames werden übersprungen", "Limitado por la inferencia; se omiten fotogramas obsoletos", "Limité par l’inférence ; les images obsolètes sont ignorées", "Ограничено скоростью инференса; устаревшие кадры пропускаются", "Limitado pela inferência; os fotogramas antigos são ignorados", "추론 속도 제한으로 오래된 프레임을 건너뜁니다"],
+  ["Running at the configured cadence", "按设定频率运行", "設定した頻度で実行中", "Läuft mit der eingestellten Rate", "Funcionando a la frecuencia configurada", "Fonctionnement à la fréquence configurée", "Работа с заданной частотой", "A funcionar à frequência configurada", "설정한 속도로 실행 중"],
+  ["{recognized} recognized · {unknown} not enrolled · {timing}", "已识别 {recognized} · 未入库 {unknown} · {timing}", "認識済み {recognized} · 未登録 {unknown} · {timing}", "{recognized} erkannt · {unknown} nicht registriert · {timing}", "{recognized} reconocidos · {unknown} no registrados · {timing}", "{recognized} reconnus · {unknown} non inscrits · {timing}", "Распознано: {recognized} · Не зарегистрировано: {unknown} · {timing}", "{recognized} reconhecidos · {unknown} não registados · {timing}", "인식 {recognized} · 미등록 {unknown} · {timing}"],
+  ["Enter a valid RTSP URL.", "请输入有效的 RTSP 地址。", "有効な RTSP URL を入力してください。", "Geben Sie eine gültige RTSP-URL ein.", "Introduzca una URL RTSP válida.", "Saisissez une URL RTSP valide.", "Введите корректный RTSP-адрес.", "Introduza um URL RTSP válido.", "올바른 RTSP URL을 입력하세요."],
+  ["RTSP camera could not start", "RTSP 摄像头无法启动", "RTSP カメラを開始できませんでした", "RTSP-Kamera konnte nicht gestartet werden", "No se pudo iniciar la cámara RTSP", "Impossible de démarrer la caméra RTSP", "Не удалось запустить RTSP-камеру", "Não foi possível iniciar a câmara RTSP", "RTSP 카메라를 시작할 수 없음"],
+  ["Recognition stopped.", "识别已停止。", "認識を停止しました。", "Erkennung gestoppt.", "Reconocimiento detenido.", "Reconnaissance arrêtée.", "Распознавание остановлено.", "Reconhecimento parado.", "인식이 중지되었습니다."],
+  ["Waiting for the first RTSP frame…", "正在等待首个 RTSP 视频帧…", "最初の RTSP フレームを待機中…", "Warten auf den ersten RTSP-Frame…", "Esperando el primer fotograma RTSP…", "En attente de la première image RTSP…", "Ожидание первого RTSP-кадра…", "A aguardar o primeiro fotograma RTSP…", "첫 RTSP 프레임을 기다리는 중…"],
+  ["RTSP camera could not stop", "RTSP 摄像头无法停止", "RTSP カメラを停止できませんでした", "RTSP-Kamera konnte nicht gestoppt werden", "No se pudo detener la cámara RTSP", "Impossible d’arrêter la caméra RTSP", "Не удалось остановить RTSP-камеру", "Não foi possível parar a câmara RTSP", "RTSP 카메라를 중지할 수 없음"],
+  ["RTSP MONITORS", "RTSP 监控器", "RTSP モニター", "RTSP-MONITORE", "MONITORES RTSP", "MONITEURS RTSP", "RTSP-МОНИТОРЫ", "MONITORES RTSP", "RTSP 모니터"],
+  ["Camera monitoring", "摄像头监控", "カメラ監視", "Kameraüberwachung", "Monitorización de cámaras", "Surveillance caméra", "Мониторинг камер", "Monitorização de câmaras", "카메라 모니터링"],
+  ["Create persistent server-side Monitors, inspect their current state, and pull recent in-memory events. Video frames and events are not saved.", "创建持久的服务端监控任务，查看当前状态并拉取最近的内存事件。视频帧和事件不会保存。", "永続的なサーバー側モニターを作成し、現在の状態と直近のメモリー内イベントを確認します。映像フレームとイベントは保存されません。", "Erstellen Sie dauerhafte serverseitige Monitore, prüfen Sie den aktuellen Zustand und rufen Sie letzte Speicherereignisse ab. Videobilder und Ereignisse werden nicht gespeichert.", "Cree monitores persistentes en el servidor, revise su estado y consulte los eventos recientes en memoria. Los fotogramas y eventos no se guardan.", "Créez des moniteurs persistants côté serveur, consultez leur état et récupérez les événements récents en mémoire. Les images et événements ne sont pas enregistrés.", "Создавайте постоянные серверные мониторы, проверяйте их состояние и получайте последние события из памяти. Кадры и события не сохраняются.", "Crie monitores persistentes no servidor, consulte o estado atual e obtenha eventos recentes em memória. Os fotogramas e eventos não são guardados.", "서버 측 영구 모니터를 만들고 현재 상태와 최근 메모리 이벤트를 조회합니다. 영상 프레임과 이벤트는 저장되지 않습니다."],
+  ["New Monitor", "新建监控任务", "新規モニター", "Neuer Monitor", "Nuevo monitor", "Nouveau moniteur", "Новый монитор", "Novo monitor", "새 모니터"],
+  ["CONFIGURATION", "配置", "設定", "KONFIGURATION", "CONFIGURACIÓN", "CONFIGURATION", "КОНФИГУРАЦИЯ", "CONFIGURAÇÃO", "구성"],
+  ["Monitors", "监控任务", "モニター", "Monitore", "Monitores", "Moniteurs", "Мониторы", "Monitores", "모니터"],
+  ["No Monitors", "暂无监控任务", "モニターなし", "Keine Monitore", "No hay monitores", "Aucun moniteur", "Нет мониторов", "Sem monitores", "모니터 없음"],
+  ["Create a Monitor to connect an RTSP camera.", "创建监控任务以连接 RTSP 摄像头。", "RTSP カメラへ接続するモニターを作成してください。", "Erstellen Sie einen Monitor, um eine RTSP-Kamera zu verbinden.", "Cree un monitor para conectar una cámara RTSP.", "Créez un moniteur pour connecter une caméra RTSP.", "Создайте монитор для подключения RTSP-камеры.", "Crie um monitor para ligar uma câmara RTSP.", "RTSP 카메라를 연결할 모니터를 만드세요."],
+  ["NEW MONITOR", "新建监控任务", "新規モニター", "NEUER MONITOR", "NUEVO MONITOR", "NOUVEAU MONITEUR", "НОВЫЙ МОНИТОР", "NOVO MONITOR", "새 모니터"],
+  ["Create Monitor", "创建监控任务", "モニターを作成", "Monitor erstellen", "Crear monitor", "Créer le moniteur", "Создать монитор", "Criar monitor", "모니터 만들기"],
+  ["EDIT MONITOR", "编辑监控任务", "モニターを編集", "MONITOR BEARBEITEN", "EDITAR MONITOR", "MODIFIER LE MONITEUR", "ИЗМЕНИТЬ МОНИТОР", "EDITAR MONITOR", "모니터 편집"],
+  ["Edit Monitor", "编辑监控任务", "モニターを編集", "Monitor bearbeiten", "Editar monitor", "Modifier le moniteur", "Изменить монитор", "Editar monitor", "모니터 편집"],
+  ["Save changes", "保存修改", "変更を保存", "Änderungen speichern", "Guardar cambios", "Enregistrer les modifications", "Сохранить изменения", "Guardar alterações", "변경 사항 저장"],
+  ["Leave blank to keep the current RTSP URL", "留空以保留当前 RTSP 地址", "現在の RTSP URL を維持するには空欄のままにします", "Leer lassen, um die aktuelle RTSP-URL beizubehalten", "Déjelo vacío para conservar la URL RTSP actual", "Laissez vide pour conserver l’URL RTSP actuelle", "Оставьте пустым, чтобы сохранить текущий RTSP-адрес", "Deixe em branco para manter o URL RTSP atual", "현재 RTSP URL을 유지하려면 비워 두세요"],
+  ["Monitor ID", "监控任务 ID", "モニター ID", "Monitor-ID", "ID del monitor", "ID du moniteur", "ID монитора", "ID do monitor", "모니터 ID"],
+  ["Company front gate", "公司前门", "会社正面入口", "Firmenhaupteingang", "Entrada principal de la empresa", "Entrée principale de l’entreprise", "Главный вход компании", "Entrada principal da empresa", "회사 정문"],
+  ["Optional location or purpose", "可选的位置或用途", "任意の場所または用途", "Optionaler Standort oder Zweck", "Ubicación o finalidad opcional", "Emplacement ou usage facultatif", "Необязательное расположение или назначение", "Localização ou finalidade opcional", "선택 사항인 위치 또는 용도"],
+  ["The credential is encrypted in /data and is never returned by the API.", "凭据加密保存在 /data 中，API 永远不会返回原始凭据。", "認証情報は /data 内で暗号化され、API から返されることはありません。", "Die Zugangsdaten werden verschlüsselt in /data gespeichert und nie von der API zurückgegeben.", "La credencial se cifra en /data y la API nunca la devuelve.", "L’identifiant est chiffré dans /data et n’est jamais renvoyé par l’API.", "Учётные данные шифруются в /data и никогда не возвращаются API.", "A credencial é encriptada em /data e nunca é devolvida pela API.", "자격 증명은 /data에 암호화되며 API가 반환하지 않습니다."],
+  ["Match threshold", "匹配阈值", "一致しきい値", "Trefferschwelle", "Umbral de coincidencia", "Seuil de correspondance", "Порог совпадения", "Limiar de correspondência", "일치 임계값"],
+  ["Recent event capacity", "最近事件容量", "直近イベント容量", "Kapazität letzter Ereignisse", "Capacidad de eventos recientes", "Capacité des événements récents", "Ёмкость последних событий", "Capacidade de eventos recentes", "최근 이벤트 용량"],
+  ["Confirmation frames", "确认帧数", "確認フレーム数", "Bestätigungsbilder", "Fotogramas de confirmación", "Images de confirmation", "Кадры подтверждения", "Fotogramas de confirmação", "확인 프레임 수"],
+  ["Absence timeout", "离开超时", "不在判定タイムアウト", "Abwesenheits-Timeout", "Tiempo de ausencia", "Délai d’absence", "Тайм-аут отсутствия", "Tempo limite de ausência", "부재 시간 제한"],
+  ["Seconds before an exit event", "产生离开事件前的秒数", "退出イベントまでの秒数", "Sekunden bis zum Austrittsereignis", "Segundos antes de un evento de salida", "Secondes avant un événement de sortie", "Секунд до события выхода", "Segundos antes de um evento de saída", "퇴장 이벤트까지의 초"],
+  ["Event cooldown", "事件冷却时间", "イベント抑制時間", "Ereignis-Abklingzeit", "Pausa entre eventos", "Délai entre événements", "Интервал между событиями", "Intervalo entre eventos", "이벤트 대기 시간"],
+  ["Seconds between repeated events", "重复事件之间的秒数", "重複イベント間の秒数", "Sekunden zwischen wiederholten Ereignissen", "Segundos entre eventos repetidos", "Secondes entre événements répétés", "Секунд между повторными событиями", "Segundos entre eventos repetidos", "반복 이벤트 사이의 초"],
+  ["Start after creation", "创建后启动", "作成後に開始", "Nach Erstellung starten", "Iniciar tras la creación", "Démarrer après création", "Запустить после создания", "Iniciar após a criação", "생성 후 시작"],
+  ["Run independently of this browser.", "独立于当前浏览器运行。", "このブラウザーとは独立して実行します。", "Unabhängig von diesem Browser ausführen.", "Se ejecuta independientemente del navegador.", "Fonctionne indépendamment de ce navigateur.", "Работает независимо от браузера.", "Executa independentemente deste navegador.", "브라우저와 독립적으로 실행됩니다."],
+  ["Enabled", "已启用", "有効", "Aktiviert", "Activado", "Activé", "Включён", "Ativado", "사용"],
+  ["Run this Monitor on the server.", "在服务器上运行此监控任务。", "このモニターをサーバー上で実行します。", "Diesen Monitor auf dem Server ausführen.", "Ejecutar este monitor en el servidor.", "Exécuter ce moniteur sur le serveur.", "Запускать этот монитор на сервере.", "Executar este monitor no servidor.", "서버에서 이 모니터를 실행합니다."],
+  ["Emit unknown-person events", "推送陌生人事件", "未登録人物イベントを生成", "Ereignisse für unbekannte Personen ausgeben", "Emitir eventos de personas desconocidas", "Émettre les événements de personnes inconnues", "Создавать события неизвестных лиц", "Emitir eventos de pessoas desconhecidas", "미등록 인물 이벤트 생성"],
+  ["Include confirmed faces that are not enrolled.", "包括已确认但未入库的人脸。", "確認済みで未登録の顔を含めます。", "Bestätigte, nicht registrierte Gesichter einschließen.", "Incluye rostros confirmados no registrados.", "Inclut les visages confirmés non enregistrés.", "Включать подтверждённые незарегистрированные лица.", "Inclui rostos confirmados não registados.", "확인되었지만 등록되지 않은 얼굴을 포함합니다."],
+  ["Enable Web video preview", "启用 Web 视频预览", "Web 映像プレビューを有効化", "Web-Videovorschau aktivieren", "Activar vista previa web", "Activer l’aperçu vidéo Web", "Включить веб-просмотр видео", "Ativar pré-visualização Web", "웹 영상 미리보기 사용"],
+  ["Off by default. Recognition and events work without preview.", "默认关闭。不开启预览也能正常识别并产生事件。", "既定では無効です。プレビューなしでも認識とイベントは動作します。", "Standardmäßig aus. Erkennung und Ereignisse funktionieren ohne Vorschau.", "Desactivada por defecto. El reconocimiento y los eventos funcionan sin vista previa.", "Désactivé par défaut. La reconnaissance et les événements fonctionnent sans aperçu.", "По умолчанию выключено. Распознавание и события работают без просмотра.", "Desativada por predefinição. O reconhecimento e os eventos funcionam sem pré-visualização.", "기본값은 꺼짐입니다. 미리보기 없이도 인식과 이벤트가 동작합니다."],
+  ["RTSP camera preview", "RTSP 摄像头预览", "RTSP カメラプレビュー", "RTSP-Kameravorschau", "Vista previa de cámara RTSP", "Aperçu caméra RTSP", "Предпросмотр RTSP-камеры", "Pré-visualização da câmara RTSP", "RTSP 카메라 미리보기"],
+  ["Recognition overlay", "识别标注层", "認識オーバーレイ", "Erkennungsüberlagerung", "Capa de reconocimiento", "Calque de reconnaissance", "Слой распознавания", "Sobreposição de reconhecimento", "인식 오버레이"],
+  ["No Monitor selected", "未选择监控任务", "モニターが選択されていません", "Kein Monitor ausgewählt", "No hay monitor seleccionado", "Aucun moniteur sélectionné", "Монитор не выбран", "Nenhum monitor selecionado", "선택한 모니터 없음"],
+  ["Select a Monitor to inspect its live recognition state.", "选择监控任务以查看实时识别状态。", "モニターを選択してリアルタイム認識状態を確認してください。", "Wählen Sie einen Monitor, um den Live-Erkennungszustand zu prüfen.", "Seleccione un monitor para revisar su estado de reconocimiento en vivo.", "Sélectionnez un moniteur pour consulter son état de reconnaissance en direct.", "Выберите монитор для просмотра текущего состояния распознавания.", "Selecione um monitor para consultar o estado de reconhecimento em direto.", "실시간 인식 상태를 보려면 모니터를 선택하세요."],
+  ["CURRENT STATE", "当前状态", "現在の状態", "AKTUELLER ZUSTAND", "ESTADO ACTUAL", "ÉTAT ACTUEL", "ТЕКУЩЕЕ СОСТОЯНИЕ", "ESTADO ATUAL", "현재 상태"],
+  ["Source", "源地址", "ソース", "Quelle", "Fuente", "Source", "Источник", "Origem", "소스"],
+  ["Collection default", "人员库默认值", "コレクション既定値", "Collection-Standard", "Valor predeterminado de Collection", "Valeur par défaut de la Collection", "Значение Collection", "Predefinição da Collection", "Collection 기본값"],
+  ["Select a Monitor", "选择监控任务", "モニターを選択", "Monitor auswählen", "Seleccionar un monitor", "Sélectionner un moniteur", "Выберите монитор", "Selecionar monitor", "모니터 선택"],
+  ["Start", "启动", "開始", "Starten", "Iniciar", "Démarrer", "Запустить", "Iniciar", "시작"],
+  ["Stop", "停止", "停止", "Stoppen", "Detener", "Arrêter", "Остановить", "Parar", "중지"],
+  ["Running", "运行中", "実行中", "Läuft", "En ejecución", "En cours", "Работает", "Em execução", "실행 중"],
+  ["Starting", "正在启动", "開始中", "Wird gestartet", "Iniciando", "Démarrage", "Запускается", "A iniciar", "시작 중"],
+  ["Degraded", "降级运行", "縮退運転", "Eingeschränkt", "Degradado", "Dégradé", "Ограниченный режим", "Degradado", "성능 저하"],
+  ["Stopped", "已停止", "停止済み", "Gestoppt", "Detenido", "Arrêté", "Остановлен", "Parado", "중지됨"],
+  ["Error", "错误", "エラー", "Fehler", "Error", "Erreur", "Ошибка", "Erro", "오류"],
+  ["Enable preview", "启用预览", "プレビューを有効化", "Vorschau aktivieren", "Activar vista previa", "Activer l’aperçu", "Включить просмотр", "Ativar pré-visualização", "미리보기 사용"],
+  ["Disable preview", "关闭预览", "プレビューを無効化", "Vorschau deaktivieren", "Desactivar vista previa", "Désactiver l’aperçu", "Отключить просмотр", "Desativar pré-visualização", "미리보기 끄기"],
+  ["Preview is disabled", "预览已关闭", "プレビューは無効です", "Vorschau ist deaktiviert", "La vista previa está desactivada", "L’aperçu est désactivé", "Просмотр выключен", "A pré-visualização está desativada", "미리보기가 꺼져 있습니다"],
+  ["Recognition and events continue without video preview.", "不开启视频预览也会继续识别并产生事件。", "映像プレビューがなくても認識とイベントは継続します。", "Erkennung und Ereignisse laufen ohne Videovorschau weiter.", "El reconocimiento y los eventos continúan sin vista previa.", "La reconnaissance et les événements continuent sans aperçu vidéo.", "Распознавание и события работают без просмотра видео.", "O reconhecimento e os eventos continuam sem pré-visualização.", "영상 미리보기 없이도 인식과 이벤트가 계속됩니다."],
+  ["Start the Monitor to receive frames.", "启动监控任务后即可接收视频帧。", "フレームを受信するにはモニターを開始してください。", "Starten Sie den Monitor, um Frames zu empfangen.", "Inicie el Monitor para recibir fotogramas.", "Démarrez le Monitor pour recevoir des images.", "Запустите Monitor для получения кадров.", "Inicie o Monitor para receber fotogramas.", "프레임을 받으려면 Monitor를 시작하세요."],
+  ["Waiting for preview", "正在等待预览", "プレビュー待機中", "Warten auf Vorschau", "Esperando vista previa", "En attente de l’aperçu", "Ожидание просмотра", "A aguardar pré-visualização", "미리보기 대기 중"],
+  ["The preview will appear after the RTSP source connects.", "RTSP源连接成功后将显示预览。", "RTSP ソースへの接続後にプレビューが表示されます。", "Die Vorschau erscheint nach Verbindung der RTSP-Quelle.", "La vista previa aparecerá al conectar la fuente RTSP.", "L’aperçu apparaîtra après connexion de la source RTSP.", "Просмотр появится после подключения RTSP-источника.", "A pré-visualização surge após ligar a origem RTSP.", "RTSP 소스가 연결되면 미리보기가 표시됩니다."],
+  ["Select a Monitor to inspect it.", "选择一个监控任务进行查看。", "確認するモニターを選択してください。", "Wählen Sie einen Monitor zur Ansicht.", "Seleccione un monitor para revisarlo.", "Sélectionnez un moniteur à consulter.", "Выберите монитор для просмотра.", "Selecione um monitor para o consultar.", "확인할 모니터를 선택하세요."],
+  ["IN-MEMORY BUFFER", "内存缓冲区", "メモリーバッファー", "SPEICHERPUFFER", "BÚFER EN MEMORIA", "TAMPON EN MÉMOIRE", "БУФЕР В ПАМЯТИ", "BUFFER EM MEMÓRIA", "메모리 버퍼"],
+  ["Recent events", "最近事件", "直近のイベント", "Letzte Ereignisse", "Eventos recientes", "Événements récents", "Последние события", "Eventos recentes", "최근 이벤트"],
+  ["Not enrolled", "未入库", "未登録", "Nicht registriert", "No registrado", "Non enregistré", "Не зарегистрировано", "Não registado", "미등록"],
+  ["Monitor is stopped.", "监控任务已停止。", "モニターは停止しています。", "Der Monitor ist gestoppt.", "El monitor está detenido.", "Le moniteur est arrêté.", "Монитор остановлен.", "O monitor está parado.", "모니터가 중지되었습니다."],
+  ["Unknown person", "陌生人", "不明な人物", "Unbekannte Person", "Persona desconocida", "Personne inconnue", "Неизвестный человек", "Pessoa desconhecida", "알 수 없는 사람"],
+  ["Known person entered", "已入库人员进入", "登録人物が入りました", "Registrierte Person eingetreten", "Persona registrada entró", "Personne inscrite entrée", "Зарегистрированный человек вошёл", "Pessoa registada entrou", "등록 인물 입장"],
+  ["Known person left", "已入库人员离开", "登録人物が退出しました", "Registrierte Person gegangen", "Persona registrada salió", "Personne inscrite sortie", "Зарегистрированный человек вышел", "Pessoa registada saiu", "등록 인물 퇴장"],
+  ["Unknown person entered", "陌生人进入", "未登録人物が入りました", "Unbekannte Person eingetreten", "Persona desconocida entró", "Personne inconnue entrée", "Неизвестный человек вошёл", "Pessoa desconhecida entrou", "미등록 인물 입장"],
+  ["Unknown person left", "陌生人离开", "未登録人物が退出しました", "Unbekannte Person gegangen", "Persona desconocida salió", "Personne inconnue sortie", "Неизвестный человек вышел", "Pessoa desconhecida saiu", "미등록 인물 퇴장"],
+  ["Monitor error", "监控任务错误", "モニターエラー", "Monitorfehler", "Error del Monitor", "Erreur du Monitor", "Ошибка Monitor", "Erro do Monitor", "Monitor 오류"],
+  ["Monitor recovered", "监控任务已恢复", "モニターが復旧しました", "Monitor wiederhergestellt", "Monitor recuperado", "Monitor rétabli", "Monitor восстановлен", "Monitor recuperado", "Monitor 복구됨"],
+  ["Event", "事件", "イベント", "Ereignis", "Evento", "Événement", "Событие", "Evento", "이벤트"],
+  ["Similarity {score}", "相似度 {score}", "類似度 {score}", "Ähnlichkeit {score}", "Similitud {score}", "Similarité {score}", "Сходство {score}", "Similaridade {score}", "유사도 {score}"],
+  ["Delete Monitor {id}?", "删除监控任务 {id}？", "モニター {id} を削除しますか？", "Monitor {id} löschen?", "¿Eliminar el monitor {id}?", "Supprimer le moniteur {id} ?", "Удалить монитор {id}?", "Eliminar o monitor {id}?", "모니터 {id}을(를) 삭제할까요?"],
+  ["Monitor created", "监控任务已创建", "モニターを作成しました", "Monitor erstellt", "Monitor creado", "Moniteur créé", "Монитор создан", "Monitor criado", "모니터가 생성되었습니다"],
+  ["Monitor updated", "监控任务已更新", "モニターを更新しました", "Monitor aktualisiert", "Monitor actualizado", "Moniteur mis à jour", "Монитор обновлён", "Monitor atualizado", "모니터가 업데이트되었습니다"],
+  ["Monitor deleted", "监控任务已删除", "モニターを削除しました", "Monitor gelöscht", "Monitor eliminado", "Moniteur supprimé", "Монитор удалён", "Monitor eliminado", "모니터가 삭제되었습니다"],
+  ["Could not load Monitors", "无法加载监控任务", "モニターを読み込めませんでした", "Monitore konnten nicht geladen werden", "No se pudieron cargar los monitores", "Impossible de charger les moniteurs", "Не удалось загрузить мониторы", "Não foi possível carregar os monitores", "모니터를 불러올 수 없습니다"],
+  ["Monitor could not be created", "无法创建监控任务", "モニターを作成できませんでした", "Monitor konnte nicht erstellt werden", "No se pudo crear el monitor", "Impossible de créer le moniteur", "Не удалось создать монитор", "Não foi possível criar o monitor", "모니터를 만들 수 없습니다"],
+  ["Monitor could not be updated", "无法更新监控任务", "モニターを更新できませんでした", "Monitor konnte nicht aktualisiert werden", "No se pudo actualizar el monitor", "Impossible de mettre à jour le moniteur", "Не удалось обновить монитор", "Não foi possível atualizar o monitor", "모니터를 업데이트할 수 없습니다"],
+  ["Monitor preview could not be updated", "无法更新监控预览配置", "モニタープレビューを更新できませんでした", "Monitorvorschau konnte nicht aktualisiert werden", "No se pudo actualizar la vista previa", "Impossible de mettre à jour l’aperçu", "Не удалось обновить просмотр монитора", "Não foi possível atualizar a pré-visualização", "모니터 미리보기를 업데이트할 수 없습니다"],
+  ["Monitor could not be deleted", "无法删除监控任务", "モニターを削除できませんでした", "Monitor konnte nicht gelöscht werden", "No se pudo eliminar el monitor", "Impossible de supprimer le moniteur", "Не удалось удалить монитор", "Não foi possível eliminar o monitor", "모니터를 삭제할 수 없습니다"],
+  ["Stopping…", "正在停止…", "停止中…", "Wird gestoppt…", "Deteniendo…", "Arrêt…", "Остановка…", "A parar…", "중지 중…"],
+  ["Deleting…", "正在删除…", "削除中…", "Wird gelöscht…", "Eliminando…", "Suppression…", "Удаление…", "A eliminar…", "삭제 중…"],
+];
+
+export const SHARED_UI_MESSAGES = new Set([
+  "InsightFace Server", "InsightFace", "INSIGHTFACE SERVER",
+  "English", "中文", "日本語", "Deutsch", "Español", "Français", "Русский", "Português", "한국어",
+  "☰", "○", "●", "↻", "—", "×", "＋", "⌗", "⌕", "▶", "✓",
+  "employees", "employee-001", "Alice", "HR-1001", "[[0.0123, -0.0456, ...]]",
+  "FP32", "FP16", "BF16", "INT8 ×736", "INT8 ×1000", "0.40", "1 FPS", "README", "OpenAPI JSON",
+  "cuDNN", "/data", "/models", "rtsp://camera.example:8554/stream", "front-gate", "0",
+]);
+
+for (const row of rows) {
+  if (row.length !== localeCodes.length + 1) throw new Error(`Incomplete UI translation row: ${row[0]}`);
+}
+
+export const UI_CATALOGS = Object.freeze(Object.fromEntries(
+  localeCodes.map((code, localeIndex) => [
+    code,
+    Object.freeze(Object.fromEntries(rows.map((row) => [row[0], row[localeIndex + 1]]))),
+  ]),
+));
diff --git a/server/migrations/0001_initial.sql b/server/migrations/0001_initial.sql
new file mode 100644
index 000000000..9ea7f9089
--- /dev/null
+++ b/server/migrations/0001_initial.sql
@@ -0,0 +1,60 @@
+CREATE TABLE collections (
+    id TEXT PRIMARY KEY,
+    name TEXT NOT NULL,
+    description TEXT NOT NULL DEFAULT '',
+    default_threshold REAL NOT NULL CHECK(default_threshold BETWEEN 0 AND 1),
+    model_id TEXT NOT NULL,
+    model_version TEXT NOT NULL,
+    model_digest TEXT NOT NULL,
+    embedding_dimension INTEGER NOT NULL,
+    preprocessing_version TEXT NOT NULL,
+    metadata_json TEXT NOT NULL DEFAULT '{}',
+    created_at TEXT NOT NULL,
+    updated_at TEXT NOT NULL
+);
+
+CREATE TABLE persons (
+    collection_id TEXT NOT NULL REFERENCES collections(id) ON DELETE CASCADE,
+    id TEXT NOT NULL,
+    name TEXT,
+    external_id TEXT,
+    metadata_json TEXT NOT NULL DEFAULT '{}',
+    created_at TEXT NOT NULL,
+    updated_at TEXT NOT NULL,
+    PRIMARY KEY(collection_id, id),
+    UNIQUE(collection_id, external_id)
+);
+
+CREATE TABLE face_samples (
+    id TEXT PRIMARY KEY,
+    collection_id TEXT NOT NULL,
+    person_id TEXT NOT NULL,
+    embedding BLOB NOT NULL,
+    embedding_dimension INTEGER NOT NULL,
+    bounding_box_json TEXT NOT NULL,
+    landmarks_json TEXT,
+    detection_score REAL NOT NULL CHECK(detection_score BETWEEN 0 AND 1),
+    quality_json TEXT NOT NULL,
+    model_id TEXT NOT NULL,
+    model_version TEXT NOT NULL,
+    model_digest TEXT NOT NULL,
+    preprocessing_version TEXT NOT NULL,
+    crop_path TEXT,
+    created_at TEXT NOT NULL,
+    FOREIGN KEY(collection_id, person_id)
+        REFERENCES persons(collection_id, id) ON DELETE CASCADE
+);
+
+CREATE INDEX face_samples_collection_person
+    ON face_samples(collection_id, person_id, id);
+CREATE INDEX persons_collection_external
+    ON persons(collection_id, external_id);
+
+CREATE TABLE api_keys (
+    id TEXT PRIMARY KEY,
+    label TEXT NOT NULL,
+    salt BLOB NOT NULL,
+    digest BLOB NOT NULL,
+    active INTEGER NOT NULL DEFAULT 1,
+    created_at TEXT NOT NULL
+);
diff --git a/server/migrations/0002_native_search.sql b/server/migrations/0002_native_search.sql
new file mode 100644
index 000000000..cfe6fba0c
--- /dev/null
+++ b/server/migrations/0002_native_search.sql
@@ -0,0 +1,87 @@
+ALTER TABLE collections ADD COLUMN search_profile TEXT NOT NULL DEFAULT 'fp32_v1'
+    CHECK(search_profile IN ('fp32_v1', 'fp16_v1', 'bf16_v1', 'int8_x1000_v1'));
+ALTER TABLE collections ADD COLUMN capacity_rows INTEGER NOT NULL DEFAULT 100000
+    CHECK(capacity_rows BETWEEN 1 AND 9223372036854775807);
+ALTER TABLE collections ADD COLUMN max_faces_per_person INTEGER NOT NULL DEFAULT 20
+    CHECK(max_faces_per_person BETWEEN 1 AND 1000000);
+ALTER TABLE collections ADD COLUMN load_policy TEXT NOT NULL DEFAULT 'lazy'
+    CHECK(load_policy IN ('lazy', 'eager'));
+ALTER TABLE collections ADD COLUMN search_revision INTEGER NOT NULL DEFAULT 0
+    CHECK(search_revision >= 0);
+
+-- Native search uses unsigned 64-bit identifiers. SQLite INTEGER is signed, so
+-- generated identifiers intentionally stay in the positive signed 63-bit range.
+-- AUTOINCREMENT makes an identifier stable and prevents its reuse after delete.
+CREATE TABLE search_person_ids (
+    numeric_id INTEGER PRIMARY KEY AUTOINCREMENT
+        CHECK(numeric_id BETWEEN 1 AND 9223372036854775807),
+    collection_id TEXT NOT NULL,
+    person_id TEXT NOT NULL,
+    UNIQUE(collection_id, person_id),
+    UNIQUE(collection_id, numeric_id),
+    FOREIGN KEY(collection_id, person_id)
+        REFERENCES persons(collection_id, id) ON DELETE CASCADE
+);
+
+CREATE TABLE search_vector_ids (
+    numeric_id INTEGER PRIMARY KEY AUTOINCREMENT
+        CHECK(numeric_id BETWEEN 1 AND 9223372036854775807),
+    collection_id TEXT NOT NULL,
+    person_id TEXT NOT NULL,
+    person_numeric_id INTEGER NOT NULL,
+    face_id TEXT NOT NULL UNIQUE,
+    UNIQUE(collection_id, numeric_id),
+    FOREIGN KEY(collection_id, person_id)
+        REFERENCES persons(collection_id, id) ON DELETE CASCADE,
+    FOREIGN KEY(collection_id, person_numeric_id)
+        REFERENCES search_person_ids(collection_id, numeric_id) ON DELETE CASCADE,
+    FOREIGN KEY(face_id)
+        REFERENCES face_samples(id) ON DELETE CASCADE
+);
+
+CREATE INDEX search_vector_ids_collection_person
+    ON search_vector_ids(collection_id, person_numeric_id, numeric_id);
+
+-- Backfill stable IDs for databases created before native search support.
+INSERT INTO search_person_ids(collection_id, person_id)
+SELECT collection_id, id FROM persons ORDER BY collection_id, id;
+
+INSERT INTO search_vector_ids(
+    collection_id, person_id, person_numeric_id, face_id
+)
+SELECT f.collection_id, f.person_id, p.numeric_id, f.id
+FROM face_samples f
+JOIN search_person_ids p
+  ON p.collection_id=f.collection_id AND p.person_id=f.person_id
+ORDER BY f.collection_id, f.person_id, f.id;
+
+-- This pending-only durable outbox is deliberately not linked by foreign keys.
+-- Delete events must survive deletion of their Collection, Person, FaceSample
+-- and ID mappings until the in-memory index acknowledges their global seq.
+CREATE TABLE search_changes (
+    seq INTEGER PRIMARY KEY AUTOINCREMENT
+        CHECK(seq BETWEEN 1 AND 9223372036854775807),
+    collection_id TEXT NOT NULL,
+    revision INTEGER NOT NULL CHECK(revision >= 1),
+    operation TEXT NOT NULL CHECK(operation IN ('add', 'delete', 'clear')),
+    vector_id INTEGER,
+    person_numeric_id INTEGER,
+    face_id TEXT,
+    created_at TEXT NOT NULL,
+    CHECK(
+        (operation IN ('add', 'delete')
+          AND vector_id IS NOT NULL
+          AND person_numeric_id IS NOT NULL
+          AND face_id IS NOT NULL)
+        OR
+        (operation='clear'
+          AND vector_id IS NULL
+          AND person_numeric_id IS NULL
+          AND face_id IS NULL)
+    )
+);
+
+CREATE INDEX search_changes_collection_seq
+    ON search_changes(collection_id, seq);
+CREATE INDEX search_changes_collection_revision
+    ON search_changes(collection_id, revision, seq);
diff --git a/server/migrations/0003_int8_x736.sql b/server/migrations/0003_int8_x736.sql
new file mode 100644
index 000000000..29dae85d9
--- /dev/null
+++ b/server/migrations/0003_int8_x736.sql
@@ -0,0 +1,17 @@
+-- Keep the profile string as the durable quantization contract. SQLite cannot
+-- widen an existing CHECK constraint in place, so replace only this column;
+-- all Collection, Person, FaceSample, and stable search-ID rows stay intact.
+ALTER TABLE collections RENAME COLUMN search_profile TO search_profile_legacy;
+
+ALTER TABLE collections ADD COLUMN search_profile TEXT NOT NULL DEFAULT 'fp32_v1'
+    CHECK(search_profile IN (
+        'fp32_v1',
+        'fp16_v1',
+        'bf16_v1',
+        'int8_x1000_v1',
+        'int8_x736_v1'
+    ));
+
+UPDATE collections SET search_profile = search_profile_legacy;
+
+ALTER TABLE collections DROP COLUMN search_profile_legacy;
diff --git a/server/migrations/0004_collection_face_crops.sql b/server/migrations/0004_collection_face_crops.sql
new file mode 100644
index 000000000..d88a6c7eb
--- /dev/null
+++ b/server/migrations/0004_collection_face_crops.sql
@@ -0,0 +1,6 @@
+ALTER TABLE collections
+    ADD COLUMN save_face_crops INTEGER NOT NULL DEFAULT 0
+    CHECK(save_face_crops IN (0, 1));
+
+ALTER TABLE face_samples ADD COLUMN crop_image BLOB;
+ALTER TABLE face_samples ADD COLUMN crop_media_type TEXT;
diff --git a/server/migrations/0005_external_trusted_embeddings.sql b/server/migrations/0005_external_trusted_embeddings.sql
new file mode 100644
index 000000000..39f163639
--- /dev/null
+++ b/server/migrations/0005_external_trusted_embeddings.sql
@@ -0,0 +1,5 @@
+ALTER TABLE face_samples
+    ADD COLUMN embedding_source TEXT NOT NULL DEFAULT 'server'
+    CHECK(embedding_source IN ('server', 'external_trusted'));
+
+ALTER TABLE face_samples ADD COLUMN embedding_contract_id TEXT;
diff --git a/server/migrations/0006_collection_detection_profiles.sql b/server/migrations/0006_collection_detection_profiles.sql
new file mode 100644
index 000000000..7aef02988
--- /dev/null
+++ b/server/migrations/0006_collection_detection_profiles.sql
@@ -0,0 +1,19 @@
+ALTER TABLE collections
+    ADD COLUMN detector_input_sizes_json TEXT NOT NULL
+    DEFAULT '[[96,96],[512,512]]';
+
+ALTER TABLE collections
+    ADD COLUMN detector_threshold REAL NOT NULL DEFAULT 0.50
+    CHECK(detector_threshold BETWEEN 0 AND 1);
+
+ALTER TABLE collections
+    ADD COLUMN detector_nms_threshold REAL NOT NULL DEFAULT 0.40
+    CHECK(detector_nms_threshold BETWEEN 0 AND 1);
+
+ALTER TABLE collections
+    ADD COLUMN single_face_selection TEXT NOT NULL DEFAULT 'largest'
+    CHECK(single_face_selection IN ('largest', 'center_largest'));
+
+ALTER TABLE collections
+    ADD COLUMN detection_revision INTEGER NOT NULL DEFAULT 1
+    CHECK(detection_revision >= 1);
diff --git a/server/migrations/0007_monitors.sql b/server/migrations/0007_monitors.sql
new file mode 100644
index 000000000..cee10144c
--- /dev/null
+++ b/server/migrations/0007_monitors.sql
@@ -0,0 +1,33 @@
+CREATE TABLE monitors (
+    id TEXT PRIMARY KEY,
+    name TEXT NOT NULL,
+    description TEXT NOT NULL DEFAULT '',
+    enabled INTEGER NOT NULL DEFAULT 1
+        CHECK(enabled IN (0, 1)),
+    source_type TEXT NOT NULL DEFAULT 'rtsp'
+        CHECK(source_type = 'rtsp'),
+    source_url_ciphertext TEXT NOT NULL,
+    collection_id TEXT NOT NULL
+        REFERENCES collections(id) ON DELETE RESTRICT,
+    inference_fps REAL NOT NULL DEFAULT 2.0
+        CHECK(inference_fps BETWEEN 0.1 AND 30.0),
+    match_threshold REAL
+        CHECK(match_threshold IS NULL OR match_threshold BETWEEN 0.0 AND 1.0),
+    event_buffer_size INTEGER NOT NULL DEFAULT 1000
+        CHECK(event_buffer_size BETWEEN 10 AND 10000),
+    confirm_frames INTEGER NOT NULL DEFAULT 3
+        CHECK(confirm_frames BETWEEN 1 AND 100),
+    absence_timeout_seconds REAL NOT NULL DEFAULT 3.0
+        CHECK(absence_timeout_seconds BETWEEN 0.1 AND 300.0),
+    cooldown_seconds REAL NOT NULL DEFAULT 10.0
+        CHECK(cooldown_seconds BETWEEN 0.0 AND 3600.0),
+    emit_unknown INTEGER NOT NULL DEFAULT 1
+        CHECK(emit_unknown IN (0, 1)),
+    preview_enabled INTEGER NOT NULL DEFAULT 0
+        CHECK(preview_enabled IN (0, 1)),
+    created_at TEXT NOT NULL,
+    updated_at TEXT NOT NULL
+);
+
+CREATE INDEX monitors_collection_id ON monitors(collection_id, id);
+CREATE INDEX monitors_enabled ON monitors(enabled, id);
diff --git a/server/native/search/CMakeLists.txt b/server/native/search/CMakeLists.txt
new file mode 100644
index 000000000..f2fe87146
--- /dev/null
+++ b/server/native/search/CMakeLists.txt
@@ -0,0 +1,24 @@
+cmake_minimum_required(VERSION 3.22)
+project(insightface_search_native VERSION 2.0.0 LANGUAGES CXX)
+
+include(GNUInstallDirs)
+
+option(IFS_SEARCH_BUILD_CPU "Build the CPU exact-search library" ON)
+option(IFS_SEARCH_BUILD_CUDA "Build the CUDA exact-search library" OFF)
+option(IFS_SEARCH_BUILD_TESTS "Build native smoke tests" ON)
+
+if(IFS_SEARCH_BUILD_TESTS)
+    enable_testing()
+endif()
+
+if(IFS_SEARCH_BUILD_CPU)
+    add_subdirectory(cpu)
+endif()
+
+if(IFS_SEARCH_BUILD_CUDA)
+    enable_language(CUDA)
+    add_subdirectory(cuda)
+endif()
+
+install(FILES include/ifs_search.h
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
diff --git a/server/native/search/README.md b/server/native/search/README.md
new file mode 100644
index 000000000..66d02bb76
--- /dev/null
+++ b/server/native/search/README.md
@@ -0,0 +1,113 @@
+# InsightFace Server native exact search
+
+This directory contains the product-facing native search ABI used by the
+Server. It is derived from the standalone benchmark implementation, but does
+not contain benchmark datasets, extracted embeddings, or result files.
+
+## ABI and score contract
+
+Both `libifs_search_cpu` and `libifs_search_cuda` export the same C ABI from
+`include/ifs_search.h`. ABI v2 is fixed to 512-dimensional input. Every input
+vector and query must be finite, FP32, and L2-normalized before it crosses the
+ABI boundary.
+
+All returned scores use **raw cosine** semantics. Because the input is
+normalized, FP32 inner product equals cosine similarity. FP16 and BF16 return
+their low-precision approximation. INT8 uses the per-index scale encoded by
+the profile (`S=736` or legacy-compatible `S=1000`):
+
+```text
+q = clamp(round_half_away_from_zero(x * S), -128, 127)
+score_internal = int32_dot(q_database, q_query) / (S * S)
+similarity = clamp(score_internal, -1, 1)
+```
+
+CPU, CUDA, and the NumPy reference deliberately perform the multiply and
+half-away-from-zero rounding with FP32 semantics. The native index uses the
+unclipped internal score for exact ordering and grouped best-Face selection,
+then clamps scores returned by the ABI to the public cosine range. The unscaled
+INT32 accumulator is never exposed by the production ABI.
+
+Profiles:
+
+- `FP32_V1`: CPU and CUDA
+- `FP16_V1`: CUDA only; CPU returns `IFS_SEARCH_UNSUPPORTED`
+- `BF16_V1`: CPU and CUDA
+- `INT8_X736_V1`: CPU and CUDA; recommended INT8 profile
+- `INT8_X1000_V1`: CPU and CUDA
+
+The two INT8 scales are separate per-index contracts and can coexist in one
+process. Profile code 3 remains x1000; x736 appends code 4. Existing x1000
+Collections are never silently reinterpreted as x736.
+
+Unsupported profiles and CUDA failures are fail-closed. There is no dtype or
+CPU fallback inside either native library.
+
+The CUDA capability mask is device-specific. In particular, BF16 is exposed
+only on Ampere/SM80 or newer; Turing still supports FP32, FP16, and INT8. Index
+creation rechecks the selected device and rejects unsupported profiles.
+
+## Capacity and deletion
+
+`reserve_rows` reserves contiguous storage at creation. `max_rows` is the hard
+logical/physical row limit checked before add. Configure them to the same value
+for a production Collection. CUDA reserves vector/group metadata, score and
+delete workspaces, and face/group candidate buffers for exact Top-K up to 100,
+so a within-capacity query does not trigger a first-query workspace allocation.
+Automatic growth replaces all of those buffers together and preserves the
+stable group metadata.
+
+CPU deletion reuses freed slots. CUDA deletion uses tombstones, so
+`physical_rows` continues increasing after delete/add cycles. The Server must
+rebuild a CUDA collection generation before tombstones consume its capacity.
+
+## Person Top-K
+
+ABI v2 accepts a stable numeric group/Person ID for every FaceSample and
+implements strict grouped Top-K. Every live face score participates, each
+Person retains its highest-scoring face (then the lowest vector ID for an exact
+score tie), and Persons are sorted by score descending and group ID ascending.
+It never uses a fixed face-level oversample, so it is exact for the Server's
+`max FaceSample score` Person policy.
+
+The CUDA implementation keeps row-to-group metadata on the device. A two-pass
+GPU reduction first finds each Person's best score and then its deterministic
+best FaceSample; the existing multi-stage exact Top-K reduction selects at most
+100 Persons. Only those final K `(group_id, vector_id, score)` records cross
+PCIe. Incremental add, tombstone delete, group reactivation, explicit reserve,
+and automatic growth all preserve the device group metadata.
+
+CUDA sets `IFS_SEARCH_CAP_GROUPED_DEVICE_RESIDENT` together with
+`IFS_SEARCH_CAP_GROUPED_PERSON_TOPK`, and does not set the host-reference flag.
+CPU continues to set `IFS_SEARCH_CAP_GROUPED_HOST_REFERENCE`: its exact grouped
+implementation still materializes face scores in host memory. Both paths have
+identical ordering and raw-cosine score semantics.
+
+## Build and smoke test
+
+CPU:
+
+```bash
+cmake -S server/native/search -B build/native-search \
+  -DIFS_SEARCH_BUILD_CPU=ON \
+  -DIFS_SEARCH_BUILD_CUDA=OFF \
+  -DIFS_SEARCH_BUILD_TESTS=ON
+cmake --build build/native-search -j
+ctest --test-dir build/native-search --output-on-failure
+```
+
+CUDA 12.9:
+
+```bash
+cmake -S server/native/search -B build/native-search-cuda \
+  -DIFS_SEARCH_BUILD_CPU=OFF \
+  -DIFS_SEARCH_BUILD_CUDA=ON \
+  -DIFS_SEARCH_BUILD_TESTS=ON
+cmake --build build/native-search-cuda -j
+ctest --test-dir build/native-search-cuda --output-on-failure
+```
+
+The default CUDA list contains SASS for 75, 80, 86, 89, 90, 100, 103, and 120,
+plus PTX for the newest target. It is intentionally explicit rather than
+`native`; release builds must contain the architectures claimed by the Server
+compatibility document. Hardware certification remains release-gating work.
diff --git a/server/native/search/cpu/CMakeLists.txt b/server/native/search/cpu/CMakeLists.txt
new file mode 100644
index 000000000..10c7f77f1
--- /dev/null
+++ b/server/native/search/cpu/CMakeLists.txt
@@ -0,0 +1,44 @@
+add_library(ifs_search_cpu SHARED
+    src/api_v2.cpp
+    src/index_legacy.cpp
+    src/kernels_portable.cpp
+    src/kernels_x86.cpp)
+
+target_compile_features(ifs_search_cpu PRIVATE cxx_std_17)
+target_compile_definitions(ifs_search_cpu PRIVATE IFS_SEARCH_BUILDING=1)
+target_include_directories(ifs_search_cpu
+    PUBLIC ${PROJECT_SOURCE_DIR}/include
+    PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+set_target_properties(ifs_search_cpu PROPERTIES
+    CXX_VISIBILITY_PRESET hidden
+    VISIBILITY_INLINES_HIDDEN YES
+    OUTPUT_NAME ifs_search_cpu
+    VERSION ${PROJECT_VERSION}
+    SOVERSION 2)
+install(TARGETS ifs_search_cpu
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+if(MSVC)
+    target_compile_options(ifs_search_cpu PRIVATE /O2 /W4)
+else()
+    # ISA kernels use function-level target attributes and runtime CPUID. Do
+    # not use -march=native: a release .so must start on older x86_64 hosts.
+    target_compile_options(ifs_search_cpu PRIVATE -O3 -Wall -Wextra -Wpedantic)
+    find_package(OpenMP)
+    if(OpenMP_CXX_FOUND)
+        target_link_libraries(ifs_search_cpu PRIVATE OpenMP::OpenMP_CXX)
+    endif()
+endif()
+
+if(IFS_SEARCH_BUILD_TESTS)
+    add_executable(ifs_search_cpu_smoke tests/smoke.cpp)
+    target_compile_features(ifs_search_cpu_smoke PRIVATE cxx_std_17)
+    target_link_libraries(ifs_search_cpu_smoke PRIVATE ifs_search_cpu)
+    if(MSVC)
+        target_compile_options(ifs_search_cpu_smoke PRIVATE /UNDEBUG)
+    else()
+        target_compile_options(ifs_search_cpu_smoke PRIVATE -UNDEBUG)
+    endif()
+    add_test(NAME ifs_search_cpu_smoke COMMAND ifs_search_cpu_smoke)
+endif()
diff --git a/server/native/search/cpu/src/api_v2.cpp b/server/native/search/cpu/src/api_v2.cpp
new file mode 100644
index 000000000..e4252c27e
--- /dev/null
+++ b/server/native/search/cpu/src/api_v2.cpp
@@ -0,0 +1,646 @@
+#include "ifs_search.h"
+
+#include "ifs_cpu_legacy.h"
+
+#include <algorithm>
+#include <chrono>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <shared_mutex>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+namespace {
+
+constexpr double kDefaultGrowthFactor = 2.0;
+constexpr double kNormalizedSquaredTolerance = 1.0e-3;
+
+thread_local std::string g_last_error;
+
+struct CpuIndex {
+    ifs_cpu_index_t core = nullptr;
+    ifs_search_profile_t profile = IFS_SEARCH_PROFILE_FP32_V1;
+    uint64_t max_rows = 0;
+    uint64_t physical_rows = 0;
+    uint64_t live_rows = 0;
+    uint64_t tombstone_rows = 0;
+    uint64_t reallocations = 0;
+    double growth_factor = kDefaultGrowthFactor;
+    std::unordered_map<uint64_t, uint64_t> vector_to_group;
+    mutable std::shared_mutex state_mutex;
+
+    ~CpuIndex() {
+        ifs_cpu_index_destroy(core);
+    }
+};
+
+ifs_search_status_t fail(ifs_search_status_t status, const std::string &message) {
+    g_last_error = message;
+    return status;
+}
+
+void clear_error() {
+    g_last_error.clear();
+}
+
+float public_cosine(float score) {
+    return std::clamp(score, -1.0f, 1.0f);
+}
+
+CpuIndex *as_index(ifs_search_index_t index) {
+    return static_cast<CpuIndex *>(index);
+}
+
+bool valid_profile(uint32_t profile) {
+    return profile == IFS_SEARCH_PROFILE_FP32_V1 ||
+           profile == IFS_SEARCH_PROFILE_BF16_V1 ||
+           profile == IFS_SEARCH_PROFILE_INT8_X1000_V1 ||
+           profile == IFS_SEARCH_PROFILE_INT8_X736_V1;
+}
+
+uint32_t int8_scale_for(uint32_t profile) {
+    switch (profile) {
+        case IFS_SEARCH_PROFILE_INT8_X1000_V1:
+            return IFS_SEARCH_INT8_X1000_SCALE;
+        case IFS_SEARCH_PROFILE_INT8_X736_V1:
+            return IFS_SEARCH_INT8_X736_SCALE;
+        default:
+            return 0;
+    }
+}
+
+float int8_score_divisor(ifs_search_profile_t profile) {
+    const float scale = static_cast<float>(int8_scale_for(profile));
+    return scale * scale;
+}
+
+ifs_cpu_storage_t storage_for(uint32_t profile) {
+    switch (profile) {
+        case IFS_SEARCH_PROFILE_FP32_V1: return IFS_CPU_STORAGE_FP32;
+        case IFS_SEARCH_PROFILE_BF16_V1: return IFS_CPU_STORAGE_BF16;
+        case IFS_SEARCH_PROFILE_INT8_X1000_V1: return IFS_CPU_STORAGE_INT8;
+        case IFS_SEARCH_PROFILE_INT8_X736_V1: return IFS_CPU_STORAGE_INT8;
+        default: throw std::invalid_argument("profile is not supported by the CPU backend");
+    }
+}
+
+uint64_t bytes_per_vector(ifs_search_profile_t profile) {
+    switch (profile) {
+        case IFS_SEARCH_PROFILE_FP32_V1:
+            return IFS_SEARCH_DIMENSION * sizeof(float);
+        case IFS_SEARCH_PROFILE_BF16_V1:
+            return IFS_SEARCH_DIMENSION * sizeof(uint16_t);
+        case IFS_SEARCH_PROFILE_INT8_X1000_V1:
+        case IFS_SEARCH_PROFILE_INT8_X736_V1:
+            return IFS_SEARCH_DIMENSION * sizeof(int8_t);
+        default:
+            return 0;
+    }
+}
+
+bool validate_normalized_rows(const float *vectors, uint64_t count,
+                              std::string *message) {
+    if (count && vectors == nullptr) {
+        *message = "vectors must not be null for a non-empty batch";
+        return false;
+    }
+    if (count > std::numeric_limits<size_t>::max() / IFS_SEARCH_DIMENSION) {
+        *message = "vector batch is too large for this process";
+        return false;
+    }
+    for (uint64_t row = 0; row < count; ++row) {
+        double squared_norm = 0.0;
+        const float *vector = vectors + static_cast<size_t>(row) * IFS_SEARCH_DIMENSION;
+        for (size_t column = 0; column < IFS_SEARCH_DIMENSION; ++column) {
+            const float value = vector[column];
+            if (!std::isfinite(value)) {
+                *message = "vector contains a non-finite value";
+                return false;
+            }
+            squared_norm += static_cast<double>(value) * value;
+        }
+        if (std::abs(squared_norm - 1.0) > kNormalizedSquaredTolerance) {
+            *message = "vector must be L2-normalized before indexing";
+            return false;
+        }
+    }
+    return true;
+}
+
+ifs_search_status_t map_legacy_status(ifs_cpu_status_t status) {
+    switch (status) {
+        case IFS_CPU_OK: return IFS_SEARCH_OK;
+        case IFS_CPU_INVALID_ARGUMENT: return IFS_SEARCH_INVALID_ARGUMENT;
+        case IFS_CPU_OUT_OF_MEMORY: return IFS_SEARCH_OUT_OF_MEMORY;
+        case IFS_CPU_DUPLICATE_ID: return IFS_SEARCH_DUPLICATE_ID;
+        case IFS_CPU_ID_NOT_FOUND: return IFS_SEARCH_ID_NOT_FOUND;
+        case IFS_CPU_INTERNAL_ERROR: return IFS_SEARCH_INTERNAL_ERROR;
+    }
+    return IFS_SEARCH_INTERNAL_ERROR;
+}
+
+ifs_search_status_t legacy_failure(ifs_cpu_status_t status,
+                                   const char *fallback) {
+    const char *legacy = ifs_cpu_last_error();
+    return fail(map_legacy_status(status),
+                legacy && legacy[0] ? legacy : fallback);
+}
+
+ifs_search_status_t reserve_locked(CpuIndex *index, uint64_t rows,
+                                   bool count_reallocation) {
+    if (index->max_rows && rows > index->max_rows) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "requested capacity exceeds the configured max_rows");
+    }
+    if (rows > std::numeric_limits<size_t>::max()) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "requested capacity exceeds the process address space");
+    }
+    const size_t before = ifs_cpu_index_capacity(index->core);
+    if (rows <= before) return IFS_SEARCH_OK;
+    const ifs_cpu_status_t status = ifs_cpu_index_reserve(
+        index->core, static_cast<size_t>(rows));
+    if (status != IFS_CPU_OK) return legacy_failure(status, "CPU reserve failed");
+    if (count_reallocation) ++index->reallocations;
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ensure_add_capacity(CpuIndex *index, uint64_t count) {
+    const uint64_t reusable = std::min(count, index->tombstone_rows);
+    const uint64_t append_rows = count - reusable;
+    if (append_rows > std::numeric_limits<uint64_t>::max() - index->physical_rows) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED, "index row count overflow");
+    }
+    const uint64_t required = index->physical_rows + append_rows;
+    if (index->max_rows && required > index->max_rows) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "add would exceed the configured max_rows");
+    }
+    const uint64_t current = static_cast<uint64_t>(
+        ifs_cpu_index_capacity(index->core));
+    if (required <= current) return IFS_SEARCH_OK;
+    uint64_t grown = current ? current : 1;
+    while (grown < required) {
+        const long double candidate = std::ceil(
+            static_cast<long double>(grown) * index->growth_factor);
+        if (candidate > std::numeric_limits<uint64_t>::max()) {
+            grown = required;
+            break;
+        }
+        const uint64_t next = static_cast<uint64_t>(candidate);
+        if (next <= grown) {
+            grown = required;
+            break;
+        }
+        grown = next;
+    }
+    if (index->max_rows) grown = std::min(grown, index->max_rows);
+    return reserve_locked(index, std::max(grown, required), true);
+}
+
+void zero_timings(ifs_search_timings_t *timings) {
+    if (!timings) return;
+    timings->kernel_ms = 0.0;
+    timings->topk_ms = 0.0;
+    timings->total_ms = 0.0;
+}
+
+}  // namespace
+
+extern "C" {
+
+uint32_t ifs_search_abi_version(void) { return IFS_SEARCH_ABI_VERSION; }
+
+uint32_t ifs_search_dimension(void) { return IFS_SEARCH_DIMENSION; }
+
+const char *ifs_search_build_info(void) {
+    static thread_local std::string info;
+    const char *features = ifs_cpu_runtime_features();
+    info = "ifs-search-cpu ABI=2 d512 exact-flat raw-cosine "
+           "profiles=fp32,bf16,int8_x1000,int8_x736 "
+           "grouped_topk=host-reference; ";
+    info += features && features[0] ? features : "runtime-features-unavailable";
+    return info.c_str();
+}
+
+const char *ifs_search_last_error(void) { return g_last_error.c_str(); }
+
+const char *ifs_search_status_string(ifs_search_status_t status) {
+    switch (status) {
+        case IFS_SEARCH_OK: return "ok";
+        case IFS_SEARCH_INVALID_ARGUMENT: return "invalid_argument";
+        case IFS_SEARCH_OUT_OF_MEMORY: return "out_of_memory";
+        case IFS_SEARCH_UNSUPPORTED: return "unsupported";
+        case IFS_SEARCH_DUPLICATE_ID: return "duplicate_id";
+        case IFS_SEARCH_ID_NOT_FOUND: return "id_not_found";
+        case IFS_SEARCH_CAPACITY_EXCEEDED: return "capacity_exceeded";
+        case IFS_SEARCH_BACKEND_ERROR: return "backend_error";
+        case IFS_SEARCH_INTERNAL_ERROR: return "internal_error";
+    }
+    return "unknown_status";
+}
+
+ifs_search_status_t ifs_search_get_capabilities(
+    int32_t device, ifs_search_capabilities_t *out) {
+    clear_error();
+    if (!out || out->struct_size < sizeof(*out)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "capabilities struct_size is too small");
+    }
+    if (device != -1 && device != 0) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "the CPU backend accepts device -1 or 0 only");
+    }
+    const uint32_t requested_size = out->struct_size;
+    *out = {};
+    out->struct_size = requested_size;
+    out->abi_version = IFS_SEARCH_ABI_VERSION;
+    out->dimension = IFS_SEARCH_DIMENSION;
+    out->backend = IFS_SEARCH_BACKEND_CPU;
+    out->profile_mask =
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_FP32_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_BF16_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X1000_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X736_V1);
+    out->flags = IFS_SEARCH_CAP_EXACT_FLAT_SCAN |
+                 IFS_SEARCH_CAP_BATCH_ADD |
+                 IFS_SEARCH_CAP_BATCH_DELETE |
+                 IFS_SEARCH_CAP_RESERVE |
+                 IFS_SEARCH_CAP_GROUPED_PERSON_TOPK |
+                 IFS_SEARCH_CAP_DELETED_SLOT_REUSE |
+                 IFS_SEARCH_CAP_GROUPED_HOST_REFERENCE;
+    out->device = -1;
+    out->compute_capability_major = -1;
+    out->compute_capability_minor = -1;
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_create(
+    const ifs_search_create_options_t *options,
+    ifs_search_index_t *out_index) {
+    clear_error();
+    if (!out_index) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_index must not be null");
+    }
+    *out_index = nullptr;
+    if (!options || options->struct_size < sizeof(*options)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "create options struct_size is too small");
+    }
+    if (options->profile == IFS_SEARCH_PROFILE_FP16_V1) {
+        return fail(IFS_SEARCH_UNSUPPORTED,
+                    "fp16_v1 is GPU-only and is not supported by the CPU backend");
+    }
+    if (!valid_profile(options->profile)) {
+        return fail(IFS_SEARCH_UNSUPPORTED, "unsupported CPU search profile");
+    }
+    if (options->device != -1 && options->device != 0) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "the CPU backend accepts device -1 or 0 only");
+    }
+    if (options->topk_mode == IFS_SEARCH_TOPK_DEVICE) {
+        return fail(IFS_SEARCH_UNSUPPORTED,
+                    "device Top-K is not supported by the CPU backend");
+    }
+    if (options->topk_mode > IFS_SEARCH_TOPK_DEVICE) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "invalid Top-K mode");
+    }
+    if (options->max_rows && options->reserve_rows > options->max_rows) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "reserve_rows must not exceed max_rows");
+    }
+    const double growth = options->growth_factor == 0.0
+        ? kDefaultGrowthFactor : options->growth_factor;
+    if (!(growth >= 1.1 && growth <= 4.0)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "growth_factor must be between 1.1 and 4.0");
+    }
+    if (options->reserve_rows > std::numeric_limits<size_t>::max()) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "reserve_rows exceeds the process address space");
+    }
+
+    try {
+        std::unique_ptr<CpuIndex> index(new CpuIndex());
+        index->profile = static_cast<ifs_search_profile_t>(options->profile);
+        index->max_rows = options->max_rows;
+        index->growth_factor = growth;
+        const ifs_cpu_status_t status = ifs_cpu_index_create(
+            storage_for(options->profile),
+            static_cast<size_t>(options->reserve_rows),
+            int8_scale_for(options->profile), &index->core);
+        if (status != IFS_CPU_OK) {
+            return legacy_failure(status, "CPU index creation failed");
+        }
+        *out_index = index.release();
+        return IFS_SEARCH_OK;
+    } catch (const std::bad_alloc &) {
+        return fail(IFS_SEARCH_OUT_OF_MEMORY, "CPU index allocation failed");
+    } catch (const std::exception &error) {
+        return fail(IFS_SEARCH_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return fail(IFS_SEARCH_INTERNAL_ERROR, "unknown CPU creation failure");
+    }
+}
+
+void ifs_search_destroy(ifs_search_index_t index) {
+    clear_error();
+    delete as_index(index);
+}
+
+ifs_search_status_t ifs_search_reserve(ifs_search_index_t opaque,
+                                       uint64_t rows) {
+    clear_error();
+    if (!opaque) return fail(IFS_SEARCH_INVALID_ARGUMENT, "index is null");
+    CpuIndex *index = as_index(opaque);
+    std::unique_lock<std::shared_mutex> guard(index->state_mutex);
+    return reserve_locked(index, rows, false);
+}
+
+ifs_search_status_t ifs_search_add_batch(
+    ifs_search_index_t opaque, const uint64_t *ids,
+    const uint64_t *group_ids, const float *vectors, uint64_t count) {
+    clear_error();
+    if (!opaque || (count && (!ids || !group_ids || !vectors))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index, vector IDs, group IDs, and vectors are required for a non-empty add");
+    }
+    if (count > std::numeric_limits<size_t>::max()) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "batch count exceeds the process address space");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(vectors, count, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+    CpuIndex *index = as_index(opaque);
+    std::unique_lock<std::shared_mutex> guard(index->state_mutex);
+    try {
+        index->vector_to_group.reserve(
+            index->vector_to_group.size() + static_cast<size_t>(count));
+        std::unordered_set<uint64_t> batch_ids;
+        batch_ids.reserve(static_cast<size_t>(count));
+        for (uint64_t offset = 0; offset < count; ++offset) {
+            if (index->vector_to_group.find(ids[offset]) !=
+                index->vector_to_group.end()) {
+                return fail(IFS_SEARCH_DUPLICATE_ID,
+                            "duplicate active vector ID in batch add");
+            }
+            if (!batch_ids.insert(ids[offset]).second) {
+                return fail(IFS_SEARCH_DUPLICATE_ID,
+                            "duplicate vector ID within batch add");
+            }
+        }
+    } catch (const std::bad_alloc &) {
+        return fail(IFS_SEARCH_OUT_OF_MEMORY,
+                    "unable to reserve vector-to-group mapping");
+    }
+    ifs_search_status_t capacity_status = ensure_add_capacity(index, count);
+    if (capacity_status != IFS_SEARCH_OK) return capacity_status;
+    try {
+        for (uint64_t offset = 0; offset < count; ++offset) {
+            index->vector_to_group.emplace(ids[offset], group_ids[offset]);
+        }
+    } catch (const std::bad_alloc &) {
+        for (uint64_t offset = 0; offset < count; ++offset) {
+            index->vector_to_group.erase(ids[offset]);
+        }
+        return fail(IFS_SEARCH_OUT_OF_MEMORY,
+                    "unable to populate vector-to-group mapping");
+    } catch (...) {
+        for (uint64_t offset = 0; offset < count; ++offset) {
+            index->vector_to_group.erase(ids[offset]);
+        }
+        return fail(IFS_SEARCH_INTERNAL_ERROR,
+                    "unable to populate vector-to-group mapping");
+    }
+    const ifs_cpu_status_t status = ifs_cpu_index_add_batch(
+        index->core, ids, vectors, static_cast<size_t>(count));
+    if (status != IFS_CPU_OK) {
+        for (uint64_t offset = 0; offset < count; ++offset) {
+            index->vector_to_group.erase(ids[offset]);
+        }
+        return legacy_failure(status, "CPU batch add failed");
+    }
+    const uint64_t reused = std::min(count, index->tombstone_rows);
+    index->tombstone_rows -= reused;
+    index->physical_rows += count - reused;
+    index->live_rows += count;
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_delete_batch(
+    ifs_search_index_t opaque, const uint64_t *ids, uint64_t count,
+    uint64_t *out_removed) {
+    clear_error();
+    if (!out_removed) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_removed must not be null");
+    }
+    *out_removed = 0;
+    if (!opaque || (count && !ids)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index and IDs are required for a non-empty delete");
+    }
+    CpuIndex *index = as_index(opaque);
+    std::unique_lock<std::shared_mutex> guard(index->state_mutex);
+    uint64_t removed = 0;
+    for (uint64_t offset = 0; offset < count; ++offset) {
+        const ifs_cpu_status_t status = ifs_cpu_index_delete(index->core, ids[offset]);
+        if (status == IFS_CPU_OK) {
+            index->vector_to_group.erase(ids[offset]);
+            ++removed;
+        } else if (status != IFS_CPU_ID_NOT_FOUND) {
+            return legacy_failure(status, "CPU batch delete failed");
+        }
+    }
+    index->live_rows -= removed;
+    index->tombstone_rows += removed;
+    *out_removed = removed;
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_topk(
+    ifs_search_index_t opaque, const float *query, uint64_t top_k,
+    uint64_t *out_ids, float *out_scores, uint64_t *out_count,
+    ifs_search_timings_t *timings) {
+    clear_error();
+    if (!out_count) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_count must not be null");
+    }
+    *out_count = 0;
+    if (timings && timings->struct_size < sizeof(*timings)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "timings struct_size is too small");
+    }
+    zero_timings(timings);
+    if (!opaque || !query || (top_k && (!out_ids || !out_scores))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "invalid index, query, or Top-K output buffers");
+    }
+    if (top_k > std::numeric_limits<size_t>::max()) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "top_k exceeds the process address space");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(query, 1, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+    CpuIndex *index = as_index(opaque);
+    size_t count = 0;
+    const auto begin = std::chrono::steady_clock::now();
+    const ifs_cpu_status_t status = ifs_cpu_index_search(
+        index->core, query, static_cast<size_t>(top_k), out_ids, out_scores, &count);
+    const auto end = std::chrono::steady_clock::now();
+    if (status != IFS_CPU_OK) return legacy_failure(status, "CPU search failed");
+    if (int8_scale_for(index->profile) != 0) {
+        const float divisor = int8_score_divisor(index->profile);
+        for (size_t offset = 0; offset < count; ++offset) {
+            out_scores[offset] /= divisor;
+        }
+    }
+    for (size_t offset = 0; offset < count; ++offset) {
+        out_scores[offset] = public_cosine(out_scores[offset]);
+    }
+    *out_count = static_cast<uint64_t>(count);
+    if (timings) {
+        timings->total_ms = std::chrono::duration<double, std::milli>(
+            end - begin).count();
+        timings->kernel_ms = timings->total_ms;
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_grouped_topk(
+    ifs_search_index_t opaque, const float *query, uint64_t top_k,
+    uint64_t *out_group_ids, uint64_t *out_vector_ids, float *out_scores,
+    uint64_t *out_count, ifs_search_timings_t *timings) {
+    clear_error();
+    if (!out_count) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_count must not be null");
+    }
+    *out_count = 0;
+    if (timings && timings->struct_size < sizeof(*timings)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "timings struct_size is too small");
+    }
+    zero_timings(timings);
+    if (!opaque || !query ||
+        (top_k && (!out_group_ids || !out_vector_ids || !out_scores))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "invalid grouped Top-K input or output buffers");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(query, 1, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+
+    struct GroupCandidate {
+        uint64_t group_id;
+        uint64_t vector_id;
+        float score;
+    };
+    const auto better = [](const GroupCandidate &left,
+                           const GroupCandidate &right) {
+        if (left.score != right.score) return left.score > right.score;
+        return left.group_id < right.group_id;
+    };
+
+    CpuIndex *index = as_index(opaque);
+    std::shared_lock<std::shared_mutex> guard(index->state_mutex);
+    const auto begin = std::chrono::steady_clock::now();
+    try {
+        const size_t live = static_cast<size_t>(index->live_rows);
+        std::vector<uint64_t> face_ids(live);
+        std::vector<float> face_scores(live);
+        size_t face_count = 0;
+        const ifs_cpu_status_t status = ifs_cpu_index_search(
+            index->core, query, live, face_ids.data(), face_scores.data(),
+            &face_count);
+        if (status != IFS_CPU_OK) {
+            return legacy_failure(status, "CPU grouped reference search failed");
+        }
+        if (int8_scale_for(index->profile) != 0) {
+            const float divisor = int8_score_divisor(index->profile);
+            for (size_t offset = 0; offset < face_count; ++offset) {
+                face_scores[offset] /= divisor;
+            }
+        }
+
+        std::unordered_map<uint64_t, GroupCandidate> best_by_group;
+        best_by_group.reserve(index->vector_to_group.size());
+        for (size_t offset = 0; offset < face_count; ++offset) {
+            const auto group = index->vector_to_group.find(face_ids[offset]);
+            if (group == index->vector_to_group.end()) {
+                return fail(IFS_SEARCH_INTERNAL_ERROR,
+                            "active vector is missing its group mapping");
+            }
+            /* face results are score descending, vector ID ascending, so the
+             * first face observed for a group is its deterministic maximum. */
+            best_by_group.emplace(
+                group->second,
+                GroupCandidate{group->second, face_ids[offset],
+                               face_scores[offset]});
+        }
+        std::vector<GroupCandidate> groups;
+        groups.reserve(best_by_group.size());
+        for (const auto &entry : best_by_group) groups.push_back(entry.second);
+        std::sort(groups.begin(), groups.end(), better);
+        const size_t result_count = std::min<size_t>(
+            static_cast<size_t>(std::min<uint64_t>(
+                top_k, std::numeric_limits<size_t>::max())), groups.size());
+        for (size_t offset = 0; offset < result_count; ++offset) {
+            out_group_ids[offset] = groups[offset].group_id;
+            out_vector_ids[offset] = groups[offset].vector_id;
+            out_scores[offset] = public_cosine(groups[offset].score);
+        }
+        *out_count = static_cast<uint64_t>(result_count);
+    } catch (const std::bad_alloc &) {
+        return fail(IFS_SEARCH_OUT_OF_MEMORY,
+                    "CPU grouped reference allocation failed");
+    } catch (const std::exception &error) {
+        return fail(IFS_SEARCH_INTERNAL_ERROR, error.what());
+    }
+    const auto end = std::chrono::steady_clock::now();
+    if (timings) {
+        timings->total_ms = std::chrono::duration<double, std::milli>(
+            end - begin).count();
+        timings->kernel_ms = timings->total_ms;
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_get_stats(
+    ifs_search_index_t opaque, ifs_search_stats_t *out) {
+    clear_error();
+    if (!opaque || !out || out->struct_size < sizeof(*out)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index and a complete stats structure are required");
+    }
+    CpuIndex *index = as_index(opaque);
+    std::shared_lock<std::shared_mutex> guard(index->state_mutex);
+    const uint32_t requested_size = out->struct_size;
+    *out = {};
+    out->struct_size = requested_size;
+    out->backend = IFS_SEARCH_BACKEND_CPU;
+    out->profile = index->profile;
+    out->device = -1;
+    out->physical_rows = index->physical_rows;
+    out->live_rows = index->live_rows;
+    out->capacity_rows = static_cast<uint64_t>(
+        ifs_cpu_index_capacity(index->core));
+    out->max_rows = index->max_rows;
+    out->tombstone_rows = index->tombstone_rows;
+    out->reallocations = index->reallocations;
+    out->bytes_per_vector = bytes_per_vector(index->profile);
+    return IFS_SEARCH_OK;
+}
+
+}  // extern "C"
diff --git a/server/native/search/cpu/src/ifs_cpu_legacy.h b/server/native/search/cpu/src/ifs_cpu_legacy.h
new file mode 100644
index 000000000..7a67d167f
--- /dev/null
+++ b/server/native/search/cpu/src/ifs_cpu_legacy.h
@@ -0,0 +1,92 @@
+#ifndef IFS_CPU_H
+#define IFS_CPU_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+/* Internal compatibility API imported from the standalone benchmark. The
+ * product library exports only ifs_search.h ABI v2. */
+#define IFS_CPU_API
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define IFS_CPU_DIMENSION 512u
+
+typedef void *ifs_cpu_index_t;
+
+typedef enum ifs_cpu_storage {
+    IFS_CPU_STORAGE_FP32 = 0,
+    IFS_CPU_STORAGE_BF16 = 1,
+    IFS_CPU_STORAGE_INT8 = 2
+} ifs_cpu_storage_t;
+
+typedef enum ifs_cpu_status {
+    IFS_CPU_OK = 0,
+    IFS_CPU_INVALID_ARGUMENT = 1,
+    IFS_CPU_OUT_OF_MEMORY = 2,
+    IFS_CPU_DUPLICATE_ID = 3,
+    IFS_CPU_ID_NOT_FOUND = 4,
+    IFS_CPU_INTERNAL_ERROR = 5
+} ifs_cpu_status_t;
+
+/* Error text is thread-local and remains valid until the next API call on the
+ * same thread. Successful calls clear it. */
+IFS_CPU_API const char *ifs_cpu_last_error(void);
+IFS_CPU_API const char *ifs_cpu_version(void);
+IFS_CPU_API uint32_t ifs_cpu_dimension(void);
+
+/* Human-readable process-wide dispatch diagnostics. */
+IFS_CPU_API const char *ifs_cpu_runtime_features(void);
+IFS_CPU_API const char *ifs_cpu_kernel_name(ifs_cpu_storage_t storage);
+
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_create(
+    ifs_cpu_storage_t storage,
+    size_t initial_capacity,
+    uint32_t int8_scale,
+    ifs_cpu_index_t *out_index);
+
+IFS_CPU_API void ifs_cpu_index_destroy(ifs_cpu_index_t index);
+
+IFS_CPU_API size_t ifs_cpu_index_size(ifs_cpu_index_t index);
+IFS_CPU_API size_t ifs_cpu_index_capacity(ifs_cpu_index_t index);
+IFS_CPU_API ifs_cpu_storage_t ifs_cpu_index_storage(ifs_cpu_index_t index);
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_reserve(
+    ifs_cpu_index_t index,
+    size_t rows);
+
+/* Vectors must contain exactly IFS_CPU_DIMENSION finite FP32 values. IDs are
+ * unique among active entries; a deleted ID may be added again. */
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_add(
+    ifs_cpu_index_t index,
+    uint64_t id,
+    const float *vector512);
+
+/* Row-major vectors[count][512]. This performs one validation and write lock
+ * for the complete batch and is the intended ingestion path for large bases. */
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_add_batch(
+    ifs_cpu_index_t index,
+    const uint64_t *ids,
+    const float *vectors,
+    size_t count);
+
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_delete(
+    ifs_cpu_index_t index,
+    uint64_t id);
+
+/* Results are ordered by descending score, then ascending ID. For INT8,
+ * out_scores contains the exact INT32 accumulator converted to float. */
+IFS_CPU_API ifs_cpu_status_t ifs_cpu_index_search(
+    ifs_cpu_index_t index,
+    const float *query512,
+    size_t top_k,
+    uint64_t *out_ids,
+    float *out_scores,
+    size_t *out_count);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/server/native/search/cpu/src/index_legacy.cpp b/server/native/search/cpu/src/index_legacy.cpp
new file mode 100644
index 000000000..e480d239c
--- /dev/null
+++ b/server/native/search/cpu/src/index_legacy.cpp
@@ -0,0 +1,690 @@
+#include "ifs_cpu_legacy.h"
+
+#include "kernels.hpp"
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <queue>
+#include <shared_mutex>
+#include <stdexcept>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+namespace ifs_cpu {
+namespace {
+
+class Index {
+public:
+    explicit Index(ifs_cpu_storage_t storage, std::size_t initial_capacity,
+                   std::uint32_t int8_scale)
+        : storage_(storage), int8_scale_(int8_scale) {
+        reserve_capacity(initial_capacity);
+    }
+
+    bool add(std::uint64_t id, const float *input) {
+        std::array<float, kDimension> fp32{};
+        std::array<std::uint16_t, kDimension> bf16{};
+        std::array<std::int8_t, kDimension> int8{};
+        for (std::size_t offset = 0; offset < kDimension; ++offset) {
+            if (!std::isfinite(input[offset])) {
+                throw std::invalid_argument("vector contains a non-finite value");
+            }
+            switch (storage_) {
+                case IFS_CPU_STORAGE_FP32:
+                    fp32[offset] = input[offset];
+                    break;
+                case IFS_CPU_STORAGE_BF16:
+                    bf16[offset] = float_to_bf16(input[offset]);
+                    break;
+                case IFS_CPU_STORAGE_INT8:
+                    int8[offset] = quantize_int8(input[offset], int8_scale_);
+                    break;
+            }
+        }
+
+        std::unique_lock<std::shared_mutex> lock(mutex_);
+        if (id_to_slot_.find(id) != id_to_slot_.end()) {
+            return false;
+        }
+
+        std::size_t slot = 0;
+        if (!free_slots_.empty()) {
+            slot = free_slots_.back();
+            free_slots_.pop_back();
+            ids_[slot] = id;
+            active_[slot] = 1;
+            write_slot(slot, fp32, bf16, int8);
+        } else {
+            ensure_capacity(ids_.size() + 1);
+            slot = ids_.size();
+            ids_.push_back(id);
+            active_.push_back(1);
+            append_slot(fp32, bf16, int8);
+        }
+
+        try {
+            id_to_slot_.emplace(id, slot);
+        } catch (...) {
+            active_[slot] = 0;
+            free_slots_.push_back(slot);
+            throw;
+        }
+        ++active_count_;
+        return true;
+    }
+
+    bool add_batch(const std::uint64_t *ids, const float *inputs,
+                   std::size_t count) {
+        if (count == 0) {
+            return true;
+        }
+        std::unordered_map<std::uint64_t, std::uint8_t> batch_ids;
+        batch_ids.reserve(count);
+        for (std::size_t row = 0; row < count; ++row) {
+            if (!batch_ids.emplace(ids[row], 1).second) {
+                return false;
+            }
+            const float *input = inputs + row * kDimension;
+            for (std::size_t offset = 0; offset < kDimension; ++offset) {
+                if (!std::isfinite(input[offset])) {
+                    throw std::invalid_argument("vector batch contains a non-finite value");
+                }
+            }
+        }
+
+        std::unique_lock<std::shared_mutex> lock(mutex_);
+        for (std::size_t row = 0; row < count; ++row) {
+            if (id_to_slot_.find(ids[row]) != id_to_slot_.end()) {
+                return false;
+            }
+        }
+        const std::size_t reusable = std::min(count, free_slots_.size());
+        if (count - reusable > std::numeric_limits<std::size_t>::max() - ids_.size()) {
+            throw std::length_error("batch row count overflows index size");
+        }
+        ensure_capacity(ids_.size() + count - reusable);
+        id_to_slot_.reserve(id_to_slot_.size() + count);
+
+        for (std::size_t row = 0; row < count; ++row) {
+            std::array<float, kDimension> fp32{};
+            std::array<std::uint16_t, kDimension> bf16{};
+            std::array<std::int8_t, kDimension> int8{};
+            const float *input = inputs + row * kDimension;
+            for (std::size_t offset = 0; offset < kDimension; ++offset) {
+                switch (storage_) {
+                    case IFS_CPU_STORAGE_FP32:
+                        fp32[offset] = input[offset];
+                        break;
+                    case IFS_CPU_STORAGE_BF16:
+                        bf16[offset] = float_to_bf16(input[offset]);
+                        break;
+                    case IFS_CPU_STORAGE_INT8:
+                        int8[offset] = quantize_int8(input[offset], int8_scale_);
+                        break;
+                }
+            }
+            std::size_t slot = 0;
+            if (!free_slots_.empty()) {
+                slot = free_slots_.back();
+                free_slots_.pop_back();
+                ids_[slot] = ids[row];
+                active_[slot] = 1;
+                write_slot(slot, fp32, bf16, int8);
+            } else {
+                slot = ids_.size();
+                ids_.push_back(ids[row]);
+                active_.push_back(1);
+                append_slot(fp32, bf16, int8);
+            }
+            id_to_slot_.emplace(ids[row], slot);
+            ++active_count_;
+        }
+        return true;
+    }
+
+    bool erase(std::uint64_t id) {
+        std::unique_lock<std::shared_mutex> lock(mutex_);
+        const auto found = id_to_slot_.find(id);
+        if (found == id_to_slot_.end()) {
+            return false;
+        }
+        const std::size_t slot = found->second;
+        id_to_slot_.erase(found);
+        active_[slot] = 0;
+        free_slots_.push_back(slot);
+        --active_count_;
+        return true;
+    }
+
+    std::vector<std::pair<std::uint64_t, float>> search(
+        const float *query,
+        std::size_t top_k) const {
+        std::array<float, kDimension> fp32{};
+        std::array<std::uint16_t, kDimension> bf16{};
+        std::array<std::int8_t, kDimension> int8{};
+        std::int32_t query_sum = 0;
+        for (std::size_t offset = 0; offset < kDimension; ++offset) {
+            if (!std::isfinite(query[offset])) {
+                throw std::invalid_argument("query contains a non-finite value");
+            }
+            switch (storage_) {
+                case IFS_CPU_STORAGE_FP32:
+                    fp32[offset] = query[offset];
+                    break;
+                case IFS_CPU_STORAGE_BF16:
+                    bf16[offset] = float_to_bf16(query[offset]);
+                    break;
+                case IFS_CPU_STORAGE_INT8:
+                    int8[offset] = quantize_int8(query[offset], int8_scale_);
+                    query_sum += static_cast<std::int32_t>(int8[offset]);
+                    break;
+            }
+        }
+
+        if (top_k == 0) {
+            return {};
+        }
+
+        struct Candidate {
+            std::uint64_t id;
+            float score;
+        };
+        const auto better = [](const Candidate &left, const Candidate &right) {
+            return left.score > right.score ||
+                   (left.score == right.score && left.id < right.id);
+        };
+        struct BetterComparator {
+            bool operator()(const Candidate &left, const Candidate &right) const {
+                return left.score > right.score ||
+                       (left.score == right.score && left.id < right.id);
+            }
+        };
+
+        std::shared_lock<std::shared_mutex> lock(mutex_);
+        auto score_slot = [&](std::size_t slot) {
+            if (active_[slot] == 0) {
+                return Candidate{0, -std::numeric_limits<float>::infinity()};
+            }
+            const std::size_t base = slot * kDimension;
+            float score = 0.0f;
+            switch (storage_) {
+                case IFS_CPU_STORAGE_FP32:
+                    score = dispatch().fp32(fp32_.data() + base, fp32.data());
+                    break;
+                case IFS_CPU_STORAGE_BF16:
+                    score = dispatch().bf16(bf16_.data() + base, bf16.data());
+                    break;
+                case IFS_CPU_STORAGE_INT8:
+                    score = static_cast<float>(dispatch().int8(
+                        int8_.data() + base, int8.data(), query_sum));
+                    break;
+            }
+            return Candidate{ids_[slot], score};
+        };
+        auto consider = [&](auto &heap, const Candidate &candidate) {
+            if (!std::isfinite(candidate.score)) return;
+            if (heap.size() < top_k) heap.push(candidate);
+            else if (better(candidate, heap.top())) {
+                heap.pop();
+                heap.push(candidate);
+            }
+        };
+
+        std::priority_queue<Candidate, std::vector<Candidate>, BetterComparator> heap;
+#ifdef _OPENMP
+        const int thread_count = std::max(1, omp_get_max_threads());
+        std::vector<std::vector<Candidate>> partials(
+            static_cast<std::size_t>(thread_count));
+#pragma omp parallel
+        {
+            std::priority_queue<Candidate, std::vector<Candidate>, BetterComparator>
+                local_heap;
+#pragma omp for schedule(static)
+            for (std::int64_t signed_slot = 0;
+                 signed_slot < static_cast<std::int64_t>(ids_.size()); ++signed_slot) {
+                consider(local_heap, score_slot(static_cast<std::size_t>(signed_slot)));
+            }
+            auto &partial = partials[static_cast<std::size_t>(omp_get_thread_num())];
+            partial.reserve(local_heap.size());
+            while (!local_heap.empty()) {
+                partial.push_back(local_heap.top());
+                local_heap.pop();
+            }
+        }
+        for (const auto &partial : partials) {
+            for (const Candidate &candidate : partial) consider(heap, candidate);
+        }
+#else
+        for (std::size_t slot = 0; slot < ids_.size(); ++slot) {
+            consider(heap, score_slot(slot));
+        }
+#endif
+
+        std::vector<Candidate> candidates;
+        candidates.reserve(heap.size());
+        while (!heap.empty()) {
+            candidates.push_back(heap.top());
+            heap.pop();
+        }
+        std::sort(candidates.begin(), candidates.end(), better);
+
+        std::vector<std::pair<std::uint64_t, float>> results;
+        results.reserve(candidates.size());
+        for (const Candidate &candidate : candidates) {
+            results.emplace_back(candidate.id, candidate.score);
+        }
+        return results;
+    }
+
+    std::size_t size() const {
+        std::shared_lock<std::shared_mutex> lock(mutex_);
+        return active_count_;
+    }
+
+    std::size_t capacity() const {
+        std::shared_lock<std::shared_mutex> lock(mutex_);
+        return capacity_;
+    }
+
+    ifs_cpu_storage_t storage() const noexcept {
+        return storage_;
+    }
+
+    void reserve(std::size_t rows) {
+        std::unique_lock<std::shared_mutex> lock(mutex_);
+        reserve_capacity(rows);
+    }
+
+private:
+    void ensure_capacity(std::size_t required) {
+        if (required <= capacity_) {
+            return;
+        }
+        std::size_t grown = capacity_ == 0 ? 1 : capacity_;
+        while (grown < required) {
+            if (grown > std::numeric_limits<std::size_t>::max() / 2) {
+                grown = required;
+                break;
+            }
+            grown *= 2;
+        }
+        reserve_capacity(grown);
+    }
+
+    void reserve_capacity(std::size_t requested) {
+        if (requested < capacity_) {
+            return;
+        }
+        if (requested > std::numeric_limits<std::size_t>::max() / kDimension) {
+            throw std::length_error("capacity exceeds addressable storage");
+        }
+        const std::size_t elements = requested * kDimension;
+        ids_.reserve(requested);
+        active_.reserve(requested);
+        free_slots_.reserve(requested);
+        id_to_slot_.reserve(requested);
+        switch (storage_) {
+            case IFS_CPU_STORAGE_FP32:
+                fp32_.reserve(elements);
+                break;
+            case IFS_CPU_STORAGE_BF16:
+                bf16_.reserve(elements);
+                break;
+            case IFS_CPU_STORAGE_INT8:
+                int8_.reserve(elements);
+                break;
+        }
+        capacity_ = requested;
+    }
+
+    void write_slot(
+        std::size_t slot,
+        const std::array<float, kDimension> &fp32,
+        const std::array<std::uint16_t, kDimension> &bf16,
+        const std::array<std::int8_t, kDimension> &int8) {
+        const std::size_t base = slot * kDimension;
+        switch (storage_) {
+            case IFS_CPU_STORAGE_FP32:
+                std::copy(fp32.begin(), fp32.end(), fp32_.begin() + base);
+                break;
+            case IFS_CPU_STORAGE_BF16:
+                std::copy(bf16.begin(), bf16.end(), bf16_.begin() + base);
+                break;
+            case IFS_CPU_STORAGE_INT8:
+                std::copy(int8.begin(), int8.end(), int8_.begin() + base);
+                break;
+        }
+    }
+
+    void append_slot(
+        const std::array<float, kDimension> &fp32,
+        const std::array<std::uint16_t, kDimension> &bf16,
+        const std::array<std::int8_t, kDimension> &int8) {
+        switch (storage_) {
+            case IFS_CPU_STORAGE_FP32:
+                fp32_.insert(fp32_.end(), fp32.begin(), fp32.end());
+                break;
+            case IFS_CPU_STORAGE_BF16:
+                bf16_.insert(bf16_.end(), bf16.begin(), bf16.end());
+                break;
+            case IFS_CPU_STORAGE_INT8:
+                int8_.insert(int8_.end(), int8.begin(), int8.end());
+                break;
+        }
+    }
+
+    ifs_cpu_storage_t storage_;
+    std::uint32_t int8_scale_;
+    mutable std::shared_mutex mutex_;
+    std::size_t capacity_ = 0;
+    std::size_t active_count_ = 0;
+    std::vector<std::uint64_t> ids_;
+    std::vector<std::uint8_t> active_;
+    std::vector<std::size_t> free_slots_;
+    std::unordered_map<std::uint64_t, std::size_t> id_to_slot_;
+    std::vector<float> fp32_;
+    std::vector<std::uint16_t> bf16_;
+    std::vector<std::int8_t> int8_;
+};
+
+thread_local std::string last_error;
+
+void clear_error() {
+    last_error.clear();
+}
+
+ifs_cpu_status_t fail(ifs_cpu_status_t status, const std::string &message) {
+    last_error = message;
+    return status;
+}
+
+Index *as_index(ifs_cpu_index_t index) {
+    return static_cast<Index *>(index);
+}
+
+bool valid_storage(ifs_cpu_storage_t storage) {
+    return storage == IFS_CPU_STORAGE_FP32 || storage == IFS_CPU_STORAGE_BF16 ||
+           storage == IFS_CPU_STORAGE_INT8;
+}
+
+}  // namespace
+}  // namespace ifs_cpu
+
+extern "C" {
+
+const char *ifs_cpu_last_error(void) {
+    return ifs_cpu::last_error.c_str();
+}
+
+const char *ifs_cpu_version(void) {
+    ifs_cpu::clear_error();
+    return "0.1.0";
+}
+
+std::uint32_t ifs_cpu_dimension(void) {
+    ifs_cpu::clear_error();
+    return static_cast<std::uint32_t>(ifs_cpu::kDimension);
+}
+
+const char *ifs_cpu_runtime_features(void) {
+    ifs_cpu::clear_error();
+    try {
+        return ifs_cpu::dispatch().features.c_str();
+    } catch (const std::exception &error) {
+        ifs_cpu::last_error = error.what();
+        return "";
+    } catch (...) {
+        ifs_cpu::last_error = "unknown runtime dispatch failure";
+        return "";
+    }
+}
+
+const char *ifs_cpu_kernel_name(ifs_cpu_storage_t storage) {
+    ifs_cpu::clear_error();
+    try {
+        const ifs_cpu::Dispatch &selected = ifs_cpu::dispatch();
+        switch (storage) {
+            case IFS_CPU_STORAGE_FP32:
+                return selected.fp32_name;
+            case IFS_CPU_STORAGE_BF16:
+                return selected.bf16_name;
+            case IFS_CPU_STORAGE_INT8:
+                return selected.int8_name;
+        }
+        ifs_cpu::last_error = "invalid storage mode";
+        return "invalid-storage";
+    } catch (const std::exception &error) {
+        ifs_cpu::last_error = error.what();
+        return "";
+    } catch (...) {
+        ifs_cpu::last_error = "unknown runtime dispatch failure";
+        return "";
+    }
+}
+
+ifs_cpu_status_t ifs_cpu_index_create(
+    ifs_cpu_storage_t storage,
+    std::size_t initial_capacity,
+    std::uint32_t int8_scale,
+    ifs_cpu_index_t *out_index) {
+    ifs_cpu::clear_error();
+    if (out_index == nullptr) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, "out_index must not be null");
+    }
+    *out_index = nullptr;
+    if (!ifs_cpu::valid_storage(storage)) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, "invalid storage mode");
+    }
+    if ((storage == IFS_CPU_STORAGE_INT8) != (int8_scale > 0)) {
+        return ifs_cpu::fail(
+            IFS_CPU_INVALID_ARGUMENT,
+            "INT8 storage requires a positive scale and other storage requires zero");
+    }
+    try {
+        *out_index = new ifs_cpu::Index(storage, initial_capacity, int8_scale);
+        return IFS_CPU_OK;
+    } catch (const std::bad_alloc &) {
+        return ifs_cpu::fail(IFS_CPU_OUT_OF_MEMORY, "unable to reserve index capacity");
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown index creation failure");
+    }
+}
+
+void ifs_cpu_index_destroy(ifs_cpu_index_t index) {
+    ifs_cpu::clear_error();
+    delete ifs_cpu::as_index(index);
+}
+
+std::size_t ifs_cpu_index_size(ifs_cpu_index_t index) {
+    ifs_cpu::clear_error();
+    if (index == nullptr) {
+        ifs_cpu::last_error = "index must not be null";
+        return 0;
+    }
+    try {
+        return ifs_cpu::as_index(index)->size();
+    } catch (const std::exception &error) {
+        ifs_cpu::last_error = error.what();
+    } catch (...) {
+        ifs_cpu::last_error = "unknown size failure";
+    }
+    return 0;
+}
+
+std::size_t ifs_cpu_index_capacity(ifs_cpu_index_t index) {
+    ifs_cpu::clear_error();
+    if (index == nullptr) {
+        ifs_cpu::last_error = "index must not be null";
+        return 0;
+    }
+    try {
+        return ifs_cpu::as_index(index)->capacity();
+    } catch (const std::exception &error) {
+        ifs_cpu::last_error = error.what();
+    } catch (...) {
+        ifs_cpu::last_error = "unknown capacity failure";
+    }
+    return 0;
+}
+
+ifs_cpu_storage_t ifs_cpu_index_storage(ifs_cpu_index_t index) {
+    ifs_cpu::clear_error();
+    if (index == nullptr) {
+        ifs_cpu::last_error = "index must not be null";
+        return IFS_CPU_STORAGE_FP32;
+    }
+    try {
+        return ifs_cpu::as_index(index)->storage();
+    } catch (const std::exception &error) {
+        ifs_cpu::last_error = error.what();
+    } catch (...) {
+        ifs_cpu::last_error = "unknown storage failure";
+    }
+    return IFS_CPU_STORAGE_FP32;
+}
+
+ifs_cpu_status_t ifs_cpu_index_reserve(
+    ifs_cpu_index_t index,
+    std::size_t rows) {
+    ifs_cpu::clear_error();
+    if (index == nullptr) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, "index must not be null");
+    }
+    try {
+        ifs_cpu::as_index(index)->reserve(rows);
+        return IFS_CPU_OK;
+    } catch (const std::bad_alloc &) {
+        return ifs_cpu::fail(IFS_CPU_OUT_OF_MEMORY, "unable to reserve index capacity");
+    } catch (const std::invalid_argument &error) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, error.what());
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown reserve failure");
+    }
+}
+
+ifs_cpu_status_t ifs_cpu_index_add(
+    ifs_cpu_index_t index,
+    std::uint64_t id,
+    const float *vector512) {
+    ifs_cpu::clear_error();
+    if (index == nullptr || vector512 == nullptr) {
+        return ifs_cpu::fail(
+            IFS_CPU_INVALID_ARGUMENT, "index and vector512 must not be null");
+    }
+    try {
+        if (!ifs_cpu::as_index(index)->add(id, vector512)) {
+            return ifs_cpu::fail(IFS_CPU_DUPLICATE_ID, "ID is already active");
+        }
+        return IFS_CPU_OK;
+    } catch (const std::invalid_argument &error) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, error.what());
+    } catch (const std::bad_alloc &) {
+        return ifs_cpu::fail(IFS_CPU_OUT_OF_MEMORY, "index growth failed");
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown add failure");
+    }
+}
+
+ifs_cpu_status_t ifs_cpu_index_add_batch(
+    ifs_cpu_index_t index,
+    const std::uint64_t *ids,
+    const float *vectors,
+    std::size_t count) {
+    ifs_cpu::clear_error();
+    if (index == nullptr || (count > 0 && (ids == nullptr || vectors == nullptr))) {
+        return ifs_cpu::fail(
+            IFS_CPU_INVALID_ARGUMENT,
+            "index, ids, and vectors must not be null for a non-empty batch");
+    }
+    try {
+        if (!ifs_cpu::as_index(index)->add_batch(ids, vectors, count)) {
+            return ifs_cpu::fail(IFS_CPU_DUPLICATE_ID, "duplicate active ID in batch add");
+        }
+        return IFS_CPU_OK;
+    } catch (const std::bad_alloc &) {
+        return ifs_cpu::fail(IFS_CPU_OUT_OF_MEMORY, "unable to grow index for batch");
+    } catch (const std::invalid_argument &error) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, error.what());
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown batch add failure");
+    }
+}
+
+ifs_cpu_status_t ifs_cpu_index_delete(ifs_cpu_index_t index, std::uint64_t id) {
+    ifs_cpu::clear_error();
+    if (index == nullptr) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, "index must not be null");
+    }
+    try {
+        if (!ifs_cpu::as_index(index)->erase(id)) {
+            return ifs_cpu::fail(IFS_CPU_ID_NOT_FOUND, "ID is not active");
+        }
+        return IFS_CPU_OK;
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown delete failure");
+    }
+}
+
+ifs_cpu_status_t ifs_cpu_index_search(
+    ifs_cpu_index_t index,
+    const float *query512,
+    std::size_t top_k,
+    std::uint64_t *out_ids,
+    float *out_scores,
+    std::size_t *out_count) {
+    ifs_cpu::clear_error();
+    if (out_count == nullptr) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, "out_count must not be null");
+    }
+    *out_count = 0;
+    if (index == nullptr || query512 == nullptr) {
+        return ifs_cpu::fail(
+            IFS_CPU_INVALID_ARGUMENT, "index and query512 must not be null");
+    }
+    if (top_k > 0 && (out_ids == nullptr || out_scores == nullptr)) {
+        return ifs_cpu::fail(
+            IFS_CPU_INVALID_ARGUMENT,
+            "out_ids and out_scores must not be null when top_k is positive");
+    }
+    try {
+        const auto results = ifs_cpu::as_index(index)->search(query512, top_k);
+        for (std::size_t offset = 0; offset < results.size(); ++offset) {
+            out_ids[offset] = results[offset].first;
+            out_scores[offset] = results[offset].second;
+        }
+        *out_count = results.size();
+        return IFS_CPU_OK;
+    } catch (const std::invalid_argument &error) {
+        return ifs_cpu::fail(IFS_CPU_INVALID_ARGUMENT, error.what());
+    } catch (const std::bad_alloc &) {
+        return ifs_cpu::fail(IFS_CPU_OUT_OF_MEMORY, "search allocation failed");
+    } catch (const std::exception &error) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, error.what());
+    } catch (...) {
+        return ifs_cpu::fail(IFS_CPU_INTERNAL_ERROR, "unknown search failure");
+    }
+}
+
+}  // extern "C"
diff --git a/server/native/search/cpu/src/kernels.hpp b/server/native/search/cpu/src/kernels.hpp
new file mode 100644
index 000000000..81f93b88f
--- /dev/null
+++ b/server/native/search/cpu/src/kernels.hpp
@@ -0,0 +1,39 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+namespace ifs_cpu {
+
+constexpr std::size_t kDimension = 512;
+
+using DotF32 = float (*)(const float *, const float *);
+using DotBF16 = float (*)(const std::uint16_t *, const std::uint16_t *);
+using DotInt8 = std::int32_t (*)(const std::int8_t *, const std::int8_t *, std::int32_t);
+
+struct Dispatch {
+    DotF32 fp32;
+    DotBF16 bf16;
+    DotInt8 int8;
+    const char *fp32_name;
+    const char *bf16_name;
+    const char *int8_name;
+    std::string features;
+};
+
+float dot_f32_portable(const float *left, const float *right);
+float dot_bf16_portable(const std::uint16_t *left, const std::uint16_t *right);
+std::int32_t dot_int8_portable(
+    const std::int8_t *left,
+    const std::int8_t *right,
+    std::int32_t query_sum);
+
+std::uint16_t float_to_bf16(float value);
+float bf16_to_float(std::uint16_t value);
+std::int8_t quantize_int8(float value, std::uint32_t scale);
+
+Dispatch make_dispatch();
+const Dispatch &dispatch();
+
+}  // namespace ifs_cpu
diff --git a/server/native/search/cpu/src/kernels_portable.cpp b/server/native/search/cpu/src/kernels_portable.cpp
new file mode 100644
index 000000000..7d4711d8a
--- /dev/null
+++ b/server/native/search/cpu/src/kernels_portable.cpp
@@ -0,0 +1,71 @@
+#include "kernels.hpp"
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+
+namespace ifs_cpu {
+
+std::uint16_t float_to_bf16(float value) {
+    std::uint32_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    const std::uint32_t rounding_bias = 0x7fffu + ((bits >> 16u) & 1u);
+    bits += rounding_bias;
+    return static_cast<std::uint16_t>(bits >> 16u);
+}
+
+float bf16_to_float(std::uint16_t value) {
+    const std::uint32_t bits = static_cast<std::uint32_t>(value) << 16u;
+    float result = 0.0f;
+    std::memcpy(&result, &bits, sizeof(result));
+    return result;
+}
+
+std::int8_t quantize_int8(float value, std::uint32_t scale) {
+    // Keep the multiplication in FP32 so the CPU encoder has the same
+    // half-integer behavior as the CUDA kernel and NumPy reference backend.
+    const float scaled = value * static_cast<float>(scale);
+    if (scaled >= 127.0f) {
+        return 127;
+    }
+    if (scaled <= -128.0f) {
+        return -128;
+    }
+    const long rounded = std::lroundf(scaled);
+    return static_cast<std::int8_t>(std::clamp<long>(rounded, -128, 127));
+}
+
+float dot_f32_portable(const float *left, const float *right) {
+    float accumulator = 0.0f;
+    for (std::size_t index = 0; index < kDimension; ++index) {
+        accumulator += left[index] * right[index];
+    }
+    return accumulator;
+}
+
+float dot_bf16_portable(const std::uint16_t *left, const std::uint16_t *right) {
+    float accumulator = 0.0f;
+    for (std::size_t index = 0; index < kDimension; ++index) {
+        accumulator += bf16_to_float(left[index]) * bf16_to_float(right[index]);
+    }
+    return accumulator;
+}
+
+std::int32_t dot_int8_portable(
+    const std::int8_t *left,
+    const std::int8_t *right,
+    std::int32_t /*query_sum*/) {
+    std::int32_t accumulator = 0;
+    for (std::size_t index = 0; index < kDimension; ++index) {
+        accumulator += static_cast<std::int32_t>(left[index]) *
+                       static_cast<std::int32_t>(right[index]);
+    }
+    return accumulator;
+}
+
+const Dispatch &dispatch() {
+    static const Dispatch selected = make_dispatch();
+    return selected;
+}
+
+}  // namespace ifs_cpu
diff --git a/server/native/search/cpu/src/kernels_x86.cpp b/server/native/search/cpu/src/kernels_x86.cpp
new file mode 100644
index 000000000..eacb0af70
--- /dev/null
+++ b/server/native/search/cpu/src/kernels_x86.cpp
@@ -0,0 +1,310 @@
+#include "kernels.hpp"
+
+#include <array>
+#include <cstdint>
+#include <sstream>
+#include <string>
+
+#if defined(__x86_64__) || defined(_M_X64)
+#  define IFS_CPU_X86_64 1
+#else
+#  define IFS_CPU_X86_64 0
+#endif
+
+#if IFS_CPU_X86_64 && (defined(__GNUC__) || defined(__clang__))
+#  include <cpuid.h>
+#  include <immintrin.h>
+#  define IFS_CPU_X86_TARGETS 1
+#  define IFS_TARGET(value) __attribute__((target(value), noinline))
+#else
+#  define IFS_CPU_X86_TARGETS 0
+#endif
+
+namespace ifs_cpu {
+namespace {
+
+struct CpuFeatures {
+    bool x86_64 = false;
+    bool os_avx = false;
+    bool os_avx512 = false;
+    bool fma = false;
+    bool avx2 = false;
+    bool avx512f = false;
+    bool avx_vnni = false;
+    bool avx512_vnni = false;
+    bool avx512_bf16 = false;
+};
+
+#if IFS_CPU_X86_TARGETS
+
+std::uint64_t read_xcr0() {
+    std::uint32_t low = 0;
+    std::uint32_t high = 0;
+    __asm__ volatile("xgetbv" : "=a"(low), "=d"(high) : "c"(0));
+    return (static_cast<std::uint64_t>(high) << 32u) | low;
+}
+
+CpuFeatures detect_features() {
+    CpuFeatures features;
+    features.x86_64 = true;
+    const unsigned int max_leaf = __get_cpuid_max(0, nullptr);
+    if (max_leaf < 1) {
+        return features;
+    }
+
+    unsigned int eax = 0;
+    unsigned int ebx = 0;
+    unsigned int ecx = 0;
+    unsigned int edx = 0;
+    __cpuid_count(1, 0, eax, ebx, ecx, edx);
+    const bool cpu_avx = (ecx & bit_AVX) != 0;
+    const bool osxsave = (ecx & bit_OSXSAVE) != 0;
+    features.fma = (ecx & bit_FMA) != 0;
+    if (cpu_avx && osxsave) {
+        const std::uint64_t xcr0 = read_xcr0();
+        features.os_avx = (xcr0 & 0x6u) == 0x6u;
+        features.os_avx512 = features.os_avx && (xcr0 & 0xe6u) == 0xe6u;
+    }
+
+    if (max_leaf < 7) {
+        return features;
+    }
+    __cpuid_count(7, 0, eax, ebx, ecx, edx);
+    const unsigned int max_subleaf = eax;
+    features.avx2 = features.os_avx && (ebx & bit_AVX2) != 0;
+    features.avx512f = features.os_avx512 && (ebx & bit_AVX512F) != 0;
+    features.avx512_vnni = features.avx512f && (ecx & (1u << 11u)) != 0;
+
+    if (max_subleaf >= 1) {
+        __cpuid_count(7, 1, eax, ebx, ecx, edx);
+        features.avx_vnni = features.avx2 && (eax & (1u << 4u)) != 0;
+        features.avx512_bf16 = features.avx512f && (eax & (1u << 5u)) != 0;
+    }
+    return features;
+}
+
+IFS_TARGET("avx2,fma")
+float dot_f32_avx2_fma(const float *left, const float *right) {
+    __m256 accumulator = _mm256_setzero_ps();
+    for (std::size_t index = 0; index < kDimension; index += 8) {
+        const __m256 a = _mm256_loadu_ps(left + index);
+        const __m256 b = _mm256_loadu_ps(right + index);
+        accumulator = _mm256_fmadd_ps(a, b, accumulator);
+    }
+    alignas(32) std::array<float, 8> lanes{};
+    _mm256_store_ps(lanes.data(), accumulator);
+    float result = 0.0f;
+    for (float value : lanes) {
+        result += value;
+    }
+    return result;
+}
+
+IFS_TARGET("avx512f")
+float dot_f32_avx512f(const float *left, const float *right) {
+    __m512 accumulator = _mm512_setzero_ps();
+    for (std::size_t index = 0; index < kDimension; index += 16) {
+        const __m512 a = _mm512_loadu_ps(left + index);
+        const __m512 b = _mm512_loadu_ps(right + index);
+        accumulator = _mm512_add_ps(accumulator, _mm512_mul_ps(a, b));
+    }
+    alignas(64) std::array<float, 16> lanes{};
+    _mm512_store_ps(lanes.data(), accumulator);
+    float result = 0.0f;
+    for (float value : lanes) {
+        result += value;
+    }
+    return result;
+}
+
+IFS_TARGET("avx2,fma")
+float dot_bf16_avx2_fma(const std::uint16_t *left, const std::uint16_t *right) {
+    __m256 accumulator = _mm256_setzero_ps();
+    for (std::size_t index = 0; index < kDimension; index += 8) {
+        const __m128i a16 = _mm_loadu_si128(
+            reinterpret_cast<const __m128i *>(left + index));
+        const __m128i b16 = _mm_loadu_si128(
+            reinterpret_cast<const __m128i *>(right + index));
+        const __m256 a = _mm256_castsi256_ps(
+            _mm256_slli_epi32(_mm256_cvtepu16_epi32(a16), 16));
+        const __m256 b = _mm256_castsi256_ps(
+            _mm256_slli_epi32(_mm256_cvtepu16_epi32(b16), 16));
+        accumulator = _mm256_fmadd_ps(a, b, accumulator);
+    }
+    alignas(32) std::array<float, 8> lanes{};
+    _mm256_store_ps(lanes.data(), accumulator);
+    float result = 0.0f;
+    for (float value : lanes) {
+        result += value;
+    }
+    return result;
+}
+
+IFS_TARGET("avx512bf16,avx512f")
+float dot_bf16_avx512_bf16(
+    const std::uint16_t *left,
+    const std::uint16_t *right) {
+    __m512 accumulator = _mm512_setzero_ps();
+    for (std::size_t index = 0; index < kDimension; index += 32) {
+        const __m512bh a = (__m512bh)_mm512_loadu_si512(
+            reinterpret_cast<const void *>(left + index));
+        const __m512bh b = (__m512bh)_mm512_loadu_si512(
+            reinterpret_cast<const void *>(right + index));
+        accumulator = _mm512_dpbf16_ps(accumulator, a, b);
+    }
+    alignas(64) std::array<float, 16> lanes{};
+    _mm512_store_ps(lanes.data(), accumulator);
+    float result = 0.0f;
+    for (float value : lanes) {
+        result += value;
+    }
+    return result;
+}
+
+IFS_TARGET("avx2")
+std::int32_t dot_int8_avx2(
+    const std::int8_t *left,
+    const std::int8_t *right,
+    std::int32_t /*query_sum*/) {
+    __m256i accumulator = _mm256_setzero_si256();
+    for (std::size_t index = 0; index < kDimension; index += 32) {
+        for (std::size_t half = 0; half < 32; half += 16) {
+            const __m128i a8 = _mm_loadu_si128(
+                reinterpret_cast<const __m128i *>(left + index + half));
+            const __m128i b8 = _mm_loadu_si128(
+                reinterpret_cast<const __m128i *>(right + index + half));
+            const __m256i a16 = _mm256_cvtepi8_epi16(a8);
+            const __m256i b16 = _mm256_cvtepi8_epi16(b8);
+            accumulator = _mm256_add_epi32(
+                accumulator, _mm256_madd_epi16(a16, b16));
+        }
+    }
+    alignas(32) std::array<std::int32_t, 8> lanes{};
+    _mm256_store_si256(reinterpret_cast<__m256i *>(lanes.data()), accumulator);
+    std::int32_t result = 0;
+    for (std::int32_t value : lanes) {
+        result += value;
+    }
+    return result;
+}
+
+IFS_TARGET("avxvnni,avx2")
+std::int32_t dot_int8_avx_vnni(
+    const std::int8_t *left,
+    const std::int8_t *right,
+    std::int32_t query_sum) {
+    const __m256i sign = _mm256_set1_epi8(static_cast<char>(0x80));
+    __m256i accumulator = _mm256_setzero_si256();
+    for (std::size_t index = 0; index < kDimension; index += 32) {
+        const __m256i signed_left = _mm256_loadu_si256(
+            reinterpret_cast<const __m256i *>(left + index));
+        const __m256i signed_right = _mm256_loadu_si256(
+            reinterpret_cast<const __m256i *>(right + index));
+        const __m256i unsigned_left = _mm256_xor_si256(signed_left, sign);
+        accumulator = _mm256_dpbusd_epi32(
+            accumulator, unsigned_left, signed_right);
+    }
+    alignas(32) std::array<std::int32_t, 8> lanes{};
+    _mm256_store_si256(reinterpret_cast<__m256i *>(lanes.data()), accumulator);
+    std::int32_t result = 0;
+    for (std::int32_t value : lanes) {
+        result += value;
+    }
+    return result - 128 * query_sum;
+}
+
+IFS_TARGET("avx512vnni,avx512f")
+std::int32_t dot_int8_avx512_vnni(
+    const std::int8_t *left,
+    const std::int8_t *right,
+    std::int32_t query_sum) {
+    const __m512i sign = _mm512_set1_epi8(static_cast<char>(0x80));
+    __m512i accumulator = _mm512_setzero_si512();
+    for (std::size_t index = 0; index < kDimension; index += 64) {
+        const __m512i signed_left = _mm512_loadu_si512(
+            reinterpret_cast<const void *>(left + index));
+        const __m512i signed_right = _mm512_loadu_si512(
+            reinterpret_cast<const void *>(right + index));
+        const __m512i unsigned_left = _mm512_xor_si512(signed_left, sign);
+        accumulator = _mm512_dpbusd_epi32(
+            accumulator, unsigned_left, signed_right);
+    }
+    alignas(64) std::array<std::int32_t, 16> lanes{};
+    _mm512_store_si512(reinterpret_cast<void *>(lanes.data()), accumulator);
+    std::int32_t result = 0;
+    for (std::int32_t value : lanes) {
+        result += value;
+    }
+    return result - 128 * query_sum;
+}
+
+#else
+
+CpuFeatures detect_features() {
+    CpuFeatures features;
+    features.x86_64 = IFS_CPU_X86_64 != 0;
+    return features;
+}
+
+#endif
+
+std::string feature_string(const CpuFeatures &features) {
+    std::ostringstream output;
+    output << "arch=" << (features.x86_64 ? "x86_64" : "non-x86_64")
+           << ";os_avx=" << features.os_avx
+           << ";os_avx512=" << features.os_avx512
+           << ";fma=" << features.fma
+           << ";avx2=" << features.avx2
+           << ";avx512f=" << features.avx512f
+           << ";avx_vnni=" << features.avx_vnni
+           << ";avx512_vnni=" << features.avx512_vnni
+           << ";avx512_bf16=" << features.avx512_bf16;
+    return output.str();
+}
+
+}  // namespace
+
+Dispatch make_dispatch() {
+    const CpuFeatures features = detect_features();
+    Dispatch selected{
+        dot_f32_portable,
+        dot_bf16_portable,
+        dot_int8_portable,
+        "portable-fp32",
+        "portable-bf16-fp32-accumulate",
+        "portable-int8-int32-accumulate",
+        feature_string(features),
+    };
+
+#if IFS_CPU_X86_TARGETS
+    if (features.avx2 && features.fma) {
+        selected.fp32 = dot_f32_avx2_fma;
+        selected.fp32_name = "avx2-fma-fp32";
+        selected.bf16 = dot_bf16_avx2_fma;
+        selected.bf16_name = "avx2-fma-bf16-fp32-accumulate";
+    }
+    if (features.avx512f) {
+        selected.fp32 = dot_f32_avx512f;
+        selected.fp32_name = "avx512f-fp32";
+    }
+    if (features.avx512_bf16) {
+        selected.bf16 = dot_bf16_avx512_bf16;
+        selected.bf16_name = "avx512-bf16-fp32-accumulate";
+    }
+    if (features.avx2) {
+        selected.int8 = dot_int8_avx2;
+        selected.int8_name = "avx2-int8-int32-accumulate";
+    }
+    if (features.avx_vnni) {
+        selected.int8 = dot_int8_avx_vnni;
+        selected.int8_name = "avx-vnni-signed-int8-int32-accumulate";
+    }
+    if (features.avx512_vnni) {
+        selected.int8 = dot_int8_avx512_vnni;
+        selected.int8_name = "avx512-vnni-signed-int8-int32-accumulate";
+    }
+#endif
+    return selected;
+}
+
+}  // namespace ifs_cpu
diff --git a/server/native/search/cpu/tests/smoke.cpp b/server/native/search/cpu/tests/smoke.cpp
new file mode 100644
index 000000000..791b64a19
--- /dev/null
+++ b/server/native/search/cpu/tests/smoke.cpp
@@ -0,0 +1,294 @@
+#include "ifs_search.h"
+
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+namespace {
+
+using Vector = std::array<float, IFS_SEARCH_DIMENSION>;
+
+Vector constant(float sign = 1.0f) {
+    Vector result{};
+    const float value = sign / std::sqrt(static_cast<float>(result.size()));
+    result.fill(value);
+    return result;
+}
+
+Vector split() {
+    Vector result = constant();
+    for (size_t i = result.size() / 2; i < result.size(); ++i) {
+        result[i] = -result[i];
+    }
+    return result;
+}
+
+float tolerance(uint32_t profile) {
+    if (profile == IFS_SEARCH_PROFILE_FP32_V1) return 1.0e-5f;
+    if (profile == IFS_SEARCH_PROFILE_BF16_V1) return 2.0e-2f;
+    return 3.0e-2f;
+}
+
+void normalize(Vector *value) {
+    double squared = 0.0;
+    for (float item : *value) squared += static_cast<double>(item) * item;
+    const float inverse = 1.0f / std::sqrt(static_cast<float>(squared));
+    for (float &item : *value) item *= inverse;
+}
+
+void test_int8_quantization_float32_contract(
+    uint32_t profile, float expected_score, float expected_group_score) {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = profile;
+    options.reserve_rows = 1;
+    options.max_rows = 1;
+    options.device = -1;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    Vector stored{};
+    Vector query{};
+    stored[0] = query[0] = 0.12449999898672104f;
+    const float tail = std::sqrt(
+        (1.0f - stored[0] * stored[0]) /
+        static_cast<float>(stored.size() - 1));
+    for (size_t i = 1; i < stored.size(); ++i) {
+        stored[i] = tail;
+        query[i] = i <= 255 ? tail : -tail;
+    }
+    normalize(&stored);
+    normalize(&query);
+    const uint64_t id = 1;
+    const uint64_t group = 1;
+    assert(ifs_search_add_batch(index, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+    uint64_t out_id = 0;
+    float score = 0.0f;
+    uint64_t count = 0;
+    assert(ifs_search_topk(index, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    // Multiplication remains FP32 and rounding is half away from zero on both
+    // supported scales, matching CUDA and the NumPy reference backend.
+    assert(std::abs(score - expected_score) < 1.0e-6f);
+    ifs_search_destroy(index);
+
+    options.reserve_rows = 2;
+    options.max_rows = 2;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    Vector first{};
+    Vector second{};
+    for (size_t i = 0; i < first.size(); ++i) {
+        first[i] = i < 99
+            ? 0.045f
+            : std::sqrt((1.0f - 99.0f * 0.045f * 0.045f) / 413.0f);
+        second[i] = i < 100
+            ? 0.045f
+            : std::sqrt((1.0f - 100.0f * 0.045f * 0.045f) / 412.0f);
+    }
+    normalize(&first);
+    normalize(&second);
+    std::vector<float> rows;
+    rows.insert(rows.end(), first.begin(), first.end());
+    rows.insert(rows.end(), second.begin(), second.end());
+    const std::array<uint64_t, 2> ids{10, 20};
+    const std::array<uint64_t, 2> groups{1, 1};
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(), rows.data(), 2) ==
+           IFS_SEARCH_OK);
+    uint64_t out_group = 0;
+    assert(ifs_search_grouped_topk(index, first.data(), 1, &out_group, &out_id,
+                                   &score, &count, nullptr) == IFS_SEARCH_OK);
+    // Group selection uses the unbounded accumulator before the public ABI
+    // clamps the returned raw-cosine approximation.
+    assert(count == 1 && out_group == 1 && out_id == 20);
+    assert(std::abs(score - expected_group_score) < 1.0e-6f);
+    ifs_search_destroy(index);
+}
+
+void test_int8_scales_are_per_index() {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.reserve_rows = 1;
+    options.max_rows = 1;
+    options.device = -1;
+    ifs_search_index_t x1000 = nullptr;
+    ifs_search_index_t x736 = nullptr;
+    options.profile = IFS_SEARCH_PROFILE_INT8_X1000_V1;
+    assert(ifs_search_create(&options, &x1000) == IFS_SEARCH_OK);
+    options.profile = IFS_SEARCH_PROFILE_INT8_X736_V1;
+    assert(ifs_search_create(&options, &x736) == IFS_SEARCH_OK);
+
+    Vector stored{};
+    Vector query{};
+    stored[0] = query[0] = 0.12449999898672104f;
+    const float tail = std::sqrt(
+        (1.0f - stored[0] * stored[0]) /
+        static_cast<float>(stored.size() - 1));
+    for (size_t i = 1; i < stored.size(); ++i) {
+        stored[i] = tail;
+        query[i] = i <= 255 ? tail : -tail;
+    }
+    normalize(&stored);
+    normalize(&query);
+    const uint64_t id = 1;
+    const uint64_t group = 1;
+    assert(ifs_search_add_batch(x1000, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+    assert(ifs_search_add_batch(x736, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+
+    uint64_t out_id = 0;
+    uint64_t count = 0;
+    float score = 0.0f;
+    assert(ifs_search_topk(x1000, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    assert(std::abs(score - 0.013689f) < 1.0e-6f);
+    assert(ifs_search_topk(x736, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    assert(std::abs(score - 7440.0f / 541696.0f) < 1.0e-6f);
+
+    ifs_search_destroy(x736);
+    ifs_search_destroy(x1000);
+}
+
+void test_profile(uint32_t profile) {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = profile;
+    options.reserve_rows = 3;
+    options.max_rows = 3;
+    options.device = -1;
+    options.topk_mode = IFS_SEARCH_TOPK_AUTO;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    assert(index != nullptr);
+
+    const Vector positive = constant();
+    const Vector orthogonal = split();
+    const Vector duplicate = positive;
+    std::vector<float> rows;
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    rows.insert(rows.end(), orthogonal.begin(), orthogonal.end());
+    rows.insert(rows.end(), duplicate.begin(), duplicate.end());
+    const std::array<uint64_t, 3> vector_ids{10, 20, 30};
+    const std::array<uint64_t, 3> group_ids{100, 200, 300};
+    assert(ifs_search_add_batch(index, vector_ids.data(), group_ids.data(),
+                                rows.data(), 3) == IFS_SEARCH_OK);
+
+    std::array<uint64_t, 3> ids{};
+    std::array<float, 3> scores{};
+    uint64_t count = 0;
+    ifs_search_timings_t timings{};
+    timings.struct_size = sizeof(timings);
+    assert(ifs_search_topk(index, positive.data(), 3, ids.data(), scores.data(),
+                           &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 3);
+    assert(ids[0] == 10 && ids[1] == 30 && ids[2] == 20);
+    assert(std::abs(scores[0] - 1.0f) < tolerance(profile));
+    assert(std::abs(scores[2]) < tolerance(profile));
+
+    std::array<uint64_t, 3> out_groups{};
+    std::array<uint64_t, 3> out_faces{};
+    assert(ifs_search_grouped_topk(
+               index, positive.data(), 3, out_groups.data(), out_faces.data(),
+               scores.data(), &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 3);
+    assert(out_groups[0] == 100 && out_faces[0] == 10);
+    assert(out_groups[1] == 300 && out_faces[1] == 30);
+    assert(out_groups[2] == 200 && out_faces[2] == 20);
+
+    uint64_t removed = 0;
+    const uint64_t deleted = 10;
+    assert(ifs_search_delete_batch(index, &deleted, 1, &removed) == IFS_SEARCH_OK);
+    assert(removed == 1);
+    const uint64_t replacement_id = 40;
+    const uint64_t replacement_group = 400;
+    assert(ifs_search_add_batch(index, &replacement_id, &replacement_group,
+                                positive.data(), 1) == IFS_SEARCH_OK);
+
+    const uint64_t overflow_id = 50;
+    const uint64_t overflow_group = 500;
+    assert(ifs_search_add_batch(index, &overflow_id, &overflow_group,
+                                positive.data(), 1) ==
+           IFS_SEARCH_CAPACITY_EXCEEDED);
+
+    ifs_search_stats_t stats{};
+    stats.struct_size = sizeof(stats);
+    assert(ifs_search_get_stats(index, &stats) == IFS_SEARCH_OK);
+    assert(stats.live_rows == 3);
+    assert(stats.physical_rows == 3);
+    assert(stats.tombstone_rows == 0);
+    assert(stats.capacity_rows == 3);
+    assert(stats.max_rows == 3);
+    ifs_search_destroy(index);
+}
+
+}  // namespace
+
+int main() {
+    static_assert(IFS_SEARCH_ABI_VERSION == 2u, "unexpected search ABI");
+    static_assert(IFS_SEARCH_DIMENSION == 512u, "fixed dimension changed");
+    static_assert(IFS_SEARCH_PROFILE_INT8_X1000_V1 == 3,
+                  "persisted x1000 profile code changed");
+    static_assert(IFS_SEARCH_PROFILE_INT8_X736_V1 == 4,
+                  "x736 profile must append after existing codes");
+    assert(ifs_search_abi_version() == 2u);
+    assert(ifs_search_dimension() == 512u);
+
+    ifs_search_capabilities_t capabilities{};
+    capabilities.struct_size = sizeof(capabilities);
+    assert(ifs_search_get_capabilities(-1, &capabilities) == IFS_SEARCH_OK);
+    assert(capabilities.backend == IFS_SEARCH_BACKEND_CPU);
+    assert(capabilities.compute_capability_major == -1);
+    assert(capabilities.compute_capability_minor == -1);
+    assert((capabilities.flags & IFS_SEARCH_CAP_GROUPED_PERSON_TOPK) != 0);
+    assert((capabilities.flags & IFS_SEARCH_CAP_GROUPED_HOST_REFERENCE) != 0);
+    assert((capabilities.profile_mask &
+            (UINT64_C(1) << IFS_SEARCH_PROFILE_FP16_V1)) == 0);
+    assert((capabilities.profile_mask &
+            (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X736_V1)) != 0);
+    assert(std::string(ifs_search_build_info()).find(
+               "grouped_topk=host-reference") != std::string::npos);
+
+    test_profile(IFS_SEARCH_PROFILE_FP32_V1);
+    test_profile(IFS_SEARCH_PROFILE_BF16_V1);
+    test_profile(IFS_SEARCH_PROFILE_INT8_X1000_V1);
+    test_profile(IFS_SEARCH_PROFILE_INT8_X736_V1);
+    test_int8_quantization_float32_contract(
+        IFS_SEARCH_PROFILE_INT8_X1000_V1, 0.013689f, 1.0f);
+    test_int8_quantization_float32_contract(
+        IFS_SEARCH_PROFILE_INT8_X736_V1, 7440.0f / 541696.0f,
+        530755.0f / 541696.0f);
+    test_int8_scales_are_per_index();
+
+    ifs_search_create_options_t fp16{};
+    fp16.struct_size = sizeof(fp16);
+    fp16.profile = IFS_SEARCH_PROFILE_FP16_V1;
+    fp16.device = -1;
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&fp16, &index) == IFS_SEARCH_UNSUPPORTED);
+    assert(index == nullptr);
+
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = IFS_SEARCH_PROFILE_FP32_V1;
+    options.reserve_rows = 1;
+    options.max_rows = 1;
+    options.device = -1;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    Vector invalid{};
+    invalid[0] = 0.5f;
+    const uint64_t id = 1;
+    const uint64_t group = 1;
+    assert(ifs_search_add_batch(index, &id, &group, invalid.data(), 1) ==
+           IFS_SEARCH_INVALID_ARGUMENT);
+    ifs_search_destroy(index);
+    return 0;
+}
diff --git a/server/native/search/cuda/CMakeLists.txt b/server/native/search/cuda/CMakeLists.txt
new file mode 100644
index 000000000..310a9dfdf
--- /dev/null
+++ b/server/native/search/cuda/CMakeLists.txt
@@ -0,0 +1,41 @@
+find_package(CUDAToolkit 12.9 REQUIRED)
+
+set(IFS_SEARCH_CUDA_ARCHITECTURES
+    "75-real;80-real;86-real;89-real;90-real;100-real;103-real;120"
+    CACHE STRING "CUDA SASS targets plus PTX for the newest target")
+
+add_library(ifs_search_cuda SHARED
+    src/api_v2.cpp
+    src/ifs_cuda_legacy.cu)
+
+target_compile_features(ifs_search_cuda PRIVATE cxx_std_17)
+set_target_properties(ifs_search_cuda PROPERTIES
+    CUDA_STANDARD 17
+    CUDA_STANDARD_REQUIRED YES
+    CXX_VISIBILITY_PRESET hidden
+    CUDA_VISIBILITY_PRESET hidden
+    VISIBILITY_INLINES_HIDDEN YES
+    OUTPUT_NAME ifs_search_cuda
+    VERSION ${PROJECT_VERSION}
+    SOVERSION 2
+    CUDA_ARCHITECTURES "${IFS_SEARCH_CUDA_ARCHITECTURES}")
+install(TARGETS ifs_search_cuda
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+target_compile_definitions(ifs_search_cuda PRIVATE IFS_SEARCH_BUILDING=1)
+target_include_directories(ifs_search_cuda
+    PUBLIC ${PROJECT_SOURCE_DIR}/include
+    PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+target_link_libraries(ifs_search_cuda PRIVATE CUDA::cudart CUDA::cublas)
+target_compile_options(ifs_search_cuda PRIVATE
+    $<$<COMPILE_LANGUAGE:CXX>:-O3;-Wall;-Wextra;-Wpedantic>
+    $<$<COMPILE_LANGUAGE:CUDA>:-O3;--expt-relaxed-constexpr>)
+
+if(IFS_SEARCH_BUILD_TESTS)
+    add_executable(ifs_search_cuda_smoke tests/smoke.cpp)
+    target_compile_features(ifs_search_cuda_smoke PRIVATE cxx_std_17)
+    target_link_libraries(ifs_search_cuda_smoke PRIVATE ifs_search_cuda)
+    target_compile_options(ifs_search_cuda_smoke PRIVATE -UNDEBUG)
+    add_test(NAME ifs_search_cuda_smoke COMMAND ifs_search_cuda_smoke)
+    set_tests_properties(ifs_search_cuda_smoke PROPERTIES SKIP_RETURN_CODE 77)
+endif()
diff --git a/server/native/search/cuda/src/api_v2.cpp b/server/native/search/cuda/src/api_v2.cpp
new file mode 100644
index 000000000..f6725e3eb
--- /dev/null
+++ b/server/native/search/cuda/src/api_v2.cpp
@@ -0,0 +1,514 @@
+#include "ifs_search.h"
+
+#include "ifs_cuda_legacy.h"
+
+#include <cuda_runtime_api.h>
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <string>
+#include <utility>
+
+namespace {
+
+constexpr double kDefaultGrowthFactor = 1.5;
+constexpr double kNormalizedSquaredTolerance = 1.0e-3;
+constexpr uint64_t kDeviceTopKLimit = 100;
+
+thread_local std::string g_last_error;
+
+struct CudaIndex {
+    void *core = nullptr;
+    ifs_search_profile_t profile = IFS_SEARCH_PROFILE_FP32_V1;
+    int32_t device = 0;
+    ifs_search_topk_mode_t topk_mode = IFS_SEARCH_TOPK_AUTO;
+    uint64_t max_rows = 0;
+    mutable std::mutex operation_mutex;
+
+    ~CudaIndex() { ifs_cuda_destroy(core); }
+};
+
+ifs_search_status_t fail(ifs_search_status_t status, const std::string &message) {
+    g_last_error = message;
+    return status;
+}
+
+void clear_error() { g_last_error.clear(); }
+
+float public_cosine(float score) {
+    return std::clamp(score, -1.0f, 1.0f);
+}
+
+CudaIndex *as_index(ifs_search_index_t index) {
+    return static_cast<CudaIndex *>(index);
+}
+
+bool valid_profile(uint32_t profile) {
+    return profile == IFS_SEARCH_PROFILE_FP32_V1 ||
+           profile == IFS_SEARCH_PROFILE_FP16_V1 ||
+           profile == IFS_SEARCH_PROFILE_BF16_V1 ||
+           profile == IFS_SEARCH_PROFILE_INT8_X1000_V1 ||
+           profile == IFS_SEARCH_PROFILE_INT8_X736_V1;
+}
+
+uint32_t int8_scale_for(uint32_t profile) {
+    switch (profile) {
+        case IFS_SEARCH_PROFILE_INT8_X1000_V1:
+            return IFS_SEARCH_INT8_X1000_SCALE;
+        case IFS_SEARCH_PROFILE_INT8_X736_V1:
+            return IFS_SEARCH_INT8_X736_SCALE;
+        default:
+            return 0;
+    }
+}
+
+uint64_t profile_mask_for_compute_capability(int major, int minor) {
+    const int capability = major * 10 + minor;
+    if (capability < 75) return 0;
+    uint64_t mask =
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_FP32_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_FP16_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X1000_V1) |
+        (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X736_V1);
+    /* Native BF16 matrix operations require Ampere (SM80) or newer. */
+    if (capability >= 80) {
+        mask |= UINT64_C(1) << IFS_SEARCH_PROFILE_BF16_V1;
+    }
+    return mask;
+}
+
+int legacy_dtype(uint32_t profile) {
+    switch (profile) {
+        case IFS_SEARCH_PROFILE_FP32_V1: return IFS_CUDA_FP32;
+        case IFS_SEARCH_PROFILE_FP16_V1: return IFS_CUDA_FP16;
+        case IFS_SEARCH_PROFILE_BF16_V1: return IFS_CUDA_BF16;
+        case IFS_SEARCH_PROFILE_INT8_X1000_V1: return IFS_CUDA_INT8;
+        case IFS_SEARCH_PROFILE_INT8_X736_V1: return IFS_CUDA_INT8;
+        default: return -1;
+    }
+}
+
+ifs_search_status_t classify_legacy_error(const char *error) {
+    if (!error) return IFS_SEARCH_BACKEND_ERROR;
+    if (std::strstr(error, "duplicate")) return IFS_SEARCH_DUPLICATE_ID;
+    if (std::strstr(error, "out of memory") ||
+        std::strstr(error, "allocation") || std::strstr(error, "allocate")) {
+        return IFS_SEARCH_OUT_OF_MEMORY;
+    }
+    if (std::strstr(error, "unsupported")) return IFS_SEARCH_UNSUPPORTED;
+    if (std::strstr(error, "null") || std::strstr(error, "must be") ||
+        std::strstr(error, "invalid")) {
+        return IFS_SEARCH_INVALID_ARGUMENT;
+    }
+    return IFS_SEARCH_BACKEND_ERROR;
+}
+
+ifs_search_status_t legacy_failure(const char *fallback) {
+    const char *legacy = ifs_cuda_last_error();
+    const std::string message = legacy && legacy[0] ? legacy : fallback;
+    return fail(classify_legacy_error(message.c_str()), message);
+}
+
+bool validate_normalized_rows(const float *vectors, uint64_t count,
+                              std::string *message) {
+    if (count && !vectors) {
+        *message = "vectors must not be null for a non-empty batch";
+        return false;
+    }
+    if (count > std::numeric_limits<size_t>::max() / IFS_SEARCH_DIMENSION) {
+        *message = "vector batch is too large for this process";
+        return false;
+    }
+    for (uint64_t row = 0; row < count; ++row) {
+        const float *vector = vectors + static_cast<size_t>(row) * IFS_SEARCH_DIMENSION;
+        double squared_norm = 0.0;
+        for (size_t column = 0; column < IFS_SEARCH_DIMENSION; ++column) {
+            const float value = vector[column];
+            if (!std::isfinite(value)) {
+                *message = "vector contains a non-finite value";
+                return false;
+            }
+            squared_norm += static_cast<double>(value) * value;
+        }
+        if (std::abs(squared_norm - 1.0) > kNormalizedSquaredTolerance) {
+            *message = "vector must be L2-normalized before indexing";
+            return false;
+        }
+    }
+    return true;
+}
+
+int resolve_topk_mode(ifs_search_topk_mode_t configured, uint64_t top_k) {
+    if (configured == IFS_SEARCH_TOPK_HOST) return IFS_CUDA_TOPK_HOST;
+    if (configured == IFS_SEARCH_TOPK_DEVICE) return IFS_CUDA_TOPK_DEVICE;
+    return top_k <= kDeviceTopKLimit
+        ? IFS_CUDA_TOPK_DEVICE : IFS_CUDA_TOPK_HOST;
+}
+
+void zero_timings(ifs_search_timings_t *timings) {
+    if (!timings) return;
+    timings->kernel_ms = 0.0;
+    timings->topk_ms = 0.0;
+    timings->total_ms = 0.0;
+}
+
+}  // namespace
+
+extern "C" {
+
+uint32_t ifs_search_abi_version(void) { return IFS_SEARCH_ABI_VERSION; }
+uint32_t ifs_search_dimension(void) { return IFS_SEARCH_DIMENSION; }
+
+const char *ifs_search_build_info(void) {
+    static thread_local std::string info;
+    info = "ifs-search-cuda ABI=2 d512 exact-flat raw-cosine "
+           "profiles=fp32,fp16,bf16,int8_x1000,int8_x736 "
+           "grouped_topk=device-resident-exact; ";
+    const char *legacy = ifs_cuda_build_info();
+    info += legacy && legacy[0] ? legacy : "legacy-build-info-unavailable";
+    return info.c_str();
+}
+
+const char *ifs_search_last_error(void) { return g_last_error.c_str(); }
+
+const char *ifs_search_status_string(ifs_search_status_t status) {
+    switch (status) {
+        case IFS_SEARCH_OK: return "ok";
+        case IFS_SEARCH_INVALID_ARGUMENT: return "invalid_argument";
+        case IFS_SEARCH_OUT_OF_MEMORY: return "out_of_memory";
+        case IFS_SEARCH_UNSUPPORTED: return "unsupported";
+        case IFS_SEARCH_DUPLICATE_ID: return "duplicate_id";
+        case IFS_SEARCH_ID_NOT_FOUND: return "id_not_found";
+        case IFS_SEARCH_CAPACITY_EXCEEDED: return "capacity_exceeded";
+        case IFS_SEARCH_BACKEND_ERROR: return "backend_error";
+        case IFS_SEARCH_INTERNAL_ERROR: return "internal_error";
+    }
+    return "unknown_status";
+}
+
+ifs_search_status_t ifs_search_get_capabilities(
+    int32_t device, ifs_search_capabilities_t *out) {
+    clear_error();
+    if (!out || out->struct_size < sizeof(*out)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "capabilities struct_size is too small");
+    }
+    if (device < 0) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "CUDA device ordinal must be non-negative");
+    }
+    cudaDeviceProp properties{};
+    cudaError_t status = cudaGetDeviceProperties(&properties, device);
+    if (status != cudaSuccess) {
+        return fail(IFS_SEARCH_BACKEND_ERROR,
+                    std::string("cudaGetDeviceProperties: ") +
+                    cudaGetErrorString(status));
+    }
+    int runtime_version = 0;
+    status = cudaRuntimeGetVersion(&runtime_version);
+    if (status != cudaSuccess) {
+        return fail(IFS_SEARCH_BACKEND_ERROR,
+                    std::string("cudaRuntimeGetVersion: ") +
+                    cudaGetErrorString(status));
+    }
+    int driver_version = 0;
+    status = cudaDriverGetVersion(&driver_version);
+    if (status != cudaSuccess) {
+        return fail(IFS_SEARCH_BACKEND_ERROR,
+                    std::string("cudaDriverGetVersion: ") +
+                    cudaGetErrorString(status));
+    }
+    const uint32_t requested_size = out->struct_size;
+    *out = {};
+    out->struct_size = requested_size;
+    out->abi_version = IFS_SEARCH_ABI_VERSION;
+    out->dimension = IFS_SEARCH_DIMENSION;
+    out->backend = IFS_SEARCH_BACKEND_CUDA;
+    out->profile_mask = profile_mask_for_compute_capability(
+        properties.major, properties.minor);
+    out->flags = IFS_SEARCH_CAP_EXACT_FLAT_SCAN |
+                 IFS_SEARCH_CAP_BATCH_ADD |
+                 IFS_SEARCH_CAP_BATCH_DELETE |
+                 IFS_SEARCH_CAP_RESERVE |
+                 IFS_SEARCH_CAP_DEVICE_TOPK |
+                 IFS_SEARCH_CAP_GROUPED_PERSON_TOPK |
+                 IFS_SEARCH_CAP_TOMBSTONE_DELETE |
+                 IFS_SEARCH_CAP_GROUPED_DEVICE_RESIDENT;
+    out->device_topk_limit = kDeviceTopKLimit;
+    out->device = device;
+    out->compute_capability_major = properties.major;
+    out->compute_capability_minor = properties.minor;
+    out->cuda_runtime_version = runtime_version;
+    out->cuda_driver_version = driver_version;
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_create(
+    const ifs_search_create_options_t *options,
+    ifs_search_index_t *out_index) {
+    clear_error();
+    if (!out_index) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_index must not be null");
+    }
+    *out_index = nullptr;
+    if (!options || options->struct_size < sizeof(*options)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "create options struct_size is too small");
+    }
+    if (!valid_profile(options->profile)) {
+        return fail(IFS_SEARCH_UNSUPPORTED, "unsupported CUDA search profile");
+    }
+    if (options->device < 0) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "CUDA device ordinal must be non-negative");
+    }
+    cudaDeviceProp properties{};
+    const cudaError_t property_status = cudaGetDeviceProperties(
+        &properties, options->device);
+    if (property_status != cudaSuccess) {
+        return fail(IFS_SEARCH_BACKEND_ERROR,
+                    std::string("cudaGetDeviceProperties: ") +
+                    cudaGetErrorString(property_status));
+    }
+    const uint64_t supported_profiles = profile_mask_for_compute_capability(
+        properties.major, properties.minor);
+    if ((supported_profiles & (UINT64_C(1) << options->profile)) == 0) {
+        return fail(IFS_SEARCH_UNSUPPORTED,
+                    "requested profile is unsupported by this CUDA compute capability");
+    }
+    if (options->topk_mode > IFS_SEARCH_TOPK_DEVICE) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "invalid Top-K mode");
+    }
+    if (options->max_rows && options->reserve_rows > options->max_rows) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "reserve_rows must not exceed max_rows");
+    }
+    const double growth = options->growth_factor == 0.0
+        ? kDefaultGrowthFactor : options->growth_factor;
+    if (!(growth >= 1.1 && growth <= 4.0)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "growth_factor must be between 1.1 and 4.0");
+    }
+    try {
+        std::unique_ptr<CudaIndex> index(new CudaIndex());
+        index->profile = static_cast<ifs_search_profile_t>(options->profile);
+        index->device = options->device;
+        index->topk_mode = static_cast<ifs_search_topk_mode_t>(
+            options->topk_mode);
+        index->max_rows = options->max_rows;
+        index->core = ifs_cuda_create(legacy_dtype(options->profile),
+                                      options->reserve_rows, options->device,
+                                      growth, int8_scale_for(options->profile));
+        if (!index->core) return legacy_failure("CUDA index creation failed");
+        *out_index = index.release();
+        return IFS_SEARCH_OK;
+    } catch (const std::bad_alloc &) {
+        return fail(IFS_SEARCH_OUT_OF_MEMORY, "CUDA wrapper allocation failed");
+    } catch (const std::exception &error) {
+        return fail(IFS_SEARCH_INTERNAL_ERROR, error.what());
+    }
+}
+
+void ifs_search_destroy(ifs_search_index_t index) {
+    clear_error();
+    delete as_index(index);
+}
+
+ifs_search_status_t ifs_search_reserve(ifs_search_index_t opaque,
+                                       uint64_t rows) {
+    clear_error();
+    if (!opaque) return fail(IFS_SEARCH_INVALID_ARGUMENT, "index is null");
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    if (index->max_rows && rows > index->max_rows) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "requested capacity exceeds the configured max_rows");
+    }
+    if (ifs_cuda_reserve(index->core, rows) != 0) {
+        return legacy_failure("CUDA reserve failed");
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_add_batch(
+    ifs_search_index_t opaque, const uint64_t *ids,
+    const uint64_t *group_ids, const float *vectors, uint64_t count) {
+    clear_error();
+    if (!opaque || (count && (!ids || !group_ids || !vectors))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index, vector IDs, group IDs, and vectors are required for a non-empty add");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(vectors, count, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    ifs_cuda_stats stats{};
+    if (ifs_cuda_get_stats(index->core, &stats) != 0) {
+        return legacy_failure("CUDA stats failed before add");
+    }
+    if (count > std::numeric_limits<uint64_t>::max() - stats.size ||
+        (index->max_rows && stats.size + count > index->max_rows)) {
+        return fail(IFS_SEARCH_CAPACITY_EXCEEDED,
+                    "add would exceed CUDA max_rows; tombstones require a rebuild");
+    }
+    if (ifs_cuda_add(index->core, vectors, ids, group_ids, count) != 0) {
+        return legacy_failure("CUDA batch add failed");
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_delete_batch(
+    ifs_search_index_t opaque, const uint64_t *ids, uint64_t count,
+    uint64_t *out_removed) {
+    clear_error();
+    if (!out_removed) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_removed must not be null");
+    }
+    *out_removed = 0;
+    if (!opaque || (count && !ids)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index and IDs are required for a non-empty delete");
+    }
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    if (ifs_cuda_remove(index->core, ids, count, out_removed) != 0) {
+        return legacy_failure("CUDA batch delete failed");
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_topk(
+    ifs_search_index_t opaque, const float *query, uint64_t top_k,
+    uint64_t *out_ids, float *out_scores, uint64_t *out_count,
+    ifs_search_timings_t *timings) {
+    clear_error();
+    if (!out_count) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_count must not be null");
+    }
+    *out_count = 0;
+    if (timings && timings->struct_size < sizeof(*timings)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "timings struct_size is too small");
+    }
+    zero_timings(timings);
+    if (!opaque || !query || (top_k && (!out_ids || !out_scores))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "invalid index, query, or Top-K output buffers");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(query, 1, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    const int mode = resolve_topk_mode(index->topk_mode, top_k);
+    if (mode == IFS_CUDA_TOPK_DEVICE && top_k > kDeviceTopKLimit) {
+        return fail(IFS_SEARCH_UNSUPPORTED,
+                    "explicit CUDA device Top-K supports at most 100 results");
+    }
+    double kernel_ms = 0.0;
+    double topk_ms = 0.0;
+    double total_ms = 0.0;
+    if (ifs_cuda_search_ex(index->core, query, top_k, mode, out_ids,
+                           out_scores, out_count, &kernel_ms, &topk_ms,
+                           &total_ms) != 0) {
+        return legacy_failure("CUDA search failed");
+    }
+    for (uint64_t offset = 0; offset < *out_count; ++offset) {
+        out_scores[offset] = public_cosine(out_scores[offset]);
+    }
+    if (timings) {
+        timings->kernel_ms = kernel_ms;
+        timings->topk_ms = topk_ms;
+        timings->total_ms = total_ms;
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_grouped_topk(
+    ifs_search_index_t opaque, const float *query, uint64_t top_k,
+    uint64_t *out_group_ids, uint64_t *out_vector_ids, float *out_scores,
+    uint64_t *out_count, ifs_search_timings_t *timings) {
+    clear_error();
+    if (!out_count) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, "out_count must not be null");
+    }
+    *out_count = 0;
+    if (timings && timings->struct_size < sizeof(*timings)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "timings struct_size is too small");
+    }
+    zero_timings(timings);
+    if (!opaque || !query ||
+        (top_k && (!out_group_ids || !out_vector_ids || !out_scores))) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "invalid grouped Top-K input or output buffers");
+    }
+    std::string validation_error;
+    if (!validate_normalized_rows(query, 1, &validation_error)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT, validation_error);
+    }
+
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    if (top_k > kDeviceTopKLimit) {
+        return fail(IFS_SEARCH_UNSUPPORTED,
+                    "CUDA grouped device Top-K supports at most 100 results");
+    }
+    double kernel_ms = 0.0;
+    double topk_ms = 0.0;
+    double total_ms = 0.0;
+    if (ifs_cuda_grouped_search(
+            index->core, query, top_k, out_group_ids, out_vector_ids,
+            out_scores, out_count, &kernel_ms, &topk_ms, &total_ms) != 0) {
+        return legacy_failure("CUDA GPU-resident grouped search failed");
+    }
+    for (uint64_t offset = 0; offset < *out_count; ++offset) {
+        out_scores[offset] = public_cosine(out_scores[offset]);
+    }
+    if (timings) {
+        timings->kernel_ms = kernel_ms;
+        timings->topk_ms = topk_ms;
+        timings->total_ms = total_ms;
+    }
+    return IFS_SEARCH_OK;
+}
+
+ifs_search_status_t ifs_search_get_stats(
+    ifs_search_index_t opaque, ifs_search_stats_t *out) {
+    clear_error();
+    if (!opaque || !out || out->struct_size < sizeof(*out)) {
+        return fail(IFS_SEARCH_INVALID_ARGUMENT,
+                    "index and a complete stats structure are required");
+    }
+    CudaIndex *index = as_index(opaque);
+    std::lock_guard<std::mutex> guard(index->operation_mutex);
+    ifs_cuda_stats legacy{};
+    if (ifs_cuda_get_stats(index->core, &legacy) != 0) {
+        return legacy_failure("CUDA stats failed");
+    }
+    const uint32_t requested_size = out->struct_size;
+    *out = {};
+    out->struct_size = requested_size;
+    out->backend = IFS_SEARCH_BACKEND_CUDA;
+    out->profile = index->profile;
+    out->device = index->device;
+    out->physical_rows = legacy.size;
+    out->live_rows = legacy.live_size;
+    out->capacity_rows = legacy.capacity;
+    out->max_rows = index->max_rows;
+    out->tombstone_rows = legacy.size - legacy.live_size;
+    out->reallocations = legacy.reallocations;
+    out->bytes_per_vector = legacy.bytes_per_vector;
+    return IFS_SEARCH_OK;
+}
+
+}  // extern "C"
diff --git a/server/native/search/cuda/src/ifs_cuda_legacy.cu b/server/native/search/cuda/src/ifs_cuda_legacy.cu
new file mode 100644
index 000000000..1333f8a1f
--- /dev/null
+++ b/server/native/search/cuda/src/ifs_cuda_legacy.cu
@@ -0,0 +1,1590 @@
+#include "ifs_cuda_legacy.h"
+
+#include <cublas_v2.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+#include <algorithm>
+#include <chrono>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <queue>
+#include <sstream>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+namespace {
+
+constexpr int kDimension = 512;
+constexpr int kTopKBlockSize = 256;
+constexpr uint64_t kTopKChunkRows = 8192;
+constexpr uint64_t kDeviceTopKLimit = 100;
+thread_local std::string g_last_error;
+
+void check_cuda(cudaError_t status, const char *what) {
+    if (status != cudaSuccess) {
+        std::ostringstream message;
+        message << what << ": " << cudaGetErrorString(status);
+        throw std::runtime_error(message.str());
+    }
+}
+
+void check_cublas(cublasStatus_t status, const char *what) {
+    if (status != CUBLAS_STATUS_SUCCESS) {
+        std::ostringstream message;
+        message << what << ": cuBLAS status " << static_cast<int>(status);
+        throw std::runtime_error(message.str());
+    }
+}
+
+size_t element_bytes(int dtype) {
+    switch (dtype) {
+        case IFS_CUDA_FP32: return sizeof(float);
+        case IFS_CUDA_FP16: return sizeof(__half);
+        case IFS_CUDA_BF16: return sizeof(__nv_bfloat16);
+        case IFS_CUDA_INT8: return sizeof(int8_t);
+        default: throw std::invalid_argument("unsupported CUDA index dtype");
+    }
+}
+
+__global__ void float_to_half(const float *input, __half *output, size_t count) {
+    const size_t i = static_cast<size_t>(blockIdx.x) * blockDim.x + threadIdx.x;
+    if (i < count) output[i] = __float2half_rn(input[i]);
+}
+
+__global__ void float_to_bfloat16(const float *input, __nv_bfloat16 *output,
+                                  size_t count) {
+    const size_t i = static_cast<size_t>(blockIdx.x) * blockDim.x + threadIdx.x;
+    if (i < count) output[i] = __float2bfloat16_rn(input[i]);
+}
+
+__global__ void float_to_int8(const float *input, int8_t *output,
+                              size_t count, float scale) {
+    const size_t i = static_cast<size_t>(blockIdx.x) * blockDim.x + threadIdx.x;
+    if (i < count) {
+        const float scaled = input[i] * scale;
+        int value = 0;
+        if (scaled >= 127.0f) value = 127;
+        else if (scaled <= -128.0f) value = -128;
+        else value = static_cast<int>(roundf(scaled));
+        output[i] = static_cast<int8_t>(value);
+    }
+}
+
+__global__ void mark_dead(uint8_t *alive, const uint64_t *slots,
+                          uint64_t count) {
+    const uint64_t i = static_cast<uint64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
+    if (i < count) alive[slots[i]] = 0;
+}
+
+template <typename Score>
+struct alignas(16) DeviceCandidate {
+    Score score;
+    uint64_t id;
+};
+
+template <typename Score>
+__device__ __forceinline__ Score worst_score();
+
+template <>
+__device__ __forceinline__ float worst_score<float>() {
+    return -INFINITY;
+}
+
+template <>
+__device__ __forceinline__ int32_t worst_score<int32_t>() {
+    return INT32_MIN;
+}
+
+template <typename Score>
+__device__ __forceinline__ DeviceCandidate<Score> worst_candidate() {
+    return DeviceCandidate<Score>{worst_score<Score>(), UINT64_MAX};
+}
+
+template <typename Score>
+__device__ __forceinline__ bool device_better(
+    const DeviceCandidate<Score> &left,
+    const DeviceCandidate<Score> &right) {
+    if (left.score != right.score) return left.score > right.score;
+    return left.id < right.id;
+}
+
+template <typename Score>
+__device__ __forceinline__ void bitonic_sort_best_first(
+    DeviceCandidate<Score> *values) {
+    const unsigned int tid = threadIdx.x;
+    #pragma unroll
+    for (unsigned int sequence = 2; sequence <= kTopKBlockSize;
+         sequence <<= 1) {
+        #pragma unroll
+        for (unsigned int stride = sequence >> 1; stride > 0; stride >>= 1) {
+            const unsigned int peer = tid ^ stride;
+            if (peer > tid) {
+                const DeviceCandidate<Score> mine = values[tid];
+                const DeviceCandidate<Score> theirs = values[peer];
+                const bool best_first_segment = (tid & sequence) == 0;
+                const bool swap = best_first_segment
+                    ? device_better(theirs, mine)
+                    : device_better(mine, theirs);
+                if (swap) {
+                    values[tid] = theirs;
+                    values[peer] = mine;
+                }
+            }
+            __syncthreads();
+        }
+    }
+}
+
+template <typename Score>
+__global__ void topk_from_scores(
+    const Score *scores, const uint64_t *ids, const uint8_t *alive,
+    uint64_t count, uint64_t k, DeviceCandidate<Score> *output) {
+    __shared__ DeviceCandidate<Score> candidates[kTopKBlockSize];
+    const uint64_t range_begin = static_cast<uint64_t>(blockIdx.x) * kTopKChunkRows;
+    const uint64_t range_end = min(count, range_begin + kTopKChunkRows);
+    const uint64_t first = range_begin + threadIdx.x;
+    candidates[threadIdx.x] = first < range_end && alive[first]
+        ? DeviceCandidate<Score>{scores[first], ids[first]}
+        : worst_candidate<Score>();
+    __syncthreads();
+    bitonic_sort_best_first(candidates);
+
+    uint64_t cursor = range_begin + kTopKBlockSize;
+    const uint64_t fresh_per_round = kTopKBlockSize - k;
+    while (cursor < range_end) {
+        if (threadIdx.x >= k) {
+            const uint64_t row = cursor + threadIdx.x - k;
+            candidates[threadIdx.x] = row < range_end && alive[row]
+                ? DeviceCandidate<Score>{scores[row], ids[row]}
+                : worst_candidate<Score>();
+        }
+        __syncthreads();
+        bitonic_sort_best_first(candidates);
+        cursor += fresh_per_round;
+    }
+    if (threadIdx.x < k) {
+        output[static_cast<uint64_t>(blockIdx.x) * k + threadIdx.x] =
+            candidates[threadIdx.x];
+    }
+}
+
+template <typename Score>
+__global__ void topk_from_candidates(
+    const DeviceCandidate<Score> *input, uint64_t count, uint64_t k,
+    DeviceCandidate<Score> *output) {
+    __shared__ DeviceCandidate<Score> candidates[kTopKBlockSize];
+    const uint64_t range_begin = static_cast<uint64_t>(blockIdx.x) * kTopKChunkRows;
+    const uint64_t range_end = min(count, range_begin + kTopKChunkRows);
+    const uint64_t first = range_begin + threadIdx.x;
+    candidates[threadIdx.x] = first < range_end
+        ? input[first] : worst_candidate<Score>();
+    __syncthreads();
+    bitonic_sort_best_first(candidates);
+
+    uint64_t cursor = range_begin + kTopKBlockSize;
+    const uint64_t fresh_per_round = kTopKBlockSize - k;
+    while (cursor < range_end) {
+        if (threadIdx.x >= k) {
+            const uint64_t row = cursor + threadIdx.x - k;
+            candidates[threadIdx.x] = row < range_end
+                ? input[row] : worst_candidate<Score>();
+        }
+        __syncthreads();
+        bitonic_sort_best_first(candidates);
+        cursor += fresh_per_round;
+    }
+    if (threadIdx.x < k) {
+        output[static_cast<uint64_t>(blockIdx.x) * k + threadIdx.x] =
+            candidates[threadIdx.x];
+    }
+}
+
+template <typename Score>
+__device__ __forceinline__ uint32_t ordered_score_key(Score score);
+
+template <>
+__device__ __forceinline__ uint32_t ordered_score_key<float>(float score) {
+    /* Canonicalize signed zero so tie-breaking agrees with C++ float equality
+     * and the host reference path. */
+    if (score == 0.0f) return UINT32_C(0x80000000);
+    const uint32_t bits = __float_as_uint(score);
+    return (bits & UINT32_C(0x80000000)) ? ~bits
+                                         : bits ^ UINT32_C(0x80000000);
+}
+
+template <>
+__device__ __forceinline__ uint32_t ordered_score_key<int32_t>(int32_t score) {
+    return static_cast<uint32_t>(score) ^ UINT32_C(0x80000000);
+}
+
+template <typename Score>
+__device__ __forceinline__ Score score_from_ordered_key(uint32_t key);
+
+template <>
+__device__ __forceinline__ float score_from_ordered_key<float>(uint32_t key) {
+    const uint32_t bits = (key & UINT32_C(0x80000000))
+        ? key ^ UINT32_C(0x80000000) : ~key;
+    return __uint_as_float(bits);
+}
+
+template <>
+__device__ __forceinline__ int32_t score_from_ordered_key<int32_t>(
+    uint32_t key) {
+    return static_cast<int32_t>(key ^ UINT32_C(0x80000000));
+}
+
+template <typename Score>
+__global__ void reduce_group_best_scores(
+    const Score *scores, const uint32_t *row_group_slots,
+    const uint8_t *alive, uint64_t count, uint32_t *group_best_keys) {
+    const uint64_t row = static_cast<uint64_t>(blockIdx.x) * blockDim.x +
+                         threadIdx.x;
+    if (row < count && alive[row]) {
+        atomicMax(group_best_keys + row_group_slots[row],
+                  ordered_score_key<Score>(scores[row]));
+    }
+}
+
+template <typename Score>
+__global__ void reduce_group_best_vector_ids(
+    const Score *scores, const uint64_t *vector_ids,
+    const uint32_t *row_group_slots, const uint8_t *alive, uint64_t count,
+    const uint32_t *group_best_keys, uint64_t *group_best_vector_ids) {
+    const uint64_t row = static_cast<uint64_t>(blockIdx.x) * blockDim.x +
+                         threadIdx.x;
+    if (row < count && alive[row]) {
+        const uint32_t group_slot = row_group_slots[row];
+        if (ordered_score_key<Score>(scores[row]) ==
+            group_best_keys[group_slot]) {
+            atomicMin(reinterpret_cast<unsigned long long *>(
+                          group_best_vector_ids + group_slot),
+                      static_cast<unsigned long long>(vector_ids[row]));
+        }
+    }
+}
+
+template <typename Score>
+struct alignas(16) DeviceGroupCandidate {
+    Score score;
+    uint64_t group_id;
+    uint64_t vector_id;
+};
+
+template <typename Score>
+__device__ __forceinline__ DeviceGroupCandidate<Score>
+worst_group_candidate() {
+    return DeviceGroupCandidate<Score>{worst_score<Score>(), UINT64_MAX,
+                                       UINT64_MAX};
+}
+
+template <typename Score>
+__device__ __forceinline__ bool device_group_better(
+    const DeviceGroupCandidate<Score> &left,
+    const DeviceGroupCandidate<Score> &right) {
+    if (left.score != right.score) return left.score > right.score;
+    if (left.group_id != right.group_id) {
+        return left.group_id < right.group_id;
+    }
+    return left.vector_id < right.vector_id;
+}
+
+template <typename Score>
+__device__ __forceinline__ void bitonic_sort_groups_best_first(
+    DeviceGroupCandidate<Score> *values) {
+    const unsigned int tid = threadIdx.x;
+    #pragma unroll
+    for (unsigned int sequence = 2; sequence <= kTopKBlockSize;
+         sequence <<= 1) {
+        #pragma unroll
+        for (unsigned int stride = sequence >> 1; stride > 0; stride >>= 1) {
+            const unsigned int peer = tid ^ stride;
+            if (peer > tid) {
+                const DeviceGroupCandidate<Score> mine = values[tid];
+                const DeviceGroupCandidate<Score> theirs = values[peer];
+                const bool best_first_segment = (tid & sequence) == 0;
+                const bool swap = best_first_segment
+                    ? device_group_better(theirs, mine)
+                    : device_group_better(mine, theirs);
+                if (swap) {
+                    values[tid] = theirs;
+                    values[peer] = mine;
+                }
+            }
+            __syncthreads();
+        }
+    }
+}
+
+template <typename Score>
+__global__ void grouped_topk_from_best(
+    const uint32_t *best_keys, const uint64_t *best_vector_ids,
+    const uint64_t *group_ids, uint64_t group_count, uint64_t k,
+    DeviceGroupCandidate<Score> *output) {
+    __shared__ DeviceGroupCandidate<Score> candidates[kTopKBlockSize];
+    const uint64_t range_begin =
+        static_cast<uint64_t>(blockIdx.x) * kTopKChunkRows;
+    const uint64_t range_end = min(group_count, range_begin + kTopKChunkRows);
+    const uint64_t first = range_begin + threadIdx.x;
+    candidates[threadIdx.x] =
+        first < range_end && best_vector_ids[first] != UINT64_MAX
+        ? DeviceGroupCandidate<Score>{score_from_ordered_key<Score>(
+                                          best_keys[first]),
+                                      group_ids[first],
+                                      best_vector_ids[first]}
+        : worst_group_candidate<Score>();
+    __syncthreads();
+    bitonic_sort_groups_best_first(candidates);
+
+    uint64_t cursor = range_begin + kTopKBlockSize;
+    const uint64_t fresh_per_round = kTopKBlockSize - k;
+    while (cursor < range_end) {
+        if (threadIdx.x >= k) {
+            const uint64_t group_slot = cursor + threadIdx.x - k;
+            candidates[threadIdx.x] =
+                group_slot < range_end &&
+                    best_vector_ids[group_slot] != UINT64_MAX
+                ? DeviceGroupCandidate<Score>{
+                      score_from_ordered_key<Score>(best_keys[group_slot]),
+                      group_ids[group_slot], best_vector_ids[group_slot]}
+                : worst_group_candidate<Score>();
+        }
+        __syncthreads();
+        bitonic_sort_groups_best_first(candidates);
+        cursor += fresh_per_round;
+    }
+    if (threadIdx.x < k) {
+        output[static_cast<uint64_t>(blockIdx.x) * k + threadIdx.x] =
+            candidates[threadIdx.x];
+    }
+}
+
+template <typename Score>
+__global__ void grouped_topk_from_candidates(
+    const DeviceGroupCandidate<Score> *input, uint64_t count, uint64_t k,
+    DeviceGroupCandidate<Score> *output) {
+    __shared__ DeviceGroupCandidate<Score> candidates[kTopKBlockSize];
+    const uint64_t range_begin =
+        static_cast<uint64_t>(blockIdx.x) * kTopKChunkRows;
+    const uint64_t range_end = min(count, range_begin + kTopKChunkRows);
+    const uint64_t first = range_begin + threadIdx.x;
+    candidates[threadIdx.x] = first < range_end
+        ? input[first] : worst_group_candidate<Score>();
+    __syncthreads();
+    bitonic_sort_groups_best_first(candidates);
+
+    uint64_t cursor = range_begin + kTopKBlockSize;
+    const uint64_t fresh_per_round = kTopKBlockSize - k;
+    while (cursor < range_end) {
+        if (threadIdx.x >= k) {
+            const uint64_t row = cursor + threadIdx.x - k;
+            candidates[threadIdx.x] = row < range_end
+                ? input[row] : worst_group_candidate<Score>();
+        }
+        __syncthreads();
+        bitonic_sort_groups_best_first(candidates);
+        cursor += fresh_per_round;
+    }
+    if (threadIdx.x < k) {
+        output[static_cast<uint64_t>(blockIdx.x) * k + threadIdx.x] =
+            candidates[threadIdx.x];
+    }
+}
+
+struct Candidate {
+    float score;
+    uint64_t id;
+};
+
+bool better(const Candidate &left, const Candidate &right) {
+    if (left.score != right.score) return left.score > right.score;
+    return left.id < right.id;
+}
+
+/* priority_queue top is the worst retained candidate. */
+struct BetterComparator {
+    bool operator()(const Candidate &left, const Candidate &right) const {
+        return better(left, right);
+    }
+};
+
+class CudaFlatIndex {
+public:
+    CudaFlatIndex(int dtype, uint64_t reserve_rows, int device,
+                  double growth_factor, uint32_t int8_scale)
+        : dtype_(dtype), device_(device), growth_factor_(growth_factor),
+          int8_scale_(int8_scale),
+          int8_score_divisor_(static_cast<float>(int8_scale) *
+                              static_cast<float>(int8_scale)),
+          bytes_per_vector_(static_cast<uint64_t>(kDimension) * element_bytes(dtype)) {
+        if (!(growth_factor_ >= 1.1 && growth_factor_ <= 4.0)) {
+            throw std::invalid_argument("growth_factor must be between 1.1 and 4.0");
+        }
+        if ((dtype_ == IFS_CUDA_INT8) != (int8_scale_ > 0)) {
+            throw std::invalid_argument(
+                "INT8 dtype requires a positive scale and other dtypes require zero");
+        }
+        check_cuda(cudaSetDevice(device_), "cudaSetDevice");
+        check_cublas(cublasCreate(&cublas_), "cublasCreate");
+        check_cublas(cublasSetMathMode(cublas_, CUBLAS_TENSOR_OP_MATH),
+                     "cublasSetMathMode");
+        try {
+            allocate_query();
+            if (reserve_rows > 0) reserve_impl(reserve_rows, false);
+        } catch (...) {
+            cleanup();
+            throw;
+        }
+    }
+
+    ~CudaFlatIndex() { cleanup(); }
+
+    void reserve(uint64_t rows) {
+        std::lock_guard<std::mutex> guard(mutex_);
+        set_device();
+        reserve_impl(rows, false);
+    }
+
+    void add(const float *vectors, const uint64_t *ids,
+             const uint64_t *group_ids, uint64_t count) {
+        if (count == 0) return;
+        if (!vectors || !ids || !group_ids) {
+            throw std::invalid_argument("add received a null pointer");
+        }
+        std::lock_guard<std::mutex> guard(mutex_);
+        set_device();
+
+        std::unordered_set<uint64_t> batch_ids;
+        batch_ids.reserve(static_cast<size_t>(count));
+        for (uint64_t i = 0; i < count; ++i) {
+            if (ids[i] == UINT64_MAX) {
+                throw std::invalid_argument(
+                    "invalid vector ID: UINT64_MAX is reserved");
+            }
+            if (id_to_slot_.count(ids[i]) || !batch_ids.insert(ids[i]).second) {
+                throw std::invalid_argument("duplicate external ID in add");
+            }
+        }
+        if (count > std::numeric_limits<uint64_t>::max() - size_) {
+            throw std::overflow_error("index row count overflow");
+        }
+        const uint64_t required = size_ + count;
+        if (required > capacity_) {
+            uint64_t grown = capacity_ ? capacity_ : 1;
+            while (grown < required) {
+                const uint64_t next = static_cast<uint64_t>(
+                    std::ceil(static_cast<double>(grown) * growth_factor_));
+                if (next <= grown) { grown = required; break; }
+                grown = next;
+            }
+            reserve_impl(std::max(grown, required), true);
+        }
+
+        const size_t old_group_count = group_ids_.size();
+        std::vector<uint64_t> new_group_ids;
+        std::unordered_map<uint64_t, uint32_t> pending_group_slots;
+        std::vector<uint32_t> row_group_slots(static_cast<size_t>(count));
+        new_group_ids.reserve(static_cast<size_t>(count));
+        pending_group_slots.reserve(static_cast<size_t>(count));
+        for (uint64_t i = 0; i < count; ++i) {
+            const auto existing = group_to_slot_.find(group_ids[i]);
+            if (existing != group_to_slot_.end()) {
+                row_group_slots[static_cast<size_t>(i)] = existing->second;
+                continue;
+            }
+            const auto pending = pending_group_slots.find(group_ids[i]);
+            if (pending != pending_group_slots.end()) {
+                row_group_slots[static_cast<size_t>(i)] = pending->second;
+                continue;
+            }
+            const uint64_t next_slot = old_group_count + new_group_ids.size();
+            if (next_slot >= static_cast<uint64_t>(UINT32_MAX)) {
+                throw std::overflow_error("CUDA group slot count exceeds UINT32_MAX");
+            }
+            const uint32_t slot = static_cast<uint32_t>(next_slot);
+            pending_group_slots.emplace(group_ids[i], slot);
+            new_group_ids.push_back(group_ids[i]);
+            row_group_slots[static_cast<size_t>(i)] = slot;
+        }
+
+        ids_.reserve(static_cast<size_t>(required));
+        alive_.reserve(static_cast<size_t>(required));
+        row_group_slots_.reserve(static_cast<size_t>(required));
+        id_to_slot_.reserve(id_to_slot_.size() + static_cast<size_t>(count));
+        group_ids_.reserve(old_group_count + new_group_ids.size());
+        group_live_counts_.reserve(old_group_count + new_group_ids.size());
+        group_to_slot_.reserve(group_to_slot_.size() + new_group_ids.size());
+
+        try {
+            for (const auto &entry : pending_group_slots) {
+                group_to_slot_.emplace(entry.first, entry.second);
+            }
+            group_ids_.insert(group_ids_.end(), new_group_ids.begin(),
+                              new_group_ids.end());
+            group_live_counts_.insert(group_live_counts_.end(),
+                                      new_group_ids.size(), UINT64_C(0));
+            for (uint64_t i = 0; i < count; ++i) {
+                id_to_slot_.emplace(ids[i], size_ + i);
+            }
+        } catch (...) {
+            for (uint64_t i = 0; i < count; ++i) id_to_slot_.erase(ids[i]);
+            for (const auto &entry : pending_group_slots) {
+                group_to_slot_.erase(entry.first);
+            }
+            group_ids_.resize(old_group_count);
+            group_live_counts_.resize(old_group_count);
+            throw;
+        }
+
+        const auto rollback_metadata = [&] {
+            for (uint64_t i = 0; i < count; ++i) id_to_slot_.erase(ids[i]);
+            for (const auto &entry : pending_group_slots) {
+                group_to_slot_.erase(entry.first);
+            }
+            group_ids_.resize(old_group_count);
+            group_live_counts_.resize(old_group_count);
+        };
+
+        try {
+            const size_t elements = checked_elements(count);
+            char *destination = static_cast<char *>(data_) +
+                                static_cast<size_t>(size_) * bytes_per_vector_;
+            if (dtype_ == IFS_CUDA_FP32) {
+                check_cuda(cudaMemcpy(destination, vectors,
+                                      elements * sizeof(float),
+                                      cudaMemcpyHostToDevice),
+                           "copy FP32 database rows to GPU");
+            } else {
+                ensure_staging(elements);
+                check_cuda(cudaMemcpy(staging_, vectors,
+                                      elements * sizeof(float),
+                                      cudaMemcpyHostToDevice),
+                           "copy conversion staging rows to GPU");
+                convert(staging_, destination, elements);
+            }
+            check_cuda(cudaMemcpy(device_ids_ + size_, ids,
+                                  static_cast<size_t>(count) * sizeof(uint64_t),
+                                  cudaMemcpyHostToDevice),
+                       "copy external IDs to GPU");
+            check_cuda(cudaMemcpy(device_row_group_slots_ + size_,
+                                  row_group_slots.data(),
+                                  static_cast<size_t>(count) * sizeof(uint32_t),
+                                  cudaMemcpyHostToDevice),
+                       "copy row group slots to GPU");
+            if (!new_group_ids.empty()) {
+                check_cuda(cudaMemcpy(device_group_ids_ + old_group_count,
+                                      new_group_ids.data(),
+                                      new_group_ids.size() * sizeof(uint64_t),
+                                      cudaMemcpyHostToDevice),
+                           "copy stable group IDs to GPU");
+            }
+            check_cuda(cudaMemset(device_alive_ + size_, 1,
+                                  static_cast<size_t>(count)),
+                       "initialize live-row flags on GPU");
+            check_cuda(cudaDeviceSynchronize(), "synchronize database add");
+        } catch (...) {
+            rollback_metadata();
+            throw;
+        }
+
+        for (uint64_t i = 0; i < count; ++i) {
+            ids_.push_back(ids[i]);
+            alive_.push_back(1);
+            const uint32_t group_slot = row_group_slots[static_cast<size_t>(i)];
+            row_group_slots_.push_back(group_slot);
+            if (group_live_counts_[group_slot]++ == 0) ++active_group_count_;
+        }
+        size_ = required;
+        live_size_ += count;
+    }
+
+    uint64_t remove(const uint64_t *ids, uint64_t count) {
+        if (count && !ids) throw std::invalid_argument("remove received a null pointer");
+        std::lock_guard<std::mutex> guard(mutex_);
+        set_device();
+        std::vector<uint64_t> removed_slots;
+        removed_slots.reserve(static_cast<size_t>(count));
+        std::unordered_set<uint64_t> scheduled_slots;
+        scheduled_slots.reserve(static_cast<size_t>(count));
+        for (uint64_t i = 0; i < count; ++i) {
+            const auto found = id_to_slot_.find(ids[i]);
+            if (found == id_to_slot_.end()) continue;
+            const uint64_t slot = found->second;
+            if (alive_[static_cast<size_t>(slot)] &&
+                scheduled_slots.insert(slot).second) {
+                const uint32_t group_slot =
+                    row_group_slots_[static_cast<size_t>(slot)];
+                if (group_live_counts_[group_slot] == 0) {
+                    throw std::logic_error("CUDA group live count underflow");
+                }
+                removed_slots.push_back(slot);
+            }
+        }
+        if (!removed_slots.empty()) {
+            check_cuda(cudaMemcpy(device_delete_slots_, removed_slots.data(),
+                                  removed_slots.size() * sizeof(uint64_t),
+                                  cudaMemcpyHostToDevice),
+                       "copy tombstone update slots");
+            constexpr int threads = 256;
+            const int blocks = static_cast<int>(
+                (removed_slots.size() + threads - 1) / threads);
+            mark_dead<<<blocks, threads>>>(device_alive_, device_delete_slots_,
+                                           removed_slots.size());
+            check_cuda(cudaGetLastError(), "launch tombstone update kernel");
+            check_cuda(cudaDeviceSynchronize(),
+                       "synchronize tombstone update kernel");
+
+            for (const uint64_t slot : removed_slots) {
+                alive_[static_cast<size_t>(slot)] = 0;
+                id_to_slot_.erase(ids_[static_cast<size_t>(slot)]);
+                const uint32_t group_slot =
+                    row_group_slots_[static_cast<size_t>(slot)];
+                if (--group_live_counts_[group_slot] == 0) {
+                    --active_group_count_;
+                }
+            }
+            live_size_ -= static_cast<uint64_t>(removed_slots.size());
+        }
+        return static_cast<uint64_t>(removed_slots.size());
+    }
+
+    void search(const float *query, uint64_t k, int topk_mode,
+                uint64_t *out_ids,
+                float *out_scores, uint64_t *out_count, double *kernel_ms,
+                double *topk_ms, double *total_ms) {
+        if (!query || !out_count) throw std::invalid_argument("search received a null pointer");
+        if (k && (!out_ids || !out_scores)) {
+            throw std::invalid_argument("search output buffers are null");
+        }
+        if (topk_mode != IFS_CUDA_TOPK_HOST &&
+            topk_mode != IFS_CUDA_TOPK_DEVICE) {
+            throw std::invalid_argument("unsupported CUDA Top-K mode");
+        }
+        if (topk_mode == IFS_CUDA_TOPK_DEVICE && k > kDeviceTopKLimit) {
+            throw std::invalid_argument("device Top-K supports k <= 100");
+        }
+        std::lock_guard<std::mutex> guard(mutex_);
+        set_device();
+        const auto total_start = std::chrono::steady_clock::now();
+        if (k == 0 || live_size_ == 0) {
+            *out_count = 0;
+            if (kernel_ms) *kernel_ms = 0.0;
+            if (topk_ms) *topk_ms = 0.0;
+            if (total_ms) *total_ms = 0.0;
+            return;
+        }
+        compute_scores(query, kernel_ms);
+
+        if (topk_mode == IFS_CUDA_TOPK_DEVICE) {
+            if (dtype_ == IFS_CUDA_INT8) {
+                select_device_topk<int32_t>(k, out_ids, out_scores, out_count,
+                                            topk_ms);
+            } else {
+                select_device_topk<float>(k, out_ids, out_scores, out_count,
+                                          topk_ms);
+            }
+        } else {
+            const auto selection_start = std::chrono::steady_clock::now();
+            select_host_topk(k, out_ids, out_scores, out_count);
+            if (topk_ms) {
+                const auto selection_end = std::chrono::steady_clock::now();
+                *topk_ms = std::chrono::duration<double, std::milli>(
+                    selection_end - selection_start).count();
+            }
+        }
+        if (total_ms) {
+            const auto total_end = std::chrono::steady_clock::now();
+            *total_ms = std::chrono::duration<double, std::milli>(
+                            total_end - total_start).count();
+        }
+    }
+
+    void grouped_search(const float *query, uint64_t k,
+                        uint64_t *out_group_ids, uint64_t *out_vector_ids,
+                        float *out_scores, uint64_t *out_count,
+                        double *kernel_ms, double *topk_ms,
+                        double *total_ms) {
+        if (!query || !out_count) {
+            throw std::invalid_argument("grouped search received a null pointer");
+        }
+        if (k && (!out_group_ids || !out_vector_ids || !out_scores)) {
+            throw std::invalid_argument("grouped search output buffers are null");
+        }
+        if (k > kDeviceTopKLimit) {
+            throw std::invalid_argument("grouped device Top-K supports k <= 100");
+        }
+        std::lock_guard<std::mutex> guard(mutex_);
+        set_device();
+        const auto total_start = std::chrono::steady_clock::now();
+        if (k == 0 || live_size_ == 0 || active_group_count_ == 0) {
+            *out_count = 0;
+            if (kernel_ms) *kernel_ms = 0.0;
+            if (topk_ms) *topk_ms = 0.0;
+            if (total_ms) *total_ms = 0.0;
+            return;
+        }
+        compute_scores(query, kernel_ms);
+        if (dtype_ == IFS_CUDA_INT8) {
+            select_device_grouped_topk<int32_t>(
+                k, out_group_ids, out_vector_ids, out_scores, out_count,
+                topk_ms);
+        } else {
+            select_device_grouped_topk<float>(
+                k, out_group_ids, out_vector_ids, out_scores, out_count,
+                topk_ms);
+        }
+        if (total_ms) {
+            const auto total_end = std::chrono::steady_clock::now();
+            *total_ms = std::chrono::duration<double, std::milli>(
+                            total_end - total_start).count();
+        }
+    }
+
+    ifs_cuda_stats stats() const {
+        std::lock_guard<std::mutex> guard(mutex_);
+        return ifs_cuda_stats{size_, live_size_, capacity_, reallocations_,
+                              bytes_per_vector_, device_, dtype_};
+    }
+
+private:
+    void set_device() const { check_cuda(cudaSetDevice(device_), "cudaSetDevice"); }
+
+    size_t checked_elements(uint64_t rows) const {
+        if (rows > std::numeric_limits<size_t>::max() / kDimension) {
+            throw std::overflow_error("vector element count overflow");
+        }
+        return static_cast<size_t>(rows) * kDimension;
+    }
+
+    void compute_scores(const float *query, double *kernel_ms) {
+        if (size_ > static_cast<uint64_t>(std::numeric_limits<int>::max())) {
+            throw std::overflow_error(
+                "cuBLAS flat search supports at most INT_MAX rows");
+        }
+        upload_query(query);
+        ensure_scores(size_);
+
+        cudaEvent_t begin = nullptr;
+        cudaEvent_t end = nullptr;
+        check_cuda(cudaEventCreate(&begin), "create CUDA timing event");
+        try {
+            check_cuda(cudaEventCreate(&end), "create CUDA timing event");
+            check_cuda(cudaEventRecord(begin), "record CUDA timing start");
+            run_gemm();
+            check_cuda(cudaEventRecord(end), "record CUDA timing end");
+            check_cuda(cudaEventSynchronize(end),
+                       "synchronize CUDA dot product");
+            float elapsed = 0.0f;
+            check_cuda(cudaEventElapsedTime(&elapsed, begin, end),
+                       "measure CUDA dot product");
+            if (kernel_ms) *kernel_ms = static_cast<double>(elapsed);
+        } catch (...) {
+            if (begin) cudaEventDestroy(begin);
+            if (end) cudaEventDestroy(end);
+            throw;
+        }
+        cudaEventDestroy(begin);
+        cudaEventDestroy(end);
+    }
+
+    void reserve_impl(uint64_t rows, bool automatic) {
+        if (rows <= capacity_) return;
+        if (rows > std::numeric_limits<size_t>::max() / bytes_per_vector_) {
+            throw std::overflow_error("GPU index allocation size overflow");
+        }
+        if (rows > std::numeric_limits<size_t>::max() / sizeof(uint64_t)) {
+            throw std::overflow_error("GPU index metadata allocation overflow");
+        }
+        void *next = nullptr;
+        uint64_t *next_ids = nullptr;
+        uint8_t *next_alive = nullptr;
+        uint32_t *next_row_group_slots = nullptr;
+        uint64_t *next_group_ids = nullptr;
+        uint32_t *next_group_best_keys = nullptr;
+        uint64_t *next_group_best_vector_ids = nullptr;
+        uint64_t *next_delete_slots = nullptr;
+        void *next_scores = nullptr;
+        void *next_topk_first = nullptr;
+        void *next_topk_second = nullptr;
+        void *next_group_topk_first = nullptr;
+        void *next_group_topk_second = nullptr;
+        const uint64_t topk_blocks = rows / kTopKChunkRows +
+            (rows % kTopKChunkRows != 0 ? 1 : 0);
+        if (topk_blocks > std::numeric_limits<uint64_t>::max() /
+                              kDeviceTopKLimit) {
+            throw std::overflow_error("GPU Top-K reserve size overflow");
+        }
+        const uint64_t topk_entries = topk_blocks * kDeviceTopKLimit;
+        if (topk_entries > std::numeric_limits<size_t>::max() /
+                               sizeof(DeviceGroupCandidate<float>)) {
+            throw std::overflow_error("GPU Top-K reserve byte size overflow");
+        }
+        check_cuda(cudaMalloc(&next, static_cast<size_t>(rows) * bytes_per_vector_),
+                   "allocate GPU index storage");
+        try {
+            check_cuda(cudaMalloc(&next_ids,
+                                  static_cast<size_t>(rows) * sizeof(uint64_t)),
+                       "allocate GPU external IDs");
+            check_cuda(cudaMalloc(&next_alive, static_cast<size_t>(rows)),
+                       "allocate GPU live-row flags");
+            check_cuda(cudaMalloc(&next_row_group_slots,
+                                  static_cast<size_t>(rows) * sizeof(uint32_t)),
+                       "allocate GPU row group slots");
+            check_cuda(cudaMalloc(&next_group_ids,
+                                  static_cast<size_t>(rows) * sizeof(uint64_t)),
+                       "allocate GPU stable group IDs");
+            check_cuda(cudaMalloc(&next_group_best_keys,
+                                  static_cast<size_t>(rows) * sizeof(uint32_t)),
+                       "allocate GPU group score reduction storage");
+            check_cuda(cudaMalloc(&next_group_best_vector_ids,
+                                  static_cast<size_t>(rows) * sizeof(uint64_t)),
+                       "allocate GPU group vector reduction storage");
+            check_cuda(cudaMalloc(&next_delete_slots,
+                                  static_cast<size_t>(rows) * sizeof(uint64_t)),
+                       "allocate GPU tombstone update slots");
+            const size_t score_bytes = dtype_ == IFS_CUDA_INT8
+                ? sizeof(int32_t) : sizeof(float);
+            check_cuda(cudaMalloc(&next_scores,
+                                  static_cast<size_t>(rows) * score_bytes),
+                       "allocate GPU score workspace");
+            const size_t face_topk_bytes = static_cast<size_t>(topk_entries) *
+                                           sizeof(DeviceCandidate<float>);
+            const size_t group_topk_bytes = static_cast<size_t>(topk_entries) *
+                                            sizeof(DeviceGroupCandidate<float>);
+            check_cuda(cudaMalloc(&next_topk_first, face_topk_bytes),
+                       "allocate first reserved GPU Top-K workspace");
+            check_cuda(cudaMalloc(&next_topk_second, face_topk_bytes),
+                       "allocate second reserved GPU Top-K workspace");
+            check_cuda(cudaMalloc(&next_group_topk_first, group_topk_bytes),
+                       "allocate first reserved GPU grouped Top-K workspace");
+            check_cuda(cudaMalloc(&next_group_topk_second, group_topk_bytes),
+                       "allocate second reserved GPU grouped Top-K workspace");
+            if (data_ && size_ > 0) {
+                check_cuda(cudaMemcpy(next, data_,
+                                      static_cast<size_t>(size_) * bytes_per_vector_,
+                                      cudaMemcpyDeviceToDevice),
+                           "copy GPU index during growth");
+                check_cuda(cudaMemcpy(next_ids, device_ids_,
+                                      static_cast<size_t>(size_) * sizeof(uint64_t),
+                                      cudaMemcpyDeviceToDevice),
+                           "copy GPU external IDs during growth");
+                check_cuda(cudaMemcpy(next_alive, device_alive_,
+                                      static_cast<size_t>(size_),
+                                      cudaMemcpyDeviceToDevice),
+                           "copy GPU live-row flags during growth");
+                check_cuda(cudaMemcpy(
+                               next_row_group_slots, device_row_group_slots_,
+                               static_cast<size_t>(size_) * sizeof(uint32_t),
+                               cudaMemcpyDeviceToDevice),
+                           "copy GPU row group slots during growth");
+            }
+            if (device_group_ids_ && !group_ids_.empty()) {
+                check_cuda(cudaMemcpy(
+                               next_group_ids, device_group_ids_,
+                               group_ids_.size() * sizeof(uint64_t),
+                               cudaMemcpyDeviceToDevice),
+                           "copy GPU stable group IDs during growth");
+            }
+        } catch (...) {
+            cudaFree(next);
+            if (next_ids) cudaFree(next_ids);
+            if (next_alive) cudaFree(next_alive);
+            if (next_row_group_slots) cudaFree(next_row_group_slots);
+            if (next_group_ids) cudaFree(next_group_ids);
+            if (next_group_best_keys) cudaFree(next_group_best_keys);
+            if (next_group_best_vector_ids) {
+                cudaFree(next_group_best_vector_ids);
+            }
+            if (next_delete_slots) cudaFree(next_delete_slots);
+            if (next_scores) cudaFree(next_scores);
+            if (next_topk_first) cudaFree(next_topk_first);
+            if (next_topk_second) cudaFree(next_topk_second);
+            if (next_group_topk_first) cudaFree(next_group_topk_first);
+            if (next_group_topk_second) cudaFree(next_group_topk_second);
+            throw;
+        }
+        if (data_) check_cuda(cudaFree(data_), "free previous GPU index storage");
+        if (device_ids_) {
+            check_cuda(cudaFree(device_ids_), "free previous GPU external IDs");
+        }
+        if (device_alive_) {
+            check_cuda(cudaFree(device_alive_), "free previous GPU live-row flags");
+        }
+        if (device_row_group_slots_) {
+            check_cuda(cudaFree(device_row_group_slots_),
+                       "free previous GPU row group slots");
+        }
+        if (device_group_ids_) {
+            check_cuda(cudaFree(device_group_ids_),
+                       "free previous GPU stable group IDs");
+        }
+        if (group_best_keys_) {
+            check_cuda(cudaFree(group_best_keys_),
+                       "free previous GPU group score reduction storage");
+        }
+        if (group_best_vector_ids_) {
+            check_cuda(cudaFree(group_best_vector_ids_),
+                       "free previous GPU group vector reduction storage");
+        }
+        if (device_delete_slots_) {
+            check_cuda(cudaFree(device_delete_slots_),
+                       "free previous GPU tombstone update slots");
+        }
+        if (scores_) {
+            check_cuda(cudaFree(scores_), "free previous GPU score workspace");
+        }
+        if (topk_first_) {
+            check_cuda(cudaFree(topk_first_),
+                       "free first previous GPU Top-K workspace");
+        }
+        if (topk_second_) {
+            check_cuda(cudaFree(topk_second_),
+                       "free second previous GPU Top-K workspace");
+        }
+        if (group_topk_first_) {
+            check_cuda(cudaFree(group_topk_first_),
+                       "free first previous GPU grouped Top-K workspace");
+        }
+        if (group_topk_second_) {
+            check_cuda(cudaFree(group_topk_second_),
+                       "free second previous GPU grouped Top-K workspace");
+        }
+        data_ = next;
+        device_ids_ = next_ids;
+        device_alive_ = next_alive;
+        device_row_group_slots_ = next_row_group_slots;
+        device_group_ids_ = next_group_ids;
+        group_best_keys_ = next_group_best_keys;
+        group_best_vector_ids_ = next_group_best_vector_ids;
+        device_delete_slots_ = next_delete_slots;
+        scores_ = next_scores;
+        scores_capacity_ = rows;
+        topk_first_ = next_topk_first;
+        topk_second_ = next_topk_second;
+        topk_capacity_entries_ = topk_entries;
+        group_topk_first_ = next_group_topk_first;
+        group_topk_second_ = next_group_topk_second;
+        group_topk_capacity_entries_ = topk_entries;
+        capacity_ = rows;
+        if (automatic) ++reallocations_;
+    }
+
+    void allocate_query() {
+        check_cuda(cudaMalloc(&query_, static_cast<size_t>(kDimension) * element_bytes(dtype_)),
+                   "allocate GPU query");
+    }
+
+    void ensure_staging(size_t elements) {
+        if (elements <= staging_elements_) return;
+        if (staging_) check_cuda(cudaFree(staging_), "free old conversion staging");
+        check_cuda(cudaMalloc(&staging_, elements * sizeof(float)),
+                   "allocate conversion staging");
+        staging_elements_ = elements;
+    }
+
+    void ensure_scores(uint64_t rows) {
+        if (rows <= scores_capacity_) return;
+        if (scores_) check_cuda(cudaFree(scores_), "free old score buffer");
+        const size_t score_bytes = dtype_ == IFS_CUDA_INT8 ? sizeof(int32_t) : sizeof(float);
+        check_cuda(cudaMalloc(&scores_, static_cast<size_t>(rows) * score_bytes),
+                   "allocate GPU score buffer");
+        scores_capacity_ = rows;
+    }
+
+    void convert(const float *input, void *output, size_t elements) {
+        constexpr int threads = 256;
+        const int blocks = static_cast<int>((elements + threads - 1) / threads);
+        switch (dtype_) {
+            case IFS_CUDA_FP16:
+                float_to_half<<<blocks, threads>>>(input, static_cast<__half *>(output), elements);
+                break;
+            case IFS_CUDA_BF16:
+                float_to_bfloat16<<<blocks, threads>>>(
+                    input, static_cast<__nv_bfloat16 *>(output), elements);
+                break;
+            case IFS_CUDA_INT8:
+                float_to_int8<<<blocks, threads>>>(
+                    input, static_cast<int8_t *>(output), elements,
+                    static_cast<float>(int8_scale_));
+                break;
+            default:
+                throw std::logic_error("conversion requested for FP32 index");
+        }
+        check_cuda(cudaGetLastError(), "launch storage conversion kernel");
+    }
+
+    void upload_query(const float *query) {
+        if (dtype_ == IFS_CUDA_FP32) {
+            check_cuda(cudaMemcpy(query_, query, kDimension * sizeof(float),
+                                  cudaMemcpyHostToDevice),
+                       "copy FP32 query to GPU");
+            return;
+        }
+        ensure_staging(kDimension);
+        check_cuda(cudaMemcpy(staging_, query, kDimension * sizeof(float),
+                              cudaMemcpyHostToDevice),
+                   "copy query conversion staging to GPU");
+        convert(staging_, query_, kDimension);
+    }
+
+    void run_gemm() {
+        const int rows = static_cast<int>(size_);
+        const int dimension = kDimension;
+        if (dtype_ == IFS_CUDA_INT8) {
+            const int32_t alpha = 1;
+            const int32_t beta = 0;
+            check_cublas(cublasGemmEx(
+                cublas_, CUBLAS_OP_T, CUBLAS_OP_N, rows, 1, dimension,
+                &alpha, data_, CUDA_R_8I, dimension, query_, CUDA_R_8I, dimension,
+                &beta, scores_, CUDA_R_32I, rows, CUBLAS_COMPUTE_32I,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP),
+                "INT8 exact flat inner-product GEMM");
+            return;
+        }
+        const float alpha = 1.0f;
+        const float beta = 0.0f;
+        cudaDataType_t storage_type = CUDA_R_32F;
+        if (dtype_ == IFS_CUDA_FP16) storage_type = CUDA_R_16F;
+        if (dtype_ == IFS_CUDA_BF16) storage_type = CUDA_R_16BF;
+        check_cublas(cublasGemmEx(
+            cublas_, CUBLAS_OP_T, CUBLAS_OP_N, rows, 1, dimension,
+            &alpha, data_, storage_type, dimension, query_, storage_type, dimension,
+            &beta, scores_, CUDA_R_32F, rows, CUBLAS_COMPUTE_32F,
+            CUBLAS_GEMM_DEFAULT_TENSOR_OP),
+            "floating-point exact flat inner-product GEMM");
+    }
+
+    void select_host_topk(uint64_t k, uint64_t *out_ids,
+                          float *out_scores, uint64_t *out_count) {
+        const size_t row_count = static_cast<size_t>(size_);
+        std::priority_queue<Candidate, std::vector<Candidate>, BetterComparator> heap;
+        if (dtype_ == IFS_CUDA_INT8) {
+            host_int_scores_.resize(row_count);
+            check_cuda(cudaMemcpy(host_int_scores_.data(), scores_,
+                                  row_count * sizeof(int32_t),
+                                  cudaMemcpyDeviceToHost),
+                       "copy INT8 dot scores to host");
+            for (size_t slot = 0; slot < row_count; ++slot) {
+                if (!alive_[slot]) continue;
+                consider(heap, k, Candidate{
+                    static_cast<float>(host_int_scores_[slot]) /
+                        int8_score_divisor_,
+                    ids_[slot]});
+            }
+        } else {
+            host_float_scores_.resize(row_count);
+            check_cuda(cudaMemcpy(host_float_scores_.data(), scores_,
+                                  row_count * sizeof(float),
+                                  cudaMemcpyDeviceToHost),
+                       "copy floating-point dot scores to host");
+            for (size_t slot = 0; slot < row_count; ++slot) {
+                if (!alive_[slot]) continue;
+                consider(heap, k,
+                         Candidate{host_float_scores_[slot], ids_[slot]});
+            }
+        }
+        std::vector<Candidate> ordered;
+        ordered.reserve(heap.size());
+        while (!heap.empty()) {
+            ordered.push_back(heap.top());
+            heap.pop();
+        }
+        std::sort(ordered.begin(), ordered.end(), better);
+        for (size_t i = 0; i < ordered.size(); ++i) {
+            out_ids[i] = ordered[i].id;
+            out_scores[i] = ordered[i].score;
+        }
+        *out_count = static_cast<uint64_t>(ordered.size());
+    }
+
+    void ensure_topk_buffers(uint64_t entries) {
+        if (entries <= topk_capacity_entries_) return;
+        if (entries > std::numeric_limits<size_t>::max() /
+                          sizeof(DeviceCandidate<float>)) {
+            throw std::overflow_error("GPU Top-K candidate allocation overflow");
+        }
+        void *first = nullptr;
+        void *second = nullptr;
+        const size_t bytes = static_cast<size_t>(entries) *
+                             sizeof(DeviceCandidate<float>);
+        check_cuda(cudaMalloc(&first, bytes),
+                   "allocate first GPU Top-K candidate buffer");
+        try {
+            check_cuda(cudaMalloc(&second, bytes),
+                       "allocate second GPU Top-K candidate buffer");
+        } catch (...) {
+            cudaFree(first);
+            throw;
+        }
+        if (topk_first_) check_cuda(cudaFree(topk_first_),
+                                    "free first GPU Top-K candidate buffer");
+        if (topk_second_) check_cuda(cudaFree(topk_second_),
+                                     "free second GPU Top-K candidate buffer");
+        topk_first_ = first;
+        topk_second_ = second;
+        topk_capacity_entries_ = entries;
+    }
+
+    template <typename Score>
+    void select_device_topk(uint64_t k, uint64_t *out_ids,
+                            float *out_scores, uint64_t *out_count,
+                            double *topk_ms) {
+        static_assert(sizeof(DeviceCandidate<Score>) ==
+                      sizeof(DeviceCandidate<float>),
+                      "all device candidate layouts must have equal size");
+        const uint64_t first_blocks =
+            (size_ + kTopKChunkRows - 1) / kTopKChunkRows;
+        if (first_blocks > std::numeric_limits<unsigned int>::max()) {
+            throw std::overflow_error("GPU Top-K grid dimension overflow");
+        }
+        if (first_blocks > std::numeric_limits<uint64_t>::max() / k) {
+            throw std::overflow_error("GPU Top-K candidate count overflow");
+        }
+        ensure_topk_buffers(first_blocks * k);
+        auto *first = static_cast<DeviceCandidate<Score> *>(topk_first_);
+        auto *second = static_cast<DeviceCandidate<Score> *>(topk_second_);
+
+        cudaEvent_t begin = nullptr;
+        cudaEvent_t end = nullptr;
+        check_cuda(cudaEventCreate(&begin), "create GPU Top-K timing event");
+        try {
+            check_cuda(cudaEventCreate(&end), "create GPU Top-K timing event");
+            check_cuda(cudaEventRecord(begin), "record GPU Top-K start");
+            topk_from_scores<Score><<<static_cast<unsigned int>(first_blocks),
+                                      kTopKBlockSize>>>(
+                static_cast<const Score *>(scores_), device_ids_, device_alive_,
+                size_, k, first);
+            check_cuda(cudaGetLastError(), "launch first GPU Top-K stage");
+
+            uint64_t candidate_count = first_blocks * k;
+            DeviceCandidate<Score> *current = first;
+            DeviceCandidate<Score> *next = second;
+            while (candidate_count > k) {
+                const uint64_t blocks =
+                    (candidate_count + kTopKChunkRows - 1) /
+                    kTopKChunkRows;
+                if (blocks > std::numeric_limits<unsigned int>::max()) {
+                    throw std::overflow_error("GPU Top-K reduction grid overflow");
+                }
+                topk_from_candidates<Score><<<static_cast<unsigned int>(blocks),
+                                              kTopKBlockSize>>>(
+                    current, candidate_count, k, next);
+                check_cuda(cudaGetLastError(),
+                           "launch GPU Top-K reduction stage");
+                candidate_count = blocks * k;
+                std::swap(current, next);
+            }
+            check_cuda(cudaEventRecord(end), "record GPU Top-K end");
+            check_cuda(cudaEventSynchronize(end),
+                       "synchronize GPU Top-K selection");
+            float elapsed = 0.0f;
+            check_cuda(cudaEventElapsedTime(&elapsed, begin, end),
+                       "measure GPU Top-K selection");
+            if (topk_ms) *topk_ms = static_cast<double>(elapsed);
+
+            const uint64_t result_count = std::min(k, live_size_);
+            std::vector<DeviceCandidate<Score>> selected(
+                static_cast<size_t>(result_count));
+            check_cuda(cudaMemcpy(selected.data(), current,
+                                  static_cast<size_t>(result_count) *
+                                      sizeof(DeviceCandidate<Score>),
+                                  cudaMemcpyDeviceToHost),
+                       "copy final GPU Top-K candidates to host");
+            for (size_t i = 0; i < selected.size(); ++i) {
+                out_ids[i] = selected[i].id;
+                if constexpr (std::is_same<Score, int32_t>::value) {
+                    out_scores[i] =
+                        static_cast<float>(selected[i].score) /
+                        int8_score_divisor_;
+                } else {
+                    out_scores[i] = selected[i].score;
+                }
+            }
+            *out_count = result_count;
+        } catch (...) {
+            if (begin) cudaEventDestroy(begin);
+            if (end) cudaEventDestroy(end);
+            throw;
+        }
+        cudaEventDestroy(begin);
+        cudaEventDestroy(end);
+    }
+
+    void ensure_group_topk_buffers(uint64_t entries) {
+        if (entries <= group_topk_capacity_entries_) return;
+        if (entries > std::numeric_limits<size_t>::max() /
+                          sizeof(DeviceGroupCandidate<float>)) {
+            throw std::overflow_error(
+                "GPU grouped Top-K candidate allocation overflow");
+        }
+        void *first = nullptr;
+        void *second = nullptr;
+        const size_t bytes = static_cast<size_t>(entries) *
+                             sizeof(DeviceGroupCandidate<float>);
+        check_cuda(cudaMalloc(&first, bytes),
+                   "allocate first GPU grouped Top-K candidate buffer");
+        try {
+            check_cuda(cudaMalloc(&second, bytes),
+                       "allocate second GPU grouped Top-K candidate buffer");
+        } catch (...) {
+            cudaFree(first);
+            throw;
+        }
+        if (group_topk_first_) {
+            check_cuda(cudaFree(group_topk_first_),
+                       "free first GPU grouped Top-K candidate buffer");
+        }
+        if (group_topk_second_) {
+            check_cuda(cudaFree(group_topk_second_),
+                       "free second GPU grouped Top-K candidate buffer");
+        }
+        group_topk_first_ = first;
+        group_topk_second_ = second;
+        group_topk_capacity_entries_ = entries;
+    }
+
+    template <typename Score>
+    void select_device_grouped_topk(
+        uint64_t k, uint64_t *out_group_ids, uint64_t *out_vector_ids,
+        float *out_scores, uint64_t *out_count, double *topk_ms) {
+        static_assert(sizeof(DeviceGroupCandidate<Score>) ==
+                          sizeof(DeviceGroupCandidate<float>),
+                      "all grouped candidate layouts must have equal size");
+        const uint64_t group_count = static_cast<uint64_t>(group_ids_.size());
+        const uint64_t first_blocks =
+            (group_count + kTopKChunkRows - 1) / kTopKChunkRows;
+        if (first_blocks > std::numeric_limits<unsigned int>::max()) {
+            throw std::overflow_error(
+                "GPU grouped Top-K grid dimension overflow");
+        }
+        if (first_blocks > std::numeric_limits<uint64_t>::max() / k) {
+            throw std::overflow_error(
+                "GPU grouped Top-K candidate count overflow");
+        }
+        ensure_group_topk_buffers(first_blocks * k);
+        auto *first = static_cast<DeviceGroupCandidate<Score> *>(
+            group_topk_first_);
+        auto *second = static_cast<DeviceGroupCandidate<Score> *>(
+            group_topk_second_);
+
+        cudaEvent_t begin = nullptr;
+        cudaEvent_t end = nullptr;
+        check_cuda(cudaEventCreate(&begin),
+                   "create GPU grouped Top-K timing event");
+        try {
+            check_cuda(cudaEventCreate(&end),
+                       "create GPU grouped Top-K timing event");
+            check_cuda(cudaEventRecord(begin),
+                       "record GPU grouped Top-K start");
+            check_cuda(cudaMemset(group_best_keys_, 0,
+                                  static_cast<size_t>(group_count) *
+                                      sizeof(uint32_t)),
+                       "reset GPU group best scores");
+            check_cuda(cudaMemset(group_best_vector_ids_, 0xff,
+                                  static_cast<size_t>(group_count) *
+                                      sizeof(uint64_t)),
+                       "reset GPU group best vector IDs");
+
+            constexpr int reduction_threads = 256;
+            const uint64_t reduction_blocks =
+                (size_ + reduction_threads - 1) / reduction_threads;
+            if (reduction_blocks > std::numeric_limits<unsigned int>::max()) {
+                throw std::overflow_error(
+                    "GPU group reduction grid dimension overflow");
+            }
+            reduce_group_best_scores<Score>
+                <<<static_cast<unsigned int>(reduction_blocks),
+                   reduction_threads>>>(
+                    static_cast<const Score *>(scores_),
+                    device_row_group_slots_, device_alive_, size_,
+                    group_best_keys_);
+            check_cuda(cudaGetLastError(),
+                       "launch GPU group score reduction");
+            reduce_group_best_vector_ids<Score>
+                <<<static_cast<unsigned int>(reduction_blocks),
+                   reduction_threads>>>(
+                    static_cast<const Score *>(scores_), device_ids_,
+                    device_row_group_slots_, device_alive_, size_,
+                    group_best_keys_, group_best_vector_ids_);
+            check_cuda(cudaGetLastError(),
+                       "launch GPU group vector reduction");
+
+            grouped_topk_from_best<Score>
+                <<<static_cast<unsigned int>(first_blocks),
+                   kTopKBlockSize>>>(
+                    group_best_keys_, group_best_vector_ids_,
+                    device_group_ids_, group_count, k, first);
+            check_cuda(cudaGetLastError(),
+                       "launch first GPU grouped Top-K stage");
+
+            uint64_t candidate_count = first_blocks * k;
+            DeviceGroupCandidate<Score> *current = first;
+            DeviceGroupCandidate<Score> *next = second;
+            while (candidate_count > k) {
+                const uint64_t blocks =
+                    (candidate_count + kTopKChunkRows - 1) /
+                    kTopKChunkRows;
+                if (blocks > std::numeric_limits<unsigned int>::max()) {
+                    throw std::overflow_error(
+                        "GPU grouped Top-K reduction grid overflow");
+                }
+                grouped_topk_from_candidates<Score>
+                    <<<static_cast<unsigned int>(blocks),
+                       kTopKBlockSize>>>(current, candidate_count, k, next);
+                check_cuda(cudaGetLastError(),
+                           "launch GPU grouped Top-K reduction stage");
+                candidate_count = blocks * k;
+                std::swap(current, next);
+            }
+            check_cuda(cudaEventRecord(end),
+                       "record GPU grouped Top-K end");
+            check_cuda(cudaEventSynchronize(end),
+                       "synchronize GPU grouped Top-K selection");
+            float elapsed = 0.0f;
+            check_cuda(cudaEventElapsedTime(&elapsed, begin, end),
+                       "measure GPU grouped Top-K selection");
+            if (topk_ms) *topk_ms = static_cast<double>(elapsed);
+
+            const uint64_t result_count = std::min(k, active_group_count_);
+            std::vector<DeviceGroupCandidate<Score>> selected(
+                static_cast<size_t>(result_count));
+            check_cuda(cudaMemcpy(
+                           selected.data(), current,
+                           static_cast<size_t>(result_count) *
+                               sizeof(DeviceGroupCandidate<Score>),
+                           cudaMemcpyDeviceToHost),
+                       "copy final GPU grouped Top-K candidates to host");
+            for (size_t i = 0; i < selected.size(); ++i) {
+                out_group_ids[i] = selected[i].group_id;
+                out_vector_ids[i] = selected[i].vector_id;
+                if constexpr (std::is_same<Score, int32_t>::value) {
+                    out_scores[i] =
+                        static_cast<float>(selected[i].score) /
+                        int8_score_divisor_;
+                } else {
+                    out_scores[i] = selected[i].score;
+                }
+            }
+            *out_count = result_count;
+        } catch (...) {
+            if (begin) cudaEventDestroy(begin);
+            if (end) cudaEventDestroy(end);
+            throw;
+        }
+        cudaEventDestroy(begin);
+        cudaEventDestroy(end);
+    }
+
+    static void consider(
+        std::priority_queue<Candidate, std::vector<Candidate>, BetterComparator> &heap,
+        uint64_t k, const Candidate &candidate) {
+        if (heap.size() < static_cast<size_t>(k)) {
+            heap.push(candidate);
+        } else if (better(candidate, heap.top())) {
+            heap.pop();
+            heap.push(candidate);
+        }
+    }
+
+    void cleanup() noexcept {
+        if (device_ >= 0) cudaSetDevice(device_);
+        if (group_topk_second_) cudaFree(group_topk_second_);
+        if (group_topk_first_) cudaFree(group_topk_first_);
+        if (topk_second_) cudaFree(topk_second_);
+        if (topk_first_) cudaFree(topk_first_);
+        if (group_best_vector_ids_) cudaFree(group_best_vector_ids_);
+        if (group_best_keys_) cudaFree(group_best_keys_);
+        if (device_delete_slots_) cudaFree(device_delete_slots_);
+        if (scores_) cudaFree(scores_);
+        if (staging_) cudaFree(staging_);
+        if (query_) cudaFree(query_);
+        if (device_group_ids_) cudaFree(device_group_ids_);
+        if (device_row_group_slots_) cudaFree(device_row_group_slots_);
+        if (device_alive_) cudaFree(device_alive_);
+        if (device_ids_) cudaFree(device_ids_);
+        if (data_) cudaFree(data_);
+        if (cublas_) cublasDestroy(cublas_);
+        scores_ = nullptr;
+        staging_ = nullptr;
+        query_ = nullptr;
+        data_ = nullptr;
+        device_ids_ = nullptr;
+        device_alive_ = nullptr;
+        device_row_group_slots_ = nullptr;
+        device_group_ids_ = nullptr;
+        group_best_keys_ = nullptr;
+        group_best_vector_ids_ = nullptr;
+        device_delete_slots_ = nullptr;
+        topk_first_ = nullptr;
+        topk_second_ = nullptr;
+        group_topk_first_ = nullptr;
+        group_topk_second_ = nullptr;
+        cublas_ = nullptr;
+    }
+
+    int dtype_;
+    int device_;
+    double growth_factor_;
+    uint32_t int8_scale_;
+    float int8_score_divisor_;
+    uint64_t bytes_per_vector_;
+    uint64_t size_ = 0;
+    uint64_t live_size_ = 0;
+    uint64_t capacity_ = 0;
+    uint64_t reallocations_ = 0;
+    void *data_ = nullptr;
+    uint64_t *device_ids_ = nullptr;
+    uint8_t *device_alive_ = nullptr;
+    uint32_t *device_row_group_slots_ = nullptr;
+    uint64_t *device_group_ids_ = nullptr;
+    uint32_t *group_best_keys_ = nullptr;
+    uint64_t *group_best_vector_ids_ = nullptr;
+    uint64_t *device_delete_slots_ = nullptr;
+    void *query_ = nullptr;
+    float *staging_ = nullptr;
+    size_t staging_elements_ = 0;
+    void *scores_ = nullptr;
+    uint64_t scores_capacity_ = 0;
+    void *topk_first_ = nullptr;
+    void *topk_second_ = nullptr;
+    uint64_t topk_capacity_entries_ = 0;
+    void *group_topk_first_ = nullptr;
+    void *group_topk_second_ = nullptr;
+    uint64_t group_topk_capacity_entries_ = 0;
+    cublasHandle_t cublas_ = nullptr;
+    std::vector<uint64_t> ids_;
+    std::vector<uint8_t> alive_;
+    std::unordered_map<uint64_t, uint64_t> id_to_slot_;
+    std::vector<uint32_t> row_group_slots_;
+    std::unordered_map<uint64_t, uint32_t> group_to_slot_;
+    std::vector<uint64_t> group_ids_;
+    std::vector<uint64_t> group_live_counts_;
+    uint64_t active_group_count_ = 0;
+    std::vector<float> host_float_scores_;
+    std::vector<int32_t> host_int_scores_;
+    mutable std::mutex mutex_;
+};
+
+template <typename Function>
+int protect(Function &&function) {
+    try {
+        function();
+        g_last_error.clear();
+        return 0;
+    } catch (const std::exception &error) {
+        g_last_error = error.what();
+        return -1;
+    } catch (...) {
+        g_last_error = "unknown C++ exception";
+        return -1;
+    }
+}
+
+CudaFlatIndex &index_from(void *handle) {
+    if (!handle) throw std::invalid_argument("index handle is null");
+    return *static_cast<CudaFlatIndex *>(handle);
+}
+
+}  // namespace
+
+extern "C" {
+
+void *ifs_cuda_create(int dtype, uint64_t reserve_rows, int device,
+                      double growth_factor, uint32_t int8_scale) {
+    try {
+        std::unique_ptr<CudaFlatIndex> index(
+            new CudaFlatIndex(
+                dtype, reserve_rows, device, growth_factor, int8_scale));
+        g_last_error.clear();
+        return index.release();
+    } catch (const std::exception &error) {
+        g_last_error = error.what();
+        return nullptr;
+    } catch (...) {
+        g_last_error = "unknown C++ exception";
+        return nullptr;
+    }
+}
+
+void ifs_cuda_destroy(void *handle) {
+    delete static_cast<CudaFlatIndex *>(handle);
+}
+
+int ifs_cuda_reserve(void *handle, uint64_t rows) {
+    return protect([&] { index_from(handle).reserve(rows); });
+}
+
+int ifs_cuda_add(void *handle, const float *vectors, const uint64_t *ids,
+                 const uint64_t *group_ids, uint64_t count) {
+    return protect(
+        [&] { index_from(handle).add(vectors, ids, group_ids, count); });
+}
+
+int ifs_cuda_remove(void *handle, const uint64_t *ids, uint64_t count,
+                    uint64_t *removed) {
+    return protect([&] {
+        if (!removed) throw std::invalid_argument("removed output pointer is null");
+        *removed = index_from(handle).remove(ids, count);
+    });
+}
+
+int ifs_cuda_search(void *handle, const float *query, uint64_t k,
+                    uint64_t *out_ids, float *out_scores,
+                    uint64_t *out_count, double *kernel_ms,
+                    double *total_ms) {
+    return protect([&] {
+        double ignored_topk_ms = 0.0;
+        index_from(handle).search(query, k, IFS_CUDA_TOPK_HOST, out_ids,
+                                  out_scores, out_count, kernel_ms,
+                                  &ignored_topk_ms, total_ms);
+    });
+}
+
+int ifs_cuda_search_ex(void *handle, const float *query, uint64_t k,
+                       int topk_mode, uint64_t *out_ids,
+                       float *out_scores, uint64_t *out_count,
+                       double *kernel_ms, double *topk_ms,
+                       double *total_ms) {
+    return protect([&] {
+        index_from(handle).search(query, k, topk_mode, out_ids, out_scores,
+                                  out_count, kernel_ms, topk_ms, total_ms);
+    });
+}
+
+int ifs_cuda_grouped_search(void *handle, const float *query, uint64_t k,
+                            uint64_t *out_group_ids,
+                            uint64_t *out_vector_ids, float *out_scores,
+                            uint64_t *out_count, double *kernel_ms,
+                            double *topk_ms, double *total_ms) {
+    return protect([&] {
+        index_from(handle).grouped_search(
+            query, k, out_group_ids, out_vector_ids, out_scores, out_count,
+            kernel_ms, topk_ms, total_ms);
+    });
+}
+
+int ifs_cuda_get_stats(void *handle, ifs_cuda_stats *out) {
+    return protect([&] {
+        if (!out) throw std::invalid_argument("stats output pointer is null");
+        *out = index_from(handle).stats();
+    });
+}
+
+const char *ifs_cuda_last_error(void) { return g_last_error.c_str(); }
+
+const char *ifs_cuda_build_info(void) {
+    return "ifs_cuda flat-ip d512; cuBLAS; FP32/FP16/BF16/INT8(per-index-scale); "
+           "float-accumulation except INT8 int32; exact scan; exact CUDA "
+           "face/grouped Top-K<=100; grouped device-resident";
+}
+
+}  // extern "C"
diff --git a/server/native/search/cuda/src/ifs_cuda_legacy.h b/server/native/search/cuda/src/ifs_cuda_legacy.h
new file mode 100644
index 000000000..1bf4211d3
--- /dev/null
+++ b/server/native/search/cuda/src/ifs_cuda_legacy.h
@@ -0,0 +1,96 @@
+#ifndef IFS_CUDA_H
+#define IFS_CUDA_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum ifs_cuda_dtype {
+    IFS_CUDA_FP32 = 0,
+    IFS_CUDA_FP16 = 1,
+    IFS_CUDA_BF16 = 2,
+    IFS_CUDA_INT8 = 3,
+};
+
+enum ifs_cuda_topk_mode {
+    /* Compatibility path: copy every score to the host, then select Top-K. */
+    IFS_CUDA_TOPK_HOST = 0,
+    /* Exact multi-stage CUDA Top-K; only the final K candidates cross PCIe. */
+    IFS_CUDA_TOPK_DEVICE = 1,
+};
+
+typedef struct ifs_cuda_stats {
+    uint64_t size;
+    uint64_t live_size;
+    uint64_t capacity;
+    uint64_t reallocations;
+    uint64_t bytes_per_vector;
+    int device;
+    int dtype;
+} ifs_cuda_stats;
+
+/*
+ * Exact, append-friendly, flat inner-product index for fixed 512-D vectors.
+ * Input vectors and queries are always normalized float32. Storage is selected
+ * by dtype. INT8 conversion rounds half away from zero after multiplying by
+ * the positive per-index scale, then clamps to [-128, 127]. Non-INT8 indexes
+ * require int8_scale=0.
+ */
+void *ifs_cuda_create(int dtype, uint64_t reserve_rows, int device,
+                      double growth_factor, uint32_t int8_scale);
+void ifs_cuda_destroy(void *handle);
+
+int ifs_cuda_reserve(void *handle, uint64_t rows);
+int ifs_cuda_add(void *handle, const float *vectors, const uint64_t *ids,
+                 const uint64_t *group_ids, uint64_t count);
+int ifs_cuda_remove(void *handle, const uint64_t *ids, uint64_t count,
+                    uint64_t *removed);
+
+/*
+ * Returns up to k live results, sorted by descending dot product and then
+ * ascending stable external ID. out_count may be less than k.
+ * kernel_ms measures cuBLAS dot-product work only; total_ms includes device to
+ * host score transfer and exact host Top-K selection.
+ */
+int ifs_cuda_search(void *handle, const float *query, uint64_t k,
+                    uint64_t *out_ids, float *out_scores,
+                    uint64_t *out_count, double *kernel_ms,
+                    double *total_ms);
+
+/*
+ * Extended search with a selectable Top-K implementation. topk_ms reports
+ * host selection time in HOST mode and CUDA selection time in DEVICE mode.
+ * DEVICE mode supports k in [1, 100], performs exact selection on the GPU,
+ * skips deleted rows, and preserves the same score-descending/ID-ascending
+ * ordering as the compatibility path.
+ */
+int ifs_cuda_search_ex(void *handle, const float *query, uint64_t k,
+                       int topk_mode, uint64_t *out_ids,
+                       float *out_scores, uint64_t *out_count,
+                       double *kernel_ms, double *topk_ms,
+                       double *total_ms);
+
+/*
+ * Exact GPU-resident Person Top-K. Every live row contributes to its group;
+ * each group keeps the highest score (then the smallest vector ID on a tie).
+ * Groups are returned score-descending/group-ID-ascending. k must be <= 100,
+ * and only the final k candidates are copied to host memory.
+ */
+int ifs_cuda_grouped_search(void *handle, const float *query, uint64_t k,
+                            uint64_t *out_group_ids,
+                            uint64_t *out_vector_ids, float *out_scores,
+                            uint64_t *out_count, double *kernel_ms,
+                            double *topk_ms, double *total_ms);
+
+int ifs_cuda_get_stats(void *handle, ifs_cuda_stats *out);
+const char *ifs_cuda_last_error(void);
+const char *ifs_cuda_build_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/server/native/search/cuda/tests/smoke.cpp b/server/native/search/cuda/tests/smoke.cpp
new file mode 100644
index 000000000..5452c2ceb
--- /dev/null
+++ b/server/native/search/cuda/tests/smoke.cpp
@@ -0,0 +1,553 @@
+#include "ifs_search.h"
+
+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstdint>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+namespace {
+
+using Vector = std::array<float, IFS_SEARCH_DIMENSION>;
+
+Vector constant() {
+    Vector result{};
+    result.fill(1.0f / std::sqrt(static_cast<float>(result.size())));
+    return result;
+}
+
+Vector split() {
+    Vector result = constant();
+    for (size_t i = result.size() / 2; i < result.size(); ++i) {
+        result[i] = -result[i];
+    }
+    return result;
+}
+
+Vector wave(uint64_t seed) {
+    Vector result{};
+    double squared_norm = 0.0;
+    for (size_t i = 0; i < result.size(); ++i) {
+        const double value =
+            std::sin((static_cast<double>(seed) + 1.0) *
+                     (static_cast<double>(i) + 1.0) * 0.0017) +
+            0.5 * std::cos((static_cast<double>(seed) + 3.0) *
+                           (static_cast<double>(i) + 1.0) * 0.0009);
+        result[i] = static_cast<float>(value);
+        squared_norm += value * value;
+    }
+    const float inverse_norm =
+        static_cast<float>(1.0 / std::sqrt(squared_norm));
+    for (float &value : result) value *= inverse_norm;
+    return result;
+}
+
+Vector sparse(uint64_t seed) {
+    Vector result{};
+    const size_t first = static_cast<size_t>(seed % result.size());
+    size_t second = static_cast<size_t>((seed * 131 + 17) % result.size());
+    if (second == first) second = (second + 1) % result.size();
+    result[first] = std::sqrt(0.75f);
+    result[second] = seed % 2 == 0 ? 0.5f : -0.5f;
+    return result;
+}
+
+float tolerance(uint32_t profile) {
+    if (profile == IFS_SEARCH_PROFILE_FP32_V1) return 1.0e-5f;
+    if (profile == IFS_SEARCH_PROFILE_FP16_V1) return 5.0e-3f;
+    if (profile == IFS_SEARCH_PROFILE_BF16_V1) return 2.0e-2f;
+    return 3.0e-2f;
+}
+
+void normalize(Vector *value) {
+    double squared = 0.0;
+    for (float item : *value) squared += static_cast<double>(item) * item;
+    const float inverse = 1.0f / std::sqrt(static_cast<float>(squared));
+    for (float &item : *value) item *= inverse;
+}
+
+void test_int8_numeric_contract(
+    uint32_t profile, float expected_score, float expected_group_score) {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = profile;
+    options.reserve_rows = 1;
+    options.max_rows = 1;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_DEVICE;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    Vector stored{};
+    Vector query{};
+    stored[0] = query[0] = 0.12449999898672104f;
+    const float tail = std::sqrt(
+        (1.0f - stored[0] * stored[0]) /
+        static_cast<float>(stored.size() - 1));
+    for (size_t i = 1; i < stored.size(); ++i) {
+        stored[i] = tail;
+        query[i] = i <= 255 ? tail : -tail;
+    }
+    normalize(&stored);
+    normalize(&query);
+    const uint64_t id = 1;
+    const uint64_t group = 1;
+    assert(ifs_search_add_batch(index, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+    uint64_t out_id = 0;
+    float score = 0.0f;
+    uint64_t count = 0;
+    assert(ifs_search_topk(index, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    assert(std::abs(score - expected_score) < 1.0e-6f);
+    ifs_search_destroy(index);
+
+    options.reserve_rows = 2;
+    options.max_rows = 2;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    Vector first{};
+    Vector second{};
+    for (size_t i = 0; i < first.size(); ++i) {
+        first[i] = i < 99
+            ? 0.045f
+            : std::sqrt((1.0f - 99.0f * 0.045f * 0.045f) / 413.0f);
+        second[i] = i < 100
+            ? 0.045f
+            : std::sqrt((1.0f - 100.0f * 0.045f * 0.045f) / 412.0f);
+    }
+    normalize(&first);
+    normalize(&second);
+    std::vector<float> rows;
+    rows.insert(rows.end(), first.begin(), first.end());
+    rows.insert(rows.end(), second.begin(), second.end());
+    const std::array<uint64_t, 2> ids{10, 20};
+    const std::array<uint64_t, 2> groups{1, 1};
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(), rows.data(), 2) ==
+           IFS_SEARCH_OK);
+    uint64_t out_group = 0;
+    assert(ifs_search_grouped_topk(index, first.data(), 1, &out_group, &out_id,
+                                   &score, &count, nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_group == 1 && out_id == 20);
+    assert(std::abs(score - expected_group_score) < 1.0e-6f);
+    ifs_search_destroy(index);
+}
+
+void test_int8_scales_are_per_index() {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.reserve_rows = 1;
+    options.max_rows = 1;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_DEVICE;
+    ifs_search_index_t x1000 = nullptr;
+    ifs_search_index_t x736 = nullptr;
+    options.profile = IFS_SEARCH_PROFILE_INT8_X1000_V1;
+    assert(ifs_search_create(&options, &x1000) == IFS_SEARCH_OK);
+    options.profile = IFS_SEARCH_PROFILE_INT8_X736_V1;
+    assert(ifs_search_create(&options, &x736) == IFS_SEARCH_OK);
+
+    Vector stored{};
+    Vector query{};
+    stored[0] = query[0] = 0.12449999898672104f;
+    const float tail = std::sqrt(
+        (1.0f - stored[0] * stored[0]) /
+        static_cast<float>(stored.size() - 1));
+    for (size_t i = 1; i < stored.size(); ++i) {
+        stored[i] = tail;
+        query[i] = i <= 255 ? tail : -tail;
+    }
+    normalize(&stored);
+    normalize(&query);
+    const uint64_t id = 1;
+    const uint64_t group = 1;
+    assert(ifs_search_add_batch(x1000, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+    assert(ifs_search_add_batch(x736, &id, &group, stored.data(), 1) ==
+           IFS_SEARCH_OK);
+
+    uint64_t out_id = 0;
+    uint64_t count = 0;
+    float score = 0.0f;
+    assert(ifs_search_topk(x1000, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    assert(std::abs(score - 0.013689f) < 1.0e-6f);
+    assert(ifs_search_topk(x736, query.data(), 1, &out_id, &score, &count,
+                           nullptr) == IFS_SEARCH_OK);
+    assert(count == 1 && out_id == id);
+    assert(std::abs(score - 7440.0f / 541696.0f) < 1.0e-6f);
+
+    ifs_search_destroy(x736);
+    ifs_search_destroy(x1000);
+}
+
+void test_profile(uint32_t profile) {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = profile;
+    options.reserve_rows = 6;
+    options.max_rows = 6;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_DEVICE;
+    options.growth_factor = 1.5;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    const Vector positive = constant();
+    const Vector orthogonal = split();
+    std::vector<float> rows;
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    rows.insert(rows.end(), orthogonal.begin(), orthogonal.end());
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    const std::array<uint64_t, 4> ids{10, 20, 30, 5};
+    const std::array<uint64_t, 4> groups{100, 200, 300, 100};
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(), rows.data(), 4) ==
+           IFS_SEARCH_OK);
+
+    std::array<uint64_t, 6> out_ids{};
+    std::array<float, 6> scores{};
+    uint64_t count = 0;
+    ifs_search_timings_t timings{};
+    timings.struct_size = sizeof(timings);
+    assert(ifs_search_topk(index, positive.data(), 4, out_ids.data(),
+                           scores.data(), &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 4);
+    assert(out_ids[0] == 5 && out_ids[1] == 10 && out_ids[2] == 30 &&
+           out_ids[3] == 20);
+    assert(std::abs(scores[0] - 1.0f) < tolerance(profile));
+    assert(std::abs(scores[3]) < tolerance(profile));
+
+    std::array<uint64_t, 6> out_groups{};
+    assert(ifs_search_grouped_topk(
+               index, positive.data(), 6, out_groups.data(), out_ids.data(),
+               scores.data(), &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 3);
+    assert(out_groups[0] == 100 && out_ids[0] == 5);
+    assert(out_groups[1] == 300 && out_ids[1] == 30);
+    assert(out_groups[2] == 200 && out_ids[2] == 20);
+
+    uint64_t removed = 0;
+    const uint64_t deleted = 5;
+    assert(ifs_search_delete_batch(index, &deleted, 1, &removed) == IFS_SEARCH_OK);
+    assert(removed == 1);
+    assert(ifs_search_grouped_topk(
+               index, positive.data(), 6, out_groups.data(), out_ids.data(),
+               scores.data(), &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 3);
+    assert(out_groups[0] == 100 && out_ids[0] == 10);
+
+    const uint64_t deleted_last_in_group = 10;
+    assert(ifs_search_delete_batch(index, &deleted_last_in_group, 1, &removed) ==
+           IFS_SEARCH_OK);
+    assert(removed == 1);
+    const uint64_t replacement_id = 40;
+    const uint64_t replacement_group = 400;
+    assert(ifs_search_add_batch(index, &replacement_id, &replacement_group,
+                                positive.data(), 1) == IFS_SEARCH_OK);
+    const uint64_t reactivated_id = 50;
+    const uint64_t reactivated_group = 100;
+    assert(ifs_search_add_batch(index, &reactivated_id, &reactivated_group,
+                                positive.data(), 1) == IFS_SEARCH_OK);
+    assert(ifs_search_grouped_topk(
+               index, positive.data(), 6, out_groups.data(), out_ids.data(),
+               scores.data(), &count, &timings) == IFS_SEARCH_OK);
+    assert(count == 4);
+    assert(out_groups[0] == 100 && out_ids[0] == 50);
+    assert(out_groups[1] == 300 && out_ids[1] == 30);
+    assert(out_groups[2] == 400 && out_ids[2] == 40);
+    assert(out_groups[3] == 200 && out_ids[3] == 20);
+
+    const uint64_t overflow_id = 60;
+    const uint64_t overflow_group = 600;
+    assert(ifs_search_add_batch(index, &overflow_id, &overflow_group,
+                                positive.data(), 1) ==
+           IFS_SEARCH_CAPACITY_EXCEEDED);
+
+    ifs_search_stats_t stats{};
+    stats.struct_size = sizeof(stats);
+    assert(ifs_search_get_stats(index, &stats) == IFS_SEARCH_OK);
+    assert(stats.physical_rows == 6);
+    assert(stats.live_rows == 4);
+    assert(stats.tombstone_rows == 2);
+    assert(stats.capacity_rows == 6);
+    ifs_search_destroy(index);
+}
+
+void test_group_metadata_survives_growth() {
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = IFS_SEARCH_PROFILE_FP32_V1;
+    options.reserve_rows = 1;
+    options.max_rows = 8;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_DEVICE;
+    options.growth_factor = 1.5;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    const Vector positive = constant();
+    const Vector orthogonal = split();
+    std::vector<float> rows;
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    rows.insert(rows.end(), orthogonal.begin(), orthogonal.end());
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    rows.insert(rows.end(), positive.begin(), positive.end());
+    const std::array<uint64_t, 4> ids{8, 7, 6, 5};
+    const std::array<uint64_t, 4> groups{900, 800, 900, 700};
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(), rows.data(), 4) ==
+           IFS_SEARCH_OK);
+
+    std::array<uint64_t, 4> out_groups{};
+    std::array<uint64_t, 4> out_ids{};
+    std::array<float, 4> scores{};
+    uint64_t count = 0;
+    assert(ifs_search_grouped_topk(
+               index, positive.data(), 4, out_groups.data(), out_ids.data(),
+               scores.data(), &count, nullptr) == IFS_SEARCH_OK);
+    assert(count == 3);
+    assert(out_groups[0] == 700 && out_ids[0] == 5);
+    assert(out_groups[1] == 900 && out_ids[1] == 6);
+    assert(out_groups[2] == 800 && out_ids[2] == 7);
+
+    ifs_search_stats_t stats{};
+    stats.struct_size = sizeof(stats);
+    assert(ifs_search_get_stats(index, &stats) == IFS_SEARCH_OK);
+    assert(stats.reallocations > 0);
+    ifs_search_destroy(index);
+}
+
+void test_grouped_matches_exact_face_reference(uint32_t profile) {
+    constexpr size_t initial_rows = 513;
+    constexpr size_t added_rows = 23;
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = profile;
+    options.reserve_rows = 3;
+    options.max_rows = 700;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_HOST;
+    options.growth_factor = 1.25;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    std::vector<uint64_t> ids(initial_rows);
+    std::vector<uint64_t> groups(initial_rows);
+    std::vector<float> vectors;
+    vectors.reserve(initial_rows * IFS_SEARCH_DIMENSION);
+    std::unordered_map<uint64_t, uint64_t> group_by_id;
+    for (size_t i = 0; i < initial_rows; ++i) {
+        ids[i] = 1000 + i;
+        groups[i] = 100 + (i * 37) % 137;
+        group_by_id.emplace(ids[i], groups[i]);
+        const Vector value = wave(i + 11);
+        vectors.insert(vectors.end(), value.begin(), value.end());
+    }
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(),
+                                vectors.data(), initial_rows) == IFS_SEARCH_OK);
+
+    std::vector<uint64_t> deleted;
+    for (size_t i = 0; i < initial_rows; i += 17) deleted.push_back(ids[i]);
+    uint64_t removed = 0;
+    assert(ifs_search_delete_batch(index, deleted.data(), deleted.size(),
+                                   &removed) == IFS_SEARCH_OK);
+    assert(removed == deleted.size());
+
+    std::vector<uint64_t> added_ids(added_rows);
+    std::vector<uint64_t> added_groups(added_rows);
+    std::vector<float> added_vectors;
+    added_vectors.reserve(added_rows * IFS_SEARCH_DIMENSION);
+    for (size_t i = 0; i < added_rows; ++i) {
+        added_ids[i] = 5000 + i;
+        added_groups[i] = i % 2 == 0 ? 100 + (i * 19) % 137 : 1000 + i;
+        group_by_id.emplace(added_ids[i], added_groups[i]);
+        const Vector value = wave(i + 3000);
+        added_vectors.insert(added_vectors.end(), value.begin(), value.end());
+    }
+    assert(ifs_search_add_batch(index, added_ids.data(), added_groups.data(),
+                                added_vectors.data(), added_rows) ==
+           IFS_SEARCH_OK);
+
+    ifs_search_stats_t stats{};
+    stats.struct_size = sizeof(stats);
+    assert(ifs_search_get_stats(index, &stats) == IFS_SEARCH_OK);
+    assert(stats.reallocations > 0);
+    const size_t live_rows = static_cast<size_t>(stats.live_rows);
+    const Vector query = wave(9999);
+    std::vector<uint64_t> face_ids(live_rows);
+    std::vector<float> face_scores(live_rows);
+    uint64_t face_count = 0;
+    assert(ifs_search_topk(index, query.data(), live_rows, face_ids.data(),
+                           face_scores.data(), &face_count, nullptr) ==
+           IFS_SEARCH_OK);
+    assert(face_count == live_rows);
+
+    struct Expected {
+        uint64_t group_id;
+        uint64_t vector_id;
+        float score;
+    };
+    std::unordered_map<uint64_t, Expected> best_by_group;
+    for (size_t i = 0; i < live_rows; ++i) {
+        const uint64_t group_id = group_by_id.at(face_ids[i]);
+        const Expected candidate{group_id, face_ids[i], face_scores[i]};
+        const auto found = best_by_group.find(group_id);
+        if (found == best_by_group.end() ||
+            candidate.score > found->second.score ||
+            (candidate.score == found->second.score &&
+             candidate.vector_id < found->second.vector_id)) {
+            best_by_group[group_id] = candidate;
+        }
+    }
+    std::vector<Expected> expected;
+    expected.reserve(best_by_group.size());
+    for (const auto &entry : best_by_group) expected.push_back(entry.second);
+    std::sort(expected.begin(), expected.end(),
+              [](const Expected &left, const Expected &right) {
+                  if (left.score != right.score) {
+                      return left.score > right.score;
+                  }
+                  if (left.group_id != right.group_id) {
+                      return left.group_id < right.group_id;
+                  }
+                  return left.vector_id < right.vector_id;
+              });
+
+    constexpr size_t grouped_k = 100;
+    std::array<uint64_t, grouped_k> out_groups{};
+    std::array<uint64_t, grouped_k> out_ids{};
+    std::array<float, grouped_k> out_scores{};
+    uint64_t grouped_count = 0;
+    assert(ifs_search_grouped_topk(
+               index, query.data(), grouped_k, out_groups.data(),
+               out_ids.data(), out_scores.data(), &grouped_count, nullptr) ==
+           IFS_SEARCH_OK);
+    assert(grouped_count == std::min(grouped_k, expected.size()));
+    for (size_t i = 0; i < static_cast<size_t>(grouped_count); ++i) {
+        assert(out_groups[i] == expected[i].group_id);
+        assert(out_ids[i] == expected[i].vector_id);
+        assert(std::abs(out_scores[i] - expected[i].score) < 1.0e-6f);
+    }
+    ifs_search_destroy(index);
+}
+
+void test_grouped_multistage_and_k_boundaries() {
+    constexpr size_t row_count = 9001;
+    ifs_search_create_options_t options{};
+    options.struct_size = sizeof(options);
+    options.profile = IFS_SEARCH_PROFILE_FP32_V1;
+    options.reserve_rows = row_count;
+    options.max_rows = row_count;
+    options.device = 0;
+    options.topk_mode = IFS_SEARCH_TOPK_HOST;
+    options.growth_factor = 1.5;
+
+    ifs_search_index_t index = nullptr;
+    assert(ifs_search_create(&options, &index) == IFS_SEARCH_OK);
+    const Vector query = constant();
+    const uint64_t reserved_id = UINT64_MAX;
+    const uint64_t reserved_group = 1;
+    assert(ifs_search_add_batch(index, &reserved_id, &reserved_group,
+                                query.data(), 1) ==
+           IFS_SEARCH_INVALID_ARGUMENT);
+    std::vector<uint64_t> ids(row_count);
+    std::vector<uint64_t> groups(row_count);
+    std::vector<float> vectors;
+    vectors.reserve(row_count * IFS_SEARCH_DIMENSION);
+    for (size_t i = 0; i < row_count; ++i) {
+        ids[i] = 10000 + i;
+        groups[i] = ids[i];
+        const Vector value = sparse(i);
+        vectors.insert(vectors.end(), value.begin(), value.end());
+    }
+    assert(ifs_search_add_batch(index, ids.data(), groups.data(),
+                                vectors.data(), row_count) == IFS_SEARCH_OK);
+
+    std::array<uint64_t, 101> face_ids{};
+    std::array<float, 101> face_scores{};
+    uint64_t face_count = 0;
+    assert(ifs_search_topk(index, query.data(), 100, face_ids.data(),
+                           face_scores.data(), &face_count, nullptr) ==
+           IFS_SEARCH_OK);
+    assert(face_count == 100);
+
+    std::array<uint64_t, 101> out_groups{};
+    std::array<uint64_t, 101> out_ids{};
+    std::array<float, 101> out_scores{};
+    uint64_t grouped_count = 0;
+    assert(ifs_search_grouped_topk(
+               index, query.data(), 100, out_groups.data(), out_ids.data(),
+               out_scores.data(), &grouped_count, nullptr) == IFS_SEARCH_OK);
+    assert(grouped_count == 100);
+    for (size_t i = 0; i < 100; ++i) {
+        assert(out_groups[i] == face_ids[i]);
+        assert(out_ids[i] == face_ids[i]);
+        assert(out_scores[i] == face_scores[i]);
+    }
+
+    assert(ifs_search_grouped_topk(
+               index, query.data(), 1, out_groups.data(), out_ids.data(),
+               out_scores.data(), &grouped_count, nullptr) == IFS_SEARCH_OK);
+    assert(grouped_count == 1 && out_ids[0] == face_ids[0]);
+    assert(ifs_search_grouped_topk(index, query.data(), 0, nullptr, nullptr,
+                                   nullptr, &grouped_count, nullptr) ==
+           IFS_SEARCH_OK);
+    assert(grouped_count == 0);
+    assert(ifs_search_grouped_topk(
+               index, query.data(), 101, out_groups.data(), out_ids.data(),
+               out_scores.data(), &grouped_count, nullptr) ==
+           IFS_SEARCH_UNSUPPORTED);
+
+    ifs_search_destroy(index);
+}
+
+}  // namespace
+
+int main() {
+    static_assert(IFS_SEARCH_PROFILE_INT8_X1000_V1 == 3,
+                  "persisted x1000 profile code changed");
+    static_assert(IFS_SEARCH_PROFILE_INT8_X736_V1 == 4,
+                  "x736 profile must append after existing codes");
+    ifs_search_capabilities_t capabilities{};
+    capabilities.struct_size = sizeof(capabilities);
+    const ifs_search_status_t capability_status =
+        ifs_search_get_capabilities(0, &capabilities);
+    if (capability_status != IFS_SEARCH_OK) {
+        std::fprintf(stderr,
+                     "SKIP: CUDA device 0 is unavailable in this build environment: %s\n",
+                     ifs_search_last_error());
+        return 77;
+    }
+    assert(capabilities.backend == IFS_SEARCH_BACKEND_CUDA);
+    assert((capabilities.flags & IFS_SEARCH_CAP_DEVICE_TOPK) != 0);
+    assert((capabilities.flags & IFS_SEARCH_CAP_GROUPED_PERSON_TOPK) != 0);
+    assert((capabilities.flags & IFS_SEARCH_CAP_GROUPED_DEVICE_RESIDENT) != 0);
+    assert((capabilities.flags & IFS_SEARCH_CAP_GROUPED_HOST_REFERENCE) == 0);
+    assert(capabilities.device_topk_limit == 100);
+    assert((capabilities.profile_mask &
+            (UINT64_C(1) << IFS_SEARCH_PROFILE_INT8_X736_V1)) != 0);
+    assert(std::string(ifs_search_build_info()).find(
+               "grouped_topk=device-resident-exact") != std::string::npos);
+
+    for (uint32_t profile = IFS_SEARCH_PROFILE_FP32_V1;
+         profile <= IFS_SEARCH_PROFILE_INT8_X736_V1; ++profile) {
+        if ((capabilities.profile_mask & (UINT64_C(1) << profile)) != 0) {
+            test_profile(profile);
+            test_grouped_matches_exact_face_reference(profile);
+        }
+    }
+    test_group_metadata_survives_growth();
+    test_grouped_multistage_and_k_boundaries();
+    test_int8_numeric_contract(
+        IFS_SEARCH_PROFILE_INT8_X1000_V1, 0.013689f, 1.0f);
+    test_int8_numeric_contract(
+        IFS_SEARCH_PROFILE_INT8_X736_V1, 7440.0f / 541696.0f,
+        530755.0f / 541696.0f);
+    test_int8_scales_are_per_index();
+    return 0;
+}
diff --git a/server/native/search/include/ifs_search.h b/server/native/search/include/ifs_search.h
new file mode 100644
index 000000000..6a5a040b4
--- /dev/null
+++ b/server/native/search/include/ifs_search.h
@@ -0,0 +1,211 @@
+#ifndef IFS_SEARCH_H
+#define IFS_SEARCH_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#if defined(_WIN32)
+#  if defined(IFS_SEARCH_BUILDING)
+#    define IFS_SEARCH_API __declspec(dllexport)
+#  else
+#    define IFS_SEARCH_API __declspec(dllimport)
+#  endif
+#else
+#  define IFS_SEARCH_API __attribute__((visibility("default")))
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define IFS_SEARCH_ABI_VERSION 2u
+#define IFS_SEARCH_DIMENSION 512u
+#define IFS_SEARCH_INT8_X1000_SCALE 1000u
+#define IFS_SEARCH_INT8_X736_SCALE 736u
+/* Source-compatibility alias for ABI-v2 clients built before x736 existed. */
+#define IFS_SEARCH_INT8_SCALE IFS_SEARCH_INT8_X1000_SCALE
+
+typedef void *ifs_search_index_t;
+
+typedef enum ifs_search_status {
+    IFS_SEARCH_OK = 0,
+    IFS_SEARCH_INVALID_ARGUMENT = 1,
+    IFS_SEARCH_OUT_OF_MEMORY = 2,
+    IFS_SEARCH_UNSUPPORTED = 3,
+    IFS_SEARCH_DUPLICATE_ID = 4,
+    IFS_SEARCH_ID_NOT_FOUND = 5,
+    IFS_SEARCH_CAPACITY_EXCEEDED = 6,
+    IFS_SEARCH_BACKEND_ERROR = 7,
+    IFS_SEARCH_INTERNAL_ERROR = 8
+} ifs_search_status_t;
+
+typedef enum ifs_search_backend {
+    IFS_SEARCH_BACKEND_CPU = 1,
+    IFS_SEARCH_BACKEND_CUDA = 2
+} ifs_search_backend_t;
+
+/* These values are persisted by the Server. Do not renumber them. */
+typedef enum ifs_search_profile {
+    IFS_SEARCH_PROFILE_FP32_V1 = 0,
+    IFS_SEARCH_PROFILE_FP16_V1 = 1,
+    IFS_SEARCH_PROFILE_BF16_V1 = 2,
+    IFS_SEARCH_PROFILE_INT8_X1000_V1 = 3,
+    IFS_SEARCH_PROFILE_INT8_X736_V1 = 4
+} ifs_search_profile_t;
+
+typedef enum ifs_search_topk_mode {
+    IFS_SEARCH_TOPK_AUTO = 0,
+    IFS_SEARCH_TOPK_HOST = 1,
+    IFS_SEARCH_TOPK_DEVICE = 2
+} ifs_search_topk_mode_t;
+
+enum ifs_search_capability_flag {
+    IFS_SEARCH_CAP_EXACT_FLAT_SCAN = UINT64_C(1) << 0,
+    IFS_SEARCH_CAP_BATCH_ADD = UINT64_C(1) << 1,
+    IFS_SEARCH_CAP_BATCH_DELETE = UINT64_C(1) << 2,
+    IFS_SEARCH_CAP_RESERVE = UINT64_C(1) << 3,
+    IFS_SEARCH_CAP_DEVICE_TOPK = UINT64_C(1) << 4,
+    IFS_SEARCH_CAP_GROUPED_PERSON_TOPK = UINT64_C(1) << 5,
+    IFS_SEARCH_CAP_DELETED_SLOT_REUSE = UINT64_C(1) << 6,
+    IFS_SEARCH_CAP_TOMBSTONE_DELETE = UINT64_C(1) << 7,
+    /* Grouped Top-K is exact but aggregates after transferring/selecting all
+     * face scores on the host. It is a correctness reference path. */
+    IFS_SEARCH_CAP_GROUPED_HOST_REFERENCE = UINT64_C(1) << 8,
+    /* Group reduction and exact Top-K selection remain on the accelerator;
+     * only the final K group candidates cross the device boundary. */
+    IFS_SEARCH_CAP_GROUPED_DEVICE_RESIDENT = UINT64_C(1) << 9
+};
+
+typedef struct ifs_search_capabilities {
+    uint32_t struct_size;
+    uint32_t abi_version;
+    uint32_t dimension;
+    uint32_t backend;
+    uint64_t profile_mask;
+    uint64_t flags;
+    uint64_t device_topk_limit;
+    int32_t device;
+    int32_t compute_capability_major;
+    int32_t compute_capability_minor;
+    int32_t cuda_runtime_version;
+    int32_t cuda_driver_version;
+} ifs_search_capabilities_t;
+
+typedef struct ifs_search_create_options {
+    uint32_t struct_size;
+    uint32_t profile;
+    uint64_t reserve_rows;
+    /* Zero means no logical hard limit. A non-zero limit is checked before
+     * every add. Set reserve_rows == max_rows to guarantee no data reallocation
+     * while the collection remains within its configured capacity. */
+    uint64_t max_rows;
+    int32_t device;
+    uint32_t topk_mode;
+    double growth_factor;
+    uint64_t reserved[4];
+} ifs_search_create_options_t;
+
+typedef struct ifs_search_stats {
+    uint32_t struct_size;
+    uint32_t backend;
+    uint32_t profile;
+    int32_t device;
+    uint64_t physical_rows;
+    uint64_t live_rows;
+    uint64_t capacity_rows;
+    uint64_t max_rows;
+    uint64_t tombstone_rows;
+    /* Number of automatic data-storage growth events after creation. Explicit
+     * reserve calls are not counted. */
+    uint64_t reallocations;
+    uint64_t bytes_per_vector;
+} ifs_search_stats_t;
+
+typedef struct ifs_search_timings {
+    uint32_t struct_size;
+    uint32_t reserved;
+    double kernel_ms;
+    double topk_ms;
+    double total_ms;
+} ifs_search_timings_t;
+
+IFS_SEARCH_API uint32_t ifs_search_abi_version(void);
+IFS_SEARCH_API uint32_t ifs_search_dimension(void);
+IFS_SEARCH_API const char *ifs_search_build_info(void);
+IFS_SEARCH_API const char *ifs_search_last_error(void);
+IFS_SEARCH_API const char *ifs_search_status_string(ifs_search_status_t status);
+
+/* The CPU library accepts device=-1. The CUDA library requires a valid CUDA
+ * device ordinal. profile_mask is authoritative: unsupported profiles fail
+ * creation instead of falling back to another representation or backend. */
+IFS_SEARCH_API ifs_search_status_t ifs_search_get_capabilities(
+    int32_t device,
+    ifs_search_capabilities_t *out_capabilities);
+
+IFS_SEARCH_API ifs_search_status_t ifs_search_create(
+    const ifs_search_create_options_t *options,
+    ifs_search_index_t *out_index);
+IFS_SEARCH_API void ifs_search_destroy(ifs_search_index_t index);
+
+IFS_SEARCH_API ifs_search_status_t ifs_search_reserve(
+    ifs_search_index_t index,
+    uint64_t rows);
+
+/* All vectors and queries are row-major finite FP32 input with exactly 512
+ * values per row and must already be L2-normalized. Low-precision profiles
+ * quantize that normalized FP32 representation internally. UINT64_MAX is
+ * reserved as an internal sentinel and is not a valid vector ID. */
+IFS_SEARCH_API ifs_search_status_t ifs_search_add_batch(
+    ifs_search_index_t index,
+    const uint64_t *vector_ids,
+    const uint64_t *group_ids,
+    const float *vectors,
+    uint64_t count);
+
+/* Missing IDs are ignored. out_removed receives the number actually removed.
+ * CPU indexes reuse deleted slots. CUDA indexes retain tombstones until the
+ * Server rebuilds the collection generation. */
+IFS_SEARCH_API ifs_search_status_t ifs_search_delete_batch(
+    ifs_search_index_t index,
+    const uint64_t *vector_ids,
+    uint64_t count,
+    uint64_t *out_removed);
+
+/* Exact face-vector Top-K. Results are score descending, vector ID ascending.
+ * Scores always use raw cosine/inner-product semantics. INT8 scores are the
+ * int32 accumulator divided by the selected profile's scale squared, never
+ * the unscaled accumulator. */
+IFS_SEARCH_API ifs_search_status_t ifs_search_topk(
+    ifs_search_index_t index,
+    const float *query512,
+    uint64_t top_k,
+    uint64_t *out_vector_ids,
+    float *out_cosine_scores,
+    uint64_t *out_count,
+    ifs_search_timings_t *out_timings);
+
+/* Exact grouped Top-K. Each vector's group ID is supplied at add time. Every
+ * live face score participates, each group retains its maximum-scoring face,
+ * then groups are sorted by score descending, group ID ascending, and vector
+ * ID ascending. This is not a fixed face-candidate oversample and therefore
+ * cannot omit a Person. Capability flags distinguish host-reference and
+ * accelerator-resident implementations. */
+IFS_SEARCH_API ifs_search_status_t ifs_search_grouped_topk(
+    ifs_search_index_t index,
+    const float *query512,
+    uint64_t top_k,
+    uint64_t *out_group_ids,
+    uint64_t *out_vector_ids,
+    float *out_cosine_scores,
+    uint64_t *out_count,
+    ifs_search_timings_t *out_timings);
+
+IFS_SEARCH_API ifs_search_status_t ifs_search_get_stats(
+    ifs_search_index_t index,
+    ifs_search_stats_t *out_stats);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/server/pyproject.toml b/server/pyproject.toml
new file mode 100644
index 000000000..7e9058b13
--- /dev/null
+++ b/server/pyproject.toml
@@ -0,0 +1,56 @@
+[build-system]
+requires = ["setuptools==83.0.0", "wheel==0.46.2"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "insightface-server"
+version = "0.2.0"
+description = "Simple self-hosted InsightFace REST API and management UI"
+readme = "README.md"
+requires-python = ">=3.11,<3.12"
+license-files = ["LICENSING.md"]
+authors = [{name = "InsightFace Contributors"}]
+dynamic = ["dependencies"]
+
+[tool.setuptools]
+package-dir = {"" = "backend"}
+
+[tool.setuptools.packages.find]
+where = ["backend"]
+
+[tool.setuptools.package-data]
+"insightface_server.licensing" = [
+  "trusted_keys/*.pem",
+  "defaults/*/MODEL.LICENSE",
+]
+
+[tool.setuptools.dynamic]
+dependencies = {file = ["requirements.common.lock"]}
+
+[tool.pytest.ini_options]
+addopts = "-ra --strict-markers"
+testpaths = ["tests"]
+markers = [
+  "cpu: requires local real ONNX models",
+  "cuda: requires an NVIDIA GPU and CUDA image",
+  "consistency: compares CPU and CUDA inference outputs",
+  "docker: requires Docker Engine",
+  "e2e: exercises the browser UI",
+]
+
+[tool.ruff]
+target-version = "py311"
+line-length = 100
+
+[tool.ruff.lint]
+select = ["E", "F", "I", "B", "UP"]
+ignore = ["E501"]
+
+[tool.ruff.lint.per-file-ignores]
+"backend/insightface_server/app.py" = ["B008"]
+
+[tool.mypy]
+python_version = "3.11"
+ignore_missing_imports = true
+check_untyped_defs = true
+warn_unused_ignores = true
diff --git a/server/requirements.common.lock b/server/requirements.common.lock
new file mode 100644
index 000000000..dc623de41
--- /dev/null
+++ b/server/requirements.common.lock
@@ -0,0 +1,40 @@
+# Runtime closure for CPython 3.11. All direct and transitive dependencies are
+# pinned so CPU and CUDA images execute the same application code.
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.2
+certifi==2026.6.17
+cffi==2.1.0
+charset-normalizer==3.4.9
+click==8.4.2
+cryptography==49.0.0
+fastapi==0.140.0
+flatbuffers==25.12.19
+h11==0.16.0
+idna==3.18
+imageio==2.37.3
+lazy-loader==0.5
+ml-dtypes==0.5.4
+networkx==3.6.1
+numpy==1.26.4
+onnx==1.22.0
+opencv-python-headless==4.11.0.86
+packaging==26.2
+pillow==12.3.0
+protobuf==7.35.1
+pycparser==3.0
+pydantic==2.13.4
+pydantic-core==2.46.4
+python-multipart==0.0.32
+requests==2.34.2
+rfc8785==0.1.4
+regex==2024.11.6
+scikit-image==0.25.0
+scipy==1.15.1
+starlette==1.3.1
+tifffile==2026.3.3
+tqdm==4.67.1
+typing-extensions==4.16.0
+typing-inspection==0.4.2
+urllib3==2.7.0
+uvicorn==0.34.0
diff --git a/server/requirements.cpu.lock b/server/requirements.cpu.lock
new file mode 100644
index 000000000..0299f2034
--- /dev/null
+++ b/server/requirements.cpu.lock
@@ -0,0 +1,2 @@
+-r requirements.common.lock
+onnxruntime==1.27.0
diff --git a/server/requirements.cuda12.lock b/server/requirements.cuda12.lock
new file mode 100644
index 000000000..c9593feb2
--- /dev/null
+++ b/server/requirements.cuda12.lock
@@ -0,0 +1,4 @@
+-r requirements.common.lock
+# ONNX Runtime 1.27's PyPI GPU wheel targets CUDA 13. This fixed CUDA 12
+# wheel is published by Microsoft and protected by its SHA-256 fragment.
+onnxruntime-gpu @ https://aiinfra.pkgs.visualstudio.com/2692857e-05ef-43b4-ba9c-ccf1c22c437c/_packaging/9387c3aa-d9ad-4513-968c-383f6f7f53b8/pypi/download/onnxruntime-gpu/1.27/onnxruntime_gpu-1.27.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=00238cee89f8f4f5c65dae0de9178f1ca903e34068029c2ccfee358a028f84ba
diff --git a/server/requirements.dev.lock b/server/requirements.dev.lock
new file mode 100644
index 000000000..3ff06ff91
--- /dev/null
+++ b/server/requirements.dev.lock
@@ -0,0 +1,14 @@
+-r requirements.cpu.lock
+coverage==7.15.1
+httpcore==1.0.9
+httpcore2==2.9.1
+httpx==0.28.1
+httpx2==2.9.1
+iniconfig==2.3.0
+mypy==1.14.1
+mypy-extensions==1.1.0
+pluggy==1.6.0
+pytest==9.0.3
+pytest-cov==6.0.0
+ruff==0.9.3
+truststore==0.10.4
diff --git a/server/scripts/create_model_manifest.py b/server/scripts/create_model_manifest.py
new file mode 100644
index 000000000..ffe0c9c96
--- /dev/null
+++ b/server/scripts/create_model_manifest.py
@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+"""Create the compact v1 manifest expected by InsightFace Server."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--models-dir", type=Path, required=True)
+    parser.add_argument("--detector", default="det_10g.onnx")
+    parser.add_argument("--recognizer", default="w600k_r50.onnx")
+    parser.add_argument("--model-id", default="buffalo_l")
+    parser.add_argument("--model-version", default="v0.7")
+    parser.add_argument("--display-name", default="Buffalo_L")
+    parser.add_argument("--license", default="MODEL.LICENSE")
+    parser.add_argument("--output", default="manifest.json")
+    args = parser.parse_args()
+    detector = (args.models_dir / args.detector).resolve()
+    recognizer = (args.models_dir / args.recognizer).resolve()
+    root = args.models_dir.resolve()
+    license_path = (args.models_dir / args.license).resolve()
+    for package_file in (detector, recognizer, license_path):
+        if package_file.parent != root or not package_file.is_file():
+            parser.error(
+                f"package file does not exist directly in --models-dir: {package_file}"
+            )
+    manifest = {
+        "manifest_version": 1,
+        "model_id": args.model_id,
+        "model_version": args.model_version,
+        "display_name": args.display_name,
+        "files": {
+            "detector": detector.name,
+            "recognizer": recognizer.name,
+        },
+        "recognition": {
+            "input_size": [112, 112],
+            "embedding_dimension": 512,
+            "preprocessing": "insightface-arcface-1",
+        },
+        "license": license_path.name,
+    }
+    output = root / args.output
+    output.write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+    print(output)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/server/scripts/release_preflight.py b/server/scripts/release_preflight.py
new file mode 100644
index 000000000..1b126aa74
--- /dev/null
+++ b/server/scripts/release_preflight.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+"""Validate that every public InsightFace Server release surface agrees.
+
+The normal mode is intentionally strict and is intended for the owner-operated
+manual release process. ``--relaxed`` keeps all structural checks enabled while
+turning the dirty-worktree condition expected during local development into a
+warning.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+import tomllib
+from dataclasses import asdict, dataclass
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+SEMVER = re.compile(r"^[0-9]+\.[0-9]+\.[0-9]+$")
+README_NAMES = (
+    "README.md",
+    "README.zh-CN.md",
+    "README.ja.md",
+    "README.de.md",
+    "README.es.md",
+    "README.fr.md",
+    "README.ru.md",
+    "README.pt.md",
+    "README.ko.md",
+)
+
+
+@dataclass(frozen=True)
+class Check:
+    name: str
+    status: str
+    detail: str
+
+
+class Preflight:
+    def __init__(self, root: Path, version: str, *, relaxed: bool) -> None:
+        self.root = root.resolve()
+        self.server = self.root / "server"
+        self.version = version
+        self.relaxed = relaxed
+        self.checks: list[Check] = []
+
+    def pass_(self, name: str, detail: str) -> None:
+        self.checks.append(Check(name, "pass", detail))
+
+    def warn(self, name: str, detail: str) -> None:
+        self.checks.append(Check(name, "warning", detail))
+
+    def fail(self, name: str, detail: str) -> None:
+        self.checks.append(Check(name, "fail", detail))
+
+    def require(self, name: str, condition: bool, detail: str) -> None:
+        if condition:
+            self.pass_(name, detail)
+        else:
+            self.fail(name, detail)
+
+    def _text(self, relative: str) -> str:
+        return (self.root / relative).read_text(encoding="utf-8")
+
+    def _toml(self, relative: str) -> dict[str, Any]:
+        with (self.root / relative).open("rb") as handle:
+            return tomllib.load(handle)
+
+    def _git(self, *args: str) -> subprocess.CompletedProcess[str]:
+        return subprocess.run(
+            ("git", *args),
+            cwd=self.root,
+            check=False,
+            capture_output=True,
+            text=True,
+        )
+
+    def _check_exact_version(self, name: str, actual: object) -> None:
+        self.require(
+            name,
+            actual == self.version,
+            f"expected {self.version!r}; found {actual!r}",
+        )
+
+    def run(self) -> list[Check]:
+        self.require(
+            "version-format",
+            bool(SEMVER.fullmatch(self.version)),
+            f"release version is {self.version!r}",
+        )
+        self._check_package_versions()
+        self._check_container_versions()
+        self._check_document_versions()
+        self._check_license_metadata()
+        self._check_git_state()
+        return self.checks
+
+    def _check_package_versions(self) -> None:
+        server_project = self._toml("server/pyproject.toml")["project"]
+        sdk_project = self._toml("server/sdk/python/pyproject.toml")["project"]
+        self._check_exact_version("server-pyproject-version", server_project["version"])
+        self._check_exact_version("sdk-pyproject-version", sdk_project["version"])
+
+        backend_init = self._text("server/backend/insightface_server/__init__.py")
+        sdk_init = self._text(
+            "server/sdk/python/src/insightface_server/__init__.py"
+        )
+        client = self._text(
+            "server/sdk/python/src/insightface_server/client.py"
+        )
+        model_packages = self._text(
+            "server/backend/insightface_server/models/packages.py"
+        )
+        self.require(
+            "backend-runtime-version",
+            f'__version__ = "{self.version}"' in backend_init,
+            "backend __version__ matches",
+        )
+        self.require(
+            "sdk-runtime-version",
+            f'__version__ = "{self.version}"' in sdk_init,
+            "SDK __version__ matches",
+        )
+        self.require(
+            "sdk-user-agent-version",
+            f"insightface-server-python/{self.version}" in client,
+            "SDK User-Agent matches",
+        )
+        self.require(
+            "model-installer-user-agent-version",
+            f"InsightFace-Server-Model-Installer/{self.version}" in model_packages,
+            "model installer User-Agent matches",
+        )
+
+    def _check_container_versions(self) -> None:
+        makefile = self._text("server/Makefile")
+        self.require(
+            "make-version",
+            re.search(
+                rf"^SERVER_VERSION\s*\?=\s*{re.escape(self.version)}$",
+                makefile,
+                re.MULTILINE,
+            )
+            is not None,
+            "Makefile SERVER_VERSION matches",
+        )
+        for variant, dockerfile in (
+            ("cpu", "server/docker/Dockerfile.cpu"),
+            ("cuda12", "server/docker/Dockerfile.cuda12"),
+        ):
+            text = self._text(dockerfile)
+            self.require(
+                f"docker-{variant}-version",
+                f"ARG INSIGHTFACE_SERVER_VERSION={self.version}" in text,
+                f"{dockerfile} default version matches",
+            )
+            self.require(
+                f"docker-{variant}-source-label",
+                'org.opencontainers.image.source="https://github.com/deepinsight/insightface"'
+                in text,
+                f"{dockerfile} links the source repository",
+            )
+
+        for variant, compose in (
+            ("cpu", "server/deploy/compose.cpu.yml"),
+            ("cuda12", "server/deploy/compose.cuda12.yml"),
+        ):
+            expected = (
+                f"ghcr.io/deepinsight/insightface-server:"
+                f"{self.version}-{variant}"
+            )
+            text = self._text(compose)
+            self.require(
+                f"compose-{variant}-version",
+                text.count(expected) == 2,
+                f"{compose} has exactly two {expected} references",
+            )
+            self.require(
+                f"compose-{variant}-no-latest",
+                "insightface-server:latest" not in text,
+                f"{compose} does not use latest",
+            )
+
+    def _check_document_versions(self) -> None:
+        for name in README_NAMES:
+            text = self._text(f"server/{name}")
+            self.require(
+                f"{name}-version",
+                self.version in text,
+                f"{name} names {self.version}",
+            )
+            self.require(
+                f"{name}-variant-tags",
+                f"{self.version}-cpu" in text and f"{self.version}-cuda12" in text,
+                f"{name} names both immutable image variants",
+            )
+
+        snapshot = json.loads(
+            self._text("server/docs/openapi.snapshot.json")
+        )
+        self._check_exact_version(
+            "openapi-version", snapshot.get("info", {}).get("version")
+        )
+
+    def _check_license_metadata(self) -> None:
+        for relative in (
+            "server/pyproject.toml",
+            "server/sdk/python/pyproject.toml",
+        ):
+            project = self._toml(relative)["project"]
+            self.require(
+                f"{relative}-legacy-license-field",
+                "license" not in project,
+                f"{relative} has no legacy project.license declaration",
+            )
+            classifiers = project.get("classifiers", [])
+            self.require(
+                f"{relative}-mit-classifier",
+                "License :: OSI Approved :: MIT License" not in classifiers,
+                f"{relative} has no MIT classifier",
+            )
+
+    def _check_git_state(self) -> None:
+        revision = self._git("rev-parse", "--verify", "HEAD")
+        sha = revision.stdout.strip()
+        self.require(
+            "git-revision",
+            revision.returncode == 0 and bool(re.fullmatch(r"[0-9a-f]{40}", sha)),
+            f"source revision is {sha or 'unavailable'}",
+        )
+
+        status = self._git("status", "--porcelain=v1", "--untracked-files=all")
+        dirty_lines = [line for line in status.stdout.splitlines() if line.strip()]
+        if dirty_lines and self.relaxed:
+            self.warn(
+                "git-clean",
+                f"relaxed precheck accepts {len(dirty_lines)} dirty paths; "
+                "release rejects them",
+            )
+        else:
+            self.require(
+                "git-clean",
+                status.returncode == 0 and not dirty_lines,
+                f"worktree has {len(dirty_lines)} dirty paths",
+            )
+
+        tag = f"server-v{self.version}"
+        existing_tag = self._git("show-ref", "--verify", "--quiet", f"refs/tags/{tag}")
+        if existing_tag.returncode != 0:
+            self.pass_("git-tag-idempotent", f"release tag {tag} is absent")
+            self.pass_("git-tag-annotated", f"release tag {tag} is not created yet")
+        else:
+            tag_revision = self._git("rev-list", "-n", "1", tag).stdout.strip()
+            tag_type = self._git("cat-file", "-t", tag).stdout.strip()
+            self.require(
+                "git-tag-idempotent",
+                tag_revision == sha,
+                f"existing release tag points to {tag_revision or 'unavailable'}",
+            )
+            self.require(
+                "git-tag-annotated",
+                tag_type == "tag",
+                f"existing release tag object type is {tag_type or 'unavailable'}",
+            )
+
+    def report(self) -> dict[str, Any]:
+        failures = sum(check.status == "fail" for check in self.checks)
+        warnings = sum(check.status == "warning" for check in self.checks)
+        revision = self._git("rev-parse", "--verify", "HEAD").stdout.strip()
+        return {
+            "schema_version": 1,
+            "version": self.version,
+            "mode": "precheck" if self.relaxed else "release",
+            "source_revision": revision,
+            "generated_at": datetime.now(UTC)
+            .replace(microsecond=0)
+            .isoformat()
+            .replace("+00:00", "Z"),
+            "summary": {
+                "checks": len(self.checks),
+                "failures": failures,
+                "warnings": warnings,
+            },
+            "checks": [asdict(check) for check in self.checks],
+        }
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--version", required=True)
+    parser.add_argument(
+        "--root",
+        type=Path,
+        default=Path(__file__).resolve().parents[2],
+        help="repository root (defaults to the root containing server/)",
+    )
+    parser.add_argument(
+        "--relaxed",
+        action="store_true",
+        help="warn instead of failing for a dirty tree",
+    )
+    parser.add_argument("--output", type=Path)
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    preflight = Preflight(args.root, args.version, relaxed=args.relaxed)
+    preflight.run()
+    report = preflight.report()
+    rendered = json.dumps(report, indent=2, sort_keys=True)
+    if args.output is not None:
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        args.output.write_text(rendered + "\n", encoding="utf-8")
+    print(rendered)
+    return 1 if report["summary"]["failures"] else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/server/scripts/smoke_test.py b/server/scripts/smoke_test.py
new file mode 100644
index 000000000..f202939d5
--- /dev/null
+++ b/server/scripts/smoke_test.py
@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+"""Small dependency-free health and collection API smoke test."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import mimetypes
+import os
+import secrets
+import urllib.error
+import urllib.request
+from pathlib import Path
+from typing import Any
+
+
+def request(
+    base: str,
+    path: str,
+    key: str | None,
+    payload: dict | None = None,
+    *,
+    body: bytes | None = None,
+    content_type: str | None = None,
+    timeout: float = 10.0,
+) -> dict[str, Any]:
+    headers = {"accept": "application/json"}
+    if key:
+        headers["authorization"] = f"Bearer {key}"
+    data = None
+    if payload is not None:
+        if body is not None:
+            raise ValueError("payload and body are mutually exclusive")
+        headers["content-type"] = "application/json"
+        data = json.dumps(payload).encode()
+    elif body is not None:
+        if content_type is None:
+            raise ValueError("content_type is required with body")
+        headers["content-type"] = content_type
+        data = body
+    call = urllib.request.Request(base.rstrip("/") + path, headers=headers, data=data)
+    try:
+        with urllib.request.urlopen(call, timeout=timeout) as response:
+            response_body = response.read()
+    except urllib.error.HTTPError as exc:
+        raise SystemExit(f"HTTP {exc.code}: {exc.read().decode(errors='replace')}") from exc
+    return json.loads(response_body) if response_body else {}
+
+
+def multipart_image(
+    path: Path,
+    *,
+    fields: dict[str, str] | None = None,
+) -> tuple[bytes, str]:
+    boundary = f"insightface-release-{secrets.token_hex(16)}"
+    media_type = mimetypes.guess_type(path.name)[0] or "application/octet-stream"
+    parts = []
+    for name, value in sorted((fields or {}).items()):
+        parts.append(
+            (
+                f"--{boundary}\r\n"
+                f'Content-Disposition: form-data; name="{name}"\r\n\r\n'
+                f"{value}\r\n"
+            ).encode()
+        )
+    parts.append(
+        (
+            f"--{boundary}\r\n"
+            'Content-Disposition: form-data; name="image"; '
+            'filename="release-smoke-image"\r\n'
+            f"Content-Type: {media_type}\r\n\r\n"
+        ).encode()
+        + path.read_bytes()
+        + f"\r\n--{boundary}--\r\n".encode()
+    )
+    body = b"".join(parts)
+    return body, f"multipart/form-data; boundary={boundary}"
+
+
+def require_single_detected_face(response: dict[str, Any]) -> None:
+    faces = response.get("faces")
+    if not isinstance(faces, list):
+        raise SystemExit("the release smoke detect response has no face list")
+    if len(faces) != 1:
+        raise SystemExit(
+            f"the release smoke image must detect exactly one face; found {len(faces)}"
+        )
+
+
+def inference_report(response: dict[str, Any]) -> dict[str, Any]:
+    faces = response.get("faces")
+    if not isinstance(faces, list) or not faces:
+        raise SystemExit("the release smoke image did not produce a face")
+    if len(faces) != 1:
+        raise SystemExit(
+            f"the release smoke image must produce exactly one face; found {len(faces)}"
+        )
+    face = faces[0]
+    embedding = face.get("embedding")
+    if not isinstance(embedding, list) or not embedding:
+        raise SystemExit("the release smoke image did not produce an embedding")
+    values = [float(value) for value in embedding]
+    if not all(math.isfinite(value) for value in values):
+        raise SystemExit("the release smoke embedding contains a non-finite value")
+    norm = math.sqrt(sum(value * value for value in values))
+    if not math.isclose(norm, 1.0, abs_tol=1.0e-5):
+        raise SystemExit(f"the release smoke embedding is not normalized: {norm}")
+    return {
+        "face_count": len(faces),
+        "bbox": face["bbox"],
+        "detection_score": face["detection_score"],
+        "embedding_dimension": len(values),
+        "embedding_norm": norm,
+        "embedding": values,
+    }
+
+
+def redacted_report(report: dict[str, Any]) -> dict[str, Any]:
+    """Return log-safe evidence without face geometry or feature material."""
+    printable = json.loads(json.dumps(report))
+    if "inference" in printable:
+        inference = printable["inference"]
+        printable["inference"] = {
+            "face_count": inference["face_count"],
+            "embedding_dimension": inference["embedding_dimension"],
+            "embedding_norm": inference["embedding_norm"],
+        }
+    return printable
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-url", default="http://127.0.0.1:8080")
+    parser.add_argument("--api-key", default=os.getenv("INSIGHTFACE_API_KEY"))
+    parser.add_argument("--image", type=Path)
+    parser.add_argument("--output", type=Path)
+    args = parser.parse_args()
+    health = request(args.base_url, "/v1/health", None)
+    system = request(args.base_url, "/v1/system", args.api_key)
+    report: dict[str, Any] = {"health": health, "system": system}
+    if args.image is not None:
+        if not args.image.is_file():
+            raise SystemExit(f"release smoke image does not exist: {args.image}")
+        detect_body, detect_content_type = multipart_image(
+            args.image,
+            fields={"max_faces": "2"},
+        )
+        detection = request(
+            args.base_url,
+            "/v1/detect",
+            args.api_key,
+            body=detect_body,
+            content_type=detect_content_type,
+            timeout=65.0,
+        )
+        require_single_detected_face(detection)
+        body, content_type = multipart_image(args.image)
+        response = request(
+            args.base_url,
+            "/v1/embeddings",
+            args.api_key,
+            body=body,
+            content_type=content_type,
+            timeout=65.0,
+        )
+        report["inference"] = inference_report(response)
+    if args.output is not None:
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        args.output.write_text(
+            json.dumps(report, indent=2, sort_keys=True) + "\n",
+            encoding="utf-8",
+        )
+        args.output.chmod(0o600)
+    print(json.dumps(redacted_report(report), indent=2, sort_keys=True))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/server/scripts/update_openapi_snapshot.py b/server/scripts/update_openapi_snapshot.py
new file mode 100644
index 000000000..a5aec1a6f
--- /dev/null
+++ b/server/scripts/update_openapi_snapshot.py
@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+"""Regenerate the reviewed public OpenAPI contract snapshot.
+
+Run this only together with the matching API implementation, all localized
+``docs/api*.md`` files, SDK/UI changes, tests, and migration guidance when
+needed.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+SERVER_DIR = Path(__file__).resolve().parents[1]
+BACKEND_DIR = SERVER_DIR / "backend"
+if str(BACKEND_DIR) not in sys.path:
+    sys.path.insert(0, str(BACKEND_DIR))
+
+from insightface_server.app import create_app  # noqa: E402
+from insightface_server.config import Settings  # noqa: E402
+
+
+def main() -> None:
+    root = SERVER_DIR / "build" / "openapi-snapshot"
+    settings = Settings(
+        data_dir=root / "data",
+        models_dir=root / "models",
+        inference_mode="mock",
+        execution_provider="CPUExecutionProvider",
+        auth_enabled=False,
+        startup_api_key=None,
+        max_image_bytes=2 * 1024 * 1024,
+        max_image_pixels=2_000_000,
+        max_request_bytes=8 * 1024 * 1024,
+        max_registration_images=20,
+        request_timeout_seconds=10,
+        detector_threshold=0.5,
+        registration_min_score=0.6,
+        registration_min_quality=0.35,
+        registration_min_face_size=40,
+        default_threshold=0.4,
+        cors_origins=(),
+        log_level="WARNING",
+        save_face_crops=False,
+        device_id=0,
+    )
+    document = json.dumps(
+        create_app(settings).openapi(),
+        ensure_ascii=False,
+        indent=2,
+        sort_keys=True,
+    )
+    output = SERVER_DIR / "docs" / "openapi.snapshot.json"
+    output.write_text(f"{document}\n", encoding="utf-8")
+    print(f"updated {output}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/server/scripts/verify_runtime.py b/server/scripts/verify_runtime.py
new file mode 100644
index 000000000..4bb885c03
--- /dev/null
+++ b/server/scripts/verify_runtime.py
@@ -0,0 +1,190 @@
+#!/usr/bin/env python3
+"""Fail-fast verification for the pinned CPU and CUDA container runtimes."""
+
+from __future__ import annotations
+
+import argparse
+import ctypes
+import json
+import os
+import platform
+import re
+import subprocess
+import sys
+from typing import Any
+
+
+class VerificationError(RuntimeError):
+    pass
+
+
+def version(value: str) -> tuple[int, ...]:
+    parts = tuple(int(item) for item in re.findall(r"\d+", value))
+    if not parts:
+        raise VerificationError(f"Unable to parse version {value!r}")
+    return parts
+
+
+def cuda_version() -> str:
+    try:
+        library = ctypes.CDLL("libcudart.so.12")
+        value = ctypes.c_int()
+        status = int(library.cudaRuntimeGetVersion(ctypes.byref(value)))
+    except (AttributeError, OSError) as exc:
+        raise VerificationError(f"Unable to load CUDA 12 runtime: {exc}") from exc
+    if status != 0 or value.value <= 0:
+        raise VerificationError(f"cudaRuntimeGetVersion failed with status {status}")
+    return f"{value.value // 1000}.{(value.value % 1000) // 10}.{value.value % 10}"
+
+
+def cudnn_version() -> str:
+    try:
+        library = ctypes.CDLL("libcudnn.so.9")
+        getter = library.cudnnGetVersion
+        getter.restype = ctypes.c_size_t
+        raw = int(getter())
+    except (AttributeError, OSError) as exc:
+        raise VerificationError(f"Unable to load cuDNN 9: {exc}") from exc
+    if raw <= 0:
+        raise VerificationError("cudnnGetVersion returned an invalid value")
+    return f"{raw // 10000}.{(raw % 10000) // 100}.{raw % 100}"
+
+
+def package_version(name: str) -> str:
+    try:
+        result = subprocess.run(
+            ["dpkg-query", "-W", "-f=${Version}", name],
+            check=True,
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except (FileNotFoundError, subprocess.SubprocessError) as exc:
+        raise VerificationError(f"Unable to query {name}: {exc}") from exc
+    return result.stdout.strip()
+
+
+def gpu_details() -> list[dict[str, str]]:
+    try:
+        result = subprocess.run(
+            [
+                "nvidia-smi",
+                "--query-gpu=name,compute_cap,driver_version",
+                "--format=csv,noheader,nounits",
+            ],
+            check=True,
+            capture_output=True,
+            text=True,
+            timeout=15,
+        )
+    except (FileNotFoundError, subprocess.SubprocessError) as exc:
+        raise VerificationError(
+            "No usable NVIDIA GPU was exposed by NVIDIA Container Toolkit"
+        ) from exc
+    values: list[dict[str, str]] = []
+    supported = {(7, 5), (8, 0), (8, 6), (8, 9), (9, 0), (10, 0), (10, 3), (12, 0)}
+    for line in result.stdout.splitlines():
+        fields = [field.strip() for field in line.split(",")]
+        if len(fields) != 3:
+            raise VerificationError(f"Unexpected nvidia-smi output: {line!r}")
+        capability = version(fields[1])[:2]
+        driver = version(fields[2])
+        if capability not in supported:
+            raise VerificationError(
+                f"Unsupported Compute Capability {fields[1]} on {fields[0]}"
+            )
+        minimum = (570, 26) if capability >= (10, 0) else (535,)
+        if driver < minimum:
+            required = ".".join(str(item) for item in minimum)
+            raise VerificationError(
+                f"{fields[0]} requires Driver {required} or newer; found {fields[2]}"
+            )
+        values.append(
+            {"name": fields[0], "compute_capability": fields[1], "driver": fields[2]}
+        )
+    if not values:
+        raise VerificationError("No NVIDIA GPU was exposed to the container")
+    return values
+
+
+def verify(args: argparse.Namespace) -> dict[str, Any]:
+    if platform.system() != "Linux" or platform.machine() not in {"x86_64", "amd64"}:
+        raise VerificationError("The image supports only Linux x86_64")
+    if version(platform.python_version())[:2] != (3, 11):
+        raise VerificationError(f"Python 3.11 is required; found {platform.python_version()}")
+    try:
+        import onnxruntime as ort
+    except ImportError as exc:
+        raise VerificationError(f"Unable to import ONNX Runtime: {exc}") from exc
+    if ort.__version__ != "1.27.0":
+        raise VerificationError(f"ONNX Runtime 1.27.0 is required; found {ort.__version__}")
+    providers = list(ort.get_available_providers())
+    if args.provider not in providers:
+        raise VerificationError(f"{args.provider} is unavailable; providers={providers}")
+    if os.getenv("INSIGHTFACE_EXECUTION_PROVIDER") != args.provider:
+        raise VerificationError("Configured Execution Provider does not match the image")
+    if os.getenv("INSIGHTFACE_INFERENCE_MODE") != "onnx":
+        raise VerificationError("Published images require ONNX inference mode")
+
+    output: dict[str, Any] = {
+        "mode": args.mode,
+        "os": platform.platform(),
+        "architecture": platform.machine(),
+        "python": platform.python_version(),
+        "onnx_runtime": ort.__version__,
+        "available_execution_providers": providers,
+        "required_execution_provider": args.provider,
+    }
+    if args.provider == "CUDAExecutionProvider":
+        cuda = cuda_version()
+        cudnn = cudnn_version()
+        package = package_version("libcudnn9-cuda-12")
+        if version(cuda)[:2] != (12, 9) or os.getenv("CUDA_VERSION") != "12.9.1":
+            raise VerificationError(
+                f"CUDA Runtime/Image 12.9.1 is required; found {cuda}/{os.getenv('CUDA_VERSION')}"
+            )
+        if cudnn != "9.24.0" or package != "9.24.0.43-1":
+            raise VerificationError(
+                f"cuDNN 9.24.0 package 9.24.0.43-1 is required; found {cudnn}/{package}"
+            )
+        if os.getenv("INSIGHTFACE_STRICT_CUDA") != "1":
+            raise VerificationError("The CUDA image requires strict Provider auditing")
+        output.update(
+            {
+                "cuda_runtime": cuda,
+                "cuda_image": os.getenv("CUDA_VERSION"),
+                "cudnn": cudnn,
+                "cudnn_package": package,
+                "gpus": gpu_details() if args.mode == "startup" else [],
+                "strict_cuda": True,
+                "model_session_validation": "application_lifespan_before_readiness",
+            }
+        )
+    return output
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mode", required=True, choices=("build", "startup"))
+    parser.add_argument("--provider", required=True)
+    parser.add_argument("command", nargs=argparse.REMAINDER)
+    args = parser.parse_args()
+    try:
+        details = verify(args)
+    except VerificationError as exc:
+        print(f"InsightFace runtime verification failed: {exc}", file=sys.stderr, flush=True)
+        return 1
+    print("InsightFace runtime verification: " + json.dumps(details, sort_keys=True), flush=True)
+    command = list(args.command)
+    if command and command[0] == "--":
+        command.pop(0)
+    if args.mode == "startup" and not command:
+        print("No server command supplied", file=sys.stderr)
+        return 1
+    if command:
+        os.execvp(command[0], command)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/server/sdk/python/README.md b/server/sdk/python/README.md
new file mode 100644
index 000000000..6d27c5bea
--- /dev/null
+++ b/server/sdk/python/README.md
@@ -0,0 +1,37 @@
+# InsightFace Server Python client
+
+The lightweight client for the self-hosted InsightFace Server REST API. It uses
+`httpx`, contains no inference runtime, and accepts image paths, bytes, or binary
+file-like objects.
+
+```bash
+python -m pip install ./server/sdk/python
+```
+
+```python
+from insightface_server import Client
+
+with Client("http://localhost:8080", api_key="replace-me") as client:
+    faces = client.detect("photo.jpg")
+    print(faces.faces)
+```
+
+The system detection profile is startup-only. A Collection copies it at
+creation and may override input sizes, detector/NMS thresholds, and the
+`largest` or `center_largest` single-face strategy. Pass `collection=` to
+stateless Detect, Compare, or Embeddings calls to use that Collection profile.
+
+Trusted upstream extractors may pass `external_embeddings` together with the
+required images and the Collection's `embedding_contract_id`. This selects
+`external_trusted`: image detection and quality review still run, while the
+server neither re-extracts nor falls back to another feature.
+
+Persistent RTSP monitoring is also available through `create_monitor`,
+`update_monitor`, `monitor_state`, and cursor-based `monitor_events`. Monitor
+preview is off by default; recognition and in-memory events do not require it.
+The client waits up to 65 seconds by default, slightly longer than the server's
+60-second request deadline. Pass `timeout=` to `Client` when an application
+needs a different fail-fast policy.
+
+See `server/docs/user-guide.md` for complete SDK and operating workflows, and
+`server/docs/api.md` for the full HTTP contract.
diff --git a/server/sdk/python/pyproject.toml b/server/sdk/python/pyproject.toml
new file mode 100644
index 000000000..6baf862fd
--- /dev/null
+++ b/server/sdk/python/pyproject.toml
@@ -0,0 +1,31 @@
+[build-system]
+requires = ["setuptools==83.0.0", "wheel==0.46.2"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "insightface-server-client"
+version = "0.2.0"
+description = "Typed Python client for InsightFace Server"
+readme = "README.md"
+requires-python = ">=3.9"
+authors = [{ name = "InsightFace contributors" }]
+dependencies = ["httpx==0.28.1"]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Intended Audience :: Developers",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3 :: Only",
+  "Typing :: Typed",
+]
+
+[project.optional-dependencies]
+test = ["pytest==9.0.3"]
+
+[tool.setuptools]
+package-dir = { "" = "src" }
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-data]
+insightface_server = ["py.typed"]
diff --git a/server/sdk/python/src/insightface_server/__init__.py b/server/sdk/python/src/insightface_server/__init__.py
new file mode 100644
index 000000000..89ffb299b
--- /dev/null
+++ b/server/sdk/python/src/insightface_server/__init__.py
@@ -0,0 +1,101 @@
+"""Python SDK for InsightFace Server."""
+
+from .client import (
+    Client,
+    EmbeddingMode,
+    ExternalEmbedding,
+    ImageInput,
+    ReviewMode,
+    SearchLoadPolicy,
+    SearchProfile,
+)
+from .exceptions import (
+    AuthenticationError,
+    ConflictError,
+    InsightFaceServerError,
+    NotFoundError,
+    PayloadTooLargeError,
+    RateLimitError,
+    ServerError,
+    ServiceUnavailableError,
+    TransportError,
+    ValidationError,
+)
+from .results import (
+    ApiResult,
+    BoundingBox,
+    Collection,
+    CollectionPage,
+    CollectionResult,
+    CompareResult,
+    DetectResult,
+    EmbeddingsResult,
+    FaceObservation,
+    FacePage,
+    FaceRegistrationResult,
+    FaceSample,
+    HealthResult,
+    Match,
+    ModelsResult,
+    Monitor,
+    MonitorEventPage,
+    MonitorPage,
+    MonitorResult,
+    MonitorStateResult,
+    Person,
+    PersonPage,
+    PersonRegistrationResult,
+    PersonResult,
+    Quality,
+    SearchResult,
+    SystemResult,
+)
+
+__all__ = [
+    "ApiResult",
+    "AuthenticationError",
+    "BoundingBox",
+    "Client",
+    "CollectionPage",
+    "CollectionResult",
+    "Collection",
+    "CompareResult",
+    "ConflictError",
+    "DetectResult",
+    "EmbeddingMode",
+    "EmbeddingsResult",
+    "ExternalEmbedding",
+    "FacePage",
+    "FaceObservation",
+    "FaceRegistrationResult",
+    "FaceSample",
+    "HealthResult",
+    "ImageInput",
+    "InsightFaceServerError",
+    "NotFoundError",
+    "ModelsResult",
+    "Match",
+    "Monitor",
+    "MonitorEventPage",
+    "MonitorPage",
+    "MonitorResult",
+    "MonitorStateResult",
+    "PayloadTooLargeError",
+    "PersonPage",
+    "Person",
+    "PersonRegistrationResult",
+    "PersonResult",
+    "RateLimitError",
+    "Quality",
+    "ReviewMode",
+    "SearchResult",
+    "SearchLoadPolicy",
+    "SearchProfile",
+    "ServerError",
+    "ServiceUnavailableError",
+    "SystemResult",
+    "TransportError",
+    "ValidationError",
+]
+
+__version__ = "0.2.0"
diff --git a/server/sdk/python/src/insightface_server/client.py b/server/sdk/python/src/insightface_server/client.py
new file mode 100644
index 000000000..4170fac32
--- /dev/null
+++ b/server/sdk/python/src/insightface_server/client.py
@@ -0,0 +1,883 @@
+"""Synchronous client for the InsightFace Server v1 REST API."""
+
+from __future__ import annotations
+
+import json
+import math
+import mimetypes
+from collections.abc import Iterable, Mapping
+from pathlib import Path
+from typing import (
+    Any,
+    BinaryIO,
+    Dict,
+    Literal,
+    Optional,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+    cast,
+)
+from urllib.parse import quote
+
+import httpx
+
+from .exceptions import (
+    AuthenticationError,
+    ConflictError,
+    InsightFaceServerError,
+    NotFoundError,
+    PayloadTooLargeError,
+    RateLimitError,
+    ServerError,
+    ServiceUnavailableError,
+    TransportError,
+    ValidationError,
+)
+from .results import (
+    ApiResult,
+    CollectionPage,
+    CollectionResult,
+    CompareResult,
+    DetectResult,
+    EmbeddingsResult,
+    FacePage,
+    FaceRegistrationResult,
+    HealthResult,
+    ModelsResult,
+    MonitorEventPage,
+    MonitorPage,
+    MonitorResult,
+    MonitorStateResult,
+    PersonPage,
+    PersonRegistrationResult,
+    PersonResult,
+    SearchResult,
+    SystemResult,
+)
+
+ImageInput = Union[str, Path, bytes, bytearray, memoryview, BinaryIO]
+Upload = Tuple[str, bytes, str]
+ResultT = TypeVar("ResultT", bound=ApiResult)
+SearchProfile = Literal[
+    "fp32_v1", "fp16_v1", "bf16_v1", "int8_x736_v1", "int8_x1000_v1"
+]
+SearchLoadPolicy = Literal["eager", "lazy"]
+ReviewMode = Literal["off", "standard", "strict"]
+EmbeddingMode = Literal["server", "external_trusted"]
+SingleFaceSelection = Literal["largest", "center_largest"]
+ExternalEmbedding = Iterable[float]
+_UNSET = object()
+
+
+class Client:
+    """A small synchronous client for InsightFace Server.
+
+    Args:
+        base_url: Server origin, for example ``http://localhost:8080``.
+        api_key: Optional bearer token. Omit only when server authentication is
+            explicitly disabled for development.
+        timeout: Overall HTTP timeout in seconds or an ``httpx.Timeout``.
+        transport: Optional httpx transport, primarily useful for tests.
+    """
+
+    def __init__(
+        self,
+        base_url: str,
+        *,
+        api_key: Optional[str] = None,
+        timeout: Union[float, httpx.Timeout] = 65.0,
+        transport: Optional[httpx.BaseTransport] = None,
+    ) -> None:
+        normalized = base_url.rstrip("/")
+        if not normalized:
+            raise ValueError("base_url must not be empty")
+        headers = {
+            "Accept": "application/json",
+            "User-Agent": "insightface-server-python/0.2.0",
+        }
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+        self._http = httpx.Client(
+            base_url=normalized,
+            headers=headers,
+            timeout=timeout,
+            transport=transport,
+            follow_redirects=False,
+        )
+
+    def __enter__(self) -> "Client":
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        self.close()
+
+    def close(self) -> None:
+        self._http.close()
+
+    def health(self) -> HealthResult:
+        return self._request("GET", "/v1/health", result_type=HealthResult)
+
+    def system(self) -> SystemResult:
+        return self._request("GET", "/v1/system", result_type=SystemResult)
+
+    def models(self) -> ModelsResult:
+        return self._request("GET", "/v1/models", result_type=ModelsResult)
+
+    def create_monitor(
+        self,
+        monitor_id: str,
+        *,
+        name: str,
+        rtsp_url: str,
+        collection: str,
+        description: str = "",
+        enabled: bool = True,
+        inference_fps: float = 2.0,
+        match_threshold: Optional[float] = None,
+        event_buffer_size: int = 1000,
+        confirm_frames: int = 3,
+        absence_timeout_seconds: float = 3.0,
+        cooldown_seconds: float = 10.0,
+        emit_unknown: bool = True,
+        preview_enabled: bool = False,
+    ) -> MonitorResult:
+        return self._request(
+            "POST",
+            "/v1/monitors",
+            json={
+                "id": monitor_id,
+                "name": name,
+                "description": description,
+                "enabled": enabled,
+                "source": {"type": "rtsp", "url": rtsp_url},
+                "collection_id": collection,
+                "inference_fps": inference_fps,
+                "match_threshold": match_threshold,
+                "event_buffer_size": event_buffer_size,
+                "event_policy": {
+                    "confirm_frames": confirm_frames,
+                    "absence_timeout_seconds": absence_timeout_seconds,
+                    "cooldown_seconds": cooldown_seconds,
+                    "emit_unknown": emit_unknown,
+                },
+                "preview_enabled": preview_enabled,
+            },
+            result_type=MonitorResult,
+        )
+
+    def list_monitors(
+        self,
+        *,
+        limit: int = 50,
+        cursor: Optional[str] = None,
+    ) -> MonitorPage:
+        return self._request(
+            "GET",
+            "/v1/monitors",
+            params=self._without_none(limit=limit, cursor=cursor),
+            result_type=MonitorPage,
+        )
+
+    def get_monitor(self, monitor_id: str) -> MonitorResult:
+        return self._request(
+            "GET",
+            self._monitor_path(monitor_id),
+            result_type=MonitorResult,
+        )
+
+    def update_monitor(
+        self,
+        monitor_id: str,
+        *,
+        name: object = _UNSET,
+        description: object = _UNSET,
+        enabled: object = _UNSET,
+        rtsp_url: object = _UNSET,
+        collection: object = _UNSET,
+        inference_fps: object = _UNSET,
+        match_threshold: object = _UNSET,
+        event_buffer_size: object = _UNSET,
+        confirm_frames: object = _UNSET,
+        absence_timeout_seconds: object = _UNSET,
+        cooldown_seconds: object = _UNSET,
+        emit_unknown: object = _UNSET,
+        preview_enabled: object = _UNSET,
+    ) -> MonitorResult:
+        payload = {
+            key: value
+            for key, value in {
+                "name": name,
+                "description": description,
+                "enabled": enabled,
+                "collection_id": collection,
+                "inference_fps": inference_fps,
+                "match_threshold": match_threshold,
+                "event_buffer_size": event_buffer_size,
+                "preview_enabled": preview_enabled,
+            }.items()
+            if value is not _UNSET
+        }
+        if rtsp_url is not _UNSET:
+            payload["source"] = {"type": "rtsp", "url": rtsp_url}
+        policy = {
+            key: value
+            for key, value in {
+                "confirm_frames": confirm_frames,
+                "absence_timeout_seconds": absence_timeout_seconds,
+                "cooldown_seconds": cooldown_seconds,
+                "emit_unknown": emit_unknown,
+            }.items()
+            if value is not _UNSET
+        }
+        if policy:
+            payload["event_policy"] = policy
+        if not payload:
+            raise ValueError("at least one field must be supplied")
+        return self._request(
+            "PATCH",
+            self._monitor_path(monitor_id),
+            json=payload,
+            result_type=MonitorResult,
+        )
+
+    def delete_monitor(self, monitor_id: str) -> ApiResult:
+        return self._request("DELETE", self._monitor_path(monitor_id))
+
+    def monitor_state(self, monitor_id: str) -> MonitorStateResult:
+        return self._request(
+            "GET",
+            f"{self._monitor_path(monitor_id)}/state",
+            result_type=MonitorStateResult,
+        )
+
+    def monitor_events(
+        self,
+        monitor_id: str,
+        *,
+        limit: int = 100,
+        cursor: Optional[str] = None,
+    ) -> MonitorEventPage:
+        return self._request(
+            "GET",
+            f"{self._monitor_path(monitor_id)}/events",
+            params=self._without_none(limit=limit, cursor=cursor),
+            result_type=MonitorEventPage,
+        )
+
+    def detect(
+        self,
+        image: ImageInput,
+        *,
+        max_faces: Optional[int] = None,
+        collection: Optional[str] = None,
+    ) -> DetectResult:
+        data = self._without_none(max_faces=max_faces, collection_id=collection)
+        return self._request(
+            "POST",
+            "/v1/detect",
+            data=data,
+            files={"image": self._prepare_image(image, "image.jpg")},
+            result_type=DetectResult,
+        )
+
+    def compare(
+        self,
+        source: ImageInput,
+        target: ImageInput,
+        *,
+        threshold: Optional[float] = None,
+        collection: Optional[str] = None,
+    ) -> CompareResult:
+        return self._request(
+            "POST",
+            "/v1/compare",
+            data=self._without_none(threshold=threshold, collection_id=collection),
+            files={
+                "source": self._prepare_image(source, "source.jpg"),
+                "target": self._prepare_image(target, "target.jpg"),
+            },
+            result_type=CompareResult,
+        )
+
+    def embeddings(
+        self,
+        image: ImageInput,
+        *,
+        collection: Optional[str] = None,
+    ) -> EmbeddingsResult:
+        return self._request(
+            "POST",
+            "/v1/embeddings",
+            data=self._without_none(collection_id=collection),
+            files={"image": self._prepare_image(image, "image.jpg")},
+            result_type=EmbeddingsResult,
+        )
+
+    def create_collection(
+        self,
+        collection_id: str,
+        *,
+        name: str,
+        description: str = "",
+        threshold: float = 0.4,
+        metadata: Optional[Mapping[str, Any]] = None,
+        save_face_crops: Optional[bool] = None,
+        search_profile: Optional[SearchProfile] = None,
+        capacity_rows: Optional[int] = None,
+        max_faces_per_person: Optional[int] = None,
+        load_policy: Optional[SearchLoadPolicy] = None,
+        detector_input_sizes: Optional[Iterable[Tuple[int, int]]] = None,
+        detector_threshold: Optional[float] = None,
+        detector_nms_threshold: Optional[float] = None,
+        single_face_selection: Optional[SingleFaceSelection] = None,
+    ) -> CollectionResult:
+        payload = {
+            "id": collection_id,
+            "name": name,
+            "description": description,
+            "threshold": threshold,
+            "metadata": dict(metadata or {}),
+        }
+        if save_face_crops is not None:
+            payload["save_face_crops"] = save_face_crops
+        search = self._without_none(
+            profile=search_profile,
+            capacity_rows=capacity_rows,
+            max_faces_per_person=max_faces_per_person,
+            load_policy=load_policy,
+        )
+        if search:
+            payload["search"] = search
+        detection = self._without_none(
+            input_sizes=(
+                [list(size) for size in detector_input_sizes]
+                if detector_input_sizes is not None
+                else None
+            ),
+            threshold=detector_threshold,
+            nms_threshold=detector_nms_threshold,
+            single_face_selection=single_face_selection,
+        )
+        if detection:
+            payload["detection"] = detection
+        return self._request(
+            "POST", "/v1/collections", json=payload, result_type=CollectionResult
+        )
+
+    def list_collections(
+        self, *, limit: int = 50, cursor: Optional[str] = None
+    ) -> CollectionPage:
+        return self._request(
+            "GET",
+            "/v1/collections",
+            params=self._without_none(limit=limit, cursor=cursor),
+            result_type=CollectionPage,
+        )
+
+    def get_collection(self, collection_id: str) -> CollectionResult:
+        return self._request(
+            "GET", self._collection_path(collection_id), result_type=CollectionResult
+        )
+
+    def update_collection(
+        self,
+        collection_id: str,
+        *,
+        name: object = _UNSET,
+        description: object = _UNSET,
+        threshold: object = _UNSET,
+        metadata: object = _UNSET,
+        save_face_crops: object = _UNSET,
+        capacity_rows: object = _UNSET,
+        max_faces_per_person: object = _UNSET,
+        load_policy: object = _UNSET,
+        detector_input_sizes: object = _UNSET,
+        detector_threshold: object = _UNSET,
+        detector_nms_threshold: object = _UNSET,
+        single_face_selection: object = _UNSET,
+    ) -> CollectionResult:
+        payload = {
+            key: value
+            for key, value in {
+                "name": name,
+                "description": description,
+                "threshold": threshold,
+                "metadata": metadata,
+                "save_face_crops": save_face_crops,
+            }.items()
+            if value is not _UNSET
+        }
+        search = {
+            key: value
+            for key, value in {
+                "capacity_rows": capacity_rows,
+                "max_faces_per_person": max_faces_per_person,
+                "load_policy": load_policy,
+            }.items()
+            if value is not _UNSET
+        }
+        if search:
+            payload["search"] = search
+        detection = {
+            key: value
+            for key, value in {
+                "input_sizes": detector_input_sizes,
+                "threshold": detector_threshold,
+                "nms_threshold": detector_nms_threshold,
+                "single_face_selection": single_face_selection,
+            }.items()
+            if value is not _UNSET
+        }
+        if detection:
+            if (
+                "input_sizes" in detection
+                and detection["input_sizes"] is not None
+            ):
+                input_sizes = cast(
+                    Iterable[Tuple[int, int]], detection["input_sizes"]
+                )
+                detection["input_sizes"] = [
+                    list(size) for size in input_sizes
+                ]
+            payload["detection"] = detection
+        if not payload:
+            raise ValueError("at least one field must be supplied")
+        return self._request(
+            "PATCH",
+            self._collection_path(collection_id),
+            json=payload,
+            result_type=CollectionResult,
+        )
+
+    def delete_collection(self, collection_id: str, *, force: bool = False) -> ApiResult:
+        return self._request(
+            "DELETE",
+            self._collection_path(collection_id),
+            params={"force": str(force).lower()},
+        )
+
+    def create_person(
+        self,
+        collection: str,
+        *,
+        images: Iterable[ImageInput],
+        person_id: Optional[str] = None,
+        name: Optional[str] = None,
+        external_id: Optional[str] = None,
+        metadata: Optional[Mapping[str, Any]] = None,
+        review_mode: ReviewMode = "off",
+        external_embeddings: Optional[Iterable[ExternalEmbedding]] = None,
+        embedding_contract_id: Optional[str] = None,
+    ) -> PersonRegistrationResult:
+        uploads = [
+            self._prepare_image(image, f"image-{index}.jpg")
+            for index, image in enumerate(images)
+        ]
+        if not uploads:
+            raise ValueError("images must contain at least one image")
+        data = self._without_none(id=person_id, name=name, external_id=external_id)
+        data["metadata"] = json.dumps(
+            dict(metadata or {}), ensure_ascii=False, separators=(",", ":")
+        )
+        data.update(
+            self._enrollment_fields(
+                image_count=len(uploads),
+                review_mode=review_mode,
+                external_embeddings=external_embeddings,
+                embedding_contract_id=embedding_contract_id,
+            )
+        )
+        files = [("images", upload) for upload in uploads]
+        return self._request(
+            "POST",
+            f"{self._collection_path(collection)}/persons",
+            data=data,
+            files=files,
+            result_type=PersonRegistrationResult,
+        )
+
+    def add_person(
+        self,
+        collection: str,
+        *,
+        images: Iterable[ImageInput],
+        person_id: Optional[str] = None,
+        name: Optional[str] = None,
+        external_id: Optional[str] = None,
+        metadata: Optional[Mapping[str, Any]] = None,
+        review_mode: ReviewMode = "off",
+        external_embeddings: Optional[Iterable[ExternalEmbedding]] = None,
+        embedding_contract_id: Optional[str] = None,
+    ) -> PersonRegistrationResult:
+        """Alias for :meth:`create_person` matching the common SDK workflow."""
+
+        return self.create_person(
+            collection,
+            images=images,
+            person_id=person_id,
+            name=name,
+            external_id=external_id,
+            metadata=metadata,
+            review_mode=review_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+        )
+
+    def list_persons(
+        self,
+        collection: str,
+        *,
+        limit: int = 50,
+        cursor: Optional[str] = None,
+        search: Optional[str] = None,
+    ) -> PersonPage:
+        return self._request(
+            "GET",
+            f"{self._collection_path(collection)}/persons",
+            params=self._without_none(limit=limit, cursor=cursor, search=search),
+            result_type=PersonPage,
+        )
+
+    def get_person(self, collection: str, person_id: str) -> PersonResult:
+        return self._request(
+            "GET", self._person_path(collection, person_id), result_type=PersonResult
+        )
+
+    def update_person(
+        self,
+        collection: str,
+        person_id: str,
+        *,
+        name: object = _UNSET,
+        external_id: object = _UNSET,
+        metadata: object = _UNSET,
+    ) -> PersonResult:
+        payload = self._patch_payload(name=name, external_id=external_id, metadata=metadata)
+        return self._request(
+            "PATCH",
+            self._person_path(collection, person_id),
+            json=payload,
+            result_type=PersonResult,
+        )
+
+    def delete_person(self, collection: str, person_id: str) -> ApiResult:
+        return self._request("DELETE", self._person_path(collection, person_id))
+
+    def list_faces(
+        self,
+        collection: str,
+        person_id: str,
+        *,
+        limit: int = 50,
+        cursor: Optional[str] = None,
+    ) -> FacePage:
+        return self._request(
+            "GET",
+            f"{self._person_path(collection, person_id)}/faces",
+            params=self._without_none(limit=limit, cursor=cursor),
+            result_type=FacePage,
+        )
+
+    def add_faces(
+        self,
+        collection: str,
+        person_id: str,
+        images: Iterable[ImageInput],
+        *,
+        review_mode: ReviewMode = "off",
+        external_embeddings: Optional[Iterable[ExternalEmbedding]] = None,
+        embedding_contract_id: Optional[str] = None,
+    ) -> FaceRegistrationResult:
+        """Register one or more additional samples for an existing person."""
+
+        uploads = [
+            self._prepare_image(image, f"image-{index}.jpg")
+            for index, image in enumerate(images)
+        ]
+        if not uploads:
+            raise ValueError("images must contain at least one image")
+        data = self._enrollment_fields(
+            image_count=len(uploads),
+            review_mode=review_mode,
+            external_embeddings=external_embeddings,
+            embedding_contract_id=embedding_contract_id,
+        )
+        return self._request(
+            "POST",
+            f"{self._person_path(collection, person_id)}/faces",
+            data=data,
+            files=[("images", upload) for upload in uploads],
+            result_type=FaceRegistrationResult,
+        )
+
+    def delete_face(self, collection: str, person_id: str, face_id: str) -> ApiResult:
+        return self._request(
+            "DELETE",
+            f"{self._person_path(collection, person_id)}/faces/{quote(face_id, safe='')}",
+        )
+
+    def get_face_crop(self, collection: str, person_id: str, face_id: str) -> bytes:
+        """Download a stored 112x112 bounding-box face crop as JPEG bytes.
+
+        A crop exists only when the Collection had ``save_face_crops`` enabled
+        when this FaceSample was registered. Original uploads are never
+        returned by this endpoint.
+        """
+
+        path = (
+            f"{self._person_path(collection, person_id)}/faces/"
+            f"{quote(face_id, safe='')}/image"
+        )
+        try:
+            response = self._http.get(path, headers={"Accept": "image/jpeg"})
+        except httpx.HTTPError as exc:
+            raise TransportError(
+                "Unable to complete the request.", code="transport_error"
+            ) from exc
+        if not response.is_success:
+            self._raise_api_error(response)
+        media_type = response.headers.get("content-type", "").split(";", 1)[0].lower()
+        if media_type != "image/jpeg" or not response.content:
+            raise ServerError(
+                "Server returned an invalid face crop response.",
+                code="invalid_response",
+                status_code=response.status_code,
+                request_id=response.headers.get("x-request-id"),
+            )
+        return bytes(response.content)
+
+    def search(
+        self,
+        collection: str,
+        image: ImageInput,
+        *,
+        limit: int = 5,
+        threshold: Optional[float] = None,
+    ) -> SearchResult:
+        data = self._without_none(limit=limit, threshold=threshold)
+        return self._request(
+            "POST",
+            f"{self._collection_path(collection)}/search",
+            data=data,
+            files={"image": self._prepare_image(image, "image.jpg")},
+            result_type=SearchResult,
+        )
+
+    @staticmethod
+    def _collection_path(collection_id: str) -> str:
+        return f"/v1/collections/{quote(collection_id, safe='')}"
+
+    @staticmethod
+    def _monitor_path(monitor_id: str) -> str:
+        return f"/v1/monitors/{quote(monitor_id, safe='')}"
+
+    @classmethod
+    def _person_path(cls, collection: str, person_id: str) -> str:
+        return f"{cls._collection_path(collection)}/persons/{quote(person_id, safe='')}"
+
+    @staticmethod
+    def _without_none(**values: Any) -> Dict[str, Any]:
+        return {key: value for key, value in values.items() if value is not None}
+
+    @staticmethod
+    def _enrollment_fields(
+        *,
+        image_count: int,
+        review_mode: ReviewMode,
+        external_embeddings: Optional[Iterable[ExternalEmbedding]],
+        embedding_contract_id: Optional[str],
+    ) -> Dict[str, str]:
+        """Build multipart fields for server or trusted-external enrollment.
+
+        Conversion uses only Python's iteration protocol, so lists, tuples and
+        NumPy arrays work without making NumPy an SDK dependency.
+        """
+
+        fields: Dict[str, str] = {"review_mode": review_mode}
+        if external_embeddings is None:
+            if embedding_contract_id is not None:
+                raise ValueError(
+                    "embedding_contract_id requires external_embeddings"
+                )
+            fields["embedding_mode"] = "server"
+            return fields
+
+        contract_id = str(embedding_contract_id or "").strip()
+        if not contract_id:
+            raise ValueError(
+                "embedding_contract_id is required with external_embeddings"
+            )
+
+        vectors = []
+        for index, vector in enumerate(external_embeddings):
+            try:
+                values = [float(value) for value in vector]
+            except (TypeError, ValueError) as exc:
+                raise ValueError(
+                    f"external_embeddings[{index}] must be a numeric vector"
+                ) from exc
+            if not values:
+                raise ValueError(
+                    f"external_embeddings[{index}] must not be empty"
+                )
+            if not all(math.isfinite(value) for value in values):
+                raise ValueError(
+                    f"external_embeddings[{index}] contains NaN or infinity"
+                )
+            norm = math.sqrt(sum(value * value for value in values))
+            if not math.isfinite(norm) or abs(norm - 1.0) > 2e-4:
+                raise ValueError(
+                    f"external_embeddings[{index}] must be L2-normalized "
+                    "within 0.0002 of unit norm"
+                )
+            vectors.append(values)
+
+        if len(vectors) != image_count:
+            raise ValueError(
+                "external_embeddings count must equal images count "
+                f"({len(vectors)} != {image_count})"
+            )
+
+        fields.update(
+            embedding_mode="external_trusted",
+            embedding_contract_id=contract_id,
+            external_embeddings=json.dumps(vectors, separators=(",", ":")),
+        )
+        return fields
+
+    @staticmethod
+    def _patch_payload(**values: object) -> Dict[str, Any]:
+        payload = {key: value for key, value in values.items() if value is not _UNSET}
+        if not payload:
+            raise ValueError("at least one field must be supplied")
+        return payload
+
+    @staticmethod
+    def _prepare_image(image: ImageInput, fallback_name: str) -> Upload:
+        filename = fallback_name
+        if isinstance(image, (str, Path)):
+            path = Path(image)
+            filename = path.name
+            try:
+                content = path.read_bytes()
+            except OSError as exc:
+                raise ValueError(f"cannot read image file: {path}") from exc
+        elif isinstance(image, (bytes, bytearray, memoryview)):
+            content = bytes(image)
+        elif hasattr(image, "read"):
+            stream = image
+            stream_name = getattr(stream, "name", None)
+            if isinstance(stream_name, str):
+                filename = Path(stream_name).name
+            old_position: Optional[int] = None
+            if hasattr(stream, "tell"):
+                try:
+                    old_position = stream.tell()
+                except (OSError, ValueError):
+                    # Non-seekable streams are still valid image inputs.
+                    old_position = None
+            try:
+                content = stream.read()
+            except (OSError, ValueError) as exc:
+                raise ValueError("cannot read image stream") from exc
+            if old_position is not None and hasattr(stream, "seek"):
+                try:
+                    stream.seek(old_position)
+                except (OSError, ValueError):
+                    pass
+            if not isinstance(content, (bytes, bytearray, memoryview)):
+                raise TypeError("image file-like object must return bytes")
+            content = bytes(content)
+        else:
+            raise TypeError("image must be a path, bytes, or binary file-like object")
+        if not content:
+            raise ValueError("image must not be empty")
+        content_type = mimetypes.guess_type(filename)[0] or "application/octet-stream"
+        return filename, content, content_type
+
+    def _request(
+        self,
+        method: str,
+        path: str,
+        *,
+        result_type: Type[ResultT] = ApiResult,  # type: ignore[assignment]
+        **kwargs: Any,
+    ) -> ResultT:
+        try:
+            response = self._http.request(method, path, **kwargs)
+        except httpx.HTTPError as exc:
+            raise TransportError(
+                "Unable to complete the request.", code="transport_error"
+            ) from exc
+        if not response.is_success:
+            self._raise_api_error(response)
+        body: Dict[str, Any]
+        if response.status_code == 204 or not response.content:
+            body = {}
+        else:
+            try:
+                decoded = response.json()
+            except ValueError as exc:
+                raise ServerError(
+                    "Server returned an invalid JSON response.",
+                    code="invalid_response",
+                    status_code=response.status_code,
+                    request_id=response.headers.get("x-request-id"),
+                ) from exc
+            if not isinstance(decoded, dict):
+                raise ServerError(
+                    "Server returned an unexpected JSON response.",
+                    code="invalid_response",
+                    status_code=response.status_code,
+                    request_id=response.headers.get("x-request-id"),
+                )
+            body = decoded
+        request_id = response.headers.get("x-request-id")
+        if not request_id:
+            candidate = body.get("request_id")
+            request_id = candidate if isinstance(candidate, str) else None
+        return result_type(body, response.status_code, request_id)
+
+    @staticmethod
+    def _raise_api_error(response: httpx.Response) -> None:
+        request_id = response.headers.get("x-request-id")
+        code = "http_error"
+        message = f"Server returned HTTP {response.status_code}."
+        details: Mapping[str, Any] = {}
+        try:
+            body = response.json()
+        except ValueError:
+            body = None
+        if isinstance(body, dict):
+            candidate_request_id = body.get("request_id")
+            if not request_id and isinstance(candidate_request_id, str):
+                request_id = candidate_request_id
+            error = body.get("error")
+            if isinstance(error, dict):
+                if isinstance(error.get("code"), str):
+                    code = error["code"]
+                if isinstance(error.get("message"), str):
+                    message = error["message"]
+                if isinstance(error.get("details"), dict):
+                    details = error["details"]
+        exception_type: Type[InsightFaceServerError]
+        exception_type = {
+            400: ValidationError,
+            401: AuthenticationError,
+            403: AuthenticationError,
+            404: NotFoundError,
+            409: ConflictError,
+            413: PayloadTooLargeError,
+            422: ValidationError,
+            429: RateLimitError,
+            503: ServiceUnavailableError,
+        }.get(response.status_code, ServerError)
+        raise exception_type(
+            message,
+            code=code,
+            status_code=response.status_code,
+            request_id=request_id,
+            details=details,
+        )
diff --git a/server/sdk/python/src/insightface_server/exceptions.py b/server/sdk/python/src/insightface_server/exceptions.py
new file mode 100644
index 000000000..7c925df51
--- /dev/null
+++ b/server/sdk/python/src/insightface_server/exceptions.py
@@ -0,0 +1,69 @@
+"""Typed errors raised by the InsightFace Server client."""
+
+from __future__ import annotations
+
+from typing import Any, Mapping, Optional
+
+
+class InsightFaceServerError(Exception):
+    """Base class for API and client errors.
+
+    Attributes are deliberately safe to log: response bodies containing image
+    or embedding data are not retained.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        code: str = "client_error",
+        status_code: Optional[int] = None,
+        request_id: Optional[str] = None,
+        details: Optional[Mapping[str, Any]] = None,
+    ) -> None:
+        super().__init__(message)
+        self.message = message
+        self.code = code
+        self.status_code = status_code
+        self.request_id = request_id
+        self.details = dict(details or {})
+
+    def __str__(self) -> str:
+        suffix = f" (request_id={self.request_id})" if self.request_id else ""
+        return f"{self.code}: {self.message}{suffix}"
+
+
+class TransportError(InsightFaceServerError):
+    """The server could not be reached or the request timed out."""
+
+
+class AuthenticationError(InsightFaceServerError):
+    """The API key is missing or invalid."""
+
+
+class NotFoundError(InsightFaceServerError):
+    """The requested resource does not exist."""
+
+
+class ConflictError(InsightFaceServerError):
+    """The requested change conflicts with current server state."""
+
+
+class PayloadTooLargeError(InsightFaceServerError):
+    """An uploaded image or request body exceeds a server limit."""
+
+
+class ValidationError(InsightFaceServerError):
+    """The request or image cannot be processed as supplied."""
+
+
+class RateLimitError(InsightFaceServerError):
+    """The server rejected the request due to rate limiting."""
+
+
+class ServiceUnavailableError(InsightFaceServerError):
+    """The model, GPU, or another required service is unavailable."""
+
+
+class ServerError(InsightFaceServerError):
+    """The server returned an unexpected error."""
diff --git a/server/sdk/python/src/insightface_server/py.typed b/server/sdk/python/src/insightface_server/py.typed
new file mode 100644
index 000000000..8b1378917
--- /dev/null
+++ b/server/sdk/python/src/insightface_server/py.typed
@@ -0,0 +1 @@
+
diff --git a/server/sdk/python/src/insightface_server/results.py b/server/sdk/python/src/insightface_server/results.py
new file mode 100644
index 000000000..94ae82aee
--- /dev/null
+++ b/server/sdk/python/src/insightface_server/results.py
@@ -0,0 +1,311 @@
+"""Typed, mapping-compatible API results."""
+
+from __future__ import annotations
+
+from collections.abc import Iterator, Mapping
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, TypedDict, cast
+
+
+class PixelBox(TypedDict):
+    x: int
+    y: int
+    width: int
+    height: int
+
+
+class NormalizedBox(TypedDict):
+    left: float
+    top: float
+    width: float
+    height: float
+
+
+class BoundingBox(TypedDict):
+    pixels: PixelBox
+    normalized: NormalizedBox
+
+
+class Quality(TypedDict, total=False):
+    score: float
+    sharpness: float
+    brightness: float
+
+
+class FaceObservation(TypedDict, total=False):
+    id: str
+    bbox: BoundingBox
+    landmarks: List[List[float]]
+    detection_score: float
+    quality: Quality
+    embedding: List[float]
+
+
+class Person(TypedDict, total=False):
+    id: str
+    name: Optional[str]
+    external_id: Optional[str]
+    metadata: Dict[str, Any]
+    face_count: int
+    created_at: str
+    updated_at: str
+
+
+class Collection(TypedDict, total=False):
+    id: str
+    name: str
+    description: str
+    default_threshold: float
+    model_id: str
+    model_version: str
+    model_digest: str
+    embedding_dimension: int
+    preprocessing_version: str
+    embedding_contract_id: str
+    search_profile: str
+    capacity_rows: int
+    max_faces_per_person: int
+    load_policy: str
+    search_revision: int
+    save_face_crops: bool
+    metadata: Dict[str, Any]
+    person_count: int
+    face_count: int
+    created_at: str
+    updated_at: str
+    detection: Dict[str, Any]
+    detection_revision: int
+
+
+class FaceSample(TypedDict, total=False):
+    id: str
+    person_id: str
+    bounding_box: BoundingBox
+    landmarks: List[List[float]]
+    detection_score: float
+    quality: Quality
+    model_id: str
+    model_version: str
+    model_digest: str
+    preprocessing_version: str
+    embedding_source: str
+    embedding_contract_id: Optional[str]
+    has_crop: bool
+    created_at: str
+
+
+class Match(TypedDict, total=False):
+    person: Person
+    similarity: float
+    matched_face_id: str
+
+
+class Monitor(TypedDict, total=False):
+    id: str
+    name: str
+    description: str
+    enabled: bool
+    source: Dict[str, str]
+    collection_id: str
+    inference_fps: float
+    match_threshold: Optional[float]
+    event_buffer_size: int
+    event_policy: Dict[str, Any]
+    preview_enabled: bool
+    runtime: Dict[str, Any]
+    created_at: str
+    updated_at: str
+
+
+@dataclass(frozen=True)
+class ApiResult(Mapping[str, Any]):
+    """A successful response with request metadata.
+
+    It behaves like a read-only mapping for callers that prefer raw JSON while
+    specialized subclasses expose typed convenience properties.
+    """
+
+    data: Dict[str, Any]
+    status_code: int
+    request_id: Optional[str]
+
+    def __getitem__(self, key: str) -> Any:
+        return self.data[key]
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(self.data)
+
+    def __len__(self) -> int:
+        return len(self.data)
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Return a shallow copy of the response JSON."""
+
+        return dict(self.data)
+
+
+class DetectResult(ApiResult):
+    @property
+    def faces(self) -> List[FaceObservation]:
+        return cast(List[FaceObservation], self.data.get("faces", []))
+
+    @property
+    def processing_ms(self) -> float:
+        return float(self.data.get("processing_ms", 0.0))
+
+
+class HealthResult(ApiResult):
+    @property
+    def status(self) -> str:
+        return str(self.data.get("status", "unknown"))
+
+
+class SystemResult(ApiResult):
+    @property
+    def execution_provider(self) -> Optional[str]:
+        value = self.data.get("execution_provider")
+        return str(value) if value is not None else None
+
+
+class ModelsResult(ApiResult):
+    @property
+    def models(self) -> List[Dict[str, Any]]:
+        return cast(List[Dict[str, Any]], self.data.get("models", []))
+
+
+class CompareResult(ApiResult):
+    @property
+    def matched(self) -> bool:
+        return bool(self.data.get("matched", False))
+
+    @property
+    def similarity(self) -> float:
+        return float(self.data["similarity"])
+
+    @property
+    def threshold(self) -> float:
+        return float(self.data["threshold"])
+
+
+class EmbeddingsResult(ApiResult):
+    @property
+    def faces(self) -> List[FaceObservation]:
+        # The API represents each embedding together with its detected face.
+        return cast(List[FaceObservation], self.data.get("faces", []))
+
+
+class CollectionResult(ApiResult):
+    @property
+    def collection(self) -> Collection:
+        value = self.data.get("collection", self.data)
+        return cast(Collection, value)
+
+
+class CollectionPage(ApiResult):
+    @property
+    def collections(self) -> List[Collection]:
+        return cast(List[Collection], self.data.get("collections", []))
+
+    @property
+    def next_cursor(self) -> Optional[str]:
+        return cast(Optional[str], self.data.get("next_cursor"))
+
+
+class MonitorResult(ApiResult):
+    @property
+    def monitor(self) -> Monitor:
+        return cast(Monitor, self.data.get("monitor", self.data))
+
+
+class MonitorPage(ApiResult):
+    @property
+    def monitors(self) -> List[Monitor]:
+        return cast(List[Monitor], self.data.get("monitors", []))
+
+    @property
+    def next_cursor(self) -> Optional[str]:
+        return cast(Optional[str], self.data.get("next_cursor"))
+
+
+class MonitorStateResult(ApiResult):
+    @property
+    def state(self) -> Dict[str, Any]:
+        return cast(Dict[str, Any], self.data.get("state", self.data))
+
+
+class MonitorEventPage(ApiResult):
+    @property
+    def events(self) -> List[Dict[str, Any]]:
+        return cast(List[Dict[str, Any]], self.data.get("events", []))
+
+    @property
+    def next_cursor(self) -> Optional[str]:
+        return cast(Optional[str], self.data.get("next_cursor"))
+
+    @property
+    def has_more(self) -> bool:
+        return bool(self.data.get("has_more", False))
+
+    @property
+    def truncated(self) -> bool:
+        return bool(self.data.get("truncated", False))
+
+    @property
+    def stream_reset(self) -> bool:
+        return bool(self.data.get("stream_reset", False))
+
+
+class PersonResult(ApiResult):
+    @property
+    def person(self) -> Person:
+        return cast(Person, self.data.get("person", self.data))
+
+
+class PersonRegistrationResult(PersonResult):
+    @property
+    def faces(self) -> List[FaceSample]:
+        return cast(List[FaceSample], self.data.get("faces", []))
+
+    @property
+    def rejected_images(self) -> List[Dict[str, Any]]:
+        return cast(List[Dict[str, Any]], self.data.get("rejected_images", []))
+
+
+class FaceRegistrationResult(ApiResult):
+    @property
+    def faces(self) -> List[FaceSample]:
+        return cast(List[FaceSample], self.data.get("faces", []))
+
+    @property
+    def rejected_images(self) -> List[Dict[str, Any]]:
+        return cast(List[Dict[str, Any]], self.data.get("rejected_images", []))
+
+
+class PersonPage(ApiResult):
+    @property
+    def persons(self) -> List[Person]:
+        return cast(List[Person], self.data.get("persons", []))
+
+    @property
+    def next_cursor(self) -> Optional[str]:
+        return cast(Optional[str], self.data.get("next_cursor"))
+
+
+class FacePage(ApiResult):
+    @property
+    def faces(self) -> List[FaceSample]:
+        return cast(List[FaceSample], self.data.get("faces", []))
+
+    @property
+    def next_cursor(self) -> Optional[str]:
+        return cast(Optional[str], self.data.get("next_cursor"))
+
+
+class SearchResult(ApiResult):
+    @property
+    def matches(self) -> List[Match]:
+        return cast(List[Match], self.data.get("matches", []))
+
+    @property
+    def threshold(self) -> float:
+        return float(self.data["threshold"])
diff --git a/server/tests/api/test_detection_profiles.py b/server/tests/api/test_detection_profiles.py
new file mode 100644
index 000000000..d32f22b7c
--- /dev/null
+++ b/server/tests/api/test_detection_profiles.py
@@ -0,0 +1,174 @@
+from __future__ import annotations
+
+import numpy as np
+from fastapi.testclient import TestClient
+from insightface_server.config import DetectionProfile
+from insightface_server.inference import FaceObservation
+from insightface_server.services.core import FaceService
+
+
+def _observation(
+    bbox: tuple[float, float, float, float], embedding_index: int
+) -> FaceObservation:
+    embedding = np.zeros(512, dtype=np.float32)
+    embedding[embedding_index] = 1.0
+    left, top, right, bottom = bbox
+    landmarks = np.asarray(
+        [
+            [left + 10, top + 10],
+            [right - 10, top + 10],
+            [(left + right) / 2, (top + bottom) / 2],
+            [left + 12, bottom - 10],
+            [right - 12, bottom - 10],
+        ],
+        dtype=np.float32,
+    )
+    return FaceObservation(
+        bbox=bbox,
+        detection_score=0.99,
+        landmarks=landmarks,
+        embedding=embedding,
+        quality_score=0.9,
+        sharpness=0.9,
+        brightness=0.7,
+    )
+
+
+def test_collection_detection_profile_snapshot_patch_and_revision(
+    client: TestClient,
+) -> None:
+    system_profile = client.get("/v1/system").json()["safe_config"]["detection"]
+    inherited = client.post(
+        "/v1/collections",
+        json={"id": "inherited", "name": "Inherited"},
+    )
+    configured = client.post(
+        "/v1/collections",
+        json={
+            "id": "configured",
+            "name": "Configured",
+            "detection": {
+                "input_sizes": [[128, 128], [640, 640]],
+                "threshold": 0.42,
+                "nms_threshold": 0.35,
+                "single_face_selection": "center_largest",
+            },
+        },
+    )
+
+    assert inherited.status_code == 201, inherited.text
+    assert inherited.json()["collection"]["detection"] == system_profile
+    assert inherited.json()["collection"]["detection_revision"] == 1
+    assert configured.status_code == 201, configured.text
+    item = configured.json()["collection"]
+    assert item["detection"] == {
+        "input_sizes": [[128, 128], [640, 640]],
+        "threshold": 0.42,
+        "nms_threshold": 0.35,
+        "single_face_selection": "center_largest",
+    }
+
+    patched = client.patch(
+        "/v1/collections/configured",
+        json={"detection": {"single_face_selection": "largest"}},
+    )
+    unchanged = client.patch(
+        "/v1/collections/configured",
+        json={"detection": {"single_face_selection": "largest"}},
+    )
+
+    assert patched.status_code == 200, patched.text
+    assert patched.json()["collection"]["detection_revision"] == 2
+    assert patched.json()["collection"]["detection"]["threshold"] == 0.42
+    assert unchanged.status_code == 200, unchanged.text
+    assert unchanged.json()["collection"]["detection_revision"] == 2
+
+
+def test_center_largest_uses_one_center_weighted_area_score() -> None:
+    edge_large = _observation((0.0, 5.0, 100.0, 95.0), 0)
+    center_small = _observation((108.0, 30.0, 148.0, 70.0), 1)
+    near_center_large = _observation((70.0, 5.0, 180.0, 95.0), 2)
+    above_center = _observation((108.0, 0.0, 148.0, 40.0), 3)
+    below_center = _observation((108.0, 60.0, 148.0, 100.0), 4)
+
+    def score(face: FaceObservation) -> float:
+        return FaceService._selection_score(
+            face,
+            image_width=256,
+            image_height=100,
+            strategy="center_largest",
+        )
+
+    assert score(edge_large) == -3168.0
+    assert score(center_small) == 1600.0
+    assert score(near_center_large) == 9882.0
+    assert score(above_center) == score(below_center) == -200.0
+    assert max([edge_large, center_small], key=score) is center_small
+    # Unlike the previous lexicographic rule, exact center proximity does not
+    # automatically beat a much larger face. Both terms share one score.
+    assert max([center_small, near_center_large], key=score) is near_center_large
+
+
+def test_collection_profile_reaches_detect_compare_embeddings_search_and_enrollment(
+    client: TestClient, image_bytes, monkeypatch
+) -> None:
+    created = client.post(
+        "/v1/collections",
+        json={
+            "id": "centered",
+            "name": "Centered",
+            "detection": {
+                "threshold": 0.42,
+                "single_face_selection": "center_largest",
+            },
+        },
+    )
+    assert created.status_code == 201, created.text
+
+    edge_large = _observation((0.0, 5.0, 100.0, 95.0), 0)
+    center_small = _observation((108.0, 30.0, 148.0, 70.0), 1)
+    profiles: list[DetectionProfile] = []
+
+    def analyze(_image, *, detection_profile=None, **_kwargs):
+        assert detection_profile is not None
+        profiles.append(detection_profile)
+        return [edge_large, center_small]
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", analyze)
+    image = ("wide.png", image_bytes(12, width=256, height=100), "image/png")
+
+    detected = client.post(
+        "/v1/detect",
+        data={"collection_id": "centered"},
+        files={"image": image},
+    )
+    compared = client.post(
+        "/v1/compare",
+        data={"collection_id": "centered"},
+        files={"source": image, "target": image},
+    )
+    embedded = client.post(
+        "/v1/embeddings",
+        data={"collection_id": "centered"},
+        files={"image": image},
+    )
+    searched = client.post(
+        "/v1/collections/centered/search",
+        files={"image": image},
+    )
+    enrolled = client.post(
+        "/v1/collections/centered/persons",
+        data={"id": "center-person", "review_mode": "off"},
+        files={"images": image},
+    )
+
+    for response in (detected, compared, embedded, searched, enrolled):
+        assert response.status_code in {200, 201}, response.text
+    assert len(detected.json()["faces"]) == 2
+    assert compared.json()["source_face"]["bbox"]["pixels"]["x"] == 108
+    assert embedded.json()["faces"][0]["bbox"]["pixels"]["x"] == 108
+    assert searched.json()["searched_face"]["bbox"]["pixels"]["x"] == 108
+    assert enrolled.json()["faces"][0]["bounding_box"]["pixels"]["x"] == 108
+    assert profiles
+    assert all(profile.threshold == 0.42 for profile in profiles)
+    assert all(profile.single_face_selection == "center_largest" for profile in profiles)
diff --git a/server/tests/api/test_documentation_contract.py b/server/tests/api/test_documentation_contract.py
new file mode 100644
index 000000000..71b656c65
--- /dev/null
+++ b/server/tests/api/test_documentation_contract.py
@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+
+from fastapi.routing import APIRoute
+from fastapi.testclient import TestClient
+
+SERVER_DIR = Path(__file__).resolve().parents[2]
+DOC_SUFFIXES = ("", ".zh-CN", ".ja", ".de", ".es", ".fr", ".ru", ".pt", ".ko")
+DOCUMENTED_OPERATION = re.compile(
+    r"^### `(?P<method>GET|POST|PATCH|DELETE) (?P<path>/v1/[^`]+)`$",
+    re.MULTILINE,
+)
+
+
+def _public_operations(client: TestClient) -> set[str]:
+    operations: set[str] = set()
+    for route in client.app.routes:
+        if not isinstance(route, APIRoute) or not route.path.startswith("/v1/"):
+            continue
+        for method in route.methods & {"GET", "POST", "PATCH", "DELETE"}:
+            operations.add(f"{method} {route.path}")
+    return operations
+
+
+def _documented_sections(markdown: str) -> dict[str, str]:
+    matches = list(DOCUMENTED_OPERATION.finditer(markdown))
+    return {
+        f"{match.group('method')} {match.group('path')}": markdown[
+            match.end() : matches[index + 1].start() if index + 1 < len(matches) else len(markdown)
+        ]
+        for index, match in enumerate(matches)
+    }
+
+
+def test_every_locale_documents_exactly_every_public_operation(client: TestClient) -> None:
+    expected = _public_operations(client)
+    assert len(expected) == 29
+    for suffix in DOC_SUFFIXES:
+        path = SERVER_DIR / "docs" / f"api{suffix}.md"
+        sections = _documented_sections(path.read_text(encoding="utf-8"))
+        assert set(sections) == expected, path.name
+        for operation, content in sections.items():
+            # CJK conveys the same guidance with fewer code points than Latin
+            # scripts, so route coverage and labeled guidance are the primary
+            # contract; this floor only catches empty placeholder sections.
+            assert len(content.strip()) >= 70, f"{path.name}: {operation} is too brief"
+            assert content.count("**") >= 4, (
+                f"{path.name}: {operation} lacks usage/result guidance"
+            )
+
+
+def test_reviewed_openapi_snapshot_matches_runtime_contract(client: TestClient) -> None:
+    expected = json.loads(
+        (SERVER_DIR / "docs" / "openapi.snapshot.json").read_text(encoding="utf-8")
+    )
+    assert client.app.openapi() == expected
diff --git a/server/tests/api/test_external_trusted.py b/server/tests/api/test_external_trusted.py
new file mode 100644
index 000000000..f8c79897e
--- /dev/null
+++ b/server/tests/api/test_external_trusted.py
@@ -0,0 +1,349 @@
+from __future__ import annotations
+
+import json
+from typing import Any
+
+import numpy as np
+import pytest
+from fastapi.testclient import TestClient
+
+
+def _unit_vector(dimension: int, index: int = 0) -> list[float]:
+    values = [0.0] * dimension
+    values[index] = 1.0
+    return values
+
+
+def _trusted_data(
+    contract_id: str,
+    embeddings: list[Any],
+    **values: str,
+) -> dict[str, str]:
+    return {
+        "embedding_mode": "external_trusted",
+        "embedding_contract_id": contract_id,
+        "external_embeddings": json.dumps(embeddings),
+        **values,
+    }
+
+
+def test_external_trusted_skips_recognizer_persists_and_indexes_supplied_vector(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    dimension = int(collection["embedding_dimension"])
+    supplied = _unit_vector(dimension, 7)
+    calls: list[bool] = []
+    original_analyze = client.app.state.engine.analyze
+
+    def tracked_analyze(image, *, require_embeddings=True, **kwargs):
+        calls.append(bool(require_embeddings))
+        observations = original_analyze(image, require_embeddings=require_embeddings, **kwargs)
+        if not require_embeddings:
+            assert all(face.embedding is None for face in observations)
+        return observations
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", tracked_analyze)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [supplied], id="alice"),
+        files={"images": ("alice.png", image_bytes(501), "image/png")},
+    )
+
+    assert response.status_code == 201, response.text
+    face = response.json()["faces"][0]
+    assert face["embedding_source"] == "external_trusted"
+    assert face["embedding_contract_id"] == contract_id
+    assert calls == [False]
+    stored = client.app.state.repository.list_face_embeddings("employees", "alice")
+    assert len(stored) == 1
+    np.testing.assert_allclose(stored[0], np.asarray(supplied, dtype=np.float32))
+    matches = client.app.state.search_indexes.search(
+        "employees",
+        np.asarray(supplied, dtype=np.float32),
+        limit=1,
+        threshold=-1.0,
+    )
+    assert matches[0]["person"]["id"] == "alice"
+    assert matches[0]["matched_face_id"] == face["id"]
+
+
+def test_external_trusted_preserves_original_image_mapping_and_partial_success(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    dimension = int(collection["embedding_dimension"])
+    valid = _unit_vector(dimension)
+    calls: list[bool] = []
+    original_analyze = client.app.state.engine.analyze
+
+    def tracked_analyze(image, *, require_embeddings=True, **kwargs):
+        calls.append(bool(require_embeddings))
+        observations = original_analyze(image, require_embeddings=require_embeddings, **kwargs)
+        # Even if an engine violates require_embeddings=False, the service must
+        # never replace a rejected trusted vector with this recognizer output.
+        for face in observations:
+            face.embedding = np.asarray(_unit_vector(dimension), dtype=np.float32)
+        return observations
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", tracked_analyze)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(
+            contract_id,
+            [valid, valid, [0.0] * dimension],
+            id="alice",
+        ),
+        files=[
+            ("images", ("accepted.png", image_bytes(502), "image/png")),
+            ("images", ("invalid.png", b"not-an-image", "image/png")),
+            ("images", ("bad-vector.png", image_bytes(503), "image/png")),
+        ],
+    )
+
+    assert response.status_code == 201, response.text
+    body = response.json()
+    assert len(body["faces"]) == 1
+    assert {(item["index"], item["reason"]) for item in body["rejected_images"]} == {
+        (1, "invalid_image"),
+        (2, "invalid_external_embedding"),
+    }
+    # Invalid image bytes never reach inference, while both decoded images run
+    # detector/quality without recognition. The invalid vector never falls back.
+    assert calls == [False, False]
+
+
+def test_external_trusted_review_modes_match_image_registration_semantics(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    vector = _unit_vector(int(collection["embedding_dimension"]))
+    small = image_bytes(504, width=32, height=32)
+
+    off = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [vector], id="off", review_mode="off"),
+        files={"images": ("small-off.png", small, "image/png")},
+    )
+    standard = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [vector], id="standard", review_mode="standard"),
+        files={"images": ("small-standard.png", small, "image/png")},
+    )
+
+    assert off.status_code == 201, off.text
+    assert standard.status_code == 422
+    assert standard.json()["error"]["details"]["rejected_images"] == [
+        {"index": 0, "filename": "small-standard.png", "reason": "face_too_small"}
+    ]
+
+
+def test_external_trusted_off_binds_vector_to_largest_detected_face(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    vector = _unit_vector(int(collection["embedding_dimension"]))
+    multiple = image_bytes(506, width=256, height=100)
+    calls: list[bool] = []
+    original_analyze = client.app.state.engine.analyze
+
+    def tracked_analyze(image, *, require_embeddings=True, **kwargs):
+        calls.append(bool(require_embeddings))
+        return original_analyze(image, require_embeddings=require_embeddings, **kwargs)
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", tracked_analyze)
+
+    off = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [vector], id="off-largest", review_mode="off"),
+        files={"images": ("multiple-off.png", multiple, "image/png")},
+    )
+    standard = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(
+            contract_id,
+            [vector],
+            id="standard-multiple",
+            review_mode="standard",
+        ),
+        files={"images": ("multiple-standard.png", multiple, "image/png")},
+    )
+
+    assert off.status_code == 201, off.text
+    assert len(off.json()["faces"]) == 1
+    assert off.json()["faces"][0]["bounding_box"]["pixels"]["width"] == 62
+    assert off.json()["rejected_images"] == []
+    np.testing.assert_allclose(
+        client.app.state.repository.list_face_embeddings("employees", "off-largest")[0],
+        np.asarray(vector, dtype=np.float32),
+    )
+    assert standard.status_code == 422
+    assert standard.json()["error"]["details"]["rejected_images"][0]["reason"] == (
+        "multiple_faces"
+    )
+    assert calls == [False, False]
+
+
+def test_external_trusted_strict_review_uses_supplied_embedding(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    dimension = int(collection["embedding_dimension"])
+    first = _unit_vector(dimension, 0)
+    second = _unit_vector(dimension, 1)
+    created = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [first], id="alice"),
+        files={"images": ("first.png", image_bytes(505), "image/png")},
+    )
+    assert created.status_code == 201, created.text
+
+    observed_queries: list[np.ndarray] = []
+
+    def no_other(_collection_id, query, *, exclude_person_id):
+        assert exclude_person_id == "alice"
+        observed_queries.append(np.asarray(query, dtype=np.float32).copy())
+        return None
+
+    monkeypatch.setattr(
+        client.app.state.search_indexes,
+        "best_other_person",
+        no_other,
+        raising=False,
+    )
+    response = client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        data=_trusted_data(contract_id, [second], review_mode="strict"),
+        files={"images": ("second.png", image_bytes(506), "image/png")},
+    )
+
+    assert response.status_code == 201, response.text
+    assert len(response.json()["faces"]) == 1
+    assert len(observed_queries) == 1
+    np.testing.assert_array_equal(observed_queries[0], np.asarray(second, dtype=np.float32))
+    stored = client.app.state.repository.list_face_embeddings("employees", "alice")
+    assert any(np.array_equal(value, np.asarray(second, dtype=np.float32)) for value in stored)
+
+
+@pytest.mark.parametrize(
+    ("data", "expected_code"),
+    [
+        (
+            {"external_embeddings": "[]"},
+            "unexpected_external_embedding",
+        ),
+        (
+            {"embedding_mode": "external_trusted", "embedding_contract_id": "value"},
+            "missing_external_embeddings",
+        ),
+        (
+            {
+                "embedding_mode": "external_trusted",
+                "embedding_contract_id": "value",
+                "external_embeddings": "not-json",
+            },
+            "invalid_external_embeddings",
+        ),
+        (
+            {
+                "embedding_mode": "external_trusted",
+                "embedding_contract_id": "value",
+                "external_embeddings": "[]",
+            },
+            "external_embedding_count_mismatch",
+        ),
+    ],
+)
+def test_external_trusted_rejects_ambiguous_or_malformed_multipart_contract(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    data: dict[str, str],
+    expected_code: str,
+) -> None:
+    create_collection(client)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "alice", **data},
+        files={"images": ("alice.png", image_bytes(507), "image/png")},
+    )
+    assert response.status_code == 400
+    assert response.json()["error"]["code"] == expected_code
+
+
+def test_external_trusted_rejects_contract_mismatch_before_inference(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    vector = _unit_vector(int(collection["embedding_dimension"]))
+    calls = 0
+
+    def unexpected_analyze(*_args, **_kwargs):
+        nonlocal calls
+        calls += 1
+        raise AssertionError("contract mismatch must not run inference")
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", unexpected_analyze)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data("ifsemb-v1-sha256:" + "0" * 64, [vector], id="alice"),
+        files={"images": ("alice.png", image_bytes(508), "image/png")},
+    )
+    assert response.status_code == 409
+    assert response.json()["error"]["code"] == "embedding_contract_mismatch"
+    assert calls == 0
+
+
+def test_external_trusted_invalid_vector_is_per_image_rejection_without_fallback(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    collection = create_collection(client)
+    contract_id = str(collection["embedding_contract_id"])
+    dimension = int(collection["embedding_dimension"])
+    calls: list[bool] = []
+    original_analyze = client.app.state.engine.analyze
+
+    def tracked_analyze(image, *, require_embeddings=True, **kwargs):
+        calls.append(bool(require_embeddings))
+        return original_analyze(image, require_embeddings=require_embeddings, **kwargs)
+
+    monkeypatch.setattr(client.app.state.engine, "analyze", tracked_analyze)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data=_trusted_data(contract_id, [[0.0] * dimension], id="alice"),
+        files={"images": ("alice.png", image_bytes(509), "image/png")},
+    )
+    assert response.status_code == 422
+    assert response.json()["error"]["details"]["rejected_images"] == [
+        {
+            "index": 0,
+            "filename": "alice.png",
+            "reason": "invalid_external_embedding",
+        }
+    ]
+    assert calls == [False]
diff --git a/server/tests/api/test_face_crop_database.py b/server/tests/api/test_face_crop_database.py
new file mode 100644
index 000000000..7e1041618
--- /dev/null
+++ b/server/tests/api/test_face_crop_database.py
@@ -0,0 +1,214 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+
+from fastapi.testclient import TestClient
+from insightface_server.app import create_app
+from insightface_server.config import Settings
+
+
+def _create_collection(
+    client: TestClient, collection_id: str, *, save_face_crops: bool | None = None
+) -> dict[str, object]:
+    payload: dict[str, object] = {"id": collection_id, "name": collection_id}
+    if save_face_crops is not None:
+        payload["save_face_crops"] = save_face_crops
+    response = client.post("/v1/collections", json=payload)
+    assert response.status_code == 201, response.text
+    return response.json()["collection"]
+
+
+def _create_person(
+    client: TestClient,
+    image_bytes: Callable[..., bytes],
+    collection_id: str,
+    person_id: str,
+    *,
+    seed: int,
+    headers: dict[str, str] | None = None,
+) -> dict[str, object]:
+    response = client.post(
+        f"/v1/collections/{collection_id}/persons",
+        data={"id": person_id},
+        files={"images": (f"{person_id}.png", image_bytes(seed), "image/png")},
+        headers=headers,
+    )
+    assert response.status_code == 201, response.text
+    return response.json()
+
+
+def test_collection_crop_setting_defaults_and_can_be_changed(
+    client: TestClient,
+) -> None:
+    inherited = _create_collection(client, "inherited")
+    explicit = _create_collection(client, "explicit", save_face_crops=True)
+
+    enabled = client.patch(
+        "/v1/collections/inherited", json={"save_face_crops": True}
+    )
+    disabled = client.patch(
+        "/v1/collections/explicit", json={"save_face_crops": False}
+    )
+    invalid = client.patch(
+        "/v1/collections/explicit", json={"save_face_crops": None}
+    )
+
+    assert inherited["save_face_crops"] is False
+    assert explicit["save_face_crops"] is True
+    assert enabled.status_code == 200
+    assert enabled.json()["collection"]["save_face_crops"] is True
+    assert disabled.status_code == 200
+    assert disabled.json()["collection"]["save_face_crops"] is False
+    assert invalid.status_code == 400
+    assert invalid.json()["error"]["code"] == "invalid_request"
+
+
+def test_collection_crop_default_can_be_enabled_by_deployment_setting(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(save_face_crops=True)
+    with TestClient(create_app(settings)) as client:
+        collection = _create_collection(client, "inherited")
+
+    assert collection["save_face_crops"] is True
+
+
+def test_saved_crop_is_private_jpeg_and_disabling_only_affects_new_faces(
+    client: TestClient, image_bytes: Callable[..., bytes]
+) -> None:
+    _create_collection(client, "employees", save_face_crops=True)
+    created = _create_person(
+        client, image_bytes, "employees", "alice", seed=101
+    )
+    first = created["faces"][0]
+    first_id = first["id"]
+    first_url = f"/v1/collections/employees/persons/alice/faces/{first_id}/image"
+
+    downloaded = client.get(first_url)
+    listed = client.get("/v1/collections/employees/persons/alice/faces")
+    disabled = client.patch(
+        "/v1/collections/employees", json={"save_face_crops": False}
+    )
+    added = client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        files={"images": ("alice-2.png", image_bytes(102), "image/png")},
+    )
+    second = added.json()["faces"][0]
+    missing = client.get(
+        f"/v1/collections/employees/persons/alice/faces/{second['id']}/image"
+    )
+    old_after_disable = client.get(first_url)
+
+    assert first["has_crop"] is True
+    assert downloaded.status_code == 200
+    assert downloaded.headers["content-type"] == "image/jpeg"
+    assert downloaded.headers["cache-control"] == "no-store"
+    assert downloaded.headers["x-request-id"]
+    assert downloaded.content.startswith(b"\xff\xd8")
+    assert downloaded.content.endswith(b"\xff\xd9")
+    assert listed.status_code == 200
+    assert listed.json()["faces"][0]["has_crop"] is True
+    assert all(
+        "crop_image" not in face and "crop_path" not in face
+        for face in listed.json()["faces"]
+    )
+    assert disabled.status_code == 200
+    assert second["has_crop"] is False
+    assert missing.status_code == 404
+    assert missing.json()["error"]["code"] == "face_image_not_found"
+    assert old_after_disable.status_code == 200
+    assert old_after_disable.content == downloaded.content
+
+
+def test_disabled_collection_does_not_store_a_crop(
+    client: TestClient, image_bytes: Callable[..., bytes]
+) -> None:
+    _create_collection(client, "employees")
+    created = _create_person(client, image_bytes, "employees", "alice", seed=103)
+    face = created["faces"][0]
+
+    response = client.get(
+        f"/v1/collections/employees/persons/alice/faces/{face['id']}/image"
+    )
+    unknown = client.get(
+        "/v1/collections/employees/persons/alice/faces/unknown-face/image"
+    )
+
+    assert face["has_crop"] is False
+    assert response.status_code == 404
+    assert response.json()["error"]["code"] == "face_image_not_found"
+    assert unknown.status_code == 404
+    assert unknown.json()["error"]["code"] == "face_not_found"
+
+
+def test_face_image_requires_bearer_authentication(
+    make_settings: Callable[..., Settings], image_bytes: Callable[..., bytes]
+) -> None:
+    settings = make_settings(auth_enabled=True, startup_api_key="test-secret")
+    headers = {"Authorization": "Bearer test-secret"}
+    with TestClient(create_app(settings)) as client:
+        collection = client.post(
+            "/v1/collections",
+            json={"id": "employees", "name": "Employees", "save_face_crops": True},
+            headers=headers,
+        )
+        assert collection.status_code == 201, collection.text
+        created = _create_person(
+            client,
+            image_bytes,
+            "employees",
+            "alice",
+            seed=104,
+            headers=headers,
+        )
+        face_id = created["faces"][0]["id"]
+        url = f"/v1/collections/employees/persons/alice/faces/{face_id}/image"
+
+        unauthenticated = client.get(url)
+        authenticated = client.get(url, headers=headers)
+
+    assert unauthenticated.status_code == 401
+    assert authenticated.status_code == 200
+    assert authenticated.headers["cache-control"] == "no-store"
+
+
+def test_strict_rejection_does_not_persist_encoded_crop(
+    client: TestClient, image_bytes: Callable[..., bytes]
+) -> None:
+    _create_collection(client, "employees", save_face_crops=True)
+    _create_person(client, image_bytes, "employees", "alice", seed=105)
+    _create_person(client, image_bytes, "employees", "bob", seed=106)
+
+    rejected = client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        data={"review_mode": "strict"},
+        files={"images": ("bob.png", image_bytes(106), "image/png")},
+    )
+
+    assert rejected.status_code == 201, rejected.text
+    assert rejected.json()["faces"] == []
+    assert rejected.json()["rejected_images"][0]["reason"] == (
+        "identity_similarity_conflict"
+    )
+    with client.app.state.database.read() as connection:
+        persisted = connection.execute(
+            "SELECT count(*) FROM face_samples WHERE crop_image IS NOT NULL"
+        ).fetchone()[0]
+    assert persisted == 2
+
+
+def test_openapi_documents_collection_crop_flag_and_jpeg_download(
+    client: TestClient,
+) -> None:
+    schema = client.get("/openapi.json").json()
+    collection_create = schema["components"]["schemas"]["CollectionCreate"]
+    collection_patch = schema["components"]["schemas"]["CollectionPatch"]
+    operation = schema["paths"][
+        "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image"
+    ]["get"]
+
+    assert "save_face_crops" in collection_create["properties"]
+    assert "save_face_crops" in collection_patch["properties"]
+    assert operation["security"] == [{"bearerAuth": []}]
+    jpeg = operation["responses"]["200"]["content"]["image/jpeg"]["schema"]
+    assert jpeg == {"type": "string", "format": "binary"}
diff --git a/server/tests/api/test_face_operations.py b/server/tests/api/test_face_operations.py
new file mode 100644
index 000000000..a3a8d5c27
--- /dev/null
+++ b/server/tests/api/test_face_operations.py
@@ -0,0 +1,157 @@
+from __future__ import annotations
+
+import math
+import time
+from collections.abc import Callable
+from io import BytesIO
+
+from fastapi.testclient import TestClient
+from insightface_server.app import create_app
+from insightface_server.config import Settings
+from PIL import Image
+
+
+def test_detect_sorts_faces_and_empty_image_is_success(client: TestClient, image_bytes) -> None:
+    two_faces = client.post(
+        "/v1/detect",
+        data={"max_faces": "2"},
+        files={"image": ("wide.png", image_bytes(4, width=256, height=100), "image/png")},
+    )
+    blank = client.post(
+        "/v1/detect",
+        files={"image": ("blank.png", image_bytes(blank=True), "image/png")},
+    )
+
+    assert two_faces.status_code == 200
+    faces = two_faces.json()["faces"]
+    assert len(faces) == 2
+    areas = [face["bbox"]["pixels"]["width"] * face["bbox"]["pixels"]["height"] for face in faces]
+    assert areas == sorted(areas, reverse=True)
+    assert all(0 <= face["detection_score"] <= 1 for face in faces)
+    assert all(len(face["landmarks"]) == 5 for face in faces)
+    assert blank.status_code == 200
+    assert blank.json()["faces"] == []
+
+
+def test_detect_respects_max_faces_and_rejects_request_threshold_override(
+    client: TestClient, image_bytes
+) -> None:
+    maximum = client.post(
+        "/v1/detect",
+        data={"max_faces": "1"},
+        files={"image": ("wide.png", image_bytes(9, width=256, height=100), "image/png")},
+    )
+    strict = client.post(
+        "/v1/detect",
+        data={"min_score": "1.0"},
+        files={"image": ("face.png", image_bytes(9), "image/png")},
+    )
+
+    assert maximum.status_code == 200
+    assert len(maximum.json()["faces"]) == 1
+    assert strict.status_code == 400
+    assert strict.json()["error"]["code"] == "request_detection_override_not_supported"
+
+
+def test_detect_accepts_webp(client: TestClient, image_bytes) -> None:
+    encoded = BytesIO()
+    with Image.open(BytesIO(image_bytes(17))) as image:
+        image.save(encoded, format="WEBP", quality=90)
+
+    response = client.post(
+        "/v1/detect",
+        files={"image": ("face.webp", encoded.getvalue(), "image/webp")},
+    )
+
+    assert response.status_code == 200
+    assert len(response.json()["faces"]) == 1
+
+
+def test_compare_same_synthetic_image_and_face_not_found(client: TestClient, image_bytes) -> None:
+    sample = image_bytes(21)
+    compared = client.post(
+        "/v1/compare",
+        data={"threshold": "0.99"},
+        files={
+            "source": ("source.png", sample, "image/png"),
+            "target": ("target.png", sample, "image/png"),
+        },
+    )
+    missing = client.post(
+        "/v1/compare",
+        files={
+            "source": ("source.png", sample, "image/png"),
+            "target": ("blank.png", image_bytes(blank=True), "image/png"),
+        },
+    )
+
+    assert compared.status_code == 200
+    assert compared.json()["matched"] is True
+    assert math.isclose(compared.json()["similarity"], 1.0)
+    assert compared.json()["threshold"] == 0.99
+    assert compared.json()["source_face"]["bbox"] == compared.json()["target_face"]["bbox"]
+    assert missing.status_code == 422
+    assert missing.json()["error"]["code"] == "face_not_found"
+
+
+def test_compare_returns_raw_negative_cosine_without_affine_mapping(
+    client: TestClient, image_bytes
+) -> None:
+    compared = client.post(
+        "/v1/compare",
+        data={"threshold": "0"},
+        files={
+            "source": ("source.png", image_bytes(3), "image/png"),
+            "target": ("target.png", image_bytes(20), "image/png"),
+        },
+    )
+
+    assert compared.status_code == 200
+    assert compared.json()["similarity"] < 0.0
+    assert compared.json()["matched"] is False
+    assert compared.json()["threshold"] == 0.0
+
+
+def test_embeddings_uses_profile_selection_and_rejects_request_override(
+    client: TestClient, image_bytes
+) -> None:
+    selected = client.post(
+        "/v1/embeddings",
+        files={"image": ("wide.png", image_bytes(8, width=256, height=100), "image/png")},
+    )
+    invalid = client.post(
+        "/v1/embeddings",
+        data={"face_selection": "nearest"},
+        files={"image": ("face.png", image_bytes(8), "image/png")},
+    )
+
+    assert selected.status_code == 200
+    assert len(selected.json()["faces"]) == 1
+    for face in selected.json()["faces"]:
+        embedding = face["embedding"]
+        assert len(embedding) == 512
+        assert math.isclose(sum(value * value for value in embedding), 1.0, abs_tol=1e-5)
+    assert invalid.status_code == 400
+    assert invalid.json()["error"]["code"] == "request_detection_override_not_supported"
+
+
+def test_inference_timeout_returns_without_waiting_for_worker(
+    make_settings: Callable[..., Settings], image_bytes
+) -> None:
+    settings = make_settings(request_timeout_seconds=1)
+    with TestClient(create_app(settings)) as client:
+        def slow_detect(*_args, **_kwargs):
+            time.sleep(2.0)
+            return []
+
+        client.app.state.service.detect = slow_detect
+        started = time.monotonic()
+        response = client.post(
+            "/v1/detect",
+            files={"image": ("face.png", image_bytes(7), "image/png")},
+        )
+        elapsed = time.monotonic() - started
+
+    assert response.status_code == 503
+    assert response.json()["error"]["code"] == "request_timeout"
+    assert elapsed < 1.6
diff --git a/server/tests/api/test_resources.py b/server/tests/api/test_resources.py
new file mode 100644
index 000000000..2ce6f31db
--- /dev/null
+++ b/server/tests/api/test_resources.py
@@ -0,0 +1,761 @@
+from __future__ import annotations
+
+import time
+from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor
+from threading import Event, Lock
+
+import pytest
+from fastapi.testclient import TestClient
+from insightface_server.app import create_app
+from insightface_server.config import Settings
+from insightface_server.search import SearchMutationCommittedError
+
+
+def test_detect_embeddings_and_search_share_global_inference_budget(
+    make_settings: Callable[..., Settings],
+    create_collection,
+    create_person,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    settings = make_settings(inference_max_concurrency=2)
+    with TestClient(create_app(settings)) as client:
+        create_collection(client)
+        create_person(client, "employees", "alice", seed=77)
+        engine = client.app.state.engine
+        original_observation = engine._observation
+        release = Event()
+        two_active = Event()
+        state_lock = Lock()
+        active = 0
+        observed_peak = 0
+
+        def blocked_observation(*args, **kwargs):
+            nonlocal active, observed_peak
+            with state_lock:
+                active += 1
+                observed_peak = max(observed_peak, active)
+                if active == 2:
+                    two_active.set()
+            try:
+                assert release.wait(timeout=5)
+                return original_observation(*args, **kwargs)
+            finally:
+                with state_lock:
+                    active -= 1
+
+        monkeypatch.setattr(engine, "_observation", blocked_observation)
+
+        def detect():
+            return client.post(
+                "/v1/detect",
+                files={"image": ("detect.png", image_bytes(71), "image/png")},
+            )
+
+        def embeddings():
+            return client.post(
+                "/v1/embeddings",
+                files={"image": ("embedding.png", image_bytes(72), "image/png")},
+            )
+
+        def search():
+            return client.post(
+                "/v1/collections/employees/search",
+                files={"image": ("search.png", image_bytes(77), "image/png")},
+            )
+
+        with ThreadPoolExecutor(max_workers=3) as pool:
+            futures = [pool.submit(operation) for operation in (detect, embeddings, search)]
+            assert two_active.wait(timeout=5)
+            deadline = time.monotonic() + 5
+            while (
+                engine.runtime_summary()["inference_concurrency"]["waiting"] < 1
+                and time.monotonic() < deadline
+            ):
+                time.sleep(0.01)
+            concurrency = engine.runtime_summary()["inference_concurrency"]
+            assert concurrency["active"] == 2
+            assert concurrency["waiting"] == 1
+            release.set()
+            responses = [future.result(timeout=5) for future in futures]
+
+        assert [response.status_code for response in responses] == [200, 200, 200]
+        assert observed_peak == 2
+        assert engine.runtime_summary()["inference_concurrency"]["peak_active"] == 2
+
+
+def test_collection_crud_conflict_and_force_delete(
+    client: TestClient, create_collection, create_person
+) -> None:
+    collection = create_collection(client, threshold=0.7)
+    assert collection["id"] == "employees"
+    assert collection["default_threshold"] == 0.7
+    assert collection["person_count"] == 0
+
+    duplicate = client.post("/v1/collections", json={"id": "employees", "name": "Duplicate"})
+    patched = client.patch(
+        "/v1/collections/employees",
+        json={"name": "Team", "threshold": 0.72, "metadata": {"site": "north"}},
+    )
+    create_person(client, "employees", "person-1")
+    protected = client.delete("/v1/collections/employees")
+    deleted = client.delete("/v1/collections/employees?force=true")
+
+    assert duplicate.status_code == 409
+    assert duplicate.json()["error"]["code"] == "collection_exists"
+    assert patched.status_code == 200
+    assert patched.json()["collection"]["name"] == "Team"
+    assert patched.json()["collection"]["default_threshold"] == 0.72
+    assert protected.status_code == 409
+    assert protected.json()["error"]["code"] == "collection_not_empty"
+    assert deleted.status_code == 204
+    assert deleted.content == b""
+    assert deleted.headers["x-request-id"]
+    assert client.get("/v1/collections/employees").status_code == 404
+
+
+def test_collection_from_an_old_model_can_still_be_deleted(
+    client: TestClient, create_collection
+) -> None:
+    create_collection(client)
+    with client.app.state.database.write() as connection:
+        connection.execute(
+            "UPDATE collections SET model_digest=? WHERE id=?",
+            ("0" * 64, "employees"),
+        )
+
+    incompatible = client.get("/v1/collections/employees")
+    deleted = client.delete("/v1/collections/employees")
+
+    assert incompatible.status_code == 409
+    assert incompatible.json()["error"]["code"] == "collection_model_mismatch"
+    assert deleted.status_code == 204
+    assert client.app.state.repository.get_collection("employees") is None
+
+
+def test_force_false_conflict_keeps_collection_index_searchable(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client)
+    create_person(client, "employees", "alice", seed=41)
+    query = {
+        "image": ("query.png", image_bytes(41), "image/png"),
+    }
+    before = client.post(
+        "/v1/collections/employees/search",
+        data={"threshold": "0.99"},
+        files=query,
+    )
+
+    protected = client.delete("/v1/collections/employees")
+    after = client.post(
+        "/v1/collections/employees/search",
+        data={"threshold": "0.99"},
+        files=query,
+    )
+
+    assert before.status_code == 200
+    assert protected.status_code == 409
+    assert protected.json()["error"]["code"] == "collection_not_empty"
+    assert after.status_code == 200
+    assert [match["person"]["id"] for match in after.json()["matches"]] == ["alice"]
+
+
+def test_force_delete_racing_search_returns_not_found_instead_of_500(
+    client: TestClient,
+    create_collection,
+    create_person,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    create_collection(client)
+    create_person(client, "employees", "alice", seed=42)
+    delete_entered = Event()
+    search_entered = Event()
+    allow_delete = Event()
+    original_delete = client.app.state.repository.delete_collection_with_mutation
+    original_search = client.app.state.search_indexes.search
+
+    def blocked_delete(collection_id: str, *, force: bool):
+        delete_entered.set()
+        assert allow_delete.wait(timeout=5)
+        return original_delete(collection_id, force=force)
+
+    def signaled_search(*args, **kwargs):
+        search_entered.set()
+        return original_search(*args, **kwargs)
+
+    monkeypatch.setattr(
+        client.app.state.repository,
+        "delete_collection_with_mutation",
+        blocked_delete,
+    )
+    monkeypatch.setattr(client.app.state.search_indexes, "search", signaled_search)
+
+    with ThreadPoolExecutor(max_workers=2) as pool:
+        deleting = pool.submit(
+            client.delete, "/v1/collections/employees?force=true"
+        )
+        assert delete_entered.wait(timeout=5)
+        searching = pool.submit(
+            client.post,
+            "/v1/collections/employees/search",
+            data={"threshold": "0.0"},
+            files={"image": ("query.png", image_bytes(42), "image/png")},
+        )
+        assert search_entered.wait(timeout=5)
+        allow_delete.set()
+        deleted = deleting.result(timeout=5)
+        searched = searching.result(timeout=5)
+
+    assert deleted.status_code == 204
+    assert searched.status_code == 404
+    assert searched.json()["error"]["code"] == "collection_not_found"
+
+
+@pytest.mark.parametrize(
+    "profile",
+    ("fp32_v1", "fp16_v1", "bf16_v1", "int8_x736_v1", "int8_x1000_v1"),
+)
+def test_collection_search_profiles_and_capacity_are_persisted(
+    client: TestClient, profile: str
+) -> None:
+    collection_id = profile.replace("_v1", "")
+    response = client.post(
+        "/v1/collections",
+        json={
+            "id": collection_id,
+            "name": profile,
+            "threshold": 0.25,
+            "search": {
+                "profile": profile,
+                "capacity_rows": 1234,
+                "max_faces_per_person": 12,
+                "load_policy": "eager",
+            },
+        },
+    )
+
+    assert response.status_code == 201, response.text
+    collection = response.json()["collection"]
+    assert collection["default_threshold"] == 0.25
+    assert collection["search_profile"] == profile
+    assert collection["capacity_rows"] == 1234
+    assert collection["max_faces_per_person"] == 12
+    assert collection["load_policy"] == "eager"
+
+
+def test_collection_search_defaults_default_id_and_validation(client: TestClient) -> None:
+    regular = client.post(
+        "/v1/collections", json={"id": "regular", "name": "Regular"}
+    )
+    default = client.post(
+        "/v1/collections", json={"id": "_default", "name": "Default"}
+    )
+    unsupported = client.post(
+        "/v1/collections",
+        json={
+            "id": "rerank",
+            "name": "Rerank",
+            "search": {"profile": "int8_x1000_rerank_v1"},
+        },
+    )
+    out_of_range = client.post(
+        "/v1/collections",
+        json={"id": "bad-score", "name": "Bad", "threshold": -0.01},
+    )
+
+    assert regular.status_code == 201
+    regular_collection = regular.json()["collection"]
+    assert regular_collection["search_profile"] == "fp32_v1"
+    assert regular_collection["capacity_rows"] == 100_000
+    assert regular_collection["max_faces_per_person"] == 20
+    assert regular_collection["load_policy"] == "lazy"
+    assert regular_collection["default_threshold"] == 0.4
+    assert default.status_code == 201
+    assert default.json()["collection"]["load_policy"] == "eager"
+    assert unsupported.status_code == 400
+    assert out_of_range.status_code == 400
+
+
+def test_collection_search_operational_fields_can_be_patched(client: TestClient) -> None:
+    created = client.post(
+        "/v1/collections", json={"id": "patchable", "name": "Patchable"}
+    )
+    assert created.status_code == 201
+
+    patched = client.patch(
+        "/v1/collections/patchable",
+        json={
+            "search": {
+                "capacity_rows": 2_000,
+                "max_faces_per_person": 30,
+                "load_policy": "eager",
+            }
+        },
+    )
+    forbidden_profile = client.patch(
+        "/v1/collections/patchable",
+        json={"search": {"profile": "bf16_v1"}},
+    )
+
+    assert patched.status_code == 200, patched.text
+    collection = patched.json()["collection"]
+    assert collection["capacity_rows"] == 2_000
+    assert collection["max_faces_per_person"] == 30
+    assert collection["load_policy"] == "eager"
+    assert forbidden_profile.status_code == 400
+
+
+def test_deployment_max_search_capacity_guards_create_patch_and_is_reported(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(
+        default_search_capacity_rows=8,
+        max_search_capacity_rows=10,
+    )
+    with TestClient(create_app(settings)) as client:
+        rejected_create = client.post(
+            "/v1/collections",
+            json={
+                "id": "too-large",
+                "name": "Too large",
+                "search": {"capacity_rows": 11},
+            },
+        )
+        created = client.post(
+            "/v1/collections",
+            json={
+                "id": "bounded",
+                "name": "Bounded",
+                "search": {"capacity_rows": 10},
+            },
+        )
+        rejected_patch = client.patch(
+            "/v1/collections/bounded",
+            json={"search": {"capacity_rows": 11}},
+        )
+        persisted = client.get("/v1/collections/bounded")
+        system = client.get("/v1/system")
+
+    assert rejected_create.status_code == 400
+    assert rejected_create.json()["error"] == {
+        "code": "search_capacity_too_large",
+        "message": "capacity_rows exceeds this deployment's configured maximum.",
+        "details": {"max_capacity_rows": 10},
+    }
+    assert created.status_code == 201
+    assert rejected_patch.status_code == 400
+    assert rejected_patch.json()["error"]["code"] == "search_capacity_too_large"
+    assert persisted.json()["collection"]["capacity_rows"] == 10
+    assert system.json()["safe_config"]["max_search_capacity_rows"] == 10
+
+
+def test_collection_capacity_is_a_hard_409_and_does_not_commit_extra_face(
+    client: TestClient, create_person, image_bytes
+) -> None:
+    created = client.post(
+        "/v1/collections",
+        json={
+            "id": "bounded",
+            "name": "Bounded",
+            "search": {"capacity_rows": 1},
+        },
+    )
+    assert created.status_code == 201
+    create_person(client, "bounded", "alice", seed=31)
+
+    overflow = client.post(
+        "/v1/collections/bounded/persons/alice/faces",
+        files={"images": ("second.png", image_bytes(32), "image/png")},
+    )
+    person = client.get("/v1/collections/bounded/persons/alice")
+    faces = client.get("/v1/collections/bounded/persons/alice/faces")
+
+    assert overflow.status_code == 409
+    assert overflow.json()["error"]["code"] == "collection_capacity_exceeded"
+    assert person.json()["person"]["face_count"] == 1
+    assert len(faces.json()["faces"]) == 1
+
+
+def test_person_registration_supports_partial_success(
+    client: TestClient, create_collection, image_bytes
+) -> None:
+    create_collection(client)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data={
+            "id": "employee-001",
+            "name": "Alice",
+            "external_id": "HR-1001",
+            "metadata": '{"department":"sales"}',
+        },
+        files=[
+            ("images", ("accepted.png", image_bytes(1), "image/png")),
+            ("images", ("blank.png", image_bytes(blank=True), "image/png")),
+            (
+                "images",
+                ("multiple.png", image_bytes(2, width=256, height=100), "image/png"),
+            ),
+            ("images", ("invalid.png", b"not-an-image", "image/png")),
+        ],
+    )
+
+    assert response.status_code == 201
+    body = response.json()
+    assert body["person"]["id"] == "employee-001"
+    assert body["person"]["metadata"] == {"department": "sales"}
+    assert body["person"]["face_count"] == 2
+    assert len(body["faces"]) == 2
+    assert [(item["index"], item["reason"]) for item in body["rejected_images"]] == [
+        (3, "invalid_image"),
+        (1, "face_not_found"),
+    ]
+
+
+def test_registration_off_selects_largest_face_but_standard_rejects_multiple(
+    client: TestClient, create_collection, image_bytes
+) -> None:
+    create_collection(client)
+    multiple = image_bytes(202, width=256, height=100)
+
+    off = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "off-largest", "review_mode": "off"},
+        files={"images": ("multiple-off.png", multiple, "image/png")},
+    )
+    standard = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "standard-multiple", "review_mode": "standard"},
+        files={"images": ("multiple-standard.png", multiple, "image/png")},
+    )
+
+    assert off.status_code == 201, off.text
+    assert len(off.json()["faces"]) == 1
+    # The mock detector returns 62px and 46px faces, in descending area order.
+    assert off.json()["faces"][0]["bounding_box"]["pixels"]["width"] == 62
+    assert off.json()["rejected_images"] == []
+    assert standard.status_code == 422
+    assert standard.json()["error"]["details"]["rejected_images"] == [
+        {
+            "index": 0,
+            "filename": "multiple-standard.png",
+            "reason": "multiple_faces",
+        }
+    ]
+
+
+def test_registration_review_mode_defaults_off_and_standard_enables_quality(
+    client: TestClient, create_collection, image_bytes
+) -> None:
+    create_collection(client)
+    small = image_bytes(101, width=32, height=32)
+
+    default_off = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "default-off"},
+        files={"images": ("small.png", small, "image/png")},
+    )
+    standard = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "standard", "review_mode": "standard"},
+        files={"images": ("small.png", small, "image/png")},
+    )
+    invalid = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "invalid", "review_mode": "aggressive"},
+        files={"images": ("small.png", small, "image/png")},
+    )
+
+    assert default_off.status_code == 201, default_off.text
+    assert standard.status_code == 422
+    assert standard.json()["error"]["details"]["rejected_images"] == [
+        {"index": 0, "filename": "small.png", "reason": "face_too_small"}
+    ]
+    assert invalid.status_code == 400
+
+
+def test_strict_review_bootstraps_then_rejects_profile_similarity_tie(
+    client: TestClient,
+    create_collection,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    create_collection(client)
+    calls = 0
+
+    def tied_other(collection_id, _query, *, exclude_person_id):
+        nonlocal calls
+        calls += 1
+        assert collection_id == "employees"
+        assert exclude_person_id == "alice"
+        return {
+            "person_id": "bob",
+            "face_id": "bob-face",
+            "similarity": 1.0,
+        }
+
+    monkeypatch.setattr(
+        client.app.state.search_indexes,
+        "best_other_person",
+        tied_other,
+        raising=False,
+    )
+    sample = image_bytes(102)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "alice", "review_mode": "strict"},
+        files=[
+            ("images", ("bootstrap.png", sample, "image/png")),
+            ("images", ("tie.png", sample, "image/png")),
+        ],
+    )
+
+    assert response.status_code == 201, response.text
+    body = response.json()
+    assert len(body["faces"]) == 1
+    assert calls == 1
+    assert body["rejected_images"] == [
+        {
+            "index": 1,
+            "filename": "tie.png",
+            "reason": "identity_similarity_conflict",
+            "same_person_similarity": pytest.approx(1.0),
+            "other_person_similarity": 1.0,
+            "other_person_id": "bob",
+            "matched_face_id": "bob-face",
+        }
+    ]
+
+
+def test_strict_review_existing_person_accepts_only_when_same_is_greater(
+    client: TestClient,
+    create_collection,
+    create_person,
+    image_bytes,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    create_collection(client)
+    create_person(client, "employees", "alice", seed=103)
+
+    monkeypatch.setattr(
+        client.app.state.search_indexes,
+        "best_other_person",
+        lambda *_args, **_kwargs: {
+            "person_id": "bob",
+            "face_id": "bob-face",
+            "similarity": 0.5,
+        },
+        raising=False,
+    )
+    accepted = client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        data={"review_mode": "strict"},
+        files={"images": ("same.png", image_bytes(103), "image/png")},
+    )
+
+    monkeypatch.setattr(
+        client.app.state.search_indexes,
+        "best_other_person",
+        lambda *_args, **_kwargs: {
+            "person_id": "bob",
+            "face_id": "bob-face",
+            "similarity": 1.0,
+        },
+        raising=False,
+    )
+    rejected = client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        data={"review_mode": "strict"},
+        files={"images": ("tie.png", image_bytes(103), "image/png")},
+    )
+
+    assert accepted.status_code == 201, accepted.text
+    assert len(accepted.json()["faces"]) == 1
+    assert rejected.status_code == 201, rejected.text
+    assert rejected.json()["faces"] == []
+    assert rejected.json()["rejected_images"][0]["reason"] == (
+        "identity_similarity_conflict"
+    )
+    assert rejected.json()["rejected_images"][0]["same_person_similarity"] == (
+        pytest.approx(1.0)
+    )
+
+
+def test_registration_all_fail_is_422_with_reasons(
+    client: TestClient, create_collection, image_bytes
+) -> None:
+    create_collection(client)
+    response = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "no-face", "review_mode": "standard"},
+        files=[
+            ("images", ("blank.png", image_bytes(blank=True), "image/png")),
+            (
+                "images",
+                ("multiple.png", image_bytes(3, width=256, height=100), "image/png"),
+            ),
+        ],
+    )
+
+    assert response.status_code == 422
+    assert response.json()["error"]["code"] == "registration_failed"
+    assert {item["reason"] for item in response.json()["error"]["details"]["rejected_images"]} == {
+        "face_not_found",
+        "multiple_faces",
+    }
+
+
+def test_person_and_face_crud(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client)
+    created = create_person(client, "employees", "employee-001", seed=5)
+    first_face_id = created["faces"][0]["id"]
+
+    patched = client.patch(
+        "/v1/collections/employees/persons/employee-001",
+        json={"name": "Alice Updated", "metadata": {"level": 2}},
+    )
+    added = client.post(
+        "/v1/collections/employees/persons/employee-001/faces",
+        files={"images": ("second.png", image_bytes(6), "image/png")},
+    )
+    listed = client.get("/v1/collections/employees/persons/employee-001/faces")
+    deleted_face = client.delete(
+        f"/v1/collections/employees/persons/employee-001/faces/{first_face_id}"
+    )
+    person = client.get("/v1/collections/employees/persons/employee-001")
+    deleted_person = client.delete("/v1/collections/employees/persons/employee-001")
+
+    assert patched.status_code == 200
+    assert patched.json()["person"]["name"] == "Alice Updated"
+    assert patched.json()["person"]["metadata"] == {"level": 2}
+    assert added.status_code == 201
+    assert len(added.json()["faces"]) == 1
+    assert listed.status_code == 200
+    assert len(listed.json()["faces"]) == 2
+    assert all(
+        "embedding" not in face
+        and "crop_path" not in face
+        and "embedding_dimension" not in face
+        for face in listed.json()["faces"]
+    )
+    assert deleted_face.status_code == 204
+    assert person.json()["person"]["face_count"] == 1
+    assert deleted_person.status_code == 204
+    assert client.get("/v1/collections/employees/persons/employee-001").status_code == 404
+
+
+def test_committed_face_delete_failure_still_removes_persisted_crop_blob(
+    make_settings: Callable[..., Settings], image_bytes, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    settings = make_settings(save_face_crops=True)
+    with TestClient(create_app(settings)) as client:
+        collection = client.post(
+            "/v1/collections", json={"id": "employees", "name": "Employees"}
+        )
+        person = client.post(
+            "/v1/collections/employees/persons",
+            data={"id": "alice"},
+            files={"images": ("alice.png", image_bytes(51), "image/png")},
+        )
+        assert collection.status_code == 201
+        assert person.status_code == 201
+        face_id = person.json()["faces"][0]["id"]
+        stored = client.app.state.repository.get_face("employees", "alice", face_id)
+        crop = client.app.state.repository.get_face_crop(
+            "employees", "alice", face_id
+        )
+        assert stored is not None and stored["has_crop"] is True
+        assert crop is not None and crop["bytes"].startswith(b"\xff\xd8")
+
+        def commit_then_fail(collection_id, operation, **_kwargs):
+            result, mutation = operation()
+            assert mutation is not None
+            raise SearchMutationCommittedError(
+                collection_id,
+                mutation.revision,
+                "injected post-commit index failure",
+                committed_result=result,
+            )
+
+        monkeypatch.setattr(
+            client.app.state.search_indexes, "run_mutation", commit_then_fail
+        )
+        deleted = client.delete(
+            f"/v1/collections/employees/persons/alice/faces/{face_id}"
+        )
+
+        assert deleted.status_code == 503
+        assert deleted.json()["error"]["code"] == "search_index_unavailable"
+        assert deleted.json()["error"]["details"]["write_committed"] is True
+        assert client.app.state.repository.get_face("employees", "alice", face_id) is None
+        assert (
+            client.app.state.repository.get_face_crop("employees", "alice", face_id)
+            is None
+        )
+
+
+def test_invalid_metadata_and_duplicate_external_id_are_clear_errors(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client)
+    create_person(client, "employees", "one", seed=1)
+    invalid = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "bad-meta", "metadata": "[]"},
+        files={"images": ("face.png", image_bytes(2), "image/png")},
+    )
+    duplicate = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "two", "external_id": "external-one"},
+        files={"images": ("face.png", image_bytes(2), "image/png")},
+    )
+
+    assert invalid.status_code == 400
+    assert invalid.json()["error"]["code"] == "invalid_metadata"
+    assert duplicate.status_code == 409
+    assert duplicate.json()["error"]["code"] == "person_exists"
+
+
+def test_invalid_person_id_and_null_collection_patch_are_client_errors(
+    client: TestClient, create_collection, image_bytes
+) -> None:
+    create_collection(client)
+    invalid_person = client.post(
+        "/v1/collections/employees/persons",
+        data={"id": "bad id"},
+        files={"images": ("face.png", image_bytes(4), "image/png")},
+    )
+
+    assert invalid_person.status_code == 400
+    assert invalid_person.json()["error"]["code"] == "invalid_person_id"
+    for field in ("name", "description", "threshold", "metadata"):
+        response = client.patch(
+            "/v1/collections/employees",
+            json={field: None},
+        )
+        assert response.status_code == 400
+        assert response.json()["error"]["code"] == "invalid_request"
+
+
+def test_configured_default_threshold_is_used(
+    make_settings: Callable[..., Settings], image_bytes
+) -> None:
+    settings = make_settings(default_threshold=0.73)
+    with TestClient(create_app(settings)) as client:
+        collection = client.post(
+            "/v1/collections", json={"id": "defaults", "name": "Defaults"}
+        )
+        compared = client.post(
+            "/v1/compare",
+            files={
+                "source": ("source.png", image_bytes(1), "image/png"),
+                "target": ("target.png", image_bytes(1), "image/png"),
+            },
+        )
+
+    assert collection.json()["collection"]["default_threshold"] == 0.73
+    assert compared.json()["threshold"] == 0.73
diff --git a/server/tests/api/test_response_contracts.py b/server/tests/api/test_response_contracts.py
new file mode 100644
index 000000000..c2ba44a58
--- /dev/null
+++ b/server/tests/api/test_response_contracts.py
@@ -0,0 +1,253 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+
+from fastapi.routing import APIRoute
+from fastapi.testclient import TestClient
+from pydantic import BaseModel
+
+
+def _assert_response_model(
+    client: TestClient,
+    response,
+    *,
+    method: str,
+    path: str,
+) -> None:
+    assert response.status_code in {200, 201}, response.text
+    route = next(
+        route
+        for route in client.app.routes
+        if isinstance(route, APIRoute)
+        and route.path == path
+        and method in route.methods
+    )
+    model = route.response_model
+    assert isinstance(model, type) and issubclass(model, BaseModel)
+    model.model_validate(response.json())
+
+
+def test_every_json_success_response_conforms_to_its_public_model(
+    client: TestClient,
+    image_bytes: Callable[..., bytes],
+) -> None:
+    responses = [
+        ("GET", "/v1/health", client.get("/v1/health")),
+        ("GET", "/v1/system", client.get("/v1/system")),
+        ("GET", "/v1/models", client.get("/v1/models")),
+        (
+            "POST",
+            "/v1/detect",
+            client.post(
+                "/v1/detect",
+                files={"image": ("detect.png", image_bytes(1), "image/png")},
+            ),
+        ),
+        (
+            "POST",
+            "/v1/compare",
+            client.post(
+                "/v1/compare",
+                files={
+                    "source": ("source.png", image_bytes(2), "image/png"),
+                    "target": ("target.png", image_bytes(3), "image/png"),
+                },
+            ),
+        ),
+        (
+            "POST",
+            "/v1/embeddings",
+            client.post(
+                "/v1/embeddings",
+                files={"image": ("embedding.png", image_bytes(4), "image/png")},
+            ),
+        ),
+    ]
+
+    responses.append(
+        (
+            "POST",
+            "/v1/collections",
+            client.post(
+                "/v1/collections",
+                json={"id": "contracts", "name": "Contract fixtures"},
+            ),
+        )
+    )
+    responses.extend(
+        [
+            ("GET", "/v1/collections", client.get("/v1/collections")),
+            (
+                "GET",
+                "/v1/collections/{collection_id}",
+                client.get("/v1/collections/contracts"),
+            ),
+            (
+                "PATCH",
+                "/v1/collections/{collection_id}",
+                client.patch(
+                    "/v1/collections/contracts",
+                    json={"description": "Updated"},
+                ),
+            ),
+        ]
+    )
+
+    responses.append(
+        (
+            "POST",
+            "/v1/collections/{collection_id}/persons",
+            client.post(
+                "/v1/collections/contracts/persons",
+                data={"id": "alice", "name": "Alice"},
+                files={"images": ("alice.png", image_bytes(5), "image/png")},
+            ),
+        )
+    )
+    responses.extend(
+        [
+            (
+                "GET",
+                "/v1/collections/{collection_id}/persons",
+                client.get("/v1/collections/contracts/persons"),
+            ),
+            (
+                "GET",
+                "/v1/collections/{collection_id}/persons/{person_id}",
+                client.get("/v1/collections/contracts/persons/alice"),
+            ),
+            (
+                "PATCH",
+                "/v1/collections/{collection_id}/persons/{person_id}",
+                client.patch(
+                    "/v1/collections/contracts/persons/alice",
+                    json={"metadata": {"team": "vision"}},
+                ),
+            ),
+            (
+                "POST",
+                "/v1/collections/{collection_id}/persons/{person_id}/faces",
+                client.post(
+                    "/v1/collections/contracts/persons/alice/faces",
+                    files={"images": ("alice-2.png", image_bytes(6), "image/png")},
+                ),
+            ),
+            (
+                "GET",
+                "/v1/collections/{collection_id}/persons/{person_id}/faces",
+                client.get("/v1/collections/contracts/persons/alice/faces"),
+            ),
+            (
+                "POST",
+                "/v1/collections/{collection_id}/search",
+                client.post(
+                    "/v1/collections/contracts/search",
+                    files={"image": ("query.png", image_bytes(5), "image/png")},
+                ),
+            ),
+        ]
+    )
+
+    monitor = {
+        "id": "contract-monitor",
+        "name": "Contract monitor",
+        "enabled": False,
+        "source": {"type": "rtsp", "url": "rtsp://camera.test/live"},
+        "collection_id": "contracts",
+    }
+    responses.append(
+        (
+            "POST",
+            "/v1/monitors",
+            client.post("/v1/monitors", json=monitor),
+        )
+    )
+    responses.extend(
+        [
+            ("GET", "/v1/monitors", client.get("/v1/monitors")),
+            (
+                "GET",
+                "/v1/monitors/{monitor_id}",
+                client.get("/v1/monitors/contract-monitor"),
+            ),
+            (
+                "PATCH",
+                "/v1/monitors/{monitor_id}",
+                client.patch(
+                    "/v1/monitors/contract-monitor",
+                    json={"event_policy": {"emit_unknown": False}},
+                ),
+            ),
+            (
+                "GET",
+                "/v1/monitors/{monitor_id}/state",
+                client.get("/v1/monitors/contract-monitor/state"),
+            ),
+            (
+                "GET",
+                "/v1/monitors/{monitor_id}/events",
+                client.get("/v1/monitors/contract-monitor/events"),
+            ),
+        ]
+    )
+
+    assert len(responses) == 23
+    for method, path, response in responses:
+        _assert_response_model(
+            client,
+            response,
+            method=method,
+            path=path,
+        )
+
+
+def test_openapi_has_typed_success_and_special_response_contracts(
+    client: TestClient,
+) -> None:
+    schema = client.app.openapi()
+    json_operations = 0
+    for route in client.app.routes:
+        if (
+            not isinstance(route, APIRoute)
+            or not route.path.startswith("/v1/")
+            or route.status_code == 204
+            or route.path.endswith("/image")
+            or route.path.endswith("/preview.mjpeg")
+        ):
+            continue
+        for method in route.methods & {"GET", "POST", "PATCH", "DELETE"}:
+            success = schema["paths"][route.path][method.lower()]["responses"][
+                str(route.status_code or 200)
+            ]
+            response_schema = success["content"]["application/json"]["schema"]
+            assert response_schema, f"{method} {route.path}"
+            assert response_schema != {"type": "object"}, f"{method} {route.path}"
+            json_operations += 1
+    assert json_operations == 23
+
+    health_503 = schema["paths"]["/v1/health"]["get"]["responses"]["503"]
+    assert health_503["content"]["application/json"]["schema"] == {
+        "$ref": "#/components/schemas/HealthResponse"
+    }
+    for method in ("post", "patch"):
+        path = "/v1/monitors" if method == "post" else "/v1/monitors/{monitor_id}"
+        rate_limited = schema["paths"][path][method]["responses"]["429"]
+        assert rate_limited["content"]["application/json"]["schema"] == {
+            "$ref": "#/components/schemas/ErrorEnvelope"
+        }
+
+    jpeg = schema["paths"][
+        "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}/image"
+    ]["get"]["responses"]["200"]["content"]["image/jpeg"]["schema"]
+    assert jpeg == {"type": "string", "format": "binary"}
+    preview = schema["paths"]["/v1/monitors/{monitor_id}/preview.mjpeg"]["get"][
+        "responses"
+    ]["200"]["content"]["multipart/x-mixed-replace"]["schema"]
+    assert preview == {"type": "string", "format": "binary"}
+
+    for route in client.app.routes:
+        if not isinstance(route, APIRoute) or route.status_code != 204:
+            continue
+        for method in route.methods & {"DELETE"}:
+            response = schema["paths"][route.path][method.lower()]["responses"]["204"]
+            assert "content" not in response
diff --git a/server/tests/api/test_rtsp_streams.py b/server/tests/api/test_rtsp_streams.py
new file mode 100644
index 000000000..748f25c1e
--- /dev/null
+++ b/server/tests/api/test_rtsp_streams.py
@@ -0,0 +1,232 @@
+from __future__ import annotations
+
+import sqlite3
+
+from fastapi.testclient import TestClient
+from insightface_server.services.rtsp import MonitorSession
+
+
+def _monitor_payload(**overrides):
+    payload = {
+        "id": "front-gate",
+        "name": "Front gate",
+        "description": "Synthetic API test Monitor",
+        "enabled": False,
+        "source": {
+            "type": "rtsp",
+            "url": "rtsp://viewer:secret@192.168.1.20:8554/live?token=private",
+        },
+        "collection_id": "employees",
+        "inference_fps": 2.0,
+        "match_threshold": None,
+        "event_buffer_size": 1000,
+        "event_policy": {
+            "confirm_frames": 3,
+            "absence_timeout_seconds": 3,
+            "cooldown_seconds": 10,
+            "emit_unknown": True,
+        },
+        "preview_enabled": False,
+    }
+    payload.update(overrides)
+    return payload
+
+
+def test_monitor_crud_persists_config_without_exposing_credentials(
+    client: TestClient, create_collection
+) -> None:
+    create_collection(client)
+
+    created = client.post("/v1/monitors", json=_monitor_payload())
+
+    assert created.status_code == 201, created.text
+    monitor = created.json()["monitor"]
+    assert monitor["id"] == "front-gate"
+    assert monitor["source"] == {
+        "type": "rtsp",
+        "url": "rtsp://192.168.1.20:8554/live",
+    }
+    assert monitor["preview_enabled"] is False
+    assert monitor["runtime"]["status"] == "stopped"
+    assert "secret" not in created.text
+    assert "private" not in created.text
+
+    database_path = client.app.state.database.path
+    assert b"viewer" not in database_path.read_bytes()
+    assert b"secret" not in database_path.read_bytes()
+    with sqlite3.connect(database_path) as connection:
+        stored = connection.execute(
+            "SELECT source_url_ciphertext FROM monitors WHERE id='front-gate'"
+        ).fetchone()[0]
+    assert stored.startswith("v1.")
+
+    listed = client.get("/v1/monitors")
+    fetched = client.get("/v1/monitors/front-gate")
+    patched = client.patch(
+        "/v1/monitors/front-gate",
+        json={
+            "name": "Main entrance",
+            "event_buffer_size": 500,
+            "event_policy": {"confirm_frames": 5},
+            "preview_enabled": True,
+        },
+    )
+    state = client.get("/v1/monitors/front-gate/state")
+    events = client.get("/v1/monitors/front-gate/events", params={"limit": 25})
+
+    assert listed.status_code == 200
+    assert [item["id"] for item in listed.json()["monitors"]] == ["front-gate"]
+    assert fetched.status_code == 200
+    assert patched.status_code == 200
+    assert patched.json()["monitor"]["name"] == "Main entrance"
+    assert patched.json()["monitor"]["event_buffer_size"] == 500
+    assert patched.json()["monitor"]["event_policy"] == {
+        "confirm_frames": 5,
+        "absence_timeout_seconds": 3.0,
+        "cooldown_seconds": 10.0,
+        "emit_unknown": True,
+    }
+    assert patched.json()["monitor"]["preview_enabled"] is True
+    assert state.status_code == 200
+    assert state.json()["state"]["status"] == "stopped"
+    assert state.json()["state"]["preview"]["enabled"] is True
+    assert events.status_code == 200
+    assert events.json()["events"] == []
+    assert isinstance(events.json()["next_cursor"], str)
+
+    deleted = client.delete("/v1/monitors/front-gate")
+    assert deleted.status_code == 204
+    assert client.get("/v1/monitors/front-gate").status_code == 404
+
+
+def test_monitor_validation_conflicts_and_collection_delete_guard(
+    client: TestClient, create_collection
+) -> None:
+    create_collection(client)
+
+    invalid_scheme = client.post(
+        "/v1/monitors",
+        json=_monitor_payload(source={"type": "rtsp", "url": "http://camera/live"}),
+    )
+    missing_collection = client.post(
+        "/v1/monitors",
+        json=_monitor_payload(collection_id="missing"),
+    )
+    created = client.post("/v1/monitors", json=_monitor_payload())
+    duplicate = client.post("/v1/monitors", json=_monitor_payload())
+    collection_delete = client.delete(
+        "/v1/collections/employees",
+        params={"force": "true"},
+    )
+    empty_patch = client.patch("/v1/monitors/front-gate", json={})
+    null_collection = client.patch(
+        "/v1/monitors/front-gate",
+        json={"collection_id": None},
+    )
+    empty_policy = client.patch(
+        "/v1/monitors/front-gate",
+        json={"event_policy": {}},
+    )
+
+    assert invalid_scheme.status_code == 400
+    assert invalid_scheme.json()["error"]["code"] == "invalid_request"
+    assert missing_collection.status_code == 404
+    assert missing_collection.json()["error"]["code"] == "collection_not_found"
+    assert created.status_code == 201
+    assert duplicate.status_code == 409
+    assert duplicate.json()["error"]["code"] == "monitor_exists"
+    assert collection_delete.status_code == 409
+    assert collection_delete.json()["error"]["code"] == "collection_in_use"
+    assert empty_patch.status_code == 400
+    assert null_collection.status_code == 400
+    assert empty_policy.status_code == 400
+
+
+def test_monitor_preview_has_explicit_disabled_and_unavailable_errors(
+    client: TestClient, create_collection
+) -> None:
+    create_collection(client)
+    client.post("/v1/monitors", json=_monitor_payload())
+
+    disabled = client.get("/v1/monitors/front-gate/preview.mjpeg")
+    client.patch(
+        "/v1/monitors/front-gate",
+        json={"preview_enabled": True},
+    )
+    unavailable = client.get("/v1/monitors/front-gate/preview.mjpeg")
+
+    assert disabled.status_code == 409
+    assert disabled.json()["error"]["code"] == "preview_disabled"
+    assert unavailable.status_code == 503
+    assert unavailable.json()["error"]["code"] == "stream_unavailable"
+
+
+def test_monitor_event_cursor_is_incremental_signed_and_epoch_aware(
+    client: TestClient, create_collection
+) -> None:
+    create_collection(client)
+    client.post("/v1/monitors", json=_monitor_payload())
+    manager = client.app.state.monitors
+    row = client.app.state.repository.get_monitor("front-gate")
+    assert row is not None
+    session = MonitorSession(
+        manager.service,
+        manager._options(row),
+        max_faces=manager.max_faces,
+        preview_fps=manager.preview_fps,
+        jpeg_quality=manager.jpeg_quality,
+        open_timeout_seconds=1.0,
+        read_timeout_seconds=1.0,
+        reconnect_delay_seconds=0.1,
+        capture_factory=lambda *_args: None,
+    )
+    manager._sessions["front-gate"] = session
+    with session._condition:
+        for index in range(3):
+            session._append_event(f"test_{index}", now=float(index))
+
+    first = client.get("/v1/monitors/front-gate/events", params={"limit": 2})
+    cursor = first.json()["next_cursor"]
+    with session._condition:
+        session._append_event("test_3", now=3.0)
+    incremental = client.get(
+        "/v1/monitors/front-gate/events",
+        params={"limit": 10, "cursor": cursor},
+    )
+    invalid = client.get(
+        "/v1/monitors/front-gate/events",
+        params={"cursor": f"{cursor}altered"},
+    )
+    with session._condition:
+        session.stream_epoch = "replacement-epoch"
+        session._events.clear()
+        session._event_sequence = 0
+        session._append_event("test_new", now=4.0)
+    reset = client.get(
+        "/v1/monitors/front-gate/events",
+        params={"limit": 1, "cursor": incremental.json()["next_cursor"]},
+    )
+
+    assert [item["sequence"] for item in first.json()["events"]] == [2, 3]
+    assert [item["sequence"] for item in incremental.json()["events"]] == [4]
+    assert incremental.json()["stream_reset"] is False
+    assert invalid.status_code == 400
+    assert invalid.json()["error"]["code"] == "invalid_cursor"
+    assert reset.status_code == 200
+    assert reset.json()["stream_reset"] is True
+    assert [item["sequence"] for item in reset.json()["events"]] == [1]
+
+
+def test_monitor_routes_replace_ephemeral_rtsp_routes_in_openapi(
+    client: TestClient,
+) -> None:
+    paths = client.get("/openapi.json").json()["paths"]
+    for path in (
+        "/v1/monitors",
+        "/v1/monitors/{monitor_id}",
+        "/v1/monitors/{monitor_id}/state",
+        "/v1/monitors/{monitor_id}/events",
+        "/v1/monitors/{monitor_id}/preview.mjpeg",
+    ):
+        assert path in paths
+    assert "/v1/streams/rtsp" not in paths
diff --git a/server/tests/api/test_search_pagination.py b/server/tests/api/test_search_pagination.py
new file mode 100644
index 000000000..8422730f7
--- /dev/null
+++ b/server/tests/api/test_search_pagination.py
@@ -0,0 +1,134 @@
+from __future__ import annotations
+
+from fastapi.testclient import TestClient
+
+
+def test_search_uses_best_face_per_person_and_sorts_by_similarity(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client, threshold=0.0)
+    query = image_bytes(10)
+    alice = create_person(client, "employees", "alice", seed=10)
+    exact_face_id = alice["faces"][0]["id"]
+    create_person(client, "employees", "bob", seed=11)
+    # Add an inferior second sample to Alice; the person score must still use her exact face.
+    client.post(
+        "/v1/collections/employees/persons/alice/faces",
+        files={"images": ("alternate.png", image_bytes(12), "image/png")},
+    )
+
+    response = client.post(
+        "/v1/collections/employees/search",
+        data={"limit": "5", "threshold": "0"},
+        files={"image": ("query.png", query, "image/png")},
+    )
+
+    assert response.status_code == 200
+    matches = response.json()["matches"]
+    assert [item["person"]["id"] for item in matches] == ["alice", "bob"]
+    assert matches[0]["similarity"] == 1.0
+    assert matches[0]["matched_face_id"] == exact_face_id
+    assert matches[0]["similarity"] >= matches[1]["similarity"]
+
+
+def test_search_threshold_limit_and_no_face_behavior(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client, threshold=0.0)
+    create_person(client, "employees", "alice", seed=1)
+    create_person(client, "employees", "bob", seed=2)
+
+    limited = client.post(
+        "/v1/collections/employees/search",
+        data={"limit": "1", "threshold": "0"},
+        files={"image": ("query.png", image_bytes(1), "image/png")},
+    )
+    none = client.post(
+        "/v1/collections/employees/search",
+        data={"threshold": "0.999999"},
+        files={"image": ("other.png", image_bytes(99), "image/png")},
+    )
+    no_face = client.post(
+        "/v1/collections/employees/search",
+        files={"image": ("blank.png", image_bytes(blank=True), "image/png")},
+    )
+
+    assert limited.status_code == 200
+    assert len(limited.json()["matches"]) == 1
+    assert limited.json()["matches"][0]["person"]["id"] == "alice"
+    assert none.status_code == 200
+    assert none.json()["matches"] == []
+    assert no_face.status_code == 422
+    assert no_face.json()["error"]["code"] == "face_not_found"
+
+
+def test_deleted_face_is_immediately_absent_from_search(
+    client: TestClient, create_collection, create_person, image_bytes
+) -> None:
+    create_collection(client)
+    query = image_bytes(23)
+    created = create_person(client, "employees", "alice", seed=23)
+    face_id = created["faces"][0]["id"]
+
+    before = client.post(
+        "/v1/collections/employees/search",
+        data={"threshold": "0.99"},
+        files={"image": ("query.png", query, "image/png")},
+    )
+    deleted = client.delete(
+        f"/v1/collections/employees/persons/alice/faces/{face_id}"
+    )
+    after = client.post(
+        "/v1/collections/employees/search",
+        data={"threshold": "0.99"},
+        files={"image": ("query.png", query, "image/png")},
+    )
+
+    assert before.status_code == 200
+    assert before.json()["matches"][0]["matched_face_id"] == face_id
+    assert deleted.status_code == 204
+    assert after.status_code == 200
+    assert after.json()["matches"] == []
+
+
+def test_collection_cursor_is_opaque_paginated_and_tamper_evident(
+    client: TestClient, create_collection
+) -> None:
+    for collection_id in ("alpha", "bravo", "charlie"):
+        create_collection(client, collection_id)
+
+    first = client.get("/v1/collections?limit=1")
+    token = first.json()["next_cursor"]
+    second = client.get("/v1/collections", params={"limit": 1, "cursor": token})
+    payload, signature = token.split(".", 1)
+    tampered_signature = ("A" if signature[0] != "A" else "B") + signature[1:]
+    tampered = f"{payload}.{tampered_signature}"
+    rejected = client.get("/v1/collections", params={"cursor": tampered})
+
+    assert [item["id"] for item in first.json()["collections"]] == ["alpha"]
+    assert [item["id"] for item in second.json()["collections"]] == ["bravo"]
+    assert "alpha" not in token
+    assert rejected.status_code == 400
+    assert rejected.json()["error"]["code"] == "invalid_cursor"
+
+
+def test_cursor_cannot_be_reused_across_scopes(
+    client: TestClient, create_collection, create_person
+) -> None:
+    create_collection(client)
+    create_person(client, "employees", "alpha", seed=1)
+    create_person(client, "employees", "bravo", seed=2)
+    first = client.get("/v1/collections/employees/persons?limit=1")
+    token = first.json()["next_cursor"]
+
+    wrong_search = client.get(
+        "/v1/collections/employees/persons",
+        params={"limit": 1, "search": "a", "cursor": token},
+    )
+    wrong_collection = client.get("/v1/collections", params={"limit": 1, "cursor": token})
+
+    assert token
+    assert wrong_search.status_code == 400
+    assert wrong_search.json()["error"]["code"] == "invalid_cursor"
+    assert wrong_collection.status_code == 400
+    assert wrong_collection.json()["error"]["code"] == "invalid_cursor"
diff --git a/server/tests/api/test_system_auth_errors.py b/server/tests/api/test_system_auth_errors.py
new file mode 100644
index 000000000..19778fa7d
--- /dev/null
+++ b/server/tests/api/test_system_auth_errors.py
@@ -0,0 +1,323 @@
+from __future__ import annotations
+
+import sqlite3
+import uuid
+from collections.abc import Callable
+from dataclasses import replace
+
+import pytest
+from fastapi.testclient import TestClient
+from insightface_server.app import create_app
+from insightface_server.config import Settings
+
+
+def assert_request_id(response) -> str:
+    request_id = response.headers["x-request-id"]
+    uuid.UUID(request_id)
+    if response.content:
+        assert response.json()["request_id"] == request_id
+    return request_id
+
+
+def test_health_is_public_and_has_correlated_request_id(client: TestClient) -> None:
+    response = client.get("/v1/health")
+
+    assert response.status_code == 200
+    assert response.json()["status"] == "ready"
+    assert response.json()["auth_enabled"] is False
+    assert_request_id(response)
+    assert response.headers["x-content-type-options"] == "nosniff"
+    assert response.headers["permissions-policy"] == (
+        "camera=(), microphone=(), geolocation=()"
+    )
+
+
+def test_versioned_multilingual_markdown_is_served_read_only(client: TestClient) -> None:
+    assert client.get("/").status_code == 200
+    assert client.get("/docs").status_code == 200
+
+    english = client.get("/guide-content/en/user-guide.md")
+    assert english.status_code == 200
+    assert english.text.startswith("# InsightFace Server user guide")
+    assert english.headers["x-content-type-options"] == "nosniff"
+    assert "text/markdown" in english.headers["content-type"]
+
+    readme = client.get("/guide-content/zh/readme.md")
+    assert readme.status_code == 404
+
+    api_guide = client.get("/guide-content/zh/api.md")
+    assert api_guide.status_code == 200
+    assert "REST API使用手册" in api_guide.text
+    assert "POST /v1/collections/{collection_id}/search" in api_guide.text
+
+    maintainer = client.get("/guide-content/zh/maintainer.md")
+    assert maintainer.status_code == 200
+    assert maintainer.text.startswith("# InsightFace Server maintainer guide")
+    assert "English only" in maintainer.text
+
+    screenshot = client.get("/guide-images/customer/dashboard-en.jpg")
+    assert screenshot.status_code == 200
+    assert screenshot.headers["content-type"] == "image/jpeg"
+    assert screenshot.content.startswith(b"\xff\xd8")
+
+    assert client.get("/guide-content/it/user-guide.md").status_code == 404
+    assert client.get("/guide-content/en/not-a-document.md").status_code == 404
+
+
+def test_api_only_mode_does_not_register_web_ui_routes(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(web_ui_disabled=True)
+    with TestClient(create_app(settings)) as api_client:
+        assert api_client.get("/v1/health").status_code == 200
+        schema = api_client.get("/openapi.json")
+        assert schema.status_code == 200
+        assert "/v1/health" in schema.json()["paths"]
+        for path in (
+            "/",
+            "/docs",
+            "/assets/app.mjs",
+            "/guide-content/en/user-guide.md",
+            "/guide-content/en/api.md",
+            "/guide-content/en/maintainer.md",
+            "/guide-content/en/readme.md",
+            "/guide-images/customer/dashboard-en.jpg",
+        ):
+            response = api_client.get(path)
+            assert response.status_code == 404
+            assert response.json()["error"]["code"] == "route_not_found"
+
+
+def test_detector_configuration_is_visible_but_not_mutable_at_runtime(
+    client: TestClient,
+) -> None:
+    system = client.get("/v1/system")
+    mutation = client.patch(
+        "/v1/system/config",
+        json={"detector_input_sizes": [[128, 128]]},
+    )
+
+    assert system.status_code == 200
+    assert system.json()["safe_config"]["detector_input_sizes"] == [
+        [96, 96],
+        [512, 512],
+    ]
+    assert system.json()["safe_config"]["web_ui_disabled"] is False
+    assert system.json()["safe_config"]["config_file"] is None
+    assert mutation.status_code == 404
+    assert mutation.json()["error"]["code"] == "route_not_found"
+
+
+def test_authentication_rejects_missing_and_invalid_keys(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(auth_enabled=True, startup_api_key="correct-secret")
+    with TestClient(create_app(settings)) as client:
+        health = client.get("/v1/health")
+        assert health.status_code == 200
+        assert health.json()["auth_enabled"] is True
+
+        missing = client.get("/v1/system")
+        invalid = client.get("/v1/system", headers={"Authorization": "Bearer wrong-secret"})
+        valid = client.get("/v1/system", headers={"Authorization": "Bearer correct-secret"})
+
+    for response in (missing, invalid):
+        assert response.status_code == 401
+        assert response.json()["error"] == {
+            "code": "unauthorized",
+            "message": "A valid Bearer API key is required.",
+            "details": {},
+        }
+        assert_request_id(response)
+    assert valid.status_code == 200
+    assert valid.json()["api_key"] == {
+        "authentication_enabled": True,
+        "configured": True,
+    }
+
+    database = sqlite3.connect(settings.data_dir / "insightface-server.db")
+    row = database.execute("SELECT salt,digest FROM api_keys").fetchone()
+    database.close()
+    assert row is not None
+    assert b"correct-secret" not in row[0]
+    assert b"correct-secret" not in row[1]
+
+
+def test_startup_api_key_rotates_and_is_optional_after_initialization(
+    make_settings: Callable[..., Settings],
+) -> None:
+    initial = make_settings(auth_enabled=True, startup_api_key="first-secret")
+    with TestClient(create_app(initial)) as client:
+        assert client.get(
+            "/v1/system", headers={"Authorization": "Bearer first-secret"}
+        ).status_code == 200
+
+    rotated = replace(initial, startup_api_key="second-secret")
+    with TestClient(create_app(rotated)) as client:
+        assert client.get(
+            "/v1/system", headers={"Authorization": "Bearer first-secret"}
+        ).status_code == 401
+        assert client.get(
+            "/v1/system", headers={"Authorization": "Bearer second-secret"}
+        ).status_code == 200
+
+    retained = replace(initial, startup_api_key=None)
+    with TestClient(create_app(retained)) as client:
+        assert client.get(
+            "/v1/system", headers={"Authorization": "Bearer second-secret"}
+        ).status_code == 200
+
+    database = sqlite3.connect(initial.data_dir / "insightface-server.db")
+    rows = database.execute(
+        "SELECT label,active FROM api_keys ORDER BY active,id"
+    ).fetchall()
+    database.close()
+    assert rows == [("startup", 0), ("startup", 1)]
+
+
+def test_new_authenticated_database_requires_a_startup_key(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(auth_enabled=True, startup_api_key=None)
+    with pytest.raises(RuntimeError, match="set INSIGHTFACE_API_KEY"):
+        with TestClient(create_app(settings)):
+            pass
+
+
+def test_validation_and_domain_errors_use_one_error_shape(client: TestClient, image_bytes) -> None:
+    invalid_id = client.post("/v1/collections", json={"id": "bad id", "name": "Invalid"})
+    invalid_image = client.post(
+        "/v1/detect", files={"image": ("bad.jpg", b"not-an-image", "image/jpeg")}
+    )
+
+    assert invalid_id.status_code == 400
+    assert invalid_id.json()["error"]["code"] == "invalid_request"
+    assert invalid_image.status_code == 422
+    assert invalid_image.json()["error"] == {
+        "code": "invalid_image",
+        "message": "The image could not be decoded.",
+        "details": {},
+    }
+    assert_request_id(invalid_id)
+    assert_request_id(invalid_image)
+
+
+def test_request_body_limit_returns_413(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(max_request_bytes=32)
+    with TestClient(create_app(settings)) as client:
+        response = client.post(
+            "/v1/collections",
+            content=b"{" + b"x" * 100 + b"}",
+            headers={"content-type": "application/json"},
+        )
+
+    assert response.status_code == 413
+    assert response.json()["error"]["code"] == "request_too_large"
+    assert_request_id(response)
+
+
+def test_streamed_request_body_cannot_bypass_limit(
+    make_settings: Callable[..., Settings],
+) -> None:
+    settings = make_settings(max_request_bytes=48)
+
+    def chunks():
+        yield b'{"id":"streamed","name":"'
+        yield b"x" * 128
+        yield b'"}'
+
+    with TestClient(create_app(settings)) as client:
+        request = client.build_request(
+            "POST",
+            "/v1/collections",
+            content=chunks(),
+            headers={"content-type": "application/json"},
+        )
+        response = client.send(request)
+
+    assert response.status_code == 413
+    assert response.json()["error"]["code"] == "request_too_large"
+    assert_request_id(response)
+
+
+def test_unknown_collection_is_a_structured_404(client: TestClient) -> None:
+    response = client.get("/v1/collections/missing")
+
+    assert response.status_code == 404
+    assert response.json()["error"] == {
+        "code": "collection_not_found",
+        "message": "Collection 'missing' was not found.",
+        "details": {},
+    }
+    assert_request_id(response)
+
+
+def test_framework_404_and_405_use_the_standard_error_shape(client: TestClient) -> None:
+    missing = client.get("/v1/not-a-route")
+    wrong_method = client.put("/v1/health")
+
+    assert missing.status_code == 404
+    assert missing.json()["error"]["code"] == "route_not_found"
+    assert wrong_method.status_code == 405
+    assert wrong_method.json()["error"]["code"] == "method_not_allowed"
+    assert_request_id(missing)
+    assert_request_id(wrong_method)
+
+
+def test_openapi_declares_bearer_auth_and_custom_errors(client: TestClient) -> None:
+    schema = client.get("/openapi.json").json()
+
+    assert schema["components"]["securitySchemes"]["bearerAuth"] == {
+        "type": "http",
+        "scheme": "bearer",
+    }
+    assert schema["paths"]["/v1/system"]["get"]["security"] == [{"bearerAuth": []}]
+    assert "security" not in schema["paths"]["/v1/health"]["get"]
+    assert "400" in schema["paths"]["/v1/collections"]["post"]["responses"]
+
+
+def test_openapi_documents_resource_creation_and_deletion_statuses(
+    client: TestClient,
+) -> None:
+    paths = client.get("/openapi.json").json()["paths"]
+
+    assert "201" in paths["/v1/collections"]["post"]["responses"]
+    assert "204" in paths["/v1/collections/{collection_id}"]["delete"]["responses"]
+    assert "201" in paths["/v1/collections/{collection_id}/persons"]["post"]["responses"]
+    assert (
+        "204"
+        in paths["/v1/collections/{collection_id}/persons/{person_id}"]["delete"][
+            "responses"
+        ]
+    )
+    assert (
+        "201"
+        in paths["/v1/collections/{collection_id}/persons/{person_id}/faces"]["post"][
+            "responses"
+        ]
+    )
+    assert (
+        "204"
+        in paths[
+            "/v1/collections/{collection_id}/persons/{person_id}/faces/{face_id}"
+        ]["delete"]["responses"]
+    )
+
+
+def test_openapi_documents_enrollment_review_modes(client: TestClient) -> None:
+    schema = client.get("/openapi.json").json()
+    for path in (
+        "/v1/collections/{collection_id}/persons",
+        "/v1/collections/{collection_id}/persons/{person_id}/faces",
+    ):
+        request_schema = schema["paths"][path]["post"]["requestBody"]["content"][
+            "multipart/form-data"
+        ]["schema"]
+        reference = request_schema["$ref"].rsplit("/", 1)[-1]
+        review = schema["components"]["schemas"][reference]["properties"][
+            "review_mode"
+        ]
+        assert review["enum"] == ["off", "standard", "strict"]
+        assert review["default"] == "off"
diff --git a/server/tests/conftest.py b/server/tests/conftest.py
new file mode 100644
index 000000000..709985297
--- /dev/null
+++ b/server/tests/conftest.py
@@ -0,0 +1,139 @@
+from __future__ import annotations
+
+import sys
+from collections.abc import Callable
+from dataclasses import replace
+from io import BytesIO
+from pathlib import Path
+
+import numpy as np
+import pytest
+from fastapi.testclient import TestClient
+from PIL import Image
+
+SERVER_DIR = Path(__file__).resolve().parents[1]
+BACKEND_DIR = SERVER_DIR / "backend"
+if str(BACKEND_DIR) not in sys.path:
+    sys.path.insert(0, str(BACKEND_DIR))
+
+from insightface_server.app import create_app  # noqa: E402
+from insightface_server.config import Settings  # noqa: E402
+
+
+def synthetic_image(
+    seed: int = 1,
+    *,
+    width: int = 128,
+    height: int = 128,
+    blank: bool = False,
+) -> bytes:
+    """Create a deterministic synthetic PNG; it does not depict a real person."""
+
+    if blank:
+        pixels = np.full((height, width, 3), 127, dtype=np.uint8)
+    else:
+        generator = np.random.default_rng(seed)
+        pixels = generator.integers(24, 232, (height, width, 3), dtype=np.uint8)
+        # Add stable high-contrast geometry so the mock quality heuristics are exercised.
+        pixels[::8, :, :] = 245
+        pixels[:, ::8, :] = 12
+    stream = BytesIO()
+    Image.fromarray(pixels, mode="RGB").save(stream, format="PNG")
+    return stream.getvalue()
+
+
+@pytest.fixture
+def image_bytes() -> Callable[..., bytes]:
+    return synthetic_image
+
+
+@pytest.fixture
+def make_settings(tmp_path: Path) -> Callable[..., Settings]:
+    counter = 0
+
+    def factory(**overrides: object) -> Settings:
+        nonlocal counter
+        counter += 1
+        root = tmp_path / f"instance-{counter}"
+        defaults = Settings(
+            data_dir=root / "data",
+            models_dir=root / "models",
+            inference_mode="mock",
+            execution_provider="CPUExecutionProvider",
+            auth_enabled=False,
+            startup_api_key=None,
+            max_image_bytes=2 * 1024 * 1024,
+            max_image_pixels=2_000_000,
+            max_request_bytes=8 * 1024 * 1024,
+            max_registration_images=20,
+            request_timeout_seconds=10,
+            detector_threshold=0.5,
+            registration_min_score=0.6,
+            registration_min_quality=0.35,
+            registration_min_face_size=40,
+            default_threshold=0.4,
+            cors_origins=(),
+            log_level="WARNING",
+            save_face_crops=False,
+            device_id=0,
+        )
+        return replace(defaults, **overrides)
+
+    return factory
+
+
+@pytest.fixture
+def client(make_settings: Callable[..., Settings]):
+    settings = make_settings()
+    with TestClient(create_app(settings)) as test_client:
+        yield test_client
+
+
+@pytest.fixture
+def create_collection():
+    def factory(
+        client: TestClient,
+        collection_id: str = "employees",
+        *,
+        threshold: float = 0.4,
+    ) -> dict[str, object]:
+        response = client.post(
+            "/v1/collections",
+            json={
+                "id": collection_id,
+                "name": collection_id.title(),
+                "description": "Synthetic test collection",
+                "threshold": threshold,
+                "metadata": {"fixture": True},
+            },
+        )
+        assert response.status_code == 201, response.text
+        return response.json()["collection"]
+
+    return factory
+
+
+@pytest.fixture
+def create_person(image_bytes: Callable[..., bytes]):
+    def factory(
+        client: TestClient,
+        collection_id: str,
+        person_id: str,
+        *,
+        seed: int = 1,
+        name: str | None = None,
+    ) -> dict[str, object]:
+        response = client.post(
+            f"/v1/collections/{collection_id}/persons",
+            data={
+                "id": person_id,
+                "name": name or person_id.title(),
+                "external_id": f"external-{person_id}",
+                "metadata": '{"synthetic":true}',
+            },
+            files={"images": (f"{person_id}.png", image_bytes(seed), "image/png")},
+        )
+        assert response.status_code == 201, response.text
+        return response.json()
+
+    return factory
diff --git a/server/tests/cpu/test_real_inference.py b/server/tests/cpu/test_real_inference.py
new file mode 100644
index 000000000..ec1909030
--- /dev/null
+++ b/server/tests/cpu/test_real_inference.py
@@ -0,0 +1,134 @@
+"""Opt-in CPU checks against the private models and images mounted by a developer.
+
+Enable with ``INSIGHTFACE_RUN_REAL_CPU=1`` and provide a JSON array of at least
+two non-public image paths in ``INSIGHTFACE_TEST_IMAGES_JSON``. The model bundle
+is read from ``INSIGHTFACE_MODELS_DIR`` (``/models`` by default).
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from types import SimpleNamespace
+
+import cv2
+import numpy as np
+import pytest
+from insightface_server.inference import cosine_similarity, create_engine
+
+pytestmark = pytest.mark.cpu
+
+
+def _require_enabled() -> None:
+    if os.getenv("INSIGHTFACE_RUN_REAL_CPU") != "1":
+        pytest.skip(
+            "real CPU inference is opt-in; set INSIGHTFACE_RUN_REAL_CPU=1",
+            allow_module_level=True,
+        )
+
+
+_require_enabled()
+
+
+def _private_images() -> list[np.ndarray]:
+    raw = os.getenv("INSIGHTFACE_TEST_IMAGES_JSON")
+    if raw is None:
+        pytest.fail(
+            "INSIGHTFACE_TEST_IMAGES_JSON must be a JSON array of at least two "
+            "private image paths when real CPU tests are enabled"
+        )
+    try:
+        paths = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        pytest.fail(f"INSIGHTFACE_TEST_IMAGES_JSON is invalid JSON: {exc}")
+    if not isinstance(paths, list) or len(paths) < 2 or not all(
+        isinstance(path, str) and path for path in paths
+    ):
+        pytest.fail("INSIGHTFACE_TEST_IMAGES_JSON must contain at least two paths")
+
+    images: list[np.ndarray] = []
+    payloads: list[bytes] = []
+    for value in paths:
+        path = Path(value)
+        if not path.is_file():
+            pytest.fail(f"private test image does not exist: {path}")
+        payload = path.read_bytes()
+        image = cv2.imdecode(np.frombuffer(payload, np.uint8), cv2.IMREAD_COLOR)
+        if image is None:
+            pytest.fail(f"private test image is not decodable: {path}")
+        payloads.append(payload)
+        images.append(image)
+    if len(set(payloads)) != len(payloads):
+        pytest.fail("real CPU Top-K checks require distinct private images")
+    return images
+
+
+@pytest.fixture(scope="module")
+def engine():
+    models_dir = Path(os.getenv("INSIGHTFACE_MODELS_DIR", "/models"))
+    if not (models_dir / "manifest.json").is_file():
+        pytest.fail(f"model manifest is missing: {models_dir / 'manifest.json'}")
+    instance = create_engine(
+        SimpleNamespace(
+            inference_mode="onnx",
+            execution_provider="CPUExecutionProvider",
+            models_dir=models_dir,
+            device_id=0,
+            detector_threshold=0.5,
+        )
+    )
+    instance.startup()
+    yield instance
+    instance.close()
+
+
+def test_real_cpu_provider_warmup_and_detection_order(engine) -> None:
+    runtime = engine.runtime_summary()
+    assert runtime["execution_provider"] == "CPUExecutionProvider"
+    assert runtime["onnx_runtime_version"] == "1.27.0"
+    assert runtime["detector_session_providers"][0] == "CPUExecutionProvider"
+    assert runtime["recognizer_session_providers"][0] == "CPUExecutionProvider"
+
+    for image in _private_images():
+        first = engine.analyze(image, require_embeddings=True)
+        second = engine.analyze(image, require_embeddings=True)
+        assert first, "each opt-in CPU image must contain a usable face"
+        assert len(first) == len(second)
+        assert [face.area for face in first] == sorted(
+            (face.area for face in first), reverse=True
+        )
+        np.testing.assert_allclose(
+            [face.bbox for face in first],
+            [face.bbox for face in second],
+            atol=1e-4,
+            rtol=0.0,
+        )
+        for face in first:
+            assert face.embedding is not None
+            assert face.embedding.shape == (engine.summary.embedding_dimension,)
+            assert np.linalg.norm(face.embedding) == pytest.approx(1.0, abs=1e-5)
+
+
+def test_real_cpu_cosine_top_k_and_threshold_are_deterministic(engine) -> None:
+    observations = [
+        engine.analyze(image, require_embeddings=True, max_faces=1)[0]
+        for image in _private_images()
+    ]
+    query = observations[0].embedding
+    assert query is not None
+    similarities = []
+    for index, observation in enumerate(observations):
+        assert observation.embedding is not None
+        similarities.append((index, cosine_similarity(query, observation.embedding)))
+
+    first_ranking = sorted(similarities, key=lambda item: (-item[1], item[0]))
+    second_ranking = sorted(similarities, key=lambda item: (-item[1], item[0]))
+    assert first_ranking == second_ranking
+    assert first_ranking[0][0] == 0
+    assert first_ranking[0][1] == pytest.approx(1.0, abs=1e-6)
+
+    threshold = float(os.getenv("INSIGHTFACE_TEST_THRESHOLD", "0.3"))
+    assert 0.0 <= threshold <= 1.0
+    decisions = [similarity >= threshold for _, similarity in first_ranking]
+    assert decisions[0] is True
diff --git a/server/tests/cuda/test_real_cuda.py b/server/tests/cuda/test_real_cuda.py
new file mode 100644
index 000000000..279b6d735
--- /dev/null
+++ b/server/tests/cuda/test_real_cuda.py
@@ -0,0 +1,97 @@
+"""Opt-in strict CUDA checks using private models and a private face image."""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from types import SimpleNamespace
+
+import cv2
+import numpy as np
+import pytest
+from insightface_server.inference import create_engine
+
+pytestmark = pytest.mark.cuda
+
+
+if os.getenv("INSIGHTFACE_RUN_REAL_CUDA") != "1":
+    pytest.skip(
+        "real CUDA inference is opt-in; set INSIGHTFACE_RUN_REAL_CUDA=1",
+        allow_module_level=True,
+    )
+
+
+def _private_image() -> np.ndarray:
+    raw = os.getenv("INSIGHTFACE_TEST_IMAGES_JSON")
+    if raw is None:
+        pytest.fail(
+            "INSIGHTFACE_TEST_IMAGES_JSON must contain a private image path when "
+            "real CUDA tests are enabled"
+        )
+    try:
+        paths = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        pytest.fail(f"INSIGHTFACE_TEST_IMAGES_JSON is invalid JSON: {exc}")
+    if not isinstance(paths, list) or not paths or not isinstance(paths[0], str):
+        pytest.fail("INSIGHTFACE_TEST_IMAGES_JSON must be a non-empty JSON array")
+    path = Path(paths[0])
+    if not path.is_file():
+        pytest.fail(f"private test image does not exist: {path}")
+    image = cv2.imdecode(np.frombuffer(path.read_bytes(), np.uint8), cv2.IMREAD_COLOR)
+    if image is None:
+        pytest.fail(f"private test image is not decodable: {path}")
+    return image
+
+
+@pytest.fixture(scope="module")
+def engine():
+    models_dir = Path(os.getenv("INSIGHTFACE_MODELS_DIR", "/models"))
+    if not (models_dir / "manifest.json").is_file():
+        pytest.fail(f"model manifest is missing: {models_dir / 'manifest.json'}")
+    instance = create_engine(
+        SimpleNamespace(
+            inference_mode="onnx",
+            execution_provider="CUDAExecutionProvider",
+            models_dir=models_dir,
+            device_id=int(os.getenv("INSIGHTFACE_DEVICE_ID", "0")),
+            detector_threshold=0.5,
+        )
+    )
+    instance.startup()
+    yield instance
+    instance.close()
+
+
+def test_real_cuda_is_primary_and_strict_warmup_ran(engine) -> None:
+    runtime = engine.runtime_summary()
+    assert runtime["execution_provider"] == "CUDAExecutionProvider"
+    assert runtime["detector_session_providers"][0] == "CUDAExecutionProvider"
+    assert runtime["recognizer_session_providers"][0] == "CUDAExecutionProvider"
+    assert "CUDAExecutionProvider" in runtime["available_execution_providers"]
+    assert runtime["onnx_runtime_version"] == "1.27.0"
+    assert runtime["cuda_runtime_version"] == "12.9"
+    assert runtime["cudnn_version"] == "9.24.0"
+    assert runtime["gpus"]
+    for gpu in runtime["gpus"]:
+        assert gpu["name"]
+        assert gpu["compute_capability"]
+        assert gpu["driver_version"]
+
+    audits = runtime["strict_provider_audit"]
+    assert set(audits) == {"detector", "recognizer"}
+    for audit in audits.values():
+        assert audit["accepted"] is True
+        assert audit["cuda_kernel_count"] > 0
+        assert "CPU limited to small integer shape metadata" in audit["policy"]
+
+
+def test_real_cuda_inference_returns_normalized_embedding(engine) -> None:
+    observations = engine.analyze(
+        _private_image(), require_embeddings=True, max_faces=1
+    )
+    assert observations, "the opt-in CUDA image must contain a usable face"
+    embedding = observations[0].embedding
+    assert embedding is not None
+    assert embedding.shape == (engine.summary.embedding_dimension,)
+    assert np.linalg.norm(embedding) == pytest.approx(1.0, abs=1e-5)
diff --git a/server/tests/docker/test_container_contract.py b/server/tests/docker/test_container_contract.py
new file mode 100644
index 000000000..34d1ef24c
--- /dev/null
+++ b/server/tests/docker/test_container_contract.py
@@ -0,0 +1,275 @@
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import time
+import urllib.error
+import urllib.request
+import uuid
+from pathlib import Path
+
+import pytest
+
+SERVER_DIR = Path(__file__).resolve().parents[2]
+REPOSITORY_DIR = SERVER_DIR.parent
+
+
+def _read(relative: str) -> str:
+    return (REPOSITORY_DIR / relative).read_text(encoding="utf-8")
+
+
+def test_dockerfiles_are_pinned_non_root_and_model_free() -> None:
+    cpu = _read("server/docker/Dockerfile.cpu")
+    cuda = _read("server/docker/Dockerfile.cuda12")
+    cpu_lock = _read("server/requirements.cpu.lock")
+    cuda_lock = _read("server/requirements.cuda12.lock")
+
+    assert re.search(r"^FROM python:3\.11-[^\s]+@sha256:[0-9a-f]{64}$", cpu, re.MULTILINE)
+    assert re.search(
+        r"^FROM nvcr\.io/nvidia/cuda:12\.9\.1-runtime-ubuntu22\.04"
+        r"@sha256:[0-9a-f]{64}$",
+        cuda,
+        re.MULTILINE,
+    )
+    assert "onnxruntime==1.27.0" in cpu_lock
+    assert "onnxruntime-gpu" in cuda_lock
+    assert "onnxruntime_gpu-1.27.0-cp311" in cuda_lock
+    assert "sha256=" in cuda_lock
+    assert "libcudnn9-cuda-12=9.24.0.43-1" in cuda
+
+    for dockerfile in (cpu, cuda):
+        assert "USER 10001:10001" in dockerfile
+        assert 'VOLUME ["/data"]' in dockerfile
+        assert "HEALTHCHECK" in dockerfile
+        assert "/v1/health" in dockerfile
+        assert "INSIGHTFACE_CONFIG_FILE=/etc/insightface/server.toml" in dockerfile
+        assert "COPY server/config/server.toml /etc/insightface/server.toml" in dockerfile
+        assert (
+            "COPY server/LICENSING.md /opt/insightface/server/LICENSING.md"
+            in dockerfile
+        )
+        assert "server/LICENSE " not in dockerfile
+        assert "LICENSE-NOTICE.md" not in dockerfile
+        assert "COPY server/README" not in dockerfile
+        assert "COPY server/docs/api*.md /opt/insightface/server/docs/" in dockerfile
+        assert "COPY server/docs/user-guide*.md /opt/insightface/server/docs/" in dockerfile
+        assert (
+            "COPY server/docs/maintainer-guide.md "
+            "/opt/insightface/server/docs/maintainer-guide.md"
+        ) in dockerfile
+        assert "COPY server/docs/images /opt/insightface/server/docs/images" in dockerfile
+        assert "COPY models" not in dockerfile
+        assert "latest" not in dockerfile.lower()
+
+
+def test_offline_license_public_key_is_packaged_but_private_issuer_is_excluded() -> None:
+    public_key = (
+        SERVER_DIR
+        / "backend"
+        / "insightface_server"
+        / "licensing"
+        / "trusted_keys"
+        / "insightface-model-license-public-ed25519.pem"
+    )
+    assert public_key.read_text(encoding="ascii").startswith("-----BEGIN PUBLIC KEY-----")
+    for relative in ("server/docker/Dockerfile.cpu", "server/docker/Dockerfile.cuda12"):
+        assert "COPY server/backend /opt/insightface/server/backend" in _read(relative)
+    dockerignore = _read(".dockerignore")
+    assert ".private/" in dockerignore
+    assert "**/*.pem" in dockerignore
+    assert (
+        "!server/backend/insightface_server/licensing/trusted_keys/*.pem"
+        in dockerignore
+    )
+    assert ".private/" in _read(".gitignore")
+
+
+def test_compose_mounts_models_read_only_and_persists_data() -> None:
+    variants = {
+        "server/deploy/compose.cpu.yml": ("0.2.0-cpu", "18097"),
+        "server/deploy/compose.cuda12.yml": ("0.2.0-cuda12", "18098"),
+    }
+    for relative, (image_tag, host_port) in variants.items():
+        compose = _read(relative)
+        assert (
+            f"image: ghcr.io/deepinsight/insightface-server:{image_tag}" in compose
+        )
+        assert f'      - "{host_port}:8080"' in compose
+        assert "x-models-path: &models-path ../.models" in compose
+        assert "source: *models-path" in compose
+        assert "INSIGHTFACE_CPU_IMAGE" not in compose
+        assert "INSIGHTFACE_CUDA_IMAGE" not in compose
+        assert compose.count("INSIGHTFACE_MODELS_DIR: /models") == 1
+        assert "INSIGHTFACE_PORT" not in compose
+        assert "target: /data" in compose
+        assert "target: /models" in compose
+        assert "target: /etc/insightface/server.toml" in compose
+        assert "source: ../config/server.toml" in compose
+        assert "read_only: true" in compose
+        assert "user: \"10001:10001\"" in compose
+        assert "  models:\n" in compose
+        assert "profiles: [tools]" in compose
+        assert 'entrypoint: ["python", "-m", "insightface_server.models_cli"]' in compose
+        assert 'user: "${INSIGHTFACE_MODELS_UID:-1000}:${INSIGHTFACE_MODELS_GID:-1000}"' in compose
+        assert 'group_add: ["10001"]' in compose
+        assert "healthcheck:\n      disable: true" in compose
+        models_service = compose.split("\n  models:\n", 1)[1]
+        assert "target: /models" in models_service
+        assert "gpus: all" not in models_service
+        for variable in (
+            "INSIGHTFACE_DEFAULT_THRESHOLD",
+            "INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE",
+            "INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS",
+            "INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS",
+            "INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON",
+            "INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY",
+            "INSIGHTFACE_SEARCH_DEVICE_ID",
+            "INSIGHTFACE_SEARCH_TOPK_MODE",
+            "INSIGHTFACE_SEARCH_BUILD_BATCH_ROWS",
+        ):
+            assert variable in compose
+
+
+def test_build_context_excludes_unrelated_repository_data() -> None:
+    dockerignore = _read(".dockerignore")
+    assert dockerignore.splitlines()[0].startswith("#")
+    assert "**" in dockerignore.splitlines()
+    assert "!server/**" in dockerignore
+    for required in (
+        "!python-package/insightface/app/**",
+        "!python-package/insightface/model_zoo/**",
+        "!python-package/insightface/utils/**",
+        "!python-package/insightface/data/__init__.py",
+        "!python-package/insightface/data/image.py",
+        "!python-package/insightface/data/pickle_object.py",
+        "!python-package/insightface/thirdparty/__init__.py",
+    ):
+        assert required in dockerignore
+    assert "!python-package/insightface/**" not in dockerignore
+    assert "!python-package/insightface/data/images" not in dockerignore
+    assert "**/*.onnx" in dockerignore
+    assert "**/__pycache__/" in dockerignore
+
+
+def _docker(*args: str, check: bool = True) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["docker", *args],
+        check=check,
+        capture_output=True,
+        text=True,
+        timeout=120,
+    )
+
+
+def _request(url: str, *, method: str = "GET", body: dict[str, object] | None = None):
+    payload = None if body is None else json.dumps(body).encode("utf-8")
+    request = urllib.request.Request(
+        url,
+        data=payload,
+        method=method,
+        headers={"content-type": "application/json"} if payload is not None else {},
+    )
+    # Never send local CI traffic through a developer or runner proxy.
+    opener = urllib.request.build_opener(urllib.request.ProxyHandler({}))
+    with opener.open(request, timeout=5) as response:
+        return response.status, json.loads(response.read())
+
+
+def _wait_ready(base_url: str, *, container: str) -> None:
+    deadline = time.monotonic() + 60
+    last_error: Exception | None = None
+    while time.monotonic() < deadline:
+        try:
+            status, body = _request(f"{base_url}/v1/health")
+            if status == 200 and body.get("status") == "ready":
+                return
+        except (OSError, urllib.error.URLError, json.JSONDecodeError) as exc:
+            last_error = exc
+        time.sleep(0.5)
+    logs = _docker("logs", container, check=False).stdout
+    pytest.fail(f"mock container did not become ready: {last_error}\n{logs}")
+
+
+@pytest.mark.docker
+def test_mock_container_restart_preserves_unique_named_volume() -> None:
+    if os.getenv("INSIGHTFACE_RUN_DOCKER_TESTS") != "1":
+        pytest.skip(
+            "Docker integration is opt-in; set INSIGHTFACE_RUN_DOCKER_TESTS=1"
+        )
+    image = os.getenv("INSIGHTFACE_TEST_CPU_IMAGE")
+    if not image:
+        pytest.fail("INSIGHTFACE_TEST_CPU_IMAGE is required for Docker integration")
+    if _docker("image", "inspect", image, check=False).returncode != 0:
+        pytest.fail(f"Docker image does not exist: {image}")
+
+    unique = uuid.uuid4().hex
+    container = f"insightface-simple-ci-{unique}"
+    volume = f"insightface-simple-ci-data-{unique}"
+    created_container = False
+    created_volume = False
+    try:
+        _docker("volume", "create", volume)
+        created_volume = True
+        result = _docker(
+            "run",
+            "--detach",
+            "--name",
+            container,
+            "--read-only",
+            "--tmpfs",
+            "/tmp:rw,noexec,nosuid,nodev",
+            "--mount",
+            f"source={volume},target=/data",
+            "--env",
+            "INSIGHTFACE_INFERENCE_MODE=mock",
+            "--env",
+            "INSIGHTFACE_EXECUTION_PROVIDER=CPUExecutionProvider",
+            "--env",
+            "INSIGHTFACE_AUTH_ENABLED=false",
+            "--publish",
+            "127.0.0.1::8080",
+            "--entrypoint",
+            "python",
+            image,
+            "-m",
+            "uvicorn",
+            "insightface_server.app:app",
+            "--host",
+            "0.0.0.0",
+            "--port",
+            "8080",
+            "--no-access-log",
+        )
+        created_container = result.returncode == 0
+        port_output = _docker("port", container, "8080/tcp").stdout.strip()
+        port = int(port_output.rsplit(":", 1)[1])
+        base_url = f"http://127.0.0.1:{port}"
+        _wait_ready(base_url, container=container)
+
+        status, created = _request(
+            f"{base_url}/v1/collections",
+            method="POST",
+            body={
+                "id": f"persistent-{unique[:12]}",
+                "name": "CI persistence",
+                "threshold": 0.68,
+            },
+        )
+        assert status == 201
+        collection_id = created["collection"]["id"]
+
+        _docker("restart", container)
+        _wait_ready(base_url, container=container)
+        status, persisted = _request(f"{base_url}/v1/collections/{collection_id}")
+        assert status == 200
+        assert persisted["collection"]["id"] == collection_id
+        assert persisted["collection"]["person_count"] == 0
+        assert persisted["collection"]["face_count"] == 0
+    finally:
+        # Names are generated by this test. Never enumerate or prune shared state.
+        if created_container:
+            _docker("rm", "--force", container, check=False)
+        if created_volume:
+            _docker("volume", "rm", volume, check=False)
diff --git a/server/tests/frontend/core.test.mjs b/server/tests/frontend/core.test.mjs
new file mode 100644
index 000000000..bced06e0c
--- /dev/null
+++ b/server/tests/frontend/core.test.mjs
@@ -0,0 +1,378 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { ApiClient, ApiError } from "../../frontend/api.mjs";
+import { LANGUAGES, detectLocale, normalizeLocale, t } from "../../frontend/i18n.mjs";
+import { renderMarkdown } from "../../frontend/markdown.mjs";
+import {
+  DEFAULT_THRESHOLD,
+  SEARCH_PROFILES,
+  applySearchProfileAvailability,
+  authenticationEnabledFromHealth,
+  bestPersonMatches,
+  clamp,
+  faceLandmarks,
+  facePixels,
+  filterPeople,
+  formatCosine,
+  formatScore,
+  listItems,
+  normalizeThreshold,
+  parseExternalEmbeddings,
+  parseMetadata,
+  searchProfilesFromSystem,
+} from "../../frontend/core.mjs";
+
+function profileSelect(value = "fp32_v1") {
+  return {
+    value,
+    disabled: false,
+    options: SEARCH_PROFILES.map((profile) => ({ value: profile, disabled: false })),
+  };
+}
+
+test("locales match the public InsightFace website and normalize browser variants", () => {
+  assert.deepEqual(LANGUAGES.map(({ code }) => code), ["en", "zh", "ja", "de", "es", "fr", "ru", "pt", "ko"]);
+  assert.equal(normalizeLocale("zh-CN"), "zh");
+  assert.equal(normalizeLocale("pt_BR"), "pt");
+  assert.equal(normalizeLocale("it-IT"), "en");
+  assert.equal(detectLocale({ stored: "ja", languages: ["de-DE"] }), "ja");
+  assert.equal(detectLocale({ languages: ["it-IT", "fr-FR"] }), "fr");
+  assert.equal(t("Dashboard", {}, "zh"), "仪表盘");
+  assert.equal(t("Loading documentation…", {}, "ko"), "문서 로드 중…");
+  assert.equal(t("Open-source model license", {}, "zh"), "开源模型许可");
+  assert.equal(t("Camera monitoring", {}, "zh"), "摄像头监控");
+  assert.equal(t("No file selected", {}, "zh"), "未选择文件");
+  assert.equal(t("JPEG, PNG, or WebP", {}, "zh"), "JPEG、PNG 或 WebP");
+  assert.match(t("Commercial use requires a separate license.", {}, "ja"), /別途ライセンス/);
+  assert.equal(t(t("Detection {score}", { score: "0.998" }, "de"), {}, "zh"), "检测 0.998");
+  assert.equal(t(t("People in {collection}", { collection: "lfw" }, "ja"), {}, "ko"), "lfw의 인물");
+});
+
+test("bundled Markdown renderer supports portable guide links without executing HTML", () => {
+  const rendered = renderMarkdown("# Guide\n\n![Dashboard](docs/images/customer/dashboard-en.jpg)\n\n[API](api.zh-CN.md)\n\n[Maintainer](maintainer-guide.md)\n\n| A | B |\n| --- | --- |\n| 1 | 2 |\n\n```html\n<script>alert(1)</script>\n```\n\n[unsafe](javascript:alert(1))");
+  assert.match(rendered, /<h1>Guide<\/h1>/);
+  assert.match(rendered, /<table>/);
+  assert.match(rendered, /&lt;script&gt;alert\(1\)&lt;\/script&gt;/);
+  assert.doesNotMatch(rendered, /<script>/);
+  assert.match(rendered, /href="#"/);
+  assert.match(rendered, /href="api\.zh-CN\.md"/);
+  assert.match(rendered, /href="maintainer-guide\.md"/);
+  assert.match(rendered, /<img src="\/guide-images\/customer\/dashboard-en\.jpg" alt="Dashboard" loading="lazy">/);
+});
+
+test("numeric helpers clamp thresholds and format scores", () => {
+  assert.equal(DEFAULT_THRESHOLD, 0.4);
+  assert.equal(clamp(2, 0, 1), 1);
+  assert.equal(clamp(-1, 0, 1), 0);
+  assert.equal(normalizeThreshold("", 0.3), 0.3);
+  assert.equal(normalizeThreshold(1.5), 1);
+  assert.equal(normalizeThreshold(-1.5), 0);
+  assert.equal(formatScore(0.8234, 3), "0.823");
+  assert.equal(formatScore(0.9984), "0.998");
+  assert.equal(formatScore(undefined), "—");
+  assert.equal(formatCosine(0.8234), "0.8234");
+  assert.equal(formatCosine(-0.125), "-0.1250");
+  assert.equal(formatCosine(undefined), "—");
+});
+
+test("authentication controls follow public health capability with a secure legacy fallback", () => {
+  assert.equal(authenticationEnabledFromHealth({ auth_enabled: false }), false);
+  assert.equal(authenticationEnabledFromHealth({ auth_enabled: true }), true);
+  assert.equal(authenticationEnabledFromHealth({}), true);
+  assert.equal(authenticationEnabledFromHealth(null), true);
+});
+
+test("search profile capabilities preserve server order-independent public profiles", () => {
+  assert.equal(searchProfilesFromSystem({}), null);
+  assert.deepEqual(searchProfilesFromSystem({ search: { profiles: [] } }), []);
+  assert.deepEqual(
+    searchProfilesFromSystem({ search: { profiles: ["int8_x1000_v1", "int8_x736_v1", "fp32_v1", "private_profile", "fp32_v1"] } }),
+    ["fp32_v1", "int8_x736_v1", "int8_x1000_v1"],
+  );
+});
+
+test("create Collection profile control disables unsupported options and selects first supported", () => {
+  const select = profileSelect("fp16_v1");
+  assert.equal(applySearchProfileAvailability(select, ["bf16_v1", "int8_x736_v1", "int8_x1000_v1"]), "bf16_v1");
+  assert.deepEqual(
+    select.options.map((option) => [option.value, option.disabled]),
+    [
+      ["fp32_v1", true],
+      ["fp16_v1", true],
+      ["bf16_v1", false],
+      ["int8_x736_v1", false],
+      ["int8_x1000_v1", false],
+    ],
+  );
+  assert.equal(select.disabled, false);
+
+  assert.equal(applySearchProfileAvailability(select, ["int8_x1000_v1"]), "int8_x1000_v1");
+  assert.equal(applySearchProfileAvailability(select, []), "");
+  assert.equal(select.disabled, true);
+});
+
+test("unknown search capabilities keep create profiles available for older servers", () => {
+  const select = profileSelect("bf16_v1");
+  select.options.forEach((option) => { option.disabled = true; });
+  assert.equal(applySearchProfileAvailability(select, null), "bf16_v1");
+  assert.equal(select.disabled, false);
+  assert.ok(select.options.every((option) => option.disabled === false));
+});
+
+test("metadata accepts only JSON objects", () => {
+  assert.deepEqual(parseMetadata(""), {});
+  assert.deepEqual(parseMetadata('{"department":"sales"}'), { department: "sales" });
+  assert.throws(() => parseMetadata("not-json"), /valid JSON/);
+  assert.throws(() => parseMetadata("[]"), /JSON object/);
+  assert.throws(() => parseMetadata("null"), /JSON object/);
+});
+
+test("trusted external embeddings require one finite vector per image", () => {
+  assert.deepEqual(
+    parseExternalEmbeddings("[[1,0],[0,1]]", 2, 2),
+    [[1, 0], [0, 1]],
+  );
+  assert.throws(() => parseExternalEmbeddings("no", 1, 2), /valid JSON/);
+  assert.throws(() => parseExternalEmbeddings("{}", 1, 2), /array of vectors/);
+  assert.throws(() => parseExternalEmbeddings("[[1,2]]", 2, 2), /one external embedding for each image/);
+  assert.throws(() => parseExternalEmbeddings("[[]]", 1, 2), /non-empty array/);
+  assert.throws(() => parseExternalEmbeddings("[[1]]", 1, 2), /contain 2 values/);
+  assert.throws(() => parseExternalEmbeddings('[[1,"2"]]', 1, 2), /non-finite or non-numeric/);
+  assert.throws(() => parseExternalEmbeddings("[[0.5,0.5]]", 1, 2), /L2-normalized/);
+});
+
+test("listItems supports direct lists and conventional response keys", () => {
+  assert.deepEqual(listItems([1, 2]), [1, 2]);
+  assert.deepEqual(listItems({ collections: [{ id: "employees" }] }, ["collections"]), [{ id: "employees" }]);
+  assert.deepEqual(listItems({ items: [3] }), [3]);
+  assert.deepEqual(listItems(null), []);
+});
+
+test("facePixels handles pixel and normalized boxes", () => {
+  assert.deepEqual(facePixels({ bbox: { pixels: { x: 10, y: 20, width: 30, height: 40 } } }, 200, 100), { x: 10, y: 20, width: 30, height: 40 });
+  assert.deepEqual(facePixels({ bbox: { normalized: { left: 0.1, top: 0.2, width: 0.3, height: 0.4 } } }, 200, 100), { x: 20, y: 20, width: 60, height: 40 });
+  assert.equal(facePixels({}, 200, 100), null);
+});
+
+test("faceLandmarks normalizes tuple and object points", () => {
+  assert.deepEqual(faceLandmarks({ landmarks: [[1, 2], { x: 3, y: 4 }, null] }), [{ x: 1, y: 2 }, { x: 3, y: 4 }]);
+  assert.deepEqual(faceLandmarks({}), []);
+});
+
+test("bestPersonMatches keeps the best face per person and sorts descending", () => {
+  const samples = [
+    { person: { id: "alice" }, similarity: 0.72, matched_face_id: "a1" },
+    { person: { id: "bob" }, similarity: 0.81, matched_face_id: "b1" },
+    { person: { id: "alice" }, similarity: 0.91, matched_face_id: "a2" },
+    { person: { id: "carol" }, similarity: 0.4, matched_face_id: "c1" },
+  ];
+  assert.deepEqual(bestPersonMatches(samples, 0.68, 2).map((item) => item.matched_face_id), ["a2", "b1"]);
+});
+
+test("filterPeople searches ID, name, and external ID without changing input", () => {
+  const people = [
+    { id: "employee-001", name: "Alice", external_id: "HR-1001" },
+    { id: "employee-002", name: "Bob", external_id: "HR-1002" },
+  ];
+  assert.deepEqual(filterPeople(people, "alice"), [people[0]]);
+  assert.deepEqual(filterPeople(people, "1002"), [people[1]]);
+  assert.notEqual(filterPeople(people, ""), people);
+});
+
+function fakeResponse(status, payload, headers = {}) {
+  return {
+    status,
+    ok: status >= 200 && status < 300,
+    headers: new Headers(headers),
+    text: async () => payload === undefined ? "" : JSON.stringify(payload),
+  };
+}
+
+test("ApiClient keeps authentication in memory and creates multipart detect requests", async () => {
+  const calls = [];
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (url, options) => {
+      calls.push({ url, options });
+      return fakeResponse(200, { faces: [] }, { "x-request-id": "request-1" });
+    },
+  });
+  client.setApiKey("secret-value");
+  const image = new Blob(["image"], { type: "image/jpeg" });
+  const result = await client.detect(image, { maxFaces: 4, collectionId: "employees" });
+  assert.equal(result.request_id, "request-1");
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].url.pathname, "/v1/detect");
+  assert.equal(calls[0].options.method, "POST");
+  assert.equal(calls[0].options.headers.get("Authorization"), "Bearer secret-value");
+  assert.equal(calls[0].options.headers.get("Content-Type"), null, "browser must set the multipart boundary");
+  assert.equal(calls[0].options.body.get("max_faces"), "4");
+  assert.equal(calls[0].options.body.get("collection_id"), "employees");
+  client.clearApiKey();
+  assert.equal(client.hasApiKey, false);
+});
+
+test("ApiClient uses repeated images for multi-photo registration", async () => {
+  let body;
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (_url, options) => { body = options.body; return fakeResponse(201, { person: { id: "alice" }, faces: [] }); },
+  });
+  const first = new Blob(["one"], { type: "image/jpeg" });
+  const second = new Blob(["two"], { type: "image/png" });
+  await client.createPerson("employees", { id: "alice", metadata: { team: "sales" }, images: [first, second], reviewMode: "strict" });
+  assert.equal(body.getAll("images").length, 2);
+  assert.equal(body.get("metadata"), '{"team":"sales"}');
+  assert.equal(body.get("review_mode"), "strict");
+  assert.equal(body.get("embedding_mode"), "server");
+});
+
+test("ApiClient sends trusted embeddings and Collection contract only when selected", async () => {
+  const bodies = [];
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (_url, options) => { bodies.push(options.body); return fakeResponse(201, { faces: [], rejected_images: [] }); },
+  });
+  const images = [new Blob(["one"]), new Blob(["two"])];
+  await client.createPerson("employees", {
+    images,
+    embeddingMode: "external_trusted",
+    externalEmbeddings: [[0.1, -0.2], [0.3, 0.4]],
+    embeddingContractId: "contract-v1",
+  });
+  await client.addFaces("employees", "alice", [images[0]], {
+    embeddingMode: "external_trusted",
+    externalEmbeddings: [[0.1, -0.2]],
+    embeddingContractId: "contract-v1",
+  });
+
+  for (const body of bodies) {
+    assert.equal(body.get("embedding_mode"), "external_trusted");
+    assert.equal(body.get("embedding_contract_id"), "contract-v1");
+  }
+  assert.equal(bodies[0].get("external_embeddings"), "[[0.1,-0.2],[0.3,0.4]]");
+});
+
+test("ApiClient defaults additional FaceSample review to off", async () => {
+  let body;
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (_url, options) => { body = options.body; return fakeResponse(201, { faces: [], rejected_images: [] }); },
+  });
+  await client.addFaces("employees", "alice", [new Blob(["face"], { type: "image/jpeg" })]);
+  assert.equal(body.get("review_mode"), "off");
+  assert.equal(body.get("embedding_mode"), "server");
+});
+
+test("ApiClient downloads a face crop with in-memory Bearer authentication", async () => {
+  let request;
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (url, options) => {
+      request = { url, options };
+      return new Response(new Blob(["jpeg-crop"], { type: "image/jpeg" }), {
+        status: 200,
+        headers: { "content-type": "image/jpeg", "x-request-id": "crop-request" },
+      });
+    },
+  });
+  client.setApiKey("secret-value");
+
+  const crop = await client.getFaceCrop("team a", "alice/b", "face/1");
+
+  assert.equal(await crop.text(), "jpeg-crop");
+  assert.equal(request.url.pathname, "/v1/collections/team%20a/persons/alice%2Fb/faces/face%2F1/image");
+  assert.equal(request.options.headers.get("Authorization"), "Bearer secret-value");
+  assert.equal(request.options.headers.get("Accept"), "image/jpeg");
+});
+
+test("ApiClient creates, polls, updates, and deletes persistent Monitors", async () => {
+  const calls = [];
+  const responses = [
+    fakeResponse(201, { monitor: { id: "front-gate" } }),
+    fakeResponse(200, { state: { monitor_id: "front-gate", status: "running" } }),
+    fakeResponse(200, { events: [], next_cursor: "cursor-1" }),
+    fakeResponse(200, { monitor: { id: "front-gate", preview_enabled: true } }),
+    fakeResponse(204),
+  ];
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (url, options) => {
+      calls.push({ url, options });
+      return responses.shift();
+    },
+  });
+
+  await client.createMonitor({
+    id: "front-gate",
+    name: "Front gate",
+    url: "rtsp://camera.test/live",
+    collectionId: "employees",
+    inferenceFps: 2.5,
+    matchThreshold: 0.45,
+    previewEnabled: false,
+  });
+  await client.monitorState("front-gate");
+  await client.monitorEvents("front-gate", { cursor: "old", limit: 25 });
+  await client.updateMonitor("front-gate", { preview_enabled: true });
+  await client.deleteMonitor("front-gate");
+
+  assert.equal(calls[0].url.pathname, "/v1/monitors");
+  assert.deepEqual(JSON.parse(calls[0].options.body), {
+    id: "front-gate",
+    name: "Front gate",
+    description: "",
+    enabled: true,
+    source: { type: "rtsp", url: "rtsp://camera.test/live" },
+    collection_id: "employees",
+    inference_fps: 2.5,
+    match_threshold: 0.45,
+    event_buffer_size: 1000,
+    event_policy: {
+      confirm_frames: 3,
+      absence_timeout_seconds: 3,
+      cooldown_seconds: 10,
+      emit_unknown: true,
+    },
+    preview_enabled: false,
+  });
+  assert.equal(calls[1].url.pathname, "/v1/monitors/front-gate/state");
+  assert.equal(calls[1].options.method, "GET");
+  assert.equal(calls[2].url.searchParams.get("cursor"), "old");
+  assert.equal(calls[2].url.searchParams.get("limit"), "25");
+  assert.equal(calls[3].options.method, "PATCH");
+  assert.equal(calls[4].options.method, "DELETE");
+});
+
+test("ApiClient preserves collection-default threshold when creating a Monitor", async () => {
+  let request;
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async (url, options) => {
+      request = { url: new URL(url), options };
+      return fakeResponse(201, { monitor: { id: "front-gate" } });
+    },
+  });
+
+  await client.createMonitor({
+    id: "front-gate",
+    name: "Front gate",
+    url: "rtsp://camera.test/live",
+    collectionId: "employees",
+    matchThreshold: null,
+  });
+
+  assert.equal(JSON.parse(request.options.body).match_threshold, null);
+});
+
+test("ApiClient converts standard API errors and preserves request ID", async () => {
+  const client = new ApiClient("http://server.test", {
+    fetchFn: async () => fakeResponse(422, {
+      error: { code: "face_not_found", message: "No usable face was detected.", details: { field: "image" } },
+      request_id: "request-body-id",
+    }),
+  });
+  await assert.rejects(
+    client.health(),
+    (error) => error instanceof ApiError && error.status === 422 && error.code === "face_not_found" && error.requestId === "request-body-id",
+  );
+});
+
+test("ApiClient handles empty 204 responses", async () => {
+  const client = new ApiClient("http://server.test", { fetchFn: async () => fakeResponse(204) });
+  assert.equal(await client.deletePerson("employees", "alice"), null);
+});
diff --git a/server/tests/frontend/static.test.mjs b/server/tests/frontend/static.test.mjs
new file mode 100644
index 000000000..6fb1d2c07
--- /dev/null
+++ b/server/tests/frontend/static.test.mjs
@@ -0,0 +1,436 @@
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import test from "node:test";
+import { hasTranslation, LANGUAGES } from "../../frontend/i18n.mjs";
+
+const frontend = new URL("../../frontend/", import.meta.url);
+const server = new URL("../../", import.meta.url);
+
+async function source(name) {
+  return readFile(new URL(name, frontend), "utf8");
+}
+
+function decodeEntities(value) {
+  return value
+    .replaceAll("&amp;", "&")
+    .replaceAll("&lt;", "<")
+    .replaceAll("&gt;", ">")
+    .replaceAll("&quot;", '"')
+    .replaceAll("&#39;", "'")
+    .replaceAll("&nbsp;", " ");
+}
+
+function staticUiStrings(html) {
+  const visible = html
+    .replace(/<!--.*?-->/gs, "")
+    .replace(/<(script|style|code|pre|textarea)\b[^>]*>[\s\S]*?<\/\1>/gi, "<ignored></ignored>");
+  const messages = new Set();
+  for (const match of visible.matchAll(/\b(?:aria-label|placeholder|title)="([^"]+)"/g)) {
+    messages.add(decodeEntities(match[1]).trim());
+  }
+  for (const part of visible.split(/<[^>]+>/g)) {
+    const message = decodeEntities(part).replace(/\s+/g, " ").trim();
+    if (message) messages.add(message);
+  }
+  return [...messages];
+}
+
+function translatedLiteralStrings(script) {
+  const messages = new Set();
+  const patterns = [
+    /\bt\(\s*("(?:[^"\\]|\\.)*")/g,
+    /\btext:\s*("(?:[^"\\]|\\.)*")/g,
+    /\bstatusRow\(\s*("(?:[^"\\]|\\.)*")/g,
+    /\btoast\(\s*("(?:[^"\\]|\\.)*")/g,
+  ];
+  for (const pattern of patterns) {
+    for (const match of script.matchAll(pattern)) messages.add(JSON.parse(match[1]));
+  }
+  return messages;
+}
+
+test("main page exposes every requested product surface", async () => {
+  const html = await source("index.html");
+  for (const page of ["dashboard", "collections", "people", "detect", "compare", "search", "video", "system", "help"]) {
+    assert.match(html, new RegExp(`data-page="${page}"`));
+  }
+  const primaryNavigation = html.match(/<nav class="nav-list">([\s\S]*?)<\/nav>/)?.[1] || "";
+  assert.doesNotMatch(primaryNavigation, /href="\/docs"/);
+  assert.match(html, /href="\/docs"/);
+  assert.match(html, /DEVELOPER TOOLS/);
+  assert.match(html, /Developer schema/);
+  assert.match(html, /Content-Security-Policy/);
+});
+
+test("console and API reference expose the official nine-language selector", async () => {
+  const index = await source("index.html");
+  const docs = await source("openapi.html");
+  for (const code of ["en", "zh", "ja", "de", "es", "fr", "ru", "pt", "ko"]) {
+    assert.match(index, new RegExp(`<option value="${code}">`));
+    assert.match(docs, new RegExp(`<option value="${code}">`));
+  }
+  assert.match(await source("app.mjs"), /guide-content\/\$\{documentLocale\}/);
+  assert.match(await source("openapi.js"), /initializeI18n/);
+});
+
+test("repository README and bundled help sources exist for every supported locale", async () => {
+  const suffixes = ["", ".zh-CN", ".ja", ".de", ".es", ".fr", ".ru", ".pt", ".ko"];
+  const retiredDocumentation = [
+    "deployment.md",
+    "sdk.md",
+    "external-embeddings.md",
+    "models.md",
+    "gpu-support.md",
+    "security.md",
+    "search.md",
+    "development.md",
+    "container-versioning.md",
+  ];
+  const requiredReadmeMarkers = [
+    "Rekognition",
+    "0.2.0-cpu",
+    "0.2.0-cuda12",
+    "dashboard-en.jpg",
+    "collections-en.jpg",
+    "monitoring-en.jpg",
+    "models install buffalo_l",
+    "--accept-license",
+    "buffalo_m",
+    "buffalo_sc",
+    "antelopev2",
+    "MODEL.LICENSE",
+    "LICENSING.md",
+    "external_trusted",
+    "review_mode",
+    "center_largest",
+    "RTX 5090",
+    "INT8",
+    "50M+",
+    "Top-5",
+    "/v1/embeddings",
+    "preview.mjpeg",
+    "collection_model_mismatch",
+    "x-request-id",
+    "server/.models",
+    "make -C server build-cpu",
+    "make -C server build-cuda12",
+    "--pull never",
+    "INSIGHTFACE_AUTH_ENABLED",
+    "0.4",
+    "112x112",
+    "bounding-box",
+  ];
+  for (const suffix of suffixes) {
+    const readme = await readFile(new URL(`README${suffix}.md`, server), "utf8");
+    const guide = await readFile(new URL(`docs/user-guide${suffix}.md`, server), "utf8");
+    const api = await readFile(new URL(`docs/api${suffix}.md`, server), "utf8");
+    assert.match(readme, /^# InsightFace Server/m);
+    assert.match(guide, /^# .*InsightFace Server|^# InsightFace Server/m);
+    assert.match(api, /^# .*InsightFace Server|^# InsightFace Server/m);
+    assert.ok(readme.length > 1000, `README${suffix}.md is unexpectedly short`);
+    assert.ok(
+      readme.split(/\r?\n/).length <= 320,
+      `README${suffix}.md must remain an overview rather than duplicate detailed guides`,
+    );
+    assert.equal(
+      (readme.match(/^## /gm) || []).length,
+      9,
+      `README${suffix}.md must use the shared nine-section information architecture`,
+    );
+    assert.ok(guide.length > 3500, `user-guide${suffix}.md is unexpectedly short`);
+    assert.ok(api.length > 7000, `api${suffix}.md is unexpectedly short`);
+    assert.equal((api.match(/^### `(?:GET|POST|PATCH|DELETE) \/v1\//gm) || []).length, 29);
+    for (const marker of requiredReadmeMarkers) {
+      assert.ok(
+        readme.includes(marker),
+        `README${suffix}.md does not cover required marker: ${marker}`,
+      );
+    }
+    for (const value of [
+      /15[.,]8M/,
+      /30[.,]7M/,
+      /58[.,]9M/,
+      /3[.,]84 ms/,
+      /260[.,]81/,
+    ]) {
+      assert.match(
+        readme,
+        value,
+        `README${suffix}.md is missing a published RTX 5090 benchmark value`,
+      );
+    }
+    assert.ok(readme.includes(`docs/user-guide${suffix}.md`));
+    assert.ok(readme.includes(`docs/api${suffix}.md`));
+    assert.ok(readme.includes("docs/maintainer-guide.md"));
+    assert.ok(!readme.includes("LICENSE-NOTICE.md"));
+    assert.ok(!readme.includes("[LICENSE](LICENSE)"));
+    for (const retired of retiredDocumentation) {
+      assert.ok(
+        !readme.includes(`docs/${retired}`),
+        `README${suffix}.md still links retired fragmented documentation: ${retired}`,
+      );
+    }
+    for (const marker of [
+      "0.4",
+      "center_largest",
+      "external_trusted",
+      "bounding-box crop",
+      "INSIGHTFACE_AUTH_ENABLED",
+      "models verify",
+      "MODEL.LICENSE",
+      "server/config/server.toml",
+      "max_concurrency",
+      "fp32_v1",
+      "int8_x736_v1",
+      "CUDA Runtime 12.9.1",
+      "docker compose down -v",
+    ]) {
+      assert.ok(
+        guide.includes(marker),
+        `user-guide${suffix}.md is missing the shared behavior marker: ${marker}`,
+      );
+    }
+    assert.ok(
+      readme.indexOf("https://www.insightface.ai") <
+        readme.indexOf("dashboard-en.jpg"),
+      `README${suffix}.md must show the commercial model-license notice near the top`,
+    );
+  }
+  const maintainer = await readFile(new URL("docs/maintainer-guide.md", server), "utf8");
+  const licensing = await readFile(new URL("LICENSING.md", server), "utf8");
+  assert.match(licensing, /source code and its Python SDK are licensed under MIT/i);
+  assert.match(licensing, /non-commercial\s+academic research only/i);
+  assert.match(licensing, /https:\/\/www\.insightface\.ai/);
+  assert.match(licensing, /MODEL\.LICENSE/);
+  assert.doesNotMatch(licensing, /Permission is hereby granted/);
+  assert.ok(maintainer.length > 12000);
+  for (const marker of [
+    "Maintainer reference — English only",
+    "durable barrier",
+    "int8_x736_v1",
+    "device-resident",
+    "Public API documentation change gate",
+    "Container versioning and release",
+  ]) {
+    assert.ok(maintainer.includes(marker), `maintainer-guide.md is missing: ${marker}`);
+  }
+  for (const retired of retiredDocumentation) {
+    await assert.rejects(readFile(new URL(`docs/${retired}`, server), "utf8"));
+  }
+  const index = await source("index.html");
+  const app = await source("app.mjs");
+  const markdown = await source("markdown.mjs");
+  assert.match(index, /id="documentation-switcher"/);
+  assert.match(index, /id="documentation-toc"/);
+  assert.match(index, /id="documentation-filter"/);
+  assert.doesNotMatch(index, /data-document=/);
+  assert.doesNotMatch(index, />README</);
+  assert.match(app, /\{ name: "user-guide", label: "User guide", localized: true \}/);
+  assert.match(app, /\{ name: "api", label: "API guide", localized: true \}/);
+  assert.match(app, /\{ name: "maintainer", label: "Maintainer guide", localized: false \}/);
+  assert.match(app, /documentDefinition\.localized \? locale\(\) : "en"/);
+  assert.match(app, /documentationTarget/);
+  assert.match(markdown, /maintainer-guide/);
+  assert.doesNotMatch(app, /name: "readme"/);
+  assert.match(app, /renderDocumentationToc/);
+});
+
+test("frontend assets use the server's /assets mount and no external CDN", async () => {
+  const index = await source("index.html");
+  const docs = await source("openapi.html");
+  assert.doesNotMatch(`${index}\n${docs}`, /\/static\//);
+  assert.match(index, /href="\/assets\/styles\.css\?v=0\.2\.0-r13"/);
+  assert.match(index, /src="\/assets\/app\.mjs\?v=0\.2\.0-r13"/);
+  assert.match(docs, /href="\/assets\/openapi\.css"/);
+  assert.match(docs, /src="\/assets\/openapi\.js\?v=0\.2\.0-r13"/);
+  const externalReferences = [...`${index}\n${docs}`.matchAll(/(?:src|href)="(https?:\/\/[^\"]+)"/g)].map((match) => match[1]);
+  assert.deepEqual(externalReferences, ["https://www.insightface.ai"]);
+});
+
+test("System page shows the open-source model license and commercial link", async () => {
+  const html = await source("index.html");
+  assert.match(html, /non-commercial research use only/);
+  assert.match(html, /Commercial use requires a separate license/);
+  assert.match(html, /href="https:\/\/www\.insightface\.ai"/);
+});
+
+test("console and API reference use the InsightFace.ai brand system", async () => {
+  const index = await source("index.html");
+  const docs = await source("openapi.html");
+  const styles = `${await source("styles.css")}\n${await source("openapi.css")}`;
+  const mark = await source("insightface-mark.svg");
+
+  assert.match(index, /InsightFace Server dashboard/);
+  assert.match(`${index}\n${docs}`, /\/assets\/insightface-mark\.svg/);
+  assert.match(mark, /fill="#6366f1"/);
+  assert.match(mark, /<circle cx="12" cy="8" r="4"/);
+  for (const token of ["#09090b", "#111113", "#6366f1", "#818cf8", "#22d3ee"]) {
+    assert.match(styles, new RegExp(token));
+  }
+});
+
+test("API key and images are not persisted by browser code", async () => {
+  const scripts = `${await source("app.mjs")}\n${await source("api.mjs")}`;
+  assert.doesNotMatch(scripts, /localStorage|sessionStorage|indexedDB/);
+  assert.doesNotMatch(scripts, /console\.(?:log|info|debug)/);
+});
+
+test("API key UI is shown only when the public health response enables authentication", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  assert.match(html, /class="sidebar-footer" data-auth-ui hidden/);
+  assert.match(html, /id="auth-state" data-auth-ui hidden/);
+  assert.match(html, /class="metric-card" data-auth-ui hidden><span>API key in UI<\/span><strong class="metric-text" id="system-auth"/);
+  assert.match(html, /id="api-key-dialog" data-auth-ui hidden/);
+  assert.match(app, /authenticationEnabledFromHealth/);
+  assert.match(app, /\$\$\('\[data-auth-ui\]'\)/);
+  assert.match(app, /node\.hidden = !enabled/);
+});
+
+test("System page shows startup-only detector configuration", async () => {
+  const app = await source("app.mjs");
+  const api = await source("api.mjs");
+  const html = await source("index.html");
+  assert.match(app, /Detector input sizes/);
+  assert.match(app, /Detection threshold/);
+  assert.match(app, /Single-face selection/);
+  assert.match(`${app}\n${html}`, /center_largest/);
+  assert.match(app, /Startup config/);
+  assert.doesNotMatch(api, /\/v1\/system\/config/);
+  assert.doesNotMatch(app, /updateSystemConfig|getSystemConfig/);
+});
+
+test("enrollment forms expose review modes with off selected by default", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  for (const value of ["off", "standard", "strict"]) {
+    assert.match(`${html}\n${app}`, new RegExp(`<option value="${value}"`));
+  }
+  assert.match(html, /<select name="review_mode"><option value="off" selected>/);
+  assert.match(app, /<select name="review_mode"><option value="off" selected>/);
+  assert.match(app, /reviewMode: form\.elements\.review_mode\.value/);
+});
+
+test("enrollment UI makes trusted external features explicit and shows the contract", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  const api = await source("api.mjs");
+  assert.match(html, /<select name="embedding_mode"><option value="server" selected>/);
+  assert.match(`${html}\n${app}`, /value="external_trusted"/);
+  assert.match(`${html}\n${app}`, /name="external_embeddings"/);
+  assert.match(app, /collection\?\.embedding_contract_id/);
+  assert.match(app, /parseExternalEmbeddings/);
+  assert.match(api, /form\.append\("embedding_mode", embeddingMode\)/);
+  assert.match(api, /form\.append\("embedding_contract_id", embeddingContractId/);
+});
+
+test("camera monitoring uses persistent Monitors, memory events, and raw lazy preview", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  const api = await source("api.mjs");
+  assert.match(html, /id="rtsp-url" type="url"/);
+  assert.match(html, /id="rtsp-preview"/);
+  assert.match(html, /legend-swatch recognized/);
+  assert.match(html, /legend-swatch not-enrolled/);
+  assert.doesNotMatch(html, /type="file" accept="video\//);
+  assert.doesNotMatch(app, /mediaDevices\.getUserMedia/);
+  assert.doesNotMatch(app, /new VideoFrameController/);
+  assert.match(html, /name="preview_enabled" type="checkbox"/);
+  assert.doesNotMatch(html, /name="preview_enabled"[^>]*checked/);
+  assert.match(html, /id="rtsp-overlay"/);
+  assert.match(html, /id="monitor-edit"/);
+  assert.match(app, /client\.createMonitor/);
+  assert.match(app, /client\.updateMonitor\(editingMonitorId, common\)/);
+  assert.match(app, /Leave blank to keep the current RTSP URL/);
+  assert.match(app, /client\.monitorState/);
+  assert.match(app, /client\.monitorEvents/);
+  assert.match(app, /client\.monitorPreview/);
+  assert.match(app, /window\.setTimeout\(\(\) => void pollMonitor/);
+  assert.match(api, /\/v1\/monitors/);
+  assert.doesNotMatch(api, /\/v1\/streams\/rtsp/);
+  assert.match(app, /URL\.createObjectURL\(new Blob\(\[jpeg\]/);
+  assert.match(app, /drawMonitorOverlay/);
+});
+
+test("image uploads advertise JPEG, PNG, and WebP and search fields share a baseline", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  const styles = await source("styles.css");
+  assert.equal([...html.matchAll(/accept="image\/jpeg,image\/png,image\/webp"/g)].length, 5);
+  assert.match(app, /accept="image\/jpeg,image\/png,image\/webp"/);
+  assert.doesNotMatch(`${html}\n${app}`, /accept="image\/jpeg,image\/png"/);
+  assert.match(html, /class="form-grid two search-parameters"/);
+  assert.match(styles, /\.search-parameters \{ align-items: end; \}/);
+});
+
+test("form-grid controls stay top-aligned when a sibling has helper text", async () => {
+  const html = await source("index.html");
+  const styles = await source("styles.css");
+  assert.match(styles, /\.form-grid > label \{[\s\S]*?align-self: start;/);
+  assert.match(styles, /grid-template-rows: minmax\(2\.65em, auto\) auto auto/);
+  assert.match(styles, /\.field-caption \{[^}]*align-self: end/s);
+  assert.ok([...html.matchAll(/class="field-caption"/g)].length >= 20);
+  assert.match(styles, /\.compare-action-bar \{[^}]*grid-template-columns:[^;}]*minmax\(190px,[^;}]*minmax\(250px/s);
+});
+
+test("all static console and API-reference copy is translated in every supported locale", async () => {
+  const messages = staticUiStrings(`${await source("index.html")}\n${await source("openapi.html")}`);
+  for (const language of LANGUAGES.filter(({ code }) => code !== "en")) {
+    const missing = messages.filter((message) => !hasTranslation(message, language.code));
+    assert.deepEqual(missing, [], `${language.code} is missing: ${missing.join(" | ")}`);
+  }
+});
+
+test("dynamic result, status, and action copy is translated in every supported locale", async () => {
+  const app = await source("app.mjs");
+  const messages = translatedLiteralStrings(`${app}\n${await source("openapi.js")}`);
+  for (const template of app.matchAll(/innerHTML\s*=\s*`([\s\S]*?)`/g)) {
+    for (const message of staticUiStrings(template[1])) messages.add(message);
+  }
+  for (const message of [
+    "Detection failed", "Detecting…", "Comparison failed", "Comparing…", "Search failed", "Searching…",
+    "Could not load collections", "Could not load people", "Could not load face samples", "Could not load system diagnostics", "Refreshing…",
+    "Collection could not be created", "Creating…", "Collection could not be updated", "Saving…", "Person could not be registered", "Registering…",
+    "Face samples could not be added", "Uploading…", "Face sample could not be deleted", "Video frame failed", "Video could not start",
+  ]) messages.add(message);
+  for (const language of LANGUAGES.filter(({ code }) => code !== "en")) {
+    const missing = [...messages].filter((message) => !hasTranslation(message, language.code));
+    assert.deepEqual(missing, [], `${language.code} is missing dynamic copy: ${missing.join(" | ")}`);
+  }
+});
+
+test("self-hosted schema explorer reads the local OpenAPI document", async () => {
+  const html = await source("openapi.html");
+  const script = await source("openapi.js");
+  assert.match(html, /OpenAPI Schema/);
+  assert.match(html, /DEVELOPER TOOL · OPENAPI 3/);
+  assert.match(script, /fetch\("\/openapi\.json"/);
+  assert.match(script, /components\?\.schemas/);
+});
+
+test("Collection creation follows runtime search profiles without changing the pinned edit field", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  assert.match(html, /<select name="search_profile" required>/);
+  for (const profile of ["fp32_v1", "fp16_v1", "bf16_v1", "int8_x736_v1", "int8_x1000_v1"]) {
+    assert.match(html, new RegExp(`<option value="${profile}">`));
+  }
+  assert.match(html, /<input name="search_profile" readonly>/);
+  assert.match(app, /searchProfilesFromSystem\(state\.system\)/);
+  assert.match(app, /state\.system = result\.system;[\s\S]*?updateCreateCollectionProfiles\(\);[\s\S]*?renderSystem\(\);/);
+});
+
+test("Collection crop storage is opt-in and previews use revocable authenticated blobs", async () => {
+  const html = await source("index.html");
+  const app = await source("app.mjs");
+  const api = await source("api.mjs");
+  assert.match(html, /name="save_face_crops" type="checkbox"/);
+  assert.doesNotMatch(html, /name="save_face_crops"[^>]*checked/);
+  assert.match(app, /save_face_crops: form\.elements\.save_face_crops\.checked/);
+  assert.match(app, /state\.system\?\.safe_config\?\.save_face_crops/);
+  assert.match(app, /face\.has_crop/);
+  assert.match(api, /getFaceCrop\(/);
+  assert.match(api, /Authorization/);
+  assert.match(app, /URL\.createObjectURL\(blob\)/);
+  assert.match(app, /URL\.revokeObjectURL\(url\)/);
+});
diff --git a/server/tests/integration/test_consistency.py b/server/tests/integration/test_consistency.py
new file mode 100644
index 000000000..4bf189fd5
--- /dev/null
+++ b/server/tests/integration/test_consistency.py
@@ -0,0 +1,140 @@
+"""Opt-in comparison of real CPU and CUDA model execution.
+
+Run this on an NVIDIA host with ``onnxruntime-gpu==1.27.0``, private models,
+``INSIGHTFACE_RUN_REAL_CONSISTENCY=1`` and at least two private image paths in
+``INSIGHTFACE_TEST_IMAGES_JSON``.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from contextlib import ExitStack
+from pathlib import Path
+from types import SimpleNamespace
+
+import cv2
+import numpy as np
+import pytest
+from insightface_server.inference import cosine_similarity, create_engine
+
+pytestmark = pytest.mark.consistency
+
+
+if os.getenv("INSIGHTFACE_RUN_REAL_CONSISTENCY") != "1":
+    pytest.skip(
+        "CPU/CUDA consistency is opt-in; set INSIGHTFACE_RUN_REAL_CONSISTENCY=1",
+        allow_module_level=True,
+    )
+
+
+def _images() -> list[np.ndarray]:
+    raw = os.getenv("INSIGHTFACE_TEST_IMAGES_JSON")
+    if raw is None:
+        pytest.fail(
+            "INSIGHTFACE_TEST_IMAGES_JSON must be a JSON array of at least two "
+            "private image paths"
+        )
+    try:
+        paths = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        pytest.fail(f"INSIGHTFACE_TEST_IMAGES_JSON is invalid JSON: {exc}")
+    if not isinstance(paths, list) or len(paths) < 2 or not all(
+        isinstance(path, str) and path for path in paths
+    ):
+        pytest.fail("consistency tests require at least two private image paths")
+    images: list[np.ndarray] = []
+    for value in paths:
+        path = Path(value)
+        if not path.is_file():
+            pytest.fail(f"private test image does not exist: {path}")
+        image = cv2.imdecode(np.frombuffer(path.read_bytes(), np.uint8), cv2.IMREAD_COLOR)
+        if image is None:
+            pytest.fail(f"private test image is not decodable: {path}")
+        images.append(image)
+    return images
+
+
+def _engine(provider: str):
+    models_dir = Path(os.getenv("INSIGHTFACE_MODELS_DIR", "/models"))
+    if not (models_dir / "manifest.json").is_file():
+        pytest.fail(f"model manifest is missing: {models_dir / 'manifest.json'}")
+    instance = create_engine(
+        SimpleNamespace(
+            inference_mode="onnx",
+            execution_provider=provider,
+            models_dir=models_dir,
+            device_id=int(os.getenv("INSIGHTFACE_DEVICE_ID", "0")),
+            detector_threshold=0.5,
+        )
+    )
+    instance.startup()
+    return instance
+
+
+def test_cpu_cuda_real_inference_consistency() -> None:
+    images = _images()
+    box_tolerance = float(os.getenv("INSIGHTFACE_BOX_TOLERANCE_PIXELS", "3.0"))
+    similarity_tolerance = float(os.getenv("INSIGHTFACE_SIMILARITY_TOLERANCE", "0.02"))
+    threshold = float(os.getenv("INSIGHTFACE_TEST_THRESHOLD", "0.3"))
+    assert box_tolerance >= 0.0
+    assert 0.0 <= similarity_tolerance <= 1.0
+    assert 0.0 <= threshold <= 1.0
+
+    cpu = _engine("CPUExecutionProvider")
+    cuda = _engine("CUDAExecutionProvider")
+    with ExitStack() as stack:
+        stack.callback(cuda.close)
+        stack.callback(cpu.close)
+        assert cuda.runtime_summary()["strict_provider_audit"]
+        assert cpu.summary.model_digest == cuda.summary.model_digest
+        assert cpu.summary.embedding_dimension == cuda.summary.embedding_dimension
+
+        cpu_faces = [cpu.analyze(image, require_embeddings=True) for image in images]
+        cuda_faces = [cuda.analyze(image, require_embeddings=True) for image in images]
+        for cpu_result, cuda_result in zip(cpu_faces, cuda_faces, strict=True):
+            assert cpu_result and cuda_result
+            assert len(cpu_result) == len(cuda_result)
+            assert [face.area for face in cpu_result] == sorted(
+                (face.area for face in cpu_result), reverse=True
+            )
+            assert [face.area for face in cuda_result] == sorted(
+                (face.area for face in cuda_result), reverse=True
+            )
+            for cpu_face, cuda_face in zip(cpu_result, cuda_result, strict=True):
+                np.testing.assert_allclose(
+                    cpu_face.bbox, cuda_face.bbox, atol=box_tolerance, rtol=0.0
+                )
+                assert cpu_face.embedding is not None
+                assert cuda_face.embedding is not None
+                assert cpu_face.embedding.shape == cuda_face.embedding.shape
+                assert cpu_face.embedding.shape == (cpu.summary.embedding_dimension,)
+                assert np.linalg.norm(cpu_face.embedding) == pytest.approx(1.0, abs=1e-5)
+                assert np.linalg.norm(cuda_face.embedding) == pytest.approx(1.0, abs=1e-5)
+
+        cpu_query = cpu_faces[0][0].embedding
+        cuda_query = cuda_faces[0][0].embedding
+        assert cpu_query is not None and cuda_query is not None
+        cpu_scores: list[tuple[int, float]] = []
+        cuda_scores: list[tuple[int, float]] = []
+        for index, (cpu_result, cuda_result) in enumerate(
+            zip(cpu_faces, cuda_faces, strict=True)
+        ):
+            cpu_embedding = cpu_result[0].embedding
+            cuda_embedding = cuda_result[0].embedding
+            assert cpu_embedding is not None and cuda_embedding is not None
+            cpu_scores.append((index, cosine_similarity(cpu_query, cpu_embedding)))
+            cuda_scores.append((index, cosine_similarity(cuda_query, cuda_embedding)))
+
+        for (_, cpu_score), (_, cuda_score) in zip(cpu_scores, cuda_scores, strict=True):
+            assert abs(cpu_score - cuda_score) <= similarity_tolerance
+            assert (cpu_score >= threshold) == (cuda_score >= threshold)
+
+        cpu_ranking = [
+            index for index, _ in sorted(cpu_scores, key=lambda item: (-item[1], item[0]))
+        ]
+        cuda_ranking = [
+            index for index, _ in sorted(cuda_scores, key=lambda item: (-item[1], item[0]))
+        ]
+        assert cpu_ranking == cuda_ranking
+        assert cpu_ranking[0] == 0
diff --git a/server/tests/integration/test_persistence_concurrency.py b/server/tests/integration/test_persistence_concurrency.py
new file mode 100644
index 000000000..b24385338
--- /dev/null
+++ b/server/tests/integration/test_persistence_concurrency.py
@@ -0,0 +1,158 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import pytest
+from fastapi.testclient import TestClient
+from insightface_server.app import create_app
+from insightface_server.config import Settings
+from insightface_server.storage import Database, Repository
+
+
+def test_database_migration_is_idempotent_and_data_survives_restart(
+    make_settings: Callable[..., Settings], image_bytes
+) -> None:
+    settings = make_settings()
+    query = image_bytes(1)
+    app = create_app(settings)
+    with TestClient(app) as first:
+        collection = first.post("/v1/collections", json={"id": "employees", "name": "Employees"})
+        person = first.post(
+            "/v1/collections/employees/persons",
+            data={"id": "alice"},
+            files={"images": ("alice.png", query, "image/png")},
+        )
+        assert collection.status_code == 201
+        assert person.status_code == 201
+        face_id = person.json()["faces"][0]["id"]
+        monitor = first.post(
+            "/v1/monitors",
+            json={
+                "id": "front-gate",
+                "name": "Front gate",
+                "enabled": False,
+                "source": {
+                    "type": "rtsp",
+                    "url": "rtsp://viewer:secret@camera.test/live",
+                },
+                "collection_id": "employees",
+            },
+        )
+        assert monitor.status_code == 201
+
+    with TestClient(create_app(settings)) as restarted:
+        person = restarted.get("/v1/collections/employees/persons/alice")
+        faces = restarted.get("/v1/collections/employees/persons/alice/faces")
+        searched = restarted.post(
+            "/v1/collections/employees/search",
+            data={"threshold": "0.99"},
+            files={"image": ("query.png", query, "image/png")},
+        )
+        status = restarted.get("/v1/system")
+        restored_monitor = restarted.get("/v1/monitors/front-gate")
+
+    assert status.status_code == 200
+    assert status.json()["database"]["migration_count"] == 7
+    assert restored_monitor.status_code == 200
+    assert restored_monitor.json()["monitor"]["source"]["url"] == (
+        "rtsp://camera.test/live"
+    )
+    assert "secret" not in restored_monitor.text
+    assert status.json()["database"]["quick_check"] == "ok"
+    assert person.status_code == 200
+    assert person.json()["person"]["face_count"] == 1
+    assert len(faces.json()["faces"]) == 1
+    assert searched.status_code == 200
+    assert searched.json()["matches"] == [
+        {
+            "person": {
+                "id": "alice",
+                "name": None,
+                "external_id": None,
+                "metadata": {},
+                "face_count": 1,
+                "created_at": person.json()["person"]["created_at"],
+                "updated_at": person.json()["person"]["updated_at"],
+            },
+            "similarity": pytest.approx(1.0),
+            "matched_face_id": face_id,
+        }
+    ]
+    runtime_collection = status.json()["search"]["collections"][0]
+    assert runtime_collection["collection_id"] == "employees"
+    assert runtime_collection["state"] == "ready"
+    assert runtime_collection["live_rows"] == 1
+    assert runtime_collection["applied_revision"] == 1
+
+
+def test_sqlite_serializes_concurrent_writers(tmp_path: Path) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "data" / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+
+    def create(index: int) -> None:
+        repository.create_collection(
+            {
+                "id": f"collection-{index:02d}",
+                "name": f"Collection {index}",
+                "description": "",
+                "default_threshold": 0.68,
+                "metadata": {},
+                "model_id": "mock",
+                "model_version": "1",
+                "model_digest": "a" * 64,
+                "embedding_dimension": 8,
+                "preprocessing_version": "1",
+            }
+        )
+
+    with ThreadPoolExecutor(max_workers=8) as pool:
+        list(pool.map(create, range(30)))
+
+    assert repository.stats() == {
+        "collection_count": 30,
+        "person_count": 0,
+        "face_count": 0,
+    }
+    assert [item["id"] for item in repository.list_collections("", 100)] == [
+        f"collection-{index:02d}" for index in range(30)
+    ]
+
+
+def test_database_process_lock_rejects_second_server_for_same_data_store(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    path = tmp_path / "data" / "server.db"
+    first = Database(path, migrations)
+    second = Database(path, migrations)
+
+    first.acquire_process_lock()
+    try:
+        with pytest.raises(RuntimeError, match="already in use"):
+            second.acquire_process_lock()
+    finally:
+        first.release_process_lock()
+
+    # A clean shutdown releases the lease so the next Server can start normally.
+    second.acquire_process_lock()
+    second.release_process_lock()
+
+
+def test_concurrent_read_requests_are_stable(client: TestClient, image_bytes) -> None:
+    content = image_bytes(17)
+
+    def detect(_: int) -> tuple[int, int]:
+        response = client.post(
+            "/v1/detect",
+            files={"image": ("synthetic.png", content, "image/png")},
+        )
+        return response.status_code, len(response.json()["faces"])
+
+    with ThreadPoolExecutor(max_workers=8) as pool:
+        results = list(pool.map(detect, range(24)))
+
+    assert results == [(200, 1)] * 24
diff --git a/server/tests/sdk/test_client.py b/server/tests/sdk/test_client.py
new file mode 100644
index 000000000..9b75111ed
--- /dev/null
+++ b/server/tests/sdk/test_client.py
@@ -0,0 +1,678 @@
+from __future__ import annotations
+
+import importlib.util
+import io
+import json
+import sys
+from collections.abc import Callable
+from pathlib import Path
+from types import ModuleType
+
+import httpx
+import pytest
+
+
+def load_sdk() -> tuple[ModuleType, ModuleType]:
+    """Load the SDK under a private name so backend tests can share a process.
+
+    The server application and installable client intentionally both use the
+    public package name ``insightface_server``, but they are never installed in
+    the same runtime environment. Using a private alias here prevents SDK test
+    collection from shadowing the backend package in an all-tests pytest run.
+    """
+
+    package_name = "_insightface_server_sdk_under_test"
+    package_dir = Path(__file__).parents[2] / "sdk" / "python" / "src" / "insightface_server"
+    spec = importlib.util.spec_from_file_location(
+        package_name,
+        package_dir / "__init__.py",
+        submodule_search_locations=[str(package_dir)],
+    )
+    if spec is None or spec.loader is None:
+        raise RuntimeError("Unable to load the SDK package for tests")
+    package = importlib.util.module_from_spec(spec)
+    sys.modules[package_name] = package
+    spec.loader.exec_module(package)
+    results = sys.modules[f"{package_name}.results"]
+    return package, results
+
+
+sdk, sdk_results = load_sdk()
+Client = sdk.Client
+AuthenticationError = sdk.AuthenticationError
+ConflictError = sdk.ConflictError
+NotFoundError = sdk.NotFoundError
+PayloadTooLargeError = sdk.PayloadTooLargeError
+RateLimitError = sdk.RateLimitError
+ServerError = sdk.ServerError
+ServiceUnavailableError = sdk.ServiceUnavailableError
+TransportError = sdk.TransportError
+ValidationError = sdk.ValidationError
+CollectionPage = sdk_results.CollectionPage
+CompareResult = sdk_results.CompareResult
+DetectResult = sdk_results.DetectResult
+FaceRegistrationResult = sdk_results.FaceRegistrationResult
+MonitorEventPage = sdk_results.MonitorEventPage
+MonitorPage = sdk_results.MonitorPage
+PersonRegistrationResult = sdk_results.PersonRegistrationResult
+SearchResult = sdk_results.SearchResult
+
+Handler = Callable[[httpx.Request], httpx.Response]
+
+
+def client_for(handler: Handler, *, api_key: str | None = "secret") -> Client:
+    return Client(
+        "http://server.test/",
+        api_key=api_key,
+        transport=httpx.MockTransport(handler),
+    )
+
+
+def response(
+    status: int,
+    payload: object | None = None,
+    *,
+    request_id: str = "req-123",
+) -> httpx.Response:
+    headers = {"x-request-id": request_id}
+    if payload is None:
+        return httpx.Response(status, headers=headers)
+    return httpx.Response(status, headers=headers, json=payload)
+
+
+def test_health_sends_auth_and_exposes_request_metadata() -> None:
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.method == "GET"
+        assert request.url == httpx.URL("http://server.test/v1/health")
+        assert request.headers["authorization"] == "Bearer secret"
+        assert request.headers["user-agent"] == "insightface-server-python/0.2.0"
+        return response(200, {"status": "ok"})
+
+    with client_for(handler) as client:
+        result = client.health()
+
+    assert result["status"] == "ok"
+    assert result.status_code == 200
+    assert result.request_id == "req-123"
+    assert result.to_dict() == {"status": "ok"}
+
+
+def test_auth_header_is_optional_for_development_mode() -> None:
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert "authorization" not in request.headers
+        return response(200, {"status": "ok"})
+
+    with client_for(handler, api_key=None) as client:
+        client.health()
+
+
+def test_default_timeout_exceeds_the_server_request_deadline() -> None:
+    with client_for(lambda _request: response(200, {"status": "ok"})) as client:
+        assert client._http.timeout.connect == 65.0
+        assert client._http.timeout.read == 65.0
+        assert client._http.timeout.write == 65.0
+        assert client._http.timeout.pool == 65.0
+
+
+def test_detect_accepts_path_and_multipart_fields(tmp_path: Path) -> None:
+    image = tmp_path / "portrait.png"
+    image.write_bytes(b"fake-png-content")
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.url.path == "/v1/detect"
+        content_type = request.headers["content-type"]
+        assert content_type.startswith("multipart/form-data; boundary=")
+        body = request.content
+        assert b'name="image"; filename="portrait.png"' in body
+        assert b"fake-png-content" in body
+        assert b'name="max_faces"' in body and b"\r\n3\r\n" in body
+        assert b'name="collection_id"' in body and b"\r\nemployees\r\n" in body
+        return response(
+            200,
+            {"faces": [{"detection_score": 0.99}], "processing_ms": 4.5},
+        )
+
+    with client_for(handler) as client:
+        result = client.detect(image, max_faces=3, collection="employees")
+
+    assert isinstance(result, DetectResult)
+    assert result.faces[0]["detection_score"] == 0.99
+    assert result.processing_ms == 4.5
+
+
+def test_compare_accepts_bytes_and_file_like_without_moving_stream() -> None:
+    stream = io.BytesIO(b"target-bytes")
+    stream.name = "target.jpeg"  # type: ignore[attr-defined]
+    stream.seek(2)
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        body = request.content
+        assert b'name="source"; filename="source.jpg"' in body
+        assert b"source-bytes" in body
+        assert b'name="target"; filename="target.jpeg"' in body
+        assert b"rget-bytes" in body
+        assert b'name="threshold"' in body and b"\r\n0.68\r\n" in body
+        return response(
+            200,
+            {"matched": True, "similarity": 0.82, "threshold": 0.68},
+        )
+
+    with client_for(handler) as client:
+        result = client.compare(b"source-bytes", stream, threshold=0.68)
+
+    assert isinstance(result, CompareResult)
+    assert result.matched is True
+    assert result.similarity == 0.82
+    assert stream.tell() == 2
+
+
+def test_detect_accepts_non_seekable_binary_stream() -> None:
+    class NonSeekable:
+        def tell(self) -> int:
+            raise OSError("not seekable")
+
+        def read(self) -> bytes:
+            return b"stream-image"
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert b"stream-image" in request.content
+        return response(200, {"faces": [], "processing_ms": 1.0})
+
+    with client_for(handler) as client:
+        result = client.detect(NonSeekable())  # type: ignore[arg-type]
+
+    assert result.faces == []
+
+
+def test_embeddings_and_models_use_expected_routes() -> None:
+    seen: list[str] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen.append(request.url.path)
+        if request.url.path == "/v1/embeddings":
+            assert b'name="collection_id"' in request.content
+            assert b"\r\nemployees\r\n" in request.content
+            return response(200, {"faces": [{"embedding": [0.0, 1.0]}]})
+        return response(200, {"models": []})
+
+    with client_for(handler) as client:
+        assert len(client.embeddings(b"image", collection="employees").faces) == 1
+        assert client.models()["models"] == []
+        client.system()
+
+    assert seen == ["/v1/embeddings", "/v1/models", "/v1/system"]
+
+
+def test_collection_crud_serialization_and_pagination() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        if request.method == "DELETE":
+            return response(204)
+        if request.method == "GET" and request.url.path == "/v1/collections":
+            return response(200, {"collections": [{"id": "employees"}], "next_cursor": "n"})
+        return response(201 if request.method == "POST" else 200, {"id": "employees"})
+
+    with client_for(handler) as client:
+        created = client.create_collection(
+            "employees",
+            name="Company Employees",
+            description="staff",
+            threshold=0.68,
+            metadata={"site": "A"},
+            save_face_crops=True,
+            search_profile="bf16_v1",
+            capacity_rows=200_000,
+            max_faces_per_person=25,
+            load_policy="eager",
+            detector_input_sizes=[(96, 96), (512, 512)],
+            detector_threshold=0.5,
+            detector_nms_threshold=0.4,
+            single_face_selection="center_largest",
+        )
+        page = client.list_collections(limit=10, cursor="opaque")
+        fetched = client.get_collection("employees")
+        updated = client.update_collection(
+            "employees",
+            threshold=0.7,
+            save_face_crops=False,
+            capacity_rows=300_000,
+            max_faces_per_person=30,
+            load_policy="lazy",
+            detector_threshold=0.45,
+            single_face_selection="largest",
+        )
+        deleted = client.delete_collection("employees", force=True)
+
+    assert created.collection["id"] == "employees"
+    assert fetched.collection["id"] == "employees"
+    assert updated.collection["id"] == "employees"
+    assert isinstance(page, CollectionPage)
+    assert page.collections[0]["id"] == "employees"
+    assert page.next_cursor == "n"
+    assert deleted.status_code == 204 and deleted.to_dict() == {}
+    assert json.loads(requests[0].content) == {
+        "id": "employees",
+        "name": "Company Employees",
+        "description": "staff",
+        "threshold": 0.68,
+        "metadata": {"site": "A"},
+        "save_face_crops": True,
+        "detection": {
+            "input_sizes": [[96, 96], [512, 512]],
+            "threshold": 0.5,
+            "nms_threshold": 0.4,
+            "single_face_selection": "center_largest",
+        },
+        "search": {
+            "profile": "bf16_v1",
+            "capacity_rows": 200_000,
+            "max_faces_per_person": 25,
+            "load_policy": "eager",
+        },
+    }
+    assert dict(requests[1].url.params) == {"limit": "10", "cursor": "opaque"}
+    assert json.loads(requests[3].content) == {
+        "threshold": 0.7,
+        "save_face_crops": False,
+        "detection": {
+            "threshold": 0.45,
+            "single_face_selection": "largest",
+        },
+        "search": {
+            "capacity_rows": 300_000,
+            "max_faces_per_person": 30,
+            "load_policy": "lazy",
+        },
+    }
+    assert dict(requests[4].url.params) == {"force": "true"}
+
+
+def test_collection_crop_default_and_authenticated_download() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        if request.method == "POST":
+            payload = json.loads(request.content)
+            assert "save_face_crops" not in payload
+            assert payload["threshold"] == 0.4
+            return response(201, {"id": "employees", "save_face_crops": False})
+        assert request.headers["authorization"] == "Bearer secret"
+        assert request.headers["accept"] == "image/jpeg"
+        return httpx.Response(
+            200,
+            headers={"content-type": "image/jpeg", "x-request-id": "crop-request"},
+            content=b"jpeg-crop-bytes",
+        )
+
+    with client_for(handler) as client:
+        client.create_collection("employees", name="Employees")
+        crop = client.get_face_crop("team a", "alice/b", "face/1")
+
+    assert crop == b"jpeg-crop-bytes"
+    assert requests[1].url.raw_path == (
+        b"/v1/collections/team%20a/persons/alice%2Fb/faces/face%2F1/image"
+    )
+
+
+def test_monitor_crud_state_and_event_cursor() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        if request.method == "DELETE":
+            return response(204)
+        if request.url.path.endswith("/state"):
+            return response(200, {"state": {"status": "running"}})
+        if request.url.path.endswith("/events"):
+            return response(
+                200,
+                {
+                    "events": [],
+                    "next_cursor": "next",
+                    "has_more": False,
+                    "truncated": True,
+                    "stream_reset": False,
+                },
+            )
+        if request.url.path == "/v1/monitors" and request.method == "GET":
+            return response(
+                200,
+                {"monitors": [{"id": "front-gate"}], "next_cursor": "page"},
+            )
+        return response(201 if request.method == "POST" else 200, {
+            "monitor": {"id": "front-gate", "preview_enabled": False}
+        })
+
+    with client_for(handler) as client:
+        created = client.create_monitor(
+            "front-gate",
+            name="Front gate",
+            rtsp_url="rtsp://camera.test/live",
+            collection="employees",
+            inference_fps=2.5,
+            match_threshold=0.45,
+        )
+        page = client.list_monitors(limit=10, cursor="list-cursor")
+        client.get_monitor("front-gate")
+        client.update_monitor(
+            "front-gate",
+            preview_enabled=True,
+            confirm_frames=4,
+        )
+        state_result = client.monitor_state("front-gate")
+        events = client.monitor_events(
+            "front-gate",
+            limit=25,
+            cursor="event-cursor",
+        )
+        deleted = client.delete_monitor("front-gate")
+
+    assert created.monitor["id"] == "front-gate"
+    assert isinstance(page, MonitorPage)
+    assert page.next_cursor == "page"
+    assert state_result.state["status"] == "running"
+    assert isinstance(events, MonitorEventPage)
+    assert events.next_cursor == "next"
+    assert events.has_more is False
+    assert events.truncated is True
+    assert events.stream_reset is False
+    assert deleted.status_code == 204
+    assert json.loads(requests[0].content) == {
+        "id": "front-gate",
+        "name": "Front gate",
+        "description": "",
+        "enabled": True,
+        "source": {"type": "rtsp", "url": "rtsp://camera.test/live"},
+        "collection_id": "employees",
+        "inference_fps": 2.5,
+        "match_threshold": 0.45,
+        "event_buffer_size": 1000,
+        "event_policy": {
+            "confirm_frames": 3,
+            "absence_timeout_seconds": 3.0,
+            "cooldown_seconds": 10.0,
+            "emit_unknown": True,
+        },
+        "preview_enabled": False,
+    }
+    assert dict(requests[1].url.params) == {
+        "limit": "10",
+        "cursor": "list-cursor",
+    }
+    assert json.loads(requests[3].content) == {
+        "preview_enabled": True,
+        "event_policy": {"confirm_frames": 4},
+    }
+    assert dict(requests[5].url.params) == {
+        "limit": "25",
+        "cursor": "event-cursor",
+    }
+
+
+def test_monitor_patch_preserves_explicit_false_without_defaulting_other_policy_fields() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        return response(200, {"monitor": {"id": "front-gate"}})
+
+    with client_for(handler) as client:
+        client.update_monitor("front-gate", emit_unknown=False)
+
+    assert json.loads(requests[0].content) == {
+        "event_policy": {"emit_unknown": False},
+    }
+
+
+def test_person_registration_sends_repeated_images_and_metadata() -> None:
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert request.url.path == "/v1/collections/employees/persons"
+        body = request.content
+        assert body.count(b'name="images";') == 2
+        assert b"first-image" in body and b"second-image" in body
+        assert b'name="id"' in body and b"employee-001" in body
+        assert b'name="external_id"' in body and b"HR-1001" in body
+        assert b'{"department":"sales"}' in body
+        assert b'name="review_mode"' in body and b"\r\nstrict\r\n" in body
+        return response(
+            201,
+            {
+                "person": {"id": "employee-001", "face_count": 1},
+                "faces": [{"id": "face-1"}],
+                "rejected_images": [{"index": 1, "reason": "multiple_faces"}],
+            },
+        )
+
+    with client_for(handler) as client:
+        result = client.add_person(
+            "employees",
+            person_id="employee-001",
+            name="Alice",
+            external_id="HR-1001",
+            metadata={"department": "sales"},
+            images=[b"first-image", io.BytesIO(b"second-image")],
+            review_mode="strict",
+        )
+
+    assert isinstance(result, PersonRegistrationResult)
+    assert result.person["id"] == "employee-001"
+    assert result.faces[0]["id"] == "face-1"
+    assert result.rejected_images[0]["reason"] == "multiple_faces"
+
+
+def test_enrollment_review_mode_defaults_to_off() -> None:
+    def handler(request: httpx.Request) -> httpx.Response:
+        assert b'name="review_mode"' in request.content
+        assert b"\r\noff\r\n" in request.content
+        assert b'name="embedding_mode"' in request.content
+        assert b"\r\nserver\r\n" in request.content
+        return response(
+            201,
+            {"person": {"id": "alice"}, "faces": [], "rejected_images": []},
+        )
+
+    with client_for(handler) as client:
+        client.create_person("employees", images=[b"image"])
+
+
+def test_external_trusted_registration_serializes_vectors_and_contract() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        return response(
+            201,
+            {"person": {"id": "alice"}, "faces": [], "rejected_images": []},
+        )
+
+    class FloatLike:
+        def __init__(self, value: float) -> None:
+            self.value = value
+
+        def __float__(self) -> float:
+            return self.value
+
+    with client_for(handler) as client:
+        client.add_person(
+            "employees",
+            person_id="alice",
+            images=[b"first", b"second"],
+            external_embeddings=[
+                [FloatLike(1.0), FloatLike(0.0)],
+                (value for value in (0.0, 1.0)),
+            ],
+            embedding_contract_id="contract-v1",
+            review_mode="strict",
+        )
+        client.add_faces(
+            "employees",
+            "alice",
+            [b"third"],
+            external_embeddings=[[1.0, 0.0]],
+            embedding_contract_id="contract-v1",
+        )
+
+    for request in requests:
+        body = request.content
+        assert b'name="embedding_mode"' in body
+        assert b"\r\nexternal_trusted\r\n" in body
+        assert b'name="embedding_contract_id"' in body
+        assert b"\r\ncontract-v1\r\n" in body
+    assert b'[[1.0,0.0],[0.0,1.0]]' in requests[0].content
+    assert b'[[1.0,0.0]]' in requests[1].content
+
+
+@pytest.mark.parametrize(
+    ("external_embeddings", "contract_id", "message"),
+    [
+        ([[1.0, 0.0]], None, "embedding_contract_id is required"),
+        (None, "contract-v1", "requires external_embeddings"),
+        ([[1.0, 0.0]], "contract-v1", "count must equal images count"),
+        ([[float("nan")], [1.0]], "contract-v1", "NaN or infinity"),
+        ([[] , [1.0]], "contract-v1", "must not be empty"),
+        ([[0.5, 0.5], [1.0, 0.0]], "contract-v1", "L2-normalized"),
+    ],
+)
+def test_external_trusted_registration_rejects_invalid_client_input(
+    external_embeddings: object,
+    contract_id: str | None,
+    message: str,
+) -> None:
+    def handler(_: httpx.Request) -> httpx.Response:
+        raise AssertionError("HTTP must not be called")
+
+    with client_for(handler) as client, pytest.raises(ValueError, match=message):
+        client.create_person(
+            "employees",
+            images=[b"first", b"second"],
+            external_embeddings=external_embeddings,  # type: ignore[arg-type]
+            embedding_contract_id=contract_id,
+        )
+
+
+def test_person_face_crud_and_search_routes_are_encoded() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        path = request.url.path
+        if path.endswith("/search"):
+            return response(200, {"matches": [], "threshold": 0.68})
+        if request.method == "DELETE":
+            return response(204)
+        if path.endswith("/faces") and request.method == "GET":
+            return response(200, {"faces": [], "next_cursor": None})
+        if path.endswith("/faces") and request.method == "POST":
+            return response(200, {"faces": [], "rejected_images": []})
+        if request.method == "GET" and path.endswith("/persons"):
+            return response(200, {"persons": []})
+        return response(200, {"id": "alice/b"})
+
+    with client_for(handler) as client:
+        client.list_persons("team a", limit=7, search="alice")
+        client.get_person("team a", "alice/b")
+        client.update_person("team a", "alice/b", name="Alice")
+        added = client.add_faces(
+            "team a", "alice/b", [b"face"], review_mode="standard"
+        )
+        client.list_faces("team a", "alice/b", limit=3)
+        client.delete_face("team a", "alice/b", "face/1")
+        client.delete_person("team a", "alice/b")
+        search = client.search("team a", b"query", limit=2, threshold=0.68)
+
+    assert isinstance(search, SearchResult)
+    assert isinstance(added, FaceRegistrationResult)
+    assert added.faces == [] and added.rejected_images == []
+    assert search.matches == []
+    assert search.threshold == 0.68
+    assert dict(requests[0].url.params) == {"limit": "7", "search": "alice"}
+    assert requests[1].url.raw_path == b"/v1/collections/team%20a/persons/alice%2Fb"
+    assert requests[5].url.raw_path.endswith(b"/faces/face%2F1")
+    assert b'name="review_mode"' in requests[3].content
+    assert b"\r\nstandard\r\n" in requests[3].content
+    assert b'name="limit"' in requests[7].content
+    assert b'name="face_selection"' not in requests[7].content
+
+
+def test_empty_images_and_empty_patch_fail_before_http() -> None:
+    def handler(_: httpx.Request) -> httpx.Response:
+        raise AssertionError("HTTP must not be called")
+
+    with client_for(handler) as client:
+        with pytest.raises(ValueError, match="at least one image"):
+            client.create_person("employees", images=[])
+        with pytest.raises(ValueError, match="at least one field"):
+            client.update_person("employees", "alice")
+        with pytest.raises(ValueError, match="must not be empty"):
+            client.detect(b"")
+        with pytest.raises(TypeError, match="path, bytes"):
+            client.detect(object())  # type: ignore[arg-type]
+
+
+@pytest.mark.parametrize(
+    ("status", "exception_type"),
+    [
+        (400, ValidationError),
+        (401, AuthenticationError),
+        (404, NotFoundError),
+        (409, ConflictError),
+        (413, PayloadTooLargeError),
+        (422, ValidationError),
+        (429, RateLimitError),
+        (500, ServerError),
+        (503, ServiceUnavailableError),
+    ],
+)
+def test_api_errors_are_typed(
+    status: int, exception_type: type[Exception]
+) -> None:
+    def handler(_: httpx.Request) -> httpx.Response:
+        return response(
+            status,
+            {
+                "error": {
+                    "code": "face_not_found",
+                    "message": "No usable face was detected.",
+                    "details": {"image": "source"},
+                },
+                "request_id": "body-request",
+            },
+            request_id="header-request",
+        )
+
+    with client_for(handler) as client:
+        with pytest.raises(exception_type) as captured:
+            client.detect(b"image")
+
+    error = captured.value
+    assert error.code == "face_not_found"
+    assert error.status_code == status
+    assert error.request_id == "header-request"
+    assert error.details == {"image": "source"}
+    assert "header-request" in str(error)
+
+
+def test_transport_and_invalid_success_response_are_safe() -> None:
+    def unavailable(request: httpx.Request) -> httpx.Response:
+        raise httpx.ConnectError("secret low-level reason", request=request)
+
+    with client_for(unavailable) as client:
+        with pytest.raises(TransportError, match="Unable to complete") as captured:
+            client.health()
+    assert "secret low-level reason" not in str(captured.value)
+
+    def invalid(_: httpx.Request) -> httpx.Response:
+        return httpx.Response(
+            200,
+            headers={"x-request-id": "bad-json"},
+            content=b"not-json",
+        )
+
+    with client_for(invalid) as client:
+        with pytest.raises(ServerError) as invalid_response:
+            client.health()
+    assert invalid_response.value.code == "invalid_response"
+    assert invalid_response.value.request_id == "bad-json"
diff --git a/server/tests/unit/test_config_search.py b/server/tests/unit/test_config_search.py
new file mode 100644
index 000000000..42c29ee1d
--- /dev/null
+++ b/server/tests/unit/test_config_search.py
@@ -0,0 +1,83 @@
+from __future__ import annotations
+
+import pytest
+from insightface_server.config import Settings
+
+SEARCH_ENVIRONMENT = (
+    "INSIGHTFACE_DEFAULT_THRESHOLD",
+    "INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE",
+    "INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS",
+    "INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS",
+    "INSIGHTFACE_COLLECTION_DEFAULT_MAX_FACES_PER_PERSON",
+    "INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY",
+)
+
+SUPPORTED_PROFILES = (
+    "fp32_v1",
+    "fp16_v1",
+    "bf16_v1",
+    "int8_x1000_v1",
+    "int8_x736_v1",
+)
+
+
+def test_search_collection_defaults(monkeypatch: pytest.MonkeyPatch) -> None:
+    for name in SEARCH_ENVIRONMENT:
+        monkeypatch.delenv(name, raising=False)
+
+    settings = Settings.from_env()
+
+    assert settings.default_threshold == 0.4
+    assert settings.default_search_profile == "fp32_v1"
+    assert settings.default_search_capacity_rows == 100_000
+    assert settings.max_search_capacity_rows == 10_000_000
+    assert settings.default_max_faces_per_person == 20
+    assert settings.default_search_load_policy == "lazy"
+
+
+def test_search_collection_environment_is_validated(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv(
+        "INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE", "int8_x1000_rerank_v1"
+    )
+    with pytest.raises(ValueError, match="DEFAULT_SEARCH_PROFILE"):
+        Settings.from_env()
+
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE", "bf16_v1")
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_DEFAULT_LOAD_POLICY", "warm")
+    with pytest.raises(ValueError, match="DEFAULT_LOAD_POLICY"):
+        Settings.from_env()
+
+
+@pytest.mark.parametrize("profile", SUPPORTED_PROFILES)
+def test_all_public_search_profiles_are_valid_configuration(
+    monkeypatch: pytest.MonkeyPatch, profile: str
+) -> None:
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_DEFAULT_SEARCH_PROFILE", profile)
+
+    assert Settings.from_env().default_search_profile == profile
+
+
+def test_threshold_must_be_nonnegative_raw_cosine(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("INSIGHTFACE_DEFAULT_THRESHOLD", "-0.01")
+    with pytest.raises(ValueError, match="INSIGHTFACE_DEFAULT_THRESHOLD"):
+        Settings.from_env()
+
+
+def test_deployment_capacity_limit_is_configurable_and_bounds_default(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS", "125")
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_MAX_CAPACITY_ROWS", "250")
+
+    settings = Settings.from_env()
+
+    assert settings.default_search_capacity_rows == 125
+    assert settings.max_search_capacity_rows == 250
+
+    monkeypatch.setenv("INSIGHTFACE_COLLECTION_DEFAULT_CAPACITY_ROWS", "251")
+    with pytest.raises(ValueError, match="DEFAULT_CAPACITY_ROWS must not exceed"):
+        Settings.from_env()
diff --git a/server/tests/unit/test_detection_config.py b/server/tests/unit/test_detection_config.py
new file mode 100644
index 000000000..ff5d59c63
--- /dev/null
+++ b/server/tests/unit/test_detection_config.py
@@ -0,0 +1,194 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+from insightface_server.config import (
+    DEFAULT_CPU_INFERENCE_MAX_CONCURRENCY,
+    DEFAULT_CUDA_INFERENCE_MAX_CONCURRENCY,
+    DEFAULT_DETECTOR_INPUT_SIZES,
+    DEFAULT_DETECTOR_NMS_THRESHOLD,
+    DEFAULT_DETECTOR_THRESHOLD,
+    DEFAULT_MAX_DETECTED_FACES,
+    DetectionProfile,
+    ServerFileConfig,
+    Settings,
+    load_detector_input_sizes,
+    load_server_config,
+    normalize_detector_input_sizes,
+)
+
+
+def test_detector_input_sizes_default_and_custom_toml(tmp_path: Path) -> None:
+    assert load_detector_input_sizes(None) == ((96, 96), (512, 512))
+    assert DEFAULT_DETECTOR_INPUT_SIZES == ((96, 96), (512, 512))
+
+    config = tmp_path / "server.toml"
+    config.write_text(
+        "[detection]\ninput_sizes = [[128, 128], [640, 640]]\n",
+        encoding="utf-8",
+    )
+
+    assert load_detector_input_sizes(config) == ((128, 128), (640, 640))
+
+
+def test_web_ui_defaults_enabled_and_can_be_disabled_in_toml(tmp_path: Path) -> None:
+    assert load_server_config(None) == ServerFileConfig()
+
+    config = tmp_path / "server.toml"
+    config.write_text(
+        "[detection]\n"
+        "input_sizes = [[96, 96]]\n"
+        "threshold = 0.42\n"
+        "nms_threshold = 0.35\n"
+        'single_face_selection = "center_largest"\n'
+        "max_detected_faces = 25\n"
+        "\n[web]\ndisabled = true\n",
+        encoding="utf-8",
+    )
+
+    assert load_server_config(config) == ServerFileConfig(
+        detection=DetectionProfile(
+            input_sizes=((96, 96),),
+            threshold=0.42,
+            nms_threshold=0.35,
+            single_face_selection="center_largest",
+        ),
+        max_detected_faces=25,
+        web_ui_disabled=True,
+    )
+
+
+def test_settings_loads_detector_sizes_once_from_selected_file(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    config = tmp_path / "server.toml"
+    config.write_text(
+        "[detection]\ninput_sizes = [[96, 96], [512, 512]]\n"
+        "\n[web]\ndisabled = true\n",
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("INSIGHTFACE_CONFIG_FILE", str(config))
+
+    settings = Settings.from_env()
+
+    assert settings.config_file == config
+    assert settings.detector_input_sizes == ((96, 96), (512, 512))
+    assert settings.detector_threshold == DEFAULT_DETECTOR_THRESHOLD
+    assert settings.detector_nms_threshold == DEFAULT_DETECTOR_NMS_THRESHOLD
+    assert settings.max_detected_faces == DEFAULT_MAX_DETECTED_FACES
+    assert settings.web_ui_disabled is True
+    assert settings.inference_max_concurrency == DEFAULT_CPU_INFERENCE_MAX_CONCURRENCY
+
+    config.write_text(
+        "[detection]\ninput_sizes = [[128, 128]]\n",
+        encoding="utf-8",
+    )
+    assert settings.detector_input_sizes == ((96, 96), (512, 512))
+
+
+def test_inference_concurrency_defaults_by_provider_and_supports_startup_override(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.delenv("INSIGHTFACE_CONFIG_FILE", raising=False)
+    monkeypatch.delenv("INSIGHTFACE_INFERENCE_MAX_CONCURRENCY", raising=False)
+    monkeypatch.setenv("INSIGHTFACE_EXECUTION_PROVIDER", "CPUExecutionProvider")
+    assert (
+        Settings.from_env().inference_max_concurrency
+        == DEFAULT_CPU_INFERENCE_MAX_CONCURRENCY
+    )
+
+    monkeypatch.setenv("INSIGHTFACE_EXECUTION_PROVIDER", "CUDAExecutionProvider")
+    assert (
+        Settings.from_env().inference_max_concurrency
+        == DEFAULT_CUDA_INFERENCE_MAX_CONCURRENCY
+    )
+
+    config = tmp_path / "server.toml"
+    config.write_text("[inference]\nmax_concurrency = 11\n", encoding="utf-8")
+    monkeypatch.setenv("INSIGHTFACE_CONFIG_FILE", str(config))
+    assert Settings.from_env().inference_max_concurrency == 11
+
+    monkeypatch.setenv("INSIGHTFACE_INFERENCE_MAX_CONCURRENCY", "3")
+    assert Settings.from_env().inference_max_concurrency == 3
+
+    monkeypatch.setenv("INSIGHTFACE_INFERENCE_MAX_CONCURRENCY", "auto")
+    assert (
+        Settings.from_env().inference_max_concurrency
+        == DEFAULT_CUDA_INFERENCE_MAX_CONCURRENCY
+    )
+
+    monkeypatch.setenv("INSIGHTFACE_INFERENCE_MAX_CONCURRENCY", "many")
+    with pytest.raises(ValueError, match="INSIGHTFACE_INFERENCE_MAX_CONCURRENCY"):
+        Settings.from_env()
+
+
+@pytest.mark.parametrize(
+    "document",
+    [
+        "[inference]\nmax_concurrency = 0\n",
+        "[inference]\nmax_concurrency = 257\n",
+        "[inference]\nmax_concurrency = true\n",
+        "[inference]\nworkers = 4\n",
+    ],
+)
+def test_inference_concurrency_file_validation(
+    tmp_path: Path, document: str
+) -> None:
+    config = tmp_path / "server.toml"
+    config.write_text(document, encoding="utf-8")
+    with pytest.raises(ValueError, match="inference"):
+        load_server_config(config)
+
+
+def test_detection_profile_defaults_and_selection_validation() -> None:
+    assert DetectionProfile().as_dict() == {
+        "input_sizes": [[96, 96], [512, 512]],
+        "threshold": 0.5,
+        "nms_threshold": 0.4,
+        "single_face_selection": "largest",
+    }
+    with pytest.raises(ValueError, match="single_face_selection"):
+        DetectionProfile.from_mapping({"single_face_selection": "nearest"})
+
+
+@pytest.mark.parametrize(
+    ("value", "message"),
+    [
+        ([], "non-empty"),
+        ([[100, 96]], "multiples of 32"),
+        ([[96, 96], [96, 96]], "duplicates"),
+        ([[True, 96]], "integers"),
+        ([[2048, 2048], [32, 32]], "combined detector input pixels"),
+    ],
+)
+def test_detector_input_size_validation(value: object, message: str) -> None:
+    with pytest.raises(ValueError, match=message):
+        normalize_detector_input_sizes(value)
+
+
+def test_detector_config_rejects_missing_invalid_and_unknown_files(
+    tmp_path: Path,
+) -> None:
+    with pytest.raises(ValueError, match="Unable to read"):
+        load_detector_input_sizes(tmp_path / "missing.toml")
+
+    invalid = tmp_path / "invalid.toml"
+    invalid.write_text("[detection\n", encoding="utf-8")
+    with pytest.raises(ValueError, match="not valid TOML"):
+        load_detector_input_sizes(invalid)
+
+    unknown = tmp_path / "unknown.toml"
+    unknown.write_text("[detection]\nresolution = 512\n", encoding="utf-8")
+    with pytest.raises(ValueError, match=r"Unsupported \[detection\] setting"):
+        load_detector_input_sizes(unknown)
+
+    invalid_web = tmp_path / "invalid-web.toml"
+    invalid_web.write_text("[web]\ndisabled = 'no'\n", encoding="utf-8")
+    with pytest.raises(ValueError, match=r"\[web\]\.disabled must be a boolean"):
+        load_server_config(invalid_web)
+
+    unknown_web = tmp_path / "unknown-web.toml"
+    unknown_web.write_text("[web]\nenabled = true\n", encoding="utf-8")
+    with pytest.raises(ValueError, match=r"Unsupported \[web\] setting"):
+        load_server_config(unknown_web)
diff --git a/server/tests/unit/test_inference_core.py b/server/tests/unit/test_inference_core.py
new file mode 100644
index 000000000..3cebe1a56
--- /dev/null
+++ b/server/tests/unit/test_inference_core.py
@@ -0,0 +1,255 @@
+from __future__ import annotations
+
+import json
+import threading
+import time
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+from insightface_server.inference import (
+    EngineSummary,
+    FaceObservation,
+    InferenceEngine,
+    cosine_similarity,
+    create_engine,
+    l2_normalize,
+    raw_cosine_similarity,
+)
+from insightface_server.inference.concurrency import InferenceConcurrencyLimiter
+from insightface_server.inference.mock import MockInferenceEngine
+from insightface_server.inference.quality import (
+    RegistrationQualityPolicy,
+    enrich_quality,
+    registration_rejection_reasons,
+)
+from insightface_server.services.core import FaceService
+
+
+def test_l2_normalization_and_raw_cosine_range() -> None:
+    vector = l2_normalize(np.array([3.0, 4.0], dtype=np.float32))
+    assert np.linalg.norm(vector) == pytest.approx(1.0)
+    assert raw_cosine_similarity(vector, vector) == pytest.approx(1.0)
+    assert cosine_similarity(vector, vector) == pytest.approx(1.0)
+    assert cosine_similarity(vector, -vector) == pytest.approx(-1.0)
+    assert cosine_similarity(np.array([1, 0]), np.array([0, 1])) == pytest.approx(0.0)
+
+
+def test_embedding_validation_rejects_invalid_values() -> None:
+    with pytest.raises(ValueError, match="zero or non-finite"):
+        l2_normalize(np.zeros(4, dtype=np.float32))
+    with pytest.raises(ValueError, match="dimensions"):
+        cosine_similarity(np.ones(2), np.ones(3))
+
+
+def test_external_trusted_embedding_validation_is_strict_and_canonical() -> None:
+    accepted = FaceService._trusted_embedding([1.0001, 0.0, 0.0], 3)
+    assert accepted is not None
+    assert accepted.dtype == np.float32
+    np.testing.assert_array_equal(accepted, np.array([1.0, 0.0, 0.0], np.float32))
+
+    assert FaceService._trusted_embedding([1.001, 0.0, 0.0], 3) is None
+    assert FaceService._trusted_embedding([0.0, 0.0, 0.0], 3) is None
+    assert FaceService._trusted_embedding([1.0, 0.0], 3) is None
+    assert FaceService._trusted_embedding([True, 0.0, 0.0], 3) is None
+    assert FaceService._trusted_embedding([float("nan"), 0.0, 0.0], 3) is None
+
+
+def test_face_observation_bbox_helpers() -> None:
+    face = FaceObservation(
+        bbox=(10, 20, 50, 80),
+        detection_score=0.9,
+        landmarks=None,
+        embedding=None,
+    )
+    assert face.area == 2400
+    assert face.confidence == 0.9
+    assert face.bbox_xywh == (10, 20, 40, 60)
+    assert face.normalized_bbox(100, 200) == pytest.approx((0.1, 0.1, 0.4, 0.3))
+    clipped = FaceObservation((-10, -20, 50, 80), 0.9, None, None)
+    assert clipped.normalized_bbox(100, 200) == pytest.approx((0, 0, 0.5, 0.4))
+
+
+def test_quality_metrics_and_registration_reasons_are_normalized() -> None:
+    image = np.zeros((100, 100, 3), dtype=np.uint8)
+    image[20:80, 20:80] = np.indices((60, 60)).sum(axis=0)[..., None] % 2 * 255
+    landmarks = np.array([[35, 40], [65, 40], [50, 52], [39, 66], [61, 66]], dtype=np.float32)
+    face = FaceObservation((20, 20, 80, 80), 0.99, landmarks, None)
+    enrich_quality(image, face)
+    assert 0.0 <= face.brightness <= 1.0
+    assert 0.0 <= face.sharpness <= 1.0
+    assert 0.0 <= face.quality_score <= 1.0
+    assert registration_rejection_reasons(face) == []
+
+    weak = FaceObservation((0, 0, 12, 12), 0.1, landmarks, None)
+    enrich_quality(np.zeros((20, 20, 3), dtype=np.uint8), weak)
+    reasons = registration_rejection_reasons(weak, RegistrationQualityPolicy(min_quality_score=0.9))
+    assert reasons == ["face_too_small", "low_detection_score", "low_quality"]
+
+
+def test_mock_engine_is_deterministic_sorted_and_l2_normalized() -> None:
+    engine = MockInferenceEngine(embedding_dimension=32)
+    assert isinstance(engine, InferenceEngine)
+    engine.startup()
+    generator = np.random.default_rng(42)
+    image = generator.integers(0, 256, (80, 180, 3), dtype=np.uint8)
+    first = engine.analyze(image)
+    second = engine.analyze(image)
+    assert len(first) == 2
+    assert first[0].area > first[1].area
+    assert first[0].embedding is not None
+    assert np.linalg.norm(first[0].embedding) == pytest.approx(1.0)
+    np.testing.assert_array_equal(first[0].embedding, second[0].embedding)
+    assert all(0.0 <= face.detection_score <= 1.0 for face in first)
+    assert len(engine.analyze(image, max_faces=1)) == 1
+    assert engine.analyze(image, min_score=1.0) == []
+
+
+def test_mock_engine_blank_image_and_lifecycle() -> None:
+    engine = MockInferenceEngine()
+    with pytest.raises(RuntimeError, match="not been started"):
+        engine.analyze(np.zeros((64, 64, 3), dtype=np.uint8))
+    engine.startup()
+    assert engine.analyze(np.zeros((64, 64, 3), dtype=np.uint8)) == []
+    assert engine.runtime_summary()["mode"] == "mock"
+    assert engine.runtime_summary()["cpu"]["logical_cores"]
+    engine.close()
+    with pytest.raises(RuntimeError, match="closed"):
+        engine.startup()
+
+
+def test_global_inference_limiter_allows_parallel_work_and_bounds_active_count() -> None:
+    limiter = InferenceConcurrencyLimiter(2)
+    release = threading.Event()
+    two_active = threading.Event()
+    state_lock = threading.Lock()
+    active = 0
+    observed_peak = 0
+
+    def work() -> None:
+        nonlocal active, observed_peak
+        with limiter.slot():
+            with state_lock:
+                active += 1
+                observed_peak = max(observed_peak, active)
+                if active == 2:
+                    two_active.set()
+            assert release.wait(timeout=5)
+            with state_lock:
+                active -= 1
+
+    with ThreadPoolExecutor(max_workers=4) as pool:
+        futures = [pool.submit(work) for _ in range(4)]
+        assert two_active.wait(timeout=5)
+        deadline = time.monotonic() + 5
+        while limiter.summary()["waiting"] < 2 and time.monotonic() < deadline:
+            time.sleep(0.01)
+        assert limiter.summary()["active"] == 2
+        assert limiter.summary()["waiting"] == 2
+        release.set()
+        for future in futures:
+            future.result(timeout=5)
+
+    assert observed_peak == 2
+    assert limiter.summary() == {
+        "max_concurrency": 2,
+        "active": 0,
+        "waiting": 0,
+        "peak_active": 2,
+    }
+
+
+def test_factory_and_summary_use_simple_public_contract() -> None:
+    engine = create_engine(
+        SimpleNamespace(
+            inference_mode="mock",
+            execution_provider="CPUExecutionProvider",
+            embedding_dimension=16,
+        )
+    )
+    assert engine.summary.embedding_dimension == 16
+    summary = EngineSummary("id", "1", "a" * 64, 16, "1", "CPUExecutionProvider")
+    assert summary.as_dict()["model_id"] == "id"
+    with pytest.raises(ValueError, match="Unsupported inference_mode"):
+        create_engine({"inference_mode": "remote"})
+
+
+def _write_manifest(directory: Path) -> None:
+    detector = directory / "detector.onnx"
+    recognizer = directory / "recognizer.onnx"
+    detector.write_bytes(b"detector")
+    recognizer.write_bytes(b"recognizer")
+    manifest = {
+        "manifest_version": 1,
+        "model_id": "buffalo_l",
+        "model_version": "v0.7",
+        "files": {"detector": detector.name, "recognizer": recognizer.name},
+        "recognition": {
+            "input_size": [112, 112],
+            "embedding_dimension": 512,
+            "preprocessing": "insightface-arcface-1",
+        },
+        "license": "MODEL.LICENSE",
+    }
+    (directory / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    default_license = (
+        Path(__file__).resolve().parents[2]
+        / "backend"
+        / "insightface_server"
+        / "licensing"
+        / "defaults"
+        / "buffalo_l"
+        / "MODEL.LICENSE"
+    )
+    (directory / "MODEL.LICENSE").write_bytes(default_license.read_bytes())
+
+
+def test_onnx_engine_startup_is_once_per_process(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    from insightface_server.inference.onnx_engine import OnnxInsightFaceEngine
+
+    _write_manifest(tmp_path)
+    engine = OnnxInsightFaceEngine(
+        SimpleNamespace(
+            models_dir=tmp_path,
+            execution_provider="CPUExecutionProvider",
+            detector_threshold=0.5,
+            device_id=0,
+        )
+    )
+    assert len(engine.summary.model_digest) == 64
+    assert engine.summary.model_digest != engine.bundle.recognizer.sha256
+    calls = 0
+
+    def fake_startup() -> None:
+        nonlocal calls
+        calls += 1
+
+    monkeypatch.setattr(engine, "_startup_once", fake_startup)
+    engine.startup()
+    engine.startup()
+    assert calls == 1
+
+
+def test_onnx_engine_refuses_tampered_model_license(tmp_path: Path) -> None:
+    from insightface_server.inference.onnx_engine import OnnxInsightFaceEngine
+
+    _write_manifest(tmp_path)
+    license_path = tmp_path / "MODEL.LICENSE"
+    document = json.loads(license_path.read_text(encoding="utf-8"))
+    document["grant"] = "commercial"
+    document["customer"] = "Tampered Customer"
+    license_path.write_text(json.dumps(document), encoding="utf-8")
+
+    with pytest.raises(RuntimeError, match="signature verification failed"):
+        OnnxInsightFaceEngine(
+            SimpleNamespace(
+                models_dir=tmp_path,
+                execution_provider="CPUExecutionProvider",
+                detector_threshold=0.5,
+                device_id=0,
+            )
+        )
diff --git a/server/tests/unit/test_inference_cuda_audit.py b/server/tests/unit/test_inference_cuda_audit.py
new file mode 100644
index 000000000..828b2b0a4
--- /dev/null
+++ b/server/tests/unit/test_inference_cuda_audit.py
@@ -0,0 +1,96 @@
+from __future__ import annotations
+
+import pytest
+from insightface_server.inference.onnx_engine import (
+    audit_cuda_profile,
+    validate_gpu_requirements,
+)
+
+
+def _event(provider: str, operator: str, **extra: object) -> dict[str, object]:
+    args: dict[str, object] = {"provider": provider, "op_name": operator}
+    args.update(extra)
+    return {"cat": "Node", "name": f"{operator}_kernel", "args": args}
+
+
+def test_cuda_profile_allows_only_small_integer_shape_metadata() -> None:
+    profile = [
+        _event("CUDAExecutionProvider", "Conv"),
+        _event(
+            "CPUExecutionProvider",
+            "Gather",
+            output_size="8",
+            output_type_shape=[{"int64": [1]}],
+        ),
+    ]
+    result = audit_cuda_profile(profile, model_name="detector")
+    assert result["accepted"] is True
+    assert result["cuda_kernel_count"] == 1
+    assert result["cpu_shape_kernel_count"] == 1
+
+
+@pytest.mark.parametrize(
+    "cpu_event",
+    [
+        _event(
+            "CPUExecutionProvider",
+            "Conv",
+            output_size="1024",
+            output_type_shape=[{"float": [1, 16]}],
+        ),
+        _event(
+            "CPUExecutionProvider",
+            "Gather",
+            output_size="8192",
+            output_type_shape=[{"int64": [1024]}],
+        ),
+    ],
+)
+def test_cuda_profile_rejects_compute_or_large_cpu_fallback(
+    cpu_event: dict[str, object],
+) -> None:
+    with pytest.raises(RuntimeError, match="rejected"):
+        audit_cuda_profile(
+            [_event("CUDAExecutionProvider", "Conv"), cpu_event],
+            model_name="detector",
+        )
+
+
+def test_cuda_profile_requires_actual_cuda_kernel() -> None:
+    with pytest.raises(RuntimeError, match="no CUDA kernels"):
+        audit_cuda_profile([], model_name="recognizer")
+
+
+def test_gpu_driver_and_blackwell_runtime_requirements() -> None:
+    validate_gpu_requirements(
+        [
+            {
+                "name": "RTX 5090",
+                "compute_capability": "12.0",
+                "driver_version": "580.105.08",
+            }
+        ],
+        "12.9",
+    )
+    with pytest.raises(RuntimeError, match="570.26"):
+        validate_gpu_requirements(
+            [
+                {
+                    "name": "RTX 5090",
+                    "compute_capability": "12.0",
+                    "driver_version": "565.0",
+                }
+            ],
+            "12.9",
+        )
+    with pytest.raises(RuntimeError, match="CUDA Runtime 12.8"):
+        validate_gpu_requirements(
+            [
+                {
+                    "name": "RTX 5090",
+                    "compute_capability": "12.0",
+                    "driver_version": "580.0",
+                }
+            ],
+            "12.7",
+        )
diff --git a/server/tests/unit/test_inference_manifest.py b/server/tests/unit/test_inference_manifest.py
new file mode 100644
index 000000000..de33a52f7
--- /dev/null
+++ b/server/tests/unit/test_inference_manifest.py
@@ -0,0 +1,125 @@
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+
+import pytest
+from insightface_server.models import load_manifest
+
+
+def _models(tmp_path: Path) -> tuple[Path, Path]:
+    detector = tmp_path / "detector.onnx"
+    recognizer = tmp_path / "recognizer.onnx"
+    detector.write_bytes(b"detector")
+    recognizer.write_bytes(b"recognizer")
+    (tmp_path / "MODEL.LICENSE").write_text("{}", encoding="utf-8")
+    return detector, recognizer
+
+
+def _manifest(detector: Path, recognizer: Path) -> dict[str, object]:
+    return {
+        "manifest_version": 1,
+        "model_id": "test_model",
+        "model_version": "v1",
+        "display_name": "Test model",
+        "files": {
+            "detector": detector.name,
+            "recognizer": recognizer.name,
+        },
+        "recognition": {
+            "input_size": [112, 112],
+            "embedding_dimension": 512,
+            "preprocessing": "arcface-v1",
+        },
+        "license": "MODEL.LICENSE",
+    }
+
+
+def test_loads_compact_manifest_and_calculates_diagnostic_digest(tmp_path: Path) -> None:
+    detector, recognizer = _models(tmp_path)
+    manifest = _manifest(detector, recognizer)
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+
+    bundle = load_manifest(tmp_path)
+
+    assert bundle.model_id == "test_model"
+    assert bundle.model_version == "v1"
+    assert bundle.detector.task == "face_detection"
+    assert bundle.recognizer.embedding_dimension == 512
+    assert bundle.recognizer.public_summary()["model_id"] == "test_model"
+    assert bundle.recognizer.sha256 == hashlib.sha256(b"recognizer").hexdigest()
+    assert bundle.license_path.name == "MODEL.LICENSE"
+
+
+def test_converted_model_content_does_not_require_manifest_hash_update(tmp_path: Path) -> None:
+    detector, recognizer = _models(tmp_path)
+    manifest = _manifest(detector, recognizer)
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    original = load_manifest(tmp_path).recognizer.sha256
+
+    recognizer.write_bytes(b"converted-fp16-recognizer")
+    converted = load_manifest(tmp_path)
+
+    assert converted.model_id == "test_model"
+    assert converted.recognizer.sha256 != original
+
+
+def test_rejects_path_escape_and_wrong_license_filename(tmp_path: Path) -> None:
+    detector, recognizer = _models(tmp_path)
+    manifest = _manifest(detector, recognizer)
+    files = manifest["files"]
+    assert isinstance(files, dict)
+    files["detector"] = "../escaped.onnx"
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    with pytest.raises(RuntimeError, match="Unsafe .onnx path"):
+        load_manifest(tmp_path)
+
+    manifest = _manifest(detector, recognizer)
+    manifest["license"] = "license.json"
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    with pytest.raises(RuntimeError, match="MODEL.LICENSE"):
+        load_manifest(tmp_path)
+
+
+def test_recognition_contract_is_required(tmp_path: Path) -> None:
+    detector, recognizer = _models(tmp_path)
+    manifest = _manifest(detector, recognizer)
+    recognition = manifest["recognition"]
+    assert isinstance(recognition, dict)
+    del recognition["embedding_dimension"]
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    with pytest.raises(RuntimeError, match="recognition must contain exactly"):
+        load_manifest(tmp_path)
+
+
+def test_legacy_manifest_is_read_without_enforcing_declared_sha256(tmp_path: Path) -> None:
+    detector, recognizer = _models(tmp_path)
+    legacy = {
+        "package": {"name": "buffalo_l", "release": "v0.7"},
+        "models": [
+            {
+                "model_id": "scrfd-detection",
+                "model_version": "1",
+                "task": "face_detection",
+                "file": detector.name,
+                "input_size": [640, 640],
+                "preprocessing_version": "insightface-scrfd-1",
+                "sha256": "0" * 64,
+            },
+            {
+                "model_id": "buffalo_l-recognition",
+                "model_version": "1",
+                "task": "face_recognition",
+                "file": recognizer.name,
+                "input_size": [112, 112],
+                "embedding_dimension": 512,
+                "preprocessing_version": "insightface-arcface-1",
+                "sha256": "0" * 64,
+            },
+        ],
+    }
+    (tmp_path / "manifest.json").write_text(json.dumps(legacy), encoding="utf-8")
+    bundle = load_manifest(tmp_path)
+    assert bundle.legacy_manifest is True
+    assert bundle.model_id == "buffalo_l"
diff --git a/server/tests/unit/test_model_license.py b/server/tests/unit/test_model_license.py
new file mode 100644
index 000000000..3eb399004
--- /dev/null
+++ b/server/tests/unit/test_model_license.py
@@ -0,0 +1,164 @@
+from __future__ import annotations
+
+import base64
+import json
+from datetime import UTC, datetime
+from pathlib import Path
+
+import pytest
+from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PrivateKey
+from insightface_server.licensing import (
+    ModelLicenseError,
+    canonical_license_bytes,
+    verify_model_license,
+)
+from insightface_server.licensing.model_license import _trusted_keys
+
+
+def _write_license(
+    path: Path,
+    private_key: Ed25519PrivateKey,
+    **overrides: object,
+) -> None:
+    document: dict[str, object] = {
+        "license_version": 1,
+        "license_id": "test-license-1",
+        "issuer": "InsightFace",
+        "model_id": "buffalo_l",
+        "grant": "non-commercial",
+        "valid_from": "2026-01-01T00:00:00Z",
+        **overrides,
+    }
+    signature = private_key.sign(canonical_license_bytes(document))
+    document["signature"] = base64.urlsafe_b64encode(signature).rstrip(b"=").decode()
+    path.write_text(json.dumps(document, indent=2) + "\n", encoding="utf-8")
+
+
+def test_verifies_signed_model_scoped_license(tmp_path: Path) -> None:
+    private_key = Ed25519PrivateKey.generate()
+    path = tmp_path / "MODEL.LICENSE"
+    _write_license(path, private_key)
+
+    result = verify_model_license(
+        path,
+        expected_model_id="buffalo_l",
+        now=datetime(2026, 7, 22, tzinfo=UTC),
+        public_keys=(private_key.public_key(),),
+    )
+
+    assert result.issuer == "InsightFace"
+    assert result.model_id == "buffalo_l"
+    assert result.grant == "non-commercial"
+    assert result.commercial_use_permitted is False
+    assert result.public_summary()["signature_valid"] is True
+
+
+def test_rejects_tampering_wrong_model_and_expiration(tmp_path: Path) -> None:
+    private_key = Ed25519PrivateKey.generate()
+    path = tmp_path / "MODEL.LICENSE"
+    _write_license(path, private_key)
+    document = json.loads(path.read_text(encoding="utf-8"))
+    document["grant"] = "commercial"
+    document["customer"] = "Tampered Customer"
+    path.write_text(json.dumps(document), encoding="utf-8")
+    with pytest.raises(ModelLicenseError, match="signature verification failed"):
+        verify_model_license(
+            path,
+            expected_model_id="buffalo_l",
+            now=datetime(2026, 7, 22, tzinfo=UTC),
+            public_keys=(private_key.public_key(),),
+        )
+
+    _write_license(path, private_key, model_id="private_recognition")
+    with pytest.raises(ModelLicenseError, match="not the active model"):
+        verify_model_license(
+            path,
+            expected_model_id="buffalo_l",
+            now=datetime(2026, 7, 22, tzinfo=UTC),
+            public_keys=(private_key.public_key(),),
+        )
+
+    _write_license(path, private_key, issuer="Another Issuer")
+    with pytest.raises(ModelLicenseError, match="issuer must be 'InsightFace'"):
+        verify_model_license(
+            path,
+            expected_model_id="buffalo_l",
+            now=datetime(2026, 7, 22, tzinfo=UTC),
+            public_keys=(private_key.public_key(),),
+        )
+
+    _write_license(
+        path,
+        private_key,
+        valid_from="2025-01-01T00:00:00Z",
+        valid_until="2026-01-01T00:00:00Z",
+    )
+    with pytest.raises(ModelLicenseError, match="expired"):
+        verify_model_license(
+            path,
+            expected_model_id="buffalo_l",
+            now=datetime(2026, 7, 22, tzinfo=UTC),
+            public_keys=(private_key.public_key(),),
+        )
+
+
+def test_commercial_license_requires_customer_and_preserves_optional_fields(
+    tmp_path: Path,
+) -> None:
+    private_key = Ed25519PrivateKey.generate()
+    path = tmp_path / "MODEL.LICENSE"
+    _write_license(path, private_key, grant="commercial")
+    with pytest.raises(ModelLicenseError, match="identify the customer"):
+        verify_model_license(
+            path,
+            expected_model_id="buffalo_l",
+            now=datetime(2026, 7, 22, tzinfo=UTC),
+            public_keys=(private_key.public_key(),),
+        )
+
+    _write_license(
+        path,
+        private_key,
+        grant="commercial",
+        customer="Example Customer",
+        reference="CONTRACT-2026-001",
+    )
+    result = verify_model_license(
+        path,
+        expected_model_id="buffalo_l",
+        now=datetime(2026, 7, 22, tzinfo=UTC),
+        public_keys=(private_key.public_key(),),
+    )
+    assert result.commercial_use_permitted is True
+    assert result.customer == "Example Customer"
+    assert result.reference == "CONTRACT-2026-001"
+
+
+def test_bundled_public_license_verifies_with_active_public_key() -> None:
+    license_path = (
+        Path(__file__).resolve().parents[2]
+        / "backend"
+        / "insightface_server"
+        / "licensing"
+        / "defaults"
+        / "buffalo_l"
+        / "MODEL.LICENSE"
+    )
+    result = verify_model_license(license_path, expected_model_id="buffalo_l")
+    assert result.issuer == "InsightFace"
+    assert result.grant == "non-commercial"
+
+
+def test_trusted_keys_ignore_macos_appledouble_files(tmp_path: Path) -> None:
+    bundled_key = (
+        Path(__file__).resolve().parents[2]
+        / "backend"
+        / "insightface_server"
+        / "licensing"
+        / "trusted_keys"
+        / "insightface-model-license-public-ed25519.pem"
+    )
+    (tmp_path / bundled_key.name).write_bytes(bundled_key.read_bytes())
+    (tmp_path / f"._{bundled_key.name}").write_bytes(b"AppleDouble metadata")
+
+    assert len(_trusted_keys(tmp_path)) == 1
diff --git a/server/tests/unit/test_model_packages.py b/server/tests/unit/test_model_packages.py
new file mode 100644
index 000000000..3befb223b
--- /dev/null
+++ b/server/tests/unit/test_model_packages.py
@@ -0,0 +1,242 @@
+from __future__ import annotations
+
+import hashlib
+import json
+import shutil
+import zipfile
+from datetime import UTC, datetime
+from pathlib import Path
+
+import pytest
+from insightface_server.licensing import verify_model_license
+from insightface_server.models import load_manifest
+from insightface_server.models.packages import (
+    DEFAULT_LICENSES_DIR,
+    MODEL_LICENSE_FILENAME,
+    PACKAGES,
+    ModelPackage,
+    ModelPackageError,
+    PackageFile,
+    _verify_model_license,
+    extract_required_models,
+    install_package,
+    license_notice,
+    model_license_document,
+    verify_installed,
+)
+
+
+def _sha256(value: bytes) -> str:
+    return hashlib.sha256(value).hexdigest()
+
+
+def test_catalog_contains_all_supported_packages_with_signed_licenses() -> None:
+    expected = {
+        "buffalo_l": (
+            "80ffe37d8a5940d59a7384c201a2a38d4741f2f3c51eef46ebb28218a7b0ca2f",
+            ("det_10g.onnx", "w600k_r50.onnx"),
+        ),
+        "buffalo_m": (
+            "d98264bd8f2dc75cbc2ddce2a14e636e02bb857b3051c234b737bf3b614edca9",
+            ("det_2.5g.onnx", "w600k_r50.onnx"),
+        ),
+        "buffalo_sc": (
+            "57d31b56b6ffa911c8a73cfc1707c73cab76efe7f13b675a05223bf42de47c72",
+            ("det_500m.onnx", "w600k_mbf.onnx"),
+        ),
+        "antelopev2": (
+            "8e182f14fc6e80b3bfa375b33eb6cff7ee05d8ef7633e738d1c89021dcf0c5c5",
+            ("scrfd_10g_bnkps.onnx", "glintr100.onnx"),
+        ),
+    }
+
+    assert set(PACKAGES) == set(expected)
+    for name, package in PACKAGES.items():
+        archive_sha256, filenames = expected[name]
+        assert package.archive_sha256 == archive_sha256
+        assert tuple(item.filename for item in package.files) == filenames
+        assert tuple(item.task for item in package.files) == (
+            "face_detection",
+            "face_recognition",
+        )
+        assert package.files[1].embedding_dimension == 512
+        license_info = verify_model_license(
+            DEFAULT_LICENSES_DIR / name / MODEL_LICENSE_FILENAME,
+            expected_model_id=name,
+        )
+        assert license_info.issuer == "InsightFace"
+        assert license_info.grant == "non-commercial"
+        assert license_info.valid_from == datetime(2021, 9, 22, tzinfo=UTC)
+        assert license_info.valid_until is None
+
+
+def _package(tmp_path: Path) -> tuple[ModelPackage, Path]:
+    detector = b"synthetic-detector"
+    recognizer = b"synthetic-recognizer"
+    archive_path = tmp_path / "test.zip"
+    with zipfile.ZipFile(archive_path, "w") as archive:
+        archive.writestr("nested/det.onnx", detector)
+        archive.writestr("nested/rec.onnx", recognizer)
+        archive.writestr("ignored.onnx", b"not installed")
+    package = ModelPackage(
+        name="buffalo_l",
+        release="v1",
+        display_name="Synthetic Buffalo_L",
+        url="https://example.invalid/test.zip",
+        archive_sha256=hashlib.sha256(archive_path.read_bytes()).hexdigest(),
+        files=(
+            PackageFile(
+                filename="det.onnx",
+                sha256=_sha256(detector),
+                model_id="test-detector",
+                model_version="1",
+                task="face_detection",
+                input_size=(640, 640),
+                preprocessing_version="test-detector-1",
+                input_mean=127.5,
+                input_std=128.0,
+            ),
+            PackageFile(
+                filename="rec.onnx",
+                sha256=_sha256(recognizer),
+                model_id="test-recognizer",
+                model_version="1",
+                task="face_recognition",
+                input_size=(112, 112),
+                preprocessing_version="test-recognizer-1",
+                input_mean=127.5,
+                input_std=127.5,
+                embedding_dimension=512,
+            ),
+        ),
+    )
+    return package, archive_path
+
+
+def test_install_is_verified_atomic_and_idempotent(tmp_path: Path) -> None:
+    package, archive = _package(tmp_path)
+    models_dir = tmp_path / "models"
+    models_dir.mkdir()
+    downloads = 0
+
+    def downloader(_package: ModelPackage, destination: Path) -> None:
+        nonlocal downloads
+        downloads += 1
+        shutil.copyfile(archive, destination)
+
+    assert install_package(package, models_dir, downloader=downloader) == "installed"
+    bundle = load_manifest(models_dir)
+    assert bundle.detector.sha256 == package.files[0].sha256
+    assert bundle.recognizer.sha256 == package.files[1].sha256
+    assert not (models_dir / "ignored.onnx").exists()
+
+    root_manifest_path = models_dir / "manifest.json"
+    root_manifest = json.loads(root_manifest_path.read_text(encoding="utf-8"))
+    relative_license = Path(root_manifest["license"])
+    license_path = models_dir / relative_license
+    assert license_path.read_text(encoding="utf-8") == model_license_document(package)
+    assert root_manifest["manifest_version"] == 1
+    assert root_manifest["model_id"] == "buffalo_l"
+    assert _verify_model_license(package, models_dir).issuer == "InsightFace"
+
+    # Re-running install upgrades a recognized legacy bundle without downloading again.
+    license_path.unlink()
+    bundle_dir = relative_license.parent
+    legacy_manifest = {
+        "package": {"name": package.name, "release": package.release},
+        "models": [
+            {
+                "model_id": item.model_id,
+                "model_version": item.model_version,
+                "task": item.task,
+                "file": (bundle_dir / item.filename).as_posix(),
+                "input_size": list(item.input_size),
+                "preprocessing_version": item.preprocessing_version,
+                "sha256": item.sha256,
+                **(
+                    {"embedding_dimension": item.embedding_dimension}
+                    if item.embedding_dimension is not None
+                    else {}
+                ),
+            }
+            for item in package.files
+        ],
+    }
+    root_manifest_path.write_text(json.dumps(legacy_manifest), encoding="utf-8")
+    assert install_package(package, models_dir, downloader=downloader) == "already_installed"
+    assert downloads == 1
+    assert license_path.read_text(encoding="utf-8") == model_license_document(package)
+    _verify_model_license(package, models_dir)
+
+    installed, summaries, license_info = verify_installed(models_dir)
+    assert installed == "buffalo_l"
+    assert license_info.issuer == "InsightFace"
+    assert [summary["task"] for summary in summaries] == [
+        "face_detection",
+        "face_recognition",
+    ]
+
+
+def test_extract_rejects_duplicate_required_filename(tmp_path: Path) -> None:
+    package, _archive = _package(tmp_path)
+    duplicate = tmp_path / "duplicate.zip"
+    with zipfile.ZipFile(duplicate, "w") as archive:
+        archive.writestr("a/det.onnx", b"synthetic-detector")
+        archive.writestr("b/det.onnx", b"synthetic-detector")
+        archive.writestr("rec.onnx", b"synthetic-recognizer")
+    output = tmp_path / "output"
+    output.mkdir()
+    with pytest.raises(ModelPackageError, match="exactly one det.onnx"):
+        extract_required_models(package, duplicate, output)
+
+
+def test_extract_rejects_model_digest_mismatch(tmp_path: Path) -> None:
+    package, archive = _package(tmp_path)
+    bad_file = PackageFile(
+        filename=package.files[0].filename,
+        sha256="0" * 64,
+        model_id=package.files[0].model_id,
+        model_version="1",
+        task="face_detection",
+        input_size=(640, 640),
+        preprocessing_version="test",
+        input_mean=127.5,
+        input_std=128.0,
+    )
+    bad_package = ModelPackage(
+        name="bad",
+        release="v1",
+        display_name="Bad",
+        url=package.url,
+        archive_sha256=package.archive_sha256,
+        files=(bad_file, package.files[1]),
+    )
+    output = tmp_path / "output"
+    output.mkdir()
+    with pytest.raises(ModelPackageError, match="SHA-256 mismatch"):
+        extract_required_models(bad_package, archive, output)
+
+
+def test_license_notice_is_explicit(tmp_path: Path) -> None:
+    package, _archive = _package(tmp_path)
+    notice = license_notice(package)
+    assert "non-commercial research use only" in notice
+    assert "Commercial use requires a separate license" in notice
+    assert "https://www.insightface.ai" in notice
+
+
+def test_verify_rejects_missing_model_license(tmp_path: Path) -> None:
+    package, archive = _package(tmp_path)
+    models_dir = tmp_path / "models"
+    models_dir.mkdir()
+
+    def downloader(_package: ModelPackage, destination: Path) -> None:
+        shutil.copyfile(archive, destination)
+
+    install_package(package, models_dir, downloader=downloader)
+    manifest = json.loads((models_dir / "manifest.json").read_text(encoding="utf-8"))
+    license_path = models_dir / manifest["license"]
+    assert license_path.name == MODEL_LICENSE_FILENAME
+    license_path.unlink()
+    with pytest.raises(RuntimeError, match="license file is missing"):
+        _verify_model_license(package, models_dir)
diff --git a/server/tests/unit/test_models_cli.py b/server/tests/unit/test_models_cli.py
new file mode 100644
index 000000000..8422f8748
--- /dev/null
+++ b/server/tests/unit/test_models_cli.py
@@ -0,0 +1,113 @@
+from __future__ import annotations
+
+import io
+from datetime import UTC, datetime
+from pathlib import Path
+
+import pytest
+from insightface_server import models_cli
+from insightface_server.licensing import ModelLicense
+
+
+class _NonInteractiveInput(io.StringIO):
+    def isatty(self) -> bool:
+        return False
+
+
+def test_install_requires_explicit_noninteractive_acceptance(
+    tmp_path: Path, monkeypatch, capsys
+) -> None:
+    monkeypatch.setattr(models_cli.sys, "stdin", _NonInteractiveInput())
+    result = models_cli.main(
+        ["--models-dir", str(tmp_path), "install", "buffalo_l"]
+    )
+    captured = capsys.readouterr()
+    assert result == 2
+    assert "non-commercial research use only" in captured.out
+    assert "add --accept-license" in captured.err
+
+
+def test_successful_install_prints_license_after_result(
+    tmp_path: Path, monkeypatch, capsys
+) -> None:
+    monkeypatch.setattr(models_cli, "install_package", lambda *_args, **_kwargs: "installed")
+    result = models_cli.main(
+        [
+            "--models-dir",
+            str(tmp_path),
+            "install",
+            "buffalo_l",
+            "--accept-license",
+        ]
+    )
+    captured = capsys.readouterr()
+    assert result == 0
+    assert "is installed and verified" in captured.out
+    assert captured.out.rfind("LICENSE NOTICE") > captured.out.find("is installed and verified")
+    assert "Commercial use requires a separate license" in captured.out
+    assert "https://www.insightface.ai" in captured.out
+
+
+def test_info_reports_pinned_source_and_hashes(capsys) -> None:
+    assert models_cli.main(["info", "buffalo_l"]) == 0
+    output = capsys.readouterr().out
+    assert "/releases/download/v0.7/buffalo_l.zip" in output
+    assert "80ffe37d8a5940d59a7384c201a2a38d4741f2f3c51eef46ebb28218a7b0ca2f" in output
+    assert "det_10g.onnx SHA-256" in output
+    assert "w600k_r50.onnx SHA-256" in output
+
+
+def test_list_reports_every_supported_package(tmp_path: Path, capsys) -> None:
+    assert models_cli.main(["--models-dir", str(tmp_path), "list"]) == 0
+    output = capsys.readouterr().out
+    for name in ("buffalo_l", "buffalo_m", "buffalo_sc", "antelopev2"):
+        assert f"{name}\tv0.7\tnot installed" in output
+
+
+@pytest.mark.parametrize(
+    ("name", "detector", "recognizer"),
+    (
+        ("buffalo_m", "det_2.5g.onnx", "w600k_r50.onnx"),
+        ("buffalo_sc", "det_500m.onnx", "w600k_mbf.onnx"),
+        ("antelopev2", "scrfd_10g_bnkps.onnx", "glintr100.onnx"),
+    ),
+)
+def test_info_reports_new_catalog_packages(
+    name: str, detector: str, recognizer: str, capsys
+) -> None:
+    assert models_cli.main(["info", name]) == 0
+    output = capsys.readouterr().out
+    assert f"/releases/download/v0.7/{name}.zip" in output
+    assert f"{detector} SHA-256" in output
+    assert f"{recognizer} SHA-256" in output
+    assert "Commercial use requires a separate license" in output
+
+
+def test_verify_prints_installed_package_license(monkeypatch, capsys) -> None:
+    monkeypatch.setattr(
+        models_cli,
+        "verify_installed",
+        lambda _models_dir: (
+            "buffalo_l",
+            (
+                {"file": "det_10g.onnx", "sha256": "a" * 64},
+                {"file": "w600k_r50.onnx", "sha256": "b" * 64},
+            ),
+            ModelLicense(
+                license_id="buffalo_l-public-v1",
+                issuer="InsightFace",
+                model_id="buffalo_l",
+                grant="non-commercial",
+                valid_from=datetime(2026, 7, 22, tzinfo=UTC),
+                valid_until=None,
+            ),
+        ),
+    )
+    assert models_cli.main(["verify", "buffalo_l"]) == 0
+    output = capsys.readouterr().out
+    assert "Installed package: buffalo_l" in output
+    assert "LICENSE VERIFIED" in output
+    assert "Issuer: InsightFace" in output
+    assert "Model ID: buffalo_l" in output
+    assert "Signature: VALID" in output
+    assert "Commercial use: NOT PERMITTED" in output
diff --git a/server/tests/unit/test_multires_scrfd.py b/server/tests/unit/test_multires_scrfd.py
new file mode 100644
index 000000000..4eb3461c7
--- /dev/null
+++ b/server/tests/unit/test_multires_scrfd.py
@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+REPOSITORY_DIR = Path(__file__).resolve().parents[3]
+PYTHON_PACKAGE = REPOSITORY_DIR / "python-package"
+if str(PYTHON_PACKAGE) not in sys.path:
+    sys.path.insert(0, str(PYTHON_PACKAGE))
+
+from insightface.model_zoo.scrfd import SCRFD  # noqa: E402
+
+
+def test_multires_candidates_are_merged_before_one_global_nms() -> None:
+    detector = object.__new__(SCRFD)
+    detector.use_kps = True
+    nms_inputs: list[np.ndarray] = []
+
+    def candidates(_image: np.ndarray, size: tuple[int, int]):
+        offset = 0.0 if size == (96, 96) else 1.0
+        score = 0.90 if size == (96, 96) else 0.80
+        boxes = np.asarray(
+            [[10.0 + offset, 10.0 + offset, 50.0 + offset, 50.0 + offset, score]],
+            dtype=np.float32,
+        )
+        landmarks = np.full((1, 5, 2), offset, dtype=np.float32)
+        return boxes, landmarks
+
+    def global_nms(merged: np.ndarray) -> list[int]:
+        nms_inputs.append(merged.copy())
+        return [0]
+
+    detector._detect_candidates = candidates
+    detector.nms = global_nms
+
+    boxes, landmarks = detector.detect(
+        np.zeros((128, 128, 3), dtype=np.uint8),
+        input_size=[(96, 96), (512, 512)],
+    )
+
+    assert len(nms_inputs) == 1
+    assert nms_inputs[0].shape == (2, 5)
+    assert nms_inputs[0][:, 4].tolist() == pytest.approx([0.90, 0.80])
+    assert boxes.shape == (1, 5)
+    assert landmarks.shape == (1, 5, 2)
diff --git a/server/tests/unit/test_release_preflight.py b/server/tests/unit/test_release_preflight.py
new file mode 100644
index 000000000..d336c181d
--- /dev/null
+++ b/server/tests/unit/test_release_preflight.py
@@ -0,0 +1,195 @@
+from __future__ import annotations
+
+import subprocess
+from pathlib import Path
+
+from scripts.release_preflight import Preflight
+
+
+def write(root: Path, relative: str, content: str) -> None:
+    path = root / relative
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(content, encoding="utf-8")
+
+
+def fixture_repository(tmp_path: Path) -> Path:
+    version = "1.2.3"
+    write(
+        tmp_path,
+        "server/pyproject.toml",
+        f'[project]\nname="server"\nversion="{version}"\n',
+    )
+    write(
+        tmp_path,
+        "server/sdk/python/pyproject.toml",
+        f'[project]\nname="sdk"\nversion="{version}"\nclassifiers=[]\n',
+    )
+    write(
+        tmp_path,
+        "server/backend/insightface_server/__init__.py",
+        f'__version__ = "{version}"\n',
+    )
+    write(
+        tmp_path,
+        "server/sdk/python/src/insightface_server/__init__.py",
+        f'__version__ = "{version}"\n',
+    )
+    write(
+        tmp_path,
+        "server/sdk/python/src/insightface_server/client.py",
+        f'insightface-server-python/{version}\n',
+    )
+    write(
+        tmp_path,
+        "server/backend/insightface_server/models/packages.py",
+        f"InsightFace-Server-Model-Installer/{version}\n",
+    )
+    write(tmp_path, "server/Makefile", f"SERVER_VERSION ?= {version}\n")
+    for name in ("Dockerfile.cpu", "Dockerfile.cuda12"):
+        write(
+            tmp_path,
+            f"server/docker/{name}",
+            "\n".join(
+                (
+                    'LABEL org.opencontainers.image.source="https://github.com/deepinsight/insightface"',
+                    f"ARG INSIGHTFACE_SERVER_VERSION={version}",
+                )
+            )
+            + "\n",
+        )
+    for variant in ("cpu", "cuda12"):
+        image = f"ghcr.io/deepinsight/insightface-server:{version}-{variant}"
+        write(
+            tmp_path,
+            f"server/deploy/compose.{variant}.yml",
+            f"server:\n  image: {image}\nmodels:\n  image: {image}\n",
+        )
+    for name in (
+        "README.md",
+        "README.zh-CN.md",
+        "README.ja.md",
+        "README.de.md",
+        "README.es.md",
+        "README.fr.md",
+        "README.ru.md",
+        "README.pt.md",
+        "README.ko.md",
+    ):
+        write(
+            tmp_path,
+            f"server/{name}",
+            f"{version} {version}-cpu {version}-cuda12\n",
+        )
+    write(
+        tmp_path,
+        "server/docs/openapi.snapshot.json",
+        f'{{"info": {{"version": "{version}"}}}}\n',
+    )
+    subprocess.run(("git", "init"), cwd=tmp_path, check=True, capture_output=True)
+    subprocess.run(
+        ("git", "config", "user.email", "test@example.invalid"),
+        cwd=tmp_path,
+        check=True,
+    )
+    subprocess.run(
+        ("git", "config", "user.name", "Release Test"),
+        cwd=tmp_path,
+        check=True,
+    )
+    subprocess.run(("git", "add", "."), cwd=tmp_path, check=True)
+    subprocess.run(
+        ("git", "commit", "-m", "fixture"),
+        cwd=tmp_path,
+        check=True,
+        capture_output=True,
+    )
+    return tmp_path
+
+
+def statuses(preflight: Preflight) -> dict[str, str]:
+    return {check.name: check.status for check in preflight.run()}
+
+
+def test_formal_preflight_accepts_consistent_clean_release(tmp_path: Path) -> None:
+    root = fixture_repository(tmp_path)
+    result = statuses(Preflight(root, "1.2.3", relaxed=False))
+
+    assert set(result.values()) == {"pass"}
+
+
+def test_relaxed_precheck_warns_for_dirty_tree(tmp_path: Path) -> None:
+    root = fixture_repository(tmp_path)
+    write(root, "scratch.txt", "dirty\n")
+
+    result = statuses(Preflight(root, "1.2.3", relaxed=True))
+
+    assert result["git-clean"] == "warning"
+    assert "fail" not in result.values()
+
+
+def test_formal_preflight_rejects_version_drift(tmp_path: Path) -> None:
+    root = fixture_repository(tmp_path)
+    write(
+        root,
+        "server/sdk/python/src/insightface_server/__init__.py",
+        '__version__ = "9.9.9"\n',
+    )
+
+    result = statuses(Preflight(root, "1.2.3", relaxed=False))
+
+    assert result["sdk-runtime-version"] == "fail"
+    assert result["git-clean"] == "fail"
+
+
+def test_formal_preflight_accepts_same_source_release_tag_for_resume(
+    tmp_path: Path,
+) -> None:
+    root = fixture_repository(tmp_path)
+    subprocess.run(
+        ("git", "tag", "-a", "server-v1.2.3", "-m", "release"),
+        cwd=root,
+        check=True,
+    )
+
+    result = statuses(Preflight(root, "1.2.3", relaxed=False))
+
+    assert result["git-tag-idempotent"] == "pass"
+    assert result["git-tag-annotated"] == "pass"
+    assert "fail" not in result.values()
+
+
+def test_formal_preflight_rejects_lightweight_release_tag(tmp_path: Path) -> None:
+    root = fixture_repository(tmp_path)
+    subprocess.run(
+        ("git", "tag", "server-v1.2.3"),
+        cwd=root,
+        check=True,
+    )
+
+    result = statuses(Preflight(root, "1.2.3", relaxed=False))
+
+    assert result["git-tag-idempotent"] == "pass"
+    assert result["git-tag-annotated"] == "fail"
+
+
+def test_formal_preflight_rejects_release_tag_from_another_source(
+    tmp_path: Path,
+) -> None:
+    root = fixture_repository(tmp_path)
+    subprocess.run(
+        ("git", "tag", "-a", "server-v1.2.3", "-m", "release"),
+        cwd=root,
+        check=True,
+    )
+    write(root, "next.txt", "new source\n")
+    subprocess.run(("git", "add", "next.txt"), cwd=root, check=True)
+    subprocess.run(
+        ("git", "commit", "-m", "next source"),
+        cwd=root,
+        check=True,
+        capture_output=True,
+    )
+
+    result = statuses(Preflight(root, "1.2.3", relaxed=False))
+
+    assert result["git-tag-idempotent"] == "fail"
diff --git a/server/tests/unit/test_rtsp_manager.py b/server/tests/unit/test_rtsp_manager.py
new file mode 100644
index 000000000..6da3e52f1
--- /dev/null
+++ b/server/tests/unit/test_rtsp_manager.py
@@ -0,0 +1,532 @@
+from __future__ import annotations
+
+import threading
+import time
+from collections.abc import Callable
+from dataclasses import replace
+from typing import Any, cast
+
+import pytest
+from fastapi.testclient import TestClient
+from insightface_server.request_context import REQUEST_DEADLINE
+from insightface_server.services.rtsp import (
+    MonitorLimitError,
+    MonitorManager,
+    MonitorOptions,
+)
+
+
+def _manager_payload(monitor_id: str, *, enabled: bool) -> dict[str, Any]:
+    return {
+        "id": monitor_id,
+        "name": monitor_id.title(),
+        "description": "Synthetic lifecycle test Monitor",
+        "enabled": enabled,
+        "source_type": "rtsp",
+        "url": f"rtsp://camera.example/{monitor_id}",
+        "collection_id": "employees",
+        "inference_fps": 2.0,
+        "match_threshold": None,
+        "event_buffer_size": 100,
+        "confirm_frames": 3,
+        "absence_timeout_seconds": 3.0,
+        "cooldown_seconds": 10.0,
+        "emit_unknown": True,
+        "preview_enabled": False,
+    }
+
+
+class _ControlledSession:
+    def __init__(
+        self,
+        options: MonitorOptions,
+        on_start: Callable[[_ControlledSession], None] | None,
+    ) -> None:
+        self.options = options
+        self.on_start = on_start
+        self.start_calls = 0
+        self.stop_calls = 0
+        self._stopped = False
+
+    def start(self) -> None:
+        self.start_calls += 1
+        if self.on_start is not None:
+            self.on_start(self)
+
+    def stop(self) -> None:
+        self.stop_calls += 1
+        self._stopped = True
+
+    @property
+    def stopped(self) -> bool:
+        return self._stopped
+
+    def apply_live_options(
+        self,
+        *,
+        preview_enabled: bool | None = None,
+        event_buffer_size: int | None = None,
+        name: str | None = None,
+        description: str | None = None,
+        updated_at: str | None = None,
+    ) -> None:
+        self.options = replace(
+            self.options,
+            preview_enabled=(
+                self.options.preview_enabled
+                if preview_enabled is None
+                else preview_enabled
+            ),
+            event_buffer_size=(
+                self.options.event_buffer_size
+                if event_buffer_size is None
+                else event_buffer_size
+            ),
+            name=self.options.name if name is None else name,
+            description=(
+                self.options.description
+                if description is None
+                else description
+            ),
+            updated_at=(
+                self.options.updated_at if updated_at is None else updated_at
+            ),
+        )
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "status": "stopped" if self.stopped else "starting",
+            "connected": False,
+            "stream_epoch": self.options.id,
+            "last_frame_at": None,
+            "last_error": None,
+            "preview_active": False,
+            "preview_viewers": 0,
+        }
+
+
+class _SessionFactory:
+    def __init__(
+        self,
+        on_start: Callable[[_ControlledSession], None] | None = None,
+    ) -> None:
+        self.on_start = on_start
+        self.sessions: list[_ControlledSession] = []
+        self._lock = threading.Lock()
+
+    def __call__(
+        self,
+        _service,
+        options: MonitorOptions,
+        **_kwargs: Any,
+    ) -> _ControlledSession:
+        session = _ControlledSession(options, self.on_start)
+        with self._lock:
+            self.sessions.append(session)
+        return session
+
+
+def _manager(
+    client: TestClient,
+    *,
+    max_monitors: int,
+    factory: _SessionFactory,
+) -> MonitorManager:
+    manager: MonitorManager = cast(Any, client.app).state.monitors
+    manager.max_monitors = max_monitors
+    manager.session_factory = cast(Any, factory)
+    return manager
+
+
+def _concurrent(
+    operations: list[Callable[[], Any]],
+) -> list[Any | BaseException]:
+    ready = threading.Barrier(len(operations) + 1)
+    results: list[Any | BaseException | None] = [None] * len(operations)
+
+    def run(index: int, operation: Callable[[], Any]) -> None:
+        ready.wait()
+        try:
+            results[index] = operation()
+        except BaseException as exc:  # retain worker failures for parent assertions
+            results[index] = exc
+
+    threads = [
+        threading.Thread(target=run, args=(index, operation), daemon=True)
+        for index, operation in enumerate(operations)
+    ]
+    for thread in threads:
+        thread.start()
+    ready.wait()
+    for thread in threads:
+        thread.join(timeout=5.0)
+        assert not thread.is_alive()
+    return [result for result in results if result is not None]
+
+
+def test_concurrent_enabled_creates_cannot_oversubscribe_capacity(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+    start_entered = threading.Event()
+    release_start = threading.Event()
+    first_start_lock = threading.Lock()
+    first_start = True
+
+    def block_first_start(_session: _ControlledSession) -> None:
+        nonlocal first_start
+        with first_start_lock:
+            should_block = first_start
+            first_start = False
+        if should_block:
+            start_entered.set()
+            assert release_start.wait(timeout=5.0)
+
+    factory = _SessionFactory(block_first_start)
+    manager = _manager(client, max_monitors=1, factory=factory)
+    ready = threading.Barrier(3)
+    results: list[Any | BaseException | None] = [None, None]
+
+    def create(index: int, monitor_id: str) -> None:
+        ready.wait()
+        try:
+            results[index] = manager.create(
+                _manager_payload(monitor_id, enabled=True)
+            )
+        except BaseException as exc:
+            results[index] = exc
+
+    threads = [
+        threading.Thread(target=create, args=(0, "north"), daemon=True),
+        threading.Thread(target=create, args=(1, "south"), daemon=True),
+    ]
+    for thread in threads:
+        thread.start()
+    ready.wait()
+    assert start_entered.wait(timeout=5.0)
+    release_start.set()
+    for thread in threads:
+        thread.join(timeout=5.0)
+        assert not thread.is_alive()
+
+    assert sum(isinstance(result, dict) for result in results) == 1
+    assert sum(isinstance(result, MonitorLimitError) for result in results) == 1
+    rows = manager.repository.list_monitors("", 10)
+    assert len(rows) == 1
+    assert rows[0]["enabled"] is True
+    assert len(manager._sessions) == 1  # noqa: SLF001 - lifecycle invariant
+    assert len(factory.sessions) == 1
+
+
+def test_concurrent_enables_leave_exactly_one_durable_and_running(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+    factory = _SessionFactory()
+    manager = _manager(client, max_monitors=1, factory=factory)
+    manager.create(_manager_payload("north", enabled=False))
+    manager.create(_manager_payload("south", enabled=False))
+
+    results = _concurrent(
+        [
+            lambda: manager.update("north", {"enabled": True}),
+            lambda: manager.update("south", {"enabled": True}),
+        ]
+    )
+
+    assert sum(isinstance(result, dict) for result in results) == 1
+    assert sum(isinstance(result, MonitorLimitError) for result in results) == 1
+    rows = manager.repository.list_monitors("", 10)
+    assert sum(bool(row["enabled"]) for row in rows) == 1
+    assert len(manager._sessions) == 1  # noqa: SLF001 - lifecycle invariant
+    running_id = next(iter(manager._sessions))  # noqa: SLF001
+    running_row = manager.repository.get_monitor(running_id)
+    assert running_row is not None
+    assert running_row["enabled"] is True
+
+
+def test_startup_serializes_capacity_against_a_concurrent_create(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+    app_manager = _manager(
+        client,
+        max_monitors=1,
+        factory=_SessionFactory(),
+    )
+    app_manager.create(_manager_payload("restored", enabled=False))
+    restored = app_manager.repository.update_monitor(
+        "restored",
+        {"enabled": True},
+    )
+    assert restored is not None
+
+    start_entered = threading.Event()
+    release_start = threading.Event()
+
+    def block_restore(_session: _ControlledSession) -> None:
+        start_entered.set()
+        assert release_start.wait(timeout=5.0)
+
+    factory = _SessionFactory(block_restore)
+    manager = MonitorManager(
+        app_manager.service,
+        app_manager.repository,
+        app_manager.cursors,
+        app_manager.credentials,
+        max_monitors=1,
+        session_factory=cast(Any, factory),
+    )
+    results: list[Any | BaseException | None] = [None, None]
+
+    def startup() -> None:
+        try:
+            manager.startup()
+            results[0] = "started"
+        except BaseException as exc:
+            results[0] = exc
+
+    def create() -> None:
+        try:
+            results[1] = manager.create(
+                _manager_payload("late", enabled=True)
+            )
+        except BaseException as exc:
+            results[1] = exc
+
+    startup_thread = threading.Thread(target=startup, daemon=True)
+    startup_thread.start()
+    assert start_entered.wait(timeout=5.0)
+    create_thread = threading.Thread(target=create, daemon=True)
+    create_thread.start()
+    release_start.set()
+    for thread in (startup_thread, create_thread):
+        thread.join(timeout=5.0)
+        assert not thread.is_alive()
+
+    assert results[0] == "started"
+    assert isinstance(results[1], MonitorLimitError)
+    assert app_manager.repository.get_monitor("late") is None
+    assert set(manager._sessions) == {"restored"}  # noqa: SLF001
+    manager.close()
+
+
+def test_restart_and_delete_are_serialized_without_leaking_either_session(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+    replacement_entered = threading.Event()
+    release_replacement = threading.Event()
+
+    def block_replacement(session: _ControlledSession) -> None:
+        if session.options.url.endswith("/replacement"):
+            replacement_entered.set()
+            assert release_replacement.wait(timeout=5.0)
+
+    factory = _SessionFactory(block_replacement)
+    manager = _manager(client, max_monitors=1, factory=factory)
+    manager.create(_manager_payload("front-gate", enabled=True))
+    results: list[Any | BaseException | None] = [None, None]
+
+    def update() -> None:
+        try:
+            results[0] = manager.update(
+                "front-gate",
+                {"url": "rtsp://camera.example/replacement"},
+            )
+        except BaseException as exc:
+            results[0] = exc
+
+    delete_started = threading.Event()
+
+    def delete() -> None:
+        delete_started.set()
+        try:
+            manager.delete("front-gate")
+            results[1] = "deleted"
+        except BaseException as exc:
+            results[1] = exc
+
+    update_thread = threading.Thread(target=update, daemon=True)
+    update_thread.start()
+    assert replacement_entered.wait(timeout=5.0)
+    delete_thread = threading.Thread(target=delete, daemon=True)
+    delete_thread.start()
+    assert delete_started.wait(timeout=5.0)
+    release_replacement.set()
+    for thread in (update_thread, delete_thread):
+        thread.join(timeout=5.0)
+        assert not thread.is_alive()
+
+    assert isinstance(results[0], dict)
+    assert results[1] == "deleted"
+    assert manager.repository.get_monitor("front-gate") is None
+    assert "front-gate" not in manager._sessions  # noqa: SLF001
+    assert len(factory.sessions) == 2
+    assert all(session.stopped for session in factory.sessions)
+
+
+def test_create_start_failure_removes_durable_row_and_candidate(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+
+    def fail_start(_session: _ControlledSession) -> None:
+        raise RuntimeError("synthetic start failure")
+
+    factory = _SessionFactory(fail_start)
+    manager = _manager(client, max_monitors=1, factory=factory)
+
+    with pytest.raises(RuntimeError, match="synthetic start failure"):
+        manager.create(_manager_payload("front-gate", enabled=True))
+
+    assert manager.repository.get_monitor("front-gate") is None
+    assert "front-gate" not in manager._sessions  # noqa: SLF001
+    assert len(factory.sessions) == 1
+    assert factory.sessions[0].stopped
+
+
+def test_create_start_failure_compensates_after_request_deadline_expires(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+
+    def expire_request_then_fail(_session: _ControlledSession) -> None:
+        REQUEST_DEADLINE.set(time.monotonic() - 1.0)
+        raise RuntimeError("synthetic delayed start failure")
+
+    factory = _SessionFactory(expire_request_then_fail)
+    manager = _manager(client, max_monitors=1, factory=factory)
+    deadline_token = REQUEST_DEADLINE.set(time.monotonic() + 5.0)
+    try:
+        with pytest.raises(RuntimeError, match="synthetic delayed start failure"):
+            manager.create(_manager_payload("front-gate", enabled=True))
+    finally:
+        REQUEST_DEADLINE.reset(deadline_token)
+
+    assert manager.repository.get_monitor("front-gate") is None
+    assert "front-gate" not in manager._sessions  # noqa: SLF001
+    assert len(factory.sessions) == 1
+    assert factory.sessions[0].stopped
+
+
+def test_restart_start_failure_preserves_durable_config_and_old_session(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+
+    def fail_replacement(session: _ControlledSession) -> None:
+        if session.options.url.endswith("/replacement"):
+            raise RuntimeError("synthetic replacement failure")
+
+    factory = _SessionFactory(fail_replacement)
+    manager = _manager(client, max_monitors=1, factory=factory)
+    manager.create(_manager_payload("front-gate", enabled=True))
+    old_session = manager._sessions["front-gate"]  # noqa: SLF001
+
+    with pytest.raises(RuntimeError, match="synthetic replacement failure"):
+        manager.update(
+            "front-gate",
+            {"url": "rtsp://camera.example/replacement"},
+        )
+
+    row = manager.repository.get_monitor("front-gate")
+    assert row is not None
+    assert manager.credentials.decrypt(
+        row["source_url_ciphertext"],
+        scope="monitor:front-gate",
+    ) == "rtsp://camera.example/front-gate"
+    assert manager._sessions["front-gate"] is old_session  # noqa: SLF001
+    assert not old_session.stopped
+    assert len(factory.sessions) == 2
+    assert factory.sessions[1].stopped
+
+
+def test_restart_database_failure_stops_candidate_and_preserves_old_session(
+    client: TestClient,
+    create_collection,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    create_collection(client)
+    factory = _SessionFactory()
+    manager = _manager(client, max_monitors=1, factory=factory)
+    manager.create(_manager_payload("front-gate", enabled=True))
+    old_session = manager._sessions["front-gate"]  # noqa: SLF001
+
+    def fail_update(_monitor_id: str, _changes: dict[str, Any]):
+        raise RuntimeError("synthetic database failure")
+
+    monkeypatch.setattr(manager.repository, "update_monitor", fail_update)
+
+    with pytest.raises(RuntimeError, match="synthetic database failure"):
+        manager.update(
+            "front-gate",
+            {"url": "rtsp://camera.example/replacement"},
+        )
+
+    row = manager.repository.get_monitor("front-gate")
+    assert row is not None
+    assert manager.credentials.decrypt(
+        row["source_url_ciphertext"],
+        scope="monitor:front-gate",
+    ) == "rtsp://camera.example/front-gate"
+    assert manager._sessions["front-gate"] is old_session  # noqa: SLF001
+    assert not old_session.stopped
+    assert len(factory.sessions) == 2
+    assert factory.sessions[1].stopped
+
+
+def test_close_waits_for_inflight_create_and_prevents_late_session_publish(
+    client: TestClient,
+    create_collection,
+) -> None:
+    create_collection(client)
+    start_entered = threading.Event()
+    release_start = threading.Event()
+
+    def block_start(_session: _ControlledSession) -> None:
+        start_entered.set()
+        assert release_start.wait(timeout=5.0)
+
+    factory = _SessionFactory(block_start)
+    manager = _manager(client, max_monitors=1, factory=factory)
+    results: list[Any | BaseException | None] = [None, None]
+
+    def create() -> None:
+        try:
+            results[0] = manager.create(
+                _manager_payload("front-gate", enabled=True)
+            )
+        except BaseException as exc:
+            results[0] = exc
+
+    def close() -> None:
+        try:
+            manager.close()
+            results[1] = "closed"
+        except BaseException as exc:
+            results[1] = exc
+
+    create_thread = threading.Thread(target=create, daemon=True)
+    create_thread.start()
+    assert start_entered.wait(timeout=5.0)
+    close_thread = threading.Thread(target=close, daemon=True)
+    close_thread.start()
+    release_start.set()
+    for thread in (create_thread, close_thread):
+        thread.join(timeout=5.0)
+        assert not thread.is_alive()
+
+    assert isinstance(results[0], dict)
+    assert results[1] == "closed"
+    assert not manager._sessions  # noqa: SLF001
+    assert factory.sessions[0].stopped
+    with pytest.raises(RuntimeError, match="closed"):
+        manager.create(_manager_payload("too-late", enabled=True))
diff --git a/server/tests/unit/test_rtsp_stream.py b/server/tests/unit/test_rtsp_stream.py
new file mode 100644
index 000000000..b5b806906
--- /dev/null
+++ b/server/tests/unit/test_rtsp_stream.py
@@ -0,0 +1,232 @@
+from __future__ import annotations
+
+import threading
+import time
+
+import cv2
+import numpy as np
+import pytest
+from insightface_server.services.rtsp import (
+    MonitorOptions,
+    MonitorPreviewDisabledError,
+    MonitorSession,
+    next_inference_time,
+    redacted_rtsp_source,
+)
+
+
+def _options(**changes) -> MonitorOptions:
+    values = {
+        "id": "front-gate",
+        "name": "Front gate",
+        "description": "",
+        "enabled": True,
+        "url": "rtsp://camera.example/live",
+        "collection_id": "employees",
+        "inference_fps": 30.0,
+        "match_threshold": None,
+        "event_buffer_size": 10,
+        "confirm_frames": 2,
+        "absence_timeout_seconds": 0.1,
+        "cooldown_seconds": 0.0,
+        "emit_unknown": True,
+        "preview_enabled": False,
+        "created_at": "2026-07-23T00:00:00Z",
+        "updated_at": "2026-07-23T00:00:00Z",
+    }
+    values.update(changes)
+    return MonitorOptions(**values)
+
+
+def test_next_inference_time_never_schedules_catch_up_work() -> None:
+    assert next_inference_time(started=10.0, finished=10.2, inference_fps=2.0) == 10.5
+    assert next_inference_time(started=10.0, finished=10.8, inference_fps=2.0) == 10.8
+
+
+def test_rtsp_source_redaction_removes_credentials_and_query() -> None:
+    source = redacted_rtsp_source(
+        "rtsp://camera-user:secret@192.168.1.20:8554/live?token=private"
+    )
+    assert source == "rtsp://192.168.1.20:8554/live"
+    assert "secret" not in source
+    assert "token" not in source
+
+
+class _FastCapture:
+    def __init__(self) -> None:
+        self.released = False
+        self.value = 40
+
+    def isOpened(self) -> bool:
+        return not self.released
+
+    def get(self, property_id: int) -> float:
+        return 25.0 if property_id == cv2.CAP_PROP_FPS else 0.0
+
+    def read(self):
+        if self.released:
+            return False, None
+        time.sleep(0.002)
+        self.value = 40 + ((self.value + 1) % 4)
+        return True, np.full((80, 120, 3), self.value, dtype=np.uint8)
+
+    def release(self) -> None:
+        self.released = True
+
+
+class _SlowService:
+    def __init__(self) -> None:
+        self.active = 0
+        self.maximum_active = 0
+        self.lock = threading.Lock()
+
+    def search_all_faces(self, *_args, **_kwargs):
+        with self.lock:
+            self.active += 1
+            self.maximum_active = max(self.maximum_active, self.active)
+        try:
+            time.sleep(0.05)
+            return (
+                [
+                    {
+                        "face": {
+                            "bbox": {
+                                "pixels": {
+                                    "x": 10,
+                                    "y": 10,
+                                    "width": 30,
+                                    "height": 30,
+                                },
+                                "normalized": {
+                                    "left": 1 / 12,
+                                    "top": 1 / 8,
+                                    "width": 0.25,
+                                    "height": 0.375,
+                                },
+                            },
+                            "landmarks": [],
+                            "detection_score": 0.99,
+                            "quality": {"score": 0.9},
+                        },
+                        "match": {
+                            "person": {
+                                "id": "alice",
+                                "name": "Alice",
+                                "external_id": "A-1",
+                                "metadata": {"must_not_leak": True},
+                            },
+                            "similarity": 0.88,
+                            "matched_face_id": "face-1",
+                        },
+                    }
+                ],
+                0.4,
+            )
+        finally:
+            with self.lock:
+                self.active -= 1
+
+
+def _session(options: MonitorOptions) -> tuple[MonitorSession, _SlowService]:
+    service = _SlowService()
+    session = MonitorSession(
+        service,  # type: ignore[arg-type]
+        options,
+        max_faces=100,
+        preview_fps=30.0,
+        jpeg_quality=90,
+        open_timeout_seconds=1.0,
+        read_timeout_seconds=1.0,
+        reconnect_delay_seconds=0.1,
+        capture_factory=lambda *_args: _FastCapture(),
+    )
+    return session, service
+
+
+def test_session_keeps_latest_frame_and_emits_confirmed_memory_events() -> None:
+    session, service = _session(_options())
+    session.start()
+    deadline = time.monotonic() + 2.0
+    while session.state()["inference"]["processed_frames"] < 3 and time.monotonic() < deadline:
+        time.sleep(0.01)
+    state = session.state()
+    events = session.event_page(cursor_epoch=None, after_sequence=None, limit=10)
+    session.stop()
+
+    assert state["inference"]["processed_frames"] >= 3
+    assert state["inference"]["decoded_frames"] > state["inference"]["processed_frames"]
+    assert state["inference"]["dropped_frames"] > 0
+    assert state["inference"]["capacity_limited"] is True
+    assert service.maximum_active == 1
+    assert state["matched_faces"] == 1
+    assert state["faces"][0]["person"] == {
+        "id": "alice",
+        "name": "Alice",
+        "external_id": "A-1",
+    }
+    assert any(event["type"] == "person_enter" for event in events["events"])
+
+
+def test_preview_is_disabled_by_default_and_raw_when_explicitly_enabled() -> None:
+    disabled, _ = _session(_options(preview_enabled=False))
+    with pytest.raises(MonitorPreviewDisabledError):
+        next(disabled.iter_mjpeg())
+
+    session, _ = _session(_options(preview_enabled=True, confirm_frames=1))
+    session.start()
+    deadline = time.monotonic() + 2.0
+    while not session.connected and time.monotonic() < deadline:
+        time.sleep(0.01)
+    frame = next(session.iter_mjpeg())
+    session.stop()
+
+    header, jpeg = frame.split(b"\r\n\r\n", 1)
+    decoded = cv2.imdecode(
+        np.frombuffer(jpeg.rstrip(b"\r\n"), dtype=np.uint8),
+        cv2.IMREAD_COLOR,
+    )
+    assert b"Content-Type: image/jpeg" in header
+    assert decoded is not None
+    # The server relays the unannotated source. A drawn green/amber rectangle
+    # would create a large channel spread on this otherwise uniform frame.
+    assert int(decoded.max()) - int(decoded.min()) < 10
+
+
+def test_event_ring_reports_truncation_reset_and_resizes_to_latest_items() -> None:
+    session, _ = _session(_options(event_buffer_size=3))
+    with session._condition:  # noqa: SLF001 - deterministic unit-level state exercise
+        for index in range(5):
+            session._append_event(  # noqa: SLF001
+                f"test_{index}",
+                now=float(index),
+            )
+    page = session.event_page(
+        cursor_epoch=session.stream_epoch,
+        after_sequence=1,
+        limit=10,
+    )
+    reset = session.event_page(
+        cursor_epoch="old-epoch",
+        after_sequence=5,
+        limit=2,
+    )
+    session.apply_live_options(event_buffer_size=2)
+    resized = session.event_page(cursor_epoch=None, after_sequence=None, limit=10)
+
+    assert [item["sequence"] for item in page["events"]] == [3, 4, 5]
+    assert page["truncated"] is True
+    assert reset["stream_reset"] is True
+    assert [item["sequence"] for item in reset["events"]] == [4, 5]
+    assert [item["sequence"] for item in resized["events"]] == [4, 5]
+
+
+@pytest.mark.parametrize(
+    ("fps", "expected"),
+    [(0.5, 2.0), (1.0, 1.0), (10.0, 0.1)],
+)
+def test_requested_rate_maps_to_start_interval(fps: float, expected: float) -> None:
+    assert next_inference_time(
+        started=4.0,
+        finished=4.01,
+        inference_fps=fps,
+    ) == pytest.approx(4.0 + expected)
diff --git a/server/tests/unit/test_search_manager.py b/server/tests/unit/test_search_manager.py
new file mode 100644
index 000000000..85759fc65
--- /dev/null
+++ b/server/tests/unit/test_search_manager.py
@@ -0,0 +1,732 @@
+from __future__ import annotations
+
+import threading
+from pathlib import Path
+
+import numpy as np
+import pytest
+from insightface_server.search import ReferenceSearchBackend, SearchIndexManager
+from insightface_server.search.base import SearchIndexError
+from insightface_server.search.reference import ReferenceSearchIndex, profile_similarity
+from insightface_server.storage import Database, Repository
+
+
+def unit(values: list[float]) -> np.ndarray:
+    vector = np.asarray(values, dtype=np.float32)
+    return vector / np.linalg.norm(vector)
+
+
+def collection_item() -> dict[str, object]:
+    return {
+        "id": "employees",
+        "name": "Employees",
+        "description": "",
+        "default_threshold": 0.3,
+        "metadata": {},
+        "model_id": "recognition",
+        "model_version": "1",
+        "model_digest": "a" * 64,
+        "embedding_dimension": 4,
+        "preprocessing_version": "1",
+        "search_profile": "fp32_v1",
+        "capacity_rows": 4,
+        "max_faces_per_person": 3,
+        "load_policy": "lazy",
+    }
+
+
+def face(face_id: str, embedding: np.ndarray) -> dict[str, object]:
+    return {
+        "id": face_id,
+        "embedding": embedding,
+        "embedding_dimension": 4,
+        "bounding_box": {"pixels": {"x": 1, "y": 2, "width": 3, "height": 4}},
+        "landmarks": None,
+        "detection_score": 0.9,
+        "quality": {"score": 0.8},
+        "model_id": "recognition",
+        "model_version": "1",
+        "model_digest": "a" * 64,
+        "preprocessing_version": "1",
+        "crop_path": None,
+        "created_at": "2026-01-01T00:00:00.000Z",
+    }
+
+
+class SelectivelyFailingBackend(ReferenceSearchBackend):
+    """Reference backend with deterministic allocation failures for rollback tests."""
+
+    def __init__(self, *failed_capacities: int) -> None:
+        self.failed_capacities = set(failed_capacities)
+
+    def create_index(self, *, profile: str, dimension: int, capacity_rows: int):
+        if capacity_rows in self.failed_capacities:
+            raise SearchIndexError(f"injected allocation failure for {capacity_rows}")
+        return super().create_index(
+            profile=profile,
+            dimension=dimension,
+            capacity_rows=capacity_rows,
+        )
+
+
+class CountingReferenceIndex(ReferenceSearchIndex):
+    def __init__(
+        self,
+        owner: CountingReferenceBackend,
+        handle_id: int,
+        *,
+        profile: str,
+        dimension: int,
+        capacity_rows: int,
+    ) -> None:
+        self.owner = owner
+        self.handle_id = handle_id
+        super().__init__(
+            profile=profile,
+            dimension=dimension,
+            capacity_rows=capacity_rows,
+        )
+
+    def close(self) -> None:
+        if not self._closed:
+            self.owner.closed_handles.add(self.handle_id)
+            self.owner.active_handles.remove(self.handle_id)
+        super().close()
+
+
+class CountingReferenceBackend(ReferenceSearchBackend):
+    def __init__(self) -> None:
+        self.created_handles = 0
+        self.active_handles: set[int] = set()
+        self.closed_handles: set[int] = set()
+
+    def create_index(self, *, profile: str, dimension: int, capacity_rows: int):
+        self.created_handles += 1
+        handle_id = self.created_handles
+        self.active_handles.add(handle_id)
+        return CountingReferenceIndex(
+            self,
+            handle_id,
+            profile=profile,
+            dimension=dimension,
+            capacity_rows=capacity_rows,
+        )
+
+
+class ConcurrentReadReferenceIndex(ReferenceSearchIndex):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.barrier = threading.Barrier(2)
+        self.concurrent_probe_enabled = False
+
+    def _scores(self, query: np.ndarray, slots: np.ndarray) -> np.ndarray:
+        if self.concurrent_probe_enabled:
+            self.barrier.wait(timeout=5)
+        return super()._scores(query, slots)
+
+
+class ConcurrentReadReferenceBackend(ReferenceSearchBackend):
+    def __init__(self) -> None:
+        self.index: ConcurrentReadReferenceIndex | None = None
+
+    def create_index(self, *, profile: str, dimension: int, capacity_rows: int):
+        self.index = ConcurrentReadReferenceIndex(
+            profile=profile,
+            dimension=dimension,
+            capacity_rows=capacity_rows,
+        )
+        return self.index
+
+
+def test_same_collection_searches_use_one_generation_concurrently(tmp_path: Path) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    backend = ConcurrentReadReferenceBackend()
+    manager = SearchIndexManager(repository, backend)
+    manager.ensure_ready("employees")
+    assert backend.index is not None
+    backend.index.concurrent_probe_enabled = True
+
+    def search() -> list[dict[str, object]]:
+        return manager.search(
+            "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.0
+        )
+
+    failures: list[BaseException] = []
+    results: list[list[dict[str, object]]] = []
+
+    def run() -> None:
+        try:
+            results.append(search())
+        except BaseException as exc:  # pragma: no cover - asserted below
+            failures.append(exc)
+
+    first = threading.Thread(target=run)
+    second = threading.Thread(target=run)
+    first.start()
+    second.start()
+    first.join(timeout=6)
+    second.join(timeout=6)
+
+    assert not first.is_alive()
+    assert not second.is_alive()
+    assert failures == []
+    assert len(results) == 2
+    assert all(result[0]["person"]["id"] == "alice" for result in results)
+    manager.close()
+
+
+def test_manager_builds_groups_and_applies_add_delete_mutations(tmp_path: Path) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "bob", "name": "Bob", "metadata": {}},
+        [face("bob-1", unit([0, 1, 0, 0]))],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend(), build_batch_rows=1)
+
+    initial = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=5, threshold=0.0
+    )
+    assert [item["person"]["id"] for item in initial] == ["alice", "bob"]
+    assert initial[0]["similarity"] == 1.0
+    assert initial[1]["similarity"] == 0.0
+    assert repository.get_search_changes("employees") == []
+
+    def add():
+        mutation = repository.add_faces_with_mutation(
+            "employees", "bob", [face("bob-2", unit([0.8, 0.6, 0, 0]))]
+        )
+        return None, mutation
+
+    manager.run_mutation("employees", add, expected_additions=1)
+    assert repository.get_search_changes("employees") == []
+    updated = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=5, threshold=0.0
+    )
+    assert [item["person"]["id"] for item in updated] == ["alice", "bob"]
+    assert updated[1]["matched_face_id"] == "bob-2"
+    assert updated[1]["similarity"] == pytest.approx(0.8)
+
+    def delete():
+        deleted, crop, mutation = repository.delete_face_with_mutation(
+            "employees", "alice", "alice-1"
+        )
+        return (deleted, crop), mutation
+
+    assert manager.run_mutation("employees", delete) == (True, None)
+    assert repository.get_search_changes("employees") == []
+    after_delete = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=5, threshold=0.0
+    )
+    assert [item["person"]["id"] for item in after_delete] == ["bob"]
+    assert manager.runtime_summary()["collections"][0]["applied_revision"] == 4
+    manager.close()
+
+
+def test_best_other_person_excludes_target_and_returns_profile_score(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "bob", "name": "Bob", "metadata": {}},
+        [face("bob-1", unit([0.8, 0.6, 0, 0]))],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+
+    other = manager.best_other_person(
+        "employees", unit([1, 0, 0, 0]), exclude_person_id="alice"
+    )
+    assert other == {
+        "person_id": "bob",
+        "face_id": "bob-1",
+        "similarity": pytest.approx(0.8),
+    }
+    assert manager.best_other_person(
+        "employees", unit([1, 0, 0, 0]), exclude_person_id="nobody"
+    ) == {
+        "person_id": "alice",
+        "face_id": "alice-1",
+        "similarity": 1.0,
+    }
+    manager.close()
+
+
+def test_best_other_person_preserves_unclipped_int8_review_score(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    values = {
+        **collection_item(),
+        "embedding_dimension": 512,
+        "search_profile": "int8_x736_v1",
+    }
+    repository.create_collection(values)
+    embedding = unit([1.0] * 512)
+    sample = {**face("alice-1", embedding), "embedding_dimension": 512}
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [sample],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+
+    other = manager.best_other_person(
+        "employees", embedding, exclude_person_id="nobody"
+    )
+
+    assert other is not None
+    assert other["similarity"] == pytest.approx(
+        profile_similarity("int8_x736_v1", embedding, embedding)
+    )
+    assert other["similarity"] > 1.0
+    manager.close()
+
+
+def test_startup_discards_pre_restart_events_for_lazy_full_rebuild(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    assert len(repository.get_search_changes("employees")) == 1
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+
+    manager.startup()
+
+    assert repository.get_search_changes("employees") == []
+    assert manager.runtime_summary()["collections"] == []
+    matches = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.3
+    )
+    assert matches[0]["person"]["id"] == "alice"
+    manager.close()
+
+
+def test_outbox_acknowledgement_failure_retains_event_without_failing_search(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+
+    def fail_acknowledgement(_collection_id: str, _revision: int) -> int:
+        raise OSError("injected acknowledgement failure")
+
+    acknowledge = repository.acknowledge_search_changes
+    monkeypatch.setattr(repository, "acknowledge_search_changes", fail_acknowledgement)
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+
+    matches = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.3
+    )
+
+    assert matches[0]["person"]["id"] == "alice"
+    assert len(repository.get_search_changes("employees")) == 1
+
+    monkeypatch.setattr(repository, "acknowledge_search_changes", acknowledge)
+    retry = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.3
+    )
+    assert retry[0]["person"]["id"] == "alice"
+    assert repository.get_search_changes("employees") == []
+    manager.close()
+
+
+def test_manager_compares_threshold_against_unrounded_raw_cosine(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    query = unit([1, 0, 0, 0])
+    just_below = np.asarray(
+        [0.2999996, np.sqrt(1.0 - 0.2999996**2), 0, 0], dtype=np.float32
+    )
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", just_below)],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+
+    assert round(float(np.dot(query, just_below)), 6) == 0.3
+    assert manager.search("employees", query, limit=1, threshold=0.3) == []
+    manager.close()
+
+
+def test_eager_create_allocation_failure_does_not_commit_collection(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    values = {**collection_item(), "load_policy": "eager"}
+    manager = SearchIndexManager(repository, SelectivelyFailingBackend(4))
+
+    with pytest.raises(SearchIndexError, match="injected allocation failure"):
+        manager.run_create(values, lambda: repository.create_collection(values))
+
+    assert repository.get_collection("employees") is None
+    assert manager.runtime_summary()["collections"] == []
+    manager.close()
+
+
+def test_capacity_patch_staging_failure_preserves_old_config_and_generation(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    backend = SelectivelyFailingBackend()
+    manager = SearchIndexManager(repository, backend)
+    assert manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.0
+    )[0]["person"]["id"] == "alice"
+    original_runtime = manager.runtime_summary()["collections"][0]
+    backend.failed_capacities.add(8)
+    update_called = False
+
+    def update():
+        nonlocal update_called
+        update_called = True
+        return repository.update_collection("employees", {"capacity_rows": 8})
+
+    with pytest.raises(SearchIndexError, match="injected allocation failure"):
+        manager.run_configuration_update(
+            "employees", update, changes={"capacity_rows": 8}
+        )
+
+    persisted = repository.get_collection("employees")
+    assert persisted is not None
+    assert persisted["capacity_rows"] == 4
+    assert update_called is False
+    runtime = manager.runtime_summary()["collections"][0]
+    assert runtime["state"] == "ready"
+    assert runtime["capacity_rows"] == 4
+    assert runtime["generation"] == original_runtime["generation"]
+    assert manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.0
+    )[0]["matched_face_id"] == "alice-1"
+    manager.close()
+
+
+def test_load_policy_transitions_materialize_and_release_index(tmp_path: Path) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    values = collection_item()
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+    manager.run_create(values, lambda: repository.create_collection(values))
+
+    initial = manager.runtime_summary()["collections"][0]
+    assert initial["state"] == "unloaded"
+
+    eager = manager.run_configuration_update(
+        "employees",
+        lambda: repository.update_collection("employees", {"load_policy": "eager"}),
+        changes={"load_policy": "eager"},
+    )
+    assert eager is not None and eager["load_policy"] == "eager"
+    ready = manager.runtime_summary()["collections"][0]
+    assert ready["state"] == "ready"
+    assert ready["capacity_rows"] == 4
+
+    lazy = manager.run_configuration_update(
+        "employees",
+        lambda: repository.update_collection("employees", {"load_policy": "lazy"}),
+        changes={"load_policy": "lazy"},
+    )
+    assert lazy is not None and lazy["load_policy"] == "lazy"
+    unloaded = manager.runtime_summary()["collections"][0]
+    assert unloaded["state"] == "unloaded"
+    assert "capacity_rows" not in unloaded
+    manager.close()
+
+
+@pytest.mark.parametrize("load_policy", ["eager", "lazy"])
+def test_unchanged_search_configuration_keeps_loaded_generation(
+    tmp_path: Path, load_policy: str
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    values = {**collection_item(), "load_policy": load_policy}
+    repository.create_collection(values)
+    backend = CountingReferenceBackend()
+    manager = SearchIndexManager(repository, backend)
+    manager.ensure_ready("employees")
+    before = manager.runtime_summary()["collections"][0]
+
+    unchanged = {
+        "capacity_rows": values["capacity_rows"],
+        "max_faces_per_person": values["max_faces_per_person"],
+        "load_policy": values["load_policy"],
+    }
+    updated = manager.run_configuration_update(
+        "employees",
+        lambda: repository.update_collection("employees", unchanged),
+        changes=unchanged,
+    )
+
+    assert updated is not None
+    assert backend.created_handles == 1
+    assert backend.closed_handles == set()
+    assert backend.active_handles == {1}
+    after = manager.runtime_summary()["collections"][0]
+    assert after["state"] == "ready"
+    assert after["generation"] == before["generation"]
+    assert after["capacity_rows"] == before["capacity_rows"]
+    manager.close()
+    assert backend.active_handles == set()
+    assert backend.closed_handles == {1}
+
+
+def test_collection_drop_serializes_with_search_and_returns_missing_not_index_error(
+    tmp_path: Path,
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+    manager.ensure_ready("employees")
+    delete_entered = threading.Event()
+    allow_delete = threading.Event()
+    failures: list[BaseException] = []
+    search_outcome: list[str] = []
+
+    def drop_operation():
+        delete_entered.set()
+        assert allow_delete.wait(timeout=5)
+        status, crops, mutation = repository.delete_collection_with_mutation(
+            "employees", force=True
+        )
+        return (status, crops), mutation
+
+    def drop() -> None:
+        try:
+            manager.run_drop("employees", drop_operation)
+        except BaseException as exc:  # pragma: no cover - asserted below
+            failures.append(exc)
+
+    def search() -> None:
+        try:
+            manager.search(
+                "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.0
+            )
+            search_outcome.append("served-before-delete")
+        except KeyError:
+            # FaceService maps this expected post-delete result to HTTP 404.
+            search_outcome.append("collection-not-found")
+        except BaseException as exc:  # pragma: no cover - asserted below
+            failures.append(exc)
+
+    drop_thread = threading.Thread(target=drop)
+    search_thread = threading.Thread(target=search)
+    drop_thread.start()
+    assert delete_entered.wait(timeout=5)
+    search_thread.start()
+    allow_delete.set()
+    drop_thread.join(timeout=5)
+    search_thread.join(timeout=5)
+
+    assert not drop_thread.is_alive()
+    assert not search_thread.is_alive()
+    assert failures == []
+    assert search_outcome == ["collection-not-found"]
+    manager.close()
+
+
+def test_collection_drop_finishes_after_derived_index_close_failure(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face("alice-1", unit([1, 0, 0, 0]))],
+    )
+    manager = SearchIndexManager(repository, ReferenceSearchBackend())
+    manager.ensure_ready("employees")
+    index = manager._slots["employees"].index
+    assert index is not None
+
+    def fail_close() -> None:
+        raise OSError("injected close failure")
+
+    monkeypatch.setattr(index, "close", fail_close)
+
+    def drop_operation():
+        status, crops, mutation = repository.delete_collection_with_mutation(
+            "employees", force=True
+        )
+        return (status, crops), mutation
+
+    result = manager.run_drop("employees", drop_operation)
+
+    assert result == ("deleted", [])
+    assert repository.get_collection("employees") is None
+    assert repository.get_search_changes("employees") == []
+    assert manager.runtime_summary()["collections"] == []
+    manager.close()
+
+
+def test_concurrent_drop_and_same_id_eager_recreate_registers_only_new_slot(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    values = {**collection_item(), "load_policy": "eager"}
+    repository.create_collection(values)
+    backend = CountingReferenceBackend()
+    manager = SearchIndexManager(repository, backend)
+    manager.ensure_ready("employees")
+
+    drop_entered = threading.Event()
+    allow_drop = threading.Event()
+    creator_captured_slot = threading.Event()
+    failures: list[BaseException] = []
+    outcomes: list[str] = []
+    original_slot = manager._slot
+
+    def observed_slot(collection_id: str):
+        slot = original_slot(collection_id)
+        if threading.current_thread().name == "creator":
+            creator_captured_slot.set()
+        return slot
+
+    monkeypatch.setattr(manager, "_slot", observed_slot)
+
+    def drop_operation():
+        drop_entered.set()
+        assert allow_drop.wait(timeout=5)
+        status, crops, mutation = repository.delete_collection_with_mutation(
+            "employees", force=True
+        )
+        return (status, crops), mutation
+
+    def drop() -> None:
+        try:
+            manager.run_drop("employees", drop_operation)
+            outcomes.append("dropped")
+        except BaseException as exc:  # pragma: no cover - asserted below
+            failures.append(exc)
+
+    def recreate() -> None:
+        try:
+            manager.run_create(values, lambda: repository.create_collection(values))
+            outcomes.append("recreated")
+        except BaseException as exc:  # pragma: no cover - asserted below
+            failures.append(exc)
+
+    drop_thread = threading.Thread(target=drop, name="dropper")
+    create_thread = threading.Thread(target=recreate, name="creator")
+    drop_thread.start()
+    assert drop_entered.wait(timeout=5)
+    create_thread.start()
+    assert creator_captured_slot.wait(timeout=5)
+    allow_drop.set()
+    drop_thread.join(timeout=5)
+    create_thread.join(timeout=5)
+
+    assert not drop_thread.is_alive()
+    assert not create_thread.is_alive()
+    assert failures == []
+    assert sorted(outcomes) == ["dropped", "recreated"]
+    assert backend.created_handles == 2
+    assert backend.closed_handles == {1}
+    assert backend.active_handles == {2}
+    runtime = manager.runtime_summary()["collections"]
+    assert len(runtime) == 1
+    assert runtime[0]["collection_id"] == "employees"
+    assert runtime[0]["state"] == "ready"
+
+    def add_person():
+        created, mutation = repository.create_person_with_faces_with_mutation(
+            "employees",
+            {"id": "alice", "name": "Alice", "metadata": {}},
+            [face("alice-1", unit([1, 0, 0, 0]))],
+        )
+        return created, mutation
+
+    manager.run_mutation("employees", add_person, expected_additions=1)
+    matches = manager.search(
+        "employees", unit([1, 0, 0, 0]), limit=1, threshold=0.3
+    )
+    assert matches[0]["person"]["id"] == "alice"
+    assert matches[0]["similarity"] == pytest.approx(1.0)
+    assert backend.created_handles == 2
+    assert backend.active_handles == {2}
+
+    manager.close()
+    assert backend.active_handles == set()
+    assert backend.closed_handles == {1, 2}
diff --git a/server/tests/unit/test_search_reference.py b/server/tests/unit/test_search_reference.py
new file mode 100644
index 000000000..3c1208c2e
--- /dev/null
+++ b/server/tests/unit/test_search_reference.py
@@ -0,0 +1,300 @@
+from __future__ import annotations
+
+import numpy as np
+import pytest
+from insightface_server.search import (
+    IndexRecord,
+    NativeCapabilities,
+    ReferenceSearchIndex,
+    SearchIndexCapacityError,
+)
+from insightface_server.search.reference import (
+    profile_similarity,
+    quantize_bf16_to_fp32,
+    quantize_int8_x736,
+    quantize_int8_x1000,
+)
+
+
+def unit(value: list[float]) -> np.ndarray:
+    vector = np.asarray(value, dtype=np.float32)
+    return vector / np.linalg.norm(vector)
+
+
+@pytest.mark.parametrize(
+    ("flags", "expected"),
+    [
+        (1 << 9, "device_exact"),
+        ((1 << 8) | (1 << 9), "device_exact"),
+        (1 << 8, "host_reference"),
+        (0, "unavailable"),
+    ],
+)
+def test_native_capability_reports_grouped_topk_location(
+    flags: int, expected: str
+) -> None:
+    capabilities = NativeCapabilities(
+        backend="native_cuda",
+        profiles=("fp32_v1",),
+        flags=flags,
+        device=0,
+        compute_capability="12.0",
+        cuda_runtime_version=12090,
+        cuda_driver_version=13000,
+        build_info="test",
+    )
+
+    assert capabilities.grouped_person_topk_mode == expected
+
+
+@pytest.mark.parametrize(
+    "profile",
+    ["fp32_v1", "fp16_v1", "bf16_v1", "int8_x1000_v1", "int8_x736_v1"],
+)
+def test_reference_profiles_return_raw_cosine_and_group_by_person(profile: str) -> None:
+    query = unit([1.0] * 512)
+    alternate = unit([1.0] * 384 + [-1.0] * 128)
+    orthogonal = unit([1.0] * 256 + [-1.0] * 256)
+    index = ReferenceSearchIndex(profile=profile, dimension=512, capacity_rows=4)
+    index.add_batch(
+        [
+            IndexRecord(1, 10, query),
+            IndexRecord(2, 10, alternate),
+            IndexRecord(3, 20, orthogonal),
+        ]
+    )
+
+    hits = index.search_persons(query, 10)
+
+    assert [hit.person_numeric_id for hit in hits] == [10, 20]
+    assert hits[0].vector_id == 1
+    assert hits[0].cosine == pytest.approx(1.0, abs=2e-2)
+    assert hits[1].cosine == pytest.approx(0.0, abs=2e-2)
+
+
+@pytest.mark.parametrize(
+    "profile",
+    ["fp32_v1", "fp16_v1", "bf16_v1", "int8_x1000_v1", "int8_x736_v1"],
+)
+def test_reference_profiles_keep_negative_raw_cosine(profile: str) -> None:
+    positive = unit([1.0] * 512)
+    negative = -positive
+    index = ReferenceSearchIndex(profile=profile, dimension=512, capacity_rows=1)
+    index.add_batch([IndexRecord(1, 1, negative)])
+
+    hit = index.search_persons(positive, 1)[0]
+
+    assert hit.cosine == pytest.approx(-1.0, abs=3e-2)
+
+
+def test_int8_x1000_quantization_rounds_and_saturates_to_int8() -> None:
+    values = np.asarray(
+        [
+            -1.0,
+            -0.1276,
+            -0.1274,
+            -0.12449999898672104,
+            -0.0005,
+            0.0005,
+            0.12449999898672104,
+            0.1266,
+            0.1276,
+            1.0,
+        ],
+        dtype=np.float32,
+    )
+
+    assert quantize_int8_x1000(values).tolist() == [
+        -128,
+        -128,
+        -127,
+        -125,
+        -1,
+        1,
+        125,
+        127,
+        127,
+        127,
+    ]
+
+
+def test_int8_x736_quantization_rounds_and_saturates_to_int8() -> None:
+    scale = np.float32(736.0)
+    values = np.asarray(
+        [
+            -1.0,
+            -128.0 / scale,
+            -127.6 / scale,
+            -127.4 / scale,
+            -0.5 / scale,
+            0.5 / scale,
+            124.49999898672104 / scale,
+            126.6 / scale,
+            127.6 / scale,
+            1.0,
+        ],
+        dtype=np.float32,
+    )
+
+    assert quantize_int8_x736(values).tolist() == [
+        -128,
+        -128,
+        -128,
+        -127,
+        -1,
+        1,
+        125,
+        127,
+        127,
+        127,
+    ]
+
+
+@pytest.mark.parametrize(
+    "profile",
+    ["fp32_v1", "fp16_v1", "bf16_v1", "int8_x1000_v1", "int8_x736_v1"],
+)
+def test_profile_similarity_matches_profile_encoding(profile: str) -> None:
+    left = unit([1.0] * 384 + [-1.0] * 128)
+    right = unit([1.0] * 256 + [-1.0] * 256)
+
+    if profile == "fp16_v1":
+        expected = float(
+            np.dot(
+                left.astype(np.float16).astype(np.float32),
+                right.astype(np.float16).astype(np.float32),
+            )
+        )
+    elif profile == "bf16_v1":
+        expected = float(
+            np.dot(quantize_bf16_to_fp32(left), quantize_bf16_to_fp32(right))
+        )
+    elif profile.startswith("int8_"):
+        scale = 1000 if "x1000" in profile else 736
+        quantize = quantize_int8_x1000 if scale == 1000 else quantize_int8_x736
+        expected = float(
+            np.dot(
+                quantize(left).astype(np.int32),
+                quantize(right).astype(np.int32),
+            )
+        ) / float(scale * scale)
+    else:
+        expected = float(np.dot(left, right))
+
+    assert profile_similarity(profile, left, right) == pytest.approx(expected)
+
+
+def test_profile_similarity_returns_unclipped_int8_raw_score() -> None:
+    value = unit([1.0] * 512)
+
+    score = profile_similarity("int8_x736_v1", value, value)
+
+    assert score == pytest.approx(557_568 / 541_696)
+    assert score > 1.0
+
+
+def test_profile_similarity_validates_profile_shape_and_normalization() -> None:
+    with pytest.raises(ValueError, match="unsupported search profile"):
+        profile_similarity("int8_v0", unit([1.0, 0.0]), unit([1.0, 0.0]))
+    with pytest.raises(ValueError, match="same non-empty shape"):
+        profile_similarity("fp32_v1", unit([1.0, 0.0]), unit([1.0, 0.0, 0.0]))
+    with pytest.raises(ValueError, match="L2-normalized"):
+        profile_similarity(
+            "fp32_v1",
+            np.asarray([2.0, 0.0], dtype=np.float32),
+            unit([1.0, 0.0]),
+        )
+
+
+def test_int8_group_selection_uses_unclipped_accumulator_before_public_score() -> None:
+    first = np.concatenate(
+        [
+            np.full(99, 0.045, dtype=np.float32),
+            np.full(
+                413,
+                np.sqrt((1.0 - 99 * 0.045**2) / 413),
+                dtype=np.float32,
+            ),
+        ]
+    )
+    second = np.concatenate(
+        [
+            np.full(100, 0.045, dtype=np.float32),
+            np.full(
+                412,
+                np.sqrt((1.0 - 100 * 0.045**2) / 412),
+                dtype=np.float32,
+            ),
+        ]
+    )
+    first /= np.linalg.norm(first)
+    second /= np.linalg.norm(second)
+    index = ReferenceSearchIndex(
+        profile="int8_x1000_v1", dimension=512, capacity_rows=2
+    )
+    index.add_batch([IndexRecord(10, 1, first), IndexRecord(20, 1, second)])
+
+    hit = index.search_persons(first, 1)[0]
+
+    # Quantized second·first (1.000087) beats first·first (1.000043). Selection
+    # therefore happens before the public raw-cosine range is clamped.
+    assert hit.vector_id == 20
+    assert hit.cosine == 1.0
+
+
+def test_int8_x736_uses_541696_divisor_and_unclipped_accumulator_order() -> None:
+    first = np.concatenate(
+        [
+            np.full(99, 0.045, dtype=np.float32),
+            np.full(
+                413,
+                np.sqrt((1.0 - 99 * 0.045**2) / 413),
+                dtype=np.float32,
+            ),
+        ]
+    )
+    second = np.concatenate(
+        [
+            np.full(100, 0.045, dtype=np.float32),
+            np.full(
+                412,
+                np.sqrt((1.0 - 100 * 0.045**2) / 412),
+                dtype=np.float32,
+            ),
+        ]
+    )
+    first /= np.linalg.norm(first)
+    second /= np.linalg.norm(second)
+    index = ReferenceSearchIndex(
+        profile="int8_x736_v1", dimension=512, capacity_rows=2
+    )
+    index.add_batch([IndexRecord(10, 1, first), IndexRecord(20, 1, second)])
+
+    hit = index.search_persons(first, 1)[0]
+
+    assert hit.vector_id == 20
+    assert hit.cosine == pytest.approx(530_755 / 541_696)
+
+
+def test_reference_delete_reuses_reserved_slot_and_capacity_is_hard() -> None:
+    index = ReferenceSearchIndex(profile="fp32_v1", dimension=2, capacity_rows=2)
+    index.add_batch(
+        [IndexRecord(1, 1, unit([1, 0])), IndexRecord(2, 2, unit([0, 1]))]
+    )
+    with pytest.raises(SearchIndexCapacityError):
+        index.add_batch([IndexRecord(3, 3, unit([-1, 0]))])
+
+    assert index.remove_batch([1, 99]) == {1}
+    index.add_batch([IndexRecord(3, 3, unit([-1, 0]))])
+
+    stats = index.stats()
+    assert stats.live_rows == 2
+    assert stats.physical_rows == 2
+    assert stats.reallocations == 0
+    assert [hit.vector_id for hit in index.search_persons(unit([-1, 0]), 2)] == [3, 2]
+
+
+def test_reference_rejects_non_normalized_vectors() -> None:
+    index = ReferenceSearchIndex(profile="fp32_v1", dimension=2, capacity_rows=1)
+    with pytest.raises(ValueError, match="L2-normalized"):
+        index.add_batch([IndexRecord(1, 1, np.array([2.0, 0.0], dtype=np.float32))])
diff --git a/server/tests/unit/test_smoke_test.py b/server/tests/unit/test_smoke_test.py
new file mode 100644
index 000000000..0b6816af1
--- /dev/null
+++ b/server/tests/unit/test_smoke_test.py
@@ -0,0 +1,109 @@
+from __future__ import annotations
+
+import pytest
+
+from scripts.smoke_test import (
+    inference_report,
+    multipart_image,
+    redacted_report,
+    require_single_detected_face,
+)
+
+
+def test_multipart_image_uses_a_fixed_filename_and_preserves_bytes(tmp_path) -> None:
+    image = tmp_path / 'unsafe"\r\nname.jpg'
+    image.write_bytes(b"\xff\xd8release-smoke\xff\xd9")
+
+    body, content_type = multipart_image(image, fields={"max_faces": "2"})
+
+    assert content_type.startswith("multipart/form-data; boundary=")
+    assert b'name="max_faces"' in body
+    assert b"\r\n\r\n2\r\n" in body
+    assert b'filename="release-smoke-image"' in body
+    assert image.name.encode() not in body
+    assert image.read_bytes() in body
+
+
+def test_inference_report_requires_a_normalized_finite_embedding() -> None:
+    report = inference_report(
+        {
+            "faces": [
+                {
+                    "bbox": {"pixels": {"x": 1, "y": 2, "width": 3, "height": 4}},
+                    "detection_score": 0.9,
+                    "embedding": [0.6, 0.8],
+                }
+            ]
+        }
+    )
+
+    assert report["embedding_dimension"] == 2
+    assert report["embedding_norm"] == pytest.approx(1.0)
+    assert report["face_count"] == 1
+
+    with pytest.raises(SystemExit, match="not normalized"):
+        inference_report(
+            {
+                "faces": [
+                    {
+                        "bbox": {},
+                        "detection_score": 0.9,
+                        "embedding": [1.0, 1.0],
+                    }
+                ]
+            }
+        )
+
+
+def test_release_image_must_detect_exactly_one_face() -> None:
+    require_single_detected_face({"faces": [{"bbox": {}}]})
+
+    with pytest.raises(SystemExit, match="found 0"):
+        require_single_detected_face({"faces": []})
+
+    with pytest.raises(SystemExit, match="found 2"):
+        require_single_detected_face({"faces": [{"bbox": {}}, {"bbox": {}}]})
+
+    with pytest.raises(SystemExit, match="exactly one face"):
+        inference_report(
+            {
+                "faces": [
+                    {
+                        "bbox": {},
+                        "detection_score": 0.9,
+                        "embedding": [0.6, 0.8],
+                    },
+                    {
+                        "bbox": {},
+                        "detection_score": 0.8,
+                        "embedding": [0.6, 0.8],
+                    },
+                ]
+            }
+        )
+
+
+def test_redacted_report_removes_biometric_material() -> None:
+    report = {
+        "health": {"status": "ready"},
+        "inference": {
+            "face_count": 1,
+            "bbox": {"pixels": {"x": 1, "y": 2, "width": 3, "height": 4}},
+            "detection_score": 0.9,
+            "embedding_dimension": 2,
+            "embedding_norm": 1.0,
+            "embedding_sha256": "stable-feature-fingerprint",
+            "embedding": [0.6, 0.8],
+        },
+    }
+
+    redacted = redacted_report(report)
+
+    assert redacted == {
+        "health": {"status": "ready"},
+        "inference": {
+            "face_count": 1,
+            "embedding_dimension": 2,
+            "embedding_norm": 1.0,
+        },
+    }
diff --git a/server/tests/unit/test_storage_core.py b/server/tests/unit/test_storage_core.py
new file mode 100644
index 000000000..29adbc447
--- /dev/null
+++ b/server/tests/unit/test_storage_core.py
@@ -0,0 +1,813 @@
+from __future__ import annotations
+
+import shutil
+import sqlite3
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pytest
+from insightface_server.errors import ApiError
+from insightface_server.models import (
+    EMBEDDING_CONTRACT_PREFIX,
+    embedding_contract_id,
+)
+from insightface_server.storage import (
+    CollectionCapacityExceeded,
+    CursorCodec,
+    Database,
+    FaceCropStore,
+    MaxFacesPerPersonExceeded,
+    Repository,
+)
+
+
+def collection_item(collection_id: str = "employees") -> dict[str, object]:
+    return {
+        "id": collection_id,
+        "name": "Employees",
+        "description": "",
+        "default_threshold": 0.68,
+        "metadata": {},
+        "model_id": "recognition",
+        "model_version": "1",
+        "model_digest": "a" * 64,
+        "embedding_dimension": 4,
+        "preprocessing_version": "1",
+    }
+
+
+def face_item(face_id: str, embedding: np.ndarray | None = None) -> dict[str, object]:
+    return {
+        "id": face_id,
+        "embedding": (
+            embedding
+            if embedding is not None
+            else np.asarray([0.5, -0.5, 0.5, -0.5], dtype=np.float32)
+        ),
+        "embedding_dimension": 4,
+        "bounding_box": {"pixels": {"x": 1, "y": 2, "width": 3, "height": 4}},
+        "landmarks": None,
+        "detection_score": 0.9,
+        "quality": {"score": 0.8},
+        "model_id": "recognition",
+        "model_version": "1",
+        "model_digest": "a" * 64,
+        "preprocessing_version": "1",
+        "crop_path": None,
+        "created_at": "2026-01-01T00:00:00.000Z",
+    }
+
+
+@pytest.fixture
+def repository(tmp_path: Path) -> Repository:
+    migrations = Path(__file__).resolve().parents[2] / "migrations"
+    database = Database(tmp_path / "server.db", migrations)
+    database.initialize()
+    return Repository(database)
+
+
+def test_cursor_secret_persists_and_tokens_are_scoped(tmp_path: Path) -> None:
+    path = tmp_path / "cursor.key"
+    first = CursorCodec(path)
+    token = first.encode("people", "person-9")
+
+    assert CursorCodec(path).decode(token, "people") == "person-9"
+    assert path.stat().st_mode & 0o777 == 0o600
+    with pytest.raises(ApiError, match="cursor is invalid") as wrong_scope:
+        first.decode(token, "collections")
+    assert wrong_scope.value.code == "invalid_cursor"
+    with pytest.raises(ApiError) as tampered:
+        first.decode(token + "A", "people")
+    assert tampered.value.code == "invalid_cursor"
+
+
+def test_api_key_sync_creates_noops_and_atomically_rotates(repository: Repository) -> None:
+    assert repository.sync_api_key("secret-key") == "created"
+    with repository.database.read() as connection:
+        original = connection.execute(
+            "SELECT id,salt,digest,active,created_at FROM api_keys"
+        ).fetchone()
+
+    assert original is not None
+    assert repository.sync_api_key("secret-key") == "unchanged"
+    with repository.database.read() as connection:
+        unchanged = connection.execute(
+            "SELECT id,salt,digest,active,created_at FROM api_keys"
+        ).fetchone()
+    assert unchanged is not None
+    assert tuple(unchanged) == tuple(original)
+
+    assert repository.sync_api_key("replacement-key") == "rotated"
+    assert repository.has_api_keys()
+    assert repository.verify_api_key("replacement-key")
+    assert not repository.verify_api_key("secret-key")
+    assert not repository.verify_api_key("wrong")
+    with repository.database.read() as connection:
+        rows = connection.execute(
+            "SELECT label,salt,digest,active FROM api_keys ORDER BY active,id"
+        ).fetchall()
+    assert len(rows) == 2
+    assert [row["active"] for row in rows] == [0, 1]
+    assert rows[1]["label"] == "startup"
+    for row in rows:
+        assert row["salt"] not in {b"secret-key", b"replacement-key"}
+        assert row["digest"] not in {b"secret-key", b"replacement-key"}
+
+
+def test_api_key_sync_rejects_an_empty_key(repository: Repository) -> None:
+    with pytest.raises(ValueError, match="must not be empty"):
+        repository.sync_api_key("")
+    assert not repository.has_api_keys()
+
+
+def test_repository_round_trips_embedding_and_optional_landmarks(
+    repository: Repository,
+) -> None:
+    repository.create_collection(collection_item())
+    embedding = np.asarray([0.5, -0.5, 0.5, -0.5], dtype=np.float32)
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [
+            {
+                "id": "face-1",
+                "embedding": embedding,
+                "embedding_dimension": 4,
+                "bounding_box": {"pixels": {"x": 1, "y": 2, "width": 3, "height": 4}},
+                "landmarks": None,
+                "detection_score": 0.9,
+                "quality": {"score": 0.8},
+                "model_id": "recognition",
+                "model_version": "1",
+                "model_digest": "a" * 64,
+                "preprocessing_version": "1",
+                "crop_path": None,
+                "created_at": "2026-01-01T00:00:00.000Z",
+            }
+        ],
+    )
+
+    private = repository.get_face("employees", "alice", "face-1")
+    public = repository.list_faces("employees", "alice", "", 10)[0]
+    enrollment_embeddings = repository.list_face_embeddings("employees", "alice")
+
+    assert private is not None
+    np.testing.assert_array_equal(private["embedding"], embedding)
+    assert private["landmarks"] is None
+    assert private["embedding_source"] == "server"
+    assert private["embedding_contract_id"] is None
+    assert "landmarks_json" not in private
+    assert "embedding" not in public
+    assert "collection_id" not in public
+    assert len(enrollment_embeddings) == 1
+    np.testing.assert_array_equal(enrollment_embeddings[0], embedding)
+    enrollment_embeddings[0][0] = 0.0
+    np.testing.assert_array_equal(
+        repository.list_face_embeddings("employees", "alice")[0], embedding
+    )
+
+
+def test_embedding_contract_id_is_stable_and_collection_derived(
+    repository: Repository,
+) -> None:
+    contract_id = embedding_contract_id(
+        model_id="recognition",
+        model_version="1",
+        model_digest="A" * 64,
+        embedding_dimension=4,
+        preprocessing_version="1",
+    )
+    created = repository.create_collection(collection_item())
+
+    assert contract_id == (
+        "ifsemb-v1-sha256:"
+        "0473aa8e9422b084c939259ce447572a82ff9d23dc3a341bd22cf4806b4494b5"
+    )
+    assert contract_id.startswith(EMBEDDING_CONTRACT_PREFIX)
+    assert created["embedding_contract_id"] == contract_id
+    assert repository.get_collection("employees")["embedding_contract_id"] == contract_id  # type: ignore[index]
+    listed = repository.list_collections("", 10)
+    assert listed[0]["embedding_contract_id"] == contract_id
+
+
+def test_repository_persists_external_trusted_embedding_provenance(
+    repository: Repository,
+) -> None:
+    collection = repository.create_collection(collection_item())
+    face = face_item("trusted-face")
+    face.update(
+        embedding_source="external_trusted",
+        embedding_contract_id=collection["embedding_contract_id"],
+    )
+    repository.create_person_with_faces(
+        "employees", {"id": "alice", "metadata": {}}, [face]
+    )
+
+    private = repository.get_face("employees", "alice", "trusted-face")
+    public = repository.list_faces("employees", "alice", "", 10)[0]
+    indexed = repository.all_faces("employees")[0]
+    for stored in (private, public, indexed):
+        assert stored is not None
+        assert stored["embedding_source"] == "external_trusted"
+        assert stored["embedding_contract_id"] == collection["embedding_contract_id"]
+    assert "embedding" not in public
+
+
+def test_migration_is_applied_once_and_ignores_non_migration_files(tmp_path: Path) -> None:
+    migrations = tmp_path / "migrations"
+    migrations.mkdir()
+    (migrations / "0001_first.sql").write_text(
+        "CREATE TABLE values_table(id INTEGER PRIMARY KEY);", encoding="utf-8"
+    )
+    (migrations / "README.sql").write_text("this is not SQL", encoding="utf-8")
+    database = Database(tmp_path / "database.db", migrations)
+
+    database.initialize()
+    database.initialize()
+
+    assert database.status()["migration_count"] == 1
+    with database.read() as connection:
+        assert (
+            connection.execute(
+                "SELECT count(*) FROM schema_migrations WHERE version='0001_first.sql'"
+            ).fetchone()[0]
+            == 1
+        )
+
+
+def test_failed_migration_rolls_back_schema_and_can_be_repaired(tmp_path: Path) -> None:
+    migrations = tmp_path / "migrations"
+    migrations.mkdir()
+    migration = migrations / "0001_atomic.sql"
+    migration.write_text(
+        "CREATE TABLE should_rollback(id INTEGER); BROKEN SQL;", encoding="utf-8"
+    )
+    database = Database(tmp_path / "atomic.db", migrations)
+
+    with pytest.raises(sqlite3.DatabaseError):
+        database.initialize()
+    with database.read() as connection:
+        assert connection.execute(
+            "SELECT count(*) FROM sqlite_master WHERE name='should_rollback'"
+        ).fetchone()[0] == 0
+        assert connection.execute("SELECT count(*) FROM schema_migrations").fetchone()[0] == 0
+
+    migration.write_text(
+        "CREATE TABLE should_rollback(id INTEGER);", encoding="utf-8"
+    )
+    database.initialize()
+    assert database.status()["migration_count"] == 1
+
+
+def test_repository_enforces_external_id_uniqueness(repository: Repository) -> None:
+    repository.create_collection(collection_item())
+    repository.create_person_with_faces(
+        "employees",
+        {"id": "one", "external_id": "same", "metadata": {}},
+        [],
+    )
+
+    with pytest.raises(sqlite3.IntegrityError):
+        repository.create_person_with_faces(
+            "employees",
+            {"id": "two", "external_id": "same", "metadata": {}},
+            [],
+        )
+
+
+def test_native_search_migration_backfills_legacy_rows(tmp_path: Path) -> None:
+    source_migrations = Path(__file__).resolve().parents[2] / "migrations"
+    migrations = tmp_path / "migrations"
+    migrations.mkdir()
+    shutil.copy(source_migrations / "0001_initial.sql", migrations)
+    database = Database(tmp_path / "legacy.db", migrations)
+    database.initialize()
+
+    now = "2026-01-01T00:00:00.000Z"
+    embedding = np.asarray([1.0, 0.0, 0.0, 0.0], dtype=np.float32)
+    with database.write() as connection:
+        connection.execute(
+            """INSERT INTO collections(
+                   id,name,description,default_threshold,model_id,model_version,
+                   model_digest,embedding_dimension,preprocessing_version,
+                   metadata_json,created_at,updated_at
+               ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?)""",
+            ("legacy", "Legacy", "", 0.3, "recognition", "1", "a" * 64, 4, "1", "{}", now, now),
+        )
+        connection.execute(
+            """INSERT INTO persons(
+                   collection_id,id,name,external_id,metadata_json,created_at,updated_at
+               ) VALUES(?,?,?,?,?,?,?)""",
+            ("legacy", "alice", "Alice", None, "{}", now, now),
+        )
+        connection.execute(
+            """INSERT INTO face_samples(
+                   id,collection_id,person_id,embedding,embedding_dimension,
+                   bounding_box_json,landmarks_json,detection_score,quality_json,
+                   model_id,model_version,model_digest,preprocessing_version,
+                   crop_path,created_at
+               ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
+            (
+                "legacy-face",
+                "legacy",
+                "alice",
+                embedding.tobytes(),
+                4,
+                "{}",
+                None,
+                0.9,
+                "{}",
+                "recognition",
+                "1",
+                "a" * 64,
+                "1",
+                None,
+                now,
+            ),
+        )
+
+    shutil.copy(source_migrations / "0002_native_search.sql", migrations)
+    database.initialize()
+    repository = Repository(database)
+
+    collection = repository.get_collection("legacy")
+    assert collection is not None
+    assert collection["search_profile"] == "fp32_v1"
+    assert collection["capacity_rows"] == 100000
+    assert collection["max_faces_per_person"] == 20
+    assert collection["load_policy"] == "lazy"
+    assert collection["search_revision"] == 0
+    batch = next(repository.iter_index_faces("legacy", batch_size=1))
+    assert batch.revision == 0
+    assert batch.max_seq == 0
+    assert len(batch.faces) == 1
+    assert batch.faces[0].vector_id > 0
+    assert batch.faces[0].person_numeric_id > 0
+    np.testing.assert_array_equal(batch.faces[0].embedding, embedding)
+
+
+def test_int8_x736_migration_preserves_x1000_collection_and_child_rows(
+    tmp_path: Path,
+) -> None:
+    source_migrations = Path(__file__).resolve().parents[2] / "migrations"
+    migrations = tmp_path / "migrations"
+    migrations.mkdir()
+    shutil.copy(source_migrations / "0001_initial.sql", migrations)
+    shutil.copy(source_migrations / "0002_native_search.sql", migrations)
+    # Repository always targets the current storage schema; stage migration 0004
+    # through 0006 here while deliberately withholding 0003 to exercise its x1000
+    # backfill.
+    shutil.copy(source_migrations / "0004_collection_face_crops.sql", migrations)
+    shutil.copy(source_migrations / "0005_external_trusted_embeddings.sql", migrations)
+    shutil.copy(source_migrations / "0006_collection_detection_profiles.sql", migrations)
+    database = Database(tmp_path / "x1000.db", migrations)
+    database.initialize()
+    repository = Repository(database)
+    item = collection_item("legacy-int8")
+    item.update(
+        search_profile="int8_x1000_v1",
+        capacity_rows=10,
+        max_faces_per_person=5,
+        load_policy="eager",
+    )
+    repository.create_collection(item)
+    repository.create_person_with_faces(
+        "legacy-int8",
+        {"id": "alice", "name": "Alice", "metadata": {}},
+        [face_item("legacy-int8-face")],
+    )
+    before = next(repository.iter_index_faces("legacy-int8", batch_size=10)).faces[0]
+
+    shutil.copy(source_migrations / "0003_int8_x736.sql", migrations)
+    shutil.copy(source_migrations / "0004_collection_face_crops.sql", migrations)
+    shutil.copy(source_migrations / "0005_external_trusted_embeddings.sql", migrations)
+    database.initialize()
+    migrated = Repository(database)
+
+    collection = migrated.get_collection("legacy-int8")
+    assert collection is not None
+    assert collection["search_profile"] == "int8_x1000_v1"
+    assert collection["person_count"] == 1
+    assert collection["face_count"] == 1
+    assert migrated.get_person("legacy-int8", "alice") is not None
+    face = migrated.get_face("legacy-int8", "alice", "legacy-int8-face")
+    assert face is not None
+    after = next(migrated.iter_index_faces("legacy-int8", batch_size=10)).faces[0]
+    assert after.vector_id == before.vector_id
+    assert after.person_numeric_id == before.person_numeric_id
+    np.testing.assert_array_equal(after.embedding, before.embedding)
+    assert database.status()["migration_count"] == 6
+    with database.read() as connection:
+        assert connection.execute("PRAGMA foreign_key_check").fetchall() == []
+
+    new_item = collection_item("new-int8")
+    new_item["search_profile"] = "int8_x736_v1"
+    assert migrated.create_collection(new_item)["search_profile"] == "int8_x736_v1"
+
+
+def test_collection_update_returns_from_its_write_transaction(
+    repository: Repository, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    repository.create_collection(collection_item())
+    original_get_collection = repository.get_collection
+
+    def forbidden_post_commit_read(collection_id: str):
+        raise AssertionError(f"unexpected post-transaction read for {collection_id}")
+
+    monkeypatch.setattr(repository, "get_collection", forbidden_post_commit_read)
+    updated = repository.update_collection(
+        "employees", {"name": "Updated", "capacity_rows": 12}
+    )
+
+    assert updated is not None
+    assert updated["name"] == "Updated"
+    assert updated["capacity_rows"] == 12
+    monkeypatch.setattr(repository, "get_collection", original_get_collection)
+    persisted = repository.get_collection("employees")
+    assert persisted is not None
+    assert persisted["name"] == "Updated"
+    assert persisted["capacity_rows"] == 12
+
+
+def test_face_mutation_is_durable_and_immediately_replayable(repository: Repository) -> None:
+    item = collection_item()
+    item.update(
+        search_profile="bf16_v1",
+        capacity_rows=10,
+        max_faces_per_person=4,
+        load_policy="eager",
+    )
+    repository.create_collection(item)
+    person, mutation = repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "alice", "name": "Alice", "metadata": {"team": "infra"}},
+        [face_item("face-1"), face_item("face-2")],
+    )
+
+    assert person["face_count"] == 2
+    assert mutation is not None
+    assert mutation.revision == 1
+    assert mutation.max_seq > 0
+    assert [change.operation for change in mutation.changes] == ["add", "add"]
+    assert all(change.revision == 1 for change in mutation.changes)
+    assert mutation.changes[0].seq < mutation.changes[1].seq
+    assert len({change.vector_id for change in mutation.changes}) == 2
+    assert len({change.person_numeric_id for change in mutation.changes}) == 1
+
+    state = repository.get_search_state("employees")
+    assert state is not None
+    assert state.revision == mutation.revision
+    assert state.max_seq == mutation.max_seq
+    batches = list(repository.iter_index_faces("employees", batch_size=1))
+    assert {batch.revision for batch in batches} == {mutation.revision}
+    assert {batch.max_seq for batch in batches} == {mutation.max_seq}
+    assert repository.get_search_changes("employees") == list(mutation.changes)
+
+    vector_ids = [int(change.vector_id) for change in mutation.changes if change.vector_id]
+    native_rows = repository.get_index_faces_by_vector_ids("employees", vector_ids)
+    assert set(native_rows) == set(vector_ids)
+    hydrated = repository.get_index_face_mappings(
+        "employees", list(reversed(vector_ids)), include_embedding=True
+    )
+    assert set(hydrated) == set(vector_ids)
+    assert {value["person"]["id"] for value in hydrated.values()} == {"alice"}
+    assert all("embedding" in value for value in hydrated.values())
+
+    next_mutation = repository.add_faces_with_mutation(
+        "employees", "alice", [face_item("face-3")]
+    )
+    assert next_mutation is not None
+    assert next_mutation.revision == 2
+    assert next_mutation.max_seq > mutation.max_seq
+    assert repository.get_search_changes(
+        "employees", after_seq=mutation.max_seq
+    ) == list(next_mutation.changes)
+
+
+def test_search_change_acknowledgement_is_revision_bounded(repository: Repository) -> None:
+    repository.create_collection(collection_item())
+    _, first = repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "alice", "metadata": {}},
+        [face_item("face-1")],
+    )
+    second = repository.add_faces_with_mutation(
+        "employees", "alice", [face_item("face-2")]
+    )
+    assert first is not None and second is not None
+
+    assert repository.acknowledge_search_changes("employees", first.max_seq) == 1
+    remaining = repository.get_search_changes("employees")
+    assert [change.revision for change in remaining] == [second.revision]
+    assert repository.acknowledge_search_changes("employees", first.max_seq) == 0
+    assert repository.reset_search_changes_before_full_rebuild() == 1
+    assert repository.get_search_changes("employees") == []
+
+    with pytest.raises(ValueError, match="must not be negative"):
+        repository.acknowledge_search_changes("employees", -1)
+
+
+def test_search_change_sequence_prevents_same_collection_id_aba_ack(
+    repository: Repository,
+) -> None:
+    repository.create_collection(collection_item())
+    _, added = repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "old-person", "metadata": {}},
+        [face_item("old-face")],
+    )
+    assert added is not None
+    status, _, dropped = repository.delete_collection_with_mutation(
+        "employees", force=True
+    )
+    assert status == "deleted" and dropped is not None
+
+    repository.create_collection(collection_item())
+    _, recreated = repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "new-person", "metadata": {}},
+        [face_item("new-face")],
+    )
+    assert recreated is not None and recreated.revision == 1
+    assert recreated.max_seq > dropped.max_seq
+
+    repository.acknowledge_search_changes("employees", dropped.max_seq)
+
+    assert repository.get_search_changes("employees") == list(recreated.changes)
+
+
+def test_face_limits_are_checked_inside_the_write_transaction(repository: Repository) -> None:
+    item = collection_item()
+    item.update(capacity_rows=2, max_faces_per_person=1)
+    repository.create_collection(item)
+    repository.create_person_with_faces_with_mutation(
+        "employees", {"id": "alice", "metadata": {}}, [face_item("face-1")]
+    )
+
+    with pytest.raises(MaxFacesPerPersonExceeded):
+        repository.add_faces_with_mutation(
+            "employees", "alice", [face_item("face-2")]
+        )
+    assert repository.get_person("employees", "alice")["face_count"] == 1  # type: ignore[index]
+    assert repository.get_search_revision("employees") == 1
+    assert len(repository.get_search_changes("employees")) == 1
+
+    repository.create_person_with_faces_with_mutation(
+        "employees", {"id": "bob", "metadata": {}}, [face_item("face-2")]
+    )
+    with pytest.raises(CollectionCapacityExceeded):
+        repository.create_person_with_faces_with_mutation(
+            "employees", {"id": "carol", "metadata": {}}, [face_item("face-3")]
+        )
+    assert repository.get_person("employees", "carol") is None
+    assert repository.get_search_revision("employees") == 2
+
+    with pytest.raises(CollectionCapacityExceeded):
+        repository.update_collection("employees", {"capacity_rows": 1})
+    with pytest.raises(MaxFacesPerPersonExceeded):
+        repository.update_collection("employees", {"max_faces_per_person": 0})
+    collection = repository.get_collection("employees")
+    assert collection is not None
+    assert collection["capacity_rows"] == 2
+    assert collection["max_faces_per_person"] == 1
+
+
+def test_cascading_deletes_produce_durable_search_changes(repository: Repository) -> None:
+    repository.create_collection(collection_item())
+    _, added = repository.create_person_with_faces_with_mutation(
+        "employees",
+        {"id": "alice", "metadata": {}},
+        [face_item("face-1"), face_item("face-2"), face_item("face-3")],
+    )
+    assert added is not None
+    add_ids = [change.vector_id for change in added.changes]
+
+    deleted, _, face_delete = repository.delete_face_with_mutation(
+        "employees", "alice", "face-1"
+    )
+    assert deleted and face_delete is not None
+    assert face_delete.revision == 2
+    assert face_delete.changes[0].operation == "delete"
+    assert face_delete.changes[0].vector_id == add_ids[0]
+
+    deleted, _, person_delete = repository.delete_person_with_mutation(
+        "employees", "alice"
+    )
+    assert deleted and person_delete is not None
+    assert person_delete.revision == 3
+    assert [change.vector_id for change in person_delete.changes] == add_ids[1:]
+    assert repository.get_index_faces_by_vector_ids(
+        "employees", [int(value) for value in add_ids if value is not None]
+    ) == {}
+
+    status, _, collection_delete = repository.delete_collection_with_mutation(
+        "employees", force=True
+    )
+    assert status == "deleted"
+    assert collection_delete is not None
+    assert collection_delete.revision == 4
+    assert [change.operation for change in collection_delete.changes] == ["clear"]
+    assert repository.get_collection("employees") is None
+    assert repository.get_search_revision("employees") is None
+    changes = repository.get_search_changes("employees")
+    assert [change.operation for change in changes] == [
+        "add",
+        "add",
+        "add",
+        "delete",
+        "delete",
+        "delete",
+        "clear",
+    ]
+
+
+def test_collection_delete_uses_one_clear_event_for_all_faces(repository: Repository) -> None:
+    repository.create_collection(collection_item("large"))
+    _, added = repository.create_person_with_faces_with_mutation(
+        "large",
+        {"id": "alice", "metadata": {}},
+        [face_item("large-face-1"), face_item("large-face-2")],
+    )
+    assert added is not None
+
+    status, _, deleted = repository.delete_collection_with_mutation("large", force=True)
+
+    assert status == "deleted"
+    assert deleted is not None
+    assert deleted.revision == 2
+    assert [change.operation for change in deleted.changes] == ["clear"]
+    assert [change.operation for change in repository.get_search_changes("large")] == [
+        "add",
+        "add",
+        "clear",
+    ]
+
+
+def test_collection_crop_storage_defaults_off_and_can_be_updated(
+    repository: Repository,
+) -> None:
+    created = repository.create_collection(collection_item())
+
+    assert created["save_face_crops"] is False
+    updated = repository.update_collection("employees", {"save_face_crops": True})
+    assert updated is not None
+    assert updated["save_face_crops"] is True
+    with repository.database.read() as connection:
+        stored = connection.execute(
+            "SELECT save_face_crops FROM collections WHERE id='employees'"
+        ).fetchone()[0]
+    assert stored == 1
+
+
+def test_crop_blob_is_only_loaded_by_dedicated_repository_method(
+    repository: Repository,
+) -> None:
+    item = collection_item()
+    item["save_face_crops"] = True
+    repository.create_collection(item)
+    crop_bytes = b"\xff\xd8private-jpeg\xff\xd9"
+    face = face_item("face-crop")
+    face.update(crop_image=crop_bytes, crop_media_type="image/jpeg")
+    repository.create_person_with_faces(
+        "employees", {"id": "alice", "metadata": {}}, [face]
+    )
+
+    private = repository.get_face("employees", "alice", "face-crop")
+    listed = repository.list_faces("employees", "alice", "", 10)[0]
+    loaded_for_index = repository.all_faces("employees")[0]
+    crop = repository.get_face_crop("employees", "alice", "face-crop")
+
+    for ordinary in (private, listed, loaded_for_index):
+        assert ordinary is not None
+        assert ordinary["has_crop"] is True
+        assert "crop_image" not in ordinary
+        assert "crop_media_type" not in ordinary
+        assert "crop_path" not in ordinary
+    assert crop == {
+        "bytes": crop_bytes,
+        "media_type": "image/jpeg",
+        "crop_path": None,
+    }
+
+
+def test_get_face_crop_supports_legacy_disk_paths(repository: Repository) -> None:
+    repository.create_collection(collection_item())
+    face = face_item("legacy-crop")
+    face["crop_path"] = "face-crops/legacy/legacy-crop.jpg"
+    repository.create_person_with_faces(
+        "employees", {"id": "alice", "metadata": {}}, [face]
+    )
+
+    public = repository.list_faces("employees", "alice", "", 10)[0]
+    crop = repository.get_face_crop("employees", "alice", "legacy-crop")
+
+    assert public["has_crop"] is True
+    assert "crop_path" not in public
+    assert crop == {
+        "bytes": None,
+        "media_type": "image/jpeg",
+        "crop_path": "face-crops/legacy/legacy-crop.jpg",
+    }
+
+
+def test_collection_face_crop_migration_preserves_legacy_rows(tmp_path: Path) -> None:
+    source_migrations = Path(__file__).resolve().parents[2] / "migrations"
+    migrations = tmp_path / "migrations"
+    migrations.mkdir()
+    for name in (
+        "0001_initial.sql",
+        "0002_native_search.sql",
+        "0003_int8_x736.sql",
+    ):
+        shutil.copy(source_migrations / name, migrations)
+    database = Database(tmp_path / "legacy-crops.db", migrations)
+    database.initialize()
+    now = "2026-01-01T00:00:00.000Z"
+    embedding = np.asarray([0.5, -0.5, 0.5, -0.5], dtype=np.float32)
+    with database.write() as connection:
+        connection.execute(
+            """INSERT INTO collections(
+                   id,name,description,default_threshold,model_id,model_version,
+                   model_digest,embedding_dimension,preprocessing_version,
+                   metadata_json,created_at,updated_at
+               ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?)""",
+            (
+                "employees",
+                "Employees",
+                "",
+                0.68,
+                "recognition",
+                "1",
+                "a" * 64,
+                4,
+                "1",
+                "{}",
+                now,
+                now,
+            ),
+        )
+        connection.execute(
+            """INSERT INTO persons(
+                   collection_id,id,name,external_id,metadata_json,created_at,updated_at
+               ) VALUES(?,?,?,?,?,?,?)""",
+            ("employees", "alice", "Alice", None, "{}", now, now),
+        )
+        connection.execute(
+            """INSERT INTO face_samples(
+                   id,collection_id,person_id,embedding,embedding_dimension,
+                   bounding_box_json,landmarks_json,detection_score,quality_json,
+                   model_id,model_version,model_digest,preprocessing_version,
+                   crop_path,created_at
+               ) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
+            (
+                "face-1",
+                "employees",
+                "alice",
+                embedding.tobytes(),
+                4,
+                "{}",
+                None,
+                0.9,
+                "{}",
+                "recognition",
+                "1",
+                "a" * 64,
+                "1",
+                None,
+                now,
+            ),
+        )
+
+    shutil.copy(source_migrations / "0004_collection_face_crops.sql", migrations)
+    shutil.copy(source_migrations / "0005_external_trusted_embeddings.sql", migrations)
+    database.initialize()
+    migrated = Repository(database)
+
+    collection = migrated.get_collection("employees")
+    face = migrated.get_face("employees", "alice", "face-1")
+    assert collection is not None and collection["save_face_crops"] is False
+    assert face is not None and face["has_crop"] is False
+    assert face["embedding_source"] == "server"
+    assert face["embedding_contract_id"] is None
+    assert migrated.get_face_crop("employees", "alice", "face-1") is None
+    assert database.status()["migration_count"] == 5
+    with database.read() as connection:
+        assert connection.execute("PRAGMA foreign_key_check").fetchall() == []
+
+
+def test_face_crop_store_encodes_and_safely_reads_legacy_jpeg(tmp_path: Path) -> None:
+    pixels = np.zeros((40, 50, 3), dtype=np.uint8)
+    pixels[5:35, 10:45] = (20, 120, 240)
+    encoded = FaceCropStore.encode(pixels, (10.0, 5.0, 45.0, 35.0))
+
+    decoded = cv2.imdecode(np.frombuffer(encoded, dtype=np.uint8), cv2.IMREAD_COLOR)
+    assert decoded.shape == (112, 112, 3)
+
+    store = FaceCropStore(tmp_path, enabled=True)
+    path = store.save("employees", "face-1", pixels, (10.0, 5.0, 45.0, 35.0))
+    assert path is not None
+    assert store.read(path) == encoded
+    assert store.read("face-crops/missing.jpg") is None
+    with pytest.raises(RuntimeError, match="outside managed storage"):
+        store.read("../secret")
diff --git a/src/age_iter.py b/src/age_iter.py
deleted file mode 100644
index 114c66651..000000000
--- a/src/age_iter.py
+++ /dev/null
@@ -1,269 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-
-import mxnet as mx
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_preprocess
-import multiprocessing
-
-logger = logging.getLogger()
-
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        if path_imgrec:
-            logging.info('loading recordio %s...',
-                         path_imgrec)
-            path_imgidx = path_imgrec[0:-4]+".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-            s = self.imgrec.read_idx(0)
-            header, _ = recordio.unpack(s)
-            self.imgidx = list(self.imgrec.keys)
-            if shuffle:
-              self.seq = self.imgidx
-              self.oseq = self.imgidx
-              print(len(self.seq))
-            else:
-              self.seq = None
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        self.provide_label = [(label_name, (batch_size,102))]
-        #print(self.provide_label[0][1])
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.shuffle:
-          random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        """Helper function for reading in next sample."""
-        #set total batch size, for example, 1800, and maximum size for each people, for example 45
-        if self.seq is not None:
-          while True:
-            if self.cur >= len(self.seq):
-                raise StopIteration
-            idx = self.seq[self.cur]
-            self.cur += 1
-            if self.imgrec is not None:
-              s = self.imgrec.read_idx(idx)
-              header, img = recordio.unpack(s)
-              label = header.label
-              return label, img, None, None
-            else:
-              label, fname, bbox, landmark = self.imglist[idx]
-              return label, self.read_image(fname), bbox, landmark
-        else:
-            s = self.imgrec.read()
-            if s is None:
-                raise StopIteration
-            header, img = recordio.unpack(s)
-            return header.label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = np.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      random.shuffle(augs)
-      for aug in augs:
-        #print(img.shape)
-        img = aug(img, x)
-        #print(img.shape)
-      return img
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in xrange(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                #if label[1]>=0.0 or label[2]>=0.0:
-                #  print(label[0:10])
-                _data = self.imdecode(s)
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.nd_mean is not None:
-                    _data = _data.astype('float32')
-                    _data -= self.nd_mean
-                    _data *= 0.0078125
-                if self.cutoff>0:
-                  centerh = random.randint(0, _data.shape[0]-1)
-                  centerw = random.randint(0, _data.shape[1]-1)
-                  half = self.cutoff//2
-                  starth = max(0, centerh-half)
-                  endh = min(_data.shape[0], centerh+half)
-                  startw = max(0, centerw-half)
-                  endw = min(_data.shape[1], centerw+half)
-                  _data = _data.astype('float32')
-                  #print(starth, endh, startw, endw, _data.shape)
-                  _data[starth:endh, startw:endw, :] = 127.5
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        return io.DataBatch([batch_data], [batch_label], batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/src/align/align_celeb.py b/src/align/align_celeb.py
deleted file mode 100644
index ae83b06ce..000000000
--- a/src/align/align_celeb.py
+++ /dev/null
@@ -1,225 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import cv2
-import argparse
-import tensorflow as tf
-import numpy as np
-import base64
-#import facenet
-import detect_face
-from easydict import EasyDict as edict
-import random
-from time import sleep
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-import face_preprocess
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-
-def main(args):
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    datamap = {}
-    pp = 0
-    datasize = 0
-    verr = 0
-    for line in open(args.input_dir+"_clean_list.txt", 'r'):
-      pp+=1
-      if pp%10000==0:
-        print('loading list', pp)
-      line = line.strip()[2:]
-      if not line.startswith('m.'):
-        continue
-      vec = line.split('/')
-      assert len(vec)==2
-      #print(line)
-      person = vec[0]
-      img = vec[1]
-      try:
-        img_id = int(img.split('.')[0])
-      except ValueError:
-        #print('value error', line)
-        verr+=1
-        continue
-      if not person in datamap:
-        labelid = len(datamap)
-        datamap[person] = [labelid, {img_id : 1}]
-      else:
-        datamap[person][1][img_id] = 1
-      datasize+=1
-
-    print('dataset size', args.name, datasize)
-    print('dataset err', verr)
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        sess = tf.Session()
-        with sess.as_default():
-            pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-    
-    minsize = 100 # minimum size of face
-    #threshold = [ 0.6, 0.7, 0.7 ]  # three steps's threshold
-    threshold = [ 0.6, 0.6, 0.3 ]  # three steps's threshold
-    factor = 0.709 # scale factor
-
-    print(minsize)
-    print(threshold)
-    print(factor)
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    #random_key = np.random.randint(0, high=99999)
-    #bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    output_filename = os.path.join(output_dir, 'faceinsight_align_%s.lst' % args.name)
-    
-    with open(output_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof_successfully_aligned = 0
-        nrof_changed = 0
-        nrof_iou3 = 0
-        nrof_force = 0
-        for line in open(args.input_dir, 'r'):
-            vec = line.strip().split()
-            person = vec[0]
-            img_id = int(vec[1])
-            v = datamap.get(person, None)
-            if v is None:
-              continue
-            #TODO
-            #if not img_id in v[1]:
-            #  continue
-            labelid = v[0]
-            img_str = base64.b64decode(vec[-1])
-            nparr = np.fromstring(img_str, np.uint8)
-            img = cv2.imdecode(nparr, cv2.CV_LOAD_IMAGE_COLOR)
-            img = img[...,::-1] #to rgb
-            if nrof_images_total%100==0:
-              print("Processing %d, (%d)" % (nrof_images_total, nrof_successfully_aligned))
-            nrof_images_total += 1
-            target_dir = os.path.join(output_dir, person)
-            if not os.path.exists(target_dir):
-              os.makedirs(target_dir)
-            target_path = os.path.join(target_dir, "%d.jpg"%img_id)
-            _minsize = minsize
-            fimage = edict()
-            fimage.bbox = None
-            fimage.image_path = target_path
-            fimage.classname = str(labelid)
-            if fimage.bbox is not None:
-              _bb = fimage.bbox
-              _minsize = min( [_bb[2]-_bb[0], _bb[3]-_bb[1], img.shape[0]//2, img.shape[1]//2] )
-            else:
-              _minsize = min(img.shape[0]//5, img.shape[1]//5)
-            bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, threshold, factor)
-            bindex = -1
-            nrof_faces = bounding_boxes.shape[0]
-            if fimage.bbox is None and nrof_faces>0:
-              det = bounding_boxes[:,0:4]
-              img_size = np.asarray(img.shape)[0:2]
-              bindex = 0
-              if nrof_faces>1:
-                bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-                img_center = img_size / 2
-                offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-                offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-                bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-            if fimage.bbox is not None:
-              if nrof_faces>0:
-                assert(bounding_boxes.shape[0]==points.shape[1])
-                det = bounding_boxes[:,0:4]
-                img_size = np.asarray(img.shape)[0:2]
-                index2 = [0.0, 0]
-                for i in xrange(det.shape[0]):
-                  _det = det[i]
-                  iou = IOU(fimage.bbox, _det)
-                  if iou>index2[0]:
-                    index2[0] = iou
-                    index2[1] = i
-                if index2[0]>-0.3:
-                  bindex = index2[1]
-                  nrof_iou3+=1
-              if bindex<0:
-                bounding_boxes, points = detect_face.detect_face_force(img, fimage.bbox, pnet, rnet, onet)
-                bindex = 0
-                nrof_force+=1
-                    
-            if bindex>=0:
-
-                det = bounding_boxes[:,0:4]
-                det = det[bindex,:]
-                points = points[:, bindex]
-                landmark = points.reshape((2,5)).T
-                #points need to be transpose, points = points.reshape( (5,2) ).transpose()
-                det = np.squeeze(det)
-                bb = det
-                points = list(points.flatten())
-                assert(len(points)==10)
-                warped = face_preprocess.preprocess(img, bbox=bb, landmark = landmark, image_size=args.image_size)
-                misc.imsave(target_path, warped)
-                nrof_successfully_aligned += 1
-                oline = '%d\t%s\t%d' % (1,fimage.image_path, int(fimage.classname))
-                #oline = '%d\t%s\t%d\t%d\t%d\t%d\t%d\t' % (0,fimage.image_path, int(fimage.classname), bb[0], bb[1], bb[2], bb[3])
-                #oline += '\t'.join([str(x) for x in points])
-                text_file.write("%s\n"%oline)
-                            
-    print('Total number of images: %d' % nrof_images_total)
-    print('Number of successfully aligned images: %d' % nrof_successfully_aligned)
-    print('Number of changed: %d' % nrof_changed)
-    print('Number of iou3: %d' % nrof_iou3)
-    print('Number of force: %d' % nrof_force)
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('--input-dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('--name', type=str, default='celeb', help='')
-    parser.add_argument('--output-dir', type=str, help='Directory with aligned face thumbnails.')
-    parser.add_argument('--image-size', type=str, help='Image size (height, width) in pixels.', default='112,112')
-    #parser.add_argument('--margin', type=int,
-    #    help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
diff --git a/src/align/align_dataset.py b/src/align/align_dataset.py
deleted file mode 100644
index e74224a1d..000000000
--- a/src/align/align_dataset.py
+++ /dev/null
@@ -1,137 +0,0 @@
-"""Performs face alignment and stores face thumbnails in the output directory."""
-
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import random
-import align_dlib  # @UnresolvedImport
-import facenet
-
-def main(args):
-    align = align_dlib.AlignDlib(os.path.expanduser(args.dlib_face_predictor))
-    landmarkIndices = align_dlib.AlignDlib.OUTER_EYES_AND_NOSE
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    # Store some git revision info in a text file in the log directory
-    src_path,_ = os.path.split(os.path.realpath(__file__))
-    facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    dataset = facenet.get_dataset(args.input_dir)
-    random.shuffle(dataset)
-    # Scale the image such that the face fills the frame when cropped to crop_size
-    scale = float(args.face_size) / args.image_size
-    nrof_images_total = 0
-    nrof_prealigned_images = 0
-    nrof_successfully_aligned = 0
-    for cls in dataset:
-        output_class_dir = os.path.join(output_dir, cls.name)
-        if not os.path.exists(output_class_dir):
-            os.makedirs(output_class_dir)
-        random.shuffle(cls.image_paths)
-        for image_path in cls.image_paths:
-            nrof_images_total += 1
-            filename = os.path.splitext(os.path.split(image_path)[1])[0]
-            output_filename = os.path.join(output_class_dir, filename+'.png')
-            if not os.path.exists(output_filename):
-                try:
-                    img = misc.imread(image_path)
-                except (IOError, ValueError, IndexError) as e:
-                    errorMessage = '{}: {}'.format(image_path, e)
-                    print(errorMessage)
-                else:
-                    if img.ndim == 2:
-                        img = facenet.to_rgb(img)
-                    if args.use_center_crop:
-                        scaled = misc.imresize(img, args.prealigned_scale, interp='bilinear')
-                        sz1 = scaled.shape[1]/2
-                        sz2 = args.image_size/2
-                        aligned = scaled[(sz1-sz2):(sz1+sz2),(sz1-sz2):(sz1+sz2),:]
-                    else:
-                        aligned = align.align(args.image_size, img, landmarkIndices=landmarkIndices, 
-                                              skipMulti=False, scale=scale)
-                    if aligned is not None:
-                        print(image_path)
-                        nrof_successfully_aligned += 1
-                        misc.imsave(output_filename, aligned)
-                    elif args.prealigned_dir:
-                        # Face detection failed. Use center crop from pre-aligned dataset
-                        class_name = os.path.split(output_class_dir)[1]
-                        image_path_without_ext = os.path.join(os.path.expanduser(args.prealigned_dir), 
-                                                              class_name, filename)
-                        # Find the extension of the image
-                        exts = ('jpg', 'png')
-                        for ext in exts:
-                            temp_path = image_path_without_ext + '.' + ext
-                            image_path = ''
-                            if os.path.exists(temp_path):
-                                image_path = temp_path
-                                break
-                        try:
-                            img = misc.imread(image_path)
-                        except (IOError, ValueError, IndexError) as e:
-                            errorMessage = '{}: {}'.format(image_path, e)
-                            print(errorMessage)
-                        else:
-                            scaled = misc.imresize(img, args.prealigned_scale, interp='bilinear')
-                            sz1 = scaled.shape[1]/2
-                            sz2 = args.image_size/2
-                            cropped = scaled[(sz1-sz2):(sz1+sz2),(sz1-sz2):(sz1+sz2),:]
-                            print(image_path)
-                            nrof_prealigned_images += 1
-                            misc.imsave(output_filename, cropped)
-                    else:
-                        print('Unable to align "%s"' % image_path)
-                            
-    print('Total number of images: %d' % nrof_images_total)
-    print('Number of successfully aligned images: %d' % nrof_successfully_aligned)
-    print('Number of pre-aligned images: %d' % nrof_prealigned_images)
-            
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('input_dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('output_dir', type=str, help='Directory with aligned face thumbnails.')
-    parser.add_argument('--dlib_face_predictor', type=str,
-        help='File containing the dlib face predictor.', default='../data/shape_predictor_68_face_landmarks.dat')
-    parser.add_argument('--image_size', type=int,
-        help='Image size (height, width) in pixels.', default=110)
-    parser.add_argument('--face_size', type=int,
-        help='Size of the face thumbnail (height, width) in pixels.', default=96)
-    parser.add_argument('--use_center_crop', 
-        help='Use the center crop of the original image after scaling the image using prealigned_scale.', action='store_true')
-    parser.add_argument('--prealigned_dir', type=str,
-        help='Replace image with a pre-aligned version when face detection fails.', default='')
-    parser.add_argument('--prealigned_scale', type=float,
-        help='The amount of scaling to apply to prealigned images before taking the center crop.', default=0.87)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
diff --git a/src/align/align_dataset_mtcnn.py b/src/align/align_dataset_mtcnn.py
deleted file mode 100644
index d2a3eea98..000000000
--- a/src/align/align_dataset_mtcnn.py
+++ /dev/null
@@ -1,143 +0,0 @@
-"""Performs face alignment and stores face thumbnails in the output directory."""
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-import facenet
-import align.detect_face
-import random
-from time import sleep
-
-def main(args):
-    sleep(random.random())
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    # Store some git revision info in a text file in the log directory
-    src_path,_ = os.path.split(os.path.realpath(__file__))
-    facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    dataset = facenet.get_dataset(args.input_dir)
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        with sess.as_default():
-            pnet, rnet, onet = align.detect_face.create_mtcnn(sess, None)
-    
-    minsize = 20 # minimum size of face
-    threshold = [ 0.6, 0.7, 0.7 ]  # three steps's threshold
-    factor = 0.709 # scale factor
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    random_key = np.random.randint(0, high=99999)
-    bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    
-    with open(bounding_boxes_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof_successfully_aligned = 0
-        if args.random_order:
-            random.shuffle(dataset)
-        for cls in dataset:
-            output_class_dir = os.path.join(output_dir, cls.name)
-            if not os.path.exists(output_class_dir):
-                os.makedirs(output_class_dir)
-                if args.random_order:
-                    random.shuffle(cls.image_paths)
-            for image_path in cls.image_paths:
-                nrof_images_total += 1
-                filename = os.path.splitext(os.path.split(image_path)[1])[0]
-                output_filename = os.path.join(output_class_dir, filename+'.png')
-                print(image_path)
-                if not os.path.exists(output_filename):
-                    try:
-                        img = misc.imread(image_path)
-                    except (IOError, ValueError, IndexError) as e:
-                        errorMessage = '{}: {}'.format(image_path, e)
-                        print(errorMessage)
-                    else:
-                        if img.ndim<2:
-                            print('Unable to align "%s"' % image_path)
-                            text_file.write('%s\n' % (output_filename))
-                            continue
-                        if img.ndim == 2:
-                            img = facenet.to_rgb(img)
-                        img = img[:,:,0:3]
-    
-                        bounding_boxes, _ = align.detect_face.detect_face(img, minsize, pnet, rnet, onet, threshold, factor)
-                        nrof_faces = bounding_boxes.shape[0]
-                        if nrof_faces>0:
-                            det = bounding_boxes[:,0:4]
-                            img_size = np.asarray(img.shape)[0:2]
-                            if nrof_faces>1:
-                                bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-                                img_center = img_size / 2
-                                offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-                                offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-                                index = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-                                det = det[index,:]
-                            det = np.squeeze(det)
-                            bb = np.zeros(4, dtype=np.int32)
-                            bb[0] = np.maximum(det[0]-args.margin/2, 0)
-                            bb[1] = np.maximum(det[1]-args.margin/2, 0)
-                            bb[2] = np.minimum(det[2]+args.margin/2, img_size[1])
-                            bb[3] = np.minimum(det[3]+args.margin/2, img_size[0])
-                            cropped = img[bb[1]:bb[3],bb[0]:bb[2],:]
-                            scaled = misc.imresize(cropped, (args.image_size, args.image_size), interp='bilinear')
-                            nrof_successfully_aligned += 1
-                            misc.imsave(output_filename, scaled)
-                            text_file.write('%s %d %d %d %d\n' % (output_filename, bb[0], bb[1], bb[2], bb[3]))
-                        else:
-                            print('Unable to align "%s"' % image_path)
-                            text_file.write('%s\n' % (output_filename))
-                            
-    print('Total number of images: %d' % nrof_images_total)
-    print('Number of successfully aligned images: %d' % nrof_successfully_aligned)
-            
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('input_dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('output_dir', type=str, help='Directory with aligned face thumbnails.')
-    parser.add_argument('--image_size', type=int,
-        help='Image size (height, width) in pixels.', default=182)
-    parser.add_argument('--margin', type=int,
-        help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    parser.add_argument('--random_order', 
-        help='Shuffles the order of images to enable alignment using multiple processes.', action='store_true')
-    parser.add_argument('--gpu_memory_fraction', type=float,
-        help='Upper bound on the amount of GPU memory that will be used by the process.', default=1.0)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
diff --git a/src/align/align_dlib.py b/src/align/align_dlib.py
deleted file mode 100644
index e5e133743..000000000
--- a/src/align/align_dlib.py
+++ /dev/null
@@ -1,204 +0,0 @@
-# Copyright 2015-2016 Carnegie Mellon University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""Module for dlib-based alignment."""
-
-# NOTE: This file has been copied from the openface project.
-#  https://github.com/cmusatyalab/openface/blob/master/openface/align_dlib.py
-
-import cv2
-import dlib
-import numpy as np
-
-TEMPLATE = np.float32([
-    (0.0792396913815, 0.339223741112), (0.0829219487236, 0.456955367943),
-    (0.0967927109165, 0.575648016728), (0.122141515615, 0.691921601066),
-    (0.168687863544, 0.800341263616), (0.239789390707, 0.895732504778),
-    (0.325662452515, 0.977068762493), (0.422318282013, 1.04329000149),
-    (0.531777802068, 1.06080371126), (0.641296298053, 1.03981924107),
-    (0.738105872266, 0.972268833998), (0.824444363295, 0.889624082279),
-    (0.894792677532, 0.792494155836), (0.939395486253, 0.681546643421),
-    (0.96111933829, 0.562238253072), (0.970579841181, 0.441758925744),
-    (0.971193274221, 0.322118743967), (0.163846223133, 0.249151738053),
-    (0.21780354657, 0.204255863861), (0.291299351124, 0.192367318323),
-    (0.367460241458, 0.203582210627), (0.4392945113, 0.233135599851),
-    (0.586445962425, 0.228141644834), (0.660152671635, 0.195923841854),
-    (0.737466449096, 0.182360984545), (0.813236546239, 0.192828009114),
-    (0.8707571886, 0.235293377042), (0.51534533827, 0.31863546193),
-    (0.516221448289, 0.396200446263), (0.517118861835, 0.473797687758),
-    (0.51816430343, 0.553157797772), (0.433701156035, 0.604054457668),
-    (0.475501237769, 0.62076344024), (0.520712933176, 0.634268222208),
-    (0.565874114041, 0.618796581487), (0.607054002672, 0.60157671656),
-    (0.252418718401, 0.331052263829), (0.298663015648, 0.302646354002),
-    (0.355749724218, 0.303020650651), (0.403718978315, 0.33867711083),
-    (0.352507175597, 0.349987615384), (0.296791759886, 0.350478978225),
-    (0.631326076346, 0.334136672344), (0.679073381078, 0.29645404267),
-    (0.73597236153, 0.294721285802), (0.782865376271, 0.321305281656),
-    (0.740312274764, 0.341849376713), (0.68499850091, 0.343734332172),
-    (0.353167761422, 0.746189164237), (0.414587777921, 0.719053835073),
-    (0.477677654595, 0.706835892494), (0.522732900812, 0.717092275768),
-    (0.569832064287, 0.705414478982), (0.635195811927, 0.71565572516),
-    (0.69951672331, 0.739419187253), (0.639447159575, 0.805236879972),
-    (0.576410514055, 0.835436670169), (0.525398405766, 0.841706377792),
-    (0.47641545769, 0.837505914975), (0.41379548902, 0.810045601727),
-    (0.380084785646, 0.749979603086), (0.477955996282, 0.74513234612),
-    (0.523389793327, 0.748924302636), (0.571057789237, 0.74332894691),
-    (0.672409137852, 0.744177032192), (0.572539621444, 0.776609286626),
-    (0.5240106503, 0.783370783245), (0.477561227414, 0.778476346951)])
-
-INV_TEMPLATE = np.float32([
-                            (-0.04099179660567834, -0.008425234314031194, 2.575498465013183),
-                            (0.04062510634554352, -0.009678089746831375, -1.2534351452524177),
-                            (0.0003666902601348179, 0.01810332406086298, -0.32206331976076663)])
-
-TPL_MIN, TPL_MAX = np.min(TEMPLATE, axis=0), np.max(TEMPLATE, axis=0)
-MINMAX_TEMPLATE = (TEMPLATE - TPL_MIN) / (TPL_MAX - TPL_MIN)
-
-
-class AlignDlib:
-    """
-    Use `dlib's landmark estimation <http://blog.dlib.net/2014/08/real-time-face-pose-estimation.html>`_ to align faces.
-
-    The alignment preprocess faces for input into a neural network.
-    Faces are resized to the same size (such as 96x96) and transformed
-    to make landmarks (such as the eyes and nose) appear at the same
-    location on every image.
-
-    Normalized landmarks:
-
-    .. image:: ../images/dlib-landmark-mean.png
-    """
-
-    #: Landmark indices corresponding to the inner eyes and bottom lip.
-    INNER_EYES_AND_BOTTOM_LIP = [39, 42, 57]
-
-    #: Landmark indices corresponding to the outer eyes and nose.
-    OUTER_EYES_AND_NOSE = [36, 45, 33]
-
-    def __init__(self, facePredictor):
-        """
-        Instantiate an 'AlignDlib' object.
-
-        :param facePredictor: The path to dlib's
-        :type facePredictor: str
-        """
-        assert facePredictor is not None
-
-        #pylint: disable=no-member
-        self.detector = dlib.get_frontal_face_detector()
-        self.predictor = dlib.shape_predictor(facePredictor)
-
-    def getAllFaceBoundingBoxes(self, rgbImg):
-        """
-        Find all face bounding boxes in an image.
-
-        :param rgbImg: RGB image to process. Shape: (height, width, 3)
-        :type rgbImg: numpy.ndarray
-        :return: All face bounding boxes in an image.
-        :rtype: dlib.rectangles
-        """
-        assert rgbImg is not None
-
-        try:
-            return self.detector(rgbImg, 1)
-        except Exception as e: #pylint: disable=broad-except
-            print("Warning: {}".format(e))
-            # In rare cases, exceptions are thrown.
-            return []
-
-    def getLargestFaceBoundingBox(self, rgbImg, skipMulti=False):
-        """
-        Find the largest face bounding box in an image.
-
-        :param rgbImg: RGB image to process. Shape: (height, width, 3)
-        :type rgbImg: numpy.ndarray
-        :param skipMulti: Skip image if more than one face detected.
-        :type skipMulti: bool
-        :return: The largest face bounding box in an image, or None.
-        :rtype: dlib.rectangle
-        """
-        assert rgbImg is not None
-
-        faces = self.getAllFaceBoundingBoxes(rgbImg)
-        if (not skipMulti and len(faces) > 0) or len(faces) == 1:
-            return max(faces, key=lambda rect: rect.width() * rect.height())
-        else:
-            return None
-
-    def findLandmarks(self, rgbImg, bb):
-        """
-        Find the landmarks of a face.
-
-        :param rgbImg: RGB image to process. Shape: (height, width, 3)
-        :type rgbImg: numpy.ndarray
-        :param bb: Bounding box around the face to find landmarks for.
-        :type bb: dlib.rectangle
-        :return: Detected landmark locations.
-        :rtype: list of (x,y) tuples
-        """
-        assert rgbImg is not None
-        assert bb is not None
-
-        points = self.predictor(rgbImg, bb)
-        #return list(map(lambda p: (p.x, p.y), points.parts()))
-        return [(p.x, p.y) for p in points.parts()]
-
-    #pylint: disable=dangerous-default-value
-    def align(self, imgDim, rgbImg, bb=None,
-              landmarks=None, landmarkIndices=INNER_EYES_AND_BOTTOM_LIP,
-              skipMulti=False, scale=1.0):
-        r"""align(imgDim, rgbImg, bb=None, landmarks=None, landmarkIndices=INNER_EYES_AND_BOTTOM_LIP)
-
-        Transform and align a face in an image.
-
-        :param imgDim: The edge length in pixels of the square the image is resized to.
-        :type imgDim: int
-        :param rgbImg: RGB image to process. Shape: (height, width, 3)
-        :type rgbImg: numpy.ndarray
-        :param bb: Bounding box around the face to align. \
-                   Defaults to the largest face.
-        :type bb: dlib.rectangle
-        :param landmarks: Detected landmark locations. \
-                          Landmarks found on `bb` if not provided.
-        :type landmarks: list of (x,y) tuples
-        :param landmarkIndices: The indices to transform to.
-        :type landmarkIndices: list of ints
-        :param skipMulti: Skip image if more than one face detected.
-        :type skipMulti: bool
-        :param scale: Scale image before cropping to the size given by imgDim.
-        :type scale: float
-        :return: The aligned RGB image. Shape: (imgDim, imgDim, 3)
-        :rtype: numpy.ndarray
-        """
-        assert imgDim is not None
-        assert rgbImg is not None
-        assert landmarkIndices is not None
-
-        if bb is None:
-            bb = self.getLargestFaceBoundingBox(rgbImg, skipMulti)
-            if bb is None:
-                return
-
-        if landmarks is None:
-            landmarks = self.findLandmarks(rgbImg, bb)
-
-        npLandmarks = np.float32(landmarks)
-        npLandmarkIndices = np.array(landmarkIndices)
-
-        #pylint: disable=maybe-no-member
-        H = cv2.getAffineTransform(npLandmarks[npLandmarkIndices],
-                                   imgDim * MINMAX_TEMPLATE[npLandmarkIndices]*scale + imgDim*(1-scale)/2)
-        thumbnail = cv2.warpAffine(rgbImg, H, (imgDim, imgDim))
-        
-        return thumbnail
diff --git a/src/align/align_facescrub.py b/src/align/align_facescrub.py
deleted file mode 100644
index 890d0b0d9..000000000
--- a/src/align/align_facescrub.py
+++ /dev/null
@@ -1,289 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import json
-import argparse
-import tensorflow as tf
-import numpy as np
-#import facenet
-import detect_face
-import random
-from time import sleep
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-from skimage import transform as trans
-import cv2
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-
-def main(args):
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    # Store some git revision info in a text file in the log directory
-    #facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    image_dir = os.path.join(args.input_dir, 'facescrub')
-    dataset = face_image.get_dataset('facescrub', image_dir)
-    print('dataset size', len(dataset))
-    bbox = {}
-    for label_file in ['facescrub_actors.txt',  'facescrub_actresses.txt']:
-      label_file = os.path.join(args.input_dir, label_file)
-      pp = 0
-      for line in open(label_file, 'r'):
-        pp+=1
-        if pp==1:
-          continue
-        vec = line.split("\t")
-        key = (vec[0], int(vec[2]))
-        value = [int(x) for x in vec[4].split(',')]
-        bbox[key] = value
-    print('bbox size', len(bbox))
-
-    valid_key = {}
-    json_data = open(os.path.join(args.input_dir, 'facescrub_uncropped_features_list.json')).read()
-    json_data = json.loads(json_data)['path']
-    for _data in json_data:
-      key = _data.split('/')[-1]
-      pos = key.rfind('.')
-      if pos<0:
-        print(_data)
-      else:
-        key = key[0:pos]
-      keys = key.split('_')
-      #print(key)
-      if len(keys)!=2:
-        print('err', key, _data)
-        continue
-      #assert len(keys)==2
-      key = (keys[0], int(keys[1]))
-      valid_key[key] = 1
-      #print(key)
-    print('valid keys', len(valid_key))
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        sess = tf.Session()
-        with sess.as_default():
-            pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-    
-    minsize = 100 # minimum size of face
-    threshold = [ 0.6, 0.7, 0.7 ]  # three steps's threshold
-    factor = 0.709 # scale factor
-    image_size = [112,96]
-    image_size = [112,112]
-    src = np.array([
-      [30.2946, 51.6963],
-      [65.5318, 51.5014],
-      [48.0252, 71.7366],
-      [33.5493, 92.3655],
-      [62.7299, 92.2041] ], dtype=np.float32 )
-    if image_size[1]==112:
-      src[:,0] += 8.0
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    #random_key = np.random.randint(0, high=99999)
-    #bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    #output_filename = os.path.join(output_dir, 'faceinsight_align_%s.lst' % args.name)
-    if not os.path.exists(args.output_dir):
-      os.makedirs(args.output_dir)
-
-    output_filename = os.path.join(args.output_dir, 'lst')
-    
-    
-    with open(output_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof = np.zeros( (5,), dtype=np.int32)
-        for fimage in dataset:
-            if nrof_images_total%100==0:
-              print("Processing %d, (%s)" % (nrof_images_total, nrof))
-            nrof_images_total += 1
-            #if nrof_images_total<950000:
-            #  continue
-            image_path = fimage.image_path
-            if not os.path.exists(image_path):
-              print('image not found (%s)'%image_path)
-              continue
-            #print(image_path)
-            filename = os.path.splitext(os.path.split(image_path)[1])[0]
-            _paths = fimage.image_path.split('/')
-            print(fimage.image_path)
-            a,b = _paths[-2], _paths[-1]
-            pb = b.rfind('.')
-            bname = b[0:pb]
-            pb = bname.rfind('_')
-            body = bname[(pb+1):]
-            img_id = int(body)
-            key = (a, img_id)
-            if not key in valid_key:
-              continue
-            #print(b, img_id)
-            assert key in bbox
-            fimage.bbox = bbox[key]
-            try:
-                img = misc.imread(image_path)
-            except (IOError, ValueError, IndexError) as e:
-                errorMessage = '{}: {}'.format(image_path, e)
-                print(errorMessage)
-            else:
-                if img.ndim<2:
-                    print('Unable to align "%s", img dim error' % image_path)
-                    #text_file.write('%s\n' % (output_filename))
-                    continue
-                if img.ndim == 2:
-                    img = to_rgb(img)
-                img = img[:,:,0:3]
-                tb = bname.replace(' ','_')+".png"
-                ta = a.replace(' ','_')
-                target_dir = os.path.join(args.output_dir, ta)
-                if not os.path.exists(target_dir):
-                  os.makedirs(target_dir)
-
-                target_file = os.path.join(target_dir, tb)
-                warped = None
-                if fimage.landmark is not None:
-                  dst = fimage.landmark.astype(np.float32)
-
-                  tform = trans.SimilarityTransform()
-                  tform.estimate(dst, src[0:3,:]*1.5+image_size[0]*0.25)
-                  M = tform.params[0:2,:]
-                  warped0 = cv2.warpAffine(img,M,(image_size[1]*2,image_size[0]*2), borderValue = 0.0)
-                  _minsize = image_size[0]
-                  bounding_boxes, points = detect_face.detect_face(warped0, _minsize, pnet, rnet, onet, threshold, factor)
-                  if bounding_boxes.shape[0]>0:
-                    bindex = 0
-                    det = bounding_boxes[bindex,0:4]
-                    #points need to be transpose, points = points.reshape( (5,2) ).transpose()
-                    dst = points[:, bindex].reshape( (2,5) ).T
-                    tform = trans.SimilarityTransform()
-                    tform.estimate(dst, src)
-                    M = tform.params[0:2,:]
-                    warped = cv2.warpAffine(warped0,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                    nrof[0]+=1
-                #assert fimage.bbox is not None
-                if warped is None and fimage.bbox is not None:
-                  _minsize = img.shape[0]//4
-                  bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, threshold, factor)
-                  if bounding_boxes.shape[0]>0:
-                    det = bounding_boxes[:,0:4]
-                    bindex = -1
-                    index2 = [0.0, 0]
-                    for i in xrange(det.shape[0]):
-                      _det = det[i]
-                      iou = IOU(fimage.bbox, _det)
-                      if iou>index2[0]:
-                        index2[0] = iou
-                        index2[1] = i
-                    if index2[0]>0.3:
-                      bindex = index2[1]
-                    if bindex>=0:
-                      dst = points[:, bindex].reshape( (2,5) ).T
-                      tform = trans.SimilarityTransform()
-                      tform.estimate(dst, src)
-                      M = tform.params[0:2,:]
-                      warped = cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                      nrof[1]+=1
-                      #print('1',target_file,index2[0])
-                if warped is None and fimage.bbox is not None:
-                  bb = fimage.bbox
-                  #croped = img[bb[1]:bb[3],bb[0]:bb[2],:]
-                  bounding_boxes, points = detect_face.detect_face_force(img, bb, pnet, rnet, onet)
-                  assert bounding_boxes.shape[0]==1
-                  _box = bounding_boxes[0]
-                  if _box[4]>=0.3:
-                    dst = points[:, 0].reshape( (2,5) ).T
-                    tform = trans.SimilarityTransform()
-                    tform.estimate(dst, src)
-                    M = tform.params[0:2,:]
-                    warped = cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                    nrof[2]+=1
-                    #print('2',target_file)
-
-                if warped is None:
-                  roi = np.zeros( (4,), dtype=np.int32)
-                  roi[0] = int(img.shape[1]*0.06)
-                  roi[1] = int(img.shape[0]*0.06)
-                  roi[2] = img.shape[1]-roi[0]
-                  roi[3] = img.shape[0]-roi[1]
-                  if fimage.bbox is not None:
-                    bb = fimage.bbox
-                    h = bb[3]-bb[1]
-                    w = bb[2]-bb[0]
-                    x = bb[0]
-                    y = bb[1]
-                    #roi = np.copy(bb)
-                    _w = int( (float(h)/image_size[0])*image_size[1] )
-                    x += (w-_w)//2
-                    #x = min( max(0,x), img.shape[1] )
-                    x = max(0,x)
-                    xw = x+_w
-                    xw = min(xw, img.shape[1])
-                    roi = np.array( (x, y, xw, y+h), dtype=np.int32)
-                    nrof[3]+=1
-                  else:
-                    nrof[4]+=1
-                  #print('3',bb,roi,img.shape)
-                  #print('3',target_file)
-                  warped = img[roi[1]:roi[3],roi[0]:roi[2],:]
-                  #print(warped.shape)
-                  warped = cv2.resize(warped, (image_size[1], image_size[0]))
-                bgr = warped[...,::-1]
-                cv2.imwrite(target_file, bgr)
-                oline = '%d\t%s\t%d\n' % (1,target_file, int(fimage.classname))
-                text_file.write(oline)
-                            
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('--input-dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('--output-dir', type=str, help='Directory with aligned face thumbnails.')
-    #parser.add_argument('--image_size', type=int,
-    #    help='Image size (height, width) in pixels.', default=182)
-    #parser.add_argument('--margin', type=int,
-    #    help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
-
diff --git a/src/align/align_insight.py b/src/align/align_insight.py
deleted file mode 100644
index 363bf3fa0..000000000
--- a/src/align/align_insight.py
+++ /dev/null
@@ -1,251 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-#import facenet
-import detect_face
-import random
-from time import sleep
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-
-def main(args):
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    # Store some git revision info in a text file in the log directory
-    src_path,_ = os.path.split(os.path.realpath(__file__))
-    #facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    dataset = face_image.get_dataset(args.name, args.input_dir)
-    print('dataset size', args.name, len(dataset))
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        sess = tf.Session()
-        with sess.as_default():
-            pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-    
-    minsize = 100 # minimum size of face
-    threshold = [ 0.6, 0.7, 0.7 ]  # three steps's threshold
-    factor = 0.709 # scale factor
-    if args.name=='lfw' or args.name=='webface' or args.name=='vgg':
-      minsize = 20
-      threshold = [0.6,0.7,0.9]
-      factor = 0.85
-    if args.name=='ytf':
-      minsize = 20
-      threshold = [0.6,0.7,0.4]
-      factor = 0.85
-
-    print(minsize)
-    print(threshold)
-    print(factor)
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    #random_key = np.random.randint(0, high=99999)
-    #bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    output_filename = os.path.join(output_dir, 'faceinsight_align_%s.lst' % args.name)
-    
-    with open(output_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof_successfully_aligned = 0
-        nrof_changed = 0
-        nrof_iou3 = 0
-        nrof_force = 0
-        for fimage in dataset:
-            if nrof_images_total%100==0:
-              print("Processing %d, (%d)" % (nrof_images_total, nrof_successfully_aligned))
-            nrof_images_total += 1
-            image_path = fimage.image_path
-            if not os.path.exists(image_path):
-              print('image not found (%s)'%image_path)
-              continue
-            filename = os.path.splitext(os.path.split(image_path)[1])[0]
-            #print(image_path)
-            try:
-                img = misc.imread(image_path)
-            except (IOError, ValueError, IndexError) as e:
-                errorMessage = '{}: {}'.format(image_path, e)
-                print(errorMessage)
-            else:
-                if img.ndim<2:
-                    print('Unable to align "%s", img dim error' % image_path)
-                    #text_file.write('%s\n' % (output_filename))
-                    continue
-                if img.ndim == 2:
-                    img = to_rgb(img)
-                img = img[:,:,0:3]
-                _minsize = minsize
-                if fimage.bbox is not None:
-                  _bb = fimage.bbox
-                  _minsize = min( [_bb[2]-_bb[0], _bb[3]-_bb[1], img.shape[0]//2, img.shape[1]//2] )
-
-                bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, threshold, factor)
-                bindex = -1
-                nrof_faces = bounding_boxes.shape[0]
-                if fimage.bbox is None and nrof_faces>0:
-                  det = bounding_boxes[:,0:4]
-                  img_size = np.asarray(img.shape)[0:2]
-                  bindex = 0
-                  if nrof_faces>1:
-                    bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-                    img_center = img_size / 2
-                    offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-                    offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-                    bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-                if fimage.bbox is not None:
-                  if nrof_faces>0:
-                    assert(bounding_boxes.shape[0]==points.shape[1])
-                    det = bounding_boxes[:,0:4]
-                    img_size = np.asarray(img.shape)[0:2]
-                    index2 = [0.0, 0]
-                    for i in xrange(det.shape[0]):
-                      _det = det[i]
-                      iou = IOU(fimage.bbox, _det)
-                      if iou>index2[0]:
-                        index2[0] = iou
-                        index2[1] = i
-                    if index2[0]>-0.3:
-                      bindex = index2[1]
-                      nrof_iou3+=1
-                  if bindex<0:
-                    bounding_boxes, points = detect_face.detect_face_force(img, fimage.bbox, pnet, rnet, onet)
-                    bindex = 0
-                    nrof_force+=1
-                  #if bindex<0:
-                  #  _img = img[fimage.bbox[1]:fimage.bbox[3], fimage.bbox[0]:fimage.bbox[2],:]
-                  #  woffset = fimage.bbox[0]
-                  #  hoffset = fimage.bbox[1]
-                  #  _minsize = min( [_img.shape[0]//3, _img.shape[1]//3] )
-                  #  bounding_boxes, points = detect_face.detect_face(_img, _minsize, pnet, rnet, onet, [0.6,0.7,0.01], factor)
-                  #  nrof_faces = bounding_boxes.shape[0]
-                  #  print(nrof_faces)
-                  #  if nrof_faces>0:
-                  #    #print(points.shape)
-                  #    #assert(nrof_faces>0)
-                  #    bounding_boxes[:,0]+=woffset
-                  #    bounding_boxes[:,2]+=woffset
-                  #    bounding_boxes[:,1]+=hoffset
-                  #    bounding_boxes[:,3]+=hoffset
-                  #    points[0:5,:] += woffset
-                  #    points[5:10,:] += hoffset
-                  #    bindex = 0
-                  #    score = bounding_boxes[bindex,4]
-                  #    print(score)
-                  #    if score<=0.0:
-                  #      bindex = -1
-                  #    else:
-                  #      nrof_force+=1
-                  #if bindex<0:
-                  #  _bb = fimage.bbox
-                  #  _minsize = min( [_bb[2]-_bb[0], _bb[3]-_bb[1], img.shape[0]//2, img.shape[1]//2] )
-                  #  bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, [0.6,0.7,0.1], factor)
-                  #  nrof_faces = bounding_boxes.shape[0]
-                  #  print(nrof_faces)
-                  #  if nrof_faces>0:
-                  #    bindex = 0
-                #if fimage.bbox is not None and bounding_boxes.shape[0]==0:
-                #  bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, [0.6,0.7,0.3], factor)
-
-
-                #print(bounding_boxes.shape, points.shape)
-                #print(nrof_faces, points.shape)
-                        
-                if bindex>=0:
-
-                    det = bounding_boxes[:,0:4]
-                    det = det[bindex,:]
-                    points = points[:, bindex]
-                    #points need to be transpose, points = points.reshape( (5,2) ).transpose()
-                    det = np.squeeze(det)
-                    #bb = np.zeros(4, dtype=np.int32)
-                    #bb[0] = np.maximum(det[0]-args.margin/2, 0)
-                    #bb[1] = np.maximum(det[1]-args.margin/2, 0)
-                    #bb[2] = np.minimum(det[2]+args.margin/2, img_size[1])
-                    #bb[3] = np.minimum(det[3]+args.margin/2, img_size[0])
-                    bb = det
-                    #print(points.shape)
-                    points = list(points.flatten())
-                    assert(len(points)==10)
-                    #cropped = img[bb[1]:bb[3],bb[0]:bb[2],:]
-                    #scaled = misc.imresize(cropped, (args.image_size, args.image_size), interp='bilinear')
-                    #misc.imsave(output_filename, scaled)
-                    nrof_successfully_aligned += 1
-                    oline = '%d\t%s\t%d\t%d\t%d\t%d\t%d\t' % (0,fimage.image_path, int(fimage.classname), bb[0], bb[1], bb[2], bb[3])
-                    oline += '\t'.join([str(x) for x in points])
-                    text_file.write("%s\n"%oline)
-                else:
-                    print('Unable to align "%s", no face detected' % image_path)
-                    if args.force>0:
-                      if fimage.bbox is None:
-                        oline = '%d\t%s\t%d\n' % (0,fimage.image_path, int(fimage.classname))
-                      else:
-                        bb = fimage.bbox
-                        oline = '%d\t%s\t%d\t%d\t%d\t%d\t%d\n' % (0,fimage.image_path, int(fimage.classname), bb[0], bb[1], bb[2], bb[3])
-                      text_file.write(oline)
-                      #text_file.write('%s\n' % (output_filename))
-                            
-    print('Total number of images: %d' % nrof_images_total)
-    print('Number of successfully aligned images: %d' % nrof_successfully_aligned)
-    print('Number of changed: %d' % nrof_changed)
-    print('Number of iou3: %d' % nrof_iou3)
-    print('Number of force: %d' % nrof_force)
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('--input-dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('--name', type=str, help='dataset name, can be facescrub, megaface, webface, celeb.')
-    parser.add_argument('--output-dir', type=str, help='Directory with aligned face thumbnails.')
-    parser.add_argument('--force', type=int, help='force to output if no faces detected.', default=1)
-    #parser.add_argument('--margin', type=int,
-    #    help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
diff --git a/src/align/align_lfw.py b/src/align/align_lfw.py
deleted file mode 100644
index b80d8879d..000000000
--- a/src/align/align_lfw.py
+++ /dev/null
@@ -1,161 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-#import facenet
-import detect_face
-import random
-from time import sleep
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-import face_preprocess
-from skimage import transform as trans
-import cv2
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-
-def main(args):
-    #facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    dataset = face_image.get_dataset('lfw', args.input_dir)
-    print('dataset size', 'lfw', len(dataset))
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        sess = tf.Session()
-        with sess.as_default():
-            pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-    
-    minsize = 20
-    threshold = [0.6,0.7,0.9]
-    factor = 0.85
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    #random_key = np.random.randint(0, high=99999)
-    #bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    #output_filename = os.path.join(output_dir, 'faceinsight_align_%s.lst' % args.name)
-
-    if not os.path.exists(args.output_dir):
-      os.makedirs(args.output_dir)
-
-    output_filename = os.path.join(args.output_dir, 'lst')
-    
-    
-    with open(output_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof = np.zeros( (5,), dtype=np.int32)
-        for fimage in dataset:
-            if nrof_images_total%100==0:
-              print("Processing %d, (%s)" % (nrof_images_total, nrof))
-            nrof_images_total += 1
-            #if nrof_images_total<950000:
-            #  continue
-            image_path = fimage.image_path
-            if not os.path.exists(image_path):
-              print('image not found (%s)'%image_path)
-              continue
-            filename = os.path.splitext(os.path.split(image_path)[1])[0]
-            #print(image_path)
-            try:
-                img = misc.imread(image_path)
-            except (IOError, ValueError, IndexError) as e:
-                errorMessage = '{}: {}'.format(image_path, e)
-                print(errorMessage)
-            else:
-                if img.ndim<2:
-                    print('Unable to align "%s", img dim error' % image_path)
-                    #text_file.write('%s\n' % (output_filename))
-                    continue
-                if img.ndim == 2:
-                    img = to_rgb(img)
-                img = img[:,:,0:3]
-                _paths = fimage.image_path.split('/')
-                a,b = _paths[-2], _paths[-1]
-                target_dir = os.path.join(args.output_dir, a)
-                if not os.path.exists(target_dir):
-                  os.makedirs(target_dir)
-                target_file = os.path.join(target_dir, b)
-                _minsize = minsize
-                _bbox = None
-                _landmark = None
-                bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, threshold, factor)
-                nrof_faces = bounding_boxes.shape[0]
-                if nrof_faces>0:
-                  det = bounding_boxes[:,0:4]
-                  img_size = np.asarray(img.shape)[0:2]
-                  bindex = 0
-                  if nrof_faces>1:
-                      bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-                      img_center = img_size / 2
-                      offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-                      offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-                      bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-                  _bbox = bounding_boxes[bindex, 0:4]
-                  _landmark = points[:, bindex].reshape( (2,5) ).T
-                  nrof[0]+=1
-                else:
-                  nrof[1]+=1
-                warped = face_preprocess.preprocess(img, bbox=_bbox, landmark = _landmark, image_size=args.image_size)
-                bgr = warped[...,::-1]
-                #print(bgr.shape)
-                cv2.imwrite(target_file, bgr)
-                oline = '%d\t%s\t%d\n' % (1,target_file, int(fimage.classname))
-                text_file.write(oline)
-                            
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('--input-dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('--output-dir', type=str, help='Directory with aligned face thumbnails.')
-    parser.add_argument('--image-size', type=str, help='Image size (height, width) in pixels.', default='112,112')
-    #parser.add_argument('--margin', type=int,
-    #    help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/src/align/align_megaface.py b/src/align/align_megaface.py
deleted file mode 100644
index c22da4ee5..000000000
--- a/src/align/align_megaface.py
+++ /dev/null
@@ -1,240 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-#import facenet
-import detect_face
-import random
-from time import sleep
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-from skimage import transform as trans
-import cv2
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-
-def main(args):
-    output_dir = os.path.expanduser(args.output_dir)
-    if not os.path.exists(output_dir):
-        os.makedirs(output_dir)
-    # Store some git revision info in a text file in the log directory
-    src_path,_ = os.path.split(os.path.realpath(__file__))
-    #facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
-    dataset = face_image.get_dataset(args.name, args.input_dir)
-    print('dataset size', args.name, len(dataset))
-    
-    print('Creating networks and loading parameters')
-    
-    with tf.Graph().as_default():
-        #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-        #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-        sess = tf.Session()
-        with sess.as_default():
-            pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-    
-    minsize = 100 # minimum size of face
-    threshold = [ 0.6, 0.7, 0.7 ]  # three steps's threshold
-    factor = 0.709 # scale factor
-    #image_size = [112,96]
-    image_size = [112,112]
-    src = np.array([
-      [30.2946, 51.6963],
-      [65.5318, 51.5014],
-      [48.0252, 71.7366],
-      [33.5493, 92.3655],
-      [62.7299, 92.2041] ], dtype=np.float32 )
-
-    if image_size[1]==112:
-      src[:,0] += 8.0
-
-    # Add a random key to the filename to allow alignment using multiple processes
-    #random_key = np.random.randint(0, high=99999)
-    #bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
-    #output_filename = os.path.join(output_dir, 'faceinsight_align_%s.lst' % args.name)
-    if not os.path.exists(args.output_dir):
-      os.makedirs(args.output_dir)
-
-    output_filename = os.path.join(args.output_dir, 'lst')
-    
-    
-    with open(output_filename, "w") as text_file:
-        nrof_images_total = 0
-        nrof = np.zeros( (5,), dtype=np.int32)
-        for fimage in dataset:
-            if nrof_images_total%100==0:
-              print("Processing %d, (%s)" % (nrof_images_total, nrof))
-            nrof_images_total += 1
-            #if nrof_images_total<950000:
-            #  continue
-            image_path = fimage.image_path
-            if not os.path.exists(image_path):
-              print('image not found (%s)'%image_path)
-              continue
-            filename = os.path.splitext(os.path.split(image_path)[1])[0]
-            #print(image_path)
-            try:
-                img = misc.imread(image_path)
-            except (IOError, ValueError, IndexError) as e:
-                errorMessage = '{}: {}'.format(image_path, e)
-                print(errorMessage)
-            else:
-                if img.ndim<2:
-                    print('Unable to align "%s", img dim error' % image_path)
-                    #text_file.write('%s\n' % (output_filename))
-                    continue
-                if img.ndim == 2:
-                    img = to_rgb(img)
-                img = img[:,:,0:3]
-                _paths = fimage.image_path.split('/')
-                a,b,c = _paths[-3], _paths[-2], _paths[-1]
-                target_dir = os.path.join(args.output_dir, a, b)
-                if not os.path.exists(target_dir):
-                  os.makedirs(target_dir)
-                target_file = os.path.join(target_dir, c)
-                warped = None
-                if fimage.landmark is not None:
-                  dst = fimage.landmark.astype(np.float32)
-
-                  tform = trans.SimilarityTransform()
-                  tform.estimate(dst, src[0:3,:]*1.5+image_size[0]*0.25)
-                  M = tform.params[0:2,:]
-                  warped0 = cv2.warpAffine(img,M,(image_size[1]*2,image_size[0]*2), borderValue = 0.0)
-                  _minsize = image_size[0]
-                  bounding_boxes, points = detect_face.detect_face(warped0, _minsize, pnet, rnet, onet, threshold, factor)
-                  if bounding_boxes.shape[0]>0:
-                    bindex = 0
-                    det = bounding_boxes[bindex,0:4]
-                    #points need to be transpose, points = points.reshape( (5,2) ).transpose()
-                    dst = points[:, bindex].reshape( (2,5) ).T
-                    tform = trans.SimilarityTransform()
-                    tform.estimate(dst, src)
-                    M = tform.params[0:2,:]
-                    warped = cv2.warpAffine(warped0,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                    nrof[0]+=1
-                #assert fimage.bbox is not None
-                if warped is None and fimage.bbox is not None:
-                  _minsize = img.shape[0]//4
-                  bounding_boxes, points = detect_face.detect_face(img, _minsize, pnet, rnet, onet, threshold, factor)
-                  if bounding_boxes.shape[0]>0:
-                    det = bounding_boxes[:,0:4]
-                    bindex = -1
-                    index2 = [0.0, 0]
-                    for i in xrange(det.shape[0]):
-                      _det = det[i]
-                      iou = IOU(fimage.bbox, _det)
-                      if iou>index2[0]:
-                        index2[0] = iou
-                        index2[1] = i
-                    if index2[0]>0.3:
-                      bindex = index2[1]
-                    if bindex>=0:
-                      dst = points[:, bindex].reshape( (2,5) ).T
-                      tform = trans.SimilarityTransform()
-                      tform.estimate(dst, src)
-                      M = tform.params[0:2,:]
-                      warped = cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                      nrof[1]+=1
-                      #print('1',target_file,index2[0])
-                if warped is None and fimage.bbox is not None:
-                  bb = fimage.bbox
-                  #croped = img[bb[1]:bb[3],bb[0]:bb[2],:]
-                  bounding_boxes, points = detect_face.detect_face_force(img, bb, pnet, rnet, onet)
-                  assert bounding_boxes.shape[0]==1
-                  _box = bounding_boxes[0]
-                  if _box[4]>=0.3:
-                    dst = points[:, 0].reshape( (2,5) ).T
-                    tform = trans.SimilarityTransform()
-                    tform.estimate(dst, src)
-                    M = tform.params[0:2,:]
-                    warped = cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue = 0.0)
-                    nrof[2]+=1
-                    #print('2',target_file)
-
-                if warped is None:
-                  roi = np.zeros( (4,), dtype=np.int32)
-                  roi[0] = int(img.shape[1]*0.06)
-                  roi[1] = int(img.shape[0]*0.06)
-                  roi[2] = img.shape[1]-roi[0]
-                  roi[3] = img.shape[0]-roi[1]
-                  if fimage.bbox is not None:
-                    bb = fimage.bbox
-                    h = bb[3]-bb[1]
-                    w = bb[2]-bb[0]
-                    x = bb[0]
-                    y = bb[1]
-                    #roi = np.copy(bb)
-                    _w = int( (float(h)/image_size[0])*image_size[1] )
-                    x += (w-_w)//2
-                    #x = min( max(0,x), img.shape[1] )
-                    x = max(0,x)
-                    xw = x+_w
-                    xw = min(xw, img.shape[1])
-                    roi = np.array( (x, y, xw, y+h), dtype=np.int32)
-                    nrof[3]+=1
-                  else:
-                    nrof[4]+=1
-                  #print('3',bb,roi,img.shape)
-                  #print('3',target_file)
-                  warped = img[roi[1]:roi[3],roi[0]:roi[2],:]
-                  #print(warped.shape)
-                  warped = cv2.resize(warped, (image_size[1], image_size[0]))
-                bgr = warped[...,::-1]
-                cv2.imwrite(target_file, bgr)
-                oline = '%d\t%s\t%d\n' % (1,target_file, int(fimage.classname))
-                text_file.write(oline)
-                            
-
-def parse_arguments(argv):
-    parser = argparse.ArgumentParser()
-    
-    parser.add_argument('--input-dir', type=str, help='Directory with unaligned images.')
-    parser.add_argument('--name', type=str, help='dataset name, can be facescrub, megaface, webface, celeb.')
-    parser.add_argument('--output-dir', type=str, help='Directory with aligned face thumbnails.')
-    #parser.add_argument('--image_size', type=str, help='Image size (height, width) in pixels.', default='112,112')
-    #parser.add_argument('--margin', type=int,
-    #    help='Margin for the crop around the bounding box (height, width) in pixels.', default=44)
-    return parser.parse_args(argv)
-
-if __name__ == '__main__':
-    main(parse_arguments(sys.argv[1:]))
-
diff --git a/src/align/detect_face.py b/src/align/detect_face.py
deleted file mode 100644
index 47af5f5e4..000000000
--- a/src/align/detect_face.py
+++ /dev/null
@@ -1,848 +0,0 @@
-""" Tensorflow implementation of the face detection / alignment algorithm found at
-https://github.com/kpzhang93/MTCNN_face_detection_alignment
-"""
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-from six import string_types, iteritems
-
-import numpy as np
-import tensorflow as tf
-#from math import floor
-import cv2
-import os
-
-def layer(op):
-    '''Decorator for composable network layers.'''
-
-    def layer_decorated(self, *args, **kwargs):
-        # Automatically set a name if not provided.
-        name = kwargs.setdefault('name', self.get_unique_name(op.__name__))
-        # Figure out the layer inputs.
-        if len(self.terminals) == 0:
-            raise RuntimeError('No input variables found for layer %s.' % name)
-        elif len(self.terminals) == 1:
-            layer_input = self.terminals[0]
-        else:
-            layer_input = list(self.terminals)
-        # Perform the operation and get the output.
-        layer_output = op(self, layer_input, *args, **kwargs)
-        # Add to layer LUT.
-        self.layers[name] = layer_output
-        # This output is now the input for the next layer.
-        self.feed(layer_output)
-        # Return self for chained calls.
-        return self
-
-    return layer_decorated
-
-class Network(object):
-
-    def __init__(self, inputs, trainable=True):
-        # The input nodes for this network
-        self.inputs = inputs
-        # The current list of terminal nodes
-        self.terminals = []
-        # Mapping from layer names to layers
-        self.layers = dict(inputs)
-        # If true, the resulting variables are set as trainable
-        self.trainable = trainable
-
-        self.setup()
-
-    def setup(self):
-        '''Construct the network. '''
-        raise NotImplementedError('Must be implemented by the subclass.')
-
-    def load(self, data_path, session, ignore_missing=False):
-        '''Load network weights.
-        data_path: The path to the numpy-serialized network weights
-        session: The current TensorFlow session
-        ignore_missing: If true, serialized weights for missing layers are ignored.
-        '''
-        data_dict = np.load(data_path, encoding='latin1').item() #pylint: disable=no-member
-
-        for op_name in data_dict:
-            with tf.variable_scope(op_name, reuse=True):
-                for param_name, data in iteritems(data_dict[op_name]):
-                    try:
-                        var = tf.get_variable(param_name)
-                        session.run(var.assign(data))
-                    except ValueError:
-                        if not ignore_missing:
-                            raise
-
-    def feed(self, *args):
-        '''Set the input(s) for the next operation by replacing the terminal nodes.
-        The arguments can be either layer names or the actual layers.
-        '''
-        assert len(args) != 0
-        self.terminals = []
-        for fed_layer in args:
-            if isinstance(fed_layer, string_types):
-                try:
-                    fed_layer = self.layers[fed_layer]
-                except KeyError:
-                    raise KeyError('Unknown layer name fed: %s' % fed_layer)
-            self.terminals.append(fed_layer)
-        return self
-
-    def get_output(self):
-        '''Returns the current network output.'''
-        return self.terminals[-1]
-
-    def get_unique_name(self, prefix):
-        '''Returns an index-suffixed unique name for the given prefix.
-        This is used for auto-generating layer names based on the type-prefix.
-        '''
-        ident = sum(t.startswith(prefix) for t, _ in self.layers.items()) + 1
-        return '%s_%d' % (prefix, ident)
-
-    def make_var(self, name, shape):
-        '''Creates a new TensorFlow variable.'''
-        return tf.get_variable(name, shape, trainable=self.trainable)
-
-    def validate_padding(self, padding):
-        '''Verifies that the padding is one of the supported ones.'''
-        assert padding in ('SAME', 'VALID')
-
-    @layer
-    def conv(self,
-             inp,
-             k_h,
-             k_w,
-             c_o,
-             s_h,
-             s_w,
-             name,
-             relu=True,
-             padding='SAME',
-             group=1,
-             biased=True):
-        # Verify that the padding is acceptable
-        self.validate_padding(padding)
-        # Get the number of channels in the input
-        c_i = int(inp.get_shape()[-1])
-        # Verify that the grouping parameter is valid
-        assert c_i % group == 0
-        assert c_o % group == 0
-        # Convolution for a given input and kernel
-        convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)
-        with tf.variable_scope(name) as scope:
-            kernel = self.make_var('weights', shape=[k_h, k_w, c_i // group, c_o])
-            # This is the common-case. Convolve the input without any further complications.
-            output = convolve(inp, kernel)
-            # Add the biases
-            if biased:
-                biases = self.make_var('biases', [c_o])
-                output = tf.nn.bias_add(output, biases)
-            if relu:
-                # ReLU non-linearity
-                output = tf.nn.relu(output, name=scope.name)
-            return output
-
-    @layer
-    def prelu(self, inp, name):
-        with tf.variable_scope(name):
-            i = int(inp.get_shape()[-1])
-            alpha = self.make_var('alpha', shape=(i,))
-            output = tf.nn.relu(inp) + tf.multiply(alpha, -tf.nn.relu(-inp))
-        return output
-
-    @layer
-    def max_pool(self, inp, k_h, k_w, s_h, s_w, name, padding='SAME'):
-        self.validate_padding(padding)
-        return tf.nn.max_pool(inp,
-                              ksize=[1, k_h, k_w, 1],
-                              strides=[1, s_h, s_w, 1],
-                              padding=padding,
-                              name=name)
-
-    @layer
-    def fc(self, inp, num_out, name, relu=True):
-        with tf.variable_scope(name):
-            input_shape = inp.get_shape()
-            if input_shape.ndims == 4:
-                # The input is spatial. Vectorize it first.
-                dim = 1
-                for d in input_shape[1:].as_list():
-                    dim *= int(d)
-                feed_in = tf.reshape(inp, [-1, dim])
-            else:
-                feed_in, dim = (inp, input_shape[-1].value)
-            weights = self.make_var('weights', shape=[dim, num_out])
-            biases = self.make_var('biases', [num_out])
-            op = tf.nn.relu_layer if relu else tf.nn.xw_plus_b
-            fc = op(feed_in, weights, biases, name=name)
-            return fc
-
-
-    """
-    Multi dimensional softmax,
-    refer to https://github.com/tensorflow/tensorflow/issues/210
-    compute softmax along the dimension of target
-    the native softmax only supports batch_size x dimension
-    """
-    @layer
-    def softmax(self, target, axis, name=None):
-        max_axis = tf.reduce_max(target, axis, keep_dims=True)
-        target_exp = tf.exp(target-max_axis)
-        normalize = tf.reduce_sum(target_exp, axis, keep_dims=True)
-        softmax = tf.div(target_exp, normalize, name)
-        return softmax
-    
-class PNet(Network):
-    def setup(self):
-        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
-             .conv(3, 3, 10, 1, 1, padding='VALID', relu=False, name='conv1')
-             .prelu(name='PReLU1')
-             .max_pool(2, 2, 2, 2, name='pool1')
-             .conv(3, 3, 16, 1, 1, padding='VALID', relu=False, name='conv2')
-             .prelu(name='PReLU2')
-             .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv3')
-             .prelu(name='PReLU3')
-             .conv(1, 1, 2, 1, 1, relu=False, name='conv4-1')
-             .softmax(3,name='prob1'))
-
-        (self.feed('PReLU3') #pylint: disable=no-value-for-parameter
-             .conv(1, 1, 4, 1, 1, relu=False, name='conv4-2'))
-        
-class RNet(Network):
-    def setup(self):
-        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
-             .conv(3, 3, 28, 1, 1, padding='VALID', relu=False, name='conv1')
-             .prelu(name='prelu1')
-             .max_pool(3, 3, 2, 2, name='pool1')
-             .conv(3, 3, 48, 1, 1, padding='VALID', relu=False, name='conv2')
-             .prelu(name='prelu2')
-             .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
-             .conv(2, 2, 64, 1, 1, padding='VALID', relu=False, name='conv3')
-             .prelu(name='prelu3')
-             .fc(128, relu=False, name='conv4')
-             .prelu(name='prelu4')
-             .fc(2, relu=False, name='conv5-1')
-             .softmax(1,name='prob1'))
-
-        (self.feed('prelu4') #pylint: disable=no-value-for-parameter
-             .fc(4, relu=False, name='conv5-2'))
-
-class ONet(Network):
-    def setup(self):
-        (self.feed('data') #pylint: disable=no-value-for-parameter, no-member
-             .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv1')
-             .prelu(name='prelu1')
-             .max_pool(3, 3, 2, 2, name='pool1')
-             .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv2')
-             .prelu(name='prelu2')
-             .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')
-             .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv3')
-             .prelu(name='prelu3')
-             .max_pool(2, 2, 2, 2, name='pool3')
-             .conv(2, 2, 128, 1, 1, padding='VALID', relu=False, name='conv4')
-             .prelu(name='prelu4')
-             .fc(256, relu=False, name='conv5')
-             .prelu(name='prelu5')
-             .fc(2, relu=False, name='conv6-1')
-             .softmax(1, name='prob1'))
-
-        (self.feed('prelu5') #pylint: disable=no-value-for-parameter
-             .fc(4, relu=False, name='conv6-2'))
-
-        (self.feed('prelu5') #pylint: disable=no-value-for-parameter
-             .fc(10, relu=False, name='conv6-3'))
-
-def create_mtcnn(sess, model_path):
-    if not model_path:
-        model_path,_ = os.path.split(os.path.realpath(__file__))
-
-    with tf.variable_scope('pnet'):
-        data = tf.placeholder(tf.float32, (None,None,None,3), 'input')
-        pnet = PNet({'data':data})
-        pnet.load(os.path.join(model_path, 'det1.npy'), sess)
-    with tf.variable_scope('rnet'):
-        data = tf.placeholder(tf.float32, (None,24,24,3), 'input')
-        rnet = RNet({'data':data})
-        rnet.load(os.path.join(model_path, 'det2.npy'), sess)
-    with tf.variable_scope('onet'):
-        data = tf.placeholder(tf.float32, (None,48,48,3), 'input')
-        onet = ONet({'data':data})
-        onet.load(os.path.join(model_path, 'det3.npy'), sess)
-        
-    pnet_fun = lambda img : sess.run(('pnet/conv4-2/BiasAdd:0', 'pnet/prob1:0'), feed_dict={'pnet/input:0':img})
-    rnet_fun = lambda img : sess.run(('rnet/conv5-2/conv5-2:0', 'rnet/prob1:0'), feed_dict={'rnet/input:0':img})
-    onet_fun = lambda img : sess.run(('onet/conv6-2/conv6-2:0', 'onet/conv6-3/conv6-3:0', 'onet/prob1:0'), feed_dict={'onet/input:0':img})
-    return pnet_fun, rnet_fun, onet_fun
-
-def detect_face(img, minsize, pnet, rnet, onet, threshold, factor):
-    # im: input image
-    # minsize: minimum of faces' size
-    # pnet, rnet, onet: caffemodel
-    # threshold: threshold=[th1 th2 th3], th1-3 are three steps's threshold
-    # fastresize: resize img from last scale (using in high-resolution images) if fastresize==true
-    factor_count=0
-    total_boxes=np.empty((0,9))
-    points=[]
-    h=img.shape[0]
-    w=img.shape[1]
-    minl=np.amin([h, w])
-    m=12.0/minsize
-    minl=minl*m
-    # creat scale pyramid
-    scales=[]
-    while minl>=12:
-        scales += [m*np.power(factor, factor_count)]
-        minl = minl*factor
-        factor_count += 1
-
-    # first stage
-    for j in range(len(scales)):
-        scale=scales[j]
-        hs=int(np.ceil(h*scale))
-        ws=int(np.ceil(w*scale))
-        im_data = imresample(img, (hs, ws))
-        im_data = (im_data-127.5)*0.0078125
-        img_x = np.expand_dims(im_data, 0)
-        img_y = np.transpose(img_x, (0,2,1,3))
-        out = pnet(img_y)
-        out0 = np.transpose(out[0], (0,2,1,3))
-        out1 = np.transpose(out[1], (0,2,1,3))
-        
-        boxes, _ = generateBoundingBox(out1[0,:,:,1].copy(), out0[0,:,:,:].copy(), scale, threshold[0])
-        
-        # inter-scale nms
-        pick = nms(boxes.copy(), 0.5, 'Union')
-        if boxes.size>0 and pick.size>0:
-            boxes = boxes[pick,:]
-            total_boxes = np.append(total_boxes, boxes, axis=0)
-
-    numbox = total_boxes.shape[0]
-    if numbox>0:
-        pick = nms(total_boxes.copy(), 0.7, 'Union')
-        total_boxes = total_boxes[pick,:]
-        regw = total_boxes[:,2]-total_boxes[:,0]
-        regh = total_boxes[:,3]-total_boxes[:,1]
-        qq1 = total_boxes[:,0]+total_boxes[:,5]*regw
-        qq2 = total_boxes[:,1]+total_boxes[:,6]*regh
-        qq3 = total_boxes[:,2]+total_boxes[:,7]*regw
-        qq4 = total_boxes[:,3]+total_boxes[:,8]*regh
-        total_boxes = np.transpose(np.vstack([qq1, qq2, qq3, qq4, total_boxes[:,4]]))
-        total_boxes = rerec(total_boxes.copy())
-        total_boxes[:,0:4] = np.fix(total_boxes[:,0:4]).astype(np.int32)
-        dy,edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
-
-    numbox = total_boxes.shape[0]
-    if numbox>0:
-        # second stage
-        tempimg = np.zeros((24,24,3,numbox))
-        for k in range(0,numbox):
-            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
-            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
-            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
-                tempimg[:,:,:,k] = imresample(tmp, (24, 24))
-            else:
-                return np.empty()
-        tempimg = (tempimg-127.5)*0.0078125
-        tempimg1 = np.transpose(tempimg, (3,1,0,2))
-        out = rnet(tempimg1)
-        out0 = np.transpose(out[0])
-        out1 = np.transpose(out[1])
-        score = out1[1,:]
-        ipass = np.where(score>threshold[1])
-        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
-        mv = out0[:,ipass[0]]
-        if total_boxes.shape[0]>0:
-            pick = nms(total_boxes, 0.7, 'Union')
-            total_boxes = total_boxes[pick,:]
-            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv[:,pick]))
-            total_boxes = rerec(total_boxes.copy())
-
-    numbox = total_boxes.shape[0]
-    if numbox>0:
-        # third stage
-        total_boxes = np.fix(total_boxes).astype(np.int32)
-        dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
-        tempimg = np.zeros((48,48,3,numbox))
-        for k in range(0,numbox):
-            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
-            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
-            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
-                tempimg[:,:,:,k] = imresample(tmp, (48, 48))
-            else:
-                return np.empty()
-        tempimg = (tempimg-127.5)*0.0078125
-        tempimg1 = np.transpose(tempimg, (3,1,0,2))
-        out = onet(tempimg1)
-        out0 = np.transpose(out[0])
-        out1 = np.transpose(out[1])
-        out2 = np.transpose(out[2])
-        score = out2[1,:]
-        points = out1
-        ipass = np.where(score>threshold[2])
-        points = points[:,ipass[0]]
-        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
-        mv = out0[:,ipass[0]]
-
-        w = total_boxes[:,2]-total_boxes[:,0]+1
-        h = total_boxes[:,3]-total_boxes[:,1]+1
-        points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1
-        points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1
-        if total_boxes.shape[0]>0:
-            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv))
-            pick = nms(total_boxes.copy(), 0.7, 'Min')
-            total_boxes = total_boxes[pick,:]
-            points = points[:,pick]
-                
-    return total_boxes, points
-
-def detect_face_force(img, bbox, pnet, rnet, onet):
-    total_boxes = np.zeros( (1,5), dtype=np.float32)
-    total_boxes[0,0:4] = bbox
-    threshold = [0.0,0.0,0.0]
-    h=img.shape[0]
-    w=img.shape[1]
-    numbox = total_boxes.shape[0]
-    if numbox>0:
-        dy,edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
-        # second stage
-        tempimg = np.zeros((24,24,3,numbox))
-        for k in range(0,numbox):
-            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
-            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
-            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
-                tempimg[:,:,:,k] = imresample(tmp, (24, 24))
-            else:
-                return np.empty()
-        tempimg = (tempimg-127.5)*0.0078125
-        tempimg1 = np.transpose(tempimg, (3,1,0,2))
-        out = rnet(tempimg1)
-        out0 = np.transpose(out[0])
-        out1 = np.transpose(out[1])
-        score = out1[1,:]
-        ipass = np.where(score>threshold[1])
-        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
-        mv = out0[:,ipass[0]]
-        if total_boxes.shape[0]>0:
-            pick = nms(total_boxes, 0.7, 'Union')
-            total_boxes = total_boxes[pick,:]
-            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv[:,pick]))
-            total_boxes = rerec(total_boxes.copy())
-
-    numbox = total_boxes.shape[0]
-    if numbox>0:
-        # third stage
-        total_boxes = np.fix(total_boxes).astype(np.int32)
-        dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)
-        tempimg = np.zeros((48,48,3,numbox))
-        for k in range(0,numbox):
-            tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))
-            tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]
-            if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:
-                tempimg[:,:,:,k] = imresample(tmp, (48, 48))
-            else:
-                return np.empty()
-        tempimg = (tempimg-127.5)*0.0078125
-        tempimg1 = np.transpose(tempimg, (3,1,0,2))
-        out = onet(tempimg1)
-        out0 = np.transpose(out[0])
-        out1 = np.transpose(out[1])
-        out2 = np.transpose(out[2])
-        score = out2[1,:]
-        points = out1
-        ipass = np.where(score>threshold[2])
-        points = points[:,ipass[0]]
-        total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])
-        mv = out0[:,ipass[0]]
-
-        w = total_boxes[:,2]-total_boxes[:,0]+1
-        h = total_boxes[:,3]-total_boxes[:,1]+1
-        points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1
-        points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1
-        if total_boxes.shape[0]>0:
-            total_boxes = bbreg(total_boxes.copy(), np.transpose(mv))
-            pick = nms(total_boxes.copy(), 0.7, 'Min')
-            total_boxes = total_boxes[pick,:]
-            points = points[:,pick]
-                
-    return total_boxes, points
-
-def bulk_detect_face(images, detection_window_size_ratio, pnet, rnet, onet, threshold, factor):
-    # im: input image
-    # minsize: minimum of faces' size
-    # pnet, rnet, onet: caffemodel
-    # threshold: threshold=[th1 th2 th3], th1-3 are three steps's threshold [0-1]
-
-    all_scales = [None] * len(images)
-    images_with_boxes = [None] * len(images)
-
-    for i in range(len(images)):
-        images_with_boxes[i] = {'total_boxes': np.empty((0, 9))}
-
-    # create scale pyramid
-    for index, img in enumerate(images):
-        all_scales[index] = []
-        h = img.shape[0]
-        w = img.shape[1]
-        minsize = int(detection_window_size_ratio * np.minimum(w, h))
-        factor_count = 0
-        minl = np.amin([h, w])
-        if minsize <= 12:
-            minsize = 12
-
-        m = 12.0 / minsize
-        minl = minl * m
-        while minl >= 12:
-            all_scales[index].append(m * np.power(factor, factor_count))
-            minl = minl * factor
-            factor_count += 1
-
-    # # # # # # # # # # # # #
-    # first stage - fast proposal network (pnet) to obtain face candidates
-    # # # # # # # # # # # # #
-
-    images_obj_per_resolution = {}
-
-    # TODO: use some type of rounding to number module 8 to increase probability that pyramid images will have the same resolution across input images
-
-    for index, scales in enumerate(all_scales):
-        h = images[index].shape[0]
-        w = images[index].shape[1]
-
-        for scale in scales:
-            hs = int(np.ceil(h * scale))
-            ws = int(np.ceil(w * scale))
-
-            if (ws, hs) not in images_obj_per_resolution:
-                images_obj_per_resolution[(ws, hs)] = []
-
-            im_data = imresample(images[index], (hs, ws))
-            im_data = (im_data - 127.5) * 0.0078125
-            img_y = np.transpose(im_data, (1, 0, 2))  # caffe uses different dimensions ordering
-            images_obj_per_resolution[(ws, hs)].append({'scale': scale, 'image': img_y, 'index': index})
-
-    for resolution in images_obj_per_resolution:
-        images_per_resolution = [i['image'] for i in images_obj_per_resolution[resolution]]
-        outs = pnet(images_per_resolution)
-
-        for index in range(len(outs[0])):
-            scale = images_obj_per_resolution[resolution][index]['scale']
-            image_index = images_obj_per_resolution[resolution][index]['index']
-            out0 = np.transpose(outs[0][index], (1, 0, 2))
-            out1 = np.transpose(outs[1][index], (1, 0, 2))
-
-            boxes, _ = generateBoundingBox(out1[:, :, 1].copy(), out0[:, :, :].copy(), scale, threshold[0])
-
-            # inter-scale nms
-            pick = nms(boxes.copy(), 0.5, 'Union')
-            if boxes.size > 0 and pick.size > 0:
-                boxes = boxes[pick, :]
-                images_with_boxes[image_index]['total_boxes'] = np.append(images_with_boxes[image_index]['total_boxes'],
-                                                                          boxes,
-                                                                          axis=0)
-
-    for index, image_obj in enumerate(images_with_boxes):
-        numbox = image_obj['total_boxes'].shape[0]
-        if numbox > 0:
-            h = images[index].shape[0]
-            w = images[index].shape[1]
-            pick = nms(image_obj['total_boxes'].copy(), 0.7, 'Union')
-            image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]
-            regw = image_obj['total_boxes'][:, 2] - image_obj['total_boxes'][:, 0]
-            regh = image_obj['total_boxes'][:, 3] - image_obj['total_boxes'][:, 1]
-            qq1 = image_obj['total_boxes'][:, 0] + image_obj['total_boxes'][:, 5] * regw
-            qq2 = image_obj['total_boxes'][:, 1] + image_obj['total_boxes'][:, 6] * regh
-            qq3 = image_obj['total_boxes'][:, 2] + image_obj['total_boxes'][:, 7] * regw
-            qq4 = image_obj['total_boxes'][:, 3] + image_obj['total_boxes'][:, 8] * regh
-            image_obj['total_boxes'] = np.transpose(np.vstack([qq1, qq2, qq3, qq4, image_obj['total_boxes'][:, 4]]))
-            image_obj['total_boxes'] = rerec(image_obj['total_boxes'].copy())
-            image_obj['total_boxes'][:, 0:4] = np.fix(image_obj['total_boxes'][:, 0:4]).astype(np.int32)
-            dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(image_obj['total_boxes'].copy(), w, h)
-
-            numbox = image_obj['total_boxes'].shape[0]
-            tempimg = np.zeros((24, 24, 3, numbox))
-
-            if numbox > 0:
-                for k in range(0, numbox):
-                    tmp = np.zeros((int(tmph[k]), int(tmpw[k]), 3))
-                    tmp[dy[k] - 1:edy[k], dx[k] - 1:edx[k], :] = images[index][y[k] - 1:ey[k], x[k] - 1:ex[k], :]
-                    if tmp.shape[0] > 0 and tmp.shape[1] > 0 or tmp.shape[0] == 0 and tmp.shape[1] == 0:
-                        tempimg[:, :, :, k] = imresample(tmp, (24, 24))
-                    else:
-                        return np.empty()
-
-                tempimg = (tempimg - 127.5) * 0.0078125
-                image_obj['rnet_input'] = np.transpose(tempimg, (3, 1, 0, 2))
-
-    # # # # # # # # # # # # #
-    # second stage - refinement of face candidates with rnet
-    # # # # # # # # # # # # #
-
-    bulk_rnet_input = np.empty((0, 24, 24, 3))
-    for index, image_obj in enumerate(images_with_boxes):
-        if 'rnet_input' in image_obj:
-            bulk_rnet_input = np.append(bulk_rnet_input, image_obj['rnet_input'], axis=0)
-
-    out = rnet(bulk_rnet_input)
-    out0 = np.transpose(out[0])
-    out1 = np.transpose(out[1])
-    score = out1[1, :]
-
-    i = 0
-    for index, image_obj in enumerate(images_with_boxes):
-        if 'rnet_input' not in image_obj:
-            continue
-
-        rnet_input_count = image_obj['rnet_input'].shape[0]
-        score_per_image = score[i:i + rnet_input_count]
-        out0_per_image = out0[:, i:i + rnet_input_count]
-
-        ipass = np.where(score_per_image > threshold[1])
-        image_obj['total_boxes'] = np.hstack([image_obj['total_boxes'][ipass[0], 0:4].copy(),
-                                              np.expand_dims(score_per_image[ipass].copy(), 1)])
-
-        mv = out0_per_image[:, ipass[0]]
-
-        if image_obj['total_boxes'].shape[0] > 0:
-            h = images[index].shape[0]
-            w = images[index].shape[1]
-            pick = nms(image_obj['total_boxes'], 0.7, 'Union')
-            image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]
-            image_obj['total_boxes'] = bbreg(image_obj['total_boxes'].copy(), np.transpose(mv[:, pick]))
-            image_obj['total_boxes'] = rerec(image_obj['total_boxes'].copy())
-
-            numbox = image_obj['total_boxes'].shape[0]
-
-            if numbox > 0:
-                tempimg = np.zeros((48, 48, 3, numbox))
-                image_obj['total_boxes'] = np.fix(image_obj['total_boxes']).astype(np.int32)
-                dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(image_obj['total_boxes'].copy(), w, h)
-
-                for k in range(0, numbox):
-                    tmp = np.zeros((int(tmph[k]), int(tmpw[k]), 3))
-                    tmp[dy[k] - 1:edy[k], dx[k] - 1:edx[k], :] = images[index][y[k] - 1:ey[k], x[k] - 1:ex[k], :]
-                    if tmp.shape[0] > 0 and tmp.shape[1] > 0 or tmp.shape[0] == 0 and tmp.shape[1] == 0:
-                        tempimg[:, :, :, k] = imresample(tmp, (48, 48))
-                    else:
-                        return np.empty()
-                tempimg = (tempimg - 127.5) * 0.0078125
-                image_obj['onet_input'] = np.transpose(tempimg, (3, 1, 0, 2))
-
-        i += rnet_input_count
-
-    # # # # # # # # # # # # #
-    # third stage - further refinement and facial landmarks positions with onet
-    # # # # # # # # # # # # #
-
-    bulk_onet_input = np.empty((0, 48, 48, 3))
-    for index, image_obj in enumerate(images_with_boxes):
-        if 'onet_input' in image_obj:
-            bulk_onet_input = np.append(bulk_onet_input, image_obj['onet_input'], axis=0)
-
-    out = onet(bulk_onet_input)
-
-    out0 = np.transpose(out[0])
-    out1 = np.transpose(out[1])
-    out2 = np.transpose(out[2])
-    score = out2[1, :]
-    points = out1
-
-    i = 0
-    ret = []
-    for index, image_obj in enumerate(images_with_boxes):
-        if 'onet_input' not in image_obj:
-            ret.append(None)
-            continue
-
-        onet_input_count = image_obj['onet_input'].shape[0]
-
-        out0_per_image = out0[:, i:i + onet_input_count]
-        score_per_image = score[i:i + onet_input_count]
-        points_per_image = points[:, i:i + onet_input_count]
-
-        ipass = np.where(score_per_image > threshold[2])
-        points_per_image = points_per_image[:, ipass[0]]
-
-        image_obj['total_boxes'] = np.hstack([image_obj['total_boxes'][ipass[0], 0:4].copy(),
-                                              np.expand_dims(score_per_image[ipass].copy(), 1)])
-        mv = out0_per_image[:, ipass[0]]
-
-        w = image_obj['total_boxes'][:, 2] - image_obj['total_boxes'][:, 0] + 1
-        h = image_obj['total_boxes'][:, 3] - image_obj['total_boxes'][:, 1] + 1
-        points_per_image[0:5, :] = np.tile(w, (5, 1)) * points_per_image[0:5, :] + np.tile(
-            image_obj['total_boxes'][:, 0], (5, 1)) - 1
-        points_per_image[5:10, :] = np.tile(h, (5, 1)) * points_per_image[5:10, :] + np.tile(
-            image_obj['total_boxes'][:, 1], (5, 1)) - 1
-
-        if image_obj['total_boxes'].shape[0] > 0:
-            image_obj['total_boxes'] = bbreg(image_obj['total_boxes'].copy(), np.transpose(mv))
-            pick = nms(image_obj['total_boxes'].copy(), 0.7, 'Min')
-            image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]
-            points_per_image = points_per_image[:, pick]
-
-            ret.append((image_obj['total_boxes'], points_per_image))
-        else:
-            ret.append(None)
-
-        i += onet_input_count
-
-    return ret
-
-
-# function [boundingbox] = bbreg(boundingbox,reg)
-def bbreg(boundingbox,reg):
-    # calibrate bounding boxes
-    if reg.shape[1]==1:
-        reg = np.reshape(reg, (reg.shape[2], reg.shape[3]))
-
-    w = boundingbox[:,2]-boundingbox[:,0]+1
-    h = boundingbox[:,3]-boundingbox[:,1]+1
-    b1 = boundingbox[:,0]+reg[:,0]*w
-    b2 = boundingbox[:,1]+reg[:,1]*h
-    b3 = boundingbox[:,2]+reg[:,2]*w
-    b4 = boundingbox[:,3]+reg[:,3]*h
-    boundingbox[:,0:4] = np.transpose(np.vstack([b1, b2, b3, b4 ]))
-    return boundingbox
- 
-def generateBoundingBox(imap, reg, scale, t):
-    # use heatmap to generate bounding boxes
-    stride=2
-    cellsize=12
-
-    imap = np.transpose(imap)
-    dx1 = np.transpose(reg[:,:,0])
-    dy1 = np.transpose(reg[:,:,1])
-    dx2 = np.transpose(reg[:,:,2])
-    dy2 = np.transpose(reg[:,:,3])
-    y, x = np.where(imap >= t)
-    if y.shape[0]==1:
-        dx1 = np.flipud(dx1)
-        dy1 = np.flipud(dy1)
-        dx2 = np.flipud(dx2)
-        dy2 = np.flipud(dy2)
-    score = imap[(y,x)]
-    reg = np.transpose(np.vstack([ dx1[(y,x)], dy1[(y,x)], dx2[(y,x)], dy2[(y,x)] ]))
-    if reg.size==0:
-        reg = np.empty((0,3))
-    bb = np.transpose(np.vstack([y,x]))
-    q1 = np.fix((stride*bb+1)/scale)
-    q2 = np.fix((stride*bb+cellsize-1+1)/scale)
-    boundingbox = np.hstack([q1, q2, np.expand_dims(score,1), reg])
-    return boundingbox, reg
- 
-# function pick = nms(boxes,threshold,type)
-def nms(boxes, threshold, method):
-    if boxes.size==0:
-        return np.empty((0,3))
-    x1 = boxes[:,0]
-    y1 = boxes[:,1]
-    x2 = boxes[:,2]
-    y2 = boxes[:,3]
-    s = boxes[:,4]
-    area = (x2-x1+1) * (y2-y1+1)
-    I = np.argsort(s)
-    pick = np.zeros_like(s, dtype=np.int16)
-    counter = 0
-    while I.size>0:
-        i = I[-1]
-        pick[counter] = i
-        counter += 1
-        idx = I[0:-1]
-        xx1 = np.maximum(x1[i], x1[idx])
-        yy1 = np.maximum(y1[i], y1[idx])
-        xx2 = np.minimum(x2[i], x2[idx])
-        yy2 = np.minimum(y2[i], y2[idx])
-        w = np.maximum(0.0, xx2-xx1+1)
-        h = np.maximum(0.0, yy2-yy1+1)
-        inter = w * h
-        if method is 'Min':
-            o = inter / np.minimum(area[i], area[idx])
-        else:
-            o = inter / (area[i] + area[idx] - inter)
-        I = I[np.where(o<=threshold)]
-    pick = pick[0:counter]
-    return pick
-
-# function [dy edy dx edx y ey x ex tmpw tmph] = pad(total_boxes,w,h)
-def pad(total_boxes, w, h):
-    # compute the padding coordinates (pad the bounding boxes to square)
-    tmpw = (total_boxes[:,2]-total_boxes[:,0]+1).astype(np.int32)
-    tmph = (total_boxes[:,3]-total_boxes[:,1]+1).astype(np.int32)
-    numbox = total_boxes.shape[0]
-
-    dx = np.ones((numbox), dtype=np.int32)
-    dy = np.ones((numbox), dtype=np.int32)
-    edx = tmpw.copy().astype(np.int32)
-    edy = tmph.copy().astype(np.int32)
-
-    x = total_boxes[:,0].copy().astype(np.int32)
-    y = total_boxes[:,1].copy().astype(np.int32)
-    ex = total_boxes[:,2].copy().astype(np.int32)
-    ey = total_boxes[:,3].copy().astype(np.int32)
-
-    tmp = np.where(ex>w)
-    edx.flat[tmp] = np.expand_dims(-ex[tmp]+w+tmpw[tmp],1)
-    ex[tmp] = w
-    
-    tmp = np.where(ey>h)
-    edy.flat[tmp] = np.expand_dims(-ey[tmp]+h+tmph[tmp],1)
-    ey[tmp] = h
-
-    tmp = np.where(x<1)
-    dx.flat[tmp] = np.expand_dims(2-x[tmp],1)
-    x[tmp] = 1
-
-    tmp = np.where(y<1)
-    dy.flat[tmp] = np.expand_dims(2-y[tmp],1)
-    y[tmp] = 1
-    
-    return dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph
-
-# function [bboxA] = rerec(bboxA)
-def rerec(bboxA):
-    # convert bboxA to square
-    h = bboxA[:,3]-bboxA[:,1]
-    w = bboxA[:,2]-bboxA[:,0]
-    l = np.maximum(w, h)
-    bboxA[:,0] = bboxA[:,0]+w*0.5-l*0.5
-    bboxA[:,1] = bboxA[:,1]+h*0.5-l*0.5
-    bboxA[:,2:4] = bboxA[:,0:2] + np.transpose(np.tile(l,(2,1)))
-    return bboxA
-
-def imresample(img, sz):
-    im_data = cv2.resize(img, (sz[1], sz[0]), interpolation=cv2.INTER_AREA) #@UndefinedVariable
-    return im_data
-
-    # This method is kept for debugging purpose
-#     h=img.shape[0]
-#     w=img.shape[1]
-#     hs, ws = sz
-#     dx = float(w) / ws
-#     dy = float(h) / hs
-#     im_data = np.zeros((hs,ws,3))
-#     for a1 in range(0,hs):
-#         for a2 in range(0,ws):
-#             for a3 in range(0,3):
-#                 im_data[a1,a2,a3] = img[int(floor(a1*dy)),int(floor(a2*dx)),a3]
-#     return im_data
-
diff --git a/src/api/app.py b/src/api/app.py
deleted file mode 100644
index 719af827b..000000000
--- a/src/api/app.py
+++ /dev/null
@@ -1,99 +0,0 @@
-from flask import Flask
-import face_model
-import argparse
-import json
-import base64
-#import requests
-import numpy as np
-import urllib
-import cv2
-from flask import Flask, render_template, request, jsonify
-
-
-parser = argparse.ArgumentParser(description='do verification')
-# general
-parser.add_argument('--image-size', default='112,112', help='')
-parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-parser.add_argument('--threshold', default=1.24, type=float, help='ver dist threshold')
-args = parser.parse_args()
-
-model = face_model.FaceModel(args)
-
-app = Flask(__name__)
-
-@app.route('/')
-def hello_world():
-  return 'Hello, This is InsightFace!'
-
-def image_resize(image):
-  m = min(image.shape[0], image.shape[1])
-  f = 640.0/m
-  if f<1.0:
-    image = cv2.resize(image, (int(image.shape[1]*f), int(image.shape[0]*f)))
-  return image
-
-def get_image(data):
-  image = None
-  if 'url' in data:
-    url = data['url']
-    if url.startswith('http'):
-      resp = urllib.urlopen(url)
-      image = np.asarray(bytearray(resp.read()), dtype="uint8")
-      image = cv2.imdecode(image, cv2.IMREAD_COLOR)
-    else:
-      image = cv2.imread(url, cv2.IMREAD_COLOR)
-    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-    image = image_resize(image)
-  elif 'data' in data:
-    _bin = data['data']
-    if _bin is not None:
-      if not isinstance(_bin, list):
-        _bin = base64.b64decode(_bin)
-        _bin = np.fromstring(_bin, np.uint8)
-        image = cv2.imdecode(_bin, cv2.IMREAD_COLOR)
-        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-        image = image_resize(image)
-      else:
-        image = []
-        for __bin in _bin:
-          __bin = base64.b64decode(__bin)
-          __bin = np.fromstring(__bin, np.uint8)
-          _image = cv2.imdecode(__bin, cv2.IMREAD_COLOR)
-          _image = cv2.cvtColor(_image, cv2.COLOR_BGR2RGB)
-          _image = image_resize(_image)
-          image.append(_image)
-
-  return image
-
-@app.route('/ver', methods=['POST'])
-def ver():
-  try:
-    data = request.data
-    values = json.loads(data)
-    source_image = get_image(values['source'])
-    if source_image is None:
-      print('source image is None')
-      return '-1'
-    assert not isinstance(source_image, list)
-    print(source_image.shape)
-    target_image = get_image(values['target'])
-    if target_image is None:
-      print('target image is None')
-      return '-1'
-    #print(target_image.shape)
-    if not isinstance(target_image, list):
-      target_image = [target_image]
-    #print('before call')
-    #ret = model.is_same_id(source_image, target_image)
-    ret = model.sim(source_image, target_image)
-  except Exception as ex:
-    print(ex)
-    return '-1'
-
-  #return str(int(ret))
-  print('sim', ret)
-  return "%1.3f"%ret
-
-if __name__ == '__main__':
-    app.run('0.0.0.0', port=18080, debug=False)
diff --git a/src/api/face_model.py b/src/api/face_model.py
deleted file mode 100644
index 0d6a251d8..000000000
--- a/src/api/face_model.py
+++ /dev/null
@@ -1,203 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from scipy import misc
-import sys
-import os
-import argparse
-import tensorflow as tf
-import numpy as np
-import mxnet as mx
-import random
-import sklearn
-from sklearn.decomposition import PCA
-from time import sleep
-from easydict import EasyDict as edict
-#import facenet
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'align'))
-import detect_face
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-import face_preprocess
-
-def ch_dev(arg_params, aux_params, ctx):
-  new_args = dict()
-  new_auxs = dict()
-  for k, v in arg_params.items():
-    new_args[k] = v.as_in_context(ctx)
-  for k, v in aux_params.items():
-    new_auxs[k] = v.as_in_context(ctx)
-  return new_args, new_auxs
-
-def do_flip(data):
-  for idx in xrange(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-class FaceModel:
-  def __init__(self, args):
-    model = edict()
-    with tf.Graph().as_default():
-      config = tf.ConfigProto()
-      config.gpu_options.per_process_gpu_memory_fraction = 0.2
-      sess = tf.Session(config=config)
-      #sess = tf.Session()
-      with sess.as_default():
-        self.pnet, self.rnet, self.onet = detect_face.create_mtcnn(sess, None)
-
-    self.threshold = args.threshold
-    self.det_minsize = 50
-    self.det_threshold = [0.4,0.6,0.6]
-    self.det_factor = 0.9
-    _vec = args.image_size.split(',')
-    assert len(_vec)==2
-    self.image_size = (int(_vec[0]), int(_vec[1]))
-    _vec = args.model.split(',')
-    assert len(_vec)==2
-    prefix = _vec[0]
-    epoch = int(_vec[1])
-    print('loading',prefix, epoch)
-    self.model = edict()
-    self.model.ctx = mx.gpu(args.gpu)
-    self.model.sym, self.model.arg_params, self.model.aux_params = mx.model.load_checkpoint(prefix, epoch)
-    self.model.arg_params, self.model.aux_params = ch_dev(self.model.arg_params, self.model.aux_params, self.model.ctx)
-    all_layers = self.model.sym.get_internals()
-    self.model.sym = all_layers['fc1_output']
-
-  def get_aligned_face(self, img, force = False):
-    #print('before det', img.shape)
-    bounding_boxes, points = detect_face.detect_face(img, self.det_minsize, self.pnet, self.rnet, self.onet, self.det_threshold, self.det_factor)
-    #if bounding_boxes.shape[0]==0:
-    #  fimg = np.copy(img)
-    #  do_flip(fimg)
-    #  bounding_boxes, points = detect_face.detect_face(fimg, self.det_minsize, self.pnet, self.rnet, self.onet, self.det_threshold, self.det_factor)
-    if bounding_boxes.shape[0]==0 and force:
-      print('force det', img.shape)
-      bounding_boxes, points = detect_face.detect_face(img, self.det_minsize, self.pnet, self.rnet, self.onet, [0.3, 0.3, 0.1], self.det_factor)
-      #bounding_boxes, points = detect_face.detect_face_force(img, None, self.pnet, self.rnet, self.onet)
-    #print('after det')
-    if bounding_boxes.shape[0]==0:
-      return None
-    bindex = 0
-    nrof_faces = bounding_boxes.shape[0]
-    det = bounding_boxes[:,0:4]
-    img_size = np.asarray(img.shape)[0:2]
-    if nrof_faces>1:
-      bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-      img_center = img_size / 2
-      offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-      offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-      bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-    det = bounding_boxes[:,0:4]
-    det = det[bindex,:]
-    points = points[:, bindex]
-    landmark = points.reshape((2,5)).T
-    #points need to be transpose, points = points.reshape( (5,2) ).transpose()
-    det = np.squeeze(det)
-    bb = det
-    points = list(points.flatten())
-    assert(len(points)==10)
-    str_image_size = "%d,%d"%(self.image_size[0], self.image_size[1])
-    warped = face_preprocess.preprocess(img, bbox=bb, landmark = landmark, image_size=str_image_size)
-    warped = np.transpose(warped, (2,0,1))
-    print(warped.shape)
-    return warped
-
-  def get_all_faces(self, img):
-    str_image_size = "%d,%d"%(self.image_size[0], self.image_size[1])
-    bounding_boxes, points = detect_face.detect_face(img, self.det_minsize, self.pnet, self.rnet, self.onet, self.det_threshold, self.det_factor)
-    ret = []
-    for i in xrange(bounding_boxes.shape[0]):
-      bbox = bounding_boxes[i,0:4]
-      landmark = points[:, i].reshape((2,5)).T
-      aligned = face_preprocess.preprocess(img, bbox=bbox, landmark = landmark, image_size=str_image_size)
-      aligned = np.transpose(aligned, (2,0,1))
-      ret.append(aligned)
-    return ret
-
-  def get_feature_impl(self, face_img, norm):
-    embedding = None
-    for flipid in [0,1]:
-      _img = np.copy(face_img)
-      if flipid==1:
-        do_flip(_img)
-      #nimg = np.zeros(_img.shape, dtype=np.float32)
-      #nimg[:,ppatch[1]:ppatch[3],ppatch[0]:ppatch[2]] = _img[:, ppatch[1]:ppatch[3], ppatch[0]:ppatch[2]]
-      #_img = nimg
-      input_blob = np.expand_dims(_img, axis=0)
-      self.model.arg_params["data"] = mx.nd.array(input_blob, self.model.ctx)
-      self.model.arg_params["softmax_label"] = mx.nd.empty((1,), self.model.ctx)
-      exe = self.model.sym.bind(self.model.ctx, self.model.arg_params ,args_grad=None, grad_req="null", aux_states=self.model.aux_params)
-      exe.forward(is_train=False)
-      _embedding = exe.outputs[0].asnumpy()
-      #print(_embedding.shape)
-      if embedding is None:
-        embedding = _embedding
-      else:
-        embedding += _embedding
-    if norm:
-      embedding = sklearn.preprocessing.normalize(embedding)
-    return embedding
-
-  def get_feature(self, face_img, norm=True):
-    #aligned_face = self.get_aligned_face(img, force)
-    #if aligned_face is None:
-    #  return None
-    return self.get_feature_impl(face_img, norm)
-
-  def is_same_id(self, source_img, target_img_list):
-    source_face = self.get_aligned_face(source_img, True)
-    print('source face', source_face.shape)
-    target_face_list = []
-    pp = 0
-    for img in target_img_list:
-      target_force = False
-      if pp==len(target_img_list)-1 and len(target_face_list)==0:
-        target_force = True
-      target_face = self.get_aligned_face(img, target_force)
-      if target_face is not None:
-        target_face_list.append(target_face)
-      pp+=1
-    print('target face', len(target_face_list)) 
-    source_feature = self.get_feature(source_face, True)
-    target_feature = None
-    for target_face in target_face_list:
-      _feature = self.get_feature(target_face, False)
-      if target_feature is None:
-        target_feature = _feature
-      else:
-        target_feature += _feature
-    target_feature = sklearn.preprocessing.normalize(target_feature)
-    #sim = np.dot(source_feature, target_feature.T)
-    diff = np.subtract(source_feature, target_feature)
-    dist = np.sum(np.square(diff),1)
-    print('dist', dist)
-    #print(sim, dist)
-    if dist<=self.threshold:
-      return True
-    else:
-      return False
-
-  def sim(self, source_img, target_img_list):
-    print('sim start')
-    source_face = self.get_aligned_face(source_img, True)
-    print('source face', source_face.shape)
-    target_face_list = []
-    pp = 0
-    for img in target_img_list:
-      target_force = False
-      if pp==len(target_img_list)-1 and len(target_face_list)==0:
-        target_force = True
-      target_face = self.get_aligned_face(img, target_force)
-      if target_face is not None:
-        target_face_list.append(target_face)
-      pp+=1
-    print('target face', len(target_face_list)) 
-    source_feature = self.get_feature(source_face, True)
-    target_feature = None
-    sim_list = []
-    for target_face in target_face_list:
-      _feature = self.get_feature(target_face, True)
-      _sim = np.dot(source_feature, _feature.T)
-      sim_list.append(_sim)
-    return np.max(sim_list)
diff --git a/src/common/face_image.py b/src/common/face_image.py
deleted file mode 100644
index a580dc666..000000000
--- a/src/common/face_image.py
+++ /dev/null
@@ -1,271 +0,0 @@
-
-from easydict import EasyDict as edict
-import os
-import json
-import numpy as np
-
-
-def load_property(data_dir):
-  prop = edict()
-  for line in open(os.path.join(data_dir, 'property')):
-    vec = line.strip().split(',')
-    assert len(vec)==3
-    prop.num_classes = int(vec[0])
-    prop.image_size = [int(vec[1]), int(vec[2])]
-  return prop
-
-
-
-def get_dataset_webface(input_dir):
-  clean_list_file = input_dir+"_clean_list.txt"
-  ret = []
-  for line in open(clean_list_file, 'r'):
-    vec = line.strip().split()
-    assert len(vec)==2
-    fimage = edict()
-    fimage.id = vec[0].replace("\\", '/')
-    fimage.classname = vec[1]
-    fimage.image_path = os.path.join(input_dir, fimage.id)
-    ret.append(fimage)
-  return ret
-
-def get_dataset_celeb(input_dir):
-  clean_list_file = input_dir+"_clean_list.txt"
-  ret = []
-  dir2label = {}
-  for line in open(clean_list_file, 'r'):
-    line = line.strip()
-    if not line.startswith('./m.'):
-      continue
-    line = line[2:]
-    vec = line.split('/')
-    assert len(vec)==2
-    if vec[0] in dir2label:
-      label = dir2label[vec[0]]
-    else:
-      label = len(dir2label)
-      dir2label[vec[0]] = label
-
-    fimage = edict()
-    fimage.id = line
-    fimage.classname = str(label)
-    fimage.image_path = os.path.join(input_dir, fimage.id)
-    ret.append(fimage)
-  return ret
-
-def _get_dataset_celeb(input_dir):
-  list_file = input_dir+"_original_list.txt"
-  ret = []
-  for line in open(list_file, 'r'):
-    vec = line.strip().split()
-    assert len(vec)==2
-    fimage = edict()
-    fimage.id = vec[0]
-    fimage.classname = vec[1]
-    fimage.image_path = os.path.join(input_dir, fimage.id)
-    ret.append(fimage)
-  return ret
-
-def get_dataset_facescrub(input_dir):
-  ret = []
-  label = 0
-  person_names = []
-  for person_name in os.listdir(input_dir):
-    person_names.append(person_name)
-  person_names = sorted(person_names)
-  for person_name in person_names:
-    subdir = os.path.join(input_dir, person_name)
-    if not os.path.isdir(subdir):
-      continue
-    for _img in os.listdir(subdir):
-      fimage = edict()
-      fimage.id = os.path.join(person_name, _img)
-      fimage.classname = str(label)
-      fimage.image_path = os.path.join(subdir, _img)
-      fimage.landmark = None
-      fimage.bbox = None
-      ret.append(fimage)
-    label += 1
-  return ret
-
-def get_dataset_megaface(input_dir):
-  ret = []
-  label = 0
-  for prefixdir in os.listdir(input_dir):
-    _prefixdir = os.path.join(input_dir, prefixdir)
-    for subdir in os.listdir(_prefixdir):
-      _subdir = os.path.join(_prefixdir, subdir)
-      if not os.path.isdir(_subdir):
-        continue
-      for img in os.listdir(_subdir):
-        if not img.endswith('.jpg.jpg') and img.endswith('.jpg'):
-          fimage = edict()
-          fimage.id = os.path.join(prefixdir, subdir, img)
-          fimage.classname = str(label)
-          fimage.image_path = os.path.join(_subdir, img)
-          json_file = fimage.image_path+".json"
-          data = None
-          fimage.bbox = None
-          fimage.landmark = None
-          if os.path.exists(json_file):
-            with open(json_file, 'r') as f:
-              data = f.read()
-              data = json.loads(data)
-            assert data is not None
-            if 'bounding_box' in data:
-              fimage.bbox = np.zeros( (4,), dtype=np.float32 )
-              bb = data['bounding_box']
-              fimage.bbox[0] = bb['x']
-              fimage.bbox[1] = bb['y']
-              fimage.bbox[2] = bb['x']+bb['width']
-              fimage.bbox[3] = bb['y']+bb['height']
-              #print('bb')
-            if 'landmarks' in data:
-              landmarks = data['landmarks']
-              if '1' in landmarks and '0' in landmarks and '2' in landmarks:
-                fimage.landmark = np.zeros( (3,2), dtype=np.float32 )
-                fimage.landmark[0][0] = landmarks['1']['x']
-                fimage.landmark[0][1] = landmarks['1']['y']
-                fimage.landmark[1][0] = landmarks['0']['x']
-                fimage.landmark[1][1] = landmarks['0']['y']
-                fimage.landmark[2][0] = landmarks['2']['x']
-                fimage.landmark[2][1] = landmarks['2']['y']
-              #print('lm')
-
-          ret.append(fimage)
-      label+=1
-  return ret
-
-def get_dataset_fgnet(input_dir):
-  ret = []
-  label = 0
-  for subdir in os.listdir(input_dir):
-    _subdir = os.path.join(input_dir, subdir)
-    if not os.path.isdir(_subdir):
-      continue
-    for img in os.listdir(_subdir):
-      if img.endswith('.JPG'):
-        fimage = edict()
-        fimage.id = os.path.join(_subdir, img)
-        fimage.classname = str(label)
-        fimage.image_path = os.path.join(_subdir, img)
-        json_file = fimage.image_path+".json"
-        data = None
-        fimage.bbox = None
-        fimage.landmark = None
-        if os.path.exists(json_file):
-          with open(json_file, 'r') as f:
-            data = f.read()
-            data = json.loads(data)
-          assert data is not None
-          if 'bounding_box' in data:
-            fimage.bbox = np.zeros( (4,), dtype=np.float32 )
-            bb = data['bounding_box']
-            fimage.bbox[0] = bb['x']
-            fimage.bbox[1] = bb['y']
-            fimage.bbox[2] = bb['x']+bb['width']
-            fimage.bbox[3] = bb['y']+bb['height']
-            #print('bb')
-          if 'landmarks' in data:
-            landmarks = data['landmarks']
-            if '1' in landmarks and '0' in landmarks and '2' in landmarks:
-              fimage.landmark = np.zeros( (3,2), dtype=np.float32 )
-              fimage.landmark[0][0] = landmarks['1']['x']
-              fimage.landmark[0][1] = landmarks['1']['y']
-              fimage.landmark[1][0] = landmarks['0']['x']
-              fimage.landmark[1][1] = landmarks['0']['y']
-              fimage.landmark[2][0] = landmarks['2']['x']
-              fimage.landmark[2][1] = landmarks['2']['y']
-            #print('lm')
-
-        #fimage.landmark = None
-        ret.append(fimage)
-    label+=1
-  return ret
-
-def get_dataset_ytf(input_dir):
-  ret = []
-  label = 0
-  person_names = []
-  for person_name in os.listdir(input_dir):
-    person_names.append(person_name)
-  person_names = sorted(person_names)
-  for person_name in person_names:
-    _subdir = os.path.join(input_dir, person_name)
-    if not os.path.isdir(_subdir):
-      continue
-    for _subdir2 in os.listdir(_subdir):
-      _subdir2 = os.path.join(_subdir, _subdir2)
-      if not os.path.isdir(_subdir2):
-        continue
-      _ret = []
-      for img in os.listdir(_subdir2):
-        fimage = edict()
-        fimage.id = os.path.join(_subdir2, img)
-        fimage.classname = str(label)
-        fimage.image_path = os.path.join(_subdir2, img)
-        fimage.bbox = None
-        fimage.landmark = None
-        _ret.append(fimage)
-      ret += _ret
-    label+=1
-  return ret
-
-def get_dataset_clfw(input_dir):
-  ret = []
-  label = 0
-  for img in os.listdir(input_dir):
-    fimage = edict()
-    fimage.id = img
-    fimage.classname = str(0)
-    fimage.image_path = os.path.join(input_dir, img)
-    fimage.bbox = None
-    fimage.landmark = None
-    ret.append(fimage)
-  return ret
-
-def get_dataset_common(input_dir, min_images = 1):
-  ret = []
-  label = 0
-  person_names = []
-  for person_name in os.listdir(input_dir):
-    person_names.append(person_name)
-  person_names = sorted(person_names)
-  for person_name in person_names:
-    _subdir = os.path.join(input_dir, person_name)
-    if not os.path.isdir(_subdir):
-      continue
-    _ret = []
-    for img in os.listdir(_subdir):
-      if not img.endswith('.jpg') and not img.endswith('.png'):
-          continue
-      fimage = edict()
-      fimage.id = os.path.join(person_name, img)
-      fimage.classname = str(label)
-      fimage.image_path = os.path.join(_subdir, img)
-      fimage.bbox = None
-      fimage.landmark = None
-      _ret.append(fimage)
-    if len(_ret)>=min_images:
-      ret += _ret
-      label+=1
-  return ret
-
-def get_dataset(name, input_dir):
-  if name=='webface' or name=='lfw' or name=='vgg':
-    return get_dataset_common(input_dir)
-  if name=='celeb':
-    return get_dataset_celeb(input_dir)
-  if name=='facescrub':
-    return get_dataset_facescrub(input_dir)
-  if name=='megaface':
-    return get_dataset_megaface(input_dir)
-  if name=='fgnet':
-    return get_dataset_fgnet(input_dir)
-  if name=='ytf':
-    return get_dataset_ytf(input_dir)
-  if name=='clfw':
-    return get_dataset_clfw(input_dir)
-  return None
-
-
diff --git a/src/common/face_preprocess.py b/src/common/face_preprocess.py
deleted file mode 100644
index 0b59828a8..000000000
--- a/src/common/face_preprocess.py
+++ /dev/null
@@ -1,113 +0,0 @@
-
-import cv2
-import numpy as np
-from skimage import transform as trans
-
-def parse_lst_line(line):
-  vec = line.strip().split("\t")
-  assert len(vec)>=3
-  aligned = int(vec[0])
-  image_path = vec[1]
-  label = int(vec[2])
-  bbox = None
-  landmark = None
-  #print(vec)
-  if len(vec)>3:
-    bbox = np.zeros( (4,), dtype=np.int32)
-    for i in xrange(3,7):
-      bbox[i-3] = int(vec[i])
-    landmark = None
-    if len(vec)>7:
-      _l = []
-      for i in xrange(7,17):
-        _l.append(float(vec[i]))
-      landmark = np.array(_l).reshape( (2,5) ).T
-  #print(aligned)
-  return image_path, label, bbox, landmark, aligned
-
-
-
-
-def read_image(img_path, **kwargs):
-  mode = kwargs.get('mode', 'rgb')
-  layout = kwargs.get('layout', 'HWC')
-  if mode=='gray':
-    img = cv2.imread(img_path, cv2.CV_LOAD_IMAGE_GRAYSCALE)
-  else:
-    img = cv2.imread(img_path, cv2.CV_LOAD_IMAGE_COLOR)
-    if mode=='rgb':
-      #print('to rgb')
-      img = img[...,::-1]
-    if layout=='CHW':
-      img = np.transpose(img, (2,0,1))
-  return img
-
-
-def preprocess(img, bbox=None, landmark=None, **kwargs):
-  if isinstance(img, str):
-    img = read_image(img, **kwargs)
-  M = None
-  image_size = []
-  str_image_size = kwargs.get('image_size', '')
-  if len(str_image_size)>0:
-    image_size = [int(x) for x in str_image_size.split(',')]
-    if len(image_size)==1:
-      image_size = [image_size[0], image_size[0]]
-    assert len(image_size)==2
-    assert image_size[0]==112
-    assert image_size[0]==112 or image_size[1]==96
-  if landmark is not None:
-    assert len(image_size)==2
-    src = np.array([
-      [30.2946, 51.6963],
-      [65.5318, 51.5014],
-      [48.0252, 71.7366],
-      [33.5493, 92.3655],
-      [62.7299, 92.2041] ], dtype=np.float32 )
-    if image_size[1]==112:
-      src[:,0] += 8.0
-    dst = landmark.astype(np.float32)
-
-    tform = trans.SimilarityTransform()
-    tform.estimate(dst, src)
-    M = tform.params[0:2,:]
-    #M = cv2.estimateRigidTransform( dst.reshape(1,5,2), src.reshape(1,5,2), False)
-
-  if M is None:
-    if bbox is None: #use center crop
-      det = np.zeros(4, dtype=np.int32)
-      det[0] = int(img.shape[1]*0.0625)
-      det[1] = int(img.shape[0]*0.0625)
-      det[2] = img.shape[1] - det[0]
-      det[3] = img.shape[0] - det[1]
-    else:
-      det = bbox
-    margin = kwargs.get('margin', 44)
-    bb = np.zeros(4, dtype=np.int32)
-    bb[0] = np.maximum(det[0]-margin/2, 0)
-    bb[1] = np.maximum(det[1]-margin/2, 0)
-    bb[2] = np.minimum(det[2]+margin/2, img.shape[1])
-    bb[3] = np.minimum(det[3]+margin/2, img.shape[0])
-    ret = img[bb[1]:bb[3],bb[0]:bb[2],:]
-    if len(image_size)>0:
-      ret = cv2.resize(ret, (image_size[1], image_size[0]))
-    return ret 
-  else: #do align using landmark
-    assert len(image_size)==2
-
-    #src = src[0:3,:]
-    #dst = dst[0:3,:]
-
-
-    #print(src.shape, dst.shape)
-    #print(src)
-    #print(dst)
-    #print(M)
-    warped = cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue = 0.0)
-
-    #tform3 = trans.ProjectiveTransform()
-    #tform3.estimate(src, dst)
-    #warped = trans.warp(img, tform3, output_shape=_shape)
-    return warped
-
-
diff --git a/src/common/noise_sgd.py b/src/common/noise_sgd.py
deleted file mode 100644
index 70024ec7d..000000000
--- a/src/common/noise_sgd.py
+++ /dev/null
@@ -1,38 +0,0 @@
-import mxnet.optimizer as optimizer
-from mxnet import ndarray as nd
-
-class NoiseSGD(optimizer.SGD):
-    """Noise SGD.
-
-
-    This optimizer accepts the same arguments as :class:`.SGD`.
-    """
-    def __init__(self, scale, **kwargs):
-        super(NoiseSGD, self).__init__(**kwargs)
-        print('init noise sgd with', scale)
-        self.scale = scale
-
-    def update(self, index, weight, grad, state):
-        assert(isinstance(weight, NDArray))
-        assert(isinstance(grad, NDArray))
-        self._update_count(index)
-        lr = self._get_lr(index)
-        wd = self._get_wd(index)
-
-        grad = grad * self.rescale_grad
-        if self.clip_gradient is not None:
-            grad = clip(grad, -self.clip_gradient, self.clip_gradient)
-        noise = nd.random.normal(scale = self.scale, shape = grad.shape, dtype=grad.dtype, ctx = grad.context)
-        grad += noise
-
-        if state is not None:
-            mom = state
-            mom[:] *= self.momentum
-            grad += wd * weight
-            mom[:] += grad
-            grad[:] += self.momentum * mom
-            weight[:] += -lr * grad
-        else:
-            assert self.momentum == 0.0
-            weight[:] += -lr * (grad + wd * weight)
-
diff --git a/src/data.py b/src/data.py
deleted file mode 100644
index 1db9243dd..000000000
--- a/src/data.py
+++ /dev/null
@@ -1,1016 +0,0 @@
-# THIS FILE IS FOR EXPERIMENTS, USE image_iter.py FOR NORMAL IMAGE LOADING.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-
-import mxnet as mx
-from mxnet import ndarray as nd
-#from . import _ndarray_internal as _internal
-#from mxnet._ndarray_internal import _cvimresize as imresize
-#from ._ndarray_internal import _cvcopyMakeBorder as copyMakeBorder
-from mxnet import io
-from mxnet import recordio
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_preprocess
-import multiprocessing
-
-logger = logging.getLogger()
-
-def pick_triplets_impl(q_in, q_out):
-  more = True
-  while more:
-      deq = q_in.get()
-      if deq is None:
-        more = False
-      else:
-        embeddings, emb_start_idx, nrof_images, alpha = deq
-        print('running', emb_start_idx, nrof_images, os.getpid())
-        for j in xrange(1,nrof_images):
-            a_idx = emb_start_idx + j - 1
-            neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
-            for pair in xrange(j, nrof_images): # For every possible positive pair.
-                p_idx = emb_start_idx + pair
-                pos_dist_sqr = np.sum(np.square(embeddings[a_idx]-embeddings[p_idx]))
-                neg_dists_sqr[emb_start_idx:emb_start_idx+nrof_images] = np.NaN
-                all_neg = np.where(np.logical_and(neg_dists_sqr-pos_dist_sqr<alpha, pos_dist_sqr<neg_dists_sqr))[0]  # FaceNet selection
-                #all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
-                nrof_random_negs = all_neg.shape[0]
-                if nrof_random_negs>0:
-                    rnd_idx = np.random.randint(nrof_random_negs)
-                    n_idx = all_neg[rnd_idx]
-                    #triplets.append( (a_idx, p_idx, n_idx) )
-                    q_out.put( (a_idx, p_idx, n_idx) )
-        #emb_start_idx += nrof_images
-  print('exit',os.getpid())
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0,
-                 c2c_threshold = 0.0, output_c2c = 0, c2c_mode = -10, limit = 0,
-                 ctx_num = 0, images_per_identity = 0, data_extra = None, hard_mining = False, 
-                 triplet_params = None, coco_mode = False,
-                 mx_model = None,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        if path_imgrec:
-            logging.info('loading recordio %s...',
-                         path_imgrec)
-            path_imgidx = path_imgrec[0:-4]+".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-            s = self.imgrec.read_idx(0)
-            header, _ = recordio.unpack(s)
-            self.idx2cos = {}
-            self.idx2flag = {}
-            self.idx2meancos = {}
-            self.c2c_auto = False
-            #if output_c2c or c2c_threshold>0.0 or c2c_mode>=-5:
-            #  path_c2c = os.path.join(os.path.dirname(path_imgrec), 'c2c')
-            #  print(path_c2c)
-            #  if os.path.exists(path_c2c):
-            #    for line in open(path_c2c, 'r'):
-            #      vec = line.strip().split(',')
-            #      idx = int(vec[0])
-            #      self.idx2cos[idx] = float(vec[1])
-            #      self.idx2flag[idx] = 1
-            #      if len(vec)>2:
-            #        self.idx2flag[idx] = int(vec[2])
-            #  else:
-            #    self.c2c_auto = True
-            #    self.c2c_step = 10000
-            if header.flag>0:
-              print('header0 label', header.label)
-              self.header0 = (int(header.label[0]), int(header.label[1]))
-              #assert(header.flag==1)
-              self.imgidx = range(1, int(header.label[0]))
-              if c2c_mode==0:
-                imgidx2 = []
-                for idx in self.imgidx:
-                  c = self.idx2cos[idx]
-                  f = self.idx2flag[idx]
-                  if f!=1:
-                    continue
-                  imgidx2.append(idx)
-                print('idx count', len(self.imgidx), len(imgidx2))
-                self.imgidx = imgidx2
-              elif c2c_mode==1:
-                imgidx2 = []
-                tmp = []
-                for idx in self.imgidx:
-                  c = self.idx2cos[idx]
-                  f = self.idx2flag[idx]
-                  if f==1:
-                    imgidx2.append(idx)
-                  else:
-                    tmp.append( (idx, c) )
-                tmp = sorted(tmp, key = lambda x:x[1])
-                tmp = tmp[250000:300000]
-                for _t in tmp:
-                  imgidx2.append(_t[0])
-                print('idx count', len(self.imgidx), len(imgidx2))
-                self.imgidx = imgidx2
-              elif c2c_mode==2:
-                imgidx2 = []
-                tmp = []
-                for idx in self.imgidx:
-                  c = self.idx2cos[idx]
-                  f = self.idx2flag[idx]
-                  if f==1:
-                    imgidx2.append(idx)
-                  else:
-                    tmp.append( (idx, c) )
-                tmp = sorted(tmp, key = lambda x:x[1])
-                tmp = tmp[200000:300000]
-                for _t in tmp:
-                  imgidx2.append(_t[0])
-                print('idx count', len(self.imgidx), len(imgidx2))
-                self.imgidx = imgidx2
-              elif c2c_mode==-2:
-                imgidx2 = []
-                for idx in self.imgidx:
-                  c = self.idx2cos[idx]
-                  f = self.idx2flag[idx]
-                  if f==2:
-                    continue
-                  if c<0.73:
-                    continue
-                  imgidx2.append(idx)
-                print('idx count', len(self.imgidx), len(imgidx2))
-                self.imgidx = imgidx2
-              elif c2c_threshold>0.0:
-                imgidx2 = []
-                for idx in self.imgidx:
-                  c = self.idx2cos[idx]
-                  f = self.idx2flag[idx]
-                  if c<c2c_threshold:
-                    continue
-                  imgidx2.append(idx)
-                print(len(self.imgidx), len(imgidx2))
-                self.imgidx = imgidx2
-              self.id2range = {}
-              self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-              c2c_stat = [0,0]
-              for identity in self.seq_identity:
-                s = self.imgrec.read_idx(identity)
-                header, _ = recordio.unpack(s)
-                a,b = int(header.label[0]), int(header.label[1])
-                self.id2range[identity] = (a,b)
-                count = b-a
-                if count>=output_c2c:
-                  c2c_stat[1]+=1
-                else:
-                  c2c_stat[0]+=1
-                for ii in xrange(a,b):
-                  self.idx2flag[ii] = count
-                if len(self.idx2cos)>0:
-                  m = 0.0
-                  for ii in xrange(a,b):
-                    m+=self.idx2cos[ii]
-                  m/=(b-a)
-                  for ii in xrange(a,b):
-                    self.idx2meancos[ii] = m
-                  #self.idx2meancos[identity] = m
-
-              print('id2range', len(self.id2range))
-              print(len(self.idx2cos), len(self.idx2meancos), len(self.idx2flag))
-              print('c2c_stat', c2c_stat)
-              if limit>0 and limit<len(self.imgidx):
-                random.seed(727)
-                prob = float(limit)/len(self.imgidx)
-                new_imgidx = []
-                new_ids = 0
-                for identity in self.seq_identity:
-                  s = self.imgrec.read_idx(identity)
-                  header, _ = recordio.unpack(s)
-                  a,b = int(header.label[0]), int(header.label[1])
-                  found = False
-                  for _idx in xrange(a,b):
-                    if random.random()<prob:
-                      found = True
-                      new_imgidx.append(_idx)
-                  if found:
-                    new_ids+=1
-                self.imgidx = new_imgidx
-                print('new ids', new_ids)
-                random.seed(None)
-                #random.Random(727).shuffle(self.imgidx)
-                #self.imgidx = self.imgidx[0:limit]
-            else:
-              self.imgidx = list(self.imgrec.keys)
-            if shuffle:
-              self.seq = self.imgidx
-              self.oseq = self.imgidx
-              print(len(self.seq))
-            else:
-              self.seq = None
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        #self.cast_aug = mx.image.CastAug()
-        #self.color_aug = mx.image.ColorJitterAug(0.4, 0.4, 0.4)
-        self.ctx_num = ctx_num 
-        self.c2c_threshold = c2c_threshold
-        self.output_c2c = output_c2c
-        self.per_batch_size = int(self.batch_size/self.ctx_num)
-        self.images_per_identity = images_per_identity
-        if self.images_per_identity>0:
-          self.identities = int(self.per_batch_size/self.images_per_identity)
-          self.per_identities = self.identities
-          self.repeat = 3000000.0/(self.images_per_identity*len(self.id2range))
-          self.repeat = int(self.repeat)
-          print(self.images_per_identity, self.identities, self.repeat)
-        self.data_extra = None
-        if data_extra is not None:
-          self.data_extra = nd.array(data_extra)
-          self.provide_data = [(data_name, (batch_size,) + data_shape), ('extra', data_extra.shape)]
-        self.hard_mining = hard_mining
-        self.mx_model = mx_model
-        if self.hard_mining:
-          assert self.images_per_identity>0
-          assert self.mx_model is not None
-        self.triplet_params = triplet_params
-        self.triplet_mode = False
-        self.coco_mode = coco_mode
-        if len(label_name)>0:
-          if output_c2c:
-            self.provide_label = [(label_name, (batch_size,2))]
-          else:
-            self.provide_label = [(label_name, (batch_size,))]
-        else:
-          self.provide_label = []
-        print(self.provide_label[0][1])
-        if self.coco_mode:
-          assert self.triplet_params is None
-          assert self.images_per_identity>0
-        if self.triplet_params is not None:
-          assert self.images_per_identity>0
-          assert self.mx_model is not None
-          self.triplet_bag_size = self.triplet_params[0]
-          self.triplet_alpha = self.triplet_params[1]
-          self.triplet_max_ap = self.triplet_params[2]
-          assert self.triplet_bag_size>0
-          assert self.triplet_alpha>=0.0
-          assert self.triplet_alpha<=1.0
-          self.triplet_mode = True
-          self.triplet_oseq_cur = 0
-          self.triplet_oseq_reset()
-          self.seq_min_size = self.batch_size*2
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-        self.times = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
-        #self.reset()
-
-
-    def ____pick_triplets(self, embeddings, nrof_images_per_class):
-      emb_start_idx = 0
-      people_per_batch = len(nrof_images_per_class)
-      nrof_threads = 8
-      q_in = multiprocessing.Queue()
-      q_out = multiprocessing.Queue()
-      processes = [multiprocessing.Process(target=pick_triplets_impl, args=(q_in, q_out)) \
-                      for i in range(nrof_threads)]
-      for p in processes:
-          p.start()
-      
-      # VGG Face: Choosing good triplets is crucial and should strike a balance between
-      #  selecting informative (i.e. challenging) examples and swamping training with examples that
-      #  are too hard. This is achieve by extending each pair (a, p) to a triplet (a, p, n) by sampling
-      #  the image n at random, but only between the ones that violate the triplet loss margin. The
-      #  latter is a form of hard-negative mining, but it is not as aggressive (and much cheaper) than
-      #  choosing the maximally violating example, as often done in structured output learning.
-
-      for i in xrange(people_per_batch):
-          nrof_images = int(nrof_images_per_class[i])
-          job = (embeddings, emb_start_idx, nrof_images, self.triplet_alpha)
-          emb_start_idx+=nrof_images
-          q_in.put(job)
-      for i in xrange(nrof_threads):
-        q_in.put(None)
-      print('joining')
-      for p in processes:
-          p.join()
-      print('joined')
-      q_out.put(None)
-
-      triplets = []
-      more = True
-      while more:
-        triplet = q_out.get()
-        if triplet is None:
-          more = False
-        else:
-          triplets.append(triplets)
-      np.random.shuffle(triplets)
-      return triplets
-
-    #cal pairwise dists on single gpu
-    def _pairwise_dists(self, embeddings):
-      nd_embedding = mx.nd.array(embeddings, mx.gpu(0))
-      pdists = []
-      for idx in xrange(embeddings.shape[0]):
-        a_embedding = nd_embedding[idx]
-        body = mx.nd.broadcast_sub(a_embedding, nd_embedding)
-        body = body*body
-        body = mx.nd.sum_axis(body, axis=1)
-        ret = body.asnumpy()
-        #print(ret.shape)
-        pdists.append(ret)
-      return pdists
-
-    def pairwise_dists(self, embeddings):
-      nd_embedding_list = []
-      for i in xrange(self.ctx_num):
-        nd_embedding = mx.nd.array(embeddings, mx.gpu(i))
-        nd_embedding_list.append(nd_embedding)
-      nd_pdists = []
-      pdists = []
-      for idx in xrange(embeddings.shape[0]):
-        emb_idx = idx%self.ctx_num
-        nd_embedding = nd_embedding_list[emb_idx]
-        a_embedding = nd_embedding[idx]
-        body = mx.nd.broadcast_sub(a_embedding, nd_embedding)
-        body = body*body
-        body = mx.nd.sum_axis(body, axis=1)
-        nd_pdists.append(body)
-        if len(nd_pdists)==self.ctx_num or idx==embeddings.shape[0]-1:
-          for x in nd_pdists:
-            pdists.append(x.asnumpy())
-          nd_pdists = []
-      return pdists
-
-    def pick_triplets(self, embeddings, nrof_images_per_class):
-      emb_start_idx = 0
-      triplets = []
-      people_per_batch = len(nrof_images_per_class)
-      #self.time_reset()
-      pdists = self.pairwise_dists(embeddings)
-      #self.times[3] += self.time_elapsed()
-
-      for i in xrange(people_per_batch):
-          nrof_images = int(nrof_images_per_class[i])
-          for j in xrange(1,nrof_images):
-              #self.time_reset()
-              a_idx = emb_start_idx + j - 1
-              #neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
-              neg_dists_sqr = pdists[a_idx]
-              #self.times[3] += self.time_elapsed()
-
-              for pair in xrange(j, nrof_images): # For every possible positive pair.
-                  p_idx = emb_start_idx + pair
-                  #self.time_reset()
-                  pos_dist_sqr = np.sum(np.square(embeddings[a_idx]-embeddings[p_idx]))
-                  #self.times[4] += self.time_elapsed()
-                  #self.time_reset()
-                  neg_dists_sqr[emb_start_idx:emb_start_idx+nrof_images] = np.NaN
-                  if self.triplet_max_ap>0.0:
-                    if pos_dist_sqr>self.triplet_max_ap:
-                      continue
-                  all_neg = np.where(np.logical_and(neg_dists_sqr-pos_dist_sqr<self.triplet_alpha, pos_dist_sqr<neg_dists_sqr))[0]  # FaceNet selection
-                  #self.times[5] += self.time_elapsed()
-                  #self.time_reset()
-                  #all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
-                  nrof_random_negs = all_neg.shape[0]
-                  if nrof_random_negs>0:
-                      rnd_idx = np.random.randint(nrof_random_negs)
-                      n_idx = all_neg[rnd_idx]
-                      triplets.append( (a_idx, p_idx, n_idx) )
-          emb_start_idx += nrof_images
-      np.random.shuffle(triplets)
-      return triplets
-
-    def __pick_triplets(self, embeddings, nrof_images_per_class):
-      emb_start_idx = 0
-      triplets = []
-      people_per_batch = len(nrof_images_per_class)
-      
-      for i in xrange(people_per_batch):
-          nrof_images = int(nrof_images_per_class[i])
-          if nrof_images<2:
-            continue
-          for j in xrange(1,nrof_images):
-              a_idx = emb_start_idx + j - 1
-              pcount = nrof_images-1
-              dists_a2all = np.sum(np.square(embeddings[a_idx] - embeddings), 1) #(N,)
-              #print(a_idx, dists_a2all.shape)
-              ba = emb_start_idx
-              bb = emb_start_idx+nrof_images
-              sorted_idx = np.argsort(dists_a2all)
-              #print('assert', sorted_idx[0], a_idx)
-              #assert sorted_idx[0]==a_idx
-              #for idx in sorted_idx:
-              #  print(idx, dists_a2all[idx])
-              p2n_map = {}
-              pfound = 0
-              for idx in sorted_idx:
-                if idx==a_idx: #is anchor
-                  continue
-                if idx<bb and idx>=ba: #is pos
-                  p2n_map[idx] = [dists_a2all[idx], []] #ap, [neg_list]
-                  pfound+=1
-                else: # is neg
-                  an = dists_a2all[idx]
-                  if pfound==pcount and len(p2n_map)==0:
-                    break
-                  to_del = []
-                  for p_idx in p2n_map:
-                    v = p2n_map[p_idx]
-                    an_ap = an - v[0]
-                    if an_ap<self.triplet_alpha:
-                      v[1].append(idx)
-                    else:
-                      #output
-                      if len(v[1])>0:
-                        n_idx = random.choice(v[1])
-                        triplets.append( (a_idx, p_idx, n_idx) )
-                      to_del.append(p_idx)
-                  for _del in to_del:
-                    del p2n_map[_del]
-              for p_idx,v in p2n_map.iteritems():
-                if len(v[1])>0:
-                  n_idx = random.choice(v[1])
-                  triplets.append( (a_idx, p_idx, n_idx) )
-          emb_start_idx += nrof_images
-      np.random.shuffle(triplets)
-      return triplets
-
-    def triplet_oseq_reset(self):
-      #reset self.oseq by identities seq
-      self.triplet_oseq_cur = 0
-      ids = []
-      for k in self.id2range:
-        ids.append(k)
-      random.shuffle(ids)
-      self.oseq = []
-      for _id in ids:
-        v = self.id2range[_id]
-        _list = range(*v)
-        random.shuffle(_list)
-        if len(_list)>self.images_per_identity:
-          _list = _list[0:self.images_per_identity]
-        self.oseq += _list
-      print('oseq', len(self.oseq))
-
-    def time_reset(self):
-      self.time_now = datetime.datetime.now()
-
-    def time_elapsed(self):
-      time_now = datetime.datetime.now()
-      diff = time_now - self.time_now
-      return diff.total_seconds()
-
-
-    def select_triplets(self):
-      self.seq = []
-      while len(self.seq)<self.seq_min_size:
-        self.time_reset()
-        embeddings = None
-        bag_size = self.triplet_bag_size
-        batch_size = self.batch_size
-        #data = np.zeros( (bag_size,)+self.data_shape )
-        #label = np.zeros( (bag_size,) )
-        tag = []
-        #idx = np.zeros( (bag_size,) )
-        print('eval %d images..'%bag_size, self.triplet_oseq_cur)
-        print('triplet time stat', self.times)
-        if self.triplet_oseq_cur+bag_size>len(self.oseq):
-          self.triplet_oseq_reset()
-          print('eval %d images..'%bag_size, self.triplet_oseq_cur)
-        self.times[0] += self.time_elapsed()
-        self.time_reset()
-        #print(data.shape)
-        data = nd.zeros( self.provide_data[0][1] )
-        label = nd.zeros( self.provide_label[0][1] )
-        ba = 0
-        while True:
-          bb = min(ba+batch_size, bag_size)
-          if ba>=bb:
-            break
-          #_batch = self.data_iter.next()
-          #_data = _batch.data[0].asnumpy()
-          #print(_data.shape)
-          #_label = _batch.label[0].asnumpy()
-          #data[ba:bb,:,:,:] = _data
-          #label[ba:bb] = _label
-          for i in xrange(ba, bb):
-            _idx = self.oseq[i+self.triplet_oseq_cur]
-            s = self.imgrec.read_idx(_idx)
-            header, img = recordio.unpack(s)
-            img = self.imdecode(img)
-            data[i-ba][:] = self.postprocess_data(img)
-            label[i-ba][:] = header.label
-            tag.append( ( int(header.label), _idx) )
-            #idx[i] = _idx
-
-          db = mx.io.DataBatch(data=(data,), label=(label,))
-          self.mx_model.forward(db, is_train=False)
-          net_out = self.mx_model.get_outputs()
-          #print('eval for selecting triplets',ba,bb)
-          #print(net_out)
-          #print(len(net_out))
-          #print(net_out[0].asnumpy())
-          net_out = net_out[0].asnumpy()
-          #print(net_out)
-          #print('net_out', net_out.shape)
-          if embeddings is None:
-            embeddings = np.zeros( (bag_size, net_out.shape[1]))
-          embeddings[ba:bb,:] = net_out
-          ba = bb
-        assert len(tag)==bag_size
-        self.triplet_oseq_cur+=bag_size
-        embeddings = sklearn.preprocessing.normalize(embeddings)
-        self.times[1] += self.time_elapsed()
-        self.time_reset()
-        nrof_images_per_class = [1]
-        for i in xrange(1, bag_size):
-          if tag[i][0]==tag[i-1][0]:
-            nrof_images_per_class[-1]+=1
-          else:
-            nrof_images_per_class.append(1)
-          
-        triplets = self.pick_triplets(embeddings, nrof_images_per_class) # shape=(T,3)
-        print('found triplets', len(triplets))
-        ba = 0
-        while True:
-          bb = ba+self.per_batch_size//3
-          if bb>len(triplets):
-            break
-          _triplets = triplets[ba:bb]
-          for i in xrange(3):
-            for triplet in _triplets:
-              _pos = triplet[i]
-              _idx = tag[_pos][1]
-              self.seq.append(_idx)
-          ba = bb
-        self.times[2] += self.time_elapsed()
-
-    def triplet_reset(self):
-      self.select_triplets()
-
-    def hard_mining_reset(self):
-      #import faiss
-      from annoy import AnnoyIndex
-      data = nd.zeros( self.provide_data[0][1] )
-      label = nd.zeros( self.provide_label[0][1] )
-      #label = np.zeros( self.provide_label[0][1] )
-      X = None
-      ba = 0
-      batch_num = 0
-      while ba<len(self.oseq):
-        batch_num+=1
-        if batch_num%10==0:
-          print('loading batch',batch_num, ba)
-        bb = min(ba+self.batch_size, len(self.oseq))
-        _count = bb-ba
-        for i in xrange(_count):
-          idx = self.oseq[i+ba]
-          s = self.imgrec.read_idx(idx)
-          header, img = recordio.unpack(s)
-          img = self.imdecode(img)
-          data[i][:] = self.postprocess_data(img)
-          label[i][:] = header.label
-        db = mx.io.DataBatch(data=(data,self.data_extra), label=(label,))
-        self.mx_model.forward(db, is_train=False)
-        net_out = self.mx_model.get_outputs()
-        embedding = net_out[0].asnumpy()
-        nembedding = sklearn.preprocessing.normalize(embedding)
-        if _count<self.batch_size:
-          nembedding = nembedding[0:_count,:]
-        if X is None:
-          X = np.zeros( (len(self.id2range), nembedding.shape[1]), dtype=np.float32 )
-        nplabel = label.asnumpy()
-        for i in xrange(_count):
-          ilabel = int(nplabel[i])
-          #print(ilabel, ilabel.__class__)
-          X[ilabel] += nembedding[i]
-        ba = bb
-      X = sklearn.preprocessing.normalize(X)
-      d = X.shape[1]
-      t = AnnoyIndex(d, metric='euclidean')
-      for i in xrange(X.shape[0]):
-        t.add_item(i, X[i])
-      print('start to build index')
-      t.build(20)
-      print(X.shape)
-      k = self.per_identities
-      self.seq = []
-      for i in xrange(X.shape[0]):
-        nnlist = t.get_nns_by_item(i, k)
-        assert nnlist[0]==i
-        for _label in nnlist:
-          assert _label<len(self.id2range)
-          _id = self.header0[0]+_label
-          v = self.id2range[_id]
-          _list = range(*v)
-          if len(_list)<self.images_per_identity:
-            random.shuffle(_list)
-          else:
-            _list = np.random.choice(_list, self.images_per_identity, replace=False)
-          for i in xrange(self.images_per_identity):
-            _idx = _list[i%len(_list)]
-            self.seq.append(_idx)
-      #faiss_params = [20,5]
-      #quantizer = faiss.IndexFlatL2(d)  # the other index
-      #index = faiss.IndexIVFFlat(quantizer, d, faiss_params[0], faiss.METRIC_L2)
-      #assert not index.is_trained
-      #index.train(X)
-      #index.add(X)
-      #assert index.is_trained
-      #print('trained')
-      #index.nprobe = faiss_params[1]
-      #D, I = index.search(X, k)     # actual search
-      #print(I.shape)
-      #self.seq = []
-      #for i in xrange(I.shape[0]):
-      #  #assert I[i][0]==i
-      #  for j in xrange(k):
-      #    _label = I[i][j]
-      #    assert _label<len(self.id2range)
-      #    _id = self.header0[0]+_label
-      #    v = self.id2range[_id]
-      #    _list = range(*v)
-      #    if len(_list)<self.images_per_identity:
-      #      random.shuffle(_list)
-      #    else:
-      #      _list = np.random.choice(_list, self.images_per_identity, replace=False)
-      #    for i in xrange(self.images_per_identity):
-      #      _idx = _list[i%len(_list)]
-      #      self.seq.append(_idx)
-    def reset_c2c(self):
-      self.select_triplets()
-      for identity,v in self.id2range.iteritems():
-        _list = range(*v)
-      
-        for idx in _list:
-          s = imgrec.read_idx(idx)
-          ocontents.append(s)
-        embeddings = None
-        #print(len(ocontents))
-        ba = 0
-        while True:
-          bb = min(ba+args.batch_size, len(ocontents))
-          if ba>=bb:
-            break
-          _batch_size = bb-ba
-          _batch_size2 = max(_batch_size, args.ctx_num)
-          data = nd.zeros( (_batch_size2,3, image_size[0], image_size[1]) )
-          label = nd.zeros( (_batch_size2,) )
-          count = bb-ba
-          ii=0
-          for i in xrange(ba, bb):
-            header, img = mx.recordio.unpack(ocontents[i])
-            img = mx.image.imdecode(img)
-            img = nd.transpose(img, axes=(2, 0, 1))
-            data[ii][:] = img
-            label[ii][:] = header.label
-            ii+=1
-          while ii<_batch_size2:
-            data[ii][:] = data[0][:]
-            label[ii][:] = label[0][:]
-            ii+=1
-          db = mx.io.DataBatch(data=(data,), label=(label,))
-          self.mx_model.forward(db, is_train=False)
-          net_out = self.mx_model.get_outputs()
-          net_out = net_out[0].asnumpy()
-          model.forward(db, is_train=False)
-          net_out = model.get_outputs()
-          net_out = net_out[0].asnumpy()
-          if embeddings is None:
-            embeddings = np.zeros( (len(ocontents), net_out.shape[1]))
-          embeddings[ba:bb,:] = net_out[0:_batch_size,:]
-          ba = bb
-        embeddings = sklearn.preprocessing.normalize(embeddings)
-        embedding = np.mean(embeddings, axis=0, keepdims=True)
-        embedding = sklearn.preprocessing.normalize(embedding)
-        sims = np.dot(embeddings, embedding).flatten()
-        assert len(sims)==len(_list)
-        for i in xrange(len(_list)):
-          _idx = _list[i]
-          self.idx2cos[_idx] = sims[i]
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        if self.c2c_auto:
-          self.reset_c2c()
-        self.cur = 0
-        if self.images_per_identity>0:
-          if self.triplet_mode:
-            self.triplet_reset()
-          elif not self.hard_mining:
-            self.seq = []
-            idlist = []
-            for _id,v in self.id2range.iteritems():
-              idlist.append((_id,range(*v)))
-            for r in xrange(self.repeat):
-              if r%10==0:
-                print('repeat', r)
-              if self.shuffle:
-                random.shuffle(idlist)
-              for item in idlist:
-                _id = item[0]
-                _list = item[1]
-                #random.shuffle(_list)
-                if len(_list)<self.images_per_identity:
-                  random.shuffle(_list)
-                else:
-                  _list = np.random.choice(_list, self.images_per_identity, replace=False)
-                for i in xrange(self.images_per_identity):
-                  _idx = _list[i%len(_list)]
-                  self.seq.append(_idx)
-          else:
-            self.hard_mining_reset()
-          print('seq len', len(self.seq))
-        else:
-          if self.shuffle:
-              random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        """Helper function for reading in next sample."""
-        #set total batch size, for example, 1800, and maximum size for each people, for example 45
-        if self.seq is not None:
-          while True:
-            if self.cur >= len(self.seq):
-                raise StopIteration
-            idx = self.seq[self.cur]
-            self.cur += 1
-            if self.imgrec is not None:
-              s = self.imgrec.read_idx(idx)
-              header, img = recordio.unpack(s)
-              label = header.label
-              if self.output_c2c:
-                count = self.idx2flag[idx]
-                if self.output_c2c==1:
-                  v = np.random.uniform(0.4, 0.5)
-                elif self.output_c2c==2:
-                  v = np.random.uniform(0.4, 0.5)
-                  if count>=self.output_c2c:
-                    v = np.random.uniform(0.3, 0.4)
-                elif self.output_c2c==3:
-                  v = (9.5 - math.log(2.0+count))/10.0
-                  v = min(max(v, 0.3), 0.5)
-                elif self.output_c2c==4:
-                  mu = 0.0
-                  sigma = 0.1
-                  mrange = [0.4,0.5]
-                  v = numpy.random.normal(mu, sigma)
-                  v = math.abs(v)*-1.0+mrange[1]
-                  v = max(v, mrange[0])
-                elif self.output_c2c==5:
-                  v = np.random.uniform(0.41, 0.51)
-                  if count>=175:
-                    v = np.random.uniform(0.37, 0.47)
-                elif self.output_c2c==6:
-                  v = np.random.uniform(0.41, 0.51)
-                  if count>=175:
-                    v = np.random.uniform(0.38, 0.48)
-                else:
-                  assert False
-
-                label = [label, v]
-              else:
-                if not isinstance(label, numbers.Number):
-                  label = label[0]
-              return label, img, None, None
-            else:
-              label, fname, bbox, landmark = self.imglist[idx]
-              return label, self.read_image(fname), bbox, landmark
-        else:
-            s = self.imgrec.read()
-            if s is None:
-                raise StopIteration
-            header, img = recordio.unpack(s)
-            return header.label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = np.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      random.shuffle(augs)
-      for aug in augs:
-        #print(img.shape)
-        img = aug(img, x)
-        #print(img.shape)
-      return img
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in xrange(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                _data = self.imdecode(s)
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.nd_mean is not None:
-                    _data = _data.astype('float32')
-                    _data -= self.nd_mean
-                    _data *= 0.0078125
-                if self.cutoff>0:
-                  centerh = random.randint(0, _data.shape[0]-1)
-                  centerw = random.randint(0, _data.shape[1]-1)
-                  half = self.cutoff//2
-                  starth = max(0, centerh-half)
-                  endh = min(_data.shape[0], centerh+half)
-                  startw = max(0, centerw-half)
-                  endw = min(_data.shape[1], centerw+half)
-                  _data = _data.astype('float32')
-                  #print(starth, endh, startw, endw, _data.shape)
-                  _data[starth:endh, startw:endw, :] = 127.5
-                #_npdata = _data.asnumpy()
-                #if landmark is not None:
-                #  _npdata = face_preprocess.preprocess(_npdata, bbox = bbox, landmark=landmark, image_size=self.image_size)
-                #if self.rand_mirror:
-                #  _npdata = self.mirror_aug(_npdata)
-                #if self.mean is not None:
-                #  _npdata = _npdata.astype(np.float32)
-                #  _npdata -= self.mean
-                #  _npdata *= 0.0078125
-                #nimg = np.zeros(_npdata.shape, dtype=np.float32)
-                #nimg[self.patch[1]:self.patch[3],self.patch[0]:self.patch[2],:] = _npdata[self.patch[1]:self.patch[3], self.patch[0]:self.patch[2], :]
-                #_data = mx.nd.array(nimg)
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    if self.provide_label is not None:
-                      if not self.coco_mode:
-                        if len(batch_label.shape)==1:
-                          batch_label[i][:] = label
-                        else:
-                          for ll in xrange(batch_label.shape[1]):
-                            v = label[ll]
-                            if ll>0:
-                              #c2c = v
-                              #_param = [0.5, 0.4, 0.85, 0.75]
-                              #_a = (_param[1]-_param[0])/(_param[3]-_param[2])
-                              #m = _param[1]+_a*(c2c-_param[3])
-                              #m = min(_param[0], max(_param[1],m))
-                              #v = math.cos(m)
-                              #v = v*v
-                              m = v
-                              v = math.cos(m)
-                              v = v*v
-                              #print('m', i,m,v)
-
-                            batch_label[i][ll] = v
-                      else:
-                        batch_label[i][:] = (i%self.per_batch_size)//self.images_per_identity
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        #print('next end', batch_size, i)
-        _label = None
-        if self.provide_label is not None:
-          _label = [batch_label]
-        if self.data_extra is not None:
-          return io.DataBatch([batch_data, self.data_extra], _label, batch_size - i)
-        else:
-          return io.DataBatch([batch_data], _label, batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/src/data/age_merge.py b/src/data/age_merge.py
deleted file mode 100644
index 27a84970a..000000000
--- a/src/data/age_merge.py
+++ /dev/null
@@ -1,125 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-import numpy as np
-
-
-
-def main(args):
-    if not os.path.exists(args.output):
-        os.makedirs(args.output)
-    train_writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'train.idx'), os.path.join(args.output, 'train.rec'), 'w')
-    val_writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'val.idx'), os.path.join(args.output, 'val.rec'), 'w')
-    train_widx = [0]
-    val_widx = [0]
-    stat = [0,0]
-    #for ds in ['ms1m', 'megaage', 'imdb']:
-    for ds in ['megaage', 'imdb']:
-        for n in ['train', 'val']:
-            #if ds=='ms1m' or ds=='imdb':
-            #  continue
-            repeat = 1
-            if args.mode=='age':
-              if args.lite:
-                if ds!='megaage':
-                  continue
-              if n=='val' and ds!='megaage':
-                continue
-              if n=='train' and ds=='megaage':
-                if args.lite==0:
-                  repeat = 10
-              if n=='train' and ds=='imdb':
-                repeat = 1
-            elif args.mode=='gender':
-              if ds!='imdb':
-                continue
-            else:
-              if n=='train' and ds=='megaage':
-                repeat = 10
-            writer = train_writer
-            widx = train_widx
-            if n=='val':
-                writer = val_writer
-                widx = val_widx
-            path = os.path.join(args.input, ds, '%s.rec'%n)
-            if not os.path.exists(path):
-                continue
-            imgrec = mx.recordio.MXIndexedRecordIO(path[:-3]+'idx', path, 'r')  # pylint: disable=redefined-variable-type
-            if ds=='ms1m':
-                s = imgrec.read_idx(0)
-                header, _ = mx.recordio.unpack(s)
-                assert header.flag>0
-                print('header0 label', header.label)
-                header0 = (int(header.label[0]), int(header.label[1]))
-                #assert(header.flag==1)
-                imgidx = range(1, int(header.label[0]))
-            else:
-                imgidx = list(imgrec.keys)
-            for idx in imgidx:
-                if ds=='megaage' and idx==0:
-                  continue
-                print('info', ds, n, idx)
-                s = imgrec.read_idx(idx)
-                _header, _content = mx.recordio.unpack(s)
-                stat[0]+=1
-                try:
-                    img = mx.image.imdecode(_content)
-                except:
-                    stat[1]+=1
-                    print('error', ds, n, idx)
-                    continue
-                #print(img.shape)
-                if ds=='ms1m':
-                    nlabel = [_header.label]
-                    nlabel += [-1]*101
-                elif ds=='megaage':
-                    #nlabel = [-1, -1]
-                    nlabel = []
-                    age_label = [0]*100
-                    age = int(_header.label[0])
-                    if age>100 or age<0:
-                        continue
-                    age = max(0, min(100, age))
-                    #print('age', age)
-
-                    for a in xrange(0, age):
-                        age_label[a] = 1
-                    nlabel += age_label
-                elif ds=='imdb':
-                    gender = int(_header.label[1])
-                    if args.mode=='gender':
-                      nlabel = [gender]
-                    else:
-                      age_label = [0]*100
-                      age = int(_header.label[0])
-                      age = max(0, min(100, age))
-                      for a in xrange(0, age):
-                          age_label[a] = 1
-                      nlabel = age_label
-                    #nlabel += age_label
-                for r in xrange(repeat):
-                    nheader = mx.recordio.IRHeader(0, nlabel, widx[0], 0)
-                    s = mx.recordio.pack(nheader, _content)
-                    writer.write_idx(widx[0], s)
-                    widx[0]+=1
-    print('stat', stat)
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do dataset merge')
-  # general
-  parser.add_argument('--input', default='', type=str, help='')
-  parser.add_argument('--output', default='', type=str, help='')
-  parser.add_argument('--mode', default='age', type=str, help='')
-  parser.add_argument('--lite', default=1, type=int, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/src/data/agedb2pack.py b/src/data/agedb2pack.py
deleted file mode 100644
index e3abceb14..000000000
--- a/src/data/agedb2pack.py
+++ /dev/null
@@ -1,158 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-#import mxnet as mx
-#from mxnet import ndarray as nd
-import argparse
-import cv2
-import pickle
-import numpy as np
-import sys
-from scipy import misc
-import os
-import tensorflow as tf
-from scipy.io import loadmat
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'align'))
-#sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import detect_face
-import face_image
-import face_preprocess
-#import lfw
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-parser = argparse.ArgumentParser(description='Package AgeDB images')
-# general
-parser.add_argument('--data-dir', default='', help='')
-parser.add_argument('--image-size', type=str, default='112,96', help='')
-parser.add_argument('--output', default='./', help='path to save.')
-args = parser.parse_args()
-
-with tf.Graph().as_default():
-    #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-    #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-    sess = tf.Session()
-    with sess.as_default():
-        pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-
-minsize = 20
-threshold = [0.6,0.7,0.9]
-factor = 0.85
-#minsize = 15
-threshold = [0.6,0.7,0.7]
-#factor = 0.9
-#factor = 0.7
-
-for part in [ ('04_FINAL_protocol_30_years.mat', 'agedb_30') ]:
-  mat_file = os.path.join(args.data_dir, part[0])
-  mat_data = loadmat(mat_file)
-  print(mat_data.__class__)
-  data = mat_data['splits']
-
-  bins = []
-  issame_list = []
-  nrof = [0, 0, 0]
-  print('processing', part[1])
-  pp = 0
-  for i in xrange(data.shape[0]):
-    split = data[i][0][0][0][0]
-    print(split.shape)
-    for c in xrange(split.shape[1]):
-      last_name = ''
-      for r in xrange(split.shape[0]):
-        pp+=1
-        if pp%10==0:
-          print('processing', pp, nrof)
-        item = split[r][c][0][0]
-        path = str(item[0][0])
-        vec = path.split('_')
-        assert len(vec)>=5
-        name = vec[0]
-        if r==1:
-          issame = False
-          if name==last_name:
-            issame = True
-          print(issame)
-          issame_list.append(issame)
-        last_name = name
-        age = int(item[1])
-        #print(path, age)
-        #sys.exit(0)
-        img_path = os.path.join(args.data_dir, '03_Protocol_Images', path+".jpg")
-        #print(img_path)
-        img = misc.imread(img_path)
-        #print(img.shape)
-        if img.ndim == 2:
-          img = to_rgb(img)
-        img = img[:,:,0:3]
-        _bbox = None
-        _landmark = None
-        bounding_boxes, points = detect_face.detect_face(img, minsize, pnet, rnet, onet, threshold, factor)
-        nrof_faces = bounding_boxes.shape[0]
-        if nrof_faces>0:
-          nrof[0]+=1
-        else:
-          bounding_boxes, points = detect_face.detect_face_force(img, minsize, pnet, rnet, onet)
-          nrof_faces = bounding_boxes.shape[0]
-          if nrof_faces>0:
-            nrof[1]+=1
-          else:
-            nrof[2]+=1
-        if nrof_faces>0:
-          det = bounding_boxes[:,0:4]
-          img_size = np.asarray(img.shape)[0:2]
-          bindex = 0
-          if nrof_faces>1:
-              bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-              img_center = img_size / 2
-              offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-              offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-              bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-          _bbox = bounding_boxes[bindex, 0:4]
-          _landmark = points[:, bindex].reshape( (2,5) ).T
-        warped = face_preprocess.preprocess(img, bbox=_bbox, landmark = _landmark, image_size=args.image_size)
-        warped = warped[...,::-1] #to bgr
-        _, s = cv2.imencode('.jpg', warped)
-        bins.append(s)
-  print(nrof)
-  outname = os.path.join(args.output, part[1]+'.bin')
-  with open(outname, 'wb') as f:
-    pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
-
-
-
-
diff --git a/src/data/agedb2pack2.py b/src/data/agedb2pack2.py
deleted file mode 100644
index b0a1d1948..000000000
--- a/src/data/agedb2pack2.py
+++ /dev/null
@@ -1,143 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-#import mxnet as mx
-#from mxnet import ndarray as nd
-import argparse
-import cv2
-import pickle
-import numpy as np
-import sys
-from scipy import misc
-import os
-import tensorflow as tf
-from scipy.io import loadmat
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'align'))
-#sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import detect_face
-import face_image
-import face_preprocess
-#import lfw
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-parser = argparse.ArgumentParser(description='Package AgeDB images')
-# general
-parser.add_argument('--data-dir', default='', help='')
-parser.add_argument('--image-size', type=str, default='112,96', help='')
-parser.add_argument('--output', default='./', help='path to save.')
-args = parser.parse_args()
-
-
-for part in [ ('04_FINAL_protocol_30_years.mat', 'agedb_30') ]:
-  mat_file = os.path.join(args.data_dir, part[0])
-  mat_data = loadmat(mat_file)
-  print(mat_data.__class__)
-  data = mat_data['splits']
-
-  bins = []
-  issame_list = []
-  nrof = [0, 0, 0]
-  print('processing', part[1])
-  pp = 0
-  for i in xrange(data.shape[0]):
-    split = data[i][0][0][0][0]
-    print(split.shape)
-    for c in xrange(split.shape[1]):
-      last_name = ''
-      for r in xrange(split.shape[0]):
-        pp+=1
-        if pp%10==0:
-          print('processing', pp, nrof)
-        item = split[r][c][0][0]
-        path = str(item[0][0])
-        vec = path.split('_')
-        assert len(vec)>=5
-        name = vec[0]
-        if r==1:
-          issame = False
-          if name==last_name:
-            issame = True
-          #print(issame)
-          issame_list.append(issame)
-        last_name = name
-        age = int(item[1])
-        #print(path, age)
-        #sys.exit(0)
-        img_path = os.path.join(args.data_dir, '03_Protocol_Images', path+".jpg")
-        #print(img_path)
-        img = misc.imread(img_path)
-        if img.ndim == 2:
-          img = to_rgb(img)
-        assert img.ndim==3
-        assert img.shape[2]==3
-        #img = img[:,:,0:3]
-        all_landmark = np.zeros( (68,2), dtype=np.float32)
-        pts_file = img_path[0:-3]+"pts"
-        pp=0
-
-        for line in open(pts_file, 'r'):
-          pp+=1
-          pointid = pp-3
-          if pointid<1 or pointid>68:
-            continue
-          point = [float(x) for x in line.strip().split()]
-          assert len(point)==2
-          point = np.array(point).reshape( (1,2) )
-          #print(pointid)
-          all_landmark[pointid-1,:] = point
-
-        
-        _landmark = np.zeros( (5,2), dtype=np.float32)
-        _landmark[0,:] = (all_landmark[36,:]+all_landmark[39,:])/2
-        _landmark[1,:] = (all_landmark[42,:]+all_landmark[45,:])/2
-        _landmark[2,:] = all_landmark[33,:]
-        _landmark[3,:] = all_landmark[48,:]
-        _landmark[4,:] = all_landmark[54,:]
-        _bbox = None
-        warped = face_preprocess.preprocess(img, bbox=_bbox, landmark = _landmark, image_size=args.image_size)
-        warped = warped[...,::-1] #to bgr
-        _, s = cv2.imencode('.jpg', warped)
-        bins.append(s)
-  print(nrof)
-  outname = os.path.join(args.output, part[1]+'.bin')
-  with open(outname, 'wb') as f:
-    pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
-
-
-
-
diff --git a/src/data/cfp2pack.py b/src/data/cfp2pack.py
deleted file mode 100644
index 5cbf2ef66..000000000
--- a/src/data/cfp2pack.py
+++ /dev/null
@@ -1,145 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-#import mxnet as mx
-#from mxnet import ndarray as nd
-import argparse
-import cv2
-import pickle
-import numpy as np
-import sys
-from scipy import misc
-import os
-import tensorflow as tf
-from scipy.io import loadmat
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'align'))
-#sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import detect_face
-import face_image
-import face_preprocess
-#import lfw
-
-def to_rgb(img):
-    w, h = img.shape
-    ret = np.empty((w, h, 3), dtype=np.uint8)
-    ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
-    return ret
-
-
-def IOU(Reframe,GTframe):
-  x1 = Reframe[0];
-  y1 = Reframe[1];
-  width1 = Reframe[2]-Reframe[0];
-  height1 = Reframe[3]-Reframe[1];
-
-  x2 = GTframe[0]
-  y2 = GTframe[1]
-  width2 = GTframe[2]-GTframe[0]
-  height2 = GTframe[3]-GTframe[1]
-
-  endx = max(x1+width1,x2+width2)
-  startx = min(x1,x2)
-  width = width1+width2-(endx-startx)
-
-  endy = max(y1+height1,y2+height2)
-  starty = min(y1,y2)
-  height = height1+height2-(endy-starty)
-
-  if width <=0 or height <= 0:
-    ratio = 0
-  else:
-    Area = width*height
-    Area1 = width1*height1
-    Area2 = width2*height2
-    ratio = Area*1./(Area1+Area2-Area)
-  return ratio
-
-parser = argparse.ArgumentParser(description='Package CFP images')
-# general
-parser.add_argument('--data-dir', default='', help='')
-parser.add_argument('--image-size', type=str, default='112,96', help='')
-parser.add_argument('--output', default='./', help='path to save.')
-args = parser.parse_args()
-
-with tf.Graph().as_default():
-    #gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
-    #sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
-    sess = tf.Session()
-    with sess.as_default():
-        pnet, rnet, onet = detect_face.create_mtcnn(sess, None)
-
-minsize = 20
-threshold = [0.6,0.7,0.9]
-factor = 0.85
-#minsize = 15
-threshold = [0.6,0.7,0.7]
-factor = 0.9
-#factor = 0.7
-
-for part in [ ('CFP_frontal_paris.mat', 'cfp_ff'), ('CFP_profile_paris.mat', 'cfp_fp') ]:
-  mat_file = os.path.join(args.data_dir, part[0])
-  mat_data = loadmat(mat_file)
-  pairs = mat_data['pairs']
-  bins = []
-  issame_list = []
-  nrof = [0, 0, 0]
-  print('processing', part[1], pairs.shape[1])
-  for i in xrange(pairs.shape[1]):
-    if i%10==0:
-      print('processing', i, nrof)
-    pair = pairs[0][i]
-    #print(str(pair[0]), str(pair[1]), int(pair[2]), int(pair[3]))
-    issame = False
-    if int(pair[3])>0:
-      issame = True
-    issame_list.append(issame)
-    fold = int(pair[2])
-    paths = [str(pair[0][0]), str(pair[1][0])]
-    for path in paths:
-      img_path = os.path.join(args.data_dir, 'CFP', path)
-      txt_path = os.path.join(args.data_dir, 'CFP', path[0:-3]+"txt")
-      landmark = []
-      for line in open(txt_path, 'r'):
-        vec = line.strip().split(',')
-        x = np.array( [float(x) for x in vec], dtype=np.float32)
-        landmark.append(x)
-      landmark = np.array(landmark, dtype=np.float32)
-      #print(landmark.shape)
-      img = misc.imread(img_path)
-      _bbox = None
-      _landmark = None
-      bounding_boxes, points = detect_face.detect_face(img, minsize, pnet, rnet, onet, threshold, factor)
-      nrof_faces = bounding_boxes.shape[0]
-      if nrof_faces>0:
-        nrof[0]+=1
-      else:
-        bounding_boxes, points = detect_face.detect_face_force(img, minsize, pnet, rnet, onet)
-        nrof_faces = bounding_boxes.shape[0]
-        if nrof_faces>0:
-          nrof[1]+=1
-        else:
-          nrof[2]+=1
-      if nrof_faces>0:
-        det = bounding_boxes[:,0:4]
-        img_size = np.asarray(img.shape)[0:2]
-        bindex = 0
-        if nrof_faces>1:
-            bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
-            img_center = img_size / 2
-            offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
-            offset_dist_squared = np.sum(np.power(offsets,2.0),0)
-            bindex = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
-        _bbox = bounding_boxes[bindex, 0:4]
-        _landmark = points[:, bindex].reshape( (2,5) ).T
-      warped = face_preprocess.preprocess(img, bbox=_bbox, landmark = _landmark, image_size=args.image_size)
-      warped = warped[...,::-1] #to bgr
-      _, s = cv2.imencode('.jpg', warped)
-      bins.append(s)
-  print(nrof)
-  outname = os.path.join(args.output, part[1]+'.bin')
-  with open(outname, 'wb') as f:
-    pickle.dump((bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
-
-
-
diff --git a/src/data/dataset_c2c.py b/src/data/dataset_c2c.py
deleted file mode 100644
index 7529ae49c..000000000
--- a/src/data/dataset_c2c.py
+++ /dev/null
@@ -1,136 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'common'))
-import face_image
-
-
-def main(args):
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in xrange(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  ctx_num = len(ctx)
-  path_imgrec = os.path.join(args.input, 'train.rec')
-  path_imgidx = os.path.join(args.input, 'train.idx')
-  imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-  outf = open(os.path.join(args.input, 'c2c'), 'w')
-  s = imgrec.read_idx(0)
-  header, _ = mx.recordio.unpack(s)
-  assert header.flag>0
-  print('header0 label', header.label)
-  header0 = (int(header.label[0]), int(header.label[1]))
-  #assert(header.flag==1)
-  imgidx = range(1, int(header.label[0]))
-  id2range = {}
-  seq_identity = range(int(header.label[0]), int(header.label[1]))
-  for identity in seq_identity:
-    s = imgrec.read_idx(identity)
-    header, _ = mx.recordio.unpack(s)
-    id2range[identity] = (int(header.label[0]), int(header.label[1]))
-  print('id2range', len(id2range))
-  prop = face_image.load_property(args.input)
-  image_size = prop.image_size
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  model = mx.mod.Module.load(prefix, epoch, context = ctx)
-  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-  nrof_images = 0
-  nrof_removed = 0
-  idx = 1
-  id2label = {}
-  pp = 0
-  for _id, v in id2range.iteritems():
-    pp+=1
-    if pp%100==0:
-      print('processing id', pp)
-    _list = range(*v)
-    ocontents = []
-    for i in xrange(len(_list)):
-      _idx = _list[i]
-      #print('_idx', _id, _idx)
-      s = imgrec.read_idx(_idx)
-      ocontents.append(s)
-    #continue
-    embeddings = None
-    headers = [None]*len(ocontents)
-    #print(len(ocontents))
-    ba = 0
-    while True:
-      bb = min(ba+args.batch_size, len(ocontents))
-      if ba>=bb:
-        break
-      _batch_size = bb-ba
-      _batch_size2 = max(_batch_size, ctx_num)
-      data = nd.zeros( (_batch_size2,3, image_size[0], image_size[1]) )
-      label = nd.zeros( (_batch_size2,) )
-      count = bb-ba
-      ii=0
-      for i in xrange(ba, bb):
-        header, img = mx.recordio.unpack(ocontents[i])
-        headers[i] = header
-        img = mx.image.imdecode(img)
-        img = nd.transpose(img, axes=(2, 0, 1))
-        data[ii][:] = img
-        label[ii][:] = header.label[0]
-        ii+=1
-      while ii<_batch_size2:
-        data[ii][:] = data[0][:]
-        label[ii][:] = label[0][:]
-        ii+=1
-      db = mx.io.DataBatch(data=(data,), label=(label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      net_out = net_out[0].asnumpy()
-      if embeddings is None:
-        embeddings = np.zeros( (len(ocontents), net_out.shape[1]))
-      embeddings[ba:bb,:] = net_out[0:_batch_size,:]
-      ba = bb
-    embeddings = sklearn.preprocessing.normalize(embeddings)
-    emb_mean = np.mean(embeddings, axis=0, keepdims=True)
-    emb_mean = sklearn.preprocessing.normalize(emb_mean)
-    sim = np.dot(embeddings, emb_mean.T)
-    #print(sim.shape)
-    sims = sim.flatten()
-    assert len(_list)==len(sims)
-    assert len(_list)==len(ocontents)
-    for i in xrange(len(ocontents)):
-      _sim = sims[i]
-      _idx = _list[i]
-      _header = headers[i]
-      #TODO
-      outf.write("%d,%f,%d\n"%(_idx, _sim, int(_header.label[1])))
-  outf.close()
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='')
-  # general
-  parser.add_argument('--input', default='', type=str, help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/src/data/dataset_clean.py b/src/data/dataset_clean.py
deleted file mode 100644
index 83e8131a0..000000000
--- a/src/data/dataset_clean.py
+++ /dev/null
@@ -1,188 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'common'))
-import face_image
-
-
-def do_clean(args):
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in xrange(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  ctx_num = len(ctx)
-  path_imgrec = os.path.join(args.input, 'train.rec')
-  path_imgidx = os.path.join(args.input, 'train.idx')
-  imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-  s = imgrec.read_idx(0)
-  header, _ = mx.recordio.unpack(s)
-  assert header.flag>0
-  print('header0 label', header.label)
-  header0 = (int(header.label[0]), int(header.label[1]))
-  #assert(header.flag==1)
-  imgidx = range(1, int(header.label[0]))
-  id2range = {}
-  seq_identity = range(int(header.label[0]), int(header.label[1]))
-  for identity in seq_identity:
-    s = imgrec.read_idx(identity)
-    header, _ = mx.recordio.unpack(s)
-    id2range[identity] = (int(header.label[0]), int(header.label[1]))
-  print('id2range', len(id2range))
-  prop = face_image.load_property(args.input)
-  image_size = prop.image_size
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  model = mx.mod.Module.load(prefix, epoch, context = ctx)
-  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-  if args.test==0:
-    if not os.path.exists(args.output):
-      os.makedirs(args.output)
-    writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'train.idx'), os.path.join(args.output, 'train.rec'), 'w')
-  nrof_images = 0
-  nrof_removed = 0
-  idx = 1
-  id2label = {}
-  pp = 0
-  for _id, v in id2range.iteritems():
-    pp+=1
-    if pp%100==0:
-      print('stat', nrof_images, nrof_removed)
-    _list = range(*v)
-    ocontents = []
-    for i in xrange(len(_list)):
-      _idx = _list[i]
-      s = imgrec.read_idx(_idx)
-      ocontents.append(s)
-    if len(ocontents)>15:
-      nrof_removed+=len(ocontents)
-      continue
-    embeddings = None
-    #print(len(ocontents))
-    ba = 0
-    while True:
-      bb = min(ba+args.batch_size, len(ocontents))
-      if ba>=bb:
-        break
-      _batch_size = bb-ba
-      _batch_size2 = max(_batch_size, ctx_num)
-      data = nd.zeros( (_batch_size2,3, image_size[0], image_size[1]) )
-      label = nd.zeros( (_batch_size2,) )
-      count = bb-ba
-      ii=0
-      for i in xrange(ba, bb):
-        header, img = mx.recordio.unpack(ocontents[i])
-        img = mx.image.imdecode(img)
-        img = nd.transpose(img, axes=(2, 0, 1))
-        data[ii][:] = img
-        label[ii][:] = header.label
-        ii+=1
-      while ii<_batch_size2:
-        data[ii][:] = data[0][:]
-        label[ii][:] = label[0][:]
-        ii+=1
-      db = mx.io.DataBatch(data=(data,), label=(label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      net_out = net_out[0].asnumpy()
-      if embeddings is None:
-        embeddings = np.zeros( (len(ocontents), net_out.shape[1]))
-      embeddings[ba:bb,:] = net_out[0:_batch_size,:]
-      ba = bb
-    embeddings = sklearn.preprocessing.normalize(embeddings)
-    contents = []
-    if args.mode==1:
-      emb_mean = np.mean(embeddings, axis=0, keepdims=True)
-      emb_mean = sklearn.preprocessing.normalize(emb_mean)
-      sim = np.dot(embeddings, emb_mean.T)
-      #print(sim.shape)
-      sim = sim.flatten()
-      #print(sim.flatten())
-      x = np.argsort(sim)
-      for ix in xrange(len(x)):
-        _idx = x[ix]
-        _sim = sim[_idx]
-        #if ix<int(len(x)*0.3) and _sim<args.threshold:
-        if _sim<args.threshold:
-          continue
-        contents.append(ocontents[_idx])
-    else:
-      y_pred = DBSCAN(eps = args.threshold, min_samples = 2).fit_predict(embeddings)
-      #print(y_pred)
-      gmap = {}
-      for _idx in xrange(embeddings.shape[0]):
-        label = int(y_pred[_idx])
-        if label not in gmap:
-          gmap[label] = []
-        gmap[label].append(_idx)
-      assert len(gmap)>0
-      _max = [0, 0]
-      for label in xrange(10):
-        if not label in gmap:
-          break
-        glist = gmap[label]
-        if len(glist)>_max[1]:
-          _max[0] = label
-          _max[1] = len(glist)
-      if _max[1]>0:
-        glist = gmap[_max[0]]
-        for _idx in glist:
-          contents.append(ocontents[_idx])
-
-    nrof_removed+=(len(ocontents)-len(contents))
-    if len(contents)==0:
-      continue
-    #assert len(contents)>0
-    id2label[_id] = (idx, idx+len(contents))
-    nrof_images += len(contents)
-    for content in contents:
-      if args.test==0:
-        writer.write_idx(idx, content)
-      idx+=1
-  id_idx = idx
-  if args.test==0:
-    for _id, _label in id2label.iteritems():
-      _header = mx.recordio.IRHeader(1, _label, idx, 0)
-      s = mx.recordio.pack(_header, '')
-      writer.write_idx(idx, s)
-      idx+=1
-    _header = mx.recordio.IRHeader(1, (id_idx, idx), 0, 0)
-    s = mx.recordio.pack(_header, '')
-    writer.write_idx(0, s)
-  print(nrof_images, nrof_removed)
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do data clean')
-  # general
-  parser.add_argument('--input', default='', type=str, help='')
-  parser.add_argument('--output', default='', type=str, help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--threshold', default=0.6, type=float, help='')
-  parser.add_argument('--mode', default=1, type=int, help='')
-  parser.add_argument('--test', default=0, type=int, help='')
-  args = parser.parse_args()
-  do_clean(args)
-
diff --git a/src/data/dataset_info.py b/src/data/dataset_info.py
deleted file mode 100644
index 60c08dd4b..000000000
--- a/src/data/dataset_info.py
+++ /dev/null
@@ -1,40 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'common'))
-import face_image
-
-
-def main(args):
-  path_imgrec = os.path.join(args.input, 'train.rec')
-  path_imgidx = os.path.join(args.input, 'train.idx')
-  imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-  s = imgrec.read_idx(0)
-  header, _ = mx.recordio.unpack(s)
-  assert header.flag>0
-  print('header0 label', header.label)
-  header0 = (int(header.label[0]), int(header.label[1]))
-  print('identities', header0[1]-header0[0])
-  print('images', header0[0])
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='')
-  # general
-  parser.add_argument('--input', default='', type=str, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/src/data/dataset_merge.py b/src/data/dataset_merge.py
deleted file mode 100644
index 5b05e0686..000000000
--- a/src/data/dataset_merge.py
+++ /dev/null
@@ -1,293 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'common'))
-import face_image
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'eval'))
-import verification
-
-def ch_dev(arg_params, aux_params, ctx):
-  new_args = dict()
-  new_auxs = dict()
-  for k, v in arg_params.items():
-    new_args[k] = v.as_in_context(ctx)
-  for k, v in aux_params.items():
-    new_auxs[k] = v.as_in_context(ctx)
-  return new_args, new_auxs
-
-def get_embedding(args, imgrec, id, image_size, model):
-  s = imgrec.read_idx(id)
-  header, _ = mx.recordio.unpack(s)
-  ocontents = []
-  for idx in xrange(int(header.label[0]), int(header.label[1])):
-    #print('idx', idx)
-    s = imgrec.read_idx(idx)
-    ocontents.append(s)
-  embeddings = None
-  #print(len(ocontents))
-  ba = 0
-  while True:
-    bb = min(ba+args.batch_size, len(ocontents))
-    if ba>=bb:
-      break
-    _batch_size = bb-ba
-    _batch_size2 = max(_batch_size, args.ctx_num)
-    data = nd.zeros( (_batch_size2,3, image_size[0], image_size[1]) )
-    #label = nd.zeros( (_batch_size2,) )
-    count = bb-ba
-    ii=0
-    for i in xrange(ba, bb):
-      header, img = mx.recordio.unpack(ocontents[i])
-      #print(header.label.shape, header.label)
-      img = mx.image.imdecode(img)
-      img = nd.transpose(img, axes=(2, 0, 1))
-      data[ii][:] = img
-      #label[ii][:] = header.label
-      ii+=1
-    while ii<_batch_size2:
-      data[ii][:] = data[0][:]
-      #label[ii][:] = label[0][:]
-      ii+=1
-    #db = mx.io.DataBatch(data=(data,), label=(label,))
-    db = mx.io.DataBatch(data=(data,))
-    model.forward(db, is_train=False)
-    net_out = model.get_outputs()
-    net_out = net_out[0].asnumpy()
-    if embeddings is None:
-      embeddings = np.zeros( (len(ocontents), net_out.shape[1]))
-    embeddings[ba:bb,:] = net_out[0:_batch_size,:]
-    ba = bb
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  embedding = np.mean(embeddings, axis=0, keepdims=True)
-  embedding = sklearn.preprocessing.normalize(embedding).flatten()
-  return embedding
-
-def main(args):
-  include_datasets = args.include.split(',')
-  prop = face_image.load_property(include_datasets[0])
-  image_size = prop.image_size
-  print('image_size', image_size)
-  model = None
-  if len(args.model)>0:
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in xrange(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    args.ctx_num = len(ctx)
-    vec = args.model.split(',')
-    prefix = vec[0]
-    epoch = int(vec[1])
-    print('loading',prefix, epoch)
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    #model = mx.mod.Module.load(prefix, epoch, context = ctx)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-  else:
-    assert args.param1==0.0
-  rec_list = []
-  for ds in include_datasets:
-    path_imgrec = os.path.join(ds, 'train.rec')
-    path_imgidx = os.path.join(ds, 'train.idx')
-    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-    rec_list.append(imgrec)
-  id_list_map = {}
-  all_id_list = []
-  test_limit = 0
-  for ds_id in xrange(len(rec_list)):
-    id_list = []
-    imgrec = rec_list[ds_id]
-    s = imgrec.read_idx(0)
-    header, _ = mx.recordio.unpack(s)
-    assert header.flag>0
-    print('header0 label', header.label)
-    header0 = (int(header.label[0]), int(header.label[1]))
-    #assert(header.flag==1)
-    imgidx = range(1, int(header.label[0]))
-    id2range = {}
-    seq_identity = range(int(header.label[0]), int(header.label[1]))
-    pp=0
-    for identity in seq_identity:
-      pp+=1
-      if pp%10==0:
-        print('processing id', pp)
-      if model is not None:
-        embedding = get_embedding(args, imgrec, identity, image_size, model)
-      else:
-        embedding = None
-      #print(embedding.shape)
-      id_list.append( [ds_id, identity, embedding] )
-      if test_limit>0 and pp>=test_limit:
-        break
-    id_list_map[ds_id] = id_list
-    if ds_id==0 or model is None:
-      all_id_list += id_list
-      print(ds_id, len(id_list))
-    else:
-      X = []
-      for id_item in all_id_list:
-        X.append(id_item[2])
-      X = np.array(X)
-      for i in xrange(len(id_list)):
-        id_item = id_list[i]
-        y = id_item[2]
-        sim = np.dot(X, y.T)
-        idx = np.where(sim>=args.param1)[0]
-        if len(idx)>0:
-          continue
-        all_id_list.append(id_item)
-      print(ds_id, len(id_list), len(all_id_list))
-
-
-  if len(args.exclude)>0:
-    if os.path.isdir(args.exclude):
-      _path_imgrec = os.path.join(args.exclude, 'train.rec')
-      _path_imgidx = os.path.join(args.exclude, 'train.idx')
-      _imgrec = mx.recordio.MXIndexedRecordIO(_path_imgidx, _path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-      _ds_id = len(rec_list)
-      _id_list = []
-      s = _imgrec.read_idx(0)
-      header, _ = mx.recordio.unpack(s)
-      assert header.flag>0
-      print('header0 label', header.label)
-      header0 = (int(header.label[0]), int(header.label[1]))
-      #assert(header.flag==1)
-      imgidx = range(1, int(header.label[0]))
-      seq_identity = range(int(header.label[0]), int(header.label[1]))
-      pp=0
-      for identity in seq_identity:
-        pp+=1
-        if pp%10==0:
-          print('processing ex id', pp)
-        embedding = get_embedding(args, _imgrec, identity, image_size, model)
-        #print(embedding.shape)
-        _id_list.append( (_ds_id, identity, embedding) )
-        if test_limit>0 and pp>=test_limit:
-          break
-    else:
-      _id_list = []
-      data_set = verification.load_bin(args.exclude, image_size)[0][0]
-      print(data_set.shape)
-      data = nd.zeros( (1,3,image_size[0], image_size[1]))
-      for i in xrange(data_set.shape[0]):
-        data[0] = data_set[i]
-        db = mx.io.DataBatch(data=(data,))
-        model.forward(db, is_train=False)
-        net_out = model.get_outputs()
-        embedding = net_out[0].asnumpy().flatten()
-        _norm=np.linalg.norm(embedding)
-        embedding /= _norm
-        _id_list.append( (i, i, embedding) )
-
-    #X = []
-    #for id_item in all_id_list:
-    #  X.append(id_item[2])
-    #X = np.array(X)
-    #param1 = 0.3
-    #while param1<=1.01:
-    #  emap = {}
-    #  for id_item in _id_list:
-    #    y = id_item[2]
-    #    sim = np.dot(X, y.T)
-    #    #print(sim.shape)
-    #    #print(sim)
-    #    idx = np.where(sim>=param1)[0]
-    #    for j in idx:
-    #      emap[j] = 1
-    #  exclude_removed = len(emap)
-    #  print(param1, exclude_removed)
-    #  param1+=0.05
-
-      X = []
-      for id_item in all_id_list:
-        X.append(id_item[2])
-      X = np.array(X)
-      emap = {}
-      for id_item in _id_list:
-        y = id_item[2]
-        sim = np.dot(X, y.T)
-        idx = np.where(sim>=args.param2)[0]
-        for j in idx:
-          emap[j] = 1
-          all_id_list[j][1] = -1
-      print('exclude', len(emap))
-
-  if args.test>0:
-    return
-
-  if not os.path.exists(args.output):
-    os.makedirs(args.output)
-  writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'train.idx'), os.path.join(args.output, 'train.rec'), 'w')
-  idx = 1
-  identities = []
-  nlabel = -1
-  for id_item in all_id_list:
-    if id_item[1]<0:
-      continue
-    nlabel+=1
-    ds_id = id_item[0]
-    imgrec = rec_list[ds_id]
-    id = id_item[1]
-    s = imgrec.read_idx(id)
-    header, _ = mx.recordio.unpack(s)
-    a, b = int(header.label[0]), int(header.label[1])
-    identities.append( (idx, idx+b-a) )
-    for _idx in xrange(a,b):
-      s = imgrec.read_idx(_idx)
-      _header, _content = mx.recordio.unpack(s)
-      nheader = mx.recordio.IRHeader(0, nlabel, idx, 0)
-      s = mx.recordio.pack(nheader, _content)
-      writer.write_idx(idx, s)
-      idx+=1
-  id_idx = idx
-  for id_label in identities:
-    _header = mx.recordio.IRHeader(1, id_label, idx, 0)
-    s = mx.recordio.pack(_header, '')
-    writer.write_idx(idx, s)
-    idx+=1
-  _header = mx.recordio.IRHeader(1, (id_idx, idx), 0, 0)
-  s = mx.recordio.pack(_header, '')
-  writer.write_idx(0, s)
-  with open(os.path.join(args.output, 'property'), 'w') as f:
-    f.write("%d,%d,%d"%(len(identities), image_size[0], image_size[1]))
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do dataset merge')
-  # general
-  parser.add_argument('--include', default='', type=str, help='')
-  parser.add_argument('--exclude', default='', type=str, help='')
-  parser.add_argument('--output', default='', type=str, help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--param1', default=0.3, type=float, help='')
-  parser.add_argument('--param2', default=0.4, type=float, help='')
-  parser.add_argument('--mode', default=1, type=int, help='')
-  parser.add_argument('--test', default=0, type=int, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/src/data/dataset_relabel.py b/src/data/dataset_relabel.py
deleted file mode 100644
index 0fb29bb26..000000000
--- a/src/data/dataset_relabel.py
+++ /dev/null
@@ -1,70 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-
-sys.path.append(os.path.join(os.path.dirname(__file__),'..', 'common'))
-import face_image
-
-def main(args):
-  include_datasets = args.include.split(',')
-  rec_list = []
-  for ds in include_datasets:
-    path_imgrec = os.path.join(ds, 'train.rec')
-    path_imgidx = os.path.join(ds, 'train.idx')
-    imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-    rec_list.append(imgrec)
-  if not os.path.exists(args.output):
-    os.makedirs(args.output)
-  writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'train.idx'), os.path.join(args.output, 'train.rec'), 'w')
-  for ds_id in xrange(len(rec_list)):
-    id_list = []
-    imgrec = rec_list[ds_id]
-    s = imgrec.read_idx(0)
-    writer.write_idx(0, s)
-    header, _ = mx.recordio.unpack(s)
-    assert header.flag>0
-    print('header0 label', header.label)
-    header0 = (int(header.label[0]), int(header.label[1]))
-    seq_identity = range(int(header.label[0]), int(header.label[1]))
-    pp=0
-    nlabel = -1
-    for identity in seq_identity:
-      pp+=1
-      if pp%10==0:
-        print('processing id', pp)
-      s = imgrec.read_idx(identity)
-      writer.write_idx(identity, s)
-      header, _ = mx.recordio.unpack(s)
-      nlabel+=1
-      for _idx in xrange(int(header.label[0]), int(header.label[1])):
-        s = imgrec.read_idx(_idx)
-        _header, _content = mx.recordio.unpack(s)
-        nheader = mx.recordio.IRHeader(0, nlabel, _idx, 0)
-        s = mx.recordio.pack(nheader, _content)
-        writer.write_idx(_idx, s)
-
-    print('max label', nlabel)
-
-
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do dataset merge')
-  # general
-  parser.add_argument('--include', default='', type=str, help='')
-  parser.add_argument('--output', default='', type=str, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/src/data/dir2lst.py b/src/data/dir2lst.py
deleted file mode 100644
index 0ff7b724a..000000000
--- a/src/data/dir2lst.py
+++ /dev/null
@@ -1,14 +0,0 @@
-
-import sys
-import os
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-
-input_dir = sys.argv[1]
-
-dataset = face_image.get_dataset_common(input_dir, 2)
-
-for item in dataset:
-  print("%d\t%s\t%d" % (1, item.image_path, int(item.classname)))
-
-
diff --git a/src/data/dir2lst_ytf.py b/src/data/dir2lst_ytf.py
deleted file mode 100644
index 3f498ec0b..000000000
--- a/src/data/dir2lst_ytf.py
+++ /dev/null
@@ -1,32 +0,0 @@
-import sys
-import os
-from easydict import EasyDict as edict
-
-input_dir = '/raid5data/dplearn/YTF/aligned_images_DB'
-ret = []
-label = 0
-person_names = []
-for person_name in os.listdir(input_dir):
-  person_names.append(person_name)
-person_names = sorted(person_names)
-for person_name in person_names:
-  _subdir = os.path.join(input_dir, person_name)
-  if not os.path.isdir(_subdir):
-    continue
-  for _subdir2 in os.listdir(_subdir):
-    _subdir2 = os.path.join(_subdir, _subdir2)
-    if not os.path.isdir(_subdir2):
-      continue
-    _ret = []
-    for img in os.listdir(_subdir2):
-      fimage = edict()
-      fimage.id = os.path.join(_subdir2, img)
-      fimage.classname = str(label)
-      fimage.image_path = os.path.join(_subdir2, img)
-      fimage.bbox = None
-      fimage.landmark = None
-      _ret.append(fimage)
-    ret += _ret
-  label+=1
-for item in ret:
-  print("%d\t%s\t%d" % (1, item.image_path, int(item.classname)))
diff --git a/src/data/dir2rec.py b/src/data/dir2rec.py
deleted file mode 100644
index c57d5d3aa..000000000
--- a/src/data/dir2rec.py
+++ /dev/null
@@ -1,205 +0,0 @@
-
-# -*- coding: utf-8 -*-
-from __future__ import print_function
-import os
-import sys
-
-import mxnet as mx
-import random
-import argparse
-import cv2
-import time
-import traceback
-import numpy as np
-#from builtins import range
-from easydict import EasyDict as edict
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_preprocess
-import face_image
-
-try:
-    import multiprocessing
-except ImportError:
-    multiprocessing = None
-
-
-
-def read_label(path_in):
-  identities = []
-  last = [-1, -1]
-  _id = 1
-  dir2label = {}
-  for line in open(path_in, 'r'):
-    line = line.strip().split()
-    item = edict()
-    item.flag = 0
-    item.image_path = os.path.join(args.input, 'images', line[0])
-    image_dir = line[0].split('/')[0]
-    if image_dir in dir2label:
-      label = dir2label[image_dir]
-    else:
-      label = len(dir2label)
-      dir2label[image_dir] = label
-    item.bbox = np.array( [float(x) for x in line[1:5]], dtype=np.float32 )
-    item.landmark = np.array( [float(x) for x in line[5:15]], dtype=np.float32 ).reshape( (5,2) )
-    item.aligned = False
-    item.id = _id
-    item.label = label
-    yield item
-    if label!=last[0]:
-      if last[1]>=0:
-        identities.append( (last[1], _id) )
-      last[0] = label
-      last[1] = _id
-    _id+=1
-  identities.append( (last[1], _id) )
-  item = edict()
-  item.flag = 2
-  item.id = 0
-  item.label = [float(_id), float(_id+len(identities))]
-  yield item
-  for identity in identities:
-    item = edict()
-    item.flag = 2
-    item.id = _id
-    _id+=1
-    item.label = [float(identity[0]), float(identity[1])]
-    yield item
-
-
-
-def image_encode(args, i, item, q_out):
-    #print('flag', item.flag)
-    if item.flag==0:
-      oitem = [item.id, item.label]
-      fullpath = item.image_path
-      header = mx.recordio.IRHeader(item.flag, item.label, item.id, 0)
-      #print('write', item.flag, item.id, item.label)
-      if item.aligned:
-        with open(fullpath, 'rb') as fin:
-            img = fin.read()
-        s = mx.recordio.pack(header, img)
-        q_out.put((i, s, oitem))
-      else:
-        img = cv2.imread(fullpath, args.color)
-        assert item.landmark is not None
-        img = face_preprocess.preprocess(img, bbox = item.bbox, landmark=item.landmark, image_size='%d,%d'%(args.image_h, args.image_w))
-        s = mx.recordio.pack_img(header, img, quality=args.quality, img_fmt=args.encoding)
-        q_out.put((i, s, oitem))
-    else: 
-      oitem = [item.id, -1]
-      header = mx.recordio.IRHeader(item.flag, item.label, item.id, 0)
-      #print('write', item.flag, item.id, item.label)
-      s = mx.recordio.pack(header, '')
-      q_out.put((i, s, oitem))
-
-
-def read_worker(args, q_in, q_out):
-    while True:
-        deq = q_in.get()
-        if deq is None:
-            break
-        i, item = deq
-        image_encode(args, i, item, q_out)
-
-def write_worker(q_out, output_dir):
-    pre_time = time.time()
-    count = 0
-    record = mx.recordio.MXIndexedRecordIO(os.path.join(output_dir, 'train.idx'),
-                                           os.path.join(output_dir, 'train.rec'), 'w')
-    buf = {}
-    more = True
-    max_label = 0
-    while more:
-        deq = q_out.get()
-        if deq is not None:
-            i, s, item = deq
-            buf[i] = (s, item)
-        else:
-            more = False
-        while count in buf:
-            s, item = buf[count]
-            label = item[1]
-            #print('label', label)
-            max_label = max(max_label, label)
-            del buf[count]
-            if s is not None:
-                #print('write idx', item[0])
-                record.write_idx(item[0], s)
-
-            if count % 1000 == 0:
-                cur_time = time.time()
-                print('time:', cur_time - pre_time, ' count:', count)
-                pre_time = cur_time
-            count += 1
-    print('writing', output_dir, 'property', max_label)
-    with open(os.path.join(output_dir, 'property'), 'w') as outf:
-      outf.write("%d,112,112"%(max_label+1))
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(
-        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
-        description='Create an image list or \
-        make a record database by reading from an image list')
-    parser.add_argument('--input', help='input data dir.')
-    parser.add_argument('--output', help='outputdata dir.')
-    #parser.add_argument('root', help='path to folder containing images.')
-
-    cgroup = parser.add_argument_group('Options for creating image lists')
-    cgroup.add_argument('--exts', nargs='+', default=['.jpeg', '.jpg'],
-                        help='list of acceptable image extensions.')
-    cgroup.add_argument('--chunks', type=int, default=1, help='number of chunks.')
-    cgroup.add_argument('--train-ratio', type=float, default=1.0,
-                        help='Ratio of images to use for training.')
-    cgroup.add_argument('--test-ratio', type=float, default=0,
-                        help='Ratio of images to use for testing.')
-
-    rgroup = parser.add_argument_group('Options for creating database')
-    rgroup.add_argument('--quality', type=int, default=95,
-                        help='JPEG quality for encoding, 1-100; or PNG compression for encoding, 1-9')
-    rgroup.add_argument('--num-thread', type=int, default=1,
-                        help='number of thread to use for encoding. order of images will be different\
-        from the input list if >1. the input list will be modified to match the\
-        resulting order.')
-    rgroup.add_argument('--color', type=int, default=1, choices=[-1, 0, 1],
-                        help='specify the color mode of the loaded image.\
-        1: Loads a color image. Any transparency of image will be neglected. It is the default flag.\
-        0: Loads image in grayscale mode.\
-        -1:Loads image as such including alpha channel.')
-    rgroup.add_argument('--encoding', type=str, default='.jpg', choices=['.jpg', '.png'],
-                        help='specify the encoding of the images.')
-    args = parser.parse_args()
-    return args
-
-if __name__ == '__main__':
-    args = parse_args()
-    label_file = os.path.join(args.input, 'label.txt')
-    assert os.path.exists(label_file)
-    if not os.path.exists(args.output):
-      os.makedirs(args.output)
-    image_size = [112, 112]
-    print('image_size', image_size)
-    args.image_h = image_size[0]
-    args.image_w = image_size[1]
-    image_list = read_label(label_file)
-    # -- write_record -- #
-    q_in = [multiprocessing.Queue(1024) for i in range(args.num_thread)]
-    q_out = multiprocessing.Queue(1024)
-    read_process = [multiprocessing.Process(target=read_worker, args=(args, q_in[i], q_out)) \
-                    for i in range(args.num_thread)]
-    for p in read_process:
-        p.start()
-    write_process = multiprocessing.Process(target=write_worker, args=(q_out, args.output))
-    write_process.start()
-
-    for i, item in enumerate(image_list):
-        q_in[i % len(q_in)].put((i, item))
-    for q in q_in:
-        q.put(None)
-    for p in read_process:
-        p.join()
-
-    q_out.put(None)
-    write_process.join()
-
diff --git a/src/data/face2rec2.py b/src/data/face2rec2.py
deleted file mode 100644
index fc1507a7d..000000000
--- a/src/data/face2rec2.py
+++ /dev/null
@@ -1,267 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-# -*- coding: utf-8 -*-
-from __future__ import print_function
-import os
-import sys
-
-#curr_path = os.path.abspath(os.path.dirname(__file__))
-#sys.path.append(os.path.join(curr_path, "../python"))
-import mxnet as mx
-import random
-import argparse
-import cv2
-import time
-import traceback
-#from builtins import range
-from easydict import EasyDict as edict
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_preprocess
-import face_image
-
-try:
-    import multiprocessing
-except ImportError:
-    multiprocessing = None
-
-
-
-def read_list(path_in):
-    with open(path_in) as fin:
-        identities = []
-        last = [-1, -1]
-        _id = 1
-        while True:
-            line = fin.readline()
-            if not line:
-                break
-            item = edict()
-            item.flag = 0
-            item.image_path, label, item.bbox, item.landmark, item.aligned = face_preprocess.parse_lst_line(line)
-            if not item.aligned and item.landmark is None:
-              #print('ignore line', line)
-              continue
-            item.id = _id
-            item.label = [label, item.aligned]
-            yield item
-            if label!=last[0]:
-              if last[1]>=0:
-                identities.append( (last[1], _id) )
-              last[0] = label
-              last[1] = _id
-            _id+=1
-        identities.append( (last[1], _id) )
-        item = edict()
-        item.flag = 2
-        item.id = 0
-        item.label = [float(_id), float(_id+len(identities))]
-        yield item
-        for identity in identities:
-          item = edict()
-          item.flag = 2
-          item.id = _id
-          _id+=1
-          item.label = [float(identity[0]), float(identity[1])]
-          yield item
-
-
-
-def image_encode(args, i, item, q_out):
-    oitem = [item.id]
-    #print('flag', item.flag)
-    if item.flag==0:
-      fullpath = item.image_path
-      header = mx.recordio.IRHeader(item.flag, item.label, item.id, 0)
-      #print('write', item.flag, item.id, item.label)
-      if item.aligned:
-        with open(fullpath, 'rb') as fin:
-            img = fin.read()
-        s = mx.recordio.pack(header, img)
-        q_out.put((i, s, oitem))
-      else:
-        img = cv2.imread(fullpath, args.color)
-        assert item.landmark is not None
-        img = face_preprocess.preprocess(img, bbox = item.bbox, landmark=item.landmark, image_size='%d,%d'%(args.image_h, args.image_w))
-        s = mx.recordio.pack_img(header, img, quality=args.quality, img_fmt=args.encoding)
-        q_out.put((i, s, oitem))
-    else: 
-      header = mx.recordio.IRHeader(item.flag, item.label, item.id, 0)
-      #print('write', item.flag, item.id, item.label)
-      s = mx.recordio.pack(header, '')
-      q_out.put((i, s, oitem))
-
-
-def read_worker(args, q_in, q_out):
-    while True:
-        deq = q_in.get()
-        if deq is None:
-            break
-        i, item = deq
-        image_encode(args, i, item, q_out)
-
-def write_worker(q_out, fname, working_dir):
-    pre_time = time.time()
-    count = 0
-    fname = os.path.basename(fname)
-    fname_rec = os.path.splitext(fname)[0] + '.rec'
-    fname_idx = os.path.splitext(fname)[0] + '.idx'
-    record = mx.recordio.MXIndexedRecordIO(os.path.join(working_dir, fname_idx),
-                                           os.path.join(working_dir, fname_rec), 'w')
-    buf = {}
-    more = True
-    while more:
-        deq = q_out.get()
-        if deq is not None:
-            i, s, item = deq
-            buf[i] = (s, item)
-        else:
-            more = False
-        while count in buf:
-            s, item = buf[count]
-            del buf[count]
-            if s is not None:
-                #print('write idx', item[0])
-                record.write_idx(item[0], s)
-
-            if count % 1000 == 0:
-                cur_time = time.time()
-                print('time:', cur_time - pre_time, ' count:', count)
-                pre_time = cur_time
-            count += 1
-
-def parse_args():
-    parser = argparse.ArgumentParser(
-        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
-        description='Create an image list or \
-        make a record database by reading from an image list')
-    parser.add_argument('prefix', help='prefix of input/output lst and rec files.')
-    #parser.add_argument('root', help='path to folder containing images.')
-
-    cgroup = parser.add_argument_group('Options for creating image lists')
-    cgroup.add_argument('--list', type=bool, default=False,
-                        help='If this is set im2rec will create image list(s) by traversing root folder\
-        and output to <prefix>.lst.\
-        Otherwise im2rec will read <prefix>.lst and create a database at <prefix>.rec')
-    cgroup.add_argument('--exts', nargs='+', default=['.jpeg', '.jpg'],
-                        help='list of acceptable image extensions.')
-    cgroup.add_argument('--chunks', type=int, default=1, help='number of chunks.')
-    cgroup.add_argument('--train-ratio', type=float, default=1.0,
-                        help='Ratio of images to use for training.')
-    cgroup.add_argument('--test-ratio', type=float, default=0,
-                        help='Ratio of images to use for testing.')
-    cgroup.add_argument('--recursive', type=bool, default=False,
-                        help='If true recursively walk through subdirs and assign an unique label\
-        to images in each folder. Otherwise only include images in the root folder\
-        and give them label 0.')
-    cgroup.add_argument('--shuffle', type=bool, default=True, help='If this is set as True, \
-        im2rec will randomize the image order in <prefix>.lst')
-
-    rgroup = parser.add_argument_group('Options for creating database')
-    rgroup.add_argument('--quality', type=int, default=95,
-                        help='JPEG quality for encoding, 1-100; or PNG compression for encoding, 1-9')
-    rgroup.add_argument('--num-thread', type=int, default=1,
-                        help='number of thread to use for encoding. order of images will be different\
-        from the input list if >1. the input list will be modified to match the\
-        resulting order.')
-    rgroup.add_argument('--color', type=int, default=1, choices=[-1, 0, 1],
-                        help='specify the color mode of the loaded image.\
-        1: Loads a color image. Any transparency of image will be neglected. It is the default flag.\
-        0: Loads image in grayscale mode.\
-        -1:Loads image as such including alpha channel.')
-    rgroup.add_argument('--encoding', type=str, default='.jpg', choices=['.jpg', '.png'],
-                        help='specify the encoding of the images.')
-    rgroup.add_argument('--pack-label', type=bool, default=False,
-        help='Whether to also pack multi dimensional label in the record file')
-    args = parser.parse_args()
-    args.prefix = os.path.abspath(args.prefix)
-    #args.root = os.path.abspath(args.root)
-    return args
-
-if __name__ == '__main__':
-    args = parse_args()
-    if args.list:
-        pass
-        #make_list(args)
-    else:
-        if os.path.isdir(args.prefix):
-            working_dir = args.prefix
-        else:
-            working_dir = os.path.dirname(args.prefix)
-        prop = face_image.load_property(working_dir)
-        image_size = prop.image_size
-        print('image_size', image_size)
-        args.image_h = image_size[0]
-        args.image_w = image_size[1]
-        files = [os.path.join(working_dir, fname) for fname in os.listdir(working_dir)
-                    if os.path.isfile(os.path.join(working_dir, fname))]
-        count = 0
-        for fname in files:
-            if fname.startswith(args.prefix) and fname.endswith('.lst'):
-                print('Creating .rec file from', fname, 'in', working_dir)
-                count += 1
-                image_list = read_list(fname)
-                # -- write_record -- #
-                if args.num_thread > 1 and multiprocessing is not None:
-                    q_in = [multiprocessing.Queue(1024) for i in range(args.num_thread)]
-                    q_out = multiprocessing.Queue(1024)
-                    read_process = [multiprocessing.Process(target=read_worker, args=(args, q_in[i], q_out)) \
-                                    for i in range(args.num_thread)]
-                    for p in read_process:
-                        p.start()
-                    write_process = multiprocessing.Process(target=write_worker, args=(q_out, fname, working_dir))
-                    write_process.start()
-
-                    for i, item in enumerate(image_list):
-                        q_in[i % len(q_in)].put((i, item))
-                    for q in q_in:
-                        q.put(None)
-                    for p in read_process:
-                        p.join()
-
-                    q_out.put(None)
-                    write_process.join()
-                else:
-                    print('multiprocessing not available, fall back to single threaded encoding')
-                    try:
-                        import Queue as queue
-                    except ImportError:
-                        import queue
-                    q_out = queue.Queue()
-                    fname = os.path.basename(fname)
-                    fname_rec = os.path.splitext(fname)[0] + '.rec'
-                    fname_idx = os.path.splitext(fname)[0] + '.idx'
-                    record = mx.recordio.MXIndexedRecordIO(os.path.join(working_dir, fname_idx),
-                                                           os.path.join(working_dir, fname_rec), 'w')
-                    cnt = 0
-                    pre_time = time.time()
-                    for i, item in enumerate(image_list):
-                        image_encode(args, i, item, q_out)
-                        if q_out.empty():
-                            continue
-                        _, s, item = q_out.get()
-                        #header, _ = mx.recordio.unpack(s)
-                        #print('write header label', header.label)
-                        record.write_idx(item[0], s)
-                        if cnt % 1000 == 0:
-                            cur_time = time.time()
-                            print('time:', cur_time - pre_time, ' count:', cnt)
-                            pre_time = cur_time
-                        cnt += 1
-        if not count:
-            print('Did not find and list file with prefix %s'%args.prefix)
-
diff --git a/src/data/glint2lst.py b/src/data/glint2lst.py
deleted file mode 100644
index 60f7cbea9..000000000
--- a/src/data/glint2lst.py
+++ /dev/null
@@ -1,42 +0,0 @@
-
-import sys
-import os
-import numpy as np
-
-input_dir = sys.argv[1]
-targets = sys.argv[2]
-targets = targets.strip().split(',')
-lmap = {}
-
-for ds in targets:
-  #image_dir = os.path.join(input_dir, ds)
-  lmk_file = os.path.join(input_dir, "%s_lmk"%(ds))
-  if not os.path.exists(lmk_file):
-    lmk_file = os.path.join(input_dir, "%s_lmk.txt"%(ds))
-    if not os.path.exists(lmk_file):
-      continue
-  #print(ds)
-  idx = 0
-  for line in open(lmk_file, 'r'):
-    idx+=1
-    vec = line.strip().split(' ')
-    assert len(vec)==12 or len(vec)==11
-    image_file = os.path.join(input_dir, vec[0])
-    assert image_file.endswith('.jpg')
-    vlabel = -1 #test mode
-    if len(vec)==12:
-      label = int(vec[1])
-      if label in lmap:
-        vlabel = lmap[label]
-      else:
-        vlabel = len(lmap)
-        lmap[label] = vlabel
-      lmk = np.array([float(x) for x in vec[2:]], dtype=np.float32)
-    else:
-      lmk = np.array([float(x) for x in vec[1:]], dtype=np.float32)
-    lmk = lmk.reshape( (5,2) ).T
-    lmk_str = "\t".join( [str(x) for x in lmk.flatten()] )
-    print("0\t%s\t%d\t0\t0\t0\t0\t%s"%(image_file, vlabel, lmk_str))
-    #if idx>10:
-    #  break
-
diff --git a/src/data/lfw2pack.py b/src/data/lfw2pack.py
deleted file mode 100644
index 68ff92668..000000000
--- a/src/data/lfw2pack.py
+++ /dev/null
@@ -1,35 +0,0 @@
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import pickle
-import sys
-import os
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'eval'))
-import lfw
-
-parser = argparse.ArgumentParser(description='Package LFW images')
-# general
-parser.add_argument('--data-dir', default='', help='')
-parser.add_argument('--image-size', type=str, default='112,96', help='')
-parser.add_argument('--output', default='', help='path to save.')
-args = parser.parse_args()
-lfw_dir = args.data_dir
-image_size = [int(x) for x in args.image_size.split(',')]
-lfw_pairs = lfw.read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
-lfw_paths, issame_list = lfw.get_paths(lfw_dir, lfw_pairs, 'jpg')
-lfw_bins = []
-#lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
-i = 0
-for path in lfw_paths:
-  with open(path, 'rb') as fin:
-    _bin = fin.read()
-    lfw_bins.append(_bin)
-    #img = mx.image.imdecode(_bin)
-    #img = nd.transpose(img, axes=(2, 0, 1))
-    #lfw_data[i][:] = img
-    i+=1
-    if i%1000==0:
-      print('loading lfw', i)
-
-with open(args.output, 'wb') as f:
-  pickle.dump((lfw_bins, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
diff --git a/src/eval/do_ver.sh b/src/eval/do_ver.sh
deleted file mode 100755
index 31f84af13..000000000
--- a/src/eval/do_ver.sh
+++ /dev/null
@@ -1,3 +0,0 @@
-
-#python -u verification.py --gpu 0 --data-dir /opt/jiaguo/faces_vgg_112x112 --image-size 112,112 --model '../../model/softmax1010d3-r101-p0_0_96_112_0,21|22|32' --target agedb_30
-python -u verification.py --gpu 0 --data-dir /opt/jiaguo/faces_normed --image-size 112,96 --model '../../model31/sphere-m51-p0_0_96_112_0,90' --target agedb_30 --batch-size 128
diff --git a/src/eval/gen_glint.py b/src/eval/gen_glint.py
deleted file mode 100644
index 6068a57cc..000000000
--- a/src/eval/gen_glint.py
+++ /dev/null
@@ -1,155 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-
-from datetime import datetime
-import os.path
-from easydict import EasyDict as edict
-import time
-import json
-import sys
-import numpy as np
-import importlib
-import itertools
-import argparse
-import struct
-import cv2
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_preprocess
-import sklearn
-from sklearn.preprocessing import normalize
-import mxnet as mx
-from mxnet import ndarray as nd
-
-image_shape = None
-net = None
-data_size = 1862120
-emb_size = 0
-use_flip = True
-
-
-
-def do_flip(data):
-  for idx in xrange(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_feature(buffer):
-  global emb_size
-  if use_flip:
-    input_blob = np.zeros( (len(buffer)*2, 3, image_shape[1], image_shape[2] ) )
-  else:
-    input_blob = np.zeros( (len(buffer), 3, image_shape[1], image_shape[2] ) )
-  idx = 0
-  for item in buffer:
-    img = face_preprocess.read_image(item[0], mode='rgb')
-    img = face_preprocess.preprocess(img, bbox=None, landmark=item[1], image_size='%d,%d'%(image_shape[1], image_shape[2]))
-    img = np.transpose( img, (2,0,1) )
-    attempts = [0,1] if use_flip else [0]
-    for flipid in attempts:
-      _img = np.copy(img)
-      if flipid==1:
-        do_flip(_img)
-      input_blob[idx] = _img
-      idx+=1
-  data = mx.nd.array(input_blob)
-  db = mx.io.DataBatch(data=(data,))
-  net.model.forward(db, is_train=False)
-  _embedding = net.model.get_outputs()[0].asnumpy()
-  if emb_size==0:
-    emb_size = _embedding.shape[1]
-    print('set emb_size to ', emb_size)
-  embedding = np.zeros( (len(buffer), emb_size), dtype=np.float32 )
-  if use_flip:
-    embedding1 = _embedding[0::2]
-    embedding2 = _embedding[1::2]
-    embedding = embedding1+embedding2
-  else:
-    embedding = _embedding
-  embedding = sklearn.preprocessing.normalize(embedding)
-  return embedding
-
-
-
-def write_bin(path, m):
-  rows, cols = m.shape
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', rows,cols,cols*4,5))
-    f.write(m.data)
-
-def main(args):
-  global image_shape
-  global net
-
-  print(args)
-  ctx = []
-  cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-  if len(cvd)>0:
-    for i in xrange(len(cvd.split(','))):
-      ctx.append(mx.gpu(i))
-  if len(ctx)==0:
-    ctx = [mx.cpu()]
-    print('use cpu')
-  else:
-    print('gpu num:', len(ctx))
-  image_shape = [int(x) for x in args.image_size.split(',')]
-  vec = args.model.split(',')
-  assert len(vec)>1
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  net = edict()
-  net.ctx = ctx
-  net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(prefix, epoch)
-  #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
-  all_layers = net.sym.get_internals()
-  net.sym = all_layers['fc1_output']
-  net.model = mx.mod.Module(symbol=net.sym, context=net.ctx, label_names = None)
-  net.model.bind(data_shapes=[('data', (args.batch_size, 3, image_shape[1], image_shape[2]))])
-  net.model.set_params(net.arg_params, net.aux_params)
-
-  features_all = None
-
-  i = 0
-  fstart = 0
-  buffer = []
-  for line in open(args.input, 'r'):
-    if i%1000==0:
-      print("processing ",i)
-    i+=1
-    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
-    buffer.append( (image_path, landmark) )
-    if len(buffer)==args.batch_size:
-      embedding = get_feature(buffer)
-      buffer = []
-      fend = fstart+embedding.shape[0]
-      if features_all is None:
-        features_all = np.zeros( (data_size, emb_size), dtype=np.float32 )
-      #print('writing', fstart, fend)
-      features_all[fstart:fend,:] = embedding
-      fstart = fend
-  if len(buffer)>0:
-    embedding = get_feature(buffer)
-    fend = fstart+embedding.shape[0]
-    print('writing', fstart, fend)
-    features_all[fstart:fend,:] = embedding
-  write_bin(args.output, features_all)
-  #os.system("bypy upload %s"%args.output)
-
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--batch_size', type=int, help='', default=32)
-  parser.add_argument('--image_size', type=str, help='', default='3,112,112')
-  parser.add_argument('--input', type=str, help='', default='')
-  parser.add_argument('--output', type=str, help='', default='')
-  parser.add_argument('--model', type=str, help='', default='')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/src/eval/lfw.py b/src/eval/lfw.py
deleted file mode 100644
index e2d624616..000000000
--- a/src/eval/lfw.py
+++ /dev/null
@@ -1,279 +0,0 @@
-"""Helper for evaluation on the Labeled Faces in the Wild dataset 
-"""
-
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import numpy as np
-from scipy import misc
-from sklearn.model_selection import KFold
-from scipy import interpolate
-import sklearn
-from sklearn.decomposition import PCA
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca = 0):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-    
-    tprs = np.zeros((nrof_folds,nrof_thresholds))
-    fprs = np.zeros((nrof_folds,nrof_thresholds))
-    accuracy = np.zeros((nrof_folds))
-    indices = np.arange(nrof_pairs)
-    #print('pca', pca)
-    
-    if pca==0:
-      diff = np.subtract(embeddings1, embeddings2)
-      dist = np.sum(np.square(diff),1)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-        #print('train_set', train_set)
-        #print('test_set', test_set)
-        if pca>0:
-          print('doing pca on', fold_idx)
-          embed1_train = embeddings1[train_set]
-          embed2_train = embeddings2[train_set]
-          _embed_train = np.concatenate( (embed1_train, embed2_train), axis=0 )
-          #print(_embed_train.shape)
-          pca_model = PCA(n_components=pca)
-          pca_model.fit(_embed_train)
-          embed1 = pca_model.transform(embeddings1)
-          embed2 = pca_model.transform(embeddings2)
-          embed1 = sklearn.preprocessing.normalize(embed1)
-          embed2 = sklearn.preprocessing.normalize(embed2)
-          #print(embed1.shape, embed2.shape)
-          diff = np.subtract(embed1, embed2)
-          dist = np.sum(np.square(diff),1)
-        
-        # Find the best threshold for the fold
-        acc_train = np.zeros((nrof_thresholds))
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
-        best_threshold_index = np.argmax(acc_train)
-        for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-          
-    tpr = np.mean(tprs,0)
-    fpr = np.mean(fprs,0)
-    return tpr, fpr, accuracy
-
-def calculate_accuracy(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    tp = np.sum(np.logical_and(predict_issame, actual_issame))
-    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
-    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
-  
-    tpr = 0 if (tp+fn==0) else float(tp) / float(tp+fn)
-    fpr = 0 if (fp+tn==0) else float(fp) / float(fp+tn)
-    acc = float(tp+tn)/dist.size
-    return tpr, fpr, acc
-
-
-  
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-    
-    val = np.zeros(nrof_folds)
-    far = np.zeros(nrof_folds)
-    
-    diff = np.subtract(embeddings1, embeddings2)
-    dist = np.sum(np.square(diff),1)
-    indices = np.arange(nrof_pairs)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
-        # Find the threshold that gives FAR = far_target
-        far_train = np.zeros(nrof_thresholds)
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
-        if np.max(far_train)>=far_target:
-            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
-            threshold = f(far_target)
-        else:
-            threshold = 0.0
-    
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
-  
-    val_mean = np.mean(val)
-    far_mean = np.mean(far)
-    val_std = np.std(val)
-    return val_mean, val_std, far_mean
-
-
-def calculate_val_far(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    n_same = np.sum(actual_issame)
-    n_diff = np.sum(np.logical_not(actual_issame))
-    val = float(true_accept) / float(n_same)
-    far = float(false_accept) / float(n_diff)
-    return val, far
-
-def evaluate(embeddings, actual_issame, nrof_folds=10, pca = 0):
-    # Calculate evaluation metrics
-    thresholds = np.arange(0, 4, 0.01)
-    embeddings1 = embeddings[0::2]
-    embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), nrof_folds=nrof_folds, pca = pca)
-    thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
-    return tpr, fpr, accuracy, val, val_std, far
-
-def get_paths(lfw_dir, pairs, file_ext):
-    nrof_skipped_pairs = 0
-    path_list = []
-    issame_list = []
-    for pair in pairs:
-        if len(pair) == 3:
-            path0 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[1])+'.'+file_ext)
-            path1 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[2])+'.'+file_ext)
-            issame = True
-        elif len(pair) == 4:
-            path0 = os.path.join(lfw_dir, pair[0], pair[0] + '_' + '%04d' % int(pair[1])+'.'+file_ext)
-            path1 = os.path.join(lfw_dir, pair[2], pair[2] + '_' + '%04d' % int(pair[3])+'.'+file_ext)
-            issame = False
-        if os.path.exists(path0) and os.path.exists(path1):    # Only add the pair if both paths exist
-            path_list += (path0,path1)
-            issame_list.append(issame)
-        else:
-            print('not exists', path0, path1)
-            nrof_skipped_pairs += 1
-    if nrof_skipped_pairs>0:
-        print('Skipped %d image pairs' % nrof_skipped_pairs)
-    
-    return path_list, issame_list
-
-def read_pairs(pairs_filename):
-    pairs = []
-    with open(pairs_filename, 'r') as f:
-        for line in f.readlines()[1:]:
-            pair = line.strip().split()
-            pairs.append(pair)
-    return np.array(pairs)
-
-
-def load_dataset(lfw_dir, image_size):
-  lfw_pairs = read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
-  lfw_paths, issame_list = get_paths(lfw_dir, lfw_pairs, 'jpg')
-  lfw_data_list = []
-  for flip in [0,1]:
-    lfw_data = nd.empty((len(lfw_paths), 3, image_size[0], image_size[1]))
-    lfw_data_list.append(lfw_data)
-  i = 0
-  for path in lfw_paths:
-    with open(path, 'rb') as fin:
-      _bin = fin.read()
-      img = mx.image.imdecode(_bin)
-      img = nd.transpose(img, axes=(2, 0, 1))
-      for flip in [0,1]:
-        if flip==1:
-          img = mx.ndarray.flip(data=img, axis=2)
-        lfw_data_list[flip][i][:] = img
-      i+=1
-      if i%1000==0:
-        print('loading lfw', i)
-  print(lfw_data_list[0].shape)
-  print(lfw_data_list[1].shape)
-  return (lfw_data_list, issame_list)
-
-def test(lfw_set, mx_model, batch_size):
-  print('testing lfw..')
-  lfw_data_list = lfw_set[0]
-  issame_list = lfw_set[1]
-  model = mx_model
-  embeddings_list = []
-  for i in xrange( len(lfw_data_list) ):
-    lfw_data = lfw_data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<lfw_data.shape[0]:
-      bb = min(ba+batch_size, lfw_data.shape[0])
-      _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
-      _label = nd.ones( (bb-ba,) )
-      #print(_data.shape, _label.shape)
-      db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      #_arg, _aux = model.get_params()
-      #__arg = {}
-      #for k,v in _arg.iteritems():
-      #  __arg[k] = v.as_in_context(_ctx)
-      #_arg = __arg
-      #_arg["data"] = _data.as_in_context(_ctx)
-      #_arg["softmax_label"] = _label.as_in_context(_ctx)
-      #for k,v in _arg.iteritems():
-      #  print(k,v.context)
-      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
-      #exe.forward(is_train=False)
-      #net_out = exe.outputs
-      _embeddings = net_out[0].asnumpy()
-      #print(_embeddings.shape)
-      if embeddings is None:
-        embeddings = np.zeros( (lfw_data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings
-      ba = bb
-    embeddings_list.append(embeddings)
-
-  _xnorm = 0.0
-  _xnorm_cnt = 0
-  for embed in embeddings_list:
-    for i in xrange(embed.shape[0]):
-      _em = embed[i]
-      _norm=np.linalg.norm(_em)
-      #print(_em.shape, _norm)
-      _xnorm+=_norm
-      _xnorm_cnt+=1
-  _xnorm /= _xnorm_cnt
-
-  embeddings = embeddings_list[0].copy()
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  acc1, std1 = np.mean(accuracy), np.std(accuracy)
-  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-  #embeddings = np.concatenate(embeddings_list, axis=1)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  print(embeddings.shape)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  return acc1, std1, acc2, std2, _xnorm, embeddings_list
-
diff --git a/src/eval/verification.py b/src/eval/verification.py
deleted file mode 100644
index 39f513e60..000000000
--- a/src/eval/verification.py
+++ /dev/null
@@ -1,590 +0,0 @@
-"""Helper for evaluation on the Labeled Faces in the Wild dataset 
-"""
-
-# MIT License
-# 
-# Copyright (c) 2016 David Sandberg
-# 
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-# 
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-# 
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import argparse
-import sys
-import numpy as np
-from scipy import misc
-from sklearn.model_selection import KFold
-from scipy import interpolate
-import sklearn
-import cv2
-import math
-import datetime
-import pickle
-from sklearn.decomposition import PCA
-import mxnet as mx
-from mxnet import ndarray as nd
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-
-
-class LFold:
-  def __init__(self, n_splits = 2, shuffle = False):
-    self.n_splits = n_splits
-    if self.n_splits>1:
-      self.k_fold = KFold(n_splits = n_splits, shuffle = shuffle)
-
-  def split(self, indices):
-    if self.n_splits>1:
-      return self.k_fold.split(indices)
-    else:
-      return [(indices, indices)]
-
-
-def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_folds=10, pca = 0):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    tprs = np.zeros((nrof_folds,nrof_thresholds))
-    fprs = np.zeros((nrof_folds,nrof_thresholds))
-    accuracy = np.zeros((nrof_folds))
-    indices = np.arange(nrof_pairs)
-    #print('pca', pca)
-    
-    if pca==0:
-      diff = np.subtract(embeddings1, embeddings2)
-      dist = np.sum(np.square(diff),1)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-        #print('train_set', train_set)
-        #print('test_set', test_set)
-        if pca>0:
-          print('doing pca on', fold_idx)
-          embed1_train = embeddings1[train_set]
-          embed2_train = embeddings2[train_set]
-          _embed_train = np.concatenate( (embed1_train, embed2_train), axis=0 )
-          #print(_embed_train.shape)
-          pca_model = PCA(n_components=pca)
-          pca_model.fit(_embed_train)
-          embed1 = pca_model.transform(embeddings1)
-          embed2 = pca_model.transform(embeddings2)
-          embed1 = sklearn.preprocessing.normalize(embed1)
-          embed2 = sklearn.preprocessing.normalize(embed2)
-          #print(embed1.shape, embed2.shape)
-          diff = np.subtract(embed1, embed2)
-          dist = np.sum(np.square(diff),1)
-        
-        # Find the best threshold for the fold
-        acc_train = np.zeros((nrof_thresholds))
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
-        best_threshold_index = np.argmax(acc_train)
-        #print('threshold', thresholds[best_threshold_index])
-        for threshold_idx, threshold in enumerate(thresholds):
-            tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-        _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-          
-    tpr = np.mean(tprs,0)
-    fpr = np.mean(fprs,0)
-    return tpr, fpr, accuracy
-
-def calculate_accuracy(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    tp = np.sum(np.logical_and(predict_issame, actual_issame))
-    fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    tn = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))
-    fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
-  
-    tpr = 0 if (tp+fn==0) else float(tp) / float(tp+fn)
-    fpr = 0 if (fp+tn==0) else float(fp) / float(fp+tn)
-    acc = float(tp+tn)/dist.size
-    return tpr, fpr, acc
-
-
-  
-def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_target, nrof_folds=10):
-    assert(embeddings1.shape[0] == embeddings2.shape[0])
-    assert(embeddings1.shape[1] == embeddings2.shape[1])
-    nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-    nrof_thresholds = len(thresholds)
-    k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-    
-    val = np.zeros(nrof_folds)
-    far = np.zeros(nrof_folds)
-    
-    diff = np.subtract(embeddings1, embeddings2)
-    dist = np.sum(np.square(diff),1)
-    indices = np.arange(nrof_pairs)
-    
-    for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
-        # Find the threshold that gives FAR = far_target
-        far_train = np.zeros(nrof_thresholds)
-        for threshold_idx, threshold in enumerate(thresholds):
-            _, far_train[threshold_idx] = calculate_val_far(threshold, dist[train_set], actual_issame[train_set])
-        if np.max(far_train)>=far_target:
-            f = interpolate.interp1d(far_train, thresholds, kind='slinear')
-            threshold = f(far_target)
-        else:
-            threshold = 0.0
-    
-        val[fold_idx], far[fold_idx] = calculate_val_far(threshold, dist[test_set], actual_issame[test_set])
-  
-    val_mean = np.mean(val)
-    far_mean = np.mean(far)
-    val_std = np.std(val)
-    return val_mean, val_std, far_mean
-
-
-def calculate_val_far(threshold, dist, actual_issame):
-    predict_issame = np.less(dist, threshold)
-    true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
-    false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
-    n_same = np.sum(actual_issame)
-    n_diff = np.sum(np.logical_not(actual_issame))
-    #print(true_accept, false_accept)
-    #print(n_same, n_diff)
-    val = float(true_accept) / float(n_same)
-    far = float(false_accept) / float(n_diff)
-    return val, far
-
-def evaluate(embeddings, actual_issame, nrof_folds=10, pca = 0):
-    # Calculate evaluation metrics
-    thresholds = np.arange(0, 4, 0.01)
-    embeddings1 = embeddings[0::2]
-    embeddings2 = embeddings[1::2]
-    tpr, fpr, accuracy = calculate_roc(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), nrof_folds=nrof_folds, pca = pca)
-    thresholds = np.arange(0, 4, 0.001)
-    val, val_std, far = calculate_val(thresholds, embeddings1, embeddings2,
-        np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
-    return tpr, fpr, accuracy, val, val_std, far
-
-def load_bin(path, image_size):
-  bins, issame_list = pickle.load(open(path, 'rb'))
-  data_list = []
-  for flip in [0,1]:
-    data = nd.empty((len(issame_list)*2, 3, image_size[0], image_size[1]))
-    data_list.append(data)
-  for i in xrange(len(issame_list)*2):
-    _bin = bins[i]
-    img = mx.image.imdecode(_bin)
-    if img.shape[1]!=image_size[0]:
-      img = mx.image.resize_short(img, image_size[0])
-    img = nd.transpose(img, axes=(2, 0, 1))
-    for flip in [0,1]:
-      if flip==1:
-        img = mx.ndarray.flip(data=img, axis=2)
-      data_list[flip][i][:] = img
-    if i%1000==0:
-      print('loading bin', i)
-  print(data_list[0].shape)
-  return (data_list, issame_list)
-
-def test(data_set, mx_model, batch_size, nfolds=10, data_extra = None, label_shape = None):
-  print('testing verification..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in xrange( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      #_arg, _aux = model.get_params()
-      #__arg = {}
-      #for k,v in _arg.iteritems():
-      #  __arg[k] = v.as_in_context(_ctx)
-      #_arg = __arg
-      #_arg["data"] = _data.as_in_context(_ctx)
-      #_arg["softmax_label"] = _label.as_in_context(_ctx)
-      #for k,v in _arg.iteritems():
-      #  print(k,v.context)
-      #exe = sym.bind(_ctx, _arg ,args_grad=None, grad_req="null", aux_states=_aux)
-      #exe.forward(is_train=False)
-      #net_out = exe.outputs
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      #print(_embeddings.shape)
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-
-  _xnorm = 0.0
-  _xnorm_cnt = 0
-  for embed in embeddings_list:
-    for i in xrange(embed.shape[0]):
-      _em = embed[i]
-      _norm=np.linalg.norm(_em)
-      #print(_em.shape, _norm)
-      _xnorm+=_norm
-      _xnorm_cnt+=1
-  _xnorm /= _xnorm_cnt
-
-  embeddings = embeddings_list[0].copy()
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  acc1 = 0.0
-  std1 = 0.0
-  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  #acc1, std1 = np.mean(accuracy), np.std(accuracy)
-
-  #print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-  #embeddings = np.concatenate(embeddings_list, axis=1)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  print(embeddings.shape)
-  print('infer time', time_consumed)
-  _, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=nfolds)
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  return acc1, std1, acc2, std2, _xnorm, embeddings_list
-
-def test_badcase(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
-  print('testing verification badcase..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in xrange( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  thresholds = np.arange(0, 4, 0.01)
-  actual_issame = np.asarray(issame_list)
-  nrof_folds = 10
-  embeddings1 = embeddings[0::2]
-  embeddings2 = embeddings[1::2]
-  assert(embeddings1.shape[0] == embeddings2.shape[0])
-  assert(embeddings1.shape[1] == embeddings2.shape[1])
-  nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-  nrof_thresholds = len(thresholds)
-  k_fold = LFold(n_splits=nrof_folds, shuffle=False)
-  
-  tprs = np.zeros((nrof_folds,nrof_thresholds))
-  fprs = np.zeros((nrof_folds,nrof_thresholds))
-  accuracy = np.zeros((nrof_folds))
-  indices = np.arange(nrof_pairs)
-  
-  diff = np.subtract(embeddings1, embeddings2)
-  dist = np.sum(np.square(diff),1)
-  data = data_list[0]
-
-  pouts = []
-  nouts = []
-  
-  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-       
-      # Find the best threshold for the fold
-      acc_train = np.zeros((nrof_thresholds))
-      #print(train_set)
-      #print(train_set.__class__)
-      for threshold_idx, threshold in enumerate(thresholds):
-          p2 = dist[train_set]
-          p3 = actual_issame[train_set]
-          _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
-      best_threshold_index = np.argmax(acc_train)
-      for threshold_idx, threshold in enumerate(thresholds):
-          tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-      _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-      best_threshold = thresholds[best_threshold_index]
-      for iid in test_set:
-        ida = iid*2
-        idb = ida+1
-        asame = actual_issame[iid]
-        _dist = dist[iid]
-        violate = _dist - best_threshold
-        if not asame:
-          violate *= -1.0
-        if violate>0.0:
-          imga = data[ida].asnumpy().transpose( (1,2,0) )[...,::-1] #to bgr
-          imgb = data[idb].asnumpy().transpose( (1,2,0) )[...,::-1]
-          #print(imga.shape, imgb.shape, violate, asame, _dist)
-          if asame:
-            pouts.append( (imga, imgb, _dist, best_threshold, ida) )
-          else:
-            nouts.append( (imga, imgb, _dist, best_threshold, ida) )
-
-        
-  tpr = np.mean(tprs,0)
-  fpr = np.mean(fprs,0)
-  acc = np.mean(accuracy)
-  pouts = sorted(pouts, key = lambda x: x[2], reverse=True)
-  nouts = sorted(nouts, key = lambda x: x[2], reverse=False)
-  print(len(pouts), len(nouts))
-  print('acc', acc)
-  gap = 10
-  image_shape = (112,224,3)
-  out_dir = "./badcases"
-  if not os.path.exists(out_dir):
-    os.makedirs(out_dir)
-  if len(nouts)>0:
-    threshold = nouts[0][3]
-  else:
-    threshold = pouts[-1][3]
-  
-  for item in [(pouts, 'positive(false_negative).png'), (nouts, 'negative(false_positive).png')]:
-    cols = 4
-    rows = 8000
-    outs = item[0]
-    if len(outs)==0:
-      continue
-    #if len(outs)==9:
-    #  cols = 3
-    #  rows = 3
-
-    _rows = int(math.ceil(len(outs)/cols))
-    rows = min(rows, _rows)
-    hack = {}
-
-    if name.startswith('cfp') and item[1].startswith('pos'):
-      hack = {0:'manual/238_13.jpg.jpg', 6:'manual/088_14.jpg.jpg', 10:'manual/470_14.jpg.jpg', 25:'manual/238_13.jpg.jpg', 28:'manual/143_11.jpg.jpg'}
-
-    filename = item[1]
-    if len(name)>0:
-      filename = name+"_"+filename
-    filename = os.path.join(out_dir, filename)
-    img = np.zeros( (image_shape[0]*rows+20, image_shape[1]*cols+(cols-1)*gap, 3), dtype=np.uint8 )
-    img[:,:,:] = 255
-    text_color = (0,0,153)
-    text_color = (255,178,102)
-    text_color = (153,255,51)
-    for outi, out in enumerate(outs):
-      row = outi//cols
-      col = outi%cols
-      if row==rows:
-        break
-      imga = out[0].copy()
-      imgb = out[1].copy()
-      if outi in hack:
-        idx = out[4]
-        print('noise idx',idx)
-        aa = hack[outi]
-        imgb = cv2.imread(aa)
-        #if aa==1:
-        #  imgb = cv2.transpose(imgb)
-        #  imgb = cv2.flip(imgb, 1)
-        #elif aa==3:
-        #  imgb = cv2.transpose(imgb)
-        #  imgb = cv2.flip(imgb, 0)
-        #else:
-        #  for ii in xrange(2):
-        #    imgb = cv2.transpose(imgb)
-        #    imgb = cv2.flip(imgb, 1)
-      dist = out[2]
-      _img = np.concatenate( (imga, imgb), axis=1 )
-      k = "%.3f"%dist
-      #print(k)
-      font = cv2.FONT_HERSHEY_SIMPLEX
-      cv2.putText(_img,k,(80,image_shape[0]//2+7), font, 0.6, text_color, 2)
-      #_filename = filename+"_%d.png"%outi
-      #cv2.imwrite(_filename, _img)
-      img[row*image_shape[0]:(row+1)*image_shape[0], (col*image_shape[1]+gap*col):((col+1)*image_shape[1]+gap*col),:] = _img
-    #threshold = outs[0][3]
-    font = cv2.FONT_HERSHEY_SIMPLEX
-    k = "threshold: %.3f"%threshold
-    cv2.putText(img,k,(img.shape[1]//2-70,img.shape[0]-5), font, 0.6, text_color, 2)
-    cv2.imwrite(filename, img)
-
-def dumpR(data_set, mx_model, batch_size, name='', data_extra = None, label_shape = None):
-  print('dump verification embedding..')
-  data_list = data_set[0]
-  issame_list = data_set[1]
-  model = mx_model
-  embeddings_list = []
-  if data_extra is not None:
-    _data_extra = nd.array(data_extra)
-  time_consumed = 0.0
-  if label_shape is None:
-    _label = nd.ones( (batch_size,) )
-  else:
-    _label = nd.ones( label_shape )
-  for i in xrange( len(data_list) ):
-    data = data_list[i]
-    embeddings = None
-    ba = 0
-    while ba<data.shape[0]:
-      bb = min(ba+batch_size, data.shape[0])
-      count = bb-ba
-      _data = nd.slice_axis(data, axis=0, begin=bb-batch_size, end=bb)
-      #print(_data.shape, _label.shape)
-      time0 = datetime.datetime.now()
-      if data_extra is None:
-        db = mx.io.DataBatch(data=(_data,), label=(_label,))
-      else:
-        db = mx.io.DataBatch(data=(_data,_data_extra), label=(_label,))
-      model.forward(db, is_train=False)
-      net_out = model.get_outputs()
-      _embeddings = net_out[0].asnumpy()
-      time_now = datetime.datetime.now()
-      diff = time_now - time0
-      time_consumed+=diff.total_seconds()
-      if embeddings is None:
-        embeddings = np.zeros( (data.shape[0], _embeddings.shape[1]) )
-      embeddings[ba:bb,:] = _embeddings[(batch_size-count):,:]
-      ba = bb
-    embeddings_list.append(embeddings)
-  embeddings = embeddings_list[0] + embeddings_list[1]
-  embeddings = sklearn.preprocessing.normalize(embeddings)
-  actual_issame = np.asarray(issame_list)
-  outname = os.path.join('temp.bin')
-  with open(outname, 'wb') as f:
-    pickle.dump((embeddings, issame_list), f, protocol=pickle.HIGHEST_PROTOCOL)
-
-if __name__ == '__main__':
-
-  parser = argparse.ArgumentParser(description='do verification')
-  # general
-  parser.add_argument('--data-dir', default='', help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--target', default='lfw,cfp_ff,cfp_fp,agedb_30', help='test targets.')
-  parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--max', default='', type=str, help='')
-  parser.add_argument('--mode', default=0, type=int, help='')
-  parser.add_argument('--nfolds', default=10, type=int, help='')
-  args = parser.parse_args()
-
-  prop = face_image.load_property(args.data_dir)
-  image_size = prop.image_size
-  print('image_size', image_size)
-  ctx = mx.gpu(args.gpu)
-  nets = []
-  vec = args.model.split(',')
-  prefix = args.model.split(',')[0]
-  epochs = []
-  if len(vec)==1:
-    pdir = os.path.dirname(prefix)
-    for fname in os.listdir(pdir):
-      if not fname.endswith('.params'):
-        continue
-      _file = os.path.join(pdir, fname)
-      if _file.startswith(prefix):
-        epoch = int(fname.split('.')[0].split('-')[1])
-        epochs.append(epoch)
-    epochs = sorted(epochs, reverse=True)
-    if len(args.max)>0:
-      _max = [int(x) for x in args.max.split(',')]
-      assert len(_max)==2
-      if len(epochs)>_max[1]:
-        epochs = epochs[_max[0]:_max[1]]
-
-  else:
-    epochs = [int(x) for x in vec[1].split('|')]
-  print('model number', len(epochs))
-  time0 = datetime.datetime.now()
-  for epoch in epochs:
-    print('loading',prefix, epoch)
-    sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-    #arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
-    all_layers = sym.get_internals()
-    sym = all_layers['fc1_output']
-    model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-    #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-    model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-    model.set_params(arg_params, aux_params)
-    nets.append(model)
-  time_now = datetime.datetime.now()
-  diff = time_now - time0
-  print('model loading time', diff.total_seconds())
-
-  ver_list = []
-  ver_name_list = []
-  for name in args.target.split(','):
-    path = os.path.join(args.data_dir,name+".bin")
-    if os.path.exists(path):
-      print('loading.. ', name)
-      data_set = load_bin(path, image_size)
-      ver_list.append(data_set)
-      ver_name_list.append(name)
-
-  if args.mode==0:
-    for i in xrange(len(ver_list)):
-      results = []
-      for model in nets:
-        acc1, std1, acc2, std2, xnorm, embeddings_list = test(ver_list[i], model, args.batch_size, args.nfolds)
-        print('[%s]XNorm: %f' % (ver_name_list[i], xnorm))
-        print('[%s]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], acc1, std1))
-        print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], acc2, std2))
-        results.append(acc2)
-      print('Max of [%s] is %1.5f' % (ver_name_list[i], np.max(results)))
-  elif args.mode==1:
-    model = nets[0]
-    test_badcase(ver_list[0], model, args.batch_size, args.target)
-  else:
-    model = nets[0]
-    dumpR(ver_list[0], model, args.batch_size, args.target)
-
-
diff --git a/src/eval/ytf.py b/src/eval/ytf.py
deleted file mode 100644
index 793dab350..000000000
--- a/src/eval/ytf.py
+++ /dev/null
@@ -1,302 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import cv2
-import argparse
-import sys
-import numpy as np
-from scipy import misc
-from sklearn.model_selection import KFold
-from scipy import interpolate
-import sklearn
-from sklearn.metrics.pairwise import euclidean_distances
-import datetime
-import pickle
-from sklearn.decomposition import PCA
-import mxnet as mx
-from mxnet import ndarray as nd
-from verification import evaluate
-from verification import calculate_accuracy
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-
-
-model = None
-feature_cache = {}
-image_size = [112,112]
-
-def get_feature(name, vid, args):
-  global feature_cache
-  key = (name,vid)
-  if key in feature_cache:
-    return feature_cache[key]
-
-  input_dir = os.path.join(args.image_dir, name, str(vid))
-  data = nd.zeros( (1 ,3, image_size[0], image_size[1]) )
-  F = []
-  for img in os.listdir(input_dir):
-    img = os.path.join(input_dir, img)
-    img = cv2.imread(img)
-    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-    img = np.transpose(img, (2,0,1))
-    data[0][:] = img
-    db = mx.io.DataBatch(data=(data,))
-    model.forward(db, is_train=False)
-    net_out = model.get_outputs()[0].asnumpy().flatten()
-    F.append(net_out)
-  F = np.array(F)
-  F = sklearn.preprocessing.normalize(F)
-  feature = np.mean(F, axis=0, keepdims=True)
-  feature = sklearn.preprocessing.normalize(feature).flatten()
-
-  feature_cache[key] = feature
-  return feature
-
-def get_feature_set(name, vid, args):
-  global feature_cache
-  key = (name,vid)
-  if key in feature_cache:
-    return feature_cache[key]
-
-  input_dir = os.path.join(args.image_dir, name, str(vid))
-  data = nd.zeros( (1 ,3, image_size[0], image_size[1]) )
-  F = []
-  for img in os.listdir(input_dir):
-    img = os.path.join(input_dir, img)
-    img = cv2.imread(img)
-    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-    img = np.transpose(img, (2,0,1))
-    data[0][:] = img
-    db = mx.io.DataBatch(data=(data,))
-    model.forward(db, is_train=False)
-    net_out = model.get_outputs()[0].asnumpy().flatten()
-    F.append(net_out)
-  F = np.array(F)
-  F = sklearn.preprocessing.normalize(F)
-
-  feature_cache[key] = F
-  return F
-
-def main(args):
-  global model
-  ctx = mx.gpu(args.gpu)
-  args.ctx_num = 1
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  args.image_dir = os.path.join(args.data_dir, 'images')
-  pairs_file = os.path.join(args.data_dir, 'splits2.txt')
-  embeddings = []
-  issame_list = []
-  data = []
-  pp = 0
-  for line in open(pairs_file, 'r'):
-    line = line.strip()
-    if line.startswith('split'):
-      continue
-    pp+=1
-    if pp%10==0:
-      print('processing', pp)
-    vec = line.split(',')
-    assert len(vec)>=5
-    issame_list.append(int(vec[-1]))
-    for i in [2,3]:
-      _str = vec[i].strip()
-      _vec = _str.split('/')
-      assert len(_vec)==2
-      name = _vec[0]
-      vid = int(_vec[1])
-      feature = get_feature(name, vid, args)
-      print('feature', feature.shape)
-      embeddings.append(feature)
-      data.append( (name, vid) )
-    #if len(issame_list)==20:
-    #  break
-  embeddings = np.array(embeddings)
-  print(embeddings.shape)
-  thresholds = np.arange(0, 4, 0.01)
-  actual_issame = np.asarray(issame_list)
-  nrof_folds = 10
-  embeddings1 = embeddings[0::2]
-  embeddings2 = embeddings[1::2]
-  assert(embeddings1.shape[0] == embeddings2.shape[0])
-  assert(embeddings1.shape[1] == embeddings2.shape[1])
-  nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
-  nrof_thresholds = len(thresholds)
-  k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-  
-  tprs = np.zeros((nrof_folds,nrof_thresholds))
-  fprs = np.zeros((nrof_folds,nrof_thresholds))
-  accuracy = np.zeros((nrof_folds))
-  indices = np.arange(nrof_pairs)
-  
-  diff = np.subtract(embeddings1, embeddings2)
-  dist = np.sum(np.square(diff),1)
-  pouts = []
-  nouts = []
-  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      # Find the best threshold for the fold
-      acc_train = np.zeros((nrof_thresholds))
-      #print(train_set)
-      #print(train_set.__class__)
-      for threshold_idx, threshold in enumerate(thresholds):
-          p2 = dist[train_set]
-          p3 = actual_issame[train_set]
-          _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
-      best_threshold_index = np.argmax(acc_train)
-      for threshold_idx, threshold in enumerate(thresholds):
-          tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-      _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-      best_threshold = thresholds[best_threshold_index]
-      for iid in test_set:
-        ida = iid*2
-        idb = ida+1
-        asame = actual_issame[iid]
-        _dist = dist[iid]
-        violate = _dist - best_threshold
-        if not asame:
-          violate *= -1.0
-        if violate>0.0:
-          dataa = data[ida]
-          datab = data[idb]
-          #print(imga.shape, imgb.shape, violate, asame, _dist)
-          if asame:
-            pouts.append( (dataa, datab, _dist, best_threshold, ida) )
-          else:
-            nouts.append( (dataa, datab, _dist, best_threshold, ida) )
-
-        
-  tpr = np.mean(tprs,0)
-  fpr = np.mean(fprs,0)
-  acc = np.mean(accuracy)
-  pouts = sorted(pouts, key = lambda x: x[2], reverse=True)
-  nouts = sorted(nouts, key = lambda x: x[2], reverse=False)
-  print(len(pouts), len(nouts))
-  print('acc', acc)
-  if len(nouts)>0:
-    threshold = nouts[0][3]
-  else:
-    threshold = pouts[-1][3]
-  #print('threshold', threshold)
-  print('positive(false negative):')
-  for out in pouts:
-    print("\t%s\t%s\t(distance:%f, threshold:%f)"%(out[0], out[1], out[2], out[3]))
-  print('negative(false positive):')
-  for out in nouts:
-    print("\t%s\t%s\t(distance:%f, threshold:%f)"%(out[0], out[1], out[2], out[3]))
-
-
-
-
-
-  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
-  #acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  #print('acc', acc2)
-
-def main2(args):
-  global model
-  ctx = mx.gpu(args.gpu)
-  args.ctx_num = 1
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  args.image_dir = os.path.join(args.data_dir, 'images')
-  pairs_file = os.path.join(args.data_dir, 'splits2.txt')
-  issame_list = []
-  dist = []
-  pp = 0
-  for line in open(pairs_file, 'r'):
-    line = line.strip()
-    if line.startswith('split'):
-      continue
-    pp+=1
-    if pp%10==0:
-      print('processing', pp)
-    vec = line.split(',')
-    assert len(vec)>=5
-    issame_list.append(int(vec[-1]))
-    feature_sets = []
-    for i in [2,3]:
-      _str = vec[i].strip()
-      _vec = _str.split('/')
-      assert len(_vec)==2
-      name = _vec[0]
-      vid = int(_vec[1])
-      feature = get_feature_set(name, vid, args)
-      print('feature', len(feature))
-      feature_sets.append(feature)
-    X = feature_sets[0]
-    Y = feature_sets[1]
-    _dist = euclidean_distances(X, Y)
-    _dist = _dist*_dist
-    #_tmp = np.eye(_dist.shape[0], dtype=np.float32)
-    #_dist += _tmp
-    if args.mode==2:
-      _dist = np.amin(_dist)
-    elif args.mode==3:
-      _dist = np.mean(_dist)
-    else:
-      _dist = np.amax(_dist)
-    print(_dist)
-    dist.append(_dist)
-    #if len(dist)==10:
-    #  break
-
-  dist = np.array(dist)
-  nrof_folds = 10
-  thresholds = np.arange(0, 4, 0.01)
-  actual_issame = np.array(issame_list)
-  nrof_pairs = len(actual_issame)
-  nrof_thresholds = len(thresholds)
-  k_fold = KFold(n_splits=nrof_folds, shuffle=False)
-  
-  tprs = np.zeros((nrof_folds,nrof_thresholds))
-  fprs = np.zeros((nrof_folds,nrof_thresholds))
-  accuracy = np.zeros((nrof_folds))
-  indices = np.arange(nrof_pairs)
-  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
-      
-    # Find the best threshold for the fold
-    acc_train = np.zeros((nrof_thresholds))
-    for threshold_idx, threshold in enumerate(thresholds):
-        _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
-    best_threshold_index = np.argmax(acc_train)
-    for threshold_idx, threshold in enumerate(thresholds):
-        tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
-    _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
-  acc2, std2 = np.mean(accuracy), np.std(accuracy)
-  print('acc', acc2)
-
-if __name__ == '__main__':
-
-  parser = argparse.ArgumentParser(description='do verification')
-  # general
-  parser.add_argument('--data-dir', default='/raid5data/dplearn/YTF', help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--gpu', default=0, type=int, help='gpu id')
-  parser.add_argument('--batch-size', default=32, type=int, help='')
-  parser.add_argument('--mode', default=1, type=int, help='')
-  args = parser.parse_args()
-  if args.mode>=2:
-    main2(args)
-  else:
-    main(args)
-
diff --git a/src/eval/ytf_badcases.py b/src/eval/ytf_badcases.py
deleted file mode 100644
index a2525c314..000000000
--- a/src/eval/ytf_badcases.py
+++ /dev/null
@@ -1,75 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import sys
-import os
-import numpy as np
-import cv2
-
-pairs_file = '/raid5data/dplearn/YTF/splits2.txt'
-stat = [0,0]
-for line in open(pairs_file, 'r'):
-  line = line.strip()
-  if line.startswith('split'):
-    continue
-  vec = line.split(',')
-  issame = int(vec[-1])
-  if issame:
-    stat[0]+=1
-  else:
-    stat[1]+=1
-print('stat', stat)
-
-image_dir = '/raid5data/dplearn/YTF/images'
-
-def get_img(name, vid):
-  input_dir = os.path.join(image_dir, name, str(vid))
-  paths = []
-  for img in os.listdir(input_dir):
-    path = os.path.join(input_dir, img)
-    paths.append(path)
-  paths = sorted(paths)
-  parts = 8
-  assert len(paths)>=parts
-  gap = len(paths)//parts
-  img = None
-  for i in xrange(parts):
-    idx = gap*i
-    path = paths[idx]
-    _img = cv2.imread(path)
-    #print(_img.shape)
-    if img is None:
-      img = _img
-    else:
-      img = np.concatenate( (img, _img), axis=1)
-  return img
-
-
-text_color = (153,255,51)
-for input in ['ytf_false_positive', 'ytf_false_negative']:
-  all_img = None
-  pp = 0
-  for line in open(input+".log", 'r'):
-    if line.startswith("\t"):
-      break
-    vec = line.strip().split(',')
-    img1 = get_img(vec[0], int(vec[1]))
-    img2 = get_img(vec[2], int(vec[3]))
-    img = np.concatenate( (img1, img2), axis=0)
-    if all_img is None:
-      all_img = img
-    else:
-      all_img = np.concatenate( (all_img, img), axis=0)
-    blank_img = np.zeros( (20, 112*8,3), dtype=np.uint8)
-    blank_img[:,:,:] = 255
-    font = cv2.FONT_HERSHEY_SIMPLEX
-    k = "centre-distance:%.3f"%(float(vec[4]))
-    #print(k)
-    cv2.putText(blank_img,k,(350,blank_img.shape[0]-4), font, 0.6, text_color, 2)
-    all_img = np.concatenate( (all_img, blank_img), axis=0)
-    pp+=1
-
-  filename = os.path.join('badcases', input+".png")
-  cv2.imwrite(filename, all_img)
-
diff --git a/src/image_iter.py b/src/image_iter.py
deleted file mode 100644
index a03cf1409..000000000
--- a/src/image_iter.py
+++ /dev/null
@@ -1,316 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-from PIL import Image
-from io import BytesIO
-
-import mxnet as mx
-from mxnet import ndarray as nd
-from mxnet import io
-from mxnet import recordio
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_preprocess
-import multiprocessing
-
-logger = logging.getLogger()
-
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None, mean = None,
-                 rand_mirror = False, cutoff = 0, color_jittering = 0,
-                 images_filter = 0,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        if path_imgrec:
-            logging.info('loading recordio %s...',
-                         path_imgrec)
-            path_imgidx = path_imgrec[0:-4]+".idx"
-            self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-            s = self.imgrec.read_idx(0)
-            header, _ = recordio.unpack(s)
-            if header.flag>0:
-              print('header0 label', header.label)
-              self.header0 = (int(header.label[0]), int(header.label[1]))
-              #assert(header.flag==1)
-              #self.imgidx = range(1, int(header.label[0]))
-              self.imgidx = []
-              self.id2range = {}
-              self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-              for identity in self.seq_identity:
-                s = self.imgrec.read_idx(identity)
-                header, _ = recordio.unpack(s)
-                a,b = int(header.label[0]), int(header.label[1])
-                count = b-a
-                if count<images_filter:
-                  continue
-                self.id2range[identity] = (a,b)
-                self.imgidx += range(a, b)
-              print('id2range', len(self.id2range))
-            else:
-              self.imgidx = list(self.imgrec.keys)
-            if shuffle:
-              self.seq = self.imgidx
-              self.oseq = self.imgidx
-              print(len(self.seq))
-            else:
-              self.seq = None
-
-        self.mean = mean
-        self.nd_mean = None
-        if self.mean:
-          self.mean = np.array(self.mean, dtype=np.float32).reshape(1,1,3)
-          self.nd_mean = mx.nd.array(self.mean).reshape((1,1,3))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        self.color_jittering = color_jittering
-        self.CJA = mx.image.ColorJitterAug(0.125, 0.125, 0.125)
-        self.provide_label = [(label_name, (batch_size,))]
-        #print(self.provide_label[0][1])
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.shuffle:
-          random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-        """Helper function for reading in next sample."""
-        #set total batch size, for example, 1800, and maximum size for each people, for example 45
-        if self.seq is not None:
-          while True:
-            if self.cur >= len(self.seq):
-                raise StopIteration
-            idx = self.seq[self.cur]
-            self.cur += 1
-            if self.imgrec is not None:
-              s = self.imgrec.read_idx(idx)
-              header, img = recordio.unpack(s)
-              label = header.label
-              if not isinstance(label, numbers.Number):
-                label = label[0]
-              return label, img, None, None
-            else:
-              label, fname, bbox, landmark = self.imglist[idx]
-              return label, self.read_image(fname), bbox, landmark
-        else:
-            s = self.imgrec.read()
-            if s is None:
-                raise StopIteration
-            header, img = recordio.unpack(s)
-            return header.label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = nd.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = nd.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      #augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      #random.shuffle(augs)
-      #for aug in augs:
-      #  #print(img.shape)
-      #  img = aug(img, x)
-      #  #print(img.shape)
-      #return img
-      return self.CJA(img)
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in xrange(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-    def compress_aug(self, img):
-      buf = BytesIO()
-      img = Image.fromarray(img.asnumpy(), 'RGB')
-      q = random.randint(2, 20)
-      img.save(buf, format='JPEG', quality=q)
-      buf = buf.getvalue()
-      img = Image.open(BytesIO(buf))
-      return nd.array(np.asarray(img, 'float32'))
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                _data = self.imdecode(s)
-                if _data.shape[0]!=self.data_shape[1]:
-                  _data = mx.image.resize_short(_data, self.data_shape[1])
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.color_jittering>0:
-                  if self.color_jittering>1:
-                    _rd = random.randint(0,1)
-                    if _rd==1:
-                      _data = self.compress_aug(_data)
-                  #print('do color aug')
-                  _data = _data.astype('float32', copy=False)
-                  #print(_data.__class__)
-                  _data = self.color_aug(_data, 0.125)
-                if self.nd_mean is not None:
-                  _data = _data.astype('float32', copy=False)
-                  _data -= self.nd_mean
-                  _data *= 0.0078125
-                if self.cutoff>0:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    #print('do cutoff aug', self.cutoff)
-                    centerh = random.randint(0, _data.shape[0]-1)
-                    centerw = random.randint(0, _data.shape[1]-1)
-                    half = self.cutoff//2
-                    starth = max(0, centerh-half)
-                    endh = min(_data.shape[0], centerh+half)
-                    startw = max(0, centerw-half)
-                    endw = min(_data.shape[1], centerw+half)
-                    #print(starth, endh, startw, endw, _data.shape)
-                    _data[starth:endh, startw:endw, :] = 128
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        return io.DataBatch([batch_data], [batch_label], batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/src/losses/center_loss.py b/src/losses/center_loss.py
deleted file mode 100644
index b14bda392..000000000
--- a/src/losses/center_loss.py
+++ /dev/null
@@ -1,123 +0,0 @@
-import os
-
-# MXNET_CPU_WORKER_NTHREADS must be greater than 1 for custom op to work on CPU
-#os.environ['MXNET_CPU_WORKER_NTHREADS'] = '2'
-import mxnet as mx
-
-
-# define metric of accuracy
-class Accuracy(mx.metric.EvalMetric):
-    def __init__(self, num=None):
-        super(Accuracy, self).__init__('accuracy', num)
-
-    def update(self, labels, preds):
-        mx.metric.check_label_shapes(labels, preds)
-
-        if self.num is not None:
-            assert len(labels) == self.num
-
-        pred_label = mx.nd.argmax_channel(preds[0]).asnumpy().astype('int32')
-        label = labels[0].asnumpy().astype('int32')
-
-        mx.metric.check_label_shapes(label, pred_label)
-
-        self.sum_metric += (pred_label.flat == label.flat).sum()
-        self.num_inst += len(pred_label.flat)
-
-
-# define some metric of center_loss
-class CenterLossMetric(mx.metric.EvalMetric):
-    def __init__(self):
-        super(CenterLossMetric, self).__init__('center_loss')
-
-    def update(self, labels, preds):
-        self.sum_metric += preds[1].asnumpy()[0]
-        self.num_inst += 1
-
-
-# see details:
-# <A Discriminative Feature Learning Approach for Deep Face Recogfnition>
-class CenterLoss(mx.operator.CustomOp):
-    def __init__(self, ctx, shapes, dtypes, num_class, alpha, scale=1.0):
-        if not len(shapes[0]) == 2:
-            raise ValueError('dim for input_data shoudl be 2 for CenterLoss')
-
-        self.alpha = alpha
-        self.batch_size = shapes[0][0]
-        self.num_class = num_class
-        self.scale = scale
-
-    def forward(self, is_train, req, in_data, out_data, aux):
-        labels = in_data[1].asnumpy()
-        diff = aux[0]
-        center = aux[1]
-
-        # store x_i - c_yi
-        for i in range(self.batch_size):
-            diff[i] = in_data[0][i] - center[int(labels[i])]
-
-        loss = mx.nd.sum(mx.nd.square(diff)) / self.batch_size / 2
-        self.assign(out_data[0], req[0], loss)
-
-    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
-        diff = aux[0]
-        center = aux[1]
-        sum_ = aux[2]
-
-        # back grad is just scale * ( x_i - c_yi)
-        grad_scale = float(self.scale/self.batch_size)
-        self.assign(in_grad[0], req[0], diff * grad_scale)
-
-        # update the center
-        labels = in_data[1].asnumpy()
-        label_occur = dict()
-        for i, label in enumerate(labels):
-            label_occur.setdefault(int(label), []).append(i)
-
-        for label, sample_index in label_occur.items():
-            sum_[:] = 0
-            for i in sample_index:
-                sum_ = sum_ + diff[i]
-            delta_c = sum_ / (1 + len(sample_index))
-            center[label] += self.alpha * delta_c
-
-
-@mx.operator.register("centerloss")
-class CenterLossProp(mx.operator.CustomOpProp):
-    def __init__(self, num_class, alpha, scale=1.0, batchsize=64):
-        super(CenterLossProp, self).__init__(need_top_grad=False)
-
-        # convert it to numbers
-        self.num_class = int(num_class)
-        self.alpha = float(alpha)
-        self.scale = float(scale)
-        self.batchsize = int(batchsize)
-
-    def list_arguments(self):
-        return ['data', 'label']
-
-    def list_outputs(self):
-        return ['output']
-
-    def list_auxiliary_states(self):
-        # call them 'bias' for zero initialization
-        return ['diff_bias', 'center_bias', 'sum_bias']
-
-    def infer_shape(self, in_shape):
-        data_shape = in_shape[0]
-        label_shape = (in_shape[0][0],)
-
-        # store diff , same shape as input batch
-        diff_shape = [self.batchsize, data_shape[1]]
-
-        # store the center of each class , should be ( num_class, d )
-        center_shape = [self.num_class, diff_shape[1]]
-
-        # computation buf
-        sum_shape = [diff_shape[1],]
-
-        output_shape = [1, ]
-        return [data_shape, label_shape], [output_shape], [diff_shape, center_shape, sum_shape]
-
-    def create_operator(self, ctx, shapes, dtypes):
-        return CenterLoss(ctx, shapes, dtypes, self.num_class, self.alpha, self.scale)
diff --git a/src/megaface/README.md b/src/megaface/README.md
deleted file mode 100644
index 6f97dfa9b..000000000
--- a/src/megaface/README.md
+++ /dev/null
@@ -1,8 +0,0 @@
-[2018.12.26] Now you can take a look at new megaface testing tool at ``https://github.com/deepinsight/insightface/tree/master/Evaluation/Megaface``. It is more easy to use.
-
-Please strictly follow these rules if you want to use our MegaFace noises list.
-
-* Please cite our paper and git repo if you want to use this list in your paper.
-* Please include the information like `We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface` if you want to submit the result to MegaFace challenge.
-* To be fair, if you want to submit MegaFace result, please ensure there's no training set overlaps with FaceScrub identities. You can do this by removing identities from your training set whose cosine similarity is larger than 0.4 with any FaceScrub identity by comparing their centre feature vectors. 
-* If you find more overlaps noise, please open an issue at InsightFace.
diff --git a/src/megaface/facescrub_noises.txt b/src/megaface/facescrub_noises.txt
deleted file mode 100644
index da28c103b..000000000
--- a/src/megaface/facescrub_noises.txt
+++ /dev/null
@@ -1,606 +0,0 @@
-#Please strictly follow the rules in https://github.com/deepinsight/insightface/tree/master/src/megaface/README.md if you want to use this list.
-Aaron_Eckhart_136.png
-Aaron_Eckhart_221.png
-Adam_McKay_468.png
-Adam_McKay_478.png
-Adam_Sandler_558.png
-Adrienne_Barbeau_4259.png
-Adrienne_Frantz_27683.png
-Adrienne_Frantz_27693.png
-Adrienne_Frantz_27699.png
-Adrienne_Frantz_27701.png
-Adrienne_Frantz_27829.png
-Aisha_Hinds_35768.png
-Alec_Baldwin_2121.png
-Alice_Krige_40717.png
-Alley_Mills_52029.png
-Alley_Mills_52039.png
-Alley_Mills_52044.png
-Allison_Janney_37357.png
-Allison_Janney_37447.png
-Alyson_Hannigan_32288.png
-Alyssa_Milano_51641.png
-America_Ferrera_25859.png
-America_Ferrera_25870.png
-Amy_Davidson_18553.png
-Andrea_Bogart_8427.png
-Andrea_Bowen_8992.png
-Andrea_Bowen_9090.png
-Andy_Richter_3141.png
-Angell_Conwell_15508.png
-Anne_Hathaway_33648.png
-Anne_Hathaway_33685.png
-Annie_Ilonzeh_37123.png
-Anthony_Hopkins_3562.png
-Anthony_Hopkins_3595.png
-Anthony_Hopkins_3694.png
-Antonio_Banderas_4016.png
-Arnold_Vosloo_4287.png
-Ashley_Benson_4144.png
-Ashley_Johnson_37631.png
-Ashley_Johnson_37649.png
-Audra_McDonald_49657.png
-Audrey_Landers_42201.png
-Audrey_Landers_42246.png
-Barbara_Carrera_11961.png
-Ben_Affleck_4700.png
-Ben_Affleck_4706.png
-Ben_Affleck_4762.png
-Ben_Affleck_4885.png
-Ben_Kingsley_5146.png
-Ben_McKenzie_5219.png
-Ben_Stiller_5419.png
-Ben_Stiller_5432.png
-Bernard_Hill_5869.png
-Bernard_Hill_5969.png
-Bernie_Mac_6083.png
-Billy_Bob_Thornton_7010.png
-Billy_Bob_Thornton_7016.png
-Billy_Bob_Thornton_7028.png
-Billy_Bob_Thornton_7047.png
-Billy_Bob_Thornton_7069.png
-Billy_Bob_Thornton_7093.png
-Billy_Bob_Thornton_7106.png
-Billy_Bob_Thornton_7140.png
-Billy_Bob_Thornton_7161.png
-Billy_Boyd_7281.png
-Billy_Boyd_7334.png
-Billy_Burke_7418.png
-Billy_Burke_7556.png
-Billy_Burke_7558.png
-Billy_Zane_7799.png
-Billy_Zane_7821.png
-Bobbie_Eakes_23690.png
-Bobbie_Eakes_23702.png
-Brad_Pitt_8285.png
-Bradley_Cooper_8416.png
-Brianna_Brown_10455.png
-Brianna_Brown_10482.png
-Brooke_Langton_42515.png
-Brooke_Langton_42603.png
-Brooke_Langton_42604.png
-Brooke_Langton_42628.png
-Brooke_Langton_42634.png
-Brooke_Langton_42643.png
-Bruce_Greenwood_8831.png
-Candice_Bergen_6028.png
-Candice_Bergen_6055.png
-Candice_Bergen_6080.png
-Carla_Gallo_28085.png
-Carla_Gallo_28139.png
-Carmen_Electra_24315.png
-Caroline_Dhavernas_21720.png
-Cary_Elwes_9746.png
-Casey_Affleck_9925.png
-Cathy_Lee_Crosby_16548.png
-Cathy_Lee_Crosby_16564.png
-Chase_Masterson_48256.png
-Chazz_Palminteri_10758.png
-Cheryl_Hines_35834.png
-Chris_Evans_10942.png
-Chris_Evans_10943.png
-Chris_Evans_10990.png
-Chris_Evans_11000.png
-Chris_Evans_11039.png
-Chris_Kattan_11206.png
-Chris_Klein_11329.png
-Chris_Klein_11346.png
-Chris_Klein_11471.png
-Christa_Miller_51872.png
-Christian_Bale_11961.png
-Christina_Applegate_2776.png
-Christopher_Lloyd_12504.png
-Christopher_Reeve_12692.png
-Christopher_Reeve_12737.png
-Chyler_Leigh_43366.png
-Chyler_Leigh_43374.png
-Ciara_Bravo_10094.png
-Ciara_Bravo_10161.png
-Ciara_Bravo_10214.png
-Clint_Eastwood_12918.png
-Colin_Farrell_13196.png
-Colin_Farrell_13268.png
-Colin_Farrell_13282.png
-Colin_Firth_13455.png
-Colin_Firth_13566.png
-Courteney_Cox_16360.png
-Crystal_Chappell_13859.png
-Crystal_Chappell_13921.png
-Crystal_Chappell_13932.png
-Crystal_Chappell_13949.png
-Crystal_Chappell_13961.png
-Dan_Lauria_13992.png
-Dan_Lauria_14073.png
-Dana_Delany_20425.png
-Dana_Delany_20561.png
-Dana_Delany_20621.png
-Dana_Delany_20632.png
-Daniel_Day-Lewis_14356.png
-Daniel_Day-Lewis_14361.png
-Daniel_Day-Lewis_14397.png
-Daniel_Day-Lewis_14545.png
-Daniel_Radcliffe_14770.png
-David_Schwimmer_15942.png
-David_Wenham_16210.png
-David_Wenham_16221.png
-David_Wenham_16224.png
-David_Wenham_16235.png
-David_Wenham_16241.png
-David_Wenham_16261.png
-David_Wenham_16295.png
-David_Wenham_16300.png
-David_Wenham_16327.png
-Dean_Cain_16435.png
-Debra_Messing_50555.png
-Debra_Messing_50597.png
-Debra_Messing_50703.png
-Debra_Messing_50746.png
-Delta_Burke_11111.png
-Delta_Burke_11176.png
-Denzel_Washington_16698.png
-Desmond_Harrington_17178.png
-Diahann_Carroll_12351.png
-Diahann_Carroll_12397.png
-Dianna_Agron_269.png
-Dianna_Agron_363.png
-Dianna_Agron_375.png
-Dianna_Agron_396.png
-Dustin_Hoffman_17932.png
-Elizabeth_Berkley_6214.png
-Elizabeth_Berkley_6259.png
-Elizabeth_Hendrickson_34413.png
-Ellen_Greene_30978.png
-Ellen_Greene_31003.png
-Ellen_Greene_31005.png
-Emily_Deschanel_20666.png
-Erin_Cummings_17368.png
-Erin_Cummings_17400.png
-Erin_Cummings_17457.png
-Ethan_Hawke_19548.png
-Eva_Longoria_45027.png
-Farah_Fath_25299.png
-Farah_Fath_25343.png
-Farah_Fath_25365.png
-Farah_Fath_25401.png
-Farrah_Fawcett_25594.png
-Florencia_Lozano_45592.png
-Florencia_Lozano_45594.png
-Florencia_Lozano_45606.png
-Florencia_Lozano_45632.png
-Florencia_Lozano_45640.png
-Fran_Drescher_22955.png
-Fran_Drescher_23030.png
-Gabrielle_Carteris_12657.png
-Gates_McFadden_49864.png
-Gates_McFadden_49868.png
-Glenn_Close_14672.png
-Glenn_Close_14739.png
-Harrison_Ford_22656.png
-Hayden_Christensen_23089.png
-Heather_Locklear_44410.png
-Heather_Locklear_44498.png
-Heather_Locklear_44504.png
-Heather_Locklear_44527.png
-Holly_Marie_Combs_15225.png
-Ian_Holm_24214.png
-Jackee_Harry_32964.png
-Jackee_Harry_33011.png
-James_Brolin_26065.png
-James_Frain_26352.png
-James_Franco_26574.png
-James_Marsden_26731.png
-James_Marsden_26756.png
-James_McAvoy_26932.png
-Jamie_Lee_Curtis_17869.png
-Jamie_Lee_Curtis_18018.png
-Jamie_Lee_Curtis_18056.png
-Jamie_Luner_46356.png
-Jamie_Luner_46423.png
-Jane_Curtin_17787.png
-Jane_Curtin_17820.png
-Jane_Leeves_42833.png
-Jane_Lynch_46639.png
-Jane_Lynch_46743.png
-January_Jones_38213.png
-January_Jones_38384.png
-January_Jones_38414.png
-Jasmine_Guy_31496.png
-Jasmine_Guy_31562.png
-Jasmine_Guy_31594.png
-Jasmine_Guy_31618.png
-Jasmine_Guy_31678.png
-Jason_Biggs_28339.png
-Jason_Lee_55690.png
-Jason_Lee_55703.png
-Jeanne_Cooper_15551.png
-Jeanne_Cooper_15652.png
-Jeanne_Cooper_15696.png
-Jenilee_Harrison_32833.png
-Jenilee_Harrison_32864.png
-Jenilee_Harrison_32895.png
-Jennette_McCurdy_49314.png
-Jennie_Garth_28672.png
-Jennie_Garth_28686.png
-Jeremy_Sisto_29948.png
-Jessica_Biel_7413.png
-Jessica_Capshaw_11891.png
-Jessica_Capshaw_11909.png
-Jessica_Leccia_42714.png
-Jessica_Leccia_42719.png
-Jill_Eikenberry_24108.png
-Jill_Eikenberry_24159.png
-Jill_Eikenberry_24166.png
-Jill_Eikenberry_24179.png
-Jill_Eikenberry_24188.png
-Jill_Hennessy_35087.png
-Jim_Carrey_30976.png
-Jim_Carrey_31137.png
-Jim_Carrey_31150.png
-Joan_Collins_14963.png
-Joanna_Kerns_40119.png
-Joanna_Kerns_40127.png
-Joanna_Kerns_40145.png
-Joanna_Kerns_40162.png
-Joanna_Kerns_40170.png
-Joanna_Kerns_40177.png
-John_Malkovich_32908.png
-John_Malkovich_32934.png
-Jon_Hamm_33859.png
-Jonathan_Rhys_Meyers_34545.png
-Jonathan_Sadowski_55304.png
-Josh_Brolin_34866.png
-Josh_Duhamel_35073.png
-Josie_Bissett_7155.png
-Josie_Bissett_7222.png
-Josie_Bissett_7243.png
-Joyce_DeWitt_20190.png
-Julia_Louis-Dreyfus_45687.png
-Julia_Louis-Dreyfus_45785.png
-Julie_Marie_Berman_6437.png
-Justin_Long_35979.png
-Justin_Long_35983.png
-Justine_Bateman_4879.png
-Kassie_DePaiva_19841.png
-Katherine_Helmond_34045.png
-Kathy_Baker_3934.png
-Kathy_Baker_3984.png
-Kathy_Griffin_31087.png
-Kathy_Griffin_31113.png
-Kellan_Lutz_36883.png
-Kevin_Costner_37976.png
-Kim_Cattrall_13087.png
-Kim_Delaney_20329.png
-Kim_Delaney_20363.png
-Kim_Fields_26573.png
-Kimberlin_Brown_10564.png
-Kimberly_McCullough_49225.png
-Kirstie_Alley_1565.png
-Kit_Harington_38601.png
-Kit_Harington_38638.png
-Kristen_Alderson_615.png
-Kristen_Alderson_714.png
-Kristen_Johnston_37751.png
-Kristin_Chenoweth_14161.png
-Kristin_Chenoweth_14220.png
-Kristin_Chenoweth_14242.png
-Kristy_McNichol_50183.png
-Kristy_McNichol_50205.png
-Kristy_McNichol_50256.png
-Kristy_McNichol_50272.png
-Kristy_McNichol_50295.png
-Kristy_McNichol_50304.png
-Kristy_McNichol_50314.png
-Lacey_Chabert_13239.png
-Lacey_Chabert_13291.png
-Laura_Innes_37302.png
-Laura_Innes_37329.png
-Laura_Leighton_43101.png
-Laura_Leighton_43169.png
-Lauralee_Bell_5169.png
-Lauralee_Bell_5206.png
-Lauralee_Bell_5272.png
-Lauren_Holly_36179.png
-Lauren_Holly_36191.png
-Lauren_Holly_36211.png
-Lauren_Koslow_40373.png
-Lauren_Koslow_40426.png
-Lauren_Koslow_40428.png
-Lauren_Koslow_40434.png
-Lauren_Koslow_40453.png
-Laurie_Metcalf_50846.png
-Laurie_Metcalf_50878.png
-Laurie_Metcalf_50883.png
-Laurie_Metcalf_50897.png
-Laurie_Metcalf_50916.png
-Laurie_Metcalf_50917.png
-Laurie_Metcalf_50920.png
-Laurie_Metcalf_50946.png
-Laurie_Metcalf_50973.png
-Laurie_Metcalf_50984.png
-Laurie_Metcalf_50986.png
-Laurie_Metcalf_50989.png
-Laurie_Metcalf_51019.png
-Lea_Michele_51327.png
-Lea_Michele_51337.png
-Lea_Michele_51362.png
-Lea_Michele_51363.png
-Lea_Michele_51427.png
-Lecy_Goranson_30511.png
-Lecy_Goranson_30538.png
-Lecy_Goranson_30539.png
-Lesley-Anne_Down_22118.png
-Lesley-Anne_Down_22144.png
-Lesley-Anne_Down_22249.png
-Lexi_Ainsworth_126.png
-Lexi_Ainsworth_129.png
-Lexi_Ainsworth_150.png
-Lexi_Ainsworth_28.png
-Lexi_Ainsworth_47.png
-Lexi_Ainsworth_54.png
-Liev_Schreiber_39673.png
-Linda_Evans_24646.png
-Linda_Evans_24670.png
-Linda_Gray_30789.png
-Linda_Gray_30823.png
-Linda_Gray_30895.png
-Lindsay_Hartley_33089.png
-Lindsay_Hartley_33091.png
-Lindsay_Hartley_33182.png
-Lindsay_Hartley_33188.png
-Lindsay_Hartley_33192.png
-Lisa_Bonet_8489.png
-Lisa_Bonet_8547.png
-Lisa_Kudrow_40922.png
-Lisa_LoCicero_44328.png
-Lisa_LoCicero_44334.png
-Loni_Anderson_2152.png
-Loni_Anderson_2190.png
-Lorraine_Bracco_9632.png
-Lorraine_Bracco_9746.png
-Lourdes_Benedicto_5324.png
-Lourdes_Benedicto_5360.png
-Mary_Beth_Evans_24742.png
-Mary_Crosby_16814.png
-Mary_Crosby_16826.png
-Mary_Crosby_16827.png
-Mary_Crosby_16835.png
-Mary_Crosby_16856.png
-Mary_Crosby_16867.png
-Mary_Crosby_16876.png
-Matt_Czuchry_41100.png
-Matt_Dillon_41441.png
-Matt_Dillon_41584.png
-Melina_Kanakaredes_39464.png
-Melissa_Archer_3127.png
-Melissa_Archer_3169.png
-Melissa_Benoist_5425.png
-Melissa_Benoist_5490.png
-Melissa_Benoist_5507.png
-Melissa_Benoist_5523.png
-Melissa_Benoist_5538.png
-Michael_Douglas_43514.png
-Michael_Douglas_43520.png
-Michael_Douglas_43593.png
-Michael_Landes_43643.png
-Michael_Landes_43712.png
-Michael_Vartan_43752.png
-Michael_Vartan_43873.png
-Mila_Kunis_41190.png
-Miranda_Cosgrove_15892.png
-Molly_Burnett_11287.png
-Molly_Burnett_11323.png
-Morena_Baccarin_3662.png
-Natalia_Livingston_44183.png
-Natalie_Hall_31866.png
-Natalie_Hall_31875.png
-Natalie_Hall_31898.png
-Natalie_Hall_31914.png
-Natalie_Martinez_48106.png
-Natalie_Martinez_48117.png
-Natalie_Martinez_48191.png
-Neve_Campbell_11584.png
-Nicole_Eggert_23928.png
-Nicole_Eggert_24088.png
-Nicole_de_Boer_19422.png
-Nicole_de_Boer_19448.png
-Olivia_d'Abo_18077.png
-Olivia_d'Abo_18126.png
-Olivia_d'Abo_18133.png
-Olivia_d'Abo_18166.png
-Olivia_d'Abo_18188.png
-Olivia_d'Abo_18202.png
-Olivia_d'Abo_18224.png
-Olivia_d'Abo_18228.png
-Olivia_d'Abo_18229.png
-Pamela_Sue_Martin_47876.png
-Pamela_Sue_Martin_48018.png
-Pamela_Sue_Martin_48027.png
-Patricia_Arquette_3360.png
-Patricia_Kalember_39346.png
-Patricia_Kalember_39361.png
-Patricia_Kalember_39365.png
-Patricia_Kalember_39366.png
-Patricia_Kalember_39367.png
-Patricia_Kalember_39369.png
-Peggy_Lipton_43900.png
-Peggy_Lipton_43935.png
-Peggy_Lipton_43962.png
-Peggy_Lipton_44032.png
-Peggy_Lipton_44070.png
-Peggy_Lipton_44071.png
-Peggy_McCay_48867.png
-Peggy_McCay_48870.png
-Peggy_McCay_48883.png
-Peggy_McCay_48897.png
-Peri_Gilpin_29456.png
-Philip_Seymour_Hoffman_47737.png
-Philip_Seymour_Hoffman_47778.png
-Portia_Doubleday_22036.png
-Portia_Doubleday_22101.png
-Portia_de_Rossi_19903.png
-Portia_de_Rossi_19908.png
-Portia_de_Rossi_19957.png
-Portia_de_Rossi_19963.png
-Portia_de_Rossi_19997.png
-Portia_de_Rossi_20036.png
-Portia_de_Rossi_20069.png
-Rachel_Dratch_22646.png
-Rachel_Dratch_22751.png
-Rebecca_Budig_10668.png
-Rebecca_Budig_10714.png
-Rebecca_Budig_10756.png
-Rebecca_Budig_10785.png
-Rebecca_Budig_10792.png
-Rebecca_Budig_10805.png
-Rebecca_Herbst_35381.png
-Rebecca_Herbst_35391.png
-Rebecca_Herbst_35407.png
-Rebecca_Herbst_35454.png
-Richard_E._Grant_48348.png
-Richard_Gere_48555.png
-Richard_Gere_48559.png
-Richard_Madden_48806.png
-Robert_Di_Niro_49242.png
-Robert_Downey_Jr._49395.png
-Robert_Downey_Jr._49485.png
-Robert_Duvall_49683.png
-Robert_Duvall_49684.png
-Robert_Duvall_49744.png
-Robert_Knepper_49865.png
-Roma_Downey_22357.png
-Roma_Downey_22378.png
-Roseanne_Barr_4520.png
-Roseanne_Barr_4546.png
-Roseanne_Barr_4622.png
-Rue_McClanahan_48968.png
-Rue_McClanahan_49092.png
-Rue_McClanahan_49151.png
-Rupert_Friend_50825.png
-Ryan_Gosling_51576.png
-S._Epatha_Merkerson_50436.png
-S._Epatha_Merkerson_50464.png
-Sara_Gilbert_29260.png
-Sara_Gilbert_29333.png
-Sarah_Drew_23164.png
-Sarah_Drew_23166.png
-Sarah_Drew_23174.png
-Sarah_Drew_23209.png
-Sarah_Drew_23245.png
-Sarah_Drew_23272.png
-Sarah_Drew_23286.png
-Sarah_Hyland_36867.png
-Sean_Bean_52769.png
-Selena_Gomez_30458.png
-Selena_Gomez_30471.png
-Seth_Rogen_52915.png
-Seth_Rogen_53117.png
-Seth_Rogen_53123.png
-Shannen_Doherty_21826.png
-Shannon_Kane_39638.png
-Shannon_Kane_39729.png
-Shannon_Kane_39732.png
-Sharon_Case_12719.png
-Sharon_Case_12779.png
-Sharon_Case_12889.png
-Sharon_Gless_29936.png
-Sharon_Gless_29976.png
-Sharon_Gless_29987.png
-Sharon_Gless_29989.png
-Shelley_Hack_31728.png
-Shelley_Hack_31736.png
-Shelley_Hack_31826.png
-Sherilyn_Fenn_25694.png
-Shirley_Jones_38512.png
-Shirley_Jones_38556.png
-Staci_Keanan_39993.png
-Staci_Keanan_40018.png
-Staci_Keanan_40038.png
-Staci_Keanan_40068.png
-Stana_Katic_39947.png
-Susan_Flannery_27075.png
-Susan_Flannery_27095.png
-Susan_Flannery_27128.png
-Swoosie_Kurtz_41518.png
-Swoosie_Kurtz_41605.png
-Tamala_Jones_38648.png
-Tamara_Braun_9874.png
-Tatyana_M._Ali_1228.png
-Tatyana_M._Ali_1266.png
-Tatyana_M._Ali_1290.png
-Tatyana_M._Ali_1349.png
-Taylor_Atelian_3507.png
-Taylor_Atelian_3517.png
-Taylor_Atelian_3526.png
-Taylor_Atelian_3532.png
-Taylor_Atelian_3533.png
-Taylor_Atelian_3553.png
-Taylor_Atelian_3559.png
-Taylor_Atelian_3566.png
-Taylor_Atelian_3581.png
-Taylor_Lautner_54170.png
-Taylor_Lautner_54238.png
-Taylor_Lautner_54258.png
-Taylor_Lautner_54279.png
-Taylor_Lautner_54292.png
-Taylor_Lautner_54296.png
-Taylor_Lautner_54328.png
-Tempestt_Bledsoe_8014.png
-Tempestt_Bledsoe_8062.png
-Tempestt_Bledsoe_8063.png
-Teri_Hatcher_33393.png
-Teri_Hatcher_33395.png
-Teri_Hatcher_33401.png
-Terry_Farrell_25136.png
-Terry_Farrell_25203.png
-Tia_Carrere_12136.png
-Tia_Carrere_12240.png
-Tobey_Maguire_54398.png
-Tom_Hanks_54730.png
-Tyne_Daly_18276.png
-Tyne_Daly_18293.png
-Tyne_Daly_18322.png
-Tyne_Daly_18362.png
-Tyne_Daly_18370.png
-Tyne_Daly_18399.png
-Tyne_Daly_18420.png
-Valerie_Cruz_17158.png
-Valerie_Harper_32638.png
-Valerie_Harper_32705.png
-Valerie_Harper_32771.png
-Vanessa_Marcil_47131.png
-Vanessa_Marcil_47158.png
-Vanessa_Marcil_47182.png
-Veronica_Hamel_32010.png
-Veronica_Hamel_32016.png
-Veronica_Hamel_32020.png
-Veronica_Hamel_32030.png
-Victor_Garber_54809.png
-Victor_Garber_54894.png
-Victoria_Justice_39219.png
-Wanda_De_Jesus_19687.png
-Wendie_Malick_46848.png
-Yasmine_Bleeth_8219.png
-Yasmine_Bleeth_8238.png
-Yasmine_Bleeth_8259.png
-Yasmine_Bleeth_8350.png
-Zooey_Deschanel_20937.png
diff --git a/src/megaface/gen_megaface.py b/src/megaface/gen_megaface.py
deleted file mode 100644
index 28e0e0917..000000000
--- a/src/megaface/gen_megaface.py
+++ /dev/null
@@ -1,236 +0,0 @@
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-#os.environ["CUDA_VISIBLE_DEVICES"] = "7"
-
-from datetime import datetime
-import os.path
-from easydict import EasyDict as edict
-import time
-import json
-import sys
-import numpy as np
-import importlib
-import itertools
-import argparse
-import struct
-import cv2
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_preprocess
-from sklearn.preprocessing import normalize
-#import facenet
-#import lfw
-import mxnet as mx
-from mxnet import ndarray as nd
-#from caffe.proto import caffe_pb2
-
-megaface_out = '/raid5data/dplearn/megaface/MegaFace_Features'
-#facescrub_out = '/raid5data/dplearn/megaface/FaceScrubSubset_Features'
-facescrub_out = '/raid5data/dplearn/megaface/FaceScrub_Features'
-
-
-def do_flip(data):
-  for idx in xrange(data.shape[0]):
-    data[idx,:,:] = np.fliplr(data[idx,:,:])
-
-def get_feature(image_path, bbox, landmark, nets, image_shape, use_align, aligned, use_mean):
-  img = face_preprocess.read_image(image_path, mode='rgb')
-  #print(img.shape)
-  if img is None:
-    print('parse image',image_path,'error')
-    return None
-  if not aligned:
-    _landmark = landmark
-    if not use_align:
-      _landmark = None
-    #cv2.imwrite("./align/origin_%s"%image_path.split('/')[-1], img)
-    img = face_preprocess.preprocess(img, bbox=bbox, landmark=_landmark, image_size='%d,%d'%(image_shape[1], image_shape[2]))
-  else:
-    assert img.shape==(image_shape[1],image_shape[2],image_shape[0])
-    #print('already aligned', image_path, img.shape)
-    #img = cv2.resize(img, (image_shape[2], image_shape[1]))
-  #cv2.imwrite("./align/%s"%image_path.split('/')[-1], img)
-  if use_mean>0:
-    v_mean = np.array([127.5,127.5,127.5], dtype=np.float32).reshape( (1,1,3) )
-    img = img.astype(np.float32) - v_mean
-    img *= 0.0078125
-  img = np.transpose( img, (2,0,1) )
-  F = None
-  for net in nets:
-    embedding = None
-    #ppatch = net.patch
-    for flipid in [0,1]:
-      _img = np.copy(img)
-      if flipid==1:
-        do_flip(_img)
-      #nimg = np.zeros(_img.shape, dtype=np.float32)
-      #nimg[:,ppatch[1]:ppatch[3],ppatch[0]:ppatch[2]] = _img[:, ppatch[1]:ppatch[3], ppatch[0]:ppatch[2]]
-      #_img = nimg
-      input_blob = np.expand_dims(_img, axis=0)
-      data = mx.nd.array(input_blob)
-      db = mx.io.DataBatch(data=(data,))
-      net.model.forward(db, is_train=False)
-      _embedding = net.model.get_outputs()[0].asnumpy().flatten()
-      #print(_embedding.shape)
-      if embedding is None:
-        embedding = _embedding
-      else:
-        embedding += _embedding
-    _norm=np.linalg.norm(embedding)
-    embedding /= _norm
-    if F is None:
-      F = embedding
-    else:
-      #F += embedding
-      F = np.concatenate((F,embedding), axis=0)
-  _norm=np.linalg.norm(F)
-  F /= _norm
-  #print(F.shape)
-  return F
-
-
-def write_bin(path, feature):
-  feature = list(feature)
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', len(feature),1,4,5))
-    f.write(struct.pack("%df"%len(feature), *feature))
-
-def main(args):
-
-  print(args)
-  gpuid = args.gpu
-  ctx = mx.gpu(gpuid)
-  nets = []
-  image_shape = [int(x) for x in args.image_size.split(',')]
-  for model in args.model.split('|'):
-    vec = model.split(',')
-    assert len(vec)>1
-    prefix = vec[0]
-    epoch = int(vec[1])
-    print('loading',prefix, epoch)
-    net = edict()
-    net.ctx = ctx
-    net.sym, net.arg_params, net.aux_params = mx.model.load_checkpoint(prefix, epoch)
-    #net.arg_params, net.aux_params = ch_dev(net.arg_params, net.aux_params, net.ctx)
-    all_layers = net.sym.get_internals()
-    net.sym = all_layers['fc1_output']
-    net.model = mx.mod.Module(symbol=net.sym, context=net.ctx, label_names = None)
-    net.model.bind(data_shapes=[('data', (1, 3, image_shape[1], image_shape[2]))])
-    net.model.set_params(net.arg_params, net.aux_params)
-    #_pp = prefix.rfind('p')+1
-    #_pp = prefix[_pp:]
-    #net.patch = [int(x) for x in _pp.split('_')]
-    #assert len(net.patch)==5
-    #print('patch', net.patch)
-    nets.append(net)
-
-  #megaface_lst = "/raid5data/dplearn/faceinsight_align_megaface.lst"
-  megaface_lst = "/raid5data/dplearn/megaface/megaface_mtcnn_112x112/lst"
-  #facescrub_lst = "/raid5data/dplearn/faceinsight_align_facescrub.lst"
-  facescrub_lst = "/raid5data/dplearn/megaface/facescrubr/small_lst"
-  if args.fsall>0:
-    facescrub_lst = "/raid5data/dplearn/megaface/facescrubr/lst"
-
-  if args.skip==0:
-    i = 0
-    succ = 0
-    for line in open(facescrub_lst, 'r'):
-      if i%1000==0:
-        print("writing fs",i, succ)
-      i+=1
-      image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
-      _path = image_path.split('/')
-      a,b = _path[-2], _path[-1]
-      #a = a.replace(' ', '_')
-      #b = b.replace(' ', '_')
-      out_dir = os.path.join(facescrub_out, a)
-      if not os.path.exists(out_dir):
-        os.makedirs(out_dir)
-      #file, ext = os.path.splitext(b)
-      #image_id = int(file.split('_')[-1])
-      #if image_id==40499 or image_id==10788 or image_id==2367:
-      #  b = file
-      #if len(ext)==0:
-      #  print(image_path)
-      #  image_path = image_path+".jpg"
-      #if facescrub_aligned_root is not None:
-      #  _vec = image_path.split('/')
-      #  _image_path = os.path.join(facescrub_aligned_root, _vec[-2], _vec[-1])
-      #  _base, _ext = os.path.splitext(_image_path)
-      #  if _ext=='.gif':
-      #    _image_path = _base+".jpg"
-      #    print('changing', _image_path)
-      #  if os.path.exists(_image_path):
-      #    image_path = _image_path
-      #    bbox = None
-      #    landmark = None
-      #    aligned = True
-      #  else:
-      #    print("not aligned:",_image_path)
-      feature = get_feature(image_path, bbox, landmark, nets, image_shape, True, aligned, args.mean)
-      if feature is None:
-        print('feature none', image_path)
-        continue
-      #print(np.linalg.norm(feature))
-      out_path = os.path.join(out_dir, b+"_%s_%dx%d.bin"%(args.algo, image_shape[1], image_shape[2]))
-      write_bin(out_path, feature)
-      succ+=1
-    print('fs stat',i, succ)
-
-  #return
-  if args.mf==0:
-    return
-  i = 0
-  succ = 0
-  for line in open(megaface_lst, 'r'):
-    if i%1000==0:
-      print("writing mf",i, succ)
-    i+=1
-    if i<=args.skip:
-      continue
-    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
-    assert aligned==True
-    _path = image_path.split('/')
-    a1, a2, b = _path[-3], _path[-2], _path[-1]
-    out_dir = os.path.join(megaface_out, a1, a2)
-    if not os.path.exists(out_dir):
-      os.makedirs(out_dir)
-      #continue
-    #print(landmark)
-    feature = get_feature(image_path, bbox, landmark, nets, image_shape, True, aligned, args.mean)
-    if feature is None:
-      continue
-    out_path = os.path.join(out_dir, b+"_%s_%dx%d.bin"%(args.algo, image_shape[1], image_shape[2]))
-    #print(out_path)
-    write_bin(out_path, feature)
-    succ+=1
-  print('mf stat',i, succ)
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--batch_size', type=int, help='', default=100)
-  parser.add_argument('--image_size', type=str, help='', default='3,112,112')
-  parser.add_argument('--gpu', type=int, help='', default=0)
-  parser.add_argument('--mean', type=int, help='', default=0)
-  parser.add_argument('--seed', type=int, help='', default=727)
-  parser.add_argument('--skip', type=int, help='', default=0)
-  parser.add_argument('--concat', type=int, help='', default=0)
-  parser.add_argument('--fsall', type=int, help='', default=0)
-  parser.add_argument('--mf', type=int, help='', default=1)
-  parser.add_argument('--algo', type=str, help='', default='mxsphereface20c')
-  #parser.add_argument('--model', type=str, help='', default='../model/sphereface-20-p0_0_96_112_0,22|../model/sphereface-20-p0_0_96_95_0,21|../model/sphereface-20-p0_0_80_95_0,21')
-  #parser.add_argument('--model', type=str, help='', default='../model/sphereface-s60-p0_0_96_112_0,31|../model/sphereface-s60-p0_0_96_95_0,21|../model/sphereface2-s60-p0_0_96_112_0,21|../model/sphereface3-s60-p0_0_96_95_0,23')
-  #parser.add_argument('--model', type=str, help='', default='../model/sphereface-s60-p0_0_96_112_0,31|../model/sphereface-s60-p0_0_96_95_0,21|../model/sphereface2-s60-p0_0_96_112_0,21|../model/sphereface3-s60-p0_0_96_95_0,23|../model/sphereface-20-p0_0_96_112_0,22|../model/sphereface-20-p0_0_96_95_0,21|../model/sphereface-20-p0_0_80_95_0,21')
-  #parser.add_argument('--model', type=str, help='', default='../model/spherefacei-s60-p0_0_96_112_0,135')
-  #parser.add_argument('--model', type=str, help='', default='../model/spherefacei-s60-p0_0_96_95_0,95')
-  parser.add_argument('--model', type=str, help='', default='../model/spherefacei-s60-p0_15_96_112_0,95')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
diff --git a/src/megaface/megaface_noises.txt b/src/megaface/megaface_noises.txt
deleted file mode 100644
index f4bf15621..000000000
--- a/src/megaface/megaface_noises.txt
+++ /dev/null
@@ -1,720 +0,0 @@
-#Please strictly follow the rules in https://github.com/deepinsight/insightface/tree/master/src/megaface/README.md if you want to use this list.
-375/37531198@N00/4174611825_0.jpg
-375/37531198@N00/4174611825_1.jpg
-925/92596048@N00/3910487430_2.jpg
-587/58736276@N00/3023685783_2.jpg
-905/90553008@N00/824787574_0.jpg
-350/35034345966@N01/3086429616_0.jpg
-388/38847947@N03/4351812234_0.jpg
-788/78865207@N05/8267101710_0.jpg
-495/49503172960@N01/2674784671_0.jpg
-481/48122140@N00/3488870947_0.jpg
-770/77029195@N00/490108118_3.jpg
-566/56692742@N05/8202697602_1.jpg
-546/54613470@N00/6155519468_0.jpg
-486/48600102544@N01/206574155_6.jpg
-857/85705006@N00/2052818739_0.jpg
-345/34531071@N07/3374444062_1.jpg
-714/71490499@N06/6460912773_0.jpg
-790/79045485@N00/8312851597_9.jpg
-497/49703429@N00/487110767_5.jpg
-588/58819758@N00/502852994_0.jpg
-603/60370779@N07/12733436605_1.jpg
-351/35197415@N06/3392955585_0.jpg
-414/41464693@N02/4431128027_1.jpg
-745/74536253@N07/7744893278_0.jpg
-745/74536253@N07/7744893278_1.jpg
-745/74536253@N07/7744893278_2.jpg
-745/74536253@N07/7746445288_0.jpg
-720/72093892@N00/8394540214_0.jpg
-943/94306126@N05/8585087758_0.jpg
-316/31635293@N06/10300923803_0.jpg
-316/31635293@N06/10300923803_1.jpg
-461/46151867@N00/1361048872_0.jpg
-316/31658559@N04/11248071845_2.jpg
-510/51035763730@N01/3487351916_0.jpg
-381/38176611@N04/6845223441_1.jpg
-812/8127814@N07/1169159373_0.jpg
-596/59689442@N00/397368405_0.jpg
-457/45702277@N04/4404748345_2.jpg
-457/45702277@N04/4405512456_0.jpg
-315/31557769@N04/5183029470_0.jpg
-105/105042823@N05/10224905845_1.jpg
-233/23302147@N08/6717536683_0.jpg
-943/94324445@N02/8585299005_5.jpg
-865/86525452@N06/9026905203_1.jpg
-575/57593906@N00/183103326_0.jpg
-575/57593906@N00/183103326_1.jpg
-513/51313580@N08/6077728920_10.jpg
-794/79447530@N00/2646557330_3.jpg
-838/8385182@N06/8791791352_2.jpg
-834/83452322@N00/1379291226_0.jpg
-299/29932833@N06/4605882096_0.jpg
-125/12523672@N02/1368391674_0.jpg
-251/25193498@N00/4015722351_2.jpg
-207/20741443@N00/5045852254_2.jpg
-313/31369133@N04/5050193084_3.jpg
-313/31369133@N04/5050193084_5.jpg
-504/50440126@N03/8061480293_0.jpg
-353/35336901@N00/4353383267_0.jpg
-102/10249456@N08/839447586_7.jpg
-621/62158648@N07/5765099754_0.jpg
-779/77991982@N00/3351774300_0.jpg
-779/77991982@N00/3351774300_2.jpg
-207/20741443@N00/5045852254_0.jpg
-579/57954142@N07/6082189380_0.jpg
-210/21091679@N08/5811404199_1.jpg
-555/55530018@N04/5481680731_1.jpg
-830/8302818@N06/7180492423_4.jpg
-108/108788939@N08/11207124956_0.jpg
-358/35803015@N03/5571012744_5.jpg
-125/12508217@N08/3458308016_3.jpg
-526/52617239@N07/9550573423_3.jpg
-760/7600622@N07/442516245_2.jpg
-838/83855583@N00/139393859_0.jpg
-357/35775623@N00/4387679624_0.jpg
-508/50824803@N04/5101695889_0.jpg
-813/81369265@N00/5962625787_1.jpg
-849/84987970@N00/6503971169_3.jpg
-366/36692718@N00/7508265904_1.jpg
-101/101878765@N05/9792203325_2.jpg
-149/14994959@N04/3611150762_2.jpg
-257/25777683@N04/3781609919_2.jpg
-206/20625597@N07/3362909564_1.jpg
-768/7686538@N06/1370874019_0.jpg
-286/28675521@N05/2677267827_0.jpg
-958/95818633@N00/317066357_0.jpg
-114/11461909@N06/3569039014_2.jpg
-114/11461909@N06/3926753297_2.jpg
-827/8271358@N08/4748146669_0.jpg
-341/34120602@N05/5979478480_5.jpg
-114/11461909@N06/8281364025_1.jpg
-467/46721620@N00/3849799981_0.jpg
-407/40719493@N06/4008300252_1.jpg
-764/76491372@N00/540480752_1.jpg
-330/33019425@N00/275100215_2.jpg
-316/31677763@N00/273039584_0.jpg
-213/21349468@N00/82691007_2.jpg
-425/42568333@N05/4079636121_3.jpg
-972/97205184@N04/10076409423_0.jpg
-827/8279509@N08/2712311284_0.jpg
-311/31150907@N03/4133932870_0.jpg
-412/41259870@N06/5812928114_0.jpg
-332/33213804@N04/7599995760_0.jpg
-375/37552553@N03/8157072811_0.jpg
-737/7376233@N08/8712184839_0.jpg
-665/66598017@N08/9614369266_3.jpg
-352/35237099299@N01/16090400_0.jpg
-395/39545243@N00/1121668561_11.jpg
-529/52968060@N00/4836487807_1.jpg
-466/46648278@N00/68839900_0.jpg
-768/76841197@N00/349452085_0.jpg
-971/97193933@N00/7441014914_1.jpg
-711/71164686@N00/2199159152_3.jpg
-956/95633051@N00/2504698743_0.jpg
-850/85094310@N00/2536647566_0.jpg
-850/85094310@N00/2536647566_1.jpg
-546/54673576@N00/3231373075_1.jpg
-419/41942641@N02/3911619734_0.jpg
-706/70654989@N00/435992217_0.jpg
-665/66588862@N00/4990604401_0.jpg
-665/66588862@N00/4990604401_1.jpg
-175/17514301@N00/504360950_0.jpg
-242/24256726@N00/2284033395_1.jpg
-388/38847947@N03/4278156225_0.jpg
-120/12057709@N00/4817523632_5.jpg
-958/95876452@N07/8757383318_5.jpg
-958/95876452@N07/9136979141_0.jpg
-511/51135741@N00/1197314078_2.jpg
-732/73257494@N00/254360075_0.jpg
-217/21725313@N04/2735944275_4.jpg
-859/85928975@N00/3054044988_0.jpg
-803/8035175@N05/511570673_2.jpg
-169/16922208@N08/5320713367_0.jpg
-370/37016273@N02/8089047614_0.jpg
-108/10807869@N03/9369308884_2.jpg
-859/85928975@N00/3320166109_0.jpg
-365/36521958172@N01/100498068_4.jpg
-724/7243800@N05/2227611248_0.jpg
-724/7243800@N05/2227611248_3.jpg
-117/11738031@N00/2998834476_0.jpg
-795/7959345@N03/5237906433_0.jpg
-236/23634892@N07/8479172294_0.jpg
-334/33484476@N00/1631862642_0.jpg
-793/7934170@N07/4553285915_3.jpg
-603/60370779@N07/12733436605_0.jpg
-943/94366076@N00/3048525915_0.jpg
-277/27721729@N07/3509070727_2.jpg
-759/75933558@N00/3620382660_0.jpg
-460/46041863@N08/4441052512_0.jpg
-129/12989638@N00/6212397379_0.jpg
-337/33757663@N00/7358306448_0.jpg
-943/94306126@N05/8585087758_1.jpg
-953/95388136@N08/8862410806_0.jpg
-223/22320444@N08/5673545350_0.jpg
-434/43411679@N00/262399750_0.jpg
-635/63534101@N00/2646352652_1.jpg
-225/22591084@N00/6196535614_0.jpg
-635/63587775@N03/6196638593_25.jpg
-353/35399223@N08/6016093673_0.jpg
-410/41063459@N02/3864724318_0.jpg
-822/8220615@N05/4552969733_3.jpg
-549/54949957@N00/5233621265_1.jpg
-701/70177760@N04/9477029774_4.jpg
-105/10564470@N04/2648385175_0.jpg
-199/19909714@N00/3830545099_2.jpg
-758/75818171@N02/7146473317_0.jpg
-233/23354880@N03/2452852823_0.jpg
-399/39915396@N00/5193727577_1.jpg
-575/57503924@N07/6296483205_0.jpg
-430/43093552@N00/6705048843_0.jpg
-417/41739894@N04/4830148462_0.jpg
-417/41739894@N04/4830148462_1.jpg
-101/101386609@N04/13523151125_0.jpg
-381/38117284@N00/1590829110_0.jpg
-827/8279509@N08/2712311284_1.jpg
-379/37963250@N06/3588816064_0.jpg
-510/51035607419@N01/3883605563_0.jpg
-311/31150907@N03/4133932870_1.jpg
-311/31150907@N03/4333226349_0.jpg
-409/40989913@N03/5368897603_1.jpg
-332/33213804@N04/7599995760_2.jpg
-519/51973188@N06/8553574872_0.jpg
-665/66598017@N08/9614369266_1.jpg
-131/13117300@N05/2323998041_0.jpg
-489/48949906@N00/3917465141_1.jpg
-489/48949906@N00/3917465141_2.jpg
-251/25117187@N03/9118076515_0.jpg
-225/22547477@N05/10576991153_1.jpg
-169/16926338@N06/3551147251_1.jpg
-120/12057709@N00/4817523632_1.jpg
-327/32721085@N04/5095989423_2.jpg
-404/40406266@N08/7499062102_1.jpg
-227/22766186@N07/5625438347_0.jpg
-227/22766186@N07/5626032050_2.jpg
-302/30237597@N00/3761331400_0.jpg
-412/41259870@N06/5811920381_0.jpg
-369/36934384@N07/7153191901_0.jpg
-369/36934384@N07/7153191901_1.jpg
-332/33213804@N04/7599995760_6.jpg
-526/52604186@N03/9982748374_2.jpg
-125/12508217@N08/3458308016_0.jpg
-793/79383703@N08/8188874944_2.jpg
-328/32820037@N04/12550334735_0.jpg
-554/55456932@N08/5239999140_0.jpg
-193/19323934@N05/5236281857_0.jpg
-300/30035643@N03/6318064786_1.jpg
-260/26071009@N04/8133430835_0.jpg
-395/39545243@N00/1121668561_23.jpg
-927/92723020@N00/198034899_0.jpg
-732/73257494@N00/254360075_1.jpg
-768/76815233@N00/318365457_0.jpg
-768/76815233@N00/318365457_1.jpg
-758/75818171@N02/6811401849_0.jpg
-960/96009072@N00/7213113090_1.jpg
-440/44042276@N00/11078882754_1.jpg
-563/56379629@N00/231943094_0.jpg
-984/98411817@N00/2372250142_0.jpg
-984/98411817@N00/2372250142_4.jpg
-758/75815807@N00/3141877049_8.jpg
-779/77991982@N00/3351774300_1.jpg
-956/95668756@N00/9137041215_1.jpg
-664/66487272@N00/3175332074_0.jpg
-714/71484472@N08/7749429212_0.jpg
-622/62292236@N03/12280954993_0.jpg
-716/71635685@N00/2171921996_1.jpg
-703/70335407@N00/2629913923_0.jpg
-375/37526864@N03/3575384689_0.jpg
-441/44124482892@N01/4076872185_0.jpg
-760/7600622@N07/442516245_2.jpg
-923/92399379@N00/3745627039_0.jpg
-236/23634892@N07/8479172294_1.jpg
-620/62092813@N00/12460977104_8.jpg
-486/48600102544@N01/206574155_2.jpg
-486/48600102544@N01/206574155_9.jpg
-120/12057709@N00/4817523632_4.jpg
-421/42197860@N05/7964010316_0.jpg
-790/79045485@N00/8312851597_13.jpg
-943/94366076@N00/3048525915_2.jpg
-904/9047144@N06/4039658431_2.jpg
-222/22292214@N00/4243141797_0.jpg
-736/73694732@N00/4431487460_0.jpg
-527/52706816@N04/4862401074_0.jpg
-696/69654695@N04/8112732403_0.jpg
-147/14754516@N00/6746387961_0.jpg
-497/49703429@N00/487110767_7.jpg
-145/14537247@N02/3736211228_8.jpg
-286/28629203@N00/3064397634_3.jpg
-332/33213624@N05/3106373748_0.jpg
-334/33415234@N02/3114165513_1.jpg
-314/31492856@N08/3175678730_1.jpg
-333/33369864@N07/3244523696_4.jpg
-309/30975003@N06/3322588992_1.jpg
-262/26212498@N08/3372483710_1.jpg
-373/37310116@N04/3434757683_0.jpg
-246/24652718@N07/3504274334_0.jpg
-311/31170710@N02/3541648434_0.jpg
-868/8683186@N05/3611020792_0.jpg
-868/8683186@N05/3611020792_6.jpg
-145/14537247@N02/3736211228_1.jpg
-145/14537247@N02/3736211228_10.jpg
-145/14537247@N02/3736211228_3.jpg
-145/14537247@N02/3736211228_7.jpg
-403/40384578@N06/4101631794_0.jpg
-495/49503010002@N01/4159540475_1.jpg
-868/8683186@N05/4808516991_6.jpg
-412/41275121@N06/4903601396_5.jpg
-529/52947003@N08/5062807160_1.jpg
-529/52947003@N08/5062807160_10.jpg
-529/52947003@N08/5062807160_6.jpg
-556/55642376@N02/5907157097_2.jpg
-556/55642376@N02/5907157097_3.jpg
-294/29475291@N05/6520878321_0.jpg
-788/78832981@N04/7439193650_0.jpg
-861/86125876@N07/7890416460_0.jpg
-495/49503124519@N01/8054014324_8.jpg
-170/17040371@N08/8095753999_5.jpg
-170/17040371@N08/8095788948_4.jpg
-336/33672038@N05/8147806935_1.jpg
-340/34085730@N06/8292774555_1.jpg
-268/26881063@N08/13878266544_2.jpg
-309/30932831@N00/2773069134_2.jpg
-276/27674701@N00/306654282_1.jpg
-402/40272616@N07/4800636882_1.jpg
-980/98055082@N00/992566662_1.jpg
-232/23206546@N04/8621794327_0.jpg
-232/23206546@N04/8621794327_1.jpg
-224/22491837@N08/3626048130_0.jpg
-224/22491837@N08/3641245154_2.jpg
-950/95026061@N00/2542285242_0.jpg
-759/75923111@N02/8430338965_0.jpg
-106/10647915@N04/6503063251_1.jpg
-759/7598168@N03/5226476291_0.jpg
-105/105042823@N05/10224905845_2.jpg
-450/45019754@N02/7606879826_2.jpg
-943/94324445@N02/8585299005_4.jpg
-310/31083236@N06/2986649565_0.jpg
-441/44124482892@N01/4076872185_4.jpg
-292/29276771@N03/11641854816_0.jpg
-363/36382552@N04/5425274524_3.jpg
-767/76795962@N03/12297733704_0.jpg
-121/121110687@N08/13323427194_2.jpg
-229/22969849@N00/1349964647_17.jpg
-371/37147296@N03/3720755210_3.jpg
-366/36684003@N06/4437979063_0.jpg
-533/53384351@N06/4978715463_0.jpg
-125/12572929@N05/5221666460_1.jpg
-641/64158255@N00/5489160164_0.jpg
-299/29912007@N05/5921111759_2.jpg
-388/38820321@N06/6281589012_0.jpg
-266/26657772@N02/7323904516_0.jpg
-812/81253095@N08/7563494124_0.jpg
-812/81253095@N08/7563494124_7.jpg
-131/13117300@N05/2323998041_1.jpg
-239/23927487@N05/4499177940_1.jpg
-624/62477948@N02/5691704999_3.jpg
-458/45842803@N00/2702028061_3.jpg
-213/21385420@N00/9351830159_2.jpg
-111/11149280@N02/9354553803_0.jpg
-297/29785808@N04/2869578452_0.jpg
-581/58167807@N00/4216179051_0.jpg
-309/30932831@N00/2773069134_0.jpg
-113/113235388@N08/11701952496_5.jpg
-431/43152922@N00/3293843245_0.jpg
-270/27088911@N08/5601971419_0.jpg
-291/29123483@N00/8613949068_1.jpg
-619/61992454@N00/119165477_2.jpg
-260/26071009@N04/8133430835_1.jpg
-561/56198214@N00/11354524454_0.jpg
-295/29599105@N00/5802624346_6.jpg
-330/33085931@N08/9400799871_0.jpg
-341/34128229@N06/3267947070_0.jpg
-213/21385420@N00/9351830159_0.jpg
-758/75815807@N00/3141877049_7.jpg
-529/52909086@N08/4912165639_0.jpg
-923/92319630@N00/2022738688_0.jpg
-318/31826784@N06/3346271475_0.jpg
-437/43791698@N07/4031782636_4.jpg
-269/26966164@N08/2699439482_3.jpg
-525/52545972@N00/7991923484_0.jpg
-250/25030443@N03/2852010154_6.jpg
-886/88664590@N00/702044241_0.jpg
-658/65873073@N00/8667101819_1.jpg
-274/27433266@N00/5972034863_0.jpg
-292/29233640@N07/3691901778_5.jpg
-122/12254431@N02/5261928850_0.jpg
-765/76562640@N00/2561153799_0.jpg
-664/66487272@N00/3175332074_3.jpg
-349/34946027@N06/3899131678_3.jpg
-914/91409706@N00/7296129156_3.jpg
-914/91409706@N00/7296129156_5.jpg
-329/32996060@N05/7496967290_1.jpg
-241/24112094@N00/7545432624_0.jpg
-588/58871905@N03/8212289784_0.jpg
-284/28426408@N00/4985494194_0.jpg
-315/31594932@N00/8242400525_1.jpg
-315/31594932@N00/8243469114_0.jpg
-500/50016899@N03/4605388504_0.jpg
-263/26357527@N05/7223540266_2.jpg
-498/49814762@N00/8014123645_0.jpg
-498/49814762@N00/8014123645_1.jpg
-875/87504239@N08/10670731404_4.jpg
-875/87504239@N08/10852751085_3.jpg
-372/37244828@N04/13173856803_10.jpg
-371/37147296@N03/3720755210_1.jpg
-125/12572929@N05/5221666460_2.jpg
-641/64158255@N00/5489160164_2.jpg
-356/35658425@N08/5726282580_0.jpg
-766/76681787@N00/5957052480_2.jpg
-812/81253095@N08/7563494124_10.jpg
-812/81253095@N08/7563494124_6.jpg
-699/69958247@N05/13388922034_0.jpg
-748/74896593@N07/6764119963_0.jpg
-846/84655869@N08/7775105362_0.jpg
-181/18155385@N00/239228291_7.jpg
-421/42103613@N08/4040845679_1.jpg
-313/31369133@N04/5050193084_4.jpg
-785/78573292@N00/2256084892_0.jpg
-167/16759096@N00/489481531_0.jpg
-957/95723376@N06/8735477642_1.jpg
-863/8630870@N02/5322350325_0.jpg
-863/8630870@N02/5322350325_2.jpg
-863/8630870@N02/5322350325_3.jpg
-863/8630870@N02/5322350325_4.jpg
-863/8630870@N02/5322350325_5.jpg
-735/73531427@N00/218215108_0.jpg
-309/30932831@N00/2773069134_3.jpg
-727/72707136@N00/33992579_1.jpg
-762/76236359@N00/6234006071_5.jpg
-736/73626930@N00/9428192473_0.jpg
-105/105042823@N05/10224905845_4.jpg
-284/28476480@N04/3975041778_0.jpg
-943/94324445@N02/8585299005_1.jpg
-601/60125017@N00/344049639_2.jpg
-876/87619178@N03/8590731823_0.jpg
-277/27712137@N04/2623505924_0.jpg
-701/70154022@N00/3925362192_1.jpg
-459/45975847@N07/4691684655_2.jpg
-459/45975847@N07/4692316090_0.jpg
-514/51460103@N07/4823559383_0.jpg
-220/22072051@N03/5011260955_0.jpg
-535/53533856@N08/5361207739_0.jpg
-535/53533856@N08/5361207739_1.jpg
-535/53533856@N08/5361207739_4.jpg
-528/52858230@N08/5564688606_0.jpg
-574/57438188@N07/5627487722_0.jpg
-176/17694278@N04/6508529521_0.jpg
-362/36204621@N08/6698329979_0.jpg
-697/69794372@N08/6785690647_2.jpg
-100/100739634@N06/9717775343_0.jpg
-267/26728047@N05/5524462661_5.jpg
-156/15699085@N05/6408711257_0.jpg
-327/32721085@N04/5095989423_0.jpg
-669/66964839@N00/8741217920_1.jpg
-608/60842486@N00/2702417524_0.jpg
-206/20625597@N07/3362909564_5.jpg
-554/55456932@N08/5239999140_1.jpg
-616/61628022@N02/6999143216_0.jpg
-685/68558611@N06/6390155109_1.jpg
-770/77029195@N00/490108118_4.jpg
-258/25827417@N00/2574466319_0.jpg
-878/87855339@N00/3244971644_0.jpg
-878/87855339@N00/3244971644_1.jpg
-654/65497908@N00/6813913952_1.jpg
-868/8683186@N05/3611020792_2.jpg
-868/8683186@N05/4808516991_9.jpg
-294/29475291@N05/6520878321_1.jpg
-716/71678139@N05/6807828131_0.jpg
-716/71678139@N05/6807828131_1.jpg
-716/71678139@N05/6807828131_2.jpg
-716/71678139@N05/6807828131_3.jpg
-716/71678139@N05/6807828131_4.jpg
-716/71678139@N05/6807828131_5.jpg
-716/71678139@N05/6807828131_6.jpg
-170/17040371@N08/8095753999_3.jpg
-170/17040371@N08/8095788948_3.jpg
-336/33672038@N05/8147806935_2.jpg
-441/44124395142@N01/407637002_0.jpg
-795/79543373@N00/2529893768_1.jpg
-575/57545119@N04/8673142329_0.jpg
-359/35909637@N06/3917967715_0.jpg
-407/40719493@N06/4008300252_0.jpg
-764/76491372@N00/540480752_0.jpg
-221/22132798@N08/8018992056_1.jpg
-402/40245280@N00/1936928371_2.jpg
-114/114772050@N03/12008186175_0.jpg
-636/63601558@N00/199623056_0.jpg
-185/18502090@N00/2524917681_47.jpg
-793/79398354@N00/2622290309_0.jpg
-898/8982863@N07/10257490644_5.jpg
-898/8982863@N07/10257490644_8.jpg
-458/45861060@N00/7900853588_1.jpg
-102/102077743@N07/12190186456_0.jpg
-228/22882274@N04/3632475832_2.jpg
-233/23357263@N03/3953791578_5.jpg
-317/31769130@N03/11011170923_0.jpg
-371/37195519@N02/3424150218_0.jpg
-795/79589933@N00/3352567863_0.jpg
-623/62362697@N00/8577869641_3.jpg
-365/36521958172@N01/100498068_1.jpg
-898/8982863@N07/10257490644_10.jpg
-102/102077743@N07/12190186456_2.jpg
-354/35468148224@N01/12347079_0.jpg
-905/90516586@N00/159075833_0.jpg
-773/77364737@N00/2503183115_0.jpg
-975/97584199@N00/2632241655_1.jpg
-527/52772894@N00/90756428_0.jpg
-527/52772894@N00/90756428_2.jpg
-972/97205184@N04/10076409423_0.jpg
-101/101386609@N04/13523144805_0.jpg
-827/8279509@N08/2712311284_0.jpg
-379/37963250@N06/3588816064_1.jpg
-379/37963250@N06/3588816064_2.jpg
-510/51035607419@N01/3883605563_1.jpg
-311/31150907@N03/4133932870_0.jpg
-409/40989913@N03/5368897603_0.jpg
-412/41259870@N06/5812928114_0.jpg
-332/33213804@N04/7599995760_0.jpg
-332/33213804@N04/7599995760_1.jpg
-375/37552553@N03/8157072811_0.jpg
-665/66598017@N08/9614369266_3.jpg
-241/24112094@N00/3994684107_2.jpg
-388/38847947@N03/4351812234_5.jpg
-510/51035620166@N01/237428859_2.jpg
-450/45019754@N02/7606879826_0.jpg
-664/66474562@N00/19195378_2.jpg
-287/28771658@N03/3905862519_0.jpg
-123/12360228@N04/7166710688_1.jpg
-123/12360228@N04/7166710688_2.jpg
-302/30245869@N07/2843650382_4.jpg
-972/97205184@N04/10079707226_0.jpg
-570/57023246@N00/4623144791_0.jpg
-332/33213804@N04/7599995760_4.jpg
-487/48782814@N07/13847705683_0.jpg
-595/59533494@N04/5479759436_0.jpg
-938/93841400@N00/8327059841_1.jpg
-824/8240241@N06/4114864483_0.jpg
-513/51313580@N08/6077728920_5.jpg
-289/28917877@N00/6333929606_0.jpg
-395/39545243@N00/1121668561_35.jpg
-640/64073015@N00/2831021702_0.jpg
-141/14195956@N04/3948300818_3.jpg
-539/53906287@N00/5101060235_7.jpg
-500/50034633@N05/4837007562_1.jpg
-254/25414047@N00/145668608_3.jpg
-489/48949906@N00/3917465141_0.jpg
-808/80812769@N00/442495313_0.jpg
-459/45964884@N08/5242276755_2.jpg
-103/10394437@N03/4568233297_5.jpg
-169/16989146@N06/7641494140_4.jpg
-597/59751999@N00/3289666452_1.jpg
-437/43709093@N07/4256796727_0.jpg
-445/44521275@N00/506634574_2.jpg
-716/71616417@N05/6509071195_2.jpg
-716/71616417@N05/6509074321_2.jpg
-524/52407821@N00/7984244356_0.jpg
-366/36624962@N03/8288304168_3.jpg
-697/69756126@N00/4662338037_0.jpg
-631/63177605@N08/6792309851_2.jpg
-616/61670399@N07/10102978284_1.jpg
-105/105042823@N05/10225015423_0.jpg
-273/27357514@N00/11343847725_0.jpg
-463/46327697@N00/2124534031_4.jpg
-984/98411817@N00/2372250142_6.jpg
-257/25797798@N07/3211844955_1.jpg
-111/11134149@N02/3317710642_0.jpg
-388/38847947@N03/4178270889_3.jpg
-822/8220615@N05/4552969733_5.jpg
-611/61109191@N00/5543733500_2.jpg
-806/80604586@N07/9070005912_4.jpg
-241/24113869@N07/9379713988_1.jpg
-748/74888386@N00/9692493485_0.jpg
-695/69501650@N00/46407370_1.jpg
-950/95082307@N00/8773335932_0.jpg
-510/51035620166@N01/237428859_2.jpg
-434/43411679@N00/262399750_3.jpg
-634/63465779@N07/6362029047_0.jpg
-987/98736785@N00/4499296_3.jpg
-715/7155702@N03/2302080410_0.jpg
-715/7155702@N03/2302080410_1.jpg
-715/7155702@N03/2302080410_3.jpg
-715/7155702@N03/2302080410_4.jpg
-715/7155702@N03/2302080410_5.jpg
-715/7155702@N03/3631806363_2.jpg
-874/87472210@N00/3105959620_1.jpg
-246/24638567@N00/3321513653_0.jpg
-945/94545836@N00/1443145007_1.jpg
-362/36298222@N08/5935447071_0.jpg
-222/22284790@N06/2176524544_0.jpg
-747/7478277@N04/3155354570_0.jpg
-747/7478277@N04/3155354570_1.jpg
-375/37531198@N00/4174611825_2.jpg
-375/37531198@N00/4174611825_4.jpg
-768/76841197@N00/349452085_2.jpg
-768/76841197@N00/349452085_3.jpg
-468/46868174@N08/6511978997_0.jpg
-100/10058188@N05/8295748234_0.jpg
-108/108285828@N06/10779570026_1.jpg
-108/108285828@N06/10779778303_0.jpg
-168/16873194@N05/1801191486_0.jpg
-857/8572970@N04/2079107678_1.jpg
-472/47253165@N00/360990372_0.jpg
-647/64767764@N00/450008059_0.jpg
-729/72971773@N00/58588039_0.jpg
-911/91148289@N00/7916666268_1.jpg
-416/41608212@N00/8718283094_0.jpg
-911/91113759@N00/2539574881_0.jpg
-293/29322132@N08/3569343093_3.jpg
-876/87671668@N00/9013803280_0.jpg
-381/38176611@N04/8638753913_4.jpg
-976/97644214@N00/6173832483_1.jpg
-232/23206546@N04/8621794327_6.jpg
-563/56379629@N00/231943094_3.jpg
-864/86429182@N00/330808508_1.jpg
-353/35336901@N00/4353383267_1.jpg
-631/63122283@N06/7978644677_2.jpg
-932/93211492@N06/8487757725_206.jpg
-575/57545119@N04/8673142329_3.jpg
-354/35479068@N06/3306075644_0.jpg
-277/27721729@N07/3509070727_0.jpg
-188/18899983@N00/4100299699_2.jpg
-213/21349468@N00/82691007_1.jpg
-110/11021333@N02/2957267195_2.jpg
-925/9259500@N03/5710775169_0.jpg
-486/48600102544@N01/206574155_11.jpg
-486/48600102544@N01/206574155_12.jpg
-486/48600102544@N01/206574155_7.jpg
-339/33959433@N05/8396364213_0.jpg
-997/9975353@N03/12630170935_1.jpg
-247/24733288@N07/5846541456_0.jpg
-872/8729914@N05/5099947628_0.jpg
-889/88976184@N05/8244475370_1.jpg
-889/88976184@N05/8244475370_3.jpg
-630/63095109@N00/1079079769_0.jpg
-664/66422694@N00/182828985_0.jpg
-645/64503524@N00/2480285692_0.jpg
-864/86429182@N00/330808508_0.jpg
-447/44754496@N00/4287694240_1.jpg
-363/36382552@N04/5425274524_6.jpg
-210/21091679@N08/5811404199_2.jpg
-562/56213435@N08/9008148399_0.jpg
-695/69501650@N00/46407370_0.jpg
-889/88976184@N05/8244475370_0.jpg
-889/88976184@N05/8244475370_2.jpg
-889/88976184@N05/8244475370_4.jpg
-889/88976184@N05/8244475370_5.jpg
-527/52706816@N04/4862401074_1.jpg
-794/79428123@N06/7349516920_1.jpg
-790/79024059@N04/7351250586_0.jpg
-693/69318970@N00/9266107252_1.jpg
-943/94324445@N02/8585410655_1.jpg
-599/59947594@N00/5684370261_0.jpg
-849/84986567@N00/177640275_0.jpg
-345/34514184@N03/3562741086_0.jpg
-158/15803691@N00/6914149962_2.jpg
-817/81708918@N03/7855914580_5.jpg
-309/30975003@N06/3322588992_8.jpg
-900/90054085@N00/43987316_0.jpg
-136/13675730@N06/2452368334_0.jpg
-386/38675455@N00/6803844225_0.jpg
-717/71763791@N00/5532357066_0.jpg
-606/60604521@N07/5584865343_0.jpg
-395/39545243@N00/1121668561_29.jpg
-391/39196943@N05/3992501716_2.jpg
-207/20719052@N04/8646466859_0.jpg
-207/20719052@N04/8646466859_1.jpg
-758/75815807@N00/3141877049_0.jpg
-471/47170787@N05/6477792561_2.jpg
-317/31734244@N00/542792072_0.jpg
-931/9319402@N03/3358320101_0.jpg
-362/36223735@N03/3784170734_3.jpg
-534/53400644@N05/5044502598_0.jpg
-513/51313580@N08/6077728920_1.jpg
-137/13774680@N04/12313323523_0.jpg
-636/63608960@N08/13553141815_2.jpg
-434/43463445@N07/5723825501_0.jpg
-237/23740675@N05/6877915376_0.jpg
-868/8683186@N05/3611020792_4.jpg
-216/21653251@N04/5633592631_0.jpg
-216/21653251@N04/5633592631_1.jpg
-170/17040371@N08/8095753999_1.jpg
-170/17040371@N08/8095788948_1.jpg
-620/62092813@N00/12460977104_4.jpg
-236/23686718@N08/10662589695_3.jpg
-264/26451473@N00/1342290947_5.jpg
-636/63608960@N08/13553141815_6.jpg
-409/40926212@N05/13257463775_7.jpg
-449/44915880@N03/4248497494_1.jpg
-235/23588194@N00/4494098639_1.jpg
-532/53201041@N07/4911217327_0.jpg
-532/53201041@N07/4911259327_0.jpg
-606/60625084@N04/6450812495_0.jpg
-650/65013293@N02/7724906252_3.jpg
-560/56021903@N02/8227297222_2.jpg
-359/35904083@N08/4345934789_0.jpg
-593/59319260@N00/8743143610_2.jpg
-271/27125798@N05/2710711437_0.jpg
-402/40210933@N02/6043303646_0.jpg
-122/12280910@N06/6299940378_1.jpg
-207/20775043@N07/3627508299_0.jpg
-595/59592751@N08/9096179270_0.jpg
-395/39545243@N00/1121668561_14.jpg
-296/29625288@N00/12891226005_0.jpg
-548/54858071@N00/205845735_1.jpg
-751/75148156@N00/3485469298_1.jpg
-407/40764492@N08/3797968770_0.jpg
-333/33369864@N07/3244523696_0.jpg
-312/31236383@N03/3259555928_1.jpg
-385/38524878@N04/3563411595_0.jpg
-259/25955358@N06/3921002900_2.jpg
-788/78832981@N04/7447167480_1.jpg
-696/69696568@N05/6872955321_1.jpg
-757/75741170@N02/6829396274_0.jpg
-724/72448202@N00/359801202_1.jpg
-969/96934953@N00/4315376370_0.jpg
-133/13361855@N00/989708629_1.jpg
-664/66478195@N00/11649050243_3.jpg
-122/12276055@N02/4214939371_0.jpg
-463/46327697@N00/2124534031_3.jpg
-534/53400644@N05/5044502598_1.jpg
-669/66944684@N03/6306275408_0.jpg
-710/71006820@N00/4037980088_0.jpg
-766/76656964@N04/9363467181_3.jpg
-112/11262617@N07/12873348265_0.jpg
-409/40926212@N05/13257463775_4.jpg
-409/40926212@N05/13835647984_0.jpg
-528/52858230@N08/5636750038_0.jpg
-528/52858230@N08/5636750038_1.jpg
-203/20335591@N04/9752849313_1.jpg
-350/35034345966@N01/3086429616_1.jpg
-736/73694732@N00/4430689233_0.jpg
-605/60517117@N00/10370054943_0.jpg
-421/42173774@N03/3889203507_0.jpg
-511/51135741@N00/1197314078_3.jpg
-267/26753167@N04/5922076258_2.jpg
-330/33062815@N00/8606375434_1.jpg
-108/10807869@N03/9369308884_0.jpg
-761/76151808@N00/6945480767_0.jpg
-757/75741170@N02/6829396274_1.jpg
-220/22057861@N07/2126995093_0.jpg
-665/66572814@N00/8242014606_0.jpg
-849/84986567@N00/177640275_1.jpg
-444/44460493@N03/4085897885_0.jpg
-330/33085931@N08/9400799871_1.jpg
-533/53314395@N00/5174087013_11.jpg
-864/86429182@N00/330808508_2.jpg
-151/15133426@N08/3646529532_0.jpg
-302/30223854@N00/6535512351_0.jpg
-929/92903356@N04/8497572867_0.jpg
-412/41232325@N03/9045866185_0.jpg
-582/58246614@N00/2842715287_1.jpg
-513/51304493@N05/4932744035_0.jpg
-301/30120696@N02/3068639840_0.jpg
-407/40764492@N08/3797968770_0.jpg
-231/23100621@N04/3986465605_0.jpg
-288/28891066@N00/4210148052_2.jpg
-940/94066910@N00/448278027_1.jpg
-532/53282124@N00/3858373020_0.jpg
-597/59751999@N00/3289666452_2.jpg
-365/36521958172@N01/100498068_4.jpg
-293/29322132@N08/3569343093_3.jpg
diff --git a/src/megaface/remove_noises.py b/src/megaface/remove_noises.py
deleted file mode 100644
index c69bcefa7..000000000
--- a/src/megaface/remove_noises.py
+++ /dev/null
@@ -1,175 +0,0 @@
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-
-import os.path
-import datetime
-from easydict import EasyDict as edict
-import time
-import json
-import shutil
-import sys
-import numpy as np
-import importlib
-import itertools
-import argparse
-import struct
-import cv2
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_preprocess
-from sklearn.preprocessing import normalize
-import mxnet as mx
-from mxnet import ndarray as nd
-
-
-feature_dim = 512
-feature_ext = 1
-
-def load_bin(path, fill = 0.0):
-  with open(path, 'rb') as f:
-    bb = f.read(4*4)
-    #print(len(bb))
-    v = struct.unpack('4i', bb)
-    #print(v[0])
-    bb = f.read(v[0]*4)
-    v = struct.unpack("%df"%(v[0]), bb)
-    feature = np.full( (feature_dim+feature_ext,), fill, dtype=np.float32)
-    feature[0:feature_dim] = v
-    #feature = np.array( v, dtype=np.float32)
-  #print(feature.shape)
-  #print(np.linalg.norm(feature))
-  return feature
-
-def write_bin(path, feature):
-  feature = list(feature)
-  with open(path, 'wb') as f:
-    f.write(struct.pack('4i', len(feature),1,4,5))
-    f.write(struct.pack("%df"%len(feature), *feature))
-
-def main(args):
-
-  out_algo = args.suffix
-  if len(args.algo)>0:
-    out_algo = args.algo
-
-  fs_noise_map = {}
-  for line in open(args.facescrub_noises, 'r'):
-    if line.startswith('#'):
-      continue
-    line = line.strip()
-    fname = line.split('.')[0]
-    p = fname.rfind('_')
-    fname = fname[0:p]
-    fs_noise_map[line] = fname
-
-  print(len(fs_noise_map))
-
-  i=0
-  fname2center = {}
-  noises = []
-  for line in open(args.facescrub_lst, 'r'):
-    if i%1000==0:
-      print("reading fs",i)
-    i+=1
-    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
-    assert aligned==True
-    _path = image_path.split('/')
-    a, b = _path[-2], _path[-1]
-    feature_path = os.path.join(args.facescrub_feature_dir, a, "%s_%s.bin"%(b, args.suffix))
-    feature_dir_out = os.path.join(args.facescrub_feature_dir_out, a)
-    if not os.path.exists(feature_dir_out):
-      os.makedirs(feature_dir_out)
-    feature_path_out = os.path.join(args.facescrub_feature_dir_out, a, "%s_%s.bin"%(b, out_algo))
-    #print(b)
-    if not b in fs_noise_map:
-      #shutil.copyfile(feature_path, feature_path_out)
-      feature = load_bin(feature_path)
-      write_bin(feature_path_out, feature)
-      if not a in fname2center:
-        fname2center[a] = np.zeros((feature_dim+feature_ext,), dtype=np.float32)
-      fname2center[a] += feature
-    else:
-      #print('n', b)
-      noises.append( (a,b) )
-  print(len(noises))
-
-  for k in noises:
-    a,b = k
-    assert a in fname2center
-    center = fname2center[a]
-    g = np.zeros( (feature_dim+feature_ext,), dtype=np.float32)
-    g2 = np.random.uniform(-0.001, 0.001, (feature_dim,))
-    g[0:feature_dim] = g2
-    f = center+g
-    _norm=np.linalg.norm(f)
-    f /= _norm
-    feature_path_out = os.path.join(args.facescrub_feature_dir_out, a, "%s_%s.bin"%(b, out_algo))
-    write_bin(feature_path_out, f)
-
-  mf_noise_map = {}
-  for line in open(args.megaface_noises, 'r'):
-    if line.startswith('#'):
-      continue
-    line = line.strip()
-    _vec = line.split("\t")
-    if len(_vec)>1:
-      line = _vec[1]
-    mf_noise_map[line] = 1
-
-  print(len(mf_noise_map))
-
-  i=0
-  nrof_noises = 0
-  for line in open(args.megaface_lst, 'r'):
-    if i%1000==0:
-      print("reading mf",i)
-    i+=1
-    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
-    assert aligned==True
-    _path = image_path.split('/')
-    a1, a2, b = _path[-3], _path[-2], _path[-1]
-    feature_path = os.path.join(args.megaface_feature_dir, a1, a2, "%s_%s.bin"%(b, args.suffix))
-    feature_dir_out = os.path.join(args.megaface_feature_dir_out, a1, a2)
-    if not os.path.exists(feature_dir_out):
-      os.makedirs(feature_dir_out)
-    feature_path_out = os.path.join(args.megaface_feature_dir_out, a1, a2, "%s_%s.bin"%(b, out_algo))
-    bb = '/'.join([a1, a2, b])
-    #print(b)
-    if not bb in mf_noise_map:
-      feature = load_bin(feature_path)
-      write_bin(feature_path_out, feature)
-      #shutil.copyfile(feature_path, feature_path_out)
-    else:
-      feature = load_bin(feature_path, 100.0)
-      write_bin(feature_path_out, feature)
-      #g = np.random.uniform(-0.001, 0.001, (feature_dim,))
-      #print('n', bb)
-      #write_bin(feature_path_out, g)
-      nrof_noises+=1
-  print(nrof_noises)
-
-
-def parse_arguments(argv):
-  parser = argparse.ArgumentParser()
-  
-  parser.add_argument('--facescrub-noises', type=str, help='', default='./facescrub_noises.txt')
-  parser.add_argument('--megaface-noises', type=str, help='', default='./megaface_noises.txt')
-  parser.add_argument('--suffix', type=str, help='', default='r100_cm_112x112')
-  parser.add_argument('--algo', type=str, help='', default='')
-  parser.add_argument('--megaface-lst', type=str, help='', default='/raid5data/dplearn/megaface/megaface_mtcnn_112x112/lst')
-  parser.add_argument('--facescrub-lst', type=str, help='', default='/raid5data/dplearn/megaface/facescrubr/small_lst')
-  parser.add_argument('--megaface-feature-dir', type=str, help='', default='/raid5data/dplearn/megaface/MegaFace_Features')
-  parser.add_argument('--facescrub-feature-dir', type=str, help='', default='/raid5data/dplearn/megaface/FaceScrub_Features')
-  #parser.add_argument('--megaface-feature-dir-out', type=str, help='', default='/raid5data/dplearn/megaface/MegaFace_Features_cm')
-  #parser.add_argument('--facescrub-feature-dir-out', type=str, help='', default='/raid5data/dplearn/megaface/FaceScrub_Features_cm')
-  parser.add_argument('--megaface-feature-dir-out', type=str, help='', default='/opt/jiaguo/MegaFace_Features_cm')
-  parser.add_argument('--facescrub-feature-dir-out', type=str, help='', default='/opt/jiaguo/FaceScrub_Features_cm')
-  return parser.parse_args(argv)
-
-if __name__ == '__main__':
-  main(parse_arguments(sys.argv[1:]))
-
-
diff --git a/src/symbols/fdensenet.py b/src/symbols/fdensenet.py
deleted file mode 100644
index 66545dc0f..000000000
--- a/src/symbols/fdensenet.py
+++ /dev/null
@@ -1,192 +0,0 @@
-"""
-Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py 
-Original author Wei Wu
-Referenced https://github.com/bamos/densenet.pytorch/blob/master/densenet.py
-Original author bamos
-Referenced https://github.com/andreasveit/densenet-pytorch/blob/master/densenet.py
-Original author andreasveit
-Referenced https://github.com/Nicatio/Densenet/blob/master/mxnet/symbol_densenet.py
-Original author Nicatio
-
-Implemented the following paper:     DenseNet-BC
-Gao Huang, Zhuang Liu, Kilian Q. Weinberger, Laurens van der Maaten. "Densely Connected Convolutional Networks"
-
-Coded by Lin Xiong Mar-1, 2017
-"""
-import mxnet as mx
-import math
-import symbol_utils
-
-def BasicBlock(data, growth_rate, stride, name, bottle_neck=True, drop_out=0.0, bn_mom=0.9, workspace=512):
-    """Return BaiscBlock Unit symbol for building DenseBlock
-    Parameters
-    ----------
-    data : str
-        Input data
-    growth_rate : int
-        Number of output channels
-    stride : tupe
-        Stride used in convolution
-    drop_out : float
-        Probability of an element to be zeroed. Default = 0.2
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    # import pdb
-    # pdb.set_trace()
-
-    if bottle_neck:
-        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-        bn1   = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1  = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
-        conv1 = mx.sym.Convolution(data=act1, num_filter=int(growth_rate*4), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        if drop_out > 0:
-            conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
-        bn2   = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2  = mx.sym.Activation(data=bn2, act_type='relu', name=name + '_relu2')
-        conv2 = mx.sym.Convolution(data=act2, num_filter=int(growth_rate), kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        if drop_out > 0:
-            conv2 = mx.symbol.Dropout(data=conv2, p=drop_out, name=name + '_dp2')
-        #return mx.symbol.Concat(data, conv2, name=name + '_concat0')
-        return conv2
-    else:
-        bn1   = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1  = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
-        conv1 = mx.sym.Convolution(data=act1, num_filter=int(growth_rate), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        if drop_out > 0:
-            conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
-        #return mx.symbol.Concat(data, conv1, name=name + '_concat0')
-        return conv1
-		
-def DenseBlock(units_num, data, growth_rate, name, bottle_neck=True, drop_out=0.0, bn_mom=0.9, workspace=512):
-    """Return DenseBlock Unit symbol for building DenseNet
-    Parameters
-    ----------
-    units_num : int
-        the number of BasicBlock in each DenseBlock
-    data : str	
-        Input data
-    growth_rate : int
-        Number of output channels
-    drop_out : float
-        Probability of an element to be zeroed. Default = 0.2
-    workspace : int
-        Workspace used in convolution operator
-    """
-    # import pdb
-    # pdb.set_trace()
-
-    for i in range(units_num):
-        Block = BasicBlock(data, growth_rate=growth_rate, stride=(1,1), name=name + '_unit%d' % (i+1), 
-                            bottle_neck=bottle_neck, drop_out=drop_out, 
-                            bn_mom=bn_mom, workspace=workspace)
-        data = mx.symbol.Concat(data, Block, name=name + '_concat%d' % (i+1))
-    return data
-	
-def TransitionBlock(num_stage, data, num_filter, stride, name, drop_out=0.0, bn_mom=0.9, workspace=512):
-    """Return TransitionBlock Unit symbol for building DenseNet
-    Parameters
-    ----------
-    num_stage : int
-        Number of stage
-    data : str
-        Input data
-    num : int
-        Number of output channels
-    stride : tupe
-        Stride used in convolution
-    name : str
-        Base name of the operators
-    drop_out : float
-        Probability of an element to be zeroed. Default = 0.2
-    workspace : int
-        Workspace used in convolution operator
-    """
-    bn1   = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    act1  = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
-    conv1 = mx.sym.Convolution(data=act1, num_filter=num_filter,
-                                kernel=(1,1), stride=stride, pad=(0,0), no_bias=True, 
-                                workspace=workspace, name=name + '_conv1')
-    if drop_out > 0:
-        conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
-    return mx.symbol.Pooling(conv1, global_pool=False, kernel=(2,2), stride=(2,2), pool_type='avg', name=name + '_pool%d' % (num_stage+1))
-
-def get_symbol(num_classes, num_layers, **kwargs):
-    """Return DenseNet symbol of imagenet
-    Parameters
-    ----------
-    units : list
-        Number of units in each stage
-    num_stage : int
-        Number of stage
-    growth_rate : int
-        Number of output channels
-    num_class : int
-        Ouput size of symbol
-    data_type : str
-        the type of dataset
-    reduction : float
-        Compression ratio. Default = 0.5
-    drop_out : float
-        Probability of an element to be zeroed. Default = 0.2
-    workspace : int
-        Workspace used in convolution operator
-    """
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    print(version_input, version_output, version_unit)
-
-    num_stage = 4
-    reduction = 0.5
-    drop_out = 0.2
-    bottle_neck = True
-    bn_mom = 0.9
-    workspace = 256
-    growth_rate = 32
-    if num_layers == 121:
-        units = [6, 12, 24, 16]
-    elif num_layers == 169:
-        units = [6, 12, 32, 32]
-    elif num_layers == 201:
-        units = [6, 12, 48, 32]
-    elif num_layers == 161:
-        units = [6, 12, 36, 24]
-        growth_rate = 48
-    else:
-        raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
-    num_unit = len(units)
-    assert(num_unit == num_stage)
-    init_channels = 2 * growth_rate
-    n_channels = init_channels
-
-    data = mx.sym.Variable(name='data')
-    #data = data-127.5
-    #data = data*0.0078125
-    #if version_input==0:
-    #  body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
-    #                            no_bias=True, name="conv0", workspace=workspace)
-    #else:
-    #  body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(3,3), stride=(1,1), pad=(1,1),
-    #                            no_bias=True, name="conv0", workspace=workspace)
-    #body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-    #body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
-    #body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    body = symbol_utils.get_head(data, version_input, growth_rate*2)
-
-    for i in range(num_stage-1):
-        body = DenseBlock(units[i], body, growth_rate=growth_rate, name='DBstage%d' % (i + 1), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
-        n_channels += units[i]*growth_rate
-        n_channels = int(math.floor(n_channels*reduction))
-        body = TransitionBlock(i, body, n_channels, stride=(1,1), name='TBstage%d' % (i + 1), drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
-    body = DenseBlock(units[num_stage-1], body, growth_rate=growth_rate, name='DBstage%d' % (num_stage), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
diff --git a/src/symbols/fdpn.py b/src/symbols/fdpn.py
deleted file mode 100644
index 0544f8347..000000000
--- a/src/symbols/fdpn.py
+++ /dev/null
@@ -1,229 +0,0 @@
-import mxnet as mx
-import symbol_utils
-
-bn_momentum = 0.9
-
-def BK(data):
-    return mx.symbol.BlockGrad(data=data)
-
-# - - - - - - - - - - - - - - - - - - - - - - -
-# Fundamental Elements
-def BN(data, fix_gamma=False, momentum=bn_momentum, name=None):
-    bn     = mx.symbol.BatchNorm( data=data, fix_gamma=fix_gamma, momentum=bn_momentum, name=('%s__bn'%name))
-    return bn
-
-def AC(data, act_type='relu', name=None):
-    act    = mx.symbol.Activation(data=data, act_type=act_type, name=('%s__%s' % (name, act_type)))
-    return act
-
-def BN_AC(data, momentum=bn_momentum, name=None):
-    bn     = BN(data=data, name=name, fix_gamma=False, momentum=momentum)
-    bn_ac  = AC(data=bn,   name=name)
-    return bn_ac
-
-def Conv(data, num_filter, kernel, stride=(1,1), pad=(0, 0), name=None, no_bias=True, w=None, b=None, attr=None, num_group=1):
-    Convolution = mx.symbol.Convolution
-    if w is None:
-        conv     = Convolution(data=data, num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=('%s__conv' %name), no_bias=no_bias, attr=attr)
-    else:
-        if b is None:
-            conv = Convolution(data=data, num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=('%s__conv' %name), no_bias=no_bias, weight=w, attr=attr)
-        else:
-            conv = Convolution(data=data, num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=('%s__conv' %name), no_bias=False, bias=b, weight=w, attr=attr)
-    return conv
-
-# - - - - - - - - - - - - - - - - - - - - - - -
-# Standard Common functions < CVPR >
-def Conv_BN(   data, num_filter,  kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
-    cov    = Conv(   data=data,   num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=name, w=w, b=b, no_bias=no_bias, attr=attr)
-    cov_bn = BN(     data=cov,    name=('%s__bn' % name))
-    return cov_bn
-
-def Conv_BN_AC(data, num_filter,  kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
-    cov_bn = Conv_BN(data=data,   num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=name, w=w, b=b, no_bias=no_bias, attr=attr)
-    cov_ba = AC(     data=cov_bn, name=('%s__ac' % name))
-    return cov_ba
-
-# - - - - - - - - - - - - - - - - - - - - - - -
-# Standard Common functions < ECCV >
-def BN_Conv(   data, num_filter,  kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
-    bn     = BN(     data=data,   name=('%s__bn' % name))
-    bn_cov = Conv(   data=bn,     num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=name, w=w, b=b, no_bias=no_bias, attr=attr)
-    return bn_cov
-
-def AC_Conv(   data, num_filter,  kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
-    ac     = AC(     data=data,   name=('%s__ac' % name))
-    ac_cov = Conv(   data=ac,     num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=name, w=w, b=b, no_bias=no_bias, attr=attr)
-    return ac_cov
-
-def BN_AC_Conv(data, num_filter,  kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
-    bn     = BN(     data=data,   name=('%s__bn' % name))
-    ba_cov = AC_Conv(data=bn,     num_filter=num_filter, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name=name, w=w, b=b, no_bias=no_bias, attr=attr)
-    return ba_cov
-
-def DualPathFactory(data, num_1x1_a, num_3x3_b, num_1x1_c, name, inc, G, _type='normal'):
-    kw = 3
-    kh = 3
-    pw = (kw-1)/2
-    ph = (kh-1)/2
-
-    # type
-    if _type is 'proj':
-        key_stride = 1
-        has_proj   = True
-    if _type is 'down':
-        key_stride = 2
-        has_proj   = True
-    if _type is 'normal':
-        key_stride = 1
-        has_proj   = False
-
-    # PROJ
-    if type(data) is list:
-        data_in  = mx.symbol.Concat(*[data[0], data[1]],  name=('%s_cat-input' % name))
-    else:
-        data_in  = data
-
-    if has_proj:
-        c1x1_w   = BN_AC_Conv( data=data_in, num_filter=(num_1x1_c+2*inc), kernel=( 1, 1), stride=(key_stride, key_stride), name=('%s_c1x1-w(s/%d)' %(name, key_stride)), pad=(0, 0))
-        data_o1  = mx.symbol.slice_axis(data=c1x1_w, axis=1, begin=0,         end=num_1x1_c,         name=('%s_c1x1-w(s/%d)-split1' %(name, key_stride)))
-        data_o2  = mx.symbol.slice_axis(data=c1x1_w, axis=1, begin=num_1x1_c, end=(num_1x1_c+2*inc), name=('%s_c1x1-w(s/%d)-split2' %(name, key_stride)))
-    else:
-        data_o1  = data[0]
-        data_o2  = data[1]
-        
-    # MAIN
-    c1x1_a = BN_AC_Conv( data=data_in, num_filter=num_1x1_a,       kernel=( 1,  1), pad=( 0,  0), name=('%s_c1x1-a'   % name))
-    c3x3_b = BN_AC_Conv( data=c1x1_a,  num_filter=num_3x3_b,       kernel=(kw, kh), pad=(pw, ph), name=('%s_c%dx%d-b' % (name,kw,kh)), stride=(key_stride,key_stride), num_group=G)
-    c1x1_c = BN_AC_Conv( data=c3x3_b,  num_filter=(num_1x1_c+inc), kernel=( 1,  1), pad=( 0,  0), name=('%s_c1x1-c'   % name))
-    c1x1_c1= mx.symbol.slice_axis(data=c1x1_c, axis=1, begin=0,         end=num_1x1_c,       name=('%s_c1x1-c-split1' % name))
-    c1x1_c2= mx.symbol.slice_axis(data=c1x1_c, axis=1, begin=num_1x1_c, end=(num_1x1_c+inc), name=('%s_c1x1-c-split2' % name))
-
-    # OUTPUTS
-    summ   = mx.symbol.ElementWiseSum(*[data_o1, c1x1_c1],                        name=('%s_sum' % name))
-    dense  = mx.symbol.Concat(        *[data_o2, c1x1_c2],                        name=('%s_cat' % name))
-
-    return [summ, dense]
-
-k_R = 160
-
-G   = 40
-
-k_sec  = {  2: 4, \
-            3: 8, \
-            4: 28, \
-            5: 3   }
-
-inc_sec= {  2: 16, \
-            3: 32, \
-            4: 32, \
-            5: 128 }
-
-def get_symbol(num_classes = 1000, num_layers=92, **kwargs):
-    if num_layers==68:
-      k_R = 128
-      G   = 32
-      k_sec  = {  2: 3,  \
-                  3: 4,  \
-                  4: 12, \
-                  5: 3   }
-      inc_sec= {  2: 16, \
-                  3: 32, \
-                  4: 32, \
-                  5: 64  }
-    elif num_layers==92:
-      k_R = 96
-      G   = 32
-      k_sec  = {  2: 3, \
-                  3: 4, \
-                  4: 20, \
-                  5: 3   }
-      inc_sec= {  2: 16, \
-                  3: 32, \
-                  4: 24, \
-                  5: 128 }
-    elif num_layers==107:
-      k_R = 200
-      G   = 50
-      k_sec  = {  2: 4, \
-                  3: 8, \
-                  4: 20, \
-                  5: 3   }
-      inc_sec= {  2: 20, \
-                  3: 64, \
-                  4: 64, \
-                  5: 128 }
-    elif num_layers==131:
-      k_R = 160
-      G   = 40
-      k_sec  = {  2: 4, \
-                  3: 8, \
-                  4: 28, \
-                  5: 3   }
-      inc_sec= {  2: 16, \
-                  3: 32, \
-                  4: 32, \
-                  5: 128 }
-    else:
-      raise ValueError("no experiments done on dpn num_layers {}, you can do it yourself".format(num_layers))
-
-    version_se = kwargs.get('version_se', 1)
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    print(version_se, version_input, version_output, version_unit)
-
-    ## define Dual Path Network
-    data = mx.symbol.Variable(name="data")
-    #data = data-127.5
-    #data = data*0.0078125
-    #if version_input==0:
-    #  conv1_x_1  = Conv(data=data,  num_filter=128,  kernel=(7, 7), name='conv1_x_1', pad=(3,3), stride=(2,2))
-    #else:
-    #  conv1_x_1  = Conv(data=data,  num_filter=128,  kernel=(3, 3), name='conv1_x_1', pad=(3,3), stride=(1,1))
-    #conv1_x_1  = BN_AC(conv1_x_1, name='conv1_x_1__relu-sp')
-    #conv1_x_x  = mx.symbol.Pooling(data=conv1_x_1, pool_type="max", kernel=(3, 3),  pad=(1,1), stride=(2,2), name="pool1")
-    conv1_x_x = symbol_utils.get_head(data, version_input, 128)
-
-    # conv2
-    bw = 256
-    inc= inc_sec[2]
-    R  = (k_R*bw)/256 
-    conv2_x_x  = DualPathFactory(     conv1_x_x,   R,   R,   bw,  'conv2_x__1',           inc,   G,  'proj'  )
-    for i_ly in range(2, k_sec[2]+1):
-        conv2_x_x  = DualPathFactory( conv2_x_x,   R,   R,   bw, ('conv2_x__%d'% i_ly),   inc,   G,  'normal')
-
-    # conv3
-    bw = 512
-    inc= inc_sec[3]
-    R  = (k_R*bw)/256
-    conv3_x_x  = DualPathFactory(     conv2_x_x,   R,   R,   bw,  'conv3_x__1',           inc,   G,  'down'  )
-    for i_ly in range(2, k_sec[3]+1):
-        conv3_x_x  = DualPathFactory( conv3_x_x,   R,   R,   bw, ('conv3_x__%d'% i_ly),   inc,   G,  'normal')
-
-    # conv4
-    bw = 1024
-    inc= inc_sec[4]
-    R  = (k_R*bw)/256
-    conv4_x_x  = DualPathFactory(     conv3_x_x,   R,   R,   bw,  'conv4_x__1',           inc,   G,  'down'  )
-    for i_ly in range(2, k_sec[4]+1):
-        conv4_x_x  = DualPathFactory( conv4_x_x,   R,   R,   bw, ('conv4_x__%d'% i_ly),   inc,   G,  'normal')
-
-    # conv5
-    bw = 2048
-    inc= inc_sec[5]
-    R  = (k_R*bw)/256
-    conv5_x_x  = DualPathFactory(     conv4_x_x,   R,   R,   bw,  'conv5_x__1',           inc,   G,  'down'  )
-    for i_ly in range(2, k_sec[5]+1):
-        conv5_x_x  = DualPathFactory( conv5_x_x,   R,   R,   bw, ('conv5_x__%d'% i_ly),   inc,   G,  'normal')
-
-    # output: concat
-    conv5_x_x  = mx.symbol.Concat(*[conv5_x_x[0], conv5_x_x[1]],  name='conv5_x_x_cat-final')
-    #conv5_x_x = BN_AC(conv5_x_x, name='conv5_x_x__relu-sp')
-    before_pool = conv5_x_x
-    fc1 = symbol_utils.get_fc1(before_pool, num_classes, fc_type)
-    return fc1
-
-
diff --git a/src/symbols/finception_resnet_v2.py b/src/symbols/finception_resnet_v2.py
deleted file mode 100644
index bd6baad1c..000000000
--- a/src/symbols/finception_resnet_v2.py
+++ /dev/null
@@ -1,168 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-"""
-Contains the definition of the Inception Resnet V2 architecture.
-As described in http://arxiv.org/abs/1602.07261.
-Inception-v4, Inception-ResNet and the Impact of Residual Connections
-on Learning
-Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi
-"""
-import mxnet as mx
-import symbol_utils
-
-
-def ConvFactory(data, num_filter, kernel, stride=(1, 1), pad=(0, 0), act_type="relu", mirror_attr={}, with_act=True):
-    conv = mx.symbol.Convolution(
-        data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad)
-    bn = mx.symbol.BatchNorm(data=conv)
-    if with_act:
-        act = mx.symbol.Activation(
-            data=bn, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return bn
-
-
-def block35(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}):
-    tower_conv = ConvFactory(net, 32, (1, 1))
-    tower_conv1_0 = ConvFactory(net, 32, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 32, (3, 3), pad=(1, 1))
-    tower_conv2_0 = ConvFactory(net, 32, (1, 1))
-    tower_conv2_1 = ConvFactory(tower_conv2_0, 48, (3, 3), pad=(1, 1))
-    tower_conv2_2 = ConvFactory(tower_conv2_1, 64, (3, 3), pad=(1, 1))
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False)
-
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-
-def block17(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}):
-    tower_conv = ConvFactory(net, 192, (1, 1))
-    tower_conv1_0 = ConvFactory(net, 129, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 160, (1, 7), pad=(1, 2))
-    tower_conv1_2 = ConvFactory(tower_conv1_1, 192, (7, 1), pad=(2, 1))
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False)
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-
-def block8(net, input_num_channels, scale=1.0, with_act=True, act_type='relu', mirror_attr={}):
-    tower_conv = ConvFactory(net, 192, (1, 1))
-    tower_conv1_0 = ConvFactory(net, 192, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 224, (1, 3), pad=(0, 1))
-    tower_conv1_2 = ConvFactory(tower_conv1_1, 256, (3, 1), pad=(1, 0))
-    tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_2])
-    tower_out = ConvFactory(
-        tower_mixed, input_num_channels, (1, 1), with_act=False)
-    net = net+scale * tower_out
-    if with_act:
-        act = mx.symbol.Activation(
-            data=net, act_type=act_type, attr=mirror_attr)
-        return act
-    else:
-        return net
-
-
-def repeat(inputs, repetitions, layer, *args, **kwargs):
-    outputs = inputs
-    for i in range(repetitions):
-        outputs = layer(outputs, *args, **kwargs)
-    return outputs
-
-
-def get_symbol(num_classes=1000, **kwargs):
-    data = mx.symbol.Variable(name='data')
-    data = data-127.5
-    data = data*0.0078125
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    print(version_input, version_output, version_unit)
-
-    if version_input==0:
-      conv1a_3_3 = ConvFactory(data=data, num_filter=32,
-                               kernel=(3, 3), stride=(2, 2))
-    else:
-      conv1a_3_3 = ConvFactory(data=data, num_filter=32,
-                               kernel=(3, 3), stride=(1, 1))
-    conv2a_3_3 = ConvFactory(conv1a_3_3, 32, (3, 3))
-    conv2b_3_3 = ConvFactory(conv2a_3_3, 64, (3, 3), pad=(1, 1))
-    maxpool3a_3_3 = mx.symbol.Pooling(
-        data=conv2b_3_3, kernel=(3, 3), stride=(2, 2), pool_type='max')
-    conv3b_1_1 = ConvFactory(conv2b_3_3, 80, (1, 1))
-    conv4a_3_3 = ConvFactory(conv3b_1_1, 192, (3, 3))
-    maxpool5a_3_3 = mx.symbol.Pooling(
-        data=conv4a_3_3, kernel=(3, 3), stride=(2, 2), pool_type='max')
-
-    tower_conv = ConvFactory(maxpool5a_3_3, 96, (1, 1))
-    tower_conv1_0 = ConvFactory(maxpool5a_3_3, 48, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 64, (5, 5), pad=(2, 2))
-
-    tower_conv2_0 = ConvFactory(maxpool5a_3_3, 64, (1, 1))
-    tower_conv2_1 = ConvFactory(tower_conv2_0, 96, (3, 3), pad=(1, 1))
-    tower_conv2_2 = ConvFactory(tower_conv2_1, 96, (3, 3), pad=(1, 1))
-
-    tower_pool3_0 = mx.symbol.Pooling(data=maxpool5a_3_3, kernel=(
-        3, 3), stride=(1, 1), pad=(1, 1), pool_type='avg')
-    tower_conv3_1 = ConvFactory(tower_pool3_0, 64, (1, 1))
-    tower_5b_out = mx.symbol.Concat(
-        *[tower_conv, tower_conv1_1, tower_conv2_2, tower_conv3_1])
-    net = repeat(tower_5b_out, 10, block35, scale=0.17, input_num_channels=320)
-    tower_conv = ConvFactory(net, 384, (3, 3), stride=(2, 2))
-    tower_conv1_0 = ConvFactory(net, 256, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1_0, 256, (3, 3), pad=(1, 1))
-    tower_conv1_2 = ConvFactory(tower_conv1_1, 384, (3, 3), stride=(2, 2))
-    tower_pool = mx.symbol.Pooling(net, kernel=(
-        3, 3), stride=(2, 2), pool_type='max')
-    net = mx.symbol.Concat(*[tower_conv, tower_conv1_2, tower_pool])
-    net = repeat(net, 20, block17, scale=0.1, input_num_channels=1088)
-    tower_conv = ConvFactory(net, 256, (1, 1))
-    tower_conv0_1 = ConvFactory(tower_conv, 384, (3, 3), stride=(2, 2))
-    tower_conv1 = ConvFactory(net, 256, (1, 1))
-    tower_conv1_1 = ConvFactory(tower_conv1, 288, (3, 3), stride=(2, 2))
-    tower_conv2 = ConvFactory(net, 256, (1, 1))
-    tower_conv2_1 = ConvFactory(tower_conv2, 288, (3, 3), pad=(1, 1))
-    tower_conv2_2 = ConvFactory(tower_conv2_1, 320, (3, 3),  stride=(2, 2))
-    tower_pool = mx.symbol.Pooling(net, kernel=(
-        3, 3), stride=(2, 2), pool_type='max')
-    net = mx.symbol.Concat(
-        *[tower_conv0_1, tower_conv1_1, tower_conv2_2, tower_pool])
-
-    net = repeat(net, 9, block8, scale=0.2, input_num_channels=2080)
-    net = block8(net, with_act=False, input_num_channels=2080)
-
-    net = ConvFactory(data = net, num_filter=1536, kernel=(1, 1))
-    body = net
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
diff --git a/src/symbols/fmobilefacenet.py b/src/symbols/fmobilefacenet.py
deleted file mode 100644
index ee31fc95d..000000000
--- a/src/symbols/fmobilefacenet.py
+++ /dev/null
@@ -1,67 +0,0 @@
-
-import mxnet as mx
-import symbol_utils
-
-bn_mom = 0.9
-#bn_mom = 0.9997
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    #body = mx.sym.Activation(data=data, act_type='relu', name=name)
-    return body
-
-def Conv(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=bn_mom)
-    act = Act(data=bn, act_type='relu', name='%s%s_relu' %(name, suffix))
-    return act
-    
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=bn_mom)    
-    return bn
-
-def ConvOnly(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    return conv    
-
-    
-def DResidual(data, num_out=1, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=1, name=None, suffix=''):
-    conv = Conv(data=data, num_filter=num_group, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name='%s%s_conv_sep' %(name, suffix))
-    conv_dw = Conv(data=conv, num_filter=num_group, num_group=num_group, kernel=kernel, pad=pad, stride=stride, name='%s%s_conv_dw' %(name, suffix))
-    proj = Linear(data=conv_dw, num_filter=num_out, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name='%s%s_conv_proj' %(name, suffix))
-    return proj
-    
-def Residual(data, num_block=1, num_out=1, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=1, name=None, suffix=''):
-    identity=data
-    for i in range(num_block):
-    	shortcut=identity
-    	conv=DResidual(data=identity, num_out=num_out, kernel=kernel, stride=stride, pad=pad, num_group=num_group, name='%s%s_block' %(name, suffix), suffix='%d'%i)
-    	identity=conv+shortcut
-    return identity
-        
-
-def get_symbol(num_classes, **kwargs):
-    global bn_mom
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    wd_mult = kwargs.get('wd_mult', 1.)
-    version_output = kwargs.get('version_output', 'GNAP')
-    #assert version_output=='GDC' or version_output=='GNAP'
-    fc_type = version_output
-    data = mx.symbol.Variable(name="data")
-    data = data-127.5
-    data = data*0.0078125
-    conv_1 = Conv(data, num_filter=64, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_1")
-    conv_2_dw = Conv(conv_1, num_group=64, num_filter=64, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_2_dw")
-    conv_23 = DResidual(conv_2_dw, num_out=64, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=128, name="dconv_23")
-    conv_3 = Residual(conv_23, num_block=4, num_out=64, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=128, name="res_3")
-    conv_34 = DResidual(conv_3, num_out=128, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=256, name="dconv_34")
-    conv_4 = Residual(conv_34, num_block=6, num_out=128, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=256, name="res_4")
-    conv_45 = DResidual(conv_4, num_out=128, kernel=(3, 3), stride=(2, 2), pad=(1, 1), num_group=512, name="dconv_45")
-    conv_5 = Residual(conv_45, num_block=2, num_out=128, kernel=(3, 3), stride=(1, 1), pad=(1, 1), num_group=256, name="res_5")
-    conv_6_sep = Conv(conv_5, num_filter=512, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_6sep")
-
-    fc1 = symbol_utils.get_fc1(conv_6_sep, num_classes, fc_type)
-    return fc1
-
diff --git a/src/symbols/fmobilenet.py b/src/symbols/fmobilenet.py
deleted file mode 100644
index 5a2d183e5..000000000
--- a/src/symbols/fmobilenet.py
+++ /dev/null
@@ -1,84 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-import mxnet as mx
-import symbol_utils
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    #act = mx.sym.Activation(data=bn, act_type='relu', name='%s%s_relu' %(name, suffix))
-    return body
-
-def Conv(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=True)
-    act = Act(data=bn, act_type='relu', name='%s%s_relu' %(name, suffix))
-    return act
-
-def ConvOnly(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    return conv
-
-def get_symbol(num_classes, **kwargs):
-    data = mx.symbol.Variable(name="data") # 224
-    data = data-127.5
-    data = data*0.0078125
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    #version_unit = kwargs.get('version_unit', 3)
-    version_multiplier = kwargs.get('version_multiplier', 1.0)
-    bf = int(32*version_multiplier)
-    print(version_input, version_output, version_multiplier, bf)
-    if version_input==0:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_1") # 224/112
-    else:
-      conv_1 = Conv(data, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_1") # 224/112
-    conv_2_dw = Conv(conv_1, num_group=bf, num_filter=bf, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_2_dw") # 112/112
-    conv_2 = Conv(conv_2_dw, num_filter=bf*2, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_2") # 112/112
-    conv_3_dw = Conv(conv_2, num_group=bf*2, num_filter=bf*2, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_3_dw") # 112/56
-    conv_3 = Conv(conv_3_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_3") # 56/56
-    conv_4_dw = Conv(conv_3, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_4_dw") # 56/56
-    conv_4 = Conv(conv_4_dw, num_filter=bf*4, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_4") # 56/56
-    conv_5_dw = Conv(conv_4, num_group=bf*4, num_filter=bf*4, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_5_dw") # 56/28
-    conv_5 = Conv(conv_5_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_5") # 28/28
-    conv_6_dw = Conv(conv_5, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_6_dw") # 28/28
-    conv_6 = Conv(conv_6_dw, num_filter=bf*8, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_6") # 28/28
-    conv_7_dw = Conv(conv_6, num_group=bf*8, num_filter=bf*8, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_7_dw") # 28/14
-    conv_7 = Conv(conv_7_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_7") # 14/14
-
-    conv_8_dw = Conv(conv_7, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_8_dw") # 14/14
-    conv_8 = Conv(conv_8_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_8") # 14/14
-    conv_9_dw = Conv(conv_8, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_9_dw") # 14/14
-    conv_9 = Conv(conv_9_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_9") # 14/14
-    conv_10_dw = Conv(conv_9, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_10_dw") # 14/14
-    conv_10 = Conv(conv_10_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_10") # 14/14
-    conv_11_dw = Conv(conv_10, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_11_dw") # 14/14
-    conv_11 = Conv(conv_11_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_11") # 14/14
-    conv_12_dw = Conv(conv_11, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_12_dw") # 14/14
-    conv_12 = Conv(conv_12_dw, num_filter=bf*16, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_12") # 14/14
-
-    conv_13_dw = Conv(conv_12, num_group=bf*16, num_filter=bf*16, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_13_dw") # 14/7
-    conv_13 = Conv(conv_13_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_13") # 7/7
-    conv_14_dw = Conv(conv_13, num_group=bf*32, num_filter=bf*32, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_14_dw") # 7/7
-    conv_14 = Conv(conv_14_dw, num_filter=bf*32, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_14") # 7/7
-    body = conv_14
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
diff --git a/src/symbols/fmobilenetv2.py b/src/symbols/fmobilenetv2.py
deleted file mode 100644
index 11d8db0f8..000000000
--- a/src/symbols/fmobilenetv2.py
+++ /dev/null
@@ -1,106 +0,0 @@
-import mxnet as mx
-import mxnet.ndarray as nd
-import mxnet.gluon as gluon
-import mxnet.gluon.nn as nn
-import mxnet.autograd as ag
-import symbol_utils
-
-def ConvBlock(channels, kernel_size, strides):
-    out = nn.HybridSequential()
-    out.add(
-        nn.Conv2D(channels, kernel_size, strides=strides, padding=1, use_bias=False),
-        nn.BatchNorm(scale=True),
-        nn.Activation('relu')
-    )
-    return out
-
-def Conv1x1(channels, is_linear=False):
-    out = nn.HybridSequential()
-    out.add(
-        nn.Conv2D(channels, 1, padding=0, use_bias=False),
-        nn.BatchNorm(scale=True)
-    )
-    if not is_linear:
-        out.add(nn.Activation('relu'))
-    return out
-
-def DWise(channels, stride):
-    out = nn.HybridSequential()
-    out.add(
-        nn.Conv2D(channels, 3, strides=stride, padding=1, groups=channels, use_bias=False),
-        nn.BatchNorm(scale=True),
-        nn.Activation('relu')
-    )
-    return out
-
-class InvertedResidual(nn.HybridBlock):
-    def __init__(self, t, e, c, s, same_shape=True, **kwargs):
-        super(InvertedResidual, self).__init__(**kwargs)
-        self.same_shape = same_shape
-        self.stride = s
-        with self.name_scope(): 
-            self.bottleneck = nn.HybridSequential()
-            self.bottleneck.add(
-                Conv1x1(e*t),
-                DWise(e*t, self.stride),
-                Conv1x1(c, is_linear=True)
-            )
-            if self.stride == 1 and not self.same_shape:
-                self.conv_res = Conv1x1(c)
-    def hybrid_forward(self, F, x):
-        out = self.bottleneck(x)
-        #if self.stride == 1 and self.same_shape:
-        #    out = F.elemwise_add(out, x)
-        if self.stride == 1:
-            if not self.same_shape:
-                x = self.conv_res(x)
-            out = F.elemwise_add(out, x)
-        return out
-
-class MobilenetV2(nn.HybridBlock):
-    def __init__(self, num_classes=1000, width_mult=1.0, **kwargs):
-        super(MobilenetV2, self).__init__(**kwargs)
-        
-        self.w = width_mult
-        
-        self.cn = [int(x*self.w) for x in [32, 16, 24, 32, 64, 96, 160, 320]]
-        
-        def InvertedResidualSequence(t, cn_id, n, s):
-            seq = nn.HybridSequential()
-            seq.add(InvertedResidual(t, self.cn[cn_id-1], self.cn[cn_id], s, same_shape=False))
-            for _ in range(n-1):
-                seq.add(InvertedResidual(t, self.cn[cn_id-1], self.cn[cn_id], 1))
-            return seq
-        
-        self.b0 = ConvBlock(self.cn[0], 3, 1)
-        self.b1 = InvertedResidualSequence(1, 1, 1, 1)
-        self.b2 = InvertedResidualSequence(6, 2, 2, 2)
-        self.b3 = InvertedResidualSequence(6, 3, 3, 2)
-        self.b4 = InvertedResidualSequence(6, 4, 4, 1)
-        self.b5 = InvertedResidualSequence(6, 5, 3, 2)
-        self.b6 = InvertedResidualSequence(6, 6, 3, 2)
-        self.b7 = InvertedResidualSequence(6, 7, 1, 1)
-
-        self.last_channels = int(1280*self.w) if self.w > 1.0 else 1280
-        with self.name_scope():
-            self.features = nn.HybridSequential()
-            with self.features.name_scope():
-                self.features.add(self.b0, self.b1, self.b2, self.b3, self.b4, self.b5, self.b6, self.b7)
-                self.features.add(Conv1x1(self.last_channels))
-                #self.features.add(nn.GlobalAvgPool2D())
-                #self.features.add(nn.Flatten())
-            #self.output = nn.Dense(num_classes)
-    def hybrid_forward(self, F, x):
-        x = self.features(x)
-        #x = self.output(x)
-        return x
-
-def get_symbol(num_classes):
-  net = MobilenetV2(num_classes, 1)
-  data = mx.sym.Variable(name='data')
-  data = data-127.5
-  data = data*0.0078125
-  body = net(data)
-  fc1 = symbol_utils.get_fc1(body, num_classes, 'E')
-  return fc1
-
diff --git a/src/symbols/fnasnet.py b/src/symbols/fnasnet.py
deleted file mode 100644
index 713d964e5..000000000
--- a/src/symbols/fnasnet.py
+++ /dev/null
@@ -1,575 +0,0 @@
-
-from mxnet.gluon import nn
-import mxnet as mx
-import symbol_utils
-
-class MaxPoolPad(nn.HybridBlock):
-
-    def __init__(self):
-        super(MaxPoolPad, self).__init__()
-        self.pool = nn.MaxPool2D(pool_size=(3,3), strides=2, padding=1)
-
-    def hybrid_forward(self, F, x):
-        x = F.pad(x, mode="constant", constant_value=0, pad_width=(0, 0, 0, 0, 1, 0, 1, 0))
-        x = self.pool(x)
-        #x = x[:, :, 1:, 1:]
-        x = F.slice(x, begin=(None, None, 1, 1), end=(None, None, None, None))
-        return x
-
-class AvgPoolPad(nn.HybridBlock):
-
-    def __init__(self, strides=2, padding=1):
-        super(AvgPoolPad, self).__init__()
-        self.pool = nn.AvgPool2D(pool_size=(3,3), strides=strides, padding=padding)
-
-    def hybrid_forward(self, F, x):
-        x = F.pad(x, mode="constant", constant_value=0, pad_width=(0, 0, 0, 0, 1, 0, 1, 0))
-        x = self.pool(x)
-        #x = x[:, :, 1:, 1:]
-        x = F.slice(x, begin=(None, None, 1, 1), end=(None, None, None, None))
-        return x
-
-class SeparableConv2d(nn.HybridBlock):
-
-    def __init__(self, in_channels, out_channels, dw_kernel, dw_stride, dw_padding, bias=False):
-        super(SeparableConv2d, self).__init__()
-        self.depthwise_conv2d = nn.Conv2D(channels=in_channels, kernel_size=dw_kernel,
-                                          strides=dw_stride,
-                                          padding=dw_padding,
-                                          use_bias=bias,
-                                          groups=in_channels)
-        self.pointwise_conv2d = nn.Conv2D(channels=out_channels, kernel_size=1, strides=1, use_bias=bias)
-
-    def hybrid_forward(self, F, x):
-        x = self.depthwise_conv2d(x)
-        x = self.pointwise_conv2d(x)
-        return x
-
-class BranchSeparables(nn.HybridBlock):
-
-    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, bias=False):
-        super(BranchSeparables, self).__init__()
-        self.relu = nn.Activation(activation='relu')
-        self.separable_1 = SeparableConv2d(in_channels, in_channels, kernel_size, stride, padding, bias=bias)
-        self.bn_sep_1 = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-        self.relu1 = nn.Activation(activation='relu')
-        self.separable_2 = SeparableConv2d(in_channels, out_channels, kernel_size, 1, padding, bias=bias)
-        self.bn_sep_2 = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-
-    def hybrid_forward(self, F, x):
-        x = self.relu(x)
-        x = self.separable_1(x)
-        x = self.bn_sep_1(x)
-        x = self.relu1(x)
-        x = self.separable_2(x)
-        x = self.bn_sep_2(x)
-        return x
-
-class BranchSeparablesStem(nn.HybridBlock):
-
-    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, bias=False):
-        super(BranchSeparablesStem, self).__init__()
-        self.relu = nn.Activation(activation='relu')
-        self.separable_1 = SeparableConv2d(in_channels, out_channels, kernel_size, stride, padding, bias=bias)
-        self.bn_sep_1 = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-        self.relu1 = nn.Activation(activation='relu')
-        self.separable_2 = SeparableConv2d(out_channels, out_channels, kernel_size, 1, padding, bias=bias)
-        self.bn_sep_2 = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-
-    def hybrid_forward(self, F, x):
-        x = self.relu(x)
-        x = self.separable_1(x)
-        x = self.bn_sep_1(x)
-        x = self.relu1(x)
-        x = self.separable_2(x)
-        x = self.bn_sep_2(x)
-        return x
-
-class BranchSeparablesReduction(BranchSeparables):
-
-    z_padding = 1
-    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, z_padding=1, bias=False):
-        BranchSeparables.__init__(self, in_channels, out_channels, kernel_size, stride, padding, bias)
-        #self.padding = nn.ZeroPad2d((z_padding, 0, z_padding, 0))
-        self.z_padding = z_padding
-    def hybrid_forward(self, F, x):
-        x = self.relu(x)
-        x = F.pad(x, mode="constant", constant_value=0, pad_width=(0, 0, 0, 0, self.z_padding, 0, self.z_padding, 0))
-        x = self.separable_1(x)
-        #x = x[:, :, 1:, 1:]
-        x = F.slice(x, begin=(None, None, 1, 1), end=(None, None, None, None))
-        x = self.bn_sep_1(x)
-        x = self.relu1(x)
-        x = self.separable_2(x)
-        x = self.bn_sep_2(x)
-        return x
-
-class CellStem0(nn.HybridBlock):
-
-    def __init__(self):
-        super(CellStem0, self).__init__()
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation(activation='relu'))
-        self.conv_1x1.add(nn.Conv2D(42, 1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.comb_iter_0_left = BranchSeparables(42, 42, 5, 2, 2)
-        self.comb_iter_0_right = BranchSeparablesStem(96, 42, 7, 2, 3, bias=False)
-
-        self.comb_iter_1_left = nn.MaxPool2D(pool_size=3, strides=2, padding=1)
-        self.comb_iter_1_right = BranchSeparablesStem(96, 42, 7, 2, 3, bias=False)
-
-        self.comb_iter_2_left = nn.AvgPool2D(pool_size=3, strides=2, padding=1)
-        self.comb_iter_2_right = BranchSeparablesStem(96, 42, 5, 2, 2, bias=False)
-
-        self.comb_iter_3_right = nn.AvgPool2D(pool_size=3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparables(42, 42, 3, 1, 1, bias=False)
-        self.comb_iter_4_right = nn.MaxPool2D(pool_size=3, strides=2, padding=1)
-
-    def hybrid_forward(self, F, x):
-        x1 = self.conv_1x1(x)
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x1)
-        x_comb_iter_0_right = self.comb_iter_0_right(x)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x1)
-        x_comb_iter_1_right = self.comb_iter_1_right(x)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x1)
-        x_comb_iter_2_right = self.comb_iter_2_right(x)
-        x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right
-
-        x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0)
-        x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0)
-        x_comb_iter_4_right = self.comb_iter_4_right(x1)
-        x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right
-
-        x_out = F.concat(*[x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class CellStem1(nn.HybridBlock):
-
-    def __init__(self):
-        super(CellStem1, self).__init__()
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation('relu'))
-        self.conv_1x1.add(nn.Conv2D(channels=84, kernel_size=1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.relu = nn.Activation('relu')
-        self.path_1 = nn.HybridSequential()
-        self.path_1.add(nn.AvgPool2D(pool_size=1, strides=2))
-        self.path_1.add(nn.Conv2D(channels=42, kernel_size=1, strides=1, use_bias=False))
-
-        #self.path_2 = nn.ModuleList()
-        #self.path_2.add_module('pad', nn.ZeroPad2d((0, 1, 0, 1)))
-        #self.path_2.add_module('avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False))
-        self.path_2_avgpool = nn.AvgPool2D(pool_size=1, strides=2)
-        self.path_2_conv = nn.Conv2D(channels=42, kernel_size=1, strides=1, use_bias=False)
-
-        self.final_path_bn = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-
-        self.comb_iter_0_left = BranchSeparables(84, 84, 5, 2, 2, bias=False)
-        self.comb_iter_0_right = BranchSeparables(84, 84, 7, 2, 3, bias=False)
-
-        self.comb_iter_1_left = nn.MaxPool2D(pool_size=3, strides=2, padding=1)
-        self.comb_iter_1_right = BranchSeparables(84, 84, 7, 2, 3, bias=False)
-
-        self.comb_iter_2_left = nn.AvgPool2D(pool_size=3, strides=2, padding=1)
-        self.comb_iter_2_right = BranchSeparables(84, 84, 5, 2, 2, bias=False)
-
-        self.comb_iter_3_right = nn.AvgPool2D(pool_size=3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparables(84, 84, 3, 1, 1, bias=False)
-        self.comb_iter_4_right = nn.MaxPool2D(pool_size=3, strides=2, padding=1)
-
-    def hybrid_forward(self, F, x_conv0, x_stem_0):
-        x_left = self.conv_1x1(x_stem_0)
-
-        x_relu = self.relu(x_conv0)
-        # path 1
-        x_path1 = self.path_1(x_relu)
-        # path 2
-        x_path2 = F.pad(x_relu, mode="constant", constant_value=0, pad_width=(0, 0, 0, 0, 0, 1, 0, 1))
-        #x_path2 = x_path2[:, :, 1:, 1:]
-        x_path2 = F.slice(x_path2, begin=(None, None, 1, 1), end=(None, None, None, None))
-        x_path2 = self.path_2_avgpool(x_path2)
-        x_path2 = self.path_2_conv(x_path2)
-        # final path
-        x_right = self.final_path_bn(F.concat(*[x_path1, x_path2], dim=1))
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x_left)
-        x_comb_iter_0_right = self.comb_iter_0_right(x_right)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x_left)
-        x_comb_iter_1_right = self.comb_iter_1_right(x_right)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x_left)
-        x_comb_iter_2_right = self.comb_iter_2_right(x_right)
-        x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right
-
-        x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0)
-        x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0)
-        x_comb_iter_4_right = self.comb_iter_4_right(x_left)
-        x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right
-
-        x_out = F.concat(*[x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class FirstCell(nn.HybridBlock):
-
-    def __init__(self, in_channels_left, out_channels_left, in_channels_right, out_channels_right):
-        super(FirstCell, self).__init__()
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation(activation='relu'))
-        self.conv_1x1.add(nn.Conv2D(channels=out_channels_right, kernel_size=1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.relu = nn.Activation(activation='relu')
-        self.path_1 = nn.HybridSequential()
-        self.path_1.add(nn.AvgPool2D(pool_size=1, strides=2))
-        self.path_1.add(nn.Conv2D(channels=out_channels_left, kernel_size=1, strides=1, use_bias=False))
-        #self.path_2 = nn.ModuleList()
-        #self.path_2.add_module('pad', nn.ZeroPad2d((0, 1, 0, 1)))
-        #self.path_2.add_module('avgpool', nn.AvgPool2d(1, stride=2, count_include_pad=False))
-        self.path_2_avgpool = nn.AvgPool2D(pool_size=1, strides=2)
-        #self.path_2.add_module('conv', nn.Conv2d(in_channels_left, out_channels_left, 1, stride=1, bias=False))
-        self.path_2_conv = nn.Conv2D(channels=out_channels_left, kernel_size=1, strides=1, use_bias=False)
-        self.final_path_bn = nn.BatchNorm(epsilon=0.001, momentum=0.1)
-
-        self.comb_iter_0_left = BranchSeparables(out_channels_right, out_channels_right, 5, 1, 2, bias=False)
-        self.comb_iter_0_right = BranchSeparables(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-
-        self.comb_iter_1_left = BranchSeparables(out_channels_right, out_channels_right, 5, 1, 2, bias=False)
-        self.comb_iter_1_right = BranchSeparables(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-
-        self.comb_iter_2_left = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_3_left = nn.AvgPool2D(3, strides=1, padding=1)
-        self.comb_iter_3_right = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparables(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-
-    def hybrid_forward(self, F, x, x_prev):
-        x_relu = self.relu(x_prev)
-        # path 1
-        x_path1 = self.path_1(x_relu)
-        # path 2
-        x_path2 = F.pad(x_relu, mode="constant", constant_value=0, pad_width=(0, 0, 0, 0, 0, 1, 0, 1))
-        #x_path2 = x_path2[:, :, 1:, 1:]
-        x_path2 = F.slice(x_path2, begin=(None, None, 1, 1), end=(None, None, None, None))
-        x_path2 = self.path_2_avgpool(x_path2)
-        x_path2 = self.path_2_conv(x_path2)
-        # final path
-        x_left = self.final_path_bn(F.concat(*[x_path1, x_path2], dim=1))
-        x_right = self.conv_1x1(x)
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x_right)
-        x_comb_iter_0_right = self.comb_iter_0_right(x_left)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x_left)
-        x_comb_iter_1_right = self.comb_iter_1_right(x_left)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x_right)
-        x_comb_iter_2 = x_comb_iter_2_left + x_left
-
-        x_comb_iter_3_left = self.comb_iter_3_left(x_left)
-        x_comb_iter_3_right = self.comb_iter_3_right(x_left)
-        x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_right)
-        x_comb_iter_4 = x_comb_iter_4_left + x_right
-
-        x_out = F.concat(*[x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class NormalCell(nn.HybridBlock):
-
-    def __init__(self, in_channels_left, out_channels_left, in_channels_right, out_channels_right):
-        super(NormalCell, self).__init__()
-        self.conv_prev_1x1 = nn.HybridSequential()
-        self.conv_prev_1x1.add(nn.Activation(activation='relu'))
-        self.conv_prev_1x1.add(nn.Conv2D(channels=out_channels_left, kernel_size=1, strides=1, use_bias=False))
-        self.conv_prev_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation(activation='relu'))
-        self.conv_1x1.add(nn.Conv2D(channels=out_channels_right, kernel_size=1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.comb_iter_0_left = BranchSeparables(out_channels_right, out_channels_right, 5, 1, 2, bias=False)
-        self.comb_iter_0_right = BranchSeparables(out_channels_left, out_channels_left, 3, 1, 1, bias=False)
-
-        self.comb_iter_1_left = BranchSeparables(out_channels_left, out_channels_left, 5, 1, 2, bias=False)
-        self.comb_iter_1_right = BranchSeparables(out_channels_left, out_channels_left, 3, 1, 1, bias=False)
-
-        self.comb_iter_2_left = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_3_left = nn.AvgPool2D(3, strides=1, padding=1)
-        self.comb_iter_3_right = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparables(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-
-    def hybrid_forward(self, F, x, x_prev):
-        x_left = self.conv_prev_1x1(x_prev)
-        x_right = self.conv_1x1(x)
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x_right)
-        x_comb_iter_0_right = self.comb_iter_0_right(x_left)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x_left)
-        x_comb_iter_1_right = self.comb_iter_1_right(x_left)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x_right)
-        x_comb_iter_2 = x_comb_iter_2_left + x_left
-
-        x_comb_iter_3_left = self.comb_iter_3_left(x_left)
-        x_comb_iter_3_right = self.comb_iter_3_right(x_left)
-        x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_right)
-        x_comb_iter_4 = x_comb_iter_4_left + x_right
-
-        x_out = F.concat(*[x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class ReductionCell0(nn.HybridBlock):
-
-    def __init__(self, in_channels_left, out_channels_left, in_channels_right, out_channels_right):
-        super(ReductionCell0, self).__init__()
-        self.conv_prev_1x1 = nn.HybridSequential()
-        self.conv_prev_1x1.add(nn.Activation(activation='relu'))
-        self.conv_prev_1x1.add(nn.Conv2D(channels=out_channels_left, kernel_size=1, strides=1, use_bias=False))
-        self.conv_prev_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation(activation='relu'))
-        self.conv_1x1.add(nn.Conv2D(channels=out_channels_right, kernel_size=1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.comb_iter_0_left = BranchSeparablesReduction(out_channels_right, out_channels_right, 5, 2, 2, bias=False)
-        self.comb_iter_0_right = BranchSeparablesReduction(out_channels_right, out_channels_right, 7, 2, 3, bias=False)
-
-        self.comb_iter_1_left = MaxPoolPad()
-        self.comb_iter_1_right = BranchSeparablesReduction(out_channels_right, out_channels_right, 7, 2, 3, bias=False)
-
-        self.comb_iter_2_left = AvgPoolPad()
-        self.comb_iter_2_right = BranchSeparablesReduction(out_channels_right, out_channels_right, 5, 2, 2, bias=False)
-
-        self.comb_iter_3_right = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparablesReduction(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-        self.comb_iter_4_right = MaxPoolPad()
-
-    def hybrid_forward(self, F, x, x_prev):
-        x_left = self.conv_prev_1x1(x_prev)
-        x_right = self.conv_1x1(x)
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x_right)
-        x_comb_iter_0_right = self.comb_iter_0_right(x_left)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x_right)
-        x_comb_iter_1_right = self.comb_iter_1_right(x_left)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x_right)
-        x_comb_iter_2_right = self.comb_iter_2_right(x_left)
-        x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right
-
-        x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0)
-        x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0)
-        x_comb_iter_4_right = self.comb_iter_4_right(x_right)
-        x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right
-
-        x_out = F.concat(*[x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class ReductionCell1(nn.HybridBlock):
-
-    def __init__(self, in_channels_left, out_channels_left, in_channels_right, out_channels_right):
-        super(ReductionCell1, self).__init__()
-        self.conv_prev_1x1 = nn.HybridSequential()
-        self.conv_prev_1x1.add(nn.Activation(activation='relu'))
-        self.conv_prev_1x1.add(nn.Conv2D(channels=out_channels_left, kernel_size=1, strides=1, use_bias=False))
-        self.conv_prev_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.conv_1x1 = nn.HybridSequential()
-        self.conv_1x1.add(nn.Activation(activation='relu'))
-        self.conv_1x1.add(nn.Conv2D(channels=out_channels_right, kernel_size=1, strides=1, use_bias=False))
-        self.conv_1x1.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.comb_iter_0_left = BranchSeparables(out_channels_right, out_channels_right, 5, 2, 2, bias=False)
-        self.comb_iter_0_right = BranchSeparables(out_channels_right, out_channels_right, 7, 2, 3, bias=False)
-
-        self.comb_iter_1_left = nn.MaxPool2D(3, strides=2, padding=1)
-        self.comb_iter_1_right = BranchSeparables(out_channels_right, out_channels_right, 7, 2, 3, bias=False)
-
-        self.comb_iter_2_left = nn.AvgPool2D(3, strides=2, padding=1)
-        self.comb_iter_2_right = BranchSeparables(out_channels_right, out_channels_right, 5, 2, 2, bias=False)
-
-        self.comb_iter_3_right = nn.AvgPool2D(3, strides=1, padding=1)
-
-        self.comb_iter_4_left = BranchSeparables(out_channels_right, out_channels_right, 3, 1, 1, bias=False)
-        self.comb_iter_4_right = nn.MaxPool2D(3, strides=2, padding=1)
-
-    def hybrid_forward(self, F, x, x_prev):
-        x_left = self.conv_prev_1x1(x_prev)
-        x_right = self.conv_1x1(x)
-
-        x_comb_iter_0_left = self.comb_iter_0_left(x_right)
-        x_comb_iter_0_right = self.comb_iter_0_right(x_left)
-        x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
-
-        x_comb_iter_1_left = self.comb_iter_1_left(x_right)
-        x_comb_iter_1_right = self.comb_iter_1_right(x_left)
-        x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
-
-        x_comb_iter_2_left = self.comb_iter_2_left(x_right)
-        x_comb_iter_2_right = self.comb_iter_2_right(x_left)
-        x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right
-
-        x_comb_iter_3_right = self.comb_iter_3_right(x_comb_iter_0)
-        x_comb_iter_3 = x_comb_iter_3_right + x_comb_iter_1
-
-        x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0)
-        x_comb_iter_4_right = self.comb_iter_4_right(x_right)
-        x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right
-
-        x_out = F.concat(*[x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], dim=1)
-        return x_out
-
-class NASNetALarge(nn.HybridBlock):
-
-    def __init__(self, num_classes=1001):
-        super(NASNetALarge, self).__init__()
-        self.num_classes = num_classes
-
-        self.conv0 = nn.HybridSequential()
-        self.conv0.add(nn.Conv2D(channels=96, kernel_size=3, padding=0, strides=1, use_bias=False))
-        self.conv0.add(nn.BatchNorm(epsilon=0.001, momentum=0.1))
-
-        self.cell_stem_0 = CellStem0()
-        self.cell_stem_1 = CellStem1()
-
-        self.cell_0 = FirstCell(in_channels_left=168, out_channels_left=84,
-                                in_channels_right=336, out_channels_right=168)
-        self.cell_1 = NormalCell(in_channels_left=336, out_channels_left=168,
-                                 in_channels_right=1008, out_channels_right=168)
-        self.cell_2 = NormalCell(in_channels_left=1008, out_channels_left=168,
-                                 in_channels_right=1008, out_channels_right=168)
-        self.cell_3 = NormalCell(in_channels_left=1008, out_channels_left=168,
-                                 in_channels_right=1008, out_channels_right=168)
-        self.cell_4 = NormalCell(in_channels_left=1008, out_channels_left=168,
-                                 in_channels_right=1008, out_channels_right=168)
-        self.cell_5 = NormalCell(in_channels_left=1008, out_channels_left=168,
-                                 in_channels_right=1008, out_channels_right=168)
-
-        self.reduction_cell_0 = ReductionCell0(in_channels_left=1008, out_channels_left=336,
-                                               in_channels_right=1008, out_channels_right=336)
-
-        self.cell_6 = FirstCell(in_channels_left=1008, out_channels_left=168,
-                                in_channels_right=1344, out_channels_right=336)
-        self.cell_7 = NormalCell(in_channels_left=1344, out_channels_left=336,
-                                 in_channels_right=2016, out_channels_right=336)
-        self.cell_8 = NormalCell(in_channels_left=2016, out_channels_left=336,
-                                 in_channels_right=2016, out_channels_right=336)
-        self.cell_9 = NormalCell(in_channels_left=2016, out_channels_left=336,
-                                 in_channels_right=2016, out_channels_right=336)
-        self.cell_10 = NormalCell(in_channels_left=2016, out_channels_left=336,
-                                  in_channels_right=2016, out_channels_right=336)
-        self.cell_11 = NormalCell(in_channels_left=2016, out_channels_left=336,
-                                  in_channels_right=2016, out_channels_right=336)
-
-        self.reduction_cell_1 = ReductionCell1(in_channels_left=2016, out_channels_left=672,
-                                               in_channels_right=2016, out_channels_right=672)
-
-        self.cell_12 = FirstCell(in_channels_left=2016, out_channels_left=336,
-                                 in_channels_right=2688, out_channels_right=672)
-        self.cell_13 = NormalCell(in_channels_left=2688, out_channels_left=672,
-                                  in_channels_right=4032, out_channels_right=672)
-        self.cell_14 = NormalCell(in_channels_left=4032, out_channels_left=672,
-                                  in_channels_right=4032, out_channels_right=672)
-        self.cell_15 = NormalCell(in_channels_left=4032, out_channels_left=672,
-                                  in_channels_right=4032, out_channels_right=672)
-        self.cell_16 = NormalCell(in_channels_left=4032, out_channels_left=672,
-                                  in_channels_right=4032, out_channels_right=672)
-        self.cell_17 = NormalCell(in_channels_left=4032, out_channels_left=672,
-                                  in_channels_right=4032, out_channels_right=672)
-
-        self.relu = nn.Activation(activation='relu')
-        self.avgpool = nn.AvgPool2D(pool_size=11, strides=1, padding=0)
-        self.flatten = nn.Flatten()
-        self.dropout = nn.Dropout(0.5)
-        self.dense= nn.Dense(num_classes)
-
-    def features(self, x):
-        x_conv0 = self.conv0(x)
-        x_stem_0 = self.cell_stem_0(x_conv0)
-        x_stem_1 = self.cell_stem_1(x_conv0, x_stem_0)
-
-        x_cell_0 = self.cell_0(x_stem_1, x_stem_0)
-        x_cell_1 = self.cell_1(x_cell_0, x_stem_1)
-        x_cell_2 = self.cell_2(x_cell_1, x_cell_0)
-        x_cell_3 = self.cell_3(x_cell_2, x_cell_1)
-        x_cell_4 = self.cell_4(x_cell_3, x_cell_2)
-        x_cell_5 = self.cell_5(x_cell_4, x_cell_3)
-
-        x_reduction_cell_0 = self.reduction_cell_0(x_cell_5, x_cell_4)
-
-        x_cell_6 = self.cell_6(x_reduction_cell_0, x_cell_4)
-        x_cell_7 = self.cell_7(x_cell_6, x_reduction_cell_0)
-        x_cell_8 = self.cell_8(x_cell_7, x_cell_6)
-        x_cell_9 = self.cell_9(x_cell_8, x_cell_7)
-        x_cell_10 = self.cell_10(x_cell_9, x_cell_8)
-        x_cell_11 = self.cell_11(x_cell_10, x_cell_9)
-
-        x_reduction_cell_1 = self.reduction_cell_1(x_cell_11, x_cell_10)
-
-        x_cell_12 = self.cell_12(x_reduction_cell_1, x_cell_10)
-        x_cell_13 = self.cell_13(x_cell_12, x_reduction_cell_1)
-        x_cell_14 = self.cell_14(x_cell_13, x_cell_12)
-        x_cell_15 = self.cell_15(x_cell_14, x_cell_13)
-        x_cell_16 = self.cell_16(x_cell_15, x_cell_14)
-        x_cell_17 = self.cell_17(x_cell_16, x_cell_15)
-
-        return x_cell_17
-
-    def classifier(self, x):
-        x = self.relu(x)
-        x = self.avgpool(x)
-        x = self.flatten(x)
-        x = self.dropout(x)
-        x = self.dense(x)
-        return x
-
-    def hybrid_forward(self, F, x):
-        x = self.features(x)
-        x = self.classifier(x)
-        return x
-
-
-def get_symbol(num_classes):
-  model = NASNetALarge(num_classes)
-  data = mx.sym.Variable(name='data')
-  data = data-127.5
-  data = data*0.0078125
-  body = model.features(data)
-  fc1 = symbol_utils.get_fc1(body, num_classes, 'E')
-  return fc1
-
diff --git a/src/symbols/fresnet.py b/src/symbols/fresnet.py
deleted file mode 100644
index 4a7fd974d..000000000
--- a/src/symbols/fresnet.py
+++ /dev/null
@@ -1,601 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-'''
-Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
-Original author Wei Wu
-
-Implemented the following paper:
-
-Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Identity Mappings in Deep Residual Networks"
-'''
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import mxnet as mx
-import numpy as np
-import symbol_utils
-import sklearn
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-
-def Act(data, act_type, name):
-    if act_type=='prelu':
-      body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    else:
-      body = mx.symbol.Activation(data=data, act_type=act_type, name=name)
-    return body
-
-def residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=stride, pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit1')
-    if bottle_neck:
-        conv1 = Conv(data=data, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn3 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn3 + shortcut, act_type=act_type, name=name + '_relu3')
-    else:
-        conv1 = Conv(data=data, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn1 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn2 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn2 = mx.symbol.broadcast_mul(bn2, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return Act(data=bn2 + shortcut, act_type=act_type, name=name + '_relu3')
-
-def residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit2')
-    if bottle_neck:
-        # the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=int(num_filter*0.25), kernel=(3,3), stride=stride, pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act3 = Act(data=bn3, act_type=act_type, name=name + '_relu3')
-        conv3 = Conv(data=act3, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv3 = mx.symbol.broadcast_mul(conv3, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv3 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn1')
-        act1 = Act(data=bn1, act_type=act_type, name=name + '_relu1')
-        conv1 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_bn2')
-        act2 = Act(data=bn2, act_type=act_type, name=name + '_relu2')
-        conv2 = Conv(data=act2, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=conv2, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          conv2 = mx.symbol.broadcast_mul(conv2, body)
-        if dim_match:
-            shortcut = data
-        else:
-            shortcut = Conv(data=act1, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return conv2 + shortcut
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    #print('in unit3')
-    if bottle_neck:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=int(num_filter*0.25), kernel=(1,1), stride=(1,1), pad=(0,0),
-                                   no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=int(num_filter*0.25), kernel=(3,3), stride=(1,1), pad=(1,1),
-                                   no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-        conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                                   workspace=workspace, name=name + '_conv3')
-        bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn4 = mx.symbol.broadcast_mul(bn4, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn4 + shortcut
-    else:
-        bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-        conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv1')
-        bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-        act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-        conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                      no_bias=True, workspace=workspace, name=name + '_conv2')
-        bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-        if use_se:
-          #se begin
-          body = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-          body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv1", workspace=workspace)
-          body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-          body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                    name=name+"_se_conv2", workspace=workspace)
-          body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-          bn3 = mx.symbol.broadcast_mul(bn3, body)
-          #se end
-
-        if dim_match:
-            shortcut = data
-        else:
-            conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                            workspace=workspace, name=name+'_conv1sc')
-            shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-        if memonger:
-            shortcut._set_attr(mirror_stage='True')
-        return bn3 + shortcut
-
-def residual_unit_v3_x(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-    
-    """Return ResNeXt Unit symbol for building ResNeXt
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    assert(bottle_neck)
-    use_se = kwargs.get('version_se', 1)
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    act_type = kwargs.get('version_act', 'prelu')
-    num_group = 32
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(1,1), stride=(1,1), pad=(0,0),
-                               no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type=act_type, name=name + '_relu1')
-    conv2 = Conv(data=act1, num_group=num_group, num_filter=int(num_filter*0.5), kernel=(3,3), stride=(1,1), pad=(1,1),
-                               no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-    act2 = Act(data=bn3, act_type=act_type, name=name + '_relu2')
-    conv3 = Conv(data=act2, num_filter=num_filter, kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
-                               workspace=workspace, name=name + '_conv3')
-    bn4 = mx.sym.BatchNorm(data=conv3, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn4')
-
-    if use_se:
-      #se begin
-      body = mx.sym.Pooling(data=bn4, global_pool=True, kernel=(7, 7), pool_type='avg', name=name+'_se_pool1')
-      body = Conv(data=body, num_filter=num_filter//16, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv1", workspace=workspace)
-      body = Act(data=body, act_type=act_type, name=name+'_se_relu1')
-      body = Conv(data=body, num_filter=num_filter, kernel=(1,1), stride=(1,1), pad=(0,0),
-                                name=name+"_se_conv2", workspace=workspace)
-      body = mx.symbol.Activation(data=body, act_type='sigmoid', name=name+"_se_sigmoid")
-      bn4 = mx.symbol.broadcast_mul(bn4, body)
-      #se end
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn4 + shortcut
-
-
-def residual_unit(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs):
-  uv = kwargs.get('version_unit', 3)
-  version_input = kwargs.get('version_input', 1)
-  if uv==1:
-    if version_input==0:
-      return residual_unit_v1(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-    else:
-      return residual_unit_v1_L(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==2:
-    return residual_unit_v2(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  elif uv==4:
-    return residual_unit_v4(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-  else:
-    return residual_unit_v3(data, num_filter, stride, dim_match, name, bottle_neck, **kwargs)
-
-def resnet(units, num_stages, filter_list, num_classes, bottle_neck, **kwargs):
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    """Return ResNet symbol of
-    Parameters
-    ----------
-    units : list
-        Number of units in each stage
-    num_stages : int
-        Number of stage
-    filter_list : list
-        Channel size of each stage
-    num_classes : int
-        Ouput size of symbol
-    dataset : str
-        Dataset type, only cifar10 and imagenet supports
-    workspace : int
-        Workspace used in convolution operator
-    """
-    version_se = kwargs.get('version_se', 1)
-    version_input = kwargs.get('version_input', 1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    act_type = kwargs.get('version_act', 'prelu')
-    print(version_se, version_input, version_output, version_unit, act_type)
-    num_unit = len(units)
-    assert(num_unit == num_stages)
-    data = mx.sym.Variable(name='data')
-    if version_input==0:
-      #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-      #body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    elif version_input==2:
-      data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-      body = Conv(data=data, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1,1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-    else:
-      data = mx.sym.identity(data=data, name='id')
-      data = data-127.5
-      data = data*0.0078125
-      body = data
-      body = Conv(data=body, num_filter=filter_list[0], kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type=act_type, name='relu0')
-
-    for i in range(num_stages):
-      #if version_input==0:
-      #  body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False,
-      #                       name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      #else:
-      #  body = residual_unit(body, filter_list[i+1], (2, 2), False,
-      #    name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      body = residual_unit(body, filter_list[i+1], (2, 2), False,
-        name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, **kwargs)
-      for j in range(units[i]-1):
-        body = residual_unit(body, filter_list[i+1], (1,1), True, name='stage%d_unit%d' % (i+1, j+2),
-          bottle_neck=bottle_neck, **kwargs)
-
-    fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)
-    return fc1
-
-def get_symbol(num_classes, num_layers, **kwargs):
-    """
-    Adapted from https://github.com/tornadomeet/ResNet/blob/master/train_resnet.py
-    Original author Wei Wu
-    """
-    if num_layers >= 101:
-        filter_list = [64, 256, 512, 1024, 2048]
-        bottle_neck = True
-    else:
-        filter_list = [64, 64, 128, 256, 512]
-        bottle_neck = False
-    num_stages = 4
-    if num_layers == 18:
-        units = [2, 2, 2, 2]
-    elif num_layers == 34:
-        units = [3, 4, 6, 3]
-    elif num_layers == 49:
-        units = [3, 4, 14, 3]
-    elif num_layers == 50:
-        units = [3, 4, 14, 3]
-    elif num_layers == 74:
-        units = [3, 6, 24, 3]
-    elif num_layers == 90:
-        units = [3, 8, 30, 3]
-    elif num_layers == 100:
-        units = [3, 13, 30, 3]
-    elif num_layers == 124:
-        units = [3, 13, 40, 5]
-    elif num_layers == 101:
-        units = [3, 4, 23, 3]
-    elif num_layers == 152:
-        units = [3, 8, 36, 3]
-    elif num_layers == 200:
-        units = [3, 24, 36, 3]
-    elif num_layers == 269:
-        units = [3, 30, 48, 8]
-    else:
-        raise ValueError("no experiments done on num_layers {}, you can do it yourself".format(num_layers))
-
-    return resnet(units       = units,
-                  num_stages  = num_stages,
-                  filter_list = filter_list,
-                  num_classes = num_classes,
-                  bottle_neck = bottle_neck,
-                  **kwargs)
-
diff --git a/src/symbols/fxception.py b/src/symbols/fxception.py
deleted file mode 100644
index 83be4c124..000000000
--- a/src/symbols/fxception.py
+++ /dev/null
@@ -1,151 +0,0 @@
-# -*- coding: utf-8 -*-
-"""
-
-Xception network, suitable for images with around 299 x 299 (original version)
-
-Reference:
-
-François Chollet. Xception: Deep Learning with Depthwise Separable Convlutions. arXiv preprint. https://arxiv.org/pdf/1610.02357v3.pdf
-
-I refered one version of MXNet from u1234x1234 https://github.com/u1234x1234/mxnet-xception/blob/master/symbol_xception.py
-
-Modified by Lin Xiong, Sep-3, 2017 for images 224 x 224
-There are some slightly differences with u1234x1234's version (pooling layer) and original version (no dropout layer).
-
-In order to accelerate computation, we use smaller parameters than original paper.
-
-"""
-
-import mxnet as mx
-import symbol_utils
-
-def Conv(data, num_filter, kernel=(1, 1), stride=(1, 1), pad=(0, 0), name=None, suffix='', withRelu=False, withBn=True, bn_mom=0.9, workspace=256):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad,
-                              name='%s%s_conv2d' % (name, suffix), workspace=workspace)
-    if withBn:
-        conv = mx.sym.BatchNorm(data=conv, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='%s%s_bn' % (name, suffix))
-    if withRelu:
-        conv = mx.sym.Activation(data=conv, act_type='relu', name='%s%s_relu' % (name, suffix))
-    return conv
-
-def Separable_Conv(data, num_in_channel, num_out_channel, kernel=(3, 3), stride=(1, 1), pad=(1, 1), name=None, suffix='', depth_mult=1, withBn=True, bn_mom=0.9, workspace=256):
-    # original version of Separable Convolution
-    # depthwise convolution
-    #channels       = mx.sym.split(data=data, axis=1, num_outputs=num_in_channel) # for new version of mxnet > 0.8
-    channels       = mx.sym.SliceChannel(data=data, axis=1, num_outputs=num_in_channel) # for old version of mxnet <= 0.8
-    depthwise_outs = [mx.sym.Convolution(data=channels[i], num_filter=depth_mult, kernel=kernel, 
-                           stride=stride, pad=pad, name=name+'_depthwise_kernel_'+str(i), workspace=workspace)
-                           for i in range(num_in_channel)]
-    depthwise_out = mx.sym.Concat(*depthwise_outs)
-    # pointwise convolution
-    pointwise_out = Conv(data=depthwise_out, num_filter=num_out_channel, name=name+'_pointwise_kernel', withBn=False, bn_mom=0.9, workspace=256)
-    if withBn:
-        pointwise_out = mx.sym.BatchNorm(data=pointwise_out, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='%s%s_bn' % (name, suffix))
-    return pointwise_out
-
-
-
-def Circle_Middle(name, data, 
-                  num_filter,
-                  bn_mom=0.9,
-                  round=8):
-    b = data
-    for i in xrange(round):
-        residual = b
-        prefix = name + '_block' + ('_%d' % i)
-
-        b = mx.sym.Activation(data=b, act_type='relu', name=prefix + '_sepconv1_relu')
-        b = Separable_Conv(data=b, num_in_channel=num_filter, num_out_channel=num_filter, name=prefix + '_sepconv1', withBn=True, bn_mom=bn_mom, workspace=256)
-        b = mx.sym.Activation(data=b, act_type='relu', name=prefix + '_sepconv2_relu')
-        b = Separable_Conv(data=b, num_in_channel=num_filter, num_out_channel=num_filter, name=prefix + '_sepconv2', withBn=True, bn_mom=bn_mom, workspace=256)
-        b = mx.sym.Activation(data=b, act_type='relu', name=prefix + '_sepconv3_relu')
-        b = Separable_Conv(data=b, num_in_channel=num_filter, num_out_channel=num_filter, name=prefix + '_sepconv3', withBn=True, bn_mom=bn_mom, workspace=256)
-
-        b = b + residual
-
-    return b
-
-
-def get_symbol(num_classes=1000, **kwargs):
-    # input shape 229*229*3 (old)
-    # input shape 224*224*3 (new)
-    
-    #filter_list=[64, 128, 256, 728, 1024, 1536, 2048]     # original version
-    filter_list=[64, 64, 128, 364, 512, 768, 1024]  # smaller one
-
-    # Entry flow
-    data = mx.sym.Variable('data')
-    data = data-127.5
-    data = data*0.0078125
-    version_input = kwargs.get('version_input',1)
-    assert version_input>=0
-    version_output = kwargs.get('version_output', 'E')
-    fc_type = version_output
-    version_unit = kwargs.get('version_unit', 3)
-    print(version_input, version_output, version_unit)
-
-    if version_input>=2:
-      filter_list=[64, 128, 256, 728, 1024, 1536, 2048]     # original version
-
-    # block 1
-    if version_input==0:
-      block1 = Conv(data=data, num_filter=int(filter_list[0]*0.5), kernel=(3, 3), stride=(2, 2), pad=(1, 1), name='Entry_flow_b1_conv1', 
-                    withRelu=True, withBn=True, bn_mom=0.9, workspace=256)
-    else:
-      block1 = Conv(data=data, num_filter=int(filter_list[0]*0.5), kernel=(3, 3), stride=(1,1), pad=(1, 1), name='Entry_flow_b1_conv1', 
-                    withRelu=True, withBn=True, bn_mom=0.9, workspace=256)
-    block1 = Conv(data=block1, num_filter=filter_list[0], kernel=(3, 3), pad=(1, 1), name='Entry_flow_b1_conv2',
-                  withRelu=True, withBn=True, bn_mom=0.9, workspace=256)
-
-    # block 2
-    rs2    = Conv(data=block1, num_filter=filter_list[1], stride=(2, 2), name='Entry_flow_b2_conv1',
-                  withBn=True, bn_mom=0.9, workspace=256)
-    block2 = Separable_Conv(block1, num_in_channel=filter_list[0], num_out_channel=filter_list[1], name='Entry_flow_b2_sepconv1', withBn=True, bn_mom=0.9, workspace=256)
-    block2 = mx.sym.Activation(data=block2, act_type='relu', name='Entry_flow_b2_sepconv1_relu')
-    block2 = Separable_Conv(block2, num_in_channel=filter_list[1], num_out_channel=filter_list[1], name='Entry_flow_b2_sepconv2', withBn=True, bn_mom=0.9, workspace=256)
-    block2 = mx.sym.Pooling(data=block2, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max', name='Entry_flow_b2_pool')
-    block2 = block2 + rs2
-
-    # block 3
-    rs3    = Conv(data=block2, num_filter=filter_list[2], stride=(2, 2), name='Entry_flow_b3_conv1',
-                  withBn=True, bn_mom=0.9, workspace=256)
-    block3 = mx.sym.Activation(data=block2, act_type='relu', name='Entry_flow_b3_sepconv1_relu')
-    block3 = Separable_Conv(block3, num_in_channel=filter_list[1], num_out_channel=filter_list[2], name='Entry_flow_b3_sepconv1', withBn=True, bn_mom=0.9, workspace=256)
-    block3 = mx.sym.Activation(data=block3, act_type='relu', name='Entry_flow_b3_sepconv2_relu')
-    block3 = Separable_Conv(block3, num_in_channel=filter_list[2], num_out_channel=filter_list[2], name='Entry_flow_b3_sepconv2', withBn=True, bn_mom=0.9, workspace=256)
-    block3 = mx.sym.Pooling(data=block3, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max', name='Entry_flow_b3_pool')
-    block3 = block3 + rs3
-
-    # block 4
-    rs4    = Conv(data=block3, num_filter=filter_list[3], stride=(2, 2), name='Entry_flow_b4_conv1',
-                  withBn=True, bn_mom=0.9, workspace=256)
-    block4 = mx.sym.Activation(data=block3, act_type='relu', name='Entry_flow_b4_sepconv1_relu')
-    block4 = Separable_Conv(block4, num_in_channel=filter_list[2], num_out_channel=filter_list[3], name='Entry_flow_b4_sepconv1', withBn=True, bn_mom=0.9, workspace=256)
-    block4 = mx.sym.Activation(data=block4, act_type='relu', name='Entry_flow_b4_sepconv2_relu')
-    block4 = Separable_Conv(block4, num_in_channel=filter_list[3], num_out_channel=filter_list[3], name='Entry_flow_b4_sepconv2', withBn=True, bn_mom=0.9, workspace=256)
-    block4 = mx.sym.Pooling(data=block4, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max', name='Entry_flow_b4_pool')
-    block4 = block4 + rs4
-
-    # Middle flow
-    block_m_f = Circle_Middle('Middle_flow', block4,
-                      filter_list[3],
-                      0.9,
-                      8)
-    # Exit flow
-    rs5    = Conv(data=block_m_f, num_filter=filter_list[4], stride=(2, 2), name='Exit_flow_b5_conv1',
-                  withBn=True, bn_mom=0.9, workspace=256)
-    block5 = mx.sym.Activation(data=block_m_f, act_type='relu', name='Exit_flow_b5_sepconv1_relu')
-    block5 = Separable_Conv(block5, num_in_channel=filter_list[3], num_out_channel=filter_list[3], name='Exit_flow_b5_sepconv1', withBn=True, bn_mom=0.9, workspace=256)
-    block5 = mx.sym.Activation(data=block5, act_type='relu', name='Exit_flow_b5_sepconv2_relu')
-    block5 = Separable_Conv(block5, num_in_channel=filter_list[3], num_out_channel=filter_list[4], name='Exit_flow_b5_sepconv2', withBn=True, bn_mom=0.9, workspace=256)
-    block5 = mx.sym.Pooling(data=block5, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max', name='Entry_flow_b5_pool')
-    block5 = block5 + rs5
-
-    block6 = Separable_Conv(block5, num_in_channel=filter_list[4], num_out_channel=filter_list[5], name='Exit_flow_b6_sepconv1', withBn=True, bn_mom=0.9, workspace=256)
-    block6 = mx.sym.Activation(data=block6, act_type='relu', name='Exit_flow_b6_sepconv1_relu')
-    block6 = Separable_Conv(block6, num_in_channel=filter_list[5], num_out_channel=filter_list[6], name='Exit_flow_b6_sepconv2', withBn=True, bn_mom=0.9, workspace=256)
-    block6 = mx.sym.Activation(data=block6, act_type='relu', name='Exit_flow_b6_sepconv2_relu')
-    fc1 = symbol_utils.get_fc1(block6, num_classes, fc_type)
-    return fc1
-
-
diff --git a/src/symbols/spherenet.py b/src/symbols/spherenet.py
deleted file mode 100644
index 29e1fc7dd..000000000
--- a/src/symbols/spherenet.py
+++ /dev/null
@@ -1,62 +0,0 @@
-import mxnet as mx
-import numpy as np
-import math
-from mxnet.base import _Null
-
-def conv_main(data, units, filters, workspace):
-  body = data
-  for i in xrange(len(units)):
-    f = filters[i]
-    _weight = mx.symbol.Variable("conv%d_%d_weight"%(i+1, 1), init=mx.init.Normal(0.01))
-    _bias = mx.symbol.Variable("conv%d_%d_bias"%(i+1, 1), lr_mult=2.0, wd_mult=0.0, init=mx.init.Constant(0.0))
-    body = mx.sym.Convolution(data=body, weight = _weight, bias = _bias, num_filter=f, kernel=(3, 3), stride=(2,2), pad=(1, 1),
-                              name= "conv%d_%d"%(i+1, 1), workspace=workspace)
-
-    body = mx.sym.LeakyReLU(data = body, act_type='prelu', name = "relu%d_%d" % (i+1, 1))
-    idx = 2
-    for j in xrange(units[i]):
-      _weight = mx.symbol.Variable("conv%d_%d_weight"%(i+1, idx), init=mx.init.Normal(0.01))
-      _body = mx.sym.Convolution(data=body, weight=_weight, no_bias=True, num_filter=f, kernel=(3, 3), stride=(1,1), pad=(1, 1),
-                                name= "conv%d_%d"%(i+1, idx), workspace=workspace)
-
-      _body = mx.sym.LeakyReLU(data = _body, act_type='prelu', name = "relu%d_%d" % (i+1, idx))
-      idx+=1
-      _weight = mx.symbol.Variable("conv%d_%d_weight"%(i+1, idx), init=mx.init.Normal(0.01))
-      _body = mx.sym.Convolution(data=_body, weight=_weight, no_bias=True, num_filter=f, kernel=(3, 3), stride=(1,1), pad=(1, 1),
-                                name= "conv%d_%d"%(i+1, idx), workspace=workspace)
-      _body = mx.sym.LeakyReLU(data = _body, act_type='prelu', name = "relu%d_%d" % (i+1, idx))
-      idx+=1
-      body = body+_body
-
-  return body
-
-def get_symbol(num_classes, num_layers, conv_workspace=256, **kwargs):
-  if num_layers==64:
-    units = [3,8,16,3]
-    filters = [64,128,256,512]
-  elif num_layers==20:
-    units = [1,2,4,1]
-    filters = [64,128,256,512]
-    #filters = [64, 256, 512, 1024]
-  elif num_layers==36:
-    units = [2,4,8,2]
-    filters = [64,128,256,512]
-    #filters = [64, 256, 512, 1024]
-  elif num_layers==60:
-    units = [3,8,14,3]
-    filters = [64,128,256,512]
-  elif num_layers==104:
-    units = [3,8,36,3]
-    filters = [64,128,256,512]
-    #filters = [64, 256, 512, 1024]
-  data = mx.symbol.Variable('data')
-  data = data-127.5
-  data = data*0.0078125
-  body = conv_main(data = data, units = units, filters = filters, workspace = conv_workspace)
-
-  _weight = mx.symbol.Variable("fc1_weight", lr_mult=1.0)
-  _bias = mx.symbol.Variable("fc1_bias", lr_mult=2.0, wd_mult=0.0)
-  fc1 = mx.sym.FullyConnected(data=body, weight=_weight, bias=_bias, num_hidden=num_classes, name='fc1')
-  return fc1
-  
-
diff --git a/src/symbols/symbol_utils.py b/src/symbols/symbol_utils.py
deleted file mode 100644
index e23b01bbc..000000000
--- a/src/symbols/symbol_utils.py
+++ /dev/null
@@ -1,182 +0,0 @@
-import mxnet as mx
-
-def Conv(**kwargs):
-    #name = kwargs.get('name')
-    #_weight = mx.symbol.Variable(name+'_weight')
-    #_bias = mx.symbol.Variable(name+'_bias', lr_mult=2.0, wd_mult=0.0)
-    #body = mx.sym.Convolution(weight = _weight, bias = _bias, **kwargs)
-    body = mx.sym.Convolution(**kwargs)
-    return body
-
-def Act(data, act_type, name):
-    #ignore param act_type, set it in this function 
-    body = mx.sym.LeakyReLU(data = data, act_type='prelu', name = name)
-    return body
-
-bn_mom = 0.9
-def Linear(data, num_filter=1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_group=1, name=None, suffix=''):
-    conv = mx.sym.Convolution(data=data, num_filter=num_filter, kernel=kernel, num_group=num_group, stride=stride, pad=pad, no_bias=True, name='%s%s_conv2d' %(name, suffix))
-    bn = mx.sym.BatchNorm(data=conv, name='%s%s_batchnorm' %(name, suffix), fix_gamma=False,momentum=bn_mom)    
-    return bn
-
-def get_fc1(last_conv, num_classes, fc_type):
-  bn_mom = 0.9
-  body = last_conv
-  if fc_type=='Z':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = body
-  elif fc_type=='E':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GAP':
-    bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    relu1 = Act(data=bn1, act_type='relu', name='relu1')
-    # Although kernel is not used here when global_pool=True, we should put one
-    pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
-    flat = mx.sym.Flatten(data=pool1)
-    fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='GNAP': #mobilefacenet++
-    filters_in = 512 # param in mobilefacenet
-    if num_classes>filters_in:
-      body = mx.sym.Convolution(data=last_conv, num_filter=num_classes, kernel=(1,1), stride=(1,1), pad=(0,0), no_bias=True, name='convx')
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=0.9, name='convx_bn')
-      body = Act(data=body, act_type='relu', name='convx_relu')
-      filters_in = num_classes
-    else:
-      body = last_conv
-    body = mx.sym.BatchNorm(data=body, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6f')  
-
-    spatial_norm=body*body
-    spatial_norm=mx.sym.sum(data=spatial_norm, axis=1, keepdims=True)
-    spatial_sqrt=mx.sym.sqrt(spatial_norm)
-    #spatial_mean=mx.sym.mean(spatial_sqrt, axis=(1,2,3), keepdims=True)
-    spatial_mean=mx.sym.mean(spatial_sqrt)
-    spatial_div_inverse=mx.sym.broadcast_div(spatial_mean, spatial_sqrt)
-
-    spatial_attention_inverse=mx.symbol.tile(spatial_div_inverse, reps=(1,filters_in,1,1))   
-    body=body*spatial_attention_inverse
-    #body = mx.sym.broadcast_mul(body, spatial_div_inverse)
-
-    fc1 = mx.sym.Pooling(body, kernel=(7, 7), global_pool=True, pool_type='avg')
-    if num_classes<filters_in:
-      fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='bn6w')
-      fc1 = mx.sym.FullyConnected(data=fc1, num_hidden=num_classes, name='pre_fc1')
-    else:
-      fc1 = mx.sym.Flatten(data=fc1)
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=0.9, name='fc1')
-  elif fc_type=="GDC": #mobilefacenet_v1
-    conv_6_dw = Linear(last_conv, num_filter=512, num_group=512, kernel=(7,7), pad=(0, 0), stride=(1, 1), name="conv_6dw7_7")  
-    conv_6_f = mx.sym.FullyConnected(data=conv_6_dw, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=conv_6_f, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='F':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    body = mx.symbol.Dropout(data=body, p=0.4)
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='G':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='H':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='fc1')
-  elif fc_type=='I':
-    body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  elif fc_type=='J':
-    fc1 = mx.sym.FullyConnected(data=body, num_hidden=num_classes, name='pre_fc1')
-    fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-  else:
-    bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
-    relu1 = Act(data=bn1, act_type='relu', name='relu1')
-    # Although kernel is not used here when global_pool=True, we should put one
-    pool1 = mx.sym.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
-    flat = mx.sym.Flatten(data=pool1)
-    if len(fc_type)>1:
-      if fc_type[1]=='X':
-        print('dropout mode')
-        flat = mx.symbol.Dropout(data=flat, p=0.2)
-      fc_type = fc_type[0]
-    if fc_type=='A':
-      fc1 = flat
-    else:
-      #B-D
-      #B
-      fc1 = mx.sym.FullyConnected(data=flat, num_hidden=num_classes, name='pre_fc1')
-      if fc_type=='C':
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-      elif fc_type=='D':
-        fc1 = mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='fc1')
-        fc1 = Act(data=fc1, act_type='relu', name='fc1_relu')
-  return fc1
-
-def residual_unit_v3(data, num_filter, stride, dim_match, name, **kwargs):
-    
-    """Return ResNet Unit symbol for building ResNet
-    Parameters
-    ----------
-    data : str
-        Input data
-    num_filter : int
-        Number of output channels
-    bnf : int
-        Bottle neck channels factor with regard to num_filter
-    stride : tuple
-        Stride used in convolution
-    dim_match : Boolean
-        True means channel number between input and output is the same, otherwise means differ
-    name : str
-        Base name of the operators
-    workspace : int
-        Workspace used in convolution operator
-    """
-    bn_mom = kwargs.get('bn_mom', 0.9)
-    workspace = kwargs.get('workspace', 256)
-    memonger = kwargs.get('memonger', False)
-    #print('in unit3')
-    bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
-    conv1 = Conv(data=bn1, num_filter=num_filter, kernel=(3,3), stride=(1,1), pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv1')
-    bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
-    act1 = Act(data=bn2, act_type='relu', name=name + '_relu1')
-    conv2 = Conv(data=act1, num_filter=num_filter, kernel=(3,3), stride=stride, pad=(1,1),
-                                  no_bias=True, workspace=workspace, name=name + '_conv2')
-    bn3 = mx.sym.BatchNorm(data=conv2, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn3')
-
-    if dim_match:
-        shortcut = data
-    else:
-        conv1sc = Conv(data=data, num_filter=num_filter, kernel=(1,1), stride=stride, no_bias=True,
-                                        workspace=workspace, name=name+'_conv1sc')
-        shortcut = mx.sym.BatchNorm(data=conv1sc, fix_gamma=False, momentum=bn_mom, eps=2e-5, name=name + '_sc')
-    if memonger:
-        shortcut._set_attr(mirror_stage='True')
-    return bn3 + shortcut
-
-
-def get_head(data, version_input, num_filter):
-    bn_mom = 0.9
-    workspace = 256
-    kwargs = {'bn_mom': bn_mom, 'workspace' : workspace}
-    data = data-127.5
-    data = data*0.0078125
-    #data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
-    if version_input==0:
-      body = Conv(data=data, num_filter=num_filter, kernel=(7, 7), stride=(2,2), pad=(3, 3),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type='relu', name='relu0')
-      body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
-    else:
-      body = data
-      _num_filter = min(num_filter, 64)
-      body = Conv(data=body, num_filter=_num_filter, kernel=(3,3), stride=(1,1), pad=(1, 1),
-                                no_bias=True, name="conv0", workspace=workspace)
-      body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
-      body = Act(data=body, act_type='relu', name='relu0')
-      body = residual_unit_v3(body, _num_filter, (2, 2), False, name='head', **kwargs)
-    return body
-
-
diff --git a/src/train.py b/src/train.py
deleted file mode 100644
index da87b2945..000000000
--- a/src/train.py
+++ /dev/null
@@ -1,1026 +0,0 @@
-# THIS FILE IS FOR EXPERIMENTS, USE train_softmax.py FOR NORMAL TRAINING.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import numpy as np
-from data import FaceImageIter
-from data import FaceImageIterList
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_image
-from noise_sgd import NoiseSGD
-sys.path.append(os.path.join(os.path.dirname(__file__), 'eval'))
-sys.path.append(os.path.join(os.path.dirname(__file__), 'symbols'))
-import fresnet
-import finception_resnet_v2
-import fmobilenet 
-import fmobilenetv2
-import fxception
-import fdensenet
-import fdpn
-import fnasnet
-import spherenet
-#import lfw
-import verification
-import sklearn
-sys.path.append(os.path.join(os.path.dirname(__file__), 'losses'))
-import center_loss
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-class AccMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(AccMetric, self).__init__(
-        'acc', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    if args.loss_type>=2 and args.loss_type<=7 and args.margin_verbose>0:
-      if self.count%args.ctx_num==0:
-        mbatch = self.count//args.ctx_num
-        _verbose = args.margin_verbose
-        if mbatch==1 or mbatch%_verbose==0:
-          a = 0.0
-          b = 0.0
-          if len(preds)>=4:
-            a = preds[-2].asnumpy()[0]
-            b = preds[-1].asnumpy()[0]
-          elif len(preds)==3:
-            a = preds[-1].asnumpy()[0]
-            b = a
-          print('[%d][MARGIN]%f,%f'%(mbatch,a,b))
-    if args.logits_verbose>0:
-      if self.count%args.ctx_num==0:
-        mbatch = self.count//args.ctx_num
-        _verbose = args.logits_verbose
-        if mbatch==1 or mbatch%_verbose==0:
-          a = 0.0
-          b = 0.0
-          if len(preds)>=3:
-            v = preds[-1].asnumpy()
-            v = np.sort(v)
-            num = len(v)//10
-            a = np.mean(v[0:num])
-            b = np.mean(v[-1*num:])
-          print('[LOGITS] %d,%f,%f'%(mbatch,a,b))
-    #loss = preds[2].asnumpy()[0]
-    #if len(self.losses)==20:
-    #  print('ce loss', sum(self.losses)/len(self.losses))
-    #  self.losses = []
-    #self.losses.append(loss)
-    preds = [preds[1]] #use softmax output
-    for label, pred_label in zip(labels, preds):
-        #print(pred_label)
-        #print(label.shape, pred_label.shape)
-        if pred_label.shape != label.shape:
-            pred_label = mx.ndarray.argmax(pred_label, axis=self.axis)
-        pred_label = pred_label.asnumpy().astype('int32').flatten()
-        label = label.asnumpy()
-        if label.ndim==2:
-          label = label[:,0]
-        label = label.astype('int32').flatten()
-        #print(label)
-        #print('label',label)
-        #print('pred_label', pred_label)
-        assert label.shape==pred_label.shape
-        self.sum_metric += (pred_label.flat == label.flat).sum()
-        self.num_inst += len(pred_label.flat)
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[-1].asnumpy()[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-    gt_label = preds[-2].asnumpy()
-    #print(gt_label)
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--data-dir', default='', help='training set directory')
-  parser.add_argument('--prefix', default='../model/model', help='directory to save model.')
-  parser.add_argument('--pretrained', default='', help='pretrained model to load')
-  parser.add_argument('--ckpt', type=int, default=1, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--network', default='r50', help='specify network')
-  parser.add_argument('--version-se', type=int, default=0, help='whether to use se in network')
-  parser.add_argument('--version-input', type=int, default=1, help='network input config')
-  parser.add_argument('--version-output', type=str, default='E', help='network embedding output config')
-  parser.add_argument('--version-unit', type=int, default=3, help='resnet unit config')
-  parser.add_argument('--version-act', type=str, default='prelu', help='network activation config')
-  parser.add_argument('--end-epoch', type=int, default=100000, help='training epoch size.')
-  parser.add_argument('--noise-sgd', type=float, default=0.0, help='')
-  parser.add_argument('--lr', type=float, default=0.1, help='start learning rate')
-  parser.add_argument('--wd', type=float, default=0.0005, help='weight decay')
-  parser.add_argument('--mom', type=float, default=0.9, help='momentum')
-  parser.add_argument('--emb-size', type=int, default=512, help='embedding length')
-  parser.add_argument('--per-batch-size', type=int, default=128, help='batch size in each context')
-  parser.add_argument('--margin-m', type=float, default=0.5, help='')
-  parser.add_argument('--margin-s', type=float, default=64.0, help='')
-  parser.add_argument('--margin-a', type=float, default=0.0, help='')
-  parser.add_argument('--margin-b', type=float, default=0.0, help='')
-  parser.add_argument('--easy-margin', type=int, default=0, help='')
-  parser.add_argument('--margin-verbose', type=int, default=0, help='')
-  parser.add_argument('--logits-verbose', type=int, default=0, help='')
-  parser.add_argument('--c2c-threshold', type=float, default=0.0, help='')
-  parser.add_argument('--c2c-mode', type=int, default=-10, help='')
-  parser.add_argument('--output-c2c', type=int, default=0, help='')
-  parser.add_argument('--train-limit', type=int, default=0, help='')
-  parser.add_argument('--margin', type=int, default=4, help='')
-  parser.add_argument('--beta', type=float, default=1000., help='')
-  parser.add_argument('--beta-min', type=float, default=5., help='')
-  parser.add_argument('--beta-freeze', type=int, default=0, help='')
-  parser.add_argument('--gamma', type=float, default=0.12, help='')
-  parser.add_argument('--power', type=float, default=1.0, help='')
-  parser.add_argument('--scale', type=float, default=0.9993, help='')
-  parser.add_argument('--center-alpha', type=float, default=0.5, help='')
-  parser.add_argument('--center-scale', type=float, default=0.003, help='')
-  parser.add_argument('--images-per-identity', type=int, default=0, help='')
-  parser.add_argument('--triplet-bag-size', type=int, default=3600, help='')
-  parser.add_argument('--triplet-alpha', type=float, default=0.3, help='')
-  parser.add_argument('--triplet-max-ap', type=float, default=0.0, help='')
-  parser.add_argument('--verbose', type=int, default=2000, help='')
-  parser.add_argument('--loss-type', type=int, default=4, help='')
-  parser.add_argument('--incay', type=float, default=0.0, help='feature incay')
-  parser.add_argument('--use-deformable', type=int, default=0, help='')
-  parser.add_argument('--rand-mirror', type=int, default=1, help='')
-  parser.add_argument('--cutoff', type=int, default=0, help='')
-  parser.add_argument('--patch', type=str, default='0_0_96_112_0',help='')
-  parser.add_argument('--lr-steps', type=str, default='', help='')
-  parser.add_argument('--max-steps', type=int, default=0, help='')
-  parser.add_argument('--target', type=str, default='lfw,cfp_fp,agedb_30,cplfw,calfw', help='')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol(args, arg_params, aux_params):
-  data_shape = (args.image_channel,args.image_h,args.image_w)
-  image_shape = ",".join([str(x) for x in data_shape])
-  margin_symbols = []
-  if args.network[0]=='d':
-    embedding = fdensenet.get_symbol(args.emb_size, args.num_layers,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='m':
-    print('init mobilenet', args.num_layers)
-    if args.num_layers==1:
-      embedding = fmobilenet.get_symbol(args.emb_size, 
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit)
-    else:
-      embedding = fmobilenetv2.get_symbol(args.emb_size)
-  elif args.network[0]=='i':
-    print('init inception-resnet-v2', args.num_layers)
-    embedding = finception_resnet_v2.get_symbol(args.emb_size,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='x':
-    print('init xception', args.num_layers)
-    embedding = fxception.get_symbol(args.emb_size,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='p':
-    print('init dpn', args.num_layers)
-    embedding = fdpn.get_symbol(args.emb_size, args.num_layers,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='n':
-    print('init nasnet', args.num_layers)
-    embedding = fnasnet.get_symbol(args.emb_size)
-  elif args.network[0]=='s':
-    print('init spherenet', args.num_layers)
-    embedding = spherenet.get_symbol(args.emb_size, args.num_layers)
-  else:
-    print('init resnet', args.num_layers)
-    embedding = fresnet.get_symbol(args.emb_size, args.num_layers, 
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit,
-        version_act=args.version_act)
-  all_label = mx.symbol.Variable('softmax_label')
-  if not args.output_c2c:
-    gt_label = all_label
-  else:
-    gt_label = mx.symbol.slice_axis(all_label, axis=1, begin=0, end=1)
-    gt_label = mx.symbol.reshape(gt_label, (args.per_batch_size,))
-    c2c_label = mx.symbol.slice_axis(all_label, axis=1, begin=1, end=2)
-    c2c_label = mx.symbol.reshape(c2c_label, (args.per_batch_size,))
-  assert args.loss_type>=0
-  extra_loss = None
-  if args.loss_type==0: #softmax
-    _weight = mx.symbol.Variable('fc7_weight')
-    _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
-    fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=args.num_classes, name='fc7')
-  elif args.loss_type==1: #sphere
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    fc7 = mx.sym.LSoftmax(data=embedding, label=gt_label, num_hidden=args.num_classes,
-                          weight = _weight,
-                          beta=args.beta, margin=args.margin, scale=args.scale,
-                          beta_min=args.beta_min, verbose=1000, name='fc7')
-  elif args.loss_type==8: #centerloss, TODO
-    _weight = mx.symbol.Variable('fc7_weight')
-    _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
-    fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=args.num_classes, name='fc7')
-    print('center-loss', args.center_alpha, args.center_scale)
-    extra_loss = mx.symbol.Custom(data=embedding, label=gt_label, name='center_loss', op_type='centerloss',\
-          num_class=args.num_classes, alpha=args.center_alpha, scale=args.center_scale, batchsize=args.per_batch_size)
-  elif args.loss_type==2:
-    s = args.margin_s
-    m = args.margin_m
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    if s>0.0:
-      nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-      fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-      if m>0.0:
-        if args.margin_verbose>0:
-          zy = mx.sym.pick(fc7, gt_label, axis=1)
-          cos_t = zy/s
-          margin_symbols.append(mx.symbol.mean(cos_t))
-
-        s_m = s*m
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = s_m, off_value = 0.0)
-        fc7 = fc7-gt_one_hot
-
-        if args.margin_verbose>0:
-          new_zy = mx.sym.pick(fc7, gt_label, axis=1)
-          new_cos_t = new_zy/s
-          margin_symbols.append(mx.symbol.mean(new_cos_t))
-    else:
-      fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-      if m>0.0:
-        body = embedding*embedding
-        body = mx.sym.sum_axis(body, axis=1, keepdims=True)
-        body = mx.sym.sqrt(body)
-        body = body*m
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-        body = mx.sym.broadcast_mul(gt_one_hot, body)
-        fc7 = fc7-body
-
-  elif args.loss_type==3:
-    s = args.margin_s
-    m = args.margin_m
-    assert args.margin==2 or args.margin==4
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    if args.margin_verbose>0:
-      margin_symbols.append(mx.symbol.mean(cos_t))
-    if m>1.0:
-      t = mx.sym.arccos(cos_t)
-      t = t*m
-      body = mx.sym.cos(t)
-      new_zy = body*s
-      if args.margin_verbose>0:
-        new_cos_t = new_zy/s
-        margin_symbols.append(mx.symbol.mean(new_cos_t))
-      diff = new_zy - zy
-      diff = mx.sym.expand_dims(diff, 1)
-      gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-      body = mx.sym.broadcast_mul(gt_one_hot, diff)
-      fc7 = fc7+body
-
-    #threshold = math.cos(args.margin_m)
-    #cond_v = cos_t - threshold
-    #cond = mx.symbol.Activation(data=cond_v, act_type='relu')
-    #body = cos_t
-    #for i in xrange(args.margin//2):
-    #  body = body*body
-    #  body = body*2-1
-    #new_zy = body*s
-    #zy_keep = zy
-    #new_zy = mx.sym.where(cond, new_zy, zy_keep)
-    #if args.margin_verbose>0:
-    #  new_cos_t = new_zy/s
-    #  margin_symbols.append(mx.symbol.mean(new_cos_t))
-    #diff = new_zy - zy
-    #diff = mx.sym.expand_dims(diff, 1)
-    #gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-    #body = mx.sym.broadcast_mul(gt_one_hot, diff)
-    #fc7 = fc7+body
-  elif args.loss_type==4:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert m>=0.0
-    assert m<(math.pi/2)
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    if args.margin_verbose>0:
-      margin_symbols.append(mx.symbol.mean(cos_t))
-    if args.output_c2c==0:
-      cos_m = math.cos(m)
-      sin_m = math.sin(m)
-      mm = math.sin(math.pi-m)*m
-      #threshold = 0.0
-      threshold = math.cos(math.pi-m)
-      if args.easy_margin:
-        cond = mx.symbol.Activation(data=cos_t, act_type='relu')
-      else:
-        cond_v = cos_t - threshold
-        cond = mx.symbol.Activation(data=cond_v, act_type='relu')
-      body = cos_t*cos_t
-      body = 1.0-body
-      sin_t = mx.sym.sqrt(body)
-      new_zy = cos_t*cos_m
-      b = sin_t*sin_m
-      new_zy = new_zy - b
-      new_zy = new_zy*s
-      if args.easy_margin:
-        zy_keep = zy
-      else:
-        zy_keep = zy - s*mm
-      new_zy = mx.sym.where(cond, new_zy, zy_keep)
-    else:
-      #set c2c as cosm^2 in data.py
-      cos_m = mx.sym.sqrt(c2c_label)
-      sin_m = 1.0-c2c_label
-      sin_m = mx.sym.sqrt(sin_m)
-      body = cos_t*cos_t
-      body = 1.0-body
-      sin_t = mx.sym.sqrt(body)
-      new_zy = cos_t*cos_m
-      b = sin_t*sin_m
-      new_zy = new_zy - b
-      new_zy = new_zy*s
-
-    if args.margin_verbose>0:
-      new_cos_t = new_zy/s
-      margin_symbols.append(mx.symbol.mean(new_cos_t))
-    diff = new_zy - zy
-    diff = mx.sym.expand_dims(diff, 1)
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-    body = mx.sym.broadcast_mul(gt_one_hot, diff)
-    fc7 = fc7+body
-  elif args.loss_type==5:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    if args.margin_a!=1.0 or args.margin_m!=0.0 or args.margin_b!=0.0:
-      if args.margin_a==1.0 and args.margin_m==0.0:
-        s_m = s*args.margin_b
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = s_m, off_value = 0.0)
-        fc7 = fc7-gt_one_hot
-      else:
-        zy = mx.sym.pick(fc7, gt_label, axis=1)
-        cos_t = zy/s
-        t = mx.sym.arccos(cos_t)
-        if args.margin_a!=1.0:
-          t = t*args.margin_a
-        if args.margin_m>0.0:
-          t = t+args.margin_m
-        body = mx.sym.cos(t)
-        if args.margin_b>0.0:
-          body = body - args.margin_b
-        new_zy = body*s
-        diff = new_zy - zy
-        diff = mx.sym.expand_dims(diff, 1)
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-        body = mx.sym.broadcast_mul(gt_one_hot, diff)
-        fc7 = fc7+body
-  elif args.loss_type==6:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert m>=0.0
-    assert m<(math.pi/2)
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    t = mx.sym.arccos(cos_t)
-    if args.margin_verbose>0:
-      margin_symbols.append(mx.symbol.mean(t))
-    t_min = mx.sym.min(t)
-    ta = mx.sym.broadcast_div(t_min, t)
-
-    a1 = args.margin_a
-    r1 = ta-a1
-    r1 = mx.symbol.Activation(data=r1, act_type='relu')
-    r1 = r1+a1
-
-    r2 = mx.symbol.zeros(shape=(args.per_batch_size,))
-
-    cond = t-1.0
-    cond = mx.symbol.Activation(data=cond, act_type='relu')
-    r = mx.sym.where(cond, r2, r1)
-    var_m = r*m
-    t = t+var_m
-    body = mx.sym.cos(t)
-    new_zy = body*s
-    if args.margin_verbose>0:
-      #new_cos_t = new_zy/s
-      #margin_symbols.append(mx.symbol.mean(new_cos_t))
-      margin_symbols.append(mx.symbol.mean(t))
-    diff = new_zy - zy
-    diff = mx.sym.expand_dims(diff, 1)
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-    body = mx.sym.broadcast_mul(gt_one_hot, diff)
-    fc7 = fc7+body
-  elif args.loss_type==7:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert m>=0.0
-    assert m<(math.pi/2)
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    t = mx.sym.arccos(cos_t)
-    if args.margin_verbose>0:
-      margin_symbols.append(mx.symbol.mean(t))
-    var_m = mx.sym.random.uniform(low=args.margin_b, high=args.margin_m, shape=(1,))
-    t = mx.sym.broadcast_add(t,var_m)
-    body = mx.sym.cos(t)
-    new_zy = body*s
-    if args.margin_verbose>0:
-      #new_cos_t = new_zy/s
-      #margin_symbols.append(mx.symbol.mean(new_cos_t))
-      margin_symbols.append(mx.symbol.mean(t))
-    diff = new_zy - zy
-    diff = mx.sym.expand_dims(diff, 1)
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-    body = mx.sym.broadcast_mul(gt_one_hot, diff)
-    fc7 = fc7+body
-  elif args.loss_type==10: #marginal loss
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-    params = [1.2, 0.3, 1.0]
-    n1 = mx.sym.expand_dims(nembedding, axis=1) #N,1,C
-    n2 = mx.sym.expand_dims(nembedding, axis=0) #1,N,C
-    body = mx.sym.broadcast_sub(n1, n2) #N,N,C
-    body = body * body
-    body = mx.sym.sum(body, axis=2) # N,N
-    #body = mx.sym.sqrt(body)
-    body = body - params[0]
-    mask = mx.sym.Variable('extra')
-    body = body*mask
-    body = body+params[1]
-    #body = mx.sym.maximum(body, 0.0)
-    body = mx.symbol.Activation(data=body, act_type='relu')
-    body = mx.sym.sum(body)
-    body = body/(args.per_batch_size*args.per_batch_size-args.per_batch_size)
-    extra_loss = mx.symbol.MakeLoss(body, grad_scale=params[2])
-  elif args.loss_type==11: #npair loss
-    params = [0.9, 0.2]
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-    nembedding = mx.sym.transpose(nembedding)
-    nembedding = mx.symbol.reshape(nembedding, (args.emb_size, args.per_identities, args.images_per_identity))
-    nembedding = mx.sym.transpose(nembedding, axes=(2,1,0)) #2*id*512
-    #nembedding = mx.symbol.reshape(nembedding, (args.emb_size, args.images_per_identity, args.per_identities))
-    #nembedding = mx.sym.transpose(nembedding, axes=(1,2,0)) #2*id*512
-    n1 = mx.symbol.slice_axis(nembedding, axis=0, begin=0, end=1)
-    n2 = mx.symbol.slice_axis(nembedding, axis=0, begin=1, end=2)
-    #n1 = []
-    #n2 = []
-    #for i in xrange(args.per_identities):
-    #  _n1 = mx.symbol.slice_axis(nembedding, axis=0, begin=2*i, end=2*i+1)
-    #  _n2 = mx.symbol.slice_axis(nembedding, axis=0, begin=2*i+1, end=2*i+2)
-    #  n1.append(_n1)
-    #  n2.append(_n2)
-    #n1 = mx.sym.concat(*n1, dim=0)
-    #n2 = mx.sym.concat(*n2, dim=0)
-    #rembeddings = mx.symbol.reshape(nembedding, (args.images_per_identity, args.per_identities, 512))
-    #n1 = mx.symbol.slice_axis(rembeddings, axis=0, begin=0, end=1)
-    #n2 = mx.symbol.slice_axis(rembeddings, axis=0, begin=1, end=2)
-    n1 = mx.symbol.reshape(n1, (args.per_identities, args.emb_size))
-    n2 = mx.symbol.reshape(n2, (args.per_identities, args.emb_size))
-    cosine_matrix = mx.symbol.dot(lhs=n1, rhs=n2, transpose_b = True) #id*id, id=N of N-pair
-    data_extra = mx.sym.Variable('extra')
-    data_extra = mx.sym.slice_axis(data_extra, axis=0, begin=0, end=args.per_identities)
-    mask = cosine_matrix * data_extra
-    #body = mx.sym.mean(mask)
-    fii = mx.sym.sum_axis(mask, axis=1)
-    fij_fii = mx.sym.broadcast_sub(cosine_matrix, fii)
-    fij_fii = mx.sym.exp(fij_fii)
-    row = mx.sym.sum_axis(fij_fii, axis=1)
-    row = mx.sym.log(row)
-    body = mx.sym.mean(row)
-    extra_loss = mx.sym.MakeLoss(body)
-  elif args.loss_type==12: #triplet loss
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-    anchor = mx.symbol.slice_axis(nembedding, axis=0, begin=0, end=args.per_batch_size//3)
-    positive = mx.symbol.slice_axis(nembedding, axis=0, begin=args.per_batch_size//3, end=2*args.per_batch_size//3)
-    negative = mx.symbol.slice_axis(nembedding, axis=0, begin=2*args.per_batch_size//3, end=args.per_batch_size)
-    ap = anchor - positive
-    an = anchor - negative
-    ap = ap*ap
-    an = an*an
-    ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-    an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-    triplet_loss = mx.symbol.Activation(data = (ap-an+args.triplet_alpha), act_type='relu')
-    triplet_loss = mx.symbol.mean(triplet_loss)
-    #triplet_loss = mx.symbol.sum(triplet_loss)/(args.per_batch_size//3)
-    extra_loss = mx.symbol.MakeLoss(triplet_loss)
-  elif args.loss_type==13: #triplet loss with angular margin
-    m = args.margin_m
-    sin_m = math.sin(m)
-    cos_m = math.cos(m)
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-    anchor = mx.symbol.slice_axis(nembedding, axis=0, begin=0, end=args.per_batch_size//3)
-    positive = mx.symbol.slice_axis(nembedding, axis=0, begin=args.per_batch_size//3, end=2*args.per_batch_size//3)
-    negative = mx.symbol.slice_axis(nembedding, axis=0, begin=2*args.per_batch_size//3, end=args.per_batch_size)
-    ap = anchor * positive
-    an = anchor * negative
-    ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-    an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-
-    ap = mx.symbol.arccos(ap)
-    an = mx.symbol.arccos(an)
-    triplet_loss = mx.symbol.Activation(data = (ap-an+args.margin_m), act_type='relu')
-
-    #body = ap*ap
-    #body = 1.0-body
-    #body = mx.symbol.sqrt(body)
-    #body = body*sin_m
-    #ap = ap*cos_m
-    #ap = ap-body
-    #triplet_loss = mx.symbol.Activation(data = (an-ap), act_type='relu')
-
-    triplet_loss = mx.symbol.mean(triplet_loss)
-    extra_loss = mx.symbol.MakeLoss(triplet_loss)
-  elif args.loss_type==9: #coco loss
-    centroids = []
-    for i in xrange(args.per_identities):
-      xs = mx.symbol.slice_axis(embedding, axis=0, begin=i*args.images_per_identity, end=(i+1)*args.images_per_identity)
-      mean = mx.symbol.mean(xs, axis=0, keepdims=True)
-      mean = mx.symbol.L2Normalization(mean, mode='instance')
-      centroids.append(mean)
-    centroids = mx.symbol.concat(*centroids, dim=0)
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*args.coco_scale
-    fc7 = mx.symbol.dot(nembedding, centroids, transpose_b = True) #(batchsize, per_identities)
-    #extra_loss = mx.symbol.softmax_cross_entropy(fc7, gt_label, name='softmax_ce')/args.per_batch_size
-    #extra_loss = mx.symbol.BlockGrad(extra_loss)
-  else:
-    #embedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*float(args.loss_type)
-    embedding = embedding * 5
-    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance') * 2
-    fc7 = mx.sym.LSoftmax(data=embedding, label=gt_label, num_hidden=args.num_classes,
-                          weight = _weight,
-                          beta=args.beta, margin=args.margin, scale=args.scale,
-                          beta_min=args.beta_min, verbose=100, name='fc7')
-
-    #fc7 = mx.sym.Custom(data=embedding, label=gt_label, weight=_weight, num_hidden=args.num_classes,
-    #                       beta=args.beta, margin=args.margin, scale=args.scale,
-    #                       op_type='ASoftmax', name='fc7')
-  if args.loss_type<=1 and args.incay>0.0:
-    params = [1.e-10]
-    sel = mx.symbol.argmax(data = fc7, axis=1)
-    sel = (sel==gt_label)
-    norm = embedding*embedding
-    norm = mx.symbol.sum(norm, axis=1)
-    norm = norm+params[0]
-    feature_incay = sel/norm
-    feature_incay = mx.symbol.mean(feature_incay) * args.incay
-    extra_loss = mx.symbol.MakeLoss(feature_incay)
-  #out = softmax
-  #l2_embedding = mx.symbol.L2Normalization(embedding)
-
-  #ce = mx.symbol.softmax_cross_entropy(fc7, gt_label, name='softmax_ce')/args.per_batch_size
-  #out = mx.symbol.Group([mx.symbol.BlockGrad(embedding), softmax, mx.symbol.BlockGrad(ce)])
-  out_list = [mx.symbol.BlockGrad(embedding)]
-  softmax = None
-  if args.loss_type<10:
-    softmax = mx.symbol.SoftmaxOutput(data=fc7, label = gt_label, name='softmax', normalization='valid')
-    out_list.append(softmax)
-    if args.logits_verbose>0:
-      logits = mx.symbol.softmax(data = fc7)
-      logits = mx.sym.pick(logits, gt_label, axis=1)
-      margin_symbols.append(logits)
-      #logit_max = mx.sym.max(logits)
-      #logit_min = mx.sym.min(logits)
-      #margin_symbols.append(logit_max)
-      #margin_symbols.append(logit_min)
-  if softmax is None:
-    out_list.append(mx.sym.BlockGrad(gt_label))
-  if extra_loss is not None:
-    out_list.append(extra_loss)
-  for _sym in margin_symbols:
-    _sym = mx.sym.BlockGrad(_sym)
-    out_list.append(_sym)
-  out = mx.symbol.Group(out_list)
-  return (out, arg_params, aux_params)
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in xrange(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = args.prefix
-    prefix_dir = os.path.dirname(prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    end_epoch = args.end_epoch
-    args.ctx_num = len(ctx)
-    args.num_layers = int(args.network[1:])
-    print('num_layers', args.num_layers)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-      if args.loss_type==10:
-        args.per_batch_size = 256
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = 3
-    ppatch = [int(x) for x in args.patch.split('_')]
-    assert len(ppatch)==5
-
-
-    os.environ['BETA'] = str(args.beta)
-    data_dir_list = args.data_dir.split(',')
-    if args.loss_type!=12 and args.loss_type!=13:
-      assert len(data_dir_list)==1
-    data_dir = data_dir_list[0]
-    args.use_val = False
-    path_imgrec = None
-    path_imglist = None
-    val_rec = None
-    prop = face_image.load_property(data_dir)
-    args.num_classes = prop.num_classes
-    image_size = prop.image_size
-    args.image_h = image_size[0]
-    args.image_w = image_size[1]
-    print('image_size', image_size)
-
-    assert(args.num_classes>0)
-    print('num_classes', args.num_classes)
-    args.coco_scale = 0.5*math.log(float(args.num_classes-1))+3
-
-    #path_imglist = "/raid5data/dplearn/MS-Celeb-Aligned/lst2"
-    path_imgrec = os.path.join(data_dir, "train.rec")
-    val_rec = os.path.join(data_dir, "val.rec")
-    if os.path.exists(val_rec) and args.loss_type<10:
-      args.use_val = True
-    else:
-      val_rec = None
-    #args.use_val = False
-
-    if args.loss_type==1 and args.num_classes>20000:
-      args.beta_freeze = 5000
-      args.gamma = 0.06
-
-    if args.loss_type<9:
-      assert args.images_per_identity==0
-    else:
-      if args.images_per_identity==0:
-        if args.loss_type==11:
-          args.images_per_identity = 2
-        elif args.loss_type==10 or args.loss_type==9:
-          args.images_per_identity = 16
-        elif args.loss_type==12 or args.loss_type==13:
-          args.images_per_identity = 5
-          assert args.per_batch_size%3==0
-      assert args.images_per_identity>=2
-      args.per_identities = int(args.per_batch_size/args.images_per_identity)
-
-    print('Called with argument:', args)
-
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-    mean = None
-
-
-
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    if len(args.pretrained)==0:
-      arg_params = None
-      aux_params = None
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-    else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      _, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-
-    data_extra = None
-    hard_mining = False
-    triplet_params = None
-    coco_mode = False
-    if args.loss_type==10:
-      hard_mining = True
-      _shape = (args.batch_size, args.per_batch_size)
-      data_extra = np.full(_shape, -1.0, dtype=np.float32)
-      c = 0
-      while c<args.batch_size:
-        a = 0
-        while a<args.per_batch_size:
-          b = a+args.images_per_identity
-          data_extra[(c+a):(c+b),a:b] = 1.0
-          #print(c+a, c+b, a, b)
-          a = b
-        c += args.per_batch_size
-    elif args.loss_type==11:
-      data_extra = np.zeros( (args.batch_size, args.per_identities), dtype=np.float32)
-      c = 0
-      while c<args.batch_size:
-        for i in xrange(args.per_identities):
-          data_extra[c+i][i] = 1.0
-        c+=args.per_batch_size
-    elif args.loss_type==12 or args.loss_type==13:
-      triplet_params = [args.triplet_bag_size, args.triplet_alpha, args.triplet_max_ap]
-    elif args.loss_type==9:
-      coco_mode = True
-
-    label_name = 'softmax_label'
-    label_shape = (args.batch_size,)
-    if args.output_c2c:
-      label_shape = (args.batch_size,2)
-    if data_extra is None:
-      model = mx.mod.Module(
-          context       = ctx,
-          symbol        = sym,
-      )
-    else:
-      data_names = ('data', 'extra')
-      #label_name = ''
-      model = mx.mod.Module(
-          context       = ctx,
-          symbol        = sym,
-          data_names    = data_names,
-          label_names   = (label_name,),
-      )
-
-    if args.use_val:
-      val_dataiter = FaceImageIter(
-          batch_size           = args.batch_size,
-          data_shape           = data_shape,
-          path_imgrec          = val_rec,
-          #path_imglist         = val_path,
-          shuffle              = False,
-          rand_mirror          = False,
-          mean                 = mean,
-          ctx_num              = args.ctx_num,
-          data_extra           = data_extra,
-      )
-    else:
-      val_dataiter = None
-
-    if len(data_dir_list)==1 and args.loss_type!=12 and args.loss_type!=13:
-      train_dataiter = FaceImageIter(
-          batch_size           = args.batch_size,
-          data_shape           = data_shape,
-          path_imgrec          = path_imgrec,
-          shuffle              = True,
-          rand_mirror          = args.rand_mirror,
-          mean                 = mean,
-          cutoff               = args.cutoff,
-          c2c_threshold        = args.c2c_threshold,
-          output_c2c           = args.output_c2c,
-          c2c_mode             = args.c2c_mode,
-          limit                = args.train_limit,
-          ctx_num              = args.ctx_num,
-          images_per_identity  = args.images_per_identity,
-          data_extra           = data_extra,
-          hard_mining          = hard_mining,
-          triplet_params       = triplet_params,
-          coco_mode            = coco_mode,
-          mx_model             = model,
-          label_name           = label_name,
-      )
-    else:
-      iter_list = []
-      for _data_dir in data_dir_list:
-        _path_imgrec = os.path.join(_data_dir, "train.rec")
-        _dataiter = FaceImageIter(
-            batch_size           = args.batch_size,
-            data_shape           = data_shape,
-            path_imgrec          = _path_imgrec,
-            shuffle              = True,
-            rand_mirror          = args.rand_mirror,
-            mean                 = mean,
-            cutoff               = args.cutoff,
-            c2c_threshold        = args.c2c_threshold,
-            output_c2c           = args.output_c2c,
-            c2c_mode             = args.c2c_mode,
-            limit                = args.train_limit,
-            ctx_num              = args.ctx_num,
-            images_per_identity  = args.images_per_identity,
-            data_extra           = data_extra,
-            hard_mining          = hard_mining,
-            triplet_params       = triplet_params,
-            coco_mode            = coco_mode,
-            mx_model             = model,
-            label_name           = label_name,
-        )
-        iter_list.append(_dataiter)
-      iter_list.append(_dataiter)
-      train_dataiter = FaceImageIterList(iter_list)
-
-    if args.loss_type<10:
-      _metric = AccMetric()
-    else:
-      _metric = LossValueMetric()
-    eval_metrics = [mx.metric.create(_metric)]
-
-    if args.network[0]=='r':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    elif args.network[0]=='i' or args.network[0]=='x':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2) #inception
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-    _rescale = 1.0/args.ctx_num
-    if args.noise_sgd>0.0:
-      print('use noise sgd')
-      opt = NoiseSGD(scale = args.noise_sgd, learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    else:
-      opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    som = 20
-    if args.loss_type==12 or args.loss_type==13:
-      som = 2
-    _cb = mx.callback.Speedometer(args.batch_size, som)
-
-    ver_list = []
-    ver_name_list = []
-    for name in args.target.split(','):
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in xrange(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, data_extra, label_shape)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-    def val_test():
-      acc = AccMetric()
-      val_metric = mx.metric.create(acc)
-      val_metric.reset()
-      val_dataiter.reset()
-      for i, eval_batch in enumerate(val_dataiter):
-        model.forward(eval_batch, is_train=False)
-        model.update_metric(val_metric, eval_batch.label)
-      acc_value = val_metric.get_name_value()[0][1]
-      print('VACC: %f'%(acc_value))
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in xrange(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    if len(args.lr_steps)==0:
-      lr_steps = [40000, 60000, 80000]
-      if args.loss_type>=1 and args.loss_type<=7:
-        lr_steps = [100000, 140000, 160000]
-      p = 512.0/args.batch_size
-      for l in xrange(len(lr_steps)):
-        lr_steps[l] = int(lr_steps[l]*p)
-    else:
-      lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for _lr in lr_steps:
-        if mbatch==args.beta_freeze+_lr:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        if len(acc_list)>0:
-          lfw_score = acc_list[0]
-          if lfw_score>highest_acc[0]:
-            highest_acc[0] = lfw_score
-            if lfw_score>=0.998:
-              do_save = True
-          if acc_list[-1]>=highest_acc[-1]:
-            highest_acc[-1] = acc_list[-1]
-            if lfw_score>=0.99:
-              do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt>1:
-          do_save = True
-        #for i in xrange(len(acc_list)):
-        #  acc = acc_list[i]
-        #  if acc>=highest_acc[i]:
-        #    highest_acc[i] = acc
-        #    if lfw_score>=0.99:
-        #      do_save = True
-        #if args.loss_type==1 and mbatch>lr_steps[-1] and mbatch%10000==0:
-        #  do_save = True
-        if do_save:
-          print('saving', msave)
-          if val_dataiter is not None:
-            val_test()
-          arg, aux = model.get_params()
-          mx.model.save_checkpoint(prefix, msave, model.symbol, arg, aux)
-          #if acc>=highest_acc[0]:
-          #  lfw_npy = "%s-lfw-%04d" % (prefix, msave)
-          #  X = np.concatenate(embeddings_list, axis=0)
-          #  print('saving lfw npy', X.shape)
-          #  np.save(lfw_npy, X)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if mbatch<=args.beta_freeze:
-        _beta = args.beta
-      else:
-        move = max(0, mbatch-args.beta_freeze)
-        _beta = max(args.beta_min, args.beta*math.pow(1+args.gamma*move, -1.0*args.power))
-      #print('beta', _beta)
-      os.environ['BETA'] = str(_beta)
-      if args.max_steps>0 and mbatch>args.max_steps:
-        sys.exit(0)
-
-    #epoch_cb = mx.callback.do_checkpoint(prefix, 1)
-    epoch_cb = None
-
-
-
-    #def _epoch_callback(epoch, sym, arg_params, aux_params):
-    #  print('epoch-end', epoch)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = end_epoch,
-        eval_data          = val_dataiter,
-        eval_metric        = eval_metrics,
-        kvstore            = 'device',
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    #time.sleep(3600*6.5)
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/src/train.sh b/src/train.sh
deleted file mode 100755
index 6b3c6ae6b..000000000
--- a/src/train.sh
+++ /dev/null
@@ -1,15 +0,0 @@
-#!/usr/bin/env bash
-export MXNET_CPU_WORKER_NTHREADS=24
-export MXNET_CUDNN_AUTOTUNE_DEFAULT=0
-export MXNET_ENGINE_TYPE=ThreadedEnginePerDevice
-
-DATA_DIR=/opt/jiaguo/faces_vgg_112x112
-
-NETWORK=r50
-JOB=softmax1e3
-MODELDIR="../model-$NETWORK-$JOB"
-mkdir -p "$MODELDIR"
-PREFIX="$MODELDIR/model"
-LOGFILE="$MODELDIR/log"
-CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --data-dir $DATA_DIR --network "$NETWORK" --loss-type 0 --prefix "$PREFIX" --per-batch-size 128 > "$LOGFILE" 2>&1 &
-
diff --git a/src/train_softmax.py b/src/train_softmax.py
deleted file mode 100644
index 229669ec5..000000000
--- a/src/train_softmax.py
+++ /dev/null
@@ -1,598 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import numpy as np
-from image_iter import FaceImageIter
-from image_iter import FaceImageIterList
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_image
-sys.path.append(os.path.join(os.path.dirname(__file__), 'eval'))
-sys.path.append(os.path.join(os.path.dirname(__file__), 'symbols'))
-import fresnet
-import finception_resnet_v2
-import fmobilenet 
-import fmobilenetv2
-import fmobilefacenet
-import fxception
-import fdensenet
-import fdpn
-import fnasnet
-import spherenet
-import verification
-import sklearn
-#sys.path.append(os.path.join(os.path.dirname(__file__), 'losses'))
-#import center_loss
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-class AccMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(AccMetric, self).__init__(
-        'acc', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-    self.count = 0
-
-  def update(self, labels, preds):
-    self.count+=1
-    label = labels[0]
-    pred_label = preds[1]
-    if pred_label.shape != label.shape:
-        pred_label = mx.ndarray.argmax(pred_label, axis=self.axis)
-    pred_label = pred_label.asnumpy().astype('int32').flatten()
-    label = label.asnumpy()
-    if label.ndim==2:
-      label = label[:,0]
-    label = label.astype('int32').flatten()
-    assert label.shape==pred_label.shape
-    self.sum_metric += (pred_label.flat == label.flat).sum()
-    self.num_inst += len(pred_label.flat)
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[-1].asnumpy()[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-    gt_label = preds[-2].asnumpy()
-    #print(gt_label)
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--data-dir', default='', help='training set directory')
-  parser.add_argument('--prefix', default='../model/model', help='directory to save model.')
-  parser.add_argument('--pretrained', default='', help='pretrained model to load')
-  parser.add_argument('--ckpt', type=int, default=1, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--loss-type', type=int, default=4, help='loss type')
-  parser.add_argument('--verbose', type=int, default=2000, help='do verification testing and model saving every verbose batches')
-  parser.add_argument('--max-steps', type=int, default=0, help='max training batches')
-  parser.add_argument('--end-epoch', type=int, default=100000, help='training epoch size.')
-  parser.add_argument('--network', default='r50', help='specify network')
-  parser.add_argument('--image-size', default='112,112', help='specify input image height and width')
-  parser.add_argument('--version-se', type=int, default=0, help='whether to use se in network')
-  parser.add_argument('--version-input', type=int, default=1, help='network input config')
-  parser.add_argument('--version-output', type=str, default='E', help='network embedding output config')
-  parser.add_argument('--version-unit', type=int, default=3, help='resnet unit config')
-  parser.add_argument('--version-multiplier', type=float, default=1.0, help='filters multiplier')
-  parser.add_argument('--version-act', type=str, default='prelu', help='network activation config')
-  parser.add_argument('--use-deformable', type=int, default=0, help='use deformable cnn in network')
-  parser.add_argument('--lr', type=float, default=0.1, help='start learning rate')
-  parser.add_argument('--lr-steps', type=str, default='', help='steps of lr changing')
-  parser.add_argument('--wd', type=float, default=0.0005, help='weight decay')
-  parser.add_argument('--fc7-wd-mult', type=float, default=1.0, help='weight decay mult for fc7')
-  parser.add_argument('--fc7-lr-mult', type=float, default=1.0, help='lr mult for fc7')
-  parser.add_argument("--fc7-no-bias", default=False, action="store_true" , help="fc7 no bias flag")
-  parser.add_argument('--bn-mom', type=float, default=0.9, help='bn mom')
-  parser.add_argument('--mom', type=float, default=0.9, help='momentum')
-  parser.add_argument('--emb-size', type=int, default=512, help='embedding length')
-  parser.add_argument('--per-batch-size', type=int, default=128, help='batch size in each context')
-  parser.add_argument('--margin-m', type=float, default=0.5, help='margin for loss')
-  parser.add_argument('--margin-s', type=float, default=64.0, help='scale for feature')
-  parser.add_argument('--margin-a', type=float, default=1.0, help='')
-  parser.add_argument('--margin-b', type=float, default=0.0, help='')
-  parser.add_argument('--easy-margin', type=int, default=0, help='')
-  parser.add_argument('--margin', type=int, default=4, help='margin for sphere')
-  parser.add_argument('--beta', type=float, default=1000., help='param for sphere')
-  parser.add_argument('--beta-min', type=float, default=5., help='param for sphere')
-  parser.add_argument('--beta-freeze', type=int, default=0, help='param for sphere')
-  parser.add_argument('--gamma', type=float, default=0.12, help='param for sphere')
-  parser.add_argument('--power', type=float, default=1.0, help='param for sphere')
-  parser.add_argument('--scale', type=float, default=0.9993, help='param for sphere')
-  parser.add_argument('--rand-mirror', type=int, default=1, help='if do random mirror in training')
-  parser.add_argument('--cutoff', type=int, default=0, help='cut off aug')
-  parser.add_argument('--color', type=int, default=0, help='color jittering aug')
-  parser.add_argument('--images-filter', type=int, default=0, help='minimum images per identity filter')
-  parser.add_argument('--target', type=str, default='lfw,cfp_fp,agedb_30', help='verification targets')
-  parser.add_argument('--ce-loss', default=False, action='store_true', help='if output ce loss')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol(args, arg_params, aux_params):
-  data_shape = (args.image_channel,args.image_h,args.image_w)
-  image_shape = ",".join([str(x) for x in data_shape])
-  margin_symbols = []
-  if args.network[0]=='d':
-    embedding = fdensenet.get_symbol(args.emb_size, args.num_layers,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='m':
-    print('init mobilenet', args.num_layers)
-    if args.num_layers==1:
-      embedding = fmobilenet.get_symbol(args.emb_size, 
-          version_input=args.version_input, 
-          version_output=args.version_output,
-          version_multiplier = args.version_multiplier)
-    else:
-      embedding = fmobilenetv2.get_symbol(args.emb_size)
-  elif args.network[0]=='i':
-    print('init inception-resnet-v2', args.num_layers)
-    embedding = finception_resnet_v2.get_symbol(args.emb_size,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='x':
-    print('init xception', args.num_layers)
-    embedding = fxception.get_symbol(args.emb_size,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='p':
-    print('init dpn', args.num_layers)
-    embedding = fdpn.get_symbol(args.emb_size, args.num_layers,
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit)
-  elif args.network[0]=='n':
-    print('init nasnet', args.num_layers)
-    embedding = fnasnet.get_symbol(args.emb_size)
-  elif args.network[0]=='s':
-    print('init spherenet', args.num_layers)
-    embedding = spherenet.get_symbol(args.emb_size, args.num_layers)
-  elif args.network[0]=='y':
-    print('init mobilefacenet', args.num_layers)
-    embedding = fmobilefacenet.get_symbol(args.emb_size, bn_mom = args.bn_mom, version_output=args.version_output)
-  else:
-    print('init resnet', args.num_layers)
-    embedding = fresnet.get_symbol(args.emb_size, args.num_layers, 
-        version_se=args.version_se, version_input=args.version_input, 
-        version_output=args.version_output, version_unit=args.version_unit,
-        version_act=args.version_act)
-  all_label = mx.symbol.Variable('softmax_label')
-  gt_label = all_label
-  extra_loss = None
-  _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=args.fc7_lr_mult, wd_mult=args.fc7_wd_mult)
-  if args.loss_type==0: #softmax
-    if args.fc7_no_bias:
-      fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    else:
-      _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
-      fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=args.num_classes, name='fc7')
-  elif args.loss_type==1: #sphere
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    fc7 = mx.sym.LSoftmax(data=embedding, label=gt_label, num_hidden=args.num_classes,
-                          weight = _weight,
-                          beta=args.beta, margin=args.margin, scale=args.scale,
-                          beta_min=args.beta_min, verbose=1000, name='fc7')
-  elif args.loss_type==2:
-    s = args.margin_s
-    m = args.margin_m
-    assert(s>0.0)
-    assert(m>0.0)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    s_m = s*m
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = s_m, off_value = 0.0)
-    fc7 = fc7-gt_one_hot
-  elif args.loss_type==4:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert m>=0.0
-    assert m<(math.pi/2)
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    cos_m = math.cos(m)
-    sin_m = math.sin(m)
-    mm = math.sin(math.pi-m)*m
-    #threshold = 0.0
-    threshold = math.cos(math.pi-m)
-    if args.easy_margin:
-      cond = mx.symbol.Activation(data=cos_t, act_type='relu')
-    else:
-      cond_v = cos_t - threshold
-      cond = mx.symbol.Activation(data=cond_v, act_type='relu')
-    body = cos_t*cos_t
-    body = 1.0-body
-    sin_t = mx.sym.sqrt(body)
-    new_zy = cos_t*cos_m
-    b = sin_t*sin_m
-    new_zy = new_zy - b
-    new_zy = new_zy*s
-    if args.easy_margin:
-      zy_keep = zy
-    else:
-      zy_keep = zy - s*mm
-    new_zy = mx.sym.where(cond, new_zy, zy_keep)
-
-    diff = new_zy - zy
-    diff = mx.sym.expand_dims(diff, 1)
-    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-    body = mx.sym.broadcast_mul(gt_one_hot, diff)
-    fc7 = fc7+body
-  elif args.loss_type==5:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    if args.margin_a!=1.0 or args.margin_m!=0.0 or args.margin_b!=0.0:
-      if args.margin_a==1.0 and args.margin_m==0.0:
-        s_m = s*args.margin_b
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = s_m, off_value = 0.0)
-        fc7 = fc7-gt_one_hot
-      else:
-        zy = mx.sym.pick(fc7, gt_label, axis=1)
-        cos_t = zy/s
-        t = mx.sym.arccos(cos_t)
-        if args.margin_a!=1.0:
-          t = t*args.margin_a
-        if args.margin_m>0.0:
-          t = t+args.margin_m
-        body = mx.sym.cos(t)
-        if args.margin_b>0.0:
-          body = body - args.margin_b
-        new_zy = body*s
-        diff = new_zy - zy
-        diff = mx.sym.expand_dims(diff, 1)
-        gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-        body = mx.sym.broadcast_mul(gt_one_hot, diff)
-        fc7 = fc7+body
-  elif args.loss_type==6:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert args.margin_b>0.0
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    t = mx.sym.arccos(cos_t)
-    intra_loss = t/np.pi
-    intra_loss = mx.sym.mean(intra_loss)
-    #intra_loss = mx.sym.exp(cos_t*-1.0)
-    intra_loss = mx.sym.MakeLoss(intra_loss, name='intra_loss', grad_scale = args.margin_b)
-    if m>0.0:
-      t = t+m
-      body = mx.sym.cos(t)
-      new_zy = body*s
-      diff = new_zy - zy
-      diff = mx.sym.expand_dims(diff, 1)
-      gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-      body = mx.sym.broadcast_mul(gt_one_hot, diff)
-      fc7 = fc7+body
-  elif args.loss_type==7:
-    s = args.margin_s
-    m = args.margin_m
-    assert s>0.0
-    assert args.margin_b>0.0
-    assert args.margin_a>0.0
-    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
-    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s
-    fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
-    zy = mx.sym.pick(fc7, gt_label, axis=1)
-    cos_t = zy/s
-    t = mx.sym.arccos(cos_t)
-
-    #counter_weight = mx.sym.take(_weight, gt_label, axis=1)
-    #counter_cos = mx.sym.dot(counter_weight, _weight, transpose_a=True)
-    counter_weight = mx.sym.take(_weight, gt_label, axis=0)
-    counter_cos = mx.sym.dot(counter_weight, _weight, transpose_b=True)
-    #counter_cos = mx.sym.minimum(counter_cos, 1.0)
-    #counter_angle = mx.sym.arccos(counter_cos)
-    #counter_angle = counter_angle * -1.0
-    #counter_angle = counter_angle/np.pi #[0,1]
-    #inter_loss = mx.sym.exp(counter_angle)
-
-    #counter_cos = mx.sym.dot(_weight, _weight, transpose_b=True)
-    #counter_cos = mx.sym.minimum(counter_cos, 1.0)
-    #counter_angle = mx.sym.arccos(counter_cos)
-    #counter_angle = mx.sym.sort(counter_angle, axis=1)
-    #counter_angle = mx.sym.slice_axis(counter_angle, axis=1, begin=0,end=int(args.margin_a))
-
-    #inter_loss = counter_angle*-1.0 # [-1,0]
-    #inter_loss = inter_loss+1.0 # [0,1]
-    inter_loss = counter_cos
-    inter_loss = mx.sym.mean(inter_loss)
-    inter_loss = mx.sym.MakeLoss(inter_loss, name='inter_loss', grad_scale = args.margin_b)
-    if m>0.0:
-      t = t+m
-      body = mx.sym.cos(t)
-      new_zy = body*s
-      diff = new_zy - zy
-      diff = mx.sym.expand_dims(diff, 1)
-      gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
-      body = mx.sym.broadcast_mul(gt_one_hot, diff)
-      fc7 = fc7+body
-  out_list = [mx.symbol.BlockGrad(embedding)]
-  softmax = mx.symbol.SoftmaxOutput(data=fc7, label = gt_label, name='softmax', normalization='valid')
-  out_list.append(softmax)
-  if args.loss_type==6:
-    out_list.append(intra_loss)
-  if args.loss_type==7:
-    out_list.append(inter_loss)
-    #out_list.append(mx.sym.BlockGrad(counter_weight))
-    #out_list.append(intra_loss)
-  if args.ce_loss:
-    #ce_loss = mx.symbol.softmax_cross_entropy(data=fc7, label = gt_label, name='ce_loss')/args.per_batch_size
-    body = mx.symbol.SoftmaxActivation(data=fc7)
-    body = mx.symbol.log(body)
-    _label = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = -1.0, off_value = 0.0)
-    body = body*_label
-    ce_loss = mx.symbol.sum(body)/args.per_batch_size
-    out_list.append(mx.symbol.BlockGrad(ce_loss))
-  out = mx.symbol.Group(out_list)
-  return (out, arg_params, aux_params)
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in xrange(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = args.prefix
-    prefix_dir = os.path.dirname(prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    end_epoch = args.end_epoch
-    args.ctx_num = len(ctx)
-    args.num_layers = int(args.network[1:])
-    print('num_layers', args.num_layers)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.rescale_threshold = 0
-    args.image_channel = 3
-
-    os.environ['BETA'] = str(args.beta)
-    data_dir_list = args.data_dir.split(',')
-    assert len(data_dir_list)==1
-    data_dir = data_dir_list[0]
-    path_imgrec = None
-    path_imglist = None
-    prop = face_image.load_property(data_dir)
-    args.num_classes = prop.num_classes
-    #image_size = prop.image_size
-    image_size = [int(x) for x in args.image_size.split(',')]
-    assert len(image_size)==2
-    assert image_size[0]==image_size[1]
-    args.image_h = image_size[0]
-    args.image_w = image_size[1]
-    print('image_size', image_size)
-    assert(args.num_classes>0)
-    print('num_classes', args.num_classes)
-    path_imgrec = os.path.join(data_dir, "train.rec")
-
-    if args.loss_type==1 and args.num_classes>20000:
-      args.beta_freeze = 5000
-      args.gamma = 0.06
-
-    print('Called with argument:', args)
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-    mean = None
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    if len(args.pretrained)==0:
-      arg_params = None
-      aux_params = None
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-      if args.network[0]=='s':
-        data_shape_dict = {'data' : (args.per_batch_size,)+data_shape}
-        spherenet.init_weights(sym, data_shape_dict, args.num_layers)
-    else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      _, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-
-    #label_name = 'softmax_label'
-    #label_shape = (args.batch_size,)
-    model = mx.mod.Module(
-        context       = ctx,
-        symbol        = sym,
-    )
-    val_dataiter = None
-
-    train_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec,
-        shuffle              = True,
-        rand_mirror          = args.rand_mirror,
-        mean                 = mean,
-        cutoff               = args.cutoff,
-        color_jittering      = args.color,
-        images_filter        = args.images_filter,
-    )
-
-    metric1 = AccMetric()
-    eval_metrics = [mx.metric.create(metric1)]
-    if args.ce_loss:
-      metric2 = LossValueMetric()
-      eval_metrics.append( mx.metric.create(metric2) )
-
-    if args.network[0]=='r' or args.network[0]=='y':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    elif args.network[0]=='i' or args.network[0]=='x':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2) #inception
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-    #initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    _rescale = 1.0/args.ctx_num
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    som = 20
-    _cb = mx.callback.Speedometer(args.batch_size, som)
-
-    ver_list = []
-    ver_name_list = []
-    for name in args.target.split(','):
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in xrange(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, None, None)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in xrange(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    if len(args.lr_steps)==0:
-      lr_steps = [40000, 60000, 80000]
-      if args.loss_type>=1 and args.loss_type<=7:
-        lr_steps = [100000, 140000, 160000]
-      p = 512.0/args.batch_size
-      for l in xrange(len(lr_steps)):
-        lr_steps[l] = int(lr_steps[l]*p)
-    else:
-      lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for _lr in lr_steps:
-        if mbatch==args.beta_freeze+_lr:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        is_highest = False
-        if len(acc_list)>0:
-          #lfw_score = acc_list[0]
-          #if lfw_score>highest_acc[0]:
-          #  highest_acc[0] = lfw_score
-          #  if lfw_score>=0.998:
-          #    do_save = True
-          score = sum(acc_list)
-          if acc_list[-1]>=highest_acc[-1]:
-            if acc_list[-1]>highest_acc[-1]:
-              is_highest = True
-            else:
-              if score>=highest_acc[0]:
-                is_highest = True
-                highest_acc[0] = score
-            highest_acc[-1] = acc_list[-1]
-            #if lfw_score>=0.99:
-            #  do_save = True
-        if is_highest:
-          do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt==2:
-          do_save = True
-        elif args.ckpt==3:
-          msave = 1
-
-        if do_save:
-          print('saving', msave)
-          arg, aux = model.get_params()
-          mx.model.save_checkpoint(prefix, msave, model.symbol, arg, aux)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if mbatch<=args.beta_freeze:
-        _beta = args.beta
-      else:
-        move = max(0, mbatch-args.beta_freeze)
-        _beta = max(args.beta_min, args.beta*math.pow(1+args.gamma*move, -1.0*args.power))
-      #print('beta', _beta)
-      os.environ['BETA'] = str(_beta)
-      if args.max_steps>0 and mbatch>args.max_steps:
-        sys.exit(0)
-
-    epoch_cb = None
-    train_dataiter = mx.io.PrefetchingIter(train_dataiter)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = end_epoch,
-        eval_data          = val_dataiter,
-        eval_metric        = eval_metrics,
-        kvstore            = 'device',
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    #time.sleep(3600*6.5)
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/src/train_triplet.py b/src/train_triplet.py
deleted file mode 100644
index 512708912..000000000
--- a/src/train_triplet.py
+++ /dev/null
@@ -1,402 +0,0 @@
-# THIS FILE IS FOR EXPERIMENTS, USE train_softmax.py FOR NORMAL TRAINING.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import math
-import random
-import logging
-import pickle
-import numpy as np
-from triplet_image_iter import FaceImageIter
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_image
-from noise_sgd import NoiseSGD
-sys.path.append(os.path.join(os.path.dirname(__file__), 'eval'))
-sys.path.append(os.path.join(os.path.dirname(__file__), 'symbols'))
-import fresnet
-import finception_resnet_v2
-import fmobilenet 
-import fmobilenetv2
-import fxception
-import fdensenet
-import fdpn
-import fnasnet
-import spherenet
-#import lfw
-import verification
-import sklearn
-sys.path.append(os.path.join(os.path.dirname(__file__), 'losses'))
-import center_loss
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-
-args = None
-
-
-class LossValueMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(LossValueMetric, self).__init__(
-        'lossvalue', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    loss = preds[-1].asnumpy()[0]
-    self.sum_metric += loss
-    self.num_inst += 1.0
-    gt_label = preds[-2].asnumpy()
-    #print(gt_label)
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--data-dir', default='', help='training set directory')
-  parser.add_argument('--prefix', default='../model/model', help='directory to save model.')
-  parser.add_argument('--pretrained', default='', help='pretrained model to load')
-  parser.add_argument('--ckpt', type=int, default=3, help='checkpoint saving option. 0: discard saving. 1: save when necessary. 2: always save')
-  parser.add_argument('--network', default='r50', help='specify network')
-  parser.add_argument('--version-se', type=int, default=0, help='whether to use se in network')
-  parser.add_argument('--version-input', type=int, default=1, help='network input config')
-  parser.add_argument('--version-output', type=str, default='E', help='network embedding output config')
-  parser.add_argument('--version-unit', type=int, default=3, help='resnet unit config')
-  parser.add_argument('--version-act', type=str, default='prelu', help='network activation config')
-  parser.add_argument('--end-epoch', type=int, default=100000, help='training epoch size.')
-  parser.add_argument('--noise-sgd', type=float, default=0.0, help='')
-  parser.add_argument('--lr', type=float, default=0.1, help='start learning rate')
-  parser.add_argument('--wd', type=float, default=0.0005, help='weight decay')
-  parser.add_argument('--mom', type=float, default=0.9, help='momentum')
-  parser.add_argument('--emb-size', type=int, default=512, help='embedding length')
-  parser.add_argument('--per-batch-size', type=int, default=128, help='batch size in each context')
-  parser.add_argument('--images-per-identity', type=int, default=5, help='')
-  parser.add_argument('--triplet-bag-size', type=int, default=3600, help='')
-  parser.add_argument('--triplet-alpha', type=float, default=0.3, help='')
-  parser.add_argument('--triplet-max-ap', type=float, default=0.0, help='')
-  parser.add_argument('--verbose', type=int, default=2000, help='')
-  parser.add_argument('--loss-type', type=int, default=1, help='')
-  parser.add_argument('--use-deformable', type=int, default=0, help='')
-  parser.add_argument('--rand-mirror', type=int, default=1, help='')
-  parser.add_argument('--cutoff', type=int, default=0, help='')
-  parser.add_argument('--lr-steps', type=str, default='', help='')
-  parser.add_argument('--max-steps', type=int, default=0, help='')
-  parser.add_argument('--target', type=str, default='lfw,cfp_fp,agedb_30', help='')
-  args = parser.parse_args()
-  return args
-
-
-def get_symbol(args, arg_params, aux_params, sym_embedding=None):
-  if sym_embedding is None:
-    if args.network[0]=='d':
-      embedding = fdensenet.get_symbol(args.emb_size, args.num_layers,
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit)
-    elif args.network[0]=='m':
-      print('init mobilenet', args.num_layers)
-      if args.num_layers==1:
-        embedding = fmobilenet.get_symbol(args.emb_size, 
-            version_se=args.version_se, version_input=args.version_input, 
-            version_output=args.version_output, version_unit=args.version_unit)
-      else:
-        embedding = fmobilenetv2.get_symbol(args.emb_size)
-    elif args.network[0]=='i':
-      print('init inception-resnet-v2', args.num_layers)
-      embedding = finception_resnet_v2.get_symbol(args.emb_size,
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit)
-    elif args.network[0]=='x':
-      print('init xception', args.num_layers)
-      embedding = fxception.get_symbol(args.emb_size,
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit)
-    elif args.network[0]=='p':
-      print('init dpn', args.num_layers)
-      embedding = fdpn.get_symbol(args.emb_size, args.num_layers,
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit)
-    elif args.network[0]=='n':
-      print('init nasnet', args.num_layers)
-      embedding = fnasnet.get_symbol(args.emb_size)
-    elif args.network[0]=='s':
-      print('init spherenet', args.num_layers)
-      embedding = spherenet.get_symbol(args.emb_size, args.num_layers)
-    else:
-      print('init resnet', args.num_layers)
-      embedding = fresnet.get_symbol(args.emb_size, args.num_layers, 
-          version_se=args.version_se, version_input=args.version_input, 
-          version_output=args.version_output, version_unit=args.version_unit,
-          version_act=args.version_act)
-  else:
-    embedding = sym_embedding
-
-  assert args.loss_type==1 or args.loss_type==2
-  gt_label = mx.symbol.Variable('softmax_label')
-  nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
-  anchor = mx.symbol.slice_axis(nembedding, axis=0, begin=0, end=args.per_batch_size//3)
-  positive = mx.symbol.slice_axis(nembedding, axis=0, begin=args.per_batch_size//3, end=2*args.per_batch_size//3)
-  negative = mx.symbol.slice_axis(nembedding, axis=0, begin=2*args.per_batch_size//3, end=args.per_batch_size)
-  if args.loss_type==1:
-    ap = anchor - positive
-    an = anchor - negative
-    ap = ap*ap
-    an = an*an
-    ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-    an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-    triplet_loss = mx.symbol.Activation(data = (ap-an+args.triplet_alpha), act_type='relu')
-    triplet_loss = mx.symbol.mean(triplet_loss)
-  elif args.loss_type==2:
-    ap = anchor*positive
-    an = anchor*negative
-    ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-    an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-    ap = mx.sym.arccos(ap)
-    an = mx.sym.arccos(an)
-    triplet_loss = mx.symbol.Activation(data = (ap-an+args.triplet_alpha), act_type='relu')
-    triplet_loss = mx.symbol.mean(triplet_loss)
-  #triplet_loss = mx.symbol.sum(triplet_loss)/(args.per_batch_size//3)
-  triplet_loss = mx.symbol.MakeLoss(triplet_loss)
-  out_list = [mx.symbol.BlockGrad(embedding)]
-  out_list.append(mx.sym.BlockGrad(gt_label))
-  out_list.append(triplet_loss)
-  out = mx.symbol.Group(out_list)
-  return (out, arg_params, aux_params)
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in xrange(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = args.prefix
-    prefix_dir = os.path.dirname(prefix)
-    if not os.path.exists(prefix_dir):
-      os.makedirs(prefix_dir)
-    end_epoch = args.end_epoch
-    args.ctx_num = len(ctx)
-    args.num_layers = int(args.network[1:])
-    print('num_layers', args.num_layers)
-    if args.per_batch_size==0:
-      args.per_batch_size = 128
-    args.batch_size = args.per_batch_size*args.ctx_num
-    args.image_channel = 3
-
-
-    data_dir_list = args.data_dir.split(',')
-    assert len(data_dir_list)==1
-    data_dir = data_dir_list[0]
-    path_imgrec = None
-    path_imglist = None
-    prop = face_image.load_property(data_dir)
-    args.num_classes = prop.num_classes
-    image_size = prop.image_size
-    args.image_h = image_size[0]
-    args.image_w = image_size[1]
-    print('image_size', image_size)
-
-    assert(args.num_classes>0)
-    print('num_classes', args.num_classes)
-
-    #path_imglist = "/raid5data/dplearn/MS-Celeb-Aligned/lst2"
-    path_imgrec = os.path.join(data_dir, "train.rec")
-
-    assert args.images_per_identity>=2
-    assert args.triplet_bag_size%args.batch_size==0
-
-    print('Called with argument:', args)
-
-    data_shape = (args.image_channel,image_size[0],image_size[1])
-    mean = None
-
-
-
-
-    begin_epoch = 0
-    base_lr = args.lr
-    base_wd = args.wd
-    base_mom = args.mom
-    if len(args.pretrained)==0:
-      arg_params = None
-      aux_params = None
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-    else:
-      vec = args.pretrained.split(',')
-      print('loading', vec)
-      sym, arg_params, aux_params = mx.model.load_checkpoint(vec[0], int(vec[1]))
-      all_layers = sym.get_internals()
-      sym = all_layers['fc1_output']
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params, sym_embedding = sym)
-
-    data_extra = None
-    hard_mining = False
-    triplet_params = [args.triplet_bag_size, args.triplet_alpha, args.triplet_max_ap]
-    model = mx.mod.Module(
-        context       = ctx,
-        symbol        = sym,
-        #data_names = ('data',),
-        #label_names = None,
-        #label_names = ('softmax_label',),
-    )
-    label_shape = (args.batch_size,)
-
-    val_dataiter = None
-
-    train_dataiter = FaceImageIter(
-        batch_size           = args.batch_size,
-        data_shape           = data_shape,
-        path_imgrec          = path_imgrec,
-        shuffle              = True,
-        rand_mirror          = args.rand_mirror,
-        mean                 = mean,
-        cutoff               = args.cutoff,
-        ctx_num              = args.ctx_num,
-        images_per_identity  = args.images_per_identity,
-        triplet_params       = triplet_params,
-        mx_model             = model,
-    )
-
-    _metric = LossValueMetric()
-    eval_metrics = [mx.metric.create(_metric)]
-
-    if args.network[0]=='r':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
-    elif args.network[0]=='i' or args.network[0]=='x':
-      initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2) #inception
-    else:
-      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
-    _rescale = 1.0/args.ctx_num
-    if args.noise_sgd>0.0:
-      print('use noise sgd')
-      opt = NoiseSGD(scale = args.noise_sgd, learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    else:
-      opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
-    som = 2
-    _cb = mx.callback.Speedometer(args.batch_size, som)
-
-    ver_list = []
-    ver_name_list = []
-    for name in args.target.split(','):
-      path = os.path.join(data_dir,name+".bin")
-      if os.path.exists(path):
-        data_set = verification.load_bin(path, image_size)
-        ver_list.append(data_set)
-        ver_name_list.append(name)
-        print('ver', name)
-
-
-
-    def ver_test(nbatch):
-      results = []
-      for i in xrange(len(ver_list)):
-        acc1, std1, acc2, std2, xnorm, embeddings_list = verification.test(ver_list[i], model, args.batch_size, 10, None, label_shape)
-        print('[%s][%d]XNorm: %f' % (ver_name_list[i], nbatch, xnorm))
-        #print('[%s][%d]Accuracy: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc1, std1))
-        print('[%s][%d]Accuracy-Flip: %1.5f+-%1.5f' % (ver_name_list[i], nbatch, acc2, std2))
-        results.append(acc2)
-      return results
-
-
-    highest_acc = [0.0, 0.0]  #lfw and target
-    #for i in xrange(len(ver_list)):
-    #  highest_acc.append(0.0)
-    global_step = [0]
-    save_step = [0]
-    if len(args.lr_steps)==0:
-      lr_steps = [1000000000]
-    else:
-      lr_steps = [int(x) for x in args.lr_steps.split(',')]
-    print('lr_steps', lr_steps)
-    def _batch_callback(param):
-      #global global_step
-      global_step[0]+=1
-      mbatch = global_step[0]
-      for _lr in lr_steps:
-        if mbatch==_lr:
-          opt.lr *= 0.1
-          print('lr change to', opt.lr)
-          break
-
-      _cb(param)
-      if mbatch%1000==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-
-      if mbatch>=0 and mbatch%args.verbose==0:
-        acc_list = ver_test(mbatch)
-        save_step[0]+=1
-        msave = save_step[0]
-        do_save = False
-        is_highest = False
-        if len(acc_list)>0:
-          #lfw_score = acc_list[0]
-          #if lfw_score>highest_acc[0]:
-          #  highest_acc[0] = lfw_score
-          #  if lfw_score>=0.998:
-          #    do_save = True
-          score = sum(acc_list)
-          if acc_list[-1]>=highest_acc[-1]:
-            if acc_list[-1]>highest_acc[-1]:
-              is_highest = True
-            else:
-              if score>=highest_acc[0]:
-                is_highest = True
-                highest_acc[0] = score
-            highest_acc[-1] = acc_list[-1]
-            #if lfw_score>=0.99:
-            #  do_save = True
-        if is_highest:
-          do_save = True
-        if args.ckpt==0:
-          do_save = False
-        elif args.ckpt==2:
-          do_save = True
-        elif args.ckpt==3:
-          msave = 1
-
-        if do_save:
-          print('saving', msave)
-          arg, aux = model.get_params()
-          mx.model.save_checkpoint(prefix, msave, model.symbol, arg, aux)
-        print('[%d]Accuracy-Highest: %1.5f'%(mbatch, highest_acc[-1]))
-      if args.max_steps>0 and mbatch>args.max_steps:
-        sys.exit(0)
-
-    #epoch_cb = mx.callback.do_checkpoint(prefix, 1)
-    epoch_cb = None
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = end_epoch,
-        eval_data          = val_dataiter,
-        eval_metric        = eval_metrics,
-        kvstore            = 'device',
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    #time.sleep(3600*6.5)
-    global args
-    args = parse_args()
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
-
diff --git a/src/triplet_image_iter.py b/src/triplet_image_iter.py
deleted file mode 100644
index d121b1110..000000000
--- a/src/triplet_image_iter.py
+++ /dev/null
@@ -1,608 +0,0 @@
-# THIS FILE IS FOR EXPERIMENTS, USE image_iter.py FOR NORMAL IMAGE LOADING.
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import random
-import logging
-import sys
-import numbers
-import math
-import sklearn
-import datetime
-import numpy as np
-import cv2
-
-import mxnet as mx
-from mxnet import ndarray as nd
-#from . import _ndarray_internal as _internal
-#from mxnet._ndarray_internal import _cvimresize as imresize
-#from ._ndarray_internal import _cvcopyMakeBorder as copyMakeBorder
-from mxnet import io
-from mxnet import recordio
-sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
-import face_preprocess
-
-logger = logging.getLogger()
-
-class FaceImageIter(io.DataIter):
-
-    def __init__(self, batch_size, data_shape,
-                 path_imgrec = None,
-                 shuffle=False, aug_list=None,
-                 rand_mirror = False, cutoff = 0,
-                 ctx_num = 0, images_per_identity = 0,
-                 triplet_params = None,
-                 mx_model = None,
-                 data_name='data', label_name='softmax_label', **kwargs):
-        super(FaceImageIter, self).__init__()
-        assert path_imgrec
-        assert shuffle
-        logging.info('loading recordio %s...',
-                     path_imgrec)
-        path_imgidx = path_imgrec[0:-4]+".idx"
-        self.imgrec = recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-        s = self.imgrec.read_idx(0)
-        header, _ = recordio.unpack(s)
-        assert header.flag>0
-        print('header0 label', header.label)
-        self.header0 = (int(header.label[0]), int(header.label[1]))
-        #assert(header.flag==1)
-        self.imgidx = range(1, int(header.label[0]))
-        self.id2range = {}
-        self.seq_identity = range(int(header.label[0]), int(header.label[1]))
-        for identity in self.seq_identity:
-          s = self.imgrec.read_idx(identity)
-          header, _ = recordio.unpack(s)
-          a,b = int(header.label[0]), int(header.label[1])
-          self.id2range[identity] = (a,b)
-
-        print('id2range', len(self.id2range))
-        self.seq = self.imgidx
-        print(len(self.seq))
-
-        self.check_data_shape(data_shape)
-        self.provide_data = [(data_name, (batch_size,) + data_shape)]
-        self.batch_size = batch_size
-        self.data_shape = data_shape
-        self.shuffle = shuffle
-        self.image_size = '%d,%d'%(data_shape[1],data_shape[2])
-        self.rand_mirror = rand_mirror
-        print('rand_mirror', rand_mirror)
-        self.cutoff = cutoff
-        #self.cast_aug = mx.image.CastAug()
-        #self.color_aug = mx.image.ColorJitterAug(0.4, 0.4, 0.4)
-        self.ctx_num = ctx_num 
-        self.per_batch_size = int(self.batch_size/self.ctx_num)
-        self.images_per_identity = images_per_identity
-        if self.images_per_identity>0:
-          self.identities = int(self.per_batch_size/self.images_per_identity)
-          self.per_identities = self.identities
-          self.repeat = 3000000.0/(self.images_per_identity*len(self.id2range))
-          self.repeat = int(self.repeat)
-          print(self.images_per_identity, self.identities, self.repeat)
-        self.mx_model = mx_model
-        self.triplet_params = triplet_params
-        self.triplet_mode = False
-        #self.provide_label = None
-        self.provide_label = [(label_name, (batch_size,))]
-        if self.triplet_params is not None:
-          assert self.images_per_identity>0
-          assert self.mx_model is not None
-          self.triplet_bag_size = self.triplet_params[0]
-          self.triplet_alpha = self.triplet_params[1]
-          self.triplet_max_ap = self.triplet_params[2]
-          assert self.triplet_bag_size>0
-          assert self.triplet_alpha>=0.0
-          assert self.triplet_alpha<=1.0
-          self.triplet_mode = True
-          self.triplet_cur = 0
-          self.triplet_seq = []
-          self.triplet_reset()
-          self.seq_min_size = self.batch_size*2
-        self.cur = 0
-        self.nbatch = 0
-        self.is_init = False
-        self.times = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
-        #self.reset()
-
-
-
-    def pairwise_dists(self, embeddings):
-      nd_embedding_list = []
-      for i in xrange(self.ctx_num):
-        nd_embedding = mx.nd.array(embeddings, mx.gpu(i))
-        nd_embedding_list.append(nd_embedding)
-      nd_pdists = []
-      pdists = []
-      for idx in xrange(embeddings.shape[0]):
-        emb_idx = idx%self.ctx_num
-        nd_embedding = nd_embedding_list[emb_idx]
-        a_embedding = nd_embedding[idx]
-        body = mx.nd.broadcast_sub(a_embedding, nd_embedding)
-        body = body*body
-        body = mx.nd.sum_axis(body, axis=1)
-        nd_pdists.append(body)
-        if len(nd_pdists)==self.ctx_num or idx==embeddings.shape[0]-1:
-          for x in nd_pdists:
-            pdists.append(x.asnumpy())
-          nd_pdists = []
-      return pdists
-
-    def pick_triplets(self, embeddings, nrof_images_per_class):
-      emb_start_idx = 0
-      triplets = []
-      people_per_batch = len(nrof_images_per_class)
-      #self.time_reset()
-      pdists = self.pairwise_dists(embeddings)
-      #self.times[3] += self.time_elapsed()
-
-      for i in xrange(people_per_batch):
-          nrof_images = int(nrof_images_per_class[i])
-          for j in xrange(1,nrof_images):
-              #self.time_reset()
-              a_idx = emb_start_idx + j - 1
-              #neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
-              neg_dists_sqr = pdists[a_idx]
-              #self.times[3] += self.time_elapsed()
-
-              for pair in xrange(j, nrof_images): # For every possible positive pair.
-                  p_idx = emb_start_idx + pair
-                  #self.time_reset()
-                  pos_dist_sqr = np.sum(np.square(embeddings[a_idx]-embeddings[p_idx]))
-                  #self.times[4] += self.time_elapsed()
-                  #self.time_reset()
-                  neg_dists_sqr[emb_start_idx:emb_start_idx+nrof_images] = np.NaN
-                  if self.triplet_max_ap>0.0:
-                    if pos_dist_sqr>self.triplet_max_ap:
-                      continue
-                  all_neg = np.where(np.logical_and(neg_dists_sqr-pos_dist_sqr<self.triplet_alpha, pos_dist_sqr<neg_dists_sqr))[0]  # FaceNet selection
-                  #self.times[5] += self.time_elapsed()
-                  #self.time_reset()
-                  #all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
-                  nrof_random_negs = all_neg.shape[0]
-                  if nrof_random_negs>0:
-                      rnd_idx = np.random.randint(nrof_random_negs)
-                      n_idx = all_neg[rnd_idx]
-                      triplets.append( (a_idx, p_idx, n_idx) )
-          emb_start_idx += nrof_images
-      np.random.shuffle(triplets)
-      return triplets
-
-    def triplet_reset(self):
-      #reset self.oseq by identities seq
-      self.triplet_cur = 0
-      ids = []
-      for k in self.id2range:
-        ids.append(k)
-      random.shuffle(ids)
-      self.triplet_seq = []
-      for _id in ids:
-        v = self.id2range[_id]
-        _list = range(*v)
-        random.shuffle(_list)
-        if len(_list)>self.images_per_identity:
-          _list = _list[0:self.images_per_identity]
-        self.triplet_seq += _list
-      print('triplet_seq', len(self.triplet_seq))
-      assert len(self.triplet_seq)>=self.triplet_bag_size
-
-    def time_reset(self):
-      self.time_now = datetime.datetime.now()
-
-    def time_elapsed(self):
-      time_now = datetime.datetime.now()
-      diff = time_now - self.time_now
-      return diff.total_seconds()
-
-
-    def select_triplets(self):
-      self.seq = []
-      while len(self.seq)<self.seq_min_size:
-        self.time_reset()
-        embeddings = None
-        bag_size = self.triplet_bag_size
-        batch_size = self.batch_size
-        #data = np.zeros( (bag_size,)+self.data_shape )
-        #label = np.zeros( (bag_size,) )
-        tag = []
-        #idx = np.zeros( (bag_size,) )
-        print('eval %d images..'%bag_size, self.triplet_cur)
-        print('triplet time stat', self.times)
-        if self.triplet_cur+bag_size>len(self.triplet_seq):
-          self.triplet_reset()
-          #bag_size = min(bag_size, len(self.triplet_seq))
-          print('eval %d images..'%bag_size, self.triplet_cur)
-        self.times[0] += self.time_elapsed()
-        self.time_reset()
-        #print(data.shape)
-        data = nd.zeros( self.provide_data[0][1] )
-        label = None
-        if self.provide_label is not None:
-          label = nd.zeros( self.provide_label[0][1] )
-        ba = 0
-        while True:
-          bb = min(ba+batch_size, bag_size)
-          if ba>=bb:
-            break
-          _count = bb-ba
-          #data = nd.zeros( (_count,)+self.data_shape )
-          #_batch = self.data_iter.next()
-          #_data = _batch.data[0].asnumpy()
-          #print(_data.shape)
-          #_label = _batch.label[0].asnumpy()
-          #data[ba:bb,:,:,:] = _data
-          #label[ba:bb] = _label
-          for i in xrange(ba, bb):
-            #print(ba, bb, self.triplet_cur, i, len(self.triplet_seq))
-            _idx = self.triplet_seq[i+self.triplet_cur]
-            s = self.imgrec.read_idx(_idx)
-            header, img = recordio.unpack(s)
-            img = self.imdecode(img)
-            data[i-ba][:] = self.postprocess_data(img)
-            _label = header.label
-            if not isinstance(_label, numbers.Number):
-              _label = _label[0]
-            if label is not None:
-              label[i-ba][:] = _label
-            tag.append( ( int(_label), _idx) )
-            #idx[i] = _idx
-
-          db = mx.io.DataBatch(data=(data,))
-          self.mx_model.forward(db, is_train=False)
-          net_out = self.mx_model.get_outputs()
-          #print('eval for selecting triplets',ba,bb)
-          #print(net_out)
-          #print(len(net_out))
-          #print(net_out[0].asnumpy())
-          net_out = net_out[0].asnumpy()
-          #print(net_out)
-          #print('net_out', net_out.shape)
-          if embeddings is None:
-            embeddings = np.zeros( (bag_size, net_out.shape[1]))
-          embeddings[ba:bb,:] = net_out
-          ba = bb
-        assert len(tag)==bag_size
-        self.triplet_cur+=bag_size
-        embeddings = sklearn.preprocessing.normalize(embeddings)
-        self.times[1] += self.time_elapsed()
-        self.time_reset()
-        nrof_images_per_class = [1]
-        for i in xrange(1, bag_size):
-          if tag[i][0]==tag[i-1][0]:
-            nrof_images_per_class[-1]+=1
-          else:
-            nrof_images_per_class.append(1)
-          
-        triplets = self.pick_triplets(embeddings, nrof_images_per_class) # shape=(T,3)
-        print('found triplets', len(triplets))
-        ba = 0
-        while True:
-          bb = ba+self.per_batch_size//3
-          if bb>len(triplets):
-            break
-          _triplets = triplets[ba:bb]
-          for i in xrange(3):
-            for triplet in _triplets:
-              _pos = triplet[i]
-              _idx = tag[_pos][1]
-              self.seq.append(_idx)
-          ba = bb
-        self.times[2] += self.time_elapsed()
-
-    def hard_mining_reset(self):
-      #import faiss
-      from annoy import AnnoyIndex
-      data = nd.zeros( self.provide_data[0][1] )
-      label = nd.zeros( self.provide_label[0][1] )
-      #label = np.zeros( self.provide_label[0][1] )
-      X = None
-      ba = 0
-      batch_num = 0
-      while ba<len(self.oseq):
-        batch_num+=1
-        if batch_num%10==0:
-          print('loading batch',batch_num, ba)
-        bb = min(ba+self.batch_size, len(self.oseq))
-        _count = bb-ba
-        for i in xrange(_count):
-          idx = self.oseq[i+ba]
-          s = self.imgrec.read_idx(idx)
-          header, img = recordio.unpack(s)
-          img = self.imdecode(img)
-          data[i][:] = self.postprocess_data(img)
-          label[i][:] = header.label
-        db = mx.io.DataBatch(data=(data,self.data_extra), label=(label,))
-        self.mx_model.forward(db, is_train=False)
-        net_out = self.mx_model.get_outputs()
-        embedding = net_out[0].asnumpy()
-        nembedding = sklearn.preprocessing.normalize(embedding)
-        if _count<self.batch_size:
-          nembedding = nembedding[0:_count,:]
-        if X is None:
-          X = np.zeros( (len(self.id2range), nembedding.shape[1]), dtype=np.float32 )
-        nplabel = label.asnumpy()
-        for i in xrange(_count):
-          ilabel = int(nplabel[i])
-          #print(ilabel, ilabel.__class__)
-          X[ilabel] += nembedding[i]
-        ba = bb
-      X = sklearn.preprocessing.normalize(X)
-      d = X.shape[1]
-      t = AnnoyIndex(d, metric='euclidean')
-      for i in xrange(X.shape[0]):
-        t.add_item(i, X[i])
-      print('start to build index')
-      t.build(20)
-      print(X.shape)
-      k = self.per_identities
-      self.seq = []
-      for i in xrange(X.shape[0]):
-        nnlist = t.get_nns_by_item(i, k)
-        assert nnlist[0]==i
-        for _label in nnlist:
-          assert _label<len(self.id2range)
-          _id = self.header0[0]+_label
-          v = self.id2range[_id]
-          _list = range(*v)
-          if len(_list)<self.images_per_identity:
-            random.shuffle(_list)
-          else:
-            _list = np.random.choice(_list, self.images_per_identity, replace=False)
-          for i in xrange(self.images_per_identity):
-            _idx = _list[i%len(_list)]
-            self.seq.append(_idx)
-      #faiss_params = [20,5]
-      #quantizer = faiss.IndexFlatL2(d)  # the other index
-      #index = faiss.IndexIVFFlat(quantizer, d, faiss_params[0], faiss.METRIC_L2)
-      #assert not index.is_trained
-      #index.train(X)
-      #index.add(X)
-      #assert index.is_trained
-      #print('trained')
-      #index.nprobe = faiss_params[1]
-      #D, I = index.search(X, k)     # actual search
-      #print(I.shape)
-      #self.seq = []
-      #for i in xrange(I.shape[0]):
-      #  #assert I[i][0]==i
-      #  for j in xrange(k):
-      #    _label = I[i][j]
-      #    assert _label<len(self.id2range)
-      #    _id = self.header0[0]+_label
-      #    v = self.id2range[_id]
-      #    _list = range(*v)
-      #    if len(_list)<self.images_per_identity:
-      #      random.shuffle(_list)
-      #    else:
-      #      _list = np.random.choice(_list, self.images_per_identity, replace=False)
-      #    for i in xrange(self.images_per_identity):
-      #      _idx = _list[i%len(_list)]
-      #      self.seq.append(_idx)
-
-    def reset(self):
-        """Resets the iterator to the beginning of the data."""
-        print('call reset()')
-        self.cur = 0
-        if self.images_per_identity>0:
-          if self.triplet_mode:
-            self.select_triplets()
-          elif not self.hard_mining:
-            self.seq = []
-            idlist = []
-            for _id,v in self.id2range.iteritems():
-              idlist.append((_id,range(*v)))
-            for r in xrange(self.repeat):
-              if r%10==0:
-                print('repeat', r)
-              if self.shuffle:
-                random.shuffle(idlist)
-              for item in idlist:
-                _id = item[0]
-                _list = item[1]
-                #random.shuffle(_list)
-                if len(_list)<self.images_per_identity:
-                  random.shuffle(_list)
-                else:
-                  _list = np.random.choice(_list, self.images_per_identity, replace=False)
-                for i in xrange(self.images_per_identity):
-                  _idx = _list[i%len(_list)]
-                  self.seq.append(_idx)
-          else:
-            self.hard_mining_reset()
-          print('seq len', len(self.seq))
-        else:
-          if self.shuffle:
-              random.shuffle(self.seq)
-        if self.seq is None and self.imgrec is not None:
-            self.imgrec.reset()
-
-    def num_samples(self):
-      return len(self.seq)
-
-    def next_sample(self):
-      while True:
-        if self.cur >= len(self.seq):
-            raise StopIteration
-        idx = self.seq[self.cur]
-        self.cur += 1
-        s = self.imgrec.read_idx(idx)
-        header, img = recordio.unpack(s)
-        label = header.label
-        if not isinstance(label, numbers.Number):
-          label = label[0]
-        return label, img, None, None
-
-    def brightness_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      src *= alpha
-      return src
-
-    def contrast_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = (3.0 * (1.0 - alpha) / gray.size) * np.sum(gray)
-      src *= alpha
-      src += gray
-      return src
-
-    def saturation_aug(self, src, x):
-      alpha = 1.0 + random.uniform(-x, x)
-      coef = np.array([[[0.299, 0.587, 0.114]]])
-      gray = src * coef
-      gray = np.sum(gray, axis=2, keepdims=True)
-      gray *= (1.0 - alpha)
-      src *= alpha
-      src += gray
-      return src
-
-    def color_aug(self, img, x):
-      augs = [self.brightness_aug, self.contrast_aug, self.saturation_aug]
-      random.shuffle(augs)
-      for aug in augs:
-        #print(img.shape)
-        img = aug(img, x)
-        #print(img.shape)
-      return img
-
-    def mirror_aug(self, img):
-      _rd = random.randint(0,1)
-      if _rd==1:
-        for c in xrange(img.shape[2]):
-          img[:,:,c] = np.fliplr(img[:,:,c])
-      return img
-
-
-    def next(self):
-        if not self.is_init:
-          self.reset()
-          self.is_init = True
-        """Returns the next batch of data."""
-        #print('in next', self.cur, self.labelcur)
-        self.nbatch+=1
-        batch_size = self.batch_size
-        c, h, w = self.data_shape
-        batch_data = nd.empty((batch_size, c, h, w))
-        if self.provide_label is not None:
-          batch_label = nd.empty(self.provide_label[0][1])
-        i = 0
-        try:
-            while i < batch_size:
-                label, s, bbox, landmark = self.next_sample()
-                _data = self.imdecode(s)
-                if self.rand_mirror:
-                  _rd = random.randint(0,1)
-                  if _rd==1:
-                    _data = mx.ndarray.flip(data=_data, axis=1)
-                if self.cutoff>0:
-                  centerh = random.randint(0, _data.shape[0]-1)
-                  centerw = random.randint(0, _data.shape[1]-1)
-                  half = self.cutoff//2
-                  starth = max(0, centerh-half)
-                  endh = min(_data.shape[0], centerh+half)
-                  startw = max(0, centerw-half)
-                  endw = min(_data.shape[1], centerw+half)
-                  _data = _data.astype('float32')
-                  #print(starth, endh, startw, endw, _data.shape)
-                  _data[starth:endh, startw:endw, :] = 127.5
-                #_npdata = _data.asnumpy()
-                #if landmark is not None:
-                #  _npdata = face_preprocess.preprocess(_npdata, bbox = bbox, landmark=landmark, image_size=self.image_size)
-                #if self.rand_mirror:
-                #  _npdata = self.mirror_aug(_npdata)
-                #if self.mean is not None:
-                #  _npdata = _npdata.astype(np.float32)
-                #  _npdata -= self.mean
-                #  _npdata *= 0.0078125
-                #nimg = np.zeros(_npdata.shape, dtype=np.float32)
-                #nimg[self.patch[1]:self.patch[3],self.patch[0]:self.patch[2],:] = _npdata[self.patch[1]:self.patch[3], self.patch[0]:self.patch[2], :]
-                #_data = mx.nd.array(nimg)
-                data = [_data]
-                try:
-                    self.check_valid_image(data)
-                except RuntimeError as e:
-                    logging.debug('Invalid image, skipping:  %s', str(e))
-                    continue
-                #print('aa',data[0].shape)
-                #data = self.augmentation_transform(data)
-                #print('bb',data[0].shape)
-                for datum in data:
-                    assert i < batch_size, 'Batch size must be multiples of augmenter output length'
-                    #print(datum.shape)
-                    batch_data[i][:] = self.postprocess_data(datum)
-                    if self.provide_label is not None:
-                      batch_label[i][:] = label
-                    i += 1
-        except StopIteration:
-            if i<batch_size:
-                raise StopIteration
-
-        #print('next end', batch_size, i)
-        _label = None
-        if self.provide_label is not None:
-          _label = [batch_label]
-        return io.DataBatch([batch_data], _label, batch_size - i)
-
-    def check_data_shape(self, data_shape):
-        """Checks if the input data shape is valid"""
-        if not len(data_shape) == 3:
-            raise ValueError('data_shape should have length 3, with dimensions CxHxW')
-        if not data_shape[0] == 3:
-            raise ValueError('This iterator expects inputs to have 3 channels.')
-
-    def check_valid_image(self, data):
-        """Checks if the input data is valid"""
-        if len(data[0].shape) == 0:
-            raise RuntimeError('Data shape is wrong')
-
-    def imdecode(self, s):
-        """Decodes a string or byte string to an NDArray.
-        See mx.img.imdecode for more details."""
-        img = mx.image.imdecode(s) #mx.ndarray
-        return img
-
-    def read_image(self, fname):
-        """Reads an input image `fname` and returns the decoded raw bytes.
-
-        Example usage:
-        ----------
-        >>> dataIter.read_image('Face.jpg') # returns decoded raw bytes.
-        """
-        with open(os.path.join(self.path_root, fname), 'rb') as fin:
-            img = fin.read()
-        return img
-
-    def augmentation_transform(self, data):
-        """Transforms input data with specified augmentation."""
-        for aug in self.auglist:
-            data = [ret for src in data for ret in aug(src)]
-        return data
-
-    def postprocess_data(self, datum):
-        """Final postprocessing step before image is loaded into the batch."""
-        return nd.transpose(datum, axes=(2, 0, 1))
-
-class FaceImageIterList(io.DataIter):
-  def __init__(self, iter_list):
-    assert len(iter_list)>0
-    self.provide_data = iter_list[0].provide_data
-    self.provide_label = iter_list[0].provide_label
-    self.iter_list = iter_list
-    self.cur_iter = None
-
-  def reset(self):
-    self.cur_iter.reset()
-
-  def next(self):
-    self.cur_iter = random.choice(self.iter_list)
-    while True:
-      try:
-        ret = self.cur_iter.next()
-      except StopIteration:
-        self.cur_iter.reset()
-        continue
-      return ret
-
-
diff --git a/src/utils/benchmark.py b/src/utils/benchmark.py
deleted file mode 100644
index 057194d8e..000000000
--- a/src/utils/benchmark.py
+++ /dev/null
@@ -1,101 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-import os
-import sys
-import datetime
-import mxnet as mx
-from mxnet import ndarray as nd
-import random
-import argparse
-import cv2
-import time
-import sklearn
-from sklearn.decomposition import PCA
-from easydict import EasyDict as edict
-from sklearn.cluster import DBSCAN
-import numpy as np
-sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
-import face_image
-
-def ch_dev(arg_params, aux_params, ctx):
-  new_args = dict()
-  new_auxs = dict()
-  for k, v in arg_params.items():
-    new_args[k] = v.as_in_context(ctx)
-  for k, v in aux_params.items():
-    new_auxs[k] = v.as_in_context(ctx)
-  return new_args, new_auxs
-
-
-def main(args):
-  ctx = mx.gpu(args.gpu)
-  args.ctx_num = 1
-  prop = face_image.load_property(args.data)
-  image_size = prop.image_size
-  print('image_size', image_size)
-  vec = args.model.split(',')
-  prefix = vec[0]
-  epoch = int(vec[1])
-  print('loading',prefix, epoch)
-  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
-  arg_params, aux_params = ch_dev(arg_params, aux_params, ctx)
-  all_layers = sym.get_internals()
-  sym = all_layers['fc1_output']
-  #model = mx.mod.Module.load(prefix, epoch, context = ctx)
-  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
-  #model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
-  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
-  model.set_params(arg_params, aux_params)
-  path_imgrec = os.path.join(args.data, 'train.rec')
-  path_imgidx = os.path.join(args.data, 'train.idx')
-  imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r')  # pylint: disable=redefined-variable-type
-  s = imgrec.read_idx(0)
-  header, _ = mx.recordio.unpack(s)
-  assert header.flag>0
-  print('header0 label', header.label)
-  header0 = (int(header.label[0]), int(header.label[1]))
-  #assert(header.flag==1)
-  imgidx = range(1, int(header.label[0]))
-  stat = []
-  count = 0
-  data = nd.zeros( (1 ,3, image_size[0], image_size[1]) )
-  label = nd.zeros( (1,) )
-  for idx in imgidx:
-    if len(stat)%100==0:
-      print('processing', len(stat))
-    s = imgrec.read_idx(idx)
-    header, img = mx.recordio.unpack(s)
-    img = mx.image.imdecode(img)
-    img = nd.transpose(img, axes=(2, 0, 1))
-    data[0][:] = img
-    #input_blob = np.expand_dims(img.asnumpy(), axis=0)
-    #arg_params["data"] = mx.nd.array(input_blob, ctx)
-    #arg_params["softmax_label"] = mx.nd.empty((1,), ctx)
-    time_now = datetime.datetime.now()
-    #exe = sym.bind(ctx, arg_params ,args_grad=None, grad_req="null", aux_states=aux_params)
-    #exe.forward(is_train=False)
-    #_embedding = exe.outputs[0].asnumpy().flatten()
-    #db = mx.io.DataBatch(data=(data,), label=(label,))
-    db = mx.io.DataBatch(data=(data,))
-    model.forward(db, is_train=False)
-    net_out = model.get_outputs()[0].asnumpy()
-    time_now2 = datetime.datetime.now()
-    diff = time_now2 - time_now
-    stat.append(diff.total_seconds())
-    if len(stat)==args.param1:
-      break
-  stat = stat[10:]
-  print('avg infer time', np.mean(stat))
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description='do network benchmark')
-  # general
-  parser.add_argument('--gpu', default=0, type=int, help='')
-  parser.add_argument('--data', default='', type=str, help='')
-  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
-  parser.add_argument('--batch-size', default=1, type=int, help='')
-  parser.add_argument('--param1', default=1010, type=int, help='')
-  args = parser.parse_args()
-  main(args)
-
diff --git a/tools/onnx2caffe/LICENSE b/tools/onnx2caffe/LICENSE
new file mode 100644
index 000000000..879e87b38
--- /dev/null
+++ b/tools/onnx2caffe/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2019 MTlab, Meitu Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/tools/onnx2caffe/MyCaffe.py b/tools/onnx2caffe/MyCaffe.py
new file mode 100644
index 000000000..0cbaf7b72
--- /dev/null
+++ b/tools/onnx2caffe/MyCaffe.py
@@ -0,0 +1,125 @@
+from collections import OrderedDict, Counter
+
+from caffe.proto import caffe_pb2
+from google import protobuf
+import six
+
+def param_name_dict():
+    """Find out the correspondence between layer names and parameter names."""
+
+    layer = caffe_pb2.LayerParameter()
+    # get all parameter names (typically underscore case) and corresponding
+    # type names (typically camel case), which contain the layer names
+    # (note that not all parameters correspond to layers, but we'll ignore that)
+    param_names = [f.name for f in layer.DESCRIPTOR.fields if f.name.endswith('_param')]
+    param_type_names = [type(getattr(layer, s)).__name__ for s in param_names]
+    # strip the final '_param' or 'Parameter'
+    param_names = [s[:-len('_param')] for s in param_names]
+    param_type_names = [s[:-len('Parameter')] for s in param_type_names]
+    return dict(zip(param_type_names, param_names))
+
+def assign_proto(proto, name, val):
+    """Assign a Python object to a protobuf message, based on the Python
+    type (in recursive fashion). Lists become repeated fields/messages, dicts
+    become messages, and other types are assigned directly. For convenience,
+    repeated fields whose values are not lists are converted to single-element
+    lists; e.g., `my_repeated_int_field=3` is converted to
+    `my_repeated_int_field=[3]`."""
+
+    is_repeated_field = hasattr(getattr(proto, name), 'extend')
+    if is_repeated_field and not isinstance(val, list):
+        val = [val]
+    if isinstance(val, list):
+        if isinstance(val[0], dict):
+            for item in val:
+                proto_item = getattr(proto, name).add()
+                for k, v in six.iteritems(item):
+                    assign_proto(proto_item, k, v)
+        else:
+            getattr(proto, name).extend(val)
+    elif isinstance(val, dict):
+        for k, v in six.iteritems(val):
+            assign_proto(getattr(proto, name), k, v)
+    else:
+        setattr(proto, name, val)
+
+class Function(object):
+    """A Function specifies a layer, its parameters, and its inputs (which
+    are Tops from other layers)."""
+
+    def __init__(self, type_name, layer_name, inputs,outputs, **params):
+        self.type_name = type_name
+        self.inputs = inputs
+        self.outputs = outputs
+        self.params = params
+        self.layer_name = layer_name
+        self.ntop = self.params.get('ntop', 1)
+        # use del to make sure kwargs are not double-processed as layer params
+        if 'ntop' in self.params:
+            del self.params['ntop']
+        self.in_place = self.params.get('in_place', False)
+        if 'in_place' in self.params:
+            del self.params['in_place']
+        # self.tops = tuple(Top(self, n) for n in range(self.ntop))l
+
+    def _get_name(self, names, autonames):
+        if self not in names and self.ntop > 0:
+            names[self] = self._get_top_name(self.tops[0], names, autonames)
+        elif self not in names:
+            autonames[self.type_name] += 1
+            names[self] = self.type_name + str(autonames[self.type_name])
+        return names[self]
+
+    def _get_top_name(self, top, names, autonames):
+        if top not in names:
+            autonames[top.fn.type_name] += 1
+            names[top] = top.fn.type_name + str(autonames[top.fn.type_name])
+        return names[top]
+
+    def _to_proto(self):
+        bottom_names = []
+        for inp in self.inputs:
+            # inp._to_proto(layers, names, autonames)
+            bottom_names.append(inp)
+        layer = caffe_pb2.LayerParameter()
+        layer.type = self.type_name
+        layer.bottom.extend(bottom_names)
+
+        if self.in_place:
+            layer.top.extend(layer.bottom)
+        else:
+            for top in self.outputs:
+                layer.top.append(top)
+        layer.name = self.layer_name
+        # print(self.type_name + "...")
+        for k, v in six.iteritems(self.params):
+            # special case to handle generic *params
+            # print("generating "+k+"...")
+
+            if k.endswith('param'):
+                assign_proto(layer, k, v)
+            else:
+                try:
+                    assign_proto(getattr(layer,
+                        _param_names[self.type_name] + '_param'), k, v)
+                except (AttributeError, KeyError):
+                    assign_proto(layer, k, v)
+
+        return layer
+
+class Layers(object):
+    """A Layers object is a pseudo-module which generates functions that specify
+    layers; e.g., Layers().Convolution(bottom, kernel_size=3) will produce a Top
+    specifying a 3x3 convolution applied to bottom."""
+
+    def __getattr__(self, name):
+        def layer_fn(*args, **kwargs):
+            fn = Function(name, args, kwargs)
+            return fn
+        return layer_fn
+
+
+
+
+_param_names = param_name_dict()
+
diff --git a/tools/onnx2caffe/README.md b/tools/onnx2caffe/README.md
new file mode 100644
index 000000000..02b8238fe
--- /dev/null
+++ b/tools/onnx2caffe/README.md
@@ -0,0 +1,37 @@
+# Convert ONNX to Caffe
+
+This tool is modified from [onnx2caffe](https://github.com/MTlab/onnx2caffe) by MTlab.
+
+We added some OPs to support one-stage mmdetection models.
+
+### Dependencies
+* pycaffe (with builtin Upsample and Permute layers)
+* onnx  
+
+
+### How to use
+To convert onnx model to caffe:
+```
+python convertCaffe.py ./model/mmdet.onnx ./model/a.prototxt ./model/a.caffemodel
+```
+
+### Current support operation
+* Conv
+* ConvTranspose
+* BatchNormalization
+* MaxPool
+* AveragePool
+* Relu
+* Sigmoid
+* Dropout
+* Gemm (InnerProduct only)
+* Add
+* Mul
+* Reshape
+* Upsample
+* Concat
+* Flatten
+* **Resize**
+* **Permute**
+* **Scale**
+
diff --git a/tools/onnx2caffe/convertCaffe.py b/tools/onnx2caffe/convertCaffe.py
new file mode 100644
index 000000000..cd5277663
--- /dev/null
+++ b/tools/onnx2caffe/convertCaffe.py
@@ -0,0 +1,114 @@
+#from __future__ import print_function
+import sys
+import caffe
+import onnx
+import numpy as np
+from caffe.proto import caffe_pb2
+caffe.set_mode_cpu()
+from onnx2caffe._transformers import ConvAddFuser,ConstantsToInitializers
+from onnx2caffe._graph import Graph
+
+import onnx2caffe._operators as cvt
+import onnx2caffe._weightloader as wlr
+from onnx2caffe._error_utils import ErrorHandling
+from collections import OrderedDict
+from onnx import shape_inference
+import importlib
+
+USE_DECONV_AS_UPSAMPLE = True
+
+transformers = [
+    ConstantsToInitializers(),
+    ConvAddFuser(),
+]
+
+def convertToCaffe(graph, prototxt_save_path, caffe_model_save_path):
+    exist_edges = []
+    layers = []
+    exist_nodes = []
+    err = ErrorHandling()
+    for i in graph.inputs:
+        edge_name = i[0]
+        input_layer = cvt.make_input(i)
+        layers.append(input_layer)
+        exist_edges.append(i[0])
+        graph.channel_dims[edge_name] = graph.shape_dict[edge_name][1]
+
+
+    for id, node in enumerate(graph.nodes):
+        node_name = node.name
+        op_type = node.op_type
+        inputs = node.inputs
+        inputs_tensor = node.input_tensors
+        input_non_exist_flag = False
+
+        for inp in inputs:
+            if inp not in exist_edges and inp not in inputs_tensor:
+                input_non_exist_flag = True
+                break
+        if input_non_exist_flag:
+            continue
+
+        if op_type not in cvt._ONNX_NODE_REGISTRY:
+            err.unsupported_op(node)
+            continue
+        converter_fn = cvt._ONNX_NODE_REGISTRY[op_type]
+        layer = converter_fn(node,graph,err)
+        if type(layer)==tuple:
+            for l in layer:
+                layers.append(l)
+        else:
+            layers.append(layer)
+        outs = node.outputs
+        for out in outs:
+            exist_edges.append(out)
+
+    net = caffe_pb2.NetParameter()
+    for id,layer in enumerate(layers):
+        layers[id] = layer._to_proto()
+    net.layer.extend(layers)
+
+    with open(prototxt_save_path, 'w') as f:
+        print(net,file=f)
+
+    caffe.set_mode_cpu()
+    deploy = prototxt_save_path
+    net = caffe.Net(deploy,
+                    caffe.TEST)
+
+    for id, node in enumerate(graph.nodes):
+        node_name = node.name
+        op_type = node.op_type
+        inputs = node.inputs
+        inputs_tensor = node.input_tensors
+        input_non_exist_flag = False
+        if op_type not in wlr._ONNX_NODE_REGISTRY:
+            err.unsupported_op(node)
+            continue
+        converter_fn = wlr._ONNX_NODE_REGISTRY[op_type]
+        converter_fn(net, node, graph, err)
+
+    net.save(caffe_model_save_path)
+    return net
+
+def getGraph(onnx_path):
+    model = onnx.load(onnx_path)
+    output_names = [node.name for node in model.graph.output]
+    model = shape_inference.infer_shapes(model)
+    model_graph = model.graph
+    graph = Graph.from_onnx(model_graph)
+    graph = graph.transformed(transformers)
+    graph.channel_dims = {}
+
+    return graph, output_names
+
+if __name__ == "__main__":
+    cvt.USE_DECONV_AS_UPSAMPLE = USE_DECONV_AS_UPSAMPLE
+    wlr.USE_DECONV_AS_UPSAMPLE = USE_DECONV_AS_UPSAMPLE
+    onnx_path = sys.argv[1]
+    prototxt_path = sys.argv[2]
+    caffemodel_path = sys.argv[3]
+    graph, output_names = getGraph(onnx_path)
+    convertToCaffe(graph, prototxt_path, caffemodel_path)
+    print('output_names:', output_names)
+
diff --git a/tools/onnx2caffe/onnx2caffe/__init__.py b/tools/onnx2caffe/onnx2caffe/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/tools/onnx2caffe/onnx2caffe/_error_utils.py b/tools/onnx2caffe/onnx2caffe/_error_utils.py
new file mode 100644
index 000000000..46f34c05d
--- /dev/null
+++ b/tools/onnx2caffe/onnx2caffe/_error_utils.py
@@ -0,0 +1,64 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from typing import Dict, Text, Any, Callable
+from ._graph import Node, Graph
+
+class ErrorHandling(object):
+  '''
+  To handle errors and addition of custom layers
+  '''
+
+  def __init__(self,
+               add_custom_layers = False, # type: bool
+               custom_conversion_functions = dict(), # type: Dict[Text, Any]
+               custom_layer_nodes = [], # type : List[Node]
+               ):
+      # type: (...) -> None
+      self.add_custom_layers = add_custom_layers
+      self.custom_conversion_functions = custom_conversion_functions
+      self.custom_layer_nodes = custom_layer_nodes
+
+
+  def unsupported_op(self,
+                     node,  # type: Node
+                    ):
+      # type: (...) -> Callable[[Any, Node, Graph, ErrorHandling], None]
+      '''
+      Either raise an error for an unsupported op type or return custom layer add function
+      '''
+      if self.add_custom_layers:
+        from ._operators import _convert_custom
+        return _convert_custom
+      else:
+        raise TypeError(
+          "ONNX node of type {} is not supported.\n".format(node.op_type,)
+        )
+
+
+  def unsupported_op_configuration(self,
+                                   node, # type: Node
+                                   err_message, # type: Text
+                                   ):
+      raise TypeError(
+        "Error while converting op of type: {}. Error message: {}\n".format(node.op_type, err_message, )
+      )
+
+
+  def missing_initializer(self,
+                          node, # type: Node
+                          err_message, # type: Text
+                          ):
+      # type: (...) -> None
+      '''
+      Missing initializer error
+      '''
+      raise ValueError(
+        "Missing initializer error in op of type {}, with input name = {}, "
+        "output name = {}. Error message: {}\n".
+        format(node.op_type, node.inputs[0], node.outputs[0], err_message)
+      )
+
+
+
diff --git a/tools/onnx2caffe/onnx2caffe/_graph.py b/tools/onnx2caffe/onnx2caffe/_graph.py
new file mode 100644
index 000000000..89e81dbc8
--- /dev/null
+++ b/tools/onnx2caffe/onnx2caffe/_graph.py
@@ -0,0 +1,225 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+from onnx import numpy_helper, ValueInfoProto, AttributeProto, GraphProto, NodeProto, TensorProto, TensorShapeProto
+from typing import Any, Text, Iterable, List, Dict, Sequence, Optional, Tuple, Union
+from typing_extensions import Protocol
+import numpy as np
+
+
+class Transformer(Protocol):
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        pass
+
+
+EdgeInfo = Tuple[Text, Any, TensorShapeProto]
+AttributeValue = Any # TODO Union[Sequence[float], Sequence[int], Sequence[Text], Sequence[TensorProto], Sequence[GraphProto]]
+
+def _input_from_onnx_input(input):  # type: (ValueInfoProto) -> EdgeInfo
+    name = input.name
+    type = input.type.tensor_type.elem_type
+    shape = tuple([d.dim_value for d in input.type.tensor_type.shape.dim])
+    return (name, type, shape)
+
+
+def _convertAttributeProto(onnx_arg):  # type: (AttributeProto) -> AttributeValue
+    """
+    Convert an ONNX AttributeProto into an appropriate Python object
+    for the type.
+    NB: Tensor attribute gets returned as numpy array
+    """
+    if onnx_arg.HasField('f'):
+        return onnx_arg.f
+    elif onnx_arg.HasField('i'):
+        return onnx_arg.i
+    elif onnx_arg.HasField('s'):
+        return onnx_arg.s
+    elif onnx_arg.HasField('t'):
+        return numpy_helper.to_array(onnx_arg.t)
+    elif len(onnx_arg.floats):
+        return list(onnx_arg.floats)
+    elif len(onnx_arg.ints):
+        return list(onnx_arg.ints)
+    elif len(onnx_arg.strings):
+        return list(onnx_arg.strings)
+    else:
+        raise ValueError("Unsupported ONNX attribute: {}".format(onnx_arg))
+
+
+class Attributes(Dict[Text, Any]):
+    @staticmethod
+    def from_onnx(args):  # type: (Iterable[AttributeProto]) -> Attributes
+        d = Attributes()
+        for arg in args:
+            d[arg.name] = _convertAttributeProto(arg)
+        return d
+
+
+class Node(object):
+    def __init__(self,
+                 name,  # type: Optional[Text]
+                 op_type,  # type: Text
+                 attrs,  # type: Dict[Text, AttributeValue]
+                 inputs,  # type: List[Text]
+                 outputs,  # type: List[Text]
+                 ):
+        # type: (...) -> None
+        self.name = name
+        self.op_type = op_type
+        self.attrs = attrs
+        self.inputs = inputs
+        self.outputs = outputs
+        self.input_tensors = {}  # type: Dict[Text, np._ArrayLike[Any]]
+        self.parents = []  # type: List[Node]
+        self.children = []  # type: List[Node]
+        self.metadata = {}  # type: Dict[Any, Any]
+
+    def add_parent(self, parent_node):  # type: (Node) -> None
+        assert parent_node not in self.parents
+        self.parents.append(parent_node)
+        if self not in parent_node.children:
+            parent_node.children.append(self)
+
+    def add_child(self, child_node):  # type: (Node) -> None
+        assert child_node not in self.children
+        self.children.append(child_node)
+        if self not in child_node.parents:
+            child_node.parents.append(self)
+
+    def get_only_parent(self):  # type: () -> Node
+        if len(self.parents) != 1:
+            raise ValueError('Node ({}) expected to have 1 parent. Found {}.'
+                             .format(self, len(self.parents)))
+        return self.parents[0]
+
+    @staticmethod
+    def from_onnx(node):  # type: (NodeProto) -> Node
+        attrs = Attributes.from_onnx(node.attribute)
+        name = Text(node.name)
+        if len(name) == 0:
+            name = "_".join(node.output)
+        return Node(
+            name, node.op_type, attrs, list(node.input), list(node.output)
+        )
+
+
+class Graph(object):
+    def __init__(self,
+                 nodes,  # type: List[Node]
+                 inputs,  # type: List[EdgeInfo]
+                 outputs,  # type: List[EdgeInfo]
+                 shape_dict, # type: Dict[Text,Tuple[int,...]]
+                 ):
+        # type: (...) -> None
+        self.nodes = nodes
+        self.inputs = inputs
+        self.outputs = outputs
+        self.shape_dict = shape_dict  # data blob name to its shape
+
+        # data blob name to the list of op types it feeds into
+        self.blob_to_op_type = {} # type: Dict[Text, List[Text]]
+        # data blob name to the op_type that generates it
+        self.blob_from_op_type = {}  # type: Dict[Text, Text]
+
+        for node_ in nodes:
+            for input_ in node_.inputs:
+                if input_ in self.blob_to_op_type:
+                    self.blob_to_op_type[input_].append(node_.op_type)
+                else:
+                    self.blob_to_op_type[input_] = [node_.op_type]
+            for output_ in node_.outputs:
+                if output_ in self.blob_from_op_type:
+                    raise ValueError("Data blob: %s, is generated by more than 1 op" %(output_))
+                self.blob_from_op_type[output_] = node_.op_type
+
+
+    def transformed(self, transformers):  # type: (Iterable[Transformer]) -> Graph
+        graph = self
+        for transformer in transformers:
+            graph = transformer(graph)
+        return graph
+
+    def has_edge_name(self, name):  # type: (Text) -> bool
+        '''
+        Check if name is already used for graph inputs/outputs or for nodes
+        inputs/outputs
+        '''
+        names = set()
+        for input in self.inputs:
+            names.add(input[0])
+        for output in self.outputs:
+            names.add(output[0])
+        for node in self.nodes:
+            names.update(node.inputs)
+            names.update(node.outputs)
+        return name in names
+
+    def get_unique_edge_name(self, name):  # type: (Text) -> Text
+        n_ = name
+        i = 0
+        while self.has_edge_name(n_):
+            n_ = "{}_{}".format(name, i)
+            i += 1
+        return n_
+
+    @staticmethod
+    def from_onnx(graph):  # type: (GraphProto) -> Graph
+        input_tensors = {
+            t.name: numpy_helper.to_array(t) for t in graph.initializer
+        }
+        nodes_ = []
+        nodes_by_input = {}  # type: Dict[Text, List[Node]]
+        nodes_by_output = {}
+        for node in graph.node:
+            node_ = Node.from_onnx(node)
+            for input_ in node_.inputs:
+                if input_ in input_tensors:
+                    node_.input_tensors[input_] = input_tensors[input_]
+                else:
+                    if input_ in nodes_by_input:
+                        input_nodes = nodes_by_input[input_]
+                    else:
+                        input_nodes = []
+                        nodes_by_input[input_] = input_nodes
+                    input_nodes.append(node_)
+            for output_ in node_.outputs:
+                nodes_by_output[output_] = node_
+            nodes_.append(node_)
+
+        inputs = []
+        for i in graph.input:
+            if i.name not in input_tensors:
+                inputs.append(_input_from_onnx_input(i))
+
+        outputs = []
+        for o in graph.output:
+            outputs.append(_input_from_onnx_input(o))
+
+        for node_ in nodes_:
+            for input_ in node_.inputs:
+                if input_ in nodes_by_output:
+                    node_.parents.append(nodes_by_output[input_])
+            for output_ in node_.outputs:
+                if output_ in nodes_by_input:
+                    node_.children.extend(nodes_by_input[output_])
+
+        # Dictionary to hold the "value_info" field from ONNX graph
+        shape_dict = {} # type: Dict[Text,Tuple[int,...]]
+
+        def extract_value_info(shape_dict, # type: Dict[Text,Tuple[int,...]]
+                               value_info, # type: ValueInfoProto[...]
+                               ):
+            # type: (...) -> None
+            shape_dict[value_info.name] = tuple([int(dim.dim_value) for dim in value_info.type.tensor_type.shape.dim])
+
+        for value_info in graph.value_info:
+            extract_value_info(shape_dict, value_info)
+        for value_info in graph.input:
+            extract_value_info(shape_dict, value_info)
+        for value_info in graph.output:
+            extract_value_info(shape_dict, value_info)
+
+
+        return Graph(nodes_, inputs, outputs, shape_dict)
diff --git a/tools/onnx2caffe/onnx2caffe/_operators.py b/tools/onnx2caffe/onnx2caffe/_operators.py
new file mode 100644
index 000000000..815ff963a
--- /dev/null
+++ b/tools/onnx2caffe/onnx2caffe/_operators.py
@@ -0,0 +1,485 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+from caffe import params as P
+import math
+import numpy as np
+from ._graph import Node, Graph
+from MyCaffe import Function as myf
+
+USE_DECONV_AS_UPSAMPLE = False
+
+def _compare(a, b, encoding="utf8"): #type: (Text, Text, Text) -> bool
+    if isinstance(a, bytes):
+        a = a.decode(encoding)
+    if isinstance(b, bytes):
+        b = b.decode(encoding)
+    return a == b
+
+def make_input(input):
+    name = input[0]
+    output = input[0]
+    output = [output]
+    shape = input[2]
+    shape = list(shape)
+    input_layer = myf("Input", name, [], output, input_param=dict(shape=dict(dim=shape)))
+    return input_layer
+
+def _convert_conv(node, graph, err):
+    weight_name = node.inputs[1]
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    node_name = node.name
+    W = None
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name,))
+    is_deconv = False
+    if node.op_type.endswith("Transpose"):
+        is_deconv = True
+    bias_flag = False
+    bias = None
+    if len(node.inputs) > 2:
+        bias = node.input_tensors[node.inputs[2]]
+        bias_flag = True
+    dilations = node.attrs.get("dilations", [1, 1])
+    # groups = 1
+    groups = node.attrs.get("group", 1)
+    kernel_shape = node.attrs["kernel_shape"]
+    pads = node.attrs.get("pads", [0, 0, 0, 0])
+    strides = node.attrs["strides"]
+
+    layer = myf("Convolution", node_name, [input_name], [output_name],
+                kernel_h = kernel_shape[0],kernel_w = kernel_shape[1],
+                stride_h=strides[0], stride_w = strides[1], group = groups,
+                pad_h = pads[0], pad_w = pads[1],
+                num_output=W.shape[0],  dilation = dilations[0], bias_term = bias_flag)
+
+    graph.channel_dims[output_name] = W.shape[0]
+    return layer
+
+def _convert_relu(node,graph,err):
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    name = str(node.name)
+
+    if input_name==output_name:
+        inplace = True
+    else:
+        inplace = False
+
+    layer = myf("ReLU",name,[input_name],[output_name],in_place=inplace)
+    # l_top_relu1 = L.ReLU(l_bottom, name=name, in_place=True)
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+
+    return layer
+
+def _convert_sigmoid(node,graph,err):
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    name = str(node.name)
+
+    if input_name==output_name:
+        inplace = True
+    else:
+        inplace = False
+
+    layer = myf("Sigmoid",name,[input_name],[output_name],in_place=inplace)
+    # l_top_relu1 = L.ReLU(l_bottom, name=name, in_place=True)
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+
+    return layer
+
+def _convert_BatchNorm(node,graph,err):
+    epsilon = node.attrs.get("epsilon", 1e-5)
+    scale = node.input_tensors[node.inputs[1]]
+    bias = node.input_tensors[node.inputs[2]]
+    mean = node.input_tensors[node.inputs[3]]
+    var = node.input_tensors[node.inputs[4]]
+    node_name = node.name
+
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+
+    if input_name==output_name:
+        inplace = True
+    else:
+        inplace = False
+
+    bn_layer = myf("BatchNorm", node_name+"_bn",[input_name],[output_name],eps = epsilon, use_global_stats = True, in_place=inplace)
+    scale_layer = myf("Scale", node_name, [output_name],[output_name],in_place=True,bias_term=True)
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+
+    return bn_layer,scale_layer
+
+def _convert_Add(node,graph,err):
+    input_name_list = [str(i) for i in node.inputs]
+    output_name = str(node.outputs[0])
+    node_name = node.name
+
+    max_dim = 0
+    for name in input_name_list:
+        if graph.channel_dims[name]>max_dim:
+            max_dim = graph.channel_dims[name]
+
+    if 'broadcast' in node.attrs:
+        if node.attrs['broadcast'] == 1:
+            input_node_number = len(input_name_list)
+            if input_node_number !=2:
+                return err.unsupported_op_configuration(node, "Broadcast Add must has 2 input, not {}".format(input_node_number))
+            axis = node.attrs['axis']
+            flat_layer = myf("Flatten",node_name+'_flat',[input_name_list[1]],[output_name+'_flat'])
+            layer = myf("Bias", node_name, [input_name_list[0],output_name+'_flat'], [output_name], axis = axis)
+            # layer = myf("Bias", node_name, input_name_list, [output_name], bias_term = False, axis = axis)
+            graph.channel_dims[output_name] = graph.channel_dims[input_name_list[0]]
+            return flat_layer,layer
+
+    layer = myf("Eltwise",node_name,input_name_list,[output_name],operation=P.Eltwise.SUM)
+    graph.channel_dims[output_name] = graph.channel_dims[input_name_list[0]]
+    return layer
+
+def _convert_Mul(node,graph,err):
+    input_name_list = [str(i) for i in node.inputs]
+    output_name = str(node.outputs[0])
+    node_name = node.name
+    print('Mul:', node.name, node.attrs, input_name_list, output_name)
+    if len(node.attrs)==0:
+        assert len(node.input_tensors)==1
+        assert len(input_name_list)==2
+        inp_tensor = node.input_tensors[input_name_list[1]]
+        scale_value = float(inp_tensor)
+        print(scale_value)
+        layer = myf("Scale", node_name, [input_name_list[0]], [output_name], bias_term = False, 
+                scale_param = dict(filler = dict(value=scale_value), bias_term=False))
+        return layer
+        #layer = myf("Reshape", node_name, [input_name], [output_name], reshape_param = dict(shape=dict(dim=list(shape))))
+    #print(len(node.input_tensors))
+
+    # max_dim = 0
+    # for name in input_name_list:
+    #     if graph.channel_dims[name]>max_dim:
+    #         max_dim = graph.channel_dims[name]
+
+    if 'broadcast' in node.attrs:
+        if node.attrs['broadcast'] == 1:
+            input_node_number = len(input_name_list)
+            if input_node_number !=2:
+                return err.unsupported_op_configuration(node, "Broadcast Mul must has 2 input, not {}".format(input_node_number))
+            axis = node.attrs['axis']
+            flat_layer = myf("Flatten",node_name+'_flat',[input_name_list[1]],[output_name+'_flat'])
+            layer = myf("Scale", node_name, [input_name_list[0],output_name+'_flat'], [output_name], bias_term = False, axis = axis)
+            graph.channel_dims[output_name] = graph.channel_dims[input_name_list[0]]
+            return flat_layer,layer
+
+    layer = myf("Eltwise",node_name,input_name_list,[output_name],operation=P.Eltwise.PROD)
+    graph.channel_dims[output_name] = graph.channel_dims[input_name_list[0]]
+    return layer
+
+def _convert_Reshape(node,graph,err):
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    if len(node.inputs)==1:
+        shape = tuple(node.attrs.get('shape', ()))
+    else:
+        shape = tuple(node.input_tensors[node.inputs[1]])
+    # if shape == ():
+
+    #print('reshape to', shape)
+
+    if input_name==output_name:
+        inplace = True
+    else:
+        inplace = False
+
+    graph.channel_dims[output_name] = shape[1]
+    layer = myf("Reshape", node_name, [input_name], [output_name], reshape_param = dict(shape=dict(dim=list(shape))))
+    return layer
+
+    #if len(shape) == 2:
+    #    layer = myf("Flatten",node_name,[input_name],[output_name],in_place=inplace)
+    #    graph.channel_dims[output_name] = shape[1]
+    #    return layer
+    #elif len(shape) == 4:
+    #    graph.channel_dims[output_name] = shape[1]
+    #    layer = myf("Reshape", node_name, [input_name], [output_name], reshape_param = dict(shape=dict(dim=list(shape))))
+    #    return layer
+    #else:
+    #    return err.unsupported_op_configuration(node, "Reshape dimention number shall be 2 or 4")
+
+def _convert_Flatten(node,graph,err):
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    # shape = tuple(node.attrs.get('shape', ()))
+    if input_name==output_name:
+        inplace = True
+    else:
+        inplace = False
+    layer = myf("Flatten", node_name, [input_name], [output_name], in_place=inplace)
+    # graph.channel_dims[output_name] = shape[1]
+    return layer
+
+def _convert_pool(node,graph,err):
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    if node.op_type.endswith("MaxPool"):
+        pool_type = P.Pooling.MAX
+    elif node.op_type.endswith("AveragePool"):
+        pool_type = P.Pooling.AVE
+    else:
+        return err.unsupported_op_configuration(node,  "Unsupported pool type")
+
+    kernel_shape = node.attrs["kernel_shape"]
+    strides = node.attrs.get('strides', [1, 1])
+    pads = node.attrs.get('pads', [0, 0, 0, 0])
+
+    layer = myf("Pooling",node_name,[input_name],[output_name],pooling_param = dict(pool = pool_type,
+                                                                                    kernel_h = kernel_shape[0],
+                                                                                    kernel_w = kernel_shape[1],
+                                                                                    stride_h = strides[0],
+                                                                                    stride_w = strides[1],
+                                                                                    pad_h = pads[0],
+                                                                                    pad_w = pads[1]))
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_dropout(node,graph,err):
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    ratio = node.attrs.get('ratio', 0.5)
+    layer = myf("Dropout", node_name, [input_name], [output_name], dropout_ratio =ratio)
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_gemm(node,graph,err):
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    weight_name = node.inputs[1]
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name, ))
+        return
+
+    if node.attrs["broadcast"] != 1 or node.attrs["transB"] != 1:
+        return err.unsupported_op_configuration(node,"Gemm is supported only for inner_product layer")
+
+    b = None
+    bias_flag = False
+    if len(node.inputs) > 2:
+        b = node.input_tensors[node.inputs[2]]
+
+    if len(W.shape) != 2 or (b is not None and len(b.shape) != 1):
+        return err.unsupported_op_configuration(node, "Gemm is supported only for inner_product layer")
+    if b is not None:
+        bias_flag = True
+        if W.shape[0] != b.shape[0]:
+            return err.unsupported_op_configuration(node,
+                                                    "Gemm is supported only for inner_product layer")
+
+    layer = myf("InnerProduct",node_name,[input_name],[output_name],num_output = W.shape[0],bias_term = bias_flag)
+    graph.channel_dims[output_name] = W.shape[0]
+
+    return layer
+
+def _convert_upsample(node,graph,err):
+    factor = int(node.attrs["height_scale"])
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    # input_shape = graph.shape_dict[input_name]
+    # channels = input_shape[1]
+    channels = graph.channel_dims[input_name]
+    pad = int(math.ceil((factor - 1) / 2.))
+    # layer = myf("Deconvolution", node_name, [input_name], [output_name],
+    #             kernel_size=2 * factor - factor % 2,
+    #             stride=factor, group=channels,
+    #             pad = pad, num_output=channels, bias_term = False)
+    mode = node.attrs["mode"]
+    #https://github.com/pytorch/pytorch/issues/6900
+    if mode=="bilinear":
+        layer = myf("Deconvolution", node_name, [input_name], [output_name],
+                    convolution_param=dict(
+                        num_output=channels,
+                        kernel_size=2 * factor - factor % 2,
+                        stride=factor,
+                        pad=pad,
+                        group=channels,
+                        bias_term=False,
+                        weight_filler=dict(type="bilinear_upsampling")
+                    ))
+    else:
+        layer = myf("Deconvolution", node_name, [input_name], [output_name],
+                    convolution_param=dict(
+                        num_output=channels,
+                        kernel_size=factor,
+                        stride=factor,
+                        group=channels,
+                        bias_term=False,
+                    ))
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_resize(node,graph,err):
+    if not USE_DECONV_AS_UPSAMPLE:
+        #print(node, graph)
+        node_name = node.name
+        input_name = str(node.inputs[0])
+        output_name = str(node.outputs[0])
+        #print(node.attrs, node_name, input_name, output_name)
+        layer = myf("Upsample", node_name, [input_name], [output_name],
+                    upsample_param=dict(
+                        scale = 2
+                    ))
+
+        graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    else:
+        print('add resize deconv operator')
+        factor = 2
+        node_name = node.name
+        input_name = str(node.inputs[0])
+        output_name = str(node.outputs[0])
+        # input_shape = graph.shape_dict[input_name]
+        # channels = input_shape[1]
+        channels = graph.channel_dims[input_name]
+        pad = int(math.ceil((factor - 1) / 2.))
+        layer = myf("Deconvolution", node_name, [input_name], [output_name],
+                    convolution_param=dict(
+                        num_output=channels,
+                        kernel_size=factor,
+                        stride=factor,
+                        group=channels,
+                        bias_term=False,
+                    ))
+        graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_transpose(node,graph,err):
+    #print(node, graph)
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    #print(node.attrs, node_name, input_name, output_name)
+    layer = myf("Permute", node_name, [input_name], [output_name],
+                permute_param=dict(
+                    order = node.attrs['perm']
+                ))
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_softmax(node,graph,err):
+    #print(node, graph)
+    node_name = node.name
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    #print(node.attrs, node_name, input_name, output_name)
+    layer = myf("Softmax", node_name, [input_name], [output_name],
+                softmax_param=dict(
+                    axis = node.attrs['axis']
+                ))
+
+    graph.channel_dims[output_name] = graph.channel_dims[input_name]
+    return layer
+
+def _convert_concat(node,graph,err):
+    node_name = node.name
+    input_name_list = [str(i) for i in node.inputs]
+    output_name = str(node.outputs[0])
+    axis = node.attrs.get("axis", 1)
+
+    layer = myf('Concat', node_name, input_name_list, [output_name], axis = axis)
+    if axis == 1:
+        dim = 0
+        for name in input_name_list:
+            dim+=graph.channel_dims[name]
+        graph.channel_dims[output_name] = dim
+    else:
+        graph.channel_dims[output_name] = graph.channel_dims[input_name_list[0]]
+
+    return layer
+
+def _convert_conv_transpose(node,graph,err):
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    node_name = node.name
+    weight_name = node.inputs[1]
+    W = None
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name,))
+    bias_flag = False
+    bias = None
+    if len(node.inputs) > 2:
+        bias = node.input_tensors[node.inputs[2]]
+        bias_flag = True
+    dilations = node.attrs.get("dilations", [1, 1])
+    # groups = 1
+    groups = node.attrs.get("group", 1)
+    kernel_shape = node.attrs["kernel_shape"]
+    pads = node.attrs.get("pads", [0, 0, 0, 0])
+    strides = node.attrs["strides"]
+
+    layer = myf('Deconvolution', node_name, [input_name], [output_name],
+                convolution_param=dict(
+                    num_output=W.shape[1],
+                    kernel_h=kernel_shape[0],kernel_w=kernel_shape[1],
+                    stride_h=strides[0],stride_w = strides[1],
+                    group=groups,
+                    pad_h=pads[0], pad_w=pads[1],
+                    bias_term=bias_flag,
+                ))
+
+    graph.channel_dims[output_name] = W.shape[1]
+    return layer
+
+    # l_top = L.Deconvolution(
+    #     l_bottom,
+    #     name=name,
+    #     convolution_param=dict(
+    #         num_output=W.shape[1],
+    #         kernel_h=kernel_h,
+    #         kernel_w=kernel_w,
+    #         stride_h=stride_h,
+    #         stride_w=stride_w,
+    #         pad_h=pad_h,
+    #         pad_w=pad_w,
+    #         group=groups,
+    #         bias_term=bias_term))
+
+
+
+_ONNX_NODE_REGISTRY = {
+    "Conv": _convert_conv,
+    "Relu": _convert_relu,
+    "BatchNormalization": _convert_BatchNorm,
+    "Add": _convert_Add,
+    "Mul": _convert_Mul,
+    "Reshape": _convert_Reshape,
+    "MaxPool": _convert_pool,
+    "AveragePool": _convert_pool,
+    "Dropout": _convert_dropout,
+    "Gemm": _convert_gemm,
+    "Upsample": _convert_upsample,
+    "Concat": _convert_concat,
+    "ConvTranspose": _convert_conv_transpose,
+    "Sigmoid": _convert_sigmoid,
+    "Flatten": _convert_Flatten,
+    "Resize": _convert_resize,
+    "Transpose": _convert_transpose,
+    "Softmax": _convert_softmax,
+}
diff --git a/tools/onnx2caffe/onnx2caffe/_transformers.py b/tools/onnx2caffe/onnx2caffe/_transformers.py
new file mode 100644
index 000000000..4f8c8159b
--- /dev/null
+++ b/tools/onnx2caffe/onnx2caffe/_transformers.py
@@ -0,0 +1,520 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+from typing import Sequence, Text, Dict, List
+import numpy as np
+
+from onnx import TensorProto
+
+from ._graph import Graph, Node
+
+
+class NodesFuser(object):
+    '''
+    An abstract helper for merging nodes
+    '''
+    def __init__(self,
+                 num_nodes,  # type: int
+                 ):
+        # type: (...) -> None
+        assert num_nodes >= 2, "Algorithm only works if fusing multiple nodes"
+        self.num_nodes = num_nodes
+
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        nodes = graph.nodes
+        merged_nodes = {}
+        for node in nodes:
+            nodes_window = []  # type: List[Node]
+            n = node
+            for _ in range(self.num_nodes - 1):
+                if len(n.parents) != 1:
+                    # We're only fusing nodes with single parents
+                    break
+                p = n.get_only_parent()
+                if len(p.children) != 1:
+                    # We can only fuse a node if its parent's
+                    # value isn't used by any other node.
+                    break
+                nodes_window.insert(0, n)
+                n = p
+            if len(nodes_window) > 0:
+                # add parent of chained nodes
+                first = nodes_window[0]
+                p = first.get_only_parent()
+                if len(p.children) == 1:
+                    nodes_window.insert(0, p)
+            if len(nodes_window) != self.num_nodes:
+                continue
+            if not self.is_eligible(graph, nodes_window):
+                continue
+            merged = self.merge(graph, nodes_window)
+            first, last = nodes_window[0], nodes_window[-1]
+            for parent in first.parents:
+                parent.children.remove(first)
+                if merged[0] not in parent.children:
+                    parent.add_child(merged[0])
+            for child in last.children:
+                child.parents.remove(last)
+                if merged[-1] not in child.parents:
+                    child.add_parent(merged[-1])
+            for n in nodes_window:
+                merged_nodes[n.name] = merged
+
+        transformed_nodes = []
+        added_merged = []  # type: List[Node]
+        for node in nodes:
+            if node.name in merged_nodes:
+                merged = merged_nodes[node.name]
+                if merged[0] not in added_merged:
+                    for n in merged:
+                        transformed_nodes.append(n)
+                    added_merged.append(merged[0])
+            else:
+                transformed_nodes.append(node)
+        return Graph(transformed_nodes, graph.inputs, graph.outputs, graph.shape_dict)
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        '''Returns true if this subset of nodes is eligible for fusion.'''
+        raise NotImplementedError('Must be implemented by subclass.')
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        '''Merge nodes'''
+        nodes[0].outputs = nodes[-1].outputs
+        return [nodes[0]]
+
+
+class ConvAddFuser(NodesFuser):
+    '''
+    Fuses Add layer into parent convolution layer.
+    '''
+    def __init__(self):  # type: () -> None
+        super(ConvAddFuser, self).__init__(2)
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        parent, child = nodes[0], nodes[1]
+        if parent.op_type != 'Conv':
+            return False
+        if child.op_type != 'Add':
+            return False
+        if 'broadcast' not in child.attrs:
+            return False
+        if 'axis' not in child.attrs:
+            return False
+        if parent.inputs[1] not in parent.input_tensors:
+            return False
+        if len(parent.inputs) > 2 and parent.inputs[2] not in parent.input_tensors:
+            return False
+        if child.inputs[1] not in child.input_tensors:
+            return False
+
+        broadcast = child.attrs['broadcast']
+        if broadcast != 1:
+            return False
+
+        axis = child.attrs['axis']
+        if axis != 1:
+            return False
+
+        return True
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        parent, child = nodes[0], nodes[1]
+        output_channels = parent.input_tensors[parent.inputs[1]].shape[0]
+        if len(parent.inputs) > 2:
+            bias_input_name = parent.inputs[2]
+            bias = parent.input_tensors[bias_input_name]
+        else:
+            bias_input_name = "{}_bias".format(parent.name,)
+            parent.inputs.append(bias_input_name)
+            bias = np.zeros(
+                (output_channels,), dtype=np.float32
+            )
+            parent.input_tensors[bias_input_name] = bias
+        bias = bias + child.input_tensors[child.inputs[1]]
+        parent.input_tensors[bias_input_name] = bias
+        parent.outputs = child.outputs
+        parent.children.remove(child)
+        child.parents.remove(parent)
+        return [parent]
+
+
+class BNBroadcastedMulFuser(NodesFuser):
+    '''
+    Fuses Mul into BatchNorm
+    '''
+    def __init__(self):  # type: () -> None
+        super(BNBroadcastedMulFuser, self).__init__(2)
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        parent, child = nodes[0], nodes[1]
+        if parent.op_type != 'BatchNormalization':
+            return False
+        if child.op_type != 'Mul':
+            return False
+        if "broadcast" not in child.attrs:
+            return False
+        if child.attrs["broadcast"] != 1:
+            return False
+        if "axis" not in child.attrs:
+            return False
+        if child.attrs["axis"] != 1:
+            return False
+        if child.inputs[1] not in child.input_tensors:
+            return False
+        if parent.inputs[1] not in parent.input_tensors:
+            return False
+        if parent.inputs[2] not in parent.input_tensors:
+            return False
+        return True
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        parent, child = nodes[0], nodes[1]
+        weight = parent.input_tensors[parent.inputs[1]]
+        bias = parent.input_tensors[parent.inputs[2]]
+        W = child.input_tensors[child.inputs[1]]
+        parent.input_tensors[parent.inputs[1]] = np.multiply(weight, W)
+        parent.input_tensors[parent.inputs[2]] = np.multiply(bias, W)
+        parent.outputs = child.outputs
+        parent.children.remove(child)
+        child.parents.remove(parent)
+        return [parent]
+
+
+class BNBroadcastedAddFuser(NodesFuser):
+    '''
+    Fuses Add into BatchNorm
+    '''
+    def __init__(self):  # type: () -> None
+        super(BNBroadcastedAddFuser, self).__init__(2)
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        parent, child = nodes[0], nodes[1]
+        if parent.op_type != 'BatchNormalization':
+            return False
+        if child.op_type != 'Add':
+            return False
+        if "broadcast" not in child.attrs:
+            return False
+        if child.attrs["broadcast"] != 1:
+            return False
+        if "axis" not in child.attrs:
+            return False
+        if child.attrs["axis"] != 1:
+            return False
+        if len(child.inputs) != 2:
+            return False
+        if child.inputs[1] not in child.input_tensors:
+            return False
+        if parent.inputs[2] not in parent.input_tensors:
+            return False
+        return True
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        parent, child = nodes[0], nodes[1]
+        bias = parent.input_tensors[parent.inputs[2]]
+        b = child.input_tensors[child.inputs[1]]
+        parent.input_tensors[parent.inputs[2]] = bias + b
+        parent.outputs = child.outputs
+        parent.children.remove(child)
+        child.parents.remove(parent)
+        return [parent]
+
+
+class DropoutRemover(NodesFuser):
+    '''
+    Removes Dropout layer
+    '''
+    def __init__(self):  # type: () -> None
+        super(DropoutRemover, self).__init__(2)
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        child = nodes[1]
+        return child.op_type == "Dropout"
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        parent, child = nodes[0], nodes[1]
+        parent.children.remove(child)
+        child.parents.remove(parent)
+        parent.outputs = child.outputs
+        return [parent]
+
+
+class ReshapeInitTensorFuser(object):
+    '''
+    Fuses Reshape operator if it is used only to reshape blob in
+    graph initializer. We can reshape here instead of runtime.
+    '''
+
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        nodes = graph.nodes
+        removed = []
+        for node in nodes:
+            if node.op_type != 'Reshape':
+                continue
+            if not (len(node.input_tensors) == 2 or len(node.input_tensors) == 1):
+                continue
+            tensor_name = node.inputs[0]
+            if tensor_name not in node.input_tensors:
+                continue
+            if len(node.inputs) > 1:
+                shape_name = node.inputs[1]
+                if shape_name not in node.input_tensors:
+                    continue
+            is_non_constant_parent = False
+            if len(node.parents) > 0:
+                for parent in node.parents:
+                    if parent.op_type != 'Constant':
+                        is_non_constant_parent = True
+                        break
+            if is_non_constant_parent:
+                continue
+
+            removed.append(node)
+            output_name = node.outputs[0]
+
+            tensor = node.input_tensors[tensor_name]
+            if 'shape' in node.attrs:
+                shape = tuple(node.attrs["shape"])
+            else:
+                shape = node.input_tensors[shape_name] # type: ignore
+
+            # ONNX spec supports setting dimension to '0', in which case
+            # it should be taken from old dimension.
+            # This isn't supported in numpy, so don't transform.
+            # TODO Should we support this case?
+            if any([s == 0 for s in shape]):
+                continue
+
+            reshaped_tensor = tensor.reshape(shape)
+
+            for child in node.children:
+                child.parents.remove(node)
+                child.input_tensors[output_name] = reshaped_tensor
+
+        transformed_nodes = [node for node in nodes if node not in removed]
+        return Graph(transformed_nodes, graph.inputs, graph.outputs, graph.shape_dict)
+
+
+class OutputRenamer(object):
+    '''
+    Rename outputs according to mapping
+    '''
+    def __init__(self,
+                 mapping,  # type: Dict[Text, Text]
+                 ):
+        # type: (...) -> None
+        self.mapping = mapping
+
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        mapping = self.mapping.copy()
+        nodes = graph.nodes
+        for node in nodes:
+            for i in range(len(node.outputs)):
+                output = node.outputs[i]
+                if output not in mapping:
+                    continue
+                node.outputs[i] = mapping[output]
+                for child in node.children:
+                    for j in range(len(child.inputs)):
+                        input_ = child.inputs[j]
+                        if input_ != output:
+                            continue
+                        child.inputs[j] = mapping[output]
+                del mapping[output]
+                if len(mapping) == 0:
+                    break
+        return graph
+
+
+class PixelShuffleFuser(NodesFuser):
+    '''
+    Fuses 3 operators reshape->transpose->reshape which is equivalent to
+    pytorch's pixel_shuffle layer
+    '''
+    def __init__(self):  # type: () -> None
+        super(PixelShuffleFuser, self).__init__(3)
+        self.num_added = 0
+
+    def is_eligible(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> bool
+        if nodes[0].op_type != 'Reshape':
+            return False
+        if nodes[1].op_type != 'Transpose':
+            return False
+        if nodes[2].op_type != 'Reshape':
+            return False
+        if nodes[0].inputs[1] not in nodes[0].input_tensors:
+            return False
+        if nodes[2].inputs[1] not in nodes[2].input_tensors:
+            return False
+
+        shape = nodes[0].input_tensors[nodes[0].inputs[1]]
+        if len(shape) != 6:
+            return False
+        if shape[0] != 1 or shape[2] != shape[3]:
+            return False
+
+        input_channels = shape[1]
+        scale_factor = shape[2]
+        input_height = shape[4]
+        input_width = shape[5]
+
+        if nodes[1].attrs.get('perm', []) != [0, 1, 4, 2, 5, 3]:
+            return False
+
+        shape = nodes[2].input_tensors[nodes[2].inputs[1]]
+        if len(shape) != 4:
+            return False
+
+        output_channels = shape[1]
+        output_height = shape[2]
+        output_width = shape[3]
+        if input_channels != output_channels:
+            return False
+        if (input_height * scale_factor) != output_height:
+            return False
+        if (input_width * scale_factor) != output_width:
+            return False
+
+        return True
+
+    def get_unique_edge_name(self, graph, name):  # type: (Graph, Text) -> Text
+        self.num_added += 1
+        return graph.get_unique_edge_name(name + '_' + str(self.num_added))
+
+    def merge(self, graph, nodes):  # type: (Graph, Sequence[Node]) -> Sequence[Node]
+        '''
+        Pixel shuffle is implemented using 3 operators:
+            - Reshape(1, channels, scale, scale, height, width)
+            - Transpose(0, 1, 4, 2, 5, 3)
+            - Reshape(1, channels, height * scale, width * scale)
+        CoreML Reshape and Transpose layers don't support tensors with more
+        than 4 dimensions. Thus we change above sequence of operators to the
+        following equivalent sequence:
+            - Reshape(channels, scale * scale, height, width)
+            - Transpose(0, 2, 1, 3)
+            - Reshape(channels * height, scale, scale, width)
+            - Transpose(0, 1, 3, 2)
+            - Reshape(1, channels, height * scale, width * scale)
+        '''
+        reshape_1 = nodes[0]
+        transpose_1 = nodes[1]
+        transpose_1.children = []
+
+        shape = reshape_1.input_tensors[reshape_1.inputs[1]]
+
+        channels = shape[1]
+        scale = shape[2]
+        height = shape[4]
+        width = shape[5]
+
+        reshape_1.input_tensors[reshape_1.inputs[1]] = np.asarray([channels, scale * scale, height, width])
+        transpose_1.attrs['perm'] = [0, 2, 1, 3]
+
+        reshape_output_name = 'pixel_shuffle_reshape'
+        transpose_output_name = 'pixel_shuffle_transpose'
+
+        transpose_1.outputs = [
+            self.get_unique_edge_name(graph, transpose_output_name)
+        ]
+
+        shape_name_second_reshape = self.get_unique_edge_name(graph, reshape_output_name)
+        output_name_second_reshape = self.get_unique_edge_name(graph, reshape_output_name)
+        reshape_2 = Node(
+            reshape_output_name,
+            'Reshape',
+            {},
+            [transpose_1.outputs[0], shape_name_second_reshape],
+            [output_name_second_reshape]
+        )
+        reshape_2.input_tensors[shape_name_second_reshape] = np.asarray([channels * height, scale, scale, width])
+        transpose_1.add_child(reshape_2)
+
+        transpose_2 = Node(
+            transpose_output_name,
+            'Transpose',
+            {'perm': [0, 1, 3, 2]},
+            reshape_2.outputs,
+            [self.get_unique_edge_name(graph, transpose_output_name)]
+        )
+        reshape_2.add_child(transpose_2)
+
+        final_reshape = nodes[2]
+        final_reshape.inputs = [transpose_2.outputs[0], nodes[2].inputs[1]]
+        final_reshape.parents = []
+        transpose_2.add_child(final_reshape)
+        return [reshape_1, transpose_1, reshape_2, transpose_2, final_reshape]
+
+
+class AddModelInputsOutputs(object):
+    '''
+    Expose hidden states of recurrent layers as model inputs and outputs
+    '''
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        input_names = [str(input_[0]) for input_ in graph.inputs]
+        output_names = [str(output_[0]) for output_ in graph.outputs]
+        for node in graph.nodes:
+            if str(node.op_type) == 'LSTM':
+                input_h = node.inputs[5] if len(node.inputs) > 5 else node.inputs[0] + '_h_input'
+                input_c = node.inputs[6] if len(node.inputs) > 6 else node.inputs[0] + '_c_input'
+                output_h = node.outputs[1] if len(node.outputs) > 1 else node.outputs[0] + '_h_output'
+                output_c = node.outputs[2] if len(node.outputs) > 2 else node.outputs[0] + '_c_output'
+                h = node.attrs["hidden_size"]
+                for input_ in [str(input_h), str(input_c)]:
+                    if input_ not in input_names:
+                        graph.inputs.append(tuple((input_, TensorProto.FLOAT, (h,))))  #type: ignore
+                    if input_ not in graph.blob_to_op_type:
+                        graph.blob_to_op_type[input_] = ['LSTM']
+                for output_ in [str(output_h), str(output_c)]:
+                    if output_ not in output_names:
+                        graph.outputs.append(tuple((output_, TensorProto.FLOAT, (h,))))  #type: ignore
+                    graph.blob_from_op_type[output_] = 'LSTM'
+        return graph
+
+
+class ConstantsToInitializers(object):
+    '''
+    Takes onnx Constant nodes and puts the tensor into graph initializers instead.
+    '''
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        output_names = [str(output_[0]) for output_ in graph.outputs]
+        remaining_nodes = []
+        for node in graph.nodes:
+            if node.op_type != 'Constant' or node.name in output_names:
+                remaining_nodes.append(node)
+                continue
+            for child in node.children:
+                child.input_tensors[node.outputs[0]] = node.attrs["value"]
+
+        graph.nodes = remaining_nodes
+        return graph
+
+
+class ImageScalerRemover(object):
+    '''
+    Removes ImageScaler layer if connected to a model input and single parent child nodes
+    '''
+
+    def __call__(self, graph):  # type: (Graph) -> Graph
+        input_names = [str(input_[0]) for input_ in graph.inputs]
+        nodes_to_be_removed = []
+        for node in graph.nodes:
+            if (node.op_type != 'ImageScaler') or (len(node.parents) != 0) or (node.inputs[0] not in input_names):
+                continue
+            is_eligible = True
+            for child in node.children:
+                if not (len(child.parents) == 1 and child.inputs[0] == node.outputs[0]):
+                    is_eligible = False
+                    break
+                child.inputs[0] = node.inputs[0]
+                child.parents = []
+            if not is_eligible:
+                continue
+            nodes_to_be_removed.append(node.name)
+
+        transformed_nodes = []
+        for node in graph.nodes:
+            if node.name not in nodes_to_be_removed:
+                transformed_nodes.append(node)
+        return Graph(transformed_nodes, graph.inputs, graph.outputs, graph.shape_dict)
\ No newline at end of file
diff --git a/tools/onnx2caffe/onnx2caffe/_weightloader.py b/tools/onnx2caffe/onnx2caffe/_weightloader.py
new file mode 100644
index 000000000..0e1129ee9
--- /dev/null
+++ b/tools/onnx2caffe/onnx2caffe/_weightloader.py
@@ -0,0 +1,162 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+# from caffe import params as P
+import numpy as np
+from ._graph import Node, Graph
+
+USE_DECONV_AS_UPSAMPLE = False
+
+def _convert_conv(net, node, graph, err):
+    weight_name = node.inputs[1]
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    node_name = node.name
+    W = None
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name,))
+    bias_flag = False
+    bias = None
+    if len(node.inputs) > 2:
+        bias = node.input_tensors[node.inputs[2]]
+        bias_flag = True
+    # net.params[node_name][0].data = W
+    # if bias_flag:
+    #     net.params[node_name][1].data = bias
+    np.copyto(net.params[node_name][0].data,W,casting='same_kind')
+    if bias_flag:
+        np.copyto(net.params[node_name][1].data, bias, casting='same_kind')
+
+def _convert_relu(net, node, graph, err):
+    pass
+
+def _convert_sigmoid(net, node, graph, err):
+    pass
+
+def _convert_BatchNorm(net, node, graph, err):
+    scale = node.input_tensors[node.inputs[1]]
+    bias = node.input_tensors[node.inputs[2]]
+    mean = node.input_tensors[node.inputs[3]]
+    var = node.input_tensors[node.inputs[4]]
+    node_name = node.name
+    np.copyto(net.params[node_name + '_bn'][0].data, mean, casting='same_kind')
+    np.copyto(net.params[node_name + '_bn'][1].data, var, casting='same_kind')
+    net.params[node_name + '_bn'][2].data[...] = 1.0
+    np.copyto(net.params[node_name][0].data, scale, casting='same_kind')
+    np.copyto(net.params[node_name][1].data, bias, casting='same_kind')
+    # net.params[node_name+'_bn'][1].data = var
+    # net.params[node_name][0].data = scale
+    # net.params[node_name][1].data = bias
+
+def _convert_Add(net, node, graph, err):
+    pass
+
+def _convert_Mul(net, node, graph, err):
+    pass
+
+def _convert_Reshape(net, node, graph, err):
+    pass
+
+def _convert_Flatten(net, node, graph, err):
+    pass
+
+def _convert_pool(net, node, graph, err):
+    pass
+
+def _convert_dropout(net, node, graph, err):
+    pass
+
+def _convert_gemm(net, node, graph, err):
+    node_name = node.name
+    weight_name = node.inputs[1]
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name, ))
+    if node.attrs["broadcast"] != 1 or node.attrs["transB"] != 1:
+        return err.unsupported_op_configuration(node, "Gemm is supported only for inner_product layer")
+    b = None
+    if len(node.inputs) > 2:
+        b = node.input_tensors[node.inputs[2]]
+    if len(W.shape) != 2 or (b is not None and len(b.shape) != 1):
+        return err.unsupported_op_configuration(node, "Gemm is supported only for inner_product layer")
+    if b is not None:
+        if W.shape[0] != b.shape[0]:
+            return err.unsupported_op_configuration(node, "Gemm is supported only for inner_product layer")
+    net.params[node_name][0].data[...] = W
+    net.params[node_name][1].data[...] = b
+
+def _convert_upsample(net, node, graph, err):
+    mode = node.attrs["mode"]
+    node_name = node.name
+    if mode == "nearest":
+        caffe_params = net.params[node_name][0].data
+        weights = np.ones(caffe_params.shape).astype("float32")
+        np.copyto(net.params[node_name][0].data, weights, casting='same_kind')
+        # net.params[node_name][0].data[]
+
+def _convert_resize(net, node, graph, err):
+    if USE_DECONV_AS_UPSAMPLE:
+        print('add resize deconv param')
+        node_name = node.name
+        caffe_params = net.params[node_name][0].data
+        weights = np.ones(caffe_params.shape).astype("float32")
+        np.copyto(net.params[node_name][0].data, weights, casting='same_kind')
+
+def _convert_transpose(net, node, graph, err):
+    pass
+def _convert_concat(net, node, graph, err):
+    pass
+def _convert_softmax(net, node, graph, err):
+    pass
+
+def _convert_conv_transpose(net, node, graph, err):
+    weight_name = node.inputs[1]
+    input_name = str(node.inputs[0])
+    output_name = str(node.outputs[0])
+    node_name = node.name
+    W = None
+    if weight_name in node.input_tensors:
+        W = node.input_tensors[weight_name]
+    else:
+        err.missing_initializer(node,
+                                "Weight tensor: {} not found in the graph initializer".format(weight_name,))
+    bias_flag = False
+    bias = None
+    if len(node.inputs) > 2:
+        bias = node.input_tensors[node.inputs[2]]
+        bias_flag = True
+    # net.params[node_name][0].data = W
+    # if bias_flag:
+    #     net.params[node_name][1].data = bias
+    np.copyto(net.params[node_name][0].data,W,casting='same_kind')
+    if bias_flag:
+        np.copyto(net.params[node_name][1].data, bias, casting='same_kind')
+
+_ONNX_NODE_REGISTRY = {
+    "Conv": _convert_conv,
+    "Relu": _convert_relu,
+    "BatchNormalization": _convert_BatchNorm,
+    "Add": _convert_Add,
+    "Mul": _convert_Mul,
+    "Reshape": _convert_Reshape,
+    "MaxPool": _convert_pool,
+    "AveragePool": _convert_pool,
+    "Dropout": _convert_dropout,
+    "Gemm": _convert_gemm,
+    "Upsample": _convert_upsample,
+    "Concat": _convert_concat,
+    "ConvTranspose": _convert_conv_transpose,
+    "Sigmoid": _convert_sigmoid,
+    "Flatten": _convert_Flatten,
+    "Resize": _convert_resize,
+    "Transpose": _convert_transpose,
+    "Softmax": _convert_softmax,
+}
+
+
diff --git a/web-demos/README.md b/web-demos/README.md
new file mode 100644
index 000000000..8765a57fa
--- /dev/null
+++ b/web-demos/README.md
@@ -0,0 +1,13 @@
+InsightFace Web Demos
+---
+
+## UPDATE
+
+``2023.04.01``:  Please use our swapping demo via Discord bot, see detail [here](swapping_discord/).
+
+
+## Swapping Demo on Discord
+
+Check the tutorial [here](swapping_discord/).
+
+
diff --git a/web-demos/src_recognition/arcface_onnx.py b/web-demos/src_recognition/arcface_onnx.py
new file mode 100644
index 000000000..870e7d641
--- /dev/null
+++ b/web-demos/src_recognition/arcface_onnx.py
@@ -0,0 +1,91 @@
+# -*- coding: utf-8 -*-
+# @Organization  : insightface.ai
+# @Author        : Jia Guo
+# @Time          : 2021-05-04
+# @Function      : 
+
+import numpy as np
+import cv2
+import onnx
+import onnxruntime
+import face_align
+
+__all__ = [
+    'ArcFaceONNX',
+]
+
+
+class ArcFaceONNX:
+    def __init__(self, model_file=None, session=None):
+        assert model_file is not None
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'recognition'
+        find_sub = False
+        find_mul = False
+        model = onnx.load(self.model_file)
+        graph = model.graph
+        for nid, node in enumerate(graph.node[:8]):
+            #print(nid, node.name)
+            if node.name.startswith('Sub') or node.name.startswith('_minus'):
+                find_sub = True
+            if node.name.startswith('Mul') or node.name.startswith('_mul'):
+                find_mul = True
+        if find_sub and find_mul:
+            #mxnet arcface model
+            input_mean = 0.0
+            input_std = 1.0
+        else:
+            input_mean = 127.5
+            input_std = 127.5
+        self.input_mean = input_mean
+        self.input_std = input_std
+        #print('input mean and std:', self.input_mean, self.input_std)
+        if self.session is None:
+            self.session = onnxruntime.InferenceSession(self.model_file, providers=['CUDAExecutionProvider'])
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        input_name = input_cfg.name
+        self.input_size = tuple(input_shape[2:4][::-1])
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        output_names = []
+        for out in outputs:
+            output_names.append(out.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        assert len(self.output_names)==1
+        self.output_shape = outputs[0].shape
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+
+    def get(self, img, kps):
+        aimg = face_align.norm_crop(img, landmark=kps, image_size=self.input_size[0])
+        embedding = self.get_feat(aimg).flatten()
+        return embedding
+
+    def compute_sim(self, feat1, feat2):
+        from numpy.linalg import norm
+        feat1 = feat1.ravel()
+        feat2 = feat2.ravel()
+        sim = np.dot(feat1, feat2) / (norm(feat1) * norm(feat2))
+        return sim
+
+    def get_feat(self, imgs):
+        if not isinstance(imgs, list):
+            imgs = [imgs]
+        input_size = self.input_size
+        
+        blob = cv2.dnn.blobFromImages(imgs, 1.0 / self.input_std, input_size,
+                                      (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+    def forward(self, batch_data):
+        blob = (batch_data - self.input_mean) / self.input_std
+        net_out = self.session.run(self.output_names, {self.input_name: blob})[0]
+        return net_out
+
+
diff --git a/web-demos/src_recognition/face_align.py b/web-demos/src_recognition/face_align.py
new file mode 100644
index 000000000..c4e9a7c90
--- /dev/null
+++ b/web-demos/src_recognition/face_align.py
@@ -0,0 +1,141 @@
+import cv2
+import numpy as np
+from skimage import transform as trans
+
+src1 = np.array([[51.642, 50.115], [57.617, 49.990], [35.740, 69.007],
+                 [51.157, 89.050], [57.025, 89.702]],
+                dtype=np.float32)
+#<--left
+src2 = np.array([[45.031, 50.118], [65.568, 50.872], [39.677, 68.111],
+                 [45.177, 86.190], [64.246, 86.758]],
+                dtype=np.float32)
+
+#---frontal
+src3 = np.array([[39.730, 51.138], [72.270, 51.138], [56.000, 68.493],
+                 [42.463, 87.010], [69.537, 87.010]],
+                dtype=np.float32)
+
+#-->right
+src4 = np.array([[46.845, 50.872], [67.382, 50.118], [72.737, 68.111],
+                 [48.167, 86.758], [67.236, 86.190]],
+                dtype=np.float32)
+
+#-->right profile
+src5 = np.array([[54.796, 49.990], [60.771, 50.115], [76.673, 69.007],
+                 [55.388, 89.702], [61.257, 89.050]],
+                dtype=np.float32)
+
+src = np.array([src1, src2, src3, src4, src5])
+src_map = {112: src, 224: src * 2}
+
+arcface_src = np.array(
+    [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+     [41.5493, 92.3655], [70.7299, 92.2041]],
+    dtype=np.float32)
+
+arcface_src = np.expand_dims(arcface_src, axis=0)
+
+# In[66]:
+
+
+# lmk is prediction; src is template
+def estimate_norm(lmk, image_size=112, mode='arcface'):
+    assert lmk.shape == (5, 2)
+    tform = trans.SimilarityTransform()
+    lmk_tran = np.insert(lmk, 2, values=np.ones(5), axis=1)
+    min_M = []
+    min_index = []
+    min_error = float('inf')
+    if mode == 'arcface':
+        if image_size == 112:
+            src = arcface_src
+        else:
+            src = float(image_size) / 112 * arcface_src
+    else:
+        src = src_map[image_size]
+    for i in np.arange(src.shape[0]):
+        tform.estimate(lmk, src[i])
+        M = tform.params[0:2, :]
+        results = np.dot(M, lmk_tran.T)
+        results = results.T
+        error = np.sum(np.sqrt(np.sum((results - src[i])**2, axis=1)))
+        #         print(error)
+        if error < min_error:
+            min_error = error
+            min_M = M
+            min_index = i
+    return min_M, min_index
+
+
+def norm_crop(img, landmark, image_size=112, mode='arcface'):
+    M, pose_index = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped
+
+def square_crop(im, S):
+    if im.shape[0] > im.shape[1]:
+        height = S
+        width = int(float(im.shape[1]) / im.shape[0] * S)
+        scale = float(S) / im.shape[0]
+    else:
+        width = S
+        height = int(float(im.shape[0]) / im.shape[1] * S)
+        scale = float(S) / im.shape[1]
+    resized_im = cv2.resize(im, (width, height))
+    det_im = np.zeros((S, S, 3), dtype=np.uint8)
+    det_im[:resized_im.shape[0], :resized_im.shape[1], :] = resized_im
+    return det_im, scale
+
+
+def transform(data, center, output_size, scale, rotation):
+    scale_ratio = scale
+    rot = float(rotation) * np.pi / 180.0
+    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
+    t1 = trans.SimilarityTransform(scale=scale_ratio)
+    cx = center[0] * scale_ratio
+    cy = center[1] * scale_ratio
+    t2 = trans.SimilarityTransform(translation=(-1 * cx, -1 * cy))
+    t3 = trans.SimilarityTransform(rotation=rot)
+    t4 = trans.SimilarityTransform(translation=(output_size / 2,
+                                                output_size / 2))
+    t = t1 + t2 + t3 + t4
+    M = t.params[0:2]
+    cropped = cv2.warpAffine(data,
+                             M, (output_size, output_size),
+                             borderValue=0.0)
+    return cropped, M
+
+
+def trans_points2d(pts, M):
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i] = new_pt[0:2]
+
+    return new_pts
+
+
+def trans_points3d(pts, M):
+    scale = np.sqrt(M[0][0] * M[0][0] + M[0][1] * M[0][1])
+    #print(scale)
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i][0:2] = new_pt[0:2]
+        new_pts[i][2] = pts[i][2] * scale
+
+    return new_pts
+
+
+def trans_points(pts, M):
+    if pts.shape[1] == 2:
+        return trans_points2d(pts, M)
+    else:
+        return trans_points3d(pts, M)
+
diff --git a/web-demos/src_recognition/main.py b/web-demos/src_recognition/main.py
new file mode 100644
index 000000000..818843373
--- /dev/null
+++ b/web-demos/src_recognition/main.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python
+
+import os
+import os.path as osp
+import argparse
+import cv2
+import numpy as np
+import onnxruntime
+from scrfd import SCRFD
+from arcface_onnx import ArcFaceONNX
+
+onnxruntime.set_default_logger_severity(3)
+
+assets_dir = osp.expanduser('~/.insightface/models/buffalo_l')
+
+detector = SCRFD(os.path.join(assets_dir, 'det_10g.onnx'))
+detector.prepare(0)
+model_path = os.path.join(assets_dir, 'w600k_r50.onnx')
+rec = ArcFaceONNX(model_path)
+rec.prepare(0)
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument('img1', type=str)
+    parser.add_argument('img2', type=str)
+    return parser.parse_args()
+
+
+def func(args):
+    image1 = cv2.imread(args.img1)
+    image2 = cv2.imread(args.img2)
+    bboxes1, kpss1 = detector.autodetect(image1, max_num=1)
+    if bboxes1.shape[0]==0:
+        return -1.0, "Face not found in Image-1"
+    bboxes2, kpss2 = detector.autodetect(image2, max_num=1)
+    if bboxes2.shape[0]==0:
+        return -1.0, "Face not found in Image-2"
+    kps1 = kpss1[0]
+    kps2 = kpss2[0]
+    feat1 = rec.get(image1, kps1)
+    feat2 = rec.get(image2, kps2)
+    sim = rec.compute_sim(feat1, feat2)
+    if sim<0.2:
+        conclu = 'They are NOT the same person'
+    elif sim>=0.2 and sim<0.28:
+        conclu = 'They are LIKELY TO be the same person'
+    else:
+        conclu = 'They ARE the same person'
+    return sim, conclu
+
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    output = func(args)
+    print('sim: %.4f, message: %s'%(output[0], output[1]))
+
diff --git a/web-demos/src_recognition/scrfd.py b/web-demos/src_recognition/scrfd.py
new file mode 100644
index 000000000..cc04996d2
--- /dev/null
+++ b/web-demos/src_recognition/scrfd.py
@@ -0,0 +1,329 @@
+
+from __future__ import division
+import datetime
+import numpy as np
+#import onnx
+import onnxruntime
+import os
+import os.path as osp
+import cv2
+import sys
+
+def softmax(z):
+    assert len(z.shape) == 2
+    s = np.max(z, axis=1)
+    s = s[:, np.newaxis] # necessary step to do broadcasting
+    e_x = np.exp(z - s)
+    div = np.sum(e_x, axis=1)
+    div = div[:, np.newaxis] # dito
+    return e_x / div
+
+def distance2bbox(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    x1 = points[:, 0] - distance[:, 0]
+    y1 = points[:, 1] - distance[:, 1]
+    x2 = points[:, 0] + distance[:, 2]
+    y2 = points[:, 1] + distance[:, 3]
+    if max_shape is not None:
+        x1 = x1.clamp(min=0, max=max_shape[1])
+        y1 = y1.clamp(min=0, max=max_shape[0])
+        x2 = x2.clamp(min=0, max=max_shape[1])
+        y2 = y2.clamp(min=0, max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def distance2kps(points, distance, max_shape=None):
+    """Decode distance prediction to bounding box.
+
+    Args:
+        points (Tensor): Shape (n, 2), [x, y].
+        distance (Tensor): Distance from the given point to 4
+            boundaries (left, top, right, bottom).
+        max_shape (tuple): Shape of the image.
+
+    Returns:
+        Tensor: Decoded bboxes.
+    """
+    preds = []
+    for i in range(0, distance.shape[1], 2):
+        px = points[:, i%2] + distance[:, i]
+        py = points[:, i%2+1] + distance[:, i+1]
+        if max_shape is not None:
+            px = px.clamp(min=0, max=max_shape[1])
+            py = py.clamp(min=0, max=max_shape[0])
+        preds.append(px)
+        preds.append(py)
+    return np.stack(preds, axis=-1)
+
+class SCRFD:
+    def __init__(self, model_file=None, session=None):
+        import onnxruntime
+        self.model_file = model_file
+        self.session = session
+        self.taskname = 'detection'
+        self.batched = False
+        if self.session is None:
+            assert self.model_file is not None
+            assert osp.exists(self.model_file)
+            self.session = onnxruntime.InferenceSession(self.model_file, providers=['CUDAExecutionProvider'])
+        self.center_cache = {}
+        self.nms_thresh = 0.4
+        self.det_thresh = 0.5
+        self._init_vars()
+
+    def _init_vars(self):
+        input_cfg = self.session.get_inputs()[0]
+        input_shape = input_cfg.shape
+        #print(input_shape)
+        if isinstance(input_shape[2], str):
+            self.input_size = None
+        else:
+            self.input_size = tuple(input_shape[2:4][::-1])
+        #print('image_size:', self.image_size)
+        input_name = input_cfg.name
+        self.input_shape = input_shape
+        outputs = self.session.get_outputs()
+        if len(outputs[0].shape) == 3:
+            self.batched = True
+        output_names = []
+        for o in outputs:
+            output_names.append(o.name)
+        self.input_name = input_name
+        self.output_names = output_names
+        self.input_mean = 127.5
+        self.input_std = 128.0
+        #print(self.output_names)
+        #assert len(outputs)==10 or len(outputs)==15
+        self.use_kps = False
+        self._anchor_ratio = 1.0
+        self._num_anchors = 1
+        if len(outputs)==6:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+        elif len(outputs)==9:
+            self.fmc = 3
+            self._feat_stride_fpn = [8, 16, 32]
+            self._num_anchors = 2
+            self.use_kps = True
+        elif len(outputs)==10:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+        elif len(outputs)==15:
+            self.fmc = 5
+            self._feat_stride_fpn = [8, 16, 32, 64, 128]
+            self._num_anchors = 1
+            self.use_kps = True
+
+    def prepare(self, ctx_id, **kwargs):
+        if ctx_id<0:
+            self.session.set_providers(['CPUExecutionProvider'])
+        nms_thresh = kwargs.get('nms_thresh', None)
+        if nms_thresh is not None:
+            self.nms_thresh = nms_thresh
+        det_thresh = kwargs.get('det_thresh', None)
+        if det_thresh is not None:
+            self.det_thresh = det_thresh
+        input_size = kwargs.get('input_size', None)
+        if input_size is not None:
+            if self.input_size is not None:
+                print('warning: det_size is already set in scrfd model, ignore')
+            else:
+                self.input_size = input_size
+
+    def forward(self, img, threshold):
+        scores_list = []
+        bboxes_list = []
+        kpss_list = []
+        input_size = tuple(img.shape[0:2][::-1])
+        blob = cv2.dnn.blobFromImage(img, 1.0/self.input_std, input_size, (self.input_mean, self.input_mean, self.input_mean), swapRB=True)
+        net_outs = self.session.run(self.output_names, {self.input_name : blob})
+
+        input_height = blob.shape[2]
+        input_width = blob.shape[3]
+        fmc = self.fmc
+        for idx, stride in enumerate(self._feat_stride_fpn):
+            # If model support batch dim, take first output
+            if self.batched:
+                scores = net_outs[idx][0]
+                bbox_preds = net_outs[idx + fmc][0]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2][0] * stride
+            # If model doesn't support batching take output as is
+            else:
+                scores = net_outs[idx]
+                bbox_preds = net_outs[idx + fmc]
+                bbox_preds = bbox_preds * stride
+                if self.use_kps:
+                    kps_preds = net_outs[idx + fmc * 2] * stride
+
+            height = input_height // stride
+            width = input_width // stride
+            K = height * width
+            key = (height, width, stride)
+            if key in self.center_cache:
+                anchor_centers = self.center_cache[key]
+            else:
+                #solution-1, c style:
+                #anchor_centers = np.zeros( (height, width, 2), dtype=np.float32 )
+                #for i in range(height):
+                #    anchor_centers[i, :, 1] = i
+                #for i in range(width):
+                #    anchor_centers[:, i, 0] = i
+
+                #solution-2:
+                #ax = np.arange(width, dtype=np.float32)
+                #ay = np.arange(height, dtype=np.float32)
+                #xv, yv = np.meshgrid(np.arange(width), np.arange(height))
+                #anchor_centers = np.stack([xv, yv], axis=-1).astype(np.float32)
+
+                #solution-3:
+                anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+                #print(anchor_centers.shape)
+
+                anchor_centers = (anchor_centers * stride).reshape( (-1, 2) )
+                if self._num_anchors>1:
+                    anchor_centers = np.stack([anchor_centers]*self._num_anchors, axis=1).reshape( (-1,2) )
+                if len(self.center_cache)<100:
+                    self.center_cache[key] = anchor_centers
+
+            pos_inds = np.where(scores>=threshold)[0]
+            bboxes = distance2bbox(anchor_centers, bbox_preds)
+            pos_scores = scores[pos_inds]
+            pos_bboxes = bboxes[pos_inds]
+            scores_list.append(pos_scores)
+            bboxes_list.append(pos_bboxes)
+            if self.use_kps:
+                kpss = distance2kps(anchor_centers, kps_preds)
+                #kpss = kps_preds
+                kpss = kpss.reshape( (kpss.shape[0], -1, 2) )
+                pos_kpss = kpss[pos_inds]
+                kpss_list.append(pos_kpss)
+        return scores_list, bboxes_list, kpss_list
+
+    def detect(self, img, input_size = None, thresh=None, max_num=0, metric='default'):
+        assert input_size is not None or self.input_size is not None
+        input_size = self.input_size if input_size is None else input_size
+            
+        im_ratio = float(img.shape[0]) / img.shape[1]
+        model_ratio = float(input_size[1]) / input_size[0]
+        if im_ratio>model_ratio:
+            new_height = input_size[1]
+            new_width = int(new_height / im_ratio)
+        else:
+            new_width = input_size[0]
+            new_height = int(new_width * im_ratio)
+        det_scale = float(new_height) / img.shape[0]
+        resized_img = cv2.resize(img, (new_width, new_height))
+        det_img = np.zeros( (input_size[1], input_size[0], 3), dtype=np.uint8 )
+        det_img[:new_height, :new_width, :] = resized_img
+        det_thresh = thresh if thresh is not None else self.det_thresh
+
+        scores_list, bboxes_list, kpss_list = self.forward(det_img, det_thresh)
+
+        scores = np.vstack(scores_list)
+        scores_ravel = scores.ravel()
+        order = scores_ravel.argsort()[::-1]
+        bboxes = np.vstack(bboxes_list) / det_scale
+        if self.use_kps:
+            kpss = np.vstack(kpss_list) / det_scale
+        pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
+        pre_det = pre_det[order, :]
+        keep = self.nms(pre_det)
+        det = pre_det[keep, :]
+        if self.use_kps:
+            kpss = kpss[order,:,:]
+            kpss = kpss[keep,:,:]
+        else:
+            kpss = None
+        if max_num > 0 and det.shape[0] > max_num:
+            area = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                    det[:, 1])
+            img_center = img.shape[0] // 2, img.shape[1] // 2
+            offsets = np.vstack([
+                (det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2 - img_center[0]
+            ])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            if metric=='max':
+                values = area
+            else:
+                values = area - offset_dist_squared * 2.0  # some extra weight on the centering
+            bindex = np.argsort(
+                values)[::-1]  # some extra weight on the centering
+            bindex = bindex[0:max_num]
+            det = det[bindex, :]
+            if kpss is not None:
+                kpss = kpss[bindex, :]
+        return det, kpss
+
+    def autodetect(self, img, max_num=0, metric='max'):
+        bboxes, kpss = self.detect(img, input_size=(640, 640), thresh=0.5)
+        bboxes2, kpss2 = self.detect(img, input_size=(128, 128), thresh=0.5)
+        bboxes_all = np.concatenate([bboxes, bboxes2], axis=0)
+        kpss_all = np.concatenate([kpss, kpss2], axis=0)
+        keep = self.nms(bboxes_all)
+        det = bboxes_all[keep,:]
+        kpss = kpss_all[keep,:]
+        if max_num > 0 and det.shape[0] > max_num:
+            area = (det[:, 2] - det[:, 0]) * (det[:, 3] -
+                                                    det[:, 1])
+            img_center = img.shape[0] // 2, img.shape[1] // 2
+            offsets = np.vstack([
+                (det[:, 0] + det[:, 2]) / 2 - img_center[1],
+                (det[:, 1] + det[:, 3]) / 2 - img_center[0]
+            ])
+            offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
+            if metric=='max':
+                values = area
+            else:
+                values = area - offset_dist_squared * 2.0  # some extra weight on the centering
+            bindex = np.argsort(
+                values)[::-1]  # some extra weight on the centering
+            bindex = bindex[0:max_num]
+            det = det[bindex, :]
+            if kpss is not None:
+                kpss = kpss[bindex, :]
+        return det, kpss
+
+    def nms(self, dets):
+        thresh = self.nms_thresh
+        x1 = dets[:, 0]
+        y1 = dets[:, 1]
+        x2 = dets[:, 2]
+        y2 = dets[:, 3]
+        scores = dets[:, 4]
+
+        areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+        order = scores.argsort()[::-1]
+
+        keep = []
+        while order.size > 0:
+            i = order[0]
+            keep.append(i)
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+
+            w = np.maximum(0.0, xx2 - xx1 + 1)
+            h = np.maximum(0.0, yy2 - yy1 + 1)
+            inter = w * h
+            ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+            inds = np.where(ovr <= thresh)[0]
+            order = order[inds + 1]
+
+        return keep
+
diff --git a/web-demos/swapping_discord/README.md b/web-demos/swapping_discord/README.md
new file mode 100644
index 000000000..060bf5ac9
--- /dev/null
+++ b/web-demos/swapping_discord/README.md
@@ -0,0 +1,233 @@
+# Using Midjourney and the Picsi.AI by InsightFaceSwap Bot to create a personalized portrait and do face swapping
+
+We have integrated our most advanced face-swapping models: **inswapper_cyn** and **inswapper_dax**, into the [Picsi.Ai face-swapping](https://www.picsi.ai)(InsightFace discord bot) service. These models outperform almost all similar commercial products and our open-source model inswapper_128. Please visit the [Picsi.Ai](https://www.picsi.ai) website to use the service and get help.
+
+## Important Links
+
+1. InsightFace(insightfaceswap) Discord bot invitation link: https://discord.com/api/oauth2/authorize?client_id=1090660574196674713&permissions=274877945856&scope=bot
+2. Discord discussion server(to get help): https://discord.gg/Ym3X8U59ZN
+3. Product homepage: https://www.picsi.ai
+4. Subscribe to unlock advanced features: https://www.picsi.ai/subscribe
+5. User guide: https://www.picsi.ai/docs
+
+## ChangeLog
+
+**`2024-06-07`**: Please visit https://www.patreon.com/picsi to view all recent feature updates.
+
+**`2023-09-09`**: **Introducing HiFidelity Mode (BETA) and Sharpen Options**
+1) Use the ``-f`` command to enable HiFidelity mode: ``/setid example -f`` or ``/swapid example -f``
+2) Add the ``-s`` command for an extra layer of sharpness: ``/setid joey -f -s`` or ``/swapid joey -f -s``
+3) These new features can be combined with existing functionalities like Oldify for even more jaw-dropping results! For instance: ``/swapid pamela -o -f -s``
+4) Sample outputs(zoom in to view the details):
+     <div align="left">
+         <img src="https://github.com/nttstar/insightface-resources/blob/master/images/v0.4_2_2.jpg?raw=true" width="800"/>
+     </div>
+5) For examples in detail, please jump to https://www.patreon.com/posts/89036144.
+
+**`2023-08-27`** 
+1) **Enhanced GIF Quality**: Improved resolution, reduced noise, and enhanced sharpness for GIF outputs. This aims to provide a clearer and better visual experience.
+2) **Increased GIF Size Limit**: The maximum allowable GIF file size has been increased from 7MB to 10MB, allowing for more detailed and creative GIFs.
+3) **Extended Frame Limit for Pro Members**: Pro members can now utilize up to 75 frames for GIFs at a flat rate of 30 credits. This expands the possibilities for more complex and intricate GIFs.
+4) **GIF Support for Basic Members**: Basic members now have access to GIF support, limited to 20 frames at a cost of 20 credits.
+5) **URL Support for GIF**: Added the ability to directly work on GIFs using URL links, eliminating the need to download and re-upload GIF files. Provides an easier and faster way to create funny GIFs.
+6) For examples in detail, please jump to https://www.patreon.com/posts/88351201.
+
+
+
+**`2023-08-25`** 
+ Time Travel Has Never Been So Easy! Introducing Oldifying Faces.
+  1) Use a saved face and transfer it into your target image, then apply the oldifying effect. For instance:
+     
+       ``/swapid johndoe --oldify 300``
+     
+     This will take the saved face named johndoe, and then oldify it with an intensity of 300.
+     
+     Note that we can use ``-o`` as a shorthand for the ``--oldify`` argument.
+  3) You can directly oldify a face in the attached picture without transfer it with one of your saved faces:
+     
+       ``/swapid _ --oldify 200``
+     
+  4) Use the --oldify option to set the transformation intensity, ranging from 1 to 1000. The default intensity is 300 if none is specified.
+
+       ``/swapid _ --oldify``
+     
+  5) Special reminder: Due to the additional arguments parsing, please make sure that the input for idname does not contain any spaces. For example, ``/setid A,B`` is allowed, but ``/setid A, B`` is incorrect.
+      <div align="left">
+         <img src="https://github.com/nttstar/insightface-resources/blob/master/images/v0.3_image.jpg?raw=true" width="640"/>
+      </div>
+
+**`2023-08-02`** 
+  We have deployed a new model and optimized three aspects:
+  1) The new model performs better in handling skin shading under complex lighting conditions, reducing the likelihood of generating black or white erroneous pixels on the skin.
+  2) We have optimized the handling of glasses in the Saved/Source photo. When the Source photo contains glasses, we will generate the image based on the version without glasses to avoid any ghosting effects caused by glasses in the resulting photo. For target images that originally have glasses or sunglasses, this process will not affect the final results.
+  3) We have optimized the handling of bangs/fringe. When the source photo has thick bangs/fringe, we will try to minimize the impact on the generated result.
+      <div align="left">
+         <img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/picsi_20230802.jpg" width="640"/>
+      </div>
+
+**`2023-06-02`**
+  1) The length limit for idname has been increased to 10, and the maximum number of idnames that can be saved has been increased to 20.
+  2) Remove the 'greedy' prefer option, now '--nogreedy' and '--greedy' produce the same result.
+  3) The feature of ID mixing has been added. You can use the symbol "+" to link multiple idnames (up to 3) to generate interesting results. For example, ``/setid father+mother`` might generate an image similar to their son, and ``/setid mother+father`` might generate a photo like their daughter (that is, the order of the "+" link will affect the result). You can also use ``/setid mother+mother+father`` to enhance the features of the mother ID. There's an example [here](https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/240_14.jpeg)
+
+**`2023-05-17`**
+  1) The maximum command usage per image is set to 2, meaning that even if there are 4 faces in a single image, it will only consume 2 commands.
+  2) Now we use a queue in our backend. When there are too many users online, the requests will be queued and processed one by one which may slow to respond.
+  3) The support for GIFs has been temporarily removed, in order to ensure fast response time.
+
+**`2023-05-13`**
+  Now we support the **greedy** mode as the default option, which can provide higher identity similarity. You can use the ``/setid --nogreedy``(put ``--nogreedy`` in the ``idname`` field) command to disable it (and use ``/setid --greedy`` to enable again). In addition, the ``/listid`` command can be used to view the current ID name and prefer options. For more information, please refer to the instruction of the ``/setid`` command on this page.
+
+**`2023-05-08`**
+  1) The maximum pixel output has now been changed to 2048, previously it was 1920.
+  2) The number of command statistics have been changed from the number of images to the number of faces (i.e. if there are 2 faces in one image, it will consume 2 commands).
+
+**`2023-04-27`**
+  1) Now we support swapping on GIFs. The usage is the same as static images. A few extra key points: 1) Uploaded gifs cannot exceed 5MB in size; 2) Performing one gif face swap will consume 5 command opportunities (i.e. a maximum of 10 gifs can be operated per day); 3) Up to the first 15 frames can be operated; 4) Supports single-person swapping only in GIFs; 5) The frames may be dynamically resized to a lower resolution.
+  2) Add FAQ.
+
+**`2023-04-18`**
+  Now we support Discord application commands(AKA. slash commands), please remember joining our [Discord group](https://discord.gg/Ym3X8U59ZN) to get notification.
+
+## Disclaimer
+
+By using this service, you acknowledge that you have read, understood, and agreed to the terms and conditions outlined in this disclaimer.
+
+We would like to emphasize that our service is intended for research and legal AI creation purposes only. We do not condone or promote the use of our service for any illegal or unethical activities. We strictly prohibit the use of our service to process the facial features of individuals without their express permission or consent. Additionally, we do not allow the usage of features of political figures, public officials, or any other public figures without their permission.
+
+We also do not assume any responsibility or liability for the consequences that may arise from the use of our service. Our service is provided on an "as is" basis, and we do not guarantee the accuracy, completeness, or reliability of the results obtained through the use of our service.
+
+By using our service, you agree to indemnify and hold us harmless from any claim or demand, including reasonable attorneys' fees, made by any third-party due to or arising out of your use of the service, your violation of these terms and conditions, or your violation of any rights of another.
+
+In summary, we strictly prohibit the use of our service for any illegal or unethical activities and we are not responsible for any consequences that may arise from the use of our service. If you agree to these terms and conditions, please proceed to use our service.
+
+## License
+
+Our service does not claim any intellectual property rights over the original images or the transformed AI-generated images. Any use of these AI-generated images should respect the copyrights and trademarks of the original images and should not infringe upon the rights of the original copyright owners.
+
+As long as the images do not infringe on any copyrights, paid users can use the generated images for commercial purposes. Free members can not. It is crucial to indicate that these images were altered and generated by Picsi.Ai - Powered by InsightFace, in a visible and accessible manner, to ensure compliance with our licensing terms, legal obligations, and ethical considerations. If a digital picture, this must also be included in the meta and exif data of the photo.
+
+## Introduction
+
+For over 99% of people, using Midjourney to create your own portraits is not feasible unless you're a famous celebrity with thousands or millions of photos online. But now, with the InsightFaceSwap Discord bot, you can accomplish this task easily with just a few steps.
+
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd0.jpg" width="800"/>
+</div>
+
+## Discord Slash Commands
+
+InsightFaceSwap bot can help you with the following commands:
+
+### /saveid ``name`` ``upload-ID-image``
+
+Used to upload and register your own ID photo or numpy feature for subsequent facial replacement and editing. You can upload up to 10 instances permanently and use them without having to upload them repeatedly.
+
+(Front-view, high quality, no glasses, no heavy bangs ID photos are prefered.）
+
+### /setid ``name/prefer``
+
+This command can be used to do two things.
+
+1) Set default identity name(s), for image generation using context menu. If you need to set multiple ID names, please use commas to separate them.
+2) Set prefer options, e.g. use ``/setid --greedy`` to enable greedy mode and ``/setid --nogreedy`` to disable. (The prefer options are placed in the ``idname`` field of ``/setid`` command, don't worry about it)
+
+Note that you can not set current id names and prefer options in one ``/setid`` command simultaneously, but call them separately.
+
+### /listid
+
+List all registered identity names, default identity names and prefer options.
+
+### /delid ``name``
+
+Delete specific identity name.
+
+### /delall 
+
+Delete all registered names.
+
+### /swapid ``name(s)`` ``upload-image``
+
+Replace the face with the registered identity name(s) on target image.
+
+## Discord Context Menu
+
+### Apps/INSwapper
+
+Replace the face with the current/default identity name(s) on target image. Current/default identity name(s) can be set via ``/savevid`` and ``/setid`` slash commands.
+
+
+   
+
+## Step-by-step guide:
+
+1. Refer to [this link](https://docs.midjourney.com/docs/invite-the-bot) to register Discord app, create a new chat room, and invite the Midjourney bot to the chat room.
+2. Invite the InsightFaceSwap bot to the chat room by this link: <https://discord.com/api/oauth2/authorize?client_id=1090660574196674713&permissions=274877945856&scope=bot>.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd1.jpg" width="480"/>
+</div>
+3. Use ``/saveid`` command to register your identity name and feature. Here 'mnls' is the registered name, which can be any alphabets or numbers up to 8 characters long. If everything goes well, the bot will tell you that the save was successful. Note that the newly created identity will be automatically set as the default identity.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd2.jpg" width="640"/>
+</div>
+4. Next, we can experiment with creating the portrait. Let's start chanting the Midjourney prompt and enlarge one of the outputs.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd3.jpg" width="640"/>
+</div>
+5. After the enlargement is complete, we can simply use the ``INSwapper`` context menu to generate our portrait. Right click on the target image and then select ``Apps-INSwapper`` menu. Note that we can also use ``/setid`` command to change the default identity name.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd4.jpg" width="640"/>
+</div>
+6. Generally, the task is completed in less than a second and we can see the result.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd5.jpg" width="640"/>
+</div>
+7. In addition to processing photos generated by Midjourney, we can also process locally uploaded photos by using ``/swapid`` command explicitly.
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd6.jpg" width="640"/>
+</div>
+8. Hit to complete!
+<div align="left">
+<img src="https://raw.githubusercontent.com/nttstar/insightface-resources/master/images/swapd7.jpg" width="640"/>
+</div>
+9. Note that the ``INSwapper`` context menu can also work on user uploaded images in your Discord channel.
+
+
+## FAQ
+
+Q: Why "application did not respond"?
+
+A: This error indicates that the server was overloaded at the time. Please try again.
+
+Q: Why is the service sometimes slow to respond?
+
+A: We used a queue in our backend. When there are too many users online, the requests will be queued and processed one by one.
+
+Q: Can I list my registered ID list?
+
+A: Yes, use ``/listid`` command.
+
+Q: Are there any restrictions on ID names?
+
+A: All ID names can only be alphabets and numbers, and cannot exceed 10 characters. The total number of registered IDs cannot exceed 20.
+
+Q: Can I delete my registered IDs?
+
+A: You can use ``/delid`` and ``/delall`` commands to delete registered IDs.
+
+Q: Support multi-facial replacement?
+
+A: Yes, you can input a comma splitted idname list, such as ``/setid me,you,him,her``. You can also use the ``_`` symbol to indicate no-replacement(e.g. ``/setid me,_,him``).
+
+Q: How to get good results?
+
+A: 1) Select front-view, high quality, no glasses, no heavy bangs ID photos; 2) Try greedy mode if you need higher identity similarity; 3) For the target image, please ensure that the facial features are proportionate to those of real humans, otherwise it may cause overflow effects.
+
+## Other notes:
+
+1. Front-view, high quality, no glasses, no heavy bangs ID photos are prefered.
+2. Each Discord account can execute 50 commands per day.
+3. This is in early development stage, so we cannot guarantee that the result will be great in every cases.
+4. If there's any problem, please join our Discord group: [link](https://discord.gg/Ym3X8U59ZN)
+
+
diff --git a/web-demos/swapping_discord/privacy.md b/web-demos/swapping_discord/privacy.md
new file mode 100644
index 000000000..58f1520f9
--- /dev/null
+++ b/web-demos/swapping_discord/privacy.md
@@ -0,0 +1,24 @@
+## Privacy Policy
+This Privacy Policy describes how your personal information is collected and used in our Discord bot application.
+
+## Information We Collect
+When you visit our bot, we collect certain information about your USER-ID and your manually uploaded identity features, the identity images will not be saved.
+We collect the information using the API of Discord application.
+
+## How We Use Your Information
+We use the information that we collect to help process the image generation.
+
+## Sharing Your Information
+We will not share your information with any external organizations, individuals, or companies.
+
+## Behavioural Advertising
+We will not do any advertisements or marketing communications by using your information.
+
+## Your Rights
+You can list your saved information and delete your information in any time.
+
+## Changes
+We may update this privacy policy from time to time in order to reflect, for example, changes to our practices or for other operational, legal or regulatory reasons.
+
+## Contact Us
+For more information about our privacy practices, if you have questions, or if you would like to make a complaint, please email us at ``contact#insightface.ai``.